Merge git://git.infradead.org/mtd-2.6
[linux-2.6] / fs / jffs2 / gc.c
1 /*
2  * JFFS2 -- Journalling Flash File System, Version 2.
3  *
4  * Copyright © 2001-2007 Red Hat, Inc.
5  *
6  * Created by David Woodhouse <dwmw2@infradead.org>
7  *
8  * For licensing information, see the file 'LICENCE' in this directory.
9  *
10  */
11
12 #include <linux/kernel.h>
13 #include <linux/mtd/mtd.h>
14 #include <linux/slab.h>
15 #include <linux/pagemap.h>
16 #include <linux/crc32.h>
17 #include <linux/compiler.h>
18 #include <linux/stat.h>
19 #include "nodelist.h"
20 #include "compr.h"
21
22 static int jffs2_garbage_collect_pristine(struct jffs2_sb_info *c,
23                                           struct jffs2_inode_cache *ic,
24                                           struct jffs2_raw_node_ref *raw);
25 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
26                                         struct jffs2_inode_info *f, struct jffs2_full_dnode *fd);
27 static int jffs2_garbage_collect_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
28                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd);
29 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
30                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd);
31 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
32                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
33                                       uint32_t start, uint32_t end);
34 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
35                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
36                                        uint32_t start, uint32_t end);
37 static int jffs2_garbage_collect_live(struct jffs2_sb_info *c,  struct jffs2_eraseblock *jeb,
38                                struct jffs2_raw_node_ref *raw, struct jffs2_inode_info *f);
39
40 /* Called with erase_completion_lock held */
41 static struct jffs2_eraseblock *jffs2_find_gc_block(struct jffs2_sb_info *c)
42 {
43         struct jffs2_eraseblock *ret;
44         struct list_head *nextlist = NULL;
45         int n = jiffies % 128;
46
47         /* Pick an eraseblock to garbage collect next. This is where we'll
48            put the clever wear-levelling algorithms. Eventually.  */
49         /* We possibly want to favour the dirtier blocks more when the
50            number of free blocks is low. */
51 again:
52         if (!list_empty(&c->bad_used_list) && c->nr_free_blocks > c->resv_blocks_gcbad) {
53                 D1(printk(KERN_DEBUG "Picking block from bad_used_list to GC next\n"));
54                 nextlist = &c->bad_used_list;
55         } else if (n < 50 && !list_empty(&c->erasable_list)) {
56                 /* Note that most of them will have gone directly to be erased.
57                    So don't favour the erasable_list _too_ much. */
58                 D1(printk(KERN_DEBUG "Picking block from erasable_list to GC next\n"));
59                 nextlist = &c->erasable_list;
60         } else if (n < 110 && !list_empty(&c->very_dirty_list)) {
61                 /* Most of the time, pick one off the very_dirty list */
62                 D1(printk(KERN_DEBUG "Picking block from very_dirty_list to GC next\n"));
63                 nextlist = &c->very_dirty_list;
64         } else if (n < 126 && !list_empty(&c->dirty_list)) {
65                 D1(printk(KERN_DEBUG "Picking block from dirty_list to GC next\n"));
66                 nextlist = &c->dirty_list;
67         } else if (!list_empty(&c->clean_list)) {
68                 D1(printk(KERN_DEBUG "Picking block from clean_list to GC next\n"));
69                 nextlist = &c->clean_list;
70         } else if (!list_empty(&c->dirty_list)) {
71                 D1(printk(KERN_DEBUG "Picking block from dirty_list to GC next (clean_list was empty)\n"));
72
73                 nextlist = &c->dirty_list;
74         } else if (!list_empty(&c->very_dirty_list)) {
75                 D1(printk(KERN_DEBUG "Picking block from very_dirty_list to GC next (clean_list and dirty_list were empty)\n"));
76                 nextlist = &c->very_dirty_list;
77         } else if (!list_empty(&c->erasable_list)) {
78                 D1(printk(KERN_DEBUG "Picking block from erasable_list to GC next (clean_list and {very_,}dirty_list were empty)\n"));
79
80                 nextlist = &c->erasable_list;
81         } else if (!list_empty(&c->erasable_pending_wbuf_list)) {
82                 /* There are blocks are wating for the wbuf sync */
83                 D1(printk(KERN_DEBUG "Synching wbuf in order to reuse erasable_pending_wbuf_list blocks\n"));
84                 spin_unlock(&c->erase_completion_lock);
85                 jffs2_flush_wbuf_pad(c);
86                 spin_lock(&c->erase_completion_lock);
87                 goto again;
88         } else {
89                 /* Eep. All were empty */
90                 D1(printk(KERN_NOTICE "jffs2: No clean, dirty _or_ erasable blocks to GC from! Where are they all?\n"));
91                 return NULL;
92         }
93
94         ret = list_entry(nextlist->next, struct jffs2_eraseblock, list);
95         list_del(&ret->list);
96         c->gcblock = ret;
97         ret->gc_node = ret->first_node;
98         if (!ret->gc_node) {
99                 printk(KERN_WARNING "Eep. ret->gc_node for block at 0x%08x is NULL\n", ret->offset);
100                 BUG();
101         }
102
103         /* Have we accidentally picked a clean block with wasted space ? */
104         if (ret->wasted_size) {
105                 D1(printk(KERN_DEBUG "Converting wasted_size %08x to dirty_size\n", ret->wasted_size));
106                 ret->dirty_size += ret->wasted_size;
107                 c->wasted_size -= ret->wasted_size;
108                 c->dirty_size += ret->wasted_size;
109                 ret->wasted_size = 0;
110         }
111
112         return ret;
113 }
114
115 /* jffs2_garbage_collect_pass
116  * Make a single attempt to progress GC. Move one node, and possibly
117  * start erasing one eraseblock.
118  */
119 int jffs2_garbage_collect_pass(struct jffs2_sb_info *c)
120 {
121         struct jffs2_inode_info *f;
122         struct jffs2_inode_cache *ic;
123         struct jffs2_eraseblock *jeb;
124         struct jffs2_raw_node_ref *raw;
125         uint32_t gcblock_dirty;
126         int ret = 0, inum, nlink;
127         int xattr = 0;
128
129         if (mutex_lock_interruptible(&c->alloc_sem))
130                 return -EINTR;
131
132         for (;;) {
133                 spin_lock(&c->erase_completion_lock);
134                 if (!c->unchecked_size)
135                         break;
136
137                 /* We can't start doing GC yet. We haven't finished checking
138                    the node CRCs etc. Do it now. */
139
140                 /* checked_ino is protected by the alloc_sem */
141                 if (c->checked_ino > c->highest_ino && xattr) {
142                         printk(KERN_CRIT "Checked all inodes but still 0x%x bytes of unchecked space?\n",
143                                c->unchecked_size);
144                         jffs2_dbg_dump_block_lists_nolock(c);
145                         spin_unlock(&c->erase_completion_lock);
146                         mutex_unlock(&c->alloc_sem);
147                         return -ENOSPC;
148                 }
149
150                 spin_unlock(&c->erase_completion_lock);
151
152                 if (!xattr)
153                         xattr = jffs2_verify_xattr(c);
154
155                 spin_lock(&c->inocache_lock);
156
157                 ic = jffs2_get_ino_cache(c, c->checked_ino++);
158
159                 if (!ic) {
160                         spin_unlock(&c->inocache_lock);
161                         continue;
162                 }
163
164                 if (!ic->nlink) {
165                         D1(printk(KERN_DEBUG "Skipping check of ino #%d with nlink zero\n",
166                                   ic->ino));
167                         spin_unlock(&c->inocache_lock);
168                         jffs2_xattr_delete_inode(c, ic);
169                         continue;
170                 }
171                 switch(ic->state) {
172                 case INO_STATE_CHECKEDABSENT:
173                 case INO_STATE_PRESENT:
174                         D1(printk(KERN_DEBUG "Skipping ino #%u already checked\n", ic->ino));
175                         spin_unlock(&c->inocache_lock);
176                         continue;
177
178                 case INO_STATE_GC:
179                 case INO_STATE_CHECKING:
180                         printk(KERN_WARNING "Inode #%u is in state %d during CRC check phase!\n", ic->ino, ic->state);
181                         spin_unlock(&c->inocache_lock);
182                         BUG();
183
184                 case INO_STATE_READING:
185                         /* We need to wait for it to finish, lest we move on
186                            and trigger the BUG() above while we haven't yet
187                            finished checking all its nodes */
188                         D1(printk(KERN_DEBUG "Waiting for ino #%u to finish reading\n", ic->ino));
189                         /* We need to come back again for the _same_ inode. We've
190                          made no progress in this case, but that should be OK */
191                         c->checked_ino--;
192
193                         mutex_unlock(&c->alloc_sem);
194                         sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
195                         return 0;
196
197                 default:
198                         BUG();
199
200                 case INO_STATE_UNCHECKED:
201                         ;
202                 }
203                 ic->state = INO_STATE_CHECKING;
204                 spin_unlock(&c->inocache_lock);
205
206                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() triggering inode scan of ino#%u\n", ic->ino));
207
208                 ret = jffs2_do_crccheck_inode(c, ic);
209                 if (ret)
210                         printk(KERN_WARNING "Returned error for crccheck of ino #%u. Expect badness...\n", ic->ino);
211
212                 jffs2_set_inocache_state(c, ic, INO_STATE_CHECKEDABSENT);
213                 mutex_unlock(&c->alloc_sem);
214                 return ret;
215         }
216
217         /* First, work out which block we're garbage-collecting */
218         jeb = c->gcblock;
219
220         if (!jeb)
221                 jeb = jffs2_find_gc_block(c);
222
223         if (!jeb) {
224                 /* Couldn't find a free block. But maybe we can just erase one and make 'progress'? */
225                 if (!list_empty(&c->erase_pending_list)) {
226                         spin_unlock(&c->erase_completion_lock);
227                         mutex_unlock(&c->alloc_sem);
228                         return -EAGAIN;
229                 }
230                 D1(printk(KERN_NOTICE "jffs2: Couldn't find erase block to garbage collect!\n"));
231                 spin_unlock(&c->erase_completion_lock);
232                 mutex_unlock(&c->alloc_sem);
233                 return -EIO;
234         }
235
236         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));
237         D1(if (c->nextblock)
238            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));
239
240         if (!jeb->used_size) {
241                 mutex_unlock(&c->alloc_sem);
242                 goto eraseit;
243         }
244
245         raw = jeb->gc_node;
246         gcblock_dirty = jeb->dirty_size;
247
248         while(ref_obsolete(raw)) {
249                 D1(printk(KERN_DEBUG "Node at 0x%08x is obsolete... skipping\n", ref_offset(raw)));
250                 raw = ref_next(raw);
251                 if (unlikely(!raw)) {
252                         printk(KERN_WARNING "eep. End of raw list while still supposedly nodes to GC\n");
253                         printk(KERN_WARNING "erase block at 0x%08x. free_size 0x%08x, dirty_size 0x%08x, used_size 0x%08x\n",
254                                jeb->offset, jeb->free_size, jeb->dirty_size, jeb->used_size);
255                         jeb->gc_node = raw;
256                         spin_unlock(&c->erase_completion_lock);
257                         mutex_unlock(&c->alloc_sem);
258                         BUG();
259                 }
260         }
261         jeb->gc_node = raw;
262
263         D1(printk(KERN_DEBUG "Going to garbage collect node at 0x%08x\n", ref_offset(raw)));
264
265         if (!raw->next_in_ino) {
266                 /* Inode-less node. Clean marker, snapshot or something like that */
267                 spin_unlock(&c->erase_completion_lock);
268                 if (ref_flags(raw) == REF_PRISTINE) {
269                         /* It's an unknown node with JFFS2_FEATURE_RWCOMPAT_COPY */
270                         jffs2_garbage_collect_pristine(c, NULL, raw);
271                 } else {
272                         /* Just mark it obsolete */
273                         jffs2_mark_node_obsolete(c, raw);
274                 }
275                 mutex_unlock(&c->alloc_sem);
276                 goto eraseit_lock;
277         }
278
279         ic = jffs2_raw_ref_to_ic(raw);
280
281 #ifdef CONFIG_JFFS2_FS_XATTR
282         /* When 'ic' refers xattr_datum/xattr_ref, this node is GCed as xattr.
283          * We can decide whether this node is inode or xattr by ic->class.     */
284         if (ic->class == RAWNODE_CLASS_XATTR_DATUM
285             || ic->class == RAWNODE_CLASS_XATTR_REF) {
286                 spin_unlock(&c->erase_completion_lock);
287
288                 if (ic->class == RAWNODE_CLASS_XATTR_DATUM) {
289                         ret = jffs2_garbage_collect_xattr_datum(c, (struct jffs2_xattr_datum *)ic, raw);
290                 } else {
291                         ret = jffs2_garbage_collect_xattr_ref(c, (struct jffs2_xattr_ref *)ic, raw);
292                 }
293                 goto test_gcnode;
294         }
295 #endif
296
297         /* We need to hold the inocache. Either the erase_completion_lock or
298            the inocache_lock are sufficient; we trade down since the inocache_lock
299            causes less contention. */
300         spin_lock(&c->inocache_lock);
301
302         spin_unlock(&c->erase_completion_lock);
303
304         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));
305
306         /* Three possibilities:
307            1. Inode is already in-core. We must iget it and do proper
308               updating to its fragtree, etc.
309            2. Inode is not in-core, node is REF_PRISTINE. We lock the
310               inocache to prevent a read_inode(), copy the node intact.
311            3. Inode is not in-core, node is not pristine. We must iget()
312               and take the slow path.
313         */
314
315         switch(ic->state) {
316         case INO_STATE_CHECKEDABSENT:
317                 /* It's been checked, but it's not currently in-core.
318                    We can just copy any pristine nodes, but have
319                    to prevent anyone else from doing read_inode() while
320                    we're at it, so we set the state accordingly */
321                 if (ref_flags(raw) == REF_PRISTINE)
322                         ic->state = INO_STATE_GC;
323                 else {
324                         D1(printk(KERN_DEBUG "Ino #%u is absent but node not REF_PRISTINE. Reading.\n",
325                                   ic->ino));
326                 }
327                 break;
328
329         case INO_STATE_PRESENT:
330                 /* It's in-core. GC must iget() it. */
331                 break;
332
333         case INO_STATE_UNCHECKED:
334         case INO_STATE_CHECKING:
335         case INO_STATE_GC:
336                 /* Should never happen. We should have finished checking
337                    by the time we actually start doing any GC, and since
338                    we're holding the alloc_sem, no other garbage collection
339                    can happen.
340                 */
341                 printk(KERN_CRIT "Inode #%u already in state %d in jffs2_garbage_collect_pass()!\n",
342                        ic->ino, ic->state);
343                 mutex_unlock(&c->alloc_sem);
344                 spin_unlock(&c->inocache_lock);
345                 BUG();
346
347         case INO_STATE_READING:
348                 /* Someone's currently trying to read it. We must wait for
349                    them to finish and then go through the full iget() route
350                    to do the GC. However, sometimes read_inode() needs to get
351                    the alloc_sem() (for marking nodes invalid) so we must
352                    drop the alloc_sem before sleeping. */
353
354                 mutex_unlock(&c->alloc_sem);
355                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() waiting for ino #%u in state %d\n",
356                           ic->ino, ic->state));
357                 sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
358                 /* And because we dropped the alloc_sem we must start again from the
359                    beginning. Ponder chance of livelock here -- we're returning success
360                    without actually making any progress.
361
362                    Q: What are the chances that the inode is back in INO_STATE_READING
363                    again by the time we next enter this function? And that this happens
364                    enough times to cause a real delay?
365
366                    A: Small enough that I don't care :)
367                 */
368                 return 0;
369         }
370
371         /* OK. Now if the inode is in state INO_STATE_GC, we are going to copy the
372            node intact, and we don't have to muck about with the fragtree etc.
373            because we know it's not in-core. If it _was_ in-core, we go through
374            all the iget() crap anyway */
375
376         if (ic->state == INO_STATE_GC) {
377                 spin_unlock(&c->inocache_lock);
378
379                 ret = jffs2_garbage_collect_pristine(c, ic, raw);
380
381                 spin_lock(&c->inocache_lock);
382                 ic->state = INO_STATE_CHECKEDABSENT;
383                 wake_up(&c->inocache_wq);
384
385                 if (ret != -EBADFD) {
386                         spin_unlock(&c->inocache_lock);
387                         goto test_gcnode;
388                 }
389
390                 /* Fall through if it wanted us to, with inocache_lock held */
391         }
392
393         /* Prevent the fairly unlikely race where the gcblock is
394            entirely obsoleted by the final close of a file which had
395            the only valid nodes in the block, followed by erasure,
396            followed by freeing of the ic because the erased block(s)
397            held _all_ the nodes of that inode.... never been seen but
398            it's vaguely possible. */
399
400         inum = ic->ino;
401         nlink = ic->nlink;
402         spin_unlock(&c->inocache_lock);
403
404         f = jffs2_gc_fetch_inode(c, inum, nlink);
405         if (IS_ERR(f)) {
406                 ret = PTR_ERR(f);
407                 goto release_sem;
408         }
409         if (!f) {
410                 ret = 0;
411                 goto release_sem;
412         }
413
414         ret = jffs2_garbage_collect_live(c, jeb, raw, f);
415
416         jffs2_gc_release_inode(c, f);
417
418  test_gcnode:
419         if (jeb->dirty_size == gcblock_dirty && !ref_obsolete(jeb->gc_node)) {
420                 /* Eep. This really should never happen. GC is broken */
421                 printk(KERN_ERR "Error garbage collecting node at %08x!\n", ref_offset(jeb->gc_node));
422                 ret = -ENOSPC;
423         }
424  release_sem:
425         mutex_unlock(&c->alloc_sem);
426
427  eraseit_lock:
428         /* If we've finished this block, start it erasing */
429         spin_lock(&c->erase_completion_lock);
430
431  eraseit:
432         if (c->gcblock && !c->gcblock->used_size) {
433                 D1(printk(KERN_DEBUG "Block at 0x%08x completely obsoleted by GC. Moving to erase_pending_list\n", c->gcblock->offset));
434                 /* We're GC'ing an empty block? */
435                 list_add_tail(&c->gcblock->list, &c->erase_pending_list);
436                 c->gcblock = NULL;
437                 c->nr_erasing_blocks++;
438                 jffs2_erase_pending_trigger(c);
439         }
440         spin_unlock(&c->erase_completion_lock);
441
442         return ret;
443 }
444
445 static int jffs2_garbage_collect_live(struct jffs2_sb_info *c,  struct jffs2_eraseblock *jeb,
446                                       struct jffs2_raw_node_ref *raw, struct jffs2_inode_info *f)
447 {
448         struct jffs2_node_frag *frag;
449         struct jffs2_full_dnode *fn = NULL;
450         struct jffs2_full_dirent *fd;
451         uint32_t start = 0, end = 0, nrfrags = 0;
452         int ret = 0;
453
454         mutex_lock(&f->sem);
455
456         /* Now we have the lock for this inode. Check that it's still the one at the head
457            of the list. */
458
459         spin_lock(&c->erase_completion_lock);
460
461         if (c->gcblock != jeb) {
462                 spin_unlock(&c->erase_completion_lock);
463                 D1(printk(KERN_DEBUG "GC block is no longer gcblock. Restart\n"));
464                 goto upnout;
465         }
466         if (ref_obsolete(raw)) {
467                 spin_unlock(&c->erase_completion_lock);
468                 D1(printk(KERN_DEBUG "node to be GC'd was obsoleted in the meantime.\n"));
469                 /* They'll call again */
470                 goto upnout;
471         }
472         spin_unlock(&c->erase_completion_lock);
473
474         /* OK. Looks safe. And nobody can get us now because we have the semaphore. Move the block */
475         if (f->metadata && f->metadata->raw == raw) {
476                 fn = f->metadata;
477                 ret = jffs2_garbage_collect_metadata(c, jeb, f, fn);
478                 goto upnout;
479         }
480
481         /* FIXME. Read node and do lookup? */
482         for (frag = frag_first(&f->fragtree); frag; frag = frag_next(frag)) {
483                 if (frag->node && frag->node->raw == raw) {
484                         fn = frag->node;
485                         end = frag->ofs + frag->size;
486                         if (!nrfrags++)
487                                 start = frag->ofs;
488                         if (nrfrags == frag->node->frags)
489                                 break; /* We've found them all */
490                 }
491         }
492         if (fn) {
493                 if (ref_flags(raw) == REF_PRISTINE) {
494                         ret = jffs2_garbage_collect_pristine(c, f->inocache, raw);
495                         if (!ret) {
496                                 /* Urgh. Return it sensibly. */
497                                 frag->node->raw = f->inocache->nodes;
498                         }
499                         if (ret != -EBADFD)
500                                 goto upnout;
501                 }
502                 /* We found a datanode. Do the GC */
503                 if((start >> PAGE_CACHE_SHIFT) < ((end-1) >> PAGE_CACHE_SHIFT)) {
504                         /* It crosses a page boundary. Therefore, it must be a hole. */
505                         ret = jffs2_garbage_collect_hole(c, jeb, f, fn, start, end);
506                 } else {
507                         /* It could still be a hole. But we GC the page this way anyway */
508                         ret = jffs2_garbage_collect_dnode(c, jeb, f, fn, start, end);
509                 }
510                 goto upnout;
511         }
512
513         /* Wasn't a dnode. Try dirent */
514         for (fd = f->dents; fd; fd=fd->next) {
515                 if (fd->raw == raw)
516                         break;
517         }
518
519         if (fd && fd->ino) {
520                 ret = jffs2_garbage_collect_dirent(c, jeb, f, fd);
521         } else if (fd) {
522                 ret = jffs2_garbage_collect_deletion_dirent(c, jeb, f, fd);
523         } else {
524                 printk(KERN_WARNING "Raw node at 0x%08x wasn't in node lists for ino #%u\n",
525                        ref_offset(raw), f->inocache->ino);
526                 if (ref_obsolete(raw)) {
527                         printk(KERN_WARNING "But it's obsolete so we don't mind too much\n");
528                 } else {
529                         jffs2_dbg_dump_node(c, ref_offset(raw));
530                         BUG();
531                 }
532         }
533  upnout:
534         mutex_unlock(&f->sem);
535
536         return ret;
537 }
538
539 static int jffs2_garbage_collect_pristine(struct jffs2_sb_info *c,
540                                           struct jffs2_inode_cache *ic,
541                                           struct jffs2_raw_node_ref *raw)
542 {
543         union jffs2_node_union *node;
544         size_t retlen;
545         int ret;
546         uint32_t phys_ofs, alloclen;
547         uint32_t crc, rawlen;
548         int retried = 0;
549
550         D1(printk(KERN_DEBUG "Going to GC REF_PRISTINE node at 0x%08x\n", ref_offset(raw)));
551
552         alloclen = rawlen = ref_totlen(c, c->gcblock, raw);
553
554         /* Ask for a small amount of space (or the totlen if smaller) because we
555            don't want to force wastage of the end of a block if splitting would
556            work. */
557         if (ic && alloclen > sizeof(struct jffs2_raw_inode) + JFFS2_MIN_DATA_LEN)
558                 alloclen = sizeof(struct jffs2_raw_inode) + JFFS2_MIN_DATA_LEN;
559
560         ret = jffs2_reserve_space_gc(c, alloclen, &alloclen, rawlen);
561         /* 'rawlen' is not the exact summary size; it is only an upper estimation */
562
563         if (ret)
564                 return ret;
565
566         if (alloclen < rawlen) {
567                 /* Doesn't fit untouched. We'll go the old route and split it */
568                 return -EBADFD;
569         }
570
571         node = kmalloc(rawlen, GFP_KERNEL);
572         if (!node)
573                 return -ENOMEM;
574
575         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)node);
576         if (!ret && retlen != rawlen)
577                 ret = -EIO;
578         if (ret)
579                 goto out_node;
580
581         crc = crc32(0, node, sizeof(struct jffs2_unknown_node)-4);
582         if (je32_to_cpu(node->u.hdr_crc) != crc) {
583                 printk(KERN_WARNING "Header CRC failed on REF_PRISTINE node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
584                        ref_offset(raw), je32_to_cpu(node->u.hdr_crc), crc);
585                 goto bail;
586         }
587
588         switch(je16_to_cpu(node->u.nodetype)) {
589         case JFFS2_NODETYPE_INODE:
590                 crc = crc32(0, node, sizeof(node->i)-8);
591                 if (je32_to_cpu(node->i.node_crc) != crc) {
592                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
593                                ref_offset(raw), je32_to_cpu(node->i.node_crc), crc);
594                         goto bail;
595                 }
596
597                 if (je32_to_cpu(node->i.dsize)) {
598                         crc = crc32(0, node->i.data, je32_to_cpu(node->i.csize));
599                         if (je32_to_cpu(node->i.data_crc) != crc) {
600                                 printk(KERN_WARNING "Data CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
601                                        ref_offset(raw), je32_to_cpu(node->i.data_crc), crc);
602                                 goto bail;
603                         }
604                 }
605                 break;
606
607         case JFFS2_NODETYPE_DIRENT:
608                 crc = crc32(0, node, sizeof(node->d)-8);
609                 if (je32_to_cpu(node->d.node_crc) != crc) {
610                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE dirent node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
611                                ref_offset(raw), je32_to_cpu(node->d.node_crc), crc);
612                         goto bail;
613                 }
614
615                 if (strnlen(node->d.name, node->d.nsize) != node->d.nsize) {
616                         printk(KERN_WARNING "Name in dirent node at 0x%08x contains zeroes\n", ref_offset(raw));
617                         goto bail;
618                 }
619
620                 if (node->d.nsize) {
621                         crc = crc32(0, node->d.name, node->d.nsize);
622                         if (je32_to_cpu(node->d.name_crc) != crc) {
623                                 printk(KERN_WARNING "Name CRC failed on REF_PRISTINE dirent node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
624                                        ref_offset(raw), je32_to_cpu(node->d.name_crc), crc);
625                                 goto bail;
626                         }
627                 }
628                 break;
629         default:
630                 /* If it's inode-less, we don't _know_ what it is. Just copy it intact */
631                 if (ic) {
632                         printk(KERN_WARNING "Unknown node type for REF_PRISTINE node at 0x%08x: 0x%04x\n",
633                                ref_offset(raw), je16_to_cpu(node->u.nodetype));
634                         goto bail;
635                 }
636         }
637
638         /* OK, all the CRCs are good; this node can just be copied as-is. */
639  retry:
640         phys_ofs = write_ofs(c);
641
642         ret = jffs2_flash_write(c, phys_ofs, rawlen, &retlen, (char *)node);
643
644         if (ret || (retlen != rawlen)) {
645                 printk(KERN_NOTICE "Write of %d bytes at 0x%08x failed. returned %d, retlen %zd\n",
646                        rawlen, phys_ofs, ret, retlen);
647                 if (retlen) {
648                         jffs2_add_physical_node_ref(c, phys_ofs | REF_OBSOLETE, rawlen, NULL);
649                 } else {
650                         printk(KERN_NOTICE "Not marking the space at 0x%08x as dirty because the flash driver returned retlen zero\n", phys_ofs);
651                 }
652                 if (!retried) {
653                         /* Try to reallocate space and retry */
654                         uint32_t dummy;
655                         struct jffs2_eraseblock *jeb = &c->blocks[phys_ofs / c->sector_size];
656
657                         retried = 1;
658
659                         D1(printk(KERN_DEBUG "Retrying failed write of REF_PRISTINE node.\n"));
660
661                         jffs2_dbg_acct_sanity_check(c,jeb);
662                         jffs2_dbg_acct_paranoia_check(c, jeb);
663
664                         ret = jffs2_reserve_space_gc(c, rawlen, &dummy, rawlen);
665                                                 /* this is not the exact summary size of it,
666                                                         it is only an upper estimation */
667
668                         if (!ret) {
669                                 D1(printk(KERN_DEBUG "Allocated space at 0x%08x to retry failed write.\n", phys_ofs));
670
671                                 jffs2_dbg_acct_sanity_check(c,jeb);
672                                 jffs2_dbg_acct_paranoia_check(c, jeb);
673
674                                 goto retry;
675                         }
676                         D1(printk(KERN_DEBUG "Failed to allocate space to retry failed write: %d!\n", ret));
677                 }
678
679                 if (!ret)
680                         ret = -EIO;
681                 goto out_node;
682         }
683         jffs2_add_physical_node_ref(c, phys_ofs | REF_PRISTINE, rawlen, ic);
684
685         jffs2_mark_node_obsolete(c, raw);
686         D1(printk(KERN_DEBUG "WHEEE! GC REF_PRISTINE node at 0x%08x succeeded\n", ref_offset(raw)));
687
688  out_node:
689         kfree(node);
690         return ret;
691  bail:
692         ret = -EBADFD;
693         goto out_node;
694 }
695
696 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
697                                         struct jffs2_inode_info *f, struct jffs2_full_dnode *fn)
698 {
699         struct jffs2_full_dnode *new_fn;
700         struct jffs2_raw_inode ri;
701         struct jffs2_node_frag *last_frag;
702         union jffs2_device_node dev;
703         char *mdata = NULL, mdatalen = 0;
704         uint32_t alloclen, ilen;
705         int ret;
706
707         if (S_ISBLK(JFFS2_F_I_MODE(f)) ||
708             S_ISCHR(JFFS2_F_I_MODE(f)) ) {
709                 /* For these, we don't actually need to read the old node */
710                 mdatalen = jffs2_encode_dev(&dev, JFFS2_F_I_RDEV(f));
711                 mdata = (char *)&dev;
712                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bytes of kdev_t\n", mdatalen));
713         } else if (S_ISLNK(JFFS2_F_I_MODE(f))) {
714                 mdatalen = fn->size;
715                 mdata = kmalloc(fn->size, GFP_KERNEL);
716                 if (!mdata) {
717                         printk(KERN_WARNING "kmalloc of mdata failed in jffs2_garbage_collect_metadata()\n");
718                         return -ENOMEM;
719                 }
720                 ret = jffs2_read_dnode(c, f, fn, mdata, 0, mdatalen);
721                 if (ret) {
722                         printk(KERN_WARNING "read of old metadata failed in jffs2_garbage_collect_metadata(): %d\n", ret);
723                         kfree(mdata);
724                         return ret;
725                 }
726                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bites of symlink target\n", mdatalen));
727
728         }
729
730         ret = jffs2_reserve_space_gc(c, sizeof(ri) + mdatalen, &alloclen,
731                                 JFFS2_SUMMARY_INODE_SIZE);
732         if (ret) {
733                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_metadata failed: %d\n",
734                        sizeof(ri)+ mdatalen, ret);
735                 goto out;
736         }
737
738         last_frag = frag_last(&f->fragtree);
739         if (last_frag)
740                 /* Fetch the inode length from the fragtree rather then
741                  * from i_size since i_size may have not been updated yet */
742                 ilen = last_frag->ofs + last_frag->size;
743         else
744                 ilen = JFFS2_F_I_SIZE(f);
745
746         memset(&ri, 0, sizeof(ri));
747         ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
748         ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
749         ri.totlen = cpu_to_je32(sizeof(ri) + mdatalen);
750         ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
751
752         ri.ino = cpu_to_je32(f->inocache->ino);
753         ri.version = cpu_to_je32(++f->highest_version);
754         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
755         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
756         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
757         ri.isize = cpu_to_je32(ilen);
758         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
759         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
760         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
761         ri.offset = cpu_to_je32(0);
762         ri.csize = cpu_to_je32(mdatalen);
763         ri.dsize = cpu_to_je32(mdatalen);
764         ri.compr = JFFS2_COMPR_NONE;
765         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
766         ri.data_crc = cpu_to_je32(crc32(0, mdata, mdatalen));
767
768         new_fn = jffs2_write_dnode(c, f, &ri, mdata, mdatalen, ALLOC_GC);
769
770         if (IS_ERR(new_fn)) {
771                 printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
772                 ret = PTR_ERR(new_fn);
773                 goto out;
774         }
775         jffs2_mark_node_obsolete(c, fn->raw);
776         jffs2_free_full_dnode(fn);
777         f->metadata = new_fn;
778  out:
779         if (S_ISLNK(JFFS2_F_I_MODE(f)))
780                 kfree(mdata);
781         return ret;
782 }
783
784 static int jffs2_garbage_collect_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
785                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
786 {
787         struct jffs2_full_dirent *new_fd;
788         struct jffs2_raw_dirent rd;
789         uint32_t alloclen;
790         int ret;
791
792         rd.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
793         rd.nodetype = cpu_to_je16(JFFS2_NODETYPE_DIRENT);
794         rd.nsize = strlen(fd->name);
795         rd.totlen = cpu_to_je32(sizeof(rd) + rd.nsize);
796         rd.hdr_crc = cpu_to_je32(crc32(0, &rd, sizeof(struct jffs2_unknown_node)-4));
797
798         rd.pino = cpu_to_je32(f->inocache->ino);
799         rd.version = cpu_to_je32(++f->highest_version);
800         rd.ino = cpu_to_je32(fd->ino);
801         /* If the times on this inode were set by explicit utime() they can be different,
802            so refrain from splatting them. */
803         if (JFFS2_F_I_MTIME(f) == JFFS2_F_I_CTIME(f))
804                 rd.mctime = cpu_to_je32(JFFS2_F_I_MTIME(f));
805         else
806                 rd.mctime = cpu_to_je32(0);
807         rd.type = fd->type;
808         rd.node_crc = cpu_to_je32(crc32(0, &rd, sizeof(rd)-8));
809         rd.name_crc = cpu_to_je32(crc32(0, fd->name, rd.nsize));
810
811         ret = jffs2_reserve_space_gc(c, sizeof(rd)+rd.nsize, &alloclen,
812                                 JFFS2_SUMMARY_DIRENT_SIZE(rd.nsize));
813         if (ret) {
814                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dirent failed: %d\n",
815                        sizeof(rd)+rd.nsize, ret);
816                 return ret;
817         }
818         new_fd = jffs2_write_dirent(c, f, &rd, fd->name, rd.nsize, ALLOC_GC);
819
820         if (IS_ERR(new_fd)) {
821                 printk(KERN_WARNING "jffs2_write_dirent in garbage_collect_dirent failed: %ld\n", PTR_ERR(new_fd));
822                 return PTR_ERR(new_fd);
823         }
824         jffs2_add_fd_to_list(c, new_fd, &f->dents);
825         return 0;
826 }
827
828 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
829                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
830 {
831         struct jffs2_full_dirent **fdp = &f->dents;
832         int found = 0;
833
834         /* On a medium where we can't actually mark nodes obsolete
835            pernamently, such as NAND flash, we need to work out
836            whether this deletion dirent is still needed to actively
837            delete a 'real' dirent with the same name that's still
838            somewhere else on the flash. */
839         if (!jffs2_can_mark_obsolete(c)) {
840                 struct jffs2_raw_dirent *rd;
841                 struct jffs2_raw_node_ref *raw;
842                 int ret;
843                 size_t retlen;
844                 int name_len = strlen(fd->name);
845                 uint32_t name_crc = crc32(0, fd->name, name_len);
846                 uint32_t rawlen = ref_totlen(c, jeb, fd->raw);
847
848                 rd = kmalloc(rawlen, GFP_KERNEL);
849                 if (!rd)
850                         return -ENOMEM;
851
852                 /* Prevent the erase code from nicking the obsolete node refs while
853                    we're looking at them. I really don't like this extra lock but
854                    can't see any alternative. Suggestions on a postcard to... */
855                 mutex_lock(&c->erase_free_sem);
856
857                 for (raw = f->inocache->nodes; raw != (void *)f->inocache; raw = raw->next_in_ino) {
858
859                         cond_resched();
860
861                         /* We only care about obsolete ones */
862                         if (!(ref_obsolete(raw)))
863                                 continue;
864
865                         /* Any dirent with the same name is going to have the same length... */
866                         if (ref_totlen(c, NULL, raw) != rawlen)
867                                 continue;
868
869                         /* Doesn't matter if there's one in the same erase block. We're going to
870                            delete it too at the same time. */
871                         if (SECTOR_ADDR(raw->flash_offset) == SECTOR_ADDR(fd->raw->flash_offset))
872                                 continue;
873
874                         D1(printk(KERN_DEBUG "Check potential deletion dirent at %08x\n", ref_offset(raw)));
875
876                         /* This is an obsolete node belonging to the same directory, and it's of the right
877                            length. We need to take a closer look...*/
878                         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)rd);
879                         if (ret) {
880                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Read error (%d) reading obsolete node at %08x\n", ret, ref_offset(raw));
881                                 /* If we can't read it, we don't need to continue to obsolete it. Continue */
882                                 continue;
883                         }
884                         if (retlen != rawlen) {
885                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Short read (%zd not %u) reading header from obsolete node at %08x\n",
886                                        retlen, rawlen, ref_offset(raw));
887                                 continue;
888                         }
889
890                         if (je16_to_cpu(rd->nodetype) != JFFS2_NODETYPE_DIRENT)
891                                 continue;
892
893                         /* If the name CRC doesn't match, skip */
894                         if (je32_to_cpu(rd->name_crc) != name_crc)
895                                 continue;
896
897                         /* If the name length doesn't match, or it's another deletion dirent, skip */
898                         if (rd->nsize != name_len || !je32_to_cpu(rd->ino))
899                                 continue;
900
901                         /* OK, check the actual name now */
902                         if (memcmp(rd->name, fd->name, name_len))
903                                 continue;
904
905                         /* OK. The name really does match. There really is still an older node on
906                            the flash which our deletion dirent obsoletes. So we have to write out
907                            a new deletion dirent to replace it */
908                         mutex_unlock(&c->erase_free_sem);
909
910                         D1(printk(KERN_DEBUG "Deletion dirent at %08x still obsoletes real dirent \"%s\" at %08x for ino #%u\n",
911                                   ref_offset(fd->raw), fd->name, ref_offset(raw), je32_to_cpu(rd->ino)));
912                         kfree(rd);
913
914                         return jffs2_garbage_collect_dirent(c, jeb, f, fd);
915                 }
916
917                 mutex_unlock(&c->erase_free_sem);
918                 kfree(rd);
919         }
920
921         /* FIXME: If we're deleting a dirent which contains the current mtime and ctime,
922            we should update the metadata node with those times accordingly */
923
924         /* No need for it any more. Just mark it obsolete and remove it from the list */
925         while (*fdp) {
926                 if ((*fdp) == fd) {
927                         found = 1;
928                         *fdp = fd->next;
929                         break;
930                 }
931                 fdp = &(*fdp)->next;
932         }
933         if (!found) {
934                 printk(KERN_WARNING "Deletion dirent \"%s\" not found in list for ino #%u\n", fd->name, f->inocache->ino);
935         }
936         jffs2_mark_node_obsolete(c, fd->raw);
937         jffs2_free_full_dirent(fd);
938         return 0;
939 }
940
941 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
942                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
943                                       uint32_t start, uint32_t end)
944 {
945         struct jffs2_raw_inode ri;
946         struct jffs2_node_frag *frag;
947         struct jffs2_full_dnode *new_fn;
948         uint32_t alloclen, ilen;
949         int ret;
950
951         D1(printk(KERN_DEBUG "Writing replacement hole node for ino #%u from offset 0x%x to 0x%x\n",
952                   f->inocache->ino, start, end));
953
954         memset(&ri, 0, sizeof(ri));
955
956         if(fn->frags > 1) {
957                 size_t readlen;
958                 uint32_t crc;
959                 /* It's partially obsoleted by a later write. So we have to
960                    write it out again with the _same_ version as before */
961                 ret = jffs2_flash_read(c, ref_offset(fn->raw), sizeof(ri), &readlen, (char *)&ri);
962                 if (readlen != sizeof(ri) || ret) {
963                         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);
964                         goto fill;
965                 }
966                 if (je16_to_cpu(ri.nodetype) != JFFS2_NODETYPE_INODE) {
967                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had node type 0x%04x instead of JFFS2_NODETYPE_INODE(0x%04x)\n",
968                                ref_offset(fn->raw),
969                                je16_to_cpu(ri.nodetype), JFFS2_NODETYPE_INODE);
970                         return -EIO;
971                 }
972                 if (je32_to_cpu(ri.totlen) != sizeof(ri)) {
973                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had totlen 0x%x instead of expected 0x%zx\n",
974                                ref_offset(fn->raw),
975                                je32_to_cpu(ri.totlen), sizeof(ri));
976                         return -EIO;
977                 }
978                 crc = crc32(0, &ri, sizeof(ri)-8);
979                 if (crc != je32_to_cpu(ri.node_crc)) {
980                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had CRC 0x%08x which doesn't match calculated CRC 0x%08x\n",
981                                ref_offset(fn->raw),
982                                je32_to_cpu(ri.node_crc), crc);
983                         /* FIXME: We could possibly deal with this by writing new holes for each frag */
984                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
985                                start, end, f->inocache->ino);
986                         goto fill;
987                 }
988                 if (ri.compr != JFFS2_COMPR_ZERO) {
989                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node 0x%08x wasn't a hole node!\n", ref_offset(fn->raw));
990                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
991                                start, end, f->inocache->ino);
992                         goto fill;
993                 }
994         } else {
995         fill:
996                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
997                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
998                 ri.totlen = cpu_to_je32(sizeof(ri));
999                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
1000
1001                 ri.ino = cpu_to_je32(f->inocache->ino);
1002                 ri.version = cpu_to_je32(++f->highest_version);
1003                 ri.offset = cpu_to_je32(start);
1004                 ri.dsize = cpu_to_je32(end - start);
1005                 ri.csize = cpu_to_je32(0);
1006                 ri.compr = JFFS2_COMPR_ZERO;
1007         }
1008
1009         frag = frag_last(&f->fragtree);
1010         if (frag)
1011                 /* Fetch the inode length from the fragtree rather then
1012                  * from i_size since i_size may have not been updated yet */
1013                 ilen = frag->ofs + frag->size;
1014         else
1015                 ilen = JFFS2_F_I_SIZE(f);
1016
1017         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1018         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1019         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1020         ri.isize = cpu_to_je32(ilen);
1021         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1022         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1023         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1024         ri.data_crc = cpu_to_je32(0);
1025         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1026
1027         ret = jffs2_reserve_space_gc(c, sizeof(ri), &alloclen,
1028                                      JFFS2_SUMMARY_INODE_SIZE);
1029         if (ret) {
1030                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_hole failed: %d\n",
1031                        sizeof(ri), ret);
1032                 return ret;
1033         }
1034         new_fn = jffs2_write_dnode(c, f, &ri, NULL, 0, ALLOC_GC);
1035
1036         if (IS_ERR(new_fn)) {
1037                 printk(KERN_WARNING "Error writing new hole node: %ld\n", PTR_ERR(new_fn));
1038                 return PTR_ERR(new_fn);
1039         }
1040         if (je32_to_cpu(ri.version) == f->highest_version) {
1041                 jffs2_add_full_dnode_to_inode(c, f, new_fn);
1042                 if (f->metadata) {
1043                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1044                         jffs2_free_full_dnode(f->metadata);
1045                         f->metadata = NULL;
1046                 }
1047                 return 0;
1048         }
1049
1050         /*
1051          * We should only get here in the case where the node we are
1052          * replacing had more than one frag, so we kept the same version
1053          * number as before. (Except in case of error -- see 'goto fill;'
1054          * above.)
1055          */
1056         D1(if(unlikely(fn->frags <= 1)) {
1057                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Replacing fn with %d frag(s) but new ver %d != highest_version %d of ino #%d\n",
1058                        fn->frags, je32_to_cpu(ri.version), f->highest_version,
1059                        je32_to_cpu(ri.ino));
1060         });
1061
1062         /* This is a partially-overlapped hole node. Mark it REF_NORMAL not REF_PRISTINE */
1063         mark_ref_normal(new_fn->raw);
1064
1065         for (frag = jffs2_lookup_node_frag(&f->fragtree, fn->ofs);
1066              frag; frag = frag_next(frag)) {
1067                 if (frag->ofs > fn->size + fn->ofs)
1068                         break;
1069                 if (frag->node == fn) {
1070                         frag->node = new_fn;
1071                         new_fn->frags++;
1072                         fn->frags--;
1073                 }
1074         }
1075         if (fn->frags) {
1076                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Old node still has frags!\n");
1077                 BUG();
1078         }
1079         if (!new_fn->frags) {
1080                 printk(KERN_WARNING "jffs2_garbage_collect_hole: New node has no frags!\n");
1081                 BUG();
1082         }
1083
1084         jffs2_mark_node_obsolete(c, fn->raw);
1085         jffs2_free_full_dnode(fn);
1086
1087         return 0;
1088 }
1089
1090 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *orig_jeb,
1091                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
1092                                        uint32_t start, uint32_t end)
1093 {
1094         struct jffs2_full_dnode *new_fn;
1095         struct jffs2_raw_inode ri;
1096         uint32_t alloclen, offset, orig_end, orig_start;
1097         int ret = 0;
1098         unsigned char *comprbuf = NULL, *writebuf;
1099         unsigned long pg;
1100         unsigned char *pg_ptr;
1101
1102         memset(&ri, 0, sizeof(ri));
1103
1104         D1(printk(KERN_DEBUG "Writing replacement dnode for ino #%u from offset 0x%x to 0x%x\n",
1105                   f->inocache->ino, start, end));
1106
1107         orig_end = end;
1108         orig_start = start;
1109
1110         if (c->nr_free_blocks + c->nr_erasing_blocks > c->resv_blocks_gcmerge) {
1111                 /* Attempt to do some merging. But only expand to cover logically
1112                    adjacent frags if the block containing them is already considered
1113                    to be dirty. Otherwise we end up with GC just going round in
1114                    circles dirtying the nodes it already wrote out, especially
1115                    on NAND where we have small eraseblocks and hence a much higher
1116                    chance of nodes having to be split to cross boundaries. */
1117
1118                 struct jffs2_node_frag *frag;
1119                 uint32_t min, max;
1120
1121                 min = start & ~(PAGE_CACHE_SIZE-1);
1122                 max = min + PAGE_CACHE_SIZE;
1123
1124                 frag = jffs2_lookup_node_frag(&f->fragtree, start);
1125
1126                 /* BUG_ON(!frag) but that'll happen anyway... */
1127
1128                 BUG_ON(frag->ofs != start);
1129
1130                 /* First grow down... */
1131                 while((frag = frag_prev(frag)) && frag->ofs >= min) {
1132
1133                         /* If the previous frag doesn't even reach the beginning, there's
1134                            excessive fragmentation. Just merge. */
1135                         if (frag->ofs > min) {
1136                                 D1(printk(KERN_DEBUG "Expanding down to cover partial frag (0x%x-0x%x)\n",
1137                                           frag->ofs, frag->ofs+frag->size));
1138                                 start = frag->ofs;
1139                                 continue;
1140                         }
1141                         /* OK. This frag holds the first byte of the page. */
1142                         if (!frag->node || !frag->node->raw) {
1143                                 D1(printk(KERN_DEBUG "First frag in page is hole (0x%x-0x%x). Not expanding down.\n",
1144                                           frag->ofs, frag->ofs+frag->size));
1145                                 break;
1146                         } else {
1147
1148                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1149                                    in a block which is still considered clean? If so, don't obsolete it.
1150                                    If not, cover it anyway. */
1151
1152                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1153                                 struct jffs2_eraseblock *jeb;
1154
1155                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1156
1157                                 if (jeb == c->gcblock) {
1158                                         D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1159                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1160                                         start = frag->ofs;
1161                                         break;
1162                                 }
1163                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1164                                         D1(printk(KERN_DEBUG "Not expanding down to cover frag (0x%x-0x%x) in clean block %08x\n",
1165                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1166                                         break;
1167                                 }
1168
1169                                 D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in dirty block %08x\n",
1170                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1171                                 start = frag->ofs;
1172                                 break;
1173                         }
1174                 }
1175
1176                 /* ... then up */
1177
1178                 /* Find last frag which is actually part of the node we're to GC. */
1179                 frag = jffs2_lookup_node_frag(&f->fragtree, end-1);
1180
1181                 while((frag = frag_next(frag)) && frag->ofs+frag->size <= max) {
1182
1183                         /* If the previous frag doesn't even reach the beginning, there's lots
1184                            of fragmentation. Just merge. */
1185                         if (frag->ofs+frag->size < max) {
1186                                 D1(printk(KERN_DEBUG "Expanding up to cover partial frag (0x%x-0x%x)\n",
1187                                           frag->ofs, frag->ofs+frag->size));
1188                                 end = frag->ofs + frag->size;
1189                                 continue;
1190                         }
1191
1192                         if (!frag->node || !frag->node->raw) {
1193                                 D1(printk(KERN_DEBUG "Last frag in page is hole (0x%x-0x%x). Not expanding up.\n",
1194                                           frag->ofs, frag->ofs+frag->size));
1195                                 break;
1196                         } else {
1197
1198                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1199                                    in a block which is still considered clean? If so, don't obsolete it.
1200                                    If not, cover it anyway. */
1201
1202                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1203                                 struct jffs2_eraseblock *jeb;
1204
1205                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1206
1207                                 if (jeb == c->gcblock) {
1208                                         D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1209                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1210                                         end = frag->ofs + frag->size;
1211                                         break;
1212                                 }
1213                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1214                                         D1(printk(KERN_DEBUG "Not expanding up to cover frag (0x%x-0x%x) in clean block %08x\n",
1215                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1216                                         break;
1217                                 }
1218
1219                                 D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in dirty block %08x\n",
1220                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1221                                 end = frag->ofs + frag->size;
1222                                 break;
1223                         }
1224                 }
1225                 D1(printk(KERN_DEBUG "Expanded dnode to write from (0x%x-0x%x) to (0x%x-0x%x)\n",
1226                           orig_start, orig_end, start, end));
1227
1228                 D1(BUG_ON(end > frag_last(&f->fragtree)->ofs + frag_last(&f->fragtree)->size));
1229                 BUG_ON(end < orig_end);
1230                 BUG_ON(start > orig_start);
1231         }
1232
1233         /* First, use readpage() to read the appropriate page into the page cache */
1234         /* Q: What happens if we actually try to GC the _same_ page for which commit_write()
1235          *    triggered garbage collection in the first place?
1236          * A: I _think_ it's OK. read_cache_page shouldn't deadlock, we'll write out the
1237          *    page OK. We'll actually write it out again in commit_write, which is a little
1238          *    suboptimal, but at least we're correct.
1239          */
1240         pg_ptr = jffs2_gc_fetch_page(c, f, start, &pg);
1241
1242         if (IS_ERR(pg_ptr)) {
1243                 printk(KERN_WARNING "read_cache_page() returned error: %ld\n", PTR_ERR(pg_ptr));
1244                 return PTR_ERR(pg_ptr);
1245         }
1246
1247         offset = start;
1248         while(offset < orig_end) {
1249                 uint32_t datalen;
1250                 uint32_t cdatalen;
1251                 uint16_t comprtype = JFFS2_COMPR_NONE;
1252
1253                 ret = jffs2_reserve_space_gc(c, sizeof(ri) + JFFS2_MIN_DATA_LEN,
1254                                         &alloclen, JFFS2_SUMMARY_INODE_SIZE);
1255
1256                 if (ret) {
1257                         printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dnode failed: %d\n",
1258                                sizeof(ri)+ JFFS2_MIN_DATA_LEN, ret);
1259                         break;
1260                 }
1261                 cdatalen = min_t(uint32_t, alloclen - sizeof(ri), end - offset);
1262                 datalen = end - offset;
1263
1264                 writebuf = pg_ptr + (offset & (PAGE_CACHE_SIZE -1));
1265
1266                 comprtype = jffs2_compress(c, f, writebuf, &comprbuf, &datalen, &cdatalen);
1267
1268                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
1269                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
1270                 ri.totlen = cpu_to_je32(sizeof(ri) + cdatalen);
1271                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
1272
1273                 ri.ino = cpu_to_je32(f->inocache->ino);
1274                 ri.version = cpu_to_je32(++f->highest_version);
1275                 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1276                 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1277                 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1278                 ri.isize = cpu_to_je32(JFFS2_F_I_SIZE(f));
1279                 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1280                 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1281                 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1282                 ri.offset = cpu_to_je32(offset);
1283                 ri.csize = cpu_to_je32(cdatalen);
1284                 ri.dsize = cpu_to_je32(datalen);
1285                 ri.compr = comprtype & 0xff;
1286                 ri.usercompr = (comprtype >> 8) & 0xff;
1287                 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1288                 ri.data_crc = cpu_to_je32(crc32(0, comprbuf, cdatalen));
1289
1290                 new_fn = jffs2_write_dnode(c, f, &ri, comprbuf, cdatalen, ALLOC_GC);
1291
1292                 jffs2_free_comprbuf(comprbuf, writebuf);
1293
1294                 if (IS_ERR(new_fn)) {
1295                         printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
1296                         ret = PTR_ERR(new_fn);
1297                         break;
1298                 }
1299                 ret = jffs2_add_full_dnode_to_inode(c, f, new_fn);
1300                 offset += datalen;
1301                 if (f->metadata) {
1302                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1303                         jffs2_free_full_dnode(f->metadata);
1304                         f->metadata = NULL;
1305                 }
1306         }
1307
1308         jffs2_gc_release_page(c, pg_ptr, &pg);
1309         return ret;
1310 }