drivers/mtd/mtdswap.c

   1 /*
   2  * Swap block device support for MTDs
   3  * Turns an MTD device into a swap device with block wear leveling
   4  *
   5  * Copyright © 2007,2011 Nokia Corporation. All rights reserved.
   6  *
   7  * Authors: Jarkko Lavinen <jarkko.lavinen@nokia.com>
   8  *
   9  * Based on Richard Purdie's earlier implementation in 2007. Background
  10  * support and lock-less operation written by Adrian Hunter.
  11  *
  12  * This program is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU General Public License
  14  * version 2 as published by the Free Software Foundation.
  15  *
  16  * This program is distributed in the hope that it will be useful, but
  17  * WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU General Public License
  22  * along with this program; if not, write to the Free Software
  23  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA
  24  * 02110-1301 USA
  25  */
  26
  27 #include <linux/kernel.h>
  28 #include <linux/module.h>
  29 #include <linux/mtd/mtd.h>
  30 #include <linux/mtd/blktrans.h>
  31 #include <linux/rbtree.h>
  32 #include <linux/sched.h>
  33 #include <linux/slab.h>
  34 #include <linux/vmalloc.h>
  35 #include <linux/genhd.h>
  36 #include <linux/swap.h>
  37 #include <linux/debugfs.h>
  38 #include <linux/seq_file.h>
  39 #include <linux/device.h>
  40 #include <linux/math64.h>
  41
  42 #define MTDSWAP_PREFIX "mtdswap"
  43
  44 /*
  45  * The number of free eraseblocks when GC should stop
  46  */
  47 #define CLEAN_BLOCK_THRESHOLD   20
  48
  49 /*
  50  * Number of free eraseblocks below which GC can also collect low frag
  51  * blocks.
  52  */
  53 #define LOW_FRAG_GC_TRESHOLD    5
  54
  55 /*
  56  * Wear level cost amortization. We want to do wear leveling on the background
  57  * without disturbing gc too much. This is made by defining max GC frequency.
  58  * Frequency value 6 means 1/6 of the GC passes will pick an erase block based
  59  * on the biggest wear difference rather than the biggest dirtiness.
  60  *
  61  * The lower freq2 should be chosen so that it makes sure the maximum erase
  62  * difference will decrease even if a malicious application is deliberately
  63  * trying to make erase differences large.
  64  */
  65 #define MAX_ERASE_DIFF          4000
  66 #define COLLECT_NONDIRTY_BASE   MAX_ERASE_DIFF
  67 #define COLLECT_NONDIRTY_FREQ1  6
  68 #define COLLECT_NONDIRTY_FREQ2  4
  69
  70 #define PAGE_UNDEF              UINT_MAX
  71 #define BLOCK_UNDEF             UINT_MAX
  72 #define BLOCK_ERROR             (UINT_MAX - 1)
  73 #define BLOCK_MAX               (UINT_MAX - 2)
  74
  75 #define EBLOCK_BAD              (1 << 0)
  76 #define EBLOCK_NOMAGIC          (1 << 1)
  77 #define EBLOCK_BITFLIP          (1 << 2)
  78 #define EBLOCK_FAILED           (1 << 3)
  79 #define EBLOCK_READERR          (1 << 4)
  80 #define EBLOCK_IDX_SHIFT        5
  81
  82 struct swap_eb {
  83         struct rb_node rb;
  84         struct rb_root *root;
  85
  86         unsigned int flags;
  87         unsigned int active_count;
  88         unsigned int erase_count;
  89         unsigned int pad;               /* speeds up pointer decrement */
  90 };
  91
  92 #define MTDSWAP_ECNT_MIN(rbroot) (rb_entry(rb_first(rbroot), struct swap_eb, \
  93                                 rb)->erase_count)
  94 #define MTDSWAP_ECNT_MAX(rbroot) (rb_entry(rb_last(rbroot), struct swap_eb, \
  95                                 rb)->erase_count)
  96
  97 struct mtdswap_tree {
  98         struct rb_root root;
  99         unsigned int count;
 100 };
 101
 102 enum {
 103         MTDSWAP_CLEAN,
 104         MTDSWAP_USED,
 105         MTDSWAP_LOWFRAG,
 106         MTDSWAP_HIFRAG,
 107         MTDSWAP_DIRTY,
 108         MTDSWAP_BITFLIP,
 109         MTDSWAP_FAILING,
 110         MTDSWAP_TREE_CNT,
 111 };
 112
 113 struct mtdswap_dev {
 114         struct mtd_blktrans_dev *mbd_dev;
 115         struct mtd_info *mtd;
 116         struct device *dev;
 117
 118         unsigned int *page_data;
 119         unsigned int *revmap;
 120
 121         unsigned int eblks;
 122         unsigned int spare_eblks;
 123         unsigned int pages_per_eblk;
 124         unsigned int max_erase_count;
 125         struct swap_eb *eb_data;
 126
 127         struct mtdswap_tree trees[MTDSWAP_TREE_CNT];
 128
 129         unsigned long long sect_read_count;
 130         unsigned long long sect_write_count;
 131         unsigned long long mtd_write_count;
 132         unsigned long long mtd_read_count;
 133         unsigned long long discard_count;
 134         unsigned long long discard_page_count;
 135
 136         unsigned int curr_write_pos;
 137         struct swap_eb *curr_write;
 138
 139         char *page_buf;
 140         char *oob_buf;
 141
 142         struct dentry *debugfs_root;
 143 };
 144
 145 struct mtdswap_oobdata {
 146         __le16 magic;
 147         __le32 count;
 148 } __attribute__((packed));
 149
 150 #define MTDSWAP_MAGIC_CLEAN     0x2095
 151 #define MTDSWAP_MAGIC_DIRTY     (MTDSWAP_MAGIC_CLEAN + 1)
 152 #define MTDSWAP_TYPE_CLEAN      0
 153 #define MTDSWAP_TYPE_DIRTY      1
 154 #define MTDSWAP_OOBSIZE         sizeof(struct mtdswap_oobdata)
 155
 156 #define MTDSWAP_ERASE_RETRIES   3 /* Before marking erase block bad */
 157 #define MTDSWAP_IO_RETRIES      3
 158
 159 enum {
 160         MTDSWAP_SCANNED_CLEAN,
 161         MTDSWAP_SCANNED_DIRTY,
 162         MTDSWAP_SCANNED_BITFLIP,
 163         MTDSWAP_SCANNED_BAD,
 164 };
 165
 166 /*
 167  * In the worst case mtdswap_writesect() has allocated the last clean
 168  * page from the current block and is then pre-empted by the GC
 169  * thread. The thread can consume a full erase block when moving a
 170  * block.
 171  */
 172 #define MIN_SPARE_EBLOCKS       2
 173 #define MIN_ERASE_BLOCKS        (MIN_SPARE_EBLOCKS + 1)
 174
 175 #define TREE_ROOT(d, name) (&d->trees[MTDSWAP_ ## name].root)
 176 #define TREE_EMPTY(d, name) (TREE_ROOT(d, name)->rb_node == NULL)
 177 #define TREE_NONEMPTY(d, name) (!TREE_EMPTY(d, name))
 178 #define TREE_COUNT(d, name) (d->trees[MTDSWAP_ ## name].count)
 179
 180 #define MTDSWAP_MBD_TO_MTDSWAP(dev) ((struct mtdswap_dev *)dev->priv)
 181
 182 static char partitions[128] = "";
 183 module_param_string(partitions, partitions, sizeof(partitions), 0444);
 184 MODULE_PARM_DESC(partitions, "MTD partition numbers to use as swap "
 185                 "partitions=\"1,3,5\"");
 186
 187 static unsigned int spare_eblocks = 10;
 188 module_param(spare_eblocks, uint, 0444);
 189 MODULE_PARM_DESC(spare_eblocks, "Percentage of spare erase blocks for "
 190                 "garbage collection (default 10%)");
 191
 192 static bool header; /* false */
 193 module_param(header, bool, 0444);
 194 MODULE_PARM_DESC(header,
 195                 "Include builtin swap header (default 0, without header)");
 196
 197 static int mtdswap_gc(struct mtdswap_dev *d, unsigned int background);
 198
 199 static loff_t mtdswap_eb_offset(struct mtdswap_dev *d, struct swap_eb *eb)
 200 {
 201         return (loff_t)(eb - d->eb_data) * d->mtd->erasesize;
 202 }
 203
 204 static void mtdswap_eb_detach(struct mtdswap_dev *d, struct swap_eb *eb)
 205 {
 206         unsigned int oldidx;
 207         struct mtdswap_tree *tp;
 208
 209         if (eb->root) {
 210                 tp = container_of(eb->root, struct mtdswap_tree, root);
 211                 oldidx = tp - &d->trees[0];
 212
 213                 d->trees[oldidx].count--;
 214                 rb_erase(&eb->rb, eb->root);
 215         }
 216 }
 217
 218 static void __mtdswap_rb_add(struct rb_root *root, struct swap_eb *eb)
 219 {
 220         struct rb_node **p, *parent = NULL;
 221         struct swap_eb *cur;
 222
 223         p = &root->rb_node;
 224         while (*p) {
 225                 parent = *p;
 226                 cur = rb_entry(parent, struct swap_eb, rb);
 227                 if (eb->erase_count > cur->erase_count)
 228                         p = &(*p)->rb_right;
 229                 else
 230                         p = &(*p)->rb_left;
 231         }
 232
 233         rb_link_node(&eb->rb, parent, p);
 234         rb_insert_color(&eb->rb, root);
 235 }
 236
 237 static void mtdswap_rb_add(struct mtdswap_dev *d, struct swap_eb *eb, int idx)
 238 {
 239         struct rb_root *root;
 240
 241         if (eb->root == &d->trees[idx].root)
 242                 return;
 243
 244         mtdswap_eb_detach(d, eb);
 245         root = &d->trees[idx].root;
 246         __mtdswap_rb_add(root, eb);
 247         eb->root = root;
 248         d->trees[idx].count++;
 249 }
 250
 251 static struct rb_node *mtdswap_rb_index(struct rb_root *root, unsigned int idx)
 252 {
 253         struct rb_node *p;
 254         unsigned int i;
 255
 256         p = rb_first(root);
 257         i = 0;
 258         while (i < idx && p) {
 259                 p = rb_next(p);
 260                 i++;
 261         }
 262
 263         return p;
 264 }
 265
 266 static int mtdswap_handle_badblock(struct mtdswap_dev *d, struct swap_eb *eb)
 267 {
 268         int ret;
 269         loff_t offset;
 270
 271         d->spare_eblks--;
 272         eb->flags |= EBLOCK_BAD;
 273         mtdswap_eb_detach(d, eb);
 274         eb->root = NULL;
 275
 276         /* badblocks not supported */
 277         if (!d->mtd->block_markbad)
 278                 return 1;
 279
 280         offset = mtdswap_eb_offset(d, eb);
 281         dev_warn(d->dev, "Marking bad block at %08llx\n", offset);
 282         ret = d->mtd->block_markbad(d->mtd, offset);
 283
 284         if (ret) {
 285                 dev_warn(d->dev, "Mark block bad failed for block at %08llx "
 286                         "error %d\n", offset, ret);
 287                 return ret;
 288         }
 289
 290         return 1;
 291
 292 }
 293
 294 static int mtdswap_handle_write_error(struct mtdswap_dev *d, struct swap_eb *eb)
 295 {
 296         unsigned int marked = eb->flags & EBLOCK_FAILED;
 297         struct swap_eb *curr_write = d->curr_write;
 298
 299         eb->flags |= EBLOCK_FAILED;
 300         if (curr_write == eb) {
 301                 d->curr_write = NULL;
 302
 303                 if (!marked && d->curr_write_pos != 0) {
 304                         mtdswap_rb_add(d, eb, MTDSWAP_FAILING);
 305                         return 0;
 306                 }
 307         }
 308
 309         return mtdswap_handle_badblock(d, eb);
 310 }
 311
 312 static int mtdswap_read_oob(struct mtdswap_dev *d, loff_t from,
 313                         struct mtd_oob_ops *ops)
 314 {
 315         int ret = d->mtd->read_oob(d->mtd, from, ops);
 316
 317         if (mtd_is_bitflip(ret))
 318                 return ret;
 319
 320         if (ret) {
 321                 dev_warn(d->dev, "Read OOB failed %d for block at %08llx\n",
 322                         ret, from);
 323                 return ret;
 324         }
 325
 326         if (ops->oobretlen < ops->ooblen) {
 327                 dev_warn(d->dev, "Read OOB return short read (%zd bytes not "
 328                         "%zd) for block at %08llx\n",
 329                         ops->oobretlen, ops->ooblen, from);
 330                 return -EIO;
 331         }
 332
 333         return 0;
 334 }
 335
 336 static int mtdswap_read_markers(struct mtdswap_dev *d, struct swap_eb *eb)
 337 {
 338         struct mtdswap_oobdata *data, *data2;
 339         int ret;
 340         loff_t offset;
 341         struct mtd_oob_ops ops;
 342
 343         offset = mtdswap_eb_offset(d, eb);
 344
 345         /* Check first if the block is bad. */
 346         if (d->mtd->block_isbad && d->mtd->block_isbad(d->mtd, offset))
 347                 return MTDSWAP_SCANNED_BAD;
 348
 349         ops.ooblen = 2 * d->mtd->ecclayout->oobavail;
 350         ops.oobbuf = d->oob_buf;
 351         ops.ooboffs = 0;
 352         ops.datbuf = NULL;
 353         ops.mode = MTD_OPS_AUTO_OOB;
 354
 355         ret = mtdswap_read_oob(d, offset, &ops);
 356
 357         if (ret && !mtd_is_bitflip(ret))
 358                 return ret;
 359
 360         data = (struct mtdswap_oobdata *)d->oob_buf;
 361         data2 = (struct mtdswap_oobdata *)
 362                 (d->oob_buf + d->mtd->ecclayout->oobavail);
 363
 364         if (le16_to_cpu(data->magic) == MTDSWAP_MAGIC_CLEAN) {
 365                 eb->erase_count = le32_to_cpu(data->count);
 366                 if (mtd_is_bitflip(ret))
 367                         ret = MTDSWAP_SCANNED_BITFLIP;
 368                 else {
 369                         if (le16_to_cpu(data2->magic) == MTDSWAP_MAGIC_DIRTY)
 370                                 ret = MTDSWAP_SCANNED_DIRTY;
 371                         else
 372                                 ret = MTDSWAP_SCANNED_CLEAN;
 373                 }
 374         } else {
 375                 eb->flags |= EBLOCK_NOMAGIC;
 376                 ret = MTDSWAP_SCANNED_DIRTY;
 377         }
 378
 379         return ret;
 380 }
 381
 382 static int mtdswap_write_marker(struct mtdswap_dev *d, struct swap_eb *eb,
 383                                 u16 marker)
 384 {
 385         struct mtdswap_oobdata n;
 386         int ret;
 387         loff_t offset;
 388         struct mtd_oob_ops ops;
 389
 390         ops.ooboffs = 0;
 391         ops.oobbuf = (uint8_t *)&n;
 392         ops.mode = MTD_OPS_AUTO_OOB;
 393         ops.datbuf = NULL;
 394
 395         if (marker == MTDSWAP_TYPE_CLEAN) {
 396                 n.magic = cpu_to_le16(MTDSWAP_MAGIC_CLEAN);
 397                 n.count = cpu_to_le32(eb->erase_count);
 398                 ops.ooblen = MTDSWAP_OOBSIZE;
 399                 offset = mtdswap_eb_offset(d, eb);
 400         } else {
 401                 n.magic = cpu_to_le16(MTDSWAP_MAGIC_DIRTY);
 402                 ops.ooblen = sizeof(n.magic);
 403                 offset = mtdswap_eb_offset(d, eb) + d->mtd->writesize;
 404         }
 405
 406         ret = d->mtd->write_oob(d->mtd, offset , &ops);
 407
 408         if (ret) {
 409                 dev_warn(d->dev, "Write OOB failed for block at %08llx "
 410                         "error %d\n", offset, ret);
 411                 if (ret == -EIO || mtd_is_eccerr(ret))
 412                         mtdswap_handle_write_error(d, eb);
 413                 return ret;
 414         }
 415
 416         if (ops.oobretlen != ops.ooblen) {
 417                 dev_warn(d->dev, "Short OOB write for block at %08llx: "
 418                         "%zd not %zd\n",
 419                         offset, ops.oobretlen, ops.ooblen);
 420                 return ret;
 421         }
 422
 423         return 0;
 424 }
 425
 426 /*
 427  * Are there any erase blocks without MAGIC_CLEAN header, presumably
 428  * because power was cut off after erase but before header write? We
 429  * need to guestimate the erase count.
 430  */
 431 static void mtdswap_check_counts(struct mtdswap_dev *d)
 432 {
 433         struct rb_root hist_root = RB_ROOT;
 434         struct rb_node *medrb;
 435         struct swap_eb *eb;
 436         unsigned int i, cnt, median;
 437
 438         cnt = 0;
 439         for (i = 0; i < d->eblks; i++) {
 440                 eb = d->eb_data + i;
 441
 442                 if (eb->flags & (EBLOCK_NOMAGIC | EBLOCK_BAD | EBLOCK_READERR))
 443                         continue;
 444
 445                 __mtdswap_rb_add(&hist_root, eb);
 446                 cnt++;
 447         }
 448
 449         if (cnt == 0)
 450                 return;
 451
 452         medrb = mtdswap_rb_index(&hist_root, cnt / 2);
 453         median = rb_entry(medrb, struct swap_eb, rb)->erase_count;
 454
 455         d->max_erase_count = MTDSWAP_ECNT_MAX(&hist_root);
 456
 457         for (i = 0; i < d->eblks; i++) {
 458                 eb = d->eb_data + i;
 459
 460                 if (eb->flags & (EBLOCK_NOMAGIC | EBLOCK_READERR))
 461                         eb->erase_count = median;
 462
 463                 if (eb->flags & (EBLOCK_NOMAGIC | EBLOCK_BAD | EBLOCK_READERR))
 464                         continue;
 465
 466                 rb_erase(&eb->rb, &hist_root);
 467         }
 468 }
 469
 470 static void mtdswap_scan_eblks(struct mtdswap_dev *d)
 471 {
 472         int status;
 473         unsigned int i, idx;
 474         struct swap_eb *eb;
 475
 476         for (i = 0; i < d->eblks; i++) {
 477                 eb = d->eb_data + i;
 478
 479                 status = mtdswap_read_markers(d, eb);
 480                 if (status < 0)
 481                         eb->flags |= EBLOCK_READERR;
 482                 else if (status == MTDSWAP_SCANNED_BAD) {
 483                         eb->flags |= EBLOCK_BAD;
 484                         continue;
 485                 }
 486
 487                 switch (status) {
 488                 case MTDSWAP_SCANNED_CLEAN:
 489                         idx = MTDSWAP_CLEAN;
 490                         break;
 491                 case MTDSWAP_SCANNED_DIRTY:
 492                 case MTDSWAP_SCANNED_BITFLIP:
 493                         idx = MTDSWAP_DIRTY;
 494                         break;
 495                 default:
 496                         idx = MTDSWAP_FAILING;
 497                 }
 498
 499                 eb->flags |= (idx << EBLOCK_IDX_SHIFT);
 500         }
 501
 502         mtdswap_check_counts(d);
 503
 504         for (i = 0; i < d->eblks; i++) {
 505                 eb = d->eb_data + i;
 506
 507                 if (eb->flags & EBLOCK_BAD)
 508                         continue;
 509
 510                 idx = eb->flags >> EBLOCK_IDX_SHIFT;
 511                 mtdswap_rb_add(d, eb, idx);
 512         }
 513 }
 514
 515 /*
 516  * Place eblk into a tree corresponding to its number of active blocks
 517  * it contains.
 518  */
 519 static void mtdswap_store_eb(struct mtdswap_dev *d, struct swap_eb *eb)
 520 {
 521         unsigned int weight = eb->active_count;
 522         unsigned int maxweight = d->pages_per_eblk;
 523
 524         if (eb == d->curr_write)
 525                 return;
 526
 527         if (eb->flags & EBLOCK_BITFLIP)
 528                 mtdswap_rb_add(d, eb, MTDSWAP_BITFLIP);
 529         else if (eb->flags & (EBLOCK_READERR | EBLOCK_FAILED))
 530                 mtdswap_rb_add(d, eb, MTDSWAP_FAILING);
 531         if (weight == maxweight)
 532                 mtdswap_rb_add(d, eb, MTDSWAP_USED);
 533         else if (weight == 0)
 534                 mtdswap_rb_add(d, eb, MTDSWAP_DIRTY);
 535         else if (weight > (maxweight/2))
 536                 mtdswap_rb_add(d, eb, MTDSWAP_LOWFRAG);
 537         else
 538                 mtdswap_rb_add(d, eb, MTDSWAP_HIFRAG);
 539 }
 540
 541
 542 static void mtdswap_erase_callback(struct erase_info *done)
 543 {
 544         wait_queue_head_t *wait_q = (wait_queue_head_t *)done->priv;
 545         wake_up(wait_q);
 546 }
 547
 548 static int mtdswap_erase_block(struct mtdswap_dev *d, struct swap_eb *eb)
 549 {
 550         struct mtd_info *mtd = d->mtd;
 551         struct erase_info erase;
 552         wait_queue_head_t wq;
 553         unsigned int retries = 0;
 554         int ret;
 555
 556         eb->erase_count++;
 557         if (eb->erase_count > d->max_erase_count)
 558                 d->max_erase_count = eb->erase_count;
 559
 560 retry:
 561         init_waitqueue_head(&wq);
 562         memset(&erase, 0, sizeof(struct erase_info));
 563
 564         erase.mtd       = mtd;
 565         erase.callback  = mtdswap_erase_callback;
 566         erase.addr      = mtdswap_eb_offset(d, eb);
 567         erase.len       = mtd->erasesize;
 568         erase.priv      = (u_long)&wq;
 569
 570         ret = mtd->erase(mtd, &erase);
 571         if (ret) {
 572                 if (retries++ < MTDSWAP_ERASE_RETRIES) {
 573                         dev_warn(d->dev,
 574                                 "erase of erase block %#llx on %s failed",
 575                                 erase.addr, mtd->name);
 576                         yield();
 577                         goto retry;
 578                 }
 579
 580                 dev_err(d->dev, "Cannot erase erase block %#llx on %s\n",
 581                         erase.addr, mtd->name);
 582
 583                 mtdswap_handle_badblock(d, eb);
 584                 return -EIO;
 585         }
 586
 587         ret = wait_event_interruptible(wq, erase.state == MTD_ERASE_DONE ||
 588                                            erase.state == MTD_ERASE_FAILED);
 589         if (ret) {
 590                 dev_err(d->dev, "Interrupted erase block %#llx erassure on %s",
 591                         erase.addr, mtd->name);
 592                 return -EINTR;
 593         }
 594
 595         if (erase.state == MTD_ERASE_FAILED) {
 596                 if (retries++ < MTDSWAP_ERASE_RETRIES) {
 597                         dev_warn(d->dev,
 598                                 "erase of erase block %#llx on %s failed",
 599                                 erase.addr, mtd->name);
 600                         yield();
 601                         goto retry;
 602                 }
 603
 604                 mtdswap_handle_badblock(d, eb);
 605                 return -EIO;
 606         }
 607
 608         return 0;
 609 }
 610
 611 static int mtdswap_map_free_block(struct mtdswap_dev *d, unsigned int page,
 612                                 unsigned int *block)
 613 {
 614         int ret;
 615         struct swap_eb *old_eb = d->curr_write;
 616         struct rb_root *clean_root;
 617         struct swap_eb *eb;
 618
 619         if (old_eb == NULL || d->curr_write_pos >= d->pages_per_eblk) {
 620                 do {
 621                         if (TREE_EMPTY(d, CLEAN))
 622                                 return -ENOSPC;
 623
 624                         clean_root = TREE_ROOT(d, CLEAN);
 625                         eb = rb_entry(rb_first(clean_root), struct swap_eb, rb);
 626                         rb_erase(&eb->rb, clean_root);
 627                         eb->root = NULL;
 628                         TREE_COUNT(d, CLEAN)--;
 629
 630                         ret = mtdswap_write_marker(d, eb, MTDSWAP_TYPE_DIRTY);
 631                 } while (ret == -EIO || mtd_is_eccerr(ret));
 632
 633                 if (ret)
 634                         return ret;
 635
 636                 d->curr_write_pos = 0;
 637                 d->curr_write = eb;
 638                 if (old_eb)
 639                         mtdswap_store_eb(d, old_eb);
 640         }
 641
 642         *block = (d->curr_write - d->eb_data) * d->pages_per_eblk +
 643                 d->curr_write_pos;
 644
 645         d->curr_write->active_count++;
 646         d->revmap[*block] = page;
 647         d->curr_write_pos++;
 648
 649         return 0;
 650 }
 651
 652 static unsigned int mtdswap_free_page_cnt(struct mtdswap_dev *d)
 653 {
 654         return TREE_COUNT(d, CLEAN) * d->pages_per_eblk +
 655                 d->pages_per_eblk - d->curr_write_pos;
 656 }
 657
 658 static unsigned int mtdswap_enough_free_pages(struct mtdswap_dev *d)
 659 {
 660         return mtdswap_free_page_cnt(d) > d->pages_per_eblk;
 661 }
 662
 663 static int mtdswap_write_block(struct mtdswap_dev *d, char *buf,
 664                         unsigned int page, unsigned int *bp, int gc_context)
 665 {
 666         struct mtd_info *mtd = d->mtd;
 667         struct swap_eb *eb;
 668         size_t retlen;
 669         loff_t writepos;
 670         int ret;
 671
 672 retry:
 673         if (!gc_context)
 674                 while (!mtdswap_enough_free_pages(d))
 675                         if (mtdswap_gc(d, 0) > 0)
 676                                 return -ENOSPC;
 677
 678         ret = mtdswap_map_free_block(d, page, bp);
 679         eb = d->eb_data + (*bp / d->pages_per_eblk);
 680
 681         if (ret == -EIO || mtd_is_eccerr(ret)) {
 682                 d->curr_write = NULL;
 683                 eb->active_count--;
 684                 d->revmap[*bp] = PAGE_UNDEF;
 685                 goto retry;
 686         }
 687
 688         if (ret < 0)
 689                 return ret;
 690
 691         writepos = (loff_t)*bp << PAGE_SHIFT;
 692         ret =  mtd->write(mtd, writepos, PAGE_SIZE, &retlen, buf);
 693         if (ret == -EIO || mtd_is_eccerr(ret)) {
 694                 d->curr_write_pos--;
 695                 eb->active_count--;
 696                 d->revmap[*bp] = PAGE_UNDEF;
 697                 mtdswap_handle_write_error(d, eb);
 698                 goto retry;
 699         }
 700
 701         if (ret < 0) {
 702                 dev_err(d->dev, "Write to MTD device failed: %d (%zd written)",
 703                         ret, retlen);
 704                 goto err;
 705         }
 706
 707         if (retlen != PAGE_SIZE) {
 708                 dev_err(d->dev, "Short write to MTD device: %zd written",
 709                         retlen);
 710                 ret = -EIO;
 711                 goto err;
 712         }
 713
 714         return ret;
 715
 716 err:
 717         d->curr_write_pos--;
 718         eb->active_count--;
 719         d->revmap[*bp] = PAGE_UNDEF;
 720
 721         return ret;
 722 }
 723
 724 static int mtdswap_move_block(struct mtdswap_dev *d, unsigned int oldblock,
 725                 unsigned int *newblock)
 726 {
 727         struct mtd_info *mtd = d->mtd;
 728         struct swap_eb *eb, *oldeb;
 729         int ret;
 730         size_t retlen;
 731         unsigned int page, retries;
 732         loff_t readpos;
 733
 734         page = d->revmap[oldblock];
 735         readpos = (loff_t) oldblock << PAGE_SHIFT;
 736         retries = 0;
 737
 738 retry:
 739         ret = mtd->read(mtd, readpos, PAGE_SIZE, &retlen, d->page_buf);
 740
 741         if (ret < 0 && !mtd_is_bitflip(ret)) {
 742                 oldeb = d->eb_data + oldblock / d->pages_per_eblk;
 743                 oldeb->flags |= EBLOCK_READERR;
 744
 745                 dev_err(d->dev, "Read Error: %d (block %u)\n", ret,
 746                         oldblock);
 747                 retries++;
 748                 if (retries < MTDSWAP_IO_RETRIES)
 749                         goto retry;
 750
 751                 goto read_error;
 752         }
 753
 754         if (retlen != PAGE_SIZE) {
 755                 dev_err(d->dev, "Short read: %zd (block %u)\n", retlen,
 756                        oldblock);
 757                 ret = -EIO;
 758                 goto read_error;
 759         }
 760
 761         ret = mtdswap_write_block(d, d->page_buf, page, newblock, 1);
 762         if (ret < 0) {
 763                 d->page_data[page] = BLOCK_ERROR;
 764                 dev_err(d->dev, "Write error: %d\n", ret);
 765                 return ret;
 766         }
 767
 768         eb = d->eb_data + *newblock / d->pages_per_eblk;
 769         d->page_data[page] = *newblock;
 770         d->revmap[oldblock] = PAGE_UNDEF;
 771         eb = d->eb_data + oldblock / d->pages_per_eblk;
 772         eb->active_count--;
 773
 774         return 0;
 775
 776 read_error:
 777         d->page_data[page] = BLOCK_ERROR;
 778         d->revmap[oldblock] = PAGE_UNDEF;
 779         return ret;
 780 }
 781
 782 static int mtdswap_gc_eblock(struct mtdswap_dev *d, struct swap_eb *eb)
 783 {
 784         unsigned int i, block, eblk_base, newblock;
 785         int ret, errcode;
 786
 787         errcode = 0;
 788         eblk_base = (eb - d->eb_data) * d->pages_per_eblk;
 789
 790         for (i = 0; i < d->pages_per_eblk; i++) {
 791                 if (d->spare_eblks < MIN_SPARE_EBLOCKS)
 792                         return -ENOSPC;
 793
 794                 block = eblk_base + i;
 795                 if (d->revmap[block] == PAGE_UNDEF)
 796                         continue;
 797
 798                 ret = mtdswap_move_block(d, block, &newblock);
 799                 if (ret < 0 && !errcode)
 800                         errcode = ret;
 801         }
 802
 803         return errcode;
 804 }
 805
 806 static int __mtdswap_choose_gc_tree(struct mtdswap_dev *d)
 807 {
 808         int idx, stopat;
 809
 810         if (TREE_COUNT(d, CLEAN) < LOW_FRAG_GC_TRESHOLD)
 811                 stopat = MTDSWAP_LOWFRAG;
 812         else
 813                 stopat = MTDSWAP_HIFRAG;
 814
 815         for (idx = MTDSWAP_BITFLIP; idx >= stopat; idx--)
 816                 if (d->trees[idx].root.rb_node != NULL)
 817                         return idx;
 818
 819         return -1;
 820 }
 821
 822 static int mtdswap_wlfreq(unsigned int maxdiff)
 823 {
 824         unsigned int h, x, y, dist, base;
 825
 826         /*
 827          * Calculate linear ramp down from f1 to f2 when maxdiff goes from
 828          * MAX_ERASE_DIFF to MAX_ERASE_DIFF + COLLECT_NONDIRTY_BASE.  Similar
 829          * to triangle with height f1 - f1 and width COLLECT_NONDIRTY_BASE.
 830          */
 831
 832         dist = maxdiff - MAX_ERASE_DIFF;
 833         if (dist > COLLECT_NONDIRTY_BASE)
 834                 dist = COLLECT_NONDIRTY_BASE;
 835
 836         /*
 837          * Modelling the slop as right angular triangle with base
 838          * COLLECT_NONDIRTY_BASE and height freq1 - freq2. The ratio y/x is
 839          * equal to the ratio h/base.
 840          */
 841         h = COLLECT_NONDIRTY_FREQ1 - COLLECT_NONDIRTY_FREQ2;
 842         base = COLLECT_NONDIRTY_BASE;
 843
 844         x = dist - base;
 845         y = (x * h + base / 2) / base;
 846
 847         return COLLECT_NONDIRTY_FREQ2 + y;
 848 }
 849
 850 static int mtdswap_choose_wl_tree(struct mtdswap_dev *d)
 851 {
 852         static unsigned int pick_cnt;
 853         unsigned int i, idx = -1, wear, max;
 854         struct rb_root *root;
 855
 856         max = 0;
 857         for (i = 0; i <= MTDSWAP_DIRTY; i++) {
 858                 root = &d->trees[i].root;
 859                 if (root->rb_node == NULL)
 860                         continue;
 861
 862                 wear = d->max_erase_count - MTDSWAP_ECNT_MIN(root);
 863                 if (wear > max) {
 864                         max = wear;
 865                         idx = i;
 866                 }
 867         }
 868
 869         if (max > MAX_ERASE_DIFF && pick_cnt >= mtdswap_wlfreq(max) - 1) {
 870                 pick_cnt = 0;
 871                 return idx;
 872         }
 873
 874         pick_cnt++;
 875         return -1;
 876 }
 877
 878 static int mtdswap_choose_gc_tree(struct mtdswap_dev *d,
 879                                 unsigned int background)
 880 {
 881         int idx;
 882
 883         if (TREE_NONEMPTY(d, FAILING) &&
 884                 (background || (TREE_EMPTY(d, CLEAN) && TREE_EMPTY(d, DIRTY))))
 885                 return MTDSWAP_FAILING;
 886
 887         idx = mtdswap_choose_wl_tree(d);
 888         if (idx >= MTDSWAP_CLEAN)
 889                 return idx;
 890
 891         return __mtdswap_choose_gc_tree(d);
 892 }
 893
 894 static struct swap_eb *mtdswap_pick_gc_eblk(struct mtdswap_dev *d,
 895                                         unsigned int background)
 896 {
 897         struct rb_root *rp = NULL;
 898         struct swap_eb *eb = NULL;
 899         int idx;
 900
 901         if (background && TREE_COUNT(d, CLEAN) > CLEAN_BLOCK_THRESHOLD &&
 902                 TREE_EMPTY(d, DIRTY) && TREE_EMPTY(d, FAILING))
 903                 return NULL;
 904
 905         idx = mtdswap_choose_gc_tree(d, background);
 906         if (idx < 0)
 907                 return NULL;
 908
 909         rp = &d->trees[idx].root;
 910         eb = rb_entry(rb_first(rp), struct swap_eb, rb);
 911
 912         rb_erase(&eb->rb, rp);
 913         eb->root = NULL;
 914         d->trees[idx].count--;
 915         return eb;
 916 }
 917
 918 static unsigned int mtdswap_test_patt(unsigned int i)
 919 {
 920         return i % 2 ? 0x55555555 : 0xAAAAAAAA;
 921 }
 922
 923 static unsigned int mtdswap_eblk_passes(struct mtdswap_dev *d,
 924                                         struct swap_eb *eb)
 925 {
 926         struct mtd_info *mtd = d->mtd;
 927         unsigned int test, i, j, patt, mtd_pages;
 928         loff_t base, pos;
 929         unsigned int *p1 = (unsigned int *)d->page_buf;
 930         unsigned char *p2 = (unsigned char *)d->oob_buf;
 931         struct mtd_oob_ops ops;
 932         int ret;
 933
 934         ops.mode = MTD_OPS_AUTO_OOB;
 935         ops.len = mtd->writesize;
 936         ops.ooblen = mtd->ecclayout->oobavail;
 937         ops.ooboffs = 0;
 938         ops.datbuf = d->page_buf;
 939         ops.oobbuf = d->oob_buf;
 940         base = mtdswap_eb_offset(d, eb);
 941         mtd_pages = d->pages_per_eblk * PAGE_SIZE / mtd->writesize;
 942
 943         for (test = 0; test < 2; test++) {
 944                 pos = base;
 945                 for (i = 0; i < mtd_pages; i++) {
 946                         patt = mtdswap_test_patt(test + i);
 947                         memset(d->page_buf, patt, mtd->writesize);
 948                         memset(d->oob_buf, patt, mtd->ecclayout->oobavail);
 949                         ret = mtd->write_oob(mtd, pos, &ops);
 950                         if (ret)
 951                                 goto error;
 952
 953                         pos += mtd->writesize;
 954                 }
 955
 956                 pos = base;
 957                 for (i = 0; i < mtd_pages; i++) {
 958                         ret = mtd->read_oob(mtd, pos, &ops);
 959                         if (ret)
 960                                 goto error;
 961
 962                         patt = mtdswap_test_patt(test + i);
 963                         for (j = 0; j < mtd->writesize/sizeof(int); j++)
 964                                 if (p1[j] != patt)
 965                                         goto error;
 966
 967                         for (j = 0; j < mtd->ecclayout->oobavail; j++)
 968                                 if (p2[j] != (unsigned char)patt)
 969                                         goto error;
 970
 971                         pos += mtd->writesize;
 972                 }
 973
 974                 ret = mtdswap_erase_block(d, eb);
 975                 if (ret)
 976                         goto error;
 977         }
 978
 979         eb->flags &= ~EBLOCK_READERR;
 980         return 1;
 981
 982 error:
 983         mtdswap_handle_badblock(d, eb);
 984         return 0;
 985 }
 986
 987 static int mtdswap_gc(struct mtdswap_dev *d, unsigned int background)
 988 {
 989         struct swap_eb *eb;
 990         int ret;
 991
 992         if (d->spare_eblks < MIN_SPARE_EBLOCKS)
 993                 return 1;
 994
 995         eb = mtdswap_pick_gc_eblk(d, background);
 996         if (!eb)
 997                 return 1;
 998
 999         ret = mtdswap_gc_eblock(d, eb);
1000         if (ret == -ENOSPC)
1001                 return 1;
1002
1003         if (eb->flags & EBLOCK_FAILED) {
1004                 mtdswap_handle_badblock(d, eb);
1005                 return 0;
1006         }
1007
1008         eb->flags &= ~EBLOCK_BITFLIP;
1009         ret = mtdswap_erase_block(d, eb);
1010         if ((eb->flags & EBLOCK_READERR) &&
1011                 (ret || !mtdswap_eblk_passes(d, eb)))
1012                 return 0;
1013
1014         if (ret == 0)
1015                 ret = mtdswap_write_marker(d, eb, MTDSWAP_TYPE_CLEAN);
1016
1017         if (ret == 0)
1018                 mtdswap_rb_add(d, eb, MTDSWAP_CLEAN);
1019         else if (ret != -EIO && !mtd_is_eccerr(ret))
1020                 mtdswap_rb_add(d, eb, MTDSWAP_DIRTY);
1021
1022         return 0;
1023 }
1024
1025 static void mtdswap_background(struct mtd_blktrans_dev *dev)
1026 {
1027         struct mtdswap_dev *d = MTDSWAP_MBD_TO_MTDSWAP(dev);
1028         int ret;
1029
1030         while (1) {
1031                 ret = mtdswap_gc(d, 1);
1032                 if (ret || mtd_blktrans_cease_background(dev))
1033                         return;
1034         }
1035 }
1036
1037 static void mtdswap_cleanup(struct mtdswap_dev *d)
1038 {
1039         vfree(d->eb_data);
1040         vfree(d->revmap);
1041         vfree(d->page_data);
1042         kfree(d->oob_buf);
1043         kfree(d->page_buf);
1044 }
1045
1046 static int mtdswap_flush(struct mtd_blktrans_dev *dev)
1047 {
1048         struct mtdswap_dev *d = MTDSWAP_MBD_TO_MTDSWAP(dev);
1049
1050         if (d->mtd->sync)
1051                 d->mtd->sync(d->mtd);
1052         return 0;
1053 }
1054
1055 static unsigned int mtdswap_badblocks(struct mtd_info *mtd, uint64_t size)
1056 {
1057         loff_t offset;
1058         unsigned int badcnt;
1059
1060         badcnt = 0;
1061
1062         if (mtd->block_isbad)
1063                 for (offset = 0; offset < size; offset += mtd->erasesize)
1064                         if (mtd->block_isbad(mtd, offset))
1065                                 badcnt++;
1066
1067         return badcnt;
1068 }
1069
1070 static int mtdswap_writesect(struct mtd_blktrans_dev *dev,
1071                         unsigned long page, char *buf)
1072 {
1073         struct mtdswap_dev *d = MTDSWAP_MBD_TO_MTDSWAP(dev);
1074         unsigned int newblock, mapped;
1075         struct swap_eb *eb;
1076         int ret;
1077
1078         d->sect_write_count++;
1079
1080         if (d->spare_eblks < MIN_SPARE_EBLOCKS)
1081                 return -ENOSPC;
1082
1083         if (header) {
1084                 /* Ignore writes to the header page */
1085                 if (unlikely(page == 0))
1086                         return 0;
1087
1088                 page--;
1089         }
1090
1091         mapped = d->page_data[page];
1092         if (mapped <= BLOCK_MAX) {
1093                 eb = d->eb_data + (mapped / d->pages_per_eblk);
1094                 eb->active_count--;
1095                 mtdswap_store_eb(d, eb);
1096                 d->page_data[page] = BLOCK_UNDEF;
1097                 d->revmap[mapped] = PAGE_UNDEF;
1098         }
1099
1100         ret = mtdswap_write_block(d, buf, page, &newblock, 0);
1101         d->mtd_write_count++;
1102
1103         if (ret < 0)
1104                 return ret;
1105
1106         eb = d->eb_data + (newblock / d->pages_per_eblk);
1107         d->page_data[page] = newblock;
1108
1109         return 0;
1110 }
1111
1112 /* Provide a dummy swap header for the kernel */
1113 static int mtdswap_auto_header(struct mtdswap_dev *d, char *buf)
1114 {
1115         union swap_header *hd = (union swap_header *)(buf);
1116
1117         memset(buf, 0, PAGE_SIZE - 10);
1118
1119         hd->info.version = 1;
1120         hd->info.last_page = d->mbd_dev->size - 1;
1121         hd->info.nr_badpages = 0;
1122
1123         memcpy(buf + PAGE_SIZE - 10, "SWAPSPACE2", 10);
1124
1125         return 0;
1126 }
1127
1128 static int mtdswap_readsect(struct mtd_blktrans_dev *dev,
1129                         unsigned long page, char *buf)
1130 {
1131         struct mtdswap_dev *d = MTDSWAP_MBD_TO_MTDSWAP(dev);
1132         struct mtd_info *mtd = d->mtd;
1133         unsigned int realblock, retries;
1134         loff_t readpos;
1135         struct swap_eb *eb;
1136         size_t retlen;
1137         int ret;
1138
1139         d->sect_read_count++;
1140
1141         if (header) {
1142                 if (unlikely(page == 0))
1143                         return mtdswap_auto_header(d, buf);
1144
1145                 page--;
1146         }
1147
1148         realblock = d->page_data[page];
1149         if (realblock > BLOCK_MAX) {
1150                 memset(buf, 0x0, PAGE_SIZE);
1151                 if (realblock == BLOCK_UNDEF)
1152                         return 0;
1153                 else
1154                         return -EIO;
1155         }
1156
1157         eb = d->eb_data + (realblock / d->pages_per_eblk);
1158         BUG_ON(d->revmap[realblock] == PAGE_UNDEF);
1159
1160         readpos = (loff_t)realblock << PAGE_SHIFT;
1161         retries = 0;
1162
1163 retry:
1164         ret = mtd->read(mtd, readpos, PAGE_SIZE, &retlen, buf);
1165
1166         d->mtd_read_count++;
1167         if (mtd_is_bitflip(ret)) {
1168                 eb->flags |= EBLOCK_BITFLIP;
1169                 mtdswap_rb_add(d, eb, MTDSWAP_BITFLIP);
1170                 ret = 0;
1171         }
1172
1173         if (ret < 0) {
1174                 dev_err(d->dev, "Read error %d\n", ret);
1175                 eb->flags |= EBLOCK_READERR;
1176                 mtdswap_rb_add(d, eb, MTDSWAP_FAILING);
1177                 retries++;
1178                 if (retries < MTDSWAP_IO_RETRIES)
1179                         goto retry;
1180
1181                 return ret;
1182         }
1183
1184         if (retlen != PAGE_SIZE) {
1185                 dev_err(d->dev, "Short read %zd\n", retlen);
1186                 return -EIO;
1187         }
1188
1189         return 0;
1190 }
1191
1192 static int mtdswap_discard(struct mtd_blktrans_dev *dev, unsigned long first,
1193                         unsigned nr_pages)
1194 {
1195         struct mtdswap_dev *d = MTDSWAP_MBD_TO_MTDSWAP(dev);
1196         unsigned long page;
1197         struct swap_eb *eb;
1198         unsigned int mapped;
1199
1200         d->discard_count++;
1201
1202         for (page = first; page < first + nr_pages; page++) {
1203                 mapped = d->page_data[page];
1204                 if (mapped <= BLOCK_MAX) {
1205                         eb = d->eb_data + (mapped / d->pages_per_eblk);
1206                         eb->active_count--;
1207                         mtdswap_store_eb(d, eb);
1208                         d->page_data[page] = BLOCK_UNDEF;
1209                         d->revmap[mapped] = PAGE_UNDEF;
1210                         d->discard_page_count++;
1211                 } else if (mapped == BLOCK_ERROR) {
1212                         d->page_data[page] = BLOCK_UNDEF;
1213                         d->discard_page_count++;
1214                 }
1215         }
1216
1217         return 0;
1218 }
1219
1220 static int mtdswap_show(struct seq_file *s, void *data)
1221 {
1222         struct mtdswap_dev *d = (struct mtdswap_dev *) s->private;
1223         unsigned long sum;
1224         unsigned int count[MTDSWAP_TREE_CNT];
1225         unsigned int min[MTDSWAP_TREE_CNT];
1226         unsigned int max[MTDSWAP_TREE_CNT];
1227         unsigned int i, cw = 0, cwp = 0, cwecount = 0, bb_cnt, mapped, pages;
1228         uint64_t use_size;
1229         char *name[] = {"clean", "used", "low", "high", "dirty", "bitflip",
1230                         "failing"};
1231
1232         mutex_lock(&d->mbd_dev->lock);
1233
1234         for (i = 0; i < MTDSWAP_TREE_CNT; i++) {
1235                 struct rb_root *root = &d->trees[i].root;
1236
1237                 if (root->rb_node) {
1238                         count[i] = d->trees[i].count;
1239                         min[i] = rb_entry(rb_first(root), struct swap_eb,
1240                                         rb)->erase_count;
1241                         max[i] = rb_entry(rb_last(root), struct swap_eb,
1242                                         rb)->erase_count;
1243                 } else
1244                         count[i] = 0;
1245         }
1246
1247         if (d->curr_write) {
1248                 cw = 1;
1249                 cwp = d->curr_write_pos;
1250                 cwecount = d->curr_write->erase_count;
1251         }
1252
1253         sum = 0;
1254         for (i = 0; i < d->eblks; i++)
1255                 sum += d->eb_data[i].erase_count;
1256
1257         use_size = (uint64_t)d->eblks * d->mtd->erasesize;
1258         bb_cnt = mtdswap_badblocks(d->mtd, use_size);
1259
1260         mapped = 0;
1261         pages = d->mbd_dev->size;
1262         for (i = 0; i < pages; i++)
1263                 if (d->page_data[i] != BLOCK_UNDEF)
1264                         mapped++;
1265
1266         mutex_unlock(&d->mbd_dev->lock);
1267
1268         for (i = 0; i < MTDSWAP_TREE_CNT; i++) {
1269                 if (!count[i])
1270                         continue;
1271
1272                 if (min[i] != max[i])
1273                         seq_printf(s, "%s:\t%5d erase blocks, erased min %d, "
1274                                 "max %d times\n",
1275                                 name[i], count[i], min[i], max[i]);
1276                 else
1277                         seq_printf(s, "%s:\t%5d erase blocks, all erased %d "
1278                                 "times\n", name[i], count[i], min[i]);
1279         }
1280
1281         if (bb_cnt)
1282                 seq_printf(s, "bad:\t%5u erase blocks\n", bb_cnt);
1283
1284         if (cw)
1285                 seq_printf(s, "current erase block: %u pages used, %u free, "
1286                         "erased %u times\n",
1287                         cwp, d->pages_per_eblk - cwp, cwecount);
1288
1289         seq_printf(s, "total erasures: %lu\n", sum);
1290
1291         seq_printf(s, "\n");
1292
1293         seq_printf(s, "mtdswap_readsect count: %llu\n", d->sect_read_count);
1294         seq_printf(s, "mtdswap_writesect count: %llu\n", d->sect_write_count);
1295         seq_printf(s, "mtdswap_discard count: %llu\n", d->discard_count);
1296         seq_printf(s, "mtd read count: %llu\n", d->mtd_read_count);
1297         seq_printf(s, "mtd write count: %llu\n", d->mtd_write_count);
1298         seq_printf(s, "discarded pages count: %llu\n", d->discard_page_count);
1299
1300         seq_printf(s, "\n");
1301         seq_printf(s, "total pages: %u\n", pages);
1302         seq_printf(s, "pages mapped: %u\n", mapped);
1303
1304         return 0;
1305 }
1306
1307 static int mtdswap_open(struct inode *inode, struct file *file)
1308 {
1309         return single_open(file, mtdswap_show, inode->i_private);
1310 }
1311
1312 static const struct file_operations mtdswap_fops = {
1313         .open           = mtdswap_open,
1314         .read           = seq_read,
1315         .llseek         = seq_lseek,
1316         .release        = single_release,
1317 };
1318
1319 static int mtdswap_add_debugfs(struct mtdswap_dev *d)
1320 {
1321         struct gendisk *gd = d->mbd_dev->disk;
1322         struct device *dev = disk_to_dev(gd);
1323
1324         struct dentry *root;
1325         struct dentry *dent;
1326
1327         root = debugfs_create_dir(gd->disk_name, NULL);
1328         if (IS_ERR(root))
1329                 return 0;
1330
1331         if (!root) {
1332                 dev_err(dev, "failed to initialize debugfs\n");
1333                 return -1;
1334         }
1335
1336         d->debugfs_root = root;
1337
1338         dent = debugfs_create_file("stats", S_IRUSR, root, d,
1339                                 &mtdswap_fops);
1340         if (!dent) {
1341                 dev_err(d->dev, "debugfs_create_file failed\n");
1342                 debugfs_remove_recursive(root);
1343                 d->debugfs_root = NULL;
1344                 return -1;
1345         }
1346
1347         return 0;
1348 }
1349
1350 static int mtdswap_init(struct mtdswap_dev *d, unsigned int eblocks,
1351                         unsigned int spare_cnt)
1352 {
1353         struct mtd_info *mtd = d->mbd_dev->mtd;
1354         unsigned int i, eblk_bytes, pages, blocks;
1355         int ret = -ENOMEM;
1356
1357         d->mtd = mtd;
1358         d->eblks = eblocks;
1359         d->spare_eblks = spare_cnt;
1360         d->pages_per_eblk = mtd->erasesize >> PAGE_SHIFT;
1361
1362         pages = d->mbd_dev->size;
1363         blocks = eblocks * d->pages_per_eblk;
1364
1365         for (i = 0; i < MTDSWAP_TREE_CNT; i++)
1366                 d->trees[i].root = RB_ROOT;
1367
1368         d->page_data = vmalloc(sizeof(int)*pages);
1369         if (!d->page_data)
1370                 goto page_data_fail;
1371
1372         d->revmap = vmalloc(sizeof(int)*blocks);
1373         if (!d->revmap)
1374                 goto revmap_fail;
1375
1376         eblk_bytes = sizeof(struct swap_eb)*d->eblks;
1377         d->eb_data = vzalloc(eblk_bytes);
1378         if (!d->eb_data)
1379                 goto eb_data_fail;
1380
1381         for (i = 0; i < pages; i++)
1382                 d->page_data[i] = BLOCK_UNDEF;
1383
1384         for (i = 0; i < blocks; i++)
1385                 d->revmap[i] = PAGE_UNDEF;
1386
1387         d->page_buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
1388         if (!d->page_buf)
1389                 goto page_buf_fail;
1390
1391         d->oob_buf = kmalloc(2 * mtd->ecclayout->oobavail, GFP_KERNEL);
1392         if (!d->oob_buf)
1393                 goto oob_buf_fail;
1394
1395         mtdswap_scan_eblks(d);
1396
1397         return 0;
1398
1399 oob_buf_fail:
1400         kfree(d->page_buf);
1401 page_buf_fail:
1402         vfree(d->eb_data);
1403 eb_data_fail:
1404         vfree(d->revmap);
1405 revmap_fail:
1406         vfree(d->page_data);
1407 page_data_fail:
1408         printk(KERN_ERR "%s: init failed (%d)\n", MTDSWAP_PREFIX, ret);
1409         return ret;
1410 }
1411
1412 static void mtdswap_add_mtd(struct mtd_blktrans_ops *tr, struct mtd_info *mtd)
1413 {
1414         struct mtdswap_dev *d;
1415         struct mtd_blktrans_dev *mbd_dev;
1416         char *parts;
1417         char *this_opt;
1418         unsigned long part;
1419         unsigned int eblocks, eavailable, bad_blocks, spare_cnt;
1420         uint64_t swap_size, use_size, size_limit;
1421         struct nand_ecclayout *oinfo;
1422         int ret;
1423
1424         parts = &partitions[0];
1425         if (!*parts)
1426                 return;
1427
1428         while ((this_opt = strsep(&parts, ",")) != NULL) {
1429                 if (strict_strtoul(this_opt, 0, &part) < 0)
1430                         return;
1431
1432                 if (mtd->index == part)
1433                         break;
1434         }
1435
1436         if (mtd->index != part)
1437                 return;
1438
1439         if (mtd->erasesize < PAGE_SIZE || mtd->erasesize % PAGE_SIZE) {
1440                 printk(KERN_ERR "%s: Erase size %u not multiple of PAGE_SIZE "
1441                         "%lu\n", MTDSWAP_PREFIX, mtd->erasesize, PAGE_SIZE);
1442                 return;
1443         }
1444
1445         if (PAGE_SIZE % mtd->writesize || mtd->writesize > PAGE_SIZE) {
1446                 printk(KERN_ERR "%s: PAGE_SIZE %lu not multiple of write size"
1447                         " %u\n", MTDSWAP_PREFIX, PAGE_SIZE, mtd->writesize);
1448                 return;
1449         }
1450
1451         oinfo = mtd->ecclayout;
1452         if (!oinfo) {
1453                 printk(KERN_ERR "%s: mtd%d does not have OOB\n",
1454                         MTDSWAP_PREFIX, mtd->index);
1455                 return;
1456         }
1457
1458         if (!mtd->oobsize || oinfo->oobavail < MTDSWAP_OOBSIZE) {
1459                 printk(KERN_ERR "%s: Not enough free bytes in OOB, "
1460                         "%d available, %zu needed.\n",
1461                         MTDSWAP_PREFIX, oinfo->oobavail, MTDSWAP_OOBSIZE);
1462                 return;
1463         }
1464
1465         if (spare_eblocks > 100)
1466                 spare_eblocks = 100;
1467
1468         use_size = mtd->size;
1469         size_limit = (uint64_t) BLOCK_MAX * PAGE_SIZE;
1470
1471         if (mtd->size > size_limit) {
1472                 printk(KERN_WARNING "%s: Device too large. Limiting size to "
1473                         "%llu bytes\n", MTDSWAP_PREFIX, size_limit);
1474                 use_size = size_limit;
1475         }
1476
1477         eblocks = mtd_div_by_eb(use_size, mtd);
1478         use_size = eblocks * mtd->erasesize;
1479         bad_blocks = mtdswap_badblocks(mtd, use_size);
1480         eavailable = eblocks - bad_blocks;
1481
1482         if (eavailable < MIN_ERASE_BLOCKS) {
1483                 printk(KERN_ERR "%s: Not enough erase blocks. %u available, "
1484                         "%d needed\n", MTDSWAP_PREFIX, eavailable,
1485                         MIN_ERASE_BLOCKS);
1486                 return;
1487         }
1488
1489         spare_cnt = div_u64((uint64_t)eavailable * spare_eblocks, 100);
1490
1491         if (spare_cnt < MIN_SPARE_EBLOCKS)
1492                 spare_cnt = MIN_SPARE_EBLOCKS;
1493
1494         if (spare_cnt > eavailable - 1)
1495                 spare_cnt = eavailable - 1;
1496
1497         swap_size = (uint64_t)(eavailable - spare_cnt) * mtd->erasesize +
1498                 (header ? PAGE_SIZE : 0);
1499
1500         printk(KERN_INFO "%s: Enabling MTD swap on device %lu, size %llu KB, "
1501                 "%u spare, %u bad blocks\n",
1502                 MTDSWAP_PREFIX, part, swap_size / 1024, spare_cnt, bad_blocks);
1503
1504         d = kzalloc(sizeof(struct mtdswap_dev), GFP_KERNEL);
1505         if (!d)
1506                 return;
1507
1508         mbd_dev = kzalloc(sizeof(struct mtd_blktrans_dev), GFP_KERNEL);
1509         if (!mbd_dev) {
1510                 kfree(d);
1511                 return;
1512         }
1513
1514         d->mbd_dev = mbd_dev;
1515         mbd_dev->priv = d;
1516
1517         mbd_dev->mtd = mtd;
1518         mbd_dev->devnum = mtd->index;
1519         mbd_dev->size = swap_size >> PAGE_SHIFT;
1520         mbd_dev->tr = tr;
1521
1522         if (!(mtd->flags & MTD_WRITEABLE))
1523                 mbd_dev->readonly = 1;
1524
1525         if (mtdswap_init(d, eblocks, spare_cnt) < 0)
1526                 goto init_failed;
1527
1528         if (add_mtd_blktrans_dev(mbd_dev) < 0)
1529                 goto cleanup;
1530
1531         d->dev = disk_to_dev(mbd_dev->disk);
1532
1533         ret = mtdswap_add_debugfs(d);
1534         if (ret < 0)
1535                 goto debugfs_failed;
1536
1537         return;
1538
1539 debugfs_failed:
1540         del_mtd_blktrans_dev(mbd_dev);
1541
1542 cleanup:
1543         mtdswap_cleanup(d);
1544
1545 init_failed:
1546         kfree(mbd_dev);
1547         kfree(d);
1548 }
1549
1550 static void mtdswap_remove_dev(struct mtd_blktrans_dev *dev)
1551 {
1552         struct mtdswap_dev *d = MTDSWAP_MBD_TO_MTDSWAP(dev);
1553
1554         debugfs_remove_recursive(d->debugfs_root);
1555         del_mtd_blktrans_dev(dev);
1556         mtdswap_cleanup(d);
1557         kfree(d);
1558 }
1559
1560 static struct mtd_blktrans_ops mtdswap_ops = {
1561         .name           = "mtdswap",
1562         .major          = 0,
1563         .part_bits      = 0,
1564         .blksize        = PAGE_SIZE,
1565         .flush          = mtdswap_flush,
1566         .readsect       = mtdswap_readsect,
1567         .writesect      = mtdswap_writesect,
1568         .discard        = mtdswap_discard,
1569         .background     = mtdswap_background,
1570         .add_mtd        = mtdswap_add_mtd,
1571         .remove_dev     = mtdswap_remove_dev,
1572         .owner          = THIS_MODULE,
1573 };
1574
1575 static int __init mtdswap_modinit(void)
1576 {
1577         return register_mtd_blktrans(&mtdswap_ops);
1578 }
1579
1580 static void __exit mtdswap_modexit(void)
1581 {
1582         deregister_mtd_blktrans(&mtdswap_ops);
1583 }
1584
1585 module_init(mtdswap_modinit);
1586 module_exit(mtdswap_modexit);
1587
1588
1589 MODULE_LICENSE("GPL");
1590 MODULE_AUTHOR("Jarkko Lavinen <jarkko.lavinen@nokia.com>");
1591 MODULE_DESCRIPTION("Block device access to an MTD suitable for using as "
1592                 "swap space");