From mboxrd@z Thu Jan  1 00:00:00 1970
From: peter@hurleysoftware.com (Peter Hurley)
Date: Thu, 25 Sep 2014 07:31:32 -0400
Subject: [PATCH 09/16] tty: serial: 8250_dma: Add a TX trigger workaround
 for AM33xx
In-Reply-To: <20140925104208.GE19396@linutronix.de>
References: <2cf997d9-e7de-465b-915d-3a0a5e7b95f7@email.android.com>
 <54129F6C.7040508@linutronix.de>
 <20140912094010.GE28458@ci00147.xsens-tech.local>
 <5412C21A.8020203@linutronix.de>
 <20140912102816.GF28458@ci00147.xsens-tech.local>
 <541716DC.9020904@linutronix.de>
 <20140917102812.GA20438@ci00147.xsens-tech.local>
 <20140921204100.GA10111@linutronix.de> <5421A7F7.3070109@hurleysoftware.com>
 <20140924075346.GB19565@linutronix.de> <20140925104208.GE19396@linutronix.de>
Message-ID: <5423FD14.9060604@hurleysoftware.com>
To: linux-arm-kernel@lists.infradead.org
List-Id: linux-arm-kernel.lists.infradead.org

On 09/25/2014 06:42 AM, Sebastian Andrzej Siewior wrote:
> * Sebastian Andrzej Siewior | 2014-09-24 09:53:46 [+0200]:
> 
>> * Peter Hurley | 2014-09-23 13:03:51 [-0400]:
>>
>>> But DMA is cheating if the UART driver's tx_empty() method is saying the
>>> transmitter is empty while TX DMA is still running.
>> This shouldn't be the case. But I will check this once I able to.
> 
> I added 
> 
> |#define BOTH_EMPTY      (UART_LSR_TEMT | UART_LSR_THRE)
> |    trace_printk("delay <%d>\n", (lsr & BOTH_EMPTY) == BOTH_EMPTY ? 1 : 0);
> 
> in my set_termios() and the trace shows
> |              vi-949   [000] d...    70.477002: omap8250_restore_regs: delay <0>
> 
> so no, it does not wait until TX FIFO is empty. It looks like it uses
> TCSANOW instead of TCSADRAIN. And since this looks "legal" I will delay
> it until TX-DMA is complete because it is known to stall the DMA operation.

I just verified that GNU readline6 uses ioctl(TCSETSW, ...) to do the set_termios
(which is the ioctl that libc should use for tcsetattr(TCSADRAIN)).

Maybe this userspace is using a readline()-alike that has a bug by not using the
correct tcsetattr() action?
Or maybe the glibc-equivalent has the bug, and tcsetattr(TCSADRAIN) is not using
ioctl(TCSETSW)?

Regards,
Peter Hurley