SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/jffs2/gc.c

   1 /*
   2  * JFFS2 -- Journalling Flash File System, Version 2.
   3  *
   4  * Copyright (C) 2001-2003 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  * $Id: gc.c,v 1.155 2005/11/07 11:14:39 gleixner Exp $
  11  *
  12  */
  13
  14 #include <linux/kernel.h>
  15 #include <linux/mtd/mtd.h>
  16 #include <linux/slab.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/crc32.h>
  19 #include <linux/compiler.h>
  20 #include <linux/stat.h>
  21 #include "nodelist.h"
  22 #include "compr.h"
  23
  24 static int jffs2_garbage_collect_pristine(struct jffs2_sb_info *c,
  25                                           struct jffs2_inode_cache *ic,
  26                                           struct jffs2_raw_node_ref *raw);
  27 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
  28                                         struct jffs2_inode_info *f, struct jffs2_full_dnode *fd);
  29 static int jffs2_garbage_collect_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_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
  32                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd);
  33 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
  34                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
  35                                       uint32_t start, uint32_t end);
  36 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
  37                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
  38                                        uint32_t start, uint32_t end);
  39 static int jffs2_garbage_collect_live(struct jffs2_sb_info *c,  struct jffs2_eraseblock *jeb,
  40                                struct jffs2_raw_node_ref *raw, struct jffs2_inode_info *f);
  41
  42 /* Called with erase_completion_lock held */
  43 static struct jffs2_eraseblock *jffs2_find_gc_block(struct jffs2_sb_info *c)
  44 {
  45         struct jffs2_eraseblock *ret;
  46         struct list_head *nextlist = NULL;
  47         int n = jiffies % 128;
  48
  49         /* Pick an eraseblock to garbage collect next. This is where we'll
  50            put the clever wear-levelling algorithms. Eventually.  */
  51         /* We possibly want to favour the dirtier blocks more when the
  52            number of free blocks is low. */
  53 again:
  54         if (!list_empty(&c->bad_used_list) && c->nr_free_blocks > c->resv_blocks_gcbad) {
  55                 D1(printk(KERN_DEBUG "Picking block from bad_used_list to GC next\n"));
  56                 nextlist = &c->bad_used_list;
  57         } else if (n < 50 && !list_empty(&c->erasable_list)) {
  58                 /* Note that most of them will have gone directly to be erased.
  59                    So don't favour the erasable_list _too_ much. */
  60                 D1(printk(KERN_DEBUG "Picking block from erasable_list to GC next\n"));
  61                 nextlist = &c->erasable_list;
  62         } else if (n < 110 && !list_empty(&c->very_dirty_list)) {
  63                 /* Most of the time, pick one off the very_dirty list */
  64                 D1(printk(KERN_DEBUG "Picking block from very_dirty_list to GC next\n"));
  65                 nextlist = &c->very_dirty_list;
  66         } else if (n < 126 && !list_empty(&c->dirty_list)) {
  67                 D1(printk(KERN_DEBUG "Picking block from dirty_list to GC next\n"));
  68                 nextlist = &c->dirty_list;
  69         } else if (!list_empty(&c->clean_list)) {
  70                 D1(printk(KERN_DEBUG "Picking block from clean_list to GC next\n"));
  71                 nextlist = &c->clean_list;
  72         } else if (!list_empty(&c->dirty_list)) {
  73                 D1(printk(KERN_DEBUG "Picking block from dirty_list to GC next (clean_list was empty)\n"));
  74
  75                 nextlist = &c->dirty_list;
  76         } else if (!list_empty(&c->very_dirty_list)) {
  77                 D1(printk(KERN_DEBUG "Picking block from very_dirty_list to GC next (clean_list and dirty_list were empty)\n"));
  78                 nextlist = &c->very_dirty_list;
  79         } else if (!list_empty(&c->erasable_list)) {
  80                 D1(printk(KERN_DEBUG "Picking block from erasable_list to GC next (clean_list and {very_,}dirty_list were empty)\n"));
  81
  82                 nextlist = &c->erasable_list;
  83         } else if (!list_empty(&c->erasable_pending_wbuf_list)) {
  84                 /* There are blocks are wating for the wbuf sync */
  85                 D1(printk(KERN_DEBUG "Synching wbuf in order to reuse erasable_pending_wbuf_list blocks\n"));
  86                 spin_unlock(&c->erase_completion_lock);
  87                 jffs2_flush_wbuf_pad(c);
  88                 spin_lock(&c->erase_completion_lock);
  89                 goto again;
  90         } else {
  91                 /* Eep. All were empty */
  92                 D1(printk(KERN_NOTICE "jffs2: No clean, dirty _or_ erasable blocks to GC from! Where are they all?\n"));
  93                 return NULL;
  94         }
  95
  96         ret = list_entry(nextlist->next, struct jffs2_eraseblock, list);
  97         list_del(&ret->list);
  98         c->gcblock = ret;
  99         ret->gc_node = ret->first_node;
 100         if (!ret->gc_node) {
 101                 printk(KERN_WARNING "Eep. ret->gc_node for block at 0x%08x is NULL\n", ret->offset);
 102                 BUG();
 103         }
 104
 105         /* Have we accidentally picked a clean block with wasted space ? */
 106         if (ret->wasted_size) {
 107                 D1(printk(KERN_DEBUG "Converting wasted_size %08x to dirty_size\n", ret->wasted_size));
 108                 ret->dirty_size += ret->wasted_size;
 109                 c->wasted_size -= ret->wasted_size;
 110                 c->dirty_size += ret->wasted_size;
 111                 ret->wasted_size = 0;
 112         }
 113
 114         return ret;
 115 }
 116
 117 /* jffs2_garbage_collect_pass
 118  * Make a single attempt to progress GC. Move one node, and possibly
 119  * start erasing one eraseblock.
 120  */
 121 int jffs2_garbage_collect_pass(struct jffs2_sb_info *c)
 122 {
 123         struct jffs2_inode_info *f;
 124         struct jffs2_inode_cache *ic;
 125         struct jffs2_eraseblock *jeb;
 126         struct jffs2_raw_node_ref *raw;
 127         int ret = 0, inum, nlink;
 128         int xattr = 0;
 129
 130         if (down_interruptible(&c->alloc_sem))
 131                 return -EINTR;
 132
 133         for (;;) {
 134                 spin_lock(&c->erase_completion_lock);
 135                 if (!c->unchecked_size)
 136                         break;
 137
 138                 /* We can't start doing GC yet. We haven't finished checking
 139                    the node CRCs etc. Do it now. */
 140
 141                 /* checked_ino is protected by the alloc_sem */
 142                 if (c->checked_ino > c->highest_ino && xattr) {
 143                         printk(KERN_CRIT "Checked all inodes but still 0x%x bytes of unchecked space?\n",
 144                                c->unchecked_size);
 145                         jffs2_dbg_dump_block_lists_nolock(c);
 146                         spin_unlock(&c->erase_completion_lock);
 147                         BUG();
 148                 }
 149
 150                 spin_unlock(&c->erase_completion_lock);
 151
 152                 if (!xattr)
 153                         xattr = jffs2_verify_xattr(c);
 154
 155                 spin_lock(&c->inocache_lock);
 156
 157                 ic = jffs2_get_ino_cache(c, c->checked_ino++);
 158
 159                 if (!ic) {
 160                         spin_unlock(&c->inocache_lock);
 161                         continue;
 162                 }
 163
 164                 if (!ic->nlink) {
 165                         D1(printk(KERN_DEBUG "Skipping check of ino #%d with nlink zero\n",
 166                                   ic->ino));
 167                         spin_unlock(&c->inocache_lock);
 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                         /* We only care about obsolete ones */
 841                         if (!(ref_obsolete(raw)))
 842                                 continue;
 843
 844                         /* Any dirent with the same name is going to have the same length... */
 845                         if (ref_totlen(c, NULL, raw) != rawlen)
 846                                 continue;
 847
 848                         /* Doesn't matter if there's one in the same erase block. We're going to
 849                            delete it too at the same time. */
 850                         if (SECTOR_ADDR(raw->flash_offset) == SECTOR_ADDR(fd->raw->flash_offset))
 851                                 continue;
 852
 853                         D1(printk(KERN_DEBUG "Check potential deletion dirent at %08x\n", ref_offset(raw)));
 854
 855                         /* This is an obsolete node belonging to the same directory, and it's of the right
 856                            length. We need to take a closer look...*/
 857                         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)rd);
 858                         if (ret) {
 859                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Read error (%d) reading obsolete node at %08x\n", ret, ref_offset(raw));
 860                                 /* If we can't read it, we don't need to continue to obsolete it. Continue */
 861                                 continue;
 862                         }
 863                         if (retlen != rawlen) {
 864                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Short read (%zd not %u) reading header from obsolete node at %08x\n",
 865                                        retlen, rawlen, ref_offset(raw));
 866                                 continue;
 867                         }
 868
 869                         if (je16_to_cpu(rd->nodetype) != JFFS2_NODETYPE_DIRENT)
 870                                 continue;
 871
 872                         /* If the name CRC doesn't match, skip */
 873                         if (je32_to_cpu(rd->name_crc) != name_crc)
 874                                 continue;
 875
 876                         /* If the name length doesn't match, or it's another deletion dirent, skip */
 877                         if (rd->nsize != name_len || !je32_to_cpu(rd->ino))
 878                                 continue;
 879
 880                         /* OK, check the actual name now */
 881                         if (memcmp(rd->name, fd->name, name_len))
 882                                 continue;
 883
 884                         /* OK. The name really does match. There really is still an older node on
 885                            the flash which our deletion dirent obsoletes. So we have to write out
 886                            a new deletion dirent to replace it */
 887                         up(&c->erase_free_sem);
 888
 889                         D1(printk(KERN_DEBUG "Deletion dirent at %08x still obsoletes real dirent \"%s\" at %08x for ino #%u\n",
 890                                   ref_offset(fd->raw), fd->name, ref_offset(raw), je32_to_cpu(rd->ino)));
 891                         kfree(rd);
 892
 893                         return jffs2_garbage_collect_dirent(c, jeb, f, fd);
 894                 }
 895
 896                 up(&c->erase_free_sem);
 897                 kfree(rd);
 898         }
 899
 900         /* FIXME: If we're deleting a dirent which contains the current mtime and ctime,
 901            we should update the metadata node with those times accordingly */
 902
 903         /* No need for it any more. Just mark it obsolete and remove it from the list */
 904         while (*fdp) {
 905                 if ((*fdp) == fd) {
 906                         found = 1;
 907                         *fdp = fd->next;
 908                         break;
 909                 }
 910                 fdp = &(*fdp)->next;
 911         }
 912         if (!found) {
 913                 printk(KERN_WARNING "Deletion dirent \"%s\" not found in list for ino #%u\n", fd->name, f->inocache->ino);
 914         }
 915         jffs2_mark_node_obsolete(c, fd->raw);
 916         jffs2_free_full_dirent(fd);
 917         return 0;
 918 }
 919
 920 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 921                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
 922                                       uint32_t start, uint32_t end)
 923 {
 924         struct jffs2_raw_inode ri;
 925         struct jffs2_node_frag *frag;
 926         struct jffs2_full_dnode *new_fn;
 927         uint32_t alloclen, ilen;
 928         int ret;
 929
 930         D1(printk(KERN_DEBUG "Writing replacement hole node for ino #%u from offset 0x%x to 0x%x\n",
 931                   f->inocache->ino, start, end));
 932
 933         memset(&ri, 0, sizeof(ri));
 934
 935         if(fn->frags > 1) {
 936                 size_t readlen;
 937                 uint32_t crc;
 938                 /* It's partially obsoleted by a later write. So we have to
 939                    write it out again with the _same_ version as before */
 940                 ret = jffs2_flash_read(c, ref_offset(fn->raw), sizeof(ri), &readlen, (char *)&ri);
 941                 if (readlen != sizeof(ri) || ret) {
 942                         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);
 943                         goto fill;
 944                 }
 945                 if (je16_to_cpu(ri.nodetype) != JFFS2_NODETYPE_INODE) {
 946                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had node type 0x%04x instead of JFFS2_NODETYPE_INODE(0x%04x)\n",
 947                                ref_offset(fn->raw),
 948                                je16_to_cpu(ri.nodetype), JFFS2_NODETYPE_INODE);
 949                         return -EIO;
 950                 }
 951                 if (je32_to_cpu(ri.totlen) != sizeof(ri)) {
 952                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had totlen 0x%x instead of expected 0x%zx\n",
 953                                ref_offset(fn->raw),
 954                                je32_to_cpu(ri.totlen), sizeof(ri));
 955                         return -EIO;
 956                 }
 957                 crc = crc32(0, &ri, sizeof(ri)-8);
 958                 if (crc != je32_to_cpu(ri.node_crc)) {
 959                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had CRC 0x%08x which doesn't match calculated CRC 0x%08x\n",
 960                                ref_offset(fn->raw),
 961                                je32_to_cpu(ri.node_crc), crc);
 962                         /* FIXME: We could possibly deal with this by writing new holes for each frag */
 963                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
 964                                start, end, f->inocache->ino);
 965                         goto fill;
 966                 }
 967                 if (ri.compr != JFFS2_COMPR_ZERO) {
 968                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node 0x%08x wasn't a hole node!\n", ref_offset(fn->raw));
 969                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
 970                                start, end, f->inocache->ino);
 971                         goto fill;
 972                 }
 973         } else {
 974         fill:
 975                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
 976                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
 977                 ri.totlen = cpu_to_je32(sizeof(ri));
 978                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
 979
 980                 ri.ino = cpu_to_je32(f->inocache->ino);
 981                 ri.version = cpu_to_je32(++f->highest_version);
 982                 ri.offset = cpu_to_je32(start);
 983                 ri.dsize = cpu_to_je32(end - start);
 984                 ri.csize = cpu_to_je32(0);
 985                 ri.compr = JFFS2_COMPR_ZERO;
 986         }
 987
 988         frag = frag_last(&f->fragtree);
 989         if (frag)
 990                 /* Fetch the inode length from the fragtree rather then
 991                  * from i_size since i_size may have not been updated yet */
 992                 ilen = frag->ofs + frag->size;
 993         else
 994                 ilen = JFFS2_F_I_SIZE(f);
 995
 996         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
 997         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
 998         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
 999         ri.isize = cpu_to_je32(ilen);
1000         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1001         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1002         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1003         ri.data_crc = cpu_to_je32(0);
1004         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1005
1006         ret = jffs2_reserve_space_gc(c, sizeof(ri), &alloclen,
1007                                      JFFS2_SUMMARY_INODE_SIZE);
1008         if (ret) {
1009                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_hole failed: %d\n",
1010                        sizeof(ri), ret);
1011                 return ret;
1012         }
1013         new_fn = jffs2_write_dnode(c, f, &ri, NULL, 0, ALLOC_GC);
1014
1015         if (IS_ERR(new_fn)) {
1016                 printk(KERN_WARNING "Error writing new hole node: %ld\n", PTR_ERR(new_fn));
1017                 return PTR_ERR(new_fn);
1018         }
1019         if (je32_to_cpu(ri.version) == f->highest_version) {
1020                 jffs2_add_full_dnode_to_inode(c, f, new_fn);
1021                 if (f->metadata) {
1022                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1023                         jffs2_free_full_dnode(f->metadata);
1024                         f->metadata = NULL;
1025                 }
1026                 return 0;
1027         }
1028
1029         /*
1030          * We should only get here in the case where the node we are
1031          * replacing had more than one frag, so we kept the same version
1032          * number as before. (Except in case of error -- see 'goto fill;'
1033          * above.)
1034          */
1035         D1(if(unlikely(fn->frags <= 1)) {
1036                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Replacing fn with %d frag(s) but new ver %d != highest_version %d of ino #%d\n",
1037                        fn->frags, je32_to_cpu(ri.version), f->highest_version,
1038                        je32_to_cpu(ri.ino));
1039         });
1040
1041         /* This is a partially-overlapped hole node. Mark it REF_NORMAL not REF_PRISTINE */
1042         mark_ref_normal(new_fn->raw);
1043
1044         for (frag = jffs2_lookup_node_frag(&f->fragtree, fn->ofs);
1045              frag; frag = frag_next(frag)) {
1046                 if (frag->ofs > fn->size + fn->ofs)
1047                         break;
1048                 if (frag->node == fn) {
1049                         frag->node = new_fn;
1050                         new_fn->frags++;
1051                         fn->frags--;
1052                 }
1053         }
1054         if (fn->frags) {
1055                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Old node still has frags!\n");
1056                 BUG();
1057         }
1058         if (!new_fn->frags) {
1059                 printk(KERN_WARNING "jffs2_garbage_collect_hole: New node has no frags!\n");
1060                 BUG();
1061         }
1062
1063         jffs2_mark_node_obsolete(c, fn->raw);
1064         jffs2_free_full_dnode(fn);
1065
1066         return 0;
1067 }
1068
1069 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
1070                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
1071                                        uint32_t start, uint32_t end)
1072 {
1073         struct jffs2_full_dnode *new_fn;
1074         struct jffs2_raw_inode ri;
1075         uint32_t alloclen, offset, orig_end, orig_start;
1076         int ret = 0;
1077         unsigned char *comprbuf = NULL, *writebuf;
1078         unsigned long pg;
1079         unsigned char *pg_ptr;
1080
1081         memset(&ri, 0, sizeof(ri));
1082
1083         D1(printk(KERN_DEBUG "Writing replacement dnode for ino #%u from offset 0x%x to 0x%x\n",
1084                   f->inocache->ino, start, end));
1085
1086         orig_end = end;
1087         orig_start = start;
1088
1089         if (c->nr_free_blocks + c->nr_erasing_blocks > c->resv_blocks_gcmerge) {
1090                 /* Attempt to do some merging. But only expand to cover logically
1091                    adjacent frags if the block containing them is already considered
1092                    to be dirty. Otherwise we end up with GC just going round in
1093                    circles dirtying the nodes it already wrote out, especially
1094                    on NAND where we have small eraseblocks and hence a much higher
1095                    chance of nodes having to be split to cross boundaries. */
1096
1097                 struct jffs2_node_frag *frag;
1098                 uint32_t min, max;
1099
1100                 min = start & ~(PAGE_CACHE_SIZE-1);
1101                 max = min + PAGE_CACHE_SIZE;
1102
1103                 frag = jffs2_lookup_node_frag(&f->fragtree, start);
1104
1105                 /* BUG_ON(!frag) but that'll happen anyway... */
1106
1107                 BUG_ON(frag->ofs != start);
1108
1109                 /* First grow down... */
1110                 while((frag = frag_prev(frag)) && frag->ofs >= min) {
1111
1112                         /* If the previous frag doesn't even reach the beginning, there's
1113                            excessive fragmentation. Just merge. */
1114                         if (frag->ofs > min) {
1115                                 D1(printk(KERN_DEBUG "Expanding down to cover partial frag (0x%x-0x%x)\n",
1116                                           frag->ofs, frag->ofs+frag->size));
1117                                 start = frag->ofs;
1118                                 continue;
1119                         }
1120                         /* OK. This frag holds the first byte of the page. */
1121                         if (!frag->node || !frag->node->raw) {
1122                                 D1(printk(KERN_DEBUG "First frag in page is hole (0x%x-0x%x). Not expanding down.\n",
1123                                           frag->ofs, frag->ofs+frag->size));
1124                                 break;
1125                         } else {
1126
1127                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1128                                    in a block which is still considered clean? If so, don't obsolete it.
1129                                    If not, cover it anyway. */
1130
1131                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1132                                 struct jffs2_eraseblock *jeb;
1133
1134                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1135
1136                                 if (jeb == c->gcblock) {
1137                                         D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1138                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1139                                         start = frag->ofs;
1140                                         break;
1141                                 }
1142                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1143                                         D1(printk(KERN_DEBUG "Not expanding down to cover frag (0x%x-0x%x) in clean block %08x\n",
1144                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1145                                         break;
1146                                 }
1147
1148                                 D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in dirty block %08x\n",
1149                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1150                                 start = frag->ofs;
1151                                 break;
1152                         }
1153                 }
1154
1155                 /* ... then up */
1156
1157                 /* Find last frag which is actually part of the node we're to GC. */
1158                 frag = jffs2_lookup_node_frag(&f->fragtree, end-1);
1159
1160                 while((frag = frag_next(frag)) && frag->ofs+frag->size <= max) {
1161
1162                         /* If the previous frag doesn't even reach the beginning, there's lots
1163                            of fragmentation. Just merge. */
1164                         if (frag->ofs+frag->size < max) {
1165                                 D1(printk(KERN_DEBUG "Expanding up to cover partial frag (0x%x-0x%x)\n",
1166                                           frag->ofs, frag->ofs+frag->size));
1167                                 end = frag->ofs + frag->size;
1168                                 continue;
1169                         }
1170
1171                         if (!frag->node || !frag->node->raw) {
1172                                 D1(printk(KERN_DEBUG "Last frag in page is hole (0x%x-0x%x). Not expanding up.\n",
1173                                           frag->ofs, frag->ofs+frag->size));
1174                                 break;
1175                         } else {
1176
1177                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1178                                    in a block which is still considered clean? If so, don't obsolete it.
1179                                    If not, cover it anyway. */
1180
1181                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1182                                 struct jffs2_eraseblock *jeb;
1183
1184                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1185
1186                                 if (jeb == c->gcblock) {
1187                                         D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1188                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1189                                         end = frag->ofs + frag->size;
1190                                         break;
1191                                 }
1192                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1193                                         D1(printk(KERN_DEBUG "Not expanding up to cover frag (0x%x-0x%x) in clean block %08x\n",
1194                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1195                                         break;
1196                                 }
1197
1198                                 D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in dirty block %08x\n",
1199                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1200                                 end = frag->ofs + frag->size;
1201                                 break;
1202                         }
1203                 }
1204                 D1(printk(KERN_DEBUG "Expanded dnode to write from (0x%x-0x%x) to (0x%x-0x%x)\n",
1205                           orig_start, orig_end, start, end));
1206
1207                 D1(BUG_ON(end > frag_last(&f->fragtree)->ofs + frag_last(&f->fragtree)->size));
1208                 BUG_ON(end < orig_end);
1209                 BUG_ON(start > orig_start);
1210         }
1211
1212         /* First, use readpage() to read the appropriate page into the page cache */
1213         /* Q: What happens if we actually try to GC the _same_ page for which commit_write()
1214          *    triggered garbage collection in the first place?
1215          * A: I _think_ it's OK. read_cache_page shouldn't deadlock, we'll write out the
1216          *    page OK. We'll actually write it out again in commit_write, which is a little
1217          *    suboptimal, but at least we're correct.
1218          */
1219         pg_ptr = jffs2_gc_fetch_page(c, f, start, &pg);
1220
1221         if (IS_ERR(pg_ptr)) {
1222                 printk(KERN_WARNING "read_cache_page() returned error: %ld\n", PTR_ERR(pg_ptr));
1223                 return PTR_ERR(pg_ptr);
1224         }
1225
1226         offset = start;
1227         while(offset < orig_end) {
1228                 uint32_t datalen;
1229                 uint32_t cdatalen;
1230                 uint16_t comprtype = JFFS2_COMPR_NONE;
1231
1232                 ret = jffs2_reserve_space_gc(c, sizeof(ri) + JFFS2_MIN_DATA_LEN,
1233                                         &alloclen, JFFS2_SUMMARY_INODE_SIZE);
1234
1235                 if (ret) {
1236                         printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dnode failed: %d\n",
1237                                sizeof(ri)+ JFFS2_MIN_DATA_LEN, ret);
1238                         break;
1239                 }
1240                 cdatalen = min_t(uint32_t, alloclen - sizeof(ri), end - offset);
1241                 datalen = end - offset;
1242
1243                 writebuf = pg_ptr + (offset & (PAGE_CACHE_SIZE -1));
1244
1245                 comprtype = jffs2_compress(c, f, writebuf, &comprbuf, &datalen, &cdatalen);
1246
1247                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
1248                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
1249                 ri.totlen = cpu_to_je32(sizeof(ri) + cdatalen);
1250                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
1251
1252                 ri.ino = cpu_to_je32(f->inocache->ino);
1253                 ri.version = cpu_to_je32(++f->highest_version);
1254                 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1255                 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1256                 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1257                 ri.isize = cpu_to_je32(JFFS2_F_I_SIZE(f));
1258                 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1259                 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1260                 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1261                 ri.offset = cpu_to_je32(offset);
1262                 ri.csize = cpu_to_je32(cdatalen);
1263                 ri.dsize = cpu_to_je32(datalen);
1264                 ri.compr = comprtype & 0xff;
1265                 ri.usercompr = (comprtype >> 8) & 0xff;
1266                 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1267                 ri.data_crc = cpu_to_je32(crc32(0, comprbuf, cdatalen));
1268
1269                 new_fn = jffs2_write_dnode(c, f, &ri, comprbuf, cdatalen, ALLOC_GC);
1270
1271                 jffs2_free_comprbuf(comprbuf, writebuf);
1272
1273                 if (IS_ERR(new_fn)) {
1274                         printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
1275                         ret = PTR_ERR(new_fn);
1276                         break;
1277                 }
1278                 ret = jffs2_add_full_dnode_to_inode(c, f, new_fn);
1279                 offset += datalen;
1280                 if (f->metadata) {
1281                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1282                         jffs2_free_full_dnode(f->metadata);
1283                         f->metadata = NULL;
1284                 }
1285         }
1286
1287         jffs2_gc_release_page(c, pg_ptr, &pg);
1288         return ret;
1289 }