Linux Audio

Check our new training course

Loading...
v3.1
  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 "drmP.h"
 31#include "drm.h"
 32#include "i915_drm.h"
 33#include "i915_drv.h"
 34
 35/** @file i915_gem_tiling.c
 
 
 
 
 
 
 
 
 36 *
 37 * Support for managing tiling state of buffer objects.
 
 
 
 
 
 
 
 
 
 38 *
 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
 42 * the order of 30%.
 43 *
 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.
 49 *
 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.
 57 *
 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
 62 * decode.
 63 *
 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
 66 * contents.
 67 *
 68 * If we don't have interleaved memory, all tiling is safe and no swizzling is
 69 * required.
 70 *
 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!
 75 *
 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.
 82 */
 83
 84/**
 85 * Detects bit 6 swizzling of address lookup between IGD access and CPU
 86 * access through main memory.
 
 
 
 
 
 
 87 */
 88void
 89i915_gem_detect_bit_6_swizzle(struct drm_device *dev)
 90{
 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;
 94
 95	if (INTEL_INFO(dev)->gen >= 5) {
 96		/* On Ironlake whatever DRAM config, GPU always do
 97		 * same swizzling setup.
 98		 */
 99		swizzle_x = I915_BIT_6_SWIZZLE_9_10;
100		swizzle_y = I915_BIT_6_SWIZZLE_9;
101	} else if (IS_GEN2(dev)) {
102		/* As far as we know, the 865 doesn't have these bit 6
103		 * swizzling issues.
104		 */
105		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
106		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
107	} else if (IS_MOBILE(dev)) {
108		uint32_t dcc;
109
110		/* On mobile 9xx chipsets, channel interleave by the CPU is
111		 * determined by DCC.  For single-channel, neither the CPU
112		 * nor the GPU do swizzling.  For dual channel interleaved,
113		 * the GPU's interleave is bit 9 and 10 for X tiled, and bit
114		 * 9 for Y tiled.  The CPU's interleave is independent, and
115		 * can be based on either bit 11 (haven't seen this yet) or
116		 * bit 17 (common).
117		 */
118		dcc = I915_READ(DCC);
119		switch (dcc & DCC_ADDRESSING_MODE_MASK) {
120		case DCC_ADDRESSING_MODE_SINGLE_CHANNEL:
121		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_ASYMMETRIC:
122			swizzle_x = I915_BIT_6_SWIZZLE_NONE;
123			swizzle_y = I915_BIT_6_SWIZZLE_NONE;
124			break;
125		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_INTERLEAVED:
126			if (dcc & DCC_CHANNEL_XOR_DISABLE) {
127				/* This is the base swizzling by the GPU for
128				 * tiled buffers.
129				 */
130				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
131				swizzle_y = I915_BIT_6_SWIZZLE_9;
132			} else if ((dcc & DCC_CHANNEL_XOR_BIT_17) == 0) {
133				/* Bit 11 swizzling by the CPU in addition. */
134				swizzle_x = I915_BIT_6_SWIZZLE_9_10_11;
135				swizzle_y = I915_BIT_6_SWIZZLE_9_11;
136			} else {
137				/* Bit 17 swizzling by the CPU in addition. */
138				swizzle_x = I915_BIT_6_SWIZZLE_9_10_17;
139				swizzle_y = I915_BIT_6_SWIZZLE_9_17;
140			}
141			break;
142		}
143		if (dcc == 0xffffffff) {
144			DRM_ERROR("Couldn't read from MCHBAR.  "
145				  "Disabling tiling.\n");
146			swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
147			swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
148		}
149	} else {
150		/* The 965, G33, and newer, have a very flexible memory
151		 * configuration.  It will enable dual-channel mode
152		 * (interleaving) on as much memory as it can, and the GPU
153		 * will additionally sometimes enable different bit 6
154		 * swizzling for tiled objects from the CPU.
155		 *
156		 * Here's what I found on the G965:
157		 *    slot fill         memory size  swizzling
158		 * 0A   0B   1A   1B    1-ch   2-ch
159		 * 512  0    0    0     512    0     O
160		 * 512  0    512  0     16     1008  X
161		 * 512  0    0    512   16     1008  X
162		 * 0    512  0    512   16     1008  X
163		 * 1024 1024 1024 0     2048   1024  O
164		 *
165		 * We could probably detect this based on either the DRB
166		 * matching, which was the case for the swizzling required in
167		 * the table above, or from the 1-ch value being less than
168		 * the minimum size of a rank.
169		 */
170		if (I915_READ16(C0DRB3) != I915_READ16(C1DRB3)) {
171			swizzle_x = I915_BIT_6_SWIZZLE_NONE;
172			swizzle_y = I915_BIT_6_SWIZZLE_NONE;
173		} else {
174			swizzle_x = I915_BIT_6_SWIZZLE_9_10;
175			swizzle_y = I915_BIT_6_SWIZZLE_9;
176		}
177	}
178
179	dev_priv->mm.bit_6_swizzle_x = swizzle_x;
180	dev_priv->mm.bit_6_swizzle_y = swizzle_y;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
181}
182
183/* Check pitch constriants for all chips & tiling formats */
184static bool
185i915_tiling_ok(struct drm_device *dev, int stride, int size, int tiling_mode)
 
186{
187	int tile_width;
 
188
189	/* Linear is always fine */
190	if (tiling_mode == I915_TILING_NONE)
191		return true;
192
193	if (IS_GEN2(dev) ||
194	    (tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev)))
195		tile_width = 128;
196	else
197		tile_width = 512;
198
199	/* check maximum stride & object size */
200	if (INTEL_INFO(dev)->gen >= 4) {
201		/* i965 stores the end address of the gtt mapping in the fence
202		 * reg, so dont bother to check the size */
 
 
 
203		if (stride / 128 > I965_FENCE_MAX_PITCH_VAL)
204			return false;
205	} else {
206		if (stride > 8192)
207			return false;
208
209		if (IS_GEN3(dev)) {
210			if (size > I830_FENCE_MAX_SIZE_VAL << 20)
211				return false;
212		} else {
213			if (size > I830_FENCE_MAX_SIZE_VAL << 19)
214				return false;
215		}
216	}
217
218	/* 965+ just needs multiples of tile width */
219	if (INTEL_INFO(dev)->gen >= 4) {
220		if (stride & (tile_width - 1))
221			return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
222		return true;
223	}
224
225	/* Pre-965 needs power of two tile widths */
226	if (stride < tile_width)
227		return false;
228
229	if (stride & (stride - 1))
 
230		return false;
231
232	return true;
233}
234
235/* Is the current GTT allocation valid for the change in tiling? */
236static bool
237i915_gem_object_fence_ok(struct drm_i915_gem_object *obj, int tiling_mode)
 
238{
239	u32 size;
 
240
241	if (tiling_mode == I915_TILING_NONE)
242		return true;
243
244	if (INTEL_INFO(obj->base.dev)->gen >= 4)
245		return true;
246
247	if (INTEL_INFO(obj->base.dev)->gen == 3) {
248		if (obj->gtt_offset & ~I915_FENCE_START_MASK)
249			return false;
250	} else {
251		if (obj->gtt_offset & ~I830_FENCE_START_MASK)
252			return false;
 
253	}
254
255	/*
256	 * Previous chips need to be aligned to the size of the smallest
257	 * fence register that can contain the object.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
258	 */
259	if (INTEL_INFO(obj->base.dev)->gen == 3)
260		size = 1024*1024;
261	else
262		size = 512*1024;
263
264	while (size < obj->base.size)
265		size <<= 1;
 
 
 
 
 
 
 
266
267	if (obj->gtt_space->size != size)
268		return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
269
270	if (obj->gtt_offset & (size - 1))
271		return false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
272
273	return true;
274}
275
276/**
 
 
 
 
 
277 * Sets the tiling mode of an object, returning the required swizzling of
278 * bit 6 of addresses in the object.
 
 
 
 
 
279 */
280int
281i915_gem_set_tiling(struct drm_device *dev, void *data,
282		   struct drm_file *file)
283{
284	struct drm_i915_gem_set_tiling *args = data;
285	drm_i915_private_t *dev_priv = dev->dev_private;
286	struct drm_i915_gem_object *obj;
287	int ret = 0;
288
289	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
290	if (&obj->base == NULL)
291		return -ENOENT;
292
293	if (!i915_tiling_ok(dev,
294			    args->stride, obj->base.size, args->tiling_mode)) {
295		drm_gem_object_unreference_unlocked(&obj->base);
296		return -EINVAL;
 
 
 
297	}
298
299	if (obj->pin_count) {
300		drm_gem_object_unreference_unlocked(&obj->base);
301		return -EBUSY;
302	}
303
304	if (args->tiling_mode == I915_TILING_NONE) {
305		args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
306		args->stride = 0;
307	} else {
308		if (args->tiling_mode == I915_TILING_X)
309			args->swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
310		else
311			args->swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
312
313		/* Hide bit 17 swizzling from the user.  This prevents old Mesa
314		 * from aborting the application on sw fallbacks to bit 17,
315		 * and we use the pread/pwrite bit17 paths to swizzle for it.
316		 * If there was a user that was relying on the swizzle
317		 * information for drm_intel_bo_map()ed reads/writes this would
318		 * break it, but we don't have any of those.
319		 */
320		if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_17)
321			args->swizzle_mode = I915_BIT_6_SWIZZLE_9;
322		if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_10_17)
323			args->swizzle_mode = I915_BIT_6_SWIZZLE_9_10;
324
325		/* If we can't handle the swizzling, make it untiled. */
326		if (args->swizzle_mode == I915_BIT_6_SWIZZLE_UNKNOWN) {
327			args->tiling_mode = I915_TILING_NONE;
328			args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
329			args->stride = 0;
330		}
331	}
332
333	mutex_lock(&dev->struct_mutex);
334	if (args->tiling_mode != obj->tiling_mode ||
335	    args->stride != obj->stride) {
336		/* We need to rebind the object if its current allocation
337		 * no longer meets the alignment restrictions for its new
338		 * tiling mode. Otherwise we can just leave it alone, but
339		 * need to ensure that any fence register is cleared.
340		 */
341		i915_gem_release_mmap(obj);
342
343		obj->map_and_fenceable =
344			obj->gtt_space == NULL ||
345			(obj->gtt_offset + obj->base.size <= dev_priv->mm.gtt_mappable_end &&
346			 i915_gem_object_fence_ok(obj, args->tiling_mode));
347
348		/* Rebind if we need a change of alignment */
349		if (!obj->map_and_fenceable) {
350			u32 unfenced_alignment =
351				i915_gem_get_unfenced_gtt_alignment(dev,
352								    obj->base.size,
353								    args->tiling_mode);
354			if (obj->gtt_offset & (unfenced_alignment - 1))
355				ret = i915_gem_object_unbind(obj);
356		}
357
358		if (ret == 0) {
359			obj->tiling_changed = true;
360			obj->tiling_mode = args->tiling_mode;
361			obj->stride = args->stride;
362		}
363	}
364	/* we have to maintain this existing ABI... */
365	args->stride = obj->stride;
366	args->tiling_mode = obj->tiling_mode;
367	drm_gem_object_unreference(&obj->base);
368	mutex_unlock(&dev->struct_mutex);
369
370	return ret;
 
 
 
 
 
 
371}
372
373/**
 
 
 
 
 
374 * Returns the current tiling mode and required bit 6 swizzling for the object.
 
 
 
 
 
375 */
376int
377i915_gem_get_tiling(struct drm_device *dev, void *data,
378		   struct drm_file *file)
379{
380	struct drm_i915_gem_get_tiling *args = data;
381	drm_i915_private_t *dev_priv = dev->dev_private;
382	struct drm_i915_gem_object *obj;
 
383
384	obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
385	if (&obj->base == NULL)
386		return -ENOENT;
 
 
 
 
 
 
 
387
388	mutex_lock(&dev->struct_mutex);
389
390	args->tiling_mode = obj->tiling_mode;
391	switch (obj->tiling_mode) {
392	case I915_TILING_X:
393		args->swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
394		break;
395	case I915_TILING_Y:
396		args->swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
397		break;
 
398	case I915_TILING_NONE:
399		args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
400		break;
401	default:
402		DRM_ERROR("unknown tiling mode\n");
403	}
404
405	/* Hide bit 17 from the user -- see comment in i915_gem_set_tiling */
 
 
 
 
406	if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_17)
407		args->swizzle_mode = I915_BIT_6_SWIZZLE_9;
408	if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_10_17)
409		args->swizzle_mode = I915_BIT_6_SWIZZLE_9_10;
410
411	drm_gem_object_unreference(&obj->base);
412	mutex_unlock(&dev->struct_mutex);
413
414	return 0;
415}
416
417/**
418 * Swap every 64 bytes of this page around, to account for it having a new
419 * bit 17 of its physical address and therefore being interpreted differently
420 * by the GPU.
421 */
422static void
423i915_gem_swizzle_page(struct page *page)
424{
425	char temp[64];
426	char *vaddr;
427	int i;
428
429	vaddr = kmap(page);
430
431	for (i = 0; i < PAGE_SIZE; i += 128) {
432		memcpy(temp, &vaddr[i], 64);
433		memcpy(&vaddr[i], &vaddr[i + 64], 64);
434		memcpy(&vaddr[i + 64], temp, 64);
435	}
436
437	kunmap(page);
438}
439
440void
441i915_gem_object_do_bit_17_swizzle(struct drm_i915_gem_object *obj)
442{
443	struct drm_device *dev = obj->base.dev;
444	drm_i915_private_t *dev_priv = dev->dev_private;
445	int page_count = obj->base.size >> PAGE_SHIFT;
446	int i;
447
448	if (dev_priv->mm.bit_6_swizzle_x != I915_BIT_6_SWIZZLE_9_10_17)
449		return;
450
451	if (obj->bit_17 == NULL)
452		return;
453
454	for (i = 0; i < page_count; i++) {
455		char new_bit_17 = page_to_phys(obj->pages[i]) >> 17;
456		if ((new_bit_17 & 0x1) !=
457		    (test_bit(i, obj->bit_17) != 0)) {
458			i915_gem_swizzle_page(obj->pages[i]);
459			set_page_dirty(obj->pages[i]);
460		}
461	}
462}
463
464void
465i915_gem_object_save_bit_17_swizzle(struct drm_i915_gem_object *obj)
466{
467	struct drm_device *dev = obj->base.dev;
468	drm_i915_private_t *dev_priv = dev->dev_private;
469	int page_count = obj->base.size >> PAGE_SHIFT;
470	int i;
471
472	if (dev_priv->mm.bit_6_swizzle_x != I915_BIT_6_SWIZZLE_9_10_17)
473		return;
474
475	if (obj->bit_17 == NULL) {
476		obj->bit_17 = kmalloc(BITS_TO_LONGS(page_count) *
477					   sizeof(long), GFP_KERNEL);
478		if (obj->bit_17 == NULL) {
479			DRM_ERROR("Failed to allocate memory for bit 17 "
480				  "record\n");
481			return;
482		}
483	}
484
485	for (i = 0; i < page_count; i++) {
486		if (page_to_phys(obj->pages[i]) & (1 << 17))
487			__set_bit(i, obj->bit_17);
488		else
489			__clear_bit(i, obj->bit_17);
490	}
491}
v4.17
  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/**
 35 * DOC: buffer object tiling
 36 *
 37 * i915_gem_set_tiling_ioctl() and i915_gem_get_tiling_ioctl() is the userspace
 38 * interface to declare fence register requirements.
 39 *
 40 * In principle GEM doesn't care at all about the internal data layout of an
 41 * object, and hence it also doesn't care about tiling or swizzling. There's two
 42 * exceptions:
 43 *
 44 * - For X and Y tiling the hardware provides detilers for CPU access, so called
 45 *   fences. Since there's only a limited amount of them the kernel must manage
 46 *   these, and therefore userspace must tell the kernel the object tiling if it
 47 *   wants to use fences for detiling.
 48 * - On gen3 and gen4 platforms have a swizzling pattern for tiled objects which
 49 *   depends upon the physical page frame number. When swapping such objects the
 50 *   page frame number might change and the kernel must be able to fix this up
 51 *   and hence now the tiling. Note that on a subset of platforms with
 52 *   asymmetric memory channel population the swizzling pattern changes in an
 53 *   unknown way, and for those the kernel simply forbids swapping completely.
 54 *
 55 * Since neither of this applies for new tiling layouts on modern platforms like
 56 * W, Ys and Yf tiling GEM only allows object tiling to be set to X or Y tiled.
 57 * Anything else can be handled in userspace entirely without the kernel's
 58 * invovlement.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 59 */
 60
 61/**
 62 * i915_gem_fence_size - required global GTT size for a fence
 63 * @i915: i915 device
 64 * @size: object size
 65 * @tiling: tiling mode
 66 * @stride: tiling stride
 67 *
 68 * Return the required global GTT size for a fence (view of a tiled object),
 69 * taking into account potential fence register mapping.
 70 */
 71u32 i915_gem_fence_size(struct drm_i915_private *i915,
 72			u32 size, unsigned int tiling, unsigned int stride)
 73{
 74	u32 ggtt_size;
 75
 76	GEM_BUG_ON(!size);
 77
 78	if (tiling == I915_TILING_NONE)
 79		return size;
 80
 81	GEM_BUG_ON(!stride);
 82
 83	if (INTEL_GEN(i915) >= 4) {
 84		stride *= i915_gem_tile_height(tiling);
 85		GEM_BUG_ON(!IS_ALIGNED(stride, I965_FENCE_PAGE));
 86		return roundup(size, stride);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 87	}
 88
 89	/* Previous chips need a power-of-two fence region when tiling */
 90	if (IS_GEN3(i915))
 91		ggtt_size = 1024*1024;
 92	else
 93		ggtt_size = 512*1024;
 94
 95	while (ggtt_size < size)
 96		ggtt_size <<= 1;
 97
 98	return ggtt_size;
 99}
100
101/**
102 * i915_gem_fence_alignment - required global GTT alignment for a fence
103 * @i915: i915 device
104 * @size: object size
105 * @tiling: tiling mode
106 * @stride: tiling stride
107 *
108 * Return the required global GTT alignment for a fence (a view of a tiled
109 * object), taking into account potential fence register mapping.
110 */
111u32 i915_gem_fence_alignment(struct drm_i915_private *i915, u32 size,
112			     unsigned int tiling, unsigned int stride)
113{
114	GEM_BUG_ON(!size);
115
116	/*
117	 * Minimum alignment is 4k (GTT page size), but might be greater
118	 * if a fence register is needed for the object.
119	 */
120	if (tiling == I915_TILING_NONE)
121		return I915_GTT_MIN_ALIGNMENT;
122
123	if (INTEL_GEN(i915) >= 4)
124		return I965_FENCE_PAGE;
125
126	/*
127	 * Previous chips need to be aligned to the size of the smallest
128	 * fence register that can contain the object.
129	 */
130	return i915_gem_fence_size(i915, size, tiling, stride);
131}
132
133/* Check pitch constriants for all chips & tiling formats */
134static bool
135i915_tiling_ok(struct drm_i915_gem_object *obj,
136	       unsigned int tiling, unsigned int stride)
137{
138	struct drm_i915_private *i915 = to_i915(obj->base.dev);
139	unsigned int tile_width;
140
141	/* Linear is always fine */
142	if (tiling == I915_TILING_NONE)
143		return true;
144
145	if (tiling > I915_TILING_LAST)
146		return false;
 
 
 
147
148	/* check maximum stride & object size */
149	/* i965+ stores the end address of the gtt mapping in the fence
150	 * reg, so dont bother to check the size */
151	if (INTEL_GEN(i915) >= 7) {
152		if (stride / 128 > GEN7_FENCE_MAX_PITCH_VAL)
153			return false;
154	} else if (INTEL_GEN(i915) >= 4) {
155		if (stride / 128 > I965_FENCE_MAX_PITCH_VAL)
156			return false;
157	} else {
158		if (stride > 8192)
159			return false;
160
161		if (!is_power_of_2(stride))
162			return false;
 
 
 
 
 
163	}
164
165	if (IS_GEN2(i915) ||
166	    (tiling == I915_TILING_Y && HAS_128_BYTE_Y_TILING(i915)))
167		tile_width = 128;
168	else
169		tile_width = 512;
170
171	if (!stride || !IS_ALIGNED(stride, tile_width))
172		return false;
173
174	return true;
175}
176
177static bool i915_vma_fence_prepare(struct i915_vma *vma,
178				   int tiling_mode, unsigned int stride)
179{
180	struct drm_i915_private *i915 = vma->vm->i915;
181	u32 size, alignment;
182
183	if (!i915_vma_is_map_and_fenceable(vma))
184		return true;
 
185
186	size = i915_gem_fence_size(i915, vma->size, tiling_mode, stride);
187	if (vma->node.size < size)
188		return false;
189
190	alignment = i915_gem_fence_alignment(i915, vma->size, tiling_mode, stride);
191	if (!IS_ALIGNED(vma->node.start, alignment))
192		return false;
193
194	return true;
195}
196
197/* Make the current GTT allocation valid for the change in tiling. */
198static int
199i915_gem_object_fence_prepare(struct drm_i915_gem_object *obj,
200			      int tiling_mode, unsigned int stride)
201{
202	struct i915_vma *vma;
203	int ret;
204
205	if (tiling_mode == I915_TILING_NONE)
206		return 0;
 
 
 
207
208	for_each_ggtt_vma(vma, obj) {
209		if (i915_vma_fence_prepare(vma, tiling_mode, stride))
210			continue;
211
212		ret = i915_vma_unbind(vma);
213		if (ret)
214			return ret;
215	}
216
217	return 0;
218}
219
220int
221i915_gem_object_set_tiling(struct drm_i915_gem_object *obj,
222			   unsigned int tiling, unsigned int stride)
223{
224	struct drm_i915_private *i915 = to_i915(obj->base.dev);
225	struct i915_vma *vma;
226	int err;
227
228	/* Make sure we don't cross-contaminate obj->tiling_and_stride */
229	BUILD_BUG_ON(I915_TILING_LAST & STRIDE_MASK);
230
231	GEM_BUG_ON(!i915_tiling_ok(obj, tiling, stride));
232	GEM_BUG_ON(!stride ^ (tiling == I915_TILING_NONE));
233	lockdep_assert_held(&i915->drm.struct_mutex);
234
235	if ((tiling | stride) == obj->tiling_and_stride)
236		return 0;
237
238	if (i915_gem_object_is_framebuffer(obj))
239		return -EBUSY;
240
241	/* We need to rebind the object if its current allocation
242	 * no longer meets the alignment restrictions for its new
243	 * tiling mode. Otherwise we can just leave it alone, but
244	 * need to ensure that any fence register is updated before
245	 * the next fenced (either through the GTT or by the BLT unit
246	 * on older GPUs) access.
247	 *
248	 * After updating the tiling parameters, we then flag whether
249	 * we need to update an associated fence register. Note this
250	 * has to also include the unfenced register the GPU uses
251	 * whilst executing a fenced command for an untiled object.
252	 */
 
 
 
 
253
254	err = i915_gem_object_fence_prepare(obj, tiling, stride);
255	if (err)
256		return err;
257
258	i915_gem_object_lock(obj);
259	if (i915_gem_object_is_framebuffer(obj)) {
260		i915_gem_object_unlock(obj);
261		return -EBUSY;
262	}
263
264	/* If the memory has unknown (i.e. varying) swizzling, we pin the
265	 * pages to prevent them being swapped out and causing corruption
266	 * due to the change in swizzling.
267	 */
268	mutex_lock(&obj->mm.lock);
269	if (i915_gem_object_has_pages(obj) &&
270	    obj->mm.madv == I915_MADV_WILLNEED &&
271	    i915->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
272		if (tiling == I915_TILING_NONE) {
273			GEM_BUG_ON(!obj->mm.quirked);
274			__i915_gem_object_unpin_pages(obj);
275			obj->mm.quirked = false;
276		}
277		if (!i915_gem_object_is_tiled(obj)) {
278			GEM_BUG_ON(obj->mm.quirked);
279			__i915_gem_object_pin_pages(obj);
280			obj->mm.quirked = true;
281		}
282	}
283	mutex_unlock(&obj->mm.lock);
284
285	for_each_ggtt_vma(vma, obj) {
286		vma->fence_size =
287			i915_gem_fence_size(i915, vma->size, tiling, stride);
288		vma->fence_alignment =
289			i915_gem_fence_alignment(i915,
290						 vma->size, tiling, stride);
291
292		if (vma->fence)
293			vma->fence->dirty = true;
294	}
295
296	obj->tiling_and_stride = tiling | stride;
297	i915_gem_object_unlock(obj);
298
299	/* Force the fence to be reacquired for GTT access */
300	i915_gem_release_mmap(obj);
301
302	/* Try to preallocate memory required to save swizzling on put-pages */
303	if (i915_gem_object_needs_bit17_swizzle(obj)) {
304		if (!obj->bit_17) {
305			obj->bit_17 = kcalloc(BITS_TO_LONGS(obj->base.size >> PAGE_SHIFT),
306					      sizeof(long), GFP_KERNEL);
307		}
308	} else {
309		kfree(obj->bit_17);
310		obj->bit_17 = NULL;
311	}
312
313	return 0;
314}
315
316/**
317 * i915_gem_set_tiling_ioctl - IOCTL handler to set tiling mode
318 * @dev: DRM device
319 * @data: data pointer for the ioctl
320 * @file: DRM file for the ioctl call
321 *
322 * Sets the tiling mode of an object, returning the required swizzling of
323 * bit 6 of addresses in the object.
324 *
325 * Called by the user via ioctl.
326 *
327 * Returns:
328 * Zero on success, negative errno on failure.
329 */
330int
331i915_gem_set_tiling_ioctl(struct drm_device *dev, void *data,
332			  struct drm_file *file)
333{
334	struct drm_i915_gem_set_tiling *args = data;
 
335	struct drm_i915_gem_object *obj;
336	int err;
337
338	obj = i915_gem_object_lookup(file, args->handle);
339	if (!obj)
340		return -ENOENT;
341
342	/*
343	 * The tiling mode of proxy objects is handled by its generator, and
344	 * not allowed to be changed by userspace.
345	 */
346	if (i915_gem_object_is_proxy(obj)) {
347		err = -ENXIO;
348		goto err;
349	}
350
351	if (!i915_tiling_ok(obj, args->tiling_mode, args->stride)) {
352		err = -EINVAL;
353		goto err;
354	}
355
356	if (args->tiling_mode == I915_TILING_NONE) {
357		args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
358		args->stride = 0;
359	} else {
360		if (args->tiling_mode == I915_TILING_X)
361			args->swizzle_mode = to_i915(dev)->mm.bit_6_swizzle_x;
362		else
363			args->swizzle_mode = to_i915(dev)->mm.bit_6_swizzle_y;
364
365		/* Hide bit 17 swizzling from the user.  This prevents old Mesa
366		 * from aborting the application on sw fallbacks to bit 17,
367		 * and we use the pread/pwrite bit17 paths to swizzle for it.
368		 * If there was a user that was relying on the swizzle
369		 * information for drm_intel_bo_map()ed reads/writes this would
370		 * break it, but we don't have any of those.
371		 */
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;
376
377		/* If we can't handle the swizzling, make it untiled. */
378		if (args->swizzle_mode == I915_BIT_6_SWIZZLE_UNKNOWN) {
379			args->tiling_mode = I915_TILING_NONE;
380			args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
381			args->stride = 0;
382		}
383	}
384
385	err = mutex_lock_interruptible(&dev->struct_mutex);
386	if (err)
387		goto err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
388
389	err = i915_gem_object_set_tiling(obj, args->tiling_mode, args->stride);
 
 
 
 
 
 
 
 
 
390	mutex_unlock(&dev->struct_mutex);
391
392	/* We have to maintain this existing ABI... */
393	args->stride = i915_gem_object_get_stride(obj);
394	args->tiling_mode = i915_gem_object_get_tiling(obj);
395
396err:
397	i915_gem_object_put(obj);
398	return err;
399}
400
401/**
402 * i915_gem_get_tiling_ioctl - IOCTL handler to get tiling mode
403 * @dev: DRM device
404 * @data: data pointer for the ioctl
405 * @file: DRM file for the ioctl call
406 *
407 * Returns the current tiling mode and required bit 6 swizzling for the object.
408 *
409 * Called by the user via ioctl.
410 *
411 * Returns:
412 * Zero on success, negative errno on failure.
413 */
414int
415i915_gem_get_tiling_ioctl(struct drm_device *dev, void *data,
416			  struct drm_file *file)
417{
418	struct drm_i915_gem_get_tiling *args = data;
419	struct drm_i915_private *dev_priv = to_i915(dev);
420	struct drm_i915_gem_object *obj;
421	int err = -ENOENT;
422
423	rcu_read_lock();
424	obj = i915_gem_object_lookup_rcu(file, args->handle);
425	if (obj) {
426		args->tiling_mode =
427			READ_ONCE(obj->tiling_and_stride) & TILING_MASK;
428		err = 0;
429	}
430	rcu_read_unlock();
431	if (unlikely(err))
432		return err;
433
434	switch (args->tiling_mode) {
 
 
 
435	case I915_TILING_X:
436		args->swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
437		break;
438	case I915_TILING_Y:
439		args->swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
440		break;
441	default:
442	case I915_TILING_NONE:
443		args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
444		break;
 
 
445	}
446
447	/* Hide bit 17 from the user -- see comment in i915_gem_set_tiling */
448	if (dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES)
449		args->phys_swizzle_mode = I915_BIT_6_SWIZZLE_UNKNOWN;
450	else
451		args->phys_swizzle_mode = args->swizzle_mode;
452	if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_17)
453		args->swizzle_mode = I915_BIT_6_SWIZZLE_9;
454	if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_10_17)
455		args->swizzle_mode = I915_BIT_6_SWIZZLE_9_10;
456
 
 
 
457	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
458}