Skip to contentSkip to conversation

iommu/vt-d: boost the mapping process

PATCH In review Longpeng(Mike) 2 participants 6 messages Oct 8, 2021

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

v3 cover iommu/vt-d: boost the mapping process

Hi guys,

We found that the __domain_mapping() would take too long when the memory region is too large, we try to make it faster in this patchset. The performance number can be found in PATCH 2, please review when you free, thanks.

Changes v2 -> v3:
- make first_pte_in_page() neater [Baolu]
- remove meaningless BUG_ON() in __domain_mapping() [Baolu]

Changes v1 -> v2:
- Fix compile warning on i386 [Baolu]

Longpeng(Mike) (2):
iommu/vt-d: convert the return type of first_pte_in_page to bool iommu/vt-d: avoid duplicated removing in __domain_mapping

 drivers/iommu/intel/iommu.c | 11 ++++++-----
 include/linux/intel-iommu.h | 10 ++++++++--
 2 files changed, 14 insertions(+), 7 deletions(-)
Signature · 1 line
1.8.3.1
#1 lore raw

PATCH v3 1/2 iommu/vt-d: convert the return type of first_pte_in_page to bool

first_pte_in_page() returns boolean value, so let's convert its return type to bool. In addition, use 'IS_ALIGNED' to make the code more readable and neater.

---
 include/linux/intel-iommu.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
include/linux/intel-iommu.h +2 -2
index 05a65eb..9bcabc7 100644
@@ -708,9 +708,9 @@ static inline bool dma_pte_superpage(struct dma_pte *pte)
 	return (pte->val & DMA_PTE_LARGE_PAGE);
 }
 
-static inline int first_pte_in_page(struct dma_pte *pte)
+static inline bool first_pte_in_page(struct dma_pte *pte)
 {
-	return !((unsigned long)pte & ~VTD_PAGE_MASK);
+	return IS_ALIGNED((unsigned long)pte, VTD_PAGE_SIZE);
 }
 
 extern struct dmar_drhd_unit * dmar_find_matched_drhd_unit(struct pci_dev *dev);
Signature · 1 line
1.8.3.1
#2 lore raw

PATCH v3 2/2 iommu/vt-d: avoid duplicated removing in __domain_mapping

In reply to Longpeng(Mike)

__domain_mapping() always removes the pages in the range from 'iov_pfn' to 'end_pfn', but the 'end_pfn' is always the last pfn of the range that the caller wants to map.

This would introduce too many duplicated removing and leads the map operation take too long, for example:

Map iova=0x100000,nr_pages=0x7d61800

    iov_pfn: 0x100000, end_pfn: 0x7e617ff
    iov_pfn: 0x140000, end_pfn: 0x7e617ff
    iov_pfn: 0x180000, end_pfn: 0x7e617ff
    iov_pfn: 0x1c0000, end_pfn: 0x7e617ff
    iov_pfn: 0x200000, end_pfn: 0x7e617ff
    ...

it takes about 50ms in total.

We can reduce the cost by recalculate the 'end_pfn' and limit it to the boundary of the end of this pte page.

Map iova=0x100000,nr_pages=0x7d61800

    iov_pfn: 0x100000, end_pfn: 0x13ffff
    iov_pfn: 0x140000, end_pfn: 0x17ffff
    iov_pfn: 0x180000, end_pfn: 0x1bffff
    iov_pfn: 0x1c0000, end_pfn: 0x1fffff
    iov_pfn: 0x200000, end_pfn: 0x23ffff
    ...

it only need 9ms now.

---
 drivers/iommu/intel/iommu.c | 11 ++++++-----
 include/linux/intel-iommu.h |  6 ++++++
 2 files changed, 12 insertions(+), 5 deletions(-)
drivers/iommu/intel/iommu.c +6 -5
index d75f59a..46edae6 100644
@@ -2354,12 +2354,17 @@ static void switch_to_super_page(struct dmar_domain *domain,
 				return -ENOMEM;
 			first_pte = pte;
 
+			lvl_pages = lvl_to_nr_pages(largepage_lvl);
+
 			/* It is large page*/
 			if (largepage_lvl > 1) {
 				unsigned long end_pfn;
+				unsigned long pages_to_remove;
 
 				pteval |= DMA_PTE_LARGE_PAGE;
-				end_pfn = ((iov_pfn + nr_pages) & level_mask(largepage_lvl)) - 1;
+				pages_to_remove = min_t(unsigned long, nr_pages,
+							nr_pte_to_next_page(pte) * lvl_pages);
+				end_pfn = iov_pfn + pages_to_remove - 1;
 				switch_to_super_page(domain, iov_pfn, end_pfn, largepage_lvl);
 			} else {
 				pteval &= ~(uint64_t)DMA_PTE_LARGE_PAGE;
@@ -2381,10 +2386,6 @@ static void switch_to_super_page(struct dmar_domain *domain,
 			WARN_ON(1);
 		}
 
-		lvl_pages = lvl_to_nr_pages(largepage_lvl);
-
-		BUG_ON(nr_pages < lvl_pages);
-
 		nr_pages -= lvl_pages;
 		iov_pfn += lvl_pages;
 		phys_pfn += lvl_pages;
include/linux/intel-iommu.h +6 -0
index 9bcabc7..b29b2a3 100644
@@ -713,6 +713,12 @@ static inline bool first_pte_in_page(struct dma_pte *pte)
 	return IS_ALIGNED((unsigned long)pte, VTD_PAGE_SIZE);
 }
 
+static inline int nr_pte_to_next_page(struct dma_pte *pte)
+{
+	return first_pte_in_page(pte) ? BIT_ULL(VTD_STRIDE_SHIFT) :
+		(struct dma_pte *)ALIGN((unsigned long)pte, VTD_PAGE_SIZE) - pte;
+}
+
 extern struct dmar_drhd_unit * dmar_find_matched_drhd_unit(struct pci_dev *dev);
 extern int dmar_find_matched_atsr_unit(struct pci_dev *dev);
 
Signature · 1 line
1.8.3.1
maintainer
#3 lore raw

Re: [PATCH v3 2/2] iommu/vt-d: avoid duplicated removing in __domain_mapping

Quoted 73 lines · Longpeng(Mike)

__domain_mapping() always removes the pages in the range from 'iov_pfn' to 'end_pfn', but the 'end_pfn' is always the last pfn of the range that the caller wants to map.

This would introduce too many duplicated removing and leads the map operation take too long, for example:

   Map iova=0x100000,nr_pages=0x7d61800
     iov_pfn: 0x100000, end_pfn: 0x7e617ff
     iov_pfn: 0x140000, end_pfn: 0x7e617ff
     iov_pfn: 0x180000, end_pfn: 0x7e617ff
     iov_pfn: 0x1c0000, end_pfn: 0x7e617ff
     iov_pfn: 0x200000, end_pfn: 0x7e617ff
     ...
   it takes about 50ms in total.

We can reduce the cost by recalculate the 'end_pfn' and limit it to the boundary of the end of this pte page.

   Map iova=0x100000,nr_pages=0x7d61800
     iov_pfn: 0x100000, end_pfn: 0x13ffff
     iov_pfn: 0x140000, end_pfn: 0x17ffff
     iov_pfn: 0x180000, end_pfn: 0x1bffff
     iov_pfn: 0x1c0000, end_pfn: 0x1fffff
     iov_pfn: 0x200000, end_pfn: 0x23ffff
     ...
   it only need 9ms now.
---
  drivers/iommu/intel/iommu.c | 11 ++++++-----
  include/linux/intel-iommu.h |  6 ++++++
  2 files changed, 12 insertions(+), 5 deletions(-)
drivers/iommu/intel/iommu.c +6 -5
index d75f59a..46edae6 100644
@@ -2354,12 +2354,17 @@ static void switch_to_super_page(struct dmar_domain *domain,
  				return -ENOMEM;
  			first_pte = pte;
 
+			lvl_pages = lvl_to_nr_pages(largepage_lvl);
+
  			/* It is large page*/
  			if (largepage_lvl > 1) {
  				unsigned long end_pfn;
+				unsigned long pages_to_remove;
 
  				pteval |= DMA_PTE_LARGE_PAGE;
-				end_pfn = ((iov_pfn + nr_pages) & level_mask(largepage_lvl)) - 1;
+				pages_to_remove = min_t(unsigned long, nr_pages,
+							nr_pte_to_next_page(pte) * lvl_pages);
+				end_pfn = iov_pfn + pages_to_remove - 1;
  				switch_to_super_page(domain, iov_pfn, end_pfn, largepage_lvl);
  			} else {
  				pteval &= ~(uint64_t)DMA_PTE_LARGE_PAGE;
@@ -2381,10 +2386,6 @@ static void switch_to_super_page(struct dmar_domain *domain,
  			WARN_ON(1);
  		}
 
-		lvl_pages = lvl_to_nr_pages(largepage_lvl);
-
-		BUG_ON(nr_pages < lvl_pages);
-
  		nr_pages -= lvl_pages;
  		iov_pfn += lvl_pages;
  		phys_pfn += lvl_pages;
include/linux/intel-iommu.h +0 -0
index 9bcabc7..b29b2a3 100644

include/linux/intel-iommu.h static inline bool first_pte_in_page(struct dma_pte *pte) · line 719

Quoted 7 lines
include/linux/intel-iommu.h +4 -0
@@ -713,6 +713,12 @@ static inline bool first_pte_in_page(struct dma_pte *pte)
  	return IS_ALIGNED((unsigned long)pte, VTD_PAGE_SIZE);
  }
 
+static inline int nr_pte_to_next_page(struct dma_pte *pte)
+{
+	return first_pte_in_page(pte) ? BIT_ULL(VTD_STRIDE_SHIFT) :
+		(struct dma_pte *)ALIGN((unsigned long)pte, VTD_PAGE_SIZE) - pte;

We should make it like this to avoid the 0day warning:

	(struct dma_pte *)(uintptr_t)VTD_PAGE_ALIGN((unsigned long)pte) - pte;

Can you please test this line of change? No need to send a new version. I will handle it if it passes your test.

include/linux/intel-iommu.h static inline bool first_pte_in_page(struct dma_pte *pte) · line 724

Quoted 6 lines
include/linux/intel-iommu.h +2 -0
+}
+
  extern struct dmar_drhd_unit * dmar_find_matched_drhd_unit(struct pci_dev *dev);
  extern int dmar_find_matched_atsr_unit(struct pci_dev *dev);
 

Best regards,
baolu

maintainer
#4 lore raw

Re: [PATCH v3 2/2] iommu/vt-d: avoid duplicated removing in __domain_mapping

Quoted 85 lines · Lu Baolu
Quoted 73 lines · Longpeng(Mike)

__domain_mapping() always removes the pages in the range from 'iov_pfn' to 'end_pfn', but the 'end_pfn' is always the last pfn of the range that the caller wants to map.

This would introduce too many duplicated removing and leads the map operation take too long, for example:

   Map iova=0x100000,nr_pages=0x7d61800
     iov_pfn: 0x100000, end_pfn: 0x7e617ff
     iov_pfn: 0x140000, end_pfn: 0x7e617ff
     iov_pfn: 0x180000, end_pfn: 0x7e617ff
     iov_pfn: 0x1c0000, end_pfn: 0x7e617ff
     iov_pfn: 0x200000, end_pfn: 0x7e617ff
     ...
   it takes about 50ms in total.

We can reduce the cost by recalculate the 'end_pfn' and limit it to the boundary of the end of this pte page.

   Map iova=0x100000,nr_pages=0x7d61800
     iov_pfn: 0x100000, end_pfn: 0x13ffff
     iov_pfn: 0x140000, end_pfn: 0x17ffff
     iov_pfn: 0x180000, end_pfn: 0x1bffff
     iov_pfn: 0x1c0000, end_pfn: 0x1fffff
     iov_pfn: 0x200000, end_pfn: 0x23ffff
     ...
   it only need 9ms now.

---
  drivers/iommu/intel/iommu.c | 11 ++++++-----
  include/linux/intel-iommu.h |  6 ++++++

  2 files changed, 12 insertions(+), 5 deletions(-)
drivers/iommu/intel/iommu.c +6 -5
index d75f59a..46edae6 100644
@@ -2354,12 +2354,17 @@ static void switch_to_super_page(struct 
 dmar_domain *domain,
                   return -ENOMEM;
               first_pte = pte;
+            lvl_pages = lvl_to_nr_pages(largepage_lvl);
+
               /* It is large page*/
               if (largepage_lvl > 1) {
                   unsigned long end_pfn;
+                unsigned long pages_to_remove;
                   pteval |= DMA_PTE_LARGE_PAGE;
-                end_pfn = ((iov_pfn + nr_pages) & 
 level_mask(largepage_lvl)) - 1;
+                pages_to_remove = min_t(unsigned long, nr_pages,
+                            nr_pte_to_next_page(pte) * lvl_pages);
+                end_pfn = iov_pfn + pages_to_remove - 1;
                   switch_to_super_page(domain, iov_pfn, end_pfn, largepage_lvl);
               } else {
                   pteval &= ~(uint64_t)DMA_PTE_LARGE_PAGE;
@@ -2381,10 +2386,6 @@ static void switch_to_super_page(struct 
 dmar_domain *domain,
               WARN_ON(1);
           }
-        lvl_pages = lvl_to_nr_pages(largepage_lvl);
-
-        BUG_ON(nr_pages < lvl_pages);
-
           nr_pages -= lvl_pages;
           iov_pfn += lvl_pages;
           phys_pfn += lvl_pages;
include/linux/intel-iommu.h +0 -0
index 9bcabc7..b29b2a3 100644

include/linux/intel-iommu.h static inline bool first_pte_in_page(struct · line 720

Quoted 8 lines
include/linux/intel-iommu.h +4 -0
@@ -713,6 +713,12 @@ static inline bool first_pte_in_page(struct 
 dma_pte *pte)
       return IS_ALIGNED((unsigned long)pte, VTD_PAGE_SIZE);
   }
+static inline int nr_pte_to_next_page(struct dma_pte *pte)
+{
+    return first_pte_in_page(pte) ? BIT_ULL(VTD_STRIDE_SHIFT) :
+        (struct dma_pte *)ALIGN((unsigned long)pte, VTD_PAGE_SIZE) - 
 pte;

We should make it like this to avoid the 0day warning:

Quoted 4 lines

    (struct dma_pte *)(uintptr_t)VTD_PAGE_ALIGN((unsigned long)pte) - pte;

Can you please test this line of change? No need to send a new version. I will handle it if it passes your test.

Just realized that ALIGN() has already done the type cast. Please ignore above comment. Sorry for the noise.

Best regards,
baolu

#5 lore raw

RE: [PATCH v3 2/2] iommu/vt-d: avoid duplicated removing in __domain_mapping

Quoted 100 lines

-----Original Message-----
From: Lu Baolu [mailto:[email protected]]
Sent: Friday, October 8, 2021 10:44 AM
To: Longpeng (Mike, Cloud Infrastructure Service Product Dept.) <[email protected]>; [email protected]; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]; Gonglei (Arei) <[email protected]>
Subject: Re: [PATCH v3 2/2] iommu/vt-d: avoid duplicated removing in __domain_mapping

Quoted 86 lines · Lu Baolu
Quoted 74 lines · Longpeng(Mike)

__domain_mapping() always removes the pages in the range from 'iov_pfn' to 'end_pfn', but the 'end_pfn' is always the last pfn of the range that the caller wants to map.

This would introduce too many duplicated removing and leads the map operation take too long, for example:

   Map iova=0x100000,nr_pages=0x7d61800
     iov_pfn: 0x100000, end_pfn: 0x7e617ff
     iov_pfn: 0x140000, end_pfn: 0x7e617ff
     iov_pfn: 0x180000, end_pfn: 0x7e617ff
     iov_pfn: 0x1c0000, end_pfn: 0x7e617ff
     iov_pfn: 0x200000, end_pfn: 0x7e617ff
     ...
   it takes about 50ms in total.

We can reduce the cost by recalculate the 'end_pfn' and limit it to the boundary of the end of this pte page.

   Map iova=0x100000,nr_pages=0x7d61800
     iov_pfn: 0x100000, end_pfn: 0x13ffff
     iov_pfn: 0x140000, end_pfn: 0x17ffff
     iov_pfn: 0x180000, end_pfn: 0x1bffff
     iov_pfn: 0x1c0000, end_pfn: 0x1fffff
     iov_pfn: 0x200000, end_pfn: 0x23ffff
     ...
   it only need 9ms now.

---
  drivers/iommu/intel/iommu.c | 11 ++++++-----
  include/linux/intel-iommu.h |  6 ++++++

  2 files changed, 12 insertions(+), 5 deletions(-)
drivers/iommu/intel/iommu.c +6 -1
index d75f59a..46edae6 100644
@@ -2354,12 +2354,17 @@ static void switch_to_super_page(struct
 dmar_domain *domain,
                   return -ENOMEM;
               first_pte = pte;
+            lvl_pages = lvl_to_nr_pages(largepage_lvl);
+
               /* It is large page*/
               if (largepage_lvl > 1) {
                   unsigned long end_pfn;
+                unsigned long pages_to_remove;
                   pteval |= DMA_PTE_LARGE_PAGE;
-                end_pfn = ((iov_pfn + nr_pages) &
 level_mask(largepage_lvl)) - 1;
+                pages_to_remove = min_t(unsigned long, nr_pages,
+                            nr_pte_to_next_page(pte) * lvl_pages);
+                end_pfn = iov_pfn + pages_to_remove - 1;
                   switch_to_super_page(domain, iov_pfn, end_pfn,
 largepage_lvl);
               } else {

                  pteval &= ~(uint64_t)DMA_PTE_LARGE_PAGE;

drivers/iommu/intel/iommu.c +0 -4
@@ -2381,10 +2386,6 @@ static void switch_to_super_page(struct
 dmar_domain *domain,
               WARN_ON(1);
           }
-        lvl_pages = lvl_to_nr_pages(largepage_lvl);
-
-        BUG_ON(nr_pages < lvl_pages);
-
           nr_pages -= lvl_pages;
           iov_pfn += lvl_pages;
           phys_pfn += lvl_pages;
include/linux/intel-iommu.h +0 -0
index 9bcabc7..b29b2a3 100644

include/linux/intel-iommu.h static inline bool first_pte_in_page(struct · line 720

Quoted 8 lines
include/linux/intel-iommu.h +4 -0
@@ -713,6 +713,12 @@ static inline bool first_pte_in_page(struct
 dma_pte *pte)
       return IS_ALIGNED((unsigned long)pte, VTD_PAGE_SIZE);
   }
+static inline int nr_pte_to_next_page(struct dma_pte *pte)
+{
+    return first_pte_in_page(pte) ? BIT_ULL(VTD_STRIDE_SHIFT) :
+        (struct dma_pte *)ALIGN((unsigned long)pte, VTD_PAGE_SIZE) -
 pte;

We should make it like this to avoid the 0day warning:

Quoted 8 lines
Quoted 4 lines

    (struct dma_pte *)(uintptr_t)VTD_PAGE_ALIGN((unsigned long)pte) - pte;

Can you please test this line of change? No need to send a new version. I will handle it if it passes your test.

Just realized that ALIGN() has already done the type cast. Please ignore above comment. Sorry for the noise.

Hi Baolu,

Our testing is completed, no compile warning on both X86 64bit and 32bit arch, and the system is working as expected.

Please add:

Quoted 2 lines

Best regards,
baolu