fix default dma_mmap_* pgprot
Subsystems (6)
Hi all,
As Shawn pointed out we've had issues with the dma mmap pgprots ever since the dma_common_mmap helper was added beyong the initial architectures - we default to uncached mappings, but for devices that are DMA coherent, or if the DMA_ATTR_NON_CONSISTENT is set (and supported) this can lead to aliasing of cache attributes. This patch fixes that. My explanation of why this hasn't been much of an issue is that the dma_mmap_ helpers aren't used widely and mostly just in architecture specific drivers.
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
PATCH 1/1 dma-mapping: fix page attributes for dma_mmap_*
All the way back to introducing dma_common_mmap we've defaulyed to mark the pages as uncached. But this is wrong for DMA coherent devices or if using DMA_ATTR_NON_CONSISTENT. Later on DMA_ATTR_WRITE_COMBINE also got incorrect treatment as that flag is only treated special on the alloc side for non-coherent devices.
Introduce a new dma_mmap_pgprot helper that deals with the check for coherent devices and DMA_ATTR_NON_CONSISTENT so that only the remapping cases even reach arch_dma_mmap_pgprot and we thus ensure no aliasing of page attributes happens.
--- arch/arm/mm/dma-mapping.c | 4 +--- arch/arm64/mm/dma-mapping.c | 4 +--- arch/powerpc/kernel/Makefile | 3 +-- arch/powerpc/kernel/dma-common.c | 17 ----------------- drivers/iommu/dma-iommu.c | 6 +++--- include/linux/dma-mapping.h | 1 + include/linux/dma-noncoherent.h | 5 ----- kernel/dma/mapping.c | 11 ++++++++++- kernel/dma/remap.c | 2 +- 9 files changed, 18 insertions(+), 35 deletions(-) delete mode 100644 arch/powerpc/kernel/dma-common.c
arch/arm/mm/dma-mapping.c +1 -3
@@ -2397,9 +2397,7 @@ long arch_dma_coherent_to_pfn(struct device *dev, void *cpu_addr,pgprot_t arch_dma_mmap_pgprot (struct device *dev,pgprot_t prot,unsigned long attrs) { -if (!dev_is_dma_coherent (dev)) -return __get_dma_pgprot (attrs, prot); -return prot; +return __get_dma_pgprot (attrs, prot); } void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
arch/arm64/mm/dma-mapping.c +1 -3
@@ -14,9 +14,7 @@pgprot_t arch_dma_mmap_pgprot (struct device *dev,pgprot_t prot,unsigned long attrs) { -if (!dev_is_dma_coherent (dev) || (attrs &DMA_ATTR_WRITE_COMBINE )) -return pgprot_writecombine (prot); -return prot; +return pgprot_writecombine (prot); } void arch_sync_dma_for_device(struct device *dev, phys_addr_t paddr,
arch/powerpc/kernel/Makefile +1 -2
@@ -49,8 +49,7 @@ obj-y := cputable.o ptrace.o syscalls.o \ signal.o sysfs.o cacheinfo.o time.o \ prom.o traps.o setup-common.o \ udbg.o misc.o io.o misc_$(BITS).o \ - of_platform.o prom_parse.o \ - dma-common.o + of_platform.o prom_parse.o obj-$(CONFIG_PPC64) += setup_64.o sys_ppc32.o \ signal_64.o ptrace32.o \ paca.o nvram_64.o firmware.o
arch/powerpc/kernel/dma-common.c deleted +0 -17
@@ -1,17 +0,0 @@ -// SPDX-License-Identifier: GPL-2.0-or-later -/* -* Contains common dma routines for all powerpc platforms. -* -* Copyright (C) 2019 Shawn Anastasio. -*/ - -#include <linux/mm.h> -#include <linux/dma-noncoherent.h> - -pgprot_t arch_dma_mmap_pgprot (struct device *dev,pgprot_t prot, -unsigned long attrs) -{ -if (!dev_is_dma_coherent (dev)) -return pgprot_noncached (prot); -return prot; -}
drivers/iommu/dma-iommu.c +3 -3
@@ -574,7 +574,7 @@ static void *iommu_dma_alloc_remap(struct device *dev, size_t size,struct iova_domain *iovad = &cookie->iovad;bool coherent =dev_is_dma_coherent (dev);int ioprot =dma_info_to_prot (DMA_BIDIRECTIONAL , coherent, attrs); -pgprot_t prot =arch_dma_mmap_pgprot (dev,PAGE_KERNEL , attrs); +pgprot_t prot =dma_mmap_pgprot (dev,PAGE_KERNEL , attrs);unsigned int count, min_size, alloc_sizes = domain->pgsize_bitmap;struct page **pages;struct sg_table sgt;
@@ -975,7 +975,7 @@ static void *iommu_dma_alloc_pages(struct device *dev, size_t size,return NULL ;if (IS_ENABLED (CONFIG_DMA_REMAP ) && (!coherent ||PageHighMem (page))) { -pgprot_t prot =arch_dma_mmap_pgprot (dev,PAGE_KERNEL , attrs); +pgprot_t prot =dma_mmap_pgprot (dev,PAGE_KERNEL , attrs); cpu_addr =dma_common_contiguous_remap (page, alloc_size,VM_USERMAP , prot,__builtin_return_address (0 ));
@@ -1035,7 +1035,7 @@ static int iommu_dma_mmap(struct device *dev, struct vm_area_struct *vma,unsigned long pfn, off = vma->vm_pgoff;int ret; - vma->vm_page_prot =arch_dma_mmap_pgprot (dev, vma->vm_page_prot, attrs); + vma->vm_page_prot =dma_mmap_pgprot (dev, vma->vm_page_prot, attrs);if (dma_mmap_from_dev_coherent (dev, vma, cpu_addr, size, &ret))return ret;
include/linux/dma-mapping.h +1 -0
@@ -611,6 +611,7 @@ static inline void dma_sync_single_range_for_device(struct device *dev,#define dma_get_sgtable (d, t, v, h, s) dma_get_sgtable_attrs(d, t, v, h, s, 0)#define dma_mmap_coherent (d, v, c, h, s) dma_mmap_attrs(d, v, c, h, s, 0) +pgprot_t dma_mmap_pgprot (struct device *dev,pgprot_t prot,unsigned long attrs);extern int dma_common_mmap (struct device *dev,struct vm_area_struct *vma,void *cpu_addr,dma_addr_t dma_addr,size_t size,unsigned long attrs);
include/linux/dma-noncoherent.h +0 -5
@@ -42,13 +42,8 @@ void arch_dma_free(struct device *dev, size_t size, void *cpu_addr, dma_addr_t dma_addr, unsigned long attrs);long arch_dma_coherent_to_pfn (struct device *dev,void *cpu_addr,dma_addr_t dma_addr); - -#ifdef CONFIG_ARCH_HAS_DMA_MMAP_PGPROT pgprot_t arch_dma_mmap_pgprot (struct device *dev,pgprot_t prot,unsigned long attrs); -#else -# define arch_dma_mmap_pgprot (dev, prot, attrs) pgprot_noncached(prot) -#endif #ifdef CONFIG_DMA_NONCOHERENT_CACHE_SYNC void arch_dma_cache_sync(struct device *dev, void *vaddr, size_t size,
kernel/dma/mapping.c +10 -1
@@ -162,6 +162,15 @@ int dma_get_sgtable_attrs(struct device *dev, struct sg_table *sgt, }EXPORT_SYMBOL (dma_get_sgtable_attrs); +pgprot_t dma_mmap_pgprot (struct device *dev,pgprot_t prot,unsigned long attrs) +{ +if (dev_is_dma_coherent (dev) || (attrs &DMA_ATTR_NON_CONSISTENT )) +return prot; +if (IS_ENABLED (CONFIG_ARCH_HAS_DMA_MMAP_PGPROT )) +return arch_dma_mmap_pgprot (dev, prot, attrs); +return pgprot_noncached (prot); +} +/* * Create userspace mapping for the DMA-coherent memory. */
@@ -175,7 +184,7 @@ int dma_common_mmap(struct device *dev, struct vm_area_struct *vma,unsigned long pfn;int ret = -ENXIO ; - vma->vm_page_prot =arch_dma_mmap_pgprot (dev, vma->vm_page_prot, attrs); + vma->vm_page_prot =dma_mmap_pgprot (dev, vma->vm_page_prot, attrs);if (dma_mmap_from_dev_coherent (dev, vma, cpu_addr, size, &ret))return ret;
kernel/dma/remap.c +1 -1
@@ -218,7 +218,7 @@ void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,/* create a coherent mapping */ ret =dma_common_contiguous_remap (page, size,VM_USERMAP , -arch_dma_mmap_pgprot (dev,PAGE_KERNEL , attrs), +dma_mmap_pgprot (dev,PAGE_KERNEL , attrs),__builtin_return_address (0 ));if (!ret) {__dma_direct_free_pages (dev, size, page);
Signature · 7 lines
2.20.1 _______________________________________________ linux-arm-kernel mailing list [email protected] http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Re: [PATCH] dma-mapping: fix page attributes for dma_mmap_*
Quoted 45 lines · Christoph Hellwig
All the way back to introducing
dma_common_mmapwe've defaulyed to mark the pages as uncached. But this is wrong for DMA coherent devices or if usingDMA_ATTR_NON_CONSISTENT. Later onDMA_ATTR_WRITE_COMBINEalso got incorrect treatment as that flag is only treated special on the alloc side for non-coherent devices.Introduce a new
dma_mmap_pgprothelper that deals with the check for coherent devices andDMA_ATTR_NON_CONSISTENTso that only the remapping cases even reacharch_dma_mmap_pgprotand we thus ensure no aliasing of page attributes happens.Signed-off-by Christoph Hellwig <[email protected]>--- arch/arm/mm/dma-mapping.c | 4 +--- arch/arm64/mm/dma-mapping.c | 4 +--- arch/powerpc/kernel/Makefile | 3 +-- arch/powerpc/kernel/dma-common.c | 17 ----------------- drivers/iommu/dma-iommu.c | 6 +++--- include/linux/dma-mapping.h | 1 + include/linux/dma-noncoherent.h | 5 ----- kernel/dma/mapping.c | 11 ++++++++++- kernel/dma/remap.c | 2 +- 9 files changed, 18 insertions(+), 35 deletions(-) delete mode 100644 arch/powerpc/kernel/dma-common.carch/arm/mm/dma-mapping.c +1
-3@@ -2397,9 +2397,7 @@ long arch_dma_coherent_to_pfn(struct device *dev, void *cpu_addr,pgprot_t arch_dma_mmap_pgprot (struct device *dev,pgprot_t prot,unsigned long attrs) { -if (!dev_is_dma_coherent (dev)) -return __get_dma_pgprot (attrs, prot); -return prot; +return __get_dma_pgprot (attrs, prot); } void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,arch/arm64/mm/dma-mapping.c +0
-0
arch/arm64/mm/dma-mapping.c line 18
Quoted 8 lines
arch/arm64/mm/dma-mapping.c +1
-3@@ -14,9 +14,7 @@pgprot_t arch_dma_mmap_pgprot (struct device *dev,pgprot_t prot,unsigned long attrs) { -if (!dev_is_dma_coherent (dev) || (attrs &DMA_ATTR_WRITE_COMBINE )) -return pgprot_writecombine (prot); -return prot; +return pgprot_writecombine (prot); }
Seems like a sensible cleanup to me:
Although arch_dma_mmap_pgprot() is a bit of a misnomer now that it only gets involved in the non-coherent case.
Will
Re: [PATCH] dma-mapping: fix page attributes for dma_mmap_*
Quoted 8 lines · Will Deacon
arch/arm64/mm/dma-mapping.c line 18
Quoted 5 lines
arch/arm64/mm/dma-mapping.c +1
-3-if (!dev_is_dma_coherent (dev) || (attrs &DMA_ATTR_WRITE_COMBINE )) -return pgprot_writecombine (prot); -return prot; +return pgprot_writecombine (prot); }Seems like a sensible cleanup to me:
Quoted 4 lines
Acked-by Will Deacon <[email protected]>Although
arch_dma_mmap_pgprot()is a bit of a misnomer now that it only gets involved in the non-coherent case.
A better name is welcome. My other idea would be to just remove it entirely and do something like:
#ifndef pgprot_dmacoherent
#define pgprot_dmacoherent pgprot_noncached
#endif
pgprot_t dma_mmap_pgprot(struct device *dev, pgprot_t prot, unsigned long attrs)
{
if (dev_is_dma_coherent(dev) || (attrs & DMA_ATTR_NON_CONSISTENT))
return prot;
#ifdef pgprot_writecombine
if (attrs & DMA_ATTR_WRITE_COMBINE)
return pgprot_writecombine(prot);
#endif
return pgprot_dmacoherent(prot);
}But my worry is how this interacts with architectures that have an uncached segment (mips, nios2, microblaze, extensa) where we'd have the kernel access DMA_ATTR_WRITE_COMBINE mappigns using the uncached segment, and userspace mmaps using pgprot_writecombine, which could lead to aliasing issues. But then again mips already supports DMA_ATTR_WRITE_COMBINE, so this must be ok somehow. I guess I'll need to field that question to the relevant parties.
Will Deacon
maintainerRe: [PATCH] dma-mapping: fix page attributes for dma_mmap_*
Quoted 9 lines · Christoph Hellwig
Quoted 8 lines · Will Deacon
arch/arm64/mm/dma-mapping.c line 18
Quoted 5 lines
arch/arm64/mm/dma-mapping.c +1
-3-if (!dev_is_dma_coherent (dev) || (attrs &DMA_ATTR_WRITE_COMBINE )) -return pgprot_writecombine (prot); -return prot; +return pgprot_writecombine (prot); }Seems like a sensible cleanup to me:
Quoted 6 lines
Quoted 4 lines
Acked-by Will Deacon <[email protected]>Although
arch_dma_mmap_pgprot()is a bit of a misnomer now that it only gets involved in the non-coherent case.A better name is welcome.
How about arch_dma_noncoherent_mmap_pgprot() ? Too long?
Quoted 7 lines
My other idea would be to just remove it entirely and do something like:
#ifndef
pgprot_dmacoherent
#definepgprot_dmacoherentpgprot_noncached
#endif
pgprot_tdma_mmap_pgprot(struct device *dev, pgprot_t prot, unsigned long attrs)
Quoted 9 lines
{
if (dev_is_dma_coherent(dev) || (attrs & DMA_ATTR_NON_CONSISTENT))
return prot;
#ifdef pgprot_writecombine
if (attrs & DMA_ATTR_WRITE_COMBINE)
return pgprot_writecombine(prot);
#endif
return pgprot_dmacoherent(prot);
}Oh, I prefer that!
Quoted 7 lines
But my worry is how this interacts with architectures that have an uncached segment (mips, nios2, microblaze, extensa) where we'd have the kernel access
DMA_ATTR_WRITE_COMBINEmappigns using the uncached segment, and userspace mmaps usingpgprot_writecombine, which could lead to aliasing issues. But then again mips already supportsDMA_ATTR_WRITE_COMBINE, so this must be ok somehow. I guess I'll need to field that question to the relevant parties.
Or it's always been busted and happens to work out in practice...
Will
Re: [PATCH] dma-mapping: fix page attributes for dma_mmap_*
Quoted 6 lines · Will Deacon
Quoted 4 lines
Quoted 2 lines
Although
arch_dma_mmap_pgprot()is a bit of a misnomer now that it only gets involved in the non-coherent case.A better name is welcome.
How about
arch_dma_noncoherent_mmap_pgprot()? Too long?
Sounds a little long yes. And doesn't fix the additional problem that we don't just it for mmap but also for the in-kernel remapping these days.
Quoted 9 lines
Quoted 7 lines
But my worry is how this interacts with architectures that have an uncached segment (mips, nios2, microblaze, extensa) where we'd have the kernel access
DMA_ATTR_WRITE_COMBINEmappigns using the uncached segment, and userspace mmaps usingpgprot_writecombine, which could lead to aliasing issues. But then again mips already supportsDMA_ATTR_WRITE_COMBINE, so this must be ok somehow. I guess I'll need to field that question to the relevant parties.Or it's always been busted and happens to work out in practice...
I've sent a ping to the mips folks. While we'are at it: arm64 and arm32 (optionally) map dma coherent allocations as write combine. I suspect this hasn't always just been busted but intentional (of course!), but is there any chance to get a quote from the arm architecture spec on why this is fine as it looks rather confusion?
Also if we assume mips is buggy DMA_ATTR_WRITE_COMBINE really just seems to be there for old arm platforms, which makes the scope pretty limited.
Will Deacon
maintainerRe: [PATCH] dma-mapping: fix page attributes for dma_mmap_*
Quoted 11 lines · Christoph Hellwig
Quoted 6 lines · Will Deacon
Quoted 4 lines
Quoted 2 lines
Although
arch_dma_mmap_pgprot()is a bit of a misnomer now that it only gets involved in the non-coherent case.A better name is welcome.
How about
arch_dma_noncoherent_mmap_pgprot()? Too long?Sounds a little long yes. And doesn't fix the additional problem that we don't just it for mmap but also for the in-kernel remapping these days.
Hmm. Maybe just arch_dma_noncoherent_pgprot() then.
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
But my worry is how this interacts with architectures that have an uncached segment (mips, nios2, microblaze, extensa) where we'd have the kernel access
DMA_ATTR_WRITE_COMBINEmappigns using the uncached segment, and userspace mmaps usingpgprot_writecombine, which could lead to aliasing issues. But then again mips already supportsDMA_ATTR_WRITE_COMBINE, so this must be ok somehow. I guess I'll need to field that question to the relevant parties.
Quoted 7 lines
Quoted 1 line
Or it's always been busted and happens to work out in practice...
I've sent a ping to the mips folks. While we'are at it: arm64 and arm32 (optionally) map dma coherent allocations as write combine. I suspect this hasn't always just been busted but intentional (of course!), but is there any chance to get a quote from the arm architecture spec on why this is fine as it looks rather confusion?
So this boils down to a terminology mismatch. The Arm architecture doesn't have anything called "write combine", so in Linux we instead provide what the Arm architecture calls "Normal non-cacheable" memory for pgprot_writecombine(). Amongst other things, this memory type permits speculation, unaligned accesses and merging of writes. I found something in the architecture spec about non-cachable memory, but it's written in Armglish[1].
pgprot_noncached(), on the other hand, provides what the architecture calls Strongly Ordered or Device-nGnRnE memory. This is intended for mapping MMIO (i.e. PCI config space) and therefore forbids speculation, preserves access size, requires strict alignment and also forces write responses to come from the endpoint.
I think the naming mismatch is historical, but on arm64 we wanted to use the same names as arm32 so that any drivers using these things directly would get the same behaviour.
Will
[1]
B2.4.4 Implication of caches for the application programmer
[...]
Data coherency issues
Software can ensure the data coherency of caches in the following ways:
* By not using the caches in situations where coherency issues can arise. This can be achieved by:
- Using Non-cacheable or, in some cases, Write-Through Cacheable memory.
- Not enabling caches in the system.
* By using cache maintenance instructions to manage the coherency issues in software.
* By using hardware coherency mechanisms to ensure the coherency of data accesses to memory for cacheable locations by observers within the different shareability domains.
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
Re: [PATCH] dma-mapping: fix page attributes for dma_mmap_*
Quoted 14 lines · Will Deacon
So this boils down to a terminology mismatch. The Arm architecture doesn't have anything called "write combine", so in Linux we instead provide what the Arm architecture calls "Normal non-cacheable" memory for
pgprot_writecombine(). Amongst other things, this memory type permits speculation, unaligned accesses and merging of writes. I found something in the architecture spec about non-cachable memory, but it's written in Armglish[1].
pgprot_noncached(), on the other hand, provides what the architecture calls Strongly Ordered or Device-nGnRnE memory. This is intended for mapping MMIO (i.e. PCI config space) and therefore forbids speculation, preserves access size, requires strict alignment and also forces write responses to come from the endpoint.
Quoted 3 lines
I think the naming mismatch is historical, but on arm64 we wanted to use the same names as arm32 so that any drivers using these things directly would get the same behaviour.
That all makes sense, but it totally needs a comment. I'll try to draft one based on this. I've also looked at the arm32 code a bit more, and it seems arm always (?) supported Normal non-cacheable attribute, but Linux only optionally uses it for arm v6+ because of fears of drivers missing barriers. The other really weird things is that in arm32 pgprot_dmacoherent incudes the L_PTE_XN bit, which from my understanding is the no-execture bit, but pgprot_writecombine does not. This seems to not very unintentional. So minus that the whole DMA_ATTR_WRITE_COMBІNE seems to be about flagging old arm specific drivers as having the proper barriers in places and otherwise is a no-op.
Here is my tentative plan:
- respin this patch with a small fix to handle the DMA_ATTR_NON_CONSISTENT (as in ignore it unless actually supported), but keep the name as-is to avoid churn. This should allow 5.3 inclusion and backports
- remove DMA_ATTR_WRITE_COMBINE support from mips, probably also 5.3 material.
- move all architectures but arm over to just define pgprot_dmacoherent, including a comment with the above explanation for arm64.
- make DMA_ATTR_WRITE_COMBINE a no-op and schedule it for removal, thus removing the last instances of arch_dma_mmap_pgprot
Will Deacon
maintainerRe: [PATCH] dma-mapping: fix page attributes for dma_mmap_*
Quoted 19 lines · Christoph Hellwig
Quoted 14 lines · Will Deacon
So this boils down to a terminology mismatch. The Arm architecture doesn't have anything called "write combine", so in Linux we instead provide what the Arm architecture calls "Normal non-cacheable" memory for
pgprot_writecombine(). Amongst other things, this memory type permits speculation, unaligned accesses and merging of writes. I found something in the architecture spec about non-cachable memory, but it's written in Armglish[1].
pgprot_noncached(), on the other hand, provides what the architecture calls Strongly Ordered or Device-nGnRnE memory. This is intended for mapping MMIO (i.e. PCI config space) and therefore forbids speculation, preserves access size, requires strict alignment and also forces write responses to come from the endpoint.Quoted 3 lines
I think the naming mismatch is historical, but on arm64 we wanted to use the same names as arm32 so that any drivers using these things directly would get the same behaviour.
Quoted 5 lines
That all makes sense, but it totally needs a comment. I'll try to draft one based on this. I've also looked at the arm32 code a bit more, and it seems arm always (?) supported Normal non-cacheable attribute, but Linux only optionally uses it for arm v6+ because of fears of drivers missing barriers.
I think it was also to do with aliasing, but I don't recall all of the details.
Quoted 6 lines
The other really weird things is that in arm32
pgprot_dmacoherentincudes theL_PTE_XNbit, which from my understanding is the no-execture bit, butpgprot_writecombinedoes not. This seems to not very unintentional. So minus that the whole DMA_ATTR_WRITE_COMBІNE seems to be about flagging old arm specific drivers as having the proper barriers in places and otherwise is a no-op.
I think it only matters for Armv7 CPUs, but yes, we should probably be setting L_PTE_XN for both of these memory types.
Quoted 11 lines
Here is my tentative plan:
- respin this patch with a small fix to handle the
DMA_ATTR_NON_CONSISTENT(as in ignore it unless actually supported), but keep the name as-is to avoid churn. This should allow 5.3 inclusion and backports
- removeDMA_ATTR_WRITE_COMBINEsupport from mips, probably also 5.3 material.
- move all architectures but arm over to just definepgprot_dmacoherent, including a comment with the above explanation for arm64.
That would be great, thanks.
Quoted 2 lines
- make
DMA_ATTR_WRITE_COMBINEa no-op and schedule it for removal, thus removing the last instances ofarch_dma_mmap_pgprot
All sounds good to me, although I suppose 32-bit Arm platforms without CONFIG_ARM_DMA_MEM_BUFFERABLE may run into issues if DMA_ATTR_WRITE_COMBINE disappears. Only one way to find out...
Will
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
Russell King - ARM Linux admin
maintainerRe: [PATCH] dma-mapping: fix page attributes for dma_mmap_*
Quoted 20 lines · Will Deacon
Quoted 19 lines · Christoph Hellwig
Quoted 17 lines · Will Deacon
So this boils down to a terminology mismatch. The Arm architecture doesn't have anything called "write combine", so in Linux we instead provide what the Arm architecture calls "Normal non-cacheable" memory for
pgprot_writecombine(). Amongst other things, this memory type permits speculation, unaligned accesses and merging of writes. I found something in the architecture spec about non-cachable memory, but it's written in Armglish[1].
pgprot_noncached(), on the other hand, provides what the architecture calls Strongly Ordered or Device-nGnRnE memory. This is intended for mapping MMIO (i.e. PCI config space) and therefore forbids speculation, preserves access size, requires strict alignment and also forces write responses to come from the endpoint.I think the naming mismatch is historical, but on arm64 we wanted to use the same names as arm32 so that any drivers using these things directly would get the same behaviour.
Quoted 8 lines
Quoted 5 lines
That all makes sense, but it totally needs a comment. I'll try to draft one based on this. I've also looked at the arm32 code a bit more, and it seems arm always (?) supported Normal non-cacheable attribute, but Linux only optionally uses it for arm v6+ because of fears of drivers missing barriers.
I think it was also to do with aliasing, but I don't recall all of the details.
ARMv6+ is where the architecture significantly changed to introduce the idea of [Normal, Device, Strongly Ordered] where Normal has the cache attributes.
Before that, we had just "uncached/unbuffered, uncached/buffered, cached/unbuffered, cached/buffered" modes.
The write buffer (enabled by buffered modes) has no architected guarantees about how long writes will sit in it, and there is only the "drain write buffer" instruction to push writes out.
Up to and including ARMv5, we took the easy approach of just using the "uncached/unbuffered" mode since that is (a) the safest, and (b) avoids write buffers that alias when there are multiple different mappings.
We could have used a different approach, making all IO writes contain a "drain write buffer" instruction, and map DMA memory as "buffered", but as there were no Linux barriers defined to order memory accesses to DMA memory (so, for example, ring buffers can be updated in the correct order) back in those days, using the uncached/unbuffered mode was the sanest and most reliable solution.
Quoted 10 lines
Quoted 7 lines
The other really weird things is that in arm32
pgprot_dmacoherentincudes theL_PTE_XNbit, which from my understanding is the no-execture bit, butpgprot_writecombinedoes not. This seems to not very unintentional. So minus that the whole DMA_ATTR_WRITE_COMBІNE seems to be about flagging old arm specific drivers as having the proper barriers in places and otherwise is a no-op.I think it only matters for Armv7 CPUs, but yes, we should probably be setting
L_PTE_XNfor both of these memory types.
Conventionally, pgprot_writecombine() has only been used to change the memory type and not the permissions. Since writecombine memory is still capable of being executed, I don't see any reason to set XN for it.
If the user wishes to mmap() using PROT_READ|PROT_EXEC, then is there really a reason for writecombine to set XN overriding the user?
That said, pgprot_writecombine() is mostly used for framebuffers, which arguably shouldn't be executable anyway - but who'd want to mmap() the framebuffer with PROT_EXEC?
Quoted 13 lines
Quoted 12 lines
Here is my tentative plan:
- respin this patch with a small fix to handle the
DMA_ATTR_NON_CONSISTENT(as in ignore it unless actually supported), but keep the name as-is to avoid churn. This should allow 5.3 inclusion and backports
- removeDMA_ATTR_WRITE_COMBINEsupport from mips, probably also 5.3 material.
- move all architectures but arm over to just definepgprot_dmacoherent, including a comment with the above explanation for arm64.
Quoted 8 lines
That would be great, thanks.
Quoted 2 lines
- make
DMA_ATTR_WRITE_COMBINEa no-op and schedule it for removal, thus removing the last instances ofarch_dma_mmap_pgprotAll sounds good to me, although I suppose 32-bit Arm platforms without
CONFIG_ARM_DMA_MEM_BUFFERABLEmay run into issues ifDMA_ATTR_WRITE_COMBINEdisappears. Only one way to find out...
Looking at the results of grep, I think only OMAP2+ and Exynos may be affected.
However, removing writecombine support from the DMA API is going to have a huge impact for framebuffers on earlier ARMs - that's where we do expect framebuffers to be mapped "uncached/buffered" for performance reasons and not "uncached/unbuffered". It's quite literally the difference between console scrolling being usable and totally unusable.
Given what I've said above, switching to using buffered mode for normal DMA mappings is data-corrupting risky - as in your filesystem could get fried. I don't think we should play fast and loose with people's data by randomly changing that "because we'd like to", and I don't see that screwing the console is really an option either.
Signature · 7 lines
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/ FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up According to speedtest.net: 11.9Mbps down 500kbps up _______________________________________________ iommu mailing list [email protected] https://lists.linuxfoundation.org/mailman/listinfo/iommu
Russell King - ARM Linux admin
maintainerRe: [PATCH] dma-mapping: fix page attributes for dma_mmap_*
Quoted 107 lines · Russell King - ARM Linux admin
Quoted 20 lines · Will Deacon
Quoted 19 lines · Christoph Hellwig
Quoted 17 lines · Will Deacon
So this boils down to a terminology mismatch. The Arm architecture doesn't have anything called "write combine", so in Linux we instead provide what the Arm architecture calls "Normal non-cacheable" memory for
pgprot_writecombine(). Amongst other things, this memory type permits speculation, unaligned accesses and merging of writes. I found something in the architecture spec about non-cachable memory, but it's written in Armglish[1].
pgprot_noncached(), on the other hand, provides what the architecture calls Strongly Ordered or Device-nGnRnE memory. This is intended for mapping MMIO (i.e. PCI config space) and therefore forbids speculation, preserves access size, requires strict alignment and also forces write responses to come from the endpoint.I think the naming mismatch is historical, but on arm64 we wanted to use the same names as arm32 so that any drivers using these things directly would get the same behaviour.
Quoted 8 lines
Quoted 5 lines
That all makes sense, but it totally needs a comment. I'll try to draft one based on this. I've also looked at the arm32 code a bit more, and it seems arm always (?) supported Normal non-cacheable attribute, but Linux only optionally uses it for arm v6+ because of fears of drivers missing barriers.
I think it was also to do with aliasing, but I don't recall all of the details.
ARMv6+ is where the architecture significantly changed to introduce the idea of [Normal, Device, Strongly Ordered] where Normal has the cache attributes.
Before that, we had just "uncached/unbuffered, uncached/buffered, cached/unbuffered, cached/buffered" modes.
The write buffer (enabled by buffered modes) has no architected guarantees about how long writes will sit in it, and there is only the "drain write buffer" instruction to push writes out.
Up to and including ARMv5, we took the easy approach of just using the "uncached/unbuffered" mode since that is (a) the safest, and (b) avoids write buffers that alias when there are multiple different mappings.
We could have used a different approach, making all IO writes contain a "drain write buffer" instruction, and map DMA memory as "buffered", but as there were no Linux barriers defined to order memory accesses to DMA memory (so, for example, ring buffers can be updated in the correct order) back in those days, using the uncached/unbuffered mode was the sanest and most reliable solution.
Quoted 10 lines
Quoted 7 lines
The other really weird things is that in arm32
pgprot_dmacoherentincudes theL_PTE_XNbit, which from my understanding is the no-execture bit, butpgprot_writecombinedoes not. This seems to not very unintentional. So minus that the whole DMA_ATTR_WRITE_COMBІNE seems to be about flagging old arm specific drivers as having the proper barriers in places and otherwise is a no-op.I think it only matters for Armv7 CPUs, but yes, we should probably be setting
L_PTE_XNfor both of these memory types.Conventionally,
pgprot_writecombine()has only been used to change the memory type and not the permissions. Since writecombine memory is still capable of being executed, I don't see any reason to set XN for it.If the user wishes to
mmap()using PROT_READ|PROT_EXEC, then is there really a reason for writecombine to set XN overriding the user?That said,
pgprot_writecombine()is mostly used for framebuffers, which arguably shouldn't be executable anyway - but who'd want tommap()the framebuffer withPROT_EXEC?Quoted 13 lines
Quoted 12 lines
Here is my tentative plan:
- respin this patch with a small fix to handle the
DMA_ATTR_NON_CONSISTENT(as in ignore it unless actually supported), but keep the name as-is to avoid churn. This should allow 5.3 inclusion and backports
- removeDMA_ATTR_WRITE_COMBINEsupport from mips, probably also 5.3 material.
- move all architectures but arm over to just definepgprot_dmacoherent, including a comment with the above explanation for arm64.Quoted 8 lines
That would be great, thanks.
Quoted 2 lines
- make
DMA_ATTR_WRITE_COMBINEa no-op and schedule it for removal, thus removing the last instances ofarch_dma_mmap_pgprotAll sounds good to me, although I suppose 32-bit Arm platforms without
CONFIG_ARM_DMA_MEM_BUFFERABLEmay run into issues ifDMA_ATTR_WRITE_COMBINEdisappears. Only one way to find out...Looking at the results of grep, I think only OMAP2+ and Exynos may be affected.
However, removing writecombine support from the DMA API is going to have a huge impact for framebuffers on earlier ARMs - that's where we do expect framebuffers to be mapped "uncached/buffered" for performance reasons and not "uncached/unbuffered". It's quite literally the difference between console scrolling being usable and totally unusable.
Quoted 5 lines
Given what I've said above, switching to using buffered mode for normal DMA mappings is data-corrupting risky - as in your filesystem could get fried. I don't think we should play fast and loose with people's data by randomly changing that "because we'd like to", and I don't see that screwing the console is really an option either.
Sorry, I forgot to explain - the reason is dma_alloc_writecombine() internally uses DMA_ATTR_WRITE_COMBINE, which I'd forgotten about when grepping - so there's potentially way more users than my greps above found.
Signature · 3 lines
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/ FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up According to speedtest.net: 11.9Mbps down 500kbps up
Re: [PATCH] dma-mapping: fix page attributes for dma_mmap_*
In reply to Russell King - ARM Linux adminQuoted 6 lines · Russell King - ARM Linux admin
We could have used a different approach, making all IO writes contain a "drain write buffer" instruction, and map DMA memory as "buffered", but as there were no Linux barriers defined to order memory accesses to DMA memory (so, for example, ring buffers can be updated in the correct order) back in those days, using the uncached/unbuffered mode was the sanest and most reliable solution.
Absolutely makes sense so far.
Quoted 15 lines
Quoted 9 lines
Quoted 6 lines
The other really weird things is that in arm32
pgprot_dmacoherentincudes theL_PTE_XNbit, which from my understanding is the no-execture bit, butpgprot_writecombinedoes not. This seems to not very unintentional. So minus that the whole DMA_ATTR_WRITE_COMBІNE seems to be about flagging old arm specific drivers as having the proper barriers in places and otherwise is a no-op.I think it only matters for Armv7 CPUs, but yes, we should probably be setting
L_PTE_XNfor both of these memory types.Conventionally,
pgprot_writecombine()has only been used to change the memory type and not the permissions. Since writecombine memory is still capable of being executed, I don't see any reason to set XN for it.
Quoted 6 lines
If the user wishes to
mmap()using PROT_READ|PROT_EXEC, then is there really a reason for writecombine to set XN overriding the user?That said,
pgprot_writecombine()is mostly used for framebuffers, which arguably shouldn't be executable anyway - but who'd want tommap()the framebuffer withPROT_EXEC?
Well, I was mostly taking about DMA_ATTR_WRITE_COMBINE, which really should include the NX bit even if pgprot_writecombine doesn't, right?
Quoted 9 lines
Quoted 6 lines
Quoted 2 lines
- make
DMA_ATTR_WRITE_COMBINEa no-op and schedule it for removal, thus removing the last instances ofarch_dma_mmap_pgprotAll sounds good to me, although I suppose 32-bit Arm platforms without
CONFIG_ARM_DMA_MEM_BUFFERABLEmay run into issues ifDMA_ATTR_WRITE_COMBINEdisappears. Only one way to find out...Looking at the results of grep, I think only OMAP2+ and Exynos may be affected.
As you mentioned later we also have the dma_alloc_wc wrapper, and a single instance of dma_alloc_writecombine.
Exynos looks like purely ARM v7 from Kconfig, so it shouldn't even be affected.
Quoted 11 lines
However, removing writecombine support from the DMA API is going to have a huge impact for framebuffers on earlier ARMs - that's where we do expect framebuffers to be mapped "uncached/buffered" for performance reasons and not "uncached/unbuffered". It's quite literally the difference between console scrolling being usable and totally unusable.
Given what I've said above, switching to using buffered mode for normal DMA mappings is data-corrupting risky - as in your filesystem could get fried. I don't think we should play fast and loose with people's data by randomly changing that "because we'd like to", and I don't see that screwing the console is really an option either.
Oh well. If we can't make dma_alloc_wc generally safe I fear we'll
have to keep it, but maybe literally limit it to the pre ARM v6 platforms. While pretty much all callers seems platform specific, there actually are a decent number that can only work on ARM v7 or arm64.