net: thunderbolt: Stop using zero to mean no valid DMA mapping
authorMika Westerberg <mika.westerberg@linux.intel.com>
Fri, 24 Nov 2017 11:05:36 +0000 (14:05 +0300)
committerDavid S. Miller <davem@davemloft.net>
Sat, 25 Nov 2017 14:56:02 +0000 (23:56 +0900)
Commit 86dabda426ac ("net: thunderbolt: Clear finished Tx frame bus
address in tbnet_tx_callback()") fixed a DMA-API violation where the
driver called dma_unmap_page() in tbnet_free_buffers() for a bus address
that might already be unmapped. The fix was to zero out the bus address
of a frame in tbnet_tx_callback().

However, as pointed out by David Miller, zero might well be valid
mapping (at least in theory) so it is not good idea to use it here.

It turns out that we don't need the whole map/unmap dance for Tx buffers
at all. Instead we can map the buffers when they are initially allocated
and unmap them when the interface is brought down. In between we just
DMA sync the buffers for the CPU or device as needed.

Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/thunderbolt.c

index 228d4aa6d9ae3e08d688296cc4d8e81e6dcde81e..ca5e375de27c131534ce1bd2768b787be2fb908f 100644 (file)
@@ -335,7 +335,7 @@ static void tbnet_free_buffers(struct tbnet_ring *ring)
                if (ring->ring->is_tx) {
                        dir = DMA_TO_DEVICE;
                        order = 0;
-                       size = tbnet_frame_size(tf);
+                       size = TBNET_FRAME_SIZE;
                } else {
                        dir = DMA_FROM_DEVICE;
                        order = TBNET_RX_PAGE_ORDER;
@@ -512,6 +512,7 @@ err_free:
 static struct tbnet_frame *tbnet_get_tx_buffer(struct tbnet *net)
 {
        struct tbnet_ring *ring = &net->tx_ring;
+       struct device *dma_dev = tb_ring_dma_device(ring->ring);
        struct tbnet_frame *tf;
        unsigned int index;
 
@@ -522,7 +523,9 @@ static struct tbnet_frame *tbnet_get_tx_buffer(struct tbnet *net)
 
        tf = &ring->frames[index];
        tf->frame.size = 0;
-       tf->frame.buffer_phy = 0;
+
+       dma_sync_single_for_cpu(dma_dev, tf->frame.buffer_phy,
+                               tbnet_frame_size(tf), DMA_TO_DEVICE);
 
        return tf;
 }
@@ -531,13 +534,8 @@ static void tbnet_tx_callback(struct tb_ring *ring, struct ring_frame *frame,
                              bool canceled)
 {
        struct tbnet_frame *tf = container_of(frame, typeof(*tf), frame);
-       struct device *dma_dev = tb_ring_dma_device(ring);
        struct tbnet *net = netdev_priv(tf->dev);
 
-       dma_unmap_page(dma_dev, tf->frame.buffer_phy, tbnet_frame_size(tf),
-                      DMA_TO_DEVICE);
-       tf->frame.buffer_phy = 0;
-
        /* Return buffer to the ring */
        net->tx_ring.prod++;
 
@@ -548,10 +546,12 @@ static void tbnet_tx_callback(struct tb_ring *ring, struct ring_frame *frame,
 static int tbnet_alloc_tx_buffers(struct tbnet *net)
 {
        struct tbnet_ring *ring = &net->tx_ring;
+       struct device *dma_dev = tb_ring_dma_device(ring->ring);
        unsigned int i;
 
        for (i = 0; i < TBNET_RING_SIZE; i++) {
                struct tbnet_frame *tf = &ring->frames[i];
+               dma_addr_t dma_addr;
 
                tf->page = alloc_page(GFP_KERNEL);
                if (!tf->page) {
@@ -559,7 +559,17 @@ static int tbnet_alloc_tx_buffers(struct tbnet *net)
                        return -ENOMEM;
                }
 
+               dma_addr = dma_map_page(dma_dev, tf->page, 0, TBNET_FRAME_SIZE,
+                                       DMA_TO_DEVICE);
+               if (dma_mapping_error(dma_dev, dma_addr)) {
+                       __free_page(tf->page);
+                       tf->page = NULL;
+                       tbnet_free_buffers(ring);
+                       return -ENOMEM;
+               }
+
                tf->dev = net->dev;
+               tf->frame.buffer_phy = dma_addr;
                tf->frame.callback = tbnet_tx_callback;
                tf->frame.sof = TBIP_PDF_FRAME_START;
                tf->frame.eof = TBIP_PDF_FRAME_END;
@@ -881,19 +891,6 @@ static int tbnet_stop(struct net_device *dev)
        return 0;
 }
 
-static bool tbnet_xmit_map(struct device *dma_dev, struct tbnet_frame *tf)
-{
-       dma_addr_t dma_addr;
-
-       dma_addr = dma_map_page(dma_dev, tf->page, 0, tbnet_frame_size(tf),
-                               DMA_TO_DEVICE);
-       if (dma_mapping_error(dma_dev, dma_addr))
-               return false;
-
-       tf->frame.buffer_phy = dma_addr;
-       return true;
-}
-
 static bool tbnet_xmit_csum_and_map(struct tbnet *net, struct sk_buff *skb,
        struct tbnet_frame **frames, u32 frame_count)
 {
@@ -908,13 +905,14 @@ static bool tbnet_xmit_csum_and_map(struct tbnet *net, struct sk_buff *skb,
 
        if (skb->ip_summed != CHECKSUM_PARTIAL) {
                /* No need to calculate checksum so we just update the
-                * total frame count and map the frames for DMA.
+                * total frame count and sync the frames for DMA.
                 */
                for (i = 0; i < frame_count; i++) {
                        hdr = page_address(frames[i]->page);
                        hdr->frame_count = cpu_to_le32(frame_count);
-                       if (!tbnet_xmit_map(dma_dev, frames[i]))
-                               goto err_unmap;
+                       dma_sync_single_for_device(dma_dev,
+                               frames[i]->frame.buffer_phy,
+                               tbnet_frame_size(frames[i]), DMA_TO_DEVICE);
                }
 
                return true;
@@ -983,21 +981,14 @@ static bool tbnet_xmit_csum_and_map(struct tbnet *net, struct sk_buff *skb,
        *tucso = csum_fold(wsum);
 
        /* Checksum is finally calculated and we don't touch the memory
-        * anymore, so DMA map the frames now.
+        * anymore, so DMA sync the frames now.
         */
        for (i = 0; i < frame_count; i++) {
-               if (!tbnet_xmit_map(dma_dev, frames[i]))
-                       goto err_unmap;
+               dma_sync_single_for_device(dma_dev, frames[i]->frame.buffer_phy,
+                       tbnet_frame_size(frames[i]), DMA_TO_DEVICE);
        }
 
        return true;
-
-err_unmap:
-       while (i--)
-               dma_unmap_page(dma_dev, frames[i]->frame.buffer_phy,
-                              tbnet_frame_size(frames[i]), DMA_TO_DEVICE);
-
-       return false;
 }
 
 static void *tbnet_kmap_frag(struct sk_buff *skb, unsigned int frag_num,