On 1/30/2024 2:22 PM, Tian, Kevin wrote:
From: Ethan Zhao <haifeng.zhao@xxxxxxxxxxxxxxx>
Sent: Tuesday, January 30, 2024 1:13 PM
On 1/29/2024 5:21 PM, Yi Liu wrote:
On 2024/1/29 17:06, Tian, Kevin wrote:
From: Ethan Zhao <haifeng.zhao@xxxxxxxxxxxxxxx>
Sent: Monday, January 29, 2024 11:49 AM
Because surprise removal could happen anytime, e.g. user could
request safe
removal to EP(endpoint device) via sysfs and brings its link down to do
surprise removal cocurrently. such aggressive cases would cause ATS
invalidation request issued to non-existence target device, then deadly
loop to retry that request after ITE fault triggered in interrupt
context.
this patch aims to optimize the ITE handling by checking the target
device
presence state to avoid retrying the timeout request blindly, thus
avoid
hard lockup or system hang.
Signed-off-by: Ethan Zhao <haifeng.zhao@xxxxxxxxxxxxxxx>
---
drivers/iommu/intel/dmar.c | 18 ++++++++++++++++++
1 file changed, 18 insertions(+)
diff --git a/drivers/iommu/intel/dmar.c b/drivers/iommu/intel/dmar.c
index 814134e9aa5a..2e214b43725c 100644
--- a/drivers/iommu/intel/dmar.c
+++ b/drivers/iommu/intel/dmar.c
@@ -1272,6 +1272,7 @@ static int qi_check_fault(struct intel_iommu
*iommu, int index, int wait_index,
{
u32 fault;
int head, tail;
+ u64 iqe_err, ite_sid;
struct q_inval *qi = iommu->qi;
int shift = qi_shift(iommu);
@@ -1316,6 +1317,13 @@ static int qi_check_fault(struct intel_iommu
*iommu, int index, int wait_index,
tail = readl(iommu->reg + DMAR_IQT_REG);
tail = ((tail >> shift) - 1 + QI_LENGTH) % QI_LENGTH;
+ /*
+ * SID field is valid only when the ITE field is Set in
FSTS_REG
+ * see Intel VT-d spec r4.1, section 11.4.9.9
+ */
+ iqe_err = dmar_readq(iommu->reg + DMAR_IQER_REG);
+ ite_sid = DMAR_IQER_REG_ITESID(iqe_err);
+
writel(DMA_FSTS_ITE, iommu->reg + DMAR_FSTS_REG);
pr_info("Invalidation Time-out Error (ITE) cleared\n");
@@ -1325,6 +1333,16 @@ static int qi_check_fault(struct intel_iommu
*iommu, int index, int wait_index,
head = (head - 2 + QI_LENGTH) % QI_LENGTH;
} while (head != tail);
+ /*
+ * If got ITE, we need to check if the sid of ITE is the
same as
+ * current ATS invalidation target device, if yes, don't
try this
+ * request anymore if the target device isn't present.
+ * 0 value of ite_sid means old VT-d device, no ite_sid value.
+ */
+ if (pdev && ite_sid && !pci_device_is_present(pdev) &&
+ ite_sid == pci_dev_id(pci_physfn(pdev)))
+ return -ETIMEDOUT;
+
since the hardware already reports source id leading to timeout,
can't we
just find the pci_dev according to reported ite_sid? this is a slow
path (either
due to device in bad state or removed) hence it's not necessary to
add more
intelligence to pass the pci_dev in, leading to only a partial fix
can be backported.
It's also more future-proof, say if one day the driver allows
batching invalidation
requests for multiple devices then no need to pass in a list of devices.
Then it's easier to backport a full fix.
May consider pci_get_domain_bus_and_slot() or
pci_find_bus()/pci_get_slot(). But I doubt if the pci_dev is still
tracked
in the bus or a kind of dev list in the device hot removal case. So Ethan
may need to check.
Perhaps it is too late to call pci_find_bus() or
pci_get_domain_bus_and_slot() to get the
device instance from this notifier registered as
BUS_NOTIFY_REMOVED_DEVICE
action. if the device is still there in bus list, *must* be a bug of
device subsystem as
*removed* device.
Ethan, looks your reply is not formatted well. Can you fix your mail
client like how you write the commit msg?
Sorry for the mail format.
Here we need consider two situations.
One is that the device is not bound to a driver or bound to a driver
which doesn't do active work to the device when it's removed. In
that case one may observe the timeout situation only in the removal
path as the stack dump in your patch02 shows.
When iommu_bus_notifier() got called for hotplug removal cases to
flush devTLB (ATS invalidation), driver was already unloaded.
whatever safe removal or surprise removal. so in theory no active
driver working there.
pciehp_ist()
pciehp_disable_slot()
remove_board()
pciehp_unconfigure_device()
pci_stop_and_remove_bus_device()
pci_stop_bus_device()--->here unload driver
pci_remove_bus_device()->here qi_flush_dev_iotlb() got called.
patch02 can fix that case by checking whether the device is present
to skip sending the invalidation requests. So the logic being discussed
here doesn't matter.
The 2nd situation is more tricky. The device might be bound to
a driver which is doing active work to the device with in-fly
ATS invalidation requests. In this case in-fly requests must be aborted
before the driver can be detached from the removed device. Conceptually
a device is removed from the bus only after its driver is detached.
Some tricky situations:
1. The ATS invalidation request is issued from driver driver, while it is
in handling, device is removed. this momment, the device instance still
exists in the bus list. yes, if searching it by BDF, could get it.
2. The ATS invalidation request is issued from iommu_bus_notifier()
for surprise removal reason, as shown in above calltrace, device was
already removed from bus list. if searching it by BDF, return NULL.
3. The ATS invlidation request is issued from iommu_bus_notifier()
for safe removal, when is in handling, device is removed or link
is down. also as #2, device was already removed from bus list.
if searching it by BDF. got NULL.
...
so, searching device by BDF, only works for the ATS invalidation
request is from device driver.
Thanks,
Ethan
From this angle you can still find the pci_dev from the bus when handling
timeout error for in-fly invalidation requests. But I'm not a PCI person
so let's wait for the inputs from Bjorn and Lukas.
