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/string.h>
29#include <linux/bitops.h>
30#include <drm/drmP.h>
31#include <drm/i915_drm.h>
32#include "i915_drv.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 */
87void
88i915_gem_detect_bit_6_swizzle(struct drm_device *dev)
89{
90 struct drm_i915_private *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 */
204static bool
205i915_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 /* i965+ stores the end address of the gtt mapping in the fence
221 * reg, so dont bother to check the size */
222 if (INTEL_INFO(dev)->gen >= 7) {
223 if (stride / 128 > GEN7_FENCE_MAX_PITCH_VAL)
224 return false;
225 } else if (INTEL_INFO(dev)->gen >= 4) {
226 if (stride / 128 > I965_FENCE_MAX_PITCH_VAL)
227 return false;
228 } else {
229 if (stride > 8192)
230 return false;
231
232 if (IS_GEN3(dev)) {
233 if (size > I830_FENCE_MAX_SIZE_VAL << 20)
234 return false;
235 } else {
236 if (size > I830_FENCE_MAX_SIZE_VAL << 19)
237 return false;
238 }
239 }
240
241 if (stride < tile_width)
242 return false;
243
244 /* 965+ just needs multiples of tile width */
245 if (INTEL_INFO(dev)->gen >= 4) {
246 if (stride & (tile_width - 1))
247 return false;
248 return true;
249 }
250
251 /* Pre-965 needs power of two tile widths */
252 if (stride & (stride - 1))
253 return false;
254
255 return true;
256}
257
258/* Is the current GTT allocation valid for the change in tiling? */
259static bool
260i915_gem_object_fence_ok(struct drm_i915_gem_object *obj, int tiling_mode)
261{
262 u32 size;
263
264 if (tiling_mode == I915_TILING_NONE)
265 return true;
266
267 if (INTEL_INFO(obj->base.dev)->gen >= 4)
268 return true;
269
270 if (INTEL_INFO(obj->base.dev)->gen == 3) {
271 if (i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK)
272 return false;
273 } else {
274 if (i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK)
275 return false;
276 }
277
278 size = i915_gem_get_gtt_size(obj->base.dev, obj->base.size, tiling_mode);
279 if (i915_gem_obj_ggtt_size(obj) != size)
280 return false;
281
282 if (i915_gem_obj_ggtt_offset(obj) & (size - 1))
283 return false;
284
285 return true;
286}
287
288/**
289 * Sets the tiling mode of an object, returning the required swizzling of
290 * bit 6 of addresses in the object.
291 */
292int
293i915_gem_set_tiling(struct drm_device *dev, void *data,
294 struct drm_file *file)
295{
296 struct drm_i915_gem_set_tiling *args = data;
297 struct drm_i915_private *dev_priv = dev->dev_private;
298 struct drm_gem_object *gobj;
299 struct drm_i915_gem_object *obj;
300 int ret = 0;
301
302 gobj = drm_gem_object_lookup(dev, file, args->handle);
303 if (gobj == NULL)
304 return -ENOENT;
305 obj = to_intel_bo(gobj);
306
307 if (!i915_tiling_ok(dev,
308 args->stride, obj->base.size, args->tiling_mode)) {
309 drm_gem_object_unreference_unlocked(&obj->base);
310 return -EINVAL;
311 }
312
313 if (i915_gem_obj_is_pinned(obj) || obj->framebuffer_references) {
314 drm_gem_object_unreference_unlocked(&obj->base);
315 return -EBUSY;
316 }
317
318 if (args->tiling_mode == I915_TILING_NONE) {
319 args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
320 args->stride = 0;
321 } else {
322 if (args->tiling_mode == I915_TILING_X)
323 args->swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
324 else
325 args->swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
326
327 /* Hide bit 17 swizzling from the user. This prevents old Mesa
328 * from aborting the application on sw fallbacks to bit 17,
329 * and we use the pread/pwrite bit17 paths to swizzle for it.
330 * If there was a user that was relying on the swizzle
331 * information for drm_intel_bo_map()ed reads/writes this would
332 * break it, but we don't have any of those.
333 */
334 if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_17)
335 args->swizzle_mode = I915_BIT_6_SWIZZLE_9;
336 if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_10_17)
337 args->swizzle_mode = I915_BIT_6_SWIZZLE_9_10;
338
339 /* If we can't handle the swizzling, make it untiled. */
340 if (args->swizzle_mode == I915_BIT_6_SWIZZLE_UNKNOWN) {
341 args->tiling_mode = I915_TILING_NONE;
342 args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
343 args->stride = 0;
344 }
345 }
346
347 mutex_lock(&dev->struct_mutex);
348 if (args->tiling_mode != obj->tiling_mode ||
349 args->stride != obj->stride) {
350 /* We need to rebind the object if its current allocation
351 * no longer meets the alignment restrictions for its new
352 * tiling mode. Otherwise we can just leave it alone, but
353 * need to ensure that any fence register is updated before
354 * the next fenced (either through the GTT or by the BLT unit
355 * on older GPUs) access.
356 *
357 * After updating the tiling parameters, we then flag whether
358 * we need to update an associated fence register. Note this
359 * has to also include the unfenced register the GPU uses
360 * whilst executing a fenced command for an untiled object.
361 */
362
363 obj->map_and_fenceable =
364 !i915_gem_obj_ggtt_bound(obj) ||
365 (i915_gem_obj_ggtt_offset(obj) +
366 obj->base.size <= dev_priv->gtt.mappable_end &&
367 i915_gem_object_fence_ok(obj, args->tiling_mode));
368
369 /* Rebind if we need a change of alignment */
370 if (!obj->map_and_fenceable) {
371 u32 unfenced_align =
372 i915_gem_get_gtt_alignment(dev, obj->base.size,
373 args->tiling_mode,
374 false);
375 if (i915_gem_obj_ggtt_offset(obj) & (unfenced_align - 1))
376 ret = i915_gem_object_ggtt_unbind(obj);
377 }
378
379 if (ret == 0) {
380 obj->fence_dirty =
381 obj->fenced_gpu_access ||
382 obj->fence_reg != I915_FENCE_REG_NONE;
383
384 obj->tiling_mode = args->tiling_mode;
385 obj->stride = args->stride;
386
387 /* Force the fence to be reacquired for GTT access */
388 i915_gem_release_mmap(obj);
389 }
390 }
391 /* we have to maintain this existing ABI... */
392 args->stride = obj->stride;
393 args->tiling_mode = obj->tiling_mode;
394
395 /* Try to preallocate memory required to save swizzling on put-pages */
396 if (i915_gem_object_needs_bit17_swizzle(obj)) {
397 if (obj->bit_17 == NULL) {
398 obj->bit_17 = kcalloc(BITS_TO_LONGS(obj->base.size >> PAGE_SHIFT),
399 sizeof(long), GFP_KERNEL);
400 }
401 } else {
402 kfree(obj->bit_17);
403 obj->bit_17 = NULL;
404 }
405
406 drm_gem_object_unreference(&obj->base);
407 mutex_unlock(&dev->struct_mutex);
408
409 return ret;
410}
411
412/**
413 * Returns the current tiling mode and required bit 6 swizzling for the object.
414 */
415int
416i915_gem_get_tiling(struct drm_device *dev, void *data,
417 struct drm_file *file)
418{
419 struct drm_i915_gem_get_tiling *args = data;
420 struct drm_i915_private *dev_priv = dev->dev_private;
421 struct drm_gem_object *gobj;
422 struct drm_i915_gem_object *obj;
423
424 gobj = drm_gem_object_lookup(dev, file, args->handle);
425 if (gobj == NULL)
426 return -ENOENT;
427 obj = to_intel_bo(gobj);
428
429 mutex_lock(&dev->struct_mutex);
430
431 args->tiling_mode = obj->tiling_mode;
432 switch (obj->tiling_mode) {
433 case I915_TILING_X:
434 args->swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
435 break;
436 case I915_TILING_Y:
437 args->swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
438 break;
439 case I915_TILING_NONE:
440 args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
441 break;
442 default:
443 DRM_ERROR("unknown tiling mode\n");
444 }
445
446 /* Hide bit 17 from the user -- see comment in i915_gem_set_tiling */
447 if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_17)
448 args->swizzle_mode = I915_BIT_6_SWIZZLE_9;
449 if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_10_17)
450 args->swizzle_mode = I915_BIT_6_SWIZZLE_9_10;
451
452 drm_gem_object_unreference(&obj->base);
453 mutex_unlock(&dev->struct_mutex);
454
455 return 0;
456}
457
458/**
459 * Swap every 64 bytes of this page around, to account for it having a new
460 * bit 17 of its physical address and therefore being interpreted differently
461 * by the GPU.
462 */
463static void
464i915_gem_swizzle_page(struct page *page)
465{
466 char temp[64];
467 char *vaddr;
468 int i;
469
470 vaddr = kmap(page);
471
472 for (i = 0; i < PAGE_SIZE; i += 128) {
473 memcpy(temp, &vaddr[i], 64);
474 memcpy(&vaddr[i], &vaddr[i + 64], 64);
475 memcpy(&vaddr[i + 64], temp, 64);
476 }
477
478 kunmap(page);
479}
480
481void
482i915_gem_object_do_bit_17_swizzle(struct drm_i915_gem_object *obj)
483{
484#ifdef __NetBSD__
485 struct vm_page *page;
486#else
487 struct sg_page_iter sg_iter;
488#endif
489 int i;
490
491 if (obj->bit_17 == NULL)
492 return;
493
494#ifdef __NetBSD__
495 i = 0;
496 TAILQ_FOREACH(page, &obj->igo_pageq, pageq.queue) {
497 unsigned char new_bit_17 = VM_PAGE_TO_PHYS(page) >> 17;
498 if ((new_bit_17 & 0x1) !=
499 (test_bit(i, obj->bit_17) != 0)) {
500 i915_gem_swizzle_page(container_of(page, struct page,
501 p_vmp));
502 page->flags &= ~PG_CLEAN;
503 }
504 i += 1;
505 }
506#else
507 i = 0;
508 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
509 struct page *page = sg_page_iter_page(&sg_iter);
510 char new_bit_17 = page_to_phys(page) >> 17;
511 if ((new_bit_17 & 0x1) !=
512 (test_bit(i, obj->bit_17) != 0)) {
513 i915_gem_swizzle_page(page);
514 set_page_dirty(page);
515 }
516 i++;
517 }
518#endif
519}
520
521void
522i915_gem_object_save_bit_17_swizzle(struct drm_i915_gem_object *obj)
523{
524#ifdef __NetBSD__
525 struct vm_page *page;
526#else
527 struct sg_page_iter sg_iter;
528#endif
529 int page_count = obj->base.size >> PAGE_SHIFT;
530 int i;
531
532 if (obj->bit_17 == NULL) {
533 obj->bit_17 = kcalloc(BITS_TO_LONGS(page_count),
534 sizeof(long), GFP_KERNEL);
535 if (obj->bit_17 == NULL) {
536 DRM_ERROR("Failed to allocate memory for bit 17 "
537 "record\n");
538 return;
539 }
540 }
541
542 i = 0;
543#ifdef __NetBSD__
544 TAILQ_FOREACH(page, &obj->igo_pageq, pageq.queue) {
545 if (ISSET(VM_PAGE_TO_PHYS(page), __BIT(17)))
546 __set_bit(i, obj->bit_17);
547 else
548 __clear_bit(i, obj->bit_17);
549 i += 1;
550 }
551#else
552 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) {
553 if (page_to_phys(sg_page_iter_page(&sg_iter)) & (1 << 17))
554 __set_bit(i, obj->bit_17);
555 else
556 __clear_bit(i, obj->bit_17);
557 i++;
558 }
559#endif
560}
561