drivers/mtd/ubi/wl.c

   1 /*
   2  * Copyright (c) International Business Machines Corp., 2006
   3  *
   4  * SPDX-License-Identifier:     GPL-2.0+
   5  *
   6  * Authors: Artem Bityutskiy (Битюцкий Артём), Thomas Gleixner
   7  */
   8
   9 /*
  10  * UBI wear-leveling unit.
  11  *
  12  * This unit is responsible for wear-leveling. It works in terms of physical
  13  * eraseblocks and erase counters and knows nothing about logical eraseblocks,
  14  * volumes, etc. From this unit's perspective all physical eraseblocks are of
  15  * two types - used and free. Used physical eraseblocks are those that were
  16  * "get" by the 'ubi_wl_get_peb()' function, and free physical eraseblocks are
  17  * those that were put by the 'ubi_wl_put_peb()' function.
  18  *
  19  * Physical eraseblocks returned by 'ubi_wl_get_peb()' have only erase counter
  20  * header. The rest of the physical eraseblock contains only 0xFF bytes.
  21  *
  22  * When physical eraseblocks are returned to the WL unit by means of the
  23  * 'ubi_wl_put_peb()' function, they are scheduled for erasure. The erasure is
  24  * done asynchronously in context of the per-UBI device background thread,
  25  * which is also managed by the WL unit.
  26  *
  27  * The wear-leveling is ensured by means of moving the contents of used
  28  * physical eraseblocks with low erase counter to free physical eraseblocks
  29  * with high erase counter.
  30  *
  31  * The 'ubi_wl_get_peb()' function accepts data type hints which help to pick
  32  * an "optimal" physical eraseblock. For example, when it is known that the
  33  * physical eraseblock will be "put" soon because it contains short-term data,
  34  * the WL unit may pick a free physical eraseblock with low erase counter, and
  35  * so forth.
  36  *
  37  * If the WL unit fails to erase a physical eraseblock, it marks it as bad.
  38  *
  39  * This unit is also responsible for scrubbing. If a bit-flip is detected in a
  40  * physical eraseblock, it has to be moved. Technically this is the same as
  41  * moving it for wear-leveling reasons.
  42  *
  43  * As it was said, for the UBI unit all physical eraseblocks are either "free"
  44  * or "used". Free eraseblock are kept in the @wl->free RB-tree, while used
  45  * eraseblocks are kept in a set of different RB-trees: @wl->used,
  46  * @wl->prot.pnum, @wl->prot.aec, and @wl->scrub.
  47  *
  48  * Note, in this implementation, we keep a small in-RAM object for each physical
  49  * eraseblock. This is surely not a scalable solution. But it appears to be good
  50  * enough for moderately large flashes and it is simple. In future, one may
  51  * re-work this unit and make it more scalable.
  52  *
  53  * At the moment this unit does not utilize the sequence number, which was
  54  * introduced relatively recently. But it would be wise to do this because the
  55  * sequence number of a logical eraseblock characterizes how old is it. For
  56  * example, when we move a PEB with low erase counter, and we need to pick the
  57  * target PEB, we pick a PEB with the highest EC if our PEB is "old" and we
  58  * pick target PEB with an average EC if our PEB is not very "old". This is a
  59  * room for future re-works of the WL unit.
  60  *
  61  * FIXME: looks too complex, should be simplified (later).
  62  */
  63
  64 #ifdef UBI_LINUX
  65 #include <linux/slab.h>
  66 #include <linux/crc32.h>
  67 #include <linux/freezer.h>
  68 #include <linux/kthread.h>
  69 #endif
  70
  71 #include <ubi_uboot.h>
  72 #include "ubi.h"
  73
  74 /* Number of physical eraseblocks reserved for wear-leveling purposes */
  75 #define WL_RESERVED_PEBS 1
  76
  77 /*
  78  * How many erase cycles are short term, unknown, and long term physical
  79  * eraseblocks protected.
  80  */
  81 #define ST_PROTECTION 16
  82 #define U_PROTECTION  10
  83 #define LT_PROTECTION 4
  84
  85 /*
  86  * Maximum difference between two erase counters. If this threshold is
  87  * exceeded, the WL unit starts moving data from used physical eraseblocks with
  88  * low erase counter to free physical eraseblocks with high erase counter.
  89  */
  90 #define UBI_WL_THRESHOLD CONFIG_MTD_UBI_WL_THRESHOLD
  91
  92 /*
  93  * When a physical eraseblock is moved, the WL unit has to pick the target
  94  * physical eraseblock to move to. The simplest way would be just to pick the
  95  * one with the highest erase counter. But in certain workloads this could lead
  96  * to an unlimited wear of one or few physical eraseblock. Indeed, imagine a
  97  * situation when the picked physical eraseblock is constantly erased after the
  98  * data is written to it. So, we have a constant which limits the highest erase
  99  * counter of the free physical eraseblock to pick. Namely, the WL unit does
 100  * not pick eraseblocks with erase counter greater then the lowest erase
 101  * counter plus %WL_FREE_MAX_DIFF.
 102  */
 103 #define WL_FREE_MAX_DIFF (2*UBI_WL_THRESHOLD)
 104
 105 /*
 106  * Maximum number of consecutive background thread failures which is enough to
 107  * switch to read-only mode.
 108  */
 109 #define WL_MAX_FAILURES 32
 110
 111 /**
 112  * struct ubi_wl_prot_entry - PEB protection entry.
 113  * @rb_pnum: link in the @wl->prot.pnum RB-tree
 114  * @rb_aec: link in the @wl->prot.aec RB-tree
 115  * @abs_ec: the absolute erase counter value when the protection ends
 116  * @e: the wear-leveling entry of the physical eraseblock under protection
 117  *
 118  * When the WL unit returns a physical eraseblock, the physical eraseblock is
 119  * protected from being moved for some "time". For this reason, the physical
 120  * eraseblock is not directly moved from the @wl->free tree to the @wl->used
 121  * tree. There is one more tree in between where this physical eraseblock is
 122  * temporarily stored (@wl->prot).
 123  *
 124  * All this protection stuff is needed because:
 125  *  o we don't want to move physical eraseblocks just after we have given them
 126  *    to the user; instead, we first want to let users fill them up with data;
 127  *
 128  *  o there is a chance that the user will put the physical eraseblock very
 129  *    soon, so it makes sense not to move it for some time, but wait; this is
 130  *    especially important in case of "short term" physical eraseblocks.
 131  *
 132  * Physical eraseblocks stay protected only for limited time. But the "time" is
 133  * measured in erase cycles in this case. This is implemented with help of the
 134  * absolute erase counter (@wl->abs_ec). When it reaches certain value, the
 135  * physical eraseblocks are moved from the protection trees (@wl->prot.*) to
 136  * the @wl->used tree.
 137  *
 138  * Protected physical eraseblocks are searched by physical eraseblock number
 139  * (when they are put) and by the absolute erase counter (to check if it is
 140  * time to move them to the @wl->used tree). So there are actually 2 RB-trees
 141  * storing the protected physical eraseblocks: @wl->prot.pnum and
 142  * @wl->prot.aec. They are referred to as the "protection" trees. The
 143  * first one is indexed by the physical eraseblock number. The second one is
 144  * indexed by the absolute erase counter. Both trees store
 145  * &struct ubi_wl_prot_entry objects.
 146  *
 147  * Each physical eraseblock has 2 main states: free and used. The former state
 148  * corresponds to the @wl->free tree. The latter state is split up on several
 149  * sub-states:
 150  * o the WL movement is allowed (@wl->used tree);
 151  * o the WL movement is temporarily prohibited (@wl->prot.pnum and
 152  * @wl->prot.aec trees);
 153  * o scrubbing is needed (@wl->scrub tree).
 154  *
 155  * Depending on the sub-state, wear-leveling entries of the used physical
 156  * eraseblocks may be kept in one of those trees.
 157  */
 158 struct ubi_wl_prot_entry {
 159         struct rb_node rb_pnum;
 160         struct rb_node rb_aec;
 161         unsigned long long abs_ec;
 162         struct ubi_wl_entry *e;
 163 };
 164
 165 /**
 166  * struct ubi_work - UBI work description data structure.
 167  * @list: a link in the list of pending works
 168  * @func: worker function
 169  * @priv: private data of the worker function
 170  *
 171  * @e: physical eraseblock to erase
 172  * @torture: if the physical eraseblock has to be tortured
 173  *
 174  * The @func pointer points to the worker function. If the @cancel argument is
 175  * not zero, the worker has to free the resources and exit immediately. The
 176  * worker has to return zero in case of success and a negative error code in
 177  * case of failure.
 178  */
 179 struct ubi_work {
 180         struct list_head list;
 181         int (*func)(struct ubi_device *ubi, struct ubi_work *wrk, int cancel);
 182         /* The below fields are only relevant to erasure works */
 183         struct ubi_wl_entry *e;
 184         int torture;
 185 };
 186
 187 #ifdef CONFIG_MTD_UBI_DEBUG_PARANOID
 188 static int paranoid_check_ec(struct ubi_device *ubi, int pnum, int ec);
 189 static int paranoid_check_in_wl_tree(struct ubi_wl_entry *e,
 190                                      struct rb_root *root);
 191 #else
 192 #define paranoid_check_ec(ubi, pnum, ec) 0
 193 #define paranoid_check_in_wl_tree(e, root)
 194 #endif
 195
 196 /**
 197  * wl_tree_add - add a wear-leveling entry to a WL RB-tree.
 198  * @e: the wear-leveling entry to add
 199  * @root: the root of the tree
 200  *
 201  * Note, we use (erase counter, physical eraseblock number) pairs as keys in
 202  * the @ubi->used and @ubi->free RB-trees.
 203  */
 204 static void wl_tree_add(struct ubi_wl_entry *e, struct rb_root *root)
 205 {
 206         struct rb_node **p, *parent = NULL;
 207
 208         p = &root->rb_node;
 209         while (*p) {
 210                 struct ubi_wl_entry *e1;
 211
 212                 parent = *p;
 213                 e1 = rb_entry(parent, struct ubi_wl_entry, rb);
 214
 215                 if (e->ec < e1->ec)
 216                         p = &(*p)->rb_left;
 217                 else if (e->ec > e1->ec)
 218                         p = &(*p)->rb_right;
 219                 else {
 220                         ubi_assert(e->pnum != e1->pnum);
 221                         if (e->pnum < e1->pnum)
 222                                 p = &(*p)->rb_left;
 223                         else
 224                                 p = &(*p)->rb_right;
 225                 }
 226         }
 227
 228         rb_link_node(&e->rb, parent, p);
 229         rb_insert_color(&e->rb, root);
 230 }
 231
 232 /**
 233  * do_work - do one pending work.
 234  * @ubi: UBI device description object
 235  *
 236  * This function returns zero in case of success and a negative error code in
 237  * case of failure.
 238  */
 239 static int do_work(struct ubi_device *ubi)
 240 {
 241         int err;
 242         struct ubi_work *wrk;
 243
 244         cond_resched();
 245
 246         /*
 247          * @ubi->work_sem is used to synchronize with the workers. Workers take
 248          * it in read mode, so many of them may be doing works at a time. But
 249          * the queue flush code has to be sure the whole queue of works is
 250          * done, and it takes the mutex in write mode.
 251          */
 252         down_read(&ubi->work_sem);
 253         spin_lock(&ubi->wl_lock);
 254         if (list_empty(&ubi->works)) {
 255                 spin_unlock(&ubi->wl_lock);
 256                 up_read(&ubi->work_sem);
 257                 return 0;
 258         }
 259
 260         wrk = list_entry(ubi->works.next, struct ubi_work, list);
 261         list_del(&wrk->list);
 262         ubi->works_count -= 1;
 263         ubi_assert(ubi->works_count >= 0);
 264         spin_unlock(&ubi->wl_lock);
 265
 266         /*
 267          * Call the worker function. Do not touch the work structure
 268          * after this call as it will have been freed or reused by that
 269          * time by the worker function.
 270          */
 271         err = wrk->func(ubi, wrk, 0);
 272         if (err)
 273                 ubi_err("work failed with error code %d", err);
 274         up_read(&ubi->work_sem);
 275
 276         return err;
 277 }
 278
 279 /**
 280  * produce_free_peb - produce a free physical eraseblock.
 281  * @ubi: UBI device description object
 282  *
 283  * This function tries to make a free PEB by means of synchronous execution of
 284  * pending works. This may be needed if, for example the background thread is
 285  * disabled. Returns zero in case of success and a negative error code in case
 286  * of failure.
 287  */
 288 static int produce_free_peb(struct ubi_device *ubi)
 289 {
 290         int err;
 291
 292         spin_lock(&ubi->wl_lock);
 293         while (!ubi->free.rb_node) {
 294                 spin_unlock(&ubi->wl_lock);
 295
 296                 dbg_wl("do one work synchronously");
 297                 err = do_work(ubi);
 298                 if (err)
 299                         return err;
 300
 301                 spin_lock(&ubi->wl_lock);
 302         }
 303         spin_unlock(&ubi->wl_lock);
 304
 305         return 0;
 306 }
 307
 308 /**
 309  * in_wl_tree - check if wear-leveling entry is present in a WL RB-tree.
 310  * @e: the wear-leveling entry to check
 311  * @root: the root of the tree
 312  *
 313  * This function returns non-zero if @e is in the @root RB-tree and zero if it
 314  * is not.
 315  */
 316 static int in_wl_tree(struct ubi_wl_entry *e, struct rb_root *root)
 317 {
 318         struct rb_node *p;
 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, rb);
 325
 326                 if (e->pnum == e1->pnum) {
 327                         ubi_assert(e == e1);
 328                         return 1;
 329                 }
 330
 331                 if (e->ec < e1->ec)
 332                         p = p->rb_left;
 333                 else if (e->ec > e1->ec)
 334                         p = p->rb_right;
 335                 else {
 336                         ubi_assert(e->pnum != e1->pnum);
 337                         if (e->pnum < e1->pnum)
 338                                 p = p->rb_left;
 339                         else
 340                                 p = p->rb_right;
 341                 }
 342         }
 343
 344         return 0;
 345 }
 346
 347 /**
 348  * prot_tree_add - add physical eraseblock to protection trees.
 349  * @ubi: UBI device description object
 350  * @e: the physical eraseblock to add
 351  * @pe: protection entry object to use
 352  * @abs_ec: absolute erase counter value when this physical eraseblock has
 353  * to be removed from the protection trees.
 354  *
 355  * @wl->lock has to be locked.
 356  */
 357 static void prot_tree_add(struct ubi_device *ubi, struct ubi_wl_entry *e,
 358                           struct ubi_wl_prot_entry *pe, int abs_ec)
 359 {
 360         struct rb_node **p, *parent = NULL;
 361         struct ubi_wl_prot_entry *pe1;
 362
 363         pe->e = e;
 364         pe->abs_ec = ubi->abs_ec + abs_ec;
 365
 366         p = &ubi->prot.pnum.rb_node;
 367         while (*p) {
 368                 parent = *p;
 369                 pe1 = rb_entry(parent, struct ubi_wl_prot_entry, rb_pnum);
 370
 371                 if (e->pnum < pe1->e->pnum)
 372                         p = &(*p)->rb_left;
 373                 else
 374                         p = &(*p)->rb_right;
 375         }
 376         rb_link_node(&pe->rb_pnum, parent, p);
 377         rb_insert_color(&pe->rb_pnum, &ubi->prot.pnum);
 378
 379         p = &ubi->prot.aec.rb_node;
 380         parent = NULL;
 381         while (*p) {
 382                 parent = *p;
 383                 pe1 = rb_entry(parent, struct ubi_wl_prot_entry, rb_aec);
 384
 385                 if (pe->abs_ec < pe1->abs_ec)
 386                         p = &(*p)->rb_left;
 387                 else
 388                         p = &(*p)->rb_right;
 389         }
 390         rb_link_node(&pe->rb_aec, parent, p);
 391         rb_insert_color(&pe->rb_aec, &ubi->prot.aec);
 392 }
 393
 394 /**
 395  * find_wl_entry - find wear-leveling entry closest to certain erase counter.
 396  * @root: the RB-tree where to look for
 397  * @max: highest possible erase counter
 398  *
 399  * This function looks for a wear leveling entry with erase counter closest to
 400  * @max and less then @max.
 401  */
 402 static struct ubi_wl_entry *find_wl_entry(struct rb_root *root, int max)
 403 {
 404         struct rb_node *p;
 405         struct ubi_wl_entry *e;
 406
 407         e = rb_entry(rb_first(root), struct ubi_wl_entry, rb);
 408         max += e->ec;
 409
 410         p = root->rb_node;
 411         while (p) {
 412                 struct ubi_wl_entry *e1;
 413
 414                 e1 = rb_entry(p, struct ubi_wl_entry, rb);
 415                 if (e1->ec >= max)
 416                         p = p->rb_left;
 417                 else {
 418                         p = p->rb_right;
 419                         e = e1;
 420                 }
 421         }
 422
 423         return e;
 424 }
 425
 426 /**
 427  * ubi_wl_get_peb - get a physical eraseblock.
 428  * @ubi: UBI device description object
 429  * @dtype: type of data which will be stored in this physical eraseblock
 430  *
 431  * This function returns a physical eraseblock in case of success and a
 432  * negative error code in case of failure. Might sleep.
 433  */
 434 int ubi_wl_get_peb(struct ubi_device *ubi, int dtype)
 435 {
 436         int err, protect, medium_ec;
 437         struct ubi_wl_entry *e, *first, *last;
 438         struct ubi_wl_prot_entry *pe;
 439
 440         ubi_assert(dtype == UBI_LONGTERM || dtype == UBI_SHORTTERM ||
 441                    dtype == UBI_UNKNOWN);
 442
 443         pe = kmalloc(sizeof(struct ubi_wl_prot_entry), GFP_NOFS);
 444         if (!pe)
 445                 return -ENOMEM;
 446
 447 retry:
 448         spin_lock(&ubi->wl_lock);
 449         if (!ubi->free.rb_node) {
 450                 if (ubi->works_count == 0) {
 451                         ubi_assert(list_empty(&ubi->works));
 452                         ubi_err("no free eraseblocks");
 453                         spin_unlock(&ubi->wl_lock);
 454                         kfree(pe);
 455                         return -ENOSPC;
 456                 }
 457                 spin_unlock(&ubi->wl_lock);
 458
 459                 err = produce_free_peb(ubi);
 460                 if (err < 0) {
 461                         kfree(pe);
 462                         return err;
 463                 }
 464                 goto retry;
 465         }
 466
 467         switch (dtype) {
 468                 case UBI_LONGTERM:
 469                         /*
 470                          * For long term data we pick a physical eraseblock
 471                          * with high erase counter. But the highest erase
 472                          * counter we can pick is bounded by the the lowest
 473                          * erase counter plus %WL_FREE_MAX_DIFF.
 474                          */
 475                         e = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
 476                         protect = LT_PROTECTION;
 477                         break;
 478                 case UBI_UNKNOWN:
 479                         /*
 480                          * For unknown data we pick a physical eraseblock with
 481                          * medium erase counter. But we by no means can pick a
 482                          * physical eraseblock with erase counter greater or
 483                          * equivalent than the lowest erase counter plus
 484                          * %WL_FREE_MAX_DIFF.
 485                          */
 486                         first = rb_entry(rb_first(&ubi->free),
 487                                          struct ubi_wl_entry, rb);
 488                         last = rb_entry(rb_last(&ubi->free),
 489                                         struct ubi_wl_entry, rb);
 490
 491                         if (last->ec - first->ec < WL_FREE_MAX_DIFF)
 492                                 e = rb_entry(ubi->free.rb_node,
 493                                                 struct ubi_wl_entry, rb);
 494                         else {
 495                                 medium_ec = (first->ec + WL_FREE_MAX_DIFF)/2;
 496                                 e = find_wl_entry(&ubi->free, medium_ec);
 497                         }
 498                         protect = U_PROTECTION;
 499                         break;
 500                 case UBI_SHORTTERM:
 501                         /*
 502                          * For short term data we pick a physical eraseblock
 503                          * with the lowest erase counter as we expect it will
 504                          * be erased soon.
 505                          */
 506                         e = rb_entry(rb_first(&ubi->free),
 507                                      struct ubi_wl_entry, rb);
 508                         protect = ST_PROTECTION;
 509                         break;
 510                 default:
 511                         protect = 0;
 512                         e = NULL;
 513                         BUG();
 514         }
 515
 516         /*
 517          * Move the physical eraseblock to the protection trees where it will
 518          * be protected from being moved for some time.
 519          */
 520         paranoid_check_in_wl_tree(e, &ubi->free);
 521         rb_erase(&e->rb, &ubi->free);
 522         prot_tree_add(ubi, e, pe, protect);
 523
 524         dbg_wl("PEB %d EC %d, protection %d", e->pnum, e->ec, protect);
 525         spin_unlock(&ubi->wl_lock);
 526
 527         return e->pnum;
 528 }
 529
 530 /**
 531  * prot_tree_del - remove a physical eraseblock from the protection trees
 532  * @ubi: UBI device description object
 533  * @pnum: the physical eraseblock to remove
 534  *
 535  * This function returns PEB @pnum from the protection trees and returns zero
 536  * in case of success and %-ENODEV if the PEB was not found in the protection
 537  * trees.
 538  */
 539 static int prot_tree_del(struct ubi_device *ubi, int pnum)
 540 {
 541         struct rb_node *p;
 542         struct ubi_wl_prot_entry *pe = NULL;
 543
 544         p = ubi->prot.pnum.rb_node;
 545         while (p) {
 546
 547                 pe = rb_entry(p, struct ubi_wl_prot_entry, rb_pnum);
 548
 549                 if (pnum == pe->e->pnum)
 550                         goto found;
 551
 552                 if (pnum < pe->e->pnum)
 553                         p = p->rb_left;
 554                 else
 555                         p = p->rb_right;
 556         }
 557
 558         return -ENODEV;
 559
 560 found:
 561         ubi_assert(pe->e->pnum == pnum);
 562         rb_erase(&pe->rb_aec, &ubi->prot.aec);
 563         rb_erase(&pe->rb_pnum, &ubi->prot.pnum);
 564         kfree(pe);
 565         return 0;
 566 }
 567
 568 /**
 569  * sync_erase - synchronously erase a physical eraseblock.
 570  * @ubi: UBI device description object
 571  * @e: the the physical eraseblock to erase
 572  * @torture: if the physical eraseblock has to be tortured
 573  *
 574  * This function returns zero in case of success and a negative error code in
 575  * case of failure.
 576  */
 577 static int sync_erase(struct ubi_device *ubi, struct ubi_wl_entry *e, int torture)
 578 {
 579         int err;
 580         struct ubi_ec_hdr *ec_hdr;
 581         unsigned long long ec = e->ec;
 582
 583         dbg_wl("erase PEB %d, old EC %llu", e->pnum, ec);
 584
 585         err = paranoid_check_ec(ubi, e->pnum, e->ec);
 586         if (err > 0)
 587                 return -EINVAL;
 588
 589         ec_hdr = kzalloc(ubi->ec_hdr_alsize, GFP_NOFS);
 590         if (!ec_hdr)
 591                 return -ENOMEM;
 592
 593         err = ubi_io_sync_erase(ubi, e->pnum, torture);
 594         if (err < 0)
 595                 goto out_free;
 596
 597         ec += err;
 598         if (ec > UBI_MAX_ERASECOUNTER) {
 599                 /*
 600                  * Erase counter overflow. Upgrade UBI and use 64-bit
 601                  * erase counters internally.
 602                  */
 603                 ubi_err("erase counter overflow at PEB %d, EC %llu",
 604                         e->pnum, ec);
 605                 err = -EINVAL;
 606                 goto out_free;
 607         }
 608
 609         dbg_wl("erased PEB %d, new EC %llu", e->pnum, ec);
 610
 611         ec_hdr->ec = cpu_to_be64(ec);
 612
 613         err = ubi_io_write_ec_hdr(ubi, e->pnum, ec_hdr);
 614         if (err)
 615                 goto out_free;
 616
 617         e->ec = ec;
 618         spin_lock(&ubi->wl_lock);
 619         if (e->ec > ubi->max_ec)
 620                 ubi->max_ec = e->ec;
 621         spin_unlock(&ubi->wl_lock);
 622
 623 out_free:
 624         kfree(ec_hdr);
 625         return err;
 626 }
 627
 628 /**
 629  * check_protection_over - check if it is time to stop protecting some
 630  * physical eraseblocks.
 631  * @ubi: UBI device description object
 632  *
 633  * This function is called after each erase operation, when the absolute erase
 634  * counter is incremented, to check if some physical eraseblock  have not to be
 635  * protected any longer. These physical eraseblocks are moved from the
 636  * protection trees to the used tree.
 637  */
 638 static void check_protection_over(struct ubi_device *ubi)
 639 {
 640         struct ubi_wl_prot_entry *pe;
 641
 642         /*
 643          * There may be several protected physical eraseblock to remove,
 644          * process them all.
 645          */
 646         while (1) {
 647                 spin_lock(&ubi->wl_lock);
 648                 if (!ubi->prot.aec.rb_node) {
 649                         spin_unlock(&ubi->wl_lock);
 650                         break;
 651                 }
 652
 653                 pe = rb_entry(rb_first(&ubi->prot.aec),
 654                               struct ubi_wl_prot_entry, rb_aec);
 655
 656                 if (pe->abs_ec > ubi->abs_ec) {
 657                         spin_unlock(&ubi->wl_lock);
 658                         break;
 659                 }
 660
 661                 dbg_wl("PEB %d protection over, abs_ec %llu, PEB abs_ec %llu",
 662                        pe->e->pnum, ubi->abs_ec, pe->abs_ec);
 663                 rb_erase(&pe->rb_aec, &ubi->prot.aec);
 664                 rb_erase(&pe->rb_pnum, &ubi->prot.pnum);
 665                 wl_tree_add(pe->e, &ubi->used);
 666                 spin_unlock(&ubi->wl_lock);
 667
 668                 kfree(pe);
 669                 cond_resched();
 670         }
 671 }
 672
 673 /**
 674  * schedule_ubi_work - schedule a work.
 675  * @ubi: UBI device description object
 676  * @wrk: the work to schedule
 677  *
 678  * This function enqueues a work defined by @wrk to the tail of the pending
 679  * works list.
 680  */
 681 static void schedule_ubi_work(struct ubi_device *ubi, struct ubi_work *wrk)
 682 {
 683         spin_lock(&ubi->wl_lock);
 684         list_add_tail(&wrk->list, &ubi->works);
 685         ubi_assert(ubi->works_count >= 0);
 686         ubi->works_count += 1;
 687
 688         /*
 689          * U-Boot special: We have no bgt_thread in U-Boot!
 690          * So just call do_work() here directly.
 691          */
 692         do_work(ubi);
 693
 694         spin_unlock(&ubi->wl_lock);
 695 }
 696
 697 static int erase_worker(struct ubi_device *ubi, struct ubi_work *wl_wrk,
 698                         int cancel);
 699
 700 /**
 701  * schedule_erase - schedule an erase work.
 702  * @ubi: UBI device description object
 703  * @e: the WL entry of the physical eraseblock to erase
 704  * @torture: if the physical eraseblock has to be tortured
 705  *
 706  * This function returns zero in case of success and a %-ENOMEM in case of
 707  * failure.
 708  */
 709 static int schedule_erase(struct ubi_device *ubi, struct ubi_wl_entry *e,
 710                           int torture)
 711 {
 712         struct ubi_work *wl_wrk;
 713
 714         dbg_wl("schedule erasure of PEB %d, EC %d, torture %d",
 715                e->pnum, e->ec, torture);
 716
 717         wl_wrk = kmalloc(sizeof(struct ubi_work), GFP_NOFS);
 718         if (!wl_wrk)
 719                 return -ENOMEM;
 720
 721         wl_wrk->func = &erase_worker;
 722         wl_wrk->e = e;
 723         wl_wrk->torture = torture;
 724
 725         schedule_ubi_work(ubi, wl_wrk);
 726         return 0;
 727 }
 728
 729 /**
 730  * wear_leveling_worker - wear-leveling worker function.
 731  * @ubi: UBI device description object
 732  * @wrk: the work object
 733  * @cancel: non-zero if the worker has to free memory and exit
 734  *
 735  * This function copies a more worn out physical eraseblock to a less worn out
 736  * one. Returns zero in case of success and a negative error code in case of
 737  * failure.
 738  */
 739 static int wear_leveling_worker(struct ubi_device *ubi, struct ubi_work *wrk,
 740                                 int cancel)
 741 {
 742         int err, put = 0, scrubbing = 0, protect = 0;
 743         struct ubi_wl_prot_entry *uninitialized_var(pe);
 744         struct ubi_wl_entry *e1, *e2;
 745         struct ubi_vid_hdr *vid_hdr;
 746
 747         kfree(wrk);
 748
 749         if (cancel)
 750                 return 0;
 751
 752         vid_hdr = ubi_zalloc_vid_hdr(ubi, GFP_NOFS);
 753         if (!vid_hdr)
 754                 return -ENOMEM;
 755
 756         mutex_lock(&ubi->move_mutex);
 757         spin_lock(&ubi->wl_lock);
 758         ubi_assert(!ubi->move_from && !ubi->move_to);
 759         ubi_assert(!ubi->move_to_put);
 760
 761         if (!ubi->free.rb_node ||
 762             (!ubi->used.rb_node && !ubi->scrub.rb_node)) {
 763                 /*
 764                  * No free physical eraseblocks? Well, they must be waiting in
 765                  * the queue to be erased. Cancel movement - it will be
 766                  * triggered again when a free physical eraseblock appears.
 767                  *
 768                  * No used physical eraseblocks? They must be temporarily
 769                  * protected from being moved. They will be moved to the
 770                  * @ubi->used tree later and the wear-leveling will be
 771                  * triggered again.
 772                  */
 773                 dbg_wl("cancel WL, a list is empty: free %d, used %d",
 774                        !ubi->free.rb_node, !ubi->used.rb_node);
 775                 goto out_cancel;
 776         }
 777
 778         if (!ubi->scrub.rb_node) {
 779                 /*
 780                  * Now pick the least worn-out used physical eraseblock and a
 781                  * highly worn-out free physical eraseblock. If the erase
 782                  * counters differ much enough, start wear-leveling.
 783                  */
 784                 e1 = rb_entry(rb_first(&ubi->used), struct ubi_wl_entry, rb);
 785                 e2 = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
 786
 787                 if (!(e2->ec - e1->ec >= UBI_WL_THRESHOLD)) {
 788                         dbg_wl("no WL needed: min used EC %d, max free EC %d",
 789                                e1->ec, e2->ec);
 790                         goto out_cancel;
 791                 }
 792                 paranoid_check_in_wl_tree(e1, &ubi->used);
 793                 rb_erase(&e1->rb, &ubi->used);
 794                 dbg_wl("move PEB %d EC %d to PEB %d EC %d",
 795                        e1->pnum, e1->ec, e2->pnum, e2->ec);
 796         } else {
 797                 /* Perform scrubbing */
 798                 scrubbing = 1;
 799                 e1 = rb_entry(rb_first(&ubi->scrub), struct ubi_wl_entry, rb);
 800                 e2 = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
 801                 paranoid_check_in_wl_tree(e1, &ubi->scrub);
 802                 rb_erase(&e1->rb, &ubi->scrub);
 803                 dbg_wl("scrub PEB %d to PEB %d", e1->pnum, e2->pnum);
 804         }
 805
 806         paranoid_check_in_wl_tree(e2, &ubi->free);
 807         rb_erase(&e2->rb, &ubi->free);
 808         ubi->move_from = e1;
 809         ubi->move_to = e2;
 810         spin_unlock(&ubi->wl_lock);
 811
 812         /*
 813          * Now we are going to copy physical eraseblock @e1->pnum to @e2->pnum.
 814          * We so far do not know which logical eraseblock our physical
 815          * eraseblock (@e1) belongs to. We have to read the volume identifier
 816          * header first.
 817          *
 818          * Note, we are protected from this PEB being unmapped and erased. The
 819          * 'ubi_wl_put_peb()' would wait for moving to be finished if the PEB
 820          * which is being moved was unmapped.
 821          */
 822
 823         err = ubi_io_read_vid_hdr(ubi, e1->pnum, vid_hdr, 0);
 824         if (err && err != UBI_IO_BITFLIPS) {
 825                 if (err == UBI_IO_PEB_FREE) {
 826                         /*
 827                          * We are trying to move PEB without a VID header. UBI
 828                          * always write VID headers shortly after the PEB was
 829                          * given, so we have a situation when it did not have
 830                          * chance to write it down because it was preempted.
 831                          * Just re-schedule the work, so that next time it will
 832                          * likely have the VID header in place.
 833                          */
 834                         dbg_wl("PEB %d has no VID header", e1->pnum);
 835                         goto out_not_moved;
 836                 }
 837
 838                 ubi_err("error %d while reading VID header from PEB %d",
 839                         err, e1->pnum);
 840                 if (err > 0)
 841                         err = -EIO;
 842                 goto out_error;
 843         }
 844
 845         err = ubi_eba_copy_leb(ubi, e1->pnum, e2->pnum, vid_hdr);
 846         if (err) {
 847
 848                 if (err < 0)
 849                         goto out_error;
 850                 if (err == 1)
 851                         goto out_not_moved;
 852
 853                 /*
 854                  * For some reason the LEB was not moved - it might be because
 855                  * the volume is being deleted. We should prevent this PEB from
 856                  * being selected for wear-levelling movement for some "time",
 857                  * so put it to the protection tree.
 858                  */
 859
 860                 dbg_wl("cancelled moving PEB %d", e1->pnum);
 861                 pe = kmalloc(sizeof(struct ubi_wl_prot_entry), GFP_NOFS);
 862                 if (!pe) {
 863                         err = -ENOMEM;
 864                         goto out_error;
 865                 }
 866
 867                 protect = 1;
 868         }
 869
 870         ubi_free_vid_hdr(ubi, vid_hdr);
 871         spin_lock(&ubi->wl_lock);
 872         if (protect)
 873                 prot_tree_add(ubi, e1, pe, protect);
 874         if (!ubi->move_to_put)
 875                 wl_tree_add(e2, &ubi->used);
 876         else
 877                 put = 1;
 878         ubi->move_from = ubi->move_to = NULL;
 879         ubi->move_to_put = ubi->wl_scheduled = 0;
 880         spin_unlock(&ubi->wl_lock);
 881
 882         if (put) {
 883                 /*
 884                  * Well, the target PEB was put meanwhile, schedule it for
 885                  * erasure.
 886                  */
 887                 dbg_wl("PEB %d was put meanwhile, erase", e2->pnum);
 888                 err = schedule_erase(ubi, e2, 0);
 889                 if (err)
 890                         goto out_error;
 891         }
 892
 893         if (!protect) {
 894                 err = schedule_erase(ubi, e1, 0);
 895                 if (err)
 896                         goto out_error;
 897         }
 898
 899
 900         dbg_wl("done");
 901         mutex_unlock(&ubi->move_mutex);
 902         return 0;
 903
 904         /*
 905          * For some reasons the LEB was not moved, might be an error, might be
 906          * something else. @e1 was not changed, so return it back. @e2 might
 907          * be changed, schedule it for erasure.
 908          */
 909 out_not_moved:
 910         ubi_free_vid_hdr(ubi, vid_hdr);
 911         spin_lock(&ubi->wl_lock);
 912         if (scrubbing)
 913                 wl_tree_add(e1, &ubi->scrub);
 914         else
 915                 wl_tree_add(e1, &ubi->used);
 916         ubi->move_from = ubi->move_to = NULL;
 917         ubi->move_to_put = ubi->wl_scheduled = 0;
 918         spin_unlock(&ubi->wl_lock);
 919
 920         err = schedule_erase(ubi, e2, 0);
 921         if (err)
 922                 goto out_error;
 923
 924         mutex_unlock(&ubi->move_mutex);
 925         return 0;
 926
 927 out_error:
 928         ubi_err("error %d while moving PEB %d to PEB %d",
 929                 err, e1->pnum, e2->pnum);
 930
 931         ubi_free_vid_hdr(ubi, vid_hdr);
 932         spin_lock(&ubi->wl_lock);
 933         ubi->move_from = ubi->move_to = NULL;
 934         ubi->move_to_put = ubi->wl_scheduled = 0;
 935         spin_unlock(&ubi->wl_lock);
 936
 937         kmem_cache_free(ubi_wl_entry_slab, e1);
 938         kmem_cache_free(ubi_wl_entry_slab, e2);
 939         ubi_ro_mode(ubi);
 940
 941         mutex_unlock(&ubi->move_mutex);
 942         return err;
 943
 944 out_cancel:
 945         ubi->wl_scheduled = 0;
 946         spin_unlock(&ubi->wl_lock);
 947         mutex_unlock(&ubi->move_mutex);
 948         ubi_free_vid_hdr(ubi, vid_hdr);
 949         return 0;
 950 }
 951
 952 /**
 953  * ensure_wear_leveling - schedule wear-leveling if it is needed.
 954  * @ubi: UBI device description object
 955  *
 956  * This function checks if it is time to start wear-leveling and schedules it
 957  * if yes. This function returns zero in case of success and a negative error
 958  * code in case of failure.
 959  */
 960 static int ensure_wear_leveling(struct ubi_device *ubi)
 961 {
 962         int err = 0;
 963         struct ubi_wl_entry *e1;
 964         struct ubi_wl_entry *e2;
 965         struct ubi_work *wrk;
 966
 967         spin_lock(&ubi->wl_lock);
 968         if (ubi->wl_scheduled)
 969                 /* Wear-leveling is already in the work queue */
 970                 goto out_unlock;
 971
 972         /*
 973          * If the ubi->scrub tree is not empty, scrubbing is needed, and the
 974          * the WL worker has to be scheduled anyway.
 975          */
 976         if (!ubi->scrub.rb_node) {
 977                 if (!ubi->used.rb_node || !ubi->free.rb_node)
 978                         /* No physical eraseblocks - no deal */
 979                         goto out_unlock;
 980
 981                 /*
 982                  * We schedule wear-leveling only if the difference between the
 983                  * lowest erase counter of used physical eraseblocks and a high
 984                  * erase counter of free physical eraseblocks is greater then
 985                  * %UBI_WL_THRESHOLD.
 986                  */
 987                 e1 = rb_entry(rb_first(&ubi->used), struct ubi_wl_entry, rb);
 988                 e2 = find_wl_entry(&ubi->free, WL_FREE_MAX_DIFF);
 989
 990                 if (!(e2->ec - e1->ec >= UBI_WL_THRESHOLD))
 991                         goto out_unlock;
 992                 dbg_wl("schedule wear-leveling");
 993         } else
 994                 dbg_wl("schedule scrubbing");
 995
 996         ubi->wl_scheduled = 1;
 997         spin_unlock(&ubi->wl_lock);
 998
 999         wrk = kmalloc(sizeof(struct ubi_work), GFP_NOFS);
1000         if (!wrk) {
1001                 err = -ENOMEM;
1002                 goto out_cancel;
1003         }
1004
1005         wrk->func = &wear_leveling_worker;
1006         schedule_ubi_work(ubi, wrk);
1007         return err;
1008
1009 out_cancel:
1010         spin_lock(&ubi->wl_lock);
1011         ubi->wl_scheduled = 0;
1012 out_unlock:
1013         spin_unlock(&ubi->wl_lock);
1014         return err;
1015 }
1016
1017 /**
1018  * erase_worker - physical eraseblock erase worker function.
1019  * @ubi: UBI device description object
1020  * @wl_wrk: the work object
1021  * @cancel: non-zero if the worker has to free memory and exit
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 cancel)
1030 {
1031         struct ubi_wl_entry *e = wl_wrk->e;
1032         int pnum = e->pnum, err, need;
1033
1034         if (cancel) {
1035                 dbg_wl("cancel erasure of PEB %d EC %d", pnum, e->ec);
1036                 kfree(wl_wrk);
1037                 kmem_cache_free(ubi_wl_entry_slab, e);
1038                 return 0;
1039         }
1040
1041         dbg_wl("erase PEB %d EC %d", pnum, e->ec);
1042
1043         err = sync_erase(ubi, e, wl_wrk->torture);
1044         if (!err) {
1045                 /* Fine, we've erased it successfully */
1046                 kfree(wl_wrk);
1047
1048                 spin_lock(&ubi->wl_lock);
1049                 ubi->abs_ec += 1;
1050                 wl_tree_add(e, &ubi->free);
1051                 spin_unlock(&ubi->wl_lock);
1052
1053                 /*
1054                  * One more erase operation has happened, take care about protected
1055                  * physical eraseblocks.
1056                  */
1057                 check_protection_over(ubi);
1058
1059                 /* And take care about wear-leveling */
1060                 err = ensure_wear_leveling(ubi);
1061                 return err;
1062         }
1063
1064         ubi_err("failed to erase PEB %d, error %d", pnum, err);
1065         kfree(wl_wrk);
1066         kmem_cache_free(ubi_wl_entry_slab, e);
1067
1068         if (err == -EINTR || err == -ENOMEM || err == -EAGAIN ||
1069             err == -EBUSY) {
1070                 int err1;
1071
1072                 /* Re-schedule the LEB for erasure */
1073                 err1 = schedule_erase(ubi, e, 0);
1074                 if (err1) {
1075                         err = err1;
1076                         goto out_ro;
1077                 }
1078                 return err;
1079         } else if (err != -EIO) {
1080                 /*
1081                  * If this is not %-EIO, we have no idea what to do. Scheduling
1082                  * this physical eraseblock for erasure again would cause
1083                  * errors again and again. Well, lets switch to RO mode.
1084                  */
1085                 goto out_ro;
1086         }
1087
1088         /* It is %-EIO, the PEB went bad */
1089
1090         if (!ubi->bad_allowed) {
1091                 ubi_err("bad physical eraseblock %d detected", pnum);
1092                 goto out_ro;
1093         }
1094
1095         spin_lock(&ubi->volumes_lock);
1096         need = ubi->beb_rsvd_level - ubi->beb_rsvd_pebs + 1;
1097         if (need > 0) {
1098                 need = ubi->avail_pebs >= need ? need : ubi->avail_pebs;
1099                 ubi->avail_pebs -= need;
1100                 ubi->rsvd_pebs += need;
1101                 ubi->beb_rsvd_pebs += need;
1102                 if (need > 0)
1103                         ubi_msg("reserve more %d PEBs", need);
1104         }
1105
1106         if (ubi->beb_rsvd_pebs == 0) {
1107                 spin_unlock(&ubi->volumes_lock);
1108                 ubi_err("no reserved physical eraseblocks");
1109                 goto out_ro;
1110         }
1111
1112         spin_unlock(&ubi->volumes_lock);
1113         ubi_msg("mark PEB %d as bad", pnum);
1114
1115         err = ubi_io_mark_bad(ubi, pnum);
1116         if (err)
1117                 goto out_ro;
1118
1119         spin_lock(&ubi->volumes_lock);
1120         ubi->beb_rsvd_pebs -= 1;
1121         ubi->bad_peb_count += 1;
1122         ubi->good_peb_count -= 1;
1123         ubi_calculate_reserved(ubi);
1124         if (ubi->beb_rsvd_pebs == 0)
1125                 ubi_warn("last PEB from the reserved pool was used");
1126         spin_unlock(&ubi->volumes_lock);
1127
1128         return err;
1129
1130 out_ro:
1131         ubi_ro_mode(ubi);
1132         return err;
1133 }
1134
1135 /**
1136  * ubi_wl_put_peb - return a physical eraseblock to the wear-leveling unit.
1137  * @ubi: UBI device description object
1138  * @pnum: physical eraseblock to return
1139  * @torture: if this physical eraseblock has to be tortured
1140  *
1141  * This function is called to return physical eraseblock @pnum to the pool of
1142  * free physical eraseblocks. The @torture flag has to be set if an I/O error
1143  * occurred to this @pnum and it has to be tested. This function returns zero
1144  * in case of success, and a negative error code in case of failure.
1145  */
1146 int ubi_wl_put_peb(struct ubi_device *ubi, int pnum, int torture)
1147 {
1148         int err;
1149         struct ubi_wl_entry *e;
1150
1151         dbg_wl("PEB %d", pnum);
1152         ubi_assert(pnum >= 0);
1153         ubi_assert(pnum < ubi->peb_count);
1154
1155 retry:
1156         spin_lock(&ubi->wl_lock);
1157         e = ubi->lookuptbl[pnum];
1158         if (e == ubi->move_from) {
1159                 /*
1160                  * User is putting the physical eraseblock which was selected to
1161                  * be moved. It will be scheduled for erasure in the
1162                  * wear-leveling worker.
1163                  */
1164                 dbg_wl("PEB %d is being moved, wait", pnum);
1165                 spin_unlock(&ubi->wl_lock);
1166
1167                 /* Wait for the WL worker by taking the @ubi->move_mutex */
1168                 mutex_lock(&ubi->move_mutex);
1169                 mutex_unlock(&ubi->move_mutex);
1170                 goto retry;
1171         } else if (e == ubi->move_to) {
1172                 /*
1173                  * User is putting the physical eraseblock which was selected
1174                  * as the target the data is moved to. It may happen if the EBA
1175                  * unit already re-mapped the LEB in 'ubi_eba_copy_leb()' but
1176                  * the WL unit has not put the PEB to the "used" tree yet, but
1177                  * it is about to do this. So we just set a flag which will
1178                  * tell the WL worker that the PEB is not needed anymore and
1179                  * should be scheduled for erasure.
1180                  */
1181                 dbg_wl("PEB %d is the target of data moving", pnum);
1182                 ubi_assert(!ubi->move_to_put);
1183                 ubi->move_to_put = 1;
1184                 spin_unlock(&ubi->wl_lock);
1185                 return 0;
1186         } else {
1187                 if (in_wl_tree(e, &ubi->used)) {
1188                         paranoid_check_in_wl_tree(e, &ubi->used);
1189                         rb_erase(&e->rb, &ubi->used);
1190                 } else if (in_wl_tree(e, &ubi->scrub)) {
1191                         paranoid_check_in_wl_tree(e, &ubi->scrub);
1192                         rb_erase(&e->rb, &ubi->scrub);
1193                 } else {
1194                         err = prot_tree_del(ubi, e->pnum);
1195                         if (err) {
1196                                 ubi_err("PEB %d not found", pnum);
1197                                 ubi_ro_mode(ubi);
1198                                 spin_unlock(&ubi->wl_lock);
1199                                 return err;
1200                         }
1201                 }
1202         }
1203         spin_unlock(&ubi->wl_lock);
1204
1205         err = schedule_erase(ubi, e, torture);
1206         if (err) {
1207                 spin_lock(&ubi->wl_lock);
1208                 wl_tree_add(e, &ubi->used);
1209                 spin_unlock(&ubi->wl_lock);
1210         }
1211
1212         return err;
1213 }
1214
1215 /**
1216  * ubi_wl_scrub_peb - schedule a physical eraseblock for scrubbing.
1217  * @ubi: UBI device description object
1218  * @pnum: the physical eraseblock to schedule
1219  *
1220  * If a bit-flip in a physical eraseblock is detected, this physical eraseblock
1221  * needs scrubbing. This function schedules a physical eraseblock for
1222  * scrubbing which is done in background. This function returns zero in case of
1223  * success and a negative error code in case of failure.
1224  */
1225 int ubi_wl_scrub_peb(struct ubi_device *ubi, int pnum)
1226 {
1227         struct ubi_wl_entry *e;
1228
1229         ubi_msg("schedule PEB %d for scrubbing", pnum);
1230
1231 retry:
1232         spin_lock(&ubi->wl_lock);
1233         e = ubi->lookuptbl[pnum];
1234         if (e == ubi->move_from || in_wl_tree(e, &ubi->scrub)) {
1235                 spin_unlock(&ubi->wl_lock);
1236                 return 0;
1237         }
1238
1239         if (e == ubi->move_to) {
1240                 /*
1241                  * This physical eraseblock was used to move data to. The data
1242                  * was moved but the PEB was not yet inserted to the proper
1243                  * tree. We should just wait a little and let the WL worker
1244                  * proceed.
1245                  */
1246                 spin_unlock(&ubi->wl_lock);
1247                 dbg_wl("the PEB %d is not in proper tree, retry", pnum);
1248                 yield();
1249                 goto retry;
1250         }
1251
1252         if (in_wl_tree(e, &ubi->used)) {
1253                 paranoid_check_in_wl_tree(e, &ubi->used);
1254                 rb_erase(&e->rb, &ubi->used);
1255         } else {
1256                 int err;
1257
1258                 err = prot_tree_del(ubi, e->pnum);
1259                 if (err) {
1260                         ubi_err("PEB %d not found", pnum);
1261                         ubi_ro_mode(ubi);
1262                         spin_unlock(&ubi->wl_lock);
1263                         return err;
1264                 }
1265         }
1266
1267         wl_tree_add(e, &ubi->scrub);
1268         spin_unlock(&ubi->wl_lock);
1269
1270         /*
1271          * Technically scrubbing is the same as wear-leveling, so it is done
1272          * by the WL worker.
1273          */
1274         return ensure_wear_leveling(ubi);
1275 }
1276
1277 /**
1278  * ubi_wl_flush - flush all pending works.
1279  * @ubi: UBI device description object
1280  *
1281  * This function returns zero in case of success and a negative error code in
1282  * case of failure.
1283  */
1284 int ubi_wl_flush(struct ubi_device *ubi)
1285 {
1286         int err;
1287
1288         /*
1289          * Erase while the pending works queue is not empty, but not more then
1290          * the number of currently pending works.
1291          */
1292         dbg_wl("flush (%d pending works)", ubi->works_count);
1293         while (ubi->works_count) {
1294                 err = do_work(ubi);
1295                 if (err)
1296                         return err;
1297         }
1298
1299         /*
1300          * Make sure all the works which have been done in parallel are
1301          * finished.
1302          */
1303         down_write(&ubi->work_sem);
1304         up_write(&ubi->work_sem);
1305
1306         /*
1307          * And in case last was the WL worker and it cancelled the LEB
1308          * movement, flush again.
1309          */
1310         while (ubi->works_count) {
1311                 dbg_wl("flush more (%d pending works)", ubi->works_count);
1312                 err = do_work(ubi);
1313                 if (err)
1314                         return err;
1315         }
1316
1317         return 0;
1318 }
1319
1320 /**
1321  * tree_destroy - destroy an RB-tree.
1322  * @root: the root of the tree to destroy
1323  */
1324 static void tree_destroy(struct rb_root *root)
1325 {
1326         struct rb_node *rb;
1327         struct ubi_wl_entry *e;
1328
1329         rb = root->rb_node;
1330         while (rb) {
1331                 if (rb->rb_left)
1332                         rb = rb->rb_left;
1333                 else if (rb->rb_right)
1334                         rb = rb->rb_right;
1335                 else {
1336                         e = rb_entry(rb, struct ubi_wl_entry, rb);
1337
1338                         rb = rb_parent(rb);
1339                         if (rb) {
1340                                 if (rb->rb_left == &e->rb)
1341                                         rb->rb_left = NULL;
1342                                 else
1343                                         rb->rb_right = NULL;
1344                         }
1345
1346                         kmem_cache_free(ubi_wl_entry_slab, e);
1347                 }
1348         }
1349 }
1350
1351 /**
1352  * ubi_thread - UBI background thread.
1353  * @u: the UBI device description object pointer
1354  */
1355 int ubi_thread(void *u)
1356 {
1357         int failures = 0;
1358         struct ubi_device *ubi = u;
1359
1360         ubi_msg("background thread \"%s\" started, PID %d",
1361                 ubi->bgt_name, task_pid_nr(current));
1362
1363         set_freezable();
1364         for (;;) {
1365                 int err;
1366
1367                 if (kthread_should_stop())
1368                         break;
1369
1370                 if (try_to_freeze())
1371                         continue;
1372
1373                 spin_lock(&ubi->wl_lock);
1374                 if (list_empty(&ubi->works) || ubi->ro_mode ||
1375                                !ubi->thread_enabled) {
1376                         set_current_state(TASK_INTERRUPTIBLE);
1377                         spin_unlock(&ubi->wl_lock);
1378                         schedule();
1379                         continue;
1380                 }
1381                 spin_unlock(&ubi->wl_lock);
1382
1383                 err = do_work(ubi);
1384                 if (err) {
1385                         ubi_err("%s: work failed with error code %d",
1386                                 ubi->bgt_name, err);
1387                         if (failures++ > WL_MAX_FAILURES) {
1388                                 /*
1389                                  * Too many failures, disable the thread and
1390                                  * switch to read-only mode.
1391                                  */
1392                                 ubi_msg("%s: %d consecutive failures",
1393                                         ubi->bgt_name, WL_MAX_FAILURES);
1394                                 ubi_ro_mode(ubi);
1395                                 break;
1396                         }
1397                 } else
1398                         failures = 0;
1399
1400                 cond_resched();
1401         }
1402
1403         dbg_wl("background thread \"%s\" is killed", ubi->bgt_name);
1404         return 0;
1405 }
1406
1407 /**
1408  * cancel_pending - cancel all pending works.
1409  * @ubi: UBI device description object
1410  */
1411 static void cancel_pending(struct ubi_device *ubi)
1412 {
1413         while (!list_empty(&ubi->works)) {
1414                 struct ubi_work *wrk;
1415
1416                 wrk = list_entry(ubi->works.next, struct ubi_work, list);
1417                 list_del(&wrk->list);
1418                 wrk->func(ubi, wrk, 1);
1419                 ubi->works_count -= 1;
1420                 ubi_assert(ubi->works_count >= 0);
1421         }
1422 }
1423
1424 /**
1425  * ubi_wl_init_scan - initialize the wear-leveling unit using scanning
1426  * information.
1427  * @ubi: UBI device description object
1428  * @si: scanning information
1429  *
1430  * This function returns zero in case of success, and a negative error code in
1431  * case of failure.
1432  */
1433 int ubi_wl_init_scan(struct ubi_device *ubi, struct ubi_scan_info *si)
1434 {
1435         int err;
1436         struct rb_node *rb1, *rb2;
1437         struct ubi_scan_volume *sv;
1438         struct ubi_scan_leb *seb, *tmp;
1439         struct ubi_wl_entry *e;
1440
1441
1442         ubi->used = ubi->free = ubi->scrub = RB_ROOT;
1443         ubi->prot.pnum = ubi->prot.aec = RB_ROOT;
1444         spin_lock_init(&ubi->wl_lock);
1445         mutex_init(&ubi->move_mutex);
1446         init_rwsem(&ubi->work_sem);
1447         ubi->max_ec = si->max_ec;
1448         INIT_LIST_HEAD(&ubi->works);
1449
1450         sprintf(ubi->bgt_name, UBI_BGT_NAME_PATTERN, ubi->ubi_num);
1451
1452         err = -ENOMEM;
1453         ubi->lookuptbl = kzalloc(ubi->peb_count * sizeof(void *), GFP_KERNEL);
1454         if (!ubi->lookuptbl)
1455                 return err;
1456
1457         list_for_each_entry_safe(seb, tmp, &si->erase, u.list) {
1458                 cond_resched();
1459
1460                 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1461                 if (!e)
1462                         goto out_free;
1463
1464                 e->pnum = seb->pnum;
1465                 e->ec = seb->ec;
1466                 ubi->lookuptbl[e->pnum] = e;
1467                 if (schedule_erase(ubi, e, 0)) {
1468                         kmem_cache_free(ubi_wl_entry_slab, e);
1469                         goto out_free;
1470                 }
1471         }
1472
1473         list_for_each_entry(seb, &si->free, u.list) {
1474                 cond_resched();
1475
1476                 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1477                 if (!e)
1478                         goto out_free;
1479
1480                 e->pnum = seb->pnum;
1481                 e->ec = seb->ec;
1482                 ubi_assert(e->ec >= 0);
1483                 wl_tree_add(e, &ubi->free);
1484                 ubi->lookuptbl[e->pnum] = e;
1485         }
1486
1487         list_for_each_entry(seb, &si->corr, u.list) {
1488                 cond_resched();
1489
1490                 e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1491                 if (!e)
1492                         goto out_free;
1493
1494                 e->pnum = seb->pnum;
1495                 e->ec = seb->ec;
1496                 ubi->lookuptbl[e->pnum] = e;
1497                 if (schedule_erase(ubi, e, 0)) {
1498                         kmem_cache_free(ubi_wl_entry_slab, e);
1499                         goto out_free;
1500                 }
1501         }
1502
1503         ubi_rb_for_each_entry(rb1, sv, &si->volumes, rb) {
1504                 ubi_rb_for_each_entry(rb2, seb, &sv->root, u.rb) {
1505                         cond_resched();
1506
1507                         e = kmem_cache_alloc(ubi_wl_entry_slab, GFP_KERNEL);
1508                         if (!e)
1509                                 goto out_free;
1510
1511                         e->pnum = seb->pnum;
1512                         e->ec = seb->ec;
1513                         ubi->lookuptbl[e->pnum] = e;
1514                         if (!seb->scrub) {
1515                                 dbg_wl("add PEB %d EC %d to the used tree",
1516                                        e->pnum, e->ec);
1517                                 wl_tree_add(e, &ubi->used);
1518                         } else {
1519                                 dbg_wl("add PEB %d EC %d to the scrub tree",
1520                                        e->pnum, e->ec);
1521                                 wl_tree_add(e, &ubi->scrub);
1522                         }
1523                 }
1524         }
1525
1526         if (ubi->avail_pebs < WL_RESERVED_PEBS) {
1527                 ubi_err("no enough physical eraseblocks (%d, need %d)",
1528                         ubi->avail_pebs, WL_RESERVED_PEBS);
1529                 err = -ENOSPC;
1530                 goto out_free;
1531         }
1532         ubi->avail_pebs -= WL_RESERVED_PEBS;
1533         ubi->rsvd_pebs += WL_RESERVED_PEBS;
1534
1535         /* Schedule wear-leveling if needed */
1536         err = ensure_wear_leveling(ubi);
1537         if (err)
1538                 goto out_free;
1539
1540         return 0;
1541
1542 out_free:
1543         cancel_pending(ubi);
1544         tree_destroy(&ubi->used);
1545         tree_destroy(&ubi->free);
1546         tree_destroy(&ubi->scrub);
1547         kfree(ubi->lookuptbl);
1548         return err;
1549 }
1550
1551 /**
1552  * protection_trees_destroy - destroy the protection RB-trees.
1553  * @ubi: UBI device description object
1554  */
1555 static void protection_trees_destroy(struct ubi_device *ubi)
1556 {
1557         struct rb_node *rb;
1558         struct ubi_wl_prot_entry *pe;
1559
1560         rb = ubi->prot.aec.rb_node;
1561         while (rb) {
1562                 if (rb->rb_left)
1563                         rb = rb->rb_left;
1564                 else if (rb->rb_right)
1565                         rb = rb->rb_right;
1566                 else {
1567                         pe = rb_entry(rb, struct ubi_wl_prot_entry, rb_aec);
1568
1569                         rb = rb_parent(rb);
1570                         if (rb) {
1571                                 if (rb->rb_left == &pe->rb_aec)
1572                                         rb->rb_left = NULL;
1573                                 else
1574                                         rb->rb_right = NULL;
1575                         }
1576
1577                         kmem_cache_free(ubi_wl_entry_slab, pe->e);
1578                         kfree(pe);
1579                 }
1580         }
1581 }
1582
1583 /**
1584  * ubi_wl_close - close the wear-leveling unit.
1585  * @ubi: UBI device description object
1586  */
1587 void ubi_wl_close(struct ubi_device *ubi)
1588 {
1589         dbg_wl("close the UBI wear-leveling unit");
1590
1591         cancel_pending(ubi);
1592         protection_trees_destroy(ubi);
1593         tree_destroy(&ubi->used);
1594         tree_destroy(&ubi->free);
1595         tree_destroy(&ubi->scrub);
1596         kfree(ubi->lookuptbl);
1597 }
1598
1599 #ifdef CONFIG_MTD_UBI_DEBUG_PARANOID
1600
1601 /**
1602  * paranoid_check_ec - make sure that the erase counter of a physical eraseblock
1603  * is correct.
1604  * @ubi: UBI device description object
1605  * @pnum: the physical eraseblock number to check
1606  * @ec: the erase counter to check
1607  *
1608  * This function returns zero if the erase counter of physical eraseblock @pnum
1609  * is equivalent to @ec, %1 if not, and a negative error code if an error
1610  * occurred.
1611  */
1612 static int paranoid_check_ec(struct ubi_device *ubi, int pnum, int ec)
1613 {
1614         int err;
1615         long long read_ec;
1616         struct ubi_ec_hdr *ec_hdr;
1617
1618         ec_hdr = kzalloc(ubi->ec_hdr_alsize, GFP_NOFS);
1619         if (!ec_hdr)
1620                 return -ENOMEM;
1621
1622         err = ubi_io_read_ec_hdr(ubi, pnum, ec_hdr, 0);
1623         if (err && err != UBI_IO_BITFLIPS) {
1624                 /* The header does not have to exist */
1625                 err = 0;
1626                 goto out_free;
1627         }
1628
1629         read_ec = be64_to_cpu(ec_hdr->ec);
1630         if (ec != read_ec) {
1631                 ubi_err("paranoid check failed for PEB %d", pnum);
1632                 ubi_err("read EC is %lld, should be %d", read_ec, ec);
1633                 ubi_dbg_dump_stack();
1634                 err = 1;
1635         } else
1636                 err = 0;
1637
1638 out_free:
1639         kfree(ec_hdr);
1640         return err;
1641 }
1642
1643 /**
1644  * paranoid_check_in_wl_tree - make sure that a wear-leveling entry is present
1645  * in a WL RB-tree.
1646  * @e: the wear-leveling entry to check
1647  * @root: the root of the tree
1648  *
1649  * This function returns zero if @e is in the @root RB-tree and %1 if it
1650  * is not.
1651  */
1652 static int paranoid_check_in_wl_tree(struct ubi_wl_entry *e,
1653                                      struct rb_root *root)
1654 {
1655         if (in_wl_tree(e, root))
1656                 return 0;
1657
1658         ubi_err("paranoid check failed for PEB %d, EC %d, RB-tree %p ",
1659                 e->pnum, e->ec, root);
1660         ubi_dbg_dump_stack();
1661         return 1;
1662 }
1663
1664 #endif /* CONFIG_MTD_UBI_DEBUG_PARANOID */