On Sat, 2006-08-12 at 12:47 +0400, Evgeniy Polyakov wrote:

On Fri, Aug 11, 2006 at 11:42:50PM -0400, Rik van Riel (riel@redhat.com) wrote:

quoted

quoted

Dropping these non-essential packets makes sure the reserve memory 
doesn't get stuck in some random blocked user-space process, hence
you can make progress.

In short:
 - every incoming packet needs to be received at the packet level
 - when memory is low, we only deliver data to memory critical sockets
 - packets to other sockets get dropped, so the memory can be reused
   for receiving other packets, including the packets needed for the
   memory critical sockets to make progress

Forwarding packets while in low memory mode should not be a problem
at all, since forwarded packets get freed quickly.

The memory pool for receiving packets does not need much accounting
of any kind, since every packet will end up coming from that pool
when normal allocations start failing.   Maybe Evgeniy's allocator
can do something smarter internally, and mark skbuffs as MEMALLOC
when the number of available skbuffs is getting low?

As you described above, memory for each packet must be allocated (either
from SLAB or from reserve), so network needs special allocator in OOM
condition, and that allocator should be separated from SLAB's one which 
got OOM, so my purpose is just to use that different allocator (with
additional features) for netroking always. Since every piece of
networking is limited (socket queues, socket numbers, hardware queues,
hardware wire speeds an so on) there is always a maximum amount of
memory it can consume and can never exceed, so if network allocator will 
get that amount of memory at the begining, it will never meet OOM, 
so it will _always_ work and thus can allow to make slow progress for 
OOM-capable things like block devices and swap issues. 
There are no special reserve and no need to switch to/from it and 
no possibility to have OOM by design.

I'm not sure if the network stack is bounded as you say; for instance
imagine you taking a lot of packets for blocked user-space processes,
these will just accumulate in the network stack and go nowhere. In that
case memory usage is very much unbounded.

Even if blocked sockets would only accept a limited amount of packets,
it would then become a function of the amount of open sockets, which is
again unbounded.

In any scheme you need to bound the amount of memory, and in low memory
situations it is very usefull to return memory as soon as possible.

Hence this approach, it uses the global memory reserve, and a page based
allocator that does indeed return memory instantly (yes the one I have
show so far is ghastly, I've since written a nice one).