| Message ID | 20180612095328.5215-7-jack@suse.cz |
|---|---|
| State | Accepted, archived |
| Headers | show |
| Series | e2fsprogs: Handle s_inodes_count overflow better | expand |
On Jun 12, 2018, at 3:53 AM, Jan Kara <jack@suse.cz> wrote: > > When s_inodes_count would overflow given number of groups and inodes per > group, we cannot currently fix the breakage in e2fsck as that requires > trimming number of groups or inodes per group which both means data & > inode migration etc. Just trimming sb->s_inodes_count is not enough as > kernel's inode allocation code is not able to handle filesystems where > not all inodes in the last group are usable. So don't pretend we can fix > s_inodes_count overflow by just trimming the s_inodes_count value. > > When s_inodes_count is just wrong but will not overflow, let's fix it. > Also move this check before we use s_inodes_count for checking > s_first_ino. > > Signed-off-by: Jan Kara <jack@suse.cz> I guess that still leaves an open issue of what a user can do if the superblock (correctly or incorrectly) has groups * inodes_per_group that overflows 2^32, since it seems e2fsck would abort in that case. I'd think that one option would be to reduce the group count until the number of inodes < 2^32, which might lose some files at the end of the filesystem, but more likely would not lose anything since the kernel would have written those inodes into the wrong place anyway. The existing code takes the approach of truncating the last group, which at least makes the filesystem usable again, though it risks corrupting inode #1 (EXT2_BAD_INO) on older kernels should the last inode be allocated. That probably isn't a huge danger in the end. Cheers, Andreas > --- > e2fsck/problem.c | 5 +++++ > e2fsck/problem.h | 2 ++ > e2fsck/super.c | 27 ++++++++++++++++----------- > 3 files changed, 23 insertions(+), 11 deletions(-) > > diff --git a/e2fsck/problem.c b/e2fsck/problem.c > index edc9d51fcf31..8de558edc5dd 100644 > --- a/e2fsck/problem.c > +++ b/e2fsck/problem.c > @@ -184,6 +184,11 @@ static struct e2fsck_problem problem_table[] = { > N_("@i count in @S is %i, @s %j.\n"), > PROMPT_FIX, 0 }, > > + /* Too many inodes in the filesystem */ > + { PR_0_INODE_COUNT_BIG, > + N_("@S would have too many inodes (%N).\n"), > + PROMPT_NONE, PR_AFTER_CODE, PR_0_SB_CORRUPT }, > + > { PR_0_HURD_CLEAR_FILETYPE, > N_("The Hurd does not support the filetype feature.\n"), > PROMPT_CLEAR, 0 }, > diff --git a/e2fsck/problem.h b/e2fsck/problem.h > index 482d111a0f2c..b6bca668bb1b 100644 > --- a/e2fsck/problem.h > +++ b/e2fsck/problem.h > @@ -282,6 +282,8 @@ struct problem_context { > /* Invalid quota inode number */ > #define PR_0_INVALID_QUOTA_INO 0x00004F > > +/* Inode count in the superblock incorrect */ > +#define PR_0_INODE_COUNT_BIG 0x000050 > > /* > * Pass 1 errors > diff --git a/e2fsck/super.c b/e2fsck/super.c > index 9c0e0e7b35b4..60a19a7e2767 100644 > --- a/e2fsck/super.c > +++ b/e2fsck/super.c > @@ -646,6 +646,22 @@ void check_super_block(e2fsck_t ctx) > check_super_value(ctx, "desc_size", > sb->s_desc_size, MAX_CHECK | LOG2_CHECK, 0, > EXT2_MAX_DESC_SIZE); > + > + should_be = (__u64)sb->s_inodes_per_group * fs->group_desc_count; > + if (should_be > ~0U) { > + pctx.num = should_be; > + fix_problem(ctx, PR_0_INODE_COUNT_BIG, &pctx); > + ctx->flags |= E2F_FLAG_ABORT; > + return; > + } > + if (sb->s_inodes_count != should_be) { > + pctx.ino = sb->s_inodes_count; > + pctx.ino2 = should_be; > + if (fix_problem(ctx, PR_0_INODE_COUNT_WRONG, &pctx)) { > + sb->s_inodes_count = should_be; > + ext2fs_mark_super_dirty(fs); > + } > + } > if (sb->s_rev_level > EXT2_GOOD_OLD_REV) > check_super_value(ctx, "first_ino", sb->s_first_ino, > MIN_CHECK | MAX_CHECK, > @@ -683,17 +699,6 @@ void check_super_block(e2fsck_t ctx) > return; > } > > - should_be = (blk64_t)sb->s_inodes_per_group * fs->group_desc_count; > - if (should_be > UINT_MAX) > - should_be = UINT_MAX; > - if (sb->s_inodes_count != should_be) { > - pctx.ino = sb->s_inodes_count; > - pctx.ino2 = should_be; > - if (fix_problem(ctx, PR_0_INODE_COUNT_WRONG, &pctx)) { > - sb->s_inodes_count = should_be; > - ext2fs_mark_super_dirty(fs); > - } > - } > if (EXT2_INODE_SIZE(sb) > EXT2_GOOD_OLD_INODE_SIZE) { > unsigned min = > sizeof(((struct ext2_inode_large *) 0)->i_extra_isize) + > -- > 2.13.7 > Cheers, Andreas
On Tue 12-06-18 13:04:00, Andreas Dilger wrote: > On Jun 12, 2018, at 3:53 AM, Jan Kara <jack@suse.cz> wrote: > > > > When s_inodes_count would overflow given number of groups and inodes per > > group, we cannot currently fix the breakage in e2fsck as that requires > > trimming number of groups or inodes per group which both means data & > > inode migration etc. Just trimming sb->s_inodes_count is not enough as > > kernel's inode allocation code is not able to handle filesystems where > > not all inodes in the last group are usable. So don't pretend we can fix > > s_inodes_count overflow by just trimming the s_inodes_count value. > > > > When s_inodes_count is just wrong but will not overflow, let's fix it. > > Also move this check before we use s_inodes_count for checking > > s_first_ino. > > > > Signed-off-by: Jan Kara <jack@suse.cz> > > I guess that still leaves an open issue of what a user can do if the > superblock (correctly or incorrectly) has groups * inodes_per_group > that overflows 2^32, since it seems e2fsck would abort in that case. Yes, e2fsck would abort as in the case of other non-trivial superblock corruptions. My thinking was that in that case user can use debugfs (hence my following debugfs changes) to fix the filesystem. > I'd think that one option would be to reduce the group count until > the number of inodes < 2^32, which might lose some files at the end > of the filesystem, but more likely would not lose anything since the > kernel would have written those inodes into the wrong place anyway. > > The existing code takes the approach of truncating the last group, > which at least makes the filesystem usable again, though it risks > corrupting inode #1 (EXT2_BAD_INO) on older kernels should the last > inode be allocated. That probably isn't a huge danger in the end. Actually currently e2fsck never succeeded in even opening the filesystem with corrupted s_inodes_count so it could not fix anything. So my change is at least an improvement that it reports what's wrong and fixes the "easily fixable" corruptions. Reducing number of block groups is doable (although libext2fs code is not quite prepared for that so it may need some fixups). I was just a bit worried that if the s_inodes_count corruption is not due to resize2fs bug, reducing the group count may actually be a wrong thing to do and could cause serious damage to the filesystem. What do people think? Honza
On Wed, Jun 13, 2018 at 02:11:05PM +0200, Jan Kara wrote: > > Actually currently e2fsck never succeeded in even opening the filesystem > with corrupted s_inodes_count so it could not fix anything. So my change is > at least an improvement that it reports what's wrong and fixes the "easily > fixable" corruptions. > > Reducing number of block groups is doable (although libext2fs code is not > quite prepared for that so it may need some fixups). I was just a bit > worried that if the s_inodes_count corruption is not due to resize2fs bug, > reducing the group count may actually be a wrong thing to do and could > cause serious damage to the filesystem. What do people think? The primary principle should be "first, do no harm". It's probably going to be easier to get e2undo with resize2fs working well and by default than trying to figure out how to fix the s_inodes_count corruption case. And it will have a better bang for the buck, since future bugs may not involve s_inodes_count at all. So that's probably a better investment of developer time, in my view. - Ted
On Tue, Jun 12, 2018 at 11:53:24AM +0200, Jan Kara wrote: > When s_inodes_count would overflow given number of groups and inodes per > group, we cannot currently fix the breakage in e2fsck as that requires > trimming number of groups or inodes per group which both means data & > inode migration etc. Just trimming sb->s_inodes_count is not enough as > kernel's inode allocation code is not able to handle filesystems where > not all inodes in the last group are usable. So don't pretend we can fix > s_inodes_count overflow by just trimming the s_inodes_count value. > > When s_inodes_count is just wrong but will not overflow, let's fix it. > Also move this check before we use s_inodes_count for checking > s_first_ino. > > Signed-off-by: Jan Kara <jack@suse.cz> Applied with the following fix-up patch. (The entries in the problem array must be in problem_code sort order.) - Ted diff --git a/e2fsck/problem.c b/e2fsck/problem.c index 8de558edc..efe98c920 100644 --- a/e2fsck/problem.c +++ b/e2fsck/problem.c @@ -184,11 +184,6 @@ static struct e2fsck_problem problem_table[] = { N_("@i count in @S is %i, @s %j.\n"), PROMPT_FIX, 0 }, - /* Too many inodes in the filesystem */ - { PR_0_INODE_COUNT_BIG, - N_("@S would have too many inodes (%N).\n"), - PROMPT_NONE, PR_AFTER_CODE, PR_0_SB_CORRUPT }, - { PR_0_HURD_CLEAR_FILETYPE, N_("The Hurd does not support the filetype feature.\n"), PROMPT_CLEAR, 0 }, @@ -498,6 +493,11 @@ static struct e2fsck_problem problem_table[] = { N_("Invalid %U @q @i %i. "), PROMPT_FIX, 0 }, + /* Too many inodes in the filesystem */ + { PR_0_INODE_COUNT_BIG, + N_("@S would have too many inodes (%N).\n"), + PROMPT_NONE, PR_AFTER_CODE, PR_0_SB_CORRUPT }, + /* Pass 1 errors */ /* Pass 1: Checking inodes, blocks, and sizes */
On Tue 19-06-18 11:25:35, Theodore Y. Ts'o wrote: > On Tue, Jun 12, 2018 at 11:53:24AM +0200, Jan Kara wrote: > > When s_inodes_count would overflow given number of groups and inodes per > > group, we cannot currently fix the breakage in e2fsck as that requires > > trimming number of groups or inodes per group which both means data & > > inode migration etc. Just trimming sb->s_inodes_count is not enough as > > kernel's inode allocation code is not able to handle filesystems where > > not all inodes in the last group are usable. So don't pretend we can fix > > s_inodes_count overflow by just trimming the s_inodes_count value. > > > > When s_inodes_count is just wrong but will not overflow, let's fix it. > > Also move this check before we use s_inodes_count for checking > > s_first_ino. > > > > Signed-off-by: Jan Kara <jack@suse.cz> > > Applied with the following fix-up patch. (The entries in the problem > array must be in problem_code sort order.) Thanks for fixing that up! Honza > diff --git a/e2fsck/problem.c b/e2fsck/problem.c > index 8de558edc..efe98c920 100644 > --- a/e2fsck/problem.c > +++ b/e2fsck/problem.c > @@ -184,11 +184,6 @@ static struct e2fsck_problem problem_table[] = { > N_("@i count in @S is %i, @s %j.\n"), > PROMPT_FIX, 0 }, > > - /* Too many inodes in the filesystem */ > - { PR_0_INODE_COUNT_BIG, > - N_("@S would have too many inodes (%N).\n"), > - PROMPT_NONE, PR_AFTER_CODE, PR_0_SB_CORRUPT }, > - > { PR_0_HURD_CLEAR_FILETYPE, > N_("The Hurd does not support the filetype feature.\n"), > PROMPT_CLEAR, 0 }, > @@ -498,6 +493,11 @@ static struct e2fsck_problem problem_table[] = { > N_("Invalid %U @q @i %i. "), > PROMPT_FIX, 0 }, > > + /* Too many inodes in the filesystem */ > + { PR_0_INODE_COUNT_BIG, > + N_("@S would have too many inodes (%N).\n"), > + PROMPT_NONE, PR_AFTER_CODE, PR_0_SB_CORRUPT }, > + > /* Pass 1 errors */ > > /* Pass 1: Checking inodes, blocks, and sizes */
diff --git a/e2fsck/problem.c b/e2fsck/problem.c index edc9d51fcf31..8de558edc5dd 100644 --- a/e2fsck/problem.c +++ b/e2fsck/problem.c @@ -184,6 +184,11 @@ static struct e2fsck_problem problem_table[] = { N_("@i count in @S is %i, @s %j.\n"), PROMPT_FIX, 0 }, + /* Too many inodes in the filesystem */ + { PR_0_INODE_COUNT_BIG, + N_("@S would have too many inodes (%N).\n"), + PROMPT_NONE, PR_AFTER_CODE, PR_0_SB_CORRUPT }, + { PR_0_HURD_CLEAR_FILETYPE, N_("The Hurd does not support the filetype feature.\n"), PROMPT_CLEAR, 0 }, diff --git a/e2fsck/problem.h b/e2fsck/problem.h index 482d111a0f2c..b6bca668bb1b 100644 --- a/e2fsck/problem.h +++ b/e2fsck/problem.h @@ -282,6 +282,8 @@ struct problem_context { /* Invalid quota inode number */ #define PR_0_INVALID_QUOTA_INO 0x00004F +/* Inode count in the superblock incorrect */ +#define PR_0_INODE_COUNT_BIG 0x000050 /* * Pass 1 errors diff --git a/e2fsck/super.c b/e2fsck/super.c index 9c0e0e7b35b4..60a19a7e2767 100644 --- a/e2fsck/super.c +++ b/e2fsck/super.c @@ -646,6 +646,22 @@ void check_super_block(e2fsck_t ctx) check_super_value(ctx, "desc_size", sb->s_desc_size, MAX_CHECK | LOG2_CHECK, 0, EXT2_MAX_DESC_SIZE); + + should_be = (__u64)sb->s_inodes_per_group * fs->group_desc_count; + if (should_be > ~0U) { + pctx.num = should_be; + fix_problem(ctx, PR_0_INODE_COUNT_BIG, &pctx); + ctx->flags |= E2F_FLAG_ABORT; + return; + } + if (sb->s_inodes_count != should_be) { + pctx.ino = sb->s_inodes_count; + pctx.ino2 = should_be; + if (fix_problem(ctx, PR_0_INODE_COUNT_WRONG, &pctx)) { + sb->s_inodes_count = should_be; + ext2fs_mark_super_dirty(fs); + } + } if (sb->s_rev_level > EXT2_GOOD_OLD_REV) check_super_value(ctx, "first_ino", sb->s_first_ino, MIN_CHECK | MAX_CHECK, @@ -683,17 +699,6 @@ void check_super_block(e2fsck_t ctx) return; } - should_be = (blk64_t)sb->s_inodes_per_group * fs->group_desc_count; - if (should_be > UINT_MAX) - should_be = UINT_MAX; - if (sb->s_inodes_count != should_be) { - pctx.ino = sb->s_inodes_count; - pctx.ino2 = should_be; - if (fix_problem(ctx, PR_0_INODE_COUNT_WRONG, &pctx)) { - sb->s_inodes_count = should_be; - ext2fs_mark_super_dirty(fs); - } - } if (EXT2_INODE_SIZE(sb) > EXT2_GOOD_OLD_INODE_SIZE) { unsigned min = sizeof(((struct ext2_inode_large *) 0)->i_extra_isize) +
When s_inodes_count would overflow given number of groups and inodes per group, we cannot currently fix the breakage in e2fsck as that requires trimming number of groups or inodes per group which both means data & inode migration etc. Just trimming sb->s_inodes_count is not enough as kernel's inode allocation code is not able to handle filesystems where not all inodes in the last group are usable. So don't pretend we can fix s_inodes_count overflow by just trimming the s_inodes_count value. When s_inodes_count is just wrong but will not overflow, let's fix it. Also move this check before we use s_inodes_count for checking s_first_ino. Signed-off-by: Jan Kara <jack@suse.cz> --- e2fsck/problem.c | 5 +++++ e2fsck/problem.h | 2 ++ e2fsck/super.c | 27 ++++++++++++++++----------- 3 files changed, 23 insertions(+), 11 deletions(-)