Re: System freezes after OOM

[Mikulas Patocka]

On Wed, 13 Jul 2016, Milan Broz wrote:

> On 07/13/2016 02:50 PM, Michal Hocko wrote:
> > On Wed 13-07-16 13:10:06, Michal Hocko wrote:
> >> On Tue 12-07-16 19:44:11, Mikulas Patocka wrote:
> > [...]
> >>> As long as swapping is in progress, the free memory is below the limit 
> >>> (because the swapping activity itself consumes any memory over the limit). 
> >>> And that triggered the OOM killer prematurely.
> >>
> >> I am not sure I understand the last part. Are you saing that we trigger
> >> OOM because the initiated swapout will not be able to finish the IO thus
> >> release the page in time?
> >>
> >> The oom detection checks waits for an ongoing writeout if there is no
> >> reclaim progress and at least half of the reclaimable memory is either
> >> dirty or under writeback. Pages under swaout are marked as under
> >> writeback AFAIR. The writeout path (dm-crypt worker in this case) should
> >> be able to allocate a memory from the mempool, hand over to the crypt
> >> layer and finish the IO. Is it possible this might take a lot of time?
> > 
> > I am not familiar with the crypto API but from what I understood from
> > crypt_convert the encryption is done asynchronously. Then I got lost in
> > the indirection. Who is completing the request and from what kind of
> > context? Is it possible it wouldn't be runable for a long time?
> 
> If you mean crypt_convert in dm-crypt, then it can do asynchronous completion
> but usually (with AES-NI ans sw implementations) it run the operation completely
> synchronously.
> Asynchronous processing is quite rare, usually only on some specific hardware
> crypto accelerators.
> 
> Once the encryption is finished, the cloned bio is sent to the block
> layer for processing.
> (There is also some magic with sorting writes but Mikulas knows this better.)

dm-crypt receives requests in crypt_map, then it distributes write 
requests to multiple encryption threads. Encryption is done usually 
synchronously; asynchronous completion is used only when using some PCI 
cards that accelerate encryption. When encryption finishes, the encrypted 
pages are submitted to a thread dmcrypt_write that sorts the requests 
using rbtree and submits them.

The block layer has a deficiency that it cannot merge adjacent requests 
submitted by the different threads.

If we submitted requests directly from encryption threads, lack of merging 
degraded performance seriously.

Mikulas

> Milan
> p.s. I added cc to dm-devel, some dmcrypt people reads only this list.
>