provide generic support for uncached segements in dma-direct

provide generic support for uncached segements in dma-direct

[Christoph Hellwig]

Hi all,

can you take a look at this series?  It lifts the support for mips-style
uncached segements to the dma-direct layer, thus removing the need
to have arch_dma_alloc/free routines for these architectures.

[PATCH 1/7] MIPS: remove the _dma_cache_wback_inv export

[Christoph Hellwig]

This export is not used in modular code, which is a good thing as
everyone should use the proper DMA API instead.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 arch/mips/mm/cache.c | 2 --
 1 file changed, 2 deletions(-)

diff --git a/arch/mips/mm/cache.c b/arch/mips/mm/cache.c
index 3da216988672..33b409391ddb 100644
--- a/arch/mips/mm/cache.c
+++ b/arch/mips/mm/cache.c
@@ -62,8 +62,6 @@ void (*_dma_cache_wback_inv)(unsigned long start, unsigned long size);
 void (*_dma_cache_wback)(unsigned long start, unsigned long size);
 void (*_dma_cache_inv)(unsigned long start, unsigned long size);
 
-EXPORT_SYMBOL(_dma_cache_wback_inv);
-
 #endif /* CONFIG_DMA_NONCOHERENT */
 
 /*
-- 
2.20.1

[PATCH 2/7] au1100fb: fix DMA API abuse

[Christoph Hellwig]

Virtual addresses return from dma(m)_alloc_coherent are opaque in what
backs then, and drivers must not poke into them.  Switch the driver
to use the generic DMA API mmap helper to avoid these games.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 drivers/video/fbdev/au1100fb.c | 24 ++++--------------------
 drivers/video/fbdev/au1100fb.h |  1 +
 2 files changed, 5 insertions(+), 20 deletions(-)

diff --git a/drivers/video/fbdev/au1100fb.c b/drivers/video/fbdev/au1100fb.c
index 0adf0683cf08..99941ae1f3a1 100644
--- a/drivers/video/fbdev/au1100fb.c
+++ b/drivers/video/fbdev/au1100fb.c
@@ -340,14 +340,12 @@ int au1100fb_fb_pan_display(struct fb_var_screeninfo *var, struct fb_info *fbi)
  */
 int au1100fb_fb_mmap(struct fb_info *fbi, struct vm_area_struct *vma)
 {
-	struct au1100fb_device *fbdev;
-
-	fbdev = to_au1100fb_device(fbi);
+	struct au1100fb_device *fbdev = to_au1100fb_device(fbi);
 
-	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 	pgprot_val(vma->vm_page_prot) |= (6 << 9); //CCA=6
 
-	return vm_iomap_memory(vma, fbdev->fb_phys, fbdev->fb_len);
+	return dma_mmap_coherent(fbdev->dev, vma, fbdev->fb_mem, fbdev->fb_phys,
+			fbdev->fb_len);
 }
 
 static struct fb_ops au1100fb_ops =
@@ -412,7 +410,6 @@ static int au1100fb_drv_probe(struct platform_device *dev)
 {
 	struct au1100fb_device *fbdev;
 	struct resource *regs_res;
-	unsigned long page;
 	struct clk *c;
 
 	/* Allocate new device private */
@@ -424,6 +421,7 @@ static int au1100fb_drv_probe(struct platform_device *dev)
 		goto failed;
 
 	platform_set_drvdata(dev, (void *)fbdev);
+	fbdev->dev = &dev->dev;
 
 	/* Allocate region for our registers and map them */
 	regs_res = platform_get_resource(dev, IORESOURCE_MEM, 0);
@@ -472,20 +470,6 @@ static int au1100fb_drv_probe(struct platform_device *dev)
 	au1100fb_fix.smem_start = fbdev->fb_phys;
 	au1100fb_fix.smem_len = fbdev->fb_len;
 
-	/*
-	 * Set page reserved so that mmap will work. This is necessary
-	 * since we'll be remapping normal memory.
-	 */
-	for (page = (unsigned long)fbdev->fb_mem;
-	     page < PAGE_ALIGN((unsigned long)fbdev->fb_mem + fbdev->fb_len);
-	     page += PAGE_SIZE) {
-#ifdef CONFIG_DMA_NONCOHERENT
-		SetPageReserved(virt_to_page(CAC_ADDR((void *)page)));
-#else
-		SetPageReserved(virt_to_page(page));
-#endif
-	}
-
 	print_dbg("Framebuffer memory map at %p", fbdev->fb_mem);
 	print_dbg("phys=0x%08x, size=%dK", fbdev->fb_phys, fbdev->fb_len / 1024);
 
diff --git a/drivers/video/fbdev/au1100fb.h b/drivers/video/fbdev/au1100fb.h
index 9af19939a9c6..e7239bceefd3 100644
--- a/drivers/video/fbdev/au1100fb.h
+++ b/drivers/video/fbdev/au1100fb.h
@@ -110,6 +110,7 @@ struct au1100fb_device {
 	dma_addr_t    		fb_phys;
 	int			panel_idx;
 	struct clk		*lcdclk;
+	struct device		*dev;
 };
 
 /********************************************************************/
-- 
2.20.1

Re: [PATCH 2/7] au1100fb: fix DMA API abuse

[Bartlomiej Zolnierkiewicz]

On 04/30/2019 01:00 PM, Christoph Hellwig wrote:
> Virtual addresses return from dma(m)_alloc_coherent are opaque in what
> backs then, and drivers must not poke into them.  Switch the driver
> to use the generic DMA API mmap helper to avoid these games.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>

Acked-by: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>

Best regards,
--
Bartlomiej Zolnierkiewicz
Samsung R&D Institute Poland
Samsung Electronics

Re: [PATCH 2/7] au1100fb: fix DMA API abuse

[Christoph Hellwig]

On Mon, May 06, 2019 at 03:49:35PM +0200, Bartlomiej Zolnierkiewicz wrote:
> 
> On 04/30/2019 01:00 PM, Christoph Hellwig wrote:
> > Virtual addresses return from dma(m)_alloc_coherent are opaque in what
> > backs then, and drivers must not poke into them.  Switch the driver
> > to use the generic DMA API mmap helper to avoid these games.
> > 
> > Signed-off-by: Christoph Hellwig <hch@lst.de>
> 
> Acked-by: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>

Given this actually fixes issues with the current driver, what do you
think of taking it for 5.2 through your tree?  Also au1200fb has
basically the same issues, just without abusing the CAC_ADDR helper,
I could send you a patch for that one as well.

Re: [PATCH 2/7] au1100fb: fix DMA API abuse

[Christoph Hellwig]

FYI, I've merged this patch into the dma-mapping tree to make progress
with the generic uncached segment support for mips.

[PATCH 3/7] dma-mapping: add a Kconfig symbol to indicate arch_dma_prep_coherent presence

[Christoph Hellwig]

Add a Kconfig symbol that indicates an architecture provides a
arch_dma_prep_coherent implementation, and provide a stub otherwise.

This will allow the generic dma-iommu code to use it while still
allowing to be built for cache coherent architectures.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Robin Murphy <robin.murphy@arm.com>
---
 arch/arm64/Kconfig              | 1 +
 arch/csky/Kconfig               | 1 +
 include/linux/dma-noncoherent.h | 6 ++++++
 kernel/dma/Kconfig              | 3 +++
 4 files changed, 11 insertions(+)

diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
index 7e34b9eba5de..adda078d6df7 100644
--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig
@@ -13,6 +13,7 @@ config ARM64
 	select ARCH_HAS_DEVMEM_IS_ALLOWED
 	select ARCH_HAS_DMA_COHERENT_TO_PFN
 	select ARCH_HAS_DMA_MMAP_PGPROT
+	select ARCH_HAS_DMA_PREP_COHERENT
 	select ARCH_HAS_ACPI_TABLE_UPGRADE if ACPI
 	select ARCH_HAS_ELF_RANDOMIZE
 	select ARCH_HAS_FAST_MULTIPLIER
diff --git a/arch/csky/Kconfig b/arch/csky/Kconfig
index 725a115759c9..2c3178848b7d 100644
--- a/arch/csky/Kconfig
+++ b/arch/csky/Kconfig
@@ -1,6 +1,7 @@
 config CSKY
 	def_bool y
 	select ARCH_32BIT_OFF_T
+	select ARCH_HAS_DMA_PREP_COHERENT
 	select ARCH_HAS_SYNC_DMA_FOR_CPU
 	select ARCH_HAS_SYNC_DMA_FOR_DEVICE
 	select ARCH_USE_BUILTIN_BSWAP
diff --git a/include/linux/dma-noncoherent.h b/include/linux/dma-noncoherent.h
index 69b36ed31a99..9741767e400f 100644
--- a/include/linux/dma-noncoherent.h
+++ b/include/linux/dma-noncoherent.h
@@ -72,6 +72,12 @@ static inline void arch_sync_dma_for_cpu_all(struct device *dev)
 }
 #endif /* CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU_ALL */
 
+#ifdef CONFIG_ARCH_HAS_DMA_PREP_COHERENT
 void arch_dma_prep_coherent(struct page *page, size_t size);
+#else
+static inline void arch_dma_prep_coherent(struct page *page, size_t size)
+{
+}
+#endif /* CONFIG_ARCH_HAS_DMA_PREP_COHERENT */
 
 #endif /* _LINUX_DMA_NONCOHERENT_H */
diff --git a/kernel/dma/Kconfig b/kernel/dma/Kconfig
index 52b704e2b97a..83d711f8d665 100644
--- a/kernel/dma/Kconfig
+++ b/kernel/dma/Kconfig
@@ -38,6 +38,9 @@ config ARCH_HAS_SYNC_DMA_FOR_CPU
 config ARCH_HAS_SYNC_DMA_FOR_CPU_ALL
 	bool
 
+config ARCH_HAS_DMA_PREP_COHERENT
+	bool
+
 config ARCH_HAS_DMA_COHERENT_TO_PFN
 	bool
 
-- 
2.20.1

[PATCH 4/7] dma-direct: provide generic support for uncached kernel segments

[Christoph Hellwig]

A few architectures support uncached kernel segments.  In that case we get
an uncached mapping for a given physica address by using an offset in the
uncached segement.  Implement support for this scheme in the generic
dma-direct code instead of duplicating it in arch hooks.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 arch/Kconfig                    |  8 ++++++++
 include/linux/dma-noncoherent.h |  3 +++
 kernel/dma/direct.c             | 18 ++++++++++++++++--
 3 files changed, 27 insertions(+), 2 deletions(-)

diff --git a/arch/Kconfig b/arch/Kconfig
index 33687dddd86a..ea22a8c894ec 100644
--- a/arch/Kconfig
+++ b/arch/Kconfig
@@ -249,6 +249,14 @@ config ARCH_HAS_FORTIFY_SOURCE
 config ARCH_HAS_SET_MEMORY
 	bool
 
+#
+# Select if arch has an uncached kernel segment and provides the
+# uncached_kernel_address / cached_kernel_address symbols to use it
+#
+config ARCH_HAS_UNCACHED_SEGMENT
+	select ARCH_HAS_DMA_PREP_COHERENT
+	bool
+
 # Select if arch init_task must go in the __init_task_data section
 config ARCH_TASK_STRUCT_ON_STACK
        bool
diff --git a/include/linux/dma-noncoherent.h b/include/linux/dma-noncoherent.h
index 9741767e400f..7e0126a04e02 100644
--- a/include/linux/dma-noncoherent.h
+++ b/include/linux/dma-noncoherent.h
@@ -80,4 +80,7 @@ static inline void arch_dma_prep_coherent(struct page *page, size_t size)
 }
 #endif /* CONFIG_ARCH_HAS_DMA_PREP_COHERENT */
 
+void *uncached_kernel_address(void *addr);
+void *cached_kernel_address(void *addr);
+
 #endif /* _LINUX_DMA_NONCOHERENT_H */
diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
index 2c2772e9702a..d15a535c3e67 100644
--- a/kernel/dma/direct.c
+++ b/kernel/dma/direct.c
@@ -164,6 +164,13 @@ void *dma_direct_alloc_pages(struct device *dev, size_t size,
 	}
 
 	ret = page_address(page);
+
+	if (IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
+	    !dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_NON_CONSISTENT)) {
+		arch_dma_prep_coherent(page, size);
+		ret = uncached_kernel_address(ret);
+	}
+
 	if (force_dma_unencrypted()) {
 		set_memory_decrypted((unsigned long)ret, 1 << get_order(size));
 		*dma_handle = __phys_to_dma(dev, page_to_phys(page));
@@ -171,6 +178,7 @@ void *dma_direct_alloc_pages(struct device *dev, size_t size,
 		*dma_handle = phys_to_dma(dev, page_to_phys(page));
 	}
 	memset(ret, 0, size);
+
 	return ret;
 }
 
@@ -189,13 +197,18 @@ void dma_direct_free_pages(struct device *dev, size_t size, void *cpu_addr,
 
 	if (force_dma_unencrypted())
 		set_memory_encrypted((unsigned long)cpu_addr, 1 << page_order);
+
+	if (IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
+	    !dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_NON_CONSISTENT))
+		cpu_addr = cached_kernel_address(cpu_addr);
 	__dma_direct_free_pages(dev, size, virt_to_page(cpu_addr));
 }
 
 void *dma_direct_alloc(struct device *dev, size_t size,
 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
 {
-	if (!dev_is_dma_coherent(dev))
+	if (!IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
+	    !dev_is_dma_coherent(dev))
 		return arch_dma_alloc(dev, size, dma_handle, gfp, attrs);
 	return dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
 }
@@ -203,7 +216,8 @@ void *dma_direct_alloc(struct device *dev, size_t size,
 void dma_direct_free(struct device *dev, size_t size,
 		void *cpu_addr, dma_addr_t dma_addr, unsigned long attrs)
 {
-	if (!dev_is_dma_coherent(dev))
+	if (!IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
+	    !dev_is_dma_coherent(dev))
 		arch_dma_free(dev, size, cpu_addr, dma_addr, attrs);
 	else
 		dma_direct_free_pages(dev, size, cpu_addr, dma_addr, attrs);
-- 
2.20.1

Re: [PATCH 4/7] dma-direct: provide generic support for uncached kernel segments

[Paul Burton]

Hi Christoph,

On Tue, Apr 30, 2019 at 07:00:29AM -0400, Christoph Hellwig wrote:
> diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
> index 2c2772e9702a..d15a535c3e67 100644
> --- a/kernel/dma/direct.c
> +++ b/kernel/dma/direct.c
> @@ -164,6 +164,13 @@ void *dma_direct_alloc_pages(struct device *dev, size_t size,
>  	}
>  
>  	ret = page_address(page);
> +
> +	if (IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
> +	    !dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_NON_CONSISTENT)) {
> +		arch_dma_prep_coherent(page, size);
> +		ret = uncached_kernel_address(ret);
> +	}
> +
>  	if (force_dma_unencrypted()) {
>  		set_memory_decrypted((unsigned long)ret, 1 << get_order(size));
>  		*dma_handle = __phys_to_dma(dev, page_to_phys(page));
> @@ -171,6 +178,7 @@ void *dma_direct_alloc_pages(struct device *dev, size_t size,
>  		*dma_handle = phys_to_dma(dev, page_to_phys(page));
>  	}
>  	memset(ret, 0, size);
> +
>  	return ret;
>  }

I'm not so sure about this part though.

On MIPS we currently don't clear the allocated memory with memset. Is
doing that really necessary?

If it is necessary then as-is this code will clear the allocated memory
using uncached writes which will be pretty slow. It would be much more
efficient to perform the memset before arch_dma_prep_coherent() & before
converting ret to an uncached address.

Thanks,
    Paul

Re: [PATCH 4/7] dma-direct: provide generic support for uncached kernel segments

[Christoph Hellwig]

[-- Attachment #1: Type: text/plain, Size: 564 bytes --]

On Wed, May 01, 2019 at 05:18:59PM +0000, Paul Burton wrote:
> I'm not so sure about this part though.
> 
> On MIPS we currently don't clear the allocated memory with memset. Is
> doing that really necessary?

We are clearling it on mips, it is inside dma_direct_alloc_pages.

> If it is necessary then as-is this code will clear the allocated memory
> using uncached writes which will be pretty slow. It would be much more
> efficient to perform the memset before arch_dma_prep_coherent() & before
> converting ret to an uncached address.

Yes, we could do that.

[-- Attachment #2: 0001-dma-direct-provide-generic-support-for-uncached-kern.patch --]
[-- Type: text/x-patch, Size: 3565 bytes --]

From 247ca658ebeb7c8d04918747ec8a0da45c36bcb8 Mon Sep 17 00:00:00 2001
From: Christoph Hellwig <hch@lst.de>
Date: Sun, 28 Apr 2019 13:23:26 -0500
Subject: dma-direct: provide generic support for uncached kernel segments

A few architectures support uncached kernel segments.  In that case we get
an uncached mapping for a given physica address by using an offset in the
uncached segement.  Implement support for this scheme in the generic
dma-direct code instead of duplicating it in arch hooks.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 arch/Kconfig                    |  8 ++++++++
 include/linux/dma-noncoherent.h |  3 +++
 kernel/dma/direct.c             | 17 +++++++++++++++--
 3 files changed, 26 insertions(+), 2 deletions(-)

diff --git a/arch/Kconfig b/arch/Kconfig
index 33687dddd86a..ea22a8c894ec 100644
--- a/arch/Kconfig
+++ b/arch/Kconfig
@@ -249,6 +249,14 @@ config ARCH_HAS_FORTIFY_SOURCE
 config ARCH_HAS_SET_MEMORY
 	bool
 
+#
+# Select if arch has an uncached kernel segment and provides the
+# uncached_kernel_address / cached_kernel_address symbols to use it
+#
+config ARCH_HAS_UNCACHED_SEGMENT
+	select ARCH_HAS_DMA_PREP_COHERENT
+	bool
+
 # Select if arch init_task must go in the __init_task_data section
 config ARCH_TASK_STRUCT_ON_STACK
        bool
diff --git a/include/linux/dma-noncoherent.h b/include/linux/dma-noncoherent.h
index 9741767e400f..7e0126a04e02 100644
--- a/include/linux/dma-noncoherent.h
+++ b/include/linux/dma-noncoherent.h
@@ -80,4 +80,7 @@ static inline void arch_dma_prep_coherent(struct page *page, size_t size)
 }
 #endif /* CONFIG_ARCH_HAS_DMA_PREP_COHERENT */
 
+void *uncached_kernel_address(void *addr);
+void *cached_kernel_address(void *addr);
+
 #endif /* _LINUX_DMA_NONCOHERENT_H */
diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
index 2c2772e9702a..6688e1cee7d1 100644
--- a/kernel/dma/direct.c
+++ b/kernel/dma/direct.c
@@ -171,6 +171,13 @@ void *dma_direct_alloc_pages(struct device *dev, size_t size,
 		*dma_handle = phys_to_dma(dev, page_to_phys(page));
 	}
 	memset(ret, 0, size);
+
+	if (IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
+	    !dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_NON_CONSISTENT)) {
+		arch_dma_prep_coherent(page, size);
+		ret = uncached_kernel_address(ret);
+	}
+
 	return ret;
 }
 
@@ -189,13 +196,18 @@ void dma_direct_free_pages(struct device *dev, size_t size, void *cpu_addr,
 
 	if (force_dma_unencrypted())
 		set_memory_encrypted((unsigned long)cpu_addr, 1 << page_order);
+
+	if (IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
+	    !dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_NON_CONSISTENT))
+		cpu_addr = cached_kernel_address(cpu_addr);
 	__dma_direct_free_pages(dev, size, virt_to_page(cpu_addr));
 }
 
 void *dma_direct_alloc(struct device *dev, size_t size,
 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
 {
-	if (!dev_is_dma_coherent(dev))
+	if (!IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
+	    !dev_is_dma_coherent(dev))
 		return arch_dma_alloc(dev, size, dma_handle, gfp, attrs);
 	return dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
 }
@@ -203,7 +215,8 @@ void *dma_direct_alloc(struct device *dev, size_t size,
 void dma_direct_free(struct device *dev, size_t size,
 		void *cpu_addr, dma_addr_t dma_addr, unsigned long attrs)
 {
-	if (!dev_is_dma_coherent(dev))
+	if (!IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
+	    !dev_is_dma_coherent(dev))
 		arch_dma_free(dev, size, cpu_addr, dma_addr, attrs);
 	else
 		dma_direct_free_pages(dev, size, cpu_addr, dma_addr, attrs);
-- 
2.20.1

Re: [PATCH 4/7] dma-direct: provide generic support for uncached kernel segments

[Paul Burton]

Hi Christoph,

On Wed, May 01, 2019 at 07:29:12PM +0200, Christoph Hellwig wrote:
> On Wed, May 01, 2019 at 05:18:59PM +0000, Paul Burton wrote:
> > I'm not so sure about this part though.
> > 
> > On MIPS we currently don't clear the allocated memory with memset. Is
> > doing that really necessary?
> 
> We are clearling it on mips, it is inside dma_direct_alloc_pages.

Ah, of course, I clearly require more caffeine :)

> > If it is necessary then as-is this code will clear the allocated memory
> > using uncached writes which will be pretty slow. It would be much more
> > efficient to perform the memset before arch_dma_prep_coherent() & before
> > converting ret to an uncached address.
> 
> Yes, we could do that.

Great; using cached writes would match the existing MIPS behavior.

Thanks,
    Paul

Re: [PATCH 4/7] dma-direct: provide generic support for uncached kernel segments

[Christoph Hellwig]

On Wed, May 01, 2019 at 05:40:34PM +0000, Paul Burton wrote:
> > > If it is necessary then as-is this code will clear the allocated memory
> > > using uncached writes which will be pretty slow. It would be much more
> > > efficient to perform the memset before arch_dma_prep_coherent() & before
> > > converting ret to an uncached address.
> > 
> > Yes, we could do that.
> 
> Great; using cached writes would match the existing MIPS behavior.

Can you test the stack with the two updated patches and ack them if
they are fine?  That would allow getting at least the infrastructure
and mips in for this merge window.

Re: [PATCH 4/7] dma-direct: provide generic support for uncached kernel segments

[Paul Burton]

Hi Christoph,

On Wed, May 01, 2019 at 07:49:05PM +0200, Christoph Hellwig wrote:
> On Wed, May 01, 2019 at 05:40:34PM +0000, Paul Burton wrote:
> > > > If it is necessary then as-is this code will clear the allocated memory
> > > > using uncached writes which will be pretty slow. It would be much more
> > > > efficient to perform the memset before arch_dma_prep_coherent() & before
> > > > converting ret to an uncached address.
> > > 
> > > Yes, we could do that.
> > 
> > Great; using cached writes would match the existing MIPS behavior.
> 
> Can you test the stack with the two updated patches and ack them if
> they are fine?  That would allow getting at least the infrastructure
> and mips in for this merge window.

Did you send a v2 of this patch?

If so it hasn't showed up in my inbox, nor on the linux-mips archive on
lore.kernel.org.

Thanks,
    Paul

Re: [PATCH 4/7] dma-direct: provide generic support for uncached kernel segments

[Christoph Hellwig]

On Thu, May 02, 2019 at 12:08:01AM +0000, Paul Burton wrote:
> > Can you test the stack with the two updated patches and ack them if
> > they are fine?  That would allow getting at least the infrastructure
> > and mips in for this merge window.
> 
> Did you send a v2 of this patch?
> 
> If so it hasn't showed up in my inbox, nor on the linux-mips archive on
> lore.kernel.org.

I did earlier in this thread.  Here it is again:

---
From 247ca658ebeb7c8d04918747ec8a0da45c36bcb8 Mon Sep 17 00:00:00 2001
From: Christoph Hellwig <hch@lst.de>
Date: Sun, 28 Apr 2019 13:23:26 -0500
Subject: dma-direct: provide generic support for uncached kernel segments

A few architectures support uncached kernel segments.  In that case we get
an uncached mapping for a given physica address by using an offset in the
uncached segement.  Implement support for this scheme in the generic
dma-direct code instead of duplicating it in arch hooks.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 arch/Kconfig                    |  8 ++++++++
 include/linux/dma-noncoherent.h |  3 +++
 kernel/dma/direct.c             | 17 +++++++++++++++--
 3 files changed, 26 insertions(+), 2 deletions(-)

diff --git a/arch/Kconfig b/arch/Kconfig
index 33687dddd86a..ea22a8c894ec 100644
--- a/arch/Kconfig
+++ b/arch/Kconfig
@@ -249,6 +249,14 @@ config ARCH_HAS_FORTIFY_SOURCE
 config ARCH_HAS_SET_MEMORY
 	bool
 
+#
+# Select if arch has an uncached kernel segment and provides the
+# uncached_kernel_address / cached_kernel_address symbols to use it
+#
+config ARCH_HAS_UNCACHED_SEGMENT
+	select ARCH_HAS_DMA_PREP_COHERENT
+	bool
+
 # Select if arch init_task must go in the __init_task_data section
 config ARCH_TASK_STRUCT_ON_STACK
        bool
diff --git a/include/linux/dma-noncoherent.h b/include/linux/dma-noncoherent.h
index 9741767e400f..7e0126a04e02 100644
--- a/include/linux/dma-noncoherent.h
+++ b/include/linux/dma-noncoherent.h
@@ -80,4 +80,7 @@ static inline void arch_dma_prep_coherent(struct page *page, size_t size)
 }
 #endif /* CONFIG_ARCH_HAS_DMA_PREP_COHERENT */
 
+void *uncached_kernel_address(void *addr);
+void *cached_kernel_address(void *addr);
+
 #endif /* _LINUX_DMA_NONCOHERENT_H */
diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
index 2c2772e9702a..6688e1cee7d1 100644
--- a/kernel/dma/direct.c
+++ b/kernel/dma/direct.c
@@ -171,6 +171,13 @@ void *dma_direct_alloc_pages(struct device *dev, size_t size,
 		*dma_handle = phys_to_dma(dev, page_to_phys(page));
 	}
 	memset(ret, 0, size);
+
+	if (IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
+	    !dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_NON_CONSISTENT)) {
+		arch_dma_prep_coherent(page, size);
+		ret = uncached_kernel_address(ret);
+	}
+
 	return ret;
 }
 
@@ -189,13 +196,18 @@ void dma_direct_free_pages(struct device *dev, size_t size, void *cpu_addr,
 
 	if (force_dma_unencrypted())
 		set_memory_encrypted((unsigned long)cpu_addr, 1 << page_order);
+
+	if (IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
+	    !dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_NON_CONSISTENT))
+		cpu_addr = cached_kernel_address(cpu_addr);
 	__dma_direct_free_pages(dev, size, virt_to_page(cpu_addr));
 }
 
 void *dma_direct_alloc(struct device *dev, size_t size,
 		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
 {
-	if (!dev_is_dma_coherent(dev))
+	if (!IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
+	    !dev_is_dma_coherent(dev))
 		return arch_dma_alloc(dev, size, dma_handle, gfp, attrs);
 	return dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
 }
@@ -203,7 +215,8 @@ void *dma_direct_alloc(struct device *dev, size_t size,
 void dma_direct_free(struct device *dev, size_t size,
 		void *cpu_addr, dma_addr_t dma_addr, unsigned long attrs)
 {
-	if (!dev_is_dma_coherent(dev))
+	if (!IS_ENABLED(CONFIG_ARCH_HAS_UNCACHED_SEGMENT) &&
+	    !dev_is_dma_coherent(dev))
 		arch_dma_free(dev, size, cpu_addr, dma_addr, attrs);
 	else
 		dma_direct_free_pages(dev, size, cpu_addr, dma_addr, attrs);
-- 
2.20.1

[PATCH 5/7] MIPS: use the generic uncached segment support in dma-direct

[Christoph Hellwig]

Stop providing our arch alloc/free hooks and just expose the segment
offset instead.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 arch/mips/Kconfig              |  1 +
 arch/mips/include/asm/page.h   |  3 ---
 arch/mips/jazz/jazzdma.c       |  6 ------
 arch/mips/mm/dma-noncoherent.c | 27 ++++++++++-----------------
 4 files changed, 11 insertions(+), 26 deletions(-)

diff --git a/arch/mips/Kconfig b/arch/mips/Kconfig
index 4a5f5b0ee9a9..cde4b490f3c7 100644
--- a/arch/mips/Kconfig
+++ b/arch/mips/Kconfig
@@ -9,6 +9,7 @@ config MIPS
 	select ARCH_HAS_ELF_RANDOMIZE
 	select ARCH_HAS_TICK_BROADCAST if GENERIC_CLOCKEVENTS_BROADCAST
 	select ARCH_HAS_UBSAN_SANITIZE_ALL
+	select ARCH_HAS_UNCACHED_SEGMENT
 	select ARCH_SUPPORTS_UPROBES
 	select ARCH_USE_BUILTIN_BSWAP
 	select ARCH_USE_CMPXCHG_LOCKREF if 64BIT
diff --git a/arch/mips/include/asm/page.h b/arch/mips/include/asm/page.h
index 6b31c93b5eaa..23e0f1386e04 100644
--- a/arch/mips/include/asm/page.h
+++ b/arch/mips/include/asm/page.h
@@ -258,9 +258,6 @@ extern int __virt_addr_valid(const volatile void *kaddr);
 	 ((current->personality & READ_IMPLIES_EXEC) ? VM_EXEC : 0) | \
 	 VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC)
 
-#define UNCAC_ADDR(addr)	(UNCAC_BASE + __pa(addr))
-#define CAC_ADDR(addr)		((unsigned long)__va((addr) - UNCAC_BASE))
-
 #include <asm-generic/memory_model.h>
 #include <asm-generic/getorder.h>
 
diff --git a/arch/mips/jazz/jazzdma.c b/arch/mips/jazz/jazzdma.c
index bedb5047aff3..1804dc9d8136 100644
--- a/arch/mips/jazz/jazzdma.c
+++ b/arch/mips/jazz/jazzdma.c
@@ -575,10 +575,6 @@ static void *jazz_dma_alloc(struct device *dev, size_t size,
 		return NULL;
 	}
 
-	if (!(attrs & DMA_ATTR_NON_CONSISTENT)) {
-		dma_cache_wback_inv((unsigned long)ret, size);
-		ret = (void *)UNCAC_ADDR(ret);
-	}
 	return ret;
 }
 
@@ -586,8 +582,6 @@ static void jazz_dma_free(struct device *dev, size_t size, void *vaddr,
 		dma_addr_t dma_handle, unsigned long attrs)
 {
 	vdma_free(dma_handle);
-	if (!(attrs & DMA_ATTR_NON_CONSISTENT))
-		vaddr = (void *)CAC_ADDR((unsigned long)vaddr);
 	dma_direct_free_pages(dev, size, vaddr, dma_handle, attrs);
 }
 
diff --git a/arch/mips/mm/dma-noncoherent.c b/arch/mips/mm/dma-noncoherent.c
index f9549d2fbea3..f739f42c9d3c 100644
--- a/arch/mips/mm/dma-noncoherent.c
+++ b/arch/mips/mm/dma-noncoherent.c
@@ -44,33 +44,26 @@ static inline bool cpu_needs_post_dma_flush(struct device *dev)
 	}
 }
 
-void *arch_dma_alloc(struct device *dev, size_t size,
-		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
+void arch_dma_prep_coherent(struct page *page, size_t size)
 {
-	void *ret;
-
-	ret = dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
-	if (ret && !(attrs & DMA_ATTR_NON_CONSISTENT)) {
-		dma_cache_wback_inv((unsigned long) ret, size);
-		ret = (void *)UNCAC_ADDR(ret);
-	}
+	if (!PageHighMem(page))
+		dma_cache_wback_inv((unsigned long)page_address(page), size);
+}
 
-	return ret;
+void *uncached_kernel_address(void *addr)
+{
+	return (void *)(__pa(addr) + UNCAC_BASE);
 }
 
-void arch_dma_free(struct device *dev, size_t size, void *cpu_addr,
-		dma_addr_t dma_addr, unsigned long attrs)
+void *cached_kernel_address(void *addr)
 {
-	if (!(attrs & DMA_ATTR_NON_CONSISTENT))
-		cpu_addr = (void *)CAC_ADDR((unsigned long)cpu_addr);
-	dma_direct_free_pages(dev, size, cpu_addr, dma_addr, attrs);
+	return __va(addr) - UNCAC_BASE;
 }
 
 long arch_dma_coherent_to_pfn(struct device *dev, void *cpu_addr,
 		dma_addr_t dma_addr)
 {
-	unsigned long addr = CAC_ADDR((unsigned long)cpu_addr);
-	return page_to_pfn(virt_to_page((void *)addr));
+	return page_to_pfn(virt_to_page(cached_kernel_address(cpu_addr)));
 }
 
 pgprot_t arch_dma_mmap_pgprot(struct device *dev, pgprot_t prot,
-- 
2.20.1

Re: [PATCH 5/7] MIPS: use the generic uncached segment support in dma-direct

[Paul Burton]

Hi Christoph,

On Tue, Apr 30, 2019 at 07:00:30AM -0400, Christoph Hellwig wrote:
> diff --git a/arch/mips/mm/dma-noncoherent.c b/arch/mips/mm/dma-noncoherent.c
> index f9549d2fbea3..f739f42c9d3c 100644
> --- a/arch/mips/mm/dma-noncoherent.c
> +++ b/arch/mips/mm/dma-noncoherent.c
> @@ -44,33 +44,26 @@ static inline bool cpu_needs_post_dma_flush(struct device *dev)
>  	}
>  }
>  
> -void *arch_dma_alloc(struct device *dev, size_t size,
> -		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
> +void arch_dma_prep_coherent(struct page *page, size_t size)
>  {
> -	void *ret;
> -
> -	ret = dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
> -	if (ret && !(attrs & DMA_ATTR_NON_CONSISTENT)) {
> -		dma_cache_wback_inv((unsigned long) ret, size);
> -		ret = (void *)UNCAC_ADDR(ret);
> -	}
> +	if (!PageHighMem(page))
> +		dma_cache_wback_inv((unsigned long)page_address(page), size);
> +}

This series looks like a nice cleanup to me - the one thing that puzzles
me is the !PageHighMem check above.

As far as I can see arch_dma_prep_coherent() should never be called with
a highmem page, so would it make more sense to either drop this check or
perhaps wrap it in a WARN_ON()?

Thanks,
    Paul

Re: [PATCH 5/7] MIPS: use the generic uncached segment support in dma-direct

[Christoph Hellwig]

On Tue, Apr 30, 2019 at 08:10:43PM +0000, Paul Burton wrote:
> This series looks like a nice cleanup to me - the one thing that puzzles
> me is the !PageHighMem check above.
> 
> As far as I can see arch_dma_prep_coherent() should never be called with
> a highmem page, so would it make more sense to either drop this check or
> perhaps wrap it in a WARN_ON()?

dma_alloc_from_contigous can return highmem pages depending on where
the CMA area is located.  But given that these pages don't have a
direct kernel mapping we also shouldn't have to flush the caches
for them.

Re: [PATCH 5/7] MIPS: use the generic uncached segment support in dma-direct

[Paul Burton]

Hi Christoph,

On Tue, Apr 30, 2019 at 10:29:47PM +0200, Christoph Hellwig wrote:
> On Tue, Apr 30, 2019 at 08:10:43PM +0000, Paul Burton wrote:
> > This series looks like a nice cleanup to me - the one thing that puzzles
> > me is the !PageHighMem check above.
> > 
> > As far as I can see arch_dma_prep_coherent() should never be called with
> > a highmem page, so would it make more sense to either drop this check or
> > perhaps wrap it in a WARN_ON()?
> 
> dma_alloc_from_contigous can return highmem pages depending on where
> the CMA area is located.  But given that these pages don't have a
> direct kernel mapping we also shouldn't have to flush the caches
> for them.

Right but dma_direct_alloc_pages() already checks for the PageHighMem
case & returns before ever calling arch_dma_prep_coherent(), no?

Thanks,
    Paul

Re: [PATCH 5/7] MIPS: use the generic uncached segment support in dma-direct

[Christoph Hellwig]

On Tue, Apr 30, 2019 at 09:11:07PM +0000, Paul Burton wrote:
> Right but dma_direct_alloc_pages() already checks for the PageHighMem
> case & returns before ever calling arch_dma_prep_coherent(), no?

True.  And of course it can't be remapped into the uncached segment
anyway.  So yes, we should drop it.  Eventually I want to add generic
support for DMA_ATTR_NO_KERNEL_MAPPING, but that'll involved auditing
all instances anyway.

[PATCH 5/7 v2] MIPS: use the generic uncached segment support in dma-direct

[Christoph Hellwig]

Stop providing our arch alloc/free hooks and just expose the segment
offset instead.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 arch/mips/Kconfig              |  1 +
 arch/mips/include/asm/page.h   |  3 ---
 arch/mips/jazz/jazzdma.c       |  6 ------
 arch/mips/mm/dma-noncoherent.c | 26 +++++++++-----------------
 4 files changed, 10 insertions(+), 26 deletions(-)

diff --git a/arch/mips/Kconfig b/arch/mips/Kconfig
index 4a5f5b0ee9a9..cde4b490f3c7 100644
--- a/arch/mips/Kconfig
+++ b/arch/mips/Kconfig
@@ -9,6 +9,7 @@ config MIPS
 	select ARCH_HAS_ELF_RANDOMIZE
 	select ARCH_HAS_TICK_BROADCAST if GENERIC_CLOCKEVENTS_BROADCAST
 	select ARCH_HAS_UBSAN_SANITIZE_ALL
+	select ARCH_HAS_UNCACHED_SEGMENT
 	select ARCH_SUPPORTS_UPROBES
 	select ARCH_USE_BUILTIN_BSWAP
 	select ARCH_USE_CMPXCHG_LOCKREF if 64BIT
diff --git a/arch/mips/include/asm/page.h b/arch/mips/include/asm/page.h
index 6b31c93b5eaa..23e0f1386e04 100644
--- a/arch/mips/include/asm/page.h
+++ b/arch/mips/include/asm/page.h
@@ -258,9 +258,6 @@ extern int __virt_addr_valid(const volatile void *kaddr);
 	 ((current->personality & READ_IMPLIES_EXEC) ? VM_EXEC : 0) | \
 	 VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC)
 
-#define UNCAC_ADDR(addr)	(UNCAC_BASE + __pa(addr))
-#define CAC_ADDR(addr)		((unsigned long)__va((addr) - UNCAC_BASE))
-
 #include <asm-generic/memory_model.h>
 #include <asm-generic/getorder.h>
 
diff --git a/arch/mips/jazz/jazzdma.c b/arch/mips/jazz/jazzdma.c
index bedb5047aff3..1804dc9d8136 100644
--- a/arch/mips/jazz/jazzdma.c
+++ b/arch/mips/jazz/jazzdma.c
@@ -575,10 +575,6 @@ static void *jazz_dma_alloc(struct device *dev, size_t size,
 		return NULL;
 	}
 
-	if (!(attrs & DMA_ATTR_NON_CONSISTENT)) {
-		dma_cache_wback_inv((unsigned long)ret, size);
-		ret = (void *)UNCAC_ADDR(ret);
-	}
 	return ret;
 }
 
@@ -586,8 +582,6 @@ static void jazz_dma_free(struct device *dev, size_t size, void *vaddr,
 		dma_addr_t dma_handle, unsigned long attrs)
 {
 	vdma_free(dma_handle);
-	if (!(attrs & DMA_ATTR_NON_CONSISTENT))
-		vaddr = (void *)CAC_ADDR((unsigned long)vaddr);
 	dma_direct_free_pages(dev, size, vaddr, dma_handle, attrs);
 }
 
diff --git a/arch/mips/mm/dma-noncoherent.c b/arch/mips/mm/dma-noncoherent.c
index f9549d2fbea3..ed56c6fa7be2 100644
--- a/arch/mips/mm/dma-noncoherent.c
+++ b/arch/mips/mm/dma-noncoherent.c
@@ -44,33 +44,25 @@ static inline bool cpu_needs_post_dma_flush(struct device *dev)
 	}
 }
 
-void *arch_dma_alloc(struct device *dev, size_t size,
-		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
+void arch_dma_prep_coherent(struct page *page, size_t size)
 {
-	void *ret;
-
-	ret = dma_direct_alloc_pages(dev, size, dma_handle, gfp, attrs);
-	if (ret && !(attrs & DMA_ATTR_NON_CONSISTENT)) {
-		dma_cache_wback_inv((unsigned long) ret, size);
-		ret = (void *)UNCAC_ADDR(ret);
-	}
+	dma_cache_wback_inv((unsigned long)page_address(page), size);
+}
 
-	return ret;
+void *uncached_kernel_address(void *addr)
+{
+	return (void *)(__pa(addr) + UNCAC_BASE);
 }
 
-void arch_dma_free(struct device *dev, size_t size, void *cpu_addr,
-		dma_addr_t dma_addr, unsigned long attrs)
+void *cached_kernel_address(void *addr)
 {
-	if (!(attrs & DMA_ATTR_NON_CONSISTENT))
-		cpu_addr = (void *)CAC_ADDR((unsigned long)cpu_addr);
-	dma_direct_free_pages(dev, size, cpu_addr, dma_addr, attrs);
+	return __va(addr) - UNCAC_BASE;
 }
 
 long arch_dma_coherent_to_pfn(struct device *dev, void *cpu_addr,
 		dma_addr_t dma_addr)
 {
-	unsigned long addr = CAC_ADDR((unsigned long)cpu_addr);
-	return page_to_pfn(virt_to_page((void *)addr));
+	return page_to_pfn(virt_to_page(cached_kernel_address(cpu_addr)));
 }
 
 pgprot_t arch_dma_mmap_pgprot(struct device *dev, pgprot_t prot,
-- 
2.20.1

Re: [PATCH 5/7 v2] MIPS: use the generic uncached segment support in dma-direct

[Paul Burton]

Hi Christoph,

On Wed, May 01, 2019 at 03:13:39PM +0200, Christoph Hellwig wrote:
> Stop providing our arch alloc/free hooks and just expose the segment
> offset instead.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> ---
>  arch/mips/Kconfig              |  1 +
>  arch/mips/include/asm/page.h   |  3 ---
>  arch/mips/jazz/jazzdma.c       |  6 ------
>  arch/mips/mm/dma-noncoherent.c | 26 +++++++++-----------------
>  4 files changed, 10 insertions(+), 26 deletions(-)

This one looks good to me now, for patches 1 & 5:

  Acked-by: Paul Burton <paul.burton@mips.com>

Thanks,
    Paul

Re: [PATCH 5/7 v2] MIPS: use the generic uncached segment support in dma-direct

[Christoph Hellwig]

On Wed, May 01, 2019 at 05:13:57PM +0000, Paul Burton wrote:
> Hi Christoph,
> 
> On Wed, May 01, 2019 at 03:13:39PM +0200, Christoph Hellwig wrote:
> > Stop providing our arch alloc/free hooks and just expose the segment
> > offset instead.
> > 
> > Signed-off-by: Christoph Hellwig <hch@lst.de>
> > ---
> >  arch/mips/Kconfig              |  1 +
> >  arch/mips/include/asm/page.h   |  3 ---
> >  arch/mips/jazz/jazzdma.c       |  6 ------
> >  arch/mips/mm/dma-noncoherent.c | 26 +++++++++-----------------
> >  4 files changed, 10 insertions(+), 26 deletions(-)
> 
> This one looks good to me now, for patches 1 & 5:
> 
>   Acked-by: Paul Burton <paul.burton@mips.com>

Thanks, I've merged thos into the dma-mapping tree.

Re: [PATCH 5/7 v2] MIPS: use the generic uncached segment support in dma-direct

[Arnd Bergmann]

On Mon, Jun 3, 2019 at 8:50 AM Christoph Hellwig <hch@lst.de> wrote:
>
> On Wed, May 01, 2019 at 05:13:57PM +0000, Paul Burton wrote:
> > Hi Christoph,
> >
> > On Wed, May 01, 2019 at 03:13:39PM +0200, Christoph Hellwig wrote:
> > > Stop providing our arch alloc/free hooks and just expose the segment
> > > offset instead.
> > >
> > > Signed-off-by: Christoph Hellwig <hch@lst.de>
> > > ---
> > >  arch/mips/Kconfig              |  1 +
> > >  arch/mips/include/asm/page.h   |  3 ---
> > >  arch/mips/jazz/jazzdma.c       |  6 ------
> > >  arch/mips/mm/dma-noncoherent.c | 26 +++++++++-----------------
> > >  4 files changed, 10 insertions(+), 26 deletions(-)
> >
> > This one looks good to me now, for patches 1 & 5:
> >
> >   Acked-by: Paul Burton <paul.burton@mips.com>
>
> Thanks, I've merged thos into the dma-mapping tree.

I think this is the cause of some kernelci failures in current
linux-next builds:

https://kernelci.org/build/next/branch/master/kernel/next-20190702/

bigsur_defconfig ‐ mips3 warnings — 1 error
cavium_octeon_defconfig ‐ mips3 warnings — 1 error
ip27_defconfig ‐ mips3 warnings — 1 error
loongson3_defconfig ‐ mips3 warnings — 1 error
mips_paravirt_defconfig ‐ mips3 warnings — 1 error
nlm_xlp_defconfig ‐ mips3 warnings — 1 error
nlm_xlr_defconfig ‐ mips1 warning — 1 error

/home/buildslave/workspace/workspace/kernel-build@8/linux/build/../kernel/dma/direct.c:144:
undefined reference to `arch_dma_prep_coherent'
2/home/buildslave/workspace/kernel-build/linux/build/../kernel/dma/direct.c:144:
undefined reference to `arch_dma_prep_coherent'
2(.text+0xafc): undefined reference to `arch_dma_prep_coherent'
1direct.c:(.text+0x934): undefined reference to `arch_dma_prep_coherent'
1(.text+0xb84): undefined reference to `arch_dma_prep_coherent'

I haven't looked into the details, but I suspect all machines
with cache-coherent DMA are broken.

       Arnd

Re: [PATCH 5/7 v2] MIPS: use the generic uncached segment support in dma-direct

[Christoph Hellwig]

On Wed, Jul 03, 2019 at 10:54:05AM +0200, Arnd Bergmann wrote:
> I think this is the cause of some kernelci failures in current
> linux-next builds:

Yes, Guenther reported this already and I sent a fix.  I've been waiting
for an ACK from the mips maintaines, but given the breakage I
might as well just pull it in without an ACK.

[PATCH 6/7] nios2: use the generic uncached segment support in dma-direct

[Christoph Hellwig]

Stop providing our own arch alloc/free hooks and just expose the segment
offset and use the generic dma-direct allocator.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 arch/nios2/Kconfig            |  1 +
 arch/nios2/include/asm/page.h |  6 ------
 arch/nios2/mm/dma-mapping.c   | 34 +++++++++++++++-------------------
 3 files changed, 16 insertions(+), 25 deletions(-)

diff --git a/arch/nios2/Kconfig b/arch/nios2/Kconfig
index 4ef15a61b7bc..c87af5c77d24 100644
--- a/arch/nios2/Kconfig
+++ b/arch/nios2/Kconfig
@@ -4,6 +4,7 @@ config NIOS2
 	select ARCH_32BIT_OFF_T
 	select ARCH_HAS_SYNC_DMA_FOR_CPU
 	select ARCH_HAS_SYNC_DMA_FOR_DEVICE
+	select ARCH_HAS_UNCACHED_SEGMENT
 	select ARCH_NO_SWAP
 	select TIMER_OF
 	select GENERIC_ATOMIC64
diff --git a/arch/nios2/include/asm/page.h b/arch/nios2/include/asm/page.h
index f1fbdc47bdaf..79fcac61f6ef 100644
--- a/arch/nios2/include/asm/page.h
+++ b/arch/nios2/include/asm/page.h
@@ -101,12 +101,6 @@ static inline bool pfn_valid(unsigned long pfn)
 # define VM_DATA_DEFAULT_FLAGS	(VM_READ | VM_WRITE | \
 				 VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC)
 
-# define UNCAC_ADDR(addr)	\
-	((void *)((unsigned)(addr) | CONFIG_NIOS2_IO_REGION_BASE))
-# define CAC_ADDR(addr)		\
-	((void *)(((unsigned)(addr) & ~CONFIG_NIOS2_IO_REGION_BASE) |	\
-		CONFIG_NIOS2_KERNEL_REGION_BASE))
-
 #include <asm-generic/memory_model.h>
 
 #include <asm-generic/getorder.h>
diff --git a/arch/nios2/mm/dma-mapping.c b/arch/nios2/mm/dma-mapping.c
index 4af9e5b5ba1c..9cb238664584 100644
--- a/arch/nios2/mm/dma-mapping.c
+++ b/arch/nios2/mm/dma-mapping.c
@@ -60,32 +60,28 @@ void arch_sync_dma_for_cpu(struct device *dev, phys_addr_t paddr,
 	}
 }
 
-void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
-		gfp_t gfp, unsigned long attrs)
+void arch_dma_prep_coherent(struct page *page, size_t size)
 {
-	void *ret;
+	unsigned long start = (unsigned long)page_address(page);
 
-	/* optimized page clearing */
-	gfp |= __GFP_ZERO;
+	flush_dcache_range(start, start + size);
+}
 
-	if (dev == NULL || (dev->coherent_dma_mask < 0xffffffff))
-		gfp |= GFP_DMA;
+void *uncached_kernel_address(void *ptr)
+{
+	unsigned long addr = (unsigned long)ptr;
 
-	ret = (void *) __get_free_pages(gfp, get_order(size));
-	if (ret != NULL) {
-		*dma_handle = virt_to_phys(ret);
-		flush_dcache_range((unsigned long) ret,
-			(unsigned long) ret + size);
-		ret = UNCAC_ADDR(ret);
-	}
+	addr |= CONFIG_NIOS2_IO_REGION_BASE;
 
-	return ret;
+	return (void *)ptr;
 }
 
-void arch_dma_free(struct device *dev, size_t size, void *vaddr,
-		dma_addr_t dma_handle, unsigned long attrs)
+void *cached_kernel_address(void *ptr)
 {
-	unsigned long addr = (unsigned long) CAC_ADDR((unsigned long) vaddr);
+	unsigned long addr = (unsigned long)ptr;
+
+	addr &= ~CONFIG_NIOS2_IO_REGION_BASE;
+	addr |= CONFIG_NIOS2_KERNEL_REGION_BASE;
 
-	free_pages(addr, get_order(size));
+	return (void *)ptr;
 }
-- 
2.20.1

[PATCH 7/7] microblaze: use the generic uncached segment support in dma-direct

[Christoph Hellwig]

Stop providing our own arch alloc/free hooks for nommu platforms and
just expose the segment offset and use the generic dma-direct
allocator.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 arch/microblaze/Kconfig         |  2 +
 arch/microblaze/mm/consistent.c | 97 +++++++++++++++------------------
 2 files changed, 47 insertions(+), 52 deletions(-)

diff --git a/arch/microblaze/Kconfig b/arch/microblaze/Kconfig
index a51b965b3b82..b0df6b53e2bf 100644
--- a/arch/microblaze/Kconfig
+++ b/arch/microblaze/Kconfig
@@ -3,9 +3,11 @@ config MICROBLAZE
 	select ARCH_32BIT_OFF_T
 	select ARCH_NO_SWAP
 	select ARCH_HAS_DMA_COHERENT_TO_PFN if MMU
+	select ARCH_HAS_DMA_PREP_COHERENT
 	select ARCH_HAS_GCOV_PROFILE_ALL
 	select ARCH_HAS_SYNC_DMA_FOR_CPU
 	select ARCH_HAS_SYNC_DMA_FOR_DEVICE
+	select ARCH_HAS_UNCACHED_SEGMENT if !MMU
 	select ARCH_MIGHT_HAVE_PC_PARPORT
 	select ARCH_NO_COHERENT_DMA_MMAP if !MMU
 	select ARCH_WANT_IPC_PARSE_VERSION
diff --git a/arch/microblaze/mm/consistent.c b/arch/microblaze/mm/consistent.c
index 3002cbca3059..a525194a754e 100644
--- a/arch/microblaze/mm/consistent.c
+++ b/arch/microblaze/mm/consistent.c
@@ -45,21 +45,52 @@
 #include <asm/cpuinfo.h>
 #include <asm/tlbflush.h>
 
-#ifndef CONFIG_MMU
-/* I have to use dcache values because I can't relate on ram size */
-# define UNCACHED_SHADOW_MASK (cpuinfo.dcache_high - cpuinfo.dcache_base + 1)
-#endif
+void arch_dma_prep_coherent(struct page *page, size_t size)
+{
+	phys_addr_t paddr = page_to_phys(page);
+
+	flush_dcache_range(paddr, paddr + size);
+}
 
+#ifndef CONFIG_MMU
 /*
- * Consistent memory allocators. Used for DMA devices that want to
- * share uncached memory with the processor core.
- * My crufty no-MMU approach is simple. In the HW platform we can optionally
- * mirror the DDR up above the processor cacheable region.  So, memory accessed
- * in this mirror region will not be cached.  It's alloced from the same
- * pool as normal memory, but the handle we return is shifted up into the
- * uncached region.  This will no doubt cause big problems if memory allocated
- * here is not also freed properly. -- JW
+ * Consistent memory allocators. Used for DMA devices that want to share
+ * uncached memory with the processor core.  My crufty no-MMU approach is
+ * simple.  In the HW platform we can optionally mirror the DDR up above the
+ * processor cacheable region.  So, memory accessed in this mirror region will
+ * not be cached.  It's alloced from the same pool as normal memory, but the
+ * handle we return is shifted up into the uncached region.  This will no doubt
+ * cause big problems if memory allocated here is not also freed properly. -- JW
+ *
+ * I have to use dcache values because I can't relate on ram size:
  */
+#ifdef CONFIG_XILINX_UNCACHED_SHADOW
+#define UNCACHED_SHADOW_MASK (cpuinfo.dcache_high - cpuinfo.dcache_base + 1)
+#else
+#define UNCACHED_SHADOW_MASK 0
+#endif /* CONFIG_XILINX_UNCACHED_SHADOW */
+
+void *uncached_kernel_address(void *ptr)
+{
+	unsigned long addr = (unsigned long)ptr;
+	
+	addr |= UNCACHED_SHADOW_MASK;
+	if (addr > cpuinfo.dcache_base && addr < cpuinfo.dcache_high)
+		pr_warn("ERROR: Your cache coherent area is CACHED!!!\n");
+	return (void *)addr;
+}
+
+void *cached_kernel_address(void *ptr)
+{
+	unsigned long addr = (unsigned long)ptr;
+
+	addr &= ~UNCACHED_SHADOW_MASK;
+
+	return (void *)addr;
+}
+
+#else /* CONFIG_MMU */
+
 void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
 		gfp_t gfp, unsigned long attrs)
 {
@@ -67,12 +98,9 @@ void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
 	void *ret;
 	unsigned int i, err = 0;
 	struct page *page, *end;
-
-#ifdef CONFIG_MMU
 	phys_addr_t pa;
 	struct vm_struct *area;
 	unsigned long va;
-#endif
 
 	if (in_interrupt())
 		BUG();
@@ -89,26 +117,8 @@ void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
 	 * we need to ensure that there are no cachelines in use,
 	 * or worse dirty in this area.
 	 */
-	flush_dcache_range(virt_to_phys((void *)vaddr),
-					virt_to_phys((void *)vaddr) + size);
-
-#ifndef CONFIG_MMU
-	ret = (void *)vaddr;
-	/*
-	 * Here's the magic!  Note if the uncached shadow is not implemented,
-	 * it's up to the calling code to also test that condition and make
-	 * other arranegments, such as manually flushing the cache and so on.
-	 */
-# ifdef CONFIG_XILINX_UNCACHED_SHADOW
-	ret = (void *)((unsigned) ret | UNCACHED_SHADOW_MASK);
-# endif
-	if ((unsigned int)ret > cpuinfo.dcache_base &&
-				(unsigned int)ret < cpuinfo.dcache_high)
-		pr_warn("ERROR: Your cache coherent area is CACHED!!!\n");
+	arch_dma_prep_coherent(virt_to_page((unsigned long)vaddr), size);
 
-	/* dma_handle is same as physical (shadowed) address */
-	*dma_handle = (dma_addr_t)ret;
-#else
 	/* Allocate some common virtual space to map the new pages. */
 	area = get_vm_area(size, VM_ALLOC);
 	if (!area) {
@@ -120,7 +130,6 @@ void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
 
 	/* This gives us the real physical address of the first page. */
 	*dma_handle = pa = __virt_to_phys(vaddr);
-#endif
 
 	/*
 	 * free wasted pages.  We skip the first page since we know
@@ -134,10 +143,8 @@ void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
 	split_page(page, order);
 
 	for (i = 0; i < size && err == 0; i += PAGE_SIZE) {
-#ifdef CONFIG_MMU
 		/* MS: This is the whole magic - use cache inhibit pages */
 		err = map_page(va + i, pa + i, _PAGE_KERNEL | _PAGE_NO_CACHE);
-#endif
 
 		SetPageReserved(page);
 		page++;
@@ -157,7 +164,6 @@ void *arch_dma_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
 	return ret;
 }
 
-#ifdef CONFIG_MMU
 static pte_t *consistent_virt_to_pte(void *vaddr)
 {
 	unsigned long addr = (unsigned long)vaddr;
@@ -175,7 +181,6 @@ long arch_dma_coherent_to_pfn(struct device *dev, void *vaddr,
 
 	return pte_pfn(*ptep);
 }
-#endif
 
 /*
  * free page(s) as defined by the above mapping.
@@ -190,18 +195,6 @@ void arch_dma_free(struct device *dev, size_t size, void *vaddr,
 
 	size = PAGE_ALIGN(size);
 
-#ifndef CONFIG_MMU
-	/* Clear SHADOW_MASK bit in address, and free as per usual */
-# ifdef CONFIG_XILINX_UNCACHED_SHADOW
-	vaddr = (void *)((unsigned)vaddr & ~UNCACHED_SHADOW_MASK);
-# endif
-	page = virt_to_page(vaddr);
-
-	do {
-		__free_reserved_page(page);
-		page++;
-	} while (size -= PAGE_SIZE);
-#else
 	do {
 		pte_t *ptep = consistent_virt_to_pte(vaddr);
 		unsigned long pfn;
@@ -219,5 +212,5 @@ void arch_dma_free(struct device *dev, size_t size, void *vaddr,
 
 	/* flush tlb */
 	flush_tlb_all();
-#endif
 }
+#endif /* CONFIG_MMU */
-- 
2.20.1