1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2020 Intel Corporation
4  */
5 
6 #include <linux/stop_machine.h>
7 
8 #include <asm/set_memory.h>
9 #include <asm/smp.h>
10 
11 #include <drm/i915_drm.h>
12 
13 #include "intel_gt.h"
14 #include "i915_drv.h"
15 #include "i915_scatterlist.h"
16 #include "i915_vgpu.h"
17 
18 #include "intel_gtt.h"
19 
20 static int
21 i915_get_ggtt_vma_pages(struct i915_vma *vma);
22 
i915_ggtt_color_adjust(const struct drm_mm_node * node,unsigned long color,u64 * start,u64 * end)23 static void i915_ggtt_color_adjust(const struct drm_mm_node *node,
24 				   unsigned long color,
25 				   u64 *start,
26 				   u64 *end)
27 {
28 	if (i915_node_color_differs(node, color))
29 		*start += I915_GTT_PAGE_SIZE;
30 
31 	/*
32 	 * Also leave a space between the unallocated reserved node after the
33 	 * GTT and any objects within the GTT, i.e. we use the color adjustment
34 	 * to insert a guard page to prevent prefetches crossing over the
35 	 * GTT boundary.
36 	 */
37 	node = list_next_entry(node, node_list);
38 	if (node->color != color)
39 		*end -= I915_GTT_PAGE_SIZE;
40 }
41 
ggtt_init_hw(struct i915_ggtt * ggtt)42 static int ggtt_init_hw(struct i915_ggtt *ggtt)
43 {
44 	struct drm_i915_private *i915 = ggtt->vm.i915;
45 
46 	i915_address_space_init(&ggtt->vm, VM_CLASS_GGTT);
47 
48 	ggtt->vm.is_ggtt = true;
49 
50 	/* Only VLV supports read-only GGTT mappings */
51 	ggtt->vm.has_read_only = IS_VALLEYVIEW(i915);
52 
53 	if (!HAS_LLC(i915) && !HAS_PPGTT(i915))
54 		ggtt->vm.mm.color_adjust = i915_ggtt_color_adjust;
55 
56 	if (ggtt->mappable_end) {
57 		if (!io_mapping_init_wc(&ggtt->iomap,
58 					ggtt->gmadr.start,
59 					ggtt->mappable_end)) {
60 			ggtt->vm.cleanup(&ggtt->vm);
61 			return -EIO;
62 		}
63 
64 		ggtt->mtrr = arch_phys_wc_add(ggtt->gmadr.start,
65 					      ggtt->mappable_end);
66 	}
67 
68 	intel_ggtt_init_fences(ggtt);
69 
70 	return 0;
71 }
72 
73 /**
74  * i915_ggtt_init_hw - Initialize GGTT hardware
75  * @i915: i915 device
76  */
i915_ggtt_init_hw(struct drm_i915_private * i915)77 int i915_ggtt_init_hw(struct drm_i915_private *i915)
78 {
79 	int ret;
80 
81 	/*
82 	 * Note that we use page colouring to enforce a guard page at the
83 	 * end of the address space. This is required as the CS may prefetch
84 	 * beyond the end of the batch buffer, across the page boundary,
85 	 * and beyond the end of the GTT if we do not provide a guard.
86 	 */
87 	ret = ggtt_init_hw(&i915->ggtt);
88 	if (ret)
89 		return ret;
90 
91 	return 0;
92 }
93 
94 /*
95  * Certain Gen5 chipsets require require idling the GPU before
96  * unmapping anything from the GTT when VT-d is enabled.
97  */
needs_idle_maps(struct drm_i915_private * i915)98 static bool needs_idle_maps(struct drm_i915_private *i915)
99 {
100 	/*
101 	 * Query intel_iommu to see if we need the workaround. Presumably that
102 	 * was loaded first.
103 	 */
104 	return IS_GEN(i915, 5) && IS_MOBILE(i915) && intel_vtd_active();
105 }
106 
i915_ggtt_suspend(struct i915_ggtt * ggtt)107 void i915_ggtt_suspend(struct i915_ggtt *ggtt)
108 {
109 	struct i915_vma *vma, *vn;
110 	int open;
111 
112 	mutex_lock(&ggtt->vm.mutex);
113 
114 	/* Skip rewriting PTE on VMA unbind. */
115 	open = atomic_xchg(&ggtt->vm.open, 0);
116 
117 	list_for_each_entry_safe(vma, vn, &ggtt->vm.bound_list, vm_link) {
118 		GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
119 		i915_vma_wait_for_bind(vma);
120 
121 		if (i915_vma_is_pinned(vma))
122 			continue;
123 
124 		if (!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) {
125 			__i915_vma_evict(vma);
126 			drm_mm_remove_node(&vma->node);
127 		}
128 	}
129 
130 	ggtt->vm.clear_range(&ggtt->vm, 0, ggtt->vm.total);
131 	ggtt->invalidate(ggtt);
132 	atomic_set(&ggtt->vm.open, open);
133 
134 	mutex_unlock(&ggtt->vm.mutex);
135 
136 	intel_gt_check_and_clear_faults(ggtt->vm.gt);
137 }
138 
gen6_ggtt_invalidate(struct i915_ggtt * ggtt)139 void gen6_ggtt_invalidate(struct i915_ggtt *ggtt)
140 {
141 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
142 
143 	spin_lock_irq(&uncore->lock);
144 	intel_uncore_write_fw(uncore, GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
145 	intel_uncore_read_fw(uncore, GFX_FLSH_CNTL_GEN6);
146 	spin_unlock_irq(&uncore->lock);
147 }
148 
gen8_ggtt_invalidate(struct i915_ggtt * ggtt)149 static void gen8_ggtt_invalidate(struct i915_ggtt *ggtt)
150 {
151 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
152 
153 	/*
154 	 * Note that as an uncached mmio write, this will flush the
155 	 * WCB of the writes into the GGTT before it triggers the invalidate.
156 	 */
157 	intel_uncore_write_fw(uncore, GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
158 }
159 
guc_ggtt_invalidate(struct i915_ggtt * ggtt)160 static void guc_ggtt_invalidate(struct i915_ggtt *ggtt)
161 {
162 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
163 	struct drm_i915_private *i915 = ggtt->vm.i915;
164 
165 	gen8_ggtt_invalidate(ggtt);
166 
167 	if (INTEL_GEN(i915) >= 12)
168 		intel_uncore_write_fw(uncore, GEN12_GUC_TLB_INV_CR,
169 				      GEN12_GUC_TLB_INV_CR_INVALIDATE);
170 	else
171 		intel_uncore_write_fw(uncore, GEN8_GTCR, GEN8_GTCR_INVALIDATE);
172 }
173 
gmch_ggtt_invalidate(struct i915_ggtt * ggtt)174 static void gmch_ggtt_invalidate(struct i915_ggtt *ggtt)
175 {
176 	intel_gtt_chipset_flush();
177 }
178 
gen8_ggtt_pte_encode(dma_addr_t addr,enum i915_cache_level level,u32 flags)179 static u64 gen8_ggtt_pte_encode(dma_addr_t addr,
180 				enum i915_cache_level level,
181 				u32 flags)
182 {
183 	return addr | _PAGE_PRESENT;
184 }
185 
gen8_set_pte(void __iomem * addr,gen8_pte_t pte)186 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
187 {
188 	writeq(pte, addr);
189 }
190 
gen8_ggtt_insert_page(struct i915_address_space * vm,dma_addr_t addr,u64 offset,enum i915_cache_level level,u32 unused)191 static void gen8_ggtt_insert_page(struct i915_address_space *vm,
192 				  dma_addr_t addr,
193 				  u64 offset,
194 				  enum i915_cache_level level,
195 				  u32 unused)
196 {
197 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
198 	gen8_pte_t __iomem *pte =
199 		(gen8_pte_t __iomem *)ggtt->gsm + offset / I915_GTT_PAGE_SIZE;
200 
201 	gen8_set_pte(pte, gen8_ggtt_pte_encode(addr, level, 0));
202 
203 	ggtt->invalidate(ggtt);
204 }
205 
gen8_ggtt_insert_entries(struct i915_address_space * vm,struct i915_vma * vma,enum i915_cache_level level,u32 flags)206 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
207 				     struct i915_vma *vma,
208 				     enum i915_cache_level level,
209 				     u32 flags)
210 {
211 	const gen8_pte_t pte_encode = gen8_ggtt_pte_encode(0, level, 0);
212 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
213 	gen8_pte_t __iomem *gte;
214 	gen8_pte_t __iomem *end;
215 	struct sgt_iter iter;
216 	dma_addr_t addr;
217 
218 	/*
219 	 * Note that we ignore PTE_READ_ONLY here. The caller must be careful
220 	 * not to allow the user to override access to a read only page.
221 	 */
222 
223 	gte = (gen8_pte_t __iomem *)ggtt->gsm;
224 	gte += vma->node.start / I915_GTT_PAGE_SIZE;
225 	end = gte + vma->node.size / I915_GTT_PAGE_SIZE;
226 
227 	for_each_sgt_daddr(addr, iter, vma->pages)
228 		gen8_set_pte(gte++, pte_encode | addr);
229 	GEM_BUG_ON(gte > end);
230 
231 	/* Fill the allocated but "unused" space beyond the end of the buffer */
232 	while (gte < end)
233 		gen8_set_pte(gte++, vm->scratch[0]->encode);
234 
235 	/*
236 	 * We want to flush the TLBs only after we're certain all the PTE
237 	 * updates have finished.
238 	 */
239 	ggtt->invalidate(ggtt);
240 }
241 
gen6_ggtt_insert_page(struct i915_address_space * vm,dma_addr_t addr,u64 offset,enum i915_cache_level level,u32 flags)242 static void gen6_ggtt_insert_page(struct i915_address_space *vm,
243 				  dma_addr_t addr,
244 				  u64 offset,
245 				  enum i915_cache_level level,
246 				  u32 flags)
247 {
248 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
249 	gen6_pte_t __iomem *pte =
250 		(gen6_pte_t __iomem *)ggtt->gsm + offset / I915_GTT_PAGE_SIZE;
251 
252 	iowrite32(vm->pte_encode(addr, level, flags), pte);
253 
254 	ggtt->invalidate(ggtt);
255 }
256 
257 /*
258  * Binds an object into the global gtt with the specified cache level.
259  * The object will be accessible to the GPU via commands whose operands
260  * reference offsets within the global GTT as well as accessible by the GPU
261  * through the GMADR mapped BAR (i915->mm.gtt->gtt).
262  */
gen6_ggtt_insert_entries(struct i915_address_space * vm,struct i915_vma * vma,enum i915_cache_level level,u32 flags)263 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
264 				     struct i915_vma *vma,
265 				     enum i915_cache_level level,
266 				     u32 flags)
267 {
268 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
269 	gen6_pte_t __iomem *gte;
270 	gen6_pte_t __iomem *end;
271 	struct sgt_iter iter;
272 	dma_addr_t addr;
273 
274 	gte = (gen6_pte_t __iomem *)ggtt->gsm;
275 	gte += vma->node.start / I915_GTT_PAGE_SIZE;
276 	end = gte + vma->node.size / I915_GTT_PAGE_SIZE;
277 
278 	for_each_sgt_daddr(addr, iter, vma->pages)
279 		iowrite32(vm->pte_encode(addr, level, flags), gte++);
280 	GEM_BUG_ON(gte > end);
281 
282 	/* Fill the allocated but "unused" space beyond the end of the buffer */
283 	while (gte < end)
284 		iowrite32(vm->scratch[0]->encode, gte++);
285 
286 	/*
287 	 * We want to flush the TLBs only after we're certain all the PTE
288 	 * updates have finished.
289 	 */
290 	ggtt->invalidate(ggtt);
291 }
292 
nop_clear_range(struct i915_address_space * vm,u64 start,u64 length)293 static void nop_clear_range(struct i915_address_space *vm,
294 			    u64 start, u64 length)
295 {
296 }
297 
gen8_ggtt_clear_range(struct i915_address_space * vm,u64 start,u64 length)298 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
299 				  u64 start, u64 length)
300 {
301 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
302 	unsigned int first_entry = start / I915_GTT_PAGE_SIZE;
303 	unsigned int num_entries = length / I915_GTT_PAGE_SIZE;
304 	const gen8_pte_t scratch_pte = vm->scratch[0]->encode;
305 	gen8_pte_t __iomem *gtt_base =
306 		(gen8_pte_t __iomem *)ggtt->gsm + first_entry;
307 	const int max_entries = ggtt_total_entries(ggtt) - first_entry;
308 	int i;
309 
310 	if (WARN(num_entries > max_entries,
311 		 "First entry = %d; Num entries = %d (max=%d)\n",
312 		 first_entry, num_entries, max_entries))
313 		num_entries = max_entries;
314 
315 	for (i = 0; i < num_entries; i++)
316 		gen8_set_pte(&gtt_base[i], scratch_pte);
317 }
318 
bxt_vtd_ggtt_wa(struct i915_address_space * vm)319 static void bxt_vtd_ggtt_wa(struct i915_address_space *vm)
320 {
321 	/*
322 	 * Make sure the internal GAM fifo has been cleared of all GTT
323 	 * writes before exiting stop_machine(). This guarantees that
324 	 * any aperture accesses waiting to start in another process
325 	 * cannot back up behind the GTT writes causing a hang.
326 	 * The register can be any arbitrary GAM register.
327 	 */
328 	intel_uncore_posting_read_fw(vm->gt->uncore, GFX_FLSH_CNTL_GEN6);
329 }
330 
331 struct insert_page {
332 	struct i915_address_space *vm;
333 	dma_addr_t addr;
334 	u64 offset;
335 	enum i915_cache_level level;
336 };
337 
bxt_vtd_ggtt_insert_page__cb(void * _arg)338 static int bxt_vtd_ggtt_insert_page__cb(void *_arg)
339 {
340 	struct insert_page *arg = _arg;
341 
342 	gen8_ggtt_insert_page(arg->vm, arg->addr, arg->offset, arg->level, 0);
343 	bxt_vtd_ggtt_wa(arg->vm);
344 
345 	return 0;
346 }
347 
bxt_vtd_ggtt_insert_page__BKL(struct i915_address_space * vm,dma_addr_t addr,u64 offset,enum i915_cache_level level,u32 unused)348 static void bxt_vtd_ggtt_insert_page__BKL(struct i915_address_space *vm,
349 					  dma_addr_t addr,
350 					  u64 offset,
351 					  enum i915_cache_level level,
352 					  u32 unused)
353 {
354 	struct insert_page arg = { vm, addr, offset, level };
355 
356 	stop_machine(bxt_vtd_ggtt_insert_page__cb, &arg, NULL);
357 }
358 
359 struct insert_entries {
360 	struct i915_address_space *vm;
361 	struct i915_vma *vma;
362 	enum i915_cache_level level;
363 	u32 flags;
364 };
365 
bxt_vtd_ggtt_insert_entries__cb(void * _arg)366 static int bxt_vtd_ggtt_insert_entries__cb(void *_arg)
367 {
368 	struct insert_entries *arg = _arg;
369 
370 	gen8_ggtt_insert_entries(arg->vm, arg->vma, arg->level, arg->flags);
371 	bxt_vtd_ggtt_wa(arg->vm);
372 
373 	return 0;
374 }
375 
bxt_vtd_ggtt_insert_entries__BKL(struct i915_address_space * vm,struct i915_vma * vma,enum i915_cache_level level,u32 flags)376 static void bxt_vtd_ggtt_insert_entries__BKL(struct i915_address_space *vm,
377 					     struct i915_vma *vma,
378 					     enum i915_cache_level level,
379 					     u32 flags)
380 {
381 	struct insert_entries arg = { vm, vma, level, flags };
382 
383 	stop_machine(bxt_vtd_ggtt_insert_entries__cb, &arg, NULL);
384 }
385 
gen6_ggtt_clear_range(struct i915_address_space * vm,u64 start,u64 length)386 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
387 				  u64 start, u64 length)
388 {
389 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
390 	unsigned int first_entry = start / I915_GTT_PAGE_SIZE;
391 	unsigned int num_entries = length / I915_GTT_PAGE_SIZE;
392 	gen6_pte_t scratch_pte, __iomem *gtt_base =
393 		(gen6_pte_t __iomem *)ggtt->gsm + first_entry;
394 	const int max_entries = ggtt_total_entries(ggtt) - first_entry;
395 	int i;
396 
397 	if (WARN(num_entries > max_entries,
398 		 "First entry = %d; Num entries = %d (max=%d)\n",
399 		 first_entry, num_entries, max_entries))
400 		num_entries = max_entries;
401 
402 	scratch_pte = vm->scratch[0]->encode;
403 	for (i = 0; i < num_entries; i++)
404 		iowrite32(scratch_pte, &gtt_base[i]);
405 }
406 
i915_ggtt_insert_page(struct i915_address_space * vm,dma_addr_t addr,u64 offset,enum i915_cache_level cache_level,u32 unused)407 static void i915_ggtt_insert_page(struct i915_address_space *vm,
408 				  dma_addr_t addr,
409 				  u64 offset,
410 				  enum i915_cache_level cache_level,
411 				  u32 unused)
412 {
413 	unsigned int flags = (cache_level == I915_CACHE_NONE) ?
414 		AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
415 
416 	intel_gtt_insert_page(addr, offset >> PAGE_SHIFT, flags);
417 }
418 
i915_ggtt_insert_entries(struct i915_address_space * vm,struct i915_vma * vma,enum i915_cache_level cache_level,u32 unused)419 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
420 				     struct i915_vma *vma,
421 				     enum i915_cache_level cache_level,
422 				     u32 unused)
423 {
424 	unsigned int flags = (cache_level == I915_CACHE_NONE) ?
425 		AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
426 
427 	intel_gtt_insert_sg_entries(vma->pages, vma->node.start >> PAGE_SHIFT,
428 				    flags);
429 }
430 
i915_ggtt_clear_range(struct i915_address_space * vm,u64 start,u64 length)431 static void i915_ggtt_clear_range(struct i915_address_space *vm,
432 				  u64 start, u64 length)
433 {
434 	intel_gtt_clear_range(start >> PAGE_SHIFT, length >> PAGE_SHIFT);
435 }
436 
ggtt_bind_vma(struct i915_address_space * vm,struct i915_vm_pt_stash * stash,struct i915_vma * vma,enum i915_cache_level cache_level,u32 flags)437 static void ggtt_bind_vma(struct i915_address_space *vm,
438 			  struct i915_vm_pt_stash *stash,
439 			  struct i915_vma *vma,
440 			  enum i915_cache_level cache_level,
441 			  u32 flags)
442 {
443 	struct drm_i915_gem_object *obj = vma->obj;
444 	u32 pte_flags;
445 
446 	if (i915_vma_is_bound(vma, ~flags & I915_VMA_BIND_MASK))
447 		return;
448 
449 	/* Applicable to VLV (gen8+ do not support RO in the GGTT) */
450 	pte_flags = 0;
451 	if (i915_gem_object_is_readonly(obj))
452 		pte_flags |= PTE_READ_ONLY;
453 
454 	vm->insert_entries(vm, vma, cache_level, pte_flags);
455 	vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
456 }
457 
ggtt_unbind_vma(struct i915_address_space * vm,struct i915_vma * vma)458 static void ggtt_unbind_vma(struct i915_address_space *vm, struct i915_vma *vma)
459 {
460 	vm->clear_range(vm, vma->node.start, vma->size);
461 }
462 
ggtt_reserve_guc_top(struct i915_ggtt * ggtt)463 static int ggtt_reserve_guc_top(struct i915_ggtt *ggtt)
464 {
465 	u64 size;
466 	int ret;
467 
468 	if (!intel_uc_uses_guc(&ggtt->vm.gt->uc))
469 		return 0;
470 
471 	GEM_BUG_ON(ggtt->vm.total <= GUC_GGTT_TOP);
472 	size = ggtt->vm.total - GUC_GGTT_TOP;
473 
474 	ret = i915_gem_gtt_reserve(&ggtt->vm, &ggtt->uc_fw, size,
475 				   GUC_GGTT_TOP, I915_COLOR_UNEVICTABLE,
476 				   PIN_NOEVICT);
477 	if (ret)
478 		drm_dbg(&ggtt->vm.i915->drm,
479 			"Failed to reserve top of GGTT for GuC\n");
480 
481 	return ret;
482 }
483 
ggtt_release_guc_top(struct i915_ggtt * ggtt)484 static void ggtt_release_guc_top(struct i915_ggtt *ggtt)
485 {
486 	if (drm_mm_node_allocated(&ggtt->uc_fw))
487 		drm_mm_remove_node(&ggtt->uc_fw);
488 }
489 
cleanup_init_ggtt(struct i915_ggtt * ggtt)490 static void cleanup_init_ggtt(struct i915_ggtt *ggtt)
491 {
492 	ggtt_release_guc_top(ggtt);
493 	if (drm_mm_node_allocated(&ggtt->error_capture))
494 		drm_mm_remove_node(&ggtt->error_capture);
495 	mutex_destroy(&ggtt->error_mutex);
496 }
497 
init_ggtt(struct i915_ggtt * ggtt)498 static int init_ggtt(struct i915_ggtt *ggtt)
499 {
500 	/*
501 	 * Let GEM Manage all of the aperture.
502 	 *
503 	 * However, leave one page at the end still bound to the scratch page.
504 	 * There are a number of places where the hardware apparently prefetches
505 	 * past the end of the object, and we've seen multiple hangs with the
506 	 * GPU head pointer stuck in a batchbuffer bound at the last page of the
507 	 * aperture.  One page should be enough to keep any prefetching inside
508 	 * of the aperture.
509 	 */
510 	unsigned long hole_start, hole_end;
511 	struct drm_mm_node *entry;
512 	int ret;
513 
514 	/*
515 	 * GuC requires all resources that we're sharing with it to be placed in
516 	 * non-WOPCM memory. If GuC is not present or not in use we still need a
517 	 * small bias as ring wraparound at offset 0 sometimes hangs. No idea
518 	 * why.
519 	 */
520 	ggtt->pin_bias = max_t(u32, I915_GTT_PAGE_SIZE,
521 			       intel_wopcm_guc_size(&ggtt->vm.i915->wopcm));
522 
523 	ret = intel_vgt_balloon(ggtt);
524 	if (ret)
525 		return ret;
526 
527 	mutex_init(&ggtt->error_mutex);
528 	if (ggtt->mappable_end) {
529 		/* Reserve a mappable slot for our lockless error capture */
530 		ret = drm_mm_insert_node_in_range(&ggtt->vm.mm,
531 						  &ggtt->error_capture,
532 						  PAGE_SIZE, 0,
533 						  I915_COLOR_UNEVICTABLE,
534 						  0, ggtt->mappable_end,
535 						  DRM_MM_INSERT_LOW);
536 		if (ret)
537 			return ret;
538 	}
539 
540 	/*
541 	 * The upper portion of the GuC address space has a sizeable hole
542 	 * (several MB) that is inaccessible by GuC. Reserve this range within
543 	 * GGTT as it can comfortably hold GuC/HuC firmware images.
544 	 */
545 	ret = ggtt_reserve_guc_top(ggtt);
546 	if (ret)
547 		goto err;
548 
549 	/* Clear any non-preallocated blocks */
550 	drm_mm_for_each_hole(entry, &ggtt->vm.mm, hole_start, hole_end) {
551 		drm_dbg_kms(&ggtt->vm.i915->drm,
552 			    "clearing unused GTT space: [%lx, %lx]\n",
553 			    hole_start, hole_end);
554 		ggtt->vm.clear_range(&ggtt->vm, hole_start,
555 				     hole_end - hole_start);
556 	}
557 
558 	/* And finally clear the reserved guard page */
559 	ggtt->vm.clear_range(&ggtt->vm, ggtt->vm.total - PAGE_SIZE, PAGE_SIZE);
560 
561 	return 0;
562 
563 err:
564 	cleanup_init_ggtt(ggtt);
565 	return ret;
566 }
567 
aliasing_gtt_bind_vma(struct i915_address_space * vm,struct i915_vm_pt_stash * stash,struct i915_vma * vma,enum i915_cache_level cache_level,u32 flags)568 static void aliasing_gtt_bind_vma(struct i915_address_space *vm,
569 				  struct i915_vm_pt_stash *stash,
570 				  struct i915_vma *vma,
571 				  enum i915_cache_level cache_level,
572 				  u32 flags)
573 {
574 	u32 pte_flags;
575 
576 	/* Currently applicable only to VLV */
577 	pte_flags = 0;
578 	if (i915_gem_object_is_readonly(vma->obj))
579 		pte_flags |= PTE_READ_ONLY;
580 
581 	if (flags & I915_VMA_LOCAL_BIND)
582 		ppgtt_bind_vma(&i915_vm_to_ggtt(vm)->alias->vm,
583 			       stash, vma, cache_level, flags);
584 
585 	if (flags & I915_VMA_GLOBAL_BIND)
586 		vm->insert_entries(vm, vma, cache_level, pte_flags);
587 }
588 
aliasing_gtt_unbind_vma(struct i915_address_space * vm,struct i915_vma * vma)589 static void aliasing_gtt_unbind_vma(struct i915_address_space *vm,
590 				    struct i915_vma *vma)
591 {
592 	if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
593 		vm->clear_range(vm, vma->node.start, vma->size);
594 
595 	if (i915_vma_is_bound(vma, I915_VMA_LOCAL_BIND))
596 		ppgtt_unbind_vma(&i915_vm_to_ggtt(vm)->alias->vm, vma);
597 }
598 
init_aliasing_ppgtt(struct i915_ggtt * ggtt)599 static int init_aliasing_ppgtt(struct i915_ggtt *ggtt)
600 {
601 	struct i915_vm_pt_stash stash = {};
602 	struct i915_ppgtt *ppgtt;
603 	int err;
604 
605 	ppgtt = i915_ppgtt_create(ggtt->vm.gt);
606 	if (IS_ERR(ppgtt))
607 		return PTR_ERR(ppgtt);
608 
609 	if (GEM_WARN_ON(ppgtt->vm.total < ggtt->vm.total)) {
610 		err = -ENODEV;
611 		goto err_ppgtt;
612 	}
613 
614 	err = i915_vm_alloc_pt_stash(&ppgtt->vm, &stash, ggtt->vm.total);
615 	if (err)
616 		goto err_ppgtt;
617 
618 	err = i915_vm_pin_pt_stash(&ppgtt->vm, &stash);
619 	if (err)
620 		goto err_stash;
621 
622 	/*
623 	 * Note we only pre-allocate as far as the end of the global
624 	 * GTT. On 48b / 4-level page-tables, the difference is very,
625 	 * very significant! We have to preallocate as GVT/vgpu does
626 	 * not like the page directory disappearing.
627 	 */
628 	ppgtt->vm.allocate_va_range(&ppgtt->vm, &stash, 0, ggtt->vm.total);
629 
630 	ggtt->alias = ppgtt;
631 	ggtt->vm.bind_async_flags |= ppgtt->vm.bind_async_flags;
632 
633 	GEM_BUG_ON(ggtt->vm.vma_ops.bind_vma != ggtt_bind_vma);
634 	ggtt->vm.vma_ops.bind_vma = aliasing_gtt_bind_vma;
635 
636 	GEM_BUG_ON(ggtt->vm.vma_ops.unbind_vma != ggtt_unbind_vma);
637 	ggtt->vm.vma_ops.unbind_vma = aliasing_gtt_unbind_vma;
638 
639 	i915_vm_free_pt_stash(&ppgtt->vm, &stash);
640 	return 0;
641 
642 err_stash:
643 	i915_vm_free_pt_stash(&ppgtt->vm, &stash);
644 err_ppgtt:
645 	i915_vm_put(&ppgtt->vm);
646 	return err;
647 }
648 
fini_aliasing_ppgtt(struct i915_ggtt * ggtt)649 static void fini_aliasing_ppgtt(struct i915_ggtt *ggtt)
650 {
651 	struct i915_ppgtt *ppgtt;
652 
653 	ppgtt = fetch_and_zero(&ggtt->alias);
654 	if (!ppgtt)
655 		return;
656 
657 	i915_vm_put(&ppgtt->vm);
658 
659 	ggtt->vm.vma_ops.bind_vma   = ggtt_bind_vma;
660 	ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma;
661 }
662 
i915_init_ggtt(struct drm_i915_private * i915)663 int i915_init_ggtt(struct drm_i915_private *i915)
664 {
665 	int ret;
666 
667 	ret = init_ggtt(&i915->ggtt);
668 	if (ret)
669 		return ret;
670 
671 	if (INTEL_PPGTT(i915) == INTEL_PPGTT_ALIASING) {
672 		ret = init_aliasing_ppgtt(&i915->ggtt);
673 		if (ret)
674 			cleanup_init_ggtt(&i915->ggtt);
675 	}
676 
677 	return 0;
678 }
679 
ggtt_cleanup_hw(struct i915_ggtt * ggtt)680 static void ggtt_cleanup_hw(struct i915_ggtt *ggtt)
681 {
682 	struct i915_vma *vma, *vn;
683 
684 	atomic_set(&ggtt->vm.open, 0);
685 
686 	rcu_barrier(); /* flush the RCU'ed__i915_vm_release */
687 	flush_workqueue(ggtt->vm.i915->wq);
688 
689 	mutex_lock(&ggtt->vm.mutex);
690 
691 	list_for_each_entry_safe(vma, vn, &ggtt->vm.bound_list, vm_link)
692 		WARN_ON(__i915_vma_unbind(vma));
693 
694 	if (drm_mm_node_allocated(&ggtt->error_capture))
695 		drm_mm_remove_node(&ggtt->error_capture);
696 	mutex_destroy(&ggtt->error_mutex);
697 
698 	ggtt_release_guc_top(ggtt);
699 	intel_vgt_deballoon(ggtt);
700 
701 	ggtt->vm.cleanup(&ggtt->vm);
702 
703 	mutex_unlock(&ggtt->vm.mutex);
704 	i915_address_space_fini(&ggtt->vm);
705 
706 	arch_phys_wc_del(ggtt->mtrr);
707 
708 	if (ggtt->iomap.size)
709 		io_mapping_fini(&ggtt->iomap);
710 }
711 
712 /**
713  * i915_ggtt_driver_release - Clean up GGTT hardware initialization
714  * @i915: i915 device
715  */
i915_ggtt_driver_release(struct drm_i915_private * i915)716 void i915_ggtt_driver_release(struct drm_i915_private *i915)
717 {
718 	struct i915_ggtt *ggtt = &i915->ggtt;
719 
720 	fini_aliasing_ppgtt(ggtt);
721 
722 	intel_ggtt_fini_fences(ggtt);
723 	ggtt_cleanup_hw(ggtt);
724 }
725 
gen6_get_total_gtt_size(u16 snb_gmch_ctl)726 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
727 {
728 	snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
729 	snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
730 	return snb_gmch_ctl << 20;
731 }
732 
gen8_get_total_gtt_size(u16 bdw_gmch_ctl)733 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
734 {
735 	bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
736 	bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
737 	if (bdw_gmch_ctl)
738 		bdw_gmch_ctl = 1 << bdw_gmch_ctl;
739 
740 #ifdef CONFIG_X86_32
741 	/* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * I915_GTT_PAGE_SIZE */
742 	if (bdw_gmch_ctl > 4)
743 		bdw_gmch_ctl = 4;
744 #endif
745 
746 	return bdw_gmch_ctl << 20;
747 }
748 
chv_get_total_gtt_size(u16 gmch_ctrl)749 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
750 {
751 	gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
752 	gmch_ctrl &= SNB_GMCH_GGMS_MASK;
753 
754 	if (gmch_ctrl)
755 		return 1 << (20 + gmch_ctrl);
756 
757 	return 0;
758 }
759 
ggtt_probe_common(struct i915_ggtt * ggtt,u64 size)760 static int ggtt_probe_common(struct i915_ggtt *ggtt, u64 size)
761 {
762 	struct drm_i915_private *i915 = ggtt->vm.i915;
763 	struct pci_dev *pdev = i915->drm.pdev;
764 	phys_addr_t phys_addr;
765 	int ret;
766 
767 	/* For Modern GENs the PTEs and register space are split in the BAR */
768 	phys_addr = pci_resource_start(pdev, 0) + pci_resource_len(pdev, 0) / 2;
769 
770 	/*
771 	 * On BXT+/CNL+ writes larger than 64 bit to the GTT pagetable range
772 	 * will be dropped. For WC mappings in general we have 64 byte burst
773 	 * writes when the WC buffer is flushed, so we can't use it, but have to
774 	 * resort to an uncached mapping. The WC issue is easily caught by the
775 	 * readback check when writing GTT PTE entries.
776 	 */
777 	if (IS_GEN9_LP(i915) || INTEL_GEN(i915) >= 10)
778 		ggtt->gsm = ioremap(phys_addr, size);
779 	else
780 		ggtt->gsm = ioremap_wc(phys_addr, size);
781 	if (!ggtt->gsm) {
782 		drm_err(&i915->drm, "Failed to map the ggtt page table\n");
783 		return -ENOMEM;
784 	}
785 
786 	ret = setup_scratch_page(&ggtt->vm);
787 	if (ret) {
788 		drm_err(&i915->drm, "Scratch setup failed\n");
789 		/* iounmap will also get called at remove, but meh */
790 		iounmap(ggtt->gsm);
791 		return ret;
792 	}
793 
794 	ggtt->vm.scratch[0]->encode =
795 		ggtt->vm.pte_encode(px_dma(ggtt->vm.scratch[0]),
796 				    I915_CACHE_NONE, 0);
797 
798 	return 0;
799 }
800 
ggtt_set_pages(struct i915_vma * vma)801 int ggtt_set_pages(struct i915_vma *vma)
802 {
803 	int ret;
804 
805 	GEM_BUG_ON(vma->pages);
806 
807 	ret = i915_get_ggtt_vma_pages(vma);
808 	if (ret)
809 		return ret;
810 
811 	vma->page_sizes = vma->obj->mm.page_sizes;
812 
813 	return 0;
814 }
815 
gen6_gmch_remove(struct i915_address_space * vm)816 static void gen6_gmch_remove(struct i915_address_space *vm)
817 {
818 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
819 
820 	iounmap(ggtt->gsm);
821 	free_scratch(vm);
822 }
823 
pci_resource(struct pci_dev * pdev,int bar)824 static struct resource pci_resource(struct pci_dev *pdev, int bar)
825 {
826 	return (struct resource)DEFINE_RES_MEM(pci_resource_start(pdev, bar),
827 					       pci_resource_len(pdev, bar));
828 }
829 
gen8_gmch_probe(struct i915_ggtt * ggtt)830 static int gen8_gmch_probe(struct i915_ggtt *ggtt)
831 {
832 	struct drm_i915_private *i915 = ggtt->vm.i915;
833 	struct pci_dev *pdev = i915->drm.pdev;
834 	unsigned int size;
835 	u16 snb_gmch_ctl;
836 
837 	/* TODO: We're not aware of mappable constraints on gen8 yet */
838 	if (!IS_DGFX(i915)) {
839 		ggtt->gmadr = pci_resource(pdev, 2);
840 		ggtt->mappable_end = resource_size(&ggtt->gmadr);
841 	}
842 
843 	pci_read_config_word(pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
844 	if (IS_CHERRYVIEW(i915))
845 		size = chv_get_total_gtt_size(snb_gmch_ctl);
846 	else
847 		size = gen8_get_total_gtt_size(snb_gmch_ctl);
848 
849 	ggtt->vm.alloc_pt_dma = alloc_pt_dma;
850 
851 	ggtt->vm.total = (size / sizeof(gen8_pte_t)) * I915_GTT_PAGE_SIZE;
852 	ggtt->vm.cleanup = gen6_gmch_remove;
853 	ggtt->vm.insert_page = gen8_ggtt_insert_page;
854 	ggtt->vm.clear_range = nop_clear_range;
855 	if (intel_scanout_needs_vtd_wa(i915))
856 		ggtt->vm.clear_range = gen8_ggtt_clear_range;
857 
858 	ggtt->vm.insert_entries = gen8_ggtt_insert_entries;
859 
860 	/* Serialize GTT updates with aperture access on BXT if VT-d is on. */
861 	if (intel_ggtt_update_needs_vtd_wa(i915) ||
862 	    IS_CHERRYVIEW(i915) /* fails with concurrent use/update */) {
863 		ggtt->vm.insert_entries = bxt_vtd_ggtt_insert_entries__BKL;
864 		ggtt->vm.insert_page    = bxt_vtd_ggtt_insert_page__BKL;
865 		ggtt->vm.bind_async_flags =
866 			I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
867 	}
868 
869 	ggtt->invalidate = gen8_ggtt_invalidate;
870 
871 	ggtt->vm.vma_ops.bind_vma    = ggtt_bind_vma;
872 	ggtt->vm.vma_ops.unbind_vma  = ggtt_unbind_vma;
873 	ggtt->vm.vma_ops.set_pages   = ggtt_set_pages;
874 	ggtt->vm.vma_ops.clear_pages = clear_pages;
875 
876 	ggtt->vm.pte_encode = gen8_ggtt_pte_encode;
877 
878 	setup_private_pat(ggtt->vm.gt->uncore);
879 
880 	return ggtt_probe_common(ggtt, size);
881 }
882 
snb_pte_encode(dma_addr_t addr,enum i915_cache_level level,u32 flags)883 static u64 snb_pte_encode(dma_addr_t addr,
884 			  enum i915_cache_level level,
885 			  u32 flags)
886 {
887 	gen6_pte_t pte = GEN6_PTE_ADDR_ENCODE(addr) | GEN6_PTE_VALID;
888 
889 	switch (level) {
890 	case I915_CACHE_L3_LLC:
891 	case I915_CACHE_LLC:
892 		pte |= GEN6_PTE_CACHE_LLC;
893 		break;
894 	case I915_CACHE_NONE:
895 		pte |= GEN6_PTE_UNCACHED;
896 		break;
897 	default:
898 		MISSING_CASE(level);
899 	}
900 
901 	return pte;
902 }
903 
ivb_pte_encode(dma_addr_t addr,enum i915_cache_level level,u32 flags)904 static u64 ivb_pte_encode(dma_addr_t addr,
905 			  enum i915_cache_level level,
906 			  u32 flags)
907 {
908 	gen6_pte_t pte = GEN6_PTE_ADDR_ENCODE(addr) | GEN6_PTE_VALID;
909 
910 	switch (level) {
911 	case I915_CACHE_L3_LLC:
912 		pte |= GEN7_PTE_CACHE_L3_LLC;
913 		break;
914 	case I915_CACHE_LLC:
915 		pte |= GEN6_PTE_CACHE_LLC;
916 		break;
917 	case I915_CACHE_NONE:
918 		pte |= GEN6_PTE_UNCACHED;
919 		break;
920 	default:
921 		MISSING_CASE(level);
922 	}
923 
924 	return pte;
925 }
926 
byt_pte_encode(dma_addr_t addr,enum i915_cache_level level,u32 flags)927 static u64 byt_pte_encode(dma_addr_t addr,
928 			  enum i915_cache_level level,
929 			  u32 flags)
930 {
931 	gen6_pte_t pte = GEN6_PTE_ADDR_ENCODE(addr) | GEN6_PTE_VALID;
932 
933 	if (!(flags & PTE_READ_ONLY))
934 		pte |= BYT_PTE_WRITEABLE;
935 
936 	if (level != I915_CACHE_NONE)
937 		pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
938 
939 	return pte;
940 }
941 
hsw_pte_encode(dma_addr_t addr,enum i915_cache_level level,u32 flags)942 static u64 hsw_pte_encode(dma_addr_t addr,
943 			  enum i915_cache_level level,
944 			  u32 flags)
945 {
946 	gen6_pte_t pte = HSW_PTE_ADDR_ENCODE(addr) | GEN6_PTE_VALID;
947 
948 	if (level != I915_CACHE_NONE)
949 		pte |= HSW_WB_LLC_AGE3;
950 
951 	return pte;
952 }
953 
iris_pte_encode(dma_addr_t addr,enum i915_cache_level level,u32 flags)954 static u64 iris_pte_encode(dma_addr_t addr,
955 			   enum i915_cache_level level,
956 			   u32 flags)
957 {
958 	gen6_pte_t pte = HSW_PTE_ADDR_ENCODE(addr) | GEN6_PTE_VALID;
959 
960 	switch (level) {
961 	case I915_CACHE_NONE:
962 		break;
963 	case I915_CACHE_WT:
964 		pte |= HSW_WT_ELLC_LLC_AGE3;
965 		break;
966 	default:
967 		pte |= HSW_WB_ELLC_LLC_AGE3;
968 		break;
969 	}
970 
971 	return pte;
972 }
973 
gen6_gmch_probe(struct i915_ggtt * ggtt)974 static int gen6_gmch_probe(struct i915_ggtt *ggtt)
975 {
976 	struct drm_i915_private *i915 = ggtt->vm.i915;
977 	struct pci_dev *pdev = i915->drm.pdev;
978 	unsigned int size;
979 	u16 snb_gmch_ctl;
980 
981 	ggtt->gmadr = pci_resource(pdev, 2);
982 	ggtt->mappable_end = resource_size(&ggtt->gmadr);
983 
984 	/*
985 	 * 64/512MB is the current min/max we actually know of, but this is
986 	 * just a coarse sanity check.
987 	 */
988 	if (ggtt->mappable_end < (64<<20) || ggtt->mappable_end > (512<<20)) {
989 		drm_err(&i915->drm, "Unknown GMADR size (%pa)\n",
990 			&ggtt->mappable_end);
991 		return -ENXIO;
992 	}
993 
994 	pci_read_config_word(pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
995 
996 	size = gen6_get_total_gtt_size(snb_gmch_ctl);
997 	ggtt->vm.total = (size / sizeof(gen6_pte_t)) * I915_GTT_PAGE_SIZE;
998 
999 	ggtt->vm.alloc_pt_dma = alloc_pt_dma;
1000 
1001 	ggtt->vm.clear_range = nop_clear_range;
1002 	if (!HAS_FULL_PPGTT(i915) || intel_scanout_needs_vtd_wa(i915))
1003 		ggtt->vm.clear_range = gen6_ggtt_clear_range;
1004 	ggtt->vm.insert_page = gen6_ggtt_insert_page;
1005 	ggtt->vm.insert_entries = gen6_ggtt_insert_entries;
1006 	ggtt->vm.cleanup = gen6_gmch_remove;
1007 
1008 	ggtt->invalidate = gen6_ggtt_invalidate;
1009 
1010 	if (HAS_EDRAM(i915))
1011 		ggtt->vm.pte_encode = iris_pte_encode;
1012 	else if (IS_HASWELL(i915))
1013 		ggtt->vm.pte_encode = hsw_pte_encode;
1014 	else if (IS_VALLEYVIEW(i915))
1015 		ggtt->vm.pte_encode = byt_pte_encode;
1016 	else if (INTEL_GEN(i915) >= 7)
1017 		ggtt->vm.pte_encode = ivb_pte_encode;
1018 	else
1019 		ggtt->vm.pte_encode = snb_pte_encode;
1020 
1021 	ggtt->vm.vma_ops.bind_vma    = ggtt_bind_vma;
1022 	ggtt->vm.vma_ops.unbind_vma  = ggtt_unbind_vma;
1023 	ggtt->vm.vma_ops.set_pages   = ggtt_set_pages;
1024 	ggtt->vm.vma_ops.clear_pages = clear_pages;
1025 
1026 	return ggtt_probe_common(ggtt, size);
1027 }
1028 
i915_gmch_remove(struct i915_address_space * vm)1029 static void i915_gmch_remove(struct i915_address_space *vm)
1030 {
1031 	intel_gmch_remove();
1032 }
1033 
i915_gmch_probe(struct i915_ggtt * ggtt)1034 static int i915_gmch_probe(struct i915_ggtt *ggtt)
1035 {
1036 	struct drm_i915_private *i915 = ggtt->vm.i915;
1037 	phys_addr_t gmadr_base;
1038 	int ret;
1039 
1040 	ret = intel_gmch_probe(i915->bridge_dev, i915->drm.pdev, NULL);
1041 	if (!ret) {
1042 		drm_err(&i915->drm, "failed to set up gmch\n");
1043 		return -EIO;
1044 	}
1045 
1046 	intel_gtt_get(&ggtt->vm.total, &gmadr_base, &ggtt->mappable_end);
1047 
1048 	ggtt->gmadr =
1049 		(struct resource)DEFINE_RES_MEM(gmadr_base, ggtt->mappable_end);
1050 
1051 	ggtt->vm.alloc_pt_dma = alloc_pt_dma;
1052 
1053 	ggtt->do_idle_maps = needs_idle_maps(i915);
1054 	ggtt->vm.insert_page = i915_ggtt_insert_page;
1055 	ggtt->vm.insert_entries = i915_ggtt_insert_entries;
1056 	ggtt->vm.clear_range = i915_ggtt_clear_range;
1057 	ggtt->vm.cleanup = i915_gmch_remove;
1058 
1059 	ggtt->invalidate = gmch_ggtt_invalidate;
1060 
1061 	ggtt->vm.vma_ops.bind_vma    = ggtt_bind_vma;
1062 	ggtt->vm.vma_ops.unbind_vma  = ggtt_unbind_vma;
1063 	ggtt->vm.vma_ops.set_pages   = ggtt_set_pages;
1064 	ggtt->vm.vma_ops.clear_pages = clear_pages;
1065 
1066 	if (unlikely(ggtt->do_idle_maps))
1067 		drm_notice(&i915->drm,
1068 			   "Applying Ironlake quirks for intel_iommu\n");
1069 
1070 	return 0;
1071 }
1072 
ggtt_probe_hw(struct i915_ggtt * ggtt,struct intel_gt * gt)1073 static int ggtt_probe_hw(struct i915_ggtt *ggtt, struct intel_gt *gt)
1074 {
1075 	struct drm_i915_private *i915 = gt->i915;
1076 	int ret;
1077 
1078 	ggtt->vm.gt = gt;
1079 	ggtt->vm.i915 = i915;
1080 	ggtt->vm.dma = &i915->drm.pdev->dev;
1081 
1082 	if (INTEL_GEN(i915) <= 5)
1083 		ret = i915_gmch_probe(ggtt);
1084 	else if (INTEL_GEN(i915) < 8)
1085 		ret = gen6_gmch_probe(ggtt);
1086 	else
1087 		ret = gen8_gmch_probe(ggtt);
1088 	if (ret)
1089 		return ret;
1090 
1091 	if ((ggtt->vm.total - 1) >> 32) {
1092 		drm_err(&i915->drm,
1093 			"We never expected a Global GTT with more than 32bits"
1094 			" of address space! Found %lldM!\n",
1095 			ggtt->vm.total >> 20);
1096 		ggtt->vm.total = 1ULL << 32;
1097 		ggtt->mappable_end =
1098 			min_t(u64, ggtt->mappable_end, ggtt->vm.total);
1099 	}
1100 
1101 	if (ggtt->mappable_end > ggtt->vm.total) {
1102 		drm_err(&i915->drm,
1103 			"mappable aperture extends past end of GGTT,"
1104 			" aperture=%pa, total=%llx\n",
1105 			&ggtt->mappable_end, ggtt->vm.total);
1106 		ggtt->mappable_end = ggtt->vm.total;
1107 	}
1108 
1109 	/* GMADR is the PCI mmio aperture into the global GTT. */
1110 	drm_dbg(&i915->drm, "GGTT size = %lluM\n", ggtt->vm.total >> 20);
1111 	drm_dbg(&i915->drm, "GMADR size = %lluM\n",
1112 		(u64)ggtt->mappable_end >> 20);
1113 	drm_dbg(&i915->drm, "DSM size = %lluM\n",
1114 		(u64)resource_size(&intel_graphics_stolen_res) >> 20);
1115 
1116 	return 0;
1117 }
1118 
1119 /**
1120  * i915_ggtt_probe_hw - Probe GGTT hardware location
1121  * @i915: i915 device
1122  */
i915_ggtt_probe_hw(struct drm_i915_private * i915)1123 int i915_ggtt_probe_hw(struct drm_i915_private *i915)
1124 {
1125 	int ret;
1126 
1127 	ret = ggtt_probe_hw(&i915->ggtt, &i915->gt);
1128 	if (ret)
1129 		return ret;
1130 
1131 	if (intel_vtd_active())
1132 		drm_info(&i915->drm, "VT-d active for gfx access\n");
1133 
1134 	return 0;
1135 }
1136 
i915_ggtt_enable_hw(struct drm_i915_private * i915)1137 int i915_ggtt_enable_hw(struct drm_i915_private *i915)
1138 {
1139 	if (INTEL_GEN(i915) < 6 && !intel_enable_gtt())
1140 		return -EIO;
1141 
1142 	return 0;
1143 }
1144 
i915_ggtt_enable_guc(struct i915_ggtt * ggtt)1145 void i915_ggtt_enable_guc(struct i915_ggtt *ggtt)
1146 {
1147 	GEM_BUG_ON(ggtt->invalidate != gen8_ggtt_invalidate);
1148 
1149 	ggtt->invalidate = guc_ggtt_invalidate;
1150 
1151 	ggtt->invalidate(ggtt);
1152 }
1153 
i915_ggtt_disable_guc(struct i915_ggtt * ggtt)1154 void i915_ggtt_disable_guc(struct i915_ggtt *ggtt)
1155 {
1156 	/* XXX Temporary pardon for error unload */
1157 	if (ggtt->invalidate == gen8_ggtt_invalidate)
1158 		return;
1159 
1160 	/* We should only be called after i915_ggtt_enable_guc() */
1161 	GEM_BUG_ON(ggtt->invalidate != guc_ggtt_invalidate);
1162 
1163 	ggtt->invalidate = gen8_ggtt_invalidate;
1164 
1165 	ggtt->invalidate(ggtt);
1166 }
1167 
i915_ggtt_resume(struct i915_ggtt * ggtt)1168 void i915_ggtt_resume(struct i915_ggtt *ggtt)
1169 {
1170 	struct i915_vma *vma;
1171 	bool flush = false;
1172 	int open;
1173 
1174 	intel_gt_check_and_clear_faults(ggtt->vm.gt);
1175 
1176 	/* First fill our portion of the GTT with scratch pages */
1177 	ggtt->vm.clear_range(&ggtt->vm, 0, ggtt->vm.total);
1178 
1179 	/* Skip rewriting PTE on VMA unbind. */
1180 	open = atomic_xchg(&ggtt->vm.open, 0);
1181 
1182 	/* clflush objects bound into the GGTT and rebind them. */
1183 	list_for_each_entry(vma, &ggtt->vm.bound_list, vm_link) {
1184 		struct drm_i915_gem_object *obj = vma->obj;
1185 		unsigned int was_bound =
1186 			atomic_read(&vma->flags) & I915_VMA_BIND_MASK;
1187 
1188 		GEM_BUG_ON(!was_bound);
1189 		vma->ops->bind_vma(&ggtt->vm, NULL, vma,
1190 				   obj ? obj->cache_level : 0,
1191 				   was_bound);
1192 		if (obj) { /* only used during resume => exclusive access */
1193 			flush |= fetch_and_zero(&obj->write_domain);
1194 			obj->read_domains |= I915_GEM_DOMAIN_GTT;
1195 		}
1196 	}
1197 
1198 	atomic_set(&ggtt->vm.open, open);
1199 	ggtt->invalidate(ggtt);
1200 
1201 	if (flush)
1202 		wbinvd_on_all_cpus();
1203 
1204 	if (INTEL_GEN(ggtt->vm.i915) >= 8)
1205 		setup_private_pat(ggtt->vm.gt->uncore);
1206 
1207 	intel_ggtt_restore_fences(ggtt);
1208 }
1209 
1210 static struct scatterlist *
rotate_pages(struct drm_i915_gem_object * obj,unsigned int offset,unsigned int width,unsigned int height,unsigned int stride,struct sg_table * st,struct scatterlist * sg)1211 rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
1212 	     unsigned int width, unsigned int height,
1213 	     unsigned int stride,
1214 	     struct sg_table *st, struct scatterlist *sg)
1215 {
1216 	unsigned int column, row;
1217 	unsigned int src_idx;
1218 
1219 	for (column = 0; column < width; column++) {
1220 		src_idx = stride * (height - 1) + column + offset;
1221 		for (row = 0; row < height; row++) {
1222 			st->nents++;
1223 			/*
1224 			 * We don't need the pages, but need to initialize
1225 			 * the entries so the sg list can be happily traversed.
1226 			 * The only thing we need are DMA addresses.
1227 			 */
1228 			sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0);
1229 			sg_dma_address(sg) =
1230 				i915_gem_object_get_dma_address(obj, src_idx);
1231 			sg_dma_len(sg) = I915_GTT_PAGE_SIZE;
1232 			sg = sg_next(sg);
1233 			src_idx -= stride;
1234 		}
1235 	}
1236 
1237 	return sg;
1238 }
1239 
1240 static noinline struct sg_table *
intel_rotate_pages(struct intel_rotation_info * rot_info,struct drm_i915_gem_object * obj)1241 intel_rotate_pages(struct intel_rotation_info *rot_info,
1242 		   struct drm_i915_gem_object *obj)
1243 {
1244 	unsigned int size = intel_rotation_info_size(rot_info);
1245 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
1246 	struct sg_table *st;
1247 	struct scatterlist *sg;
1248 	int ret = -ENOMEM;
1249 	int i;
1250 
1251 	/* Allocate target SG list. */
1252 	st = kmalloc(sizeof(*st), GFP_KERNEL);
1253 	if (!st)
1254 		goto err_st_alloc;
1255 
1256 	ret = sg_alloc_table(st, size, GFP_KERNEL);
1257 	if (ret)
1258 		goto err_sg_alloc;
1259 
1260 	st->nents = 0;
1261 	sg = st->sgl;
1262 
1263 	for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++) {
1264 		sg = rotate_pages(obj, rot_info->plane[i].offset,
1265 				  rot_info->plane[i].width, rot_info->plane[i].height,
1266 				  rot_info->plane[i].stride, st, sg);
1267 	}
1268 
1269 	return st;
1270 
1271 err_sg_alloc:
1272 	kfree(st);
1273 err_st_alloc:
1274 
1275 	drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",
1276 		obj->base.size, rot_info->plane[0].width,
1277 		rot_info->plane[0].height, size);
1278 
1279 	return ERR_PTR(ret);
1280 }
1281 
1282 static struct scatterlist *
remap_pages(struct drm_i915_gem_object * obj,unsigned int offset,unsigned int width,unsigned int height,unsigned int stride,struct sg_table * st,struct scatterlist * sg)1283 remap_pages(struct drm_i915_gem_object *obj, unsigned int offset,
1284 	    unsigned int width, unsigned int height,
1285 	    unsigned int stride,
1286 	    struct sg_table *st, struct scatterlist *sg)
1287 {
1288 	unsigned int row;
1289 
1290 	for (row = 0; row < height; row++) {
1291 		unsigned int left = width * I915_GTT_PAGE_SIZE;
1292 
1293 		while (left) {
1294 			dma_addr_t addr;
1295 			unsigned int length;
1296 
1297 			/*
1298 			 * We don't need the pages, but need to initialize
1299 			 * the entries so the sg list can be happily traversed.
1300 			 * The only thing we need are DMA addresses.
1301 			 */
1302 
1303 			addr = i915_gem_object_get_dma_address_len(obj, offset, &length);
1304 
1305 			length = min(left, length);
1306 
1307 			st->nents++;
1308 
1309 			sg_set_page(sg, NULL, length, 0);
1310 			sg_dma_address(sg) = addr;
1311 			sg_dma_len(sg) = length;
1312 			sg = sg_next(sg);
1313 
1314 			offset += length / I915_GTT_PAGE_SIZE;
1315 			left -= length;
1316 		}
1317 
1318 		offset += stride - width;
1319 	}
1320 
1321 	return sg;
1322 }
1323 
1324 static noinline struct sg_table *
intel_remap_pages(struct intel_remapped_info * rem_info,struct drm_i915_gem_object * obj)1325 intel_remap_pages(struct intel_remapped_info *rem_info,
1326 		  struct drm_i915_gem_object *obj)
1327 {
1328 	unsigned int size = intel_remapped_info_size(rem_info);
1329 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
1330 	struct sg_table *st;
1331 	struct scatterlist *sg;
1332 	int ret = -ENOMEM;
1333 	int i;
1334 
1335 	/* Allocate target SG list. */
1336 	st = kmalloc(sizeof(*st), GFP_KERNEL);
1337 	if (!st)
1338 		goto err_st_alloc;
1339 
1340 	ret = sg_alloc_table(st, size, GFP_KERNEL);
1341 	if (ret)
1342 		goto err_sg_alloc;
1343 
1344 	st->nents = 0;
1345 	sg = st->sgl;
1346 
1347 	for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) {
1348 		sg = remap_pages(obj, rem_info->plane[i].offset,
1349 				 rem_info->plane[i].width, rem_info->plane[i].height,
1350 				 rem_info->plane[i].stride, st, sg);
1351 	}
1352 
1353 	i915_sg_trim(st);
1354 
1355 	return st;
1356 
1357 err_sg_alloc:
1358 	kfree(st);
1359 err_st_alloc:
1360 
1361 	drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n",
1362 		obj->base.size, rem_info->plane[0].width,
1363 		rem_info->plane[0].height, size);
1364 
1365 	return ERR_PTR(ret);
1366 }
1367 
1368 static noinline struct sg_table *
intel_partial_pages(const struct i915_ggtt_view * view,struct drm_i915_gem_object * obj)1369 intel_partial_pages(const struct i915_ggtt_view *view,
1370 		    struct drm_i915_gem_object *obj)
1371 {
1372 	struct sg_table *st;
1373 	struct scatterlist *sg, *iter;
1374 	unsigned int count = view->partial.size;
1375 	unsigned int offset;
1376 	int ret = -ENOMEM;
1377 
1378 	st = kmalloc(sizeof(*st), GFP_KERNEL);
1379 	if (!st)
1380 		goto err_st_alloc;
1381 
1382 	ret = sg_alloc_table(st, count, GFP_KERNEL);
1383 	if (ret)
1384 		goto err_sg_alloc;
1385 
1386 	iter = i915_gem_object_get_sg(obj, view->partial.offset, &offset);
1387 	GEM_BUG_ON(!iter);
1388 
1389 	sg = st->sgl;
1390 	st->nents = 0;
1391 	do {
1392 		unsigned int len;
1393 
1394 		len = min(iter->length - (offset << PAGE_SHIFT),
1395 			  count << PAGE_SHIFT);
1396 		sg_set_page(sg, NULL, len, 0);
1397 		sg_dma_address(sg) =
1398 			sg_dma_address(iter) + (offset << PAGE_SHIFT);
1399 		sg_dma_len(sg) = len;
1400 
1401 		st->nents++;
1402 		count -= len >> PAGE_SHIFT;
1403 		if (count == 0) {
1404 			sg_mark_end(sg);
1405 			i915_sg_trim(st); /* Drop any unused tail entries. */
1406 
1407 			return st;
1408 		}
1409 
1410 		sg = __sg_next(sg);
1411 		iter = __sg_next(iter);
1412 		offset = 0;
1413 	} while (1);
1414 
1415 err_sg_alloc:
1416 	kfree(st);
1417 err_st_alloc:
1418 	return ERR_PTR(ret);
1419 }
1420 
1421 static int
i915_get_ggtt_vma_pages(struct i915_vma * vma)1422 i915_get_ggtt_vma_pages(struct i915_vma *vma)
1423 {
1424 	int ret;
1425 
1426 	/*
1427 	 * The vma->pages are only valid within the lifespan of the borrowed
1428 	 * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so
1429 	 * must be the vma->pages. A simple rule is that vma->pages must only
1430 	 * be accessed when the obj->mm.pages are pinned.
1431 	 */
1432 	GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));
1433 
1434 	switch (vma->ggtt_view.type) {
1435 	default:
1436 		GEM_BUG_ON(vma->ggtt_view.type);
1437 		fallthrough;
1438 	case I915_GGTT_VIEW_NORMAL:
1439 		vma->pages = vma->obj->mm.pages;
1440 		return 0;
1441 
1442 	case I915_GGTT_VIEW_ROTATED:
1443 		vma->pages =
1444 			intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj);
1445 		break;
1446 
1447 	case I915_GGTT_VIEW_REMAPPED:
1448 		vma->pages =
1449 			intel_remap_pages(&vma->ggtt_view.remapped, vma->obj);
1450 		break;
1451 
1452 	case I915_GGTT_VIEW_PARTIAL:
1453 		vma->pages = intel_partial_pages(&vma->ggtt_view, vma->obj);
1454 		break;
1455 	}
1456 
1457 	ret = 0;
1458 	if (IS_ERR(vma->pages)) {
1459 		ret = PTR_ERR(vma->pages);
1460 		vma->pages = NULL;
1461 		drm_err(&vma->vm->i915->drm,
1462 			"Failed to get pages for VMA view type %u (%d)!\n",
1463 			vma->ggtt_view.type, ret);
1464 	}
1465 	return ret;
1466 }
1467