Checks are missing when delta equals 0 in __xfs_ag_resv_free() and
__xfs_ag_resv_init().
the case that the delta equals 0 is reachable with the command
sequence below:
# mkfs.xfs -f /dev/sdb5
# mount /dev/sdb5 /mnt/scratch/
where /dev/sdb5 is my disk for test. And if the patch below is
applied:
====
xfs_mod_freecounter(
if (rsvd)
ASSERT(has_resv_pool);
+ if (delta == 0)
+ dump_stack();
+
if (delta > 0) {
/*
* If the reserve pool is depleted, put blocks back into it
====
the following stack will be shown in the message:
=> xfs_mod_freecounter+0x84/0x2b8
=> __xfs_ag_resv_free+0xc4/0x188
=> xfs_ag_resv_free+0x24/0x50
=> xfs_fs_unreserve_ag_blocks+0x40/0x160
=> xfs_mountfs+0x500/0x900
=> xfs_fs_fill_super+0x3d8/0x810
=> get_tree_bdev+0x164/0x258
=> xfs_fs_get_tree+0x20/0x30
=> vfs_get_tree+0x30/0xf8
=> path_mount+0x3c4/0xa58
=> do_mount+0x74/0x98
=> xfs_mod_freecounter+0x84/0x2b8
=> __xfs_ag_resv_init+0x64/0x1d0
=> xfs_ag_resv_init+0x108/0x1c8
=> xfs_fs_reserve_ag_blocks+0x4c/0x110
=> xfs_mountfs+0x57c/0x900
=> xfs_fs_fill_super+0x3d8/0x810
=> get_tree_bdev+0x164/0x258
=> xfs_fs_get_tree+0x20/0x30
=> vfs_get_tree+0x30/0xf8
=> path_mount+0x3c4/0xa58
=> do_mount+0x74/0x98
After applying this patch, we can avoid to call xfs_mod_fdblocks when
delta equals 0.
Signed-off-by: Shida Zhang <[email protected]>
---
Changes from v1:
-Add checks before calling xfs_mod_fdblocks instead.
Changes from v2:
-Rephrase the commit description.
fs/xfs/libxfs/xfs_ag_resv.c | 16 +++++++++++++---
1 file changed, 13 insertions(+), 3 deletions(-)
diff --git a/fs/xfs/libxfs/xfs_ag_resv.c b/fs/xfs/libxfs/xfs_ag_resv.c
index fe94058d4e9e..c8fa032e4b00 100644
--- a/fs/xfs/libxfs/xfs_ag_resv.c
+++ b/fs/xfs/libxfs/xfs_ag_resv.c
@@ -149,7 +149,12 @@ __xfs_ag_resv_free(
oldresv = resv->ar_orig_reserved;
else
oldresv = resv->ar_reserved;
- error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
+
+ if (oldresv)
+ error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
+ else
+ error = 0;
+
resv->ar_reserved = 0;
resv->ar_asked = 0;
resv->ar_orig_reserved = 0;
@@ -215,8 +220,13 @@ __xfs_ag_resv_init(
if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_AG_RESV_FAIL))
error = -ENOSPC;
- else
- error = xfs_mod_fdblocks(mp, -(int64_t)hidden_space, true);
+ else {
+ error = 0;
+ if (hidden_space)
+ error = xfs_mod_fdblocks(mp, -(int64_t)hidden_space,
+ true);
+ }
+
if (error) {
trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
error, _RET_IP_);
--
2.25.1
On Tue, Jun 21, 2022 at 04:42:38PM +0800, Shida Zhang wrote:
> Checks are missing when delta equals 0 in __xfs_ag_resv_free() and
> __xfs_ag_resv_init().
>
> the case that the delta equals 0 is reachable with the command
> sequence below:
>
> # mkfs.xfs -f /dev/sdb5
> # mount /dev/sdb5 /mnt/scratch/
>
> where /dev/sdb5 is my disk for test. And if the patch below is
> applied:
>
> ====
> xfs_mod_freecounter(
> if (rsvd)
> ASSERT(has_resv_pool);
>
> + if (delta == 0)
> + dump_stack();
> +
> if (delta > 0) {
> /*
> * If the reserve pool is depleted, put blocks back into it
> ====
>
> the following stack will be shown in the message:
>
> => xfs_mod_freecounter+0x84/0x2b8
> => __xfs_ag_resv_free+0xc4/0x188
> => xfs_ag_resv_free+0x24/0x50
> => xfs_fs_unreserve_ag_blocks+0x40/0x160
> => xfs_mountfs+0x500/0x900
> => xfs_fs_fill_super+0x3d8/0x810
> => get_tree_bdev+0x164/0x258
> => xfs_fs_get_tree+0x20/0x30
> => vfs_get_tree+0x30/0xf8
> => path_mount+0x3c4/0xa58
> => do_mount+0x74/0x98
>
> => xfs_mod_freecounter+0x84/0x2b8
> => __xfs_ag_resv_init+0x64/0x1d0
> => xfs_ag_resv_init+0x108/0x1c8
> => xfs_fs_reserve_ag_blocks+0x4c/0x110
> => xfs_mountfs+0x57c/0x900
> => xfs_fs_fill_super+0x3d8/0x810
> => get_tree_bdev+0x164/0x258
> => xfs_fs_get_tree+0x20/0x30
> => vfs_get_tree+0x30/0xf8
> => path_mount+0x3c4/0xa58
> => do_mount+0x74/0x98
>
> After applying this patch, we can avoid to call xfs_mod_fdblocks when
> delta equals 0.
>
> Signed-off-by: Shida Zhang <[email protected]>
> ---
> Changes from v1:
> -Add checks before calling xfs_mod_fdblocks instead.
> Changes from v2:
> -Rephrase the commit description.
>
> fs/xfs/libxfs/xfs_ag_resv.c | 16 +++++++++++++---
> 1 file changed, 13 insertions(+), 3 deletions(-)
>
> diff --git a/fs/xfs/libxfs/xfs_ag_resv.c b/fs/xfs/libxfs/xfs_ag_resv.c
> index fe94058d4e9e..c8fa032e4b00 100644
> --- a/fs/xfs/libxfs/xfs_ag_resv.c
> +++ b/fs/xfs/libxfs/xfs_ag_resv.c
> @@ -149,7 +149,12 @@ __xfs_ag_resv_free(
> oldresv = resv->ar_orig_reserved;
> else
> oldresv = resv->ar_reserved;
> - error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
> +
> + if (oldresv)
> + error = xfs_mod_fdblocks(pag->pag_mount, oldresv, true);
> + else
> + error = 0;
> +
> resv->ar_reserved = 0;
> resv->ar_asked = 0;
> resv->ar_orig_reserved = 0;
> @@ -215,8 +220,13 @@ __xfs_ag_resv_init(
>
> if (XFS_TEST_ERROR(false, mp, XFS_ERRTAG_AG_RESV_FAIL))
> error = -ENOSPC;
> - else
> - error = xfs_mod_fdblocks(mp, -(int64_t)hidden_space, true);
> + else {
> + error = 0;
> + if (hidden_space)
> + error = xfs_mod_fdblocks(mp, -(int64_t)hidden_space,
> + true);
I understand that calling __xfs_ag_resv_init on an AG with a maximally
sized data structure can result in @hidden_space being zero here, but
why does that matter enough to change the code? Are you experiencing
problems when this happens? Unnecessary slowdowns at mount time?
Something else?
This is v3 of a patch and I still can't tell why I should care ...?
--D
> + }
> +
> if (error) {
> trace_xfs_ag_resv_init_error(pag->pag_mount, pag->pag_agno,
> error, _RET_IP_);
> --
> 2.25.1
>
> I understand that calling __xfs_ag_resv_init on an AG with a maximally
> sized data structure can result in @hidden_space being zero here, but
> why does that matter enough to change the code? Are you experiencing
> problems when this happens? Unnecessary slowdowns at mount time?
> Something else?
>
> This is v3 of a patch and I still can't tell why I should care ...?
After applying this patch, we can avoid to call xfs_mod_fdblocks when
delta equals 0. So we can reduce unnecessary operations here.
On Wed, Jun 22, 2022 at 01:57:31PM +0800, Stephen Zhang wrote:
> > I understand that calling __xfs_ag_resv_init on an AG with a maximally
> > sized data structure can result in @hidden_space being zero here, but
> > why does that matter enough to change the code? Are you experiencing
> > problems when this happens? Unnecessary slowdowns at mount time?
> > Something else?
> >
> > This is v3 of a patch and I still can't tell why I should care ...?
>
> After applying this patch, we can avoid to call xfs_mod_fdblocks when
> delta equals 0. So we can reduce unnecessary operations here.
Yeah, I get that, but what is the real world impact of those unnecessary
operations? Have you run fstests to make sure this change doesn't trip
over some weird subtlety in the code? Do the anticipated benefits
justify diverting my time to figuring out if we've really covered all
the corner cases?
IOWS: don't waste our time on theoretical improvements. There are
/plenty/ of things in 5.19 that need real attention, like generic/522
corrupting things and recoveryloop tests that trip over log recovery.
--D