mbox series

[00/13] preempt: Make preempt count unconditional

Message ID 20200914204209.256266093@linutronix.de (mailing list archive)
Headers show
Series preempt: Make preempt count unconditional | expand

Message

Thomas Gleixner Sept. 14, 2020, 8:42 p.m. UTC
Folks!

While working on various preempt count related things, I stumbled (again)
over the inconsistency of our preempt count handling.

The handling of preempt_count() is inconsistent accross kernel
configurations. On kernels which have PREEMPT_COUNT=n
preempt_disable/enable() and the lock/unlock functions are not affecting
the preempt count, only local_bh_disable/enable() and _bh variants of
locking, soft interrupt delivery, hard interrupt and NMI context affect it.

It's therefore impossible to have a consistent set of checks which provide
information about the context in which a function is called. In many cases
it makes sense to have seperate functions for seperate contexts, but there
are valid reasons to avoid that and handle different calling contexts
conditionally.

The lack of such indicators which work on all kernel configuratios is a
constant source of trouble because developers either do not understand the
implications or try to work around this inconsistency in weird
ways. Neither seem these issues be catched by reviewers and testing.

Recently merged code does:

	 gfp = preemptible() ? GFP_KERNEL : GFP_ATOMIC;

Looks obviously correct, except for the fact that preemptible() is
unconditionally false for CONFIF_PREEMPT_COUNT=n, i.e. all allocations in
that code use GFP_ATOMIC on such kernels.

Attempts to make preempt count unconditional and consistent have been
rejected in the past with handwaving performance arguments.

Freshly conducted benchmarks did not reveal any measurable impact from
enabling preempt count unconditionally. On kernels with CONFIG_PREEMPT_NONE
or CONFIG_PREEMPT_VOLUNTARY the preempt count is only incremented and
decremented but the result of the decrement is not tested. Contrary to that
enabling CONFIG_PREEMPT which tests the result has a small but measurable
impact due to the conditional branch/call.

It's about time to make essential functionality of the kernel consistent
accross the various preemption models.

The series is also available from git:

   git://git.kernel.org/pub/scm/linux/kernel/git/tglx/devel.git preempt

That's the first part of a larger effort related to preempt count:

 1) The analysis of the usage sites of in_interrupt(), in_atomic(),
    in_softirq() is still ongoing, but so far the number of buggy users is
    clearly the vast majority. There will be seperate patch series
    (currently 46 and counting) to address these issues once the analysis
    is complete in the next days.

 2) The long discussed state tracking of local irq disable in preempt count
    which accounts interrupt disabled sections as atomic and avoids issuing
    costly instructions (sti, cli, popf or their non X86 counterparts) when
    the state does not change, i.e. nested irq_save() or irq_restore(). I
    have this working on X86 already and contrary to my earlier attempts
    this was reasonably straight forward due to the recent entry/exit code
    consolidation.

    What I've not done yet is to optimize the preempt count handling
    of the [un]lock_irq* operations so they handle the interrupt disabled
    state and the preempt count modification in one go. That's an obvious
    add on, but correctness first ...

 3) Lazy interrupt disabling as a straight forward extension to #2. This
    avoids the actual disabling at the CPU level completely and catches an
    incoming interrupt in the low level entry code, modifies the interrupt
    disabled state on the return stack, notes the interrupt as pending in
    software and raises it again when interrupts are reenabled. This has
    still a few issues which I'm hunting down (cpuidle is unhappy ...)

Thanks,

	tglx
---
 arch/arm/include/asm/assembler.h                                 |   11 --
 arch/arm/kernel/iwmmxt.S                                         |    2 
 arch/arm/mach-ep93xx/crunch-bits.S                               |    2 
 arch/xtensa/kernel/entry.S                                       |    2 
 drivers/gpu/drm/i915/Kconfig.debug                               |    1 
 drivers/gpu/drm/i915/i915_utils.h                                |    3 
 include/linux/bit_spinlock.h                                     |    4 -
 include/linux/lockdep.h                                          |    6 -
 include/linux/pagemap.h                                          |    4 -
 include/linux/preempt.h                                          |   37 +---------
 include/linux/uaccess.h                                          |    6 -
 kernel/Kconfig.preempt                                           |    4 -
 kernel/sched/core.c                                              |    6 -
 lib/Kconfig.debug                                                |    3 
 lib/Kconfig.debug.rej                                            |   14 +--
 tools/testing/selftests/rcutorture/configs/rcu/SRCU-t            |    1 
 tools/testing/selftests/rcutorture/configs/rcu/SRCU-u            |    1 
 tools/testing/selftests/rcutorture/configs/rcu/TINY01            |    1 
 tools/testing/selftests/rcutorture/doc/TINY_RCU.txt              |    5 -
 tools/testing/selftests/rcutorture/doc/TREE_RCU-kconfig.txt      |    1 
 tools/testing/selftests/rcutorture/formal/srcu-cbmc/src/config.h |    1 
 21 files changed, 23 insertions(+), 92 deletions(-)

Comments

Thomas Gleixner Sept. 14, 2020, 9:55 p.m. UTC | #1
On Mon, Sep 14 2020 at 13:59, Linus Torvalds wrote:
> On Mon, Sep 14, 2020 at 1:45 PM Thomas Gleixner <tglx@linutronix.de> wrote:
>>
>> Recently merged code does:
>>
>>          gfp = preemptible() ? GFP_KERNEL : GFP_ATOMIC;
>>
>> Looks obviously correct, except for the fact that preemptible() is
>> unconditionally false for CONFIF_PREEMPT_COUNT=n, i.e. all allocations in
>> that code use GFP_ATOMIC on such kernels.
>
> I don't think this is a good reason to entirely get rid of the
> no-preempt thing.

I did not say that this is a good reason. It just illustrates the
problem.

> The above is just garbage. It's bogus. You can't do it.
>
> Blaming the no-preempt code for this bug is extremely unfair, imho.

I'm not blaming the no-preempt code. I'm blaming inconsistency and there
is no real good argument for inconsistent behaviour, TBH.

> And the no-preempt code does help make for much better code generation
> for simple spinlocks.

Yes it does generate better code, but I tried hard to spot a difference
in various metrics exposed by perf. It's all in the noise and I only
can spot a difference when the actual preemption check after the
decrement which still depends on CONFIG_PREEMPT is in place, but that's
not the case for PREEMPT_NONE or PREEMPT_VOLUNTARY kernels where the
decrement is just a decrement w/o any conditional behind it.

> Where is that horribly buggy recent code? It's not in that exact
> format, certainly, since 'grep' doesn't find it.

Bah, that was stuff in next which got dropped again.

But just look at any check which uses preemptible(), especially those
which check !preemptible():

In the X86 #GP handler we have:

	/*
	 * To be potentially processing a kprobe fault and to trust the result
	 * from kprobe_running(), we have to be non-preemptible.
	 */
	if (!preemptible() &&
	    kprobe_running() &&
	    kprobe_fault_handler(regs, X86_TRAP_GP))
		goto exit;

and a similar check in the S390 code in kprobe_exceptions_notify(). That
all magically 'works' because that code might have been actually tested
with lockdep enabled which enforces PREEMPT_COUNT...

The SG code has some interesting usage as well:

		if (miter->__flags & SG_MITER_ATOMIC) {
			WARN_ON_ONCE(preemptible());
			kunmap_atomic(miter->addr);

How is that WARN_ON_ONCE() supposed to catch anything? Especially as
calling code does:

	flags = SG_MITER_TO_SG;
	if (!preemptible())
		flags |= SG_MITER_ATOMIC;

which is equally useless on kernels which have PREEMPT_COUNT=n.

There are bugs which are related to in_atomic() or other in_***() usage
all over the place as well.

Inconsistency at the core level is a clear recipe for disaster and at
some point we have to bite the bullet and accept that consistency is
more important than the non measurable 3 cycles?

Thanks,

        tglx
Herbert Xu Sept. 15, 2020, 6:22 a.m. UTC | #2
On Tue, Sep 15, 2020 at 09:20:59AM +0300, Ard Biesheuvel wrote:
>
> The documentation of kmap_atomic() states the following:
> 
>  * The use of kmap_atomic/kunmap_atomic is discouraged - kmap/kunmap
>  * gives a more generic (and caching) interface. But kmap_atomic can
>  * be used in IRQ contexts, so in some (very limited) cases we need
>  * it.
> 
> so if this is no longer accurate, perhaps we should fix it?

This hasn't been accurate for at least ten years :)

> But another reason I tried to avoid kmap_atomic() is that it disables
> preemption unconditionally, even on 64-bit architectures where HIGHMEM
> is irrelevant. So using kmap_atomic() here means that the bulk of
> WireGuard packet encryption runs with preemption disabled, essentially
> for legacy reasons.

Agreed.  We should definitely fix that.

Thanks,
Paul E. McKenney Sept. 15, 2020, 5:25 p.m. UTC | #3
On Mon, Sep 14, 2020 at 01:59:15PM -0700, Linus Torvalds wrote:
> On Mon, Sep 14, 2020 at 1:45 PM Thomas Gleixner <tglx@linutronix.de> wrote:
> >
> > Recently merged code does:
> >
> >          gfp = preemptible() ? GFP_KERNEL : GFP_ATOMIC;
> >
> > Looks obviously correct, except for the fact that preemptible() is
> > unconditionally false for CONFIF_PREEMPT_COUNT=n, i.e. all allocations in
> > that code use GFP_ATOMIC on such kernels.
> 
> I don't think this is a good reason to entirely get rid of the no-preempt thing.
> 
> The above is just garbage. It's bogus. You can't do it.
> 
> Blaming the no-preempt code for this bug is extremely unfair, imho.
> 
> And the no-preempt code does help make for much better code generation
> for simple spinlocks.
> 
> Where is that horribly buggy recent code? It's not in that exact
> format, certainly, since 'grep' doesn't find it.

It would be convenient for that "gfp =" code to work, as this would
allow better cache locality while invoking RCU callbacks, and would
further provide better robustness to callback floods.  The full story
is quite long, but here are alternatives have not yet been proven to be
abject failures:

1.	Use workqueues to do the allocations in a clean context.
	While waiting for the allocations, the callbacks are queued
	in the old cache-busting manner.  This functions correctly,
	but in the meantime (which on busy systems can be some time)
	the cache locality and robustness are lost.

2.	Provide the ability to allocate memory in raw atomic context.
	This is extremely effective, especially when used in combination
	with #1 above, but as you might suspect, the MM guys don't like
	it much.

In contrast, with Thomas's patch series, call_rcu() and kvfree_rcu()
could just look at preemptible() to see whether or not it was safe to
allocate memory, even in !PREEMPT kernels -- and in the common case,
it almost always would be safe.  It is quite possible that this approach
would work in isolation, or failing that, that adding #1 above would do
the trick.

I understand that this is all very hand-wavy, and I do apologize for that.
If you really want the full sad story with performance numbers and the
works, let me know!

							Thanx, Paul
Linus Torvalds Sept. 15, 2020, 5:29 p.m. UTC | #4
On Tue, Sep 15, 2020 at 12:24 AM Thomas Gleixner <tglx@linutronix.de> wrote:
>
> Alternatively we just make highmem a bit more expensive by making these
> maps preemptible. RT is doing this for a long time and it's not that
> horrible.

Ack.

In fact, I've wanted to start just removing kmap support entirely. At
some point it's not so much about "I have an old machine that wants
HIGHMEM" but about "I have an old CPU, and I'll just run an old
kernel".

It's not that 32-bit is irrelevant, it's that 32-bit with large
amounts of memory is irrelevant.

Last time this was discussed, iirc the main issue was some
questionable old ARM chips that were still very common in embedded
environments, even with large memory.

But we could definitely start de-emphasizing HIGHMEM.

                     Linus
Linus Torvalds Sept. 15, 2020, 5:35 p.m. UTC | #5
On Tue, Sep 15, 2020 at 1:39 AM Thomas Gleixner <tglx@linutronix.de> wrote:
>
> OTOH, having a working 'preemptible()' or maybe better named
> 'can_schedule()' check makes tons of sense to make decisions about
> allocation modes or other things.

No. I think that those kinds of decisions about actual behavior are
always simply fundamentally wrong.

Note that this is very different from having warnings about invalid
use. THAT is correct. It may not warn in all configurations, but that
doesn't matter: what matters is that it warns in common enough
configurations that developers will catch it.

So having a warning in "might_sleep()" that doesn't always trigger,
because you have a limited configuration that can't even detect the
situation, that's fine and dandy and intentional.

But having code like

       if (can_schedule())
           .. do something different ..

is fundamentally complete and utter garbage.

It's one thing if you test for "am I in hardware interrupt context".
Those tests aren't great either, but at least they make sense.

But a driver - or some library routine - making a difference based on
some nebulous "can I schedule" is fundamentally and basically WRONG.

If some code changes behavior, it needs to be explicit to the *caller*
of that code.

So this is why GFP_ATOMIC is fine, but "if (!can_schedule())
do_something_atomic()" is pure shite.

And I am not IN THE LEAST interested in trying to help people doing
pure shite. We need to fix them. Like the crypto code is getting
fixed.

                   Linus
Thomas Gleixner Sept. 15, 2020, 7:57 p.m. UTC | #6
On Tue, Sep 15 2020 at 10:35, Linus Torvalds wrote:
> On Tue, Sep 15, 2020 at 1:39 AM Thomas Gleixner <tglx@linutronix.de> wrote:
>>
>> OTOH, having a working 'preemptible()' or maybe better named
>> 'can_schedule()' check makes tons of sense to make decisions about
>> allocation modes or other things.
>
> No. I think that those kinds of decisions about actual behavior are
> always simply fundamentally wrong.
>
> Note that this is very different from having warnings about invalid
> use. THAT is correct. It may not warn in all configurations, but that
> doesn't matter: what matters is that it warns in common enough
> configurations that developers will catch it.

You wish. I just found a 7 year old bug in a 10G network driver which
surely would have been found if people would enable debug configs and
not just run the crap on their PREEMPT_NONE, all debug off kernel. And
that driver is not subject to bitrot, it gets regular bug fixes from
people who seem to care (distro folks).

> So having a warning in "might_sleep()" that doesn't always trigger,
> because you have a limited configuration that can't even detect the
> situation, that's fine and dandy and intentional.

and lets people get away with their crap.

> But having code like
>
>        if (can_schedule())
>            .. do something different ..
>
> is fundamentally complete and utter garbage.
>
> It's one thing if you test for "am I in hardware interrupt context".
> Those tests aren't great either, but at least they make sense.

They make sense in limited situations like exception handlers and such
which really have to know from which context an exception was raised.

But with the above reasoning such checks do not make sense in any other
general code. 'in hard interrupt context' is just another context where
you can't do stuff which you can do when in preemptible task context.

Most tests are way broader than a single context. in_interrupt() is true
for hard interrupt, soft interrupt delivery and all BH disabled
contexts, which is completely ill defined.

> But a driver - or some library routine - making a difference based on
> some nebulous "can I schedule" is fundamentally and basically WRONG.
>
> If some code changes behavior, it needs to be explicit to the *caller*
> of that code.

I'm fine with that, but then we have to be consequent and ban _all_ of
these and not just declare can_schedule() to be a bad one.

Thanks,

        tglx
Daniel Vetter Sept. 16, 2020, 7:37 a.m. UTC | #7
On Tue, Sep 15, 2020 at 7:35 PM Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> On Tue, Sep 15, 2020 at 1:39 AM Thomas Gleixner <tglx@linutronix.de> wrote:
> >
> > OTOH, having a working 'preemptible()' or maybe better named
> > 'can_schedule()' check makes tons of sense to make decisions about
> > allocation modes or other things.
>
> No. I think that those kinds of decisions about actual behavior are
> always simply fundamentally wrong.
>
> Note that this is very different from having warnings about invalid
> use. THAT is correct. It may not warn in all configurations, but that
> doesn't matter: what matters is that it warns in common enough
> configurations that developers will catch it.
>
> So having a warning in "might_sleep()" that doesn't always trigger,
> because you have a limited configuration that can't even detect the
> situation, that's fine and dandy and intentional.
>
> But having code like
>
>        if (can_schedule())
>            .. do something different ..
>
> is fundamentally complete and utter garbage.
>
> It's one thing if you test for "am I in hardware interrupt context".
> Those tests aren't great either, but at least they make sense.
>
> But a driver - or some library routine - making a difference based on
> some nebulous "can I schedule" is fundamentally and basically WRONG.
>
> If some code changes behavior, it needs to be explicit to the *caller*
> of that code.
>
> So this is why GFP_ATOMIC is fine, but "if (!can_schedule())
> do_something_atomic()" is pure shite.
>
> And I am not IN THE LEAST interested in trying to help people doing
> pure shite. We need to fix them. Like the crypto code is getting
> fixed.

Just figured I'll throw my +1 in from reading too many (gpu) drivers.
Code that tries to cleverly adjust its behaviour depending upon the
context it's running in is harder to understand and blows up in more
interesting ways. We still have drm_can_sleep() and it's mostly just
used for debug code, and I've largely ended up just deleting
everything that used it because when you're driver is blowing up the
last thing you want is to realize your debug code and output can't be
relied upon. Or worse, that the only Oops you have is the one in the
debug code, because the real one scrolled away - the original idea
behind drm_can_sleep was to make all the modeset code work
automagically both in normal ioctl/kworker context and in the panic
handlers or kgdb callbacks. Wishful thinking at best.

Also at least for me that extends to everything, e.g. I much prefer
explicit spin_lock and spin_lock_irq vs magic spin_lock_irqsave for
locks shared with interrupt handlers, since the former two gives me
clear information from which contexts such function can be called.
Other end is the memalloc_no*_save/restore functions, where I recently
made a real big fool of myself because I didn't realize how much that
impacts everything that's run within - suddenly "GFP_KERNEL for small
stuff never fails" is wrong everywhere.

It's all great for debugging and sanity checks (and we run with all
that stuff enabled in our CI), but really semantic changes depending
upon magic context checks freak my out :-)
-Daniel
Paul E. McKenney Sept. 16, 2020, 3:29 p.m. UTC | #8
On Wed, Sep 16, 2020 at 09:37:17AM +0200, Daniel Vetter wrote:
> On Tue, Sep 15, 2020 at 7:35 PM Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
> >
> > On Tue, Sep 15, 2020 at 1:39 AM Thomas Gleixner <tglx@linutronix.de> wrote:
> > >
> > > OTOH, having a working 'preemptible()' or maybe better named
> > > 'can_schedule()' check makes tons of sense to make decisions about
> > > allocation modes or other things.
> >
> > No. I think that those kinds of decisions about actual behavior are
> > always simply fundamentally wrong.
> >
> > Note that this is very different from having warnings about invalid
> > use. THAT is correct. It may not warn in all configurations, but that
> > doesn't matter: what matters is that it warns in common enough
> > configurations that developers will catch it.
> >
> > So having a warning in "might_sleep()" that doesn't always trigger,
> > because you have a limited configuration that can't even detect the
> > situation, that's fine and dandy and intentional.
> >
> > But having code like
> >
> >        if (can_schedule())
> >            .. do something different ..
> >
> > is fundamentally complete and utter garbage.
> >
> > It's one thing if you test for "am I in hardware interrupt context".
> > Those tests aren't great either, but at least they make sense.
> >
> > But a driver - or some library routine - making a difference based on
> > some nebulous "can I schedule" is fundamentally and basically WRONG.
> >
> > If some code changes behavior, it needs to be explicit to the *caller*
> > of that code.
> >
> > So this is why GFP_ATOMIC is fine, but "if (!can_schedule())
> > do_something_atomic()" is pure shite.
> >
> > And I am not IN THE LEAST interested in trying to help people doing
> > pure shite. We need to fix them. Like the crypto code is getting
> > fixed.
> 
> Just figured I'll throw my +1 in from reading too many (gpu) drivers.
> Code that tries to cleverly adjust its behaviour depending upon the
> context it's running in is harder to understand and blows up in more
> interesting ways. We still have drm_can_sleep() and it's mostly just
> used for debug code, and I've largely ended up just deleting
> everything that used it because when you're driver is blowing up the
> last thing you want is to realize your debug code and output can't be
> relied upon. Or worse, that the only Oops you have is the one in the
> debug code, because the real one scrolled away - the original idea
> behind drm_can_sleep was to make all the modeset code work
> automagically both in normal ioctl/kworker context and in the panic
> handlers or kgdb callbacks. Wishful thinking at best.
> 
> Also at least for me that extends to everything, e.g. I much prefer
> explicit spin_lock and spin_lock_irq vs magic spin_lock_irqsave for
> locks shared with interrupt handlers, since the former two gives me
> clear information from which contexts such function can be called.
> Other end is the memalloc_no*_save/restore functions, where I recently
> made a real big fool of myself because I didn't realize how much that
> impacts everything that's run within - suddenly "GFP_KERNEL for small
> stuff never fails" is wrong everywhere.
> 
> It's all great for debugging and sanity checks (and we run with all
> that stuff enabled in our CI), but really semantic changes depending
> upon magic context checks freak my out :-)

All fair, but some of us need to write code that must handle being
invoked from a wide variety of contexts.  Now perhaps you like the idea of
call_rcu() for schedulable contexts, call_rcu_nosched() when preemption
is disabled, call_rcu_irqs_are_disabled() when interrupts are disabled,
call_rcu_raw_atomic() from contexts where (for example) raw spinlocks
are held, and so on.  However, from what I can see, most people instead
consistently prefer that the RCU API instead be consolidated.

Some in-flight cache-efficiency work for kvfree_rcu() and call_rcu()
needs to be able to allocate memory occasionally.  It can do that when
invoked from some contexts, but not when invoked from others.  Right now,
in !PREEMPT kernels, it cannot tell, and must either do things to the
memory allocators that some of the MM hate or must unnecessarily invoke
workqueues.  Thomas's patches would allow the code to just allocate in
the common case when these primitives are invoked from contexts where
allocation is permitted.

If we want to restrict access to the can_schedule() or whatever primitive,
fine and good.  We can add a check to checkpatch.pl, for example.  Maybe
we can go back to the old brlock approach of requiring certain people's
review for each addition to the kernel.

But there really are use cases that it would greatly help.

							Thanx, Paul
Linus Torvalds Sept. 16, 2020, 6:32 p.m. UTC | #9
On Wed, Sep 16, 2020 at 8:29 AM Paul E. McKenney <paulmck@kernel.org> wrote:
>
> All fair, but some of us need to write code that must handle being
> invoked from a wide variety of contexts.

Note that I think that core functionality is different from random drivers.

Of course core code can (and will) look at things like

        if (in_interrupt())
            .. schedule work asynchronously ..

because core code ends up being called from odd places, and code like
that is expected to have understanding of the rules it plays with.

But something like RCU is a very different beast from some "walk the
scatter-gather list" code.

RCU does its work in the background, and works with lots of different
things. And it's so core and used everywhere that it knows about these
things. I mean, we literally have special code explicitly to let RCU
know "we entered kernel context now".

But something like a driver list walking thing should not be doing
different things behind peoples back depending on whether they hold
spinlocks or not. It should either just work regardless, or there
should be a flag (or special interface) for the "you're being called
in a crtitical region".

Because dynamically changing behavior really is very confusing.

               Linus
Linus Torvalds Sept. 16, 2020, 6:34 p.m. UTC | #10
On Tue, Sep 15, 2020 at 12:57 PM Thomas Gleixner <tglx@linutronix.de> wrote:
>
> You wish. I just found a 7 year old bug in a 10G network driver which
> surely would have been found if people would enable debug configs and
> not just run the crap on their PREEMPT_NONE, all debug off kernel. And
> that driver is not subject to bitrot, it gets regular bug fixes from
> people who seem to care (distro folks).

That driver clearly cannot be very well maintained. All the distro
kernels have the basic debug checks in place, afaik.

Is it some wonderful "enterprise hardware" garbage again that only
gets used in special data centers?

Becasue the "enterprise" people really are special. Very much in the
"short bus" special kind of way. The fact that they have fooled so
much of the industry into thinking that they are the competent and
serious people is a disgrace.

              Linus
Matthew Wilcox (Oracle) Sept. 16, 2020, 7:23 p.m. UTC | #11
On Mon, Sep 14, 2020 at 11:55:24PM +0200, Thomas Gleixner wrote:
> But just look at any check which uses preemptible(), especially those
> which check !preemptible():

hmm.

+++ b/include/linux/preempt.h
@@ -180,7 +180,9 @@ do { \
 
 #define preempt_enable_no_resched() sched_preempt_enable_no_resched()
 
+#ifndef MODULE
 #define preemptible()  (preempt_count() == 0 && !irqs_disabled())
+#endif
 
 #ifdef CONFIG_PREEMPTION
 #define preempt_enable() \


$ git grep -w preemptible drivers
(slightly trimmed by hand to remove, eg, comments)
drivers/firmware/arm_sdei.c:    WARN_ON_ONCE(preemptible());
drivers/firmware/arm_sdei.c:    WARN_ON_ONCE(preemptible());
drivers/firmware/arm_sdei.c:    WARN_ON_ONCE(preemptible());
drivers/firmware/arm_sdei.c:    WARN_ON_ONCE(preemptible());
drivers/firmware/arm_sdei.c:    WARN_ON(preemptible());
drivers/firmware/efi/efi-pstore.c:                            preemptible(), record->size, record->psi->buf);
drivers/irqchip/irq-gic-v4.c:   WARN_ON(preemptible());
drivers/irqchip/irq-gic-v4.c:   WARN_ON(preemptible());
drivers/scsi/hisi_sas/hisi_sas_main.c:          if (!preemptible())
drivers/xen/time.c:     BUG_ON(preemptible());

That only looks like two drivers that need more than WARNectomies.

Although maybe rcu_read_load_sched_held() or rcu_read_lock_any_held()
might get called from a module ...
Daniel Vetter Sept. 16, 2020, 8:29 p.m. UTC | #12
On Wed, Sep 16, 2020 at 5:29 PM Paul E. McKenney <paulmck@kernel.org> wrote:
>
> On Wed, Sep 16, 2020 at 09:37:17AM +0200, Daniel Vetter wrote:
> > On Tue, Sep 15, 2020 at 7:35 PM Linus Torvalds
> > <torvalds@linux-foundation.org> wrote:
> > >
> > > On Tue, Sep 15, 2020 at 1:39 AM Thomas Gleixner <tglx@linutronix.de> wrote:
> > > >
> > > > OTOH, having a working 'preemptible()' or maybe better named
> > > > 'can_schedule()' check makes tons of sense to make decisions about
> > > > allocation modes or other things.
> > >
> > > No. I think that those kinds of decisions about actual behavior are
> > > always simply fundamentally wrong.
> > >
> > > Note that this is very different from having warnings about invalid
> > > use. THAT is correct. It may not warn in all configurations, but that
> > > doesn't matter: what matters is that it warns in common enough
> > > configurations that developers will catch it.
> > >
> > > So having a warning in "might_sleep()" that doesn't always trigger,
> > > because you have a limited configuration that can't even detect the
> > > situation, that's fine and dandy and intentional.
> > >
> > > But having code like
> > >
> > >        if (can_schedule())
> > >            .. do something different ..
> > >
> > > is fundamentally complete and utter garbage.
> > >
> > > It's one thing if you test for "am I in hardware interrupt context".
> > > Those tests aren't great either, but at least they make sense.
> > >
> > > But a driver - or some library routine - making a difference based on
> > > some nebulous "can I schedule" is fundamentally and basically WRONG.
> > >
> > > If some code changes behavior, it needs to be explicit to the *caller*
> > > of that code.
> > >
> > > So this is why GFP_ATOMIC is fine, but "if (!can_schedule())
> > > do_something_atomic()" is pure shite.
> > >
> > > And I am not IN THE LEAST interested in trying to help people doing
> > > pure shite. We need to fix them. Like the crypto code is getting
> > > fixed.
> >
> > Just figured I'll throw my +1 in from reading too many (gpu) drivers.
> > Code that tries to cleverly adjust its behaviour depending upon the
> > context it's running in is harder to understand and blows up in more
> > interesting ways. We still have drm_can_sleep() and it's mostly just
> > used for debug code, and I've largely ended up just deleting
> > everything that used it because when you're driver is blowing up the
> > last thing you want is to realize your debug code and output can't be
> > relied upon. Or worse, that the only Oops you have is the one in the
> > debug code, because the real one scrolled away - the original idea
> > behind drm_can_sleep was to make all the modeset code work
> > automagically both in normal ioctl/kworker context and in the panic
> > handlers or kgdb callbacks. Wishful thinking at best.
> >
> > Also at least for me that extends to everything, e.g. I much prefer
> > explicit spin_lock and spin_lock_irq vs magic spin_lock_irqsave for
> > locks shared with interrupt handlers, since the former two gives me
> > clear information from which contexts such function can be called.
> > Other end is the memalloc_no*_save/restore functions, where I recently
> > made a real big fool of myself because I didn't realize how much that
> > impacts everything that's run within - suddenly "GFP_KERNEL for small
> > stuff never fails" is wrong everywhere.
> >
> > It's all great for debugging and sanity checks (and we run with all
> > that stuff enabled in our CI), but really semantic changes depending
> > upon magic context checks freak my out :-)
>
> All fair, but some of us need to write code that must handle being
> invoked from a wide variety of contexts.  Now perhaps you like the idea of
> call_rcu() for schedulable contexts, call_rcu_nosched() when preemption
> is disabled, call_rcu_irqs_are_disabled() when interrupts are disabled,
> call_rcu_raw_atomic() from contexts where (for example) raw spinlocks
> are held, and so on.  However, from what I can see, most people instead
> consistently prefer that the RCU API instead be consolidated.
>
> Some in-flight cache-efficiency work for kvfree_rcu() and call_rcu()
> needs to be able to allocate memory occasionally.  It can do that when
> invoked from some contexts, but not when invoked from others.  Right now,
> in !PREEMPT kernels, it cannot tell, and must either do things to the
> memory allocators that some of the MM hate or must unnecessarily invoke
> workqueues.  Thomas's patches would allow the code to just allocate in
> the common case when these primitives are invoked from contexts where
> allocation is permitted.
>
> If we want to restrict access to the can_schedule() or whatever primitive,
> fine and good.  We can add a check to checkpatch.pl, for example.  Maybe
> we can go back to the old brlock approach of requiring certain people's
> review for each addition to the kernel.
>
> But there really are use cases that it would greatly help.

We can deadlock in random fun places if random stuff we're calling
suddenly starts allocating. Sometimes. Maybe once in a blue moon, to
make it extra fun to reproduce. Maybe most driver subsystems are less
brittle, but gpu drivers definitely need to know about the details for
exactly this example. And yes gpu drivers use rcu for freeing
dma_fence structures, and that tends to happen in code that we only
recently figured out should really not allocate memory.

I think minimally you need to throw in an unconditional
fs_reclaim_acquire();fs_reclaim_release(); so that everyone who runs
with full debugging knows what might happen. It's kinda like
might_sleep, but a lot more specific. might_sleep() alone is not
enough, because in the specific code paths I'm thinking of (and
created special lockdep annotations for just recently) sleeping is
allowed, but any memory allocations with GFP_RECLAIM set are no-go.

Cheers, Daniel




--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Paul E. McKenney Sept. 16, 2020, 8:43 p.m. UTC | #13
On Wed, Sep 16, 2020 at 11:32:00AM -0700, Linus Torvalds wrote:
> On Wed, Sep 16, 2020 at 8:29 AM Paul E. McKenney <paulmck@kernel.org> wrote:
> >
> > All fair, but some of us need to write code that must handle being
> > invoked from a wide variety of contexts.
> 
> Note that I think that core functionality is different from random drivers.
> 
> Of course core code can (and will) look at things like
> 
>         if (in_interrupt())
>             .. schedule work asynchronously ..
> 
> because core code ends up being called from odd places, and code like
> that is expected to have understanding of the rules it plays with.
> 
> But something like RCU is a very different beast from some "walk the
> scatter-gather list" code.
> 
> RCU does its work in the background, and works with lots of different
> things. And it's so core and used everywhere that it knows about these
> things. I mean, we literally have special code explicitly to let RCU
> know "we entered kernel context now".
> 
> But something like a driver list walking thing should not be doing
> different things behind peoples back depending on whether they hold
> spinlocks or not. It should either just work regardless, or there
> should be a flag (or special interface) for the "you're being called
> in a crtitical region".
> 
> Because dynamically changing behavior really is very confusing.

Whew!  I feel much better now.  ;-)

							Thanx, Paul
Paul E. McKenney Sept. 16, 2020, 8:48 p.m. UTC | #14
On Wed, Sep 16, 2020 at 08:23:52PM +0100, Matthew Wilcox wrote:
> On Mon, Sep 14, 2020 at 11:55:24PM +0200, Thomas Gleixner wrote:
> > But just look at any check which uses preemptible(), especially those
> > which check !preemptible():
> 
> hmm.
> 
> +++ b/include/linux/preempt.h
> @@ -180,7 +180,9 @@ do { \
>  
>  #define preempt_enable_no_resched() sched_preempt_enable_no_resched()
>  
> +#ifndef MODULE
>  #define preemptible()  (preempt_count() == 0 && !irqs_disabled())
> +#endif
>  
>  #ifdef CONFIG_PREEMPTION
>  #define preempt_enable() \
> 
> 
> $ git grep -w preemptible drivers
> (slightly trimmed by hand to remove, eg, comments)
> drivers/firmware/arm_sdei.c:    WARN_ON_ONCE(preemptible());
> drivers/firmware/arm_sdei.c:    WARN_ON_ONCE(preemptible());
> drivers/firmware/arm_sdei.c:    WARN_ON_ONCE(preemptible());
> drivers/firmware/arm_sdei.c:    WARN_ON_ONCE(preemptible());
> drivers/firmware/arm_sdei.c:    WARN_ON(preemptible());
> drivers/firmware/efi/efi-pstore.c:                            preemptible(), record->size, record->psi->buf);
> drivers/irqchip/irq-gic-v4.c:   WARN_ON(preemptible());
> drivers/irqchip/irq-gic-v4.c:   WARN_ON(preemptible());
> drivers/scsi/hisi_sas/hisi_sas_main.c:          if (!preemptible())
> drivers/xen/time.c:     BUG_ON(preemptible());
> 
> That only looks like two drivers that need more than WARNectomies.

I could easily imagine someone thinking that these did something in
CONFIG_PREEMPT_NONE=y kernels.  In fact, I could easily imagine myself
making that mistake.  :-/

> Although maybe rcu_read_load_sched_held() or rcu_read_lock_any_held()
> might get called from a module ...

But yes, from the rcutorture module for certain and also from any other
RCU-using module that includes the usual RCU debug checks.

							Thanx, Paul
Paul E. McKenney Sept. 16, 2020, 8:58 p.m. UTC | #15
On Wed, Sep 16, 2020 at 10:29:06PM +0200, Daniel Vetter wrote:
> On Wed, Sep 16, 2020 at 5:29 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> >
> > On Wed, Sep 16, 2020 at 09:37:17AM +0200, Daniel Vetter wrote:
> > > On Tue, Sep 15, 2020 at 7:35 PM Linus Torvalds
> > > <torvalds@linux-foundation.org> wrote:
> > > >
> > > > On Tue, Sep 15, 2020 at 1:39 AM Thomas Gleixner <tglx@linutronix.de> wrote:
> > > > >
> > > > > OTOH, having a working 'preemptible()' or maybe better named
> > > > > 'can_schedule()' check makes tons of sense to make decisions about
> > > > > allocation modes or other things.
> > > >
> > > > No. I think that those kinds of decisions about actual behavior are
> > > > always simply fundamentally wrong.
> > > >
> > > > Note that this is very different from having warnings about invalid
> > > > use. THAT is correct. It may not warn in all configurations, but that
> > > > doesn't matter: what matters is that it warns in common enough
> > > > configurations that developers will catch it.
> > > >
> > > > So having a warning in "might_sleep()" that doesn't always trigger,
> > > > because you have a limited configuration that can't even detect the
> > > > situation, that's fine and dandy and intentional.
> > > >
> > > > But having code like
> > > >
> > > >        if (can_schedule())
> > > >            .. do something different ..
> > > >
> > > > is fundamentally complete and utter garbage.
> > > >
> > > > It's one thing if you test for "am I in hardware interrupt context".
> > > > Those tests aren't great either, but at least they make sense.
> > > >
> > > > But a driver - or some library routine - making a difference based on
> > > > some nebulous "can I schedule" is fundamentally and basically WRONG.
> > > >
> > > > If some code changes behavior, it needs to be explicit to the *caller*
> > > > of that code.
> > > >
> > > > So this is why GFP_ATOMIC is fine, but "if (!can_schedule())
> > > > do_something_atomic()" is pure shite.
> > > >
> > > > And I am not IN THE LEAST interested in trying to help people doing
> > > > pure shite. We need to fix them. Like the crypto code is getting
> > > > fixed.
> > >
> > > Just figured I'll throw my +1 in from reading too many (gpu) drivers.
> > > Code that tries to cleverly adjust its behaviour depending upon the
> > > context it's running in is harder to understand and blows up in more
> > > interesting ways. We still have drm_can_sleep() and it's mostly just
> > > used for debug code, and I've largely ended up just deleting
> > > everything that used it because when you're driver is blowing up the
> > > last thing you want is to realize your debug code and output can't be
> > > relied upon. Or worse, that the only Oops you have is the one in the
> > > debug code, because the real one scrolled away - the original idea
> > > behind drm_can_sleep was to make all the modeset code work
> > > automagically both in normal ioctl/kworker context and in the panic
> > > handlers or kgdb callbacks. Wishful thinking at best.
> > >
> > > Also at least for me that extends to everything, e.g. I much prefer
> > > explicit spin_lock and spin_lock_irq vs magic spin_lock_irqsave for
> > > locks shared with interrupt handlers, since the former two gives me
> > > clear information from which contexts such function can be called.
> > > Other end is the memalloc_no*_save/restore functions, where I recently
> > > made a real big fool of myself because I didn't realize how much that
> > > impacts everything that's run within - suddenly "GFP_KERNEL for small
> > > stuff never fails" is wrong everywhere.
> > >
> > > It's all great for debugging and sanity checks (and we run with all
> > > that stuff enabled in our CI), but really semantic changes depending
> > > upon magic context checks freak my out :-)
> >
> > All fair, but some of us need to write code that must handle being
> > invoked from a wide variety of contexts.  Now perhaps you like the idea of
> > call_rcu() for schedulable contexts, call_rcu_nosched() when preemption
> > is disabled, call_rcu_irqs_are_disabled() when interrupts are disabled,
> > call_rcu_raw_atomic() from contexts where (for example) raw spinlocks
> > are held, and so on.  However, from what I can see, most people instead
> > consistently prefer that the RCU API instead be consolidated.
> >
> > Some in-flight cache-efficiency work for kvfree_rcu() and call_rcu()
> > needs to be able to allocate memory occasionally.  It can do that when
> > invoked from some contexts, but not when invoked from others.  Right now,
> > in !PREEMPT kernels, it cannot tell, and must either do things to the
> > memory allocators that some of the MM hate or must unnecessarily invoke
> > workqueues.  Thomas's patches would allow the code to just allocate in
> > the common case when these primitives are invoked from contexts where
> > allocation is permitted.
> >
> > If we want to restrict access to the can_schedule() or whatever primitive,
> > fine and good.  We can add a check to checkpatch.pl, for example.  Maybe
> > we can go back to the old brlock approach of requiring certain people's
> > review for each addition to the kernel.
> >
> > But there really are use cases that it would greatly help.
> 
> We can deadlock in random fun places if random stuff we're calling
> suddenly starts allocating. Sometimes. Maybe once in a blue moon, to
> make it extra fun to reproduce. Maybe most driver subsystems are less
> brittle, but gpu drivers definitely need to know about the details for
> exactly this example. And yes gpu drivers use rcu for freeing
> dma_fence structures, and that tends to happen in code that we only
> recently figured out should really not allocate memory.
> 
> I think minimally you need to throw in an unconditional
> fs_reclaim_acquire();fs_reclaim_release(); so that everyone who runs
> with full debugging knows what might happen. It's kinda like
> might_sleep, but a lot more specific. might_sleep() alone is not
> enough, because in the specific code paths I'm thinking of (and
> created special lockdep annotations for just recently) sleeping is
> allowed, but any memory allocations with GFP_RECLAIM set are no-go.

Completely agreed!  Any allocation on any free path must be handled
-extremely- carefully.  To that end...

First, there is always a fallback in case the allocation fails.  Which
might have performance or corner-case robustness issues, but which will
at least allow forward progress.  Second, we consulted with a number of
MM experts to arrive at appropriate GFP_* flags (and their patience is
greatly appreciated).  Third, the paths that can allocate will do so about
one time of 500, so any issues should be spotted sooner rather than later.

So you are quite right to be concerned, but I believe we will be doing the
right things.  And based on his previous track record, I am also quite
certain that Mr. Murphy will be on hand to provide me any additional
education that I might require.

Finally, I have noted down your point about fs_reclaim_acquire() and
fs_reclaim_release().  Whether or not they prove to be needed, I do
appreciate your calling them to my attention.

							Thanx, Paul
Daniel Vetter Sept. 16, 2020, 9:43 p.m. UTC | #16
On Wed, Sep 16, 2020 at 10:58 PM Paul E. McKenney <paulmck@kernel.org> wrote:
>
> On Wed, Sep 16, 2020 at 10:29:06PM +0200, Daniel Vetter wrote:
> > On Wed, Sep 16, 2020 at 5:29 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> > >
> > > On Wed, Sep 16, 2020 at 09:37:17AM +0200, Daniel Vetter wrote:
> > > > On Tue, Sep 15, 2020 at 7:35 PM Linus Torvalds
> > > > <torvalds@linux-foundation.org> wrote:
> > > > >
> > > > > On Tue, Sep 15, 2020 at 1:39 AM Thomas Gleixner <tglx@linutronix.de> wrote:
> > > > > >
> > > > > > OTOH, having a working 'preemptible()' or maybe better named
> > > > > > 'can_schedule()' check makes tons of sense to make decisions about
> > > > > > allocation modes or other things.
> > > > >
> > > > > No. I think that those kinds of decisions about actual behavior are
> > > > > always simply fundamentally wrong.
> > > > >
> > > > > Note that this is very different from having warnings about invalid
> > > > > use. THAT is correct. It may not warn in all configurations, but that
> > > > > doesn't matter: what matters is that it warns in common enough
> > > > > configurations that developers will catch it.
> > > > >
> > > > > So having a warning in "might_sleep()" that doesn't always trigger,
> > > > > because you have a limited configuration that can't even detect the
> > > > > situation, that's fine and dandy and intentional.
> > > > >
> > > > > But having code like
> > > > >
> > > > >        if (can_schedule())
> > > > >            .. do something different ..
> > > > >
> > > > > is fundamentally complete and utter garbage.
> > > > >
> > > > > It's one thing if you test for "am I in hardware interrupt context".
> > > > > Those tests aren't great either, but at least they make sense.
> > > > >
> > > > > But a driver - or some library routine - making a difference based on
> > > > > some nebulous "can I schedule" is fundamentally and basically WRONG.
> > > > >
> > > > > If some code changes behavior, it needs to be explicit to the *caller*
> > > > > of that code.
> > > > >
> > > > > So this is why GFP_ATOMIC is fine, but "if (!can_schedule())
> > > > > do_something_atomic()" is pure shite.
> > > > >
> > > > > And I am not IN THE LEAST interested in trying to help people doing
> > > > > pure shite. We need to fix them. Like the crypto code is getting
> > > > > fixed.
> > > >
> > > > Just figured I'll throw my +1 in from reading too many (gpu) drivers.
> > > > Code that tries to cleverly adjust its behaviour depending upon the
> > > > context it's running in is harder to understand and blows up in more
> > > > interesting ways. We still have drm_can_sleep() and it's mostly just
> > > > used for debug code, and I've largely ended up just deleting
> > > > everything that used it because when you're driver is blowing up the
> > > > last thing you want is to realize your debug code and output can't be
> > > > relied upon. Or worse, that the only Oops you have is the one in the
> > > > debug code, because the real one scrolled away - the original idea
> > > > behind drm_can_sleep was to make all the modeset code work
> > > > automagically both in normal ioctl/kworker context and in the panic
> > > > handlers or kgdb callbacks. Wishful thinking at best.
> > > >
> > > > Also at least for me that extends to everything, e.g. I much prefer
> > > > explicit spin_lock and spin_lock_irq vs magic spin_lock_irqsave for
> > > > locks shared with interrupt handlers, since the former two gives me
> > > > clear information from which contexts such function can be called.
> > > > Other end is the memalloc_no*_save/restore functions, where I recently
> > > > made a real big fool of myself because I didn't realize how much that
> > > > impacts everything that's run within - suddenly "GFP_KERNEL for small
> > > > stuff never fails" is wrong everywhere.
> > > >
> > > > It's all great for debugging and sanity checks (and we run with all
> > > > that stuff enabled in our CI), but really semantic changes depending
> > > > upon magic context checks freak my out :-)
> > >
> > > All fair, but some of us need to write code that must handle being
> > > invoked from a wide variety of contexts.  Now perhaps you like the idea of
> > > call_rcu() for schedulable contexts, call_rcu_nosched() when preemption
> > > is disabled, call_rcu_irqs_are_disabled() when interrupts are disabled,
> > > call_rcu_raw_atomic() from contexts where (for example) raw spinlocks
> > > are held, and so on.  However, from what I can see, most people instead
> > > consistently prefer that the RCU API instead be consolidated.
> > >
> > > Some in-flight cache-efficiency work for kvfree_rcu() and call_rcu()
> > > needs to be able to allocate memory occasionally.  It can do that when
> > > invoked from some contexts, but not when invoked from others.  Right now,
> > > in !PREEMPT kernels, it cannot tell, and must either do things to the
> > > memory allocators that some of the MM hate or must unnecessarily invoke
> > > workqueues.  Thomas's patches would allow the code to just allocate in
> > > the common case when these primitives are invoked from contexts where
> > > allocation is permitted.
> > >
> > > If we want to restrict access to the can_schedule() or whatever primitive,
> > > fine and good.  We can add a check to checkpatch.pl, for example.  Maybe
> > > we can go back to the old brlock approach of requiring certain people's
> > > review for each addition to the kernel.
> > >
> > > But there really are use cases that it would greatly help.
> >
> > We can deadlock in random fun places if random stuff we're calling
> > suddenly starts allocating. Sometimes. Maybe once in a blue moon, to
> > make it extra fun to reproduce. Maybe most driver subsystems are less
> > brittle, but gpu drivers definitely need to know about the details for
> > exactly this example. And yes gpu drivers use rcu for freeing
> > dma_fence structures, and that tends to happen in code that we only
> > recently figured out should really not allocate memory.
> >
> > I think minimally you need to throw in an unconditional
> > fs_reclaim_acquire();fs_reclaim_release(); so that everyone who runs
> > with full debugging knows what might happen. It's kinda like
> > might_sleep, but a lot more specific. might_sleep() alone is not
> > enough, because in the specific code paths I'm thinking of (and
> > created special lockdep annotations for just recently) sleeping is
> > allowed, but any memory allocations with GFP_RECLAIM set are no-go.
>
> Completely agreed!  Any allocation on any free path must be handled
> -extremely- carefully.  To that end...
>
> First, there is always a fallback in case the allocation fails.  Which
> might have performance or corner-case robustness issues, but which will
> at least allow forward progress.  Second, we consulted with a number of
> MM experts to arrive at appropriate GFP_* flags (and their patience is
> greatly appreciated).  Third, the paths that can allocate will do so about
> one time of 500, so any issues should be spotted sooner rather than later.
>
> So you are quite right to be concerned, but I believe we will be doing the
> right things.  And based on his previous track record, I am also quite
> certain that Mr. Murphy will be on hand to provide me any additional
> education that I might require.
>
> Finally, I have noted down your point about fs_reclaim_acquire() and
> fs_reclaim_release().  Whether or not they prove to be needed, I do
> appreciate your calling them to my attention.

I just realized that since these dma_fence structs are refcounted and
userspace can hold references (directly, it can pass them around
behind file descriptors) we might never hit such a path until slightly
unusual or evil userspace does something interesting. Do you have
links to those patches? Some googling didn't turn up anything. I can
then figure out whether it's better to risk not spotting issues with
call_rcu vs slapping a memalloc_noio_save/restore around all these
critical section which force-degrades any allocation to GFP_ATOMIC at
most, but has the risk that we run into code that assumes "GFP_KERNEL
never fails for small stuff" and has a decidedly less tested fallback
path than rcu code.
-Daniel
Paul E. McKenney Sept. 16, 2020, 10:39 p.m. UTC | #17
On Wed, Sep 16, 2020 at 11:43:02PM +0200, Daniel Vetter wrote:
> On Wed, Sep 16, 2020 at 10:58 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> >
> > On Wed, Sep 16, 2020 at 10:29:06PM +0200, Daniel Vetter wrote:
> > > On Wed, Sep 16, 2020 at 5:29 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> > > >
> > > > On Wed, Sep 16, 2020 at 09:37:17AM +0200, Daniel Vetter wrote:
> > > > > On Tue, Sep 15, 2020 at 7:35 PM Linus Torvalds
> > > > > <torvalds@linux-foundation.org> wrote:
> > > > > >
> > > > > > On Tue, Sep 15, 2020 at 1:39 AM Thomas Gleixner <tglx@linutronix.de> wrote:
> > > > > > >
> > > > > > > OTOH, having a working 'preemptible()' or maybe better named
> > > > > > > 'can_schedule()' check makes tons of sense to make decisions about
> > > > > > > allocation modes or other things.
> > > > > >
> > > > > > No. I think that those kinds of decisions about actual behavior are
> > > > > > always simply fundamentally wrong.
> > > > > >
> > > > > > Note that this is very different from having warnings about invalid
> > > > > > use. THAT is correct. It may not warn in all configurations, but that
> > > > > > doesn't matter: what matters is that it warns in common enough
> > > > > > configurations that developers will catch it.
> > > > > >
> > > > > > So having a warning in "might_sleep()" that doesn't always trigger,
> > > > > > because you have a limited configuration that can't even detect the
> > > > > > situation, that's fine and dandy and intentional.
> > > > > >
> > > > > > But having code like
> > > > > >
> > > > > >        if (can_schedule())
> > > > > >            .. do something different ..
> > > > > >
> > > > > > is fundamentally complete and utter garbage.
> > > > > >
> > > > > > It's one thing if you test for "am I in hardware interrupt context".
> > > > > > Those tests aren't great either, but at least they make sense.
> > > > > >
> > > > > > But a driver - or some library routine - making a difference based on
> > > > > > some nebulous "can I schedule" is fundamentally and basically WRONG.
> > > > > >
> > > > > > If some code changes behavior, it needs to be explicit to the *caller*
> > > > > > of that code.
> > > > > >
> > > > > > So this is why GFP_ATOMIC is fine, but "if (!can_schedule())
> > > > > > do_something_atomic()" is pure shite.
> > > > > >
> > > > > > And I am not IN THE LEAST interested in trying to help people doing
> > > > > > pure shite. We need to fix them. Like the crypto code is getting
> > > > > > fixed.
> > > > >
> > > > > Just figured I'll throw my +1 in from reading too many (gpu) drivers.
> > > > > Code that tries to cleverly adjust its behaviour depending upon the
> > > > > context it's running in is harder to understand and blows up in more
> > > > > interesting ways. We still have drm_can_sleep() and it's mostly just
> > > > > used for debug code, and I've largely ended up just deleting
> > > > > everything that used it because when you're driver is blowing up the
> > > > > last thing you want is to realize your debug code and output can't be
> > > > > relied upon. Or worse, that the only Oops you have is the one in the
> > > > > debug code, because the real one scrolled away - the original idea
> > > > > behind drm_can_sleep was to make all the modeset code work
> > > > > automagically both in normal ioctl/kworker context and in the panic
> > > > > handlers or kgdb callbacks. Wishful thinking at best.
> > > > >
> > > > > Also at least for me that extends to everything, e.g. I much prefer
> > > > > explicit spin_lock and spin_lock_irq vs magic spin_lock_irqsave for
> > > > > locks shared with interrupt handlers, since the former two gives me
> > > > > clear information from which contexts such function can be called.
> > > > > Other end is the memalloc_no*_save/restore functions, where I recently
> > > > > made a real big fool of myself because I didn't realize how much that
> > > > > impacts everything that's run within - suddenly "GFP_KERNEL for small
> > > > > stuff never fails" is wrong everywhere.
> > > > >
> > > > > It's all great for debugging and sanity checks (and we run with all
> > > > > that stuff enabled in our CI), but really semantic changes depending
> > > > > upon magic context checks freak my out :-)
> > > >
> > > > All fair, but some of us need to write code that must handle being
> > > > invoked from a wide variety of contexts.  Now perhaps you like the idea of
> > > > call_rcu() for schedulable contexts, call_rcu_nosched() when preemption
> > > > is disabled, call_rcu_irqs_are_disabled() when interrupts are disabled,
> > > > call_rcu_raw_atomic() from contexts where (for example) raw spinlocks
> > > > are held, and so on.  However, from what I can see, most people instead
> > > > consistently prefer that the RCU API instead be consolidated.
> > > >
> > > > Some in-flight cache-efficiency work for kvfree_rcu() and call_rcu()
> > > > needs to be able to allocate memory occasionally.  It can do that when
> > > > invoked from some contexts, but not when invoked from others.  Right now,
> > > > in !PREEMPT kernels, it cannot tell, and must either do things to the
> > > > memory allocators that some of the MM hate or must unnecessarily invoke
> > > > workqueues.  Thomas's patches would allow the code to just allocate in
> > > > the common case when these primitives are invoked from contexts where
> > > > allocation is permitted.
> > > >
> > > > If we want to restrict access to the can_schedule() or whatever primitive,
> > > > fine and good.  We can add a check to checkpatch.pl, for example.  Maybe
> > > > we can go back to the old brlock approach of requiring certain people's
> > > > review for each addition to the kernel.
> > > >
> > > > But there really are use cases that it would greatly help.
> > >
> > > We can deadlock in random fun places if random stuff we're calling
> > > suddenly starts allocating. Sometimes. Maybe once in a blue moon, to
> > > make it extra fun to reproduce. Maybe most driver subsystems are less
> > > brittle, but gpu drivers definitely need to know about the details for
> > > exactly this example. And yes gpu drivers use rcu for freeing
> > > dma_fence structures, and that tends to happen in code that we only
> > > recently figured out should really not allocate memory.
> > >
> > > I think minimally you need to throw in an unconditional
> > > fs_reclaim_acquire();fs_reclaim_release(); so that everyone who runs
> > > with full debugging knows what might happen. It's kinda like
> > > might_sleep, but a lot more specific. might_sleep() alone is not
> > > enough, because in the specific code paths I'm thinking of (and
> > > created special lockdep annotations for just recently) sleeping is
> > > allowed, but any memory allocations with GFP_RECLAIM set are no-go.
> >
> > Completely agreed!  Any allocation on any free path must be handled
> > -extremely- carefully.  To that end...
> >
> > First, there is always a fallback in case the allocation fails.  Which
> > might have performance or corner-case robustness issues, but which will
> > at least allow forward progress.  Second, we consulted with a number of
> > MM experts to arrive at appropriate GFP_* flags (and their patience is
> > greatly appreciated).  Third, the paths that can allocate will do so about
> > one time of 500, so any issues should be spotted sooner rather than later.
> >
> > So you are quite right to be concerned, but I believe we will be doing the
> > right things.  And based on his previous track record, I am also quite
> > certain that Mr. Murphy will be on hand to provide me any additional
> > education that I might require.
> >
> > Finally, I have noted down your point about fs_reclaim_acquire() and
> > fs_reclaim_release().  Whether or not they prove to be needed, I do
> > appreciate your calling them to my attention.
> 
> I just realized that since these dma_fence structs are refcounted and
> userspace can hold references (directly, it can pass them around
> behind file descriptors) we might never hit such a path until slightly
> unusual or evil userspace does something interesting. Do you have
> links to those patches? Some googling didn't turn up anything. I can
> then figure out whether it's better to risk not spotting issues with
> call_rcu vs slapping a memalloc_noio_save/restore around all these
> critical section which force-degrades any allocation to GFP_ATOMIC at
> most, but has the risk that we run into code that assumes "GFP_KERNEL
> never fails for small stuff" and has a decidedly less tested fallback
> path than rcu code.

Here is the previous early draft version, which will change considerably
for the next version:

	lore.kernel.org/lkml/20200809204354.20137-1-urezki@gmail.com

This does kvfree_rcu(), but we expect to handle call_rcu() similarly.

The version in preparation will use workqueues to do the allocation in a
known-safe environment and also use lockless access to certain portions
of the allocator caches (as noted earlier, this last is not much loved
by some of the MM guys).  Given Thomas's patch, we could with high
probability allocate directly, perhaps even not needing memory-allocator
modifications.

Either way, kvfree_rcu(), and later call_rcu(), will avoid asking the
allocator to do anything that the calling context prohibits.  So what
types of bugs are you looking for?  Where reclaim calls back into the
driver or some such?

							Thanx, Paul
Ard Biesheuvel Sept. 17, 2020, 6:38 a.m. UTC | #18
On Wed, 16 Sep 2020 at 21:32, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> But something like a driver list walking thing should not be doing
> different things behind peoples back depending on whether they hold
> spinlocks or not. It should either just work regardless, or there
> should be a flag (or special interface) for the "you're being called
> in a crtitical region".
>
> Because dynamically changing behavior really is very confusing.
>

By the same reasoning, I don't think a generic crypto library should
be playing tricks with preemption en/disabling under the hood when
iterating over some data that is all directly accessible via the
linear map on the platforms that most people care about. And using
kmap_atomic() unconditionally achieves exactly that.

As I argued before, the fact that kmap_atomic() can be called from an
atomic context, and the fact that its implementation on HIGHMEM
platforms requires preemption to be disabled until the next kunmap()
are two different things, and I don't agree with your assertion that
the name kmap_atomic() implies the latter semantics. If we can avoid
disabling preemption on HIGHMEM, as Thomas suggests, we surely don't
need it on !HIGHMEM either, and given that kmap_atomic() is preferred
today anyway, we can just merge the two implementations. Are there any
existing debug features that could help us spot [ab]use of things like
raw per-CPU data within kmap_atomic regions?

Re your point about deprecating HIGHMEM: some work is underway on ARM
to implement a 3.75/3.75 GB kernel/user split on recent LPAE capable
hardware (which shouldn't suffer from the performance issues that
plagued the 4/4 split on i686), and so hopefully, there is a path
forward for ARM that does not rely on HIGHMEM as it does today.
Daniel Vetter Sept. 17, 2020, 7:52 a.m. UTC | #19
On Thu, Sep 17, 2020 at 12:39 AM Paul E. McKenney <paulmck@kernel.org> wrote:
>
> On Wed, Sep 16, 2020 at 11:43:02PM +0200, Daniel Vetter wrote:
> > On Wed, Sep 16, 2020 at 10:58 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> > >
> > > On Wed, Sep 16, 2020 at 10:29:06PM +0200, Daniel Vetter wrote:
> > > > On Wed, Sep 16, 2020 at 5:29 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> > > > >
> > > > > On Wed, Sep 16, 2020 at 09:37:17AM +0200, Daniel Vetter wrote:
> > > > > > On Tue, Sep 15, 2020 at 7:35 PM Linus Torvalds
> > > > > > <torvalds@linux-foundation.org> wrote:
> > > > > > >
> > > > > > > On Tue, Sep 15, 2020 at 1:39 AM Thomas Gleixner <tglx@linutronix.de> wrote:
> > > > > > > >
> > > > > > > > OTOH, having a working 'preemptible()' or maybe better named
> > > > > > > > 'can_schedule()' check makes tons of sense to make decisions about
> > > > > > > > allocation modes or other things.
> > > > > > >
> > > > > > > No. I think that those kinds of decisions about actual behavior are
> > > > > > > always simply fundamentally wrong.
> > > > > > >
> > > > > > > Note that this is very different from having warnings about invalid
> > > > > > > use. THAT is correct. It may not warn in all configurations, but that
> > > > > > > doesn't matter: what matters is that it warns in common enough
> > > > > > > configurations that developers will catch it.
> > > > > > >
> > > > > > > So having a warning in "might_sleep()" that doesn't always trigger,
> > > > > > > because you have a limited configuration that can't even detect the
> > > > > > > situation, that's fine and dandy and intentional.
> > > > > > >
> > > > > > > But having code like
> > > > > > >
> > > > > > >        if (can_schedule())
> > > > > > >            .. do something different ..
> > > > > > >
> > > > > > > is fundamentally complete and utter garbage.
> > > > > > >
> > > > > > > It's one thing if you test for "am I in hardware interrupt context".
> > > > > > > Those tests aren't great either, but at least they make sense.
> > > > > > >
> > > > > > > But a driver - or some library routine - making a difference based on
> > > > > > > some nebulous "can I schedule" is fundamentally and basically WRONG.
> > > > > > >
> > > > > > > If some code changes behavior, it needs to be explicit to the *caller*
> > > > > > > of that code.
> > > > > > >
> > > > > > > So this is why GFP_ATOMIC is fine, but "if (!can_schedule())
> > > > > > > do_something_atomic()" is pure shite.
> > > > > > >
> > > > > > > And I am not IN THE LEAST interested in trying to help people doing
> > > > > > > pure shite. We need to fix them. Like the crypto code is getting
> > > > > > > fixed.
> > > > > >
> > > > > > Just figured I'll throw my +1 in from reading too many (gpu) drivers.
> > > > > > Code that tries to cleverly adjust its behaviour depending upon the
> > > > > > context it's running in is harder to understand and blows up in more
> > > > > > interesting ways. We still have drm_can_sleep() and it's mostly just
> > > > > > used for debug code, and I've largely ended up just deleting
> > > > > > everything that used it because when you're driver is blowing up the
> > > > > > last thing you want is to realize your debug code and output can't be
> > > > > > relied upon. Or worse, that the only Oops you have is the one in the
> > > > > > debug code, because the real one scrolled away - the original idea
> > > > > > behind drm_can_sleep was to make all the modeset code work
> > > > > > automagically both in normal ioctl/kworker context and in the panic
> > > > > > handlers or kgdb callbacks. Wishful thinking at best.
> > > > > >
> > > > > > Also at least for me that extends to everything, e.g. I much prefer
> > > > > > explicit spin_lock and spin_lock_irq vs magic spin_lock_irqsave for
> > > > > > locks shared with interrupt handlers, since the former two gives me
> > > > > > clear information from which contexts such function can be called.
> > > > > > Other end is the memalloc_no*_save/restore functions, where I recently
> > > > > > made a real big fool of myself because I didn't realize how much that
> > > > > > impacts everything that's run within - suddenly "GFP_KERNEL for small
> > > > > > stuff never fails" is wrong everywhere.
> > > > > >
> > > > > > It's all great for debugging and sanity checks (and we run with all
> > > > > > that stuff enabled in our CI), but really semantic changes depending
> > > > > > upon magic context checks freak my out :-)
> > > > >
> > > > > All fair, but some of us need to write code that must handle being
> > > > > invoked from a wide variety of contexts.  Now perhaps you like the idea of
> > > > > call_rcu() for schedulable contexts, call_rcu_nosched() when preemption
> > > > > is disabled, call_rcu_irqs_are_disabled() when interrupts are disabled,
> > > > > call_rcu_raw_atomic() from contexts where (for example) raw spinlocks
> > > > > are held, and so on.  However, from what I can see, most people instead
> > > > > consistently prefer that the RCU API instead be consolidated.
> > > > >
> > > > > Some in-flight cache-efficiency work for kvfree_rcu() and call_rcu()
> > > > > needs to be able to allocate memory occasionally.  It can do that when
> > > > > invoked from some contexts, but not when invoked from others.  Right now,
> > > > > in !PREEMPT kernels, it cannot tell, and must either do things to the
> > > > > memory allocators that some of the MM hate or must unnecessarily invoke
> > > > > workqueues.  Thomas's patches would allow the code to just allocate in
> > > > > the common case when these primitives are invoked from contexts where
> > > > > allocation is permitted.
> > > > >
> > > > > If we want to restrict access to the can_schedule() or whatever primitive,
> > > > > fine and good.  We can add a check to checkpatch.pl, for example.  Maybe
> > > > > we can go back to the old brlock approach of requiring certain people's
> > > > > review for each addition to the kernel.
> > > > >
> > > > > But there really are use cases that it would greatly help.
> > > >
> > > > We can deadlock in random fun places if random stuff we're calling
> > > > suddenly starts allocating. Sometimes. Maybe once in a blue moon, to
> > > > make it extra fun to reproduce. Maybe most driver subsystems are less
> > > > brittle, but gpu drivers definitely need to know about the details for
> > > > exactly this example. And yes gpu drivers use rcu for freeing
> > > > dma_fence structures, and that tends to happen in code that we only
> > > > recently figured out should really not allocate memory.
> > > >
> > > > I think minimally you need to throw in an unconditional
> > > > fs_reclaim_acquire();fs_reclaim_release(); so that everyone who runs
> > > > with full debugging knows what might happen. It's kinda like
> > > > might_sleep, but a lot more specific. might_sleep() alone is not
> > > > enough, because in the specific code paths I'm thinking of (and
> > > > created special lockdep annotations for just recently) sleeping is
> > > > allowed, but any memory allocations with GFP_RECLAIM set are no-go.
> > >
> > > Completely agreed!  Any allocation on any free path must be handled
> > > -extremely- carefully.  To that end...
> > >
> > > First, there is always a fallback in case the allocation fails.  Which
> > > might have performance or corner-case robustness issues, but which will
> > > at least allow forward progress.  Second, we consulted with a number of
> > > MM experts to arrive at appropriate GFP_* flags (and their patience is
> > > greatly appreciated).  Third, the paths that can allocate will do so about
> > > one time of 500, so any issues should be spotted sooner rather than later.
> > >
> > > So you are quite right to be concerned, but I believe we will be doing the
> > > right things.  And based on his previous track record, I am also quite
> > > certain that Mr. Murphy will be on hand to provide me any additional
> > > education that I might require.
> > >
> > > Finally, I have noted down your point about fs_reclaim_acquire() and
> > > fs_reclaim_release().  Whether or not they prove to be needed, I do
> > > appreciate your calling them to my attention.
> >
> > I just realized that since these dma_fence structs are refcounted and
> > userspace can hold references (directly, it can pass them around
> > behind file descriptors) we might never hit such a path until slightly
> > unusual or evil userspace does something interesting. Do you have
> > links to those patches? Some googling didn't turn up anything. I can
> > then figure out whether it's better to risk not spotting issues with
> > call_rcu vs slapping a memalloc_noio_save/restore around all these
> > critical section which force-degrades any allocation to GFP_ATOMIC at
> > most, but has the risk that we run into code that assumes "GFP_KERNEL
> > never fails for small stuff" and has a decidedly less tested fallback
> > path than rcu code.
>
> Here is the previous early draft version, which will change considerably
> for the next version:
>
>         lore.kernel.org/lkml/20200809204354.20137-1-urezki@gmail.com
>
> This does kvfree_rcu(), but we expect to handle call_rcu() similarly.
>
> The version in preparation will use workqueues to do the allocation in a
> known-safe environment and also use lockless access to certain portions
> of the allocator caches (as noted earlier, this last is not much loved
> by some of the MM guys).  Given Thomas's patch, we could with high
> probability allocate directly, perhaps even not needing memory-allocator
> modifications.
>
> Either way, kvfree_rcu(), and later call_rcu(), will avoid asking the
> allocator to do anything that the calling context prohibits.  So what
> types of bugs are you looking for?  Where reclaim calls back into the
> driver or some such?

Yeah pretty much. It's a problem for gpu, fs, block drivers and really
anything else that's remotely involved in memory reclaim somehow.
Generally this is all handled explicitly by passing gfp_t flags down
any call chain, but in some cases it's instead solved with the
memalloc_no* functions. E.g. sunrpc uses that to make sure the network
stack (which generally just assumes it can allocate memory) doesn't,
to avoid recursions back into nfs/sunrpc. To my knowledge there's no
way to check at runtime with which gfp flags you're allowed to
allocate memory, a preemptible check is definitely not enough.
Disabled preemption implies only GFP_ATOMIC is allowed (ignoring nmi
and stuff like that), but the inverse is not true.

So if you want the automagic in call_rcu I think either
- we need to replace all explicit gfp flags with the context marking
memalloc_no* across the entire kernel, or at least anywhere rcu might
be used.
- audit all callchains and make sure a call_rcu_noalloc is used
anywhere there might be a problem. probably better to have a
call_rcu_gfp with explicit gfp flags parameter, since generally that
needs to be passed down.

But at least to me the lockless magic in mm sounds a lot safer, since
it contains the complexity and doesn't leak it out to callers of
call_rcu.
-Daniel
Paul E. McKenney Sept. 17, 2020, 4:28 p.m. UTC | #20
On Thu, Sep 17, 2020 at 09:52:30AM +0200, Daniel Vetter wrote:
> On Thu, Sep 17, 2020 at 12:39 AM Paul E. McKenney <paulmck@kernel.org> wrote:
> >
> > On Wed, Sep 16, 2020 at 11:43:02PM +0200, Daniel Vetter wrote:
> > > On Wed, Sep 16, 2020 at 10:58 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> > > >
> > > > On Wed, Sep 16, 2020 at 10:29:06PM +0200, Daniel Vetter wrote:
> > > > > On Wed, Sep 16, 2020 at 5:29 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> > > > > >
> > > > > > On Wed, Sep 16, 2020 at 09:37:17AM +0200, Daniel Vetter wrote:
> > > > > > > On Tue, Sep 15, 2020 at 7:35 PM Linus Torvalds
> > > > > > > <torvalds@linux-foundation.org> wrote:
> > > > > > > >
> > > > > > > > On Tue, Sep 15, 2020 at 1:39 AM Thomas Gleixner <tglx@linutronix.de> wrote:
> > > > > > > > >
> > > > > > > > > OTOH, having a working 'preemptible()' or maybe better named
> > > > > > > > > 'can_schedule()' check makes tons of sense to make decisions about
> > > > > > > > > allocation modes or other things.
> > > > > > > >
> > > > > > > > No. I think that those kinds of decisions about actual behavior are
> > > > > > > > always simply fundamentally wrong.
> > > > > > > >
> > > > > > > > Note that this is very different from having warnings about invalid
> > > > > > > > use. THAT is correct. It may not warn in all configurations, but that
> > > > > > > > doesn't matter: what matters is that it warns in common enough
> > > > > > > > configurations that developers will catch it.
> > > > > > > >
> > > > > > > > So having a warning in "might_sleep()" that doesn't always trigger,
> > > > > > > > because you have a limited configuration that can't even detect the
> > > > > > > > situation, that's fine and dandy and intentional.
> > > > > > > >
> > > > > > > > But having code like
> > > > > > > >
> > > > > > > >        if (can_schedule())
> > > > > > > >            .. do something different ..
> > > > > > > >
> > > > > > > > is fundamentally complete and utter garbage.
> > > > > > > >
> > > > > > > > It's one thing if you test for "am I in hardware interrupt context".
> > > > > > > > Those tests aren't great either, but at least they make sense.
> > > > > > > >
> > > > > > > > But a driver - or some library routine - making a difference based on
> > > > > > > > some nebulous "can I schedule" is fundamentally and basically WRONG.
> > > > > > > >
> > > > > > > > If some code changes behavior, it needs to be explicit to the *caller*
> > > > > > > > of that code.
> > > > > > > >
> > > > > > > > So this is why GFP_ATOMIC is fine, but "if (!can_schedule())
> > > > > > > > do_something_atomic()" is pure shite.
> > > > > > > >
> > > > > > > > And I am not IN THE LEAST interested in trying to help people doing
> > > > > > > > pure shite. We need to fix them. Like the crypto code is getting
> > > > > > > > fixed.
> > > > > > >
> > > > > > > Just figured I'll throw my +1 in from reading too many (gpu) drivers.
> > > > > > > Code that tries to cleverly adjust its behaviour depending upon the
> > > > > > > context it's running in is harder to understand and blows up in more
> > > > > > > interesting ways. We still have drm_can_sleep() and it's mostly just
> > > > > > > used for debug code, and I've largely ended up just deleting
> > > > > > > everything that used it because when you're driver is blowing up the
> > > > > > > last thing you want is to realize your debug code and output can't be
> > > > > > > relied upon. Or worse, that the only Oops you have is the one in the
> > > > > > > debug code, because the real one scrolled away - the original idea
> > > > > > > behind drm_can_sleep was to make all the modeset code work
> > > > > > > automagically both in normal ioctl/kworker context and in the panic
> > > > > > > handlers or kgdb callbacks. Wishful thinking at best.
> > > > > > >
> > > > > > > Also at least for me that extends to everything, e.g. I much prefer
> > > > > > > explicit spin_lock and spin_lock_irq vs magic spin_lock_irqsave for
> > > > > > > locks shared with interrupt handlers, since the former two gives me
> > > > > > > clear information from which contexts such function can be called.
> > > > > > > Other end is the memalloc_no*_save/restore functions, where I recently
> > > > > > > made a real big fool of myself because I didn't realize how much that
> > > > > > > impacts everything that's run within - suddenly "GFP_KERNEL for small
> > > > > > > stuff never fails" is wrong everywhere.
> > > > > > >
> > > > > > > It's all great for debugging and sanity checks (and we run with all
> > > > > > > that stuff enabled in our CI), but really semantic changes depending
> > > > > > > upon magic context checks freak my out :-)
> > > > > >
> > > > > > All fair, but some of us need to write code that must handle being
> > > > > > invoked from a wide variety of contexts.  Now perhaps you like the idea of
> > > > > > call_rcu() for schedulable contexts, call_rcu_nosched() when preemption
> > > > > > is disabled, call_rcu_irqs_are_disabled() when interrupts are disabled,
> > > > > > call_rcu_raw_atomic() from contexts where (for example) raw spinlocks
> > > > > > are held, and so on.  However, from what I can see, most people instead
> > > > > > consistently prefer that the RCU API instead be consolidated.
> > > > > >
> > > > > > Some in-flight cache-efficiency work for kvfree_rcu() and call_rcu()
> > > > > > needs to be able to allocate memory occasionally.  It can do that when
> > > > > > invoked from some contexts, but not when invoked from others.  Right now,
> > > > > > in !PREEMPT kernels, it cannot tell, and must either do things to the
> > > > > > memory allocators that some of the MM hate or must unnecessarily invoke
> > > > > > workqueues.  Thomas's patches would allow the code to just allocate in
> > > > > > the common case when these primitives are invoked from contexts where
> > > > > > allocation is permitted.
> > > > > >
> > > > > > If we want to restrict access to the can_schedule() or whatever primitive,
> > > > > > fine and good.  We can add a check to checkpatch.pl, for example.  Maybe
> > > > > > we can go back to the old brlock approach of requiring certain people's
> > > > > > review for each addition to the kernel.
> > > > > >
> > > > > > But there really are use cases that it would greatly help.
> > > > >
> > > > > We can deadlock in random fun places if random stuff we're calling
> > > > > suddenly starts allocating. Sometimes. Maybe once in a blue moon, to
> > > > > make it extra fun to reproduce. Maybe most driver subsystems are less
> > > > > brittle, but gpu drivers definitely need to know about the details for
> > > > > exactly this example. And yes gpu drivers use rcu for freeing
> > > > > dma_fence structures, and that tends to happen in code that we only
> > > > > recently figured out should really not allocate memory.
> > > > >
> > > > > I think minimally you need to throw in an unconditional
> > > > > fs_reclaim_acquire();fs_reclaim_release(); so that everyone who runs
> > > > > with full debugging knows what might happen. It's kinda like
> > > > > might_sleep, but a lot more specific. might_sleep() alone is not
> > > > > enough, because in the specific code paths I'm thinking of (and
> > > > > created special lockdep annotations for just recently) sleeping is
> > > > > allowed, but any memory allocations with GFP_RECLAIM set are no-go.
> > > >
> > > > Completely agreed!  Any allocation on any free path must be handled
> > > > -extremely- carefully.  To that end...
> > > >
> > > > First, there is always a fallback in case the allocation fails.  Which
> > > > might have performance or corner-case robustness issues, but which will
> > > > at least allow forward progress.  Second, we consulted with a number of
> > > > MM experts to arrive at appropriate GFP_* flags (and their patience is
> > > > greatly appreciated).  Third, the paths that can allocate will do so about
> > > > one time of 500, so any issues should be spotted sooner rather than later.
> > > >
> > > > So you are quite right to be concerned, but I believe we will be doing the
> > > > right things.  And based on his previous track record, I am also quite
> > > > certain that Mr. Murphy will be on hand to provide me any additional
> > > > education that I might require.
> > > >
> > > > Finally, I have noted down your point about fs_reclaim_acquire() and
> > > > fs_reclaim_release().  Whether or not they prove to be needed, I do
> > > > appreciate your calling them to my attention.
> > >
> > > I just realized that since these dma_fence structs are refcounted and
> > > userspace can hold references (directly, it can pass them around
> > > behind file descriptors) we might never hit such a path until slightly
> > > unusual or evil userspace does something interesting. Do you have
> > > links to those patches? Some googling didn't turn up anything. I can
> > > then figure out whether it's better to risk not spotting issues with
> > > call_rcu vs slapping a memalloc_noio_save/restore around all these
> > > critical section which force-degrades any allocation to GFP_ATOMIC at
> > > most, but has the risk that we run into code that assumes "GFP_KERNEL
> > > never fails for small stuff" and has a decidedly less tested fallback
> > > path than rcu code.
> >
> > Here is the previous early draft version, which will change considerably
> > for the next version:
> >
> >         lore.kernel.org/lkml/20200809204354.20137-1-urezki@gmail.com
> >
> > This does kvfree_rcu(), but we expect to handle call_rcu() similarly.
> >
> > The version in preparation will use workqueues to do the allocation in a
> > known-safe environment and also use lockless access to certain portions
> > of the allocator caches (as noted earlier, this last is not much loved
> > by some of the MM guys).  Given Thomas's patch, we could with high
> > probability allocate directly, perhaps even not needing memory-allocator
> > modifications.
> >
> > Either way, kvfree_rcu(), and later call_rcu(), will avoid asking the
> > allocator to do anything that the calling context prohibits.  So what
> > types of bugs are you looking for?  Where reclaim calls back into the
> > driver or some such?
> 
> Yeah pretty much. It's a problem for gpu, fs, block drivers and really
> anything else that's remotely involved in memory reclaim somehow.
> Generally this is all handled explicitly by passing gfp_t flags down
> any call chain, but in some cases it's instead solved with the
> memalloc_no* functions. E.g. sunrpc uses that to make sure the network
> stack (which generally just assumes it can allocate memory) doesn't,
> to avoid recursions back into nfs/sunrpc. To my knowledge there's no
> way to check at runtime with which gfp flags you're allowed to
> allocate memory, a preemptible check is definitely not enough.
> Disabled preemption implies only GFP_ATOMIC is allowed (ignoring nmi
> and stuff like that), but the inverse is not true.

Thank you for the confirmation!

> So if you want the automagic in call_rcu I think either
> - we need to replace all explicit gfp flags with the context marking
> memalloc_no* across the entire kernel, or at least anywhere rcu might
> be used.
> - audit all callchains and make sure a call_rcu_noalloc is used
> anywhere there might be a problem. probably better to have a
> call_rcu_gfp with explicit gfp flags parameter, since generally that
> needs to be passed down.
> 
> But at least to me the lockless magic in mm sounds a lot safer, since
> it contains the complexity and doesn't leak it out to callers of
> call_rcu.

Agreed, I greatly prefer Peter Zijlstra's lockless-allocation patch
myself.

In the meantime, it looks like we will start by causing the allocation to
happen in a safe environment.  That may have issues with delays, but is
at least something that can be done entirely within the confines of RCU.

							Thanx, Paul
Michal Hocko Sept. 29, 2020, 8:19 a.m. UTC | #21
On Wed 16-09-20 23:43:02, Daniel Vetter wrote:
> I can
> then figure out whether it's better to risk not spotting issues with
> call_rcu vs slapping a memalloc_noio_save/restore around all these
> critical section which force-degrades any allocation to GFP_ATOMIC at

did you mean memalloc_noreclaim_* here?

> most, but has the risk that we run into code that assumes "GFP_KERNEL
> never fails for small stuff" and has a decidedly less tested fallback
> path than rcu code.

Even if the above then please note that memalloc_noreclaim_* or
PF_MEMALLOC should be used with an extreme care. Essentially only for
internal memory reclaimers. It grants access to _all_ the available
memory so any abuse can be detrimental to the overall system operation.
Allocation failure in this mode means that we are out of memory and any
code relying on such an allocation has to carefuly consider failure.
This is not a random allocation mode.
Daniel Vetter Sept. 29, 2020, 9 a.m. UTC | #22
On Tue, Sep 29, 2020 at 10:19:38AM +0200, Michal Hocko wrote:
> On Wed 16-09-20 23:43:02, Daniel Vetter wrote:
> > I can
> > then figure out whether it's better to risk not spotting issues with
> > call_rcu vs slapping a memalloc_noio_save/restore around all these
> > critical section which force-degrades any allocation to GFP_ATOMIC at
> 
> did you mean memalloc_noreclaim_* here?

Yeah I picked the wrong one of that family of functions.

> > most, but has the risk that we run into code that assumes "GFP_KERNEL
> > never fails for small stuff" and has a decidedly less tested fallback
> > path than rcu code.
> 
> Even if the above then please note that memalloc_noreclaim_* or
> PF_MEMALLOC should be used with an extreme care. Essentially only for
> internal memory reclaimers. It grants access to _all_ the available
> memory so any abuse can be detrimental to the overall system operation.
> Allocation failure in this mode means that we are out of memory and any
> code relying on such an allocation has to carefuly consider failure.
> This is not a random allocation mode.

Agreed, that's why I don't like having these kind of automagic critical
sections. It's a bit a shotgun approach. Paul said that the code would
handle failures, but the problem is that it applies everywhere.

Anyway my understanding is that call_rcu will be reworked and gain a pile
of tricks so that these problems for the callchains leading to call_rcu
all disappear.
-Daniel
Michal Hocko Sept. 29, 2020, 2:54 p.m. UTC | #23
On Tue 29-09-20 11:00:03, Daniel Vetter wrote:
> On Tue, Sep 29, 2020 at 10:19:38AM +0200, Michal Hocko wrote:
> > On Wed 16-09-20 23:43:02, Daniel Vetter wrote:
> > > I can
> > > then figure out whether it's better to risk not spotting issues with
> > > call_rcu vs slapping a memalloc_noio_save/restore around all these
> > > critical section which force-degrades any allocation to GFP_ATOMIC at
> > 
> > did you mean memalloc_noreclaim_* here?
> 
> Yeah I picked the wrong one of that family of functions.
> 
> > > most, but has the risk that we run into code that assumes "GFP_KERNEL
> > > never fails for small stuff" and has a decidedly less tested fallback
> > > path than rcu code.
> > 
> > Even if the above then please note that memalloc_noreclaim_* or
> > PF_MEMALLOC should be used with an extreme care. Essentially only for
> > internal memory reclaimers. It grants access to _all_ the available
> > memory so any abuse can be detrimental to the overall system operation.
> > Allocation failure in this mode means that we are out of memory and any
> > code relying on such an allocation has to carefuly consider failure.
> > This is not a random allocation mode.
> 
> Agreed, that's why I don't like having these kind of automagic critical
> sections. It's a bit a shotgun approach. Paul said that the code would
> handle failures, but the problem is that it applies everywhere.

Ohh, in the ideal world we wouldn't need anything like that. But then
the reality fires:
* PF_MEMALLOC (resp memalloc_noreclaim_* for that matter) is primarily used
  to make sure that allocations from inside the memory reclaim - yeah that
  happens - will not recurse.
* PF_MEMALLOC_NO{FS,IO} (resp memalloc_no{fs,io}*) are used to mark no
  fs/io reclaim recursion critical sections because controling that for
  each allocation inside fs transaction (or other sensitive) or IO
  contexts turned out to be unmaintainable and people simply fallen into
  using NOFS/NOIO unconditionally which is causing reclaim imbalance
  problems.
* PF_MEMALLOC_NOCMA (resp memalloc_nocma*) is used for long term pinning
  when CMA pages cannot be pinned because that would break the CMA
  guarantees. Communicating this to all potential allocations during
  pinning is simply unfeasible.

So you are absolutely right that these critical sections with side
effects on all allocations are far from ideal from the API point of view
but they are mostly mirroring a demand for functionality which is
_practically_ impossible to achieve with our current code base. Not that
we couldn't get back to drawing board and come up with a saner thing and
rework the world...