Re: [PATCH] nios2: enable HAVE_COPY_THREAD_TLS, switch to kernel_clone_args

[Christian Brauner]

On Fri, Jun 12, 2020 at 06:20:48AM +0000, Tan, Ley Foon wrote:
> 
> 
> > -----Original Message-----
> > From: Christian Brauner <christian.brauner@xxxxxxxxxx>
> > Sent: Sunday, May 24, 2020 6:06 PM
> > To: Tan, Ley Foon <ley.foon.tan@xxxxxxxxx>; Arnd Bergmann
> > <arnd@xxxxxxxx>
> > Cc: linux-kernel@xxxxxxxxxxxxxxx; Thomas Gleixner <tglx@xxxxxxxxxxxxx>;
> > Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>; Christian Brauner
> > <christian.brauner@xxxxxxxxxx>
> > Subject: [PATCH] nios2: enable HAVE_COPY_THREAD_TLS, switch to
> > kernel_clone_args
> > 
> > This is part of a larger series that aims at getting rid of the
> > copy_thread()/copy_thread_tls() split that makes the process creation
> > codepaths in the kernel more convoluted and error-prone than they need to
> > be.
> > I'm converting all the remaining arches that haven't yet switched and am
> > collecting individual acks. Once I have them, I'll send the whole series
> > removing the copy_thread()/copy_thread_tls() split, the
> > HAVE_COPY_THREAD_TLS define and the legacy do_fork() helper. The only
> > kernel-wide process creation entry point for anything not going directly
> > through the syscall path will then be based on struct kernel_clone_args.
> > No more danger of weird process creation abi quirks between architectures
> > hopefully, and easier to maintain overall.
> > It also unblocks implementing clone3() on architectures not support
> > copy_thread_tls(). Any architecture that wants to implement clone3() will
> > need to select HAVE_COPY_THREAD_TLS and thus need to implement
> > copy_thread_tls(). So both goals are connected but independently beneficial.
> > 
> > HAVE_COPY_THREAD_TLS means that a given architecture supports
> > CLONE_SETTLS and not setting it should usually mean that the architectures
> > doesn't implement it but that's not how things are. In fact all architectures
> > support CLONE_TLS it's just that they don't follow the calling convention that
> > HAVE_COPY_THREAD_TLS implies. That means all architectures can be
> > switched over to select HAVE_COPY_THREAD_TLS. Once that is done we can
> > remove that macro (yay, less code), the unnecessary do_fork() export in
> > kernel/fork.c, and also rename copy_thread_tls() back to copy_thread(). At
> > this point
> > copy_thread() becomes the main architecture specific part of process
> > creation but it will be the same layout and calling convention for all
> > architectures. (Once that is done we can probably cleanup each
> > copy_thread() function even more but that's for the future.)
> > 
> > Since nios2 does support CLONE_SETTLS there's no reason to not select
> > HAVE_COPY_THREAD_TLS. This brings us one step closer to getting rid of the
> > copy_thread()/copy_thread_tls() split we still have and ultimately the
> > HAVE_COPY_THREAD_TLS define in general. A lot of architectures have
> > already converted and nios2 is one of the few hat haven't yet. This also
> > unblocks implementing the clone3() syscall on nios2. Once that is done we
> > can get of another ARCH_WANTS_* macro.
> > 
> > Once Any architecture that supports HAVE_COPY_THREAD_TLS cannot call
> > the
> > do_fork() helper anymore. This is fine and intended since it should be
> > removed in favor of the new, cleaner _do_fork() calling convention based on
> > struct kernel_clone_args. In fact, most architectures have already switched.
> > With this patch, nios2 joins the other arches which can't use the fork(),
> > vfork(), clone(), clone3() syscalls directly and who follow the new process
> > creation calling convention that is based on struct kernel_clone_args which
> > we introduced a while back. This means less custom assembly in the
> > architectures entry path to set up the registers before calling into the process
> > creation helper and it is easier to to support new features without having to
> > adapt calling conventions. It also unifies all process creation paths between
> > fork(), vfork(), clone(), and clone3(). (We can't fix the ABI nightmare that
> > legacy
> > clone() is but we can prevent stuff like this happening in the future.)
> > 
> > For some more context, please see:
> > commit 606e9ad20094f6d500166881d301f31a51bc8aa7
> > Merge: ac61145a725a 457677c70c76
> > Author: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
> > Date:   Sat Jan 11 15:33:48 2020 -0800
> > 
> >     Merge tag 'clone3-tls-v5.5-rc6' of
> > git://git.kernel.org/pub/scm/linux/kernel/git/brauner/linux
> > 
> >     Pull thread fixes from Christian Brauner:
> >      "This contains a series of patches to fix CLONE_SETTLS when used with
> >       clone3().
> > 
> >       The clone3() syscall passes the tls argument through struct clone_args
> >       instead of a register. This means, all architectures that do not
> >       implement copy_thread_tls() but still support CLONE_SETTLS via
> >       copy_thread() expecting the tls to be located in a register argument
> >       based on clone() are currently unfortunately broken. Their tls value
> >       will be garbage.
> > 
> >       The patch series fixes this on all architectures that currently define
> >       __ARCH_WANT_SYS_CLONE3. It also adds a compile-time check to
> > ensure
> >       that any architecture that enables clone3() in the future is forced to
> >       also implement copy_thread_tls().
> > 
> >       My ultimate goal is to get rid of the copy_thread()/copy_thread_tls()
> >       split and just have copy_thread_tls() at some point in the not too
> >       distant future (Maybe even renaming copy_thread_tls() back to simply
> >       copy_thread() once the old function is ripped from all arches). This
> >       is dependent now on all arches supporting clone3().
> > 
> >       While all relevant arches do that now there are still four missing:
> >       ia64, m68k, sh and sparc. They have the system call reserved, but not
> >       implemented. Once they all implement clone3() we can get rid of
> >       ARCH_WANT_SYS_CLONE3 and HAVE_COPY_THREAD_TLS.
> > 
> > Note that in the meantime, m68k has already switched to the new calling
> > convention. And I've got sparc patches acked by Dave and ia64 is already
> > done too. You can find a link to a booting qemu nios2 system with all the
> > changes here at [1].
> > 
> > [1]: https://asciinema.org/a/333353
> > Cc: linux-kernel@xxxxxxxxxxxxxxx
> > Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > Cc: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> > Cc: Ley Foon Tan <ley.foon.tan@xxxxxxxxx>
> > Signed-off-by: Christian Brauner <christian.brauner@xxxxxxxxxx>
> Acked-by: Ley Foon Tan <ley.foon.tan@xxxxxxxxx>

Thanks!

> 
> Assume this patch will merge with your series.

If people agree with it, then yep. Just waiting for rc1 before sending
it out.

Christian