freedreno: Fix missing rsc->seqno updates
[mesa.git] / src / gallium / drivers / freedreno / freedreno_resource.c
index 46deed39b8cf2a80c54f18e7a68076a1d87f20aa..803054c4f2fa0d569cf0d1bfc84c2974b073062c 100644 (file)
@@ -34,6 +34,8 @@
 #include "util/set.h"
 #include "util/u_drm.h"
 
+#include "decode/util.h"
+
 #include "freedreno_resource.h"
 #include "freedreno_batch_cache.h"
 #include "freedreno_blitter.h"
@@ -48,7 +50,7 @@
 #include <errno.h>
 
 /* XXX this should go away, needed for 'struct winsys_handle' */
-#include "state_tracker/drm_driver.h"
+#include "frontend/drm_driver.h"
 
 /* A private modifier for now, so we have a way to request tiled but not
  * compressed.  It would perhaps be good to get real modifiers for the
  * emitted so the GPU looks at the new backing bo.
  */
 static void
-rebind_resource(struct fd_context *ctx, struct pipe_resource *prsc)
+rebind_resource_in_ctx(struct fd_context *ctx, struct fd_resource *rsc)
 {
+       struct pipe_resource *prsc = &rsc->base;
+
+       if (ctx->rebind_resource)
+               ctx->rebind_resource(ctx, rsc);
+
        /* VBOs */
-       for (unsigned i = 0; i < ctx->vtx.vertexbuf.count && !(ctx->dirty & FD_DIRTY_VTXBUF); i++) {
-               if (ctx->vtx.vertexbuf.vb[i].buffer.resource == prsc)
-                       ctx->dirty |= FD_DIRTY_VTXBUF;
+       if (rsc->dirty & FD_DIRTY_VTXBUF) {
+               struct fd_vertexbuf_stateobj *vb = &ctx->vtx.vertexbuf;
+               for (unsigned i = 0; i < vb->count && !(ctx->dirty & FD_DIRTY_VTXBUF); i++) {
+                       if (vb->vb[i].buffer.resource == prsc)
+                               ctx->dirty |= FD_DIRTY_VTXBUF;
+               }
        }
 
+       const enum fd_dirty_3d_state per_stage_dirty =
+                       FD_DIRTY_CONST | FD_DIRTY_TEX | FD_DIRTY_IMAGE | FD_DIRTY_SSBO;
+
+       if (!(rsc->dirty & per_stage_dirty))
+               return;
+
        /* per-shader-stage resources: */
        for (unsigned stage = 0; stage < PIPE_SHADER_TYPES; stage++) {
                /* Constbufs.. note that constbuf[0] is normal uniforms emitted in
                 * cmdstream rather than by pointer..
                 */
-               const unsigned num_ubos = util_last_bit(ctx->constbuf[stage].enabled_mask);
-               for (unsigned i = 1; i < num_ubos; i++) {
-                       if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_CONST)
-                               break;
-                       if (ctx->constbuf[stage].cb[i].buffer == prsc) {
-                               ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_CONST;
-                               ctx->dirty |= FD_DIRTY_CONST;
+               if ((rsc->dirty & FD_DIRTY_CONST) &&
+                               !(ctx->dirty_shader[stage] & FD_DIRTY_CONST)) {
+                       struct fd_constbuf_stateobj *cb = &ctx->constbuf[stage];
+                       const unsigned num_ubos = util_last_bit(cb->enabled_mask);
+                       for (unsigned i = 1; i < num_ubos; i++) {
+                               if (cb->cb[i].buffer == prsc) {
+                                       ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_CONST;
+                                       ctx->dirty |= FD_DIRTY_CONST;
+                                       break;
+                               }
                        }
                }
 
                /* Textures */
-               for (unsigned i = 0; i < ctx->tex[stage].num_textures; i++) {
-                       if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_TEX)
-                               break;
-                       if (ctx->tex[stage].textures[i] && (ctx->tex[stage].textures[i]->texture == prsc)) {
-                               ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_TEX;
-                               ctx->dirty |= FD_DIRTY_TEX;
+               if ((rsc->dirty & FD_DIRTY_TEX) &&
+                               !(ctx->dirty_shader[stage] & FD_DIRTY_TEX)) {
+                       struct fd_texture_stateobj *tex = &ctx->tex[stage];
+                       for (unsigned i = 0; i < tex->num_textures; i++) {
+                               if (tex->textures[i] && (tex->textures[i]->texture == prsc)) {
+                                       ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_TEX;
+                                       ctx->dirty |= FD_DIRTY_TEX;
+                                       break;
+                               }
                        }
                }
 
                /* Images */
-               const unsigned num_images = util_last_bit(ctx->shaderimg[stage].enabled_mask);
-               for (unsigned i = 0; i < num_images; i++) {
-                       if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_IMAGE)
-                               break;
-                       if (ctx->shaderimg[stage].si[i].resource == prsc) {
-                               ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_IMAGE;
-                               ctx->dirty |= FD_DIRTY_IMAGE;
+               if ((rsc->dirty & FD_DIRTY_IMAGE) &&
+                               !(ctx->dirty_shader[stage] & FD_DIRTY_IMAGE)) {
+                       struct fd_shaderimg_stateobj *si = &ctx->shaderimg[stage];
+                       const unsigned num_images = util_last_bit(si->enabled_mask);
+                       for (unsigned i = 0; i < num_images; i++) {
+                               if (si->si[i].resource == prsc) {
+                                       ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_IMAGE;
+                                       ctx->dirty |= FD_DIRTY_IMAGE;
+                                       break;
+                               }
                        }
                }
 
                /* SSBOs */
-               const unsigned num_ssbos = util_last_bit(ctx->shaderbuf[stage].enabled_mask);
-               for (unsigned i = 0; i < num_ssbos; i++) {
-                       if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_SSBO)
-                               break;
-                       if (ctx->shaderbuf[stage].sb[i].buffer == prsc) {
-                               ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_SSBO;
-                               ctx->dirty |= FD_DIRTY_SSBO;
+               if ((rsc->dirty & FD_DIRTY_SSBO) &&
+                               !(ctx->dirty_shader[stage] & FD_DIRTY_SSBO)) {
+                       struct fd_shaderbuf_stateobj *sb = &ctx->shaderbuf[stage];
+                       const unsigned num_ssbos = util_last_bit(sb->enabled_mask);
+                       for (unsigned i = 0; i < num_ssbos; i++) {
+                               if (sb->sb[i].buffer == prsc) {
+                                       ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_SSBO;
+                                       ctx->dirty |= FD_DIRTY_SSBO;
+                                       break;
+                               }
                        }
                }
        }
 }
 
+static void
+rebind_resource(struct fd_resource *rsc)
+{
+       struct fd_screen *screen = fd_screen(rsc->base.screen);
+
+       fd_screen_lock(screen);
+       fd_resource_lock(rsc);
+
+       if (rsc->dirty)
+               list_for_each_entry (struct fd_context, ctx, &screen->context_list, node)
+                       rebind_resource_in_ctx(ctx, rsc);
+
+       fd_resource_unlock(rsc);
+       fd_screen_unlock(screen);
+}
+
+static inline void
+fd_resource_set_bo(struct fd_resource *rsc, struct fd_bo *bo)
+{
+       struct fd_screen *screen = fd_screen(rsc->base.screen);
+
+       rsc->bo = bo;
+       rsc->seqno = p_atomic_inc_return(&screen->rsc_seqno);
+}
+
 static void
 realloc_bo(struct fd_resource *rsc, uint32_t size)
 {
@@ -139,8 +192,9 @@ realloc_bo(struct fd_resource *rsc, uint32_t size)
        if (rsc->bo)
                fd_bo_del(rsc->bo);
 
-       rsc->bo = fd_bo_new(screen->dev, size, flags, "%ux%ux%u@%u:%x",
+       struct fd_bo *bo = fd_bo_new(screen->dev, size, flags, "%ux%ux%u@%u:%x",
                        prsc->width0, prsc->height0, prsc->depth0, rsc->layout.cpp, prsc->bind);
+       fd_resource_set_bo(rsc, bo);
 
        /* Zero out the UBWC area on allocation.  This fixes intermittent failures
         * with UBWC, which I suspect are due to the HW having a hard time
@@ -150,11 +204,9 @@ realloc_bo(struct fd_resource *rsc, uint32_t size)
         * around the issue, but any memset value seems to.
         */
        if (rsc->layout.ubwc) {
-               void *buf = fd_bo_map(rsc->bo);
-               memset(buf, 0, rsc->layout.slices[0].offset);
+               rsc->needs_ubwc_clear = true;
        }
 
-       rsc->seqno = p_atomic_inc_return(&screen->rsc_seqno);
        util_range_set_empty(&rsc->valid_buffer_range);
        fd_bc_invalidate_resource(rsc, true);
 }
@@ -174,6 +226,9 @@ do_blit(struct fd_context *ctx, const struct pipe_blit_info *blit, bool fallback
        }
 }
 
+static void
+flush_resource(struct fd_context *ctx, struct fd_resource *rsc, unsigned usage);
+
 /**
  * @rsc: the resource to shadow
  * @level: the level to discard (if box != NULL, otherwise ignored)
@@ -191,6 +246,21 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc,
        if (prsc->next)
                return false;
 
+       /* If you have a sequence where there is a single rsc associated
+        * with the current render target, and then you end up shadowing
+        * that same rsc on the 3d pipe (u_blitter), because of how we
+        * swap the new shadow and rsc before the back-blit, you could end
+        * up confusing things into thinking that u_blitter's framebuffer
+        * state is the same as the current framebuffer state, which has
+        * the result of blitting to rsc rather than shadow.
+        *
+        * Normally we wouldn't want to unconditionally trigger a flush,
+        * since that defeats the purpose of shadowing, but this is a
+        * case where we'd have to flush anyways.
+        */
+       if (rsc->write_batch == ctx->batch)
+               flush_resource(ctx, rsc, 0);
+
        /* TODO: somehow munge dimensions and format to copy unsupported
         * render target format to something that is supported?
         */
@@ -225,8 +295,9 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc,
         * should empty/destroy rsc->batches hashset)
         */
        fd_bc_invalidate_resource(rsc, false);
+       rebind_resource(rsc);
 
-       mtx_lock(&ctx->screen->lock);
+       fd_screen_lock(ctx->screen);
 
        /* Swap the backing bo's, so shadow becomes the old buffer,
         * blit from shadow to new buffer.  From here on out, we
@@ -260,7 +331,7 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc,
        }
        swap(rsc->batch_mask, shadow->batch_mask);
 
-       mtx_unlock(&ctx->screen->lock);
+       fd_screen_unlock(ctx->screen);
 
        struct pipe_blit_info blit = {};
        blit.dst.resource = prsc;
@@ -337,7 +408,7 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc,
  * Uncompress an UBWC compressed buffer "in place".  This works basically
  * like resource shadowing, creating a new resource, and doing an uncompress
  * blit, and swapping the state between shadow and original resource so it
- * appears to the state tracker as if nothing changed.
+ * appears to the gallium frontends as if nothing changed.
  */
 void
 fd_resource_uncompress(struct fd_context *ctx, struct fd_resource *rsc)
@@ -347,18 +418,17 @@ fd_resource_uncompress(struct fd_context *ctx, struct fd_resource *rsc)
 
        /* shadow should not fail in any cases where we need to uncompress: */
        debug_assert(success);
+}
 
-       /*
-        * TODO what if rsc is used in other contexts, we don't currently
-        * have a good way to rebind_resource() in other contexts.  And an
-        * app that is reading one resource in multiple contexts, isn't
-        * going to expect that the resource is modified.
-        *
-        * Hopefully the edge cases where we need to uncompress are rare
-        * enough that they mostly only show up in deqp.
-        */
-
-       rebind_resource(ctx, &rsc->base);
+/**
+ * Debug helper to hexdump a resource.
+ */
+void
+fd_resource_dump(struct fd_resource *rsc, const char *name)
+{
+       fd_bo_cpu_prep(rsc->bo, NULL, DRM_FREEDRENO_PREP_READ);
+       printf("%s: \n", name);
+       dump_hex(fd_bo_map(rsc->bo), fd_bo_size(rsc->bo));
 }
 
 static struct fd_resource *
@@ -450,9 +520,9 @@ flush_resource(struct fd_context *ctx, struct fd_resource *rsc, unsigned usage)
 {
        struct fd_batch *write_batch = NULL;
 
-       mtx_lock(&ctx->screen->lock);
+       fd_screen_lock(ctx->screen);
        fd_batch_reference_locked(&write_batch, rsc->write_batch);
-       mtx_unlock(&ctx->screen->lock);
+       fd_screen_unlock(ctx->screen);
 
        if (usage & PIPE_TRANSFER_WRITE) {
                struct fd_batch *batch, *batches[32] = {};
@@ -463,11 +533,11 @@ flush_resource(struct fd_context *ctx, struct fd_resource *rsc, unsigned usage)
                 * to iterate the batches which reference this resource.  So
                 * we must first grab references under a lock, then flush.
                 */
-               mtx_lock(&ctx->screen->lock);
+               fd_screen_lock(ctx->screen);
                batch_mask = rsc->batch_mask;
                foreach_batch(batch, &ctx->screen->batch_cache, batch_mask)
                        fd_batch_reference_locked(&batches[batch->idx], batch);
-               mtx_unlock(&ctx->screen->lock);
+               fd_screen_unlock(ctx->screen);
 
                foreach_batch(batch, &ctx->screen->batch_cache, batch_mask)
                        fd_batch_flush(batch);
@@ -526,7 +596,6 @@ fd_resource_transfer_map(struct pipe_context *pctx,
 {
        struct fd_context *ctx = fd_context(pctx);
        struct fd_resource *rsc = fd_resource(prsc);
-       struct fdl_slice *slice = fd_resource_slice(rsc, level);
        struct fd_transfer *trans;
        struct pipe_transfer *ptrans;
        enum pipe_format format = prsc->format;
@@ -538,6 +607,11 @@ fd_resource_transfer_map(struct pipe_context *pctx,
        DBG("prsc=%p, level=%u, usage=%x, box=%dx%d+%d,%d", prsc, level, usage,
                box->width, box->height, box->x, box->y);
 
+       if ((usage & PIPE_TRANSFER_MAP_DIRECTLY) && rsc->layout.tile_mode) {
+               DBG("CANNOT MAP DIRECTLY!\n");
+               return NULL;
+       }
+
        ptrans = slab_alloc(&ctx->transfer_pool);
        if (!ptrans)
                return NULL;
@@ -550,7 +624,7 @@ fd_resource_transfer_map(struct pipe_context *pctx,
        ptrans->level = level;
        ptrans->usage = usage;
        ptrans->box = *box;
-       ptrans->stride = slice->pitch;
+       ptrans->stride = fd_resource_pitch(rsc, level);
        ptrans->layer_stride = fd_resource_layer_stride(rsc, level);
 
        /* we always need a staging texture for tiled buffers:
@@ -564,11 +638,9 @@ fd_resource_transfer_map(struct pipe_context *pctx,
 
                staging_rsc = fd_alloc_staging(ctx, rsc, level, box);
                if (staging_rsc) {
-                       struct fdl_slice *staging_slice =
-                               fd_resource_slice(staging_rsc, 0);
                        // TODO for PIPE_TRANSFER_READ, need to do untiling blit..
                        trans->staging_prsc = &staging_rsc->base;
-                       trans->base.stride = staging_slice->pitch;
+                       trans->base.stride = fd_resource_pitch(staging_rsc, 0);
                        trans->base.layer_stride = fd_resource_layer_stride(staging_rsc, 0);
                        trans->staging_box = *box;
                        trans->staging_box.x = 0;
@@ -606,8 +678,8 @@ fd_resource_transfer_map(struct pipe_context *pctx,
 
        if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
                if (needs_flush || fd_resource_busy(rsc, op)) {
+                       rebind_resource(rsc);
                        realloc_bo(rsc, fd_bo_size(rsc->bo));
-                       rebind_resource(ctx, prsc);
                }
        } else if ((usage & PIPE_TRANSFER_WRITE) &&
                           prsc->target == PIPE_BUFFER &&
@@ -650,7 +722,6 @@ fd_resource_transfer_map(struct pipe_context *pctx,
                        if (needs_flush && fd_try_shadow_resource(ctx, rsc, level,
                                                        box, DRM_FORMAT_MOD_LINEAR)) {
                                needs_flush = busy = false;
-                               rebind_resource(ctx, prsc);
                                ctx->stats.shadow_uploads++;
                        } else {
                                struct fd_resource *staging_rsc;
@@ -667,10 +738,8 @@ fd_resource_transfer_map(struct pipe_context *pctx,
                                 */
                                staging_rsc = fd_alloc_staging(ctx, rsc, level, box);
                                if (staging_rsc) {
-                                       struct fdl_slice *staging_slice =
-                                               fd_resource_slice(staging_rsc, 0);
                                        trans->staging_prsc = &staging_rsc->base;
-                                       trans->base.stride = staging_slice->pitch;
+                                       trans->base.stride = fd_resource_pitch(staging_rsc, 0);
                                        trans->base.layer_stride =
                                                fd_resource_layer_stride(staging_rsc, 0);
                                        trans->staging_box = *box;
@@ -739,6 +808,7 @@ fd_resource_destroy(struct pipe_screen *pscreen,
                renderonly_scanout_destroy(rsc->scanout, fd_screen(pscreen)->ro);
 
        util_range_destroy(&rsc->valid_buffer_range);
+       simple_mtx_destroy(&rsc->lock);
        FREE(rsc);
 }
 
@@ -767,105 +837,7 @@ fd_resource_get_handle(struct pipe_screen *pscreen,
        handle->modifier = fd_resource_modifier(rsc);
 
        return fd_screen_bo_get_handle(pscreen, rsc->bo, rsc->scanout,
-                       fd_resource_slice(rsc, 0)->pitch, handle);
-}
-
-static uint32_t
-setup_slices(struct fd_resource *rsc, uint32_t alignment, enum pipe_format format)
-{
-       struct pipe_resource *prsc = &rsc->base;
-       struct fd_screen *screen = fd_screen(prsc->screen);
-       enum util_format_layout layout = util_format_description(format)->layout;
-       uint32_t pitchalign = screen->gmem_alignw;
-       uint32_t level, size = 0;
-       uint32_t width = prsc->width0;
-       uint32_t height = prsc->height0;
-       uint32_t depth = prsc->depth0;
-       /* in layer_first layout, the level (slice) contains just one
-        * layer (since in fact the layer contains the slices)
-        */
-       uint32_t layers_in_level = rsc->layout.layer_first ? 1 : prsc->array_size;
-
-       for (level = 0; level <= prsc->last_level; level++) {
-               struct fdl_slice *slice = fd_resource_slice(rsc, level);
-               uint32_t blocks;
-
-               if (layout == UTIL_FORMAT_LAYOUT_ASTC)
-                       width = util_align_npot(width, pitchalign * util_format_get_blockwidth(format));
-               else
-                       width = align(width, pitchalign);
-               slice->pitch = util_format_get_nblocksx(format, width) * rsc->layout.cpp;
-               slice->offset = size;
-               blocks = util_format_get_nblocks(format, width, height);
-               /* 1d array and 2d array textures must all have the same layer size
-                * for each miplevel on a3xx. 3d textures can have different layer
-                * sizes for high levels, but the hw auto-sizer is buggy (or at least
-                * different than what this code does), so as soon as the layer size
-                * range gets into range, we stop reducing it.
-                */
-               if (prsc->target == PIPE_TEXTURE_3D && (
-                                       level == 1 ||
-                                       (level > 1 && fd_resource_slice(rsc, level - 1)->size0 > 0xf000)))
-                       slice->size0 = align(blocks * rsc->layout.cpp, alignment);
-               else if (level == 0 || rsc->layout.layer_first || alignment == 1)
-                       slice->size0 = align(blocks * rsc->layout.cpp, alignment);
-               else
-                       slice->size0 = fd_resource_slice(rsc, level - 1)->size0;
-
-               size += slice->size0 * depth * layers_in_level;
-
-               width = u_minify(width, 1);
-               height = u_minify(height, 1);
-               depth = u_minify(depth, 1);
-       }
-
-       return size;
-}
-
-static uint32_t
-slice_alignment(enum pipe_texture_target target)
-{
-       /* on a3xx, 2d array and 3d textures seem to want their
-        * layers aligned to page boundaries:
-        */
-       switch (target) {
-       case PIPE_TEXTURE_3D:
-       case PIPE_TEXTURE_1D_ARRAY:
-       case PIPE_TEXTURE_2D_ARRAY:
-               return 4096;
-       default:
-               return 1;
-       }
-}
-
-/* cross generation texture layout to plug in to screen->setup_slices()..
- * replace with generation specific one as-needed.
- *
- * TODO for a4xx probably can extract out the a4xx specific logic int
- * a small fd4_setup_slices() wrapper that sets up layer_first, and then
- * calls this.
- */
-uint32_t
-fd_setup_slices(struct fd_resource *rsc)
-{
-       uint32_t alignment;
-
-       alignment = slice_alignment(rsc->base.target);
-
-       struct fd_screen *screen = fd_screen(rsc->base.screen);
-       if (is_a4xx(screen)) {
-               switch (rsc->base.target) {
-               case PIPE_TEXTURE_3D:
-                       rsc->layout.layer_first = false;
-                       break;
-               default:
-                       rsc->layout.layer_first = true;
-                       alignment = 1;
-                       break;
-               }
-       }
-
-       return setup_slices(rsc, alignment, rsc->base.format);
+                       fd_resource_pitch(rsc, 0), handle);
 }
 
 /* special case to resize query buf after allocated.. */
@@ -888,6 +860,8 @@ fd_resource_layout_init(struct pipe_resource *prsc)
        struct fd_resource *rsc = fd_resource(prsc);
        struct fdl_layout *layout = &rsc->layout;
 
+       layout->format = prsc->format;
+
        layout->width0 = prsc->width0;
        layout->height0 = prsc->height0;
        layout->depth0 = prsc->depth0;
@@ -898,12 +872,16 @@ fd_resource_layout_init(struct pipe_resource *prsc)
 }
 
 /**
- * Create a new texture object, using the given template info.
+ * Helper that allocates a resource and resolves its layout (but doesn't
+ * allocate its bo).
+ *
+ * It returns a pipe_resource (as fd_resource_create_with_modifiers()
+ * would do), and also bo's minimum required size as an output argument.
  */
 static struct pipe_resource *
-fd_resource_create_with_modifiers(struct pipe_screen *pscreen,
+fd_resource_allocate_and_resolve(struct pipe_screen *pscreen,
                const struct pipe_resource *tmpl,
-               const uint64_t *modifiers, int count)
+               const uint64_t *modifiers, int count, uint32_t *psize)
 {
        struct fd_screen *screen = fd_screen(pscreen);
        struct fd_resource *rsc;
@@ -911,38 +889,6 @@ fd_resource_create_with_modifiers(struct pipe_screen *pscreen,
        enum pipe_format format = tmpl->format;
        uint32_t size;
 
-       /* when using kmsro, scanout buffers are allocated on the display device
-        * create_with_modifiers() doesn't give us usage flags, so we have to
-        * assume that all calls with modifiers are scanout-possible
-        */
-       if (screen->ro &&
-               ((tmpl->bind & PIPE_BIND_SCANOUT) ||
-                !(count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID))) {
-               struct pipe_resource scanout_templat = *tmpl;
-               struct renderonly_scanout *scanout;
-               struct winsys_handle handle;
-
-               /* apply freedreno alignment requirement */
-               scanout_templat.width0 = align(tmpl->width0, screen->gmem_alignw);
-
-               scanout = renderonly_scanout_for_resource(&scanout_templat,
-                                                                                                 screen->ro, &handle);
-               if (!scanout)
-                       return NULL;
-
-               renderonly_scanout_destroy(scanout, screen->ro);
-
-               assert(handle.type == WINSYS_HANDLE_TYPE_FD);
-               rsc = fd_resource(pscreen->resource_from_handle(pscreen, tmpl,
-                                                                                                               &handle,
-                                                                                                               PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE));
-               close(handle.handle);
-               if (!rsc)
-                       return NULL;
-
-               return &rsc->base;
-       }
-
        rsc = CALLOC_STRUCT(fd_resource);
        prsc = &rsc->base;
 
@@ -997,6 +943,8 @@ fd_resource_create_with_modifiers(struct pipe_screen *pscreen,
 
        util_range_init(&rsc->valid_buffer_range);
 
+       simple_mtx_init(&rsc->lock, mtx_plain);
+
        rsc->internal_format = format;
 
        rsc->layout.ubwc = rsc->layout.tile_mode && is_a6xx(screen) && allow_ubwc;
@@ -1027,6 +975,63 @@ fd_resource_create_with_modifiers(struct pipe_screen *pscreen,
        if (fd_mesa_debug & FD_DBG_LAYOUT)
                fdl_dump_layout(&rsc->layout);
 
+       /* Hand out the resolved size. */
+       if (psize)
+               *psize = size;
+
+       return prsc;
+}
+
+/**
+ * Create a new texture object, using the given template info.
+ */
+static struct pipe_resource *
+fd_resource_create_with_modifiers(struct pipe_screen *pscreen,
+               const struct pipe_resource *tmpl,
+               const uint64_t *modifiers, int count)
+{
+       struct fd_screen *screen = fd_screen(pscreen);
+       struct fd_resource *rsc;
+       struct pipe_resource *prsc;
+       uint32_t size;
+
+       /* when using kmsro, scanout buffers are allocated on the display device
+        * create_with_modifiers() doesn't give us usage flags, so we have to
+        * assume that all calls with modifiers are scanout-possible
+        */
+       if (screen->ro &&
+               ((tmpl->bind & PIPE_BIND_SCANOUT) ||
+                !(count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID))) {
+               struct pipe_resource scanout_templat = *tmpl;
+               struct renderonly_scanout *scanout;
+               struct winsys_handle handle;
+
+               /* note: alignment is wrong for a6xx */
+               scanout_templat.width0 = align(tmpl->width0, screen->gmem_alignw);
+
+               scanout = renderonly_scanout_for_resource(&scanout_templat,
+                                                                                                 screen->ro, &handle);
+               if (!scanout)
+                       return NULL;
+
+               renderonly_scanout_destroy(scanout, screen->ro);
+
+               assert(handle.type == WINSYS_HANDLE_TYPE_FD);
+               rsc = fd_resource(pscreen->resource_from_handle(pscreen, tmpl,
+                                                                                                               &handle,
+                                                                                                               PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE));
+               close(handle.handle);
+               if (!rsc)
+                       return NULL;
+
+               return &rsc->base;
+       }
+
+       prsc = fd_resource_allocate_and_resolve(pscreen, tmpl, modifiers, count, &size);
+       if (!prsc)
+               return NULL;
+       rsc = fd_resource(prsc);
+
        realloc_bo(rsc, size);
        if (!rsc->bo)
                goto fail;
@@ -1059,7 +1064,6 @@ fd_resource_from_handle(struct pipe_screen *pscreen,
        struct fd_resource *rsc = CALLOC_STRUCT(fd_resource);
        struct fdl_slice *slice = fd_resource_slice(rsc, 0);
        struct pipe_resource *prsc = &rsc->base;
-       uint32_t pitchalign = fd_screen(pscreen)->gmem_alignw * rsc->layout.cpp;
 
        DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
                        "nr_samples=%u, usage=%u, bind=%x, flags=%x",
@@ -1080,17 +1084,36 @@ fd_resource_from_handle(struct pipe_screen *pscreen,
 
        util_range_init(&rsc->valid_buffer_range);
 
-       rsc->bo = fd_screen_bo_from_handle(pscreen, handle);
-       if (!rsc->bo)
+       simple_mtx_init(&rsc->lock, mtx_plain);
+
+       struct fd_bo *bo = fd_screen_bo_from_handle(pscreen, handle);
+       if (!bo)
                goto fail;
 
+       fd_resource_set_bo(rsc, bo);
+
        rsc->internal_format = tmpl->format;
-       slice->pitch = handle->stride;
+       rsc->layout.pitch0 = handle->stride;
        slice->offset = handle->offset;
        slice->size0 = handle->stride * prsc->height0;
 
-       if ((slice->pitch < align(prsc->width0 * rsc->layout.cpp, pitchalign)) ||
-                       (slice->pitch & (pitchalign - 1)))
+       /* use a pitchalign of gmem_alignw pixels, because GMEM resolve for
+        * lower alignments is not implemented (but possible for a6xx at least)
+        *
+        * for UBWC-enabled resources, layout_resource_for_modifier will further
+        * validate the pitch and set the right pitchalign
+        */
+       rsc->layout.pitchalign =
+               fdl_cpp_shift(&rsc->layout) + util_logbase2(screen->gmem_alignw);
+
+       /* apply the minimum pitchalign (note: actually 4 for a3xx but doesn't matter) */
+       if (is_a6xx(screen) || is_a5xx(screen))
+               rsc->layout.pitchalign = MAX2(rsc->layout.pitchalign, 6);
+       else
+               rsc->layout.pitchalign = MAX2(rsc->layout.pitchalign, 5);
+
+       if (rsc->layout.pitch0 < (prsc->width0 * rsc->layout.cpp) ||
+               fd_resource_pitch(rsc, 0) != rsc->layout.pitch0)
                goto fail;
 
        assert(rsc->layout.cpp);
@@ -1208,12 +1231,94 @@ fd_layout_resource_for_modifier(struct fd_resource *rsc, uint64_t modifier)
 {
        switch (modifier) {
        case DRM_FORMAT_MOD_LINEAR:
+               /* The dri gallium frontend will pass DRM_FORMAT_MOD_INVALID to us
+                * when it's called through any of the non-modifier BO create entry
+                * points.  Other drivers will determine tiling from the kernel or
+                * other legacy backchannels, but for freedreno it just means
+                * LINEAR. */
+       case DRM_FORMAT_MOD_INVALID:
                return 0;
        default:
                return -1;
        }
 }
 
+static struct pipe_resource *
+fd_resource_from_memobj(struct pipe_screen *pscreen,
+                                               const struct pipe_resource *tmpl,
+                                               struct pipe_memory_object *pmemobj,
+                                               uint64_t offset)
+{
+       struct fd_screen *screen = fd_screen(pscreen);
+       struct fd_memory_object *memobj = fd_memory_object(pmemobj);
+       struct pipe_resource *prsc;
+       struct fd_resource *rsc;
+       uint32_t size;
+       assert(memobj->bo);
+
+       /* We shouldn't get a scanout buffer here. */
+       assert(!(tmpl->bind & PIPE_BIND_SCANOUT));
+
+       uint64_t modifiers = DRM_FORMAT_MOD_INVALID;
+       if (tmpl->bind & PIPE_BIND_LINEAR) {
+               modifiers = DRM_FORMAT_MOD_LINEAR;
+       } else if (is_a6xx(screen) && tmpl->width0 >= FDL_MIN_UBWC_WIDTH) {
+               modifiers = DRM_FORMAT_MOD_QCOM_COMPRESSED;
+       }
+
+       /* Allocate new pipe resource. */
+       prsc = fd_resource_allocate_and_resolve(pscreen, tmpl, &modifiers, 1, &size);
+       if (!prsc)
+               return NULL;
+       rsc = fd_resource(prsc);
+
+       /* bo's size has to be large enough, otherwise cleanup resource and fail
+        * gracefully.
+        */
+       if (fd_bo_size(memobj->bo) < size) {
+               fd_resource_destroy(pscreen, prsc);
+               return NULL;
+       }
+
+       /* Share the bo with the memory object. */
+       fd_resource_set_bo(rsc, fd_bo_ref(memobj->bo));
+
+       return prsc;
+}
+
+static struct pipe_memory_object *
+fd_memobj_create_from_handle(struct pipe_screen *pscreen,
+                                                        struct winsys_handle *whandle,
+                                                        bool dedicated)
+{
+       struct fd_memory_object *memobj = CALLOC_STRUCT(fd_memory_object);
+       if (!memobj)
+               return NULL;
+
+       struct fd_bo *bo = fd_screen_bo_from_handle(pscreen, whandle);
+       if (!bo) {
+               free(memobj);
+               return NULL;
+       }
+
+       memobj->b.dedicated = dedicated;
+       memobj->bo = bo;
+
+       return &memobj->b;
+}
+
+static void
+fd_memobj_destroy(struct pipe_screen *pscreen,
+               struct pipe_memory_object *pmemobj)
+{
+       struct fd_memory_object *memobj = fd_memory_object(pmemobj);
+
+       assert(memobj->bo);
+       fd_bo_del(memobj->bo);
+
+       free(pmemobj);
+}
+
 void
 fd_resource_screen_init(struct pipe_screen *pscreen)
 {
@@ -1232,14 +1337,17 @@ fd_resource_screen_init(struct pipe_screen *pscreen)
        pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl,
                        true, false, fake_rgtc, true);
 
-       if (!screen->setup_slices)
-               screen->setup_slices = fd_setup_slices;
        if (!screen->layout_resource_for_modifier)
                screen->layout_resource_for_modifier = fd_layout_resource_for_modifier;
        if (!screen->supported_modifiers) {
                screen->supported_modifiers = supported_modifiers;
                screen->num_supported_modifiers = ARRAY_SIZE(supported_modifiers);
        }
+
+       /* GL_EXT_memory_object */
+       pscreen->memobj_create_from_handle = fd_memobj_create_from_handle;
+       pscreen->memobj_destroy = fd_memobj_destroy;
+       pscreen->resource_from_memobj = fd_resource_from_memobj;
 }
 
 static void