[alsa-devel] [PATCH]: x86_64, audio: Fix system panic when DMA > 40 bits
The audio driver mistakenly allows 64 bit addresses to be created for the audio driver on Nvidia GPUs. Unfortunately, the hardware normally only supports up to 40 bits of DMA. This can cause system panics as well as misdirected data when the address is > 40 bits as the upper part the address is truncated.
Signed-off-by: Mike Travis travis@sgi.com Reviewed-by: Mike Habeck habeck@sgi.com --- sound/pci/hda/hda_intel.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/sound/pci/hda/hda_intel.c b/sound/pci/hda/hda_intel.c index bcd40ee..45eb165 100644 --- a/sound/pci/hda/hda_intel.c +++ b/sound/pci/hda/hda_intel.c @@ -615,7 +615,7 @@ enum { /* quirks for Nvidia */ #define AZX_DCAPS_PRESET_NVIDIA \ (AZX_DCAPS_NVIDIA_SNOOP | AZX_DCAPS_RIRB_DELAY | AZX_DCAPS_NO_MSI |\ - AZX_DCAPS_ALIGN_BUFSIZE) + AZX_DCAPS_ALIGN_BUFSIZE | AZX_DCAPS_NO_64BIT)
#define AZX_DCAPS_PRESET_CTHDA \ (AZX_DCAPS_NO_MSI | AZX_DCAPS_POSFIX_LPIB | AZX_DCAPS_4K_BDLE_BOUNDARY)
On 05/01/2013 09:04 PM, Mike Travis wrote:
The audio driver mistakenly allows 64 bit addresses to be created for the audio driver on Nvidia GPUs. Unfortunately, the hardware normally only supports up to 40 bits of DMA. This can cause system panics as well as misdirected data when the address is > 40 bits as the upper part the address is truncated.
Thanks for this patch. Stephen Warren, is this something you can confirm/deny, and do you know what range of hardware this actually applies to?
Signed-off-by: Mike Travis travis@sgi.com Reviewed-by: Mike Habeck habeck@sgi.com
sound/pci/hda/hda_intel.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/sound/pci/hda/hda_intel.c b/sound/pci/hda/hda_intel.c index bcd40ee..45eb165 100644 --- a/sound/pci/hda/hda_intel.c +++ b/sound/pci/hda/hda_intel.c @@ -615,7 +615,7 @@ enum { /* quirks for Nvidia */ #define AZX_DCAPS_PRESET_NVIDIA \ (AZX_DCAPS_NVIDIA_SNOOP | AZX_DCAPS_RIRB_DELAY | AZX_DCAPS_NO_MSI |\
AZX_DCAPS_ALIGN_BUFSIZE)
AZX_DCAPS_ALIGN_BUFSIZE | AZX_DCAPS_NO_64BIT)#define AZX_DCAPS_PRESET_CTHDA \ (AZX_DCAPS_NO_MSI | AZX_DCAPS_POSFIX_LPIB | AZX_DCAPS_4K_BDLE_BOUNDARY)
On 5/1/2013 11:20 PM, David Henningsson wrote:
On 05/01/2013 09:04 PM, Mike Travis wrote:
The audio driver mistakenly allows 64 bit addresses to be created for the audio driver on Nvidia GPUs. Unfortunately, the hardware normally only supports up to 40 bits of DMA. This can cause system panics as well as misdirected data when the address is > 40 bits as the upper part the address is truncated.
Thanks for this patch. Stephen Warren, is this something you can confirm/deny, and do you know what range of hardware this actually applies to?
The dma_mask can be read from the gpu device's pci entry. I know it's a stretch to try to link to that but past history has shown that it ranges from 36 to 40 bits.
Signed-off-by: Mike Travis travis@sgi.com Reviewed-by: Mike Habeck habeck@sgi.com
sound/pci/hda/hda_intel.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/sound/pci/hda/hda_intel.c b/sound/pci/hda/hda_intel.c index bcd40ee..45eb165 100644 --- a/sound/pci/hda/hda_intel.c +++ b/sound/pci/hda/hda_intel.c @@ -615,7 +615,7 @@ enum { /* quirks for Nvidia */ #define AZX_DCAPS_PRESET_NVIDIA \ (AZX_DCAPS_NVIDIA_SNOOP | AZX_DCAPS_RIRB_DELAY | AZX_DCAPS_NO_MSI |\
AZX_DCAPS_ALIGN_BUFSIZE)
AZX_DCAPS_ALIGN_BUFSIZE | AZX_DCAPS_NO_64BIT)#define AZX_DCAPS_PRESET_CTHDA \ (AZX_DCAPS_NO_MSI | AZX_DCAPS_POSFIX_LPIB |
AZX_DCAPS_4K_BDLE_BOUNDARY)
On 05/02/2013 08:36 AM, Mike Travis wrote:
On 5/1/2013 11:20 PM, David Henningsson wrote:
On 05/01/2013 09:04 PM, Mike Travis wrote:
The audio driver mistakenly allows 64 bit addresses to be created for the audio driver on Nvidia GPUs. Unfortunately, the hardware normally only supports up to 40 bits of DMA. This can cause system panics as well as misdirected data when the address is > 40 bits as the upper part the address is truncated.
Thanks for this patch. Stephen Warren, is this something you can confirm/deny, and do you know what range of hardware this actually applies to?
The dma_mask can be read from the gpu device's pci entry. I know it's a stretch to try to link to that but past history has shown that it ranges from 36 to 40 bits.
Ok, so the below patch is some kind of workaround, as we don't currently take the dma_mask into account correctly? Or put in another way, would it be possible/better to actually read the dma_mask and use that instead of just clamping at 32 bit?
Signed-off-by: Mike Travis travis@sgi.com Reviewed-by: Mike Habeck habeck@sgi.com
sound/pci/hda/hda_intel.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/sound/pci/hda/hda_intel.c b/sound/pci/hda/hda_intel.c index bcd40ee..45eb165 100644 --- a/sound/pci/hda/hda_intel.c +++ b/sound/pci/hda/hda_intel.c @@ -615,7 +615,7 @@ enum { /* quirks for Nvidia */ #define AZX_DCAPS_PRESET_NVIDIA \ (AZX_DCAPS_NVIDIA_SNOOP | AZX_DCAPS_RIRB_DELAY | AZX_DCAPS_NO_MSI |\
AZX_DCAPS_ALIGN_BUFSIZE)
AZX_DCAPS_ALIGN_BUFSIZE | AZX_DCAPS_NO_64BIT)#define AZX_DCAPS_PRESET_CTHDA \ (AZX_DCAPS_NO_MSI | AZX_DCAPS_POSFIX_LPIB |
AZX_DCAPS_4K_BDLE_BOUNDARY)
On 5/2/2013 1:29 AM, David Henningsson wrote:
On 05/02/2013 08:36 AM, Mike Travis wrote:
On 5/1/2013 11:20 PM, David Henningsson wrote:
On 05/01/2013 09:04 PM, Mike Travis wrote:
The audio driver mistakenly allows 64 bit addresses to be created for the audio driver on Nvidia GPUs. Unfortunately, the hardware normally only supports up to 40 bits of DMA. This can cause system panics as well as misdirected data when the address is > 40 bits as the upper part the address is truncated.
Thanks for this patch. Stephen Warren, is this something you can confirm/deny, and do you know what range of hardware this actually applies to?
The dma_mask can be read from the gpu device's pci entry. I know it's a stretch to try to link to that but past history has shown that it ranges from 36 to 40 bits.
Ok, so the below patch is some kind of workaround, as we don't currently take the dma_mask into account correctly? Or put in another way, would it be possible/better to actually read the dma_mask and use that instead of just clamping at 32 bit?\
Yes, it would. There is a strange scenario where if you request a DMA32 address it must be on the lowest node (where the DMA32 area is). But if you are running on a cpuset that excludes Node 0's memory, you'll get an error. But I couldn't think of a workable way to link the two (except that the PCI id is the same except for the function #.)
Signed-off-by: Mike Travis travis@sgi.com Reviewed-by: Mike Habeck habeck@sgi.com
sound/pci/hda/hda_intel.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/sound/pci/hda/hda_intel.c b/sound/pci/hda/hda_intel.c index bcd40ee..45eb165 100644 --- a/sound/pci/hda/hda_intel.c +++ b/sound/pci/hda/hda_intel.c @@ -615,7 +615,7 @@ enum { /* quirks for Nvidia */ #define AZX_DCAPS_PRESET_NVIDIA \ (AZX_DCAPS_NVIDIA_SNOOP | AZX_DCAPS_RIRB_DELAY | AZX_DCAPS_NO_MSI |\
AZX_DCAPS_ALIGN_BUFSIZE)
AZX_DCAPS_ALIGN_BUFSIZE | AZX_DCAPS_NO_64BIT)#define AZX_DCAPS_PRESET_CTHDA \ (AZX_DCAPS_NO_MSI | AZX_DCAPS_POSFIX_LPIB |
AZX_DCAPS_4K_BDLE_BOUNDARY)
At Wed, 01 May 2013 23:36:24 -0700, Mike Travis wrote:
On 5/1/2013 11:20 PM, David Henningsson wrote:
On 05/01/2013 09:04 PM, Mike Travis wrote:
The audio driver mistakenly allows 64 bit addresses to be created for the audio driver on Nvidia GPUs. Unfortunately, the hardware normally only supports up to 40 bits of DMA. This can cause system panics as well as misdirected data when the address is > 40 bits as the upper part the address is truncated.
Thanks for this patch. Stephen Warren, is this something you can confirm/deny, and do you know what range of hardware this actually applies to?
The dma_mask can be read from the gpu device's pci entry.
Well, this assumes that the graphics driver is initialized before the audio driver, right?
I know it's a stretch to try to link to that but past history has shown that it ranges from 36 to 40 bits.
Looking through nouveau drivers, even 32bit dma_bits is used for some devices, too, so a safe bid would be simply 32bit. Of course, if we have a table of DMA bits, we can use the more exact values.
I'll apply your original patch as a quick fix for now.
thanks,
Takashi
Signed-off-by: Mike Travis travis@sgi.com Reviewed-by: Mike Habeck habeck@sgi.com
sound/pci/hda/hda_intel.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/sound/pci/hda/hda_intel.c b/sound/pci/hda/hda_intel.c index bcd40ee..45eb165 100644 --- a/sound/pci/hda/hda_intel.c +++ b/sound/pci/hda/hda_intel.c @@ -615,7 +615,7 @@ enum { /* quirks for Nvidia */ #define AZX_DCAPS_PRESET_NVIDIA \ (AZX_DCAPS_NVIDIA_SNOOP | AZX_DCAPS_RIRB_DELAY | AZX_DCAPS_NO_MSI |\
AZX_DCAPS_ALIGN_BUFSIZE)
AZX_DCAPS_ALIGN_BUFSIZE | AZX_DCAPS_NO_64BIT)#define AZX_DCAPS_PRESET_CTHDA \ (AZX_DCAPS_NO_MSI | AZX_DCAPS_POSFIX_LPIB |
AZX_DCAPS_4K_BDLE_BOUNDARY)
On 05/02/2013 12:20 AM, David Henningsson wrote:
On 05/01/2013 09:04 PM, Mike Travis wrote:
The audio driver mistakenly allows 64 bit addresses to be created for the audio driver on Nvidia GPUs. Unfortunately, the hardware normally only supports up to 40 bits of DMA. This can cause system panics as well as misdirected data when the address is > 40 bits as the upper part the address is truncated.
Thanks for this patch. Stephen Warren, is this something you can confirm/deny, and do you know what range of hardware this actually applies to?
It's certainly plausible that the GPU's internal physical addresses are less than a full 64-bits, and this probably affects PCI accesses too.
On 5/2/2013 11:20 AM, Stephen Warren wrote:
On 05/02/2013 12:20 AM, David Henningsson wrote:
On 05/01/2013 09:04 PM, Mike Travis wrote:
The audio driver mistakenly allows 64 bit addresses to be created for the audio driver on Nvidia GPUs. Unfortunately, the hardware normally only supports up to 40 bits of DMA. This can cause system panics as well as misdirected data when the address is > 40 bits as the upper part the address is truncated.
Thanks for this patch. Stephen Warren, is this something you can confirm/deny, and do you know what range of hardware this actually applies to?
It's certainly plausible that the GPU's internal physical addresses are less than a full 64-bits, and this probably affects PCI accesses too.
I haven't noticed any other problems with MMIOH accesses. The problem of DMA32 addresses certainly affects us though audio is not a high priority on UV(*) except in specific applications (like Media Servers, etc.] Accesses to the GPU and GPGPU are much more common.
participants (4)
-
David Henningsson -
Mike Travis -
Stephen Warren -
Takashi Iwai