Skip to contentSkip to conversation

x86: fpu xstate fix memory poison overwritten

PATCH In review Xiaotian Feng 5 participants 8 messages Aug 12, 2010 – Aug 16, 2010

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

PATCH 1/1 x86: fpu xstate fix memory poison overwritten

fpu.state is allocated from task_xstate_cachep, the size of task_xstate_cachep is xstate_size. But fpu.state is an union struct, which size is bigger than xstate_size if cpu_has_xsave, so if we want to visit fpu.state->xsave, the memory we allocated for fpu.state is not enough.

This caused many poison/redzone overwritten alerts on task_xstate while using kvm.

[ 1899.399373] =============================================================================
[ 1899.399377] BUG task_xstate: Poison overwritten
[ 1899.399378] -----------------------------------------------------------------------------
[ 1899.399379]
[ 1899.399382] INFO: 0xffff88020aca2100-0xffff88020aca217f. First byte 0x0 instead of 0x6b
[ 1899.399385] INFO: Slab 0xffffea000725c300 objects=23 used=12 fp=0xffff88020aca2100 flags=0x200000000040c1
[ 1899.399387] INFO: Object 0xffff88020aca2100 @offset=8448 fp=0xffff88020aca23c0

With this patch applied, the poison overwritten alert disappeared.

Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Brian Gerst <[email protected]>
Cc Avi Kivity <[email protected]>
Cc Peter Zijlstra <[email protected]>
---
 arch/x86/kernel/process.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/process.c +1 -1
index d401f1d..609bee5 100644
@@ -59,7 +59,7 @@ void free_thread_info(struct thread_info *ti)
 void arch_task_cache_init(void)
 {
         task_xstate_cachep =
-        	kmem_cache_create("task_xstate", xstate_size,
+		kmem_cache_create("task_xstate", sizeof(union thread_xstate),
 				  __alignof__(union thread_xstate),
 				  SLAB_PANIC | SLAB_NOTRACK, NULL);
 }
Signature · 1 line
1.7.2.1
#1 lore raw
Quoted 36 lines · Xiaotian Feng

fpu.state is allocated from task_xstate_cachep, the size of task_xstate_cachep is xstate_size. But fpu.state is an union struct, which size is bigger than xstate_size if cpu_has_xsave, so if we want to visit fpu.state->xsave, the memory we allocated for fpu.state is not enough.

This caused many poison/redzone overwritten alerts on task_xstate while using kvm.

[ 1899.399373] =============================================================================
[ 1899.399377] BUG task_xstate: Poison overwritten
[ 1899.399378] -----------------------------------------------------------------------------
[ 1899.399379]
[ 1899.399382] INFO: 0xffff88020aca2100-0xffff88020aca217f. First byte 0x0 instead of 0x6b
[ 1899.399385] INFO: Slab 0xffffea000725c300 objects=23 used=12 fp=0xffff88020aca2100 flags=0x200000000040c1
[ 1899.399387] INFO: Object 0xffff88020aca2100 @offset=8448 fp=0xffff88020aca23c0

With this patch applied, the poison overwritten alert disappeared.

Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Brian Gerst <[email protected]>
Cc Avi Kivity <[email protected]>
Cc Peter Zijlstra <[email protected]>

---
 arch/x86/kernel/process.c |    2 +-

 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/process.c +0 -0
index d401f1d..609bee5 100644
@@ -59,7 +59,7 @@ void free_thread_info(struct thread_info *ti)
  void arch_task_cache_init(void)
  {
Quoted 6 lines

        task_xstate_cachep =
-               kmem_cache_create("task_xstate", xstate_size,
+               kmem_cache_create("task_xstate", sizeof(union thread_xstate), __alignof__(union thread_xstate), SLAB_PANIC | SLAB_NOTRACK, NULL); }

This isn't the right solution. The point of having a variable xstate_size is to only allocate the memory that is needed for the capabilities of the cpu. The real bug is that some code is trying to use xsave_struct when it either shouldn't be (because the cpu doesn't support xsave), or because xstate_size < sizeof(xsave_struct).

Looking at xstate_enable_boot_cpu(), the xstate_size is set from cpuid. It's possible that it is smaller than sizeof(xsave_strruct). Maybe a sanity check should be added there?

--
Brian Gerst

#2 lore raw

PATCH 1/1 kvm: fix poison overwritten caused by using wrong xstate size

fpu.state is allocated from task_xstate_cachep, the size of task_xstate_cachep is xstate_size. xstate_size is set from cpuid instruction, which is often smaller than sizeof(struct xsave_struct). kvm is using sizeof(struct xsave_struct) to fill in/out fpu.state.xsave, as what we allocated for fpu.state is xstate_size, kernel will write out of memory and caused poison/redzone/padding overwritten warnings.

Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Suresh Siddha <[email protected]>
Cc Brian Gerst <[email protected]>
Cc Avi Kivity <[email protected]>
Cc Robert Richter <[email protected]>
Cc Sheng Yang <[email protected]>
Cc Marcelo Tosatti <[email protected]>
Cc Gleb Natapov <[email protected]>
Cc Jan Kiszka <[email protected]>
---
 arch/x86/kernel/i387.c |    1 +
 arch/x86/kvm/x86.c     |    4 ++--
 2 files changed, 3 insertions(+), 2 deletions(-)
arch/x86/kernel/i387.c +1 -0
index 1f11f5c..a46cb35 100644
@@ -40,6 +40,7 @@
 
 static unsigned int		mxcsr_feature_mask __read_mostly = 0xffffffffu;
 unsigned int xstate_size;
+EXPORT_SYMBOL_GPL(xstate_size);
 unsigned int sig_xstate_ia32_size = sizeof(struct _fpstate_ia32);
 static struct i387_fxsave_struct fx_scratch __cpuinitdata;
 
arch/x86/kvm/x86.c +2 -2
index 25f1907..3a09c62 100644
@@ -2387,7 +2387,7 @@ static void kvm_vcpu_ioctl_x86_get_xsave(struct kvm_vcpu *vcpu,
 	if (cpu_has_xsave)
 		memcpy(guest_xsave->region,
 			&vcpu->arch.guest_fpu.state->xsave,
-			sizeof(struct xsave_struct));
+			xstate_size);
 	else {
 		memcpy(guest_xsave->region,
 			&vcpu->arch.guest_fpu.state->fxsave,
@@ -2405,7 +2405,7 @@ static int kvm_vcpu_ioctl_x86_set_xsave(struct kvm_vcpu *vcpu,
 
 	if (cpu_has_xsave)
 		memcpy(&vcpu->arch.guest_fpu.state->xsave,
-			guest_xsave->region, sizeof(struct xsave_struct));
+			guest_xsave->region, xstate_size);
 	else {
 		if (xstate_bv & ~XSTATE_FPSSE)
 			return -EINVAL;
Signature · 1 line
1.7.2.1
Reviewed-by on v1 1/1
#3 lore raw

Re: [PATCH] kvm: fix poison overwritten caused by using wrong xstate size

Quoted 7 lines · On Friday 13 August 2010 15:19:11 Xiaotian Feng wrote:

fpu.state is allocated from task_xstate_cachep, the size of task_xstate_cachep is xstate_size. xstate_size is set from cpuid instruction, which is often smaller than sizeof(struct xsave_struct). kvm is using sizeof(struct xsave_struct) to fill in/out fpu.state.xsave, as what we allocated for fpu.state is xstate_size, kernel will write out of memory and caused poison/redzone/padding overwritten warnings.

Quoted 12 lines
Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Suresh Siddha <[email protected]>
Cc Brian Gerst <[email protected]>
Cc Avi Kivity <[email protected]>
Cc Robert Richter <[email protected]>
Cc Sheng Yang <[email protected]>
Cc Marcelo Tosatti <[email protected]>
Cc Gleb Natapov <[email protected]>
Cc Jan Kiszka <[email protected]>

--
regards
Yang, Sheng

Quoted 40 lines
---
 arch/x86/kernel/i387.c |    1 +
 arch/x86/kvm/x86.c     |    4 ++--
 2 files changed, 3 insertions(+), 2 deletions(-)
arch/x86/kernel/i387.c +1 -0
index 1f11f5c..a46cb35 100644
@@ -40,6 +40,7 @@
 
 static unsigned int		mxcsr_feature_mask __read_mostly = 0xffffffffu;
 unsigned int xstate_size;
+EXPORT_SYMBOL_GPL(xstate_size);
 unsigned int sig_xstate_ia32_size = sizeof(struct _fpstate_ia32);
 static struct i387_fxsave_struct fx_scratch __cpuinitdata;
 
arch/x86/kvm/x86.c +2 -2
index 25f1907..3a09c62 100644
@@ -2387,7 +2387,7 @@ static void kvm_vcpu_ioctl_x86_get_xsave(struct
 kvm_vcpu *vcpu, if (cpu_has_xsave)
 		memcpy(guest_xsave->region,
 			&vcpu->arch.guest_fpu.state->xsave,
-			sizeof(struct xsave_struct));
+			xstate_size);
 	else {
 		memcpy(guest_xsave->region,
 			&vcpu->arch.guest_fpu.state->fxsave,
@@ -2405,7 +2405,7 @@ static int kvm_vcpu_ioctl_x86_set_xsave(struct
 kvm_vcpu *vcpu,
 
 	if (cpu_has_xsave)
 		memcpy(&vcpu->arch.guest_fpu.state->xsave,
-			guest_xsave->region, sizeof(struct xsave_struct));
+			guest_xsave->region, xstate_size);
 	else {
 		if (xstate_bv & ~XSTATE_FPSSE)

return -EINVAL;

#4 lore raw

Re: [PATCH] kvm: fix poison overwritten caused by using wrong xstate size

In reply to Xiaotian Feng

Avi, do you want to take this one or should I?

-hpa

Quoted 58 lines · Xiaotian Feng

fpu.state is allocated from task_xstate_cachep, the size of task_xstate_cachep is xstate_size. xstate_size is set from cpuid instruction, which is often smaller than sizeof(struct xsave_struct). kvm is using sizeof(struct xsave_struct) to fill in/out fpu.state.xsave, as what we allocated for fpu.state is xstate_size, kernel will write out of memory and caused poison/redzone/padding overwritten warnings.

Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Suresh Siddha <[email protected]>
Cc Brian Gerst <[email protected]>
Cc Avi Kivity <[email protected]>
Cc Robert Richter <[email protected]>
Cc Sheng Yang <[email protected]>
Cc Marcelo Tosatti <[email protected]>
Cc Gleb Natapov <[email protected]>
Cc Jan Kiszka <[email protected]>
---
 arch/x86/kernel/i387.c |    1 +
 arch/x86/kvm/x86.c     |    4 ++--
 2 files changed, 3 insertions(+), 2 deletions(-)
arch/x86/kernel/i387.c +1 -0
index 1f11f5c..a46cb35 100644
@@ -40,6 +40,7 @@
 
 static unsigned int		mxcsr_feature_mask __read_mostly = 0xffffffffu;
 unsigned int xstate_size;
+EXPORT_SYMBOL_GPL(xstate_size);
 unsigned int sig_xstate_ia32_size = sizeof(struct _fpstate_ia32);
 static struct i387_fxsave_struct fx_scratch __cpuinitdata;
 
arch/x86/kvm/x86.c +2 -2
index 25f1907..3a09c62 100644
@@ -2387,7 +2387,7 @@ static void kvm_vcpu_ioctl_x86_get_xsave(struct kvm_vcpu *vcpu,
 	if (cpu_has_xsave)
 		memcpy(guest_xsave->region,
 			&vcpu->arch.guest_fpu.state->xsave,
-			sizeof(struct xsave_struct));
+			xstate_size);
 	else {
 		memcpy(guest_xsave->region,
 			&vcpu->arch.guest_fpu.state->fxsave,
@@ -2405,7 +2405,7 @@ static int kvm_vcpu_ioctl_x86_set_xsave(struct kvm_vcpu *vcpu,
 
 	if (cpu_has_xsave)
 		memcpy(&vcpu->arch.guest_fpu.state->xsave,
-			guest_xsave->region, sizeof(struct xsave_struct));
+			guest_xsave->region, xstate_size);
 	else {
 		if (xstate_bv & ~XSTATE_FPSSE)
 			return -EINVAL;
#5 lore raw

Re: [PATCH] kvm: fix poison overwritten caused by using wrong xstate size

Quoted 1 line · H. Peter Anvin

Avi, do you want to take this one or should I?

I will, thanks.

Signature · 1 line
error compiling committee.c: too many arguments to function
#6 lore raw

Re: [PATCH] kvm: fix poison overwritten caused by using wrong xstate size

Feel free to add my ack.

Quoted 8 lines · Avi Kivity
Quoted 1 line · H. Peter Anvin

Avi, do you want to take this one or should I?

I will, thanks.

Signature · 1 line
error compiling committee.c: too many arguments to function
Signature · 1 line
Sent from my mobile phone.  Please pardon any lack of formatting.
Applied
#7 lore raw

Re: [PATCH] kvm: fix poison overwritten caused by using wrong xstate size

In reply to Xiaotian Feng
Quoted 6 lines · Xiaotian Feng

fpu.state is allocated from task_xstate_cachep, the size of task_xstate_cachep is xstate_size. xstate_size is set from cpuid instruction, which is often smaller than sizeof(struct xsave_struct). kvm is using sizeof(struct xsave_struct) to fill in/out fpu.state.xsave, as what we allocated for fpu.state is xstate_size, kernel will write out of memory and caused poison/redzone/padding overwritten warnings.

Thanks, applied.

Signature · 1 line
error compiling committee.c: too many arguments to function