Re: [RFC v2][PATCH] dynamically enable readprofile at runtime

[Dave Hansen]

On Wed, 2008-09-10 at 15:59 -0700, Andrew Morton wrote: 
> On Tue, 09 Sep 2008 11:05:36 -0700
> Dave Hansen <dave@xxxxxxxxxxxxxxxxxx> wrote:
> > Way too often, I have a machine that exhibits some kind of crappy
> > behavior.  The CPU looks wedged in the kernel or it is spending
> > way too much system time and I wonder what is responsible.
> > 
> > I try to run readprofile.  But, of course, Ubuntu doesn't enable
> > it by default.  Dang!
> > 
> > The reason we boot-time enable it is that it takes a big bufffer
> > that we generally can only bootmem alloc.  But, does it hurt to
> > at least try and runtime-alloc it?
> > 
> > To use:
> > echo 2 > /sys/kernel/profile
> > 
> > Then run readprofile like normal.
> > 
> > This should fix the compile issue with allmodconfig.  I've
> > compile-tested on a bunch more configs now including a few
> > more architectures.
> 
> Can it be turned off again?  afaict: no?

No.  Probably wouldn't be that hard, but I wanted to see if anyone would
take this, first. :)

> > +#ifdef CONFIG_PROFILING
> > +static ssize_t profiling_show(struct kobject *kobj,
> > +				  struct kobj_attribute *attr, char *buf)
> > +{
> > +	return sprintf(buf, "%d\n", prof_on);
> > +}
> > +static ssize_t profiling_store(struct kobject *kobj,
> > +				   struct kobj_attribute *attr,
> > +				   const char *buf, size_t count)
> > +{
> > +	int ret;
> > +
> > +	if (prof_on)
> > +		return -EEXIST;
> > +	/*
> > +	 * This eventually calls into get_option() which
> > +	 * has a ton of callers and is not const.  It is
> > +	 * easiest to cast it away here.
> > +	 */
> > +	profile_setup((char *)buf);
> > +	ret = profile_init();
> > +	if (ret)
> > +		return ret;
> > +	ret = create_proc_profile();
> > +	if (ret)
> > +		return ret;
> > +	return count;
> > +}
> > +KERNEL_ATTR_RW(profiling);
> > +#endif
> 
> Tested with CONFIG_SYSFS=n?

There's a bunch of other sysfs stuff in that file.  I'm actually not
quite sure what magic lets it build with SYSFS=n, but it does build.

> > -void __init profile_init(void)
> > +int profile_init(void)
> >  {
> > +	int buffer_bytes;
> >  	if (!prof_on)
> > -		return;
> > +		return 0;
> >  
> >  	/* only text is profiled */
> >  	prof_len = (_etext - _stext) >> prof_shift;
> > -	prof_buffer = alloc_bootmem(prof_len*sizeof(atomic_t));
> > +	buffer_bytes = prof_len*sizeof(atomic_t);
> > +	if (!slab_is_available()) {
> > +		prof_buffer = alloc_bootmem(buffer_bytes);
> > +		return 0;
> > +	}
> > +
> > +	prof_buffer = kzalloc(buffer_bytes, GFP_KERNEL);
> > +	if (prof_buffer)
> > +		return 0;
> > +
> > +	prof_buffer = alloc_pages_exact(buffer_bytes, GFP_KERNEL|__GFP_ZERO);
> > +	if (prof_buffer)
> > +		return 0;
> > +
> > +	prof_buffer = vmalloc(buffer_bytes);
> > +	if (prof_buffer)
> > +		return 0;
> > +
> > +	return -ENOMEM;
> >  }
> 
> Well that should cover it.
> 
> Did you check to see if any __GFP_NOWARNs are needed there?

No, because the two cases I've actually tested were where the buffer is
too big to be handled by the slab or buddy allocators.  It never even
tries to allocate the memory and doesn't issue the warnings.  But,
you're right, it needs them.  Added in new version.

> alloc_bootmem() will (apparently undocumentedly and secretly) zero the
> memory.
> 
> kzalloc() will zero the memory.
> 
> alloc_pages_exact(__GFP_ZERO) will zero the memory.
> 
> But what about vmalloc?  I see no documentation which says that it
> zeroes the memory, although it seems that some flavours of it will do
> this, but the nommu version does not.  Seems all messed up.

Yeah, it does seem messed up.  The right thing to do for this patch is
probably just to zero the vmalloc() result.  I'll look into vmalloc
zeroing a bit as well.

-- Dave

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/