linux-kernel - [PATCH] serial: pl011: Fix DMA ->flush

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>] [day] [month] [year] [list]

Message-Id: <20191118092547.32135-1-vincent.whitchurch@axis.com>
Date:   Mon, 18 Nov 2019 10:25:47 +0100
From:   Vincent Whitchurch <vincent.whitchurch@...s.com>
To:     linux@...linux.org.uk, gregkh@...uxfoundation.org
Cc:     jslaby@...e.com, linux-serial@...r.kernel.org,
        linux-kernel@...r.kernel.org, dmaengine@...r.kernel.org,
        Vincent Whitchurch <rabinv@...s.com>
Subject: [PATCH] serial: pl011: Fix DMA ->flush_buffer()

PL011's ->flush_buffer() implementation releases and reacquires the port
lock.  Due to a race condition here, data can end up being added to the
circular buffer but neither being discarded nor being sent out.  This
leads to, for example, tcdrain(2) waiting indefinitely.

Process A                       Process B

uart_flush_buffer()
 - acquire lock
 - circ_clear
 - pl011_flush_buffer()
 -- release lock
 -- dmaengine_terminate_all()

                                uart_write()
                                - acquire lock
                                - add chars to circ buffer
                                - start_tx()
                                -- start DMA
                                - release lock

 -- acquire lock
 -- turn off DMA
 -- release lock

                                // Data in circ buffer but DMA is off

According to the comment in the code, the releasing of the lock around
dmaengine_terminate_all() is to avoid a deadlock with the DMA engine
callback.  However, since the time this code was written, the DMA engine
API documentation seems to have been clarified to say that
dmaengine_terminate_all() (in the identically implemented but
differently named dmaengine_terminate_async() variant) does not wait for
any running complete callback to be completed and can even be called
from a complete callback.  So there is no possibility of deadlock if the
DMA engine driver implements this API correctly.

So we should be able to just remove this release and reacquire of the
lock to prevent the aforementioned race condition.

Signed-off-by: Vincent Whitchurch <vincent.whitchurch@...s.com>
---
 drivers/tty/serial/amba-pl011.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/drivers/tty/serial/amba-pl011.c b/drivers/tty/serial/amba-pl011.c
index 3a7d1a66f79c..b0b689546395 100644
--- a/drivers/tty/serial/amba-pl011.c
+++ b/drivers/tty/serial/amba-pl011.c
@@ -813,10 +813,8 @@ __acquires(&uap->port.lock)
 	if (!uap->using_tx_dma)
 		return;

-	/* Avoid deadlock with the DMA engine callback */
-	spin_unlock(&uap->port.lock);
-	dmaengine_terminate_all(uap->dmatx.chan);
-	spin_lock(&uap->port.lock);
+	dmaengine_terminate_async(uap->dmatx.chan);
+
 	if (uap->dmatx.queued) {
 		dma_unmap_sg(uap->dmatx.chan->device->dev, &uap->dmatx.sg, 1,
 			     DMA_TO_DEVICE);
-- 
2.20.0