[PATCH v2] ACPI: APEI: Handle repeated SEA error storms

[Luck, Tony]

On Wed, May 27, 2026 at 04:27:07PM +0800, Junhao He wrote:
> When hardware memory corruption occurs and a user process accesses the
> corrupted page, the CPU triggers a Synchronous External Abort (SEA).
> The kernel invokes do_sea() to handle the exception, which calls
> memory_failure() to handle the faulty page.
> 
> Scenario 1: Memory Error Interrupt First, then SEA
> The page is already poisoned by the memory error interrupt path. The
> subsequent SEA handler sends a SIGBUS to the task, which accesses the
> poisoned page. This flow is correct.
> 
> Scenario 2: SEA first, then memory error interrupt (problematic scenario)
> If a user task directly accesses corrupted memory through a PFNMAP-style
> mapping (e.g., devmem), the page may still be in the free-buddy state when
> SEA is handled. In this case, memory_failure() will poison the page without
> invoking kill_accessing_process(), and then takes the free-buddy recovery
> path.
> 
> After the CPU returns to the task context, the task re-enters the SEA
> handler due to the same access. However, ghes_estatus_cached() suppresses
> all subsequent entries during the 10-second window, preventing
> ghes_do_proc() from being called. This suppression blocks the
> MF_ACTION_REQUIRED-based SIGBUS delivery, causing the kernel to fail to
> kill the task immediately. Consequently, the process keeps re-entering
> the SEA handler, leading to an SEA storm. Later, the memory error
> interrupt path also cannot kill the task, leaving the system stuck in
> this repeated loop.
> 
> The following error logs are explained using the devmem process:
>   NOTICE:  SEA Handle
>   [Hardware Error]: Hardware error from APEI Generic Hardware Error Source: 9
>   [Hardware Error]: event severity: recoverable
>   [Hardware Error]:   section_type: ARM processor error
>   [Hardware Error]:    physical fault address: 0x0000001000093c00
>   [T54990] Memory failure: 0x1000093: recovery action for free buddy page: Recovered
>   [ T9955] EDAC MC0: 1 UE Multi-bit ECC on unknown memory
>            (page:0x1000093 offset:0xc00 grain:1 - APEI location: ...)
>   NOTICE:  SEA Handle
>   NOTICE:  SEA Handle
>   ...
>   ...        ---> SEA storm
>   ...
>   NOTICE:  SEA Handle
>   [ T9955] Memory failure: 0x1000093: already hardware poisoned
>   ghes_print_estatus: 1 callbacks suppressed
>   [Hardware Error]: Hardware error from APEI Generic Hardware Error Source: 9
>   [Hardware Error]: event severity: recoverable
>   [Hardware Error]:   section_type: ARM processor error
>   [Hardware Error]:    physical fault address: 0x0000001000093c00
>   [T54990] Memory failure: 0x1000093: already hardware poisoned
>   [T54990] 0x1000093: Sending SIGBUS to devmem:54990 due to hardware memory corruption
> 
> To resolve this, return an error when encountering the same SEA again.
> The subsequent SEA handler invocation uses arm64_notify_die() to send a
> SIGBUS signal to the task, which terminates the process and prevents it
> from re-entering the handler loop.
> 
> Signed-off-by: Junhao He <hejunhao3 at h-partners.com>
> ---
>  drivers/acpi/apei/ghes.c | 10 +++++++++-
>  1 file changed, 9 insertions(+), 1 deletion(-)
> 
> Changes in V2:
> 1. update the commit message per suggestion from Xueshuai
> 2. Add a check to only return failure on the ghes_notify_sea() path,
>    avoiding impact on other NMI-type GHES handlers.
> Link to V1 - https://lore.kernel.org/all/20251030071321.2763224-1-hejunhao3@h-partners.com/
> 
> diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c
> index 3236a3ce79d6..787664740150 100644
> --- a/drivers/acpi/apei/ghes.c
> +++ b/drivers/acpi/apei/ghes.c
> @@ -1383,8 +1383,16 @@ static int ghes_in_nmi_queue_one_entry(struct ghes *ghes,

ghes_in_nmi_queue_one_entry() is called from two places:

__ghes_sdei_callback()

I think, but I'm not sure, that you are addressing call from here. By
returning -ECANCELLED this code skips the irq_work_queue() call and
returns -ENOENT (instead of zero).

ghes_in_nmi_spool_from_list()

Effect here is more complex. Depends on whether there are multiple
ghes entries on the rcu_list. If there was just one, and you return
-ECANCELLED, then irq_work_queue() is skipped. But if there are multiple
entries on the list and at least one of them gets a zero return from
ghes_in_nmi_queue_one_entry() then irq_work_queue() is called.



So which of these callers is the one that needs this fix? And is the
other called OK with getting -ECANCELLED error?



>  	ghes_clear_estatus(ghes, &tmp_header, buf_paddr, fixmap_idx);
>  
>  	/* This error has been reported before, don't process it again. */
> -	if (ghes_estatus_cached(estatus))
> +	if (ghes_estatus_cached(estatus)) {
> +		/*
> +		 * Return failure on duplicate SEA entries so that the
> +		 * subsequent SEA handler invocation sends a SIGBUS signal to
> +		 * the task to prevent it from re-entering the handler loop.
> +		 */
> +		if (is_hest_sync_notify(ghes))
> +			rc = -ECANCELED;
>  		goto no_work;
> +	}
>  
>  	llist_add(&estatus_node->llnode, &ghes_estatus_llist);
>  
> -- 
> 2.33.0
> 

-Tony