netdev - Re: [PATCH net-next] ibmvnic: Defer tx completion processing using a wait queue

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-Id: <20160412.211202.1299929008077475122.davem@davemloft.net>
Date:	Tue, 12 Apr 2016 21:12:02 -0400 (EDT)
From:	David Miller <davem@...emloft.net>
To:	jallen@...ux.vnet.ibm.com
Cc:	eric.dumazet@...il.com, tlfalcon@...ux.vnet.ibm.com,
	netdev@...r.kernel.org, linuxppc-dev@...ts.ozlabs.org
Subject: Re: [PATCH net-next] ibmvnic: Defer tx completion processing using
 a wait queue

From: John Allen <jallen@...ux.vnet.ibm.com>
Date: Tue, 12 Apr 2016 16:00:23 -0500

> On 04/12/2016 03:12 PM, Eric Dumazet wrote:
>> On Tue, 2016-04-12 at 14:38 -0500, John Allen wrote:
>>> Moves tx completion processing out of interrupt context, deferring work
>>> using a wait queue. With this work now deferred, we must account for the
>>> possibility that skbs can be sent faster than we can process completion
>>> requests in which case the tx buffer will overflow. If the tx buffer is
>>> full, ibmvnic_xmit will return NETDEV_TX_BUSY and stop the current tx
>>> queue. Subsequently, the queue will be restarted in ibmvnic_complete_tx
>>> when all pending tx completion requests have been cleared.
>> 
>> 1) Why is this needed ?
> 
> In the current ibmvnic implementation, tx completion processing is done in
> interrupt context. Depending on the load, this can block further
> interrupts for a long time. This patch just creates a bottom half so that
> when a tx completion interrupt comes in, we can defer the majority of the
> work and exit interrupt context quickly.

You should use NAPI polling for this, not your own invented
mechanism.