Re: [patch 06/13] locking/bitspinlock: Clenaup PREEMPT_COUNT leftovers

Re: [patch 06/13] locking/bitspinlock: Clenaup PREEMPT_COUNT leftovers

[Will Deacon]

On Mon, Sep 14, 2020 at 10:42:15PM +0200, Thomas Gleixner wrote:
> CONFIG_PREEMPT_COUNT is now unconditionally enabled and will be
> removed. Cleanup the leftovers before doing so.
> 
> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
> ---
>  include/linux/bit_spinlock.h |    4 +---
>  1 file changed, 1 insertion(+), 3 deletions(-)
> 
> --- a/include/linux/bit_spinlock.h
> +++ b/include/linux/bit_spinlock.h
> @@ -90,10 +90,8 @@ static inline int bit_spin_is_locked(int
>  {
>  #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
>  	return test_bit(bitnum, addr);
> -#elif defined CONFIG_PREEMPT_COUNT
> -	return preempt_count();
>  #else
> -	return 1;
> +	return preempt_count();
>  #endif

Acked-by: Will Deacon <will@kernel.org>

Will

Re: [patch 04/13] lockdep: Clenaup PREEMPT_COUNT leftovers

[Will Deacon]

On Mon, Sep 14, 2020 at 10:42:13PM +0200, Thomas Gleixner wrote:
> CONFIG_PREEMPT_COUNT is now unconditionally enabled and will be
> removed. Cleanup the leftovers before doing so.
> 
> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Ingo Molnar <mingo@kernel.org>
> Cc: Will Deacon <will@kernel.org>
> ---
>  include/linux/lockdep.h |    6 ++----
>  lib/Kconfig.debug       |    1 -
>  2 files changed, 2 insertions(+), 5 deletions(-)
> 
> --- a/include/linux/lockdep.h
> +++ b/include/linux/lockdep.h
> @@ -585,16 +585,14 @@ do {									\
>  
>  #define lockdep_assert_preemption_enabled()				\
>  do {									\
> -	WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_COUNT)	&&		\
> -		     debug_locks			&&		\
> +	WARN_ON_ONCE(debug_locks			&&		\
>  		     (preempt_count() != 0		||		\
>  		      !raw_cpu_read(hardirqs_enabled)));		\
>  } while (0)
>  
>  #define lockdep_assert_preemption_disabled()				\
>  do {									\
> -	WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_COUNT)	&&		\
> -		     debug_locks			&&		\
> +	WARN_ON_ONCE(debug_locks			&&		\
>  		     (preempt_count() == 0		&&		\
>  		      raw_cpu_read(hardirqs_enabled)));			\
>  } while (0)
> --- a/lib/Kconfig.debug
> +++ b/lib/Kconfig.debug
> @@ -1161,7 +1161,6 @@ config PROVE_LOCKING
>  	select DEBUG_RWSEMS
>  	select DEBUG_WW_MUTEX_SLOWPATH
>  	select DEBUG_LOCK_ALLOC
> -	select PREEMPT_COUNT
>  	select TRACE_IRQFLAGS
>  	default n
>  	help

Acked-by: Will Deacon <will@kernel.org>

Will

Re: [patch 00/13] preempt: Make preempt count unconditional

[Paul E. McKenney]

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

[PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Herbert Xu]

I trimmed the cc as the mailing lists appear to be blocking this
email because of it.

On Mon, Sep 14, 2020 at 03:37:49PM -0700, Linus Torvalds wrote:
>
> So it _looks_ like this code started using kmap() - probably back when
> kmap_atomic() was so cumbersome to use - and was then converted
> (conditionally) to kmap_atomic() rather than just changed whole-sale.
> Is there actually something that wants to use those sg_miter functions
> and sleep?

I dug up the old zinc patch submissions and this wasn't present at
all in the original.  The original zinc code used blkcipher_walk
which unconditinoally does kmap_atomic.

So it's only the SG miter conversion that introduced this change,
which appears to be a simple oversight (I think Ard was working on
latency issues at that time, perhaps he was worried about keeping
preemption off unnecessarily).

---8<---
There is no reason for the chacha20poly1305 SG miter code to use
kmap instead of kmap_atomic as the critical section doesn't sleep
anyway.  So we can simply get rid of the preemptible check and
set SG_MITER_ATOMIC unconditionally.

Even if we need to reenable preemption to lower latency we should
be doing that by interrupting the SG miter walk rather than using
kmap.

Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

diff --git a/lib/crypto/chacha20poly1305.c b/lib/crypto/chacha20poly1305.c
index 431e04280332..5850f3b87359 100644
--- a/lib/crypto/chacha20poly1305.c
+++ b/lib/crypto/chacha20poly1305.c
@@ -251,9 +251,7 @@ bool chacha20poly1305_crypt_sg_inplace(struct scatterlist *src,
 			poly1305_update(&poly1305_state, pad0, 0x10 - (ad_len & 0xf));
 	}
 
-	flags = SG_MITER_TO_SG;
-	if (!preemptible())
-		flags |= SG_MITER_ATOMIC;
+	flags = SG_MITER_TO_SG | SG_MITER_ATOMIC;
 
 	sg_miter_start(&miter, src, sg_nents(src), flags);
 
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Ard Biesheuvel]

(+ Jason)

On Tue, 15 Sep 2020 at 06:30, Herbert Xu <herbert@gondor.apana.org.au> wrote:
>
> I trimmed the cc as the mailing lists appear to be blocking this
> email because of it.
>
> On Mon, Sep 14, 2020 at 03:37:49PM -0700, Linus Torvalds wrote:
> >
> > So it _looks_ like this code started using kmap() - probably back when
> > kmap_atomic() was so cumbersome to use - and was then converted
> > (conditionally) to kmap_atomic() rather than just changed whole-sale.
> > Is there actually something that wants to use those sg_miter functions
> > and sleep?
>
> I dug up the old zinc patch submissions and this wasn't present at
> all in the original.  The original zinc code used blkcipher_walk
> which unconditinoally does kmap_atomic.
>

Remember that the Zinc patchset was very vocal about not relying on
the Linux crypto API, yet it [ab]used the crypto blkcipher_walk API
(which was already deprecated at that point) in a rather horrid way,
by going around the blkcipher API itself, and creating some mock
objects that the blkcipher scatterlist walker would expect to exist.

So instead, I opted to rewrite this code using the SG miter API so that:
- src == dst, and so we only need to traverse (and kmap) a single
scatterlist instead of two in parallel (as Wireguard has no need for
the latter)
- no elaborate handling of the scatterlist elements when they are not
a multiple of the cipher chunk size (which is not needed for a stream
cipher liker ChaCha)
- no need to use scatterwalk_map_and_copy() (and do another kmap()) to
access the tag if it was covered by the last scatterlist element.

> So it's only the SG miter conversion that introduced this change,
> which appears to be a simple oversight (I think Ard was working on
> latency issues at that time, perhaps he was worried about keeping
> preemption off unnecessarily).
>

No, the problem with using kmap_atomic() is that it disables
preemption even on !HIGHMEM architectures. So using it unconditionally
here means that all chacha/poly processing will execute with
preemption disabled on 64-bit architectures as well.

This means that, even if you avoid the SIMD accelerated ciphers for
latency reasons (as they disable preemption as well), you are still
running the bulk of the WireGuard processing with preemption disabled.

> ---8<---
> There is no reason for the chacha20poly1305 SG miter code to use
> kmap instead of kmap_atomic as the critical section doesn't sleep
> anyway.  So we can simply get rid of the preemptible check and
> set SG_MITER_ATOMIC unconditionally.
>
> Even if we need to reenable preemption to lower latency we should
> be doing that by interrupting the SG miter walk rather than using
> kmap.
>

AIUI, the common case is that the entire packet is covered by a single
scatterlist element, so there is no room for latency reduction here.

The problem is really that kmap_atomic() is not simply a no-op on
!HIGHMEM architectures. If we can fix that, I have no objections to
this patch.

> Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
> Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
>
> diff --git a/lib/crypto/chacha20poly1305.c b/lib/crypto/chacha20poly1305.c
> index 431e04280332..5850f3b87359 100644
> --- a/lib/crypto/chacha20poly1305.c
> +++ b/lib/crypto/chacha20poly1305.c
> @@ -251,9 +251,7 @@ bool chacha20poly1305_crypt_sg_inplace(struct scatterlist *src,
>                         poly1305_update(&poly1305_state, pad0, 0x10 - (ad_len & 0xf));
>         }
>
> -       flags = SG_MITER_TO_SG;
> -       if (!preemptible())
> -               flags |= SG_MITER_ATOMIC;
> +       flags = SG_MITER_TO_SG | SG_MITER_ATOMIC;
>
>         sg_miter_start(&miter, src, sg_nents(src), flags);
>
> --
> Email: Herbert Xu <herbert@gondor.apana.org.au>
> Home Page: http://gondor.apana.org.au/~herbert/
> PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Herbert Xu]

On Tue, Sep 15, 2020 at 09:03:46AM +0300, Ard Biesheuvel wrote:
>
> The problem is really that kmap_atomic() is not simply a no-op on
> !HIGHMEM architectures. If we can fix that, I have no objections to
> this patch.

Yes we should definitely fix that.  However, doing so will involve
manually checking every instance of kmap_atomic.

Cheers,
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Linus Torvalds]

On Mon, Sep 14, 2020 at 8:30 PM Herbert Xu <herbert@gondor.apana.org.au> wrote:
>
> There is no reason for the chacha20poly1305 SG miter code to use
> kmap instead of kmap_atomic as the critical section doesn't sleep
> anyway.  So we can simply get rid of the preemptible check and
> set SG_MITER_ATOMIC unconditionally.

So I'd prefer to make SG_MITER_ATOMIC go away entirely, and just
remove the non-atomic case..

A quick grep seems to imply that just about all users set the ATOMIC
bit anyway. I didn't look at everything, but every case I _did_ look
at did seem to set the ATOMIC bit.

So it really did seem like there isn't a lot of reason to have the
non-atomic case, and this flag could go away - not by virtue of the
atomic case going away, but by virtue of the atomic case being the
only actual case.

I mean, I did find one case that didn't set it (cb710-mmc.c), but
pattern-matching to the other mmc cases, that one looks like it
_should_ have set the atomic flag like everybody else did.

Did I miss something?

             Linus

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Linus Torvalds]

On Mon, Sep 14, 2020 at 11:45 PM Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> I mean, I did find one case that didn't set it (cb710-mmc.c), but
> pattern-matching to the other mmc cases, that one looks like it
> _should_ have set the atomic flag like everybody else did.

Oh, and immediately after sending that out I notice
nvmet_bdev_execute_rw(), which does seem to make allocations inside
that sg_miter loop.

So those non-atomic cases do clearly exist.

It does make the case for why kmap_atomic() wants to have the
debugging code. It will "just work" on 64-bit to do it wrong, because
the address doesn't become invalid just because you sleep or get
rescheduled. But then the code that every developer tests (since
developers tend to run on good hardware) might be completely broken on
bad old hardware.

Maybe we could hide it behind a debug option, at least.

Or, alterantively, introduce a new "debug_preempt_count" that doesn't
actually disable preemption, but warns about actual sleeping
operations..

             Linus

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Herbert Xu]

On Mon, Sep 14, 2020 at 11:55:53PM -0700, Linus Torvalds wrote:
>
> Maybe we could hide it behind a debug option, at least.
> 
> Or, alterantively, introduce a new "debug_preempt_count" that doesn't
> actually disable preemption, but warns about actual sleeping
> operations..

I'm more worried about existing users of kmap_atomic relying on
the preemption disabling semantics.  Short of someone checking
on every single instance (and that would include derived cases
such as all users of sg miter), I think the safer option is to
create something brand new and then migrate the existing users
to it.  Something like

static inline void *kmap_atomic_ifhigh(struct page *page)
{
	if (PageHighMem(page))
		return kmap_atomic(page);
	return page_address(page);
}

static inline void kunmap_atomic_ifhigh(struct page *page, void *addr)
{
	if (PageHighMem(page))
		kunmap_atomic(addr);
}

Cheers,
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Ard Biesheuvel]

On Tue, 15 Sep 2020 at 10:05, Herbert Xu <herbert@gondor.apana.org.au> wrote:
>
> On Mon, Sep 14, 2020 at 11:55:53PM -0700, Linus Torvalds wrote:
> >
> > Maybe we could hide it behind a debug option, at least.
> >
> > Or, alterantively, introduce a new "debug_preempt_count" that doesn't
> > actually disable preemption, but warns about actual sleeping
> > operations..
>
> I'm more worried about existing users of kmap_atomic relying on
> the preemption disabling semantics.  Short of someone checking
> on every single instance (and that would include derived cases
> such as all users of sg miter), I think the safer option is to
> create something brand new and then migrate the existing users
> to it.  Something like
>
> static inline void *kmap_atomic_ifhigh(struct page *page)
> {
>         if (PageHighMem(page))
>                 return kmap_atomic(page);
>         return page_address(page);
> }
>
> static inline void kunmap_atomic_ifhigh(struct page *page, void *addr)
> {
>         if (PageHighMem(page))
>                 kunmap_atomic(addr);
> }
>

But we would still need to check all users of SG miter before we could
move it to this interface.

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Thomas Gleixner]

On Tue, Sep 15 2020 at 17:05, Herbert Xu wrote:
> On Mon, Sep 14, 2020 at 11:55:53PM -0700, Linus Torvalds wrote:
>>
>> Maybe we could hide it behind a debug option, at least.
>> 
>> Or, alterantively, introduce a new "debug_preempt_count" that doesn't
>> actually disable preemption, but warns about actual sleeping
>> operations..
>
> I'm more worried about existing users of kmap_atomic relying on
> the preemption disabling semantics.  Short of someone checking
> on every single instance (and that would include derived cases
> such as all users of sg miter), I think the safer option is to
> create something brand new and then migrate the existing users
> to it.  Something like
>
> static inline void *kmap_atomic_ifhigh(struct page *page)
> {
> 	if (PageHighMem(page))
> 		return kmap_atomic(page);
> 	return page_address(page);
> }
>
> static inline void kunmap_atomic_ifhigh(struct page *page, void *addr)
> {
> 	if (PageHighMem(page))
> 		kunmap_atomic(addr);
> }

Hmm, that still has the issue that the code between map and unmap must
not sleep and the conversion must carefully check whether anything in
this region relies on preemption being disabled by kmap_atomic()
regardless of highmem or not.

kmap_atomic() is at least consistent vs. preemption, the above not so
much.

I'd rather go for a preemptible/sleepable version of highmem mapping
which is in itself consistent for both highmen and not highmem.

Thanks,

        tglx

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Ard Biesheuvel]

On Tue, 15 Sep 2020 at 12:34, Thomas Gleixner <tglx@linutronix.de> wrote:
>
> On Tue, Sep 15 2020 at 17:05, Herbert Xu wrote:
> > On Mon, Sep 14, 2020 at 11:55:53PM -0700, Linus Torvalds wrote:
> >>
> >> Maybe we could hide it behind a debug option, at least.
> >>
> >> Or, alterantively, introduce a new "debug_preempt_count" that doesn't
> >> actually disable preemption, but warns about actual sleeping
> >> operations..
> >
> > I'm more worried about existing users of kmap_atomic relying on
> > the preemption disabling semantics.  Short of someone checking
> > on every single instance (and that would include derived cases
> > such as all users of sg miter), I think the safer option is to
> > create something brand new and then migrate the existing users
> > to it.  Something like
> >
> > static inline void *kmap_atomic_ifhigh(struct page *page)
> > {
> >       if (PageHighMem(page))
> >               return kmap_atomic(page);
> >       return page_address(page);
> > }
> >
> > static inline void kunmap_atomic_ifhigh(struct page *page, void *addr)
> > {
> >       if (PageHighMem(page))
> >               kunmap_atomic(addr);
> > }
>
> Hmm, that still has the issue that the code between map and unmap must
> not sleep and the conversion must carefully check whether anything in
> this region relies on preemption being disabled by kmap_atomic()
> regardless of highmem or not.
>
> kmap_atomic() is at least consistent vs. preemption, the above not so
> much.
>

But that is really the point. I don't *want* to be forced to disable
preemption in brand new code simply because some legacy highmem API
conflates being callable from atomic context with instantiating an
atomic context by disabling preemption for no good reason. IIUC, in
the past, you would really only call kmap_atomic() if you absolutely
had to, and so you would never rely on the preemption disabling
semantics accidentally. By making kmap_atomic() the preferred API even
for calls from non-atomic contexts, this line has blurred and we no
longer know why individual kmap_atomic() occurrences exist in the
first place.

> I'd rather go for a preemptible/sleepable version of highmem mapping
> which is in itself consistent for both highmen and not highmem.
>

I don't think we need to obsess about highmem, although we should
obviously take care not to regress its performance unnecessarily. What
I want to avoid is to burden a brand new subsystem with legacy highmem
baggage simply because we could not agree on how to avoid that.

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Herbert Xu]

On Tue, Sep 15, 2020 at 01:02:10PM +0300, Ard Biesheuvel wrote:
>
> > I'd rather go for a preemptible/sleepable version of highmem mapping
> > which is in itself consistent for both highmen and not highmem.
> 
> I don't think we need to obsess about highmem, although we should
> obviously take care not to regress its performance unnecessarily. What
> I want to avoid is to burden a brand new subsystem with legacy highmem
> baggage simply because we could not agree on how to avoid that.

I think what Thomas is proposing should address your concerns Ard.
As long as nobody objects to the slight performance degradation on
legacy highmem platforms it should make kmap_atomic just go away on
modern platforms.

Cheers,
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Ard Biesheuvel]

On Tue, 15 Sep 2020 at 13:05, Herbert Xu <herbert@gondor.apana.org.au> wrote:
>
> On Tue, Sep 15, 2020 at 01:02:10PM +0300, Ard Biesheuvel wrote:
> >
> > > I'd rather go for a preemptible/sleepable version of highmem mapping
> > > which is in itself consistent for both highmen and not highmem.
> >
> > I don't think we need to obsess about highmem, although we should
> > obviously take care not to regress its performance unnecessarily. What
> > I want to avoid is to burden a brand new subsystem with legacy highmem
> > baggage simply because we could not agree on how to avoid that.
>
> I think what Thomas is proposing should address your concerns Ard.
> As long as nobody objects to the slight performance degradation on
> legacy highmem platforms it should make kmap_atomic just go away on
> modern platforms.
>

But making atomic kmap preemptible/sleepable creates the exact same
problem, i.e., that we have no idea which existing callers are
currently relying on those preemption disabling semantics, so we can't
just take them away. Or am I missing something?

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Herbert Xu]

On Tue, Sep 15, 2020 at 01:08:31PM +0300, Ard Biesheuvel wrote:
>
> But making atomic kmap preemptible/sleepable creates the exact same
> problem, i.e., that we have no idea which existing callers are
> currently relying on those preemption disabling semantics, so we can't
> just take them away. Or am I missing something?

Good point.

Thomas mentioned that RT has been doing this for a while now so
perhaps someone has studied this problem already? Thomas?

Cheers,
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Thomas Gleixner]

On Tue, Sep 15 2020 at 20:10, Herbert Xu wrote:
> On Tue, Sep 15, 2020 at 01:08:31PM +0300, Ard Biesheuvel wrote:
>>
>> But making atomic kmap preemptible/sleepable creates the exact same
>> problem, i.e., that we have no idea which existing callers are
>> currently relying on those preemption disabling semantics, so we can't
>> just take them away. Or am I missing something?
>
> Good point.
>
> Thomas mentioned that RT has been doing this for a while now so
> perhaps someone has studied this problem already? Thomas?

RT is substituting preempt_disable() with migrate_disable() which pins
the task on the CPU so that per CPU stuff still works. And we did quite
some staring whether there is code which purely relies on the
preempt_disable() to prevent reentrancy, but there is almost none.

Though we don't have migrate disable on !RT and PeterZ is not a great
fan of making it available as it wreckages schedulability - though IMO
not much more than preempt disable :)

Thanks,

        tglx

Re: [PATCH] crypto: lib/chacha20poly1305 - Set SG_MITER_ATOMIC unconditionally

[Ard Biesheuvel]

On Tue, 15 Sep 2020 at 09:56, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> On Mon, Sep 14, 2020 at 11:45 PM Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
> >
> > I mean, I did find one case that didn't set it (cb710-mmc.c), but
> > pattern-matching to the other mmc cases, that one looks like it
> > _should_ have set the atomic flag like everybody else did.
>
> Oh, and immediately after sending that out I notice
> nvmet_bdev_execute_rw(), which does seem to make allocations inside
> that sg_miter loop.
>
> So those non-atomic cases do clearly exist.
>
> It does make the case for why kmap_atomic() wants to have the
> debugging code. It will "just work" on 64-bit to do it wrong, because
> the address doesn't become invalid just because you sleep or get
> rescheduled. But then the code that every developer tests (since
> developers tend to run on good hardware) might be completely broken on
> bad old hardware.
>

If we want code that is optimal on recent hardware, and yet still
correct on older 32-bit hardware, kmap() is definitely a better choice
here than kmap_atomic(), since it is a no-op on !HIGHMEM, and
tolerates sleeping on 32-bit. /That/ is why I wrote the code this way.

The problem is of course that kmap() itself might sleep.

So I would argue that the semantics in the name of kmap_atomic() are
not about the fact that it starts a non-preemptible section, but that
it can be *called from* a non-preemptible section. And starting a
non-preemptible section is unnecessary on !HIGHMEM, and should be
avoided if possible.

> Maybe we could hide it behind a debug option, at least.
>
> Or, alterantively, introduce a new "debug_preempt_count" that doesn't
> actually disable preemption, but warns about actual sleeping
> operations..
>
>              Linus

Re: [patch 00/13] preempt: Make preempt count unconditional

[Ard Biesheuvel]

On Tue, 15 Sep 2020 at 01:43, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> On Mon, Sep 14, 2020 at 3:24 PM Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
> >
> > Ard and Herbert added to participants: see
> > chacha20poly1305_crypt_sg_inplace(), which does
> >
> >         flags = SG_MITER_TO_SG;
> >         if (!preemptible())
> >                 flags |= SG_MITER_ATOMIC;
> >
> > introduced in commit d95312a3ccc0 ("crypto: lib/chacha20poly1305 -
> > reimplement crypt_from_sg() routine").
>
> As far as I can tell, the only reason for this all is to try to use
> "kmap()" rather than "kmap_atomic()".
>
> And kmap() actually has the much more complex "might_sleep()" tests,
> and apparently the "preemptible()" check wasn't even the proper full
> debug check, it was just a complete hack to catch the one that
> triggered.
>

This was not driven by a failing check.

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?

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.


> From a quick look, that code should probably just get rid of
> SG_MITER_ATOMIC entirely, and alwayse use kmap_atomic().
>
> kmap_atomic() is actually the faster and proper interface to use
> anyway (never mind that any of this matters on any sane hardware). The
> old kmap() and kunmap() interfaces should generally be avoided like
> the plague - yes, they allow sleeping in the middle and that is
> sometimes required, but if you don't need that, you should never ever
> use them.
>
> We used to have a very nasty kmap_atomic() that required people to be
> very careful and know exactly which atomic entry to use, and that was
> admitedly quite nasty.
>
> So it _looks_ like this code started using kmap() - probably back when
> kmap_atomic() was so cumbersome to use - and was then converted
> (conditionally) to kmap_atomic() rather than just changed whole-sale.
> Is there actually something that wants to use those sg_miter functions
> and sleep?
>
> Because if there is, that choice should come from the outside, not
> from inside lib/scatterlist.c trying to make some bad guess based on
> the wrong thing entirely.
>
>                  Linus

Re: [patch 00/13] preempt: Make preempt count unconditional

[Linus Torvalds]

On Mon, Sep 14, 2020 at 11:24 PM Herbert Xu <herbert@gondor.apana.org.au> wrote:
>
> 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 :)

Yeah, that used to be true a long long time ago, but the comment is very stale.

> > 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.

Well, honestly, one big reason for that is debugging.

The *semantics* of the kmap_atomic() is in the name - you can't sleep
in between it and the kunmap_atomic().

On any sane architecture, kmap_atomic() ends up being a no-op from an
implementation standpoint, and sleeping would work just fine.

But we very much want to make sure that people don't then write code
that doesn't work on the bad old 32-bit machines where it really needs
that sequence to be safe from preemption.

So it's mostly a debug thing.

I say "mostly", because there might be small other details too, like
shared code, and perhaps even a couple of users out in the wild that
depend on the pagefault_disable() inherent in the current
kmap_atomic(), who knows..

So no, the preemption disabling isn't inherent in the operation
itself. But it does have some argument for it.

                   Linus

Re: [patch 00/13] preempt: Make preempt count unconditional

[Thomas Gleixner]

On Mon, Sep 14 2020 at 23:39, Linus Torvalds wrote:
> On Mon, Sep 14, 2020 at 11:24 PM Herbert Xu <herbert@gondor.apana.org.au> wrote:
>> > 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.
>
> Well, honestly, one big reason for that is debugging.
>
> The *semantics* of the kmap_atomic() is in the name - you can't sleep
> in between it and the kunmap_atomic().
>
> On any sane architecture, kmap_atomic() ends up being a no-op from an
> implementation standpoint, and sleeping would work just fine.
>
> But we very much want to make sure that people don't then write code
> that doesn't work on the bad old 32-bit machines where it really needs
> that sequence to be safe from preemption.

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.

The approach is to keep track about the number of active maps in a task
and on an eventual context switch save them away in the task struct and
restore them when the task is scheduled back in.

Thanks,

        tglx

Re: [patch 00/13] preempt: Make preempt count unconditional

[Linus Torvalds]

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

Re: [patch 00/13] preempt: Make preempt count unconditional

[Thomas Gleixner]

On Mon, Sep 14 2020 at 15:24, Linus Torvalds wrote:
> On Mon, Sep 14, 2020 at 2:55 PM Thomas Gleixner <tglx@linutronix.de> wrote:
>>
>> 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
>
> I'm somewhat more worried about the small-device case.

I just checked on one of my old UP ARM toys which I run at home. The .text
increase is about 2% (75k) and none of the tests I ran showed any
significant difference. Couldn't verify with perf though as the PMU on
that piece of art is unusable.

> That said, the diffstat certainly has its very clear charm, and I do
> agree that it makes things simpler.
>
> I'm just not convinced people should ever EVER do things like that "if
> (preemptible())" garbage. It sounds like somebody is doing seriously
> bad things.

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.

We're currently looking through all of in_atomic(), in_interrupt()
etc. usage sites and quite some of them are historic and have the clear
intent of checking whether the code is called from task context or
hard/softirq context. Lots of them are completely broken or just work by
chance.

But there is clearly historic precendence that context checks are
useful, but they only can be useful if we have a consistent mechanism
which works everywhere.

Of course we could mandate that every interface which might be called
from one or the other context has a context argument or provides two
variants of the same thing. But I'm not really convinced whether that's
a win over having a consistent and reliable set of checks.

Thanks,

        tglx

Re: [patch 00/13] preempt: Make preempt count unconditional

[Linus Torvalds]

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

Re: [patch 00/13] preempt: Make preempt count unconditional

[Thomas Gleixner]

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

Re: [patch 00/13] preempt: Make preempt count unconditional

[Linus Torvalds]

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

Re: [patch 00/13] preempt: Make preempt count unconditional

[Daniel Vetter]

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
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Re: [patch 00/13] preempt: Make preempt count unconditional

[Paul E. McKenney]

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

Re: [patch 00/13] preempt: Make preempt count unconditional

[Linus Torvalds]

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

Re: [patch 00/13] preempt: Make preempt count unconditional

[Paul E. McKenney]

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

Re: [patch 00/13] preempt: Make preempt count unconditional

[Ard Biesheuvel]

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.

Re: [patch 00/13] preempt: Make preempt count unconditional

[Daniel Vetter]

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

Re: [patch 00/13] preempt: Make preempt count unconditional

[Paul E. McKenney]

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

Re: [patch 00/13] preempt: Make preempt count unconditional

[Daniel Vetter]

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
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Re: [patch 00/13] preempt: Make preempt count unconditional

[Paul E. McKenney]

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

Re: [patch 00/13] preempt: Make preempt count unconditional

[Daniel Vetter]

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
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Re: [patch 00/13] preempt: Make preempt count unconditional

[Paul E. McKenney]

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

Re: [patch 00/13] preempt: Make preempt count unconditional

[Michal Hocko]

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.

-- 
Michal Hocko
SUSE Labs

Re: [patch 00/13] preempt: Make preempt count unconditional

[Michal Hocko]

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.

-- 
Michal Hocko
SUSE Labs

Re: [patch 00/13] preempt: Make preempt count unconditional

[Michal Hocko]

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.

-- 
Michal Hocko
SUSE Labs

Re: [patch 00/13] preempt: Make preempt count unconditional

[Michal Hocko]

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.

-- 
Michal Hocko
SUSE Labs

Re: [patch 00/13] preempt: Make preempt count unconditional

[Michal Hocko]

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.

-- 
Michal Hocko
SUSE Labs

Re: [patch 00/13] preempt: Make preempt count unconditional

[Daniel Vetter]

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
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Re: [patch 00/13] preempt: Make preempt count unconditional

[Michal Hocko]

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...
-- 
Michal Hocko
SUSE Labs

Re: [patch 00/13] preempt: Make preempt count unconditional

[Matthew Wilcox]

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 ...

Re: [patch 00/13] preempt: Make preempt count unconditional

[Paul E. McKenney]

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

Re: [patch 03/13] preempt: Clenaup PREEMPT_COUNT leftovers

[Valentin Schneider]

On 14/09/20 21:42, Thomas Gleixner wrote:
> CONFIG_PREEMPT_COUNT is now unconditionally enabled and will be
> removed. Cleanup the leftovers before doing so.
>
> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
> Cc: Ingo Molnar <mingo@kernel.org>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Juri Lelli <juri.lelli@redhat.com>
> Cc: Vincent Guittot <vincent.guittot@linaro.org>
> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
> Cc: Steven Rostedt <rostedt@goodmis.org>
> Cc: Ben Segall <bsegall@google.com>
> Cc: Mel Gorman <mgorman@suse.de>
> Cc: Daniel Bristot de Oliveira <bristot@redhat.com>

Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>

Re: [patch 08/13] sched: Clenaup PREEMPT_COUNT leftovers

[Valentin Schneider]

On 14/09/20 21:42, Thomas Gleixner wrote:
> CONFIG_PREEMPT_COUNT is now unconditionally enabled and will be
> removed. Cleanup the leftovers before doing so.
>
> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Juri Lelli <juri.lelli@redhat.com>
> Cc: Vincent Guittot <vincent.guittot@linaro.org>
> Cc: Dietmar Eggemann <dietmar.eggemann@arm.com>
> Cc: Steven Rostedt <rostedt@goodmis.org>
> Cc: Ben Segall <bsegall@google.com>
> Cc: Mel Gorman <mgorman@suse.de>
> Cc: Daniel Bristot de Oliveira <bristot@redhat.com>

Small nit below;

Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>

> ---
>  kernel/sched/core.c |    6 +-----
>  lib/Kconfig.debug   |    1 -
>  2 files changed, 1 insertion(+), 6 deletions(-)
>
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -3706,8 +3706,7 @@ asmlinkage __visible void schedule_tail(
>        * finish_task_switch() for details.
>        *
>        * finish_task_switch() will drop rq->lock() and lower preempt_count
> -	 * and the preempt_enable() will end up enabling preemption (on
> -	 * PREEMPT_COUNT kernels).

I suppose this wanted to be s/PREEMPT_COUNT/PREEMPT/ in the first place,
which ought to be still relevant.

> +	 * and the preempt_enable() will end up enabling preemption.
>        */
>
>       rq = finish_task_switch(prev);