iommu/vt-d: Some fine tuning of SVA
Subsystems (2)
Lu Baolu
maintainercover iommu/vt-d: Some fine tuning of SVA
Hi folks,
This includes several tunings of Intel SVA implementation. I plan to target them for v5.19. Please help to review.
Best regards,
baolu
Lu Baolu (3):
iommu/vt-d: Set PGSNP bit in pasid table entry for sva binding iommu/vt-d: Drop stop marker messages
iommu/vt-d: Size Page Request Queue to avoid overflow condition
include/linux/intel-svm.h | 2 +- drivers/iommu/intel/pasid.c | 2 +- drivers/iommu/intel/svm.c | 5 +++++ 3 files changed, 7 insertions(+), 2 deletions(-)
Signature · 6 lines
2.25.1 _______________________________________________ iommu mailing list [email protected] https://lists.linuxfoundation.org/mailman/listinfo/iommu
Lu Baolu
maintainerPATCH 1/3 iommu/vt-d: Set PGSNP bit in pasid table entry for sva binding
This field make the requests snoop processor caches irrespective of other attributes in the request or other fields in paging structure entries used to translate the request. The latest VT-d specification states that this field is treated as Reserved(0) for implementations not supporting Snoop Control (SC=0 in the Extended Capability Register). Hence add a check in the code.
--- drivers/iommu/intel/pasid.c | 2 +- drivers/iommu/intel/svm.c | 1 + 2 files changed, 2 insertions(+), 1 deletion(-)
drivers/iommu/intel/pasid.c +1 -1
@@ -625,7 +625,7 @@ int intel_pasid_setup_first_level(struct intel_iommu *iommu, } } -if (flags &PASID_FLAG_PAGE_SNOOP ) +if ((flags &PASID_FLAG_PAGE_SNOOP ) &&ecap_sc_support (iommu->ecap))pasid_set_pgsnp (pte);pasid_set_domain_id (pte, did);
drivers/iommu/intel/svm.c +1 -0
@@ -394,6 +394,7 @@ static struct iommu_sva *intel_svm_bind_mm(struct intel_iommu *iommu, sflags = (flags &SVM_FLAG_SUPERVISOR_MODE ) ?PASID_FLAG_SUPERVISOR_MODE :0 ; sflags |=cpu_feature_enabled (X86_FEATURE_LA57 ) ?PASID_FLAG_FL5LP :0 ; + sflags |=PASID_FLAG_PAGE_SNOOP ;spin_lock_irqsave (&iommu->lock, iflags); ret =intel_pasid_setup_first_level (iommu, dev, mm->pgd, mm->pasid,FLPT_DEFAULT_DID , sflags);
Signature · 6 lines
2.25.1 _______________________________________________ iommu mailing list [email protected] https://lists.linuxfoundation.org/mailman/listinfo/iommu
RE: [PATCH 1/3] iommu/vt-d: Set PGSNP bit in pasid table entry for sva binding
Quoted 21 lines
From: Lu Baolu <[email protected]>
Sent: Saturday, April 16, 2022 8:31 PMThis field make the requests snoop processor caches irrespective of other attributes in the request or other fields in paging structure entries used to translate the request. The latest VT-d specification states that this field is treated as Reserved(0) for implementations not supporting Snoop Control (SC=0 in the Extended Capability Register). Hence add a check in the code.
Signed-off-by Lu Baolu <[email protected]>--- drivers/iommu/intel/pasid.c | 2 +- drivers/iommu/intel/svm.c | 1 + 2 files changed, 2 insertions(+), 1 deletion(-)drivers/iommu/intel/pasid.c +0
-0
drivers/iommu/intel/pasid.c int intel_pasid_setup_first_level(struct intel_iommu · line 631
Quoted 8 lines
drivers/iommu/intel/pasid.c +1
-1@@ -625,7 +625,7 @@ int intel_pasid_setup_first_level(struct intel_iommu *iommu, } } -if (flags &PASID_FLAG_PAGE_SNOOP ) +if ((flags &PASID_FLAG_PAGE_SNOOP ) &&ecap_sc_support (iommu- >ecap))pasid_set_pgsnp (pte);
If the caller wants snoop for some reason is it correct to simply ignore the request when lacking of hw support? Suppose certain errno should be returned here...
Quoted 15 lines
drivers/iommu/intel/pasid.c +0
-0pasid_set_domain_id (pte, did);drivers/iommu/intel/svm.c +1
-0@@ -394,6 +394,7 @@ static struct iommu_sva *intel_svm_bind_mm(structintel_iommu *iommu, sflags = (flags &SVM_FLAG_SUPERVISOR_MODE ) ?PASID_FLAG_SUPERVISOR_MODE :0 ; sflags |=cpu_feature_enabled (X86_FEATURE_LA57 ) ?PASID_FLAG_FL5LP :0 ; + sflags |=PASID_FLAG_PAGE_SNOOP ;spin_lock_irqsave (&iommu->lock, iflags);ret = intel_pasid_setup_first_level(iommu, dev,
mm->pgd, mm-
Quoted 4 lines
Quoted 1 line
pasid,
FLPT_DEFAULT_DID, sflags);
--
2.25.1
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
Lu Baolu
maintainerRe: [PATCH 1/3] iommu/vt-d: Set PGSNP bit in pasid table entry for sva binding
Quoted 25 lines · Kevin
Quoted 25 lines
From: Lu Baolu<[email protected]>
Sent: Saturday, April 16, 2022 8:31 PMThis field make the requests snoop processor caches irrespective of other attributes in the request or other fields in paging structure entries used to translate the request. The latest VT-d specification states that this field is treated as Reserved(0) for implementations not supporting Snoop Control (SC=0 in the Extended Capability Register). Hence add a check in the code.
Signed-off-by Lu Baolu<[email protected]>--- drivers/iommu/intel/pasid.c | 2 +- drivers/iommu/intel/svm.c | 1 + 2 files changed, 2 insertions(+), 1 deletion(-)drivers/iommu/intel/pasid.c +0
-0@@ -625,7 +625,7 @@ int intel_pasid_setup_first_level(struct intel_iommu *iommu, } }
Quoted 7 lines
drivers/iommu/intel/pasid.c int intel_pasid_setup_first_level(struct intel_iommu *iommu, · line 629
Quoted 4 lines
drivers/iommu/intel/pasid.c +1
-1-if (flags &PASID_FLAG_PAGE_SNOOP ) +if ((flags &PASID_FLAG_PAGE_SNOOP ) &&ecap_sc_support (iommu- > ecap))pasid_set_pgsnp (pte);If the caller wants snoop for some reason is it correct to simply ignore the request when lacking of hw support? Suppose certain errno should be returned here...
Good catch. Perhaps I should make the cap check in a separated patch.
Best regards,
baolu
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
Lu Baolu
maintainerPATCH 2/3 iommu/vt-d: Drop stop marker messages
In reply to Lu BaoluThe VT-d driver explicitly drains the pending page requests when a CPU page table (represented by a mm struct) is unbound from a PASID according to the procedures defined in the VT-d spec. Hence, there's no need to report the stop-marker message in prq_event_thread(). The stop marker messages do not need a response. This drops stop marker messages silently if any of them is found in the page request queue.
--- drivers/iommu/intel/svm.c | 4 ++++ 1 file changed, 4 insertions(+)
drivers/iommu/intel/svm.c +4 -0
@@ -758,6 +758,10 @@ static irqreturn_t prq_event_thread(int irq, void *d) goto bad_req; } +/* Drop Stop Marker message. No need for a response. */ +if (unlikely (req->lpig && !req->rd_req && !req->wr_req)) + goto prq_advance; +if (!svm || svm->pasid != req->pasid) {/* * It can't go away, because the driver is not permitted
Signature · 6 lines
2.25.1 _______________________________________________ iommu mailing list [email protected] https://lists.linuxfoundation.org/mailman/listinfo/iommu
RE: [PATCH 2/3] iommu/vt-d: Drop stop marker messages
Quoted 9 lines
From: Lu Baolu <[email protected]>
Sent: Saturday, April 16, 2022 8:31 PMThe VT-d driver explicitly drains the pending page requests when a CPU page table (represented by a mm struct) is unbound from a PASID according to the procedures defined in the VT-d spec. Hence, there's no need to report the stop-marker message in
prq_event_thread(). The stop marker messages do not need a response. This drops stop marker messages silently if any of them is found in the page request queue.
The comment for iommu_queue_iopf says:
* This module doesn't handle PCI PASID Stop Marker; IOMMU drivers must discard
* them before reporting faults. A PASID Stop Marker (LRW = 0b100) doesn't
* expect a response. It may be generated when disabling a PASID (issuing a
* PASID stop request) by some PCI devices.
So obviously the current vt-d driver behavior violates that requirement. Then should this be a bug fix instead?
Quoted 22 lines
Signed-off-by Lu Baolu <[email protected]>Reviewed-by Jacob Pan <[email protected]>--- drivers/iommu/intel/svm.c | 4 ++++ 1 file changed, 4 insertions(+)drivers/iommu/intel/svm.c +4
-0@@ -758,6 +758,10 @@ static irqreturn_t prq_event_thread(int irq, void *d) goto bad_req; } +/* Drop Stop Marker message. No need for a response. */ +if (unlikely (req->lpig && !req->rd_req && !req->wr_req)) + goto prq_advance; +if (!svm || svm->pasid != req->pasid) {/* * It can't go away, because the driver is not
Quoted 3 lines
permitted
--
2.25.1
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
Lu Baolu
maintainerRe: [PATCH 2/3] iommu/vt-d: Drop stop marker messages
Quoted 11 lines · Kevin
Quoted 9 lines
From: Lu Baolu<[email protected]>
Sent: Saturday, April 16, 2022 8:31 PMThe VT-d driver explicitly drains the pending page requests when a CPU page table (represented by a mm struct) is unbound from a PASID according to the procedures defined in the VT-d spec. Hence, there's no need to report the stop-marker message in
prq_event_thread(). The stop marker messages do not need a response. This drops stop marker messages silently if any of them is found in the page request queue.The comment for
iommu_queue_iopfsays:
Quoted 8 lines
* This module doesn't handle PCI PASID Stop Marker; IOMMU drivers must discard
* them before reporting faults. A PASID Stop Marker (LRW = 0b100) doesn't
* expect a response. It may be generated when disabling a PASID (issuing a
* PASID stop request) by some PCI devices.So obviously the current vt-d driver behavior violates that requirement. Then should this be a bug fix instead?
Our platforms have no devices issuing Stop Marker messages yet. But in a virtualization environment, a device might be emulated with this capability. Yes! Let me make it as a fix.
Best regards,
baolu
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
Lu Baolu
maintainerPATCH 3/3 iommu/vt-d: Size Page Request Queue to avoid overflow condition
In reply to Lu BaoluPRQ overflow may cause I/O throughput congestion, resulting in unnecessary degradation of IO performance. Appropriately increasing the length of PRQ can greatly reduce the occurrence of PRQ overflow. The count of maximum page requests that can be generated in parallel by a PCIe device is statically defined in the Outstanding Page Request Capacity field of the PCIe ATS configure space.
The new lenght of PRQ is calculated by summing up the value of Outstanding Page Request Capacity register across all devices where Page Requests are supported on the real PR-capable platfrom (Intel Sapphire Rapids). The result is round to the nearest higher power of 2.
The PRQ length is also double sized as the VT-d IOMMU driver only updates the Page Request Queue Head Register (PQH_REG) after processing the entire queue.
--- include/linux/intel-svm.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
include/linux/intel-svm.h +1 -1
@@ -9,7 +9,7 @@#define __INTEL_SVM_H__/* Page Request Queue depth */ -#define PRQ_ORDER 2 +#define PRQ_ORDER 4#define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x20)#define PRQ_DEPTH ((0x1000 << PRQ_ORDER) >> 5)
Signature · 6 lines
2.25.1 _______________________________________________ iommu mailing list [email protected] https://lists.linuxfoundation.org/mailman/listinfo/iommu
RE: [PATCH 3/3] iommu/vt-d: Size Page Request Queue to avoid overflow condition
Quoted 10 lines
From: Lu Baolu <[email protected]>
Sent: Saturday, April 16, 2022 8:31 PMPRQ overflow may cause I/O throughput congestion, resulting in unnecessary degradation of IO performance. Appropriately increasing the length of PRQ can greatly reduce the occurrence of PRQ overflow. The count of maximum page requests that can be generated in parallel by a PCIe device is statically defined in the Outstanding Page Request Capacity field of the PCIe ATS configure space.
Quoted 4 lines
The new lenght of PRQ is calculated by summing up the value of Outstanding Page Request Capacity register across all devices where Page Requests are supported on the real PR-capable platfrom (Intel Sapphire Rapids). The result is round to the nearest higher power of 2.
The actual requirement is usage and platform specific. What about doubling the default size and also provide an option for admin to tune?
Quoted 25 lines
The PRQ length is also double sized as the VT-d IOMMU driver only updates the Page Request Queue Head Register (
PQH_REG) after processing the entire
queue.Signed-off-by Lu Baolu <[email protected]>--- include/linux/intel-svm.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)include/linux/intel-svm.h +1
-2@@ -9,7 +9,7 @@#define __INTEL_SVM_H__/* Page Request Queue depth */ -#define PRQ_ORDER 2 +#define PRQ_ORDER 4#define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x20)#define PRQ_DEPTH ((0x1000 << PRQ_ORDER) >> 5) --
Quoted 1 line
2.25.1
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
Lu Baolu
maintainerRe: [PATCH 3/3] iommu/vt-d: Size Page Request Queue to avoid overflow condition
Quoted 10 lines · Kevin
Quoted 10 lines
From: Lu Baolu<[email protected]>
Sent: Saturday, April 16, 2022 8:31 PMPRQ overflow may cause I/O throughput congestion, resulting in unnecessary degradation of IO performance. Appropriately increasing the length of PRQ can greatly reduce the occurrence of PRQ overflow. The count of maximum page requests that can be generated in parallel by a PCIe device is statically defined in the Outstanding Page Request Capacity field of the PCIe ATS configure space.
Quoted 7 lines
Quoted 4 lines
The new lenght of PRQ is calculated by summing up the value of Outstanding Page Request Capacity register across all devices where Page Requests are supported on the real PR-capable platfrom (Intel Sapphire Rapids). The result is round to the nearest higher power of 2.
The actual requirement is usage and platform specific. What about doubling the default size and also provide an option for admin to tune?
I also have this in my list to provide an opt-in interface for user. But I don't want to include it in this series as this is only for small tuning. The opt-in interface looks more like a new feature. :-)
Best regards,
baolu