fs/jffs2/gc.c

   1 /*
   2  * JFFS2 -- Journalling Flash File System, Version 2.
   3  *
   4  * Copyright © 2001-2007 Red Hat, Inc.
   5  *
   6  * Created by David Woodhouse <dwmw2@infradead.org>
   7  *
   8  * For licensing information, see the file 'LICENCE' in this directory.
   9  *
  10  */
  11
  12 #include <linux/kernel.h>
  13 #include <linux/mtd/mtd.h>
  14 #include <linux/slab.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/crc32.h>
  17 #include <linux/compiler.h>
  18 #include <linux/stat.h>
  19 #include "nodelist.h"
  20 #include "compr.h"
  21
  22 static int jffs2_garbage_collect_pristine(struct jffs2_sb_info *c,
  23                                           struct jffs2_inode_cache *ic,
  24                                           struct jffs2_raw_node_ref *raw);
  25 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
  26                                         struct jffs2_inode_info *f, struct jffs2_full_dnode *fd);
  27 static int jffs2_garbage_collect_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
  28                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd);
  29 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
  30                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd);
  31 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
  32                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
  33                                       uint32_t start, uint32_t end);
  34 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
  35                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
  36                                        uint32_t start, uint32_t end);
  37 static int jffs2_garbage_collect_live(struct jffs2_sb_info *c,  struct jffs2_eraseblock *jeb,
  38                                struct jffs2_raw_node_ref *raw, struct jffs2_inode_info *f);
  39
  40 /* Called with erase_completion_lock held */
  41 static struct jffs2_eraseblock *jffs2_find_gc_block(struct jffs2_sb_info *c)
  42 {
  43         struct jffs2_eraseblock *ret;
  44         struct list_head *nextlist = NULL;
  45         int n = jiffies % 128;
  46
  47         /* Pick an eraseblock to garbage collect next. This is where we'll
  48            put the clever wear-levelling algorithms. Eventually.  */
  49         /* We possibly want to favour the dirtier blocks more when the
  50            number of free blocks is low. */
  51 again:
  52         if (!list_empty(&c->bad_used_list) && c->nr_free_blocks > c->resv_blocks_gcbad) {
  53                 D1(printk(KERN_DEBUG "Picking block from bad_used_list to GC next\n"));
  54                 nextlist = &c->bad_used_list;
  55         } else if (n < 50 && !list_empty(&c->erasable_list)) {
  56                 /* Note that most of them will have gone directly to be erased.
  57                    So don't favour the erasable_list _too_ much. */
  58                 D1(printk(KERN_DEBUG "Picking block from erasable_list to GC next\n"));
  59                 nextlist = &c->erasable_list;
  60         } else if (n < 110 && !list_empty(&c->very_dirty_list)) {
  61                 /* Most of the time, pick one off the very_dirty list */
  62                 D1(printk(KERN_DEBUG "Picking block from very_dirty_list to GC next\n"));
  63                 nextlist = &c->very_dirty_list;
  64         } else if (n < 126 && !list_empty(&c->dirty_list)) {
  65                 D1(printk(KERN_DEBUG "Picking block from dirty_list to GC next\n"));
  66                 nextlist = &c->dirty_list;
  67         } else if (!list_empty(&c->clean_list)) {
  68                 D1(printk(KERN_DEBUG "Picking block from clean_list to GC next\n"));
  69                 nextlist = &c->clean_list;
  70         } else if (!list_empty(&c->dirty_list)) {
  71                 D1(printk(KERN_DEBUG "Picking block from dirty_list to GC next (clean_list was empty)\n"));
  72
  73                 nextlist = &c->dirty_list;
  74         } else if (!list_empty(&c->very_dirty_list)) {
  75                 D1(printk(KERN_DEBUG "Picking block from very_dirty_list to GC next (clean_list and dirty_list were empty)\n"));
  76                 nextlist = &c->very_dirty_list;
  77         } else if (!list_empty(&c->erasable_list)) {
  78                 D1(printk(KERN_DEBUG "Picking block from erasable_list to GC next (clean_list and {very_,}dirty_list were empty)\n"));
  79
  80                 nextlist = &c->erasable_list;
  81         } else if (!list_empty(&c->erasable_pending_wbuf_list)) {
  82                 /* There are blocks are wating for the wbuf sync */
  83                 D1(printk(KERN_DEBUG "Synching wbuf in order to reuse erasable_pending_wbuf_list blocks\n"));
  84                 spin_unlock(&c->erase_completion_lock);
  85                 jffs2_flush_wbuf_pad(c);
  86                 spin_lock(&c->erase_completion_lock);
  87                 goto again;
  88         } else {
  89                 /* Eep. All were empty */
  90                 D1(printk(KERN_NOTICE "jffs2: No clean, dirty _or_ erasable blocks to GC from! Where are they all?\n"));
  91                 return NULL;
  92         }
  93
  94         ret = list_entry(nextlist->next, struct jffs2_eraseblock, list);
  95         list_del(&ret->list);
  96         c->gcblock = ret;
  97         ret->gc_node = ret->first_node;
  98         if (!ret->gc_node) {
  99                 printk(KERN_WARNING "Eep. ret->gc_node for block at 0x%08x is NULL\n", ret->offset);
 100                 BUG();
 101         }
 102
 103         /* Have we accidentally picked a clean block with wasted space ? */
 104         if (ret->wasted_size) {
 105                 D1(printk(KERN_DEBUG "Converting wasted_size %08x to dirty_size\n", ret->wasted_size));
 106                 ret->dirty_size += ret->wasted_size;
 107                 c->wasted_size -= ret->wasted_size;
 108                 c->dirty_size += ret->wasted_size;
 109                 ret->wasted_size = 0;
 110         }
 111
 112         return ret;
 113 }
 114
 115 /* jffs2_garbage_collect_pass
 116  * Make a single attempt to progress GC. Move one node, and possibly
 117  * start erasing one eraseblock.
 118  */
 119 int jffs2_garbage_collect_pass(struct jffs2_sb_info *c)
 120 {
 121         struct jffs2_inode_info *f;
 122         struct jffs2_inode_cache *ic;
 123         struct jffs2_eraseblock *jeb;
 124         struct jffs2_raw_node_ref *raw;
 125         int ret = 0, inum, nlink;
 126         int xattr = 0;
 127
 128         if (down_interruptible(&c->alloc_sem))
 129                 return -EINTR;
 130
 131         for (;;) {
 132                 spin_lock(&c->erase_completion_lock);
 133                 if (!c->unchecked_size)
 134                         break;
 135
 136                 /* We can't start doing GC yet. We haven't finished checking
 137                    the node CRCs etc. Do it now. */
 138
 139                 /* checked_ino is protected by the alloc_sem */
 140                 if (c->checked_ino > c->highest_ino && xattr) {
 141                         printk(KERN_CRIT "Checked all inodes but still 0x%x bytes of unchecked space?\n",
 142                                c->unchecked_size);
 143                         jffs2_dbg_dump_block_lists_nolock(c);
 144                         spin_unlock(&c->erase_completion_lock);
 145                         up(&c->alloc_sem);
 146                         return -ENOSPC;
 147                 }
 148
 149                 spin_unlock(&c->erase_completion_lock);
 150
 151                 if (!xattr)
 152                         xattr = jffs2_verify_xattr(c);
 153
 154                 spin_lock(&c->inocache_lock);
 155
 156                 ic = jffs2_get_ino_cache(c, c->checked_ino++);
 157
 158                 if (!ic) {
 159                         spin_unlock(&c->inocache_lock);
 160                         continue;
 161                 }
 162
 163                 if (!ic->nlink) {
 164                         D1(printk(KERN_DEBUG "Skipping check of ino #%d with nlink zero\n",
 165                                   ic->ino));
 166                         spin_unlock(&c->inocache_lock);
 167                         jffs2_xattr_delete_inode(c, ic);
 168                         continue;
 169                 }
 170                 switch(ic->state) {
 171                 case INO_STATE_CHECKEDABSENT:
 172                 case INO_STATE_PRESENT:
 173                         D1(printk(KERN_DEBUG "Skipping ino #%u already checked\n", ic->ino));
 174                         spin_unlock(&c->inocache_lock);
 175                         continue;
 176
 177                 case INO_STATE_GC:
 178                 case INO_STATE_CHECKING:
 179                         printk(KERN_WARNING "Inode #%u is in state %d during CRC check phase!\n", ic->ino, ic->state);
 180                         spin_unlock(&c->inocache_lock);
 181                         BUG();
 182
 183                 case INO_STATE_READING:
 184                         /* We need to wait for it to finish, lest we move on
 185                            and trigger the BUG() above while we haven't yet
 186                            finished checking all its nodes */
 187                         D1(printk(KERN_DEBUG "Waiting for ino #%u to finish reading\n", ic->ino));
 188                         /* We need to come back again for the _same_ inode. We've
 189                          made no progress in this case, but that should be OK */
 190                         c->checked_ino--;
 191
 192                         up(&c->alloc_sem);
 193                         sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
 194                         return 0;
 195
 196                 default:
 197                         BUG();
 198
 199                 case INO_STATE_UNCHECKED:
 200                         ;
 201                 }
 202                 ic->state = INO_STATE_CHECKING;
 203                 spin_unlock(&c->inocache_lock);
 204
 205                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() triggering inode scan of ino#%u\n", ic->ino));
 206
 207                 ret = jffs2_do_crccheck_inode(c, ic);
 208                 if (ret)
 209                         printk(KERN_WARNING "Returned error for crccheck of ino #%u. Expect badness...\n", ic->ino);
 210
 211                 jffs2_set_inocache_state(c, ic, INO_STATE_CHECKEDABSENT);
 212                 up(&c->alloc_sem);
 213                 return ret;
 214         }
 215
 216         /* First, work out which block we're garbage-collecting */
 217         jeb = c->gcblock;
 218
 219         if (!jeb)
 220                 jeb = jffs2_find_gc_block(c);
 221
 222         if (!jeb) {
 223                 D1 (printk(KERN_NOTICE "jffs2: Couldn't find erase block to garbage collect!\n"));
 224                 spin_unlock(&c->erase_completion_lock);
 225                 up(&c->alloc_sem);
 226                 return -EIO;
 227         }
 228
 229         D1(printk(KERN_DEBUG "GC from block %08x, used_size %08x, dirty_size %08x, free_size %08x\n", jeb->offset, jeb->used_size, jeb->dirty_size, jeb->free_size));
 230         D1(if (c->nextblock)
 231            printk(KERN_DEBUG "Nextblock at  %08x, used_size %08x, dirty_size %08x, wasted_size %08x, free_size %08x\n", c->nextblock->offset, c->nextblock->used_size, c->nextblock->dirty_size, c->nextblock->wasted_size, c->nextblock->free_size));
 232
 233         if (!jeb->used_size) {
 234                 up(&c->alloc_sem);
 235                 goto eraseit;
 236         }
 237
 238         raw = jeb->gc_node;
 239
 240         while(ref_obsolete(raw)) {
 241                 D1(printk(KERN_DEBUG "Node at 0x%08x is obsolete... skipping\n", ref_offset(raw)));
 242                 raw = ref_next(raw);
 243                 if (unlikely(!raw)) {
 244                         printk(KERN_WARNING "eep. End of raw list while still supposedly nodes to GC\n");
 245                         printk(KERN_WARNING "erase block at 0x%08x. free_size 0x%08x, dirty_size 0x%08x, used_size 0x%08x\n",
 246                                jeb->offset, jeb->free_size, jeb->dirty_size, jeb->used_size);
 247                         jeb->gc_node = raw;
 248                         spin_unlock(&c->erase_completion_lock);
 249                         up(&c->alloc_sem);
 250                         BUG();
 251                 }
 252         }
 253         jeb->gc_node = raw;
 254
 255         D1(printk(KERN_DEBUG "Going to garbage collect node at 0x%08x\n", ref_offset(raw)));
 256
 257         if (!raw->next_in_ino) {
 258                 /* Inode-less node. Clean marker, snapshot or something like that */
 259                 spin_unlock(&c->erase_completion_lock);
 260                 if (ref_flags(raw) == REF_PRISTINE) {
 261                         /* It's an unknown node with JFFS2_FEATURE_RWCOMPAT_COPY */
 262                         jffs2_garbage_collect_pristine(c, NULL, raw);
 263                 } else {
 264                         /* Just mark it obsolete */
 265                         jffs2_mark_node_obsolete(c, raw);
 266                 }
 267                 up(&c->alloc_sem);
 268                 goto eraseit_lock;
 269         }
 270
 271         ic = jffs2_raw_ref_to_ic(raw);
 272
 273 #ifdef CONFIG_JFFS2_FS_XATTR
 274         /* When 'ic' refers xattr_datum/xattr_ref, this node is GCed as xattr.
 275          * We can decide whether this node is inode or xattr by ic->class.     */
 276         if (ic->class == RAWNODE_CLASS_XATTR_DATUM
 277             || ic->class == RAWNODE_CLASS_XATTR_REF) {
 278                 spin_unlock(&c->erase_completion_lock);
 279
 280                 if (ic->class == RAWNODE_CLASS_XATTR_DATUM) {
 281                         ret = jffs2_garbage_collect_xattr_datum(c, (struct jffs2_xattr_datum *)ic, raw);
 282                 } else {
 283                         ret = jffs2_garbage_collect_xattr_ref(c, (struct jffs2_xattr_ref *)ic, raw);
 284                 }
 285                 goto release_sem;
 286         }
 287 #endif
 288
 289         /* We need to hold the inocache. Either the erase_completion_lock or
 290            the inocache_lock are sufficient; we trade down since the inocache_lock
 291            causes less contention. */
 292         spin_lock(&c->inocache_lock);
 293
 294         spin_unlock(&c->erase_completion_lock);
 295
 296         D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass collecting from block @0x%08x. Node @0x%08x(%d), ino #%u\n", jeb->offset, ref_offset(raw), ref_flags(raw), ic->ino));
 297
 298         /* Three possibilities:
 299            1. Inode is already in-core. We must iget it and do proper
 300               updating to its fragtree, etc.
 301            2. Inode is not in-core, node is REF_PRISTINE. We lock the
 302               inocache to prevent a read_inode(), copy the node intact.
 303            3. Inode is not in-core, node is not pristine. We must iget()
 304               and take the slow path.
 305         */
 306
 307         switch(ic->state) {
 308         case INO_STATE_CHECKEDABSENT:
 309                 /* It's been checked, but it's not currently in-core.
 310                    We can just copy any pristine nodes, but have
 311                    to prevent anyone else from doing read_inode() while
 312                    we're at it, so we set the state accordingly */
 313                 if (ref_flags(raw) == REF_PRISTINE)
 314                         ic->state = INO_STATE_GC;
 315                 else {
 316                         D1(printk(KERN_DEBUG "Ino #%u is absent but node not REF_PRISTINE. Reading.\n",
 317                                   ic->ino));
 318                 }
 319                 break;
 320
 321         case INO_STATE_PRESENT:
 322                 /* It's in-core. GC must iget() it. */
 323                 break;
 324
 325         case INO_STATE_UNCHECKED:
 326         case INO_STATE_CHECKING:
 327         case INO_STATE_GC:
 328                 /* Should never happen. We should have finished checking
 329                    by the time we actually start doing any GC, and since
 330                    we're holding the alloc_sem, no other garbage collection
 331                    can happen.
 332                 */
 333                 printk(KERN_CRIT "Inode #%u already in state %d in jffs2_garbage_collect_pass()!\n",
 334                        ic->ino, ic->state);
 335                 up(&c->alloc_sem);
 336                 spin_unlock(&c->inocache_lock);
 337                 BUG();
 338
 339         case INO_STATE_READING:
 340                 /* Someone's currently trying to read it. We must wait for
 341                    them to finish and then go through the full iget() route
 342                    to do the GC. However, sometimes read_inode() needs to get
 343                    the alloc_sem() (for marking nodes invalid) so we must
 344                    drop the alloc_sem before sleeping. */
 345
 346                 up(&c->alloc_sem);
 347                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() waiting for ino #%u in state %d\n",
 348                           ic->ino, ic->state));
 349                 sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
 350                 /* And because we dropped the alloc_sem we must start again from the
 351                    beginning. Ponder chance of livelock here -- we're returning success
 352                    without actually making any progress.
 353
 354                    Q: What are the chances that the inode is back in INO_STATE_READING
 355                    again by the time we next enter this function? And that this happens
 356                    enough times to cause a real delay?
 357
 358                    A: Small enough that I don't care :)
 359                 */
 360                 return 0;
 361         }
 362
 363         /* OK. Now if the inode is in state INO_STATE_GC, we are going to copy the
 364            node intact, and we don't have to muck about with the fragtree etc.
 365            because we know it's not in-core. If it _was_ in-core, we go through
 366            all the iget() crap anyway */
 367
 368         if (ic->state == INO_STATE_GC) {
 369                 spin_unlock(&c->inocache_lock);
 370
 371                 ret = jffs2_garbage_collect_pristine(c, ic, raw);
 372
 373                 spin_lock(&c->inocache_lock);
 374                 ic->state = INO_STATE_CHECKEDABSENT;
 375                 wake_up(&c->inocache_wq);
 376
 377                 if (ret != -EBADFD) {
 378                         spin_unlock(&c->inocache_lock);
 379                         goto release_sem;
 380                 }
 381
 382                 /* Fall through if it wanted us to, with inocache_lock held */
 383         }
 384
 385         /* Prevent the fairly unlikely race where the gcblock is
 386            entirely obsoleted by the final close of a file which had
 387            the only valid nodes in the block, followed by erasure,
 388            followed by freeing of the ic because the erased block(s)
 389            held _all_ the nodes of that inode.... never been seen but
 390            it's vaguely possible. */
 391
 392         inum = ic->ino;
 393         nlink = ic->nlink;
 394         spin_unlock(&c->inocache_lock);
 395
 396         f = jffs2_gc_fetch_inode(c, inum, nlink);
 397         if (IS_ERR(f)) {
 398                 ret = PTR_ERR(f);
 399                 goto release_sem;
 400         }
 401         if (!f) {
 402                 ret = 0;
 403                 goto release_sem;
 404         }
 405
 406         ret = jffs2_garbage_collect_live(c, jeb, raw, f);
 407
 408         jffs2_gc_release_inode(c, f);
 409
 410  release_sem:
 411         up(&c->alloc_sem);
 412
 413  eraseit_lock:
 414         /* If we've finished this block, start it erasing */
 415         spin_lock(&c->erase_completion_lock);
 416
 417  eraseit:
 418         if (c->gcblock && !c->gcblock->used_size) {
 419                 D1(printk(KERN_DEBUG "Block at 0x%08x completely obsoleted by GC. Moving to erase_pending_list\n", c->gcblock->offset));
 420                 /* We're GC'ing an empty block? */
 421                 list_add_tail(&c->gcblock->list, &c->erase_pending_list);
 422                 c->gcblock = NULL;
 423                 c->nr_erasing_blocks++;
 424                 jffs2_erase_pending_trigger(c);
 425         }
 426         spin_unlock(&c->erase_completion_lock);
 427
 428         return ret;
 429 }
 430
 431 static int jffs2_garbage_collect_live(struct jffs2_sb_info *c,  struct jffs2_eraseblock *jeb,
 432                                       struct jffs2_raw_node_ref *raw, struct jffs2_inode_info *f)
 433 {
 434         struct jffs2_node_frag *frag;
 435         struct jffs2_full_dnode *fn = NULL;
 436         struct jffs2_full_dirent *fd;
 437         uint32_t start = 0, end = 0, nrfrags = 0;
 438         int ret = 0;
 439
 440         down(&f->sem);
 441
 442         /* Now we have the lock for this inode. Check that it's still the one at the head
 443            of the list. */
 444
 445         spin_lock(&c->erase_completion_lock);
 446
 447         if (c->gcblock != jeb) {
 448                 spin_unlock(&c->erase_completion_lock);
 449                 D1(printk(KERN_DEBUG "GC block is no longer gcblock. Restart\n"));
 450                 goto upnout;
 451         }
 452         if (ref_obsolete(raw)) {
 453                 spin_unlock(&c->erase_completion_lock);
 454                 D1(printk(KERN_DEBUG "node to be GC'd was obsoleted in the meantime.\n"));
 455                 /* They'll call again */
 456                 goto upnout;
 457         }
 458         spin_unlock(&c->erase_completion_lock);
 459
 460         /* OK. Looks safe. And nobody can get us now because we have the semaphore. Move the block */
 461         if (f->metadata && f->metadata->raw == raw) {
 462                 fn = f->metadata;
 463                 ret = jffs2_garbage_collect_metadata(c, jeb, f, fn);
 464                 goto upnout;
 465         }
 466
 467         /* FIXME. Read node and do lookup? */
 468         for (frag = frag_first(&f->fragtree); frag; frag = frag_next(frag)) {
 469                 if (frag->node && frag->node->raw == raw) {
 470                         fn = frag->node;
 471                         end = frag->ofs + frag->size;
 472                         if (!nrfrags++)
 473                                 start = frag->ofs;
 474                         if (nrfrags == frag->node->frags)
 475                                 break; /* We've found them all */
 476                 }
 477         }
 478         if (fn) {
 479                 if (ref_flags(raw) == REF_PRISTINE) {
 480                         ret = jffs2_garbage_collect_pristine(c, f->inocache, raw);
 481                         if (!ret) {
 482                                 /* Urgh. Return it sensibly. */
 483                                 frag->node->raw = f->inocache->nodes;
 484                         }
 485                         if (ret != -EBADFD)
 486                                 goto upnout;
 487                 }
 488                 /* We found a datanode. Do the GC */
 489                 if((start >> PAGE_CACHE_SHIFT) < ((end-1) >> PAGE_CACHE_SHIFT)) {
 490                         /* It crosses a page boundary. Therefore, it must be a hole. */
 491                         ret = jffs2_garbage_collect_hole(c, jeb, f, fn, start, end);
 492                 } else {
 493                         /* It could still be a hole. But we GC the page this way anyway */
 494                         ret = jffs2_garbage_collect_dnode(c, jeb, f, fn, start, end);
 495                 }
 496                 goto upnout;
 497         }
 498
 499         /* Wasn't a dnode. Try dirent */
 500         for (fd = f->dents; fd; fd=fd->next) {
 501                 if (fd->raw == raw)
 502                         break;
 503         }
 504
 505         if (fd && fd->ino) {
 506                 ret = jffs2_garbage_collect_dirent(c, jeb, f, fd);
 507         } else if (fd) {
 508                 ret = jffs2_garbage_collect_deletion_dirent(c, jeb, f, fd);
 509         } else {
 510                 printk(KERN_WARNING "Raw node at 0x%08x wasn't in node lists for ino #%u\n",
 511                        ref_offset(raw), f->inocache->ino);
 512                 if (ref_obsolete(raw)) {
 513                         printk(KERN_WARNING "But it's obsolete so we don't mind too much\n");
 514                 } else {
 515                         jffs2_dbg_dump_node(c, ref_offset(raw));
 516                         BUG();
 517                 }
 518         }
 519  upnout:
 520         up(&f->sem);
 521
 522         return ret;
 523 }
 524
 525 static int jffs2_garbage_collect_pristine(struct jffs2_sb_info *c,
 526                                           struct jffs2_inode_cache *ic,
 527                                           struct jffs2_raw_node_ref *raw)
 528 {
 529         union jffs2_node_union *node;
 530         size_t retlen;
 531         int ret;
 532         uint32_t phys_ofs, alloclen;
 533         uint32_t crc, rawlen;
 534         int retried = 0;
 535
 536         D1(printk(KERN_DEBUG "Going to GC REF_PRISTINE node at 0x%08x\n", ref_offset(raw)));
 537
 538         alloclen = rawlen = ref_totlen(c, c->gcblock, raw);
 539
 540         /* Ask for a small amount of space (or the totlen if smaller) because we
 541            don't want to force wastage of the end of a block if splitting would
 542            work. */
 543         if (ic && alloclen > sizeof(struct jffs2_raw_inode) + JFFS2_MIN_DATA_LEN)
 544                 alloclen = sizeof(struct jffs2_raw_inode) + JFFS2_MIN_DATA_LEN;
 545
 546         ret = jffs2_reserve_space_gc(c, alloclen, &alloclen, rawlen);
 547         /* 'rawlen' is not the exact summary size; it is only an upper estimation */
 548
 549         if (ret)
 550                 return ret;
 551
 552         if (alloclen < rawlen) {
 553                 /* Doesn't fit untouched. We'll go the old route and split it */
 554                 return -EBADFD;
 555         }
 556
 557         node = kmalloc(rawlen, GFP_KERNEL);
 558         if (!node)
 559                 return -ENOMEM;
 560
 561         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)node);
 562         if (!ret && retlen != rawlen)
 563                 ret = -EIO;
 564         if (ret)
 565                 goto out_node;
 566
 567         crc = crc32(0, node, sizeof(struct jffs2_unknown_node)-4);
 568         if (je32_to_cpu(node->u.hdr_crc) != crc) {
 569                 printk(KERN_WARNING "Header CRC failed on REF_PRISTINE node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 570                        ref_offset(raw), je32_to_cpu(node->u.hdr_crc), crc);
 571                 goto bail;
 572         }
 573
 574         switch(je16_to_cpu(node->u.nodetype)) {
 575         case JFFS2_NODETYPE_INODE:
 576                 crc = crc32(0, node, sizeof(node->i)-8);
 577                 if (je32_to_cpu(node->i.node_crc) != crc) {
 578                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 579                                ref_offset(raw), je32_to_cpu(node->i.node_crc), crc);
 580                         goto bail;
 581                 }
 582
 583                 if (je32_to_cpu(node->i.dsize)) {
 584                         crc = crc32(0, node->i.data, je32_to_cpu(node->i.csize));
 585                         if (je32_to_cpu(node->i.data_crc) != crc) {
 586                                 printk(KERN_WARNING "Data CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 587                                        ref_offset(raw), je32_to_cpu(node->i.data_crc), crc);
 588                                 goto bail;
 589                         }
 590                 }
 591                 break;
 592
 593         case JFFS2_NODETYPE_DIRENT:
 594                 crc = crc32(0, node, sizeof(node->d)-8);
 595                 if (je32_to_cpu(node->d.node_crc) != crc) {
 596                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE dirent node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 597                                ref_offset(raw), je32_to_cpu(node->d.node_crc), crc);
 598                         goto bail;
 599                 }
 600
 601                 if (node->d.nsize) {
 602                         crc = crc32(0, node->d.name, node->d.nsize);
 603                         if (je32_to_cpu(node->d.name_crc) != crc) {
 604                                 printk(KERN_WARNING "Name CRC failed on REF_PRISTINE dirent ode at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 605                                        ref_offset(raw), je32_to_cpu(node->d.name_crc), crc);
 606                                 goto bail;
 607                         }
 608                 }
 609                 break;
 610         default:
 611                 /* If it's inode-less, we don't _know_ what it is. Just copy it intact */
 612                 if (ic) {
 613                         printk(KERN_WARNING "Unknown node type for REF_PRISTINE node at 0x%08x: 0x%04x\n",
 614                                ref_offset(raw), je16_to_cpu(node->u.nodetype));
 615                         goto bail;
 616                 }
 617         }
 618
 619         /* OK, all the CRCs are good; this node can just be copied as-is. */
 620  retry:
 621         phys_ofs = write_ofs(c);
 622
 623         ret = jffs2_flash_write(c, phys_ofs, rawlen, &retlen, (char *)node);
 624
 625         if (ret || (retlen != rawlen)) {
 626                 printk(KERN_NOTICE "Write of %d bytes at 0x%08x failed. returned %d, retlen %zd\n",
 627                        rawlen, phys_ofs, ret, retlen);
 628                 if (retlen) {
 629                         jffs2_add_physical_node_ref(c, phys_ofs | REF_OBSOLETE, rawlen, NULL);
 630                 } else {
 631                         printk(KERN_NOTICE "Not marking the space at 0x%08x as dirty because the flash driver returned retlen zero\n", phys_ofs);
 632                 }
 633                 if (!retried) {
 634                         /* Try to reallocate space and retry */
 635                         uint32_t dummy;
 636                         struct jffs2_eraseblock *jeb = &c->blocks[phys_ofs / c->sector_size];
 637
 638                         retried = 1;
 639
 640                         D1(printk(KERN_DEBUG "Retrying failed write of REF_PRISTINE node.\n"));
 641
 642                         jffs2_dbg_acct_sanity_check(c,jeb);
 643                         jffs2_dbg_acct_paranoia_check(c, jeb);
 644
 645                         ret = jffs2_reserve_space_gc(c, rawlen, &dummy, rawlen);
 646                                                 /* this is not the exact summary size of it,
 647                                                         it is only an upper estimation */
 648
 649                         if (!ret) {
 650                                 D1(printk(KERN_DEBUG "Allocated space at 0x%08x to retry failed write.\n", phys_ofs));
 651
 652                                 jffs2_dbg_acct_sanity_check(c,jeb);
 653                                 jffs2_dbg_acct_paranoia_check(c, jeb);
 654
 655                                 goto retry;
 656                         }
 657                         D1(printk(KERN_DEBUG "Failed to allocate space to retry failed write: %d!\n", ret));
 658                 }
 659
 660                 if (!ret)
 661                         ret = -EIO;
 662                 goto out_node;
 663         }
 664         jffs2_add_physical_node_ref(c, phys_ofs | REF_PRISTINE, rawlen, ic);
 665
 666         jffs2_mark_node_obsolete(c, raw);
 667         D1(printk(KERN_DEBUG "WHEEE! GC REF_PRISTINE node at 0x%08x succeeded\n", ref_offset(raw)));
 668
 669  out_node:
 670         kfree(node);
 671         return ret;
 672  bail:
 673         ret = -EBADFD;
 674         goto out_node;
 675 }
 676
 677 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 678                                         struct jffs2_inode_info *f, struct jffs2_full_dnode *fn)
 679 {
 680         struct jffs2_full_dnode *new_fn;
 681         struct jffs2_raw_inode ri;
 682         struct jffs2_node_frag *last_frag;
 683         union jffs2_device_node dev;
 684         char *mdata = NULL, mdatalen = 0;
 685         uint32_t alloclen, ilen;
 686         int ret;
 687
 688         if (S_ISBLK(JFFS2_F_I_MODE(f)) ||
 689             S_ISCHR(JFFS2_F_I_MODE(f)) ) {
 690                 /* For these, we don't actually need to read the old node */
 691                 mdatalen = jffs2_encode_dev(&dev, JFFS2_F_I_RDEV(f));
 692                 mdata = (char *)&dev;
 693                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bytes of kdev_t\n", mdatalen));
 694         } else if (S_ISLNK(JFFS2_F_I_MODE(f))) {
 695                 mdatalen = fn->size;
 696                 mdata = kmalloc(fn->size, GFP_KERNEL);
 697                 if (!mdata) {
 698                         printk(KERN_WARNING "kmalloc of mdata failed in jffs2_garbage_collect_metadata()\n");
 699                         return -ENOMEM;
 700                 }
 701                 ret = jffs2_read_dnode(c, f, fn, mdata, 0, mdatalen);
 702                 if (ret) {
 703                         printk(KERN_WARNING "read of old metadata failed in jffs2_garbage_collect_metadata(): %d\n", ret);
 704                         kfree(mdata);
 705                         return ret;
 706                 }
 707                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bites of symlink target\n", mdatalen));
 708
 709         }
 710
 711         ret = jffs2_reserve_space_gc(c, sizeof(ri) + mdatalen, &alloclen,
 712                                 JFFS2_SUMMARY_INODE_SIZE);
 713         if (ret) {
 714                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_metadata failed: %d\n",
 715                        sizeof(ri)+ mdatalen, ret);
 716                 goto out;
 717         }
 718
 719         last_frag = frag_last(&f->fragtree);
 720         if (last_frag)
 721                 /* Fetch the inode length from the fragtree rather then
 722                  * from i_size since i_size may have not been updated yet */
 723                 ilen = last_frag->ofs + last_frag->size;
 724         else
 725                 ilen = JFFS2_F_I_SIZE(f);
 726
 727         memset(&ri, 0, sizeof(ri));
 728         ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
 729         ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
 730         ri.totlen = cpu_to_je32(sizeof(ri) + mdatalen);
 731         ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
 732
 733         ri.ino = cpu_to_je32(f->inocache->ino);
 734         ri.version = cpu_to_je32(++f->highest_version);
 735         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
 736         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
 737         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
 738         ri.isize = cpu_to_je32(ilen);
 739         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
 740         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
 741         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
 742         ri.offset = cpu_to_je32(0);
 743         ri.csize = cpu_to_je32(mdatalen);
 744         ri.dsize = cpu_to_je32(mdatalen);
 745         ri.compr = JFFS2_COMPR_NONE;
 746         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
 747         ri.data_crc = cpu_to_je32(crc32(0, mdata, mdatalen));
 748
 749         new_fn = jffs2_write_dnode(c, f, &ri, mdata, mdatalen, ALLOC_GC);
 750
 751         if (IS_ERR(new_fn)) {
 752                 printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
 753                 ret = PTR_ERR(new_fn);
 754                 goto out;
 755         }
 756         jffs2_mark_node_obsolete(c, fn->raw);
 757         jffs2_free_full_dnode(fn);
 758         f->metadata = new_fn;
 759  out:
 760         if (S_ISLNK(JFFS2_F_I_MODE(f)))
 761                 kfree(mdata);
 762         return ret;
 763 }
 764
 765 static int jffs2_garbage_collect_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 766                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
 767 {
 768         struct jffs2_full_dirent *new_fd;
 769         struct jffs2_raw_dirent rd;
 770         uint32_t alloclen;
 771         int ret;
 772
 773         rd.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
 774         rd.nodetype = cpu_to_je16(JFFS2_NODETYPE_DIRENT);
 775         rd.nsize = strlen(fd->name);
 776         rd.totlen = cpu_to_je32(sizeof(rd) + rd.nsize);
 777         rd.hdr_crc = cpu_to_je32(crc32(0, &rd, sizeof(struct jffs2_unknown_node)-4));
 778
 779         rd.pino = cpu_to_je32(f->inocache->ino);
 780         rd.version = cpu_to_je32(++f->highest_version);
 781         rd.ino = cpu_to_je32(fd->ino);
 782         /* If the times on this inode were set by explicit utime() they can be different,
 783            so refrain from splatting them. */
 784         if (JFFS2_F_I_MTIME(f) == JFFS2_F_I_CTIME(f))
 785                 rd.mctime = cpu_to_je32(JFFS2_F_I_MTIME(f));
 786         else
 787                 rd.mctime = cpu_to_je32(0);
 788         rd.type = fd->type;
 789         rd.node_crc = cpu_to_je32(crc32(0, &rd, sizeof(rd)-8));
 790         rd.name_crc = cpu_to_je32(crc32(0, fd->name, rd.nsize));
 791
 792         ret = jffs2_reserve_space_gc(c, sizeof(rd)+rd.nsize, &alloclen,
 793                                 JFFS2_SUMMARY_DIRENT_SIZE(rd.nsize));
 794         if (ret) {
 795                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dirent failed: %d\n",
 796                        sizeof(rd)+rd.nsize, ret);
 797                 return ret;
 798         }
 799         new_fd = jffs2_write_dirent(c, f, &rd, fd->name, rd.nsize, ALLOC_GC);
 800
 801         if (IS_ERR(new_fd)) {
 802                 printk(KERN_WARNING "jffs2_write_dirent in garbage_collect_dirent failed: %ld\n", PTR_ERR(new_fd));
 803                 return PTR_ERR(new_fd);
 804         }
 805         jffs2_add_fd_to_list(c, new_fd, &f->dents);
 806         return 0;
 807 }
 808
 809 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 810                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
 811 {
 812         struct jffs2_full_dirent **fdp = &f->dents;
 813         int found = 0;
 814
 815         /* On a medium where we can't actually mark nodes obsolete
 816            pernamently, such as NAND flash, we need to work out
 817            whether this deletion dirent is still needed to actively
 818            delete a 'real' dirent with the same name that's still
 819            somewhere else on the flash. */
 820         if (!jffs2_can_mark_obsolete(c)) {
 821                 struct jffs2_raw_dirent *rd;
 822                 struct jffs2_raw_node_ref *raw;
 823                 int ret;
 824                 size_t retlen;
 825                 int name_len = strlen(fd->name);
 826                 uint32_t name_crc = crc32(0, fd->name, name_len);
 827                 uint32_t rawlen = ref_totlen(c, jeb, fd->raw);
 828
 829                 rd = kmalloc(rawlen, GFP_KERNEL);
 830                 if (!rd)
 831                         return -ENOMEM;
 832
 833                 /* Prevent the erase code from nicking the obsolete node refs while
 834                    we're looking at them. I really don't like this extra lock but
 835                    can't see any alternative. Suggestions on a postcard to... */
 836                 down(&c->erase_free_sem);
 837
 838                 for (raw = f->inocache->nodes; raw != (void *)f->inocache; raw = raw->next_in_ino) {
 839
 840                         cond_resched();
 841
 842                         /* We only care about obsolete ones */
 843                         if (!(ref_obsolete(raw)))
 844                                 continue;
 845
 846                         /* Any dirent with the same name is going to have the same length... */
 847                         if (ref_totlen(c, NULL, raw) != rawlen)
 848                                 continue;
 849
 850                         /* Doesn't matter if there's one in the same erase block. We're going to
 851                            delete it too at the same time. */
 852                         if (SECTOR_ADDR(raw->flash_offset) == SECTOR_ADDR(fd->raw->flash_offset))
 853                                 continue;
 854
 855                         D1(printk(KERN_DEBUG "Check potential deletion dirent at %08x\n", ref_offset(raw)));
 856
 857                         /* This is an obsolete node belonging to the same directory, and it's of the right
 858                            length. We need to take a closer look...*/
 859                         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)rd);
 860                         if (ret) {
 861                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Read error (%d) reading obsolete node at %08x\n", ret, ref_offset(raw));
 862                                 /* If we can't read it, we don't need to continue to obsolete it. Continue */
 863                                 continue;
 864                         }
 865                         if (retlen != rawlen) {
 866                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Short read (%zd not %u) reading header from obsolete node at %08x\n",
 867                                        retlen, rawlen, ref_offset(raw));
 868                                 continue;
 869                         }
 870
 871                         if (je16_to_cpu(rd->nodetype) != JFFS2_NODETYPE_DIRENT)
 872                                 continue;
 873
 874                         /* If the name CRC doesn't match, skip */
 875                         if (je32_to_cpu(rd->name_crc) != name_crc)
 876                                 continue;
 877
 878                         /* If the name length doesn't match, or it's another deletion dirent, skip */
 879                         if (rd->nsize != name_len || !je32_to_cpu(rd->ino))
 880                                 continue;
 881
 882                         /* OK, check the actual name now */
 883                         if (memcmp(rd->name, fd->name, name_len))
 884                                 continue;
 885
 886                         /* OK. The name really does match. There really is still an older node on
 887                            the flash which our deletion dirent obsoletes. So we have to write out
 888                            a new deletion dirent to replace it */
 889                         up(&c->erase_free_sem);
 890
 891                         D1(printk(KERN_DEBUG "Deletion dirent at %08x still obsoletes real dirent \"%s\" at %08x for ino #%u\n",
 892                                   ref_offset(fd->raw), fd->name, ref_offset(raw), je32_to_cpu(rd->ino)));
 893                         kfree(rd);
 894
 895                         return jffs2_garbage_collect_dirent(c, jeb, f, fd);
 896                 }
 897
 898                 up(&c->erase_free_sem);
 899                 kfree(rd);
 900         }
 901
 902         /* FIXME: If we're deleting a dirent which contains the current mtime and ctime,
 903            we should update the metadata node with those times accordingly */
 904
 905         /* No need for it any more. Just mark it obsolete and remove it from the list */
 906         while (*fdp) {
 907                 if ((*fdp) == fd) {
 908                         found = 1;
 909                         *fdp = fd->next;
 910                         break;
 911                 }
 912                 fdp = &(*fdp)->next;
 913         }
 914         if (!found) {
 915                 printk(KERN_WARNING "Deletion dirent \"%s\" not found in list for ino #%u\n", fd->name, f->inocache->ino);
 916         }
 917         jffs2_mark_node_obsolete(c, fd->raw);
 918         jffs2_free_full_dirent(fd);
 919         return 0;
 920 }
 921
 922 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 923                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
 924                                       uint32_t start, uint32_t end)
 925 {
 926         struct jffs2_raw_inode ri;
 927         struct jffs2_node_frag *frag;
 928         struct jffs2_full_dnode *new_fn;
 929         uint32_t alloclen, ilen;
 930         int ret;
 931
 932         D1(printk(KERN_DEBUG "Writing replacement hole node for ino #%u from offset 0x%x to 0x%x\n",
 933                   f->inocache->ino, start, end));
 934
 935         memset(&ri, 0, sizeof(ri));
 936
 937         if(fn->frags > 1) {
 938                 size_t readlen;
 939                 uint32_t crc;
 940                 /* It's partially obsoleted by a later write. So we have to
 941                    write it out again with the _same_ version as before */
 942                 ret = jffs2_flash_read(c, ref_offset(fn->raw), sizeof(ri), &readlen, (char *)&ri);
 943                 if (readlen != sizeof(ri) || ret) {
 944                         printk(KERN_WARNING "Node read failed in jffs2_garbage_collect_hole. Ret %d, retlen %zd. Data will be lost by writing new hole node\n", ret, readlen);
 945                         goto fill;
 946                 }
 947                 if (je16_to_cpu(ri.nodetype) != JFFS2_NODETYPE_INODE) {
 948                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had node type 0x%04x instead of JFFS2_NODETYPE_INODE(0x%04x)\n",
 949                                ref_offset(fn->raw),
 950                                je16_to_cpu(ri.nodetype), JFFS2_NODETYPE_INODE);
 951                         return -EIO;
 952                 }
 953                 if (je32_to_cpu(ri.totlen) != sizeof(ri)) {
 954                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had totlen 0x%x instead of expected 0x%zx\n",
 955                                ref_offset(fn->raw),
 956                                je32_to_cpu(ri.totlen), sizeof(ri));
 957                         return -EIO;
 958                 }
 959                 crc = crc32(0, &ri, sizeof(ri)-8);
 960                 if (crc != je32_to_cpu(ri.node_crc)) {
 961                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had CRC 0x%08x which doesn't match calculated CRC 0x%08x\n",
 962                                ref_offset(fn->raw),
 963                                je32_to_cpu(ri.node_crc), crc);
 964                         /* FIXME: We could possibly deal with this by writing new holes for each frag */
 965                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
 966                                start, end, f->inocache->ino);
 967                         goto fill;
 968                 }
 969                 if (ri.compr != JFFS2_COMPR_ZERO) {
 970                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node 0x%08x wasn't a hole node!\n", ref_offset(fn->raw));
 971                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
 972                                start, end, f->inocache->ino);
 973                         goto fill;
 974                 }
 975         } else {
 976         fill:
 977                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
 978                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
 979                 ri.totlen = cpu_to_je32(sizeof(ri));
 980                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
 981
 982                 ri.ino = cpu_to_je32(f->inocache->ino);
 983                 ri.version = cpu_to_je32(++f->highest_version);
 984                 ri.offset = cpu_to_je32(start);
 985                 ri.dsize = cpu_to_je32(end - start);
 986                 ri.csize = cpu_to_je32(0);
 987                 ri.compr = JFFS2_COMPR_ZERO;
 988         }
 989
 990         frag = frag_last(&f->fragtree);
 991         if (frag)
 992                 /* Fetch the inode length from the fragtree rather then
 993                  * from i_size since i_size may have not been updated yet */
 994                 ilen = frag->ofs + frag->size;
 995         else
 996                 ilen = JFFS2_F_I_SIZE(f);
 997
 998         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
 999         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1000         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1001         ri.isize = cpu_to_je32(ilen);
1002         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1003         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1004         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1005         ri.data_crc = cpu_to_je32(0);
1006         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1007
1008         ret = jffs2_reserve_space_gc(c, sizeof(ri), &alloclen,
1009                                      JFFS2_SUMMARY_INODE_SIZE);
1010         if (ret) {
1011                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_hole failed: %d\n",
1012                        sizeof(ri), ret);
1013                 return ret;
1014         }
1015         new_fn = jffs2_write_dnode(c, f, &ri, NULL, 0, ALLOC_GC);
1016
1017         if (IS_ERR(new_fn)) {
1018                 printk(KERN_WARNING "Error writing new hole node: %ld\n", PTR_ERR(new_fn));
1019                 return PTR_ERR(new_fn);
1020         }
1021         if (je32_to_cpu(ri.version) == f->highest_version) {
1022                 jffs2_add_full_dnode_to_inode(c, f, new_fn);
1023                 if (f->metadata) {
1024                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1025                         jffs2_free_full_dnode(f->metadata);
1026                         f->metadata = NULL;
1027                 }
1028                 return 0;
1029         }
1030
1031         /*
1032          * We should only get here in the case where the node we are
1033          * replacing had more than one frag, so we kept the same version
1034          * number as before. (Except in case of error -- see 'goto fill;'
1035          * above.)
1036          */
1037         D1(if(unlikely(fn->frags <= 1)) {
1038                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Replacing fn with %d frag(s) but new ver %d != highest_version %d of ino #%d\n",
1039                        fn->frags, je32_to_cpu(ri.version), f->highest_version,
1040                        je32_to_cpu(ri.ino));
1041         });
1042
1043         /* This is a partially-overlapped hole node. Mark it REF_NORMAL not REF_PRISTINE */
1044         mark_ref_normal(new_fn->raw);
1045
1046         for (frag = jffs2_lookup_node_frag(&f->fragtree, fn->ofs);
1047              frag; frag = frag_next(frag)) {
1048                 if (frag->ofs > fn->size + fn->ofs)
1049                         break;
1050                 if (frag->node == fn) {
1051                         frag->node = new_fn;
1052                         new_fn->frags++;
1053                         fn->frags--;
1054                 }
1055         }
1056         if (fn->frags) {
1057                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Old node still has frags!\n");
1058                 BUG();
1059         }
1060         if (!new_fn->frags) {
1061                 printk(KERN_WARNING "jffs2_garbage_collect_hole: New node has no frags!\n");
1062                 BUG();
1063         }
1064
1065         jffs2_mark_node_obsolete(c, fn->raw);
1066         jffs2_free_full_dnode(fn);
1067
1068         return 0;
1069 }
1070
1071 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
1072                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
1073                                        uint32_t start, uint32_t end)
1074 {
1075         struct jffs2_full_dnode *new_fn;
1076         struct jffs2_raw_inode ri;
1077         uint32_t alloclen, offset, orig_end, orig_start;
1078         int ret = 0;
1079         unsigned char *comprbuf = NULL, *writebuf;
1080         unsigned long pg;
1081         unsigned char *pg_ptr;
1082
1083         memset(&ri, 0, sizeof(ri));
1084
1085         D1(printk(KERN_DEBUG "Writing replacement dnode for ino #%u from offset 0x%x to 0x%x\n",
1086                   f->inocache->ino, start, end));
1087
1088         orig_end = end;
1089         orig_start = start;
1090
1091         if (c->nr_free_blocks + c->nr_erasing_blocks > c->resv_blocks_gcmerge) {
1092                 /* Attempt to do some merging. But only expand to cover logically
1093                    adjacent frags if the block containing them is already considered
1094                    to be dirty. Otherwise we end up with GC just going round in
1095                    circles dirtying the nodes it already wrote out, especially
1096                    on NAND where we have small eraseblocks and hence a much higher
1097                    chance of nodes having to be split to cross boundaries. */
1098
1099                 struct jffs2_node_frag *frag;
1100                 uint32_t min, max;
1101
1102                 min = start & ~(PAGE_CACHE_SIZE-1);
1103                 max = min + PAGE_CACHE_SIZE;
1104
1105                 frag = jffs2_lookup_node_frag(&f->fragtree, start);
1106
1107                 /* BUG_ON(!frag) but that'll happen anyway... */
1108
1109                 BUG_ON(frag->ofs != start);
1110
1111                 /* First grow down... */
1112                 while((frag = frag_prev(frag)) && frag->ofs >= min) {
1113
1114                         /* If the previous frag doesn't even reach the beginning, there's
1115                            excessive fragmentation. Just merge. */
1116                         if (frag->ofs > min) {
1117                                 D1(printk(KERN_DEBUG "Expanding down to cover partial frag (0x%x-0x%x)\n",
1118                                           frag->ofs, frag->ofs+frag->size));
1119                                 start = frag->ofs;
1120                                 continue;
1121                         }
1122                         /* OK. This frag holds the first byte of the page. */
1123                         if (!frag->node || !frag->node->raw) {
1124                                 D1(printk(KERN_DEBUG "First frag in page is hole (0x%x-0x%x). Not expanding down.\n",
1125                                           frag->ofs, frag->ofs+frag->size));
1126                                 break;
1127                         } else {
1128
1129                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1130                                    in a block which is still considered clean? If so, don't obsolete it.
1131                                    If not, cover it anyway. */
1132
1133                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1134                                 struct jffs2_eraseblock *jeb;
1135
1136                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1137
1138                                 if (jeb == c->gcblock) {
1139                                         D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1140                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1141                                         start = frag->ofs;
1142                                         break;
1143                                 }
1144                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1145                                         D1(printk(KERN_DEBUG "Not expanding down to cover frag (0x%x-0x%x) in clean block %08x\n",
1146                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1147                                         break;
1148                                 }
1149
1150                                 D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in dirty block %08x\n",
1151                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1152                                 start = frag->ofs;
1153                                 break;
1154                         }
1155                 }
1156
1157                 /* ... then up */
1158
1159                 /* Find last frag which is actually part of the node we're to GC. */
1160                 frag = jffs2_lookup_node_frag(&f->fragtree, end-1);
1161
1162                 while((frag = frag_next(frag)) && frag->ofs+frag->size <= max) {
1163
1164                         /* If the previous frag doesn't even reach the beginning, there's lots
1165                            of fragmentation. Just merge. */
1166                         if (frag->ofs+frag->size < max) {
1167                                 D1(printk(KERN_DEBUG "Expanding up to cover partial frag (0x%x-0x%x)\n",
1168                                           frag->ofs, frag->ofs+frag->size));
1169                                 end = frag->ofs + frag->size;
1170                                 continue;
1171                         }
1172
1173                         if (!frag->node || !frag->node->raw) {
1174                                 D1(printk(KERN_DEBUG "Last frag in page is hole (0x%x-0x%x). Not expanding up.\n",
1175                                           frag->ofs, frag->ofs+frag->size));
1176                                 break;
1177                         } else {
1178
1179                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1180                                    in a block which is still considered clean? If so, don't obsolete it.
1181                                    If not, cover it anyway. */
1182
1183                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1184                                 struct jffs2_eraseblock *jeb;
1185
1186                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1187
1188                                 if (jeb == c->gcblock) {
1189                                         D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1190                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1191                                         end = frag->ofs + frag->size;
1192                                         break;
1193                                 }
1194                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1195                                         D1(printk(KERN_DEBUG "Not expanding up to cover frag (0x%x-0x%x) in clean block %08x\n",
1196                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1197                                         break;
1198                                 }
1199
1200                                 D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in dirty block %08x\n",
1201                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1202                                 end = frag->ofs + frag->size;
1203                                 break;
1204                         }
1205                 }
1206                 D1(printk(KERN_DEBUG "Expanded dnode to write from (0x%x-0x%x) to (0x%x-0x%x)\n",
1207                           orig_start, orig_end, start, end));
1208
1209                 D1(BUG_ON(end > frag_last(&f->fragtree)->ofs + frag_last(&f->fragtree)->size));
1210                 BUG_ON(end < orig_end);
1211                 BUG_ON(start > orig_start);
1212         }
1213
1214         /* First, use readpage() to read the appropriate page into the page cache */
1215         /* Q: What happens if we actually try to GC the _same_ page for which commit_write()
1216          *    triggered garbage collection in the first place?
1217          * A: I _think_ it's OK. read_cache_page shouldn't deadlock, we'll write out the
1218          *    page OK. We'll actually write it out again in commit_write, which is a little
1219          *    suboptimal, but at least we're correct.
1220          */
1221         pg_ptr = jffs2_gc_fetch_page(c, f, start, &pg);
1222
1223         if (IS_ERR(pg_ptr)) {
1224                 printk(KERN_WARNING "read_cache_page() returned error: %ld\n", PTR_ERR(pg_ptr));
1225                 return PTR_ERR(pg_ptr);
1226         }
1227
1228         offset = start;
1229         while(offset < orig_end) {
1230                 uint32_t datalen;
1231                 uint32_t cdatalen;
1232                 uint16_t comprtype = JFFS2_COMPR_NONE;
1233
1234                 ret = jffs2_reserve_space_gc(c, sizeof(ri) + JFFS2_MIN_DATA_LEN,
1235                                         &alloclen, JFFS2_SUMMARY_INODE_SIZE);
1236
1237                 if (ret) {
1238                         printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dnode failed: %d\n",
1239                                sizeof(ri)+ JFFS2_MIN_DATA_LEN, ret);
1240                         break;
1241                 }
1242                 cdatalen = min_t(uint32_t, alloclen - sizeof(ri), end - offset);
1243                 datalen = end - offset;
1244
1245                 writebuf = pg_ptr + (offset & (PAGE_CACHE_SIZE -1));
1246
1247                 comprtype = jffs2_compress(c, f, writebuf, &comprbuf, &datalen, &cdatalen);
1248
1249                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
1250                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
1251                 ri.totlen = cpu_to_je32(sizeof(ri) + cdatalen);
1252                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
1253
1254                 ri.ino = cpu_to_je32(f->inocache->ino);
1255                 ri.version = cpu_to_je32(++f->highest_version);
1256                 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1257                 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1258                 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1259                 ri.isize = cpu_to_je32(JFFS2_F_I_SIZE(f));
1260                 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1261                 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1262                 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1263                 ri.offset = cpu_to_je32(offset);
1264                 ri.csize = cpu_to_je32(cdatalen);
1265                 ri.dsize = cpu_to_je32(datalen);
1266                 ri.compr = comprtype & 0xff;
1267                 ri.usercompr = (comprtype >> 8) & 0xff;
1268                 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1269                 ri.data_crc = cpu_to_je32(crc32(0, comprbuf, cdatalen));
1270
1271                 new_fn = jffs2_write_dnode(c, f, &ri, comprbuf, cdatalen, ALLOC_GC);
1272
1273                 jffs2_free_comprbuf(comprbuf, writebuf);
1274
1275                 if (IS_ERR(new_fn)) {
1276                         printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
1277                         ret = PTR_ERR(new_fn);
1278                         break;
1279                 }
1280                 ret = jffs2_add_full_dnode_to_inode(c, f, new_fn);
1281                 offset += datalen;
1282                 if (f->metadata) {
1283                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1284                         jffs2_free_full_dnode(f->metadata);
1285                         f->metadata = NULL;
1286                 }
1287         }
1288
1289         jffs2_gc_release_page(c, pg_ptr, &pg);
1290         return ret;
1291 }