[PATCH v4 1/2] ARM: dma-mapping: Don't use outer_flush_range when the L2C is coherent

[PATCH v4 1/2] ARM: dma-mapping: Don't use outer_flush_range when the L2C is coherent

[Gregory CLEMENT]

When a L2 cache controller is used in a system that provides hardware
coherency, the entire outer cache operations are useless, and can be
skipped.  Moreover, on some systems, it is harmful as it causes
deadlocks between the Marvell coherency mechanism, the Marvell PCIe
controller and the Cortex-A9.

In the current kernel implementation, the outer cache flush range
operation is triggered by the dma_alloc function.
This operation can be take place during runtime and in some
circumstances may lead to the PCIe/PL310 deadlock on Armada 375/38x
SoCs.

This patch extends the __dma_clear_buffer() function to receive a
boolean argument related to the coherency of the system. The same
things is done for the calling functions.

Reported-by: Nadav Haklai <nadavh@marvell.com>
Signed-off-by: Gregory CLEMENT <gregory.clement@free-electrons.com>
Cc: <stable@vger.kernel.org> # v3.16+
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
---
 arch/arm/mm/dma-mapping.c | 56 +++++++++++++++++++++++++++++++----------------
 1 file changed, 37 insertions(+), 19 deletions(-)

diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
index e62604384945..c8ed0c16c5a8 100644
--- a/arch/arm/mm/dma-mapping.c
+++ b/arch/arm/mm/dma-mapping.c
@@ -225,7 +225,7 @@ static u64 get_coherent_dma_mask(struct device *dev)
 	return mask;
 }
 
-static void __dma_clear_buffer(struct page *page, size_t size)
+static void __dma_clear_buffer(struct page *page, size_t size, bool is_coherent)
 {
 	/*
 	 * Ensure that the allocated pages are zeroed, and that any data
@@ -237,17 +237,21 @@ static void __dma_clear_buffer(struct page *page, size_t size)
 		while (size > 0) {
 			void *ptr = kmap_atomic(page);
 			memset(ptr, 0, PAGE_SIZE);
-			dmac_flush_range(ptr, ptr + PAGE_SIZE);
+			if (!is_coherent)
+				dmac_flush_range(ptr, ptr + PAGE_SIZE);
 			kunmap_atomic(ptr);
 			page++;
 			size -= PAGE_SIZE;
 		}
-		outer_flush_range(base, end);
+		if (!is_coherent)
+			outer_flush_range(base, end);
 	} else {
 		void *ptr = page_address(page);
 		memset(ptr, 0, size);
-		dmac_flush_range(ptr, ptr + size);
-		outer_flush_range(__pa(ptr), __pa(ptr) + size);
+		if (!is_coherent) {
+			dmac_flush_range(ptr, ptr + size);
+			outer_flush_range(__pa(ptr), __pa(ptr) + size);
+		}
 	}
 }
 
@@ -255,7 +259,8 @@ static void __dma_clear_buffer(struct page *page, size_t size)
  * Allocate a DMA buffer for 'dev' of size 'size' using the
  * specified gfp mask.  Note that 'size' must be page aligned.
  */
-static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gfp)
+static struct page *__dma_alloc_buffer(struct device *dev, size_t size,
+				gfp_t gfp, bool is_coherent)
 {
 	unsigned long order = get_order(size);
 	struct page *page, *p, *e;
@@ -271,7 +276,7 @@ static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gf
 	for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
 		__free_page(p);
 
-	__dma_clear_buffer(page, size);
+	__dma_clear_buffer(page, size, is_coherent);
 
 	return page;
 }
@@ -293,7 +298,8 @@ static void __dma_free_buffer(struct page *page, size_t size)
 
 static void *__alloc_from_contiguous(struct device *dev, size_t size,
 				     pgprot_t prot, struct page **ret_page,
-				     const void *caller, bool want_vaddr);
+				     const void *caller, bool want_vaddr,
+				     bool is_coherent);
 
 static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
 				 pgprot_t prot, struct page **ret_page,
@@ -359,9 +365,13 @@ static int __init atomic_pool_init(void)
 	if (!atomic_pool)
 		goto out;
 
+	/*
+	 * The atomic pool is only used for non-coherent allocations
+	 * so we must pass false for is_coherent.
+	 */
 	if (dev_get_cma_area(NULL))
 		ptr = __alloc_from_contiguous(NULL, atomic_pool_size, prot,
-					      &page, atomic_pool_init, true);
+					&page, atomic_pool_init, true, false);
 	else
 		ptr = __alloc_remap_buffer(NULL, atomic_pool_size, gfp, prot,
 					   &page, atomic_pool_init, true);
@@ -475,7 +485,11 @@ static void *__alloc_remap_buffer(struct device *dev, size_t size, gfp_t gfp,
 {
 	struct page *page;
 	void *ptr = NULL;
-	page = __dma_alloc_buffer(dev, size, gfp);
+	/*
+	 * __alloc_remap_buffer is only called when the device is
+	 * non-coherent
+	 */
+	page = __dma_alloc_buffer(dev, size, gfp, false);
 	if (!page)
 		return NULL;
 	if (!want_vaddr)
@@ -530,7 +544,8 @@ static int __free_from_pool(void *start, size_t size)
 
 static void *__alloc_from_contiguous(struct device *dev, size_t size,
 				     pgprot_t prot, struct page **ret_page,
-				     const void *caller, bool want_vaddr)
+				     const void *caller, bool want_vaddr,
+				     bool is_coherent)
 {
 	unsigned long order = get_order(size);
 	size_t count = size >> PAGE_SHIFT;
@@ -541,7 +556,7 @@ static void *__alloc_from_contiguous(struct device *dev, size_t size,
 	if (!page)
 		return NULL;
 
-	__dma_clear_buffer(page, size);
+	__dma_clear_buffer(page, size, is_coherent);
 
 	if (!want_vaddr)
 		goto out;
@@ -591,7 +606,7 @@ static inline pgprot_t __get_dma_pgprot(struct dma_attrs *attrs, pgprot_t prot)
 #define __get_dma_pgprot(attrs, prot)				__pgprot(0)
 #define __alloc_remap_buffer(dev, size, gfp, prot, ret, c, wv)	NULL
 #define __alloc_from_pool(size, ret_page)			NULL
-#define __alloc_from_contiguous(dev, size, prot, ret, c, wv)	NULL
+#define __alloc_from_contiguous(dev, size, prot, ret, c, wv, is_coherent)   NULL
 #define __free_from_pool(cpu_addr, size)			0
 #define __free_from_contiguous(dev, page, cpu_addr, size, wv)	do { } while (0)
 #define __dma_free_remap(cpu_addr, size)			do { } while (0)
@@ -602,7 +617,8 @@ static void *__alloc_simple_buffer(struct device *dev, size_t size, gfp_t gfp,
 				   struct page **ret_page)
 {
 	struct page *page;
-	page = __dma_alloc_buffer(dev, size, gfp);
+	/* __alloc_simple_buffer is only called when the device is coherent */
+	page = __dma_alloc_buffer(dev, size, gfp, true);
 	if (!page)
 		return NULL;
 
@@ -653,7 +669,7 @@ static void *__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
 		addr = __alloc_simple_buffer(dev, size, gfp, &page);
 	else if (dev_get_cma_area(dev) && (gfp & __GFP_WAIT))
 		addr = __alloc_from_contiguous(dev, size, prot, &page,
-					       caller, want_vaddr);
+					caller, want_vaddr, is_coherent);
 	else if (is_coherent)
 		addr = __alloc_simple_buffer(dev, size, gfp, &page);
 	else if (!(gfp & __GFP_WAIT))
@@ -1123,7 +1139,8 @@ static inline void __free_iova(struct dma_iommu_mapping *mapping,
 }
 
 static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
-					  gfp_t gfp, struct dma_attrs *attrs)
+					gfp_t gfp, struct dma_attrs *attrs,
+					bool is_coherent)
 {
 	struct page **pages;
 	int count = size >> PAGE_SHIFT;
@@ -1146,7 +1163,7 @@ static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
 		if (!page)
 			goto error;
 
-		__dma_clear_buffer(page, size);
+		__dma_clear_buffer(page, size, is_coherent);
 
 		for (i = 0; i < count; i++)
 			pages[i] = page + i;
@@ -1190,7 +1207,7 @@ static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
 				pages[i + j] = pages[i] + j;
 		}
 
-		__dma_clear_buffer(pages[i], PAGE_SIZE << order);
+		__dma_clear_buffer(pages[i], PAGE_SIZE << order, is_coherent);
 		i += 1 << order;
 		count -= 1 << order;
 	}
@@ -1373,7 +1390,8 @@ static void *arm_iommu_alloc_attrs(struct device *dev, size_t size,
 	 */
 	gfp &= ~(__GFP_COMP);
 
-	pages = __iommu_alloc_buffer(dev, size, gfp, attrs);
+	/* For now always consider we are in a non-coherent case */
+	pages = __iommu_alloc_buffer(dev, size, gfp, attrs, false);
 	if (!pages)
 		return NULL;
 
-- 
2.1.0

Re: [PATCH v4 1/2] ARM: dma-mapping: Don't use outer_flush_range when the L2C is coherent

[Russell King - ARM Linux]

On Tue, Sep 29, 2015 at 06:50:57PM +0200, Gregory CLEMENT wrote:
> When a L2 cache controller is used in a system that provides hardware

You're talking about L2 cache here, but you're also masking out the L1
cache maintanence (dmac_*) too.  It's my understanding that we don't
yet support coherency to L1 yet.

> In the current kernel implementation, the outer cache flush range
> operation is triggered by the dma_alloc function.
> This operation can be take place during runtime and in some
> circumstances may lead to the PCIe/PL310 deadlock on Armada 375/38x
> SoCs.

I wonder if that's what's causing the sporadic lockups I'm seeing on
38x with a SATA PCIe card - it happens at a very specific point during
boot while initialising the SATA card, right down to the kernel message
character that it stops at.

-- 
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.

Re: [PATCH v4 1/2] ARM: dma-mapping: Don't use outer_flush_range when the L2C is coherent

[Thomas Petazzoni]

Hello Russell,

On Tue, 29 Sep 2015 18:10:54 +0100, Russell King - ARM Linux wrote:

> > In the current kernel implementation, the outer cache flush range
> > operation is triggered by the dma_alloc function.
> > This operation can be take place during runtime and in some
> > circumstances may lead to the PCIe/PL310 deadlock on Armada 375/38x
> > SoCs.
> 
> I wonder if that's what's causing the sporadic lockups I'm seeing on
> 38x with a SATA PCIe card - it happens at a very specific point during
> boot while initialising the SATA card, right down to the kernel message
> character that it stops at.

It might very well be the case. If there's enough PCIe traffic and a
PL310 cache maintenance operation happening at the same time, the
system will lockup. I'm a bit surprised that just the initialization of
the PCIe card generates enough traffic to trigger the deadlock, but
maybe I'm underestimating the problem.

Best regards,

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com

Re: [PATCH v4 1/2] ARM: dma-mapping: Don't use outer_flush_range when the L2C is coherent

[Russell King - ARM Linux]

On Tue, Sep 29, 2015 at 07:30:31PM +0200, Thomas Petazzoni wrote:
> Hello Russell,
> 
> On Tue, 29 Sep 2015 18:10:54 +0100, Russell King - ARM Linux wrote:
> 
> > > In the current kernel implementation, the outer cache flush range
> > > operation is triggered by the dma_alloc function.
> > > This operation can be take place during runtime and in some
> > > circumstances may lead to the PCIe/PL310 deadlock on Armada 375/38x
> > > SoCs.
> > 
> > I wonder if that's what's causing the sporadic lockups I'm seeing on
> > 38x with a SATA PCIe card - it happens at a very specific point during
> > boot while initialising the SATA card, right down to the kernel message
> > character that it stops at.
> 
> It might very well be the case. If there's enough PCIe traffic and a
> PL310 cache maintenance operation happening at the same time, the
> system will lockup. I'm a bit surprised that just the initialization of
> the PCIe card generates enough traffic to trigger the deadlock, but
> maybe I'm underestimating the problem.

It isn't every boot - the board has booted around 240 kernels so far and
maybe 5% of them have needed the reset button pressed because of this.
It only happens when I have the SATA card connected.  I don't have any
drives on the SATA card at the moment though.

-- 
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.

Re: [PATCH v4 1/2] ARM: dma-mapping: Don't use outer_flush_range when the L2C is coherent

[Thomas Petazzoni]

Hello,

On Tue, 29 Sep 2015 18:48:35 +0100, Russell King - ARM Linux wrote:

> > It might very well be the case. If there's enough PCIe traffic and a
> > PL310 cache maintenance operation happening at the same time, the
> > system will lockup. I'm a bit surprised that just the initialization of
> > the PCIe card generates enough traffic to trigger the deadlock, but
> > maybe I'm underestimating the problem.
> 
> It isn't every boot - the board has booted around 240 kernels so far and
> maybe 5% of them have needed the reset button pressed because of this.
> It only happens when I have the SATA card connected.  I don't have any
> drives on the SATA card at the moment though.

Hum, ok. Then I'm not sure it's the same problem. I believe it's
unlikely that the few PCIe accesses used just to enumerate the PCIe
device and initialize it are enough to cause the deadlock. But I don't
have (and anyway wouldn't fully understand) all the details about this
issue, so I can't say for sure.

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com

Re: [PATCH v4 1/2] ARM: dma-mapping: Don't use outer_flush_range when the L2C is coherent

[Gregory CLEMENT]

Hi Russell,

 On mar., sept. 29 2015, Russell King - ARM Linux <linux@arm.linux.org.uk> wrote:

> On Tue, Sep 29, 2015 at 06:50:57PM +0200, Gregory CLEMENT wrote:
>> When a L2 cache controller is used in a system that provides hardware
>
> You're talking about L2 cache here, but you're also masking out the L1
> cache maintanence (dmac_*) too.  It's my understanding that we don't
> yet support coherency to L1 yet.

Do you suggest to not masking the dmac_* operation ?

Initially I tought that L1 and L2 cache maintanence operation were more
or less linked but I might mix up.

Thanks,

Gregory

-- 
Gregory Clement, Free Electrons
Kernel, drivers, real-time and embedded Linux
development, consulting, training and support.
http://free-electrons.com