Date: Mon, 15 Oct 2012 02:30:09 +0100
From: Al Viro <viro@ZenIV.linux.org.uk>
To: linux-kernel@vger.kernel.org
Cc: Linus Torvalds <torvalds@linux-foundation.org>, linux-arch@vger.kernel.org,
        David Miller <davem@davemloft.net>,
        Benjamin Herrenschmidt <benh@kernel.crashing.org>
Subject: Re: [RFC][CFT][CFReview] execve and kernel_thread unification work
Message-ID: <20121015013009.GB2616@ZenIV.linux.org.uk>
References: <20121001213809.GA31155@ZenIV.linux.org.uk>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20121001213809.GA31155@ZenIV.linux.org.uk>
User-Agent: Mutt/1.5.21 (2010-09-15)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 7309
Lines: 128

On Mon, Oct 01, 2012 at 10:38:09PM +0100, Al Viro wrote:
> [with apologies for folks Cc'd, resent due to mis-autoexpanded l-k address
> on the original posting ;-/  Mea culpa...]
> 
> 	There's an interesting ongoing project around kernel_thread() and
> friends, including execve() variants.  I really need help from architecture
> maintainers on that one; I'd been able to handle (and test) quite a few
> architectures on my own [alpha, arm, m68k, powerpc, s390, sparc, x86, um]
> plus two more untested [frv, mn10300].  c6x patches had been supplied by
> Mark Salter; everything else remains to be done.  Right now it's at
> minus 1.2KLoC, quite a bit of that removed from asm glue and other black magic.

Update:
	* all infrastructure is in mainline now, along with conversion for
kernel_thread() callbacks to the form that allows really simple model for
kernel_execve() _without_ flagday changes.
	* #experimental-kernel_thread is gone; this stuff is in for-next
now.
	* a lot of architecture conversions had been done and some are
even tested.  Currently missing are only 7 - avr32, hexagon, m32r, openrisc,
score, tile and xtensa.  OTOH, a lot are completely untested.  I've put
per-architecture stuff into separate branches and I promise never rebase
those once arch maintainers will be OK with the stuff in them.  IOW, they'll
be safe to pull into respective architecture trees.

Folks, *please* review the stuff in signal.git#arch-*.  All of them are
completely independent.  I'll be glad to get ACKs/fixes/replacements/etc.
I've merged some of those into for-next, but that can change at any time -
it's not final; for-next will be rebased.  Obviously, I hope to get to
the situation when all of those branches (plus currently missing ones)
get into shape that satisfies architecture maintainers.  Once that happens,
all those branches will be merged into for-next.

I think the model is about final wrt kernel_thread()/kernel_execve()/
sys_execve().  There's one possible change on top of it, but it's reasonably
well-isolated from the rest.  As it is, the model to aim for is this:
	* select GENERIC_KERNEL_THREAD and GENERIC_KERNEL_EXECVE
	* kill local kernel_thread()/kernel_execve() implementations
	* generic kernel_thread() will call your copy_thread() with
NULL regs and fn/arg passed in the pair of arguments that are blindly
passed all the way through to copy_thread() - usp and stack_size resp.
In such case copy_thread() should arrange for the newborn to be woken
up in a function that is very similar to ret_from_fork().  The only
difference is that between the call of schedule_tail() and jumping into
the "return from syscall" code it should call fn(arg), using the data
left for it by copy_thread().
	* unlike the previous variant, ret_from_kernel_execve() is not
needed at all; no need to play longjmp()-like games when kernel_thread()
callbacks had been taught to return normally all the way out when
kernel_execve() returns 0; any updates of sp/manipulations of register
windows/etc. will happen without any magic.
	* provide current_pt_regs() if needed.  Default is
task_pt_regs(current), but you might want to optimize it and unlike
task_pt_regs() it must work whenever we are in syscall or in a kernel thread.
task_pt_regs(task), OTOH, is required to work only when task can be
interrogated by tracer.
	* no more syscalls-from-kernel, which often allows for simplifications
in the syscall entry/exit logics.  I haven't done any of those; up to the
architecture maintainers.

	One thing to keep in mind is that right now on SMP architectures
there's the third caller of copy_thread(), besides fork()/clone()/vfork()
(all pass userland pt_regs, with the address being current_pt_regs()) and
kernel_thread() (pass NULL pt_regs, kthread creation time).  It's fork_idle()
and it passes zero-filled pt_regs.  Frankly, I'm not even sure we want to
call copy_thread() in that case - the stuff set up by it goes nowhere.
We do that for each possible secondary CPU on SMP and we do *not* expose
those threads to scheduler.  When CPU gets initialized we have the
secondary bootstrap take that task_struct as current.  Its kernel stack,
thread_info, etc. are set up by said secondary bootstrap, overriding whatever
copy_thread() has done.  Eventually the bootstrap reaches cpu_idle(),
which is where we schedule away.  switch_to() done by schedule() is what
completes setting the things up; at that point they are ready to be woken
up - and not in ret_from_fork(), of course.
	For the majority of architectures nothing done by copy_thread() in
that case is used afterwards, so we might as well stop calling it when
copy_process() is called by fork_idle().  I know of only one dubious case -
powerpc sets thread->ksp_limit on copy_thread() and I'm not sure if
that's get overwritten in secondary bootstrap - the value would be still
correct and I don't see any obvious places where it would be reassigned
on that codepath.  There might be other cases like that, though.  I would
argue that for this kind of stuff the right place is arch_dup_task_struct(),
not copy_thread()...  Hell knows.  Note that we are pretty much hitting
the random path in copy_thread() in that case - what zeroed pt_regs look
like to user_regs() is arch-dependent.

	This is the possible change I've mentioned above.  Not sure; I'd
really like comments on that one.

	Branches in there:
arch-blackfin - conversion; completely untested
arch-cris - conversion; completely untested
arch-h8300 - conversion; completely untested
arch-microblaze - conversion; completely untested
arch-sh - conversion; completely untested
arch-unicore32 - conversion; completely untested
arch-ia64 - conversion; tested only on ski, which is worth very little
arch-c6x - followup to mainline; while it's minor, it's pretty much done
blindly and *really* needs review by maintainer.
arch-arm - contains heroic fix by rmk and nothing else.  Seems to work fine.
arch-m68k - minor followup to stuff already in mainline; works on aranym
arch-parisc - mostly the stuff tested by parisc folks + minor followup
similar to m68k one.
arch-s390 - minor followup to mainline; works in hercules
arch-arm64 - patches from maintainer with minor followup folded
arch-frv - minor followup to mainline, needs testing
arch-mn10300 - minor followup to mainline, needs testing
arch-mips - patches from me and Ralf; works on qemu
arch-sparc - conversions for sparc32 and sparc64, plus the syscall_noerror
optimization
arch-powerpc - minor followups to mainline, need review by maintainers

"Completely untested" in the above reads "no promises it even compiles, let
alone isn't horribly broken".  Please, treat that as a possible starting
point for doing the conversion for arch in question.  I might have misread
the CPU manuals, your switch_to() implementation, etc., or just have been
temporary insane from digging through dozens of architectures.  Hopefully
temporary, that is...

And folks, for pity sake, do the remaining seven.  The merge window is
over, so...

			Al, buggering off to get some VFS work done.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/