Re: Update kernel-features.h files for Linux 5.1

[Arnd Bergmann <arnd@arndb.de>]

On Fri, May 10, 2019 at 3:19 PM Stepan Golosunov <stepan@golosunov.pp.ru> wrote:
>
> 10.05.2019 в 14:27:13 +0400 Stepan Golosunov написал(а):
> > 09.05.2019 в 23:00:37 +0000 Joseph Myers написал:
> > > Linux 5.1 adds missing syscalls to the syscall table for many Linux
> > > kernel architectures.  This patch updates the kernel-features.h
> > > headers accordingly.  I believe the statfs64 structure used by alpha
> > > matches what the new kernel syscalls use, but that should be reviewed
> > > carefully.
> > >
> > > Tested with build-many-glibcs.py.
> >
> > The newly added direct ipc syscalls are different from the old ones:
> >
> > 1. They do not accept IPC_64.  This means that __IPC_64 should be set
> > to zero when new syscalls are used.  And new syscalls can not be used
> > for compat functions like __old_semctl.

Correct, the idea is that we can stop passing that flag on most architectures
when a C library only supports new kernels.

I expect that glibc will start using the direct system calls in 5 to 10 years,
once it stops running on linux-5.0 and earlier. Until then, the easiest way
is to keep using the ipc() wrapper.

> > (Hmm.  Is __old_msgctl already buggy due to its use of __IPC_64 when
> > __ASSUME_DIRECT_SYSVIPC_SYSCALLS is defined?  Is there any
> > architecture with both __ASSUME_DIRECT_SYSVIPC_SYSCALLS and
> > __old_msgctl currently?)

My spreadsheet [1] tells me that mips-n32, mips-n64, alpha, arm, microblaze
and xtensa all provide the direct syscalls traditionally but require passing
__IPC_64 for historic reasons. Out of these, only arm-oabi (no longer supported
in glibc) also has sys_ipc().
It would have been nice to change that, but it seemed not worth adding another
set of IPC syscalls for these.

> > 2. semtimedop does not exist on 32-bit ABIs.  They have
> > semtimedop_time64 instead.

Right, we had a long debate over that, and in the end I decided not to
add two versions of semtimedop() to keep the ABI more compact.

> > (I also think that ipc with IPCOP_semtimedop was accidentally made to
> > behave like semtimedop_time64 in 32-bit builds of linux 5.1.  But that
> > would be a kernel bug.)

Can you elaborate? The code I see in mainline is

        case SEMTIMEDOP:
                if (IS_ENABLED(CONFIG_64BIT) || !IS_ENABLED(CONFIG_64BIT_TIME))
                        return ksys_semtimedop(first, ptr, second,
                                (const struct __kernel_timespec __user *)fifth);
                else if (IS_ENABLED(CONFIG_COMPAT_32BIT_TIME))
                        return compat_ksys_semtimedop(first, ptr, second,
                                (const struct old_timespec32 __user *)fifth);
                else
                        return -ENOSYS;

Since both CONFIG_64BIT_TIME and CONFIG_COMPAT_32BIT_TIME
are always set on 32-bit architectures, SEMTIMEDOP passes
an old_timespec32 argument here. Am I missing something?

> And, after rereading
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=0d6040d4681735dfc47565de288525de405a5c99
>
> 3. There is no semop. semtimedop(_time64)? should be used instead.

Right.

      Arnd

[1] https://docs.google.com/spreadsheets/d/1QxMvW5jpVG2jb4RM9CQQl27-wVpNYOa-_3K2RVKifb0/edit#gid=0