Message ID | 1522247980-31892-1-git-send-email-timur@codeaurora.org (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On 2018-03-28 15:39, Timur Tabi wrote: > From: Sameer Goel <sgoel@codeaurora.org> > > Set SMMU_GBPA to abort all incoming translations during the SMMU > reset > when SMMUEN==0. > > This prevents a race condition where a stray DMA from the crashed > primary > kernel can try to access an IOVA address as an invalid PA when SMMU > is > disabled during reset in the crash kernel. > > Signed-off-by: Sameer Goel <sgoel@codeaurora.org> > --- > drivers/iommu/arm-smmu-v3.c | 12 ++++++++++++ > 1 file changed, 12 insertions(+) > > diff --git a/drivers/iommu/arm-smmu-v3.c > b/drivers/iommu/arm-smmu-v3.c > index 3f2f1fc68b52..c04a89310c59 100644 > --- a/drivers/iommu/arm-smmu-v3.c > +++ b/drivers/iommu/arm-smmu-v3.c > @@ -2458,6 +2458,18 @@ static int arm_smmu_device_reset(struct > arm_smmu_device *smmu, bool bypass) > if (reg & CR0_SMMUEN) > dev_warn(smmu->dev, "SMMU currently enabled! Resetting...\n"); > > + /* > + * Abort all incoming translations. This can happen in a kdump case > + * where SMMU is initialized when a prior DMA is pending. Just > + * disabling the SMMU in this case might result in writes to > invalid > + * PAs. > + */ > + ret = arm_smmu_update_gbpa(smmu, 1, GBPA_ABORT); > + if (ret) { > + dev_err(smmu->dev, "GBPA not responding to update\n"); > + return ret; > + } > + > ret = arm_smmu_device_disable(smmu); > if (ret) > return ret; A tangential question: can we reliably detect that the SMMU already has valid mappings, which would indicate that we're in a pretty bad shape already by the time we set that bit? For all we know, memory could have been corrupted long before we hit this point, and this patch barely narrows the window of opportunity. At the very least, we should emit a warning and taint the kernel (we've recently added such measures to the GICv3 driver). Thanks, M.
On Wed, Mar 28, 2018 at 09:39:40AM -0500, Timur Tabi wrote: > From: Sameer Goel <sgoel@codeaurora.org> > > Set SMMU_GBPA to abort all incoming translations during the SMMU reset > when SMMUEN==0. > > This prevents a race condition where a stray DMA from the crashed primary > kernel can try to access an IOVA address as an invalid PA when SMMU is > disabled during reset in the crash kernel. > > Signed-off-by: Sameer Goel <sgoel@codeaurora.org> > --- > drivers/iommu/arm-smmu-v3.c | 12 ++++++++++++ > 1 file changed, 12 insertions(+) > > diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c > index 3f2f1fc68b52..c04a89310c59 100644 > --- a/drivers/iommu/arm-smmu-v3.c > +++ b/drivers/iommu/arm-smmu-v3.c > @@ -2458,6 +2458,18 @@ static int arm_smmu_device_reset(struct arm_smmu_device *smmu, bool bypass) > if (reg & CR0_SMMUEN) > dev_warn(smmu->dev, "SMMU currently enabled! Resetting...\n"); > > + /* > + * Abort all incoming translations. This can happen in a kdump case > + * where SMMU is initialized when a prior DMA is pending. Just > + * disabling the SMMU in this case might result in writes to invalid > + * PAs. > + */ > + ret = arm_smmu_update_gbpa(smmu, 1, GBPA_ABORT); > + if (ret) { > + dev_err(smmu->dev, "GBPA not responding to update\n"); > + return ret; > + } This needs to be predicated on the disable_bypass option, otherwise I think it will cause regressions for systems that rely on passthrough. Will
On 4/5/2018 5:26 AM, Will Deacon wrote: > On Wed, Mar 28, 2018 at 09:39:40AM -0500, Timur Tabi wrote: >> From: Sameer Goel <sgoel@codeaurora.org> >> >> Set SMMU_GBPA to abort all incoming translations during the SMMU reset >> when SMMUEN==0. >> >> This prevents a race condition where a stray DMA from the crashed primary >> kernel can try to access an IOVA address as an invalid PA when SMMU is >> disabled during reset in the crash kernel. >> >> Signed-off-by: Sameer Goel <sgoel@codeaurora.org> >> --- >> drivers/iommu/arm-smmu-v3.c | 12 ++++++++++++ >> 1 file changed, 12 insertions(+) >> >> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c >> index 3f2f1fc68b52..c04a89310c59 100644 >> --- a/drivers/iommu/arm-smmu-v3.c >> +++ b/drivers/iommu/arm-smmu-v3.c >> @@ -2458,6 +2458,18 @@ static int arm_smmu_device_reset(struct arm_smmu_device *smmu, bool bypass) >> if (reg & CR0_SMMUEN) >> dev_warn(smmu->dev, "SMMU currently enabled! Resetting...\n"); >> >> + /* >> + * Abort all incoming translations. This can happen in a kdump case >> + * where SMMU is initialized when a prior DMA is pending. Just >> + * disabling the SMMU in this case might result in writes to invalid >> + * PAs. >> + */ >> + ret = arm_smmu_update_gbpa(smmu, 1, GBPA_ABORT); >> + if (ret) { >> + dev_err(smmu->dev, "GBPA not responding to update\n"); >> + return ret; >> + } > > This needs to be predicated on the disable_bypass option, otherwise I think > it will cause regressions for systems that rely on passthrough. Ok, I'll make the change. > > Will >
On 3/28/2018 9:00 AM, Marc Zyngier wrote: > On 2018-03-28 15:39, Timur Tabi wrote: >> From: Sameer Goel <sgoel@codeaurora.org> >> >> Set SMMU_GBPA to abort all incoming translations during the SMMU reset >> when SMMUEN==0. >> >> This prevents a race condition where a stray DMA from the crashed primary >> kernel can try to access an IOVA address as an invalid PA when SMMU is >> disabled during reset in the crash kernel. >> >> Signed-off-by: Sameer Goel <sgoel@codeaurora.org> >> --- >> drivers/iommu/arm-smmu-v3.c | 12 ++++++++++++ >> 1 file changed, 12 insertions(+) >> >> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c >> index 3f2f1fc68b52..c04a89310c59 100644 >> --- a/drivers/iommu/arm-smmu-v3.c >> +++ b/drivers/iommu/arm-smmu-v3.c >> @@ -2458,6 +2458,18 @@ static int arm_smmu_device_reset(struct >> arm_smmu_device *smmu, bool bypass) >> if (reg & CR0_SMMUEN) >> dev_warn(smmu->dev, "SMMU currently enabled! Resetting...\n"); >> >> + /* >> + * Abort all incoming translations. This can happen in a kdump case >> + * where SMMU is initialized when a prior DMA is pending. Just >> + * disabling the SMMU in this case might result in writes to invalid >> + * PAs. >> + */ >> + ret = arm_smmu_update_gbpa(smmu, 1, GBPA_ABORT); >> + if (ret) { >> + dev_err(smmu->dev, "GBPA not responding to update\n"); >> + return ret; >> + } >> + >> ret = arm_smmu_device_disable(smmu); >> if (ret) >> return ret; > > A tangential question: can we reliably detect that the SMMU already has valid mappings, which would indicate that we're in a pretty bad shape already by the time we set that bit? For all we know, memory could have been corrupted long before we hit this point, and this patch barely narrows the window of opportunity. :) Yes that is correct. This only covers the kdump scenario. Trying to get some reliability when booting up the crash kernel. The system is already in a bad state. I don't think that this will happen in a normal scenario. But please point me to the GICv3 change and I'll have a look. Thanks, Sameer > > At the very least, we should emit a warning and taint the kernel (we've recently added such measures to the GICv3 driver). > > Thanks, > > M.
Hi Sammer, On 11/04/18 16:58, Goel, Sameer wrote: > > > On 3/28/2018 9:00 AM, Marc Zyngier wrote: >> On 2018-03-28 15:39, Timur Tabi wrote: >>> From: Sameer Goel <sgoel@codeaurora.org> >>> >>> Set SMMU_GBPA to abort all incoming translations during the SMMU reset >>> when SMMUEN==0. >>> >>> This prevents a race condition where a stray DMA from the crashed primary >>> kernel can try to access an IOVA address as an invalid PA when SMMU is >>> disabled during reset in the crash kernel. >>> >>> Signed-off-by: Sameer Goel <sgoel@codeaurora.org> >>> --- >>> drivers/iommu/arm-smmu-v3.c | 12 ++++++++++++ >>> 1 file changed, 12 insertions(+) >>> >>> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c >>> index 3f2f1fc68b52..c04a89310c59 100644 >>> --- a/drivers/iommu/arm-smmu-v3.c >>> +++ b/drivers/iommu/arm-smmu-v3.c >>> @@ -2458,6 +2458,18 @@ static int arm_smmu_device_reset(struct >>> arm_smmu_device *smmu, bool bypass) >>> if (reg & CR0_SMMUEN) >>> dev_warn(smmu->dev, "SMMU currently enabled! Resetting...\n"); >>> >>> + /* >>> + * Abort all incoming translations. This can happen in a kdump case >>> + * where SMMU is initialized when a prior DMA is pending. Just >>> + * disabling the SMMU in this case might result in writes to invalid >>> + * PAs. >>> + */ >>> + ret = arm_smmu_update_gbpa(smmu, 1, GBPA_ABORT); >>> + if (ret) { >>> + dev_err(smmu->dev, "GBPA not responding to update\n"); >>> + return ret; >>> + } >>> + >>> ret = arm_smmu_device_disable(smmu); >>> if (ret) >>> return ret; >> >> A tangential question: can we reliably detect that the SMMU already >> has valid mappings, which would indicate that we're in a pretty bad >> shape already by the time we set that bit? For all we know, memory >> could have been corrupted long before we hit this point, and this >> patch barely narrows the window of opportunity. > > :) Yes that is correct. This only covers the kdump scenario. Trying > to get some reliability when booting up the crash kernel. The system > is already in a bad state. I don't think that this will happen in a > normal scenario. But please point me to the GICv3 change and I'll > have a look. See this: https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/tree/drivers/irqchip/irq-gic-v3-its.c?h=irq/irqchip-4.17&id=6eb486b66a3094cdcd68dc39c9df3a29d6a51dd5#n3407 M.
On 11/04/18 17:54, Marc Zyngier wrote: > Hi Sammer, > > On 11/04/18 16:58, Goel, Sameer wrote: >> >> >> On 3/28/2018 9:00 AM, Marc Zyngier wrote: >>> On 2018-03-28 15:39, Timur Tabi wrote: >>>> From: Sameer Goel <sgoel@codeaurora.org> >>>> >>>> Set SMMU_GBPA to abort all incoming translations during the SMMU reset >>>> when SMMUEN==0. >>>> >>>> This prevents a race condition where a stray DMA from the crashed primary >>>> kernel can try to access an IOVA address as an invalid PA when SMMU is >>>> disabled during reset in the crash kernel. >>>> >>>> Signed-off-by: Sameer Goel <sgoel@codeaurora.org> >>>> --- >>>> drivers/iommu/arm-smmu-v3.c | 12 ++++++++++++ >>>> 1 file changed, 12 insertions(+) >>>> >>>> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c >>>> index 3f2f1fc68b52..c04a89310c59 100644 >>>> --- a/drivers/iommu/arm-smmu-v3.c >>>> +++ b/drivers/iommu/arm-smmu-v3.c >>>> @@ -2458,6 +2458,18 @@ static int arm_smmu_device_reset(struct >>>> arm_smmu_device *smmu, bool bypass) >>>> if (reg & CR0_SMMUEN) >>>> dev_warn(smmu->dev, "SMMU currently enabled! Resetting...\n"); >>>> >>>> + /* >>>> + * Abort all incoming translations. This can happen in a kdump case >>>> + * where SMMU is initialized when a prior DMA is pending. Just >>>> + * disabling the SMMU in this case might result in writes to invalid >>>> + * PAs. >>>> + */ >>>> + ret = arm_smmu_update_gbpa(smmu, 1, GBPA_ABORT); >>>> + if (ret) { >>>> + dev_err(smmu->dev, "GBPA not responding to update\n"); >>>> + return ret; >>>> + } >>>> + >>>> ret = arm_smmu_device_disable(smmu); >>>> if (ret) >>>> return ret; >>> >>> A tangential question: can we reliably detect that the SMMU already >>> has valid mappings, which would indicate that we're in a pretty bad >>> shape already by the time we set that bit? For all we know, memory >>> could have been corrupted long before we hit this point, and this >>> patch barely narrows the window of opportunity. >> >> :) Yes that is correct. This only covers the kdump scenario. Trying >> to get some reliability when booting up the crash kernel. The system >> is already in a bad state. I don't think that this will happen in a >> normal scenario. But please point me to the GICv3 change and I'll >> have a look. > > See this: > https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/tree/drivers/irqchip/irq-gic-v3-its.c?h=irq/irqchip-4.17&id=6eb486b66a3094cdcd68dc39c9df3a29d6a51dd5#n3407 The nearest equivalent to that is probably the top-level SMMUEN check that we already have (see the diff context above). To go beyond that you'd have to chase the old stream table pointer and scan the whole thing looking for valid contexts, then potentially walk page tables within those contexts to check for live translations if you really wanted to be sure. That would be a hell of a lot of work to do in the boot path. Robin.
On Thu, Apr 12, 2018 at 11:17:24AM +0100, Robin Murphy wrote: > On 11/04/18 17:54, Marc Zyngier wrote: > >On 11/04/18 16:58, Goel, Sameer wrote: > >>On 3/28/2018 9:00 AM, Marc Zyngier wrote: > >>>A tangential question: can we reliably detect that the SMMU already > >>>has valid mappings, which would indicate that we're in a pretty bad > >>>shape already by the time we set that bit? For all we know, memory > >>>could have been corrupted long before we hit this point, and this > >>>patch barely narrows the window of opportunity. > >> > >>:) Yes that is correct. This only covers the kdump scenario. Trying > >>to get some reliability when booting up the crash kernel. The system > >>is already in a bad state. I don't think that this will happen in a > >>normal scenario. But please point me to the GICv3 change and I'll > >>have a look. > > > >See this: > >https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/tree/drivers/irqchip/irq-gic-v3-its.c?h=irq/irqchip-4.17&id=6eb486b66a3094cdcd68dc39c9df3a29d6a51dd5#n3407 > > The nearest equivalent to that is probably the top-level SMMUEN check that > we already have (see the diff context above). To go beyond that you'd have > to chase the old stream table pointer and scan the whole thing looking for > valid contexts, then potentially walk page tables within those contexts to > check for live translations if you really wanted to be sure. That would be a > hell of a lot of work to do in the boot path. Yeah, please don't waste time writing a patch to do that! ;) Will
On 12/04/18 11:17, Robin Murphy wrote: > On 11/04/18 17:54, Marc Zyngier wrote: >> Hi Sammer, >> >> On 11/04/18 16:58, Goel, Sameer wrote: >>> >>> >>> On 3/28/2018 9:00 AM, Marc Zyngier wrote: >>>> On 2018-03-28 15:39, Timur Tabi wrote: >>>>> From: Sameer Goel <sgoel@codeaurora.org> >>>>> >>>>> Set SMMU_GBPA to abort all incoming translations during the SMMU reset >>>>> when SMMUEN==0. >>>>> >>>>> This prevents a race condition where a stray DMA from the crashed primary >>>>> kernel can try to access an IOVA address as an invalid PA when SMMU is >>>>> disabled during reset in the crash kernel. >>>>> >>>>> Signed-off-by: Sameer Goel <sgoel@codeaurora.org> >>>>> --- >>>>> drivers/iommu/arm-smmu-v3.c | 12 ++++++++++++ >>>>> 1 file changed, 12 insertions(+) >>>>> >>>>> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c >>>>> index 3f2f1fc68b52..c04a89310c59 100644 >>>>> --- a/drivers/iommu/arm-smmu-v3.c >>>>> +++ b/drivers/iommu/arm-smmu-v3.c >>>>> @@ -2458,6 +2458,18 @@ static int arm_smmu_device_reset(struct >>>>> arm_smmu_device *smmu, bool bypass) >>>>> if (reg & CR0_SMMUEN) >>>>> dev_warn(smmu->dev, "SMMU currently enabled! Resetting...\n"); >>>>> >>>>> + /* >>>>> + * Abort all incoming translations. This can happen in a kdump case >>>>> + * where SMMU is initialized when a prior DMA is pending. Just >>>>> + * disabling the SMMU in this case might result in writes to invalid >>>>> + * PAs. >>>>> + */ >>>>> + ret = arm_smmu_update_gbpa(smmu, 1, GBPA_ABORT); >>>>> + if (ret) { >>>>> + dev_err(smmu->dev, "GBPA not responding to update\n"); >>>>> + return ret; >>>>> + } >>>>> + >>>>> ret = arm_smmu_device_disable(smmu); >>>>> if (ret) >>>>> return ret; >>>> >>>> A tangential question: can we reliably detect that the SMMU already >>>> has valid mappings, which would indicate that we're in a pretty bad >>>> shape already by the time we set that bit? For all we know, memory >>>> could have been corrupted long before we hit this point, and this >>>> patch barely narrows the window of opportunity. >>> >>> :) Yes that is correct. This only covers the kdump scenario. Trying >>> to get some reliability when booting up the crash kernel. The system >>> is already in a bad state. I don't think that this will happen in a >>> normal scenario. But please point me to the GICv3 change and I'll >>> have a look. >> >> See this: >> https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/tree/drivers/irqchip/irq-gic-v3-its.c?h=irq/irqchip-4.17&id=6eb486b66a3094cdcd68dc39c9df3a29d6a51dd5#n3407 > > The nearest equivalent to that is probably the top-level SMMUEN check > that we already have (see the diff context above). To go beyond that > you'd have to chase the old stream table pointer and scan the whole > thing looking for valid contexts, then potentially walk page tables > within those contexts to check for live translations if you really > wanted to be sure. That would be a hell of a lot of work to do in the > boot path. Yeah, feels a bit too involved for sanity. I'd simply suggest you taint the kernel if you find the SMMU enabled, as you're already on shaky ground. Thanks, M.
On 4/12/2018 5:56 AM, Marc Zyngier wrote: > On 12/04/18 11:17, Robin Murphy wrote: >> On 11/04/18 17:54, Marc Zyngier wrote: >>> Hi Sammer, >>> >>> On 11/04/18 16:58, Goel, Sameer wrote: >>>> >>>> >>>> On 3/28/2018 9:00 AM, Marc Zyngier wrote: >>>>> On 2018-03-28 15:39, Timur Tabi wrote: >>>>>> From: Sameer Goel <sgoel@codeaurora.org> >>>>>> >>>>>> Set SMMU_GBPA to abort all incoming translations during the SMMU reset >>>>>> when SMMUEN==0. >>>>>> >>>>>> This prevents a race condition where a stray DMA from the crashed primary >>>>>> kernel can try to access an IOVA address as an invalid PA when SMMU is >>>>>> disabled during reset in the crash kernel. >>>>>> >>>>>> Signed-off-by: Sameer Goel <sgoel@codeaurora.org> >>>>>> --- >>>>>> drivers/iommu/arm-smmu-v3.c | 12 ++++++++++++ >>>>>> 1 file changed, 12 insertions(+) >>>>>> >>>>>> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c >>>>>> index 3f2f1fc68b52..c04a89310c59 100644 >>>>>> --- a/drivers/iommu/arm-smmu-v3.c >>>>>> +++ b/drivers/iommu/arm-smmu-v3.c >>>>>> @@ -2458,6 +2458,18 @@ static int arm_smmu_device_reset(struct >>>>>> arm_smmu_device *smmu, bool bypass) >>>>>> if (reg & CR0_SMMUEN) >>>>>> dev_warn(smmu->dev, "SMMU currently enabled! Resetting...\n"); >>>>>> >>>>>> + /* >>>>>> + * Abort all incoming translations. This can happen in a kdump case >>>>>> + * where SMMU is initialized when a prior DMA is pending. Just >>>>>> + * disabling the SMMU in this case might result in writes to invalid >>>>>> + * PAs. >>>>>> + */ >>>>>> + ret = arm_smmu_update_gbpa(smmu, 1, GBPA_ABORT); >>>>>> + if (ret) { >>>>>> + dev_err(smmu->dev, "GBPA not responding to update\n"); >>>>>> + return ret; >>>>>> + } >>>>>> + >>>>>> ret = arm_smmu_device_disable(smmu); >>>>>> if (ret) >>>>>> return ret; >>>>> >>>>> A tangential question: can we reliably detect that the SMMU already >>>>> has valid mappings, which would indicate that we're in a pretty bad >>>>> shape already by the time we set that bit? For all we know, memory >>>>> could have been corrupted long before we hit this point, and this >>>>> patch barely narrows the window of opportunity. >>>> >>>> :) Yes that is correct. This only covers the kdump scenario. Trying >>>> to get some reliability when booting up the crash kernel. The system >>>> is already in a bad state. I don't think that this will happen in a >>>> normal scenario. But please point me to the GICv3 change and I'll >>>> have a look. >>> >>> See this: >>> https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/tree/drivers/irqchip/irq-gic-v3-its.c?h=irq/irqchip-4.17&id=6eb486b66a3094cdcd68dc39c9df3a29d6a51dd5#n3407 >> >> The nearest equivalent to that is probably the top-level SMMUEN check >> that we already have (see the diff context above). To go beyond that >> you'd have to chase the old stream table pointer and scan the whole >> thing looking for valid contexts, then potentially walk page tables >> within those contexts to check for live translations if you really >> wanted to be sure. That would be a hell of a lot of work to do in the >> boot path. > Yeah, feels a bit too involved for sanity. I'd simply suggest you taint > the kernel if you find the SMMU enabled, as you're already on shaky ground. Ok. I think since this is a kdump kernel a taint is not necessary? > > Thanks, > > M. >
Hi Mark, On 4/12/2018 7:56 AM, Marc Zyngier wrote: > On 12/04/18 11:17, Robin Murphy wrote: >> On 11/04/18 17:54, Marc Zyngier wrote: >>> Hi Sammer, >>> >>> On 11/04/18 16:58, Goel, Sameer wrote: >>>> >>>> >>>> On 3/28/2018 9:00 AM, Marc Zyngier wrote: >>>>> On 2018-03-28 15:39, Timur Tabi wrote: >>>>>> From: Sameer Goel <sgoel@codeaurora.org> >>>>>> >>>>>> Set SMMU_GBPA to abort all incoming translations during the SMMU reset >>>>>> when SMMUEN==0. >>>>>> >>>>>> This prevents a race condition where a stray DMA from the crashed primary >>>>>> kernel can try to access an IOVA address as an invalid PA when SMMU is >>>>>> disabled during reset in the crash kernel. >>>>>> >>>>>> Signed-off-by: Sameer Goel <sgoel@codeaurora.org> >>>>>> --- >>>>>> drivers/iommu/arm-smmu-v3.c | 12 ++++++++++++ >>>>>> 1 file changed, 12 insertions(+) >>>>>> >>>>>> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c >>>>>> index 3f2f1fc68b52..c04a89310c59 100644 >>>>>> --- a/drivers/iommu/arm-smmu-v3.c >>>>>> +++ b/drivers/iommu/arm-smmu-v3.c >>>>>> @@ -2458,6 +2458,18 @@ static int arm_smmu_device_reset(struct >>>>>> arm_smmu_device *smmu, bool bypass) >>>>>> if (reg & CR0_SMMUEN) >>>>>> dev_warn(smmu->dev, "SMMU currently enabled! Resetting...\n"); >>>>>> >>>>>> + /* >>>>>> + * Abort all incoming translations. This can happen in a kdump case >>>>>> + * where SMMU is initialized when a prior DMA is pending. Just >>>>>> + * disabling the SMMU in this case might result in writes to invalid >>>>>> + * PAs. >>>>>> + */ >>>>>> + ret = arm_smmu_update_gbpa(smmu, 1, GBPA_ABORT); >>>>>> + if (ret) { >>>>>> + dev_err(smmu->dev, "GBPA not responding to update\n"); >>>>>> + return ret; >>>>>> + } >>>>>> + >>>>>> ret = arm_smmu_device_disable(smmu); >>>>>> if (ret) >>>>>> return ret; >>>>> >>>>> A tangential question: can we reliably detect that the SMMU already >>>>> has valid mappings, which would indicate that we're in a pretty bad >>>>> shape already by the time we set that bit? For all we know, memory >>>>> could have been corrupted long before we hit this point, and this >>>>> patch barely narrows the window of opportunity. >>>> >>>> :) Yes that is correct. This only covers the kdump scenario. Trying >>>> to get some reliability when booting up the crash kernel. The system >>>> is already in a bad state. I don't think that this will happen in a >>>> normal scenario. But please point me to the GICv3 change and I'll >>>> have a look. >>> >>> See this: >>> https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/tree/drivers/irqchip/irq-gic-v3-its.c?h=irq/irqchip-4.17&id=6eb486b66a3094cdcd68dc39c9df3a29d6a51dd5#n3407 >> >> The nearest equivalent to that is probably the top-level SMMUEN check >> that we already have (see the diff context above). To go beyond that >> you'd have to chase the old stream table pointer and scan the whole >> thing looking for valid contexts, then potentially walk page tables >> within those contexts to check for live translations if you really >> wanted to be sure. That would be a hell of a lot of work to do in the >> boot path. > Yeah, feels a bit too involved for sanity. I'd simply suggest you taint > the kernel if you find the SMMU enabled, as you're already on shaky ground. Finding the SMMU already enabled does not necessarily indicate that anything catastrophic has occurred. For instance, to support OSes without an SMMUv3 driver, boot FW may have enabled the SMMU and installed 1-to-1 mappings for DDR and MSI target addr(s) to compensate for a MSI-capable master whose default DMA attrs needed tweaking (ex: non-coherent -> coherent). If such a configuration warrants tainting the kernel, then we should similarly check GBPA for attr overrides and taint the kernel if any are found there. > > Thanks, > > M. >
diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c index 3f2f1fc68b52..c04a89310c59 100644 --- a/drivers/iommu/arm-smmu-v3.c +++ b/drivers/iommu/arm-smmu-v3.c @@ -2458,6 +2458,18 @@ static int arm_smmu_device_reset(struct arm_smmu_device *smmu, bool bypass) if (reg & CR0_SMMUEN) dev_warn(smmu->dev, "SMMU currently enabled! Resetting...\n"); + /* + * Abort all incoming translations. This can happen in a kdump case + * where SMMU is initialized when a prior DMA is pending. Just + * disabling the SMMU in this case might result in writes to invalid + * PAs. + */ + ret = arm_smmu_update_gbpa(smmu, 1, GBPA_ABORT); + if (ret) { + dev_err(smmu->dev, "GBPA not responding to update\n"); + return ret; + } + ret = arm_smmu_device_disable(smmu); if (ret) return ret;