[PATCH] arm64: mm: allow preemption in copy_to_user_page

[PATCH] arm64: mm: allow preemption in copy_to_user_page

[Mark Rutland]

Currently we disable preemption in copy_to_user_page; a behaviour that
we inherited from the 32-bit arm code. This was necessary for older
cores without broadcast data cache maintenance, and ensured that cache
lines were dirtied and cleaned by the same CPU. On these systems dirty
cache line migration was not possible, so this was sufficient to
guarantee coherency.

On contemporary systems, cache coherence protocols permit (dirty) cache
lines to migrate between CPUs as a result of speculation, prefetching,
and other behaviours. To account for this, in ARMv8 data cache
maintenance operations are broadcast and affect all data caches in the
domain associated with the VA (i.e. ISH for kernel and user mappings).

In __switch_to we ensure that tasks can be safely migrated in the middle
of a maintenance sequence, using a dsb(ish) to ensure prior explicit
memory accesses are observed and cache maintenance operations are
completed before a task can be run on another CPU.

Given the above, it is not necessary to disable preemption in
copy_to_user_page. This patch removes the preempt_{disable,enable}
calls, permitting preemption.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Will Deacon <will.deacon@arm.com>
---
 arch/arm64/mm/flush.c | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/arch/arm64/mm/flush.c b/arch/arm64/mm/flush.c
index 60585bd..dbd12ea 100644
--- a/arch/arm64/mm/flush.c
+++ b/arch/arm64/mm/flush.c
@@ -58,17 +58,13 @@ static void flush_ptrace_access(struct vm_area_struct *vma, struct page *page,
  * Copy user data from/to a page which is mapped into a different processes
  * address space.  Really, we want to allow our "user space" model to handle
  * this.
- *
- * Note that this code needs to run on the current CPU.
  */
 void copy_to_user_page(struct vm_area_struct *vma, struct page *page,
 		       unsigned long uaddr, void *dst, const void *src,
 		       unsigned long len)
 {
-	preempt_disable();
 	memcpy(dst, src, len);
 	flush_ptrace_access(vma, page, uaddr, dst, len);
-	preempt_enable();
 }
 
 void __sync_icache_dcache(pte_t pte, unsigned long addr)
-- 
1.9.1

[PATCH] arm64: mm: allow preemption in copy_to_user_page

[Catalin Marinas]

On Tue, Mar 22, 2016 at 10:11:28AM +0000, Mark Rutland wrote:
> Currently we disable preemption in copy_to_user_page; a behaviour that
> we inherited from the 32-bit arm code. This was necessary for older
> cores without broadcast data cache maintenance, and ensured that cache
> lines were dirtied and cleaned by the same CPU. On these systems dirty
> cache line migration was not possible, so this was sufficient to
> guarantee coherency.
> 
> On contemporary systems, cache coherence protocols permit (dirty) cache
> lines to migrate between CPUs as a result of speculation, prefetching,
> and other behaviours. To account for this, in ARMv8 data cache
> maintenance operations are broadcast and affect all data caches in the
> domain associated with the VA (i.e. ISH for kernel and user mappings).
> 
> In __switch_to we ensure that tasks can be safely migrated in the middle
> of a maintenance sequence, using a dsb(ish) to ensure prior explicit
> memory accesses are observed and cache maintenance operations are
> completed before a task can be run on another CPU.
> 
> Given the above, it is not necessary to disable preemption in
> copy_to_user_page. This patch removes the preempt_{disable,enable}
> calls, permitting preemption.
> 
> Signed-off-by: Mark Rutland <mark.rutland@arm.com>
> Cc: Catalin Marinas <catalin.marinas@arm.com>
> Cc: Will Deacon <will.deacon@arm.com>

Applied. Thanks.

-- 
Catalin