When e2fsck recreates a missing lost+found directory, it uses indirect
block mapping. It does this even if the file system being repaired is
extent mapped and where the original lost+found was extent mapped when
created by mkfs.
This inconsistency exists because the kernel code handling indirect
block mapping was considered more trustworthy in ext4's early days;
today, the extent handling code is robust.
This inconsistency also causes potential behavioral problems on
bigalloc file systems. The ext4 kernel code does not support indirect
block mapped directories when bigalloc is enabled. For example, it is
not possible to extend lost+found on a mounted bigalloc file system
with the current implementation, so the behavior of that directory
differs from all others, potentially surprising users.
To make the handling of lost+found more consistent, and to avoid
behavioral problems on bigalloc file systems, recreate lost+found
with extent mapping if the file system under repair is extent mapped.
Signed-off-by: Eric Whitney <[email protected]>
---
e2fsck/pass3.c | 29 +++++++++++++++++++++++++++--
e2fsck/problem.c | 10 ++++++++++
e2fsck/problem.h | 6 ++++++
3 files changed, 43 insertions(+), 2 deletions(-)
diff --git a/e2fsck/pass3.c b/e2fsck/pass3.c
index 44203ca..259c978 100644
--- a/e2fsck/pass3.c
+++ b/e2fsck/pass3.c
@@ -378,7 +378,7 @@ static int check_directory(e2fsck_t ctx, ext2_ino_t dir,
ext2_ino_t e2fsck_get_lost_and_found(e2fsck_t ctx, int fix)
{
ext2_filsys fs = ctx->fs;
- ext2_ino_t ino;
+ ext2_ino_t ino;
blk64_t blk;
errcode_t retval;
struct ext2_inode_large inode;
@@ -386,6 +386,7 @@ ext2_ino_t e2fsck_get_lost_and_found(e2fsck_t ctx, int fix)
static const char name[] = "lost+found";
struct problem_context pctx;
int will_rehash, flags;
+ ext2_extent_handle_t handle;
if (ctx->lost_and_found)
return ctx->lost_and_found;
@@ -520,7 +521,10 @@ skip_new_block:
inode.i_atime = inode.i_ctime = inode.i_mtime = ctx->now;
inode.i_links_count = 2;
ext2fs_iblk_set(fs, EXT2_INODE(&inode), 1);
- inode.i_block[0] = blk;
+ if (ext2fs_has_feature_extents(fs->super))
+ inode.i_flags |= EXT4_EXTENTS_FL;
+ else
+ inode.i_block[0] = blk;
/*
* Next, write out the inode.
@@ -553,6 +557,27 @@ skip_new_block:
}
/*
+ * If the new lost+found is extent mapped, map its first new
+ * directory block
+ */
+ if (ext2fs_has_feature_extents(fs->super)) {
+ retval = ext2fs_extent_open2(fs, ino, EXT2_INODE(&inode),
+ &handle);
+ if (retval) {
+ pctx.errcode = retval;
+ fix_problem(ctx, PR_3_ERR_LPF_OPEN_EXTENT, &pctx);
+ return 0;
+ }
+ retval = ext2fs_extent_set_bmap(handle, 0, blk, 0);
+ ext2fs_extent_free(handle);
+ if (retval) {
+ pctx.errcode = retval;
+ fix_problem(ctx, PR_3_ERR_LPF_SET_BMAP, &pctx);
+ return 0;
+ }
+ }
+
+ /*
* Finally, create the directory link
*/
pctx.errcode = ext2fs_link(fs, EXT2_ROOT_INO, name, ino, EXT2_FT_DIR);
diff --git a/e2fsck/problem.c b/e2fsck/problem.c
index c3e4f6a..736b822 100644
--- a/e2fsck/problem.c
+++ b/e2fsck/problem.c
@@ -1791,6 +1791,16 @@ static struct e2fsck_problem problem_table[] = {
N_("/@l is encrypted\n"),
PROMPT_CLEAR, 0 },
+ /* Error while creating an extent for /lost+found */
+ { PR_3_ERR_LPF_OPEN_EXTENT,
+ N_("ext2fs_@x_open2: %m while creating an @x for /@l\n"),
+ PROMPT_NONE, 0 },
+
+ /* Error while mapping a directory block for /lost+found */
+ { PR_3_ERR_LPF_SET_BMAP,
+ N_("ext2fs_@x_set_bmap: %m while mapping a @d @b for /@l\n"),
+ PROMPT_NONE, 0 },
+
/* Pass 3A Directory Optimization */
/* Pass 3A: Optimizing directories */
diff --git a/e2fsck/problem.h b/e2fsck/problem.h
index d291e26..dfdb6d0 100644
--- a/e2fsck/problem.h
+++ b/e2fsck/problem.h
@@ -1078,6 +1078,12 @@ struct problem_context {
/* Lost+found is encrypted */
#define PR_3_LPF_ENCRYPTED 0x03001B
+/* Error while creating an extent for /lost+found */
+#define PR_3_ERR_LPF_OPEN_EXTENT 0x03001C
+
+/* Error while mapping a directory block for /lost+found */
+#define PR_3_ERR_LPF_SET_BMAP 0x03001D
+
/*
* Pass 3a --- rehashing diretories
*/
--
2.1.4
On Thu, Mar 16, 2017 at 11:39:48AM -0400, Eric Whitney wrote:
> When e2fsck recreates a missing lost+found directory, it uses indirect
> block mapping. It does this even if the file system being repaired is
> extent mapped and where the original lost+found was extent mapped when
> created by mkfs.
>
> This inconsistency exists because the kernel code handling indirect
> block mapping was considered more trustworthy in ext4's early days;
> today, the extent handling code is robust.
>
> This inconsistency also causes potential behavioral problems on
> bigalloc file systems. The ext4 kernel code does not support indirect
> block mapped directories when bigalloc is enabled. For example, it is
> not possible to extend lost+found on a mounted bigalloc file system
> with the current implementation, so the behavior of that directory
> differs from all others, potentially surprising users.
>
> To make the handling of lost+found more consistent, and to avoid
> behavioral problems on bigalloc file systems, recreate lost+found
> with extent mapping if the file system under repair is extent mapped.
>
> Signed-off-by: Eric Whitney <[email protected]>
> ---
> e2fsck/pass3.c | 29 +++++++++++++++++++++++++++--
> e2fsck/problem.c | 10 ++++++++++
> e2fsck/problem.h | 6 ++++++
> 3 files changed, 43 insertions(+), 2 deletions(-)
>
> diff --git a/e2fsck/pass3.c b/e2fsck/pass3.c
> index 44203ca..259c978 100644
> --- a/e2fsck/pass3.c
> +++ b/e2fsck/pass3.c
> @@ -378,7 +378,7 @@ static int check_directory(e2fsck_t ctx, ext2_ino_t dir,
> ext2_ino_t e2fsck_get_lost_and_found(e2fsck_t ctx, int fix)
> {
> ext2_filsys fs = ctx->fs;
> - ext2_ino_t ino;
> + ext2_ino_t ino;
> blk64_t blk;
> errcode_t retval;
> struct ext2_inode_large inode;
> @@ -386,6 +386,7 @@ ext2_ino_t e2fsck_get_lost_and_found(e2fsck_t ctx, int fix)
> static const char name[] = "lost+found";
> struct problem_context pctx;
> int will_rehash, flags;
> + ext2_extent_handle_t handle;
>
> if (ctx->lost_and_found)
> return ctx->lost_and_found;
> @@ -520,7 +521,10 @@ skip_new_block:
> inode.i_atime = inode.i_ctime = inode.i_mtime = ctx->now;
> inode.i_links_count = 2;
> ext2fs_iblk_set(fs, EXT2_INODE(&inode), 1);
> - inode.i_block[0] = blk;
> + if (ext2fs_has_feature_extents(fs->super))
> + inode.i_flags |= EXT4_EXTENTS_FL;
> + else
> + inode.i_block[0] = blk;
>
> /*
> * Next, write out the inode.
> @@ -553,6 +557,27 @@ skip_new_block:
> }
>
> /*
> + * If the new lost+found is extent mapped, map its first new
> + * directory block
> + */
> + if (ext2fs_has_feature_extents(fs->super)) {
> + retval = ext2fs_extent_open2(fs, ino, EXT2_INODE(&inode),
> + &handle);
> + if (retval) {
> + pctx.errcode = retval;
> + fix_problem(ctx, PR_3_ERR_LPF_OPEN_EXTENT, &pctx);
> + return 0;
> + }
> + retval = ext2fs_extent_set_bmap(handle, 0, blk, 0);
> + ext2fs_extent_free(handle);
> + if (retval) {
> + pctx.errcode = retval;
> + fix_problem(ctx, PR_3_ERR_LPF_SET_BMAP, &pctx);
> + return 0;
> + }
> + }
You could refactor most of this function since ext2fs_bmap2() can do the
work for you -- allocating a block, setting the extents flag, and
initializing the extent tree header.
Maybe there's a specific reason to allocate the block and then the
inode instead of simply using bmap? I can't find any obvious reason,
but Ted might remember something.
--D
> +
> + /*
> * Finally, create the directory link
> */
> pctx.errcode = ext2fs_link(fs, EXT2_ROOT_INO, name, ino, EXT2_FT_DIR);
> diff --git a/e2fsck/problem.c b/e2fsck/problem.c
> index c3e4f6a..736b822 100644
> --- a/e2fsck/problem.c
> +++ b/e2fsck/problem.c
> @@ -1791,6 +1791,16 @@ static struct e2fsck_problem problem_table[] = {
> N_("/@l is encrypted\n"),
> PROMPT_CLEAR, 0 },
>
> + /* Error while creating an extent for /lost+found */
> + { PR_3_ERR_LPF_OPEN_EXTENT,
> + N_("ext2fs_@x_open2: %m while creating an @x for /@l\n"),
> + PROMPT_NONE, 0 },
> +
> + /* Error while mapping a directory block for /lost+found */
> + { PR_3_ERR_LPF_SET_BMAP,
> + N_("ext2fs_@x_set_bmap: %m while mapping a @d @b for /@l\n"),
> + PROMPT_NONE, 0 },
> +
> /* Pass 3A Directory Optimization */
>
> /* Pass 3A: Optimizing directories */
> diff --git a/e2fsck/problem.h b/e2fsck/problem.h
> index d291e26..dfdb6d0 100644
> --- a/e2fsck/problem.h
> +++ b/e2fsck/problem.h
> @@ -1078,6 +1078,12 @@ struct problem_context {
> /* Lost+found is encrypted */
> #define PR_3_LPF_ENCRYPTED 0x03001B
>
> +/* Error while creating an extent for /lost+found */
> +#define PR_3_ERR_LPF_OPEN_EXTENT 0x03001C
> +
> +/* Error while mapping a directory block for /lost+found */
> +#define PR_3_ERR_LPF_SET_BMAP 0x03001D
> +
> /*
> * Pass 3a --- rehashing diretories
> */
> --
> 2.1.4
>
* Darrick J. Wong <[email protected]>:
> On Thu, Mar 16, 2017 at 11:39:48AM -0400, Eric Whitney wrote:
> > When e2fsck recreates a missing lost+found directory, it uses indirect
> > block mapping. It does this even if the file system being repaired is
> > extent mapped and where the original lost+found was extent mapped when
> > created by mkfs.
> >
> > This inconsistency exists because the kernel code handling indirect
> > block mapping was considered more trustworthy in ext4's early days;
> > today, the extent handling code is robust.
> >
> > This inconsistency also causes potential behavioral problems on
> > bigalloc file systems. The ext4 kernel code does not support indirect
> > block mapped directories when bigalloc is enabled. For example, it is
> > not possible to extend lost+found on a mounted bigalloc file system
> > with the current implementation, so the behavior of that directory
> > differs from all others, potentially surprising users.
> >
> > To make the handling of lost+found more consistent, and to avoid
> > behavioral problems on bigalloc file systems, recreate lost+found
> > with extent mapping if the file system under repair is extent mapped.
> >
> > Signed-off-by: Eric Whitney <[email protected]>
> > ---
> > e2fsck/pass3.c | 29 +++++++++++++++++++++++++++--
> > e2fsck/problem.c | 10 ++++++++++
> > e2fsck/problem.h | 6 ++++++
> > 3 files changed, 43 insertions(+), 2 deletions(-)
> >
> > diff --git a/e2fsck/pass3.c b/e2fsck/pass3.c
> > index 44203ca..259c978 100644
> > --- a/e2fsck/pass3.c
> > +++ b/e2fsck/pass3.c
> > @@ -378,7 +378,7 @@ static int check_directory(e2fsck_t ctx, ext2_ino_t dir,
> > ext2_ino_t e2fsck_get_lost_and_found(e2fsck_t ctx, int fix)
> > {
> > ext2_filsys fs = ctx->fs;
> > - ext2_ino_t ino;
> > + ext2_ino_t ino;
> > blk64_t blk;
> > errcode_t retval;
> > struct ext2_inode_large inode;
> > @@ -386,6 +386,7 @@ ext2_ino_t e2fsck_get_lost_and_found(e2fsck_t ctx, int fix)
> > static const char name[] = "lost+found";
> > struct problem_context pctx;
> > int will_rehash, flags;
> > + ext2_extent_handle_t handle;
> >
> > if (ctx->lost_and_found)
> > return ctx->lost_and_found;
> > @@ -520,7 +521,10 @@ skip_new_block:
> > inode.i_atime = inode.i_ctime = inode.i_mtime = ctx->now;
> > inode.i_links_count = 2;
> > ext2fs_iblk_set(fs, EXT2_INODE(&inode), 1);
> > - inode.i_block[0] = blk;
> > + if (ext2fs_has_feature_extents(fs->super))
> > + inode.i_flags |= EXT4_EXTENTS_FL;
> > + else
> > + inode.i_block[0] = blk;
> >
> > /*
> > * Next, write out the inode.
> > @@ -553,6 +557,27 @@ skip_new_block:
> > }
> >
> > /*
> > + * If the new lost+found is extent mapped, map its first new
> > + * directory block
> > + */
> > + if (ext2fs_has_feature_extents(fs->super)) {
> > + retval = ext2fs_extent_open2(fs, ino, EXT2_INODE(&inode),
> > + &handle);
> > + if (retval) {
> > + pctx.errcode = retval;
> > + fix_problem(ctx, PR_3_ERR_LPF_OPEN_EXTENT, &pctx);
> > + return 0;
> > + }
> > + retval = ext2fs_extent_set_bmap(handle, 0, blk, 0);
> > + ext2fs_extent_free(handle);
> > + if (retval) {
> > + pctx.errcode = retval;
> > + fix_problem(ctx, PR_3_ERR_LPF_SET_BMAP, &pctx);
> > + return 0;
> > + }
> > + }
>
> You could refactor most of this function since ext2fs_bmap2() can do the
> work for you -- allocating a block, setting the extents flag, and
> initializing the extent tree header.
>
> Maybe there's a specific reason to allocate the block and then the
> inode instead of simply using bmap? I can't find any obvious reason,
> but Ted might remember something.
>
> --D
>
> > +
> > + /*
> > * Finally, create the directory link
> > */
> > pctx.errcode = ext2fs_link(fs, EXT2_ROOT_INO, name, ino, EXT2_FT_DIR);
> > diff --git a/e2fsck/problem.c b/e2fsck/problem.c
> > index c3e4f6a..736b822 100644
> > --- a/e2fsck/problem.c
> > +++ b/e2fsck/problem.c
> > @@ -1791,6 +1791,16 @@ static struct e2fsck_problem problem_table[] = {
> > N_("/@l is encrypted\n"),
> > PROMPT_CLEAR, 0 },
> >
> > + /* Error while creating an extent for /lost+found */
> > + { PR_3_ERR_LPF_OPEN_EXTENT,
> > + N_("ext2fs_@x_open2: %m while creating an @x for /@l\n"),
> > + PROMPT_NONE, 0 },
> > +
> > + /* Error while mapping a directory block for /lost+found */
> > + { PR_3_ERR_LPF_SET_BMAP,
> > + N_("ext2fs_@x_set_bmap: %m while mapping a @d @b for /@l\n"),
> > + PROMPT_NONE, 0 },
> > +
> > /* Pass 3A Directory Optimization */
> >
> > /* Pass 3A: Optimizing directories */
> > diff --git a/e2fsck/problem.h b/e2fsck/problem.h
> > index d291e26..dfdb6d0 100644
> > --- a/e2fsck/problem.h
> > +++ b/e2fsck/problem.h
> > @@ -1078,6 +1078,12 @@ struct problem_context {
> > /* Lost+found is encrypted */
> > #define PR_3_LPF_ENCRYPTED 0x03001B
> >
> > +/* Error while creating an extent for /lost+found */
> > +#define PR_3_ERR_LPF_OPEN_EXTENT 0x03001C
> > +
> > +/* Error while mapping a directory block for /lost+found */
> > +#define PR_3_ERR_LPF_SET_BMAP 0x03001D
> > +
> > /*
> > * Pass 3a --- rehashing diretories
> > */
> > --
> > 2.1.4
> >
Darrick:
As discussed in today's concall, I'd chosen to follow the pattern used in
mke2fs when lost+found is first created for conservatism's sake. As Ted noted,
the code in mke2fs predates ext2fs_bmap2, and it hasn't been modified to use it.
For now, I'll withdraw this patch, and will take a look at a cleaner
implementation as you suggest.
Thanks for the comment!
Eric