]> asedeno.scripts.mit.edu Git - linux.git/commitdiff
Revert "arm64: Revert L1_CACHE_SHIFT back to 6 (64-byte cache line size)"
authorWill Deacon <will.deacon@arm.com>
Tue, 27 Mar 2018 11:04:51 +0000 (12:04 +0100)
committerWill Deacon <will.deacon@arm.com>
Tue, 27 Mar 2018 11:04:51 +0000 (12:04 +0100)
This reverts commit 1f85b42a691cd8329ba82dbcaeec80ac1231b32a.

The internal dma-direct.h API has changed in -next, which collides with
us trying to use it to manage non-coherent DMA devices on systems with
unreasonably large cache writeback granules.

This isn't at all trivial to resolve, so revert our changes for now and
we can revisit this after the merge window. Effectively, this just
restores our behaviour back to that of 4.16.

Signed-off-by: Will Deacon <will.deacon@arm.com>
arch/arm64/Kconfig
arch/arm64/include/asm/cache.h
arch/arm64/include/asm/dma-direct.h [deleted file]
arch/arm64/kernel/cpufeature.c
arch/arm64/mm/dma-mapping.c
arch/arm64/mm/init.c

index fd74c58302328cacca2e6ce636b0df3a10069d74..d4d53c87267eb5fa1b0b2264809f2b7ad50bb78b 100644 (file)
@@ -17,7 +17,6 @@ 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
index 5df5cfe1c1431a763657a19339150b0f5b788159..9bbffc7a301f976458fc032c77472e2520105f00 100644 (file)
@@ -33,7 +33,7 @@
 #define ICACHE_POLICY_VIPT     2
 #define ICACHE_POLICY_PIPT     3
 
-#define L1_CACHE_SHIFT         (6)
+#define L1_CACHE_SHIFT         7
 #define L1_CACHE_BYTES         (1 << L1_CACHE_SHIFT)
 
 /*
@@ -43,7 +43,7 @@
  * cache before the transfer is done, causing old data to be seen by
  * the CPU.
  */
-#define ARCH_DMA_MINALIGN      (128)
+#define ARCH_DMA_MINALIGN      L1_CACHE_BYTES
 
 #ifndef __ASSEMBLY__
 
@@ -77,7 +77,7 @@ static inline u32 cache_type_cwg(void)
 static inline int cache_line_size(void)
 {
        u32 cwg = cache_type_cwg();
-       return cwg ? 4 << cwg : ARCH_DMA_MINALIGN;
+       return cwg ? 4 << cwg : L1_CACHE_BYTES;
 }
 
 #endif /* __ASSEMBLY__ */
diff --git a/arch/arm64/include/asm/dma-direct.h b/arch/arm64/include/asm/dma-direct.h
deleted file mode 100644 (file)
index abb1b40..0000000
+++ /dev/null
@@ -1,43 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0 */
-#ifndef __ASM_DMA_DIRECT_H
-#define __ASM_DMA_DIRECT_H
-
-#include <linux/jump_label.h>
-#include <linux/swiotlb.h>
-
-#include <asm/cache.h>
-
-DECLARE_STATIC_KEY_FALSE(swiotlb_noncoherent_bounce);
-
-static inline dma_addr_t phys_to_dma(struct device *dev, phys_addr_t paddr)
-{
-       dma_addr_t dev_addr = (dma_addr_t)paddr;
-
-       return dev_addr - ((dma_addr_t)dev->dma_pfn_offset << PAGE_SHIFT);
-}
-
-static inline phys_addr_t dma_to_phys(struct device *dev, dma_addr_t dev_addr)
-{
-       phys_addr_t paddr = (phys_addr_t)dev_addr;
-
-       return paddr + ((phys_addr_t)dev->dma_pfn_offset << PAGE_SHIFT);
-}
-
-static inline bool dma_capable(struct device *dev, dma_addr_t addr, size_t size)
-{
-       if (!dev->dma_mask)
-               return false;
-
-       /*
-        * 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;
-}
-
-#endif /* __ASM_DMA_DIRECT_H */
index 071a4548a231bef4a3d3056654674fd538418307..96b15d7b10a86b8ff6ba2779659338544ce73313 100644 (file)
@@ -1624,6 +1624,7 @@ static void __init setup_system_capabilities(void)
 void __init setup_cpu_features(void)
 {
        u32 cwg;
+       int cls;
 
        setup_system_capabilities();
        mark_const_caps_ready();
@@ -1644,9 +1645,13 @@ 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 %d\n",
-                       ARCH_DMA_MINALIGN);
+               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);
 }
 
 static bool __maybe_unused
index 1e9dac8684ca40e9e0d1e6f911c016553d419338..a96ec0181818b90e830898753ea602d77e34b2a9 100644 (file)
@@ -33,7 +33,6 @@
 #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)
@@ -505,14 +504,6 @@ static int __init arm64_dma_init(void)
            max_pfn > (arm64_dma_phys_limit >> PAGE_SHIFT))
                swiotlb = 1;
 
-       if (WARN_TAINT(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())) {
-               swiotlb = 1;
-               static_branch_enable(&swiotlb_noncoherent_bounce);
-       }
-
        return atomic_pool_init();
 }
 arch_initcall(arm64_dma_init);
@@ -891,14 +882,6 @@ 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 (static_branch_unlikely(&swiotlb_noncoherent_bounce) && !coherent)
-               iommu = NULL;
-
        if (!dev->dma_ops)
                dev->dma_ops = &arm64_swiotlb_dma_ops;
 
index 664acf1777995d0967ecfcda14ddc18b5ed34c14..9f3c47acf8ffb1170ad411985dcb1a288dfea2df 100644 (file)
@@ -586,8 +586,7 @@ 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) ||
-           ARCH_DMA_MINALIGN < cache_line_size())
+           max_pfn > (arm64_dma_phys_limit >> PAGE_SHIFT))
                swiotlb_init(1);
        else
                swiotlb_force = SWIOTLB_NO_FORCE;