arm64: Revert L1_CACHE_SHIFT back to 6 (64-byte cache line size)
Catalin Marinas
maintainerPATCH 1/1 arm64: Revert L1_CACHE_SHIFT back to 6 (64-byte cache line size)
Commit 97303480753e ("arm64: Increase the max granular size") increased the cache line size to 128 to match Cavium ThunderX, apparently for some performance benefit which could not be confirmed. This change, however, has an impact on the network packets allocation in certain circumstances, requiring slightly over a 4K page with a significant performance degradation.
This patch reverts L1_CACHE_SHIFT back to 6 (64-byte cache line) while keeping ARCH_DMA_MINALIGN at 128. The cache_line_size() function was changed to default to ARCH_DMA_MINALIGN in the absence of a meaningful CTR_EL0.CWG bit field.
In addition, if a system with ARCH_DMA_MINALIGN < CTR_EL0.CWG is detected, the kernel will force swiotlb bounce buffering for all non-coherent devices since DMA cache maintenance on sub-CWG ranges is not safe, leading to data corruption.
--- arch/arm64/Kconfig | 1 + arch/arm64/include/asm/cache.h | 6 +++--- arch/arm64/include/asm/dma-direct.h | 43 +++++++++++++++++++++++++++++++++++++ arch/arm64/kernel/cpufeature.c | 9 ++------ arch/arm64/mm/dma-mapping.c | 15 +++++++++++++ arch/arm64/mm/init.c | 3 ++- 6 files changed, 66 insertions(+), 11 deletions(-) create mode 100644 arch/arm64/include/asm/dma-direct.h
arch/arm64/Kconfig +1 -0
@@ -17,6 +17,7 @@ config ARM64 select ARCH_HAS_GIGANTIC_PAGE if (MEMORY_ISOLATION && COMPACTION) || CMA select ARCH_HAS_KCOV select ARCH_HAS_MEMBARRIER_SYNC_CORE + select ARCH_HAS_PHYS_TO_DMA select ARCH_HAS_SET_MEMORY select ARCH_HAS_SG_CHAIN select ARCH_HAS_STRICT_KERNEL_RWX
arch/arm64/include/asm/cache.h +3 -3
@@ -29,7 +29,7 @@#define ICACHE_POLICY_VIPT 2#define ICACHE_POLICY_PIPT 3 -#define L1_CACHE_SHIFT 7 +#define L1_CACHE_SHIFT (6)#define L1_CACHE_BYTES (1 << L1_CACHE_SHIFT)/*
@@ -39,7 +39,7 @@* cache before the transfer is done, causing old data to be seen by * the CPU. */ -#define ARCH_DMA_MINALIGN L1_CACHE_BYTES +#define ARCH_DMA_MINALIGN (128)#ifndef __ASSEMBLY__
@@ -73,7 +73,7 @@ static inline u32 cache_type_cwg(void)static inline int cache_line_size (void ) {u32 cwg =cache_type_cwg (); -return cwg ?4 << cwg :L1_CACHE_BYTES ; +return cwg ?4 << cwg :ARCH_DMA_MINALIGN ; }#endif /* __ASSEMBLY__ */
arch/arm64/include/asm/dma-direct.h new file +43 -0
43 lines not on this page · raw message
arch/arm64/kernel/cpufeature.c +2 -7
@@ -1383,7 +1383,6 @@ bool this_cpu_has_cap(unsigned int cap)void __init setup_cpu_features (void ) {u32 cwg; -int cls;/* Set the CPU feature capabilies */ setup_feature_capabilities ();
@@ -1403,13 +1402,9 @@ void __init setup_cpu_features(void)* Check for sane CTR_EL0.CWG value. */ cwg =cache_type_cwg (); - cls =cache_line_size ();if (!cwg) -pr_warn ("No Cache Writeback Granule information, assuming cache line size %d\n" , - cls); -if (L1_CACHE_BYTES < cls) -pr_warn ("L1_CACHE_BYTES smaller than the Cache Writeback Granule (%d < %d)\n" , -L1_CACHE_BYTES , cls); +pr_warn ("No Cache Writeback Granule information, assuming %d\n" , +ARCH_DMA_MINALIGN ); } static bool __maybe_unused
arch/arm64/mm/dma-mapping.c +15 -0
@@ -33,6 +33,7 @@#include <asm/cacheflush.h> static int swiotlb__ro_after_init ; +DEFINE_STATIC_KEY_FALSE (swiotlb_noncoherent_bounce);static pgprot_t __get_dma_pgprot (unsigned long attrs,pgprot_t prot,bool coherent)
@@ -882,6 +883,20 @@ static void __iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,void arch_setup_dma_ops (struct device *dev,u64 dma_base,u64 size,const struct iommu_ops *iommu,bool coherent) { +/* +* Enable swiotlb for buffer bouncing if ARCH_DMA_MINALIGN < CWG. +* dma_capable() forces the actual bounce if the device is +* non-coherent. +*/ +if (WARN_TAINT_ONCE (!coherent &&ARCH_DMA_MINALIGN <cache_line_size (), +TAINT_CPU_OUT_OF_SPEC , +"ARCH_DMA_MINALIGN smaller than CTR_EL0.CWG (%d < %d)" , +ARCH_DMA_MINALIGN ,cache_line_size ())) { +static_branch_enable (&swiotlb_noncoherent_bounce); + swiotlb =1 ; + iommu =NULL ; + } +if (!dev->dma_ops) dev->dma_ops = &arm64_swiotlb_dma_ops;
arch/arm64/mm/init.c +2 -1
@@ -586,7 +586,8 @@ static void __init free_unused_memmap(void)void __init mem_init (void ) {if (swiotlb_force ==SWIOTLB_FORCE || - max_pfn > (arm64_dma_phys_limit >>PAGE_SHIFT )) + max_pfn > (arm64_dma_phys_limit >>PAGE_SHIFT ) || +ARCH_DMA_MINALIGN <cache_line_size ())swiotlb_init (1 );else swiotlb_force =SWIOTLB_NO_FORCE ;
PATCH 1/1 arm64: Revert L1_CACHE_SHIFT back to 6 (64-byte cache line size)
Quoted 31 lines · Catalin Marinas
Commit
97303480753e("arm64: Increase the max granular size") increased the cache line size to 128 to match Cavium ThunderX, apparently for some performance benefit which could not be confirmed. This change, however, has an impact on the network packets allocation in certain circumstances, requiring slightly over a 4K page with a significant performance degradation.This patch reverts
L1_CACHE_SHIFTback to 6 (64-byte cache line) while keepingARCH_DMA_MINALIGNat 128. Thecache_line_size()function was changed to default toARCH_DMA_MINALIGNin the absence of a meaningful CTR_EL0.CWG bit field.In addition, if a system with
ARCH_DMA_MINALIGN< CTR_EL0.CWG is detected, the kernel will force swiotlb bounce buffering for all non-coherent devices since DMA cache maintenance on sub-CWG ranges is not safe, leading to data corruption.Cc Tirumalesh Chalamarla <[email protected]>Cc Timur Tabi <[email protected]>Cc Florian Fainelli <[email protected]>Cc Will Deacon <[email protected]>Signed-off-by Catalin Marinas <[email protected]>--- arch/arm64/Kconfig | 1 + arch/arm64/include/asm/cache.h | 6 +++--- arch/arm64/include/asm/dma-direct.h | 43 +++++++++++++++++++++++++++++++++++++ arch/arm64/kernel/cpufeature.c | 9 ++------ arch/arm64/mm/dma-mapping.c | 15 +++++++++++++ arch/arm64/mm/init.c | 3 ++- 6 files changed, 66 insertions(+), 11 deletions(-) create mode 100644 arch/arm64/include/asm/dma-direct.h
[...]
Quoted 5 lines
arch/arm64/include/asm/dma-direct.h new file +5
-0+static inline bool dma_capable (struct device *dev,dma_addr_t addr,size_t size) +{ +if (!dev->dma_mask) +return false; +
arch/arm64/include/asm/dma-direct.h line 40
Quoted 10 lines
arch/arm64/include/asm/dma-direct.h new file +10
-0+/* +* Force swiotlb buffer bouncing when ARCH_DMA_MINALIGN < CWG. The +* swiotlb bounce buffers are aligned to (1 << IO_TLB_SHIFT). +*/ +if (static_branch_unlikely (&swiotlb_noncoherent_bounce) && + !is_device_dma_coherent (dev) && + !is_swiotlb_buffer (dma_to_phys (dev, addr))) +return false; + +return addr + size -1 <= *dev->dma_mask;
I can't think of a better way to do this and hopefully this won't actually trigger in practice, so:
Will
PATCH 1/1 arm64: Revert L1_CACHE_SHIFT back to 6 (64-byte cache line size)
In reply to Catalin MarinasOn 22/02/18 16:06, Catalin Marinas wrote:
[...]
Quoted 17 lines
arch/arm64/mm/dma-mapping.c +2
-0@@ -33,6 +33,7 @@#include <asm/cacheflush.h> static int swiotlb__ro_after_init ; +DEFINE_STATIC_KEY_FALSE (swiotlb_noncoherent_bounce);static pgprot_t __get_dma_pgprot (unsigned long attrs,pgprot_t prot,bool coherent)@@ -882,6 +883,20 @@ static void __iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,void arch_setup_dma_ops (struct device *dev,u64 dma_base,u64 size,const struct iommu_ops *iommu,bool coherent) { +/*
arch/arm64/mm/dma-mapping.c static void __iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 size, · line 896
Quoted 10 lines
arch/arm64/mm/dma-mapping.c +10
-0+* Enable swiotlb for buffer bouncing if ARCH_DMA_MINALIGN < CWG. +* dma_capable() forces the actual bounce if the device is +* non-coherent. +*/ +if (WARN_TAINT_ONCE (!coherent &&ARCH_DMA_MINALIGN <cache_line_size (), +TAINT_CPU_OUT_OF_SPEC , +"ARCH_DMA_MINALIGN smaller than CTR_EL0.CWG (%d < %d)" , +ARCH_DMA_MINALIGN ,cache_line_size ())) { +static_branch_enable (&swiotlb_noncoherent_bounce); + swiotlb =1 ;
I think it's possible (if a little contrived) for the first non-coherent device to only appear after loading a module from userspace, wherein this would fall foul of __ro_after_init. That said, it would be nice to keep the logic in arm64_dma_init() consistent with its mirror in mem_init() (i.e. add the GWC check there as below), at which point the aforementioned problem no longer applies anyway.
Robin.
Quoted 26 lines
arch/arm64/mm/dma-mapping.c +3
-0+ iommu =NULL ; + } +if (!dev->dma_ops) dev->dma_ops = &arm64_swiotlb_dma_ops;arch/arm64/mm/init.c +2
-1@@ -586,7 +586,8 @@ static void __init free_unused_memmap(void)void __init mem_init (void ) {if (swiotlb_force ==SWIOTLB_FORCE || - max_pfn > (arm64_dma_phys_limit >>PAGE_SHIFT )) + max_pfn > (arm64_dma_phys_limit >>PAGE_SHIFT ) || +ARCH_DMA_MINALIGN <cache_line_size ())swiotlb_init (1 );else swiotlb_force =SWIOTLB_NO_FORCE ;_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Catalin Marinas
maintainerPATCH 1/1 arm64: Revert L1_CACHE_SHIFT back to 6 (64-byte cache line size)
Quoted 26 lines · Robin Murphy
On 22/02/18 16:06, Catalin Marinas wrote:
[...]Quoted 24 lines
arch/arm64/mm/dma-mapping.c +11
-0@@ -33,6 +33,7 @@#include <asm/cacheflush.h> static int swiotlb__ro_after_init ; +DEFINE_STATIC_KEY_FALSE (swiotlb_noncoherent_bounce);static pgprot_t __get_dma_pgprot (unsigned long attrs,pgprot_t prot,bool coherent)@@ -882,6 +883,20 @@ static void __iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,void arch_setup_dma_ops (struct device *dev,u64 dma_base,u64 size,const struct iommu_ops *iommu,bool coherent) { +/* +* Enable swiotlb for buffer bouncing if ARCH_DMA_MINALIGN < CWG. +* dma_capable() forces the actual bounce if the device is +* non-coherent. +*/ +if (WARN_TAINT_ONCE (!coherent &&ARCH_DMA_MINALIGN <cache_line_size (), +TAINT_CPU_OUT_OF_SPEC , +"ARCH_DMA_MINALIGN smaller than CTR_EL0.CWG (%d < %d)" , +ARCH_DMA_MINALIGN ,cache_line_size ())) { +static_branch_enable (&swiotlb_noncoherent_bounce);
Quoted 8 lines
arch/arm64/mm/dma-mapping.c static void __iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 size, · line 896
Quoted 1 line
arch/arm64/mm/dma-mapping.c +1
-0+ swiotlb =1 ;I think it's possible (if a little contrived) for the first non-coherent device to only appear after loading a module from userspace, wherein this would fall foul of
__ro_after_init. That said, it would be nice to keep the logic inarm64_dma_init()consistent with its mirror inmem_init()(i.e. add the GWC check there as below), at which point the aforementioned problem no longer applies anyway.
That's a good point actually. As discovered, with DT we will get at least one call to arch_setup_dma_ops() with coherent == false even when the device does not intend to do any DMA at all. So moving it to arm64_dma_init() wouldn't make any functional difference (but fixing __ro_after_init).
We still force bouncing according to whether the device is coherent or not.
Signature · 1 line
Catalin