[PATCH v2] drm/xe: flush gtt before signalling user fence on all engines

[Matthew Brost]

On Thu, May 30, 2024 at 01:17:32PM +0200, Thomas Hellström wrote:
> Hi, All.
> 
> I was looking at this patch for drm-xe-fixes but it doesn't look
> correct to me.
> 
> First, AFAICT, the "emit flush imm ggtt" means that we're flushing
> outstanding / posted writes, and then write a DW to a ggtt address, so
> we're not really "flushing gtt"
> 

So, is this a bad name? I think I agree. It could have been a holdover
from the i915 names. Maybe we should do a cleanup in xe_ring_ops soon?

Or are you saying that the existing emit_flush_imm_ggtt is not
sufficient to ensure all writes from batches are visible? If this were
true, I would think we'd have all sorts of problems popping up.

> Second, I don't think we have anything left that explicitly flushes the
> posted write of the user-fence value?
> 

I think this might be true. So there could be a case where we get an IRQ
and the user fence value is not yet visible? 

Not an expert ring programming but are instructions to store a dword
which make these immediately visible? If so, I think that is what should
be used.

I think this might be true. So, there could be a case where we get an
IRQ and the user fence value is not yet visible?

I'm not an expert in ring programming, but are instructions to store a
dword which make these immediately visible? If so, I think that is what
should be used.

We should also probably check how downstream i915 did this too.

> and finally the seqno fence now gets flushed before the user-fence.
> Perhaps that's not a bad thing, though.
>

I don't think this is an issue, I can't think of a case where this
reordering would create a problem.

Matt
 
> /Thomas
> 
> 
> On Wed, 2024-05-22 at 09:27 +0200, Andrzej Hajda wrote:
> > Tests show that user fence signalling requires kind of write barrier,
> > otherwise not all writes performed by the workload will be available
> > to userspace. It is already done for render and compute, we need it
> > also for the rest: video, gsc, copy.
> > 
> > v2: added gsc and copy engines, added fixes and r-b tags
> > 
> > Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/1488
> > Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel
> > GPUs")
> > Signed-off-by: Andrzej Hajda <andrzej.hajda at intel.com>
> > Reviewed-by: Matthew Brost <matthew.brost at intel.com>
> > ---
> > Changes in v2:
> > - Added fixes and r-b tags
> > - Link to v1:
> > https://lore.kernel.org/r/20240521-xu_flush_vcs_before_ufence-v1-1-ded38b56c8c9@intel.com
> > ---
> > Matthew,
> > 
> > I have extended patch to copy and gsc engines. I have kept your r-b,
> > since the change is similar, I hope it is OK.
> > ---
> >  drivers/gpu/drm/xe/xe_ring_ops.c | 8 ++++----
> >  1 file changed, 4 insertions(+), 4 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_ring_ops.c
> > b/drivers/gpu/drm/xe/xe_ring_ops.c
> > index a3ca718456f6..a46a1257a24f 100644
> > --- a/drivers/gpu/drm/xe/xe_ring_ops.c
> > +++ b/drivers/gpu/drm/xe/xe_ring_ops.c
> > @@ -234,13 +234,13 @@ static void __emit_job_gen12_simple(struct
> > xe_sched_job *job, struct xe_lrc *lrc
> >  
> >  	i = emit_bb_start(batch_addr, ppgtt_flag, dw, i);
> >  
> > +	i = emit_flush_imm_ggtt(xe_lrc_seqno_ggtt_addr(lrc), seqno,
> > false, dw, i);
> > +
> >  	if (job->user_fence.used)
> >  		i = emit_store_imm_ppgtt_posted(job-
> > >user_fence.addr,
> >  						job-
> > >user_fence.value,
> >  						dw, i);
> >  
> > -	i = emit_flush_imm_ggtt(xe_lrc_seqno_ggtt_addr(lrc), seqno,
> > false, dw, i);
> > -
> >  	i = emit_user_interrupt(dw, i);
> >  
> >  	xe_gt_assert(gt, i <= MAX_JOB_SIZE_DW);
> > @@ -293,13 +293,13 @@ static void __emit_job_gen12_video(struct
> > xe_sched_job *job, struct xe_lrc *lrc,
> >  
> >  	i = emit_bb_start(batch_addr, ppgtt_flag, dw, i);
> >  
> > +	i = emit_flush_imm_ggtt(xe_lrc_seqno_ggtt_addr(lrc), seqno,
> > false, dw, i);
> > +
> >  	if (job->user_fence.used)
> >  		i = emit_store_imm_ppgtt_posted(job-
> > >user_fence.addr,
> >  						job-
> > >user_fence.value,
> >  						dw, i);
> >  
> > -	i = emit_flush_imm_ggtt(xe_lrc_seqno_ggtt_addr(lrc), seqno,
> > false, dw, i);
> > -
> >  	i = emit_user_interrupt(dw, i);
> >  
> >  	xe_gt_assert(gt, i <= MAX_JOB_SIZE_DW);
> > 
> > ---
> > base-commit: 188ced1e0ff892f0948f20480e2e0122380ae46d
> > change-id: 20240521-xu_flush_vcs_before_ufence-a7b45d94cf33
> > 
> > Best regards,
>