__virt_addr_valid vs virtual percpu areas

__virt_addr_valid vs virtual percpu areas

[Jeremy Fitzhardinge]

On i386, __virt_addr_valid() has the test:

	if (system_state != SYSTEM_BOOTING && is_vmalloc_addr((void *) x))
		return false;


Why is the vmalloc area a valid virtual address while the system is 
booting?  This is biting me because I need to translate percpu addresses 
to pfns, but I only bother doing the full pagetable walk if 
virt_addr_valid() is false (otherwise I just use __pa()).

Removing this test doesn't seem to harm anything at first glance.  Is 
this OK to do in general (and can we quietly set fire to system_state 
while we're about it)?

Thanks,
    J

Re: __virt_addr_valid vs virtual percpu areas

[Jiri Slaby]

On 4.3.2009 23:57, Jeremy Fitzhardinge wrote:
> On i386, __virt_addr_valid() has the test:
>
> if (system_state != SYSTEM_BOOTING && is_vmalloc_addr((void *) x))
> return false;
>
>
> Why is the vmalloc area a valid virtual address while the system is
> booting?

It's not (in the meaning of virt_* functions), but while booting we 
don't have variables used in VMALLOC_START and VMALLOC_END ready for use 
on i386.

Maybe we can introduce more clever method/state which would say: hey, 
vmalloc framework is up and running. And use instead (system_state != 
SYSTEM_BOOTING) hack.

> This is biting me because I need to translate percpu addresses
> to pfns, but I only bother doing the full pagetable walk if
> virt_addr_valid() is false (otherwise I just use __pa()).

Do you need to bother also with vmalloc space?

> Removing this test doesn't seem to harm anything at first glance. Is
> this OK to do in general (and can we quietly set fire to system_state
> while we're about it)?

I wouldn't do that, since vmalloc addr is not virt addr, again in the 
meaning of virt_* functions. And the function wouldn't do the right 
thing, at least in the RUNNING state anymore.

Re: __virt_addr_valid vs virtual percpu areas

[Jeremy Fitzhardinge]

Jiri Slaby wrote:
> It's not (in the meaning of virt_* functions), but while booting we 
> don't have variables used in VMALLOC_START and VMALLOC_END ready for 
> use on i386.
>
> Maybe we can introduce more clever method/state which would say: hey, 
> vmalloc framework is up and running. And use instead (system_state != 
> SYSTEM_BOOTING) hack.

Yeah, I pieced that together.  I'm just testing a patch with a specific 
__vmalloc_start_set flag to test.

>> This is biting me because I need to translate percpu addresses
>> to pfns, but I only bother doing the full pagetable walk if
>> virt_addr_valid() is false (otherwise I just use __pa()).
>
> Do you need to bother also with vmalloc space?

percpu data gets mapped into the vmalloc region now (when using 4k page 
mappings, at least).

>> Removing this test doesn't seem to harm anything at first glance. Is
>> this OK to do in general (and can we quietly set fire to system_state
>> while we're about it)?
>
> I wouldn't do that, since vmalloc addr is not virt addr, again in the 
> meaning of virt_* functions. And the function wouldn't do the right 
> thing, at least in the RUNNING state anymore.

OK.

    J