Skip to contentSkip to conversation

Where do we stand with the Xen patches?

PATCH New Jeremy Fitzhardinge 7 participants 43 messages May 14, 2009 – May 22, 2009

Subsystems (8)
x86 PowerPC PCI Library code +4 more subsystems
lore
Share & tools
lore
#0 lore raw

Hi Ingo,

Over the last week or so, I've set out pull requests for the following branches in git://git.kernel.org/pub/scm/linux/kernel/git/jeremy/xen.git :

for-ingo/xen/dom0/core

You made two comments about the first post of this set:

1. The // comments in the mtrr code. Now fixed.
2. A query about when Xen can support PAT. In progress; when its done, we can remove the unconditional PAT disable.

for-ingo/xen/dom0/pci
for-ingo/xen/dom0/swiotlb

Updated with Joerg Roedel, FUJITA Tomonori and Matthew Wilcox's comments, Acked-by and Reviewed-bys as appropriate.

for-ingo/xen/dom0/apic-ops

After discussion between yourself and HPA, we resolved that using io_apic_ops was the right way to go forward with this. I replaced for-ingo/xen/dom0/apic with the new branch for-ingo/xen/dom0/apic-ops, which is identical aside from implementing and using io_apic_ops.

for-ingo/xen/dom0/mtrr

You queried the value of "extending" this interface, given that its considered to be deprecated. These changes in no way extend the interface, but just make the existing interface functional under Xen. And while we don't have PAT support, there's no other way of setting cachability attributes on memory, so not supporting it has a fairly severe performance impact on things like X.

Aside from some whitespace issues around some Impact: lines, I don't know of any outstanding problems. (I just pushed an updates to these branches to fix those, and fold a change to address Jesse's comment.)

Please tell me if you have any further issues which prevents you from pulling these changes. Otherwise I'd appreciate it if you pulled them soon, as we're already on -rc5, and I have more changes I'd like to prep for the next merge window.

Thanks,
J

#1 lore raw
Quoted 14 lines · Jeremy Fitzhardinge

Hi Ingo,

Over the last week or so, I've set out pull requests for the following branches in git://git.kernel.org/pub/scm/linux/kernel/git/jeremy/xen.git
:

for-ingo/xen/dom0/core

You made two comments about the first post of this set:

1. The // comments in the mtrr code. Now fixed.
2. A query about when Xen can support PAT. In progress; when its done, we can remove the unconditional PAT disable.

Quoted 5 lines

for-ingo/xen/dom0/pci
for-ingo/xen/dom0/swiotlb

Updated with Joerg Roedel, FUJITA Tomonori and Matthew Wilcox's comments, Acked-by and Reviewed-bys as appropriate.

These look fine but i still need to go over them one last time before pulling them.

Quoted 7 lines

for-ingo/xen/dom0/apic-ops

After discussion between yourself and HPA, we resolved that using io_apic_ops was the right way to go forward with this. I replaced for-ingo/xen/dom0/apic with the new branch for-ingo/xen/dom0/apic-ops, which is identical aside from implementing and using io_apic_ops.

Yes. Here too i still need to go over them once more before pulling them.

Quoted 8 lines

for-ingo/xen/dom0/mtrr

You queried the value of "extending" this interface, given that its considered to be deprecated. These changes in no way extend the interface, but just make the existing interface functional under Xen. And while we don't have PAT support, there's no other way of setting cachability attributes on memory, so not supporting it has a fairly severe performance impact on things like X.

Latest Xorg doesnt need /proc/mtrr. By the time this kernel (.31 or later) hits any distribution, /proc/mtrr using Xorg will be a distant memory. So i see no reason why to apply those bits at all, and i see a lot of reasons to not apply them.

Quoted 9 lines

Aside from some whitespace issues around some Impact: lines, I don't know of any outstanding problems. (I just pushed an updates to these branches to fix those, and fold a change to address Jesse's comment.)

Please tell me if you have any further issues which prevents you from pulling these changes. Otherwise I'd appreciate it if you pulled them soon, as we're already on -rc5, and I have more changes I'd like to prep for the next merge window.

As in the past, my main worry is performance overhead of paravirt in general.

The patches that dont affect any native kernel fast path are probably OK (but still pending final review).

Regarding patches that do change the fastpath i'll do a round of measurements of CONFIG_PARAVIRT against !CONFIG_PARAVIRT kernels, and make up my mind based on that.

You could accelerate this by sending some "perf stat" hard numbers to give us an idea about where we stand today.

Ingo

#2 lore raw
Quoted 5 lines · Ingo Molnar

These look fine but i still need to go over them one last time before pulling them.

Yes. Here too i still need to go over them once more before pulling them.

I've been posting these patches in fundamentally the same form for about 6 months now. I don't think you'll find anything surprising.

Quoted 10 lines
Quoted 9 lines

for-ingo/xen/dom0/mtrr

You queried the value of "extending" this interface, given that its considered to be deprecated. These changes in no way extend the interface, but just make the existing interface functional under Xen. And while we don't have PAT support, there's no other way of setting cachability attributes on memory, so not supporting it has a fairly severe performance impact on things like X.

Quoted 2 lines

Latest Xorg doesnt need /proc/mtrr. By the time this kernel (.31 or later) hits any distribution, /proc/mtrr using Xorg will be a distant memory. So i see no reason why to apply those bits at all, and i see a lot of reasons to not apply them.

In general we don't break usermode ABIs, even when using new kernels with old distros, so I don't see why this case is any different.

Besides, these changes are not only for /proc/mtrr, but also to implement the internal mtrr_add() APIs that DRM uses to set the MTRR inside the kernel, so failing to implement them will cause performance regressions on new X servers.

Given that we're talking about 120 lines of code with no runtime impact and tiny kernel size impact (when configured), I'm at a loss for what your "lot of reasons" might be. Perhaps you could be more specific.

Quoted 4 lines

As in the past, my main worry is performance overhead of paravirt in general.

The patches that dont affect any native kernel fast path are probably OK (but still pending final review).

These changes don't have anything much to do with paravirt_ops, per se, and are all specifically about Xen dom0 support. Aside from that, none of the changes are on performance-critical paths.

Quoted 3 lines

Regarding patches that do change the fastpath i'll do a round of measurements of CONFIG_PARAVIRT against !CONFIG_PARAVIRT kernels, and make up my mind based on that.

You could accelerate this by sending some "perf stat" hard numbers to give us an idea about where we stand today.

I posted them the other day; those perf stat measurements pointing out the pv-spinlock problem also showed that paravirt_ops in general has a sub-1% performance hit (and possible performance benefit) when running mmap-perf. You added them into the commit log for the patch, so I presume you read it.

J

#3 lore raw
In reply to Jeremy Fitzhardinge
Quoted 13 lines · On Thu, 14 May 2009 12:54:54 -0700 Jeremy Fitzhardinge <[email protected]> wrote:

Hi Ingo,

Over the last week or so, I've set out pull requests for the following branches in git://git.kernel.org/pub/scm/linux/kernel/git/jeremy/xen.git :

for-ingo/xen/dom0/core

You made two comments about the first post of this set:

1. The // comments in the mtrr code. Now fixed.
2. A query about when Xen can support PAT. In progress; when its done, we can remove the unconditional PAT disable.

Quoted 5 lines

for-ingo/xen/dom0/pci
for-ingo/xen/dom0/swiotlb

Updated with Joerg Roedel, FUJITA Tomonori and Matthew Wilcox's comments, Acked-by and Reviewed-bys as appropriate.

The original code added dom0-specific dma mapping stuff in the generic place, which is completely wrong. I asked you to move the hacky stuff to Xen-specific code and ack'ed the patchset.

But as I said again and again, the dom0 changes to the generic dma mapipng code is really ugly and I don't like them at all. I didn't ack'ed such changes.

#4 lore raw
Quoted 7 lines · FUJITA Tomonori
Quoted 6 lines

for-ingo/xen/dom0/pci
for-ingo/xen/dom0/swiotlb

Updated with Joerg Roedel, FUJITA Tomonori and Matthew Wilcox's comments, Acked-by and Reviewed-bys as appropriate.

Quoted 8 lines

The original code added dom0-specific dma mapping stuff in the generic place, which is completely wrong. I asked you to move the hacky stuff to Xen-specific code and ack'ed the patchset.

But as I said again and again, the dom0 changes to the generic dma mapipng code is really ugly and I don't like them at all. I didn't ack'ed such changes.

I didn't mean to imply that you all acked all the patches. I tried to be careful to add Acked tags to the actual patches you each respectively acked.

J

#5 lore raw
In reply to FUJITA Tomonori
Quoted 22 lines · FUJITA Tomonori
Quoted 13 lines · On Thu, 14 May 2009 12:54:54 -0700 Jeremy Fitzhardinge <[email protected]> wrote:

Hi Ingo,

Over the last week or so, I've set out pull requests for the following branches in git://git.kernel.org/pub/scm/linux/kernel/git/jeremy/xen.git :

for-ingo/xen/dom0/core

You made two comments about the first post of this set:

1. The // comments in the mtrr code. Now fixed.
2. A query about when Xen can support PAT. In progress; when its done, we can remove the unconditional PAT disable.

Quoted 5 lines

for-ingo/xen/dom0/pci
for-ingo/xen/dom0/swiotlb

Updated with Joerg Roedel, FUJITA Tomonori and Matthew Wilcox's comments, Acked-by and Reviewed-bys as appropriate.

Quoted 7 lines

The original code added dom0-specific dma mapping stuff in the generic place, which is completely wrong. I asked you to move the hacky stuff to Xen-specific code and ack'ed the patchset.

But as I said again and again, the dom0 changes to the generic dma mapipng code is really ugly and I don't like them at all. I didn't ack'ed such changes.

How should it be solved instead? Can you see a clean way to achieve it? (maybe you already explained it in past threads - if yes then have you got subject lines or URIs to that?)

Thanks,

Ingo

#6 lore raw
Quoted 29 lines · On Mon, 18 May 2009 10:40:52 +0200 Ingo Molnar <[email protected]> wrote:
Quoted 26 lines · FUJITA Tomonori
Quoted 13 lines · On Thu, 14 May 2009 12:54:54 -0700 Jeremy Fitzhardinge <[email protected]> wrote:

Hi Ingo,

Over the last week or so, I've set out pull requests for the following branches in git://git.kernel.org/pub/scm/linux/kernel/git/jeremy/xen.git :

for-ingo/xen/dom0/core

You made two comments about the first post of this set:

1. The // comments in the mtrr code. Now fixed.
2. A query about when Xen can support PAT. In progress; when its done, we can remove the unconditional PAT disable.

Quoted 5 lines

for-ingo/xen/dom0/pci
for-ingo/xen/dom0/swiotlb

Updated with Joerg Roedel, FUJITA Tomonori and Matthew Wilcox's comments, Acked-by and Reviewed-bys as appropriate.

The original code added dom0-specific dma mapping stuff in the generic place, which is completely wrong. I asked you to move the hacky stuff to Xen-specific code and ack'ed the patchset.

Quoted 7 lines
Quoted 3 lines

But as I said again and again, the dom0 changes to the generic dma mapipng code is really ugly and I don't like them at all. I didn't ack'ed such changes.

How should it be solved instead? Can you see a clean way to achieve it? (maybe you already explained it in past threads - if yes then have you got subject lines or URIs to that?)

If we really need to merge dom0 support, then dom0 should have the own IOMMU implementation instead of adding any hacks to swiotlb (as I said long time ago). Yeah, there might be some duplication between swiotlb and the Xen IOMMU but IMO it's better to have clean code with some duplication rather than ugly unified code; unification makes sense only if we do cleanly.

#7 lore raw
Quoted 40 lines · FUJITA Tomonori
Quoted 29 lines · On Mon, 18 May 2009 10:40:52 +0200 Ingo Molnar <[email protected]> wrote:
Quoted 26 lines · FUJITA Tomonori
Quoted 13 lines · On Thu, 14 May 2009 12:54:54 -0700 Jeremy Fitzhardinge <[email protected]> wrote:

Hi Ingo,

Over the last week or so, I've set out pull requests for the following branches in git://git.kernel.org/pub/scm/linux/kernel/git/jeremy/xen.git :

for-ingo/xen/dom0/core

You made two comments about the first post of this set:

1. The // comments in the mtrr code. Now fixed.
2. A query about when Xen can support PAT. In progress; when its done, we can remove the unconditional PAT disable.

Quoted 5 lines

for-ingo/xen/dom0/pci
for-ingo/xen/dom0/swiotlb

Updated with Joerg Roedel, FUJITA Tomonori and Matthew Wilcox's comments, Acked-by and Reviewed-bys as appropriate.

The original code added dom0-specific dma mapping stuff in the generic place, which is completely wrong. I asked you to move the hacky stuff to Xen-specific code and ack'ed the patchset.

Quoted 7 lines
Quoted 3 lines

But as I said again and again, the dom0 changes to the generic dma mapipng code is really ugly and I don't like them at all. I didn't ack'ed such changes.

How should it be solved instead? Can you see a clean way to achieve it? (maybe you already explained it in past threads - if yes then have you got subject lines or URIs to that?)

Quoted 6 lines

If we really need to merge dom0 support, then dom0 should have the own IOMMU implementation instead of adding any hacks to swiotlb (as I said long time ago). Yeah, there might be some duplication between swiotlb and the Xen IOMMU but IMO it's better to have clean code with some duplication rather than ugly unified code; unification makes sense only if we do cleanly.

Well, but since we merged PowerPC highmem support ... the precedent is there already that the swiotlb code can be librarized and generalized some more, right?

What is the fundamental difference between a DMA32 device on native hardware using lib/swiotlb.c to bounce IO directed to/from a high address over buffers in a low, DMA-capable address, and between a Xen Dom0 instance using the same functionality to DMA buffers?

The payback is significant: Linux drivers can be used basically as-is (as far as their DMA functionality goes) in a Xen dom0 guest.

There's really just two non-standard Xen features compared to the PowerPC-highmem use:

- special allocation

- non-standard virtual/physical/bus memory translations

The latter is understandable - Xen cannot do the (mostly-) linear transformations that (most) platforms can do in their address transformation primitives. But other than the non-linearity, it's a "mapping" just as much.

The special allocator is arguably a bit of a hack - Xen should _probably_ recognize GFP_DMA32 allocations at the page allocator level and rearrange buffers there.

So realizing that dom0 _has_ to have guest OS help here, the best possible design seems to be to minimalize the Xen cross section to Linux, while still keeping it fast enough. And that seems to be these two straightforward hooks: 'ready buffers for DMA' and 'map addresses'. As long as we'll want to have swiotlb code in the kernel, we'll deal with such primitives anyway - so there doesnt seem to be a significant mainteinance drag here.

Hm?

Ingo

#8 lore raw
Quoted 10 lines · On Tue, 19 May 2009 15:03:00 +0200 Ingo Molnar <[email protected]> wrote:
Quoted 6 lines

If we really need to merge dom0 support, then dom0 should have the own IOMMU implementation instead of adding any hacks to swiotlb (as I said long time ago). Yeah, there might be some duplication between swiotlb and the Xen IOMMU but IMO it's better to have clean code with some duplication rather than ugly unified code; unification makes sense only if we do cleanly.

Well, but since we merged PowerPC highmem support ... the precedent is there already that the swiotlb code can be librarized and generalized some more, right?

No,

The highmem support is ok for me (I was against the initial highmem patchset but Becky's highmem patchset was much cleaner and we merged it).

Due to dom0 support, we can't use architecture abstraction (such as arch/include/asm/):

http://marc.info/?l=linux-kernel&m=124271086910299&w=2

We also need hacks for memory allocation due to dom0 support.

Quoted 24 lines

What is the fundamental difference between a DMA32 device on native hardware using lib/swiotlb.c to bounce IO directed to/from a high address over buffers in a low, DMA-capable address, and between a Xen Dom0 instance using the same functionality to DMA buffers?

The payback is significant: Linux drivers can be used basically as-is (as far as their DMA functionality goes) in a Xen dom0 guest.

There's really just two non-standard Xen features compared to the PowerPC-highmem use:

- special allocation

- non-standard virtual/physical/bus memory translations

The latter is understandable - Xen cannot do the (mostly-) linear transformations that (most) platforms can do in their address transformation primitives. But other than the non-linearity, it's a "mapping" just as much.

The special allocator is arguably a bit of a hack - Xen should _probably_ recognize GFP_DMA32 allocations at the page allocator level and rearrange buffers there.

Quoted 7 lines

So realizing that dom0 _has_ to have guest OS help here, the best possible design seems to be to minimalize the Xen cross section to Linux, while still keeping it fast enough. And that seems to be these two straightforward hooks: 'ready buffers for DMA' and 'map addresses'. As long as we'll want to have swiotlb code in the kernel, we'll deal with such primitives anyway - so there doesnt seem to be a significant mainteinance drag here.

We need these hooks but as I wrote above, they are architecture-specific and we should handle them with the architecture abstraction (as we handle similar problems) however we can't due to dom0 support.

#9 lore raw
Quoted 5 lines · FUJITA Tomonori

We need these hooks but as I wrote above, they are architecture-specific and we should handle them with the architecture abstraction (as we handle similar problems) however we can't due to dom0 support.

I don't understand this. What exactly about the dom0 support patch prevents future abstraction here?

The dom0 hooks would simply move into the per-arch abstractions as appropriate, wouldn't they?

Ian.

Signature · 5 lines
Ian Campbell
Current Noise: Mondo Generator - Simple Exploding Man

"We are on the verge: Today our program proved Fermat's next-to-last theorem."
		-- Epigrams in Programming, ACM SIGPLAN Sept. 1982
#10 lore raw
Quoted 8 lines · Ian Campbell
Quoted 5 lines · FUJITA Tomonori

We need these hooks but as I wrote above, they are architecture-specific and we should handle them with the architecture abstraction (as we handle similar problems) however we can't due to dom0 support.

Quoted 5 lines

I don't understand this. What exactly about the dom0 support patch prevents future abstraction here?

The dom0 hooks would simply move into the per-arch abstractions as appropriate, wouldn't they?

Fujita-san's suggestion to me was that swiotlb could just use the normal (albeit deprecated) phys_to_bus()/bus_to_phys() interfaces rather than defining its own. That would be perfectly OK for Xen; we have a single global translation which is unaffected by the target device, etc.

But I'm not sure it would work for powerpc; Becky's patches which added swiotlb_bus_to_phys/phys_bus made them take a device argument, because (apparently) the bus/phys offset can differ on a per-device or per-bus basis. The powerpc side of swiotlb doesn't seem to be in mainline yet, so I'm not sure what the details are here (maybe it can be handled with a single big runtime switch, if the offset is always constant on a given machine).

(Hm, now that I look I see that they're defined as virt_to_bus/bus_to_virt, which doesn't work for highmem at all; it would need to be phys.)

But I may have misinterpreted what he meant.

J

#11 lore raw
Quoted 28 lines · On Wed, 20 May 2009 10:06:13 -0700 Jeremy Fitzhardinge <[email protected]> wrote:
Quoted 8 lines · Ian Campbell
Quoted 5 lines · FUJITA Tomonori

We need these hooks but as I wrote above, they are architecture-specific and we should handle them with the architecture abstraction (as we handle similar problems) however we can't due to dom0 support.

Quoted 5 lines

I don't understand this. What exactly about the dom0 support patch prevents future abstraction here?

The dom0 hooks would simply move into the per-arch abstractions as appropriate, wouldn't they?

Fujita-san's suggestion to me was that swiotlb could just use the normal (albeit deprecated) phys_to_bus()/bus_to_phys() interfaces rather than defining its own. That would be perfectly OK for Xen; we have a single global translation which is unaffected by the target device, etc.

But I'm not sure it would work for powerpc; Becky's patches which added swiotlb_bus_to_phys/phys_bus made them take a device argument, because (apparently) the bus/phys offset can differ on a per-device or per-bus basis. The powerpc side of swiotlb doesn't seem to be in mainline yet, so I'm not sure what the details are here (maybe it can be handled with a single big runtime switch, if the offset is always constant on a given machine).

Quoted 5 lines

(Hm, now that I look I see that they're defined as virt_to_bus/bus_to_virt, which doesn't work for highmem at all; it would need to be phys.)

But I may have misinterpreted what he meant.

What I want to remove all the __weak hacks and use the architecture abstraction. For example, the following patch is killing swiotlb_arch_address_needs_mapping() and swiotlb_arch_range_needs_mapping().

As you said, POWERPC needs a pair of conversion functions between phys and bus. I want to use arch/*/include/ for it in the same way.

=
From: FUJITA Tomonori <[email protected]>
Subject: [PATCH] make is_buffer_dma_capable architecture-specific

This moves is_buffer_dma_capable() from include/linux/dma-mapping.h to arch/*/include/asm/dma-mapping.h because it's architecture-specific; we shouldn't have added it in the generic place.

This function is used only in swiotlb (supported by x86 and IA64, and POWERPC shortly).

POWERPC needs struct device to see if an address is DMA-capable or not. How POWERPC implements is_buffer_dma_capable() is still under discussion. So this patch doesn't add POWERPC's one.

---
 arch/ia64/include/asm/dma-mapping.h |    6 +++++
 arch/x86/include/asm/dma-mapping.h  |    6 +++++
 arch/x86/kernel/pci-dma.c           |    2 +-
 arch/x86/kernel/pci-gart_64.c       |    4 +-
 arch/x86/kernel/pci-nommu.c         |    2 +-
 include/linux/dma-mapping.h         |    5 ----
 lib/swiotlb.c                       |   39 +++++++---------------------------
 7 files changed, 24 insertions(+), 40 deletions(-)
arch/ia64/include/asm/dma-mapping.h +6 -0
index 36c0009..a091648 100644
@@ -174,4 +174,10 @@ dma_cache_sync (struct device *dev, void *vaddr, size_t size,
 
 #define dma_is_consistent(d, h)	(1)	/* all we do is coherent memory... */
 
+static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
+					dma_addr_t addr, size_t size)
+{
+	return addr + size <= mask;
+}
+
 #endif /* _ASM_IA64_DMA_MAPPING_H */
arch/x86/include/asm/dma-mapping.h +6 -0
index 916cbb6..52b8d36 100644
@@ -309,4 +309,10 @@ static inline void dma_free_coherent(struct device *dev, size_t size,
 		ops->free_coherent(dev, size, vaddr, bus);
 }
 
+static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
+					dma_addr_t addr, size_t size)
+{
+	return addr + size <= mask;
+}
+
 #endif
arch/x86/kernel/pci-dma.c +1 -1
index 745579b..efb0bd6 100644
@@ -145,7 +145,7 @@ again:
 		return NULL;
 
 	addr = page_to_phys(page);
-	if (!is_buffer_dma_capable(dma_mask, addr, size)) {
+	if (!is_buffer_dma_capable(dev, dma_mask, addr, size)) {
 		__free_pages(page, get_order(size));
 
 		if (dma_mask < DMA_BIT_MASK(32) && !(flag & GFP_DMA)) {
arch/x86/kernel/pci-gart_64.c +2 -2
index 1e8920d..13f5265 100644
@@ -191,13 +191,13 @@ static inline int
 need_iommu(struct device *dev, unsigned long addr, size_t size)
 {
 	return force_iommu ||
-		!is_buffer_dma_capable(*dev->dma_mask, addr, size);
+		!is_buffer_dma_capable(dev, *dev->dma_mask, addr, size);
 }
 
 static inline int
 nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
 {
-	return !is_buffer_dma_capable(*dev->dma_mask, addr, size);
+	return !is_buffer_dma_capable(dev, *dev->dma_mask, addr, size);
 }
 
 /* Map a single continuous physical area into the IOMMU.
arch/x86/kernel/pci-nommu.c +1 -1
index 71d412a..df930f3 100644
@@ -14,7 +14,7 @@
 static int
 check_addr(char *name, struct device *hwdev, dma_addr_t bus, size_t size)
 {
-	if (hwdev && !is_buffer_dma_capable(*hwdev->dma_mask, bus, size)) {
+	if (hwdev && !is_buffer_dma_capable(hwdev, *hwdev->dma_mask, bus, size)) {
 		if (*hwdev->dma_mask >= DMA_BIT_MASK(32))
 			printk(KERN_ERR
 			    "nommu_%s: overflow %Lx+%zu of device mask %Lx\n",
include/linux/dma-mapping.h +0 -5
index 8083b6a..85dafa1 100644
@@ -96,11 +96,6 @@ static inline int is_device_dma_capable(struct device *dev)
 	return dev->dma_mask != NULL && *dev->dma_mask != DMA_MASK_NONE;
 }
 
-static inline int is_buffer_dma_capable(u64 mask, dma_addr_t addr, size_t size)
-{
-	return addr + size <= mask;
-}
-
 #ifdef CONFIG_HAS_DMA
 #include <asm/dma-mapping.h>
 #else
lib/swiotlb.c +8 -31
index bffe6d7..01c5775 100644
@@ -145,17 +145,6 @@ void * __weak swiotlb_bus_to_virt(struct device *hwdev, dma_addr_t address)
 	return phys_to_virt(swiotlb_bus_to_phys(hwdev, address));
 }
 
-int __weak swiotlb_arch_address_needs_mapping(struct device *hwdev,
-					       dma_addr_t addr, size_t size)
-{
-	return !is_buffer_dma_capable(dma_get_mask(hwdev), addr, size);
-}
-
-int __weak swiotlb_arch_range_needs_mapping(phys_addr_t paddr, size_t size)
-{
-	return 0;
-}
-
 static void swiotlb_print_info(unsigned long bytes)
 {
 	phys_addr_t pstart, pend;
@@ -315,17 +304,6 @@ cleanup1:
 	return -ENOMEM;
 }
 
-static inline int
-address_needs_mapping(struct device *hwdev, dma_addr_t addr, size_t size)
-{
-	return swiotlb_arch_address_needs_mapping(hwdev, addr, size);
-}
-
-static inline int range_needs_mapping(phys_addr_t paddr, size_t size)
-{
-	return swiotlb_force || swiotlb_arch_range_needs_mapping(paddr, size);
-}
-
 static int is_swiotlb_buffer(char *addr)
 {
 	return addr >= io_tlb_start && addr < io_tlb_end;
@@ -561,9 +539,8 @@ swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 		dma_mask = hwdev->coherent_dma_mask;
 
 	ret = (void *)__get_free_pages(flags, order);
-	if (ret &&
-	    !is_buffer_dma_capable(dma_mask, swiotlb_virt_to_bus(hwdev, ret),
-				   size)) {
+	if (ret && !is_buffer_dma_capable(hwdev, dma_mask,
+					  swiotlb_virt_to_bus(hwdev, ret), size)) {
 		/*
 		 * The allocated memory isn't reachable by the device.
 		 */
@@ -585,7 +562,7 @@ swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 	dev_addr = swiotlb_virt_to_bus(hwdev, ret);
 
 	/* Confirm address can be DMA'd by device */
-	if (!is_buffer_dma_capable(dma_mask, dev_addr, size)) {
+	if (!is_buffer_dma_capable(hwdev, dma_mask, dev_addr, size)) {
 		printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
 		       (unsigned long long)dma_mask,
 		       (unsigned long long)dev_addr);
@@ -655,8 +632,8 @@ dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
 	 * we can safely return the device addr and not worry about bounce
 	 * buffering it.
 	 */
-	if (!address_needs_mapping(dev, dev_addr, size) &&
-	    !range_needs_mapping(phys, size))
+	if (is_buffer_dma_capable(dev, dma_get_mask(dev), dev_addr, size) &&
+	    !swiotlb_force)
 		return dev_addr;
 
 	/*
@@ -673,7 +650,7 @@ dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
 	/*
 	 * Ensure that the address returned is DMA'ble
 	 */
-	if (address_needs_mapping(dev, dev_addr, size))
+	if (!is_buffer_dma_capable(dev, dma_get_mask(dev), dev_addr, size))
 		panic("map_single: bounce buffer is not DMA'ble");
 
 	return dev_addr;
@@ -819,8 +796,8 @@ swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
 		phys_addr_t paddr = sg_phys(sg);
 		dma_addr_t dev_addr = swiotlb_phys_to_bus(hwdev, paddr);
 
-		if (range_needs_mapping(paddr, sg->length) ||
-		    address_needs_mapping(hwdev, dev_addr, sg->length)) {
+		if (!is_buffer_dma_capable(hwdev, dma_get_mask(hwdev), paddr, sg->length) ||
+		    swiotlb_force) {
 			void *map = map_single(hwdev, sg_phys(sg),
 					       sg->length, dir);
 			if (!map) {
Signature · 1 line
1.6.0.6
#12 lore raw
Quoted 5 lines · FUJITA Tomonori

What I want to remove all the __weak hacks and use the architecture abstraction. For example, the following patch is killing swiotlb_arch_address_needs_mapping() and swiotlb_arch_range_needs_mapping().

I think the swiotlb_arch_address_needs_mapping()/is_buffer_dma_capable() aspects of this are fine from the Xen POV but a hook along the lines of swiotlb_arch_range_needs_mapping() is unfortunately still required to handle potential discontigousness in multipage buffers. There's no reason this can't be handled in a similar way though. e.g. the following updated patch is based on yours but also moves the swiotlb_range_needs mapping hook to dma-mapping.h. The corresponding the Xen updates will.

Subject: swiotlb: is_buffer_dma_capable and swiotlb_range_needs_mapping are arch-specific

This moves is_buffer_dma_capable() from include/linux/dma-mapping.h to arch/*/include/asm/dma-mapping.h because it's architecture-specific; we shouldn't have added it in the generic place.

This function is used only in swiotlb (supported by x86 and IA64, and POWERPC shortly).

POWERPC needs struct device to see if an address is DMA-capable or not. How POWERPC implements is_buffer_dma_capable() is still under discussion. So this patch doesn't add POWERPC's one.

The range_needs_mapping hook is needed by Xen PCI to support multipage buffers which are potentially discontiguous in the DMA address space.

Based on original patch by FUJITA Tomonori.

Cc FUJITA Tomonori <[email protected]>
arch/ia64/include/asm/dma-mapping.h +11 -0
index 36c0009..cc25a4a 100644
@@ -174,4 +174,15 @@ dma_cache_sync (struct device *dev, void *vaddr, size_t size,
 
 #define dma_is_consistent(d, h)	(1)	/* all we do is coherent memory... */
 
+static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
+					dma_addr_t addr, size_t size)
+{
+	return addr + size <= mask;
+}
+
+static int inline swiotlb_range_needs_mapping(phys_addr_t paddr, size_t size)
+{
+       return 0;
+}
+
 #endif /* _ASM_IA64_DMA_MAPPING_H */
arch/x86/include/asm/dma-mapping.h +11 -0
index 916cbb6..a80139a 100644
@@ -309,4 +309,15 @@ static inline void dma_free_coherent(struct device *dev, size_t size,
 		ops->free_coherent(dev, size, vaddr, bus);
 }
 
+static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
+					dma_addr_t addr, size_t size)
+{
+	return addr + size <= mask;
+}
+
+static int inline swiotlb_range_needs_mapping(phys_addr_t paddr, size_t size)
+{
+	return 0;
+}
+
 #endif
arch/x86/kernel/pci-dma.c +1 -1
index 2fffc22..587cc6a 100644
@@ -146,7 +146,7 @@ again:
 		return NULL;
 
 	addr = page_to_phys(page);
-	if (!is_buffer_dma_capable(dma_mask, addr, size)) {
+	if (!is_buffer_dma_capable(dev, dma_mask, addr, size)) {
 		__free_pages(page, get_order(size));
 
 		if (dma_mask < DMA_BIT_MASK(32) && !(flag & GFP_DMA)) {
arch/x86/kernel/pci-gart_64.c +2 -2
index 1e8920d..13f5265 100644
@@ -191,13 +191,13 @@ static inline int
 need_iommu(struct device *dev, unsigned long addr, size_t size)
 {
 	return force_iommu ||
-		!is_buffer_dma_capable(*dev->dma_mask, addr, size);
+		!is_buffer_dma_capable(dev, *dev->dma_mask, addr, size);
 }
 
 static inline int
 nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
 {
-	return !is_buffer_dma_capable(*dev->dma_mask, addr, size);
+	return !is_buffer_dma_capable(dev, *dev->dma_mask, addr, size);
 }
 
 /* Map a single continuous physical area into the IOMMU.
arch/x86/kernel/pci-nommu.c +1 -1
index 71d412a..df930f3 100644
@@ -14,7 +14,7 @@
 static int
 check_addr(char *name, struct device *hwdev, dma_addr_t bus, size_t size)
 {
-	if (hwdev && !is_buffer_dma_capable(*hwdev->dma_mask, bus, size)) {
+	if (hwdev && !is_buffer_dma_capable(hwdev, *hwdev->dma_mask, bus, size)) {
 		if (*hwdev->dma_mask >= DMA_BIT_MASK(32))
 			printk(KERN_ERR
 			    "nommu_%s: overflow %Lx+%zu of device mask %Lx\n",
include/linux/dma-mapping.h +0 -5
index 8083b6a..85dafa1 100644
@@ -96,11 +96,6 @@ static inline int is_device_dma_capable(struct device *dev)
 	return dev->dma_mask != NULL && *dev->dma_mask != DMA_MASK_NONE;
 }
 
-static inline int is_buffer_dma_capable(u64 mask, dma_addr_t addr, size_t size)
-{
-	return addr + size <= mask;
-}
-
 #ifdef CONFIG_HAS_DMA
 #include <asm/dma-mapping.h>
 #else
include/linux/swiotlb.h +0 -2
index cb1a663..32f4fa4 100644
@@ -32,8 +32,6 @@ extern dma_addr_t swiotlb_phys_to_bus(struct device *hwdev,
 extern phys_addr_t swiotlb_bus_to_phys(struct device *hwdev,
 				       dma_addr_t address);
 
-extern int swiotlb_arch_range_needs_mapping(phys_addr_t paddr, size_t size);
-
 extern void
 *swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 			dma_addr_t *dma_handle, gfp_t flags);
lib/swiotlb.c +7 -25
index cec5f62..bde376c 100644
@@ -147,17 +147,6 @@ void * __weak swiotlb_bus_to_virt(struct device *hwdev, dma_addr_t address)
 	return phys_to_virt(swiotlb_bus_to_phys(hwdev, address));
 }
 
-int __weak swiotlb_arch_address_needs_mapping(struct device *hwdev,
-					       dma_addr_t addr, size_t size)
-{
-	return !is_buffer_dma_capable(dma_get_mask(hwdev), addr, size);
-}
-
-int __weak swiotlb_arch_range_needs_mapping(phys_addr_t paddr, size_t size)
-{
-	return 0;
-}
-
 static void swiotlb_print_info(unsigned long bytes)
 {
 	phys_addr_t pstart, pend;
@@ -318,15 +307,9 @@ cleanup1:
 	return -ENOMEM;
 }
 
-static inline int
-address_needs_mapping(struct device *hwdev, dma_addr_t addr, size_t size)
-{
-	return swiotlb_arch_address_needs_mapping(hwdev, addr, size);
-}
-
 static inline int range_needs_mapping(phys_addr_t paddr, size_t size)
 {
-	return swiotlb_force || swiotlb_arch_range_needs_mapping(paddr, size);
+	return swiotlb_force || swiotlb_range_needs_mapping(paddr, size);
 }
 
 static int is_swiotlb_buffer(char *addr)
@@ -564,9 +547,8 @@ swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 		dma_mask = hwdev->coherent_dma_mask;
 
 	ret = (void *)__get_free_pages(flags, order);
-	if (ret &&
-	    !is_buffer_dma_capable(dma_mask, swiotlb_virt_to_bus(hwdev, ret),
-				   size)) {
+	if (ret && !is_buffer_dma_capable(hwdev, dma_mask,
+					  swiotlb_virt_to_bus(hwdev, ret), size)) {
 		/*
 		 * The allocated memory isn't reachable by the device.
 		 */
@@ -588,7 +570,7 @@ swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 	dev_addr = swiotlb_virt_to_bus(hwdev, ret);
 
 	/* Confirm address can be DMA'd by device */
-	if (!is_buffer_dma_capable(dma_mask, dev_addr, size)) {
+	if (!is_buffer_dma_capable(hwdev, dma_mask, dev_addr, size)) {
 		printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
 		       (unsigned long long)dma_mask,
 		       (unsigned long long)dev_addr);
@@ -658,7 +640,7 @@ dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
 	 * we can safely return the device addr and not worry about bounce
 	 * buffering it.
 	 */
-	if (!address_needs_mapping(dev, dev_addr, size) &&
+	if (is_buffer_dma_capable(dev, dma_get_mask(dev), dev_addr, size) &&
 	    !range_needs_mapping(phys, size))
 		return dev_addr;
 
@@ -676,7 +658,7 @@ dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
 	/*
 	 * Ensure that the address returned is DMA'ble
 	 */
-	if (address_needs_mapping(dev, dev_addr, size))
+	if (!is_buffer_dma_capable(dev, dma_get_mask(dev), dev_addr, size))
 		panic("map_single: bounce buffer is not DMA'ble");
 
 	return dev_addr;
@@ -823,7 +805,7 @@ swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
 		dma_addr_t dev_addr = swiotlb_phys_to_bus(hwdev, paddr);
 
 		if (range_needs_mapping(paddr, sg->length) ||
-		    address_needs_mapping(hwdev, dev_addr, sg->length)) {
+		    !is_buffer_dma_capable(hwdev, dma_get_mask(hwdev), dev_addr, sg->length)) {
 			void *map = map_single(hwdev, sg_phys(sg),
 					       sg->length, dir);
 			if (!map) {
Signature · 4 lines
Ian Campbell
Current Noise: Dark Fortress - Edge Of Night

Yow!  Is this sexual intercourse yet??  Is it, huh, is it??
#13 lore raw
Quoted 1 line · Ian Campbell

The corresponding the Xen updates will...

...follow.

Subject:swiotlb/xen: add implementation of swiotlb_range_needs_mapping hook

This function is now implemented via asm/dma-mapping.h rather than as a weak hook in swiotlb.c

arch/x86/include/asm/dma-mapping.h +7 -1
index a80139a..ed51bd1 100644
@@ -315,9 +315,15 @@ static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
 	return addr + size <= mask;
 }
 
+#ifdef CONFIG_PCI_XEN
+extern int xen_range_needs_mapping(phys_addr_t paddr, size_t size);
+#else
+static inline int xen_range_needs_mapping(phys_addr_t paddr, size_t size) { return 0; }
+#endif
+
 static int inline swiotlb_range_needs_mapping(phys_addr_t paddr, size_t size)
 {
-	return 0;
+	return xen_range_needs_mapping(paddr, size);
 }
 
 #endif
arch/x86/xen/pci-swiotlb.c +0 -8
index 5a46418..e23b00b 100644
@@ -43,11 +43,3 @@ phys_addr_t swiotlb_bus_to_phys(struct device *hwdev, dma_addr_t baddr)
 
 	return baddr;
 }
-
-int swiotlb_arch_range_needs_mapping(phys_addr_t paddr, size_t size)
-{
-	if (xen_pv_domain())
-		return xen_range_needs_mapping(paddr, size);
-
-	return 0;
-}
drivers/pci/xen-iommu.c +2 -0
index 41c276f..003ffe3 100644
@@ -132,6 +132,8 @@ static int range_straddles_page_boundary(phys_addr_t p, size_t size)
 
 int xen_range_needs_mapping(phys_addr_t paddr, size_t size)
 {
+	if (!xen_pv_domain())
+		return 0;
 	return range_straddles_page_boundary(paddr, size);
 }
 
Signature · 6 lines
Ian Campbell
Current Noise: Isis - Wills Dissolve

"Hey Ivan, check your six."
-- Sidewinder missile jacket patch, showing a Sidewinder driving up the tail
 of a Russian Su-27
#14 lore raw
Quoted 16 lines · On Thu, 21 May 2009 11:28:53 +0100 Ian Campbell <[email protected]> wrote:
Quoted 1 line · Ian Campbell

The corresponding the Xen updates will...

...follow.

Subject:swiotlb/xen: add implementation of swiotlb_range_needs_mapping hook

This function is now implemented via asm/dma-mapping.h rather than as a weak hook in swiotlb.c

arch/x86/include/asm/dma-mapping.h +0 -0
index a80139a..ed51bd1 100644

arch/x86/include/asm/dma-mapping.h static inline int is_buffer_dma_capable(struct device *dev, u64 mask, · line 322

Quoted 8 lines
arch/x86/include/asm/dma-mapping.h +5 -0
@@ -315,9 +315,15 @@ static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
 	return addr + size <= mask;
 }
 
+#ifdef CONFIG_PCI_XEN
+extern int xen_range_needs_mapping(phys_addr_t paddr, size_t size);
+#else
+static inline int xen_range_needs_mapping(phys_addr_t paddr, size_t size) { return 0; }
+#endif

I know Xen can do something like this but you think that this is clean?

In addition, you also the similar hack in arch/ia64/include/asm/dma-mapping.h for ia64's dom0 support, I think.

IMO, your patch just moves the ugly hacks from lib/swiotlb.c to arch/{x86|ia64}/include/asm/dma-mapping.h.

#15 lore raw
Quoted 11 lines · FUJITA Tomonori

arch/x86/include/asm/dma-mapping.h static inline int is_buffer_dma_capable(struct device *dev, u64 mask, · line 322

Quoted 5 lines · On Thu, 21 May 2009 11:28:53 +0100 Ian Campbell <[email protected]> wrote:
arch/x86/include/asm/dma-mapping.h +5 -0
+#ifdef CONFIG_PCI_XEN
+extern int xen_range_needs_mapping(phys_addr_t paddr, size_t size);
+#else
+static inline int xen_range_needs_mapping(phys_addr_t paddr, size_t size) { return 0; }
+#endif

I know Xen can do something like this but you think that this is clean?

Well, defining a static inline function when a CONFIG option is disabled is fairly idiomatic in the kernel and in general hiding these sorts of things in the headers in this way is preferred to having them in .c files. See e.g. the handling of CONFIG_PRINTK in include/linux/kernel.h or CONFIG_HIGHMEM in include/linux/highmem.h for just two examples out of many.

Quoted 5 lines

In addition, you also the similar hack in arch/ia64/include/asm/dma-mapping.h for ia64's dom0 support, I think.

IMO, your patch just moves the ugly hacks from lib/swiotlb.c to arch/{x86|ia64}/include/asm/dma-mapping.h.

I nearly suggested that for this hook it might actually be preferable to put the one line Xen hook directly into swiotlb.c. I didn't think this suggestion would go down very well though.

In any case something along these lines needs to go somewhere. I think you are slightly mischaracterising this as an "ugly hack" -- it is a necessary interface to enable a particular use-case, and it actually has a very small cross section (it's basically five or six lines of code).

If there was a cleaner way to achieve the same result we would of course go with that. I don't think duplicating swiotlb.c, as has been suggested as the alternative, just for that one hook point makes sense.

Ian.

Signature · 6 lines
Ian Campbell
Current Noise: Isis - Altered Course

"For a male and female to live continuously together is...  biologically
speaking, an extremely unnatural condition."
		-- Robert Briffault
#16 lore raw
Quoted 12 lines · Ian Campbell
Quoted 5 lines · FUJITA Tomonori

Review

Quoted 2 lines · On Thu, 21 May 2009 11:28:53 +0100 Ian Campbell <[email protected]> wrote:
(quoted diff) +2 -0
+#ifdef CONFIG_PCI_XEN
+extern int xen_range_needs_mapping(phys_addr_t paddr, size_t

size);

Quoted 2 lines

Review

Quoted 2 lines
(quoted diff) +2 -0
+#else
+static inline int xen_range_needs_mapping(phys_addr_t paddr,

size_t size) { return 0; }

Quoted 2 lines
Quoted 1 line

+#endif

Quoted 7 lines
Quoted 2 lines

I know Xen can do something like this but you think that this is clean?

Well, defining a static inline function when a CONFIG option is disabled is fairly idiomatic in the kernel and in general hiding these sorts of things in the headers in this way is preferred to having them in .c files.

Although I do concede that the function definition would probably be better placed in a xen specific header.

Ian.

#17 lore raw
In reply to Ian Campbell
Quoted 13 lines · On Thu, 21 May 2009 12:03:05 +0100 Ian Campbell <[email protected]> wrote:
Quoted 11 lines · FUJITA Tomonori

arch/x86/include/asm/dma-mapping.h static inline int is_buffer_dma_capable(struct device *dev, u64 mask, · line 322

Quoted 5 lines · On Thu, 21 May 2009 11:28:53 +0100 Ian Campbell <[email protected]> wrote:
arch/x86/include/asm/dma-mapping.h +5 -0
+#ifdef CONFIG_PCI_XEN
+extern int xen_range_needs_mapping(phys_addr_t paddr, size_t size);
+#else
+static inline int xen_range_needs_mapping(phys_addr_t paddr, size_t size) { return 0; }
+#endif

I know Xen can do something like this but you think that this is clean?

Quoted 6 lines

Well, defining a static inline function when a CONFIG option is disabled is fairly idiomatic in the kernel and in general hiding these sorts of things in the headers in this way is preferred to having them in .c files. See e.g. the handling of CONFIG_PRINTK in include/linux/kernel.h or CONFIG_HIGHMEM in include/linux/highmem.h for just two examples out of many.

Well, I know that it's idiomatic, but placing CONFIG_PCI_XEN in arch/{x86|ia64}/include/asm/ is a wrong abstraction to me.

Quoted 9 lines
Quoted 5 lines

In addition, you also the similar hack in arch/ia64/include/asm/dma-mapping.h for ia64's dom0 support, I think.

IMO, your patch just moves the ugly hacks from lib/swiotlb.c to arch/{x86|ia64}/include/asm/dma-mapping.h.

I nearly suggested that for this hook it might actually be preferable to put the one line Xen hook directly into swiotlb.c. I didn't think this suggestion would go down very well though.

Any arch or Xen specific code should not live in swiotlb.c

Quoted 4 lines

In any case something along these lines needs to go somewhere. I think you are slightly mischaracterising this as an "ugly hack" -- it is a necessary interface to enable a particular use-case, and it actually has a very small cross section (it's basically five or six lines of code).

A necessary interface? Sorry, I don't buy it. It's necessary for only Xen. And it's not fit well for swiotlb and the architecture abstraction.

Quoted 3 lines

If there was a cleaner way to achieve the same result we would of course go with that. I don't think duplicating swiotlb.c, as has been suggested as the alternative, just for that one hook point makes sense.

One hook? You need to reread swiotlb.c. There are more. All should go.

#18 lore raw
Quoted 23 lines · FUJITA Tomonori
Quoted 13 lines · On Thu, 21 May 2009 12:03:05 +0100 Ian Campbell <[email protected]> wrote:
Quoted 11 lines · FUJITA Tomonori

arch/x86/include/asm/dma-mapping.h static inline int is_buffer_dma_capable(struct device *dev, u64 mask, · line 322

Quoted 5 lines · On Thu, 21 May 2009 11:28:53 +0100 Ian Campbell <[email protected]> wrote:
arch/x86/include/asm/dma-mapping.h +5 -0
+#ifdef CONFIG_PCI_XEN
+extern int xen_range_needs_mapping(phys_addr_t paddr, size_t size);
+#else
+static inline int xen_range_needs_mapping(phys_addr_t paddr, size_t size) { return 0; }
+#endif

I know Xen can do something like this but you think that this is clean?

Quoted 6 lines

Well, defining a static inline function when a CONFIG option is disabled is fairly idiomatic in the kernel and in general hiding these sorts of things in the headers in this way is preferred to having them in .c files. See e.g. the handling of CONFIG_PRINTK in include/linux/kernel.h or CONFIG_HIGHMEM in include/linux/highmem.h for just two examples out of many.

Quoted 2 lines

Well, I know that it's idiomatic, but placing CONFIG_PCI_XEN in arch/{x86|ia64}/include/asm/ is a wrong abstraction to me.

Agreed, include/xen/swiotlb.h is the right place for it.

Quoted 11 lines
Quoted 9 lines
Quoted 5 lines

In addition, you also the similar hack in arch/ia64/include/asm/dma-mapping.h for ia64's dom0 support, I think.

IMO, your patch just moves the ugly hacks from lib/swiotlb.c to arch/{x86|ia64}/include/asm/dma-mapping.h.

I nearly suggested that for this hook it might actually be preferable to put the one line Xen hook directly into swiotlb.c. I didn't think this suggestion would go down very well though.

Any arch or Xen specific code should not live in swiotlb.c

Again agreed, which is why I didn't suggest it ;-)

Quoted 10 lines
Quoted 4 lines

In any case something along these lines needs to go somewhere. I think you are slightly mischaracterising this as an "ugly hack" -- it is a necessary interface to enable a particular use-case, and it actually has a very small cross section (it's basically five or six lines of code).

A necessary interface? Sorry, I don't buy it. It's necessary for only Xen. And it's not fit well for swiotlb and the architecture abstraction.

I said "necessary interface to enable a particular use-case". Xen is a valid use case -- I appreciate that you personally may have no interest in it but plenty of people do and plenty of people would like to see it working in the mainline kernels.

In terms of clean abstraction this is a architecture hook to allow mappings to be forced through the swiotlb. It is essentially a more flexible extension to the existing swiotlb_force flag, in fact swiotlb_force_mapping() might even be a better name for it.

IMHO, the benefits here (working Xen domain 0 plus PCI passthrough to guest domains) are well worth the costs.

Quoted 5 lines
Quoted 3 lines

If there was a cleaner way to achieve the same result we would of course go with that. I don't think duplicating swiotlb.c, as has been suggested as the alternative, just for that one hook point makes sense.

One hook? You need to reread swiotlb.c. There are more. All should go.

I meant that swiotlb_range_needs_mapping is the single contentious hook as far as I can tell. It is the only one which you appear to be disputing the very existence of (irrespective of whether it uses __weak or is moved into asm/dma-mapping.h). The other existing __weak hooks are useful to other users too and all can, AFAICT, be moved into asm/dma-mapping.h.

You have already dealt with moving swiotlb_address_needs_mapping and I am currently looking into the the phys_to_bus and bus_to_phys ones. That just leaves the alloc functions, which are next on my list after phys_to_bus and bus_to_phys. Will finishing up those patches resolve most of your objections are do you have others?

Ian.

Signature · 3 lines
Ian Campbell

This supersedes all previous notices.
#19 lore raw

swiotlb: remove __weak hooks in favour of architecture-specific functions

At the end of this series there are no more __weak functions in lib/swiotlb.c

The series adds several hook functions to the x86 architecture. Would they be preferred as a struct x86_swiotlb_ops or as individual hooks?

I was unsure what to do about powerpc in most places since the existing support seems to in-progress so it wasn't always clear where to put the implementation. If there is a tree somewhere with more complete support I'll be happy to provide additional patches.

Boot tested on x86 under xen but not even compiled for ia64 or powerpc. If someone can point me to a decent source of cross compilers I can sort that out. (http://www.kernel.org/pub/tools/crosstool/ seems to be out-of-date and only has ia64 in any case)

Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>

---

 arch/ia64/include/asm/dma-mapping.h    |   29 ++++++++++++++
 arch/ia64/kernel/pci-swiotlb.c         |   11 +++++
 arch/powerpc/include/asm/dma-mapping.h |    5 ++
 arch/x86/include/asm/agp.h             |    4 +-
 arch/x86/include/asm/dma-mapping.h     |   34 +++++++++++++++++
 arch/x86/include/asm/xen/iommu.h       |    4 ++
 arch/x86/kernel/pci-dma.c              |    6 ++-
 arch/x86/kernel/pci-gart_64.c          |    4 +-
 arch/x86/kernel/pci-nommu.c            |    3 +-
 arch/x86/kernel/pci-swiotlb.c          |   32 ++++++++++++++++
 arch/x86/xen/Makefile                  |    3 +-
 arch/x86/xen/pci-swiotlb.c             |   53 --------------------------
 drivers/pci/xen-iommu.c                |   20 ++++++++--
 include/linux/dma-mapping.h            |    5 --
 include/linux/swiotlb.h                |   10 -----
 include/xen/swiotlb.h                  |    5 --
 lib/swiotlb.c                          |   64 +++++++-------------------------
 17 files changed, 156 insertions(+), 136 deletions(-)
#20 lore raw

Re: swiotlb: remove __weak hooks in favour of architecture-specific functions

Quoted 2 lines · Ian Campbell

At the end of this series there are no more __weak functions in lib/swiotlb.c

Hmm, I was expecting the patches to be numbered in the subject by default. The correct order is:

swiotlb: make is_buffer_dma_capable architecture-specific
swiotlb: make range_needs_mapping architecture-specific swiotlb/xen: update xen for swiotlb_arch_force_mapping changes
swiotlb: make swiotlb allocation functions architecture-specific swiotlb/xen: update xen for changes to swiotlb allocation interface
swiotlb: make swiotlb phys<->bus translations architecture-specific swiotlb/xen: update xen swiotlb for phys<->bus API changes
xen: remove arch/x86/xen/pci-swiotlb.c

Also xen-devel missed out due to a typo, sorry. See the LKML archives if you are interested.

Ian.

Quoted 44 lines

The series adds several hook functions to the x86 architecture. Would they be preferred as a struct x86_swiotlb_ops or as individual hooks?

I was unsure what to do about powerpc in most places since the existing support seems to in-progress so it wasn't always clear where to put the implementation. If there is a tree somewhere with more complete support I'll be happy to provide additional patches.

Boot tested on x86 under xen but not even compiled for ia64 or powerpc. If someone can point me to a decent source of cross compilers I can sort that out. (http://www.kernel.org/pub/tools/crosstool/ seems to be out-of-date and only has ia64 in any case)

Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>

---

 arch/ia64/include/asm/dma-mapping.h    |   29 ++++++++++++++
 arch/ia64/kernel/pci-swiotlb.c         |   11 +++++
 arch/powerpc/include/asm/dma-mapping.h |    5 ++
 arch/x86/include/asm/agp.h             |    4 +-
 arch/x86/include/asm/dma-mapping.h     |   34 +++++++++++++++++
 arch/x86/include/asm/xen/iommu.h       |    4 ++
 arch/x86/kernel/pci-dma.c              |    6 ++-
 arch/x86/kernel/pci-gart_64.c          |    4 +-
 arch/x86/kernel/pci-nommu.c            |    3 +-
 arch/x86/kernel/pci-swiotlb.c          |   32 ++++++++++++++++
 arch/x86/xen/Makefile                  |    3 +-
 arch/x86/xen/pci-swiotlb.c             |   53 --------------------------
 drivers/pci/xen-iommu.c                |   20 ++++++++--
 include/linux/dma-mapping.h            |    5 --
 include/linux/swiotlb.h                |   10 -----
 include/xen/swiotlb.h                  |    5 --
 lib/swiotlb.c                          |   64 +++++++-------------------------
 17 files changed, 156 insertions(+), 136 deletions(-)
#21 lore raw

Re: swiotlb: remove __weak hooks in favour of architecture-specific functions

In reply to Ian Campbell
Quoted 11 lines · Ian Campbell

At the end of this series there are no more __weak functions in lib/swiotlb.c

The series adds several hook functions to the x86 architecture. Would they be preferred as a struct x86_swiotlb_ops or as individual hooks?

I was unsure what to do about powerpc in most places since the existing support seems to in-progress so it wasn't always clear where to put the implementation. If there is a tree somewhere with more complete support I'll be happy to provide additional patches.

Quoted 4 lines

Boot tested on x86 under xen but not even compiled for ia64 or powerpc. If someone can point me to a decent source of cross compilers I can sort that out. (http://www.kernel.org/pub/tools/crosstool/ seems to be out-of-date and only has ia64 in any case)

Try these. They are fairly current. I have used several of them.

http://bakeyournoodle.com/cross/

Signature · 3 lines
~Randy
LPC 2009, Sept. 23-25, Portland, Oregon
http://linuxplumbersconf.org/2009/
#22 lore raw

Re: swiotlb: remove __weak hooks in favour of architecture-specific functions

Quoted 3 lines · Randy Dunlap

Try these. They are fairly current. I have used several of them.

http://bakeyournoodle.com/cross/

Thanks for the pointer, these look ideal.

Ian.

#23 lore raw

Re: swiotlb: remove __weak hooks in favour of architecture-specific functions

In reply to Ian Campbell
Quoted 16 lines · On Thu, 21 May 2009 17:15:21 +0100 Ian Campbell <[email protected]> wrote:

At the end of this series there are no more __weak functions in lib/swiotlb.c

The series adds several hook functions to the x86 architecture. Would they be preferred as a struct x86_swiotlb_ops or as individual hooks?

I was unsure what to do about powerpc in most places since the existing support seems to in-progress so it wasn't always clear where to put the implementation. If there is a tree somewhere with more complete support I'll be happy to provide additional patches.

Boot tested on x86 under xen but not even compiled for ia64 or powerpc. If someone can point me to a decent source of cross compilers I can sort that out. (http://www.kernel.org/pub/tools/crosstool/ seems to be out-of-date and only has ia64 in any case)

Quoted 10 lines
Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>

Sorry, Nack.

As I wrote in another mail, this patch makes the code more difficult to understand; it just moves the hacks in lib/swiotlb.c somewhere else in a strange way instead of killing them.

Please go with the following way (that I posted yesterday):

http://marc.info/?l=xen-devel&m=124292666214380&w=2

Export the core feature of swiotlb, managing iotlb buffer and implement the Xen mapping functions. With that approach, there is not much code duplication and there is no need for ugly hooks for dom0; the phys/bus address conversion and address checking.

#24 lore raw

Re: swiotlb: remove __weak hooks in favour of architecture-specific functions

Quoted 9 lines · FUJITA Tomonori

On Thu, 21 May 2009 17:15:21 +0100
Ian Campbell <[email protected]> wrote:
Please go with the following way (that I posted yesterday):

http://marc.info/?l=xen-devel&m=124292666214380&w=2

Export the core feature of swiotlb, managing iotlb buffer and implement the Xen mapping functions.

I feel that should be a last resort, before we go down that path we should try and find a way for us to use the generic code in a clean way which makes everyone happy.

We have had several attempts at this and admittedly have not yet come up with something that satisfies everyone but I don't really think we have gotten to the point of admitting defeat and just duplicating the code.

I think the proposal to use a dma_map_range-like function which I sent a few minutes ago I think gets us closer to something which satisfies everyone's requirements, including yours for a clean abstraction.

Quoted 3 lines

With that approach, there is not
much code duplication and there is no need for ugly hooks for dom0; the phys/bus address conversion and address checking.

The phys/bus address conversion is also needed for powerpc.

I think the two address checking functions can be collapsed into a single one which satisfies the needs of both Xen and powerpc.

What dom0 specific "ugly" hooks does that leave? The alloc one? I've discussed that in another mail.

Ian.

#25 lore raw

Re: swiotlb: remove __weak hooks in favour of architecture-specific functions

Quoted 11 lines · On Fri, 22 May 2009 12:43:16 +0100 Ian Campbell <[email protected]>…
Quoted 9 lines · FUJITA Tomonori

On Thu, 21 May 2009 17:15:21 +0100
Ian Campbell <[email protected]> wrote:
Please go with the following way (that I posted yesterday):

http://marc.info/?l=xen-devel&m=124292666214380&w=2

Export the core feature of swiotlb, managing iotlb buffer and implement the Xen mapping functions.

Quoted 7 lines

I feel that should be a last resort, before we go down that path we should try and find a way for us to use the generic code in a clean way which makes everyone happy.

We have had several attempts at this and admittedly have not yet come up with something that satisfies everyone but I don't really think we have gotten to the point of admitting defeat and just duplicating the code.

There should not be much duplication.

Quoted 3 lines

I think the proposal to use a dma_map_range-like function which I sent a few minutes ago I think gets us closer to something which satisfies everyone's requirements, including yours for a clean abstraction.

Seems you don't understand the point; with dom0, we can't cleanly use arch/*/include/asm/.

You need to insert Xen's hook like this:

+++ b/arch/x86/include/asm/dma-mapping.h
@@ -309,4 +309,20 @@ static inline void dma_free_coherent(struct device *dev, size_t size,
ops->free_coherent(dev, size, vaddr, bus);
}

+static inline dma_addr_t dma_map_range(struct device *dev, u64 mask,
+ phys_addr_t addr, size_t size)
+{
+ dma_addr_t dma_addr;
+#ifdef CONFIG_XEN
+ extern int xen_range_needs_mapping(phys_addr_t paddr, size_t size);
+ if (xen_pv_domain() && xen_range_needs_mapping(addr, size))
+ return 0;
+#endif
+
+ dma_addr = swiotlb_phys_to_bus(dev, addr);
+ if (dma_addr + size <= mask)
+ return 0;
+ return dma_addr;
+}

Or you need to use a function pointer in a strange way.

arch/x86/include/asm/dma-mapping.h +9 -0
@@ -315,4 +315,13 @@ static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
 	return addr + size <= mask;
 }
 
+extern int (*x86_swiotlb_force_mapping)(phys_addr_t paddr, size_t size);
+
+static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
+{
+	if (x86_swiotlb_force_mapping)
+		return x86_swiotlb_force_mapping(paddr, size);
+	return 0;
+}
+
 

Or you could invent something more.

As you said in another mail, it's up to the x86 maintainer :

Quoted 2 lines

This case is internal to the x86 arch code and I'd really like to hear the x86 maintainer's opinion of the general approach.

But the above code looks really ugly to me.

Quoted 11 lines
Quoted 3 lines

With that approach, there is not
much code duplication and there is no need for ugly hooks for dom0; the phys/bus address conversion and address checking.

The phys/bus address conversion is also needed for powerpc.

I think the two address checking functions can be collapsed into a single one which satisfies the needs of both Xen and powerpc.

What dom0 specific "ugly" hooks does that leave? The alloc one? I've discussed that in another mail.

See above. POWERPC can use arch/*/include/asm/ cleanly for the phys/bus address conversion while dom0 can't. That's what I said again and again.

#26 lore raw

Re: swiotlb: remove __weak hooks in favour of architecture-specific functions

Quoted 25 lines · FUJITA Tomonori
Quoted 11 lines · On Fri, 22 May 2009 12:43:16 +0100 Ian Campbell <[email protected]>…
Quoted 9 lines · FUJITA Tomonori

On Thu, 21 May 2009 17:15:21 +0100
Ian Campbell <[email protected]> wrote:
Please go with the following way (that I posted yesterday):

http://marc.info/?l=xen-devel&m=124292666214380&w=2

Export the core feature of swiotlb, managing iotlb buffer and implement the Xen mapping functions.

Quoted 7 lines

I feel that should be a last resort, before we go down that path we should try and find a way for us to use the generic code in a clean way which makes everyone happy.

We have had several attempts at this and admittedly have not yet come up with something that satisfies everyone but I don't really think we have gotten to the point of admitting defeat and just duplicating the code.

There should not be much duplication.

Quoted 6 lines
Quoted 3 lines

I think the proposal to use a dma_map_range-like function which I sent a few minutes ago I think gets us closer to something which satisfies everyone's requirements, including yours for a clean abstraction.

Seems you don't understand the point; with dom0, we can't cleanly use arch/*/include/asm/.

I understand precisely what you are saying, I just fundamentally disagree with you. It is perfectly possible for an arch/*/include/asm interface for this stuff to be defined which is completely abstracted from the POV of the swiotlb code (and any other arch-external code).

Quoted 1 line

You need to insert Xen's hook like this:

As I said in the email this snippet was contained in:

Quoted 3 lines

* The Xen specific stuff in arch/x86/include/asm/dma-mapping.h clearly needs to be properly abstracted away (I just stuck it there for testing).

So obviously I am well aware that it is unacceptable as it stands.

I think we can find a way to implement this functionality which is contained entirely within the arch/x86 code and is also acceptable to the x86 maintainers. There are plenty of cases where we define similar interfaces where arch code implements an API with different backends for different configurations, hardware configurations etc etc.

Quoted 3 lines

See above. POWERPC can use arch/*/include/asm/ cleanly for the phys/bus address conversion while dom0 can't. That's what I said again and again.

And I dispute this claim, again and again.

Ian.

#27 lore raw

Re: swiotlb: remove __weak hooks in favour of architecture-specific functions

Quoted 38 lines · On Fri, 22 May 2009 15:02:04 +0100 Ian Campbell <[email protected]>…
Quoted 25 lines · FUJITA Tomonori
Quoted 11 lines · On Fri, 22 May 2009 12:43:16 +0100 Ian Campbell <[email protected]>…
Quoted 9 lines · FUJITA Tomonori

On Thu, 21 May 2009 17:15:21 +0100
Ian Campbell <[email protected]> wrote:
Please go with the following way (that I posted yesterday):

http://marc.info/?l=xen-devel&m=124292666214380&w=2

Export the core feature of swiotlb, managing iotlb buffer and implement the Xen mapping functions.

Quoted 7 lines

I feel that should be a last resort, before we go down that path we should try and find a way for us to use the generic code in a clean way which makes everyone happy.

We have had several attempts at this and admittedly have not yet come up with something that satisfies everyone but I don't really think we have gotten to the point of admitting defeat and just duplicating the code.

There should not be much duplication.

Quoted 6 lines
Quoted 3 lines

I think the proposal to use a dma_map_range-like function which I sent a few minutes ago I think gets us closer to something which satisfies everyone's requirements, including yours for a clean abstraction.

Seems you don't understand the point; with dom0, we can't cleanly use arch/*/include/asm/.

I understand precisely what you are saying, I just fundamentally disagree with you. It is perfectly possible for an arch/*/include/asm interface for this stuff to be defined which is completely abstracted from the POV of the swiotlb code (and any other arch-external code).

Quoted 8 lines
Quoted 1 line

You need to insert Xen's hook like this:

As I said in the email this snippet was contained in:

Quoted 3 lines

* The Xen specific stuff in arch/x86/include/asm/dma-mapping.h clearly needs to be properly abstracted away (I just stuck it there for testing).

So obviously I am well aware that it is unacceptable as it stands.

I don't think that you can't clean up the above much.

After all, you need a hook for Xen in arch/x86/include/asm/dma-mapping.h and define it in Xen's land.

That will not be much different from:

http://marc.info/?l=linux-kernel&m=124299344606890&w=2

And adding 'if xen' hook in common header files in arch/x86/include/ is a wrong design to me. Do you add a similar 'if xen' hook in other x86's common headers files?

Quoted 1 line

I think we can find a way to implement this functionality which is

k> contained entirely within the arch/x86 code and is also acceptable to

Quoted 1 line

the x86 maintainers. There are plenty of cases where we define similar

I'm not sure about it (hopefully, not). But if the x86 maintainers ACK the above code, I have no complaint.

Quoted 2 lines

interfaces where arch code implements an API with different backends for different configurations, hardware configurations etc etc.

#28 lore raw

PATCH 1/1 swiotlb: make is_buffer_dma_capable architecture-specific

In reply to Ian Campbell

This moves is_buffer_dma_capable() from include/linux/dma-mapping.h to arch/*/include/asm/dma-mapping.h because it's architecture-specific; we shouldn't have added it in the generic place.

This function is used only in swiotlb (supported by x86 and IA64, and POWERPC shortly).

POWERPC needs struct device to see if an address is DMA-capable or not. How POWERPC implements is_buffer_dma_capable() is still under discussion. So this patch doesn't add POWERPC's one.

Based on original patch by FUJITA Tomonori.

Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>
---
 arch/ia64/include/asm/dma-mapping.h |    6 ++++++
 arch/x86/include/asm/dma-mapping.h  |    6 ++++++
 arch/x86/kernel/pci-dma.c           |    2 +-
 arch/x86/kernel/pci-gart_64.c       |    4 ++--
 arch/x86/kernel/pci-nommu.c         |    3 ++-
 include/linux/dma-mapping.h         |    5 -----
 lib/swiotlb.c                       |   25 +++++++------------------
 7 files changed, 24 insertions(+), 27 deletions(-)
arch/ia64/include/asm/dma-mapping.h +6 -0
index 36c0009..a091648 100644
@@ -174,4 +174,10 @@ dma_cache_sync (struct device *dev, void *vaddr, size_t size,
 
 #define dma_is_consistent(d, h)	(1)	/* all we do is coherent memory... */
 
+static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
+					dma_addr_t addr, size_t size)
+{
+	return addr + size <= mask;
+}
+
 #endif /* _ASM_IA64_DMA_MAPPING_H */
arch/x86/include/asm/dma-mapping.h +6 -0
index 916cbb6..52b8d36 100644
@@ -309,4 +309,10 @@ static inline void dma_free_coherent(struct device *dev, size_t size,
 		ops->free_coherent(dev, size, vaddr, bus);
 }
 
+static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
+					dma_addr_t addr, size_t size)
+{
+	return addr + size <= mask;
+}
+
 #endif
arch/x86/kernel/pci-dma.c +1 -1
index 2fffc22..587cc6a 100644
@@ -146,7 +146,7 @@ again:
 		return NULL;
 
 	addr = page_to_phys(page);
-	if (!is_buffer_dma_capable(dma_mask, addr, size)) {
+	if (!is_buffer_dma_capable(dev, dma_mask, addr, size)) {
 		__free_pages(page, get_order(size));
 
 		if (dma_mask < DMA_BIT_MASK(32) && !(flag & GFP_DMA)) {
arch/x86/kernel/pci-gart_64.c +2 -2
index 1e8920d..13f5265 100644
@@ -191,13 +191,13 @@ static inline int
 need_iommu(struct device *dev, unsigned long addr, size_t size)
 {
 	return force_iommu ||
-		!is_buffer_dma_capable(*dev->dma_mask, addr, size);
+		!is_buffer_dma_capable(dev, *dev->dma_mask, addr, size);
 }
 
 static inline int
 nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
 {
-	return !is_buffer_dma_capable(*dev->dma_mask, addr, size);
+	return !is_buffer_dma_capable(dev, *dev->dma_mask, addr, size);
 }
 
 /* Map a single continuous physical area into the IOMMU.
arch/x86/kernel/pci-nommu.c +2 -1
index 71d412a..23fa16e 100644
@@ -14,7 +14,8 @@
 static int
 check_addr(char *name, struct device *hwdev, dma_addr_t bus, size_t size)
 {
-	if (hwdev && !is_buffer_dma_capable(*hwdev->dma_mask, bus, size)) {
+	if (hwdev &&
+	    !is_buffer_dma_capable(hwdev, *hwdev->dma_mask, bus, size)) {
 		if (*hwdev->dma_mask >= DMA_BIT_MASK(32))
 			printk(KERN_ERR
 			    "nommu_%s: overflow %Lx+%zu of device mask %Lx\n",
include/linux/dma-mapping.h +0 -5
index 8083b6a..85dafa1 100644
@@ -96,11 +96,6 @@ static inline int is_device_dma_capable(struct device *dev)
 	return dev->dma_mask != NULL && *dev->dma_mask != DMA_MASK_NONE;
 }
 
-static inline int is_buffer_dma_capable(u64 mask, dma_addr_t addr, size_t size)
-{
-	return addr + size <= mask;
-}
-
 #ifdef CONFIG_HAS_DMA
 #include <asm/dma-mapping.h>
 #else
lib/swiotlb.c +7 -18
index cec5f62..98726a2 100644
@@ -147,12 +147,6 @@ void * __weak swiotlb_bus_to_virt(struct device *hwdev, dma_addr_t address)
 	return phys_to_virt(swiotlb_bus_to_phys(hwdev, address));
 }
 
-int __weak swiotlb_arch_address_needs_mapping(struct device *hwdev,
-					       dma_addr_t addr, size_t size)
-{
-	return !is_buffer_dma_capable(dma_get_mask(hwdev), addr, size);
-}
-
 int __weak swiotlb_arch_range_needs_mapping(phys_addr_t paddr, size_t size)
 {
 	return 0;
@@ -318,12 +312,6 @@ cleanup1:
 	return -ENOMEM;
 }
 
-static inline int
-address_needs_mapping(struct device *hwdev, dma_addr_t addr, size_t size)
-{
-	return swiotlb_arch_address_needs_mapping(hwdev, addr, size);
-}
-
 static inline int range_needs_mapping(phys_addr_t paddr, size_t size)
 {
 	return swiotlb_force || swiotlb_arch_range_needs_mapping(paddr, size);
@@ -565,8 +553,8 @@ swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 
 	ret = (void *)__get_free_pages(flags, order);
 	if (ret &&
-	    !is_buffer_dma_capable(dma_mask, swiotlb_virt_to_bus(hwdev, ret),
-				   size)) {
+	    !is_buffer_dma_capable(hwdev, dma_mask,
+				   swiotlb_virt_to_bus(hwdev, ret), size)) {
 		/*
 		 * The allocated memory isn't reachable by the device.
 		 */
@@ -588,7 +576,7 @@ swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 	dev_addr = swiotlb_virt_to_bus(hwdev, ret);
 
 	/* Confirm address can be DMA'd by device */
-	if (!is_buffer_dma_capable(dma_mask, dev_addr, size)) {
+	if (!is_buffer_dma_capable(hwdev, dma_mask, dev_addr, size)) {
 		printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
 		       (unsigned long long)dma_mask,
 		       (unsigned long long)dev_addr);
@@ -658,7 +646,7 @@ dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
 	 * we can safely return the device addr and not worry about bounce
 	 * buffering it.
 	 */
-	if (!address_needs_mapping(dev, dev_addr, size) &&
+	if (is_buffer_dma_capable(dev, dma_get_mask(dev), dev_addr, size) &&
 	    !range_needs_mapping(phys, size))
 		return dev_addr;
 
@@ -676,7 +664,7 @@ dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
 	/*
 	 * Ensure that the address returned is DMA'ble
 	 */
-	if (address_needs_mapping(dev, dev_addr, size))
+	if (!is_buffer_dma_capable(dev, dma_get_mask(dev), dev_addr, size))
 		panic("map_single: bounce buffer is not DMA'ble");
 
 	return dev_addr;
@@ -823,7 +811,8 @@ swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
 		dma_addr_t dev_addr = swiotlb_phys_to_bus(hwdev, paddr);
 
 		if (range_needs_mapping(paddr, sg->length) ||
-		    address_needs_mapping(hwdev, dev_addr, sg->length)) {
+		    !is_buffer_dma_capable(hwdev, dma_get_mask(hwdev),
+					   dev_addr, sg->length)) {
 			void *map = map_single(hwdev, sg_phys(sg),
 					       sg->length, dir);
 			if (!map) {
Signature · 1 line
1.5.6.5
#29 lore raw

PATCH 1/1 swiotlb: make range_needs_mapping architecture-specific

In reply to Ian Campbell

This moves range_needs_mapping() from lib/swiotlb.c to arch/*/include/asm/dma-mapping.h and renames it swiotlb_arch_force_mapping() to more acurately reflect its usage.

Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>
---
 arch/ia64/include/asm/dma-mapping.h    |    5 +++++
 arch/powerpc/include/asm/dma-mapping.h |    5 +++++
 arch/x86/include/asm/dma-mapping.h     |    9 +++++++++
 arch/x86/kernel/pci-swiotlb.c          |    2 ++
 include/linux/swiotlb.h                |    2 --
 lib/swiotlb.c                          |   13 ++++---------
 6 files changed, 25 insertions(+), 11 deletions(-)
arch/ia64/include/asm/dma-mapping.h +5 -0
index a091648..3e8fb31 100644
@@ -180,4 +180,9 @@ static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
 	return addr + size <= mask;
 }
 
+static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
+{
+       return 0;
+}
+
 #endif /* _ASM_IA64_DMA_MAPPING_H */
arch/powerpc/include/asm/dma-mapping.h +5 -0
index c69f2b5..7e3510d 100644
@@ -429,5 +429,10 @@ static inline void dma_cache_sync(struct device *dev, void *vaddr, size_t size,
 	__dma_sync(vaddr, size, (int)direction);
 }
 
+static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
+{
+	return 0;
+}
+
 #endif /* __KERNEL__ */
 #endif	/* _ASM_DMA_MAPPING_H */
arch/x86/include/asm/dma-mapping.h +9 -0
index 52b8d36..aa9639f 100644
@@ -315,4 +315,13 @@ static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
 	return addr + size <= mask;
 }
 
+extern int (*x86_swiotlb_force_mapping)(phys_addr_t paddr, size_t size);
+
+static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
+{
+	if (x86_swiotlb_force_mapping)
+		return x86_swiotlb_force_mapping(paddr, size);
+	return 0;
+}
+
 #endif
arch/x86/kernel/pci-swiotlb.c +2 -0
index 7267376..d48912c 100644
@@ -15,6 +15,8 @@
 
 int swiotlb __read_mostly;
 
+int (*x86_swiotlb_force_mapping)(phys_addr_t paddr, size_t size);
+
 static void *x86_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 					dma_addr_t *dma_handle, gfp_t flags)
 {
include/linux/swiotlb.h +0 -2
index cb1a663..32f4fa4 100644
@@ -32,8 +32,6 @@ extern dma_addr_t swiotlb_phys_to_bus(struct device *hwdev,
 extern phys_addr_t swiotlb_bus_to_phys(struct device *hwdev,
 				       dma_addr_t address);
 
-extern int swiotlb_arch_range_needs_mapping(phys_addr_t paddr, size_t size);
-
 extern void
 *swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 			dma_addr_t *dma_handle, gfp_t flags);
lib/swiotlb.c +4 -9
index 98726a2..d311654 100644
@@ -147,11 +147,6 @@ void * __weak swiotlb_bus_to_virt(struct device *hwdev, dma_addr_t address)
 	return phys_to_virt(swiotlb_bus_to_phys(hwdev, address));
 }
 
-int __weak swiotlb_arch_range_needs_mapping(phys_addr_t paddr, size_t size)
-{
-	return 0;
-}
-
 static void swiotlb_print_info(unsigned long bytes)
 {
 	phys_addr_t pstart, pend;
@@ -312,9 +307,9 @@ cleanup1:
 	return -ENOMEM;
 }
 
-static inline int range_needs_mapping(phys_addr_t paddr, size_t size)
+static inline int force_mapping(phys_addr_t paddr, size_t size)
 {
-	return swiotlb_force || swiotlb_arch_range_needs_mapping(paddr, size);
+	return swiotlb_force || swiotlb_force_mapping(paddr, size);
 }
 
 static int is_swiotlb_buffer(char *addr)
@@ -647,7 +642,7 @@ dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
 	 * buffering it.
 	 */
 	if (is_buffer_dma_capable(dev, dma_get_mask(dev), dev_addr, size) &&
-	    !range_needs_mapping(phys, size))
+	    !force_mapping(phys, size))
 		return dev_addr;
 
 	/*
@@ -810,7 +805,7 @@ swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
 		phys_addr_t paddr = sg_phys(sg);
 		dma_addr_t dev_addr = swiotlb_phys_to_bus(hwdev, paddr);
 
-		if (range_needs_mapping(paddr, sg->length) ||
+		if (force_mapping(paddr, sg->length) ||
 		    !is_buffer_dma_capable(hwdev, dma_get_mask(hwdev),
 					   dev_addr, sg->length)) {
 			void *map = map_single(hwdev, sg_phys(sg),
Signature · 1 line
1.5.6.5
#30 lore raw

Re: [PATCH] swiotlb: make range_needs_mapping architecture-specific

Quoted 42 lines · On Thu, 21 May 2009 17:15:23 +0100 Ian Campbell <[email protected]> wrote:

This moves range_needs_mapping() from lib/swiotlb.c to arch/*/include/asm/dma-mapping.h and renames it swiotlb_arch_force_mapping() to more acurately reflect its usage.

Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>
---
 arch/ia64/include/asm/dma-mapping.h    |    5 +++++
 arch/powerpc/include/asm/dma-mapping.h |    5 +++++
 arch/x86/include/asm/dma-mapping.h     |    9 +++++++++
 arch/x86/kernel/pci-swiotlb.c          |    2 ++
 include/linux/swiotlb.h                |    2 --
 lib/swiotlb.c                          |   13 ++++---------
 6 files changed, 25 insertions(+), 11 deletions(-)
arch/ia64/include/asm/dma-mapping.h +5 -0
index a091648..3e8fb31 100644
@@ -180,4 +180,9 @@ static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
 	return addr + size <= mask;
 }
 
+static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
+{
+       return 0;
+}
+
 #endif /* _ASM_IA64_DMA_MAPPING_H */
arch/powerpc/include/asm/dma-mapping.h +0 -0
index c69f2b5..7e3510d 100644

arch/powerpc/include/asm/dma-mapping.h static inline void dma_cache_sync(struct device *dev, void *vaddr, size_t size, · line 436

Quoted 8 lines
arch/powerpc/include/asm/dma-mapping.h +5 -0
@@ -429,5 +429,10 @@ static inline void dma_cache_sync(struct device *dev, void *vaddr, size_t size,
 	__dma_sync(vaddr, size, (int)direction);
 }
 
+static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
+{
+	return 0;
+}
+

Adding a swiotlb specific function to asm/dma-mapping.h is wrong.

Quoted 38 lines
arch/powerpc/include/asm/dma-mapping.h +0 -0
 #endif /* __KERNEL__ */
 #endif	/* _ASM_DMA_MAPPING_H */
arch/x86/include/asm/dma-mapping.h +9 -0
index 52b8d36..aa9639f 100644
@@ -315,4 +315,13 @@ static inline int is_buffer_dma_capable(struct device *dev, u64 mask,
 	return addr + size <= mask;
 }
 
+extern int (*x86_swiotlb_force_mapping)(phys_addr_t paddr, size_t size);
+
+static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
+{
+	if (x86_swiotlb_force_mapping)
+		return x86_swiotlb_force_mapping(paddr, size);
+	return 0;
+}
+
 #endif
arch/x86/kernel/pci-swiotlb.c +2 -0
index 7267376..d48912c 100644
@@ -15,6 +15,8 @@
 
 int swiotlb __read_mostly;
 
+int (*x86_swiotlb_force_mapping)(phys_addr_t paddr, size_t size);
+
 static void *x86_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 					dma_addr_t *dma_handle, gfp_t flags)
 {
include/linux/swiotlb.h +0 -0
index cb1a663..32f4fa4 100644

include/linux/swiotlb.h extern dma_addr_t swiotlb_phys_to_bus(struct device *hwdev, · line 33

Quoted 2 lines
include/linux/swiotlb.h +0 -0
@@ -32,8 +32,6 @@ extern dma_addr_t swiotlb_phys_to_bus(struct device *hwdev,
 extern phys_addr_t swiotlb_bus_to_phys(struct device *hwdev,
 				       dma_addr_t address);

Ditto.

And using a function pointer for the architecture abstraction is worse than __weak.

Quoted 59 lines
include/linux/swiotlb.h +0 -2
-extern int swiotlb_arch_range_needs_mapping(phys_addr_t paddr, size_t size);
-
 extern void
 *swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 			dma_addr_t *dma_handle, gfp_t flags);
lib/swiotlb.c +4 -9
index 98726a2..d311654 100644
@@ -147,11 +147,6 @@ void * __weak swiotlb_bus_to_virt(struct device *hwdev, dma_addr_t address)
 	return phys_to_virt(swiotlb_bus_to_phys(hwdev, address));
 }
 
-int __weak swiotlb_arch_range_needs_mapping(phys_addr_t paddr, size_t size)
-{
-	return 0;
-}
-
 static void swiotlb_print_info(unsigned long bytes)
 {
 	phys_addr_t pstart, pend;
@@ -312,9 +307,9 @@ cleanup1:
 	return -ENOMEM;
 }
 
-static inline int range_needs_mapping(phys_addr_t paddr, size_t size)
+static inline int force_mapping(phys_addr_t paddr, size_t size)
 {
-	return swiotlb_force || swiotlb_arch_range_needs_mapping(paddr, size);
+	return swiotlb_force || swiotlb_force_mapping(paddr, size);
 }
 
 static int is_swiotlb_buffer(char *addr)
@@ -647,7 +642,7 @@ dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
 	 * buffering it.
 	 */
 	if (is_buffer_dma_capable(dev, dma_get_mask(dev), dev_addr, size) &&
-	    !range_needs_mapping(phys, size))
+	    !force_mapping(phys, size))
 		return dev_addr;
 
 	/*
@@ -810,7 +805,7 @@ swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
 		phys_addr_t paddr = sg_phys(sg);
 		dma_addr_t dev_addr = swiotlb_phys_to_bus(hwdev, paddr);
 
-		if (range_needs_mapping(paddr, sg->length) ||
+		if (force_mapping(paddr, sg->length) ||
 		    !is_buffer_dma_capable(hwdev, dma_get_mask(hwdev),
 					   dev_addr, sg->length)) {
 			void *map = map_single(hwdev, sg_phys(sg),
Signature · 7 lines
1.5.6.5

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/
#31 lore raw

Re: [PATCH] swiotlb: make range_needs_mapping architecture-specific

Quoted 9 lines · FUJITA Tomonori

arch/ia64/include/asm/dma-mapping.h static inline int is_buffer_dma_capable(struct device *dev, u64 mask, · line 187

Quoted 5 lines · On Thu, 21 May 2009 17:15:23 +0100 Ian Campbell <[email protected]> wrote:
arch/ia64/include/asm/dma-mapping.h +5 -0
+static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
+{
+	return 0;
+}
+

Adding a swiotlb specific function to asm/dma-mapping.h is wrong.

This one is unnecessary with the dma_map_range proposal.

Quoted 2 lines
Quoted 2 lines
arch/x86/kernel/pci-swiotlb.c +2 -0
+int (*x86_swiotlb_force_mapping)(phys_addr_t paddr, size_t size);
+

[...]

Quoted 2 lines

And using a function pointer for the architecture abstraction is worse than __weak.

This specific hook is unnecessary with the dma_map_range proposal but in general we use function pointers quite extensively for abstraction in the kernel.

This case is internal to the x86 arch code and I'd really like to hear the x86 maintainer's opinion of the general approach.

Ian.

#32 lore raw

PATCH 1/1 swiotlb/xen: update xen for swiotlb_arch_force_mapping changes

In reply to Ian Campbell

This function is now implemented via a hook in arch/x86/include/asm/dma-mapping.h rather than as a weak hook in swiotlb.c

Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>
---
 arch/x86/xen/pci-swiotlb.c |    8 --------
 drivers/pci/xen-iommu.c    |    4 +++-
 include/xen/swiotlb.h      |    1 -
 3 files changed, 3 insertions(+), 10 deletions(-)
arch/x86/xen/pci-swiotlb.c +0 -8
index 5a46418..e23b00b 100644
@@ -43,11 +43,3 @@ phys_addr_t swiotlb_bus_to_phys(struct device *hwdev, dma_addr_t baddr)
 
 	return baddr;
 }
-
-int swiotlb_arch_range_needs_mapping(phys_addr_t paddr, size_t size)
-{
-	if (xen_pv_domain())
-		return xen_range_needs_mapping(paddr, size);
-
-	return 0;
-}
drivers/pci/xen-iommu.c +3 -1
index 41c276f..f65660a 100644
@@ -130,7 +130,7 @@ static int range_straddles_page_boundary(phys_addr_t p, size_t size)
 	return 1;
 }
 
-int xen_range_needs_mapping(phys_addr_t paddr, size_t size)
+static int xen_swiotlb_force_mapping(phys_addr_t paddr, size_t size)
 {
 	return range_straddles_page_boundary(paddr, size);
 }
@@ -324,6 +324,8 @@ void __init xen_iommu_init(void)
 	force_iommu = 0;
 	dma_ops = &xen_dma_ops;
 
+	x86_swiotlb_force_mapping = &xen_swiotlb_force_mapping;
+
 	if (swiotlb) {
 		printk(KERN_INFO "Xen: Enabling DMA fallback to swiotlb\n");
 		dma_ops = &xen_swiotlb_dma_ops;
include/xen/swiotlb.h +0 -1
index 75d1da1..64137fa 100644
@@ -4,7 +4,6 @@
 extern void xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs);
 extern phys_addr_t xen_bus_to_phys(dma_addr_t daddr);
 extern dma_addr_t xen_phys_to_bus(phys_addr_t paddr);
-extern int xen_range_needs_mapping(phys_addr_t phys, size_t size);
 
 #ifdef CONFIG_PCI_XEN
 extern int xen_wants_swiotlb(void);
Signature · 1 line
1.5.6.5
#33 lore raw

PATCH 1/1 swiotlb: make swiotlb allocation functions architecture-specific

In reply to Ian Campbell

Some architectures need to allocate memory to be used as a swiotlb in a special manner.

Also make the swiotlb_late_init_with_default_size() function only available on architectures which require it.

Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>
---
 arch/ia64/include/asm/dma-mapping.h |    5 +++++
 arch/ia64/kernel/pci-swiotlb.c      |   11 +++++++++++
 arch/x86/include/asm/dma-mapping.h  |    4 ++++
 arch/x86/kernel/pci-swiotlb.c       |   12 ++++++++++++
 arch/x86/xen/pci-swiotlb.c          |   17 -----------------
 include/linux/swiotlb.h             |    3 ---
 lib/swiotlb.c                       |   12 ++----------
 7 files changed, 34 insertions(+), 30 deletions(-)
arch/ia64/include/asm/dma-mapping.h +5 -0
index 3e8fb31..9f9994d 100644
@@ -185,4 +185,9 @@ static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
        return 0;
 }
 
+#define SWIOTLB_WANT_LATE_INIT 1
+
+extern void *swiotlb_alloc_boot(size_t size, unsigned long nslabs);
+extern void *swiotlb_alloc(unsigned order, unsigned long nslabs);
+
 #endif /* _ASM_IA64_DMA_MAPPING_H */
arch/ia64/kernel/pci-swiotlb.c +11 -0
index 285aae8..1b42f21 100644
@@ -4,6 +4,7 @@
 #include <linux/cache.h>
 #include <linux/module.h>
 #include <linux/dma-mapping.h>
+#include <linux/bootmem.h>
 
 #include <asm/swiotlb.h>
 #include <asm/dma.h>
@@ -13,6 +14,16 @@
 int swiotlb __read_mostly;
 EXPORT_SYMBOL(swiotlb);
 
+void * __init swiotlb_alloc_boot(size_t size, unsigned long nslabs)
+{
+	return alloc_bootmem_low_pages(size);
+}
+
+void *swiotlb_alloc(unsigned order, unsigned long nslabs)
+{
+	return (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN, order);
+}
+
 static void *ia64_swiotlb_alloc_coherent(struct device *dev, size_t size,
 					 dma_addr_t *dma_handle, gfp_t gfp)
 {
arch/x86/include/asm/dma-mapping.h +4 -0
index aa9639f..5a3180d 100644
@@ -324,4 +324,8 @@ static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
 	return 0;
 }
 
+extern void (*x86_swiotlb_alloc_fixup)(void *buf, size_t size,
+				       unsigned long nslabs);
+extern void *swiotlb_alloc_boot(size_t size, unsigned long nslabs);
+
 #endif
arch/x86/kernel/pci-swiotlb.c +12 -0
index d48912c..7de9fdd 100644
@@ -6,6 +6,7 @@
 #include <linux/swiotlb.h>
 #include <linux/bootmem.h>
 #include <linux/dma-mapping.h>
+#include <linux/bootmem.h>
 
 #include <asm/iommu.h>
 #include <asm/swiotlb.h>
@@ -16,6 +17,17 @@
 int swiotlb __read_mostly;
 
 int (*x86_swiotlb_force_mapping)(phys_addr_t paddr, size_t size);
+void __initdata (*x86_swiotlb_alloc_fixup)(void *buf, size_t size, unsigned long nslabs);
+
+void * __init swiotlb_alloc_boot(size_t size, unsigned long nslabs)
+{
+	void *ret = alloc_bootmem_low_pages(size);
+
+	if (ret && x86_swiotlb_alloc_fixup)
+		x86_swiotlb_alloc_fixup(ret, size, nslabs);
+
+	return ret;
+}
 
 static void *x86_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 					dma_addr_t *dma_handle, gfp_t flags)
arch/x86/xen/pci-swiotlb.c +0 -17
index e23b00b..dcf2ef8 100644
@@ -11,23 +11,6 @@
  * implementations in lib/swiotlb.c.
  */
 
-void * __init swiotlb_alloc_boot(size_t size, unsigned long nslabs)
-{
-	void *ret = alloc_bootmem_low_pages(size);
-
-	if (ret && xen_pv_domain())
-		xen_swiotlb_fixup(ret, size, nslabs);
-
-	return ret;
-}
-
-void *swiotlb_alloc(unsigned order, unsigned long nslabs)
-{
-	/* Never called on x86.  Warn, just in case. */
-	WARN_ON(1);
-	return NULL;
-}
-
 dma_addr_t swiotlb_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
 {
 	if (xen_pv_domain())
include/linux/swiotlb.h +0 -3
index 32f4fa4..be8b77d 100644
@@ -24,9 +24,6 @@ struct scatterlist;
 extern void
 swiotlb_init(void);
 
-extern void *swiotlb_alloc_boot(size_t bytes, unsigned long nslabs);
-extern void *swiotlb_alloc(unsigned order, unsigned long nslabs);
-
 extern dma_addr_t swiotlb_phys_to_bus(struct device *hwdev,
 				      phys_addr_t address);
 extern phys_addr_t swiotlb_bus_to_phys(struct device *hwdev,
lib/swiotlb.c +2 -10
index d311654..640c2f1 100644
@@ -114,16 +114,6 @@ setup_io_tlb_npages(char *str)
 __setup("swiotlb=", setup_io_tlb_npages);
 /* make io_tlb_overflow tunable too? */
 
-void * __weak __init swiotlb_alloc_boot(size_t size, unsigned long nslabs)
-{
-	return alloc_bootmem_low_pages(size);
-}
-
-void * __weak swiotlb_alloc(unsigned order, unsigned long nslabs)
-{
-	return (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN, order);
-}
-
 dma_addr_t __weak swiotlb_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
 {
 	return paddr;
@@ -213,6 +203,7 @@ swiotlb_init(void)
 	swiotlb_init_with_default_size(64 * (1<<20));	/* default to 64MB */
 }
 
+#ifdef SWIOTLB_WANT_LATE_INIT
 /*
  * Systems with larger DMA zones (those that don't support ISA) can
  * initialize the swiotlb later using the slab allocator if needed.
@@ -306,6 +297,7 @@ cleanup1:
 	io_tlb_nslabs = req_nslabs;
 	return -ENOMEM;
 }
+#endif
 
 static inline int force_mapping(phys_addr_t paddr, size_t size)
 {
Signature · 1 line
1.5.6.5
#34 lore raw

Re: [PATCH] swiotlb: make swiotlb allocation functions architecture-specific

Quoted 2 lines · On Thu, 21 May 2009 17:15:25 +0100 Ian Campbell <[email protected]> wrote:

Some architectures need to allocate memory to be used as a swiotlb in a special manner.

Hmm, what architectures need a special manner? I guess only Xen.

Quoted 196 lines

Also make the swiotlb_late_init_with_default_size() function only available on architectures which require it.

Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>
---
 arch/ia64/include/asm/dma-mapping.h |    5 +++++
 arch/ia64/kernel/pci-swiotlb.c      |   11 +++++++++++
 arch/x86/include/asm/dma-mapping.h  |    4 ++++
 arch/x86/kernel/pci-swiotlb.c       |   12 ++++++++++++
 arch/x86/xen/pci-swiotlb.c          |   17 -----------------
 include/linux/swiotlb.h             |    3 ---
 lib/swiotlb.c                       |   12 ++----------
 7 files changed, 34 insertions(+), 30 deletions(-)
arch/ia64/include/asm/dma-mapping.h +5 -0
index 3e8fb31..9f9994d 100644
@@ -185,4 +185,9 @@ static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
        return 0;
 }
 
+#define SWIOTLB_WANT_LATE_INIT 1
+
+extern void *swiotlb_alloc_boot(size_t size, unsigned long nslabs);
+extern void *swiotlb_alloc(unsigned order, unsigned long nslabs);
+
 #endif /* _ASM_IA64_DMA_MAPPING_H */
arch/ia64/kernel/pci-swiotlb.c +11 -0
index 285aae8..1b42f21 100644
@@ -4,6 +4,7 @@
 #include <linux/cache.h>
 #include <linux/module.h>
 #include <linux/dma-mapping.h>
+#include <linux/bootmem.h>
 
 #include <asm/swiotlb.h>
 #include <asm/dma.h>
@@ -13,6 +14,16 @@
 int swiotlb __read_mostly;
 EXPORT_SYMBOL(swiotlb);
 
+void * __init swiotlb_alloc_boot(size_t size, unsigned long nslabs)
+{
+	return alloc_bootmem_low_pages(size);
+}
+
+void *swiotlb_alloc(unsigned order, unsigned long nslabs)
+{
+	return (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN, order);
+}
+
 static void *ia64_swiotlb_alloc_coherent(struct device *dev, size_t size,
 					 dma_addr_t *dma_handle, gfp_t gfp)
 {
arch/x86/include/asm/dma-mapping.h +4 -0
index aa9639f..5a3180d 100644
@@ -324,4 +324,8 @@ static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
 	return 0;
 }
 
+extern void (*x86_swiotlb_alloc_fixup)(void *buf, size_t size,
+				       unsigned long nslabs);
+extern void *swiotlb_alloc_boot(size_t size, unsigned long nslabs);
+
 #endif
arch/x86/kernel/pci-swiotlb.c +12 -0
index d48912c..7de9fdd 100644
@@ -6,6 +6,7 @@
 #include <linux/swiotlb.h>
 #include <linux/bootmem.h>
 #include <linux/dma-mapping.h>
+#include <linux/bootmem.h>
 
 #include <asm/iommu.h>
 #include <asm/swiotlb.h>
@@ -16,6 +17,17 @@
 int swiotlb __read_mostly;
 
 int (*x86_swiotlb_force_mapping)(phys_addr_t paddr, size_t size);
+void __initdata (*x86_swiotlb_alloc_fixup)(void *buf, size_t size, unsigned long nslabs);
+
+void * __init swiotlb_alloc_boot(size_t size, unsigned long nslabs)
+{
+	void *ret = alloc_bootmem_low_pages(size);
+
+	if (ret && x86_swiotlb_alloc_fixup)
+		x86_swiotlb_alloc_fixup(ret, size, nslabs);
+
+	return ret;
+}
 
 static void *x86_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 					dma_addr_t *dma_handle, gfp_t flags)
arch/x86/xen/pci-swiotlb.c +0 -17
index e23b00b..dcf2ef8 100644
@@ -11,23 +11,6 @@
  * implementations in lib/swiotlb.c.
  */
 
-void * __init swiotlb_alloc_boot(size_t size, unsigned long nslabs)
-{
-	void *ret = alloc_bootmem_low_pages(size);
-
-	if (ret && xen_pv_domain())
-		xen_swiotlb_fixup(ret, size, nslabs);
-
-	return ret;
-}
-
-void *swiotlb_alloc(unsigned order, unsigned long nslabs)
-{
-	/* Never called on x86.  Warn, just in case. */
-	WARN_ON(1);
-	return NULL;
-}
-
 dma_addr_t swiotlb_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
 {
 	if (xen_pv_domain())
include/linux/swiotlb.h +0 -3
index 32f4fa4..be8b77d 100644
@@ -24,9 +24,6 @@ struct scatterlist;
 extern void
 swiotlb_init(void);
 
-extern void *swiotlb_alloc_boot(size_t bytes, unsigned long nslabs);
-extern void *swiotlb_alloc(unsigned order, unsigned long nslabs);
-
 extern dma_addr_t swiotlb_phys_to_bus(struct device *hwdev,
 				      phys_addr_t address);
 extern phys_addr_t swiotlb_bus_to_phys(struct device *hwdev,
lib/swiotlb.c +2 -10
index d311654..640c2f1 100644
@@ -114,16 +114,6 @@ setup_io_tlb_npages(char *str)
 __setup("swiotlb=", setup_io_tlb_npages);
 /* make io_tlb_overflow tunable too? */
 
-void * __weak __init swiotlb_alloc_boot(size_t size, unsigned long nslabs)
-{
-	return alloc_bootmem_low_pages(size);
-}
-
-void * __weak swiotlb_alloc(unsigned order, unsigned long nslabs)
-{
-	return (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN, order);
-}
-
 dma_addr_t __weak swiotlb_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
 {
 	return paddr;
@@ -213,6 +203,7 @@ swiotlb_init(void)
 	swiotlb_init_with_default_size(64 * (1<<20));	/* default to 64MB */
 }
 
+#ifdef SWIOTLB_WANT_LATE_INIT
 /*
  * Systems with larger DMA zones (those that don't support ISA) can
  * initialize the swiotlb later using the slab allocator if needed.
@@ -306,6 +297,7 @@ cleanup1:
 	io_tlb_nslabs = req_nslabs;
 	return -ENOMEM;
 }
+#endif
 
 static inline int force_mapping(phys_addr_t paddr, size_t size)
 {
Signature · 7 lines
1.5.6.5

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/
#35 lore raw

Re: [PATCH] swiotlb: make swiotlb allocation functions architecture-specific

Quoted 7 lines · FUJITA Tomonori
Quoted 2 lines · On Thu, 21 May 2009 17:15:25 +0100 Ian Campbell <[email protected]> wrote:

Some architectures need to allocate memory to be used as a swiotlb in a special manner.

Hmm, what architectures need a special manner? I guess only Xen.

The x86 architecture, when running as a paravirtualised guest under Xen, needs it. I guess ia64 might need it for similar reasons, I'm not sure.

How does the fact that x86-Xen is (currently) the only user invalidate the requirement? You seem to have a knee-jerk reaction to anything which might be useful to Xen and I really don't understand why that should be the case.

We are talking about an initialisation path and we are not talking about changes which completely obscure all meaning or anything. The semantics of the API are pretty clear, I think.

Ian.

#36 lore raw

PATCH 1/1 swiotlb/xen: update xen for changes to swiotlb allocation interface

In reply to Ian Campbell

This function is now implemented via an explicit hook in arch/x86/kernel/pci-swiotlb.c rather than as a weak hook in swiotlb.c

Initialsation of the hook needs to happen before xen_iommu_init() is called so add detect_xen_iommu() (mirroring other IOMMU implementations) which is called early enough.

Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>
---
 arch/x86/include/asm/xen/iommu.h |    4 ++++
 arch/x86/kernel/pci-dma.c        |    4 +++-
 drivers/pci/xen-iommu.c          |   14 +++++++++++---
 include/xen/swiotlb.h            |    1 -
 4 files changed, 18 insertions(+), 5 deletions(-)
arch/x86/include/asm/xen/iommu.h +4 -0
index 75df312..22b6091 100644
@@ -1,8 +1,12 @@
 #ifndef ASM_X86__XEN_IOMMU_H
 
 #ifdef CONFIG_PCI_XEN
+extern void detect_xen_iommu(void);
 extern void xen_iommu_init(void);
 #else
+static inline void detect_xen_iommu(void)
+{
+}
 static inline void xen_iommu_init(void)
 {
 }
arch/x86/kernel/pci-dma.c +3 -1
index 587cc6a..695f8a6 100644
@@ -121,6 +121,8 @@ void __init pci_iommu_alloc(void)
 	 */
 	gart_iommu_hole_init();
 
+	detect_xen_iommu();
+
 	detect_calgary();
 
 	detect_intel_iommu();
@@ -277,7 +279,7 @@ static int __init pci_iommu_init(void)
 #endif
 
 	xen_iommu_init();
-	
+
 	calgary_iommu_init();
 
 	intel_iommu_init();
drivers/pci/xen-iommu.c +11 -3
index f65660a..2d727d1 100644
@@ -39,7 +39,8 @@ do {							\
 
 static int max_dma_bits = 32;
 
-void xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
+static void __init xen_swiotlb_alloc_fixup(void *buf, size_t size,
+					   unsigned long nslabs)
 {
 	int i, rc;
 	int dma_bits;
@@ -314,6 +315,15 @@ static struct dma_map_ops xen_swiotlb_dma_ops = {
 	.is_phys = 0,
 };
 
+void __init detect_xen_iommu(void)
+{
+	if (!xen_pv_domain())
+		return;
+
+	x86_swiotlb_force_mapping = &xen_swiotlb_force_mapping;
+	x86_swiotlb_alloc_fixup = &xen_swiotlb_alloc_fixup;
+}
+
 void __init xen_iommu_init(void)
 {
 	if (!xen_pv_domain())
@@ -324,8 +334,6 @@ void __init xen_iommu_init(void)
 	force_iommu = 0;
 	dma_ops = &xen_dma_ops;
 
-	x86_swiotlb_force_mapping = &xen_swiotlb_force_mapping;
-
 	if (swiotlb) {
 		printk(KERN_INFO "Xen: Enabling DMA fallback to swiotlb\n");
 		dma_ops = &xen_swiotlb_dma_ops;
include/xen/swiotlb.h +0 -1
index 64137fa..83ec002 100644
@@ -1,7 +1,6 @@
 #ifndef _XEN_SWIOTLB_H
 #define _XEN_SWIOTLB_H
 
-extern void xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs);
 extern phys_addr_t xen_bus_to_phys(dma_addr_t daddr);
 extern dma_addr_t xen_phys_to_bus(phys_addr_t paddr);
 
Signature · 1 line
1.5.6.5
#37 lore raw

PATCH 1/1 swiotlb: make swiotlb phys<->bus translations architecture-specific

In reply to Ian Campbell

This moves swiotlb_bus_to_phys() and swiotlb_phys_to_bus() from lib/swiotlb.c to arch/*/include/asm/dma-mapping.h. On x86 gart_to_phys and phys_to_gart translations need an exported function for use from drivers so introduce common x86_ variants to implement the interface required by both.

The __weak swiotlb_bus_to_virt is not currently overriden by any architecture therefore it can be made static again. I imagine the ability to override phys<->bus translations will be sufficient for any potential users anyhow, otherwise we can revisit this aspect.

Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>
---
 arch/ia64/include/asm/dma-mapping.h |   13 +++++++++++++
 arch/x86/include/asm/agp.h          |    4 ++--
 arch/x86/include/asm/dma-mapping.h  |   15 +++++++++++++++
 arch/x86/kernel/pci-swiotlb.c       |   18 ++++++++++++++++++
 arch/x86/xen/pci-swiotlb.c          |   15 ---------------
 include/linux/swiotlb.h             |    5 -----
 lib/swiotlb.c                       |   14 +-------------
 7 files changed, 49 insertions(+), 35 deletions(-)
arch/ia64/include/asm/dma-mapping.h +13 -0
index 9f9994d..667d04e 100644
@@ -190,4 +190,17 @@ static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
 extern void *swiotlb_alloc_boot(size_t size, unsigned long nslabs);
 extern void *swiotlb_alloc(unsigned order, unsigned long nslabs);
 
+static inline dma_addr_t swiotlb_phys_to_bus(struct device *hwdev,
+					     phys_addr_t paddr)
+{
+	return paddr;
+}
+
+static inline phys_addr_t swiotlb_bus_to_phys(struct device *hwdev,
+					      dma_addr_t baddr)
+{
+	return baddr;
+}
+
+
 #endif /* _ASM_IA64_DMA_MAPPING_H */
arch/x86/include/asm/agp.h +2 -2
index d972b14..085de80 100644
@@ -26,8 +26,8 @@
 #define flush_agp_cache() wbinvd()
 
 /* Convert a physical address to an address suitable for the GART. */
-#define phys_to_gart(x) swiotlb_phys_to_bus(NULL, (x))
-#define gart_to_phys(x) swiotlb_bus_to_phys(NULL, (x))
+#define phys_to_gart(x) x86_phys_to_bus(NULL, (x))
+#define gart_to_phys(x) x86_bus_to_phys(NULL, (x))
 
 /* GATT allocation. Returns/accepts GATT kernel virtual address. */
 #define alloc_gatt_pages(order)	({                                          \
arch/x86/include/asm/dma-mapping.h +15 -0
index 5a3180d..bccf196 100644
@@ -328,4 +328,19 @@ extern void (*x86_swiotlb_alloc_fixup)(void *buf, size_t size,
 				       unsigned long nslabs);
 extern void *swiotlb_alloc_boot(size_t size, unsigned long nslabs);
 
+extern dma_addr_t (*x86_phys_to_bus)(struct device *hwdev, phys_addr_t paddr);
+extern phys_addr_t (*x86_bus_to_phys)(struct device *hwdev, dma_addr_t baddr);
+
+static inline dma_addr_t swiotlb_phys_to_bus(struct device *hwdev,
+					     phys_addr_t paddr)
+{
+	return x86_phys_to_bus(hwdev, paddr);
+}
+
+static inline phys_addr_t swiotlb_bus_to_phys(struct device *hwdev,
+					      dma_addr_t baddr)
+{
+	return x86_bus_to_phys(hwdev, baddr);
+}
+
 #endif
arch/x86/kernel/pci-swiotlb.c +18 -0
index 7de9fdd..6cc1020 100644
@@ -29,6 +29,24 @@ void * __init swiotlb_alloc_boot(size_t size, unsigned long nslabs)
 	return ret;
 }
 
+static dma_addr_t __x86_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
+{
+	return paddr;
+}
+
+static phys_addr_t __x86_bus_to_phys(struct device *hwdev, dma_addr_t baddr)
+{
+	return baddr;
+}
+
+dma_addr_t (*x86_phys_to_bus)(struct device *hwdev,
+			      phys_addr_t paddr) = __x86_phys_to_bus;
+EXPORT_SYMBOL_GPL(x86_phys_to_bus);
+
+phys_addr_t (*x86_bus_to_phys)(struct device *hwdev,
+			       dma_addr_t baddr) = __x86_bus_to_phys;
+EXPORT_SYMBOL_GPL(x86_bus_to_phys);
+
 static void *x86_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 					dma_addr_t *dma_handle, gfp_t flags)
 {
arch/x86/xen/pci-swiotlb.c +0 -15
index dcf2ef8..92f04a3 100644
@@ -11,18 +11,3 @@
  * implementations in lib/swiotlb.c.
  */
 
-dma_addr_t swiotlb_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
-{
-	if (xen_pv_domain())
-		return xen_phys_to_bus(paddr);
-
-	return paddr;
-}
-
-phys_addr_t swiotlb_bus_to_phys(struct device *hwdev, dma_addr_t baddr)
-{
-	if (xen_pv_domain())
-		return xen_bus_to_phys(baddr);
-
-	return baddr;
-}
include/linux/swiotlb.h +0 -5
index be8b77d..b5b2245 100644
@@ -24,11 +24,6 @@ struct scatterlist;
 extern void
 swiotlb_init(void);
 
-extern dma_addr_t swiotlb_phys_to_bus(struct device *hwdev,
-				      phys_addr_t address);
-extern phys_addr_t swiotlb_bus_to_phys(struct device *hwdev,
-				       dma_addr_t address);
-
 extern void
 *swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 			dma_addr_t *dma_handle, gfp_t flags);
lib/swiotlb.c +1 -13
index 640c2f1..b5dcb68 100644
@@ -114,25 +114,13 @@ setup_io_tlb_npages(char *str)
 __setup("swiotlb=", setup_io_tlb_npages);
 /* make io_tlb_overflow tunable too? */
 
-dma_addr_t __weak swiotlb_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
-{
-	return paddr;
-}
-EXPORT_SYMBOL_GPL(swiotlb_phys_to_bus);
-
-phys_addr_t __weak swiotlb_bus_to_phys(struct device *hwdev, dma_addr_t baddr)
-{
-	return baddr;
-}
-EXPORT_SYMBOL_GPL(swiotlb_bus_to_phys);
-
 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
 				      volatile void *address)
 {
 	return swiotlb_phys_to_bus(hwdev, virt_to_phys(address));
 }
 
-void * __weak swiotlb_bus_to_virt(struct device *hwdev, dma_addr_t address)
+static void *swiotlb_bus_to_virt(struct device *hwdev, dma_addr_t address)
 {
 	return phys_to_virt(swiotlb_bus_to_phys(hwdev, address));
 }
Signature · 1 line
1.5.6.5
#38 lore raw

Re: [PATCH] swiotlb: make swiotlb phys<->bus translations architecture-specific

Quoted 42 lines · On Thu, 21 May 2009 17:15:27 +0100 Ian Campbell <[email protected]> wrote:

This moves swiotlb_bus_to_phys() and swiotlb_phys_to_bus() from lib/swiotlb.c to arch/*/include/asm/dma-mapping.h. On x86 gart_to_phys and phys_to_gart translations need an exported function for use from drivers so introduce common x86_ variants to implement the interface required by both.

The __weak swiotlb_bus_to_virt is not currently overriden by any architecture therefore it can be made static again. I imagine the ability to override phys<->bus translations will be sufficient for any potential users anyhow, otherwise we can revisit this aspect.

Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>
---
 arch/ia64/include/asm/dma-mapping.h |   13 +++++++++++++
 arch/x86/include/asm/agp.h          |    4 ++--
 arch/x86/include/asm/dma-mapping.h  |   15 +++++++++++++++
 arch/x86/kernel/pci-swiotlb.c       |   18 ++++++++++++++++++
 arch/x86/xen/pci-swiotlb.c          |   15 ---------------
 include/linux/swiotlb.h             |    5 -----
 lib/swiotlb.c                       |   14 +-------------
 7 files changed, 49 insertions(+), 35 deletions(-)
arch/ia64/include/asm/dma-mapping.h +3 -0
index 9f9994d..667d04e 100644
@@ -190,4 +190,17 @@ static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size)
 extern void *swiotlb_alloc_boot(size_t size, unsigned long nslabs);
 extern void *swiotlb_alloc(unsigned order, unsigned long nslabs);
 
+static inline dma_addr_t swiotlb_phys_to_bus(struct device *hwdev,
+					     phys_addr_t paddr)
+{

arch/ia64/include/asm/dma-mapping.h static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size) · line 205

Quoted 10 lines
arch/ia64/include/asm/dma-mapping.h +10 -0
+	return paddr;
+}
+
+static inline phys_addr_t swiotlb_bus_to_phys(struct device *hwdev,
+					      dma_addr_t baddr)
+{
+	return baddr;
+}
+
+

As I wrote in another mail, adding swiotlb specific code in dma-mapping.h is wrong.

#39 lore raw

Re: [PATCH] swiotlb: make swiotlb phys<->bus translations architecture-specific

Quoted 10 lines · FUJITA Tomonori

arch/ia64/include/asm/dma-mapping.h static inline int swiotlb_force_mapping(phys_addr_t paddr, size_t size) · line 205

Quoted 7 lines
arch/ia64/include/asm/dma-mapping.h +7 -0
+static inline phys_addr_t swiotlb_bus_to_phys(struct device *hwdev,
+                                           dma_addr_t baddr)
+{
+     return baddr;
+}
+
+

As I wrote in another mail, adding swiotlb specific code in dma-mapping.h is wrong.

They aren't really swiotlb specific, I just foolishly carried over the old name, really I think they should be called just bus_to_phys and phys_to_bus or something.

Ian.

#40 lore raw

PATCH 1/1 swiotlb/xen: update xen swiotlb for phys<->bus API changes

In reply to Ian Campbell

These functions are now implemented via explicit hooks in arch/x86/kernel/pci-swiotlb.c rather than as weak hooks in swiotlb.c

Cc FUJITA Tomonori <[email protected]>
Cc Jeremy Fitzhardinge <[email protected]>
Cc Becky Bruce <[email protected]>
Cc Olaf Kirch <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Greg KH <[email protected]>
Cc xen-devel <[email protected]>
Cc x86 maintainers <[email protected]>
---
 drivers/pci/xen-iommu.c |    6 ++++--
 include/xen/swiotlb.h   |    3 ---
 2 files changed, 4 insertions(+), 5 deletions(-)
drivers/pci/xen-iommu.c +4 -2
index 2d727d1..02de48a 100644
@@ -72,12 +72,12 @@ int xen_wants_swiotlb(void)
 	return xen_initial_domain();
 }
 
-dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
+static dma_addr_t xen_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
 {
 	return phys_to_machine(XPADDR(paddr)).maddr;
 }
 
-phys_addr_t xen_bus_to_phys(dma_addr_t daddr)
+static phys_addr_t xen_bus_to_phys(struct device *hwdev, dma_addr_t daddr)
 {
 	return machine_to_phys(XMADDR(daddr)).paddr;
 }
@@ -322,6 +322,8 @@ void __init detect_xen_iommu(void)
 
 	x86_swiotlb_force_mapping = &xen_swiotlb_force_mapping;
 	x86_swiotlb_alloc_fixup = &xen_swiotlb_alloc_fixup;
+	x86_bus_to_phys = &xen_bus_to_phys;
+	x86_phys_to_bus = &xen_phys_to_bus;
 }
 
 void __init xen_iommu_init(void)
include/xen/swiotlb.h +0 -3
index 83ec002..e6c7707 100644
@@ -1,9 +1,6 @@
 #ifndef _XEN_SWIOTLB_H
 #define _XEN_SWIOTLB_H
 
-extern phys_addr_t xen_bus_to_phys(dma_addr_t daddr);
-extern dma_addr_t xen_phys_to_bus(phys_addr_t paddr);
-
 #ifdef CONFIG_PCI_XEN
 extern int xen_wants_swiotlb(void);
 #else
Signature · 1 line
1.5.6.5
#41 lore raw
In reply to Ian Campbell
Quoted 9 lines · On Thu, 21 May 2009 12:45:35 +0100 Ian Campbell <[email protected]> wrote:
Quoted 3 lines

A necessary interface? Sorry, I don't buy it. It's necessary for only Xen. And it's not fit well for swiotlb and the architecture abstraction.

I said "necessary interface to enable a particular use-case". Xen is a valid use case -- I appreciate that you personally may have no interest in it but plenty of people do and plenty of people would like to see it working in the mainline kernels.

Quoted 7 lines

In terms of clean abstraction this is a architecture hook to allow mappings to be forced through the swiotlb. It is essentially a more flexible extension to the existing swiotlb_force flag, in fact swiotlb_force_mapping() might even be a better name for it.

IMHO, the benefits here (working Xen domain 0 plus PCI passthrough to guest domains) are well worth the costs.

All I care about is a clean design; an wrong design will hurt us.

Quoted 13 lines
Quoted 5 lines
Quoted 3 lines

If there was a cleaner way to achieve the same result we would of course go with that. I don't think duplicating swiotlb.c, as has been suggested as the alternative, just for that one hook point makes sense.

One hook? You need to reread swiotlb.c. There are more. All should go.

I meant that swiotlb_range_needs_mapping is the single contentious hook as far as I can tell. It is the only one which you appear to be disputing the very existence of (irrespective of whether it uses __weak or is moved into asm/dma-mapping.h). The other existing __weak hooks are useful to other users too and all can, AFAICT, be moved into asm/dma-mapping.h.

Quoted 5 lines

You have already dealt with moving swiotlb_address_needs_mapping and I am currently looking into the the phys_to_bus and bus_to_phys ones. That just leaves the alloc functions, which are next on my list after phys_to_bus and bus_to_phys. Will finishing up those patches resolve most of your objections are do you have others?

The latest your patch is more hacky than the current code. You just move the ugly hacks for dom0 from lib/swiotlb.c to somewhere else.

I guess that the only way to keep the Xen-specific stuff in Xen's land is something like this. Then xen/pci-swiotlb.c implements its own map/unmap functions without messing up the generic code.

I guess that do_map_single's io_tlb_daddr argument is pointless for Xen since swiotlb doesn't work if iotlb buffer is not physically continuous (you will see data corruption with some device drivers).

include/linux/swiotlb.h +11 -2
index cb1a663..e1c40ae 100644
@@ -21,8 +21,17 @@ struct scatterlist;
  */
 #define IO_TLB_SHIFT 11
 
-extern void
-swiotlb_init(void);
+extern void swiotlb_init(void);
+extern void swiotlb_register_io_tlb(void *);
+
+extern void *do_map_single(struct device *hwdev, dma_addr_t io_tlb_daddr,
+			   phys_addr_t phys, size_t size,
+			   enum dma_data_direction dir);
+extern void do_unmap_single(struct device *hwdev, char *dma_addr, size_t size,
+			    enum dma_data_direction dir);
+
+extern void sync_single(struct device *hwdev, char *dma_addr, size_t size,
+			int dir, int target);
 
 extern void *swiotlb_alloc_boot(size_t bytes, unsigned long nslabs);
 extern void *swiotlb_alloc(unsigned order, unsigned long nslabs);
lib/swiotlb.c +34 -25

112 lines not on this page · raw message

#42 lore raw
In reply to FUJITA Tomonori

lib/swiotlb.c swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems, · line 803

Quoted 11 lines · FUJITA Tomonori
lib/swiotlb.c +2 -2
@@ -819,8 +796,8 @@ swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
                phys_addr_t paddr = sg_phys(sg);
                dma_addr_t dev_addr = swiotlb_phys_to_bus(hwdev, paddr);
 
-               if (range_needs_mapping(paddr, sg->length) ||
-                   address_needs_mapping(hwdev, dev_addr, sg->length)) {
+               if (!is_buffer_dma_capable(hwdev, dma_get_mask(hwdev), paddr, sg->length) ||
+                   swiotlb_force) {
                        void *map = map_single(hwdev, sg_phys(sg),
                                               sg->length, dir);
                        if (!map) {

BTW I think there was a typo here, dev_addr becomes paddr. I fixed that in my updated version.

Ian.

Signature · 7 lines
Ian Campbell
Current Noise: Isis - In Fiction

Poorochrondria:
	Hypochrondria derived from not having medical insurance.
		-- Douglas Coupland, "Generation X: Tales for an Accelerated
		   Culture"