netdev - Re: [PATCH net 2/2] sfc: limit ARFS workitems in flight per channel

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

Message-Id: <20180413.121415.787716401343641542.davem@davemloft.net>
Date:   Fri, 13 Apr 2018 12:14:15 -0400 (EDT)
From:   David Miller <davem@...emloft.net>
To:     ecree@...arflare.com
Cc:     linux-net-drivers@...arflare.com, netdev@...r.kernel.org
Subject: Re: [PATCH net 2/2] sfc: limit ARFS workitems in flight per channel

From: Edward Cree <ecree@...arflare.com>
Date: Fri, 13 Apr 2018 16:59:07 +0100

> On 13/04/18 16:03, David Miller wrote:
>> Whilst you may not be able to program the filter into the hardware
>> synchronously, you should be able to allocate the ID and get all of
>> the software state setup.
> That's what we were doing before commit 3af0f34290f6 ("sfc: replace
>  asynchronous filter operations"), and (as mentioned in that commit)
>  that leads to (or at least the scheme we used had) race conditions
>  which I could not see a way to fix.  If the hardware resets (and
>  thus forgets all its filters) after we've done the software state
>  setup but before we reach the point of finalising the software state
>  after the hardware operation, we don't know what operations we need
>  to do to re-apply the software state to the hardware, because we
>  don't know whether the reset happened before or after the hardware
>  operation.

When an entry is successfully programmed into the chip, you update
the software state.

When the chip resets, you clear all of those state booleans to false.

Indeed, you would have to synchronize these things somehow.

Is the issue that you learn about the hardware reset asynchronously,
and therefore cannot determine if filter insertion programming
happened afterwards and thus is still in the chip?

You must have a table of all the entries, so that you can reprogram
the hardware should it reset.  Or do you not handle things that way
and it's a lossy system?

> Well, the alternative, even if the software state setup part _could_
>  be made synchronous, is to allow a potentially unbounded queue for
>  the hardware update part (I think there are even still cases in
>  which the exponential growth pathology is possible), causing the
>  filter insertions to be delayed an arbitrarily long time.  Either
>  the flow is still going by that time (in which case the backlog
>  limit approach will get a new ndo_rx_flow_steer request and insert
>  the filter too) or it isn't, in which case getting round to it
>  eventually is no better than dropping it immediately.  In fact it's
>  worse because now you waste time inserting a useless filter which
>  delays new requests even more.
> Besides, I'm fairly confident that the only cases in which you'll
>  even come close to hitting the limit are ones where ARFS wouldn't
>  do you much good anyway, such as:
> * Misconfigured interrupt affinities where ARFS is entirely pointless
> * Many short-lived flows (which greatly diminish the utility of ARFS)
> 
> So for multiple reasons, hitting the limit won't actually make
>  performance worse, although it will often be a sign that performance
>  will be bad for other reasons.

Understood, thanks for explaining.

Please respin your series with the updates you talked about and I'll
apply it.

But generally we do have this issue with various kinds of
configuration programming and async vs. sync.