Skip to contentSkip to conversation

x86-64: Calgary IOMMU - move valid_dma_direction into the callers

PATCH In review Jon Mason 4 participants 7 messages May 25, 2006 – May 26, 2006

lore
Share & tools
lore

Jon Mason

author
#0 lore raw

PATCH 2/4 x86-64: Calgary IOMMU - move valid_dma_direction into the callers

Quoted 1 line

From Andi Kleen's comments on the original Calgary patch, move

valid_dma_direction into the calling functions.

include/asm-x86_64/dma-mapping.h +17 -0
diff -r 460b603caa42 include/asm-x86_64/dma-mapping.h
@@ -56,6 +56,13 @@ extern struct dma_mapping_ops* dma_ops;
 extern struct dma_mapping_ops* dma_ops;
 extern int iommu_merge;
 
+static inline int valid_dma_direction(int dma_direction)
+{
+	return ((dma_direction == DMA_BIDIRECTIONAL) ||
+		(dma_direction == DMA_TO_DEVICE) ||
+		(dma_direction == DMA_FROM_DEVICE));
+}
+
 static inline int dma_mapping_error(dma_addr_t dma_addr)
 {
 	if (dma_ops->mapping_error)
@@ -73,6 +80,7 @@ dma_map_single(struct device *hwdev, voi
 dma_map_single(struct device *hwdev, void *ptr, size_t size,
 	       int direction)
 {
+	BUG_ON(!valid_dma_direction(direction));
 	return dma_ops->map_single(hwdev, ptr, size, direction);
 }
 
@@ -80,6 +88,7 @@ dma_unmap_single(struct device *dev, dma
 dma_unmap_single(struct device *dev, dma_addr_t addr,size_t size,
 		 int direction)
 {
+	BUG_ON(!valid_dma_direction(direction));
 	dma_ops->unmap_single(dev, addr, size, direction);
 }
 
@@ -92,6 +101,7 @@ dma_sync_single_for_cpu(struct device *h
 dma_sync_single_for_cpu(struct device *hwdev, dma_addr_t dma_handle,
 			size_t size, int direction)
 {
+	BUG_ON(!valid_dma_direction(direction));
 	if (dma_ops->sync_single_for_cpu)
 		dma_ops->sync_single_for_cpu(hwdev, dma_handle, size,
 					     direction);
@@ -102,6 +112,7 @@ dma_sync_single_for_device(struct device
 dma_sync_single_for_device(struct device *hwdev, dma_addr_t dma_handle,
 			   size_t size, int direction)
 {
+	BUG_ON(!valid_dma_direction(direction));
 	if (dma_ops->sync_single_for_device)
 		dma_ops->sync_single_for_device(hwdev, dma_handle, size,
 						direction);
@@ -112,6 +123,7 @@ dma_sync_single_range_for_cpu(struct dev
 dma_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dma_handle,
 			      unsigned long offset, size_t size, int direction)
 {
+	BUG_ON(!valid_dma_direction(direction));
 	if (dma_ops->sync_single_range_for_cpu) {
 		dma_ops->sync_single_range_for_cpu(hwdev, dma_handle, offset, size, direction);
 	}
@@ -123,6 +135,7 @@ dma_sync_single_range_for_device(struct 
 dma_sync_single_range_for_device(struct device *hwdev, dma_addr_t dma_handle,
 				 unsigned long offset, size_t size, int direction)
 {
+	BUG_ON(!valid_dma_direction(direction));
 	if (dma_ops->sync_single_range_for_device)
 		dma_ops->sync_single_range_for_device(hwdev, dma_handle,
 						      offset, size, direction);
@@ -134,6 +147,7 @@ dma_sync_sg_for_cpu(struct device *hwdev
 dma_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
 		    int nelems, int direction)
 {
+	BUG_ON(!valid_dma_direction(direction));
 	if (dma_ops->sync_sg_for_cpu)
 		dma_ops->sync_sg_for_cpu(hwdev, sg, nelems, direction);
 	flush_write_buffers();
@@ -143,6 +157,7 @@ dma_sync_sg_for_device(struct device *hw
 dma_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
 		       int nelems, int direction)
 {
+	BUG_ON(!valid_dma_direction(direction));
 	if (dma_ops->sync_sg_for_device) {
 		dma_ops->sync_sg_for_device(hwdev, sg, nelems, direction);
 	}
@@ -153,6 +168,7 @@ static inline int
 static inline int
 dma_map_sg(struct device *hwdev, struct scatterlist *sg, int nents, int direction)
 {
+	BUG_ON(!valid_dma_direction(direction));
 	return dma_ops->map_sg(hwdev, sg, nents, direction);
 }
 
@@ -160,6 +176,7 @@ dma_unmap_sg(struct device *hwdev, struc
 dma_unmap_sg(struct device *hwdev, struct scatterlist *sg, int nents,
 	     int direction)
 {
+	BUG_ON(!valid_dma_direction(direction));
 	dma_ops->unmap_sg(hwdev, sg, nents, direction);
 }
 
#1 lore raw
Quoted 5 lines · Jon Mason
Quoted 1 line

From Andi Kleen's comments on the original Calgary patch, move

valid_dma_direction into the calling functions.

Even though BUG_ON() includes unlikely(), this introduces additional tests in very hot paths.

_Why_ do we need this at all?

I would prefer to NAK the patch, and fix the odd user that gets it wrong. It becomes REALLY obvious that a driver has gotten this wrong, REALLY quickly.

I see no need to burden critical hot paths with dumb checks like this.

Jeff

#2 lore raw
Quoted 9 lines · Jeff Garzik
Quoted 5 lines · Jon Mason
Quoted 1 line

From Andi Kleen's comments on the original Calgary patch, move

valid_dma_direction into the calling functions.

Even though BUG_ON() includes unlikely(), this introduces additional tests in very hot paths.

Are they really very hot? I mean if you're calling the DMA API, you're about to frob the hardware or have already frobbed it - does this check really matter?

Quoted 1 line

_Why_ do we need this at all?

It was helpful for us during the dma-ops work and Calgary bringup and Andi requested that we move it from Calgary to common code. I think we're fine with dropping it if that's the consensus, but it did catch a few bugs early on and the cost is tiny.

Cheers,
Muli

Signature · 2 lines
Muli Ben-Yehuda
http://www.mulix.org | http://mulix.livejournal.com/
#3 lore raw
Quoted 5 lines · Muli Ben-Yehuda
Quoted 4 lines · Jeff Garzik
Quoted 3 lines · Jon Mason
Quoted 1 line

From Andi Kleen's comments on the original Calgary patch, move

valid_dma_direction into the calling functions.

Quoted 7 lines
Quoted 3 lines
Quoted 2 lines

Even though BUG_ON() includes unlikely(), this introduces additional tests in very hot paths.

Are they really very hot? I mean if you're calling the DMA API, you're about to frob the hardware or have already frobbed it - does this check really matter?

When you are adding a check that will _never_ be hit in production, to the _hottest_ paths in the kernel, you can be assured it matters...

Quoted 6 lines
Quoted 1 line

_Why_ do we need this at all?

It was helpful for us during the dma-ops work and Calgary bringup and Andi requested that we move it from Calgary to common code. I think we're fine with dropping it if that's the consensus, but it did catch a few bugs early on and the cost is tiny.

Key phrase: "early on"

These checks will only be useful during _early_ development of a new DMA platform. For _100%_ of the real world users, these checks are useless.
Not 99%, 100%.

Jeff

#4 lore raw
Quoted 11 lines · Jeff Garzik
Quoted 10 lines · Muli Ben-Yehuda
Quoted 8 lines · Jeff Garzik
Quoted 5 lines · Jon Mason
Quoted 1 line

From Andi Kleen's comments on the original Calgary patch, move

valid_dma_direction into the calling functions.

Even though BUG_ON() includes unlikely(), this introduces additional tests in very hot paths.

Quoted 6 lines
Quoted 3 lines

Are they really very hot? I mean if you're calling the DMA API, you're about to frob the hardware or have already frobbed it - does this check really matter?

When you are adding a check that will _never_ be hit in production, to the _hottest_ paths in the kernel, you can be assured it matters...

pci_dma_* shouldn't be that hot. Or at least IO usually has so much overhead that some more bugging shouldn't matter.

On the other hand if the problem of passing wrong parameters here isn't common I would be ok with dropping them.

-Andi

#5 lore raw
Quoted 9 lines · Andi Kleen
Quoted 8 lines · Jeff Garzik
Quoted 7 lines · Muli Ben-Yehuda
Quoted 6 lines · Jeff Garzik
Quoted 5 lines · Jon Mason
Quoted 1 line

From Andi Kleen's comments on the original Calgary patch, move

valid_dma_direction into the calling functions.

Quoted 8 lines
Quoted 5 lines
Quoted 4 lines
Quoted 1 line

Even though BUG_ON() includes unlikely(), this introduces additional tests in very hot paths.

Are they really very hot? I mean if you're calling the DMA API, you're about to frob the hardware or have already frobbed it - does this check really matter?

When you are adding a check that will _never_ be hit in production, to the _hottest_ paths in the kernel, you can be assured it matters...

pci_dma_* shouldn't be that hot. Or at least IO usually has so much overhead that some more bugging shouldn't matter.

I respectfully disagree with that logic. If its a key hot path -- which it is, every modern network or disk I/O runs through these paths -- then it deserves at least _some_ consideration before adding more CPU cycles.

Quoted 2 lines

On the other hand if the problem of passing wrong parameters here isn't common I would be ok with dropping them.

As the author noted, it was only used in early platform bring-up. And simply reviewing the patch... it is clear that screwing up the parameters would cause massive, noticeable problems immediately -- such as on EM64T with swiotlb.

Jeff

#6 lore raw
Quoted 8 lines
Quoted 3 lines

On the other hand if the problem of passing wrong parameters here isn't common I would be ok with dropping them.

As the author noted, it was only used in early platform bring-up. And simply reviewing the patch... it is clear that screwing up the parameters would cause massive, noticeable problems immediately -- such as on EM64T with swiotlb.

Well the point is that even if you don't run on swiotlb
(which most people even on EM64T don't use because they don't have more than 3GB of RAM) you won't screw them up.

On the other hand correct IOMMU operation needs much more than passing these parameters so only checking for this might not be very useful.

BTW we've had these checks forever - Jon just moved them to a different place. But they indeed might have outlived their usefulness.

-Andi