[PATCH 3/7] drm/omap: Use bitmaps for TILER placement

Tomi Valkeinen tomi.valkeinen at ti.com
Wed Dec 9 07:38:07 PST 2015


From: Andy Gross <andy.gross at ti.com>

Modified Tiler placement to utilize bitmaps for bookkeeping and
all placement algorithms.  This resulted in a substantial savings
in time for all Tiler reservation and free operations.  Typical
savings are in the range of 28% decrease in time taken with larger
buffers showing a 80%+ decrease.

Signed-off-by: Andy Gross <andy.gross at ti.com>
Signed-off-by: Tomi Valkeinen <tomi.valkeinen at ti.com>
---
 drivers/gpu/drm/omapdrm/Makefile         |   2 +-
 drivers/gpu/drm/omapdrm/omap_dmm_tiler.c |  14 +-
 drivers/gpu/drm/omapdrm/sita.c           | 261 ++++++++++++
 drivers/gpu/drm/omapdrm/tcm-sita.c       | 703 -------------------------------
 drivers/gpu/drm/omapdrm/tcm.h            |  26 +-
 5 files changed, 285 insertions(+), 721 deletions(-)
 create mode 100644 drivers/gpu/drm/omapdrm/sita.c
 delete mode 100644 drivers/gpu/drm/omapdrm/tcm-sita.c

diff --git a/drivers/gpu/drm/omapdrm/Makefile b/drivers/gpu/drm/omapdrm/Makefile
index 778372b062ad..4efdc9901288 100644
--- a/drivers/gpu/drm/omapdrm/Makefile
+++ b/drivers/gpu/drm/omapdrm/Makefile
@@ -16,6 +16,6 @@ omapdrm-y := omap_drv.o \
 	omap_gem.o \
 	omap_gem_dmabuf.o \
 	omap_dmm_tiler.o \
-	tcm-sita.o
+	sita.o
 
 obj-$(CONFIG_DRM_OMAP)	+= omapdrm.o
diff --git a/drivers/gpu/drm/omapdrm/omap_dmm_tiler.c b/drivers/gpu/drm/omapdrm/omap_dmm_tiler.c
index 7841970de48d..23af2ee53ac8 100644
--- a/drivers/gpu/drm/omapdrm/omap_dmm_tiler.c
+++ b/drivers/gpu/drm/omapdrm/omap_dmm_tiler.c
@@ -363,6 +363,7 @@ struct tiler_block *tiler_reserve_2d(enum tiler_fmt fmt, uint16_t w,
 	u32 min_align = 128;
 	int ret;
 	unsigned long flags;
+	size_t slot_bytes;
 
 	BUG_ON(!validfmt(fmt));
 
@@ -371,13 +372,15 @@ struct tiler_block *tiler_reserve_2d(enum tiler_fmt fmt, uint16_t w,
 	h = DIV_ROUND_UP(h, geom[fmt].slot_h);
 
 	/* convert alignment to slots */
-	min_align = max(min_align, (geom[fmt].slot_w * geom[fmt].cpp));
-	align = ALIGN(align, min_align);
-	align /= geom[fmt].slot_w * geom[fmt].cpp;
+	slot_bytes = geom[fmt].slot_w * geom[fmt].cpp;
+	min_align = max(min_align, slot_bytes);
+	align = (align > min_align) ? ALIGN(align, min_align) : min_align;
+	align /= slot_bytes;
 
 	block->fmt = fmt;
 
-	ret = tcm_reserve_2d(containers[fmt], w, h, align, &block->area);
+	ret = tcm_reserve_2d(containers[fmt], w, h, align, -1, slot_bytes,
+			&block->area);
 	if (ret) {
 		kfree(block);
 		return ERR_PTR(-ENOMEM);
@@ -739,8 +742,7 @@ static int omap_dmm_probe(struct platform_device *dev)
 	   programming during reill operations */
 	for (i = 0; i < omap_dmm->num_lut; i++) {
 		omap_dmm->tcm[i] = sita_init(omap_dmm->container_width,
-						omap_dmm->container_height,
-						NULL);
+						omap_dmm->container_height);
 
 		if (!omap_dmm->tcm[i]) {
 			dev_err(&dev->dev, "failed to allocate container\n");
diff --git a/drivers/gpu/drm/omapdrm/sita.c b/drivers/gpu/drm/omapdrm/sita.c
new file mode 100644
index 000000000000..07fc7ee67044
--- /dev/null
+++ b/drivers/gpu/drm/omapdrm/sita.c
@@ -0,0 +1,261 @@
+/*
+ * sita.c
+ *
+ * SImple Tiler Allocator (SiTA): 2D and 1D allocation(reservation) algorithm
+ *
+ * Authors: Ravi Ramachandra <r.ramachandra at ti.com>,
+ *          Lajos Molnar <molnar at ti.com>
+ *          Andy Gross <andy.gross at ti.com>
+ *
+ * Copyright (C) 2012 Texas Instruments, Inc.
+ *
+ * This package is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * THIS PACKAGE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED
+ * WARRANTIES OF MERCHANTIBILITY AND FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ */
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/errno.h>
+#include <linux/sched.h>
+#include <linux/wait.h>
+#include <linux/bitmap.h>
+#include <linux/slab.h>
+#include "tcm.h"
+
+static unsigned long mask[8];
+/*
+ * pos		position in bitmap
+ * w		width in slots
+ * h		height in slots
+ * map		ptr to bitmap
+ * stride		slots in a row
+ */
+static void free_slots(unsigned long pos, uint16_t w, uint16_t h,
+		unsigned long *map, uint16_t stride)
+{
+	int i;
+
+	for (i = 0; i < h; i++, pos += stride)
+		bitmap_clear(map, pos, w);
+}
+
+/*
+ * w		width in slots
+ * pos		ptr to position
+ * map		ptr to bitmap
+ * num_bits	number of bits in bitmap
+ */
+static int r2l_b2t_1d(uint16_t w, unsigned long *pos, unsigned long *map,
+		size_t num_bits)
+{
+	unsigned long search_count = 0;
+	unsigned long bit;
+	bool area_found = false;
+
+	*pos = num_bits - w;
+
+	while (search_count < num_bits) {
+		bit = find_next_bit(map, num_bits, *pos);
+
+		if (bit - *pos >= w) {
+			/* found a long enough free area */
+			bitmap_set(map, *pos, w);
+			area_found = true;
+			break;
+		}
+
+		search_count = num_bits - bit + w;
+		*pos = bit - w;
+	}
+
+	return (area_found) ? 0 : -ENOMEM;
+}
+
+/*
+ * w = width in slots
+ * h = height in slots
+ * a = align in slots	(mask, 2^n-1, 0 is unaligned)
+ * offset = offset in bytes from 4KiB
+ * pos = position in bitmap for buffer
+ * map = bitmap ptr
+ * num_bits = size of bitmap
+ * stride = bits in one row of container
+ */
+static int l2r_t2b(uint16_t w, uint16_t h, uint16_t a, int16_t offset,
+		unsigned long *pos, unsigned long slot_bytes,
+		unsigned long *map, size_t num_bits, size_t slot_stride)
+{
+	int i;
+	unsigned long index;
+	bool area_free;
+	unsigned long slots_per_band = PAGE_SIZE / slot_bytes;
+	unsigned long bit_offset = (offset > 0) ? offset / slot_bytes : 0;
+	unsigned long curr_bit = bit_offset;
+
+	/* reset alignment to 1 if we are matching a specific offset */
+	/* adjust alignment - 1 to get to the format expected in bitmaps */
+	a = (offset > 0) ? 0 : a - 1;
+
+	/* FIXME Return error if slots_per_band > stride */
+
+	while (curr_bit < num_bits) {
+		*pos = bitmap_find_next_zero_area(map, num_bits, curr_bit, w,
+				a);
+
+		/* skip forward if we are not at right offset */
+		if (bit_offset > 0 && (*pos % slots_per_band != bit_offset)) {
+			curr_bit = ALIGN(*pos, slots_per_band) + bit_offset;
+			continue;
+		}
+
+		/* skip forward to next row if we overlap end of row */
+		if ((*pos % slot_stride) + w > slot_stride) {
+			curr_bit = ALIGN(*pos, slot_stride) + bit_offset;
+			continue;
+		}
+
+		/* TODO: Handle overlapping 4K boundaries */
+
+		/* break out of look if we will go past end of container */
+		if ((*pos + slot_stride * h) > num_bits)
+			break;
+
+		/* generate mask that represents out matching pattern */
+		bitmap_clear(mask, 0, slot_stride);
+		bitmap_set(mask, (*pos % BITS_PER_LONG), w);
+
+		/* assume the area is free until we find an overlap */
+		area_free = true;
+
+		/* check subsequent rows to see if complete area is free */
+		for (i = 1; i < h; i++) {
+			index = *pos / BITS_PER_LONG + i * 8;
+			if (bitmap_intersects(&map[index], mask,
+				(*pos % BITS_PER_LONG) + w)) {
+				area_free = false;
+				break;
+			}
+		}
+
+		if (area_free)
+			break;
+
+		/* go forward past this match */
+		if (bit_offset > 0)
+			curr_bit = ALIGN(*pos, slots_per_band) + bit_offset;
+		else
+			curr_bit = *pos + a + 1;
+	}
+
+	if (area_free) {
+		/* set area as in-use. iterate over rows */
+		for (i = 0, index = *pos; i < h; i++, index += slot_stride)
+			bitmap_set(map, index, w);
+	}
+
+	return (area_free) ? 0 : -ENOMEM;
+}
+
+static s32 sita_reserve_1d(struct tcm *tcm, u32 num_slots,
+			   struct tcm_area *area)
+{
+	unsigned long pos;
+	int ret;
+
+	spin_lock(&(tcm->lock));
+	ret = r2l_b2t_1d(num_slots, &pos, tcm->bitmap, tcm->map_size);
+	if (!ret) {
+		area->p0.x = pos % tcm->width;
+		area->p0.y = pos / tcm->width;
+		area->p1.x = (pos + num_slots - 1) % tcm->width;
+		area->p1.y = (pos + num_slots - 1) / tcm->width;
+	}
+	spin_unlock(&(tcm->lock));
+
+	return ret;
+}
+
+static s32 sita_reserve_2d(struct tcm *tcm, u16 h, u16 w, u16 align,
+				int16_t offset, uint16_t slot_bytes,
+				struct tcm_area *area)
+{
+	unsigned long pos;
+	int ret;
+
+	spin_lock(&(tcm->lock));
+	ret = l2r_t2b(w, h, align, offset, &pos, slot_bytes, tcm->bitmap,
+			tcm->map_size, tcm->width);
+
+	if (!ret) {
+		area->p0.x = pos % tcm->width;
+		area->p0.y = pos / tcm->width;
+		area->p1.x = area->p0.x + w - 1;
+		area->p1.y = area->p0.y + h - 1;
+	}
+	spin_unlock(&(tcm->lock));
+
+	return ret;
+}
+
+static void sita_deinit(struct tcm *tcm)
+{
+	kfree(tcm);
+}
+
+static s32 sita_free(struct tcm *tcm, struct tcm_area *area)
+{
+	unsigned long pos;
+	uint16_t w, h;
+
+	pos = area->p0.x + area->p0.y * tcm->width;
+	if (area->is2d) {
+		w = area->p1.x - area->p0.x + 1;
+		h = area->p1.y - area->p0.y + 1;
+	} else {
+		w = area->p1.x + area->p1.y * tcm->width - pos + 1;
+		h = 1;
+	}
+
+	spin_lock(&(tcm->lock));
+	free_slots(pos, w, h, tcm->bitmap, tcm->width);
+	spin_unlock(&(tcm->lock));
+	return 0;
+}
+
+struct tcm *sita_init(u16 width, u16 height)
+{
+	struct tcm *tcm;
+	size_t map_size = BITS_TO_LONGS(width*height) * sizeof(unsigned long);
+
+	if (width == 0 || height == 0)
+		return NULL;
+
+	tcm = kzalloc(sizeof(*tcm) + map_size, GFP_KERNEL);
+	if (!tcm)
+		goto error;
+
+	/* Updating the pointers to SiTA implementation APIs */
+	tcm->height = height;
+	tcm->width = width;
+	tcm->reserve_2d = sita_reserve_2d;
+	tcm->reserve_1d = sita_reserve_1d;
+	tcm->free = sita_free;
+	tcm->deinit = sita_deinit;
+
+	spin_lock_init(&tcm->lock);
+	tcm->bitmap = (unsigned long *)(tcm + 1);
+	bitmap_clear(tcm->bitmap, 0, width*height);
+
+	tcm->map_size = width*height;
+
+	return tcm;
+
+error:
+	kfree(tcm);
+	return NULL;
+}
diff --git a/drivers/gpu/drm/omapdrm/tcm-sita.c b/drivers/gpu/drm/omapdrm/tcm-sita.c
deleted file mode 100644
index efb609510540..000000000000
--- a/drivers/gpu/drm/omapdrm/tcm-sita.c
+++ /dev/null
@@ -1,703 +0,0 @@
-/*
- * tcm-sita.c
- *
- * SImple Tiler Allocator (SiTA): 2D and 1D allocation(reservation) algorithm
- *
- * Authors: Ravi Ramachandra <r.ramachandra at ti.com>,
- *          Lajos Molnar <molnar at ti.com>
- *
- * Copyright (C) 2009-2010 Texas Instruments, Inc.
- *
- * This package is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License version 2 as
- * published by the Free Software Foundation.
- *
- * THIS PACKAGE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR
- * IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED
- * WARRANTIES OF MERCHANTIBILITY AND FITNESS FOR A PARTICULAR PURPOSE.
- *
- */
-#include <linux/slab.h>
-#include <linux/spinlock.h>
-
-#include "tcm-sita.h"
-
-#define ALIGN_DOWN(value, align) ((value) & ~((align) - 1))
-
-/* Individual selection criteria for different scan areas */
-static s32 CR_L2R_T2B = CR_BIAS_HORIZONTAL;
-static s32 CR_R2L_T2B = CR_DIAGONAL_BALANCE;
-
-/*********************************************
- *	TCM API - Sita Implementation
- *********************************************/
-static s32 sita_reserve_2d(struct tcm *tcm, u16 h, u16 w, u8 align,
-			   struct tcm_area *area);
-static s32 sita_reserve_1d(struct tcm *tcm, u32 slots, struct tcm_area *area);
-static s32 sita_free(struct tcm *tcm, struct tcm_area *area);
-static void sita_deinit(struct tcm *tcm);
-
-/*********************************************
- *	Main Scanner functions
- *********************************************/
-static s32 scan_areas_and_find_fit(struct tcm *tcm, u16 w, u16 h, u16 align,
-				   struct tcm_area *area);
-
-static s32 scan_l2r_t2b(struct tcm *tcm, u16 w, u16 h, u16 align,
-			struct tcm_area *field, struct tcm_area *area);
-
-static s32 scan_r2l_t2b(struct tcm *tcm, u16 w, u16 h, u16 align,
-			struct tcm_area *field, struct tcm_area *area);
-
-static s32 scan_r2l_b2t_one_dim(struct tcm *tcm, u32 num_slots,
-			struct tcm_area *field, struct tcm_area *area);
-
-/*********************************************
- *	Support Infrastructure Methods
- *********************************************/
-static s32 is_area_free(struct tcm_area ***map, u16 x0, u16 y0, u16 w, u16 h);
-
-static s32 update_candidate(struct tcm *tcm, u16 x0, u16 y0, u16 w, u16 h,
-			    struct tcm_area *field, s32 criteria,
-			    struct score *best);
-
-static void get_nearness_factor(struct tcm_area *field,
-				struct tcm_area *candidate,
-				struct nearness_factor *nf);
-
-static void get_neighbor_stats(struct tcm *tcm, struct tcm_area *area,
-			       struct neighbor_stats *stat);
-
-static void fill_area(struct tcm *tcm,
-				struct tcm_area *area, struct tcm_area *parent);
-
-
-/*********************************************/
-
-/*********************************************
- *	Utility Methods
- *********************************************/
-struct tcm *sita_init(u16 width, u16 height, struct tcm_pt *attr)
-{
-	struct tcm *tcm;
-	struct sita_pvt *pvt;
-	struct tcm_area area = {0};
-	s32 i;
-
-	if (width == 0 || height == 0)
-		return NULL;
-
-	tcm = kmalloc(sizeof(*tcm), GFP_KERNEL);
-	pvt = kmalloc(sizeof(*pvt), GFP_KERNEL);
-	if (!tcm || !pvt)
-		goto error;
-
-	memset(tcm, 0, sizeof(*tcm));
-	memset(pvt, 0, sizeof(*pvt));
-
-	/* Updating the pointers to SiTA implementation APIs */
-	tcm->height = height;
-	tcm->width = width;
-	tcm->reserve_2d = sita_reserve_2d;
-	tcm->reserve_1d = sita_reserve_1d;
-	tcm->free = sita_free;
-	tcm->deinit = sita_deinit;
-	tcm->pvt = (void *)pvt;
-
-	spin_lock_init(&(pvt->lock));
-
-	/* Creating tam map */
-	pvt->map = kmalloc(sizeof(*pvt->map) * tcm->width, GFP_KERNEL);
-	if (!pvt->map)
-		goto error;
-
-	for (i = 0; i < tcm->width; i++) {
-		pvt->map[i] =
-			kmalloc(sizeof(**pvt->map) * tcm->height,
-								GFP_KERNEL);
-		if (pvt->map[i] == NULL) {
-			while (i--)
-				kfree(pvt->map[i]);
-			kfree(pvt->map);
-			goto error;
-		}
-	}
-
-	if (attr && attr->x <= tcm->width && attr->y <= tcm->height) {
-		pvt->div_pt.x = attr->x;
-		pvt->div_pt.y = attr->y;
-
-	} else {
-		/* Defaulting to 3:1 ratio on width for 2D area split */
-		/* Defaulting to 3:1 ratio on height for 2D and 1D split */
-		pvt->div_pt.x = (tcm->width * 3) / 4;
-		pvt->div_pt.y = (tcm->height * 3) / 4;
-	}
-
-	spin_lock(&(pvt->lock));
-	assign(&area, 0, 0, width - 1, height - 1);
-	fill_area(tcm, &area, NULL);
-	spin_unlock(&(pvt->lock));
-	return tcm;
-
-error:
-	kfree(tcm);
-	kfree(pvt);
-	return NULL;
-}
-
-static void sita_deinit(struct tcm *tcm)
-{
-	struct sita_pvt *pvt = (struct sita_pvt *)tcm->pvt;
-	struct tcm_area area = {0};
-	s32 i;
-
-	area.p1.x = tcm->width - 1;
-	area.p1.y = tcm->height - 1;
-
-	spin_lock(&(pvt->lock));
-	fill_area(tcm, &area, NULL);
-	spin_unlock(&(pvt->lock));
-
-	for (i = 0; i < tcm->height; i++)
-		kfree(pvt->map[i]);
-	kfree(pvt->map);
-	kfree(pvt);
-}
-
-/**
- * Reserve a 1D area in the container
- *
- * @param num_slots	size of 1D area
- * @param area		pointer to the area that will be populated with the
- *			reserved area
- *
- * @return 0 on success, non-0 error value on failure.
- */
-static s32 sita_reserve_1d(struct tcm *tcm, u32 num_slots,
-			   struct tcm_area *area)
-{
-	s32 ret;
-	struct tcm_area field = {0};
-	struct sita_pvt *pvt = (struct sita_pvt *)tcm->pvt;
-
-	spin_lock(&(pvt->lock));
-
-	/* Scanning entire container */
-	assign(&field, tcm->width - 1, tcm->height - 1, 0, 0);
-
-	ret = scan_r2l_b2t_one_dim(tcm, num_slots, &field, area);
-	if (!ret)
-		/* update map */
-		fill_area(tcm, area, area);
-
-	spin_unlock(&(pvt->lock));
-	return ret;
-}
-
-/**
- * Reserve a 2D area in the container
- *
- * @param w	width
- * @param h	height
- * @param area	pointer to the area that will be populated with the reserved
- *		area
- *
- * @return 0 on success, non-0 error value on failure.
- */
-static s32 sita_reserve_2d(struct tcm *tcm, u16 h, u16 w, u8 align,
-			   struct tcm_area *area)
-{
-	s32 ret;
-	struct sita_pvt *pvt = (struct sita_pvt *)tcm->pvt;
-
-	/* not supporting more than 64 as alignment */
-	if (align > 64)
-		return -EINVAL;
-
-	/* we prefer 1, 32 and 64 as alignment */
-	align = align <= 1 ? 1 : align <= 32 ? 32 : 64;
-
-	spin_lock(&(pvt->lock));
-	ret = scan_areas_and_find_fit(tcm, w, h, align, area);
-	if (!ret)
-		/* update map */
-		fill_area(tcm, area, area);
-
-	spin_unlock(&(pvt->lock));
-	return ret;
-}
-
-/**
- * Unreserve a previously allocated 2D or 1D area
- * @param area	area to be freed
- * @return 0 - success
- */
-static s32 sita_free(struct tcm *tcm, struct tcm_area *area)
-{
-	struct sita_pvt *pvt = (struct sita_pvt *)tcm->pvt;
-
-	spin_lock(&(pvt->lock));
-
-	/* check that this is in fact an existing area */
-	WARN_ON(pvt->map[area->p0.x][area->p0.y] != area ||
-		pvt->map[area->p1.x][area->p1.y] != area);
-
-	/* Clear the contents of the associated tiles in the map */
-	fill_area(tcm, area, NULL);
-
-	spin_unlock(&(pvt->lock));
-
-	return 0;
-}
-
-/**
- * Note: In general the cordinates in the scan field area relevant to the can
- * sweep directions. The scan origin (e.g. top-left corner) will always be
- * the p0 member of the field.  Therfore, for a scan from top-left p0.x <= p1.x
- * and p0.y <= p1.y; whereas, for a scan from bottom-right p1.x <= p0.x and p1.y
- * <= p0.y
- */
-
-/**
- * Raster scan horizontally right to left from top to bottom to find a place for
- * a 2D area of given size inside a scan field.
- *
- * @param w	width of desired area
- * @param h	height of desired area
- * @param align	desired area alignment
- * @param area	pointer to the area that will be set to the best position
- * @param field	area to scan (inclusive)
- *
- * @return 0 on success, non-0 error value on failure.
- */
-static s32 scan_r2l_t2b(struct tcm *tcm, u16 w, u16 h, u16 align,
-			struct tcm_area *field, struct tcm_area *area)
-{
-	s32 x, y;
-	s16 start_x, end_x, start_y, end_y, found_x = -1;
-	struct tcm_area ***map = ((struct sita_pvt *)tcm->pvt)->map;
-	struct score best = {{0}, {0}, {0}, 0};
-
-	start_x = field->p0.x;
-	end_x = field->p1.x;
-	start_y = field->p0.y;
-	end_y = field->p1.y;
-
-	/* check scan area co-ordinates */
-	if (field->p0.x < field->p1.x ||
-	    field->p1.y < field->p0.y)
-		return -EINVAL;
-
-	/* check if allocation would fit in scan area */
-	if (w > LEN(start_x, end_x) || h > LEN(end_y, start_y))
-		return -ENOSPC;
-
-	/* adjust start_x and end_y, as allocation would not fit beyond */
-	start_x = ALIGN_DOWN(start_x - w + 1, align); /* - 1 to be inclusive */
-	end_y = end_y - h + 1;
-
-	/* check if allocation would still fit in scan area */
-	if (start_x < end_x)
-		return -ENOSPC;
-
-	/* scan field top-to-bottom, right-to-left */
-	for (y = start_y; y <= end_y; y++) {
-		for (x = start_x; x >= end_x; x -= align) {
-			if (is_area_free(map, x, y, w, h)) {
-				found_x = x;
-
-				/* update best candidate */
-				if (update_candidate(tcm, x, y, w, h, field,
-							CR_R2L_T2B, &best))
-					goto done;
-
-				/* change upper x bound */
-				end_x = x + 1;
-				break;
-			} else if (map[x][y] && map[x][y]->is2d) {
-				/* step over 2D areas */
-				x = ALIGN(map[x][y]->p0.x - w + 1, align);
-			}
-		}
-
-		/* break if you find a free area shouldering the scan field */
-		if (found_x == start_x)
-			break;
-	}
-
-	if (!best.a.tcm)
-		return -ENOSPC;
-done:
-	assign(area, best.a.p0.x, best.a.p0.y, best.a.p1.x, best.a.p1.y);
-	return 0;
-}
-
-/**
- * Raster scan horizontally left to right from top to bottom to find a place for
- * a 2D area of given size inside a scan field.
- *
- * @param w	width of desired area
- * @param h	height of desired area
- * @param align	desired area alignment
- * @param area	pointer to the area that will be set to the best position
- * @param field	area to scan (inclusive)
- *
- * @return 0 on success, non-0 error value on failure.
- */
-static s32 scan_l2r_t2b(struct tcm *tcm, u16 w, u16 h, u16 align,
-			struct tcm_area *field, struct tcm_area *area)
-{
-	s32 x, y;
-	s16 start_x, end_x, start_y, end_y, found_x = -1;
-	struct tcm_area ***map = ((struct sita_pvt *)tcm->pvt)->map;
-	struct score best = {{0}, {0}, {0}, 0};
-
-	start_x = field->p0.x;
-	end_x = field->p1.x;
-	start_y = field->p0.y;
-	end_y = field->p1.y;
-
-	/* check scan area co-ordinates */
-	if (field->p1.x < field->p0.x ||
-	    field->p1.y < field->p0.y)
-		return -EINVAL;
-
-	/* check if allocation would fit in scan area */
-	if (w > LEN(end_x, start_x) || h > LEN(end_y, start_y))
-		return -ENOSPC;
-
-	start_x = ALIGN(start_x, align);
-
-	/* check if allocation would still fit in scan area */
-	if (w > LEN(end_x, start_x))
-		return -ENOSPC;
-
-	/* adjust end_x and end_y, as allocation would not fit beyond */
-	end_x = end_x - w + 1; /* + 1 to be inclusive */
-	end_y = end_y - h + 1;
-
-	/* scan field top-to-bottom, left-to-right */
-	for (y = start_y; y <= end_y; y++) {
-		for (x = start_x; x <= end_x; x += align) {
-			if (is_area_free(map, x, y, w, h)) {
-				found_x = x;
-
-				/* update best candidate */
-				if (update_candidate(tcm, x, y, w, h, field,
-							CR_L2R_T2B, &best))
-					goto done;
-				/* change upper x bound */
-				end_x = x - 1;
-
-				break;
-			} else if (map[x][y] && map[x][y]->is2d) {
-				/* step over 2D areas */
-				x = ALIGN_DOWN(map[x][y]->p1.x, align);
-			}
-		}
-
-		/* break if you find a free area shouldering the scan field */
-		if (found_x == start_x)
-			break;
-	}
-
-	if (!best.a.tcm)
-		return -ENOSPC;
-done:
-	assign(area, best.a.p0.x, best.a.p0.y, best.a.p1.x, best.a.p1.y);
-	return 0;
-}
-
-/**
- * Raster scan horizontally right to left from bottom to top to find a place
- * for a 1D area of given size inside a scan field.
- *
- * @param num_slots	size of desired area
- * @param align		desired area alignment
- * @param area		pointer to the area that will be set to the best
- *			position
- * @param field		area to scan (inclusive)
- *
- * @return 0 on success, non-0 error value on failure.
- */
-static s32 scan_r2l_b2t_one_dim(struct tcm *tcm, u32 num_slots,
-				struct tcm_area *field, struct tcm_area *area)
-{
-	s32 found = 0;
-	s16 x, y;
-	struct sita_pvt *pvt = (struct sita_pvt *)tcm->pvt;
-	struct tcm_area *p;
-
-	/* check scan area co-ordinates */
-	if (field->p0.y < field->p1.y)
-		return -EINVAL;
-
-	/**
-	 * Currently we only support full width 1D scan field, which makes sense
-	 * since 1D slot-ordering spans the full container width.
-	 */
-	if (tcm->width != field->p0.x - field->p1.x + 1)
-		return -EINVAL;
-
-	/* check if allocation would fit in scan area */
-	if (num_slots > tcm->width * LEN(field->p0.y, field->p1.y))
-		return -ENOSPC;
-
-	x = field->p0.x;
-	y = field->p0.y;
-
-	/* find num_slots consecutive free slots to the left */
-	while (found < num_slots) {
-		if (y < 0)
-			return -ENOSPC;
-
-		/* remember bottom-right corner */
-		if (found == 0) {
-			area->p1.x = x;
-			area->p1.y = y;
-		}
-
-		/* skip busy regions */
-		p = pvt->map[x][y];
-		if (p) {
-			/* move to left of 2D areas, top left of 1D */
-			x = p->p0.x;
-			if (!p->is2d)
-				y = p->p0.y;
-
-			/* start over */
-			found = 0;
-		} else {
-			/* count consecutive free slots */
-			found++;
-			if (found == num_slots)
-				break;
-		}
-
-		/* move to the left */
-		if (x == 0)
-			y--;
-		x = (x ? : tcm->width) - 1;
-
-	}
-
-	/* set top-left corner */
-	area->p0.x = x;
-	area->p0.y = y;
-	return 0;
-}
-
-/**
- * Find a place for a 2D area of given size inside a scan field based on its
- * alignment needs.
- *
- * @param w	width of desired area
- * @param h	height of desired area
- * @param align	desired area alignment
- * @param area	pointer to the area that will be set to the best position
- *
- * @return 0 on success, non-0 error value on failure.
- */
-static s32 scan_areas_and_find_fit(struct tcm *tcm, u16 w, u16 h, u16 align,
-				   struct tcm_area *area)
-{
-	s32 ret = 0;
-	struct tcm_area field = {0};
-	u16 boundary_x, boundary_y;
-	struct sita_pvt *pvt = (struct sita_pvt *)tcm->pvt;
-
-	if (align > 1) {
-		/* prefer top-left corner */
-		boundary_x = pvt->div_pt.x - 1;
-		boundary_y = pvt->div_pt.y - 1;
-
-		/* expand width and height if needed */
-		if (w > pvt->div_pt.x)
-			boundary_x = tcm->width - 1;
-		if (h > pvt->div_pt.y)
-			boundary_y = tcm->height - 1;
-
-		assign(&field, 0, 0, boundary_x, boundary_y);
-		ret = scan_l2r_t2b(tcm, w, h, align, &field, area);
-
-		/* scan whole container if failed, but do not scan 2x */
-		if (ret != 0 && (boundary_x != tcm->width - 1 ||
-				 boundary_y != tcm->height - 1)) {
-			/* scan the entire container if nothing found */
-			assign(&field, 0, 0, tcm->width - 1, tcm->height - 1);
-			ret = scan_l2r_t2b(tcm, w, h, align, &field, area);
-		}
-	} else if (align == 1) {
-		/* prefer top-right corner */
-		boundary_x = pvt->div_pt.x;
-		boundary_y = pvt->div_pt.y - 1;
-
-		/* expand width and height if needed */
-		if (w > (tcm->width - pvt->div_pt.x))
-			boundary_x = 0;
-		if (h > pvt->div_pt.y)
-			boundary_y = tcm->height - 1;
-
-		assign(&field, tcm->width - 1, 0, boundary_x, boundary_y);
-		ret = scan_r2l_t2b(tcm, w, h, align, &field, area);
-
-		/* scan whole container if failed, but do not scan 2x */
-		if (ret != 0 && (boundary_x != 0 ||
-				 boundary_y != tcm->height - 1)) {
-			/* scan the entire container if nothing found */
-			assign(&field, tcm->width - 1, 0, 0, tcm->height - 1);
-			ret = scan_r2l_t2b(tcm, w, h, align, &field,
-					   area);
-		}
-	}
-
-	return ret;
-}
-
-/* check if an entire area is free */
-static s32 is_area_free(struct tcm_area ***map, u16 x0, u16 y0, u16 w, u16 h)
-{
-	u16 x = 0, y = 0;
-	for (y = y0; y < y0 + h; y++) {
-		for (x = x0; x < x0 + w; x++) {
-			if (map[x][y])
-				return false;
-		}
-	}
-	return true;
-}
-
-/* fills an area with a parent tcm_area */
-static void fill_area(struct tcm *tcm, struct tcm_area *area,
-			struct tcm_area *parent)
-{
-	s32 x, y;
-	struct sita_pvt *pvt = (struct sita_pvt *)tcm->pvt;
-	struct tcm_area a, a_;
-
-	/* set area's tcm; otherwise, enumerator considers it invalid */
-	area->tcm = tcm;
-
-	tcm_for_each_slice(a, *area, a_) {
-		for (x = a.p0.x; x <= a.p1.x; ++x)
-			for (y = a.p0.y; y <= a.p1.y; ++y)
-				pvt->map[x][y] = parent;
-
-	}
-}
-
-/**
- * Compares a candidate area to the current best area, and if it is a better
- * fit, it updates the best to this one.
- *
- * @param x0, y0, w, h		top, left, width, height of candidate area
- * @param field			scan field
- * @param criteria		scan criteria
- * @param best			best candidate and its scores
- *
- * @return 1 (true) if the candidate area is known to be the final best, so no
- * more searching should be performed
- */
-static s32 update_candidate(struct tcm *tcm, u16 x0, u16 y0, u16 w, u16 h,
-			    struct tcm_area *field, s32 criteria,
-			    struct score *best)
-{
-	struct score me;	/* score for area */
-
-	/*
-	 * NOTE: For horizontal bias we always give the first found, because our
-	 * scan is horizontal-raster-based and the first candidate will always
-	 * have the horizontal bias.
-	 */
-	bool first = criteria & CR_BIAS_HORIZONTAL;
-
-	assign(&me.a, x0, y0, x0 + w - 1, y0 + h - 1);
-
-	/* calculate score for current candidate */
-	if (!first) {
-		get_neighbor_stats(tcm, &me.a, &me.n);
-		me.neighs = me.n.edge + me.n.busy;
-		get_nearness_factor(field, &me.a, &me.f);
-	}
-
-	/* the 1st candidate is always the best */
-	if (!best->a.tcm)
-		goto better;
-
-	BUG_ON(first);
-
-	/* diagonal balance check */
-	if ((criteria & CR_DIAGONAL_BALANCE) &&
-		best->neighs <= me.neighs &&
-		(best->neighs < me.neighs ||
-		 /* this implies that neighs and occupied match */
-		 best->n.busy < me.n.busy ||
-		 (best->n.busy == me.n.busy &&
-		  /* check the nearness factor */
-		  best->f.x + best->f.y > me.f.x + me.f.y)))
-		goto better;
-
-	/* not better, keep going */
-	return 0;
-
-better:
-	/* save current area as best */
-	memcpy(best, &me, sizeof(me));
-	best->a.tcm = tcm;
-	return first;
-}
-
-/**
- * Calculate the nearness factor of an area in a search field.  The nearness
- * factor is smaller if the area is closer to the search origin.
- */
-static void get_nearness_factor(struct tcm_area *field, struct tcm_area *area,
-				struct nearness_factor *nf)
-{
-	/**
-	 * Using signed math as field coordinates may be reversed if
-	 * search direction is right-to-left or bottom-to-top.
-	 */
-	nf->x = (s32)(area->p0.x - field->p0.x) * 1000 /
-		(field->p1.x - field->p0.x);
-	nf->y = (s32)(area->p0.y - field->p0.y) * 1000 /
-		(field->p1.y - field->p0.y);
-}
-
-/* get neighbor statistics */
-static void get_neighbor_stats(struct tcm *tcm, struct tcm_area *area,
-			 struct neighbor_stats *stat)
-{
-	s16 x = 0, y = 0;
-	struct sita_pvt *pvt = (struct sita_pvt *)tcm->pvt;
-
-	/* Clearing any exisiting values */
-	memset(stat, 0, sizeof(*stat));
-
-	/* process top & bottom edges */
-	for (x = area->p0.x; x <= area->p1.x; x++) {
-		if (area->p0.y == 0)
-			stat->edge++;
-		else if (pvt->map[x][area->p0.y - 1])
-			stat->busy++;
-
-		if (area->p1.y == tcm->height - 1)
-			stat->edge++;
-		else if (pvt->map[x][area->p1.y + 1])
-			stat->busy++;
-	}
-
-	/* process left & right edges */
-	for (y = area->p0.y; y <= area->p1.y; ++y) {
-		if (area->p0.x == 0)
-			stat->edge++;
-		else if (pvt->map[area->p0.x - 1][y])
-			stat->busy++;
-
-		if (area->p1.x == tcm->width - 1)
-			stat->edge++;
-		else if (pvt->map[area->p1.x + 1][y])
-			stat->busy++;
-	}
-}
diff --git a/drivers/gpu/drm/omapdrm/tcm.h b/drivers/gpu/drm/omapdrm/tcm.h
index a8d5ce47686f..ef7df7d6fc84 100644
--- a/drivers/gpu/drm/omapdrm/tcm.h
+++ b/drivers/gpu/drm/omapdrm/tcm.h
@@ -61,18 +61,17 @@ struct tcm {
 
 	unsigned int y_offset;	/* offset to use for y coordinates */
 
-	/* 'pvt' structure shall contain any tcm details (attr) along with
-	linked list of allocated areas and mutex for mutually exclusive access
-	to the list.  It may also contain copies of width and height to notice
-	any changes to the publicly available width and height fields. */
-	void *pvt;
+	spinlock_t lock;
+	unsigned long *bitmap;
+	size_t map_size;
 
 	/* function table */
-	s32 (*reserve_2d)(struct tcm *tcm, u16 height, u16 width, u8 align,
+	s32 (*reserve_2d)(struct tcm *tcm, u16 height, u16 width, u16 align,
+			  int16_t offset, uint16_t slot_bytes,
 			  struct tcm_area *area);
 	s32 (*reserve_1d)(struct tcm *tcm, u32 slots, struct tcm_area *area);
-	s32 (*free)      (struct tcm *tcm, struct tcm_area *area);
-	void (*deinit)   (struct tcm *tcm);
+	s32 (*free)(struct tcm *tcm, struct tcm_area *area);
+	void (*deinit)(struct tcm *tcm);
 };
 
 /*=============================================================================
@@ -91,7 +90,7 @@ struct tcm {
  *
  */
 
-struct tcm *sita_init(u16 width, u16 height, struct tcm_pt *attr);
+struct tcm *sita_init(u16 width, u16 height);
 
 
 /**
@@ -120,6 +119,9 @@ static inline void tcm_deinit(struct tcm *tcm)
  *			all values may be supported by the container manager,
  *			but it must support 0 (1), 32 and 64.
  *			0 value is equivalent to 1.
+ * @param offset	Offset requirement, in bytes.  This is the offset
+ *			from a 4KiB aligned virtual address.
+ * @param slot_bytes	Width of slot in bytes
  * @param area		Pointer to where the reserved area should be stored.
  *
  * @return 0 on success.  Non-0 error code on failure.  Also,
@@ -129,7 +131,8 @@ static inline void tcm_deinit(struct tcm *tcm)
  *	    allocation.
  */
 static inline s32 tcm_reserve_2d(struct tcm *tcm, u16 width, u16 height,
-				 u16 align, struct tcm_area *area)
+				u16 align, int16_t offset, uint16_t slot_bytes,
+				struct tcm_area *area)
 {
 	/* perform rudimentary error checking */
 	s32 res = tcm  == NULL ? -ENODEV :
@@ -140,7 +143,8 @@ static inline s32 tcm_reserve_2d(struct tcm *tcm, u16 width, u16 height,
 
 	if (!res) {
 		area->is2d = true;
-		res = tcm->reserve_2d(tcm, height, width, align, area);
+		res = tcm->reserve_2d(tcm, height, width, align, offset,
+					slot_bytes, area);
 		area->tcm = res ? NULL : tcm;
 	}
 
-- 
2.5.0



More information about the dri-devel mailing list