Re: [PATCH net-next] stmmac: align RX buffers

From: Eric Dumazet
Date: Wed Aug 11 2021 - 08:54:14 EST




On 8/11/21 12:28 PM, Thierry Reding wrote:
> On Tue, Aug 10, 2021 at 08:07:47PM +0100, Marc Zyngier wrote:
>> Hi all,
>>
>> [adding Thierry, Jon and Will to the fun]
>>
>> On Mon, 14 Jun 2021 03:25:04 +0100,
>> Matteo Croce <mcroce@xxxxxxxxxxxxxxxxxxx> wrote:
>>>
>>> From: Matteo Croce <mcroce@xxxxxxxxxxxxx>
>>>
>>> On RX an SKB is allocated and the received buffer is copied into it.
>>> But on some architectures, the memcpy() needs the source and destination
>>> buffers to have the same alignment to be efficient.
>>>
>>> This is not our case, because SKB data pointer is misaligned by two bytes
>>> to compensate the ethernet header.
>>>
>>> Align the RX buffer the same way as the SKB one, so the copy is faster.
>>> An iperf3 RX test gives a decent improvement on a RISC-V machine:
>>>
>>> before:
>>> [ ID] Interval Transfer Bitrate Retr
>>> [ 5] 0.00-10.00 sec 733 MBytes 615 Mbits/sec 88 sender
>>> [ 5] 0.00-10.01 sec 730 MBytes 612 Mbits/sec receiver
>>>
>>> after:
>>> [ ID] Interval Transfer Bitrate Retr
>>> [ 5] 0.00-10.00 sec 1.10 GBytes 942 Mbits/sec 0 sender
>>> [ 5] 0.00-10.00 sec 1.09 GBytes 940 Mbits/sec receiver
>>>
>>> And the memcpy() overhead during the RX drops dramatically.
>>>
>>> before:
>>> Overhead Shared O Symbol
>>> 43.35% [kernel] [k] memcpy
>>> 33.77% [kernel] [k] __asm_copy_to_user
>>> 3.64% [kernel] [k] sifive_l2_flush64_range
>>>
>>> after:
>>> Overhead Shared O Symbol
>>> 45.40% [kernel] [k] __asm_copy_to_user
>>> 28.09% [kernel] [k] memcpy
>>> 4.27% [kernel] [k] sifive_l2_flush64_range
>>>
>>> Signed-off-by: Matteo Croce <mcroce@xxxxxxxxxxxxx>
>>
>> This patch completely breaks my Jetson TX2 system, composed of 2
>> Nvidia Denver and 4 Cortex-A57, in a very "funny" way.
>>
>> Any significant amount of traffic result in all sort of corruption
>> (ssh connections get dropped, Debian packages downloaded have the
>> wrong checksums) if any Denver core is involved in any significant way
>> (packet processing, interrupt handling). And it is all triggered by
>> this very change.
>>
>> The only way I have to make it work on a Denver core is to route the
>> interrupt to that particular core and taskset the workload to it. Any
>> other configuration involving a Denver CPU results in some sort of
>> corruption. On their own, the A57s are fine.
>>
>> This smells of memory ordering going really wrong, which this change
>> would expose. I haven't had a chance to dig into the driver yet (it
>> took me long enough to bisect it), but if someone points me at what is
>> supposed to synchronise the DMA when receiving an interrupt, I'll have
>> a look.
>
> I recall that Jon was looking into a similar issue recently, though I
> think the failure mode was slightly different. I also vaguely recall
> that CPU frequency was impacting this to some degree (lower CPU
> frequencies would increase the chances of this happening).
>
> Jon's currently out of office, but let me try and dig up the details
> on this.
>
> Thierry
>
>>
>> Thanks,
>>
>> M.
>>
>>> ---
>>> drivers/net/ethernet/stmicro/stmmac/stmmac.h | 4 ++--
>>> 1 file changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac.h b/drivers/net/ethernet/stmicro/stmmac/stmmac.h
>>> index b6cd43eda7ac..04bdb3950d63 100644
>>> --- a/drivers/net/ethernet/stmicro/stmmac/stmmac.h
>>> +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac.h
>>> @@ -338,9 +338,9 @@ static inline bool stmmac_xdp_is_enabled(struct stmmac_priv *priv)
>>> static inline unsigned int stmmac_rx_offset(struct stmmac_priv *priv)
>>> {
>>> if (stmmac_xdp_is_enabled(priv))
>>> - return XDP_PACKET_HEADROOM;
>>> + return XDP_PACKET_HEADROOM + NET_IP_ALIGN;
>>>
>>> - return 0;
>>> + return NET_SKB_PAD + NET_IP_ALIGN;
>>> }
>>>
>>> void stmmac_disable_rx_queue(struct stmmac_priv *priv, u32 queue);
>>> --
>>> 2.31.1
>>>
>>>
>>
>> --
>> Without deviation from the norm, progress is not possible.

Are you sure you do not need to adjust stmmac_set_bfsize(),
stmmac_rx_buf1_len() and stmmac_rx_buf2_len() ?

Presumably DEFAULT_BUFSIZE also want to be increased by NET_SKB_PAD

Patch for stmmac_rx_buf1_len() :

diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
index 7b8404a21544cf29668e8a14240c3971e6bce0c3..041a74e7efca3436bfe3e17f972dd156173957a9 100644
--- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
+++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
@@ -4508,12 +4508,12 @@ static unsigned int stmmac_rx_buf1_len(struct stmmac_priv *priv,

/* First descriptor, not last descriptor and not split header */
if (status & rx_not_ls)
- return priv->dma_buf_sz;
+ return priv->dma_buf_sz - NET_SKB_PAD - NET_IP_ALIGN;

plen = stmmac_get_rx_frame_len(priv, p, coe);

/* First descriptor and last descriptor and not split header */
- return min_t(unsigned int, priv->dma_buf_sz, plen);
+ return min_t(unsigned int, priv->dma_buf_sz - NET_SKB_PAD - NET_IP_ALIGN, plen);
}

static unsigned int stmmac_rx_buf2_len(struct stmmac_priv *priv,