1 files changed, 68 insertions, 220 deletions

diff --git a/drivers/video/tgafb.c b/drivers/video/tgafb.c
index 07c7df9..65ba992 100644
--- a/drivers/video/tgafb.c
+++ b/drivers/video/tgafb.c
@@ -182,6 +182,8 @@ tgafb_check_var(struct fb_var_screeninfo *var, struct fb_info *info)
 
 	if (var->xres_virtual != var->xres || var->yres_virtual != var->yres)
 		return -EINVAL;
+	if (var->xres * var->yres * (var->bits_per_pixel >> 3) > info->fix.smem_len)
+		return -EINVAL;
 	if (var->nonstd)
 		return -EINVAL;
 	if (1000000000 / var->pixclock > TGA_PLL_MAX_FREQ)
@@ -190,8 +192,8 @@ tgafb_check_var(struct fb_var_screeninfo *var, struct fb_info *info)
 		return -EINVAL;
 
 	/* Some of the acceleration routines assume the line width is
-	   a multiple of 64 bytes.  */
-	if (var->xres * (par->tga_type == TGA_TYPE_8PLANE ? 1 : 4) % 64)
+	   a multiple of 8 bytes.  */
+	if (var->xres * (par->tga_type == TGA_TYPE_8PLANE ? 1 : 4) % 8)
 		return -EINVAL;
 
 	return 0;
@@ -262,6 +264,7 @@ tgafb_set_par(struct fb_info *info)
 	par->yres = info->var.yres;
 	par->pll_freq = pll_freq = 1000000000 / info->var.pixclock;
 	par->bits_per_pixel = info->var.bits_per_pixel;
+	info->fix.line_length = par->xres * (par->bits_per_pixel >> 3);
 
 	tga_type = par->tga_type;
 
@@ -1136,222 +1139,57 @@ copyarea_line_32bpp(struct fb_info *info, u32 dy, u32 sy,
 	__raw_writel(TGA_MODE_SBM_24BPP|TGA_MODE_SIMPLE, tga_regs+TGA_MODE_REG);
 }
 
-/* The general case of forward copy in 8bpp mode.  */
+/* The (almost) general case of backward copy in 8bpp mode.  */
 static inline void
-copyarea_foreward_8bpp(struct fb_info *info, u32 dx, u32 dy, u32 sx, u32 sy,
-		       u32 height, u32 width, u32 line_length)
+copyarea_8bpp(struct fb_info *info, u32 dx, u32 dy, u32 sx, u32 sy,
+	      u32 height, u32 width, u32 line_length,
+	      const struct fb_copyarea *area)
 {
 	struct tga_par *par = (struct tga_par *) info->par;
-	unsigned long i, copied, left;
-	unsigned long dpos, spos, dalign, salign, yincr;
-	u32 smask_first, dmask_first, dmask_last;
-	int pixel_shift, need_prime, need_second;
-	unsigned long n64, n32, xincr_first;
+	unsigned i, yincr;
+	int depos, sepos, backward, last_step, step;
+	u32 mask_last;
+	unsigned n32;
 	void __iomem *tga_regs;
 	void __iomem *tga_fb;
 
-	yincr = line_length;
-	if (dy > sy) {
-		dy += height - 1;
-		sy += height - 1;
-		yincr = -yincr;
-	}
-
-	/* Compute the offsets and alignments in the frame buffer.
-	   More than anything else, these control how we do copies.  */
-	dpos = dy * line_length + dx;
-	spos = sy * line_length + sx;
-	dalign = dpos & 7;
-	salign = spos & 7;
-	dpos &= -8;
-	spos &= -8;
-
-	/* Compute the value for the PIXELSHIFT register.  This controls
-	   both non-co-aligned source and destination and copy direction.  */
-	if (dalign >= salign)
-		pixel_shift = dalign - salign;
-	else
-		pixel_shift = 8 - (salign - dalign);
-
-	/* Figure out if we need an additional priming step for the
-	   residue register.  */
-	need_prime = (salign > dalign);
-	if (need_prime)
-		dpos -= 8;
-
-	/* Begin by copying the leading unaligned destination.  Copy enough
-	   to make the next destination address 32-byte aligned.  */
-	copied = 32 - (dalign + (dpos & 31));
-	if (copied == 32)
-		copied = 0;
-	xincr_first = (copied + 7) & -8;
-	smask_first = dmask_first = (1ul << copied) - 1;
-	smask_first <<= salign;
-	dmask_first <<= dalign + need_prime*8;
-	if (need_prime && copied > 24)
-		copied -= 8;
-	left = width - copied;
-
-	/* Care for small copies.  */
-	if (copied > width) {
-		u32 t;
-		t = (1ul << width) - 1;
-		t <<= dalign + need_prime*8;
-		dmask_first &= t;
-		left = 0;
-	}
-
-	/* Attempt to use 64-byte copies.  This is only possible if the
-	   source and destination are co-aligned at 64 bytes.  */
-	n64 = need_second = 0;
-	if ((dpos & 63) == (spos & 63)
-	    && (height == 1 || line_length % 64 == 0)) {
-		/* We may need a 32-byte copy to ensure 64 byte alignment.  */
-		need_second = (dpos + xincr_first) & 63;
-		if ((need_second & 32) != need_second)
-			printk(KERN_ERR "tgafb: need_second wrong\n");
-		if (left >= need_second + 64) {
-			left -= need_second;
-			n64 = left / 64;
-			left %= 64;
-		} else
-			need_second = 0;
-	}
-
-	/* Copy trailing full 32-byte sections.  This will be the main
-	   loop if the 64 byte loop can't be used.  */
-	n32 = left / 32;
-	left %= 32;
-
-	/* Copy the trailing unaligned destination.  */
-	dmask_last = (1ul << left) - 1;
-
-	tga_regs = par->tga_regs_base;
-	tga_fb = par->tga_fb_base;
-
-	/* Set up the MODE and PIXELSHIFT registers.  */
-	__raw_writel(TGA_MODE_SBM_8BPP|TGA_MODE_COPY, tga_regs+TGA_MODE_REG);
-	__raw_writel(pixel_shift, tga_regs+TGA_PIXELSHIFT_REG);
-	wmb();
-
-	for (i = 0; i < height; ++i) {
-		unsigned long j;
-		void __iomem *sfb;
-		void __iomem *dfb;
-
-		sfb = tga_fb + spos;
-		dfb = tga_fb + dpos;
-		if (dmask_first) {
-			__raw_writel(smask_first, sfb);
-			wmb();
-			__raw_writel(dmask_first, dfb);
-			wmb();
-			sfb += xincr_first;
-			dfb += xincr_first;
-		}
-
-		if (need_second) {
-			__raw_writel(0xffffffff, sfb);
-			wmb();
-			__raw_writel(0xffffffff, dfb);
-			wmb();
-			sfb += 32;
-			dfb += 32;
-		}
-
-		if (n64 && (((unsigned long)sfb | (unsigned long)dfb) & 63))
-			printk(KERN_ERR
-			       "tgafb: misaligned copy64 (s:%p, d:%p)\n",
-			       sfb, dfb);
-
-		for (j = 0; j < n64; ++j) {
-			__raw_writel(sfb - tga_fb, tga_regs+TGA_COPY64_SRC);
-			wmb();
-			__raw_writel(dfb - tga_fb, tga_regs+TGA_COPY64_DST);
-			wmb();
-			sfb += 64;
-			dfb += 64;
-		}
-
-		for (j = 0; j < n32; ++j) {
-			__raw_writel(0xffffffff, sfb);
-			wmb();
-			__raw_writel(0xffffffff, dfb);
-			wmb();
-			sfb += 32;
-			dfb += 32;
-		}
-
-		if (dmask_last) {
-			__raw_writel(0xffffffff, sfb);
-			wmb();
-			__raw_writel(dmask_last, dfb);
-			wmb();
-		}
-
-		spos += yincr;
-		dpos += yincr;
+	/* Do acceleration only if we are aligned on 8 pixels */
+	if ((dx | sx | width) & 7) {
+		cfb_copyarea(info, area);
+		return;
 	}
 
-	/* Reset the MODE register to normal.  */
-	__raw_writel(TGA_MODE_SBM_8BPP|TGA_MODE_SIMPLE, tga_regs+TGA_MODE_REG);
-}
-
-/* The (almost) general case of backward copy in 8bpp mode.  */
-static inline void
-copyarea_backward_8bpp(struct fb_info *info, u32 dx, u32 dy, u32 sx, u32 sy,
-		       u32 height, u32 width, u32 line_length,
-		       const struct fb_copyarea *area)
-{
-	struct tga_par *par = (struct tga_par *) info->par;
-	unsigned long i, left, yincr;
-	unsigned long depos, sepos, dealign, sealign;
-	u32 mask_first, mask_last;
-	unsigned long n32;
-	void __iomem *tga_regs;
-	void __iomem *tga_fb;
-
 	yincr = line_length;
 	if (dy > sy) {
 		dy += height - 1;
 		sy += height - 1;
 		yincr = -yincr;
 	}
+	backward = dy == sy && dx > sx && dx < sx + width;
 
 	/* Compute the offsets and alignments in the frame buffer.
 	   More than anything else, these control how we do copies.  */
-	depos = dy * line_length + dx + width;
-	sepos = sy * line_length + sx + width;
-	dealign = depos & 7;
-	sealign = sepos & 7;
-
-	/* ??? The documentation appears to be incorrect (or very
-	   misleading) wrt how pixel shifting works in backward copy
-	   mode, i.e. when PIXELSHIFT is negative.  I give up for now.
-	   Do handle the common case of co-aligned backward copies,
-	   but frob everything else back on generic code.  */
-	if (dealign != sealign) {
-		cfb_copyarea(info, area);
-		return;
-	}
-
-	/* We begin the copy with the trailing pixels of the
-	   unaligned destination.  */
-	mask_first = (1ul << dealign) - 1;
-	left = width - dealign;
-
-	/* Care for small copies.  */
-	if (dealign > width) {
-		mask_first ^= (1ul << (dealign - width)) - 1;
-		left = 0;
-	}
+	depos = dy * line_length + dx;
+	sepos = sy * line_length + sx;
+	if (backward)
+		depos += width, sepos += width;
 
 	/* Next copy full words at a time.  */
-	n32 = left / 32;
-	left %= 32;
+	n32 = width / 32;
+	last_step = width % 32;
 
 	/* Finally copy the unaligned head of the span.  */
-	mask_last = -1 << (32 - left);
+	mask_last = (1ul << last_step) - 1;
+
+	if (!backward) {
+		step = 32;
+		last_step = 32;
+	} else {
+		step = -32;
+		last_step = -last_step;
+		sepos -= 32;
+		depos -= 32;
+	}
 
 	tga_regs = par->tga_regs_base;
 	tga_fb = par->tga_fb_base;
@@ -1368,25 +1206,33 @@ copyarea_backward_8bpp(struct fb_info *info, u32 dx, u32 dy, u32 sx, u32 sy,
 
 		sfb = tga_fb + sepos;
 		dfb = tga_fb + depos;
-		if (mask_first) {
-			__raw_writel(mask_first, sfb);
-			wmb();
-			__raw_writel(mask_first, dfb);
-			wmb();
-		}
 
-		for (j = 0; j < n32; ++j) {
-			sfb -= 32;
-			dfb -= 32;
+		for (j = 0; j < n32; j++) {
+			if (j < 2 && j + 1 < n32 && !backward &&
+			    !(((unsigned long)sfb | (unsigned long)dfb) & 63)) {
+				do {
+					__raw_writel(sfb - tga_fb, tga_regs+TGA_COPY64_SRC);
+					wmb();
+					__raw_writel(dfb - tga_fb, tga_regs+TGA_COPY64_DST);
+					wmb();
+					sfb += 64;
+					dfb += 64;
+					j += 2;
+				} while (j + 1 < n32);
+				j--;
+				continue;
+			}
 			__raw_writel(0xffffffff, sfb);
 			wmb();
 			__raw_writel(0xffffffff, dfb);
 			wmb();
+			sfb += step;
+			dfb += step;
 		}
 
 		if (mask_last) {
-			sfb -= 32;
-			dfb -= 32;
+			sfb += last_step - step;
+			dfb += last_step - step;
 			__raw_writel(mask_last, sfb);
 			wmb();
 			__raw_writel(mask_last, dfb);
@@ -1434,7 +1280,7 @@ tgafb_copyarea(struct fb_info *info, const struct fb_copyarea *area)
 	bpp = info->var.bits_per_pixel;
 
 	/* Detect copies of the entire line.  */
-	if (width * (bpp >> 3) == line_length) {
+	if (!(line_length & 63) && width * (bpp >> 3) == line_length) {
 		if (bpp == 8)
 			copyarea_line_8bpp(info, dy, sy, height, width);
 		else
@@ -1447,14 +1293,9 @@ tgafb_copyarea(struct fb_info *info, const struct fb_copyarea *area)
 	else if (bpp == 32)
 		cfb_copyarea(info, area);
 
-	/* Detect overlapping source and destination that requires
-	   a backward copy.  */
-	else if (dy == sy && dx > sx && dx < sx + width)
-		copyarea_backward_8bpp(info, dx, dy, sx, sy, height,
-				       width, line_length, area);
 	else
-		copyarea_foreward_8bpp(info, dx, dy, sx, sy, height,
-				       width, line_length);
+		copyarea_8bpp(info, dx, dy, sx, sy, height,
+			      width, line_length, area);
 }
 
 
@@ -1470,6 +1311,7 @@ tgafb_init_fix(struct fb_info *info)
 	int tga_bus_tc = TGA_BUS_TC(par->dev);
 	u8 tga_type = par->tga_type;
 	const char *tga_type_name = NULL;
+	unsigned memory_size;
 
 	switch (tga_type) {
 	case TGA_TYPE_8PLANE:
@@ -1477,22 +1319,27 @@ tgafb_init_fix(struct fb_info *info)
 			tga_type_name = "Digital ZLXp-E1";
 		if (tga_bus_tc)
 			tga_type_name = "Digital ZLX-E1";
+		memory_size = 2097152;
 		break;
 	case TGA_TYPE_24PLANE:
 		if (tga_bus_pci)
 			tga_type_name = "Digital ZLXp-E2";
 		if (tga_bus_tc)
 			tga_type_name = "Digital ZLX-E2";
+		memory_size = 8388608;
 		break;
 	case TGA_TYPE_24PLUSZ:
 		if (tga_bus_pci)
 			tga_type_name = "Digital ZLXp-E3";
 		if (tga_bus_tc)
 			tga_type_name = "Digital ZLX-E3";
+		memory_size = 16777216;
 		break;
 	}
-	if (!tga_type_name)
+	if (!tga_type_name) {
 		tga_type_name = "Unknown";
+		memory_size = 16777216;
+	}
 
 	strlcpy(info->fix.id, tga_type_name, sizeof(info->fix.id));
 
@@ -1502,9 +1349,8 @@ tgafb_init_fix(struct fb_info *info)
 			    ? FB_VISUAL_PSEUDOCOLOR
 			    : FB_VISUAL_DIRECTCOLOR);
 
-	info->fix.line_length = par->xres * (par->bits_per_pixel >> 3);
 	info->fix.smem_start = (size_t) par->tga_fb_base;
-	info->fix.smem_len = info->fix.line_length * par->yres;
+	info->fix.smem_len = memory_size;
 	info->fix.mmio_start = (size_t) par->tga_regs_base;
 	info->fix.mmio_len = 512;
 
@@ -1628,6 +1474,9 @@ static int tgafb_register(struct device *dev)
 		modedb_tga = &modedb_tc;
 		modedbsize_tga = 1;
 	}
+
+	tgafb_init_fix(info);
+
 	ret = fb_find_mode(&info->var, info,
 			   mode_option ? mode_option : mode_option_tga,
 			   modedb_tga, modedbsize_tga, NULL,
@@ -1645,7 +1494,6 @@ static int tgafb_register(struct device *dev)
 	}
 
 	tgafb_set_par(info);
-	tgafb_init_fix(info);
 
 	if (register_framebuffer(info) < 0) {
 		printk(KERN_ERR "tgafb: Could not register framebuffer\n");