Skip to contentSkip to conversation

powerpc/perf: Use SIER_USER_MASK while updating SPRN_SIER for EBB events

PATCH In review Athira Rajeev 2 participants 5 messages Mar 18, 2020 – Jul 15, 2020 PowerPC

lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 powerpc/perf: Use SIER_USER_MASK while updating SPRN_SIER for EBB events

Sampled Instruction Event Register (SIER), is a PMU register, captures architecture state for a given sample. And sier_user_mask defined in commit 330a1eb7775b ("powerpc/perf: Core EBB support for 64-bit book3s") defines the architected bits that needs to be saved from the SPR. Currently all of the bits from SIER are saved for EBB events. Patch fixes this by ANDing the "sier_user_mask" to data from SIER in ebb_switch_out(). This will force save only architected bits from the SIER.

Fixes 330a1eb7775b ("powerpc/perf: Core EBB support for 64-bit book3s")

---
Changelog:
v2 -> v3:
- Corrected name of SIER register in commit message as pointed by Segher Boessenkool
v1 -> v2:
- Make the commit message more clearer.

 arch/powerpc/perf/core-book3s.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/powerpc/perf/core-book3s.c +1 -1
index 3086055..48b61cc 100644
@@ -579,7 +579,7 @@ static void ebb_switch_out(unsigned long mmcr0)
 		return;
 
 	current->thread.siar  = mfspr(SPRN_SIAR);
-	current->thread.sier  = mfspr(SPRN_SIER);
+	current->thread.sier  = mfspr(SPRN_SIER) & SIER_USER_MASK;
 	current->thread.sdar  = mfspr(SPRN_SDAR);
 	current->thread.mmcr0 = mmcr0 & MMCR0_USER_MASK;
 	current->thread.mmcr2 = mfspr(SPRN_MMCR2) & MMCR2_USER_MASK;
Signature · 1 line
1.8.3.1
#1 lore raw

Hi Athira,

Quoted 1 line · Athira Rajeev

Sampled Instruction Event Register (SIER), is a PMU register,

                                                               ^
                                                               that
Quoted 1 line

captures architecture state for a given sample. And sier_user_mask

^ ^
don't think we need "architecture" SIER_USER_MASK

Quoted 2 lines

defined in commit 330a1eb7775b ("powerpc/perf: Core EBB support for 64-bit book3s") defines the architected bits that needs to be saved from the SPR.

Not quite, it defines the bits that are visible to userspace.

And I think it's true that for EBB events the bits we need/want to save are only the user visible bits.

Quoted 3 lines

Currently all of the bits from SIER are saved for EBB events. Patch fixes this by ANDing the "sier_user_mask" to data from SIER in ebb_switch_out(). This will force save only architected bits from the SIER.

s/architected/user visible/

But, why does it matter? The kernel saves the user visible bits, as well as the kernel-only bits into the thread struct. And then later the kernel restores that value into the hardware before returning to userspace.

But the hardware enforces the visibility of the bits, so userspace can't observe any bits that it shouldn't.

Or is there some other mechanism whereby userspace can see those bits? ;)

If there was, what would the security implications of that be?

cheers

Quoted 28 lines
Fixes 330a1eb7775b ("powerpc/perf: Core EBB support for 64-bit book3s")

---
Changelog:
v2 -> v3:
- Corrected name of SIER register in commit message as pointed by Segher Boessenkool
v1 -> v2:
- Make the commit message more clearer.

 arch/powerpc/perf/core-book3s.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/powerpc/perf/core-book3s.c +1 -1
index 3086055..48b61cc 100644
@@ -579,7 +579,7 @@ static void ebb_switch_out(unsigned long mmcr0)
 		return;
 
 	current->thread.siar  = mfspr(SPRN_SIAR);
-	current->thread.sier  = mfspr(SPRN_SIER);
+	current->thread.sier  = mfspr(SPRN_SIER) & SIER_USER_MASK;
 	current->thread.sdar  = mfspr(SPRN_SDAR);
 	current->thread.mmcr0 = mmcr0 & MMCR0_USER_MASK;
 	current->thread.mmcr2 = mfspr(SPRN_MMCR2) & MMCR2_USER_MASK;
Signature · 1 line
1.8.3.1
#2 lore raw
Quoted 34 lines · Michael Ellerman

Hi Athira,

Quoted 1 line · Athira Rajeev

Sampled Instruction Event Register (SIER), is a PMU register,

                                                              ^
                                                              that
Quoted 1 line

captures architecture state for a given sample. And sier_user_mask

^ ^
don't think we need "architecture" SIER_USER_MASK

Quoted 2 lines

defined in commit 330a1eb7775b ("powerpc/perf: Core EBB support for 64-bit book3s") defines the architected bits that needs to be saved from the SPR.

Not quite, it defines the bits that are visible to userspace.

And I think it's true that for EBB events the bits we need/want to save are only the user visible bits.

Quoted 3 lines

Currently all of the bits from SIER are saved for EBB events. Patch fixes this by ANDing the "sier_user_mask" to data from SIER in ebb_switch_out(). This will force save only architected bits from the SIER.

s/architected/user visible/

But, why does it matter? The kernel saves the user visible bits, as well as the kernel-only bits into the thread struct. And then later the kernel restores that value into the hardware before returning to userspace.

Quoted 6 lines

But the hardware enforces the visibility of the bits, so userspace can't observe any bits that it shouldn't.

Or is there some other mechanism whereby userspace can see those bits? ;)

If there was, what would the security implications of that be?

Hi Michael,

Thanks for your comments.

In ebb_switch_in, we set PMCC bit [MMCR0 44:45 ] to 10 which means SIER ( Group B ) register is readable in problem state. Hence the intention of the patch was to make sure we are not exposing the bits which the userspace shouldn't be reading.

But following your comment about "hardware enforcing the visibility of bits", I did try an "ebb" experiment which showed that reading SPRN_SIER didn't expose any bits other than the user visible bits. Sorry for the confusion here.

In that case, Can we drop the existing definition of SIER_USER_MASK if it is no more needed ?

Thanks
Athira

Quoted 31 lines

cheers

Quoted 28 lines
Fixes 330a1eb7775b ("powerpc/perf: Core EBB support for 64-bit book3s")

---
Changelog:
v2 -> v3:
- Corrected name of SIER register in commit message as pointed by Segher Boessenkool
v1 -> v2:
- Make the commit message more clearer.

arch/powerpc/perf/core-book3s.c | 2 +-

1 file changed, 1 insertion(+), 1 deletion(-)
arch/powerpc/perf/core-book3s.c +1 -1
index 3086055..48b61cc 100644
@@ -579,7 +579,7 @@ static void ebb_switch_out(unsigned long mmcr0)
 		return;
 
 	current->thread.siar  = mfspr(SPRN_SIAR);
-	current->thread.sier  = mfspr(SPRN_SIER);
+	current->thread.sier  = mfspr(SPRN_SIER) & SIER_USER_MASK;
 	current->thread.sdar  = mfspr(SPRN_SDAR);
 	current->thread.mmcr0 = mmcr0 & MMCR0_USER_MASK;
 	current->thread.mmcr2 = mfspr(SPRN_MMCR2) & MMCR2_USER_MASK;
Signature · 1 line
1.8.3.1
#3 lore raw
Quoted 48 lines · Athira Rajeev
Quoted 32 lines · Michael Ellerman

Hi Athira,

Quoted 1 line · Athira Rajeev

Sampled Instruction Event Register (SIER), is a PMU register,

                                                              ^
                                                              that
Quoted 1 line

captures architecture state for a given sample. And sier_user_mask

^ ^
don't think we need "architecture" SIER_USER_MASK

Quoted 2 lines

defined in commit 330a1eb7775b ("powerpc/perf: Core EBB support for 64-bit book3s") defines the architected bits that needs to be saved from the SPR.

Not quite, it defines the bits that are visible to userspace.

And I think it's true that for EBB events the bits we need/want to save are only the user visible bits.

Quoted 3 lines

Currently all of the bits from SIER are saved for EBB events. Patch fixes this by ANDing the "sier_user_mask" to data from SIER in ebb_switch_out(). This will force save only architected bits from the SIER.

s/architected/user visible/

But, why does it matter? The kernel saves the user visible bits, as well as the kernel-only bits into the thread struct. And then later the kernel restores that value into the hardware before returning to userspace.

Quoted 6 lines

But the hardware enforces the visibility of the bits, so userspace can't observe any bits that it shouldn't.

Or is there some other mechanism whereby userspace can see those bits? ;)

If there was, what would the security implications of that be?

Hi Michael,

Thanks for your comments.

In ebb_switch_in, we set PMCC bit [MMCR0 44:45 ] to 10 which means SIER ( Group B ) register is readable in problem state. Hence the intention of the patch was to make sure we are not exposing the bits which the userspace shouldn't be reading.

Quoted 4 lines

But following your comment about "hardware enforcing the visibility of bits", I did try an "ebb" experiment which showed that reading SPRN_SIER didn't expose any bits other than the user visible bits. Sorry for the confusion here.

That's OK. Thanks for following my trail of clues :)

Quoted 2 lines

In that case, Can we drop the existing definition of SIER_USER_MASK if it is no more needed ?

I think it is still needed, and I think this change to use it is good, because SIER is visible via ptrace.

What we need to do, is look at what information in SIER we are currently exposing to userspace via ptrace, and what the security implications (if any) of that are.

cheers

#4 lore raw
Quoted 53 lines · Michael Ellerman
Quoted 48 lines · Athira Rajeev <[email protected]…
Quoted 32 lines · Michael Ellerman

Hi Athira,

Quoted 1 line · Athira Rajeev

Sampled Instruction Event Register (SIER), is a PMU register,

                                                             ^
                                                             that
Quoted 1 line

captures architecture state for a given sample. And sier_user_mask

^ ^
don't think we need "architecture" SIER_USER_MASK

Quoted 2 lines

defined in commit 330a1eb7775b ("powerpc/perf: Core EBB support for 64-bit book3s") defines the architected bits that needs to be saved from the SPR.

Not quite, it defines the bits that are visible to userspace.

And I think it's true that for EBB events the bits we need/want to save are only the user visible bits.

Quoted 3 lines

Currently all of the bits from SIER are saved for EBB events. Patch fixes this by ANDing the "sier_user_mask" to data from SIER in ebb_switch_out(). This will force save only architected bits from the SIER.

s/architected/user visible/

But, why does it matter? The kernel saves the user visible bits, as well as the kernel-only bits into the thread struct. And then later the kernel restores that value into the hardware before returning to userspace.

Quoted 6 lines

But the hardware enforces the visibility of the bits, so userspace can't observe any bits that it shouldn't.

Or is there some other mechanism whereby userspace can see those bits? ;)

If there was, what would the security implications of that be?

Hi Michael,

Thanks for your comments.

In ebb_switch_in, we set PMCC bit [MMCR0 44:45 ] to 10 which means SIER ( Group B ) register is readable in problem state. Hence the intention of the patch was to make sure we are not exposing the bits which the userspace shouldn't be reading.

Quoted 7 lines
Quoted 4 lines

But following your comment about "hardware enforcing the visibility of bits", I did try an "ebb" experiment which showed that reading SPRN_SIER didn't expose any bits other than the user visible bits. Sorry for the confusion here.

That's OK. Thanks for following my trail of clues :)

Sure, Thanks for your inputs. I will come back on this after checking on information exposed via Ptrace

Quoted 11 lines
Quoted 2 lines

In that case, Can we drop the existing definition of SIER_USER_MASK if it is no more needed ?

I think it is still needed, and I think this change to use it is good, because SIER is visible via ptrace.

What we need to do, is look at what information in SIER we are currently exposing to userspace via ptrace, and what the security implications (if any) of that are.

cheers