On 11/14/24 12:46, Charlie Jenkins wrote:
On Thu, Nov 14, 2024 at 11:26:47AM +0800, Yangyu Chen wrote:
On 11/14/24 11:02, Charlie Jenkins wrote:
On Thu, Nov 14, 2024 at 10:44:37AM +0800, Yangyu Chen wrote:
On 11/14/24 10:21, Charlie Jenkins wrote:
Add a new hwprobe key "RISCV_HWPROBE_KEY_VENDOR_EXT_THEAD_0" which
allows userspace to probe for the new RISCV_ISA_VENDOR_EXT_XTHEADVECTOR
vendor extension.
Hi Charlie,
How about changing the name of the key from
"RISCV_ISA_VENDOR_EXT_XTHEADVECTOR" to "RISCV_HWPROBE_KEY_VENDOR_EXT_0" and
use marchid to identify what the vendor is, each vendor will have its own
bit definition in this value. So we can avoid adding so many hwprobe keys
for each vendor in the future.
I proposed a commit here: https://github.com/cyyself/linux/commit/36390645d85d1ac75dd71172f167719df4297f59
I actually originally had this in one of my first versions of this
series but was convinced by Conor to change it. The problem with it was
that tying vendor extensions to mvendorid means that it is enforced by
the kernel that vendors cannot share vendor extensions. It is possible
for vendor A to purchase IP that contains a vendor extension from vendor
B. This vendor extension should work on platforms created by vendor A
and vendor B. However, vendor A and vendor B have different mvendorids,
so the kernel can't support this if it is tied to mvendorid. It could
be solved by duplicating every extension that vendors have, but then
userspace software would have to keep in mind the mvendorid they are
running on and check the different extensions for the different vendors
even though the implementation of the extension is the same.
The original conversation where Conor and I agreed that it was better to
have vendor extensions not rely on mvendorid:
https://lore.kernel.org/linux-riscv/20240416-husband-flavored-96c1dad58b6e@wendy/
Thanks for your explanation. I will strongly agree with Conor's opinion if
the feature bitmask does not exist in RISC-V C-ABI.
However, as the feature mask defined in RISC-V C-ABI[1] uses the design
depending on marchid currently, should we reconsider this key for its use
case? The current target_clones and taget_version implemented in GCC[2] and
LLVM[3] also use the bitmask defined in C-ABI. I think if we use this key
depending on marchid, to make a key shared with all vendors will make this
cleaner.
Changing this will break linux userspace API. It is a non-workable
solution for the kernel to associate extensions with marchid/mvendorid
for the reasons provided. I fail to see why this ABI would require the
kernel to behave in this manner. The ABI provides the marchid to be used
by function multi-versioning and applications are free to use the
marchid to change which function they want to compile. However, if they
want to know if an extension is supported, then they need to use
hwprobe. If they want to check if xtheadvector is supported, then they> call hwprobe with the xtheadvector key. This is true no matter what the
mvendorid of the system is.
A userspace software can use either c-api defined feature masks or
directly use hwprobe syscall. If they use c-api defined feature masks as
GCC or LLVM did for compiler generated IFUNC resolver, the bitmask is
guarded by mvendorid. So my point at that time was that if the C-API
defined way became mainstream, why should we keep this key only for
T-Head to increase the maintenance overhead?
This has been discussed here before in RISC-V C-API:
https://github.com/riscv-non-isa/riscv-c-api-doc/pull/74#issuecomment-2128844747
But now (from the last email), you convinced me. So, I would like to
make the c-api change:
https://github.com/riscv-non-isa/riscv-c-api-doc/issues/96
This does not add any complexity, "clean"
code can equally be written following this scheme or following a scheme
that relies on mvendorid. Ditching the reliance on mvendorid in the
kernel allows the kernel to be as generic as possible, and allow
whatever ABIs or hardware that exist to have a resiliant way of
communicating with the kernel.
OK. I'm just concerned about when these vendors will add the hwprobe key
for their own extension, which may introduce a potential merge conflict
in the kernel tree. It can also be a disaster if the hardware vendor
ships their kernel with these under-review patches for their products
with hwprobe key conflict with mainline kernel.
But we can avoid this now by adding each key for each vendor to avoid
potential conflict in the future. This can be a separate patch for
future work, so there is nothing to change here.
Thanks,
Yangyu Chen
- CHarlie
[1] https://github.com/riscv-non-isa/riscv-c-api-doc/blob/main/src/c-api.adoc#function-multi-version
[2] https://github.com/gcc-mirror/gcc/blob/8564d0948c72df0a66d7eb47e15c6ab43e9b25ce/gcc/config/riscv/riscv.cc#L13016
[3] https://github.com/llvm/llvm-project/blob/f407dff50cdcbcfee9dd92397d3792627c3ac708/clang/lib/CodeGen/CGBuiltin.cpp#L14627
This new key will allow userspace code to probe for which thead vendor
extensions are supported. This API is modeled to be consistent with
RISCV_HWPROBE_KEY_IMA_EXT_0. The bitmask returned will have each bit
corresponding to a supported thead vendor extension of the cpumask set.
Just like RISCV_HWPROBE_KEY_IMA_EXT_0, this allows a userspace program
to determine all of the supported thead vendor extensions in one call.
Signed-off-by: Charlie Jenkins <charlie@xxxxxxxxxxxx>
Reviewed-by: Evan Green <evan@xxxxxxxxxxxx>
---
arch/riscv/include/asm/hwprobe.h | 3 +-
.../include/asm/vendor_extensions/thead_hwprobe.h | 19 +++++++++++
.../include/asm/vendor_extensions/vendor_hwprobe.h | 37 ++++++++++++++++++++++
arch/riscv/include/uapi/asm/hwprobe.h | 3 +-
arch/riscv/include/uapi/asm/vendor/thead.h | 3 ++
arch/riscv/kernel/sys_hwprobe.c | 5 +++
arch/riscv/kernel/vendor_extensions/Makefile | 1 +
.../riscv/kernel/vendor_extensions/thead_hwprobe.c | 19 +++++++++++
8 files changed, 88 insertions(+), 2 deletions(-)