Merge x86-64 update from Andi
[linux-2.6] / fs / jffs2 / gc.c
1 /*
2  * JFFS2 -- Journalling Flash File System, Version 2.
3  *
4  * Copyright (C) 2001-2003 Red Hat, Inc.
5  *
6  * Created by David Woodhouse <dwmw2@infradead.org>
7  *
8  * For licensing information, see the file 'LICENCE' in this directory.
9  *
10  * $Id: gc.c,v 1.155 2005/11/07 11:14:39 gleixner Exp $
11  *
12  */
13
14 #include <linux/kernel.h>
15 #include <linux/mtd/mtd.h>
16 #include <linux/slab.h>
17 #include <linux/pagemap.h>
18 #include <linux/crc32.h>
19 #include <linux/compiler.h>
20 #include <linux/stat.h>
21 #include "nodelist.h"
22 #include "compr.h"
23
24 static int jffs2_garbage_collect_pristine(struct jffs2_sb_info *c,
25                                           struct jffs2_inode_cache *ic,
26                                           struct jffs2_raw_node_ref *raw);
27 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
28                                         struct jffs2_inode_info *f, struct jffs2_full_dnode *fd);
29 static int jffs2_garbage_collect_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
30                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd);
31 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
32                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd);
33 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
34                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
35                                       uint32_t start, uint32_t end);
36 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
37                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
38                                        uint32_t start, uint32_t end);
39 static int jffs2_garbage_collect_live(struct jffs2_sb_info *c,  struct jffs2_eraseblock *jeb,
40                                struct jffs2_raw_node_ref *raw, struct jffs2_inode_info *f);
41
42 /* Called with erase_completion_lock held */
43 static struct jffs2_eraseblock *jffs2_find_gc_block(struct jffs2_sb_info *c)
44 {
45         struct jffs2_eraseblock *ret;
46         struct list_head *nextlist = NULL;
47         int n = jiffies % 128;
48
49         /* Pick an eraseblock to garbage collect next. This is where we'll
50            put the clever wear-levelling algorithms. Eventually.  */
51         /* We possibly want to favour the dirtier blocks more when the
52            number of free blocks is low. */
53 again:
54         if (!list_empty(&c->bad_used_list) && c->nr_free_blocks > c->resv_blocks_gcbad) {
55                 D1(printk(KERN_DEBUG "Picking block from bad_used_list to GC next\n"));
56                 nextlist = &c->bad_used_list;
57         } else if (n < 50 && !list_empty(&c->erasable_list)) {
58                 /* Note that most of them will have gone directly to be erased.
59                    So don't favour the erasable_list _too_ much. */
60                 D1(printk(KERN_DEBUG "Picking block from erasable_list to GC next\n"));
61                 nextlist = &c->erasable_list;
62         } else if (n < 110 && !list_empty(&c->very_dirty_list)) {
63                 /* Most of the time, pick one off the very_dirty list */
64                 D1(printk(KERN_DEBUG "Picking block from very_dirty_list to GC next\n"));
65                 nextlist = &c->very_dirty_list;
66         } else if (n < 126 && !list_empty(&c->dirty_list)) {
67                 D1(printk(KERN_DEBUG "Picking block from dirty_list to GC next\n"));
68                 nextlist = &c->dirty_list;
69         } else if (!list_empty(&c->clean_list)) {
70                 D1(printk(KERN_DEBUG "Picking block from clean_list to GC next\n"));
71                 nextlist = &c->clean_list;
72         } else if (!list_empty(&c->dirty_list)) {
73                 D1(printk(KERN_DEBUG "Picking block from dirty_list to GC next (clean_list was empty)\n"));
74
75                 nextlist = &c->dirty_list;
76         } else if (!list_empty(&c->very_dirty_list)) {
77                 D1(printk(KERN_DEBUG "Picking block from very_dirty_list to GC next (clean_list and dirty_list were empty)\n"));
78                 nextlist = &c->very_dirty_list;
79         } else if (!list_empty(&c->erasable_list)) {
80                 D1(printk(KERN_DEBUG "Picking block from erasable_list to GC next (clean_list and {very_,}dirty_list were empty)\n"));
81
82                 nextlist = &c->erasable_list;
83         } else if (!list_empty(&c->erasable_pending_wbuf_list)) {
84                 /* There are blocks are wating for the wbuf sync */
85                 D1(printk(KERN_DEBUG "Synching wbuf in order to reuse erasable_pending_wbuf_list blocks\n"));
86                 spin_unlock(&c->erase_completion_lock);
87                 jffs2_flush_wbuf_pad(c);
88                 spin_lock(&c->erase_completion_lock);
89                 goto again;
90         } else {
91                 /* Eep. All were empty */
92                 D1(printk(KERN_NOTICE "jffs2: No clean, dirty _or_ erasable blocks to GC from! Where are they all?\n"));
93                 return NULL;
94         }
95
96         ret = list_entry(nextlist->next, struct jffs2_eraseblock, list);
97         list_del(&ret->list);
98         c->gcblock = ret;
99         ret->gc_node = ret->first_node;
100         if (!ret->gc_node) {
101                 printk(KERN_WARNING "Eep. ret->gc_node for block at 0x%08x is NULL\n", ret->offset);
102                 BUG();
103         }
104
105         /* Have we accidentally picked a clean block with wasted space ? */
106         if (ret->wasted_size) {
107                 D1(printk(KERN_DEBUG "Converting wasted_size %08x to dirty_size\n", ret->wasted_size));
108                 ret->dirty_size += ret->wasted_size;
109                 c->wasted_size -= ret->wasted_size;
110                 c->dirty_size += ret->wasted_size;
111                 ret->wasted_size = 0;
112         }
113
114         return ret;
115 }
116
117 /* jffs2_garbage_collect_pass
118  * Make a single attempt to progress GC. Move one node, and possibly
119  * start erasing one eraseblock.
120  */
121 int jffs2_garbage_collect_pass(struct jffs2_sb_info *c)
122 {
123         struct jffs2_inode_info *f;
124         struct jffs2_inode_cache *ic;
125         struct jffs2_eraseblock *jeb;
126         struct jffs2_raw_node_ref *raw;
127         int ret = 0, inum, nlink;
128
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) +
517                                 JFFS2_MIN_DATA_LEN, rawlen), &phys_ofs, &alloclen, rawlen);
518                                 /* this is not the exact summary size of it,
519                                         it is only an upper estimation */
520
521         if (ret)
522                 return ret;
523
524         if (alloclen < rawlen) {
525                 /* Doesn't fit untouched. We'll go the old route and split it */
526                 return -EBADFD;
527         }
528
529         node = kmalloc(rawlen, GFP_KERNEL);
530         if (!node)
531                return -ENOMEM;
532
533         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)node);
534         if (!ret && retlen != rawlen)
535                 ret = -EIO;
536         if (ret)
537                 goto out_node;
538
539         crc = crc32(0, node, sizeof(struct jffs2_unknown_node)-4);
540         if (je32_to_cpu(node->u.hdr_crc) != crc) {
541                 printk(KERN_WARNING "Header CRC failed on REF_PRISTINE node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
542                        ref_offset(raw), je32_to_cpu(node->u.hdr_crc), crc);
543                 goto bail;
544         }
545
546         switch(je16_to_cpu(node->u.nodetype)) {
547         case JFFS2_NODETYPE_INODE:
548                 crc = crc32(0, node, sizeof(node->i)-8);
549                 if (je32_to_cpu(node->i.node_crc) != crc) {
550                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
551                                ref_offset(raw), je32_to_cpu(node->i.node_crc), crc);
552                         goto bail;
553                 }
554
555                 if (je32_to_cpu(node->i.dsize)) {
556                         crc = crc32(0, node->i.data, je32_to_cpu(node->i.csize));
557                         if (je32_to_cpu(node->i.data_crc) != crc) {
558                                 printk(KERN_WARNING "Data CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
559                                        ref_offset(raw), je32_to_cpu(node->i.data_crc), crc);
560                                 goto bail;
561                         }
562                 }
563                 break;
564
565         case JFFS2_NODETYPE_DIRENT:
566                 crc = crc32(0, node, sizeof(node->d)-8);
567                 if (je32_to_cpu(node->d.node_crc) != crc) {
568                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE dirent node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
569                                ref_offset(raw), je32_to_cpu(node->d.node_crc), crc);
570                         goto bail;
571                 }
572
573                 if (node->d.nsize) {
574                         crc = crc32(0, node->d.name, node->d.nsize);
575                         if (je32_to_cpu(node->d.name_crc) != crc) {
576                                 printk(KERN_WARNING "Name CRC failed on REF_PRISTINE dirent ode at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
577                                        ref_offset(raw), je32_to_cpu(node->d.name_crc), crc);
578                                 goto bail;
579                         }
580                 }
581                 break;
582         default:
583                 printk(KERN_WARNING "Unknown node type for REF_PRISTINE node at 0x%08x: 0x%04x\n",
584                        ref_offset(raw), je16_to_cpu(node->u.nodetype));
585                 goto bail;
586         }
587
588         nraw = jffs2_alloc_raw_node_ref();
589         if (!nraw) {
590                 ret = -ENOMEM;
591                 goto out_node;
592         }
593
594         /* OK, all the CRCs are good; this node can just be copied as-is. */
595  retry:
596         nraw->flash_offset = phys_ofs;
597         nraw->__totlen = rawlen;
598         nraw->next_phys = NULL;
599
600         ret = jffs2_flash_write(c, phys_ofs, rawlen, &retlen, (char *)node);
601
602         if (ret || (retlen != rawlen)) {
603                 printk(KERN_NOTICE "Write of %d bytes at 0x%08x failed. returned %d, retlen %zd\n",
604                        rawlen, phys_ofs, ret, retlen);
605                 if (retlen) {
606                         /* Doesn't belong to any inode */
607                         nraw->next_in_ino = NULL;
608
609                         nraw->flash_offset |= REF_OBSOLETE;
610                         jffs2_add_physical_node_ref(c, nraw);
611                         jffs2_mark_node_obsolete(c, nraw);
612                 } else {
613                         printk(KERN_NOTICE "Not marking the space at 0x%08x as dirty because the flash driver returned retlen zero\n", nraw->flash_offset);
614                         jffs2_free_raw_node_ref(nraw);
615                 }
616                 if (!retried && (nraw = jffs2_alloc_raw_node_ref())) {
617                         /* Try to reallocate space and retry */
618                         uint32_t dummy;
619                         struct jffs2_eraseblock *jeb = &c->blocks[phys_ofs / c->sector_size];
620
621                         retried = 1;
622
623                         D1(printk(KERN_DEBUG "Retrying failed write of REF_PRISTINE node.\n"));
624
625                         jffs2_dbg_acct_sanity_check(c,jeb);
626                         jffs2_dbg_acct_paranoia_check(c, jeb);
627
628                         ret = jffs2_reserve_space_gc(c, rawlen, &phys_ofs, &dummy, rawlen);
629                                                 /* this is not the exact summary size of it,
630                                                         it is only an upper estimation */
631
632                         if (!ret) {
633                                 D1(printk(KERN_DEBUG "Allocated space at 0x%08x to retry failed write.\n", phys_ofs));
634
635                                 jffs2_dbg_acct_sanity_check(c,jeb);
636                                 jffs2_dbg_acct_paranoia_check(c, jeb);
637
638                                 goto retry;
639                         }
640                         D1(printk(KERN_DEBUG "Failed to allocate space to retry failed write: %d!\n", ret));
641                         jffs2_free_raw_node_ref(nraw);
642                 }
643
644                 jffs2_free_raw_node_ref(nraw);
645                 if (!ret)
646                         ret = -EIO;
647                 goto out_node;
648         }
649         nraw->flash_offset |= REF_PRISTINE;
650         jffs2_add_physical_node_ref(c, nraw);
651
652         /* Link into per-inode list. This is safe because of the ic
653            state being INO_STATE_GC. Note that if we're doing this
654            for an inode which is in-core, the 'nraw' pointer is then
655            going to be fetched from ic->nodes by our caller. */
656         spin_lock(&c->erase_completion_lock);
657         nraw->next_in_ino = ic->nodes;
658         ic->nodes = nraw;
659         spin_unlock(&c->erase_completion_lock);
660
661         jffs2_mark_node_obsolete(c, raw);
662         D1(printk(KERN_DEBUG "WHEEE! GC REF_PRISTINE node at 0x%08x succeeded\n", ref_offset(raw)));
663
664  out_node:
665         kfree(node);
666         return ret;
667  bail:
668         ret = -EBADFD;
669         goto out_node;
670 }
671
672 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
673                                         struct jffs2_inode_info *f, struct jffs2_full_dnode *fn)
674 {
675         struct jffs2_full_dnode *new_fn;
676         struct jffs2_raw_inode ri;
677         struct jffs2_node_frag *last_frag;
678         jint16_t dev;
679         char *mdata = NULL, mdatalen = 0;
680         uint32_t alloclen, phys_ofs, ilen;
681         int ret;
682
683         if (S_ISBLK(JFFS2_F_I_MODE(f)) ||
684             S_ISCHR(JFFS2_F_I_MODE(f)) ) {
685                 /* For these, we don't actually need to read the old node */
686                 /* FIXME: for minor or major > 255. */
687                 dev = cpu_to_je16(((JFFS2_F_I_RDEV_MAJ(f) << 8) |
688                         JFFS2_F_I_RDEV_MIN(f)));
689                 mdata = (char *)&dev;
690                 mdatalen = sizeof(dev);
691                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bytes of kdev_t\n", mdatalen));
692         } else if (S_ISLNK(JFFS2_F_I_MODE(f))) {
693                 mdatalen = fn->size;
694                 mdata = kmalloc(fn->size, GFP_KERNEL);
695                 if (!mdata) {
696                         printk(KERN_WARNING "kmalloc of mdata failed in jffs2_garbage_collect_metadata()\n");
697                         return -ENOMEM;
698                 }
699                 ret = jffs2_read_dnode(c, f, fn, mdata, 0, mdatalen);
700                 if (ret) {
701                         printk(KERN_WARNING "read of old metadata failed in jffs2_garbage_collect_metadata(): %d\n", ret);
702                         kfree(mdata);
703                         return ret;
704                 }
705                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bites of symlink target\n", mdatalen));
706
707         }
708
709         ret = jffs2_reserve_space_gc(c, sizeof(ri) + mdatalen, &phys_ofs, &alloclen,
710                                 JFFS2_SUMMARY_INODE_SIZE);
711         if (ret) {
712                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_metadata failed: %d\n",
713                        sizeof(ri)+ mdatalen, ret);
714                 goto out;
715         }
716
717         last_frag = frag_last(&f->fragtree);
718         if (last_frag)
719                 /* Fetch the inode length from the fragtree rather then
720                  * from i_size since i_size may have not been updated yet */
721                 ilen = last_frag->ofs + last_frag->size;
722         else
723                 ilen = JFFS2_F_I_SIZE(f);
724
725         memset(&ri, 0, sizeof(ri));
726         ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
727         ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
728         ri.totlen = cpu_to_je32(sizeof(ri) + mdatalen);
729         ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
730
731         ri.ino = cpu_to_je32(f->inocache->ino);
732         ri.version = cpu_to_je32(++f->highest_version);
733         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
734         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
735         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
736         ri.isize = cpu_to_je32(ilen);
737         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
738         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
739         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
740         ri.offset = cpu_to_je32(0);
741         ri.csize = cpu_to_je32(mdatalen);
742         ri.dsize = cpu_to_je32(mdatalen);
743         ri.compr = JFFS2_COMPR_NONE;
744         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
745         ri.data_crc = cpu_to_je32(crc32(0, mdata, mdatalen));
746
747         new_fn = jffs2_write_dnode(c, f, &ri, mdata, mdatalen, phys_ofs, ALLOC_GC);
748
749         if (IS_ERR(new_fn)) {
750                 printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
751                 ret = PTR_ERR(new_fn);
752                 goto out;
753         }
754         jffs2_mark_node_obsolete(c, fn->raw);
755         jffs2_free_full_dnode(fn);
756         f->metadata = new_fn;
757  out:
758         if (S_ISLNK(JFFS2_F_I_MODE(f)))
759                 kfree(mdata);
760         return ret;
761 }
762
763 static int jffs2_garbage_collect_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
764                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
765 {
766         struct jffs2_full_dirent *new_fd;
767         struct jffs2_raw_dirent rd;
768         uint32_t alloclen, phys_ofs;
769         int ret;
770
771         rd.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
772         rd.nodetype = cpu_to_je16(JFFS2_NODETYPE_DIRENT);
773         rd.nsize = strlen(fd->name);
774         rd.totlen = cpu_to_je32(sizeof(rd) + rd.nsize);
775         rd.hdr_crc = cpu_to_je32(crc32(0, &rd, sizeof(struct jffs2_unknown_node)-4));
776
777         rd.pino = cpu_to_je32(f->inocache->ino);
778         rd.version = cpu_to_je32(++f->highest_version);
779         rd.ino = cpu_to_je32(fd->ino);
780         /* If the times on this inode were set by explicit utime() they can be different,
781            so refrain from splatting them. */
782         if (JFFS2_F_I_MTIME(f) == JFFS2_F_I_CTIME(f))
783                 rd.mctime = cpu_to_je32(JFFS2_F_I_MTIME(f));
784         else
785                 rd.mctime = cpu_to_je32(0);
786         rd.type = fd->type;
787         rd.node_crc = cpu_to_je32(crc32(0, &rd, sizeof(rd)-8));
788         rd.name_crc = cpu_to_je32(crc32(0, fd->name, rd.nsize));
789
790         ret = jffs2_reserve_space_gc(c, sizeof(rd)+rd.nsize, &phys_ofs, &alloclen,
791                                 JFFS2_SUMMARY_DIRENT_SIZE(rd.nsize));
792         if (ret) {
793                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dirent failed: %d\n",
794                        sizeof(rd)+rd.nsize, ret);
795                 return ret;
796         }
797         new_fd = jffs2_write_dirent(c, f, &rd, fd->name, rd.nsize, phys_ofs, ALLOC_GC);
798
799         if (IS_ERR(new_fd)) {
800                 printk(KERN_WARNING "jffs2_write_dirent in garbage_collect_dirent failed: %ld\n", PTR_ERR(new_fd));
801                 return PTR_ERR(new_fd);
802         }
803         jffs2_add_fd_to_list(c, new_fd, &f->dents);
804         return 0;
805 }
806
807 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
808                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
809 {
810         struct jffs2_full_dirent **fdp = &f->dents;
811         int found = 0;
812
813         /* On a medium where we can't actually mark nodes obsolete
814            pernamently, such as NAND flash, we need to work out
815            whether this deletion dirent is still needed to actively
816            delete a 'real' dirent with the same name that's still
817            somewhere else on the flash. */
818         if (!jffs2_can_mark_obsolete(c)) {
819                 struct jffs2_raw_dirent *rd;
820                 struct jffs2_raw_node_ref *raw;
821                 int ret;
822                 size_t retlen;
823                 int name_len = strlen(fd->name);
824                 uint32_t name_crc = crc32(0, fd->name, name_len);
825                 uint32_t rawlen = ref_totlen(c, jeb, fd->raw);
826
827                 rd = kmalloc(rawlen, GFP_KERNEL);
828                 if (!rd)
829                         return -ENOMEM;
830
831                 /* Prevent the erase code from nicking the obsolete node refs while
832                    we're looking at them. I really don't like this extra lock but
833                    can't see any alternative. Suggestions on a postcard to... */
834                 down(&c->erase_free_sem);
835
836                 for (raw = f->inocache->nodes; raw != (void *)f->inocache; raw = raw->next_in_ino) {
837
838                         /* We only care about obsolete ones */
839                         if (!(ref_obsolete(raw)))
840                                 continue;
841
842                         /* Any dirent with the same name is going to have the same length... */
843                         if (ref_totlen(c, NULL, raw) != rawlen)
844                                 continue;
845
846                         /* Doesn't matter if there's one in the same erase block. We're going to
847                            delete it too at the same time. */
848                         if (SECTOR_ADDR(raw->flash_offset) == SECTOR_ADDR(fd->raw->flash_offset))
849                                 continue;
850
851                         D1(printk(KERN_DEBUG "Check potential deletion dirent at %08x\n", ref_offset(raw)));
852
853                         /* This is an obsolete node belonging to the same directory, and it's of the right
854                            length. We need to take a closer look...*/
855                         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)rd);
856                         if (ret) {
857                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Read error (%d) reading obsolete node at %08x\n", ret, ref_offset(raw));
858                                 /* If we can't read it, we don't need to continue to obsolete it. Continue */
859                                 continue;
860                         }
861                         if (retlen != rawlen) {
862                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Short read (%zd not %u) reading header from obsolete node at %08x\n",
863                                        retlen, rawlen, ref_offset(raw));
864                                 continue;
865                         }
866
867                         if (je16_to_cpu(rd->nodetype) != JFFS2_NODETYPE_DIRENT)
868                                 continue;
869
870                         /* If the name CRC doesn't match, skip */
871                         if (je32_to_cpu(rd->name_crc) != name_crc)
872                                 continue;
873
874                         /* If the name length doesn't match, or it's another deletion dirent, skip */
875                         if (rd->nsize != name_len || !je32_to_cpu(rd->ino))
876                                 continue;
877
878                         /* OK, check the actual name now */
879                         if (memcmp(rd->name, fd->name, name_len))
880                                 continue;
881
882                         /* OK. The name really does match. There really is still an older node on
883                            the flash which our deletion dirent obsoletes. So we have to write out
884                            a new deletion dirent to replace it */
885                         up(&c->erase_free_sem);
886
887                         D1(printk(KERN_DEBUG "Deletion dirent at %08x still obsoletes real dirent \"%s\" at %08x for ino #%u\n",
888                                   ref_offset(fd->raw), fd->name, ref_offset(raw), je32_to_cpu(rd->ino)));
889                         kfree(rd);
890
891                         return jffs2_garbage_collect_dirent(c, jeb, f, fd);
892                 }
893
894                 up(&c->erase_free_sem);
895                 kfree(rd);
896         }
897
898         /* FIXME: If we're deleting a dirent which contains the current mtime and ctime,
899            we should update the metadata node with those times accordingly */
900
901         /* No need for it any more. Just mark it obsolete and remove it from the list */
902         while (*fdp) {
903                 if ((*fdp) == fd) {
904                         found = 1;
905                         *fdp = fd->next;
906                         break;
907                 }
908                 fdp = &(*fdp)->next;
909         }
910         if (!found) {
911                 printk(KERN_WARNING "Deletion dirent \"%s\" not found in list for ino #%u\n", fd->name, f->inocache->ino);
912         }
913         jffs2_mark_node_obsolete(c, fd->raw);
914         jffs2_free_full_dirent(fd);
915         return 0;
916 }
917
918 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
919                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
920                                       uint32_t start, uint32_t end)
921 {
922         struct jffs2_raw_inode ri;
923         struct jffs2_node_frag *frag;
924         struct jffs2_full_dnode *new_fn;
925         uint32_t alloclen, phys_ofs, ilen;
926         int ret;
927
928         D1(printk(KERN_DEBUG "Writing replacement hole node for ino #%u from offset 0x%x to 0x%x\n",
929                   f->inocache->ino, start, end));
930
931         memset(&ri, 0, sizeof(ri));
932
933         if(fn->frags > 1) {
934                 size_t readlen;
935                 uint32_t crc;
936                 /* It's partially obsoleted by a later write. So we have to
937                    write it out again with the _same_ version as before */
938                 ret = jffs2_flash_read(c, ref_offset(fn->raw), sizeof(ri), &readlen, (char *)&ri);
939                 if (readlen != sizeof(ri) || ret) {
940                         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);
941                         goto fill;
942                 }
943                 if (je16_to_cpu(ri.nodetype) != JFFS2_NODETYPE_INODE) {
944                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had node type 0x%04x instead of JFFS2_NODETYPE_INODE(0x%04x)\n",
945                                ref_offset(fn->raw),
946                                je16_to_cpu(ri.nodetype), JFFS2_NODETYPE_INODE);
947                         return -EIO;
948                 }
949                 if (je32_to_cpu(ri.totlen) != sizeof(ri)) {
950                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had totlen 0x%x instead of expected 0x%zx\n",
951                                ref_offset(fn->raw),
952                                je32_to_cpu(ri.totlen), sizeof(ri));
953                         return -EIO;
954                 }
955                 crc = crc32(0, &ri, sizeof(ri)-8);
956                 if (crc != je32_to_cpu(ri.node_crc)) {
957                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had CRC 0x%08x which doesn't match calculated CRC 0x%08x\n",
958                                ref_offset(fn->raw),
959                                je32_to_cpu(ri.node_crc), crc);
960                         /* FIXME: We could possibly deal with this by writing new holes for each frag */
961                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
962                                start, end, f->inocache->ino);
963                         goto fill;
964                 }
965                 if (ri.compr != JFFS2_COMPR_ZERO) {
966                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node 0x%08x wasn't a hole node!\n", ref_offset(fn->raw));
967                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
968                                start, end, f->inocache->ino);
969                         goto fill;
970                 }
971         } else {
972         fill:
973                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
974                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
975                 ri.totlen = cpu_to_je32(sizeof(ri));
976                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
977
978                 ri.ino = cpu_to_je32(f->inocache->ino);
979                 ri.version = cpu_to_je32(++f->highest_version);
980                 ri.offset = cpu_to_je32(start);
981                 ri.dsize = cpu_to_je32(end - start);
982                 ri.csize = cpu_to_je32(0);
983                 ri.compr = JFFS2_COMPR_ZERO;
984         }
985
986         frag = frag_last(&f->fragtree);
987         if (frag)
988                 /* Fetch the inode length from the fragtree rather then
989                  * from i_size since i_size may have not been updated yet */
990                 ilen = frag->ofs + frag->size;
991         else
992                 ilen = JFFS2_F_I_SIZE(f);
993
994         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
995         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
996         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
997         ri.isize = cpu_to_je32(ilen);
998         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
999         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1000         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1001         ri.data_crc = cpu_to_je32(0);
1002         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1003
1004         ret = jffs2_reserve_space_gc(c, sizeof(ri), &phys_ofs, &alloclen,
1005                                 JFFS2_SUMMARY_INODE_SIZE);
1006         if (ret) {
1007                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_hole failed: %d\n",
1008                        sizeof(ri), ret);
1009                 return ret;
1010         }
1011         new_fn = jffs2_write_dnode(c, f, &ri, NULL, 0, phys_ofs, ALLOC_GC);
1012
1013         if (IS_ERR(new_fn)) {
1014                 printk(KERN_WARNING "Error writing new hole node: %ld\n", PTR_ERR(new_fn));
1015                 return PTR_ERR(new_fn);
1016         }
1017         if (je32_to_cpu(ri.version) == f->highest_version) {
1018                 jffs2_add_full_dnode_to_inode(c, f, new_fn);
1019                 if (f->metadata) {
1020                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1021                         jffs2_free_full_dnode(f->metadata);
1022                         f->metadata = NULL;
1023                 }
1024                 return 0;
1025         }
1026
1027         /*
1028          * We should only get here in the case where the node we are
1029          * replacing had more than one frag, so we kept the same version
1030          * number as before. (Except in case of error -- see 'goto fill;'
1031          * above.)
1032          */
1033         D1(if(unlikely(fn->frags <= 1)) {
1034                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Replacing fn with %d frag(s) but new ver %d != highest_version %d of ino #%d\n",
1035                        fn->frags, je32_to_cpu(ri.version), f->highest_version,
1036                        je32_to_cpu(ri.ino));
1037         });
1038
1039         /* This is a partially-overlapped hole node. Mark it REF_NORMAL not REF_PRISTINE */
1040         mark_ref_normal(new_fn->raw);
1041
1042         for (frag = jffs2_lookup_node_frag(&f->fragtree, fn->ofs);
1043              frag; frag = frag_next(frag)) {
1044                 if (frag->ofs > fn->size + fn->ofs)
1045                         break;
1046                 if (frag->node == fn) {
1047                         frag->node = new_fn;
1048                         new_fn->frags++;
1049                         fn->frags--;
1050                 }
1051         }
1052         if (fn->frags) {
1053                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Old node still has frags!\n");
1054                 BUG();
1055         }
1056         if (!new_fn->frags) {
1057                 printk(KERN_WARNING "jffs2_garbage_collect_hole: New node has no frags!\n");
1058                 BUG();
1059         }
1060
1061         jffs2_mark_node_obsolete(c, fn->raw);
1062         jffs2_free_full_dnode(fn);
1063
1064         return 0;
1065 }
1066
1067 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
1068                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
1069                                        uint32_t start, uint32_t end)
1070 {
1071         struct jffs2_full_dnode *new_fn;
1072         struct jffs2_raw_inode ri;
1073         uint32_t alloclen, phys_ofs, offset, orig_end, orig_start;
1074         int ret = 0;
1075         unsigned char *comprbuf = NULL, *writebuf;
1076         unsigned long pg;
1077         unsigned char *pg_ptr;
1078
1079         memset(&ri, 0, sizeof(ri));
1080
1081         D1(printk(KERN_DEBUG "Writing replacement dnode for ino #%u from offset 0x%x to 0x%x\n",
1082                   f->inocache->ino, start, end));
1083
1084         orig_end = end;
1085         orig_start = start;
1086
1087         if (c->nr_free_blocks + c->nr_erasing_blocks > c->resv_blocks_gcmerge) {
1088                 /* Attempt to do some merging. But only expand to cover logically
1089                    adjacent frags if the block containing them is already considered
1090                    to be dirty. Otherwise we end up with GC just going round in
1091                    circles dirtying the nodes it already wrote out, especially
1092                    on NAND where we have small eraseblocks and hence a much higher
1093                    chance of nodes having to be split to cross boundaries. */
1094
1095                 struct jffs2_node_frag *frag;
1096                 uint32_t min, max;
1097
1098                 min = start & ~(PAGE_CACHE_SIZE-1);
1099                 max = min + PAGE_CACHE_SIZE;
1100
1101                 frag = jffs2_lookup_node_frag(&f->fragtree, start);
1102
1103                 /* BUG_ON(!frag) but that'll happen anyway... */
1104
1105                 BUG_ON(frag->ofs != start);
1106
1107                 /* First grow down... */
1108                 while((frag = frag_prev(frag)) && frag->ofs >= min) {
1109
1110                         /* If the previous frag doesn't even reach the beginning, there's
1111                            excessive fragmentation. Just merge. */
1112                         if (frag->ofs > min) {
1113                                 D1(printk(KERN_DEBUG "Expanding down to cover partial frag (0x%x-0x%x)\n",
1114                                           frag->ofs, frag->ofs+frag->size));
1115                                 start = frag->ofs;
1116                                 continue;
1117                         }
1118                         /* OK. This frag holds the first byte of the page. */
1119                         if (!frag->node || !frag->node->raw) {
1120                                 D1(printk(KERN_DEBUG "First frag in page is hole (0x%x-0x%x). Not expanding down.\n",
1121                                           frag->ofs, frag->ofs+frag->size));
1122                                 break;
1123                         } else {
1124
1125                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1126                                    in a block which is still considered clean? If so, don't obsolete it.
1127                                    If not, cover it anyway. */
1128
1129                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1130                                 struct jffs2_eraseblock *jeb;
1131
1132                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1133
1134                                 if (jeb == c->gcblock) {
1135                                         D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1136                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1137                                         start = frag->ofs;
1138                                         break;
1139                                 }
1140                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1141                                         D1(printk(KERN_DEBUG "Not expanding down to cover frag (0x%x-0x%x) in clean block %08x\n",
1142                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1143                                         break;
1144                                 }
1145
1146                                 D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in dirty block %08x\n",
1147                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1148                                 start = frag->ofs;
1149                                 break;
1150                         }
1151                 }
1152
1153                 /* ... then up */
1154
1155                 /* Find last frag which is actually part of the node we're to GC. */
1156                 frag = jffs2_lookup_node_frag(&f->fragtree, end-1);
1157
1158                 while((frag = frag_next(frag)) && frag->ofs+frag->size <= max) {
1159
1160                         /* If the previous frag doesn't even reach the beginning, there's lots
1161                            of fragmentation. Just merge. */
1162                         if (frag->ofs+frag->size < max) {
1163                                 D1(printk(KERN_DEBUG "Expanding up to cover partial frag (0x%x-0x%x)\n",
1164                                           frag->ofs, frag->ofs+frag->size));
1165                                 end = frag->ofs + frag->size;
1166                                 continue;
1167                         }
1168
1169                         if (!frag->node || !frag->node->raw) {
1170                                 D1(printk(KERN_DEBUG "Last frag in page is hole (0x%x-0x%x). Not expanding up.\n",
1171                                           frag->ofs, frag->ofs+frag->size));
1172                                 break;
1173                         } else {
1174
1175                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1176                                    in a block which is still considered clean? If so, don't obsolete it.
1177                                    If not, cover it anyway. */
1178
1179                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1180                                 struct jffs2_eraseblock *jeb;
1181
1182                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1183
1184                                 if (jeb == c->gcblock) {
1185                                         D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1186                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1187                                         end = frag->ofs + frag->size;
1188                                         break;
1189                                 }
1190                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1191                                         D1(printk(KERN_DEBUG "Not expanding up to cover frag (0x%x-0x%x) in clean block %08x\n",
1192                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1193                                         break;
1194                                 }
1195
1196                                 D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in dirty block %08x\n",
1197                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1198                                 end = frag->ofs + frag->size;
1199                                 break;
1200                         }
1201                 }
1202                 D1(printk(KERN_DEBUG "Expanded dnode to write from (0x%x-0x%x) to (0x%x-0x%x)\n",
1203                           orig_start, orig_end, start, end));
1204
1205                 D1(BUG_ON(end > frag_last(&f->fragtree)->ofs + frag_last(&f->fragtree)->size));
1206                 BUG_ON(end < orig_end);
1207                 BUG_ON(start > orig_start);
1208         }
1209
1210         /* First, use readpage() to read the appropriate page into the page cache */
1211         /* Q: What happens if we actually try to GC the _same_ page for which commit_write()
1212          *    triggered garbage collection in the first place?
1213          * A: I _think_ it's OK. read_cache_page shouldn't deadlock, we'll write out the
1214          *    page OK. We'll actually write it out again in commit_write, which is a little
1215          *    suboptimal, but at least we're correct.
1216          */
1217         pg_ptr = jffs2_gc_fetch_page(c, f, start, &pg);
1218
1219         if (IS_ERR(pg_ptr)) {
1220                 printk(KERN_WARNING "read_cache_page() returned error: %ld\n", PTR_ERR(pg_ptr));
1221                 return PTR_ERR(pg_ptr);
1222         }
1223
1224         offset = start;
1225         while(offset < orig_end) {
1226                 uint32_t datalen;
1227                 uint32_t cdatalen;
1228                 uint16_t comprtype = JFFS2_COMPR_NONE;
1229
1230                 ret = jffs2_reserve_space_gc(c, sizeof(ri) + JFFS2_MIN_DATA_LEN, &phys_ofs,
1231                                         &alloclen, JFFS2_SUMMARY_INODE_SIZE);
1232
1233                 if (ret) {
1234                         printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dnode failed: %d\n",
1235                                sizeof(ri)+ JFFS2_MIN_DATA_LEN, ret);
1236                         break;
1237                 }
1238                 cdatalen = min_t(uint32_t, alloclen - sizeof(ri), end - offset);
1239                 datalen = end - offset;
1240
1241                 writebuf = pg_ptr + (offset & (PAGE_CACHE_SIZE -1));
1242
1243                 comprtype = jffs2_compress(c, f, writebuf, &comprbuf, &datalen, &cdatalen);
1244
1245                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
1246                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
1247                 ri.totlen = cpu_to_je32(sizeof(ri) + cdatalen);
1248                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
1249
1250                 ri.ino = cpu_to_je32(f->inocache->ino);
1251                 ri.version = cpu_to_je32(++f->highest_version);
1252                 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1253                 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1254                 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1255                 ri.isize = cpu_to_je32(JFFS2_F_I_SIZE(f));
1256                 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1257                 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1258                 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1259                 ri.offset = cpu_to_je32(offset);
1260                 ri.csize = cpu_to_je32(cdatalen);
1261                 ri.dsize = cpu_to_je32(datalen);
1262                 ri.compr = comprtype & 0xff;
1263                 ri.usercompr = (comprtype >> 8) & 0xff;
1264                 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1265                 ri.data_crc = cpu_to_je32(crc32(0, comprbuf, cdatalen));
1266
1267                 new_fn = jffs2_write_dnode(c, f, &ri, comprbuf, cdatalen, phys_ofs, ALLOC_GC);
1268
1269                 jffs2_free_comprbuf(comprbuf, writebuf);
1270
1271                 if (IS_ERR(new_fn)) {
1272                         printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
1273                         ret = PTR_ERR(new_fn);
1274                         break;
1275                 }
1276                 ret = jffs2_add_full_dnode_to_inode(c, f, new_fn);
1277                 offset += datalen;
1278                 if (f->metadata) {
1279                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1280                         jffs2_free_full_dnode(f->metadata);
1281                         f->metadata = NULL;
1282                 }
1283         }
1284
1285         jffs2_gc_release_page(c, pg_ptr, &pg);
1286         return ret;
1287 }