x/sys/unix: TestSysvSharedMemory consistently failing on multiple builders

[bcmills]

TestSysvSharedMemory, added for #46084 in CL 327830, is consistently failing on several of the linux builders for less-common architectures:

--- FAIL: TestSysvSharedMemory (0.00s)
    sysvshm_unix_test.go:30: SysvShmGet: function not implemented
FAIL
FAIL	golang.org/x/sys/unix	0.571s

It isn't obvious to me whether we expect SysvShmGet to work on those platforms. If not, the test should call Skip on platforms / kernels on which it is not supported.

Since these builders are consistently failing (and thus not providing useful feedback about other changes), this fix needs to happen soon — either by fixing the test, adding skips, or reverting CL 327830 until someone has more time to fix it.

CC @tklauser @virtuald

greplogs --dashboard -md -l -e 'FAIL: TestSysvSharedMemory'

2021-09-30T14:19:18-969570c/04242ac/linux-ppc64le-buildlet
2021-09-30T14:19:18-969570c/04242ac/linux-ppc64le-power9osu
2021-09-30T14:19:18-969570c/04242ac/linux-s390x-ibm
2021-09-30T14:19:18-969570c/205640e/linux-mips64-rtrk
2021-09-30T14:19:18-969570c/205640e/linux-ppc64le-buildlet
2021-09-30T14:19:18-969570c/205640e/linux-ppc64le-power9osu
2021-09-30T14:19:18-969570c/205640e/linux-s390x-ibm
2021-09-30T14:19:18-969570c/6b0026a/linux-ppc64le-buildlet
2021-09-30T14:19:18-969570c/6b0026a/linux-ppc64le-power9osu
2021-09-30T14:19:18-969570c/6b0026a/linux-s390x-ibm
2021-09-30T14:19:18-969570c/7162c4c/linux-mips64le-mengzhuo
2021-09-30T14:19:18-969570c/7162c4c/linux-ppc64le-buildlet
2021-09-30T14:19:18-969570c/7162c4c/linux-ppc64le-power9osu
2021-09-30T14:19:18-969570c/7162c4c/linux-s390x-ibm
2021-09-30T14:19:18-969570c/7d57324/linux-ppc64le-buildlet
2021-09-30T14:19:18-969570c/7d57324/linux-ppc64le-power9osu
2021-09-30T14:19:18-969570c/7d57324/linux-s390x-ibm
2021-09-30T14:19:18-969570c/9112d29/linux-mips64-rtrk
2021-09-30T14:19:18-969570c/9112d29/linux-mips64le-mengzhuo
2021-09-30T14:19:18-969570c/9112d29/linux-ppc64le-buildlet
2021-09-30T14:19:18-969570c/9112d29/linux-ppc64le-power9osu
2021-09-30T14:19:18-969570c/9112d29/linux-s390x-ibm
2021-09-30T14:19:18-969570c/b043d23/linux-mips64le-mengzhuo
2021-09-30T14:19:18-969570c/b043d23/linux-ppc64le-buildlet
2021-09-30T14:19:18-969570c/b043d23/linux-ppc64le-power9osu
2021-09-30T14:19:18-969570c/b043d23/linux-s390x-ibm
2021-09-30T14:19:18-969570c/bd8009d/linux-mips64-rtrk
2021-09-30T14:19:18-969570c/bd8009d/linux-mips64le-mengzhuo
2021-09-30T14:19:18-969570c/bd8009d/linux-ppc64le-buildlet
2021-09-30T14:19:18-969570c/bd8009d/linux-ppc64le-power9osu
2021-09-30T14:19:18-969570c/bd8009d/linux-s390x-ibm
2021-09-30T14:19:18-969570c/d4aed7e/linux-mips64le-mengzhuo
2021-09-30T14:19:18-969570c/d4aed7e/linux-s390x-ibm

[gopherbot]

Change https://golang.org/cl/353392 mentions this issue: unix: skip TestSysvSharedMemory on ENOSYS

[gopherbot]

Change https://golang.org/cl/353398 mentions this issue: unix: SysvShmCtl use ipc_64 on mips64

[virtuald]

I believe the combination of 353392 and 353398 should address the test failures. However, I don't really understand the IPC_64 thing in 353398... if there was someone else who understood linux syscalls a bit better then a better fix could be made.

[virtuald]

I am a bit surprised that the tests didn't fail on my original commit in gerrit, but I assume it was just that it didn't run on all the architectures? I do think I've identified what's going on here now, and I'll push a better fix soon I hope.

After Linux 5.1, a bunch of architectures got SYS_SHMxxx syscalls (torvalds/linux@0d6040d). However, prior to that they only supported using the SYS_IPC entrypoint. Looking through glibc's source, it doesn't bother trying to detect this at compile time, instead there are two different macros __ASSUME_SYSVIPC_DEFAULT_IPC_64 and __ASSUME_DIRECT_SYSVIPC_SYSCALLS that are defined in architecture specific configuration files (kernel-features.h).

In sysdeps/unix/sysv/linux/ipc_priv.h there is this:

#ifdef __ASSUME_SYSVIPC_DEFAULT_IPC_64
# define __IPC_64      0x0
#else
# define __IPC_64      0x100
#endif

And in glibc/sysdeps/unix/sysv/linux/kernel-features.h this is defined:

/* The generic default __IPC_64 value is 0x0, however some architectures
   require a different value of 0x100.  */
#define __ASSUME_SYSVIPC_DEFAULT_IPC_64		1

It appears that __ASSUME_SYSVIPC_DEFAULT_IPC_64 is #undef for these architectures/conditions:

• alpha
• arm
• i386 (if __LINUX_KERNEL_VERSION < 0x050100)
• microblaze
• mips (if _MIPS_SIM == _ABIO32 && __LINUX_KERNEL_VERSION < 0x050100)
• ppc (if __LINUX_KERNEL_VERSION < 0x050100)
• s390 (if __LINUX_KERNEL_VERSION < 0x050100)
• sh4 (if __LINUX_KERNEL_VERSION < 0x050100)
• sparc (if __LINUX_KERNEL_VERSION < 0x050100)

__ASSUME_DIRECT_SYSVIPC_SYSCALLS is often disabled as well under similar conditions.

[bcmills]

I believe the combination of 353392 and 353398 should address the test failures.

Looks like the builder failures have indeed cleared.