6b5da0a1d319d6059ab1860ab7cc13e3f557dd40
[linux-2.6.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 1.150 2005/07/17 12:01:43 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         jffs2_dbg_dump_block_lists(c);
115         return ret;
116 }
117
118 /* jffs2_garbage_collect_pass
119  * Make a single attempt to progress GC. Move one node, and possibly
120  * start erasing one eraseblock.
121  */
122 int jffs2_garbage_collect_pass(struct jffs2_sb_info *c)
123 {
124         struct jffs2_inode_info *f;
125         struct jffs2_inode_cache *ic;
126         struct jffs2_eraseblock *jeb;
127         struct jffs2_raw_node_ref *raw;
128         int ret = 0, inum, nlink;
129
130         if (down_interruptible(&c->alloc_sem))
131                 return -EINTR;
132
133         for (;;) {
134                 spin_lock(&c->erase_completion_lock);
135                 if (!c->unchecked_size)
136                         break;
137
138                 /* We can't start doing GC yet. We haven't finished checking
139                    the node CRCs etc. Do it now. */
140                 
141                 /* checked_ino is protected by the alloc_sem */
142                 if (c->checked_ino > c->highest_ino) {
143                         printk(KERN_CRIT "Checked all inodes but still 0x%x bytes of unchecked space?\n",
144                                c->unchecked_size);
145                         jffs2_dbg_dump_block_lists(c);
146                         spin_unlock(&c->erase_completion_lock);
147                         BUG();
148                 }
149
150                 spin_unlock(&c->erase_completion_lock);
151
152                 spin_lock(&c->inocache_lock);
153
154                 ic = jffs2_get_ino_cache(c, c->checked_ino++);
155
156                 if (!ic) {
157                         spin_unlock(&c->inocache_lock);
158                         continue;
159                 }
160
161                 if (!ic->nlink) {
162                         D1(printk(KERN_DEBUG "Skipping check of ino #%d with nlink zero\n",
163                                   ic->ino));
164                         spin_unlock(&c->inocache_lock);
165                         continue;
166                 }
167                 switch(ic->state) {
168                 case INO_STATE_CHECKEDABSENT:
169                 case INO_STATE_PRESENT:
170                         D1(printk(KERN_DEBUG "Skipping ino #%u already checked\n", ic->ino));
171                         spin_unlock(&c->inocache_lock);
172                         continue;
173
174                 case INO_STATE_GC:
175                 case INO_STATE_CHECKING:
176                         printk(KERN_WARNING "Inode #%u is in state %d during CRC check phase!\n", ic->ino, ic->state);
177                         spin_unlock(&c->inocache_lock);
178                         BUG();
179
180                 case INO_STATE_READING:
181                         /* We need to wait for it to finish, lest we move on
182                            and trigger the BUG() above while we haven't yet 
183                            finished checking all its nodes */
184                         D1(printk(KERN_DEBUG "Waiting for ino #%u to finish reading\n", ic->ino));
185                         up(&c->alloc_sem);
186                         sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
187                         return 0;
188
189                 default:
190                         BUG();
191
192                 case INO_STATE_UNCHECKED:
193                         ;
194                 }
195                 ic->state = INO_STATE_CHECKING;
196                 spin_unlock(&c->inocache_lock);
197
198                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() triggering inode scan of ino#%u\n", ic->ino));
199
200                 ret = jffs2_do_crccheck_inode(c, ic);
201                 if (ret)
202                         printk(KERN_WARNING "Returned error for crccheck of ino #%u. Expect badness...\n", ic->ino);
203
204                 jffs2_set_inocache_state(c, ic, INO_STATE_CHECKEDABSENT);
205                 up(&c->alloc_sem);
206                 return ret;
207         }
208
209         /* First, work out which block we're garbage-collecting */
210         jeb = c->gcblock;
211
212         if (!jeb)
213                 jeb = jffs2_find_gc_block(c);
214
215         if (!jeb) {
216                 D1 (printk(KERN_NOTICE "jffs2: Couldn't find erase block to garbage collect!\n"));
217                 spin_unlock(&c->erase_completion_lock);
218                 up(&c->alloc_sem);
219                 return -EIO;
220         }
221
222         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));
223         D1(if (c->nextblock)
224            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));
225
226         if (!jeb->used_size) {
227                 up(&c->alloc_sem);
228                 goto eraseit;
229         }
230
231         raw = jeb->gc_node;
232                         
233         while(ref_obsolete(raw)) {
234                 D1(printk(KERN_DEBUG "Node at 0x%08x is obsolete... skipping\n", ref_offset(raw)));
235                 raw = raw->next_phys;
236                 if (unlikely(!raw)) {
237                         printk(KERN_WARNING "eep. End of raw list while still supposedly nodes to GC\n");
238                         printk(KERN_WARNING "erase block at 0x%08x. free_size 0x%08x, dirty_size 0x%08x, used_size 0x%08x\n", 
239                                jeb->offset, jeb->free_size, jeb->dirty_size, jeb->used_size);
240                         jeb->gc_node = raw;
241                         spin_unlock(&c->erase_completion_lock);
242                         up(&c->alloc_sem);
243                         BUG();
244                 }
245         }
246         jeb->gc_node = raw;
247
248         D1(printk(KERN_DEBUG "Going to garbage collect node at 0x%08x\n", ref_offset(raw)));
249
250         if (!raw->next_in_ino) {
251                 /* Inode-less node. Clean marker, snapshot or something like that */
252                 /* FIXME: If it's something that needs to be copied, including something
253                    we don't grok that has JFFS2_NODETYPE_RWCOMPAT_COPY, we should do so */
254                 spin_unlock(&c->erase_completion_lock);
255                 jffs2_mark_node_obsolete(c, raw);
256                 up(&c->alloc_sem);
257                 goto eraseit_lock;
258         }
259
260         ic = jffs2_raw_ref_to_ic(raw);
261
262         /* We need to hold the inocache. Either the erase_completion_lock or
263            the inocache_lock are sufficient; we trade down since the inocache_lock 
264            causes less contention. */
265         spin_lock(&c->inocache_lock);
266
267         spin_unlock(&c->erase_completion_lock);
268
269         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));
270
271         /* Three possibilities:
272            1. Inode is already in-core. We must iget it and do proper
273               updating to its fragtree, etc.
274            2. Inode is not in-core, node is REF_PRISTINE. We lock the
275               inocache to prevent a read_inode(), copy the node intact.
276            3. Inode is not in-core, node is not pristine. We must iget()
277               and take the slow path.
278         */
279
280         switch(ic->state) {
281         case INO_STATE_CHECKEDABSENT:
282                 /* It's been checked, but it's not currently in-core. 
283                    We can just copy any pristine nodes, but have
284                    to prevent anyone else from doing read_inode() while
285                    we're at it, so we set the state accordingly */
286                 if (ref_flags(raw) == REF_PRISTINE)
287                         ic->state = INO_STATE_GC;
288                 else {
289                         D1(printk(KERN_DEBUG "Ino #%u is absent but node not REF_PRISTINE. Reading.\n", 
290                                   ic->ino));
291                 }
292                 break;
293
294         case INO_STATE_PRESENT:
295                 /* It's in-core. GC must iget() it. */
296                 break;
297
298         case INO_STATE_UNCHECKED:
299         case INO_STATE_CHECKING:
300         case INO_STATE_GC:
301                 /* Should never happen. We should have finished checking
302                    by the time we actually start doing any GC, and since 
303                    we're holding the alloc_sem, no other garbage collection 
304                    can happen.
305                 */
306                 printk(KERN_CRIT "Inode #%u already in state %d in jffs2_garbage_collect_pass()!\n",
307                        ic->ino, ic->state);
308                 up(&c->alloc_sem);
309                 spin_unlock(&c->inocache_lock);
310                 BUG();
311
312         case INO_STATE_READING:
313                 /* Someone's currently trying to read it. We must wait for
314                    them to finish and then go through the full iget() route
315                    to do the GC. However, sometimes read_inode() needs to get
316                    the alloc_sem() (for marking nodes invalid) so we must
317                    drop the alloc_sem before sleeping. */
318
319                 up(&c->alloc_sem);
320                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() waiting for ino #%u in state %d\n",
321                           ic->ino, ic->state));
322                 sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
323                 /* And because we dropped the alloc_sem we must start again from the 
324                    beginning. Ponder chance of livelock here -- we're returning success
325                    without actually making any progress.
326
327                    Q: What are the chances that the inode is back in INO_STATE_READING 
328                    again by the time we next enter this function? And that this happens
329                    enough times to cause a real delay?
330
331                    A: Small enough that I don't care :) 
332                 */
333                 return 0;
334         }
335
336         /* OK. Now if the inode is in state INO_STATE_GC, we are going to copy the
337            node intact, and we don't have to muck about with the fragtree etc. 
338            because we know it's not in-core. If it _was_ in-core, we go through
339            all the iget() crap anyway */
340
341         if (ic->state == INO_STATE_GC) {
342                 spin_unlock(&c->inocache_lock);
343
344                 ret = jffs2_garbage_collect_pristine(c, ic, raw);
345
346                 spin_lock(&c->inocache_lock);
347                 ic->state = INO_STATE_CHECKEDABSENT;
348                 wake_up(&c->inocache_wq);
349
350                 if (ret != -EBADFD) {
351                         spin_unlock(&c->inocache_lock);
352                         goto release_sem;
353                 }
354
355                 /* Fall through if it wanted us to, with inocache_lock held */
356         }
357
358         /* Prevent the fairly unlikely race where the gcblock is
359            entirely obsoleted by the final close of a file which had
360            the only valid nodes in the block, followed by erasure,
361            followed by freeing of the ic because the erased block(s)
362            held _all_ the nodes of that inode.... never been seen but
363            it's vaguely possible. */
364
365         inum = ic->ino;
366         nlink = ic->nlink;
367         spin_unlock(&c->inocache_lock);
368
369         f = jffs2_gc_fetch_inode(c, inum, nlink);
370         if (IS_ERR(f)) {
371                 ret = PTR_ERR(f);
372                 goto release_sem;
373         }
374         if (!f) {
375                 ret = 0;
376                 goto release_sem;
377         }
378
379         ret = jffs2_garbage_collect_live(c, jeb, raw, f);
380
381         jffs2_gc_release_inode(c, f);
382
383  release_sem:
384         up(&c->alloc_sem);
385
386  eraseit_lock:
387         /* If we've finished this block, start it erasing */
388         spin_lock(&c->erase_completion_lock);
389
390  eraseit:
391         if (c->gcblock && !c->gcblock->used_size) {
392                 D1(printk(KERN_DEBUG "Block at 0x%08x completely obsoleted by GC. Moving to erase_pending_list\n", c->gcblock->offset));
393                 /* We're GC'ing an empty block? */
394                 list_add_tail(&c->gcblock->list, &c->erase_pending_list);
395                 c->gcblock = NULL;
396                 c->nr_erasing_blocks++;
397                 jffs2_erase_pending_trigger(c);
398         }
399         spin_unlock(&c->erase_completion_lock);
400
401         return ret;
402 }
403
404 static int jffs2_garbage_collect_live(struct jffs2_sb_info *c,  struct jffs2_eraseblock *jeb,
405                                       struct jffs2_raw_node_ref *raw, struct jffs2_inode_info *f)
406 {
407         struct jffs2_node_frag *frag;
408         struct jffs2_full_dnode *fn = NULL;
409         struct jffs2_full_dirent *fd;
410         uint32_t start = 0, end = 0, nrfrags = 0;
411         int ret = 0;
412
413         down(&f->sem);
414
415         /* Now we have the lock for this inode. Check that it's still the one at the head
416            of the list. */
417
418         spin_lock(&c->erase_completion_lock);
419
420         if (c->gcblock != jeb) {
421                 spin_unlock(&c->erase_completion_lock);
422                 D1(printk(KERN_DEBUG "GC block is no longer gcblock. Restart\n"));
423                 goto upnout;
424         }
425         if (ref_obsolete(raw)) {
426                 spin_unlock(&c->erase_completion_lock);
427                 D1(printk(KERN_DEBUG "node to be GC'd was obsoleted in the meantime.\n"));
428                 /* They'll call again */
429                 goto upnout;
430         }
431         spin_unlock(&c->erase_completion_lock);
432
433         /* OK. Looks safe. And nobody can get us now because we have the semaphore. Move the block */
434         if (f->metadata && f->metadata->raw == raw) {
435                 fn = f->metadata;
436                 ret = jffs2_garbage_collect_metadata(c, jeb, f, fn);
437                 goto upnout;
438         }
439
440         /* FIXME. Read node and do lookup? */
441         for (frag = frag_first(&f->fragtree); frag; frag = frag_next(frag)) {
442                 if (frag->node && frag->node->raw == raw) {
443                         fn = frag->node;
444                         end = frag->ofs + frag->size;
445                         if (!nrfrags++)
446                                 start = frag->ofs;
447                         if (nrfrags == frag->node->frags)
448                                 break; /* We've found them all */
449                 }
450         }
451         if (fn) {
452                 if (ref_flags(raw) == REF_PRISTINE) {
453                         ret = jffs2_garbage_collect_pristine(c, f->inocache, raw);
454                         if (!ret) {
455                                 /* Urgh. Return it sensibly. */
456                                 frag->node->raw = f->inocache->nodes;
457                         }       
458                         if (ret != -EBADFD)
459                                 goto upnout;
460                 }
461                 /* We found a datanode. Do the GC */
462                 if((start >> PAGE_CACHE_SHIFT) < ((end-1) >> PAGE_CACHE_SHIFT)) {
463                         /* It crosses a page boundary. Therefore, it must be a hole. */
464                         ret = jffs2_garbage_collect_hole(c, jeb, f, fn, start, end);
465                 } else {
466                         /* It could still be a hole. But we GC the page this way anyway */
467                         ret = jffs2_garbage_collect_dnode(c, jeb, f, fn, start, end);
468                 }
469                 goto upnout;
470         }
471         
472         /* Wasn't a dnode. Try dirent */
473         for (fd = f->dents; fd; fd=fd->next) {
474                 if (fd->raw == raw)
475                         break;
476         }
477
478         if (fd && fd->ino) {
479                 ret = jffs2_garbage_collect_dirent(c, jeb, f, fd);
480         } else if (fd) {
481                 ret = jffs2_garbage_collect_deletion_dirent(c, jeb, f, fd);
482         } else {
483                 printk(KERN_WARNING "Raw node at 0x%08x wasn't in node lists for ino #%u\n",
484                        ref_offset(raw), f->inocache->ino);
485                 if (ref_obsolete(raw)) {
486                         printk(KERN_WARNING "But it's obsolete so we don't mind too much\n");
487                 } else {
488                         ret = -EIO;
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         rd.mctime = cpu_to_je32(max(JFFS2_F_I_MTIME(f), JFFS2_F_I_CTIME(f)));
775         rd.type = fd->type;
776         rd.node_crc = cpu_to_je32(crc32(0, &rd, sizeof(rd)-8));
777         rd.name_crc = cpu_to_je32(crc32(0, fd->name, rd.nsize));
778         
779         ret = jffs2_reserve_space_gc(c, sizeof(rd)+rd.nsize, &phys_ofs, &alloclen);
780         if (ret) {
781                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dirent failed: %d\n",
782                        sizeof(rd)+rd.nsize, ret);
783                 return ret;
784         }
785         new_fd = jffs2_write_dirent(c, f, &rd, fd->name, rd.nsize, phys_ofs, ALLOC_GC);
786
787         if (IS_ERR(new_fd)) {
788                 printk(KERN_WARNING "jffs2_write_dirent in garbage_collect_dirent failed: %ld\n", PTR_ERR(new_fd));
789                 return PTR_ERR(new_fd);
790         }
791         jffs2_add_fd_to_list(c, new_fd, &f->dents);
792         return 0;
793 }
794
795 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb, 
796                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
797 {
798         struct jffs2_full_dirent **fdp = &f->dents;
799         int found = 0;
800
801         /* On a medium where we can't actually mark nodes obsolete
802            pernamently, such as NAND flash, we need to work out
803            whether this deletion dirent is still needed to actively
804            delete a 'real' dirent with the same name that's still
805            somewhere else on the flash. */
806         if (!jffs2_can_mark_obsolete(c)) {
807                 struct jffs2_raw_dirent *rd;
808                 struct jffs2_raw_node_ref *raw;
809                 int ret;
810                 size_t retlen;
811                 int name_len = strlen(fd->name);
812                 uint32_t name_crc = crc32(0, fd->name, name_len);
813                 uint32_t rawlen = ref_totlen(c, jeb, fd->raw);
814
815                 rd = kmalloc(rawlen, GFP_KERNEL);
816                 if (!rd)
817                         return -ENOMEM;
818
819                 /* Prevent the erase code from nicking the obsolete node refs while
820                    we're looking at them. I really don't like this extra lock but
821                    can't see any alternative. Suggestions on a postcard to... */
822                 down(&c->erase_free_sem);
823
824                 for (raw = f->inocache->nodes; raw != (void *)f->inocache; raw = raw->next_in_ino) {
825
826                         /* We only care about obsolete ones */
827                         if (!(ref_obsolete(raw)))
828                                 continue;
829
830                         /* Any dirent with the same name is going to have the same length... */
831                         if (ref_totlen(c, NULL, raw) != rawlen)
832                                 continue;
833
834                         /* Doesn't matter if there's one in the same erase block. We're going to 
835                            delete it too at the same time. */
836                         if (SECTOR_ADDR(raw->flash_offset) == SECTOR_ADDR(fd->raw->flash_offset))
837                                 continue;
838
839                         D1(printk(KERN_DEBUG "Check potential deletion dirent at %08x\n", ref_offset(raw)));
840
841                         /* This is an obsolete node belonging to the same directory, and it's of the right
842                            length. We need to take a closer look...*/
843                         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)rd);
844                         if (ret) {
845                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Read error (%d) reading obsolete node at %08x\n", ret, ref_offset(raw));
846                                 /* If we can't read it, we don't need to continue to obsolete it. Continue */
847                                 continue;
848                         }
849                         if (retlen != rawlen) {
850                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Short read (%zd not %u) reading header from obsolete node at %08x\n",
851                                        retlen, rawlen, ref_offset(raw));
852                                 continue;
853                         }
854
855                         if (je16_to_cpu(rd->nodetype) != JFFS2_NODETYPE_DIRENT)
856                                 continue;
857
858                         /* If the name CRC doesn't match, skip */
859                         if (je32_to_cpu(rd->name_crc) != name_crc)
860                                 continue;
861
862                         /* If the name length doesn't match, or it's another deletion dirent, skip */
863                         if (rd->nsize != name_len || !je32_to_cpu(rd->ino))
864                                 continue;
865
866                         /* OK, check the actual name now */
867                         if (memcmp(rd->name, fd->name, name_len))
868                                 continue;
869
870                         /* OK. The name really does match. There really is still an older node on
871                            the flash which our deletion dirent obsoletes. So we have to write out
872                            a new deletion dirent to replace it */
873                         up(&c->erase_free_sem);
874
875                         D1(printk(KERN_DEBUG "Deletion dirent at %08x still obsoletes real dirent \"%s\" at %08x for ino #%u\n",
876                                   ref_offset(fd->raw), fd->name, ref_offset(raw), je32_to_cpu(rd->ino)));
877                         kfree(rd);
878
879                         return jffs2_garbage_collect_dirent(c, jeb, f, fd);
880                 }
881
882                 up(&c->erase_free_sem);
883                 kfree(rd);
884         }
885
886         /* No need for it any more. Just mark it obsolete and remove it from the list */
887         while (*fdp) {
888                 if ((*fdp) == fd) {
889                         found = 1;
890                         *fdp = fd->next;
891                         break;
892                 }
893                 fdp = &(*fdp)->next;
894         }
895         if (!found) {
896                 printk(KERN_WARNING "Deletion dirent \"%s\" not found in list for ino #%u\n", fd->name, f->inocache->ino);
897         }
898         jffs2_mark_node_obsolete(c, fd->raw);
899         jffs2_free_full_dirent(fd);
900         return 0;
901 }
902
903 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
904                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
905                                       uint32_t start, uint32_t end)
906 {
907         struct jffs2_raw_inode ri;
908         struct jffs2_node_frag *frag;
909         struct jffs2_full_dnode *new_fn;
910         uint32_t alloclen, phys_ofs, ilen;
911         int ret;
912
913         D1(printk(KERN_DEBUG "Writing replacement hole node for ino #%u from offset 0x%x to 0x%x\n",
914                   f->inocache->ino, start, end));
915         
916         memset(&ri, 0, sizeof(ri));
917
918         if(fn->frags > 1) {
919                 size_t readlen;
920                 uint32_t crc;
921                 /* It's partially obsoleted by a later write. So we have to 
922                    write it out again with the _same_ version as before */
923                 ret = jffs2_flash_read(c, ref_offset(fn->raw), sizeof(ri), &readlen, (char *)&ri);
924                 if (readlen != sizeof(ri) || ret) {
925                         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);
926                         goto fill;
927                 }
928                 if (je16_to_cpu(ri.nodetype) != JFFS2_NODETYPE_INODE) {
929                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had node type 0x%04x instead of JFFS2_NODETYPE_INODE(0x%04x)\n",
930                                ref_offset(fn->raw),
931                                je16_to_cpu(ri.nodetype), JFFS2_NODETYPE_INODE);
932                         return -EIO;
933                 }
934                 if (je32_to_cpu(ri.totlen) != sizeof(ri)) {
935                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had totlen 0x%x instead of expected 0x%zx\n",
936                                ref_offset(fn->raw),
937                                je32_to_cpu(ri.totlen), sizeof(ri));
938                         return -EIO;
939                 }
940                 crc = crc32(0, &ri, sizeof(ri)-8);
941                 if (crc != je32_to_cpu(ri.node_crc)) {
942                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had CRC 0x%08x which doesn't match calculated CRC 0x%08x\n",
943                                ref_offset(fn->raw), 
944                                je32_to_cpu(ri.node_crc), crc);
945                         /* FIXME: We could possibly deal with this by writing new holes for each frag */
946                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n", 
947                                start, end, f->inocache->ino);
948                         goto fill;
949                 }
950                 if (ri.compr != JFFS2_COMPR_ZERO) {
951                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node 0x%08x wasn't a hole node!\n", ref_offset(fn->raw));
952                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n", 
953                                start, end, f->inocache->ino);
954                         goto fill;
955                 }
956         } else {
957         fill:
958                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
959                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
960                 ri.totlen = cpu_to_je32(sizeof(ri));
961                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
962
963                 ri.ino = cpu_to_je32(f->inocache->ino);
964                 ri.version = cpu_to_je32(++f->highest_version);
965                 ri.offset = cpu_to_je32(start);
966                 ri.dsize = cpu_to_je32(end - start);
967                 ri.csize = cpu_to_je32(0);
968                 ri.compr = JFFS2_COMPR_ZERO;
969         }
970         
971         frag = frag_last(&f->fragtree);
972         if (frag)
973                 /* Fetch the inode length from the fragtree rather then
974                  * from i_size since i_size may have not been updated yet */
975                 ilen = frag->ofs + frag->size;
976         else
977                 ilen = JFFS2_F_I_SIZE(f);
978
979         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
980         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
981         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
982         ri.isize = cpu_to_je32(ilen);
983         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
984         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
985         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
986         ri.data_crc = cpu_to_je32(0);
987         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
988
989         ret = jffs2_reserve_space_gc(c, sizeof(ri), &phys_ofs, &alloclen);
990         if (ret) {
991                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_hole failed: %d\n",
992                        sizeof(ri), ret);
993                 return ret;
994         }
995         new_fn = jffs2_write_dnode(c, f, &ri, NULL, 0, phys_ofs, ALLOC_GC);
996
997         if (IS_ERR(new_fn)) {
998                 printk(KERN_WARNING "Error writing new hole node: %ld\n", PTR_ERR(new_fn));
999                 return PTR_ERR(new_fn);
1000         }
1001         if (je32_to_cpu(ri.version) == f->highest_version) {
1002                 jffs2_add_full_dnode_to_inode(c, f, new_fn);
1003                 if (f->metadata) {
1004                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1005                         jffs2_free_full_dnode(f->metadata);
1006                         f->metadata = NULL;
1007                 }
1008                 return 0;
1009         }
1010
1011         /* 
1012          * We should only get here in the case where the node we are
1013          * replacing had more than one frag, so we kept the same version
1014          * number as before. (Except in case of error -- see 'goto fill;' 
1015          * above.)
1016          */
1017         D1(if(unlikely(fn->frags <= 1)) {
1018                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Replacing fn with %d frag(s) but new ver %d != highest_version %d of ino #%d\n",
1019                        fn->frags, je32_to_cpu(ri.version), f->highest_version,
1020                        je32_to_cpu(ri.ino));
1021         });
1022
1023         /* This is a partially-overlapped hole node. Mark it REF_NORMAL not REF_PRISTINE */
1024         mark_ref_normal(new_fn->raw);
1025
1026         for (frag = jffs2_lookup_node_frag(&f->fragtree, fn->ofs); 
1027              frag; frag = frag_next(frag)) {
1028                 if (frag->ofs > fn->size + fn->ofs)
1029                         break;
1030                 if (frag->node == fn) {
1031                         frag->node = new_fn;
1032                         new_fn->frags++;
1033                         fn->frags--;
1034                 }
1035         }
1036         if (fn->frags) {
1037                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Old node still has frags!\n");
1038                 BUG();
1039         }
1040         if (!new_fn->frags) {
1041                 printk(KERN_WARNING "jffs2_garbage_collect_hole: New node has no frags!\n");
1042                 BUG();
1043         }
1044                 
1045         jffs2_mark_node_obsolete(c, fn->raw);
1046         jffs2_free_full_dnode(fn);
1047         
1048         return 0;
1049 }
1050
1051 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
1052                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
1053                                        uint32_t start, uint32_t end)
1054 {
1055         struct jffs2_full_dnode *new_fn;
1056         struct jffs2_raw_inode ri;
1057         uint32_t alloclen, phys_ofs, offset, orig_end, orig_start;      
1058         int ret = 0;
1059         unsigned char *comprbuf = NULL, *writebuf;
1060         unsigned long pg;
1061         unsigned char *pg_ptr;
1062  
1063         memset(&ri, 0, sizeof(ri));
1064
1065         D1(printk(KERN_DEBUG "Writing replacement dnode for ino #%u from offset 0x%x to 0x%x\n",
1066                   f->inocache->ino, start, end));
1067
1068         orig_end = end;
1069         orig_start = start;
1070
1071         if (c->nr_free_blocks + c->nr_erasing_blocks > c->resv_blocks_gcmerge) {
1072                 /* Attempt to do some merging. But only expand to cover logically
1073                    adjacent frags if the block containing them is already considered
1074                    to be dirty. Otherwise we end up with GC just going round in 
1075                    circles dirtying the nodes it already wrote out, especially 
1076                    on NAND where we have small eraseblocks and hence a much higher
1077                    chance of nodes having to be split to cross boundaries. */
1078
1079                 struct jffs2_node_frag *frag;
1080                 uint32_t min, max;
1081
1082                 min = start & ~(PAGE_CACHE_SIZE-1);
1083                 max = min + PAGE_CACHE_SIZE;
1084
1085                 frag = jffs2_lookup_node_frag(&f->fragtree, start);
1086
1087                 /* BUG_ON(!frag) but that'll happen anyway... */
1088
1089                 BUG_ON(frag->ofs != start);
1090
1091                 /* First grow down... */
1092                 while((frag = frag_prev(frag)) && frag->ofs >= min) {
1093
1094                         /* If the previous frag doesn't even reach the beginning, there's
1095                            excessive fragmentation. Just merge. */
1096                         if (frag->ofs > min) {
1097                                 D1(printk(KERN_DEBUG "Expanding down to cover partial frag (0x%x-0x%x)\n",
1098                                           frag->ofs, frag->ofs+frag->size));
1099                                 start = frag->ofs;
1100                                 continue;
1101                         }
1102                         /* OK. This frag holds the first byte of the page. */
1103                         if (!frag->node || !frag->node->raw) {
1104                                 D1(printk(KERN_DEBUG "First frag in page is hole (0x%x-0x%x). Not expanding down.\n",
1105                                           frag->ofs, frag->ofs+frag->size));
1106                                 break;
1107                         } else {
1108
1109                                 /* OK, it's a frag which extends to the beginning of the page. Does it live 
1110                                    in a block which is still considered clean? If so, don't obsolete it.
1111                                    If not, cover it anyway. */
1112
1113                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1114                                 struct jffs2_eraseblock *jeb;
1115
1116                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1117
1118                                 if (jeb == c->gcblock) {
1119                                         D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1120                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1121                                         start = frag->ofs;
1122                                         break;
1123                                 }
1124                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1125                                         D1(printk(KERN_DEBUG "Not expanding down to cover frag (0x%x-0x%x) in clean block %08x\n",
1126                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1127                                         break;
1128                                 }
1129
1130                                 D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in dirty block %08x\n",
1131                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1132                                 start = frag->ofs;
1133                                 break;
1134                         }
1135                 }
1136
1137                 /* ... then up */
1138
1139                 /* Find last frag which is actually part of the node we're to GC. */
1140                 frag = jffs2_lookup_node_frag(&f->fragtree, end-1);
1141
1142                 while((frag = frag_next(frag)) && frag->ofs+frag->size <= max) {
1143
1144                         /* If the previous frag doesn't even reach the beginning, there's lots
1145                            of fragmentation. Just merge. */
1146                         if (frag->ofs+frag->size < max) {
1147                                 D1(printk(KERN_DEBUG "Expanding up to cover partial frag (0x%x-0x%x)\n",
1148                                           frag->ofs, frag->ofs+frag->size));
1149                                 end = frag->ofs + frag->size;
1150                                 continue;
1151                         }
1152
1153                         if (!frag->node || !frag->node->raw) {
1154                                 D1(printk(KERN_DEBUG "Last frag in page is hole (0x%x-0x%x). Not expanding up.\n",
1155                                           frag->ofs, frag->ofs+frag->size));
1156                                 break;
1157                         } else {
1158
1159                                 /* OK, it's a frag which extends to the beginning of the page. Does it live 
1160                                    in a block which is still considered clean? If so, don't obsolete it.
1161                                    If not, cover it anyway. */
1162
1163                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1164                                 struct jffs2_eraseblock *jeb;
1165
1166                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1167
1168                                 if (jeb == c->gcblock) {
1169                                         D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1170                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1171                                         end = frag->ofs + frag->size;
1172                                         break;
1173                                 }
1174                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1175                                         D1(printk(KERN_DEBUG "Not expanding up to cover frag (0x%x-0x%x) in clean block %08x\n",
1176                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1177                                         break;
1178                                 }
1179
1180                                 D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in dirty block %08x\n",
1181                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1182                                 end = frag->ofs + frag->size;
1183                                 break;
1184                         }
1185                 }
1186                 D1(printk(KERN_DEBUG "Expanded dnode to write from (0x%x-0x%x) to (0x%x-0x%x)\n", 
1187                           orig_start, orig_end, start, end));
1188
1189                 D1(BUG_ON(end > frag_last(&f->fragtree)->ofs + frag_last(&f->fragtree)->size));
1190                 BUG_ON(end < orig_end);
1191                 BUG_ON(start > orig_start);
1192         }
1193         
1194         /* First, use readpage() to read the appropriate page into the page cache */
1195         /* Q: What happens if we actually try to GC the _same_ page for which commit_write()
1196          *    triggered garbage collection in the first place?
1197          * A: I _think_ it's OK. read_cache_page shouldn't deadlock, we'll write out the
1198          *    page OK. We'll actually write it out again in commit_write, which is a little
1199          *    suboptimal, but at least we're correct.
1200          */
1201         pg_ptr = jffs2_gc_fetch_page(c, f, start, &pg);
1202
1203         if (IS_ERR(pg_ptr)) {
1204                 printk(KERN_WARNING "read_cache_page() returned error: %ld\n", PTR_ERR(pg_ptr));
1205                 return PTR_ERR(pg_ptr);
1206         }
1207
1208         offset = start;
1209         while(offset < orig_end) {
1210                 uint32_t datalen;
1211                 uint32_t cdatalen;
1212                 uint16_t comprtype = JFFS2_COMPR_NONE;
1213
1214                 ret = jffs2_reserve_space_gc(c, sizeof(ri) + JFFS2_MIN_DATA_LEN, &phys_ofs, &alloclen);
1215
1216                 if (ret) {
1217                         printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dnode failed: %d\n",
1218                                sizeof(ri)+ JFFS2_MIN_DATA_LEN, ret);
1219                         break;
1220                 }
1221                 cdatalen = min_t(uint32_t, alloclen - sizeof(ri), end - offset);
1222                 datalen = end - offset;
1223
1224                 writebuf = pg_ptr + (offset & (PAGE_CACHE_SIZE -1));
1225
1226                 comprtype = jffs2_compress(c, f, writebuf, &comprbuf, &datalen, &cdatalen);
1227
1228                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
1229                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
1230                 ri.totlen = cpu_to_je32(sizeof(ri) + cdatalen);
1231                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
1232
1233                 ri.ino = cpu_to_je32(f->inocache->ino);
1234                 ri.version = cpu_to_je32(++f->highest_version);
1235                 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1236                 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1237                 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1238                 ri.isize = cpu_to_je32(JFFS2_F_I_SIZE(f));
1239                 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1240                 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1241                 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1242                 ri.offset = cpu_to_je32(offset);
1243                 ri.csize = cpu_to_je32(cdatalen);
1244                 ri.dsize = cpu_to_je32(datalen);
1245                 ri.compr = comprtype & 0xff;
1246                 ri.usercompr = (comprtype >> 8) & 0xff;
1247                 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1248                 ri.data_crc = cpu_to_je32(crc32(0, comprbuf, cdatalen));
1249         
1250                 new_fn = jffs2_write_dnode(c, f, &ri, comprbuf, cdatalen, phys_ofs, ALLOC_GC);
1251
1252                 jffs2_free_comprbuf(comprbuf, writebuf);
1253
1254                 if (IS_ERR(new_fn)) {
1255                         printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
1256                         ret = PTR_ERR(new_fn);
1257                         break;
1258                 }
1259                 ret = jffs2_add_full_dnode_to_inode(c, f, new_fn);
1260                 offset += datalen;
1261                 if (f->metadata) {
1262                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1263                         jffs2_free_full_dnode(f->metadata);
1264                         f->metadata = NULL;
1265                 }
1266         }
1267
1268         jffs2_gc_release_page(c, pg_ptr, &pg);
1269         return ret;
1270 }
1271