[IB] Fix MAD layer DMA mappings to avoid touching data buffer once mapped

The MAD layer was violating the DMA API by touching data buffers used
for sends after the DMA mapping was done.  This causes problems on
non-cache-coherent architectures, because the device doing DMA won't
see updates to the payload buffers that exist only in the CPU cache.

Fix this by having all MAD consumers use ib_create_send_mad() to
allocate their send buffers, and moving the DMA mapping into the MAD
layer so it can be done just before calling send (and after any
modifications of the send buffer by the MAD layer).

Tested on a non-cache-coherent PowerPC 440SPe system.

Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

1 files changed, 9 insertions, 63 deletions

diff --git a/drivers/infiniband/hw/mthca/mthca_mad.c b/drivers/infiniband/hw/mthca/mthca_mad.c
index 9804174f7f3c..8561b297a19b 100644
--- a/drivers/infiniband/hw/mthca/mthca_mad.c
+++ b/drivers/infiniband/hw/mthca/mthca_mad.c
@@ -46,11 +46,6 @@ enum {
 	MTHCA_VENDOR_CLASS2 = 0xa
 };
 
-struct mthca_trap_mad {
-	struct ib_mad *mad;
-	DECLARE_PCI_UNMAP_ADDR(mapping)
-};
-
 static void update_sm_ah(struct mthca_dev *dev,
 			 u8 port_num, u16 lid, u8 sl)
 {
@@ -116,49 +111,14 @@ static void forward_trap(struct mthca_dev *dev,
 			 struct ib_mad *mad)
 {
 	int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
-	struct mthca_trap_mad *tmad;
-	struct ib_sge      gather_list;
-	struct ib_send_wr *bad_wr, wr = {
-		.opcode      = IB_WR_SEND,
-		.sg_list     = &gather_list,
-		.num_sge     = 1,
-		.send_flags  = IB_SEND_SIGNALED,
-		.wr	     = {
-			 .ud = {
-				 .remote_qpn  = qpn,
-				 .remote_qkey = qpn ? IB_QP1_QKEY : 0,
-				 .timeout_ms  = 0
-			 }
-		 }
-	};
+	struct ib_mad_send_buf *send_buf;
 	struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
 	int ret;
 	unsigned long flags;
 
 	if (agent) {
-		tmad = kmalloc(sizeof *tmad, GFP_KERNEL);
-		if (!tmad)
-			return;
-
-		tmad->mad = kmalloc(sizeof *tmad->mad, GFP_KERNEL);
-		if (!tmad->mad) {
-			kfree(tmad);
-			return;
-		}
-
-		memcpy(tmad->mad, mad, sizeof *mad);
-
-		wr.wr.ud.mad_hdr = &tmad->mad->mad_hdr;
-		wr.wr_id         = (unsigned long) tmad;
-
-		gather_list.addr   = dma_map_single(agent->device->dma_device,
-						    tmad->mad,
-						    sizeof *tmad->mad,
-						    DMA_TO_DEVICE);
-		gather_list.length = sizeof *tmad->mad;
-		gather_list.lkey   = to_mpd(agent->qp->pd)->ntmr.ibmr.lkey;
-		pci_unmap_addr_set(tmad, mapping, gather_list.addr);
-
+		send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
+					      IB_MGMT_MAD_DATA, GFP_ATOMIC);
 		/*
 		 * We rely here on the fact that MLX QPs don't use the
 		 * address handle after the send is posted (this is
@@ -166,21 +126,15 @@ static void forward_trap(struct mthca_dev *dev,
 		 * it's OK for our devices).
 		 */
 		spin_lock_irqsave(&dev->sm_lock, flags);
-		wr.wr.ud.ah      = dev->sm_ah[port_num - 1];
-		if (wr.wr.ud.ah)
-			ret = ib_post_send_mad(agent, &wr, &bad_wr);
+		memcpy(send_buf->mad, mad, sizeof *mad);
+		if ((send_buf->ah = dev->sm_ah[port_num - 1]))
+			ret = ib_post_send_mad(send_buf, NULL);
 		else
 			ret = -EINVAL;
 		spin_unlock_irqrestore(&dev->sm_lock, flags);
 
-		if (ret) {
-			dma_unmap_single(agent->device->dma_device,
-					 pci_unmap_addr(tmad, mapping),
-					 sizeof *tmad->mad,
-					 DMA_TO_DEVICE);
-			kfree(tmad->mad);
-			kfree(tmad);
-		}
+		if (ret)
+			ib_free_send_mad(send_buf);
 	}
 }
 
@@ -267,15 +221,7 @@ int mthca_process_mad(struct ib_device *ibdev,
 static void send_handler(struct ib_mad_agent *agent,
 			 struct ib_mad_send_wc *mad_send_wc)
 {
-	struct mthca_trap_mad *tmad =
-		(void *) (unsigned long) mad_send_wc->wr_id;
-
-	dma_unmap_single(agent->device->dma_device,
-			 pci_unmap_addr(tmad, mapping),
-			 sizeof *tmad->mad,
-			 DMA_TO_DEVICE);
-	kfree(tmad->mad);
-	kfree(tmad);
+	ib_free_send_mad(mad_send_wc->send_buf);
 }
 
 int mthca_create_agents(struct mthca_dev *dev)