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 = ref_next(raw); 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 spin_unlock(&c->erase_completion_lock); 260 if (ref_flags(raw) == REF_PRISTINE) { 261 /* It's an unknown node with JFFS2_FEATURE_RWCOMPAT_COPY */ 262 jffs2_garbage_collect_pristine(c, NULL, raw); 263 } else { 264 /* Just mark it obsolete */ 265 jffs2_mark_node_obsolete(c, raw); 266 } 267 up(&c->alloc_sem); 268 goto eraseit_lock; 269 } 270 271 ic = jffs2_raw_ref_to_ic(raw); 272 273 #ifdef CONFIG_JFFS2_FS_XATTR 274 /* When 'ic' refers xattr_datum/xattr_ref, this node is GCed as xattr. 275 * We can decide whether this node is inode or xattr by ic->class. */ 276 if (ic->class == RAWNODE_CLASS_XATTR_DATUM 277 || ic->class == RAWNODE_CLASS_XATTR_REF) { 278 BUG_ON(raw->next_in_ino != (void *)ic); 279 spin_unlock(&c->erase_completion_lock); 280 281 if (ic->class == RAWNODE_CLASS_XATTR_DATUM) { 282 ret = jffs2_garbage_collect_xattr_datum(c, (struct jffs2_xattr_datum *)ic); 283 } else { 284 ret = jffs2_garbage_collect_xattr_ref(c, (struct jffs2_xattr_ref *)ic); 285 } 286 goto release_sem; 287 } 288 #endif 289 290 /* We need to hold the inocache. Either the erase_completion_lock or 291 the inocache_lock are sufficient; we trade down since the inocache_lock 292 causes less contention. */ 293 spin_lock(&c->inocache_lock); 294 295 spin_unlock(&c->erase_completion_lock); 296 297 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)); 298 299 /* Three possibilities: 300 1. Inode is already in-core. We must iget it and do proper 301 updating to its fragtree, etc. 302 2. Inode is not in-core, node is REF_PRISTINE. We lock the 303 inocache to prevent a read_inode(), copy the node intact. 304 3. Inode is not in-core, node is not pristine. We must iget() 305 and take the slow path. 306 */ 307 308 switch(ic->state) { 309 case INO_STATE_CHECKEDABSENT: 310 /* It's been checked, but it's not currently in-core. 311 We can just copy any pristine nodes, but have 312 to prevent anyone else from doing read_inode() while 313 we're at it, so we set the state accordingly */ 314 if (ref_flags(raw) == REF_PRISTINE) 315 ic->state = INO_STATE_GC; 316 else { 317 D1(printk(KERN_DEBUG "Ino #%u is absent but node not REF_PRISTINE. Reading.\n", 318 ic->ino)); 319 } 320 break; 321 322 case INO_STATE_PRESENT: 323 /* It's in-core. GC must iget() it. */ 324 break; 325 326 case INO_STATE_UNCHECKED: 327 case INO_STATE_CHECKING: 328 case INO_STATE_GC: 329 /* Should never happen. We should have finished checking 330 by the time we actually start doing any GC, and since 331 we're holding the alloc_sem, no other garbage collection 332 can happen. 333 */ 334 printk(KERN_CRIT "Inode #%u already in state %d in jffs2_garbage_collect_pass()!\n", 335 ic->ino, ic->state); 336 up(&c->alloc_sem); 337 spin_unlock(&c->inocache_lock); 338 BUG(); 339 340 case INO_STATE_READING: 341 /* Someone's currently trying to read it. We must wait for 342 them to finish and then go through the full iget() route 343 to do the GC. However, sometimes read_inode() needs to get 344 the alloc_sem() (for marking nodes invalid) so we must 345 drop the alloc_sem before sleeping. */ 346 347 up(&c->alloc_sem); 348 D1(printk(KERN_DEBUG "jffs2_garbage_collect_pass() waiting for ino #%u in state %d\n", 349 ic->ino, ic->state)); 350 sleep_on_spinunlock(&c->inocache_wq, &c->inocache_lock); 351 /* And because we dropped the alloc_sem we must start again from the 352 beginning. Ponder chance of livelock here -- we're returning success 353 without actually making any progress. 354 355 Q: What are the chances that the inode is back in INO_STATE_READING 356 again by the time we next enter this function? And that this happens 357 enough times to cause a real delay? 358 359 A: Small enough that I don't care :) 360 */ 361 return 0; 362 } 363 364 /* OK. Now if the inode is in state INO_STATE_GC, we are going to copy the 365 node intact, and we don't have to muck about with the fragtree etc. 366 because we know it's not in-core. If it _was_ in-core, we go through 367 all the iget() crap anyway */ 368 369 if (ic->state == INO_STATE_GC) { 370 spin_unlock(&c->inocache_lock); 371 372 ret = jffs2_garbage_collect_pristine(c, ic, raw); 373 374 spin_lock(&c->inocache_lock); 375 ic->state = INO_STATE_CHECKEDABSENT; 376 wake_up(&c->inocache_wq); 377 378 if (ret != -EBADFD) { 379 spin_unlock(&c->inocache_lock); 380 goto release_sem; 381 } 382 383 /* Fall through if it wanted us to, with inocache_lock held */ 384 } 385 386 /* Prevent the fairly unlikely race where the gcblock is 387 entirely obsoleted by the final close of a file which had 388 the only valid nodes in the block, followed by erasure, 389 followed by freeing of the ic because the erased block(s) 390 held _all_ the nodes of that inode.... never been seen but 391 it's vaguely possible. */ 392 393 inum = ic->ino; 394 nlink = ic->nlink; 395 spin_unlock(&c->inocache_lock); 396 397 f = jffs2_gc_fetch_inode(c, inum, nlink); 398 if (IS_ERR(f)) { 399 ret = PTR_ERR(f); 400 goto release_sem; 401 } 402 if (!f) { 403 ret = 0; 404 goto release_sem; 405 } 406 407 ret = jffs2_garbage_collect_live(c, jeb, raw, f); 408 409 jffs2_gc_release_inode(c, f); 410 411 release_sem: 412 up(&c->alloc_sem); 413 414 eraseit_lock: 415 /* If we've finished this block, start it erasing */ 416 spin_lock(&c->erase_completion_lock); 417 418 eraseit: 419 if (c->gcblock && !c->gcblock->used_size) { 420 D1(printk(KERN_DEBUG "Block at 0x%08x completely obsoleted by GC. Moving to erase_pending_list\n", c->gcblock->offset)); 421 /* We're GC'ing an empty block? */ 422 list_add_tail(&c->gcblock->list, &c->erase_pending_list); 423 c->gcblock = NULL; 424 c->nr_erasing_blocks++; 425 jffs2_erase_pending_trigger(c); 426 } 427 spin_unlock(&c->erase_completion_lock); 428 429 return ret; 430 } 431 432 static int jffs2_garbage_collect_live(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb, 433 struct jffs2_raw_node_ref *raw, struct jffs2_inode_info *f) 434 { 435 struct jffs2_node_frag *frag; 436 struct jffs2_full_dnode *fn = NULL; 437 struct jffs2_full_dirent *fd; 438 uint32_t start = 0, end = 0, nrfrags = 0; 439 int ret = 0; 440 441 down(&f->sem); 442 443 /* Now we have the lock for this inode. Check that it's still the one at the head 444 of the list. */ 445 446 spin_lock(&c->erase_completion_lock); 447 448 if (c->gcblock != jeb) { 449 spin_unlock(&c->erase_completion_lock); 450 D1(printk(KERN_DEBUG "GC block is no longer gcblock. Restart\n")); 451 goto upnout; 452 } 453 if (ref_obsolete(raw)) { 454 spin_unlock(&c->erase_completion_lock); 455 D1(printk(KERN_DEBUG "node to be GC'd was obsoleted in the meantime.\n")); 456 /* They'll call again */ 457 goto upnout; 458 } 459 spin_unlock(&c->erase_completion_lock); 460 461 /* OK. Looks safe. And nobody can get us now because we have the semaphore. Move the block */ 462 if (f->metadata && f->metadata->raw == raw) { 463 fn = f->metadata; 464 ret = jffs2_garbage_collect_metadata(c, jeb, f, fn); 465 goto upnout; 466 } 467 468 /* FIXME. Read node and do lookup? */ 469 for (frag = frag_first(&f->fragtree); frag; frag = frag_next(frag)) { 470 if (frag->node && frag->node->raw == raw) { 471 fn = frag->node; 472 end = frag->ofs + frag->size; 473 if (!nrfrags++) 474 start = frag->ofs; 475 if (nrfrags == frag->node->frags) 476 break; /* We've found them all */ 477 } 478 } 479 if (fn) { 480 if (ref_flags(raw) == REF_PRISTINE) { 481 ret = jffs2_garbage_collect_pristine(c, f->inocache, raw); 482 if (!ret) { 483 /* Urgh. Return it sensibly. */ 484 frag->node->raw = f->inocache->nodes; 485 } 486 if (ret != -EBADFD) 487 goto upnout; 488 } 489 /* We found a datanode. Do the GC */ 490 if((start >> PAGE_CACHE_SHIFT) < ((end-1) >> PAGE_CACHE_SHIFT)) { 491 /* It crosses a page boundary. Therefore, it must be a hole. */ 492 ret = jffs2_garbage_collect_hole(c, jeb, f, fn, start, end); 493 } else { 494 /* It could still be a hole. But we GC the page this way anyway */ 495 ret = jffs2_garbage_collect_dnode(c, jeb, f, fn, start, end); 496 } 497 goto upnout; 498 } 499 500 /* Wasn't a dnode. Try dirent */ 501 for (fd = f->dents; fd; fd=fd->next) { 502 if (fd->raw == raw) 503 break; 504 } 505 506 if (fd && fd->ino) { 507 ret = jffs2_garbage_collect_dirent(c, jeb, f, fd); 508 } else if (fd) { 509 ret = jffs2_garbage_collect_deletion_dirent(c, jeb, f, fd); 510 } else { 511 printk(KERN_WARNING "Raw node at 0x%08x wasn't in node lists for ino #%u\n", 512 ref_offset(raw), f->inocache->ino); 513 if (ref_obsolete(raw)) { 514 printk(KERN_WARNING "But it's obsolete so we don't mind too much\n"); 515 } else { 516 jffs2_dbg_dump_node(c, ref_offset(raw)); 517 BUG(); 518 } 519 } 520 upnout: 521 up(&f->sem); 522 523 return ret; 524 } 525 526 static int jffs2_garbage_collect_pristine(struct jffs2_sb_info *c, 527 struct jffs2_inode_cache *ic, 528 struct jffs2_raw_node_ref *raw) 529 { 530 union jffs2_node_union *node; 531 size_t retlen; 532 int ret; 533 uint32_t phys_ofs, alloclen; 534 uint32_t crc, rawlen; 535 int retried = 0; 536 537 D1(printk(KERN_DEBUG "Going to GC REF_PRISTINE node at 0x%08x\n", ref_offset(raw))); 538 539 alloclen = rawlen = ref_totlen(c, c->gcblock, raw); 540 541 /* Ask for a small amount of space (or the totlen if smaller) because we 542 don't want to force wastage of the end of a block if splitting would 543 work. */ 544 if (ic && alloclen > sizeof(struct jffs2_raw_inode) + JFFS2_MIN_DATA_LEN) 545 alloclen = sizeof(struct jffs2_raw_inode) + JFFS2_MIN_DATA_LEN; 546 547 ret = jffs2_reserve_space_gc(c, alloclen, &alloclen, rawlen); 548 /* 'rawlen' is not the exact summary size; it is only an upper estimation */ 549 550 if (ret) 551 return ret; 552 553 if (alloclen < rawlen) { 554 /* Doesn't fit untouched. We'll go the old route and split it */ 555 return -EBADFD; 556 } 557 558 node = kmalloc(rawlen, GFP_KERNEL); 559 if (!node) 560 return -ENOMEM; 561 562 ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)node); 563 if (!ret && retlen != rawlen) 564 ret = -EIO; 565 if (ret) 566 goto out_node; 567 568 crc = crc32(0, node, sizeof(struct jffs2_unknown_node)-4); 569 if (je32_to_cpu(node->u.hdr_crc) != crc) { 570 printk(KERN_WARNING "Header CRC failed on REF_PRISTINE node at 0x%08x: Read 0x%08x, calculated 0x%08x\n", 571 ref_offset(raw), je32_to_cpu(node->u.hdr_crc), crc); 572 goto bail; 573 } 574 575 switch(je16_to_cpu(node->u.nodetype)) { 576 case JFFS2_NODETYPE_INODE: 577 crc = crc32(0, node, sizeof(node->i)-8); 578 if (je32_to_cpu(node->i.node_crc) != crc) { 579 printk(KERN_WARNING "Node CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n", 580 ref_offset(raw), je32_to_cpu(node->i.node_crc), crc); 581 goto bail; 582 } 583 584 if (je32_to_cpu(node->i.dsize)) { 585 crc = crc32(0, node->i.data, je32_to_cpu(node->i.csize)); 586 if (je32_to_cpu(node->i.data_crc) != crc) { 587 printk(KERN_WARNING "Data CRC failed on REF_PRISTINE data node at 0x%08x: Read 0x%08x, calculated 0x%08x\n", 588 ref_offset(raw), je32_to_cpu(node->i.data_crc), crc); 589 goto bail; 590 } 591 } 592 break; 593 594 case JFFS2_NODETYPE_DIRENT: 595 crc = crc32(0, node, sizeof(node->d)-8); 596 if (je32_to_cpu(node->d.node_crc) != crc) { 597 printk(KERN_WARNING "Node CRC failed on REF_PRISTINE dirent node at 0x%08x: Read 0x%08x, calculated 0x%08x\n", 598 ref_offset(raw), je32_to_cpu(node->d.node_crc), crc); 599 goto bail; 600 } 601 602 if (node->d.nsize) { 603 crc = crc32(0, node->d.name, node->d.nsize); 604 if (je32_to_cpu(node->d.name_crc) != crc) { 605 printk(KERN_WARNING "Name CRC failed on REF_PRISTINE dirent ode at 0x%08x: Read 0x%08x, calculated 0x%08x\n", 606 ref_offset(raw), je32_to_cpu(node->d.name_crc), crc); 607 goto bail; 608 } 609 } 610 break; 611 default: 612 /* If it's inode-less, we don't _know_ what it is. Just copy it intact */ 613 if (ic) { 614 printk(KERN_WARNING "Unknown node type for REF_PRISTINE node at 0x%08x: 0x%04x\n", 615 ref_offset(raw), je16_to_cpu(node->u.nodetype)); 616 goto bail; 617 } 618 } 619 620 /* OK, all the CRCs are good; this node can just be copied as-is. */ 621 retry: 622 phys_ofs = write_ofs(c); 623 624 ret = jffs2_flash_write(c, phys_ofs, rawlen, &retlen, (char *)node); 625 626 if (ret || (retlen != rawlen)) { 627 printk(KERN_NOTICE "Write of %d bytes at 0x%08x failed. returned %d, retlen %zd\n", 628 rawlen, phys_ofs, ret, retlen); 629 if (retlen) { 630 jffs2_add_physical_node_ref(c, phys_ofs | REF_OBSOLETE, rawlen, NULL); 631 } else { 632 printk(KERN_NOTICE "Not marking the space at 0x%08x as dirty because the flash driver returned retlen zero\n", phys_ofs); 633 } 634 if (!retried) { 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, &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 } 660 661 if (!ret) 662 ret = -EIO; 663 goto out_node; 664 } 665 jffs2_add_physical_node_ref(c, phys_ofs | REF_PRISTINE, rawlen, ic); 666 667 jffs2_mark_node_obsolete(c, raw); 668 D1(printk(KERN_DEBUG "WHEEE! GC REF_PRISTINE node at 0x%08x succeeded\n", ref_offset(raw))); 669 670 out_node: 671 kfree(node); 672 return ret; 673 bail: 674 ret = -EBADFD; 675 goto out_node; 676 } 677 678 static int jffs2_garbage_collect_metadata(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb, 679 struct jffs2_inode_info *f, struct jffs2_full_dnode *fn) 680 { 681 struct jffs2_full_dnode *new_fn; 682 struct jffs2_raw_inode ri; 683 struct jffs2_node_frag *last_frag; 684 union jffs2_device_node dev; 685 char *mdata = NULL, mdatalen = 0; 686 uint32_t alloclen, ilen; 687 int ret; 688 689 if (S_ISBLK(JFFS2_F_I_MODE(f)) || 690 S_ISCHR(JFFS2_F_I_MODE(f)) ) { 691 /* For these, we don't actually need to read the old node */ 692 mdatalen = jffs2_encode_dev(&dev, JFFS2_F_I_RDEV(f)); 693 mdata = (char *)&dev; 694 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bytes of kdev_t\n", mdatalen)); 695 } else if (S_ISLNK(JFFS2_F_I_MODE(f))) { 696 mdatalen = fn->size; 697 mdata = kmalloc(fn->size, GFP_KERNEL); 698 if (!mdata) { 699 printk(KERN_WARNING "kmalloc of mdata failed in jffs2_garbage_collect_metadata()\n"); 700 return -ENOMEM; 701 } 702 ret = jffs2_read_dnode(c, f, fn, mdata, 0, mdatalen); 703 if (ret) { 704 printk(KERN_WARNING "read of old metadata failed in jffs2_garbage_collect_metadata(): %d\n", ret); 705 kfree(mdata); 706 return ret; 707 } 708 D1(printk(KERN_DEBUG "jffs2_garbage_collect_metadata(): Writing %d bites of symlink target\n", mdatalen)); 709 710 } 711 712 ret = jffs2_reserve_space_gc(c, sizeof(ri) + mdatalen, &alloclen, 713 JFFS2_SUMMARY_INODE_SIZE); 714 if (ret) { 715 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_metadata failed: %d\n", 716 sizeof(ri)+ mdatalen, ret); 717 goto out; 718 } 719 720 last_frag = frag_last(&f->fragtree); 721 if (last_frag) 722 /* Fetch the inode length from the fragtree rather then 723 * from i_size since i_size may have not been updated yet */ 724 ilen = last_frag->ofs + last_frag->size; 725 else 726 ilen = JFFS2_F_I_SIZE(f); 727 728 memset(&ri, 0, sizeof(ri)); 729 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK); 730 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE); 731 ri.totlen = cpu_to_je32(sizeof(ri) + mdatalen); 732 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4)); 733 734 ri.ino = cpu_to_je32(f->inocache->ino); 735 ri.version = cpu_to_je32(++f->highest_version); 736 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f)); 737 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f)); 738 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f)); 739 ri.isize = cpu_to_je32(ilen); 740 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f)); 741 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f)); 742 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f)); 743 ri.offset = cpu_to_je32(0); 744 ri.csize = cpu_to_je32(mdatalen); 745 ri.dsize = cpu_to_je32(mdatalen); 746 ri.compr = JFFS2_COMPR_NONE; 747 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8)); 748 ri.data_crc = cpu_to_je32(crc32(0, mdata, mdatalen)); 749 750 new_fn = jffs2_write_dnode(c, f, &ri, mdata, mdatalen, ALLOC_GC); 751 752 if (IS_ERR(new_fn)) { 753 printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn)); 754 ret = PTR_ERR(new_fn); 755 goto out; 756 } 757 jffs2_mark_node_obsolete(c, fn->raw); 758 jffs2_free_full_dnode(fn); 759 f->metadata = new_fn; 760 out: 761 if (S_ISLNK(JFFS2_F_I_MODE(f))) 762 kfree(mdata); 763 return ret; 764 } 765 766 static int jffs2_garbage_collect_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb, 767 struct jffs2_inode_info *f, struct jffs2_full_dirent *fd) 768 { 769 struct jffs2_full_dirent *new_fd; 770 struct jffs2_raw_dirent rd; 771 uint32_t alloclen; 772 int ret; 773 774 rd.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK); 775 rd.nodetype = cpu_to_je16(JFFS2_NODETYPE_DIRENT); 776 rd.nsize = strlen(fd->name); 777 rd.totlen = cpu_to_je32(sizeof(rd) + rd.nsize); 778 rd.hdr_crc = cpu_to_je32(crc32(0, &rd, sizeof(struct jffs2_unknown_node)-4)); 779 780 rd.pino = cpu_to_je32(f->inocache->ino); 781 rd.version = cpu_to_je32(++f->highest_version); 782 rd.ino = cpu_to_je32(fd->ino); 783 /* If the times on this inode were set by explicit utime() they can be different, 784 so refrain from splatting them. */ 785 if (JFFS2_F_I_MTIME(f) == JFFS2_F_I_CTIME(f)) 786 rd.mctime = cpu_to_je32(JFFS2_F_I_MTIME(f)); 787 else 788 rd.mctime = cpu_to_je32(0); 789 rd.type = fd->type; 790 rd.node_crc = cpu_to_je32(crc32(0, &rd, sizeof(rd)-8)); 791 rd.name_crc = cpu_to_je32(crc32(0, fd->name, rd.nsize)); 792 793 ret = jffs2_reserve_space_gc(c, sizeof(rd)+rd.nsize, &alloclen, 794 JFFS2_SUMMARY_DIRENT_SIZE(rd.nsize)); 795 if (ret) { 796 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dirent failed: %d\n", 797 sizeof(rd)+rd.nsize, ret); 798 return ret; 799 } 800 new_fd = jffs2_write_dirent(c, f, &rd, fd->name, rd.nsize, ALLOC_GC); 801 802 if (IS_ERR(new_fd)) { 803 printk(KERN_WARNING "jffs2_write_dirent in garbage_collect_dirent failed: %ld\n", PTR_ERR(new_fd)); 804 return PTR_ERR(new_fd); 805 } 806 jffs2_add_fd_to_list(c, new_fd, &f->dents); 807 return 0; 808 } 809 810 static int jffs2_garbage_collect_deletion_dirent(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb, 811 struct jffs2_inode_info *f, struct jffs2_full_dirent *fd) 812 { 813 struct jffs2_full_dirent **fdp = &f->dents; 814 int found = 0; 815 816 /* On a medium where we can't actually mark nodes obsolete 817 pernamently, such as NAND flash, we need to work out 818 whether this deletion dirent is still needed to actively 819 delete a 'real' dirent with the same name that's still 820 somewhere else on the flash. */ 821 if (!jffs2_can_mark_obsolete(c)) { 822 struct jffs2_raw_dirent *rd; 823 struct jffs2_raw_node_ref *raw; 824 int ret; 825 size_t retlen; 826 int name_len = strlen(fd->name); 827 uint32_t name_crc = crc32(0, fd->name, name_len); 828 uint32_t rawlen = ref_totlen(c, jeb, fd->raw); 829 830 rd = kmalloc(rawlen, GFP_KERNEL); 831 if (!rd) 832 return -ENOMEM; 833 834 /* Prevent the erase code from nicking the obsolete node refs while 835 we're looking at them. I really don't like this extra lock but 836 can't see any alternative. Suggestions on a postcard to... */ 837 down(&c->erase_free_sem); 838 839 for (raw = f->inocache->nodes; raw != (void *)f->inocache; raw = raw->next_in_ino) { 840 841 /* We only care about obsolete ones */ 842 if (!(ref_obsolete(raw))) 843 continue; 844 845 /* Any dirent with the same name is going to have the same length... */ 846 if (ref_totlen(c, NULL, raw) != rawlen) 847 continue; 848 849 /* Doesn't matter if there's one in the same erase block. We're going to 850 delete it too at the same time. */ 851 if (SECTOR_ADDR(raw->flash_offset) == SECTOR_ADDR(fd->raw->flash_offset)) 852 continue; 853 854 D1(printk(KERN_DEBUG "Check potential deletion dirent at %08x\n", ref_offset(raw))); 855 856 /* This is an obsolete node belonging to the same directory, and it's of the right 857 length. We need to take a closer look...*/ 858 ret = jffs2_flash_read(c, ref_offset(raw), rawlen, &retlen, (char *)rd); 859 if (ret) { 860 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Read error (%d) reading obsolete node at %08x\n", ret, ref_offset(raw)); 861 /* If we can't read it, we don't need to continue to obsolete it. Continue */ 862 continue; 863 } 864 if (retlen != rawlen) { 865 printk(KERN_WARNING "jffs2_g_c_deletion_dirent(): Short read (%zd not %u) reading header from obsolete node at %08x\n", 866 retlen, rawlen, ref_offset(raw)); 867 continue; 868 } 869 870 if (je16_to_cpu(rd->nodetype) != JFFS2_NODETYPE_DIRENT) 871 continue; 872 873 /* If the name CRC doesn't match, skip */ 874 if (je32_to_cpu(rd->name_crc) != name_crc) 875 continue; 876 877 /* If the name length doesn't match, or it's another deletion dirent, skip */ 878 if (rd->nsize != name_len || !je32_to_cpu(rd->ino)) 879 continue; 880 881 /* OK, check the actual name now */ 882 if (memcmp(rd->name, fd->name, name_len)) 883 continue; 884 885 /* OK. The name really does match. There really is still an older node on 886 the flash which our deletion dirent obsoletes. So we have to write out 887 a new deletion dirent to replace it */ 888 up(&c->erase_free_sem); 889 890 D1(printk(KERN_DEBUG "Deletion dirent at %08x still obsoletes real dirent \"%s\" at %08x for ino #%u\n", 891 ref_offset(fd->raw), fd->name, ref_offset(raw), je32_to_cpu(rd->ino))); 892 kfree(rd); 893 894 return jffs2_garbage_collect_dirent(c, jeb, f, fd); 895 } 896 897 up(&c->erase_free_sem); 898 kfree(rd); 899 } 900 901 /* FIXME: If we're deleting a dirent which contains the current mtime and ctime, 902 we should update the metadata node with those times accordingly */ 903 904 /* No need for it any more. Just mark it obsolete and remove it from the list */ 905 while (*fdp) { 906 if ((*fdp) == fd) { 907 found = 1; 908 *fdp = fd->next; 909 break; 910 } 911 fdp = &(*fdp)->next; 912 } 913 if (!found) { 914 printk(KERN_WARNING "Deletion dirent \"%s\" not found in list for ino #%u\n", fd->name, f->inocache->ino); 915 } 916 jffs2_mark_node_obsolete(c, fd->raw); 917 jffs2_free_full_dirent(fd); 918 return 0; 919 } 920 921 static int jffs2_garbage_collect_hole(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb, 922 struct jffs2_inode_info *f, struct jffs2_full_dnode *fn, 923 uint32_t start, uint32_t end) 924 { 925 struct jffs2_raw_inode ri; 926 struct jffs2_node_frag *frag; 927 struct jffs2_full_dnode *new_fn; 928 uint32_t alloclen, ilen; 929 int ret; 930 931 D1(printk(KERN_DEBUG "Writing replacement hole node for ino #%u from offset 0x%x to 0x%x\n", 932 f->inocache->ino, start, end)); 933 934 memset(&ri, 0, sizeof(ri)); 935 936 if(fn->frags > 1) { 937 size_t readlen; 938 uint32_t crc; 939 /* It's partially obsoleted by a later write. So we have to 940 write it out again with the _same_ version as before */ 941 ret = jffs2_flash_read(c, ref_offset(fn->raw), sizeof(ri), &readlen, (char *)&ri); 942 if (readlen != sizeof(ri) || ret) { 943 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); 944 goto fill; 945 } 946 if (je16_to_cpu(ri.nodetype) != JFFS2_NODETYPE_INODE) { 947 printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had node type 0x%04x instead of JFFS2_NODETYPE_INODE(0x%04x)\n", 948 ref_offset(fn->raw), 949 je16_to_cpu(ri.nodetype), JFFS2_NODETYPE_INODE); 950 return -EIO; 951 } 952 if (je32_to_cpu(ri.totlen) != sizeof(ri)) { 953 printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had totlen 0x%x instead of expected 0x%zx\n", 954 ref_offset(fn->raw), 955 je32_to_cpu(ri.totlen), sizeof(ri)); 956 return -EIO; 957 } 958 crc = crc32(0, &ri, sizeof(ri)-8); 959 if (crc != je32_to_cpu(ri.node_crc)) { 960 printk(KERN_WARNING "jffs2_garbage_collect_hole: Node at 0x%08x had CRC 0x%08x which doesn't match calculated CRC 0x%08x\n", 961 ref_offset(fn->raw), 962 je32_to_cpu(ri.node_crc), crc); 963 /* FIXME: We could possibly deal with this by writing new holes for each frag */ 964 printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n", 965 start, end, f->inocache->ino); 966 goto fill; 967 } 968 if (ri.compr != JFFS2_COMPR_ZERO) { 969 printk(KERN_WARNING "jffs2_garbage_collect_hole: Node 0x%08x wasn't a hole node!\n", ref_offset(fn->raw)); 970 printk(KERN_WARNING "Data in the range 0x%08x to 0x%08x of inode #%u will be lost\n", 971 start, end, f->inocache->ino); 972 goto fill; 973 } 974 } else { 975 fill: 976 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK); 977 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE); 978 ri.totlen = cpu_to_je32(sizeof(ri)); 979 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4)); 980 981 ri.ino = cpu_to_je32(f->inocache->ino); 982 ri.version = cpu_to_je32(++f->highest_version); 983 ri.offset = cpu_to_je32(start); 984 ri.dsize = cpu_to_je32(end - start); 985 ri.csize = cpu_to_je32(0); 986 ri.compr = JFFS2_COMPR_ZERO; 987 } 988 989 frag = frag_last(&f->fragtree); 990 if (frag) 991 /* Fetch the inode length from the fragtree rather then 992 * from i_size since i_size may have not been updated yet */ 993 ilen = frag->ofs + frag->size; 994 else 995 ilen = JFFS2_F_I_SIZE(f); 996 997 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f)); 998 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f)); 999 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f)); 1000 ri.isize = cpu_to_je32(ilen); 1001 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f)); 1002 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f)); 1003 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f)); 1004 ri.data_crc = cpu_to_je32(0); 1005 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8)); 1006 1007 ret = jffs2_reserve_space_gc(c, sizeof(ri), &alloclen, 1008 JFFS2_SUMMARY_INODE_SIZE); 1009 if (ret) { 1010 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_hole failed: %d\n", 1011 sizeof(ri), ret); 1012 return ret; 1013 } 1014 new_fn = jffs2_write_dnode(c, f, &ri, NULL, 0, ALLOC_GC); 1015 1016 if (IS_ERR(new_fn)) { 1017 printk(KERN_WARNING "Error writing new hole node: %ld\n", PTR_ERR(new_fn)); 1018 return PTR_ERR(new_fn); 1019 } 1020 if (je32_to_cpu(ri.version) == f->highest_version) { 1021 jffs2_add_full_dnode_to_inode(c, f, new_fn); 1022 if (f->metadata) { 1023 jffs2_mark_node_obsolete(c, f->metadata->raw); 1024 jffs2_free_full_dnode(f->metadata); 1025 f->metadata = NULL; 1026 } 1027 return 0; 1028 } 1029 1030 /* 1031 * We should only get here in the case where the node we are 1032 * replacing had more than one frag, so we kept the same version 1033 * number as before. (Except in case of error -- see 'goto fill;' 1034 * above.) 1035 */ 1036 D1(if(unlikely(fn->frags <= 1)) { 1037 printk(KERN_WARNING "jffs2_garbage_collect_hole: Replacing fn with %d frag(s) but new ver %d != highest_version %d of ino #%d\n", 1038 fn->frags, je32_to_cpu(ri.version), f->highest_version, 1039 je32_to_cpu(ri.ino)); 1040 }); 1041 1042 /* This is a partially-overlapped hole node. Mark it REF_NORMAL not REF_PRISTINE */ 1043 mark_ref_normal(new_fn->raw); 1044 1045 for (frag = jffs2_lookup_node_frag(&f->fragtree, fn->ofs); 1046 frag; frag = frag_next(frag)) { 1047 if (frag->ofs > fn->size + fn->ofs) 1048 break; 1049 if (frag->node == fn) { 1050 frag->node = new_fn; 1051 new_fn->frags++; 1052 fn->frags--; 1053 } 1054 } 1055 if (fn->frags) { 1056 printk(KERN_WARNING "jffs2_garbage_collect_hole: Old node still has frags!\n"); 1057 BUG(); 1058 } 1059 if (!new_fn->frags) { 1060 printk(KERN_WARNING "jffs2_garbage_collect_hole: New node has no frags!\n"); 1061 BUG(); 1062 } 1063 1064 jffs2_mark_node_obsolete(c, fn->raw); 1065 jffs2_free_full_dnode(fn); 1066 1067 return 0; 1068 } 1069 1070 static int jffs2_garbage_collect_dnode(struct jffs2_sb_info *c, struct jffs2_eraseblock *jeb, 1071 struct jffs2_inode_info *f, struct jffs2_full_dnode *fn, 1072 uint32_t start, uint32_t end) 1073 { 1074 struct jffs2_full_dnode *new_fn; 1075 struct jffs2_raw_inode ri; 1076 uint32_t alloclen, offset, orig_end, orig_start; 1077 int ret = 0; 1078 unsigned char *comprbuf = NULL, *writebuf; 1079 unsigned long pg; 1080 unsigned char *pg_ptr; 1081 1082 memset(&ri, 0, sizeof(ri)); 1083 1084 D1(printk(KERN_DEBUG "Writing replacement dnode for ino #%u from offset 0x%x to 0x%x\n", 1085 f->inocache->ino, start, end)); 1086 1087 orig_end = end; 1088 orig_start = start; 1089 1090 if (c->nr_free_blocks + c->nr_erasing_blocks > c->resv_blocks_gcmerge) { 1091 /* Attempt to do some merging. But only expand to cover logically 1092 adjacent frags if the block containing them is already considered 1093 to be dirty. Otherwise we end up with GC just going round in 1094 circles dirtying the nodes it already wrote out, especially 1095 on NAND where we have small eraseblocks and hence a much higher 1096 chance of nodes having to be split to cross boundaries. */ 1097 1098 struct jffs2_node_frag *frag; 1099 uint32_t min, max; 1100 1101 min = start & ~(PAGE_CACHE_SIZE-1); 1102 max = min + PAGE_CACHE_SIZE; 1103 1104 frag = jffs2_lookup_node_frag(&f->fragtree, start); 1105 1106 /* BUG_ON(!frag) but that'll happen anyway... */ 1107 1108 BUG_ON(frag->ofs != start); 1109 1110 /* First grow down... */ 1111 while((frag = frag_prev(frag)) && frag->ofs >= min) { 1112 1113 /* If the previous frag doesn't even reach the beginning, there's 1114 excessive fragmentation. Just merge. */ 1115 if (frag->ofs > min) { 1116 D1(printk(KERN_DEBUG "Expanding down to cover partial frag (0x%x-0x%x)\n", 1117 frag->ofs, frag->ofs+frag->size)); 1118 start = frag->ofs; 1119 continue; 1120 } 1121 /* OK. This frag holds the first byte of the page. */ 1122 if (!frag->node || !frag->node->raw) { 1123 D1(printk(KERN_DEBUG "First frag in page is hole (0x%x-0x%x). Not expanding down.\n", 1124 frag->ofs, frag->ofs+frag->size)); 1125 break; 1126 } else { 1127 1128 /* OK, it's a frag which extends to the beginning of the page. Does it live 1129 in a block which is still considered clean? If so, don't obsolete it. 1130 If not, cover it anyway. */ 1131 1132 struct jffs2_raw_node_ref *raw = frag->node->raw; 1133 struct jffs2_eraseblock *jeb; 1134 1135 jeb = &c->blocks[raw->flash_offset / c->sector_size]; 1136 1137 if (jeb == c->gcblock) { 1138 D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in gcblock at %08x\n", 1139 frag->ofs, frag->ofs+frag->size, ref_offset(raw))); 1140 start = frag->ofs; 1141 break; 1142 } 1143 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) { 1144 D1(printk(KERN_DEBUG "Not expanding down to cover frag (0x%x-0x%x) in clean block %08x\n", 1145 frag->ofs, frag->ofs+frag->size, jeb->offset)); 1146 break; 1147 } 1148 1149 D1(printk(KERN_DEBUG "Expanding down to cover frag (0x%x-0x%x) in dirty block %08x\n", 1150 frag->ofs, frag->ofs+frag->size, jeb->offset)); 1151 start = frag->ofs; 1152 break; 1153 } 1154 } 1155 1156 /* ... then up */ 1157 1158 /* Find last frag which is actually part of the node we're to GC. */ 1159 frag = jffs2_lookup_node_frag(&f->fragtree, end-1); 1160 1161 while((frag = frag_next(frag)) && frag->ofs+frag->size <= max) { 1162 1163 /* If the previous frag doesn't even reach the beginning, there's lots 1164 of fragmentation. Just merge. */ 1165 if (frag->ofs+frag->size < max) { 1166 D1(printk(KERN_DEBUG "Expanding up to cover partial frag (0x%x-0x%x)\n", 1167 frag->ofs, frag->ofs+frag->size)); 1168 end = frag->ofs + frag->size; 1169 continue; 1170 } 1171 1172 if (!frag->node || !frag->node->raw) { 1173 D1(printk(KERN_DEBUG "Last frag in page is hole (0x%x-0x%x). Not expanding up.\n", 1174 frag->ofs, frag->ofs+frag->size)); 1175 break; 1176 } else { 1177 1178 /* OK, it's a frag which extends to the beginning of the page. Does it live 1179 in a block which is still considered clean? If so, don't obsolete it. 1180 If not, cover it anyway. */ 1181 1182 struct jffs2_raw_node_ref *raw = frag->node->raw; 1183 struct jffs2_eraseblock *jeb; 1184 1185 jeb = &c->blocks[raw->flash_offset / c->sector_size]; 1186 1187 if (jeb == c->gcblock) { 1188 D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in gcblock at %08x\n", 1189 frag->ofs, frag->ofs+frag->size, ref_offset(raw))); 1190 end = frag->ofs + frag->size; 1191 break; 1192 } 1193 if (!ISDIRTY(jeb->dirty_size + jeb->wasted_size)) { 1194 D1(printk(KERN_DEBUG "Not expanding up to cover frag (0x%x-0x%x) in clean block %08x\n", 1195 frag->ofs, frag->ofs+frag->size, jeb->offset)); 1196 break; 1197 } 1198 1199 D1(printk(KERN_DEBUG "Expanding up to cover frag (0x%x-0x%x) in dirty block %08x\n", 1200 frag->ofs, frag->ofs+frag->size, jeb->offset)); 1201 end = frag->ofs + frag->size; 1202 break; 1203 } 1204 } 1205 D1(printk(KERN_DEBUG "Expanded dnode to write from (0x%x-0x%x) to (0x%x-0x%x)\n", 1206 orig_start, orig_end, start, end)); 1207 1208 D1(BUG_ON(end > frag_last(&f->fragtree)->ofs + frag_last(&f->fragtree)->size)); 1209 BUG_ON(end < orig_end); 1210 BUG_ON(start > orig_start); 1211 } 1212 1213 /* First, use readpage() to read the appropriate page into the page cache */ 1214 /* Q: What happens if we actually try to GC the _same_ page for which commit_write() 1215 * triggered garbage collection in the first place? 1216 * A: I _think_ it's OK. read_cache_page shouldn't deadlock, we'll write out the 1217 * page OK. We'll actually write it out again in commit_write, which is a little 1218 * suboptimal, but at least we're correct. 1219 */ 1220 pg_ptr = jffs2_gc_fetch_page(c, f, start, &pg); 1221 1222 if (IS_ERR(pg_ptr)) { 1223 printk(KERN_WARNING "read_cache_page() returned error: %ld\n", PTR_ERR(pg_ptr)); 1224 return PTR_ERR(pg_ptr); 1225 } 1226 1227 offset = start; 1228 while(offset < orig_end) { 1229 uint32_t datalen; 1230 uint32_t cdatalen; 1231 uint16_t comprtype = JFFS2_COMPR_NONE; 1232 1233 ret = jffs2_reserve_space_gc(c, sizeof(ri) + JFFS2_MIN_DATA_LEN, 1234 &alloclen, JFFS2_SUMMARY_INODE_SIZE); 1235 1236 if (ret) { 1237 printk(KERN_WARNING "jffs2_reserve_space_gc of %zd bytes for garbage_collect_dnode failed: %d\n", 1238 sizeof(ri)+ JFFS2_MIN_DATA_LEN, ret); 1239 break; 1240 } 1241 cdatalen = min_t(uint32_t, alloclen - sizeof(ri), end - offset); 1242 datalen = end - offset; 1243 1244 writebuf = pg_ptr + (offset & (PAGE_CACHE_SIZE -1)); 1245 1246 comprtype = jffs2_compress(c, f, writebuf, &comprbuf, &datalen, &cdatalen); 1247 1248 ri.magic = cpu_to_je16(JFFS2_MAGIC_BITMASK); 1249 ri.nodetype = cpu_to_je16(JFFS2_NODETYPE_INODE); 1250 ri.totlen = cpu_to_je32(sizeof(ri) + cdatalen); 1251 ri.hdr_crc = cpu_to_je32(crc32(0, &ri, sizeof(struct jffs2_unknown_node)-4)); 1252 1253 ri.ino = cpu_to_je32(f->inocache->ino); 1254 ri.version = cpu_to_je32(++f->highest_version); 1255 ri.mode = cpu_to_jemode(JFFS2_F_I_MODE(f)); 1256 ri.uid = cpu_to_je16(JFFS2_F_I_UID(f)); 1257 ri.gid = cpu_to_je16(JFFS2_F_I_GID(f)); 1258 ri.isize = cpu_to_je32(JFFS2_F_I_SIZE(f)); 1259 ri.atime = cpu_to_je32(JFFS2_F_I_ATIME(f)); 1260 ri.ctime = cpu_to_je32(JFFS2_F_I_CTIME(f)); 1261 ri.mtime = cpu_to_je32(JFFS2_F_I_MTIME(f)); 1262 ri.offset = cpu_to_je32(offset); 1263 ri.csize = cpu_to_je32(cdatalen); 1264 ri.dsize = cpu_to_je32(datalen); 1265 ri.compr = comprtype & 0xff; 1266 ri.usercompr = (comprtype >> 8) & 0xff; 1267 ri.node_crc = cpu_to_je32(crc32(0, &ri, sizeof(ri)-8)); 1268 ri.data_crc = cpu_to_je32(crc32(0, comprbuf, cdatalen)); 1269 1270 new_fn = jffs2_write_dnode(c, f, &ri, comprbuf, cdatalen, ALLOC_GC); 1271 1272 jffs2_free_comprbuf(comprbuf, writebuf); 1273 1274 if (IS_ERR(new_fn)) { 1275 printk(KERN_WARNING "Error writing new dnode: %ld\n", PTR_ERR(new_fn)); 1276 ret = PTR_ERR(new_fn); 1277 break; 1278 } 1279 ret = jffs2_add_full_dnode_to_inode(c, f, new_fn); 1280 offset += datalen; 1281 if (f->metadata) { 1282 jffs2_mark_node_obsolete(c, f->metadata->raw); 1283 jffs2_free_full_dnode(f->metadata); 1284 f->metadata = NULL; 1285 } 1286 } 1287 1288 jffs2_gc_release_page(c, pg_ptr, &pg); 1289 return ret; 1290 } 1291