[PATCH] arm64: Enable vmalloc-huge with ptdump

arm64 disables vmalloc-huge when kernel page table dumping is enabled,
because an intermediate table may be removed, potentially causing the
ptdump code to dereference an invalid address. We want to be able to
analyze block vs page mappings for kernel mappings with ptdump, so to
enable vmalloc-huge with ptdump, synchronize between page table removal in
pmd_free_pte_page()/pud_free_pmd_page() and ptdump pagetable walking. We
use mmap_read_lock and not write lock because we don't need to synchronize
between two different vm_structs; two vmalloc objects running this same
code path will point to different page tables, hence there is no race. 

Signed-off-by: Dev Jain <dev.jain@arm.com>
---
 arch/arm64/include/asm/vmalloc.h | 6 ++----
 arch/arm64/mm/mmu.c              | 7 +++++++
 2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/include/asm/vmalloc.h b/arch/arm64/include/asm/vmalloc.h
index 38fafffe699f..28b7173d8693 100644
--- a/arch/arm64/include/asm/vmalloc.h
+++ b/arch/arm64/include/asm/vmalloc.h

@@ -12,15 +12,13 @@ static inline bool arch_vmap_pud_supported(pgprot_t prot)
 	/*
 	 * SW table walks can't handle removal of intermediate entries.
 	 */
-	return pud_sect_supported() &&
-	       !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+	return pud_sect_supported();
 }
 
 #define arch_vmap_pmd_supported arch_vmap_pmd_supported
 static inline bool arch_vmap_pmd_supported(pgprot_t prot)
 {
-	/* See arch_vmap_pud_supported() */
-	return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+	return true;
 }
 
 #endif

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index ea6695d53fb9..798cebd9e147 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1261,7 +1261,11 @@ int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
 	}
 
 	table = pte_offset_kernel(pmdp, addr);
+
+	/* Synchronize against ptdump_walk_pgd() */
+	mmap_read_lock(&init_mm);
 	pmd_clear(pmdp);
+	mmap_read_unlock(&init_mm);
 	__flush_tlb_kernel_pgtable(addr);
 	pte_free_kernel(NULL, table);
 	return 1;

@@ -1289,7 +1293,10 @@ int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
 		pmd_free_pte_page(pmdp, next);
 	} while (pmdp++, next += PMD_SIZE, next != end);
 
+	/* Synchronize against ptdump_walk_pgd() */
+	mmap_read_lock(&init_mm);
 	pud_clear(pudp);
+	mmap_read_unlock(&init_mm);
 	__flush_tlb_kernel_pgtable(addr);
 	pmd_free(NULL, table);
 	return 1;

-- 
2.30.2

On 30/05/25 1:50 pm, Dev Jain wrote:

arm64 disables vmalloc-huge when kernel page table dumping is enabled,
because an intermediate table may be removed, potentially causing the
ptdump code to dereference an invalid address. We want to be able to
analyze block vs page mappings for kernel mappings with ptdump, so to
enable vmalloc-huge with ptdump, synchronize between page table removal in
pmd_free_pte_page()/pud_free_pmd_page() and ptdump pagetable walking. We
use mmap_read_lock and not write lock because we don't need to synchronize
between two different vm_structs; two vmalloc objects running this same
code path will point to different page tables, hence there is no race.

I mean, there *is* a race, but there is no problem :)

On 30/05/2025 09:20, Dev Jain wrote:

So this works because ptdump_walk_pgd() takes the write_lock (which is mutually
exclusive with any read_lock holders) for the duration of the table walk, so it
will either consistently see the pgtables before or after this removal. It will
never disappear during the walk, correct?

I guess there is a risk of this showing up as contention with other init_mm
write_lock holders. But I expect that pmd_free_pte_page()/pud_free_pmd_page()
are called sufficiently rarely that the risk is very small. Let's fix any perf
problem if/when we see it.

 	__flush_tlb_kernel_pgtable(addr);

And the tlbi doesn't need to be serialized because there is no security issue.
The walker can be trusted to only dereference memory that it sees as it walks
the pgtable (obviously).

Hmm, so pud_free_pmd_page() is now going to cause us to acquire and release the
(upto) lock 513 times (for a 4K kernel). I wonder if there is an argument for
clearing the pud first (under the lock), then the pmds can all be cleared
without a lock, since the walker won't be able to see the pmds once the pud is
cleared.

Thanks,
Ryan

 	__flush_tlb_kernel_pgtable(addr);
 	pmd_free(NULL, table);
 	return 1;

On 5/30/25 14:10, Ryan Roberts wrote:

On 30/05/2025 09:20, Dev Jain wrote:

quoted

arm64 disables vmalloc-huge when kernel page table dumping is enabled,
because an intermediate table may be removed, potentially causing the
ptdump code to dereference an invalid address. We want to be able to
analyze block vs page mappings for kernel mappings with ptdump, so to
enable vmalloc-huge with ptdump, synchronize between page table removal in
pmd_free_pte_page()/pud_free_pmd_page() and ptdump pagetable walking. We
use mmap_read_lock and not write lock because we don't need to synchronize
between two different vm_structs; two vmalloc objects running this same
code path will point to different page tables, hence there is no race. 

Signed-off-by: Dev Jain <dev.jain@arm.com>
---
 arch/arm64/include/asm/vmalloc.h | 6 ++----
 arch/arm64/mm/mmu.c              | 7 +++++++
 2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/include/asm/vmalloc.h b/arch/arm64/include/asm/vmalloc.h
index 38fafffe699f..28b7173d8693 100644
--- a/arch/arm64/include/asm/vmalloc.h
+++ b/arch/arm64/include/asm/vmalloc.h

@@ -12,15 +12,13 @@ static inline bool arch_vmap_pud_supported(pgprot_t prot)
 	/*
 	 * SW table walks can't handle removal of intermediate entries.
 	 */
-	return pud_sect_supported() &&
-	       !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+	return pud_sect_supported();
 }
 
 #define arch_vmap_pmd_supported arch_vmap_pmd_supported
 static inline bool arch_vmap_pmd_supported(pgprot_t prot)
 {
-	/* See arch_vmap_pud_supported() */
-	return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+	return true;
 }
 
 #endif

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index ea6695d53fb9..798cebd9e147 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1261,7 +1261,11 @@ int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
 	}
 
 	table = pte_offset_kernel(pmdp, addr);
+
+	/* Synchronize against ptdump_walk_pgd() */
+	mmap_read_lock(&init_mm);
 	pmd_clear(pmdp);
+	mmap_read_unlock(&init_mm);

So this works because ptdump_walk_pgd() takes the write_lock (which is mutually
exclusive with any read_lock holders) for the duration of the table walk, so it
will either consistently see the pgtables before or after this removal. It will
never disappear during the walk, correct?

Agreed.

I guess there is a risk of this showing up as contention with other init_mm
write_lock holders. But I expect that pmd_free_pte_page()/pud_free_pmd_page()
are called sufficiently rarely that the risk is very small. Let's fix any perf
problem if/when we see it.

Checking against CONFIG_PTDUMP_DEBUGFS being enabled is simple enough without much
cost. So why not make this conditional only for scenarios, where this read lock is
really required. Something like

--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1293,11 +1293,15 @@ int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
                pmd_free_pte_page(pmdp, next);
        } while (pmdp++, next += PMD_SIZE, next != end);
 
-       /* Synchronize against ptdump_walk_pgd() */
-       mmap_read_lock(&init_mm);
-       pud_clear(pudp);
-       mmap_read_unlock(&init_mm);
        __flush_tlb_kernel_pgtable(addr);
+       if (IS_ENABLED(CONFIG_PTDUMP_DEBUGFS)) {
+               /* Synchronize against ptdump_walk_pgd() */
+               mmap_read_lock(&init_mm);
+               pud_clear(pudp);
+               mmap_read_unlock(&init_mm);
+       } else {
+               pud_clear(pudp);
+       }
        pmd_free(NULL, table);
        return 1;
 }

quoted

 	__flush_tlb_kernel_pgtable(addr);

And the tlbi doesn't need to be serialized because there is no security issue.
The walker can be trusted to only dereference memory that it sees as it walks
the pgtable (obviously).

Agreed.

quoted

 	pte_free_kernel(NULL, table);
 	return 1;

@@ -1289,7 +1293,10 @@ int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
 		pmd_free_pte_page(pmdp, next);
 	} while (pmdp++, next += PMD_SIZE, next != end);
 
+	/* Synchronize against ptdump_walk_pgd() */
+	mmap_read_lock(&init_mm);
 	pud_clear(pudp);
+	mmap_read_unlock(&init_mm);

Hmm, so pud_free_pmd_page() is now going to cause us to acquire and release the
(upto) lock 513 times (for a 4K kernel). I wonder if there is an argument for
clearing the pud first (under the lock), then the pmds can all be cleared
without a lock, since the walker won't be able to see the pmds once the pud is
cleared.

Makes sense if pud_free_pmd_page() would have been the only caller but seems like
vmap_try_huge_pmd() calls pmd_free_pte_page() directly as well.

Thanks,
Ryan

quoted

 	__flush_tlb_kernel_pgtable(addr);
 	pmd_free(NULL, table);
 	return 1;

On 30/05/25 2:10 pm, Ryan Roberts wrote:

On 30/05/2025 09:20, Dev Jain wrote:

quoted

arm64 disables vmalloc-huge when kernel page table dumping is enabled,
because an intermediate table may be removed, potentially causing the
ptdump code to dereference an invalid address. We want to be able to
analyze block vs page mappings for kernel mappings with ptdump, so to
enable vmalloc-huge with ptdump, synchronize between page table removal in
pmd_free_pte_page()/pud_free_pmd_page() and ptdump pagetable walking. We
use mmap_read_lock and not write lock because we don't need to synchronize
between two different vm_structs; two vmalloc objects running this same
code path will point to different page tables, hence there is no race.

My "correction" from race->no problem was incorrect after all :) There will
be no race too since the vm_struct object has exclusive access to whatever
table it is clearing.

quoted

Signed-off-by: Dev Jain <dev.jain@arm.com>
---
  arch/arm64/include/asm/vmalloc.h | 6 ++----
  arch/arm64/mm/mmu.c              | 7 +++++++
  2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/include/asm/vmalloc.h b/arch/arm64/include/asm/vmalloc.h
index 38fafffe699f..28b7173d8693 100644
--- a/arch/arm64/include/asm/vmalloc.h
+++ b/arch/arm64/include/asm/vmalloc.h

@@ -12,15 +12,13 @@ static inline bool arch_vmap_pud_supported(pgprot_t prot)
  	/*
  	 * SW table walks can't handle removal of intermediate entries.
  	 */
-	return pud_sect_supported() &&
-	       !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+	return pud_sect_supported();
  }
  
  #define arch_vmap_pmd_supported arch_vmap_pmd_supported
  static inline bool arch_vmap_pmd_supported(pgprot_t prot)
  {
-	/* See arch_vmap_pud_supported() */
-	return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+	return true;
  }
  
  #endif

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index ea6695d53fb9..798cebd9e147 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1261,7 +1261,11 @@ int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
  	}
  
  	table = pte_offset_kernel(pmdp, addr);
+
+	/* Synchronize against ptdump_walk_pgd() */
+	mmap_read_lock(&init_mm);
  	pmd_clear(pmdp);
+	mmap_read_unlock(&init_mm);

So this works because ptdump_walk_pgd() takes the write_lock (which is mutually
exclusive with any read_lock holders) for the duration of the table walk, so it
will either consistently see the pgtables before or after this removal. It will
never disappear during the walk, correct?

I guess there is a risk of this showing up as contention with other init_mm
write_lock holders. But I expect that pmd_free_pte_page()/pud_free_pmd_page()
are called sufficiently rarely that the risk is very small. Let's fix any perf
problem if/when we see it.

We can avoid all of that by my initial approach - to wrap the lock around CONFIG_PTDUMP_DEBUGFS.
I don't have a strong opinion, just putting it out there.

quoted

  	__flush_tlb_kernel_pgtable(addr);

And the tlbi doesn't need to be serialized because there is no security issue.
The walker can be trusted to only dereference memory that it sees as it walks
the pgtable (obviously).

quoted

  	pte_free_kernel(NULL, table);
  	return 1;

@@ -1289,7 +1293,10 @@ int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
  		pmd_free_pte_page(pmdp, next);
  	} while (pmdp++, next += PMD_SIZE, next != end);
  
+	/* Synchronize against ptdump_walk_pgd() */
+	mmap_read_lock(&init_mm);
  	pud_clear(pudp);
+	mmap_read_unlock(&init_mm);

Hmm, so pud_free_pmd_page() is now going to cause us to acquire and release the
(upto) lock 513 times (for a 4K kernel). I wonder if there is an argument for
clearing the pud first (under the lock), then the pmds can all be cleared
without a lock, since the walker won't be able to see the pmds once the pud is
cleared.

Yes, we can isolate the PMD table in case the caller of pmd_free_pte_page is
pud_free_pmd_page. In this case, vm_struct_1 has exclusive access to the entire
pmd page, hence no race will occur. But, in case of vmap_try_huge_pmd() being the
caller, we cannot drop the locks around pmd_free_pte_page. So we can have something
like

#ifdef CONFIG_PTDUMP_DEBUGFS
static inline void ptdump_synchronize_lock(bool flag)
{
	if (flag)
		mmap_read_lock(&init_mm);
}

and pass false when the caller is pud_free_pmd_page.

Thanks,
Ryan

quoted

  	__flush_tlb_kernel_pgtable(addr);
  	pmd_free(NULL, table);
  	return 1;

On 30/05/2025 10:14, Dev Jain wrote:

On 30/05/25 2:10 pm, Ryan Roberts wrote:

quoted

On 30/05/2025 09:20, Dev Jain wrote:

quoted

arm64 disables vmalloc-huge when kernel page table dumping is enabled,
because an intermediate table may be removed, potentially causing the
ptdump code to dereference an invalid address. We want to be able to
analyze block vs page mappings for kernel mappings with ptdump, so to
enable vmalloc-huge with ptdump, synchronize between page table removal in
pmd_free_pte_page()/pud_free_pmd_page() and ptdump pagetable walking. We
use mmap_read_lock and not write lock because we don't need to synchronize
between two different vm_structs; two vmalloc objects running this same
code path will point to different page tables, hence there is no race.

My "correction" from race->no problem was incorrect after all :) There will
be no race too since the vm_struct object has exclusive access to whatever
table it is clearing.

quoted

quoted

Signed-off-by: Dev Jain <dev.jain@arm.com>
---
  arch/arm64/include/asm/vmalloc.h | 6 ++----
  arch/arm64/mm/mmu.c              | 7 +++++++
  2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/include/asm/vmalloc.h b/arch/arm64/include/asm/vmalloc.h
index 38fafffe699f..28b7173d8693 100644
--- a/arch/arm64/include/asm/vmalloc.h
+++ b/arch/arm64/include/asm/vmalloc.h

@@ -12,15 +12,13 @@ static inline bool arch_vmap_pud_supported(pgprot_t prot)

      /*
       * SW table walks can't handle removal of intermediate entries.
       */
-    return pud_sect_supported() &&
-           !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return pud_sect_supported();
  }
    #define arch_vmap_pmd_supported arch_vmap_pmd_supported
  static inline bool arch_vmap_pmd_supported(pgprot_t prot)
  {
-    /* See arch_vmap_pud_supported() */
-    return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return true;
  }
    #endif

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index ea6695d53fb9..798cebd9e147 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1261,7 +1261,11 @@ int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)

      }
        table = pte_offset_kernel(pmdp, addr);
+
+    /* Synchronize against ptdump_walk_pgd() */
+    mmap_read_lock(&init_mm);
      pmd_clear(pmdp);
+    mmap_read_unlock(&init_mm);

So this works because ptdump_walk_pgd() takes the write_lock (which is mutually
exclusive with any read_lock holders) for the duration of the table walk, so it
will either consistently see the pgtables before or after this removal. It will
never disappear during the walk, correct?

I guess there is a risk of this showing up as contention with other init_mm
write_lock holders. But I expect that pmd_free_pte_page()/pud_free_pmd_page()
are called sufficiently rarely that the risk is very small. Let's fix any perf
problem if/when we see it.

We can avoid all of that by my initial approach - to wrap the lock around
CONFIG_PTDUMP_DEBUGFS.
I don't have a strong opinion, just putting it out there.

quoted

quoted

      __flush_tlb_kernel_pgtable(addr);

And the tlbi doesn't need to be serialized because there is no security issue.
The walker can be trusted to only dereference memory that it sees as it walks
the pgtable (obviously).

quoted

      pte_free_kernel(NULL, table);
      return 1;

@@ -1289,7 +1293,10 @@ int pud_free_pmd_page(pud_t *pudp, unsigned long addr)

          pmd_free_pte_page(pmdp, next);
      } while (pmdp++, next += PMD_SIZE, next != end);
  +    /* Synchronize against ptdump_walk_pgd() */
+    mmap_read_lock(&init_mm);
      pud_clear(pudp);
+    mmap_read_unlock(&init_mm);

Hmm, so pud_free_pmd_page() is now going to cause us to acquire and release the
(upto) lock 513 times (for a 4K kernel). I wonder if there is an argument for
clearing the pud first (under the lock), then the pmds can all be cleared
without a lock, since the walker won't be able to see the pmds once the pud is
cleared.

Yes, we can isolate the PMD table in case the caller of pmd_free_pte_page is
pud_free_pmd_page. In this case, vm_struct_1 has exclusive access to the entire
pmd page, hence no race will occur. But, in case of vmap_try_huge_pmd() being the
caller, we cannot drop the locks around pmd_free_pte_page. So we can have something
like

#ifdef CONFIG_PTDUMP_DEBUGFS
static inline void ptdump_synchronize_lock(bool flag)
{
    if (flag)
        mmap_read_lock(&init_mm);
}

and pass false when the caller is pud_free_pmd_page.

of something like this? (completely untested):

---8<---

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index 8fcf59ba39db..1f3a922167e4 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1267,7 +1267,7 @@ int pmd_clear_huge(pmd_t *pmdp)
        return 1;
 }

-int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
+static int __pmd_free_pte_page(pmd_t *pmdp, unsigned long addr, bool lock)
 {
        pte_t *table;
        pmd_t pmd;

@@ -1280,12 +1280,23 @@ int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
        }

        table = pte_offset_kernel(pmdp, addr);
+
+       if (lock)
+               mmap_read_lock(&init_mm);
        pmd_clear(pmdp);
+       if (lock)
+               mmap_read_unlock(&init_mm);
+
        __flush_tlb_kernel_pgtable(addr);
        pte_free_kernel(NULL, table);
        return 1;
 }

+int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
+{
+       return __pmd_free_pte_page(pmdp, addr, true);
+}
+
 int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
 {
        pmd_t *table;

@@ -1300,15 +1311,19 @@ int pud_free_pmd_page(pud_t *pudp, unsigned long addr)
                return 1;
        }

+       /* Synchronize against ptdump_walk_pgd() */
+       mmap_read_lock(&init_mm);
+       pud_clear(pudp);
+       mmap_read_unlock(&init_mm);
+
        table = pmd_offset(pudp, addr);
        pmdp = table;
        next = addr;
        end = addr + PUD_SIZE;
        do {
-               pmd_free_pte_page(pmdp, next);
+               __pmd_free_pte_page(pmdp, next, false);
        } while (pmdp++, next += PMD_SIZE, next != end);

-       pud_clear(pudp);
        __flush_tlb_kernel_pgtable(addr);
        pmd_free(NULL, table);
        return 1;
---8<---

quoted

Thanks,
Ryan

quoted

      __flush_tlb_kernel_pgtable(addr);
      pmd_free(NULL, table);
      return 1;

On 30/05/25 3:17 pm, Ryan Roberts wrote:

This looks good too! Although would like to first decide on whether we want
to wrap around CONFIG_PTDUMP_DEBUGFS.

quoted

quoted

Thanks,
Ryan

quoted

       __flush_tlb_kernel_pgtable(addr);
       pmd_free(NULL, table);
       return 1;

On 30/05/2025 10:14, Dev Jain wrote:

On 30/05/25 2:10 pm, Ryan Roberts wrote:

quoted

On 30/05/2025 09:20, Dev Jain wrote:

quoted

arm64 disables vmalloc-huge when kernel page table dumping is enabled,
because an intermediate table may be removed, potentially causing the
ptdump code to dereference an invalid address. We want to be able to
analyze block vs page mappings for kernel mappings with ptdump, so to
enable vmalloc-huge with ptdump, synchronize between page table removal in
pmd_free_pte_page()/pud_free_pmd_page() and ptdump pagetable walking. We
use mmap_read_lock and not write lock because we don't need to synchronize
between two different vm_structs; two vmalloc objects running this same
code path will point to different page tables, hence there is no race.

My "correction" from race->no problem was incorrect after all :) There will
be no race too since the vm_struct object has exclusive access to whatever
table it is clearing.

quoted

quoted

Signed-off-by: Dev Jain <dev.jain@arm.com>
---
  arch/arm64/include/asm/vmalloc.h | 6 ++----
  arch/arm64/mm/mmu.c              | 7 +++++++
  2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/include/asm/vmalloc.h b/arch/arm64/include/asm/vmalloc.h
index 38fafffe699f..28b7173d8693 100644
--- a/arch/arm64/include/asm/vmalloc.h
+++ b/arch/arm64/include/asm/vmalloc.h

@@ -12,15 +12,13 @@ static inline bool arch_vmap_pud_supported(pgprot_t prot)

      /*
       * SW table walks can't handle removal of intermediate entries.
       */
-    return pud_sect_supported() &&
-           !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return pud_sect_supported();
  }
    #define arch_vmap_pmd_supported arch_vmap_pmd_supported
  static inline bool arch_vmap_pmd_supported(pgprot_t prot)
  {
-    /* See arch_vmap_pud_supported() */
-    return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return true;
  }
    #endif

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index ea6695d53fb9..798cebd9e147 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1261,7 +1261,11 @@ int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)

      }
        table = pte_offset_kernel(pmdp, addr);
+
+    /* Synchronize against ptdump_walk_pgd() */
+    mmap_read_lock(&init_mm);
      pmd_clear(pmdp);
+    mmap_read_unlock(&init_mm);

So this works because ptdump_walk_pgd() takes the write_lock (which is mutually
exclusive with any read_lock holders) for the duration of the table walk, so it
will either consistently see the pgtables before or after this removal. It will
never disappear during the walk, correct?

I guess there is a risk of this showing up as contention with other init_mm
write_lock holders. But I expect that pmd_free_pte_page()/pud_free_pmd_page()
are called sufficiently rarely that the risk is very small. Let's fix any perf
problem if/when we see it.

We can avoid all of that by my initial approach - to wrap the lock around
CONFIG_PTDUMP_DEBUGFS.
I don't have a strong opinion, just putting it out there.

(I wrote then failed to send earlier):

It's ugly though. Personally I'd prefer to keep it simple unless we have clear
evidence that its needed. I was just laying out my justification for not needing
to doing the conditional wrapping in this comment.

quoted

quoted

      __flush_tlb_kernel_pgtable(addr);

And the tlbi doesn't need to be serialized because there is no security issue.
The walker can be trusted to only dereference memory that it sees as it walks
the pgtable (obviously).

quoted

      pte_free_kernel(NULL, table);
      return 1;

@@ -1289,7 +1293,10 @@ int pud_free_pmd_page(pud_t *pudp, unsigned long addr)

          pmd_free_pte_page(pmdp, next);
      } while (pmdp++, next += PMD_SIZE, next != end);
  +    /* Synchronize against ptdump_walk_pgd() */
+    mmap_read_lock(&init_mm);
      pud_clear(pudp);
+    mmap_read_unlock(&init_mm);

Hmm, so pud_free_pmd_page() is now going to cause us to acquire and release the
(upto) lock 513 times (for a 4K kernel). I wonder if there is an argument for
clearing the pud first (under the lock), then the pmds can all be cleared
without a lock, since the walker won't be able to see the pmds once the pud is
cleared.

Yes, we can isolate the PMD table in case the caller of pmd_free_pte_page is
pud_free_pmd_page. In this case, vm_struct_1 has exclusive access to the entire
pmd page, hence no race will occur. But, in case of vmap_try_huge_pmd() being the
caller, we cannot drop the locks around pmd_free_pte_page. So we can have something
like

#ifdef CONFIG_PTDUMP_DEBUGFS
static inline void ptdump_synchronize_lock(bool flag)
{
    if (flag)
        mmap_read_lock(&init_mm);
}

and pass false when the caller is pud_free_pmd_page.

quoted

Thanks,
Ryan

quoted

      __flush_tlb_kernel_pgtable(addr);
      pmd_free(NULL, table);
      return 1;

On Fri, May 30, 2025 at 12:50:40PM +0100, Ryan Roberts wrote:

On 30/05/2025 10:14, Dev Jain wrote:

quoted

On 30/05/25 2:10 pm, Ryan Roberts wrote:

quoted

On 30/05/2025 09:20, Dev Jain wrote:

quoted

arm64 disables vmalloc-huge when kernel page table dumping is enabled,
because an intermediate table may be removed, potentially causing the
ptdump code to dereference an invalid address. We want to be able to
analyze block vs page mappings for kernel mappings with ptdump, so to
enable vmalloc-huge with ptdump, synchronize between page table removal in
pmd_free_pte_page()/pud_free_pmd_page() and ptdump pagetable walking. We
use mmap_read_lock and not write lock because we don't need to synchronize
between two different vm_structs; two vmalloc objects running this same
code path will point to different page tables, hence there is no race.

My "correction" from race->no problem was incorrect after all :) There will
be no race too since the vm_struct object has exclusive access to whatever
table it is clearing.

quoted

quoted

Signed-off-by: Dev Jain <dev.jain@arm.com>
---
  arch/arm64/include/asm/vmalloc.h | 6 ++----
  arch/arm64/mm/mmu.c              | 7 +++++++
  2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/include/asm/vmalloc.h b/arch/arm64/include/asm/vmalloc.h
index 38fafffe699f..28b7173d8693 100644
--- a/arch/arm64/include/asm/vmalloc.h
+++ b/arch/arm64/include/asm/vmalloc.h

@@ -12,15 +12,13 @@ static inline bool arch_vmap_pud_supported(pgprot_t prot)

      /*
       * SW table walks can't handle removal of intermediate entries.
       */
-    return pud_sect_supported() &&
-           !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return pud_sect_supported();
  }
    #define arch_vmap_pmd_supported arch_vmap_pmd_supported
  static inline bool arch_vmap_pmd_supported(pgprot_t prot)
  {
-    /* See arch_vmap_pud_supported() */
-    return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return true;
  }
    #endif

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index ea6695d53fb9..798cebd9e147 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1261,7 +1261,11 @@ int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)

      }
        table = pte_offset_kernel(pmdp, addr);
+
+    /* Synchronize against ptdump_walk_pgd() */
+    mmap_read_lock(&init_mm);
      pmd_clear(pmdp);
+    mmap_read_unlock(&init_mm);

So this works because ptdump_walk_pgd() takes the write_lock (which is mutually
exclusive with any read_lock holders) for the duration of the table walk, so it
will either consistently see the pgtables before or after this removal. It will
never disappear during the walk, correct?

I guess there is a risk of this showing up as contention with other init_mm
write_lock holders. But I expect that pmd_free_pte_page()/pud_free_pmd_page()
are called sufficiently rarely that the risk is very small. Let's fix any perf
problem if/when we see it.

We can avoid all of that by my initial approach - to wrap the lock around
CONFIG_PTDUMP_DEBUGFS.
I don't have a strong opinion, just putting it out there.

(I wrote then failed to send earlier):

It's ugly though. Personally I'd prefer to keep it simple unless we have clear
evidence that its needed. I was just laying out my justification for not needing
to doing the conditional wrapping in this comment.

I really don't think we should be adding unconditional locking overhead
to core mm routines purely to facilitate a rarely used debug option.

Instead, can we either adopt something like the RCU-like walk used by
fast GUP or stick the locking behind a static key that's only enabled
when a ptdump walk is in progress (a bit like how
hugetlb_vmemmap_optimize_folio() manipulates hugetlb_optimize_vmemmap_key)?

Will

On 30/05/2025 13:35, Will Deacon wrote:

On Fri, May 30, 2025 at 12:50:40PM +0100, Ryan Roberts wrote:

quoted

On 30/05/2025 10:14, Dev Jain wrote:

quoted

On 30/05/25 2:10 pm, Ryan Roberts wrote:

quoted

On 30/05/2025 09:20, Dev Jain wrote:

quoted

arm64 disables vmalloc-huge when kernel page table dumping is enabled,
because an intermediate table may be removed, potentially causing the
ptdump code to dereference an invalid address. We want to be able to
analyze block vs page mappings for kernel mappings with ptdump, so to
enable vmalloc-huge with ptdump, synchronize between page table removal in
pmd_free_pte_page()/pud_free_pmd_page() and ptdump pagetable walking. We
use mmap_read_lock and not write lock because we don't need to synchronize
between two different vm_structs; two vmalloc objects running this same
code path will point to different page tables, hence there is no race.

My "correction" from race->no problem was incorrect after all :) There will
be no race too since the vm_struct object has exclusive access to whatever
table it is clearing.

quoted

quoted

Signed-off-by: Dev Jain <dev.jain@arm.com>
---
  arch/arm64/include/asm/vmalloc.h | 6 ++----
  arch/arm64/mm/mmu.c              | 7 +++++++
  2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/include/asm/vmalloc.h b/arch/arm64/include/asm/vmalloc.h
index 38fafffe699f..28b7173d8693 100644
--- a/arch/arm64/include/asm/vmalloc.h
+++ b/arch/arm64/include/asm/vmalloc.h

@@ -12,15 +12,13 @@ static inline bool arch_vmap_pud_supported(pgprot_t prot)

      /*
       * SW table walks can't handle removal of intermediate entries.
       */
-    return pud_sect_supported() &&
-           !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return pud_sect_supported();
  }
    #define arch_vmap_pmd_supported arch_vmap_pmd_supported
  static inline bool arch_vmap_pmd_supported(pgprot_t prot)
  {
-    /* See arch_vmap_pud_supported() */
-    return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return true;
  }
    #endif

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index ea6695d53fb9..798cebd9e147 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1261,7 +1261,11 @@ int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)

      }
        table = pte_offset_kernel(pmdp, addr);
+
+    /* Synchronize against ptdump_walk_pgd() */
+    mmap_read_lock(&init_mm);
      pmd_clear(pmdp);
+    mmap_read_unlock(&init_mm);

So this works because ptdump_walk_pgd() takes the write_lock (which is mutually
exclusive with any read_lock holders) for the duration of the table walk, so it
will either consistently see the pgtables before or after this removal. It will
never disappear during the walk, correct?

I guess there is a risk of this showing up as contention with other init_mm
write_lock holders. But I expect that pmd_free_pte_page()/pud_free_pmd_page()
are called sufficiently rarely that the risk is very small. Let's fix any perf
problem if/when we see it.

We can avoid all of that by my initial approach - to wrap the lock around
CONFIG_PTDUMP_DEBUGFS.
I don't have a strong opinion, just putting it out there.

(I wrote then failed to send earlier):

It's ugly though. Personally I'd prefer to keep it simple unless we have clear
evidence that its needed. I was just laying out my justification for not needing
to doing the conditional wrapping in this comment.

I really don't think we should be adding unconditional locking overhead
to core mm routines purely to facilitate a rarely used debug option.

Instead, can we either adopt something like the RCU-like walk used by
fast GUP or stick the locking behind a static key that's only enabled
when a ptdump walk is in progress (a bit like how
hugetlb_vmemmap_optimize_folio() manipulates hugetlb_optimize_vmemmap_key)?

My sense is that the static key will be less effort and can be contained fully
in arm64. I think we would need to enable the key around the call to
ptdump_walk_pgd() in both ptdump_walk() and ptdump_check_wx(). Then where Dev is
currently taking the read lock, that would be contingent on the key being
enabled and the unlock would be contingent on having taken the lock.

Does that sound like an acceptable approach?

Will

On Fri, May 30, 2025 at 02:11:36PM +0100, Ryan Roberts wrote:

On 30/05/2025 13:35, Will Deacon wrote:

quoted

On Fri, May 30, 2025 at 12:50:40PM +0100, Ryan Roberts wrote:

quoted

On 30/05/2025 10:14, Dev Jain wrote:

quoted

On 30/05/25 2:10 pm, Ryan Roberts wrote:

quoted

On 30/05/2025 09:20, Dev Jain wrote:

quoted

arm64 disables vmalloc-huge when kernel page table dumping is enabled,
because an intermediate table may be removed, potentially causing the
ptdump code to dereference an invalid address. We want to be able to
analyze block vs page mappings for kernel mappings with ptdump, so to
enable vmalloc-huge with ptdump, synchronize between page table removal in
pmd_free_pte_page()/pud_free_pmd_page() and ptdump pagetable walking. We
use mmap_read_lock and not write lock because we don't need to synchronize
between two different vm_structs; two vmalloc objects running this same
code path will point to different page tables, hence there is no race.

My "correction" from race->no problem was incorrect after all :) There will
be no race too since the vm_struct object has exclusive access to whatever
table it is clearing.

quoted

quoted

Signed-off-by: Dev Jain <dev.jain@arm.com>
---
  arch/arm64/include/asm/vmalloc.h | 6 ++----
  arch/arm64/mm/mmu.c              | 7 +++++++
  2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/include/asm/vmalloc.h b/arch/arm64/include/asm/vmalloc.h
index 38fafffe699f..28b7173d8693 100644
--- a/arch/arm64/include/asm/vmalloc.h
+++ b/arch/arm64/include/asm/vmalloc.h

@@ -12,15 +12,13 @@ static inline bool arch_vmap_pud_supported(pgprot_t prot)

      /*
       * SW table walks can't handle removal of intermediate entries.
       */
-    return pud_sect_supported() &&
-           !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return pud_sect_supported();
  }
    #define arch_vmap_pmd_supported arch_vmap_pmd_supported
  static inline bool arch_vmap_pmd_supported(pgprot_t prot)
  {
-    /* See arch_vmap_pud_supported() */
-    return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return true;
  }
    #endif

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index ea6695d53fb9..798cebd9e147 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1261,7 +1261,11 @@ int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)

      }
        table = pte_offset_kernel(pmdp, addr);
+
+    /* Synchronize against ptdump_walk_pgd() */
+    mmap_read_lock(&init_mm);
      pmd_clear(pmdp);
+    mmap_read_unlock(&init_mm);

So this works because ptdump_walk_pgd() takes the write_lock (which is mutually
exclusive with any read_lock holders) for the duration of the table walk, so it
will either consistently see the pgtables before or after this removal. It will
never disappear during the walk, correct?

I guess there is a risk of this showing up as contention with other init_mm
write_lock holders. But I expect that pmd_free_pte_page()/pud_free_pmd_page()
are called sufficiently rarely that the risk is very small. Let's fix any perf
problem if/when we see it.

We can avoid all of that by my initial approach - to wrap the lock around
CONFIG_PTDUMP_DEBUGFS.
I don't have a strong opinion, just putting it out there.

(I wrote then failed to send earlier):

It's ugly though. Personally I'd prefer to keep it simple unless we have clear
evidence that its needed. I was just laying out my justification for not needing
to doing the conditional wrapping in this comment.

I really don't think we should be adding unconditional locking overhead
to core mm routines purely to facilitate a rarely used debug option.

Instead, can we either adopt something like the RCU-like walk used by
fast GUP or stick the locking behind a static key that's only enabled
when a ptdump walk is in progress (a bit like how
hugetlb_vmemmap_optimize_folio() manipulates hugetlb_optimize_vmemmap_key)?

My sense is that the static key will be less effort and can be contained fully
in arm64. I think we would need to enable the key around the call to
ptdump_walk_pgd() in both ptdump_walk() and ptdump_check_wx(). Then where Dev is
currently taking the read lock, that would be contingent on the key being
enabled and the unlock would be contingent on having taken the lock.

Does that sound like an acceptable approach?

Yup, and I think you'll probably need something like a synchronize_rcu()
when flipping the key to deal with any pre-existing page-table freers.

Will

On 30/05/2025 14:36, Will Deacon wrote:

On Fri, May 30, 2025 at 02:11:36PM +0100, Ryan Roberts wrote:

quoted

On 30/05/2025 13:35, Will Deacon wrote:

quoted

On Fri, May 30, 2025 at 12:50:40PM +0100, Ryan Roberts wrote:

quoted

On 30/05/2025 10:14, Dev Jain wrote:

quoted

On 30/05/25 2:10 pm, Ryan Roberts wrote:

quoted

On 30/05/2025 09:20, Dev Jain wrote:

quoted

arm64 disables vmalloc-huge when kernel page table dumping is enabled,
because an intermediate table may be removed, potentially causing the
ptdump code to dereference an invalid address. We want to be able to
analyze block vs page mappings for kernel mappings with ptdump, so to
enable vmalloc-huge with ptdump, synchronize between page table removal in
pmd_free_pte_page()/pud_free_pmd_page() and ptdump pagetable walking. We
use mmap_read_lock and not write lock because we don't need to synchronize
between two different vm_structs; two vmalloc objects running this same
code path will point to different page tables, hence there is no race.

My "correction" from race->no problem was incorrect after all :) There will
be no race too since the vm_struct object has exclusive access to whatever
table it is clearing.

quoted

quoted

Signed-off-by: Dev Jain <dev.jain@arm.com>
---
  arch/arm64/include/asm/vmalloc.h | 6 ++----
  arch/arm64/mm/mmu.c              | 7 +++++++
  2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/include/asm/vmalloc.h b/arch/arm64/include/asm/vmalloc.h
index 38fafffe699f..28b7173d8693 100644
--- a/arch/arm64/include/asm/vmalloc.h
+++ b/arch/arm64/include/asm/vmalloc.h

@@ -12,15 +12,13 @@ static inline bool arch_vmap_pud_supported(pgprot_t prot)

      /*
       * SW table walks can't handle removal of intermediate entries.
       */
-    return pud_sect_supported() &&
-           !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return pud_sect_supported();
  }
    #define arch_vmap_pmd_supported arch_vmap_pmd_supported
  static inline bool arch_vmap_pmd_supported(pgprot_t prot)
  {
-    /* See arch_vmap_pud_supported() */
-    return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return true;
  }
    #endif

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index ea6695d53fb9..798cebd9e147 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1261,7 +1261,11 @@ int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)

      }
        table = pte_offset_kernel(pmdp, addr);
+
+    /* Synchronize against ptdump_walk_pgd() */
+    mmap_read_lock(&init_mm);
      pmd_clear(pmdp);
+    mmap_read_unlock(&init_mm);

So this works because ptdump_walk_pgd() takes the write_lock (which is mutually
exclusive with any read_lock holders) for the duration of the table walk, so it
will either consistently see the pgtables before or after this removal. It will
never disappear during the walk, correct?

I guess there is a risk of this showing up as contention with other init_mm
write_lock holders. But I expect that pmd_free_pte_page()/pud_free_pmd_page()
are called sufficiently rarely that the risk is very small. Let's fix any perf
problem if/when we see it.

We can avoid all of that by my initial approach - to wrap the lock around
CONFIG_PTDUMP_DEBUGFS.
I don't have a strong opinion, just putting it out there.

(I wrote then failed to send earlier):

It's ugly though. Personally I'd prefer to keep it simple unless we have clear
evidence that its needed. I was just laying out my justification for not needing
to doing the conditional wrapping in this comment.

I really don't think we should be adding unconditional locking overhead
to core mm routines purely to facilitate a rarely used debug option.

Instead, can we either adopt something like the RCU-like walk used by
fast GUP or stick the locking behind a static key that's only enabled
when a ptdump walk is in progress (a bit like how
hugetlb_vmemmap_optimize_folio() manipulates hugetlb_optimize_vmemmap_key)?

My sense is that the static key will be less effort and can be contained fully
in arm64. I think we would need to enable the key around the call to
ptdump_walk_pgd() in both ptdump_walk() and ptdump_check_wx(). Then where Dev is
currently taking the read lock, that would be contingent on the key being
enabled and the unlock would be contingent on having taken the lock.

Does that sound like an acceptable approach?

Yup, and I think you'll probably need something like a synchronize_rcu()
when flipping the key to deal with any pre-existing page-table freers.

Yeah good point.

Will

On 30/05/25 7:06 pm, Will Deacon wrote:

On Fri, May 30, 2025 at 02:11:36PM +0100, Ryan Roberts wrote:

quoted

On 30/05/2025 13:35, Will Deacon wrote:

quoted

On Fri, May 30, 2025 at 12:50:40PM +0100, Ryan Roberts wrote:

quoted

On 30/05/2025 10:14, Dev Jain wrote:

quoted

On 30/05/25 2:10 pm, Ryan Roberts wrote:

quoted

On 30/05/2025 09:20, Dev Jain wrote:

quoted

arm64 disables vmalloc-huge when kernel page table dumping is enabled,
because an intermediate table may be removed, potentially causing the
ptdump code to dereference an invalid address. We want to be able to
analyze block vs page mappings for kernel mappings with ptdump, so to
enable vmalloc-huge with ptdump, synchronize between page table removal in
pmd_free_pte_page()/pud_free_pmd_page() and ptdump pagetable walking. We
use mmap_read_lock and not write lock because we don't need to synchronize
between two different vm_structs; two vmalloc objects running this same
code path will point to different page tables, hence there is no race.

My "correction" from race->no problem was incorrect after all :) There will
be no race too since the vm_struct object has exclusive access to whatever
table it is clearing.

quoted

quoted

Signed-off-by: Dev Jain <dev.jain@arm.com>
---
   arch/arm64/include/asm/vmalloc.h | 6 ++----
   arch/arm64/mm/mmu.c              | 7 +++++++
   2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/include/asm/vmalloc.h b/arch/arm64/include/asm/vmalloc.h
index 38fafffe699f..28b7173d8693 100644
--- a/arch/arm64/include/asm/vmalloc.h
+++ b/arch/arm64/include/asm/vmalloc.h

@@ -12,15 +12,13 @@ static inline bool arch_vmap_pud_supported(pgprot_t prot)
       /*
        * SW table walks can't handle removal of intermediate entries.
        */
-    return pud_sect_supported() &&
-           !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return pud_sect_supported();
   }
     #define arch_vmap_pmd_supported arch_vmap_pmd_supported
   static inline bool arch_vmap_pmd_supported(pgprot_t prot)
   {
-    /* See arch_vmap_pud_supported() */
-    return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return true;
   }
     #endif

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index ea6695d53fb9..798cebd9e147 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1261,7 +1261,11 @@ int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)
       }
         table = pte_offset_kernel(pmdp, addr);
+
+    /* Synchronize against ptdump_walk_pgd() */
+    mmap_read_lock(&init_mm);
       pmd_clear(pmdp);
+    mmap_read_unlock(&init_mm);

So this works because ptdump_walk_pgd() takes the write_lock (which is mutually
exclusive with any read_lock holders) for the duration of the table walk, so it
will either consistently see the pgtables before or after this removal. It will
never disappear during the walk, correct?

I guess there is a risk of this showing up as contention with other init_mm
write_lock holders. But I expect that pmd_free_pte_page()/pud_free_pmd_page()
are called sufficiently rarely that the risk is very small. Let's fix any perf
problem if/when we see it.

We can avoid all of that by my initial approach - to wrap the lock around
CONFIG_PTDUMP_DEBUGFS.
I don't have a strong opinion, just putting it out there.

(I wrote then failed to send earlier):

It's ugly though. Personally I'd prefer to keep it simple unless we have clear
evidence that its needed. I was just laying out my justification for not needing
to doing the conditional wrapping in this comment.

I really don't think we should be adding unconditional locking overhead
to core mm routines purely to facilitate a rarely used debug option.

Instead, can we either adopt something like the RCU-like walk used by
fast GUP or stick the locking behind a static key that's only enabled
when a ptdump walk is in progress (a bit like how
hugetlb_vmemmap_optimize_folio() manipulates hugetlb_optimize_vmemmap_key)?

My sense is that the static key will be less effort and can be contained fully
in arm64. I think we would need to enable the key around the call to
ptdump_walk_pgd() in both ptdump_walk() and ptdump_check_wx(). Then where Dev is
currently taking the read lock, that would be contingent on the key being
enabled and the unlock would be contingent on having taken the lock.

Does that sound like an acceptable approach?

Yup, and I think you'll probably need something like a synchronize_rcu()
when flipping the key to deal with any pre-existing page-table freers.

IIUC, you mean to say that we need to ensure any existing readers having
a reference to the isolated table in pmd_free_pte_page and friend, have exited.
But the problem is that we take an mmap_write_lock() around ptdump_walk_pgd() so
this is a sleepable code path.

Will

On 05/06/25 10:18 am, Dev Jain wrote:

On 30/05/25 7:06 pm, Will Deacon wrote:

quoted

On Fri, May 30, 2025 at 02:11:36PM +0100, Ryan Roberts wrote:

quoted

On 30/05/2025 13:35, Will Deacon wrote:

quoted

On Fri, May 30, 2025 at 12:50:40PM +0100, Ryan Roberts wrote:

quoted

On 30/05/2025 10:14, Dev Jain wrote:

quoted

On 30/05/25 2:10 pm, Ryan Roberts wrote:

quoted

On 30/05/2025 09:20, Dev Jain wrote:

quoted

arm64 disables vmalloc-huge when kernel page table dumping is enabled,
because an intermediate table may be removed, potentially causing the
ptdump code to dereference an invalid address. We want to be able to
analyze block vs page mappings for kernel mappings with ptdump, so to
enable vmalloc-huge with ptdump, synchronize between page table removal in
pmd_free_pte_page()/pud_free_pmd_page() and ptdump pagetable walking. We
use mmap_read_lock and not write lock because we don't need to synchronize
between two different vm_structs; two vmalloc objects running this same
code path will point to different page tables, hence there is no race.

My "correction" from race->no problem was incorrect after all :) There will
be no race too since the vm_struct object has exclusive access to whatever
table it is clearing.

quoted

quoted

Signed-off-by: Dev Jain <dev.jain@arm.com>
---
   arch/arm64/include/asm/vmalloc.h | 6 ++----
   arch/arm64/mm/mmu.c              | 7 +++++++
   2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/include/asm/vmalloc.h b/arch/arm64/include/asm/vmalloc.h
index 38fafffe699f..28b7173d8693 100644
--- a/arch/arm64/include/asm/vmalloc.h
+++ b/arch/arm64/include/asm/vmalloc.h

@@ -12,15 +12,13 @@ static inline bool arch_vmap_pud_supported(pgprot_t prot)

       /*
        * SW table walks can't handle removal of intermediate entries.
        */
-    return pud_sect_supported() &&
-           !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return pud_sect_supported();
   }
     #define arch_vmap_pmd_supported arch_vmap_pmd_supported
   static inline bool arch_vmap_pmd_supported(pgprot_t prot)
   {
-    /* See arch_vmap_pud_supported() */
-    return !IS_ENABLED(CONFIG_PTDUMP_DEBUGFS);
+    return true;
   }
     #endif

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index ea6695d53fb9..798cebd9e147 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c

@@ -1261,7 +1261,11 @@ int pmd_free_pte_page(pmd_t *pmdp, unsigned long addr)

       }
         table = pte_offset_kernel(pmdp, addr);
+
+    /* Synchronize against ptdump_walk_pgd() */
+    mmap_read_lock(&init_mm);
       pmd_clear(pmdp);
+    mmap_read_unlock(&init_mm);

So this works because ptdump_walk_pgd() takes the write_lock (which is mutually
exclusive with any read_lock holders) for the duration of the table walk, so it
will either consistently see the pgtables before or after this removal. It will
never disappear during the walk, correct?

I guess there is a risk of this showing up as contention with other init_mm
write_lock holders. But I expect that pmd_free_pte_page()/pud_free_pmd_page()
are called sufficiently rarely that the risk is very small. Let's fix any perf
problem if/when we see it.

We can avoid all of that by my initial approach - to wrap the lock around
CONFIG_PTDUMP_DEBUGFS.
I don't have a strong opinion, just putting it out there.

(I wrote then failed to send earlier):

It's ugly though. Personally I'd prefer to keep it simple unless we have clear
evidence that its needed. I was just laying out my justification for not needing
to doing the conditional wrapping in this comment.

I really don't think we should be adding unconditional locking overhead
to core mm routines purely to facilitate a rarely used debug option.

Instead, can we either adopt something like the RCU-like walk used by
fast GUP or stick the locking behind a static key that's only enabled
when a ptdump walk is in progress (a bit like how
hugetlb_vmemmap_optimize_folio() manipulates hugetlb_optimize_vmemmap_key)?

My sense is that the static key will be less effort and can be contained fully
in arm64. I think we would need to enable the key around the call to
ptdump_walk_pgd() in both ptdump_walk() and ptdump_check_wx(). Then where Dev is
currently taking the read lock, that would be contingent on the key being
enabled and the unlock would be contingent on having taken the lock.

Does that sound like an acceptable approach?

Yup, and I think you'll probably need something like a synchronize_rcu()
when flipping the key to deal with any pre-existing page-table freers.

IIUC, you mean to say that we need to ensure any existing readers having
a reference to the isolated table in pmd_free_pte_page and friend, have exited.
But the problem is that we take an mmap_write_lock() around ptdump_walk_pgd() so
this is a sleepable code path.

The mmap_write_lock around ptdump_walk_pgd() is definitely restrictive. I think that

was put because ptdump is rarely used, I think we could have done RCU-freeing of pagetables to

synchronize between ptdump and vmalloc pagetable lazy freeing/ hotunplug pagetable freeing.

quoted

Will

On Thu, Jun 05, 2025 at 01:46:01PM +0530, Dev Jain wrote:

On 05/06/25 10:18 am, Dev Jain wrote:

quoted

On 30/05/25 7:06 pm, Will Deacon wrote:

quoted

On Fri, May 30, 2025 at 02:11:36PM +0100, Ryan Roberts wrote:

quoted

On 30/05/2025 13:35, Will Deacon wrote:

quoted

I really don't think we should be adding unconditional
locking overhead
to core mm routines purely to facilitate a rarely used debug option.

Instead, can we either adopt something like the RCU-like walk used by
fast GUP or stick the locking behind a static key that's only enabled
when a ptdump walk is in progress (a bit like how
hugetlb_vmemmap_optimize_folio() manipulates
hugetlb_optimize_vmemmap_key)?

My sense is that the static key will be less effort and can be
contained fully
in arm64. I think we would need to enable the key around the call to
ptdump_walk_pgd() in both ptdump_walk() and ptdump_check_wx().
Then where Dev is
currently taking the read lock, that would be contingent on the
key being
enabled and the unlock would be contingent on having taken the lock.

Does that sound like an acceptable approach?

Yup, and I think you'll probably need something like a synchronize_rcu()
when flipping the key to deal with any pre-existing page-table freers.

IIUC, you mean to say that we need to ensure any existing readers having
a reference to the isolated table in pmd_free_pte_page and friend, have
exited.
But the problem is that we take an mmap_write_lock() around
ptdump_walk_pgd() so
this is a sleepable code path.

The mmap_write_lock around ptdump_walk_pgd() is definitely restrictive. I
think that

was put because ptdump is rarely used, I think we could have done
RCU-freeing of pagetables to

synchronize between ptdump and vmalloc pagetable lazy freeing/ hotunplug
pagetable freeing.

The other idea was to use a static key like the HVO code does, which
shouldn't place any RCU requirements on the debug walk.

Will

On 11/06/25 3:03 pm, Will Deacon wrote:

On Thu, Jun 05, 2025 at 01:46:01PM +0530, Dev Jain wrote:

quoted

On 05/06/25 10:18 am, Dev Jain wrote:

quoted

On 30/05/25 7:06 pm, Will Deacon wrote:

quoted

On Fri, May 30, 2025 at 02:11:36PM +0100, Ryan Roberts wrote:

quoted

On 30/05/2025 13:35, Will Deacon wrote:

quoted

I really don't think we should be adding unconditional
locking overhead
to core mm routines purely to facilitate a rarely used debug option.

Instead, can we either adopt something like the RCU-like walk used by
fast GUP or stick the locking behind a static key that's only enabled
when a ptdump walk is in progress (a bit like how
hugetlb_vmemmap_optimize_folio() manipulates
hugetlb_optimize_vmemmap_key)?

My sense is that the static key will be less effort and can be
contained fully
in arm64. I think we would need to enable the key around the call to
ptdump_walk_pgd() in both ptdump_walk() and ptdump_check_wx().
Then where Dev is
currently taking the read lock, that would be contingent on the
key being
enabled and the unlock would be contingent on having taken the lock.

Does that sound like an acceptable approach?

Yup, and I think you'll probably need something like a synchronize_rcu()
when flipping the key to deal with any pre-existing page-table freers.

IIUC, you mean to say that we need to ensure any existing readers having
a reference to the isolated table in pmd_free_pte_page and friend, have
exited.
But the problem is that we take an mmap_write_lock() around
ptdump_walk_pgd() so
this is a sleepable code path.

The mmap_write_lock around ptdump_walk_pgd() is definitely restrictive. I
think that

was put because ptdump is rarely used, I think we could have done
RCU-freeing of pagetables to

synchronize between ptdump and vmalloc pagetable lazy freeing/ hotunplug
pagetable freeing.

The other idea was to use a static key like the HVO code does, which
shouldn't place any RCU requirements on the debug walk.

Thanks for your suggestion. I shall look into this.

Will