Only allocate DRI2 pixmaps with GEM

The driver is currently set up (via EXA) to handle all pixmap allocations
itself, and it passes them to GEM.

However, running non-3D apps tends to generate a large number of small
pixmap allocations that we're never going to pass to GL, and it is
expensive and pointless to send them through GEM in this case.

To solve this, fbturbo and my earlier armsoc work for r3p2 strips out EXA,
falling back to X's fast default implementation of pixmap allocation in
the fb layer, then we would migrate such pixmaps to GEM the first time
DRI2 interaction is requested.

However, especially with growing interest in using the CPU cache for a
future performance increase, EXA's PrepareAccess and FinishAccess hooks
are looking more and more useful. In order to use them, we also have to
take over allocation of all pixmaps, but we can do it better: we can
use usage_hint to determine which pixmaps are backing buffers for windows
(hence likely candidates for DRI2 interaction) and which aren't.

The only remaining complication is to detect which pixmap is going to be
used for scanout, this is a bit hacky but it seems to always be the first
pixmap created after ScreenInit.

Performance differences (at 1080p):

                |  BEFORE   BEFORE  |  AFTER   AFTER
                |  RAW X    MUTTER  |  RAW X   MUTTER
----------------+-------------------+----------------
glxgears (fps)  |  148     140      |  148     137
es2gears (fps)  |  313     290      |  313     295
glmark-es2      |  91      68       |  91      73
gtkperf (time)  |  10.7    13.6     |  7.3     9.7

src/armsoc_dri2.c

@@ -141,7 +141,7 @@ static PixmapPtr
 createpix(DrawablePtr pDraw)
 {
 	ScreenPtr pScreen = pDraw->pScreen;
-	int flags = canflip(pDraw) ? ARMSOC_CREATE_PIXMAP_SCANOUT : 0;
+	int flags = canflip(pDraw) ? ARMSOC_CREATE_PIXMAP_SCANOUT : CREATE_PIXMAP_USAGE_BACKING_PIXMAP;
 	return pScreen->CreatePixmap(pScreen,
 			pDraw->width, pDraw->height, pDraw->depth, flags);
 }

src/armsoc_driver.c

@@ -989,6 +989,8 @@ ARMSOCScreenInit(SCREEN_INIT_ARGS_DECL)
 
 	TRACE_ENTER();
 
+	pARMSOC->created_scanout_pixmap = FALSE;
+
 	/* set drm master before allocating scanout buffer */
 	if (ARMSOCSetDRMMaster()) {
 		ERROR_MSG("Cannot get DRM master: %s", strerror(errno));

src/armsoc_driver.h

@@ -180,6 +180,11 @@ struct ARMSOCRec {
 	/* Size of the swap chain. Set to 1 if DRI2SwapLimit unsupported,
 	 * driNumBufs if early display enabled, otherwise driNumBufs-1 */
 	unsigned int                       swap_chain_size;
+
+	/* The first CreatePixmap after ScreenInit ends up being the
+	 * scanout, but we don't get any usage hint indicating that it should
+	 * be accelerated. Use a flag to detect this and act accordingly. */
+	Bool				created_scanout_pixmap;
 };
 
 /*

src/armsoc_exa.c

@@ -35,6 +35,16 @@
 #include <sys/ioctl.h>
 #include <unistd.h>
 
+static Bool is_accel_pixmap(struct ARMSOCPixmapPrivRec *priv)
+{
+	/* For pixmaps that are scanout or backing for windows, we
+	 * "accelerate" them by allocating them via GEM. For all other
+	 * pixmaps (where we never expect DRI2 CreateBuffer to be called), we
+	 * just malloc them, which turns out to be much faster.
+	 */
+	return priv->usage_hint == ARMSOC_CREATE_PIXMAP_SCANOUT || priv->usage_hint == CREATE_PIXMAP_USAGE_BACKING_PIXMAP;
+}
+
 /* keep this here, instead of static-inline so submodule doesn't
  * need to know layout of ARMSOCRec.
  */
@@ -75,20 +85,40 @@ ARMSOCPixmapExchange(PixmapPtr a, PixmapPtr b)
 	}
 }
 
-_X_EXPORT void *
-ARMSOCCreatePixmap2(ScreenPtr pScreen, int width, int height,
-		int depth, int usage_hint, int bitsPerPixel,
+static void *
+CreateNoAccelPixmap(struct ARMSOCPixmapPrivRec *priv, ScreenPtr pScreen, int width, int height,
+		int depth, int bitsPerPixel,
+		int *new_fb_pitch)
+{
+	ScrnInfoPtr pScrn = xf86ScreenToScrn(pScreen);
+
+	if (width > 0 && height > 0 && depth > 0 && bitsPerPixel > 0) {
+		int pitch = ((width * bitsPerPixel + FB_MASK) >> FB_SHIFT) * sizeof(FbBits);
+		size_t datasize = pitch * height;
+		priv->unaccel = malloc(datasize);
+
+		if (!priv->unaccel) {
+			ERROR_MSG("failed to allocate %dx%d mem", width, height);
+			free(priv);
+			return NULL;
+		}
+		priv->unaccel_size = datasize;
+		*new_fb_pitch = pitch;
+	}
+
+	return priv;
+}
+
+static void *
+CreateAccelPixmap(struct ARMSOCPixmapPrivRec *priv, ScreenPtr pScreen, int width, int height,
+		int depth, int bitsPerPixel,
 		int *new_fb_pitch)
 {
-	struct ARMSOCPixmapPrivRec *priv = calloc(1, sizeof(*priv));
 	ScrnInfoPtr pScrn = xf86ScreenToScrn(pScreen);
 	struct ARMSOCRec *pARMSOC = ARMSOCPTR(pScrn);
 	enum armsoc_buf_type buf_type = ARMSOC_BO_NON_SCANOUT;
 
-	if (!priv)
-		return NULL;
-
-	if (usage_hint & ARMSOC_CREATE_PIXMAP_SCANOUT)
+	if (priv->usage_hint == ARMSOC_CREATE_PIXMAP_SCANOUT)
 		buf_type = ARMSOC_BO_SCANOUT;
 
 	if (width > 0 && height > 0 && depth > 0 && bitsPerPixel > 0) {
@@ -123,15 +153,32 @@ ARMSOCCreatePixmap2(ScreenPtr pScreen, int width, int height,
 		*new_fb_pitch = armsoc_bo_pitch(priv->bo);
 	}
 
-	/* The usage_hint field of the Pixmap passed to ModifyPixmapHeader is
-	 * not set to the usage_hint parameter passed to CreatePixmap.
-	 * It does appear to be set here so we stash it in the private
-	 * structure. However as we do not fully understand the uses of this
-	 * parameter, beware of any unexpected values!
-	 */
+	return priv;
+}
+
+_X_EXPORT void *
+ARMSOCCreatePixmap2(ScreenPtr pScreen, int width, int height,
+		int depth, int usage_hint, int bitsPerPixel,
+		int *new_fb_pitch)
+{
+	struct ARMSOCPixmapPrivRec *priv =
+				calloc(sizeof(struct ARMSOCPixmapPrivRec), 1);
+	ScrnInfoPtr pScrn = xf86ScreenToScrn(pScreen);
+	struct ARMSOCRec *pARMSOC = ARMSOCPTR(pScrn);
+
+	if (!priv)
+		return NULL;
+
+	if (!pARMSOC->created_scanout_pixmap) {
+		usage_hint = ARMSOC_CREATE_PIXMAP_SCANOUT;
+		pARMSOC->created_scanout_pixmap = TRUE;
+	}
 	priv->usage_hint = usage_hint;
 
-	return priv;
+	if (is_accel_pixmap(priv))
+		return CreateAccelPixmap(priv, pScreen, width, height, depth, bitsPerPixel, new_fb_pitch);
+	else
+		return CreateNoAccelPixmap(priv, pScreen, width, height, depth, bitsPerPixel, new_fb_pitch);
 }
 
 _X_EXPORT void
@@ -149,15 +196,87 @@ ARMSOCDestroyPixmap(ScreenPtr pScreen, void *driverPriv)
 		armsoc_bo_unreference(priv->bo);
 	}
 
+	if (priv->unaccel)
+		free(priv->unaccel);
+
 	free(priv);
 }
 
-_X_EXPORT Bool
-ARMSOCModifyPixmapHeader(PixmapPtr pPixmap, int width, int height,
+static Bool
+ModifyUnAccelPixmapHeader(struct ARMSOCPixmapPrivRec *priv, PixmapPtr pPixmap, int width, int height,
+		int depth, int bitsPerPixel, int devKind,
+		pointer pPixData)
+{
+	ScrnInfoPtr pScrn = pix2scrn(pPixmap);
+	size_t datasize;
+
+	/* Only modify specified fields, keeping all others intact. */
+	if (pPixData)
+		pPixmap->devPrivate.ptr = pPixData;
+
+	if (devKind > 0)
+		pPixmap->devKind = devKind;
+
+	/*
+	 * Someone is messing with the memory allocation. Let's step out of
+	 * the picture.
+	 */
+	if (pPixData && pPixData != priv->unaccel) {
+		if (priv->unaccel)
+			free(priv->unaccel);
+		priv->unaccel = NULL;
+		priv->unaccel_size = 0;
+
+		/* Returning FALSE calls miModifyPixmapHeader */
+		return FALSE;
+	}
+
+	if (depth > 0)
+		pPixmap->drawable.depth = depth;
+
+	if (bitsPerPixel > 0)
+		pPixmap->drawable.bitsPerPixel = bitsPerPixel;
+
+	if (width > 0)
+		pPixmap->drawable.width = width;
+
+	if (height > 0)
+		pPixmap->drawable.height = height;
+
+	/*
+	 * X will sometimes create an empty pixmap (width/height == 0) and then
+	 * use ModifyPixmapHeader to point it at PixData. We'll hit this path
+	 * during the CreatePixmap call. Just return true and skip the allocate
+	 * in this case.
+	 */
+	if (!pPixmap->drawable.width || !pPixmap->drawable.height)
+		return TRUE;
+
+	datasize = devKind * height;
+	if (!priv->unaccel || priv->unaccel_size != datasize) {
+		/* re-allocate buffer! */
+		if (priv->unaccel)
+			free(priv->unaccel);
+		priv->unaccel = malloc(datasize);
+
+		if (!priv->unaccel) {
+			ERROR_MSG("failed to allocate %d bytes mem",
+					datasize);
+			priv->unaccel_size = 0;
+			return FALSE;
+		}
+		priv->unaccel_size = datasize;
+	}
+
+	return TRUE;
+}
+
+
+static Bool
+ModifyAccelPixmapHeader(struct ARMSOCPixmapPrivRec *priv, PixmapPtr pPixmap, int width, int height,
 		int depth, int bitsPerPixel, int devKind,
 		pointer pPixData)
 {
-	struct ARMSOCPixmapPrivRec *priv = exaGetPixmapDriverPrivate(pPixmap);
 	ScrnInfoPtr pScrn = pix2scrn(pPixmap);
 	struct ARMSOCRec *pARMSOC = ARMSOCPTR(pScrn);
 	enum armsoc_buf_type buf_type = ARMSOC_BO_NON_SCANOUT;
@@ -201,7 +320,7 @@ ARMSOCModifyPixmapHeader(PixmapPtr pPixmap, int width, int height,
 		}
 	}
 
-	if (priv->usage_hint & ARMSOC_CREATE_PIXMAP_SCANOUT)
+	if (priv->usage_hint == ARMSOC_CREATE_PIXMAP_SCANOUT)
 		buf_type = ARMSOC_BO_SCANOUT;
 
 	if (depth > 0)
@@ -265,6 +384,18 @@ ARMSOCModifyPixmapHeader(PixmapPtr pPixmap, int width, int height,
 	return TRUE;
 }
 
+_X_EXPORT Bool
+ARMSOCModifyPixmapHeader(PixmapPtr pPixmap, int width, int height,
+		int depth, int bitsPerPixel, int devKind,
+		pointer pPixData)
+{
+	struct ARMSOCPixmapPrivRec *priv = exaGetPixmapDriverPrivate(pPixmap);
+	if (is_accel_pixmap(priv))
+		return ModifyAccelPixmapHeader(priv, pPixmap, width, height, depth, bitsPerPixel, devKind, pPixData);
+	else
+		return ModifyUnAccelPixmapHeader(priv, pPixmap, width, height, depth, bitsPerPixel, devKind, pPixData);
+}
+
 /**
  * WaitMarker is a required EXA callback but synchronization is
  * performed during ARMSOCPrepareAccess so this function does not
@@ -335,6 +466,11 @@ ARMSOCPrepareAccess(PixmapPtr pPixmap, int index)
 	int ret;
 	struct ARMSOCPixmapPrivRec *priv = exaGetPixmapDriverPrivate(pPixmap);
 
+	if (!is_accel_pixmap(priv)) {
+		pPixmap->devPrivate.ptr = priv->unaccel;
+		return TRUE;
+	}
+
 	pPixmap->devPrivate.ptr = armsoc_bo_map(priv->bo);
 	if (!pPixmap->devPrivate.ptr) {
 		xf86DrvMsg(-1, X_ERROR, "%s: Failed to map buffer\n", __func__);
@@ -426,7 +562,8 @@ ARMSOCFinishAccess(PixmapPtr pPixmap, int index)
 		 * do a more precise cache flush..
 		 */
 		pPixmap->devPrivate.ptr = NULL;
-		armsoc_bo_cpu_fini(priv->bo, idx2op(index));
+		if (is_accel_pixmap(priv))
+			armsoc_bo_cpu_fini(priv->bo, idx2op(index));
 	}
 }
 
@@ -454,7 +591,7 @@ ARMSOCPixmapIsOffscreen(PixmapPtr pPixmap)
 	 * wrap this function.
 	 */
 	struct ARMSOCPixmapPrivRec *priv = exaGetPixmapDriverPrivate(pPixmap);
-	return priv && priv->bo;
+	return priv && (priv->bo || priv->unaccel);
 }
 
 void ARMSOCRegisterExternalAccess(PixmapPtr pPixmap)

src/armsoc_exa.h

@@ -103,6 +103,8 @@ struct ARMSOCPixmapPrivRec {
 	 */
 	int ext_access_cnt;
 	struct armsoc_bo *bo;
+	unsigned char *unaccel;
+	size_t unaccel_size;
 	int usage_hint;
 };