Date: Tue, 1 Dec 2015 10:47:50 -0500 (EST)
From: Alan Stern <stern@rowland.harvard.edu>
To: Mathias Nyman <mathias.nyman@linux.intel.com>
cc: daniel@quora.org, <linux-usb@vger.kernel.org>,
        <linux-kernel@vger.kernel.org>
Subject: Re: [TESTPATCH] xhci: fix usb2 resume timing and races.
In-Reply-To: <1448896592-26421-1-git-send-email-mathias.nyman@linux.intel.com>
Message-ID: <Pine.LNX.4.44L0.1512011043220.1697-100000@iolanthe.rowland.org>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2339
Lines: 52

On Mon, 30 Nov 2015, Mathias Nyman wrote:

> usb2 ports need to signal resume for 20ms before moving to U0 state.
> Both device and host can initiate resume.
> 
> On host initated resume port is set to resume state, sleep 20ms,
> and finally set port to U0 state.

That's an odd approach.  The assumption in usbcore is that the HCD will 
not sleep here.

> On device initated resume a port status interrupt with a port in resume
> state in issued. The interrupt handler tags a resume_done[port]
> timestamp with current time + 20ms, and kick roothub timer.
> Root hub timer requests for port status, finds the port in resume state,
> checks if resume_done[port] timestamp passed, and set port to U0 state.

ehci-hcd does the same thing, except that it also uses this resume_done 
timestamp with host-initiated resumes.

> There are a few issues with this approach,
> 1. A host initated resume will also generate a resume event, the event
>    handler will find the port in resume state, believe it's a device
>    initated and act accordingly.
> 
> 2. A port status request might cut the 20ms resume signalling short if a
>    get_port_status request is handled during the 20ms host resume.
>    The port will be found in resume state. The timestamp is not set leading
>    to time_after_eq(jiffoes, timestamp) returning true, as timestamp = 0.
>    get_port_status will proceed with moving the port to U0.
> 
> 3. If an error, or anything else happends to the port during device
>    initated 20ms resume signalling it will leave all device resume
>    parameters hanging uncleared preventing further resume.
> 
> Fix this by using the existing resuming_ports bitfield to indicate if
> resume signalling timing is taken care of.
> Also check if the resume_done[port] is set  before using it in time
> comparison. Also clear out any resume signalling related variables if port
> is not in U0 or Resume state.

Wouldn't it be better to change the host-initiated resume mechanism to
be consisten with device-initiated resumes?  Or would that be too big a 
change for the time being?

Alan Stern

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