Re: [PATCH 16/19] riscv: hwprobe: Add vendor extension probing

From: Evan Green
Date: Fri Apr 12 2024 - 17:43:52 EST

On Fri, Apr 12, 2024 at 1:20 PM Charlie Jenkins <charlie@xxxxxxxxxxxx> wrote:
>
> On Fri, Apr 12, 2024 at 12:07:46PM -0700, Evan Green wrote:
> > On Fri, Apr 12, 2024 at 11:17 AM Charlie Jenkins <charlie@xxxxxxxxxxxx> wrote:
> > >
> > > On Fri, Apr 12, 2024 at 10:05:21AM -0700, Evan Green wrote:
> > > > On Thu, Apr 11, 2024 at 9:12 PM Charlie Jenkins <charlie@xxxxxxxxxxxx> wrote:
> > > > >
> > > > > Add a new hwprobe key "RISCV_HWPROBE_KEY_VENDOR_EXT_0" which allows
> > > > > userspace to probe for the new RISCV_ISA_VENDOR_EXT_XTHEADVECTOR vendor
> > > > > extension.
> > > > >
> > > > > Signed-off-by: Charlie Jenkins <charlie@xxxxxxxxxxxx>
> > > > > ---
> > > > > arch/riscv/include/asm/hwprobe.h | 4 +--
> > > > > arch/riscv/include/uapi/asm/hwprobe.h | 10 +++++-
> > > > > arch/riscv/kernel/sys_hwprobe.c | 59 +++++++++++++++++++++++++++++++++--
> > > > > 3 files changed, 68 insertions(+), 5 deletions(-)
> > > > >
> > > > > diff --git a/arch/riscv/include/asm/hwprobe.h b/arch/riscv/include/asm/hwprobe.h
> > > > > index 630507dff5ea..e68496b4f8de 100644
> > > > > --- a/arch/riscv/include/asm/hwprobe.h
> > > > > +++ b/arch/riscv/include/asm/hwprobe.h
> > > > > @@ -1,6 +1,6 @@
> > > > > /* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
> > > > > /*
> > > > > - * Copyright 2023 Rivos, Inc
> > > > > + * Copyright 2023-2024 Rivos, Inc
> > > > > */
> > > > >
> > > > > #ifndef _ASM_HWPROBE_H
> > > > > @@ -8,7 +8,7 @@
> > > > >
> > > > > #include <uapi/asm/hwprobe.h>
> > > > >
> > > > > -#define RISCV_HWPROBE_MAX_KEY 6
> > > > > +#define RISCV_HWPROBE_MAX_KEY 7
> > > > >
> > > > > static inline bool riscv_hwprobe_key_is_valid(__s64 key)
> > > > > {
> > > > > diff --git a/arch/riscv/include/uapi/asm/hwprobe.h b/arch/riscv/include/uapi/asm/hwprobe.h
> > > > > index 9f2a8e3ff204..6614d3adfc75 100644
> > > > > --- a/arch/riscv/include/uapi/asm/hwprobe.h
> > > > > +++ b/arch/riscv/include/uapi/asm/hwprobe.h
> > > > > @@ -1,6 +1,6 @@
> > > > > /* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
> > > > > /*
> > > > > - * Copyright 2023 Rivos, Inc
> > > > > + * Copyright 2023-2024 Rivos, Inc
> > > > > */
> > > > >
> > > > > #ifndef _UAPI_ASM_HWPROBE_H
> > > > > @@ -67,6 +67,14 @@ struct riscv_hwprobe {
> > > > > #define RISCV_HWPROBE_MISALIGNED_UNSUPPORTED (4 << 0)
> > > > > #define RISCV_HWPROBE_MISALIGNED_MASK (7 << 0)
> > > > > #define RISCV_HWPROBE_KEY_ZICBOZ_BLOCK_SIZE 6
> > > > > +/*
> > > > > + * It is not possible for one CPU to have multiple vendor ids, so each vendor
> > > > > + * has its own vendor extension "namespace". The keys for each vendor starts
> > > > > + * at zero.
> > > > > + */
> > > > > +#define RISCV_HWPROBE_KEY_VENDOR_EXT_0 7
> > > > > + /* T-Head */
> > > > > +#define RISCV_HWPROBE_VENDOR_EXT_XTHEADVECTOR (1 << 0)
> > > > > /* Increase RISCV_HWPROBE_MAX_KEY when adding items. */
> > > > >
> > > > > /* Flags */
> > > > > diff --git a/arch/riscv/kernel/sys_hwprobe.c b/arch/riscv/kernel/sys_hwprobe.c
> > > > > index e0a42c851511..365ce7380443 100644
> > > > > --- a/arch/riscv/kernel/sys_hwprobe.c
> > > > > +++ b/arch/riscv/kernel/sys_hwprobe.c
> > > > > @@ -69,7 +69,8 @@ static void hwprobe_isa_ext0(struct riscv_hwprobe *pair,
> > > > > if (riscv_isa_extension_available(NULL, c))
> > > > > pair->value |= RISCV_HWPROBE_IMA_C;
> > > > >
> > > > > - if (has_vector() && !riscv_has_vendor_extension_unlikely(RISCV_ISA_VENDOR_EXT_XTHEADVECTOR))
> > > > > + if (has_vector() &&
> > > > > + !__riscv_isa_vendor_extension_available(NULL, RISCV_ISA_VENDOR_EXT_XTHEADVECTOR))
> > > > > pair->value |= RISCV_HWPROBE_IMA_V;
> > > > >
> > > > > /*
> > > > > @@ -112,7 +113,8 @@ static void hwprobe_isa_ext0(struct riscv_hwprobe *pair,
> > > > > EXT_KEY(ZACAS);
> > > > > EXT_KEY(ZICOND);
> > > > >
> > > > > - if (has_vector() && !riscv_has_vendor_extension_unlikely(RISCV_ISA_VENDOR_EXT_XTHEADVECTOR)) {
> > > > > + if (has_vector() &&
> > > > > + !riscv_has_vendor_extension_unlikely(RISCV_ISA_VENDOR_EXT_XTHEADVECTOR)) {
> > > > > EXT_KEY(ZVBB);
> > > > > EXT_KEY(ZVBC);
> > > > > EXT_KEY(ZVKB);
> > > > > @@ -139,6 +141,55 @@ static void hwprobe_isa_ext0(struct riscv_hwprobe *pair,
> > > > > pair->value &= ~missing;
> > > > > }
> > > > >
> > > > > +static void hwprobe_isa_vendor_ext0(struct riscv_hwprobe *pair,
> > > > > + const struct cpumask *cpus)
> > > > > +{
> > > > > + int cpu;
> > > > > + u64 missing = 0;
> > > > > +
> > > > > + pair->value = 0;
> > > > > +
> > > > > + struct riscv_hwprobe mvendorid = {
> > > > > + .key = RISCV_HWPROBE_KEY_MVENDORID,
> > > > > + .value = 0
> > > > > + };
> > > > > +
> > > > > + hwprobe_arch_id(&mvendorid, cpus);
> > > > > +
> > > > > + /* Set value to zero if CPUs in the set do not have the same vendor. */
> > > > > + if (mvendorid.value == -1ULL)
> > > > > + return;
> > > > > +
> > > > > + /*
> > > > > + * Loop through and record vendor extensions that 1) anyone has, and
> > > > > + * 2) anyone doesn't have.
> > > > > + */
> > > > > + for_each_cpu(cpu, cpus) {
> > > > > + struct riscv_isainfo *isavendorinfo = &hart_isa_vendor[cpu];
> > > > > +
> > > > > +#define VENDOR_EXT_KEY(ext) \
> > > > > + do { \
> > > > > + if (__riscv_isa_vendor_extension_available(isavendorinfo->isa, \
> > > > > + RISCV_ISA_VENDOR_EXT_##ext)) \
> > > > > + pair->value |= RISCV_HWPROBE_VENDOR_EXT_##ext; \
> > > > > + else \
> > > > > + missing |= RISCV_HWPROBE_VENDOR_EXT_##ext; \
> > > > > + } while (false)
> > > > > +
> > > > > + /*
> > > > > + * Only use VENDOR_EXT_KEY() for extensions which can be exposed to userspace,
> > > > > + * regardless of the kernel's configuration, as no other checks, besides
> > > > > + * presence in the hart_vendor_isa bitmap, are made.
> > > > > + */
> > > > > + VENDOR_EXT_KEY(XTHEADVECTOR);
> > > > > +
> > > > > +#undef VENDOR_EXT_KEY
> > > >
> > > > Hey Charlie,
> > > > Thanks for writing this up! At the very least I think the
> > > > THEAD-specific stuff should probably end up in its own file, otherwise
> > > > it'll get chaotic with vendors clamoring to add stuff right here.
> > >
> > > Great idea!
> > >
> > > > What do you think about this approach:
> > > > * We leave RISCV_HWPROBE_MAX_KEY as the max key for the "generic
> > > > world", eg 6-ish
> > > > * We define that any key above 0x8000000000000000 is in the vendor
> > > > space, so the meaning of the keys depends first on the mvendorid
> > > > value.
> > > > * In the kernel code, each new vendor adds on to a global struct,
> > > > which might look something like:
> > > > struct hwprobe_vendor_space vendor_space[] = {
> > > > {
> > > > .mvendorid = VENDOR_THEAD,
> > > > .max_hwprobe_key = THEAD_MAX_HWPROBE_KEY, // currently
> > > > 1 or 0x8000000000000001 with what you've got.
> > > > .hwprobe_fn = thead_hwprobe
> > > > },
> > > > ...
> > > > };
> > > >
> > > > * A hwprobe_thead.c implements thead_hwprobe(), and is called
> > > > whenever the generic hwprobe encounters a key >=0x8000000000000000.
> > > > * Generic code for setting up the VDSO can then still call the
> > > > vendor-specific hwprobe_fn() repeatedly with an "all CPUs" mask from
> > > > the base to max_hwprobe_key and set up the cached tables in userspace.
> > > > * Since the VDSO data has limited space we may have to cap the number
> > > > of vendor keys we cache to be lower than max_hwprobe_key. Since the
> > > > data itself is not exposed to usermode we can raise this cap later if
> > > > needed.
> > >
> > > I know vendor extensions are kind of the "wild west" of riscv, but in
> > > spite of that I want to design a consistent API. The issue I had with
> > > having this "vendor space" for exposing vendor extensions was that this
> > > is something that is inherently the same for all vendors. I see a vendor
> > > space like this more applicable for something like
> > > "RISCV_HWPROBE_KEY_ZICBOZ_BLOCK_SIZE" where a vendor has a specific
> > > value they would like to expose. I do agree that having a vendor space
> > > is a good design choice, but I am not convinced that vendor extensions
> > > are the proper use-case.
> > >
> > > By having RISCV_HWPROBE_KEY_VENDOR_EXT_0 we can expose the vendor
> > > extensions in the same way that standard extensions are exposed, with a
> > > bitmask representing each extension. If these are instead in the vendor
> > > space, each vendor would probably be inclined to introduce a key like
> > > RISCV_HWPROBE_KEY_THEAD_EXT_0 that returns a bitmask of all of the thead
> > > vendor extensions. This duplicated effort is what I am trying to avoid.
> > > The alternative would be that vendors have a separate key for each
> > > vendor extension they would like to expose, but that is strictly less
> > > efficient than the existing bitmask probing.
> > >
> > > Do you think that having the vendor space is appropriate for vendor
> > > extensions given my concerns?
> >
> > I do see what you're going for. It's tidy for a bitmask to just let
> > anyone allocate the next bit, but leaves you with the same problem
> > when a vendor decides they want to expose an enum, or decides they
> > want to expose a bazillion things. I think a generalized version of
>
> This patch is strictly to expose if a vendor extension is supported,
> how does exposing enums factor in here?
>
> > the approach you've written would be: simply let vendors allocate keys
> > from the same global space we're already using. My worry was that it
>
> I am missing how my proposal suggests allowing vendors to allocate keys
> in a global space.
>
> > would turn into an expansive suburban sprawl of mostly dead bits, or
> > in the case of vendor-specific keys, full of "if (mvendor_id() !=
> > MINE) return 0;". My hope with the vendored keyspace is it would keep
>
> An application will always need to check vendorid before calling hwprobe
> with a vendor-specific feature? If that hwprobe support is a key above
> 1<<63, then the application will need to pass that vendor-specific key
> and interpret the vendor-specific value. If that hwprobe support is what
> I have proposed here, then the user calls the standardized vendor
> extension hwprobe endpoint and then needs to interpret the result based
> on the vendor of the cpumask. In both cases they need to check the
> vendorid of the cpumask. In the test case I added I failed to check the
> vendorid but I should have had that.
>
> > the sprawl from polluting the general array of (hopefully valuable)
> > info with stuff that's likely to become less relevant as time passes.
> > It also lowers the bar a bit to make it easier for vendors to expose
> > bits, as they don't consume global space for everyone for all of time,
> > just themselves.
>
> The vendor keys are tied directly to the vendor. So as it grows we would
> have something like:
>
> #define RISCV_HWPROBE_KEY_VENDOR_EXT_0 7
> /* T-Head */
> #define RISCV_HWPROBE_VENDOR_EXT_XTHEADVECTOR (1 << 0)
> #define RISCV_HWPROBE_VENDOR_EXT_XTHEAD2 (2 << 0)
> #define RISCV_HWPROBE_VENDOR_EXT_XTHEAD3 (3 << 0)
> /* Vendor 2 */
> #define RISCV_HWPROBE_VENDOR_EXT_XVENDOR1 (1 << 0)
> #define RISCV_HWPROBE_VENDOR_EXT_XVENDOR2 (2 << 0)
> /* Vendor 3 */
> ...
>
> The keys overlap between vendors. To determine which extension a vendor
> supports, hwprobe gets data from hart_isa_vendor[cpu]. If the vendor is
> vendor 2, it is not possible for a vendor extension from vendor 3 to end
> up in there. Only the extensions from that vendor can be supported by
> that vendor's hardware.

Gotcha. You're right I had misinterpreted this, thinking XTHEADVECTOR
was a valid bit regardless of mvendorid, and that other vendors would
have to choose new bits for their features and always return 0 for
XTHEADVECTOR. With your explanation, it seems like you're allocating
keys (in no particular order) whose meaning will change based on
mvendorid.

I guess I'm still not convinced that saving each vendor from having to
add a VENDOR_EXT key in their keyspace is worth the sacrifice of
spraying the vendor-specific keys across the generic keyspace. Are
there advantages to having a single key whose category is similar but
whose bits are entirely vendor-defined? Maybe if I were userspace and
my feature could be satisfied equivalently by XTHEADVECTOR or
XRIVOSOTHERTHING, then I could do one hwprobe call instead of two? But
I don't think the vendors are going to be consistent enough for that
equivalency to ever prove useful. The advantages in my head of the
separate vendor keyspace are:
* Keeps the kernel code simple: if key >= (1 >> 63)
vendor_config->do_hwprobe(), rather than having all these little calls
in each specific switch case for vendor_config->do_vendor_ext0(),
vendor_config->do_vendor_ext1(), etc.
* It extends easily into passing other forms of vendor hwprobe info
later, rather than solving only the case of risc-v extensions now, and
then having to do this all again for each additional category of
vendor data.
* Similarly, it discourages future vendors from trying to squint and
find a way to make a vaguely generic sounding category for their own
hwprobe key which will ultimately only ever be filled in by them
anyway.

-Evan