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 /**
 607  * do_sync_erase - run the erase worker synchronously.
 608  * @ubi: UBI device description object
 609  * @e: the WL entry of the physical eraseblock to erase
 610  * @vol_id: the volume ID that last used this PEB
 611  * @lnum: the last used logical eraseblock number for the PEB
 612  * @torture: if the physical eraseblock has to be tortured
 613  *
 614  */
 615 static int do_sync_erase(struct ubi_device *ubi, struct ubi_wl_entry *e,
 616                          int vol_id, int lnum, int torture)
 617 {
 618         struct ubi_work *wl_wrk;
 619
 620         dbg_wl("sync erase of PEB %i", e->pnum);
 621
 622         wl_wrk = kmalloc(sizeof(struct ubi_work), GFP_NOFS);
 623         if (!wl_wrk)
 624                 return -ENOMEM;
 625
 626         wl_wrk->e = e;
 627         wl_wrk->vol_id = vol_id;
 628         wl_wrk->lnum = lnum;
 629         wl_wrk->torture = torture;
 630
 631         return erase_worker(ubi, wl_wrk, 0);
 632 }
 633
 634 /**
 635  * wear_leveling_worker - wear-leveling worker function.
 636  * @ubi: UBI device description object
 637  * @wrk: the work object
 638  * @shutdown: non-zero if the worker has to free memory and exit
 639  * because the WL-subsystem is shutting down
 640  *
 641  * This function copies a more worn out physical eraseblock to a less worn out
 642  * one. Returns zero in case of success and a negative error code in case of
 643  * failure.
 644  */
 645 static int wear_leveling_worker(struct ubi_device *ubi, struct ubi_work *wrk,
 646                                 int shutdown)
 647 {
 648         int err, scrubbing = 0, torture = 0, protect = 0, erroneous = 0;
 649         int vol_id = -1, lnum = -1;
 650 #ifdef CONFIG_MTD_UBI_FASTMAP
 651         int anchor = wrk->anchor;
 652 #endif
 653         struct ubi_wl_entry *e1, *e2;
 654         struct ubi_vid_hdr *vid_hdr;
 655
 656         kfree(wrk);
 657         if (shutdown)
 658                 return 0;
 659
 660         vid_hdr = ubi_zalloc_vid_hdr(ubi, GFP_NOFS);
 661         if (!vid_hdr)
 662                 return -ENOMEM;
 663
 664         mutex_lock(&ubi->move_mutex);
 665         spin_lock(&ubi->wl_lock);
 666         ubi_assert(!ubi->move_from && !ubi->move_to);
 667         ubi_assert(!ubi->move_to_put);
 668
 669         if (!ubi->free.rb_node ||
 670             (!ubi->used.rb_node && !ubi->scrub.rb_node)) {
 671                 /*
 672                  * No free physical eraseblocks? Well, they must be waiting in
 673                  * the queue to be erased. Cancel movement - it will be
 674                  * triggered again when a free physical eraseblock appears.
 675                  *
 676                  * No used physical eraseblocks? They must be temporarily
 677                  * protected from being moved. They will be moved to the
 678                  * @ubi->used tree later and the wear-leveling will be
 679                  * triggered again.
 680                  */
 681                 dbg_wl("cancel WL, a list is empty: free %d, used %d",
 682                        !ubi->free.rb_node, !ubi->used.rb_node);
 683                 goto out_cancel;
 684         }
 685
 686 #ifdef CONFIG_MTD_UBI_FASTMAP
 687         /* Check whether we need to produce an anchor PEB */
 688         if (!anchor)
 689                 anchor = !anchor_pebs_avalible(&ubi->free);
 690
 691         if (anchor) {
 692                 e1 = find_anchor_wl_entry(&ubi->used);
 693                 if (!e1)
 694                         goto out_cancel;
 695                 e2 = get_peb_for_wl(ubi);
 696                 if (!e2)
 697                         goto out_cancel;
 698
 699                 self_check_in_wl_tree(ubi, e1, &ubi->used);
 700                 rb_erase(&e1->u.rb, &ubi->used);
 701                 dbg_wl("anchor-move PEB %d to PEB %d", e1->pnum, e2->pnum);
 702         } else if (!ubi->scrub.rb_node) {
 703 #else
 704         if (!ubi->scrub.rb_node) {
 705 #endif
 706                 /*
 707                  * Now pick the least worn-out used physical eraseblock and a
 708                  * highly worn-out free physical eraseblock. If the erase
 709                  * counters differ much enough, start wear-leveling.
 710                  */
 711                 e1 = rb_entry(rb_first(&ubi->used), struct ubi_wl_entry, u.rb);
 712                 e2 = get_peb_for_wl(ubi);
 713                 if (!e2)
 714                         goto out_cancel;
 715
 716                 if (!(e2->ec - e1->ec >= UBI_WL_THRESHOLD)) {
 717                         dbg_wl("no WL needed: min used EC %d, max free EC %d",
 718                                e1->ec, e2->ec);
 719
 720                         /* Give the unused PEB back */
 721                         wl_tree_add(e2, &ubi->free);
 722                         ubi->free_count++;
 723                         goto out_cancel;
 724                 }
 725                 self_check_in_wl_tree(ubi, e1, &ubi->used);
 726                 rb_erase(&e1->u.rb, &ubi->used);
 727                 dbg_wl("move PEB %d EC %d to PEB %d EC %d",
 728                        e1->pnum, e1->ec, e2->pnum, e2->ec);
 729         } else {
 730                 /* Perform scrubbing */
 731                 scrubbing = 1;
 732                 e1 = rb_entry(rb_first(&ubi->scrub), struct ubi_wl_entry, u.rb);
 733                 e2 = get_peb_for_wl(ubi);
 734                 if (!e2)
 735                         goto out_cancel;
 736
 737                 self_check_in_wl_tree(ubi, e1, &ubi->scrub);
 738                 rb_erase(&e1->u.rb, &ubi->scrub);
 739                 dbg_wl("scrub PEB %d to PEB %d", e1->pnum, e2->pnum);
 740         }
 741
 742         ubi->move_from = e1;
 743         ubi->move_to = e2;
 744         spin_unlock(&ubi->wl_lock);
 745
 746         /*
 747          * Now we are going to copy physical eraseblock @e1->pnum to @e2->pnum.
 748          * We so far do not know which logical eraseblock our physical
 749          * eraseblock (@e1) belongs to. We have to read the volume identifier
 750          * header first.
 751          *
 752          * Note, we are protected from this PEB being unmapped and erased. The
 753          * 'ubi_wl_put_peb()' would wait for moving to be finished if the PEB
 754          * which is being moved was unmapped.
 755          */
 756
 757         err = ubi_io_read_vid_hdr(ubi, e1->pnum, vid_hdr, 0);
 758         if (err && err != UBI_IO_BITFLIPS) {
 759                 if (err == UBI_IO_FF) {
 760                         /*
 761                          * We are trying to move PEB without a VID header. UBI
 762                          * always write VID headers shortly after the PEB was
 763                          * given, so we have a situation when it has not yet
 764                          * had a chance to write it, because it was preempted.
 765                          * So add this PEB to the protection queue so far,
 766                          * because presumably more data will be written there
 767                          * (including the missing VID header), and then we'll
 768                          * move it.
 769                          */
 770                         dbg_wl("PEB %d has no VID header", e1->pnum);
 771                         protect = 1;
 772                         goto out_not_moved;
 773                 } else if (err == UBI_IO_FF_BITFLIPS) {
 774                         /*
 775                          * The same situation as %UBI_IO_FF, but bit-flips were
 776                          * detected. It is better to schedule this PEB for
 777                          * scrubbing.
 778                          */
 779                         dbg_wl("PEB %d has no VID header but has bit-flips",
 780                                e1->pnum);
 781                         scrubbing = 1;
 782                         goto out_not_moved;
 783                 }
 784
 785                 ubi_err(ubi, "error %d while reading VID header from PEB %d",
 786                         err, e1->pnum);
 787                 goto out_error;
 788         }
 789
 790         vol_id = be32_to_cpu(vid_hdr->vol_id);
 791         lnum = be32_to_cpu(vid_hdr->lnum);
 792
 793         err = ubi_eba_copy_leb(ubi, e1->pnum, e2->pnum, vid_hdr);
 794         if (err) {
 795                 if (err == MOVE_CANCEL_RACE) {
 796                         /*
 797                          * The LEB has not been moved because the volume is
 798                          * being deleted or the PEB has been put meanwhile. We
 799                          * should prevent this PEB from being selected for
 800                          * wear-leveling movement again, so put it to the
 801                          * protection queue.
 802                          */
 803                         protect = 1;
 804                         goto out_not_moved;
 805                 }
 806                 if (err == MOVE_RETRY) {
 807                         scrubbing = 1;
 808                         goto out_not_moved;
 809                 }
 810                 if (err == MOVE_TARGET_BITFLIPS || err == MOVE_TARGET_WR_ERR ||
 811                     err == MOVE_TARGET_RD_ERR) {
 812                         /*
 813                          * Target PEB had bit-flips or write error - torture it.
 814                          */
 815                         torture = 1;
 816                         goto out_not_moved;
 817                 }
 818
 819                 if (err == MOVE_SOURCE_RD_ERR) {
 820                         /*
 821                          * An error happened while reading the source PEB. Do
 822                          * not switch to R/O mode in this case, and give the
 823                          * upper layers a possibility to recover from this,
 824                          * e.g. by unmapping corresponding LEB. Instead, just
 825                          * put this PEB to the @ubi->erroneous list to prevent
 826                          * UBI from trying to move it over and over again.
 827                          */
 828                         if (ubi->erroneous_peb_count > ubi->max_erroneous) {
 829                                 ubi_err(ubi, "too many erroneous eraseblocks (%d)",
 830                                         ubi->erroneous_peb_count);
 831                                 goto out_error;
 832                         }
 833                         erroneous = 1;
 834                         goto out_not_moved;
 835                 }
 836
 837                 if (err < 0)
 838                         goto out_error;
 839
 840                 ubi_assert(0);
 841         }
 842
 843         /* The PEB has been successfully moved */
 844         if (scrubbing)
 845                 ubi_msg(ubi, "scrubbed PEB %d (LEB %d:%d), data moved to PEB %d",
 846                         e1->pnum, vol_id, lnum, e2->pnum);
 847         ubi_free_vid_hdr(ubi, vid_hdr);
 848
 849         spin_lock(&ubi->wl_lock);
 850         if (!ubi->move_to_put) {
 851                 wl_tree_add(e2, &ubi->used);
 852                 e2 = NULL;
 853         }
 854         ubi->move_from = ubi->move_to = NULL;
 855         ubi->move_to_put = ubi->wl_scheduled = 0;
 856         spin_unlock(&ubi->wl_lock);
 857
 858         err = do_sync_erase(ubi, e1, vol_id, lnum, 0);
 859         if (err) {
 860                 if (e2)
 861                         wl_entry_destroy(ubi, e2);
 862                 goto out_ro;
 863         }
 864
 865         if (e2) {
 866                 /*
 867                  * Well, the target PEB was put meanwhile, schedule it for
 868                  * erasure.
 869                  */
 870                 dbg_wl("PEB %d (LEB %d:%d) was put meanwhile, erase",
 871                        e2->pnum, vol_id, lnum);
 872                 err = do_sync_erase(ubi, e2, vol_id, lnum, 0);
 873                 if (err)
 874                         goto out_ro;
 875         }
 876
 877         dbg_wl("done");
 878         mutex_unlock(&ubi->move_mutex);
 879         return 0;
 880
 881         /*
 882          * For some reasons the LEB was not moved, might be an error, might be
 883          * something else. @e1 was not changed, so return it back. @e2 might
 884          * have been changed, schedule it for erasure.
 885          */
 886 out_not_moved:
 887         if (vol_id != -1)
 888                 dbg_wl("cancel moving PEB %d (LEB %d:%d) to PEB %d (%d)",
 889                        e1->pnum, vol_id, lnum, e2->pnum, err);
 890         else
 891                 dbg_wl("cancel moving PEB %d to PEB %d (%d)",
 892                        e1->pnum, e2->pnum, err);
 893         spin_lock(&ubi->wl_lock);
 894         if (protect)
 895                 prot_queue_add(ubi, e1);
 896         else if (erroneous) {
 897                 wl_tree_add(e1, &ubi->erroneous);
 898                 ubi->erroneous_peb_count += 1;
 899         } else if (scrubbing)
 900                 wl_tree_add(e1, &ubi->scrub);
 901         else
 902                 wl_tree_add(e1, &ubi->used);
 903         ubi_assert(!ubi->move_to_put);
 904         ubi->move_from = ubi->move_to = NULL;
 905         ubi->wl_scheduled = 0;
 906         spin_unlock(&ubi->wl_lock);
 907
 908         ubi_free_vid_hdr(ubi, vid_hdr);
 909         err = do_sync_erase(ubi, e2, vol_id, lnum, torture);
 910         if (err)
 911                 goto out_ro;
 912
 913         mutex_unlock(&ubi->move_mutex);
 914         return 0;
 915
 916 out_error:
 917         if (vol_id != -1)
 918                 ubi_err(ubi, "error %d while moving PEB %d to PEB %d",
 919                         err, e1->pnum, e2->pnum);
 920         else
 921                 ubi_err(ubi, "error %d while moving PEB %d (LEB %d:%d) to PEB %d",
 922                         err, e1->pnum, vol_id, lnum, e2->pnum);
 923         spin_lock(&ubi->wl_lock);
 924         ubi->move_from = ubi->move_to = NULL;
 925         ubi->move_to_put = ubi->wl_scheduled = 0;
 926         spin_unlock(&ubi->wl_lock);
 927
 928         ubi_free_vid_hdr(ubi, vid_hdr);
 929         wl_entry_destroy(ubi, e1);
 930         wl_entry_destroy(ubi, e2);
 931
 932 out_ro:
 933         ubi_ro_mode(ubi);
 934         mutex_unlock(&ubi->move_mutex);
 935         ubi_assert(err != 0);
 936         return err < 0 ? err : -EIO;
 937
 938 out_cancel:
 939         ubi->wl_scheduled = 0;
 940         spin_unlock(&ubi->wl_lock);
 941         mutex_unlock(&ubi->move_mutex);
 942         ubi_free_vid_hdr(ubi, vid_hdr);
 943         return 0;
 944 }
 945
 946 /**
 947  * ensure_wear_leveling - schedule wear-leveling if it is needed.
 948  * @ubi: UBI device description object
 949  * @nested: set to non-zero if this function is called from UBI worker
 950  *
 951  * This function checks if it is time to start wear-leveling and schedules it
 952  * if yes. This function returns zero in case of success and a negative error
 953  * code in case of failure.
 954  */
 955 static int ensure_wear_leveling(struct ubi_device *ubi, int nested)
 956 {
 957         int err = 0;
 958         struct ubi_wl_entry *e1;
 959         struct ubi_wl_entry *e2;
 960         struct ubi_work *wrk;
 961
 962         spin_lock(&ubi->wl_lock);
 963         if (ubi->wl_scheduled)
 964                 /* Wear-leveling is already in the work queue */
 965                 goto out_unlock;
 966
 967         /*
 968          * If the ubi->scrub tree is not empty, scrubbing is needed, and the
 969          * the WL worker has to be scheduled anyway.
 970          */
 971         if (!ubi->scrub.rb_node) {
 972                 if (!ubi->used.rb_node || !ubi->free.rb_node)
 973                         /* No physical eraseblocks - no deal */
 974                         goto out_unlock;
 975
 976                 /*
 977                  * We schedule wear-leveling only if the difference between the
 978                  * lowest erase counter of used physical eraseblocks and a high
 979                  * erase counter of free physical eraseblocks is greater than
 980                  * %UBI_WL_THRESHOLD.
 981                  */
 982                 e1 = rb_entry(rb_first(&ubi->used), struct ubi_wl_entry, u.rb);
 983                 e2 = find_wl_entry(ubi, &ubi->free, WL_FREE_MAX_DIFF);
 984
 985                 if (!(e2->ec - e1->ec >= UBI_WL_THRESHOLD))
 986                         goto out_unlock;
 987                 dbg_wl("schedule wear-leveling");
 988         } else
 989                 dbg_wl("schedule scrubbing");
 990
 991         ubi->wl_scheduled = 1;
 992         spin_unlock(&ubi->wl_lock);
 993
 994         wrk = kmalloc(sizeof(struct ubi_work), GFP_NOFS);
 995         if (!wrk) {
 996                 err = -ENOMEM;
 997                 goto out_cancel;
 998         }
 999
1000         wrk->anchor = 0;
1001         wrk->func = &wear_leveling_worker;
1002         if (nested)
1003                 __schedule_ubi_work(ubi, wrk);
1004         else
1005                 schedule_ubi_work(ubi, wrk);
1006         return err;
1007
1008 out_cancel:
1009         spin_lock(&ubi->wl_lock);
1010         ubi->wl_scheduled = 0;
1011 out_unlock:
1012         spin_unlock(&ubi->wl_lock);
1013         return err;
1014 }
1015
1016 /**
1017  * erase_worker - physical eraseblock erase worker function.
1018  * @ubi: UBI device description object
1019  * @wl_wrk: the work object
1020  * @shutdown: non-zero if the worker has to free memory and exit
1021  * because the WL sub-system is shutting down
1022  *
1023  * This function erases a physical eraseblock and perform torture testing if
1024  * needed. It also takes care about marking the physical eraseblock bad if
1025  * needed. Returns zero in case of success and a negative error code in case of
1026  * failure.
1027  */
1028 static int erase_worker(struct ubi_device *ubi, struct ubi_work *wl_wrk,
1029                         int shutdown)
1030 {
1031         struct ubi_wl_entry *e = wl_wrk->e;
1032         int pnum = e->pnum;
1033         int vol_id = wl_wrk->vol_id;
1034         int lnum = wl_wrk->lnum;
1035         int err, available_consumed = 0;
1036
1037         if (shutdown) {
1038                 dbg_wl("cancel erasure of PEB %d EC %d", pnum, e->ec);
1039                 kfree(wl_wrk);
1040                 wl_entry_destroy(ubi, e);
1041                 return 0;
1042         }
1043
1044         dbg_wl("erase PEB %d EC %d LEB %d:%d",
1045                pnum, e->ec, wl_wrk->vol_id, wl_wrk->lnum);
1046
1047         err = sync_erase(ubi, e, wl_wrk->torture);
1048         if (!err) {
1049                 /* Fine, we've erased it successfully */
1050                 kfree(wl_wrk);
1051
1052                 spin_lock(&ubi->wl_lock);
1053                 wl_tree_add(e, &ubi->free);
1054                 ubi->free_count++;
1055                 spin_unlock(&ubi->wl_lock);
1056
1057                 /*
1058                  * One more erase operation has happened, take care about
1059                  * protected physical eraseblocks.
1060                  */
1061                 serve_prot_queue(ubi);
1062
1063                 /* And take care about wear-leveling */
1064                 err = ensure_wear_leveling(ubi, 1);
1065                 return err;
1066         }
1067
1068         ubi_err(ubi, "failed to erase PEB %d, error %d", pnum, err);
1069         kfree(wl_wrk);
1070
1071         if (err == -EINTR || err == -ENOMEM || err == -EAGAIN ||
1072             err == -EBUSY) {
1073                 int err1;
1074
1075                 /* Re-schedule the LEB for erasure */
1076                 err1 = schedule_erase(ubi, e, vol_id, lnum, 0);
1077                 if (err1) {
1078                         err = err1;
1079                         goto out_ro;
1080                 }
1081                 return err;
1082         }
1083
1084         wl_entry_destroy(ubi, e);
1085         if (err != -EIO)
1086                 /*
1087                  * If this is not %-EIO, we have no idea what to do. Scheduling
1088                  * this physical eraseblock for erasure again would cause
1089                  * errors again and again. Well, lets switch to R/O mode.
1090                  */
1091                 goto out_ro;
1092
1093         /* It is %-EIO, the PEB went bad */
1094
1095         if (!ubi->bad_allowed) {
1096                 ubi_err(ubi, "bad physical eraseblock %d detected", pnum);
1097                 goto out_ro;
1098         }
1099
1100         spin_lock(&ubi->volumes_lock);
1101         if (ubi->beb_rsvd_pebs == 0) {
1102                 if (ubi->avail_pebs == 0) {
1103                         spin_unlock(&ubi->volumes_lock);
1104                         ubi_err(ubi, "no reserved/available physical eraseblocks");
1105                         goto out_ro;
1106                 }
1107                 ubi->avail_pebs -= 1;
1108                 available_consumed = 1;
1109         }
1110         spin_unlock(&ubi->volumes_lock);
1111
1112         ubi_msg(ubi, "mark PEB %d as bad", pnum);
1113         err = ubi_io_mark_bad(ubi, pnum);
1114         if (err)
1115                 goto out_ro;
1116
1117         spin_lock(&ubi->volumes_lock);
1118         if (ubi->beb_rsvd_pebs > 0) {
1119                 if (available_consumed) {
1120                         /*
1121                          * The amount of reserved PEBs increased since we last
1122                          * checked.
1123                          */
1124                         ubi->avail_pebs += 1;
1125                         available_consumed = 0;
1126                 }
1127                 ubi->beb_rsvd_pebs -= 1;
1128         }
1129         ubi->bad_peb_count += 1;
1130         ubi->good_peb_count -= 1;
1131         ubi_calculate_reserved(ubi);
1132         if (available_consumed)
1133                 ubi_warn(ubi, "no PEBs in the reserved pool, used an available PEB");
1134         else if (ubi->beb_rsvd_pebs)
1135                 ubi_msg(ubi, "%d PEBs left in the reserve",
1136                         ubi->beb_rsvd_pebs);
1137         else
1138                 ubi_warn(ubi, "last PEB from the reserve was used");
1139         spin_unlock(&ubi->volumes_lock);
1140
1141         return err;
1142
1143 out_ro:
1144         if (available_consumed) {
1145                 spin_lock(&ubi->volumes_lock);
1146                 ubi->avail_pebs += 1;
1147                 spin_unlock(&ubi->volumes_lock);
1148         }
1149         ubi_ro_mode(ubi);
1150         return err;
1151 }
1152
1153 /**
1154  * ubi_wl_put_peb - return a PEB to the wear-leveling sub-system.
1155  * @ubi: UBI device description object
1156  * @vol_id: the volume ID that last used this PEB
1157  * @lnum: the last used logical eraseblock number for the PEB
1158  * @pnum: physical eraseblock to return
1159  * @torture: if this physical eraseblock has to be tortured
1160  *
1161  * This function is called to return physical eraseblock @pnum to the pool of
1162  * free physical eraseblocks. The @torture flag has to be set if an I/O error
1163  * occurred to this @pnum and it has to be tested. This function returns zero
1164  * in case of success, and a negative error code in case of failure.
1165  */
1166 int ubi_wl_put_peb(struct ubi_device *ubi, int vol_id, int lnum,
1167                    int pnum, int torture)
1168 {
1169         int err;
1170         struct ubi_wl_entry *e;
1171
1172         dbg_wl("PEB %d", pnum);
1173         ubi_assert(pnum >= 0);
1174         ubi_assert(pnum < ubi->peb_count);
1175
1176         down_read(&ubi->fm_protect);
1177
1178 retry:
1179         spin_lock(&ubi->wl_lock);
1180         e = ubi->lookuptbl[pnum];
1181         if (e == ubi->move_from) {
1182                 /*
1183                  * User is putting the physical eraseblock which was selected to
1184                  * be moved. It will be scheduled for erasure in the
1185                  * wear-leveling worker.
1186                  */
1187                 dbg_wl("PEB %d is being moved, wait", pnum);
1188                 spin_unlock(&ubi->wl_lock);
1189
1190                 /* Wait for the WL worker by taking the @ubi->move_mutex */
1191                 mutex_lock(&ubi->move_mutex);
1192                 mutex_unlock(&ubi->move_mutex);
1193                 goto retry;
1194         } else if (e == ubi->move_to) {
1195                 /*
1196                  * User is putting the physical eraseblock which was selected
1197                  * as the target the data is moved to. It may happen if the EBA
1198                  * sub-system already re-mapped the LEB in 'ubi_eba_copy_leb()'
1199                  * but the WL sub-system has not put the PEB to the "used" tree
1200                  * yet, but it is about to do this. So we just set a flag which
1201                  * will tell the WL worker that the PEB is not needed anymore
1202                  * and should be scheduled for erasure.
1203                  */
1204                 dbg_wl("PEB %d is the target of data moving", pnum);
1205                 ubi_assert(!ubi->move_to_put);
1206                 ubi->move_to_put = 1;
1207                 spin_unlock(&ubi->wl_lock);
1208                 up_read(&ubi->fm_protect);
1209                 return 0;
1210         } else {
1211                 if (in_wl_tree(e, &ubi->used)) {
1212                         self_check_in_wl_tree(ubi, e, &ubi->used);
1213                         rb_erase(&e->u.rb, &ubi->used);
1214                 } else if (in_wl_tree(e, &ubi->scrub)) {
1215                         self_check_in_wl_tree(ubi, e, &ubi->scrub);
1216                         rb_erase(&e->u.rb, &ubi->scrub);
1217                 } else if (in_wl_tree(e, &ubi->erroneous)) {
1218                         self_check_in_wl_tree(ubi, e, &ubi->erroneous);
1219                         rb_erase(&e->u.rb, &ubi->erroneous);
1220                         ubi->erroneous_peb_count -= 1;
1221                         ubi_assert(ubi->erroneous_peb_count >= 0);
1222                         /* Erroneous PEBs should be tortured */
1223                         torture = 1;
1224                 } else {
1225                         err = prot_queue_del(ubi, e->pnum);
1226                         if (err) {
1227                                 ubi_err(ubi, "PEB %d not found", pnum);
1228                                 ubi_ro_mode(ubi);
1229                                 spin_unlock(&ubi->wl_lock);
1230                                 up_read(&ubi->fm_protect);
1231                                 return err;
1232                         }
1233                 }
1234         }
1235         spin_unlock(&ubi->wl_lock);
1236
1237         err = schedule_erase(ubi, e, vol_id, lnum, torture);
1238         if (err) {
1239                 spin_lock(&ubi->wl_lock);
1240                 wl_tree_add(e, &ubi->used);
1241                 spin_unlock(&ubi->wl_lock);
1242         }
1243
1244         up_read(&ubi->fm_protect);
1245         return err;
1246 }
1247
1248 /**
1249  * ubi_wl_scrub_peb - schedule a physical eraseblock for scrubbing.
1250  * @ubi: UBI device description object
1251  * @pnum: the physical eraseblock to schedule
1252  *
1253  * If a bit-flip in a physical eraseblock is detected, this physical eraseblock
1254  * needs scrubbing. This function schedules a physical eraseblock for
1255  * scrubbing which is done in background. This function returns zero in case of
1256  * success and a negative error code in case of failure.
1257  */
1258 int ubi_wl_scrub_peb(struct ubi_device *ubi, int pnum)
1259 {
1260         struct ubi_wl_entry *e;
1261
1262         ubi_msg(ubi, "schedule PEB %d for scrubbing", pnum);
1263
1264 retry:
1265         spin_lock(&ubi->wl_lock);
1266         e = ubi->lookuptbl[pnum];
1267         if (e == ubi->move_from || in_wl_tree(e, &ubi->scrub) ||
1268                                    in_wl_tree(e, &ubi->erroneous)) {
1269                 spin_unlock(&ubi->wl_lock);
1270                 return 0;
1271         }
1272
1273         if (e == ubi->move_to) {
1274                 /*
1275                  * This physical eraseblock was used to move data to. The data
1276                  * was moved but the PEB was not yet inserted to the proper
1277                  * tree. We should just wait a little and let the WL worker
1278                  * proceed.
1279                  */
1280                 spin_unlock(&ubi->wl_lock);
1281                 dbg_wl("the PEB %d is not in proper tree, retry", pnum);
1282                 yield();
1283                 goto retry;
1284         }
1285
1286         if (in_wl_tree(e, &ubi->used)) {
1287                 self_check_in_wl_tree(ubi, e, &ubi->used);
1288                 rb_erase(&e->u.rb, &ubi->used);
1289         } else {
1290                 int err;
1291
1292                 err = prot_queue_del(ubi, e->pnum);
1293                 if (err) {
1294                         ubi_err(ubi, "PEB %d not found", pnum);
1295                         ubi_ro_mode(ubi);
1296                         spin_unlock(&ubi->wl_lock);
1297                         return err;
1298                 }
1299         }
1300
1301         wl_tree_add(e, &ubi->scrub);
1302         spin_unlock(&ubi->wl_lock);
1303
1304         /*
1305          * Technically scrubbing is the same as wear-leveling, so it is done
1306          * by the WL worker.
1307          */
1308         return ensure_wear_leveling(ubi, 0);
1309 }
1310
1311 /**
1312  * ubi_wl_flush - flush all pending works.
1313  * @ubi: UBI device description object
1314  * @vol_id: the volume id to flush for
1315  * @lnum: the logical eraseblock number to flush for
1316  *
1317  * This function executes all pending works for a particular volume id /
1318  * logical eraseblock number pair. If either value is set to %UBI_ALL, then it
1319  * acts as a wildcard for all of the corresponding volume numbers or logical
1320  * eraseblock numbers. It returns zero in case of success and a negative error
1321  * code in case of failure.
1322  */
1323 int ubi_wl_flush(struct ubi_device *ubi, int vol_id, int lnum)
1324 {
1325         int err = 0;
1326         int found = 1;
1327
1328         /*
1329          * Erase while the pending works queue is not empty, but not more than
1330          * the number of currently pending works.
1331          */
1332         dbg_wl("flush pending work for LEB %d:%d (%d pending works)",
1333                vol_id, lnum, ubi->works_count);
1334
1335         while (found) {
1336                 struct ubi_work *wrk, *tmp;
1337                 found = 0;
1338
1339                 down_read(&ubi->work_sem);
1340                 spin_lock(&ubi->wl_lock);
1341                 list_for_each_entry_safe(wrk, tmp, &ubi->works, list) {
1342                         if ((vol_id == UBI_ALL || wrk->vol_id == vol_id) &&
1343                             (lnum == UBI_ALL || wrk->lnum == lnum)) {
1344                                 list_del(&wrk->list);
1345                                 ubi->works_count -= 1;
1346                                 ubi_assert(ubi->works_count >= 0);
1347                                 spin_unlock(&ubi->wl_lock);
1348
1349                                 err = wrk->func(ubi, wrk, 0);
1350                                 if (err) {
1351                                         up_read(&ubi->work_sem);
1352                                         return err;
1353                                 }
1354
1355                                 spin_lock(&ubi->wl_lock);
1356                                 found = 1;
1357                                 break;
1358                         }
1359                 }
1360                 spin_unlock(&ubi->wl_lock);
1361                 up_read(&ubi->work_sem);
1362         }
1363
1364         /*
1365          * Make sure all the works which have been done in parallel are
1366          * finished.
1367          */
1368         down_write(&ubi->work_sem);
1369         up_write(&ubi->work_sem);
1370
1371         return err;
1372 }
1373
1374 /**
1375  * tree_destroy - destroy an RB-tree.
1376  * @ubi: UBI device description object
1377  * @root: the root of the tree to destroy
1378  */
1379 static void tree_destroy(struct ubi_device *ubi, struct rb_root *root)
1380 {
1381         struct rb_node *rb;
1382         struct ubi_wl_entry *e;
1383
1384         rb = root->rb_node;
1385         while (rb) {
1386                 if (rb->rb_left)
1387                         rb = rb->rb_left;
1388                 else if (rb->rb_right)
1389                         rb = rb->rb_right;
1390                 else {
1391                         e = rb_entry(rb, struct ubi_wl_entry, u.rb);
1392
1393                         rb = rb_parent(rb);
1394                         if (rb) {
1395                                 if (rb->rb_left == &e->u.rb)
1396                                         rb->rb_left = NULL;
1397                                 else
1398                                         rb->rb_right = NULL;
1399                         }
1400
1401                         wl_entry_destroy(ubi, e);
1402                 }
1403         }
1404 }
1405
1406 /**
1407  * ubi_thread - UBI background thread.
1408  * @u: the UBI device description object pointer
1409  */
1410 int ubi_thread(void *u)
1411 {
1412         int failures = 0;
1413         struct ubi_device *ubi = u;
1414
1415         ubi_msg(ubi, "background thread \"%s\" started, PID %d",
1416                 ubi->bgt_name, task_pid_nr(current));
1417
1418         set_freezable();
1419         for (;;) {
1420                 int err;
1421
1422                 if (kthread_should_stop())
1423                         break;
1424
1425                 if (try_to_freeze())
1426                         continue;
1427
1428                 spin_lock(&ubi->wl_lock);
1429                 if (list_empty(&ubi->works) || ubi->ro_mode ||
1430                     !ubi->thread_enabled || ubi_dbg_is_bgt_disabled(ubi)) {
1431                         set_current_state(TASK_INTERRUPTIBLE);
1432                         spin_unlock(&ubi->wl_lock);
1433                         schedule();
1434                         continue;
1435                 }
1436                 spin_unlock(&ubi->wl_lock);
1437
1438                 err = do_work(ubi);
1439                 if (err) {
1440                         ubi_err(ubi, "%s: work failed with error code %d",
1441                                 ubi->bgt_name, err);
1442                         if (failures++ > WL_MAX_FAILURES) {
1443                                 /*
1444                                  * Too many failures, disable the thread and
1445                                  * switch to read-only mode.
1446                                  */
1447                                 ubi_msg(ubi, "%s: %d consecutive failures",
1448                                         ubi->bgt_name, WL_MAX_FAILURES);
1449                                 ubi_ro_mode(ubi);
1450                                 ubi->thread_enabled = 0;
1451                                 continue;
1452                         }
1453                 } else
1454                         failures = 0;
1455
1456                 cond_resched();
1457         }
1458
1459         dbg_wl("background thread \"%s\" is killed", ubi->bgt_name);
1460         return 0;
1461 }
1462
1463 /**
1464  * shutdown_work - shutdown all pending works.
1465  * @ubi: UBI device description object
1466  */
1467 static void shutdown_work(struct ubi_device *ubi)
1468 {
1469 #ifdef CONFIG_MTD_UBI_FASTMAP
1470         flush_work(&ubi->fm_work);
1471 #endif
1472         while (!list_empty(&ubi->works)) {
1473                 struct ubi_work *wrk;
1474
1475                 wrk = list_entry(ubi->works.next, struct ubi_work, list);
1476                 list_del(&wrk->list);
1477                 wrk->func(ubi, wrk, 1);
1478                 ubi->works_count -= 1;
1479                 ubi_assert(ubi->works_count >= 0);
1480         }
1481 }
1482
1483 /**
1484  * ubi_wl_init - initialize the WL sub-system using attaching information.
1485  * @ubi: UBI device description object
1486  * @ai: attaching information
1487  *
1488  * This function returns zero in case of success, and a negative error code in
1489  * case of failure.
1490  */
1491 int ubi_wl_init(struct ubi_device *ubi, struct ubi_attach_info *ai)
1492 {
1493         int err, i, reserved_pebs, found_pebs = 0;
1494         struct rb_node *rb1, *rb2;
1495         struct ubi_ainf_volume *av;
1496         struct ubi_ainf_peb *aeb, *tmp;
1497         struct ubi_wl_entry *e;
1498
1499         ubi->used = ubi->erroneous = ubi->free = ubi->scrub = RB_ROOT;
1500         spin_lock_init(&ubi->wl_lock);
1501         mutex_init(&ubi->move_mutex);
1502         init_rwsem(&ubi->work_sem);
1503         ubi->max_ec = ai->max_ec;
1504         INIT_LIST_HEAD(&ubi->works);
1505
1506         sprintf(ubi->bgt_name, UBI_BGT_NAME_PATTERN, ubi->ubi_num);
1507
1508         err = -ENOMEM;
1509         ubi->lookuptbl = kzalloc(ubi->peb_count * sizeof(void *), GFP_KERNEL);
1510         if (!ubi->lookuptbl)
1511                 return err;
1512
1513         for (i = 0; i < UBI_PROT_QUEUE_LEN; i++)
1514                 INIT_LIST_HEAD(&ubi->pq[i]);
1515         ubi->pq_head = 0;
1516
1517         list_for_each_entry_safe(aeb, tmp, &ai->erase, u.list) {
1518                 cond_resched();
1519
1520                 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1521                 if (!e)
1522                         goto out_free;
1523
1524                 e->pnum = aeb->pnum;
1525                 e->ec = aeb->ec;
1526                 ubi->lookuptbl[e->pnum] = e;
1527                 if (schedule_erase(ubi, e, aeb->vol_id, aeb->lnum, 0)) {
1528                         wl_entry_destroy(ubi, e);
1529                         goto out_free;
1530                 }
1531
1532                 found_pebs++;
1533         }
1534
1535         ubi->free_count = 0;
1536         list_for_each_entry(aeb, &ai->free, u.list) {
1537                 cond_resched();
1538
1539                 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1540                 if (!e)
1541                         goto out_free;
1542
1543                 e->pnum = aeb->pnum;
1544                 e->ec = aeb->ec;
1545                 ubi_assert(e->ec >= 0);
1546
1547                 wl_tree_add(e, &ubi->free);
1548                 ubi->free_count++;
1549
1550                 ubi->lookuptbl[e->pnum] = e;
1551
1552                 found_pebs++;
1553         }
1554
1555         ubi_rb_for_each_entry(rb1, av, &ai->volumes, rb) {
1556                 ubi_rb_for_each_entry(rb2, aeb, &av->root, u.rb) {
1557                         cond_resched();
1558
1559                         e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1560                         if (!e)
1561                                 goto out_free;
1562
1563                         e->pnum = aeb->pnum;
1564                         e->ec = aeb->ec;
1565                         ubi->lookuptbl[e->pnum] = e;
1566
1567                         if (!aeb->scrub) {
1568                                 dbg_wl("add PEB %d EC %d to the used tree",
1569                                        e->pnum, e->ec);
1570                                 wl_tree_add(e, &ubi->used);
1571                         } else {
1572                                 dbg_wl("add PEB %d EC %d to the scrub tree",
1573                                        e->pnum, e->ec);
1574                                 wl_tree_add(e, &ubi->scrub);
1575                         }
1576
1577                         found_pebs++;
1578                 }
1579         }
1580
1581         dbg_wl("found %i PEBs", found_pebs);
1582
1583         if (ubi->fm) {
1584                 ubi_assert(ubi->good_peb_count ==
1585                            found_pebs + ubi->fm->used_blocks);
1586
1587                 for (i = 0; i < ubi->fm->used_blocks; i++) {
1588                         e = ubi->fm->e[i];
1589                         ubi->lookuptbl[e->pnum] = e;
1590                 }
1591         }
1592         else
1593                 ubi_assert(ubi->good_peb_count == found_pebs);
1594
1595         reserved_pebs = WL_RESERVED_PEBS;
1596         ubi_fastmap_init(ubi, &reserved_pebs);
1597
1598         if (ubi->avail_pebs < reserved_pebs) {
1599                 ubi_err(ubi, "no enough physical eraseblocks (%d, need %d)",
1600                         ubi->avail_pebs, reserved_pebs);
1601                 if (ubi->corr_peb_count)
1602                         ubi_err(ubi, "%d PEBs are corrupted and not used",
1603                                 ubi->corr_peb_count);
1604                 err = -ENOSPC;
1605                 goto out_free;
1606         }
1607         ubi->avail_pebs -= reserved_pebs;
1608         ubi->rsvd_pebs += reserved_pebs;
1609
1610         /* Schedule wear-leveling if needed */
1611         err = ensure_wear_leveling(ubi, 0);
1612         if (err)
1613                 goto out_free;
1614
1615         return 0;
1616
1617 out_free:
1618         shutdown_work(ubi);
1619         tree_destroy(ubi, &ubi->used);
1620         tree_destroy(ubi, &ubi->free);
1621         tree_destroy(ubi, &ubi->scrub);
1622         kfree(ubi->lookuptbl);
1623         return err;
1624 }
1625
1626 /**
1627  * protection_queue_destroy - destroy the protection queue.
1628  * @ubi: UBI device description object
1629  */
1630 static void protection_queue_destroy(struct ubi_device *ubi)
1631 {
1632         int i;
1633         struct ubi_wl_entry *e, *tmp;
1634
1635         for (i = 0; i < UBI_PROT_QUEUE_LEN; ++i) {
1636                 list_for_each_entry_safe(e, tmp, &ubi->pq[i], u.list) {
1637                         list_del(&e->u.list);
1638                         wl_entry_destroy(ubi, e);
1639                 }
1640         }
1641 }
1642
1643 /**
1644  * ubi_wl_close - close the wear-leveling sub-system.
1645  * @ubi: UBI device description object
1646  */
1647 void ubi_wl_close(struct ubi_device *ubi)
1648 {
1649         dbg_wl("close the WL sub-system");
1650         ubi_fastmap_close(ubi);
1651         shutdown_work(ubi);
1652         protection_queue_destroy(ubi);
1653         tree_destroy(ubi, &ubi->used);
1654         tree_destroy(ubi, &ubi->erroneous);
1655         tree_destroy(ubi, &ubi->free);
1656         tree_destroy(ubi, &ubi->scrub);
1657         kfree(ubi->lookuptbl);
1658 }
1659
1660 /**
1661  * self_check_ec - make sure that the erase counter of a PEB is correct.
1662  * @ubi: UBI device description object
1663  * @pnum: the physical eraseblock number to check
1664  * @ec: the erase counter to check
1665  *
1666  * This function returns zero if the erase counter of physical eraseblock @pnum
1667  * is equivalent to @ec, and a negative error code if not or if an error
1668  * occurred.
1669  */
1670 static int self_check_ec(struct ubi_device *ubi, int pnum, int ec)
1671 {
1672         int err;
1673         long long read_ec;
1674         struct ubi_ec_hdr *ec_hdr;
1675
1676         if (!ubi_dbg_chk_gen(ubi))
1677                 return 0;
1678
1679         ec_hdr = kzalloc(ubi->ec_hdr_alsize, GFP_NOFS);
1680         if (!ec_hdr)
1681                 return -ENOMEM;
1682
1683         err = ubi_io_read_ec_hdr(ubi, pnum, ec_hdr, 0);
1684         if (err && err != UBI_IO_BITFLIPS) {
1685                 /* The header does not have to exist */
1686                 err = 0;
1687                 goto out_free;
1688         }
1689
1690         read_ec = be64_to_cpu(ec_hdr->ec);
1691         if (ec != read_ec && read_ec - ec > 1) {
1692                 ubi_err(ubi, "self-check failed for PEB %d", pnum);
1693                 ubi_err(ubi, "read EC is %lld, should be %d", read_ec, ec);
1694                 dump_stack();
1695                 err = 1;
1696         } else
1697                 err = 0;
1698
1699 out_free:
1700         kfree(ec_hdr);
1701         return err;
1702 }
1703
1704 /**
1705  * self_check_in_wl_tree - check that wear-leveling entry is in WL RB-tree.
1706  * @ubi: UBI device description object
1707  * @e: the wear-leveling entry to check
1708  * @root: the root of the tree
1709  *
1710  * This function returns zero if @e is in the @root RB-tree and %-EINVAL if it
1711  * is not.
1712  */
1713 static int self_check_in_wl_tree(const struct ubi_device *ubi,
1714                                  struct ubi_wl_entry *e, struct rb_root *root)
1715 {
1716         if (!ubi_dbg_chk_gen(ubi))
1717                 return 0;
1718
1719         if (in_wl_tree(e, root))
1720                 return 0;
1721
1722         ubi_err(ubi, "self-check failed for PEB %d, EC %d, RB-tree %p ",
1723                 e->pnum, e->ec, root);
1724         dump_stack();
1725         return -EINVAL;
1726 }
1727
1728 /**
1729  * self_check_in_pq - check if wear-leveling entry is in the protection
1730  *                        queue.
1731  * @ubi: UBI device description object
1732  * @e: the wear-leveling entry to check
1733  *
1734  * This function returns zero if @e is in @ubi->pq and %-EINVAL if it is not.
1735  */
1736 static int self_check_in_pq(const struct ubi_device *ubi,
1737                             struct ubi_wl_entry *e)
1738 {
1739         struct ubi_wl_entry *p;
1740         int i;
1741
1742         if (!ubi_dbg_chk_gen(ubi))
1743                 return 0;
1744
1745         for (i = 0; i < UBI_PROT_QUEUE_LEN; ++i)
1746                 list_for_each_entry(p, &ubi->pq[i], u.list)
1747                         if (p == e)
1748                                 return 0;
1749
1750         ubi_err(ubi, "self-check failed for PEB %d, EC %d, Protect queue",
1751                 e->pnum, e->ec);
1752         dump_stack();
1753         return -EINVAL;
1754 }
1755 #ifndef CONFIG_MTD_UBI_FASTMAP
1756 static struct ubi_wl_entry *get_peb_for_wl(struct ubi_device *ubi)
1757 {
1758         struct ubi_wl_entry *e;
1759
1760         e = find_wl_entry(ubi, &ubi->free, WL_FREE_MAX_DIFF);
1761         self_check_in_wl_tree(ubi, e, &ubi->free);
1762         ubi->free_count--;
1763         ubi_assert(ubi->free_count >= 0);
1764         rb_erase(&e->u.rb, &ubi->free);
1765
1766         return e;
1767 }
1768
1769 /**
1770  * produce_free_peb - produce a free physical eraseblock.
1771  * @ubi: UBI device description object
1772  *
1773  * This function tries to make a free PEB by means of synchronous execution of
1774  * pending works. This may be needed if, for example the background thread is
1775  * disabled. Returns zero in case of success and a negative error code in case
1776  * of failure.
1777  */
1778 static int produce_free_peb(struct ubi_device *ubi)
1779 {
1780         int err;
1781
1782         while (!ubi->free.rb_node && ubi->works_count) {
1783                 spin_unlock(&ubi->wl_lock);
1784
1785                 dbg_wl("do one work synchronously");
1786                 err = do_work(ubi);
1787
1788                 spin_lock(&ubi->wl_lock);
1789                 if (err)
1790                         return err;
1791         }
1792
1793         return 0;
1794 }
1795
1796 /**
1797  * ubi_wl_get_peb - get a physical eraseblock.
1798  * @ubi: UBI device description object
1799  *
1800  * This function returns a physical eraseblock in case of success and a
1801  * negative error code in case of failure.
1802  * Returns with ubi->fm_eba_sem held in read mode!
1803  */
1804 int ubi_wl_get_peb(struct ubi_device *ubi)
1805 {
1806         int err;
1807         struct ubi_wl_entry *e;
1808
1809 retry:
1810         down_read(&ubi->fm_eba_sem);
1811         spin_lock(&ubi->wl_lock);
1812         if (!ubi->free.rb_node) {
1813                 if (ubi->works_count == 0) {
1814                         ubi_err(ubi, "no free eraseblocks");
1815                         ubi_assert(list_empty(&ubi->works));
1816                         spin_unlock(&ubi->wl_lock);
1817                         return -ENOSPC;
1818                 }
1819
1820                 err = produce_free_peb(ubi);
1821                 if (err < 0) {
1822                         spin_unlock(&ubi->wl_lock);
1823                         return err;
1824                 }
1825                 spin_unlock(&ubi->wl_lock);
1826                 up_read(&ubi->fm_eba_sem);
1827                 goto retry;
1828
1829         }
1830         e = wl_get_wle(ubi);
1831         prot_queue_add(ubi, e);
1832         spin_unlock(&ubi->wl_lock);
1833
1834         err = ubi_self_check_all_ff(ubi, e->pnum, ubi->vid_hdr_aloffset,
1835                                     ubi->peb_size - ubi->vid_hdr_aloffset);
1836         if (err) {
1837                 ubi_err(ubi, "new PEB %d does not contain all 0xFF bytes", e->pnum);
1838                 return err;
1839         }
1840
1841         return e->pnum;
1842 }
1843 #else
1844 #include "fastmap-wl.c"
1845 #endif