Merge branch 'dmabuf' into 'master'

Allow scanning out fullscreen surfaces on secondary GPUs

See merge request wlroots/wlroots!4055

backend/drm/backend.c

@@ -53,9 +53,7 @@ static void backend_destroy(struct wlr_backend *backend) {
 	wl_list_remove(&drm->dev_change.link);
 	wl_list_remove(&drm->dev_remove.link);
 
-	if (drm->parent) {
-		finish_drm_renderer(&drm->mgpu_renderer);
-	}
+	finish_drm_renderer(&drm->mgpu_renderer);
 
 	finish_drm_resources(drm);
 
@@ -210,22 +208,20 @@ struct wlr_backend *wlr_drm_backend_create(struct wlr_session *session,
 		goto error_event;
 	}
 
-	if (drm->parent) {
-		if (!init_drm_renderer(drm, &drm->mgpu_renderer)) {
-			wlr_log(WLR_ERROR, "Failed to initialize renderer");
-			goto error_resources;
-		}
-
-		// We'll perform a multi-GPU copy for all submitted buffers, we need
-		// to be able to texture from them
-		struct wlr_renderer *renderer = drm->mgpu_renderer.wlr_rend;
-		const struct wlr_drm_format_set *texture_formats =
-			wlr_renderer_get_dmabuf_texture_formats(renderer);
-		if (texture_formats == NULL) {
-			wlr_log(WLR_ERROR, "Failed to query renderer texture formats");
-			goto error_mgpu_renderer;
-		}
+	if (!init_drm_renderer(drm, &drm->mgpu_renderer)) {
+		wlr_log(WLR_ERROR, "Failed to initialize renderer");
+		goto error_resources;
+	}
 
+	// We'll perform a multi-GPU copy for all submitted buffers, we need
+	// to be able to texture from them
+	struct wlr_renderer *renderer = drm->mgpu_renderer.wlr_rend;
+	const struct wlr_drm_format_set *texture_formats =
+		wlr_renderer_get_dmabuf_texture_formats(renderer);
+	// Some configurations (alpine CI job) will have a renderer here that does not
+	// support dmabuf formats. We don't want to fail creation of the drm backend
+	// as a result of this, we simply don't populate the format set in that case.
+	if (texture_formats) {
 		// Forbid implicit modifiers, because their meaning changes from one
 		// GPU to another.
 		for (size_t i = 0; i < texture_formats->len; i++) {
@@ -245,8 +241,6 @@ struct wlr_backend *wlr_drm_backend_create(struct wlr_session *session,
 
 	return &drm->backend;
 
-error_mgpu_renderer:
-	finish_drm_renderer(&drm->mgpu_renderer);
 error_resources:
 	finish_drm_resources(drm);
 error_event:

backend/drm/drm.c

@@ -576,6 +576,7 @@ static void drm_connector_state_finish(struct wlr_drm_connector_state *state) {
 
 static bool drm_connector_state_update_primary_fb(struct wlr_drm_connector *conn,
 		struct wlr_drm_connector_state *state) {
+	bool ok;
 	struct wlr_drm_backend *drm = conn->backend;
 
 	assert(state->base->committed & WLR_OUTPUT_STATE_BUFFER);
@@ -585,34 +586,63 @@ static bool drm_connector_state_update_primary_fb(struct wlr_drm_connector *conn
 
 	struct wlr_drm_plane *plane = crtc->primary;
 	struct wlr_buffer *source_buf = state->base->buffer;
+	struct wlr_buffer *local_buf = wlr_buffer_lock(source_buf);
 
-	struct wlr_buffer *local_buf;
-	if (drm->parent) {
+	/*
+	 * First try to import the buffer. We can have a decent degree of
+	 * confidence this will work for a couple reasons:
+	 * 1. Apps running on the dGPU in PRIME setups will be submitting
+	 *    buffers with linear modifiers, so that they can be imported
+	 *    on the primary GPU. This means they are directly imporatable
+	 *    here as well. This gives a nice FPS boost.
+	 * 2. When the dGPU app supports reacting to dmabuf feedback it will
+	 *    be using dGPU modifiers, again meaning it can be imported into
+	 *    the dGPU directly for an additional nice perf boost.
+	 *
+	 * The fallback drm_surface_blit path will only be hit when the
+	 * app is running fullscreen with dGPU (non-linear) modifiers and
+	 * we start using rendered composition again. For a frame we will
+	 * do the fallback before the app reallocs its buffers back to
+	 * linear to be compatible with the primary GPU.
+	 */
+	ok = drm_fb_import(&state->primary_fb, drm, local_buf,
+		&crtc->primary->formats);
+
+	/*
+	 * If trying to import this buffer directly didn't work then try
+	 * to perform a blit to a mgpu drm surface and import that instead.
+	 */
+	if (!ok && drm->parent) {
 		struct wlr_drm_format format = {0};
 		if (!drm_plane_pick_render_format(plane, &format, &drm->mgpu_renderer)) {
 			wlr_log(WLR_ERROR, "Failed to pick primary plane format");
-			return false;
+			ok = false;
+			goto release_buf;
 		}
 
 		// TODO: fallback to modifier-less buffer allocation
-		bool ok = init_drm_surface(&plane->mgpu_surf, &drm->mgpu_renderer,
+		ok = init_drm_surface(&plane->mgpu_surf, &drm->mgpu_renderer,
 			source_buf->width, source_buf->height, &format);
 		wlr_drm_format_finish(&format);
 		if (!ok) {
-			return false;
+			ok = false;
+			goto release_buf;
 		}
 
-		local_buf = drm_surface_blit(&plane->mgpu_surf, source_buf);
-		if (local_buf == NULL) {
-			return false;
+		struct wlr_buffer *drm_buf = drm_surface_blit(&plane->mgpu_surf,
+				&drm->parent->mgpu_renderer, source_buf);
+		if (drm_buf == NULL) {
+			ok = false;
+			goto release_buf;
 		}
-	} else {
-		local_buf = wlr_buffer_lock(source_buf);
+		ok = drm_fb_import(&state->primary_fb, drm, drm_buf,
+				&plane->formats);
+		wlr_buffer_unlock(drm_buf);
 	}
 
-	bool ok = drm_fb_import(&state->primary_fb, drm, local_buf,
-		&plane->formats);
+release_buf:
 	wlr_buffer_unlock(local_buf);
+
 	if (!ok) {
 		wlr_drm_conn_log(conn, WLR_DEBUG,
 			"Failed to import buffer for scan-out");
@@ -1010,7 +1040,7 @@ static bool drm_connector_set_cursor(struct wlr_output *output,
 				return false;
 			}
 
-			local_buf = drm_surface_blit(&plane->mgpu_surf, buffer);
+			local_buf = drm_surface_blit(&plane->mgpu_surf, &drm->parent->mgpu_renderer, buffer);
 			if (local_buf == NULL) {
 				return false;
 			}

backend/drm/renderer.c

@@ -3,6 +3,7 @@
 #include <wlr/render/swapchain.h>
 #include <wlr/render/wlr_renderer.h>
 #include <wlr/util/log.h>
+#include "backend/backend.h"
 #include "backend/drm/drm.h"
 #include "backend/drm/fb.h"
 #include "backend/drm/renderer.h"
@@ -74,7 +75,7 @@ bool init_drm_surface(struct wlr_drm_surface *surf,
 }
 
 struct wlr_buffer *drm_surface_blit(struct wlr_drm_surface *surf,
-		struct wlr_buffer *buffer) {
+		struct wlr_drm_renderer *parent_renderer, struct wlr_buffer *buffer) {
 	struct wlr_renderer *renderer = surf->renderer->wlr_rend;
 
 	if (surf->swapchain->width != buffer->width ||
@@ -83,11 +84,23 @@ struct wlr_buffer *drm_surface_blit(struct wlr_drm_surface *surf,
 		return NULL;
 	}
 
-	struct wlr_texture *tex = wlr_texture_from_buffer(renderer, buffer);
-	if (tex == NULL) {
-		wlr_log(WLR_ERROR, "Failed to import source buffer into multi-GPU renderer");
+	struct wlr_texture_set *set = wlr_texture_set_create(renderer, NULL);
+	if (set == NULL) {
+		wlr_log(WLR_ERROR, "Failed to import source buffer multi-GPU texture set");
 		return NULL;
 	}
+	/* Add the parent renderer so the texture set can use it for copies */
+	wlr_texture_set_add_renderer(set, parent_renderer->wlr_rend, parent_renderer->allocator);
+	if (!wlr_texture_set_import_buffer(set, buffer)) {
+		wlr_log(WLR_ERROR, "Failed to import source buffer multi-GPU texture set");
+		goto error_tex;
+	}
+
+	struct wlr_texture *tex = wlr_texture_set_get_tex_for_renderer(set, renderer);
+	if (tex == NULL) {
+		wlr_log(WLR_ERROR, "Failed to export source buffer for multi-GPU renderer");
+		goto error_tex;
+	}
 
 	struct wlr_buffer *dst = wlr_swapchain_acquire(surf->swapchain, NULL);
 	if (!dst) {
@@ -110,14 +123,14 @@ struct wlr_buffer *drm_surface_blit(struct wlr_drm_surface *surf,
 		goto error_dst;
 	}
 
-	wlr_texture_destroy(tex);
+	wlr_texture_set_destroy(set);
 
 	return dst;
 
 error_dst:
 	wlr_buffer_unlock(dst);
 error_tex:
-	wlr_texture_destroy(tex);
+	wlr_texture_set_destroy(set);
 	return NULL;
 }
 

backend/multi/backend.c

@@ -2,11 +2,15 @@
 #include <stdbool.h>
 #include <stdlib.h>
 #include <time.h>
+#include <xf86drm.h>
+#include <fcntl.h>
 #include <wlr/backend/interface.h>
 #include <wlr/types/wlr_buffer.h>
 #include <wlr/util/log.h>
+#include "render/wlr_renderer.h"
 #include "backend/backend.h"
 #include "backend/multi.h"
+#include "render/allocator/allocator.h"
 
 struct subbackend_state {
 	struct wlr_backend *backend;
@@ -58,6 +62,7 @@ static void multi_backend_destroy(struct wlr_backend *wlr_backend) {
 			wl_container_of(backend->backends.next, sub, link);
 		wlr_backend_destroy(sub->backend);
 	}
+	wlr_multi_gpu_destroy(backend->multi_gpu);
 
 	free(backend);
 }
@@ -118,6 +123,7 @@ struct wlr_backend *wlr_multi_backend_create(struct wl_event_loop *loop) {
 	}
 
 	wl_list_init(&backend->backends);
+	backend->multi_gpu = wlr_multi_gpu_create();
 	wlr_backend_init(&backend->backend, &backend_impl);
 
 	wl_signal_init(&backend->events.backend_add);
@@ -225,3 +231,101 @@ void wlr_multi_for_each_backend(struct wlr_backend *_backend,
 		callback(sub->backend, data);
 	}
 }
+
+/*
+ * Create a wlr_multi_gpu struct and populate it with a renderer and allocator for each
+ * device in the system. This is done by finding all DRM nodes using drmGetDevices2.
+ */
+struct wlr_multi_gpu *wlr_multi_gpu_create(void) {
+	int flags = 0;
+	struct wlr_multi_gpu *multi_gpu = NULL;
+	int devices_len = drmGetDevices2(flags, NULL, 0);
+
+	if (devices_len < 0) {
+		wlr_log(WLR_ERROR, "drmGetDevices2 failed: %s", strerror(-devices_len));
+		return NULL;
+	}
+	drmDevice **devices = calloc(devices_len, sizeof(*devices));
+	if (devices == NULL) {
+		wlr_log_errno(WLR_ERROR, "Allocation failed");
+		goto out;
+	}
+	devices_len = drmGetDevices2(flags, devices, devices_len);
+	if (devices_len < 0) {
+		wlr_log(WLR_ERROR, "drmGetDevices2 failed: %s", strerror(-devices_len));
+		goto out;
+	}
+
+	multi_gpu = calloc(1, sizeof(struct wlr_multi_gpu));
+	if (!multi_gpu) {
+		goto out;
+	}
+	wl_list_init(&multi_gpu->devices);
+
+	for (int i = 0; i < devices_len; i++) {
+		drmDevice *dev = devices[i];
+		if (dev->available_nodes & (1 << DRM_NODE_RENDER)) {
+			const char *name = dev->nodes[DRM_NODE_RENDER];
+			wlr_log(WLR_DEBUG, "Opening DRM render node '%s'", name);
+			int fd = open(name, O_RDWR | O_CLOEXEC);
+			if (fd < 0) {
+				wlr_log_errno(WLR_ERROR, "Failed to open '%s'", name);
+				goto out;
+			}
+
+			// Create a renderer/allocator and add it as a new device
+			struct wlr_renderer *renderer = renderer_autocreate_with_drm_fd(fd);
+			if (!renderer) {
+				wlr_log(WLR_ERROR, "Failed to create multi-GPU renderer");
+				goto fail;
+			}
+
+			struct wlr_allocator *allocator =
+				allocator_autocreate_with_drm_fd(WLR_BUFFER_CAP_DMABUF, renderer, fd);
+			if (!allocator) {
+				wlr_log(WLR_ERROR, "Failed to create multi-GPU allocator");
+				wlr_renderer_destroy(renderer);
+				goto fail;
+			}
+
+			struct wlr_multi_gpu_device *device = calloc(1, sizeof(struct wlr_multi_gpu_device));
+			if (!device) {
+				wlr_allocator_destroy(allocator);
+				wlr_renderer_destroy(renderer);
+				goto fail;
+			}
+			wl_list_insert(&multi_gpu->devices, &device->link);
+			device->renderer = renderer;
+			device->allocator = allocator;
+		}
+	}
+
+	goto out;
+
+fail:
+	wlr_multi_gpu_destroy(multi_gpu);
+	multi_gpu = NULL;
+
+out:
+	for (int i = 0; i < devices_len; i++) {
+		drmFreeDevice(&devices[i]);
+	}
+	if (devices) {
+		free(devices);
+	}
+
+	return multi_gpu;
+}
+
+void wlr_multi_gpu_destroy(struct wlr_multi_gpu *multi_gpu) {
+	struct wlr_multi_gpu_device *device;
+	// Remove and destroy all devices
+	wl_list_for_each(device, &multi_gpu->devices, link) {
+		wlr_allocator_destroy(device->allocator);
+		wlr_renderer_destroy(device->renderer);
+		wl_list_remove(&device->link);
+		free(device);
+	}
+
+	free(multi_gpu);
+}

include/backend/drm/drm.h

@@ -107,6 +107,7 @@ struct wlr_drm_backend {
 
 	/* Only initialized on multi-GPU setups */
 	struct wlr_drm_renderer mgpu_renderer;
+	struct wlr_multi_gpu *multi_gpu;
 
 	struct wlr_session *session;
 

include/backend/drm/renderer.h

@@ -32,7 +32,7 @@ bool init_drm_surface(struct wlr_drm_surface *surf,
 void finish_drm_surface(struct wlr_drm_surface *surf);
 
 struct wlr_buffer *drm_surface_blit(struct wlr_drm_surface *surf,
-	struct wlr_buffer *buffer);
+	struct wlr_drm_renderer *parent_renderer, struct wlr_buffer *buffer);
 
 bool drm_plane_pick_render_format(struct wlr_drm_plane *plane,
 	struct wlr_drm_format *fmt, struct wlr_drm_renderer *renderer);

include/backend/multi.h

@@ -4,10 +4,31 @@
 #include <wayland-util.h>
 #include <wlr/backend/interface.h>
 #include <wlr/backend/multi.h>
+#include <wlr/render/allocator.h>
+#include <wlr/render/wlr_renderer.h>
+
+struct wlr_multi_gpu_device {
+    struct wlr_renderer *renderer;
+    struct wlr_allocator *allocator;
+    struct wl_list link;
+};
+
+/*
+ * Helper struct for tracking multiple renderers. This solves the
+ * problem of us having many renderers (primary, plus individual
+ * secondary GPU drm renderers) but not tracking them in one location.
+ * We can use this struct to access renderers for each GPU in
+ * the system all from one place. Will be populated by the renderer
+ * the compositor makes, plus every time a drm mgpu renderer is made.
+ */
+struct wlr_multi_gpu {
+	struct wl_list devices;
+};
 
 struct wlr_multi_backend {
 	struct wlr_backend backend;
 
+	struct wlr_multi_gpu *multi_gpu;
 	struct wl_list backends;
 
 	struct wl_listener event_loop_destroy;

include/wlr/backend/multi.h

@@ -32,4 +32,7 @@ bool wlr_multi_is_empty(struct wlr_backend *backend);
 void wlr_multi_for_each_backend(struct wlr_backend *backend,
 		void (*callback)(struct wlr_backend *backend, void *data), void *data);
 
+struct wlr_multi_gpu *wlr_multi_gpu_create(void);
+void wlr_multi_gpu_destroy(struct wlr_multi_gpu *multi_gpu);
+
 #endif

include/wlr/render/wlr_renderer.h

@@ -14,13 +14,10 @@
 #include <wlr/render/pass.h>
 #include <wlr/render/wlr_texture.h>
 #include <wlr/util/box.h>
+#include <wlr/backend/multi.h>
 
-struct wlr_backend;
 struct wlr_renderer_impl;
-struct wlr_drm_format_set;
 struct wlr_buffer;
-struct wlr_box;
-struct wlr_fbox;
 
 /**
  * A renderer for basic 2D operations.
@@ -39,6 +36,9 @@ struct wlr_renderer {
 	// private state
 
 	const struct wlr_renderer_impl *impl;
+
+	/* The GPU list we are a part of, may be null if not created from multi backend */
+	struct wlr_multi_gpu *multi_gpu;
 };
 
 /**

include/wlr/render/wlr_texture.h

@@ -18,6 +18,7 @@
 struct wlr_buffer;
 struct wlr_renderer;
 struct wlr_texture_impl;
+struct wlr_multi_gpu;
 
 struct wlr_texture {
 	const struct wlr_texture_impl *impl;
@@ -82,4 +83,123 @@ void wlr_texture_destroy(struct wlr_texture *texture);
 struct wlr_texture *wlr_texture_from_buffer(struct wlr_renderer *renderer,
 	struct wlr_buffer *buffer);
 
+struct wlr_texture_renderer_pair {
+	struct wlr_renderer *renderer;
+	struct wlr_texture *texture;
+	struct wlr_allocator *allocator;
+};
+
+/**
+ * The texture set provides a mapping between renderers and the texture
+ * imported into them. You can use it to query a texture for a particular
+ * renderer and it will handle importing and any blitting that needs to
+ * take place.
+ */
+struct wlr_texture_set {
+	/* The buffer this texture set was made from */
+	struct wlr_buffer *buffer;
+	struct wl_listener buffer_release;
+
+	/**
+	 * Index into pairings of the device that this texture directly
+	 * imports into. This texture is "native" to that device, and
+	 * will have to be blitted to other gpus.
+	 *
+	 * This will be -1 if no buffer has been imported yet.
+	 */
+	int32_t native_pair;
+	struct wlr_multi_gpu *multi_gpu;
+	/*
+	 * This will cache the result of creating a linear-layout version of
+	 * this texture on the native device. This can then be imported into
+	 * the other GPUs.
+	 */
+	uint32_t format;
+	void *pixel_data;
+
+	uint32_t width;
+	uint32_t height;
+
+	/* This is the size of the pairings array */
+	int pairing_count;
+	struct wlr_texture_renderer_pair *pairings;
+};
+
+/**
+ * Create an empty texture set. When setting up our wlr_multi_gpu struct we put
+ * all renderers into a list. This lets us iterate them from here. If this
+ * request is made on a renderer not in the multi-GPU set, then the list will
+ * be of length 1, and the renderer will be the only entry in the set.
+ *
+ * A buffer must be imported for this set to be used.
+ */
+struct wlr_texture_set *wlr_texture_set_create(struct wlr_renderer *renderer,
+	struct wlr_allocator *allocator);
+
+/**
+ * Add a renderer to the set. This adds an entry to the set tracking this renderer
+ * in the set's internal list. No texture is created for this renderer.
+ */
+void wlr_texture_set_add_renderer(struct wlr_texture_set *set, struct wlr_renderer *renderer,
+	struct wlr_allocator *allocator);
+
+/*
+ * Imports a buffer into the texture set. This initializes the native_pair
+ * internal state and returns true if the buffer was imported on at least one
+ * of the renderers in the set.
+ *
+ * This should only be called once per texture set initialization.
+ */
+bool wlr_texture_set_import_buffer(struct wlr_texture_set *set, struct wlr_buffer *buffer);
+
+/**
+ * Create a new texture set from a DMA-BUF. The returned texture is immutable.
+ * The dmabuf will be imported on only one of the mgpu renderers in the system,
+ * no copies will be made. Returns NULL if the dmabuf could not be imported into
+ * any renderer.
+ */
+struct wlr_texture_set *wlr_texture_set_from_dmabuf(struct wlr_renderer *renderer,
+	struct wlr_dmabuf_attributes *attribs);
+
+/**
+ * Create a new texture set from a buffer.
+ */
+struct wlr_texture_set *wlr_texture_set_from_buffer(struct wlr_renderer *renderer,
+	struct wlr_buffer *buffer);
+
+/**
+ * Request a wlr_texture for this resource that is compatible with the given
+ * renderer. This allows for on-demand cross-GPU blits in multi-GPU setups.
+ * The texture will have been imported into the renderer that corresponds to
+ * its native device. If a texture is requeseted with a different renderer,
+ * this function will perform a blit and return the appropriate texture.
+ *
+ * Textures are cached, so if multiple requests with a non-native renderer
+ * are made there will be only one blit.
+ */
+struct wlr_texture *wlr_texture_set_get_tex_for_renderer(struct wlr_texture_set *set,
+	struct wlr_renderer *renderer);
+
+/**
+ * Get the wlr_texture corresponding to the texture's local GPU. This is the GPU it
+ * is directly importable into.
+ */
+struct wlr_texture *wlr_texture_set_get_native_texture(struct wlr_texture_set *set);
+
+/**
+ * Get the linear pixel data for the backing texture.
+ */
+void *wlr_texture_set_get_linear_data(struct wlr_texture_set *set);
+
+/**
+  * Update all textures in a set with the contents of the next buffer. This will call
+  * wlr_texture_update_from_buffer for each texture in the set.
+  */
+bool wlr_texture_set_update_from_buffer(struct wlr_texture_set *set,
+		struct wlr_buffer *next, const pixman_region32_t *damage);
+
+/**
+ * Destroys the texture set and all textures held inside it.
+ */
+void wlr_texture_set_destroy(struct wlr_texture_set *set);
 #endif

include/wlr/types/wlr_buffer.h

@@ -142,7 +142,7 @@ struct wlr_client_buffer {
 	 * The buffer's texture, if any. A buffer will not have a texture if the
 	 * client destroys the buffer before it has been released.
 	 */
-	struct wlr_texture *texture;
+	struct wlr_texture_set *texture_set;
 	/**
 	 * The buffer this client buffer was created from. NULL if destroyed.
 	 */

include/wlr/types/wlr_linux_dmabuf_v1.h

@@ -63,6 +63,9 @@ struct wlr_linux_dmabuf_v1 {
 
 	int main_device_fd; // to sanity check FDs sent by clients, -1 if unavailable
 
+	// This is only set when the compositor isn't providing a custom renderer.
+	struct wlr_renderer *main_renderer;
+
 	struct wl_listener display_destroy;
 
 	bool (*check_dmabuf_callback)(struct wlr_dmabuf_attributes *attribs, void *data);

render/wlr_renderer.c

@@ -25,6 +25,7 @@
 #endif // WLR_HAS_VULKAN_RENDERER
 
 #include "backend/backend.h"
+#include "backend/multi.h"
 #include "render/pixel_format.h"
 #include "render/wlr_renderer.h"
 #include "util/env.h"
@@ -285,6 +286,13 @@ out:
 	if (own_drm_fd && drm_fd >= 0) {
 		close(drm_fd);
 	}
+	// If we have a multi GPU environment, then track this renderer
+	// for cross-GPU imports.
+	if (renderer && backend && wlr_backend_is_multi(backend)) {
+		struct wlr_multi_backend *multi = (struct wlr_multi_backend *)backend;
+		renderer->multi_gpu = multi->multi_gpu;
+	}
+
 	return renderer;
 }
 

render/wlr_texture.c

@@ -3,10 +3,17 @@
 #include <stdbool.h>
 #include <stdlib.h>
 #include <string.h>
+#include <drm_fourcc.h>
 #include <wlr/render/interface.h>
 #include <wlr/render/wlr_texture.h>
 #include "render/pixel_format.h"
+#include <wlr/types/wlr_matrix.h>
+#include <wlr/util/log.h>
 #include "types/wlr_buffer.h"
+#include "backend/multi.h"
+#include "backend/drm/drm.h"
+#include "render/drm_format_set.h"
+#include "render/wlr_renderer.h"
 
 void wlr_texture_init(struct wlr_texture *texture, struct wlr_renderer *renderer,
 		const struct wlr_texture_impl *impl, uint32_t width, uint32_t height) {
@@ -116,6 +123,26 @@ struct wlr_texture *wlr_texture_from_buffer(struct wlr_renderer *renderer,
 	if (!renderer->impl->texture_from_buffer) {
 		return NULL;
 	}
+
+	struct wlr_dmabuf_attributes dmabuf;
+	/*
+	 * If this is a dmabuf backed buffer then get the format/modifier for it and
+	 * compare it with the set supported by the renderer
+	 */
+	if (wlr_buffer_get_dmabuf(buffer, &dmabuf)) {
+		const struct wlr_drm_format_set *formats = wlr_renderer_get_dmabuf_texture_formats(renderer);
+		if (!formats) {
+			wlr_log(WLR_DEBUG, "Could not get DRM format set for renderer");
+			return NULL;
+		}
+
+		if (!wlr_drm_format_set_has(formats, dmabuf.format, dmabuf.modifier)) {
+			wlr_log(WLR_DEBUG, "Renderer could not import buffer with format 0x%x and modifier 0x%lx",
+					dmabuf.format, dmabuf.modifier);
+			return NULL;
+		}
+	}
+
 	return renderer->impl->texture_from_buffer(renderer, buffer);
 }
 
@@ -135,3 +162,362 @@ bool wlr_texture_update_from_buffer(struct wlr_texture *texture,
 	}
 	return texture->impl->update_from_buffer(texture, buffer, damage);
 }
+
+struct wlr_texture_set *wlr_texture_set_from_dmabuf(struct wlr_renderer *renderer,
+	struct wlr_dmabuf_attributes *attribs) {
+	struct wlr_dmabuf_buffer *buffer = dmabuf_buffer_create(attribs);
+	if (buffer == NULL) {
+		return NULL;
+	}
+
+	struct wlr_texture_set *set =
+		wlr_texture_set_from_buffer(renderer, &buffer->base);
+
+	// By this point, the renderer should have locked the buffer if it still
+	// needs to access it in the future.
+	dmabuf_buffer_drop(buffer);
+
+	return set;
+}
+
+static void texture_set_handle_buffer_release(struct wl_listener *listener, void *data) {
+	struct wlr_texture_set *set = wl_container_of(listener, set, buffer_release);
+	set->buffer = NULL;
+	wl_list_remove(&set->buffer_release.link);
+}
+
+static void wlr_texture_set_add_pair(struct wlr_texture_set *set, struct wlr_renderer *renderer,
+		struct wlr_allocator *allocator) {
+
+	set->pairings = realloc(set->pairings,
+			sizeof(struct wlr_texture_renderer_pair) * (set->pairing_count + 1));
+	if (!set->pairings) {
+		return;
+	}
+
+	memset(&set->pairings[set->pairing_count], 0, sizeof(struct wlr_texture_renderer_pair));
+	set->pairings[set->pairing_count].renderer = renderer;
+	set->pairings[set->pairing_count].allocator = allocator;
+	set->pairing_count++;
+}
+
+void wlr_texture_set_add_renderer(struct wlr_texture_set *set, struct wlr_renderer *renderer,
+		struct wlr_allocator *allocator) {
+	if (!renderer) {
+		return;
+	}
+
+	wlr_texture_set_add_pair(set, renderer, allocator);
+
+	if (renderer->multi_gpu) {
+		set->multi_gpu = renderer->multi_gpu;
+		/* Now add each mgpu renderer to the set */
+		struct wlr_multi_gpu_device *device;
+		wl_list_for_each(device, &renderer->multi_gpu->devices, link) {
+			wlr_texture_set_add_pair(set, device->renderer, device->allocator);
+		}
+	}
+}
+
+/*
+ * When setting up our wlr_multi_gpu struct we put all renderers into a list. This lets us
+ * iterate them from here. If this request is made on a renderer not in the multi-GPU set,
+ * then the list will be of length 1, and the renderer will be the only entry in the set.
+ */
+struct wlr_texture_set *wlr_texture_set_create(struct wlr_renderer *renderer,
+		struct wlr_allocator *allocator) {
+	struct wlr_texture_set *set = calloc(1, sizeof(struct wlr_texture_set));
+	if (!set) {
+		return NULL;
+	}
+	set->native_pair = -1;
+
+	wlr_texture_set_add_renderer(set, renderer, allocator);
+
+	return set;
+}
+
+/*
+ * Helper for importing a buffer into the texture set. This initializes
+ * the native_pair internal state.
+ */
+bool wlr_texture_set_import_buffer(struct wlr_texture_set *set, struct wlr_buffer *buffer) {
+	set->buffer = buffer;
+	// Don't lock our buffer since it gets in the way of releasing shm buffers immediately
+	// Instead keep a reference to the buffer but register a handler to notify us when
+	// it is released and clear the pointer.
+	set->buffer_release.notify = texture_set_handle_buffer_release;
+	wl_signal_add(&set->buffer->events.release, &set->buffer_release);
+
+	buffer = wlr_buffer_lock(buffer);
+	bool ret = false;
+
+	/*
+	 * For each renderer, try to create a texture. Go in order, since the first 
+	 * entry is always the "primary" renderer that the user created this texture set with.
+	 * The odds are highest that it is importable into that renderer, so start with that
+	 * one.
+	 */
+	for (int i = 0; i < set->pairing_count; i++) {
+		assert(!set->pairings[i].texture);
+		set->pairings[i].texture = wlr_texture_from_buffer(set->pairings[i].renderer, buffer);
+		/* If we got a match, mark this renderer as the "native" one the buffer is local to */
+		if (set->pairings[i].texture) {
+			/* Cache the width and height so other places don't have to search for it in pairings */
+			set->width = set->pairings[i].texture->width;
+			set->height = set->pairings[i].texture->height;
+			set->native_pair = i;
+			ret = true;
+			goto buffer_unlock;
+		}
+	}
+
+buffer_unlock:
+	wlr_buffer_unlock(buffer);
+	return ret;
+}
+
+struct wlr_texture_set *wlr_texture_set_from_buffer(struct wlr_renderer *renderer,
+		struct wlr_buffer *buffer) {
+	/* Get an empty texture set */
+	struct wlr_texture_set *set = wlr_texture_set_create(renderer, NULL);
+	if (!set) {
+		return NULL;
+	}
+
+	if (!wlr_texture_set_import_buffer(set, buffer)) {
+		goto fail;
+	}
+
+	return set;
+
+fail:
+	/* If the buffer couldn't be imported into any renderer in the system, return NULL */
+	wlr_texture_set_destroy(set);
+	return NULL;
+}
+
+static struct wlr_buffer *texture_set_blit_gpu_buffer(struct wlr_texture_set *set,
+	struct wlr_renderer *renderer) {
+	struct wlr_renderer *native_renderer = set->pairings[set->native_pair].renderer;
+	struct wlr_allocator *native_allocator = set->pairings[set->native_pair].allocator;
+	struct wlr_texture *native_texture = set->pairings[set->native_pair].texture;
+	assert(native_texture);
+
+	// If the user didn't give us an allocator for this renderer then this path can't be used.
+	if (!native_allocator) {
+		return NULL;
+	}
+
+	// Now intersect our DRM formats
+	const struct wlr_drm_format_set *src_formats = wlr_renderer_get_render_formats(native_renderer);
+	if (!src_formats) {
+		wlr_log(WLR_ERROR, "Failed to get primary renderer DRM formats");
+		return NULL;
+	}
+
+	const struct wlr_drm_format_set *dst_formats = wlr_renderer_get_dmabuf_texture_formats(renderer);
+	if (!dst_formats) {
+		wlr_log(WLR_ERROR, "Failed to get destination renderer DRM formats");
+		return NULL;
+	}
+
+	// Get the argb8 mods to use for our new buffer
+	struct wlr_drm_format argb_format = {0};
+	if (!wlr_drm_format_intersect(&argb_format,
+				wlr_drm_format_set_get(dst_formats, DRM_FORMAT_ARGB8888),
+				wlr_drm_format_set_get(src_formats, DRM_FORMAT_ARGB8888))
+			|| argb_format.len == 0) {
+		wlr_log(WLR_ERROR, "Failed to intersect DRM formats");
+		return NULL;
+	}
+
+	// Allocate a new buffer on the source renderer, we will blit the original texture
+	// to this and then return it so the caller can import it.
+	struct wlr_buffer *buffer = wlr_allocator_create_buffer(
+		native_allocator, set->width, set->height, &argb_format);
+	wlr_drm_format_finish(&argb_format);
+	if (!buffer) {
+		wlr_log(WLR_ERROR, "Failed to allocate buffer on source GPU");
+		return NULL;
+	}
+
+	struct wlr_render_pass *pass = wlr_renderer_begin_buffer_pass(native_renderer, buffer, NULL);
+	if (!pass) {
+		wlr_log(WLR_ERROR, "Failed to create a render pass");
+		goto drop_buffer;
+	}
+
+	wlr_render_pass_add_texture(pass, &(struct wlr_render_texture_options) {
+		.texture = native_texture,
+	});
+
+	if (!wlr_render_pass_submit(pass)) {
+		wlr_log(WLR_ERROR, "Failed to render to buffer");
+		goto drop_buffer;
+	}
+
+	return buffer;
+
+drop_buffer:
+	wlr_buffer_drop(buffer);
+	return NULL;
+}
+
+void *wlr_texture_set_get_linear_data(struct wlr_texture_set *set) {
+	struct wlr_renderer *native_renderer = set->pairings[set->native_pair].renderer;
+	struct wlr_texture *native_texture = set->pairings[set->native_pair].texture;
+	assert(native_texture);
+	int stride = native_texture->width * 4;
+
+	if (set->pixel_data) {
+		return set->pixel_data;
+	}
+
+	set->pixel_data = malloc(native_texture->height * stride);
+	if (!set->pixel_data) {
+		return NULL;
+	}
+
+	struct wlr_buffer *buffer = set->buffer;
+	if (!set->buffer) {
+		// If the buffer this set was created with has already been released, blit ourselves
+		// a new one.
+		buffer = texture_set_blit_gpu_buffer(set, native_renderer);
+		if (!buffer) {
+			wlr_log(WLR_DEBUG, "Cannot get linear data, wlr_texture_set's buffer was released");
+			return NULL;
+		}
+	}
+	wlr_buffer_lock(buffer);
+
+	/* Make a buffer with a linear layout and the same format */
+	set->format = wlr_texture_preferred_read_format(native_texture);
+    if (set->format == DRM_FORMAT_INVALID) {
+		wlr_buffer_unlock(buffer);
+		return NULL;
+	}
+
+	bool result = wlr_texture_read_pixels(native_texture, &(struct wlr_texture_read_pixels_options) {
+		.format = DRM_FORMAT_ARGB8888,
+		.stride = stride,
+		.data = set->pixel_data,
+	});
+	wlr_buffer_unlock(buffer);
+	if (!result) {
+		return NULL;
+	}
+
+	wlr_log(WLR_DEBUG, "Copied GPU vidmem buffer to linear sysmem buffer");
+	return set->pixel_data;
+}
+
+struct wlr_texture *wlr_texture_set_get_tex_for_renderer(struct wlr_texture_set *set,
+	struct wlr_renderer *renderer) {
+	/* Find the entry for this renderer */
+	struct wlr_texture_renderer_pair *pair = NULL;
+	for (int i = 0; i < set->pairing_count; i++) {
+		if (set->pairings[i].renderer == renderer) {
+			pair = &set->pairings[i];
+		}
+	}
+
+	/*
+	 * If we have not seen this renderer then add an entry for it so
+	 * we can cache the results of this copy.
+	 */
+	if (!pair) {
+		wlr_texture_set_add_pair(set, renderer, NULL);
+		pair = &set->pairings[set->pairing_count - 1];
+	}
+
+	/* If we already have a texture for this renderer, return it */
+	if (pair->texture) {
+        return pair->texture;
+	}
+
+	/*
+	 * First try to directly import the texture. We must have a valid buffer
+	 * to lock in order to do this. If the buffer has been released (as is the
+	 * case with shm buffers) then we will have to perform a fallback copy.
+	 */
+	if (set->buffer) {
+		wlr_buffer_lock(set->buffer);
+		pair->texture = wlr_texture_from_buffer(renderer, set->buffer);
+		wlr_buffer_unlock(set->buffer);
+		if (pair->texture) {
+	        return pair->texture;
+		}
+	}
+
+	/*
+	 * Directly importing didn't work. The next thing to try is blitting to a compatible
+	 * GPU texture and then importing that.
+	 */
+	struct wlr_buffer *buffer = texture_set_blit_gpu_buffer(set, renderer);
+	if (buffer) {
+		pair->texture = wlr_texture_from_buffer(renderer, buffer);
+		wlr_buffer_drop(buffer);
+		if (pair->texture) {
+	        return pair->texture;
+		}
+	}
+
+	/*
+	 * If the above didn't work then we can try a CPU fallback. This is much more expensive
+	 * but should always work. The reason we need this is that sometimes we have to copy
+	 * from GPU A to GPU B, but GPU A can't render to any modifiers that GPU B supports. This
+	 * happens on NVIDIA (among others) where you cannot render to a linear texture, but need
+	 * to convert to linear so that you can import it anywhere.
+	 *
+	 * Get our linear pixel data so we can import it into the target renderer.
+	 * */
+	void *pixel_data = wlr_texture_set_get_linear_data(set);
+	if (!pixel_data) {
+        return NULL;
+	}
+
+	/* import the linear texture into our renderer */
+	uint32_t stride = set->width * 4;
+	pair->texture = wlr_texture_from_pixels(renderer, DRM_FORMAT_ARGB8888, stride, set->width,
+			set->height, pixel_data);
+
+    return pair->texture;
+}
+
+struct wlr_texture *wlr_texture_set_get_native_texture(struct wlr_texture_set *set) {
+	return set->pairings[set->native_pair].texture;
+}
+
+bool wlr_texture_set_update_from_buffer(struct wlr_texture_set *set,
+		struct wlr_buffer *next, const pixman_region32_t *damage) {
+	/* Call wlr_texture_write_pixels on each valid texture in the set */
+	for (int i = 0; i < set->pairing_count; i++) {
+		if (set->pairings[i].texture) {
+			if (!wlr_texture_update_from_buffer(set->pairings[i].texture,
+						next, damage)) {
+				return false;
+			}
+		}
+	}
+
+	return true;
+}
+
+void wlr_texture_set_destroy(struct wlr_texture_set *set) {
+	if (set->buffer) {
+		wl_list_remove(&set->buffer_release.link);
+	}
+	free(set->pixel_data);
+
+	for (int i = 0; i < set->pairing_count; i++) {
+		if (set->pairings[i].texture) {
+			wlr_texture_destroy(set->pairings[i].texture);
+		}
+	}
+
+	if (set) {
+		free(set->pairings);
+		free(set);
+	}
+}

types/buffer/client.c

@@ -25,7 +25,7 @@ static struct wlr_client_buffer *client_buffer_from_buffer(
 static void client_buffer_destroy(struct wlr_buffer *buffer) {
 	struct wlr_client_buffer *client_buffer = client_buffer_from_buffer(buffer);
 	wl_list_remove(&client_buffer->source_destroy.link);
-	wlr_texture_destroy(client_buffer->texture);
+	wlr_texture_set_destroy(client_buffer->texture_set);
 	free(client_buffer);
 }
 
@@ -56,21 +56,21 @@ static void client_buffer_handle_source_destroy(struct wl_listener *listener,
 
 struct wlr_client_buffer *wlr_client_buffer_create(struct wlr_buffer *buffer,
 		struct wlr_renderer *renderer) {
-	struct wlr_texture *texture = wlr_texture_from_buffer(renderer, buffer);
-	if (texture == NULL) {
+	struct wlr_texture_set *texture_set = wlr_texture_set_from_buffer(renderer, buffer);
+	if (texture_set == NULL) {
 		wlr_log(WLR_ERROR, "Failed to create texture");
 		return NULL;
 	}
 
 	struct wlr_client_buffer *client_buffer = calloc(1, sizeof(*client_buffer));
 	if (client_buffer == NULL) {
-		wlr_texture_destroy(texture);
+		wlr_texture_set_destroy(texture_set);
 		return NULL;
 	}
 	wlr_buffer_init(&client_buffer->base, &client_buffer_impl,
-		texture->width, texture->height);
+		buffer->width, buffer->height);
 	client_buffer->source = buffer;
-	client_buffer->texture = texture;
+	client_buffer->texture_set = texture_set;
 
 	wl_signal_add(&buffer->events.destroy, &client_buffer->source_destroy);
 	client_buffer->source_destroy.notify = client_buffer_handle_source_destroy;
@@ -89,5 +89,5 @@ bool wlr_client_buffer_apply_damage(struct wlr_client_buffer *client_buffer,
 		return false;
 	}
 
-	return wlr_texture_update_from_buffer(client_buffer->texture, next, damage);
+	return wlr_texture_set_update_from_buffer(client_buffer->texture_set, next, damage);
 }

types/scene/wlr_scene.c

@@ -882,7 +882,8 @@ static struct wlr_texture *scene_buffer_get_texture(
 	struct wlr_client_buffer *client_buffer =
 		wlr_client_buffer_get(scene_buffer->buffer);
 	if (client_buffer != NULL) {
-		return client_buffer->texture;
+		return wlr_texture_set_get_tex_for_renderer(client_buffer->texture_set,
+				renderer);
 	}
 
 	scene_buffer->texture =

types/wlr_compositor.c

@@ -445,7 +445,11 @@ static void surface_apply_damage(struct wlr_surface *surface) {
 }
 
 static void surface_update_opaque_region(struct wlr_surface *surface) {
-	if (!wlr_surface_has_buffer(surface)) {
+	/*
+	 * The surface's client_buffer may not have a texture imported yet,
+	 * but if it has a texture set it is tracking a valid buffer.
+	 */
+	if (!wlr_surface_has_buffer(surface) || !surface->buffer->texture_set) {
 		pixman_region32_clear(&surface->opaque_region);
 		return;
 	}
@@ -802,7 +806,8 @@ struct wlr_texture *wlr_surface_get_texture(struct wlr_surface *surface) {
 	if (surface->buffer == NULL) {
 		return NULL;
 	}
-	return surface->buffer->texture;
+	return wlr_texture_set_get_tex_for_renderer(surface->buffer->texture_set,
+			surface->renderer);
 }
 
 bool wlr_surface_has_buffer(struct wlr_surface *surface) {

types/wlr_linux_dmabuf_v1.c

@@ -211,16 +211,39 @@ static bool check_import_dmabuf(struct wlr_dmabuf_attributes *attribs, void *dat
 		return true;
 	}
 
-	// TODO: check number of planes
-	for (int i = 0; i < attribs->n_planes; i++) {
-		uint32_t handle = 0;
-		if (drmPrimeFDToHandle(linux_dmabuf->main_device_fd, attribs->fd[i], &handle) != 0) {
-			wlr_log_errno(WLR_DEBUG, "Failed to import DMA-BUF FD");
+	/*
+	 * Some compositors will be using this linux dmabuf manager with custom renderers,
+	 * while others will use a wlroots-managed wlr_renderer. When checking if a dmabuf
+	 * is valid for import we should treat these differently. In the first case we just
+	 * need to check if the dmabuf is importable into the DRM device, in the wlroots-managed
+	 * renderer case we should check if this dmabuf can be imported into the renderer.
+	 *
+	 * In the case where we have a wlr_renderer we need to check if a texture set can
+	 * be created in order to handle multi-gpu systems. The texture set will handle ensuring
+	 * that the dmabuf is importable on one GPU in the system, instead of only checking
+	 * the main device.
+	 */
+	if (linux_dmabuf->main_renderer) {
+		struct wlr_texture_set *set=
+			wlr_texture_set_from_dmabuf(linux_dmabuf->main_renderer, attribs);
+		if (!set) {
 			return false;
 		}
-		if (drmCloseBufferHandle(linux_dmabuf->main_device_fd, handle) != 0) {
-			wlr_log_errno(WLR_ERROR, "Failed to close buffer handle");
-			return false;
+		// We can import the image, good. No need to keep it since wlr_surface will
+		// import it again on commit.
+		wlr_texture_set_destroy(set);
+	} else {
+		// TODO: check number of planes
+		for (int i = 0; i < attribs->n_planes; i++) {
+			uint32_t handle = 0;
+			if (drmPrimeFDToHandle(linux_dmabuf->main_device_fd, attribs->fd[i], &handle) != 0) {
+				wlr_log_errno(WLR_DEBUG, "Failed to import DMA-BUF FD");
+				return false;
+			}
+			if (drmCloseBufferHandle(linux_dmabuf->main_device_fd, handle) != 0) {
+				wlr_log_errno(WLR_ERROR, "Failed to close buffer handle");
+				return false;
+			}
 		}
 	}
 	return true;
@@ -1001,6 +1024,9 @@ struct wlr_linux_dmabuf_v1 *wlr_linux_dmabuf_v1_create_with_renderer(struct wl_d
 	struct wlr_linux_dmabuf_v1 *linux_dmabuf =
 		wlr_linux_dmabuf_v1_create(display, version, &feedback);
 	wlr_linux_dmabuf_feedback_v1_finish(&feedback);
+
+	linux_dmabuf->main_renderer = renderer;
+
 	return linux_dmabuf;
 }
 
@@ -1070,15 +1096,6 @@ static bool devid_from_fd(int fd, dev_t *devid) {
 	return true;
 }
 
-static bool is_secondary_drm_backend(struct wlr_backend *backend) {
-#if WLR_HAS_DRM_BACKEND
-	return wlr_backend_is_drm(backend) &&
-		wlr_drm_backend_get_parent(backend) != NULL;
-#else
-	return false;
-#endif
-}
-
 bool wlr_linux_dmabuf_feedback_v1_init_with_options(struct wlr_linux_dmabuf_feedback_v1 *feedback,
 		const struct wlr_linux_dmabuf_feedback_v1_init_options *options) {
 	assert(options->main_renderer != NULL);
@@ -1121,8 +1138,7 @@ bool wlr_linux_dmabuf_feedback_v1_init_with_options(struct wlr_linux_dmabuf_feed
 			wlr_log(WLR_ERROR, "Failed to intersect renderer and scanout formats");
 			goto error;
 		}
-	} else if (options->scanout_primary_output != NULL &&
-			!is_secondary_drm_backend(options->scanout_primary_output->backend)) {
+	} else if (options->scanout_primary_output != NULL) {
 		int backend_drm_fd = wlr_backend_get_drm_fd(options->scanout_primary_output->backend);
 		if (backend_drm_fd < 0) {
 			wlr_log(WLR_ERROR, "Failed to get backend DRM FD");
@@ -1148,8 +1164,9 @@ bool wlr_linux_dmabuf_feedback_v1_init_with_options(struct wlr_linux_dmabuf_feed
 
 		tranche->target_device = backend_dev;
 		tranche->flags = ZWP_LINUX_DMABUF_FEEDBACK_V1_TRANCHE_FLAGS_SCANOUT;
-		if (!wlr_drm_format_set_intersect(&tranche->formats, scanout_formats, renderer_formats)) {
-			wlr_log(WLR_ERROR, "Failed to intersect renderer and scanout formats");
+		// Copy our scanout formats to the scanout tranche
+		if (!wlr_drm_format_set_copy(&tranche->formats, scanout_formats)) {
+			wlr_log(WLR_ERROR, "Failed to copy scanout formats");
 			goto error;
 		}
 	}