RE: [PATCH] arm64: enable ARCH_WANT_RELAX_ORDER for aarch64

[Gabriele Paoloni]

Hi Robin and all

> -----Original Message-----
> From: linux-kernel-owner@xxxxxxxxxxxxxxx [mailto:linux-kernel-
> owner@xxxxxxxxxxxxxxx] On Behalf Of Robin Murphy
> Sent: 20 March 2017 14:00
> To: Dingtianhong; Catalin Marinas; Will Deacon; linux-arm-
> kernel@xxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> Cc: alexander.duyck@xxxxxxxxx; maowenan
> Subject: Re: [PATCH] arm64: enable ARCH_WANT_RELAX_ORDER for aarch64
> 
> On 14/03/17 14:06, Ding Tianhong wrote:
> > Hi Robinï
> >
> > On 2017/3/13 21:31, Robin Murphy wrote:
> >> On 13/03/17 12:03, Ding Tianhong wrote:
> >>> The ARCH_WANT_RELAX_ORDER will enable Relaxed Ordering (RO) which
> allows
> >>> transactions that do not have any order of completion requirements
> to
> >>> complete more efficiently compare to the Stricted Ordering (SO) for
> ixbge
> >>> nic card.
> >>
> >> Which ixgbe NIC? As far as I can see we have an arch-level config
> option
> >> here which applies to one single driver, and doesn't even cover all
> the
> >> hardware supported by that driver (82598, for example, still has the
> >> #ifndef CONFIG_SPARC in the equivalent place). Looking at the
> history,
> >> I'd prefer to at least know what the "various issues with certain
> >> chipsets" were, and why they wouldn't affect ARM systems,  before
> making
> >> any judgement about whether this could be considered universally
> safe
> >> for arm64.
> >>
> >
> > Indeed, in fact if the chipsets didn't support RO mode or has some
> errata for RO mode, it may
> > occur some issues, but it looks no such aarch64 chips, maybe I miss
> something.
> >
> > There are several intel nic card could support enable relax order, so
> need another patch to rename the SPARC
> > to ARCH_WANT_RELAX_ORDER, the universal name looks more better.
> 
> I'm sure I'm not alone in disagreeing outright that it looks better,
> because ARCH_ is hardly the appropriate namespace for a driver option
> unrelated to an architecture port's interaction with core kernel code;
> plus it's further confounded by a name which both doesn't imply any
> relationship with said driver, and does overlap with the kind of CPU
> memory model terminology which *is* the purview of architecture ports.
> 
> As an equivalent example, consider how equally misleading it would be
> from the ARM maintainer perspective if CONFIG_IOMMU_IO_PGTABLE_LPAE was
> just called CONFIG_ARCH_WANT_LPAE and implemented in this manner.
> 
> Having looked into it, I see that "Relaxed Order" does actually turn
> out
> to be a specific PCIe term, but even in that context it doesn't apply
> at
> the arch level - that's going to be a matter for particular endpoints
> and particular host controllers and all the quirks in between.

I fully agree on this and to be honest I don't understand how
<<commit 1a8b6d76dc5b "net:add one common config ARCH_WANT_RELAX_ORDER
to support relax ordering">> has landed into mainline...


> 
> >>> The system will see high write-to-memory performance when RO is
> >>> enabled on the data transactions just like the SPARC did.
> >>>
> >>> The aarch64 pcie controller could both support Relaxed Ordering
> (RO)
> >>
> >> What is "the AArch64 PCIe controller", exactly? Disregarding that
> >> talking of PCIe in terms of the CPU ISA makes little sense, I can
> barely
> >> name two ARMv8-based systems which nominally use the same PCIe IP,
> and
> >> the amount of various quirks and incompatibilities I'm aware of
> leaves
> >> me with the default assumption that any such unqualified blanket
> >> statement is probably wrong. I think we need some much more
> considered
> >> reasoning here.
> >>
> >
> > Agree, till now I could only test on hip06/hip07 board and get the
> better performance,
> > maybe I could test on other aarch64 platform.
> >
> >>> and Stricted Ordering (SO), so enable ARCH_WANT_RELAX_ORDER for
> ixgbe
> >>> nic card to get much more better performance, and didn't see any
> >>> adverse effects.
> >>>
> >>> Nic Card(Ixgbe)			Disable RO	|	Enable RO
> >>> Performance(Per thread)		8.4Gb/s		|	9.4Gb/s
> >>>
> >>> Tested by Iperf on Hip06/Hip07 Soc Board.
> >>>
> >>> Signed-off-by: Ding Tianhong <dingtianhong@xxxxxxxxxx>
> >>> ---
> >>>  arch/arm64/Kconfig | 1 +
> >>>  1 file changed, 1 insertion(+)
> >>>
> >>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> >>> index 8c7c244..36249a3 100644
> >>> --- a/arch/arm64/Kconfig
> >>> +++ b/arch/arm64/Kconfig
> >>> @@ -115,6 +115,7 @@ config ARM64
> >>>  	select SPARSE_IRQ
> >>>  	select SYSCTL_EXCEPTION_TRACE
> >>>  	select THREAD_INFO_IN_TASK
> >>> +	select ARCH_WANT_RELAX_ORDER
> >>
> >> I'd say the first order of business is to rename this config option
> to
> >> IXBGE_82599_WANT_RELAXED_ORDER so that it's not entirely misleading
> and
> >
> > not only for 82599, including 82598, 82576....
> 
> So why does ixgbe_start_hw_82598() still have the original #ifndef
> CONFIG_SPARC from 887012e80aea?
> 
> It was pretty clear from the outset that this is one of those patches
> for making a particular card go faster in a particular system based on
> what's available in the test lab - there's nothing inherently wrong
> with
> that, but if it were presented merely in those terms there would
> probably be a lot less to object to.
> 
> >> ambiguous. At first glance it looks far more like something scary to
> do
> >> with memory barriers than a network driver option. Howcome this
> isn't
> >> just in drivers/net/intel/Kconfig as a "default y if SPARC" bool
> anyway?
> >
> > didn't see any essential differences, and I still need to get some
> Acked by arm maintainer.
> 
> The big difference is that had people done the sensible thing by
> adding,
> say, CONFIG_IXGBE_ALLOW_RELAXED_ORDER to drivers/net/intel/... and
> sending a self-contained patch through the net tree, architecture
> maintainers wouldn't even need to be aware, let alone ack anything.
> Then
> in future if someone sends another patch against the net tree changing
> "y if (SPARC || ARM64)" back to "y if SPARC" because it happens to
> break
> on their system, the resulting discussion and resolution can happen on
> netdev, and architecture maintainers who aren't necessarily familiar
> with particular ixgbe/PCIe hardware details *still* don't need to care.

Standard PCIe drivers uses bit 4 of the Device Control Register to
enable/disable relaxed ordering: here it is not clear what Intel means
by relaxed ordering and in which context (at least not to me) and why
it should be disabled by default.

From my perspective I would try to propose the following patch as RFC
and see what the Intel maintainer comes up with and if any other ARM64
host vendor would oppose to it.
The RFC below reverts commit 1a8b6d76dc5b and enable relaxed ordering
on SPARC and ARM64 machines...

What do you think?


diff --git a/arch/Kconfig b/arch/Kconfig
index cd211a1..e03d354 100644
--- a/arch/Kconfig
+++ b/arch/Kconfig
@@ -844,7 +844,4 @@ config STRICT_MODULE_RWX
 	  and non-text memory will be made non-executable. This provides
 	  protection against certain security exploits (e.g. writing to text)
 
-config ARCH_WANT_RELAX_ORDER
-	bool
-
 source "kernel/gcov/Kconfig"
diff --git a/arch/sparc/Kconfig b/arch/sparc/Kconfig
index 68ac5c7..cf4034c 100644
--- a/arch/sparc/Kconfig
+++ b/arch/sparc/Kconfig
@@ -44,7 +44,6 @@ config SPARC
 	select CPU_NO_EFFICIENT_FFS
 	select HAVE_ARCH_HARDENED_USERCOPY
 	select PROVE_LOCKING_SMALL if PROVE_LOCKING
-	select ARCH_WANT_RELAX_ORDER
 
 config SPARC32
 	def_bool !64BIT
diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_common.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_common.c
index c38d50c..d55dcac 100644
--- a/drivers/net/ethernet/intel/ixgbe/ixgbe_common.c
+++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_common.c
@@ -350,7 +350,8 @@ s32 ixgbe_start_hw_gen2(struct ixgbe_hw *hw)
 	}
 	IXGBE_WRITE_FLUSH(hw);
 
-#ifndef CONFIG_ARCH_WANT_RELAX_ORDER
+#if !defined(CONFIG_SPARC) && !defined(CONFIG_ARM64)
+
 	/* Disable relaxed ordering */
 	for (i = 0; i < hw->mac.max_tx_queues; i++) {
 		u32 regval;


> 
> Robin.