| Message ID | 20230327163203.2918455-1-evan@rivosinc.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | RISC-V Hardware Probing User Interface | expand |
On Mon, Mar 27, 2023 at 09:31:57AM -0700, Evan Green wrote: Hey Evan, Patchwork has a rake of complaints about the series unfortunately: https://patchwork.kernel.org/project/linux-riscv/list/?series=734234 Some of the checkpatch whinging may be spurious, but there's some definitely valid stuff in there! > Evan Green (6): > RISC-V: Move struct riscv_cpuinfo to new header > RISC-V: Add a syscall for HW probing > RISC-V: hwprobe: Add support for RISCV_HWPROBE_BASE_BEHAVIOR_IMA > RISC-V: hwprobe: Support probing of misaligned access performance > selftests: Test the new RISC-V hwprobe interface > RISC-V: Add hwprobe vDSO function and data And this one breaks the build for !MMU kernels unfortunately. Thanks, Conor.
Am Montag, 27. März 2023, 18:31:57 CEST schrieb Evan Green: > > There's been a bunch of off-list discussions about this, including at > Plumbers. The original plan was to do something involving providing an > ISA string to userspace, but ISA strings just aren't sufficient for a > stable ABI any more: in order to parse an ISA string users need the > version of the specifications that the string is written to, the version > of each extension (sometimes at a finer granularity than the RISC-V > releases/versions encode), and the expected use case for the ISA string > (ie, is it a U-mode or M-mode string). That's a lot of complexity to > try and keep ABI compatible and it's probably going to continue to grow, > as even if there's no more complexity in the specifications we'll have > to deal with the various ISA string parsing oddities that end up all > over userspace. > > Instead this patch set takes a very different approach and provides a set > of key/value pairs that encode various bits about the system. The big > advantage here is that we can clearly define what these mean so we can > ensure ABI stability, but it also allows us to encode information that's > unlikely to ever appear in an ISA string (see the misaligned access > performance, for example). The resulting interface looks a lot like > what arm64 and x86 do, and will hopefully fit well into something like > ACPI in the future. > > The actual user interface is a syscall, with a vDSO function in front of > it. The vDSO function can answer some queries without a syscall at all, > and falls back to the syscall for cases it doesn't have answers to. > Currently we prepopulate it with an array of answers for all keys and > a CPU set of "all CPUs". This can be adjusted as necessary to provide > fast answers to the most common queries. > > An example series in glibc exposing this syscall and using it in an > ifunc selector for memcpy can be found at [1]. I'm about to send a v2 > of that series out that incorporates the vDSO function. > > I was asked about the performance delta between this and something like > sysfs. I created a small test program [2] and ran it on a Nezha D1 > Allwinner board. Doing each operation 100000 times and dividing, these > operations take the following amount of time: > - open()+read()+close() of /sys/kernel/cpu_byteorder: 3.8us > - access("/sys/kernel/cpu_byteorder", R_OK): 1.3us > - riscv_hwprobe() vDSO and syscall: .0094us > - riscv_hwprobe() vDSO with no syscall: 0.0091us Looks like this series spawned a thread on one of the riscv-lists [0]. As auxvals were mentioned in that thread, I was wondering what's the difference between doing a new syscall vs. putting the keys + values as architecture auxvec elements [1] ? I'm probably missing some simple issue but from looking at that stuff I fathom RISCV_HWPROBE_KEY_BASE_BEHAVIOR could also just be AT_RISCV_BASE_BEHAVIOR ? Heiko [0] https://lists.riscv.org/g/sig-toolchains/topic/97886491 [1] https://elixir.bootlin.com/linux/latest/source/arch/riscv/include/uapi/asm/auxvec.h
On Tue, Mar 28, 2023 at 1:35 PM Heiko Stübner <heiko@sntech.de> wrote: > > Am Montag, 27. März 2023, 18:31:57 CEST schrieb Evan Green: > > > > There's been a bunch of off-list discussions about this, including at > > Plumbers. The original plan was to do something involving providing an > > ISA string to userspace, but ISA strings just aren't sufficient for a > > stable ABI any more: in order to parse an ISA string users need the > > version of the specifications that the string is written to, the version > > of each extension (sometimes at a finer granularity than the RISC-V > > releases/versions encode), and the expected use case for the ISA string > > (ie, is it a U-mode or M-mode string). That's a lot of complexity to > > try and keep ABI compatible and it's probably going to continue to grow, > > as even if there's no more complexity in the specifications we'll have > > to deal with the various ISA string parsing oddities that end up all > > over userspace. > > > > Instead this patch set takes a very different approach and provides a set > > of key/value pairs that encode various bits about the system. The big > > advantage here is that we can clearly define what these mean so we can > > ensure ABI stability, but it also allows us to encode information that's > > unlikely to ever appear in an ISA string (see the misaligned access > > performance, for example). The resulting interface looks a lot like > > what arm64 and x86 do, and will hopefully fit well into something like > > ACPI in the future. > > > > The actual user interface is a syscall, with a vDSO function in front of > > it. The vDSO function can answer some queries without a syscall at all, > > and falls back to the syscall for cases it doesn't have answers to. > > Currently we prepopulate it with an array of answers for all keys and > > a CPU set of "all CPUs". This can be adjusted as necessary to provide > > fast answers to the most common queries. > > > > An example series in glibc exposing this syscall and using it in an > > ifunc selector for memcpy can be found at [1]. I'm about to send a v2 > > of that series out that incorporates the vDSO function. > > > > I was asked about the performance delta between this and something like > > sysfs. I created a small test program [2] and ran it on a Nezha D1 > > Allwinner board. Doing each operation 100000 times and dividing, these > > operations take the following amount of time: > > - open()+read()+close() of /sys/kernel/cpu_byteorder: 3.8us > > - access("/sys/kernel/cpu_byteorder", R_OK): 1.3us > > - riscv_hwprobe() vDSO and syscall: .0094us > > - riscv_hwprobe() vDSO with no syscall: 0.0091us > > Looks like this series spawned a thread on one of the riscv-lists [0]. > > As auxvals were mentioned in that thread, I was wondering what's the > difference between doing a new syscall vs. putting the keys + values as > architecture auxvec elements [1] ? The auxvec approach would also work. The primary difference is that auxvec bits are actively copied into every new process, forever. If you predict a slow pace of new bits coming in, the auxvec approach probably makes more sense. This series was born out of a prediction that this set of "stuff" was going to be larger than traditional x86/ARM architectures, fiddly (ie bits possibly representing specific versions of various extensions), evolving regularly over time, and heterogeneous between cores. With that sort of rubber band ball in mind, a key/value interface seemed to make more sense. -Evan
On Mon, Mar 27, 2023 at 11:34 PM Conor Dooley <conor.dooley@microchip.com> wrote: > > On Mon, Mar 27, 2023 at 09:31:57AM -0700, Evan Green wrote: > > Hey Evan, > > Patchwork has a rake of complaints about the series unfortunately: > https://patchwork.kernel.org/project/linux-riscv/list/?series=734234 > > Some of the checkpatch whinging may be spurious, but there's some > definitely valid stuff in there! > > > Evan Green (6): > > RISC-V: Move struct riscv_cpuinfo to new header > > RISC-V: Add a syscall for HW probing > > RISC-V: hwprobe: Add support for RISCV_HWPROBE_BASE_BEHAVIOR_IMA > > RISC-V: hwprobe: Support probing of misaligned access performance > > selftests: Test the new RISC-V hwprobe interface > > > RISC-V: Add hwprobe vDSO function and data > > And this one breaks the build for !MMU kernels unfortunately. Drat! Ok, thanks for the heads up. I'll go track these down. -Evan
There's been a bunch of off-list discussions about this, including at Plumbers. The original plan was to do something involving providing an ISA string to userspace, but ISA strings just aren't sufficient for a stable ABI any more: in order to parse an ISA string users need the version of the specifications that the string is written to, the version of each extension (sometimes at a finer granularity than the RISC-V releases/versions encode), and the expected use case for the ISA string (ie, is it a U-mode or M-mode string). That's a lot of complexity to try and keep ABI compatible and it's probably going to continue to grow, as even if there's no more complexity in the specifications we'll have to deal with the various ISA string parsing oddities that end up all over userspace. Instead this patch set takes a very different approach and provides a set of key/value pairs that encode various bits about the system. The big advantage here is that we can clearly define what these mean so we can ensure ABI stability, but it also allows us to encode information that's unlikely to ever appear in an ISA string (see the misaligned access performance, for example). The resulting interface looks a lot like what arm64 and x86 do, and will hopefully fit well into something like ACPI in the future. The actual user interface is a syscall, with a vDSO function in front of it. The vDSO function can answer some queries without a syscall at all, and falls back to the syscall for cases it doesn't have answers to. Currently we prepopulate it with an array of answers for all keys and a CPU set of "all CPUs". This can be adjusted as necessary to provide fast answers to the most common queries. An example series in glibc exposing this syscall and using it in an ifunc selector for memcpy can be found at [1]. I'm about to send a v2 of that series out that incorporates the vDSO function. I was asked about the performance delta between this and something like sysfs. I created a small test program [2] and ran it on a Nezha D1 Allwinner board. Doing each operation 100000 times and dividing, these operations take the following amount of time: - open()+read()+close() of /sys/kernel/cpu_byteorder: 3.8us - access("/sys/kernel/cpu_byteorder", R_OK): 1.3us - riscv_hwprobe() vDSO and syscall: .0094us - riscv_hwprobe() vDSO with no syscall: 0.0091us These numbers get farther apart if we query multiple keys, as sysfs will scale linearly with the number of keys, where the dedicated syscall stays the same. To frame these numbers, I also did a tight fork/exec/wait loop, which I measured as 4.8ms. So doing 4 open/read/close operations is a delta of about 0.3%, versus a single vDSO call is a delta of essentially zero. [1] https://public-inbox.org/libc-alpha/20230206194819.1679472-1-evan@rivosinc.com/T/#t [2] https://pastebin.com/x84NEKaS Changes in v5: - Added tags - Fixed misuse of ISA_EXT_c as bitmap, changed to use riscv_isa_extension_available() (Heiko, Conor) - Document the alternatives approach in the commit message (Conor and Heiko). - Fix __init call warnings by making probe_vendor_features() and thead_feature_probe_func() __init_or_module. - Fixed compat vdso compilation failure (lkp). Changes in v4: - Used real types in syscall prototypes (Arnd) - Fixed static line break in do_riscv_hwprobe() (Conor) - Added newlines between documentation lists (Conor) - Crispen up size types to size_t, and cpu indices to int (Joe) - Fix copy_from_user() return logic bug (found via kselftests!) - Add __user to SYSCALL_DEFINE() to fix warning - More newlines in BASE_BEHAVIOR_IMA documentation (Conor) - Add newlines to CPUPERF_0 documentation (Conor) - Add UNSUPPORTED value (Conor) - Switched from DT to alternatives-based probing (Rob) - Crispen up cpu index type to always be int (Conor) - Fixed selftests commit description, no more tiny libc (Mark Brown) - Fixed selftest syscall prototype types to match v4. - Added a prototype to fix -Wmissing-prototype warning (lkp@intel.com) - Fixed rv32 build failure (lkp@intel.com) - Make vdso prototype match syscall types update Changes in v3: - Updated copyright date in cpufeature.h - Fixed typo in cpufeature.h comment (Conor) - Refactored functions so that kernel mode can query too, in preparation for the vDSO data population. - Changed the vendor/arch/imp IDs to return a value of -1 on mismatch rather than failing the whole call. - Const cpumask pointer in hwprobe_mid() - Embellished documentation WRT cpu_set and the returned values. - Renamed hwprobe_mid() to hwprobe_arch_id() (Conor) - Fixed machine ID doc warnings, changed elements to c:macro:. - Completed dangling unistd.h comment (Conor) - Fixed line breaks and minor logic optimization (Conor). - Use riscv_cached_mxxxid() (Conor) - Refactored base ISA behavior probe to allow kernel probing as well, in prep for vDSO data initialization. - Fixed doc warnings in IMA text list, use :c:macro:. - Have hwprobe_misaligned return int instead of long. - Constify cpumask pointer in hwprobe_misaligned() - Fix warnings in _PERF_O list documentation, use :c:macro:. - Move include cpufeature.h to misaligned patch. - Fix documentation mismatch for RISCV_HWPROBE_KEY_CPUPERF_0 (Conor) - Use for_each_possible_cpu() instead of NR_CPUS (Conor) - Break early in misaligned access iteration (Conor) - Increase MISALIGNED_MASK from 2 bits to 3 for possible UNSUPPORTED future value (Conor) - Introduced vDSO function Changes in v2: - Factored the move of struct riscv_cpuinfo to its own header - Changed the interface to look more like poll(). Rather than supplying key_offset and getting back an array of values with numerically contiguous keys, have the user pre-fill the key members of the array, and the kernel will fill in the corresponding values. For any key it doesn't recognize, it will set the key of that element to -1. This allows usermode to quickly ask for exactly the elements it cares about, and not get bogged down in a back and forth about newer keys that older kernels might not recognize. In other words, the kernel can communicate that it doesn't recognize some of the keys while still providing the data for the keys it does know. - Added a shortcut to the cpuset parameters that if a size of 0 and NULL is provided for the CPU set, the kernel will use a cpu mask of all online CPUs. This is convenient because I suspect most callers will only want to act on a feature if it's supported on all CPUs, and it's a headache to dynamically allocate an array of all 1s, not to mention a waste to have the kernel loop over all of the offline bits. - Fixed logic error in if(of_property_read_string...) that caused crash - Include cpufeature.h in cpufeature.h to avoid undeclared variable warning. - Added a _MASK define - Fix random checkpatch complaints - Updated the selftests to the new API and added some more. - Fixed indentation, comments in .S, and general checkpatch complaints. Evan Green (6): RISC-V: Move struct riscv_cpuinfo to new header RISC-V: Add a syscall for HW probing RISC-V: hwprobe: Add support for RISCV_HWPROBE_BASE_BEHAVIOR_IMA RISC-V: hwprobe: Support probing of misaligned access performance selftests: Test the new RISC-V hwprobe interface RISC-V: Add hwprobe vDSO function and data Documentation/riscv/hwprobe.rst | 86 +++++++ Documentation/riscv/index.rst | 1 + arch/riscv/Kconfig | 1 + arch/riscv/errata/thead/errata.c | 10 + arch/riscv/include/asm/alternative.h | 5 + arch/riscv/include/asm/cpufeature.h | 23 ++ arch/riscv/include/asm/hwprobe.h | 13 + arch/riscv/include/asm/syscall.h | 4 + arch/riscv/include/asm/vdso/data.h | 17 ++ arch/riscv/include/asm/vdso/gettimeofday.h | 8 + arch/riscv/include/uapi/asm/hwprobe.h | 37 +++ arch/riscv/include/uapi/asm/unistd.h | 9 + arch/riscv/kernel/alternative.c | 19 ++ arch/riscv/kernel/compat_vdso/Makefile | 2 +- arch/riscv/kernel/cpu.c | 8 +- arch/riscv/kernel/cpufeature.c | 3 + arch/riscv/kernel/smpboot.c | 1 + arch/riscv/kernel/sys_riscv.c | 225 +++++++++++++++++- arch/riscv/kernel/vdso.c | 6 - arch/riscv/kernel/vdso/Makefile | 4 + arch/riscv/kernel/vdso/hwprobe.c | 52 ++++ arch/riscv/kernel/vdso/sys_hwprobe.S | 15 ++ arch/riscv/kernel/vdso/vdso.lds.S | 3 + tools/testing/selftests/Makefile | 1 + tools/testing/selftests/riscv/Makefile | 58 +++++ .../testing/selftests/riscv/hwprobe/Makefile | 10 + .../testing/selftests/riscv/hwprobe/hwprobe.c | 90 +++++++ .../selftests/riscv/hwprobe/sys_hwprobe.S | 12 + 28 files changed, 709 insertions(+), 14 deletions(-) create mode 100644 Documentation/riscv/hwprobe.rst create mode 100644 arch/riscv/include/asm/cpufeature.h create mode 100644 arch/riscv/include/asm/hwprobe.h create mode 100644 arch/riscv/include/asm/vdso/data.h create mode 100644 arch/riscv/include/uapi/asm/hwprobe.h create mode 100644 arch/riscv/kernel/vdso/hwprobe.c create mode 100644 arch/riscv/kernel/vdso/sys_hwprobe.S create mode 100644 tools/testing/selftests/riscv/Makefile create mode 100644 tools/testing/selftests/riscv/hwprobe/Makefile create mode 100644 tools/testing/selftests/riscv/hwprobe/hwprobe.c create mode 100644 tools/testing/selftests/riscv/hwprobe/sys_hwprobe.S