On Mon, 14 Sep 2026 18:49:38 -0500 Kyle Meyer <[email protected]> wrote:
> Soft offlining a HugeTLB page dissolves it, permanently reducing the
> HugeTLB page pool. This can be problematic for workloads that depend on
> a fixed number of HugeTLB pages.
>
> Currently, soft offline must be disabled to prevent HugeTLB pages from
> being soft offlined.
>
> This patch allows soft offline to be disabled for HugeTLB pages while
> remaining enabled for non-HugeTLB pages.
>
> Commit 56374430c5dfc ("mm/memory-failure: userspace controls
> soft-offlining pages") introduced the following sysctl interface to
> control soft offline:
>
> /proc/sys/vm/enable_soft_offline
>
> The interface does not distinguish between page types:
>
> 0 - Soft offline is disabled
> 1 - Soft offline is enabled
>
> Convert enable_soft_offline to a bitmask and support disabling soft
> offline for HugeTLB pages:
>
> Bits:
>
> 0 - Enable soft offline
> 1 - Disable soft offline for HugeTLB pages
>
> Supported values:
>
> 0 - Soft offline is disabled
> 1 - Soft offline is enabled
> 3 - Soft offline is enabled (disabled for HugeTLB pages)
>
> Existing behavior is preserved.
>
> Update documentation and HugeTLB soft offline selftests.
I'm hitting some rejects against
https://lore.kernel.org/[email protected].
Please update this against mm.git's mm-unstalbe branch or linux-next.
I'd updated v2's changelog with this:
Tony said:
: Recap of original problem is that some BIOS keep track of error
: threshold per-rank and use this GHES mechanism to report threshold
: exceeded on the rank.
:
: Systems that stay up a long time can accumulate enough soft errors to
: trigger this threshold. But the action of taking a page offline isn't
: going to help. For a 4K page this is merely annoying. For 1G page it
: can mess things up badly.
:
: My original patch for this just skipped the GHES->offline process for
: huge pages. But I wasn't aware of the sysctl control. That provides a
: better solution.
Which is cumbersome but I think there's useful info here so please
incorporate in some way.
> Suggested-by: Tony Luck <[email protected]>
> Signed-off-by: Kyle Meyer <[email protected]>
v2 had
Reported-by: Shawn Fan <[email protected]>
which got lost.
