]> Pileus Git - ~andy/linux/blob - fs/jffs2/gc.c
[JFFS2] Correct handling of JFFS2_FEATURE_RWCOMPAT_COPY nodes.
[~andy/linux] / 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, &phys_ofs, &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;
630         nraw->__totlen = rawlen;
631         nraw->next_phys = NULL;
632
633         ret = jffs2_flash_write(c, phys_ofs, rawlen, &retlen, (char *)node);
634
635         if (ret || (retlen != rawlen)) {
636                 printk(KERN_NOTICE "Write of %d bytes at 0x%08x failed. returned %d, retlen %zd\n",
637                        rawlen, phys_ofs, ret, retlen);
638                 if (retlen) {
639                         /* Doesn't belong to any inode */
640                         nraw->next_in_ino = NULL;
641
642                         nraw->flash_offset |= REF_OBSOLETE;
643                         jffs2_add_physical_node_ref(c, nraw);
644                         jffs2_mark_node_obsolete(c, nraw);
645                 } else {
646                         printk(KERN_NOTICE "Not marking the space at 0x%08x as dirty because the flash driver returned retlen zero\n", nraw->flash_offset);
647                         jffs2_free_raw_node_ref(nraw);
648                 }
649                 if (!retried && (nraw = jffs2_alloc_raw_node_ref())) {
650                         /* Try to reallocate space and retry */
651                         uint32_t dummy;
652                         struct jffs2_eraseblock *jeb = &c->blocks[phys_ofs / c->sector_size];
653
654                         retried = 1;
655
656                         D1(printk(KERN_DEBUG "Retrying failed write of REF_PRISTINE node.\n"));
657
658                         jffs2_dbg_acct_sanity_check(c,jeb);
659                         jffs2_dbg_acct_paranoia_check(c, jeb);
660
661                         ret = jffs2_reserve_space_gc(c, rawlen, &phys_ofs, &dummy, rawlen);
662                                                 /* this is not the exact summary size of it,
663                                                         it is only an upper estimation */
664
665                         if (!ret) {
666                                 D1(printk(KERN_DEBUG "Allocated space at 0x%08x to retry failed write.\n", phys_ofs));
667
668                                 jffs2_dbg_acct_sanity_check(c,jeb);
669                                 jffs2_dbg_acct_paranoia_check(c, jeb);
670
671                                 goto retry;
672                         }
673                         D1(printk(KERN_DEBUG "Failed to allocate space to retry failed write: %d!\n", ret));
674                         jffs2_free_raw_node_ref(nraw);
675                 }
676
677                 jffs2_free_raw_node_ref(nraw);
678                 if (!ret)
679                         ret = -EIO;
680                 goto out_node;
681         }
682         nraw->flash_offset |= REF_PRISTINE;
683         jffs2_add_physical_node_ref(c, nraw);
684
685         if (ic) {
686                 /* Link into per-inode list. This is safe because of the ic
687                    state being INO_STATE_GC. Note that if we're doing this
688                    for an inode which is in-core, the 'nraw' pointer is then
689                    going to be fetched from ic->nodes by our caller. */
690                 spin_lock(&c->erase_completion_lock);
691                 nraw->next_in_ino = ic->nodes;
692                 ic->nodes = nraw;
693                 spin_unlock(&c->erase_completion_lock);
694         }
695         jffs2_mark_node_obsolete(c, raw);
696         D1(printk(KERN_DEBUG "WHEEE! GC REF_PRISTINE node at 0x%08x succeeded\n", ref_offset(raw)));
697
698  out_node:
699         kfree(node);
700         return ret;
701  bail:
702         ret = -EBADFD;
703         goto out_node;
704 }
705
706 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
707                                         struct jffs2_inode_info *f, struct jffs2_full_dnode *fn)
708 {
709         struct jffs2_full_dnode *new_fn;
710         struct jffs2_raw_inode ri;
711         struct jffs2_node_frag *last_frag;
712         union jffs2_device_node dev;
713         char *mdata = NULL, mdatalen = 0;
714         uint32_t alloclen, phys_ofs, ilen;
715         int ret;
716
717         if (S_ISBLK(JFFS2_F_I_MODE(f)) ||
718             S_ISCHR(JFFS2_F_I_MODE(f)) ) {
719                 /* For these, we don't actually need to read the old node */
720                 mdatalen = jffs2_encode_dev(&dev, JFFS2_F_I_RDEV(f));
721                 mdata = (char *)&dev;
722                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bytes of kdev_t\n", mdatalen));
723         } else if (S_ISLNK(JFFS2_F_I_MODE(f))) {
724                 mdatalen = fn->size;
725                 mdata = kmalloc(fn->size, GFP_KERNEL);
726                 if (!mdata) {
727                         printk(KERN_WARNING "kmalloc of mdata failed in jffs2_garbage_collect_metadata()\n");
728                         return -ENOMEM;
729                 }
730                 ret = jffs2_read_dnode(c, f, fn, mdata, 0, mdatalen);
731                 if (ret) {
732                         printk(KERN_WARNING "read of old metadata failed in jffs2_garbage_collect_metadata(): %d\n", ret);
733                         kfree(mdata);
734                         return ret;
735                 }
736                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bites of symlink target\n", mdatalen));
737
738         }
739
740         ret = jffs2_reserve_space_gc(c, sizeof(ri) + mdatalen, &phys_ofs, &alloclen,
741                                 JFFS2_SUMMARY_INODE_SIZE);
742         if (ret) {
743                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_metadata failed: %d\n",
744                        sizeof(ri)+ mdatalen, ret);
745                 goto out;
746         }
747
748         last_frag = frag_last(&f->fragtree);
749         if (last_frag)
750                 /* Fetch the inode length from the fragtree rather then
751                  * from i_size since i_size may have not been updated yet */
752                 ilen = last_frag->ofs + last_frag->size;
753         else
754                 ilen = JFFS2_F_I_SIZE(f);
755
756         memset(&ri, 0, sizeof(ri));
757         ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
758         ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
759         ri.totlen = cpu_to_je32(sizeof(ri) + mdatalen);
760         ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
761
762         ri.ino = cpu_to_je32(f->inocache->ino);
763         ri.version = cpu_to_je32(++f->highest_version);
764         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
765         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
766         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
767         ri.isize = cpu_to_je32(ilen);
768         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
769         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
770         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
771         ri.offset = cpu_to_je32(0);
772         ri.csize = cpu_to_je32(mdatalen);
773         ri.dsize = cpu_to_je32(mdatalen);
774         ri.compr = JFFS2_COMPR_NONE;
775         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
776         ri.data_crc = cpu_to_je32(crc32(0, mdata, mdatalen));
777
778         new_fn = jffs2_write_dnode(c, f, &ri, mdata, mdatalen, phys_ofs, ALLOC_GC);
779
780         if (IS_ERR(new_fn)) {
781                 printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
782                 ret = PTR_ERR(new_fn);
783                 goto out;
784         }
785         jffs2_mark_node_obsolete(c, fn->raw);
786         jffs2_free_full_dnode(fn);
787         f->metadata = new_fn;
788  out:
789         if (S_ISLNK(JFFS2_F_I_MODE(f)))
790                 kfree(mdata);
791         return ret;
792 }
793
794 static int jffs2_garbage_collect_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
795                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
796 {
797         struct jffs2_full_dirent *new_fd;
798         struct jffs2_raw_dirent rd;
799         uint32_t alloclen, phys_ofs;
800         int ret;
801
802         rd.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
803         rd.nodetype = cpu_to_je16(JFFS2_NODETYPE_DIRENT);
804         rd.nsize = strlen(fd->name);
805         rd.totlen = cpu_to_je32(sizeof(rd) + rd.nsize);
806         rd.hdr_crc = cpu_to_je32(crc32(0, &rd, sizeof(struct jffs2_unknown_node)-4));
807
808         rd.pino = cpu_to_je32(f->inocache->ino);
809         rd.version = cpu_to_je32(++f->highest_version);
810         rd.ino = cpu_to_je32(fd->ino);
811         /* If the times on this inode were set by explicit utime() they can be different,
812            so refrain from splatting them. */
813         if (JFFS2_F_I_MTIME(f) == JFFS2_F_I_CTIME(f))
814                 rd.mctime = cpu_to_je32(JFFS2_F_I_MTIME(f));
815         else
816                 rd.mctime = cpu_to_je32(0);
817         rd.type = fd->type;
818         rd.node_crc = cpu_to_je32(crc32(0, &rd, sizeof(rd)-8));
819         rd.name_crc = cpu_to_je32(crc32(0, fd->name, rd.nsize));
820
821         ret = jffs2_reserve_space_gc(c, sizeof(rd)+rd.nsize, &phys_ofs, &alloclen,
822                                 JFFS2_SUMMARY_DIRENT_SIZE(rd.nsize));
823         if (ret) {
824                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dirent failed: %d\n",
825                        sizeof(rd)+rd.nsize, ret);
826                 return ret;
827         }
828         new_fd = jffs2_write_dirent(c, f, &rd, fd->name, rd.nsize, phys_ofs, ALLOC_GC);
829
830         if (IS_ERR(new_fd)) {
831                 printk(KERN_WARNING "jffs2_write_dirent in garbage_collect_dirent failed: %ld\n", PTR_ERR(new_fd));
832                 return PTR_ERR(new_fd);
833         }
834         jffs2_add_fd_to_list(c, new_fd, &f->dents);
835         return 0;
836 }
837
838 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
839                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
840 {
841         struct jffs2_full_dirent **fdp = &f->dents;
842         int found = 0;
843
844         /* On a medium where we can't actually mark nodes obsolete
845            pernamently, such as NAND flash, we need to work out
846            whether this deletion dirent is still needed to actively
847            delete a 'real' dirent with the same name that's still
848            somewhere else on the flash. */
849         if (!jffs2_can_mark_obsolete(c)) {
850                 struct jffs2_raw_dirent *rd;
851                 struct jffs2_raw_node_ref *raw;
852                 int ret;
853                 size_t retlen;
854                 int name_len = strlen(fd->name);
855                 uint32_t name_crc = crc32(0, fd->name, name_len);
856                 uint32_t rawlen = ref_totlen(c, jeb, fd->raw);
857
858                 rd = kmalloc(rawlen, GFP_KERNEL);
859                 if (!rd)
860                         return -ENOMEM;
861
862                 /* Prevent the erase code from nicking the obsolete node refs while
863                    we're looking at them. I really don't like this extra lock but
864                    can't see any alternative. Suggestions on a postcard to... */
865                 down(&c->erase_free_sem);
866
867                 for (raw = f->inocache->nodes; raw != (void *)f->inocache; raw = raw->next_in_ino) {
868
869                         /* We only care about obsolete ones */
870                         if (!(ref_obsolete(raw)))
871                                 continue;
872
873                         /* Any dirent with the same name is going to have the same length... */
874                         if (ref_totlen(c, NULL, raw) != rawlen)
875                                 continue;
876
877                         /* Doesn't matter if there's one in the same erase block. We're going to
878                            delete it too at the same time. */
879                         if (SECTOR_ADDR(raw->flash_offset) == SECTOR_ADDR(fd->raw->flash_offset))
880                                 continue;
881
882                         D1(printk(KERN_DEBUG "Check potential deletion dirent at %08x\n", ref_offset(raw)));
883
884                         /* This is an obsolete node belonging to the same directory, and it's of the right
885                            length. We need to take a closer look...*/
886                         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)rd);
887                         if (ret) {
888                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Read error (%d) reading obsolete node at %08x\n", ret, ref_offset(raw));
889                                 /* If we can't read it, we don't need to continue to obsolete it. Continue */
890                                 continue;
891                         }
892                         if (retlen != rawlen) {
893                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Short read (%zd not %u) reading header from obsolete node at %08x\n",
894                                        retlen, rawlen, ref_offset(raw));
895                                 continue;
896                         }
897
898                         if (je16_to_cpu(rd->nodetype) != JFFS2_NODETYPE_DIRENT)
899                                 continue;
900
901                         /* If the name CRC doesn't match, skip */
902                         if (je32_to_cpu(rd->name_crc) != name_crc)
903                                 continue;
904
905                         /* If the name length doesn't match, or it's another deletion dirent, skip */
906                         if (rd->nsize != name_len || !je32_to_cpu(rd->ino))
907                                 continue;
908
909                         /* OK, check the actual name now */
910                         if (memcmp(rd->name, fd->name, name_len))
911                                 continue;
912
913                         /* OK. The name really does match. There really is still an older node on
914                            the flash which our deletion dirent obsoletes. So we have to write out
915                            a new deletion dirent to replace it */
916                         up(&c->erase_free_sem);
917
918                         D1(printk(KERN_DEBUG "Deletion dirent at %08x still obsoletes real dirent \"%s\" at %08x for ino #%u\n",
919                                   ref_offset(fd->raw), fd->name, ref_offset(raw), je32_to_cpu(rd->ino)));
920                         kfree(rd);
921
922                         return jffs2_garbage_collect_dirent(c, jeb, f, fd);
923                 }
924
925                 up(&c->erase_free_sem);
926                 kfree(rd);
927         }
928
929         /* FIXME: If we're deleting a dirent which contains the current mtime and ctime,
930            we should update the metadata node with those times accordingly */
931
932         /* No need for it any more. Just mark it obsolete and remove it from the list */
933         while (*fdp) {
934                 if ((*fdp) == fd) {
935                         found = 1;
936                         *fdp = fd->next;
937                         break;
938                 }
939                 fdp = &(*fdp)->next;
940         }
941         if (!found) {
942                 printk(KERN_WARNING "Deletion dirent \"%s\" not found in list for ino #%u\n", fd->name, f->inocache->ino);
943         }
944         jffs2_mark_node_obsolete(c, fd->raw);
945         jffs2_free_full_dirent(fd);
946         return 0;
947 }
948
949 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
950                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
951                                       uint32_t start, uint32_t end)
952 {
953         struct jffs2_raw_inode ri;
954         struct jffs2_node_frag *frag;
955         struct jffs2_full_dnode *new_fn;
956         uint32_t alloclen, phys_ofs, ilen;
957         int ret;
958
959         D1(printk(KERN_DEBUG "Writing replacement hole node for ino #%u from offset 0x%x to 0x%x\n",
960                   f->inocache->ino, start, end));
961
962         memset(&ri, 0, sizeof(ri));
963
964         if(fn->frags > 1) {
965                 size_t readlen;
966                 uint32_t crc;
967                 /* It's partially obsoleted by a later write. So we have to
968                    write it out again with the _same_ version as before */
969                 ret = jffs2_flash_read(c, ref_offset(fn->raw), sizeof(ri), &readlen, (char *)&ri);
970                 if (readlen != sizeof(ri) || ret) {
971                         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);
972                         goto fill;
973                 }
974                 if (je16_to_cpu(ri.nodetype) != JFFS2_NODETYPE_INODE) {
975                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had node type 0x%04x instead of JFFS2_NODETYPE_INODE(0x%04x)\n",
976                                ref_offset(fn->raw),
977                                je16_to_cpu(ri.nodetype), JFFS2_NODETYPE_INODE);
978                         return -EIO;
979                 }
980                 if (je32_to_cpu(ri.totlen) != sizeof(ri)) {
981                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had totlen 0x%x instead of expected 0x%zx\n",
982                                ref_offset(fn->raw),
983                                je32_to_cpu(ri.totlen), sizeof(ri));
984                         return -EIO;
985                 }
986                 crc = crc32(0, &ri, sizeof(ri)-8);
987                 if (crc != je32_to_cpu(ri.node_crc)) {
988                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had CRC 0x%08x which doesn't match calculated CRC 0x%08x\n",
989                                ref_offset(fn->raw),
990                                je32_to_cpu(ri.node_crc), crc);
991                         /* FIXME: We could possibly deal with this by writing new holes for each frag */
992                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
993                                start, end, f->inocache->ino);
994                         goto fill;
995                 }
996                 if (ri.compr != JFFS2_COMPR_ZERO) {
997                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node 0x%08x wasn't a hole node!\n", ref_offset(fn->raw));
998                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
999                                start, end, f->inocache->ino);
1000                         goto fill;
1001                 }
1002         } else {
1003         fill:
1004                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
1005                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
1006                 ri.totlen = cpu_to_je32(sizeof(ri));
1007                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
1008
1009                 ri.ino = cpu_to_je32(f->inocache->ino);
1010                 ri.version = cpu_to_je32(++f->highest_version);
1011                 ri.offset = cpu_to_je32(start);
1012                 ri.dsize = cpu_to_je32(end - start);
1013                 ri.csize = cpu_to_je32(0);
1014                 ri.compr = JFFS2_COMPR_ZERO;
1015         }
1016
1017         frag = frag_last(&f->fragtree);
1018         if (frag)
1019                 /* Fetch the inode length from the fragtree rather then
1020                  * from i_size since i_size may have not been updated yet */
1021                 ilen = frag->ofs + frag->size;
1022         else
1023                 ilen = JFFS2_F_I_SIZE(f);
1024
1025         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1026         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1027         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1028         ri.isize = cpu_to_je32(ilen);
1029         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1030         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1031         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1032         ri.data_crc = cpu_to_je32(0);
1033         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1034
1035         ret = jffs2_reserve_space_gc(c, sizeof(ri), &phys_ofs, &alloclen,
1036                                 JFFS2_SUMMARY_INODE_SIZE);
1037         if (ret) {
1038                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_hole failed: %d\n",
1039                        sizeof(ri), ret);
1040                 return ret;
1041         }
1042         new_fn = jffs2_write_dnode(c, f, &ri, NULL, 0, phys_ofs, ALLOC_GC);
1043
1044         if (IS_ERR(new_fn)) {
1045                 printk(KERN_WARNING "Error writing new hole node: %ld\n", PTR_ERR(new_fn));
1046                 return PTR_ERR(new_fn);
1047         }
1048         if (je32_to_cpu(ri.version) == f->highest_version) {
1049                 jffs2_add_full_dnode_to_inode(c, f, new_fn);
1050                 if (f->metadata) {
1051                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1052                         jffs2_free_full_dnode(f->metadata);
1053                         f->metadata = NULL;
1054                 }
1055                 return 0;
1056         }
1057
1058         /*
1059          * We should only get here in the case where the node we are
1060          * replacing had more than one frag, so we kept the same version
1061          * number as before. (Except in case of error -- see 'goto fill;'
1062          * above.)
1063          */
1064         D1(if(unlikely(fn->frags <= 1)) {
1065                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Replacing fn with %d frag(s) but new ver %d != highest_version %d of ino #%d\n",
1066                        fn->frags, je32_to_cpu(ri.version), f->highest_version,
1067                        je32_to_cpu(ri.ino));
1068         });
1069
1070         /* This is a partially-overlapped hole node. Mark it REF_NORMAL not REF_PRISTINE */
1071         mark_ref_normal(new_fn->raw);
1072
1073         for (frag = jffs2_lookup_node_frag(&f->fragtree, fn->ofs);
1074              frag; frag = frag_next(frag)) {
1075                 if (frag->ofs > fn->size + fn->ofs)
1076                         break;
1077                 if (frag->node == fn) {
1078                         frag->node = new_fn;
1079                         new_fn->frags++;
1080                         fn->frags--;
1081                 }
1082         }
1083         if (fn->frags) {
1084                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Old node still has frags!\n");
1085                 BUG();
1086         }
1087         if (!new_fn->frags) {
1088                 printk(KERN_WARNING "jffs2_garbage_collect_hole: New node has no frags!\n");
1089                 BUG();
1090         }
1091
1092         jffs2_mark_node_obsolete(c, fn->raw);
1093         jffs2_free_full_dnode(fn);
1094
1095         return 0;
1096 }
1097
1098 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
1099                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
1100                                        uint32_t start, uint32_t end)
1101 {
1102         struct jffs2_full_dnode *new_fn;
1103         struct jffs2_raw_inode ri;
1104         uint32_t alloclen, phys_ofs, offset, orig_end, orig_start;
1105         int ret = 0;
1106         unsigned char *comprbuf = NULL, *writebuf;
1107         unsigned long pg;
1108         unsigned char *pg_ptr;
1109
1110         memset(&ri, 0, sizeof(ri));
1111
1112         D1(printk(KERN_DEBUG "Writing replacement dnode for ino #%u from offset 0x%x to 0x%x\n",
1113                   f->inocache->ino, start, end));
1114
1115         orig_end = end;
1116         orig_start = start;
1117
1118         if (c->nr_free_blocks + c->nr_erasing_blocks > c->resv_blocks_gcmerge) {
1119                 /* Attempt to do some merging. But only expand to cover logically
1120                    adjacent frags if the block containing them is already considered
1121                    to be dirty. Otherwise we end up with GC just going round in
1122                    circles dirtying the nodes it already wrote out, especially
1123                    on NAND where we have small eraseblocks and hence a much higher
1124                    chance of nodes having to be split to cross boundaries. */
1125
1126                 struct jffs2_node_frag *frag;
1127                 uint32_t min, max;
1128
1129                 min = start & ~(PAGE_CACHE_SIZE-1);
1130                 max = min + PAGE_CACHE_SIZE;
1131
1132                 frag = jffs2_lookup_node_frag(&f->fragtree, start);
1133
1134                 /* BUG_ON(!frag) but that'll happen anyway... */
1135
1136                 BUG_ON(frag->ofs != start);
1137
1138                 /* First grow down... */
1139                 while((frag = frag_prev(frag)) && frag->ofs >= min) {
1140
1141                         /* If the previous frag doesn't even reach the beginning, there's
1142                            excessive fragmentation. Just merge. */
1143                         if (frag->ofs > min) {
1144                                 D1(printk(KERN_DEBUG "Expanding down to cover partial frag (0x%x-0x%x)\n",
1145                                           frag->ofs, frag->ofs+frag->size));
1146                                 start = frag->ofs;
1147                                 continue;
1148                         }
1149                         /* OK. This frag holds the first byte of the page. */
1150                         if (!frag->node || !frag->node->raw) {
1151                                 D1(printk(KERN_DEBUG "First frag in page is hole (0x%x-0x%x). Not expanding down.\n",
1152                                           frag->ofs, frag->ofs+frag->size));
1153                                 break;
1154                         } else {
1155
1156                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1157                                    in a block which is still considered clean? If so, don't obsolete it.
1158                                    If not, cover it anyway. */
1159
1160                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1161                                 struct jffs2_eraseblock *jeb;
1162
1163                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1164
1165                                 if (jeb == c->gcblock) {
1166                                         D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1167                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1168                                         start = frag->ofs;
1169                                         break;
1170                                 }
1171                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1172                                         D1(printk(KERN_DEBUG "Not expanding down to cover frag (0x%x-0x%x) in clean block %08x\n",
1173                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1174                                         break;
1175                                 }
1176
1177                                 D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in dirty block %08x\n",
1178                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1179                                 start = frag->ofs;
1180                                 break;
1181                         }
1182                 }
1183
1184                 /* ... then up */
1185
1186                 /* Find last frag which is actually part of the node we're to GC. */
1187                 frag = jffs2_lookup_node_frag(&f->fragtree, end-1);
1188
1189                 while((frag = frag_next(frag)) && frag->ofs+frag->size <= max) {
1190
1191                         /* If the previous frag doesn't even reach the beginning, there's lots
1192                            of fragmentation. Just merge. */
1193                         if (frag->ofs+frag->size < max) {
1194                                 D1(printk(KERN_DEBUG "Expanding up to cover partial frag (0x%x-0x%x)\n",
1195                                           frag->ofs, frag->ofs+frag->size));
1196                                 end = frag->ofs + frag->size;
1197                                 continue;
1198                         }
1199
1200                         if (!frag->node || !frag->node->raw) {
1201                                 D1(printk(KERN_DEBUG "Last frag in page is hole (0x%x-0x%x). Not expanding up.\n",
1202                                           frag->ofs, frag->ofs+frag->size));
1203                                 break;
1204                         } else {
1205
1206                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1207                                    in a block which is still considered clean? If so, don't obsolete it.
1208                                    If not, cover it anyway. */
1209
1210                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1211                                 struct jffs2_eraseblock *jeb;
1212
1213                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1214
1215                                 if (jeb == c->gcblock) {
1216                                         D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1217                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1218                                         end = frag->ofs + frag->size;
1219                                         break;
1220                                 }
1221                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1222                                         D1(printk(KERN_DEBUG "Not expanding up to cover frag (0x%x-0x%x) in clean block %08x\n",
1223                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1224                                         break;
1225                                 }
1226
1227                                 D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in dirty block %08x\n",
1228                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1229                                 end = frag->ofs + frag->size;
1230                                 break;
1231                         }
1232                 }
1233                 D1(printk(KERN_DEBUG "Expanded dnode to write from (0x%x-0x%x) to (0x%x-0x%x)\n",
1234                           orig_start, orig_end, start, end));
1235
1236                 D1(BUG_ON(end > frag_last(&f->fragtree)->ofs + frag_last(&f->fragtree)->size));
1237                 BUG_ON(end < orig_end);
1238                 BUG_ON(start > orig_start);
1239         }
1240
1241         /* First, use readpage() to read the appropriate page into the page cache */
1242         /* Q: What happens if we actually try to GC the _same_ page for which commit_write()
1243          *    triggered garbage collection in the first place?
1244          * A: I _think_ it's OK. read_cache_page shouldn't deadlock, we'll write out the
1245          *    page OK. We'll actually write it out again in commit_write, which is a little
1246          *    suboptimal, but at least we're correct.
1247          */
1248         pg_ptr = jffs2_gc_fetch_page(c, f, start, &pg);
1249
1250         if (IS_ERR(pg_ptr)) {
1251                 printk(KERN_WARNING "read_cache_page() returned error: %ld\n", PTR_ERR(pg_ptr));
1252                 return PTR_ERR(pg_ptr);
1253         }
1254
1255         offset = start;
1256         while(offset < orig_end) {
1257                 uint32_t datalen;
1258                 uint32_t cdatalen;
1259                 uint16_t comprtype = JFFS2_COMPR_NONE;
1260
1261                 ret = jffs2_reserve_space_gc(c, sizeof(ri) + JFFS2_MIN_DATA_LEN, &phys_ofs,
1262                                         &alloclen, JFFS2_SUMMARY_INODE_SIZE);
1263
1264                 if (ret) {
1265                         printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dnode failed: %d\n",
1266                                sizeof(ri)+ JFFS2_MIN_DATA_LEN, ret);
1267                         break;
1268                 }
1269                 cdatalen = min_t(uint32_t, alloclen - sizeof(ri), end - offset);
1270                 datalen = end - offset;
1271
1272                 writebuf = pg_ptr + (offset & (PAGE_CACHE_SIZE -1));
1273
1274                 comprtype = jffs2_compress(c, f, writebuf, &comprbuf, &datalen, &cdatalen);
1275
1276                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
1277                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
1278                 ri.totlen = cpu_to_je32(sizeof(ri) + cdatalen);
1279                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
1280
1281                 ri.ino = cpu_to_je32(f->inocache->ino);
1282                 ri.version = cpu_to_je32(++f->highest_version);
1283                 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1284                 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1285                 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1286                 ri.isize = cpu_to_je32(JFFS2_F_I_SIZE(f));
1287                 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1288                 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1289                 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1290                 ri.offset = cpu_to_je32(offset);
1291                 ri.csize = cpu_to_je32(cdatalen);
1292                 ri.dsize = cpu_to_je32(datalen);
1293                 ri.compr = comprtype & 0xff;
1294                 ri.usercompr = (comprtype >> 8) & 0xff;
1295                 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1296                 ri.data_crc = cpu_to_je32(crc32(0, comprbuf, cdatalen));
1297
1298                 new_fn = jffs2_write_dnode(c, f, &ri, comprbuf, cdatalen, phys_ofs, ALLOC_GC);
1299
1300                 jffs2_free_comprbuf(comprbuf, writebuf);
1301
1302                 if (IS_ERR(new_fn)) {
1303                         printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
1304                         ret = PTR_ERR(new_fn);
1305                         break;
1306                 }
1307                 ret = jffs2_add_full_dnode_to_inode(c, f, new_fn);
1308                 offset += datalen;
1309                 if (f->metadata) {
1310                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1311                         jffs2_free_full_dnode(f->metadata);
1312                         f->metadata = NULL;
1313                 }
1314         }
1315
1316         jffs2_gc_release_page(c, pg_ptr, &pg);
1317         return ret;
1318 }