panfrost: Add backend targeting the DRM driver

This backend interacts with the new DRM driver for Midgard GPUs which is
currently in development.

When using this backend, Panfrost has roughly on-par functionality as
when using the non-DRM driver from Arm.

Alyssa Rosenzweig: To do so, we implement additional routines for
runtime GPU version detection and fencing. We cleanup some duplicate
code interfering with the new driver. We fix a long-standing memory leak
which is aggravated on the new driver. Finally, we implement BO
import/export in a way compatible with the new driver. These changes are
squashed to preserve bisectability given the hard-to-track ABI shifts in
the nondrm module

Signed-off-by: Tomeu Vizoso <[email protected]>
Reviewed-by: Alyssa Rosenzweig <[email protected]>

9 files changed, 607 insertions, 74 deletions

diff --git a/include/drm-uapi/panfrost_drm.h b/include/drm-uapi/panfrost_drm.h
new file mode 100644
index 00000000000..7618f14f9e2
--- /dev/null
+++ b/include/drm-uapi/panfrost_drm.h
@@ -0,0 +1,141 @@
+/* SPDX-License-Identifier: MIT */
+/*
+ * Copyright © 2014-2018 Broadcom
+ * Copyright © 2019 Collabora ltd.
+ */
+#ifndef _PANFROST_DRM_H_
+#define _PANFROST_DRM_H_
+
+#include "drm.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+#define DRM_PANFROST_SUBMIT			0x00
+#define DRM_PANFROST_WAIT_BO			0x01
+#define DRM_PANFROST_CREATE_BO			0x02
+#define DRM_PANFROST_MMAP_BO			0x03
+#define DRM_PANFROST_GET_PARAM			0x04
+#define DRM_PANFROST_GET_BO_OFFSET		0x05
+
+#define DRM_IOCTL_PANFROST_SUBMIT		DRM_IOWR(DRM_COMMAND_BASE + DRM_PANFROST_SUBMIT, struct drm_panfrost_submit)
+#define DRM_IOCTL_PANFROST_WAIT_BO		DRM_IOWR(DRM_COMMAND_BASE + DRM_PANFROST_WAIT_BO, struct drm_panfrost_wait_bo)
+#define DRM_IOCTL_PANFROST_CREATE_BO		DRM_IOWR(DRM_COMMAND_BASE + DRM_PANFROST_CREATE_BO, struct drm_panfrost_create_bo)
+#define DRM_IOCTL_PANFROST_MMAP_BO		DRM_IOWR(DRM_COMMAND_BASE + DRM_PANFROST_MMAP_BO, struct drm_panfrost_mmap_bo)
+#define DRM_IOCTL_PANFROST_GET_PARAM		DRM_IOWR(DRM_COMMAND_BASE + DRM_PANFROST_GET_PARAM, struct drm_panfrost_get_param)
+#define DRM_IOCTL_PANFROST_GET_BO_OFFSET	DRM_IOWR(DRM_COMMAND_BASE + DRM_PANFROST_GET_BO_OFFSET, struct drm_panfrost_get_bo_offset)
+
+#define PANFROST_JD_REQ_FS (1 << 0)
+
+/**
+ * struct drm_panfrost_submit - ioctl argument for submitting commands to the 3D
+ * engine.
+ *
+ * This asks the kernel to have the GPU execute a render command list.
+ */
+struct drm_panfrost_submit {
+
+	/** Address to GPU mapping of job descriptor */
+	__u64 jc;
+
+	/** An optional array of sync objects to wait on before starting this job. */
+	__u64 in_syncs;
+
+	/** Number of sync objects to wait on before starting this job. */
+	__u32 in_sync_count;
+
+	/** An optional sync object to place the completion fence in. */
+	__u32 out_sync;
+
+	/** Pointer to a u32 array of the BOs that are referenced by the job. */
+	__u64 bo_handles;
+
+	/** Number of BO handles passed in (size is that times 4). */
+	__u32 bo_handle_count;
+
+	/** A combination of PANFROST_JD_REQ_* */
+	__u32 requirements;
+};
+
+/**
+ * struct drm_panfrost_wait_bo - ioctl argument for waiting for
+ * completion of the last DRM_PANFROST_SUBMIT_CL on a BO.
+ *
+ * This is useful for cases where multiple processes might be
+ * rendering to a BO and you want to wait for all rendering to be
+ * completed.
+ */
+struct drm_panfrost_wait_bo {
+	__u32 handle;
+	__u32 pad;
+	__s64 timeout_ns;	/* absolute */
+};
+
+/**
+ * struct drm_panfrost_create_bo - ioctl argument for creating Panfrost BOs.
+ *
+ * There are currently no values for the flags argument, but it may be
+ * used in a future extension.
+ */
+struct drm_panfrost_create_bo {
+	__u32 size;
+	__u32 flags;
+	/** Returned GEM handle for the BO. */
+	__u32 handle;
+	/**
+	 * Returned offset for the BO in the GPU address space.  This offset
+	 * is private to the DRM fd and is valid for the lifetime of the GEM
+	 * handle.
+	 *
+	 * This offset value will always be nonzero, since various HW
+	 * units treat 0 specially.
+	 */
+	__u64 offset;
+};
+
+/**
+ * struct drm_panfrost_mmap_bo - ioctl argument for mapping Panfrost BOs.
+ *
+ * This doesn't actually perform an mmap.  Instead, it returns the
+ * offset you need to use in an mmap on the DRM device node.  This
+ * means that tools like valgrind end up knowing about the mapped
+ * memory.
+ *
+ * There are currently no values for the flags argument, but it may be
+ * used in a future extension.
+ */
+struct drm_panfrost_mmap_bo {
+	/** Handle for the object being mapped. */
+	__u32 handle;
+	__u32 flags;
+	/** offset into the drm node to use for subsequent mmap call. */
+	__u64 offset;
+};
+
+enum drm_panfrost_param {
+	DRM_PANFROST_PARAM_GPU_ID,
+};
+
+struct drm_panfrost_get_param {
+	__u32 param;
+	__u32 pad;
+	__u64 value;
+};
+
+/**
+ * Returns the offset for the BO in the GPU address space for this DRM fd.
+ * This is the same value returned by drm_panfrost_create_bo, if that was called
+ * from this DRM fd.
+ */
+struct drm_panfrost_get_bo_offset {
+	__u32 handle;
+	__u32 pad;
+	__u64 offset;
+};
+
+#if defined(__cplusplus)
+}
+#endif
+
+#endif /* _PANFROST_DRM_H_ */
diff --git a/src/gallium/drivers/panfrost/pan_allocate.c b/src/gallium/drivers/panfrost/pan_allocate.c
index 3732d253978..91ace74d0e4 100644
--- a/src/gallium/drivers/panfrost/pan_allocate.c
+++ b/src/gallium/drivers/panfrost/pan_allocate.c
@@ -41,7 +41,7 @@ panfrost_allocate_chunk(struct panfrost_context *ctx, size_t size, unsigned heap
         size = ALIGN(size, ALIGNMENT);
 
         struct pipe_context *gallium = (struct pipe_context *) ctx;
-        struct panfrost_screen *screen = panfrost_screen(gallium->screen);
+        struct panfrost_screen *screen = pan_screen(gallium->screen);
 
         struct pb_slab_entry *entry = pb_slab_alloc(&screen->slabs, size, heap_id);
         struct panfrost_memory_entry *p_entry = (struct panfrost_memory_entry *) entry;
@@ -81,7 +81,7 @@ panfrost_allocate_transient(struct panfrost_context *ctx, size_t sz)
                 if (pool->entry_index >= pool->entry_count) {
                         /* Don't overflow the pool -- allocate a new one */
                         struct pipe_context *gallium = (struct pipe_context *) ctx;
-                        struct panfrost_screen *screen = panfrost_screen(gallium->screen);
+                        struct panfrost_screen *screen = pan_screen(gallium->screen);
                         struct pb_slab_entry *entry = pb_slab_alloc(&screen->slabs, pool->entry_size, HEAP_TRANSIENT);
 
                         pool->entry_count++;
diff --git a/src/gallium/drivers/panfrost/pan_context.c b/src/gallium/drivers/panfrost/pan_context.c
index 4c41969fd05..3c8a483b8f5 100644
--- a/src/gallium/drivers/panfrost/pan_context.c
+++ b/src/gallium/drivers/panfrost/pan_context.c
@@ -57,26 +57,13 @@ extern const char *pan_counters_base;
 
 /* TODO: Sample size, etc */
 
-/* True for t6XX, false for t8xx. TODO: Run-time settable for automatic
- * hardware configuration. */
-
-static bool is_t6xx = false;
-
-/* If set, we'll require the use of single render-target framebuffer
- * descriptors (SFBD), for older hardware -- specifically, <T760 hardware, If
- * false, we'll use the MFBD no matter what. New hardware -does- retain support
- * for SFBD, and in theory we could flip between them on a per-RT basis, but
- * there's no real advantage to doing so */
-
-static bool require_sfbd = false;
-
 static void
 panfrost_set_framebuffer_msaa(struct panfrost_context *ctx, bool enabled)
 {
         SET_BIT(ctx->fragment_shader_core.unknown2_3, MALI_HAS_MSAA, enabled);
         SET_BIT(ctx->fragment_shader_core.unknown2_4, MALI_NO_MSAA, !enabled);
 
-        if (require_sfbd) {
+        if (ctx->require_sfbd) {
                 SET_BIT(ctx->fragment_sfbd.format, MALI_FRAMEBUFFER_MSAA_A | MALI_FRAMEBUFFER_MSAA_B, enabled);
         } else {
                 SET_BIT(ctx->fragment_rts[0].format.flags, MALI_MFBD_FORMAT_MSAA, enabled);
@@ -97,7 +84,7 @@ panfrost_set_framebuffer_msaa(struct panfrost_context *ctx, bool enabled)
 static void
 panfrost_enable_afbc(struct panfrost_context *ctx, struct panfrost_resource *rsrc, bool ds)
 {
-        if (require_sfbd) {
+        if (ctx->require_sfbd) {
                 printf("AFBC not supported yet on SFBD\n");
                 assert(0);
         }
@@ -120,6 +107,7 @@ panfrost_enable_afbc(struct panfrost_context *ctx, struct panfrost_resource *rsr
                                true, 0, 0, 0);
 
         rsrc->bo->has_afbc = true;
+        rsrc->bo->gem_handle = rsrc->bo->afbc_slab.gem_handle;
 
         /* Compressed textured reads use a tagged pointer to the metadata */
 
@@ -156,7 +144,7 @@ panfrost_set_fragment_afbc(struct panfrost_context *ctx)
                 if (!rsrc->bo->has_afbc)
                         continue;
 
-                if (require_sfbd) {
+                if (ctx->require_sfbd) {
                         fprintf(stderr, "Color AFBC not supported on SFBD\n");
                         assert(0);
                 }
@@ -180,7 +168,7 @@ panfrost_set_fragment_afbc(struct panfrost_context *ctx)
                 struct panfrost_resource *rsrc = (struct panfrost_resource *) ctx->pipe_framebuffer.zsbuf->texture;
 
                 if (rsrc->bo->has_afbc) {
-                        if (require_sfbd) {
+                        if (ctx->require_sfbd) {
                                 fprintf(stderr, "Depth AFBC not supported on SFBD\n");
                                 assert(0);
                         }
@@ -204,7 +192,7 @@ panfrost_set_fragment_afbc(struct panfrost_context *ctx)
         /* For the special case of a depth-only FBO, we need to attach a dummy render target */
 
         if (ctx->pipe_framebuffer.nr_cbufs == 0) {
-                if (require_sfbd) {
+                if (ctx->require_sfbd) {
                         fprintf(stderr, "Depth-only FBO not supported on SFBD\n");
                         assert(0);
                 }
@@ -364,7 +352,7 @@ panfrost_new_frag_framebuffer(struct panfrost_context *ctx)
                 stride = -stride;
         }
 
-        if (require_sfbd) {
+        if (ctx->require_sfbd) {
                 struct mali_single_framebuffer fb = panfrost_emit_sfbd(ctx);
 
                 fb.framebuffer = framebuffer;
@@ -562,7 +550,7 @@ panfrost_attach_vt_sfbd(struct panfrost_context *ctx)
 static void
 panfrost_attach_vt_framebuffer(struct panfrost_context *ctx)
 {
-        mali_ptr framebuffer = require_sfbd ?
+        mali_ptr framebuffer = ctx->require_sfbd ?
                 panfrost_attach_vt_sfbd(ctx) :
                 panfrost_attach_vt_mfbd(ctx);
 
@@ -616,7 +604,7 @@ panfrost_invalidate_frame(struct panfrost_context *ctx)
         if ((++ctx->cmdstream_i) == (sizeof(ctx->transient_pools) / sizeof(ctx->transient_pools[0])))
                 ctx->cmdstream_i = 0;
 
-        if (require_sfbd)
+        if (ctx->require_sfbd)
                 ctx->vt_framebuffer_sfbd = panfrost_emit_sfbd(ctx);
         else
                 ctx->vt_framebuffer_mfbd = panfrost_emit_mfbd(ctx);
@@ -655,7 +643,7 @@ panfrost_emit_vertex_payload(struct panfrost_context *ctx)
                         .workgroups_x_shift_2 = 0x2,
                         .workgroups_x_shift_3 = 0x5,
                 },
-		.gl_enables = 0x4 | (is_t6xx ? 0 : 0x2),
+		.gl_enables = 0x4 | (ctx->is_t6xx ? 0 : 0x2),
         };
 
         memcpy(&ctx->payload_vertex, &payload, sizeof(payload));
@@ -847,7 +835,7 @@ panfrost_default_shader_backend(struct panfrost_context *ctx)
                 .unknown2_4 = MALI_NO_MSAA | 0x4e0,
         };
 
-	if (is_t6xx) {
+	if (ctx->is_t6xx) {
                 shader.unknown2_4 |= 0x10;
 	}
 
@@ -957,7 +945,7 @@ panfrost_fragment_job(struct panfrost_context *ctx)
         /* Actualize the clear late; TODO: Fix order dependency between clear
          * and afbc */
 
-        if (require_sfbd) {
+        if (ctx->require_sfbd) {
                 panfrost_clear_sfbd(job);
         } else {
                 panfrost_clear_mfbd(job);
@@ -969,7 +957,7 @@ panfrost_fragment_job(struct panfrost_context *ctx)
                 struct panfrost_resource *rsrc = (struct panfrost_resource *) ctx->pipe_framebuffer.cbufs[0]->texture;
 
                 if (rsrc->bo->has_checksum) {
-                        if (require_sfbd) {
+                        if (ctx->require_sfbd) {
                                 fprintf(stderr, "Checksumming not supported on SFBD\n");
                                 assert(0);
                         }
@@ -986,11 +974,11 @@ panfrost_fragment_job(struct panfrost_context *ctx)
         /* The frame is complete and therefore the framebuffer descriptor is
          * ready for linkage and upload */
 
-        size_t sz = require_sfbd ? sizeof(struct mali_single_framebuffer) : (sizeof(struct bifrost_framebuffer) + sizeof(struct bifrost_fb_extra) + sizeof(struct bifrost_render_target) * 1);
+        size_t sz = ctx->require_sfbd ? sizeof(struct mali_single_framebuffer) : (sizeof(struct bifrost_framebuffer) + sizeof(struct bifrost_fb_extra) + sizeof(struct bifrost_render_target) * 1);
         struct panfrost_transfer fbd_t = panfrost_allocate_transient(ctx, sz);
         off_t offset = 0;
 
-        if (require_sfbd) {
+        if (ctx->require_sfbd) {
                 /* Upload just the SFBD all at once */
                 memcpy(fbd_t.cpu, &ctx->fragment_sfbd, sizeof(ctx->fragment_sfbd));
                 offset += sizeof(ctx->fragment_sfbd);
@@ -1022,10 +1010,10 @@ panfrost_fragment_job(struct panfrost_context *ctx)
         struct mali_payload_fragment payload = {
                 .min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(0, 0),
                 .max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(ctx->pipe_framebuffer.width, ctx->pipe_framebuffer.height),
-                .framebuffer = fbd_t.gpu | (require_sfbd ? MALI_SFBD : MALI_MFBD),
+                .framebuffer = fbd_t.gpu | (ctx->require_sfbd ? MALI_SFBD : MALI_MFBD),
         };
 
-	if (!require_sfbd && ctx->fragment_mfbd.unk3 & MALI_MFBD_EXTRA) {
+	if (!ctx->require_sfbd && ctx->fragment_mfbd.unk3 & MALI_MFBD_EXTRA) {
 		/* Signal that there is an extra portion of the framebuffer
 		 * descriptor */
 
@@ -1247,7 +1235,7 @@ panfrost_emit_for_draw(struct panfrost_context *ctx, bool with_vertex_data)
 			(ctx->blend->equation.alpha_mode == 0x122) &&
 			(ctx->blend->equation.color_mask == 0xf);
 
-                if (require_sfbd) {
+                if (ctx->require_sfbd) {
                         /* When only a single render target platform is used, the blend
                          * information is inside the shader meta itself. We
                          * additionally need to signal CAN_DISCARD for nontrivial blend
@@ -1270,7 +1258,7 @@ panfrost_emit_for_draw(struct panfrost_context *ctx, bool with_vertex_data)
 
                 ctx->payload_tiler.postfix._shader_upper = (transfer.gpu) >> 4;
 
-                if (!require_sfbd) {
+                if (!ctx->require_sfbd) {
                         /* Additional blend descriptor tacked on for jobs using MFBD */
 
                         unsigned blend_count = 0;
@@ -1360,7 +1348,7 @@ panfrost_emit_for_draw(struct panfrost_context *ctx, bool with_vertex_data)
                                 int s = ctx->sampler_views[t][i]->hw.nr_mipmap_levels;
 
                                 if (!rsrc->bo->is_mipmap) {
-                                        if (is_t6xx) {
+                                        if (ctx->is_t6xx) {
                                                 /* HW ERRATA, not needed after t6XX */
                                                 ctx->sampler_views[t][i]->hw.swizzled_bitmaps[1] = rsrc->bo->gpu[0];
 
@@ -1375,7 +1363,7 @@ panfrost_emit_for_draw(struct panfrost_context *ctx, bool with_vertex_data)
                                 /* Restore */
                                 ctx->sampler_views[t][i]->hw.nr_mipmap_levels = s;
 
-				if (is_t6xx) {
+				if (ctx->is_t6xx) {
 					ctx->sampler_views[t][i]->hw.unknown3A = 0;
 				}
                         }
@@ -1534,7 +1522,8 @@ panfrost_link_jobs(struct panfrost_context *ctx)
 /* The entire frame is in memory -- send it off to the kernel! */
 
 static void
-panfrost_submit_frame(struct panfrost_context *ctx, bool flush_immediate)
+panfrost_submit_frame(struct panfrost_context *ctx, bool flush_immediate,
+		      struct pipe_fence_handle **fence)
 {
         struct pipe_context *gallium = (struct pipe_context *) ctx;
         struct panfrost_screen *screen = pan_screen(gallium->screen);
@@ -1561,14 +1550,14 @@ panfrost_submit_frame(struct panfrost_context *ctx, bool flush_immediate)
         /* If visual, we can stall a frame */
 
         if (!flush_immediate)
-                screen->driver->force_flush_fragment(ctx);
+                screen->driver->force_flush_fragment(ctx, fence);
 
         screen->last_fragment_id = fragment_id;
         screen->last_fragment_flushed = false;
 
         /* If readback, flush now (hurts the pipelined performance) */
         if (flush_immediate)
-                screen->driver->force_flush_fragment(ctx);
+                screen->driver->force_flush_fragment(ctx, fence);
 
         if (screen->driver->dump_counters && pan_counters_base) {
                 screen->driver->dump_counters(screen);
@@ -1599,7 +1588,7 @@ panfrost_flush(
         bool flush_immediate = flags & PIPE_FLUSH_END_OF_FRAME;
 
         /* Submit the frame itself */
-        panfrost_submit_frame(ctx, flush_immediate);
+        panfrost_submit_frame(ctx, flush_immediate, fence);
 
         /* Prepare for the next frame */
         panfrost_invalidate_frame(ctx);
@@ -1827,12 +1816,13 @@ panfrost_create_rasterizer_state(
         struct pipe_context *pctx,
         const struct pipe_rasterizer_state *cso)
 {
+        struct panfrost_context *ctx = pan_context(pctx);
         struct panfrost_rasterizer *so = CALLOC_STRUCT(panfrost_rasterizer);
 
         so->base = *cso;
 
         /* Bitmask, unknown meaning of the start value */
-        so->tiler_gl_enables = is_t6xx ? 0x105 : 0x7;
+        so->tiler_gl_enables = ctx->is_t6xx ? 0x105 : 0x7;
 
         so->tiler_gl_enables |= MALI_FRONT_FACE(
                                         cso->front_ccw ? MALI_CCW : MALI_CW);
@@ -2341,7 +2331,7 @@ panfrost_set_framebuffer_state(struct pipe_context *pctx,
                 if (!cb)
                         continue;
 
-                if (require_sfbd)
+                if (ctx->require_sfbd)
                         ctx->vt_framebuffer_sfbd = panfrost_emit_sfbd(ctx);
                 else
                         ctx->vt_framebuffer_mfbd = panfrost_emit_mfbd(ctx);
@@ -2376,7 +2366,7 @@ panfrost_set_framebuffer_state(struct pipe_context *pctx,
                         if (zb) {
                                 /* FBO has depth */
 
-                                if (require_sfbd)
+                                if (ctx->require_sfbd)
                                         ctx->vt_framebuffer_sfbd = panfrost_emit_sfbd(ctx);
                                 else
                                         ctx->vt_framebuffer_mfbd = panfrost_emit_mfbd(ctx);
@@ -2596,9 +2586,16 @@ static void
 panfrost_destroy(struct pipe_context *pipe)
 {
         struct panfrost_context *panfrost = pan_context(pipe);
+        struct panfrost_screen *screen = pan_screen(pipe->screen);
 
         if (panfrost->blitter)
                 util_blitter_destroy(panfrost->blitter);
+
+        screen->driver->free_slab(screen, &panfrost->scratchpad);
+        screen->driver->free_slab(screen, &panfrost->varying_mem);
+        screen->driver->free_slab(screen, &panfrost->shaders);
+        screen->driver->free_slab(screen, &panfrost->tiler_heap);
+        screen->driver->free_slab(screen, &panfrost->misc_0);
 }
 
 static struct pipe_query *
@@ -2724,8 +2721,14 @@ struct pipe_context *
 panfrost_create_context(struct pipe_screen *screen, void *priv, unsigned flags)
 {
         struct panfrost_context *ctx = CALLOC_STRUCT(panfrost_context);
+        struct panfrost_screen *pscreen = pan_screen(screen);
         memset(ctx, 0, sizeof(*ctx));
         struct pipe_context *gallium = (struct pipe_context *) ctx;
+        unsigned gpu_id;
+
+        gpu_id = pscreen->driver->query_gpu_version(pscreen);
+        ctx->is_t6xx = gpu_id <= 0x0750; /* For now, this flag means t76x or less */
+        ctx->require_sfbd = gpu_id < 0x0750; /* t76x is the first to support MFD */
 
         gallium->screen = screen;
 
@@ -2792,6 +2795,8 @@ panfrost_create_context(struct pipe_screen *screen, void *priv, unsigned flags)
 
         panfrost_resource_context_init(gallium);
 
+        pscreen->driver->init_context(ctx);
+
         panfrost_setup_hardware(ctx);
 
         /* XXX: leaks */
diff --git a/src/gallium/drivers/panfrost/pan_context.h b/src/gallium/drivers/panfrost/pan_context.h
index b1a0a09146d..091d9988698 100644
--- a/src/gallium/drivers/panfrost/pan_context.h
+++ b/src/gallium/drivers/panfrost/pan_context.h
@@ -74,6 +74,11 @@ struct panfrost_query {
         struct panfrost_transfer transfer;
 };
 
+struct panfrost_fence {
+        struct pipe_reference reference;
+        int fd;
+};
+
 #define PANFROST_MAX_TRANSIENT_ENTRIES 64
 
 struct panfrost_transient_pool {
@@ -208,6 +213,18 @@ struct panfrost_context {
         struct pipe_blend_color blend_color;
         struct pipe_depth_stencil_alpha_state *depth_stencil;
         struct pipe_stencil_ref stencil_ref;
+
+        /* True for t6XX, false for t8xx. */
+        bool is_t6xx;
+
+        /* If set, we'll require the use of single render-target framebuffer
+         * descriptors (SFBD), for older hardware -- specifically, <T760 hardware, If
+         * false, we'll use the MFBD no matter what. New hardware -does- retain support
+         * for SFBD, and in theory we could flip between them on a per-RT basis, but
+         * there's no real advantage to doing so */
+        bool require_sfbd;
+
+	uint32_t out_sync;
 };
 
 /* Corresponds to the CSO */
diff --git a/src/gallium/drivers/panfrost/pan_drm.c b/src/gallium/drivers/panfrost/pan_drm.c
index cb01cce634f..a4018a1c74a 100644
--- a/src/gallium/drivers/panfrost/pan_drm.c
+++ b/src/gallium/drivers/panfrost/pan_drm.c
@@ -21,16 +21,365 @@
  * SOFTWARE.
  *
  */
+
+#include <fcntl.h>
+#include <xf86drm.h>
+
+#include "drm-uapi/panfrost_drm.h"
+
 #include "util/u_memory.h"
+#include "util/os_time.h"
 
 #include "pan_screen.h"
+#include "pan_resource.h"
+#include "pan_context.h"
 #include "pan_drm.h"
+#include "pan_trace.h"
 
 struct panfrost_drm {
 	struct panfrost_driver base;
 	int fd;
 };
 
+static void
+panfrost_drm_allocate_slab(struct panfrost_screen *screen,
+		           struct panfrost_memory *mem,
+		           size_t pages,
+		           bool same_va,
+		           int extra_flags,
+		           int commit_count,
+		           int extent)
+{
+	struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver;
+	struct drm_panfrost_create_bo create_bo = {
+		        .size = pages * 4096,
+		        .flags = 0,  // TODO figure out proper flags..
+	};
+	struct drm_panfrost_mmap_bo mmap_bo = {0,};
+	int ret;
+
+	// TODO cache allocations
+	// TODO properly handle errors
+	// TODO take into account extra_flags
+
+	ret = drmIoctl(drm->fd, DRM_IOCTL_PANFROST_CREATE_BO, &create_bo);
+	if (ret) {
+                fprintf(stderr, "DRM_IOCTL_PANFROST_CREATE_BO failed: %d\n", ret);
+		assert(0);
+	}
+
+	mem->gpu = create_bo.offset;
+	mem->gem_handle = create_bo.handle;
+        mem->stack_bottom = 0;
+        mem->size = create_bo.size;
+
+	// TODO map and unmap on demand?
+	mmap_bo.handle = create_bo.handle;
+	ret = drmIoctl(drm->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo);
+	if (ret) {
+                fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %d\n", ret);
+		assert(0);
+	}
+
+        mem->cpu = mmap(NULL, mem->size, PROT_READ | PROT_WRITE, MAP_SHARED,
+                       drm->fd, mmap_bo.offset);
+        if (mem->cpu == MAP_FAILED) {
+                fprintf(stderr, "mmap failed: %p\n", mem->cpu);
+		assert(0);
+	}
+
+        /* Record the mmap if we're tracing */
+        if (!(extra_flags & PAN_ALLOCATE_GROWABLE))
+                pantrace_mmap(mem->gpu, mem->cpu, mem->size, NULL);
+}
+
+static void
+panfrost_drm_free_slab(struct panfrost_screen *screen, struct panfrost_memory *mem)
+{
+	struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver;
+	struct drm_gem_close gem_close = {
+		.handle = mem->gem_handle,
+	};
+	int ret;
+
+        if (munmap((void *) (uintptr_t) mem->cpu, mem->size)) {
+                perror("munmap");
+                abort();
+        }
+
+	mem->cpu = NULL;
+
+	ret = drmIoctl(drm->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
+	if (ret) {
+                fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %d\n", ret);
+		assert(0);
+	}
+
+	mem->gem_handle = -1;
+}
+
+static struct panfrost_bo *
+panfrost_drm_import_bo(struct panfrost_screen *screen, struct winsys_handle *whandle)
+{
+	struct panfrost_bo *bo = CALLOC_STRUCT(panfrost_bo);
+	struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver;
+        struct drm_panfrost_get_bo_offset get_bo_offset = {0,};
+	struct drm_panfrost_mmap_bo mmap_bo = {0,};
+        int ret, size;
+        unsigned gem_handle;
+
+	ret = drmPrimeFDToHandle(drm->fd, whandle->handle, &gem_handle);
+	assert(!ret);
+
+	get_bo_offset.handle = gem_handle;
+        ret = drmIoctl(drm->fd, DRM_IOCTL_PANFROST_GET_BO_OFFSET, &get_bo_offset);
+        assert(!ret);
+
+	bo->gem_handle = gem_handle;
+        bo->gpu[0] = (mali_ptr) get_bo_offset.offset;
+
+	// TODO map and unmap on demand?
+	mmap_bo.handle = gem_handle;
+	ret = drmIoctl(drm->fd, DRM_IOCTL_PANFROST_MMAP_BO, &mmap_bo);
+	if (ret) {
+                fprintf(stderr, "DRM_IOCTL_PANFROST_MMAP_BO failed: %d\n", ret);
+		assert(0);
+	}
+
+	size = lseek(whandle->handle, 0, SEEK_END);
+	assert(size > 0);
+        bo->cpu[0] = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED,
+                       drm->fd, mmap_bo.offset);
+        if (bo->cpu[0] == MAP_FAILED) {
+                fprintf(stderr, "mmap failed: %p\n", bo->cpu[0]);
+		assert(0);
+	}
+
+        /* Record the mmap if we're tracing */
+        pantrace_mmap(bo->gpu[0], bo->cpu[0], size, NULL);
+
+        return bo;
+}
+
+static int
+panfrost_drm_export_bo(struct panfrost_screen *screen, int gem_handle, struct winsys_handle *whandle)
+{
+	struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver;
+        struct drm_prime_handle args = {
+                .handle = gem_handle,
+                .flags = DRM_CLOEXEC,
+        };
+
+        int ret = drmIoctl(drm->fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
+        if (ret == -1)
+                return FALSE;
+
+        whandle->handle = args.fd;
+
+        return TRUE;
+}
+
+static void
+panfrost_drm_free_imported_bo(struct panfrost_screen *screen, struct panfrost_bo *bo) 
+{
+	struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver;
+	struct drm_gem_close gem_close = {
+		.handle = bo->gem_handle,
+	};
+	int ret;
+
+	ret = drmIoctl(drm->fd, DRM_IOCTL_GEM_CLOSE, &gem_close);
+	if (ret) {
+                fprintf(stderr, "DRM_IOCTL_GEM_CLOSE failed: %d\n", ret);
+		assert(0);
+	}
+
+	bo->gem_handle = -1;
+	bo->gpu[0] = (mali_ptr)NULL;
+}
+
+static int
+panfrost_drm_submit_job(struct panfrost_context *ctx, u64 job_desc, int reqs, struct pipe_surface *surf)
+{
+        struct pipe_context *gallium = (struct pipe_context *) ctx;
+        struct panfrost_screen *screen = pan_screen(gallium->screen);
+	struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver;
+        struct drm_panfrost_submit submit = {0,};
+
+        submit.in_syncs = &ctx->out_sync;
+        submit.in_sync_count = 1;
+
+        submit.out_sync = ctx->out_sync;
+
+	submit.jc = job_desc;
+	submit.requirements = reqs;
+
+	if (surf) {
+		struct panfrost_resource *res = pan_resource(surf->texture);
+		submit.bo_handles = (u64) &res->bo->gem_handle;
+		submit.bo_handle_count = 1;
+	}
+
+        /* Dump memory _before_ submitting so we're not corrupted with actual GPU results */
+        pantrace_dump_memory();
+
+	if (drmIoctl(drm->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit)) {
+	        fprintf(stderr, "Error submitting: %m\n");
+	        return errno;
+	}
+
+        /* Trace the job if we're doing that and do a memory dump. We may
+         * want to adjust this logic once we're ready to trace FBOs */
+        pantrace_submit_job(submit.jc, submit.requirements, FALSE);
+
+	return 0;
+}
+
+static int
+panfrost_drm_submit_vs_fs_job(struct panfrost_context *ctx, bool has_draws, bool is_scanout)
+{
+        struct pipe_surface *surf = ctx->pipe_framebuffer.cbufs[0];
+	int ret;
+
+        if (has_draws) {
+		ret = panfrost_drm_submit_job(ctx, ctx->set_value_job, 0, NULL);
+		assert(!ret);
+	}
+
+	ret = panfrost_drm_submit_job(ctx, panfrost_fragment_job(ctx), PANFROST_JD_REQ_FS, surf);
+	assert(!ret);
+
+        return 0;
+}
+
+static struct panfrost_fence *
+panfrost_fence_create(struct panfrost_context *ctx)
+{
+        struct pipe_context *gallium = (struct pipe_context *) ctx;
+        struct panfrost_screen *screen = pan_screen(gallium->screen);
+	struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver;
+        struct panfrost_fence *f = calloc(1, sizeof(*f));
+        if (!f)
+                return NULL;
+
+        /* Snapshot the last Panfrost's rendering's out fence.  We'd rather have
+         * another syncobj instead of a sync file, but this is all we get.
+         * (HandleToFD/FDToHandle just gives you another syncobj ID for the
+         * same syncobj).
+         */
+        drmSyncobjExportSyncFile(drm->fd, ctx->out_sync, &f->fd);
+        if (f->fd == -1) {
+                fprintf(stderr, "export failed\n");
+                free(f);
+                return NULL;
+        }
+
+        pipe_reference_init(&f->reference, 1);
+
+        return f;
+}
+
+static void
+panfrost_drm_force_flush_fragment(struct panfrost_context *ctx,
+				  struct pipe_fence_handle **fence)
+{
+        struct pipe_context *gallium = (struct pipe_context *) ctx;
+
+        if (fence) {
+                struct panfrost_fence *f = panfrost_fence_create(ctx);
+                gallium->screen->fence_reference(gallium->screen, fence, NULL);
+                *fence = (struct pipe_fence_handle *)f;
+        }
+}
+
+static void
+panfrost_drm_enable_counters(struct panfrost_screen *screen)
+{
+	fprintf(stderr, "unimplemented: %s\n", __func__);
+}
+
+static void
+panfrost_drm_dump_counters(struct panfrost_screen *screen)
+{
+	fprintf(stderr, "unimplemented: %s\n", __func__);
+}
+
+static unsigned
+panfrost_drm_query_gpu_version(struct panfrost_screen *screen)
+{
+	struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver;
+        struct drm_panfrost_get_param get_param = {0,};
+        int ret;
+
+	get_param.param = DRM_PANFROST_PARAM_GPU_ID;
+        ret = drmIoctl(drm->fd, DRM_IOCTL_PANFROST_GET_PARAM, &get_param);
+        assert(!ret);
+
+	return get_param.value;
+}
+
+static int
+panfrost_drm_init_context(struct panfrost_context *ctx)
+{
+        struct pipe_context *gallium = (struct pipe_context *) ctx;
+        struct panfrost_screen *screen = pan_screen(gallium->screen);
+	struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver;
+
+        return drmSyncobjCreate(drm->fd, DRM_SYNCOBJ_CREATE_SIGNALED,
+                                &ctx->out_sync);
+}
+
+static void
+panfrost_drm_fence_reference(struct pipe_screen *screen,
+                         struct pipe_fence_handle **ptr,
+                         struct pipe_fence_handle *fence)
+{
+        struct panfrost_fence **p = (struct panfrost_fence **)ptr;
+        struct panfrost_fence *f = (struct panfrost_fence *)fence;
+        struct panfrost_fence *old = *p;
+
+        if (pipe_reference(&(*p)->reference, &f->reference)) {
+                close(old->fd);
+                free(old);
+        }
+        *p = f;
+}
+
+static boolean
+panfrost_drm_fence_finish(struct pipe_screen *pscreen,
+                      struct pipe_context *ctx,
+                      struct pipe_fence_handle *fence,
+                      uint64_t timeout)
+{
+        struct panfrost_screen *screen = pan_screen(pscreen);
+	struct panfrost_drm *drm = (struct panfrost_drm *)screen->driver;
+        struct panfrost_fence *f = (struct panfrost_fence *)fence;
+        int ret;
+
+        unsigned syncobj;
+        ret = drmSyncobjCreate(drm->fd, 0, &syncobj);
+        if (ret) {
+                fprintf(stderr, "Failed to create syncobj to wait on: %m\n");
+                return false;
+        }
+
+        drmSyncobjImportSyncFile(drm->fd, syncobj, f->fd);
+        if (ret) {
+                fprintf(stderr, "Failed to import fence to syncobj: %m\n");
+                return false;
+        }
+
+        uint64_t abs_timeout = os_time_get_absolute_timeout(timeout);
+        if (abs_timeout == OS_TIMEOUT_INFINITE)
+                abs_timeout = INT64_MAX;
+
+        ret = drmSyncobjWait(drm->fd, &syncobj, 1, abs_timeout, 0, NULL);
+
+        drmSyncobjDestroy(drm->fd, syncobj);
+
+        return ret >= 0;
+}
+
 struct panfrost_driver *
 panfrost_create_drm_driver(int fd)
 {
@@ -38,5 +387,19 @@ panfrost_create_drm_driver(int fd)
 
 	driver->fd = fd;
 
+	driver->base.import_bo = panfrost_drm_import_bo;
+	driver->base.export_bo = panfrost_drm_export_bo;
+	driver->base.free_imported_bo = panfrost_drm_free_imported_bo;
+	driver->base.submit_vs_fs_job = panfrost_drm_submit_vs_fs_job;
+	driver->base.force_flush_fragment = panfrost_drm_force_flush_fragment;
+	driver->base.allocate_slab = panfrost_drm_allocate_slab;
+	driver->base.free_slab = panfrost_drm_free_slab;
+	driver->base.enable_counters = panfrost_drm_enable_counters;
+	driver->base.query_gpu_version = panfrost_drm_query_gpu_version;
+	driver->base.init_context = panfrost_drm_init_context;
+	driver->base.fence_reference = panfrost_drm_fence_reference;
+	driver->base.fence_finish = panfrost_drm_fence_finish;
+	driver->base.dump_counters = panfrost_drm_dump_counters;
+
         return &driver->base;
 }
diff --git a/src/gallium/drivers/panfrost/pan_resource.c b/src/gallium/drivers/panfrost/pan_resource.c
index f26f33db96b..abef927ada5 100644
--- a/src/gallium/drivers/panfrost/pan_resource.c
+++ b/src/gallium/drivers/panfrost/pan_resource.c
@@ -68,6 +68,12 @@ panfrost_resource_from_handle(struct pipe_screen *pscreen,
 
 	rsc->bo = screen->driver->import_bo(screen, whandle);
 
+	if (screen->ro) {
+		rsc->scanout =
+			renderonly_create_gpu_import_for_resource(prsc, screen->ro, NULL);
+		/* failure is expected in some cases.. */
+	}
+
         return prsc;
 }
 
@@ -87,17 +93,15 @@ panfrost_resource_get_handle(struct pipe_screen *pscreen,
         handle->stride = stride;
         handle->modifier = DRM_FORMAT_MOD_INVALID;
 
-        if (handle->type == WINSYS_HANDLE_TYPE_SHARED) {
-                printf("Missed shared handle\n");
-                return FALSE;
-        } else if (handle->type == WINSYS_HANDLE_TYPE_KMS) {
-                if (renderonly_get_handle(scanout, handle)) {
-                        return TRUE;
-                } else {
-                        printf("Missed nonrenderonly KMS handle for resource %p with scanout %p\n", pt, scanout);
-                        return FALSE;
-                }
-        } else if (handle->type == WINSYS_HANDLE_TYPE_FD) {
+	if (handle->type == WINSYS_HANDLE_TYPE_SHARED) {
+		return FALSE;
+	} else if (handle->type == WINSYS_HANDLE_TYPE_KMS) {
+		if (renderonly_get_handle(scanout, handle))
+			return TRUE;
+
+		handle->handle = rsrc->bo->gem_handle;
+		return TRUE;
+	} else if (handle->type == WINSYS_HANDLE_TYPE_FD) {
                 if (scanout) {
                         struct drm_prime_handle args = {
                                 .handle = scanout->handle,
@@ -111,14 +115,11 @@ panfrost_resource_get_handle(struct pipe_screen *pscreen,
                         handle->handle = args.fd;
 
                         return TRUE;
-                } else {
-                        printf("Missed nonscanout FD handle\n");
-                        assert(0);
-                        return FALSE;
-                }
-        }
+                } else
+			return screen->driver->export_bo(screen, rsrc->bo->gem_handle, handle);
+	}
 
-        return FALSE;
+	return FALSE;
 }
 
 static void
@@ -322,7 +323,7 @@ static void
 panfrost_resource_destroy(struct pipe_screen *screen,
                           struct pipe_resource *pt)
 {
-        struct panfrost_screen *pscreen = panfrost_screen(screen);
+        struct panfrost_screen *pscreen = pan_screen(screen);
         struct panfrost_resource *rsrc = (struct panfrost_resource *) pt;
 
 	if (rsrc->scanout)
diff --git a/src/gallium/drivers/panfrost/pan_screen.c b/src/gallium/drivers/panfrost/pan_screen.c
index b8a119fd343..b7f2e4bb2e9 100644
--- a/src/gallium/drivers/panfrost/pan_screen.c
+++ b/src/gallium/drivers/panfrost/pan_screen.c
@@ -500,21 +500,22 @@ panfrost_get_timestamp(struct pipe_screen *_screen)
 }
 
 static void
-panfrost_fence_reference(struct pipe_screen *screen,
+panfrost_fence_reference(struct pipe_screen *pscreen,
                          struct pipe_fence_handle **ptr,
                          struct pipe_fence_handle *fence)
 {
-        *ptr = fence;
+        struct panfrost_screen *screen = pan_screen(pscreen);
+        screen->driver->fence_reference(pscreen, ptr, fence);
 }
 
 static boolean
-panfrost_fence_finish(struct pipe_screen *screen,
+panfrost_fence_finish(struct pipe_screen *pscreen,
                       struct pipe_context *ctx,
                       struct pipe_fence_handle *fence,
                       uint64_t timeout)
 {
-        assert(fence);
-        return TRUE;
+        struct panfrost_screen *screen = pan_screen(pscreen);
+        return screen->driver->fence_finish(pscreen, ctx, fence, timeout);
 }
 
 static const void *
diff --git a/src/gallium/drivers/panfrost/pan_screen.h b/src/gallium/drivers/panfrost/pan_screen.h
index 0005b2feb15..882611e93e3 100644
--- a/src/gallium/drivers/panfrost/pan_screen.h
+++ b/src/gallium/drivers/panfrost/pan_screen.h
@@ -49,9 +49,11 @@ struct panfrost_screen;
 
 struct panfrost_driver {
 	struct panfrost_bo * (*import_bo) (struct panfrost_screen *screen, struct winsys_handle *whandle);
+	int (*export_bo) (struct panfrost_screen *screen, int gem_handle, struct winsys_handle *whandle);
 
 	int (*submit_vs_fs_job) (struct panfrost_context *ctx, bool has_draws, bool is_scanout);
-	void (*force_flush_fragment) (struct panfrost_context *ctx);
+	void (*force_flush_fragment) (struct panfrost_context *ctx,
+				      struct pipe_fence_handle **fence);
 	void (*allocate_slab) (struct panfrost_screen *screen,
 		               struct panfrost_memory *mem,
 		               size_t pages,
@@ -65,6 +67,15 @@ struct panfrost_driver {
                              struct panfrost_bo *bo);
         void (*enable_counters) (struct panfrost_screen *screen);
         void (*dump_counters) (struct panfrost_screen *screen);
+	unsigned (*query_gpu_version) (struct panfrost_screen *screen);
+	int (*init_context) (struct panfrost_context *ctx);
+	void (*fence_reference) (struct pipe_screen *screen,
+                         struct pipe_fence_handle **ptr,
+                         struct pipe_fence_handle *fence);
+	boolean (*fence_finish) (struct pipe_screen *screen,
+                      struct pipe_context *ctx,
+                      struct pipe_fence_handle *fence,
+                      uint64_t timeout);
 };
 
 struct panfrost_screen {
@@ -85,10 +96,4 @@ struct panfrost_screen {
 	int last_fragment_flushed;
 };
 
-static inline struct panfrost_screen *
-panfrost_screen( struct pipe_screen *pipe )
-{
-        return (struct panfrost_screen *)pipe;
-}
-
 #endif /* PAN_SCREEN_H */
diff --git a/src/gallium/drivers/panfrost/pan_wallpaper.c b/src/gallium/drivers/panfrost/pan_wallpaper.c
index 735524a49c9..48418012fe8 100644
--- a/src/gallium/drivers/panfrost/pan_wallpaper.c
+++ b/src/gallium/drivers/panfrost/pan_wallpaper.c
@@ -181,7 +181,7 @@ panfrost_draw_wallpaper(struct pipe_context *pipe)
                 .normalized_coords = 1
         };
 
-        struct pipe_resource *rsrc = panfrost_screen(pipe->screen)->display_target;
+        struct pipe_resource *rsrc = pan_screen(pipe->screen)->display_target;
         struct pipe_sampler_state *sampler_state = pipe->create_sampler_state(pipe, &state);
         struct pipe_sampler_view *sampler_view = pipe->create_sampler_view(pipe, rsrc, &tmpl);