xref: /linux/drivers/gpu/drm/i915/gt/intel_gtt.c (revision e7d759f31ca295d589f7420719c311870bb3166f)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2020 Intel Corporation
4  */
5 
6 #include <linux/slab.h> /* fault-inject.h is not standalone! */
7 
8 #include <linux/fault-inject.h>
9 #include <linux/sched/mm.h>
10 
11 #include <drm/drm_cache.h>
12 
13 #include "gem/i915_gem_internal.h"
14 #include "gem/i915_gem_lmem.h"
15 #include "i915_reg.h"
16 #include "i915_trace.h"
17 #include "i915_utils.h"
18 #include "intel_gt.h"
19 #include "intel_gt_mcr.h"
20 #include "intel_gt_print.h"
21 #include "intel_gt_regs.h"
22 #include "intel_gtt.h"
23 
24 bool i915_ggtt_require_binder(struct drm_i915_private *i915)
25 {
26 	/* Wa_13010847436 & Wa_14019519902 */
27 	return MEDIA_VER_FULL(i915) == IP_VER(13, 0);
28 }
29 
30 static bool intel_ggtt_update_needs_vtd_wa(struct drm_i915_private *i915)
31 {
32 	return IS_BROXTON(i915) && i915_vtd_active(i915);
33 }
34 
35 bool intel_vm_no_concurrent_access_wa(struct drm_i915_private *i915)
36 {
37 	return IS_CHERRYVIEW(i915) || intel_ggtt_update_needs_vtd_wa(i915);
38 }
39 
40 struct drm_i915_gem_object *alloc_pt_lmem(struct i915_address_space *vm, int sz)
41 {
42 	struct drm_i915_gem_object *obj;
43 
44 	/*
45 	 * To avoid severe over-allocation when dealing with min_page_size
46 	 * restrictions, we override that behaviour here by allowing an object
47 	 * size and page layout which can be smaller. In practice this should be
48 	 * totally fine, since GTT paging structures are not typically inserted
49 	 * into the GTT.
50 	 *
51 	 * Note that we also hit this path for the scratch page, and for this
52 	 * case it might need to be 64K, but that should work fine here since we
53 	 * used the passed in size for the page size, which should ensure it
54 	 * also has the same alignment.
55 	 */
56 	obj = __i915_gem_object_create_lmem_with_ps(vm->i915, sz, sz,
57 						    vm->lmem_pt_obj_flags);
58 	/*
59 	 * Ensure all paging structures for this vm share the same dma-resv
60 	 * object underneath, with the idea that one object_lock() will lock
61 	 * them all at once.
62 	 */
63 	if (!IS_ERR(obj)) {
64 		obj->base.resv = i915_vm_resv_get(vm);
65 		obj->shares_resv_from = vm;
66 
67 		if (vm->fpriv)
68 			i915_drm_client_add_object(vm->fpriv->client, obj);
69 	}
70 
71 	return obj;
72 }
73 
74 struct drm_i915_gem_object *alloc_pt_dma(struct i915_address_space *vm, int sz)
75 {
76 	struct drm_i915_gem_object *obj;
77 
78 	if (I915_SELFTEST_ONLY(should_fail(&vm->fault_attr, 1)))
79 		i915_gem_shrink_all(vm->i915);
80 
81 	obj = i915_gem_object_create_internal(vm->i915, sz);
82 	/*
83 	 * Ensure all paging structures for this vm share the same dma-resv
84 	 * object underneath, with the idea that one object_lock() will lock
85 	 * them all at once.
86 	 */
87 	if (!IS_ERR(obj)) {
88 		obj->base.resv = i915_vm_resv_get(vm);
89 		obj->shares_resv_from = vm;
90 
91 		if (vm->fpriv)
92 			i915_drm_client_add_object(vm->fpriv->client, obj);
93 	}
94 
95 	return obj;
96 }
97 
98 int map_pt_dma(struct i915_address_space *vm, struct drm_i915_gem_object *obj)
99 {
100 	enum i915_map_type type;
101 	void *vaddr;
102 
103 	type = intel_gt_coherent_map_type(vm->gt, obj, true);
104 	/*
105 	 * FIXME: It is suspected that some Address Translation Service (ATS)
106 	 * issue on IOMMU is causing CAT errors to occur on some MTL workloads.
107 	 * Applying a write barrier to the ppgtt set entry functions appeared
108 	 * to have no effect, so we must temporarily use I915_MAP_WC here on
109 	 * MTL until a proper ATS solution is found.
110 	 */
111 	if (IS_METEORLAKE(vm->i915))
112 		type = I915_MAP_WC;
113 
114 	vaddr = i915_gem_object_pin_map_unlocked(obj, type);
115 	if (IS_ERR(vaddr))
116 		return PTR_ERR(vaddr);
117 
118 	i915_gem_object_make_unshrinkable(obj);
119 	return 0;
120 }
121 
122 int map_pt_dma_locked(struct i915_address_space *vm, struct drm_i915_gem_object *obj)
123 {
124 	enum i915_map_type type;
125 	void *vaddr;
126 
127 	type = intel_gt_coherent_map_type(vm->gt, obj, true);
128 	/*
129 	 * FIXME: It is suspected that some Address Translation Service (ATS)
130 	 * issue on IOMMU is causing CAT errors to occur on some MTL workloads.
131 	 * Applying a write barrier to the ppgtt set entry functions appeared
132 	 * to have no effect, so we must temporarily use I915_MAP_WC here on
133 	 * MTL until a proper ATS solution is found.
134 	 */
135 	if (IS_METEORLAKE(vm->i915))
136 		type = I915_MAP_WC;
137 
138 	vaddr = i915_gem_object_pin_map(obj, type);
139 	if (IS_ERR(vaddr))
140 		return PTR_ERR(vaddr);
141 
142 	i915_gem_object_make_unshrinkable(obj);
143 	return 0;
144 }
145 
146 static void clear_vm_list(struct list_head *list)
147 {
148 	struct i915_vma *vma, *vn;
149 
150 	list_for_each_entry_safe(vma, vn, list, vm_link) {
151 		struct drm_i915_gem_object *obj = vma->obj;
152 
153 		if (!i915_gem_object_get_rcu(obj)) {
154 			/*
155 			 * Object is dying, but has not yet cleared its
156 			 * vma list.
157 			 * Unbind the dying vma to ensure our list
158 			 * is completely drained. We leave the destruction to
159 			 * the object destructor to avoid the vma
160 			 * disappearing under it.
161 			 */
162 			atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
163 			WARN_ON(__i915_vma_unbind(vma));
164 
165 			/* Remove from the unbound list */
166 			list_del_init(&vma->vm_link);
167 
168 			/*
169 			 * Delay the vm and vm mutex freeing until the
170 			 * object is done with destruction.
171 			 */
172 			i915_vm_resv_get(vma->vm);
173 			vma->vm_ddestroy = true;
174 		} else {
175 			i915_vma_destroy_locked(vma);
176 			i915_gem_object_put(obj);
177 		}
178 
179 	}
180 }
181 
182 static void __i915_vm_close(struct i915_address_space *vm)
183 {
184 	mutex_lock(&vm->mutex);
185 
186 	clear_vm_list(&vm->bound_list);
187 	clear_vm_list(&vm->unbound_list);
188 
189 	/* Check for must-fix unanticipated side-effects */
190 	GEM_BUG_ON(!list_empty(&vm->bound_list));
191 	GEM_BUG_ON(!list_empty(&vm->unbound_list));
192 
193 	mutex_unlock(&vm->mutex);
194 }
195 
196 /* lock the vm into the current ww, if we lock one, we lock all */
197 int i915_vm_lock_objects(struct i915_address_space *vm,
198 			 struct i915_gem_ww_ctx *ww)
199 {
200 	if (vm->scratch[0]->base.resv == &vm->_resv) {
201 		return i915_gem_object_lock(vm->scratch[0], ww);
202 	} else {
203 		struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
204 
205 		/* We borrowed the scratch page from ggtt, take the top level object */
206 		return i915_gem_object_lock(ppgtt->pd->pt.base, ww);
207 	}
208 }
209 
210 void i915_address_space_fini(struct i915_address_space *vm)
211 {
212 	drm_mm_takedown(&vm->mm);
213 }
214 
215 /**
216  * i915_vm_resv_release - Final struct i915_address_space destructor
217  * @kref: Pointer to the &i915_address_space.resv_ref member.
218  *
219  * This function is called when the last lock sharer no longer shares the
220  * &i915_address_space._resv lock, and also if we raced when
221  * destroying a vma by the vma destruction
222  */
223 void i915_vm_resv_release(struct kref *kref)
224 {
225 	struct i915_address_space *vm =
226 		container_of(kref, typeof(*vm), resv_ref);
227 
228 	dma_resv_fini(&vm->_resv);
229 	mutex_destroy(&vm->mutex);
230 
231 	kfree(vm);
232 }
233 
234 static void __i915_vm_release(struct work_struct *work)
235 {
236 	struct i915_address_space *vm =
237 		container_of(work, struct i915_address_space, release_work);
238 
239 	__i915_vm_close(vm);
240 
241 	/* Synchronize async unbinds. */
242 	i915_vma_resource_bind_dep_sync_all(vm);
243 
244 	vm->cleanup(vm);
245 	i915_address_space_fini(vm);
246 
247 	i915_vm_resv_put(vm);
248 }
249 
250 void i915_vm_release(struct kref *kref)
251 {
252 	struct i915_address_space *vm =
253 		container_of(kref, struct i915_address_space, ref);
254 
255 	GEM_BUG_ON(i915_is_ggtt(vm));
256 	trace_i915_ppgtt_release(vm);
257 
258 	queue_work(vm->i915->wq, &vm->release_work);
259 }
260 
261 void i915_address_space_init(struct i915_address_space *vm, int subclass)
262 {
263 	kref_init(&vm->ref);
264 
265 	/*
266 	 * Special case for GGTT that has already done an early
267 	 * kref_init here.
268 	 */
269 	if (!kref_read(&vm->resv_ref))
270 		kref_init(&vm->resv_ref);
271 
272 	vm->pending_unbind = RB_ROOT_CACHED;
273 	INIT_WORK(&vm->release_work, __i915_vm_release);
274 
275 	/*
276 	 * The vm->mutex must be reclaim safe (for use in the shrinker).
277 	 * Do a dummy acquire now under fs_reclaim so that any allocation
278 	 * attempt holding the lock is immediately reported by lockdep.
279 	 */
280 	mutex_init(&vm->mutex);
281 	lockdep_set_subclass(&vm->mutex, subclass);
282 
283 	if (!intel_vm_no_concurrent_access_wa(vm->i915)) {
284 		i915_gem_shrinker_taints_mutex(vm->i915, &vm->mutex);
285 	} else {
286 		/*
287 		 * CHV + BXT VTD workaround use stop_machine(),
288 		 * which is allowed to allocate memory. This means &vm->mutex
289 		 * is the outer lock, and in theory we can allocate memory inside
290 		 * it through stop_machine().
291 		 *
292 		 * Add the annotation for this, we use trylock in shrinker.
293 		 */
294 		mutex_acquire(&vm->mutex.dep_map, 0, 0, _THIS_IP_);
295 		might_alloc(GFP_KERNEL);
296 		mutex_release(&vm->mutex.dep_map, _THIS_IP_);
297 	}
298 	dma_resv_init(&vm->_resv);
299 
300 	GEM_BUG_ON(!vm->total);
301 	drm_mm_init(&vm->mm, 0, vm->total);
302 
303 	memset64(vm->min_alignment, I915_GTT_MIN_ALIGNMENT,
304 		 ARRAY_SIZE(vm->min_alignment));
305 
306 	if (HAS_64K_PAGES(vm->i915)) {
307 		vm->min_alignment[INTEL_MEMORY_LOCAL] = I915_GTT_PAGE_SIZE_64K;
308 		vm->min_alignment[INTEL_MEMORY_STOLEN_LOCAL] = I915_GTT_PAGE_SIZE_64K;
309 	}
310 
311 	vm->mm.head_node.color = I915_COLOR_UNEVICTABLE;
312 
313 	INIT_LIST_HEAD(&vm->bound_list);
314 	INIT_LIST_HEAD(&vm->unbound_list);
315 }
316 
317 void *__px_vaddr(struct drm_i915_gem_object *p)
318 {
319 	enum i915_map_type type;
320 
321 	GEM_BUG_ON(!i915_gem_object_has_pages(p));
322 	return page_unpack_bits(p->mm.mapping, &type);
323 }
324 
325 dma_addr_t __px_dma(struct drm_i915_gem_object *p)
326 {
327 	GEM_BUG_ON(!i915_gem_object_has_pages(p));
328 	return sg_dma_address(p->mm.pages->sgl);
329 }
330 
331 struct page *__px_page(struct drm_i915_gem_object *p)
332 {
333 	GEM_BUG_ON(!i915_gem_object_has_pages(p));
334 	return sg_page(p->mm.pages->sgl);
335 }
336 
337 void
338 fill_page_dma(struct drm_i915_gem_object *p, const u64 val, unsigned int count)
339 {
340 	void *vaddr = __px_vaddr(p);
341 
342 	memset64(vaddr, val, count);
343 	drm_clflush_virt_range(vaddr, PAGE_SIZE);
344 }
345 
346 static void poison_scratch_page(struct drm_i915_gem_object *scratch)
347 {
348 	void *vaddr = __px_vaddr(scratch);
349 	u8 val;
350 
351 	val = 0;
352 	if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
353 		val = POISON_FREE;
354 
355 	memset(vaddr, val, scratch->base.size);
356 	drm_clflush_virt_range(vaddr, scratch->base.size);
357 }
358 
359 int setup_scratch_page(struct i915_address_space *vm)
360 {
361 	unsigned long size;
362 
363 	/*
364 	 * In order to utilize 64K pages for an object with a size < 2M, we will
365 	 * need to support a 64K scratch page, given that every 16th entry for a
366 	 * page-table operating in 64K mode must point to a properly aligned 64K
367 	 * region, including any PTEs which happen to point to scratch.
368 	 *
369 	 * This is only relevant for the 48b PPGTT where we support
370 	 * huge-gtt-pages, see also i915_vma_insert(). However, as we share the
371 	 * scratch (read-only) between all vm, we create one 64k scratch page
372 	 * for all.
373 	 */
374 	size = I915_GTT_PAGE_SIZE_4K;
375 	if (i915_vm_is_4lvl(vm) &&
376 	    HAS_PAGE_SIZES(vm->i915, I915_GTT_PAGE_SIZE_64K) &&
377 	    !HAS_64K_PAGES(vm->i915))
378 		size = I915_GTT_PAGE_SIZE_64K;
379 
380 	do {
381 		struct drm_i915_gem_object *obj;
382 
383 		obj = vm->alloc_scratch_dma(vm, size);
384 		if (IS_ERR(obj))
385 			goto skip;
386 
387 		if (map_pt_dma(vm, obj))
388 			goto skip_obj;
389 
390 		/* We need a single contiguous page for our scratch */
391 		if (obj->mm.page_sizes.sg < size)
392 			goto skip_obj;
393 
394 		/* And it needs to be correspondingly aligned */
395 		if (__px_dma(obj) & (size - 1))
396 			goto skip_obj;
397 
398 		/*
399 		 * Use a non-zero scratch page for debugging.
400 		 *
401 		 * We want a value that should be reasonably obvious
402 		 * to spot in the error state, while also causing a GPU hang
403 		 * if executed. We prefer using a clear page in production, so
404 		 * should it ever be accidentally used, the effect should be
405 		 * fairly benign.
406 		 */
407 		poison_scratch_page(obj);
408 
409 		vm->scratch[0] = obj;
410 		vm->scratch_order = get_order(size);
411 		return 0;
412 
413 skip_obj:
414 		i915_gem_object_put(obj);
415 skip:
416 		if (size == I915_GTT_PAGE_SIZE_4K)
417 			return -ENOMEM;
418 
419 		size = I915_GTT_PAGE_SIZE_4K;
420 	} while (1);
421 }
422 
423 void free_scratch(struct i915_address_space *vm)
424 {
425 	int i;
426 
427 	if (!vm->scratch[0])
428 		return;
429 
430 	for (i = 0; i <= vm->top; i++)
431 		i915_gem_object_put(vm->scratch[i]);
432 }
433 
434 void gtt_write_workarounds(struct intel_gt *gt)
435 {
436 	struct drm_i915_private *i915 = gt->i915;
437 	struct intel_uncore *uncore = gt->uncore;
438 
439 	/*
440 	 * This function is for gtt related workarounds. This function is
441 	 * called on driver load and after a GPU reset, so you can place
442 	 * workarounds here even if they get overwritten by GPU reset.
443 	 */
444 	/* WaIncreaseDefaultTLBEntries:chv,bdw,skl,bxt,kbl,glk,cfl,cnl,icl */
445 	if (IS_BROADWELL(i915))
446 		intel_uncore_write(uncore,
447 				   GEN8_L3_LRA_1_GPGPU,
448 				   GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_BDW);
449 	else if (IS_CHERRYVIEW(i915))
450 		intel_uncore_write(uncore,
451 				   GEN8_L3_LRA_1_GPGPU,
452 				   GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_CHV);
453 	else if (IS_GEN9_LP(i915))
454 		intel_uncore_write(uncore,
455 				   GEN8_L3_LRA_1_GPGPU,
456 				   GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_BXT);
457 	else if (GRAPHICS_VER(i915) >= 9 && GRAPHICS_VER(i915) <= 11)
458 		intel_uncore_write(uncore,
459 				   GEN8_L3_LRA_1_GPGPU,
460 				   GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_SKL);
461 
462 	/*
463 	 * To support 64K PTEs we need to first enable the use of the
464 	 * Intermediate-Page-Size(IPS) bit of the PDE field via some magical
465 	 * mmio, otherwise the page-walker will simply ignore the IPS bit. This
466 	 * shouldn't be needed after GEN10.
467 	 *
468 	 * 64K pages were first introduced from BDW+, although technically they
469 	 * only *work* from gen9+. For pre-BDW we instead have the option for
470 	 * 32K pages, but we don't currently have any support for it in our
471 	 * driver.
472 	 */
473 	if (HAS_PAGE_SIZES(i915, I915_GTT_PAGE_SIZE_64K) &&
474 	    GRAPHICS_VER(i915) <= 10)
475 		intel_uncore_rmw(uncore,
476 				 GEN8_GAMW_ECO_DEV_RW_IA,
477 				 0,
478 				 GAMW_ECO_ENABLE_64K_IPS_FIELD);
479 
480 	if (IS_GRAPHICS_VER(i915, 8, 11)) {
481 		bool can_use_gtt_cache = true;
482 
483 		/*
484 		 * According to the BSpec if we use 2M/1G pages then we also
485 		 * need to disable the GTT cache. At least on BDW we can see
486 		 * visual corruption when using 2M pages, and not disabling the
487 		 * GTT cache.
488 		 */
489 		if (HAS_PAGE_SIZES(i915, I915_GTT_PAGE_SIZE_2M))
490 			can_use_gtt_cache = false;
491 
492 		/* WaGttCachingOffByDefault */
493 		intel_uncore_write(uncore,
494 				   HSW_GTT_CACHE_EN,
495 				   can_use_gtt_cache ? GTT_CACHE_EN_ALL : 0);
496 		gt_WARN_ON_ONCE(gt, can_use_gtt_cache &&
497 				intel_uncore_read(uncore,
498 						  HSW_GTT_CACHE_EN) == 0);
499 	}
500 }
501 
502 static void xelpmp_setup_private_ppat(struct intel_uncore *uncore)
503 {
504 	intel_uncore_write(uncore, XELPMP_PAT_INDEX(0),
505 			   MTL_PPAT_L4_0_WB);
506 	intel_uncore_write(uncore, XELPMP_PAT_INDEX(1),
507 			   MTL_PPAT_L4_1_WT);
508 	intel_uncore_write(uncore, XELPMP_PAT_INDEX(2),
509 			   MTL_PPAT_L4_3_UC);
510 	intel_uncore_write(uncore, XELPMP_PAT_INDEX(3),
511 			   MTL_PPAT_L4_0_WB | MTL_2_COH_1W);
512 	intel_uncore_write(uncore, XELPMP_PAT_INDEX(4),
513 			   MTL_PPAT_L4_0_WB | MTL_3_COH_2W);
514 
515 	/*
516 	 * Remaining PAT entries are left at the hardware-default
517 	 * fully-cached setting
518 	 */
519 }
520 
521 static void xelpg_setup_private_ppat(struct intel_gt *gt)
522 {
523 	intel_gt_mcr_multicast_write(gt, XEHP_PAT_INDEX(0),
524 				     MTL_PPAT_L4_0_WB);
525 	intel_gt_mcr_multicast_write(gt, XEHP_PAT_INDEX(1),
526 				     MTL_PPAT_L4_1_WT);
527 	intel_gt_mcr_multicast_write(gt, XEHP_PAT_INDEX(2),
528 				     MTL_PPAT_L4_3_UC);
529 	intel_gt_mcr_multicast_write(gt, XEHP_PAT_INDEX(3),
530 				     MTL_PPAT_L4_0_WB | MTL_2_COH_1W);
531 	intel_gt_mcr_multicast_write(gt, XEHP_PAT_INDEX(4),
532 				     MTL_PPAT_L4_0_WB | MTL_3_COH_2W);
533 
534 	/*
535 	 * Remaining PAT entries are left at the hardware-default
536 	 * fully-cached setting
537 	 */
538 }
539 
540 static void tgl_setup_private_ppat(struct intel_uncore *uncore)
541 {
542 	/* TGL doesn't support LLC or AGE settings */
543 	intel_uncore_write(uncore, GEN12_PAT_INDEX(0), GEN8_PPAT_WB);
544 	intel_uncore_write(uncore, GEN12_PAT_INDEX(1), GEN8_PPAT_WC);
545 	intel_uncore_write(uncore, GEN12_PAT_INDEX(2), GEN8_PPAT_WT);
546 	intel_uncore_write(uncore, GEN12_PAT_INDEX(3), GEN8_PPAT_UC);
547 	intel_uncore_write(uncore, GEN12_PAT_INDEX(4), GEN8_PPAT_WB);
548 	intel_uncore_write(uncore, GEN12_PAT_INDEX(5), GEN8_PPAT_WB);
549 	intel_uncore_write(uncore, GEN12_PAT_INDEX(6), GEN8_PPAT_WB);
550 	intel_uncore_write(uncore, GEN12_PAT_INDEX(7), GEN8_PPAT_WB);
551 }
552 
553 static void xehp_setup_private_ppat(struct intel_gt *gt)
554 {
555 	enum forcewake_domains fw;
556 	unsigned long flags;
557 
558 	fw = intel_uncore_forcewake_for_reg(gt->uncore, _MMIO(XEHP_PAT_INDEX(0).reg),
559 					    FW_REG_WRITE);
560 	intel_uncore_forcewake_get(gt->uncore, fw);
561 
562 	intel_gt_mcr_lock(gt, &flags);
563 	intel_gt_mcr_multicast_write_fw(gt, XEHP_PAT_INDEX(0), GEN8_PPAT_WB);
564 	intel_gt_mcr_multicast_write_fw(gt, XEHP_PAT_INDEX(1), GEN8_PPAT_WC);
565 	intel_gt_mcr_multicast_write_fw(gt, XEHP_PAT_INDEX(2), GEN8_PPAT_WT);
566 	intel_gt_mcr_multicast_write_fw(gt, XEHP_PAT_INDEX(3), GEN8_PPAT_UC);
567 	intel_gt_mcr_multicast_write_fw(gt, XEHP_PAT_INDEX(4), GEN8_PPAT_WB);
568 	intel_gt_mcr_multicast_write_fw(gt, XEHP_PAT_INDEX(5), GEN8_PPAT_WB);
569 	intel_gt_mcr_multicast_write_fw(gt, XEHP_PAT_INDEX(6), GEN8_PPAT_WB);
570 	intel_gt_mcr_multicast_write_fw(gt, XEHP_PAT_INDEX(7), GEN8_PPAT_WB);
571 	intel_gt_mcr_unlock(gt, flags);
572 
573 	intel_uncore_forcewake_put(gt->uncore, fw);
574 }
575 
576 static void icl_setup_private_ppat(struct intel_uncore *uncore)
577 {
578 	intel_uncore_write(uncore,
579 			   GEN10_PAT_INDEX(0),
580 			   GEN8_PPAT_WB | GEN8_PPAT_LLC);
581 	intel_uncore_write(uncore,
582 			   GEN10_PAT_INDEX(1),
583 			   GEN8_PPAT_WC | GEN8_PPAT_LLCELLC);
584 	intel_uncore_write(uncore,
585 			   GEN10_PAT_INDEX(2),
586 			   GEN8_PPAT_WB | GEN8_PPAT_ELLC_OVERRIDE);
587 	intel_uncore_write(uncore,
588 			   GEN10_PAT_INDEX(3),
589 			   GEN8_PPAT_UC);
590 	intel_uncore_write(uncore,
591 			   GEN10_PAT_INDEX(4),
592 			   GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0));
593 	intel_uncore_write(uncore,
594 			   GEN10_PAT_INDEX(5),
595 			   GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1));
596 	intel_uncore_write(uncore,
597 			   GEN10_PAT_INDEX(6),
598 			   GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2));
599 	intel_uncore_write(uncore,
600 			   GEN10_PAT_INDEX(7),
601 			   GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
602 }
603 
604 /*
605  * The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
606  * bits. When using advanced contexts each context stores its own PAT, but
607  * writing this data shouldn't be harmful even in those cases.
608  */
609 static void bdw_setup_private_ppat(struct intel_uncore *uncore)
610 {
611 	struct drm_i915_private *i915 = uncore->i915;
612 	u64 pat;
613 
614 	pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) |	/* for normal objects, no eLLC */
615 	      GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) |	/* for something pointing to ptes? */
616 	      GEN8_PPAT(3, GEN8_PPAT_UC) |			/* Uncached objects, mostly for scanout */
617 	      GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) |
618 	      GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) |
619 	      GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) |
620 	      GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
621 
622 	/* for scanout with eLLC */
623 	if (GRAPHICS_VER(i915) >= 9)
624 		pat |= GEN8_PPAT(2, GEN8_PPAT_WB | GEN8_PPAT_ELLC_OVERRIDE);
625 	else
626 		pat |= GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC);
627 
628 	intel_uncore_write(uncore, GEN8_PRIVATE_PAT_LO, lower_32_bits(pat));
629 	intel_uncore_write(uncore, GEN8_PRIVATE_PAT_HI, upper_32_bits(pat));
630 }
631 
632 static void chv_setup_private_ppat(struct intel_uncore *uncore)
633 {
634 	u64 pat;
635 
636 	/*
637 	 * Map WB on BDW to snooped on CHV.
638 	 *
639 	 * Only the snoop bit has meaning for CHV, the rest is
640 	 * ignored.
641 	 *
642 	 * The hardware will never snoop for certain types of accesses:
643 	 * - CPU GTT (GMADR->GGTT->no snoop->memory)
644 	 * - PPGTT page tables
645 	 * - some other special cycles
646 	 *
647 	 * As with BDW, we also need to consider the following for GT accesses:
648 	 * "For GGTT, there is NO pat_sel[2:0] from the entry,
649 	 * so RTL will always use the value corresponding to
650 	 * pat_sel = 000".
651 	 * Which means we must set the snoop bit in PAT entry 0
652 	 * in order to keep the global status page working.
653 	 */
654 
655 	pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) |
656 	      GEN8_PPAT(1, 0) |
657 	      GEN8_PPAT(2, 0) |
658 	      GEN8_PPAT(3, 0) |
659 	      GEN8_PPAT(4, CHV_PPAT_SNOOP) |
660 	      GEN8_PPAT(5, CHV_PPAT_SNOOP) |
661 	      GEN8_PPAT(6, CHV_PPAT_SNOOP) |
662 	      GEN8_PPAT(7, CHV_PPAT_SNOOP);
663 
664 	intel_uncore_write(uncore, GEN8_PRIVATE_PAT_LO, lower_32_bits(pat));
665 	intel_uncore_write(uncore, GEN8_PRIVATE_PAT_HI, upper_32_bits(pat));
666 }
667 
668 void setup_private_pat(struct intel_gt *gt)
669 {
670 	struct intel_uncore *uncore = gt->uncore;
671 	struct drm_i915_private *i915 = gt->i915;
672 
673 	GEM_BUG_ON(GRAPHICS_VER(i915) < 8);
674 
675 	if (gt->type == GT_MEDIA) {
676 		xelpmp_setup_private_ppat(gt->uncore);
677 		return;
678 	}
679 
680 	if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 70))
681 		xelpg_setup_private_ppat(gt);
682 	else if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50))
683 		xehp_setup_private_ppat(gt);
684 	else if (GRAPHICS_VER(i915) >= 12)
685 		tgl_setup_private_ppat(uncore);
686 	else if (GRAPHICS_VER(i915) >= 11)
687 		icl_setup_private_ppat(uncore);
688 	else if (IS_CHERRYVIEW(i915) || IS_GEN9_LP(i915))
689 		chv_setup_private_ppat(uncore);
690 	else
691 		bdw_setup_private_ppat(uncore);
692 }
693 
694 struct i915_vma *
695 __vm_create_scratch_for_read(struct i915_address_space *vm, unsigned long size)
696 {
697 	struct drm_i915_gem_object *obj;
698 	struct i915_vma *vma;
699 
700 	obj = i915_gem_object_create_internal(vm->i915, PAGE_ALIGN(size));
701 	if (IS_ERR(obj))
702 		return ERR_CAST(obj);
703 
704 	i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
705 
706 	vma = i915_vma_instance(obj, vm, NULL);
707 	if (IS_ERR(vma)) {
708 		i915_gem_object_put(obj);
709 		return vma;
710 	}
711 
712 	return vma;
713 }
714 
715 struct i915_vma *
716 __vm_create_scratch_for_read_pinned(struct i915_address_space *vm, unsigned long size)
717 {
718 	struct i915_vma *vma;
719 	int err;
720 
721 	vma = __vm_create_scratch_for_read(vm, size);
722 	if (IS_ERR(vma))
723 		return vma;
724 
725 	err = i915_vma_pin(vma, 0, 0,
726 			   i915_vma_is_ggtt(vma) ? PIN_GLOBAL : PIN_USER);
727 	if (err) {
728 		i915_vma_put(vma);
729 		return ERR_PTR(err);
730 	}
731 
732 	return vma;
733 }
734 
735 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
736 #include "selftests/mock_gtt.c"
737 #endif
738