OSDN Git Service

serial: sh-sci: Fix TX DMA buffer flushing and workqueue races
authorGeert Uytterhoeven <geert+renesas@glider.be>
Mon, 24 Jun 2019 12:35:39 +0000 (14:35 +0200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 3 Jul 2019 17:31:30 +0000 (19:31 +0200)
When uart_flush_buffer() is called, the .flush_buffer() callback zeroes
the tx_dma_len field.  This may race with the work queue function
handling transmit DMA requests:

  1. If the buffer is flushed before the first DMA API call,
     dmaengine_prep_slave_single() may be called with a zero length,
     causing the DMA request to never complete, leading to messages
     like:

        rcar-dmac e7300000.dma-controller: Channel Address Error happen

     and, with debug enabled:

sh-sci e6e88000.serial: sci_dma_tx_work_fn: ffff800639b55000: 0...0, cookie 126

     and DMA timeouts.

  2. If the buffer is flushed after the first DMA API call, but before
     the second, dma_sync_single_for_device() may be called with a zero
     length, causing the transmit data not to be flushed to RAM, and
     leading to stale data being output.

Fix this by:
  1. Letting sci_dma_tx_work_fn() return immediately if the transmit
     buffer is empty,
  2. Extending the critical section to cover all DMA preparational work,
     so tx_dma_len stays consistent for all of it,
  3. Using local copies of circ_buf.head and circ_buf.tail, to make sure
     they match the actual operation above.

Reported-by: Eugeniu Rosca <erosca@de.adit-jv.com>
Suggested-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Reviewed-by: Eugeniu Rosca <erosca@de.adit-jv.com>
Tested-by: Eugeniu Rosca <erosca@de.adit-jv.com>
Link: https://lore.kernel.org/r/20190624123540.20629-2-geert+renesas@glider.be
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/tty/serial/sh-sci.c

index abc7057..d4504da 100644 (file)
@@ -1398,6 +1398,7 @@ static void sci_dma_tx_work_fn(struct work_struct *work)
        struct circ_buf *xmit = &port->state->xmit;
        unsigned long flags;
        dma_addr_t buf;
+       int head, tail;
 
        /*
         * DMA is idle now.
@@ -1407,16 +1408,23 @@ static void sci_dma_tx_work_fn(struct work_struct *work)
         * consistent xmit buffer state.
         */
        spin_lock_irq(&port->lock);
-       buf = s->tx_dma_addr + (xmit->tail & (UART_XMIT_SIZE - 1));
+       head = xmit->head;
+       tail = xmit->tail;
+       buf = s->tx_dma_addr + (tail & (UART_XMIT_SIZE - 1));
        s->tx_dma_len = min_t(unsigned int,
-               CIRC_CNT(xmit->head, xmit->tail, UART_XMIT_SIZE),
-               CIRC_CNT_TO_END(xmit->head, xmit->tail, UART_XMIT_SIZE));
-       spin_unlock_irq(&port->lock);
+               CIRC_CNT(head, tail, UART_XMIT_SIZE),
+               CIRC_CNT_TO_END(head, tail, UART_XMIT_SIZE));
+       if (!s->tx_dma_len) {
+               /* Transmit buffer has been flushed */
+               spin_unlock_irq(&port->lock);
+               return;
+       }
 
        desc = dmaengine_prep_slave_single(chan, buf, s->tx_dma_len,
                                           DMA_MEM_TO_DEV,
                                           DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
        if (!desc) {
+               spin_unlock_irq(&port->lock);
                dev_warn(port->dev, "Failed preparing Tx DMA descriptor\n");
                goto switch_to_pio;
        }
@@ -1424,18 +1432,18 @@ static void sci_dma_tx_work_fn(struct work_struct *work)
        dma_sync_single_for_device(chan->device->dev, buf, s->tx_dma_len,
                                   DMA_TO_DEVICE);
 
-       spin_lock_irq(&port->lock);
        desc->callback = sci_dma_tx_complete;
        desc->callback_param = s;
-       spin_unlock_irq(&port->lock);
        s->cookie_tx = dmaengine_submit(desc);
        if (dma_submit_error(s->cookie_tx)) {
+               spin_unlock_irq(&port->lock);
                dev_warn(port->dev, "Failed submitting Tx DMA descriptor\n");
                goto switch_to_pio;
        }
 
+       spin_unlock_irq(&port->lock);
        dev_dbg(port->dev, "%s: %p: %d...%d, cookie %d\n",
-               __func__, xmit->buf, xmit->tail, xmit->head, s->cookie_tx);
+               __func__, xmit->buf, tail, head, s->cookie_tx);
 
        dma_async_issue_pending(chan);
        return;