2 * Copyright © 2008 Intel Corporation
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:
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
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
24 * Eric Anholt <eric@anholt.net>
28 #include "linux/string.h"
29 #include "linux/bitops.h"
35 /** @file i915_gem_tiling.c
37 * Support for managing tiling state of buffer objects.
39 * The idea behind tiling is to increase cache hit rates by rearranging
40 * pixel data so that a group of pixel accesses are in the same cacheline.
41 * Performance improvement from doing this on the back/depth buffer are on
44 * Intel architectures make this somewhat more complicated, though, by
45 * adjustments made to addressing of data when the memory is in interleaved
46 * mode (matched pairs of DIMMS) to improve memory bandwidth.
47 * For interleaved memory, the CPU sends every sequential 64 bytes
48 * to an alternate memory channel so it can get the bandwidth from both.
50 * The GPU also rearranges its accesses for increased bandwidth to interleaved
51 * memory, and it matches what the CPU does for non-tiled. However, when tiled
52 * it does it a little differently, since one walks addresses not just in the
53 * X direction but also Y. So, along with alternating channels when bit
54 * 6 of the address flips, it also alternates when other bits flip -- Bits 9
55 * (every 512 bytes, an X tile scanline) and 10 (every two X tile scanlines)
56 * are common to both the 915 and 965-class hardware.
58 * The CPU also sometimes XORs in higher bits as well, to improve
59 * bandwidth doing strided access like we do so frequently in graphics. This
60 * is called "Channel XOR Randomization" in the MCH documentation. The result
61 * is that the CPU is XORing in either bit 11 or bit 17 to bit 6 of its address
64 * All of this bit 6 XORing has an effect on our memory management,
65 * as we need to make sure that the 3d driver can correctly address object
68 * If we don't have interleaved memory, all tiling is safe and no swizzling is
71 * When bit 17 is XORed in, we simply refuse to tile at all. Bit
72 * 17 is not just a page offset, so as we page an objet out and back in,
73 * individual pages in it will have different bit 17 addresses, resulting in
74 * each 64 bytes being swapped with its neighbor!
76 * Otherwise, if interleaved, we have to tell the 3d driver what the address
77 * swizzling it needs to do is, since it's writing with the CPU to the pages
78 * (bit 6 and potentially bit 11 XORed in), and the GPU is reading from the
79 * pages (bit 6, 9, and 10 XORed in), resulting in a cumulative bit swizzling
80 * required by the CPU of XORing in bit 6, 9, 10, and potentially 11, in order
81 * to match what the GPU expects.
85 * Detects bit 6 swizzling of address lookup between IGD access and CPU
86 * access through main memory.
89 i915_gem_detect_bit_6_swizzle(struct drm_device *dev)
91 drm_i915_private_t *dev_priv = dev->dev_private;
92 uint32_t swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
93 uint32_t swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
96 /* As far as we know, the 865 doesn't have these bit 6
99 swizzle_x = I915_BIT_6_SWIZZLE_NONE;
100 swizzle_y = I915_BIT_6_SWIZZLE_NONE;
101 } else if (IS_MOBILE(dev)) {
104 /* On mobile 9xx chipsets, channel interleave by the CPU is
105 * determined by DCC. For single-channel, neither the CPU
106 * nor the GPU do swizzling. For dual channel interleaved,
107 * the GPU's interleave is bit 9 and 10 for X tiled, and bit
108 * 9 for Y tiled. The CPU's interleave is independent, and
109 * can be based on either bit 11 (haven't seen this yet) or
112 dcc = I915_READ(DCC);
113 switch (dcc & DCC_ADDRESSING_MODE_MASK) {
114 case DCC_ADDRESSING_MODE_SINGLE_CHANNEL:
115 case DCC_ADDRESSING_MODE_DUAL_CHANNEL_ASYMMETRIC:
116 swizzle_x = I915_BIT_6_SWIZZLE_NONE;
117 swizzle_y = I915_BIT_6_SWIZZLE_NONE;
119 case DCC_ADDRESSING_MODE_DUAL_CHANNEL_INTERLEAVED:
120 if (dcc & DCC_CHANNEL_XOR_DISABLE) {
121 /* This is the base swizzling by the GPU for
124 swizzle_x = I915_BIT_6_SWIZZLE_9_10;
125 swizzle_y = I915_BIT_6_SWIZZLE_9;
126 } else if ((dcc & DCC_CHANNEL_XOR_BIT_17) == 0) {
127 /* Bit 11 swizzling by the CPU in addition. */
128 swizzle_x = I915_BIT_6_SWIZZLE_9_10_11;
129 swizzle_y = I915_BIT_6_SWIZZLE_9_11;
131 /* Bit 17 swizzling by the CPU in addition. */
132 swizzle_x = I915_BIT_6_SWIZZLE_9_10_17;
133 swizzle_y = I915_BIT_6_SWIZZLE_9_17;
137 if (dcc == 0xffffffff) {
138 DRM_ERROR("Couldn't read from MCHBAR. "
139 "Disabling tiling.\n");
140 swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
141 swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
144 /* The 965, G33, and newer, have a very flexible memory
145 * configuration. It will enable dual-channel mode
146 * (interleaving) on as much memory as it can, and the GPU
147 * will additionally sometimes enable different bit 6
148 * swizzling for tiled objects from the CPU.
150 * Here's what I found on the G965:
151 * slot fill memory size swizzling
152 * 0A 0B 1A 1B 1-ch 2-ch
154 * 512 0 512 0 16 1008 X
155 * 512 0 0 512 16 1008 X
156 * 0 512 0 512 16 1008 X
157 * 1024 1024 1024 0 2048 1024 O
159 * We could probably detect this based on either the DRB
160 * matching, which was the case for the swizzling required in
161 * the table above, or from the 1-ch value being less than
162 * the minimum size of a rank.
164 if (I915_READ16(C0DRB3) != I915_READ16(C1DRB3)) {
165 swizzle_x = I915_BIT_6_SWIZZLE_NONE;
166 swizzle_y = I915_BIT_6_SWIZZLE_NONE;
168 swizzle_x = I915_BIT_6_SWIZZLE_9_10;
169 swizzle_y = I915_BIT_6_SWIZZLE_9;
173 dev_priv->mm.bit_6_swizzle_x = swizzle_x;
174 dev_priv->mm.bit_6_swizzle_y = swizzle_y;
179 * Returns the size of the fence for a tiled object of the given size.
182 i915_get_fence_size(struct drm_device *dev, int size)
188 /* The 965 can have fences at any page boundary. */
189 return ALIGN(size, 4096);
191 /* Align the size to a power of two greater than the smallest
199 for (i = start; i < size; i <<= 1)
206 /* Check pitch constriants for all chips & tiling formats */
208 i915_tiling_ok(struct drm_device *dev, int stride, int size, int tiling_mode)
212 /* Linear is always fine */
213 if (tiling_mode == I915_TILING_NONE)
216 if (tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
221 /* check maximum stride & object size */
223 /* i965 stores the end address of the gtt mapping in the fence
224 * reg, so dont bother to check the size */
225 if (stride / 128 > I965_FENCE_MAX_PITCH_VAL)
227 } else if (IS_I9XX(dev)) {
228 if (stride / tile_width > I830_FENCE_MAX_PITCH_VAL ||
229 size > (I830_FENCE_MAX_SIZE_VAL << 20))
232 if (stride / 128 > I830_FENCE_MAX_PITCH_VAL ||
233 size > (I830_FENCE_MAX_SIZE_VAL << 19))
237 /* 965+ just needs multiples of tile width */
239 if (stride & (tile_width - 1))
244 /* Pre-965 needs power of two tile widths */
245 if (stride < tile_width)
248 if (stride & (stride - 1))
251 /* We don't handle the aperture area covered by the fence being bigger
252 * than the object size.
254 if (i915_get_fence_size(dev, size) != size)
261 * Sets the tiling mode of an object, returning the required swizzling of
262 * bit 6 of addresses in the object.
265 i915_gem_set_tiling(struct drm_device *dev, void *data,
266 struct drm_file *file_priv)
268 struct drm_i915_gem_set_tiling *args = data;
269 drm_i915_private_t *dev_priv = dev->dev_private;
270 struct drm_gem_object *obj;
271 struct drm_i915_gem_object *obj_priv;
273 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
276 obj_priv = obj->driver_private;
278 if (!i915_tiling_ok(dev, args->stride, obj->size, args->tiling_mode)) {
279 drm_gem_object_unreference(obj);
283 mutex_lock(&dev->struct_mutex);
285 if (args->tiling_mode == I915_TILING_NONE) {
286 args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
288 if (args->tiling_mode == I915_TILING_X)
289 args->swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
291 args->swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
293 /* Hide bit 17 swizzling from the user. This prevents old Mesa
294 * from aborting the application on sw fallbacks to bit 17,
295 * and we use the pread/pwrite bit17 paths to swizzle for it.
296 * If there was a user that was relying on the swizzle
297 * information for drm_intel_bo_map()ed reads/writes this would
298 * break it, but we don't have any of those.
300 if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_17)
301 args->swizzle_mode = I915_BIT_6_SWIZZLE_9;
302 if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_10_17)
303 args->swizzle_mode = I915_BIT_6_SWIZZLE_9_10;
305 /* If we can't handle the swizzling, make it untiled. */
306 if (args->swizzle_mode == I915_BIT_6_SWIZZLE_UNKNOWN) {
307 args->tiling_mode = I915_TILING_NONE;
308 args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
311 if (args->tiling_mode != obj_priv->tiling_mode) {
314 /* Unbind the object, as switching tiling means we're
315 * switching the cache organization due to fencing, probably.
317 ret = i915_gem_object_unbind(obj);
319 WARN(ret != -ERESTARTSYS,
320 "failed to unbind object for tiling switch");
321 args->tiling_mode = obj_priv->tiling_mode;
322 mutex_unlock(&dev->struct_mutex);
323 drm_gem_object_unreference(obj);
327 obj_priv->tiling_mode = args->tiling_mode;
329 obj_priv->stride = args->stride;
331 drm_gem_object_unreference(obj);
332 mutex_unlock(&dev->struct_mutex);
338 * Returns the current tiling mode and required bit 6 swizzling for the object.
341 i915_gem_get_tiling(struct drm_device *dev, void *data,
342 struct drm_file *file_priv)
344 struct drm_i915_gem_get_tiling *args = data;
345 drm_i915_private_t *dev_priv = dev->dev_private;
346 struct drm_gem_object *obj;
347 struct drm_i915_gem_object *obj_priv;
349 obj = drm_gem_object_lookup(dev, file_priv, args->handle);
352 obj_priv = obj->driver_private;
354 mutex_lock(&dev->struct_mutex);
356 args->tiling_mode = obj_priv->tiling_mode;
357 switch (obj_priv->tiling_mode) {
359 args->swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
362 args->swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
364 case I915_TILING_NONE:
365 args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
368 DRM_ERROR("unknown tiling mode\n");
371 /* Hide bit 17 from the user -- see comment in i915_gem_set_tiling */
372 if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_17)
373 args->swizzle_mode = I915_BIT_6_SWIZZLE_9;
374 if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_10_17)
375 args->swizzle_mode = I915_BIT_6_SWIZZLE_9_10;
377 drm_gem_object_unreference(obj);
378 mutex_unlock(&dev->struct_mutex);
384 * Swap every 64 bytes of this page around, to account for it having a new
385 * bit 17 of its physical address and therefore being interpreted differently
389 i915_gem_swizzle_page(struct page *page)
399 for (i = 0; i < PAGE_SIZE; i += 128) {
400 memcpy(temp, &vaddr[i], 64);
401 memcpy(&vaddr[i], &vaddr[i + 64], 64);
402 memcpy(&vaddr[i + 64], temp, 64);
411 i915_gem_object_do_bit_17_swizzle(struct drm_gem_object *obj)
413 struct drm_device *dev = obj->dev;
414 drm_i915_private_t *dev_priv = dev->dev_private;
415 struct drm_i915_gem_object *obj_priv = obj->driver_private;
416 int page_count = obj->size >> PAGE_SHIFT;
419 if (dev_priv->mm.bit_6_swizzle_x != I915_BIT_6_SWIZZLE_9_10_17)
422 if (obj_priv->bit_17 == NULL)
425 for (i = 0; i < page_count; i++) {
426 char new_bit_17 = page_to_phys(obj_priv->pages[i]) >> 17;
427 if ((new_bit_17 & 0x1) !=
428 (test_bit(i, obj_priv->bit_17) != 0)) {
429 int ret = i915_gem_swizzle_page(obj_priv->pages[i]);
431 DRM_ERROR("Failed to swizzle page\n");
434 set_page_dirty(obj_priv->pages[i]);
440 i915_gem_object_save_bit_17_swizzle(struct drm_gem_object *obj)
442 struct drm_device *dev = obj->dev;
443 drm_i915_private_t *dev_priv = dev->dev_private;
444 struct drm_i915_gem_object *obj_priv = obj->driver_private;
445 int page_count = obj->size >> PAGE_SHIFT;
448 if (dev_priv->mm.bit_6_swizzle_x != I915_BIT_6_SWIZZLE_9_10_17)
451 if (obj_priv->bit_17 == NULL) {
452 obj_priv->bit_17 = kmalloc(BITS_TO_LONGS(page_count) *
453 sizeof(long), GFP_KERNEL);
454 if (obj_priv->bit_17 == NULL) {
455 DRM_ERROR("Failed to allocate memory for bit 17 "
461 for (i = 0; i < page_count; i++) {
462 if (page_to_phys(obj_priv->pages[i]) & (1 << 17))
463 __set_bit(i, obj_priv->bit_17);
465 __clear_bit(i, obj_priv->bit_17);