order 5 memory allocation failures in radeon_vm_get_bos

order 5 memory allocation failures in radeon_vm_get_bos

[Dave Airlie]

I have someone running F20 + a hacked up glamor to fix some leaks,
however he is getting
order 4/5 memory allocation failures in the kmalloc_array in radeon_vm_get_bos

radeon_vm_get_bos+0x30/0x120

I think we should fallback to vmalloc or just use drm_malloc_ab in here instead,

any reason we haven't?

Dave.

Re: order 5 memory allocation failures in radeon_vm_get_bos

[Michel Dänzer]

On 17.10.2014 09:23, Dave Airlie wrote:
> I have someone running F20 + a hacked up glamor to fix some leaks,
> however he is getting
> order 4/5 memory allocation failures in the kmalloc_array in radeon_vm_get_bos
>
> radeon_vm_get_bos+0x30/0x120
>
> I think we should fallback to vmalloc or just use drm_malloc_ab in here instead,
>
> any reason we haven't?

I don't think so, see https://bugs.freedesktop.org/show_bug.cgi?id=81991 
. I didn't know about drm_malloc_ab, that looks like the easiest fix indeed.


-- 
Earthling Michel Dänzer            |                  http://www.amd.com
Libre software enthusiast          |                Mesa and X developer

Re: order 5 memory allocation failures in radeon_vm_get_bos

[Christian König]

Am 17.10.2014 um 05:11 schrieb Michel Dänzer:
> On 17.10.2014 09:23, Dave Airlie wrote:
>> I have someone running F20 + a hacked up glamor to fix some leaks,
>> however he is getting
>> order 4/5 memory allocation failures in the kmalloc_array in 
>> radeon_vm_get_bos
>>
>> radeon_vm_get_bos+0x30/0x120
>>
>> I think we should fallback to vmalloc or just use drm_malloc_ab in 
>> here instead,
>>
>> any reason we haven't?
>
> I don't think so, see 
> https://bugs.freedesktop.org/show_bug.cgi?id=81991 . I didn't know 
> about drm_malloc_ab, that looks like the easiest fix indeed.

Mhm fixing this using another allocator function is probably a good 
idea, but on the other hand why does X want to allocate an order 4/5 in 
vm_get_bos in the first place?

Assuming 64 bytes per array element that would mean that we have over 1K 
BOs for the address space handling. Even with the lowest settings on BO 
size that covers something like 4GB of GPU address space and with normal 
settings it is more something like 256GB...

So question here is why does X need so much GPU address space? The box 
doesn't even have so much memory, doesn't it?

Cheers,
Christian.

Re: order 5 memory allocation failures in radeon_vm_get_bos

[Dave Airlie]

> Mhm fixing this using another allocator function is probably a good idea,
> but on the other hand why does X want to allocate an order 4/5 in vm_get_bos
> in the first place?
>
> Assuming 64 bytes per array element that would mean that we have over 1K BOs
> for the address space handling. Even with the lowest settings on BO size
> that covers something like 4GB of GPU address space and with normal settings
> it is more something like 256GB...
>
> So question here is why does X need so much GPU address space? The box
> doesn't even have so much memory, doesn't it?

The user I'm seeing it with us using glamor and I think he is using Xv a lot,
and has a lot of stuff open on his desktop and does a lot of gimp and
video editing.

I've gotten him a non-crashing glamor now (pre-Xserver glamor on F20)
and he only sees the alloc failures now.

Dave.