Skip to contentSkip to conversation

CPA improvements

RFC In review Hyeonggon Yoo 4 participants 18 messages Jun 14, 2022 – Jul 7, 2022

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover CPA improvements

This is some improvements of CPA.

Main purpose of this series make pages not to lose _PAGE_GLOBAL after _set_pages_np() and set_pages_p() are called.

Hyeonggon Yoo (2):
x86/mm/cpa: always fail when user address is passed x86/mm/cpa: drop pgprot_clear_protnone_bits()

 arch/x86/mm/pat/set_memory.c | 30 ++++++------------------------
 1 file changed, 6 insertions(+), 24 deletions(-)
Signature · 1 line
2.32.0
#1 lore raw

RFC 1/2 x86/mm/cpa: always fail when user address is passed

Currently CPA is not used for user mappings (only pgd of init_mm or and efi_mm is used). For simplicity, always fail when user address is passed.

Note that efi_mm uses 1:1 mapping so its address should not be considered as user address.

---
 arch/x86/mm/pat/set_memory.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/mm/pat/set_memory.c +6 -0
index 1abd5438f126..67cf969fed0d 100644
@@ -20,6 +20,7 @@
 #include <linux/kernel.h>
 #include <linux/cc_platform.h>
 #include <linux/set_memory.h>
+#include <linux/efi.h>
 
 #include <asm/e820/api.h>
 #include <asm/processor.h>
@@ -1514,6 +1515,11 @@ static int __change_page_attr(struct cpa_data *cpa, int primary)
 	pte_t *kpte, old_pte;
 
 	address = __cpa_addr(cpa, cpa->curpage);
+
+	if (WARN((IS_ENABLED(CONFIG_EFI) ? cpa->pgd != efi_mm.pgd : true)
+		  && address <= TASK_SIZE_MAX,
+		 KERN_WARNING "CPA: Got a user address"))
+		return -EINVAL;
 repeat:
 	kpte = _lookup_address_cpa(cpa, address, &level);
 	if (!kpte)
Signature · 1 line
2.32.0
#2 lore raw

Re: [RFC 1/2] x86/mm/cpa: always fail when user address is passed

Quoted 26 lines · Hyeonggon Yoo

Currently CPA is not used for user mappings (only pgd of init_mm or and efi_mm is used). For simplicity, always fail when user address is passed.

Note that efi_mm uses 1:1 mapping so its address should not be considered as user address.

---
 arch/x86/mm/pat/set_memory.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/mm/pat/set_memory.c +1 -0
index 1abd5438f126..67cf969fed0d 100644
@@ -20,6 +20,7 @@
 #include <linux/kernel.h>
 #include <linux/cc_platform.h>
 #include <linux/set_memory.h>
+#include <linux/efi.h>
 
 #include <asm/e820/api.h>
 #include <asm/processor.h>

arch/x86/mm/pat/set_memory.c static int __change_page_attr(struct cpa_data · line 1515

Quoted 1 line
arch/x86/mm/pat/set_memory.c +0 -0
@@ -1514,6 +1515,11 @@ static int __change_page_attr(struct cpa_data
 *cpa, int primary)

I guess having it here instead of __change_page_attr_set_clr() will result in the direct map alias addresses getting checked as well. Since these are determined inside of CPA, I'm not sure if it's needed as much.

arch/x86/mm/pat/set_memory.c static int __change_page_attr(struct cpa_data *cpa, int primary)

Quoted 6 lines
arch/x86/mm/pat/set_memory.c +2 -0
 	pte_t *kpte, old_pte;
 
 	address = __cpa_addr(cpa, cpa->curpage);
+
+	if (WARN((IS_ENABLED(CONFIG_EFI) ? cpa->pgd != efi_mm.pgd :
 true)

Could it be?

(!IS_ENABLED(CONFIG_EFI) || cpa->pgd != efi_mm.pgd)

Quoted 6 lines
arch/x86/mm/pat/set_memory.c +3 -0
+		  && address <= TASK_SIZE_MAX,
+		 KERN_WARNING "CPA: Got a user address"))
+		return -EINVAL;
 repeat:
arch/x86/mm/pat/set_memory.c +0 -0
 	kpte = _lookup_address_cpa(cpa, address, &level);
 	if (!kpte)
#3 lore raw

Re: [RFC 1/2] x86/mm/cpa: always fail when user address is passed

Quoted 24 lines · Rick P
Quoted 23 lines · Hyeonggon Yoo

Currently CPA is not used for user mappings (only pgd of init_mm or and efi_mm is used). For simplicity, always fail when user address is passed.

Note that efi_mm uses 1:1 mapping so its address should not be considered as user address.

---
 arch/x86/mm/pat/set_memory.c | 6 ++++++
 1 file changed, 6 insertions(+)
arch/x86/mm/pat/set_memory.c +1 -0
index 1abd5438f126..67cf969fed0d 100644
@@ -20,6 +20,7 @@
 #include <linux/kernel.h>
 #include <linux/cc_platform.h>
 #include <linux/set_memory.h>
+#include <linux/efi.h>
 
Quoted 9 lines

arch/x86/mm/pat/set_memory.c static int __change_page_attr(struct cpa_data · line 1515

Quoted 4 lines
arch/x86/mm/pat/set_memory.c +0 -0
 #include <asm/e820/api.h>
 #include <asm/processor.h>
@@ -1514,6 +1515,11 @@ static int __change_page_attr(struct cpa_data
 *cpa, int primary)

I guess having it here instead of __change_page_attr_set_clr() will result in the direct map alias addresses getting checked as well. Since these are determined inside of CPA, I'm not sure if it's needed as much.

It does not check alias address when it failed. I put it in __change_page_attr() with CPA_ARRAY in mind. Because it may not be a single continuous area.

Quoted 11 lines

arch/x86/mm/pat/set_memory.c static int __change_page_attr(struct cpa_data *cpa, int primary)

Quoted 7 lines
arch/x86/mm/pat/set_memory.c +2 -0
 	pte_t *kpte, old_pte;
 
 	address = __cpa_addr(cpa, cpa->curpage);
+
+	if (WARN((IS_ENABLED(CONFIG_EFI) ? cpa->pgd != efi_mm.pgd :
 true)

Could it be?

(!IS_ENABLED(CONFIG_EFI) || cpa->pgd != efi_mm.pgd)

Looks better, will update in v2.

Thanks!
Hyeonggon

Quoted 7 lines
Quoted 7 lines
arch/x86/mm/pat/set_memory.c +3 -0
+		  && address <= TASK_SIZE_MAX,
+		 KERN_WARNING "CPA: Got a user address"))
+		return -EINVAL;
 repeat:
arch/x86/mm/pat/set_memory.c +0 -0
 	kpte = _lookup_address_cpa(cpa, address, &level);
 	if (!kpte)
Signature · 2 lines
Thanks,
Hyeonggon
#4 lore raw

Re: [RFC 1/2] x86/mm/cpa: always fail when user address is passed

Quoted 38 lines · Hyeonggon Yoo
Quoted 25 lines · Rick P
Quoted 24 lines · Hyeonggon Yoo

Currently CPA is not used for user mappings (only pgd of init_mm or and efi_mm is used). For simplicity, always fail when user address
is passed.

Note that efi_mm uses 1:1 mapping so its address should not be considered as user address.

---
  arch/x86/mm/pat/set_memory.c | 6 ++++++
  1 file changed, 6 insertions(+)
arch/x86/mm/pat/set_memory.c +1 -0
index 1abd5438f126..67cf969fed0d 100644
@@ -20,6 +20,7 @@
  #include <linux/kernel.h>
  #include <linux/cc_platform.h>
  #include <linux/set_memory.h>
+#include <linux/efi.h>
 
Quoted 11 lines
Quoted 5 lines
arch/x86/mm/pat/set_memory.c +0 -0
  #include <asm/e820/api.h>
  #include <asm/processor.h>
@@ -1514,6 +1515,11 @@ static int __change_page_attr(struct

cpa_data
*cpa, int primary)

I guess having it here instead of __change_page_attr_set_clr() will result in the direct map alias addresses getting checked as well. Since
these are determined inside of CPA, I'm not sure if it's needed as much.

Quoted 3 lines

It does not check alias address when it failed. I put it in __change_page_attr() with CPA_ARRAY in mind. Because it may not be a single continuous area.

Makes sense.

Quoted 9 lines
Quoted 9 lines
Quoted 8 lines

     pte_t *kpte, old_pte;
  
     address = __cpa_addr(cpa, cpa->curpage);
arch/x86/mm/pat/set_memory.c +2 -0
+
+   if (WARN((IS_ENABLED(CONFIG_EFI) ? cpa->pgd != efi_mm.pgd :

true)

Quoted 5 lines
Quoted 3 lines

Could it be?

(!IS_ENABLED(CONFIG_EFI) || cpa->pgd != efi_mm.pgd)

Looks better, will update in v2.

Christoph's seems better to me.

#5 lore raw

Re: [RFC 1/2] x86/mm/cpa: always fail when user address is passed

In reply to Hyeonggon Yoo

arch/x86/mm/pat/set_memory.c static int __change_page_attr(struct cpa_data *cpa, int primary) · line 1522

Quoted 9 lines · Hyeonggon Yoo
arch/x86/mm/pat/set_memory.c +5 -0
@@ -1514,6 +1515,11 @@ static int __change_page_attr(struct cpa_data *cpa, int primary)
 	pte_t *kpte, old_pte;
 
 	address = __cpa_addr(cpa, cpa->curpage);
+
+	if (WARN((IS_ENABLED(CONFIG_EFI) ? cpa->pgd != efi_mm.pgd : true)
+		  && address <= TASK_SIZE_MAX,
+		 KERN_WARNING "CPA: Got a user address"))
+		return -EINVAL;

I was expecting this to actually go after _PAGE_USER, not necessarily userspace addresses themselves.

What does and should happen with the VDSO, for instance? It's a _PAGE_USER mapping, but it's >TASK_SIZE. Should set_page_attr() work on it?

#6 lore raw

Re: [RFC 1/2] x86/mm/cpa: always fail when user address is passed

Quoted 4 lines · Dave Hansen
Quoted 3 lines · Hyeonggon Yoo
arch/x86/mm/pat/set_memory.c +0 -0
@@ -1514,6 +1515,11 @@ static int __change_page_attr(struct cpa_data *cpa, int primary)
 	pte_t *kpte, old_pte;
 
Quoted 9 lines

arch/x86/mm/pat/set_memory.c static int __change_page_attr(struct cpa_data *cpa, int primary) · line 1522

Quoted 6 lines
arch/x86/mm/pat/set_memory.c +5 -0
 	address = __cpa_addr(cpa, cpa->curpage);
+
+	if (WARN((IS_ENABLED(CONFIG_EFI) ? cpa->pgd != efi_mm.pgd : true)
+		  && address <= TASK_SIZE_MAX,
+		 KERN_WARNING "CPA: Got a user address"))
+		return -EINVAL;

I was expecting this to actually go after _PAGE_USER, not necessarily userspace addresses themselves.

userspace ptes may not have _PAGE_USER set. (e.g. swap entry) I think it's more accurate to go after user addresses.

Quoted 2 lines

What does and should happen with the VDSO, for instance? It's a _PAGE_USER mapping, but it's >TASK_SIZE.

you mean vsyscall? AFAIK address of mapped vDSO image is < TASK_SIZE. (or please tell me I'm wrong)

Quoted 1 line

Should set_page_attr() work on it?

vsyscall does not need CPA functionalities.
So I don't think it (__change_page_attr()) should work on vsyscall.

I think renaming fault_in_kernel_space() and using it would be more robust than simply using the expression (address <= TASK_SIZE_MAX).

Signature · 2 lines
Thanks,
Hyeonggon
#7 lore raw

Re: [RFC 1/2] x86/mm/cpa: always fail when user address is passed

Quoted 15 lines · Hyeonggon Yoo
Quoted 4 lines · Dave Hansen
Quoted 3 lines · Hyeonggon Yoo
arch/x86/mm/pat/set_memory.c +0 -0
@@ -1514,6 +1515,11 @@ static int __change_page_attr(struct cpa_data *cpa, int primary)
 	pte_t *kpte, old_pte;
 
Quoted 9 lines

arch/x86/mm/pat/set_memory.c static int __change_page_attr(struct cpa_data *cpa, int primary) · line 1522

Quoted 6 lines
arch/x86/mm/pat/set_memory.c +5 -0
 	address = __cpa_addr(cpa, cpa->curpage);
+
+	if (WARN((IS_ENABLED(CONFIG_EFI) ? cpa->pgd != efi_mm.pgd : true)
+		  && address <= TASK_SIZE_MAX,
+		 KERN_WARNING "CPA: Got a user address"))
+		return -EINVAL;

I was expecting this to actually go after _PAGE_USER, not necessarily userspace addresses themselves.

Quoted 2 lines

userspace ptes may not have _PAGE_USER set. (e.g. swap entry) I think it's more accurate to go after user addresses.

It would, of course, have to be paired with _PAGE_PRESENT checks. This works both on the way in and out of the set_memory code. It shouldn't clear other bits a PTE with _PAGE_PRESENT|_PAGE_USER and also shouldn't *result* in _PAGE_USER|_PAGE_PRESENT PTEs, even if those PTEs are in the kernel address space.

Filtering on the addresses also makes sense.

Quoted 5 lines
Quoted 2 lines

What does and should happen with the VDSO, for instance? It's a _PAGE_USER mapping, but it's >TASK_SIZE.

you mean vsyscall? AFAIK address of mapped vDSO image is < TASK_SIZE. (or please tell me I'm wrong)

You're right. That was a silly thinko.

Quoted 4 lines
Quoted 1 line

Should set_page_attr() work on it?

vsyscall does not need CPA functionalities.
So I don't think it (__change_page_attr()) should work on vsyscall.

Agreed.

#8 lore raw

Re: [RFC 1/2] x86/mm/cpa: always fail when user address is passed

Quoted 16 lines · Dave Hansen
Quoted 15 lines · Hyeonggon Yoo
Quoted 13 lines · Dave Hansen

arch/x86/mm/pat/set_memory.c static int __change_page_attr(struct cpa_data *cpa, int primary) · line 1522

Quoted 9 lines · Hyeonggon Yoo
arch/x86/mm/pat/set_memory.c +5 -0
@@ -1514,6 +1515,11 @@ static int __change_page_attr(struct cpa_data *cpa, int primary)
 	pte_t *kpte, old_pte;
 
 	address = __cpa_addr(cpa, cpa->curpage);
+
+	if (WARN((IS_ENABLED(CONFIG_EFI) ? cpa->pgd != efi_mm.pgd : true)
+		  && address <= TASK_SIZE_MAX,
+		 KERN_WARNING "CPA: Got a user address"))
+		return -EINVAL;

I was expecting this to actually go after _PAGE_USER, not necessarily userspace addresses themselves.

Quoted 6 lines
Quoted 2 lines

userspace ptes may not have _PAGE_USER set. (e.g. swap entry) I think it's more accurate to go after user addresses.

It would, of course, have to be paired with _PAGE_PRESENT checks. This works both on the way in and out of the set_memory code. It shouldn't clear other bits a PTE with _PAGE_PRESENT|_PAGE_USER and

You mean nothing should not use set_memory code for PTEs with _PAGE_USER|_PAGE_PRESENT but set_memory can still be used to clear _PAGE_USER|_PAGE_PRESENT?

Can't we just simply deny any PTE/PMDs with _PAGE_PRESENT|_PAGE_USER?

Quoted 3 lines

also shouldn't
*result* in _PAGE_USER|_PAGE_PRESENT PTEs, even if those PTEs are in the kernel address space.

Makes sense.

Quoted 16 lines

Filtering on the addresses also makes sense.

Quoted 5 lines
Quoted 2 lines

What does and should happen with the VDSO, for instance? It's a _PAGE_USER mapping, but it's >TASK_SIZE.

you mean vsyscall? AFAIK address of mapped vDSO image is < TASK_SIZE. (or please tell me I'm wrong)

You're right. That was a silly thinko.

Quoted 4 lines
Quoted 1 line

Should set_page_attr() work on it?

vsyscall does not need CPA functionalities.
So I don't think it (__change_page_attr()) should work on vsyscall.

Agreed.

Signature · 2 lines
Thanks,
Hyeonggon
#9 lore raw

Re: [RFC 1/2] x86/mm/cpa: always fail when user address is passed

Quoted 6 lines · Hyeonggon Yoo
Quoted 3 lines

It would, of course, have to be paired with _PAGE_PRESENT checks. This works both on the way in and out of the set_memory code. It shouldn't clear other bits a PTE with _PAGE_PRESENT|_PAGE_USER and

You mean nothing should not use set_memory code for PTEs with _PAGE_USER|_PAGE_PRESENT but set_memory can still be used to clear _PAGE_USER|_PAGE_PRESENT?

I don't see a reason why it should be able to clear _PAGE_USER|_PAGE_PRESENT entries either.

#10 lore raw

Re: [RFC 1/2] x86/mm/cpa: always fail when user address is passed

In reply to Hyeonggon Yoo

arch/x86/mm/pat/set_memory.c static int __change_page_attr(struct cpa_data *cpa, int primary) · line 1522

Quoted 5 lines · Hyeonggon Yoo
arch/x86/mm/pat/set_memory.c +5 -0
+
+	if (WARN((IS_ENABLED(CONFIG_EFI) ? cpa->pgd != efi_mm.pgd : true)
+		  && address <= TASK_SIZE_MAX,
+		 KERN_WARNING "CPA: Got a user address"))
+		return -EINVAL;

This looks rather unreadable.

Why not something like:

static bool cpa_addr_valid(struct cpa_data *cpa, unsigned long address)

{
	if (address > TASK_SIZE_MAX)
		return true;
	if (IS_ENABLED(CONFIG_EFI) && cpa->pgd == efi_mm.pgd)
		return true;
	return false;
}

....

	if (WARN_ON_ONCE(!cpa_addr_valid(cpa, address)))
		return -EINVAL;
#11 lore raw

Re: [RFC 1/2] x86/mm/cpa: always fail when user address is passed

Quoted 20 lines · Christoph Hellwig

arch/x86/mm/pat/set_memory.c static int __change_page_attr(struct cpa_data *cpa, int primary) · line 1522

Quoted 5 lines · Hyeonggon Yoo
arch/x86/mm/pat/set_memory.c +5 -0
+
+	if (WARN((IS_ENABLED(CONFIG_EFI) ? cpa->pgd != efi_mm.pgd : true)
+		  && address <= TASK_SIZE_MAX,
+		 KERN_WARNING "CPA: Got a user address"))
+		return -EINVAL;

This looks rather unreadable.

Why not something like:

static bool cpa_addr_valid(struct cpa_data *cpa, unsigned long address)

{
	if (address > TASK_SIZE_MAX)
		return true;
	if (IS_ENABLED(CONFIG_EFI) && cpa->pgd == efi_mm.pgd)
		return true;
	return false;
}
Quoted 5 lines

....

	if (WARN_ON_ONCE(!cpa_addr_valid(cpa, address)))
		return -EINVAL;

Thank your for feedback. I'll care more about readability in v2.

Signature · 2 lines
Thanks,
Hyeonggon
#12 lore raw

RFC 2/2 x86/mm/cpa: drop pgprot_clear_protnone_bits()

In reply to Hyeonggon Yoo

commit a8aed3e0752b4 ("x86/mm/pageattr: Prevent PSE and GLOABL leftovers to confuse pmd/pte_present and pmd_huge") made CPA clear _PAGE_GLOBAL when _PAGE_PRESENT is not set. This prevents kernel crashing when kernel reads a page with !_PAGE_PRESENT and _PAGE_PROTNONE (_PAGE_GLOBAL). And then it set _PAGE_GLOBAL back when setting _PAGE_PRESENT again.

After commit d1440b23c922d ("x86/mm: Factor out pageattr _PAGE_GLOBAL setting") made kernel not set unconditionally _PAGE_GLOBAL, pages lose global flag after _set_pages_np() and _set_pages_p() are called.

But after commit 3166851142411 ("x86: skip check for spurious faults for non-present faults"), spurious_kernel_fault() does not confuse pte/pmd entries with _PAGE_PROTNONE as present anymore. So simply drop pgprot_clear_protnone_bits().

---
 arch/x86/mm/pat/set_memory.c | 24 ------------------------
 1 file changed, 24 deletions(-)
arch/x86/mm/pat/set_memory.c +0 -24
index 67cf969fed0d..8a8ce8d78694 100644
@@ -746,23 +746,6 @@ static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 #endif
 }
 
-static pgprot_t pgprot_clear_protnone_bits(pgprot_t prot)
-{
-	/*
-	 * _PAGE_GLOBAL means "global page" for present PTEs.
-	 * But, it is also used to indicate _PAGE_PROTNONE
-	 * for non-present PTEs.
-	 *
-	 * This ensures that a _PAGE_GLOBAL PTE going from
-	 * present to non-present is not confused as
-	 * _PAGE_PROTNONE.
-	 */
-	if (!(pgprot_val(prot) & _PAGE_PRESENT))
-		pgprot_val(prot) &= ~_PAGE_GLOBAL;
-
-	return prot;
-}
-
 static int __should_split_large_page(pte_t *kpte, unsigned long address,
 				     struct cpa_data *cpa)
 {
@@ -824,7 +807,6 @@ static int __should_split_large_page(pte_t *kpte, unsigned long address,
 	 * different bit positions in the two formats.
 	 */
 	req_prot = pgprot_4k_2_large(req_prot);
-	req_prot = pgprot_clear_protnone_bits(req_prot);
 	if (pgprot_val(req_prot) & _PAGE_PRESENT)
 		pgprot_val(req_prot) |= _PAGE_PSE;
 
@@ -1013,8 +995,6 @@ __split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
 		return 1;
 	}
 
-	ref_prot = pgprot_clear_protnone_bits(ref_prot);
-
 	/*
 	 * Get the target pfn from the original entry:
 	 */
@@ -1246,8 +1226,6 @@ static void populate_pte(struct cpa_data *cpa,
 
 	pte = pte_offset_kernel(pmd, start);
 
-	pgprot = pgprot_clear_protnone_bits(pgprot);
-
 	while (num_pages-- && start < end) {
 		set_pte(pte, pfn_pte(cpa->pfn, pgprot));
 
@@ -1542,8 +1520,6 @@ static int __change_page_attr(struct cpa_data *cpa, int primary)
 		new_prot = static_protections(new_prot, address, pfn, 1, 0,
 					      CPA_PROTECT);
 
-		new_prot = pgprot_clear_protnone_bits(new_prot);
-
 		/*
 		 * We need to keep the pfn from the existing PTE,
 		 * after all we're only going to change it's attributes
Signature · 1 line
2.32.0
#13 lore raw

Re: [RFC 2/2] x86/mm/cpa: drop pgprot_clear_protnone_bits()

Quoted 6 lines · Hyeonggon Yoo

commit a8aed3e0752b4 ("x86/mm/pageattr: Prevent PSE and GLOABL leftovers to confuse pmd/pte_present and pmd_huge") made CPA clear _PAGE_GLOBAL when _PAGE_PRESENT is not set. This prevents kernel crashing when kernel reads a page with !_PAGE_PRESENT and _PAGE_PROTNONE (_PAGE_GLOBAL). And then it set _PAGE_GLOBAL back when setting _PAGE_PRESENT again.

Quoted 8 lines

After commit d1440b23c922d ("x86/mm: Factor out pageattr _PAGE_GLOBAL setting") made kernel not set unconditionally _PAGE_GLOBAL, pages lose global flag after _set_pages_np() and _set_pages_p() are called.

But after commit 3166851142411 ("x86: skip check for spurious faults for non-present faults"), spurious_kernel_fault() does not confuse pte/pmd entries with _PAGE_PROTNONE as present anymore. So simply drop pgprot_clear_protnone_bits().

Looks like I forgot to Cc: Andrea Arcangeli <[email protected]>

Plus I did check that kernel does not crash when reading from/writing to non-present pages with this patch applied.

Quoted 71 lines
---
 arch/x86/mm/pat/set_memory.c | 24 ------------------------
 1 file changed, 24 deletions(-)
arch/x86/mm/pat/set_memory.c +0 -24
index 67cf969fed0d..8a8ce8d78694 100644
@@ -746,23 +746,6 @@ static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 #endif
 }
 
-static pgprot_t pgprot_clear_protnone_bits(pgprot_t prot)
-{
-	/*
-	 * _PAGE_GLOBAL means "global page" for present PTEs.
-	 * But, it is also used to indicate _PAGE_PROTNONE
-	 * for non-present PTEs.
-	 *
-	 * This ensures that a _PAGE_GLOBAL PTE going from
-	 * present to non-present is not confused as
-	 * _PAGE_PROTNONE.
-	 */
-	if (!(pgprot_val(prot) & _PAGE_PRESENT))
-		pgprot_val(prot) &= ~_PAGE_GLOBAL;
-
-	return prot;
-}
-
 static int __should_split_large_page(pte_t *kpte, unsigned long address,
 				     struct cpa_data *cpa)
 {
@@ -824,7 +807,6 @@ static int __should_split_large_page(pte_t *kpte, unsigned long address,
 	 * different bit positions in the two formats.
 	 */
 	req_prot = pgprot_4k_2_large(req_prot);
-	req_prot = pgprot_clear_protnone_bits(req_prot);
 	if (pgprot_val(req_prot) & _PAGE_PRESENT)
 		pgprot_val(req_prot) |= _PAGE_PSE;
 
@@ -1013,8 +995,6 @@ __split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
 		return 1;
 	}
 
-	ref_prot = pgprot_clear_protnone_bits(ref_prot);
-
 	/*
 	 * Get the target pfn from the original entry:
 	 */
@@ -1246,8 +1226,6 @@ static void populate_pte(struct cpa_data *cpa,
 
 	pte = pte_offset_kernel(pmd, start);
 
-	pgprot = pgprot_clear_protnone_bits(pgprot);
-
 	while (num_pages-- && start < end) {
 		set_pte(pte, pfn_pte(cpa->pfn, pgprot));
 
@@ -1542,8 +1520,6 @@ static int __change_page_attr(struct cpa_data *cpa, int primary)
 		new_prot = static_protections(new_prot, address, pfn, 1, 0,
 					      CPA_PROTECT);
 
-		new_prot = pgprot_clear_protnone_bits(new_prot);
-
 		/*
 		 * We need to keep the pfn from the existing PTE,
 		 * after all we're only going to change it's attributes
Signature · 1 line
2.32.0
Signature · 2 lines
Thanks,
Hyeonggon
#14 lore raw

Re: [RFC 2/2] x86/mm/cpa: drop pgprot_clear_protnone_bits()

Quoted 24 lines · Hyeonggon Yoo
Quoted 16 lines · Hyeonggon Yoo

commit a8aed3e0752b4 ("x86/mm/pageattr: Prevent PSE and GLOABL leftovers
to confuse pmd/pte_present and pmd_huge") made CPA clear _PAGE_GLOBAL when
_PAGE_PRESENT is not set. This prevents kernel crashing when kernel reads
a page with !_PAGE_PRESENT and _PAGE_PROTNONE (_PAGE_GLOBAL). And then it
set _PAGE_GLOBAL back when setting _PAGE_PRESENT again.

After commit d1440b23c922d ("x86/mm: Factor out pageattr _PAGE_GLOBAL
setting") made kernel not set unconditionally _PAGE_GLOBAL, pages lose
global flag after _set_pages_np() and _set_pages_p() are called.

Quoted 5 lines

But after commit 3166851142411 ("x86: skip check for spurious faults for
non-present faults"), spurious_kernel_fault() does not confuse pte/pmd entries with _PAGE_PROTNONE as present anymore. So simply drop pgprot_clear_protnone_bits().

Quoted 5 lines

Looks like I forgot to Cc: Andrea Arcangeli <[email protected]>

Plus I did check that kernel does not crash when reading from/writing to
non-present pages with this patch applied.

Thanks for the history.

I think we should still fix pte_present() to not check prot_none if the user bit is clear. The spurious fault handler infinite loop may no longer be a problem, but pte_present() still would return true for kernel NP pages, so be fragile. Today I see at least the oops message and memory hotunplug (see remove_pagetable()) that would get confused.

#15 lore raw

Re: [RFC 2/2] x86/mm/cpa: drop pgprot_clear_protnone_bits()

Quoted 27 lines · Rick P
Quoted 26 lines · Hyeonggon Yoo
Quoted 16 lines · Hyeonggon Yoo

commit a8aed3e0752b4 ("x86/mm/pageattr: Prevent PSE and GLOABL leftovers
to confuse pmd/pte_present and pmd_huge") made CPA clear _PAGE_GLOBAL when
_PAGE_PRESENT is not set. This prevents kernel crashing when kernel reads
a page with !_PAGE_PRESENT and _PAGE_PROTNONE (_PAGE_GLOBAL). And then it
set _PAGE_GLOBAL back when setting _PAGE_PRESENT again.

After commit d1440b23c922d ("x86/mm: Factor out pageattr _PAGE_GLOBAL
setting") made kernel not set unconditionally _PAGE_GLOBAL, pages lose
global flag after _set_pages_np() and _set_pages_p() are called.

Quoted 5 lines

But after commit 3166851142411 ("x86: skip check for spurious faults for
non-present faults"), spurious_kernel_fault() does not confuse pte/pmd entries with _PAGE_PROTNONE as present anymore. So simply drop pgprot_clear_protnone_bits().

Looks like I forgot to Cc: Andrea Arcangeli <[email protected]>

Quoted 8 lines
Quoted 3 lines

Plus I did check that kernel does not crash when reading from/writing to
non-present pages with this patch applied.

Thanks for the history.

I think we should still fix pte_present() to not check prot_none if the user bit is clear.

I tried, but realized it wouldn't work :(

For example, when a pte entry is used as swap entry, _PAGE_PRESENT is cleared and _PAGE_PROTNONE is set.

And other bits are used as type and offset of swap entry. In that case, _PAGE_BIT_USER bit does not represent _PAGE_USER. It is just one of bits that represents type of swap entry.

So checking if _PAGE_PROTNONE set only when _PAGE_USER is set will confuse some swap entries as non-present.

Quoted 4 lines

The spurious fault handler infinite loop may no
longer be a problem, but pte_present() still would return true for kernel NP pages, so be fragile. Today I see at least the oops message and memory hotunplug (see remove_pagetable()) that would get confused.

As explained above, I don't think it's possible to make pte_present() accurate for both kernel and user ptes.

Maybe we can implement pte_present_kernel()/pte_present_user() for when kernel knows it is user or kernel pte.

or pte_present_with_address(pte, address) if we don't know it is user pte or kernel pte.

Signature · 2 lines
Thanks,
Hyeonggon
#16 lore raw

Re: [RFC 2/2] x86/mm/cpa: drop pgprot_clear_protnone_bits()

Quoted 49 lines · Hyeonggon Yoo
Quoted 37 lines · Rick P
Quoted 29 lines · Hyeonggon Yoo
Quoted 20 lines · Hyeonggon Yoo

commit a8aed3e0752b4 ("x86/mm/pageattr: Prevent PSE and GLOABL leftovers
to confuse pmd/pte_present and pmd_huge") made CPA clear _PAGE_GLOBAL when
_PAGE_PRESENT is not set. This prevents kernel crashing when kernel
reads
a page with !_PAGE_PRESENT and _PAGE_PROTNONE (_PAGE_GLOBAL). And
then it
set _PAGE_GLOBAL back when setting _PAGE_PRESENT again.

After commit d1440b23c922d ("x86/mm: Factor out pageattr _PAGE_GLOBAL
setting") made kernel not set unconditionally _PAGE_GLOBAL, pages
lose
global flag after _set_pages_np() and _set_pages_p() are called.

Quoted 6 lines

But after commit 3166851142411 ("x86: skip check for spurious faults for
non-present faults"), spurious_kernel_fault() does not confuse pte/pmd entries with _PAGE_PROTNONE as present anymore. So simply
drop pgprot_clear_protnone_bits().

Quoted 6 lines

Looks like I forgot to Cc: Andrea Arcangeli <[email protected]>

Plus I did check that kernel does not crash when reading from/writing
to
non-present pages with this patch applied.

Quoted 5 lines

Thanks for the history.

I think we should still fix pte_present() to not check prot_none if the
user bit is clear.

I tried, but realized it wouldn't work :(

For example, when a pte entry is used as swap entry, _PAGE_PRESENT is cleared and _PAGE_PROTNONE is set.

Quoted 6 lines

And other bits are used as type and offset of swap entry. In that case, _PAGE_BIT_USER bit does not represent _PAGE_USER. It is just one of bits that represents type of swap entry.

So checking if _PAGE_PROTNONE set only when _PAGE_USER is set will confuse some swap entries as non-present.

Oooh, right. So the user bit records "when a pagetable is writeprotected by userfaultfd WP support". I'm not sure if maybe PCD is available to move that to and leave the user bit in place, but it sounds like an errata sensitive area to be tweaking.

Quoted 8 lines
Quoted 7 lines

The spurious fault handler infinite loop may no
longer be a problem, but pte_present() still would return true for kernel NP pages, so be fragile. Today I see at least the oops message
and memory hotunplug (see remove_pagetable()) that would get confused.

Quoted 6 lines

As explained above, I don't think it's possible to make pte_present()
accurate for both kernel and user ptes.

Maybe we can implement pte_present_kernel()/pte_present_user() for when kernel knows it is user or kernel pte.

This seems like a decent option to me. It seems there are only a few places that are isolated to arch/x86.

Quoted 4 lines

or pte_present_with_address(pte, address) if we don't know it is user pte or kernel pte.

#17 lore raw

Re: [RFC 2/2] x86/mm/cpa: drop pgprot_clear_protnone_bits()

Quoted 74 lines · Rick P
Quoted 49 lines · Hyeonggon Yoo
Quoted 37 lines · Rick P
Quoted 29 lines · Hyeonggon Yoo
Quoted 20 lines · Hyeonggon Yoo

commit a8aed3e0752b4 ("x86/mm/pageattr: Prevent PSE and GLOABL leftovers
to confuse pmd/pte_present and pmd_huge") made CPA clear _PAGE_GLOBAL when
_PAGE_PRESENT is not set. This prevents kernel crashing when kernel
reads
a page with !_PAGE_PRESENT and _PAGE_PROTNONE (_PAGE_GLOBAL). And
then it
set _PAGE_GLOBAL back when setting _PAGE_PRESENT again.

After commit d1440b23c922d ("x86/mm: Factor out pageattr _PAGE_GLOBAL
setting") made kernel not set unconditionally _PAGE_GLOBAL, pages
lose
global flag after _set_pages_np() and _set_pages_p() are called.

Quoted 6 lines

But after commit 3166851142411 ("x86: skip check for spurious faults for
non-present faults"), spurious_kernel_fault() does not confuse pte/pmd entries with _PAGE_PROTNONE as present anymore. So simply
drop pgprot_clear_protnone_bits().

Quoted 6 lines

Looks like I forgot to Cc: Andrea Arcangeli <[email protected]>

Plus I did check that kernel does not crash when reading from/writing
to
non-present pages with this patch applied.

Quoted 5 lines

Thanks for the history.

I think we should still fix pte_present() to not check prot_none if the
user bit is clear.

I tried, but realized it wouldn't work :(

For example, when a pte entry is used as swap entry, _PAGE_PRESENT is cleared and _PAGE_PROTNONE is set.

Quoted 6 lines

And other bits are used as type and offset of swap entry. In that case, _PAGE_BIT_USER bit does not represent _PAGE_USER. It is just one of bits that represents type of swap entry.

So checking if _PAGE_PROTNONE set only when _PAGE_USER is set will confuse some swap entries as non-present.

Oooh, right. So the user bit records "when a pagetable is writeprotected by userfaultfd WP support". I'm not sure if maybe PCD is available to move that to and leave the user bit in place, but it sounds like an errata sensitive area to be tweaking.

Quoted 12 lines
Quoted 7 lines

The spurious fault handler infinite loop may no
longer be a problem, but pte_present() still would return true for kernel NP pages, so be fragile. Today I see at least the oops message
and memory hotunplug (see remove_pagetable()) that would get confused.

As explained above, I don't think it's possible to make pte_present()
accurate for both kernel and user ptes.

Quoted 5 lines
Quoted 2 lines

Maybe we can implement pte_present_kernel()/pte_present_user() for when kernel knows it is user or kernel pte.

This seems like a decent option to me. It seems there are only a few places that are isolated to arch/x86.

But there are some places where kernel does not know if it's kernel pte or user pte.

For example show_fault_oops() can be called for both kernel and user address. Is something like this acceptable?

static inline bool pte_present_address(pte_t pte, address)

{
	if (kernel address)
		return pte_present_kernel(pte);
	return pte_present_user(pte);
}
Quoted 4 lines
Quoted 4 lines

or pte_present_with_address(pte, address) if we don't know it is user pte or kernel pte.

Signature · 2 lines
Thanks,
Hyeonggon