[PATCH] coredump: Replace opencoded set_mask_bits()

[PATCH] coredump: Replace opencoded set_mask_bits()

[Vineet Gupta]

Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: linux-fsdevel@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
---
 fs/exec.c | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/fs/exec.c b/fs/exec.c
index 1977c2a553ac..25078627f048 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -1690,15 +1690,10 @@ EXPORT_SYMBOL(set_binfmt);
  */
 void set_dumpable(struct mm_struct *mm, int value)
 {
-	unsigned long old, new;
-
 	if (WARN_ON((unsigned)value > SUID_DUMP_ROOT))
 		return;
 
-	do {
-		old = ACCESS_ONCE(mm->flags);
-		new = (old & ~MMF_DUMPABLE_MASK) | value;
-	} while (cmpxchg(&mm->flags, old, new) != old);
+	set_mask_bits(&mm->flags, MMF_DUMPABLE_MASK, value);
 }
 
 SYSCALL_DEFINE3(execve,
-- 
1.9.1

Re: [PATCH] coredump: Replace opencoded set_mask_bits()

[Oleg Nesterov]

On 08/07, Vineet Gupta wrote:
>
> --- a/fs/exec.c
> +++ b/fs/exec.c
> @@ -1690,15 +1690,10 @@ EXPORT_SYMBOL(set_binfmt);
>   */
>  void set_dumpable(struct mm_struct *mm, int value)
>  {
> -	unsigned long old, new;
> -
>  	if (WARN_ON((unsigned)value > SUID_DUMP_ROOT))
>  		return;
>  
> -	do {
> -		old = ACCESS_ONCE(mm->flags);
> -		new = (old & ~MMF_DUMPABLE_MASK) | value;
> -	} while (cmpxchg(&mm->flags, old, new) != old);
> +	set_mask_bits(&mm->flags, MMF_DUMPABLE_MASK, value);
>  }

Acked-by: Oleg Nesterov <oleg@redhat.com>

Re: [PATCH] coredump: Replace opencoded set_mask_bits()

[Vineet Gupta]

On Friday 07 August 2015 05:27 PM, Oleg Nesterov wrote:
> On 08/07, Vineet Gupta wrote:
>>
>> --- a/fs/exec.c
>> +++ b/fs/exec.c
>> @@ -1690,15 +1690,10 @@ EXPORT_SYMBOL(set_binfmt);
>>   */
>>  void set_dumpable(struct mm_struct *mm, int value)
>>  {
>> -	unsigned long old, new;
>> -
>>  	if (WARN_ON((unsigned)value > SUID_DUMP_ROOT))
>>  		return;
>>  
>> -	do {
>> -		old = ACCESS_ONCE(mm->flags);
>> -		new = (old & ~MMF_DUMPABLE_MASK) | value;
>> -	} while (cmpxchg(&mm->flags, old, new) != old);
>> +	set_mask_bits(&mm->flags, MMF_DUMPABLE_MASK, value);
>>  }
> 
> Acked-by: Oleg Nesterov <oleg@redhat.com>


I have a fundamental question though, perhaps stupid, do use cases like these
warrant the data to be atomic_t in first place. Do API like set_mask_bits() make
sense at all - or shd they be moved to atomic_* (after changing the underlying data)

See, I have such a cmpxchg loop in ARC code - originally from Peter :-)
arch/arc/kernel/smp.c. @ipi_data_ptr is NOT atomic_t

	do {
		new = old = ACCESS_ONCE(*ipi_data_ptr);
		new |= 1U << msg;
	} while (cmpxchg(ipi_data_ptr, old, new) != old);


Given that ARC (and some other RISC cores) lack native cmpxchg, we use LLSC
instructions to implement atomics including cpmxchg - the implementation itself
ensures loop is builtin making the outer loping superfluous and waste of cycles
(see e.g. cover letter @ http://www.spinics.net/lists/kernel/msg2029217.html)

So I wanted to convert that loop (and similar other cases to "some" API which
could be built conditionally based on cmpxchg or llsc. None such exist and I was
thinking of converting my case to atomic_t. Is that the right approach ?

Thx,
-Vineet

Re: [PATCH] coredump: Replace opencoded set_mask_bits()

[Peter Zijlstra]

On Fri, Aug 07, 2015 at 08:14:03PM +0530, Vineet Gupta wrote:

> See, I have such a cmpxchg loop in ARC code - originally from Peter :-)
> arch/arc/kernel/smp.c. @ipi_data_ptr is NOT atomic_t
> 
> 	do {
> 		new = old = ACCESS_ONCE(*ipi_data_ptr);
> 		new |= 1U << msg;
> 	} while (cmpxchg(ipi_data_ptr, old, new) != old);
> 

Well, you'll have atomic_or() real soon now.

Re: [PATCH] coredump: Replace opencoded set_mask_bits()

[Vineet Gupta]

On Friday 07 August 2015 08:27 PM, Peter Zijlstra wrote:
> On Fri, Aug 07, 2015 at 08:14:03PM +0530, Vineet Gupta wrote:
> 
>> > See, I have such a cmpxchg loop in ARC code - originally from Peter :-)
>> > arch/arc/kernel/smp.c. @ipi_data_ptr is NOT atomic_t
>> > 
>> > 	do {
>> > 		new = old = ACCESS_ONCE(*ipi_data_ptr);
>> > 		new |= 1U << msg;
>> > 	} while (cmpxchg(ipi_data_ptr, old, new) != old);
>> > 
> Well, you'll have atomic_or() real soon now.


Doesn't help my cause - ipi_data_ptr is not atomic_t - hence my prev question in
this thread

Re: [PATCH] coredump: Replace opencoded set_mask_bits()

[Peter Zijlstra]

On Fri, Aug 07, 2015 at 09:05:06PM +0530, Vineet Gupta wrote:
> On Friday 07 August 2015 08:27 PM, Peter Zijlstra wrote:
> > On Fri, Aug 07, 2015 at 08:14:03PM +0530, Vineet Gupta wrote:
> > 
> >> > See, I have such a cmpxchg loop in ARC code - originally from Peter :-)
> >> > arch/arc/kernel/smp.c. @ipi_data_ptr is NOT atomic_t
> >> > 
> >> > 	do {
> >> > 		new = old = ACCESS_ONCE(*ipi_data_ptr);
> >> > 		new |= 1U << msg;
> >> > 	} while (cmpxchg(ipi_data_ptr, old, new) != old);
> >> > 
> > Well, you'll have atomic_or() real soon now.
> 
> Doesn't help my cause - ipi_data_ptr is not atomic_t - hence my prev question in
> this thread

A cast will work :-)

But yes, ideally everything will be type safe because of those archs
that cannot have atomic RmW ops like !ARC_HAS_LLSC.

Mixing cmpxchg()/xchg() with regular stores is broken on those.

Fwiw, you might want to set ARCH_SUPPORTS_ATOMIC_RMW for ARC_HAS_LLSC.

Re: [PATCH] coredump: Replace opencoded set_mask_bits()

[Vineet Gupta]

On Friday 07 August 2015 09:15 PM, Peter Zijlstra wrote:
> On Fri, Aug 07, 2015 at 09:05:06PM +0530, Vineet Gupta wrote:
>> On Friday 07 August 2015 08:27 PM, Peter Zijlstra wrote:
>>> On Fri, Aug 07, 2015 at 08:14:03PM +0530, Vineet Gupta wrote:
>>>
>>>>> See, I have such a cmpxchg loop in ARC code - originally from Peter :-)
>>>>> arch/arc/kernel/smp.c. @ipi_data_ptr is NOT atomic_t
>>>>>
>>>>> 	do {
>>>>> 		new = old = ACCESS_ONCE(*ipi_data_ptr);
>>>>> 		new |= 1U << msg;
>>>>> 	} while (cmpxchg(ipi_data_ptr, old, new) != old);
>>>>>
>>> Well, you'll have atomic_or() real soon now.
>>
>> Doesn't help my cause - ipi_data_ptr is not atomic_t - hence my prev question in
>> this thread
> 
> A cast will work :-)
> 

How ? We have

typedef struct {
 int counter;
} atomic_t;

> But yes, ideally everything will be type safe because of those archs
> that cannot have atomic RmW ops like !ARC_HAS_LLSC.

Type safe - how / what ?

> 
> Mixing cmpxchg()/xchg() with regular stores is broken on those.

Right, but how does that relate to this discussion - perhaps I shd stop talking -
long friday already :-)

> 
> Fwiw, you might want to set ARCH_SUPPORTS_ATOMIC_RMW for ARC_HAS_LLSC.
> 

Sure I will !

Re: [PATCH] coredump: Replace opencoded set_mask_bits()

[Peter Zijlstra]

On Fri, Aug 07, 2015 at 09:28:05PM +0530, Vineet Gupta wrote:
> On Friday 07 August 2015 09:15 PM, Peter Zijlstra wrote:
> > On Fri, Aug 07, 2015 at 09:05:06PM +0530, Vineet Gupta wrote:
> >> On Friday 07 August 2015 08:27 PM, Peter Zijlstra wrote:
> >>> On Fri, Aug 07, 2015 at 08:14:03PM +0530, Vineet Gupta wrote:
> >>>
> >>>>> See, I have such a cmpxchg loop in ARC code - originally from Peter :-)
> >>>>> arch/arc/kernel/smp.c. @ipi_data_ptr is NOT atomic_t
> >>>>>
> >>>>> 	do {
> >>>>> 		new = old = ACCESS_ONCE(*ipi_data_ptr);
> >>>>> 		new |= 1U << msg;
> >>>>> 	} while (cmpxchg(ipi_data_ptr, old, new) != old);
> >>>>>
> >>> Well, you'll have atomic_or() real soon now.
> >>
> >> Doesn't help my cause - ipi_data_ptr is not atomic_t - hence my prev question in
> >> this thread
> > 
> > A cast will work :-)
> > 
> 
> How ? We have
> 
> typedef struct {
>  int counter;
> } atomic_t;

ARC is 32bit, right? So int and unsigned long are of the same size.
Therefore:

	atomic_or(1 << msg, (atomic_t *)ipi_data_ptr);

Ugly, yes, but it should DTRT.

> > But yes, ideally everything will be type safe because of those archs
> > that cannot have atomic RmW ops like !ARC_HAS_LLSC.
> 
> Type safe - how / what ?

All atomic stuff restricted to atomic*t and bitmap functions (and
ideally we'd also have bitmap_t to avoid passing random unsigned long *
into bitmap functions and praying it all works, we do, and it doesn't,
well mostly :-).

> > Mixing cmpxchg()/xchg() with regular stores is broken on those.
> 
> Right, but how does that relate to this discussion - perhaps I shd stop talking -
> long friday already :-)

:-)

Well, its a very good argument for why we should not use cmpxchg/xchg on
!atomic*t types, and therefore why the function at hand (set_mask_bit)
should really be on an atomic_t.

That said, it will probably make the fs code fugly for having to use
atomic_t and all its accessors all over the place.