lkml.org 
[lkml]   [2000]   [Oct]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: tcp_do_sendmsg() allocation still broken ?
On Sat, Oct 28, 2000 at 09:13:27PM -0200, Marcelo Tosatti wrote:
>
>
> On Sun, 29 Oct 2000, Andi Kleen wrote:
>
> > On Sat, Oct 28, 2000 at 07:12:44PM -0200, Marcelo Tosatti wrote:
> > >
> > > David,
> > >
> > > tcp_do_sendmsg() still allocates using GFP_KERNEL when it can't, it seems:
> >
> > tcp_do_sendmsg() should only be called from process context, because it can
> > sleep for other reasons anyways.
> >
> > If someone calls it from interrupt context it needs to be fixed.
>
> Andi,
>
> Think about nbd.

Making tcp_do_sendmsg use GFP_ATOMIC would make it too unreliable for other
situations. Penalizing the whole system just for nbd is not a good idea.

>
> It allocates memory inside its request function, and since we do write
> throttling in try_to_free_pages() we can end up with this deadlock in low
> memory conditions:
>
> nbd_do_request -> nbd_send_req -> nbd_xmit -> sock_sendmsg -> ... ->
> tcp_do_sendmsg -> tcp_send_skb -> skb_copy -> alloc_skb ->
> kmalloc(GFP_KERNEL) -> kmem_cache_grow -> get_free_pages ->
> try_to_free_pages -> shrink_mmap -> try_to_free_buffers ->
> sync_page_buffers -> ll_rw_block -> make_request -> add_request ->
> nbd_do_request -> nbd_send_req -> ...
>
> until there are no more free requests on the request queue and writer
> processes stuck on __get_request_wait.

Looks like a job for PF_MEMALLOC and making GFP_KERNEL fail earlier.

-Andi
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 12:45    [from the cache]
©2003-2011 Jasper Spaans