From: Artem Blagodarenko <[email protected]>
EXT2FS_BMAP64_RBTREE is too expensive for fragmented prtition,
that can lead to situation than e2fsck use swapfile.
This patch change EXT2FS_BMAP64_RBTREE to bimap.
Marked as RFC because it must be descussed whether this flags
should be changed by default or some additional option or
a heuristic is needed to contol this flags.
signed-off-by: Artem Blagodarenko <[email protected]>
Change-Id: I6a906b5e54cf40eaba82624d8e4c2b0f90132813
---
e2fsck/pass1.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/e2fsck/pass1.c b/e2fsck/pass1.c
index 26b9ab71..563dcdc5 100644
--- a/e2fsck/pass1.c
+++ b/e2fsck/pass1.c
@@ -1247,7 +1247,7 @@ void e2fsck_pass1(e2fsck_t ctx)
return;
}
pctx.errcode = e2fsck_allocate_subcluster_bitmap(fs,
- _("in-use block map"), EXT2FS_BMAP64_RBTREE,
+ _("in-use block map"), EXT2FS_BMAP64_BITARRAY,
"block_found_map", &ctx->block_found_map);
if (pctx.errcode) {
pctx.num = 1;
@@ -1256,7 +1256,7 @@ void e2fsck_pass1(e2fsck_t ctx)
return;
}
pctx.errcode = e2fsck_allocate_block_bitmap(fs,
- _("metadata block map"), EXT2FS_BMAP64_RBTREE,
+ _("metadata block map"), EXT2FS_BMAP64_BITARRAY,
"block_metadata_map", &ctx->block_metadata_map);
if (pctx.errcode) {
pctx.num = 1;
@@ -2456,7 +2456,7 @@ static int check_ext_attr(e2fsck_t ctx, struct problem_context *pctx,
if (!ctx->block_ea_map) {
pctx->errcode = e2fsck_allocate_block_bitmap(fs,
_("ext attr block map"),
- EXT2FS_BMAP64_RBTREE, "block_ea_map",
+ EXT2FS_BMAP64_BITARRAY, "block_ea_map",
&ctx->block_ea_map);
if (pctx->errcode) {
pctx->num = 2;
--
2.27.0
On Apr 22, 2022, at 7:42 PM, Artem Blagodarenko <[email protected]> wrote:
>
> From: Artem Blagodarenko <[email protected]>
>
> EXT2FS_BMAP64_RBTREE is too expensive for fragmented prtition,
> that can lead to situation than e2fsck use swapfile.
>
> This patch change EXT2FS_BMAP64_RBTREE to bitmap.
>
> Marked as RFC because it must be descussed whether this flags
> should be changed by default or some additional option or
> a heuristic is needed to contol this flags.
Artem, I think switching over to a bitmap would be worse for the
case of relatively empty filesystems, and would likely increase
memory usage for many full filesystems as well. For a very large
filesystem a full bitmap might take tens of GB of RAM per use.
I think a better (but more complex) solution would be to have a
hybrid model, where an rbtree is used for a group if it is more
efficient, but once there are too many fragments in a group and
it consumes as much memory than a bitmap (4KB) then it would be
flattened into a rbtree leaf that is a 4KB bitmap so it doesn't
keep expanding. That would give the best of both worlds, but I
have no idea how easy/hard this would be to implement.
Cheers, Andreas
> signed-off-by: Artem Blagodarenko <[email protected]>
>
> Change-Id: I6a906b5e54cf40eaba82624d8e4c2b0f90132813
> ---
> e2fsck/pass1.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/e2fsck/pass1.c b/e2fsck/pass1.c
> index 26b9ab71..563dcdc5 100644
> --- a/e2fsck/pass1.c
> +++ b/e2fsck/pass1.c
> @@ -1247,7 +1247,7 @@ void e2fsck_pass1(e2fsck_t ctx)
> return;
> }
> pctx.errcode = e2fsck_allocate_subcluster_bitmap(fs,
> - _("in-use block map"), EXT2FS_BMAP64_RBTREE,
> + _("in-use block map"), EXT2FS_BMAP64_BITARRAY,
> "block_found_map", &ctx->block_found_map);
> if (pctx.errcode) {
> pctx.num = 1;
> @@ -1256,7 +1256,7 @@ void e2fsck_pass1(e2fsck_t ctx)
> return;
> }
> pctx.errcode = e2fsck_allocate_block_bitmap(fs,
> - _("metadata block map"), EXT2FS_BMAP64_RBTREE,
> + _("metadata block map"), EXT2FS_BMAP64_BITARRAY,
> "block_metadata_map", &ctx->block_metadata_map);
> if (pctx.errcode) {
> pctx.num = 1;
> @@ -2456,7 +2456,7 @@ static int check_ext_attr(e2fsck_t ctx, struct problem_context *pctx,
> if (!ctx->block_ea_map) {
> pctx->errcode = e2fsck_allocate_block_bitmap(fs,
> _("ext attr block map"),
> - EXT2FS_BMAP64_RBTREE, "block_ea_map",
> + EXT2FS_BMAP64_BITARRAY, "block_ea_map",
> &ctx->block_ea_map);
> if (pctx->errcode) {
> pctx->num = 2;
> --
> 2.27.0
>
Cheers, Andreas