[PATCH] x86: Fix gfp flags for coherent DMA memory allocation

[PATCH] x86: Fix gfp flags for coherent DMA memory allocation

[Junichi Nomura]

Commit 6894258eda2f reversed the order of gfp_flags adjustment in
dma_alloc_attrs() for x86 [arch/x86/kernel/pci-dma.c]
As a result, relevant flags set by dma_alloc_coherent_gfp_flags()
is just discarded and causes coherent DMA memory allocation failure
on some devices.

Fixes: 6894258eda2f ("dma-mapping: consolidate dma_{alloc,free}_{attrs,coherent}")
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>

diff --git a/arch/x86/kernel/pci-dma.c b/arch/x86/kernel/pci-dma.c
index 84b8ef8..1b55de1 100644
--- a/arch/x86/kernel/pci-dma.c
+++ b/arch/x86/kernel/pci-dma.c
@@ -131,8 +131,8 @@ void dma_generic_free_coherent(struct device *dev, size_t size, void *vaddr,
 
 bool arch_dma_alloc_attrs(struct device **dev, gfp_t *gfp)
 {
-	*gfp = dma_alloc_coherent_gfp_flags(*dev, *gfp);
 	*gfp &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
+	*gfp = dma_alloc_coherent_gfp_flags(*dev, *gfp);
 
 	if (!*dev)
 		*dev = &x86_dma_fallback_dev;

Re: [PATCH] x86: Fix gfp flags for coherent DMA memory allocation

[Junichi Nomura]

On 09/14/15 16:38, Junichi Nomura wrote:
> Commit 6894258eda2f reversed the order of gfp_flags adjustment in
> dma_alloc_attrs() for x86 [arch/x86/kernel/pci-dma.c]
> As a result, relevant flags set by dma_alloc_coherent_gfp_flags()
> is just discarded and causes coherent DMA memory allocation failure
> on some devices.
> 
> Fixes: 6894258eda2f ("dma-mapping: consolidate dma_{alloc,free}_{attrs,coherent}")
> Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
> 
> diff --git a/arch/x86/kernel/pci-dma.c b/arch/x86/kernel/pci-dma.c
> index 84b8ef8..1b55de1 100644
> --- a/arch/x86/kernel/pci-dma.c
> +++ b/arch/x86/kernel/pci-dma.c
> @@ -131,8 +131,8 @@ void dma_generic_free_coherent(struct device *dev, size_t size, void *vaddr,
>  
>  bool arch_dma_alloc_attrs(struct device **dev, gfp_t *gfp)
>  {
> -	*gfp = dma_alloc_coherent_gfp_flags(*dev, *gfp);
>  	*gfp &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
> +	*gfp = dma_alloc_coherent_gfp_flags(*dev, *gfp);

Additional note:

I noticed this problem because megaraid driver, which uses 32-bit
dma mask for coherent buffers, failed initialization and my system
didn't boot.

But there can be other drivers which are silently and unnecessarily
falling back to swiotlb due to this problem.

megaraid_sas 0000:10:00.0: FW now in Ready state 
megaraid_sas 0000:10:00.0: firmware supports msix	: (16) 
megaraid_sas 0000:10:00.0: current msix/online cpus	: (16/32) 
megaraid_sas 0000:10:00.0: Init cmd success 
megaraid_sas 0000:10:00.0: firmware type	: Legacy(64 VD) firmware 
megaraid_sas 0000:10:00.0: controller type	: MR(512MB) 
megaraid_sas 0000:10:00.0: swiotlb buffer is full (sz: 665896 bytes) 
swiotlb: coherent allocation failed for device 0000:10:00.0 size=665896 
CPU: 0 PID: 657 Comm: modprobe Tainted: G            E   4.3.0-rc1 #1 
Hardware name: 
 ffff88046e99f098 ffff88046db8f9c8 ffffffff812a7904 ffffffffffffffff 
 00000000ffffffff ffff88046db8fa18 ffffffff812cfe74 ffff88046cd00188 
 0000000800000008 ffff88046dfce8c0 00000000000002d0 ffff88046cd00188 
Call Trace: 
 [<ffffffff812a7904>] dump_stack+0x48/0x64 
 [<ffffffff812cfe74>] swiotlb_alloc_coherent+0x154/0x170 
 [<ffffffff8104a2af>] x86_swiotlb_alloc_coherent+0x5f/0x70 
 [<ffffffffa000f9b5>] megasas_init_adapter_fusion+0x375/0x4b0 [megaraid_sas] 
 [<ffffffffa0005a06>] megasas_init_fw+0x2e6/0x940 [megaraid_sas] 
 [<ffffffffa0007795>] megasas_probe_one+0x745/0xb60 [megaraid_sas] 
 [<ffffffff812e913c>] local_pci_probe+0x4c/0xa0 
 [<ffffffff812e9213>] pci_call_probe+0x83/0xa0 
 [<ffffffff812e9744>] ? pci_match_device+0xe4/0x110 
 [<ffffffff812e982f>] pci_device_probe+0x7f/0xb0 
 [<ffffffff813ce2f3>] really_probe+0x133/0x390 
 [<ffffffff813ce5a9>] driver_probe_device+0x59/0xd0 
 [<ffffffff813ce6bd>] __driver_attach+0x9d/0xa0 
 [<ffffffff813ce620>] ? driver_probe_device+0xd0/0xd0 
 [<ffffffff813ce620>] ? driver_probe_device+0xd0/0xd0 
 [<ffffffff813cc2fa>] bus_for_each_dev+0x8a/0xb0 
 [<ffffffff813cdd6e>] driver_attach+0x1e/0x20 
 [<ffffffff813cd7cf>] bus_add_driver+0x1cf/0x290 
 [<ffffffff813cee50>] driver_register+0x60/0xe0 
 [<ffffffff812e952c>] __pci_register_driver+0x4c/0x50 
 [<ffffffffa001d0a4>] megasas_init+0xa4/0x1ba [megaraid_sas] 
 [<ffffffffa001d000>] ? 0xffffffffa001d000 
 [<ffffffff8100027d>] do_one_initcall+0xad/0x1c0 
 [<ffffffff810dba39>] do_init_module+0x69/0x210 
 [<ffffffff810de19a>] load_module+0x5aa/0x730 
 [<ffffffff810dad60>] ? mod_sysfs_teardown+0x140/0x140 
 [<ffffffff810da790>] ? module_sect_show+0x30/0x30 
 [<ffffffff810de4a2>] SyS_init_module+0x92/0xc0 
 [<ffffffff815ce697>] entry_SYSCALL_64_fastpath+0x12/0x6a 
megaraid_sas 0000:10:00.0: Could not allocate memory for map info 

-- 
Jun'ichi Nomura, NEC Corporation

Re: [PATCH] x86: Fix gfp flags for coherent DMA memory allocation

[Tony Luck]

On Mon, Sep 14, 2015 at 12:38 AM, Junichi Nomura <j-nomura@ce.jp.nec.com> wrote:
> Commit 6894258eda2f reversed the order of gfp_flags adjustment in
> dma_alloc_attrs() for x86 [arch/x86/kernel/pci-dma.c]
> As a result, relevant flags set by dma_alloc_coherent_gfp_flags()
> is just discarded and causes coherent DMA memory allocation failure
> on some devices.
>
> Fixes: 6894258eda2f ("dma-mapping: consolidate dma_{alloc,free}_{attrs,coherent}")
> Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>

Tested-by: Tony Luck <tony.luck@intel.com>

Is this en-route to Linus? I can't boot 4.3-rc1 without this patch.

-Tony

Re: [PATCH] x86: Fix gfp flags for coherent DMA memory allocation

[Christoph Hellwig]

On Mon, Sep 14, 2015 at 07:38:36AM +0000, Junichi Nomura wrote:
> Commit 6894258eda2f reversed the order of gfp_flags adjustment in
> dma_alloc_attrs() for x86 [arch/x86/kernel/pci-dma.c]
> As a result, relevant flags set by dma_alloc_coherent_gfp_flags()
> is just discarded and causes coherent DMA memory allocation failure
> on some devices.

Oops, sorry.  Thanks for fixing this:

Acked-by: Christoph Hellwig <hch@lst.de>

[tip:x86/urgent] x86/pci/dma: Fix gfp flags for coherent DMA memory allocation

[tip-bot for Junichi Nomura]

Commit-ID:  590f07874e8e3c83729b919312c65aea2533c8cf
Gitweb:     http://git.kernel.org/tip/590f07874e8e3c83729b919312c65aea2533c8cf
Author:     Junichi Nomura <j-nomura@ce.jp.nec.com>
AuthorDate: Mon, 14 Sep 2015 07:38:36 +0000
Committer:  Thomas Gleixner <tglx@linutronix.de>
CommitDate: Thu, 17 Sep 2015 16:22:11 +0200

x86/pci/dma: Fix gfp flags for coherent DMA memory allocation

Commit 6894258eda2f reversed the order of gfp_flags adjustment in
dma_alloc_attrs() for x86 [arch/x86/kernel/pci-dma.c] As a result,
relevant flags set by dma_alloc_coherent_gfp_flags() are just
discarded and cause coherent DMA memory allocation failure on some
devices.

Fixes: 6894258eda2f ("dma-mapping: consolidate dma_{alloc,free}_{attrs,coherent}")
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Tested-by: Tony Luck <tony.luck@intel.com>
Acked-by: Christoph Hellwig <hch@lst.de>
Link: http://lkml.kernel.org/r/20150914073834.GA13077@xzibit.linux.bs1.fc.nec.co.jp
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>

---
 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 84b8ef8..1b55de1 100644
--- a/arch/x86/kernel/pci-dma.c
+++ b/arch/x86/kernel/pci-dma.c
@@ -131,8 +131,8 @@ void dma_generic_free_coherent(struct device *dev, size_t size, void *vaddr,
 
 bool arch_dma_alloc_attrs(struct device **dev, gfp_t *gfp)
 {
-	*gfp = dma_alloc_coherent_gfp_flags(*dev, *gfp);
 	*gfp &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
+	*gfp = dma_alloc_coherent_gfp_flags(*dev, *gfp);
 
 	if (!*dev)
 		*dev = &x86_dma_fallback_dev;