diff --git a/src/gallium/drivers/panfrost/pan_blitter.c b/src/gallium/drivers/panfrost/pan_blitter.c
index f41061ff769..53d8c3f7dec 100644
--- a/src/gallium/drivers/panfrost/pan_blitter.c
+++ b/src/gallium/drivers/panfrost/pan_blitter.c
@@ -34,6 +34,32 @@ panfrost_blitter_draw_rectangle(struct blitter_context *blitter,
    struct panfrost_context *pctx = pan_context(ctx);
    struct panfrost_screen *scr = pan_screen(ctx->screen);
 
+   /* u_blitter allows src == dst for disjoint texel sets without any texture
+    * barrier enforcement. Mali tile-based architecture can't guarantee that a
+    * read from the dst texture will fetch up-to-date values since it depends
+    * on the tile processing order. Request a fresh batch to ensure any writes
+    * to the resource are flushed. Doing so in this callback ensures that the
+    * framebuffer state is set for the current blit.
+    *
+    * XXX This should ideally be done at the draw call handling level when it
+    * requests a batch for the current framebuffer (see prepare_draw) by first
+    * submitting any batches writing to the draw call's sampler views. This
+    * check (along with resource accesses handling) is currently done when
+    * emitting texture descriptors but it explicitly discards the special case
+    * where the batch writing to the draw call's sampler views is the current
+    * batch because it can't be submitted at this time of the draw call
+    * handling (see panfrost_batch_update_access).
+    */
+   if (pctx->has_blit_loop) {
+      panfrost_get_fresh_batch_for_fbo(pctx, "Pre-blit flush");
+
+      /* XXX On all arch, Piglit's copy-pixels test sometimes fails to copy a
+       * block of pixels. Most easily reproducible at image size
+       * 50x38. Fallback to draw_vbo for now.
+       */
+      goto fallback;
+   }
+
    if (scr->dev.arch <= 8 || depth != 0.0f || num_instances > 1)
       goto fallback;
 
@@ -77,6 +103,10 @@ panfrost_blitter_draw_rectangle(struct blitter_context *blitter,
    /* Fallback to draw_vbo. */
    util_blitter_draw_rectangle(blitter, vertex_elements_cso, get_vs, x1, y1,
                                x2, y2, depth, num_instances, type, attrib);
+
+   /* XXX Depth/Stencil blits on v9 need that too. Not sure why. */
+   if (pctx->has_blit_loop && scr->dev.arch == 9)
+      panfrost_flush_all_batches(pctx, "Post-blit flush");
 }
 
 struct blitter_context *
@@ -157,7 +187,9 @@ panfrost_blitter_blit_legalized(struct pipe_context *pipe,
    }
 
    panfrost_blitter_save(ctx, states);
+   ctx->has_blit_loop = info->src.resource == info->dst.resource;
    util_blitter_blit(ctx->blitter, info, NULL);
+   ctx->has_blit_loop = false;
 }
 
 void
@@ -183,9 +215,7 @@ panfrost_blitter_blit(struct pipe_context *pipe,
                        util_format_linear(info->src.format), false, false);
    pan_legalize_format(ctx, pan_resource(info->dst.resource),
                        util_format_linear(info->dst.format), true, false);
-   panfrost_flush_all_batches(ctx, "Blit");
    panfrost_blitter_blit_legalized(pipe, info);
-   panfrost_flush_all_batches(ctx, "Blit");
 }
 
 /* Setup HW tile buffer clears if the batch for the current FBO doesn't have
diff --git a/src/gallium/drivers/panfrost/pan_context.h b/src/gallium/drivers/panfrost/pan_context.h
index 6470921f16b..91e3d89bf9c 100644
--- a/src/gallium/drivers/panfrost/pan_context.h
+++ b/src/gallium/drivers/panfrost/pan_context.h
@@ -192,6 +192,7 @@ struct panfrost_context {
    } texture_buffer[MESA_SHADER_STAGES];
 
    struct blitter_context *blitter;
+   bool has_blit_loop;
 
    struct pan_mod_convert_shaders mod_convert_shaders;