[JFFS2] Fix JFFS2 [mc]time handling
[linux-flexiantxendom0-3.2.10.git] / 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 1.153 2005/08/17 13:46:22 dedekind 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
129         if (down_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) {
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                         BUG();
147                 }
148
149                 spin_unlock(&c->erase_completion_lock);
150
151                 spin_lock(&c->inocache_lock);
152
153                 ic = jffs2_get_ino_cache(c, c->checked_ino++);
154
155                 if (!ic) {
156                         spin_unlock(&c->inocache_lock);
157                         continue;
158                 }
159
160                 if (!ic->nlink) {
161                         D1(printk(KERN_DEBUG "Skipping check of ino #%d with nlink zero\n",
162                                   ic->ino));
163                         spin_unlock(&c->inocache_lock);
164                         continue;
165                 }
166                 switch(ic->state) {
167                 case INO_STATE_CHECKEDABSENT:
168                 case INO_STATE_PRESENT:
169                         D1(printk(KERN_DEBUG "Skipping ino #%u already checked\n", ic->ino));
170                         spin_unlock(&c->inocache_lock);
171                         continue;
172
173                 case INO_STATE_GC:
174                 case INO_STATE_CHECKING:
175                         printk(KERN_WARNING "Inode #%u is in state %d during CRC check phase!\n", ic->ino, ic->state);
176                         spin_unlock(&c->inocache_lock);
177                         BUG();
178
179                 case INO_STATE_READING:
180                         /* We need to wait for it to finish, lest we move on
181                            and trigger the BUG() above while we haven't yet 
182                            finished checking all its nodes */
183                         D1(printk(KERN_DEBUG "Waiting for ino #%u to finish reading\n", ic->ino));
184                         up(&c->alloc_sem);
185                         sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
186                         return 0;
187
188                 default:
189                         BUG();
190
191                 case INO_STATE_UNCHECKED:
192                         ;
193                 }
194                 ic->state = INO_STATE_CHECKING;
195                 spin_unlock(&c->inocache_lock);
196
197                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() triggering inode scan of ino#%u\n", ic->ino));
198
199                 ret = jffs2_do_crccheck_inode(c, ic);
200                 if (ret)
201                         printk(KERN_WARNING "Returned error for crccheck of ino #%u. Expect badness...\n", ic->ino);
202
203                 jffs2_set_inocache_state(c, ic, INO_STATE_CHECKEDABSENT);
204                 up(&c->alloc_sem);
205                 return ret;
206         }
207
208         /* First, work out which block we're garbage-collecting */
209         jeb = c->gcblock;
210
211         if (!jeb)
212                 jeb = jffs2_find_gc_block(c);
213
214         if (!jeb) {
215                 D1 (printk(KERN_NOTICE "jffs2: Couldn't find erase block to garbage collect!\n"));
216                 spin_unlock(&c->erase_completion_lock);
217                 up(&c->alloc_sem);
218                 return -EIO;
219         }
220
221         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));
222         D1(if (c->nextblock)
223            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));
224
225         if (!jeb->used_size) {
226                 up(&c->alloc_sem);
227                 goto eraseit;
228         }
229
230         raw = jeb->gc_node;
231                         
232         while(ref_obsolete(raw)) {
233                 D1(printk(KERN_DEBUG "Node at 0x%08x is obsolete... skipping\n", ref_offset(raw)));
234                 raw = raw->next_phys;
235                 if (unlikely(!raw)) {
236                         printk(KERN_WARNING "eep. End of raw list while still supposedly nodes to GC\n");
237                         printk(KERN_WARNING "erase block at 0x%08x. free_size 0x%08x, dirty_size 0x%08x, used_size 0x%08x\n", 
238                                jeb->offset, jeb->free_size, jeb->dirty_size, jeb->used_size);
239                         jeb->gc_node = raw;
240                         spin_unlock(&c->erase_completion_lock);
241                         up(&c->alloc_sem);
242                         BUG();
243                 }
244         }
245         jeb->gc_node = raw;
246
247         D1(printk(KERN_DEBUG "Going to garbage collect node at 0x%08x\n", ref_offset(raw)));
248
249         if (!raw->next_in_ino) {
250                 /* Inode-less node. Clean marker, snapshot or something like that */
251                 /* FIXME: If it's something that needs to be copied, including something
252                    we don't grok that has JFFS2_NODETYPE_RWCOMPAT_COPY, we should do so */
253                 spin_unlock(&c->erase_completion_lock);
254                 jffs2_mark_node_obsolete(c, raw);
255                 up(&c->alloc_sem);
256                 goto eraseit_lock;
257         }
258
259         ic = jffs2_raw_ref_to_ic(raw);
260
261         /* We need to hold the inocache. Either the erase_completion_lock or
262            the inocache_lock are sufficient; we trade down since the inocache_lock 
263            causes less contention. */
264         spin_lock(&c->inocache_lock);
265
266         spin_unlock(&c->erase_completion_lock);
267
268         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));
269
270         /* Three possibilities:
271            1. Inode is already in-core. We must iget it and do proper
272               updating to its fragtree, etc.
273            2. Inode is not in-core, node is REF_PRISTINE. We lock the
274               inocache to prevent a read_inode(), copy the node intact.
275            3. Inode is not in-core, node is not pristine. We must iget()
276               and take the slow path.
277         */
278
279         switch(ic->state) {
280         case INO_STATE_CHECKEDABSENT:
281                 /* It's been checked, but it's not currently in-core. 
282                    We can just copy any pristine nodes, but have
283                    to prevent anyone else from doing read_inode() while
284                    we're at it, so we set the state accordingly */
285                 if (ref_flags(raw) == REF_PRISTINE)
286                         ic->state = INO_STATE_GC;
287                 else {
288                         D1(printk(KERN_DEBUG "Ino #%u is absent but node not REF_PRISTINE. Reading.\n", 
289                                   ic->ino));
290                 }
291                 break;
292
293         case INO_STATE_PRESENT:
294                 /* It's in-core. GC must iget() it. */
295                 break;
296
297         case INO_STATE_UNCHECKED:
298         case INO_STATE_CHECKING:
299         case INO_STATE_GC:
300                 /* Should never happen. We should have finished checking
301                    by the time we actually start doing any GC, and since 
302                    we're holding the alloc_sem, no other garbage collection 
303                    can happen.
304                 */
305                 printk(KERN_CRIT "Inode #%u already in state %d in jffs2_garbage_collect_pass()!\n",
306                        ic->ino, ic->state);
307                 up(&c->alloc_sem);
308                 spin_unlock(&c->inocache_lock);
309                 BUG();
310
311         case INO_STATE_READING:
312                 /* Someone's currently trying to read it. We must wait for
313                    them to finish and then go through the full iget() route
314                    to do the GC. However, sometimes read_inode() needs to get
315                    the alloc_sem() (for marking nodes invalid) so we must
316                    drop the alloc_sem before sleeping. */
317
318                 up(&c->alloc_sem);
319                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() waiting for ino #%u in state %d\n",
320                           ic->ino, ic->state));
321                 sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
322                 /* And because we dropped the alloc_sem we must start again from the 
323                    beginning. Ponder chance of livelock here -- we're returning success
324                    without actually making any progress.
325
326                    Q: What are the chances that the inode is back in INO_STATE_READING 
327                    again by the time we next enter this function? And that this happens
328                    enough times to cause a real delay?
329
330                    A: Small enough that I don't care :) 
331                 */
332                 return 0;
333         }
334
335         /* OK. Now if the inode is in state INO_STATE_GC, we are going to copy the
336            node intact, and we don't have to muck about with the fragtree etc. 
337            because we know it's not in-core. If it _was_ in-core, we go through
338            all the iget() crap anyway */
339
340         if (ic->state == INO_STATE_GC) {
341                 spin_unlock(&c->inocache_lock);
342
343                 ret = jffs2_garbage_collect_pristine(c, ic, raw);
344
345                 spin_lock(&c->inocache_lock);
346                 ic->state = INO_STATE_CHECKEDABSENT;
347                 wake_up(&c->inocache_wq);
348
349                 if (ret != -EBADFD) {
350                         spin_unlock(&c->inocache_lock);
351                         goto release_sem;
352                 }
353
354                 /* Fall through if it wanted us to, with inocache_lock held */
355         }
356
357         /* Prevent the fairly unlikely race where the gcblock is
358            entirely obsoleted by the final close of a file which had
359            the only valid nodes in the block, followed by erasure,
360            followed by freeing of the ic because the erased block(s)
361            held _all_ the nodes of that inode.... never been seen but
362            it's vaguely possible. */
363
364         inum = ic->ino;
365         nlink = ic->nlink;
366         spin_unlock(&c->inocache_lock);
367
368         f = jffs2_gc_fetch_inode(c, inum, nlink);
369         if (IS_ERR(f)) {
370                 ret = PTR_ERR(f);
371                 goto release_sem;
372         }
373         if (!f) {
374                 ret = 0;
375                 goto release_sem;
376         }
377
378         ret = jffs2_garbage_collect_live(c, jeb, raw, f);
379
380         jffs2_gc_release_inode(c, f);
381
382  release_sem:
383         up(&c->alloc_sem);
384
385  eraseit_lock:
386         /* If we've finished this block, start it erasing */
387         spin_lock(&c->erase_completion_lock);
388
389  eraseit:
390         if (c->gcblock && !c->gcblock->used_size) {
391                 D1(printk(KERN_DEBUG "Block at 0x%08x completely obsoleted by GC. Moving to erase_pending_list\n", c->gcblock->offset));
392                 /* We're GC'ing an empty block? */
393                 list_add_tail(&c->gcblock->list, &c->erase_pending_list);
394                 c->gcblock = NULL;
395                 c->nr_erasing_blocks++;
396                 jffs2_erase_pending_trigger(c);
397         }
398         spin_unlock(&c->erase_completion_lock);
399
400         return ret;
401 }
402
403 static int jffs2_garbage_collect_live(struct jffs2_sb_info *c,  struct jffs2_eraseblock *jeb,
404                                       struct jffs2_raw_node_ref *raw, struct jffs2_inode_info *f)
405 {
406         struct jffs2_node_frag *frag;
407         struct jffs2_full_dnode *fn = NULL;
408         struct jffs2_full_dirent *fd;
409         uint32_t start = 0, end = 0, nrfrags = 0;
410         int ret = 0;
411
412         down(&f->sem);
413
414         /* Now we have the lock for this inode. Check that it's still the one at the head
415            of the list. */
416
417         spin_lock(&c->erase_completion_lock);
418
419         if (c->gcblock != jeb) {
420                 spin_unlock(&c->erase_completion_lock);
421                 D1(printk(KERN_DEBUG "GC block is no longer gcblock. Restart\n"));
422                 goto upnout;
423         }
424         if (ref_obsolete(raw)) {
425                 spin_unlock(&c->erase_completion_lock);
426                 D1(printk(KERN_DEBUG "node to be GC'd was obsoleted in the meantime.\n"));
427                 /* They'll call again */
428                 goto upnout;
429         }
430         spin_unlock(&c->erase_completion_lock);
431
432         /* OK. Looks safe. And nobody can get us now because we have the semaphore. Move the block */
433         if (f->metadata && f->metadata->raw == raw) {
434                 fn = f->metadata;
435                 ret = jffs2_garbage_collect_metadata(c, jeb, f, fn);
436                 goto upnout;
437         }
438
439         /* FIXME. Read node and do lookup? */
440         for (frag = frag_first(&f->fragtree); frag; frag = frag_next(frag)) {
441                 if (frag->node && frag->node->raw == raw) {
442                         fn = frag->node;
443                         end = frag->ofs + frag->size;
444                         if (!nrfrags++)
445                                 start = frag->ofs;
446                         if (nrfrags == frag->node->frags)
447                                 break; /* We've found them all */
448                 }
449         }
450         if (fn) {
451                 if (ref_flags(raw) == REF_PRISTINE) {
452                         ret = jffs2_garbage_collect_pristine(c, f->inocache, raw);
453                         if (!ret) {
454                                 /* Urgh. Return it sensibly. */
455                                 frag->node->raw = f->inocache->nodes;
456                         }       
457                         if (ret != -EBADFD)
458                                 goto upnout;
459                 }
460                 /* We found a datanode. Do the GC */
461                 if((start >> PAGE_CACHE_SHIFT) < ((end-1) >> PAGE_CACHE_SHIFT)) {
462                         /* It crosses a page boundary. Therefore, it must be a hole. */
463                         ret = jffs2_garbage_collect_hole(c, jeb, f, fn, start, end);
464                 } else {
465                         /* It could still be a hole. But we GC the page this way anyway */
466                         ret = jffs2_garbage_collect_dnode(c, jeb, f, fn, start, end);
467                 }
468                 goto upnout;
469         }
470         
471         /* Wasn't a dnode. Try dirent */
472         for (fd = f->dents; fd; fd=fd->next) {
473                 if (fd->raw == raw)
474                         break;
475         }
476
477         if (fd && fd->ino) {
478                 ret = jffs2_garbage_collect_dirent(c, jeb, f, fd);
479         } else if (fd) {
480                 ret = jffs2_garbage_collect_deletion_dirent(c, jeb, f, fd);
481         } else {
482                 printk(KERN_WARNING "Raw node at 0x%08x wasn't in node lists for ino #%u\n",
483                        ref_offset(raw), f->inocache->ino);
484                 if (ref_obsolete(raw)) {
485                         printk(KERN_WARNING "But it's obsolete so we don't mind too much\n");
486                 } else {
487                         jffs2_dbg_dump_node(c, ref_offset(raw));
488                         BUG();
489                 }
490         }
491  upnout:
492         up(&f->sem);
493
494         return ret;
495 }
496
497 static int jffs2_garbage_collect_pristine(struct jffs2_sb_info *c, 
498                                           struct jffs2_inode_cache *ic,
499                                           struct jffs2_raw_node_ref *raw)
500 {
501         union jffs2_node_union *node;
502         struct jffs2_raw_node_ref *nraw;
503         size_t retlen;
504         int ret;
505         uint32_t phys_ofs, alloclen;
506         uint32_t crc, rawlen;
507         int retried = 0;
508
509         D1(printk(KERN_DEBUG "Going to GC REF_PRISTINE node at 0x%08x\n", ref_offset(raw)));
510
511         rawlen = ref_totlen(c, c->gcblock, raw);
512
513         /* Ask for a small amount of space (or the totlen if smaller) because we
514            don't want to force wastage of the end of a block if splitting would
515            work. */
516         ret = jffs2_reserve_space_gc(c, min_t(uint32_t, sizeof(struct jffs2_raw_inode) + JFFS2_MIN_DATA_LEN, 
517                                               rawlen), &phys_ofs, &alloclen);
518         if (ret)
519                 return ret;
520
521         if (alloclen < rawlen) {
522                 /* Doesn't fit untouched. We'll go the old route and split it */
523                 return -EBADFD;
524         }
525
526         node = kmalloc(rawlen, GFP_KERNEL);
527         if (!node)
528                return -ENOMEM;
529
530         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)node);
531         if (!ret && retlen != rawlen)
532                 ret = -EIO;
533         if (ret)
534                 goto out_node;
535
536         crc = crc32(0, node, sizeof(struct jffs2_unknown_node)-4);
537         if (je32_to_cpu(node->u.hdr_crc) != crc) {
538                 printk(KERN_WARNING "Header CRC failed on REF_PRISTINE node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
539                        ref_offset(raw), je32_to_cpu(node->u.hdr_crc), crc);
540                 goto bail;
541         }
542
543         switch(je16_to_cpu(node->u.nodetype)) {
544         case JFFS2_NODETYPE_INODE:
545                 crc = crc32(0, node, sizeof(node->i)-8);
546                 if (je32_to_cpu(node->i.node_crc) != crc) {
547                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
548                                ref_offset(raw), je32_to_cpu(node->i.node_crc), crc);
549                         goto bail;
550                 }
551
552                 if (je32_to_cpu(node->i.dsize)) {
553                         crc = crc32(0, node->i.data, je32_to_cpu(node->i.csize));
554                         if (je32_to_cpu(node->i.data_crc) != crc) {
555                                 printk(KERN_WARNING "Data CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
556                                        ref_offset(raw), je32_to_cpu(node->i.data_crc), crc);
557                                 goto bail;
558                         }
559                 }
560                 break;
561
562         case JFFS2_NODETYPE_DIRENT:
563                 crc = crc32(0, node, sizeof(node->d)-8);
564                 if (je32_to_cpu(node->d.node_crc) != crc) {
565                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE dirent node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
566                                ref_offset(raw), je32_to_cpu(node->d.node_crc), crc);
567                         goto bail;
568                 }
569
570                 if (node->d.nsize) {
571                         crc = crc32(0, node->d.name, node->d.nsize);
572                         if (je32_to_cpu(node->d.name_crc) != crc) {
573                                 printk(KERN_WARNING "Name CRC failed on REF_PRISTINE dirent ode at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
574                                        ref_offset(raw), je32_to_cpu(node->d.name_crc), crc);
575                                 goto bail;
576                         }
577                 }
578                 break;
579         default:
580                 printk(KERN_WARNING "Unknown node type for REF_PRISTINE node at 0x%08x: 0x%04x\n", 
581                        ref_offset(raw), je16_to_cpu(node->u.nodetype));
582                 goto bail;
583         }
584
585         nraw = jffs2_alloc_raw_node_ref();
586         if (!nraw) {
587                 ret = -ENOMEM;
588                 goto out_node;
589         }
590
591         /* OK, all the CRCs are good; this node can just be copied as-is. */
592  retry:
593         nraw->flash_offset = phys_ofs;
594         nraw->__totlen = rawlen;
595         nraw->next_phys = NULL;
596
597         ret = jffs2_flash_write(c, phys_ofs, rawlen, &retlen, (char *)node);
598
599         if (ret || (retlen != rawlen)) {
600                 printk(KERN_NOTICE "Write of %d bytes at 0x%08x failed. returned %d, retlen %zd\n",
601                        rawlen, phys_ofs, ret, retlen);
602                 if (retlen) {
603                         /* Doesn't belong to any inode */
604                         nraw->next_in_ino = NULL;
605
606                         nraw->flash_offset |= REF_OBSOLETE;
607                         jffs2_add_physical_node_ref(c, nraw);
608                         jffs2_mark_node_obsolete(c, nraw);
609                 } else {
610                         printk(KERN_NOTICE "Not marking the space at 0x%08x as dirty because the flash driver returned retlen zero\n", nraw->flash_offset);
611                         jffs2_free_raw_node_ref(nraw);
612                 }
613                 if (!retried && (nraw = jffs2_alloc_raw_node_ref())) {
614                         /* Try to reallocate space and retry */
615                         uint32_t dummy;
616                         struct jffs2_eraseblock *jeb = &c->blocks[phys_ofs / c->sector_size];
617
618                         retried = 1;
619
620                         D1(printk(KERN_DEBUG "Retrying failed write of REF_PRISTINE node.\n"));
621                         
622                         jffs2_dbg_acct_sanity_check(c,jeb);
623                         jffs2_dbg_acct_paranoia_check(c, jeb);
624
625                         ret = jffs2_reserve_space_gc(c, rawlen, &phys_ofs, &dummy);
626
627                         if (!ret) {
628                                 D1(printk(KERN_DEBUG "Allocated space at 0x%08x to retry failed write.\n", phys_ofs));
629
630                                 jffs2_dbg_acct_sanity_check(c,jeb);
631                                 jffs2_dbg_acct_paranoia_check(c, jeb);
632
633                                 goto retry;
634                         }
635                         D1(printk(KERN_DEBUG "Failed to allocate space to retry failed write: %d!\n", ret));
636                         jffs2_free_raw_node_ref(nraw);
637                 }
638
639                 jffs2_free_raw_node_ref(nraw);
640                 if (!ret)
641                         ret = -EIO;
642                 goto out_node;
643         }
644         nraw->flash_offset |= REF_PRISTINE;
645         jffs2_add_physical_node_ref(c, nraw);
646
647         /* Link into per-inode list. This is safe because of the ic
648            state being INO_STATE_GC. Note that if we're doing this
649            for an inode which is in-core, the 'nraw' pointer is then
650            going to be fetched from ic->nodes by our caller. */
651         spin_lock(&c->erase_completion_lock);
652         nraw->next_in_ino = ic->nodes;
653         ic->nodes = nraw;
654         spin_unlock(&c->erase_completion_lock);
655
656         jffs2_mark_node_obsolete(c, raw);
657         D1(printk(KERN_DEBUG "WHEEE! GC REF_PRISTINE node at 0x%08x succeeded\n", ref_offset(raw)));
658
659  out_node:
660         kfree(node);
661         return ret;
662  bail:
663         ret = -EBADFD;
664         goto out_node;
665 }
666
667 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb, 
668                                         struct jffs2_inode_info *f, struct jffs2_full_dnode *fn)
669 {
670         struct jffs2_full_dnode *new_fn;
671         struct jffs2_raw_inode ri;
672         struct jffs2_node_frag *last_frag;
673         jint16_t dev;
674         char *mdata = NULL, mdatalen = 0;
675         uint32_t alloclen, phys_ofs, ilen;
676         int ret;
677
678         if (S_ISBLK(JFFS2_F_I_MODE(f)) ||
679             S_ISCHR(JFFS2_F_I_MODE(f)) ) {
680                 /* For these, we don't actually need to read the old node */
681                 /* FIXME: for minor or major > 255. */
682                 dev = cpu_to_je16(((JFFS2_F_I_RDEV_MAJ(f) << 8) | 
683                         JFFS2_F_I_RDEV_MIN(f)));
684                 mdata = (char *)&dev;
685                 mdatalen = sizeof(dev);
686                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bytes of kdev_t\n", mdatalen));
687         } else if (S_ISLNK(JFFS2_F_I_MODE(f))) {
688                 mdatalen = fn->size;
689                 mdata = kmalloc(fn->size, GFP_KERNEL);
690                 if (!mdata) {
691                         printk(KERN_WARNING "kmalloc of mdata failed in jffs2_garbage_collect_metadata()\n");
692                         return -ENOMEM;
693                 }
694                 ret = jffs2_read_dnode(c, f, fn, mdata, 0, mdatalen);
695                 if (ret) {
696                         printk(KERN_WARNING "read of old metadata failed in jffs2_garbage_collect_metadata(): %d\n", ret);
697                         kfree(mdata);
698                         return ret;
699                 }
700                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bites of symlink target\n", mdatalen));
701
702         }
703         
704         ret = jffs2_reserve_space_gc(c, sizeof(ri) + mdatalen, &phys_ofs, &alloclen);
705         if (ret) {
706                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_metadata failed: %d\n",
707                        sizeof(ri)+ mdatalen, ret);
708                 goto out;
709         }
710         
711         last_frag = frag_last(&f->fragtree);
712         if (last_frag)
713                 /* Fetch the inode length from the fragtree rather then
714                  * from i_size since i_size may have not been updated yet */
715                 ilen = last_frag->ofs + last_frag->size;
716         else
717                 ilen = JFFS2_F_I_SIZE(f);
718         
719         memset(&ri, 0, sizeof(ri));
720         ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
721         ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
722         ri.totlen = cpu_to_je32(sizeof(ri) + mdatalen);
723         ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
724
725         ri.ino = cpu_to_je32(f->inocache->ino);
726         ri.version = cpu_to_je32(++f->highest_version);
727         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
728         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
729         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
730         ri.isize = cpu_to_je32(ilen);
731         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
732         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
733         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
734         ri.offset = cpu_to_je32(0);
735         ri.csize = cpu_to_je32(mdatalen);
736         ri.dsize = cpu_to_je32(mdatalen);
737         ri.compr = JFFS2_COMPR_NONE;
738         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
739         ri.data_crc = cpu_to_je32(crc32(0, mdata, mdatalen));
740
741         new_fn = jffs2_write_dnode(c, f, &ri, mdata, mdatalen, phys_ofs, ALLOC_GC);
742
743         if (IS_ERR(new_fn)) {
744                 printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
745                 ret = PTR_ERR(new_fn);
746                 goto out;
747         }
748         jffs2_mark_node_obsolete(c, fn->raw);
749         jffs2_free_full_dnode(fn);
750         f->metadata = new_fn;
751  out:
752         if (S_ISLNK(JFFS2_F_I_MODE(f)))
753                 kfree(mdata);
754         return ret;
755 }
756
757 static int jffs2_garbage_collect_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb, 
758                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
759 {
760         struct jffs2_full_dirent *new_fd;
761         struct jffs2_raw_dirent rd;
762         uint32_t alloclen, phys_ofs;
763         int ret;
764
765         rd.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
766         rd.nodetype = cpu_to_je16(JFFS2_NODETYPE_DIRENT);
767         rd.nsize = strlen(fd->name);
768         rd.totlen = cpu_to_je32(sizeof(rd) + rd.nsize);
769         rd.hdr_crc = cpu_to_je32(crc32(0, &rd, sizeof(struct jffs2_unknown_node)-4));
770
771         rd.pino = cpu_to_je32(f->inocache->ino);
772         rd.version = cpu_to_je32(++f->highest_version);
773         rd.ino = cpu_to_je32(fd->ino);
774         /* If the times on this inode were set by explicit utime() they can be different,
775            so refrain from splatting them. */
776         if (JFFS2_F_I_MTIME(f) == JFFS2_F_I_CTIME(f))
777                 rd.mctime = cpu_to_je32(JFFS2_F_I_MTIME(f));
778         else 
779                 rd.mctime = cpu_to_je32(0);
780         rd.type = fd->type;
781         rd.node_crc = cpu_to_je32(crc32(0, &rd, sizeof(rd)-8));
782         rd.name_crc = cpu_to_je32(crc32(0, fd->name, rd.nsize));
783         
784         ret = jffs2_reserve_space_gc(c, sizeof(rd)+rd.nsize, &phys_ofs, &alloclen);
785         if (ret) {
786                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dirent failed: %d\n",
787                        sizeof(rd)+rd.nsize, ret);
788                 return ret;
789         }
790         new_fd = jffs2_write_dirent(c, f, &rd, fd->name, rd.nsize, phys_ofs, ALLOC_GC);
791
792         if (IS_ERR(new_fd)) {
793                 printk(KERN_WARNING "jffs2_write_dirent in garbage_collect_dirent failed: %ld\n", PTR_ERR(new_fd));
794                 return PTR_ERR(new_fd);
795         }
796         jffs2_add_fd_to_list(c, new_fd, &f->dents);
797         return 0;
798 }
799
800 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb, 
801                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
802 {
803         struct jffs2_full_dirent **fdp = &f->dents;
804         int found = 0;
805
806         /* On a medium where we can't actually mark nodes obsolete
807            pernamently, such as NAND flash, we need to work out
808            whether this deletion dirent is still needed to actively
809            delete a 'real' dirent with the same name that's still
810            somewhere else on the flash. */
811         if (!jffs2_can_mark_obsolete(c)) {
812                 struct jffs2_raw_dirent *rd;
813                 struct jffs2_raw_node_ref *raw;
814                 int ret;
815                 size_t retlen;
816                 int name_len = strlen(fd->name);
817                 uint32_t name_crc = crc32(0, fd->name, name_len);
818                 uint32_t rawlen = ref_totlen(c, jeb, fd->raw);
819
820                 rd = kmalloc(rawlen, GFP_KERNEL);
821                 if (!rd)
822                         return -ENOMEM;
823
824                 /* Prevent the erase code from nicking the obsolete node refs while
825                    we're looking at them. I really don't like this extra lock but
826                    can't see any alternative. Suggestions on a postcard to... */
827                 down(&c->erase_free_sem);
828
829                 for (raw = f->inocache->nodes; raw != (void *)f->inocache; raw = raw->next_in_ino) {
830
831                         /* We only care about obsolete ones */
832                         if (!(ref_obsolete(raw)))
833                                 continue;
834
835                         /* Any dirent with the same name is going to have the same length... */
836                         if (ref_totlen(c, NULL, raw) != rawlen)
837                                 continue;
838
839                         /* Doesn't matter if there's one in the same erase block. We're going to 
840                            delete it too at the same time. */
841                         if (SECTOR_ADDR(raw->flash_offset) == SECTOR_ADDR(fd->raw->flash_offset))
842                                 continue;
843
844                         D1(printk(KERN_DEBUG "Check potential deletion dirent at %08x\n", ref_offset(raw)));
845
846                         /* This is an obsolete node belonging to the same directory, and it's of the right
847                            length. We need to take a closer look...*/
848                         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)rd);
849                         if (ret) {
850                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Read error (%d) reading obsolete node at %08x\n", ret, ref_offset(raw));
851                                 /* If we can't read it, we don't need to continue to obsolete it. Continue */
852                                 continue;
853                         }
854                         if (retlen != rawlen) {
855                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Short read (%zd not %u) reading header from obsolete node at %08x\n",
856                                        retlen, rawlen, ref_offset(raw));
857                                 continue;
858                         }
859
860                         if (je16_to_cpu(rd->nodetype) != JFFS2_NODETYPE_DIRENT)
861                                 continue;
862
863                         /* If the name CRC doesn't match, skip */
864                         if (je32_to_cpu(rd->name_crc) != name_crc)
865                                 continue;
866
867                         /* If the name length doesn't match, or it's another deletion dirent, skip */
868                         if (rd->nsize != name_len || !je32_to_cpu(rd->ino))
869                                 continue;
870
871                         /* OK, check the actual name now */
872                         if (memcmp(rd->name, fd->name, name_len))
873                                 continue;
874
875                         /* OK. The name really does match. There really is still an older node on
876                            the flash which our deletion dirent obsoletes. So we have to write out
877                            a new deletion dirent to replace it */
878                         up(&c->erase_free_sem);
879
880                         D1(printk(KERN_DEBUG "Deletion dirent at %08x still obsoletes real dirent \"%s\" at %08x for ino #%u\n",
881                                   ref_offset(fd->raw), fd->name, ref_offset(raw), je32_to_cpu(rd->ino)));
882                         kfree(rd);
883
884                         return jffs2_garbage_collect_dirent(c, jeb, f, fd);
885                 }
886
887                 up(&c->erase_free_sem);
888                 kfree(rd);
889         }
890
891         /* FIXME: If we're deleting a dirent which contains the current mtime and ctime, 
892            we should update the metadata node with those times accordingly */
893
894         /* No need for it any more. Just mark it obsolete and remove it from the list */
895         while (*fdp) {
896                 if ((*fdp) == fd) {
897                         found = 1;
898                         *fdp = fd->next;
899                         break;
900                 }
901                 fdp = &(*fdp)->next;
902         }
903         if (!found) {
904                 printk(KERN_WARNING "Deletion dirent \"%s\" not found in list for ino #%u\n", fd->name, f->inocache->ino);
905         }
906         jffs2_mark_node_obsolete(c, fd->raw);
907         jffs2_free_full_dirent(fd);
908         return 0;
909 }
910
911 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
912                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
913                                       uint32_t start, uint32_t end)
914 {
915         struct jffs2_raw_inode ri;
916         struct jffs2_node_frag *frag;
917         struct jffs2_full_dnode *new_fn;
918         uint32_t alloclen, phys_ofs, ilen;
919         int ret;
920
921         D1(printk(KERN_DEBUG "Writing replacement hole node for ino #%u from offset 0x%x to 0x%x\n",
922                   f->inocache->ino, start, end));
923         
924         memset(&ri, 0, sizeof(ri));
925
926         if(fn->frags > 1) {
927                 size_t readlen;
928                 uint32_t crc;
929                 /* It's partially obsoleted by a later write. So we have to 
930                    write it out again with the _same_ version as before */
931                 ret = jffs2_flash_read(c, ref_offset(fn->raw), sizeof(ri), &readlen, (char *)&ri);
932                 if (readlen != sizeof(ri) || ret) {
933                         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);
934                         goto fill;
935                 }
936                 if (je16_to_cpu(ri.nodetype) != JFFS2_NODETYPE_INODE) {
937                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had node type 0x%04x instead of JFFS2_NODETYPE_INODE(0x%04x)\n",
938                                ref_offset(fn->raw),
939                                je16_to_cpu(ri.nodetype), JFFS2_NODETYPE_INODE);
940                         return -EIO;
941                 }
942                 if (je32_to_cpu(ri.totlen) != sizeof(ri)) {
943                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had totlen 0x%x instead of expected 0x%zx\n",
944                                ref_offset(fn->raw),
945                                je32_to_cpu(ri.totlen), sizeof(ri));
946                         return -EIO;
947                 }
948                 crc = crc32(0, &ri, sizeof(ri)-8);
949                 if (crc != je32_to_cpu(ri.node_crc)) {
950                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had CRC 0x%08x which doesn't match calculated CRC 0x%08x\n",
951                                ref_offset(fn->raw), 
952                                je32_to_cpu(ri.node_crc), crc);
953                         /* FIXME: We could possibly deal with this by writing new holes for each frag */
954                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n", 
955                                start, end, f->inocache->ino);
956                         goto fill;
957                 }
958                 if (ri.compr != JFFS2_COMPR_ZERO) {
959                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node 0x%08x wasn't a hole node!\n", ref_offset(fn->raw));
960                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n", 
961                                start, end, f->inocache->ino);
962                         goto fill;
963                 }
964         } else {
965         fill:
966                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
967                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
968                 ri.totlen = cpu_to_je32(sizeof(ri));
969                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
970
971                 ri.ino = cpu_to_je32(f->inocache->ino);
972                 ri.version = cpu_to_je32(++f->highest_version);
973                 ri.offset = cpu_to_je32(start);
974                 ri.dsize = cpu_to_je32(end - start);
975                 ri.csize = cpu_to_je32(0);
976                 ri.compr = JFFS2_COMPR_ZERO;
977         }
978         
979         frag = frag_last(&f->fragtree);
980         if (frag)
981                 /* Fetch the inode length from the fragtree rather then
982                  * from i_size since i_size may have not been updated yet */
983                 ilen = frag->ofs + frag->size;
984         else
985                 ilen = JFFS2_F_I_SIZE(f);
986
987         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
988         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
989         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
990         ri.isize = cpu_to_je32(ilen);
991         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
992         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
993         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
994         ri.data_crc = cpu_to_je32(0);
995         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
996
997         ret = jffs2_reserve_space_gc(c, sizeof(ri), &phys_ofs, &alloclen);
998         if (ret) {
999                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_hole failed: %d\n",
1000                        sizeof(ri), ret);
1001                 return ret;
1002         }
1003         new_fn = jffs2_write_dnode(c, f, &ri, NULL, 0, phys_ofs, ALLOC_GC);
1004
1005         if (IS_ERR(new_fn)) {
1006                 printk(KERN_WARNING "Error writing new hole node: %ld\n", PTR_ERR(new_fn));
1007                 return PTR_ERR(new_fn);
1008         }
1009         if (je32_to_cpu(ri.version) == f->highest_version) {
1010                 jffs2_add_full_dnode_to_inode(c, f, new_fn);
1011                 if (f->metadata) {
1012                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1013                         jffs2_free_full_dnode(f->metadata);
1014                         f->metadata = NULL;
1015                 }
1016                 return 0;
1017         }
1018
1019         /* 
1020          * We should only get here in the case where the node we are
1021          * replacing had more than one frag, so we kept the same version
1022          * number as before. (Except in case of error -- see 'goto fill;' 
1023          * above.)
1024          */
1025         D1(if(unlikely(fn->frags <= 1)) {
1026                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Replacing fn with %d frag(s) but new ver %d != highest_version %d of ino #%d\n",
1027                        fn->frags, je32_to_cpu(ri.version), f->highest_version,
1028                        je32_to_cpu(ri.ino));
1029         });
1030
1031         /* This is a partially-overlapped hole node. Mark it REF_NORMAL not REF_PRISTINE */
1032         mark_ref_normal(new_fn->raw);
1033
1034         for (frag = jffs2_lookup_node_frag(&f->fragtree, fn->ofs); 
1035              frag; frag = frag_next(frag)) {
1036                 if (frag->ofs > fn->size + fn->ofs)
1037                         break;
1038                 if (frag->node == fn) {
1039                         frag->node = new_fn;
1040                         new_fn->frags++;
1041                         fn->frags--;
1042                 }
1043         }
1044         if (fn->frags) {
1045                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Old node still has frags!\n");
1046                 BUG();
1047         }
1048         if (!new_fn->frags) {
1049                 printk(KERN_WARNING "jffs2_garbage_collect_hole: New node has no frags!\n");
1050                 BUG();
1051         }
1052                 
1053         jffs2_mark_node_obsolete(c, fn->raw);
1054         jffs2_free_full_dnode(fn);
1055         
1056         return 0;
1057 }
1058
1059 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
1060                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
1061                                        uint32_t start, uint32_t end)
1062 {
1063         struct jffs2_full_dnode *new_fn;
1064         struct jffs2_raw_inode ri;
1065         uint32_t alloclen, phys_ofs, offset, orig_end, orig_start;      
1066         int ret = 0;
1067         unsigned char *comprbuf = NULL, *writebuf;
1068         unsigned long pg;
1069         unsigned char *pg_ptr;
1070  
1071         memset(&ri, 0, sizeof(ri));
1072
1073         D1(printk(KERN_DEBUG "Writing replacement dnode for ino #%u from offset 0x%x to 0x%x\n",
1074                   f->inocache->ino, start, end));
1075
1076         orig_end = end;
1077         orig_start = start;
1078
1079         if (c->nr_free_blocks + c->nr_erasing_blocks > c->resv_blocks_gcmerge) {
1080                 /* Attempt to do some merging. But only expand to cover logically
1081                    adjacent frags if the block containing them is already considered
1082                    to be dirty. Otherwise we end up with GC just going round in 
1083                    circles dirtying the nodes it already wrote out, especially 
1084                    on NAND where we have small eraseblocks and hence a much higher
1085                    chance of nodes having to be split to cross boundaries. */
1086
1087                 struct jffs2_node_frag *frag;
1088                 uint32_t min, max;
1089
1090                 min = start & ~(PAGE_CACHE_SIZE-1);
1091                 max = min + PAGE_CACHE_SIZE;
1092
1093                 frag = jffs2_lookup_node_frag(&f->fragtree, start);
1094
1095                 /* BUG_ON(!frag) but that'll happen anyway... */
1096
1097                 BUG_ON(frag->ofs != start);
1098
1099                 /* First grow down... */
1100                 while((frag = frag_prev(frag)) && frag->ofs >= min) {
1101
1102                         /* If the previous frag doesn't even reach the beginning, there's
1103                            excessive fragmentation. Just merge. */
1104                         if (frag->ofs > min) {
1105                                 D1(printk(KERN_DEBUG "Expanding down to cover partial frag (0x%x-0x%x)\n",
1106                                           frag->ofs, frag->ofs+frag->size));
1107                                 start = frag->ofs;
1108                                 continue;
1109                         }
1110                         /* OK. This frag holds the first byte of the page. */
1111                         if (!frag->node || !frag->node->raw) {
1112                                 D1(printk(KERN_DEBUG "First frag in page is hole (0x%x-0x%x). Not expanding down.\n",
1113                                           frag->ofs, frag->ofs+frag->size));
1114                                 break;
1115                         } else {
1116
1117                                 /* OK, it's a frag which extends to the beginning of the page. Does it live 
1118                                    in a block which is still considered clean? If so, don't obsolete it.
1119                                    If not, cover it anyway. */
1120
1121                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1122                                 struct jffs2_eraseblock *jeb;
1123
1124                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1125
1126                                 if (jeb == c->gcblock) {
1127                                         D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1128                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1129                                         start = frag->ofs;
1130                                         break;
1131                                 }
1132                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1133                                         D1(printk(KERN_DEBUG "Not expanding down to cover frag (0x%x-0x%x) in clean block %08x\n",
1134                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1135                                         break;
1136                                 }
1137
1138                                 D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in dirty block %08x\n",
1139                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1140                                 start = frag->ofs;
1141                                 break;
1142                         }
1143                 }
1144
1145                 /* ... then up */
1146
1147                 /* Find last frag which is actually part of the node we're to GC. */
1148                 frag = jffs2_lookup_node_frag(&f->fragtree, end-1);
1149
1150                 while((frag = frag_next(frag)) && frag->ofs+frag->size <= max) {
1151
1152                         /* If the previous frag doesn't even reach the beginning, there's lots
1153                            of fragmentation. Just merge. */
1154                         if (frag->ofs+frag->size < max) {
1155                                 D1(printk(KERN_DEBUG "Expanding up to cover partial frag (0x%x-0x%x)\n",
1156                                           frag->ofs, frag->ofs+frag->size));
1157                                 end = frag->ofs + frag->size;
1158                                 continue;
1159                         }
1160
1161                         if (!frag->node || !frag->node->raw) {
1162                                 D1(printk(KERN_DEBUG "Last frag in page is hole (0x%x-0x%x). Not expanding up.\n",
1163                                           frag->ofs, frag->ofs+frag->size));
1164                                 break;
1165                         } else {
1166
1167                                 /* OK, it's a frag which extends to the beginning of the page. Does it live 
1168                                    in a block which is still considered clean? If so, don't obsolete it.
1169                                    If not, cover it anyway. */
1170
1171                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1172                                 struct jffs2_eraseblock *jeb;
1173
1174                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1175
1176                                 if (jeb == c->gcblock) {
1177                                         D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1178                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1179                                         end = frag->ofs + frag->size;
1180                                         break;
1181                                 }
1182                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1183                                         D1(printk(KERN_DEBUG "Not expanding up to cover frag (0x%x-0x%x) in clean block %08x\n",
1184                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1185                                         break;
1186                                 }
1187
1188                                 D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in dirty block %08x\n",
1189                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1190                                 end = frag->ofs + frag->size;
1191                                 break;
1192                         }
1193                 }
1194                 D1(printk(KERN_DEBUG "Expanded dnode to write from (0x%x-0x%x) to (0x%x-0x%x)\n", 
1195                           orig_start, orig_end, start, end));
1196
1197                 D1(BUG_ON(end > frag_last(&f->fragtree)->ofs + frag_last(&f->fragtree)->size));
1198                 BUG_ON(end < orig_end);
1199                 BUG_ON(start > orig_start);
1200         }
1201         
1202         /* First, use readpage() to read the appropriate page into the page cache */
1203         /* Q: What happens if we actually try to GC the _same_ page for which commit_write()
1204          *    triggered garbage collection in the first place?
1205          * A: I _think_ it's OK. read_cache_page shouldn't deadlock, we'll write out the
1206          *    page OK. We'll actually write it out again in commit_write, which is a little
1207          *    suboptimal, but at least we're correct.
1208          */
1209         pg_ptr = jffs2_gc_fetch_page(c, f, start, &pg);
1210
1211         if (IS_ERR(pg_ptr)) {
1212                 printk(KERN_WARNING "read_cache_page() returned error: %ld\n", PTR_ERR(pg_ptr));
1213                 return PTR_ERR(pg_ptr);
1214         }
1215
1216         offset = start;
1217         while(offset < orig_end) {
1218                 uint32_t datalen;
1219                 uint32_t cdatalen;
1220                 uint16_t comprtype = JFFS2_COMPR_NONE;
1221
1222                 ret = jffs2_reserve_space_gc(c, sizeof(ri) + JFFS2_MIN_DATA_LEN, &phys_ofs, &alloclen);
1223
1224                 if (ret) {
1225                         printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dnode failed: %d\n",
1226                                sizeof(ri)+ JFFS2_MIN_DATA_LEN, ret);
1227                         break;
1228                 }
1229                 cdatalen = min_t(uint32_t, alloclen - sizeof(ri), end - offset);
1230                 datalen = end - offset;
1231
1232                 writebuf = pg_ptr + (offset & (PAGE_CACHE_SIZE -1));
1233
1234                 comprtype = jffs2_compress(c, f, writebuf, &comprbuf, &datalen, &cdatalen);
1235
1236                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
1237                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
1238                 ri.totlen = cpu_to_je32(sizeof(ri) + cdatalen);
1239                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
1240
1241                 ri.ino = cpu_to_je32(f->inocache->ino);
1242                 ri.version = cpu_to_je32(++f->highest_version);
1243                 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1244                 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1245                 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1246                 ri.isize = cpu_to_je32(JFFS2_F_I_SIZE(f));
1247                 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1248                 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1249                 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1250                 ri.offset = cpu_to_je32(offset);
1251                 ri.csize = cpu_to_je32(cdatalen);
1252                 ri.dsize = cpu_to_je32(datalen);
1253                 ri.compr = comprtype & 0xff;
1254                 ri.usercompr = (comprtype >> 8) & 0xff;
1255                 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1256                 ri.data_crc = cpu_to_je32(crc32(0, comprbuf, cdatalen));
1257         
1258                 new_fn = jffs2_write_dnode(c, f, &ri, comprbuf, cdatalen, phys_ofs, ALLOC_GC);
1259
1260                 jffs2_free_comprbuf(comprbuf, writebuf);
1261
1262                 if (IS_ERR(new_fn)) {
1263                         printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
1264                         ret = PTR_ERR(new_fn);
1265                         break;
1266                 }
1267                 ret = jffs2_add_full_dnode_to_inode(c, f, new_fn);
1268                 offset += datalen;
1269                 if (f->metadata) {
1270                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1271                         jffs2_free_full_dnode(f->metadata);
1272                         f->metadata = NULL;
1273                 }
1274         }
1275
1276         jffs2_gc_release_page(c, pg_ptr, &pg);
1277         return ret;
1278 }
1279