Re: [dpdk-stable] [PATCH] mbuf: optimize memory loads during mbuf freeing

[Olivier Matz <olivier.matz@6wind.com>]

Hi,

On Mon, Mar 16, 2020 at 06:31:40PM +0000, Alexander Kozyrev wrote:
> Introduction of pinned external buffers doubled memory loads in the
> rte_pktmbuf_prefree_seg() function. Analysis of the generated assembly
> code shows unnecessary load of the pool field of the rte_mbuf structure.
> Here is the snippet of the assembly for "if (!RTE_MBUF_DIRECT(m))":
> Before the change the code was:
> 	movq  0x18(%rbx), %rax // load the ol_flags field
> 	test %r13, %rax	       // check if ol_flags equals to 0x60...0
> 	jz 0x9a8718 <Block 2>  // jump out to "if (m->next != NULL)"
> After the change the code becomed:
> 	movq  0x18(%rbx), %rax // load ol_flags
> 	test %r14, %rax	       // check if ol_flags equals to 0x60...0
> 	jnz 0x9bea38 <Block 2> // jump in to "if (!RTE_MBUF_HAS_EXTBUF(m)"
> 	movq  0x48(%rbx), %rax // load the pool field
> 	jmp 0x9bea78 <Block 7> // jump out to "if (m->next != NULL)"
> Look like this absolutely unneeded memory load of the pool field is an
> optimization for the external buffer case in GCC (4.8.5), since Clang
> generates the same assembly for both before and after the chenge versions.
> Plus, GCC favors the extrnal buffer case over the simple case.
> This assembly code layout causes the performance degradation because the
> rte_pktmbuf_prefree_seg() function is a part of a very hot path.
> Workaround this compilation issue by moving the check for pinned buffer
> apart from the check for external buffer and restore the initial code
> flow that favors the direct mbuf case over the external one.
> 
> Fixes: 6ef1107ad4c6 ("mbuf: detach mbuf with pinned external buffer")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Alexander Kozyrev <akozyrev@mellanox.com>
> Acked-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
> ---
>  lib/librte_mbuf/rte_mbuf.h | 14 ++++++--------
>  1 file changed, 6 insertions(+), 8 deletions(-)
> 
> diff --git a/lib/librte_mbuf/rte_mbuf.h b/lib/librte_mbuf/rte_mbuf.h
> index 34679e0..ab9d3f5 100644
> --- a/lib/librte_mbuf/rte_mbuf.h
> +++ b/lib/librte_mbuf/rte_mbuf.h
> @@ -1335,10 +1335,9 @@ static inline int __rte_pktmbuf_pinned_extbuf_decref(struct rte_mbuf *m)
>  	if (likely(rte_mbuf_refcnt_read(m) == 1)) {
>  
>  		if (!RTE_MBUF_DIRECT(m)) {
> -			if (!RTE_MBUF_HAS_EXTBUF(m) ||
> -			    !RTE_MBUF_HAS_PINNED_EXTBUF(m))
> -				rte_pktmbuf_detach(m);
> -			else if (__rte_pktmbuf_pinned_extbuf_decref(m))
> +			rte_pktmbuf_detach(m);
> +			if (RTE_MBUF_HAS_PINNED_EXTBUF(m) &&
> +			    __rte_pktmbuf_pinned_extbuf_decref(m))
>  				return NULL;
>  		}
>  
[...]

Reading the previous code again, it was correct but not easy
to understand, especially the:

  if (!RTE_MBUF_HAS_EXTBUF(m) || !RTE_MBUF_HAS_PINNED_EXTBUF(m))

Knowing we already checked it is not a direct mbuf, it is equivalent to:

  if (!RTE_MBUF_HAS_PINNED_EXTBUF(m))

I think the objective was to avoid an access to the pool flags if not
necessary.

Completely removing the test as you did is also functionally OK, because
rte_pktmbuf_detach() also does the check, and the code is even clearer.

I wonder however if doing this wouldn't avoid an access to the pool
flags for mbufs which have the IND_ATTACHED flags:

		if (!RTE_MBUF_DIRECT(m)) {
			rte_pktmbuf_detach(m);
			if (RTE_MBUF_HAS_EXTBUF(m) &&
			    RTE_MBUF_HAS_PINNED_EXTBUF(m) &&
			    __rte_pktmbuf_pinned_extbuf_decref(m))
				return NULL;
		}

What do you think?

Nit: if you wish to send a v2, there are few english fixes that could
be done (becomed, chenge, extrnal)

Thanks