syscall: Allow the sv_fetch_syscall_args to return an uninitialized sysent on error

[dchagin]

Summary

To avoid magic handling of non-existing system calls the ABI
sv_fetch_syscall_args() can simply return ENOSYS.
This will be used in the following Linux emulation layer changes.

MFC after: 1 week

[kib]

Can you explain more? Does linuxolator need syscall slot 0?

[dchagin]

In D41859#954114, @kib wrote:

Can you explain more? Does linuxolator need syscall slot 0?

I want to deorbit linux_nosys hack from syscalls.master and fix sv_fetch_syscall_args: https://reviews.freebsd.org/D41902 & https://reviews.freebsd.org/D41901
Due to native nosys() kills binary we use linux_nosys hack

[kib]

D41901 sounds as a hack, again.

Am I right that the issue is that nosys() sends signal, while Linux' nosys should not? If yes, there are two options:

1. add sysentvec flag to disallow sending signal
2. add sv_nosys method and call it from nosys() instead of direct tdsignal()

[dchagin]

In D41859#955044, @kib wrote:

D41901 sounds as a hack, again.

Am I right that the issue is that nosys() sends signal, while Linux' nosys should not? If yes, there are two options:

yes

1. add sysentvec flag to disallow sending signal
2. add sv_nosys method and call it from nosys() instead of direct tdsignal()

This way is so generic while for Linuxulator we just need to return ENOSYS. I updated D41901, now callp is initialized and fetch_syscall_args() returns ENOSYS for non-existent syscall. What do you think?

Btw, the way you proposed will not work, due to nosys marked as ABSENT and in such cases syscall_thread_enter() return ENOSYS.

Ah, this revision is not needed for D41901 now

[kib]

In D41859#955367, @dchagin wrote:

Btw, the way you proposed will not work, due to nosys marked as ABSENT and in such cases syscall_thread_enter() return ENOSYS.

Are you saying that sometimes we do not send SIGSYS for FreeBSD native processes?

[dchagin]

In D41859#955376, @kib wrote:

In D41859#955367, @dchagin wrote:

Btw, the way you proposed will not work, due to nosys marked as ABSENT and in such cases syscall_thread_enter() return ENOSYS.

Are you saying that sometimes we do not send SIGSYS for FreeBSD native processes?

yes, OBSOL syscalls, https://people.freebsd.org/~dchagin/nosys.c

[kib]

In D41859#955405, @dchagin wrote:

In D41859#955376, @kib wrote:

In D41859#955367, @dchagin wrote:

Btw, the way you proposed will not work, due to nosys marked as ABSENT and in such cases syscall_thread_enter() return ENOSYS.

Are you saying that sometimes we do not send SIGSYS for FreeBSD native processes?

yes, OBSOL syscalls, https://people.freebsd.org/~dchagin/nosys.c

And ABSENT as well?

I want to have the obsol() handling to be centralized, we need to have single place where the ENOSYS actions like psignal/log/uprintf happen, and it seems that it cannot be easily kept in nosysl(). Perhaps nosys() should become just return (ENOSYS); and dedicated code in syscallret() do the ENOSYS handling.

[dchagin]

In D41859#955406, @kib wrote:

In D41859#955405, @dchagin wrote:

In D41859#955376, @kib wrote:

In D41859#955367, @dchagin wrote:

Btw, the way you proposed will not work, due to nosys marked as ABSENT and in such cases syscall_thread_enter() return ENOSYS.

Are you saying that sometimes we do not send SIGSYS for FreeBSD native processes?

yes, OBSOL syscalls, https://people.freebsd.org/~dchagin/nosys.c

And ABSENT as well?

Yes, all SY_THR_ABSENT marked syscalls go throught syscall_thread_enter() which can return ENOSYS then sy_call() is skiped.

I want to have the obsol() handling to be centralized, we need to have single place where the ENOSYS actions like psignal/log/uprintf happen, and it seems that it cannot be easily kept in nosysl(). Perhaps nosys() should become just return (ENOSYS); and dedicated code in syscallret() do the ENOSYS handling.

I understand, however this requires one extra check on a "fast path", may be add a check to a slow path? Like in https://reviews.freebsd.org/D41929

[dchagin]

Due to D41976