Fix ext2fs_set_gdt_csum() to use access functions.
PATCH 1/1 Fix ext2fs_set_gdt_csum() to use access functions.
Following Andreas's suggestion, I found that ext2fs_set_gdt_csum() was the culprit: it used an ext2_group_desc pointer to access/set fields and it did that directly, not through access functions.
I'm testing the patch now, but it'll take a while, so again I'm sending it out for comments and to possibly try out. Even if it checks out, it will need some additional care once the flags situation settles down.
Thanks,
Nick
Quoted 1 line
From
6a3b83cda1bcd1e3594515ee888f175bf5cc7906Mon Sep 17 00:00:00 2001
From: Nick Dokos <[email protected]>
Date: Wed, 2 Sep 2009 23:00:23 -0400
Subject: [PATCH] Fix ext2fs_set_gdt_csum() to use access functions.
Replace all field accesses with calls to access functions. Most importantly, get rid of the mis-declared group descriptor pointer which caused the wrong fields to be updated.
--- lib/ext2fs/csum.c | 30 +++++++++++++++++------------- 1 files changed, 17 insertions(+), 13 deletions(-)
lib/ext2fs/csum.c +17 -13
@@ -105,7 +105,6 @@ static __u32 find_last_inode_ingrp(ext2fs_inode_bitmap bitmap,errcode_t ext2fs_set_gdt_csum (ext2_filsys fs) {struct ext2_super_block *sb = fs->super; -struct ext2_group_desc *bg = fs->group_desc;int dirty =0 ;dgrp_t i;
@@ -116,27 +115,32 @@ errcode_t ext2fs_set_gdt_csum(ext2_filsys fs) EXT4_FEATURE_RO_COMPAT_GDT_CSUM))return 0 ; -for (i =0 ; i < fs->group_desc_count; i++, bg++) { -int old_csum = bg->bg_checksum; -int old_unused = bg->bg_itable_unused; -int old_flags = bg->bg_flags; +for (i =0 ; i < fs->group_desc_count; i++) { +unsigned int old_csum =ext2fs_bg_checksum (fs, i); +int old_unused =ext2fs_bg_itable_unused (fs, i); +unsigned int old_flags =ext2fs_bg_flags (fs, i); +int old_free_inodes_count =ext2fs_bg_free_inodes_count (fs, i); + -if (bg->bg_free_inodes_count == sb->s_inodes_per_group) { - bg->bg_flags |=EXT2_BG_INODE_UNINIT ; - bg->bg_itable_unused = sb->s_inodes_per_group; +if (old_free_inodes_count == sb->s_inodes_per_group) { +ext2fs_bg_flags_set (fs, i, old_flags |EXT2_BG_INODE_UNINIT ); +ext2fs_bg_itable_unused_set (fs, i, sb->s_inodes_per_group); }else { - bg->bg_flags &= ~EXT2_BG_INODE_UNINIT ; - bg->bg_itable_unused = sb->s_inodes_per_group - +int unused = sb->s_inodes_per_group -find_last_inode_ingrp (fs->inode_map, sb->s_inodes_per_group,i); + +ext2fs_bg_flags_set (fs, i, old_flags & ~EXT2_BG_INODE_UNINIT ); + +ext2fs_bg_itable_unused_set (fs, i, unused); }ext2fs_group_desc_csum_set (fs, i); -if (old_flags != bg->bg_flags) +if (old_flags !=ext2fs_bg_flags (fs, i)) dirty =1 ; -if (old_unused != bg->bg_itable_unused) +if (old_unused !=ext2fs_bg_itable_unused (fs, i)) dirty =1 ; -if (old_csum != bg->bg_checksum) +if (old_csum !=ext2fs_bg_checksum (fs, i)) dirty =1 ; }if (dirty)
Signature · 1 line
1.6.0.6
Quoted 7 lines · Nick Dokos
Following Andreas's suggestion, I found that
ext2fs_set_gdt_csum()was the culprit: it used anext2_group_descpointer to access/set fields and it did that directly, not through access functions.I'm testing the patch now, but it'll take a while, so again I'm sending it out for comments and to possibly try out. Even if it checks out, it will need some additional care once the flags situation settles down.
cool, thanks.
yep this would explain the bug (any direct access would I guess)
With this on top of my patch stack, my testcase passes, until I allocate something high enough that I run into yet another 32-bit overflow somewhere ;)
Ted, want me to collate these into a proper patch series? I have one more direct access fix as well. (and then things like ext2ed & e2imgge etc all need a thorough going-over too....)
We really must make these structs opaque or something or it will be endless pain...
Minor comments below.
Quoted 11 lines
Thanks,
NickFrom
6a3b83cda1bcd1e3594515ee888f175bf5cc7906Mon Sep 17 00:00:00 2001
From: Nick Dokos <[email protected]>
Date: Wed, 2 Sep 2009 23:00:23 -0400
Subject: [PATCH] Fixext2fs_set_gdt_csum()to use access functions.Replace all field accesses with calls to access functions. Most importantly, get rid of the mis-declared group descriptor pointer which caused the wrong fields to be updated.
Not quite sure what you mean by this? It worked ok for the "old" size ...
Quoted 18 lines
Signed-off-by Nick Dokos <[email protected]>--- lib/ext2fs/csum.c | 30 +++++++++++++++++------------- 1 files changed, 17 insertions(+), 13 deletions(-)lib/ext2fs/csum.c +0
-1@@ -105,7 +105,6 @@ static __u32 find_last_inode_ingrp(ext2fs_inode_bitmap bitmap,errcode_t ext2fs_set_gdt_csum (ext2_filsys fs) {struct ext2_super_block *sb = fs->super; -struct ext2_group_desc *bg = fs->group_desc;int dirty =0 ;dgrp_t i;
lib/ext2fs/csum.c errcode_t ext2fs_set_gdt_csum(ext2_filsys fs) · line 120
Quoted 10 lines
lib/ext2fs/csum.c +3
-4@@ -116,27 +115,32 @@ errcode_t ext2fs_set_gdt_csum(ext2_filsys fs) EXT4_FEATURE_RO_COMPAT_GDT_CSUM))return 0 ; -for (i =0 ; i < fs->group_desc_count; i++, bg++) { -int old_csum = bg->bg_checksum; -int old_unused = bg->bg_itable_unused; -int old_flags = bg->bg_flags; +for (i =0 ; i < fs->group_desc_count; i++) { +unsigned int old_csum =ext2fs_bg_checksum (fs, i); +int old_unused =ext2fs_bg_itable_unused (fs, i);
whitespace problem here, use tabs
lib/ext2fs/csum.c errcode_t ext2fs_set_gdt_csum(ext2_filsys fs) · line 122
Quoted 2 lines
lib/ext2fs/csum.c +2
-0+unsigned int old_flags =ext2fs_bg_flags (fs, i); +int old_free_inodes_count =ext2fs_bg_free_inodes_count (fs, i);
ditto
also, strictly, csum & flags are 16 bits, unused is (today) 64 bits ....
lib/ext2fs/csum.c errcode_t ext2fs_set_gdt_csum(ext2_filsys fs) · line 123
Quoted 1 line
lib/ext2fs/csum.c +1
-0+
extra blank line?
lib/ext2fs/csum.c errcode_t ext2fs_set_gdt_csum(ext2_filsys fs) · line 126
Quoted 6 lines
lib/ext2fs/csum.c +2
-3-if (bg->bg_free_inodes_count == sb->s_inodes_per_group) { - bg->bg_flags |=EXT2_BG_INODE_UNINIT ; - bg->bg_itable_unused = sb->s_inodes_per_group; +if (old_free_inodes_count == sb->s_inodes_per_group) { +ext2fs_bg_flags_set (fs, i, old_flags |EXT2_BG_INODE_UNINIT );
better this I think (though I guess it may change ...):
+ ext2fs_bg_flag_set(fs, i, EXT2_BG_INODE_UNINIT);
lib/ext2fs/csum.c errcode_t ext2fs_set_gdt_csum(ext2_filsys fs) · line 133
Quoted 9 lines
lib/ext2fs/csum.c +4
-2+ext2fs_bg_itable_unused_set (fs, i, sb->s_inodes_per_group); }else { - bg->bg_flags &= ~EXT2_BG_INODE_UNINIT ; - bg->bg_itable_unused = sb->s_inodes_per_group - +int unused = sb->s_inodes_per_group -find_last_inode_ingrp (fs->inode_map, sb->s_inodes_per_group,i); + +ext2fs_bg_flags_set (fs, i, old_flags & ~EXT2_BG_INODE_UNINIT );
and this:
+ ext2fs_flag_clear(fs, i, EXT2_BG_INODE_UNINIT);
Quoted 16 lines
lib/ext2fs/csum.c +5
-3+ +ext2fs_bg_itable_unused_set (fs, i, unused); }ext2fs_group_desc_csum_set (fs, i); -if (old_flags != bg->bg_flags) +if (old_flags !=ext2fs_bg_flags (fs, i)) dirty =1 ; -if (old_unused != bg->bg_itable_unused) +if (old_unused !=ext2fs_bg_itable_unused (fs, i)) dirty =1 ; -if (old_csum != bg->bg_checksum) +if (old_csum !=ext2fs_bg_checksum (fs, i)) dirty =1 ; }if (dirty)
Quoted 7 lines · Eric Sandeen
Quoted 4 lines
Replace all field accesses with calls to access functions. Most importantly, get rid of the mis-declared group descriptor pointer which caused the wrong fields to be updated.
Not quite sure what you mean by this? It worked ok for the "old" size ...
Yes, it worked fine for the old ext2_group_desc structure, but it has no hope of working with the ext4_group_desc: the sizes are different, so doing bg++ gets it pointing somewhere within the first descriptor, not to the beginning of the second. That's how it ended up modifying reserved fields.
Sorry about the whitespace problems. As for the flags, I didn't worry too much about how to set them: I figured that once the dust settles, we'll follow the proper convention.
Thanks for looking it over and testing too! We'll see whether Justin's problem disappears.
Nick
Quoted 5 lines · Nick Dokos
Quoted 3 lines · Eric Sandeen
Quoted 3 lines
Replace all field accesses with calls to access functions. Most importantly, get rid of the mis-declared group descriptor pointer which caused the wrong fields to be updated.
Quoted 8 lines
Quoted 2 lines
Not quite sure what you mean by this? It worked ok for the "old" size ...
Yes, it worked fine for the old
ext2_group_descstructure, but it has no hope of working with theext4_group_desc: the sizes are different, so doing bg++ gets it pointing somewhere within the first descriptor, not to the beginning of the second. That's how it ended up modifying reserved fields.
right. Ok, I was just confused when you said "mis-declared" - thought you meant something was wrong with how bg was set or initialized.
Quoted 3 lines
Sorry about the whitespace problems. As for the flags, I didn't worry too much about how to set them: I figured that once the dust settles, we'll follow the proper convention.
sounds good.
Thanks,
-Eric
Quoted 4 lines
Thanks for looking it over and testing too! We'll see whether Justin's problem disappears.
Nick
Quoted 3 lines · Nick Dokos
Thanks for looking it over and testing too! We'll see whether Justin's problem disappears.
Yes, it did! Thanks everyone for taking the time to look into this and get out a fix!
-Justin
lib/ext2fs/csum.c static __u32 find_last_inode_ingrp(ext2fs_inode_bitmap bitmap, · line 108
Quoted 8 lines · Nick Dokos
Replace all field accesses with calls to access functions. Most importantly, get rid of the mis-declared group descriptor pointer which caused the wrong fields to be updated.
errcode_text2fs_set_gdt_csum(ext2_filsys fs)
{lib/ext2fs/csum.c +0
-1struct ext2_super_block *sb = fs->super; -struct ext2_group_desc *bg = fs->group_desc;
Given the danger of ongoing direct access to fs->group_desc (including potentially from external applications) I think the only safe way of doing this is to have an opaque fs->group_desc structure that cannot be dereferenced outside of the library.
As a potential compatibility measure, we might consider fs->group_desc to be valid for 32-bit filesystems, and leave it NULL for 64-bit filesystems and use a second (opaque) fs->group_desc64 pointer for access to filesystems with INCOMPAT_64BIT set.
Cheers, Andreas
--
Andreas Dilger
Sr. Staff Engineer, Lustre Group
Sun Microsystems of Canada, Inc.