On a VT-d capable machine Linux will enable IOMMU by default. If it
then kexec's a second kernel with intel_iommu=off, this second kernel
will leave the DMA remapping engine on with no code handling it. The
symptom is at least USB and SATA drives stop working. This patch fixes
the problem by always disabling DMA remapping when intel_iommu=off.
Signed-off-by: Wei Hu <[email protected]>
---
drivers/iommu/intel-iommu.c | 6 +++++-
1 files changed, 5 insertions(+), 1 deletions(-)
diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
index 0099667..0b8f8bb 100644
--- a/drivers/iommu/intel-iommu.c
+++ b/drivers/iommu/intel-iommu.c
@@ -3681,8 +3681,12 @@ int __init intel_iommu_init(void)
return -ENODEV;
}
- if (no_iommu || dmar_disabled)
+ if (no_iommu || dmar_disabled) {
+ struct dmar_drhd_unit *drhd;
+ for_each_drhd_unit(drhd)
+ iommu_disable_translation(drhd->iommu);
return -ENODEV;
+ }
if (iommu_init_mempool()) {
if (force_on)
--
1.7.4.4
(2013/04/20 15:58), Wei Hu wrote:
> On a VT-d capable machine Linux will enable IOMMU by default. If it
> then kexec's a second kernel with intel_iommu=off, this second kernel
> will leave the DMA remapping engine on with no code handling it. The
> symptom is at least USB and SATA drives stop working. This patch fixes
> the problem by always disabling DMA remapping when intel_iommu=off.
Even when second kernel boots up with intel_iommu=on, dma-remapping need
to be disabled as well before it is initialized and enabled again in
init_dmars(). So, how about something like this?
for_each_drhd_unit(drhd) {
struct dmar_drhd_unit *drhd;
if (drhd->ignored)
continue;
iommu = drhd->iommu;
if (iommu->gcmd & DMA_GCMD_TE)
iommu_disable_translation(iommu);
}
Note, if you agree above code and fix your patch like this, you need
additinal fix to set iommu->gcmd flag sinse gcmd is always zero here.
See first hunk of this patch.
https://lkml.org/lkml/2013/3/20/707
Thanks,
Takao Indoh
>
> Signed-off-by: Wei Hu <[email protected]>
> ---
> drivers/iommu/intel-iommu.c | 6 +++++-
> 1 files changed, 5 insertions(+), 1 deletions(-)
>
> diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
> index 0099667..0b8f8bb 100644
> --- a/drivers/iommu/intel-iommu.c
> +++ b/drivers/iommu/intel-iommu.c
> @@ -3681,8 +3681,12 @@ int __init intel_iommu_init(void)
> return -ENODEV;
> }
>
> - if (no_iommu || dmar_disabled)
> + if (no_iommu || dmar_disabled) {
> + struct dmar_drhd_unit *drhd;
> + for_each_drhd_unit(drhd)
> + iommu_disable_translation(drhd->iommu);
> return -ENODEV;
> + }
>
> if (iommu_init_mempool()) {
> if (force_on)
>
On Sun, Apr 21, 2013 at 10:31 PM, Takao Indoh
<[email protected]> wrote:
> (2013/04/20 15:58), Wei Hu wrote:
>> On a VT-d capable machine Linux will enable IOMMU by default. If it
>> then kexec's a second kernel with intel_iommu=off, this second kernel
>> will leave the DMA remapping engine on with no code handling it. The
>> symptom is at least USB and SATA drives stop working. This patch fixes
>> the problem by always disabling DMA remapping when intel_iommu=off.
>
> Even when second kernel boots up with intel_iommu=on, dma-remapping need
> to be disabled as well before it is initialized and enabled again in
> init_dmars(). So, how about something like this?
>
> for_each_drhd_unit(drhd) {
> struct dmar_drhd_unit *drhd;
> if (drhd->ignored)
> continue;
>
> iommu = drhd->iommu;
> if (iommu->gcmd & DMA_GCMD_TE)
> iommu_disable_translation(iommu);
> }
>
> Note, if you agree above code and fix your patch like this, you need
> additinal fix to set iommu->gcmd flag sinse gcmd is always zero here.
> See first hunk of this patch.
> https://lkml.org/lkml/2013/3/20/707
Thanks for your reply. I reviewed your patch. Are you suggesting
keeping your change to dmar.c, and moving your change to intel-iommu.c
(i.e. the code you just showed) up after the call to
dmar_table_init()?
(2013/04/22 15:50), Wei Hu wrote:
> On Sun, Apr 21, 2013 at 10:31 PM, Takao Indoh
> <[email protected]> wrote:
>> (2013/04/20 15:58), Wei Hu wrote:
>>> On a VT-d capable machine Linux will enable IOMMU by default. If it
>>> then kexec's a second kernel with intel_iommu=off, this second kernel
>>> will leave the DMA remapping engine on with no code handling it. The
>>> symptom is at least USB and SATA drives stop working. This patch fixes
>>> the problem by always disabling DMA remapping when intel_iommu=off.
>>
>> Even when second kernel boots up with intel_iommu=on, dma-remapping need
>> to be disabled as well before it is initialized and enabled again in
>> init_dmars(). So, how about something like this?
>>
>> for_each_drhd_unit(drhd) {
>> struct dmar_drhd_unit *drhd;
>> if (drhd->ignored)
>> continue;
>>
>> iommu = drhd->iommu;
>> if (iommu->gcmd & DMA_GCMD_TE)
>> iommu_disable_translation(iommu);
>> }
>>
>> Note, if you agree above code and fix your patch like this, you need
>> additinal fix to set iommu->gcmd flag sinse gcmd is always zero here.
>> See first hunk of this patch.
>> https://lkml.org/lkml/2013/3/20/707
>
> Thanks for your reply. I reviewed your patch. Are you suggesting
> keeping your change to dmar.c, and moving your change to intel-iommu.c
> (i.e. the code you just showed) up after the call to
> dmar_table_init()?
>
Yes, exactly.
Thanks,
Takao Indoh
On Mon, Apr 22, 2013 at 12:06 AM, Takao Indoh
<[email protected]> wrote:
> (2013/04/22 15:50), Wei Hu wrote:
>> On Sun, Apr 21, 2013 at 10:31 PM, Takao Indoh
>> <[email protected]> wrote:
>>> (2013/04/20 15:58), Wei Hu wrote:
>>>> On a VT-d capable machine Linux will enable IOMMU by default. If it
>>>> then kexec's a second kernel with intel_iommu=off, this second kernel
>>>> will leave the DMA remapping engine on with no code handling it. The
>>>> symptom is at least USB and SATA drives stop working. This patch fixes
>>>> the problem by always disabling DMA remapping when intel_iommu=off.
>>>
>>> Even when second kernel boots up with intel_iommu=on, dma-remapping need
>>> to be disabled as well before it is initialized and enabled again in
>>> init_dmars(). So, how about something like this?
>>>
>>> for_each_drhd_unit(drhd) {
>>> struct dmar_drhd_unit *drhd;
>>> if (drhd->ignored)
>>> continue;
>>>
>>> iommu = drhd->iommu;
>>> if (iommu->gcmd & DMA_GCMD_TE)
>>> iommu_disable_translation(iommu);
>>> }
>>>
>>> Note, if you agree above code and fix your patch like this, you need
>>> additinal fix to set iommu->gcmd flag sinse gcmd is always zero here.
>>> See first hunk of this patch.
>>> https://lkml.org/lkml/2013/3/20/707
>>
>> Thanks for your reply. I reviewed your patch. Are you suggesting
>> keeping your change to dmar.c, and moving your change to intel-iommu.c
>> (i.e. the code you just showed) up after the call to
>> dmar_table_init()?
>>
>
> Yes, exactly.
>
> Thanks,
> Takao Indoh
>
Sounds great. If you want to make the change and get your patch
accepted that's fine with me.