Re: [RFC] Make the memory failure blast radius more precise

[Jane Chu <jane.chu@oracle.com>]

On 6/24/2020 5:13 PM, Luck, Tony wrote:
>> Both the RFC patch and the above 5-step recovery plan look neat, step 4)
>> is nice to carry forward on icelake when a single instruction to clear
>> poison is available.
> 
> Jane,
> 
> Clearing poison has some challenges.
> 
> On persistent memory it probably works (as the DIMM is going to remap that address to a different
> part of the media to avoid the bad spot).
> 
> On DDR memory you'd need to decide whether the problem was transient, so that a simple
> overwrite fixes the problem. Or persistent ... in which case the problem will likely come back
> with the right data pattern.  To tell that you may need to run some memory test on the affected
> area.
> 
> If the error was just in a 4K page, I'd be inclined to copy the good data to a new page and
> map that in instead. Throwing away one 4K page isn't likely to be painful.
> 
> If it is in a 2M/1G page ... perhaps it is worth the effort and risk of trying to clear the poison
> in place to avoid the pain of breaking up a large page.

Thanks!  Yes I was only thinking about persistent memory, but 
memory_failure_dev_pagemap() applies to DDR as well depends on the 
underlying technology. In our use case, even if the error was just in a 
4K page, we'd like to clear the poison and reuse the page to maintain a 
contiguous 256MB extent in the filesystem.  Perhaps it is better to 
leave that to the filesystem and driver.

Regards,
-jane

> 
> -Tony
>