x86/mm: Decouple dynamic __PHYSICAL_MASK from AMD SME
PATCH 1/1 x86/mm: Decouple dynamic __PHYSICAL_MASK from AMD SME
AMD SME claims one bit from physical address to indicate whether the page is encrypted or not. To achieve that we clear out the bit from __PHYSICAL_MASK.
The capability to adjust __PHYSICAL_MASK is required beyond AMD SME. For instance for upcoming Intel Multi-Key Total Memory Encryption.
Let's factor it out into separate feature with own Kconfig handle.
It also helps with overhead of AMD SME. It saves more than 3k in .text on defconfig + AMD_MEM_ENCRYPT:
add/remove: 3/2 grow/shrink: 5/110 up/down: 189/-3753 (-3564)
We would need to return to this once we have infrastructure to patch constants in code. That's good candidate for it.
--- arch/x86/Kconfig | 4 ++++ arch/x86/boot/compressed/pagetable.c | 3 +++ arch/x86/include/asm/page_types.h | 8 +++++++- arch/x86/mm/mem_encrypt.c | 3 +++ arch/x86/mm/pgtable.c | 5 +++++ 5 files changed, 22 insertions(+), 1 deletion(-)
arch/x86/Kconfig +4 -0
@@ -332,6 +332,9 @@ config ARCH_SUPPORTS_UPROBES config FIX_EARLYCON_MEM def_bool y +config DYNAMIC_PHYSICAL_MASK + bool + config PGTABLE_LEVELS int default 5 if X86_5LEVEL
@@ -1469,6 +1472,7 @@ config ARCH_HAS_MEM_ENCRYPT config AMD_MEM_ENCRYPT bool "AMD Secure Memory Encryption (SME) support" depends on X86_64 && CPU_SUP_AMD + select DYNAMIC_PHYSICAL_MASK ---help--- Say yes to enable support for the encryption of system memory. This requires an AMD processor that supports Secure Memory
arch/x86/boot/compressed/pagetable.c +3 -0
@@ -16,6 +16,9 @@#define __pa (x) ((unsigned long)(x))#define __va (x) ((void *)((unsigned long)(x))) +/* No need in adjustable __PHYSICAL_MASK during decompresssion phase */ +#undef CONFIG_DYNAMIC_PHYSICAL_MASK +/* * The pgtable.h and mm/ident_map.c includes make use of the SME related * information which is not used in the compressed image support. Un-define
arch/x86/include/asm/page_types.h +7 -1
@@ -17,7 +17,6 @@#define PUD_PAGE_SIZE (_AC(1, UL) << PUD_SHIFT)#define PUD_PAGE_MASK (~(PUD_PAGE_SIZE-1)) -#define __PHYSICAL_MASK ((phys_addr_t)(__sme_clr((1ULL << __PHYSICAL_MASK_SHIFT) - 1)))#define __VIRTUAL_MASK ((1UL << __VIRTUAL_MASK_SHIFT) - 1)/* Cast *PAGE_MASK to a signed type so that it is sign-extended if
@@ -55,6 +54,13 @@#ifndef __ASSEMBLY__ +#ifdef CONFIG_DYNAMIC_PHYSICAL_MASK +extern phys_addr_t physical_mask; +#define __PHYSICAL_MASK physical_mask +#else +#define __PHYSICAL_MASK ((phys_addr_t)((1ULL << __PHYSICAL_MASK_SHIFT) - 1)) +#endif +extern int devmem_is_allowed (unsigned long pagenr);extern unsigned long max_low_pfn_mapped;
arch/x86/mm/mem_encrypt.c +3 -0
@@ -999,6 +999,7 @@ void __init __nostackprotector sme_enable(struct boot_params *bp)/* SEV state cannot be controlled by a command line option */ sme_me_mask = me_mask; sev_enabled = true; + physical_mask &= ~sme_me_mask;return ; }
@@ -1033,4 +1034,6 @@ void __init __nostackprotector sme_enable(struct boot_params *bp) sme_me_mask =0 ; else sme_me_mask = active_by_default ? me_mask :0 ; + + physical_mask &= ~sme_me_mask; }
arch/x86/mm/pgtable.c +5 -0
@@ -7,6 +7,11 @@#include <asm/fixmap.h> #include <asm/mtrr.h> +#ifdef CONFIG_DYNAMIC_PHYSICAL_MASK +phys_addr_t physical_mask__ro_after_init = (1ULL << __PHYSICAL_MASK_SHIFT) -1 ; +EXPORT_SYMBOL (physical_mask); +#endif +#define PGALLOC_GFP (GFP_KERNEL_ACCOUNT | __GFP_ZERO)#ifdef CONFIG_HIGHPTE
Signature · 1 line
2.15.1
Quoted 3 lines · Kirill A. Shutemov
AMD SME claims one bit from physical address to indicate whether the page is encrypted or not. To achieve that we clear out the bit from
__PHYSICAL_MASK.
I was actually working on a suggestion by Linus to use one of the software page table bits to indicate encryption and translate that to the hardware bit when writing the actual page table entry. With that, __PHYSICAL_MASK would go back to its original definition.
Thanks,
Tom
Quoted 121 lines
The capability to adjust
__PHYSICAL_MASKis required beyond AMD SME. For instance for upcoming Intel Multi-Key Total Memory Encryption.Let's factor it out into separate feature with own
Kconfighandle.It also helps with overhead of AMD SME. It saves more than 3k in .text on defconfig +
AMD_MEM_ENCRYPT:add/remove: 3/2 grow/shrink: 5/110 up/down: 189/-3753 (-3564)We would need to return to this once we have infrastructure to patch constants in code. That's good candidate for it.
Signed-off-by Kirill A. Shutemov <[email protected]>--- arch/x86/Kconfig | 4 ++++ arch/x86/boot/compressed/pagetable.c | 3 +++ arch/x86/include/asm/page_types.h | 8 +++++++- arch/x86/mm/mem_encrypt.c | 3 +++ arch/x86/mm/pgtable.c | 5 +++++ 5 files changed, 22 insertions(+), 1 deletion(-)arch/x86/Kconfig +4
-0@@ -332,6 +332,9 @@ config ARCH_SUPPORTS_UPROBES config FIX_EARLYCON_MEM def_bool y +config DYNAMIC_PHYSICAL_MASK + bool + config PGTABLE_LEVELS int default 5 if X86_5LEVEL@@ -1469,6 +1472,7 @@ config ARCH_HAS_MEM_ENCRYPT config AMD_MEM_ENCRYPT bool "AMD Secure Memory Encryption (SME) support" depends on X86_64 && CPU_SUP_AMD + select DYNAMIC_PHYSICAL_MASK ---help--- Say yes to enable support for the encryption of system memory. This requires an AMD processor that supports Secure Memoryarch/x86/boot/compressed/pagetable.c +3
-0@@ -16,6 +16,9 @@#define __pa (x) ((unsigned long)(x))#define __va (x) ((void *)((unsigned long)(x))) +/* No need in adjustable __PHYSICAL_MASK during decompresssion phase */ +#undef CONFIG_DYNAMIC_PHYSICAL_MASK +/* * The pgtable.h and mm/ident_map.c includes make use of the SME related * information which is not used in the compressed image support. Un-define arch/x86/include/asm/page_types.h +7
-1@@ -17,7 +17,6 @@#define PUD_PAGE_SIZE (_AC(1, UL) << PUD_SHIFT)#define PUD_PAGE_MASK (~(PUD_PAGE_SIZE-1)) -#define __PHYSICAL_MASK ((phys_addr_t)(__sme_clr((1ULL << __PHYSICAL_MASK_SHIFT) - 1)))#define __VIRTUAL_MASK ((1UL << __VIRTUAL_MASK_SHIFT) - 1)/* Cast *PAGE_MASK to a signed type so that it is sign-extended if @@ -55,6 +54,13 @@#ifndef __ASSEMBLY__ +#ifdef CONFIG_DYNAMIC_PHYSICAL_MASK +extern phys_addr_t physical_mask; +#define __PHYSICAL_MASK physical_mask +#else +#define __PHYSICAL_MASK ((phys_addr_t)((1ULL << __PHYSICAL_MASK_SHIFT) - 1)) +#endif +extern int devmem_is_allowed (unsigned long pagenr);extern unsigned long max_low_pfn_mapped;arch/x86/mm/mem_encrypt.c +3
-0@@ -999,6 +999,7 @@ void __init __nostackprotector sme_enable(struct boot_params *bp)/* SEV state cannot be controlled by a command line option */ sme_me_mask = me_mask; sev_enabled = true; + physical_mask &= ~sme_me_mask;return ; }@@ -1033,4 +1034,6 @@ void __init __nostackprotector sme_enable(struct boot_params *bp) sme_me_mask =0 ; else sme_me_mask = active_by_default ? me_mask :0 ; + + physical_mask &= ~sme_me_mask; }arch/x86/mm/pgtable.c +5
-0@@ -7,6 +7,11 @@#include <asm/fixmap.h> #include <asm/mtrr.h> +#ifdef CONFIG_DYNAMIC_PHYSICAL_MASK +phys_addr_t physical_mask__ro_after_init = (1ULL << __PHYSICAL_MASK_SHIFT) -1 ; +EXPORT_SYMBOL (physical_mask); +#endif +#define PGALLOC_GFP (GFP_KERNEL_ACCOUNT | __GFP_ZERO)#ifdef CONFIG_HIGHPTE
Quoted 9 lines · Tom Lendacky
Quoted 3 lines · Kirill A. Shutemov
AMD SME claims one bit from physical address to indicate whether the page is encrypted or not. To achieve that we clear out the bit from
__PHYSICAL_MASK.I was actually working on a suggestion by Linus to use one of the software page table bits to indicate encryption and translate that to the hardware bit when writing the actual page table entry. With that,
__PHYSICAL_MASKwould go back to its original definition.
But you would need to mask it on reading of pfn from page table entry, right? I expect it to have more overhead than this one.
And software bits are valuable. Do we still have a spare one for this?
Signature · 1 line
Kirill A. Shutemov
Quoted 6 lines · Kirill A. Shutemov
Quoted 5 lines · Tom Lendacky
Quoted 3 lines · Kirill A. Shutemov
AMD SME claims one bit from physical address to indicate whether the page is encrypted or not. To achieve that we clear out the bit from
__PHYSICAL_MASK.
Quoted 7 lines
Quoted 4 lines
I was actually working on a suggestion by Linus to use one of the software page table bits to indicate encryption and translate that to the hardware bit when writing the actual page table entry. With that,
__PHYSICAL_MASKwould go back to its original definition.But you would need to mask it on reading of pfn from page table entry, right? I expect it to have more overhead than this one.
When reading back an entry it would translate the hardware bit position back to the software bit position. The suggestion for changing it was to make _PAGE_ENC a constant and not tied to the sme_me_mask.
See https://marc.info/?l=linux-kernel&m=151017622615894&w=2
Quoted 2 lines
And software bits are valuable. Do we still have a spare one for this?
I was looking at possibly using bit 57 (_PAGE_BIT_SOFTW5).
Thanks,
Tom
Quoted 1 line
Quoted 27 lines · Tom Lendacky
Quoted 16 lines · Kirill A. Shutemov
Quoted 7 lines · Tom Lendacky
Quoted 5 lines · Kirill A. Shutemov
AMD SME claims one bit from physical address to indicate whether the
page is encrypted or not. To achieve that we clear out the bit from__PHYSICAL_MASK.Quoted 7 lines
I was actually working on a suggestion by Linus to use one of the software
page table bits to indicate encryption and translate that to the hardware
bit when writing the actual page table entry. With that,__PHYSICAL_MASK
would go back to its original definition.Quoted 3 lines
But you would need to mask it on reading of pfn from page table entry,
right? I expect it to have more overhead than this one.When reading back an entry it would translate the hardware bit position
back to the software bit position. The suggestion for changing it was
to make_PAGE_ENCa constant and not tied to thesme_me_mask.
Quoted 7 lines
See https://marc.info/?l=linux-kernel&m=151017622615894&w=2
Quoted 3 lines
And software bits are valuable. Do we still have a spare one for this?
I was looking at possibly using bit 57 (
_PAGE_BIT_SOFTW5).
But MK-TME supports upto 15 bits (architectually) as keyID. How is this supposed to work with MK-TME?
Thanks,
-Kai
Quoted 5 lines
Thanks,
TomQuoted 1 line
Quoted 22 lines · Kai Huang
Quoted 21 lines · Tom Lendacky
Quoted 19 lines · Kirill A. Shutemov
Quoted 7 lines · Tom Lendacky
Quoted 5 lines · Kirill A. Shutemov
AMD SME claims one bit from physical address to indicate whether the
page is encrypted or not. To achieve that we clear out the bit from__PHYSICAL_MASK.Quoted 7 lines
I was actually working on a suggestion by Linus to use one of the software
page table bits to indicate encryption and translate that to the hardware
bit when writing the actual page table entry. With that,__PHYSICAL_MASK
would go back to its original definition.But you would need to mask it on reading of pfn from page table entry,
right? I expect it to have more overhead than this one.
Quoted 5 lines
Quoted 5 lines
When reading back an entry it would translate the hardware bit position
back to the software bit position. The suggestion for changing it was
to make_PAGE_ENCa constant and not tied to thesme_me_mask.
But is it really constant? I thought it's enumerated at boot-time. Can we step onto a problem for future AMD CPUs?
In case of MKTME the bits we need to clear are not constant. Depends on CPU and BIOS settings.
By making _PAGE_ENC constant we would effectively lower maximum physical address space the kernel can handle, regardless if the system has SME enabled. I can imagine some people wouldn't be happy about this.
And I think it would collide with 5-level paging.
I would leave it as variable for now and look on this later once we would have infrastructure to patch constants in kernel text.
Signature · 1 line
Kirill A. Shutemov
Quoted 23 lines · Kirill A. Shutemov
Quoted 22 lines · Kai Huang
Quoted 21 lines · Tom Lendacky
Quoted 19 lines · Kirill A. Shutemov
Quoted 7 lines · Tom Lendacky
Quoted 5 lines · Kirill A. Shutemov
AMD SME claims one bit from physical address to indicate whether the
page is encrypted or not. To achieve that we clear out the bit from__PHYSICAL_MASK.Quoted 7 lines
I was actually working on a suggestion by Linus to use one of the software
page table bits to indicate encryption and translate that to the hardware
bit when writing the actual page table entry. With that,__PHYSICAL_MASK
would go back to its original definition.But you would need to mask it on reading of pfn from page table entry,
right? I expect it to have more overhead than this one.
Quoted 8 lines
Quoted 5 lines
Quoted 5 lines
When reading back an entry it would translate the hardware bit position
back to the software bit position. The suggestion for changing it was
to make_PAGE_ENCa constant and not tied to thesme_me_mask.But is it really constant? I thought it's enumerated at boot-time. Can we step onto a problem for future AMD CPUs?
_PAGE_ENC would be constant and it would be translated to the actual bit that was enumerated at boot-time when writing the page table entry and translated back to _PAGE_ENC when reading the page table entry.
Quoted 7 lines
In case of MKTME the bits we need to clear are not constant. Depends on CPU and BIOS settings.
By making
_PAGE_ENCconstant we would effectively lower maximum physical address space the kernel can handle, regardless if the system has SME enabled. I can imagine some people wouldn't be happy about this.
I don't see how this would lower the maximum physical address space the kernel could handle. Bit 57 is part of the reserved page table flag bits and if SME is not enabled the hardware bits are never used.
What I do see as a problem is a kernel built with support for SME, and therefore _PAGE_ENC is not zero, but SME has not been enabled by the BIOS or mem_encrypt=off is specified. In this case you can never be certain that the translation from software bit to hardware bit and back is correct. Take for example, pmd_bad(). Here, _KERNPG_TABLE would have a non-zero _PAGE_ENC or'd into it. When written to a page table entry when SME is not enabled/active, the actual hardware encryption bit would not be set. When reading back the value, since the hardware encryption bit is not set, the translation to set _PAGE_ENC bit won't be done and the comparison to _KERNPG_TABLE would fail. Of course we could just eliminate _PAGE_ENC from the comparison...
Quoted 2 lines
And I think it would collide with 5-level paging.
Does 5-level paging remove bit 57 from the reserved flags?
Quoted 3 lines
I would leave it as variable for now and look on this later once we would have infrastructure to patch constants in kernel text.
If the MK-TME support is going to use the same approach to include the mask/bits real time in _PAGE_ENC then maybe it would be best to get that in first and then look to see if something could be done along the lines of what Linus suggests or with the patchable constants.
Thanks,
Tom
Quoted 1 line
Quoted 30 lines · Tom Lendacky
Quoted 29 lines · Kirill A. Shutemov
Quoted 27 lines · Kai Huang
Quoted 26 lines · Tom Lendacky
Quoted 16 lines · Kirill A. Shutemov
Quoted 7 lines · Tom Lendacky
Quoted 5 lines · Kirill A. Shutemov
AMD SME claims one bit from physical address to indicate whether the
page is encrypted or not. To achieve that we clear out the bit from__PHYSICAL_MASK.Quoted 7 lines
I was actually working on a suggestion by Linus to use one of the software
page table bits to indicate encryption and translate that to the hardware
bit when writing the actual page table entry. With that,__PHYSICAL_MASK
would go back to its original definition.Quoted 3 lines
But you would need to mask it on reading of pfn from page table entry,
right? I expect it to have more overhead than this one.When reading back an entry it would translate the hardware bit position
back to the software bit position. The suggestion for changing it was
to make_PAGE_ENCa constant and not tied to thesme_me_mask.
Quoted 6 lines
Quoted 2 lines
But is it really constant? I thought it's enumerated at boot-time. Can we step onto a problem for future AMD CPUs?
_PAGE_ENCwould be constant and it would be translated to the actual bit that was enumerated at boot-time when writing the page table entry and translated back to_PAGE_ENCwhen reading the page table entry.
I'm confused. I thought that _PAGE_ENC would actual bit that indicate encryption.
If we need translation, I don't see how pte_pfn() can be implemented efficiently. Getting pfn from page table entry is the path that I'm worry about. It's all over the kernel and having overhead there comes with cost.
Could you write a prototype of such patch? I have hard time grasping your idea.
Quoted 10 lines
Quoted 6 lines
In case of MKTME the bits we need to clear are not constant. Depends on CPU and BIOS settings.
By making
_PAGE_ENCconstant we would effectively lower maximum physical address space the kernel can handle, regardless if the system has SME enabled. I can imagine some people wouldn't be happy about this.I don't see how this would lower the maximum physical address space the kernel could handle. Bit 57 is part of the reserved page table flag bits and if SME is not enabled the hardware bits are never used.
My statement came from confusion about _PAGE_ENC value.
Quoted 13 lines
What I do see as a problem is a kernel built with support for SME, and therefore
_PAGE_ENCis not zero, but SME has not been enabled by the BIOS or mem_encrypt=off is specified. In this case you can never be certain that the translation from software bit to hardware bit and back is correct. Take for example,pmd_bad(). Here,_KERNPG_TABLEwould have a non-zero_PAGE_ENCor'd into it. When written to a page table entry when SME is not enabled/active, the actual hardware encryption bit would not be set. When reading back the value, since the hardware encryption bit is not set, the translation to set_PAGE_ENCbit won't be done and the comparison to_KERNPG_TABLEwould fail. Of course we could just eliminate_PAGE_ENCfrom the comparison...Quoted 1 line
Quoted 3 lines
Quoted 1 line
And I think it would collide with 5-level paging.
Does 5-level paging remove bit 57 from the reserved flags?
No. The same confusion.
Quoted 7 lines
Quoted 2 lines
I would leave it as variable for now and look on this later once we would have infrastructure to patch constants in kernel text.
If the MK-TME support is going to use the same approach to include the mask/bits real time in
_PAGE_ENCthen maybe it would be best to get that in first and then look to see if something could be done along the lines of what Linus suggests or with the patchable constants.
Signature · 1 line
Kirill A. Shutemov
Quoted 31 lines · Kai Huang
Quoted 30 lines · Tom Lendacky
Quoted 16 lines · Kirill A. Shutemov
Quoted 7 lines · Tom Lendacky
Quoted 5 lines · Kirill A. Shutemov
AMD SME claims one bit from physical address to indicate whether the
page is encrypted or not. To achieve that we clear out the bit from__PHYSICAL_MASK.Quoted 7 lines
I was actually working on a suggestion by Linus to use one of the software
page table bits to indicate encryption and translate that to the hardware
bit when writing the actual page table entry. With that,__PHYSICAL_MASK
would go back to its original definition.Quoted 3 lines
But you would need to mask it on reading of pfn from page table entry,
right? I expect it to have more overhead than this one.When reading back an entry it would translate the hardware bit position
back to the software bit position. The suggestion for changing it was
to make_PAGE_ENCa constant and not tied to thesme_me_mask.See https://marc.info/?l=linux-kernel&m=151017622615894&w=2
Quoted 1 line
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
And software bits are valuable. Do we still have a spare one for this?
I was looking at possibly using bit 57 (
_PAGE_BIT_SOFTW5).But MK-TME supports upto 15 bits (architectually) as keyID. How is this supposed to work with MK-TME?
I didn't know about MK-TME when I first started looking at this. Is there any way to still use just the single bit to indicate encryption and then have logic that provides the proper bits when actually writing to the page table? I'm not sure what it would take, but it might be worth looking into.
Thanks,
Tom
Quoted 8 lines
Thanks,
-KaiQuoted 5 lines
Thanks,
TomQuoted 1 line