Re: 2.6.10-rc1-mm4 -1 EAGAIN after allocation failure was: Re: Kernel 2.6.9 Multiple Page Allocation Failures

Stefan Schmidt wrote:

On Tue, Nov 09, 2004 at 05:39:20PM -0800, Andrew Morton wrote:

quoted

Well you've definitely used up all the memory which is available for atomic
allocations.  Are you using an increased /proc/sys/vm/min_free_kbytes there?

Yes, vm.min_free_kbytes=8192.
For other vm-settings find sysctl.conf attached.

Netdev: tg3 BCM5704r03, TSO off, ~32kpps rx, ~35kpps tx, ~2 rx errors/s

quoted

As for the application collapse: dunno.  Maybe networking broke.  It would
be interesting to test Linus's current tree, at
ftp://ftp.kernel.org/pub/linux/kernel/v2.6/snapshots/patch-2.6.10-rc1-bk19.gz

Will try that tomorrow. Would you suggest printing out show_free_areas();
there too? I don't know what kind of an overhead that will generate on
subsequent stack traces.

Stefan,
Can you try the following patch, please? It is diffed against 2.6.10-rc1,
but I think it should apply to -mm kernels as well.

Basically 2.6.8 and earlier kernels had some quirks in the page allocator
that would allow for example, a large portion of "DMA" memory to be reserved
for network memory allocations (atomic allocations). After 'fixing' this
problem, 2.6.9 is effectively left with about a quarter the amount of memory
reserved for network allocations compared with 2.6.8.

The following patch roughly restores parity there. Thanks.

Nick