On 2015/11/06, 00:12, "Dilger, Andreas" <andreas.dilger@xxxxxxxxx> wrote: >Running e2fsck -fD on a large extent+htree directory (> 300k entries, >1600+ filesystem blocks) may result in the directory becoming corrupted. >This is definitely caused by a bug in the code rather than hardware, as >this corrupted multiple large directories on different systems. Thanks to a suggestion from Darrick, I was able to reproduce this problem with an e2fsck test script (attached) when shrinking an htree extent directory with only 3 index blocks referenced directly by the inode. The problem is not present on block-mapped directories but looks to be a danger for any user of the "-fD" option with extent-mapped directories. It looks like the problem is if the inode shrinks enough that one of the index blocks is dropped from the end of the file (blocks after logical block 114 were freed), but the write_directory() write_dir_block() iterator doesn't free the index block 800: : write_dir_block 113:583 - write write_dir_block 114:587 - write write_dir_block 115:591 - free write_dir_block 116:595 - free : : write_dir_block 165:791 - free write_dir_block -1:800 - skip write_dir_block 166:795 - free write_dir_block 167:799 - free write_dir_block 168:804 - free write_dir_block 169:808 - free write_dir_block 170:812 - free write_dir_block 171:813 - free write_dir_block 172:814 - free write_dir_block -1:800 - skip Pass 4: Checking reference counts Pass 5: Checking group summary information The extent tree now has a bogus index block at the end, but somehow is also missing the valid extent block that was holding the rest of the file, as shown by debugfs (after "e2fsck -fD" but before the second e2fsck that detects the corruption) and logical blocks 83-114 are lost: debugfs: stat subdir Inode: 12 Type: directory Mode: 0755 Flags: 0x81000 Generation: 0 Version: 0x00000000 User: 0 Group: 0 Size: 117760 File ACL: 0 Directory ACL: 0 Links: 2 Blockcount: 238 Fragment: Address: 0 Number: 0 Size: 0 ctime: 0x5642e764 -- Tue Nov 10 23:59:48 2015 atime: 0x5642e764 -- Tue Nov 10 23:59:48 2015 mtime: 0x5642e764 -- Tue Nov 10 23:59:48 2015 EXTENTS: (ETB0):146, (0):129, (1):133, (2):137, (3):141, (4):145, (5):150, (6):154, (7):158, (8):162, (9):166, (10):170, (11):174, (12):178, (13):182, (14):186, (15):190, (16):194, (17):198, (18):202, (19):206, (20):210, (21):214, (22):218, (23):222, (24):226, (25):230, (26):234, (27):238, (28):242, (29):246, (30):250, (31):254, (32):258, (33):262, (34):266, (35):270, (36):274, (37):278, (38):282, (39):286, (40):290, (41):294, (42):298, (43):302, (44):306, (45):310, (46):314, (47):318, (48):322, (49):326, (50):330, (51):334, (52):338, (53):342, (54):346, (55):350, (56):354, (57):358, (58):362, (59):366, (60):370, (61):374, (62):378, (63):382, (64):386, (65):390, (66):394, (67):398, (68):402, (69):406, (70):410, (71):414, (72):418, (73):422, (74):426, (75):430, (76):434, (77):438, (78):442, (79):446, (80):450, (81):454, (82):458, (ETB0):800, (172):814 debugfs: extents subdir : : 1/ 1 82/ 83 81 - 81 454 - 454 1 1/ 1 83/ 83 82 - 82 458 - 458 1 0/ 1 2/ 2 170 - 4294967410 800 4294967241 1/ 1 1/ 1 172 - 172 814 - 814 1 The i_size is correct for 115 data blocks written, and i_blocks would be correct if the second index block wouldn't have been lost. It seems the bug is in the extent handling code, but I haven't yet dug into why the last extent is kept. I tried deleting it like the other blocks, but the iteration immediately stops with an error that the index block is corrupted, and I'm not sure how to catch it the second time. Cheers, Andreas -- Andreas Dilger Lustre Principal Engineer Intel High Performance Data Division
Attachment:
script
Description: script