riscv: use hwprobe syscall for capability detection

[ZenithalHourlyRate]

This originates from #24069 (comment)

hwprobe is a linux syscall specific to riscv architecture where key/value pairs can be queried to detect whether an extension exists. This corresponds to the capability defined in openssl.

hwprobe syscall is added in Linux 6.4 and <asm/hwprobe.h> is still updating (check lkml). To workaround the compile time/runtime linux version problem, examples are

• compile/run in system where kernel<6.4
• compile/run in system where kernel>=6.4 but <asm/hwprobe.h> is does not corresponds to the hwprobe.h openssl required
• compile in a new system and run in an old system

As for non-existing syscall, __has_include(<asm/hwprobe.h>) can be used at compile time and -ENOSYS can be used at runtime

As for unknown key/value query, the syscall would clear it, see

If a key is unknown to the kernel, its key field will be cleared to -1, and its value set to 0.
https://docs.kernel.org/arch/riscv/hwprobe.html

The only problem is the possible unknown macros from <asm/hwprobe.h> (undefined macro when compiling in old system) so directly coping value is the workaround.

As the original thread discussed, OPENSSL_riscvcap, when provided, should override hwprobe interface.

Documentation for OPENSSL_riscvcap together with hwprobe would be added in another PR.

Checklist

• documentation is added or updated

[ZenithalHourlyRate]

Cherry-picked #24069 on this branch and ran openssl on Canaan Kendryte K230 where kernel=6.8.0 to test whether hwprobe works, the results matched

$ unset OPENSSL_riscvcap
$ ./openssl speed -evp chacha20-poly1305
type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes  16384 bytes
ChaCha20-Poly1305    33989.34k    52381.21k    75642.11k    79155.54k    90438.59k    90581.67k
$ export OPENSSL_riscvcap="rv64gc"
$ ./openssl speed -evp chacha20-poly1305
type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes  16384 bytes
ChaCha20-Poly1305    34001.63k    53398.65k    59937.11k    62083.66k    63053.82k    63018.33k
$ export OPENSSL_riscvcap="rv64gc_v_zbb"
$ ./openssl speed -evp chacha20-poly1305
type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes  16384 bytes
ChaCha20-Poly1305    33614.25k    52176.50k    76595.11k    79170.56k    90750.98k    90608.98k

[ZenithalHourlyRate]

Now cpuinfo can be displayed. Benchmark result using #24069

$ ./openssl info -cpusettings
OPENSSL_riscvcap=ZBA_ZBB_ZBC_ZBS_V
$ ./openssl speed -evp chacha20-poly1305
CPUINFO: OPENSSL_riscvcap=ZBA_ZBB_ZBC_ZBS_V
type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes  16384 bytes
ChaCha20-Poly1305    33657.82k    52458.13k    75558.49k    78885.42k    90753.71k    90570.75k

$ export OPENSSL_riscvcap="rv64gc_zba_zbb_zbc_zbs"
$ ./openssl info -cpusettings
OPENSSL_riscvcap=ZBA_ZBB_ZBC_ZBS env:rv64gc_zba_zbb_zbc_zbs
$ ./openssl speed -evp chacha20-poly1305
CPUINFO: OPENSSL_riscvcap=ZBA_ZBB_ZBC_ZBS env:rv64gc_zba_zbb_zbc_zbs
type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes  16384 bytes
ChaCha20-Poly1305    33741.30k    52294.55k    59968.85k    62311.77k    62792.63k    62941.87k

[ZenithalHourlyRate]

Documentation for OPENSSL_riscvcap and its interaction with hwprobe is added. Cc @kroeckx in #24069 (comment)

The limitation of the environment variable parser, as documented, would be fixed in another PR.

[ZenithalHourlyRate]

Wonder if anything should be done for this PR

[paulidale]

My approval stands

[paulidale]

@openssl/committers

[t8m]

A few nits

[ZenithalHourlyRate]

Requested changes addressed

[t8m]

@paulidale please reconfirm

[openssl-machine]

This pull request is ready to merge

[t8m]

Merged to the master branch. Thank you for your contribution.

[cyyself]

It's better to replace 4 here with RISCV_HWPROBE_KEY_IMA_EXT_0 with #define RISCV_HWPROBE_KEY_IMA_EXT_0 4 somewhere.