diff mbox series

[edk2,v2,1/4] ArmPkg/CpuDxe ARM: avoid splitting page table sections unnecessarily

Message ID 1488450976-16257-2-git-send-email-ard.biesheuvel@linaro.org
State Superseded
Headers show
Series ArmPkg, ArmVirtpkg ARM: enable strict memory protection | expand

Commit Message

Ard Biesheuvel March 2, 2017, 10:36 a.m. UTC
Currently, any range passed to CpuArchProtocol::SetMemoryAttributes is
fully broken down into page mappings if the start or the size of the
region happens to be misaliged relative to the section size of 1 MB.

This is going to hurt when we enable strict memory permissions, given
that we remap the entire RAM space non-executable (modulo the code
bits) when the CpuArchProtocol is installed.

So refactor the code to iterate over the range in a way that ensures
that all naturally aligned section sized subregions are not broken up.

Contributed-under: TianoCore Contribution Agreement 1.0
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>

---
 ArmPkg/Drivers/CpuDxe/Arm/Mmu.c | 47 ++++++++++++++++----
 1 file changed, 39 insertions(+), 8 deletions(-)

-- 
2.7.4

_______________________________________________
edk2-devel mailing list
edk2-devel@lists.01.org
https://lists.01.org/mailman/listinfo/edk2-devel

Comments

Leif Lindholm March 6, 2017, 2:12 p.m. UTC | #1
On Thu, Mar 02, 2017 at 10:36:13AM +0000, Ard Biesheuvel wrote:
> Currently, any range passed to CpuArchProtocol::SetMemoryAttributes is

> fully broken down into page mappings if the start or the size of the

> region happens to be misaliged relative to the section size of 1 MB.

> 

> This is going to hurt when we enable strict memory permissions, given


"Hurt" -> "cause unnecessary performance penalties" or "hurt" ->
"break everything"?

> that we remap the entire RAM space non-executable (modulo the code

> bits) when the CpuArchProtocol is installed.

> 

> So refactor the code to iterate over the range in a way that ensures

> that all naturally aligned section sized subregions are not broken up.

> 

> Contributed-under: TianoCore Contribution Agreement 1.0

> Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>


Many thanks for getting rid of the magic values, and in general making
the code more logical. One question below:

> ---

>  ArmPkg/Drivers/CpuDxe/Arm/Mmu.c | 47 ++++++++++++++++----

>  1 file changed, 39 insertions(+), 8 deletions(-)

> 

> diff --git a/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c b/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c

> index 89e429925ba9..ce4d529bda67 100644

> --- a/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c

> +++ b/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c

> @@ -679,6 +679,7 @@ SetMemoryAttributes (

>    )

>  {

>    EFI_STATUS    Status;

> +  UINT64        ChunkLength;

>  

>    //

>    // Ignore invocations that only modify permission bits

> @@ -687,14 +688,44 @@ SetMemoryAttributes (

>      return EFI_SUCCESS;

>    }

>  

> -  if(((BaseAddress & 0xFFFFF) == 0) && ((Length & 0xFFFFF) == 0)) {

> -    // Is the base and length a multiple of 1 MB?

> -    DEBUG ((EFI_D_PAGE, "SetMemoryAttributes(): MMU section 0x%x length 0x%x to %lx\n", (UINTN)BaseAddress, (UINTN)Length, Attributes));

> -    Status = UpdateSectionEntries (BaseAddress, Length, Attributes, VirtualMask);

> -  } else {

> -    // Base and/or length is not a multiple of 1 MB

> -    DEBUG ((EFI_D_PAGE, "SetMemoryAttributes(): MMU page 0x%x length 0x%x to %lx\n", (UINTN)BaseAddress, (UINTN)Length, Attributes));

> -    Status = UpdatePageEntries (BaseAddress, Length, Attributes, VirtualMask);

> +  while (Length > 0) {


Would this not end up returning an uninitialized Status if called with
Length == 0?

/
    Leif

> +    if ((BaseAddress % TT_DESCRIPTOR_SECTION_SIZE == 0) &&

> +        Length >= TT_DESCRIPTOR_SECTION_SIZE) {

> +

> +      ChunkLength = Length - Length % TT_DESCRIPTOR_SECTION_SIZE;

> +

> +      DEBUG ((DEBUG_PAGE,

> +        "SetMemoryAttributes(): MMU section 0x%lx length 0x%lx to %lx\n",

> +        BaseAddress, ChunkLength, Attributes));

> +

> +      Status = UpdateSectionEntries (BaseAddress, ChunkLength, Attributes,

> +                 VirtualMask);

> +    } else {

> +

> +      //

> +      // Process page by page until the next section boundary, but only if

> +      // we have more than a section's worth of area to deal with after that.

> +      //

> +      ChunkLength = TT_DESCRIPTOR_SECTION_SIZE -

> +                    (BaseAddress % TT_DESCRIPTOR_SECTION_SIZE);

> +      if (ChunkLength + TT_DESCRIPTOR_SECTION_SIZE > Length) {

> +        ChunkLength = Length;

> +      }

> +

> +      DEBUG ((DEBUG_PAGE,

> +        "SetMemoryAttributes(): MMU page 0x%lx length 0x%lx to %lx\n",

> +        BaseAddress, ChunkLength, Attributes));

> +

> +      Status = UpdatePageEntries (BaseAddress, ChunkLength, Attributes,

> +                 VirtualMask);

> +    }

> +

> +    if (EFI_ERROR (Status)) {

> +      break;

> +    }

> +

> +    BaseAddress += ChunkLength;

> +    Length -= ChunkLength;

>    }

>  

>    // Flush d-cache so descriptors make it back to uncached memory for subsequent table walks

> -- 

> 2.7.4

> 

_______________________________________________
edk2-devel mailing list
edk2-devel@lists.01.org
https://lists.01.org/mailman/listinfo/edk2-devel
Ard Biesheuvel March 6, 2017, 2:55 p.m. UTC | #2
On 6 March 2017 at 15:12, Leif Lindholm <leif.lindholm@linaro.org> wrote:
> On Thu, Mar 02, 2017 at 10:36:13AM +0000, Ard Biesheuvel wrote:

>> Currently, any range passed to CpuArchProtocol::SetMemoryAttributes is

>> fully broken down into page mappings if the start or the size of the

>> region happens to be misaliged relative to the section size of 1 MB.

>>

>> This is going to hurt when we enable strict memory permissions, given

>

> "Hurt" -> "cause unnecessary performance penalties" or "hurt" ->

> "break everything"?

>


The former. It will map all of RAM using page mappings, which uses
more space unnecessarily

>> that we remap the entire RAM space non-executable (modulo the code

>> bits) when the CpuArchProtocol is installed.

>>

>> So refactor the code to iterate over the range in a way that ensures

>> that all naturally aligned section sized subregions are not broken up.

>>

>> Contributed-under: TianoCore Contribution Agreement 1.0

>> Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>

>

> Many thanks for getting rid of the magic values, and in general making

> the code more logical. One question below:

>

>> ---

>>  ArmPkg/Drivers/CpuDxe/Arm/Mmu.c | 47 ++++++++++++++++----

>>  1 file changed, 39 insertions(+), 8 deletions(-)

>>

>> diff --git a/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c b/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c

>> index 89e429925ba9..ce4d529bda67 100644

>> --- a/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c

>> +++ b/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c

>> @@ -679,6 +679,7 @@ SetMemoryAttributes (

>>    )

>>  {

>>    EFI_STATUS    Status;

>> +  UINT64        ChunkLength;

>>

>>    //

>>    // Ignore invocations that only modify permission bits

>> @@ -687,14 +688,44 @@ SetMemoryAttributes (

>>      return EFI_SUCCESS;

>>    }

>>

>> -  if(((BaseAddress & 0xFFFFF) == 0) && ((Length & 0xFFFFF) == 0)) {

>> -    // Is the base and length a multiple of 1 MB?

>> -    DEBUG ((EFI_D_PAGE, "SetMemoryAttributes(): MMU section 0x%x length 0x%x to %lx\n", (UINTN)BaseAddress, (UINTN)Length, Attributes));

>> -    Status = UpdateSectionEntries (BaseAddress, Length, Attributes, VirtualMask);

>> -  } else {

>> -    // Base and/or length is not a multiple of 1 MB

>> -    DEBUG ((EFI_D_PAGE, "SetMemoryAttributes(): MMU page 0x%x length 0x%x to %lx\n", (UINTN)BaseAddress, (UINTN)Length, Attributes));

>> -    Status = UpdatePageEntries (BaseAddress, Length, Attributes, VirtualMask);

>> +  while (Length > 0) {

>

> Would this not end up returning an uninitialized Status if called with

> Length == 0?

>


Yes, well spotted. I will just add an early 'return EFI_SUCCESS' for this case.


>> +    if ((BaseAddress % TT_DESCRIPTOR_SECTION_SIZE == 0) &&

>> +        Length >= TT_DESCRIPTOR_SECTION_SIZE) {

>> +

>> +      ChunkLength = Length - Length % TT_DESCRIPTOR_SECTION_SIZE;

>> +

>> +      DEBUG ((DEBUG_PAGE,

>> +        "SetMemoryAttributes(): MMU section 0x%lx length 0x%lx to %lx\n",

>> +        BaseAddress, ChunkLength, Attributes));

>> +

>> +      Status = UpdateSectionEntries (BaseAddress, ChunkLength, Attributes,

>> +                 VirtualMask);

>> +    } else {

>> +

>> +      //

>> +      // Process page by page until the next section boundary, but only if

>> +      // we have more than a section's worth of area to deal with after that.

>> +      //

>> +      ChunkLength = TT_DESCRIPTOR_SECTION_SIZE -

>> +                    (BaseAddress % TT_DESCRIPTOR_SECTION_SIZE);

>> +      if (ChunkLength + TT_DESCRIPTOR_SECTION_SIZE > Length) {

>> +        ChunkLength = Length;

>> +      }

>> +

>> +      DEBUG ((DEBUG_PAGE,

>> +        "SetMemoryAttributes(): MMU page 0x%lx length 0x%lx to %lx\n",

>> +        BaseAddress, ChunkLength, Attributes));

>> +

>> +      Status = UpdatePageEntries (BaseAddress, ChunkLength, Attributes,

>> +                 VirtualMask);

>> +    }

>> +

>> +    if (EFI_ERROR (Status)) {

>> +      break;

>> +    }

>> +

>> +    BaseAddress += ChunkLength;

>> +    Length -= ChunkLength;

>>    }

>>

>>    // Flush d-cache so descriptors make it back to uncached memory for subsequent table walks

>> --

>> 2.7.4

>>

_______________________________________________
edk2-devel mailing list
edk2-devel@lists.01.org
https://lists.01.org/mailman/listinfo/edk2-devel
Leif Lindholm March 6, 2017, 4:24 p.m. UTC | #3
On Mon, Mar 06, 2017 at 03:55:42PM +0100, Ard Biesheuvel wrote:
> On 6 March 2017 at 15:12, Leif Lindholm <leif.lindholm@linaro.org> wrote:

> > On Thu, Mar 02, 2017 at 10:36:13AM +0000, Ard Biesheuvel wrote:

> >> Currently, any range passed to CpuArchProtocol::SetMemoryAttributes is

> >> fully broken down into page mappings if the start or the size of the

> >> region happens to be misaliged relative to the section size of 1 MB.

> >>

> >> This is going to hurt when we enable strict memory permissions, given

> >

> > "Hurt" -> "cause unnecessary performance penalties" or "hurt" ->

> > "break everything"?

> >

> 

> The former. It will map all of RAM using page mappings, which uses

> more space unnecessarily


OK, can you expand the statement to say "hurt performance" then?

> >> that we remap the entire RAM space non-executable (modulo the code

> >> bits) when the CpuArchProtocol is installed.

> >>

> >> So refactor the code to iterate over the range in a way that ensures

> >> that all naturally aligned section sized subregions are not broken up.

> >>

> >> Contributed-under: TianoCore Contribution Agreement 1.0

> >> Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>

> >

> > Many thanks for getting rid of the magic values, and in general making

> > the code more logical. One question below:

> >

> >> ---

> >>  ArmPkg/Drivers/CpuDxe/Arm/Mmu.c | 47 ++++++++++++++++----

> >>  1 file changed, 39 insertions(+), 8 deletions(-)

> >>

> >> diff --git a/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c b/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c

> >> index 89e429925ba9..ce4d529bda67 100644

> >> --- a/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c

> >> +++ b/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c

> >> @@ -679,6 +679,7 @@ SetMemoryAttributes (

> >>    )

> >>  {

> >>    EFI_STATUS    Status;

> >> +  UINT64        ChunkLength;

> >>

> >>    //

> >>    // Ignore invocations that only modify permission bits

> >> @@ -687,14 +688,44 @@ SetMemoryAttributes (

> >>      return EFI_SUCCESS;

> >>    }

> >>

> >> -  if(((BaseAddress & 0xFFFFF) == 0) && ((Length & 0xFFFFF) == 0)) {

> >> -    // Is the base and length a multiple of 1 MB?

> >> -    DEBUG ((EFI_D_PAGE, "SetMemoryAttributes(): MMU section 0x%x length 0x%x to %lx\n", (UINTN)BaseAddress, (UINTN)Length, Attributes));

> >> -    Status = UpdateSectionEntries (BaseAddress, Length, Attributes, VirtualMask);

> >> -  } else {

> >> -    // Base and/or length is not a multiple of 1 MB

> >> -    DEBUG ((EFI_D_PAGE, "SetMemoryAttributes(): MMU page 0x%x length 0x%x to %lx\n", (UINTN)BaseAddress, (UINTN)Length, Attributes));

> >> -    Status = UpdatePageEntries (BaseAddress, Length, Attributes, VirtualMask);

> >> +  while (Length > 0) {

> >

> > Would this not end up returning an uninitialized Status if called with

> > Length == 0?

> >

> 

> Yes, well spotted. I will just add an early 'return EFI_SUCCESS' for this case.


Works for me. If you also tweak the commit message:
Reviewed-by: Leif Lindholm <leif.lindholm@linaro.org>


> >> +    if ((BaseAddress % TT_DESCRIPTOR_SECTION_SIZE == 0) &&

> >> +        Length >= TT_DESCRIPTOR_SECTION_SIZE) {

> >> +

> >> +      ChunkLength = Length - Length % TT_DESCRIPTOR_SECTION_SIZE;

> >> +

> >> +      DEBUG ((DEBUG_PAGE,

> >> +        "SetMemoryAttributes(): MMU section 0x%lx length 0x%lx to %lx\n",

> >> +        BaseAddress, ChunkLength, Attributes));

> >> +

> >> +      Status = UpdateSectionEntries (BaseAddress, ChunkLength, Attributes,

> >> +                 VirtualMask);

> >> +    } else {

> >> +

> >> +      //

> >> +      // Process page by page until the next section boundary, but only if

> >> +      // we have more than a section's worth of area to deal with after that.

> >> +      //

> >> +      ChunkLength = TT_DESCRIPTOR_SECTION_SIZE -

> >> +                    (BaseAddress % TT_DESCRIPTOR_SECTION_SIZE);

> >> +      if (ChunkLength + TT_DESCRIPTOR_SECTION_SIZE > Length) {

> >> +        ChunkLength = Length;

> >> +      }

> >> +

> >> +      DEBUG ((DEBUG_PAGE,

> >> +        "SetMemoryAttributes(): MMU page 0x%lx length 0x%lx to %lx\n",

> >> +        BaseAddress, ChunkLength, Attributes));

> >> +

> >> +      Status = UpdatePageEntries (BaseAddress, ChunkLength, Attributes,

> >> +                 VirtualMask);

> >> +    }

> >> +

> >> +    if (EFI_ERROR (Status)) {

> >> +      break;

> >> +    }

> >> +

> >> +    BaseAddress += ChunkLength;

> >> +    Length -= ChunkLength;

> >>    }

> >>

> >>    // Flush d-cache so descriptors make it back to uncached memory for subsequent table walks

> >> --

> >> 2.7.4

> >>

_______________________________________________
edk2-devel mailing list
edk2-devel@lists.01.org
https://lists.01.org/mailman/listinfo/edk2-devel
diff mbox series

Patch

diff --git a/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c b/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c
index 89e429925ba9..ce4d529bda67 100644
--- a/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c
+++ b/ArmPkg/Drivers/CpuDxe/Arm/Mmu.c
@@ -679,6 +679,7 @@  SetMemoryAttributes (
   )
 {
   EFI_STATUS    Status;
+  UINT64        ChunkLength;
 
   //
   // Ignore invocations that only modify permission bits
@@ -687,14 +688,44 @@  SetMemoryAttributes (
     return EFI_SUCCESS;
   }
 
-  if(((BaseAddress & 0xFFFFF) == 0) && ((Length & 0xFFFFF) == 0)) {
-    // Is the base and length a multiple of 1 MB?
-    DEBUG ((EFI_D_PAGE, "SetMemoryAttributes(): MMU section 0x%x length 0x%x to %lx\n", (UINTN)BaseAddress, (UINTN)Length, Attributes));
-    Status = UpdateSectionEntries (BaseAddress, Length, Attributes, VirtualMask);
-  } else {
-    // Base and/or length is not a multiple of 1 MB
-    DEBUG ((EFI_D_PAGE, "SetMemoryAttributes(): MMU page 0x%x length 0x%x to %lx\n", (UINTN)BaseAddress, (UINTN)Length, Attributes));
-    Status = UpdatePageEntries (BaseAddress, Length, Attributes, VirtualMask);
+  while (Length > 0) {
+    if ((BaseAddress % TT_DESCRIPTOR_SECTION_SIZE == 0) &&
+        Length >= TT_DESCRIPTOR_SECTION_SIZE) {
+
+      ChunkLength = Length - Length % TT_DESCRIPTOR_SECTION_SIZE;
+
+      DEBUG ((DEBUG_PAGE,
+        "SetMemoryAttributes(): MMU section 0x%lx length 0x%lx to %lx\n",
+        BaseAddress, ChunkLength, Attributes));
+
+      Status = UpdateSectionEntries (BaseAddress, ChunkLength, Attributes,
+                 VirtualMask);
+    } else {
+
+      //
+      // Process page by page until the next section boundary, but only if
+      // we have more than a section's worth of area to deal with after that.
+      //
+      ChunkLength = TT_DESCRIPTOR_SECTION_SIZE -
+                    (BaseAddress % TT_DESCRIPTOR_SECTION_SIZE);
+      if (ChunkLength + TT_DESCRIPTOR_SECTION_SIZE > Length) {
+        ChunkLength = Length;
+      }
+
+      DEBUG ((DEBUG_PAGE,
+        "SetMemoryAttributes(): MMU page 0x%lx length 0x%lx to %lx\n",
+        BaseAddress, ChunkLength, Attributes));
+
+      Status = UpdatePageEntries (BaseAddress, ChunkLength, Attributes,
+                 VirtualMask);
+    }
+
+    if (EFI_ERROR (Status)) {
+      break;
+    }
+
+    BaseAddress += ChunkLength;
+    Length -= ChunkLength;
   }
 
   // Flush d-cache so descriptors make it back to uncached memory for subsequent table walks