[Mesa-dev] [PATCH 08/11] radeonsi: fix memory exhaustion issue with DCC statistics gathering with DRI2

[Marek Olšák]

On Fri, Jun 29, 2018 at 11:40 AM, Dylan Baker <dylan at pnwbakers.com> wrote:
> Quoting Marek Olšák (2018-06-18 16:33:09)
>> From: Marek Olšák <marek.olsak at amd.com>
>>
>> Cc: 18.1 <mesa-stable at lists.freedesktop.org>
>> ---
>>  src/gallium/drivers/radeonsi/si_blit.c | 30 +++++++++++++++++++++++---
>>  1 file changed, 27 insertions(+), 3 deletions(-)
>>
>> diff --git a/src/gallium/drivers/radeonsi/si_blit.c b/src/gallium/drivers/radeonsi/si_blit.c
>> index fe059b36577..93cf7fe9001 100644
>> --- a/src/gallium/drivers/radeonsi/si_blit.c
>> +++ b/src/gallium/drivers/radeonsi/si_blit.c
>> @@ -1320,23 +1320,47 @@ static void si_flush_resource(struct pipe_context *ctx,
>>         if (rtex->dcc_separate_buffer && !rtex->separate_dcc_dirty)
>>                 return;
>>
>>         if (!rtex->is_depth && (rtex->cmask.size || rtex->dcc_offset)) {
>>                 si_blit_decompress_color(sctx, rtex, 0, res->last_level,
>>                                          0, util_max_layer(res, 0),
>>                                          rtex->dcc_separate_buffer != NULL);
>>         }
>>
>>         /* Always do the analysis even if DCC is disabled at the moment. */
>> -       if (rtex->dcc_gather_statistics && rtex->separate_dcc_dirty) {
>> -               rtex->separate_dcc_dirty = false;
>> -               vi_separate_dcc_process_and_reset_stats(ctx, rtex);
>> +       if (rtex->dcc_gather_statistics) {
>> +               bool separate_dcc_dirty = rtex->separate_dcc_dirty;
>> +
>> +               /* If the color buffer hasn't been unbound and fast clear hasn't
>> +                * been used, separate_dcc_dirty is false, but there may have been
>> +                * new rendering. Check if the color buffer is bound and assume
>> +                * it's dirty.
>> +                *
>> +                * Note that DRI2 never unbinds window colorbuffers, which means
>> +                * the DCC pipeline statistics query would never be re-set and would
>> +                * keep adding new results until all free memory is exhausted if we
>> +                * didn't do this.
>> +                */
>> +               if (!separate_dcc_dirty) {
>> +                       for (unsigned i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
>> +                               if (sctx->framebuffer.state.cbufs[i] &&
>> +                                   sctx->framebuffer.state.cbufs[i]->texture == res) {
>> +                                       separate_dcc_dirty = true;
>> +                                       break;
>> +                               }
>> +                       }
>> +               }
>> +
>> +               if (separate_dcc_dirty) {
>> +                       rtex->separate_dcc_dirty = false;
>> +                       vi_separate_dcc_process_and_reset_stats(ctx, rtex);
>> +               }
>>         }
>>  }
>>
>>  void si_decompress_dcc(struct si_context *sctx, struct r600_texture *rtex)
>>  {
>>         if (!rtex->dcc_offset)
>>                 return;
>>
>>         si_blit_decompress_color(sctx, rtex, 0, rtex->buffer.b.b.last_level,
>>                                  0, util_max_layer(&rtex->buffer.b.b, 0),
>> --
>> 2.17.1
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
> Hi Marek,
>
> This didn't apply cleanly to 18.1 because of
> 1ba87f4438069964af6548f4fa05386be999f4de (radeonsi: rename r600_texture ->
> si_texture, rxxx -> xxx or sxxx), I've attempted to rebase the commit by
> changing "tex" to "rtex", please take a look at the commit in staging/18.1 in
> the main tree or 18.1-proposed in my tree and let me know if it looks good to
> you.

Yes, it looks good. Thanks.

Marek