Run-time feature detection on RISC-V is still quite a new feature and not all kernels support it, so knowing the kernel version will be important.

Linux kernel 6.6.36 on SpacemiT K1.

I compile and run riscv_feature.c as a demo, but it looks just fine.
(Zbc detection is only used in demo, not in libz.)

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/utsname.h>

#include <sys/auxv.h>

// see asm/hwcap.h
#define ISA_V_HWCAP (1 << ('v' - 'a'))
#define ISA_ZBC_HWCAP 45

struct riscv_cpu_features {
    int has_rvv;
    int has_zbc;
};

int is_kernel_version_greater_or_equal_to_6_5() {
    struct utsname buffer;
    uname(&buffer);

    int major, minor;
    if (sscanf(buffer.release, "%d.%d", &major, &minor) != 2) {
        // Something bad with uname()
        return 0;
    }

    if (major > 6 || major == 6 && minor >= 5)
        return 1;
    return 0;
}

void riscv_check_features_compile_time(struct riscv_cpu_features *features) {
    printf("riscv_check_features_compile_time\n");
#if defined(__riscv_v) && defined(__linux__)
    features->has_rvv = 1;
#else
    features->has_rvv = 0;
#endif

#if defined(__riscv_zbc) && defined(__linux__)
    features->has_zbc = 1;
#else
    features->has_zbc = 0;
#endif
}

void riscv_check_features_runtime(struct riscv_cpu_features *features) {
    printf("riscv_check_features_runtime\n");
    unsigned long hw_cap = getauxval(AT_HWCAP);
    features->has_rvv = hw_cap & ISA_V_HWCAP;
    features->has_zbc = hw_cap & ISA_ZBC_HWCAP;
}

void riscv_check_features(struct riscv_cpu_features *features) {
    if (is_kernel_version_greater_or_equal_to_6_5())
        riscv_check_features_runtime(features);
    else
        riscv_check_features_compile_time(features);

    printf("has_rvv = %08x\n", features->has_rvv);
    printf("has_zbc = %08x\n", features->has_zbc);
}

int main(){
    struct riscv_cpu_features rf;
    riscv_check_features(&rf);
}

~ % ./demo
riscv_check_features_runtime
has_rvv = 00200000
has_zbc = 0000002d

SSH connection to the RISC-V board using -C shows the same problem as asyncssh.

~ % ssh -C board@ip
board@ip's password:
Connection closed by ip port 22

The issue can be reproduced more efficiently in this way.

Things also happend on Linux kernel 6.1.
6.1 won't run into runtime check, so i assume that something is wrong with is_kernel_version_greater_or_equal_to_6_5().
Then i force feature checking in compile time with -DWITH_RUNTIME_CPU_DETECTION=on and

void Z_INTERNAL riscv_check_features(struct riscv_cpu_features *features) {
    // if (is_kernel_version_greater_or_equal_to_6_5())
    //     riscv_check_features_runtime(features);
    // else
        riscv_check_features_compile_time(features);
}

Connection succeed.
Any idea about is_kernel_version_greater_or_equal_to_6_5()?

@chLFF

If your CPU supports RVV, please replace the riscv_check_features function (

void Z_INTERNAL riscv_check_features(struct riscv_cpu_features *features) {
    features->has_rvv = 1;
}

If the error is reproducible, then the problem is in the implementation of zlib-ng algorithms for RISC-V.

@phprus
The replacement you mentioned in #1769 (comment) is basically the same as my modification in #1769 (comment). They both make connection work (with -DWITH_RUNTIME_CPU_DETECTION=on).

And thanks for reminding me.
With -DWITH_RVV=off -DWITH_RUNTIME_CPU_DETECTION=on, the connection succeed.
With -DWITH_RVV=on -DWITH_RUNTIME_CPU_DETECTION=off, the connection succeed.
With -DWITH_RVV=on -DWITH_RUNTIME_CPU_DETECTION=on (original setting), the connection failed.
So it seems that WITH_RVV and WITH_RUNTIME_CPU_DETECTION can not be turned on together.

#1769 (comment) is not equivalent to #1769 (comment)

For test we need to manual set of features->has_rvv to 1.

@chLFF

If your CPU supports RVV, please replace the riscv_check_features function (

zlib-ng/arch/riscv/riscv_features.c

Lines 47 to 52 in 5b04d9c

	void Z_INTERNAL riscv_check_features(struct riscv_cpu_features *features) {
	if (is_kernel_version_greater_or_equal_to_6_5())
	riscv_check_features_runtime(features);
	else
	riscv_check_features_compile_time(features);
	}

) with:

void Z_INTERNAL riscv_check_features(struct riscv_cpu_features *features) {
    features->has_rvv = 1;
}

If the error is reproducible, then the problem is in the implementation of zlib-ng algorithms for RISC-V.

I follow this to manual set of features->has_rvv to 1. No error shows up during connection.
i.e.
With -DWITH_RVV=on -DWITH_RUNTIME_CPU_DETECTION=on, is_kernel_version_greater_or_equal_to_6_5() commented out, the connection succeed.

And i don't understand why the equivalence is false. Because i also manually set features->has_rvv to 1 but only inside riscv_check_features_compile_time(features) which determined during compiling.

And i don't understand why the equivalence is false. Because i also manually set features->has_rvv to 1 but only inside riscv_check_features_compile_time(features) which determined during compiling.

riscv_check_features_compile_time has macro conditions.

To differentiate a feature detector bug and an implementation bug, we need to completely remove the feature detector from the source and set the features->has_rvv flag unconditionally.

@phprus Thanks very much, i got it. We do have rvv feature and rvv is turned on and thing goes well. So that there is no problem with RISC-V zlib-ng algorithms? As i said, is_kernel_version_greater_or_equal_to_6_5() causes the error?

@phprus Thanks very much, i got it. We do have rvv feature and rvv is turned on and thing goes well. So that there is no problem with RISC-V zlib-ng algorithms? As i said, is_kernel_version_greater_or_equal_to_6_5() causes the error?

Only function call in is_kernel_version_greater_or_equal_to_6_5() that can cause issues is uname... You could write a simple program to check what it returns for buffer.release and compare if you get the same result if you output it in stderr while running Python... If it returns garbage or string with no NULL-termination, then it can crash...

I will try to get all the feature detection fixes added in #1770, when I know the root cause of all the failures and crashes.