Skip to contentSkip to conversation

ARM: remove the .vm_mm value from gate_vma.

PATCH New Steve Capper 3 participants 4 messages May 16, 2013 – May 17, 2013 Arm (32-bit)

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ARM: remove the .vm_mm value from gate_vma.

If one reads /proc/$PID/smaps, the mmap_sem belonging to the address space of the task being examined is locked for reading. All the pages of the vmas belonging to the task's address space are then walked with this lock held.

If a gate_vma is present in the architecture, it too is examined by the fs/proc/task_mmu.c code. As gate_vma doesn't belong to the address space of the task though, its pages are not walked.

A recent cleanup (commit f6604efe) of the gate_vma initialisation code set the vm_mm value to &init_mm. Unfortunately a non-NULL vm_mm value in the gate_vma will cause the task_mmu code to attempt to walk the pages of the gate_vma (with no mmap-sem lock held). If one enables Transparent Huge Page support and vm debugging, this will then cause OOPses as pmd_trans_huge_lock is called without mmap_sem being locked.

This patch removes the .vm_mm value from gate_vma, restoring the original behaviour of the task_mmu code.

---
 arch/arm/kernel/process.c | 1 -
 1 file changed, 1 deletion(-)
arch/arm/kernel/process.c +0 -1
index f219703..282de48 100644
@@ -411,7 +411,6 @@ static struct vm_area_struct gate_vma = {
 	.vm_start	= 0xffff0000,
 	.vm_end		= 0xffff0000 + PAGE_SIZE,
 	.vm_flags	= VM_READ | VM_EXEC | VM_MAYREAD | VM_MAYEXEC,
-	.vm_mm		= &init_mm,
 };
 
 static int __init gate_vma_init(void)
Signature · 1 line
1.8.1.4
#1 lore raw

PATCH 1/1 ARM: remove the .vm_mm value from gate_vma.

Quoted 17 lines · Steve Capper

If one reads /proc/$PID/smaps, the mmap_sem belonging to the address space of the task being examined is locked for reading. All the pages of the vmas belonging to the task's address space are then walked with this lock held.

If a gate_vma is present in the architecture, it too is examined by the fs/proc/task_mmu.c code. As gate_vma doesn't belong to the address space of the task though, its pages are not walked.

A recent cleanup (commit f6604efe) of the gate_vma initialisation code set the vm_mm value to &init_mm. Unfortunately a non-NULL vm_mm value in the gate_vma will cause the task_mmu code to attempt to walk the pages of the gate_vma (with no mmap-sem lock held). If one enables Transparent Huge Page support and vm debugging, this will then cause OOPses as pmd_trans_huge_lock is called without mmap_sem being locked.

Quoted 2 lines

This patch removes the .vm_mm value from gate_vma, restoring the original behaviour of the task_mmu code.

This looks fine - x86 also sets .vm_mm to NULL here. So please put it in my patch system. Does this need to be applied to stable kernels as well? If so, please mark it with Cc: <[email protected]> (but do not send it to that address). Also, it'd be worth checking which stable kernels it needs to be applied to.

Thanks.

#2 lore raw

PATCH 1/1 ARM: remove the .vm_mm value from gate_vma.

Quoted 1 line · Russell King - ARM Linux

On Thu, May 16, 2013 at 12:06:34PM +0100, Steve Capper wrote:

[ ... ]

Quoted 8 lines

This looks fine - x86 also sets .vm_mm to NULL here. So please put it in my patch system. Does this need to be applied to stable kernels as well? If so, please mark it with Cc: <[email protected]> (but do not send it to that address). Also, it'd be worth checking which stable kernels it needs to be applied to.

Thanks.

Thank you Russell,
I couldn't see any branches in linux-stable that need this patch. The patch is now in your system with id 7727/1.

Cheers,

Signature · 1 line
Steve
#3 lore raw

PATCH 1/1 ARM: remove the .vm_mm value from gate_vma.

In reply to Steve Capper
Quoted 30 lines · Steve Capper

If one reads /proc/$PID/smaps, the mmap_sem belonging to the address space of the task being examined is locked for reading. All the pages of the vmas belonging to the task's address space are then walked with this lock held.

If a gate_vma is present in the architecture, it too is examined by the fs/proc/task_mmu.c code. As gate_vma doesn't belong to the address space of the task though, its pages are not walked.

A recent cleanup (commit f6604efe) of the gate_vma initialisation code set the vm_mm value to &init_mm. Unfortunately a non-NULL vm_mm value in the gate_vma will cause the task_mmu code to attempt to walk the pages of the gate_vma (with no mmap-sem lock held). If one enables Transparent Huge Page support and vm debugging, this will then cause OOPses as pmd_trans_huge_lock is called without mmap_sem being locked.

This patch removes the .vm_mm value from gate_vma, restoring the original behaviour of the task_mmu code.

---
 arch/arm/kernel/process.c | 1 -
 1 file changed, 1 deletion(-)
arch/arm/kernel/process.c +0 -0
index f219703..282de48 100644

arch/arm/kernel/process.c static struct vm_area_struct gate_vma = { · line 416

Quoted 7 lines
arch/arm/kernel/process.c +0 -1
@@ -411,7 +411,6 @@ static struct vm_area_struct gate_vma = {
 	.vm_start	= 0xffff0000,
 	.vm_end		= 0xffff0000 + PAGE_SIZE,
 	.vm_flags	= VM_READ | VM_EXEC | VM_MAYREAD | VM_MAYEXEC,
-	.vm_mm		= &init_mm,
 };
 
 static int __init gate_vma_init(void)

Thanks to the wonders of Mimecrap, which decided to sit on most of my email yesterday, I only just received this patch.

I see it's in the patch system, but for the record:

Cheers,

Will