Re: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

[Gavin Hu <Gavin.Hu@arm.com>]

Hi Xiaoyun,

> -----Original Message-----
> From: Li, Xiaoyun <xiaoyun.li@intel.com>
> Sent: Monday, December 23, 2019 5:36 PM
> To: Gavin Hu <Gavin.Hu@arm.com>; Wu, Jingjing <jingjing.wu@intel.com>
> Cc: dev@dpdk.org; Maslekar, Omkar <omkar.maslekar@intel.com>;
> stable@dpdk.org; nd <nd@arm.com>; jerinj@marvell.com; Honnappa
> Nagarahalli <Honnappa.Nagarahalli@arm.com>; Richardson, Bruce
> <bruce.richardson@intel.com>; nd <nd@arm.com>
> Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue
> 
> Hi
> 
> Still, stability and correctness are much more important than performance.
> As I said, with WC can benefit more than 20X perf. Comparing to this benefit,
> the difference between rte_wmb and rte_io_wmb is not that important.
> And in my test, the performance is not that bad with rte_wmb especially with
> large packets which are the normal use cases.
I agree 'sfence' is the correct barrier for WC, as it is a weak memory model.
Rte_io on x86 is a compiler barrier, that is not strong enough.  
> 
> BTW, I've searched linux kernel codes and don't see any NTB device on arm
> platform.
> So I don't think you need to consider the perf hurt to arm.
Limited the discussion to NTB, I am fine, and since WC in not used for any other NICs, that's ok.
> 
> Best Regards
> Xiaoyun Li
> 
> > -----Original Message-----
> > From: Gavin Hu [mailto:Gavin.Hu@arm.com]
> > Sent: Monday, December 23, 2019 16:38
> > To: Li, Xiaoyun <xiaoyun.li@intel.com>; Wu, Jingjing
> <jingjing.wu@intel.com>
> > Cc: dev@dpdk.org; Maslekar, Omkar <omkar.maslekar@intel.com>;
> > stable@dpdk.org; nd <nd@arm.com>; jerinj@marvell.com; Honnappa
> > Nagarahalli <Honnappa.Nagarahalli@arm.com>; Richardson, Bruce
> > <bruce.richardson@intel.com>; nd <nd@arm.com>
> > Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue
> >
> > Hi Xiaoyun,
> >
> > > -----Original Message-----
> > > From: Li, Xiaoyun <xiaoyun.li@intel.com>
> > > Sent: Monday, December 23, 2019 3:52 PM
> > > To: Gavin Hu <Gavin.Hu@arm.com>; Wu, Jingjing <jingjing.wu@intel.com>
> > > Cc: dev@dpdk.org; Maslekar, Omkar <omkar.maslekar@intel.com>;
> > > stable@dpdk.org; nd <nd@arm.com>
> > > Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier
> > > issue
> > >
> > > Hi
> > > I reconsidered and retested about this issue.
> > > I still need to use rte_wmb instead of using rte_io_wmb.
> > >
> > > Because to achieve high performance, ntb needs to turn on WC(write
> > > combining) feature. The perf difference with and without WC enabled is
> > > more than 20X.
> > > And when WC enabled, rte_io_wmb cannot make sure the instructions
> are
> > > in order only rte_wmb can make sure that.
> > >
> > > And in my retest, when sending 64 bytes packets, using rte_io_wmb will
> > > cause out-of-order issue and cause memory corruption on rx side.
> > > And using rte_wmb is fine.
> > That's true, as it is declared as 'write combine' region, even x86 is known as
> > strong ordered, it is the interconnect or PCI RC may do the reordering, 'write
> > combine', 'write coalescing', which caused this problem.
> > IMO, rte_io_*mb barriers on x86 should be promoted to stronger is WC is
> > involved(but that will sap performance for non-WC memories?).
> >
> https://code.dpdk.org/dpdk/latest/source/lib/librte_eal/common/include/ar
> ch/
> > x86/rte_atomic.h#L78
> >
> > Using rte_wmb will hurt performance for aarch64 also, as pci device
> memory
> > accesses to a single device are strongly ordered therefore the strongest
> > rte_wmb is not necessary.
> > > So I can only use v1 patch and suspend v2 patch in patchwork.
> > >
> > > Best Regards
> > > Xiaoyun Li
> > >
> > > > -----Original Message-----
> > > > From: Gavin Hu (Arm Technology China) [mailto:Gavin.Hu@arm.com]
> > > > Sent: Monday, December 16, 2019 18:50
> > > > To: Li, Xiaoyun <xiaoyun.li@intel.com>; Wu, Jingjing
> > > <jingjing.wu@intel.com>
> > > > Cc: dev@dpdk.org; Maslekar, Omkar <omkar.maslekar@intel.com>;
> > > > stable@dpdk.org; nd <nd@arm.com>
> > > > Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier
> > > issue
> > > >
> > > >
> > > >
> > > > > -----Original Message-----
> > > > > From: dev <dev-bounces@dpdk.org> On Behalf Of Xiaoyun Li
> > > > > Sent: Monday, December 16, 2019 9:59 AM
> > > > > To: jingjing.wu@intel.com
> > > > > Cc: dev@dpdk.org; omkar.maslekar@intel.com; Xiaoyun Li
> > > > > <xiaoyun.li@intel.com>; stable@dpdk.org
> > > > > Subject: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier
> > > > > issue
> > > > >
> > > > > All buffers and ring info should be written before tail register update.
> > > > > This patch relocates the write memory barrier before updating tail
> > > > > register to avoid potential issues.
> > > > >
> > > > > Fixes: 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")
> > > > > Cc: stable@dpdk.org
> > > > >
> > > > > Signed-off-by: Xiaoyun Li <xiaoyun.li@intel.com>
> > > > > ---
> > > > > v2:
> > > > >  * Replaced rte_wmb with rte_io_wmb since rte_io_wmb is enough.
> > > > > ---
> > > > >  drivers/raw/ntb/ntb.c | 4 ++--
> > > > >  1 file changed, 2 insertions(+), 2 deletions(-)
> > > > >
> > > > > diff --git a/drivers/raw/ntb/ntb.c b/drivers/raw/ntb/ntb.c index
> > > > > ad7f6abfd..c7de86f36 100644
> > > > > --- a/drivers/raw/ntb/ntb.c
> > > > > +++ b/drivers/raw/ntb/ntb.c
> > > > > @@ -683,8 +683,8 @@ ntb_enqueue_bufs(struct rte_rawdev *dev,
> > > > >  			   sizeof(struct ntb_used) * nb1);
> > > > >  		rte_memcpy(txq->tx_used_ring, tx_used + nb1,
> > > > >  			   sizeof(struct ntb_used) * nb2);
> > > > > +		rte_io_wmb();
> > > > As both txq->tx_used_ring and *txq->used_cnt are physically reside
> > > > in the
> > > PCI
> > > > device side, rte_io_wmb is correct to ensure the ordering.
> > > >
> > > > >  		*txq->used_cnt = txq->last_used;
> > > > > -		rte_wmb();
> > > > >
> > > > >  		/* update queue stats */
> > > > >  		hw->ntb_xstats[NTB_TX_BYTES_ID + off] += bytes; @@ -
> > > 789,8
> > > > +789,8 @@
> > > > > ntb_dequeue_bufs(struct rte_rawdev *dev,
> > > > >  			   sizeof(struct ntb_desc) * nb1);
> > > > >  		rte_memcpy(rxq->rx_desc_ring, rx_desc + nb1,
> > > > >  			   sizeof(struct ntb_desc) * nb2);
> > > > > +		rte_io_wmb();
> > > > >  		*rxq->avail_cnt = rxq->last_avail;
> > > > > -		rte_wmb();
> > > > >
> > > > >  		/* update queue stats */
> > > > >  		off = NTB_XSTATS_NUM * ((size_t)context + 1);
> > > > > --
> > > > > 2.17.1
> > > >
> > > > Reviewed-by: Gavin Hu <gavin.hu@arm.com>