Linux Virtualization - Jan 2014 - [PATCH net-next 1/3] net: allow > 0 order atomic page alloc in skb_page_frag_refill

>> On Thu, 2014-01-02 at 16:56 -0800, Eric Dumazet wrote:
>>
>> Hmm... it looks like I missed __GFP_NORETRY
>>
>>
>>
>> diff --git a/net/core/sock.c b/net/core/sock.c
>> index 5393b4b719d7..5f42a4d70cb2 100644
>> --- a/net/core/sock.c
>> +++ b/net/core/sock.c
>> @@ -1872,7 +1872,7 @@ bool skb_page_frag_refill(unsigned int sz, struct
page_frag *pfrag, gfp_t prio)
>>                 gfp_t gfp = prio;
>>
>>                 if (order)
>> -                       gfp |= __GFP_COMP | __GFP_NOWARN;
>> +                       gfp |= __GFP_COMP | __GFP_NOWARN |
__GFP_NORETRY;
>>                 pfrag->page = alloc_pages(gfp, order);
>>                 if (likely(pfrag->page)) {
>>                         pfrag->offset = 0;
>>
>>
>>
There is another patch needed (looks like good stable fixes):
diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index 06e72d3..d42d48c 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -378,7 +378,7 @@ refill:
                        gfp_t gfp = gfp_mask;

                        if (order)
-                               gfp |= __GFP_COMP | __GFP_NOWARN;
+                               gfp |= __GFP_COMP | __GFP_NOWARN |
__GFP_NORETRY;
                        nc->frag.page = alloc_pages(gfp, order);
                        if (likely(nc->frag.page))
                                break;

This reduces the really pathological compact/reclaim behavior but
doesn't fix it. Actually it still really quite bad because the whole
thing loops until it gets to order-0 so it's effectively trying the
allocation 4 times anyway. I typically see non-zero order allocations
very rarely without these two pieces of code. I hotpatched a running
system to get results from this quickly. Even setting the max order to
order-1 I still see bad behavior. If anything this behavior should be
conditional until this is ironed out.

Performance data: http://pastebin.ubuntu.com/6687527/

-Debabrata

On Fri, 2014-01-03 at 17:47 -0500, Debabrata Banerjee
wrote:
> >> On Thu, 2014-01-02 at 16:56 -0800, Eric Dumazet wrote:
> >>
> >> Hmm... it looks like I missed __GFP_NORETRY
> >>
> >>
> >>
> >> diff --git a/net/core/sock.c b/net/core/sock.c
> >> index 5393b4b719d7..5f42a4d70cb2 100644
> >> --- a/net/core/sock.c
> >> +++ b/net/core/sock.c
> >> @@ -1872,7 +1872,7 @@ bool skb_page_frag_refill(unsigned int sz,
struct page_frag *pfrag, gfp_t prio)
> >>                 gfp_t gfp = prio;
> >>
> >>                 if (order)
> >> -                       gfp |= __GFP_COMP | __GFP_NOWARN;
> >> +                       gfp |= __GFP_COMP | __GFP_NOWARN |
__GFP_NORETRY;
> >>                 pfrag->page = alloc_pages(gfp, order);
> >>                 if (likely(pfrag->page)) {
> >>                         pfrag->offset = 0;
> >>
> >>
> >>
> 
> There is another patch needed (looks like good stable fixes):
> diff --git a/net/core/skbuff.c b/net/core/skbuff.c
> index 06e72d3..d42d48c 100644
> --- a/net/core/skbuff.c
> +++ b/net/core/skbuff.c
> @@ -378,7 +378,7 @@ refill:
>                         gfp_t gfp = gfp_mask;
> 
>                         if (order)
> -                               gfp |= __GFP_COMP | __GFP_NOWARN;
> +                               gfp |= __GFP_COMP | __GFP_NOWARN |
> __GFP_NORETRY;
>                         nc->frag.page = alloc_pages(gfp, order);
>                         if (likely(nc->frag.page))
>                                 break;
> 
This is in GFP_ATOMIC cases, I dont think it can ever start compaction.
> This reduces the really pathological compact/reclaim behavior but
> doesn't fix it. Actually it still really quite bad because the whole
> thing loops until it gets to order-0 so it's effectively trying the
> allocation 4 times anyway. I typically see non-zero order allocations
> very rarely without these two pieces of code. I hotpatched a running
> system to get results from this quickly. Even setting the max order to
> order-1 I still see bad behavior. If anything this behavior should be
> conditional until this is ironed out.
> 
> Performance data: http://pastebin.ubuntu.com/6687527/
It seems that you shoot the messenger : If memory is fragmented, then
one order-1 allocation is going to start compaction.

It can be a simple fork().

If your workload never fork(), then yes, you never needed compaction.

It doesn't really matter to say that which memory allocation triggered
compaction, which is a normal step in mm layer.

If you believe its badly done, you should ask to mm guys to fix/improve
it, not netdev...

We are not trying to optimize the kernel behavior for hosts in deep
memory pressure.

Using order-3 pages in TCP stack improves performance for 99% of the
hosts, there might be something wrong on your side ?

On Fri, Jan 3, 2014 at 5:54 PM, Eric Dumazet <eric.dumazet at gmail.com>
wrote:
>
> This is in GFP_ATOMIC cases, I dont think it can ever start compaction.
I think that's right I probably finally got it back to normal behavior
with order-0 allocations.
>
> It seems that you shoot the messenger : If memory is fragmented, then
> one order-1 allocation is going to start compaction.
>
> It can be a simple fork().
>
> If your workload never fork(), then yes, you never needed compaction.
>
Sure but the rate of network packets in and out and subsequent
allocations would be more equivalent to a fork bomb than normal
forking. I understand mm should work more sanely in this scenario but
at the same time we see a bad regression with this code, I see we're
not alone.
>
> We are not trying to optimize the kernel behavior for hosts in deep
> memory pressure.
We're leaving about half for the kernel so I wouldn't call it
"deep".
Any server application that is using page cache and mlocked memory
will run into similar issues.
> It doesn't really matter to say that which memory allocation triggered
> compaction, which is a normal step in mm layer.
>
> If you believe its badly done, you should ask to mm guys to fix/improve
> it, not netdev...
>
> Using order-3 pages in TCP stack improves performance for 99% of the
> hosts, there might be something wrong on your side ?
>
Having lots of memory mlocked is bad right now yes, but not
necessarily an uncommon scenario. We're handing mm an almost
intractable problem. I see compaction of mlocked pages has been
discussed a few times over there, but no patch has actually made it
in.

-Debabrata