xref: /openbmc/linux/drivers/gpu/drm/i915/gt/intel_ggtt.c (revision 82df5b73)
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 
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 
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  */
77 int i915_ggtt_init_hw(struct drm_i915_private *i915)
78 {
79 	int ret;
80 
81 	stash_init(&i915->mm.wc_stash);
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(&i915->ggtt);
90 	if (ret)
91 		return ret;
92 
93 	return 0;
94 }
95 
96 /*
97  * Certain Gen5 chipsets require require idling the GPU before
98  * unmapping anything from the GTT when VT-d is enabled.
99  */
100 static bool needs_idle_maps(struct drm_i915_private *i915)
101 {
102 	/*
103 	 * Query intel_iommu to see if we need the workaround. Presumably that
104 	 * was loaded first.
105 	 */
106 	return IS_GEN(i915, 5) && IS_MOBILE(i915) && intel_vtd_active();
107 }
108 
109 void i915_ggtt_suspend(struct i915_ggtt *ggtt)
110 {
111 	struct i915_vma *vma;
112 
113 	list_for_each_entry(vma, &ggtt->vm.bound_list, vm_link)
114 		i915_vma_wait_for_bind(vma);
115 
116 	ggtt->vm.clear_range(&ggtt->vm, 0, ggtt->vm.total);
117 	ggtt->invalidate(ggtt);
118 
119 	intel_gt_check_and_clear_faults(ggtt->vm.gt);
120 }
121 
122 void gen6_ggtt_invalidate(struct i915_ggtt *ggtt)
123 {
124 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
125 
126 	spin_lock_irq(&uncore->lock);
127 	intel_uncore_write_fw(uncore, GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
128 	intel_uncore_read_fw(uncore, GFX_FLSH_CNTL_GEN6);
129 	spin_unlock_irq(&uncore->lock);
130 }
131 
132 static void gen8_ggtt_invalidate(struct i915_ggtt *ggtt)
133 {
134 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
135 
136 	/*
137 	 * Note that as an uncached mmio write, this will flush the
138 	 * WCB of the writes into the GGTT before it triggers the invalidate.
139 	 */
140 	intel_uncore_write_fw(uncore, GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
141 }
142 
143 static void guc_ggtt_invalidate(struct i915_ggtt *ggtt)
144 {
145 	struct intel_uncore *uncore = ggtt->vm.gt->uncore;
146 	struct drm_i915_private *i915 = ggtt->vm.i915;
147 
148 	gen8_ggtt_invalidate(ggtt);
149 
150 	if (INTEL_GEN(i915) >= 12)
151 		intel_uncore_write_fw(uncore, GEN12_GUC_TLB_INV_CR,
152 				      GEN12_GUC_TLB_INV_CR_INVALIDATE);
153 	else
154 		intel_uncore_write_fw(uncore, GEN8_GTCR, GEN8_GTCR_INVALIDATE);
155 }
156 
157 static void gmch_ggtt_invalidate(struct i915_ggtt *ggtt)
158 {
159 	intel_gtt_chipset_flush();
160 }
161 
162 static u64 gen8_ggtt_pte_encode(dma_addr_t addr,
163 				enum i915_cache_level level,
164 				u32 flags)
165 {
166 	return addr | _PAGE_PRESENT;
167 }
168 
169 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
170 {
171 	writeq(pte, addr);
172 }
173 
174 static void gen8_ggtt_insert_page(struct i915_address_space *vm,
175 				  dma_addr_t addr,
176 				  u64 offset,
177 				  enum i915_cache_level level,
178 				  u32 unused)
179 {
180 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
181 	gen8_pte_t __iomem *pte =
182 		(gen8_pte_t __iomem *)ggtt->gsm + offset / I915_GTT_PAGE_SIZE;
183 
184 	gen8_set_pte(pte, gen8_ggtt_pte_encode(addr, level, 0));
185 
186 	ggtt->invalidate(ggtt);
187 }
188 
189 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
190 				     struct i915_vma *vma,
191 				     enum i915_cache_level level,
192 				     u32 flags)
193 {
194 	const gen8_pte_t pte_encode = gen8_ggtt_pte_encode(0, level, 0);
195 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
196 	gen8_pte_t __iomem *gte;
197 	gen8_pte_t __iomem *end;
198 	struct sgt_iter iter;
199 	dma_addr_t addr;
200 
201 	/*
202 	 * Note that we ignore PTE_READ_ONLY here. The caller must be careful
203 	 * not to allow the user to override access to a read only page.
204 	 */
205 
206 	gte = (gen8_pte_t __iomem *)ggtt->gsm;
207 	gte += vma->node.start / I915_GTT_PAGE_SIZE;
208 	end = gte + vma->node.size / I915_GTT_PAGE_SIZE;
209 
210 	for_each_sgt_daddr(addr, iter, vma->pages)
211 		gen8_set_pte(gte++, pte_encode | addr);
212 	GEM_BUG_ON(gte > end);
213 
214 	/* Fill the allocated but "unused" space beyond the end of the buffer */
215 	while (gte < end)
216 		gen8_set_pte(gte++, vm->scratch[0].encode);
217 
218 	/*
219 	 * We want to flush the TLBs only after we're certain all the PTE
220 	 * updates have finished.
221 	 */
222 	ggtt->invalidate(ggtt);
223 }
224 
225 static void gen6_ggtt_insert_page(struct i915_address_space *vm,
226 				  dma_addr_t addr,
227 				  u64 offset,
228 				  enum i915_cache_level level,
229 				  u32 flags)
230 {
231 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
232 	gen6_pte_t __iomem *pte =
233 		(gen6_pte_t __iomem *)ggtt->gsm + offset / I915_GTT_PAGE_SIZE;
234 
235 	iowrite32(vm->pte_encode(addr, level, flags), pte);
236 
237 	ggtt->invalidate(ggtt);
238 }
239 
240 /*
241  * Binds an object into the global gtt with the specified cache level.
242  * The object will be accessible to the GPU via commands whose operands
243  * reference offsets within the global GTT as well as accessible by the GPU
244  * through the GMADR mapped BAR (i915->mm.gtt->gtt).
245  */
246 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
247 				     struct i915_vma *vma,
248 				     enum i915_cache_level level,
249 				     u32 flags)
250 {
251 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
252 	gen6_pte_t __iomem *gte;
253 	gen6_pte_t __iomem *end;
254 	struct sgt_iter iter;
255 	dma_addr_t addr;
256 
257 	gte = (gen6_pte_t __iomem *)ggtt->gsm;
258 	gte += vma->node.start / I915_GTT_PAGE_SIZE;
259 	end = gte + vma->node.size / I915_GTT_PAGE_SIZE;
260 
261 	for_each_sgt_daddr(addr, iter, vma->pages)
262 		iowrite32(vm->pte_encode(addr, level, flags), gte++);
263 	GEM_BUG_ON(gte > end);
264 
265 	/* Fill the allocated but "unused" space beyond the end of the buffer */
266 	while (gte < end)
267 		iowrite32(vm->scratch[0].encode, gte++);
268 
269 	/*
270 	 * We want to flush the TLBs only after we're certain all the PTE
271 	 * updates have finished.
272 	 */
273 	ggtt->invalidate(ggtt);
274 }
275 
276 static void nop_clear_range(struct i915_address_space *vm,
277 			    u64 start, u64 length)
278 {
279 }
280 
281 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
282 				  u64 start, u64 length)
283 {
284 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
285 	unsigned int first_entry = start / I915_GTT_PAGE_SIZE;
286 	unsigned int num_entries = length / I915_GTT_PAGE_SIZE;
287 	const gen8_pte_t scratch_pte = vm->scratch[0].encode;
288 	gen8_pte_t __iomem *gtt_base =
289 		(gen8_pte_t __iomem *)ggtt->gsm + first_entry;
290 	const int max_entries = ggtt_total_entries(ggtt) - first_entry;
291 	int i;
292 
293 	if (WARN(num_entries > max_entries,
294 		 "First entry = %d; Num entries = %d (max=%d)\n",
295 		 first_entry, num_entries, max_entries))
296 		num_entries = max_entries;
297 
298 	for (i = 0; i < num_entries; i++)
299 		gen8_set_pte(&gtt_base[i], scratch_pte);
300 }
301 
302 static void bxt_vtd_ggtt_wa(struct i915_address_space *vm)
303 {
304 	/*
305 	 * Make sure the internal GAM fifo has been cleared of all GTT
306 	 * writes before exiting stop_machine(). This guarantees that
307 	 * any aperture accesses waiting to start in another process
308 	 * cannot back up behind the GTT writes causing a hang.
309 	 * The register can be any arbitrary GAM register.
310 	 */
311 	intel_uncore_posting_read_fw(vm->gt->uncore, GFX_FLSH_CNTL_GEN6);
312 }
313 
314 struct insert_page {
315 	struct i915_address_space *vm;
316 	dma_addr_t addr;
317 	u64 offset;
318 	enum i915_cache_level level;
319 };
320 
321 static int bxt_vtd_ggtt_insert_page__cb(void *_arg)
322 {
323 	struct insert_page *arg = _arg;
324 
325 	gen8_ggtt_insert_page(arg->vm, arg->addr, arg->offset, arg->level, 0);
326 	bxt_vtd_ggtt_wa(arg->vm);
327 
328 	return 0;
329 }
330 
331 static void bxt_vtd_ggtt_insert_page__BKL(struct i915_address_space *vm,
332 					  dma_addr_t addr,
333 					  u64 offset,
334 					  enum i915_cache_level level,
335 					  u32 unused)
336 {
337 	struct insert_page arg = { vm, addr, offset, level };
338 
339 	stop_machine(bxt_vtd_ggtt_insert_page__cb, &arg, NULL);
340 }
341 
342 struct insert_entries {
343 	struct i915_address_space *vm;
344 	struct i915_vma *vma;
345 	enum i915_cache_level level;
346 	u32 flags;
347 };
348 
349 static int bxt_vtd_ggtt_insert_entries__cb(void *_arg)
350 {
351 	struct insert_entries *arg = _arg;
352 
353 	gen8_ggtt_insert_entries(arg->vm, arg->vma, arg->level, arg->flags);
354 	bxt_vtd_ggtt_wa(arg->vm);
355 
356 	return 0;
357 }
358 
359 static void bxt_vtd_ggtt_insert_entries__BKL(struct i915_address_space *vm,
360 					     struct i915_vma *vma,
361 					     enum i915_cache_level level,
362 					     u32 flags)
363 {
364 	struct insert_entries arg = { vm, vma, level, flags };
365 
366 	stop_machine(bxt_vtd_ggtt_insert_entries__cb, &arg, NULL);
367 }
368 
369 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
370 				  u64 start, u64 length)
371 {
372 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
373 	unsigned int first_entry = start / I915_GTT_PAGE_SIZE;
374 	unsigned int num_entries = length / I915_GTT_PAGE_SIZE;
375 	gen6_pte_t scratch_pte, __iomem *gtt_base =
376 		(gen6_pte_t __iomem *)ggtt->gsm + first_entry;
377 	const int max_entries = ggtt_total_entries(ggtt) - first_entry;
378 	int i;
379 
380 	if (WARN(num_entries > max_entries,
381 		 "First entry = %d; Num entries = %d (max=%d)\n",
382 		 first_entry, num_entries, max_entries))
383 		num_entries = max_entries;
384 
385 	scratch_pte = vm->scratch[0].encode;
386 	for (i = 0; i < num_entries; i++)
387 		iowrite32(scratch_pte, &gtt_base[i]);
388 }
389 
390 static void i915_ggtt_insert_page(struct i915_address_space *vm,
391 				  dma_addr_t addr,
392 				  u64 offset,
393 				  enum i915_cache_level cache_level,
394 				  u32 unused)
395 {
396 	unsigned int flags = (cache_level == I915_CACHE_NONE) ?
397 		AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
398 
399 	intel_gtt_insert_page(addr, offset >> PAGE_SHIFT, flags);
400 }
401 
402 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
403 				     struct i915_vma *vma,
404 				     enum i915_cache_level cache_level,
405 				     u32 unused)
406 {
407 	unsigned int flags = (cache_level == I915_CACHE_NONE) ?
408 		AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
409 
410 	intel_gtt_insert_sg_entries(vma->pages, vma->node.start >> PAGE_SHIFT,
411 				    flags);
412 }
413 
414 static void i915_ggtt_clear_range(struct i915_address_space *vm,
415 				  u64 start, u64 length)
416 {
417 	intel_gtt_clear_range(start >> PAGE_SHIFT, length >> PAGE_SHIFT);
418 }
419 
420 static int ggtt_bind_vma(struct i915_vma *vma,
421 			 enum i915_cache_level cache_level,
422 			 u32 flags)
423 {
424 	struct drm_i915_gem_object *obj = vma->obj;
425 	u32 pte_flags;
426 
427 	/* Applicable to VLV (gen8+ do not support RO in the GGTT) */
428 	pte_flags = 0;
429 	if (i915_gem_object_is_readonly(obj))
430 		pte_flags |= PTE_READ_ONLY;
431 
432 	vma->vm->insert_entries(vma->vm, vma, cache_level, pte_flags);
433 
434 	vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
435 
436 	/*
437 	 * Without aliasing PPGTT there's no difference between
438 	 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
439 	 * upgrade to both bound if we bind either to avoid double-binding.
440 	 */
441 	atomic_or(I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND, &vma->flags);
442 
443 	return 0;
444 }
445 
446 static void ggtt_unbind_vma(struct i915_vma *vma)
447 {
448 	vma->vm->clear_range(vma->vm, vma->node.start, vma->size);
449 }
450 
451 static int ggtt_reserve_guc_top(struct i915_ggtt *ggtt)
452 {
453 	u64 size;
454 	int ret;
455 
456 	if (!intel_uc_uses_guc(&ggtt->vm.gt->uc))
457 		return 0;
458 
459 	GEM_BUG_ON(ggtt->vm.total <= GUC_GGTT_TOP);
460 	size = ggtt->vm.total - GUC_GGTT_TOP;
461 
462 	ret = i915_gem_gtt_reserve(&ggtt->vm, &ggtt->uc_fw, size,
463 				   GUC_GGTT_TOP, I915_COLOR_UNEVICTABLE,
464 				   PIN_NOEVICT);
465 	if (ret)
466 		drm_dbg(&ggtt->vm.i915->drm,
467 			"Failed to reserve top of GGTT for GuC\n");
468 
469 	return ret;
470 }
471 
472 static void ggtt_release_guc_top(struct i915_ggtt *ggtt)
473 {
474 	if (drm_mm_node_allocated(&ggtt->uc_fw))
475 		drm_mm_remove_node(&ggtt->uc_fw);
476 }
477 
478 static void cleanup_init_ggtt(struct i915_ggtt *ggtt)
479 {
480 	ggtt_release_guc_top(ggtt);
481 	if (drm_mm_node_allocated(&ggtt->error_capture))
482 		drm_mm_remove_node(&ggtt->error_capture);
483 	mutex_destroy(&ggtt->error_mutex);
484 }
485 
486 static int init_ggtt(struct i915_ggtt *ggtt)
487 {
488 	/*
489 	 * Let GEM Manage all of the aperture.
490 	 *
491 	 * However, leave one page at the end still bound to the scratch page.
492 	 * There are a number of places where the hardware apparently prefetches
493 	 * past the end of the object, and we've seen multiple hangs with the
494 	 * GPU head pointer stuck in a batchbuffer bound at the last page of the
495 	 * aperture.  One page should be enough to keep any prefetching inside
496 	 * of the aperture.
497 	 */
498 	unsigned long hole_start, hole_end;
499 	struct drm_mm_node *entry;
500 	int ret;
501 
502 	/*
503 	 * GuC requires all resources that we're sharing with it to be placed in
504 	 * non-WOPCM memory. If GuC is not present or not in use we still need a
505 	 * small bias as ring wraparound at offset 0 sometimes hangs. No idea
506 	 * why.
507 	 */
508 	ggtt->pin_bias = max_t(u32, I915_GTT_PAGE_SIZE,
509 			       intel_wopcm_guc_size(&ggtt->vm.i915->wopcm));
510 
511 	ret = intel_vgt_balloon(ggtt);
512 	if (ret)
513 		return ret;
514 
515 	mutex_init(&ggtt->error_mutex);
516 	if (ggtt->mappable_end) {
517 		/* Reserve a mappable slot for our lockless error capture */
518 		ret = drm_mm_insert_node_in_range(&ggtt->vm.mm,
519 						  &ggtt->error_capture,
520 						  PAGE_SIZE, 0,
521 						  I915_COLOR_UNEVICTABLE,
522 						  0, ggtt->mappable_end,
523 						  DRM_MM_INSERT_LOW);
524 		if (ret)
525 			return ret;
526 	}
527 
528 	/*
529 	 * The upper portion of the GuC address space has a sizeable hole
530 	 * (several MB) that is inaccessible by GuC. Reserve this range within
531 	 * GGTT as it can comfortably hold GuC/HuC firmware images.
532 	 */
533 	ret = ggtt_reserve_guc_top(ggtt);
534 	if (ret)
535 		goto err;
536 
537 	/* Clear any non-preallocated blocks */
538 	drm_mm_for_each_hole(entry, &ggtt->vm.mm, hole_start, hole_end) {
539 		drm_dbg_kms(&ggtt->vm.i915->drm,
540 			    "clearing unused GTT space: [%lx, %lx]\n",
541 			    hole_start, hole_end);
542 		ggtt->vm.clear_range(&ggtt->vm, hole_start,
543 				     hole_end - hole_start);
544 	}
545 
546 	/* And finally clear the reserved guard page */
547 	ggtt->vm.clear_range(&ggtt->vm, ggtt->vm.total - PAGE_SIZE, PAGE_SIZE);
548 
549 	return 0;
550 
551 err:
552 	cleanup_init_ggtt(ggtt);
553 	return ret;
554 }
555 
556 static int aliasing_gtt_bind_vma(struct i915_vma *vma,
557 				 enum i915_cache_level cache_level,
558 				 u32 flags)
559 {
560 	u32 pte_flags;
561 	int ret;
562 
563 	/* Currently applicable only to VLV */
564 	pte_flags = 0;
565 	if (i915_gem_object_is_readonly(vma->obj))
566 		pte_flags |= PTE_READ_ONLY;
567 
568 	if (flags & I915_VMA_LOCAL_BIND) {
569 		struct i915_ppgtt *alias = i915_vm_to_ggtt(vma->vm)->alias;
570 
571 		if (flags & I915_VMA_ALLOC) {
572 			ret = alias->vm.allocate_va_range(&alias->vm,
573 							  vma->node.start,
574 							  vma->size);
575 			if (ret)
576 				return ret;
577 
578 			set_bit(I915_VMA_ALLOC_BIT, __i915_vma_flags(vma));
579 		}
580 
581 		GEM_BUG_ON(!test_bit(I915_VMA_ALLOC_BIT,
582 				     __i915_vma_flags(vma)));
583 		alias->vm.insert_entries(&alias->vm, vma,
584 					 cache_level, pte_flags);
585 	}
586 
587 	if (flags & I915_VMA_GLOBAL_BIND)
588 		vma->vm->insert_entries(vma->vm, vma, cache_level, pte_flags);
589 
590 	return 0;
591 }
592 
593 static void aliasing_gtt_unbind_vma(struct i915_vma *vma)
594 {
595 	if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) {
596 		struct i915_address_space *vm = vma->vm;
597 
598 		vm->clear_range(vm, vma->node.start, vma->size);
599 	}
600 
601 	if (test_and_clear_bit(I915_VMA_ALLOC_BIT, __i915_vma_flags(vma))) {
602 		struct i915_address_space *vm =
603 			&i915_vm_to_ggtt(vma->vm)->alias->vm;
604 
605 		vm->clear_range(vm, vma->node.start, vma->size);
606 	}
607 }
608 
609 static int init_aliasing_ppgtt(struct i915_ggtt *ggtt)
610 {
611 	struct i915_ppgtt *ppgtt;
612 	int err;
613 
614 	ppgtt = i915_ppgtt_create(ggtt->vm.gt);
615 	if (IS_ERR(ppgtt))
616 		return PTR_ERR(ppgtt);
617 
618 	if (GEM_WARN_ON(ppgtt->vm.total < ggtt->vm.total)) {
619 		err = -ENODEV;
620 		goto err_ppgtt;
621 	}
622 
623 	/*
624 	 * Note we only pre-allocate as far as the end of the global
625 	 * GTT. On 48b / 4-level page-tables, the difference is very,
626 	 * very significant! We have to preallocate as GVT/vgpu does
627 	 * not like the page directory disappearing.
628 	 */
629 	err = ppgtt->vm.allocate_va_range(&ppgtt->vm, 0, ggtt->vm.total);
630 	if (err)
631 		goto err_ppgtt;
632 
633 	ggtt->alias = ppgtt;
634 	ggtt->vm.bind_async_flags |= ppgtt->vm.bind_async_flags;
635 
636 	GEM_BUG_ON(ggtt->vm.vma_ops.bind_vma != ggtt_bind_vma);
637 	ggtt->vm.vma_ops.bind_vma = aliasing_gtt_bind_vma;
638 
639 	GEM_BUG_ON(ggtt->vm.vma_ops.unbind_vma != ggtt_unbind_vma);
640 	ggtt->vm.vma_ops.unbind_vma = aliasing_gtt_unbind_vma;
641 
642 	return 0;
643 
644 err_ppgtt:
645 	i915_vm_put(&ppgtt->vm);
646 	return err;
647 }
648 
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 
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 
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  */
716 void i915_ggtt_driver_release(struct drm_i915_private *i915)
717 {
718 	struct i915_ggtt *ggtt = &i915->ggtt;
719 	struct pagevec *pvec;
720 
721 	fini_aliasing_ppgtt(ggtt);
722 
723 	intel_ggtt_fini_fences(ggtt);
724 	ggtt_cleanup_hw(ggtt);
725 
726 	pvec = &i915->mm.wc_stash.pvec;
727 	if (pvec->nr) {
728 		set_pages_array_wb(pvec->pages, pvec->nr);
729 		__pagevec_release(pvec);
730 	}
731 }
732 
733 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
734 {
735 	snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
736 	snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
737 	return snb_gmch_ctl << 20;
738 }
739 
740 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
741 {
742 	bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
743 	bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
744 	if (bdw_gmch_ctl)
745 		bdw_gmch_ctl = 1 << bdw_gmch_ctl;
746 
747 #ifdef CONFIG_X86_32
748 	/* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * I915_GTT_PAGE_SIZE */
749 	if (bdw_gmch_ctl > 4)
750 		bdw_gmch_ctl = 4;
751 #endif
752 
753 	return bdw_gmch_ctl << 20;
754 }
755 
756 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
757 {
758 	gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
759 	gmch_ctrl &= SNB_GMCH_GGMS_MASK;
760 
761 	if (gmch_ctrl)
762 		return 1 << (20 + gmch_ctrl);
763 
764 	return 0;
765 }
766 
767 static int ggtt_probe_common(struct i915_ggtt *ggtt, u64 size)
768 {
769 	struct drm_i915_private *i915 = ggtt->vm.i915;
770 	struct pci_dev *pdev = i915->drm.pdev;
771 	phys_addr_t phys_addr;
772 	int ret;
773 
774 	/* For Modern GENs the PTEs and register space are split in the BAR */
775 	phys_addr = pci_resource_start(pdev, 0) + pci_resource_len(pdev, 0) / 2;
776 
777 	/*
778 	 * On BXT+/CNL+ writes larger than 64 bit to the GTT pagetable range
779 	 * will be dropped. For WC mappings in general we have 64 byte burst
780 	 * writes when the WC buffer is flushed, so we can't use it, but have to
781 	 * resort to an uncached mapping. The WC issue is easily caught by the
782 	 * readback check when writing GTT PTE entries.
783 	 */
784 	if (IS_GEN9_LP(i915) || INTEL_GEN(i915) >= 10)
785 		ggtt->gsm = ioremap(phys_addr, size);
786 	else
787 		ggtt->gsm = ioremap_wc(phys_addr, size);
788 	if (!ggtt->gsm) {
789 		drm_err(&i915->drm, "Failed to map the ggtt page table\n");
790 		return -ENOMEM;
791 	}
792 
793 	ret = setup_scratch_page(&ggtt->vm, GFP_DMA32);
794 	if (ret) {
795 		drm_err(&i915->drm, "Scratch setup failed\n");
796 		/* iounmap will also get called at remove, but meh */
797 		iounmap(ggtt->gsm);
798 		return ret;
799 	}
800 
801 	ggtt->vm.scratch[0].encode =
802 		ggtt->vm.pte_encode(px_dma(&ggtt->vm.scratch[0]),
803 				    I915_CACHE_NONE, 0);
804 
805 	return 0;
806 }
807 
808 int ggtt_set_pages(struct i915_vma *vma)
809 {
810 	int ret;
811 
812 	GEM_BUG_ON(vma->pages);
813 
814 	ret = i915_get_ggtt_vma_pages(vma);
815 	if (ret)
816 		return ret;
817 
818 	vma->page_sizes = vma->obj->mm.page_sizes;
819 
820 	return 0;
821 }
822 
823 static void gen6_gmch_remove(struct i915_address_space *vm)
824 {
825 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
826 
827 	iounmap(ggtt->gsm);
828 	cleanup_scratch_page(vm);
829 }
830 
831 static struct resource pci_resource(struct pci_dev *pdev, int bar)
832 {
833 	return (struct resource)DEFINE_RES_MEM(pci_resource_start(pdev, bar),
834 					       pci_resource_len(pdev, bar));
835 }
836 
837 static int gen8_gmch_probe(struct i915_ggtt *ggtt)
838 {
839 	struct drm_i915_private *i915 = ggtt->vm.i915;
840 	struct pci_dev *pdev = i915->drm.pdev;
841 	unsigned int size;
842 	u16 snb_gmch_ctl;
843 
844 	/* TODO: We're not aware of mappable constraints on gen8 yet */
845 	if (!IS_DGFX(i915)) {
846 		ggtt->gmadr = pci_resource(pdev, 2);
847 		ggtt->mappable_end = resource_size(&ggtt->gmadr);
848 	}
849 
850 	pci_read_config_word(pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
851 	if (IS_CHERRYVIEW(i915))
852 		size = chv_get_total_gtt_size(snb_gmch_ctl);
853 	else
854 		size = gen8_get_total_gtt_size(snb_gmch_ctl);
855 
856 	ggtt->vm.total = (size / sizeof(gen8_pte_t)) * I915_GTT_PAGE_SIZE;
857 	ggtt->vm.cleanup = gen6_gmch_remove;
858 	ggtt->vm.insert_page = gen8_ggtt_insert_page;
859 	ggtt->vm.clear_range = nop_clear_range;
860 	if (intel_scanout_needs_vtd_wa(i915))
861 		ggtt->vm.clear_range = gen8_ggtt_clear_range;
862 
863 	ggtt->vm.insert_entries = gen8_ggtt_insert_entries;
864 
865 	/* Serialize GTT updates with aperture access on BXT if VT-d is on. */
866 	if (intel_ggtt_update_needs_vtd_wa(i915) ||
867 	    IS_CHERRYVIEW(i915) /* fails with concurrent use/update */) {
868 		ggtt->vm.insert_entries = bxt_vtd_ggtt_insert_entries__BKL;
869 		ggtt->vm.insert_page    = bxt_vtd_ggtt_insert_page__BKL;
870 		ggtt->vm.bind_async_flags =
871 			I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
872 	}
873 
874 	ggtt->invalidate = gen8_ggtt_invalidate;
875 
876 	ggtt->vm.vma_ops.bind_vma    = ggtt_bind_vma;
877 	ggtt->vm.vma_ops.unbind_vma  = ggtt_unbind_vma;
878 	ggtt->vm.vma_ops.set_pages   = ggtt_set_pages;
879 	ggtt->vm.vma_ops.clear_pages = clear_pages;
880 
881 	ggtt->vm.pte_encode = gen8_ggtt_pte_encode;
882 
883 	setup_private_pat(ggtt->vm.gt->uncore);
884 
885 	return ggtt_probe_common(ggtt, size);
886 }
887 
888 static u64 snb_pte_encode(dma_addr_t addr,
889 			  enum i915_cache_level level,
890 			  u32 flags)
891 {
892 	gen6_pte_t pte = GEN6_PTE_ADDR_ENCODE(addr) | GEN6_PTE_VALID;
893 
894 	switch (level) {
895 	case I915_CACHE_L3_LLC:
896 	case I915_CACHE_LLC:
897 		pte |= GEN6_PTE_CACHE_LLC;
898 		break;
899 	case I915_CACHE_NONE:
900 		pte |= GEN6_PTE_UNCACHED;
901 		break;
902 	default:
903 		MISSING_CASE(level);
904 	}
905 
906 	return pte;
907 }
908 
909 static u64 ivb_pte_encode(dma_addr_t addr,
910 			  enum i915_cache_level level,
911 			  u32 flags)
912 {
913 	gen6_pte_t pte = GEN6_PTE_ADDR_ENCODE(addr) | GEN6_PTE_VALID;
914 
915 	switch (level) {
916 	case I915_CACHE_L3_LLC:
917 		pte |= GEN7_PTE_CACHE_L3_LLC;
918 		break;
919 	case I915_CACHE_LLC:
920 		pte |= GEN6_PTE_CACHE_LLC;
921 		break;
922 	case I915_CACHE_NONE:
923 		pte |= GEN6_PTE_UNCACHED;
924 		break;
925 	default:
926 		MISSING_CASE(level);
927 	}
928 
929 	return pte;
930 }
931 
932 static u64 byt_pte_encode(dma_addr_t addr,
933 			  enum i915_cache_level level,
934 			  u32 flags)
935 {
936 	gen6_pte_t pte = GEN6_PTE_ADDR_ENCODE(addr) | GEN6_PTE_VALID;
937 
938 	if (!(flags & PTE_READ_ONLY))
939 		pte |= BYT_PTE_WRITEABLE;
940 
941 	if (level != I915_CACHE_NONE)
942 		pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
943 
944 	return pte;
945 }
946 
947 static u64 hsw_pte_encode(dma_addr_t addr,
948 			  enum i915_cache_level level,
949 			  u32 flags)
950 {
951 	gen6_pte_t pte = HSW_PTE_ADDR_ENCODE(addr) | GEN6_PTE_VALID;
952 
953 	if (level != I915_CACHE_NONE)
954 		pte |= HSW_WB_LLC_AGE3;
955 
956 	return pte;
957 }
958 
959 static u64 iris_pte_encode(dma_addr_t addr,
960 			   enum i915_cache_level level,
961 			   u32 flags)
962 {
963 	gen6_pte_t pte = HSW_PTE_ADDR_ENCODE(addr) | GEN6_PTE_VALID;
964 
965 	switch (level) {
966 	case I915_CACHE_NONE:
967 		break;
968 	case I915_CACHE_WT:
969 		pte |= HSW_WT_ELLC_LLC_AGE3;
970 		break;
971 	default:
972 		pte |= HSW_WB_ELLC_LLC_AGE3;
973 		break;
974 	}
975 
976 	return pte;
977 }
978 
979 static int gen6_gmch_probe(struct i915_ggtt *ggtt)
980 {
981 	struct drm_i915_private *i915 = ggtt->vm.i915;
982 	struct pci_dev *pdev = i915->drm.pdev;
983 	unsigned int size;
984 	u16 snb_gmch_ctl;
985 
986 	ggtt->gmadr = pci_resource(pdev, 2);
987 	ggtt->mappable_end = resource_size(&ggtt->gmadr);
988 
989 	/*
990 	 * 64/512MB is the current min/max we actually know of, but this is
991 	 * just a coarse sanity check.
992 	 */
993 	if (ggtt->mappable_end < (64<<20) || ggtt->mappable_end > (512<<20)) {
994 		drm_err(&i915->drm, "Unknown GMADR size (%pa)\n",
995 			&ggtt->mappable_end);
996 		return -ENXIO;
997 	}
998 
999 	pci_read_config_word(pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
1000 
1001 	size = gen6_get_total_gtt_size(snb_gmch_ctl);
1002 	ggtt->vm.total = (size / sizeof(gen6_pte_t)) * I915_GTT_PAGE_SIZE;
1003 
1004 	ggtt->vm.clear_range = nop_clear_range;
1005 	if (!HAS_FULL_PPGTT(i915) || intel_scanout_needs_vtd_wa(i915))
1006 		ggtt->vm.clear_range = gen6_ggtt_clear_range;
1007 	ggtt->vm.insert_page = gen6_ggtt_insert_page;
1008 	ggtt->vm.insert_entries = gen6_ggtt_insert_entries;
1009 	ggtt->vm.cleanup = gen6_gmch_remove;
1010 
1011 	ggtt->invalidate = gen6_ggtt_invalidate;
1012 
1013 	if (HAS_EDRAM(i915))
1014 		ggtt->vm.pte_encode = iris_pte_encode;
1015 	else if (IS_HASWELL(i915))
1016 		ggtt->vm.pte_encode = hsw_pte_encode;
1017 	else if (IS_VALLEYVIEW(i915))
1018 		ggtt->vm.pte_encode = byt_pte_encode;
1019 	else if (INTEL_GEN(i915) >= 7)
1020 		ggtt->vm.pte_encode = ivb_pte_encode;
1021 	else
1022 		ggtt->vm.pte_encode = snb_pte_encode;
1023 
1024 	ggtt->vm.vma_ops.bind_vma    = ggtt_bind_vma;
1025 	ggtt->vm.vma_ops.unbind_vma  = ggtt_unbind_vma;
1026 	ggtt->vm.vma_ops.set_pages   = ggtt_set_pages;
1027 	ggtt->vm.vma_ops.clear_pages = clear_pages;
1028 
1029 	return ggtt_probe_common(ggtt, size);
1030 }
1031 
1032 static void i915_gmch_remove(struct i915_address_space *vm)
1033 {
1034 	intel_gmch_remove();
1035 }
1036 
1037 static int i915_gmch_probe(struct i915_ggtt *ggtt)
1038 {
1039 	struct drm_i915_private *i915 = ggtt->vm.i915;
1040 	phys_addr_t gmadr_base;
1041 	int ret;
1042 
1043 	ret = intel_gmch_probe(i915->bridge_dev, i915->drm.pdev, NULL);
1044 	if (!ret) {
1045 		drm_err(&i915->drm, "failed to set up gmch\n");
1046 		return -EIO;
1047 	}
1048 
1049 	intel_gtt_get(&ggtt->vm.total, &gmadr_base, &ggtt->mappable_end);
1050 
1051 	ggtt->gmadr =
1052 		(struct resource)DEFINE_RES_MEM(gmadr_base, ggtt->mappable_end);
1053 
1054 	ggtt->do_idle_maps = needs_idle_maps(i915);
1055 	ggtt->vm.insert_page = i915_ggtt_insert_page;
1056 	ggtt->vm.insert_entries = i915_ggtt_insert_entries;
1057 	ggtt->vm.clear_range = i915_ggtt_clear_range;
1058 	ggtt->vm.cleanup = i915_gmch_remove;
1059 
1060 	ggtt->invalidate = gmch_ggtt_invalidate;
1061 
1062 	ggtt->vm.vma_ops.bind_vma    = ggtt_bind_vma;
1063 	ggtt->vm.vma_ops.unbind_vma  = ggtt_unbind_vma;
1064 	ggtt->vm.vma_ops.set_pages   = ggtt_set_pages;
1065 	ggtt->vm.vma_ops.clear_pages = clear_pages;
1066 
1067 	if (unlikely(ggtt->do_idle_maps))
1068 		drm_notice(&i915->drm,
1069 			   "Applying Ironlake quirks for intel_iommu\n");
1070 
1071 	return 0;
1072 }
1073 
1074 static int ggtt_probe_hw(struct i915_ggtt *ggtt, struct intel_gt *gt)
1075 {
1076 	struct drm_i915_private *i915 = gt->i915;
1077 	int ret;
1078 
1079 	ggtt->vm.gt = gt;
1080 	ggtt->vm.i915 = i915;
1081 	ggtt->vm.dma = &i915->drm.pdev->dev;
1082 
1083 	if (INTEL_GEN(i915) <= 5)
1084 		ret = i915_gmch_probe(ggtt);
1085 	else if (INTEL_GEN(i915) < 8)
1086 		ret = gen6_gmch_probe(ggtt);
1087 	else
1088 		ret = gen8_gmch_probe(ggtt);
1089 	if (ret)
1090 		return ret;
1091 
1092 	if ((ggtt->vm.total - 1) >> 32) {
1093 		drm_err(&i915->drm,
1094 			"We never expected a Global GTT with more than 32bits"
1095 			" of address space! Found %lldM!\n",
1096 			ggtt->vm.total >> 20);
1097 		ggtt->vm.total = 1ULL << 32;
1098 		ggtt->mappable_end =
1099 			min_t(u64, ggtt->mappable_end, ggtt->vm.total);
1100 	}
1101 
1102 	if (ggtt->mappable_end > ggtt->vm.total) {
1103 		drm_err(&i915->drm,
1104 			"mappable aperture extends past end of GGTT,"
1105 			" aperture=%pa, total=%llx\n",
1106 			&ggtt->mappable_end, ggtt->vm.total);
1107 		ggtt->mappable_end = ggtt->vm.total;
1108 	}
1109 
1110 	/* GMADR is the PCI mmio aperture into the global GTT. */
1111 	drm_dbg(&i915->drm, "GGTT size = %lluM\n", ggtt->vm.total >> 20);
1112 	drm_dbg(&i915->drm, "GMADR size = %lluM\n",
1113 		(u64)ggtt->mappable_end >> 20);
1114 	drm_dbg(&i915->drm, "DSM size = %lluM\n",
1115 		(u64)resource_size(&intel_graphics_stolen_res) >> 20);
1116 
1117 	return 0;
1118 }
1119 
1120 /**
1121  * i915_ggtt_probe_hw - Probe GGTT hardware location
1122  * @i915: i915 device
1123  */
1124 int i915_ggtt_probe_hw(struct drm_i915_private *i915)
1125 {
1126 	int ret;
1127 
1128 	ret = ggtt_probe_hw(&i915->ggtt, &i915->gt);
1129 	if (ret)
1130 		return ret;
1131 
1132 	if (intel_vtd_active())
1133 		drm_info(&i915->drm, "VT-d active for gfx access\n");
1134 
1135 	return 0;
1136 }
1137 
1138 int i915_ggtt_enable_hw(struct drm_i915_private *i915)
1139 {
1140 	if (INTEL_GEN(i915) < 6 && !intel_enable_gtt())
1141 		return -EIO;
1142 
1143 	return 0;
1144 }
1145 
1146 void i915_ggtt_enable_guc(struct i915_ggtt *ggtt)
1147 {
1148 	GEM_BUG_ON(ggtt->invalidate != gen8_ggtt_invalidate);
1149 
1150 	ggtt->invalidate = guc_ggtt_invalidate;
1151 
1152 	ggtt->invalidate(ggtt);
1153 }
1154 
1155 void i915_ggtt_disable_guc(struct i915_ggtt *ggtt)
1156 {
1157 	/* XXX Temporary pardon for error unload */
1158 	if (ggtt->invalidate == gen8_ggtt_invalidate)
1159 		return;
1160 
1161 	/* We should only be called after i915_ggtt_enable_guc() */
1162 	GEM_BUG_ON(ggtt->invalidate != guc_ggtt_invalidate);
1163 
1164 	ggtt->invalidate = gen8_ggtt_invalidate;
1165 
1166 	ggtt->invalidate(ggtt);
1167 }
1168 
1169 void i915_ggtt_resume(struct i915_ggtt *ggtt)
1170 {
1171 	struct i915_vma *vma;
1172 	bool flush = false;
1173 	int open;
1174 
1175 	intel_gt_check_and_clear_faults(ggtt->vm.gt);
1176 
1177 	/* First fill our portion of the GTT with scratch pages */
1178 	ggtt->vm.clear_range(&ggtt->vm, 0, ggtt->vm.total);
1179 
1180 	/* Skip rewriting PTE on VMA unbind. */
1181 	open = atomic_xchg(&ggtt->vm.open, 0);
1182 
1183 	/* clflush objects bound into the GGTT and rebind them. */
1184 	list_for_each_entry(vma, &ggtt->vm.bound_list, vm_link) {
1185 		struct drm_i915_gem_object *obj = vma->obj;
1186 
1187 		if (!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
1188 			continue;
1189 
1190 		clear_bit(I915_VMA_GLOBAL_BIND_BIT, __i915_vma_flags(vma));
1191 		WARN_ON(i915_vma_bind(vma,
1192 				      obj ? obj->cache_level : 0,
1193 				      PIN_GLOBAL, NULL));
1194 		if (obj) { /* only used during resume => exclusive access */
1195 			flush |= fetch_and_zero(&obj->write_domain);
1196 			obj->read_domains |= I915_GEM_DOMAIN_GTT;
1197 		}
1198 	}
1199 
1200 	atomic_set(&ggtt->vm.open, open);
1201 	ggtt->invalidate(ggtt);
1202 
1203 	if (flush)
1204 		wbinvd_on_all_cpus();
1205 
1206 	if (INTEL_GEN(ggtt->vm.i915) >= 8)
1207 		setup_private_pat(ggtt->vm.gt->uncore);
1208 
1209 	intel_ggtt_restore_fences(ggtt);
1210 }
1211 
1212 static struct scatterlist *
1213 rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
1214 	     unsigned int width, unsigned int height,
1215 	     unsigned int stride,
1216 	     struct sg_table *st, struct scatterlist *sg)
1217 {
1218 	unsigned int column, row;
1219 	unsigned int src_idx;
1220 
1221 	for (column = 0; column < width; column++) {
1222 		src_idx = stride * (height - 1) + column + offset;
1223 		for (row = 0; row < height; row++) {
1224 			st->nents++;
1225 			/*
1226 			 * We don't need the pages, but need to initialize
1227 			 * the entries so the sg list can be happily traversed.
1228 			 * The only thing we need are DMA addresses.
1229 			 */
1230 			sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0);
1231 			sg_dma_address(sg) =
1232 				i915_gem_object_get_dma_address(obj, src_idx);
1233 			sg_dma_len(sg) = I915_GTT_PAGE_SIZE;
1234 			sg = sg_next(sg);
1235 			src_idx -= stride;
1236 		}
1237 	}
1238 
1239 	return sg;
1240 }
1241 
1242 static noinline struct sg_table *
1243 intel_rotate_pages(struct intel_rotation_info *rot_info,
1244 		   struct drm_i915_gem_object *obj)
1245 {
1246 	unsigned int size = intel_rotation_info_size(rot_info);
1247 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
1248 	struct sg_table *st;
1249 	struct scatterlist *sg;
1250 	int ret = -ENOMEM;
1251 	int i;
1252 
1253 	/* Allocate target SG list. */
1254 	st = kmalloc(sizeof(*st), GFP_KERNEL);
1255 	if (!st)
1256 		goto err_st_alloc;
1257 
1258 	ret = sg_alloc_table(st, size, GFP_KERNEL);
1259 	if (ret)
1260 		goto err_sg_alloc;
1261 
1262 	st->nents = 0;
1263 	sg = st->sgl;
1264 
1265 	for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++) {
1266 		sg = rotate_pages(obj, rot_info->plane[i].offset,
1267 				  rot_info->plane[i].width, rot_info->plane[i].height,
1268 				  rot_info->plane[i].stride, st, sg);
1269 	}
1270 
1271 	return st;
1272 
1273 err_sg_alloc:
1274 	kfree(st);
1275 err_st_alloc:
1276 
1277 	drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",
1278 		obj->base.size, rot_info->plane[0].width,
1279 		rot_info->plane[0].height, size);
1280 
1281 	return ERR_PTR(ret);
1282 }
1283 
1284 static struct scatterlist *
1285 remap_pages(struct drm_i915_gem_object *obj, unsigned int offset,
1286 	    unsigned int width, unsigned int height,
1287 	    unsigned int stride,
1288 	    struct sg_table *st, struct scatterlist *sg)
1289 {
1290 	unsigned int row;
1291 
1292 	for (row = 0; row < height; row++) {
1293 		unsigned int left = width * I915_GTT_PAGE_SIZE;
1294 
1295 		while (left) {
1296 			dma_addr_t addr;
1297 			unsigned int length;
1298 
1299 			/*
1300 			 * We don't need the pages, but need to initialize
1301 			 * the entries so the sg list can be happily traversed.
1302 			 * The only thing we need are DMA addresses.
1303 			 */
1304 
1305 			addr = i915_gem_object_get_dma_address_len(obj, offset, &length);
1306 
1307 			length = min(left, length);
1308 
1309 			st->nents++;
1310 
1311 			sg_set_page(sg, NULL, length, 0);
1312 			sg_dma_address(sg) = addr;
1313 			sg_dma_len(sg) = length;
1314 			sg = sg_next(sg);
1315 
1316 			offset += length / I915_GTT_PAGE_SIZE;
1317 			left -= length;
1318 		}
1319 
1320 		offset += stride - width;
1321 	}
1322 
1323 	return sg;
1324 }
1325 
1326 static noinline struct sg_table *
1327 intel_remap_pages(struct intel_remapped_info *rem_info,
1328 		  struct drm_i915_gem_object *obj)
1329 {
1330 	unsigned int size = intel_remapped_info_size(rem_info);
1331 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
1332 	struct sg_table *st;
1333 	struct scatterlist *sg;
1334 	int ret = -ENOMEM;
1335 	int i;
1336 
1337 	/* Allocate target SG list. */
1338 	st = kmalloc(sizeof(*st), GFP_KERNEL);
1339 	if (!st)
1340 		goto err_st_alloc;
1341 
1342 	ret = sg_alloc_table(st, size, GFP_KERNEL);
1343 	if (ret)
1344 		goto err_sg_alloc;
1345 
1346 	st->nents = 0;
1347 	sg = st->sgl;
1348 
1349 	for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) {
1350 		sg = remap_pages(obj, rem_info->plane[i].offset,
1351 				 rem_info->plane[i].width, rem_info->plane[i].height,
1352 				 rem_info->plane[i].stride, st, sg);
1353 	}
1354 
1355 	i915_sg_trim(st);
1356 
1357 	return st;
1358 
1359 err_sg_alloc:
1360 	kfree(st);
1361 err_st_alloc:
1362 
1363 	drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n",
1364 		obj->base.size, rem_info->plane[0].width,
1365 		rem_info->plane[0].height, size);
1366 
1367 	return ERR_PTR(ret);
1368 }
1369 
1370 static noinline struct sg_table *
1371 intel_partial_pages(const struct i915_ggtt_view *view,
1372 		    struct drm_i915_gem_object *obj)
1373 {
1374 	struct sg_table *st;
1375 	struct scatterlist *sg, *iter;
1376 	unsigned int count = view->partial.size;
1377 	unsigned int offset;
1378 	int ret = -ENOMEM;
1379 
1380 	st = kmalloc(sizeof(*st), GFP_KERNEL);
1381 	if (!st)
1382 		goto err_st_alloc;
1383 
1384 	ret = sg_alloc_table(st, count, GFP_KERNEL);
1385 	if (ret)
1386 		goto err_sg_alloc;
1387 
1388 	iter = i915_gem_object_get_sg(obj, view->partial.offset, &offset);
1389 	GEM_BUG_ON(!iter);
1390 
1391 	sg = st->sgl;
1392 	st->nents = 0;
1393 	do {
1394 		unsigned int len;
1395 
1396 		len = min(iter->length - (offset << PAGE_SHIFT),
1397 			  count << PAGE_SHIFT);
1398 		sg_set_page(sg, NULL, len, 0);
1399 		sg_dma_address(sg) =
1400 			sg_dma_address(iter) + (offset << PAGE_SHIFT);
1401 		sg_dma_len(sg) = len;
1402 
1403 		st->nents++;
1404 		count -= len >> PAGE_SHIFT;
1405 		if (count == 0) {
1406 			sg_mark_end(sg);
1407 			i915_sg_trim(st); /* Drop any unused tail entries. */
1408 
1409 			return st;
1410 		}
1411 
1412 		sg = __sg_next(sg);
1413 		iter = __sg_next(iter);
1414 		offset = 0;
1415 	} while (1);
1416 
1417 err_sg_alloc:
1418 	kfree(st);
1419 err_st_alloc:
1420 	return ERR_PTR(ret);
1421 }
1422 
1423 static int
1424 i915_get_ggtt_vma_pages(struct i915_vma *vma)
1425 {
1426 	int ret;
1427 
1428 	/*
1429 	 * The vma->pages are only valid within the lifespan of the borrowed
1430 	 * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so
1431 	 * must be the vma->pages. A simple rule is that vma->pages must only
1432 	 * be accessed when the obj->mm.pages are pinned.
1433 	 */
1434 	GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));
1435 
1436 	switch (vma->ggtt_view.type) {
1437 	default:
1438 		GEM_BUG_ON(vma->ggtt_view.type);
1439 		/* fall through */
1440 	case I915_GGTT_VIEW_NORMAL:
1441 		vma->pages = vma->obj->mm.pages;
1442 		return 0;
1443 
1444 	case I915_GGTT_VIEW_ROTATED:
1445 		vma->pages =
1446 			intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj);
1447 		break;
1448 
1449 	case I915_GGTT_VIEW_REMAPPED:
1450 		vma->pages =
1451 			intel_remap_pages(&vma->ggtt_view.remapped, vma->obj);
1452 		break;
1453 
1454 	case I915_GGTT_VIEW_PARTIAL:
1455 		vma->pages = intel_partial_pages(&vma->ggtt_view, vma->obj);
1456 		break;
1457 	}
1458 
1459 	ret = 0;
1460 	if (IS_ERR(vma->pages)) {
1461 		ret = PTR_ERR(vma->pages);
1462 		vma->pages = NULL;
1463 		drm_err(&vma->vm->i915->drm,
1464 			"Failed to get pages for VMA view type %u (%d)!\n",
1465 			vma->ggtt_view.type, ret);
1466 	}
1467 	return ret;
1468 }
1469