[patch 2.6.14-rc4] b44: alternate allocation option for DMA descriptors

[patch 2.6.14-rc4] b44: alternate allocation option for DMA descriptors

[John W. Linville]

This is a (final?) hack to support the odd DMA allocation requirements
of the b44 hardware.  The b44 hardware has a 30-bit DMA mask.  On x86,
anything less than a 32-bit DMA mask forces allocations into the 16MB
GFP_DMA range.  The memory there is somewhat limited, often resulting
in an inability to initialize the b44 driver.

This hack uses streaming DMA allocation APIs in order to provide an
alternative in case the GFP_DMA allocation fails.  It is somewhat ugly,
but not much worse than the similar existing hacks to support SKB
allocations in the same driver.  FWIW, I have received positive
feedback on this from several Fedora users.

Signed-off-by: John W. Linville <linville@tuxdriver.com>
---

 drivers/net/b44.c |  134 ++++++++++++++++++++++++++++++++++++++++++++++++++----
 drivers/net/b44.h |    2 
 2 files changed, 128 insertions(+), 8 deletions(-)

diff --git a/drivers/net/b44.c b/drivers/net/b44.c
--- a/drivers/net/b44.c
+++ b/drivers/net/b44.c
@@ -106,6 +106,29 @@ static int b44_poll(struct net_device *d
 static void b44_poll_controller(struct net_device *dev);
 #endif
 
+static int dma_desc_align_mask;
+static int dma_desc_sync_size;
+
+static inline void b44_sync_dma_desc_for_device(struct pci_dev *pdev,
+                                                dma_addr_t dma_base,
+                                                unsigned long offset,
+                                                enum dma_data_direction dir)
+{
+	dma_sync_single_range_for_device(&pdev->dev, dma_base,
+	                                 offset & dma_desc_align_mask,
+	                                 dma_desc_sync_size, dir);
+}
+
+static inline void b44_sync_dma_desc_for_cpu(struct pci_dev *pdev,
+                                             dma_addr_t dma_base,
+                                             unsigned long offset,
+                                             enum dma_data_direction dir)
+{
+	dma_sync_single_range_for_cpu(&pdev->dev, dma_base,
+	                              offset & dma_desc_align_mask,
+	                              dma_desc_sync_size, dir);
+}
+
 static inline unsigned long br32(const struct b44 *bp, unsigned long reg)
 {
 	return readl(bp->regs + reg);
@@ -668,6 +691,11 @@ static int b44_alloc_rx_skb(struct b44 *
 	dp->ctrl = cpu_to_le32(ctrl);
 	dp->addr = cpu_to_le32((u32) mapping + bp->rx_offset + bp->dma_offset);
 
+	if (bp->flags & B44_FLAG_RX_RING_HACK)
+		b44_sync_dma_desc_for_device(bp->pdev, bp->rx_ring_dma,
+		                             dest_idx * sizeof(dp),
+		                             DMA_BIDIRECTIONAL);
+
 	return RX_PKT_BUF_SZ;
 }
 
@@ -692,6 +720,11 @@ static void b44_recycle_rx(struct b44 *b
 	pci_unmap_addr_set(dest_map, mapping,
 			   pci_unmap_addr(src_map, mapping));
 
+	if (bp->flags & B44_FLAG_RX_RING_HACK)
+		b44_sync_dma_desc_for_cpu(bp->pdev, bp->rx_ring_dma,
+		                          src_idx * sizeof(src_desc),
+		                          DMA_BIDIRECTIONAL);
+
 	ctrl = src_desc->ctrl;
 	if (dest_idx == (B44_RX_RING_SIZE - 1))
 		ctrl |= cpu_to_le32(DESC_CTRL_EOT);
@@ -700,8 +733,14 @@ static void b44_recycle_rx(struct b44 *b
 
 	dest_desc->ctrl = ctrl;
 	dest_desc->addr = src_desc->addr;
+
 	src_map->skb = NULL;
 
+	if (bp->flags & B44_FLAG_RX_RING_HACK)
+		b44_sync_dma_desc_for_device(bp->pdev, bp->rx_ring_dma,
+		                             dest_idx * sizeof(dest_desc),
+		                             DMA_BIDIRECTIONAL);
+
 	pci_dma_sync_single_for_device(bp->pdev, src_desc->addr,
 				       RX_PKT_BUF_SZ,
 				       PCI_DMA_FROMDEVICE);
@@ -959,6 +998,11 @@ static int b44_start_xmit(struct sk_buff
 	bp->tx_ring[entry].ctrl = cpu_to_le32(ctrl);
 	bp->tx_ring[entry].addr = cpu_to_le32((u32) mapping+bp->dma_offset);
 
+	if (bp->flags & B44_FLAG_TX_RING_HACK)
+		b44_sync_dma_desc_for_device(bp->pdev, bp->tx_ring_dma,
+		                             entry * sizeof(bp->tx_ring[0]),
+		                             DMA_TO_DEVICE);
+
 	entry = NEXT_TX(entry);
 
 	bp->tx_prod = entry;
@@ -1064,6 +1108,16 @@ static void b44_init_rings(struct b44 *b
 	memset(bp->rx_ring, 0, B44_RX_RING_BYTES);
 	memset(bp->tx_ring, 0, B44_TX_RING_BYTES);
 
+	if (bp->flags & B44_FLAG_RX_RING_HACK)
+		dma_sync_single_for_device(&bp->pdev->dev, bp->rx_ring_dma,
+		                           DMA_TABLE_BYTES,
+		                           PCI_DMA_BIDIRECTIONAL);
+
+	if (bp->flags & B44_FLAG_TX_RING_HACK)
+		dma_sync_single_for_device(&bp->pdev->dev, bp->tx_ring_dma,
+		                           DMA_TABLE_BYTES,
+		                           PCI_DMA_TODEVICE);
+
 	for (i = 0; i < bp->rx_pending; i++) {
 		if (b44_alloc_rx_skb(bp, -1, i) < 0)
 			break;
@@ -1085,14 +1139,28 @@ static void b44_free_consistent(struct b
 		bp->tx_buffers = NULL;
 	}
 	if (bp->rx_ring) {
-		pci_free_consistent(bp->pdev, DMA_TABLE_BYTES,
-				    bp->rx_ring, bp->rx_ring_dma);
+		if (bp->flags & B44_FLAG_RX_RING_HACK) {
+			dma_unmap_single(&bp->pdev->dev, bp->rx_ring_dma,
+				         DMA_TABLE_BYTES,
+				         DMA_BIDIRECTIONAL);
+			kfree(bp->rx_ring);
+		} else
+			pci_free_consistent(bp->pdev, DMA_TABLE_BYTES,
+					    bp->rx_ring, bp->rx_ring_dma);
 		bp->rx_ring = NULL;
+		bp->flags &= ~B44_FLAG_RX_RING_HACK;
 	}
 	if (bp->tx_ring) {
-		pci_free_consistent(bp->pdev, DMA_TABLE_BYTES,
-				    bp->tx_ring, bp->tx_ring_dma);
+		if (bp->flags & B44_FLAG_TX_RING_HACK) {
+			dma_unmap_single(&bp->pdev->dev, bp->tx_ring_dma,
+				         DMA_TABLE_BYTES,
+				         DMA_TO_DEVICE);
+			kfree(bp->tx_ring);
+		} else
+			pci_free_consistent(bp->pdev, DMA_TABLE_BYTES,
+					    bp->tx_ring, bp->tx_ring_dma);
 		bp->tx_ring = NULL;
+		bp->flags &= ~B44_FLAG_TX_RING_HACK;
 	}
 }
 
@@ -1118,12 +1186,56 @@ static int b44_alloc_consistent(struct b
 
 	size = DMA_TABLE_BYTES;
 	bp->rx_ring = pci_alloc_consistent(bp->pdev, size, &bp->rx_ring_dma);
-	if (!bp->rx_ring)
-		goto out_err;
+	if (!bp->rx_ring) {
+		/* Allocation may have failed due to pci_alloc_consistent
+		   insisting on use of GFP_DMA, which is more restrictive
+		   than necessary...  */
+		struct dma_desc *rx_ring;
+		dma_addr_t rx_ring_dma;
+
+		if (!(rx_ring = (struct dma_desc *)kmalloc(size, GFP_KERNEL)))
+			goto out_err;
+
+		memset(rx_ring, 0, size);
+		rx_ring_dma = dma_map_single(&bp->pdev->dev, rx_ring,
+		                             DMA_TABLE_BYTES,
+		                             DMA_BIDIRECTIONAL);
+
+		if (rx_ring_dma + size > B44_DMA_MASK) {
+			kfree(rx_ring);
+			goto out_err;
+		}
+
+		bp->rx_ring = rx_ring;
+		bp->rx_ring_dma = rx_ring_dma;
+		bp->flags |= B44_FLAG_RX_RING_HACK;
+	}
 
 	bp->tx_ring = pci_alloc_consistent(bp->pdev, size, &bp->tx_ring_dma);
-	if (!bp->tx_ring)
-		goto out_err;
+	if (!bp->tx_ring) {
+		/* Allocation may have failed due to pci_alloc_consistent
+		   insisting on use of GFP_DMA, which is more restrictive
+		   than necessary...  */
+		struct dma_desc *tx_ring;
+		dma_addr_t tx_ring_dma;
+
+		if (!(tx_ring = (struct dma_desc *)kmalloc(size, GFP_KERNEL)))
+			goto out_err;
+
+		memset(tx_ring, 0, size);
+		tx_ring_dma = dma_map_single(&bp->pdev->dev, tx_ring,
+		                             DMA_TABLE_BYTES,
+		                             DMA_TO_DEVICE);
+
+		if (tx_ring_dma + size > B44_DMA_MASK) {
+			kfree(tx_ring);
+			goto out_err;
+		}
+
+		bp->tx_ring = tx_ring;
+		bp->tx_ring_dma = tx_ring_dma;
+		bp->flags |= B44_FLAG_TX_RING_HACK;
+	}
 
 	return 0;
 
@@ -1971,6 +2083,12 @@ static struct pci_driver b44_driver = {
 
 static int __init b44_init(void)
 {
+	unsigned int dma_desc_align_size = dma_get_cache_alignment();
+
+	/* Setup paramaters for syncing RX/TX DMA descriptors */
+	dma_desc_align_mask = ~(dma_desc_align_size - 1);
+	dma_desc_sync_size = max(dma_desc_align_size, sizeof(struct dma_desc));
+
 	return pci_module_init(&b44_driver);
 }
 
diff --git a/drivers/net/b44.h b/drivers/net/b44.h
--- a/drivers/net/b44.h
+++ b/drivers/net/b44.h
@@ -400,6 +400,8 @@ struct b44 {
 #define B44_FLAG_ADV_100HALF	0x04000000
 #define B44_FLAG_ADV_100FULL	0x08000000
 #define B44_FLAG_INTERNAL_PHY	0x10000000
+#define B44_FLAG_RX_RING_HACK	0x20000000
+#define B44_FLAG_TX_RING_HACK	0x40000000
 
 	u32			rx_offset;
 

Re: [patch 2.6.14-rc4] b44: alternate allocation option for DMA descriptors

[Jeff Garzik]

applied

Re: [patch 2.6.14-rc4] b44: alternate allocation option for DMA descriptors

[Benjamin Herrenschmidt]

On Tue, 2005-10-18 at 21:30 -0400, John W. Linville wrote:
> This is a (final?) hack to support the odd DMA allocation requirements
> of the b44 hardware.  The b44 hardware has a 30-bit DMA mask.  On x86,
> anything less than a 32-bit DMA mask forces allocations into the 16MB
> GFP_DMA range.  The memory there is somewhat limited, often resulting
> in an inability to initialize the b44 driver.
> 
> This hack uses streaming DMA allocation APIs in order to provide an
> alternative in case the GFP_DMA allocation fails.  It is somewhat ugly,
> but not much worse than the similar existing hacks to support SKB
> allocations in the same driver.  FWIW, I have received positive
> feedback on this from several Fedora users.

I'm not sure what you are trying to do here ... If pci_alloc_* failed,
you do kmalloc(...,GFP_KERNEL) which can give you memory above your DMA
mask. Then, you use dma_map_* but that won't help much more neither.
Unless you have an iommu (or swiotlb) _and_ that implements arbitrary
DMA masks support (which it typically doesn't it's often 32 bits vs. 64
bits) it won't help, you'll get into your error case.

So basically, what you are doing is: if allocation fails, you try to get
memory using GFP_KERNEL. If it happens to be in the low 2Gb of memory,
use it, if not, drop it.

Did I get that right ?

Note that the Broadcom wireless (which is currently being reverse
engineered) seem to suffer from the same stupid DMA engine...

Ben.

Re: [patch 2.6.14-rc4] b44: alternate allocation option for DMA descriptors

[John W. Linville]

On Thu, Oct 20, 2005 at 05:17:05PM +1000, Benjamin Herrenschmidt wrote:

> So basically, what you are doing is: if allocation fails, you try to get
> memory using GFP_KERNEL. If it happens to be in the low 2Gb of memory,
> use it, if not, drop it.
> 
> Did I get that right ?

Yes, that is basically correct.  I wish I had something more clever
than that...suggestions welcome...

John
-- 
John W. Linville
linville@tuxdriver.com