Skip to contentSkip to conversation

Add io-mapping functions to dynamically map large device apertures

PATCH In review Eric Anholt 3 participants 5 messages Oct 31, 2008 – Nov 3, 2008 1 message not in the archive

Subsystems (3)
lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw

PATCH 1/1 Add io-mapping functions to dynamically map large device apertures

From: Keith Packard <[email protected]>

Graphics devices have large PCI apertures which would consume a significant fraction of a 32-bit address space if mapped during driver initialization. Using ioremap at runtime is impractical as it is too slow. This new set of interfaces uses atomic mappings on 32-bit processors and a large static mapping on 64-bit processors to provide reasonable 32-bit performance and optimal 64-bit performance.

The current implementation sits atop the io_map_atomic fixmap-based mechanism for 32-bit processors.

This includes some editorial suggestions from Randy Dunlap for Documentation/io-mapping.txt

---
 Documentation/io-mapping.txt |   76 +++++++++++++++++++++++++++
 include/linux/io-mapping.h   |  118 ++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 194 insertions(+), 0 deletions(-)
 create mode 100644 Documentation/io-mapping.txt
 create mode 100644 include/linux/io-mapping.h
Documentation/io-mapping.txt new file +76 -0
new file mode 100644
index 0000000..cd2f726
@@ -0,0 +1,76 @@
+The io_mapping functions in linux/io-mapping.h provide an abstraction for
+efficiently mapping small regions of an I/O device to the CPU. The initial
+usage is to support the large graphics aperture on 32-bit processors where
+ioremap_wc cannot be used to statically map the entire aperture to the CPU
+as it would consume too much of the kernel address space.
+
+A mapping object is created during driver initialization using
+
+	struct io_mapping *io_mapping_create_wc(unsigned long base,
+						unsigned long size)
+
+		'base' is the bus address of the region to be made
+		mappable, while 'size' indicates how large a mapping region to
+		enable. Both are in bytes.
+
+		This _wc variant provides a mapping which may only be used
+		with the io_mapping_map_atomic_wc or io_mapping_map_wc.
+
+With this mapping object, individual pages can be mapped either atomically
+or not, depending on the necessary scheduling environment. Of course, atomic
+maps are more efficient:
+
+	void *io_mapping_map_atomic_wc(struct io_mapping *mapping,
+				       unsigned long offset)
+
+		'offset' is the offset within the defined mapping region.
+		Accessing addresses beyond the region specified in the
+		creation function yields undefined results. Using an offset
+		which is not page aligned yields an undefined result. The
+		return value points to a single page in CPU address space.
+
+		This _wc variant returns a write-combining map to the
+		page and may only be used with mappings created by
+		io_mapping_create_wc
+
+		Note that the task may not sleep while holding this page
+		mapped.
+
+	void io_mapping_unmap_atomic(void *vaddr)
+
+		'vaddr' must be the the value returned by the last
+		io_mapping_map_atomic_wc call. This unmaps the specified
+		page and allows the task to sleep once again.
+
+If you need to sleep while holding the lock, you can use the non-atomic
+variant, although they may be significantly slower.
+
+	void *io_mapping_map_wc(struct io_mapping *mapping,
+				unsigned long offset)
+
+		This works like io_mapping_map_atomic_wc except it allows
+		the task to sleep while holding the page mapped.
+
+	void io_mapping_unmap(void *vaddr)
+
+		This works like io_mapping_unmap_atomic, except it is used
+		for pages mapped with io_mapping_map_wc.
+
+At driver close time, the io_mapping object must be freed:
+
+	void io_mapping_free(struct io_mapping *mapping)
+
+Current Implementation:
+
+The initial implementation of these functions uses existing mapping
+mechanisms and so provides only an abstraction layer and no new
+functionality.
+
+On 64-bit processors, io_mapping_create_wc calls ioremap_wc for the whole
+range, creating a permanent kernel-visible mapping to the resource. The
+map_atomic and map functions add the requested offset to the base of the
+virtual address returned by ioremap_wc.
+
+On 32-bit processors, io_mapping_map_atomic_wc uses io_map_atomic_prot_pfn,
+which uses the fixmaps to get us a mapping to a page using an atomic fashion.
+For io_mapping_map_wc, ioremap_wc() is used to get a mapping of the region.
include/linux/io-mapping.h new file +118 -0

118 lines not on this page · raw message

Signature · 1 line
1.5.6.5
#2 lore raw

PATCH 1/1 i915: Use io-mapping interfaces instead of a variety of mapping kludges

From: Keith Packard <[email protected]>

Switch the i915 device aperture mapping to the io-mapping interface, taking advantage of the cleaner API to extend it across all of the mapping uses, including both pwrite and relocation updates.

This dramatically improves performance on 64-bit kernels which were using the same slow path as 32-bit non-HIGHMEM kernels prior to this patch.

---
 drivers/gpu/drm/i915/i915_drv.h |    3 +
 drivers/gpu/drm/i915/i915_gem.c |  174 ++++++++++++++++++---------------------
 2 files changed, 83 insertions(+), 94 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +3 -0
index cc8a9f3..572dcd0 100644
@@ -31,6 +31,7 @@
 #define _I915_DRV_H_
 
 #include "i915_reg.h"
+#include <linux/io-mapping.h>
 
 /* General customization:
  */
@@ -246,6 +247,8 @@ typedef struct drm_i915_private {
 	struct {
 		struct drm_mm gtt_space;
 
+		struct io_mapping *gtt_mapping;
+
 		/**
 		 * List of objects currently involved in rendering from the
 		 * ringbuffer.
drivers/gpu/drm/i915/i915_gem.c +80 -94

252 lines not on this page · raw message

Signature · 1 line
1.5.6.5
#3 lore raw
In reply to Eric Anholt
Quoted 12 lines · Eric Anholt

From: Keith Packard <[email protected]>

Graphics devices have large PCI apertures which would consume a significant fraction of a 32-bit address space if mapped during driver initialization. Using ioremap at runtime is impractical as it is too slow. This new set of interfaces uses atomic mappings on 32-bit processors and a large static mapping on 64-bit processors to provide reasonable 32-bit performance and optimal 64-bit performance.

The current implementation sits atop the io_map_atomic fixmap-based mechanism for 32-bit processors.

Quoted 8 lines

This includes some editorial suggestions from Randy Dunlap for Documentation/io-mapping.txt

---
 Documentation/io-mapping.txt |   76 +++++++++++++++++++++++++++
 include/linux/io-mapping.h   |  118 ++++++++++++++++++++++++++++++++++++++++++

I've applied your three patches to tip/core/resources for testing, thanks!

One small detail:

Quoted 3 lines

+++ b/include/linux/io-mapping.h

+#ifdef CONFIG_X86_64

it's ugly and inflexible to put x86 dependencies into generic headers. (even though with a high likelyhood 32-bit x86 will be the only arch to ever implement the iomap_atomic() APIs)

Instead please add a HAVE_ATOMIC_IOMAP define to arch/x86/Kconfig:

config HAVE_ATOMIC_IOMAP

         def_bool y
         depends on X86_32

... and use #ifndef HAVE_ATOMIC_IOMAP in include/linux/io-mapping.h instead of #ifdef CONFIG_X86_64.

( Other 32-bit architectures which need an atomic iomap implementation
for address space reasons can then implement the iomap_atomic*() APIs too and set this same flag, to gain the same generic io_mapping implementation. )

Please send this cleanup as a delta patch, ontop of your three patches.

Ingo

#4 lore raw
Quoted 4 lines · Ingo Molnar

it's ugly and inflexible to put x86 dependencies into generic headers. (even though with a high likelyhood 32-bit x86 will be the only arch to ever implement the iomap_atomic() APIs)

Quoted 8 lines

Instead please add a HAVE_ATOMIC_IOMAP define to arch/x86/Kconfig:

config HAVE_ATOMIC_IOMAP

         def_bool y
         depends on X86_32

... and use #ifndef HAVE_ATOMIC_IOMAP in include/linux/io-mapping.h instead of #ifdef CONFIG_X86_64.

Just to clarify the issue here: there are two separate implementations of the io_mapping API -- one for 'large address space' machines where ioremap_wc can handle the typical graphics aperture within the kernel virtual map, and the other using iomap_atomic_prot_pfn for machines with puny address spaces.

All large address space machines can provide the io_mapping API without any archtecture-specific support. For efficient 32-bit io_mapping support, we require the new iomap_atomic_prot_pfn function.

So, it seems like what I want to do is use the large address space code on any machine which supports it, and then use the iomap_atomic_prot_pfn version for small address space machines which have the iomap_atomic_prot_pfn function.

What I think you're suggesting is to just assume that machines without iomap_atomic_prot_pfn have address spaces large enough to support the ioremap_wc path. The alternative is to create a third (slow) path (which I did before the iomap_atomic_prot_pfn API was introduced) that uses ioremap_wc at run time for small address space machines without iomap_atomic_prot_pfn.

Let me know which you'd prefer and I'll get a patch out ASAP.

Signature · 1 line
[email protected]
#5 lore raw
Quoted 26 lines · Keith Packard
Quoted 4 lines · Ingo Molnar

it's ugly and inflexible to put x86 dependencies into generic headers. (even though with a high likelyhood 32-bit x86 will be the only arch to ever implement the iomap_atomic() APIs)

Quoted 8 lines

Instead please add a HAVE_ATOMIC_IOMAP define to arch/x86/Kconfig:

config HAVE_ATOMIC_IOMAP

         def_bool y
         depends on X86_32

... and use #ifndef HAVE_ATOMIC_IOMAP in include/linux/io-mapping.h instead of #ifdef CONFIG_X86_64.

Just to clarify the issue here: there are two separate implementations of the io_mapping API -- one for 'large address space' machines where ioremap_wc can handle the typical graphics aperture within the kernel virtual map, and the other using iomap_atomic_prot_pfn for machines with puny address spaces.

All large address space machines can provide the io_mapping API without any archtecture-specific support. For efficient 32-bit io_mapping support, we require the new iomap_atomic_prot_pfn function.

Quoted 4 lines

So, it seems like what I want to do is use the large address space code on any machine which supports it, and then use the iomap_atomic_prot_pfn version for small address space machines which have the iomap_atomic_prot_pfn function.

Correct.

Quoted 8 lines

What I think you're suggesting is to just assume that machines without iomap_atomic_prot_pfn have address spaces large enough to support the ioremap_wc path. The alternative is to create a third (slow) path (which I did before the iomap_atomic_prot_pfn API was introduced) that uses ioremap_wc at run time for small address space machines without iomap_atomic_prot_pfn.

Let me know which you'd prefer and I'll get a patch out ASAP.

Please lets keep it simple: i.e. always use ioremap_wc() when there's no iomap_atomic_prot_pfn() 32-bit API provided.

( and by all means ioremap_wc() will just work fine on most 32-bit
architectures out of box: they dont go about trying to map hundreds of megabytes of graphics aperture. If they nevertheless need it, they can implement iomap_atomic_prot_pfn() to add support. )

Ingo