[JFFS2][XATTR] XATTR support on JFFS2 (version. 5)
[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         int xattr = 0;
129
130         if (down_interruptible(&c->alloc_sem))
131                 return -EINTR;
132
133         for (;;) {
134                 spin_lock(&c->erase_completion_lock);
135                 if (!c->unchecked_size)
136                         break;
137
138                 /* We can't start doing GC yet. We haven't finished checking
139                    the node CRCs etc. Do it now. */
140
141                 /* checked_ino is protected by the alloc_sem */
142                 if (c->checked_ino > c->highest_ino && xattr) {
143                         printk(KERN_CRIT "Checked all inodes but still 0x%x bytes of unchecked space?\n",
144                                c->unchecked_size);
145                         jffs2_dbg_dump_block_lists_nolock(c);
146                         spin_unlock(&c->erase_completion_lock);
147                         BUG();
148                 }
149
150                 spin_unlock(&c->erase_completion_lock);
151
152                 if (!xattr)
153                         xattr = jffs2_verify_xattr(c);
154
155                 spin_lock(&c->inocache_lock);
156
157                 ic = jffs2_get_ino_cache(c, c->checked_ino++);
158
159                 if (!ic) {
160                         spin_unlock(&c->inocache_lock);
161                         continue;
162                 }
163
164                 if (!ic->nlink) {
165                         D1(printk(KERN_DEBUG "Skipping check of ino #%d with nlink zero\n",
166                                   ic->ino));
167                         spin_unlock(&c->inocache_lock);
168                         continue;
169                 }
170                 switch(ic->state) {
171                 case INO_STATE_CHECKEDABSENT:
172                 case INO_STATE_PRESENT:
173                         D1(printk(KERN_DEBUG "Skipping ino #%u already checked\n", ic->ino));
174                         spin_unlock(&c->inocache_lock);
175                         continue;
176
177                 case INO_STATE_GC:
178                 case INO_STATE_CHECKING:
179                         printk(KERN_WARNING "Inode #%u is in state %d during CRC check phase!\n", ic->ino, ic->state);
180                         spin_unlock(&c->inocache_lock);
181                         BUG();
182
183                 case INO_STATE_READING:
184                         /* We need to wait for it to finish, lest we move on
185                            and trigger the BUG() above while we haven't yet
186                            finished checking all its nodes */
187                         D1(printk(KERN_DEBUG "Waiting for ino #%u to finish reading\n", ic->ino));
188                         /* We need to come back again for the _same_ inode. We've
189                          made no progress in this case, but that should be OK */
190                         c->checked_ino--;
191
192                         up(&c->alloc_sem);
193                         sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
194                         return 0;
195
196                 default:
197                         BUG();
198
199                 case INO_STATE_UNCHECKED:
200                         ;
201                 }
202                 ic->state = INO_STATE_CHECKING;
203                 spin_unlock(&c->inocache_lock);
204
205                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() triggering inode scan of ino#%u\n", ic->ino));
206
207                 ret = jffs2_do_crccheck_inode(c, ic);
208                 if (ret)
209                         printk(KERN_WARNING "Returned error for crccheck of ino #%u. Expect badness...\n", ic->ino);
210
211                 jffs2_set_inocache_state(c, ic, INO_STATE_CHECKEDABSENT);
212                 up(&c->alloc_sem);
213                 return ret;
214         }
215
216         /* First, work out which block we're garbage-collecting */
217         jeb = c->gcblock;
218
219         if (!jeb)
220                 jeb = jffs2_find_gc_block(c);
221
222         if (!jeb) {
223                 D1 (printk(KERN_NOTICE "jffs2: Couldn't find erase block to garbage collect!\n"));
224                 spin_unlock(&c->erase_completion_lock);
225                 up(&c->alloc_sem);
226                 return -EIO;
227         }
228
229         D1(printk(KERN_DEBUG "GC from block %08x, used_size %08x, dirty_size %08x, free_size %08x\n", jeb->offset, jeb->used_size, jeb->dirty_size, jeb->free_size));
230         D1(if (c->nextblock)
231            printk(KERN_DEBUG "Nextblock at  %08x, used_size %08x, dirty_size %08x, wasted_size %08x, free_size %08x\n", c->nextblock->offset, c->nextblock->used_size, c->nextblock->dirty_size, c->nextblock->wasted_size, c->nextblock->free_size));
232
233         if (!jeb->used_size) {
234                 up(&c->alloc_sem);
235                 goto eraseit;
236         }
237
238         raw = jeb->gc_node;
239
240         while(ref_obsolete(raw)) {
241                 D1(printk(KERN_DEBUG "Node at 0x%08x is obsolete... skipping\n", ref_offset(raw)));
242                 raw = raw->next_phys;
243                 if (unlikely(!raw)) {
244                         printk(KERN_WARNING "eep. End of raw list while still supposedly nodes to GC\n");
245                         printk(KERN_WARNING "erase block at 0x%08x. free_size 0x%08x, dirty_size 0x%08x, used_size 0x%08x\n",
246                                jeb->offset, jeb->free_size, jeb->dirty_size, jeb->used_size);
247                         jeb->gc_node = raw;
248                         spin_unlock(&c->erase_completion_lock);
249                         up(&c->alloc_sem);
250                         BUG();
251                 }
252         }
253         jeb->gc_node = raw;
254
255         D1(printk(KERN_DEBUG "Going to garbage collect node at 0x%08x\n", ref_offset(raw)));
256
257         if (!raw->next_in_ino) {
258                 /* Inode-less node. Clean marker, snapshot or something like that */
259                 /* FIXME: If it's something that needs to be copied, including something
260                    we don't grok that has JFFS2_NODETYPE_RWCOMPAT_COPY, we should do so */
261                 spin_unlock(&c->erase_completion_lock);
262                 jffs2_mark_node_obsolete(c, raw);
263                 up(&c->alloc_sem);
264                 goto eraseit_lock;
265         }
266
267         ic = jffs2_raw_ref_to_ic(raw);
268
269         /* When 'ic' refers xattr_datum/xattr_ref, this node is GCed as xattr.
270            We can decide whether this node is inode or xattr by ic->class.
271            ret = 0 : ic is xattr_datum/xattr_ref, and GC was SUCCESSED.
272            ret < 0 : ic is xattr_datum/xattr_ref, but GC was FAILED.
273            ret > 0 : ic is NOT xattr_datum/xattr_ref.
274         */
275         ret = jffs2_garbage_collect_xattr(c, ic);
276         if (ret <= 0)
277                 goto release_sem;
278
279         /* We need to hold the inocache. Either the erase_completion_lock or
280            the inocache_lock are sufficient; we trade down since the inocache_lock
281            causes less contention. */
282         spin_lock(&c->inocache_lock);
283
284         spin_unlock(&c->erase_completion_lock);
285
286         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));
287
288         /* Three possibilities:
289            1. Inode is already in-core. We must iget it and do proper
290               updating to its fragtree, etc.
291            2. Inode is not in-core, node is REF_PRISTINE. We lock the
292               inocache to prevent a read_inode(), copy the node intact.
293            3. Inode is not in-core, node is not pristine. We must iget()
294               and take the slow path.
295         */
296
297         switch(ic->state) {
298         case INO_STATE_CHECKEDABSENT:
299                 /* It's been checked, but it's not currently in-core.
300                    We can just copy any pristine nodes, but have
301                    to prevent anyone else from doing read_inode() while
302                    we're at it, so we set the state accordingly */
303                 if (ref_flags(raw) == REF_PRISTINE)
304                         ic->state = INO_STATE_GC;
305                 else {
306                         D1(printk(KERN_DEBUG "Ino #%u is absent but node not REF_PRISTINE. Reading.\n",
307                                   ic->ino));
308                 }
309                 break;
310
311         case INO_STATE_PRESENT:
312                 /* It's in-core. GC must iget() it. */
313                 break;
314
315         case INO_STATE_UNCHECKED:
316         case INO_STATE_CHECKING:
317         case INO_STATE_GC:
318                 /* Should never happen. We should have finished checking
319                    by the time we actually start doing any GC, and since
320                    we're holding the alloc_sem, no other garbage collection
321                    can happen.
322                 */
323                 printk(KERN_CRIT "Inode #%u already in state %d in jffs2_garbage_collect_pass()!\n",
324                        ic->ino, ic->state);
325                 up(&c->alloc_sem);
326                 spin_unlock(&c->inocache_lock);
327                 BUG();
328
329         case INO_STATE_READING:
330                 /* Someone's currently trying to read it. We must wait for
331                    them to finish and then go through the full iget() route
332                    to do the GC. However, sometimes read_inode() needs to get
333                    the alloc_sem() (for marking nodes invalid) so we must
334                    drop the alloc_sem before sleeping. */
335
336                 up(&c->alloc_sem);
337                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() waiting for ino #%u in state %d\n",
338                           ic->ino, ic->state));
339                 sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock);
340                 /* And because we dropped the alloc_sem we must start again from the
341                    beginning. Ponder chance of livelock here -- we're returning success
342                    without actually making any progress.
343
344                    Q: What are the chances that the inode is back in INO_STATE_READING
345                    again by the time we next enter this function? And that this happens
346                    enough times to cause a real delay?
347
348                    A: Small enough that I don't care :)
349                 */
350                 return 0;
351         }
352
353         /* OK. Now if the inode is in state INO_STATE_GC, we are going to copy the
354            node intact, and we don't have to muck about with the fragtree etc.
355            because we know it's not in-core. If it _was_ in-core, we go through
356            all the iget() crap anyway */
357
358         if (ic->state == INO_STATE_GC) {
359                 spin_unlock(&c->inocache_lock);
360
361                 ret = jffs2_garbage_collect_pristine(c, ic, raw);
362
363                 spin_lock(&c->inocache_lock);
364                 ic->state = INO_STATE_CHECKEDABSENT;
365                 wake_up(&c->inocache_wq);
366
367                 if (ret != -EBADFD) {
368                         spin_unlock(&c->inocache_lock);
369                         goto release_sem;
370                 }
371
372                 /* Fall through if it wanted us to, with inocache_lock held */
373         }
374
375         /* Prevent the fairly unlikely race where the gcblock is
376            entirely obsoleted by the final close of a file which had
377            the only valid nodes in the block, followed by erasure,
378            followed by freeing of the ic because the erased block(s)
379            held _all_ the nodes of that inode.... never been seen but
380            it's vaguely possible. */
381
382         inum = ic->ino;
383         nlink = ic->nlink;
384         spin_unlock(&c->inocache_lock);
385
386         f = jffs2_gc_fetch_inode(c, inum, nlink);
387         if (IS_ERR(f)) {
388                 ret = PTR_ERR(f);
389                 goto release_sem;
390         }
391         if (!f) {
392                 ret = 0;
393                 goto release_sem;
394         }
395
396         ret = jffs2_garbage_collect_live(c, jeb, raw, f);
397
398         jffs2_gc_release_inode(c, f);
399
400  release_sem:
401         up(&c->alloc_sem);
402
403  eraseit_lock:
404         /* If we've finished this block, start it erasing */
405         spin_lock(&c->erase_completion_lock);
406
407  eraseit:
408         if (c->gcblock && !c->gcblock->used_size) {
409                 D1(printk(KERN_DEBUG "Block at 0x%08x completely obsoleted by GC. Moving to erase_pending_list\n", c->gcblock->offset));
410                 /* We're GC'ing an empty block? */
411                 list_add_tail(&c->gcblock->list, &c->erase_pending_list);
412                 c->gcblock = NULL;
413                 c->nr_erasing_blocks++;
414                 jffs2_erase_pending_trigger(c);
415         }
416         spin_unlock(&c->erase_completion_lock);
417
418         return ret;
419 }
420
421 static int jffs2_garbage_collect_live(struct jffs2_sb_info *c,  struct jffs2_eraseblock *jeb,
422                                       struct jffs2_raw_node_ref *raw, struct jffs2_inode_info *f)
423 {
424         struct jffs2_node_frag *frag;
425         struct jffs2_full_dnode *fn = NULL;
426         struct jffs2_full_dirent *fd;
427         uint32_t start = 0, end = 0, nrfrags = 0;
428         int ret = 0;
429
430         down(&f->sem);
431
432         /* Now we have the lock for this inode. Check that it's still the one at the head
433            of the list. */
434
435         spin_lock(&c->erase_completion_lock);
436
437         if (c->gcblock != jeb) {
438                 spin_unlock(&c->erase_completion_lock);
439                 D1(printk(KERN_DEBUG "GC block is no longer gcblock. Restart\n"));
440                 goto upnout;
441         }
442         if (ref_obsolete(raw)) {
443                 spin_unlock(&c->erase_completion_lock);
444                 D1(printk(KERN_DEBUG "node to be GC'd was obsoleted in the meantime.\n"));
445                 /* They'll call again */
446                 goto upnout;
447         }
448         spin_unlock(&c->erase_completion_lock);
449
450         /* OK. Looks safe. And nobody can get us now because we have the semaphore. Move the block */
451         if (f->metadata && f->metadata->raw == raw) {
452                 fn = f->metadata;
453                 ret = jffs2_garbage_collect_metadata(c, jeb, f, fn);
454                 goto upnout;
455         }
456
457         /* FIXME. Read node and do lookup? */
458         for (frag = frag_first(&f->fragtree); frag; frag = frag_next(frag)) {
459                 if (frag->node && frag->node->raw == raw) {
460                         fn = frag->node;
461                         end = frag->ofs + frag->size;
462                         if (!nrfrags++)
463                                 start = frag->ofs;
464                         if (nrfrags == frag->node->frags)
465                                 break; /* We've found them all */
466                 }
467         }
468         if (fn) {
469                 if (ref_flags(raw) == REF_PRISTINE) {
470                         ret = jffs2_garbage_collect_pristine(c, f->inocache, raw);
471                         if (!ret) {
472                                 /* Urgh. Return it sensibly. */
473                                 frag->node->raw = f->inocache->nodes;
474                         }
475                         if (ret != -EBADFD)
476                                 goto upnout;
477                 }
478                 /* We found a datanode. Do the GC */
479                 if((start >> PAGE_CACHE_SHIFT) < ((end-1) >> PAGE_CACHE_SHIFT)) {
480                         /* It crosses a page boundary. Therefore, it must be a hole. */
481                         ret = jffs2_garbage_collect_hole(c, jeb, f, fn, start, end);
482                 } else {
483                         /* It could still be a hole. But we GC the page this way anyway */
484                         ret = jffs2_garbage_collect_dnode(c, jeb, f, fn, start, end);
485                 }
486                 goto upnout;
487         }
488
489         /* Wasn't a dnode. Try dirent */
490         for (fd = f->dents; fd; fd=fd->next) {
491                 if (fd->raw == raw)
492                         break;
493         }
494
495         if (fd && fd->ino) {
496                 ret = jffs2_garbage_collect_dirent(c, jeb, f, fd);
497         } else if (fd) {
498                 ret = jffs2_garbage_collect_deletion_dirent(c, jeb, f, fd);
499         } else {
500                 printk(KERN_WARNING "Raw node at 0x%08x wasn't in node lists for ino #%u\n",
501                        ref_offset(raw), f->inocache->ino);
502                 if (ref_obsolete(raw)) {
503                         printk(KERN_WARNING "But it's obsolete so we don't mind too much\n");
504                 } else {
505                         jffs2_dbg_dump_node(c, ref_offset(raw));
506                         BUG();
507                 }
508         }
509  upnout:
510         up(&f->sem);
511
512         return ret;
513 }
514
515 static int jffs2_garbage_collect_pristine(struct jffs2_sb_info *c,
516                                           struct jffs2_inode_cache *ic,
517                                           struct jffs2_raw_node_ref *raw)
518 {
519         union jffs2_node_union *node;
520         struct jffs2_raw_node_ref *nraw;
521         size_t retlen;
522         int ret;
523         uint32_t phys_ofs, alloclen;
524         uint32_t crc, rawlen;
525         int retried = 0;
526
527         D1(printk(KERN_DEBUG "Going to GC REF_PRISTINE node at 0x%08x\n", ref_offset(raw)));
528
529         rawlen = ref_totlen(c, c->gcblock, raw);
530
531         /* Ask for a small amount of space (or the totlen if smaller) because we
532            don't want to force wastage of the end of a block if splitting would
533            work. */
534         ret = jffs2_reserve_space_gc(c, min_t(uint32_t, sizeof(struct jffs2_raw_inode) +
535                                 JFFS2_MIN_DATA_LEN, rawlen), &phys_ofs, &alloclen, rawlen);
536                                 /* this is not the exact summary size of it,
537                                         it is only an upper estimation */
538
539         if (ret)
540                 return ret;
541
542         if (alloclen < rawlen) {
543                 /* Doesn't fit untouched. We'll go the old route and split it */
544                 return -EBADFD;
545         }
546
547         node = kmalloc(rawlen, GFP_KERNEL);
548         if (!node)
549                return -ENOMEM;
550
551         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)node);
552         if (!ret && retlen != rawlen)
553                 ret = -EIO;
554         if (ret)
555                 goto out_node;
556
557         crc = crc32(0, node, sizeof(struct jffs2_unknown_node)-4);
558         if (je32_to_cpu(node->u.hdr_crc) != crc) {
559                 printk(KERN_WARNING "Header CRC failed on REF_PRISTINE node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
560                        ref_offset(raw), je32_to_cpu(node->u.hdr_crc), crc);
561                 goto bail;
562         }
563
564         switch(je16_to_cpu(node->u.nodetype)) {
565         case JFFS2_NODETYPE_INODE:
566                 crc = crc32(0, node, sizeof(node->i)-8);
567                 if (je32_to_cpu(node->i.node_crc) != crc) {
568                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
569                                ref_offset(raw), je32_to_cpu(node->i.node_crc), crc);
570                         goto bail;
571                 }
572
573                 if (je32_to_cpu(node->i.dsize)) {
574                         crc = crc32(0, node->i.data, je32_to_cpu(node->i.csize));
575                         if (je32_to_cpu(node->i.data_crc) != crc) {
576                                 printk(KERN_WARNING "Data CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
577                                        ref_offset(raw), je32_to_cpu(node->i.data_crc), crc);
578                                 goto bail;
579                         }
580                 }
581                 break;
582
583         case JFFS2_NODETYPE_DIRENT:
584                 crc = crc32(0, node, sizeof(node->d)-8);
585                 if (je32_to_cpu(node->d.node_crc) != crc) {
586                         printk(KERN_WARNING "Node CRC failed on REF_PRISTINE dirent node at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
587                                ref_offset(raw), je32_to_cpu(node->d.node_crc), crc);
588                         goto bail;
589                 }
590
591                 if (node->d.nsize) {
592                         crc = crc32(0, node->d.name, node->d.nsize);
593                         if (je32_to_cpu(node->d.name_crc) != crc) {
594                                 printk(KERN_WARNING "Name CRC failed on REF_PRISTINE dirent ode at 0x%08x: Read 0x%08x, calculated 0x%08x\n",
595                                        ref_offset(raw), je32_to_cpu(node->d.name_crc), crc);
596                                 goto bail;
597                         }
598                 }
599                 break;
600         default:
601                 printk(KERN_WARNING "Unknown node type for REF_PRISTINE node at 0x%08x: 0x%04x\n",
602                        ref_offset(raw), je16_to_cpu(node->u.nodetype));
603                 goto bail;
604         }
605
606         nraw = jffs2_alloc_raw_node_ref();
607         if (!nraw) {
608                 ret = -ENOMEM;
609                 goto out_node;
610         }
611
612         /* OK, all the CRCs are good; this node can just be copied as-is. */
613  retry:
614         nraw->flash_offset = phys_ofs;
615         nraw->__totlen = rawlen;
616         nraw->next_phys = NULL;
617
618         ret = jffs2_flash_write(c, phys_ofs, rawlen, &retlen, (char *)node);
619
620         if (ret || (retlen != rawlen)) {
621                 printk(KERN_NOTICE "Write of %d bytes at 0x%08x failed. returned %d, retlen %zd\n",
622                        rawlen, phys_ofs, ret, retlen);
623                 if (retlen) {
624                         /* Doesn't belong to any inode */
625                         nraw->next_in_ino = NULL;
626
627                         nraw->flash_offset |= REF_OBSOLETE;
628                         jffs2_add_physical_node_ref(c, nraw);
629                         jffs2_mark_node_obsolete(c, nraw);
630                 } else {
631                         printk(KERN_NOTICE "Not marking the space at 0x%08x as dirty because the flash driver returned retlen zero\n", nraw->flash_offset);
632                         jffs2_free_raw_node_ref(nraw);
633                 }
634                 if (!retried && (nraw = jffs2_alloc_raw_node_ref())) {
635                         /* Try to reallocate space and retry */
636                         uint32_t dummy;
637                         struct jffs2_eraseblock *jeb = &c->blocks[phys_ofs / c->sector_size];
638
639                         retried = 1;
640
641                         D1(printk(KERN_DEBUG "Retrying failed write of REF_PRISTINE node.\n"));
642
643                         jffs2_dbg_acct_sanity_check(c,jeb);
644                         jffs2_dbg_acct_paranoia_check(c, jeb);
645
646                         ret = jffs2_reserve_space_gc(c, rawlen, &phys_ofs, &dummy, rawlen);
647                                                 /* this is not the exact summary size of it,
648                                                         it is only an upper estimation */
649
650                         if (!ret) {
651                                 D1(printk(KERN_DEBUG "Allocated space at 0x%08x to retry failed write.\n", phys_ofs));
652
653                                 jffs2_dbg_acct_sanity_check(c,jeb);
654                                 jffs2_dbg_acct_paranoia_check(c, jeb);
655
656                                 goto retry;
657                         }
658                         D1(printk(KERN_DEBUG "Failed to allocate space to retry failed write: %d!\n", ret));
659                         jffs2_free_raw_node_ref(nraw);
660                 }
661
662                 jffs2_free_raw_node_ref(nraw);
663                 if (!ret)
664                         ret = -EIO;
665                 goto out_node;
666         }
667         nraw->flash_offset |= REF_PRISTINE;
668         jffs2_add_physical_node_ref(c, nraw);
669
670         /* Link into per-inode list. This is safe because of the ic
671            state being INO_STATE_GC. Note that if we're doing this
672            for an inode which is in-core, the 'nraw' pointer is then
673            going to be fetched from ic->nodes by our caller. */
674         spin_lock(&c->erase_completion_lock);
675         nraw->next_in_ino = ic->nodes;
676         ic->nodes = nraw;
677         spin_unlock(&c->erase_completion_lock);
678
679         jffs2_mark_node_obsolete(c, raw);
680         D1(printk(KERN_DEBUG "WHEEE! GC REF_PRISTINE node at 0x%08x succeeded\n", ref_offset(raw)));
681
682  out_node:
683         kfree(node);
684         return ret;
685  bail:
686         ret = -EBADFD;
687         goto out_node;
688 }
689
690 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
691                                         struct jffs2_inode_info *f, struct jffs2_full_dnode *fn)
692 {
693         struct jffs2_full_dnode *new_fn;
694         struct jffs2_raw_inode ri;
695         struct jffs2_node_frag *last_frag;
696         jint16_t dev;
697         char *mdata = NULL, mdatalen = 0;
698         uint32_t alloclen, phys_ofs, ilen;
699         int ret;
700
701         if (S_ISBLK(JFFS2_F_I_MODE(f)) ||
702             S_ISCHR(JFFS2_F_I_MODE(f)) ) {
703                 /* For these, we don't actually need to read the old node */
704                 /* FIXME: for minor or major > 255. */
705                 dev = cpu_to_je16(((JFFS2_F_I_RDEV_MAJ(f) << 8) |
706                         JFFS2_F_I_RDEV_MIN(f)));
707                 mdata = (char *)&dev;
708                 mdatalen = sizeof(dev);
709                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bytes of kdev_t\n", mdatalen));
710         } else if (S_ISLNK(JFFS2_F_I_MODE(f))) {
711                 mdatalen = fn->size;
712                 mdata = kmalloc(fn->size, GFP_KERNEL);
713                 if (!mdata) {
714                         printk(KERN_WARNING "kmalloc of mdata failed in jffs2_garbage_collect_metadata()\n");
715                         return -ENOMEM;
716                 }
717                 ret = jffs2_read_dnode(c, f, fn, mdata, 0, mdatalen);
718                 if (ret) {
719                         printk(KERN_WARNING "read of old metadata failed in jffs2_garbage_collect_metadata(): %d\n", ret);
720                         kfree(mdata);
721                         return ret;
722                 }
723                 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bites of symlink target\n", mdatalen));
724
725         }
726
727         ret = jffs2_reserve_space_gc(c, sizeof(ri) + mdatalen, &phys_ofs, &alloclen,
728                                 JFFS2_SUMMARY_INODE_SIZE);
729         if (ret) {
730                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_metadata failed: %d\n",
731                        sizeof(ri)+ mdatalen, ret);
732                 goto out;
733         }
734
735         last_frag = frag_last(&f->fragtree);
736         if (last_frag)
737                 /* Fetch the inode length from the fragtree rather then
738                  * from i_size since i_size may have not been updated yet */
739                 ilen = last_frag->ofs + last_frag->size;
740         else
741                 ilen = JFFS2_F_I_SIZE(f);
742
743         memset(&ri, 0, sizeof(ri));
744         ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
745         ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
746         ri.totlen = cpu_to_je32(sizeof(ri) + mdatalen);
747         ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
748
749         ri.ino = cpu_to_je32(f->inocache->ino);
750         ri.version = cpu_to_je32(++f->highest_version);
751         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
752         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
753         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
754         ri.isize = cpu_to_je32(ilen);
755         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
756         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
757         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
758         ri.offset = cpu_to_je32(0);
759         ri.csize = cpu_to_je32(mdatalen);
760         ri.dsize = cpu_to_je32(mdatalen);
761         ri.compr = JFFS2_COMPR_NONE;
762         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
763         ri.data_crc = cpu_to_je32(crc32(0, mdata, mdatalen));
764
765         new_fn = jffs2_write_dnode(c, f, &ri, mdata, mdatalen, phys_ofs, ALLOC_GC);
766
767         if (IS_ERR(new_fn)) {
768                 printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
769                 ret = PTR_ERR(new_fn);
770                 goto out;
771         }
772         jffs2_mark_node_obsolete(c, fn->raw);
773         jffs2_free_full_dnode(fn);
774         f->metadata = new_fn;
775  out:
776         if (S_ISLNK(JFFS2_F_I_MODE(f)))
777                 kfree(mdata);
778         return ret;
779 }
780
781 static int jffs2_garbage_collect_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
782                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
783 {
784         struct jffs2_full_dirent *new_fd;
785         struct jffs2_raw_dirent rd;
786         uint32_t alloclen, phys_ofs;
787         int ret;
788
789         rd.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
790         rd.nodetype = cpu_to_je16(JFFS2_NODETYPE_DIRENT);
791         rd.nsize = strlen(fd->name);
792         rd.totlen = cpu_to_je32(sizeof(rd) + rd.nsize);
793         rd.hdr_crc = cpu_to_je32(crc32(0, &rd, sizeof(struct jffs2_unknown_node)-4));
794
795         rd.pino = cpu_to_je32(f->inocache->ino);
796         rd.version = cpu_to_je32(++f->highest_version);
797         rd.ino = cpu_to_je32(fd->ino);
798         /* If the times on this inode were set by explicit utime() they can be different,
799            so refrain from splatting them. */
800         if (JFFS2_F_I_MTIME(f) == JFFS2_F_I_CTIME(f))
801                 rd.mctime = cpu_to_je32(JFFS2_F_I_MTIME(f));
802         else
803                 rd.mctime = cpu_to_je32(0);
804         rd.type = fd->type;
805         rd.node_crc = cpu_to_je32(crc32(0, &rd, sizeof(rd)-8));
806         rd.name_crc = cpu_to_je32(crc32(0, fd->name, rd.nsize));
807
808         ret = jffs2_reserve_space_gc(c, sizeof(rd)+rd.nsize, &phys_ofs, &alloclen,
809                                 JFFS2_SUMMARY_DIRENT_SIZE(rd.nsize));
810         if (ret) {
811                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dirent failed: %d\n",
812                        sizeof(rd)+rd.nsize, ret);
813                 return ret;
814         }
815         new_fd = jffs2_write_dirent(c, f, &rd, fd->name, rd.nsize, phys_ofs, ALLOC_GC);
816
817         if (IS_ERR(new_fd)) {
818                 printk(KERN_WARNING "jffs2_write_dirent in garbage_collect_dirent failed: %ld\n", PTR_ERR(new_fd));
819                 return PTR_ERR(new_fd);
820         }
821         jffs2_add_fd_to_list(c, new_fd, &f->dents);
822         return 0;
823 }
824
825 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
826                                         struct jffs2_inode_info *f, struct jffs2_full_dirent *fd)
827 {
828         struct jffs2_full_dirent **fdp = &f->dents;
829         int found = 0;
830
831         /* On a medium where we can't actually mark nodes obsolete
832            pernamently, such as NAND flash, we need to work out
833            whether this deletion dirent is still needed to actively
834            delete a 'real' dirent with the same name that's still
835            somewhere else on the flash. */
836         if (!jffs2_can_mark_obsolete(c)) {
837                 struct jffs2_raw_dirent *rd;
838                 struct jffs2_raw_node_ref *raw;
839                 int ret;
840                 size_t retlen;
841                 int name_len = strlen(fd->name);
842                 uint32_t name_crc = crc32(0, fd->name, name_len);
843                 uint32_t rawlen = ref_totlen(c, jeb, fd->raw);
844
845                 rd = kmalloc(rawlen, GFP_KERNEL);
846                 if (!rd)
847                         return -ENOMEM;
848
849                 /* Prevent the erase code from nicking the obsolete node refs while
850                    we're looking at them. I really don't like this extra lock but
851                    can't see any alternative. Suggestions on a postcard to... */
852                 down(&c->erase_free_sem);
853
854                 for (raw = f->inocache->nodes; raw != (void *)f->inocache; raw = raw->next_in_ino) {
855
856                         /* We only care about obsolete ones */
857                         if (!(ref_obsolete(raw)))
858                                 continue;
859
860                         /* Any dirent with the same name is going to have the same length... */
861                         if (ref_totlen(c, NULL, raw) != rawlen)
862                                 continue;
863
864                         /* Doesn't matter if there's one in the same erase block. We're going to
865                            delete it too at the same time. */
866                         if (SECTOR_ADDR(raw->flash_offset) == SECTOR_ADDR(fd->raw->flash_offset))
867                                 continue;
868
869                         D1(printk(KERN_DEBUG "Check potential deletion dirent at %08x\n", ref_offset(raw)));
870
871                         /* This is an obsolete node belonging to the same directory, and it's of the right
872                            length. We need to take a closer look...*/
873                         ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)rd);
874                         if (ret) {
875                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Read error (%d) reading obsolete node at %08x\n", ret, ref_offset(raw));
876                                 /* If we can't read it, we don't need to continue to obsolete it. Continue */
877                                 continue;
878                         }
879                         if (retlen != rawlen) {
880                                 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Short read (%zd not %u) reading header from obsolete node at %08x\n",
881                                        retlen, rawlen, ref_offset(raw));
882                                 continue;
883                         }
884
885                         if (je16_to_cpu(rd->nodetype) != JFFS2_NODETYPE_DIRENT)
886                                 continue;
887
888                         /* If the name CRC doesn't match, skip */
889                         if (je32_to_cpu(rd->name_crc) != name_crc)
890                                 continue;
891
892                         /* If the name length doesn't match, or it's another deletion dirent, skip */
893                         if (rd->nsize != name_len || !je32_to_cpu(rd->ino))
894                                 continue;
895
896                         /* OK, check the actual name now */
897                         if (memcmp(rd->name, fd->name, name_len))
898                                 continue;
899
900                         /* OK. The name really does match. There really is still an older node on
901                            the flash which our deletion dirent obsoletes. So we have to write out
902                            a new deletion dirent to replace it */
903                         up(&c->erase_free_sem);
904
905                         D1(printk(KERN_DEBUG "Deletion dirent at %08x still obsoletes real dirent \"%s\" at %08x for ino #%u\n",
906                                   ref_offset(fd->raw), fd->name, ref_offset(raw), je32_to_cpu(rd->ino)));
907                         kfree(rd);
908
909                         return jffs2_garbage_collect_dirent(c, jeb, f, fd);
910                 }
911
912                 up(&c->erase_free_sem);
913                 kfree(rd);
914         }
915
916         /* FIXME: If we're deleting a dirent which contains the current mtime and ctime,
917            we should update the metadata node with those times accordingly */
918
919         /* No need for it any more. Just mark it obsolete and remove it from the list */
920         while (*fdp) {
921                 if ((*fdp) == fd) {
922                         found = 1;
923                         *fdp = fd->next;
924                         break;
925                 }
926                 fdp = &(*fdp)->next;
927         }
928         if (!found) {
929                 printk(KERN_WARNING "Deletion dirent \"%s\" not found in list for ino #%u\n", fd->name, f->inocache->ino);
930         }
931         jffs2_mark_node_obsolete(c, fd->raw);
932         jffs2_free_full_dirent(fd);
933         return 0;
934 }
935
936 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
937                                       struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
938                                       uint32_t start, uint32_t end)
939 {
940         struct jffs2_raw_inode ri;
941         struct jffs2_node_frag *frag;
942         struct jffs2_full_dnode *new_fn;
943         uint32_t alloclen, phys_ofs, ilen;
944         int ret;
945
946         D1(printk(KERN_DEBUG "Writing replacement hole node for ino #%u from offset 0x%x to 0x%x\n",
947                   f->inocache->ino, start, end));
948
949         memset(&ri, 0, sizeof(ri));
950
951         if(fn->frags > 1) {
952                 size_t readlen;
953                 uint32_t crc;
954                 /* It's partially obsoleted by a later write. So we have to
955                    write it out again with the _same_ version as before */
956                 ret = jffs2_flash_read(c, ref_offset(fn->raw), sizeof(ri), &readlen, (char *)&ri);
957                 if (readlen != sizeof(ri) || ret) {
958                         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);
959                         goto fill;
960                 }
961                 if (je16_to_cpu(ri.nodetype) != JFFS2_NODETYPE_INODE) {
962                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had node type 0x%04x instead of JFFS2_NODETYPE_INODE(0x%04x)\n",
963                                ref_offset(fn->raw),
964                                je16_to_cpu(ri.nodetype), JFFS2_NODETYPE_INODE);
965                         return -EIO;
966                 }
967                 if (je32_to_cpu(ri.totlen) != sizeof(ri)) {
968                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had totlen 0x%x instead of expected 0x%zx\n",
969                                ref_offset(fn->raw),
970                                je32_to_cpu(ri.totlen), sizeof(ri));
971                         return -EIO;
972                 }
973                 crc = crc32(0, &ri, sizeof(ri)-8);
974                 if (crc != je32_to_cpu(ri.node_crc)) {
975                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had CRC 0x%08x which doesn't match calculated CRC 0x%08x\n",
976                                ref_offset(fn->raw),
977                                je32_to_cpu(ri.node_crc), crc);
978                         /* FIXME: We could possibly deal with this by writing new holes for each frag */
979                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
980                                start, end, f->inocache->ino);
981                         goto fill;
982                 }
983                 if (ri.compr != JFFS2_COMPR_ZERO) {
984                         printk(KERN_WARNING "jffs2_garbage_collect_hole: Node 0x%08x wasn't a hole node!\n", ref_offset(fn->raw));
985                         printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n",
986                                start, end, f->inocache->ino);
987                         goto fill;
988                 }
989         } else {
990         fill:
991                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
992                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
993                 ri.totlen = cpu_to_je32(sizeof(ri));
994                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
995
996                 ri.ino = cpu_to_je32(f->inocache->ino);
997                 ri.version = cpu_to_je32(++f->highest_version);
998                 ri.offset = cpu_to_je32(start);
999                 ri.dsize = cpu_to_je32(end - start);
1000                 ri.csize = cpu_to_je32(0);
1001                 ri.compr = JFFS2_COMPR_ZERO;
1002         }
1003
1004         frag = frag_last(&f->fragtree);
1005         if (frag)
1006                 /* Fetch the inode length from the fragtree rather then
1007                  * from i_size since i_size may have not been updated yet */
1008                 ilen = frag->ofs + frag->size;
1009         else
1010                 ilen = JFFS2_F_I_SIZE(f);
1011
1012         ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1013         ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1014         ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1015         ri.isize = cpu_to_je32(ilen);
1016         ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1017         ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1018         ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1019         ri.data_crc = cpu_to_je32(0);
1020         ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1021
1022         ret = jffs2_reserve_space_gc(c, sizeof(ri), &phys_ofs, &alloclen,
1023                                 JFFS2_SUMMARY_INODE_SIZE);
1024         if (ret) {
1025                 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_hole failed: %d\n",
1026                        sizeof(ri), ret);
1027                 return ret;
1028         }
1029         new_fn = jffs2_write_dnode(c, f, &ri, NULL, 0, phys_ofs, ALLOC_GC);
1030
1031         if (IS_ERR(new_fn)) {
1032                 printk(KERN_WARNING "Error writing new hole node: %ld\n", PTR_ERR(new_fn));
1033                 return PTR_ERR(new_fn);
1034         }
1035         if (je32_to_cpu(ri.version) == f->highest_version) {
1036                 jffs2_add_full_dnode_to_inode(c, f, new_fn);
1037                 if (f->metadata) {
1038                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1039                         jffs2_free_full_dnode(f->metadata);
1040                         f->metadata = NULL;
1041                 }
1042                 return 0;
1043         }
1044
1045         /*
1046          * We should only get here in the case where the node we are
1047          * replacing had more than one frag, so we kept the same version
1048          * number as before. (Except in case of error -- see 'goto fill;'
1049          * above.)
1050          */
1051         D1(if(unlikely(fn->frags <= 1)) {
1052                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Replacing fn with %d frag(s) but new ver %d != highest_version %d of ino #%d\n",
1053                        fn->frags, je32_to_cpu(ri.version), f->highest_version,
1054                        je32_to_cpu(ri.ino));
1055         });
1056
1057         /* This is a partially-overlapped hole node. Mark it REF_NORMAL not REF_PRISTINE */
1058         mark_ref_normal(new_fn->raw);
1059
1060         for (frag = jffs2_lookup_node_frag(&f->fragtree, fn->ofs);
1061              frag; frag = frag_next(frag)) {
1062                 if (frag->ofs > fn->size + fn->ofs)
1063                         break;
1064                 if (frag->node == fn) {
1065                         frag->node = new_fn;
1066                         new_fn->frags++;
1067                         fn->frags--;
1068                 }
1069         }
1070         if (fn->frags) {
1071                 printk(KERN_WARNING "jffs2_garbage_collect_hole: Old node still has frags!\n");
1072                 BUG();
1073         }
1074         if (!new_fn->frags) {
1075                 printk(KERN_WARNING "jffs2_garbage_collect_hole: New node has no frags!\n");
1076                 BUG();
1077         }
1078
1079         jffs2_mark_node_obsolete(c, fn->raw);
1080         jffs2_free_full_dnode(fn);
1081
1082         return 0;
1083 }
1084
1085 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb,
1086                                        struct jffs2_inode_info *f, struct jffs2_full_dnode *fn,
1087                                        uint32_t start, uint32_t end)
1088 {
1089         struct jffs2_full_dnode *new_fn;
1090         struct jffs2_raw_inode ri;
1091         uint32_t alloclen, phys_ofs, offset, orig_end, orig_start;
1092         int ret = 0;
1093         unsigned char *comprbuf = NULL, *writebuf;
1094         unsigned long pg;
1095         unsigned char *pg_ptr;
1096
1097         memset(&ri, 0, sizeof(ri));
1098
1099         D1(printk(KERN_DEBUG "Writing replacement dnode for ino #%u from offset 0x%x to 0x%x\n",
1100                   f->inocache->ino, start, end));
1101
1102         orig_end = end;
1103         orig_start = start;
1104
1105         if (c->nr_free_blocks + c->nr_erasing_blocks > c->resv_blocks_gcmerge) {
1106                 /* Attempt to do some merging. But only expand to cover logically
1107                    adjacent frags if the block containing them is already considered
1108                    to be dirty. Otherwise we end up with GC just going round in
1109                    circles dirtying the nodes it already wrote out, especially
1110                    on NAND where we have small eraseblocks and hence a much higher
1111                    chance of nodes having to be split to cross boundaries. */
1112
1113                 struct jffs2_node_frag *frag;
1114                 uint32_t min, max;
1115
1116                 min = start & ~(PAGE_CACHE_SIZE-1);
1117                 max = min + PAGE_CACHE_SIZE;
1118
1119                 frag = jffs2_lookup_node_frag(&f->fragtree, start);
1120
1121                 /* BUG_ON(!frag) but that'll happen anyway... */
1122
1123                 BUG_ON(frag->ofs != start);
1124
1125                 /* First grow down... */
1126                 while((frag = frag_prev(frag)) && frag->ofs >= min) {
1127
1128                         /* If the previous frag doesn't even reach the beginning, there's
1129                            excessive fragmentation. Just merge. */
1130                         if (frag->ofs > min) {
1131                                 D1(printk(KERN_DEBUG "Expanding down to cover partial frag (0x%x-0x%x)\n",
1132                                           frag->ofs, frag->ofs+frag->size));
1133                                 start = frag->ofs;
1134                                 continue;
1135                         }
1136                         /* OK. This frag holds the first byte of the page. */
1137                         if (!frag->node || !frag->node->raw) {
1138                                 D1(printk(KERN_DEBUG "First frag in page is hole (0x%x-0x%x). Not expanding down.\n",
1139                                           frag->ofs, frag->ofs+frag->size));
1140                                 break;
1141                         } else {
1142
1143                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1144                                    in a block which is still considered clean? If so, don't obsolete it.
1145                                    If not, cover it anyway. */
1146
1147                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1148                                 struct jffs2_eraseblock *jeb;
1149
1150                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1151
1152                                 if (jeb == c->gcblock) {
1153                                         D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1154                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1155                                         start = frag->ofs;
1156                                         break;
1157                                 }
1158                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1159                                         D1(printk(KERN_DEBUG "Not expanding down to cover frag (0x%x-0x%x) in clean block %08x\n",
1160                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1161                                         break;
1162                                 }
1163
1164                                 D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in dirty block %08x\n",
1165                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1166                                 start = frag->ofs;
1167                                 break;
1168                         }
1169                 }
1170
1171                 /* ... then up */
1172
1173                 /* Find last frag which is actually part of the node we're to GC. */
1174                 frag = jffs2_lookup_node_frag(&f->fragtree, end-1);
1175
1176                 while((frag = frag_next(frag)) && frag->ofs+frag->size <= max) {
1177
1178                         /* If the previous frag doesn't even reach the beginning, there's lots
1179                            of fragmentation. Just merge. */
1180                         if (frag->ofs+frag->size < max) {
1181                                 D1(printk(KERN_DEBUG "Expanding up to cover partial frag (0x%x-0x%x)\n",
1182                                           frag->ofs, frag->ofs+frag->size));
1183                                 end = frag->ofs + frag->size;
1184                                 continue;
1185                         }
1186
1187                         if (!frag->node || !frag->node->raw) {
1188                                 D1(printk(KERN_DEBUG "Last frag in page is hole (0x%x-0x%x). Not expanding up.\n",
1189                                           frag->ofs, frag->ofs+frag->size));
1190                                 break;
1191                         } else {
1192
1193                                 /* OK, it's a frag which extends to the beginning of the page. Does it live
1194                                    in a block which is still considered clean? If so, don't obsolete it.
1195                                    If not, cover it anyway. */
1196
1197                                 struct jffs2_raw_node_ref *raw = frag->node->raw;
1198                                 struct jffs2_eraseblock *jeb;
1199
1200                                 jeb = &c->blocks[raw->flash_offset / c->sector_size];
1201
1202                                 if (jeb == c->gcblock) {
1203                                         D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in gcblock at %08x\n",
1204                                                   frag->ofs, frag->ofs+frag->size, ref_offset(raw)));
1205                                         end = frag->ofs + frag->size;
1206                                         break;
1207                                 }
1208                                 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) {
1209                                         D1(printk(KERN_DEBUG "Not expanding up to cover frag (0x%x-0x%x) in clean block %08x\n",
1210                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1211                                         break;
1212                                 }
1213
1214                                 D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in dirty block %08x\n",
1215                                                   frag->ofs, frag->ofs+frag->size, jeb->offset));
1216                                 end = frag->ofs + frag->size;
1217                                 break;
1218                         }
1219                 }
1220                 D1(printk(KERN_DEBUG "Expanded dnode to write from (0x%x-0x%x) to (0x%x-0x%x)\n",
1221                           orig_start, orig_end, start, end));
1222
1223                 D1(BUG_ON(end > frag_last(&f->fragtree)->ofs + frag_last(&f->fragtree)->size));
1224                 BUG_ON(end < orig_end);
1225                 BUG_ON(start > orig_start);
1226         }
1227
1228         /* First, use readpage() to read the appropriate page into the page cache */
1229         /* Q: What happens if we actually try to GC the _same_ page for which commit_write()
1230          *    triggered garbage collection in the first place?
1231          * A: I _think_ it's OK. read_cache_page shouldn't deadlock, we'll write out the
1232          *    page OK. We'll actually write it out again in commit_write, which is a little
1233          *    suboptimal, but at least we're correct.
1234          */
1235         pg_ptr = jffs2_gc_fetch_page(c, f, start, &pg);
1236
1237         if (IS_ERR(pg_ptr)) {
1238                 printk(KERN_WARNING "read_cache_page() returned error: %ld\n", PTR_ERR(pg_ptr));
1239                 return PTR_ERR(pg_ptr);
1240         }
1241
1242         offset = start;
1243         while(offset < orig_end) {
1244                 uint32_t datalen;
1245                 uint32_t cdatalen;
1246                 uint16_t comprtype = JFFS2_COMPR_NONE;
1247
1248                 ret = jffs2_reserve_space_gc(c, sizeof(ri) + JFFS2_MIN_DATA_LEN, &phys_ofs,
1249                                         &alloclen, JFFS2_SUMMARY_INODE_SIZE);
1250
1251                 if (ret) {
1252                         printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dnode failed: %d\n",
1253                                sizeof(ri)+ JFFS2_MIN_DATA_LEN, ret);
1254                         break;
1255                 }
1256                 cdatalen = min_t(uint32_t, alloclen - sizeof(ri), end - offset);
1257                 datalen = end - offset;
1258
1259                 writebuf = pg_ptr + (offset & (PAGE_CACHE_SIZE -1));
1260
1261                 comprtype = jffs2_compress(c, f, writebuf, &comprbuf, &datalen, &cdatalen);
1262
1263                 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK);
1264                 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE);
1265                 ri.totlen = cpu_to_je32(sizeof(ri) + cdatalen);
1266                 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4));
1267
1268                 ri.ino = cpu_to_je32(f->inocache->ino);
1269                 ri.version = cpu_to_je32(++f->highest_version);
1270                 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f));
1271                 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f));
1272                 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f));
1273                 ri.isize = cpu_to_je32(JFFS2_F_I_SIZE(f));
1274                 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f));
1275                 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f));
1276                 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f));
1277                 ri.offset = cpu_to_je32(offset);
1278                 ri.csize = cpu_to_je32(cdatalen);
1279                 ri.dsize = cpu_to_je32(datalen);
1280                 ri.compr = comprtype & 0xff;
1281                 ri.usercompr = (comprtype >> 8) & 0xff;
1282                 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8));
1283                 ri.data_crc = cpu_to_je32(crc32(0, comprbuf, cdatalen));
1284
1285                 new_fn = jffs2_write_dnode(c, f, &ri, comprbuf, cdatalen, phys_ofs, ALLOC_GC);
1286
1287                 jffs2_free_comprbuf(comprbuf, writebuf);
1288
1289                 if (IS_ERR(new_fn)) {
1290                         printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn));
1291                         ret = PTR_ERR(new_fn);
1292                         break;
1293                 }
1294                 ret = jffs2_add_full_dnode_to_inode(c, f, new_fn);
1295                 offset += datalen;
1296                 if (f->metadata) {
1297                         jffs2_mark_node_obsolete(c, f->metadata->raw);
1298                         jffs2_free_full_dnode(f->metadata);
1299                         f->metadata = NULL;
1300                 }
1301         }
1302
1303         jffs2_gc_release_page(c, pg_ptr, &pg);
1304         return ret;
1305 }