Re: [PATCH] fdtable: Avoid triggering OOMs from alloc_fdmem

Eric Dumazet [off-list ref] writes:

On Tue, 2014-02-04 at 09:22 -0800, Eric W. Biederman wrote:

quoted

The two code paths below certainly look good canidates for having
__GFP_NORETRY added to them.  The same issues I ran into with
alloc_fdmem are likely to show up there as well.

Yes, this is what I thought : a write into TCP socket should be more
frequent than the alloc_fdmem() case ;)

But then, maybe your workload was only using UDP ?

As I have heard it described one tcp connection per small requestion,
and someone goofed and started creating new connections when the server
was bogged down.  But since all of the requests and replies were small I
don't expect even TCP would allocate more than a 4KiB page in that
worload.

I had oodles of 4KiB and 8KiB pages.  What size of memory allocation did
you see failing?  

Eric

On Tue, 2014-02-04 at 10:57 -0800, Eric W. Biederman wrote:

As I have heard it described one tcp connection per small requestion,
and someone goofed and started creating new connections when the server
was bogged down.  But since all of the requests and replies were small I
don't expect even TCP would allocate more than a 4KiB page in that
worload.

Right, small writes uses regular skb (no page fragments).

I had oodles of 4KiB and 8KiB pages.  What size of memory allocation did
you see failing?  

We got some reports of order-3 allocations failing.


--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>