1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2020 Intel Corporation
4  */
5 
6 #include <linux/types.h>
7 #include <asm/set_memory.h>
8 #include <asm/smp.h>
9 
10 #include <drm/i915_drm.h>
11 
12 #include "gem/i915_gem_lmem.h"
13 
14 #include "intel_gt.h"
15 #include "intel_gt_gmch.h"
16 #include "intel_gt_regs.h"
17 #include "i915_drv.h"
18 #include "i915_scatterlist.h"
19 #include "i915_utils.h"
20 #include "i915_vgpu.h"
21 
22 #include "intel_gtt.h"
23 #include "gen8_ppgtt.h"
24 
i915_ggtt_color_adjust(const struct drm_mm_node * node,unsigned long color,u64 * start,u64 * end)25 static void i915_ggtt_color_adjust(const struct drm_mm_node *node,
26 				   unsigned long color,
27 				   u64 *start,
28 				   u64 *end)
29 {
30 	if (i915_node_color_differs(node, color))
31 		*start += I915_GTT_PAGE_SIZE;
32 
33 	/*
34 	 * Also leave a space between the unallocated reserved node after the
35 	 * GTT and any objects within the GTT, i.e. we use the color adjustment
36 	 * to insert a guard page to prevent prefetches crossing over the
37 	 * GTT boundary.
38 	 */
39 	node = list_next_entry(node, node_list);
40 	if (node->color != color)
41 		*end -= I915_GTT_PAGE_SIZE;
42 }
43 
ggtt_init_hw(struct i915_ggtt * ggtt)44 static int ggtt_init_hw(struct i915_ggtt *ggtt)
45 {
46 	struct drm_i915_private *i915 = ggtt->vm.i915;
47 
48 	i915_address_space_init(&ggtt->vm, VM_CLASS_GGTT);
49 
50 	ggtt->vm.is_ggtt = true;
51 
52 	/* Only VLV supports read-only GGTT mappings */
53 	ggtt->vm.has_read_only = IS_VALLEYVIEW(i915);
54 
55 	if (!HAS_LLC(i915) && !HAS_PPGTT(i915))
56 		ggtt->vm.mm.color_adjust = i915_ggtt_color_adjust;
57 
58 	if (ggtt->mappable_end) {
59 		if (!io_mapping_init_wc(&ggtt->iomap,
60 					ggtt->gmadr.start,
61 					ggtt->mappable_end)) {
62 			ggtt->vm.cleanup(&ggtt->vm);
63 			return -EIO;
64 		}
65 
66 		ggtt->mtrr = arch_phys_wc_add(ggtt->gmadr.start,
67 					      ggtt->mappable_end);
68 	}
69 
70 	intel_ggtt_init_fences(ggtt);
71 
72 	return 0;
73 }
74 
75 /**
76  * i915_ggtt_init_hw - Initialize GGTT hardware
77  * @i915: i915 device
78  */
i915_ggtt_init_hw(struct drm_i915_private * i915)79 int i915_ggtt_init_hw(struct drm_i915_private *i915)
80 {
81 	int ret;
82 
83 	/*
84 	 * Note that we use page colouring to enforce a guard page at the
85 	 * end of the address space. This is required as the CS may prefetch
86 	 * beyond the end of the batch buffer, across the page boundary,
87 	 * and beyond the end of the GTT if we do not provide a guard.
88 	 */
89 	ret = ggtt_init_hw(to_gt(i915)->ggtt);
90 	if (ret)
91 		return ret;
92 
93 	return 0;
94 }
95 
96 /**
97  * i915_ggtt_suspend_vm - Suspend the memory mappings for a GGTT or DPT VM
98  * @vm: The VM to suspend the mappings for
99  *
100  * Suspend the memory mappings for all objects mapped to HW via the GGTT or a
101  * DPT page table.
102  */
i915_ggtt_suspend_vm(struct i915_address_space * vm)103 void i915_ggtt_suspend_vm(struct i915_address_space *vm)
104 {
105 	struct i915_vma *vma, *vn;
106 	int save_skip_rewrite;
107 
108 	drm_WARN_ON(&vm->i915->drm, !vm->is_ggtt && !vm->is_dpt);
109 
110 retry:
111 	i915_gem_drain_freed_objects(vm->i915);
112 
113 	mutex_lock(&vm->mutex);
114 
115 	/*
116 	 * Skip rewriting PTE on VMA unbind.
117 	 * FIXME: Use an argument to i915_vma_unbind() instead?
118 	 */
119 	save_skip_rewrite = vm->skip_pte_rewrite;
120 	vm->skip_pte_rewrite = true;
121 
122 	list_for_each_entry_safe(vma, vn, &vm->bound_list, vm_link) {
123 		struct drm_i915_gem_object *obj = vma->obj;
124 
125 		GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
126 
127 		if (i915_vma_is_pinned(vma) || !i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
128 			continue;
129 
130 		/* unlikely to race when GPU is idle, so no worry about slowpath.. */
131 		if (WARN_ON(!i915_gem_object_trylock(obj, NULL))) {
132 			/*
133 			 * No dead objects should appear here, GPU should be
134 			 * completely idle, and userspace suspended
135 			 */
136 			i915_gem_object_get(obj);
137 
138 			mutex_unlock(&vm->mutex);
139 
140 			i915_gem_object_lock(obj, NULL);
141 			GEM_WARN_ON(i915_vma_unbind(vma));
142 			i915_gem_object_unlock(obj);
143 			i915_gem_object_put(obj);
144 
145 			vm->skip_pte_rewrite = save_skip_rewrite;
146 			goto retry;
147 		}
148 
149 		if (!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) {
150 			i915_vma_wait_for_bind(vma);
151 
152 			__i915_vma_evict(vma, false);
153 			drm_mm_remove_node(&vma->node);
154 		}
155 
156 		i915_gem_object_unlock(obj);
157 	}
158 
159 	vm->clear_range(vm, 0, vm->total);
160 
161 	vm->skip_pte_rewrite = save_skip_rewrite;
162 
163 	mutex_unlock(&vm->mutex);
164 }
165 
i915_ggtt_suspend(struct i915_ggtt * ggtt)166 void i915_ggtt_suspend(struct i915_ggtt *ggtt)
167 {
168 	i915_ggtt_suspend_vm(&ggtt->vm);
169 	ggtt->invalidate(ggtt);
170 
171 	intel_gt_check_and_clear_faults(ggtt->vm.gt);
172 }
173 
gen6_ggtt_invalidate(struct i915_ggtt * ggtt)174 void gen6_ggtt_invalidate(struct i915_ggtt *ggtt)
175 {
176 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
177 
178 	spin_lock_irq(&uncore->lock);
179 	intel_uncore_write_fw(uncore, GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
180 	intel_uncore_read_fw(uncore, GFX_FLSH_CNTL_GEN6);
181 	spin_unlock_irq(&uncore->lock);
182 }
183 
gen8_ggtt_invalidate(struct i915_ggtt * ggtt)184 void gen8_ggtt_invalidate(struct i915_ggtt *ggtt)
185 {
186 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
187 
188 	/*
189 	 * Note that as an uncached mmio write, this will flush the
190 	 * WCB of the writes into the GGTT before it triggers the invalidate.
191 	 */
192 	intel_uncore_write_fw(uncore, GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
193 }
194 
guc_ggtt_invalidate(struct i915_ggtt * ggtt)195 static void guc_ggtt_invalidate(struct i915_ggtt *ggtt)
196 {
197 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
198 	struct drm_i915_private *i915 = ggtt->vm.i915;
199 
200 	gen8_ggtt_invalidate(ggtt);
201 
202 	if (GRAPHICS_VER(i915) >= 12)
203 		intel_uncore_write_fw(uncore, GEN12_GUC_TLB_INV_CR,
204 				      GEN12_GUC_TLB_INV_CR_INVALIDATE);
205 	else
206 		intel_uncore_write_fw(uncore, GEN8_GTCR, GEN8_GTCR_INVALIDATE);
207 }
208 
gen8_ggtt_pte_encode(dma_addr_t addr,enum i915_cache_level level,u32 flags)209 u64 gen8_ggtt_pte_encode(dma_addr_t addr,
210 			 enum i915_cache_level level,
211 			 u32 flags)
212 {
213 	gen8_pte_t pte = addr | GEN8_PAGE_PRESENT;
214 
215 	if (flags & PTE_LM)
216 		pte |= GEN12_GGTT_PTE_LM;
217 
218 	return pte;
219 }
220 
intel_ggtt_bind_vma(struct i915_address_space * vm,struct i915_vm_pt_stash * stash,struct i915_vma_resource * vma_res,enum i915_cache_level cache_level,u32 flags)221 void intel_ggtt_bind_vma(struct i915_address_space *vm,
222 			  struct i915_vm_pt_stash *stash,
223 			  struct i915_vma_resource *vma_res,
224 			  enum i915_cache_level cache_level,
225 			  u32 flags)
226 {
227 	u32 pte_flags;
228 
229 	if (vma_res->bound_flags & (~flags & I915_VMA_BIND_MASK))
230 		return;
231 
232 	vma_res->bound_flags |= flags;
233 
234 	/* Applicable to VLV (gen8+ do not support RO in the GGTT) */
235 	pte_flags = 0;
236 	if (vma_res->bi.readonly)
237 		pte_flags |= PTE_READ_ONLY;
238 	if (vma_res->bi.lmem)
239 		pte_flags |= PTE_LM;
240 
241 	vm->insert_entries(vm, vma_res, cache_level, pte_flags);
242 	vma_res->page_sizes_gtt = I915_GTT_PAGE_SIZE;
243 }
244 
intel_ggtt_unbind_vma(struct i915_address_space * vm,struct i915_vma_resource * vma_res)245 void intel_ggtt_unbind_vma(struct i915_address_space *vm,
246 			    struct i915_vma_resource *vma_res)
247 {
248 	vm->clear_range(vm, vma_res->start, vma_res->vma_size);
249 }
250 
ggtt_reserve_guc_top(struct i915_ggtt * ggtt)251 static int ggtt_reserve_guc_top(struct i915_ggtt *ggtt)
252 {
253 	u64 size;
254 	int ret;
255 
256 	if (!intel_uc_uses_guc(&ggtt->vm.gt->uc))
257 		return 0;
258 
259 	GEM_BUG_ON(ggtt->vm.total <= GUC_GGTT_TOP);
260 	size = ggtt->vm.total - GUC_GGTT_TOP;
261 
262 	ret = i915_gem_gtt_reserve(&ggtt->vm, NULL, &ggtt->uc_fw, size,
263 				   GUC_GGTT_TOP, I915_COLOR_UNEVICTABLE,
264 				   PIN_NOEVICT);
265 	if (ret)
266 		drm_dbg(&ggtt->vm.i915->drm,
267 			"Failed to reserve top of GGTT for GuC\n");
268 
269 	return ret;
270 }
271 
ggtt_release_guc_top(struct i915_ggtt * ggtt)272 static void ggtt_release_guc_top(struct i915_ggtt *ggtt)
273 {
274 	if (drm_mm_node_allocated(&ggtt->uc_fw))
275 		drm_mm_remove_node(&ggtt->uc_fw);
276 }
277 
cleanup_init_ggtt(struct i915_ggtt * ggtt)278 static void cleanup_init_ggtt(struct i915_ggtt *ggtt)
279 {
280 	ggtt_release_guc_top(ggtt);
281 	if (drm_mm_node_allocated(&ggtt->error_capture))
282 		drm_mm_remove_node(&ggtt->error_capture);
283 	mutex_destroy(&ggtt->error_mutex);
284 }
285 
init_ggtt(struct i915_ggtt * ggtt)286 static int init_ggtt(struct i915_ggtt *ggtt)
287 {
288 	/*
289 	 * Let GEM Manage all of the aperture.
290 	 *
291 	 * However, leave one page at the end still bound to the scratch page.
292 	 * There are a number of places where the hardware apparently prefetches
293 	 * past the end of the object, and we've seen multiple hangs with the
294 	 * GPU head pointer stuck in a batchbuffer bound at the last page of the
295 	 * aperture.  One page should be enough to keep any prefetching inside
296 	 * of the aperture.
297 	 */
298 	unsigned long hole_start, hole_end;
299 	struct drm_mm_node *entry;
300 	int ret;
301 
302 	/*
303 	 * GuC requires all resources that we're sharing with it to be placed in
304 	 * non-WOPCM memory. If GuC is not present or not in use we still need a
305 	 * small bias as ring wraparound at offset 0 sometimes hangs. No idea
306 	 * why.
307 	 */
308 	ggtt->pin_bias = max_t(u32, I915_GTT_PAGE_SIZE,
309 			       intel_wopcm_guc_size(&ggtt->vm.i915->wopcm));
310 
311 	ret = intel_vgt_balloon(ggtt);
312 	if (ret)
313 		return ret;
314 
315 	mutex_init(&ggtt->error_mutex);
316 	if (ggtt->mappable_end) {
317 		/*
318 		 * Reserve a mappable slot for our lockless error capture.
319 		 *
320 		 * We strongly prefer taking address 0x0 in order to protect
321 		 * other critical buffers against accidental overwrites,
322 		 * as writing to address 0 is a very common mistake.
323 		 *
324 		 * Since 0 may already be in use by the system (e.g. the BIOS
325 		 * framebuffer), we let the reservation fail quietly and hope
326 		 * 0 remains reserved always.
327 		 *
328 		 * If we fail to reserve 0, and then fail to find any space
329 		 * for an error-capture, remain silent. We can afford not
330 		 * to reserve an error_capture node as we have fallback
331 		 * paths, and we trust that 0 will remain reserved. However,
332 		 * the only likely reason for failure to insert is a driver
333 		 * bug, which we expect to cause other failures...
334 		 */
335 		ggtt->error_capture.size = I915_GTT_PAGE_SIZE;
336 		ggtt->error_capture.color = I915_COLOR_UNEVICTABLE;
337 		if (drm_mm_reserve_node(&ggtt->vm.mm, &ggtt->error_capture))
338 			drm_mm_insert_node_in_range(&ggtt->vm.mm,
339 						    &ggtt->error_capture,
340 						    ggtt->error_capture.size, 0,
341 						    ggtt->error_capture.color,
342 						    0, ggtt->mappable_end,
343 						    DRM_MM_INSERT_LOW);
344 	}
345 	if (drm_mm_node_allocated(&ggtt->error_capture))
346 		drm_dbg(&ggtt->vm.i915->drm,
347 			"Reserved GGTT:[%llx, %llx] for use by error capture\n",
348 			ggtt->error_capture.start,
349 			ggtt->error_capture.start + ggtt->error_capture.size);
350 
351 	/*
352 	 * The upper portion of the GuC address space has a sizeable hole
353 	 * (several MB) that is inaccessible by GuC. Reserve this range within
354 	 * GGTT as it can comfortably hold GuC/HuC firmware images.
355 	 */
356 	ret = ggtt_reserve_guc_top(ggtt);
357 	if (ret)
358 		goto err;
359 
360 	/* Clear any non-preallocated blocks */
361 	drm_mm_for_each_hole(entry, &ggtt->vm.mm, hole_start, hole_end) {
362 		drm_dbg(&ggtt->vm.i915->drm,
363 			"clearing unused GTT space: [%lx, %lx]\n",
364 			hole_start, hole_end);
365 		ggtt->vm.clear_range(&ggtt->vm, hole_start,
366 				     hole_end - hole_start);
367 	}
368 
369 	/* And finally clear the reserved guard page */
370 	ggtt->vm.clear_range(&ggtt->vm, ggtt->vm.total - PAGE_SIZE, PAGE_SIZE);
371 
372 	return 0;
373 
374 err:
375 	cleanup_init_ggtt(ggtt);
376 	return ret;
377 }
378 
aliasing_gtt_bind_vma(struct i915_address_space * vm,struct i915_vm_pt_stash * stash,struct i915_vma_resource * vma_res,enum i915_cache_level cache_level,u32 flags)379 static void aliasing_gtt_bind_vma(struct i915_address_space *vm,
380 				  struct i915_vm_pt_stash *stash,
381 				  struct i915_vma_resource *vma_res,
382 				  enum i915_cache_level cache_level,
383 				  u32 flags)
384 {
385 	u32 pte_flags;
386 
387 	/* Currently applicable only to VLV */
388 	pte_flags = 0;
389 	if (vma_res->bi.readonly)
390 		pte_flags |= PTE_READ_ONLY;
391 
392 	if (flags & I915_VMA_LOCAL_BIND)
393 		ppgtt_bind_vma(&i915_vm_to_ggtt(vm)->alias->vm,
394 			       stash, vma_res, cache_level, flags);
395 
396 	if (flags & I915_VMA_GLOBAL_BIND)
397 		vm->insert_entries(vm, vma_res, cache_level, pte_flags);
398 
399 	vma_res->bound_flags |= flags;
400 }
401 
aliasing_gtt_unbind_vma(struct i915_address_space * vm,struct i915_vma_resource * vma_res)402 static void aliasing_gtt_unbind_vma(struct i915_address_space *vm,
403 				    struct i915_vma_resource *vma_res)
404 {
405 	if (vma_res->bound_flags & I915_VMA_GLOBAL_BIND)
406 		vm->clear_range(vm, vma_res->start, vma_res->vma_size);
407 
408 	if (vma_res->bound_flags & I915_VMA_LOCAL_BIND)
409 		ppgtt_unbind_vma(&i915_vm_to_ggtt(vm)->alias->vm, vma_res);
410 }
411 
init_aliasing_ppgtt(struct i915_ggtt * ggtt)412 static int init_aliasing_ppgtt(struct i915_ggtt *ggtt)
413 {
414 	struct i915_vm_pt_stash stash = {};
415 	struct i915_ppgtt *ppgtt;
416 	int err;
417 
418 	ppgtt = i915_ppgtt_create(ggtt->vm.gt, 0);
419 	if (IS_ERR(ppgtt))
420 		return PTR_ERR(ppgtt);
421 
422 	if (GEM_WARN_ON(ppgtt->vm.total < ggtt->vm.total)) {
423 		err = -ENODEV;
424 		goto err_ppgtt;
425 	}
426 
427 	err = i915_vm_alloc_pt_stash(&ppgtt->vm, &stash, ggtt->vm.total);
428 	if (err)
429 		goto err_ppgtt;
430 
431 	i915_gem_object_lock(ppgtt->vm.scratch[0], NULL);
432 	err = i915_vm_map_pt_stash(&ppgtt->vm, &stash);
433 	i915_gem_object_unlock(ppgtt->vm.scratch[0]);
434 	if (err)
435 		goto err_stash;
436 
437 	/*
438 	 * Note we only pre-allocate as far as the end of the global
439 	 * GTT. On 48b / 4-level page-tables, the difference is very,
440 	 * very significant! We have to preallocate as GVT/vgpu does
441 	 * not like the page directory disappearing.
442 	 */
443 	ppgtt->vm.allocate_va_range(&ppgtt->vm, &stash, 0, ggtt->vm.total);
444 
445 	ggtt->alias = ppgtt;
446 	ggtt->vm.bind_async_flags |= ppgtt->vm.bind_async_flags;
447 
448 	GEM_BUG_ON(ggtt->vm.vma_ops.bind_vma != intel_ggtt_bind_vma);
449 	ggtt->vm.vma_ops.bind_vma = aliasing_gtt_bind_vma;
450 
451 	GEM_BUG_ON(ggtt->vm.vma_ops.unbind_vma != intel_ggtt_unbind_vma);
452 	ggtt->vm.vma_ops.unbind_vma = aliasing_gtt_unbind_vma;
453 
454 	i915_vm_free_pt_stash(&ppgtt->vm, &stash);
455 	return 0;
456 
457 err_stash:
458 	i915_vm_free_pt_stash(&ppgtt->vm, &stash);
459 err_ppgtt:
460 	i915_vm_put(&ppgtt->vm);
461 	return err;
462 }
463 
fini_aliasing_ppgtt(struct i915_ggtt * ggtt)464 static void fini_aliasing_ppgtt(struct i915_ggtt *ggtt)
465 {
466 	struct i915_ppgtt *ppgtt;
467 
468 	ppgtt = fetch_and_zero(&ggtt->alias);
469 	if (!ppgtt)
470 		return;
471 
472 	i915_vm_put(&ppgtt->vm);
473 
474 	ggtt->vm.vma_ops.bind_vma   = intel_ggtt_bind_vma;
475 	ggtt->vm.vma_ops.unbind_vma = intel_ggtt_unbind_vma;
476 }
477 
i915_init_ggtt(struct drm_i915_private * i915)478 int i915_init_ggtt(struct drm_i915_private *i915)
479 {
480 	int ret;
481 
482 	ret = init_ggtt(to_gt(i915)->ggtt);
483 	if (ret)
484 		return ret;
485 
486 	if (INTEL_PPGTT(i915) == INTEL_PPGTT_ALIASING) {
487 		ret = init_aliasing_ppgtt(to_gt(i915)->ggtt);
488 		if (ret)
489 			cleanup_init_ggtt(to_gt(i915)->ggtt);
490 	}
491 
492 	return 0;
493 }
494 
ggtt_cleanup_hw(struct i915_ggtt * ggtt)495 static void ggtt_cleanup_hw(struct i915_ggtt *ggtt)
496 {
497 	struct i915_vma *vma, *vn;
498 
499 	flush_workqueue(ggtt->vm.i915->wq);
500 	i915_gem_drain_freed_objects(ggtt->vm.i915);
501 
502 	mutex_lock(&ggtt->vm.mutex);
503 
504 	ggtt->vm.skip_pte_rewrite = true;
505 
506 	list_for_each_entry_safe(vma, vn, &ggtt->vm.bound_list, vm_link) {
507 		struct drm_i915_gem_object *obj = vma->obj;
508 		bool trylock;
509 
510 		trylock = i915_gem_object_trylock(obj, NULL);
511 		WARN_ON(!trylock);
512 
513 		WARN_ON(__i915_vma_unbind(vma));
514 		if (trylock)
515 			i915_gem_object_unlock(obj);
516 	}
517 
518 	if (drm_mm_node_allocated(&ggtt->error_capture))
519 		drm_mm_remove_node(&ggtt->error_capture);
520 	mutex_destroy(&ggtt->error_mutex);
521 
522 	ggtt_release_guc_top(ggtt);
523 	intel_vgt_deballoon(ggtt);
524 
525 	ggtt->vm.cleanup(&ggtt->vm);
526 
527 	mutex_unlock(&ggtt->vm.mutex);
528 	i915_address_space_fini(&ggtt->vm);
529 
530 	arch_phys_wc_del(ggtt->mtrr);
531 
532 	if (ggtt->iomap.size)
533 		io_mapping_fini(&ggtt->iomap);
534 }
535 
536 /**
537  * i915_ggtt_driver_release - Clean up GGTT hardware initialization
538  * @i915: i915 device
539  */
i915_ggtt_driver_release(struct drm_i915_private * i915)540 void i915_ggtt_driver_release(struct drm_i915_private *i915)
541 {
542 	struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
543 
544 	fini_aliasing_ppgtt(ggtt);
545 
546 	intel_ggtt_fini_fences(ggtt);
547 	ggtt_cleanup_hw(ggtt);
548 }
549 
550 /**
551  * i915_ggtt_driver_late_release - Cleanup of GGTT that needs to be done after
552  * all free objects have been drained.
553  * @i915: i915 device
554  */
i915_ggtt_driver_late_release(struct drm_i915_private * i915)555 void i915_ggtt_driver_late_release(struct drm_i915_private *i915)
556 {
557 	struct i915_ggtt *ggtt = to_gt(i915)->ggtt;
558 
559 	GEM_WARN_ON(kref_read(&ggtt->vm.resv_ref) != 1);
560 	dma_resv_fini(&ggtt->vm._resv);
561 }
562 
intel_pci_resource(struct pci_dev * pdev,int bar)563 struct resource intel_pci_resource(struct pci_dev *pdev, int bar)
564 {
565 	return (struct resource)DEFINE_RES_MEM(pci_resource_start(pdev, bar),
566 					       pci_resource_len(pdev, bar));
567 }
568 
ggtt_probe_hw(struct i915_ggtt * ggtt,struct intel_gt * gt)569 static int ggtt_probe_hw(struct i915_ggtt *ggtt, struct intel_gt *gt)
570 {
571 	struct drm_i915_private *i915 = gt->i915;
572 	int ret;
573 
574 	ggtt->vm.gt = gt;
575 	ggtt->vm.i915 = i915;
576 	ggtt->vm.dma = i915->drm.dev;
577 	dma_resv_init(&ggtt->vm._resv);
578 
579 	if (GRAPHICS_VER(i915) <= 5)
580 		ret = intel_gt_gmch_gen5_probe(ggtt);
581 	else if (GRAPHICS_VER(i915) < 8)
582 		ret = intel_gt_gmch_gen6_probe(ggtt);
583 	else
584 		ret = intel_gt_gmch_gen8_probe(ggtt);
585 	if (ret) {
586 		dma_resv_fini(&ggtt->vm._resv);
587 		return ret;
588 	}
589 
590 	if ((ggtt->vm.total - 1) >> 32) {
591 		drm_err(&i915->drm,
592 			"We never expected a Global GTT with more than 32bits"
593 			" of address space! Found %lldM!\n",
594 			ggtt->vm.total >> 20);
595 		ggtt->vm.total = 1ULL << 32;
596 		ggtt->mappable_end =
597 			min_t(u64, ggtt->mappable_end, ggtt->vm.total);
598 	}
599 
600 	if (ggtt->mappable_end > ggtt->vm.total) {
601 		drm_err(&i915->drm,
602 			"mappable aperture extends past end of GGTT,"
603 			" aperture=%pa, total=%llx\n",
604 			&ggtt->mappable_end, ggtt->vm.total);
605 		ggtt->mappable_end = ggtt->vm.total;
606 	}
607 
608 	/* GMADR is the PCI mmio aperture into the global GTT. */
609 	drm_dbg(&i915->drm, "GGTT size = %lluM\n", ggtt->vm.total >> 20);
610 	drm_dbg(&i915->drm, "GMADR size = %lluM\n",
611 		(u64)ggtt->mappable_end >> 20);
612 	drm_dbg(&i915->drm, "DSM size = %lluM\n",
613 		(u64)resource_size(&intel_graphics_stolen_res) >> 20);
614 
615 	return 0;
616 }
617 
618 /**
619  * i915_ggtt_probe_hw - Probe GGTT hardware location
620  * @i915: i915 device
621  */
i915_ggtt_probe_hw(struct drm_i915_private * i915)622 int i915_ggtt_probe_hw(struct drm_i915_private *i915)
623 {
624 	int ret;
625 
626 	ret = ggtt_probe_hw(to_gt(i915)->ggtt, to_gt(i915));
627 	if (ret)
628 		return ret;
629 
630 	if (i915_vtd_active(i915))
631 		drm_info(&i915->drm, "VT-d active for gfx access\n");
632 
633 	return 0;
634 }
635 
i915_ggtt_enable_hw(struct drm_i915_private * i915)636 int i915_ggtt_enable_hw(struct drm_i915_private *i915)
637 {
638 	return intel_gt_gmch_gen5_enable_hw(i915);
639 }
640 
i915_ggtt_enable_guc(struct i915_ggtt * ggtt)641 void i915_ggtt_enable_guc(struct i915_ggtt *ggtt)
642 {
643 	GEM_BUG_ON(ggtt->invalidate != gen8_ggtt_invalidate);
644 
645 	ggtt->invalidate = guc_ggtt_invalidate;
646 
647 	ggtt->invalidate(ggtt);
648 }
649 
i915_ggtt_disable_guc(struct i915_ggtt * ggtt)650 void i915_ggtt_disable_guc(struct i915_ggtt *ggtt)
651 {
652 	/* XXX Temporary pardon for error unload */
653 	if (ggtt->invalidate == gen8_ggtt_invalidate)
654 		return;
655 
656 	/* We should only be called after i915_ggtt_enable_guc() */
657 	GEM_BUG_ON(ggtt->invalidate != guc_ggtt_invalidate);
658 
659 	ggtt->invalidate = gen8_ggtt_invalidate;
660 
661 	ggtt->invalidate(ggtt);
662 }
663 
664 /**
665  * i915_ggtt_resume_vm - Restore the memory mappings for a GGTT or DPT VM
666  * @vm: The VM to restore the mappings for
667  *
668  * Restore the memory mappings for all objects mapped to HW via the GGTT or a
669  * DPT page table.
670  *
671  * Returns %true if restoring the mapping for any object that was in a write
672  * domain before suspend.
673  */
i915_ggtt_resume_vm(struct i915_address_space * vm)674 bool i915_ggtt_resume_vm(struct i915_address_space *vm)
675 {
676 	struct i915_vma *vma;
677 	bool write_domain_objs = false;
678 
679 	drm_WARN_ON(&vm->i915->drm, !vm->is_ggtt && !vm->is_dpt);
680 
681 	/* First fill our portion of the GTT with scratch pages */
682 	vm->clear_range(vm, 0, vm->total);
683 
684 	/* clflush objects bound into the GGTT and rebind them. */
685 	list_for_each_entry(vma, &vm->bound_list, vm_link) {
686 		struct drm_i915_gem_object *obj = vma->obj;
687 		unsigned int was_bound =
688 			atomic_read(&vma->flags) & I915_VMA_BIND_MASK;
689 
690 		GEM_BUG_ON(!was_bound);
691 		vma->ops->bind_vma(vm, NULL, vma->resource,
692 				   obj ? obj->cache_level : 0,
693 				   was_bound);
694 		if (obj) { /* only used during resume => exclusive access */
695 			write_domain_objs |= fetch_and_zero(&obj->write_domain);
696 			obj->read_domains |= I915_GEM_DOMAIN_GTT;
697 		}
698 	}
699 
700 	return write_domain_objs;
701 }
702 
i915_ggtt_resume(struct i915_ggtt * ggtt)703 void i915_ggtt_resume(struct i915_ggtt *ggtt)
704 {
705 	bool flush;
706 
707 	intel_gt_check_and_clear_faults(ggtt->vm.gt);
708 
709 	flush = i915_ggtt_resume_vm(&ggtt->vm);
710 
711 	ggtt->invalidate(ggtt);
712 
713 	if (flush)
714 		wbinvd_on_all_cpus();
715 
716 	if (GRAPHICS_VER(ggtt->vm.i915) >= 8)
717 		setup_private_pat(ggtt->vm.gt->uncore);
718 
719 	intel_ggtt_restore_fences(ggtt);
720 }
721