xref: /dflybsd-src/sys/dev/drm/i915/i915_gem_tiling.c (revision 2a1ad637466621af45d5a17185b33f3dcaaa1b1c)
1 /*
2  * Copyright © 2008 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  * Authors:
24  *    Eric Anholt <eric@anholt.net>
25  *
26  */
27 
28 #include <linux/bitops.h>
29 #include <drm/drmP.h>
30 #include <drm/i915_drm.h>
31 #include "i915_drv.h"
32 #include <linux/highmem.h>
33 
34 /** @file i915_gem_tiling.c
35  *
36  * Support for managing tiling state of buffer objects.
37  *
38  * The idea behind tiling is to increase cache hit rates by rearranging
39  * pixel data so that a group of pixel accesses are in the same cacheline.
40  * Performance improvement from doing this on the back/depth buffer are on
41  * the order of 30%.
42  *
43  * Intel architectures make this somewhat more complicated, though, by
44  * adjustments made to addressing of data when the memory is in interleaved
45  * mode (matched pairs of DIMMS) to improve memory bandwidth.
46  * For interleaved memory, the CPU sends every sequential 64 bytes
47  * to an alternate memory channel so it can get the bandwidth from both.
48  *
49  * The GPU also rearranges its accesses for increased bandwidth to interleaved
50  * memory, and it matches what the CPU does for non-tiled.  However, when tiled
51  * it does it a little differently, since one walks addresses not just in the
52  * X direction but also Y.  So, along with alternating channels when bit
53  * 6 of the address flips, it also alternates when other bits flip --  Bits 9
54  * (every 512 bytes, an X tile scanline) and 10 (every two X tile scanlines)
55  * are common to both the 915 and 965-class hardware.
56  *
57  * The CPU also sometimes XORs in higher bits as well, to improve
58  * bandwidth doing strided access like we do so frequently in graphics.  This
59  * is called "Channel XOR Randomization" in the MCH documentation.  The result
60  * is that the CPU is XORing in either bit 11 or bit 17 to bit 6 of its address
61  * decode.
62  *
63  * All of this bit 6 XORing has an effect on our memory management,
64  * as we need to make sure that the 3d driver can correctly address object
65  * contents.
66  *
67  * If we don't have interleaved memory, all tiling is safe and no swizzling is
68  * required.
69  *
70  * When bit 17 is XORed in, we simply refuse to tile at all.  Bit
71  * 17 is not just a page offset, so as we page an objet out and back in,
72  * individual pages in it will have different bit 17 addresses, resulting in
73  * each 64 bytes being swapped with its neighbor!
74  *
75  * Otherwise, if interleaved, we have to tell the 3d driver what the address
76  * swizzling it needs to do is, since it's writing with the CPU to the pages
77  * (bit 6 and potentially bit 11 XORed in), and the GPU is reading from the
78  * pages (bit 6, 9, and 10 XORed in), resulting in a cumulative bit swizzling
79  * required by the CPU of XORing in bit 6, 9, 10, and potentially 11, in order
80  * to match what the GPU expects.
81  */
82 
83 /**
84  * Detects bit 6 swizzling of address lookup between IGD access and CPU
85  * access through main memory.
86  */
87 void
88 i915_gem_detect_bit_6_swizzle(struct drm_device *dev)
89 {
90 	drm_i915_private_t *dev_priv = dev->dev_private;
91 	uint32_t swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
92 	uint32_t swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
93 
94 	if (IS_VALLEYVIEW(dev)) {
95 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
96 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
97 	} else if (INTEL_INFO(dev)->gen >= 6) {
98 		uint32_t dimm_c0, dimm_c1;
99 		dimm_c0 = I915_READ(MAD_DIMM_C0);
100 		dimm_c1 = I915_READ(MAD_DIMM_C1);
101 		dimm_c0 &= MAD_DIMM_A_SIZE_MASK | MAD_DIMM_B_SIZE_MASK;
102 		dimm_c1 &= MAD_DIMM_A_SIZE_MASK | MAD_DIMM_B_SIZE_MASK;
103 		/* Enable swizzling when the channels are populated with
104 		 * identically sized dimms. We don't need to check the 3rd
105 		 * channel because no cpu with gpu attached ships in that
106 		 * configuration. Also, swizzling only makes sense for 2
107 		 * channels anyway. */
108 		if (dimm_c0 == dimm_c1) {
109 			swizzle_x = I915_BIT_6_SWIZZLE_9_10;
110 			swizzle_y = I915_BIT_6_SWIZZLE_9;
111 		} else {
112 			swizzle_x = I915_BIT_6_SWIZZLE_NONE;
113 			swizzle_y = I915_BIT_6_SWIZZLE_NONE;
114 		}
115 	} else if (IS_GEN5(dev)) {
116 		/* On Ironlake whatever DRAM config, GPU always do
117 		 * same swizzling setup.
118 		 */
119 		swizzle_x = I915_BIT_6_SWIZZLE_9_10;
120 		swizzle_y = I915_BIT_6_SWIZZLE_9;
121 	} else if (IS_GEN2(dev)) {
122 		/* As far as we know, the 865 doesn't have these bit 6
123 		 * swizzling issues.
124 		 */
125 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
126 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
127 	} else if (IS_MOBILE(dev) || (IS_GEN3(dev) && !IS_G33(dev))) {
128 		uint32_t dcc;
129 
130 		/* On 9xx chipsets, channel interleave by the CPU is
131 		 * determined by DCC.  For single-channel, neither the CPU
132 		 * nor the GPU do swizzling.  For dual channel interleaved,
133 		 * the GPU's interleave is bit 9 and 10 for X tiled, and bit
134 		 * 9 for Y tiled.  The CPU's interleave is independent, and
135 		 * can be based on either bit 11 (haven't seen this yet) or
136 		 * bit 17 (common).
137 		 */
138 		dcc = I915_READ(DCC);
139 		switch (dcc & DCC_ADDRESSING_MODE_MASK) {
140 		case DCC_ADDRESSING_MODE_SINGLE_CHANNEL:
141 		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_ASYMMETRIC:
142 			swizzle_x = I915_BIT_6_SWIZZLE_NONE;
143 			swizzle_y = I915_BIT_6_SWIZZLE_NONE;
144 			break;
145 		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_INTERLEAVED:
146 			if (dcc & DCC_CHANNEL_XOR_DISABLE) {
147 				/* This is the base swizzling by the GPU for
148 				 * tiled buffers.
149 				 */
150 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
151 				swizzle_y = I915_BIT_6_SWIZZLE_9;
152 			} else if ((dcc & DCC_CHANNEL_XOR_BIT_17) == 0) {
153 				/* Bit 11 swizzling by the CPU in addition. */
154 				swizzle_x = I915_BIT_6_SWIZZLE_9_10_11;
155 				swizzle_y = I915_BIT_6_SWIZZLE_9_11;
156 			} else {
157 				/* Bit 17 swizzling by the CPU in addition. */
158 				swizzle_x = I915_BIT_6_SWIZZLE_9_10_17;
159 				swizzle_y = I915_BIT_6_SWIZZLE_9_17;
160 			}
161 			break;
162 		}
163 		if (dcc == 0xffffffff) {
164 			DRM_ERROR("Couldn't read from MCHBAR.  "
165 				  "Disabling tiling.\n");
166 			swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
167 			swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
168 		}
169 	} else {
170 		/* The 965, G33, and newer, have a very flexible memory
171 		 * configuration.  It will enable dual-channel mode
172 		 * (interleaving) on as much memory as it can, and the GPU
173 		 * will additionally sometimes enable different bit 6
174 		 * swizzling for tiled objects from the CPU.
175 		 *
176 		 * Here's what I found on the G965:
177 		 *    slot fill         memory size  swizzling
178 		 * 0A   0B   1A   1B    1-ch   2-ch
179 		 * 512  0    0    0     512    0     O
180 		 * 512  0    512  0     16     1008  X
181 		 * 512  0    0    512   16     1008  X
182 		 * 0    512  0    512   16     1008  X
183 		 * 1024 1024 1024 0     2048   1024  O
184 		 *
185 		 * We could probably detect this based on either the DRB
186 		 * matching, which was the case for the swizzling required in
187 		 * the table above, or from the 1-ch value being less than
188 		 * the minimum size of a rank.
189 		 */
190 		if (I915_READ16(C0DRB3) != I915_READ16(C1DRB3)) {
191 			swizzle_x = I915_BIT_6_SWIZZLE_NONE;
192 			swizzle_y = I915_BIT_6_SWIZZLE_NONE;
193 		} else {
194 			swizzle_x = I915_BIT_6_SWIZZLE_9_10;
195 			swizzle_y = I915_BIT_6_SWIZZLE_9;
196 		}
197 	}
198 
199 	dev_priv->mm.bit_6_swizzle_x = swizzle_x;
200 	dev_priv->mm.bit_6_swizzle_y = swizzle_y;
201 }
202 
203 /* Check pitch constriants for all chips & tiling formats */
204 static bool
205 i915_tiling_ok(struct drm_device *dev, int stride, int size, int tiling_mode)
206 {
207 	int tile_width;
208 
209 	/* Linear is always fine */
210 	if (tiling_mode == I915_TILING_NONE)
211 		return true;
212 
213 	if (IS_GEN2(dev) ||
214 	    (tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev)))
215 		tile_width = 128;
216 	else
217 		tile_width = 512;
218 
219 	/* check maximum stride & object size */
220 	if (INTEL_INFO(dev)->gen >= 4) {
221 		/* i965 stores the end address of the gtt mapping in the fence
222 		 * reg, so dont bother to check the size */
223 		if (stride / 128 > I965_FENCE_MAX_PITCH_VAL)
224 			return false;
225 	} else {
226 		if (stride > 8192)
227 			return false;
228 
229 		if (IS_GEN3(dev)) {
230 			if (size > I830_FENCE_MAX_SIZE_VAL << 20)
231 				return false;
232 		} else {
233 			if (size > I830_FENCE_MAX_SIZE_VAL << 19)
234 				return false;
235 		}
236 	}
237 
238 	/* 965+ just needs multiples of tile width */
239 	if (INTEL_INFO(dev)->gen >= 4) {
240 		if (stride & (tile_width - 1))
241 			return false;
242 		return true;
243 	}
244 
245 	/* Pre-965 needs power of two tile widths */
246 	if (stride < tile_width)
247 		return false;
248 
249 	if (stride & (stride - 1))
250 		return false;
251 
252 	return true;
253 }
254 
255 /* Is the current GTT allocation valid for the change in tiling? */
256 static bool
257 i915_gem_object_fence_ok(struct drm_i915_gem_object *obj, int tiling_mode)
258 {
259 	u32 size;
260 
261 	if (tiling_mode == I915_TILING_NONE)
262 		return true;
263 
264 	if (INTEL_INFO(obj->base.dev)->gen >= 4)
265 		return true;
266 
267 	if (INTEL_INFO(obj->base.dev)->gen == 3) {
268 		if (obj->gtt_offset & ~I915_FENCE_START_MASK)
269 			return false;
270 	} else {
271 		if (obj->gtt_offset & ~I830_FENCE_START_MASK)
272 			return false;
273 	}
274 
275 	size = i915_gem_get_gtt_size(obj->base.dev, obj->base.size, tiling_mode);
276 	if (obj->gtt_space->size != size)
277 		return false;
278 
279 	if (obj->gtt_offset & (size - 1))
280 		return false;
281 
282 	return true;
283 }
284 
285 /**
286  * Sets the tiling mode of an object, returning the required swizzling of
287  * bit 6 of addresses in the object.
288  */
289 int
290 i915_gem_set_tiling(struct drm_device *dev, void *data,
291 		   struct drm_file *file)
292 {
293 	struct drm_i915_gem_set_tiling *args = data;
294 	drm_i915_private_t *dev_priv = dev->dev_private;
295 	struct drm_i915_gem_object *obj;
296 	int ret = 0;
297 
298 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
299 	if (&obj->base == NULL)
300 		return -ENOENT;
301 
302 	if (!i915_tiling_ok(dev,
303 			    args->stride, obj->base.size, args->tiling_mode)) {
304 		drm_gem_object_unreference_unlocked(&obj->base);
305 		return -EINVAL;
306 	}
307 
308 	if (obj->pin_count) {
309 		drm_gem_object_unreference_unlocked(&obj->base);
310 		return -EBUSY;
311 	}
312 
313 	if (args->tiling_mode == I915_TILING_NONE) {
314 		args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
315 		args->stride = 0;
316 	} else {
317 		if (args->tiling_mode == I915_TILING_X)
318 			args->swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
319 		else
320 			args->swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
321 
322 		/* Hide bit 17 swizzling from the user.  This prevents old Mesa
323 		 * from aborting the application on sw fallbacks to bit 17,
324 		 * and we use the pread/pwrite bit17 paths to swizzle for it.
325 		 * If there was a user that was relying on the swizzle
326 		 * information for drm_intel_bo_map()ed reads/writes this would
327 		 * break it, but we don't have any of those.
328 		 */
329 		if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_17)
330 			args->swizzle_mode = I915_BIT_6_SWIZZLE_9;
331 		if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_10_17)
332 			args->swizzle_mode = I915_BIT_6_SWIZZLE_9_10;
333 
334 		/* If we can't handle the swizzling, make it untiled. */
335 		if (args->swizzle_mode == I915_BIT_6_SWIZZLE_UNKNOWN) {
336 			args->tiling_mode = I915_TILING_NONE;
337 			args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
338 			args->stride = 0;
339 		}
340 	}
341 
342 	mutex_lock(&dev->struct_mutex);
343 	if (args->tiling_mode != obj->tiling_mode ||
344 	    args->stride != obj->stride) {
345 		/* We need to rebind the object if its current allocation
346 		 * no longer meets the alignment restrictions for its new
347 		 * tiling mode. Otherwise we can just leave it alone, but
348 		 * need to ensure that any fence register is updated before
349 		 * the next fenced (either through the GTT or by the BLT unit
350 		 * on older GPUs) access.
351 		 *
352 		 * After updating the tiling parameters, we then flag whether
353 		 * we need to update an associated fence register. Note this
354 		 * has to also include the unfenced register the GPU uses
355 		 * whilst executing a fenced command for an untiled object.
356 		 */
357 
358 		obj->map_and_fenceable =
359 			obj->gtt_space == NULL ||
360 			(obj->gtt_offset + obj->base.size <= dev_priv->gtt.mappable_end &&
361 			 i915_gem_object_fence_ok(obj, args->tiling_mode));
362 
363 		/* Rebind if we need a change of alignment */
364 		if (!obj->map_and_fenceable) {
365 			u32 unfenced_alignment =
366 				i915_gem_get_gtt_alignment(dev, obj->base.size,
367 							    args->tiling_mode,
368 							    false);
369 			if (obj->gtt_offset & (unfenced_alignment - 1))
370 				ret = i915_gem_object_unbind(obj);
371 		}
372 
373 		if (ret == 0) {
374 			obj->fence_dirty =
375 				obj->fenced_gpu_access ||
376 				obj->fence_reg != I915_FENCE_REG_NONE;
377 
378 			obj->tiling_mode = args->tiling_mode;
379 			obj->stride = args->stride;
380 
381 			/* Force the fence to be reacquired for GTT access */
382 			i915_gem_release_mmap(obj);
383 		}
384 	}
385 	/* we have to maintain this existing ABI... */
386 	args->stride = obj->stride;
387 	args->tiling_mode = obj->tiling_mode;
388 
389 	/* Try to preallocate memory required to save swizzling on put-pages */
390 	if (i915_gem_object_needs_bit17_swizzle(obj)) {
391 		if (obj->bit_17 == NULL) {
392 			obj->bit_17 = kmalloc(BITS_TO_LONGS(obj->base.size >> PAGE_SHIFT) *
393 					      sizeof(long), M_DRM, M_WAITOK);
394 		}
395 	} else {
396 		drm_free(obj->bit_17, M_DRM);
397 		obj->bit_17 = NULL;
398 	}
399 
400 	drm_gem_object_unreference(&obj->base);
401 	mutex_unlock(&dev->struct_mutex);
402 
403 	return ret;
404 }
405 
406 /**
407  * Returns the current tiling mode and required bit 6 swizzling for the object.
408  */
409 int
410 i915_gem_get_tiling(struct drm_device *dev, void *data,
411 		   struct drm_file *file)
412 {
413 	struct drm_i915_gem_get_tiling *args = data;
414 	drm_i915_private_t *dev_priv = dev->dev_private;
415 	struct drm_i915_gem_object *obj;
416 
417 	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
418 	if (&obj->base == NULL)
419 		return -ENOENT;
420 
421 	mutex_lock(&dev->struct_mutex);
422 
423 	args->tiling_mode = obj->tiling_mode;
424 	switch (obj->tiling_mode) {
425 	case I915_TILING_X:
426 		args->swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
427 		break;
428 	case I915_TILING_Y:
429 		args->swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
430 		break;
431 	case I915_TILING_NONE:
432 		args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
433 		break;
434 	default:
435 		DRM_ERROR("unknown tiling mode\n");
436 	}
437 
438 	/* Hide bit 17 from the user -- see comment in i915_gem_set_tiling */
439 	if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_17)
440 		args->swizzle_mode = I915_BIT_6_SWIZZLE_9;
441 	if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_10_17)
442 		args->swizzle_mode = I915_BIT_6_SWIZZLE_9_10;
443 
444 	drm_gem_object_unreference(&obj->base);
445 	mutex_unlock(&dev->struct_mutex);
446 
447 	return 0;
448 }
449 
450 /**
451  * Swap every 64 bytes of this page around, to account for it having a new
452  * bit 17 of its physical address and therefore being interpreted differently
453  * by the GPU.
454  */
455 static void
456 i915_gem_swizzle_page(struct vm_page *page)
457 {
458 	char temp[64];
459 	char *vaddr;
460 	int i;
461 
462 	vaddr = kmap(page);
463 
464 	for (i = 0; i < PAGE_SIZE; i += 128) {
465 		memcpy(temp, &vaddr[i], 64);
466 		memcpy(&vaddr[i], &vaddr[i + 64], 64);
467 		memcpy(&vaddr[i + 64], temp, 64);
468 	}
469 
470 	kunmap(page);
471 }
472 
473 void
474 i915_gem_object_do_bit_17_swizzle(struct drm_i915_gem_object *obj)
475 {
476 	int page_count = obj->base.size >> PAGE_SHIFT;
477 	int i;
478 
479 	if (obj->bit_17 == NULL)
480 		return;
481 
482 	for (i = 0; i < page_count; i++) {
483 		char new_bit_17 = VM_PAGE_TO_PHYS(obj->pages[i]) >> 17;
484 		if ((new_bit_17 & 0x1) !=
485 		    (test_bit(i, obj->bit_17) != 0)) {
486 			i915_gem_swizzle_page(obj->pages[i]);
487 			vm_page_dirty(obj->pages[i]);
488 		}
489 	}
490 }
491 
492 void
493 i915_gem_object_save_bit_17_swizzle(struct drm_i915_gem_object *obj)
494 {
495 	int page_count = obj->base.size >> PAGE_SHIFT;
496 	int i;
497 
498 	if (obj->bit_17 == NULL) {
499 		obj->bit_17 = kmalloc(BITS_TO_LONGS(page_count) *
500 					   sizeof(long), M_DRM, M_WAITOK);
501 		if (obj->bit_17 == NULL) {
502 			DRM_ERROR("Failed to allocate memory for bit 17 "
503 				  "record\n");
504 			return;
505 		}
506 	}
507 
508 	for (i = 0; i < page_count; i++) {
509 		if (VM_PAGE_TO_PHYS(obj->pages[i]) & (1 << 17))
510 			__set_bit(i, obj->bit_17);
511 		else
512 			__clear_bit(i, obj->bit_17);
513 	}
514 }
515