lkml.org 
[lkml]   [2015]   [Sep]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH net-next RFC] net: increase LL_MAX_HEADER for Hyper-V
From
Date
On 09/16/2015 09:25 AM, David Laight wrote:
> From: Haiyang Zhang
>> Sent: 16 September 2015 17:09
>>> -----Original Message-----
>>> From: Vitaly Kuznetsov [mailto:vkuznets@redhat.com]
>>> Sent: Wednesday, September 16, 2015 11:50 AM
>>> To: netdev@vger.kernel.org
>>> Cc: David S. Miller <davem@davemloft.net>; linux-kernel@vger.kernel.org;
>>> KY Srinivasan <kys@microsoft.com>; Haiyang Zhang
>>> <haiyangz@microsoft.com>; Jason Wang <jasowang@redhat.com>
>>> Subject: [PATCH net-next RFC] net: increase LL_MAX_HEADER for Hyper-V
>>>
>>> Commit b08cc79155fc26d0d112b1470d1ece5034651a4b ("hv_netvsc: Eliminate
>>> memory allocation in the packet send path") introduced skb headroom
>>> request for Hyper-V netvsc driver:
>>>
>>> max_needed_headroom = sizeof(struct hv_netvsc_packet) +
>>> sizeof(struct rndis_message) +
>>> NDIS_VLAN_PPI_SIZE + NDIS_CSUM_PPI_SIZE +
>>> NDIS_LSO_PPI_SIZE + NDIS_HASH_PPI_SIZE;
>>> ...
>>> net->needed_headroom = max_needed_headroom;
>>>
>>> max_needed_headroom is 220 bytes, it significantly exceeds the
>>> LL_MAX_HEADER setting. This causes each skb to be cloned on send path,
>>> e.g. for IPv4 case we fall into the following clause
>>> (ip_finish_output2()):
>>>
>>> if (unlikely(skb_headroom(skb) < hh_len && dev->header_ops)) {
>>> ...
>>> skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
>>> ...
>>> }
>>>
>>> leading to a significant performance regression. Increase LL_MAX_HEADER
>>> to make it suitable for netvsc, make it 224 to be 16-aligned.
>>> Alternatively we could (partially) revert the commit which introduced
>>> skb
>>> headroom request restoring manual memory allocation on transmit path.
>>>
>>> Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
>>> ---
>>> include/linux/netdevice.h | 4 +++-
>>> 1 file changed, 3 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
>>> index 88a0069..7233790 100644
>>> --- a/include/linux/netdevice.h
>>> +++ b/include/linux/netdevice.h
>>> @@ -132,7 +132,9 @@ static inline bool dev_xmit_complete(int rc)
>>> * used.
>>> */
>>>
>>> -#if defined(CONFIG_WLAN) || IS_ENABLED(CONFIG_AX25)
>>> +#if IS_ENABLED(CONFIG_HYPERV_NET)
>>> +# define LL_MAX_HEADER 224
>>> +#elif defined(CONFIG_WLAN) || IS_ENABLED(CONFIG_AX25)
>>> # if defined(CONFIG_MAC80211_MESH)
>>> # define LL_MAX_HEADER 128
>>> # else
>> Thanks for the patch.
>> To avoid we forget to update that 224 number when we add more things
>> into netvsc header, I suggest that we define a macro in netdevice.h such
>> as:
>> #define HVNETVSC_MAX_HEADER 224
>> #define LL_MAX_HEADER HVNETVSC_MAX_HEADER
>>
>> And, put a note in netvsc code saying the header reservation shouldn't
>> exceed HVNETVSC_MAX_HEADER, or you need to update HVNETVSC_MAX_HEADER.
> Am I right in thinking this is adding an extra 96 unused bytes to the front
> of almost all skb just so that hyper-v can make its link level header
> contiguous with whatever follows (IP header ?).
>
> Doesn't sound ideal.

Actually it will probably add closer to 128 unused bytes per frame since
an skb allocation is rounded off to the nearest cache line in terms of size.

Is there a reason why this buffer info and descriptor like thing even
needs to be in the packet header? From what I can tell it was being
allocated separately before. Why not just maintain a ring of them if
you need to keep them around and want to avoid having to reallocate them
for the next transmit? Also I don't know if the layout matters but it
probably wouldn't hurt to go through with pahole and clean the
hv_netvsc_packet structure up. You could probably save yourself
something like 20+ bytes just doing that.

- Alex


\
 
 \ /
  Last update: 2015-09-16 19:41    [W:0.136 / U:0.052 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site