Skip to contentSkip to conversation

VT-d minor fixes

PATCH Superseded CLEMENT MATHIEU--DRIF 2 participants 9 messages Jul 9, 2024

v6 of 7 Superseded Latest is v7 →

lore
Share & tools
lore
#0 lore raw

v6 cover VT-d minor fixes

From: Clément Mathieu--Drif <[email protected]>

Various fixes for VT-d

This series contains fixes that will be necessary when adding in-guest (fully emulated) SVM support.

v6
intel_iommu: fix type of the mask field in VTDIOTLBPageInvInfo:
- Add 'Fixes' tag

v5
intel_iommu: fix FRCD construction macro:
- Remove empty line after 'Fixes'

intel_iommu: fix type of the mask field in VTDIOTLBPageInvInfo:
- Edit commit message after comment from Yi

intel_iommu: make types match:
- Edit commit message (s/"make types match"/"make type match"/)

v4
- Move declarations of VTD_FRCD_PV and VTD_FRCD_PP
- intel_iommu: make types match:
- edit commit message to explain that we are not fixing a bug
- intel_iommu: fix type of the mask field in VTDIOTLBPageInvInfo
- edit commit message

v3
FRCD construction macro :
- Longer sha1 for the 'Fixes' tag
- Add '.' at the end of the sentence

Make types match :
- Split into 2 patches (one for the fix and one for type matching)

Remove patch for wait descriptor handling (will be in the PRI series)

v2
Make commit author consistent

Clément Mathieu--Drif (4):
intel_iommu: fix FRCD construction macro
intel_iommu: move VTD_FRCD_PV and VTD_FRCD_PP declarations
intel_iommu: fix type of the mask field in VTDIOTLBPageInvInfo
intel_iommu: make type match

 hw/i386/intel_iommu.c          | 2 +-
 hw/i386/intel_iommu_internal.h | 6 +++---
 2 files changed, 4 insertions(+), 4 deletions(-)
Signature · 1 line
2.45.2
#1 lore raw

PATCH v6 1/4 intel_iommu: fix FRCD construction macro

From: Clément Mathieu--Drif <[email protected]>

The constant must be unsigned, otherwise the two's complement overrides the other fields when a PASID is present.

Fixes 1b2b12376c8a ("intel-iommu: PASID support")
---
 hw/i386/intel_iommu_internal.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
hw/i386/intel_iommu_internal.h +1 -1
index f8cf99bddf..cbc4030031 100644
@@ -267,7 +267,7 @@
 /* For the low 64-bit of 128-bit */
 #define VTD_FRCD_FI(val)        ((val) & ~0xfffULL)
 #define VTD_FRCD_PV(val)        (((val) & 0xffffULL) << 40)
-#define VTD_FRCD_PP(val)        (((val) & 0x1) << 31)
+#define VTD_FRCD_PP(val)        (((val) & 0x1ULL) << 31)
 #define VTD_FRCD_IR_IDX(val)    (((val) & 0xffffULL) << 48)
 
 /* DMA Remapping Fault Conditions */
Signature · 1 line
2.45.2
#2 lore raw

PATCH v6 2/4 intel_iommu: move VTD_FRCD_PV and VTD_FRCD_PP declarations

In reply to CLEMENT MATHIEU--DRIF

From: Clément Mathieu--Drif <[email protected]>

These 2 macros are for high 64-bit of the FRCD registers. Declarations have to be moved accordingly.

---
 hw/i386/intel_iommu_internal.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
hw/i386/intel_iommu_internal.h +2 -2
index cbc4030031..faea23e8d6 100644
@@ -264,10 +264,10 @@
 #define VTD_FRCD_FR(val)        (((val) & 0xffULL) << 32)
 #define VTD_FRCD_SID_MASK       0xffffULL
 #define VTD_FRCD_SID(val)       ((val) & VTD_FRCD_SID_MASK)
-/* For the low 64-bit of 128-bit */
-#define VTD_FRCD_FI(val)        ((val) & ~0xfffULL)
 #define VTD_FRCD_PV(val)        (((val) & 0xffffULL) << 40)
 #define VTD_FRCD_PP(val)        (((val) & 0x1ULL) << 31)
+/* For the low 64-bit of 128-bit */
+#define VTD_FRCD_FI(val)        ((val) & ~0xfffULL)
 #define VTD_FRCD_IR_IDX(val)    (((val) & 0xffffULL) << 48)
 
 /* DMA Remapping Fault Conditions */
Signature · 1 line
2.45.2
Reviewed-by on v6 2/4
#3 lore raw

Re: [PATCH v6 2/4] intel_iommu: move VTD_FRCD_PV and VTD_FRCD_PP declarations

Quoted 10 lines · CLEMENT MATHIEU--DRIF

From: Clément Mathieu--Drif <[email protected]>

These 2 macros are for high 64-bit of the FRCD registers. Declarations have to be moved accordingly.

---
  hw/i386/intel_iommu_internal.h | 4 ++--
  1 file changed, 2 insertions(+), 2 deletions(-)
Quoted 18 lines
hw/i386/intel_iommu_internal.h +2 -2
index cbc4030031..faea23e8d6 100644
@@ -264,10 +264,10 @@
  #define VTD_FRCD_FR(val)        (((val) & 0xffULL) << 32)
  #define VTD_FRCD_SID_MASK       0xffffULL
  #define VTD_FRCD_SID(val)       ((val) & VTD_FRCD_SID_MASK)
-/* For the low 64-bit of 128-bit */
-#define VTD_FRCD_FI(val)        ((val) & ~0xfffULL)
  #define VTD_FRCD_PV(val)        (((val) & 0xffffULL) << 40)
  #define VTD_FRCD_PP(val)        (((val) & 0x1ULL) << 31)
+/* For the low 64-bit of 128-bit */
+#define VTD_FRCD_FI(val)        ((val) & ~0xfffULL)
  #define VTD_FRCD_IR_IDX(val)    (((val) & 0xffffULL) << 48)
 
  /* DMA Remapping Fault Conditions */
Signature · 2 lines
Regards,
Yi Liu
#4 lore raw

PATCH v6 3/4 intel_iommu: fix type of the mask field in VTDIOTLBPageInvInfo

In reply to CLEMENT MATHIEU--DRIF

From: Clément Mathieu--Drif <[email protected]>

The mask we are trying to store into VTDIOTLBPageInvInfo.mask might not fit in an uint8_t. Use uint64_t to avoid overflows.

Per the below code, it can overflow as am can be larger than 8 according to the CH 6.5.2.3 IOTLB Invalidate. And you may want a fix tag as well.

info.mask = ~((1 << am) - 1);

CH 6.5.2.3 IOTLB Invalidate

Address Mask (AM): For page-selective-within-domain invalidations, the Address Mask specifies the number of low order bits of the ADDR field that must be masked for the invalidation operation. This field enables software to request invalidation of contiguous mappings for size-aligned regions. Refer to Table 19 for encodings of this field. When invalidating a large-page translation, software must use the appropriate Address Mask value (0 for 4KByte page, 9 for 2-MByte page, and 18 for 1-GByte page). Hardware implementations report the maximum supported address mask value through the Capability register.

Fixes b5a280c00840 ("intel-iommu: add IOTLB using hash table")
---
 hw/i386/intel_iommu_internal.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
hw/i386/intel_iommu_internal.h +1 -1
index faea23e8d6..5f32c36943 100644
@@ -436,7 +436,7 @@ struct VTDIOTLBPageInvInfo {
     uint16_t domain_id;
     uint32_t pasid;
     uint64_t addr;
-    uint8_t mask;
+    uint64_t mask;
 };
 typedef struct VTDIOTLBPageInvInfo VTDIOTLBPageInvInfo;
 
Signature · 1 line
2.45.2
Reviewed-by on v6 3/4
#5 lore raw

Re: [PATCH v6 3/4] intel_iommu: fix type of the mask field in VTDIOTLBPageInvInfo

Quoted 7 lines · CLEMENT MATHIEU--DRIF

From: Clément Mathieu--Drif <[email protected]>

The mask we are trying to store into VTDIOTLBPageInvInfo.mask might not fit in an uint8_t. Use uint64_t to avoid overflows.

Per the below code, it can overflow as am can be larger than 8 according to the CH 6.5.2.3 IOTLB Invalidate. And you may want a fix tag as well.

no need to paste the comment here. Especially the last sentence. :) Just highlight that overflow is possible.

Quoted 15 lines

info.mask = ~((1 << am) - 1);

CH 6.5.2.3 IOTLB Invalidate

Address Mask (AM): For page-selective-within-domain invalidations, the Address Mask specifies the number of low order bits of the ADDR field that must be masked for the invalidation operation. This field enables software to request invalidation of contiguous mappings for size-aligned regions. Refer to Table 19 for encodings of this field. When invalidating a large-page translation, software must use the appropriate Address Mask value (0 for 4KByte page, 9 for 2-MByte page, and 18 for 1-GByte page). Hardware implementations report the maximum supported address mask value through the Capability register.

Quoted 1 line
Fixes b5a280c00840 ("intel-iommu: add IOTLB using hash table")

looks like this is a very old commit. The VTD_MAMV is set as 9 at that time. If guest sets am=9 in the PSI iotlb invalidation, there is overflow. Just curious why there is no bug reported here. Have you encountered any issue before this commit? Anyhow, looks good to me. Remember to refine the commit message a bit.

Quoted 19 lines
---
  hw/i386/intel_iommu_internal.h | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
hw/i386/intel_iommu_internal.h +1 -1
index faea23e8d6..5f32c36943 100644
@@ -436,7 +436,7 @@ struct VTDIOTLBPageInvInfo {
      uint16_t domain_id;
      uint32_t pasid;
      uint64_t addr;
-    uint8_t mask;
+    uint64_t mask;
  };
  typedef struct VTDIOTLBPageInvInfo VTDIOTLBPageInvInfo;
 
Signature · 2 lines
Regards,
Yi Liu
#6 lore raw

Re: [PATCH v6 3/4] intel_iommu: fix type of the mask field in VTDIOTLBPageInvInfo

Quoted 32 lines · Yi Liu

Caution: External email. Do not open attachments or click links, unless this email comes from a known sender and you know the content is safe.

Quoted 7 lines · CLEMENT MATHIEU--DRIF

From: Clément Mathieu--Drif <[email protected]>

The mask we are trying to store into VTDIOTLBPageInvInfo.mask might not fit in an uint8_t. Use uint64_t to avoid overflows.

Per the below code, it can overflow as am can be larger than 8 according to the CH 6.5.2.3 IOTLB Invalidate. And you may want a fix tag as well.

no need to paste the comment here. Especially the last sentence. :) Just highlight that overflow is possible.

Quoted 15 lines

info.mask = ~((1 << am) - 1);

CH 6.5.2.3 IOTLB Invalidate

Address Mask (AM): For page-selective-within-domain invalidations, the Address Mask specifies the number of low order bits of the ADDR field that must be masked for the invalidation operation. This field enables software to request invalidation of contiguous mappings for size-aligned regions. Refer to Table 19 for encodings of this field. When invalidating a large-page translation, software must use the appropriate Address Mask value (0 for 4KByte page, 9 for 2-MByte page, and 18 for 1-GByte page). Hardware implementations report the maximum supported address mask value through the Capability register.

Quoted 8 lines
Quoted 1 line
Fixes b5a280c00840 ("intel-iommu: add IOTLB using hash table")

looks like this is a very old commit. The VTD_MAMV is set as 9 at that time. If guest sets am=9 in the PSI iotlb invalidation, there is overflow.
Just curious why there is no bug reported here. Have you encountered any issue before this commit? Anyhow, looks good to me. Remember to refine the commit message a bit.

I detected weird values in gdb when I started to implement pasid-based invalidations for SVM

Quoted 27 lines
Quoted 20 lines

---
  hw/i386/intel_iommu_internal.h | 2 +-

  1 file changed, 1 insertion(+), 1 deletion(-)
hw/i386/intel_iommu_internal.h +1 -1
index faea23e8d6..5f32c36943 100644
@@ -436,7 +436,7 @@ struct VTDIOTLBPageInvInfo {
       uint16_t domain_id;
       uint32_t pasid;
       uint64_t addr;
-    uint8_t mask;
+    uint64_t mask;
   };
   typedef struct VTDIOTLBPageInvInfo VTDIOTLBPageInvInfo;
 
Signature · 2 lines
Regards,
Yi Liu
#7 lore raw

Re: [PATCH v6 3/4] intel_iommu: fix type of the mask field in VTDIOTLBPageInvInfo

In reply to Yi Liu
Quoted 8 lines · Yi Liu

Caution: External email. Do not open attachments or click links, unless this email comes from a known sender and you know the content is safe.

Quoted 2 lines · CLEMENT MATHIEU--DRIF

From: Clément Mathieu--Drif <[email protected]>

Quoted 8 lines
Quoted 5 lines

The mask we are trying to store into VTDIOTLBPageInvInfo.mask might not fit in an uint8_t. Use uint64_t to avoid overflows.

Per the below code, it can overflow as am can be larger than 8 according to the CH 6.5.2.3 IOTLB Invalidate. And you may want a fix tag as well.

no need to paste the comment here. Especially the last sentence. :) Just highlight that overflow is possible.

Oops, sorry, will send v7

Quoted 51 lines
Quoted 16 lines

info.mask = ~((1 << am) - 1);

CH 6.5.2.3 IOTLB Invalidate

Address Mask (AM): For page-selective-within-domain invalidations, the Address Mask specifies the number of low order bits of the ADDR field that must be masked for the invalidation operation. This field enables software to request invalidation of contiguous mappings for size-aligned regions. Refer to Table 19 for encodings of this field. When invalidating a large-page translation, software must use the appropriate Address Mask value (0 for 4KByte page, 9 for 2-MByte page, and 18 for 1-GByte page). Hardware implementations report the maximum supported address mask value through the Capability register.

Quoted 1 line
Fixes b5a280c00840 ("intel-iommu: add IOTLB using hash table")

looks like this is a very old commit. The VTD_MAMV is set as 9 at that time. If guest sets am=9 in the PSI iotlb invalidation, there is overflow.
Just curious why there is no bug reported here. Have you encountered any issue before this commit? Anyhow, looks good to me. Remember to refine the commit message a bit.

Quoted 20 lines

---
  hw/i386/intel_iommu_internal.h | 2 +-

  1 file changed, 1 insertion(+), 1 deletion(-)
hw/i386/intel_iommu_internal.h +1 -1
index faea23e8d6..5f32c36943 100644
@@ -436,7 +436,7 @@ struct VTDIOTLBPageInvInfo {
       uint16_t domain_id;
       uint32_t pasid;
       uint64_t addr;
-    uint8_t mask;
+    uint64_t mask;
   };
   typedef struct VTDIOTLBPageInvInfo VTDIOTLBPageInvInfo;
 
Signature · 2 lines
Regards,
Yi Liu
#8 lore raw

PATCH v6 4/4 intel_iommu: make type match

In reply to CLEMENT MATHIEU--DRIF

From: Clément Mathieu--Drif <[email protected]>

The 'level' field in vtd_iotlb_key is an unsigned integer. We don't need to store level as an int in vtd_lookup_iotlb.

This is not an issue by itself, but using unsigned here seems cleaner.

---
 hw/i386/intel_iommu.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
hw/i386/intel_iommu.c +1 -1
index 37c21a0aec..be0cb39b5c 100644
@@ -358,7 +358,7 @@ static VTDIOTLBEntry *vtd_lookup_iotlb(IntelIOMMUState *s, uint16_t source_id,
 {
     struct vtd_iotlb_key key;
     VTDIOTLBEntry *entry;
-    int level;
+    unsigned level;
 
     for (level = VTD_SL_PT_LEVEL; level < VTD_SL_PML4_LEVEL; level++) {
         key.gfn = vtd_get_iotlb_gfn(addr, level);
Signature · 1 line
2.45.2