linux-kernel - Re: [PATCH net-next v1 5/6] tg3: implementation of a non-NAPI mode

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]

Date:	Fri, 16 Dec 2011 20:50:42 +0100
From:	Eric Dumazet <eric.dumazet@...il.com>
To:	David Decotigny <decot@...glers.com>
Cc:	Matt Carlson <mcarlson@...adcom.com>,
	Michael Chan <mchan@...adcom.com>, netdev@...r.kernel.org,
	linux-kernel@...r.kernel.org,
	Javier Martinez Canillas <martinez.javier@...il.com>,
	Robin Getz <rgetz@...ckfin.uclinux.org>,
	Matt Mackall <mpm@...enic.com>,
	Tom Herbert <therbert@...gle.com>
Subject: Re: [PATCH net-next v1 5/6] tg3: implementation of a non-NAPI mode

Le vendredi 16 décembre 2011 à 10:19 -0800, David Decotigny a écrit :
> From: Tom Herbert <therbert@...gle.com>
> 
> The tg3 NIC has a hard limit of 511 descriptors for the receive ring.
> Under heavy load of small packets, this device receive queue may not
> be serviced fast enough to prevent packets drops.  This could be due
> to a variety of reasons such as lengthy processing delays of packets
> in the stack, softirqs being disabled too long, etc. If the driver is
> run in non-NAPI mode, the RX queue is serviced in the device
> interrupt, which is much less likely to be deferred for a substantial
> period of time.
> 
> There are some effects in not using NAPI that need to be considered.
> It does increase the chance of live-lock in interrupt handler,
> although since the tg3 does interrupt coalescing this is very unlikely
> to occur.  Also, more code is being run with interrupts disabled
> potentially deferring other hardware interrupts.  The amount of time
> spent in the interrupt handler should be minimized by dequeuing
> packets of the device queue and queuing them to a host queue as
> quickly as possible.
> 
> The default mode of operation remains NAPI and its performances are
> kept unchanged (code unchanged). Non-NAPI mode is enabled by
> commenting-out CONFIG_TIGON3_NAPI Kconfig parameter.
> 
> 

Oh well, thats ugly :(

I suspect this was only used with RPS/RFS ?

Or interrupts stick on a given cpu ?

Because with a default setup, and IRQ serviced by multiple cpus, you
endup with possible packet reorderings.

Packet1,2,3,4 handled by CPU0 : queued on netif_rx() queue.
EndOfInterrupt
Packet4,5,6,7 handled by CPU1 : queued on netif_rx() queue.
EndOfInterrupt

CPU0/CPU1 happily merge packets...



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/