[PATCH] x86: dma-mapping: fix GFP_ATOMIC macro usage

GFP_ATOMIC is not a single gfp flag, but a macro which expands to the other
flags and LACK of __GFP_WAIT flag. To check if caller wanted to perform an
atomic allocation, the code must test __GFP_WAIT flag presence. This patch
fixes the issue introduced in v3.5-rc1

CC: stable@vger.kernel.org
Signed-off-by: Marek Szyprowski <m.szyprowski@samsung.com>
---
 arch/x86/kernel/pci-dma.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/x86/kernel/pci-dma.c b/arch/x86/kernel/pci-dma.c
index 872079a..32a81c9 100644
--- a/arch/x86/kernel/pci-dma.c
+++ b/arch/x86/kernel/pci-dma.c

@@ -100,7 +100,7 @@ void *dma_generic_alloc_coherent(struct device *dev, size_t size,
 	flag |= __GFP_ZERO;
 again:
 	page = NULL;
-	if (!(flag & GFP_ATOMIC))
+	if (flag & __GFP_WAIT)
 		page = dma_alloc_from_contiguous(dev, count, get_order(size));
 	if (!page)
 		page = alloc_pages_node(dev_to_node(dev), flag, get_order(size));

-- 
1.7.9.5

On Fri, Jan 17, 2014 at 8:46 AM, Marek Szyprowski
[off-list ref] wrote:

From that description should this not actually be:

    if (!(flag & (GFP_ATOMIC|__GFP_WAIT) == GFP_ATOMIC))

Else we will start using this pool for more than __GFP_HIGH allocations?

That said, it is possible this is right and the intent was to allow
__GFP_HIGH allocations (in general) to use this contigious pool, but I
will let someone more intimate with the code comment to that.  I would
have hoped the code would have been as below in that case:

    if (!(flag & __GFP_HIGH))

Either way once this is resolved a nice comment should be added to
make it really clear:

-apw

Hello,

On 2014-01-17 11:49, Andy Whitcroft wrote:

On Fri, Jan 17, 2014 at 8:46 AM, Marek Szyprowski
[off-list ref] wrote:

quoted

GFP_ATOMIC is not a single gfp flag, but a macro which expands to the other
flags and LACK of __GFP_WAIT flag. To check if caller wanted to perform an
atomic allocation, the code must test __GFP_WAIT flag presence. This patch
fixes the issue introduced in v3.5-rc1

CC: stable@vger.kernel.org
Signed-off-by: Marek Szyprowski <m.szyprowski@samsung.com>
---
 arch/x86/kernel/pci-dma.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/x86/kernel/pci-dma.c b/arch/x86/kernel/pci-dma.c
index 872079a..32a81c9 100644
--- a/arch/x86/kernel/pci-dma.c
+++ b/arch/x86/kernel/pci-dma.c

@@ -100,7 +100,7 @@ void *dma_generic_alloc_coherent(struct device *dev, size_t size,
        flag |= __GFP_ZERO;
 again:
        page = NULL;
-       if (!(flag & GFP_ATOMIC))
+       if (flag & __GFP_WAIT)

quoted

From that description should this not actually be:

     if (!(flag & (GFP_ATOMIC|__GFP_WAIT) == GFP_ATOMIC))

Else we will start using this pool for more than __GFP_HIGH allocations?

That said, it is possible this is right and the intent was to allow
__GFP_HIGH allocations (in general) to use this contigious pool, but I
will let someone more intimate with the code comment to that.  I would
have hoped the code would have been as below in that case:

     if (!(flag & __GFP_HIGH))

Either way once this is resolved a nice comment should be added to
make it really clear:

Exactly in this case, the GFP_ATOMIC check was (incorrectly) added by me in
commit 0a2b9a6ea936 ("X86: integrate CMA with DMA-mapping subsystem"). My
intention was to use CMA only if caller use other allocation flags than
GFP_ATOMIC, because CMA cannot be used from atomic context. The pool is not
aimed for __GFP_HIGH alocations. I will add additional comment to make clear
why __GFP_WAIT flag is being checked.

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland

GFP_ATOMIC is not a single gfp flag, but a macro which expands to the other
flags, where meaningful is the LACK of __GFP_WAIT flag. To check if caller
wants to perform anatomic allocation, the code must test for a lack of the
__GFP_WAIT flag. This patch fixes the issue introduced in v3.5-rc1.

CC: stable@vger.kernel.org
Signed-off-by: Marek Szyprowski <m.szyprowski@samsung.com>
---
 arch/x86/kernel/pci-dma.c |    4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/arch/x86/kernel/pci-dma.c b/arch/x86/kernel/pci-dma.c
index 872079a..3519a78 100644
--- a/arch/x86/kernel/pci-dma.c
+++ b/arch/x86/kernel/pci-dma.c

@@ -100,8 +100,10 @@ void *dma_generic_alloc_coherent(struct device *dev, size_t size,
 	flag |= __GFP_ZERO;
 again:
 	page = NULL;
-	if (!(flag & GFP_ATOMIC))
+	/* CMA can be used only in the context which allows sleeping */
+	if (flag & __GFP_WAIT)
 		page = dma_alloc_from_contiguous(dev, count, get_order(size));
+	/* fallback */
 	if (!page)
 		page = alloc_pages_node(dev_to_node(dev), flag, get_order(size));
 	if (!page)

-- 
1.7.9.5