lkml.org 
[lkml]   [2018]   [May]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] perf/ring_buffer: ensure atomicity and order of updates
On Thu, May 10, 2018 at 02:06:32PM +0100, Mark Rutland wrote:
> Userspace can read/write the user page at any point in time, and in
> perf_output_put_handle() we're very careful to use memory barriers to
> ensure ordering between updates to data and the user page.
>
> We don't use barriers when updating aux_head, where similar ordering
> constraints apply. This could result in userspace seeing stale data, or
> data being overwritten while userspace was still consuming it.
>
> Further, we update data_head and aux_head with plain assignments, which
> the compiler can tear, potentially resulting in userspace seeing
> erroneous values.
>
> We can solve both of these problems by using smp_store_release to update
> data_head and aux_head, so let's do so.
>
> Signed-off-by: Mark Rutland <mark.rutland@arm.com>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Will Deacon <will.deacon@arm.com>
> ---
> kernel/events/ring_buffer.c | 13 ++++++-------
> 1 file changed, 6 insertions(+), 7 deletions(-)
>
> diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
> index 6c6b3c48db71..839b207e4c77 100644
> --- a/kernel/events/ring_buffer.c
> +++ b/kernel/events/ring_buffer.c
> @@ -63,10 +63,10 @@ static void perf_output_put_handle(struct perf_output_handle *handle)
> * kernel user
> *
> * if (LOAD ->data_tail) { LOAD ->data_head
> - * (A) smp_rmb() (C)
> + * (A) smp_rmb() (C)
> * STORE $data LOAD $data
> - * smp_wmb() (B) smp_mb() (D)
> - * STORE ->data_head STORE ->data_tail
> + * smp_mb() (D)
> + * RELEASE ->data_head (B) STORE ->data_tail
> * }

One thing to be aware of here is that the choice of ordering primitive (e.g.
using fences vs acquire/release operations) has the potential to create
ABI with userspace. I don't know of any architectures which currently care,
but if were were to merge a non multi-copy atomic architecture with native
acquire/release instructions, you could see issues if e.g. userspace used
smp_rmb(); READ_ONCE but the kernel used a RELEASE store.

Anyway, that's currently theoretical, but I think it's an argument for
putting these accessors in a uapi header.

Will

\
 
 \ /
  Last update: 2018-05-23 18:42    [W:0.103 / U:1.444 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site