Re: [PATCH 1/7] drm/xe: Use ring ops TLB invalidation for rebinds

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

On Thu, 2024-03-21 at 19:09 +0000, Matthew Brost wrote:
> 
> 
> Can we simplify this too?
> 
> 	if (vm && (vm->batch_invalidate_tlb || (vm->tlb_flush_seqno
> != q->tlb_flush_seqno))) {
> 		q->tlb_flush_seqno = vm->tlb_flush_seqno;
> 		job->ring_ops_flush_tlb = true;
> 	}
> 
> I think this works as xe_sched_job_is_migration has
> emit_migration_job_gen12 which doesn't look at job-
> >ring_ops_flush_tlb,
> so no need to xe_sched_job_is_migration.
> 
> Also no need to check xe_vm_in_lr_mode as we wouldn'y increment the
> seqno above if that true.
> 
> Lastly, harmless to increment q->tlb_flush_seqno in the case of
> batch_invalidate_tlb being true.

I think I can simplify it a bit. Problem is that neither the migration
vm nor lr mode grabs the vm->resv at arm time so we would access the
seqnos unlocked and potentially get caught by clever static analyzers.

I actually had an assert for that at one point. I should probably re-
add it.

/Thomas