Re: [PATCH] net: add per device sg_max_frags for skb

From: Eric Dumazet
Date: Wed Jan 13 2016 - 09:19:56 EST


1) There are no arch with 1K page sizes. Most certainly, if we had
MAX_SKB_FRAGS=65 some assumptions in the stack would fail.

2) TCP stack has coalescing support. write(2) or sendmsg(2) should
append data into the last skb in write queue, and still use 32 KB
frags.
You get pathological skb when using sendpage() or when one thread
writes data into _multiple_ TCP sockets, since TCP stack uses
a per thread 32 KB reserve (
http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=5640f7685831e088fe6c2e1f863a6805962f8e81
)

2) As I said, implementing a limit in TCP stack is not enough. Your
patch is therefore adding complexity for all users, but is not a
general solution.

GRO, tun device, many things can still cook 'big skbs'

You need to properly implement a fallback, possibly using
ndo_features_check(), or directly from your ndo_start_xmit()

3) We currently have a very dumb way to fallback, forcing a linearize
call, likely to fail if memory is fragmented and skb big.

You could instead provide a smart helper, trying to reduce the
number of frags in a skb by chosing adjacent frags and
re-allocating/merging them.

By choosing, I mean trying to pick smallest ones to minimize copy
cost, to get one skb with X less fragment. (X=1 in your case ?)

I know for example that bnx2x could benefit from such a helper, as
it has a 13 frags limits.
(bnx2x_pkt_req_lin(), called from bnx2x ndo_start_xmit()


On Wed, Jan 13, 2016 at 5:57 AM, Hans Westgaard Ry
<hans.westgaard.ry@xxxxxxxxxx> wrote:
>
>
> On 01/08/2016 12:47 PM, Hannes Frederic Sowa wrote:
>>
>> On 08.01.2016 10:55, Hans Westgaard Ry wrote:
>>>
>>>
>>>
>>> On 01/06/2016 02:59 PM, David Laight wrote:
>>>>
>>>> From: Hans Westgaard Ry
>>>>>
>>>>> Sent: 06 January 2016 13:16
>>>>> Devices may have limits on the number of fragments in an skb they
>>>>> support. Current codebase uses a constant as maximum for number of
>>>>> fragments (MAX_SKB_FRAGS) one skb can hold and use.
>>>>>
>>>>> When enabling scatter/gather and running traffic with many small
>>>>> messages the codebase uses the maximum number of fragments and thereby
>>>>> violates the max for certain devices.
>>>>>
>>>>> An example of such a violation is when running IPoIB on a HCA
>>>>> supporting 16 SGE on an architecture with 4K pagesize. The
>>>>> MAX_SKB_FRAGS will be 17 (64K/4K+1) and because IPoIB adds yet another
>>>>> segment we end up with send_requests with 18 SGE resulting in
>>>>> kernel-panic.
>>>>>
>>>>> The patch allows the device to limit the maximum number fragments used
>>>>> in one skb.
>>>>
>>>> This doesn't seem to me to be the correct way to fix this.
>>>> Anything that adds an extra fragment (in this case IPoIB) should allow
>>>> for the skb already having the maximum number of fragments.
>>>> Fully linearising the skb is overkill, but I think the first fragment
>>>> can be added to the linear part of the skb.
>>>>
>>>> David
>>>>
>>>>
>>> When IpoIB handles a skb-request it converts fragments to SGEs to
>>> be handled by a HCA.
>>> The problem arises when the HCA have a limited number of SGEs less than
>>> MAX_SKB_FRAGS.
>>> (it gets a little worse since IPoIB need to yet another segment)
>>> I have not found any easy way of fixing this with currenct codebase.
>>
>>
>> I think because of the complex forwarding nature, a global counter which
>> driver's can reduce during initialization time is the only solution I see
>> right now without changing the layout of the skb later on.
>>
>> Unfortunately this doesn't resolve the cases were virtual machines inject
>> gso skbs, for those there still needs to be a slow path to do the
>> reformatting of the skb. :/
>>
>> Bye,
>> Hannes
>>
>>
> The use-case for this patch is an application which sends many small
> messages, by write(2) on a TCP socket which has Nagle enabled. A
> scatter-gather capable NIC (potentially also supporting tso) will then be
> asked to send an skb containing up to MAX_SKB_FRAGS worth of fragments (17
> considering a 4kb page size, hypothetically 65 considering an arch
> supporting 1kb page size).
>
> Now, if the NIC hardware supports less _gather-fragments_, said hardware
> must run with scatter-gather disabled - or - the NIC driver has to implement
> a partial linearization of the skb to reduce #frags to what the hardware
> supports. The latter is far from elegant, and must be implemented in all NIC
> drivers which have this restriction.
>
> This patch provides the flexibility to choose the maximum number of
> fragments that can be passed down to the NIC in order to
> utilize the NIC SG hardware features.
>
>
> In our view we are discussing two different issues:
>
> 1. Is it reasonable that a NIC can restrict #frags in an skb when
> transmitting?
> 2. If yes to the above, how is this implemented the best possible way.
>
> Thanks a lot for feedback on the implementation from David Laight, Eric
> Dumazet and Hannes Fredreric Sowa.
>
> What do you think?
>
> Hans
>