This patch quiesces devices before disabling IOMMU on boot to stop
ongoing DMA. In intel_iommu_init(), check context entries and if there
is entry whose present bit is set then reset corresponding device.
When IOMMU is already enabled on boot, it is disabled and new DMAR table
is created and then re-enabled in intel_iommu_init(). This causes DMAR
faults if there are in-flight DMAs.
This causes problem on kdump. Devices are working in first kernel, and
after switching to second kernel and initializing IOMMU, many DMAR
faults occur and it causes problems like driver error or PCI SERR, at
last kdump fails. This patch fixes this problem.
Changelog:
v2:
- Add CONTEXT_ENTRY_NR
v1:
https://lkml.org/lkml/2013/8/21/71
Signed-off-by: Takao Indoh <[email protected]>
---
drivers/iommu/intel-iommu.c | 56 ++++++++++++++++++++++++++++++++++++++++++-
1 files changed, 55 insertions(+), 1 deletions(-)
diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
index eec0d3e..d0e8aff 100644
--- a/drivers/iommu/intel-iommu.c
+++ b/drivers/iommu/intel-iommu.c
@@ -224,6 +224,7 @@ struct context_entry {
u64 lo;
u64 hi;
};
+#define CONTEXT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct context_entry))
static inline bool context_present(struct context_entry *context)
{
@@ -3663,6 +3664,56 @@ static struct notifier_block device_nb = {
.notifier_call = device_notifier,
};
+/* Reset PCI devices if its entry exists in DMAR table */
+static void __init iommu_reset_devices(struct intel_iommu *iommu, u16 segment)
+{
+ u64 addr;
+ struct root_entry *root;
+ struct context_entry *context;
+ int bus, devfn;
+ struct pci_dev *dev;
+
+ addr = dmar_readq(iommu->reg + DMAR_RTADDR_REG);
+ if (!addr)
+ return;
+
+ /*
+ * In the case of kdump, ioremap is needed because root-entry table
+ * exists in first kernel's memory area which is not mapped in second
+ * kernel
+ */
+ root = (struct root_entry *)ioremap(addr, PAGE_SIZE);
+ if (!root)
+ return;
+
+ for (bus = 0; bus < ROOT_ENTRY_NR; bus++) {
+ if (!root_present(&root[bus]))
+ continue;
+
+ context = (struct context_entry *)ioremap(
+ root[bus].val & VTD_PAGE_MASK, PAGE_SIZE);
+ if (!context)
+ continue;
+
+ for (devfn = 0; devfn < CONTEXT_ENTRY_NR; devfn++) {
+ if (!context_present(&context[devfn]))
+ continue;
+
+ dev = pci_get_domain_bus_and_slot(segment, bus, devfn);
+ if (!dev)
+ continue;
+
+ if (!pci_reset_bus(dev->bus)) /* go to next bus */
+ break;
+ else /* Try per-function reset */
+ pci_reset_function(dev);
+
+ }
+ iounmap(context);
+ }
+ iounmap(root);
+}
+
int __init intel_iommu_init(void)
{
int ret = 0;
@@ -3687,8 +3738,11 @@ int __init intel_iommu_init(void)
continue;
iommu = drhd->iommu;
- if (iommu->gcmd & DMA_GCMD_TE)
+ if (iommu->gcmd & DMA_GCMD_TE) {
+ if (reset_devices)
+ iommu_reset_devices(iommu, drhd->segment);
iommu_disable_translation(iommu);
+ }
}
if (dmar_dev_scope_init() < 0) {
--
1.7.1
On 09/18/13 at 03:09pm, Takao Indoh wrote:
> This patch quiesces devices before disabling IOMMU on boot to stop
> ongoing DMA. In intel_iommu_init(), check context entries and if there
> is entry whose present bit is set then reset corresponding device.
>
> When IOMMU is already enabled on boot, it is disabled and new DMAR table
> is created and then re-enabled in intel_iommu_init(). This causes DMAR
> faults if there are in-flight DMAs.
>
> This causes problem on kdump. Devices are working in first kernel, and
> after switching to second kernel and initializing IOMMU, many DMAR
> faults occur and it causes problems like driver error or PCI SERR, at
> last kdump fails. This patch fixes this problem.
>
> Changelog:
> v2:
> - Add CONTEXT_ENTRY_NR
>
> v1:
> https://lkml.org/lkml/2013/8/21/71
>
> Signed-off-by: Takao Indoh <[email protected]>
> ---
> drivers/iommu/intel-iommu.c | 56 ++++++++++++++++++++++++++++++++++++++++++-
> 1 files changed, 55 insertions(+), 1 deletions(-)
>
> diff --git a/drivers/iommu/intel-iommu.c b/drivers/iommu/intel-iommu.c
> index eec0d3e..d0e8aff 100644
> --- a/drivers/iommu/intel-iommu.c
> +++ b/drivers/iommu/intel-iommu.c
> @@ -224,6 +224,7 @@ struct context_entry {
> u64 lo;
> u64 hi;
> };
> +#define CONTEXT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct context_entry))
>
> static inline bool context_present(struct context_entry *context)
> {
> @@ -3663,6 +3664,56 @@ static struct notifier_block device_nb = {
> .notifier_call = device_notifier,
> };
>
> +/* Reset PCI devices if its entry exists in DMAR table */
> +static void __init iommu_reset_devices(struct intel_iommu *iommu, u16 segment)
> +{
> + u64 addr;
> + struct root_entry *root;
> + struct context_entry *context;
> + int bus, devfn;
> + struct pci_dev *dev;
> +
> + addr = dmar_readq(iommu->reg + DMAR_RTADDR_REG);
> + if (!addr)
> + return;
> +
> + /*
> + * In the case of kdump, ioremap is needed because root-entry table
> + * exists in first kernel's memory area which is not mapped in second
> + * kernel
> + */
> + root = (struct root_entry *)ioremap(addr, PAGE_SIZE);
> + if (!root)
> + return;
> +
> + for (bus = 0; bus < ROOT_ENTRY_NR; bus++) {
> + if (!root_present(&root[bus]))
> + continue;
> +
> + context = (struct context_entry *)ioremap(
> + root[bus].val & VTD_PAGE_MASK, PAGE_SIZE);
> + if (!context)
> + continue;
> +
> + for (devfn = 0; devfn < CONTEXT_ENTRY_NR; devfn++) {
> + if (!context_present(&context[devfn]))
> + continue;
> +
> + dev = pci_get_domain_bus_and_slot(segment, bus, devfn);
> + if (!dev)
> + continue;
> +
> + if (!pci_reset_bus(dev->bus)) /* go to next bus */
> + break;
> + else /* Try per-function reset */
> + pci_reset_function(dev);
> +
> + }
> + iounmap(context);
> + }
> + iounmap(root);
> +}
> +
> int __init intel_iommu_init(void)
> {
> int ret = 0;
> @@ -3687,8 +3738,11 @@ int __init intel_iommu_init(void)
> continue;
>
> iommu = drhd->iommu;
> - if (iommu->gcmd & DMA_GCMD_TE)
> + if (iommu->gcmd & DMA_GCMD_TE) {
> + if (reset_devices)
> + iommu_reset_devices(iommu, drhd->segment);
> iommu_disable_translation(iommu);
> + }
> }
>
> if (dmar_dev_scope_init() < 0) {
Looks good to me, thanks!
Acked-by: Baoquan He <[email protected]>
> --
> 1.7.1
>
>
On Wed, Sep 18, 2013 at 03:09:01PM +0900, Takao Indoh wrote:
> + /*
> + * In the case of kdump, ioremap is needed because root-entry table
> + * exists in first kernel's memory area which is not mapped in second
> + * kernel
> + */
> + root = (struct root_entry *)ioremap(addr, PAGE_SIZE);
> + if (!root)
> + return;
> +
> + for (bus = 0; bus < ROOT_ENTRY_NR; bus++) {
> + if (!root_present(&root[bus]))
> + continue;
> +
> + context = (struct context_entry *)ioremap(
> + root[bus].val & VTD_PAGE_MASK, PAGE_SIZE);
> + if (!context)
> + continue;
> +
> + for (devfn = 0; devfn < CONTEXT_ENTRY_NR; devfn++) {
> + if (!context_present(&context[devfn]))
> + continue;
> +
> + dev = pci_get_domain_bus_and_slot(segment, bus, devfn);
> + if (!dev)
> + continue;
> +
> + if (!pci_reset_bus(dev->bus)) /* go to next bus */
> + break;
> + else /* Try per-function reset */
> + pci_reset_function(dev);
> +
> + }
> + iounmap(context);
> + }
> + iounmap(root);
I am not convinced that this is the right approach. If a device wasn't
translated by VT-d in the old kernel doesn't mean it will not be
translated in the new kernel. How about unconditionally resetting all
PCI busses and/or functions here before IOMMU initialization proceeds?
Joerg
On Tue, 2013-09-24 at 15:16 +0200, Joerg Roedel wrote:
>
> I am not convinced that this is the right approach. If a device wasn't
> translated by VT-d in the old kernel doesn't mean it will not be
> translated in the new kernel. How about unconditionally resetting all
> PCI busses and/or functions here before IOMMU initialization proceeds?
Forget the IOMMU.
If a device is doing DMA in the old kernel, and *continues* doing DMA
under the new kernel, surely something ought to shut it down?
It's the generic kexec or machine_shutdown code which should be doing
this, not IOMMU code. *Especially* not Intel-specific IOMMU code. It
does not live here.
If anything, the IOMMU gives you a way to *survive* this kind of thing
and means that you might get away without quiescing devices when
otherwise you would have died.
--
dwmw2