Skip to contentSkip to conversation

NTB: ntb_transport: Ensure the destination buffer is mapped for TX DMA

PATCH In review Logan Gunthorpe 4 participants 5 messages Jan 19, 2019 – Feb 15, 2019 NTB driver core

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 NTB: ntb_transport: Ensure the destination buffer is mapped for TX DMA

Presently, when ntb_transport is used with DMA and the IOMMU turned on, it fails with errors from the IOMMU such as:

DMAR: DRHD: handling fault status reg 202
DMAR: [DMA Write] Request device [00:04.0] fault addr
381fc0340000 [fault reason 05] PTE Write access is not set

This is because ntb_transport does not map the BAR space with the IOMMU.

To fix this, we map the entire MW region for each QP after we assign the DMA channel. This prevents needing an extra DMA map in the fast path.

---
 drivers/ntb/ntb_transport.c | 28 ++++++++++++++++++++++++++--
 1 file changed, 26 insertions(+), 2 deletions(-)
drivers/ntb/ntb_transport.c +26 -2
index 3bfdb4562408..526b65afc16a 100644
@@ -144,7 +144,9 @@ struct ntb_transport_qp {
 	struct list_head tx_free_q;
 	spinlock_t ntb_tx_free_q_lock;
 	void __iomem *tx_mw;
-	dma_addr_t tx_mw_phys;
+	phys_addr_t tx_mw_phys;
+	size_t tx_mw_size;
+	dma_addr_t tx_mw_dma_addr;
 	unsigned int tx_index;
 	unsigned int tx_max_entry;
 	unsigned int tx_max_frame;
@@ -1049,6 +1051,7 @@ static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
 	tx_size = (unsigned int)mw_size / num_qps_mw;
 	qp_offset = tx_size * (qp_num / mw_count);
 
+	qp->tx_mw_size = tx_size;
 	qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
 	if (!qp->tx_mw)
 		return -EINVAL;
@@ -1644,7 +1647,7 @@ static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
 	dma_cookie_t cookie;
 
 	device = chan->device;
-	dest = qp->tx_mw_phys + qp->tx_max_frame * entry->tx_index;
+	dest = qp->tx_mw_dma_addr + qp->tx_max_frame * entry->tx_index;
 	buff_off = (size_t)buf & ~PAGE_MASK;
 	dest_off = (size_t)dest & ~PAGE_MASK;
 
@@ -1863,6 +1866,18 @@ ntb_transport_create_queue(void *data, struct device *client_dev,
 		qp->rx_dma_chan = NULL;
 	}
 
+	if (qp->tx_dma_chan) {
+		qp->tx_mw_dma_addr =
+			dma_map_resource(qp->tx_dma_chan->device->dev,
+					 qp->tx_mw_phys, qp->tx_mw_size,
+					 DMA_FROM_DEVICE, 0);
+		if (dma_mapping_error(qp->tx_dma_chan->device->dev,
+				      qp->tx_mw_dma_addr)) {
+			qp->tx_mw_dma_addr = 0;
+			goto err1;
+		}
+	}
+
 	dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
 		qp->tx_dma_chan ? "DMA" : "CPU");
 
@@ -1904,6 +1919,10 @@ ntb_transport_create_queue(void *data, struct device *client_dev,
 	qp->rx_alloc_entry = 0;
 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
 		kfree(entry);
+	if (qp->tx_mw_dma_addr)
+		dma_unmap_resource(qp->tx_dma_chan->device->dev,
+				   qp->tx_mw_dma_addr, qp->tx_mw_size,
+				   DMA_FROM_DEVICE, 0);
 	if (qp->tx_dma_chan)
 		dma_release_channel(qp->tx_dma_chan);
 	if (qp->rx_dma_chan)
@@ -1945,6 +1964,11 @@ void ntb_transport_free_queue(struct ntb_transport_qp *qp)
 		 */
 		dma_sync_wait(chan, qp->last_cookie);
 		dmaengine_terminate_all(chan);
+
+		dma_unmap_resource(chan->device->dev,
+				   qp->tx_mw_dma_addr, qp->tx_mw_size,
+				   DMA_FROM_DEVICE, 0);
+
 		dma_release_channel(chan);
 	}
 
Signature · 1 line
2.19.0
maintainer Reviewed-by on v1 1/1
#1 lore raw
Quoted 13 lines · Logan Gunthorpe

Presently, when ntb_transport is used with DMA and the IOMMU turned on, it fails with errors from the IOMMU such as:

DMAR: DRHD: handling fault status reg 202
DMAR: [DMA Write] Request device [00:04.0] fault addr
381fc0340000 [fault reason 05] PTE Write access is not set

This is because ntb_transport does not map the BAR space with the IOMMU.

To fix this, we map the entire MW region for each QP after we assign the DMA channel. This prevents needing an extra DMA map in the fast path.

Quoted 5 lines

Nice! I actually never encountered this on the Intel NTB with IOMMU on. It also could be that the Intel BIOS already took care of it for all embedded device BARs on the uncore. Nevertheless it's needed. Thanks!

Quoted 79 lines
---
 drivers/ntb/ntb_transport.c | 28 ++++++++++++++++++++++++++--
 1 file changed, 26 insertions(+), 2 deletions(-)
drivers/ntb/ntb_transport.c +26 -2
index 3bfdb4562408..526b65afc16a 100644
@@ -144,7 +144,9 @@ struct ntb_transport_qp {
 	struct list_head tx_free_q;
 	spinlock_t ntb_tx_free_q_lock;
 	void __iomem *tx_mw;
-	dma_addr_t tx_mw_phys;
+	phys_addr_t tx_mw_phys;
+	size_t tx_mw_size;
+	dma_addr_t tx_mw_dma_addr;
 	unsigned int tx_index;
 	unsigned int tx_max_entry;
 	unsigned int tx_max_frame;
@@ -1049,6 +1051,7 @@ static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
 	tx_size = (unsigned int)mw_size / num_qps_mw;
 	qp_offset = tx_size * (qp_num / mw_count);
 
+	qp->tx_mw_size = tx_size;
 	qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
 	if (!qp->tx_mw)
 		return -EINVAL;
@@ -1644,7 +1647,7 @@ static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
 	dma_cookie_t cookie;
 
 	device = chan->device;
-	dest = qp->tx_mw_phys + qp->tx_max_frame * entry->tx_index;
+	dest = qp->tx_mw_dma_addr + qp->tx_max_frame * entry->tx_index;
 	buff_off = (size_t)buf & ~PAGE_MASK;
 	dest_off = (size_t)dest & ~PAGE_MASK;
 
@@ -1863,6 +1866,18 @@ ntb_transport_create_queue(void *data, struct device *client_dev,
 		qp->rx_dma_chan = NULL;
 	}
 
+	if (qp->tx_dma_chan) {
+		qp->tx_mw_dma_addr =
+			dma_map_resource(qp->tx_dma_chan->device->dev,
+					 qp->tx_mw_phys, qp->tx_mw_size,
+					 DMA_FROM_DEVICE, 0);
+		if (dma_mapping_error(qp->tx_dma_chan->device->dev,
+				      qp->tx_mw_dma_addr)) {
+			qp->tx_mw_dma_addr = 0;
+			goto err1;
+		}
+	}
+
 	dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
 		qp->tx_dma_chan ? "DMA" : "CPU");
 
@@ -1904,6 +1919,10 @@ ntb_transport_create_queue(void *data, struct device *client_dev,
 	qp->rx_alloc_entry = 0;
 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
 		kfree(entry);
+	if (qp->tx_mw_dma_addr)
+		dma_unmap_resource(qp->tx_dma_chan->device->dev,
+				   qp->tx_mw_dma_addr, qp->tx_mw_size,
+				   DMA_FROM_DEVICE, 0);
 	if (qp->tx_dma_chan)
 		dma_release_channel(qp->tx_dma_chan);
 	if (qp->rx_dma_chan)
@@ -1945,6 +1964,11 @@ void ntb_transport_free_queue(struct ntb_transport_qp *qp)
 		 */
 		dma_sync_wait(chan, qp->last_cookie);
 		dmaengine_terminate_all(chan);
+
+		dma_unmap_resource(chan->device->dev,
+				   qp->tx_mw_dma_addr, qp->tx_mw_size,
+				   DMA_FROM_DEVICE, 0);
+
 		dma_release_channel(chan);
 	}
 
maintainer
#2 lore raw
Quoted 22 lines · Dave Jiang
Quoted 13 lines · Logan Gunthorpe

Presently, when ntb_transport is used with DMA and the IOMMU turned on, it fails with errors from the IOMMU such as:

DMAR: DRHD: handling fault status reg 202
DMAR: [DMA Write] Request device [00:04.0] fault addr
381fc0340000 [fault reason 05] PTE Write access is not set

This is because ntb_transport does not map the BAR space with the IOMMU.

To fix this, we map the entire MW region for each QP after we assign the DMA channel. This prevents needing an extra DMA map in the fast path.

Quoted 5 lines
Quoted 5 lines

Nice! I actually never encountered this on the Intel NTB with IOMMU on. It also could be that the Intel BIOS already took care of it for all embedded device BARs on the uncore. Nevertheless it's needed. Thanks!

Added to the ntb branch, thanks!

Quoted 80 lines
Quoted 80 lines
---
 drivers/ntb/ntb_transport.c | 28 ++++++++++++++++++++++++++--
 1 file changed, 26 insertions(+), 2 deletions(-)
drivers/ntb/ntb_transport.c +26 -2
index 3bfdb4562408..526b65afc16a 100644
@@ -144,7 +144,9 @@ struct ntb_transport_qp {
 	struct list_head tx_free_q;
 	spinlock_t ntb_tx_free_q_lock;
 	void __iomem *tx_mw;
-	dma_addr_t tx_mw_phys;
+	phys_addr_t tx_mw_phys;
+	size_t tx_mw_size;
+	dma_addr_t tx_mw_dma_addr;
 	unsigned int tx_index;
 	unsigned int tx_max_entry;
 	unsigned int tx_max_frame;
@@ -1049,6 +1051,7 @@ static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
 	tx_size = (unsigned int)mw_size / num_qps_mw;
 	qp_offset = tx_size * (qp_num / mw_count);
 
+	qp->tx_mw_size = tx_size;
 	qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
 	if (!qp->tx_mw)
 		return -EINVAL;
@@ -1644,7 +1647,7 @@ static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
 	dma_cookie_t cookie;
 
 	device = chan->device;
-	dest = qp->tx_mw_phys + qp->tx_max_frame * entry->tx_index;
+	dest = qp->tx_mw_dma_addr + qp->tx_max_frame * entry->tx_index;
 	buff_off = (size_t)buf & ~PAGE_MASK;
 	dest_off = (size_t)dest & ~PAGE_MASK;
 
@@ -1863,6 +1866,18 @@ ntb_transport_create_queue(void *data, struct device *client_dev,
 		qp->rx_dma_chan = NULL;
 	}
 
+	if (qp->tx_dma_chan) {
+		qp->tx_mw_dma_addr =
+			dma_map_resource(qp->tx_dma_chan->device->dev,
+					 qp->tx_mw_phys, qp->tx_mw_size,
+					 DMA_FROM_DEVICE, 0);
+		if (dma_mapping_error(qp->tx_dma_chan->device->dev,
+				      qp->tx_mw_dma_addr)) {
+			qp->tx_mw_dma_addr = 0;
+			goto err1;
+		}
+	}
+
 	dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
 		qp->tx_dma_chan ? "DMA" : "CPU");
 
@@ -1904,6 +1919,10 @@ ntb_transport_create_queue(void *data, struct device *client_dev,
 	qp->rx_alloc_entry = 0;
 	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
 		kfree(entry);
+	if (qp->tx_mw_dma_addr)
+		dma_unmap_resource(qp->tx_dma_chan->device->dev,
+				   qp->tx_mw_dma_addr, qp->tx_mw_size,
+				   DMA_FROM_DEVICE, 0);
 	if (qp->tx_dma_chan)
 		dma_release_channel(qp->tx_dma_chan);
 	if (qp->rx_dma_chan)
@@ -1945,6 +1964,11 @@ void ntb_transport_free_queue(struct ntb_transport_qp *qp)
 		 */
 		dma_sync_wait(chan, qp->last_cookie);
 		dmaengine_terminate_all(chan);
+
+		dma_unmap_resource(chan->device->dev,
+				   qp->tx_mw_dma_addr, qp->tx_mw_size,
+				   DMA_FROM_DEVICE, 0);
+
 		dma_release_channel(chan);
 	}
 
#3 lore raw
In reply to Logan Gunthorpe

Hi Logan ,
I took a look on the dma_map_resource() you used in this patch , looks like ops->map_resource not implemented for intel
if I see it right, this will not fix the issue for Intel NTB .

Please let me know if I am right

Thanks.
Leonid Ravich

Quoted 109 lines · Logan Gunthorpe

Presently, when ntb_transport is used with DMA and the IOMMU turned on, it fails with errors from the IOMMU such as:

DMAR: DRHD: handling fault status reg 202
DMAR: [DMA Write] Request device [00:04.0] fault addr
381fc0340000 [fault reason 05] PTE Write access is not set

This is because ntb_transport does not map the BAR space with the IOMMU.

To fix this, we map the entire MW region for each QP after we assign the DMA channel. This prevents needing an extra DMA map in the fast path.

Signed-off-by
<https://lore.kernel.org/linux-pci/[email protected]/Signed-off-by>:
Logan Gunthorpe <[email protected] <javascript:>>
Cc: Jon Mason <[email protected] <javascript:>>
Cc: Dave Jiang <[email protected] <javascript:>>
Cc: Allen Hubbe <[email protected] <javascript:>>
---

 drivers/ntb/ntb_transport.c | 28 ++++++++++++++++++++++++++-- 
 1 file changed, 26 insertions(+), 2 deletions(-) 
drivers/ntb/ntb_transport.c +4 -1
index 3bfdb4562408..526b65afc16a 100644 
@@ -144,7 +144,9 @@ struct ntb_transport_qp { 
         struct list_head tx_free_q; 
         spinlock_t ntb_tx_free_q_lock; 
         void __iomem *tx_mw; 
-        dma_addr_t tx_mw_phys; 
+        phys_addr_t tx_mw_phys; 
+        size_t tx_mw_size; 
+        dma_addr_t tx_mw_dma_addr; 
         unsigned int tx_index; 
         unsigned int tx_max_entry; 
         unsigned int tx_max_frame; 
@@ -1049,6 +1051,7 @@ static int ntb_transport_init_queue(struct 
 ntb_transport_ctx *nt, 
         tx_size = (unsigned int)mw_size / num_qps_mw; 
         qp_offset = tx_size * (qp_num / mw_count); 
 
+        qp->tx_mw_size = tx_size; 
         qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset; 
         if (!qp->tx_mw) 
                 return -EINVAL; 
drivers/ntb/ntb_transport.c +17 -1
@@ -1644,7 +1647,7 @@ static int ntb_async_tx_submit(struct 
 ntb_transport_qp *qp, 
         dma_cookie_t cookie; 
 
         device = chan->device; 
-        dest = qp->tx_mw_phys + qp->tx_max_frame * entry->tx_index; 
+        dest = qp->tx_mw_dma_addr + qp->tx_max_frame * entry->tx_index; 
         buff_off = (size_t)buf & ~PAGE_MASK; 
         dest_off = (size_t)dest & ~PAGE_MASK; 
@@ -1863,6 +1866,18 @@ ntb_transport_create_queue(void *data, struct 
 device *client_dev, 
                 qp->rx_dma_chan = NULL; 
         } 
 
+        if (qp->tx_dma_chan) { 
+                qp->tx_mw_dma_addr = 
+                        dma_map_resource(qp->tx_dma_chan->device->dev, 
+                                         qp->tx_mw_phys, qp->tx_mw_size, 
+                                         DMA_FROM_DEVICE, 0); 
+                if (dma_mapping_error(qp->tx_dma_chan->device->dev, 
+                                      qp->tx_mw_dma_addr)) { 
+                        qp->tx_mw_dma_addr = 0; 
+                        goto err1; 
+                } 
+        } 
+ 
         dev_dbg(&pdev->dev, "Using %s memcpy for TX\n", 
                 qp->tx_dma_chan ? "DMA" : "CPU"); 
@@ -1904,6 +1919,10 @@ ntb_transport_create_queue(void *data, struct 
 device *client_dev, 
         qp->rx_alloc_entry = 0; 
         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q))) 
                 kfree(entry); 
+        if (qp->tx_mw_dma_addr) 
+                dma_unmap_resource(qp->tx_dma_chan->device->dev, 
+                                   qp->tx_mw_dma_addr, qp->tx_mw_size, 
+                                   DMA_FROM_DEVICE, 0); 
         if (qp->tx_dma_chan) 
                 dma_release_channel(qp->tx_dma_chan); 
         if (qp->rx_dma_chan) 
drivers/ntb/ntb_transport.c +5 -0
@@ -1945,6 +1964,11 @@ void ntb_transport_free_queue(struct 
 ntb_transport_qp *qp) 
                  */ 
                 dma_sync_wait(chan, qp->last_cookie); 
                 dmaengine_terminate_all(chan); 
+ 
+                dma_unmap_resource(chan->device->dev, 
+                                   qp->tx_mw_dma_addr, qp->tx_mw_size, 
+                                   DMA_FROM_DEVICE, 0); 
+ 
                 dma_release_channel(chan); 
         } 
Signature · 2 lines
-- 
2.19.0