etnaviv: *ERROR* relocation N outside object in kernel log

[Wladimir J. van der Laan]

On Tue, Jul 25, 2017 at 09:47:32AM +0200, Wladimir J. van der Laan wrote:
> Lucas,
> 
> > My guess would be some stale relocation again, possibly with only a
> > stale offset or something. But that's hard to track down.
> 
> I added some further debugging and it looks like the relocation points to
> offset 000ffffc of a 00100000-sized bo.
> 
> Haven't found out which state this is for yet, but is there any specific reason
> this would not be allowed, or is the check off-by-one here?
> 
> As a reminder, the code is:
> 
>     if (r->reloc_offset >= bo->obj->base.size - sizeof(*ptr)) {
>        error...
>     }

I've checked the state: it is 00694, so a vertex stream base address.

This is the annotated command buffer around where it happens:

[  208.130431] 058e  0801019d  SET_STATE 1 at 00674
[  208.133148] 058f  00000000
[  208.135864] 0590  080601a1  SET_STATE 6 at 00684 (FE.VERTEX_STREAMS.BASE_ADDR)
[  208.138580] 0591  883dd00c
[  208.141318] 0592  884ddff4
[  208.144036] 0593  883dd014
[  208.146751] 0594  884ddff8
[  208.149489] 0595* 00000000  (warning: relocation to offset 000ffffc of size 00100000 of some bo)
[  208.152206] 0596  00000000
[  208.154922] 0597  deadbeef
[  208.157638] 0598  080601a9  SET_STATE 6 at 006a4 (FE.VERTEX_STREAMS.CONTROL)
[  208.160377] 0599  00000024
[  208.163094] 059a  00000000
[  208.165811] 059b  00000024
[  208.168528] 059c  00000000
[  208.171267] 059d  00000000  (associated stride=0)
[  208.173985] 059e  00000000
[  208.176701] 059f  deadbeef
[  208.179440] 05a0  08010e02  SET_STATE 1 at 03808 (stall token)
[  208.182159] 05a1  00000701
[  208.184876] 05a2  48000000  STALL
[  208.187591] 05a3  00000701
[  208.190330] 05a4  0803140c  SET_STATE 3 at 05030 (vs uniforms)
[  208.193049] 05a5  c3c5b342
[  208.195766] 05a6  43e09672
[  208.198481] 05a7  41a903ec
[  208.201219] 05a8  08081418  SET_STATE 8 at 05060 (vs uniforms)
[  208.203936] 05a9  bf7ff14b
[  208.206654] 05aa  3ae7a09a
[  208.209369] 05ab  bcaddd5e
[  208.212111] 05ac  bcad52a0
[  208.214828] 05ad  3cad8f20
[  208.217543] 05ae  3daac96f
[  208.220281] 05af  bf803265
[  208.222999] 05b0  bf7f9832
[  208.225715] 05b1  deadbeef
[  208.228431] 05b2  08031425 SET_STATE 3 at 05094 (vs uniforms)
[  208.231170] 05b3  c28d0830
[  208.233887] 05b4  43d82db2
[  208.236603] 05b5  43db7f98
[  208.239318] 05b6  30000000 DRAW_INDEXED_PRIMITIVES
[  208.242056] 05b7  00000004 COMMAND TYPE=TRIANGLES
[  208.244773] 05b8  00000000 START 
[  208.247489] 05b9  0000004e COUNT
[  208.250227] 05ba  00000000 OFFSET
[  208.252944] 05bb  00000000

Vertex stream 5 points to the last four bytes of a buffer, and a draw of 0x4e
triangles is being done, so on first sight it seems wrong.

However, the stream has a stride of 0, so the same value will be reused for
every vertex. I checked the vertex elements config, and it appears to be
loading a single float.

So I suspect the check is a false positive.

Regards,
Wladimir