xref: /titanic_41/usr/src/uts/intel/io/drm/i915_gem_tiling.c (revision 79a77829f1ca134b5058f1269fe5a7a52b874aa9)
1 /* BEGIN CSTYLED */
2 
3 /*
4  * Copyright (c) 2009, Intel Corporation.
5  * All Rights Reserved.
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a
8  * copy of this software and associated documentation files (the "Software"),
9  * to deal in the Software without restriction, including without limitation
10  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11  * and/or sell copies of the Software, and to permit persons to whom the
12  * Software is furnished to do so, subject to the following conditions:
13  *
14  * The above copyright notice and this permission notice (including the next
15  * paragraph) shall be included in all copies or substantial portions of the
16  * Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
21  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24  * IN THE SOFTWARE.
25  *
26  * Authors:
27  *    Eric Anholt <eric@anholt.net>
28  *
29  */
30 
31 /*
32  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
33  * Use is subject to license terms.
34  */
35 
36 #include <sys/sysmacros.h>
37 #include "drmP.h"
38 #include "drm.h"
39 #include "i915_drm.h"
40 #include "i915_drv.h"
41 
42 /** @file i915_gem_tiling.c
43  *
44  * Support for managing tiling state of buffer objects.
45  *
46  * The idea behind tiling is to increase cache hit rates by rearranging
47  * pixel data so that a group of pixel accesses are in the same cacheline.
48  * Performance improvement from doing this on the back/depth buffer are on
49  * the order of 30%.
50  *
51  * Intel architectures make this somewhat more complicated, though, by
52  * adjustments made to addressing of data when the memory is in interleaved
53  * mode (matched pairs of DIMMS) to improve memory bandwidth.
54  * For interleaved memory, the CPU sends every sequential 64 bytes
55  * to an alternate memory channel so it can get the bandwidth from both.
56  *
57  * The GPU also rearranges its accesses for increased bandwidth to interleaved
58  * memory, and it matches what the CPU does for non-tiled.  However, when tiled
59  * it does it a little differently, since one walks addresses not just in the
60  * X direction but also Y.  So, along with alternating channels when bit
61  * 6 of the address flips, it also alternates when other bits flip --  Bits 9
62  * (every 512 bytes, an X tile scanline) and 10 (every two X tile scanlines)
63  * are common to both the 915 and 965-class hardware.
64  *
65  * The CPU also sometimes XORs in higher bits as well, to improve
66  * bandwidth doing strided access like we do so frequently in graphics.  This
67  * is called "Channel XOR Randomization" in the MCH documentation.  The result
68  * is that the CPU is XORing in either bit 11 or bit 17 to bit 6 of its address
69  * decode.
70  *
71  * All of this bit 6 XORing has an effect on our memory management,
72  * as we need to make sure that the 3d driver can correctly address object
73  * contents.
74  *
75  * If we don't have interleaved memory, all tiling is safe and no swizzling is
76  * required.
77  *
78  * When bit 17 is XORed in, we simply refuse to tile at all.  Bit
79  * 17 is not just a page offset, so as we page an objet out and back in,
80  * individual pages in it will have different bit 17 addresses, resulting in
81  * each 64 bytes being swapped with its neighbor!
82  *
83  * Otherwise, if interleaved, we have to tell the 3d driver what the address
84  * swizzling it needs to do is, since it's writing with the CPU to the pages
85  * (bit 6 and potentially bit 11 XORed in), and the GPU is reading from the
86  * pages (bit 6, 9, and 10 XORed in), resulting in a cumulative bit swizzling
87  * required by the CPU of XORing in bit 6, 9, 10, and potentially 11, in order
88  * to match what the GPU expects.
89  */
90 
91 /**
92  * Detects bit 6 swizzling of address lookup between IGD access and CPU
93  * access through main memory.
94  */
95 void
i915_gem_detect_bit_6_swizzle(struct drm_device * dev)96 i915_gem_detect_bit_6_swizzle(struct drm_device *dev)
97 {
98 	drm_i915_private_t *dev_priv = dev->dev_private;
99 	uint32_t swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
100 	uint32_t swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
101 
102 	if (!IS_I9XX(dev)) {
103 		/* As far as we know, the 865 doesn't have these bit 6
104 		 * swizzling issues.
105 		 */
106 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
107 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
108 		} else if (IS_MOBILE(dev)) {
109 		uint32_t dcc;
110 
111 		/* On mobile 9xx chipsets, channel interleave by the CPU is
112 		 * determined by DCC.  For single-channel, neither the CPU
113 		 * nor the GPU do swizzling.  For dual channel interleaved,
114 		 * the GPU's interleave is bit 9 and 10 for X tiled, and bit
115 		 * 9 for Y tiled.  The CPU's interleave is independent, and
116 		 * can be based on either bit 11 (haven't seen this yet) or
117 		 * bit 17 (common).
118 		 */
119 
120 		dcc = I915_READ(DCC);
121 		switch (dcc & DCC_ADDRESSING_MODE_MASK) {
122 		case DCC_ADDRESSING_MODE_SINGLE_CHANNEL:
123 		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_ASYMMETRIC:
124 			swizzle_x = I915_BIT_6_SWIZZLE_NONE;
125 			swizzle_y = I915_BIT_6_SWIZZLE_NONE;
126 			break;
127 		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_INTERLEAVED:
128 			if (dcc & DCC_CHANNEL_XOR_DISABLE) {
129 				/* This is the base swizzling by the GPU for
130 				 * tiled buffers.
131 				 */
132 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
133 				swizzle_y = I915_BIT_6_SWIZZLE_9;
134 			} else if ((dcc & DCC_CHANNEL_XOR_BIT_17) == 0) {
135 				/* Bit 11 swizzling by the CPU in addition. */
136 				swizzle_x = I915_BIT_6_SWIZZLE_9_10_11;
137 				swizzle_y = I915_BIT_6_SWIZZLE_9_11;
138 			} else {
139 				/* Bit 17 swizzling by the CPU in addition. */
140 				swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
141 				swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
142 			}
143 			break;
144 		}
145 		if (dcc == 0xffffffff) {
146 			DRM_ERROR("Couldn't read from MCHBAR.  "
147 				  "Disabling tiling.\n");
148 			swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
149 			swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
150 		}
151 	} else {
152 		/* The 965, G33, and newer, have a very flexible memory
153 		 * configuration.  It will enable dual-channel mode
154 		 * (interleaving) on as much memory as it can, and the GPU
155 		 * will additionally sometimes enable different bit 6
156 		 * swizzling for tiled objects from the CPU.
157 		 *
158 		 * Here's what I found on the G965:
159 		 *    slot fill         memory size  swizzling
160 		 * 0A   0B   1A   1B    1-ch   2-ch
161 		 * 512  0    0    0     512    0     O
162 		 * 512  0    512  0     16     1008  X
163 		 * 512  0    0    512   16     1008  X
164 		 * 0    512  0    512   16     1008  X
165 		 * 1024 1024 1024 0     2048   1024  O
166 		 *
167 		 * We could probably detect this based on either the DRB
168 		 * matching, which was the case for the swizzling required in
169 		 * the table above, or from the 1-ch value being less than
170 		 * the minimum size of a rank.
171 		 */
172 		if (I915_READ16(C0DRB3) != I915_READ16(C1DRB3)) {
173 			swizzle_x = I915_BIT_6_SWIZZLE_NONE;
174 			swizzle_y = I915_BIT_6_SWIZZLE_NONE;
175 		} else {
176 			swizzle_x = I915_BIT_6_SWIZZLE_9_10;
177 			swizzle_y = I915_BIT_6_SWIZZLE_9;
178 		}
179 	}
180 
181        /* FIXME: check with memory config on IGDNG */
182        if (IS_IGDNG(dev)) {
183                swizzle_x = I915_BIT_6_SWIZZLE_9_10;
184                swizzle_y = I915_BIT_6_SWIZZLE_9;
185        }
186 
187 	dev_priv->mm.bit_6_swizzle_x = swizzle_x;
188 	dev_priv->mm.bit_6_swizzle_y = swizzle_y;
189 }
190 
191 
192 /**
193  * Returns the size of the fence for a tiled object of the given size.
194  */
195 static int
i915_get_fence_size(struct drm_device * dev,int size)196 i915_get_fence_size(struct drm_device *dev, int size)
197 {
198 	int i;
199 	int start;
200 
201 	if (IS_I965G(dev)) {
202 		/* The 965 can have fences at any page boundary. */
203 
204 		return (size + PAGE_SIZE-1) & ~(PAGE_SIZE-1);
205 	} else {
206 		/* Align the size to a power of two greater than the smallest
207 		 * fence size.
208 		 */
209 		if (IS_I9XX(dev))
210 			start = 1024 * 1024;
211 		else
212 			start = 512 * 1024;
213 
214 		for (i = start; i < size; i <<= 1)
215 			;
216 
217 		return i;
218 	}
219 }
220 
221 /* Check pitch constriants for all chips & tiling formats */
222 static int
i915_tiling_ok(struct drm_device * dev,int stride,int size,int tiling_mode)223 i915_tiling_ok(struct drm_device *dev, int stride, int size, int tiling_mode)
224 {
225 	int tile_width;
226 
227 	/* Linear is always fine */
228 	if (tiling_mode == I915_TILING_NONE)
229 		return 1;
230 
231 	if (tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
232 		tile_width = 128;
233 	else
234 		tile_width = 512;
235 
236 	if (stride == 0)
237 		return 0;
238 
239 	/* 965+ just needs multiples of tile width */
240 	if (IS_I965G(dev)) {
241 		if (stride & (tile_width - 1))
242 			return 0;
243 		return 1;
244 	}
245 
246 	/* Pre-965 needs power of two tile widths */
247 	if (stride < tile_width)
248 		return 0;
249 
250 	if (!ISP2(stride))
251 		return 0;
252 
253 	/* We don't handle the aperture area covered by the fence being bigger
254 	 * than the object size.
255 	 */
256 	if (i915_get_fence_size(dev, size) != size)
257 		return 0;
258 
259 	return 1;
260 }
261 
262 /**
263  * Sets the tiling mode of an object, returning the required swizzling of
264  * bit 6 of addresses in the object.
265  */
266 /*ARGSUSED*/
267 int
i915_gem_set_tiling(DRM_IOCTL_ARGS)268 i915_gem_set_tiling(DRM_IOCTL_ARGS)
269 {
270 	DRM_DEVICE;
271 	struct drm_i915_gem_set_tiling args;
272 	drm_i915_private_t *dev_priv = dev->dev_private;
273 	struct drm_gem_object *obj;
274 	struct drm_i915_gem_object *obj_priv;
275 	int ret;
276 
277 	if (dev->driver->use_gem != 1)
278 		return ENODEV;
279 
280 	DRM_COPYFROM_WITH_RETURN(&args,
281             (struct drm_i915_gem_set_tiling __user *) data, sizeof(args));
282 
283 	obj = drm_gem_object_lookup(fpriv, args.handle);
284 	if (obj == NULL)
285 		return EINVAL;
286 	obj_priv = obj->driver_private;
287 
288 	if (!i915_tiling_ok(dev, args.stride, obj->size, args.tiling_mode)) {
289 		drm_gem_object_unreference(obj);
290 		DRM_DEBUG("i915 tiling is not OK");
291 		return EINVAL;
292 	}
293 
294 	spin_lock(&dev->struct_mutex);
295 
296 	if (args.tiling_mode == I915_TILING_NONE) {
297 		args.swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
298 	} else {
299 		if (args.tiling_mode == I915_TILING_X)
300 			args.swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
301 		else
302 			args.swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
303 		/* If we can't handle the swizzling, make it untiled. */
304 		if (args.swizzle_mode == I915_BIT_6_SWIZZLE_UNKNOWN) {
305 			args.tiling_mode = I915_TILING_NONE;
306 			args.swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
307 		}
308 	}
309 
310 	if (args.tiling_mode != obj_priv->tiling_mode) {
311 		int ret;
312 
313 		/* Unbind the object, as switching tiling means we're
314 		 * switching the cache organization due to fencing, probably.
315 		 */
316 		ret = i915_gem_object_unbind(obj, 1);
317 		if (ret != 0) {
318 			args.tiling_mode = obj_priv->tiling_mode;
319 			spin_unlock(&dev->struct_mutex);
320 			drm_gem_object_unreference(obj);
321 			DRM_ERROR("tiling switch!! unbind error %d", ret);
322 			return ret;
323 		}
324 		obj_priv->tiling_mode = args.tiling_mode;
325 	}
326 	obj_priv->stride = args.stride;
327 
328 	ret = DRM_COPY_TO_USER((struct drm_i915_gem_set_tiling __user *) data, &args, sizeof(args));
329         if ( ret != 0)
330                 DRM_ERROR(" gem set tiling error! %d", ret);
331 
332 	drm_gem_object_unreference(obj);
333 	spin_unlock(&dev->struct_mutex);
334 
335 	return 0;
336 }
337 
338 /**
339  * Returns the current tiling mode and required bit 6 swizzling for the object.
340  */
341 /*ARGSUSED*/
342 int
i915_gem_get_tiling(DRM_IOCTL_ARGS)343 i915_gem_get_tiling(DRM_IOCTL_ARGS)
344 {
345 	DRM_DEVICE;
346 	struct drm_i915_gem_get_tiling args;
347 	drm_i915_private_t *dev_priv = dev->dev_private;
348 	struct drm_gem_object *obj;
349 	struct drm_i915_gem_object *obj_priv;
350 	int ret;
351 
352 	if (dev->driver->use_gem != 1)
353 		return ENODEV;
354 
355 	DRM_COPYFROM_WITH_RETURN(&args,
356 	    (struct drm_i915_gem_get_tiling __user *) data, sizeof(args));
357 
358 	obj = drm_gem_object_lookup(fpriv, args.handle);
359 	if (obj == NULL)
360 		return EINVAL;
361 	obj_priv = obj->driver_private;
362 
363 	spin_lock(&dev->struct_mutex);
364 
365 	args.tiling_mode = obj_priv->tiling_mode;
366 	switch (obj_priv->tiling_mode) {
367 	case I915_TILING_X:
368 		args.swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
369 		break;
370 	case I915_TILING_Y:
371 		args.swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
372 		break;
373 	case I915_TILING_NONE:
374 		args.swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
375 		break;
376 	default:
377 		DRM_ERROR("unknown tiling mode\n");
378 	}
379 
380 
381 
382 	ret = DRM_COPY_TO_USER((struct drm_i915_gem_get_tiling __user *) data, &args, sizeof(args));
383         if ( ret != 0)
384                 DRM_ERROR(" gem get tiling error! %d", ret);
385 
386 	drm_gem_object_unreference(obj);
387 	spin_unlock(&dev->struct_mutex);
388 
389 	return 0;
390 }
391