Re: I915 CI-run with kfence enabled, issues found

[Marco Elver]

[+Cc x86 maintainers]

On Mon, Mar 29, 2021 at 11:11AM +0000, Sarvela, Tomi P wrote:
> Hello,
> 
> I'm Tomi Sarvela, maintainer and original creator of linux i915-CI:
> https://intel-gfx-ci.01.org/
> 
> I got a hint from Martin Peres about kfence functionality in kernel, and it looked
> something we'd like to enable in future CI runs so I made a trial run on DRM-Tip.
> We've had regular KASAN-enabled runs, so the expectation was that there
> wouldn't be too many new problems exposed.
> 
> On this run two issues were found, where one is clearly kernel (GUC) issue,
> but another looked a lot like kfence issue on old platforms. Affected
> were IVB, SNB and ILK, with bug signature being:
> 
> <3> [31.556004] BUG: using smp_processor_id() in preemptible [00000000] code: ...
> <4> [31.556070] caller is invalidate_user_asid+0x13/0x50
> 
> I'm not a kernel developer myself, so I can't make hard assertions
> where the issue originates. In comparison to kernel without kfence,
> it looks like the newly enabled code is the cause because the
> "BUG: KFENCE" signature is missing from the trace
> 
> Can someone take a look at the traces and verify if the kfence issue
> exists and is not related to the rest of the kernel? 
> 
> If there is an issue tracker, I can add this information there.
> 
> Example traces:
> https://intel-gfx-ci.01.org/tree/drm-tip/kfence_1/fi-ivb-3770/igt@gem_ctx_create@xxxxxxxxxxxxxxxx
> 
> https://intel-gfx-ci.01.org/tree/drm-tip/kfence_1/fi-snb-2520m/igt@gem_ctx_create@xxxxxxxxxxxxxxxx
> 
> https://intel-gfx-ci.01.org/tree/drm-tip/kfence_1/fi-ilk-650/igt@gem_exec_create@xxxxxxxxxx
> 
> Kfence-exposed possible GUC issue:
> https://intel-gfx-ci.01.org/tree/drm-tip/kfence_1/fi-kbl-guc/igt@kms_addfb_basic@xxxxxxxxxxxxxxxxxxxxxxxxxxxxx
> 
> All results can be seen at:
> https://intel-gfx-ci.01.org/tree/drm-tip/kfence_1/index.html
> 
> CI_DRM_9910 is recent DRM-Tip commit without -rc5 pulled in yet.
> kfence_1 is same commit with kfence defaults turned on:
[...]

It looks like the code path from flush_tlb_one_kernel() to
invalidate_user_asid()'s this_cpu_ptr() has several feature checks, so
probably some feature difference between systems where it triggers and
it doesn't.

As far as I'm aware, there is no restriction on where
flush_tlb_one_kernel() is called. We could of course guard it but I
think that's wrong.

Other than that, I hope the x86 maintainers know what's going on here.

Just for reference, the stack traces in the above logs start with:

| <3> [31.556004] BUG: using smp_processor_id() in preemptible [00000000] code: dmesg/1075
| <4> [31.556070] caller is invalidate_user_asid+0x13/0x50
| <4> [31.556078] CPU: 6 PID: 1075 Comm: dmesg Not tainted 5.12.0-rc4-gda4a2b1a5479-kfence_1+ #1
| <4> [31.556081] Hardware name: Hewlett-Packard HP Pro 3500 Series/2ABF, BIOS 8.11 10/24/2012
| <4> [31.556084] Call Trace:
| <4> [31.556088]  dump_stack+0x7f/0xad
| <4> [31.556097]  check_preemption_disabled+0xc8/0xd0
| <4> [31.556104]  invalidate_user_asid+0x13/0x50
| <4> [31.556109]  flush_tlb_one_kernel+0x5/0x20
| <4> [31.556113]  kfence_protect+0x56/0x80
| 	...........

Thanks,
-- Marco