From mboxrd@z Thu Jan  1 00:00:00 1970
From: jamie@shareable.org (Jamie Lokier)
Date: Wed, 10 Feb 2010 01:16:37 +0000
Subject: [PATCH] amba-pl011: support hardware flow control
In-Reply-To: <20100209221449.GA23083@n2100.arm.linux.org.uk>
References: <1265636510-27679-1-git-send-email-rabin.vincent@stericsson.com>
	<20100208135124.GA26608@n2100.arm.linux.org.uk>
	<20100209143055.GA6015@bnru01.bnr.st.com>
	<20100209221449.GA23083@n2100.arm.linux.org.uk>
Message-ID: <20100210011637.GA8683@shareable.org>
To: linux-arm-kernel@lists.infradead.org
List-Id: linux-arm-kernel.lists.infradead.org

Russell King - ARM Linux wrote:
> This means that when the kernel's buffers fill up, the kernel calls
> down to the serial core layer to throttle the input.  This then
> calls into the set_mctrl function to de-assert RTS.  However, because
> the hardware ignores the requested software state, the RTS signal
> is not de-asserted, and the remote end continues sending data.
> 
> So, enabling hardware auto-RTS is bad news - you will lose data if
> the application stops reading data.

Surely the driver can just stop reading from the UART when the
kernel's buffers are full and hardware-RTS is enabled.  Then the
hardware will deassert RTS itself.

The problem with relying on the kernel to deassert RTS is that it's
too slow when something disables interrupts for a few ms, and you get
lost data that way instead.

The ideal combination may be a bit of both:

    1. When kernel deasserts RTS, do that and disable hardware-RTS
       so it really is deasserted.

    2. When kernel asserts RTS with mctrl, enable hardware-RTS
       so what's output depends on the receive FIFO state.

That should solve both types of overrun.

-- Jamie