Re: [PATCH v2 4/6] RISC-V: Detect unaligned vector accesses supported.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 





On 6/21/24 13:18, Charlie Jenkins wrote:
On Fri, Jun 21, 2024 at 11:06:31AM +0100, Conor Dooley wrote:
On Thu, Jun 20, 2024 at 03:14:14PM -0700, Charlie Jenkins wrote:
On Thu, Jun 20, 2024 at 05:31:28PM -0400, Jesse Taube wrote:


On 6/17/24 22:09, Charlie Jenkins wrote:
On Mon, Jun 17, 2024 at 06:43:32PM -0700, Charlie Jenkins wrote:
On Thu, Jun 13, 2024 at 03:16:13PM -0400, Jesse Taube wrote:
Run a unaligned vector access to test if the system supports
vector unaligned access. Add the result to a new key in hwprobe.
This is useful for usermode to know if vector misaligned accesses are
supported and if they are faster or slower than equivalent byte accesses.

Signed-off-by: Jesse Taube <jesse@xxxxxxxxxxxx>
---
V1 -> V2:
   - Add Kconfig options
   - Add insn_is_vector
   - Add handle_vector_misaligned_load
   - Fix build
   - Seperate vector from scalar misaligned access
   - This patch was almost completely rewritten
---
   arch/riscv/Kconfig                         |  41 +++++++
   arch/riscv/include/asm/cpufeature.h        |   7 +-
   arch/riscv/include/asm/entry-common.h      |  11 --
   arch/riscv/include/asm/hwprobe.h           |   2 +-
   arch/riscv/include/asm/vector.h            |   1 +
   arch/riscv/include/uapi/asm/hwprobe.h      |   5 +
   arch/riscv/kernel/Makefile                 |   4 +-
   arch/riscv/kernel/sys_hwprobe.c            |  41 +++++++
   arch/riscv/kernel/traps_misaligned.c       | 119 ++++++++++++++++++++-
   arch/riscv/kernel/unaligned_access_speed.c |   9 +-
   arch/riscv/kernel/vector.c                 |   2 +-
   11 files changed, 221 insertions(+), 21 deletions(-)

diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
index b94176e25be1..f12df0ca6c18 100644
--- a/arch/riscv/Kconfig
+++ b/arch/riscv/Kconfig
@@ -723,6 +723,12 @@ config RISCV_MISALIGNED
   	help
   	  Embed support for emulating misaligned loads and stores.
+config RISCV_VECTOR_MISALIGNED
+	bool
+	depends on RISCV_ISA_V
+	help
+	  Enable detecting support for vector misaligned loads and stores.
+
   choice
   	prompt "Unaligned Accesses Support"
   	default RISCV_PROBE_UNALIGNED_ACCESS
@@ -774,6 +780,41 @@ config RISCV_EFFICIENT_UNALIGNED_ACCESS
   endchoice
+choice
+	prompt "Vector unaligned Accesses Support"
+	depends on RISCV_ISA_V
+	default RISCV_PROBE_VECTOR_UNALIGNED_ACCESS
+	help
+	  This determines the level of support for vector unaligned accesses. This
+	  information is used by the kernel to perform optimizations.

I haven't actually checked the patchset, but is it actually used by the
kernel to perform optimisations?

No ;)

Right now this patch is just providing the information to userspace
through hwprobe and doesn't optimize anything in the kernel.


It is also
+	  exposed to user space via the hwprobe syscall. The hardware will be
+	  probed at boot by default.
+
+config RISCV_DETECT_VECTOR_UNALIGNED_ACCESS

This is not used anywhere, what is the reason for including it?

This is so that we can check if they are supported or not, but not check the
speed of them. Similar to RISCV_EMULATED_UNALIGNED_ACCESS.

What do you mean? It isn't used anywhere so this "check if they are
supported or not" is not guarded by this config.



+	bool "Detect support for vector unaligned accesses"
+	select RISCV_VECTOR_MISALIGNED
+	help
+	  During boot, the kernel will detect if the system supports vector
+	  unaligned accesses.
+
+config RISCV_PROBE_VECTOR_UNALIGNED_ACCESS
+	bool "Probe speed of vector unaligned accesses"
+	select RISCV_VECTOR_MISALIGNED
+	help
+	  During boot, the kernel will run a series of tests to determine the
+	  speed of vector unaligned accesses if they are supported. This probing
+	  will dynamically determine the speed of vector unaligned accesses on
+	  the underlying system if they are supported.
+
+config CONFIG_RISCV_UNALIGNED_ACCESS_UNSUPPORTED

This should not be prefixed with CONFIG and does not include VECTOR in
the name.

Huh thought it would warn fixed though

What do you mean by "warn fixed"?


I assume you meant to put
"RISCV_VEC_UNALIGNED_ACCESS_UNSUPPORTED" here?

This is to leave a faster path like SLOW or FAST to say that unaligned
access arent suported.

I am not sure what you are responding to. This comment seems to be
responding to my correction of
CONFIG_RISCV_UNALIGNED_ACCESS_UNSUPPORTED->RISCV_VEC_UNALIGNED_ACCESS_UNSUPPORTED
so I don't see how that ties into SLOW/FAST.



This was also intentionally left out on the scalar side [1]. The
implication here is that having this config will cause people to compile
kernels without unaligned access support which really shouldn't be
something we are explicitly supporting.

If somebody does want to support hardware that does not handle vector
unaligned accesses, the solution should be to add emulation support to
the kernel.

Yes but we dont have emulation support yet so I do think its a good idea.

I am hesitant because it is very likely that somebody will add support
for unaligned vector emulation. When there is emulation support, this
config option should not exist to be consistent with scalar. However if
we add this option in now, we must expect a user to enable this config,
and then

I dunno, I think there could be value in having the option here. For
scalar, we couldn't have an option that would break the uABI, so the
unsupported option wasn't okay. That's not a constraint that we have for
vector.

For vector, if you have a system that doesn't support misaligned access,
you probably don't want to emulate the accesses either, since that's
likely remove any performance gains you get from using vector in the
first place, so I can see benefit in the option.

We have the RISCV_MISALIGNED option that disables the scalar emulation,
but doesn't break the UABI because it's not saying that misaligned
scalar is not supported, but just that the kernel doesn't emulate.
Having an UNSUPPORTED option explicitly breaks the UABI which doesn't
seem like something that the kernel should support. If we are okay with
having options that break the UABI then this is fine, but I was under
the impression that we did our best to avoid that.

There definitely is value in having an option like this for hardware
that never wants to run misaligned code, but since we decided with the
scalar accesses that we should not break the UABI I think vector should
do the same.

The rational for scalar accesses was slightly different as
The base ISA spec said/says: ". The base ISA supports misaligned accesses, but these might run extremely slowly depending on the implementation."

Section: 2.6 Load and Store Instructions. Available: https://riscv.org/wp-content/uploads/2017/05/riscv-spec-v2.2.pdf


Enabling the probing is going to end up with same outcome for userspace
as having this option on such a system, so it comes down to whether you
want to allow people to skip the probing if they know their system has
this problem.

we will have to get rid of it later. Users are not always happy
when config options are removed.

I dunno, I don't think that adding emulation requires that we remove
this unsupported option.

I am probably being too paranoid about this but I am hesistant to cause
vector and scalar misaligned access implemenations to diverge. It is
inconsistent to allow an unsupported option for vector but not for
scalar when both support emulation. The probing is very fast as it just
checks if a misaligned access causes a trap and then sets the access
speed to unsupported if it does trap.

Charlie is right about probing for support being fast. There is
RISCV_DETECT_VECTOR_UNALIGNED_ACCESS to only detect support not the speed. I thought it might be a good idea to add this config option, but
I'll just drop it to shorten this thread.



Additionally, what are we doing in the kernel if we detect that
misaligned stuff isn't supported? Are we going to mandate that kernel
code is aligned only

As of now yes.

disable in-kernel vector or some other mechanism
to make sure that things like crypto code don't have/introduce code
that'll not run on these systems?

I'm not too familiar with the uses of unaligned vector accesses, but if
it significantly improves performance then I think there should be a faster unaligned access pathway, and a aligned access pathway, so both are supported. This also ties into your first question.

Thanks,
Jesse Taube

UNSUPPORTED will still be set by the quick probe so it would be possible
for the kernel/userspace to avoid running misaligned vector when it's
unsupported. Any kernel methods would probably want to always run
aligned vector unless misaligned support was determined to be FAST
anyway, I am doubtful that code will have different optimizations for
FAST, SLOW, and UNSUPPORTED but it is possible.

I would prefer consistency between scalar and vector misaligned support,
but this is not a deal breaker for this patch. I am not convinced it is
the best choice, but I am okay with leaving this option in the kernel.

- Charlie


Cheers,
Conor.






[Index of Archives]     [Device Tree Compilter]     [Device Tree Spec]     [Linux Driver Backports]     [Video for Linux]     [Linux USB Devel]     [Linux PCI Devel]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Yosemite Backpacking]


  Powered by Linux