1 /*
2  * Copyright 2014-2018 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  */
22 #include <linux/dma-buf.h>
23 #include <linux/list.h>
24 #include <linux/pagemap.h>
25 #include <linux/sched/mm.h>
26 #include <linux/sched/task.h>
27 
28 #include "amdgpu_object.h"
29 #include "amdgpu_gem.h"
30 #include "amdgpu_vm.h"
31 #include "amdgpu_amdkfd.h"
32 #include "amdgpu_dma_buf.h"
33 #include <uapi/linux/kfd_ioctl.h>
34 #include "amdgpu_xgmi.h"
35 
36 /* Userptr restore delay, just long enough to allow consecutive VM
37  * changes to accumulate
38  */
39 #define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
40 
41 /* Impose limit on how much memory KFD can use */
42 static struct {
43 	uint64_t max_system_mem_limit;
44 	uint64_t max_ttm_mem_limit;
45 	int64_t system_mem_used;
46 	int64_t ttm_mem_used;
47 	spinlock_t mem_limit_lock;
48 } kfd_mem_limit;
49 
50 static const char * const domain_bit_to_string[] = {
51 		"CPU",
52 		"GTT",
53 		"VRAM",
54 		"GDS",
55 		"GWS",
56 		"OA"
57 };
58 
59 #define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
60 
61 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work);
62 
63 static bool kfd_mem_is_attached(struct amdgpu_vm *avm,
64 		struct kgd_mem *mem)
65 {
66 	struct kfd_mem_attachment *entry;
67 
68 	list_for_each_entry(entry, &mem->attachments, list)
69 		if (entry->bo_va->base.vm == avm)
70 			return true;
71 
72 	return false;
73 }
74 
75 /* Set memory usage limits. Current, limits are
76  *  System (TTM + userptr) memory - 15/16th System RAM
77  *  TTM memory - 3/8th System RAM
78  */
79 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
80 {
81 	struct sysinfo si;
82 	uint64_t mem;
83 
84 	si_meminfo(&si);
85 	mem = si.freeram - si.freehigh;
86 	mem *= si.mem_unit;
87 
88 	spin_lock_init(&kfd_mem_limit.mem_limit_lock);
89 	kfd_mem_limit.max_system_mem_limit = mem - (mem >> 4);
90 	kfd_mem_limit.max_ttm_mem_limit = (mem >> 1) - (mem >> 3);
91 	pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
92 		(kfd_mem_limit.max_system_mem_limit >> 20),
93 		(kfd_mem_limit.max_ttm_mem_limit >> 20));
94 }
95 
96 void amdgpu_amdkfd_reserve_system_mem(uint64_t size)
97 {
98 	kfd_mem_limit.system_mem_used += size;
99 }
100 
101 /* Estimate page table size needed to represent a given memory size
102  *
103  * With 4KB pages, we need one 8 byte PTE for each 4KB of memory
104  * (factor 512, >> 9). With 2MB pages, we need one 8 byte PTE for 2MB
105  * of memory (factor 256K, >> 18). ROCm user mode tries to optimize
106  * for 2MB pages for TLB efficiency. However, small allocations and
107  * fragmented system memory still need some 4KB pages. We choose a
108  * compromise that should work in most cases without reserving too
109  * much memory for page tables unnecessarily (factor 16K, >> 14).
110  */
111 #define ESTIMATE_PT_SIZE(mem_size) ((mem_size) >> 14)
112 
113 static size_t amdgpu_amdkfd_acc_size(uint64_t size)
114 {
115 	size >>= PAGE_SHIFT;
116 	size *= sizeof(dma_addr_t) + sizeof(void *);
117 
118 	return __roundup_pow_of_two(sizeof(struct amdgpu_bo)) +
119 		__roundup_pow_of_two(sizeof(struct ttm_tt)) +
120 		PAGE_ALIGN(size);
121 }
122 
123 /**
124  * amdgpu_amdkfd_reserve_mem_limit() - Decrease available memory by size
125  * of buffer including any reserved for control structures
126  *
127  * @adev: Device to which allocated BO belongs to
128  * @size: Size of buffer, in bytes, encapsulated by B0. This should be
129  * equivalent to amdgpu_bo_size(BO)
130  * @alloc_flag: Flag used in allocating a BO as noted above
131  *
132  * Return: returns -ENOMEM in case of error, ZERO otherwise
133  */
134 static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
135 		uint64_t size, u32 alloc_flag)
136 {
137 	uint64_t reserved_for_pt =
138 		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
139 	size_t acc_size, system_mem_needed, ttm_mem_needed, vram_needed;
140 	int ret = 0;
141 
142 	acc_size = amdgpu_amdkfd_acc_size(size);
143 
144 	vram_needed = 0;
145 	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
146 		system_mem_needed = acc_size + size;
147 		ttm_mem_needed = acc_size + size;
148 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
149 		system_mem_needed = acc_size;
150 		ttm_mem_needed = acc_size;
151 		vram_needed = size;
152 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
153 		system_mem_needed = acc_size + size;
154 		ttm_mem_needed = acc_size;
155 	} else if (alloc_flag &
156 		   (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
157 		    KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
158 		system_mem_needed = acc_size;
159 		ttm_mem_needed = acc_size;
160 	} else {
161 		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
162 		return -ENOMEM;
163 	}
164 
165 	spin_lock(&kfd_mem_limit.mem_limit_lock);
166 
167 	if (kfd_mem_limit.system_mem_used + system_mem_needed >
168 	    kfd_mem_limit.max_system_mem_limit)
169 		pr_debug("Set no_system_mem_limit=1 if using shared memory\n");
170 
171 	if ((kfd_mem_limit.system_mem_used + system_mem_needed >
172 	     kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) ||
173 	    (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
174 	     kfd_mem_limit.max_ttm_mem_limit) ||
175 	    (adev->kfd.vram_used + vram_needed >
176 	     adev->gmc.real_vram_size - reserved_for_pt)) {
177 		ret = -ENOMEM;
178 		goto release;
179 	}
180 
181 	/* Update memory accounting by decreasing available system
182 	 * memory, TTM memory and GPU memory as computed above
183 	 */
184 	adev->kfd.vram_used += vram_needed;
185 	kfd_mem_limit.system_mem_used += system_mem_needed;
186 	kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
187 
188 release:
189 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
190 	return ret;
191 }
192 
193 static void unreserve_mem_limit(struct amdgpu_device *adev,
194 		uint64_t size, u32 alloc_flag)
195 {
196 	size_t acc_size;
197 
198 	acc_size = amdgpu_amdkfd_acc_size(size);
199 
200 	spin_lock(&kfd_mem_limit.mem_limit_lock);
201 
202 	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
203 		kfd_mem_limit.system_mem_used -= (acc_size + size);
204 		kfd_mem_limit.ttm_mem_used -= (acc_size + size);
205 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
206 		kfd_mem_limit.system_mem_used -= acc_size;
207 		kfd_mem_limit.ttm_mem_used -= acc_size;
208 		adev->kfd.vram_used -= size;
209 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
210 		kfd_mem_limit.system_mem_used -= (acc_size + size);
211 		kfd_mem_limit.ttm_mem_used -= acc_size;
212 	} else if (alloc_flag &
213 		   (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
214 		    KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
215 		kfd_mem_limit.system_mem_used -= acc_size;
216 		kfd_mem_limit.ttm_mem_used -= acc_size;
217 	} else {
218 		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
219 		goto release;
220 	}
221 
222 	WARN_ONCE(adev->kfd.vram_used < 0,
223 		  "KFD VRAM memory accounting unbalanced");
224 	WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
225 		  "KFD TTM memory accounting unbalanced");
226 	WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
227 		  "KFD system memory accounting unbalanced");
228 
229 release:
230 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
231 }
232 
233 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
234 {
235 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
236 	u32 alloc_flags = bo->kfd_bo->alloc_flags;
237 	u64 size = amdgpu_bo_size(bo);
238 
239 	unreserve_mem_limit(adev, size, alloc_flags);
240 
241 	kfree(bo->kfd_bo);
242 }
243 
244 /* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
245  *  reservation object.
246  *
247  * @bo: [IN] Remove eviction fence(s) from this BO
248  * @ef: [IN] This eviction fence is removed if it
249  *  is present in the shared list.
250  *
251  * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
252  */
253 static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
254 					struct amdgpu_amdkfd_fence *ef)
255 {
256 	struct dma_resv *resv = bo->tbo.base.resv;
257 	struct dma_resv_list *old, *new;
258 	unsigned int i, j, k;
259 
260 	if (!ef)
261 		return -EINVAL;
262 
263 	old = dma_resv_shared_list(resv);
264 	if (!old)
265 		return 0;
266 
267 	new = kmalloc(struct_size(new, shared, old->shared_max), GFP_KERNEL);
268 	if (!new)
269 		return -ENOMEM;
270 
271 	/* Go through all the shared fences in the resevation object and sort
272 	 * the interesting ones to the end of the list.
273 	 */
274 	for (i = 0, j = old->shared_count, k = 0; i < old->shared_count; ++i) {
275 		struct dma_fence *f;
276 
277 		f = rcu_dereference_protected(old->shared[i],
278 					      dma_resv_held(resv));
279 
280 		if (f->context == ef->base.context)
281 			RCU_INIT_POINTER(new->shared[--j], f);
282 		else
283 			RCU_INIT_POINTER(new->shared[k++], f);
284 	}
285 	new->shared_max = old->shared_max;
286 	new->shared_count = k;
287 
288 	/* Install the new fence list, seqcount provides the barriers */
289 	write_seqcount_begin(&resv->seq);
290 	RCU_INIT_POINTER(resv->fence, new);
291 	write_seqcount_end(&resv->seq);
292 
293 	/* Drop the references to the removed fences or move them to ef_list */
294 	for (i = j; i < old->shared_count; ++i) {
295 		struct dma_fence *f;
296 
297 		f = rcu_dereference_protected(new->shared[i],
298 					      dma_resv_held(resv));
299 		dma_fence_put(f);
300 	}
301 	kfree_rcu(old, rcu);
302 
303 	return 0;
304 }
305 
306 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
307 {
308 	struct amdgpu_bo *root = bo;
309 	struct amdgpu_vm_bo_base *vm_bo;
310 	struct amdgpu_vm *vm;
311 	struct amdkfd_process_info *info;
312 	struct amdgpu_amdkfd_fence *ef;
313 	int ret;
314 
315 	/* we can always get vm_bo from root PD bo.*/
316 	while (root->parent)
317 		root = root->parent;
318 
319 	vm_bo = root->vm_bo;
320 	if (!vm_bo)
321 		return 0;
322 
323 	vm = vm_bo->vm;
324 	if (!vm)
325 		return 0;
326 
327 	info = vm->process_info;
328 	if (!info || !info->eviction_fence)
329 		return 0;
330 
331 	ef = container_of(dma_fence_get(&info->eviction_fence->base),
332 			struct amdgpu_amdkfd_fence, base);
333 
334 	BUG_ON(!dma_resv_trylock(bo->tbo.base.resv));
335 	ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef);
336 	dma_resv_unlock(bo->tbo.base.resv);
337 
338 	dma_fence_put(&ef->base);
339 	return ret;
340 }
341 
342 static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
343 				     bool wait)
344 {
345 	struct ttm_operation_ctx ctx = { false, false };
346 	int ret;
347 
348 	if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
349 		 "Called with userptr BO"))
350 		return -EINVAL;
351 
352 	amdgpu_bo_placement_from_domain(bo, domain);
353 
354 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
355 	if (ret)
356 		goto validate_fail;
357 	if (wait)
358 		amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
359 
360 validate_fail:
361 	return ret;
362 }
363 
364 static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo)
365 {
366 	return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false);
367 }
368 
369 /* vm_validate_pt_pd_bos - Validate page table and directory BOs
370  *
371  * Page directories are not updated here because huge page handling
372  * during page table updates can invalidate page directory entries
373  * again. Page directories are only updated after updating page
374  * tables.
375  */
376 static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm)
377 {
378 	struct amdgpu_bo *pd = vm->root.bo;
379 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
380 	int ret;
381 
382 	ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate_vm_bo, NULL);
383 	if (ret) {
384 		pr_err("failed to validate PT BOs\n");
385 		return ret;
386 	}
387 
388 	ret = amdgpu_amdkfd_validate_vm_bo(NULL, pd);
389 	if (ret) {
390 		pr_err("failed to validate PD\n");
391 		return ret;
392 	}
393 
394 	vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo);
395 
396 	if (vm->use_cpu_for_update) {
397 		ret = amdgpu_bo_kmap(pd, NULL);
398 		if (ret) {
399 			pr_err("failed to kmap PD, ret=%d\n", ret);
400 			return ret;
401 		}
402 	}
403 
404 	return 0;
405 }
406 
407 static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
408 {
409 	struct amdgpu_bo *pd = vm->root.bo;
410 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
411 	int ret;
412 
413 	ret = amdgpu_vm_update_pdes(adev, vm, false);
414 	if (ret)
415 		return ret;
416 
417 	return amdgpu_sync_fence(sync, vm->last_update);
418 }
419 
420 static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem)
421 {
422 	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
423 	bool coherent = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT;
424 	bool uncached = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED;
425 	uint32_t mapping_flags;
426 	uint64_t pte_flags;
427 	bool snoop = false;
428 
429 	mapping_flags = AMDGPU_VM_PAGE_READABLE;
430 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE)
431 		mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
432 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE)
433 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
434 
435 	switch (adev->asic_type) {
436 	case CHIP_ARCTURUS:
437 		if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
438 			if (bo_adev == adev)
439 				mapping_flags |= coherent ?
440 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
441 			else
442 				mapping_flags |= coherent ?
443 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
444 		} else {
445 			mapping_flags |= coherent ?
446 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
447 		}
448 		break;
449 	case CHIP_ALDEBARAN:
450 		if (coherent && uncached) {
451 			if (adev->gmc.xgmi.connected_to_cpu ||
452 				!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM))
453 				snoop = true;
454 			mapping_flags |= AMDGPU_VM_MTYPE_UC;
455 		} else if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
456 			if (bo_adev == adev) {
457 				mapping_flags |= coherent ?
458 					AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
459 				if (adev->gmc.xgmi.connected_to_cpu)
460 					snoop = true;
461 			} else {
462 				mapping_flags |= coherent ?
463 					AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
464 				if (amdgpu_xgmi_same_hive(adev, bo_adev))
465 					snoop = true;
466 			}
467 		} else {
468 			snoop = true;
469 			mapping_flags |= coherent ?
470 				AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
471 		}
472 		break;
473 	default:
474 		mapping_flags |= coherent ?
475 			AMDGPU_VM_MTYPE_UC : AMDGPU_VM_MTYPE_NC;
476 	}
477 
478 	pte_flags = amdgpu_gem_va_map_flags(adev, mapping_flags);
479 	pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
480 
481 	return pte_flags;
482 }
483 
484 static int
485 kfd_mem_dmamap_userptr(struct kgd_mem *mem,
486 		       struct kfd_mem_attachment *attachment)
487 {
488 	enum dma_data_direction direction =
489 		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
490 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
491 	struct ttm_operation_ctx ctx = {.interruptible = true};
492 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
493 	struct amdgpu_device *adev = attachment->adev;
494 	struct ttm_tt *src_ttm = mem->bo->tbo.ttm;
495 	struct ttm_tt *ttm = bo->tbo.ttm;
496 	int ret;
497 
498 	ttm->sg = kmalloc(sizeof(*ttm->sg), GFP_KERNEL);
499 	if (unlikely(!ttm->sg))
500 		return -ENOMEM;
501 
502 	if (WARN_ON(ttm->num_pages != src_ttm->num_pages))
503 		return -EINVAL;
504 
505 	/* Same sequence as in amdgpu_ttm_tt_pin_userptr */
506 	ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages,
507 					ttm->num_pages, 0,
508 					(u64)ttm->num_pages << PAGE_SHIFT,
509 					GFP_KERNEL);
510 	if (unlikely(ret))
511 		goto free_sg;
512 
513 	ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
514 	if (unlikely(ret))
515 		goto release_sg;
516 
517 	drm_prime_sg_to_dma_addr_array(ttm->sg, ttm->dma_address,
518 				       ttm->num_pages);
519 
520 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
521 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
522 	if (ret)
523 		goto unmap_sg;
524 
525 	return 0;
526 
527 unmap_sg:
528 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
529 release_sg:
530 	pr_err("DMA map userptr failed: %d\n", ret);
531 	sg_free_table(ttm->sg);
532 free_sg:
533 	kfree(ttm->sg);
534 	ttm->sg = NULL;
535 	return ret;
536 }
537 
538 static int
539 kfd_mem_dmamap_dmabuf(struct kfd_mem_attachment *attachment)
540 {
541 	struct ttm_operation_ctx ctx = {.interruptible = true};
542 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
543 
544 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
545 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
546 }
547 
548 static int
549 kfd_mem_dmamap_attachment(struct kgd_mem *mem,
550 			  struct kfd_mem_attachment *attachment)
551 {
552 	switch (attachment->type) {
553 	case KFD_MEM_ATT_SHARED:
554 		return 0;
555 	case KFD_MEM_ATT_USERPTR:
556 		return kfd_mem_dmamap_userptr(mem, attachment);
557 	case KFD_MEM_ATT_DMABUF:
558 		return kfd_mem_dmamap_dmabuf(attachment);
559 	default:
560 		WARN_ON_ONCE(1);
561 	}
562 	return -EINVAL;
563 }
564 
565 static void
566 kfd_mem_dmaunmap_userptr(struct kgd_mem *mem,
567 			 struct kfd_mem_attachment *attachment)
568 {
569 	enum dma_data_direction direction =
570 		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
571 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
572 	struct ttm_operation_ctx ctx = {.interruptible = false};
573 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
574 	struct amdgpu_device *adev = attachment->adev;
575 	struct ttm_tt *ttm = bo->tbo.ttm;
576 
577 	if (unlikely(!ttm->sg))
578 		return;
579 
580 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
581 	ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
582 
583 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
584 	sg_free_table(ttm->sg);
585 	kfree(ttm->sg);
586 	ttm->sg = NULL;
587 }
588 
589 static void
590 kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment)
591 {
592 	struct ttm_operation_ctx ctx = {.interruptible = true};
593 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
594 
595 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
596 	ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
597 }
598 
599 static void
600 kfd_mem_dmaunmap_attachment(struct kgd_mem *mem,
601 			    struct kfd_mem_attachment *attachment)
602 {
603 	switch (attachment->type) {
604 	case KFD_MEM_ATT_SHARED:
605 		break;
606 	case KFD_MEM_ATT_USERPTR:
607 		kfd_mem_dmaunmap_userptr(mem, attachment);
608 		break;
609 	case KFD_MEM_ATT_DMABUF:
610 		kfd_mem_dmaunmap_dmabuf(attachment);
611 		break;
612 	default:
613 		WARN_ON_ONCE(1);
614 	}
615 }
616 
617 static int
618 kfd_mem_attach_userptr(struct amdgpu_device *adev, struct kgd_mem *mem,
619 		       struct amdgpu_bo **bo)
620 {
621 	unsigned long bo_size = mem->bo->tbo.base.size;
622 	struct drm_gem_object *gobj;
623 	int ret;
624 
625 	ret = amdgpu_bo_reserve(mem->bo, false);
626 	if (ret)
627 		return ret;
628 
629 	ret = amdgpu_gem_object_create(adev, bo_size, 1,
630 				       AMDGPU_GEM_DOMAIN_CPU,
631 				       AMDGPU_GEM_CREATE_PREEMPTIBLE,
632 				       ttm_bo_type_sg, mem->bo->tbo.base.resv,
633 				       &gobj);
634 	amdgpu_bo_unreserve(mem->bo);
635 	if (ret)
636 		return ret;
637 
638 	*bo = gem_to_amdgpu_bo(gobj);
639 	(*bo)->parent = amdgpu_bo_ref(mem->bo);
640 
641 	return 0;
642 }
643 
644 static int
645 kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem,
646 		      struct amdgpu_bo **bo)
647 {
648 	struct drm_gem_object *gobj;
649 	int ret;
650 
651 	if (!mem->dmabuf) {
652 		mem->dmabuf = amdgpu_gem_prime_export(&mem->bo->tbo.base,
653 			mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
654 				DRM_RDWR : 0);
655 		if (IS_ERR(mem->dmabuf)) {
656 			ret = PTR_ERR(mem->dmabuf);
657 			mem->dmabuf = NULL;
658 			return ret;
659 		}
660 	}
661 
662 	gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf);
663 	if (IS_ERR(gobj))
664 		return PTR_ERR(gobj);
665 
666 	*bo = gem_to_amdgpu_bo(gobj);
667 	(*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE;
668 	(*bo)->parent = amdgpu_bo_ref(mem->bo);
669 
670 	return 0;
671 }
672 
673 /* kfd_mem_attach - Add a BO to a VM
674  *
675  * Everything that needs to bo done only once when a BO is first added
676  * to a VM. It can later be mapped and unmapped many times without
677  * repeating these steps.
678  *
679  * 0. Create BO for DMA mapping, if needed
680  * 1. Allocate and initialize BO VA entry data structure
681  * 2. Add BO to the VM
682  * 3. Determine ASIC-specific PTE flags
683  * 4. Alloc page tables and directories if needed
684  * 4a.  Validate new page tables and directories
685  */
686 static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
687 		struct amdgpu_vm *vm, bool is_aql)
688 {
689 	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
690 	unsigned long bo_size = mem->bo->tbo.base.size;
691 	uint64_t va = mem->va;
692 	struct kfd_mem_attachment *attachment[2] = {NULL, NULL};
693 	struct amdgpu_bo *bo[2] = {NULL, NULL};
694 	int i, ret;
695 
696 	if (!va) {
697 		pr_err("Invalid VA when adding BO to VM\n");
698 		return -EINVAL;
699 	}
700 
701 	for (i = 0; i <= is_aql; i++) {
702 		attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL);
703 		if (unlikely(!attachment[i])) {
704 			ret = -ENOMEM;
705 			goto unwind;
706 		}
707 
708 		pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
709 			 va + bo_size, vm);
710 
711 		if (adev == bo_adev ||
712 		   (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && adev->ram_is_direct_mapped) ||
713 		   (mem->domain == AMDGPU_GEM_DOMAIN_VRAM && amdgpu_xgmi_same_hive(adev, bo_adev))) {
714 			/* Mappings on the local GPU, or VRAM mappings in the
715 			 * local hive, or userptr mapping IOMMU direct map mode
716 			 * share the original BO
717 			 */
718 			attachment[i]->type = KFD_MEM_ATT_SHARED;
719 			bo[i] = mem->bo;
720 			drm_gem_object_get(&bo[i]->tbo.base);
721 		} else if (i > 0) {
722 			/* Multiple mappings on the same GPU share the BO */
723 			attachment[i]->type = KFD_MEM_ATT_SHARED;
724 			bo[i] = bo[0];
725 			drm_gem_object_get(&bo[i]->tbo.base);
726 		} else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
727 			/* Create an SG BO to DMA-map userptrs on other GPUs */
728 			attachment[i]->type = KFD_MEM_ATT_USERPTR;
729 			ret = kfd_mem_attach_userptr(adev, mem, &bo[i]);
730 			if (ret)
731 				goto unwind;
732 		} else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT &&
733 			   mem->bo->tbo.type != ttm_bo_type_sg) {
734 			/* GTT BOs use DMA-mapping ability of dynamic-attach
735 			 * DMA bufs. TODO: The same should work for VRAM on
736 			 * large-BAR GPUs.
737 			 */
738 			attachment[i]->type = KFD_MEM_ATT_DMABUF;
739 			ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]);
740 			if (ret)
741 				goto unwind;
742 		} else {
743 			/* FIXME: Need to DMA-map other BO types:
744 			 * large-BAR VRAM, doorbells, MMIO remap
745 			 */
746 			attachment[i]->type = KFD_MEM_ATT_SHARED;
747 			bo[i] = mem->bo;
748 			drm_gem_object_get(&bo[i]->tbo.base);
749 		}
750 
751 		/* Add BO to VM internal data structures */
752 		ret = amdgpu_bo_reserve(bo[i], false);
753 		if (ret) {
754 			pr_debug("Unable to reserve BO during memory attach");
755 			goto unwind;
756 		}
757 		attachment[i]->bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]);
758 		amdgpu_bo_unreserve(bo[i]);
759 		if (unlikely(!attachment[i]->bo_va)) {
760 			ret = -ENOMEM;
761 			pr_err("Failed to add BO object to VM. ret == %d\n",
762 			       ret);
763 			goto unwind;
764 		}
765 		attachment[i]->va = va;
766 		attachment[i]->pte_flags = get_pte_flags(adev, mem);
767 		attachment[i]->adev = adev;
768 		list_add(&attachment[i]->list, &mem->attachments);
769 
770 		va += bo_size;
771 	}
772 
773 	return 0;
774 
775 unwind:
776 	for (; i >= 0; i--) {
777 		if (!attachment[i])
778 			continue;
779 		if (attachment[i]->bo_va) {
780 			amdgpu_bo_reserve(bo[i], true);
781 			amdgpu_vm_bo_del(adev, attachment[i]->bo_va);
782 			amdgpu_bo_unreserve(bo[i]);
783 			list_del(&attachment[i]->list);
784 		}
785 		if (bo[i])
786 			drm_gem_object_put(&bo[i]->tbo.base);
787 		kfree(attachment[i]);
788 	}
789 	return ret;
790 }
791 
792 static void kfd_mem_detach(struct kfd_mem_attachment *attachment)
793 {
794 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
795 
796 	pr_debug("\t remove VA 0x%llx in entry %p\n",
797 			attachment->va, attachment);
798 	amdgpu_vm_bo_del(attachment->adev, attachment->bo_va);
799 	drm_gem_object_put(&bo->tbo.base);
800 	list_del(&attachment->list);
801 	kfree(attachment);
802 }
803 
804 static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
805 				struct amdkfd_process_info *process_info,
806 				bool userptr)
807 {
808 	struct ttm_validate_buffer *entry = &mem->validate_list;
809 	struct amdgpu_bo *bo = mem->bo;
810 
811 	INIT_LIST_HEAD(&entry->head);
812 	entry->num_shared = 1;
813 	entry->bo = &bo->tbo;
814 	mutex_lock(&process_info->lock);
815 	if (userptr)
816 		list_add_tail(&entry->head, &process_info->userptr_valid_list);
817 	else
818 		list_add_tail(&entry->head, &process_info->kfd_bo_list);
819 	mutex_unlock(&process_info->lock);
820 }
821 
822 static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
823 		struct amdkfd_process_info *process_info)
824 {
825 	struct ttm_validate_buffer *bo_list_entry;
826 
827 	bo_list_entry = &mem->validate_list;
828 	mutex_lock(&process_info->lock);
829 	list_del(&bo_list_entry->head);
830 	mutex_unlock(&process_info->lock);
831 }
832 
833 /* Initializes user pages. It registers the MMU notifier and validates
834  * the userptr BO in the GTT domain.
835  *
836  * The BO must already be on the userptr_valid_list. Otherwise an
837  * eviction and restore may happen that leaves the new BO unmapped
838  * with the user mode queues running.
839  *
840  * Takes the process_info->lock to protect against concurrent restore
841  * workers.
842  *
843  * Returns 0 for success, negative errno for errors.
844  */
845 static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr,
846 			   bool criu_resume)
847 {
848 	struct amdkfd_process_info *process_info = mem->process_info;
849 	struct amdgpu_bo *bo = mem->bo;
850 	struct ttm_operation_ctx ctx = { true, false };
851 	int ret = 0;
852 
853 	mutex_lock(&process_info->lock);
854 
855 	ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
856 	if (ret) {
857 		pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
858 		goto out;
859 	}
860 
861 	ret = amdgpu_mn_register(bo, user_addr);
862 	if (ret) {
863 		pr_err("%s: Failed to register MMU notifier: %d\n",
864 		       __func__, ret);
865 		goto out;
866 	}
867 
868 	if (criu_resume) {
869 		/*
870 		 * During a CRIU restore operation, the userptr buffer objects
871 		 * will be validated in the restore_userptr_work worker at a
872 		 * later stage when it is scheduled by another ioctl called by
873 		 * CRIU master process for the target pid for restore.
874 		 */
875 		atomic_inc(&mem->invalid);
876 		mutex_unlock(&process_info->lock);
877 		return 0;
878 	}
879 
880 	ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
881 	if (ret) {
882 		pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
883 		goto unregister_out;
884 	}
885 
886 	ret = amdgpu_bo_reserve(bo, true);
887 	if (ret) {
888 		pr_err("%s: Failed to reserve BO\n", __func__);
889 		goto release_out;
890 	}
891 	amdgpu_bo_placement_from_domain(bo, mem->domain);
892 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
893 	if (ret)
894 		pr_err("%s: failed to validate BO\n", __func__);
895 	amdgpu_bo_unreserve(bo);
896 
897 release_out:
898 	amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
899 unregister_out:
900 	if (ret)
901 		amdgpu_mn_unregister(bo);
902 out:
903 	mutex_unlock(&process_info->lock);
904 	return ret;
905 }
906 
907 /* Reserving a BO and its page table BOs must happen atomically to
908  * avoid deadlocks. Some operations update multiple VMs at once. Track
909  * all the reservation info in a context structure. Optionally a sync
910  * object can track VM updates.
911  */
912 struct bo_vm_reservation_context {
913 	struct amdgpu_bo_list_entry kfd_bo; /* BO list entry for the KFD BO */
914 	unsigned int n_vms;		    /* Number of VMs reserved	    */
915 	struct amdgpu_bo_list_entry *vm_pd; /* Array of VM BO list entries  */
916 	struct ww_acquire_ctx ticket;	    /* Reservation ticket	    */
917 	struct list_head list, duplicates;  /* BO lists			    */
918 	struct amdgpu_sync *sync;	    /* Pointer to sync object	    */
919 	bool reserved;			    /* Whether BOs are reserved	    */
920 };
921 
922 enum bo_vm_match {
923 	BO_VM_NOT_MAPPED = 0,	/* Match VMs where a BO is not mapped */
924 	BO_VM_MAPPED,		/* Match VMs where a BO is mapped     */
925 	BO_VM_ALL,		/* Match all VMs a BO was added to    */
926 };
927 
928 /**
929  * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
930  * @mem: KFD BO structure.
931  * @vm: the VM to reserve.
932  * @ctx: the struct that will be used in unreserve_bo_and_vms().
933  */
934 static int reserve_bo_and_vm(struct kgd_mem *mem,
935 			      struct amdgpu_vm *vm,
936 			      struct bo_vm_reservation_context *ctx)
937 {
938 	struct amdgpu_bo *bo = mem->bo;
939 	int ret;
940 
941 	WARN_ON(!vm);
942 
943 	ctx->reserved = false;
944 	ctx->n_vms = 1;
945 	ctx->sync = &mem->sync;
946 
947 	INIT_LIST_HEAD(&ctx->list);
948 	INIT_LIST_HEAD(&ctx->duplicates);
949 
950 	ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), GFP_KERNEL);
951 	if (!ctx->vm_pd)
952 		return -ENOMEM;
953 
954 	ctx->kfd_bo.priority = 0;
955 	ctx->kfd_bo.tv.bo = &bo->tbo;
956 	ctx->kfd_bo.tv.num_shared = 1;
957 	list_add(&ctx->kfd_bo.tv.head, &ctx->list);
958 
959 	amdgpu_vm_get_pd_bo(vm, &ctx->list, &ctx->vm_pd[0]);
960 
961 	ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
962 				     false, &ctx->duplicates);
963 	if (ret) {
964 		pr_err("Failed to reserve buffers in ttm.\n");
965 		kfree(ctx->vm_pd);
966 		ctx->vm_pd = NULL;
967 		return ret;
968 	}
969 
970 	ctx->reserved = true;
971 	return 0;
972 }
973 
974 /**
975  * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
976  * @mem: KFD BO structure.
977  * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
978  * is used. Otherwise, a single VM associated with the BO.
979  * @map_type: the mapping status that will be used to filter the VMs.
980  * @ctx: the struct that will be used in unreserve_bo_and_vms().
981  *
982  * Returns 0 for success, negative for failure.
983  */
984 static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
985 				struct amdgpu_vm *vm, enum bo_vm_match map_type,
986 				struct bo_vm_reservation_context *ctx)
987 {
988 	struct amdgpu_bo *bo = mem->bo;
989 	struct kfd_mem_attachment *entry;
990 	unsigned int i;
991 	int ret;
992 
993 	ctx->reserved = false;
994 	ctx->n_vms = 0;
995 	ctx->vm_pd = NULL;
996 	ctx->sync = &mem->sync;
997 
998 	INIT_LIST_HEAD(&ctx->list);
999 	INIT_LIST_HEAD(&ctx->duplicates);
1000 
1001 	list_for_each_entry(entry, &mem->attachments, list) {
1002 		if ((vm && vm != entry->bo_va->base.vm) ||
1003 			(entry->is_mapped != map_type
1004 			&& map_type != BO_VM_ALL))
1005 			continue;
1006 
1007 		ctx->n_vms++;
1008 	}
1009 
1010 	if (ctx->n_vms != 0) {
1011 		ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd),
1012 				     GFP_KERNEL);
1013 		if (!ctx->vm_pd)
1014 			return -ENOMEM;
1015 	}
1016 
1017 	ctx->kfd_bo.priority = 0;
1018 	ctx->kfd_bo.tv.bo = &bo->tbo;
1019 	ctx->kfd_bo.tv.num_shared = 1;
1020 	list_add(&ctx->kfd_bo.tv.head, &ctx->list);
1021 
1022 	i = 0;
1023 	list_for_each_entry(entry, &mem->attachments, list) {
1024 		if ((vm && vm != entry->bo_va->base.vm) ||
1025 			(entry->is_mapped != map_type
1026 			&& map_type != BO_VM_ALL))
1027 			continue;
1028 
1029 		amdgpu_vm_get_pd_bo(entry->bo_va->base.vm, &ctx->list,
1030 				&ctx->vm_pd[i]);
1031 		i++;
1032 	}
1033 
1034 	ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
1035 				     false, &ctx->duplicates);
1036 	if (ret) {
1037 		pr_err("Failed to reserve buffers in ttm.\n");
1038 		kfree(ctx->vm_pd);
1039 		ctx->vm_pd = NULL;
1040 		return ret;
1041 	}
1042 
1043 	ctx->reserved = true;
1044 	return 0;
1045 }
1046 
1047 /**
1048  * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1049  * @ctx: Reservation context to unreserve
1050  * @wait: Optionally wait for a sync object representing pending VM updates
1051  * @intr: Whether the wait is interruptible
1052  *
1053  * Also frees any resources allocated in
1054  * reserve_bo_and_(cond_)vm(s). Returns the status from
1055  * amdgpu_sync_wait.
1056  */
1057 static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
1058 				 bool wait, bool intr)
1059 {
1060 	int ret = 0;
1061 
1062 	if (wait)
1063 		ret = amdgpu_sync_wait(ctx->sync, intr);
1064 
1065 	if (ctx->reserved)
1066 		ttm_eu_backoff_reservation(&ctx->ticket, &ctx->list);
1067 	kfree(ctx->vm_pd);
1068 
1069 	ctx->sync = NULL;
1070 
1071 	ctx->reserved = false;
1072 	ctx->vm_pd = NULL;
1073 
1074 	return ret;
1075 }
1076 
1077 static void unmap_bo_from_gpuvm(struct kgd_mem *mem,
1078 				struct kfd_mem_attachment *entry,
1079 				struct amdgpu_sync *sync)
1080 {
1081 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1082 	struct amdgpu_device *adev = entry->adev;
1083 	struct amdgpu_vm *vm = bo_va->base.vm;
1084 
1085 	amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
1086 
1087 	amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
1088 
1089 	amdgpu_sync_fence(sync, bo_va->last_pt_update);
1090 
1091 	kfd_mem_dmaunmap_attachment(mem, entry);
1092 }
1093 
1094 static int update_gpuvm_pte(struct kgd_mem *mem,
1095 			    struct kfd_mem_attachment *entry,
1096 			    struct amdgpu_sync *sync,
1097 			    bool *table_freed)
1098 {
1099 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1100 	struct amdgpu_device *adev = entry->adev;
1101 	int ret;
1102 
1103 	ret = kfd_mem_dmamap_attachment(mem, entry);
1104 	if (ret)
1105 		return ret;
1106 
1107 	/* Update the page tables  */
1108 	ret = amdgpu_vm_bo_update(adev, bo_va, false, table_freed);
1109 	if (ret) {
1110 		pr_err("amdgpu_vm_bo_update failed\n");
1111 		return ret;
1112 	}
1113 
1114 	return amdgpu_sync_fence(sync, bo_va->last_pt_update);
1115 }
1116 
1117 static int map_bo_to_gpuvm(struct kgd_mem *mem,
1118 			   struct kfd_mem_attachment *entry,
1119 			   struct amdgpu_sync *sync,
1120 			   bool no_update_pte,
1121 			   bool *table_freed)
1122 {
1123 	int ret;
1124 
1125 	/* Set virtual address for the allocation */
1126 	ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0,
1127 			       amdgpu_bo_size(entry->bo_va->base.bo),
1128 			       entry->pte_flags);
1129 	if (ret) {
1130 		pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
1131 				entry->va, ret);
1132 		return ret;
1133 	}
1134 
1135 	if (no_update_pte)
1136 		return 0;
1137 
1138 	ret = update_gpuvm_pte(mem, entry, sync, table_freed);
1139 	if (ret) {
1140 		pr_err("update_gpuvm_pte() failed\n");
1141 		goto update_gpuvm_pte_failed;
1142 	}
1143 
1144 	return 0;
1145 
1146 update_gpuvm_pte_failed:
1147 	unmap_bo_from_gpuvm(mem, entry, sync);
1148 	return ret;
1149 }
1150 
1151 static struct sg_table *create_doorbell_sg(uint64_t addr, uint32_t size)
1152 {
1153 	struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
1154 
1155 	if (!sg)
1156 		return NULL;
1157 	if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
1158 		kfree(sg);
1159 		return NULL;
1160 	}
1161 	sg->sgl->dma_address = addr;
1162 	sg->sgl->length = size;
1163 #ifdef CONFIG_NEED_SG_DMA_LENGTH
1164 	sg->sgl->dma_length = size;
1165 #endif
1166 	return sg;
1167 }
1168 
1169 static int process_validate_vms(struct amdkfd_process_info *process_info)
1170 {
1171 	struct amdgpu_vm *peer_vm;
1172 	int ret;
1173 
1174 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1175 			    vm_list_node) {
1176 		ret = vm_validate_pt_pd_bos(peer_vm);
1177 		if (ret)
1178 			return ret;
1179 	}
1180 
1181 	return 0;
1182 }
1183 
1184 static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
1185 				 struct amdgpu_sync *sync)
1186 {
1187 	struct amdgpu_vm *peer_vm;
1188 	int ret;
1189 
1190 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1191 			    vm_list_node) {
1192 		struct amdgpu_bo *pd = peer_vm->root.bo;
1193 
1194 		ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
1195 				       AMDGPU_SYNC_NE_OWNER,
1196 				       AMDGPU_FENCE_OWNER_KFD);
1197 		if (ret)
1198 			return ret;
1199 	}
1200 
1201 	return 0;
1202 }
1203 
1204 static int process_update_pds(struct amdkfd_process_info *process_info,
1205 			      struct amdgpu_sync *sync)
1206 {
1207 	struct amdgpu_vm *peer_vm;
1208 	int ret;
1209 
1210 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1211 			    vm_list_node) {
1212 		ret = vm_update_pds(peer_vm, sync);
1213 		if (ret)
1214 			return ret;
1215 	}
1216 
1217 	return 0;
1218 }
1219 
1220 static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
1221 		       struct dma_fence **ef)
1222 {
1223 	struct amdkfd_process_info *info = NULL;
1224 	int ret;
1225 
1226 	if (!*process_info) {
1227 		info = kzalloc(sizeof(*info), GFP_KERNEL);
1228 		if (!info)
1229 			return -ENOMEM;
1230 
1231 		mutex_init(&info->lock);
1232 		INIT_LIST_HEAD(&info->vm_list_head);
1233 		INIT_LIST_HEAD(&info->kfd_bo_list);
1234 		INIT_LIST_HEAD(&info->userptr_valid_list);
1235 		INIT_LIST_HEAD(&info->userptr_inval_list);
1236 
1237 		info->eviction_fence =
1238 			amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
1239 						   current->mm,
1240 						   NULL);
1241 		if (!info->eviction_fence) {
1242 			pr_err("Failed to create eviction fence\n");
1243 			ret = -ENOMEM;
1244 			goto create_evict_fence_fail;
1245 		}
1246 
1247 		info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
1248 		atomic_set(&info->evicted_bos, 0);
1249 		INIT_DELAYED_WORK(&info->restore_userptr_work,
1250 				  amdgpu_amdkfd_restore_userptr_worker);
1251 
1252 		*process_info = info;
1253 		*ef = dma_fence_get(&info->eviction_fence->base);
1254 	}
1255 
1256 	vm->process_info = *process_info;
1257 
1258 	/* Validate page directory and attach eviction fence */
1259 	ret = amdgpu_bo_reserve(vm->root.bo, true);
1260 	if (ret)
1261 		goto reserve_pd_fail;
1262 	ret = vm_validate_pt_pd_bos(vm);
1263 	if (ret) {
1264 		pr_err("validate_pt_pd_bos() failed\n");
1265 		goto validate_pd_fail;
1266 	}
1267 	ret = amdgpu_bo_sync_wait(vm->root.bo,
1268 				  AMDGPU_FENCE_OWNER_KFD, false);
1269 	if (ret)
1270 		goto wait_pd_fail;
1271 	ret = dma_resv_reserve_shared(vm->root.bo->tbo.base.resv, 1);
1272 	if (ret)
1273 		goto reserve_shared_fail;
1274 	amdgpu_bo_fence(vm->root.bo,
1275 			&vm->process_info->eviction_fence->base, true);
1276 	amdgpu_bo_unreserve(vm->root.bo);
1277 
1278 	/* Update process info */
1279 	mutex_lock(&vm->process_info->lock);
1280 	list_add_tail(&vm->vm_list_node,
1281 			&(vm->process_info->vm_list_head));
1282 	vm->process_info->n_vms++;
1283 	mutex_unlock(&vm->process_info->lock);
1284 
1285 	return 0;
1286 
1287 reserve_shared_fail:
1288 wait_pd_fail:
1289 validate_pd_fail:
1290 	amdgpu_bo_unreserve(vm->root.bo);
1291 reserve_pd_fail:
1292 	vm->process_info = NULL;
1293 	if (info) {
1294 		/* Two fence references: one in info and one in *ef */
1295 		dma_fence_put(&info->eviction_fence->base);
1296 		dma_fence_put(*ef);
1297 		*ef = NULL;
1298 		*process_info = NULL;
1299 		put_pid(info->pid);
1300 create_evict_fence_fail:
1301 		mutex_destroy(&info->lock);
1302 		kfree(info);
1303 	}
1304 	return ret;
1305 }
1306 
1307 /**
1308  * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
1309  * @bo: Handle of buffer object being pinned
1310  * @domain: Domain into which BO should be pinned
1311  *
1312  *   - USERPTR BOs are UNPINNABLE and will return error
1313  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1314  *     PIN count incremented. It is valid to PIN a BO multiple times
1315  *
1316  * Return: ZERO if successful in pinning, Non-Zero in case of error.
1317  */
1318 static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain)
1319 {
1320 	int ret = 0;
1321 
1322 	ret = amdgpu_bo_reserve(bo, false);
1323 	if (unlikely(ret))
1324 		return ret;
1325 
1326 	ret = amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1327 	if (ret)
1328 		pr_err("Error in Pinning BO to domain: %d\n", domain);
1329 
1330 	amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
1331 	amdgpu_bo_unreserve(bo);
1332 
1333 	return ret;
1334 }
1335 
1336 /**
1337  * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
1338  * @bo: Handle of buffer object being unpinned
1339  *
1340  *   - Is a illegal request for USERPTR BOs and is ignored
1341  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1342  *     PIN count decremented. Calls to UNPIN must balance calls to PIN
1343  */
1344 static void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo)
1345 {
1346 	int ret = 0;
1347 
1348 	ret = amdgpu_bo_reserve(bo, false);
1349 	if (unlikely(ret))
1350 		return;
1351 
1352 	amdgpu_bo_unpin(bo);
1353 	amdgpu_bo_unreserve(bo);
1354 }
1355 
1356 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
1357 					   struct file *filp, u32 pasid,
1358 					   void **process_info,
1359 					   struct dma_fence **ef)
1360 {
1361 	struct amdgpu_fpriv *drv_priv;
1362 	struct amdgpu_vm *avm;
1363 	int ret;
1364 
1365 	ret = amdgpu_file_to_fpriv(filp, &drv_priv);
1366 	if (ret)
1367 		return ret;
1368 	avm = &drv_priv->vm;
1369 
1370 	/* Already a compute VM? */
1371 	if (avm->process_info)
1372 		return -EINVAL;
1373 
1374 	/* Free the original amdgpu allocated pasid,
1375 	 * will be replaced with kfd allocated pasid.
1376 	 */
1377 	if (avm->pasid) {
1378 		amdgpu_pasid_free(avm->pasid);
1379 		amdgpu_vm_set_pasid(adev, avm, 0);
1380 	}
1381 
1382 	/* Convert VM into a compute VM */
1383 	ret = amdgpu_vm_make_compute(adev, avm);
1384 	if (ret)
1385 		return ret;
1386 
1387 	ret = amdgpu_vm_set_pasid(adev, avm, pasid);
1388 	if (ret)
1389 		return ret;
1390 	/* Initialize KFD part of the VM and process info */
1391 	ret = init_kfd_vm(avm, process_info, ef);
1392 	if (ret)
1393 		return ret;
1394 
1395 	amdgpu_vm_set_task_info(avm);
1396 
1397 	return 0;
1398 }
1399 
1400 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1401 				    struct amdgpu_vm *vm)
1402 {
1403 	struct amdkfd_process_info *process_info = vm->process_info;
1404 	struct amdgpu_bo *pd = vm->root.bo;
1405 
1406 	if (!process_info)
1407 		return;
1408 
1409 	/* Release eviction fence from PD */
1410 	amdgpu_bo_reserve(pd, false);
1411 	amdgpu_bo_fence(pd, NULL, false);
1412 	amdgpu_bo_unreserve(pd);
1413 
1414 	/* Update process info */
1415 	mutex_lock(&process_info->lock);
1416 	process_info->n_vms--;
1417 	list_del(&vm->vm_list_node);
1418 	mutex_unlock(&process_info->lock);
1419 
1420 	vm->process_info = NULL;
1421 
1422 	/* Release per-process resources when last compute VM is destroyed */
1423 	if (!process_info->n_vms) {
1424 		WARN_ON(!list_empty(&process_info->kfd_bo_list));
1425 		WARN_ON(!list_empty(&process_info->userptr_valid_list));
1426 		WARN_ON(!list_empty(&process_info->userptr_inval_list));
1427 
1428 		dma_fence_put(&process_info->eviction_fence->base);
1429 		cancel_delayed_work_sync(&process_info->restore_userptr_work);
1430 		put_pid(process_info->pid);
1431 		mutex_destroy(&process_info->lock);
1432 		kfree(process_info);
1433 	}
1434 }
1435 
1436 void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev,
1437 					    void *drm_priv)
1438 {
1439 	struct amdgpu_vm *avm;
1440 
1441 	if (WARN_ON(!adev || !drm_priv))
1442 		return;
1443 
1444 	avm = drm_priv_to_vm(drm_priv);
1445 
1446 	pr_debug("Releasing process vm %p\n", avm);
1447 
1448 	/* The original pasid of amdgpu vm has already been
1449 	 * released during making a amdgpu vm to a compute vm
1450 	 * The current pasid is managed by kfd and will be
1451 	 * released on kfd process destroy. Set amdgpu pasid
1452 	 * to 0 to avoid duplicate release.
1453 	 */
1454 	amdgpu_vm_release_compute(adev, avm);
1455 }
1456 
1457 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv)
1458 {
1459 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1460 	struct amdgpu_bo *pd = avm->root.bo;
1461 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1462 
1463 	if (adev->asic_type < CHIP_VEGA10)
1464 		return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1465 	return avm->pd_phys_addr;
1466 }
1467 
1468 void amdgpu_amdkfd_block_mmu_notifications(void *p)
1469 {
1470 	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1471 
1472 	mutex_lock(&pinfo->lock);
1473 	WRITE_ONCE(pinfo->block_mmu_notifications, true);
1474 	mutex_unlock(&pinfo->lock);
1475 }
1476 
1477 int amdgpu_amdkfd_criu_resume(void *p)
1478 {
1479 	int ret = 0;
1480 	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1481 
1482 	mutex_lock(&pinfo->lock);
1483 	pr_debug("scheduling work\n");
1484 	atomic_inc(&pinfo->evicted_bos);
1485 	if (!READ_ONCE(pinfo->block_mmu_notifications)) {
1486 		ret = -EINVAL;
1487 		goto out_unlock;
1488 	}
1489 	WRITE_ONCE(pinfo->block_mmu_notifications, false);
1490 	schedule_delayed_work(&pinfo->restore_userptr_work, 0);
1491 
1492 out_unlock:
1493 	mutex_unlock(&pinfo->lock);
1494 	return ret;
1495 }
1496 
1497 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1498 		struct amdgpu_device *adev, uint64_t va, uint64_t size,
1499 		void *drm_priv, struct kgd_mem **mem,
1500 		uint64_t *offset, uint32_t flags, bool criu_resume)
1501 {
1502 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1503 	enum ttm_bo_type bo_type = ttm_bo_type_device;
1504 	struct sg_table *sg = NULL;
1505 	uint64_t user_addr = 0;
1506 	struct amdgpu_bo *bo;
1507 	struct drm_gem_object *gobj = NULL;
1508 	u32 domain, alloc_domain;
1509 	u64 alloc_flags;
1510 	int ret;
1511 
1512 	/*
1513 	 * Check on which domain to allocate BO
1514 	 */
1515 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1516 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1517 		alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1518 		alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1519 			AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0;
1520 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1521 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1522 		alloc_flags = 0;
1523 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1524 		domain = AMDGPU_GEM_DOMAIN_GTT;
1525 		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1526 		alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE;
1527 		if (!offset || !*offset)
1528 			return -EINVAL;
1529 		user_addr = untagged_addr(*offset);
1530 	} else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1531 			KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1532 		domain = AMDGPU_GEM_DOMAIN_GTT;
1533 		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1534 		bo_type = ttm_bo_type_sg;
1535 		alloc_flags = 0;
1536 		if (size > UINT_MAX)
1537 			return -EINVAL;
1538 		sg = create_doorbell_sg(*offset, size);
1539 		if (!sg)
1540 			return -ENOMEM;
1541 	} else {
1542 		return -EINVAL;
1543 	}
1544 
1545 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1546 	if (!*mem) {
1547 		ret = -ENOMEM;
1548 		goto err;
1549 	}
1550 	INIT_LIST_HEAD(&(*mem)->attachments);
1551 	mutex_init(&(*mem)->lock);
1552 	(*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1553 
1554 	/* Workaround for AQL queue wraparound bug. Map the same
1555 	 * memory twice. That means we only actually allocate half
1556 	 * the memory.
1557 	 */
1558 	if ((*mem)->aql_queue)
1559 		size = size >> 1;
1560 
1561 	(*mem)->alloc_flags = flags;
1562 
1563 	amdgpu_sync_create(&(*mem)->sync);
1564 
1565 	ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, flags);
1566 	if (ret) {
1567 		pr_debug("Insufficient memory\n");
1568 		goto err_reserve_limit;
1569 	}
1570 
1571 	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n",
1572 			va, size, domain_string(alloc_domain));
1573 
1574 	ret = amdgpu_gem_object_create(adev, size, 1, alloc_domain, alloc_flags,
1575 				       bo_type, NULL, &gobj);
1576 	if (ret) {
1577 		pr_debug("Failed to create BO on domain %s. ret %d\n",
1578 			 domain_string(alloc_domain), ret);
1579 		goto err_bo_create;
1580 	}
1581 	ret = drm_vma_node_allow(&gobj->vma_node, drm_priv);
1582 	if (ret) {
1583 		pr_debug("Failed to allow vma node access. ret %d\n", ret);
1584 		goto err_node_allow;
1585 	}
1586 	bo = gem_to_amdgpu_bo(gobj);
1587 	if (bo_type == ttm_bo_type_sg) {
1588 		bo->tbo.sg = sg;
1589 		bo->tbo.ttm->sg = sg;
1590 	}
1591 	bo->kfd_bo = *mem;
1592 	(*mem)->bo = bo;
1593 	if (user_addr)
1594 		bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO;
1595 
1596 	(*mem)->va = va;
1597 	(*mem)->domain = domain;
1598 	(*mem)->mapped_to_gpu_memory = 0;
1599 	(*mem)->process_info = avm->process_info;
1600 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1601 
1602 	if (user_addr) {
1603 		pr_debug("creating userptr BO for user_addr = %llu\n", user_addr);
1604 		ret = init_user_pages(*mem, user_addr, criu_resume);
1605 		if (ret)
1606 			goto allocate_init_user_pages_failed;
1607 	} else  if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1608 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1609 		ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT);
1610 		if (ret) {
1611 			pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n");
1612 			goto err_pin_bo;
1613 		}
1614 		bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
1615 		bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
1616 	}
1617 
1618 	if (offset)
1619 		*offset = amdgpu_bo_mmap_offset(bo);
1620 
1621 	return 0;
1622 
1623 allocate_init_user_pages_failed:
1624 err_pin_bo:
1625 	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1626 	drm_vma_node_revoke(&gobj->vma_node, drm_priv);
1627 err_node_allow:
1628 	/* Don't unreserve system mem limit twice */
1629 	goto err_reserve_limit;
1630 err_bo_create:
1631 	unreserve_mem_limit(adev, size, flags);
1632 err_reserve_limit:
1633 	mutex_destroy(&(*mem)->lock);
1634 	if (gobj)
1635 		drm_gem_object_put(gobj);
1636 	else
1637 		kfree(*mem);
1638 err:
1639 	if (sg) {
1640 		sg_free_table(sg);
1641 		kfree(sg);
1642 	}
1643 	return ret;
1644 }
1645 
1646 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1647 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
1648 		uint64_t *size)
1649 {
1650 	struct amdkfd_process_info *process_info = mem->process_info;
1651 	unsigned long bo_size = mem->bo->tbo.base.size;
1652 	struct kfd_mem_attachment *entry, *tmp;
1653 	struct bo_vm_reservation_context ctx;
1654 	struct ttm_validate_buffer *bo_list_entry;
1655 	unsigned int mapped_to_gpu_memory;
1656 	int ret;
1657 	bool is_imported = false;
1658 
1659 	mutex_lock(&mem->lock);
1660 
1661 	/* Unpin MMIO/DOORBELL BO's that were pinnned during allocation */
1662 	if (mem->alloc_flags &
1663 	    (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1664 	     KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1665 		amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo);
1666 	}
1667 
1668 	mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1669 	is_imported = mem->is_imported;
1670 	mutex_unlock(&mem->lock);
1671 	/* lock is not needed after this, since mem is unused and will
1672 	 * be freed anyway
1673 	 */
1674 
1675 	if (mapped_to_gpu_memory > 0) {
1676 		pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1677 				mem->va, bo_size);
1678 		return -EBUSY;
1679 	}
1680 
1681 	/* Make sure restore workers don't access the BO any more */
1682 	bo_list_entry = &mem->validate_list;
1683 	mutex_lock(&process_info->lock);
1684 	list_del(&bo_list_entry->head);
1685 	mutex_unlock(&process_info->lock);
1686 
1687 	/* No more MMU notifiers */
1688 	amdgpu_mn_unregister(mem->bo);
1689 
1690 	ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1691 	if (unlikely(ret))
1692 		return ret;
1693 
1694 	/* The eviction fence should be removed by the last unmap.
1695 	 * TODO: Log an error condition if the bo still has the eviction fence
1696 	 * attached
1697 	 */
1698 	amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1699 					process_info->eviction_fence);
1700 	pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1701 		mem->va + bo_size * (1 + mem->aql_queue));
1702 
1703 	/* Remove from VM internal data structures */
1704 	list_for_each_entry_safe(entry, tmp, &mem->attachments, list)
1705 		kfd_mem_detach(entry);
1706 
1707 	ret = unreserve_bo_and_vms(&ctx, false, false);
1708 
1709 	/* Free the sync object */
1710 	amdgpu_sync_free(&mem->sync);
1711 
1712 	/* If the SG is not NULL, it's one we created for a doorbell or mmio
1713 	 * remap BO. We need to free it.
1714 	 */
1715 	if (mem->bo->tbo.sg) {
1716 		sg_free_table(mem->bo->tbo.sg);
1717 		kfree(mem->bo->tbo.sg);
1718 	}
1719 
1720 	/* Update the size of the BO being freed if it was allocated from
1721 	 * VRAM and is not imported.
1722 	 */
1723 	if (size) {
1724 		if ((mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM) &&
1725 		    (!is_imported))
1726 			*size = bo_size;
1727 		else
1728 			*size = 0;
1729 	}
1730 
1731 	/* Free the BO*/
1732 	drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
1733 	if (mem->dmabuf)
1734 		dma_buf_put(mem->dmabuf);
1735 	mutex_destroy(&mem->lock);
1736 
1737 	/* If this releases the last reference, it will end up calling
1738 	 * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why
1739 	 * this needs to be the last call here.
1740 	 */
1741 	drm_gem_object_put(&mem->bo->tbo.base);
1742 
1743 	return ret;
1744 }
1745 
1746 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1747 		struct amdgpu_device *adev, struct kgd_mem *mem,
1748 		void *drm_priv, bool *table_freed)
1749 {
1750 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1751 	int ret;
1752 	struct amdgpu_bo *bo;
1753 	uint32_t domain;
1754 	struct kfd_mem_attachment *entry;
1755 	struct bo_vm_reservation_context ctx;
1756 	unsigned long bo_size;
1757 	bool is_invalid_userptr = false;
1758 
1759 	bo = mem->bo;
1760 	if (!bo) {
1761 		pr_err("Invalid BO when mapping memory to GPU\n");
1762 		return -EINVAL;
1763 	}
1764 
1765 	/* Make sure restore is not running concurrently. Since we
1766 	 * don't map invalid userptr BOs, we rely on the next restore
1767 	 * worker to do the mapping
1768 	 */
1769 	mutex_lock(&mem->process_info->lock);
1770 
1771 	/* Lock mmap-sem. If we find an invalid userptr BO, we can be
1772 	 * sure that the MMU notifier is no longer running
1773 	 * concurrently and the queues are actually stopped
1774 	 */
1775 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1776 		mmap_write_lock(current->mm);
1777 		is_invalid_userptr = atomic_read(&mem->invalid);
1778 		mmap_write_unlock(current->mm);
1779 	}
1780 
1781 	mutex_lock(&mem->lock);
1782 
1783 	domain = mem->domain;
1784 	bo_size = bo->tbo.base.size;
1785 
1786 	pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
1787 			mem->va,
1788 			mem->va + bo_size * (1 + mem->aql_queue),
1789 			avm, domain_string(domain));
1790 
1791 	if (!kfd_mem_is_attached(avm, mem)) {
1792 		ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue);
1793 		if (ret)
1794 			goto out;
1795 	}
1796 
1797 	ret = reserve_bo_and_vm(mem, avm, &ctx);
1798 	if (unlikely(ret))
1799 		goto out;
1800 
1801 	/* Userptr can be marked as "not invalid", but not actually be
1802 	 * validated yet (still in the system domain). In that case
1803 	 * the queues are still stopped and we can leave mapping for
1804 	 * the next restore worker
1805 	 */
1806 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
1807 	    bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
1808 		is_invalid_userptr = true;
1809 
1810 	ret = vm_validate_pt_pd_bos(avm);
1811 	if (unlikely(ret))
1812 		goto out_unreserve;
1813 
1814 	if (mem->mapped_to_gpu_memory == 0 &&
1815 	    !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1816 		/* Validate BO only once. The eviction fence gets added to BO
1817 		 * the first time it is mapped. Validate will wait for all
1818 		 * background evictions to complete.
1819 		 */
1820 		ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
1821 		if (ret) {
1822 			pr_debug("Validate failed\n");
1823 			goto out_unreserve;
1824 		}
1825 	}
1826 
1827 	list_for_each_entry(entry, &mem->attachments, list) {
1828 		if (entry->bo_va->base.vm != avm || entry->is_mapped)
1829 			continue;
1830 
1831 		pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
1832 			 entry->va, entry->va + bo_size, entry);
1833 
1834 		ret = map_bo_to_gpuvm(mem, entry, ctx.sync,
1835 				      is_invalid_userptr, table_freed);
1836 		if (ret) {
1837 			pr_err("Failed to map bo to gpuvm\n");
1838 			goto out_unreserve;
1839 		}
1840 
1841 		ret = vm_update_pds(avm, ctx.sync);
1842 		if (ret) {
1843 			pr_err("Failed to update page directories\n");
1844 			goto out_unreserve;
1845 		}
1846 
1847 		entry->is_mapped = true;
1848 		mem->mapped_to_gpu_memory++;
1849 		pr_debug("\t INC mapping count %d\n",
1850 			 mem->mapped_to_gpu_memory);
1851 	}
1852 
1853 	if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->tbo.pin_count)
1854 		amdgpu_bo_fence(bo,
1855 				&avm->process_info->eviction_fence->base,
1856 				true);
1857 	ret = unreserve_bo_and_vms(&ctx, false, false);
1858 
1859 	goto out;
1860 
1861 out_unreserve:
1862 	unreserve_bo_and_vms(&ctx, false, false);
1863 out:
1864 	mutex_unlock(&mem->process_info->lock);
1865 	mutex_unlock(&mem->lock);
1866 	return ret;
1867 }
1868 
1869 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1870 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
1871 {
1872 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1873 	struct amdkfd_process_info *process_info = avm->process_info;
1874 	unsigned long bo_size = mem->bo->tbo.base.size;
1875 	struct kfd_mem_attachment *entry;
1876 	struct bo_vm_reservation_context ctx;
1877 	int ret;
1878 
1879 	mutex_lock(&mem->lock);
1880 
1881 	ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx);
1882 	if (unlikely(ret))
1883 		goto out;
1884 	/* If no VMs were reserved, it means the BO wasn't actually mapped */
1885 	if (ctx.n_vms == 0) {
1886 		ret = -EINVAL;
1887 		goto unreserve_out;
1888 	}
1889 
1890 	ret = vm_validate_pt_pd_bos(avm);
1891 	if (unlikely(ret))
1892 		goto unreserve_out;
1893 
1894 	pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
1895 		mem->va,
1896 		mem->va + bo_size * (1 + mem->aql_queue),
1897 		avm);
1898 
1899 	list_for_each_entry(entry, &mem->attachments, list) {
1900 		if (entry->bo_va->base.vm != avm || !entry->is_mapped)
1901 			continue;
1902 
1903 		pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
1904 			 entry->va, entry->va + bo_size, entry);
1905 
1906 		unmap_bo_from_gpuvm(mem, entry, ctx.sync);
1907 		entry->is_mapped = false;
1908 
1909 		mem->mapped_to_gpu_memory--;
1910 		pr_debug("\t DEC mapping count %d\n",
1911 			 mem->mapped_to_gpu_memory);
1912 	}
1913 
1914 	/* If BO is unmapped from all VMs, unfence it. It can be evicted if
1915 	 * required.
1916 	 */
1917 	if (mem->mapped_to_gpu_memory == 0 &&
1918 	    !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) &&
1919 	    !mem->bo->tbo.pin_count)
1920 		amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1921 						process_info->eviction_fence);
1922 
1923 unreserve_out:
1924 	unreserve_bo_and_vms(&ctx, false, false);
1925 out:
1926 	mutex_unlock(&mem->lock);
1927 	return ret;
1928 }
1929 
1930 int amdgpu_amdkfd_gpuvm_sync_memory(
1931 		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr)
1932 {
1933 	struct amdgpu_sync sync;
1934 	int ret;
1935 
1936 	amdgpu_sync_create(&sync);
1937 
1938 	mutex_lock(&mem->lock);
1939 	amdgpu_sync_clone(&mem->sync, &sync);
1940 	mutex_unlock(&mem->lock);
1941 
1942 	ret = amdgpu_sync_wait(&sync, intr);
1943 	amdgpu_sync_free(&sync);
1944 	return ret;
1945 }
1946 
1947 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct amdgpu_device *adev,
1948 		struct kgd_mem *mem, void **kptr, uint64_t *size)
1949 {
1950 	int ret;
1951 	struct amdgpu_bo *bo = mem->bo;
1952 
1953 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1954 		pr_err("userptr can't be mapped to kernel\n");
1955 		return -EINVAL;
1956 	}
1957 
1958 	/* delete kgd_mem from kfd_bo_list to avoid re-validating
1959 	 * this BO in BO's restoring after eviction.
1960 	 */
1961 	mutex_lock(&mem->process_info->lock);
1962 
1963 	ret = amdgpu_bo_reserve(bo, true);
1964 	if (ret) {
1965 		pr_err("Failed to reserve bo. ret %d\n", ret);
1966 		goto bo_reserve_failed;
1967 	}
1968 
1969 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
1970 	if (ret) {
1971 		pr_err("Failed to pin bo. ret %d\n", ret);
1972 		goto pin_failed;
1973 	}
1974 
1975 	ret = amdgpu_bo_kmap(bo, kptr);
1976 	if (ret) {
1977 		pr_err("Failed to map bo to kernel. ret %d\n", ret);
1978 		goto kmap_failed;
1979 	}
1980 
1981 	amdgpu_amdkfd_remove_eviction_fence(
1982 		bo, mem->process_info->eviction_fence);
1983 	list_del_init(&mem->validate_list.head);
1984 
1985 	if (size)
1986 		*size = amdgpu_bo_size(bo);
1987 
1988 	amdgpu_bo_unreserve(bo);
1989 
1990 	mutex_unlock(&mem->process_info->lock);
1991 	return 0;
1992 
1993 kmap_failed:
1994 	amdgpu_bo_unpin(bo);
1995 pin_failed:
1996 	amdgpu_bo_unreserve(bo);
1997 bo_reserve_failed:
1998 	mutex_unlock(&mem->process_info->lock);
1999 
2000 	return ret;
2001 }
2002 
2003 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct amdgpu_device *adev,
2004 						  struct kgd_mem *mem)
2005 {
2006 	struct amdgpu_bo *bo = mem->bo;
2007 
2008 	amdgpu_bo_reserve(bo, true);
2009 	amdgpu_bo_kunmap(bo);
2010 	amdgpu_bo_unpin(bo);
2011 	amdgpu_bo_unreserve(bo);
2012 }
2013 
2014 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
2015 					  struct kfd_vm_fault_info *mem)
2016 {
2017 	if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
2018 		*mem = *adev->gmc.vm_fault_info;
2019 		mb();
2020 		atomic_set(&adev->gmc.vm_fault_info_updated, 0);
2021 	}
2022 	return 0;
2023 }
2024 
2025 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
2026 				      struct dma_buf *dma_buf,
2027 				      uint64_t va, void *drm_priv,
2028 				      struct kgd_mem **mem, uint64_t *size,
2029 				      uint64_t *mmap_offset)
2030 {
2031 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2032 	struct drm_gem_object *obj;
2033 	struct amdgpu_bo *bo;
2034 	int ret;
2035 
2036 	if (dma_buf->ops != &amdgpu_dmabuf_ops)
2037 		/* Can't handle non-graphics buffers */
2038 		return -EINVAL;
2039 
2040 	obj = dma_buf->priv;
2041 	if (drm_to_adev(obj->dev) != adev)
2042 		/* Can't handle buffers from other devices */
2043 		return -EINVAL;
2044 
2045 	bo = gem_to_amdgpu_bo(obj);
2046 	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
2047 				    AMDGPU_GEM_DOMAIN_GTT)))
2048 		/* Only VRAM and GTT BOs are supported */
2049 		return -EINVAL;
2050 
2051 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2052 	if (!*mem)
2053 		return -ENOMEM;
2054 
2055 	ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
2056 	if (ret) {
2057 		kfree(mem);
2058 		return ret;
2059 	}
2060 
2061 	if (size)
2062 		*size = amdgpu_bo_size(bo);
2063 
2064 	if (mmap_offset)
2065 		*mmap_offset = amdgpu_bo_mmap_offset(bo);
2066 
2067 	INIT_LIST_HEAD(&(*mem)->attachments);
2068 	mutex_init(&(*mem)->lock);
2069 
2070 	(*mem)->alloc_flags =
2071 		((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2072 		KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
2073 		| KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
2074 		| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
2075 
2076 	drm_gem_object_get(&bo->tbo.base);
2077 	(*mem)->bo = bo;
2078 	(*mem)->va = va;
2079 	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2080 		AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
2081 	(*mem)->mapped_to_gpu_memory = 0;
2082 	(*mem)->process_info = avm->process_info;
2083 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
2084 	amdgpu_sync_create(&(*mem)->sync);
2085 	(*mem)->is_imported = true;
2086 
2087 	return 0;
2088 }
2089 
2090 /* Evict a userptr BO by stopping the queues if necessary
2091  *
2092  * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
2093  * cannot do any memory allocations, and cannot take any locks that
2094  * are held elsewhere while allocating memory. Therefore this is as
2095  * simple as possible, using atomic counters.
2096  *
2097  * It doesn't do anything to the BO itself. The real work happens in
2098  * restore, where we get updated page addresses. This function only
2099  * ensures that GPU access to the BO is stopped.
2100  */
2101 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem,
2102 				struct mm_struct *mm)
2103 {
2104 	struct amdkfd_process_info *process_info = mem->process_info;
2105 	int evicted_bos;
2106 	int r = 0;
2107 
2108 	/* Do not process MMU notifications until stage-4 IOCTL is received */
2109 	if (READ_ONCE(process_info->block_mmu_notifications))
2110 		return 0;
2111 
2112 	atomic_inc(&mem->invalid);
2113 	evicted_bos = atomic_inc_return(&process_info->evicted_bos);
2114 	if (evicted_bos == 1) {
2115 		/* First eviction, stop the queues */
2116 		r = kgd2kfd_quiesce_mm(mm);
2117 		if (r)
2118 			pr_err("Failed to quiesce KFD\n");
2119 		schedule_delayed_work(&process_info->restore_userptr_work,
2120 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2121 	}
2122 
2123 	return r;
2124 }
2125 
2126 /* Update invalid userptr BOs
2127  *
2128  * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
2129  * userptr_inval_list and updates user pages for all BOs that have
2130  * been invalidated since their last update.
2131  */
2132 static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
2133 				     struct mm_struct *mm)
2134 {
2135 	struct kgd_mem *mem, *tmp_mem;
2136 	struct amdgpu_bo *bo;
2137 	struct ttm_operation_ctx ctx = { false, false };
2138 	int invalid, ret;
2139 
2140 	/* Move all invalidated BOs to the userptr_inval_list and
2141 	 * release their user pages by migration to the CPU domain
2142 	 */
2143 	list_for_each_entry_safe(mem, tmp_mem,
2144 				 &process_info->userptr_valid_list,
2145 				 validate_list.head) {
2146 		if (!atomic_read(&mem->invalid))
2147 			continue; /* BO is still valid */
2148 
2149 		bo = mem->bo;
2150 
2151 		if (amdgpu_bo_reserve(bo, true))
2152 			return -EAGAIN;
2153 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
2154 		ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2155 		amdgpu_bo_unreserve(bo);
2156 		if (ret) {
2157 			pr_err("%s: Failed to invalidate userptr BO\n",
2158 			       __func__);
2159 			return -EAGAIN;
2160 		}
2161 
2162 		list_move_tail(&mem->validate_list.head,
2163 			       &process_info->userptr_inval_list);
2164 	}
2165 
2166 	if (list_empty(&process_info->userptr_inval_list))
2167 		return 0; /* All evicted userptr BOs were freed */
2168 
2169 	/* Go through userptr_inval_list and update any invalid user_pages */
2170 	list_for_each_entry(mem, &process_info->userptr_inval_list,
2171 			    validate_list.head) {
2172 		invalid = atomic_read(&mem->invalid);
2173 		if (!invalid)
2174 			/* BO hasn't been invalidated since the last
2175 			 * revalidation attempt. Keep its BO list.
2176 			 */
2177 			continue;
2178 
2179 		bo = mem->bo;
2180 
2181 		/* Get updated user pages */
2182 		ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
2183 		if (ret) {
2184 			pr_debug("Failed %d to get user pages\n", ret);
2185 
2186 			/* Return -EFAULT bad address error as success. It will
2187 			 * fail later with a VM fault if the GPU tries to access
2188 			 * it. Better than hanging indefinitely with stalled
2189 			 * user mode queues.
2190 			 *
2191 			 * Return other error -EBUSY or -ENOMEM to retry restore
2192 			 */
2193 			if (ret != -EFAULT)
2194 				return ret;
2195 		} else {
2196 
2197 			/*
2198 			 * FIXME: Cannot ignore the return code, must hold
2199 			 * notifier_lock
2200 			 */
2201 			amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
2202 		}
2203 
2204 		/* Mark the BO as valid unless it was invalidated
2205 		 * again concurrently.
2206 		 */
2207 		if (atomic_cmpxchg(&mem->invalid, invalid, 0) != invalid)
2208 			return -EAGAIN;
2209 	}
2210 
2211 	return 0;
2212 }
2213 
2214 /* Validate invalid userptr BOs
2215  *
2216  * Validates BOs on the userptr_inval_list, and moves them back to the
2217  * userptr_valid_list. Also updates GPUVM page tables with new page
2218  * addresses and waits for the page table updates to complete.
2219  */
2220 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
2221 {
2222 	struct amdgpu_bo_list_entry *pd_bo_list_entries;
2223 	struct list_head resv_list, duplicates;
2224 	struct ww_acquire_ctx ticket;
2225 	struct amdgpu_sync sync;
2226 
2227 	struct amdgpu_vm *peer_vm;
2228 	struct kgd_mem *mem, *tmp_mem;
2229 	struct amdgpu_bo *bo;
2230 	struct ttm_operation_ctx ctx = { false, false };
2231 	int i, ret;
2232 
2233 	pd_bo_list_entries = kcalloc(process_info->n_vms,
2234 				     sizeof(struct amdgpu_bo_list_entry),
2235 				     GFP_KERNEL);
2236 	if (!pd_bo_list_entries) {
2237 		pr_err("%s: Failed to allocate PD BO list entries\n", __func__);
2238 		ret = -ENOMEM;
2239 		goto out_no_mem;
2240 	}
2241 
2242 	INIT_LIST_HEAD(&resv_list);
2243 	INIT_LIST_HEAD(&duplicates);
2244 
2245 	/* Get all the page directory BOs that need to be reserved */
2246 	i = 0;
2247 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2248 			    vm_list_node)
2249 		amdgpu_vm_get_pd_bo(peer_vm, &resv_list,
2250 				    &pd_bo_list_entries[i++]);
2251 	/* Add the userptr_inval_list entries to resv_list */
2252 	list_for_each_entry(mem, &process_info->userptr_inval_list,
2253 			    validate_list.head) {
2254 		list_add_tail(&mem->resv_list.head, &resv_list);
2255 		mem->resv_list.bo = mem->validate_list.bo;
2256 		mem->resv_list.num_shared = mem->validate_list.num_shared;
2257 	}
2258 
2259 	/* Reserve all BOs and page tables for validation */
2260 	ret = ttm_eu_reserve_buffers(&ticket, &resv_list, false, &duplicates);
2261 	WARN(!list_empty(&duplicates), "Duplicates should be empty");
2262 	if (ret)
2263 		goto out_free;
2264 
2265 	amdgpu_sync_create(&sync);
2266 
2267 	ret = process_validate_vms(process_info);
2268 	if (ret)
2269 		goto unreserve_out;
2270 
2271 	/* Validate BOs and update GPUVM page tables */
2272 	list_for_each_entry_safe(mem, tmp_mem,
2273 				 &process_info->userptr_inval_list,
2274 				 validate_list.head) {
2275 		struct kfd_mem_attachment *attachment;
2276 
2277 		bo = mem->bo;
2278 
2279 		/* Validate the BO if we got user pages */
2280 		if (bo->tbo.ttm->pages[0]) {
2281 			amdgpu_bo_placement_from_domain(bo, mem->domain);
2282 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2283 			if (ret) {
2284 				pr_err("%s: failed to validate BO\n", __func__);
2285 				goto unreserve_out;
2286 			}
2287 		}
2288 
2289 		list_move_tail(&mem->validate_list.head,
2290 			       &process_info->userptr_valid_list);
2291 
2292 		/* Update mapping. If the BO was not validated
2293 		 * (because we couldn't get user pages), this will
2294 		 * clear the page table entries, which will result in
2295 		 * VM faults if the GPU tries to access the invalid
2296 		 * memory.
2297 		 */
2298 		list_for_each_entry(attachment, &mem->attachments, list) {
2299 			if (!attachment->is_mapped)
2300 				continue;
2301 
2302 			kfd_mem_dmaunmap_attachment(mem, attachment);
2303 			ret = update_gpuvm_pte(mem, attachment, &sync, NULL);
2304 			if (ret) {
2305 				pr_err("%s: update PTE failed\n", __func__);
2306 				/* make sure this gets validated again */
2307 				atomic_inc(&mem->invalid);
2308 				goto unreserve_out;
2309 			}
2310 		}
2311 	}
2312 
2313 	/* Update page directories */
2314 	ret = process_update_pds(process_info, &sync);
2315 
2316 unreserve_out:
2317 	ttm_eu_backoff_reservation(&ticket, &resv_list);
2318 	amdgpu_sync_wait(&sync, false);
2319 	amdgpu_sync_free(&sync);
2320 out_free:
2321 	kfree(pd_bo_list_entries);
2322 out_no_mem:
2323 
2324 	return ret;
2325 }
2326 
2327 /* Worker callback to restore evicted userptr BOs
2328  *
2329  * Tries to update and validate all userptr BOs. If successful and no
2330  * concurrent evictions happened, the queues are restarted. Otherwise,
2331  * reschedule for another attempt later.
2332  */
2333 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
2334 {
2335 	struct delayed_work *dwork = to_delayed_work(work);
2336 	struct amdkfd_process_info *process_info =
2337 		container_of(dwork, struct amdkfd_process_info,
2338 			     restore_userptr_work);
2339 	struct task_struct *usertask;
2340 	struct mm_struct *mm;
2341 	int evicted_bos;
2342 
2343 	evicted_bos = atomic_read(&process_info->evicted_bos);
2344 	if (!evicted_bos)
2345 		return;
2346 
2347 	/* Reference task and mm in case of concurrent process termination */
2348 	usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
2349 	if (!usertask)
2350 		return;
2351 	mm = get_task_mm(usertask);
2352 	if (!mm) {
2353 		put_task_struct(usertask);
2354 		return;
2355 	}
2356 
2357 	mutex_lock(&process_info->lock);
2358 
2359 	if (update_invalid_user_pages(process_info, mm))
2360 		goto unlock_out;
2361 	/* userptr_inval_list can be empty if all evicted userptr BOs
2362 	 * have been freed. In that case there is nothing to validate
2363 	 * and we can just restart the queues.
2364 	 */
2365 	if (!list_empty(&process_info->userptr_inval_list)) {
2366 		if (atomic_read(&process_info->evicted_bos) != evicted_bos)
2367 			goto unlock_out; /* Concurrent eviction, try again */
2368 
2369 		if (validate_invalid_user_pages(process_info))
2370 			goto unlock_out;
2371 	}
2372 	/* Final check for concurrent evicton and atomic update. If
2373 	 * another eviction happens after successful update, it will
2374 	 * be a first eviction that calls quiesce_mm. The eviction
2375 	 * reference counting inside KFD will handle this case.
2376 	 */
2377 	if (atomic_cmpxchg(&process_info->evicted_bos, evicted_bos, 0) !=
2378 	    evicted_bos)
2379 		goto unlock_out;
2380 	evicted_bos = 0;
2381 	if (kgd2kfd_resume_mm(mm)) {
2382 		pr_err("%s: Failed to resume KFD\n", __func__);
2383 		/* No recovery from this failure. Probably the CP is
2384 		 * hanging. No point trying again.
2385 		 */
2386 	}
2387 
2388 unlock_out:
2389 	mutex_unlock(&process_info->lock);
2390 	mmput(mm);
2391 	put_task_struct(usertask);
2392 
2393 	/* If validation failed, reschedule another attempt */
2394 	if (evicted_bos)
2395 		schedule_delayed_work(&process_info->restore_userptr_work,
2396 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2397 }
2398 
2399 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2400  *   KFD process identified by process_info
2401  *
2402  * @process_info: amdkfd_process_info of the KFD process
2403  *
2404  * After memory eviction, restore thread calls this function. The function
2405  * should be called when the Process is still valid. BO restore involves -
2406  *
2407  * 1.  Release old eviction fence and create new one
2408  * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2409  * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2410  *     BOs that need to be reserved.
2411  * 4.  Reserve all the BOs
2412  * 5.  Validate of PD and PT BOs.
2413  * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2414  * 7.  Add fence to all PD and PT BOs.
2415  * 8.  Unreserve all BOs
2416  */
2417 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
2418 {
2419 	struct amdgpu_bo_list_entry *pd_bo_list;
2420 	struct amdkfd_process_info *process_info = info;
2421 	struct amdgpu_vm *peer_vm;
2422 	struct kgd_mem *mem;
2423 	struct bo_vm_reservation_context ctx;
2424 	struct amdgpu_amdkfd_fence *new_fence;
2425 	int ret = 0, i;
2426 	struct list_head duplicate_save;
2427 	struct amdgpu_sync sync_obj;
2428 	unsigned long failed_size = 0;
2429 	unsigned long total_size = 0;
2430 
2431 	INIT_LIST_HEAD(&duplicate_save);
2432 	INIT_LIST_HEAD(&ctx.list);
2433 	INIT_LIST_HEAD(&ctx.duplicates);
2434 
2435 	pd_bo_list = kcalloc(process_info->n_vms,
2436 			     sizeof(struct amdgpu_bo_list_entry),
2437 			     GFP_KERNEL);
2438 	if (!pd_bo_list)
2439 		return -ENOMEM;
2440 
2441 	i = 0;
2442 	mutex_lock(&process_info->lock);
2443 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2444 			vm_list_node)
2445 		amdgpu_vm_get_pd_bo(peer_vm, &ctx.list, &pd_bo_list[i++]);
2446 
2447 	/* Reserve all BOs and page tables/directory. Add all BOs from
2448 	 * kfd_bo_list to ctx.list
2449 	 */
2450 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2451 			    validate_list.head) {
2452 
2453 		list_add_tail(&mem->resv_list.head, &ctx.list);
2454 		mem->resv_list.bo = mem->validate_list.bo;
2455 		mem->resv_list.num_shared = mem->validate_list.num_shared;
2456 	}
2457 
2458 	ret = ttm_eu_reserve_buffers(&ctx.ticket, &ctx.list,
2459 				     false, &duplicate_save);
2460 	if (ret) {
2461 		pr_debug("Memory eviction: TTM Reserve Failed. Try again\n");
2462 		goto ttm_reserve_fail;
2463 	}
2464 
2465 	amdgpu_sync_create(&sync_obj);
2466 
2467 	/* Validate PDs and PTs */
2468 	ret = process_validate_vms(process_info);
2469 	if (ret)
2470 		goto validate_map_fail;
2471 
2472 	ret = process_sync_pds_resv(process_info, &sync_obj);
2473 	if (ret) {
2474 		pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2475 		goto validate_map_fail;
2476 	}
2477 
2478 	/* Validate BOs and map them to GPUVM (update VM page tables). */
2479 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2480 			    validate_list.head) {
2481 
2482 		struct amdgpu_bo *bo = mem->bo;
2483 		uint32_t domain = mem->domain;
2484 		struct kfd_mem_attachment *attachment;
2485 
2486 		total_size += amdgpu_bo_size(bo);
2487 
2488 		ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2489 		if (ret) {
2490 			pr_debug("Memory eviction: Validate BOs failed\n");
2491 			failed_size += amdgpu_bo_size(bo);
2492 			ret = amdgpu_amdkfd_bo_validate(bo,
2493 						AMDGPU_GEM_DOMAIN_GTT, false);
2494 			if (ret) {
2495 				pr_debug("Memory eviction: Try again\n");
2496 				goto validate_map_fail;
2497 			}
2498 		}
2499 		ret = amdgpu_sync_fence(&sync_obj, bo->tbo.moving);
2500 		if (ret) {
2501 			pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2502 			goto validate_map_fail;
2503 		}
2504 		list_for_each_entry(attachment, &mem->attachments, list) {
2505 			if (!attachment->is_mapped)
2506 				continue;
2507 
2508 			kfd_mem_dmaunmap_attachment(mem, attachment);
2509 			ret = update_gpuvm_pte(mem, attachment, &sync_obj, NULL);
2510 			if (ret) {
2511 				pr_debug("Memory eviction: update PTE failed. Try again\n");
2512 				goto validate_map_fail;
2513 			}
2514 		}
2515 	}
2516 
2517 	if (failed_size)
2518 		pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2519 
2520 	/* Update page directories */
2521 	ret = process_update_pds(process_info, &sync_obj);
2522 	if (ret) {
2523 		pr_debug("Memory eviction: update PDs failed. Try again\n");
2524 		goto validate_map_fail;
2525 	}
2526 
2527 	/* Wait for validate and PT updates to finish */
2528 	amdgpu_sync_wait(&sync_obj, false);
2529 
2530 	/* Release old eviction fence and create new one, because fence only
2531 	 * goes from unsignaled to signaled, fence cannot be reused.
2532 	 * Use context and mm from the old fence.
2533 	 */
2534 	new_fence = amdgpu_amdkfd_fence_create(
2535 				process_info->eviction_fence->base.context,
2536 				process_info->eviction_fence->mm,
2537 				NULL);
2538 	if (!new_fence) {
2539 		pr_err("Failed to create eviction fence\n");
2540 		ret = -ENOMEM;
2541 		goto validate_map_fail;
2542 	}
2543 	dma_fence_put(&process_info->eviction_fence->base);
2544 	process_info->eviction_fence = new_fence;
2545 	*ef = dma_fence_get(&new_fence->base);
2546 
2547 	/* Attach new eviction fence to all BOs */
2548 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2549 		validate_list.head)
2550 		amdgpu_bo_fence(mem->bo,
2551 			&process_info->eviction_fence->base, true);
2552 
2553 	/* Attach eviction fence to PD / PT BOs */
2554 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2555 			    vm_list_node) {
2556 		struct amdgpu_bo *bo = peer_vm->root.bo;
2557 
2558 		amdgpu_bo_fence(bo, &process_info->eviction_fence->base, true);
2559 	}
2560 
2561 validate_map_fail:
2562 	ttm_eu_backoff_reservation(&ctx.ticket, &ctx.list);
2563 	amdgpu_sync_free(&sync_obj);
2564 ttm_reserve_fail:
2565 	mutex_unlock(&process_info->lock);
2566 	kfree(pd_bo_list);
2567 	return ret;
2568 }
2569 
2570 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
2571 {
2572 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2573 	struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
2574 	int ret;
2575 
2576 	if (!info || !gws)
2577 		return -EINVAL;
2578 
2579 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2580 	if (!*mem)
2581 		return -ENOMEM;
2582 
2583 	mutex_init(&(*mem)->lock);
2584 	INIT_LIST_HEAD(&(*mem)->attachments);
2585 	(*mem)->bo = amdgpu_bo_ref(gws_bo);
2586 	(*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
2587 	(*mem)->process_info = process_info;
2588 	add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
2589 	amdgpu_sync_create(&(*mem)->sync);
2590 
2591 
2592 	/* Validate gws bo the first time it is added to process */
2593 	mutex_lock(&(*mem)->process_info->lock);
2594 	ret = amdgpu_bo_reserve(gws_bo, false);
2595 	if (unlikely(ret)) {
2596 		pr_err("Reserve gws bo failed %d\n", ret);
2597 		goto bo_reservation_failure;
2598 	}
2599 
2600 	ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
2601 	if (ret) {
2602 		pr_err("GWS BO validate failed %d\n", ret);
2603 		goto bo_validation_failure;
2604 	}
2605 	/* GWS resource is shared b/t amdgpu and amdkfd
2606 	 * Add process eviction fence to bo so they can
2607 	 * evict each other.
2608 	 */
2609 	ret = dma_resv_reserve_shared(gws_bo->tbo.base.resv, 1);
2610 	if (ret)
2611 		goto reserve_shared_fail;
2612 	amdgpu_bo_fence(gws_bo, &process_info->eviction_fence->base, true);
2613 	amdgpu_bo_unreserve(gws_bo);
2614 	mutex_unlock(&(*mem)->process_info->lock);
2615 
2616 	return ret;
2617 
2618 reserve_shared_fail:
2619 bo_validation_failure:
2620 	amdgpu_bo_unreserve(gws_bo);
2621 bo_reservation_failure:
2622 	mutex_unlock(&(*mem)->process_info->lock);
2623 	amdgpu_sync_free(&(*mem)->sync);
2624 	remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
2625 	amdgpu_bo_unref(&gws_bo);
2626 	mutex_destroy(&(*mem)->lock);
2627 	kfree(*mem);
2628 	*mem = NULL;
2629 	return ret;
2630 }
2631 
2632 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
2633 {
2634 	int ret;
2635 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2636 	struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
2637 	struct amdgpu_bo *gws_bo = kgd_mem->bo;
2638 
2639 	/* Remove BO from process's validate list so restore worker won't touch
2640 	 * it anymore
2641 	 */
2642 	remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
2643 
2644 	ret = amdgpu_bo_reserve(gws_bo, false);
2645 	if (unlikely(ret)) {
2646 		pr_err("Reserve gws bo failed %d\n", ret);
2647 		//TODO add BO back to validate_list?
2648 		return ret;
2649 	}
2650 	amdgpu_amdkfd_remove_eviction_fence(gws_bo,
2651 			process_info->eviction_fence);
2652 	amdgpu_bo_unreserve(gws_bo);
2653 	amdgpu_sync_free(&kgd_mem->sync);
2654 	amdgpu_bo_unref(&gws_bo);
2655 	mutex_destroy(&kgd_mem->lock);
2656 	kfree(mem);
2657 	return 0;
2658 }
2659 
2660 /* Returns GPU-specific tiling mode information */
2661 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
2662 				struct tile_config *config)
2663 {
2664 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
2665 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
2666 	config->num_tile_configs =
2667 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
2668 	config->macro_tile_config_ptr =
2669 			adev->gfx.config.macrotile_mode_array;
2670 	config->num_macro_tile_configs =
2671 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
2672 
2673 	/* Those values are not set from GFX9 onwards */
2674 	config->num_banks = adev->gfx.config.num_banks;
2675 	config->num_ranks = adev->gfx.config.num_ranks;
2676 
2677 	return 0;
2678 }
2679 
2680 bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *mem)
2681 {
2682 	struct kfd_mem_attachment *entry;
2683 
2684 	list_for_each_entry(entry, &mem->attachments, list) {
2685 		if (entry->is_mapped && entry->adev == adev)
2686 			return true;
2687 	}
2688 	return false;
2689 }
2690