[dpdk-dev] [PATCH 1/2] mem: add write memory barrier before changing heap state

[didier.pallard]

On 04/15/2014 04:08 PM, Richardson, Bruce wrote:
>> -----Original Message-----
>> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of David Marchand
>> Sent: Tuesday, April 15, 2014 2:51 PM
>> To: dev at dpdk.org
>> Subject: [dpdk-dev] [PATCH 1/2] mem: add write memory barrier before
>> changing heap state
>>
>> From: Didier Pallard <didier.pallard at 6wind.com>
>>
>> a write memory barrier is needed before changing heap state value, else some
>> concurrent core may see state changing before all initialization values are
>> written to memory, causing unpredictable results in malloc function.
>>
>> Signed-off-by: Didier Pallard <didier.pallard at 6wind.com>
> No barrier should be necessary here. As in a number of other places, such as rings, compiler barriers can be used in place of write memory barriers, due to IA ordering rules. However, in this case, both variables referenced are volatile variables and so the assignments to them cannot be reordered by the compiler so no compiler barrier is necessary either.
>
> Regards,
> /Bruce

Hi bruce,

Indeed a compiler barrier is absolutely needed here. volatile variable 
use is absolutely not a serializing instruction from compiler point of 
view; only atomic variable use is serializing, due to asm volatile 
(memory) directive use.
Here is the assembler generated with and without rte_wmb:


With rte_wmb

  142:   f0 45 0f b1 07          lock cmpxchg %r8d,(%r15)
  147:   0f 94 c0                sete   %al
  14a:   84 c0                   test   %al,%al
  14c:   74 ea                   je     138 <malloc_heap_alloc+0x68>
  14e:   49 c7 47 10 00 00 00    movq   $0x0,0x10(%r15)
  155:   00
  156:   41 c7 47 18 00 00 00    movl   $0x0,0x18(%r15)
  15d:   00
  15e:   41 c7 47 08 00 00 00    movl   $0x0,0x8(%r15)
  165:   00
  166:   41 c7 47 1c 00 00 00    movl   $0x0,0x1c(%r15)
  16d:   00
  16e:   49 c7 47 20 00 00 00    movq   $0x0,0x20(%r15)
  175:   00
  176:   45 89 57 04             mov    %r10d,0x4(%r15)
  17a:   0f ae f8                sfence
* 17d:   41 c7 07 02 00 00 00    movl   $0x2,(%r15)**
** 184:   41 8b 37                mov    (%r15),%esi**
* 187:   83 fe 02                cmp    $0x2,%esi
  18a:   75 b4                   jne    140 <malloc_heap_alloc+0x70>
  18c:   0f 1f 40 00             nopl   0x0(%rax)
  190:   48 83 c3 3f             add    $0x3f,%rbx


Without rte_wmb

  142:   f0 45 0f b1 07          lock cmpxchg %r8d,(%r15)
  147:   0f 94 c0                sete   %al
  14a:   84 c0                   test   %al,%al
  14c:   74 ea                   je     138 <malloc_heap_alloc+0x68>
  14e:   49 c7 47 10 00 00 00    movq   $0x0,0x10(%r15)
  155:   00
  156:   41 c7 47 08 00 00 00    movl   $0x0,0x8(%r15)
  15d:   00
* 15e:   41 c7 07 02 00 00 00    movl   $0x2,(%r15)**
** 165:   41 8b 37                mov    (%r15),%esi**
* 168:   41 c7 47 18 00 00 00    movl $0x0,0x18(%r15)
  16f:   00
  170:   41 c7 47 1c 00 00 00    movl   $0x0,0x1c(%r15)
  177:   00
  178:   49 c7 47 20 00 00 00    movq   $0x0,0x20(%r15)
  17f:   00
  180:   45 89 57 04             mov    %r10d,0x4(%r15)
  184:   83 fe 02                cmp    $0x2,%esi
  187:   75 b7                   jne    140 <malloc_heap_alloc+0x70>
  189:   0f 1f 80 00 00 00 00    nopl   0x0(%rax)
  190:   48 83 c3 3f             add    $0x3f,%rbx

It's clear that the *heap->initialised = INITIALISED;* instruction has 
been reordered by the compiler.

About rte_wmb and rte_rmb use, i agree with you that on intel 
architecture those macro should do nothing more than compiler barrier, 
due to Intel architecture choices.
But for code completness, i think those memory barriers should remain in 
place in the code, and rte_*mb should map to compiler barrier on intel 
architecture.

didier