1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2008-2015 Intel Corporation
4  */
5 
6 #include "i915_drv.h"
7 #include "i915_reg.h"
8 #include "i915_scatterlist.h"
9 #include "i915_pvinfo.h"
10 #include "i915_vgpu.h"
11 #include "intel_gt_regs.h"
12 #include "intel_mchbar_regs.h"
13 
14 /**
15  * DOC: fence register handling
16  *
17  * Important to avoid confusions: "fences" in the i915 driver are not execution
18  * fences used to track command completion but hardware detiler objects which
19  * wrap a given range of the global GTT. Each platform has only a fairly limited
20  * set of these objects.
21  *
22  * Fences are used to detile GTT memory mappings. They're also connected to the
23  * hardware frontbuffer render tracking and hence interact with frontbuffer
24  * compression. Furthermore on older platforms fences are required for tiled
25  * objects used by the display engine. They can also be used by the render
26  * engine - they're required for blitter commands and are optional for render
27  * commands. But on gen4+ both display (with the exception of fbc) and rendering
28  * have their own tiling state bits and don't need fences.
29  *
30  * Also note that fences only support X and Y tiling and hence can't be used for
31  * the fancier new tiling formats like W, Ys and Yf.
32  *
33  * Finally note that because fences are such a restricted resource they're
34  * dynamically associated with objects. Furthermore fence state is committed to
35  * the hardware lazily to avoid unnecessary stalls on gen2/3. Therefore code must
36  * explicitly call i915_gem_object_get_fence() to synchronize fencing status
37  * for cpu access. Also note that some code wants an unfenced view, for those
38  * cases the fence can be removed forcefully with i915_gem_object_put_fence().
39  *
40  * Internally these functions will synchronize with userspace access by removing
41  * CPU ptes into GTT mmaps (not the GTT ptes themselves) as needed.
42  */
43 
44 #define pipelined 0
45 
46 static struct drm_i915_private *fence_to_i915(struct i915_fence_reg *fence)
47 {
48 	return fence->ggtt->vm.i915;
49 }
50 
51 static struct intel_uncore *fence_to_uncore(struct i915_fence_reg *fence)
52 {
53 	return fence->ggtt->vm.gt->uncore;
54 }
55 
56 static void i965_write_fence_reg(struct i915_fence_reg *fence)
57 {
58 	i915_reg_t fence_reg_lo, fence_reg_hi;
59 	int fence_pitch_shift;
60 	u64 val;
61 
62 	if (GRAPHICS_VER(fence_to_i915(fence)) >= 6) {
63 		fence_reg_lo = FENCE_REG_GEN6_LO(fence->id);
64 		fence_reg_hi = FENCE_REG_GEN6_HI(fence->id);
65 		fence_pitch_shift = GEN6_FENCE_PITCH_SHIFT;
66 
67 	} else {
68 		fence_reg_lo = FENCE_REG_965_LO(fence->id);
69 		fence_reg_hi = FENCE_REG_965_HI(fence->id);
70 		fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
71 	}
72 
73 	val = 0;
74 	if (fence->tiling) {
75 		unsigned int stride = fence->stride;
76 
77 		GEM_BUG_ON(!IS_ALIGNED(stride, 128));
78 
79 		val = fence->start + fence->size - I965_FENCE_PAGE;
80 		val <<= 32;
81 		val |= fence->start;
82 		val |= (u64)((stride / 128) - 1) << fence_pitch_shift;
83 		if (fence->tiling == I915_TILING_Y)
84 			val |= BIT(I965_FENCE_TILING_Y_SHIFT);
85 		val |= I965_FENCE_REG_VALID;
86 	}
87 
88 	if (!pipelined) {
89 		struct intel_uncore *uncore = fence_to_uncore(fence);
90 
91 		/*
92 		 * To w/a incoherency with non-atomic 64-bit register updates,
93 		 * we split the 64-bit update into two 32-bit writes. In order
94 		 * for a partial fence not to be evaluated between writes, we
95 		 * precede the update with write to turn off the fence register,
96 		 * and only enable the fence as the last step.
97 		 *
98 		 * For extra levels of paranoia, we make sure each step lands
99 		 * before applying the next step.
100 		 */
101 		intel_uncore_write_fw(uncore, fence_reg_lo, 0);
102 		intel_uncore_posting_read_fw(uncore, fence_reg_lo);
103 
104 		intel_uncore_write_fw(uncore, fence_reg_hi, upper_32_bits(val));
105 		intel_uncore_write_fw(uncore, fence_reg_lo, lower_32_bits(val));
106 		intel_uncore_posting_read_fw(uncore, fence_reg_lo);
107 	}
108 }
109 
110 static void i915_write_fence_reg(struct i915_fence_reg *fence)
111 {
112 	u32 val;
113 
114 	val = 0;
115 	if (fence->tiling) {
116 		unsigned int stride = fence->stride;
117 		unsigned int tiling = fence->tiling;
118 		bool is_y_tiled = tiling == I915_TILING_Y;
119 
120 		if (is_y_tiled && HAS_128_BYTE_Y_TILING(fence_to_i915(fence)))
121 			stride /= 128;
122 		else
123 			stride /= 512;
124 		GEM_BUG_ON(!is_power_of_2(stride));
125 
126 		val = fence->start;
127 		if (is_y_tiled)
128 			val |= BIT(I830_FENCE_TILING_Y_SHIFT);
129 		val |= I915_FENCE_SIZE_BITS(fence->size);
130 		val |= ilog2(stride) << I830_FENCE_PITCH_SHIFT;
131 
132 		val |= I830_FENCE_REG_VALID;
133 	}
134 
135 	if (!pipelined) {
136 		struct intel_uncore *uncore = fence_to_uncore(fence);
137 		i915_reg_t reg = FENCE_REG(fence->id);
138 
139 		intel_uncore_write_fw(uncore, reg, val);
140 		intel_uncore_posting_read_fw(uncore, reg);
141 	}
142 }
143 
144 static void i830_write_fence_reg(struct i915_fence_reg *fence)
145 {
146 	u32 val;
147 
148 	val = 0;
149 	if (fence->tiling) {
150 		unsigned int stride = fence->stride;
151 
152 		val = fence->start;
153 		if (fence->tiling == I915_TILING_Y)
154 			val |= BIT(I830_FENCE_TILING_Y_SHIFT);
155 		val |= I830_FENCE_SIZE_BITS(fence->size);
156 		val |= ilog2(stride / 128) << I830_FENCE_PITCH_SHIFT;
157 		val |= I830_FENCE_REG_VALID;
158 	}
159 
160 	if (!pipelined) {
161 		struct intel_uncore *uncore = fence_to_uncore(fence);
162 		i915_reg_t reg = FENCE_REG(fence->id);
163 
164 		intel_uncore_write_fw(uncore, reg, val);
165 		intel_uncore_posting_read_fw(uncore, reg);
166 	}
167 }
168 
169 static void fence_write(struct i915_fence_reg *fence)
170 {
171 	struct drm_i915_private *i915 = fence_to_i915(fence);
172 
173 	/*
174 	 * Previous access through the fence register is marshalled by
175 	 * the mb() inside the fault handlers (i915_gem_release_mmaps)
176 	 * and explicitly managed for internal users.
177 	 */
178 
179 	if (GRAPHICS_VER(i915) == 2)
180 		i830_write_fence_reg(fence);
181 	else if (GRAPHICS_VER(i915) == 3)
182 		i915_write_fence_reg(fence);
183 	else
184 		i965_write_fence_reg(fence);
185 
186 	/*
187 	 * Access through the fenced region afterwards is
188 	 * ordered by the posting reads whilst writing the registers.
189 	 */
190 }
191 
192 static bool gpu_uses_fence_registers(struct i915_fence_reg *fence)
193 {
194 	return GRAPHICS_VER(fence_to_i915(fence)) < 4;
195 }
196 
197 static int fence_update(struct i915_fence_reg *fence,
198 			struct i915_vma *vma)
199 {
200 	struct i915_ggtt *ggtt = fence->ggtt;
201 	struct intel_uncore *uncore = fence_to_uncore(fence);
202 	intel_wakeref_t wakeref;
203 	struct i915_vma *old;
204 	int ret;
205 
206 	fence->tiling = 0;
207 	if (vma) {
208 		GEM_BUG_ON(!i915_gem_object_get_stride(vma->obj) ||
209 			   !i915_gem_object_get_tiling(vma->obj));
210 
211 		if (!i915_vma_is_map_and_fenceable(vma))
212 			return -EINVAL;
213 
214 		if (gpu_uses_fence_registers(fence)) {
215 			/* implicit 'unfenced' GPU blits */
216 			ret = i915_vma_sync(vma);
217 			if (ret)
218 				return ret;
219 		}
220 
221 		fence->start = vma->node.start;
222 		fence->size = vma->fence_size;
223 		fence->stride = i915_gem_object_get_stride(vma->obj);
224 		fence->tiling = i915_gem_object_get_tiling(vma->obj);
225 	}
226 	WRITE_ONCE(fence->dirty, false);
227 
228 	old = xchg(&fence->vma, NULL);
229 	if (old) {
230 		/* XXX Ideally we would move the waiting to outside the mutex */
231 		ret = i915_active_wait(&fence->active);
232 		if (ret) {
233 			fence->vma = old;
234 			return ret;
235 		}
236 
237 		i915_vma_flush_writes(old);
238 
239 		/*
240 		 * Ensure that all userspace CPU access is completed before
241 		 * stealing the fence.
242 		 */
243 		if (old != vma) {
244 			GEM_BUG_ON(old->fence != fence);
245 			i915_vma_revoke_mmap(old);
246 			old->fence = NULL;
247 		}
248 
249 		list_move(&fence->link, &ggtt->fence_list);
250 	}
251 
252 	/*
253 	 * We only need to update the register itself if the device is awake.
254 	 * If the device is currently powered down, we will defer the write
255 	 * to the runtime resume, see intel_ggtt_restore_fences().
256 	 *
257 	 * This only works for removing the fence register, on acquisition
258 	 * the caller must hold the rpm wakeref. The fence register must
259 	 * be cleared before we can use any other fences to ensure that
260 	 * the new fences do not overlap the elided clears, confusing HW.
261 	 */
262 	wakeref = intel_runtime_pm_get_if_in_use(uncore->rpm);
263 	if (!wakeref) {
264 		GEM_BUG_ON(vma);
265 		return 0;
266 	}
267 
268 	WRITE_ONCE(fence->vma, vma);
269 	fence_write(fence);
270 
271 	if (vma) {
272 		vma->fence = fence;
273 		list_move_tail(&fence->link, &ggtt->fence_list);
274 	}
275 
276 	intel_runtime_pm_put(uncore->rpm, wakeref);
277 	return 0;
278 }
279 
280 /**
281  * i915_vma_revoke_fence - force-remove fence for a VMA
282  * @vma: vma to map linearly (not through a fence reg)
283  *
284  * This function force-removes any fence from the given object, which is useful
285  * if the kernel wants to do untiled GTT access.
286  */
287 void i915_vma_revoke_fence(struct i915_vma *vma)
288 {
289 	struct i915_fence_reg *fence = vma->fence;
290 	intel_wakeref_t wakeref;
291 
292 	lockdep_assert_held(&vma->vm->mutex);
293 	if (!fence)
294 		return;
295 
296 	GEM_BUG_ON(fence->vma != vma);
297 	GEM_BUG_ON(!i915_active_is_idle(&fence->active));
298 	GEM_BUG_ON(atomic_read(&fence->pin_count));
299 
300 	fence->tiling = 0;
301 	WRITE_ONCE(fence->vma, NULL);
302 	vma->fence = NULL;
303 
304 	/*
305 	 * Skip the write to HW if and only if the device is currently
306 	 * suspended.
307 	 *
308 	 * If the driver does not currently hold a wakeref (if_in_use == 0),
309 	 * the device may currently be runtime suspended, or it may be woken
310 	 * up before the suspend takes place. If the device is not suspended
311 	 * (powered down) and we skip clearing the fence register, the HW is
312 	 * left in an undefined state where we may end up with multiple
313 	 * registers overlapping.
314 	 */
315 	with_intel_runtime_pm_if_active(fence_to_uncore(fence)->rpm, wakeref)
316 		fence_write(fence);
317 }
318 
319 static bool fence_is_active(const struct i915_fence_reg *fence)
320 {
321 	return fence->vma && i915_vma_is_active(fence->vma);
322 }
323 
324 static struct i915_fence_reg *fence_find(struct i915_ggtt *ggtt)
325 {
326 	struct i915_fence_reg *active = NULL;
327 	struct i915_fence_reg *fence, *fn;
328 
329 	list_for_each_entry_safe(fence, fn, &ggtt->fence_list, link) {
330 		GEM_BUG_ON(fence->vma && fence->vma->fence != fence);
331 
332 		if (fence == active) /* now seen this fence twice */
333 			active = ERR_PTR(-EAGAIN);
334 
335 		/* Prefer idle fences so we do not have to wait on the GPU */
336 		if (active != ERR_PTR(-EAGAIN) && fence_is_active(fence)) {
337 			if (!active)
338 				active = fence;
339 
340 			list_move_tail(&fence->link, &ggtt->fence_list);
341 			continue;
342 		}
343 
344 		if (atomic_read(&fence->pin_count))
345 			continue;
346 
347 		return fence;
348 	}
349 
350 	/* Wait for completion of pending flips which consume fences */
351 	if (intel_has_pending_fb_unpin(ggtt->vm.i915))
352 		return ERR_PTR(-EAGAIN);
353 
354 	return ERR_PTR(-ENOBUFS);
355 }
356 
357 int __i915_vma_pin_fence(struct i915_vma *vma)
358 {
359 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vma->vm);
360 	struct i915_fence_reg *fence;
361 	struct i915_vma *set = i915_gem_object_is_tiled(vma->obj) ? vma : NULL;
362 	int err;
363 
364 	lockdep_assert_held(&vma->vm->mutex);
365 
366 	/* Just update our place in the LRU if our fence is getting reused. */
367 	if (vma->fence) {
368 		fence = vma->fence;
369 		GEM_BUG_ON(fence->vma != vma);
370 		atomic_inc(&fence->pin_count);
371 		if (!fence->dirty) {
372 			list_move_tail(&fence->link, &ggtt->fence_list);
373 			return 0;
374 		}
375 	} else if (set) {
376 		fence = fence_find(ggtt);
377 		if (IS_ERR(fence))
378 			return PTR_ERR(fence);
379 
380 		GEM_BUG_ON(atomic_read(&fence->pin_count));
381 		atomic_inc(&fence->pin_count);
382 	} else {
383 		return 0;
384 	}
385 
386 	err = fence_update(fence, set);
387 	if (err)
388 		goto out_unpin;
389 
390 	GEM_BUG_ON(fence->vma != set);
391 	GEM_BUG_ON(vma->fence != (set ? fence : NULL));
392 
393 	if (set)
394 		return 0;
395 
396 out_unpin:
397 	atomic_dec(&fence->pin_count);
398 	return err;
399 }
400 
401 /**
402  * i915_vma_pin_fence - set up fencing for a vma
403  * @vma: vma to map through a fence reg
404  *
405  * When mapping objects through the GTT, userspace wants to be able to write
406  * to them without having to worry about swizzling if the object is tiled.
407  * This function walks the fence regs looking for a free one for @obj,
408  * stealing one if it can't find any.
409  *
410  * It then sets up the reg based on the object's properties: address, pitch
411  * and tiling format.
412  *
413  * For an untiled surface, this removes any existing fence.
414  *
415  * Returns:
416  *
417  * 0 on success, negative error code on failure.
418  */
419 int i915_vma_pin_fence(struct i915_vma *vma)
420 {
421 	int err;
422 
423 	if (!vma->fence && !i915_gem_object_is_tiled(vma->obj))
424 		return 0;
425 
426 	/*
427 	 * Note that we revoke fences on runtime suspend. Therefore the user
428 	 * must keep the device awake whilst using the fence.
429 	 */
430 	assert_rpm_wakelock_held(vma->vm->gt->uncore->rpm);
431 	GEM_BUG_ON(!i915_vma_is_ggtt(vma));
432 
433 	err = mutex_lock_interruptible(&vma->vm->mutex);
434 	if (err)
435 		return err;
436 
437 	err = __i915_vma_pin_fence(vma);
438 	mutex_unlock(&vma->vm->mutex);
439 
440 	return err;
441 }
442 
443 /**
444  * i915_reserve_fence - Reserve a fence for vGPU
445  * @ggtt: Global GTT
446  *
447  * This function walks the fence regs looking for a free one and remove
448  * it from the fence_list. It is used to reserve fence for vGPU to use.
449  */
450 struct i915_fence_reg *i915_reserve_fence(struct i915_ggtt *ggtt)
451 {
452 	struct i915_fence_reg *fence;
453 	int count;
454 	int ret;
455 
456 	lockdep_assert_held(&ggtt->vm.mutex);
457 
458 	/* Keep at least one fence available for the display engine. */
459 	count = 0;
460 	list_for_each_entry(fence, &ggtt->fence_list, link)
461 		count += !atomic_read(&fence->pin_count);
462 	if (count <= 1)
463 		return ERR_PTR(-ENOSPC);
464 
465 	fence = fence_find(ggtt);
466 	if (IS_ERR(fence))
467 		return fence;
468 
469 	if (fence->vma) {
470 		/* Force-remove fence from VMA */
471 		ret = fence_update(fence, NULL);
472 		if (ret)
473 			return ERR_PTR(ret);
474 	}
475 
476 	list_del(&fence->link);
477 
478 	return fence;
479 }
480 
481 /**
482  * i915_unreserve_fence - Reclaim a reserved fence
483  * @fence: the fence reg
484  *
485  * This function add a reserved fence register from vGPU to the fence_list.
486  */
487 void i915_unreserve_fence(struct i915_fence_reg *fence)
488 {
489 	struct i915_ggtt *ggtt = fence->ggtt;
490 
491 	lockdep_assert_held(&ggtt->vm.mutex);
492 
493 	list_add(&fence->link, &ggtt->fence_list);
494 }
495 
496 /**
497  * intel_ggtt_restore_fences - restore fence state
498  * @ggtt: Global GTT
499  *
500  * Restore the hw fence state to match the software tracking again, to be called
501  * after a gpu reset and on resume. Note that on runtime suspend we only cancel
502  * the fences, to be reacquired by the user later.
503  */
504 void intel_ggtt_restore_fences(struct i915_ggtt *ggtt)
505 {
506 	int i;
507 
508 	for (i = 0; i < ggtt->num_fences; i++)
509 		fence_write(&ggtt->fence_regs[i]);
510 }
511 
512 /**
513  * DOC: tiling swizzling details
514  *
515  * The idea behind tiling is to increase cache hit rates by rearranging
516  * pixel data so that a group of pixel accesses are in the same cacheline.
517  * Performance improvement from doing this on the back/depth buffer are on
518  * the order of 30%.
519  *
520  * Intel architectures make this somewhat more complicated, though, by
521  * adjustments made to addressing of data when the memory is in interleaved
522  * mode (matched pairs of DIMMS) to improve memory bandwidth.
523  * For interleaved memory, the CPU sends every sequential 64 bytes
524  * to an alternate memory channel so it can get the bandwidth from both.
525  *
526  * The GPU also rearranges its accesses for increased bandwidth to interleaved
527  * memory, and it matches what the CPU does for non-tiled.  However, when tiled
528  * it does it a little differently, since one walks addresses not just in the
529  * X direction but also Y.  So, along with alternating channels when bit
530  * 6 of the address flips, it also alternates when other bits flip --  Bits 9
531  * (every 512 bytes, an X tile scanline) and 10 (every two X tile scanlines)
532  * are common to both the 915 and 965-class hardware.
533  *
534  * The CPU also sometimes XORs in higher bits as well, to improve
535  * bandwidth doing strided access like we do so frequently in graphics.  This
536  * is called "Channel XOR Randomization" in the MCH documentation.  The result
537  * is that the CPU is XORing in either bit 11 or bit 17 to bit 6 of its address
538  * decode.
539  *
540  * All of this bit 6 XORing has an effect on our memory management,
541  * as we need to make sure that the 3d driver can correctly address object
542  * contents.
543  *
544  * If we don't have interleaved memory, all tiling is safe and no swizzling is
545  * required.
546  *
547  * When bit 17 is XORed in, we simply refuse to tile at all.  Bit
548  * 17 is not just a page offset, so as we page an object out and back in,
549  * individual pages in it will have different bit 17 addresses, resulting in
550  * each 64 bytes being swapped with its neighbor!
551  *
552  * Otherwise, if interleaved, we have to tell the 3d driver what the address
553  * swizzling it needs to do is, since it's writing with the CPU to the pages
554  * (bit 6 and potentially bit 11 XORed in), and the GPU is reading from the
555  * pages (bit 6, 9, and 10 XORed in), resulting in a cumulative bit swizzling
556  * required by the CPU of XORing in bit 6, 9, 10, and potentially 11, in order
557  * to match what the GPU expects.
558  */
559 
560 /**
561  * detect_bit_6_swizzle - detect bit 6 swizzling pattern
562  * @ggtt: Global GGTT
563  *
564  * Detects bit 6 swizzling of address lookup between IGD access and CPU
565  * access through main memory.
566  */
567 static void detect_bit_6_swizzle(struct i915_ggtt *ggtt)
568 {
569 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
570 	struct drm_i915_private *i915 = ggtt->vm.i915;
571 	u32 swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
572 	u32 swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
573 
574 	if (GRAPHICS_VER(i915) >= 8 || IS_VALLEYVIEW(i915)) {
575 		/*
576 		 * On BDW+, swizzling is not used. We leave the CPU memory
577 		 * controller in charge of optimizing memory accesses without
578 		 * the extra address manipulation GPU side.
579 		 *
580 		 * VLV and CHV don't have GPU swizzling.
581 		 */
582 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
583 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
584 	} else if (GRAPHICS_VER(i915) >= 6) {
585 		if (i915->preserve_bios_swizzle) {
586 			if (intel_uncore_read(uncore, DISP_ARB_CTL) &
587 			    DISP_TILE_SURFACE_SWIZZLING) {
588 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
589 				swizzle_y = I915_BIT_6_SWIZZLE_9;
590 			} else {
591 				swizzle_x = I915_BIT_6_SWIZZLE_NONE;
592 				swizzle_y = I915_BIT_6_SWIZZLE_NONE;
593 			}
594 		} else {
595 			u32 dimm_c0, dimm_c1;
596 
597 			dimm_c0 = intel_uncore_read(uncore, MAD_DIMM_C0);
598 			dimm_c1 = intel_uncore_read(uncore, MAD_DIMM_C1);
599 			dimm_c0 &= MAD_DIMM_A_SIZE_MASK | MAD_DIMM_B_SIZE_MASK;
600 			dimm_c1 &= MAD_DIMM_A_SIZE_MASK | MAD_DIMM_B_SIZE_MASK;
601 			/*
602 			 * Enable swizzling when the channels are populated
603 			 * with identically sized dimms. We don't need to check
604 			 * the 3rd channel because no cpu with gpu attached
605 			 * ships in that configuration. Also, swizzling only
606 			 * makes sense for 2 channels anyway.
607 			 */
608 			if (dimm_c0 == dimm_c1) {
609 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
610 				swizzle_y = I915_BIT_6_SWIZZLE_9;
611 			} else {
612 				swizzle_x = I915_BIT_6_SWIZZLE_NONE;
613 				swizzle_y = I915_BIT_6_SWIZZLE_NONE;
614 			}
615 		}
616 	} else if (GRAPHICS_VER(i915) == 5) {
617 		/*
618 		 * On Ironlake whatever DRAM config, GPU always do
619 		 * same swizzling setup.
620 		 */
621 		swizzle_x = I915_BIT_6_SWIZZLE_9_10;
622 		swizzle_y = I915_BIT_6_SWIZZLE_9;
623 	} else if (GRAPHICS_VER(i915) == 2) {
624 		/*
625 		 * As far as we know, the 865 doesn't have these bit 6
626 		 * swizzling issues.
627 		 */
628 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
629 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
630 	} else if (IS_G45(i915) || IS_I965G(i915) || IS_G33(i915)) {
631 		/*
632 		 * The 965, G33, and newer, have a very flexible memory
633 		 * configuration.  It will enable dual-channel mode
634 		 * (interleaving) on as much memory as it can, and the GPU
635 		 * will additionally sometimes enable different bit 6
636 		 * swizzling for tiled objects from the CPU.
637 		 *
638 		 * Here's what I found on the G965:
639 		 *    slot fill         memory size  swizzling
640 		 * 0A   0B   1A   1B    1-ch   2-ch
641 		 * 512  0    0    0     512    0     O
642 		 * 512  0    512  0     16     1008  X
643 		 * 512  0    0    512   16     1008  X
644 		 * 0    512  0    512   16     1008  X
645 		 * 1024 1024 1024 0     2048   1024  O
646 		 *
647 		 * We could probably detect this based on either the DRB
648 		 * matching, which was the case for the swizzling required in
649 		 * the table above, or from the 1-ch value being less than
650 		 * the minimum size of a rank.
651 		 *
652 		 * Reports indicate that the swizzling actually
653 		 * varies depending upon page placement inside the
654 		 * channels, i.e. we see swizzled pages where the
655 		 * banks of memory are paired and unswizzled on the
656 		 * uneven portion, so leave that as unknown.
657 		 */
658 		if (intel_uncore_read16(uncore, C0DRB3_BW) ==
659 		    intel_uncore_read16(uncore, C1DRB3_BW)) {
660 			swizzle_x = I915_BIT_6_SWIZZLE_9_10;
661 			swizzle_y = I915_BIT_6_SWIZZLE_9;
662 		}
663 	} else {
664 		u32 dcc = intel_uncore_read(uncore, DCC);
665 
666 		/*
667 		 * On 9xx chipsets, channel interleave by the CPU is
668 		 * determined by DCC.  For single-channel, neither the CPU
669 		 * nor the GPU do swizzling.  For dual channel interleaved,
670 		 * the GPU's interleave is bit 9 and 10 for X tiled, and bit
671 		 * 9 for Y tiled.  The CPU's interleave is independent, and
672 		 * can be based on either bit 11 (haven't seen this yet) or
673 		 * bit 17 (common).
674 		 */
675 		switch (dcc & DCC_ADDRESSING_MODE_MASK) {
676 		case DCC_ADDRESSING_MODE_SINGLE_CHANNEL:
677 		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_ASYMMETRIC:
678 			swizzle_x = I915_BIT_6_SWIZZLE_NONE;
679 			swizzle_y = I915_BIT_6_SWIZZLE_NONE;
680 			break;
681 		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_INTERLEAVED:
682 			if (dcc & DCC_CHANNEL_XOR_DISABLE) {
683 				/*
684 				 * This is the base swizzling by the GPU for
685 				 * tiled buffers.
686 				 */
687 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
688 				swizzle_y = I915_BIT_6_SWIZZLE_9;
689 			} else if ((dcc & DCC_CHANNEL_XOR_BIT_17) == 0) {
690 				/* Bit 11 swizzling by the CPU in addition. */
691 				swizzle_x = I915_BIT_6_SWIZZLE_9_10_11;
692 				swizzle_y = I915_BIT_6_SWIZZLE_9_11;
693 			} else {
694 				/* Bit 17 swizzling by the CPU in addition. */
695 				swizzle_x = I915_BIT_6_SWIZZLE_9_10_17;
696 				swizzle_y = I915_BIT_6_SWIZZLE_9_17;
697 			}
698 			break;
699 		}
700 
701 		/* check for L-shaped memory aka modified enhanced addressing */
702 		if (GRAPHICS_VER(i915) == 4 &&
703 		    !(intel_uncore_read(uncore, DCC2) & DCC2_MODIFIED_ENHANCED_DISABLE)) {
704 			swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
705 			swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
706 		}
707 
708 		if (dcc == 0xffffffff) {
709 			drm_err(&i915->drm, "Couldn't read from MCHBAR.  "
710 				  "Disabling tiling.\n");
711 			swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
712 			swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
713 		}
714 	}
715 
716 	if (swizzle_x == I915_BIT_6_SWIZZLE_UNKNOWN ||
717 	    swizzle_y == I915_BIT_6_SWIZZLE_UNKNOWN) {
718 		/*
719 		 * Userspace likes to explode if it sees unknown swizzling,
720 		 * so lie. We will finish the lie when reporting through
721 		 * the get-tiling-ioctl by reporting the physical swizzle
722 		 * mode as unknown instead.
723 		 *
724 		 * As we don't strictly know what the swizzling is, it may be
725 		 * bit17 dependent, and so we need to also prevent the pages
726 		 * from being moved.
727 		 */
728 		i915->quirks |= QUIRK_PIN_SWIZZLED_PAGES;
729 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
730 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
731 	}
732 
733 	to_gt(i915)->ggtt->bit_6_swizzle_x = swizzle_x;
734 	to_gt(i915)->ggtt->bit_6_swizzle_y = swizzle_y;
735 }
736 
737 /*
738  * Swap every 64 bytes of this page around, to account for it having a new
739  * bit 17 of its physical address and therefore being interpreted differently
740  * by the GPU.
741  */
742 static void swizzle_page(struct page *page)
743 {
744 	char temp[64];
745 	char *vaddr;
746 	int i;
747 
748 	vaddr = kmap(page);
749 
750 	for (i = 0; i < PAGE_SIZE; i += 128) {
751 		memcpy(temp, &vaddr[i], 64);
752 		memcpy(&vaddr[i], &vaddr[i + 64], 64);
753 		memcpy(&vaddr[i + 64], temp, 64);
754 	}
755 
756 	kunmap(page);
757 }
758 
759 /**
760  * i915_gem_object_do_bit_17_swizzle - fixup bit 17 swizzling
761  * @obj: i915 GEM buffer object
762  * @pages: the scattergather list of physical pages
763  *
764  * This function fixes up the swizzling in case any page frame number for this
765  * object has changed in bit 17 since that state has been saved with
766  * i915_gem_object_save_bit_17_swizzle().
767  *
768  * This is called when pinning backing storage again, since the kernel is free
769  * to move unpinned backing storage around (either by directly moving pages or
770  * by swapping them out and back in again).
771  */
772 void
773 i915_gem_object_do_bit_17_swizzle(struct drm_i915_gem_object *obj,
774 				  struct sg_table *pages)
775 {
776 	struct sgt_iter sgt_iter;
777 	struct page *page;
778 	int i;
779 
780 	if (obj->bit_17 == NULL)
781 		return;
782 
783 	i = 0;
784 	for_each_sgt_page(page, sgt_iter, pages) {
785 		char new_bit_17 = page_to_phys(page) >> 17;
786 
787 		if ((new_bit_17 & 0x1) != (test_bit(i, obj->bit_17) != 0)) {
788 			swizzle_page(page);
789 			set_page_dirty(page);
790 		}
791 
792 		i++;
793 	}
794 }
795 
796 /**
797  * i915_gem_object_save_bit_17_swizzle - save bit 17 swizzling
798  * @obj: i915 GEM buffer object
799  * @pages: the scattergather list of physical pages
800  *
801  * This function saves the bit 17 of each page frame number so that swizzling
802  * can be fixed up later on with i915_gem_object_do_bit_17_swizzle(). This must
803  * be called before the backing storage can be unpinned.
804  */
805 void
806 i915_gem_object_save_bit_17_swizzle(struct drm_i915_gem_object *obj,
807 				    struct sg_table *pages)
808 {
809 	const unsigned int page_count = obj->base.size >> PAGE_SHIFT;
810 	struct sgt_iter sgt_iter;
811 	struct page *page;
812 	int i;
813 
814 	if (obj->bit_17 == NULL) {
815 		obj->bit_17 = bitmap_zalloc(page_count, GFP_KERNEL);
816 		if (obj->bit_17 == NULL) {
817 			DRM_ERROR("Failed to allocate memory for bit 17 "
818 				  "record\n");
819 			return;
820 		}
821 	}
822 
823 	i = 0;
824 
825 	for_each_sgt_page(page, sgt_iter, pages) {
826 		if (page_to_phys(page) & (1 << 17))
827 			__set_bit(i, obj->bit_17);
828 		else
829 			__clear_bit(i, obj->bit_17);
830 		i++;
831 	}
832 }
833 
834 void intel_ggtt_init_fences(struct i915_ggtt *ggtt)
835 {
836 	struct drm_i915_private *i915 = ggtt->vm.i915;
837 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
838 	int num_fences;
839 	int i;
840 
841 	INIT_LIST_HEAD(&ggtt->fence_list);
842 	INIT_LIST_HEAD(&ggtt->userfault_list);
843 	intel_wakeref_auto_init(&ggtt->userfault_wakeref, uncore->rpm);
844 
845 	detect_bit_6_swizzle(ggtt);
846 
847 	if (!i915_ggtt_has_aperture(ggtt))
848 		num_fences = 0;
849 	else if (GRAPHICS_VER(i915) >= 7 &&
850 		 !(IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)))
851 		num_fences = 32;
852 	else if (GRAPHICS_VER(i915) >= 4 ||
853 		 IS_I945G(i915) || IS_I945GM(i915) ||
854 		 IS_G33(i915) || IS_PINEVIEW(i915))
855 		num_fences = 16;
856 	else
857 		num_fences = 8;
858 
859 	if (intel_vgpu_active(i915))
860 		num_fences = intel_uncore_read(uncore,
861 					       vgtif_reg(avail_rs.fence_num));
862 	ggtt->fence_regs = kcalloc(num_fences,
863 				   sizeof(*ggtt->fence_regs),
864 				   GFP_KERNEL);
865 	if (!ggtt->fence_regs)
866 		num_fences = 0;
867 
868 	/* Initialize fence registers to zero */
869 	for (i = 0; i < num_fences; i++) {
870 		struct i915_fence_reg *fence = &ggtt->fence_regs[i];
871 
872 		i915_active_init(&fence->active, NULL, NULL, 0);
873 		fence->ggtt = ggtt;
874 		fence->id = i;
875 		list_add_tail(&fence->link, &ggtt->fence_list);
876 	}
877 	ggtt->num_fences = num_fences;
878 
879 	intel_ggtt_restore_fences(ggtt);
880 }
881 
882 void intel_ggtt_fini_fences(struct i915_ggtt *ggtt)
883 {
884 	int i;
885 
886 	for (i = 0; i < ggtt->num_fences; i++) {
887 		struct i915_fence_reg *fence = &ggtt->fence_regs[i];
888 
889 		i915_active_fini(&fence->active);
890 	}
891 
892 	kfree(ggtt->fence_regs);
893 }
894 
895 void intel_gt_init_swizzling(struct intel_gt *gt)
896 {
897 	struct drm_i915_private *i915 = gt->i915;
898 	struct intel_uncore *uncore = gt->uncore;
899 
900 	if (GRAPHICS_VER(i915) < 5 ||
901 	    to_gt(i915)->ggtt->bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
902 		return;
903 
904 	intel_uncore_rmw(uncore, DISP_ARB_CTL, 0, DISP_TILE_SURFACE_SWIZZLING);
905 
906 	if (GRAPHICS_VER(i915) == 5)
907 		return;
908 
909 	intel_uncore_rmw(uncore, TILECTL, 0, TILECTL_SWZCTL);
910 
911 	if (GRAPHICS_VER(i915) == 6)
912 		intel_uncore_write(uncore,
913 				   ARB_MODE,
914 				   _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
915 	else if (GRAPHICS_VER(i915) == 7)
916 		intel_uncore_write(uncore,
917 				   ARB_MODE,
918 				   _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
919 	else if (GRAPHICS_VER(i915) == 8)
920 		intel_uncore_write(uncore,
921 				   GAMTARBMODE,
922 				   _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
923 	else
924 		MISSING_CASE(GRAPHICS_VER(i915));
925 }
926