| blob | 4a543e1149700b08141d6769814848ea9295e2c9 |
1 /*
2 * JFFS -- Journaling Flash File System, Linux implementation.
3 *
4 * Copyright (C) 1999, 2000 Axis Communications, Inc.
5 *
6 * Created by Finn Hakansson <finn@axis.com>.
7 *
8 * This is free software; you can redistribute it and/or modify it
9 * under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
12 *
13 * $Id: intrep.c,v 1.102 2001/09/23 23:28:36 dwmw2 Exp $
14 *
15 * Ported to Linux 2.3.x and MTD:
16 * Copyright (C) 2000 Alexander Larsson (alex@cendio.se), Cendio Systems AB
17 *
18 */
20 /* This file contains the code for the internal structure of the
21 Journaling Flash File System, JFFS. */
23 /*
24 * Todo list:
25 *
26 * memcpy_to_flash() and memcpy_from_flash() functions.
27 *
28 * Implementation of hard links.
29 *
30 * Organize the source code in a better way. Against the VFS we could
31 * have jffs_ext.c, and against the block device jffs_int.c.
32 * A better file-internal organization too.
33 *
34 * A better checksum algorithm.
35 *
36 * Consider endianness stuff. ntohl() etc.
37 *
38 * Are we handling the atime, mtime, ctime members of the inode right?
39 *
40 * Remove some duplicated code. Take a look at jffs_write_node() and
41 * jffs_rewrite_data() for instance.
42 *
43 * Implement more meaning of the nlink member in various data structures.
44 * nlink could be used in conjunction with hard links for instance.
45 *
46 * Better memory management. Allocate data structures in larger chunks
47 * if possible.
48 *
49 * If too much meta data is stored, a garbage collect should be issued.
50 * We have experienced problems with too much meta data with for instance
51 * log files.
52 *
53 * Improve the calls to jffs_ioctl(). We would like to retrieve more
54 * information to be able to debug (or to supervise) JFFS during run-time.
55 *
56 */
58 #include <linux/types.h>
59 #include <linux/slab.h>
60 #include <linux/jffs.h>
61 #include <linux/fs.h>
62 #include <linux/stat.h>
63 #include <linux/pagemap.h>
64 #include <linux/mutex.h>
65 #include <asm/byteorder.h>
66 #include <linux/smp_lock.h>
67 #include <linux/time.h>
68 #include <linux/ctype.h>
75 #if defined(JFFS_MEMORY_DEBUG) && JFFS_MEMORY_DEBUG
83 #endif
94 /* Is there enough space on the flash? */
96 {
103 }
110 }
111 }
112 }
114 #if CONFIG_JFFS_FS_VERBOSE > 0
115 static __u8
117 {
119 __u8 ret;
126 }
129 }
131 static void
133 {
143 }
147 }
150 }
151 }
153 /* Print empty space */
157 }
160 }
161 }
167 }
170 }
171 }
174 }
175 }
177 /* Print the contents of a node. */
178 static void
180 {
196 }
198 #endif
200 /* Print the contents of a raw inode. */
201 static void
203 {
233 }
235 #define flash_safe_acquire(arg)
236 #define flash_safe_release(arg)
239 static int
242 {
252 }
254 }
257 static __u32
259 {
261 __u32 ret;
268 }
271 }
274 static int
277 {
287 }
289 }
292 static int
295 {
306 }
307 /* Not implemented writev. Repeatedly use write - on the not so
308 unreasonable assumption that the mtd driver doesn't care how
309 many write cycles we use. */
320 }
321 /* If res is non-zero, retlen_a is undefined, but we don't
322 care because in that case it's not going to be
323 returned anyway.
324 */
327 }
329 }
332 static int
335 {
339 /* fill up pattern */
344 /* write as many 64-byte chunks as we can */
350 }
352 /* and the rest */
358 }
361 static void
363 {
369 }
372 static int
375 {
378 wait_queue_head_t wait_q;
392 /* FIXME: Use TASK_INTERRUPTIBLE and deal with being interrupted */
405 }
413 }
415 /* This routine calculates checksums in JFFS. */
416 static __u32
418 {
423 }
426 }
429 static int
431 {
437 /* Allocate read buffer */
442 }
443 /* Loop until checksum done */
445 /* Get amount of data to read */
448 else
451 /* Perform flash read */
455 /* Compute checksum */
459 /* Update pointer and size */
462 }
464 /* Free read buffer */
467 /* Return result */
471 }
474 {
475 // down(&fmc->wlock);
476 }
479 {
480 // up(&fmc->wlock);
481 }
484 /* Create and initialize a new struct jffs_file. */
488 {
494 }
495 no_jffs_file++;
503 }
506 /* Build a control block for the file system. */
509 {
519 }
527 }
533 }
538 fail_fminit:
540 fail_hash:
544 fail_control:
548 }
551 /* Clean up all data structures associated with the file system. */
552 void
554 {
560 }
567 }
569 /* Free all files and nodes. */
575 }
580 }
583 /* This function adds a virtual root node to the in-RAM representation.
584 Called by jffs_build_fs(). */
585 static int
587 {
596 }
597 no_jffs_file++;
600 no_jffs_file--;
602 }
616 }
619 /* This is where the file system is built and initialized. */
620 int
622 {
630 }
635 /* scan_flash() wants us to try once more. A flipping
636 bits sector was detect in the middle of the scan flash.
637 Clean up old allocated memory before going in.
638 */
644 }
650 }
653 }
654 }
656 /* Add a virtual root node if no one exists. */
660 }
661 }
669 }
673 }
675 /* Remove deleted nodes. */
679 }
680 /* Remove redundant nodes. (We are not interested in the
681 return value in this case.) */
683 /* Try to build a tree from all the nodes. */
687 }
688 /* Compute the sizes of all files in the filesystem. Adjust if
689 necessary. */
693 }
702 jffs_build_fs_fail:
708 /*
709 This checks for sectors that were being erased in their previous
710 lifetimes and for some reason or the other (power fail etc.),
711 the erase cycles never completed.
712 As the flash array would have reverted back to read status,
713 these sectors are detected by the symptom of the "flipping bits",
714 i.e. bits being read back differently from the same location in
715 flash if read multiple times.
716 The only solution to this is to re-erase the entire
717 sector.
718 Unfortunately detecting "flipping bits" is not a simple exercise
719 as a bit may be read back at 1 or 0 depending on the alignment
720 of the stars in the universe.
721 The level of confidence is in direct proportion to the number of
722 scans done. By power fail testing I (Vipin) have been able to
723 proove that reading twice is not enough.
724 Maybe 4 times? Change NUM_REREADS to a higher number if you want
725 a (even) higher degree of confidence in your mount process.
726 A higher number would of course slow down your mount.
727 */
732 usually set to kernel page size */
741 __u32 offset;
746 /* Allocate read buffers */
755 }
757 CHECK_NEXT:
772 /* buffers MUST match, double word for word! */
775 ){
776 /* flipping bits detected, time to erase sector */
777 /* This will help us log some statistics etc. */
785 /* calculate start of present sector */
789 offset));
802 offset));
803 /* skip ahead to the next sector */
807 }
808 }
809 }
810 }
812 }
814 returnBack:
818 D2(printk("check_partly_erased_sector():Done checking all sectors till offset 0x%x for flipping bits.\n",
827 /* Scan the whole flash memory in order to find all nodes in the
828 file systems. */
829 static int
831 {
836 __u32 checksum;
837 __u8 tmp_accurate;
838 __u16 tmp_chksum;
839 __u32 deleted_file;
841 loff_t start;
842 loff_t test_start;
846 __u32 offset;
848 __u32 free_chunk_size1;
849 __u32 free_chunk_size2;
854 free blocks found. This is NOT allowed
855 by the current jffs design.
856 */
858 of how much free space was rejected and
859 marked dirty
860 */
867 /*
868 check and make sure that any sector does not suffer
869 from the "partly erased, bit flipping syndrome" (TM Vipin :)
870 If so, offending sectors will be erased.
871 */
876 }
878 /* Allocate read buffer */
883 }
885 /* Start the scan. */
889 /* Remember the position from where we started this scan. */
894 /* We have found 0xffffffff at this position. We have to
895 scan the rest of the flash till the end or till
896 something else than 0xffffffff is found.
897 Keep going till we do not find JFFS_EMPTY_BITMASK
898 anymore */
914 JFFS_EMPTY_BITMASK)
916 }
919 else
921 }
926 /* If some free space ends in the middle of a sector,
927 treat it as dirty rather than clean.
928 This is to handle the case where one thread
929 allocated space for a node, but didn't get to
930 actually _write_ it before power was lost, leaving
931 a gap in the log. Shifting all node writes into
932 a single kernel thread will fix the original problem.
933 */
935 /* If there was free space in previous
936 sectors, don't mark that dirty too -
937 only from the beginning of this sector
938 (or from start)
939 */
945 /* free space started in the previous sector! */
950 /*
951 Count it in if we are still under NUMFREEALLOWED *and* it is
952 at least 1 erase sector in length. This will keep us from
953 picking any little ole' space as "free".
954 */
963 /* below, space from "start" to "pos" will be marked dirty. */
966 /* Being in here means that we have found at least an entire
967 erase sector size of free space ending on a sector boundary.
968 Keep track of free spaces accepted.
969 */
970 num_free_space++;
972 num_free_spc_not_accp++;
978 }
980 }
988 /* "Flipping bits" detected. This means that our scan for them
989 did not catch this offset. See check_partly_erased_sectors() for
990 more info.
991 */
998 /* calculate start of present sector */
1002 offset));
1014 }
1019 }
1021 /* Being in here means that we have found free space that ends on an erase sector
1022 boundary.
1023 Count it in if we are still under NUMFREEALLOWED *and* it is at least 1 erase
1024 sector in length. This will keep us from picking any little ole' space as "free".
1025 */
1028 /* We really don't do anything to mark space as free, except *not*
1029 mark it dirty and just advance the "pos" location pointer.
1030 It will automatically be picked up as free space.
1031 */
1032 num_free_space++;
1036 num_free_spc_not_accp++;
1042 /* Mark this space as dirty. We already have our free space. */
1047 }
1049 }
1054 }
1058 /* We have found 0x00000000 at this position. Scan as far
1059 as possible to find out how much is dirty. */
1072 /* We have probably found a new raw inode. */
1076 bad_inode:
1077 /* We're f*cked. This is not solved yet. We have
1078 to scan for the magic pattern. */
1080 "bad inode: "
1089 /* handle these in the main switch() loop */
1094 }
1095 }
1097 cont_scan:
1098 /* First, mark as dirty the region
1099 which really does contain crap. */
1102 NULL);
1107 /* We have found the beginning of an inode. Create a
1108 node for it unless there already is one available. */
1111 /* Free read buffer */
1114 /* Release the flash device */
1118 }
1120 }
1122 /* Read the next raw inode. */
1127 /* When we compute the checksum for the inode, we never
1128 count the 'accurate' or the 'checksum' fields. */
1144 "checksum = %u, "
1150 /* Reuse this unused struct jffs_node. */
1152 }
1154 /* Check the raw inode read so far. Start with the
1155 maximum length of the filename. */
1160 }
1168 }
1170 /* The node's data segment should not exceed a
1171 certain length. */
1177 }
1181 /* This shouldn't be necessary because a node that
1182 violates the flash boundaries shouldn't be written
1183 in the first place. */
1186 }
1188 /* Read the name. */
1199 "checksum = %u, "
1204 /* Reuse this unused struct jffs_node. */
1206 }
1209 }
1210 }
1212 /* Read the data, if it exists, in order to be sure it
1213 matches the checksum. */
1217 }
1222 /* Reuse this unused struct jffs_node. */
1224 }
1230 "checksum = %u, "
1235 /* Reuse this unused struct jffs_node. */
1237 }
1238 }
1240 check_node:
1242 /* Remember the highest inode number in the whole file
1243 system. This information will be used when assigning
1244 new files new inode numbers. */
1247 }
1254 /* Compute the offset to the actual data in the
1255 on-flash node. */
1256 node->fm_offset
1262 node);
1268 /* Free read buffer */
1271 /* Release the flash device */
1275 }
1281 }
1286 GFP_KERNEL);
1292 /* Release the flash device */
1295 /* Free read buffer */
1299 }
1304 }
1307 }
1313 /* Reuse this unused struct jffs_node. */
1314 }
1315 }
1320 }
1323 /* Free read buffer */
1329 }
1331 /* Return happy */
1335 /* This is to trap the "free size accounting screwed error. */
1347 Mounting this screwed up f/s will screw us up anyway.
1348 */
1349 }
1355 /* Insert any kind of node into the file system. Take care of data
1356 insertions and deletions. Also remove redundant information. The
1357 memory allocated for the `name' is regarded as "given away" in the
1358 caller's perspective. */
1359 int
1363 {
1372 /* If there doesn't exist an associated jffs_file, then
1373 create, initialize and insert one into the file system. */
1377 }
1380 }
1388 /* Now insert the node at the correct position into the file's
1389 version list. */
1391 /* This is the first node. */
1404 }
1407 /* Insert at the end of the list. I.e. this node is the
1408 newest one so far. */
1422 }
1424 /* Insert at the bottom of the list. */
1431 }
1432 }
1436 /* Search for the insertion position starting from
1437 the tail (newest node). */
1446 }
1448 }
1451 }
1452 }
1453 }
1455 /* Deletion is irreversible. If any 'deleted' node is ever
1456 written, the file is deleted */
1460 /* Perhaps update the name. */
1465 }
1467 GFP_KERNEL))) {
1469 }
1476 }
1483 }
1484 /* Once upon a time, we would call jffs_possibly_delete_file()
1485 here. That causes an oops if someone's still got the file
1486 open, so now we only do it in jffs_delete_inode()
1487 -- dwmw2
1488 */
1491 }
1498 }
1504 /* Unlink a jffs_node from the version list it is in. */
1508 {
1513 }
1518 }
1519 }
1522 /* Unlink a jffs_node from the range list it is in. */
1525 {
1528 }
1531 }
1534 }
1537 }
1538 }
1541 /* Function used by jffs_remove_redundant_nodes() below. This function
1542 classifies what kind of information a node adds to a file. */
1545 {
1550 }
1553 }
1555 }
1558 /* Remove redundant nodes from a file. Mark the on-flash memory
1559 as dirty. */
1560 static int
1562 {
1566 __u8 newest_type;
1567 __u8 mod_type;
1572 }
1574 /* What does the `newest_node' modify? */
1580 newest_type));
1582 /* Traverse the file's nodes and determine which of them that are
1583 superfluous. Yeah, this might look very complex at first
1584 glance but it is actually very simple. */
1590 && (mod_type
1594 /* Yes, this node is redundant. Remove it. */
1596 "Removing node: ino: %u, version: %u, "
1598 mod_type));
1603 }
1606 }
1607 }
1610 }
1613 /* Insert a file into the hash table. */
1614 static int
1616 {
1623 }
1626 /* Insert a file into the file system tree. */
1627 int
1629 {
1642 }
1648 }
1649 }
1654 }
1658 }
1661 /* Remove a file from the hash table. */
1662 static int
1664 {
1670 }
1673 /* Just remove the file from the parent's children. Don't free
1674 any memory. */
1675 int
1677 {
1683 }
1687 }
1690 }
1692 }
1695 /* Find a file with its inode number. */
1698 {
1710 );
1712 }
1715 );
1717 }
1720 /* Find a file in a directory. We are comparing the names. */
1723 {
1733 }
1734 }
1738 }
1744 }
1748 });
1751 }
1754 /* Write a raw inode that takes up a certain amount of space in the flash
1755 memory. At the end of the flash device, there is often space that is
1756 impossible to use. At these times we want to mark this space as not
1757 used. In the cases when the amount of space is greater or equal than
1758 a struct jffs_raw_inode, we write a "dummy node" that takes up this
1759 space. The space after the raw inode, if it exists, is left as it is.
1760 Since this space after the raw inode contains JFFS_EMPTY_BITMASK bytes,
1761 we can compute the checksum of it; we don't have to manipulate it any
1762 further.
1764 If the space left on the device is less than the size of a struct
1765 jffs_raw_inode, this space is filled with JFFS_DIRTY_BITMASK bytes.
1766 No raw inode is written this time. */
1767 static int
1769 {
1784 raw_inode.chksum
1795 }
1796 }
1801 }
1805 }
1808 /* Write a raw inode, possibly its name and possibly some data. */
1809 int
1815 {
1821 __u32 pos;
1832 /* If this node isn't something that will eventually let
1833 GC free even more space, then don't allow it unless
1834 there's at least max_chunk_size space still available
1835 */
1840 /* Fire the retrorockets and shoot the fruiton torpedoes, sir! */
1845 });
1850 });
1855 total_size));
1859 retry:
1864 /* Deadlocks suck. */
1870 }
1872 /* First try to allocate some flash memory. */
1876 /* Just out of space. GC and try again */
1880 total_size));
1882 }
1889 }
1892 }
1900 }
1903 /* The jffs_fm struct that we got is not good enough.
1904 Make that space dirty and try again */
1912 }
1914 }
1920 });
1924 /* Increment the version number here. We can't let the caller
1925 set it beforehand, because we might have had to do GC on a node
1926 of this file - and we'd end up reusing version numbers.
1927 */
1930 D1(printk (KERN_NOTICE "jffs_write_node(): setting version of %s to %d\n", f->name, raw_inode->version));
1932 /* if the file was deleted, set the deleted bit in the raw inode */
1935 }
1937 /* Compute the checksum for the data and name chunks. */
1941 /* The checksum is calculated without the chksum and accurate
1942 fields so set them to zero first. */
1953 /* The actual raw JFFS node */
1958 /* Get name and size if there is one */
1962 iovec_cnt++;
1966 /* Add some extra padding if necessary */
1970 iovec_cnt++;
1971 }
1972 }
1974 /* Get data and size if there is any */
1978 iovec_cnt++;
1979 /* No need to pad this because we're not actually putting
1980 anything after it.
1981 */
1982 }
1991 }
2001 /* Read data from the node and write it to the buffer. 'node_offset'
2002 is how much we have read from this particular node before and which
2003 shouldn't be read again. 'max_size' is how much space there is in
2004 the buffer. */
2005 static int
2008 {
2012 __u32 r;
2026 }
2029 /* Read data from the file's nodes. Write the data to the buffer
2030 'buf'. 'read_offset' tells how much data we should skip. */
2031 int
2033 __u32 size)
2034 {
2047 }
2049 /* First find the node to read data from. */
2056 }
2059 }
2060 }
2062 /* "Cats are living proof that not everything in nature
2063 has to be useful."
2064 - Garrison Keilor ('97) */
2066 /* Fill the buffer. */
2070 /* This node does not refer to real data. */
2074 }
2076 node_offset,
2079 }
2083 }
2086 }
2089 /* Used for traversing all nodes in the hash table. */
2090 int
2092 {
2100 /* We must do _safe, because 'func' might remove the
2101 current file 'f' from the list. */
2107 }
2108 }
2111 }
2114 /* Free all nodes associated with a file. */
2115 static int
2117 {
2129 }
2131 }
2134 /* Free a file and its name. */
2135 static int
2137 {
2144 }
2146 no_jffs_file--;
2148 }
2150 static long
2152 {
2154 }
2156 /* See if a file is deleted. If so, mark that file's nodes as obsolete. */
2157 int
2159 {
2168 });
2171 /* First try to remove all older versions. Commence with
2172 the oldest node. */
2176 }
2179 }
2180 }
2181 /* Unlink the file from the filesystem. */
2184 }
2188 }
2190 }
2193 /* Used in conjunction with jffs_foreach_file() to count the number
2194 of files in the file system. */
2195 int
2197 {
2199 }
2202 /* Build up a file's range list from scratch by going through the
2203 version list. */
2204 static int
2206 {
2214 }
2216 }
2219 /* Remove an amount of data from a file. If this amount of data is
2220 zero, that could mean that a node should be split in two parts.
2221 We remove or change the appropriate nodes in the lists.
2223 Starting offset of area to be removed is node->data_offset,
2224 and the length of the area is in node->removed_size. */
2225 static int
2227 {
2239 /* A simple append; nothing to remove or no node to split. */
2241 }
2243 /* Find the node where we should begin the removal. */
2247 }
2248 }
2250 /* If there's no data in the file there's no data to
2251 remove either. */
2253 }
2256 /* XXX: Not implemented yet. */
2259 }
2261 /* See if the node has to be split into two parts. */
2263 /* Do the split. */
2271 }
2286 }
2289 }
2295 }
2298 }
2299 /* A very interesting can of worms. */
2307 }
2310 }
2312 /* No. No need to split the node. Just remove
2313 the end of the node. */
2319 }
2320 }
2322 /* Remove as many nodes as necessary. */
2334 }
2339 }
2346 }
2347 }
2349 /* Adjust the following nodes' information about offsets etc. */
2353 }
2356 /* It's possible that the removed_size is in fact
2357 * greater than the amount of data we actually thought
2358 * were present in the first place - some of the nodes
2359 * which this node originally obsoleted may already have
2360 * been deleted from the flash by subsequent garbage
2361 * collection.
2362 *
2363 * If this is the case, don't let f->size go negative.
2364 * Bad things would happen :)
2365 */
2369 }
2375 /* Insert some data into a file. Prior to the call to this function,
2376 jffs_delete_data should be called. */
2377 static int
2379 {
2384 /* Find the position where we should insert data. */
2385 retry:
2387 /* A simple append. This is the most common operation. */
2392 }
2397 }
2398 }
2400 /* Trying to insert data into the middle of the file. This
2401 means no problem because jffs_delete_data() has already
2402 prepared the range list for us. */
2405 /* Find the correct place for the insertion and then insert
2406 the node. */
2414 }
2417 }
2421 }
2425 "Couldn't find a place to insert "
2428 });
2429 }
2431 /* Adjust later nodes' offsets etc. */
2436 }
2438 }
2440 /* Okay. This is tricky. This means that we want to insert
2441 data at a place that is beyond the limits of the file as
2442 it is constructed right now. This is actually a common
2443 event that for instance could occur during the mounting
2444 of the file system if a large file have been truncated,
2445 rewritten and then only partially garbage collected. */
2449 /* We need a place holder for the data that is missing in
2450 front of this insertion. This "virtual node" will not
2451 be associated with any space on the flash device. */
2455 }
2471 /* Are there any data at all in the file yet? */
2473 virtual_node->data_offset
2476 virtual_node->data_size
2480 }
2486 }
2491 /* Insert this virtual node in the version list as well. */
2497 virtual_node->version_prev
2499 }
2502 }
2505 }
2506 }
2510 /* Make a new try to insert the node. */
2512 }
2516 }
2519 /* A new node (with data) has been added to the file and now the range
2520 list has to be modified. */
2521 static int
2523 {
2531 /* data_offset == X */
2532 /* data_size == 0 */
2533 /* remove_size == 0 */
2534 }
2536 /* data_offset == X */
2537 /* data_size == 0 */
2538 /* remove_size != 0 */
2541 }
2542 }
2543 }
2545 /* data_offset == X */
2546 /* data_size != 0 */
2547 /* remove_size == Y */
2550 }
2553 }
2554 }
2556 }
2558 /* Print the contents of a file. */
2559 #if 0
2560 int
2562 {
2580 });
2598 }
2601 void
2603 {
2614 }
2615 }
2616 }
2619 void
2621 {
2628 }
2633 }
2645 }
2646 }
2649 }
2652 #if defined(JFFS_MEMORY_DEBUG) && JFFS_MEMORY_DEBUG
2653 void
2655 {
2668 }
2669 #endif
2672 /* Rewrite `size' bytes, and begin at `node'. */
2673 static int
2675 {
2681 __u32 pos;
2682 __u32 pos_dchksum;
2683 __u32 total_name_size;
2684 __u32 total_data_size;
2685 __u32 total_size;
2691 /* Create and initialize the new node. */
2696 }
2707 retry:
2717 }
2719 /* The jffs_fm struct that we got is not big enough. */
2720 /* This should never happen, because we deal with this case
2721 in jffs_garbage_collect_next().*/
2722 printk(KERN_WARNING "jffs_rewrite_data(): Allocated node is too small (%d bytes of %d)\n", fm->size, total_size);
2728 }
2734 }
2737 /* Initialize the raw inode. */
2780 }
2783 /* Write the name to the flash memory. */
2797 }
2800 }
2802 /* Write the data. */
2811 }
2821 "jffs_read_data() "
2824 }
2831 "Write error during rewrite. "
2834 }
2839 }
2842 }
2849 /* Add the checksum. */
2861 }
2863 /* Now make the file system aware of the newly written node. */
2872 /* jffs_garbage_collect_next implements one step in the garbage collect
2873 process and is often called multiple times at each occasion of a
2874 garbage collect. */
2876 static int
2878 {
2883 __u32 size;
2884 __u32 data_size;
2885 __u32 total_name_size;
2886 __u32 extra_available;
2887 __u32 space_needed;
2891 /* Get the oldest node in the flash. */
2900 });
2902 /* Find its corresponding file too. */
2907 "No file to garbage collect! "
2909 /* FIXME: Free the offending node and recover. */
2912 }
2914 /* We always write out the name. Theoretically, we don't need
2915 to, but for now it's easier - because otherwise we'd have
2916 to keep track of how many times the current name exists on
2917 the flash and make sure it never reaches zero.
2919 The current approach means that would be possible to cause
2920 the GC to end up eating its tail by writing lots of nodes
2921 with no name for it to garbage-collect. Hence the change in
2922 inode.c to write names with _every_ node.
2924 It sucks, but it _should_ work.
2925 */
2933 /* Compute how many data it's possible to rewrite at the moment. */
2936 /* And from that, the total size of the chunk we want to write */
2940 /* If that's more than max_chunk_size, reduce it accordingly */
2945 }
2947 /* If we're asking to take up more space than free_chunk_size1
2948 but we _could_ fit in it, shrink accordingly.
2949 */
2954 /* The space left is too small to be of any
2955 use really. */
2962 "jffs_garbage_collect_next: "
2963 "Failed to allocate `dirty' "
2967 }
2972 }
2980 }
2983 /* Calculate the amount of space needed to hold the nodes
2984 which are remaining in the tail */
2987 /* From that, calculate how much 'extra' space we can use to
2988 increase the size of the node we're writing from the size
2989 of the node we're obsoleting
2990 */
2992 /* If we've gone below min_free_size for some reason,
2993 don't fuck up. This is why we have
2994 min_free_size > sector_size. Whinge about it though,
2995 just so I can convince myself my maths is right.
2996 */
3003 }
3005 /* Check that we don't use up any more 'extra' space than
3006 what's available */
3020 };
3035 }
3037 jffs_garbage_collect_next_end:
3043 /* If an obsolete node is partly going to be erased due to garbage
3044 collection, the part that isn't going to be erased must be filled
3045 with zeroes so that the scan of the flash will work smoothly next
3046 time. (The data in the file could for instance be a JFFS image
3047 which could cause enormous confusion during a scan of the flash
3048 device if we didn't do this.)
3049 There are two phases in this procedure: First, the clearing of
3050 the name and data parts of the node. Second, possibly also clearing
3051 a part of the raw inode as well. If the box is power cycled during
3052 the first phase, only the checksum of this node-to-be-cleared-at-
3053 the-end will be wrong. If the box is power cycled during, or after,
3054 the clearing of the raw inode, the information like the length of
3055 the name and data parts are zeroed. The next time the box is
3056 powered up, the scanning algorithm manages this faulty data too
3057 because:
3059 - The checksum is invalid and thus the raw inode must be discarded
3060 in any case.
3061 - If the lengths of the data part or the name part are zeroed, the
3062 scanning just continues after the raw inode. But after the inode
3063 the scanning procedure just finds zeroes which is the same as
3064 dirt.
3066 So, in the end, this could never fail. :-) Even if it does fail,
3067 the scanning algorithm should manage that too. */
3069 static int
3071 {
3074 __u32 zero_offset;
3075 __u32 zero_size;
3076 __u32 zero_offset_data;
3077 __u32 zero_size_data;
3083 }
3085 /* Where and how much shall we clear? */
3089 /* Do we have to clear the raw_inode explicitly? */
3093 }
3095 /* First, clear the name and data fields. */
3102 /* Should we clear a part of the raw inode? */
3104 /* I guess it is ok to clear the raw inode in this order. */
3107 cutting_raw_inode);
3109 }
3114 /* Try to erase as much as possible of the dirt in the flash memory. */
3115 static long
3117 {
3121 __u32 offset;
3131 }
3136 }
3142 }
3146 /* Now, let's try to do the erase. */
3152 /* XXX: Here we should allocate this area as dirty
3153 with jffs_fmalloced or something similar. Now
3154 we just report the error. */
3156 }
3158 #if 0
3159 /* Check if the erased sectors really got erased. */
3160 {
3161 __u32 pos;
3162 __u32 end;
3179 }
3180 }
3186 }
3188 /* XXX: Here we should allocate the memory
3189 with jffs_fmalloced() in order to prevent
3190 JFFS from using this area accidentally. */
3192 }
3193 }
3194 #endif
3196 /* Update the flash memory data structures. */
3200 }
3203 /* There are different criteria that should trigger a garbage collect:
3205 1. There is too much dirt in the memory.
3206 2. The free space is becoming small.
3207 3. There are many versions of a node.
3209 The garbage collect should always be done in a manner that guarantees
3210 that future garbage collects cannot be locked. E.g. Rewritten chunks
3211 should not be too large (span more than one sector in the flash memory
3212 for exemple). Of course there is a limit on how intelligent this garbage
3213 collection can be. */
3216 static int
3218 {
3223 D2(printk("***jffs_garbage_collect_now(): fmc->dirty_size = %u, fmc->free_size = 0x%x\n, fcs1=0x%x, fcs2=0x%x",
3227 // down(&fmc->gclock);
3229 /* If it is possible to garbage collect, do so. */
3241 }
3246 /* Argh */
3250 }
3253 /* Actually, we _may_ have been able to free some,
3254 * if there are many overlapping nodes which aren't
3255 * actually marked dirty because they still have
3256 * some valid data in each.
3257 */
3260 }
3262 /* Let's dare to make a garbage collect. */
3265 "has gone seriously wrong "
3268 }
3272 }
3274 gc_end:
3275 // up(&fmc->gclock);
3281 });
3290 /* Determine if it is reasonable to start garbage collection.
3291 We start a gc pass if either:
3292 - The number of free bytes < MIN_FREE_BYTES && at least one
3293 block is dirty, OR
3294 - The number of dirty bytes > MAX_DIRTY_BYTES
3295 */
3297 {
3301 /* If there's not enough dirty space to free a block, there's no point. */
3305 }
3306 #if 1
3307 /* If there is too much RAM used by the various structures, GC */
3308 if (jffs_get_node_inuse() > (c->fmc->used_size/c->fmc->max_chunk_size * 5 + jffs_get_file_count() * 2 + 50)) {
3309 /* FIXME: Provide proof that this test can be satisfied. We
3310 don't want a filesystem doing endless GC just because this
3311 condition cannot ever be false.
3312 */
3315 }
3316 #endif
3317 /* If there are fewer free bytes than the threshold, GC */
3321 }
3322 /* If there are more dirty bytes than the threshold, GC */
3326 }
3327 /* FIXME: What about the "There are many versions of a node" condition? */
3330 }
3334 {
3335 /* NOTE: We rely on the fact that we have the BKL here.
3336 * Otherwise, the gc_task could go away between the check
3337 * and the wake_up_process()
3338 */
3341 }
3344 /* Kernel threads take (void *) as arguments. Thus we pass
3345 the jffs_control data as a (void *) and then cast it. */
3346 int
3348 {
3362 siginitsetinv (¤t->blocked, sigmask(SIGHUP) | sigmask(SIGKILL) | sigmask(SIGSTOP) | sigmask(SIGCONT));
3370 /* See if we need to start gc. If we don't, go to sleep.
3372 Current implementation is a BAD THING(tm). If we try
3373 to unmount the FS, the unmount operation will sleep waiting
3374 for this thread to exit. We need to arrange to send it a
3375 sig before the umount process sleeps.
3376 */
3382 on immediately - we're a low priority
3383 background task. */
3385 /* Put_super will send a SIGKILL and then wait on the sem.
3386 */
3388 siginfo_t info;
3409 }
3410 }
3425 }
3431 /* Argh. Might as well commit suicide. */
3434 // panic()
3436 }
3438 /* Let's dare to make a garbage collect. */
3441 "has gone seriously wrong "
3443 }
3445 gc_end: