Thanks to NVIDIA for confirming this workaround, and clarifying which HW
is affected.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Tested-by: Alexander Kapshuk <alexander.kapshuk@gmail.com>
dmac->push->end = dmac->push->bgn;
dmac->max = 0x1000/4 - 1;
+ /* EVO channels are affected by a HW bug where the last 12 DWORDs
+ * of the push buffer aren't able to be used safely.
+ */
+ if (disp->oclass < GV100_DISP)
+ dmac->max -= 12;
+
args->pushbuf = nvif_handle(&dmac->_push.mem.object);
ret = nv50_chan_create(device, disp, oclass, head, data, size,