#include "radeon_reg.h"
 #include "radeon.h"
 #include "atom.h"
+#include "r100d.h"
 #include "r420d.h"
 
 int r420_mc_init(struct radeon_device *rdev)
        WREG32_PLL(R_00000D_SCLK_CNTL, sclk_cntl);
 }
 
+static void r420_cp_errata_init(struct radeon_device *rdev)
+{
+       /* RV410 and R420 can lock up if CP DMA to host memory happens
+        * while the 2D engine is busy.
+        *
+        * The proper workaround is to queue a RESYNC at the beginning
+        * of the CP init, apparently.
+        */
+       radeon_scratch_get(rdev, &rdev->config.r300.resync_scratch);
+       radeon_ring_lock(rdev, 8);
+       radeon_ring_write(rdev, PACKET0(R300_CP_RESYNC_ADDR, 1));
+       radeon_ring_write(rdev, rdev->config.r300.resync_scratch);
+       radeon_ring_write(rdev, 0xDEADBEEF);
+       radeon_ring_unlock_commit(rdev);
+}
+
+static void r420_cp_errata_fini(struct radeon_device *rdev)
+{
+       /* Catch the RESYNC we dispatched all the way back,
+        * at the very beginning of the CP init.
+        */
+       radeon_ring_lock(rdev, 8);
+       radeon_ring_write(rdev, PACKET0(R300_RB3D_DSTCACHE_CTLSTAT, 0));
+       radeon_ring_write(rdev, R300_RB3D_DC_FINISH);
+       radeon_ring_unlock_commit(rdev);
+       radeon_scratch_free(rdev, rdev->config.r300.resync_scratch);
+}
+
 static int r420_startup(struct radeon_device *rdev)
 {
        int r;
                dev_err(rdev->dev, "failled initializing CP (%d).\n", r);
                return r;
        }
+       r420_cp_errata_init(rdev);
        r = r100_wb_init(rdev);
        if (r) {
                dev_err(rdev->dev, "failled initializing WB (%d).\n", r);
 
 int r420_suspend(struct radeon_device *rdev)
 {
+       r420_cp_errata_fini(rdev);
        r100_cp_disable(rdev);
        r100_wb_disable(rdev);
        r100_irq_disable(rdev);