Merge tag 'drm-misc-next-2017-10-20' of git://anongit.freedesktop.org/drm/drm-misc into drm-next

Final drm-misc feature pull for 4.15:

UAPI Changes:
- new madvise ioctl for vc4 (Boris)

Core Changes:
- plane commit tracking fixes (Maarten)
- vgaarb improvements for fancy new platforms (aka ppc64 and arm64) by
Bjorn Helgaas

Driver Changes:
- pile of new panel drivers: Toshiba LT089AC19000, Innolux AT043TN24
- more sun4i work to support A10/A20 Tcon and hdmi outputs
- vc4: fix sleep in irq handler by making it threaded (Eric)
- udl probe/edid read fixes (Robert Tarasov)

And a bunch of misc small cleanups/refactors and doc fixes all over.

* tag 'drm-misc-next-2017-10-20' of git://anongit.freedesktop.org/drm/drm-misc: (32 commits)
drm/vc4: Fix sleeps during the IRQ handler for DSI transactions.
drm/vc4: Add the DRM_IOCTL_VC4_GEM_MADVISE ioctl
drm/panel: simple: add Toshiba LT089AC19000
dma-fence: remove duplicate word in comment
drm/panel: simple: add delays for Innolux AT043TN24
drm/panel: simple: add bus flags for Innolux AT043TN24
drm/panel: simple: fix vertical timings for Innolux AT043TN24
drm/atomic-helper: check that drivers call drm_crtc_vblank_off
drm: some KMS todo ideas
vgaarb: Factor out EFI and fallback default device selection
vgaarb: Select a default VGA device even if there's no legacy VGA
drm/bridge: adv7511: Fix a use after free
drm/sun4i: Add support for A20 display pipeline components
drm/sun4i: Add support for A10 display pipeline components
drm/sun4i: hdmi: Support HDMI controller on A10
drm/sun4i: tcon: Add support for A10 TCON
drm/sun4i: backend: Support output muxing
drm/sun4i: tcon: Move out the tcon0 common setup
drm/sun4i: tcon: Don't rely on encoders to set the TCON mode
drm/sun4i: tcon: Don't rely on encoders to enable the TCON
...

Documentation/devicetree/bindings/display/panel/toshiba,lt089ac29000.txt

···
+Toshiba 8.9" WXGA (1280x768) TFT LCD panel
+
+Required properties:
+- compatible: should be "toshiba,lt089ac29000.txt"
+- power-supply: as specified in the base binding
+
+This binding is compatible with the simple-panel binding, which is specified
+in simple-panel.txt in this directory.

Documentation/devicetree/bindings/display/sunxi/sun4i-drm.txt

···
 
 Required properties:
   - compatible: value must be one of:
+    * allwinner,sun4i-a10-hdmi
     * allwinner,sun5i-a10s-hdmi
     * allwinner,sun6i-a31-hdmi
   - reg: base address and size of memory-mapped region
···
 
 Required properties:
  - compatible: value must be either:
+   * allwinner,sun4i-a10-tcon
    * allwinner,sun5i-a13-tcon
    * allwinner,sun6i-a31-tcon
    * allwinner,sun6i-a31s-tcon
+   * allwinner,sun7i-a20-tcon
    * allwinner,sun8i-a33-tcon
    * allwinner,sun8i-v3s-tcon
  - reg: base address and size of memory-mapped region
···
 
 Required properties:
   - compatible: value must be one of:
+    * allwinner,sun4i-a10-display-backend
     * allwinner,sun5i-a13-display-backend
     * allwinner,sun6i-a31-display-backend
+    * allwinner,sun7i-a20-display-backend
     * allwinner,sun8i-a33-display-backend
   - reg: base address and size of the memory-mapped region.
   - interrupts: interrupt associated to this IP
···
 
 Required properties:
   - compatible: value must be one of:
+    * allwinner,sun4i-a10-display-frontend
     * allwinner,sun5i-a13-display-frontend
     * allwinner,sun6i-a31-display-frontend
+    * allwinner,sun7i-a20-display-frontend
     * allwinner,sun8i-a33-display-frontend
   - reg: base address and size of the memory-mapped region.
   - interrupts: interrupt associated to this IP
···
 
 Required properties:
   - compatible: value must be one of:
+    * allwinner,sun4i-a10-display-engine
     * allwinner,sun5i-a10s-display-engine
     * allwinner,sun5i-a13-display-engine
     * allwinner,sun6i-a31-display-engine
     * allwinner,sun6i-a31s-display-engine
+    * allwinner,sun7i-a20-display-engine
     * allwinner,sun8i-a33-display-engine
     * allwinner,sun8i-v3s-display-engine
 

Documentation/gpu/todo.rst

···
 
 Contact: Daniel Vetter
 
+KMS cleanups
+------------
+
+Some of these date from the very introduction of KMS in 2008 ...
+
+- drm_mode_config.crtc_idr is misnamed, since it contains all KMS object. Should
+  be renamed to drm_mode_config.object_idr.
+
+- drm_display_mode doesn't need to be derived from drm_mode_object. That's
+  leftovers from older (never merged into upstream) KMS designs where modes
+  where set using their ID, including support to add/remove modes.
+
 Better Testing
 ==============
 

arch/powerpc/kernel/pci-common.c

···
 }
 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_MOTOROLA, PCI_ANY_ID, fixup_hide_host_resource_fsl);
 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_FREESCALE, PCI_ANY_ID, fixup_hide_host_resource_fsl);
-
-static void fixup_vga(struct pci_dev *pdev)
-{
-	u16 cmd;
-
-	pci_read_config_word(pdev, PCI_COMMAND, &cmd);
-	if ((cmd & (PCI_COMMAND_IO | PCI_COMMAND_MEMORY)) || !vga_default_device())
-		vga_set_default_device(pdev);
-
-}
-DECLARE_PCI_FIXUP_CLASS_FINAL(PCI_ANY_ID, PCI_ANY_ID,
-			      PCI_CLASS_DISPLAY_VGA, 8, fixup_vga);

drivers/gpu/drm/armada/armada_crtc.c

···
 
 	if (force) {
 		/* Display is disabled, so just drop the old fb */
-		drm_framebuffer_unreference(fb);
+		drm_framebuffer_put(fb);
 		return;
 	}
 
···
 	 * the best.  The worst that will happen is the buffer gets
 	 * reused before it has finished being displayed.
 	 */
-	drm_framebuffer_unreference(fb);
+	drm_framebuffer_put(fb);
 }
 
 static void armada_drm_vblank_off(struct armada_crtc *dcrtc)
···
 	unsigned i;
 	bool interlaced;
 
-	drm_framebuffer_reference(crtc->primary->fb);
+	drm_framebuffer_get(crtc->primary->fb);
 
 	interlaced = !!(adj->flags & DRM_MODE_FLAG_INTERLACE);
 
···
 				   MAX_SCHEDULE_TIMEOUT);
 
 	/* Take a reference to the new fb as we're using it */
-	drm_framebuffer_reference(crtc->primary->fb);
+	drm_framebuffer_get(crtc->primary->fb);
 
 	/* Update the base in the CRTC */
 	armada_drm_crtc_update_regs(dcrtc, regs);
···
 	 * primary plane.
 	 */
 	if (plane->fb)
-		drm_framebuffer_unreference(plane->fb);
+		drm_framebuffer_put(plane->fb);
 
 	/* Power down the Y/U/V FIFOs */
 	sram_para1 = CFG_PDWN16x66 | CFG_PDWN32x66;
···
 
 		/* Must be a kernel-mapped object */
 		if (!obj->addr) {
-			drm_gem_object_unreference_unlocked(&obj->obj);
+			drm_gem_object_put_unlocked(&obj->obj);
 			return -EINVAL;
 		}
 
 		if (obj->obj.size < w * h * 4) {
 			DRM_ERROR("buffer is too small\n");
-			drm_gem_object_unreference_unlocked(&obj->obj);
+			drm_gem_object_put_unlocked(&obj->obj);
 			return -ENOMEM;
 		}
 	}
···
 	if (dcrtc->cursor_obj) {
 		dcrtc->cursor_obj->update = NULL;
 		dcrtc->cursor_obj->update_data = NULL;
-		drm_gem_object_unreference_unlocked(&dcrtc->cursor_obj->obj);
+		drm_gem_object_put_unlocked(&dcrtc->cursor_obj->obj);
 	}
 	dcrtc->cursor_obj = obj;
 	dcrtc->cursor_w = w;
···
 	struct armada_private *priv = crtc->dev->dev_private;
 
 	if (dcrtc->cursor_obj)
-		drm_gem_object_unreference_unlocked(&dcrtc->cursor_obj->obj);
+		drm_gem_object_put_unlocked(&dcrtc->cursor_obj->obj);
 
 	priv->dcrtc[dcrtc->num] = NULL;
 	drm_crtc_cleanup(&dcrtc->crtc);
···
 	 * Ensure that we hold a reference on the new framebuffer.
 	 * This has to match the behaviour in mode_set.
 	 */
-	drm_framebuffer_reference(fb);
+	drm_framebuffer_get(fb);
 
 	ret = armada_drm_crtc_queue_frame_work(dcrtc, work);
 	if (ret) {
 		/* Undo our reference above */
-		drm_framebuffer_unreference(fb);
+		drm_framebuffer_put(fb);
 		kfree(work);
 		return ret;
 	}

drivers/gpu/drm/armada/armada_drv.c

···
 	struct drm_framebuffer *fb;
 
 	while (kfifo_get(&priv->fb_unref, &fb))
-		drm_framebuffer_unreference(fb);
+		drm_framebuffer_put(fb);
 }
 
 /* Must be called with dev->event_lock held */

drivers/gpu/drm/armada/armada_fb.c

···
 	struct armada_framebuffer *dfb = drm_fb_to_armada_fb(fb);
 
 	drm_framebuffer_cleanup(&dfb->fb);
-	drm_gem_object_unreference_unlocked(&dfb->obj->obj);
+	drm_gem_object_put_unlocked(&dfb->obj->obj);
 	kfree(dfb);
 }
 
···
 	 * the above call, but the caller will drop their reference
 	 * to it.  Hence we need to take our own reference.
 	 */
-	drm_gem_object_reference(&obj->obj);
+	drm_gem_object_get(&obj->obj);
 
 	return dfb;
 }
···
 		goto err;
 	}
 
-	drm_gem_object_unreference_unlocked(&obj->obj);
+	drm_gem_object_put_unlocked(&obj->obj);
 
 	return &dfb->fb;
 
  err_unref:
-	drm_gem_object_unreference_unlocked(&obj->obj);
+	drm_gem_object_put_unlocked(&obj->obj);
  err:
 	DRM_ERROR("failed to initialize framebuffer: %d\n", ret);
 	return ERR_PTR(ret);

drivers/gpu/drm/armada/armada_fbdev.c

···
 
 	ret = armada_gem_linear_back(dev, obj);
 	if (ret) {
-		drm_gem_object_unreference_unlocked(&obj->obj);
+		drm_gem_object_put_unlocked(&obj->obj);
 		return ret;
 	}
 
 	ptr = armada_gem_map_object(dev, obj);
 	if (!ptr) {
-		drm_gem_object_unreference_unlocked(&obj->obj);
+		drm_gem_object_put_unlocked(&obj->obj);
 		return -ENOMEM;
 	}
 
···
 	 * A reference is now held by the framebuffer object if
 	 * successful, otherwise this drops the ref for the error path.
 	 */
-	drm_gem_object_unreference_unlocked(&obj->obj);
+	drm_gem_object_put_unlocked(&obj->obj);
 
 	if (IS_ERR(dfb))
 		return PTR_ERR(dfb);

drivers/gpu/drm/armada/armada_gem.c

···
 	/* drop reference from allocate - handle holds it now */
 	DRM_DEBUG_DRIVER("obj %p size %zu handle %#x\n", dobj, size, handle);
  err:
-	drm_gem_object_unreference_unlocked(&dobj->obj);
+	drm_gem_object_put_unlocked(&dobj->obj);
 	return ret;
 }
 
···
 	/* drop reference from allocate - handle holds it now */
 	DRM_DEBUG_DRIVER("obj %p size %zu handle %#x\n", dobj, size, handle);
  err:
-	drm_gem_object_unreference_unlocked(&dobj->obj);
+	drm_gem_object_put_unlocked(&dobj->obj);
 	return ret;
 }
 
···
 		return -ENOENT;
 
 	if (!dobj->obj.filp) {
-		drm_gem_object_unreference_unlocked(&dobj->obj);
+		drm_gem_object_put_unlocked(&dobj->obj);
 		return -EINVAL;
 	}
 
 	addr = vm_mmap(dobj->obj.filp, 0, args->size, PROT_READ | PROT_WRITE,
 		       MAP_SHARED, args->offset);
-	drm_gem_object_unreference_unlocked(&dobj->obj);
+	drm_gem_object_put_unlocked(&dobj->obj);
 	if (IS_ERR_VALUE(addr))
 		return addr;
 
···
 	}
 
  unref:
-	drm_gem_object_unreference_unlocked(&dobj->obj);
+	drm_gem_object_put_unlocked(&dobj->obj);
 	return ret;
 }
 
···
 			 * Importing our own dmabuf(s) increases the
 			 * refcount on the gem object itself.
 			 */
-			drm_gem_object_reference(obj);
+			drm_gem_object_get(obj);
 			return obj;
 		}
 	}

drivers/gpu/drm/armada/armada_overlay.c

···
 		 * Take a reference on the new framebuffer - we want to
 		 * hold on to it while the hardware is displaying it.
 		 */
-		drm_framebuffer_reference(fb);
+		drm_framebuffer_get(fb);
 
 		if (plane->fb)
 			armada_ovl_retire_fb(dplane, plane->fb);
···
 
 	fb = xchg(&dplane->old_fb, NULL);
 	if (fb)
-		drm_framebuffer_unreference(fb);
+		drm_framebuffer_put(fb);
 
 	return 0;
 }

drivers/gpu/drm/bridge/adv7511/adv7511_drv.c

···
 	adv7511_set_config_csc(adv7511, connector, adv7511->rgb,
 			       drm_detect_hdmi_monitor(edid));
 
-	kfree(edid);
-
 	cec_s_phys_addr_from_edid(adv7511->cec_adap, edid);
+
+	kfree(edid);
 
 	return count;
 }

drivers/gpu/drm/drm_atomic_helper.c

···
 
 	for_each_oldnew_crtc_in_state(old_state, crtc, old_crtc_state, new_crtc_state, i) {
 		const struct drm_crtc_helper_funcs *funcs;
+		int ret;
 
 		/* Shut down everything that needs a full modeset. */
 		if (!drm_atomic_crtc_needs_modeset(new_crtc_state))
···
 			funcs->disable(crtc);
 		else
 			funcs->dpms(crtc, DRM_MODE_DPMS_OFF);
+
+		if (!(dev->irq_enabled && dev->num_crtcs))
+			continue;
+
+		ret = drm_crtc_vblank_get(crtc);
+		WARN_ONCE(ret != -EINVAL, "driver forgot to call drm_crtc_vblank_off()\n");
+		if (ret == 0)
+			drm_crtc_vblank_put(crtc);
 	}
 }
 
···
 	}
 
 	for_each_oldnew_connector_in_state(state, conn, old_conn_state, new_conn_state, i) {
-		/* commit tracked through new_crtc_state->commit, no need to do it explicitly */
-		if (new_conn_state->crtc)
-			continue;
-
 		/* Userspace is not allowed to get ahead of the previous
 		 * commit with nonblocking ones. */
 		if (nonblock && old_conn_state->commit &&
 		    !try_wait_for_completion(&old_conn_state->commit->flip_done))
 			return -EBUSY;
+
+		/* commit tracked through new_crtc_state->commit, no need to do it explicitly */
+		if (new_conn_state->crtc)
+			continue;
 
 		commit = crtc_or_fake_commit(state, old_conn_state->crtc);
 		if (!commit)
···
 	}
 
 	for_each_oldnew_plane_in_state(state, plane, old_plane_state, new_plane_state, i) {
-		/*
-		 * Unlike connectors, always track planes explicitly for
-		 * async pageflip support.
-		 */
-
 		/* Userspace is not allowed to get ahead of the previous
 		 * commit with nonblocking ones. */
 		if (nonblock && old_plane_state->commit &&
 		    !try_wait_for_completion(&old_plane_state->commit->flip_done))
 			return -EBUSY;
 
-		commit = crtc_or_fake_commit(state, old_plane_state->crtc);
+		/*
+		 * Unlike connectors, always track planes explicitly for
+		 * async pageflip support.
+		 */
+		commit = crtc_or_fake_commit(state, new_plane_state->crtc ?: old_plane_state->crtc);
 		if (!commit)
 			return -ENOMEM;
 

drivers/gpu/drm/drm_gem_cma_helper.c

···
 	cma_obj->vaddr = dma_alloc_wc(drm->dev, size, &cma_obj->paddr,
 				      GFP_KERNEL | __GFP_NOWARN);
 	if (!cma_obj->vaddr) {
-		dev_err(drm->dev, "failed to allocate buffer with size %zu\n",
+		dev_dbg(drm->dev, "failed to allocate buffer with size %zu\n",
 			size);
 		ret = -ENOMEM;
 		goto error;

drivers/gpu/drm/panel/panel-simple.c

···
 		.width = 195,
 		.height = 117,
 	},
+	.delay = {
+		.enable = 160,
+		.disable = 160,
+	},
 };
 
 static const struct drm_display_mode innolux_at043tn24_mode = {
···
 	.htotal = 480 + 2 + 41 + 2,
 	.vdisplay = 272,
 	.vsync_start = 272 + 2,
-	.vsync_end = 272 + 2 + 11,
-	.vtotal = 272 + 2 + 11 + 2,
+	.vsync_end = 272 + 2 + 10,
+	.vtotal = 272 + 2 + 10 + 2,
 	.vrefresh = 60,
 	.flags = DRM_MODE_FLAG_NHSYNC | DRM_MODE_FLAG_NVSYNC,
 };
···
 		.height = 54,
 	},
 	.bus_format = MEDIA_BUS_FMT_RGB888_1X24,
+	.bus_flags = DRM_BUS_FLAG_DE_HIGH | DRM_BUS_FLAG_PIXDATA_POSEDGE,
 };
 
 static const struct drm_display_mode innolux_at070tn92_mode = {
···
 	.bus_format = MEDIA_BUS_FMT_RGB888_1X7X4_SPWG,
 };
 
+static const struct drm_display_mode toshiba_lt089ac29000_mode = {
+	.clock = 79500,
+	.hdisplay = 1280,
+	.hsync_start = 1280 + 192,
+	.hsync_end = 1280 + 192 + 128,
+	.htotal = 1280 + 192 + 128 + 64,
+	.vdisplay = 768,
+	.vsync_start = 768 + 20,
+	.vsync_end = 768 + 20 + 7,
+	.vtotal = 768 + 20 + 7 + 3,
+	.vrefresh = 60,
+};
+
+static const struct panel_desc toshiba_lt089ac29000 = {
+	.modes = &toshiba_lt089ac29000_mode,
+	.num_modes = 1,
+	.size = {
+		.width = 194,
+		.height = 116,
+	},
+	.bus_format = MEDIA_BUS_FMT_RGB888_1X24,
+	.bus_flags = DRM_BUS_FLAG_DE_HIGH | DRM_BUS_FLAG_PIXDATA_POSEDGE,
+};
+
 static const struct drm_display_mode tpk_f07a_0102_mode = {
 	.clock = 33260,
 	.hdisplay = 800,
···
 	}, {
 		.compatible = "tianma,tm070jdhg30",
 		.data = &tianma_tm070jdhg30,
+	}, {
+		.compatible = "toshiba,lt089ac29000",
+		.data = &toshiba_lt089ac29000,
 	}, {
 		.compatible = "tpk,f07a-0102",
 		.data = &tpk_f07a_0102,

drivers/gpu/drm/sun4i/Makefile

···
-sun4i-drm-y += sun4i_drv.o
-sun4i-drm-y += sun4i_framebuffer.o
+sun4i-backend-y			+= sun4i_backend.o sun4i_layer.o
 
-sun4i-drm-hdmi-y += sun4i_hdmi_enc.o
-sun4i-drm-hdmi-y += sun4i_hdmi_i2c.o
-sun4i-drm-hdmi-y += sun4i_hdmi_ddc_clk.o
-sun4i-drm-hdmi-y += sun4i_hdmi_tmds_clk.o
+sun4i-drm-y			+= sun4i_drv.o
+sun4i-drm-y			+= sun4i_framebuffer.o
 
-sun4i-tcon-y += sun4i_tcon.o
-sun4i-tcon-y += sun4i_rgb.o
-sun4i-tcon-y += sun4i_dotclock.o
-sun4i-tcon-y += sun4i_crtc.o
+sun4i-drm-hdmi-y		+= sun4i_hdmi_ddc_clk.o
+sun4i-drm-hdmi-y		+= sun4i_hdmi_enc.o
+sun4i-drm-hdmi-y		+= sun4i_hdmi_i2c.o
+sun4i-drm-hdmi-y		+= sun4i_hdmi_tmds_clk.o
 
-sun4i-backend-y += sun4i_backend.o sun4i_layer.o
+sun8i-mixer-y			+= sun8i_mixer.o sun8i_layer.o
 
-sun8i-mixer-y += sun8i_mixer.o sun8i_layer.o
+sun4i-tcon-y			+= sun4i_crtc.o
+sun4i-tcon-y			+= sun4i_dotclock.o
+sun4i-tcon-y			+= sun4i_tcon.o
+sun4i-tcon-y			+= sun4i_rgb.o
 
-obj-$(CONFIG_DRM_SUN4I)		+= sun4i-drm.o sun4i-tcon.o
-obj-$(CONFIG_DRM_SUN4I)		+= sun6i_drc.o
+obj-$(CONFIG_DRM_SUN4I)		+= sun4i-drm.o
+obj-$(CONFIG_DRM_SUN4I)		+= sun4i-tcon.o
 obj-$(CONFIG_DRM_SUN4I)		+= sun4i_tv.o
+obj-$(CONFIG_DRM_SUN4I)		+= sun6i_drc.o
 
-obj-$(CONFIG_DRM_SUN4I_BACKEND)		+= sun4i-backend.o
+obj-$(CONFIG_DRM_SUN4I_BACKEND)	+= sun4i-backend.o
 obj-$(CONFIG_DRM_SUN4I_HDMI)	+= sun4i-drm-hdmi.o
-obj-$(CONFIG_DRM_SUN8I_MIXER)		+= sun8i-mixer.o
+obj-$(CONFIG_DRM_SUN8I_MIXER)	+= sun8i-mixer.o

drivers/gpu/drm/sun4i/sun4i_backend.c

···
 
 #include <linux/component.h>
 #include <linux/list.h>
+#include <linux/of_device.h>
 #include <linux/of_graph.h>
 #include <linux/reset.h>
 
···
 #include "sun4i_drv.h"
 #include "sun4i_layer.h"
 #include "sunxi_engine.h"
+
+struct sun4i_backend_quirks {
+	/* backend <-> TCON muxing selection done in backend */
+	bool needs_output_muxing;
+};
 
 static const u32 sunxi_rgb2yuv_coef[12] = {
 	0x00000107, 0x00000204, 0x00000064, 0x00000108,
···
 	paddr = drm_fb_cma_get_gem_addr(fb, state, 0);
 	DRM_DEBUG_DRIVER("Setting buffer address to %pad\n", &paddr);
 
+	/*
+	 * backend DMA accesses DRAM directly, bypassing the system
+	 * bus. As such, the address range is different and the buffer
+	 * address needs to be corrected.
+	 */
+	paddr -= PHYS_OFFSET;
+
 	/* Write the 32 lower bits of the address (in bits) */
 	lo_paddr = paddr << 3;
 	DRM_DEBUG_DRIVER("Setting address lower bits to 0x%x\n", lo_paddr);
···
 	struct drm_device *drm = data;
 	struct sun4i_drv *drv = drm->dev_private;
 	struct sun4i_backend *backend;
+	const struct sun4i_backend_quirks *quirks;
 	struct resource *res;
 	void __iomem *regs;
 	int i, ret;
···
 		     SUN4I_BACKEND_MODCTL_DEBE_EN |
 		     SUN4I_BACKEND_MODCTL_START_CTL);
 
+	/* Set output selection if needed */
+	quirks = of_device_get_match_data(dev);
+	if (quirks->needs_output_muxing) {
+		/*
+		 * We assume there is no dynamic muxing of backends
+		 * and TCONs, so we select the backend with same ID.
+		 *
+		 * While dynamic selection might be interesting, since
+		 * the CRTC is tied to the TCON, while the layers are
+		 * tied to the backends, this means, we will need to
+		 * switch between groups of layers. There might not be
+		 * a way to represent this constraint in DRM.
+		 */
+		regmap_update_bits(backend->engine.regs,
+				   SUN4I_BACKEND_MODCTL_REG,
+				   SUN4I_BACKEND_MODCTL_OUT_SEL,
+				   (backend->engine.id
+				    ? SUN4I_BACKEND_MODCTL_OUT_LCD1
+				    : SUN4I_BACKEND_MODCTL_OUT_LCD0));
+	}
+
 	return 0;
 
 err_disable_ram_clk:
···
 	return 0;
 }
 
+static const struct sun4i_backend_quirks sun4i_backend_quirks = {
+	.needs_output_muxing = true,
+};
+
+static const struct sun4i_backend_quirks sun5i_backend_quirks = {
+};
+
+static const struct sun4i_backend_quirks sun6i_backend_quirks = {
+};
+
+static const struct sun4i_backend_quirks sun7i_backend_quirks = {
+	.needs_output_muxing = true,
+};
+
+static const struct sun4i_backend_quirks sun8i_a33_backend_quirks = {
+};
+
 static const struct of_device_id sun4i_backend_of_table[] = {
-	{ .compatible = "allwinner,sun5i-a13-display-backend" },
-	{ .compatible = "allwinner,sun6i-a31-display-backend" },
-	{ .compatible = "allwinner,sun8i-a33-display-backend" },
+	{
+		.compatible = "allwinner,sun4i-a10-display-backend",
+		.data = &sun4i_backend_quirks,
+	},
+	{
+		.compatible = "allwinner,sun5i-a13-display-backend",
+		.data = &sun5i_backend_quirks,
+	},
+	{
+		.compatible = "allwinner,sun6i-a31-display-backend",
+		.data = &sun6i_backend_quirks,
+	},
+	{
+		.compatible = "allwinner,sun7i-a20-display-backend",
+		.data = &sun7i_backend_quirks,
+	},
+	{
+		.compatible = "allwinner,sun8i-a33-display-backend",
+		.data = &sun8i_a33_backend_quirks,
+	},
 	{ }
 };
 MODULE_DEVICE_TABLE(of, sun4i_backend_of_table);

drivers/gpu/drm/sun4i/sun4i_backend.h

···
 #define SUN4I_BACKEND_MODCTL_LINE_SEL			BIT(29)
 #define SUN4I_BACKEND_MODCTL_ITLMOD_EN			BIT(28)
 #define SUN4I_BACKEND_MODCTL_OUT_SEL			GENMASK(22, 20)
-#define SUN4I_BACKEND_MODCTL_OUT_LCD				(0 << 20)
+#define SUN4I_BACKEND_MODCTL_OUT_LCD0				(0 << 20)
+#define SUN4I_BACKEND_MODCTL_OUT_LCD1				(1 << 20)
 #define SUN4I_BACKEND_MODCTL_OUT_FE0				(6 << 20)
 #define SUN4I_BACKEND_MODCTL_OUT_FE1				(7 << 20)
 #define SUN4I_BACKEND_MODCTL_HWC_EN			BIT(16)

drivers/gpu/drm/sun4i/sun4i_crtc.c

···
 #include "sunxi_engine.h"
 #include "sun4i_tcon.h"
 
+/*
+ * While this isn't really working in the DRM theory, in practice we
+ * can only ever have one encoder per TCON since we have a mux in our
+ * TCON.
+ */
+static struct drm_encoder *sun4i_crtc_get_encoder(struct drm_crtc *crtc)
+{
+	struct drm_encoder *encoder;
+
+	drm_for_each_encoder(encoder, crtc->dev)
+		if (encoder->crtc == crtc)
+			return encoder;
+
+	return NULL;
+}
+
 static void sun4i_crtc_atomic_begin(struct drm_crtc *crtc,
 				    struct drm_crtc_state *old_state)
 {
···
 static void sun4i_crtc_atomic_disable(struct drm_crtc *crtc,
 				      struct drm_crtc_state *old_state)
 {
+	struct drm_encoder *encoder = sun4i_crtc_get_encoder(crtc);
 	struct sun4i_crtc *scrtc = drm_crtc_to_sun4i_crtc(crtc);
 
 	DRM_DEBUG_DRIVER("Disabling the CRTC\n");
 
-	sun4i_tcon_disable(scrtc->tcon);
+	sun4i_tcon_set_status(scrtc->tcon, encoder, false);
 
 	if (crtc->state->event && !crtc->state->active) {
 		spin_lock_irq(&crtc->dev->event_lock);
···
 static void sun4i_crtc_atomic_enable(struct drm_crtc *crtc,
 				     struct drm_crtc_state *old_state)
 {
+	struct drm_encoder *encoder = sun4i_crtc_get_encoder(crtc);
 	struct sun4i_crtc *scrtc = drm_crtc_to_sun4i_crtc(crtc);
 
 	DRM_DEBUG_DRIVER("Enabling the CRTC\n");
 
-	sun4i_tcon_enable(scrtc->tcon);
+	sun4i_tcon_set_status(scrtc->tcon, encoder, true);
+}
+
+static void sun4i_crtc_mode_set_nofb(struct drm_crtc *crtc)
+{
+	struct drm_display_mode *mode = &crtc->state->adjusted_mode;
+	struct drm_encoder *encoder = sun4i_crtc_get_encoder(crtc);
+	struct sun4i_crtc *scrtc = drm_crtc_to_sun4i_crtc(crtc);
+
+	sun4i_tcon_mode_set(scrtc->tcon, encoder, mode);
 }
 
 static const struct drm_crtc_helper_funcs sun4i_crtc_helper_funcs = {
···
 	.atomic_flush	= sun4i_crtc_atomic_flush,
 	.atomic_enable	= sun4i_crtc_atomic_enable,
 	.atomic_disable	= sun4i_crtc_atomic_disable,
+	.mode_set_nofb	= sun4i_crtc_mode_set_nofb,
 };
 
 static int sun4i_crtc_enable_vblank(struct drm_crtc *crtc)

drivers/gpu/drm/sun4i/sun4i_drv.c

···
  */
 
 #include <linux/component.h>
+#include <linux/kfifo.h>
 #include <linux/of_graph.h>
 #include <linux/of_reserved_mem.h>
 
···
 
 static bool sun4i_drv_node_is_frontend(struct device_node *node)
 {
-	return of_device_is_compatible(node, "allwinner,sun5i-a13-display-frontend") ||
+	return of_device_is_compatible(node, "allwinner,sun4i-a10-display-frontend") ||
+		of_device_is_compatible(node, "allwinner,sun5i-a13-display-frontend") ||
 		of_device_is_compatible(node, "allwinner,sun6i-a31-display-frontend") ||
+		of_device_is_compatible(node, "allwinner,sun7i-a20-display-frontend") ||
 		of_device_is_compatible(node, "allwinner,sun8i-a33-display-frontend");
 }
 
 static bool sun4i_drv_node_is_tcon(struct device_node *node)
 {
-	return of_device_is_compatible(node, "allwinner,sun5i-a13-tcon") ||
+	return of_device_is_compatible(node, "allwinner,sun4i-a10-tcon") ||
+		of_device_is_compatible(node, "allwinner,sun5i-a13-tcon") ||
 		of_device_is_compatible(node, "allwinner,sun6i-a31-tcon") ||
 		of_device_is_compatible(node, "allwinner,sun6i-a31s-tcon") ||
+		of_device_is_compatible(node, "allwinner,sun7i-a20-tcon") ||
 		of_device_is_compatible(node, "allwinner,sun8i-a33-tcon") ||
 		of_device_is_compatible(node, "allwinner,sun8i-v3s-tcon");
 }
···
  * matching system handles this for us.
  */
 struct endpoint_list {
-	struct device_node *node;
-	struct list_head list;
+	DECLARE_KFIFO(fifo, struct device_node *, 16);
 };
 
-static bool node_is_in_list(struct list_head *endpoints,
-			    struct device_node *node)
-{
-	struct endpoint_list *endpoint;
-
-	list_for_each_entry(endpoint, endpoints, list)
-		if (endpoint->node == node)
-			return true;
-
-	return false;
-}
-
 static int sun4i_drv_add_endpoints(struct device *dev,
-				   struct list_head *endpoints,
+				   struct endpoint_list *list,
 				   struct component_match **match,
 				   struct device_node *node)
 {
 	struct device_node *port, *ep, *remote;
-	struct endpoint_list *endpoint;
 	int count = 0;
 
 	/*
···
 			}
 		}
 
-		/* skip downstream node if it is already in the queue */
-		if (node_is_in_list(endpoints, remote))
-			continue;
-
-		/* Add downstream nodes to the queue */
-		endpoint = kzalloc(sizeof(*endpoint), GFP_KERNEL);
-		if (!endpoint) {
-			of_node_put(remote);
-			return -ENOMEM;
-		}
-
-		endpoint->node = remote;
-		list_add_tail(&endpoint->list, endpoints);
+		kfifo_put(&list->fifo, remote);
 	}
 
 	return count;
···
 static int sun4i_drv_probe(struct platform_device *pdev)
 {
 	struct component_match *match = NULL;
-	struct device_node *np = pdev->dev.of_node;
-	struct endpoint_list *endpoint, *endpoint_temp;
+	struct device_node *np = pdev->dev.of_node, *endpoint;
+	struct endpoint_list list;
 	int i, ret, count = 0;
-	LIST_HEAD(endpoints);
+
+	INIT_KFIFO(list.fifo);
 
 	for (i = 0;; i++) {
 		struct device_node *pipeline = of_parse_phandle(np,
···
 		if (!pipeline)
 			break;
 
-		endpoint = kzalloc(sizeof(*endpoint), GFP_KERNEL);
-		if (!endpoint) {
-			ret = -ENOMEM;
-			goto err_free_endpoints;
-		}
-
-		endpoint->node = pipeline;
-		list_add_tail(&endpoint->list, &endpoints);
+		kfifo_put(&list.fifo, pipeline);
 	}
 
-	list_for_each_entry_safe(endpoint, endpoint_temp, &endpoints, list) {
+	while (kfifo_get(&list.fifo, &endpoint)) {
 		/* process this endpoint */
-		ret = sun4i_drv_add_endpoints(&pdev->dev, &endpoints, &match,
-					      endpoint->node);
+		ret = sun4i_drv_add_endpoints(&pdev->dev, &list, &match,
+					      endpoint);
 
 		/* sun4i_drv_add_endpoints can fail to allocate memory */
 		if (ret < 0)
-			goto err_free_endpoints;
+			return ret;
 
 		count += ret;
-
-		/* delete and cleanup the current entry */
-		list_del(&endpoint->list);
-		of_node_put(endpoint->node);
-		kfree(endpoint);
 	}
 
 	if (count)
···
 						       match);
 	else
 		return 0;
-
-err_free_endpoints:
-	list_for_each_entry_safe(endpoint, endpoint_temp, &endpoints, list) {
-		list_del(&endpoint->list);
-		of_node_put(endpoint->node);
-		kfree(endpoint);
-	}
-
-	return ret;
 }
 
 static int sun4i_drv_remove(struct platform_device *pdev)
···
 }
 
 static const struct of_device_id sun4i_drv_of_table[] = {
+	{ .compatible = "allwinner,sun4i-a10-display-engine" },
 	{ .compatible = "allwinner,sun5i-a10s-display-engine" },
 	{ .compatible = "allwinner,sun5i-a13-display-engine" },
 	{ .compatible = "allwinner,sun6i-a31-display-engine" },
 	{ .compatible = "allwinner,sun6i-a31s-display-engine" },
+	{ .compatible = "allwinner,sun7i-a20-display-engine" },
 	{ .compatible = "allwinner,sun8i-a33-display-engine" },
 	{ .compatible = "allwinner,sun8i-v3s-display-engine" },
 	{ }

drivers/gpu/drm/sun4i/sun4i_hdmi_ddc_clk.c

···
 #include <linux/clk-provider.h>
 #include <linux/regmap.h>
 
-#include "sun4i_tcon.h"
 #include "sun4i_hdmi.h"
 
 struct sun4i_ddc {

drivers/gpu/drm/sun4i/sun4i_hdmi_enc.c

···
 #include "sun4i_crtc.h"
 #include "sun4i_drv.h"
 #include "sun4i_hdmi.h"
-#include "sun4i_tcon.h"
 
 static inline struct sun4i_hdmi *
 drm_encoder_to_sun4i_hdmi(struct drm_encoder *encoder)
···
 static void sun4i_hdmi_disable(struct drm_encoder *encoder)
 {
 	struct sun4i_hdmi *hdmi = drm_encoder_to_sun4i_hdmi(encoder);
-	struct sun4i_crtc *crtc = drm_crtc_to_sun4i_crtc(encoder->crtc);
-	struct sun4i_tcon *tcon = crtc->tcon;
 	u32 val;
 
 	DRM_DEBUG_DRIVER("Disabling the HDMI Output\n");
···
 	val = readl(hdmi->base + SUN4I_HDMI_VID_CTRL_REG);
 	val &= ~SUN4I_HDMI_VID_CTRL_ENABLE;
 	writel(val, hdmi->base + SUN4I_HDMI_VID_CTRL_REG);
-
-	sun4i_tcon_channel_disable(tcon, 1);
 }
 
 static void sun4i_hdmi_enable(struct drm_encoder *encoder)
 {
 	struct drm_display_mode *mode = &encoder->crtc->state->adjusted_mode;
 	struct sun4i_hdmi *hdmi = drm_encoder_to_sun4i_hdmi(encoder);
-	struct sun4i_crtc *crtc = drm_crtc_to_sun4i_crtc(encoder->crtc);
-	struct sun4i_tcon *tcon = crtc->tcon;
 	u32 val = 0;
 
 	DRM_DEBUG_DRIVER("Enabling the HDMI Output\n");
-
-	sun4i_tcon_channel_enable(tcon, 1);
 
 	sun4i_hdmi_setup_avi_infoframes(hdmi, mode);
 	val |= SUN4I_HDMI_PKT_CTRL_TYPE(0, SUN4I_HDMI_PKT_AVI);
···
 				struct drm_display_mode *adjusted_mode)
 {
 	struct sun4i_hdmi *hdmi = drm_encoder_to_sun4i_hdmi(encoder);
-	struct sun4i_crtc *crtc = drm_crtc_to_sun4i_crtc(encoder->crtc);
-	struct sun4i_tcon *tcon = crtc->tcon;
 	unsigned int x, y;
 	u32 val;
 
-	sun4i_tcon1_mode_set(tcon, mode);
-	sun4i_tcon_set_mux(tcon, 1, encoder);
-
-	clk_set_rate(tcon->sclk1, mode->crtc_clock * 1000);
 	clk_set_rate(hdmi->mod_clk, mode->crtc_clock * 1000);
 	clk_set_rate(hdmi->tmds_clk, mode->crtc_clock * 1000);
 
···
 
 #define SUN4I_HDMI_PAD_CTRL1_MASK	(GENMASK(24, 7) | GENMASK(5, 0))
 #define SUN4I_HDMI_PLL_CTRL_MASK	(GENMASK(31, 8) | GENMASK(3, 0))
+
+/* Only difference from sun5i is AMP is 4 instead of 6 */
+static const struct sun4i_hdmi_variant sun4i_variant = {
+	.pad_ctrl0_init_val	= SUN4I_HDMI_PAD_CTRL0_TXEN |
+				  SUN4I_HDMI_PAD_CTRL0_CKEN |
+				  SUN4I_HDMI_PAD_CTRL0_PWENG |
+				  SUN4I_HDMI_PAD_CTRL0_PWEND |
+				  SUN4I_HDMI_PAD_CTRL0_PWENC |
+				  SUN4I_HDMI_PAD_CTRL0_LDODEN |
+				  SUN4I_HDMI_PAD_CTRL0_LDOCEN |
+				  SUN4I_HDMI_PAD_CTRL0_BIASEN,
+	.pad_ctrl1_init_val	= SUN4I_HDMI_PAD_CTRL1_REG_AMP(4) |
+				  SUN4I_HDMI_PAD_CTRL1_REG_EMP(2) |
+				  SUN4I_HDMI_PAD_CTRL1_REG_DENCK |
+				  SUN4I_HDMI_PAD_CTRL1_REG_DEN |
+				  SUN4I_HDMI_PAD_CTRL1_EMPCK_OPT |
+				  SUN4I_HDMI_PAD_CTRL1_EMP_OPT |
+				  SUN4I_HDMI_PAD_CTRL1_AMPCK_OPT |
+				  SUN4I_HDMI_PAD_CTRL1_AMP_OPT,
+	.pll_ctrl_init_val	= SUN4I_HDMI_PLL_CTRL_VCO_S(8) |
+				  SUN4I_HDMI_PLL_CTRL_CS(7) |
+				  SUN4I_HDMI_PLL_CTRL_CP_S(15) |
+				  SUN4I_HDMI_PLL_CTRL_S(7) |
+				  SUN4I_HDMI_PLL_CTRL_VCO_GAIN(4) |
+				  SUN4I_HDMI_PLL_CTRL_SDIV2 |
+				  SUN4I_HDMI_PLL_CTRL_LDO2_EN |
+				  SUN4I_HDMI_PLL_CTRL_LDO1_EN |
+				  SUN4I_HDMI_PLL_CTRL_HV_IS_33 |
+				  SUN4I_HDMI_PLL_CTRL_BWS |
+				  SUN4I_HDMI_PLL_CTRL_PLL_EN,
+
+	.ddc_clk_reg		= REG_FIELD(SUN4I_HDMI_DDC_CLK_REG, 0, 6),
+	.ddc_clk_pre_divider	= 2,
+	.ddc_clk_m_offset	= 1,
+
+	.field_ddc_en		= REG_FIELD(SUN4I_HDMI_DDC_CTRL_REG, 31, 31),
+	.field_ddc_start	= REG_FIELD(SUN4I_HDMI_DDC_CTRL_REG, 30, 30),
+	.field_ddc_reset	= REG_FIELD(SUN4I_HDMI_DDC_CTRL_REG, 0, 0),
+	.field_ddc_addr_reg	= REG_FIELD(SUN4I_HDMI_DDC_ADDR_REG, 0, 31),
+	.field_ddc_slave_addr	= REG_FIELD(SUN4I_HDMI_DDC_ADDR_REG, 0, 6),
+	.field_ddc_int_status	= REG_FIELD(SUN4I_HDMI_DDC_INT_STATUS_REG, 0, 8),
+	.field_ddc_fifo_clear	= REG_FIELD(SUN4I_HDMI_DDC_FIFO_CTRL_REG, 31, 31),
+	.field_ddc_fifo_rx_thres = REG_FIELD(SUN4I_HDMI_DDC_FIFO_CTRL_REG, 4, 7),
+	.field_ddc_fifo_tx_thres = REG_FIELD(SUN4I_HDMI_DDC_FIFO_CTRL_REG, 0, 3),
+	.field_ddc_byte_count	= REG_FIELD(SUN4I_HDMI_DDC_BYTE_COUNT_REG, 0, 9),
+	.field_ddc_cmd		= REG_FIELD(SUN4I_HDMI_DDC_CMD_REG, 0, 2),
+	.field_ddc_sda_en	= REG_FIELD(SUN4I_HDMI_DDC_LINE_CTRL_REG, 9, 9),
+	.field_ddc_sck_en	= REG_FIELD(SUN4I_HDMI_DDC_LINE_CTRL_REG, 8, 8),
+
+	.ddc_fifo_reg		= SUN4I_HDMI_DDC_FIFO_DATA_REG,
+	.ddc_fifo_has_dir	= true,
+};
 
 static const struct sun4i_hdmi_variant sun5i_variant = {
 	.pad_ctrl0_init_val	= SUN4I_HDMI_PAD_CTRL0_TXEN |
···
 }
 
 static const struct of_device_id sun4i_hdmi_of_table[] = {
+	{ .compatible = "allwinner,sun4i-a10-hdmi", .data = &sun4i_variant, },
 	{ .compatible = "allwinner,sun5i-a10s-hdmi", .data = &sun5i_variant, },
 	{ .compatible = "allwinner,sun6i-a31-hdmi", .data = &sun6i_variant, },
 	{ }

drivers/gpu/drm/sun4i/sun4i_hdmi_tmds_clk.c

···
 
 #include <linux/clk-provider.h>
 
-#include "sun4i_tcon.h"
 #include "sun4i_hdmi.h"
 
 struct sun4i_tmds {

drivers/gpu/drm/sun4i/sun4i_rgb.c

···
 
 	DRM_DEBUG_DRIVER("Enabling RGB output\n");
 
-	if (!IS_ERR(tcon->panel))
+	if (!IS_ERR(tcon->panel)) {
 		drm_panel_prepare(tcon->panel);
-
-	sun4i_tcon_channel_enable(tcon, 0);
-
-	if (!IS_ERR(tcon->panel))
 		drm_panel_enable(tcon->panel);
+	}
 }
 
 static void sun4i_rgb_encoder_disable(struct drm_encoder *encoder)
···
 
 	DRM_DEBUG_DRIVER("Disabling RGB output\n");
 
-	if (!IS_ERR(tcon->panel))
+	if (!IS_ERR(tcon->panel)) {
 		drm_panel_disable(tcon->panel);
-
-	sun4i_tcon_channel_disable(tcon, 0);
-
-	if (!IS_ERR(tcon->panel))
 		drm_panel_unprepare(tcon->panel);
-}
-
-static void sun4i_rgb_encoder_mode_set(struct drm_encoder *encoder,
-				       struct drm_display_mode *mode,
-				       struct drm_display_mode *adjusted_mode)
-{
-	struct sun4i_rgb *rgb = drm_encoder_to_sun4i_rgb(encoder);
-	struct sun4i_tcon *tcon = rgb->tcon;
-
-	sun4i_tcon0_mode_set(tcon, mode);
-	sun4i_tcon_set_mux(tcon, 0, encoder);
-
-	/* FIXME: This seems to be board specific */
-	clk_set_phase(tcon->dclk, 120);
+	}
 }
 
 static struct drm_encoder_helper_funcs sun4i_rgb_enc_helper_funcs = {
-	.mode_set	= sun4i_rgb_encoder_mode_set,
 	.disable	= sun4i_rgb_encoder_disable,
 	.enable		= sun4i_rgb_encoder_enable,
 };

drivers/gpu/drm/sun4i/sun4i_tcon.c

···
 #include "sun4i_tcon.h"
 #include "sunxi_engine.h"
 
-void sun4i_tcon_disable(struct sun4i_tcon *tcon)
+static void sun4i_tcon_channel_set_status(struct sun4i_tcon *tcon, int channel,
+					  bool enabled)
 {
-	DRM_DEBUG_DRIVER("Disabling TCON\n");
+	struct clk *clk;
 
-	/* Disable the TCON */
-	regmap_update_bits(tcon->regs, SUN4I_TCON_GCTL_REG,
-			   SUN4I_TCON_GCTL_TCON_ENABLE, 0);
-}
-EXPORT_SYMBOL(sun4i_tcon_disable);
-
-void sun4i_tcon_enable(struct sun4i_tcon *tcon)
-{
-	DRM_DEBUG_DRIVER("Enabling TCON\n");
-
-	/* Enable the TCON */
-	regmap_update_bits(tcon->regs, SUN4I_TCON_GCTL_REG,
-			   SUN4I_TCON_GCTL_TCON_ENABLE,
-			   SUN4I_TCON_GCTL_TCON_ENABLE);
-}
-EXPORT_SYMBOL(sun4i_tcon_enable);
-
-void sun4i_tcon_channel_disable(struct sun4i_tcon *tcon, int channel)
-{
-	DRM_DEBUG_DRIVER("Disabling TCON channel %d\n", channel);
-
-	/* Disable the TCON's channel */
-	if (channel == 0) {
-		regmap_update_bits(tcon->regs, SUN4I_TCON0_CTL_REG,
-				   SUN4I_TCON0_CTL_TCON_ENABLE, 0);
-		clk_disable_unprepare(tcon->dclk);
-		return;
-	}
-
-	WARN_ON(!tcon->quirks->has_channel_1);
-	regmap_update_bits(tcon->regs, SUN4I_TCON1_CTL_REG,
-			   SUN4I_TCON1_CTL_TCON_ENABLE, 0);
-	clk_disable_unprepare(tcon->sclk1);
-}
-EXPORT_SYMBOL(sun4i_tcon_channel_disable);
-
-void sun4i_tcon_channel_enable(struct sun4i_tcon *tcon, int channel)
-{
-	DRM_DEBUG_DRIVER("Enabling TCON channel %d\n", channel);
-
-	/* Enable the TCON's channel */
-	if (channel == 0) {
+	switch (channel) {
+	case 0:
 		regmap_update_bits(tcon->regs, SUN4I_TCON0_CTL_REG,
 				   SUN4I_TCON0_CTL_TCON_ENABLE,
-				   SUN4I_TCON0_CTL_TCON_ENABLE);
-		clk_prepare_enable(tcon->dclk);
+				   enabled ? SUN4I_TCON0_CTL_TCON_ENABLE : 0);
+		clk = tcon->dclk;
+		break;
+	case 1:
+		WARN_ON(!tcon->quirks->has_channel_1);
+		regmap_update_bits(tcon->regs, SUN4I_TCON1_CTL_REG,
+				   SUN4I_TCON1_CTL_TCON_ENABLE,
+				   enabled ? SUN4I_TCON1_CTL_TCON_ENABLE : 0);
+		clk = tcon->sclk1;
+		break;
+	default:
+		DRM_WARN("Unknown channel... doing nothing\n");
 		return;
 	}
 
-	WARN_ON(!tcon->quirks->has_channel_1);
-	regmap_update_bits(tcon->regs, SUN4I_TCON1_CTL_REG,
-			   SUN4I_TCON1_CTL_TCON_ENABLE,
-			   SUN4I_TCON1_CTL_TCON_ENABLE);
-	clk_prepare_enable(tcon->sclk1);
+	if (enabled)
+		clk_prepare_enable(clk);
+	else
+		clk_disable_unprepare(clk);
 }
-EXPORT_SYMBOL(sun4i_tcon_channel_enable);
+
+void sun4i_tcon_set_status(struct sun4i_tcon *tcon,
+			   const struct drm_encoder *encoder,
+			   bool enabled)
+{
+	int channel;
+
+	switch (encoder->encoder_type) {
+	case DRM_MODE_ENCODER_NONE:
+		channel = 0;
+		break;
+	case DRM_MODE_ENCODER_TMDS:
+	case DRM_MODE_ENCODER_TVDAC:
+		channel = 1;
+		break;
+	default:
+		DRM_DEBUG_DRIVER("Unknown encoder type, doing nothing...\n");
+		return;
+	}
+
+	regmap_update_bits(tcon->regs, SUN4I_TCON_GCTL_REG,
+			   SUN4I_TCON_GCTL_TCON_ENABLE,
+			   enabled ? SUN4I_TCON_GCTL_TCON_ENABLE : 0);
+
+	sun4i_tcon_channel_set_status(tcon, channel, enabled);
+}
 
 void sun4i_tcon_enable_vblank(struct sun4i_tcon *tcon, bool enable)
 {
···
 }
 
 void sun4i_tcon_set_mux(struct sun4i_tcon *tcon, int channel,
-			struct drm_encoder *encoder)
+			const struct drm_encoder *encoder)
 {
 	int ret = -ENOTSUPP;
 
···
 	DRM_DEBUG_DRIVER("Muxing encoder %s to CRTC %s: %d\n",
 			 encoder->name, encoder->crtc->name, ret);
 }
-EXPORT_SYMBOL(sun4i_tcon_set_mux);
 
-static int sun4i_tcon_get_clk_delay(struct drm_display_mode *mode,
+static int sun4i_tcon_get_clk_delay(const struct drm_display_mode *mode,
 				    int channel)
 {
 	int delay = mode->vtotal - mode->vdisplay;
···
 	return delay;
 }
 
-void sun4i_tcon0_mode_set(struct sun4i_tcon *tcon,
-			  struct drm_display_mode *mode)
+static void sun4i_tcon0_mode_set_common(struct sun4i_tcon *tcon,
+					const struct drm_display_mode *mode)
+{
+	/* Configure the dot clock */
+	clk_set_rate(tcon->dclk, mode->crtc_clock * 1000);
+
+	/* Set the resolution */
+	regmap_write(tcon->regs, SUN4I_TCON0_BASIC0_REG,
+		     SUN4I_TCON0_BASIC0_X(mode->crtc_hdisplay) |
+		     SUN4I_TCON0_BASIC0_Y(mode->crtc_vdisplay));
+}
+
+static void sun4i_tcon0_mode_set_rgb(struct sun4i_tcon *tcon,
+				     const struct drm_display_mode *mode)
 {
 	unsigned int bp, hsync, vsync;
 	u8 clk_delay;
 	u32 val = 0;
 
-	/* Configure the dot clock */
-	clk_set_rate(tcon->dclk, mode->crtc_clock * 1000);
+	sun4i_tcon0_mode_set_common(tcon, mode);
 
 	/* Adjust clock delay */
 	clk_delay = sun4i_tcon_get_clk_delay(mode, 0);
 	regmap_update_bits(tcon->regs, SUN4I_TCON0_CTL_REG,
 			   SUN4I_TCON0_CTL_CLK_DELAY_MASK,
 			   SUN4I_TCON0_CTL_CLK_DELAY(clk_delay));
-
-	/* Set the resolution */
-	regmap_write(tcon->regs, SUN4I_TCON0_BASIC0_REG,
-		     SUN4I_TCON0_BASIC0_X(mode->crtc_hdisplay) |
-		     SUN4I_TCON0_BASIC0_Y(mode->crtc_vdisplay));
 
 	/*
 	 * This is called a backporch in the register documentation,
···
 	/* Enable the output on the pins */
 	regmap_write(tcon->regs, SUN4I_TCON0_IO_TRI_REG, 0);
 }
-EXPORT_SYMBOL(sun4i_tcon0_mode_set);
 
-void sun4i_tcon1_mode_set(struct sun4i_tcon *tcon,
-			  struct drm_display_mode *mode)
+static void sun4i_tcon1_mode_set(struct sun4i_tcon *tcon,
+				 const struct drm_display_mode *mode)
 {
 	unsigned int bp, hsync, vsync, vtotal;
 	u8 clk_delay;
···
 			   SUN4I_TCON_GCTL_IOMAP_MASK,
 			   SUN4I_TCON_GCTL_IOMAP_TCON1);
 }
-EXPORT_SYMBOL(sun4i_tcon1_mode_set);
+
+void sun4i_tcon_mode_set(struct sun4i_tcon *tcon,
+			 const struct drm_encoder *encoder,
+			 const struct drm_display_mode *mode)
+{
+	switch (encoder->encoder_type) {
+	case DRM_MODE_ENCODER_NONE:
+		sun4i_tcon0_mode_set_rgb(tcon, mode);
+		sun4i_tcon_set_mux(tcon, 0, encoder);
+		break;
+	case DRM_MODE_ENCODER_TVDAC:
+	case DRM_MODE_ENCODER_TMDS:
+		sun4i_tcon1_mode_set(tcon, mode);
+		sun4i_tcon_set_mux(tcon, 1, encoder);
+		break;
+	default:
+		DRM_DEBUG_DRIVER("Unknown encoder type, doing nothing...\n");
+	}
+}
+EXPORT_SYMBOL(sun4i_tcon_mode_set);
 
 static void sun4i_tcon_finish_page_flip(struct drm_device *dev,
 					struct sun4i_crtc *scrtc)
···
 }
 
 /* platform specific TCON muxing callbacks */
+static int sun4i_a10_tcon_set_mux(struct sun4i_tcon *tcon,
+				  const struct drm_encoder *encoder)
+{
+	struct sun4i_tcon *tcon0 = sun4i_get_tcon0(encoder->dev);
+	u32 shift;
+
+	if (!tcon0)
+		return -EINVAL;
+
+	switch (encoder->encoder_type) {
+	case DRM_MODE_ENCODER_TMDS:
+		/* HDMI */
+		shift = 8;
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	regmap_update_bits(tcon0->regs, SUN4I_TCON_MUX_CTRL_REG,
+			   0x3 << shift, tcon->id << shift);
+
+	return 0;
+}
+
 static int sun5i_a13_tcon_set_mux(struct sun4i_tcon *tcon,
-				  struct drm_encoder *encoder)
+				  const struct drm_encoder *encoder)
 {
 	u32 val;
 
···
 }
 
 static int sun6i_tcon_set_mux(struct sun4i_tcon *tcon,
-			      struct drm_encoder *encoder)
+			      const struct drm_encoder *encoder)
 {
 	struct sun4i_tcon *tcon0 = sun4i_get_tcon0(encoder->dev);
 	u32 shift;
···
 	return 0;
 }
 
+static const struct sun4i_tcon_quirks sun4i_a10_quirks = {
+	.has_channel_1		= true,
+	.set_mux		= sun4i_a10_tcon_set_mux,
+};
+
 static const struct sun4i_tcon_quirks sun5i_a13_quirks = {
 	.has_channel_1		= true,
 	.set_mux		= sun5i_a13_tcon_set_mux,
···
 	.needs_de_be_mux	= true,
 };
 
+static const struct sun4i_tcon_quirks sun7i_a20_quirks = {
+	.has_channel_1		= true,
+	/* Same display pipeline structure as A10 */
+	.set_mux		= sun4i_a10_tcon_set_mux,
+};
+
 static const struct sun4i_tcon_quirks sun8i_a33_quirks = {
 	/* nothing is supported */
 };
···
 };
 
 static const struct of_device_id sun4i_tcon_of_table[] = {
+	{ .compatible = "allwinner,sun4i-a10-tcon", .data = &sun4i_a10_quirks },
 	{ .compatible = "allwinner,sun5i-a13-tcon", .data = &sun5i_a13_quirks },
 	{ .compatible = "allwinner,sun6i-a31-tcon", .data = &sun6i_a31_quirks },
 	{ .compatible = "allwinner,sun6i-a31s-tcon", .data = &sun6i_a31s_quirks },
+	{ .compatible = "allwinner,sun7i-a20-tcon", .data = &sun7i_a20_quirks },
 	{ .compatible = "allwinner,sun8i-a33-tcon", .data = &sun8i_a33_quirks },
 	{ .compatible = "allwinner,sun8i-v3s-tcon", .data = &sun8i_v3s_quirks },
 	{ }

drivers/gpu/drm/sun4i/sun4i_tcon.h

···
 	bool	needs_de_be_mux; /* sun6i needs mux to select backend */
 
 	/* callback to handle tcon muxing options */
-	int	(*set_mux)(struct sun4i_tcon *, struct drm_encoder *);
+	int	(*set_mux)(struct sun4i_tcon *, const struct drm_encoder *);
 };
 
 struct sun4i_tcon {
···
 struct drm_bridge *sun4i_tcon_find_bridge(struct device_node *node);
 struct drm_panel *sun4i_tcon_find_panel(struct device_node *node);
 
-/* Global Control */
-void sun4i_tcon_disable(struct sun4i_tcon *tcon);
-void sun4i_tcon_enable(struct sun4i_tcon *tcon);
-
-/* Channel Control */
-void sun4i_tcon_channel_disable(struct sun4i_tcon *tcon, int channel);
-void sun4i_tcon_channel_enable(struct sun4i_tcon *tcon, int channel);
-
 void sun4i_tcon_enable_vblank(struct sun4i_tcon *tcon, bool enable);
-
-/* Mode Related Controls */
-void sun4i_tcon_set_mux(struct sun4i_tcon *tcon, int channel,
-			struct drm_encoder *encoder);
-void sun4i_tcon0_mode_set(struct sun4i_tcon *tcon,
-			  struct drm_display_mode *mode);
-void sun4i_tcon1_mode_set(struct sun4i_tcon *tcon,
-			  struct drm_display_mode *mode);
+void sun4i_tcon_mode_set(struct sun4i_tcon *tcon,
+			 const struct drm_encoder *encoder,
+			 const struct drm_display_mode *mode);
+void sun4i_tcon_set_status(struct sun4i_tcon *crtc,
+			   const struct drm_encoder *encoder, bool enable);
 
 #endif /* __SUN4I_TCON_H__ */

drivers/gpu/drm/sun4i/sun4i_tv.c

···
 
 #include "sun4i_crtc.h"
 #include "sun4i_drv.h"
-#include "sun4i_tcon.h"
 #include "sunxi_engine.h"
 
 #define SUN4I_TVE_EN_REG		0x000
···
 {
 	struct sun4i_tv *tv = drm_encoder_to_sun4i_tv(encoder);
 	struct sun4i_crtc *crtc = drm_crtc_to_sun4i_crtc(encoder->crtc);
-	struct sun4i_tcon *tcon = crtc->tcon;
 
 	DRM_DEBUG_DRIVER("Disabling the TV Output\n");
-
-	sun4i_tcon_channel_disable(tcon, 1);
 
 	regmap_update_bits(tv->regs, SUN4I_TVE_EN_REG,
 			   SUN4I_TVE_EN_ENABLE,
···
 {
 	struct sun4i_tv *tv = drm_encoder_to_sun4i_tv(encoder);
 	struct sun4i_crtc *crtc = drm_crtc_to_sun4i_crtc(encoder->crtc);
-	struct sun4i_tcon *tcon = crtc->tcon;
 
 	DRM_DEBUG_DRIVER("Enabling the TV Output\n");
 
···
 	regmap_update_bits(tv->regs, SUN4I_TVE_EN_REG,
 			   SUN4I_TVE_EN_ENABLE,
 			   SUN4I_TVE_EN_ENABLE);
-
-	sun4i_tcon_channel_enable(tcon, 1);
 }
 
 static void sun4i_tv_mode_set(struct drm_encoder *encoder,
···
 			      struct drm_display_mode *adjusted_mode)
 {
 	struct sun4i_tv *tv = drm_encoder_to_sun4i_tv(encoder);
-	struct sun4i_crtc *crtc = drm_crtc_to_sun4i_crtc(encoder->crtc);
-	struct sun4i_tcon *tcon = crtc->tcon;
 	const struct tv_mode *tv_mode = sun4i_tv_find_tv_by_mode(mode);
-
-	sun4i_tcon1_mode_set(tcon, mode);
-	sun4i_tcon_set_mux(tcon, 1, encoder);
 
 	/* Enable and map the DAC to the output */
 	regmap_update_bits(tv->regs, SUN4I_TVE_EN_REG,

drivers/gpu/drm/udl/udl_connector.c

···
 #include <drm/drm_crtc.h>
 #include <drm/drm_edid.h>
 #include <drm/drm_crtc_helper.h>
+#include "udl_connector.h"
 #include "udl_drv.h"
 
-/* dummy connector to just get EDID,
-   all UDL appear to have a DVI-D */
-
-static u8 *udl_get_edid(struct udl_device *udl)
+static bool udl_get_edid_block(struct udl_device *udl, int block_idx,
+							   u8 *buff)
 {
-	u8 *block;
-	char *rbuf;
 	int ret, i;
+	u8 *read_buff;
 
-	block = kmalloc(EDID_LENGTH, GFP_KERNEL);
-	if (block == NULL)
-		return NULL;
-
-	rbuf = kmalloc(2, GFP_KERNEL);
-	if (rbuf == NULL)
-		goto error;
+	read_buff = kmalloc(2, GFP_KERNEL);
+	if (!read_buff)
+		return false;
 
 	for (i = 0; i < EDID_LENGTH; i++) {
+		int bval = (i + block_idx * EDID_LENGTH) << 8;
 		ret = usb_control_msg(udl->udev,
-				      usb_rcvctrlpipe(udl->udev, 0), (0x02),
-				      (0x80 | (0x02 << 5)), i << 8, 0xA1, rbuf, 2,
-				      HZ);
+				      usb_rcvctrlpipe(udl->udev, 0),
+					  (0x02), (0x80 | (0x02 << 5)), bval,
+					  0xA1, read_buff, 2, HZ);
 		if (ret < 1) {
 			DRM_ERROR("Read EDID byte %d failed err %x\n", i, ret);
-			goto error;
+			kfree(read_buff);
+			return false;
 		}
-		block[i] = rbuf[1];
+		buff[i] = read_buff[1];
 	}
 
-	kfree(rbuf);
-	return block;
+	kfree(read_buff);
+	return true;
+}
 
-error:
-	kfree(block);
-	kfree(rbuf);
-	return NULL;
+static bool udl_get_edid(struct udl_device *udl, u8 **result_buff,
+			 int *result_buff_size)
+{
+	int i, extensions;
+	u8 *block_buff = NULL, *buff_ptr;
+
+	block_buff = kmalloc(EDID_LENGTH, GFP_KERNEL);
+	if (block_buff == NULL)
+		return false;
+
+	if (udl_get_edid_block(udl, 0, block_buff) &&
+	    memchr_inv(block_buff, 0, EDID_LENGTH)) {
+		extensions = ((struct edid *)block_buff)->extensions;
+		if (extensions > 0) {
+			/* we have to read all extensions one by one */
+			*result_buff_size = EDID_LENGTH * (extensions + 1);
+			*result_buff = kmalloc(*result_buff_size, GFP_KERNEL);
+			buff_ptr = *result_buff;
+			if (buff_ptr == NULL) {
+				kfree(block_buff);
+				return false;
+			}
+			memcpy(buff_ptr, block_buff, EDID_LENGTH);
+			kfree(block_buff);
+			buff_ptr += EDID_LENGTH;
+			for (i = 1; i < extensions; ++i) {
+				if (udl_get_edid_block(udl, i, buff_ptr)) {
+					buff_ptr += EDID_LENGTH;
+				} else {
+					kfree(*result_buff);
+					*result_buff = NULL;
+					return false;
+				}
+			}
+			return true;
+		}
+		/* we have only base edid block */
+		*result_buff = block_buff;
+		*result_buff_size = EDID_LENGTH;
+		return true;
+	}
+
+	kfree(block_buff);
+
+	return false;
 }
 
 static int udl_get_modes(struct drm_connector *connector)
 {
-	struct udl_device *udl = connector->dev->dev_private;
-	struct edid *edid;
-	int ret;
+	struct udl_drm_connector *udl_connector =
+					container_of(connector,
+					struct udl_drm_connector,
+					connector);
 
-	edid = (struct edid *)udl_get_edid(udl);
-	if (!edid) {
-		drm_mode_connector_update_edid_property(connector, NULL);
-		return 0;
-	}
-
-	/*
-	 * We only read the main block, but if the monitor reports extension
-	 * blocks then the drm edid code expects them to be present, so patch
-	 * the extension count to 0.
-	 */
-	edid->checksum += edid->extensions;
-	edid->extensions = 0;
-
-	drm_mode_connector_update_edid_property(connector, edid);
-	ret = drm_add_edid_modes(connector, edid);
-	kfree(edid);
-	return ret;
+	drm_mode_connector_update_edid_property(connector, udl_connector->edid);
+	if (udl_connector->edid)
+		return drm_add_edid_modes(connector, udl_connector->edid);
+	return 0;
 }
 
 static int udl_mode_valid(struct drm_connector *connector,
···
 static enum drm_connector_status
 udl_detect(struct drm_connector *connector, bool force)
 {
-	if (drm_dev_is_unplugged(connector->dev))
+	u8 *edid_buff = NULL;
+	int edid_buff_size = 0;
+	struct udl_device *udl = connector->dev->dev_private;
+	struct udl_drm_connector *udl_connector =
+					container_of(connector,
+					struct udl_drm_connector,
+					connector);
+
+	/* cleanup previous edid */
+	if (udl_connector->edid != NULL) {
+		kfree(udl_connector->edid);
+		udl_connector->edid = NULL;
+	}
+
+
+	if (!udl_get_edid(udl, &edid_buff, &edid_buff_size))
 		return connector_status_disconnected;
+
+	udl_connector->edid = (struct edid *)edid_buff;
+	
 	return connector_status_connected;
 }
 
···
 
 static void udl_connector_destroy(struct drm_connector *connector)
 {
+	struct udl_drm_connector *udl_connector =
+					container_of(connector,
+					struct udl_drm_connector,
+					connector);
+
 	drm_connector_unregister(connector);
 	drm_connector_cleanup(connector);
+	kfree(udl_connector->edid);
 	kfree(connector);
 }
 
···
 
 int udl_connector_init(struct drm_device *dev, struct drm_encoder *encoder)
 {
+	struct udl_drm_connector *udl_connector;
 	struct drm_connector *connector;
 
-	connector = kzalloc(sizeof(struct drm_connector), GFP_KERNEL);
-	if (!connector)
+	udl_connector = kzalloc(sizeof(struct udl_drm_connector), GFP_KERNEL);
+	if (!udl_connector)
 		return -ENOMEM;
 
-	drm_connector_init(dev, connector, &udl_connector_funcs, DRM_MODE_CONNECTOR_DVII);
+	connector = &udl_connector->connector;
+	drm_connector_init(dev, connector, &udl_connector_funcs,
+			   DRM_MODE_CONNECTOR_DVII);
 	drm_connector_helper_add(connector, &udl_connector_helper_funcs);
 
 	drm_connector_register(connector);
 	drm_mode_connector_attach_encoder(connector, encoder);
+	connector->polled = DRM_CONNECTOR_POLL_HPD |
+		DRM_CONNECTOR_POLL_CONNECT | DRM_CONNECTOR_POLL_DISCONNECT;
 
 	return 0;
 }

drivers/gpu/drm/udl/udl_connector.h

···
+#ifndef __UDL_CONNECTOR_H__
+#define __UDL_CONNECTOR_H__
+
+#include <drm/drm_crtc.h>
+
+struct udl_drm_connector {
+	struct drm_connector connector;
+	/* last udl_detect edid */
+	struct edid *edid;
+};
+
+
+#endif //__UDL_CONNECTOR_H__

drivers/gpu/drm/udl/udl_drv.c

···
 static int udl_usb_suspend(struct usb_interface *interface,
 			   pm_message_t message)
 {
+	struct drm_device *dev = usb_get_intfdata(interface);
+
+	drm_kms_helper_poll_disable(dev);
 	return 0;
 }
 
···
 {
 	struct drm_device *dev = usb_get_intfdata(interface);
 
+	drm_kms_helper_poll_enable(dev);
 	udl_modeset_restore(dev);
 	return 0;
 }

drivers/gpu/drm/udl/udl_main.c

···
  * more details.
  */
 #include <drm/drmP.h>
+#include <drm/drm_crtc_helper.h>
 #include "udl_drv.h"
 
 /* -BULK_SIZE as per usb-skeleton. Can we get full page and avoid overhead? */
···
 	if (ret)
 		goto err_fb;
 
+	drm_kms_helper_poll_init(dev);
+
 	return 0;
 err_fb:
 	udl_fbdev_cleanup(dev);
···
 void udl_driver_unload(struct drm_device *dev)
 {
 	struct udl_device *udl = dev->dev_private;
+
+	drm_kms_helper_poll_fini(dev);
 
 	if (udl->urbs.count)
 		udl_free_urb_list(dev);

drivers/gpu/drm/vc4/vc4_bo.c

···
 			 vc4->bo_labels[i].size_allocated / 1024,
 			 vc4->bo_labels[i].num_allocated);
 	}
+
+	mutex_lock(&vc4->purgeable.lock);
+	if (vc4->purgeable.num)
+		DRM_INFO("%30s: %6zdkb BOs (%d)\n", "userspace BO cache",
+			 vc4->purgeable.size / 1024, vc4->purgeable.num);
+
+	if (vc4->purgeable.purged_num)
+		DRM_INFO("%30s: %6zdkb BOs (%d)\n", "total purged BO",
+			 vc4->purgeable.purged_size / 1024,
+			 vc4->purgeable.purged_num);
+	mutex_unlock(&vc4->purgeable.lock);
 }
 
 #ifdef CONFIG_DEBUG_FS
···
 			   vc4->bo_labels[i].num_allocated);
 	}
 	mutex_unlock(&vc4->bo_lock);
+
+	mutex_lock(&vc4->purgeable.lock);
+	if (vc4->purgeable.num)
+		seq_printf(m, "%30s: %6dkb BOs (%d)\n", "userspace BO cache",
+			   vc4->purgeable.size / 1024, vc4->purgeable.num);
+
+	if (vc4->purgeable.purged_num)
+		seq_printf(m, "%30s: %6dkb BOs (%d)\n", "total purged BO",
+			   vc4->purgeable.purged_size / 1024,
+			   vc4->purgeable.purged_num);
+	mutex_unlock(&vc4->purgeable.lock);
 
 	return 0;
 }
···
 	mutex_unlock(&vc4->bo_lock);
 }
 
+void vc4_bo_add_to_purgeable_pool(struct vc4_bo *bo)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev);
+
+	mutex_lock(&vc4->purgeable.lock);
+	list_add_tail(&bo->size_head, &vc4->purgeable.list);
+	vc4->purgeable.num++;
+	vc4->purgeable.size += bo->base.base.size;
+	mutex_unlock(&vc4->purgeable.lock);
+}
+
+static void vc4_bo_remove_from_purgeable_pool_locked(struct vc4_bo *bo)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev);
+
+	/* list_del_init() is used here because the caller might release
+	 * the purgeable lock in order to acquire the madv one and update the
+	 * madv status.
+	 * During this short period of time a user might decide to mark
+	 * the BO as unpurgeable, and if bo->madv is set to
+	 * VC4_MADV_DONTNEED it will try to remove the BO from the
+	 * purgeable list which will fail if the ->next/prev fields
+	 * are set to LIST_POISON1/LIST_POISON2 (which is what
+	 * list_del() does).
+	 * Re-initializing the list element guarantees that list_del()
+	 * will work correctly even if it's a NOP.
+	 */
+	list_del_init(&bo->size_head);
+	vc4->purgeable.num--;
+	vc4->purgeable.size -= bo->base.base.size;
+}
+
+void vc4_bo_remove_from_purgeable_pool(struct vc4_bo *bo)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(bo->base.base.dev);
+
+	mutex_lock(&vc4->purgeable.lock);
+	vc4_bo_remove_from_purgeable_pool_locked(bo);
+	mutex_unlock(&vc4->purgeable.lock);
+}
+
+static void vc4_bo_purge(struct drm_gem_object *obj)
+{
+	struct vc4_bo *bo = to_vc4_bo(obj);
+	struct drm_device *dev = obj->dev;
+
+	WARN_ON(!mutex_is_locked(&bo->madv_lock));
+	WARN_ON(bo->madv != VC4_MADV_DONTNEED);
+
+	drm_vma_node_unmap(&obj->vma_node, dev->anon_inode->i_mapping);
+
+	dma_free_wc(dev->dev, obj->size, bo->base.vaddr, bo->base.paddr);
+	bo->base.vaddr = NULL;
+	bo->madv = __VC4_MADV_PURGED;
+}
+
+static void vc4_bo_userspace_cache_purge(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	mutex_lock(&vc4->purgeable.lock);
+	while (!list_empty(&vc4->purgeable.list)) {
+		struct vc4_bo *bo = list_first_entry(&vc4->purgeable.list,
+						     struct vc4_bo, size_head);
+		struct drm_gem_object *obj = &bo->base.base;
+		size_t purged_size = 0;
+
+		vc4_bo_remove_from_purgeable_pool_locked(bo);
+
+		/* Release the purgeable lock while we're purging the BO so
+		 * that other people can continue inserting things in the
+		 * purgeable pool without having to wait for all BOs to be
+		 * purged.
+		 */
+		mutex_unlock(&vc4->purgeable.lock);
+		mutex_lock(&bo->madv_lock);
+
+		/* Since we released the purgeable pool lock before acquiring
+		 * the BO madv one, the user may have marked the BO as WILLNEED
+		 * and re-used it in the meantime.
+		 * Before purging the BO we need to make sure
+		 * - it is still marked as DONTNEED
+		 * - it has not been re-inserted in the purgeable list
+		 * - it is not used by HW blocks
+		 * If one of these conditions is not met, just skip the entry.
+		 */
+		if (bo->madv == VC4_MADV_DONTNEED &&
+		    list_empty(&bo->size_head) &&
+		    !refcount_read(&bo->usecnt)) {
+			purged_size = bo->base.base.size;
+			vc4_bo_purge(obj);
+		}
+		mutex_unlock(&bo->madv_lock);
+		mutex_lock(&vc4->purgeable.lock);
+
+		if (purged_size) {
+			vc4->purgeable.purged_size += purged_size;
+			vc4->purgeable.purged_num++;
+		}
+	}
+	mutex_unlock(&vc4->purgeable.lock);
+}
+
 static struct vc4_bo *vc4_bo_get_from_cache(struct drm_device *dev,
 					    uint32_t size,
 					    enum vc4_kernel_bo_type type)
···
 	if (!bo)
 		return ERR_PTR(-ENOMEM);
 
+	bo->madv = VC4_MADV_WILLNEED;
+	refcount_set(&bo->usecnt, 0);
+	mutex_init(&bo->madv_lock);
 	mutex_lock(&vc4->bo_lock);
 	bo->label = VC4_BO_TYPE_KERNEL;
 	vc4->bo_labels[VC4_BO_TYPE_KERNEL].num_allocated++;
···
 		 * CMA allocations we've got laying around and try again.
 		 */
 		vc4_bo_cache_purge(dev);
-
 		cma_obj = drm_gem_cma_create(dev, size);
-		if (IS_ERR(cma_obj)) {
-			DRM_ERROR("Failed to allocate from CMA:\n");
-			vc4_bo_stats_dump(vc4);
-			return ERR_PTR(-ENOMEM);
-		}
+	}
+
+	if (IS_ERR(cma_obj)) {
+		/*
+		 * Still not enough CMA memory, purge the userspace BO
+		 * cache and retry.
+		 * This is sub-optimal since we purge the whole userspace
+		 * BO cache which forces user that want to re-use the BO to
+		 * restore its initial content.
+		 * Ideally, we should purge entries one by one and retry
+		 * after each to see if CMA allocation succeeds. Or even
+		 * better, try to find an entry with at least the same
+		 * size.
+		 */
+		vc4_bo_userspace_cache_purge(dev);
+		cma_obj = drm_gem_cma_create(dev, size);
+	}
+
+	if (IS_ERR(cma_obj)) {
+		DRM_ERROR("Failed to allocate from CMA:\n");
+		vc4_bo_stats_dump(vc4);
+		return ERR_PTR(-ENOMEM);
 	}
 	bo = to_vc4_bo(&cma_obj->base);
+
+	/* By default, BOs do not support the MADV ioctl. This will be enabled
+	 * only on BOs that are exposed to userspace (V3D, V3D_SHADER and DUMB
+	 * BOs).
+	 */
+	bo->madv = __VC4_MADV_NOTSUPP;
 
 	mutex_lock(&vc4->bo_lock);
 	vc4_bo_set_label(&cma_obj->base, type);
···
 	bo = vc4_bo_create(dev, args->size, false, VC4_BO_TYPE_DUMB);
 	if (IS_ERR(bo))
 		return PTR_ERR(bo);
+
+	bo->madv = VC4_MADV_WILLNEED;
 
 	ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
 	drm_gem_object_put_unlocked(&bo->base.base);
···
 	struct vc4_bo *bo = to_vc4_bo(gem_bo);
 	struct list_head *cache_list;
 
+	/* Remove the BO from the purgeable list. */
+	mutex_lock(&bo->madv_lock);
+	if (bo->madv == VC4_MADV_DONTNEED && !refcount_read(&bo->usecnt))
+		vc4_bo_remove_from_purgeable_pool(bo);
+	mutex_unlock(&bo->madv_lock);
+
 	mutex_lock(&vc4->bo_lock);
 	/* If the object references someone else's memory, we can't cache it.
 	 */
···
 	}
 
 	/* If this object was partially constructed but CMA allocation
-	 * had failed, just free it.
+	 * had failed, just free it. Can also happen when the BO has been
+	 * purged.
 	 */
 	if (!bo->base.vaddr) {
 		vc4_bo_destroy(bo);
···
 		kfree(bo->validated_shader);
 		bo->validated_shader = NULL;
 	}
+
+	/* Reset madv and usecnt before adding the BO to the cache. */
+	bo->madv = __VC4_MADV_NOTSUPP;
+	refcount_set(&bo->usecnt, 0);
 
 	bo->t_format = false;
 	bo->free_time = jiffies;
···
 	mutex_unlock(&vc4->bo_lock);
 }
 
+int vc4_bo_inc_usecnt(struct vc4_bo *bo)
+{
+	int ret;
+
+	/* Fast path: if the BO is already retained by someone, no need to
+	 * check the madv status.
+	 */
+	if (refcount_inc_not_zero(&bo->usecnt))
+		return 0;
+
+	mutex_lock(&bo->madv_lock);
+	switch (bo->madv) {
+	case VC4_MADV_WILLNEED:
+		refcount_inc(&bo->usecnt);
+		ret = 0;
+		break;
+	case VC4_MADV_DONTNEED:
+		/* We shouldn't use a BO marked as purgeable if at least
+		 * someone else retained its content by incrementing usecnt.
+		 * Luckily the BO hasn't been purged yet, but something wrong
+		 * is happening here. Just throw an error instead of
+		 * authorizing this use case.
+		 */
+	case __VC4_MADV_PURGED:
+		/* We can't use a purged BO. */
+	default:
+		/* Invalid madv value. */
+		ret = -EINVAL;
+		break;
+	}
+	mutex_unlock(&bo->madv_lock);
+
+	return ret;
+}
+
+void vc4_bo_dec_usecnt(struct vc4_bo *bo)
+{
+	/* Fast path: if the BO is still retained by someone, no need to test
+	 * the madv value.
+	 */
+	if (refcount_dec_not_one(&bo->usecnt))
+		return;
+
+	mutex_lock(&bo->madv_lock);
+	if (refcount_dec_and_test(&bo->usecnt) &&
+	    bo->madv == VC4_MADV_DONTNEED)
+		vc4_bo_add_to_purgeable_pool(bo);
+	mutex_unlock(&bo->madv_lock);
+}
+
 static void vc4_bo_cache_time_timer(unsigned long data)
 {
 	struct drm_device *dev = (struct drm_device *)data;
···
 vc4_prime_export(struct drm_device *dev, struct drm_gem_object *obj, int flags)
 {
 	struct vc4_bo *bo = to_vc4_bo(obj);
+	struct dma_buf *dmabuf;
+	int ret;
 
 	if (bo->validated_shader) {
 		DRM_DEBUG("Attempting to export shader BO\n");
 		return ERR_PTR(-EINVAL);
 	}
 
-	return drm_gem_prime_export(dev, obj, flags);
+	/* Note: as soon as the BO is exported it becomes unpurgeable, because
+	 * noone ever decrements the usecnt even if the reference held by the
+	 * exported BO is released. This shouldn't be a problem since we don't
+	 * expect exported BOs to be marked as purgeable.
+	 */
+	ret = vc4_bo_inc_usecnt(bo);
+	if (ret) {
+		DRM_ERROR("Failed to increment BO usecnt\n");
+		return ERR_PTR(ret);
+	}
+
+	dmabuf = drm_gem_prime_export(dev, obj, flags);
+	if (IS_ERR(dmabuf))
+		vc4_bo_dec_usecnt(bo);
+
+	return dmabuf;
+}
+
+int vc4_fault(struct vm_fault *vmf)
+{
+	struct vm_area_struct *vma = vmf->vma;
+	struct drm_gem_object *obj = vma->vm_private_data;
+	struct vc4_bo *bo = to_vc4_bo(obj);
+
+	/* The only reason we would end up here is when user-space accesses
+	 * BO's memory after it's been purged.
+	 */
+	mutex_lock(&bo->madv_lock);
+	WARN_ON(bo->madv != __VC4_MADV_PURGED);
+	mutex_unlock(&bo->madv_lock);
+
+	return VM_FAULT_SIGBUS;
 }
 
 int vc4_mmap(struct file *filp, struct vm_area_struct *vma)
 {
 	struct drm_gem_object *gem_obj;
+	unsigned long vm_pgoff;
 	struct vc4_bo *bo;
 	int ret;
 
···
 		return -EINVAL;
 	}
 
+	if (bo->madv != VC4_MADV_WILLNEED) {
+		DRM_DEBUG("mmaping of %s BO not allowed\n",
+			  bo->madv == VC4_MADV_DONTNEED ?
+			  "purgeable" : "purged");
+		return -EINVAL;
+	}
+
 	/*
 	 * Clear the VM_PFNMAP flag that was set by drm_gem_mmap(), and set the
 	 * vm_pgoff (used as a fake buffer offset by DRM) to 0 as we want to map
 	 * the whole buffer.
 	 */
 	vma->vm_flags &= ~VM_PFNMAP;
-	vma->vm_pgoff = 0;
 
+	/* This ->vm_pgoff dance is needed to make all parties happy:
+	 * - dma_mmap_wc() uses ->vm_pgoff as an offset within the allocated
+	 *   mem-region, hence the need to set it to zero (the value set by
+	 *   the DRM core is a virtual offset encoding the GEM object-id)
+	 * - the mmap() core logic needs ->vm_pgoff to be restored to its
+	 *   initial value before returning from this function because it
+	 *   encodes the  offset of this GEM in the dev->anon_inode pseudo-file
+	 *   and this information will be used when we invalidate userspace
+	 *   mappings  with drm_vma_node_unmap() (called from vc4_gem_purge()).
+	 */
+	vm_pgoff = vma->vm_pgoff;
+	vma->vm_pgoff = 0;
 	ret = dma_mmap_wc(bo->base.base.dev->dev, vma, bo->base.vaddr,
 			  bo->base.paddr, vma->vm_end - vma->vm_start);
+	vma->vm_pgoff = vm_pgoff;
+
 	if (ret)
 		drm_gem_vm_close(vma);
 
···
 	if (IS_ERR(bo))
 		return PTR_ERR(bo);
 
+	bo->madv = VC4_MADV_WILLNEED;
+
 	ret = drm_gem_handle_create(file_priv, &bo->base.base, &args->handle);
 	drm_gem_object_put_unlocked(&bo->base.base);
 
···
 	bo = vc4_bo_create(dev, args->size, true, VC4_BO_TYPE_V3D_SHADER);
 	if (IS_ERR(bo))
 		return PTR_ERR(bo);
+
+	bo->madv = VC4_MADV_WILLNEED;
 
 	if (copy_from_user(bo->base.vaddr,
 			     (void __user *)(uintptr_t)args->data,

drivers/gpu/drm/vc4/vc4_drv.c

···
 	case DRM_VC4_PARAM_SUPPORTS_ETC1:
 	case DRM_VC4_PARAM_SUPPORTS_THREADED_FS:
 	case DRM_VC4_PARAM_SUPPORTS_FIXED_RCL_ORDER:
+	case DRM_VC4_PARAM_SUPPORTS_MADVISE:
 		args->value = true;
 		break;
 	default:
···
 
 	drm_fbdev_cma_restore_mode(vc4->fbdev);
 }
+
+static const struct vm_operations_struct vc4_vm_ops = {
+	.fault = vc4_fault,
+	.open = drm_gem_vm_open,
+	.close = drm_gem_vm_close,
+};
 
 static const struct file_operations vc4_drm_fops = {
 	.owner = THIS_MODULE,
···
 	DRM_IOCTL_DEF_DRV(VC4_SET_TILING, vc4_set_tiling_ioctl, DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(VC4_GET_TILING, vc4_get_tiling_ioctl, DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(VC4_LABEL_BO, vc4_label_bo_ioctl, DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(VC4_GEM_MADVISE, vc4_gem_madvise_ioctl, DRM_RENDER_ALLOW),
 };
 
 static struct drm_driver vc4_drm_driver = {
···
 
 	.gem_create_object = vc4_create_object,
 	.gem_free_object_unlocked = vc4_free_object,
-	.gem_vm_ops = &drm_gem_cma_vm_ops,
+	.gem_vm_ops = &vc4_vm_ops,
 
 	.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
 	.prime_fd_to_handle = drm_gem_prime_fd_to_handle,

drivers/gpu/drm/vc4/vc4_drv.h

···
 	/* Protects bo_cache and bo_labels. */
 	struct mutex bo_lock;
 
+	/* Purgeable BO pool. All BOs in this pool can have their memory
+	 * reclaimed if the driver is unable to allocate new BOs. We also
+	 * keep stats related to the purge mechanism here.
+	 */
+	struct {
+		struct list_head list;
+		unsigned int num;
+		size_t size;
+		unsigned int purged_num;
+		size_t purged_size;
+		struct mutex lock;
+	} purgeable;
+
 	uint64_t dma_fence_context;
 
 	/* Sequence number for the last job queued in bin_job_list.
···
 	 * for user-allocated labels.
 	 */
 	int label;
+
+	/* Count the number of active users. This is needed to determine
+	 * whether we can move the BO to the purgeable list or not (when the BO
+	 * is used by the GPU or the display engine we can't purge it).
+	 */
+	refcount_t usecnt;
+
+	/* Store purgeable/purged state here */
+	u32 madv;
+	struct mutex madv_lock;
 };
 
 static inline struct vc4_bo *
···
 			     struct drm_file *file_priv);
 int vc4_label_bo_ioctl(struct drm_device *dev, void *data,
 		       struct drm_file *file_priv);
+int vc4_fault(struct vm_fault *vmf);
 int vc4_mmap(struct file *filp, struct vm_area_struct *vma);
 struct reservation_object *vc4_prime_res_obj(struct drm_gem_object *obj);
 int vc4_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma);
···
 int vc4_bo_cache_init(struct drm_device *dev);
 void vc4_bo_cache_destroy(struct drm_device *dev);
 int vc4_bo_stats_debugfs(struct seq_file *m, void *arg);
+int vc4_bo_inc_usecnt(struct vc4_bo *bo);
+void vc4_bo_dec_usecnt(struct vc4_bo *bo);
+void vc4_bo_add_to_purgeable_pool(struct vc4_bo *bo);
+void vc4_bo_remove_from_purgeable_pool(struct vc4_bo *bo);
 
 /* vc4_crtc.c */
 extern struct platform_driver vc4_crtc_driver;
···
 int vc4_queue_seqno_cb(struct drm_device *dev,
 		       struct vc4_seqno_cb *cb, uint64_t seqno,
 		       void (*func)(struct vc4_seqno_cb *cb));
+int vc4_gem_madvise_ioctl(struct drm_device *dev, void *data,
+			  struct drm_file *file_priv);
 
 /* vc4_hdmi.c */
 extern struct platform_driver vc4_hdmi_driver;

drivers/gpu/drm/vc4/vc4_dsi.c

···
 	*ret = IRQ_HANDLED;
 }
 
+/*
+ * Initial handler for port 1 where we need the reg_dma workaround.
+ * The register DMA writes sleep, so we can't do it in the top half.
+ * Instead we use IRQF_ONESHOT so that the IRQ gets disabled in the
+ * parent interrupt contrller until our interrupt thread is done.
+ */
+static irqreturn_t vc4_dsi_irq_defer_to_thread_handler(int irq, void *data)
+{
+	struct vc4_dsi *dsi = data;
+	u32 stat = DSI_PORT_READ(INT_STAT);
+
+	if (!stat)
+		return IRQ_NONE;
+
+	return IRQ_WAKE_THREAD;
+}
+
+/*
+ * Normal IRQ handler for port 0, or the threaded IRQ handler for port
+ * 1 where we need the reg_dma workaround.
+ */
 static irqreturn_t vc4_dsi_irq_handler(int irq, void *data)
 {
 	struct vc4_dsi *dsi = data;
···
 	/* Clear any existing interrupt state. */
 	DSI_PORT_WRITE(INT_STAT, DSI_PORT_READ(INT_STAT));
 
-	ret = devm_request_irq(dev, platform_get_irq(pdev, 0),
-			       vc4_dsi_irq_handler, 0, "vc4 dsi", dsi);
+	if (dsi->reg_dma_mem)
+		ret = devm_request_threaded_irq(dev, platform_get_irq(pdev, 0),
+						vc4_dsi_irq_defer_to_thread_handler,
+						vc4_dsi_irq_handler,
+						IRQF_ONESHOT,
+						"vc4 dsi", dsi);
+	else
+		ret = devm_request_irq(dev, platform_get_irq(pdev, 0),
+				       vc4_dsi_irq_handler, 0, "vc4 dsi", dsi);
 	if (ret) {
 		if (ret != -EPROBE_DEFER)
 			dev_err(dev, "Failed to get interrupt: %d\n", ret);

drivers/gpu/drm/vc4/vc4_gem.c

···
 			continue;
 
 		for (j = 0; j < exec[i]->bo_count; j++) {
+			bo = to_vc4_bo(&exec[i]->bo[j]->base);
+
+			/* Retain BOs just in case they were marked purgeable.
+			 * This prevents the BO from being purged before
+			 * someone had a chance to dump the hang state.
+			 */
+			WARN_ON(!refcount_read(&bo->usecnt));
+			refcount_inc(&bo->usecnt);
 			drm_gem_object_get(&exec[i]->bo[j]->base);
 			kernel_state->bo[j + prev_idx] = &exec[i]->bo[j]->base;
 		}
 
 		list_for_each_entry(bo, &exec[i]->unref_list, unref_head) {
+			/* No need to retain BOs coming from the ->unref_list
+			 * because they are naturally unpurgeable.
+			 */
 			drm_gem_object_get(&bo->base.base);
 			kernel_state->bo[j + prev_idx] = &bo->base.base;
 			j++;
···
 	state->fdbgr = V3D_READ(V3D_FDBGR);
 	state->fdbgs = V3D_READ(V3D_FDBGS);
 	state->errstat = V3D_READ(V3D_ERRSTAT);
+
+	/* We need to turn purgeable BOs into unpurgeable ones so that
+	 * userspace has a chance to dump the hang state before the kernel
+	 * decides to purge those BOs.
+	 * Note that BO consistency at dump time cannot be guaranteed. For
+	 * example, if the owner of these BOs decides to re-use them or mark
+	 * them purgeable again there's nothing we can do to prevent it.
+	 */
+	for (i = 0; i < kernel_state->user_state.bo_count; i++) {
+		struct vc4_bo *bo = to_vc4_bo(kernel_state->bo[i]);
+
+		if (bo->madv == __VC4_MADV_NOTSUPP)
+			continue;
+
+		mutex_lock(&bo->madv_lock);
+		if (!WARN_ON(bo->madv == __VC4_MADV_PURGED))
+			bo->madv = VC4_MADV_WILLNEED;
+		refcount_dec(&bo->usecnt);
+		mutex_unlock(&bo->madv_lock);
+	}
 
 	spin_lock_irqsave(&vc4->job_lock, irqflags);
 	if (vc4->hang_state) {
···
  * The command validator needs to reference BOs by their index within
  * the submitted job's BO list.  This does the validation of the job's
  * BO list and reference counting for the lifetime of the job.
- *
- * Note that this function doesn't need to unreference the BOs on
- * failure, because that will happen at vc4_complete_exec() time.
  */
 static int
 vc4_cl_lookup_bos(struct drm_device *dev,
···
 			DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
 				  i, handles[i]);
 			ret = -EINVAL;
-			spin_unlock(&file_priv->table_lock);
-			goto fail;
+			break;
 		}
+
 		drm_gem_object_get(bo);
 		exec->bo[i] = (struct drm_gem_cma_object *)bo;
 	}
 	spin_unlock(&file_priv->table_lock);
 
+	if (ret)
+		goto fail_put_bo;
+
+	for (i = 0; i < exec->bo_count; i++) {
+		ret = vc4_bo_inc_usecnt(to_vc4_bo(&exec->bo[i]->base));
+		if (ret)
+			goto fail_dec_usecnt;
+	}
+
+	kvfree(handles);
+	return 0;
+
+fail_dec_usecnt:
+	/* Decrease usecnt on acquired objects.
+	 * We cannot rely on  vc4_complete_exec() to release resources here,
+	 * because vc4_complete_exec() has no information about which BO has
+	 * had its ->usecnt incremented.
+	 * To make things easier we just free everything explicitly and set
+	 * exec->bo to NULL so that vc4_complete_exec() skips the 'BO release'
+	 * step.
+	 */
+	for (i-- ; i >= 0; i--)
+		vc4_bo_dec_usecnt(to_vc4_bo(&exec->bo[i]->base));
+
+fail_put_bo:
+	/* Release any reference to acquired objects. */
+	for (i = 0; i < exec->bo_count && exec->bo[i]; i++)
+		drm_gem_object_put_unlocked(&exec->bo[i]->base);
+
 fail:
 	kvfree(handles);
+	kvfree(exec->bo);
+	exec->bo = NULL;
 	return ret;
 }
 
···
 		dma_fence_signal(exec->fence);
 
 	if (exec->bo) {
-		for (i = 0; i < exec->bo_count; i++)
+		for (i = 0; i < exec->bo_count; i++) {
+			struct vc4_bo *bo = to_vc4_bo(&exec->bo[i]->base);
+
+			vc4_bo_dec_usecnt(bo);
 			drm_gem_object_put_unlocked(&exec->bo[i]->base);
+		}
 		kvfree(exec->bo);
 	}
 
···
 	INIT_WORK(&vc4->job_done_work, vc4_job_done_work);
 
 	mutex_init(&vc4->power_lock);
+
+	INIT_LIST_HEAD(&vc4->purgeable.list);
+	mutex_init(&vc4->purgeable.lock);
 }
 
 void
···
 
 	if (vc4->hang_state)
 		vc4_free_hang_state(dev, vc4->hang_state);
+}
+
+int vc4_gem_madvise_ioctl(struct drm_device *dev, void *data,
+			  struct drm_file *file_priv)
+{
+	struct drm_vc4_gem_madvise *args = data;
+	struct drm_gem_object *gem_obj;
+	struct vc4_bo *bo;
+	int ret;
+
+	switch (args->madv) {
+	case VC4_MADV_DONTNEED:
+	case VC4_MADV_WILLNEED:
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	if (args->pad != 0)
+		return -EINVAL;
+
+	gem_obj = drm_gem_object_lookup(file_priv, args->handle);
+	if (!gem_obj) {
+		DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
+		return -ENOENT;
+	}
+
+	bo = to_vc4_bo(gem_obj);
+
+	/* Only BOs exposed to userspace can be purged. */
+	if (bo->madv == __VC4_MADV_NOTSUPP) {
+		DRM_DEBUG("madvise not supported on this BO\n");
+		ret = -EINVAL;
+		goto out_put_gem;
+	}
+
+	/* Not sure it's safe to purge imported BOs. Let's just assume it's
+	 * not until proven otherwise.
+	 */
+	if (gem_obj->import_attach) {
+		DRM_DEBUG("madvise not supported on imported BOs\n");
+		ret = -EINVAL;
+		goto out_put_gem;
+	}
+
+	mutex_lock(&bo->madv_lock);
+
+	if (args->madv == VC4_MADV_DONTNEED && bo->madv == VC4_MADV_WILLNEED &&
+	    !refcount_read(&bo->usecnt)) {
+		/* If the BO is about to be marked as purgeable, is not used
+		 * and is not already purgeable or purged, add it to the
+		 * purgeable list.
+		 */
+		vc4_bo_add_to_purgeable_pool(bo);
+	} else if (args->madv == VC4_MADV_WILLNEED &&
+		   bo->madv == VC4_MADV_DONTNEED &&
+		   !refcount_read(&bo->usecnt)) {
+		/* The BO has not been purged yet, just remove it from
+		 * the purgeable list.
+		 */
+		vc4_bo_remove_from_purgeable_pool(bo);
+	}
+
+	/* Save the purged state. */
+	args->retained = bo->madv != __VC4_MADV_PURGED;
+
+	/* Update internal madv state only if the bo was not purged. */
+	if (bo->madv != __VC4_MADV_PURGED)
+		bo->madv = args->madv;
+
+	mutex_unlock(&bo->madv_lock);
+
+	ret = 0;
+
+out_put_gem:
+	drm_gem_object_put_unlocked(gem_obj);
+
+	return ret;
 }

drivers/gpu/drm/vc4/vc4_plane.c

···
 #include <drm/drm_fb_cma_helper.h>
 #include <drm/drm_plane_helper.h>
 
+#include "uapi/drm/vc4_drm.h"
 #include "vc4_drv.h"
 #include "vc4_regs.h"
 
···
 {
 	struct vc4_bo *bo;
 	struct dma_fence *fence;
+	int ret;
 
 	if ((plane->state->fb == state->fb) || !state->fb)
 		return 0;
 
 	bo = to_vc4_bo(&drm_fb_cma_get_gem_obj(state->fb, 0)->base);
+
+	ret = vc4_bo_inc_usecnt(bo);
+	if (ret)
+		return ret;
+
 	fence = reservation_object_get_excl_rcu(bo->resv);
 	drm_atomic_set_fence_for_plane(state, fence);
 
 	return 0;
 }
 
+static void vc4_cleanup_fb(struct drm_plane *plane,
+			   struct drm_plane_state *state)
+{
+	struct vc4_bo *bo;
+
+	if (plane->state->fb == state->fb || !state->fb)
+		return;
+
+	bo = to_vc4_bo(&drm_fb_cma_get_gem_obj(state->fb, 0)->base);
+	vc4_bo_dec_usecnt(bo);
+}
+
 static const struct drm_plane_helper_funcs vc4_plane_helper_funcs = {
 	.atomic_check = vc4_plane_atomic_check,
 	.atomic_update = vc4_plane_atomic_update,
 	.prepare_fb = vc4_prepare_fb,
+	.cleanup_fb = vc4_cleanup_fb,
 };
 
 static void vc4_plane_destroy(struct drm_plane *plane)

drivers/gpu/vga/vgaarb.c

···
 	MISC_DYNAMIC_MINOR, "vga_arbiter", &vga_arb_device_fops
 };
 
-static int __init vga_arb_device_init(void)
+static void __init vga_arb_select_default_device(void)
 {
-	int rc;
 	struct pci_dev *pdev;
 	struct vga_device *vgadev;
 
-	rc = misc_register(&vga_arb_device);
-	if (rc < 0)
-		pr_err("error %d registering device\n", rc);
-
-	bus_register_notifier(&pci_bus_type, &pci_notifier);
-
-	/* We add all pci devices satisfying vga class in the arbiter by
-	 * default */
-	pdev = NULL;
-	while ((pdev =
-		pci_get_subsys(PCI_ANY_ID, PCI_ANY_ID, PCI_ANY_ID,
-			       PCI_ANY_ID, pdev)) != NULL)
-		vga_arbiter_add_pci_device(pdev);
-
+#if defined(CONFIG_X86) || defined(CONFIG_IA64)
 	list_for_each_entry(vgadev, &vga_list, list) {
 		struct device *dev = &vgadev->pdev->dev;
-#if defined(CONFIG_X86) || defined(CONFIG_IA64)
 		/*
 		 * Override vga_arbiter_add_pci_device()'s I/O based detection
 		 * as it may take the wrong device (e.g. on Apple system under
···
 				vgaarb_info(dev, "overriding boot device\n");
 			vga_set_default_device(vgadev->pdev);
 		}
+	}
 #endif
+
+	if (!vga_default_device()) {
+		list_for_each_entry(vgadev, &vga_list, list) {
+			struct device *dev = &vgadev->pdev->dev;
+			u16 cmd;
+
+			pdev = vgadev->pdev;
+			pci_read_config_word(pdev, PCI_COMMAND, &cmd);
+			if (cmd & (PCI_COMMAND_IO | PCI_COMMAND_MEMORY)) {
+				vgaarb_info(dev, "setting as boot device (VGA legacy resources not available)\n");
+				vga_set_default_device(pdev);
+				break;
+			}
+		}
+	}
+
+	if (!vga_default_device()) {
+		vgadev = list_first_entry_or_null(&vga_list,
+						  struct vga_device, list);
+		if (vgadev) {
+			struct device *dev = &vgadev->pdev->dev;
+			vgaarb_info(dev, "setting as boot device (VGA legacy resources not available)\n");
+			vga_set_default_device(vgadev->pdev);
+		}
+	}
+}
+
+static int __init vga_arb_device_init(void)
+{
+	int rc;
+	struct pci_dev *pdev;
+	struct vga_device *vgadev;
+
+	rc = misc_register(&vga_arb_device);
+	if (rc < 0)
+		pr_err("error %d registering device\n", rc);
+
+	bus_register_notifier(&pci_bus_type, &pci_notifier);
+
+	/* We add all PCI devices satisfying VGA class in the arbiter by
+	 * default */
+	pdev = NULL;
+	while ((pdev =
+		pci_get_subsys(PCI_ANY_ID, PCI_ANY_ID, PCI_ANY_ID,
+			       PCI_ANY_ID, pdev)) != NULL)
+		vga_arbiter_add_pci_device(pdev);
+
+	list_for_each_entry(vgadev, &vga_list, list) {
+		struct device *dev = &vgadev->pdev->dev;
+
 		if (vgadev->bridge_has_one_vga)
 			vgaarb_info(dev, "bridge control possible\n");
 		else
 			vgaarb_info(dev, "no bridge control possible\n");
 	}
+
+	vga_arb_select_default_device();
 
 	pr_info("loaded\n");
 	return rc;

include/linux/dma-fence.h

···
  * implementation know that there is another driver waiting on
  * the signal (ie. hw->sw case).
  *
- * This function can be called called from atomic context, but not
+ * This function can be called from atomic context, but not
  * from irq context, so normal spinlocks can be used.
  *
  * A return value of false indicates the fence already passed,

include/uapi/drm/vc4_drm.h

···
 #define DRM_VC4_SET_TILING                        0x08
 #define DRM_VC4_GET_TILING                        0x09
 #define DRM_VC4_LABEL_BO                          0x0a
+#define DRM_VC4_GEM_MADVISE                       0x0b
 
 #define DRM_IOCTL_VC4_SUBMIT_CL           DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_SUBMIT_CL, struct drm_vc4_submit_cl)
 #define DRM_IOCTL_VC4_WAIT_SEQNO          DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_WAIT_SEQNO, struct drm_vc4_wait_seqno)
···
 #define DRM_IOCTL_VC4_SET_TILING          DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_SET_TILING, struct drm_vc4_set_tiling)
 #define DRM_IOCTL_VC4_GET_TILING          DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_GET_TILING, struct drm_vc4_get_tiling)
 #define DRM_IOCTL_VC4_LABEL_BO            DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_LABEL_BO, struct drm_vc4_label_bo)
+#define DRM_IOCTL_VC4_GEM_MADVISE         DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_GEM_MADVISE, struct drm_vc4_gem_madvise)
 
 struct drm_vc4_submit_rcl_surface {
 	__u32 hindex; /* Handle index, or ~0 if not present. */
···
 #define DRM_VC4_PARAM_SUPPORTS_ETC1		4
 #define DRM_VC4_PARAM_SUPPORTS_THREADED_FS	5
 #define DRM_VC4_PARAM_SUPPORTS_FIXED_RCL_ORDER	6
+#define DRM_VC4_PARAM_SUPPORTS_MADVISE		7
 
 struct drm_vc4_get_param {
 	__u32 param;
···
 	__u32 handle;
 	__u32 len;
 	__u64 name;
+};
+
+/*
+ * States prefixed with '__' are internal states and cannot be passed to the
+ * DRM_IOCTL_VC4_GEM_MADVISE ioctl.
+ */
+#define VC4_MADV_WILLNEED			0
+#define VC4_MADV_DONTNEED			1
+#define __VC4_MADV_PURGED			2
+#define __VC4_MADV_NOTSUPP			3
+
+struct drm_vc4_gem_madvise {
+	__u32 handle;
+	__u32 madv;
+	__u32 retained;
+	__u32 pad;
 };
 
 #if defined(__cplusplus)