1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2008-2015 Intel Corporation
4  */
5 
6 #include "i915_drv.h"
7 #include "i915_scatterlist.h"
8 #include "i915_pvinfo.h"
9 #include "i915_vgpu.h"
10 
11 /**
12  * DOC: fence register handling
13  *
14  * Important to avoid confusions: "fences" in the i915 driver are not execution
15  * fences used to track command completion but hardware detiler objects which
16  * wrap a given range of the global GTT. Each platform has only a fairly limited
17  * set of these objects.
18  *
19  * Fences are used to detile GTT memory mappings. They're also connected to the
20  * hardware frontbuffer render tracking and hence interact with frontbuffer
21  * compression. Furthermore on older platforms fences are required for tiled
22  * objects used by the display engine. They can also be used by the render
23  * engine - they're required for blitter commands and are optional for render
24  * commands. But on gen4+ both display (with the exception of fbc) and rendering
25  * have their own tiling state bits and don't need fences.
26  *
27  * Also note that fences only support X and Y tiling and hence can't be used for
28  * the fancier new tiling formats like W, Ys and Yf.
29  *
30  * Finally note that because fences are such a restricted resource they're
31  * dynamically associated with objects. Furthermore fence state is committed to
32  * the hardware lazily to avoid unnecessary stalls on gen2/3. Therefore code must
33  * explicitly call i915_gem_object_get_fence() to synchronize fencing status
34  * for cpu access. Also note that some code wants an unfenced view, for those
35  * cases the fence can be removed forcefully with i915_gem_object_put_fence().
36  *
37  * Internally these functions will synchronize with userspace access by removing
38  * CPU ptes into GTT mmaps (not the GTT ptes themselves) as needed.
39  */
40 
41 #define pipelined 0
42 
43 static struct drm_i915_private *fence_to_i915(struct i915_fence_reg *fence)
44 {
45 	return fence->ggtt->vm.i915;
46 }
47 
48 static struct intel_uncore *fence_to_uncore(struct i915_fence_reg *fence)
49 {
50 	return fence->ggtt->vm.gt->uncore;
51 }
52 
53 static void i965_write_fence_reg(struct i915_fence_reg *fence)
54 {
55 	i915_reg_t fence_reg_lo, fence_reg_hi;
56 	int fence_pitch_shift;
57 	u64 val;
58 
59 	if (INTEL_GEN(fence_to_i915(fence)) >= 6) {
60 		fence_reg_lo = FENCE_REG_GEN6_LO(fence->id);
61 		fence_reg_hi = FENCE_REG_GEN6_HI(fence->id);
62 		fence_pitch_shift = GEN6_FENCE_PITCH_SHIFT;
63 
64 	} else {
65 		fence_reg_lo = FENCE_REG_965_LO(fence->id);
66 		fence_reg_hi = FENCE_REG_965_HI(fence->id);
67 		fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
68 	}
69 
70 	val = 0;
71 	if (fence->tiling) {
72 		unsigned int stride = fence->stride;
73 
74 		GEM_BUG_ON(!IS_ALIGNED(stride, 128));
75 
76 		val = fence->start + fence->size - I965_FENCE_PAGE;
77 		val <<= 32;
78 		val |= fence->start;
79 		val |= (u64)((stride / 128) - 1) << fence_pitch_shift;
80 		if (fence->tiling == I915_TILING_Y)
81 			val |= BIT(I965_FENCE_TILING_Y_SHIFT);
82 		val |= I965_FENCE_REG_VALID;
83 	}
84 
85 	if (!pipelined) {
86 		struct intel_uncore *uncore = fence_to_uncore(fence);
87 
88 		/*
89 		 * To w/a incoherency with non-atomic 64-bit register updates,
90 		 * we split the 64-bit update into two 32-bit writes. In order
91 		 * for a partial fence not to be evaluated between writes, we
92 		 * precede the update with write to turn off the fence register,
93 		 * and only enable the fence as the last step.
94 		 *
95 		 * For extra levels of paranoia, we make sure each step lands
96 		 * before applying the next step.
97 		 */
98 		intel_uncore_write_fw(uncore, fence_reg_lo, 0);
99 		intel_uncore_posting_read_fw(uncore, fence_reg_lo);
100 
101 		intel_uncore_write_fw(uncore, fence_reg_hi, upper_32_bits(val));
102 		intel_uncore_write_fw(uncore, fence_reg_lo, lower_32_bits(val));
103 		intel_uncore_posting_read_fw(uncore, fence_reg_lo);
104 	}
105 }
106 
107 static void i915_write_fence_reg(struct i915_fence_reg *fence)
108 {
109 	u32 val;
110 
111 	val = 0;
112 	if (fence->tiling) {
113 		unsigned int stride = fence->stride;
114 		unsigned int tiling = fence->tiling;
115 		bool is_y_tiled = tiling == I915_TILING_Y;
116 
117 		if (is_y_tiled && HAS_128_BYTE_Y_TILING(fence_to_i915(fence)))
118 			stride /= 128;
119 		else
120 			stride /= 512;
121 		GEM_BUG_ON(!is_power_of_2(stride));
122 
123 		val = fence->start;
124 		if (is_y_tiled)
125 			val |= BIT(I830_FENCE_TILING_Y_SHIFT);
126 		val |= I915_FENCE_SIZE_BITS(fence->size);
127 		val |= ilog2(stride) << I830_FENCE_PITCH_SHIFT;
128 
129 		val |= I830_FENCE_REG_VALID;
130 	}
131 
132 	if (!pipelined) {
133 		struct intel_uncore *uncore = fence_to_uncore(fence);
134 		i915_reg_t reg = FENCE_REG(fence->id);
135 
136 		intel_uncore_write_fw(uncore, reg, val);
137 		intel_uncore_posting_read_fw(uncore, reg);
138 	}
139 }
140 
141 static void i830_write_fence_reg(struct i915_fence_reg *fence)
142 {
143 	u32 val;
144 
145 	val = 0;
146 	if (fence->tiling) {
147 		unsigned int stride = fence->stride;
148 
149 		val = fence->start;
150 		if (fence->tiling == I915_TILING_Y)
151 			val |= BIT(I830_FENCE_TILING_Y_SHIFT);
152 		val |= I830_FENCE_SIZE_BITS(fence->size);
153 		val |= ilog2(stride / 128) << I830_FENCE_PITCH_SHIFT;
154 		val |= I830_FENCE_REG_VALID;
155 	}
156 
157 	if (!pipelined) {
158 		struct intel_uncore *uncore = fence_to_uncore(fence);
159 		i915_reg_t reg = FENCE_REG(fence->id);
160 
161 		intel_uncore_write_fw(uncore, reg, val);
162 		intel_uncore_posting_read_fw(uncore, reg);
163 	}
164 }
165 
166 static void fence_write(struct i915_fence_reg *fence)
167 {
168 	struct drm_i915_private *i915 = fence_to_i915(fence);
169 
170 	/*
171 	 * Previous access through the fence register is marshalled by
172 	 * the mb() inside the fault handlers (i915_gem_release_mmaps)
173 	 * and explicitly managed for internal users.
174 	 */
175 
176 	if (IS_GEN(i915, 2))
177 		i830_write_fence_reg(fence);
178 	else if (IS_GEN(i915, 3))
179 		i915_write_fence_reg(fence);
180 	else
181 		i965_write_fence_reg(fence);
182 
183 	/*
184 	 * Access through the fenced region afterwards is
185 	 * ordered by the posting reads whilst writing the registers.
186 	 */
187 }
188 
189 static bool gpu_uses_fence_registers(struct i915_fence_reg *fence)
190 {
191 	return INTEL_GEN(fence_to_i915(fence)) < 4;
192 }
193 
194 static int fence_update(struct i915_fence_reg *fence,
195 			struct i915_vma *vma)
196 {
197 	struct i915_ggtt *ggtt = fence->ggtt;
198 	struct intel_uncore *uncore = fence_to_uncore(fence);
199 	intel_wakeref_t wakeref;
200 	struct i915_vma *old;
201 	int ret;
202 
203 	fence->tiling = 0;
204 	if (vma) {
205 		GEM_BUG_ON(!i915_gem_object_get_stride(vma->obj) ||
206 			   !i915_gem_object_get_tiling(vma->obj));
207 
208 		if (!i915_vma_is_map_and_fenceable(vma))
209 			return -EINVAL;
210 
211 		if (gpu_uses_fence_registers(fence)) {
212 			/* implicit 'unfenced' GPU blits */
213 			ret = i915_vma_sync(vma);
214 			if (ret)
215 				return ret;
216 		}
217 
218 		fence->start = vma->node.start;
219 		fence->size = vma->fence_size;
220 		fence->stride = i915_gem_object_get_stride(vma->obj);
221 		fence->tiling = i915_gem_object_get_tiling(vma->obj);
222 	}
223 	WRITE_ONCE(fence->dirty, false);
224 
225 	old = xchg(&fence->vma, NULL);
226 	if (old) {
227 		/* XXX Ideally we would move the waiting to outside the mutex */
228 		ret = i915_active_wait(&fence->active);
229 		if (ret) {
230 			fence->vma = old;
231 			return ret;
232 		}
233 
234 		i915_vma_flush_writes(old);
235 
236 		/*
237 		 * Ensure that all userspace CPU access is completed before
238 		 * stealing the fence.
239 		 */
240 		if (old != vma) {
241 			GEM_BUG_ON(old->fence != fence);
242 			i915_vma_revoke_mmap(old);
243 			old->fence = NULL;
244 		}
245 
246 		list_move(&fence->link, &ggtt->fence_list);
247 	}
248 
249 	/*
250 	 * We only need to update the register itself if the device is awake.
251 	 * If the device is currently powered down, we will defer the write
252 	 * to the runtime resume, see intel_ggtt_restore_fences().
253 	 *
254 	 * This only works for removing the fence register, on acquisition
255 	 * the caller must hold the rpm wakeref. The fence register must
256 	 * be cleared before we can use any other fences to ensure that
257 	 * the new fences do not overlap the elided clears, confusing HW.
258 	 */
259 	wakeref = intel_runtime_pm_get_if_in_use(uncore->rpm);
260 	if (!wakeref) {
261 		GEM_BUG_ON(vma);
262 		return 0;
263 	}
264 
265 	WRITE_ONCE(fence->vma, vma);
266 	fence_write(fence);
267 
268 	if (vma) {
269 		vma->fence = fence;
270 		list_move_tail(&fence->link, &ggtt->fence_list);
271 	}
272 
273 	intel_runtime_pm_put(uncore->rpm, wakeref);
274 	return 0;
275 }
276 
277 /**
278  * i915_vma_revoke_fence - force-remove fence for a VMA
279  * @vma: vma to map linearly (not through a fence reg)
280  *
281  * This function force-removes any fence from the given object, which is useful
282  * if the kernel wants to do untiled GTT access.
283  */
284 void i915_vma_revoke_fence(struct i915_vma *vma)
285 {
286 	struct i915_fence_reg *fence = vma->fence;
287 	intel_wakeref_t wakeref;
288 
289 	lockdep_assert_held(&vma->vm->mutex);
290 	if (!fence)
291 		return;
292 
293 	GEM_BUG_ON(fence->vma != vma);
294 	GEM_BUG_ON(!i915_active_is_idle(&fence->active));
295 	GEM_BUG_ON(atomic_read(&fence->pin_count));
296 
297 	fence->tiling = 0;
298 	WRITE_ONCE(fence->vma, NULL);
299 	vma->fence = NULL;
300 
301 	/*
302 	 * Skip the write to HW if and only if the device is currently
303 	 * suspended.
304 	 *
305 	 * If the driver does not currently hold a wakeref (if_in_use == 0),
306 	 * the device may currently be runtime suspended, or it may be woken
307 	 * up before the suspend takes place. If the device is not suspended
308 	 * (powered down) and we skip clearing the fence register, the HW is
309 	 * left in an undefined state where we may end up with multiple
310 	 * registers overlapping.
311 	 */
312 	with_intel_runtime_pm_if_active(fence_to_uncore(fence)->rpm, wakeref)
313 		fence_write(fence);
314 }
315 
316 static bool fence_is_active(const struct i915_fence_reg *fence)
317 {
318 	return fence->vma && i915_vma_is_active(fence->vma);
319 }
320 
321 static struct i915_fence_reg *fence_find(struct i915_ggtt *ggtt)
322 {
323 	struct i915_fence_reg *active = NULL;
324 	struct i915_fence_reg *fence, *fn;
325 
326 	list_for_each_entry_safe(fence, fn, &ggtt->fence_list, link) {
327 		GEM_BUG_ON(fence->vma && fence->vma->fence != fence);
328 
329 		if (fence == active) /* now seen this fence twice */
330 			active = ERR_PTR(-EAGAIN);
331 
332 		/* Prefer idle fences so we do not have to wait on the GPU */
333 		if (active != ERR_PTR(-EAGAIN) && fence_is_active(fence)) {
334 			if (!active)
335 				active = fence;
336 
337 			list_move_tail(&fence->link, &ggtt->fence_list);
338 			continue;
339 		}
340 
341 		if (atomic_read(&fence->pin_count))
342 			continue;
343 
344 		return fence;
345 	}
346 
347 	/* Wait for completion of pending flips which consume fences */
348 	if (intel_has_pending_fb_unpin(ggtt->vm.i915))
349 		return ERR_PTR(-EAGAIN);
350 
351 	return ERR_PTR(-EDEADLK);
352 }
353 
354 int __i915_vma_pin_fence(struct i915_vma *vma)
355 {
356 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vma->vm);
357 	struct i915_fence_reg *fence;
358 	struct i915_vma *set = i915_gem_object_is_tiled(vma->obj) ? vma : NULL;
359 	int err;
360 
361 	lockdep_assert_held(&vma->vm->mutex);
362 
363 	/* Just update our place in the LRU if our fence is getting reused. */
364 	if (vma->fence) {
365 		fence = vma->fence;
366 		GEM_BUG_ON(fence->vma != vma);
367 		atomic_inc(&fence->pin_count);
368 		if (!fence->dirty) {
369 			list_move_tail(&fence->link, &ggtt->fence_list);
370 			return 0;
371 		}
372 	} else if (set) {
373 		fence = fence_find(ggtt);
374 		if (IS_ERR(fence))
375 			return PTR_ERR(fence);
376 
377 		GEM_BUG_ON(atomic_read(&fence->pin_count));
378 		atomic_inc(&fence->pin_count);
379 	} else {
380 		return 0;
381 	}
382 
383 	err = fence_update(fence, set);
384 	if (err)
385 		goto out_unpin;
386 
387 	GEM_BUG_ON(fence->vma != set);
388 	GEM_BUG_ON(vma->fence != (set ? fence : NULL));
389 
390 	if (set)
391 		return 0;
392 
393 out_unpin:
394 	atomic_dec(&fence->pin_count);
395 	return err;
396 }
397 
398 /**
399  * i915_vma_pin_fence - set up fencing for a vma
400  * @vma: vma to map through a fence reg
401  *
402  * When mapping objects through the GTT, userspace wants to be able to write
403  * to them without having to worry about swizzling if the object is tiled.
404  * This function walks the fence regs looking for a free one for @obj,
405  * stealing one if it can't find any.
406  *
407  * It then sets up the reg based on the object's properties: address, pitch
408  * and tiling format.
409  *
410  * For an untiled surface, this removes any existing fence.
411  *
412  * Returns:
413  *
414  * 0 on success, negative error code on failure.
415  */
416 int i915_vma_pin_fence(struct i915_vma *vma)
417 {
418 	int err;
419 
420 	if (!vma->fence && !i915_gem_object_is_tiled(vma->obj))
421 		return 0;
422 
423 	/*
424 	 * Note that we revoke fences on runtime suspend. Therefore the user
425 	 * must keep the device awake whilst using the fence.
426 	 */
427 	assert_rpm_wakelock_held(vma->vm->gt->uncore->rpm);
428 	GEM_BUG_ON(!i915_vma_is_pinned(vma));
429 	GEM_BUG_ON(!i915_vma_is_ggtt(vma));
430 
431 	err = mutex_lock_interruptible(&vma->vm->mutex);
432 	if (err)
433 		return err;
434 
435 	err = __i915_vma_pin_fence(vma);
436 	mutex_unlock(&vma->vm->mutex);
437 
438 	return err;
439 }
440 
441 /**
442  * i915_reserve_fence - Reserve a fence for vGPU
443  * @ggtt: Global GTT
444  *
445  * This function walks the fence regs looking for a free one and remove
446  * it from the fence_list. It is used to reserve fence for vGPU to use.
447  */
448 struct i915_fence_reg *i915_reserve_fence(struct i915_ggtt *ggtt)
449 {
450 	struct i915_fence_reg *fence;
451 	int count;
452 	int ret;
453 
454 	lockdep_assert_held(&ggtt->vm.mutex);
455 
456 	/* Keep at least one fence available for the display engine. */
457 	count = 0;
458 	list_for_each_entry(fence, &ggtt->fence_list, link)
459 		count += !atomic_read(&fence->pin_count);
460 	if (count <= 1)
461 		return ERR_PTR(-ENOSPC);
462 
463 	fence = fence_find(ggtt);
464 	if (IS_ERR(fence))
465 		return fence;
466 
467 	if (fence->vma) {
468 		/* Force-remove fence from VMA */
469 		ret = fence_update(fence, NULL);
470 		if (ret)
471 			return ERR_PTR(ret);
472 	}
473 
474 	list_del(&fence->link);
475 
476 	return fence;
477 }
478 
479 /**
480  * i915_unreserve_fence - Reclaim a reserved fence
481  * @fence: the fence reg
482  *
483  * This function add a reserved fence register from vGPU to the fence_list.
484  */
485 void i915_unreserve_fence(struct i915_fence_reg *fence)
486 {
487 	struct i915_ggtt *ggtt = fence->ggtt;
488 
489 	lockdep_assert_held(&ggtt->vm.mutex);
490 
491 	list_add(&fence->link, &ggtt->fence_list);
492 }
493 
494 /**
495  * intel_ggtt_restore_fences - restore fence state
496  * @ggtt: Global GTT
497  *
498  * Restore the hw fence state to match the software tracking again, to be called
499  * after a gpu reset and on resume. Note that on runtime suspend we only cancel
500  * the fences, to be reacquired by the user later.
501  */
502 void intel_ggtt_restore_fences(struct i915_ggtt *ggtt)
503 {
504 	int i;
505 
506 	for (i = 0; i < ggtt->num_fences; i++)
507 		fence_write(&ggtt->fence_regs[i]);
508 }
509 
510 /**
511  * DOC: tiling swizzling details
512  *
513  * The idea behind tiling is to increase cache hit rates by rearranging
514  * pixel data so that a group of pixel accesses are in the same cacheline.
515  * Performance improvement from doing this on the back/depth buffer are on
516  * the order of 30%.
517  *
518  * Intel architectures make this somewhat more complicated, though, by
519  * adjustments made to addressing of data when the memory is in interleaved
520  * mode (matched pairs of DIMMS) to improve memory bandwidth.
521  * For interleaved memory, the CPU sends every sequential 64 bytes
522  * to an alternate memory channel so it can get the bandwidth from both.
523  *
524  * The GPU also rearranges its accesses for increased bandwidth to interleaved
525  * memory, and it matches what the CPU does for non-tiled.  However, when tiled
526  * it does it a little differently, since one walks addresses not just in the
527  * X direction but also Y.  So, along with alternating channels when bit
528  * 6 of the address flips, it also alternates when other bits flip --  Bits 9
529  * (every 512 bytes, an X tile scanline) and 10 (every two X tile scanlines)
530  * are common to both the 915 and 965-class hardware.
531  *
532  * The CPU also sometimes XORs in higher bits as well, to improve
533  * bandwidth doing strided access like we do so frequently in graphics.  This
534  * is called "Channel XOR Randomization" in the MCH documentation.  The result
535  * is that the CPU is XORing in either bit 11 or bit 17 to bit 6 of its address
536  * decode.
537  *
538  * All of this bit 6 XORing has an effect on our memory management,
539  * as we need to make sure that the 3d driver can correctly address object
540  * contents.
541  *
542  * If we don't have interleaved memory, all tiling is safe and no swizzling is
543  * required.
544  *
545  * When bit 17 is XORed in, we simply refuse to tile at all.  Bit
546  * 17 is not just a page offset, so as we page an object out and back in,
547  * individual pages in it will have different bit 17 addresses, resulting in
548  * each 64 bytes being swapped with its neighbor!
549  *
550  * Otherwise, if interleaved, we have to tell the 3d driver what the address
551  * swizzling it needs to do is, since it's writing with the CPU to the pages
552  * (bit 6 and potentially bit 11 XORed in), and the GPU is reading from the
553  * pages (bit 6, 9, and 10 XORed in), resulting in a cumulative bit swizzling
554  * required by the CPU of XORing in bit 6, 9, 10, and potentially 11, in order
555  * to match what the GPU expects.
556  */
557 
558 /**
559  * detect_bit_6_swizzle - detect bit 6 swizzling pattern
560  * @ggtt: Global GGTT
561  *
562  * Detects bit 6 swizzling of address lookup between IGD access and CPU
563  * access through main memory.
564  */
565 static void detect_bit_6_swizzle(struct i915_ggtt *ggtt)
566 {
567 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
568 	struct drm_i915_private *i915 = ggtt->vm.i915;
569 	u32 swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
570 	u32 swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
571 
572 	if (INTEL_GEN(i915) >= 8 || IS_VALLEYVIEW(i915)) {
573 		/*
574 		 * On BDW+, swizzling is not used. We leave the CPU memory
575 		 * controller in charge of optimizing memory accesses without
576 		 * the extra address manipulation GPU side.
577 		 *
578 		 * VLV and CHV don't have GPU swizzling.
579 		 */
580 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
581 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
582 	} else if (INTEL_GEN(i915) >= 6) {
583 		if (i915->preserve_bios_swizzle) {
584 			if (intel_uncore_read(uncore, DISP_ARB_CTL) &
585 			    DISP_TILE_SURFACE_SWIZZLING) {
586 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
587 				swizzle_y = I915_BIT_6_SWIZZLE_9;
588 			} else {
589 				swizzle_x = I915_BIT_6_SWIZZLE_NONE;
590 				swizzle_y = I915_BIT_6_SWIZZLE_NONE;
591 			}
592 		} else {
593 			u32 dimm_c0, dimm_c1;
594 
595 			dimm_c0 = intel_uncore_read(uncore, MAD_DIMM_C0);
596 			dimm_c1 = intel_uncore_read(uncore, MAD_DIMM_C1);
597 			dimm_c0 &= MAD_DIMM_A_SIZE_MASK | MAD_DIMM_B_SIZE_MASK;
598 			dimm_c1 &= MAD_DIMM_A_SIZE_MASK | MAD_DIMM_B_SIZE_MASK;
599 			/*
600 			 * Enable swizzling when the channels are populated
601 			 * with identically sized dimms. We don't need to check
602 			 * the 3rd channel because no cpu with gpu attached
603 			 * ships in that configuration. Also, swizzling only
604 			 * makes sense for 2 channels anyway.
605 			 */
606 			if (dimm_c0 == dimm_c1) {
607 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
608 				swizzle_y = I915_BIT_6_SWIZZLE_9;
609 			} else {
610 				swizzle_x = I915_BIT_6_SWIZZLE_NONE;
611 				swizzle_y = I915_BIT_6_SWIZZLE_NONE;
612 			}
613 		}
614 	} else if (IS_GEN(i915, 5)) {
615 		/*
616 		 * On Ironlake whatever DRAM config, GPU always do
617 		 * same swizzling setup.
618 		 */
619 		swizzle_x = I915_BIT_6_SWIZZLE_9_10;
620 		swizzle_y = I915_BIT_6_SWIZZLE_9;
621 	} else if (IS_GEN(i915, 2)) {
622 		/*
623 		 * As far as we know, the 865 doesn't have these bit 6
624 		 * swizzling issues.
625 		 */
626 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
627 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
628 	} else if (IS_G45(i915) || IS_I965G(i915) || IS_G33(i915)) {
629 		/*
630 		 * The 965, G33, and newer, have a very flexible memory
631 		 * configuration.  It will enable dual-channel mode
632 		 * (interleaving) on as much memory as it can, and the GPU
633 		 * will additionally sometimes enable different bit 6
634 		 * swizzling for tiled objects from the CPU.
635 		 *
636 		 * Here's what I found on the G965:
637 		 *    slot fill         memory size  swizzling
638 		 * 0A   0B   1A   1B    1-ch   2-ch
639 		 * 512  0    0    0     512    0     O
640 		 * 512  0    512  0     16     1008  X
641 		 * 512  0    0    512   16     1008  X
642 		 * 0    512  0    512   16     1008  X
643 		 * 1024 1024 1024 0     2048   1024  O
644 		 *
645 		 * We could probably detect this based on either the DRB
646 		 * matching, which was the case for the swizzling required in
647 		 * the table above, or from the 1-ch value being less than
648 		 * the minimum size of a rank.
649 		 *
650 		 * Reports indicate that the swizzling actually
651 		 * varies depending upon page placement inside the
652 		 * channels, i.e. we see swizzled pages where the
653 		 * banks of memory are paired and unswizzled on the
654 		 * uneven portion, so leave that as unknown.
655 		 */
656 		if (intel_uncore_read16(uncore, C0DRB3) ==
657 		    intel_uncore_read16(uncore, C1DRB3)) {
658 			swizzle_x = I915_BIT_6_SWIZZLE_9_10;
659 			swizzle_y = I915_BIT_6_SWIZZLE_9;
660 		}
661 	} else {
662 		u32 dcc = intel_uncore_read(uncore, DCC);
663 
664 		/*
665 		 * On 9xx chipsets, channel interleave by the CPU is
666 		 * determined by DCC.  For single-channel, neither the CPU
667 		 * nor the GPU do swizzling.  For dual channel interleaved,
668 		 * the GPU's interleave is bit 9 and 10 for X tiled, and bit
669 		 * 9 for Y tiled.  The CPU's interleave is independent, and
670 		 * can be based on either bit 11 (haven't seen this yet) or
671 		 * bit 17 (common).
672 		 */
673 		switch (dcc & DCC_ADDRESSING_MODE_MASK) {
674 		case DCC_ADDRESSING_MODE_SINGLE_CHANNEL:
675 		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_ASYMMETRIC:
676 			swizzle_x = I915_BIT_6_SWIZZLE_NONE;
677 			swizzle_y = I915_BIT_6_SWIZZLE_NONE;
678 			break;
679 		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_INTERLEAVED:
680 			if (dcc & DCC_CHANNEL_XOR_DISABLE) {
681 				/*
682 				 * This is the base swizzling by the GPU for
683 				 * tiled buffers.
684 				 */
685 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
686 				swizzle_y = I915_BIT_6_SWIZZLE_9;
687 			} else if ((dcc & DCC_CHANNEL_XOR_BIT_17) == 0) {
688 				/* Bit 11 swizzling by the CPU in addition. */
689 				swizzle_x = I915_BIT_6_SWIZZLE_9_10_11;
690 				swizzle_y = I915_BIT_6_SWIZZLE_9_11;
691 			} else {
692 				/* Bit 17 swizzling by the CPU in addition. */
693 				swizzle_x = I915_BIT_6_SWIZZLE_9_10_17;
694 				swizzle_y = I915_BIT_6_SWIZZLE_9_17;
695 			}
696 			break;
697 		}
698 
699 		/* check for L-shaped memory aka modified enhanced addressing */
700 		if (IS_GEN(i915, 4) &&
701 		    !(intel_uncore_read(uncore, DCC2) & DCC2_MODIFIED_ENHANCED_DISABLE)) {
702 			swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
703 			swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
704 		}
705 
706 		if (dcc == 0xffffffff) {
707 			drm_err(&i915->drm, "Couldn't read from MCHBAR.  "
708 				  "Disabling tiling.\n");
709 			swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
710 			swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
711 		}
712 	}
713 
714 	if (swizzle_x == I915_BIT_6_SWIZZLE_UNKNOWN ||
715 	    swizzle_y == I915_BIT_6_SWIZZLE_UNKNOWN) {
716 		/*
717 		 * Userspace likes to explode if it sees unknown swizzling,
718 		 * so lie. We will finish the lie when reporting through
719 		 * the get-tiling-ioctl by reporting the physical swizzle
720 		 * mode as unknown instead.
721 		 *
722 		 * As we don't strictly know what the swizzling is, it may be
723 		 * bit17 dependent, and so we need to also prevent the pages
724 		 * from being moved.
725 		 */
726 		i915->quirks |= QUIRK_PIN_SWIZZLED_PAGES;
727 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
728 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
729 	}
730 
731 	i915->ggtt.bit_6_swizzle_x = swizzle_x;
732 	i915->ggtt.bit_6_swizzle_y = swizzle_y;
733 }
734 
735 /*
736  * Swap every 64 bytes of this page around, to account for it having a new
737  * bit 17 of its physical address and therefore being interpreted differently
738  * by the GPU.
739  */
740 static void swizzle_page(struct page *page)
741 {
742 	char temp[64];
743 	char *vaddr;
744 	int i;
745 
746 	vaddr = kmap(page);
747 
748 	for (i = 0; i < PAGE_SIZE; i += 128) {
749 		memcpy(temp, &vaddr[i], 64);
750 		memcpy(&vaddr[i], &vaddr[i + 64], 64);
751 		memcpy(&vaddr[i + 64], temp, 64);
752 	}
753 
754 	kunmap(page);
755 }
756 
757 /**
758  * i915_gem_object_do_bit_17_swizzle - fixup bit 17 swizzling
759  * @obj: i915 GEM buffer object
760  * @pages: the scattergather list of physical pages
761  *
762  * This function fixes up the swizzling in case any page frame number for this
763  * object has changed in bit 17 since that state has been saved with
764  * i915_gem_object_save_bit_17_swizzle().
765  *
766  * This is called when pinning backing storage again, since the kernel is free
767  * to move unpinned backing storage around (either by directly moving pages or
768  * by swapping them out and back in again).
769  */
770 void
771 i915_gem_object_do_bit_17_swizzle(struct drm_i915_gem_object *obj,
772 				  struct sg_table *pages)
773 {
774 	struct sgt_iter sgt_iter;
775 	struct page *page;
776 	int i;
777 
778 	if (obj->bit_17 == NULL)
779 		return;
780 
781 	i = 0;
782 	for_each_sgt_page(page, sgt_iter, pages) {
783 		char new_bit_17 = page_to_phys(page) >> 17;
784 
785 		if ((new_bit_17 & 0x1) != (test_bit(i, obj->bit_17) != 0)) {
786 			swizzle_page(page);
787 			set_page_dirty(page);
788 		}
789 
790 		i++;
791 	}
792 }
793 
794 /**
795  * i915_gem_object_save_bit_17_swizzle - save bit 17 swizzling
796  * @obj: i915 GEM buffer object
797  * @pages: the scattergather list of physical pages
798  *
799  * This function saves the bit 17 of each page frame number so that swizzling
800  * can be fixed up later on with i915_gem_object_do_bit_17_swizzle(). This must
801  * be called before the backing storage can be unpinned.
802  */
803 void
804 i915_gem_object_save_bit_17_swizzle(struct drm_i915_gem_object *obj,
805 				    struct sg_table *pages)
806 {
807 	const unsigned int page_count = obj->base.size >> PAGE_SHIFT;
808 	struct sgt_iter sgt_iter;
809 	struct page *page;
810 	int i;
811 
812 	if (obj->bit_17 == NULL) {
813 		obj->bit_17 = bitmap_zalloc(page_count, GFP_KERNEL);
814 		if (obj->bit_17 == NULL) {
815 			DRM_ERROR("Failed to allocate memory for bit 17 "
816 				  "record\n");
817 			return;
818 		}
819 	}
820 
821 	i = 0;
822 
823 	for_each_sgt_page(page, sgt_iter, pages) {
824 		if (page_to_phys(page) & (1 << 17))
825 			__set_bit(i, obj->bit_17);
826 		else
827 			__clear_bit(i, obj->bit_17);
828 		i++;
829 	}
830 }
831 
832 void intel_ggtt_init_fences(struct i915_ggtt *ggtt)
833 {
834 	struct drm_i915_private *i915 = ggtt->vm.i915;
835 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
836 	int num_fences;
837 	int i;
838 
839 	INIT_LIST_HEAD(&ggtt->fence_list);
840 	INIT_LIST_HEAD(&ggtt->userfault_list);
841 	intel_wakeref_auto_init(&ggtt->userfault_wakeref, uncore->rpm);
842 
843 	detect_bit_6_swizzle(ggtt);
844 
845 	if (!i915_ggtt_has_aperture(ggtt))
846 		num_fences = 0;
847 	else if (INTEL_GEN(i915) >= 7 &&
848 		 !(IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)))
849 		num_fences = 32;
850 	else if (INTEL_GEN(i915) >= 4 ||
851 		 IS_I945G(i915) || IS_I945GM(i915) ||
852 		 IS_G33(i915) || IS_PINEVIEW(i915))
853 		num_fences = 16;
854 	else
855 		num_fences = 8;
856 
857 	if (intel_vgpu_active(i915))
858 		num_fences = intel_uncore_read(uncore,
859 					       vgtif_reg(avail_rs.fence_num));
860 	ggtt->fence_regs = kcalloc(num_fences,
861 				   sizeof(*ggtt->fence_regs),
862 				   GFP_KERNEL);
863 	if (!ggtt->fence_regs)
864 		num_fences = 0;
865 
866 	/* Initialize fence registers to zero */
867 	for (i = 0; i < num_fences; i++) {
868 		struct i915_fence_reg *fence = &ggtt->fence_regs[i];
869 
870 		i915_active_init(&fence->active, NULL, NULL);
871 		fence->ggtt = ggtt;
872 		fence->id = i;
873 		list_add_tail(&fence->link, &ggtt->fence_list);
874 	}
875 	ggtt->num_fences = num_fences;
876 
877 	intel_ggtt_restore_fences(ggtt);
878 }
879 
880 void intel_ggtt_fini_fences(struct i915_ggtt *ggtt)
881 {
882 	int i;
883 
884 	for (i = 0; i < ggtt->num_fences; i++) {
885 		struct i915_fence_reg *fence = &ggtt->fence_regs[i];
886 
887 		i915_active_fini(&fence->active);
888 	}
889 
890 	kfree(ggtt->fence_regs);
891 }
892 
893 void intel_gt_init_swizzling(struct intel_gt *gt)
894 {
895 	struct drm_i915_private *i915 = gt->i915;
896 	struct intel_uncore *uncore = gt->uncore;
897 
898 	if (INTEL_GEN(i915) < 5 ||
899 	    i915->ggtt.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
900 		return;
901 
902 	intel_uncore_rmw(uncore, DISP_ARB_CTL, 0, DISP_TILE_SURFACE_SWIZZLING);
903 
904 	if (IS_GEN(i915, 5))
905 		return;
906 
907 	intel_uncore_rmw(uncore, TILECTL, 0, TILECTL_SWZCTL);
908 
909 	if (IS_GEN(i915, 6))
910 		intel_uncore_write(uncore,
911 				   ARB_MODE,
912 				   _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
913 	else if (IS_GEN(i915, 7))
914 		intel_uncore_write(uncore,
915 				   ARB_MODE,
916 				   _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
917 	else if (IS_GEN(i915, 8))
918 		intel_uncore_write(uncore,
919 				   GAMTARBMODE,
920 				   _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
921 	else
922 		MISSING_CASE(INTEL_GEN(i915));
923 }
924