Re: can kfree sleep?

Re: can kfree sleep?

[Manfred Spraul]

Hi Peter,

>Just a question: can kfree sleep?
>
>
>  
>
No, it never sleeps. It's safe to call kfree from arbitrary context. The 
only exception is the NMI oopser and similar arch code.

>I believe so, but slab.c does not enlighten me immediately:
>  
>
Yes, the kfree code is quite long - it must check if freeing one object 
created a freeable page and return it to the page allocator. Together 
with lots of caching and debug checks.

--
    Manfred

Re: can kfree sleep?

[Peter T. Breuer]

In article <41A0E4E9.3040902@colorfullife.com> you wrote:
> Hi Peter,

Hi!

> >Just a question: can kfree sleep?
> >
> No, it never sleeps. It's safe to call kfree from arbitrary context. The 
> only exception is the NMI oopser and similar arch code.

OK - thanks. I'll take that as read and eliminate kfree from the list
of sleep "seeds" in my code analyser.

> >I believe so, but slab.c does not enlighten me immediately:
> >
> Yes, the kfree code is quite long - it must check if freeing one object 
> created a freeable page and return it to the page allocator. Together 
> with lots of caching and debug checks.

May I ask where your knowledge of this comes from? (So I can duplicate
the learning process)!

Peter

can kfree sleep?

[Peter T. Breuer]

Just a question: can kfree sleep?

I believe so, but slab.c does not enlighten me immediately:

 void
 kfree (const void *objp)
 {
        kmem_cache_t *c;
        unsigned long flags;

        if (!objp)
                return;
        local_irq_save (flags);
        c = GET_PAGE_CACHE (virt_to_page (objp));
        __cache_free (c, (void *) objp);
        local_irq_restore (flags);
 }

 static inline void __cache_free (kmem_cache_t *cachep, void* objp)
 {
        struct array_cache *ac = ac_data(cachep);

        check_irq_off();
        objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));

        if (likely(ac->avail < ac->limit)) {
                STATS_INC_FREEHIT(cachep);
                ac_entry(ac)[ac->avail++] = objp;
                return;
        } else {
                STATS_INC_FREEMISS(cachep);
                cache_flusharray(cachep, ac);
                ac_entry(ac)[ac->avail++] = objp;
        }
 }

 ...


Peter

Re: can kfree sleep?

[Andrew Morton]

"Peter T. Breuer" <ptb@it.uc3m.es> wrote:
>
> Just a question: can kfree sleep?

Nope.  All memory freeing codepaths are atomic and may be called from any
context except NMI handlers.

Re: can kfree sleep?

[Christoph Hellwig]

On Sun, Nov 21, 2004 at 01:10:38PM -0800, Andrew Morton wrote:
> Nope.  All memory freeing codepaths are atomic and may be called from any
> context except NMI handlers.

Not true for vfree()

Re: can kfree sleep?

[Peter T. Breuer]

In article <20041121211451.GA12826@infradead.org> you wrote:
> On Sun, Nov 21, 2004 at 01:10:38PM -0800, Andrew Morton wrote:
> > Nope.  All memory freeing codepaths are atomic and may be called from any
> > context except NMI handlers.
> 
> Not true for vfree()

My interest at the moment is in what can sleep and what cannot sleep.
Are you saying that vfree can sleep or that vfree cannot be called from
at least one other context in addition to the NMI handler context (from
which it cannot be called ...)?

:)

Peter

Re: can kfree sleep?

[Christoph Hellwig]

On Sun, Nov 21, 2004 at 11:30:38PM +0100, Peter T. Breuer wrote:
> In article <20041121211451.GA12826@infradead.org> you wrote:
> > On Sun, Nov 21, 2004 at 01:10:38PM -0800, Andrew Morton wrote:
> > > Nope.  All memory freeing codepaths are atomic and may be called from any
> > > context except NMI handlers.
> > 
> > Not true for vfree()
> 
> My interest at the moment is in what can sleep and what cannot sleep.
> Are you saying that vfree can sleep or that vfree cannot be called from
> at least one other context in addition to the NMI handler context (from
> which it cannot be called ...)?

vfree can't sleep, but it can't be called from every context.

Re: can kfree sleep?

[Arjan van de Ven]

> > Not true for vfree()
> 
> My interest at the moment is in what can sleep and what cannot sleep.
> Are you saying that vfree can sleep or that vfree cannot be called from
> at least one other context in addition to the NMI handler context (from
> which it cannot be called ...)?


vfree() is not allowed to be called from irq context, and since it can do cross cpu IPI's, you have to generally be careful with it.