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 = raw->next_phys;
 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                 BUG_ON(raw->next_in_ino != (void *)ic);
 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);
 283                 } else {
 284                         ret = jffs2_garbage_collect_xattr_ref(c, (struct jffs2_xattr_ref *)ic);
 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         struct jffs2_raw_node_ref *nraw;
 532         size_t retlen;
 533         int ret;
 534         uint32_t phys_ofs, alloclen;
 535         uint32_t crc, rawlen;
 536         int retried = 0;
 537
 538         D1(printk(KERN_DEBUG "Going to GC REF_PRISTINE node at 0x%08x\n", ref_offset(raw)));
 539
 540         alloclen = rawlen = ref_totlen(c, c->gcblock, raw);
 541
 542         /* Ask for a small amount of space (or the totlen if smaller) because we
 543            don't want to force wastage of the end of a block if splitting would
 544            work. */
 545         if (ic && alloclen > sizeof(struct jffs2_raw_inode) + JFFS2_MIN_DATA_LEN)
 546                 alloclen = sizeof(struct jffs2_raw_inode) + JFFS2_MIN_DATA_LEN;
 547
 548         ret = jffs2_reserve_space_gc(c, alloclen, &alloclen, rawlen);
 549         /* 'rawlen' is not the exact summary size; it is only an upper estimation */
 550
 551         if (ret)
 552                 return ret;
 553
 554         if (alloclen < rawlen) {
 555                 /* Doesn't fit untouched. We'll go the old route and split it */
 556                 return -EBADFD;
 557         }
 558
 559         node = kmalloc(rawlen, GFP_KERNEL);
 560         if (!node)
 561                return -ENOMEM;
 562
 563         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)node);
 564         if (!ret && retlen != rawlen)
 565                 ret = -EIO;
 566         if (ret)
 567                 goto out_node;
 568
 569         crc = crc32(0, node, sizeof(struct jffs2_unknown_node)-4);
 570         if (je32_to_cpu(node->u.hdr_crc) != crc) {
 571                 printk(KERN_WARNING "Header CRC failed on REF_PRISTINE node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 572                        ref_offset(raw), je32_to_cpu(node->u.hdr_crc), crc);
 573                 goto bail;
 574         }
 575
 576         switch(je16_to_cpu(node->u.nodetype)) {
 577         case JFFS2_NODETYPE_INODE:
 578                 crc = crc32(0, node, sizeof(node->i)-8);
 579                 if (je32_to_cpu(node->i.node_crc) != crc) {
 580                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 581                                ref_offset(raw), je32_to_cpu(node->i.node_crc), crc);
 582                         goto bail;
 583                 }
 584
 585                 if (je32_to_cpu(node->i.dsize)) {
 586                         crc = crc32(0, node->i.data, je32_to_cpu(node->i.csize));
 587                         if (je32_to_cpu(node->i.data_crc) != crc) {
 588                                 printk(KERN_WARNING "Data CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 589                                        ref_offset(raw), je32_to_cpu(node->i.data_crc), crc);
 590                                 goto bail;
 591                         }
 592                 }
 593                 break;
 594
 595         case JFFS2_NODETYPE_DIRENT:
 596                 crc = crc32(0, node, sizeof(node->d)-8);
 597                 if (je32_to_cpu(node->d.node_crc) != crc) {
 598                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE dirent node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 599                                ref_offset(raw), je32_to_cpu(node->d.node_crc), crc);
 600                         goto bail;
 601                 }
 602
 603                 if (node->d.nsize) {
 604                         crc = crc32(0, node->d.name, node->d.nsize);
 605                         if (je32_to_cpu(node->d.name_crc) != crc) {
 606                                 printk(KERN_WARNING "Name CRC failed on REF_PRISTINE dirent ode at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
 607                                        ref_offset(raw), je32_to_cpu(node->d.name_crc), crc);
 608                                 goto bail;
 609                         }
 610                 }
 611                 break;
 612         default:
 613                 /* If it's inode-less, we don't _know_ what it is. Just copy it intact */
 614                 if (ic) {
 615                         printk(KERN_WARNING "Unknown node type for REF_PRISTINE node at 0x%08x: 0x%04x\n",
 616                                ref_offset(raw), je16_to_cpu(node->u.nodetype));
 617                         goto bail;
 618                 }
 619         }
 620
 621         nraw = jffs2_alloc_raw_node_ref();
 622         if (!nraw) {
 623                 ret = -ENOMEM;
 624                 goto out_node;
 625         }
 626
 627         /* OK, all the CRCs are good; this node can just be copied as-is. */
 628  retry:
 629         nraw->flash_offset = phys_ofs = write_ofs(c);
 630
 631         ret = jffs2_flash_write(c, phys_ofs, rawlen, &retlen, (char *)node);
 632
 633         if (ret || (retlen != rawlen)) {
 634                 printk(KERN_NOTICE "Write of %d bytes at 0x%08x failed. returned %d, retlen %zd\n",
 635                        rawlen, nraw->flash_offset, ret, retlen);
 636                 if (retlen) {
 637                         nraw->flash_offset |= REF_OBSOLETE;
 638                         jffs2_add_physical_node_ref(c, nraw, rawlen, NULL);
 639                         jffs2_mark_node_obsolete(c, nraw);
 640                 } else {
 641                         printk(KERN_NOTICE "Not marking the space at 0x%08x as dirty because the flash driver returned retlen zero\n", nraw->flash_offset);
 642                         jffs2_free_raw_node_ref(nraw);
 643                 }
 644                 if (!retried && (nraw = jffs2_alloc_raw_node_ref())) {
 645                         /* Try to reallocate space and retry */
 646                         uint32_t dummy;
 647                         struct jffs2_eraseblock *jeb = &c->blocks[phys_ofs / c->sector_size];
 648
 649                         retried = 1;
 650
 651                         D1(printk(KERN_DEBUG "Retrying failed write of REF_PRISTINE node.\n"));
 652
 653                         jffs2_dbg_acct_sanity_check(c,jeb);
 654                         jffs2_dbg_acct_paranoia_check(c, jeb);
 655
 656                         ret = jffs2_reserve_space_gc(c, rawlen, &dummy, rawlen);
 657                                                 /* this is not the exact summary size of it,
 658                                                         it is only an upper estimation */
 659
 660                         if (!ret) {
 661                                 D1(printk(KERN_DEBUG "Allocated space at 0x%08x to retry failed write.\n", phys_ofs));
 662
 663                                 jffs2_dbg_acct_sanity_check(c,jeb);
 664                                 jffs2_dbg_acct_paranoia_check(c, jeb);
 665
 666                                 goto retry;
 667                         }
 668                         D1(printk(KERN_DEBUG "Failed to allocate space to retry failed write: %d!\n", ret));
 669                         jffs2_free_raw_node_ref(nraw);
 670                 }
 671
 672                 jffs2_free_raw_node_ref(nraw);
 673                 if (!ret)
 674                         ret = -EIO;
 675                 goto out_node;
 676         }
 677         nraw->flash_offset |= REF_PRISTINE;
 678         jffs2_add_physical_node_ref(c, nraw, rawlen, ic);
 679
 680         jffs2_mark_node_obsolete(c, raw);
 681         D1(printk(KERN_DEBUG "WHEEE! GC REF_PRISTINE node at 0x%08x succeeded\n", ref_offset(raw)));
 682
 683  out_node:
 684         kfree(node);
 685         return ret;
 686  bail:
 687         ret = -EBADFD;
 688         goto out_node;
 689 }
 690
 691 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 692                                         struct jffs2_inode_info *f, struct jffs2_full_dnode *fn)
 693 {
 694         struct jffs2_full_dnode *new_fn;
 695         struct jffs2_raw_inode ri;
 696         struct jffs2_node_frag *last_frag;
 697         union jffs2_device_node dev;
 698         char *mdata = NULL, mdatalen = 0;
 699         uint32_t alloclen, ilen;
 700         int ret;
 701
 702         if (S_ISBLK(JFFS2_F_I_MODE(f)) ||
 703             S_ISCHR(JFFS2_F_I_MODE(f)) ) {
 704                 /* For these, we don't actually need to read the old node */
 705                 mdatalen = jffs2_encode_dev(&dev, JFFS2_F_I_RDEV(f));
 706                 mdata = (char *)&dev;
 707                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bytes of kdev_t\n", mdatalen));
 708         } else if (S_ISLNK(JFFS2_F_I_MODE(f))) {
 709                 mdatalen = fn->size;
 710                 mdata = kmalloc(fn->size, GFP_KERNEL);
 711                 if (!mdata) {
 712                         printk(KERN_WARNING "kmalloc of mdata failed in jffs2_garbage_collect_metadata()\n");
 713                         return -ENOMEM;
 714                 }
 715                 ret = jffs2_read_dnode(c, f, fn, mdata, 0, mdatalen);
 716                 if (ret) {
 717                         printk(KERN_WARNING "read of old metadata failed in jffs2_garbage_collect_metadata(): %d\n", ret);
 718                         kfree(mdata);
 719                         return ret;
 720                 }
 721                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bites of symlink target\n", mdatalen));
 722
 723         }
 724
 725         ret = jffs2_reserve_space_gc(c, sizeof(ri) + mdatalen, &alloclen,
 726                                 JFFS2_SUMMARY_INODE_SIZE);
 727         if (ret) {
 728                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_metadata failed: %d\n",
 729                        sizeof(ri)+ mdatalen, ret);
 730                 goto out;
 731         }
 732
 733         last_frag = frag_last(&f->fragtree);
 734         if (last_frag)
 735                 /* Fetch the inode length from the fragtree rather then
 736                  * from i_size since i_size may have not been updated yet */
 737                 ilen = last_frag->ofs + last_frag->size;
 738         else
 739                 ilen = JFFS2_F_I_SIZE(f);
 740
 741         memset(&ri, 0, sizeof(ri));
 742         ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
 743         ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
 744         ri.totlen = cpu_to_je32(sizeof(ri) + mdatalen);
 745         ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
 746
 747         ri.ino = cpu_to_je32(f->inocache->ino);
 748         ri.version = cpu_to_je32(++f->highest_version);
 749         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
 750         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
 751         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
 752         ri.isize = cpu_to_je32(ilen);
 753         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
 754         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
 755         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
 756         ri.offset = cpu_to_je32(0);
 757         ri.csize = cpu_to_je32(mdatalen);
 758         ri.dsize = cpu_to_je32(mdatalen);
 759         ri.compr = JFFS2_COMPR_NONE;
 760         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
 761         ri.data_crc = cpu_to_je32(crc32(0, mdata, mdatalen));
 762
 763         new_fn = jffs2_write_dnode(c, f, &ri, mdata, mdatalen, ALLOC_GC);
 764
 765         if (IS_ERR(new_fn)) {
 766                 printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
 767                 ret = PTR_ERR(new_fn);
 768                 goto out;
 769         }
 770         jffs2_mark_node_obsolete(c, fn->raw);
 771         jffs2_free_full_dnode(fn);
 772         f->metadata = new_fn;
 773  out:
 774         if (S_ISLNK(JFFS2_F_I_MODE(f)))
 775                 kfree(mdata);
 776         return ret;
 777 }
 778
 779 static int jffs2_garbage_collect_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 780                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
 781 {
 782         struct jffs2_full_dirent *new_fd;
 783         struct jffs2_raw_dirent rd;
 784         uint32_t alloclen;
 785         int ret;
 786
 787         rd.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
 788         rd.nodetype = cpu_to_je16(JFFS2_NODETYPE_DIRENT);
 789         rd.nsize = strlen(fd->name);
 790         rd.totlen = cpu_to_je32(sizeof(rd) + rd.nsize);
 791         rd.hdr_crc = cpu_to_je32(crc32(0, &rd, sizeof(struct jffs2_unknown_node)-4));
 792
 793         rd.pino = cpu_to_je32(f->inocache->ino);
 794         rd.version = cpu_to_je32(++f->highest_version);
 795         rd.ino = cpu_to_je32(fd->ino);
 796         /* If the times on this inode were set by explicit utime() they can be different,
 797            so refrain from splatting them. */
 798         if (JFFS2_F_I_MTIME(f) == JFFS2_F_I_CTIME(f))
 799                 rd.mctime = cpu_to_je32(JFFS2_F_I_MTIME(f));
 800         else
 801                 rd.mctime = cpu_to_je32(0);
 802         rd.type = fd->type;
 803         rd.node_crc = cpu_to_je32(crc32(0, &rd, sizeof(rd)-8));
 804         rd.name_crc = cpu_to_je32(crc32(0, fd->name, rd.nsize));
 805
 806         ret = jffs2_reserve_space_gc(c, sizeof(rd)+rd.nsize, &alloclen,
 807                                 JFFS2_SUMMARY_DIRENT_SIZE(rd.nsize));
 808         if (ret) {
 809                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dirent failed: %d\n",
 810                        sizeof(rd)+rd.nsize, ret);
 811                 return ret;
 812         }
 813         new_fd = jffs2_write_dirent(c, f, &rd, fd->name, rd.nsize, ALLOC_GC);
 814
 815         if (IS_ERR(new_fd)) {
 816                 printk(KERN_WARNING "jffs2_write_dirent in garbage_collect_dirent failed: %ld\n", PTR_ERR(new_fd));
 817                 return PTR_ERR(new_fd);
 818         }
 819         jffs2_add_fd_to_list(c, new_fd, &f->dents);
 820         return 0;
 821 }
 822
 823 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 824                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
 825 {
 826         struct jffs2_full_dirent **fdp = &f->dents;
 827         int found = 0;
 828
 829         /* On a medium where we can't actually mark nodes obsolete
 830            pernamently, such as NAND flash, we need to work out
 831            whether this deletion dirent is still needed to actively
 832            delete a 'real' dirent with the same name that's still
 833            somewhere else on the flash. */
 834         if (!jffs2_can_mark_obsolete(c)) {
 835                 struct jffs2_raw_dirent *rd;
 836                 struct jffs2_raw_node_ref *raw;
 837                 int ret;
 838                 size_t retlen;
 839                 int name_len = strlen(fd->name);
 840                 uint32_t name_crc = crc32(0, fd->name, name_len);
 841                 uint32_t rawlen = ref_totlen(c, jeb, fd->raw);
 842
 843                 rd = kmalloc(rawlen, GFP_KERNEL);
 844                 if (!rd)
 845                         return -ENOMEM;
 846
 847                 /* Prevent the erase code from nicking the obsolete node refs while
 848                    we're looking at them. I really don't like this extra lock but
 849                    can't see any alternative. Suggestions on a postcard to... */
 850                 down(&c->erase_free_sem);
 851
 852                 for (raw = f->inocache->nodes; raw != (void *)f->inocache; raw = raw->next_in_ino) {
 853
 854                         /* We only care about obsolete ones */
 855                         if (!(ref_obsolete(raw)))
 856                                 continue;
 857
 858                         /* Any dirent with the same name is going to have the same length... */
 859                         if (ref_totlen(c, NULL, raw) != rawlen)
 860                                 continue;
 861
 862                         /* Doesn't matter if there's one in the same erase block. We're going to
 863                            delete it too at the same time. */
 864                         if (SECTOR_ADDR(raw->flash_offset) == SECTOR_ADDR(fd->raw->flash_offset))
 865                                 continue;
 866
 867                         D1(printk(KERN_DEBUG "Check potential deletion dirent at %08x\n", ref_offset(raw)));
 868
 869                         /* This is an obsolete node belonging to the same directory, and it's of the right
 870                            length. We need to take a closer look...*/
 871                         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)rd);
 872                         if (ret) {
 873                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Read error (%d) reading obsolete node at %08x\n", ret, ref_offset(raw));
 874                                 /* If we can't read it, we don't need to continue to obsolete it. Continue */
 875                                 continue;
 876                         }
 877                         if (retlen != rawlen) {
 878                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Short read (%zd not %u) reading header from obsolete node at %08x\n",
 879                                        retlen, rawlen, ref_offset(raw));
 880                                 continue;
 881                         }
 882
 883                         if (je16_to_cpu(rd->nodetype) != JFFS2_NODETYPE_DIRENT)
 884                                 continue;
 885
 886                         /* If the name CRC doesn't match, skip */
 887                         if (je32_to_cpu(rd->name_crc) != name_crc)
 888                                 continue;
 889
 890                         /* If the name length doesn't match, or it's another deletion dirent, skip */
 891                         if (rd->nsize != name_len || !je32_to_cpu(rd->ino))
 892                                 continue;
 893
 894                         /* OK, check the actual name now */
 895                         if (memcmp(rd->name, fd->name, name_len))
 896                                 continue;
 897
 898                         /* OK. The name really does match. There really is still an older node on
 899                            the flash which our deletion dirent obsoletes. So we have to write out
 900                            a new deletion dirent to replace it */
 901                         up(&c->erase_free_sem);
 902
 903                         D1(printk(KERN_DEBUG "Deletion dirent at %08x still obsoletes real dirent \"%s\" at %08x for ino #%u\n",
 904                                   ref_offset(fd->raw), fd->name, ref_offset(raw), je32_to_cpu(rd->ino)));
 905                         kfree(rd);
 906
 907                         return jffs2_garbage_collect_dirent(c, jeb, f, fd);
 908                 }
 909
 910                 up(&c->erase_free_sem);
 911                 kfree(rd);
 912         }
 913
 914         /* FIXME: If we're deleting a dirent which contains the current mtime and ctime,
 915            we should update the metadata node with those times accordingly */
 916
 917         /* No need for it any more. Just mark it obsolete and remove it from the list */
 918         while (*fdp) {
 919                 if ((*fdp) == fd) {
 920                         found = 1;
 921                         *fdp = fd->next;
 922                         break;
 923                 }
 924                 fdp = &(*fdp)->next;
 925         }
 926         if (!found) {
 927                 printk(KERN_WARNING "Deletion dirent \"%s\" not found in list for ino #%u\n", fd->name, f->inocache->ino);
 928         }
 929         jffs2_mark_node_obsolete(c, fd->raw);
 930         jffs2_free_full_dirent(fd);
 931         return 0;
 932 }
 933
 934 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
 935                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
 936                                       uint32_t start, uint32_t end)
 937 {
 938         struct jffs2_raw_inode ri;
 939         struct jffs2_node_frag *frag;
 940         struct jffs2_full_dnode *new_fn;
 941         uint32_t alloclen, ilen;
 942         int ret;
 943
 944         D1(printk(KERN_DEBUG "Writing replacement hole node for ino #%u from offset 0x%x to 0x%x\n",
 945                   f->inocache->ino, start, end));
 946
 947         memset(&ri, 0, sizeof(ri));
 948
 949         if(fn->frags > 1) {
 950                 size_t readlen;
 951                 uint32_t crc;
 952                 /* It's partially obsoleted by a later write. So we have to
 953                    write it out again with the _same_ version as before */
 954                 ret = jffs2_flash_read(c, ref_offset(fn->raw), sizeof(ri), &readlen, (char *)&ri);
 955                 if (readlen != sizeof(ri) || ret) {
 956                         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);
 957                         goto fill;
 958                 }
 959                 if (je16_to_cpu(ri.nodetype) != JFFS2_NODETYPE_INODE) {
 960                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had node type 0x%04x instead of JFFS2_NODETYPE_INODE(0x%04x)\n",
 961                                ref_offset(fn->raw),
 962                                je16_to_cpu(ri.nodetype), JFFS2_NODETYPE_INODE);
 963                         return -EIO;
 964                 }
 965                 if (je32_to_cpu(ri.totlen) != sizeof(ri)) {
 966                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had totlen 0x%x instead of expected 0x%zx\n",
 967                                ref_offset(fn->raw),
 968                                je32_to_cpu(ri.totlen), sizeof(ri));
 969                         return -EIO;
 970                 }
 971                 crc = crc32(0, &ri, sizeof(ri)-8);
 972                 if (crc != je32_to_cpu(ri.node_crc)) {
 973                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had CRC 0x%08x which doesn't match calculated CRC 0x%08x\n",
 974                                ref_offset(fn->raw),
 975                                je32_to_cpu(ri.node_crc), crc);
 976                         /* FIXME: We could possibly deal with this by writing new holes for each frag */
 977                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
 978                                start, end, f->inocache->ino);
 979                         goto fill;
 980                 }
 981                 if (ri.compr != JFFS2_COMPR_ZERO) {
 982                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node 0x%08x wasn't a hole node!\n", ref_offset(fn->raw));
 983                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
 984                                start, end, f->inocache->ino);
 985                         goto fill;
 986                 }
 987         } else {
 988         fill:
 989                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
 990                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
 991                 ri.totlen = cpu_to_je32(sizeof(ri));
 992                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
 993
 994                 ri.ino = cpu_to_je32(f->inocache->ino);
 995                 ri.version = cpu_to_je32(++f->highest_version);
 996                 ri.offset = cpu_to_je32(start);
 997                 ri.dsize = cpu_to_je32(end - start);
 998                 ri.csize = cpu_to_je32(0);
 999                 ri.compr = JFFS2_COMPR_ZERO;
1000         }
1001
1002         frag = frag_last(&f->fragtree);
1003         if (frag)
1004                 /* Fetch the inode length from the fragtree rather then
1005                  * from i_size since i_size may have not been updated yet */
1006                 ilen = frag->ofs + frag->size;
1007         else
1008                 ilen = JFFS2_F_I_SIZE(f);
1009
1010         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1011         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1012         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1013         ri.isize = cpu_to_je32(ilen);
1014         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1015         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1016         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1017         ri.data_crc = cpu_to_je32(0);
1018         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1019
1020         ret = jffs2_reserve_space_gc(c, sizeof(ri), &alloclen,
1021                                      JFFS2_SUMMARY_INODE_SIZE);
1022         if (ret) {
1023                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_hole failed: %d\n",
1024                        sizeof(ri), ret);
1025                 return ret;
1026         }
1027         new_fn = jffs2_write_dnode(c, f, &ri, NULL, 0, ALLOC_GC);
1028
1029         if (IS_ERR(new_fn)) {
1030                 printk(KERN_WARNING "Error writing new hole node: %ld\n", PTR_ERR(new_fn));
1031                 return PTR_ERR(new_fn);
1032         }
1033         if (je32_to_cpu(ri.version) == f->highest_version) {
1034                 jffs2_add_full_dnode_to_inode(c, f, new_fn);
1035                 if (f->metadata) {
1036                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1037                         jffs2_free_full_dnode(f->metadata);
1038                         f->metadata = NULL;
1039                 }
1040                 return 0;
1041         }
1042
1043         /*
1044          * We should only get here in the case where the node we are
1045          * replacing had more than one frag, so we kept the same version
1046          * number as before. (Except in case of error -- see 'goto fill;'
1047          * above.)
1048          */
1049         D1(if(unlikely(fn->frags <= 1)) {
1050                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Replacing fn with %d frag(s) but new ver %d != highest_version %d of ino #%d\n",
1051                        fn->frags, je32_to_cpu(ri.version), f->highest_version,
1052                        je32_to_cpu(ri.ino));
1053         });
1054
1055         /* This is a partially-overlapped hole node. Mark it REF_NORMAL not REF_PRISTINE */
1056         mark_ref_normal(new_fn->raw);
1057
1058         for (frag = jffs2_lookup_node_frag(&f->fragtree, fn->ofs);
1059              frag; frag = frag_next(frag)) {
1060                 if (frag->ofs > fn->size + fn->ofs)
1061                         break;
1062                 if (frag->node == fn) {
1063                         frag->node = new_fn;
1064                         new_fn->frags++;
1065                         fn->frags--;
1066                 }
1067         }
1068         if (fn->frags) {
1069                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Old node still has frags!\n");
1070                 BUG();
1071         }
1072         if (!new_fn->frags) {
1073                 printk(KERN_WARNING "jffs2_garbage_collect_hole: New node has no frags!\n");
1074                 BUG();
1075         }
1076
1077         jffs2_mark_node_obsolete(c, fn->raw);
1078         jffs2_free_full_dnode(fn);
1079
1080         return 0;
1081 }
1082
1083 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
1084                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
1085                                        uint32_t start, uint32_t end)
1086 {
1087         struct jffs2_full_dnode *new_fn;
1088         struct jffs2_raw_inode ri;
1089         uint32_t alloclen, offset, orig_end, orig_start;
1090         int ret = 0;
1091         unsigned char *comprbuf = NULL, *writebuf;
1092         unsigned long pg;
1093         unsigned char *pg_ptr;
1094
1095         memset(&ri, 0, sizeof(ri));
1096
1097         D1(printk(KERN_DEBUG "Writing replacement dnode for ino #%u from offset 0x%x to 0x%x\n",
1098                   f->inocache->ino, start, end));
1099
1100         orig_end = end;
1101         orig_start = start;
1102
1103         if (c->nr_free_blocks + c->nr_erasing_blocks > c->resv_blocks_gcmerge) {
1104                 /* Attempt to do some merging. But only expand to cover logically
1105                    adjacent frags if the block containing them is already considered
1106                    to be dirty. Otherwise we end up with GC just going round in
1107                    circles dirtying the nodes it already wrote out, especially
1108                    on NAND where we have small eraseblocks and hence a much higher
1109                    chance of nodes having to be split to cross boundaries. */
1110
1111                 struct jffs2_node_frag *frag;
1112                 uint32_t min, max;
1113
1114                 min = start & ~(PAGE_CACHE_SIZE-1);
1115                 max = min + PAGE_CACHE_SIZE;
1116
1117                 frag = jffs2_lookup_node_frag(&f->fragtree, start);
1118
1119                 /* BUG_ON(!frag) but that'll happen anyway... */
1120
1121                 BUG_ON(frag->ofs != start);
1122
1123                 /* First grow down... */
1124                 while((frag = frag_prev(frag)) && frag->ofs >= min) {
1125
1126                         /* If the previous frag doesn't even reach the beginning, there's
1127                            excessive fragmentation. Just merge. */
1128                         if (frag->ofs > min) {
1129                                 D1(printk(KERN_DEBUG "Expanding down to cover partial frag (0x%x-0x%x)\n",
1130                                           frag->ofs, frag->ofs+frag->size));
1131                                 start = frag->ofs;
1132                                 continue;
1133                         }
1134                         /* OK. This frag holds the first byte of the page. */
1135                         if (!frag->node || !frag->node->raw) {
1136                                 D1(printk(KERN_DEBUG "First frag in page is hole (0x%x-0x%x). Not expanding down.\n",
1137                                           frag->ofs, frag->ofs+frag->size));
1138                                 break;
1139                         } else {
1140
1141                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1142                                    in a block which is still considered clean? If so, don't obsolete it.
1143                                    If not, cover it anyway. */
1144
1145                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1146                                 struct jffs2_eraseblock *jeb;
1147
1148                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1149
1150                                 if (jeb == c->gcblock) {
1151                                         D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1152                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1153                                         start = frag->ofs;
1154                                         break;
1155                                 }
1156                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1157                                         D1(printk(KERN_DEBUG "Not expanding down to cover frag (0x%x-0x%x) in clean block %08x\n",
1158                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1159                                         break;
1160                                 }
1161
1162                                 D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in dirty block %08x\n",
1163                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1164                                 start = frag->ofs;
1165                                 break;
1166                         }
1167                 }
1168
1169                 /* ... then up */
1170
1171                 /* Find last frag which is actually part of the node we're to GC. */
1172                 frag = jffs2_lookup_node_frag(&f->fragtree, end-1);
1173
1174                 while((frag = frag_next(frag)) && frag->ofs+frag->size <= max) {
1175
1176                         /* If the previous frag doesn't even reach the beginning, there's lots
1177                            of fragmentation. Just merge. */
1178                         if (frag->ofs+frag->size < max) {
1179                                 D1(printk(KERN_DEBUG "Expanding up to cover partial frag (0x%x-0x%x)\n",
1180                                           frag->ofs, frag->ofs+frag->size));
1181                                 end = frag->ofs + frag->size;
1182                                 continue;
1183                         }
1184
1185                         if (!frag->node || !frag->node->raw) {
1186                                 D1(printk(KERN_DEBUG "Last frag in page is hole (0x%x-0x%x). Not expanding up.\n",
1187                                           frag->ofs, frag->ofs+frag->size));
1188                                 break;
1189                         } else {
1190
1191                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1192                                    in a block which is still considered clean? If so, don't obsolete it.
1193                                    If not, cover it anyway. */
1194
1195                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1196                                 struct jffs2_eraseblock *jeb;
1197
1198                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1199
1200                                 if (jeb == c->gcblock) {
1201                                         D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1202                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1203                                         end = frag->ofs + frag->size;
1204                                         break;
1205                                 }
1206                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1207                                         D1(printk(KERN_DEBUG "Not expanding up to cover frag (0x%x-0x%x) in clean block %08x\n",
1208                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1209                                         break;
1210                                 }
1211
1212                                 D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in dirty block %08x\n",
1213                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1214                                 end = frag->ofs + frag->size;
1215                                 break;
1216                         }
1217                 }
1218                 D1(printk(KERN_DEBUG "Expanded dnode to write from (0x%x-0x%x) to (0x%x-0x%x)\n",
1219                           orig_start, orig_end, start, end));
1220
1221                 D1(BUG_ON(end > frag_last(&f->fragtree)->ofs + frag_last(&f->fragtree)->size));
1222                 BUG_ON(end < orig_end);
1223                 BUG_ON(start > orig_start);
1224         }
1225
1226         /* First, use readpage() to read the appropriate page into the page cache */
1227         /* Q: What happens if we actually try to GC the _same_ page for which commit_write()
1228          *    triggered garbage collection in the first place?
1229          * A: I _think_ it's OK. read_cache_page shouldn't deadlock, we'll write out the
1230          *    page OK. We'll actually write it out again in commit_write, which is a little
1231          *    suboptimal, but at least we're correct.
1232          */
1233         pg_ptr = jffs2_gc_fetch_page(c, f, start, &pg);
1234
1235         if (IS_ERR(pg_ptr)) {
1236                 printk(KERN_WARNING "read_cache_page() returned error: %ld\n", PTR_ERR(pg_ptr));
1237                 return PTR_ERR(pg_ptr);
1238         }
1239
1240         offset = start;
1241         while(offset < orig_end) {
1242                 uint32_t datalen;
1243                 uint32_t cdatalen;
1244                 uint16_t comprtype = JFFS2_COMPR_NONE;
1245
1246                 ret = jffs2_reserve_space_gc(c, sizeof(ri) + JFFS2_MIN_DATA_LEN,
1247                                         &alloclen, JFFS2_SUMMARY_INODE_SIZE);
1248
1249                 if (ret) {
1250                         printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dnode failed: %d\n",
1251                                sizeof(ri)+ JFFS2_MIN_DATA_LEN, ret);
1252                         break;
1253                 }
1254                 cdatalen = min_t(uint32_t, alloclen - sizeof(ri), end - offset);
1255                 datalen = end - offset;
1256
1257                 writebuf = pg_ptr + (offset & (PAGE_CACHE_SIZE -1));
1258
1259                 comprtype = jffs2_compress(c, f, writebuf, &comprbuf, &datalen, &cdatalen);
1260
1261                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
1262                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
1263                 ri.totlen = cpu_to_je32(sizeof(ri) + cdatalen);
1264                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
1265
1266                 ri.ino = cpu_to_je32(f->inocache->ino);
1267                 ri.version = cpu_to_je32(++f->highest_version);
1268                 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1269                 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1270                 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1271                 ri.isize = cpu_to_je32(JFFS2_F_I_SIZE(f));
1272                 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1273                 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1274                 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1275                 ri.offset = cpu_to_je32(offset);
1276                 ri.csize = cpu_to_je32(cdatalen);
1277                 ri.dsize = cpu_to_je32(datalen);
1278                 ri.compr = comprtype & 0xff;
1279                 ri.usercompr = (comprtype >> 8) & 0xff;
1280                 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1281                 ri.data_crc = cpu_to_je32(crc32(0, comprbuf, cdatalen));
1282
1283                 new_fn = jffs2_write_dnode(c, f, &ri, comprbuf, cdatalen, ALLOC_GC);
1284
1285                 jffs2_free_comprbuf(comprbuf, writebuf);
1286
1287                 if (IS_ERR(new_fn)) {
1288                         printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
1289                         ret = PTR_ERR(new_fn);
1290                         break;
1291                 }
1292                 ret = jffs2_add_full_dnode_to_inode(c, f, new_fn);
1293                 offset += datalen;
1294                 if (f->metadata) {
1295                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1296                         jffs2_free_full_dnode(f->metadata);
1297                         f->metadata = NULL;
1298                 }
1299         }
1300
1301         jffs2_gc_release_page(c, pg_ptr, &pg);
1302         return ret;
1303 }