From: Linus Torvalds <torvalds@osdl.org>
To: Ingo Molnar <mingo@elte.hu>
Cc: Andrew Morton <akpm@osdl.org>,
linux-kernel@vger.kernel.org, arjan@infradead.org
Subject: Re: [patch] spinlocks: remove 'volatile'
Date: Thu, 6 Jul 2006 12:56:02 -0700 (PDT) [thread overview]
Message-ID: <Pine.LNX.4.64.0607061237300.3869@g5.osdl.org> (raw)
In-Reply-To: <20060706081639.GA24179@elte.hu>
On Thu, 6 Jul 2006, Ingo Molnar wrote:
>
> yeah. I tried this and it indeed slashed 42K off text size (0.2%):
>
> text data bss dec filename
> 20779489 6073834 3075176 29928499 vmlinux.volatile
> 20736884 6073834 3075176 29885894 vmlinux.non-volatile
>
> i booted the resulting allyesconfig bzImage and everything seems to be
> working fine. Find patch below.
Ok, the patch itself looks fine, but looking at some of the usage of the
spinlocks, I worry that the 'volatile' may actually have hidden a real
bug.
Here's what __raw_spin_lock() looks like on x86:
alternative_smp(
__raw_spin_lock_string,
__raw_spin_lock_string_up,
"=m" (lock->slock) : : "memory");
where the actual spinlock sequence itself isn't important.
What's important is what we tell the compiler, and the "=m" in particular.
The compiler will validly think that the spinlock assembly wil overwrite
the "slock" location _WITHOUT_ caring about the old value.
And that's dangerous, because it means that in theory a sequence like
spin_lock_init(&lock);
spin_lock(&lock);
might end up having the compiler optimize away the "unnecessary"
initialization code because the compiler (correctly, as far as we've told
it) thinks that the spin_lock() will overwrite the initial value.
And the "volatile" would have hidden that bug, for totally stupid and
unrelated reasons..
Now, admittedly, the above kind of code is insane, and possibly doesn't
exist, but still..
So I _think_ that we should change the "=m" to the much more correct "+m"
at the same time (or before - it's really a bug-fix regardless) as
removing the "volatile".
It would be interesting to hear whether that actually changes any code
generation (hopefully it doesn't, but if it does, that in itself is
interesting).
Btw, grepping for "=m" shows that semaphores may have the same bug, and in
fact, we seem to have that "volatile" there too (perhaps, in fact, because
somebody hit the bug and decided to fix it the simple way with "volatile"?
Who knows, it might even have been me, back before I realized how badly
gcc messes up volatiles)
Interestingly (or perhaps not), "atomic_add()" and friends (along with the
local_t stuff that seems to largely have been copied from the atomic code)
use a combination of "=m" and "m" in the assembler dst/src to avoid the
bug. They too would probably be better off using "+m".
I think that's because the whole "+m" thing is fairly new (and not well
documented either).
Appended is my previous test-program extended to show the difference
between "=m" and "+m"..
For me, I get:
horrid:
subl $16, %esp
movl $1, 12(%esp)
movl 12(%esp), %eax
movl %eax, a1
movl $1, b1
#APP
# overwrite a1
# modify b1
#NO_APP
addl $16, %esp
ret
notbad:
movl $1, b2
#APP
# overwrite a2
# modify b2
#NO_APP
ret
which shows that "horrid" (with volatile) generates truly crapola code due
to the volatile, but that the "overwrite a1" will not optimize away the
initializer.
And "notbad" has good code, but exactly because it's good code, the
compiler has also noticed that the "=m" means that the initialization of
"a2" is unnecessary, and has thus promptly removed it.
(Removing the initializer is _correct_ for that particular source code -
it's just that with the current x86 spinlock() macros, it would be
disastrous, and shows that your "just remove volatile" patch is dangerous
because of our incorrect inline assembly)
Linus
---
struct duh {
volatile int i;
};
void horrid(void)
{
extern struct duh a1;
extern struct duh b1;
a1 = (struct duh) { 1 };
b1.i = 1;
asm("# overwrite %0":"=m" (a1.i));
asm("# modify %0":"+m" (b1.i));
}
struct gaah {
int i;
};
void notbad(void)
{
extern struct gaah a2;
extern struct gaah b2;
a2 = (struct gaah) { 1 };
b2.i = 1;
asm("# overwrite %0":"=m" (a2.i));
asm("# modify %0":"+m" (b2.i));
}
next prev parent reply other threads:[~2006-07-06 19:56 UTC|newest]
Thread overview: 163+ messages / expand[flat|nested] mbox.gz Atom feed top
2006-07-05 8:49 [patch] uninline init_waitqueue_*() functions Ingo Molnar
2006-07-05 9:31 ` Andrew Morton
2006-07-05 9:32 ` Ingo Molnar
2006-07-05 9:53 ` Andrew Morton
2006-07-05 10:26 ` Ingo Molnar
2006-07-05 11:30 ` Ingo Molnar
2006-07-05 11:46 ` Ingo Molnar
2006-07-05 17:10 ` Andrew Morton
2006-07-05 19:35 ` Ingo Molnar
2006-07-05 20:18 ` Andrew Morton
2006-07-05 20:36 ` Linus Torvalds
2006-07-05 20:47 ` Ingo Molnar
2006-07-05 21:15 ` Ingo Molnar
2006-07-05 21:21 ` Linus Torvalds
2006-07-05 21:45 ` Ingo Molnar
2006-07-05 21:58 ` Randy.Dunlap
2006-07-05 22:00 ` Ingo Molnar
2006-07-05 22:10 ` Randy.Dunlap
2006-07-05 22:27 ` Ingo Molnar
2006-07-05 23:15 ` Adrian Bunk
2006-07-05 22:09 ` Linus Torvalds
2006-07-05 22:30 ` Ingo Molnar
2006-07-05 23:11 ` Linus Torvalds
2006-07-06 8:16 ` [patch] spinlocks: remove 'volatile' Ingo Molnar
2006-07-06 8:23 ` Ingo Molnar
2006-07-06 9:27 ` Heiko Carstens
2006-07-06 9:34 ` Arjan van de Ven
2006-07-06 11:59 ` linux-os (Dick Johnson)
2006-07-06 12:01 ` Arjan van de Ven
2006-07-06 12:29 ` linux-os (Dick Johnson)
2006-07-06 12:39 ` Arjan van de Ven
2006-07-06 13:39 ` J.A. Magallón
2006-07-06 13:43 ` Arjan van de Ven
2006-07-06 14:05 ` Chase Venters
2006-07-06 14:26 ` Andreas Schwab
2006-07-06 16:40 ` Nick Piggin
2006-07-06 23:19 ` David Schwartz
2006-07-06 18:15 ` Mark Lord
2006-07-06 19:15 ` Linus Torvalds
2006-07-06 19:33 ` Chris Friesen
2006-07-06 19:37 ` Mark Lord
2006-07-06 20:28 ` Chris Friesen
2006-07-06 20:32 ` Linus Torvalds
2006-07-06 19:38 ` Arjan van de Ven
2006-07-06 19:41 ` Måns Rullgård
2006-07-06 19:42 ` Jeff Garzik
2006-07-06 19:58 ` Linus Torvalds
2006-07-06 20:27 ` Mark Lord
2006-07-06 20:40 ` Chris Friesen
2006-07-06 21:00 ` Linus Torvalds
2006-07-08 8:40 ` Avi Kivity
2006-07-08 8:51 ` Arjan van de Ven
2006-07-08 9:20 ` Avi Kivity
2006-07-08 9:51 ` Arjan van de Ven
2006-07-08 10:18 ` Avi Kivity
2006-07-08 10:28 ` Thomas Gleixner
2006-07-09 4:19 ` Benjamin Herrenschmidt
2006-07-09 12:47 ` Avi Kivity
2006-07-09 19:16 ` David Schwartz
2006-07-09 19:51 ` Theodore Tso
2006-07-09 20:40 ` [OT] 'volatile' in userspace Rutger Nijlunsing
2006-07-10 3:42 ` Theodore Tso
2006-07-10 17:00 ` Joshua Hudson
2006-07-10 17:54 ` Nick Piggin
2006-07-11 7:48 ` Jan Engelhardt
2006-07-11 11:53 ` Nick Piggin
2006-07-11 19:09 ` Björn Steinbrink
2006-07-11 20:55 ` Jeff Dike
2006-07-10 18:54 ` Jeff Dike
2006-07-10 20:09 ` Philippe Troin
2006-07-10 20:52 ` Jeff Dike
2006-07-06 16:07 ` [patch] spinlocks: remove 'volatile' Linus Torvalds
2006-07-06 16:13 ` Linus Torvalds
2006-07-06 17:04 ` Jeff Garzik
2006-07-06 17:52 ` linux-os (Dick Johnson)
2006-07-06 18:00 ` Linus Torvalds
2006-07-06 21:02 ` J.A. Magallón
2006-07-06 21:12 ` Linus Torvalds
2006-07-06 18:10 ` Michael Buesch
2006-07-06 18:16 ` Chase Venters
2006-07-07 18:16 ` Krzysztof Halasa
2006-07-07 19:51 ` linux-os (Dick Johnson)
2006-07-07 20:25 ` Linus Torvalds
2006-07-07 21:22 ` linux-os (Dick Johnson)
2006-07-07 21:48 ` Chase Venters
2006-07-08 10:00 ` Krzysztof Halasa
2006-07-08 13:41 ` Chase Venters
2006-07-08 20:09 ` Krzysztof Halasa
2006-07-08 20:40 ` Chase Venters
2006-07-08 20:47 ` Chase Venters
2006-07-09 10:57 ` Krzysztof Halasa
2006-07-07 21:59 ` Linus Torvalds
2006-07-07 22:06 ` Linus Torvalds
2006-07-08 20:49 ` Pavel Machek
2006-07-08 21:43 ` Linus Torvalds
2006-07-07 22:05 ` J.A. Magallón
2006-07-07 22:22 ` Chase Venters
2006-07-07 22:37 ` J.A. Magallón
2006-07-08 9:33 ` David Schwartz
2006-07-07 22:49 ` J.A. Magallón
2006-07-07 22:59 ` Vadim Lobanov
2006-07-07 23:18 ` Chase Venters
2006-07-07 23:36 ` Davide Libenzi
2006-07-07 22:09 ` Ingo Molnar
2006-07-08 7:36 ` Ingo Molnar
2006-07-07 20:39 ` Krzysztof Halasa
2006-07-07 23:06 ` Björn Steinbrink
2006-07-08 8:36 ` Avi Kivity
2006-07-06 19:32 ` Jan Engelhardt
2006-07-06 20:26 ` Jeremy Fitzhardinge
2006-07-06 20:55 ` Jan Engelhardt
2006-07-06 21:07 ` Linus Torvalds
2006-07-06 19:56 ` Linus Torvalds [this message]
2006-07-06 20:34 ` Linus Torvalds
2006-07-08 22:50 ` Ralf Baechle
2006-07-09 3:09 ` Linus Torvalds
2006-07-09 3:07 ` Keith Owens
2006-07-09 3:29 ` Linus Torvalds
2006-07-09 3:43 ` Keith Owens
2006-07-09 3:58 ` Linus Torvalds
2006-07-09 6:13 ` David Miller
2006-07-09 14:28 ` Roman Zippel
2006-07-09 15:27 ` Linus Torvalds
2006-07-06 8:18 ` [patch] lockdep: clean up completion initializer in smpboot.c Ingo Molnar
2006-07-06 8:23 ` [patch] uninline init_waitqueue_*() functions Ingo Molnar
2006-07-06 9:02 ` Andrew Morton
2006-07-06 9:12 ` [patch] uninline init_waitqueue_head() Ingo Molnar
2006-07-05 20:45 ` [patch] uninline init_waitqueue_*() functions Ingo Molnar
2006-07-05 10:37 ` Christoph Hellwig
2006-07-05 10:44 ` Andrew Morton
2006-07-05 10:46 ` Andrew Morton
2006-07-05 10:47 ` Ingo Molnar
2006-07-05 9:54 ` [patch] uninline init_waitqueue_*() functions, fix Ingo Molnar
2006-07-07 10:21 [patch] spinlocks: remove 'volatile' Chuck Ebbert
2006-07-07 17:09 ` Linus Torvalds
2006-07-08 3:54 Albert Cahalan
2006-07-08 5:25 ` Linus Torvalds
2006-07-08 6:39 ` Nick Piggin
2006-07-08 18:53 ` Linus Torvalds
2006-07-08 9:45 ` Joe Korty
2006-07-08 9:52 ` Arjan van de Ven
2006-07-08 9:59 ` David Schwartz
2006-07-08 10:24 ` Thomas Gleixner
2006-07-08 15:49 ` Albert Cahalan
2006-07-08 18:31 ` Thomas Gleixner
2006-07-08 19:33 ` Albert Cahalan
2006-07-08 20:11 ` Linus Torvalds
2006-07-09 21:10 ` Pavel Machek
2006-07-09 22:18 ` Linus Torvalds
2006-07-10 16:25 ` marty fouts
2006-07-08 23:10 ` David Schwartz
2006-07-08 13:57 ` Andi Kleen
2006-07-08 19:13 ` Linus Torvalds
2006-07-08 10:45 ` Krzysztof Halasa
2006-07-08 6:12 trajce nedev
2006-07-08 6:19 ` Chase Venters
2006-07-08 6:45 ` trajce nedev
2006-07-08 6:58 ` Arjan van de Ven
2006-07-08 7:02 ` Vadim Lobanov
2006-07-08 13:46 ` Chase Venters
2006-07-09 4:39 ` Benjamin Herrenschmidt
2006-07-14 3:30 ` Steven Rostedt
2006-07-08 18:23 ` Linus Torvalds
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=Pine.LNX.4.64.0607061237300.3869@g5.osdl.org \
--to=torvalds@osdl.org \
--cc=akpm@osdl.org \
--cc=arjan@infradead.org \
--cc=linux-kernel@vger.kernel.org \
--cc=mingo@elte.hu \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.