Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Fri, 16 Feb 2018 13:29:01 -0500 (EST)
From:   Alan Stern <stern@rowland.harvard.edu>
To:     Sebastian Andrzej Siewior <bigeasy@linutronix.de>
cc:     Mauro Carvalho Chehab <mchehab@s-opensource.com>,
        Frederic Weisbecker <frederic@kernel.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Peter Zijlstra <peterz@infradead.org>,
        Thomas Gleixner <tglx@linutronix.de>,
        <linux-usb@vger.kernel.org>
Subject: Re: [RFC PATCH] usb: hcd: complete URBs in threaded-IRQ context
 instead of tasklet
In-Reply-To: <20180216170450.yl5owfphuvltstnt@breakpoint.cc>
Message-ID: <Pine.LNX.4.44L0.1802161325050.1402-100000@iolanthe.rowland.org>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Fri, 16 Feb 2018, Sebastian Andrzej Siewior wrote:

> I've been going over Frederic's softirq patches and it seems that there
> were two problems. One was network related, the other was Mauro's USB
> dvb-[stc] device which was not able to stream properly over time.
> 
> Here is an attempt to let the URB complete in the threaded handler
> instead of going to the tasklet. In case the system is SMP then the
> patch [0] would be required in order to have the IRQ and its thread on
> the same CPU.
> 
> Mauro, would you mind giving it a shot?
> 
> [0] genirq: Let irq thread follow the effective hard irq affinity
>     https://git.kernel.org/tip/cbf8699996a6e7f2f674b3a2a4cef9f666ff613e
> 
> ---
> 
> The urb->complete callback was initially invoked in IRQ context after
> the HCD dropped its lock because the callback could re-queue the URB
> again. Later this completion was deferred to the tasklet allowing the
> HCD hold the lock. Also the BH handler can be interrupted by the IRQ
> handler adding more "completed" requests to its queue.
> While this batching is good in general, the softirq defers its doing for
> short period of time if it is running constantly without a break. This
> breaks some use cases where constant USB throughput is required.
> As an alternative approach to tasklet handling, I defer the URB
> completion to the HCD's threaded handler. There are two lists for
> "high-prio" proccessing and lower priority (to mimic current behaviour).
> The URBs in the high-priority list are always preffered over the URBs
> in the low-priority list.

We originally used tasklets because we didn't want to incur the delays 
associated with running in a process context.  It seems odd to be 
reversing that decision now.

> The URBs from the root-hub never create an interrupt so I currently
> process them in a workqueue (I'm not sure if an URB-enqueue in the
> completion handler would break something).

It worked okay before we changed over to using tasklets.

Alan Stern