1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2008-2015 Intel Corporation
4  */
5 
6 #include <linux/highmem.h>
7 
8 #include "display/intel_display.h"
9 #include "i915_drv.h"
10 #include "i915_reg.h"
11 #include "i915_scatterlist.h"
12 #include "i915_pvinfo.h"
13 #include "i915_vgpu.h"
14 #include "intel_gt_regs.h"
15 #include "intel_mchbar_regs.h"
16 
17 /**
18  * DOC: fence register handling
19  *
20  * Important to avoid confusions: "fences" in the i915 driver are not execution
21  * fences used to track command completion but hardware detiler objects which
22  * wrap a given range of the global GTT. Each platform has only a fairly limited
23  * set of these objects.
24  *
25  * Fences are used to detile GTT memory mappings. They're also connected to the
26  * hardware frontbuffer render tracking and hence interact with frontbuffer
27  * compression. Furthermore on older platforms fences are required for tiled
28  * objects used by the display engine. They can also be used by the render
29  * engine - they're required for blitter commands and are optional for render
30  * commands. But on gen4+ both display (with the exception of fbc) and rendering
31  * have their own tiling state bits and don't need fences.
32  *
33  * Also note that fences only support X and Y tiling and hence can't be used for
34  * the fancier new tiling formats like W, Ys and Yf.
35  *
36  * Finally note that because fences are such a restricted resource they're
37  * dynamically associated with objects. Furthermore fence state is committed to
38  * the hardware lazily to avoid unnecessary stalls on gen2/3. Therefore code must
39  * explicitly call i915_gem_object_get_fence() to synchronize fencing status
40  * for cpu access. Also note that some code wants an unfenced view, for those
41  * cases the fence can be removed forcefully with i915_gem_object_put_fence().
42  *
43  * Internally these functions will synchronize with userspace access by removing
44  * CPU ptes into GTT mmaps (not the GTT ptes themselves) as needed.
45  */
46 
47 #define pipelined 0
48 
49 static struct drm_i915_private *fence_to_i915(struct i915_fence_reg *fence)
50 {
51 	return fence->ggtt->vm.i915;
52 }
53 
54 static struct intel_uncore *fence_to_uncore(struct i915_fence_reg *fence)
55 {
56 	return fence->ggtt->vm.gt->uncore;
57 }
58 
59 static void i965_write_fence_reg(struct i915_fence_reg *fence)
60 {
61 	i915_reg_t fence_reg_lo, fence_reg_hi;
62 	int fence_pitch_shift;
63 	u64 val;
64 
65 	if (GRAPHICS_VER(fence_to_i915(fence)) >= 6) {
66 		fence_reg_lo = FENCE_REG_GEN6_LO(fence->id);
67 		fence_reg_hi = FENCE_REG_GEN6_HI(fence->id);
68 		fence_pitch_shift = GEN6_FENCE_PITCH_SHIFT;
69 
70 	} else {
71 		fence_reg_lo = FENCE_REG_965_LO(fence->id);
72 		fence_reg_hi = FENCE_REG_965_HI(fence->id);
73 		fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
74 	}
75 
76 	val = 0;
77 	if (fence->tiling) {
78 		unsigned int stride = fence->stride;
79 
80 		GEM_BUG_ON(!IS_ALIGNED(stride, 128));
81 
82 		val = fence->start + fence->size - I965_FENCE_PAGE;
83 		val <<= 32;
84 		val |= fence->start;
85 		val |= (u64)((stride / 128) - 1) << fence_pitch_shift;
86 		if (fence->tiling == I915_TILING_Y)
87 			val |= BIT(I965_FENCE_TILING_Y_SHIFT);
88 		val |= I965_FENCE_REG_VALID;
89 	}
90 
91 	if (!pipelined) {
92 		struct intel_uncore *uncore = fence_to_uncore(fence);
93 
94 		/*
95 		 * To w/a incoherency with non-atomic 64-bit register updates,
96 		 * we split the 64-bit update into two 32-bit writes. In order
97 		 * for a partial fence not to be evaluated between writes, we
98 		 * precede the update with write to turn off the fence register,
99 		 * and only enable the fence as the last step.
100 		 *
101 		 * For extra levels of paranoia, we make sure each step lands
102 		 * before applying the next step.
103 		 */
104 		intel_uncore_write_fw(uncore, fence_reg_lo, 0);
105 		intel_uncore_posting_read_fw(uncore, fence_reg_lo);
106 
107 		intel_uncore_write_fw(uncore, fence_reg_hi, upper_32_bits(val));
108 		intel_uncore_write_fw(uncore, fence_reg_lo, lower_32_bits(val));
109 		intel_uncore_posting_read_fw(uncore, fence_reg_lo);
110 	}
111 }
112 
113 static void i915_write_fence_reg(struct i915_fence_reg *fence)
114 {
115 	u32 val;
116 
117 	val = 0;
118 	if (fence->tiling) {
119 		unsigned int stride = fence->stride;
120 		unsigned int tiling = fence->tiling;
121 		bool is_y_tiled = tiling == I915_TILING_Y;
122 
123 		if (is_y_tiled && HAS_128_BYTE_Y_TILING(fence_to_i915(fence)))
124 			stride /= 128;
125 		else
126 			stride /= 512;
127 		GEM_BUG_ON(!is_power_of_2(stride));
128 
129 		val = fence->start;
130 		if (is_y_tiled)
131 			val |= BIT(I830_FENCE_TILING_Y_SHIFT);
132 		val |= I915_FENCE_SIZE_BITS(fence->size);
133 		val |= ilog2(stride) << I830_FENCE_PITCH_SHIFT;
134 
135 		val |= I830_FENCE_REG_VALID;
136 	}
137 
138 	if (!pipelined) {
139 		struct intel_uncore *uncore = fence_to_uncore(fence);
140 		i915_reg_t reg = FENCE_REG(fence->id);
141 
142 		intel_uncore_write_fw(uncore, reg, val);
143 		intel_uncore_posting_read_fw(uncore, reg);
144 	}
145 }
146 
147 static void i830_write_fence_reg(struct i915_fence_reg *fence)
148 {
149 	u32 val;
150 
151 	val = 0;
152 	if (fence->tiling) {
153 		unsigned int stride = fence->stride;
154 
155 		val = fence->start;
156 		if (fence->tiling == I915_TILING_Y)
157 			val |= BIT(I830_FENCE_TILING_Y_SHIFT);
158 		val |= I830_FENCE_SIZE_BITS(fence->size);
159 		val |= ilog2(stride / 128) << I830_FENCE_PITCH_SHIFT;
160 		val |= I830_FENCE_REG_VALID;
161 	}
162 
163 	if (!pipelined) {
164 		struct intel_uncore *uncore = fence_to_uncore(fence);
165 		i915_reg_t reg = FENCE_REG(fence->id);
166 
167 		intel_uncore_write_fw(uncore, reg, val);
168 		intel_uncore_posting_read_fw(uncore, reg);
169 	}
170 }
171 
172 static void fence_write(struct i915_fence_reg *fence)
173 {
174 	struct drm_i915_private *i915 = fence_to_i915(fence);
175 
176 	/*
177 	 * Previous access through the fence register is marshalled by
178 	 * the mb() inside the fault handlers (i915_gem_release_mmaps)
179 	 * and explicitly managed for internal users.
180 	 */
181 
182 	if (GRAPHICS_VER(i915) == 2)
183 		i830_write_fence_reg(fence);
184 	else if (GRAPHICS_VER(i915) == 3)
185 		i915_write_fence_reg(fence);
186 	else
187 		i965_write_fence_reg(fence);
188 
189 	/*
190 	 * Access through the fenced region afterwards is
191 	 * ordered by the posting reads whilst writing the registers.
192 	 */
193 }
194 
195 static bool gpu_uses_fence_registers(struct i915_fence_reg *fence)
196 {
197 	return GRAPHICS_VER(fence_to_i915(fence)) < 4;
198 }
199 
200 static int fence_update(struct i915_fence_reg *fence,
201 			struct i915_vma *vma)
202 {
203 	struct i915_ggtt *ggtt = fence->ggtt;
204 	struct intel_uncore *uncore = fence_to_uncore(fence);
205 	intel_wakeref_t wakeref;
206 	struct i915_vma *old;
207 	int ret;
208 
209 	fence->tiling = 0;
210 	if (vma) {
211 		GEM_BUG_ON(!i915_gem_object_get_stride(vma->obj) ||
212 			   !i915_gem_object_get_tiling(vma->obj));
213 
214 		if (!i915_vma_is_map_and_fenceable(vma))
215 			return -EINVAL;
216 
217 		if (gpu_uses_fence_registers(fence)) {
218 			/* implicit 'unfenced' GPU blits */
219 			ret = i915_vma_sync(vma);
220 			if (ret)
221 				return ret;
222 		}
223 
224 		GEM_BUG_ON(vma->fence_size > i915_vma_size(vma));
225 		fence->start = i915_ggtt_offset(vma);
226 		fence->size = vma->fence_size;
227 		fence->stride = i915_gem_object_get_stride(vma->obj);
228 		fence->tiling = i915_gem_object_get_tiling(vma->obj);
229 	}
230 	WRITE_ONCE(fence->dirty, false);
231 
232 	old = xchg(&fence->vma, NULL);
233 	if (old) {
234 		/* XXX Ideally we would move the waiting to outside the mutex */
235 		ret = i915_active_wait(&fence->active);
236 		if (ret) {
237 			fence->vma = old;
238 			return ret;
239 		}
240 
241 		i915_vma_flush_writes(old);
242 
243 		/*
244 		 * Ensure that all userspace CPU access is completed before
245 		 * stealing the fence.
246 		 */
247 		if (old != vma) {
248 			GEM_BUG_ON(old->fence != fence);
249 			i915_vma_revoke_mmap(old);
250 			old->fence = NULL;
251 		}
252 
253 		list_move(&fence->link, &ggtt->fence_list);
254 	}
255 
256 	/*
257 	 * We only need to update the register itself if the device is awake.
258 	 * If the device is currently powered down, we will defer the write
259 	 * to the runtime resume, see intel_ggtt_restore_fences().
260 	 *
261 	 * This only works for removing the fence register, on acquisition
262 	 * the caller must hold the rpm wakeref. The fence register must
263 	 * be cleared before we can use any other fences to ensure that
264 	 * the new fences do not overlap the elided clears, confusing HW.
265 	 */
266 	wakeref = intel_runtime_pm_get_if_in_use(uncore->rpm);
267 	if (!wakeref) {
268 		GEM_BUG_ON(vma);
269 		return 0;
270 	}
271 
272 	WRITE_ONCE(fence->vma, vma);
273 	fence_write(fence);
274 
275 	if (vma) {
276 		vma->fence = fence;
277 		list_move_tail(&fence->link, &ggtt->fence_list);
278 	}
279 
280 	intel_runtime_pm_put(uncore->rpm, wakeref);
281 	return 0;
282 }
283 
284 /**
285  * i915_vma_revoke_fence - force-remove fence for a VMA
286  * @vma: vma to map linearly (not through a fence reg)
287  *
288  * This function force-removes any fence from the given object, which is useful
289  * if the kernel wants to do untiled GTT access.
290  */
291 void i915_vma_revoke_fence(struct i915_vma *vma)
292 {
293 	struct i915_fence_reg *fence = vma->fence;
294 	intel_wakeref_t wakeref;
295 
296 	lockdep_assert_held(&vma->vm->mutex);
297 	if (!fence)
298 		return;
299 
300 	GEM_BUG_ON(fence->vma != vma);
301 	i915_active_wait(&fence->active);
302 	GEM_BUG_ON(!i915_active_is_idle(&fence->active));
303 	GEM_BUG_ON(atomic_read(&fence->pin_count));
304 
305 	fence->tiling = 0;
306 	WRITE_ONCE(fence->vma, NULL);
307 	vma->fence = NULL;
308 
309 	/*
310 	 * Skip the write to HW if and only if the device is currently
311 	 * suspended.
312 	 *
313 	 * If the driver does not currently hold a wakeref (if_in_use == 0),
314 	 * the device may currently be runtime suspended, or it may be woken
315 	 * up before the suspend takes place. If the device is not suspended
316 	 * (powered down) and we skip clearing the fence register, the HW is
317 	 * left in an undefined state where we may end up with multiple
318 	 * registers overlapping.
319 	 */
320 	with_intel_runtime_pm_if_active(fence_to_uncore(fence)->rpm, wakeref)
321 		fence_write(fence);
322 }
323 
324 static bool fence_is_active(const struct i915_fence_reg *fence)
325 {
326 	return fence->vma && i915_vma_is_active(fence->vma);
327 }
328 
329 static struct i915_fence_reg *fence_find(struct i915_ggtt *ggtt)
330 {
331 	struct i915_fence_reg *active = NULL;
332 	struct i915_fence_reg *fence, *fn;
333 
334 	list_for_each_entry_safe(fence, fn, &ggtt->fence_list, link) {
335 		GEM_BUG_ON(fence->vma && fence->vma->fence != fence);
336 
337 		if (fence == active) /* now seen this fence twice */
338 			active = ERR_PTR(-EAGAIN);
339 
340 		/* Prefer idle fences so we do not have to wait on the GPU */
341 		if (active != ERR_PTR(-EAGAIN) && fence_is_active(fence)) {
342 			if (!active)
343 				active = fence;
344 
345 			list_move_tail(&fence->link, &ggtt->fence_list);
346 			continue;
347 		}
348 
349 		if (atomic_read(&fence->pin_count))
350 			continue;
351 
352 		return fence;
353 	}
354 
355 	/* Wait for completion of pending flips which consume fences */
356 	if (intel_has_pending_fb_unpin(ggtt->vm.i915))
357 		return ERR_PTR(-EAGAIN);
358 
359 	return ERR_PTR(-ENOBUFS);
360 }
361 
362 int __i915_vma_pin_fence(struct i915_vma *vma)
363 {
364 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vma->vm);
365 	struct i915_fence_reg *fence;
366 	struct i915_vma *set = i915_gem_object_is_tiled(vma->obj) ? vma : NULL;
367 	int err;
368 
369 	lockdep_assert_held(&vma->vm->mutex);
370 
371 	/* Just update our place in the LRU if our fence is getting reused. */
372 	if (vma->fence) {
373 		fence = vma->fence;
374 		GEM_BUG_ON(fence->vma != vma);
375 		atomic_inc(&fence->pin_count);
376 		if (!fence->dirty) {
377 			list_move_tail(&fence->link, &ggtt->fence_list);
378 			return 0;
379 		}
380 	} else if (set) {
381 		fence = fence_find(ggtt);
382 		if (IS_ERR(fence))
383 			return PTR_ERR(fence);
384 
385 		GEM_BUG_ON(atomic_read(&fence->pin_count));
386 		atomic_inc(&fence->pin_count);
387 	} else {
388 		return 0;
389 	}
390 
391 	err = fence_update(fence, set);
392 	if (err)
393 		goto out_unpin;
394 
395 	GEM_BUG_ON(fence->vma != set);
396 	GEM_BUG_ON(vma->fence != (set ? fence : NULL));
397 
398 	if (set)
399 		return 0;
400 
401 out_unpin:
402 	atomic_dec(&fence->pin_count);
403 	return err;
404 }
405 
406 /**
407  * i915_vma_pin_fence - set up fencing for a vma
408  * @vma: vma to map through a fence reg
409  *
410  * When mapping objects through the GTT, userspace wants to be able to write
411  * to them without having to worry about swizzling if the object is tiled.
412  * This function walks the fence regs looking for a free one for @obj,
413  * stealing one if it can't find any.
414  *
415  * It then sets up the reg based on the object's properties: address, pitch
416  * and tiling format.
417  *
418  * For an untiled surface, this removes any existing fence.
419  *
420  * Returns:
421  *
422  * 0 on success, negative error code on failure.
423  */
424 int i915_vma_pin_fence(struct i915_vma *vma)
425 {
426 	int err;
427 
428 	if (!vma->fence && !i915_gem_object_is_tiled(vma->obj))
429 		return 0;
430 
431 	/*
432 	 * Note that we revoke fences on runtime suspend. Therefore the user
433 	 * must keep the device awake whilst using the fence.
434 	 */
435 	assert_rpm_wakelock_held(vma->vm->gt->uncore->rpm);
436 	GEM_BUG_ON(!i915_vma_is_ggtt(vma));
437 
438 	err = mutex_lock_interruptible(&vma->vm->mutex);
439 	if (err)
440 		return err;
441 
442 	err = __i915_vma_pin_fence(vma);
443 	mutex_unlock(&vma->vm->mutex);
444 
445 	return err;
446 }
447 
448 /**
449  * i915_reserve_fence - Reserve a fence for vGPU
450  * @ggtt: Global GTT
451  *
452  * This function walks the fence regs looking for a free one and remove
453  * it from the fence_list. It is used to reserve fence for vGPU to use.
454  */
455 struct i915_fence_reg *i915_reserve_fence(struct i915_ggtt *ggtt)
456 {
457 	struct i915_fence_reg *fence;
458 	int count;
459 	int ret;
460 
461 	lockdep_assert_held(&ggtt->vm.mutex);
462 
463 	/* Keep at least one fence available for the display engine. */
464 	count = 0;
465 	list_for_each_entry(fence, &ggtt->fence_list, link)
466 		count += !atomic_read(&fence->pin_count);
467 	if (count <= 1)
468 		return ERR_PTR(-ENOSPC);
469 
470 	fence = fence_find(ggtt);
471 	if (IS_ERR(fence))
472 		return fence;
473 
474 	if (fence->vma) {
475 		/* Force-remove fence from VMA */
476 		ret = fence_update(fence, NULL);
477 		if (ret)
478 			return ERR_PTR(ret);
479 	}
480 
481 	list_del(&fence->link);
482 
483 	return fence;
484 }
485 
486 /**
487  * i915_unreserve_fence - Reclaim a reserved fence
488  * @fence: the fence reg
489  *
490  * This function add a reserved fence register from vGPU to the fence_list.
491  */
492 void i915_unreserve_fence(struct i915_fence_reg *fence)
493 {
494 	struct i915_ggtt *ggtt = fence->ggtt;
495 
496 	lockdep_assert_held(&ggtt->vm.mutex);
497 
498 	list_add(&fence->link, &ggtt->fence_list);
499 }
500 
501 /**
502  * intel_ggtt_restore_fences - restore fence state
503  * @ggtt: Global GTT
504  *
505  * Restore the hw fence state to match the software tracking again, to be called
506  * after a gpu reset and on resume. Note that on runtime suspend we only cancel
507  * the fences, to be reacquired by the user later.
508  */
509 void intel_ggtt_restore_fences(struct i915_ggtt *ggtt)
510 {
511 	int i;
512 
513 	for (i = 0; i < ggtt->num_fences; i++)
514 		fence_write(&ggtt->fence_regs[i]);
515 }
516 
517 /**
518  * DOC: tiling swizzling details
519  *
520  * The idea behind tiling is to increase cache hit rates by rearranging
521  * pixel data so that a group of pixel accesses are in the same cacheline.
522  * Performance improvement from doing this on the back/depth buffer are on
523  * the order of 30%.
524  *
525  * Intel architectures make this somewhat more complicated, though, by
526  * adjustments made to addressing of data when the memory is in interleaved
527  * mode (matched pairs of DIMMS) to improve memory bandwidth.
528  * For interleaved memory, the CPU sends every sequential 64 bytes
529  * to an alternate memory channel so it can get the bandwidth from both.
530  *
531  * The GPU also rearranges its accesses for increased bandwidth to interleaved
532  * memory, and it matches what the CPU does for non-tiled.  However, when tiled
533  * it does it a little differently, since one walks addresses not just in the
534  * X direction but also Y.  So, along with alternating channels when bit
535  * 6 of the address flips, it also alternates when other bits flip --  Bits 9
536  * (every 512 bytes, an X tile scanline) and 10 (every two X tile scanlines)
537  * are common to both the 915 and 965-class hardware.
538  *
539  * The CPU also sometimes XORs in higher bits as well, to improve
540  * bandwidth doing strided access like we do so frequently in graphics.  This
541  * is called "Channel XOR Randomization" in the MCH documentation.  The result
542  * is that the CPU is XORing in either bit 11 or bit 17 to bit 6 of its address
543  * decode.
544  *
545  * All of this bit 6 XORing has an effect on our memory management,
546  * as we need to make sure that the 3d driver can correctly address object
547  * contents.
548  *
549  * If we don't have interleaved memory, all tiling is safe and no swizzling is
550  * required.
551  *
552  * When bit 17 is XORed in, we simply refuse to tile at all.  Bit
553  * 17 is not just a page offset, so as we page an object out and back in,
554  * individual pages in it will have different bit 17 addresses, resulting in
555  * each 64 bytes being swapped with its neighbor!
556  *
557  * Otherwise, if interleaved, we have to tell the 3d driver what the address
558  * swizzling it needs to do is, since it's writing with the CPU to the pages
559  * (bit 6 and potentially bit 11 XORed in), and the GPU is reading from the
560  * pages (bit 6, 9, and 10 XORed in), resulting in a cumulative bit swizzling
561  * required by the CPU of XORing in bit 6, 9, 10, and potentially 11, in order
562  * to match what the GPU expects.
563  */
564 
565 /**
566  * detect_bit_6_swizzle - detect bit 6 swizzling pattern
567  * @ggtt: Global GGTT
568  *
569  * Detects bit 6 swizzling of address lookup between IGD access and CPU
570  * access through main memory.
571  */
572 static void detect_bit_6_swizzle(struct i915_ggtt *ggtt)
573 {
574 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
575 	struct drm_i915_private *i915 = ggtt->vm.i915;
576 	u32 swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
577 	u32 swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
578 
579 	if (GRAPHICS_VER(i915) >= 8 || IS_VALLEYVIEW(i915)) {
580 		/*
581 		 * On BDW+, swizzling is not used. We leave the CPU memory
582 		 * controller in charge of optimizing memory accesses without
583 		 * the extra address manipulation GPU side.
584 		 *
585 		 * VLV and CHV don't have GPU swizzling.
586 		 */
587 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
588 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
589 	} else if (GRAPHICS_VER(i915) >= 6) {
590 		if (i915->preserve_bios_swizzle) {
591 			if (intel_uncore_read(uncore, DISP_ARB_CTL) &
592 			    DISP_TILE_SURFACE_SWIZZLING) {
593 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
594 				swizzle_y = I915_BIT_6_SWIZZLE_9;
595 			} else {
596 				swizzle_x = I915_BIT_6_SWIZZLE_NONE;
597 				swizzle_y = I915_BIT_6_SWIZZLE_NONE;
598 			}
599 		} else {
600 			u32 dimm_c0, dimm_c1;
601 
602 			dimm_c0 = intel_uncore_read(uncore, MAD_DIMM_C0);
603 			dimm_c1 = intel_uncore_read(uncore, MAD_DIMM_C1);
604 			dimm_c0 &= MAD_DIMM_A_SIZE_MASK | MAD_DIMM_B_SIZE_MASK;
605 			dimm_c1 &= MAD_DIMM_A_SIZE_MASK | MAD_DIMM_B_SIZE_MASK;
606 			/*
607 			 * Enable swizzling when the channels are populated
608 			 * with identically sized dimms. We don't need to check
609 			 * the 3rd channel because no cpu with gpu attached
610 			 * ships in that configuration. Also, swizzling only
611 			 * makes sense for 2 channels anyway.
612 			 */
613 			if (dimm_c0 == dimm_c1) {
614 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
615 				swizzle_y = I915_BIT_6_SWIZZLE_9;
616 			} else {
617 				swizzle_x = I915_BIT_6_SWIZZLE_NONE;
618 				swizzle_y = I915_BIT_6_SWIZZLE_NONE;
619 			}
620 		}
621 	} else if (GRAPHICS_VER(i915) == 5) {
622 		/*
623 		 * On Ironlake whatever DRAM config, GPU always do
624 		 * same swizzling setup.
625 		 */
626 		swizzle_x = I915_BIT_6_SWIZZLE_9_10;
627 		swizzle_y = I915_BIT_6_SWIZZLE_9;
628 	} else if (GRAPHICS_VER(i915) == 2) {
629 		/*
630 		 * As far as we know, the 865 doesn't have these bit 6
631 		 * swizzling issues.
632 		 */
633 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
634 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
635 	} else if (IS_G45(i915) || IS_I965G(i915) || IS_G33(i915)) {
636 		/*
637 		 * The 965, G33, and newer, have a very flexible memory
638 		 * configuration.  It will enable dual-channel mode
639 		 * (interleaving) on as much memory as it can, and the GPU
640 		 * will additionally sometimes enable different bit 6
641 		 * swizzling for tiled objects from the CPU.
642 		 *
643 		 * Here's what I found on the G965:
644 		 *    slot fill         memory size  swizzling
645 		 * 0A   0B   1A   1B    1-ch   2-ch
646 		 * 512  0    0    0     512    0     O
647 		 * 512  0    512  0     16     1008  X
648 		 * 512  0    0    512   16     1008  X
649 		 * 0    512  0    512   16     1008  X
650 		 * 1024 1024 1024 0     2048   1024  O
651 		 *
652 		 * We could probably detect this based on either the DRB
653 		 * matching, which was the case for the swizzling required in
654 		 * the table above, or from the 1-ch value being less than
655 		 * the minimum size of a rank.
656 		 *
657 		 * Reports indicate that the swizzling actually
658 		 * varies depending upon page placement inside the
659 		 * channels, i.e. we see swizzled pages where the
660 		 * banks of memory are paired and unswizzled on the
661 		 * uneven portion, so leave that as unknown.
662 		 */
663 		if (intel_uncore_read16(uncore, C0DRB3_BW) ==
664 		    intel_uncore_read16(uncore, C1DRB3_BW)) {
665 			swizzle_x = I915_BIT_6_SWIZZLE_9_10;
666 			swizzle_y = I915_BIT_6_SWIZZLE_9;
667 		}
668 	} else {
669 		u32 dcc = intel_uncore_read(uncore, DCC);
670 
671 		/*
672 		 * On 9xx chipsets, channel interleave by the CPU is
673 		 * determined by DCC.  For single-channel, neither the CPU
674 		 * nor the GPU do swizzling.  For dual channel interleaved,
675 		 * the GPU's interleave is bit 9 and 10 for X tiled, and bit
676 		 * 9 for Y tiled.  The CPU's interleave is independent, and
677 		 * can be based on either bit 11 (haven't seen this yet) or
678 		 * bit 17 (common).
679 		 */
680 		switch (dcc & DCC_ADDRESSING_MODE_MASK) {
681 		case DCC_ADDRESSING_MODE_SINGLE_CHANNEL:
682 		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_ASYMMETRIC:
683 			swizzle_x = I915_BIT_6_SWIZZLE_NONE;
684 			swizzle_y = I915_BIT_6_SWIZZLE_NONE;
685 			break;
686 		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_INTERLEAVED:
687 			if (dcc & DCC_CHANNEL_XOR_DISABLE) {
688 				/*
689 				 * This is the base swizzling by the GPU for
690 				 * tiled buffers.
691 				 */
692 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
693 				swizzle_y = I915_BIT_6_SWIZZLE_9;
694 			} else if ((dcc & DCC_CHANNEL_XOR_BIT_17) == 0) {
695 				/* Bit 11 swizzling by the CPU in addition. */
696 				swizzle_x = I915_BIT_6_SWIZZLE_9_10_11;
697 				swizzle_y = I915_BIT_6_SWIZZLE_9_11;
698 			} else {
699 				/* Bit 17 swizzling by the CPU in addition. */
700 				swizzle_x = I915_BIT_6_SWIZZLE_9_10_17;
701 				swizzle_y = I915_BIT_6_SWIZZLE_9_17;
702 			}
703 			break;
704 		}
705 
706 		/* check for L-shaped memory aka modified enhanced addressing */
707 		if (GRAPHICS_VER(i915) == 4 &&
708 		    !(intel_uncore_read(uncore, DCC2) & DCC2_MODIFIED_ENHANCED_DISABLE)) {
709 			swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
710 			swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
711 		}
712 
713 		if (dcc == 0xffffffff) {
714 			drm_err(&i915->drm, "Couldn't read from MCHBAR.  "
715 				  "Disabling tiling.\n");
716 			swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
717 			swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
718 		}
719 	}
720 
721 	if (swizzle_x == I915_BIT_6_SWIZZLE_UNKNOWN ||
722 	    swizzle_y == I915_BIT_6_SWIZZLE_UNKNOWN) {
723 		/*
724 		 * Userspace likes to explode if it sees unknown swizzling,
725 		 * so lie. We will finish the lie when reporting through
726 		 * the get-tiling-ioctl by reporting the physical swizzle
727 		 * mode as unknown instead.
728 		 *
729 		 * As we don't strictly know what the swizzling is, it may be
730 		 * bit17 dependent, and so we need to also prevent the pages
731 		 * from being moved.
732 		 */
733 		i915->gem_quirks |= GEM_QUIRK_PIN_SWIZZLED_PAGES;
734 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
735 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
736 	}
737 
738 	to_gt(i915)->ggtt->bit_6_swizzle_x = swizzle_x;
739 	to_gt(i915)->ggtt->bit_6_swizzle_y = swizzle_y;
740 }
741 
742 /*
743  * Swap every 64 bytes of this page around, to account for it having a new
744  * bit 17 of its physical address and therefore being interpreted differently
745  * by the GPU.
746  */
747 static void swizzle_page(struct page *page)
748 {
749 	char temp[64];
750 	char *vaddr;
751 	int i;
752 
753 	vaddr = kmap(page);
754 
755 	for (i = 0; i < PAGE_SIZE; i += 128) {
756 		memcpy(temp, &vaddr[i], 64);
757 		memcpy(&vaddr[i], &vaddr[i + 64], 64);
758 		memcpy(&vaddr[i + 64], temp, 64);
759 	}
760 
761 	kunmap(page);
762 }
763 
764 /**
765  * i915_gem_object_do_bit_17_swizzle - fixup bit 17 swizzling
766  * @obj: i915 GEM buffer object
767  * @pages: the scattergather list of physical pages
768  *
769  * This function fixes up the swizzling in case any page frame number for this
770  * object has changed in bit 17 since that state has been saved with
771  * i915_gem_object_save_bit_17_swizzle().
772  *
773  * This is called when pinning backing storage again, since the kernel is free
774  * to move unpinned backing storage around (either by directly moving pages or
775  * by swapping them out and back in again).
776  */
777 void
778 i915_gem_object_do_bit_17_swizzle(struct drm_i915_gem_object *obj,
779 				  struct sg_table *pages)
780 {
781 	struct sgt_iter sgt_iter;
782 	struct page *page;
783 	int i;
784 
785 	if (obj->bit_17 == NULL)
786 		return;
787 
788 	i = 0;
789 	for_each_sgt_page(page, sgt_iter, pages) {
790 		char new_bit_17 = page_to_phys(page) >> 17;
791 
792 		if ((new_bit_17 & 0x1) != (test_bit(i, obj->bit_17) != 0)) {
793 			swizzle_page(page);
794 			set_page_dirty(page);
795 		}
796 
797 		i++;
798 	}
799 }
800 
801 /**
802  * i915_gem_object_save_bit_17_swizzle - save bit 17 swizzling
803  * @obj: i915 GEM buffer object
804  * @pages: the scattergather list of physical pages
805  *
806  * This function saves the bit 17 of each page frame number so that swizzling
807  * can be fixed up later on with i915_gem_object_do_bit_17_swizzle(). This must
808  * be called before the backing storage can be unpinned.
809  */
810 void
811 i915_gem_object_save_bit_17_swizzle(struct drm_i915_gem_object *obj,
812 				    struct sg_table *pages)
813 {
814 	const unsigned int page_count = obj->base.size >> PAGE_SHIFT;
815 	struct sgt_iter sgt_iter;
816 	struct page *page;
817 	int i;
818 
819 	if (obj->bit_17 == NULL) {
820 		obj->bit_17 = bitmap_zalloc(page_count, GFP_KERNEL);
821 		if (obj->bit_17 == NULL) {
822 			drm_err(obj->base.dev,
823 				"Failed to allocate memory for bit 17 record\n");
824 			return;
825 		}
826 	}
827 
828 	i = 0;
829 
830 	for_each_sgt_page(page, sgt_iter, pages) {
831 		if (page_to_phys(page) & (1 << 17))
832 			__set_bit(i, obj->bit_17);
833 		else
834 			__clear_bit(i, obj->bit_17);
835 		i++;
836 	}
837 }
838 
839 void intel_ggtt_init_fences(struct i915_ggtt *ggtt)
840 {
841 	struct drm_i915_private *i915 = ggtt->vm.i915;
842 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
843 	int num_fences;
844 	int i;
845 
846 	INIT_LIST_HEAD(&ggtt->fence_list);
847 	INIT_LIST_HEAD(&ggtt->userfault_list);
848 
849 	detect_bit_6_swizzle(ggtt);
850 
851 	if (!i915_ggtt_has_aperture(ggtt))
852 		num_fences = 0;
853 	else if (GRAPHICS_VER(i915) >= 7 &&
854 		 !(IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)))
855 		num_fences = 32;
856 	else if (GRAPHICS_VER(i915) >= 4 ||
857 		 IS_I945G(i915) || IS_I945GM(i915) ||
858 		 IS_G33(i915) || IS_PINEVIEW(i915))
859 		num_fences = 16;
860 	else
861 		num_fences = 8;
862 
863 	if (intel_vgpu_active(i915))
864 		num_fences = intel_uncore_read(uncore,
865 					       vgtif_reg(avail_rs.fence_num));
866 	ggtt->fence_regs = kcalloc(num_fences,
867 				   sizeof(*ggtt->fence_regs),
868 				   GFP_KERNEL);
869 	if (!ggtt->fence_regs)
870 		num_fences = 0;
871 
872 	/* Initialize fence registers to zero */
873 	for (i = 0; i < num_fences; i++) {
874 		struct i915_fence_reg *fence = &ggtt->fence_regs[i];
875 
876 		i915_active_init(&fence->active, NULL, NULL, 0);
877 		fence->ggtt = ggtt;
878 		fence->id = i;
879 		list_add_tail(&fence->link, &ggtt->fence_list);
880 	}
881 	ggtt->num_fences = num_fences;
882 
883 	intel_ggtt_restore_fences(ggtt);
884 }
885 
886 void intel_ggtt_fini_fences(struct i915_ggtt *ggtt)
887 {
888 	int i;
889 
890 	for (i = 0; i < ggtt->num_fences; i++) {
891 		struct i915_fence_reg *fence = &ggtt->fence_regs[i];
892 
893 		i915_active_fini(&fence->active);
894 	}
895 
896 	kfree(ggtt->fence_regs);
897 }
898 
899 void intel_gt_init_swizzling(struct intel_gt *gt)
900 {
901 	struct drm_i915_private *i915 = gt->i915;
902 	struct intel_uncore *uncore = gt->uncore;
903 
904 	if (GRAPHICS_VER(i915) < 5 ||
905 	    to_gt(i915)->ggtt->bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
906 		return;
907 
908 	intel_uncore_rmw(uncore, DISP_ARB_CTL, 0, DISP_TILE_SURFACE_SWIZZLING);
909 
910 	if (GRAPHICS_VER(i915) == 5)
911 		return;
912 
913 	intel_uncore_rmw(uncore, TILECTL, 0, TILECTL_SWZCTL);
914 
915 	if (GRAPHICS_VER(i915) == 6)
916 		intel_uncore_write(uncore,
917 				   ARB_MODE,
918 				   _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
919 	else if (GRAPHICS_VER(i915) == 7)
920 		intel_uncore_write(uncore,
921 				   ARB_MODE,
922 				   _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
923 	else if (GRAPHICS_VER(i915) == 8)
924 		intel_uncore_write(uncore,
925 				   GAMTARBMODE,
926 				   _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
927 	else
928 		MISSING_CASE(GRAPHICS_VER(i915));
929 }
930