Re: 2.6.14-rc5 e1000 and page allocation failures.. still

Re: 2.6.14-rc5 e1000 and page allocation failures.. still

[Jesse Brandeburg]

On 10/23/05, Robert Hancock <hancockr@shaw.ca> wrote:
> John Bäckstrand wrote:
> > Im seeing a massive amount of page allocation failures with 2.6.14-rc5,
> > and also earlier kernels, see "E1000 - page allocation failure - saga
[snip]
> It looks like you have enough memory free - the problem is that the
> driver is allocating a block of memory with order 3, which is 8 pages.
> Quite likely there are not enough contiguous free pages to satisfy that.
>
> That's an awful big buffer size for a packet - I assume you're using
> jumbo frames or something? Ideally the driver and hardware should be
> able to allocate a buffer for those packets in multiple chunks, but I
> have no idea if this is possible.

the latest e1000 driver (6.2.15) from http://sf.net/projects/e1000
fixes this by using multiple descriptors for jumbo frames, therefore
only doing order 0 (single page) page allocations.

let us know how it goes.

BTW why is this so much more common with recent kernels?

Re: 2.6.14-rc5 e1000 and page allocation failures.. still

[Bill Davidsen]

Jesse Brandeburg wrote:
> On 10/23/05, Robert Hancock <hancockr@shaw.ca> wrote:
> 
>>John Bäckstrand wrote:
>>
>>>Im seeing a massive amount of page allocation failures with 2.6.14-rc5,
>>>and also earlier kernels, see "E1000 - page allocation failure - saga
> 
> [snip]
> 
>>It looks like you have enough memory free - the problem is that the
>>driver is allocating a block of memory with order 3, which is 8 pages.
>>Quite likely there are not enough contiguous free pages to satisfy that.
>>
>>That's an awful big buffer size for a packet - I assume you're using
>>jumbo frames or something? Ideally the driver and hardware should be
>>able to allocate a buffer for those packets in multiple chunks, but I
>>have no idea if this is possible.
> 
> 
> the latest e1000 driver (6.2.15) from http://sf.net/projects/e1000
> fixes this by using multiple descriptors for jumbo frames, therefore
> only doing order 0 (single page) page allocations.
> 
> let us know how it goes.
> 
> BTW why is this so much more common with recent kernels?

I don't know why it's more common, but I agree that it seems so. I have 
speculated that it may be related to 4k stack, but I can't even generate 
a credible wild-ass guess on that, much less find any evidence, so I 
doubt that's much if any correlation.

Getting memory a page at a time is ugly, but it will probably work just 
fine.

-- 
    -bill davidsen (davidsen@tmr.com)
"The secret to procrastination is to put things off until the
  last possible moment - but no longer"  -me