Re: [RFC 2/2] ACPI: APEI: fix reboot caused by synchronous error loop because of memory_failure() failed

[Bixuan Cui <cuibixuan@xxxxxxxxxxxxxxxxx>]

在 2022/12/5 19:51, Lv Ying 写道:
> diff --git a/mm/memory-failure.c b/mm/memory-failure.c
> index 3b6ac3694b8d..4c1c558f7161 100644
> --- a/mm/memory-failure.c
> +++ b/mm/memory-failure.c
> @@ -2266,7 +2266,11 @@ static void __memory_failure_work_func(struct work_struct *work, bool sync)
>   			break;
>   		if (entry.flags & MF_SOFT_OFFLINE)
>   			soft_offline_page(entry.pfn, entry.flags);
> -		else if (!sync || (entry.flags & MF_ACTION_REQUIRED))
> +		else if (sync) {
> +			if ((entry.flags & MF_ACTION_REQUIRED) &&
> +					memory_failure(entry.pfn, entry.flags))
> +				force_sig_mceerr(BUS_MCEERR_AR, 0, 0);
> +		} else
>   			memory_failure(entry.pfn, entry.flags);
Hi,

Some of the ideas in this patch are wrong :-(

1. As Shuai Xue said, it is wrong to judge synchronization error and 
asynchronization error through functions such as 
memory_failure_queue_kick()/ghes_proc()/ghes_proc_in_irq(), because both 
synchronization error and asynchronization error may go to the same 
notification.

2. There is no need to pass 'sync' to __memory_failure_work_func(), 
because memory_failure() can directly handle synchronous and 
asynchronous errors according to entry.flags & MF_ACTION_REQUIRED:

entry.flags & MF_ACTION_REQUIRED == 1: Action: poison page and kill task 
for synchronous error
entry.flags & MF_ACTION_REQUIRED == 0: Action: poison page for 
asynchronous error

Reference x86:
do_machine_check # MCE, synchronous
   ->kill_me_maybe
     ->memory_failure(p->mce_addr >> PAGE_SHIFT, MF_ACTION_REQUIRED);

uc_decode_notifier # CMCI, asynchronous
   ->memory_failure(pfn, 0)

At the same time, the modification here is repeated with your patch 01
 	if (sev == GHES_SEV_RECOVERABLE && sec_sev == GHES_SEV_RECOVERABLE)
-		flags = 0;
+		flags = sync ? MF_ACTION_REQUIRED : 0;

3. Why add 'force_sig_mceerr(BUS_MCEERR_AR, 0, 0)' after 
memory_failure(pfn, MF_ACTION_REQUIRED)?
The task will be killed in memory_failure():
if poisoned, kill_accessing_process()->kill_proc()
if not poisoned, hwpoison_user_mappings()->collect_procs()->kill_procs()

Reference x86 to handle synchronous error:
kill_me_maybe()
{
    int flags = MF_ACTION_REQUIRED;
    ret = memory_failure(p->mce_addr >> PAGE_SHIFT, flags);
    if (!ret) {
	...
        return;
    }
    if (ret == -EHWPOISON || ret == -EOPNOTSUPP)
        return;

    pr_err("Memory error not recovered");
    kill_me_now(cb);
}


Thanks,
Bixuan Cui