git.oblomov.eu Git - 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                         jffs2_xattr_delete_inode(c, ic);
 169                         continue;
 170                 }
 171                 switch(ic->state) {
 172                 case INO_STATE_CHECKEDABSENT:
 173                 case INO_STATE_PRESENT:
 174                         D1(printk(KERN_DEBUG "Skipping ino #%u already checked\n", ic->ino));
 175                         spin_unlock(&c->inocache_lock);
 176                         continue;
 177
 178                 case INO_STATE_GC:
 179                 case INO_STATE_CHECKING:
 180                         printk(KERN_WARNING "Inode #%u is in state %d during CRC check phase!\n", ic->ino, ic->state);
 181                         spin_unlock(&c->inocache_lock);
 182                         BUG();
 183
 184                 case INO_STATE_READING:
 185                         /* We need to wait for it to finish, lest we move on
 186                            and trigger the BUG() above while we haven't yet
 187                            finished checking all its nodes */
 188                         D1(printk(KERN_DEBUG "Waiting for ino #%u to finish reading\n", ic->ino));
 189                         /* We need to come back again for the _same_ inode. We've
 190                          made no progress in this case, but that should be OK */
 191                         c->checked_ino--;
 192
 193                         up(&c->alloc_sem);
 194                         sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
 195                         return 0;
 196
 197                 default:
 198                         BUG();
 199
 200                 case INO_STATE_UNCHECKED:
 201                         ;
 202                 }
 203                 ic->state = INO_STATE_CHECKING;
 204                 spin_unlock(&c->inocache_lock);
 205
 206                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() triggering inode scan of ino#%u\n", ic->ino));
 207
 208                 ret = jffs2_do_crccheck_inode(c, ic);
 209                 if (ret)
 210                         printk(KERN_WARNING "Returned error for crccheck of ino #%u. Expect badness...\n", ic->ino);
 211
 212                 jffs2_set_inocache_state(c, ic, INO_STATE_CHECKEDABSENT);
 213                 up(&c->alloc_sem);
 214                 return ret;
 215         }
 216
 217         /* First, work out which block we're garbage-collecting */
 218         jeb = c->gcblock;
 219
 220         if (!jeb)
 221                 jeb = jffs2_find_gc_block(c);
 222
 223         if (!jeb) {
 224                 D1 (printk(KERN_NOTICE "jffs2: Couldn't find erase block to garbage collect!\n"));
 225                 spin_unlock(&c->erase_completion_lock);
 226                 up(&c->alloc_sem);
 227                 return -EIO;
 228         }
 229
 230         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));
 231         D1(if (c->nextblock)
 232            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));
 233
 234         if (!jeb->used_size) {
 235                 up(&c->alloc_sem);
 236                 goto eraseit;
 237         }
 238
 239         raw = jeb->gc_node;
 240
 241         while(ref_obsolete(raw)) {
 242                 D1(printk(KERN_DEBUG "Node at 0x%08x is obsolete... skipping\n", ref_offset(raw)));
 243                 raw = ref_next(raw);
 244                 if (unlikely(!raw)) {
 245                         printk(KERN_WARNING "eep. End of raw list while still supposedly nodes to GC\n");
 246                         printk(KERN_WARNING "erase block at 0x%08x. free_size 0x%08x, dirty_size 0x%08x, used_size 0x%08x\n",
 247                                jeb->offset, jeb->free_size, jeb->dirty_size, jeb->used_size);
 248                         jeb->gc_node = raw;
 249                         spin_unlock(&c->erase_completion_lock);
 250                         up(&c->alloc_sem);
 251                         BUG();
 252                 }
 253         }
 254         jeb->gc_node = raw;
 255
 256         D1(printk(KERN_DEBUG "Going to garbage collect node at 0x%08x\n", ref_offset(raw)));
 257
 258         if (!raw->next_in_ino) {
 259                 /* Inode-less node. Clean marker, snapshot or something like that */
 260                 spin_unlock(&c->erase_completion_lock);
 261                 if (ref_flags(raw) == REF_PRISTINE) {
 262                         /* It's an unknown node with JFFS2_FEATURE_RWCOMPAT_COPY */
 263                         jffs2_garbage_collect_pristine(c, NULL, raw);
 264                 } else {
 265                         /* Just mark it obsolete */
 266                         jffs2_mark_node_obsolete(c, raw);
 267                 }
 268                 up(&c->alloc_sem);
 269                 goto eraseit_lock;
 270         }
 271
 272         ic = jffs2_raw_ref_to_ic(raw);
 273
 274 #ifdef CONFIG_JFFS2_FS_XATTR
 275         /* When 'ic' refers xattr_datum/xattr_ref, this node is GCed as xattr.
 276          * We can decide whether this node is inode or xattr by ic->class.     */
 277         if (ic->class == RAWNODE_CLASS_XATTR_DATUM
 278             || ic->class == RAWNODE_CLASS_XATTR_REF) {
 279                 spin_unlock(&c->erase_completion_lock);
 280
 281                 if (ic->class == RAWNODE_CLASS_XATTR_DATUM) {
 282                         ret = jffs2_garbage_collect_xattr_datum(c, (struct jffs2_xattr_datum *)ic, raw);
 283                 } else {
 284                         ret = jffs2_garbage_collect_xattr_ref(c, (struct jffs2_xattr_ref *)ic, raw);
 285                 }
 286                 goto release_sem;
 287         }
 288 #endif
 289
 290         /* We need to hold the inocache. Either the erase_completion_lock or
 291            the inocache_lock are sufficient; we trade down since the inocache_lock
 292            causes less contention. */
 293         spin_lock(&c->inocache_lock);
 294
 295         spin_unlock(&c->erase_completion_lock);
 296
 297         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));
 298
 299         /* Three possibilities:
 300            1. Inode is already in-core. We must iget it and do proper
 301               updating to its fragtree, etc.
 302            2. Inode is not in-core, node is REF_PRISTINE. We lock the
 303               inocache to prevent a read_inode(), copy the node intact.
 304            3. Inode is not in-core, node is not pristine. We must iget()
 305               and take the slow path.
 306         */
 307
 308         switch(ic->state) {
 309         case INO_STATE_CHECKEDABSENT:
 310                 /* It's been checked, but it's not currently in-core.
 311                    We can just copy any pristine nodes, but have
 312                    to prevent anyone else from doing read_inode() while
 313                    we're at it, so we set the state accordingly */
 314                 if (ref_flags(raw) == REF_PRISTINE)
 315                         ic->state = INO_STATE_GC;
 316                 else {
 317                         D1(printk(KERN_DEBUG "Ino #%u is absent but node not REF_PRISTINE. Reading.\n",
 318                                   ic->ino));
 319                 }
 320                 break;
 321
 322         case INO_STATE_PRESENT:
 323                 /* It's in-core. GC must iget() it. */
 324                 break;
 325
 326         case INO_STATE_UNCHECKED:
 327         case INO_STATE_CHECKING:
 328         case INO_STATE_GC:
 329                 /* Should never happen. We should have finished checking
 330                    by the time we actually start doing any GC, and since
 331                    we're holding the alloc_sem, no other garbage collection
 332                    can happen.
 333                 */
 334                 printk(KERN_CRIT "Inode #%u already in state %d in jffs2_garbage_collect_pass()!\n",
 335                        ic->ino, ic->state);
 336                 up(&c->alloc_sem);
 337                 spin_unlock(&c->inocache_lock);
 338                 BUG();
 339
 340         case INO_STATE_READING:
 341                 /* Someone's currently trying to read it. We must wait for
 342                    them to finish and then go through the full iget() route
 343                    to do the GC. However, sometimes read_inode() needs to get
 344                    the alloc_sem() (for marking nodes invalid) so we must
 345                    drop the alloc_sem before sleeping. */
 346
 347                 up(&c->alloc_sem);
 348                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() waiting for ino #%u in state %d\n",
 349                           ic->ino, ic->state));
 350                 sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
 351                 /* And because we dropped the alloc_sem we must start again from the
 352                    beginning. Ponder chance of livelock here -- we're returning success
 353                    without actually making any progress.
 354
 355                    Q: What are the chances that the inode is back in INO_STATE_READING
 356                    again by the time we next enter this function? And that this happens
 357                    enough times to cause a real delay?
 358
 359                    A: Small enough that I don't care :)
 360                 */
 361                 return 0;
 362         }
 363
 364         /* OK. Now if the inode is in state INO_STATE_GC, we are going to copy the
 365            node intact, and we don't have to muck about with the fragtree etc.
 366            because we know it's not in-core. If it _was_ in-core, we go through
 367            all the iget() crap anyway */
 368
 369         if (ic->state == INO_STATE_GC) {
 370                 spin_unlock(&c->inocache_lock);
 371
 372                 ret = jffs2_garbage_collect_pristine(c, ic, raw);
 373
 374                 spin_lock(&c->inocache_lock);
 375                 ic->state = INO_STATE_CHECKEDABSENT;
 376                 wake_up(&c->inocache_wq);
 377
 378                 if (ret != -EBADFD) {
 379                         spin_unlock(&c->inocache_lock);
 380                         goto release_sem;
 381                 }
 382
 383                 /* Fall through if it wanted us to, with inocache_lock held */
 384         }
 385
 386         /* Prevent the fairly unlikely race where the gcblock is
 387            entirely obsoleted by the final close of a file which had
 388            the only valid nodes in the block, followed by erasure,
 389            followed by freeing of the ic because the erased block(s)
 390            held _all_ the nodes of that inode.... never been seen but
 391            it's vaguely possible. */
 392
 393         inum = ic->ino;
 394         nlink = ic->nlink;
 395         spin_unlock(&c->inocache_lock);
 396
 397         f = jffs2_gc_fetch_inode(c, inum, nlink);
 398         if (IS_ERR(f)) {
 399                 ret = PTR_ERR(f);
 400                 goto release_sem;
 401         }
 402         if (!f) {
 403                 ret = 0;
 404                 goto release_sem;
 405         }
 406
 407         ret = jffs2_garbage_collect_live(c, jeb, raw, f);
 408
 409         jffs2_gc_release_inode(c, f);
 410
 411  release_sem:
 412         up(&c->alloc_sem);
 413
 414  eraseit_lock:
 415         /* If we've finished this block, start it erasing */
 416         spin_lock(&c->erase_completion_lock);
 417
 418  eraseit:
 419         if (c->gcblock && !c->gcblock->used_size) {
 420                 D1(printk(KERN_DEBUG "Block at 0x%08x completely obsoleted by GC. Moving to erase_pending_list\n", c->gcblock->offset));
 421                 /* We're GC'ing an empty block? */
 422                 list_add_tail(&c->gcblock->list, &c->erase_pending_list);
 423                 c->gcblock = NULL;
 424                 c->nr_erasing_blocks++;
 425                 jffs2_erase_pending_trigger(c);
 426         }
 427         spin_unlock(&c->erase_completion_lock);
 428
 429         return ret;
 430 }
 431
 432 static int jffs2_garbage_collect_live(struct jffs2_sb_info *c,  struct jffs2_eraseblock *jeb,
 433                                       struct jffs2_raw_node_ref *raw, struct jffs2_inode_info *f)
 434 {
 435         struct jffs2_node_frag *frag;
 436         struct jffs2_full_dnode *fn = NULL;
 437         struct jffs2_full_dirent *fd;
 438         uint32_t start = 0, end = 0, nrfrags = 0;
 439         int ret = 0;
 440
 441         down(&f->sem);
 442
 443         /* Now we have the lock for this inode. Check that it's still the one at the head
 444            of the list. */
 445
 446         spin_lock(&c->erase_completion_lock);
 447
 448         if (c->gcblock != jeb) {
 449                 spin_unlock(&c->erase_completion_lock);
 450                 D1(printk(KERN_DEBUG "GC block is no longer gcblock. Restart\n"));
 451                 goto upnout;
 452         }
 453         if (ref_obsolete(raw)) {
 454                 spin_unlock(&c->erase_completion_lock);
 455                 D1(printk(KERN_DEBUG "node to be GC'd was obsoleted in the meantime.\n"));
 456                 /* They'll call again */
 457                 goto upnout;
 458         }
 459         spin_unlock(&c->erase_completion_lock);
 460
 461         /* OK. Looks safe. And nobody can get us now because we have the semaphore. Move the block */
 462         if (f->metadata && f->metadata->raw == raw) {
 463                 fn = f->metadata;
 464                 ret = jffs2_garbage_collect_metadata(c, jeb, f, fn);
 465                 goto upnout;
 466         }
 467
 468         /* FIXME. Read node and do lookup? */
 469         for (frag = frag_first(&f->fragtree); frag; frag = frag_next(frag)) {
 470                 if (frag->node && frag->node->raw == raw) {
 471                         fn = frag->node;
 472                         end = frag->ofs + frag->size;
 473                         if (!nrfrags++)
 474                                 start = frag->ofs;
 475                         if (nrfrags == frag->node->frags)
 476                                 break; /* We've found them all */
 477                 }
 478         }
 479         if (fn) {
 480                 if (ref_flags(raw) == REF_PRISTINE) {
 481                         ret = jffs2_garbage_collect_pristine(c, f->inocache, raw);
 482                         if (!ret) {
 483                                 /* Urgh. Return it sensibly. */
 484                                 frag->node->raw = f->inocache->nodes;
 485                         }
 486                         if (ret != -EBADFD)
 487                                 goto upnout;
 488                 }
 489                 /* We found a datanode. Do the GC */
 490                 if((start >> PAGE_CACHE_SHIFT) < ((end-1) >> PAGE_CACHE_SHIFT)) {
 491                         /* It crosses a page boundary. Therefore, it must be a hole. */
 492                         ret = jffs2_garbage_collect_hole(c, jeb, f, fn, start, end);
 493                 } else {
 494                         /* It could still be a hole. But we GC the page this way anyway */
 495                         ret = jffs2_garbage_collect_dnode(c, jeb, f, fn, start, end);
 496                 }
 497                 goto upnout;
 498         }
 499
 500         /* Wasn't a dnode. Try dirent */
 501         for (fd = f->dents; fd; fd=fd->next) {
 502                 if (fd->raw == raw)
 503                         break;
 504         }
 505
 506         if (fd && fd->ino) {
 507                 ret = jffs2_garbage_collect_dirent(c, jeb, f, fd);
 508         } else if (fd) {
 509                 ret = jffs2_garbage_collect_deletion_dirent(c, jeb, f, fd);
 510         } else {
 511                 printk(KERN_WARNING "Raw node at 0x%08x wasn't in node lists for ino #%u\n",
 512                        ref_offset(raw), f->inocache->ino);
 513                 if (ref_obsolete(raw)) {
 514                         printk(KERN_WARNING "But it's obsolete so we don't mind too much\n");
 515                 } else {
 516                         jffs2_dbg_dump_node(c, ref_offset(raw));
 517                         BUG();
 518                 }
 519         }
 520  upnout:
 521         up(&f->sem);
 522
 523         return ret;
 524 }
 525
 526 static int jffs2_garbage_collect_pristine(struct jffs2_sb_info *c,
 527                                           struct jffs2_inode_cache *ic,
 528                                           struct jffs2_raw_node_ref *raw)
 529 {
 530         union jffs2_node_union *node;
 531         size_t retlen;
 532         int ret;
 533         uint32_t phys_ofs, alloclen;
 534         uint32_t crc, rawlen;
 535         int retried = 0;
 536
 537         D1(printk(KERN_DEBUG "Going to GC REF_PRISTINE node at 0x%08x\n", ref_offset(raw)));
 538
 539         alloclen = rawlen = ref_totlen(c, c->gcblock, raw);
 540
 541         /* Ask for a small amount of space (or the totlen if smaller) because we
 542            don't want to force wastage of the end of a block if splitting would
 543            work. */
 544         if (ic && alloclen > sizeof(struct jffs2_raw_inode) + JFFS2_MIN_DATA_LEN)
 545                 alloclen = sizeof(struct jffs2_raw_inode) + JFFS2_MIN_DATA_LEN;
 546
 547         ret = jffs2_reserve_space_gc(c, alloclen, &alloclen, rawlen);
 548         /* 'rawlen' is not the exact summary size; it is only an upper estimation */
 549
 550         if (ret)
 551                 return ret;
 552
 553         if (alloclen < rawlen) {
 554                 /* Doesn't fit untouched. We'll go the old route and split it */
 555                 return -EBADFD;
 556         }
 557
 558         node = kmalloc(rawlen, GFP_KERNEL);
 559         if (!node)
 560                return -ENOMEM;
 561
 562         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)node);
 563         if (!ret && retlen != rawlen)
 564                 ret = -EIO;
 565         if (ret)
 566                 goto out_node;
 567
 568         crc = crc32(0, node, sizeof(struct jffs2_unknown_node)-4);
 569         if (je32_to_cpu(node->u.hdr_crc) != crc) {
 570                 printk(KERN_WARNING "Header CRC failed on REF_PRISTINE node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 571                        ref_offset(raw), je32_to_cpu(node->u.hdr_crc), crc);
 572                 goto bail;
 573         }
 574
 575         switch(je16_to_cpu(node->u.nodetype)) {
 576         case JFFS2_NODETYPE_INODE:
 577                 crc = crc32(0, node, sizeof(node->i)-8);
 578                 if (je32_to_cpu(node->i.node_crc) != crc) {
 579                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 580                                ref_offset(raw), je32_to_cpu(node->i.node_crc), crc);
 581                         goto bail;
 582                 }
 583
 584                 if (je32_to_cpu(node->i.dsize)) {
 585                         crc = crc32(0, node->i.data, je32_to_cpu(node->i.csize));
 586                         if (je32_to_cpu(node->i.data_crc) != crc) {
 587                                 printk(KERN_WARNING "Data CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 588                                        ref_offset(raw), je32_to_cpu(node->i.data_crc), crc);
 589                                 goto bail;
 590                         }
 591                 }
 592                 break;
 593
 594         case JFFS2_NODETYPE_DIRENT:
 595                 crc = crc32(0, node, sizeof(node->d)-8);
 596                 if (je32_to_cpu(node->d.node_crc) != crc) {
 597                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE dirent node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 598                                ref_offset(raw), je32_to_cpu(node->d.node_crc), crc);
 599                         goto bail;
 600                 }
 601
 602                 if (node->d.nsize) {
 603                         crc = crc32(0, node->d.name, node->d.nsize);
 604                         if (je32_to_cpu(node->d.name_crc) != crc) {
 605                                 printk(KERN_WARNING "Name CRC failed on REF_PRISTINE dirent ode at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 606                                        ref_offset(raw), je32_to_cpu(node->d.name_crc), crc);
 607                                 goto bail;
 608                         }
 609                 }
 610                 break;
 611         default:
 612                 /* If it's inode-less, we don't _know_ what it is. Just copy it intact */
 613                 if (ic) {
 614                         printk(KERN_WARNING "Unknown node type for REF_PRISTINE node at 0x%08x: 0x%04x\n",
 615                                ref_offset(raw), je16_to_cpu(node->u.nodetype));
 616                         goto bail;
 617                 }
 618         }
 619
 620         /* OK, all the CRCs are good; this node can just be copied as-is. */
 621  retry:
 622         phys_ofs = write_ofs(c);
 623
 624         ret = jffs2_flash_write(c, phys_ofs, rawlen, &retlen, (char *)node);
 625
 626         if (ret || (retlen != rawlen)) {
 627                 printk(KERN_NOTICE "Write of %d bytes at 0x%08x failed. returned %d, retlen %zd\n",
 628                        rawlen, phys_ofs, ret, retlen);
 629                 if (retlen) {
 630                         jffs2_add_physical_node_ref(c, phys_ofs | REF_OBSOLETE, rawlen, NULL);
 631                 } else {
 632                         printk(KERN_NOTICE "Not marking the space at 0x%08x as dirty because the flash driver returned retlen zero\n", phys_ofs);
 633                 }
 634                 if (!retried) {
 635                         /* Try to reallocate space and retry */
 636                         uint32_t dummy;
 637                         struct jffs2_eraseblock *jeb = &c->blocks[phys_ofs / c->sector_size];
 638
 639                         retried = 1;
 640
 641                         D1(printk(KERN_DEBUG "Retrying failed write of REF_PRISTINE node.\n"));
 642
 643                         jffs2_dbg_acct_sanity_check(c,jeb);
 644                         jffs2_dbg_acct_paranoia_check(c, jeb);
 645
 646                         ret = jffs2_reserve_space_gc(c, rawlen, &dummy, rawlen);
 647                                                 /* this is not the exact summary size of it,
 648                                                         it is only an upper estimation */
 649
 650                         if (!ret) {
 651                                 D1(printk(KERN_DEBUG "Allocated space at 0x%08x to retry failed write.\n", phys_ofs));
 652
 653                                 jffs2_dbg_acct_sanity_check(c,jeb);
 654                                 jffs2_dbg_acct_paranoia_check(c, jeb);
 655
 656                                 goto retry;
 657                         }
 658                         D1(printk(KERN_DEBUG "Failed to allocate space to retry failed write: %d!\n", ret));
 659                 }
 660
 661                 if (!ret)
 662                         ret = -EIO;
 663                 goto out_node;
 664         }
 665         jffs2_add_physical_node_ref(c, phys_ofs | REF_PRISTINE, rawlen, ic);
 666
 667         jffs2_mark_node_obsolete(c, raw);
 668         D1(printk(KERN_DEBUG "WHEEE! GC REF_PRISTINE node at 0x%08x succeeded\n", ref_offset(raw)));
 669
 670  out_node:
 671         kfree(node);
 672         return ret;
 673  bail:
 674         ret = -EBADFD;
 675         goto out_node;
 676 }
 677
 678 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 679                                         struct jffs2_inode_info *f, struct jffs2_full_dnode *fn)
 680 {
 681         struct jffs2_full_dnode *new_fn;
 682         struct jffs2_raw_inode ri;
 683         struct jffs2_node_frag *last_frag;
 684         union jffs2_device_node dev;
 685         char *mdata = NULL, mdatalen = 0;
 686         uint32_t alloclen, ilen;
 687         int ret;
 688
 689         if (S_ISBLK(JFFS2_F_I_MODE(f)) ||
 690             S_ISCHR(JFFS2_F_I_MODE(f)) ) {
 691                 /* For these, we don't actually need to read the old node */
 692                 mdatalen = jffs2_encode_dev(&dev, JFFS2_F_I_RDEV(f));
 693                 mdata = (char *)&dev;
 694                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bytes of kdev_t\n", mdatalen));
 695         } else if (S_ISLNK(JFFS2_F_I_MODE(f))) {
 696                 mdatalen = fn->size;
 697                 mdata = kmalloc(fn->size, GFP_KERNEL);
 698                 if (!mdata) {
 699                         printk(KERN_WARNING "kmalloc of mdata failed in jffs2_garbage_collect_metadata()\n");
 700                         return -ENOMEM;
 701                 }
 702                 ret = jffs2_read_dnode(c, f, fn, mdata, 0, mdatalen);
 703                 if (ret) {
 704                         printk(KERN_WARNING "read of old metadata failed in jffs2_garbage_collect_metadata(): %d\n", ret);
 705                         kfree(mdata);
 706                         return ret;
 707                 }
 708                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bites of symlink target\n", mdatalen));
 709
 710         }
 711
 712         ret = jffs2_reserve_space_gc(c, sizeof(ri) + mdatalen, &alloclen,
 713                                 JFFS2_SUMMARY_INODE_SIZE);
 714         if (ret) {
 715                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_metadata failed: %d\n",
 716                        sizeof(ri)+ mdatalen, ret);
 717                 goto out;
 718         }
 719
 720         last_frag = frag_last(&f->fragtree);
 721         if (last_frag)
 722                 /* Fetch the inode length from the fragtree rather then
 723                  * from i_size since i_size may have not been updated yet */
 724                 ilen = last_frag->ofs + last_frag->size;
 725         else
 726                 ilen = JFFS2_F_I_SIZE(f);
 727
 728         memset(&ri, 0, sizeof(ri));
 729         ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
 730         ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
 731         ri.totlen = cpu_to_je32(sizeof(ri) + mdatalen);
 732         ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
 733
 734         ri.ino = cpu_to_je32(f->inocache->ino);
 735         ri.version = cpu_to_je32(++f->highest_version);
 736         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
 737         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
 738         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
 739         ri.isize = cpu_to_je32(ilen);
 740         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
 741         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
 742         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
 743         ri.offset = cpu_to_je32(0);
 744         ri.csize = cpu_to_je32(mdatalen);
 745         ri.dsize = cpu_to_je32(mdatalen);
 746         ri.compr = JFFS2_COMPR_NONE;
 747         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
 748         ri.data_crc = cpu_to_je32(crc32(0, mdata, mdatalen));
 749
 750         new_fn = jffs2_write_dnode(c, f, &ri, mdata, mdatalen, ALLOC_GC);
 751
 752         if (IS_ERR(new_fn)) {
 753                 printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
 754                 ret = PTR_ERR(new_fn);
 755                 goto out;
 756         }
 757         jffs2_mark_node_obsolete(c, fn->raw);
 758         jffs2_free_full_dnode(fn);
 759         f->metadata = new_fn;
 760  out:
 761         if (S_ISLNK(JFFS2_F_I_MODE(f)))
 762                 kfree(mdata);
 763         return ret;
 764 }
 765
 766 static int jffs2_garbage_collect_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 767                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
 768 {
 769         struct jffs2_full_dirent *new_fd;
 770         struct jffs2_raw_dirent rd;
 771         uint32_t alloclen;
 772         int ret;
 773
 774         rd.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
 775         rd.nodetype = cpu_to_je16(JFFS2_NODETYPE_DIRENT);
 776         rd.nsize = strlen(fd->name);
 777         rd.totlen = cpu_to_je32(sizeof(rd) + rd.nsize);
 778         rd.hdr_crc = cpu_to_je32(crc32(0, &rd, sizeof(struct jffs2_unknown_node)-4));
 779
 780         rd.pino = cpu_to_je32(f->inocache->ino);
 781         rd.version = cpu_to_je32(++f->highest_version);
 782         rd.ino = cpu_to_je32(fd->ino);
 783         /* If the times on this inode were set by explicit utime() they can be different,
 784            so refrain from splatting them. */
 785         if (JFFS2_F_I_MTIME(f) == JFFS2_F_I_CTIME(f))
 786                 rd.mctime = cpu_to_je32(JFFS2_F_I_MTIME(f));
 787         else
 788                 rd.mctime = cpu_to_je32(0);
 789         rd.type = fd->type;
 790         rd.node_crc = cpu_to_je32(crc32(0, &rd, sizeof(rd)-8));
 791         rd.name_crc = cpu_to_je32(crc32(0, fd->name, rd.nsize));
 792
 793         ret = jffs2_reserve_space_gc(c, sizeof(rd)+rd.nsize, &alloclen,
 794                                 JFFS2_SUMMARY_DIRENT_SIZE(rd.nsize));
 795         if (ret) {
 796                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dirent failed: %d\n",
 797                        sizeof(rd)+rd.nsize, ret);
 798                 return ret;
 799         }
 800         new_fd = jffs2_write_dirent(c, f, &rd, fd->name, rd.nsize, ALLOC_GC);
 801
 802         if (IS_ERR(new_fd)) {
 803                 printk(KERN_WARNING "jffs2_write_dirent in garbage_collect_dirent failed: %ld\n", PTR_ERR(new_fd));
 804                 return PTR_ERR(new_fd);
 805         }
 806         jffs2_add_fd_to_list(c, new_fd, &f->dents);
 807         return 0;
 808 }
 809
 810 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 811                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
 812 {
 813         struct jffs2_full_dirent **fdp = &f->dents;
 814         int found = 0;
 815
 816         /* On a medium where we can't actually mark nodes obsolete
 817            pernamently, such as NAND flash, we need to work out
 818            whether this deletion dirent is still needed to actively
 819            delete a 'real' dirent with the same name that's still
 820            somewhere else on the flash. */
 821         if (!jffs2_can_mark_obsolete(c)) {
 822                 struct jffs2_raw_dirent *rd;
 823                 struct jffs2_raw_node_ref *raw;
 824                 int ret;
 825                 size_t retlen;
 826                 int name_len = strlen(fd->name);
 827                 uint32_t name_crc = crc32(0, fd->name, name_len);
 828                 uint32_t rawlen = ref_totlen(c, jeb, fd->raw);
 829
 830                 rd = kmalloc(rawlen, GFP_KERNEL);
 831                 if (!rd)
 832                         return -ENOMEM;
 833
 834                 /* Prevent the erase code from nicking the obsolete node refs while
 835                    we're looking at them. I really don't like this extra lock but
 836                    can't see any alternative. Suggestions on a postcard to... */
 837                 down(&c->erase_free_sem);
 838
 839                 for (raw = f->inocache->nodes; raw != (void *)f->inocache; raw = raw->next_in_ino) {
 840
 841                         cond_resched();
 842
 843                         /* We only care about obsolete ones */
 844                         if (!(ref_obsolete(raw)))
 845                                 continue;
 846
 847                         /* Any dirent with the same name is going to have the same length... */
 848                         if (ref_totlen(c, NULL, raw) != rawlen)
 849                                 continue;
 850
 851                         /* Doesn't matter if there's one in the same erase block. We're going to
 852                            delete it too at the same time. */
 853                         if (SECTOR_ADDR(raw->flash_offset) == SECTOR_ADDR(fd->raw->flash_offset))
 854                                 continue;
 855
 856                         D1(printk(KERN_DEBUG "Check potential deletion dirent at %08x\n", ref_offset(raw)));
 857
 858                         /* This is an obsolete node belonging to the same directory, and it's of the right
 859                            length. We need to take a closer look...*/
 860                         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)rd);
 861                         if (ret) {
 862                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Read error (%d) reading obsolete node at %08x\n", ret, ref_offset(raw));
 863                                 /* If we can't read it, we don't need to continue to obsolete it. Continue */
 864                                 continue;
 865                         }
 866                         if (retlen != rawlen) {
 867                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Short read (%zd not %u) reading header from obsolete node at %08x\n",
 868                                        retlen, rawlen, ref_offset(raw));
 869                                 continue;
 870                         }
 871
 872                         if (je16_to_cpu(rd->nodetype) != JFFS2_NODETYPE_DIRENT)
 873                                 continue;
 874
 875                         /* If the name CRC doesn't match, skip */
 876                         if (je32_to_cpu(rd->name_crc) != name_crc)
 877                                 continue;
 878
 879                         /* If the name length doesn't match, or it's another deletion dirent, skip */
 880                         if (rd->nsize != name_len || !je32_to_cpu(rd->ino))
 881                                 continue;
 882
 883                         /* OK, check the actual name now */
 884                         if (memcmp(rd->name, fd->name, name_len))
 885                                 continue;
 886
 887                         /* OK. The name really does match. There really is still an older node on
 888                            the flash which our deletion dirent obsoletes. So we have to write out
 889                            a new deletion dirent to replace it */
 890                         up(&c->erase_free_sem);
 891
 892                         D1(printk(KERN_DEBUG "Deletion dirent at %08x still obsoletes real dirent \"%s\" at %08x for ino #%u\n",
 893                                   ref_offset(fd->raw), fd->name, ref_offset(raw), je32_to_cpu(rd->ino)));
 894                         kfree(rd);
 895
 896                         return jffs2_garbage_collect_dirent(c, jeb, f, fd);
 897                 }
 898
 899                 up(&c->erase_free_sem);
 900                 kfree(rd);
 901         }
 902
 903         /* FIXME: If we're deleting a dirent which contains the current mtime and ctime,
 904            we should update the metadata node with those times accordingly */
 905
 906         /* No need for it any more. Just mark it obsolete and remove it from the list */
 907         while (*fdp) {
 908                 if ((*fdp) == fd) {
 909                         found = 1;
 910                         *fdp = fd->next;
 911                         break;
 912                 }
 913                 fdp = &(*fdp)->next;
 914         }
 915         if (!found) {
 916                 printk(KERN_WARNING "Deletion dirent \"%s\" not found in list for ino #%u\n", fd->name, f->inocache->ino);
 917         }
 918         jffs2_mark_node_obsolete(c, fd->raw);
 919         jffs2_free_full_dirent(fd);
 920         return 0;
 921 }
 922
 923 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 924                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
 925                                       uint32_t start, uint32_t end)
 926 {
 927         struct jffs2_raw_inode ri;
 928         struct jffs2_node_frag *frag;
 929         struct jffs2_full_dnode *new_fn;
 930         uint32_t alloclen, ilen;
 931         int ret;
 932
 933         D1(printk(KERN_DEBUG "Writing replacement hole node for ino #%u from offset 0x%x to 0x%x\n",
 934                   f->inocache->ino, start, end));
 935
 936         memset(&ri, 0, sizeof(ri));
 937
 938         if(fn->frags > 1) {
 939                 size_t readlen;
 940                 uint32_t crc;
 941                 /* It's partially obsoleted by a later write. So we have to
 942                    write it out again with the _same_ version as before */
 943                 ret = jffs2_flash_read(c, ref_offset(fn->raw), sizeof(ri), &readlen, (char *)&ri);
 944                 if (readlen != sizeof(ri) || ret) {
 945                         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);
 946                         goto fill;
 947                 }
 948                 if (je16_to_cpu(ri.nodetype) != JFFS2_NODETYPE_INODE) {
 949                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had node type 0x%04x instead of JFFS2_NODETYPE_INODE(0x%04x)\n",
 950                                ref_offset(fn->raw),
 951                                je16_to_cpu(ri.nodetype), JFFS2_NODETYPE_INODE);
 952                         return -EIO;
 953                 }
 954                 if (je32_to_cpu(ri.totlen) != sizeof(ri)) {
 955                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had totlen 0x%x instead of expected 0x%zx\n",
 956                                ref_offset(fn->raw),
 957                                je32_to_cpu(ri.totlen), sizeof(ri));
 958                         return -EIO;
 959                 }
 960                 crc = crc32(0, &ri, sizeof(ri)-8);
 961                 if (crc != je32_to_cpu(ri.node_crc)) {
 962                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had CRC 0x%08x which doesn't match calculated CRC 0x%08x\n",
 963                                ref_offset(fn->raw),
 964                                je32_to_cpu(ri.node_crc), crc);
 965                         /* FIXME: We could possibly deal with this by writing new holes for each frag */
 966                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
 967                                start, end, f->inocache->ino);
 968                         goto fill;
 969                 }
 970                 if (ri.compr != JFFS2_COMPR_ZERO) {
 971                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node 0x%08x wasn't a hole node!\n", ref_offset(fn->raw));
 972                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
 973                                start, end, f->inocache->ino);
 974                         goto fill;
 975                 }
 976         } else {
 977         fill:
 978                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
 979                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
 980                 ri.totlen = cpu_to_je32(sizeof(ri));
 981                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
 982
 983                 ri.ino = cpu_to_je32(f->inocache->ino);
 984                 ri.version = cpu_to_je32(++f->highest_version);
 985                 ri.offset = cpu_to_je32(start);
 986                 ri.dsize = cpu_to_je32(end - start);
 987                 ri.csize = cpu_to_je32(0);
 988                 ri.compr = JFFS2_COMPR_ZERO;
 989         }
 990
 991         frag = frag_last(&f->fragtree);
 992         if (frag)
 993                 /* Fetch the inode length from the fragtree rather then
 994                  * from i_size since i_size may have not been updated yet */
 995                 ilen = frag->ofs + frag->size;
 996         else
 997                 ilen = JFFS2_F_I_SIZE(f);
 998
 999         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1000         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1001         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1002         ri.isize = cpu_to_je32(ilen);
1003         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1004         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1005         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1006         ri.data_crc = cpu_to_je32(0);
1007         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1008
1009         ret = jffs2_reserve_space_gc(c, sizeof(ri), &alloclen,
1010                                      JFFS2_SUMMARY_INODE_SIZE);
1011         if (ret) {
1012                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_hole failed: %d\n",
1013                        sizeof(ri), ret);
1014                 return ret;
1015         }
1016         new_fn = jffs2_write_dnode(c, f, &ri, NULL, 0, ALLOC_GC);
1017
1018         if (IS_ERR(new_fn)) {
1019                 printk(KERN_WARNING "Error writing new hole node: %ld\n", PTR_ERR(new_fn));
1020                 return PTR_ERR(new_fn);
1021         }
1022         if (je32_to_cpu(ri.version) == f->highest_version) {
1023                 jffs2_add_full_dnode_to_inode(c, f, new_fn);
1024                 if (f->metadata) {
1025                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1026                         jffs2_free_full_dnode(f->metadata);
1027                         f->metadata = NULL;
1028                 }
1029                 return 0;
1030         }
1031
1032         /*
1033          * We should only get here in the case where the node we are
1034          * replacing had more than one frag, so we kept the same version
1035          * number as before. (Except in case of error -- see 'goto fill;'
1036          * above.)
1037          */
1038         D1(if(unlikely(fn->frags <= 1)) {
1039                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Replacing fn with %d frag(s) but new ver %d != highest_version %d of ino #%d\n",
1040                        fn->frags, je32_to_cpu(ri.version), f->highest_version,
1041                        je32_to_cpu(ri.ino));
1042         });
1043
1044         /* This is a partially-overlapped hole node. Mark it REF_NORMAL not REF_PRISTINE */
1045         mark_ref_normal(new_fn->raw);
1046
1047         for (frag = jffs2_lookup_node_frag(&f->fragtree, fn->ofs);
1048              frag; frag = frag_next(frag)) {
1049                 if (frag->ofs > fn->size + fn->ofs)
1050                         break;
1051                 if (frag->node == fn) {
1052                         frag->node = new_fn;
1053                         new_fn->frags++;
1054                         fn->frags--;
1055                 }
1056         }
1057         if (fn->frags) {
1058                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Old node still has frags!\n");
1059                 BUG();
1060         }
1061         if (!new_fn->frags) {
1062                 printk(KERN_WARNING "jffs2_garbage_collect_hole: New node has no frags!\n");
1063                 BUG();
1064         }
1065
1066         jffs2_mark_node_obsolete(c, fn->raw);
1067         jffs2_free_full_dnode(fn);
1068
1069         return 0;
1070 }
1071
1072 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
1073                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
1074                                        uint32_t start, uint32_t end)
1075 {
1076         struct jffs2_full_dnode *new_fn;
1077         struct jffs2_raw_inode ri;
1078         uint32_t alloclen, offset, orig_end, orig_start;
1079         int ret = 0;
1080         unsigned char *comprbuf = NULL, *writebuf;
1081         unsigned long pg;
1082         unsigned char *pg_ptr;
1083
1084         memset(&ri, 0, sizeof(ri));
1085
1086         D1(printk(KERN_DEBUG "Writing replacement dnode for ino #%u from offset 0x%x to 0x%x\n",
1087                   f->inocache->ino, start, end));
1088
1089         orig_end = end;
1090         orig_start = start;
1091
1092         if (c->nr_free_blocks + c->nr_erasing_blocks > c->resv_blocks_gcmerge) {
1093                 /* Attempt to do some merging. But only expand to cover logically
1094                    adjacent frags if the block containing them is already considered
1095                    to be dirty. Otherwise we end up with GC just going round in
1096                    circles dirtying the nodes it already wrote out, especially
1097                    on NAND where we have small eraseblocks and hence a much higher
1098                    chance of nodes having to be split to cross boundaries. */
1099
1100                 struct jffs2_node_frag *frag;
1101                 uint32_t min, max;
1102
1103                 min = start & ~(PAGE_CACHE_SIZE-1);
1104                 max = min + PAGE_CACHE_SIZE;
1105
1106                 frag = jffs2_lookup_node_frag(&f->fragtree, start);
1107
1108                 /* BUG_ON(!frag) but that'll happen anyway... */
1109
1110                 BUG_ON(frag->ofs != start);
1111
1112                 /* First grow down... */
1113                 while((frag = frag_prev(frag)) && frag->ofs >= min) {
1114
1115                         /* If the previous frag doesn't even reach the beginning, there's
1116                            excessive fragmentation. Just merge. */
1117                         if (frag->ofs > min) {
1118                                 D1(printk(KERN_DEBUG "Expanding down to cover partial frag (0x%x-0x%x)\n",
1119                                           frag->ofs, frag->ofs+frag->size));
1120                                 start = frag->ofs;
1121                                 continue;
1122                         }
1123                         /* OK. This frag holds the first byte of the page. */
1124                         if (!frag->node || !frag->node->raw) {
1125                                 D1(printk(KERN_DEBUG "First frag in page is hole (0x%x-0x%x). Not expanding down.\n",
1126                                           frag->ofs, frag->ofs+frag->size));
1127                                 break;
1128                         } else {
1129
1130                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1131                                    in a block which is still considered clean? If so, don't obsolete it.
1132                                    If not, cover it anyway. */
1133
1134                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1135                                 struct jffs2_eraseblock *jeb;
1136
1137                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1138
1139                                 if (jeb == c->gcblock) {
1140                                         D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1141                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1142                                         start = frag->ofs;
1143                                         break;
1144                                 }
1145                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1146                                         D1(printk(KERN_DEBUG "Not expanding down to cover frag (0x%x-0x%x) in clean block %08x\n",
1147                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1148                                         break;
1149                                 }
1150
1151                                 D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in dirty block %08x\n",
1152                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1153                                 start = frag->ofs;
1154                                 break;
1155                         }
1156                 }
1157
1158                 /* ... then up */
1159
1160                 /* Find last frag which is actually part of the node we're to GC. */
1161                 frag = jffs2_lookup_node_frag(&f->fragtree, end-1);
1162
1163                 while((frag = frag_next(frag)) && frag->ofs+frag->size <= max) {
1164
1165                         /* If the previous frag doesn't even reach the beginning, there's lots
1166                            of fragmentation. Just merge. */
1167                         if (frag->ofs+frag->size < max) {
1168                                 D1(printk(KERN_DEBUG "Expanding up to cover partial frag (0x%x-0x%x)\n",
1169                                           frag->ofs, frag->ofs+frag->size));
1170                                 end = frag->ofs + frag->size;
1171                                 continue;
1172                         }
1173
1174                         if (!frag->node || !frag->node->raw) {
1175                                 D1(printk(KERN_DEBUG "Last frag in page is hole (0x%x-0x%x). Not expanding up.\n",
1176                                           frag->ofs, frag->ofs+frag->size));
1177                                 break;
1178                         } else {
1179
1180                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1181                                    in a block which is still considered clean? If so, don't obsolete it.
1182                                    If not, cover it anyway. */
1183
1184                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1185                                 struct jffs2_eraseblock *jeb;
1186
1187                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1188
1189                                 if (jeb == c->gcblock) {
1190                                         D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1191                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1192                                         end = frag->ofs + frag->size;
1193                                         break;
1194                                 }
1195                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1196                                         D1(printk(KERN_DEBUG "Not expanding up to cover frag (0x%x-0x%x) in clean block %08x\n",
1197                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1198                                         break;
1199                                 }
1200
1201                                 D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in dirty block %08x\n",
1202                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1203                                 end = frag->ofs + frag->size;
1204                                 break;
1205                         }
1206                 }
1207                 D1(printk(KERN_DEBUG "Expanded dnode to write from (0x%x-0x%x) to (0x%x-0x%x)\n",
1208                           orig_start, orig_end, start, end));
1209
1210                 D1(BUG_ON(end > frag_last(&f->fragtree)->ofs + frag_last(&f->fragtree)->size));
1211                 BUG_ON(end < orig_end);
1212                 BUG_ON(start > orig_start);
1213         }
1214
1215         /* First, use readpage() to read the appropriate page into the page cache */
1216         /* Q: What happens if we actually try to GC the _same_ page for which commit_write()
1217          *    triggered garbage collection in the first place?
1218          * A: I _think_ it's OK. read_cache_page shouldn't deadlock, we'll write out the
1219          *    page OK. We'll actually write it out again in commit_write, which is a little
1220          *    suboptimal, but at least we're correct.
1221          */
1222         pg_ptr = jffs2_gc_fetch_page(c, f, start, &pg);
1223
1224         if (IS_ERR(pg_ptr)) {
1225                 printk(KERN_WARNING "read_cache_page() returned error: %ld\n", PTR_ERR(pg_ptr));
1226                 return PTR_ERR(pg_ptr);
1227         }
1228
1229         offset = start;
1230         while(offset < orig_end) {
1231                 uint32_t datalen;
1232                 uint32_t cdatalen;
1233                 uint16_t comprtype = JFFS2_COMPR_NONE;
1234
1235                 ret = jffs2_reserve_space_gc(c, sizeof(ri) + JFFS2_MIN_DATA_LEN,
1236                                         &alloclen, JFFS2_SUMMARY_INODE_SIZE);
1237
1238                 if (ret) {
1239                         printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dnode failed: %d\n",
1240                                sizeof(ri)+ JFFS2_MIN_DATA_LEN, ret);
1241                         break;
1242                 }
1243                 cdatalen = min_t(uint32_t, alloclen - sizeof(ri), end - offset);
1244                 datalen = end - offset;
1245
1246                 writebuf = pg_ptr + (offset & (PAGE_CACHE_SIZE -1));
1247
1248                 comprtype = jffs2_compress(c, f, writebuf, &comprbuf, &datalen, &cdatalen);
1249
1250                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
1251                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
1252                 ri.totlen = cpu_to_je32(sizeof(ri) + cdatalen);
1253                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
1254
1255                 ri.ino = cpu_to_je32(f->inocache->ino);
1256                 ri.version = cpu_to_je32(++f->highest_version);
1257                 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1258                 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1259                 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1260                 ri.isize = cpu_to_je32(JFFS2_F_I_SIZE(f));
1261                 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1262                 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1263                 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1264                 ri.offset = cpu_to_je32(offset);
1265                 ri.csize = cpu_to_je32(cdatalen);
1266                 ri.dsize = cpu_to_je32(datalen);
1267                 ri.compr = comprtype & 0xff;
1268                 ri.usercompr = (comprtype >> 8) & 0xff;
1269                 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1270                 ri.data_crc = cpu_to_je32(crc32(0, comprbuf, cdatalen));
1271
1272                 new_fn = jffs2_write_dnode(c, f, &ri, comprbuf, cdatalen, ALLOC_GC);
1273
1274                 jffs2_free_comprbuf(comprbuf, writebuf);
1275
1276                 if (IS_ERR(new_fn)) {
1277                         printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
1278                         ret = PTR_ERR(new_fn);
1279                         break;
1280                 }
1281                 ret = jffs2_add_full_dnode_to_inode(c, f, new_fn);
1282                 offset += datalen;
1283                 if (f->metadata) {
1284                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1285                         jffs2_free_full_dnode(f->metadata);
1286                         f->metadata = NULL;
1287                 }
1288         }
1289
1290         jffs2_gc_release_page(c, pg_ptr, &pg);
1291         return ret;
1292 }