09:45 <marex> OK, so yesterday I continued on this MMU stuff ...

09:45 <marex> I think I am still confused by it and I didn't manage to find any useful information on this

09:45 <marex> but it seems there is a BO that is missing in the MMUv2 tables

09:47 <marex> there is apparently MTLB entry, but no STLB entry, but maybe I'm decoding the information wrong

10:17 <marex> lynxeye: austriancoder: ^

10:19 <lynxeye> marex: Did you take into account that the devcoredump only contains STLB entries for the populated STLB slots?

10:19 <marex> lynxeye: yeah

10:19 <marex> lynxeye: there is one 4k page for MTLB and then populated 4k pages for STLB, right ?

10:20 <lynxeye> marex: yep

10:23 <lynxeye> Okay, so you have a BO referenced by the current job, but no page table entries? This smells like a kernel driver bug, the page tables entries should not disappear before the BO...

10:23 <lynxeye> Is this just a single BO and all the others have pagetable entries?

10:24 <marex> lynxeye: let me decode this again and pastebin it, but yes, that's what it is

10:26 <marex> lynxeye: so e.g. I get this

10:26 <marex> [ 2263.976187] etnaviv-gpu 59000000.gpu: MMU fault status 0x00000002

10:26 <marex> [ 2263.980880] etnaviv-gpu 59000000.gpu: MMU 0 fault addr 0xfe5b0f80

10:28 <marex> and then

10:28 <marex> viv-unpack etnaviv-20210426232145.bin . ; hexdump -C mmu.bin

10:28 <marex> produces

10:29 <marex> https://paste.debian.net/hidden/79d80f99/ this

10:29 <marex> and also there is this

10:32 <marex> https://paste.debian.net/hidden/d07bd43c/

10:33 <marex> so I suppose, I am looking at MTLB entry

10:33 <marex> 00000fe0 01 40 86 fb 01 50 86 fb 01 e0 85 fb 01 f0 85 fb |.@...P..........|

10:33 <marex> 0xfe4 ^^^^^^^^^^^

10:33 <marex> that one is PRESENT

10:34 <marex> and since I have 14 MTLB entries in total, this is the 8th, then the STLB offset is 0x8000

10:36 <marex> which is 0x2 , EXCEPTION

10:36 <marex> but there is a BO which is supposed to be there, 5 bo fe4e9000 00100000 1048576

11:51 <marex> austriancoder: lynxeye: ^

12:00 <lynxeye> marex: I agree with your analysis. This looks very suspicious.

12:01 <lynxeye> However I'm currently can't see how we get into this situation. Looks like I need to do an audit on the kernel driver MMU code...

12:05 <marex> lynxeye: got any hints for debugging this ?

15:51 <marex> lynxeye: got any hints for debugging this ?

15:52 <lynxeye> marex: Not really, yet. One thing I wonder however is if you can make it more likely to hit this issue by disabling the userspace BO cache.

16:05 <mntmn> marex: sorry but which gpu are you working with atm? gc7000?

16:32 <marex> mntmn: gc400

16:35 <marex> lynxeye: is there an env var for that or do I have to hack it out ?

16:38 <mntmn> ok

16:40 <lynxeye> marex: no env var afair