Revert "vmwgfx: Get rid of device-specific DMA code"

This reverts commit 45b2457516a9db4bd1d60fbb24a1efbe2d9dd932.

Reverting this because using direct DMA for shared pixmaps should
boost performance. It should be usable both in a composited environment
and when running hosted. Need to retest the DMA flaws I saw before disabling
this code.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Jakob Bornecrantz <jakob@vmware.com>

4 files changed, 136 insertions, 26 deletions

diff --git a/vmwgfx/vmwgfx_drmi.c b/vmwgfx/vmwgfx_drmi.c
index 1e23f4a..496a16b 100644
--- a/vmwgfx/vmwgfx_drmi.c
+++ b/vmwgfx/vmwgfx_drmi.c
@@ -284,6 +284,109 @@ vmwgfx_dmabuf_destroy(struct vmwgfx_dmabuf *buf)
 }
 
 int
+vmwgfx_dma(unsigned int host_x, unsigned int host_y,
+	   RegionPtr region, struct vmwgfx_dmabuf *buf,
+	   uint32_t buf_pitch, uint32_t surface_handle, int to_surface)
+{
+    BoxPtr clips = REGION_RECTS(region);
+    unsigned int num_clips = REGION_NUM_RECTS(region);
+    struct drm_vmw_execbuf_arg arg;
+    struct drm_vmw_fence_rep rep;
+    int ret;
+    unsigned int size;
+    unsigned i;
+    SVGA3dCopyBox *cb;
+    SVGA3dCmdSurfaceDMASuffix *suffix;
+    SVGA3dCmdSurfaceDMA *body;
+    struct vmwgfx_int_dmabuf *ibuf = vmwgfx_int_dmabuf(buf);
+
+    struct {
+	SVGA3dCmdHeader header;
+	SVGA3dCmdSurfaceDMA body;
+	SVGA3dCopyBox cb;
+    } *cmd;
+
+    if (num_clips == 0)
+	return 0;
+
+    size = sizeof(*cmd) + (num_clips - 1) * sizeof(cmd->cb) +
+	sizeof(*suffix);
+    cmd = malloc(size);
+    if (!cmd)
+	return -1;
+
+    cmd->header.id = SVGA_3D_CMD_SURFACE_DMA;
+    cmd->header.size = sizeof(cmd->body) + num_clips * sizeof(cmd->cb) +
+	sizeof(*suffix);
+    cb = &cmd->cb;
+
+    suffix = (SVGA3dCmdSurfaceDMASuffix *) &cb[num_clips];
+    suffix->suffixSize = sizeof(*suffix);
+    suffix->maximumOffset = (uint32_t) -1;
+    suffix->flags.discard = 0;
+    suffix->flags.unsynchronized = 0;
+    suffix->flags.reserved = 0;
+
+    body = &cmd->body;
+    body->guest.ptr.gmrId = buf->gmr_id;
+    body->guest.ptr.offset = buf->gmr_offset;
+    body->guest.pitch = buf_pitch;
+    body->host.sid = surface_handle;
+    body->host.face = 0;
+    body->host.mipmap = 0;
+
+    body->transfer =  (to_surface ? SVGA3D_WRITE_HOST_VRAM :
+		       SVGA3D_READ_HOST_VRAM);
+
+
+    for (i=0; i < num_clips; i++, cb++, clips++) {
+	cb->x = (uint16_t) clips->x1 + host_x;
+	cb->y = (uint16_t) clips->y1 + host_y;
+	cb->z = 0;
+	cb->srcx = (uint16_t) clips->x1;
+	cb->srcy = (uint16_t) clips->y1;
+	cb->srcz = 0;
+	cb->w = (uint16_t) (clips->x2 - clips->x1);
+	cb->h = (uint16_t) (clips->y2 - clips->y1);
+	cb->d = 1;
+#if 0
+	LogMessage(X_INFO, "DMA! x: %u y: %u srcx: %u srcy: %u w: %u h: %u %s\n",
+		   cb->x, cb->y, cb->srcx, cb->srcy, cb->w, cb->h,
+		   to_surface ? "to" : "from");
+#endif
+
+    }
+
+    memset(&arg, 0, sizeof(arg));
+    memset(&rep, 0, sizeof(rep));
+
+    rep.error = -EFAULT;
+    arg.fence_rep = ((to_surface) ? 0UL : (unsigned long)&rep);
+    arg.commands = (unsigned long)cmd;
+    arg.command_size = size;
+    arg.throttle_us = 0;
+    arg.version = DRM_VMW_EXECBUF_VERSION;
+
+    ret = drmCommandWrite(ibuf->drm_fd, DRM_VMW_EXECBUF, &arg, sizeof(arg));
+    if (ret) {
+	LogMessage(X_ERROR, "DMA error %s.\n", strerror(-ret));
+    }
+
+    free(cmd);
+
+    if (rep.error == 0) {
+	ret = vmwgfx_fence_wait(ibuf->drm_fd, rep.handle, TRUE);
+	if (ret) {
+	    LogMessage(X_ERROR, "DMA from host fence wait error %s.\n",
+		       strerror(-ret));
+	    vmwgfx_fence_unref(ibuf->drm_fd, rep.handle);
+	}
+    }
+
+    return 0;
+}
+
+int
 vmwgfx_get_param(int drm_fd, uint32_t param, uint64_t *out)
 {
     struct drm_vmw_getparam_arg gp_arg;
diff --git a/vmwgfx/vmwgfx_drmi.h b/vmwgfx/vmwgfx_drmi.h
index fff728a..2435009 100644
--- a/vmwgfx/vmwgfx_drmi.h
+++ b/vmwgfx/vmwgfx_drmi.h
@@ -60,6 +60,11 @@ extern void
 vmwgfx_dmabuf_unmap(struct vmwgfx_dmabuf *buf);
 
 extern int
+vmwgfx_dma(unsigned int host_x, unsigned int host_y,
+	   RegionPtr region, struct vmwgfx_dmabuf *buf,
+	   uint32_t buf_pitch, uint32_t surface_handle, int to_surface);
+
+extern int
 vmwgfx_num_streams(int drm_fd, uint32_t *ntot, uint32_t *nfree);
 
 extern int
diff --git a/vmwgfx/vmwgfx_saa.c b/vmwgfx/vmwgfx_saa.c
index 28819c1..96275a5 100644
--- a/vmwgfx/vmwgfx_saa.c
+++ b/vmwgfx/vmwgfx_saa.c
@@ -285,40 +285,40 @@ vmwgfx_saa_dma(struct vmwgfx_saa *vsaa,
 	       Bool to_hw)
 {
     struct vmwgfx_saa_pixmap *vpix = vmwgfx_saa_pixmap(pixmap);
-    void *data = vpix->malloc;
-    int ret;
 
     if (!vpix->hw || (!vpix->gmr && !vpix->malloc))
 	return TRUE;
 
+    if (vpix->gmr && vsaa->can_optimize_dma) {
+	uint32_t handle, dummy;
 
-    if (vpix->gmr) {
-	data = vmwgfx_dmabuf_map(vpix->gmr);
-	if (!data)
+	if (xa_surface_handle(vpix->hw, xa_handle_type_shared,
+		 &handle, &dummy) != 0)
 	    goto out_err;
-    }
-
-    ret = xa_surface_dma(vsaa->xa_ctx, vpix->hw, data, pixmap->devKind,
-			 (int) to_hw,
-			 (struct xa_box *) REGION_RECTS(reg),
-			 REGION_NUM_RECTS(reg));
-    if (vpix->gmr)
-	vmwgfx_dmabuf_unmap(vpix->gmr);
-
-    if (ret)
-	goto out_err;
+	if (vmwgfx_dma(0, 0, reg, vpix->gmr, pixmap->devKind, handle,
+		       to_hw) != 0)
+	    goto out_err;
+    } else {
+	void *data = vpix->malloc;
+	int ret;
 
-    ret = xa_surface_dma(vsaa->xa_ctx, vpix->hw, data, pixmap->devKind,
-			 (int) to_hw,
-			 (struct xa_box *) REGION_RECTS(reg),
-			 REGION_NUM_RECTS(reg));
-    if (to_hw)
-	xa_context_flush(vsaa->xa_ctx);
-    if (vpix->gmr)
-	vmwgfx_dmabuf_unmap(vpix->gmr);
-    if (ret)
-	goto out_err;
+	if (vpix->gmr) {
+	    data = vmwgfx_dmabuf_map(vpix->gmr);
+	    if (!data)
+		goto out_err;
+	}
 
+	ret = xa_surface_dma(vsaa->xa_ctx, vpix->hw, data, pixmap->devKind,
+			     (int) to_hw,
+			     (struct xa_box *) REGION_RECTS(reg),
+			     REGION_NUM_RECTS(reg));
+	if (to_hw)
+	    xa_context_flush(vsaa->xa_ctx);
+	if (vpix->gmr)
+	    vmwgfx_dmabuf_unmap(vpix->gmr);
+	if (ret)
+	    goto out_err;
+    }
     return TRUE;
   out_err:
     LogMessage(X_ERROR, "DMA %s surface failed.\n",
@@ -1377,6 +1377,7 @@ vmwgfx_saa_init(ScreenPtr pScreen, int drm_fd, struct xa_tracker *xat,
 	vsaa->xa_ctx = xa_context_default(xat);
     vsaa->drm_fd = drm_fd;
     vsaa->present_flush = present_flush;
+    vsaa->can_optimize_dma = FALSE;
     vsaa->use_present_opt = direct_presents;
     vsaa->only_hw_presents = only_hw_presents;
     vsaa->rendercheck = rendercheck;
diff --git a/vmwgfx/vmwgfx_saa_priv.h b/vmwgfx/vmwgfx_saa_priv.h
index f5f45ba..16583b0 100644
--- a/vmwgfx/vmwgfx_saa_priv.h
+++ b/vmwgfx/vmwgfx_saa_priv.h
@@ -50,6 +50,7 @@ struct vmwgfx_saa {
     int ydiff;
     RegionRec present_region;
     uint32_t src_handle;
+    Bool can_optimize_dma;
     Bool use_present_opt;
     Bool only_hw_presents;
     Bool rendercheck;