1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2008 Intel Corporation
5  */
6 
7 #include <linux/string.h>
8 #include <linux/bitops.h>
9 #include <drm/i915_drm.h>
10 
11 #include "i915_drv.h"
12 #include "i915_gem.h"
13 #include "i915_gem_ioctls.h"
14 #include "i915_gem_mman.h"
15 #include "i915_gem_object.h"
16 
17 /**
18  * DOC: buffer object tiling
19  *
20  * i915_gem_set_tiling_ioctl() and i915_gem_get_tiling_ioctl() is the userspace
21  * interface to declare fence register requirements.
22  *
23  * In principle GEM doesn't care at all about the internal data layout of an
24  * object, and hence it also doesn't care about tiling or swizzling. There's two
25  * exceptions:
26  *
27  * - For X and Y tiling the hardware provides detilers for CPU access, so called
28  *   fences. Since there's only a limited amount of them the kernel must manage
29  *   these, and therefore userspace must tell the kernel the object tiling if it
30  *   wants to use fences for detiling.
31  * - On gen3 and gen4 platforms have a swizzling pattern for tiled objects which
32  *   depends upon the physical page frame number. When swapping such objects the
33  *   page frame number might change and the kernel must be able to fix this up
34  *   and hence now the tiling. Note that on a subset of platforms with
35  *   asymmetric memory channel population the swizzling pattern changes in an
36  *   unknown way, and for those the kernel simply forbids swapping completely.
37  *
38  * Since neither of this applies for new tiling layouts on modern platforms like
39  * W, Ys and Yf tiling GEM only allows object tiling to be set to X or Y tiled.
40  * Anything else can be handled in userspace entirely without the kernel's
41  * invovlement.
42  */
43 
44 /**
45  * i915_gem_fence_size - required global GTT size for a fence
46  * @i915: i915 device
47  * @size: object size
48  * @tiling: tiling mode
49  * @stride: tiling stride
50  *
51  * Return the required global GTT size for a fence (view of a tiled object),
52  * taking into account potential fence register mapping.
53  */
54 u32 i915_gem_fence_size(struct drm_i915_private *i915,
55 			u32 size, unsigned int tiling, unsigned int stride)
56 {
57 	u32 ggtt_size;
58 
59 	GEM_BUG_ON(!size);
60 
61 	if (tiling == I915_TILING_NONE)
62 		return size;
63 
64 	GEM_BUG_ON(!stride);
65 
66 	if (INTEL_GEN(i915) >= 4) {
67 		stride *= i915_gem_tile_height(tiling);
68 		GEM_BUG_ON(!IS_ALIGNED(stride, I965_FENCE_PAGE));
69 		return roundup(size, stride);
70 	}
71 
72 	/* Previous chips need a power-of-two fence region when tiling */
73 	if (IS_GEN(i915, 3))
74 		ggtt_size = 1024*1024;
75 	else
76 		ggtt_size = 512*1024;
77 
78 	while (ggtt_size < size)
79 		ggtt_size <<= 1;
80 
81 	return ggtt_size;
82 }
83 
84 /**
85  * i915_gem_fence_alignment - required global GTT alignment for a fence
86  * @i915: i915 device
87  * @size: object size
88  * @tiling: tiling mode
89  * @stride: tiling stride
90  *
91  * Return the required global GTT alignment for a fence (a view of a tiled
92  * object), taking into account potential fence register mapping.
93  */
94 u32 i915_gem_fence_alignment(struct drm_i915_private *i915, u32 size,
95 			     unsigned int tiling, unsigned int stride)
96 {
97 	GEM_BUG_ON(!size);
98 
99 	/*
100 	 * Minimum alignment is 4k (GTT page size), but might be greater
101 	 * if a fence register is needed for the object.
102 	 */
103 	if (tiling == I915_TILING_NONE)
104 		return I915_GTT_MIN_ALIGNMENT;
105 
106 	if (INTEL_GEN(i915) >= 4)
107 		return I965_FENCE_PAGE;
108 
109 	/*
110 	 * Previous chips need to be aligned to the size of the smallest
111 	 * fence register that can contain the object.
112 	 */
113 	return i915_gem_fence_size(i915, size, tiling, stride);
114 }
115 
116 /* Check pitch constriants for all chips & tiling formats */
117 static bool
118 i915_tiling_ok(struct drm_i915_gem_object *obj,
119 	       unsigned int tiling, unsigned int stride)
120 {
121 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
122 	unsigned int tile_width;
123 
124 	/* Linear is always fine */
125 	if (tiling == I915_TILING_NONE)
126 		return true;
127 
128 	if (tiling > I915_TILING_LAST)
129 		return false;
130 
131 	/* check maximum stride & object size */
132 	/* i965+ stores the end address of the gtt mapping in the fence
133 	 * reg, so dont bother to check the size */
134 	if (INTEL_GEN(i915) >= 7) {
135 		if (stride / 128 > GEN7_FENCE_MAX_PITCH_VAL)
136 			return false;
137 	} else if (INTEL_GEN(i915) >= 4) {
138 		if (stride / 128 > I965_FENCE_MAX_PITCH_VAL)
139 			return false;
140 	} else {
141 		if (stride > 8192)
142 			return false;
143 
144 		if (!is_power_of_2(stride))
145 			return false;
146 	}
147 
148 	if (IS_GEN(i915, 2) ||
149 	    (tiling == I915_TILING_Y && HAS_128_BYTE_Y_TILING(i915)))
150 		tile_width = 128;
151 	else
152 		tile_width = 512;
153 
154 	if (!stride || !IS_ALIGNED(stride, tile_width))
155 		return false;
156 
157 	return true;
158 }
159 
160 static bool i915_vma_fence_prepare(struct i915_vma *vma,
161 				   int tiling_mode, unsigned int stride)
162 {
163 	struct drm_i915_private *i915 = vma->vm->i915;
164 	u32 size, alignment;
165 
166 	if (!i915_vma_is_map_and_fenceable(vma))
167 		return true;
168 
169 	size = i915_gem_fence_size(i915, vma->size, tiling_mode, stride);
170 	if (vma->node.size < size)
171 		return false;
172 
173 	alignment = i915_gem_fence_alignment(i915, vma->size, tiling_mode, stride);
174 	if (!IS_ALIGNED(vma->node.start, alignment))
175 		return false;
176 
177 	return true;
178 }
179 
180 /* Make the current GTT allocation valid for the change in tiling. */
181 static int
182 i915_gem_object_fence_prepare(struct drm_i915_gem_object *obj,
183 			      int tiling_mode, unsigned int stride)
184 {
185 	struct i915_ggtt *ggtt = &to_i915(obj->base.dev)->ggtt;
186 	struct i915_vma *vma;
187 	int ret = 0;
188 
189 	if (tiling_mode == I915_TILING_NONE)
190 		return 0;
191 
192 	mutex_lock(&ggtt->vm.mutex);
193 	for_each_ggtt_vma(vma, obj) {
194 		if (i915_vma_fence_prepare(vma, tiling_mode, stride))
195 			continue;
196 
197 		ret = __i915_vma_unbind(vma);
198 		if (ret)
199 			break;
200 	}
201 	mutex_unlock(&ggtt->vm.mutex);
202 
203 	return ret;
204 }
205 
206 int
207 i915_gem_object_set_tiling(struct drm_i915_gem_object *obj,
208 			   unsigned int tiling, unsigned int stride)
209 {
210 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
211 	struct i915_vma *vma;
212 	int err;
213 
214 	/* Make sure we don't cross-contaminate obj->tiling_and_stride */
215 	BUILD_BUG_ON(I915_TILING_LAST & STRIDE_MASK);
216 
217 	GEM_BUG_ON(!i915_tiling_ok(obj, tiling, stride));
218 	GEM_BUG_ON(!stride ^ (tiling == I915_TILING_NONE));
219 
220 	if ((tiling | stride) == obj->tiling_and_stride)
221 		return 0;
222 
223 	if (i915_gem_object_is_framebuffer(obj))
224 		return -EBUSY;
225 
226 	/* We need to rebind the object if its current allocation
227 	 * no longer meets the alignment restrictions for its new
228 	 * tiling mode. Otherwise we can just leave it alone, but
229 	 * need to ensure that any fence register is updated before
230 	 * the next fenced (either through the GTT or by the BLT unit
231 	 * on older GPUs) access.
232 	 *
233 	 * After updating the tiling parameters, we then flag whether
234 	 * we need to update an associated fence register. Note this
235 	 * has to also include the unfenced register the GPU uses
236 	 * whilst executing a fenced command for an untiled object.
237 	 */
238 
239 	i915_gem_object_lock(obj);
240 	if (i915_gem_object_is_framebuffer(obj)) {
241 		i915_gem_object_unlock(obj);
242 		return -EBUSY;
243 	}
244 
245 	err = i915_gem_object_fence_prepare(obj, tiling, stride);
246 	if (err) {
247 		i915_gem_object_unlock(obj);
248 		return err;
249 	}
250 
251 	/* If the memory has unknown (i.e. varying) swizzling, we pin the
252 	 * pages to prevent them being swapped out and causing corruption
253 	 * due to the change in swizzling.
254 	 */
255 	mutex_lock(&obj->mm.lock);
256 	if (i915_gem_object_has_pages(obj) &&
257 	    obj->mm.madv == I915_MADV_WILLNEED &&
258 	    i915->quirks & QUIRK_PIN_SWIZZLED_PAGES) {
259 		if (tiling == I915_TILING_NONE) {
260 			GEM_BUG_ON(!obj->mm.quirked);
261 			__i915_gem_object_unpin_pages(obj);
262 			obj->mm.quirked = false;
263 		}
264 		if (!i915_gem_object_is_tiled(obj)) {
265 			GEM_BUG_ON(obj->mm.quirked);
266 			__i915_gem_object_pin_pages(obj);
267 			obj->mm.quirked = true;
268 		}
269 	}
270 	mutex_unlock(&obj->mm.lock);
271 
272 	for_each_ggtt_vma(vma, obj) {
273 		vma->fence_size =
274 			i915_gem_fence_size(i915, vma->size, tiling, stride);
275 		vma->fence_alignment =
276 			i915_gem_fence_alignment(i915,
277 						 vma->size, tiling, stride);
278 
279 		if (vma->fence)
280 			vma->fence->dirty = true;
281 	}
282 
283 	obj->tiling_and_stride = tiling | stride;
284 	i915_gem_object_unlock(obj);
285 
286 	/* Force the fence to be reacquired for GTT access */
287 	i915_gem_object_release_mmap(obj);
288 
289 	/* Try to preallocate memory required to save swizzling on put-pages */
290 	if (i915_gem_object_needs_bit17_swizzle(obj)) {
291 		if (!obj->bit_17) {
292 			obj->bit_17 = bitmap_zalloc(obj->base.size >> PAGE_SHIFT,
293 						    GFP_KERNEL);
294 		}
295 	} else {
296 		bitmap_free(obj->bit_17);
297 		obj->bit_17 = NULL;
298 	}
299 
300 	return 0;
301 }
302 
303 /**
304  * i915_gem_set_tiling_ioctl - IOCTL handler to set tiling mode
305  * @dev: DRM device
306  * @data: data pointer for the ioctl
307  * @file: DRM file for the ioctl call
308  *
309  * Sets the tiling mode of an object, returning the required swizzling of
310  * bit 6 of addresses in the object.
311  *
312  * Called by the user via ioctl.
313  *
314  * Returns:
315  * Zero on success, negative errno on failure.
316  */
317 int
318 i915_gem_set_tiling_ioctl(struct drm_device *dev, void *data,
319 			  struct drm_file *file)
320 {
321 	struct drm_i915_private *dev_priv = to_i915(dev);
322 	struct drm_i915_gem_set_tiling *args = data;
323 	struct drm_i915_gem_object *obj;
324 	int err;
325 
326 	if (!dev_priv->ggtt.num_fences)
327 		return -EOPNOTSUPP;
328 
329 	obj = i915_gem_object_lookup(file, args->handle);
330 	if (!obj)
331 		return -ENOENT;
332 
333 	/*
334 	 * The tiling mode of proxy objects is handled by its generator, and
335 	 * not allowed to be changed by userspace.
336 	 */
337 	if (i915_gem_object_is_proxy(obj)) {
338 		err = -ENXIO;
339 		goto err;
340 	}
341 
342 	if (!i915_tiling_ok(obj, args->tiling_mode, args->stride)) {
343 		err = -EINVAL;
344 		goto err;
345 	}
346 
347 	if (args->tiling_mode == I915_TILING_NONE) {
348 		args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
349 		args->stride = 0;
350 	} else {
351 		if (args->tiling_mode == I915_TILING_X)
352 			args->swizzle_mode = to_i915(dev)->ggtt.bit_6_swizzle_x;
353 		else
354 			args->swizzle_mode = to_i915(dev)->ggtt.bit_6_swizzle_y;
355 
356 		/* Hide bit 17 swizzling from the user.  This prevents old Mesa
357 		 * from aborting the application on sw fallbacks to bit 17,
358 		 * and we use the pread/pwrite bit17 paths to swizzle for it.
359 		 * If there was a user that was relying on the swizzle
360 		 * information for drm_intel_bo_map()ed reads/writes this would
361 		 * break it, but we don't have any of those.
362 		 */
363 		if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_17)
364 			args->swizzle_mode = I915_BIT_6_SWIZZLE_9;
365 		if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_10_17)
366 			args->swizzle_mode = I915_BIT_6_SWIZZLE_9_10;
367 
368 		/* If we can't handle the swizzling, make it untiled. */
369 		if (args->swizzle_mode == I915_BIT_6_SWIZZLE_UNKNOWN) {
370 			args->tiling_mode = I915_TILING_NONE;
371 			args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
372 			args->stride = 0;
373 		}
374 	}
375 
376 	err = i915_gem_object_set_tiling(obj, args->tiling_mode, args->stride);
377 
378 	/* We have to maintain this existing ABI... */
379 	args->stride = i915_gem_object_get_stride(obj);
380 	args->tiling_mode = i915_gem_object_get_tiling(obj);
381 
382 err:
383 	i915_gem_object_put(obj);
384 	return err;
385 }
386 
387 /**
388  * i915_gem_get_tiling_ioctl - IOCTL handler to get tiling mode
389  * @dev: DRM device
390  * @data: data pointer for the ioctl
391  * @file: DRM file for the ioctl call
392  *
393  * Returns the current tiling mode and required bit 6 swizzling for the object.
394  *
395  * Called by the user via ioctl.
396  *
397  * Returns:
398  * Zero on success, negative errno on failure.
399  */
400 int
401 i915_gem_get_tiling_ioctl(struct drm_device *dev, void *data,
402 			  struct drm_file *file)
403 {
404 	struct drm_i915_gem_get_tiling *args = data;
405 	struct drm_i915_private *dev_priv = to_i915(dev);
406 	struct drm_i915_gem_object *obj;
407 	int err = -ENOENT;
408 
409 	if (!dev_priv->ggtt.num_fences)
410 		return -EOPNOTSUPP;
411 
412 	rcu_read_lock();
413 	obj = i915_gem_object_lookup_rcu(file, args->handle);
414 	if (obj) {
415 		args->tiling_mode =
416 			READ_ONCE(obj->tiling_and_stride) & TILING_MASK;
417 		err = 0;
418 	}
419 	rcu_read_unlock();
420 	if (unlikely(err))
421 		return err;
422 
423 	switch (args->tiling_mode) {
424 	case I915_TILING_X:
425 		args->swizzle_mode = dev_priv->ggtt.bit_6_swizzle_x;
426 		break;
427 	case I915_TILING_Y:
428 		args->swizzle_mode = dev_priv->ggtt.bit_6_swizzle_y;
429 		break;
430 	default:
431 	case I915_TILING_NONE:
432 		args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
433 		break;
434 	}
435 
436 	/* Hide bit 17 from the user -- see comment in i915_gem_set_tiling */
437 	if (dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES)
438 		args->phys_swizzle_mode = I915_BIT_6_SWIZZLE_UNKNOWN;
439 	else
440 		args->phys_swizzle_mode = args->swizzle_mode;
441 	if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_17)
442 		args->swizzle_mode = I915_BIT_6_SWIZZLE_9;
443 	if (args->swizzle_mode == I915_BIT_6_SWIZZLE_9_10_17)
444 		args->swizzle_mode = I915_BIT_6_SWIZZLE_9_10;
445 
446 	return 0;
447 }
448