Message-ID: <53887B74.7060202@canonical.com>
Date:	Fri, 30 May 2014 14:37:08 +0200
From:	Stefan Bader <stefan.bader@...onical.com>
To:	Zoltan Kiss <zoltan.kiss@...rix.com>,
	Wei Liu <wei.liu2@...rix.com>,
	Eric Dumazet <eric.dumazet@...il.com>
CC:	netdev@...r.kernel.org, xen-devel@...ts.xen.org,
	David Vrabel <david.vrabel@...rix.com>,
	Konrad Wilk <konrad.wilk@...cle.com>,
	Boris Ostrovsky <boris.ostrovsky@...cle.com>
Subject: Re: [PATCH net-next] xen-netfront: try linearizing SKB if it occupies
 too many slots

On 30.05.2014 14:07, Zoltan Kiss wrote:
> On 30/05/14 09:06, Stefan Bader wrote:
>> On 16.05.2014 18:29, Zoltan Kiss wrote:
>>> On 16/05/14 16:34, Wei Liu wrote:
>>>> On Fri, May 16, 2014 at 08:22:19AM -0700, Eric Dumazet wrote:
>>>>> On Fri, 2014-05-16 at 15:36 +0100, Wei Liu wrote:
>>>>>> On Fri, May 16, 2014 at 07:21:08AM -0700, Eric Dumazet wrote:
>>>>>>> On Fri, 2014-05-16 at 14:11 +0100, Wei Liu wrote:
>>>>>>>
>>>>>>>> It's not that common to trigger this, I only saw a few reports. In fact
>>>>>>>> Stefan's report is the first one that comes with a method to reproduce
>>>>>>>> it.
>>>>>>>>
>>>>>>>> I tested with redis-benchmark on a guest with 256MB RAM and only saw a
>>>>>>>> few "failed to linearize", never saw a single one with 1GB guest.
>>>>>>> Well, I am just saying. This is asking order-5 allocations, and yes,
>>>>>>> this is going to fail after few days of uptime, no matter what you try.
>>>>>>>
>>>>>> Hmm... I see what you mean -- memory fragmentation leads to allocation
>>>>>> failure. Thanks.
>>>>> In the mean time, have you tried to lower gso_max_size ?
>>>>>
>>>>> Setting it witk netif_set_gso_max_size() to something like 56000 might
>>>>> avoid the problem.
>>>>>
>>>>> (Not sure if it is applicable in your case)
>>>>>
>>>> It works, at least in this Redis testcase. Could you explain a bit where
>>>> this 56000 magic number comes from? :-)
>>>>
>>>> Presumably I can derive it from some constant in core network code?
>>> I guess it just makes more unlikely to have packets with problematic layout.
>>> But the following packet would still fail:
>>> linear buffer : 80 bytes, on 2 pages
>>> 17 frags, 80 bytes each, each spanning over page boundary.
>>>
>>> I just had an idea: a modified version of xenvif_handle_frag_list function
>>> from netback would be useful for us here. It recreates the frags array on
>>> fully utilized 4k pages. Plus we can use pskb_expand_head to reduce the page
>>> number on the linear buffer (although it might not work, see my comment in
>>> the patch)
>>> The worst case linear buffer then spans N+1 pages, and has N*PAGE_SIZE+1
>>> bytes. Then the frags after this coalescing should have 16*PAGE_SIZE -
>>> (N*PAGE_SIZE+2) bytes at most, which is 16-N pages. Altogether that's 16+1
>>> page, which should definitely fit!
>>> This is what I mean:
>>>
>> I had been idly wondering about this onwards. And trying to understand the whole
>> skb handling environment, I tried to come up with some idea as well. It may be
>> totally stupid and using the wrong assumptions. It seems to work in the sense
>> that things did not blow up into my face immediately and somehow I did not see
>> dropped packages due to the number of slots either.
>> But again, I am not sure I am doing the right thing. The idea was to just try to
>> get rid of so many compound pages (which I believe are the only ones that can
>> have an offset big enough to allow some alignment savings)...
> Hi,
> 
> This probably helps in a lot of scenarios, but not for everything. E.g. if the
> skb has 15 frags, each PAGE_SIZE+1 bytes, it'll still consume 30 slots for the
> frags, and this function won't be able to help that.
> It's hard to come up with an algorithm which handles all the scenarios we can
> have here, and does that with the least possible amount of copy. I'll keep
> looking into this in the next weeks, but don't hesitate, if you have an idea!

Right, this can fail and then the packet still gets dropped. I only could test
against that regis-benchmark use case which seemed in general to generate about
4 or 5 frags but mostly compound pages then and with additional printk
statements the common case was just to have that 19 slots case (rare cases 20).
So the recovery usually needs to handle 1 or 2 pages reductions. This is one
reason I decided to give that function a target on how much it needs to do (also
to not try its luck on new memory allocation to heavily).

-Stefan
> 
> Regads,
> 
> Zoltan
>>
>> -Stefan
>>
>>
>>  From 8571b106643b32296e58526e2fbe97c330877ac8 Mon Sep 17 00:00:00 2001
>> From: Stefan Bader <stefan.bader@...onical.com>
>> Date: Thu, 29 May 2014 12:18:01 +0200
>> Subject: [PATCH] xen-netfront: Align frags to fit max slots
>>
>> In cases where the frags in a skb require more than MAX_SKB_FRAGS + 1
>> (= 18) 4K pages of grant pages, try to reduce the footprint by moving
>> the data to new pages and have it aligned to the beginning.
>> Then replace the page in the frag and release the old one. This sure is
>> more expensive in compute but should happen not too often and sounds
>> better than to just drop the packet in that case.
>>
>> Signed-off-by: Stefan Bader <stefan.bader@...onical.com>
>> ---
>>   drivers/net/xen-netfront.c | 65 +++++++++++++++++++++++++++++++++++++++++++---
>>   1 file changed, 62 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/net/xen-netfront.c b/drivers/net/xen-netfront.c
>> index 158b5e6..ad71e5c 100644
>> --- a/drivers/net/xen-netfront.c
>> +++ b/drivers/net/xen-netfront.c
>> @@ -544,6 +544,61 @@ static int xennet_count_skb_frag_slots(struct sk_buff *skb)
>>       return pages;
>>   }
>>
>> +/*
>> + * Align data to new pages in order to save slots required to
>> + * transmit this buffer.
>> + * @skb - socket buffer
>> + * @target - number of pages to save
>> + * returns the number of pages the fragments have been reduced of
>> + */
>> +static int xennet_align_frags(struct sk_buff *skb, int target)
>> +{
>> +    int i, frags = skb_shinfo(skb)->nr_frags;
>> +    int reduced = 0;
>> +
>> +    for (i = 0; i < frags; i++) {
>> +        skb_frag_t *frag = skb_shinfo(skb)->frags + i;
>> +        struct page *fpage = skb_frag_page(frag);
>> +        struct page *npage;
>> +        unsigned long size;
>> +        unsigned long offset;
>> +        gfp_t gfp;
>> +        int order;
>> +
>> +        if (!PageCompound(fpage))
>> +            continue;
>> +
>> +        size = skb_frag_size(frag);
>> +        offset = frag->page_offset & ~PAGE_MASK;
>> +
>> +        /*
>> +         * If the length of data in the last subpage of a compound
>> +         * page is smaller than the offset into the first data sub-
>> +         * page, we can save a subpage by copying data around.
>> +         */
>> +        if ( ((offset + size) & ~PAGE_MASK) > offset )
>> +            continue;
>> +
>> +        gfp = GFP_ATOMIC | __GFP_COLD;
>> +        order = PFN_UP(size);
>> +        if (order)
>> +            gfp |= __GFP_COMP | __GFP_NOWARN;
>> +
>> +        npage = alloc_pages(gfp, order);
>> +        if (!npage)
>> +            break;
>> +        memcpy(page_address(npage), skb_frag_address(frag), size);
>> +        frag->page.p = npage;
>> +        frag->page_offset = 0;
>> +        put_page(fpage);
>> +
>> +        if (++reduced >= target)
>> +            break;
>> +    }
>> +
>> +    return reduced;
>> +}
>> +
>>   static int xennet_start_xmit(struct sk_buff *skb, struct net_device *dev)
>>   {
>>       unsigned short id;
>> @@ -573,9 +628,13 @@ static int xennet_start_xmit(struct sk_buff *skb, struct
>> net_device *dev)
>>       slots = DIV_ROUND_UP(offset + len, PAGE_SIZE) +
>>           xennet_count_skb_frag_slots(skb);
>>       if (unlikely(slots > MAX_SKB_FRAGS + 1)) {
>> -        net_alert_ratelimited(
>> -            "xennet: skb rides the rocket: %d slots\n", slots);
>> -        goto drop;
>> +        slots -= xennet_align_frags(skb, slots - (MAX_SKB_FRAGS + 1));
>> +        if (slots > MAX_SKB_FRAGS + 1) {
>> +            net_alert_ratelimited(
>> +                "xennet: skb rides the rocket: %d slots\n",
>> +                slots);
>> +            goto drop;
>> +        }
>>       }
>>
>>       spin_lock_irqsave(&np->tx_lock, flags);
> 



Download attachment "signature.asc" of type "application/pgp-signature" (885 bytes)

Powered by blists - more mailing lists