1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2008-2015 Intel Corporation
4  */
5 
6 #include <linux/highmem.h>
7 
8 #include "i915_drv.h"
9 #include "i915_reg.h"
10 #include "i915_scatterlist.h"
11 #include "i915_pvinfo.h"
12 #include "i915_vgpu.h"
13 #include "intel_gt_regs.h"
14 #include "intel_mchbar_regs.h"
15 
16 /**
17  * DOC: fence register handling
18  *
19  * Important to avoid confusions: "fences" in the i915 driver are not execution
20  * fences used to track command completion but hardware detiler objects which
21  * wrap a given range of the global GTT. Each platform has only a fairly limited
22  * set of these objects.
23  *
24  * Fences are used to detile GTT memory mappings. They're also connected to the
25  * hardware frontbuffer render tracking and hence interact with frontbuffer
26  * compression. Furthermore on older platforms fences are required for tiled
27  * objects used by the display engine. They can also be used by the render
28  * engine - they're required for blitter commands and are optional for render
29  * commands. But on gen4+ both display (with the exception of fbc) and rendering
30  * have their own tiling state bits and don't need fences.
31  *
32  * Also note that fences only support X and Y tiling and hence can't be used for
33  * the fancier new tiling formats like W, Ys and Yf.
34  *
35  * Finally note that because fences are such a restricted resource they're
36  * dynamically associated with objects. Furthermore fence state is committed to
37  * the hardware lazily to avoid unnecessary stalls on gen2/3. Therefore code must
38  * explicitly call i915_gem_object_get_fence() to synchronize fencing status
39  * for cpu access. Also note that some code wants an unfenced view, for those
40  * cases the fence can be removed forcefully with i915_gem_object_put_fence().
41  *
42  * Internally these functions will synchronize with userspace access by removing
43  * CPU ptes into GTT mmaps (not the GTT ptes themselves) as needed.
44  */
45 
46 #define pipelined 0
47 
48 static struct drm_i915_private *fence_to_i915(struct i915_fence_reg *fence)
49 {
50 	return fence->ggtt->vm.i915;
51 }
52 
53 static struct intel_uncore *fence_to_uncore(struct i915_fence_reg *fence)
54 {
55 	return fence->ggtt->vm.gt->uncore;
56 }
57 
58 static void i965_write_fence_reg(struct i915_fence_reg *fence)
59 {
60 	i915_reg_t fence_reg_lo, fence_reg_hi;
61 	int fence_pitch_shift;
62 	u64 val;
63 
64 	if (GRAPHICS_VER(fence_to_i915(fence)) >= 6) {
65 		fence_reg_lo = FENCE_REG_GEN6_LO(fence->id);
66 		fence_reg_hi = FENCE_REG_GEN6_HI(fence->id);
67 		fence_pitch_shift = GEN6_FENCE_PITCH_SHIFT;
68 
69 	} else {
70 		fence_reg_lo = FENCE_REG_965_LO(fence->id);
71 		fence_reg_hi = FENCE_REG_965_HI(fence->id);
72 		fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
73 	}
74 
75 	val = 0;
76 	if (fence->tiling) {
77 		unsigned int stride = fence->stride;
78 
79 		GEM_BUG_ON(!IS_ALIGNED(stride, 128));
80 
81 		val = fence->start + fence->size - I965_FENCE_PAGE;
82 		val <<= 32;
83 		val |= fence->start;
84 		val |= (u64)((stride / 128) - 1) << fence_pitch_shift;
85 		if (fence->tiling == I915_TILING_Y)
86 			val |= BIT(I965_FENCE_TILING_Y_SHIFT);
87 		val |= I965_FENCE_REG_VALID;
88 	}
89 
90 	if (!pipelined) {
91 		struct intel_uncore *uncore = fence_to_uncore(fence);
92 
93 		/*
94 		 * To w/a incoherency with non-atomic 64-bit register updates,
95 		 * we split the 64-bit update into two 32-bit writes. In order
96 		 * for a partial fence not to be evaluated between writes, we
97 		 * precede the update with write to turn off the fence register,
98 		 * and only enable the fence as the last step.
99 		 *
100 		 * For extra levels of paranoia, we make sure each step lands
101 		 * before applying the next step.
102 		 */
103 		intel_uncore_write_fw(uncore, fence_reg_lo, 0);
104 		intel_uncore_posting_read_fw(uncore, fence_reg_lo);
105 
106 		intel_uncore_write_fw(uncore, fence_reg_hi, upper_32_bits(val));
107 		intel_uncore_write_fw(uncore, fence_reg_lo, lower_32_bits(val));
108 		intel_uncore_posting_read_fw(uncore, fence_reg_lo);
109 	}
110 }
111 
112 static void i915_write_fence_reg(struct i915_fence_reg *fence)
113 {
114 	u32 val;
115 
116 	val = 0;
117 	if (fence->tiling) {
118 		unsigned int stride = fence->stride;
119 		unsigned int tiling = fence->tiling;
120 		bool is_y_tiled = tiling == I915_TILING_Y;
121 
122 		if (is_y_tiled && HAS_128_BYTE_Y_TILING(fence_to_i915(fence)))
123 			stride /= 128;
124 		else
125 			stride /= 512;
126 		GEM_BUG_ON(!is_power_of_2(stride));
127 
128 		val = fence->start;
129 		if (is_y_tiled)
130 			val |= BIT(I830_FENCE_TILING_Y_SHIFT);
131 		val |= I915_FENCE_SIZE_BITS(fence->size);
132 		val |= ilog2(stride) << I830_FENCE_PITCH_SHIFT;
133 
134 		val |= I830_FENCE_REG_VALID;
135 	}
136 
137 	if (!pipelined) {
138 		struct intel_uncore *uncore = fence_to_uncore(fence);
139 		i915_reg_t reg = FENCE_REG(fence->id);
140 
141 		intel_uncore_write_fw(uncore, reg, val);
142 		intel_uncore_posting_read_fw(uncore, reg);
143 	}
144 }
145 
146 static void i830_write_fence_reg(struct i915_fence_reg *fence)
147 {
148 	u32 val;
149 
150 	val = 0;
151 	if (fence->tiling) {
152 		unsigned int stride = fence->stride;
153 
154 		val = fence->start;
155 		if (fence->tiling == I915_TILING_Y)
156 			val |= BIT(I830_FENCE_TILING_Y_SHIFT);
157 		val |= I830_FENCE_SIZE_BITS(fence->size);
158 		val |= ilog2(stride / 128) << I830_FENCE_PITCH_SHIFT;
159 		val |= I830_FENCE_REG_VALID;
160 	}
161 
162 	if (!pipelined) {
163 		struct intel_uncore *uncore = fence_to_uncore(fence);
164 		i915_reg_t reg = FENCE_REG(fence->id);
165 
166 		intel_uncore_write_fw(uncore, reg, val);
167 		intel_uncore_posting_read_fw(uncore, reg);
168 	}
169 }
170 
171 static void fence_write(struct i915_fence_reg *fence)
172 {
173 	struct drm_i915_private *i915 = fence_to_i915(fence);
174 
175 	/*
176 	 * Previous access through the fence register is marshalled by
177 	 * the mb() inside the fault handlers (i915_gem_release_mmaps)
178 	 * and explicitly managed for internal users.
179 	 */
180 
181 	if (GRAPHICS_VER(i915) == 2)
182 		i830_write_fence_reg(fence);
183 	else if (GRAPHICS_VER(i915) == 3)
184 		i915_write_fence_reg(fence);
185 	else
186 		i965_write_fence_reg(fence);
187 
188 	/*
189 	 * Access through the fenced region afterwards is
190 	 * ordered by the posting reads whilst writing the registers.
191 	 */
192 }
193 
194 static bool gpu_uses_fence_registers(struct i915_fence_reg *fence)
195 {
196 	return GRAPHICS_VER(fence_to_i915(fence)) < 4;
197 }
198 
199 static int fence_update(struct i915_fence_reg *fence,
200 			struct i915_vma *vma)
201 {
202 	struct i915_ggtt *ggtt = fence->ggtt;
203 	struct intel_uncore *uncore = fence_to_uncore(fence);
204 	intel_wakeref_t wakeref;
205 	struct i915_vma *old;
206 	int ret;
207 
208 	fence->tiling = 0;
209 	if (vma) {
210 		GEM_BUG_ON(!i915_gem_object_get_stride(vma->obj) ||
211 			   !i915_gem_object_get_tiling(vma->obj));
212 
213 		if (!i915_vma_is_map_and_fenceable(vma))
214 			return -EINVAL;
215 
216 		if (gpu_uses_fence_registers(fence)) {
217 			/* implicit 'unfenced' GPU blits */
218 			ret = i915_vma_sync(vma);
219 			if (ret)
220 				return ret;
221 		}
222 
223 		GEM_BUG_ON(vma->fence_size > i915_vma_size(vma));
224 		fence->start = i915_ggtt_offset(vma);
225 		fence->size = vma->fence_size;
226 		fence->stride = i915_gem_object_get_stride(vma->obj);
227 		fence->tiling = i915_gem_object_get_tiling(vma->obj);
228 	}
229 	WRITE_ONCE(fence->dirty, false);
230 
231 	old = xchg(&fence->vma, NULL);
232 	if (old) {
233 		/* XXX Ideally we would move the waiting to outside the mutex */
234 		ret = i915_active_wait(&fence->active);
235 		if (ret) {
236 			fence->vma = old;
237 			return ret;
238 		}
239 
240 		i915_vma_flush_writes(old);
241 
242 		/*
243 		 * Ensure that all userspace CPU access is completed before
244 		 * stealing the fence.
245 		 */
246 		if (old != vma) {
247 			GEM_BUG_ON(old->fence != fence);
248 			i915_vma_revoke_mmap(old);
249 			old->fence = NULL;
250 		}
251 
252 		list_move(&fence->link, &ggtt->fence_list);
253 	}
254 
255 	/*
256 	 * We only need to update the register itself if the device is awake.
257 	 * If the device is currently powered down, we will defer the write
258 	 * to the runtime resume, see intel_ggtt_restore_fences().
259 	 *
260 	 * This only works for removing the fence register, on acquisition
261 	 * the caller must hold the rpm wakeref. The fence register must
262 	 * be cleared before we can use any other fences to ensure that
263 	 * the new fences do not overlap the elided clears, confusing HW.
264 	 */
265 	wakeref = intel_runtime_pm_get_if_in_use(uncore->rpm);
266 	if (!wakeref) {
267 		GEM_BUG_ON(vma);
268 		return 0;
269 	}
270 
271 	WRITE_ONCE(fence->vma, vma);
272 	fence_write(fence);
273 
274 	if (vma) {
275 		vma->fence = fence;
276 		list_move_tail(&fence->link, &ggtt->fence_list);
277 	}
278 
279 	intel_runtime_pm_put(uncore->rpm, wakeref);
280 	return 0;
281 }
282 
283 /**
284  * i915_vma_revoke_fence - force-remove fence for a VMA
285  * @vma: vma to map linearly (not through a fence reg)
286  *
287  * This function force-removes any fence from the given object, which is useful
288  * if the kernel wants to do untiled GTT access.
289  */
290 void i915_vma_revoke_fence(struct i915_vma *vma)
291 {
292 	struct i915_fence_reg *fence = vma->fence;
293 	intel_wakeref_t wakeref;
294 
295 	lockdep_assert_held(&vma->vm->mutex);
296 	if (!fence)
297 		return;
298 
299 	GEM_BUG_ON(fence->vma != vma);
300 	GEM_BUG_ON(!i915_active_is_idle(&fence->active));
301 	GEM_BUG_ON(atomic_read(&fence->pin_count));
302 
303 	fence->tiling = 0;
304 	WRITE_ONCE(fence->vma, NULL);
305 	vma->fence = NULL;
306 
307 	/*
308 	 * Skip the write to HW if and only if the device is currently
309 	 * suspended.
310 	 *
311 	 * If the driver does not currently hold a wakeref (if_in_use == 0),
312 	 * the device may currently be runtime suspended, or it may be woken
313 	 * up before the suspend takes place. If the device is not suspended
314 	 * (powered down) and we skip clearing the fence register, the HW is
315 	 * left in an undefined state where we may end up with multiple
316 	 * registers overlapping.
317 	 */
318 	with_intel_runtime_pm_if_active(fence_to_uncore(fence)->rpm, wakeref)
319 		fence_write(fence);
320 }
321 
322 static bool fence_is_active(const struct i915_fence_reg *fence)
323 {
324 	return fence->vma && i915_vma_is_active(fence->vma);
325 }
326 
327 static struct i915_fence_reg *fence_find(struct i915_ggtt *ggtt)
328 {
329 	struct i915_fence_reg *active = NULL;
330 	struct i915_fence_reg *fence, *fn;
331 
332 	list_for_each_entry_safe(fence, fn, &ggtt->fence_list, link) {
333 		GEM_BUG_ON(fence->vma && fence->vma->fence != fence);
334 
335 		if (fence == active) /* now seen this fence twice */
336 			active = ERR_PTR(-EAGAIN);
337 
338 		/* Prefer idle fences so we do not have to wait on the GPU */
339 		if (active != ERR_PTR(-EAGAIN) && fence_is_active(fence)) {
340 			if (!active)
341 				active = fence;
342 
343 			list_move_tail(&fence->link, &ggtt->fence_list);
344 			continue;
345 		}
346 
347 		if (atomic_read(&fence->pin_count))
348 			continue;
349 
350 		return fence;
351 	}
352 
353 	/* Wait for completion of pending flips which consume fences */
354 	if (intel_has_pending_fb_unpin(ggtt->vm.i915))
355 		return ERR_PTR(-EAGAIN);
356 
357 	return ERR_PTR(-ENOBUFS);
358 }
359 
360 int __i915_vma_pin_fence(struct i915_vma *vma)
361 {
362 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vma->vm);
363 	struct i915_fence_reg *fence;
364 	struct i915_vma *set = i915_gem_object_is_tiled(vma->obj) ? vma : NULL;
365 	int err;
366 
367 	lockdep_assert_held(&vma->vm->mutex);
368 
369 	/* Just update our place in the LRU if our fence is getting reused. */
370 	if (vma->fence) {
371 		fence = vma->fence;
372 		GEM_BUG_ON(fence->vma != vma);
373 		atomic_inc(&fence->pin_count);
374 		if (!fence->dirty) {
375 			list_move_tail(&fence->link, &ggtt->fence_list);
376 			return 0;
377 		}
378 	} else if (set) {
379 		fence = fence_find(ggtt);
380 		if (IS_ERR(fence))
381 			return PTR_ERR(fence);
382 
383 		GEM_BUG_ON(atomic_read(&fence->pin_count));
384 		atomic_inc(&fence->pin_count);
385 	} else {
386 		return 0;
387 	}
388 
389 	err = fence_update(fence, set);
390 	if (err)
391 		goto out_unpin;
392 
393 	GEM_BUG_ON(fence->vma != set);
394 	GEM_BUG_ON(vma->fence != (set ? fence : NULL));
395 
396 	if (set)
397 		return 0;
398 
399 out_unpin:
400 	atomic_dec(&fence->pin_count);
401 	return err;
402 }
403 
404 /**
405  * i915_vma_pin_fence - set up fencing for a vma
406  * @vma: vma to map through a fence reg
407  *
408  * When mapping objects through the GTT, userspace wants to be able to write
409  * to them without having to worry about swizzling if the object is tiled.
410  * This function walks the fence regs looking for a free one for @obj,
411  * stealing one if it can't find any.
412  *
413  * It then sets up the reg based on the object's properties: address, pitch
414  * and tiling format.
415  *
416  * For an untiled surface, this removes any existing fence.
417  *
418  * Returns:
419  *
420  * 0 on success, negative error code on failure.
421  */
422 int i915_vma_pin_fence(struct i915_vma *vma)
423 {
424 	int err;
425 
426 	if (!vma->fence && !i915_gem_object_is_tiled(vma->obj))
427 		return 0;
428 
429 	/*
430 	 * Note that we revoke fences on runtime suspend. Therefore the user
431 	 * must keep the device awake whilst using the fence.
432 	 */
433 	assert_rpm_wakelock_held(vma->vm->gt->uncore->rpm);
434 	GEM_BUG_ON(!i915_vma_is_ggtt(vma));
435 
436 	err = mutex_lock_interruptible(&vma->vm->mutex);
437 	if (err)
438 		return err;
439 
440 	err = __i915_vma_pin_fence(vma);
441 	mutex_unlock(&vma->vm->mutex);
442 
443 	return err;
444 }
445 
446 /**
447  * i915_reserve_fence - Reserve a fence for vGPU
448  * @ggtt: Global GTT
449  *
450  * This function walks the fence regs looking for a free one and remove
451  * it from the fence_list. It is used to reserve fence for vGPU to use.
452  */
453 struct i915_fence_reg *i915_reserve_fence(struct i915_ggtt *ggtt)
454 {
455 	struct i915_fence_reg *fence;
456 	int count;
457 	int ret;
458 
459 	lockdep_assert_held(&ggtt->vm.mutex);
460 
461 	/* Keep at least one fence available for the display engine. */
462 	count = 0;
463 	list_for_each_entry(fence, &ggtt->fence_list, link)
464 		count += !atomic_read(&fence->pin_count);
465 	if (count <= 1)
466 		return ERR_PTR(-ENOSPC);
467 
468 	fence = fence_find(ggtt);
469 	if (IS_ERR(fence))
470 		return fence;
471 
472 	if (fence->vma) {
473 		/* Force-remove fence from VMA */
474 		ret = fence_update(fence, NULL);
475 		if (ret)
476 			return ERR_PTR(ret);
477 	}
478 
479 	list_del(&fence->link);
480 
481 	return fence;
482 }
483 
484 /**
485  * i915_unreserve_fence - Reclaim a reserved fence
486  * @fence: the fence reg
487  *
488  * This function add a reserved fence register from vGPU to the fence_list.
489  */
490 void i915_unreserve_fence(struct i915_fence_reg *fence)
491 {
492 	struct i915_ggtt *ggtt = fence->ggtt;
493 
494 	lockdep_assert_held(&ggtt->vm.mutex);
495 
496 	list_add(&fence->link, &ggtt->fence_list);
497 }
498 
499 /**
500  * intel_ggtt_restore_fences - restore fence state
501  * @ggtt: Global GTT
502  *
503  * Restore the hw fence state to match the software tracking again, to be called
504  * after a gpu reset and on resume. Note that on runtime suspend we only cancel
505  * the fences, to be reacquired by the user later.
506  */
507 void intel_ggtt_restore_fences(struct i915_ggtt *ggtt)
508 {
509 	int i;
510 
511 	for (i = 0; i < ggtt->num_fences; i++)
512 		fence_write(&ggtt->fence_regs[i]);
513 }
514 
515 /**
516  * DOC: tiling swizzling details
517  *
518  * The idea behind tiling is to increase cache hit rates by rearranging
519  * pixel data so that a group of pixel accesses are in the same cacheline.
520  * Performance improvement from doing this on the back/depth buffer are on
521  * the order of 30%.
522  *
523  * Intel architectures make this somewhat more complicated, though, by
524  * adjustments made to addressing of data when the memory is in interleaved
525  * mode (matched pairs of DIMMS) to improve memory bandwidth.
526  * For interleaved memory, the CPU sends every sequential 64 bytes
527  * to an alternate memory channel so it can get the bandwidth from both.
528  *
529  * The GPU also rearranges its accesses for increased bandwidth to interleaved
530  * memory, and it matches what the CPU does for non-tiled.  However, when tiled
531  * it does it a little differently, since one walks addresses not just in the
532  * X direction but also Y.  So, along with alternating channels when bit
533  * 6 of the address flips, it also alternates when other bits flip --  Bits 9
534  * (every 512 bytes, an X tile scanline) and 10 (every two X tile scanlines)
535  * are common to both the 915 and 965-class hardware.
536  *
537  * The CPU also sometimes XORs in higher bits as well, to improve
538  * bandwidth doing strided access like we do so frequently in graphics.  This
539  * is called "Channel XOR Randomization" in the MCH documentation.  The result
540  * is that the CPU is XORing in either bit 11 or bit 17 to bit 6 of its address
541  * decode.
542  *
543  * All of this bit 6 XORing has an effect on our memory management,
544  * as we need to make sure that the 3d driver can correctly address object
545  * contents.
546  *
547  * If we don't have interleaved memory, all tiling is safe and no swizzling is
548  * required.
549  *
550  * When bit 17 is XORed in, we simply refuse to tile at all.  Bit
551  * 17 is not just a page offset, so as we page an object out and back in,
552  * individual pages in it will have different bit 17 addresses, resulting in
553  * each 64 bytes being swapped with its neighbor!
554  *
555  * Otherwise, if interleaved, we have to tell the 3d driver what the address
556  * swizzling it needs to do is, since it's writing with the CPU to the pages
557  * (bit 6 and potentially bit 11 XORed in), and the GPU is reading from the
558  * pages (bit 6, 9, and 10 XORed in), resulting in a cumulative bit swizzling
559  * required by the CPU of XORing in bit 6, 9, 10, and potentially 11, in order
560  * to match what the GPU expects.
561  */
562 
563 /**
564  * detect_bit_6_swizzle - detect bit 6 swizzling pattern
565  * @ggtt: Global GGTT
566  *
567  * Detects bit 6 swizzling of address lookup between IGD access and CPU
568  * access through main memory.
569  */
570 static void detect_bit_6_swizzle(struct i915_ggtt *ggtt)
571 {
572 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
573 	struct drm_i915_private *i915 = ggtt->vm.i915;
574 	u32 swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
575 	u32 swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
576 
577 	if (GRAPHICS_VER(i915) >= 8 || IS_VALLEYVIEW(i915)) {
578 		/*
579 		 * On BDW+, swizzling is not used. We leave the CPU memory
580 		 * controller in charge of optimizing memory accesses without
581 		 * the extra address manipulation GPU side.
582 		 *
583 		 * VLV and CHV don't have GPU swizzling.
584 		 */
585 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
586 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
587 	} else if (GRAPHICS_VER(i915) >= 6) {
588 		if (i915->preserve_bios_swizzle) {
589 			if (intel_uncore_read(uncore, DISP_ARB_CTL) &
590 			    DISP_TILE_SURFACE_SWIZZLING) {
591 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
592 				swizzle_y = I915_BIT_6_SWIZZLE_9;
593 			} else {
594 				swizzle_x = I915_BIT_6_SWIZZLE_NONE;
595 				swizzle_y = I915_BIT_6_SWIZZLE_NONE;
596 			}
597 		} else {
598 			u32 dimm_c0, dimm_c1;
599 
600 			dimm_c0 = intel_uncore_read(uncore, MAD_DIMM_C0);
601 			dimm_c1 = intel_uncore_read(uncore, MAD_DIMM_C1);
602 			dimm_c0 &= MAD_DIMM_A_SIZE_MASK | MAD_DIMM_B_SIZE_MASK;
603 			dimm_c1 &= MAD_DIMM_A_SIZE_MASK | MAD_DIMM_B_SIZE_MASK;
604 			/*
605 			 * Enable swizzling when the channels are populated
606 			 * with identically sized dimms. We don't need to check
607 			 * the 3rd channel because no cpu with gpu attached
608 			 * ships in that configuration. Also, swizzling only
609 			 * makes sense for 2 channels anyway.
610 			 */
611 			if (dimm_c0 == dimm_c1) {
612 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
613 				swizzle_y = I915_BIT_6_SWIZZLE_9;
614 			} else {
615 				swizzle_x = I915_BIT_6_SWIZZLE_NONE;
616 				swizzle_y = I915_BIT_6_SWIZZLE_NONE;
617 			}
618 		}
619 	} else if (GRAPHICS_VER(i915) == 5) {
620 		/*
621 		 * On Ironlake whatever DRAM config, GPU always do
622 		 * same swizzling setup.
623 		 */
624 		swizzle_x = I915_BIT_6_SWIZZLE_9_10;
625 		swizzle_y = I915_BIT_6_SWIZZLE_9;
626 	} else if (GRAPHICS_VER(i915) == 2) {
627 		/*
628 		 * As far as we know, the 865 doesn't have these bit 6
629 		 * swizzling issues.
630 		 */
631 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
632 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
633 	} else if (IS_G45(i915) || IS_I965G(i915) || IS_G33(i915)) {
634 		/*
635 		 * The 965, G33, and newer, have a very flexible memory
636 		 * configuration.  It will enable dual-channel mode
637 		 * (interleaving) on as much memory as it can, and the GPU
638 		 * will additionally sometimes enable different bit 6
639 		 * swizzling for tiled objects from the CPU.
640 		 *
641 		 * Here's what I found on the G965:
642 		 *    slot fill         memory size  swizzling
643 		 * 0A   0B   1A   1B    1-ch   2-ch
644 		 * 512  0    0    0     512    0     O
645 		 * 512  0    512  0     16     1008  X
646 		 * 512  0    0    512   16     1008  X
647 		 * 0    512  0    512   16     1008  X
648 		 * 1024 1024 1024 0     2048   1024  O
649 		 *
650 		 * We could probably detect this based on either the DRB
651 		 * matching, which was the case for the swizzling required in
652 		 * the table above, or from the 1-ch value being less than
653 		 * the minimum size of a rank.
654 		 *
655 		 * Reports indicate that the swizzling actually
656 		 * varies depending upon page placement inside the
657 		 * channels, i.e. we see swizzled pages where the
658 		 * banks of memory are paired and unswizzled on the
659 		 * uneven portion, so leave that as unknown.
660 		 */
661 		if (intel_uncore_read16(uncore, C0DRB3_BW) ==
662 		    intel_uncore_read16(uncore, C1DRB3_BW)) {
663 			swizzle_x = I915_BIT_6_SWIZZLE_9_10;
664 			swizzle_y = I915_BIT_6_SWIZZLE_9;
665 		}
666 	} else {
667 		u32 dcc = intel_uncore_read(uncore, DCC);
668 
669 		/*
670 		 * On 9xx chipsets, channel interleave by the CPU is
671 		 * determined by DCC.  For single-channel, neither the CPU
672 		 * nor the GPU do swizzling.  For dual channel interleaved,
673 		 * the GPU's interleave is bit 9 and 10 for X tiled, and bit
674 		 * 9 for Y tiled.  The CPU's interleave is independent, and
675 		 * can be based on either bit 11 (haven't seen this yet) or
676 		 * bit 17 (common).
677 		 */
678 		switch (dcc & DCC_ADDRESSING_MODE_MASK) {
679 		case DCC_ADDRESSING_MODE_SINGLE_CHANNEL:
680 		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_ASYMMETRIC:
681 			swizzle_x = I915_BIT_6_SWIZZLE_NONE;
682 			swizzle_y = I915_BIT_6_SWIZZLE_NONE;
683 			break;
684 		case DCC_ADDRESSING_MODE_DUAL_CHANNEL_INTERLEAVED:
685 			if (dcc & DCC_CHANNEL_XOR_DISABLE) {
686 				/*
687 				 * This is the base swizzling by the GPU for
688 				 * tiled buffers.
689 				 */
690 				swizzle_x = I915_BIT_6_SWIZZLE_9_10;
691 				swizzle_y = I915_BIT_6_SWIZZLE_9;
692 			} else if ((dcc & DCC_CHANNEL_XOR_BIT_17) == 0) {
693 				/* Bit 11 swizzling by the CPU in addition. */
694 				swizzle_x = I915_BIT_6_SWIZZLE_9_10_11;
695 				swizzle_y = I915_BIT_6_SWIZZLE_9_11;
696 			} else {
697 				/* Bit 17 swizzling by the CPU in addition. */
698 				swizzle_x = I915_BIT_6_SWIZZLE_9_10_17;
699 				swizzle_y = I915_BIT_6_SWIZZLE_9_17;
700 			}
701 			break;
702 		}
703 
704 		/* check for L-shaped memory aka modified enhanced addressing */
705 		if (GRAPHICS_VER(i915) == 4 &&
706 		    !(intel_uncore_read(uncore, DCC2) & DCC2_MODIFIED_ENHANCED_DISABLE)) {
707 			swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
708 			swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
709 		}
710 
711 		if (dcc == 0xffffffff) {
712 			drm_err(&i915->drm, "Couldn't read from MCHBAR.  "
713 				  "Disabling tiling.\n");
714 			swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
715 			swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
716 		}
717 	}
718 
719 	if (swizzle_x == I915_BIT_6_SWIZZLE_UNKNOWN ||
720 	    swizzle_y == I915_BIT_6_SWIZZLE_UNKNOWN) {
721 		/*
722 		 * Userspace likes to explode if it sees unknown swizzling,
723 		 * so lie. We will finish the lie when reporting through
724 		 * the get-tiling-ioctl by reporting the physical swizzle
725 		 * mode as unknown instead.
726 		 *
727 		 * As we don't strictly know what the swizzling is, it may be
728 		 * bit17 dependent, and so we need to also prevent the pages
729 		 * from being moved.
730 		 */
731 		i915->gem_quirks |= GEM_QUIRK_PIN_SWIZZLED_PAGES;
732 		swizzle_x = I915_BIT_6_SWIZZLE_NONE;
733 		swizzle_y = I915_BIT_6_SWIZZLE_NONE;
734 	}
735 
736 	to_gt(i915)->ggtt->bit_6_swizzle_x = swizzle_x;
737 	to_gt(i915)->ggtt->bit_6_swizzle_y = swizzle_y;
738 }
739 
740 /*
741  * Swap every 64 bytes of this page around, to account for it having a new
742  * bit 17 of its physical address and therefore being interpreted differently
743  * by the GPU.
744  */
745 static void swizzle_page(struct page *page)
746 {
747 	char temp[64];
748 	char *vaddr;
749 	int i;
750 
751 	vaddr = kmap(page);
752 
753 	for (i = 0; i < PAGE_SIZE; i += 128) {
754 		memcpy(temp, &vaddr[i], 64);
755 		memcpy(&vaddr[i], &vaddr[i + 64], 64);
756 		memcpy(&vaddr[i + 64], temp, 64);
757 	}
758 
759 	kunmap(page);
760 }
761 
762 /**
763  * i915_gem_object_do_bit_17_swizzle - fixup bit 17 swizzling
764  * @obj: i915 GEM buffer object
765  * @pages: the scattergather list of physical pages
766  *
767  * This function fixes up the swizzling in case any page frame number for this
768  * object has changed in bit 17 since that state has been saved with
769  * i915_gem_object_save_bit_17_swizzle().
770  *
771  * This is called when pinning backing storage again, since the kernel is free
772  * to move unpinned backing storage around (either by directly moving pages or
773  * by swapping them out and back in again).
774  */
775 void
776 i915_gem_object_do_bit_17_swizzle(struct drm_i915_gem_object *obj,
777 				  struct sg_table *pages)
778 {
779 	struct sgt_iter sgt_iter;
780 	struct page *page;
781 	int i;
782 
783 	if (obj->bit_17 == NULL)
784 		return;
785 
786 	i = 0;
787 	for_each_sgt_page(page, sgt_iter, pages) {
788 		char new_bit_17 = page_to_phys(page) >> 17;
789 
790 		if ((new_bit_17 & 0x1) != (test_bit(i, obj->bit_17) != 0)) {
791 			swizzle_page(page);
792 			set_page_dirty(page);
793 		}
794 
795 		i++;
796 	}
797 }
798 
799 /**
800  * i915_gem_object_save_bit_17_swizzle - save bit 17 swizzling
801  * @obj: i915 GEM buffer object
802  * @pages: the scattergather list of physical pages
803  *
804  * This function saves the bit 17 of each page frame number so that swizzling
805  * can be fixed up later on with i915_gem_object_do_bit_17_swizzle(). This must
806  * be called before the backing storage can be unpinned.
807  */
808 void
809 i915_gem_object_save_bit_17_swizzle(struct drm_i915_gem_object *obj,
810 				    struct sg_table *pages)
811 {
812 	const unsigned int page_count = obj->base.size >> PAGE_SHIFT;
813 	struct sgt_iter sgt_iter;
814 	struct page *page;
815 	int i;
816 
817 	if (obj->bit_17 == NULL) {
818 		obj->bit_17 = bitmap_zalloc(page_count, GFP_KERNEL);
819 		if (obj->bit_17 == NULL) {
820 			drm_err(&to_i915(obj->base.dev)->drm,
821 				"Failed to allocate memory for bit 17 record\n");
822 			return;
823 		}
824 	}
825 
826 	i = 0;
827 
828 	for_each_sgt_page(page, sgt_iter, pages) {
829 		if (page_to_phys(page) & (1 << 17))
830 			__set_bit(i, obj->bit_17);
831 		else
832 			__clear_bit(i, obj->bit_17);
833 		i++;
834 	}
835 }
836 
837 void intel_ggtt_init_fences(struct i915_ggtt *ggtt)
838 {
839 	struct drm_i915_private *i915 = ggtt->vm.i915;
840 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
841 	int num_fences;
842 	int i;
843 
844 	INIT_LIST_HEAD(&ggtt->fence_list);
845 	INIT_LIST_HEAD(&ggtt->userfault_list);
846 
847 	detect_bit_6_swizzle(ggtt);
848 
849 	if (!i915_ggtt_has_aperture(ggtt))
850 		num_fences = 0;
851 	else if (GRAPHICS_VER(i915) >= 7 &&
852 		 !(IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)))
853 		num_fences = 32;
854 	else if (GRAPHICS_VER(i915) >= 4 ||
855 		 IS_I945G(i915) || IS_I945GM(i915) ||
856 		 IS_G33(i915) || IS_PINEVIEW(i915))
857 		num_fences = 16;
858 	else
859 		num_fences = 8;
860 
861 	if (intel_vgpu_active(i915))
862 		num_fences = intel_uncore_read(uncore,
863 					       vgtif_reg(avail_rs.fence_num));
864 	ggtt->fence_regs = kcalloc(num_fences,
865 				   sizeof(*ggtt->fence_regs),
866 				   GFP_KERNEL);
867 	if (!ggtt->fence_regs)
868 		num_fences = 0;
869 
870 	/* Initialize fence registers to zero */
871 	for (i = 0; i < num_fences; i++) {
872 		struct i915_fence_reg *fence = &ggtt->fence_regs[i];
873 
874 		i915_active_init(&fence->active, NULL, NULL, 0);
875 		fence->ggtt = ggtt;
876 		fence->id = i;
877 		list_add_tail(&fence->link, &ggtt->fence_list);
878 	}
879 	ggtt->num_fences = num_fences;
880 
881 	intel_ggtt_restore_fences(ggtt);
882 }
883 
884 void intel_ggtt_fini_fences(struct i915_ggtt *ggtt)
885 {
886 	int i;
887 
888 	for (i = 0; i < ggtt->num_fences; i++) {
889 		struct i915_fence_reg *fence = &ggtt->fence_regs[i];
890 
891 		i915_active_fini(&fence->active);
892 	}
893 
894 	kfree(ggtt->fence_regs);
895 }
896 
897 void intel_gt_init_swizzling(struct intel_gt *gt)
898 {
899 	struct drm_i915_private *i915 = gt->i915;
900 	struct intel_uncore *uncore = gt->uncore;
901 
902 	if (GRAPHICS_VER(i915) < 5 ||
903 	    to_gt(i915)->ggtt->bit_6_swizzle_x == I915_BIT_6_SWIZZLE_NONE)
904 		return;
905 
906 	intel_uncore_rmw(uncore, DISP_ARB_CTL, 0, DISP_TILE_SURFACE_SWIZZLING);
907 
908 	if (GRAPHICS_VER(i915) == 5)
909 		return;
910 
911 	intel_uncore_rmw(uncore, TILECTL, 0, TILECTL_SWZCTL);
912 
913 	if (GRAPHICS_VER(i915) == 6)
914 		intel_uncore_write(uncore,
915 				   ARB_MODE,
916 				   _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB));
917 	else if (GRAPHICS_VER(i915) == 7)
918 		intel_uncore_write(uncore,
919 				   ARB_MODE,
920 				   _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB));
921 	else if (GRAPHICS_VER(i915) == 8)
922 		intel_uncore_write(uncore,
923 				   GAMTARBMODE,
924 				   _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW));
925 	else
926 		MISSING_CASE(GRAPHICS_VER(i915));
927 }
928