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_rmv(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_rmv(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 {
847 	struct amdkfd_process_info *process_info = mem->process_info;
848 	struct amdgpu_bo *bo = mem->bo;
849 	struct ttm_operation_ctx ctx = { true, false };
850 	int ret = 0;
851 
852 	mutex_lock(&process_info->lock);
853 
854 	ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
855 	if (ret) {
856 		pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
857 		goto out;
858 	}
859 
860 	ret = amdgpu_mn_register(bo, user_addr);
861 	if (ret) {
862 		pr_err("%s: Failed to register MMU notifier: %d\n",
863 		       __func__, ret);
864 		goto out;
865 	}
866 
867 	ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
868 	if (ret) {
869 		pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
870 		goto unregister_out;
871 	}
872 
873 	ret = amdgpu_bo_reserve(bo, true);
874 	if (ret) {
875 		pr_err("%s: Failed to reserve BO\n", __func__);
876 		goto release_out;
877 	}
878 	amdgpu_bo_placement_from_domain(bo, mem->domain);
879 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
880 	if (ret)
881 		pr_err("%s: failed to validate BO\n", __func__);
882 	amdgpu_bo_unreserve(bo);
883 
884 release_out:
885 	amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
886 unregister_out:
887 	if (ret)
888 		amdgpu_mn_unregister(bo);
889 out:
890 	mutex_unlock(&process_info->lock);
891 	return ret;
892 }
893 
894 /* Reserving a BO and its page table BOs must happen atomically to
895  * avoid deadlocks. Some operations update multiple VMs at once. Track
896  * all the reservation info in a context structure. Optionally a sync
897  * object can track VM updates.
898  */
899 struct bo_vm_reservation_context {
900 	struct amdgpu_bo_list_entry kfd_bo; /* BO list entry for the KFD BO */
901 	unsigned int n_vms;		    /* Number of VMs reserved	    */
902 	struct amdgpu_bo_list_entry *vm_pd; /* Array of VM BO list entries  */
903 	struct ww_acquire_ctx ticket;	    /* Reservation ticket	    */
904 	struct list_head list, duplicates;  /* BO lists			    */
905 	struct amdgpu_sync *sync;	    /* Pointer to sync object	    */
906 	bool reserved;			    /* Whether BOs are reserved	    */
907 };
908 
909 enum bo_vm_match {
910 	BO_VM_NOT_MAPPED = 0,	/* Match VMs where a BO is not mapped */
911 	BO_VM_MAPPED,		/* Match VMs where a BO is mapped     */
912 	BO_VM_ALL,		/* Match all VMs a BO was added to    */
913 };
914 
915 /**
916  * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
917  * @mem: KFD BO structure.
918  * @vm: the VM to reserve.
919  * @ctx: the struct that will be used in unreserve_bo_and_vms().
920  */
921 static int reserve_bo_and_vm(struct kgd_mem *mem,
922 			      struct amdgpu_vm *vm,
923 			      struct bo_vm_reservation_context *ctx)
924 {
925 	struct amdgpu_bo *bo = mem->bo;
926 	int ret;
927 
928 	WARN_ON(!vm);
929 
930 	ctx->reserved = false;
931 	ctx->n_vms = 1;
932 	ctx->sync = &mem->sync;
933 
934 	INIT_LIST_HEAD(&ctx->list);
935 	INIT_LIST_HEAD(&ctx->duplicates);
936 
937 	ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), GFP_KERNEL);
938 	if (!ctx->vm_pd)
939 		return -ENOMEM;
940 
941 	ctx->kfd_bo.priority = 0;
942 	ctx->kfd_bo.tv.bo = &bo->tbo;
943 	ctx->kfd_bo.tv.num_shared = 1;
944 	list_add(&ctx->kfd_bo.tv.head, &ctx->list);
945 
946 	amdgpu_vm_get_pd_bo(vm, &ctx->list, &ctx->vm_pd[0]);
947 
948 	ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
949 				     false, &ctx->duplicates);
950 	if (ret) {
951 		pr_err("Failed to reserve buffers in ttm.\n");
952 		kfree(ctx->vm_pd);
953 		ctx->vm_pd = NULL;
954 		return ret;
955 	}
956 
957 	ctx->reserved = true;
958 	return 0;
959 }
960 
961 /**
962  * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
963  * @mem: KFD BO structure.
964  * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
965  * is used. Otherwise, a single VM associated with the BO.
966  * @map_type: the mapping status that will be used to filter the VMs.
967  * @ctx: the struct that will be used in unreserve_bo_and_vms().
968  *
969  * Returns 0 for success, negative for failure.
970  */
971 static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
972 				struct amdgpu_vm *vm, enum bo_vm_match map_type,
973 				struct bo_vm_reservation_context *ctx)
974 {
975 	struct amdgpu_bo *bo = mem->bo;
976 	struct kfd_mem_attachment *entry;
977 	unsigned int i;
978 	int ret;
979 
980 	ctx->reserved = false;
981 	ctx->n_vms = 0;
982 	ctx->vm_pd = NULL;
983 	ctx->sync = &mem->sync;
984 
985 	INIT_LIST_HEAD(&ctx->list);
986 	INIT_LIST_HEAD(&ctx->duplicates);
987 
988 	list_for_each_entry(entry, &mem->attachments, list) {
989 		if ((vm && vm != entry->bo_va->base.vm) ||
990 			(entry->is_mapped != map_type
991 			&& map_type != BO_VM_ALL))
992 			continue;
993 
994 		ctx->n_vms++;
995 	}
996 
997 	if (ctx->n_vms != 0) {
998 		ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd),
999 				     GFP_KERNEL);
1000 		if (!ctx->vm_pd)
1001 			return -ENOMEM;
1002 	}
1003 
1004 	ctx->kfd_bo.priority = 0;
1005 	ctx->kfd_bo.tv.bo = &bo->tbo;
1006 	ctx->kfd_bo.tv.num_shared = 1;
1007 	list_add(&ctx->kfd_bo.tv.head, &ctx->list);
1008 
1009 	i = 0;
1010 	list_for_each_entry(entry, &mem->attachments, list) {
1011 		if ((vm && vm != entry->bo_va->base.vm) ||
1012 			(entry->is_mapped != map_type
1013 			&& map_type != BO_VM_ALL))
1014 			continue;
1015 
1016 		amdgpu_vm_get_pd_bo(entry->bo_va->base.vm, &ctx->list,
1017 				&ctx->vm_pd[i]);
1018 		i++;
1019 	}
1020 
1021 	ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
1022 				     false, &ctx->duplicates);
1023 	if (ret) {
1024 		pr_err("Failed to reserve buffers in ttm.\n");
1025 		kfree(ctx->vm_pd);
1026 		ctx->vm_pd = NULL;
1027 		return ret;
1028 	}
1029 
1030 	ctx->reserved = true;
1031 	return 0;
1032 }
1033 
1034 /**
1035  * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1036  * @ctx: Reservation context to unreserve
1037  * @wait: Optionally wait for a sync object representing pending VM updates
1038  * @intr: Whether the wait is interruptible
1039  *
1040  * Also frees any resources allocated in
1041  * reserve_bo_and_(cond_)vm(s). Returns the status from
1042  * amdgpu_sync_wait.
1043  */
1044 static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
1045 				 bool wait, bool intr)
1046 {
1047 	int ret = 0;
1048 
1049 	if (wait)
1050 		ret = amdgpu_sync_wait(ctx->sync, intr);
1051 
1052 	if (ctx->reserved)
1053 		ttm_eu_backoff_reservation(&ctx->ticket, &ctx->list);
1054 	kfree(ctx->vm_pd);
1055 
1056 	ctx->sync = NULL;
1057 
1058 	ctx->reserved = false;
1059 	ctx->vm_pd = NULL;
1060 
1061 	return ret;
1062 }
1063 
1064 static void unmap_bo_from_gpuvm(struct kgd_mem *mem,
1065 				struct kfd_mem_attachment *entry,
1066 				struct amdgpu_sync *sync)
1067 {
1068 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1069 	struct amdgpu_device *adev = entry->adev;
1070 	struct amdgpu_vm *vm = bo_va->base.vm;
1071 
1072 	amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
1073 
1074 	amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
1075 
1076 	amdgpu_sync_fence(sync, bo_va->last_pt_update);
1077 
1078 	kfd_mem_dmaunmap_attachment(mem, entry);
1079 }
1080 
1081 static int update_gpuvm_pte(struct kgd_mem *mem,
1082 			    struct kfd_mem_attachment *entry,
1083 			    struct amdgpu_sync *sync,
1084 			    bool *table_freed)
1085 {
1086 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1087 	struct amdgpu_device *adev = entry->adev;
1088 	int ret;
1089 
1090 	ret = kfd_mem_dmamap_attachment(mem, entry);
1091 	if (ret)
1092 		return ret;
1093 
1094 	/* Update the page tables  */
1095 	ret = amdgpu_vm_bo_update(adev, bo_va, false, table_freed);
1096 	if (ret) {
1097 		pr_err("amdgpu_vm_bo_update failed\n");
1098 		return ret;
1099 	}
1100 
1101 	return amdgpu_sync_fence(sync, bo_va->last_pt_update);
1102 }
1103 
1104 static int map_bo_to_gpuvm(struct kgd_mem *mem,
1105 			   struct kfd_mem_attachment *entry,
1106 			   struct amdgpu_sync *sync,
1107 			   bool no_update_pte,
1108 			   bool *table_freed)
1109 {
1110 	int ret;
1111 
1112 	/* Set virtual address for the allocation */
1113 	ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0,
1114 			       amdgpu_bo_size(entry->bo_va->base.bo),
1115 			       entry->pte_flags);
1116 	if (ret) {
1117 		pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
1118 				entry->va, ret);
1119 		return ret;
1120 	}
1121 
1122 	if (no_update_pte)
1123 		return 0;
1124 
1125 	ret = update_gpuvm_pte(mem, entry, sync, table_freed);
1126 	if (ret) {
1127 		pr_err("update_gpuvm_pte() failed\n");
1128 		goto update_gpuvm_pte_failed;
1129 	}
1130 
1131 	return 0;
1132 
1133 update_gpuvm_pte_failed:
1134 	unmap_bo_from_gpuvm(mem, entry, sync);
1135 	return ret;
1136 }
1137 
1138 static struct sg_table *create_doorbell_sg(uint64_t addr, uint32_t size)
1139 {
1140 	struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
1141 
1142 	if (!sg)
1143 		return NULL;
1144 	if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
1145 		kfree(sg);
1146 		return NULL;
1147 	}
1148 	sg->sgl->dma_address = addr;
1149 	sg->sgl->length = size;
1150 #ifdef CONFIG_NEED_SG_DMA_LENGTH
1151 	sg->sgl->dma_length = size;
1152 #endif
1153 	return sg;
1154 }
1155 
1156 static int process_validate_vms(struct amdkfd_process_info *process_info)
1157 {
1158 	struct amdgpu_vm *peer_vm;
1159 	int ret;
1160 
1161 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1162 			    vm_list_node) {
1163 		ret = vm_validate_pt_pd_bos(peer_vm);
1164 		if (ret)
1165 			return ret;
1166 	}
1167 
1168 	return 0;
1169 }
1170 
1171 static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
1172 				 struct amdgpu_sync *sync)
1173 {
1174 	struct amdgpu_vm *peer_vm;
1175 	int ret;
1176 
1177 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1178 			    vm_list_node) {
1179 		struct amdgpu_bo *pd = peer_vm->root.bo;
1180 
1181 		ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
1182 				       AMDGPU_SYNC_NE_OWNER,
1183 				       AMDGPU_FENCE_OWNER_KFD);
1184 		if (ret)
1185 			return ret;
1186 	}
1187 
1188 	return 0;
1189 }
1190 
1191 static int process_update_pds(struct amdkfd_process_info *process_info,
1192 			      struct amdgpu_sync *sync)
1193 {
1194 	struct amdgpu_vm *peer_vm;
1195 	int ret;
1196 
1197 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1198 			    vm_list_node) {
1199 		ret = vm_update_pds(peer_vm, sync);
1200 		if (ret)
1201 			return ret;
1202 	}
1203 
1204 	return 0;
1205 }
1206 
1207 static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
1208 		       struct dma_fence **ef)
1209 {
1210 	struct amdkfd_process_info *info = NULL;
1211 	int ret;
1212 
1213 	if (!*process_info) {
1214 		info = kzalloc(sizeof(*info), GFP_KERNEL);
1215 		if (!info)
1216 			return -ENOMEM;
1217 
1218 		mutex_init(&info->lock);
1219 		INIT_LIST_HEAD(&info->vm_list_head);
1220 		INIT_LIST_HEAD(&info->kfd_bo_list);
1221 		INIT_LIST_HEAD(&info->userptr_valid_list);
1222 		INIT_LIST_HEAD(&info->userptr_inval_list);
1223 
1224 		info->eviction_fence =
1225 			amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
1226 						   current->mm,
1227 						   NULL);
1228 		if (!info->eviction_fence) {
1229 			pr_err("Failed to create eviction fence\n");
1230 			ret = -ENOMEM;
1231 			goto create_evict_fence_fail;
1232 		}
1233 
1234 		info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
1235 		atomic_set(&info->evicted_bos, 0);
1236 		INIT_DELAYED_WORK(&info->restore_userptr_work,
1237 				  amdgpu_amdkfd_restore_userptr_worker);
1238 
1239 		*process_info = info;
1240 		*ef = dma_fence_get(&info->eviction_fence->base);
1241 	}
1242 
1243 	vm->process_info = *process_info;
1244 
1245 	/* Validate page directory and attach eviction fence */
1246 	ret = amdgpu_bo_reserve(vm->root.bo, true);
1247 	if (ret)
1248 		goto reserve_pd_fail;
1249 	ret = vm_validate_pt_pd_bos(vm);
1250 	if (ret) {
1251 		pr_err("validate_pt_pd_bos() failed\n");
1252 		goto validate_pd_fail;
1253 	}
1254 	ret = amdgpu_bo_sync_wait(vm->root.bo,
1255 				  AMDGPU_FENCE_OWNER_KFD, false);
1256 	if (ret)
1257 		goto wait_pd_fail;
1258 	ret = dma_resv_reserve_shared(vm->root.bo->tbo.base.resv, 1);
1259 	if (ret)
1260 		goto reserve_shared_fail;
1261 	amdgpu_bo_fence(vm->root.bo,
1262 			&vm->process_info->eviction_fence->base, true);
1263 	amdgpu_bo_unreserve(vm->root.bo);
1264 
1265 	/* Update process info */
1266 	mutex_lock(&vm->process_info->lock);
1267 	list_add_tail(&vm->vm_list_node,
1268 			&(vm->process_info->vm_list_head));
1269 	vm->process_info->n_vms++;
1270 	mutex_unlock(&vm->process_info->lock);
1271 
1272 	return 0;
1273 
1274 reserve_shared_fail:
1275 wait_pd_fail:
1276 validate_pd_fail:
1277 	amdgpu_bo_unreserve(vm->root.bo);
1278 reserve_pd_fail:
1279 	vm->process_info = NULL;
1280 	if (info) {
1281 		/* Two fence references: one in info and one in *ef */
1282 		dma_fence_put(&info->eviction_fence->base);
1283 		dma_fence_put(*ef);
1284 		*ef = NULL;
1285 		*process_info = NULL;
1286 		put_pid(info->pid);
1287 create_evict_fence_fail:
1288 		mutex_destroy(&info->lock);
1289 		kfree(info);
1290 	}
1291 	return ret;
1292 }
1293 
1294 /**
1295  * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
1296  * @bo: Handle of buffer object being pinned
1297  * @domain: Domain into which BO should be pinned
1298  *
1299  *   - USERPTR BOs are UNPINNABLE and will return error
1300  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1301  *     PIN count incremented. It is valid to PIN a BO multiple times
1302  *
1303  * Return: ZERO if successful in pinning, Non-Zero in case of error.
1304  */
1305 static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain)
1306 {
1307 	int ret = 0;
1308 
1309 	ret = amdgpu_bo_reserve(bo, false);
1310 	if (unlikely(ret))
1311 		return ret;
1312 
1313 	ret = amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1314 	if (ret)
1315 		pr_err("Error in Pinning BO to domain: %d\n", domain);
1316 
1317 	amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
1318 	amdgpu_bo_unreserve(bo);
1319 
1320 	return ret;
1321 }
1322 
1323 /**
1324  * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
1325  * @bo: Handle of buffer object being unpinned
1326  *
1327  *   - Is a illegal request for USERPTR BOs and is ignored
1328  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1329  *     PIN count decremented. Calls to UNPIN must balance calls to PIN
1330  */
1331 static void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo)
1332 {
1333 	int ret = 0;
1334 
1335 	ret = amdgpu_bo_reserve(bo, false);
1336 	if (unlikely(ret))
1337 		return;
1338 
1339 	amdgpu_bo_unpin(bo);
1340 	amdgpu_bo_unreserve(bo);
1341 }
1342 
1343 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
1344 					   struct file *filp, u32 pasid,
1345 					   void **process_info,
1346 					   struct dma_fence **ef)
1347 {
1348 	struct amdgpu_fpriv *drv_priv;
1349 	struct amdgpu_vm *avm;
1350 	int ret;
1351 
1352 	ret = amdgpu_file_to_fpriv(filp, &drv_priv);
1353 	if (ret)
1354 		return ret;
1355 	avm = &drv_priv->vm;
1356 
1357 	/* Already a compute VM? */
1358 	if (avm->process_info)
1359 		return -EINVAL;
1360 
1361 	/* Free the original amdgpu allocated pasid,
1362 	 * will be replaced with kfd allocated pasid.
1363 	 */
1364 	if (avm->pasid) {
1365 		amdgpu_pasid_free(avm->pasid);
1366 		amdgpu_vm_set_pasid(adev, avm, 0);
1367 	}
1368 
1369 	/* Convert VM into a compute VM */
1370 	ret = amdgpu_vm_make_compute(adev, avm);
1371 	if (ret)
1372 		return ret;
1373 
1374 	ret = amdgpu_vm_set_pasid(adev, avm, pasid);
1375 	if (ret)
1376 		return ret;
1377 	/* Initialize KFD part of the VM and process info */
1378 	ret = init_kfd_vm(avm, process_info, ef);
1379 	if (ret)
1380 		return ret;
1381 
1382 	amdgpu_vm_set_task_info(avm);
1383 
1384 	return 0;
1385 }
1386 
1387 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1388 				    struct amdgpu_vm *vm)
1389 {
1390 	struct amdkfd_process_info *process_info = vm->process_info;
1391 	struct amdgpu_bo *pd = vm->root.bo;
1392 
1393 	if (!process_info)
1394 		return;
1395 
1396 	/* Release eviction fence from PD */
1397 	amdgpu_bo_reserve(pd, false);
1398 	amdgpu_bo_fence(pd, NULL, false);
1399 	amdgpu_bo_unreserve(pd);
1400 
1401 	/* Update process info */
1402 	mutex_lock(&process_info->lock);
1403 	process_info->n_vms--;
1404 	list_del(&vm->vm_list_node);
1405 	mutex_unlock(&process_info->lock);
1406 
1407 	vm->process_info = NULL;
1408 
1409 	/* Release per-process resources when last compute VM is destroyed */
1410 	if (!process_info->n_vms) {
1411 		WARN_ON(!list_empty(&process_info->kfd_bo_list));
1412 		WARN_ON(!list_empty(&process_info->userptr_valid_list));
1413 		WARN_ON(!list_empty(&process_info->userptr_inval_list));
1414 
1415 		dma_fence_put(&process_info->eviction_fence->base);
1416 		cancel_delayed_work_sync(&process_info->restore_userptr_work);
1417 		put_pid(process_info->pid);
1418 		mutex_destroy(&process_info->lock);
1419 		kfree(process_info);
1420 	}
1421 }
1422 
1423 void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev,
1424 					    void *drm_priv)
1425 {
1426 	struct amdgpu_vm *avm;
1427 
1428 	if (WARN_ON(!adev || !drm_priv))
1429 		return;
1430 
1431 	avm = drm_priv_to_vm(drm_priv);
1432 
1433 	pr_debug("Releasing process vm %p\n", avm);
1434 
1435 	/* The original pasid of amdgpu vm has already been
1436 	 * released during making a amdgpu vm to a compute vm
1437 	 * The current pasid is managed by kfd and will be
1438 	 * released on kfd process destroy. Set amdgpu pasid
1439 	 * to 0 to avoid duplicate release.
1440 	 */
1441 	amdgpu_vm_release_compute(adev, avm);
1442 }
1443 
1444 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv)
1445 {
1446 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1447 	struct amdgpu_bo *pd = avm->root.bo;
1448 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1449 
1450 	if (adev->asic_type < CHIP_VEGA10)
1451 		return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1452 	return avm->pd_phys_addr;
1453 }
1454 
1455 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1456 		struct amdgpu_device *adev, uint64_t va, uint64_t size,
1457 		void *drm_priv, struct kgd_mem **mem,
1458 		uint64_t *offset, uint32_t flags)
1459 {
1460 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1461 	enum ttm_bo_type bo_type = ttm_bo_type_device;
1462 	struct sg_table *sg = NULL;
1463 	uint64_t user_addr = 0;
1464 	struct amdgpu_bo *bo;
1465 	struct drm_gem_object *gobj = NULL;
1466 	u32 domain, alloc_domain;
1467 	u64 alloc_flags;
1468 	int ret;
1469 
1470 	/*
1471 	 * Check on which domain to allocate BO
1472 	 */
1473 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1474 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1475 		alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1476 		alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1477 			AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0;
1478 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1479 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1480 		alloc_flags = 0;
1481 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1482 		domain = AMDGPU_GEM_DOMAIN_GTT;
1483 		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1484 		alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE;
1485 		if (!offset || !*offset)
1486 			return -EINVAL;
1487 		user_addr = untagged_addr(*offset);
1488 	} else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1489 			KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1490 		domain = AMDGPU_GEM_DOMAIN_GTT;
1491 		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1492 		bo_type = ttm_bo_type_sg;
1493 		alloc_flags = 0;
1494 		if (size > UINT_MAX)
1495 			return -EINVAL;
1496 		sg = create_doorbell_sg(*offset, size);
1497 		if (!sg)
1498 			return -ENOMEM;
1499 	} else {
1500 		return -EINVAL;
1501 	}
1502 
1503 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1504 	if (!*mem) {
1505 		ret = -ENOMEM;
1506 		goto err;
1507 	}
1508 	INIT_LIST_HEAD(&(*mem)->attachments);
1509 	mutex_init(&(*mem)->lock);
1510 	(*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1511 
1512 	/* Workaround for AQL queue wraparound bug. Map the same
1513 	 * memory twice. That means we only actually allocate half
1514 	 * the memory.
1515 	 */
1516 	if ((*mem)->aql_queue)
1517 		size = size >> 1;
1518 
1519 	(*mem)->alloc_flags = flags;
1520 
1521 	amdgpu_sync_create(&(*mem)->sync);
1522 
1523 	ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, flags);
1524 	if (ret) {
1525 		pr_debug("Insufficient memory\n");
1526 		goto err_reserve_limit;
1527 	}
1528 
1529 	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n",
1530 			va, size, domain_string(alloc_domain));
1531 
1532 	ret = amdgpu_gem_object_create(adev, size, 1, alloc_domain, alloc_flags,
1533 				       bo_type, NULL, &gobj);
1534 	if (ret) {
1535 		pr_debug("Failed to create BO on domain %s. ret %d\n",
1536 			 domain_string(alloc_domain), ret);
1537 		goto err_bo_create;
1538 	}
1539 	ret = drm_vma_node_allow(&gobj->vma_node, drm_priv);
1540 	if (ret) {
1541 		pr_debug("Failed to allow vma node access. ret %d\n", ret);
1542 		goto err_node_allow;
1543 	}
1544 	bo = gem_to_amdgpu_bo(gobj);
1545 	if (bo_type == ttm_bo_type_sg) {
1546 		bo->tbo.sg = sg;
1547 		bo->tbo.ttm->sg = sg;
1548 	}
1549 	bo->kfd_bo = *mem;
1550 	(*mem)->bo = bo;
1551 	if (user_addr)
1552 		bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO;
1553 
1554 	(*mem)->va = va;
1555 	(*mem)->domain = domain;
1556 	(*mem)->mapped_to_gpu_memory = 0;
1557 	(*mem)->process_info = avm->process_info;
1558 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1559 
1560 	if (user_addr) {
1561 		ret = init_user_pages(*mem, user_addr);
1562 		if (ret)
1563 			goto allocate_init_user_pages_failed;
1564 	} else  if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1565 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1566 		ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT);
1567 		if (ret) {
1568 			pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n");
1569 			goto err_pin_bo;
1570 		}
1571 		bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
1572 		bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
1573 	}
1574 
1575 	if (offset)
1576 		*offset = amdgpu_bo_mmap_offset(bo);
1577 
1578 	return 0;
1579 
1580 allocate_init_user_pages_failed:
1581 err_pin_bo:
1582 	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1583 	drm_vma_node_revoke(&gobj->vma_node, drm_priv);
1584 err_node_allow:
1585 	/* Don't unreserve system mem limit twice */
1586 	goto err_reserve_limit;
1587 err_bo_create:
1588 	unreserve_mem_limit(adev, size, flags);
1589 err_reserve_limit:
1590 	mutex_destroy(&(*mem)->lock);
1591 	if (gobj)
1592 		drm_gem_object_put(gobj);
1593 	else
1594 		kfree(*mem);
1595 err:
1596 	if (sg) {
1597 		sg_free_table(sg);
1598 		kfree(sg);
1599 	}
1600 	return ret;
1601 }
1602 
1603 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1604 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
1605 		uint64_t *size)
1606 {
1607 	struct amdkfd_process_info *process_info = mem->process_info;
1608 	unsigned long bo_size = mem->bo->tbo.base.size;
1609 	struct kfd_mem_attachment *entry, *tmp;
1610 	struct bo_vm_reservation_context ctx;
1611 	struct ttm_validate_buffer *bo_list_entry;
1612 	unsigned int mapped_to_gpu_memory;
1613 	int ret;
1614 	bool is_imported = false;
1615 
1616 	mutex_lock(&mem->lock);
1617 
1618 	/* Unpin MMIO/DOORBELL BO's that were pinnned during allocation */
1619 	if (mem->alloc_flags &
1620 	    (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1621 	     KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1622 		amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo);
1623 	}
1624 
1625 	mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1626 	is_imported = mem->is_imported;
1627 	mutex_unlock(&mem->lock);
1628 	/* lock is not needed after this, since mem is unused and will
1629 	 * be freed anyway
1630 	 */
1631 
1632 	if (mapped_to_gpu_memory > 0) {
1633 		pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1634 				mem->va, bo_size);
1635 		return -EBUSY;
1636 	}
1637 
1638 	/* Make sure restore workers don't access the BO any more */
1639 	bo_list_entry = &mem->validate_list;
1640 	mutex_lock(&process_info->lock);
1641 	list_del(&bo_list_entry->head);
1642 	mutex_unlock(&process_info->lock);
1643 
1644 	/* No more MMU notifiers */
1645 	amdgpu_mn_unregister(mem->bo);
1646 
1647 	ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1648 	if (unlikely(ret))
1649 		return ret;
1650 
1651 	/* The eviction fence should be removed by the last unmap.
1652 	 * TODO: Log an error condition if the bo still has the eviction fence
1653 	 * attached
1654 	 */
1655 	amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1656 					process_info->eviction_fence);
1657 	pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1658 		mem->va + bo_size * (1 + mem->aql_queue));
1659 
1660 	/* Remove from VM internal data structures */
1661 	list_for_each_entry_safe(entry, tmp, &mem->attachments, list)
1662 		kfd_mem_detach(entry);
1663 
1664 	ret = unreserve_bo_and_vms(&ctx, false, false);
1665 
1666 	/* Free the sync object */
1667 	amdgpu_sync_free(&mem->sync);
1668 
1669 	/* If the SG is not NULL, it's one we created for a doorbell or mmio
1670 	 * remap BO. We need to free it.
1671 	 */
1672 	if (mem->bo->tbo.sg) {
1673 		sg_free_table(mem->bo->tbo.sg);
1674 		kfree(mem->bo->tbo.sg);
1675 	}
1676 
1677 	/* Update the size of the BO being freed if it was allocated from
1678 	 * VRAM and is not imported.
1679 	 */
1680 	if (size) {
1681 		if ((mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM) &&
1682 		    (!is_imported))
1683 			*size = bo_size;
1684 		else
1685 			*size = 0;
1686 	}
1687 
1688 	/* Free the BO*/
1689 	drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
1690 	if (mem->dmabuf)
1691 		dma_buf_put(mem->dmabuf);
1692 	mutex_destroy(&mem->lock);
1693 
1694 	/* If this releases the last reference, it will end up calling
1695 	 * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why
1696 	 * this needs to be the last call here.
1697 	 */
1698 	drm_gem_object_put(&mem->bo->tbo.base);
1699 
1700 	return ret;
1701 }
1702 
1703 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1704 		struct amdgpu_device *adev, struct kgd_mem *mem,
1705 		void *drm_priv, bool *table_freed)
1706 {
1707 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1708 	int ret;
1709 	struct amdgpu_bo *bo;
1710 	uint32_t domain;
1711 	struct kfd_mem_attachment *entry;
1712 	struct bo_vm_reservation_context ctx;
1713 	unsigned long bo_size;
1714 	bool is_invalid_userptr = false;
1715 
1716 	bo = mem->bo;
1717 	if (!bo) {
1718 		pr_err("Invalid BO when mapping memory to GPU\n");
1719 		return -EINVAL;
1720 	}
1721 
1722 	/* Make sure restore is not running concurrently. Since we
1723 	 * don't map invalid userptr BOs, we rely on the next restore
1724 	 * worker to do the mapping
1725 	 */
1726 	mutex_lock(&mem->process_info->lock);
1727 
1728 	/* Lock mmap-sem. If we find an invalid userptr BO, we can be
1729 	 * sure that the MMU notifier is no longer running
1730 	 * concurrently and the queues are actually stopped
1731 	 */
1732 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1733 		mmap_write_lock(current->mm);
1734 		is_invalid_userptr = atomic_read(&mem->invalid);
1735 		mmap_write_unlock(current->mm);
1736 	}
1737 
1738 	mutex_lock(&mem->lock);
1739 
1740 	domain = mem->domain;
1741 	bo_size = bo->tbo.base.size;
1742 
1743 	pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
1744 			mem->va,
1745 			mem->va + bo_size * (1 + mem->aql_queue),
1746 			avm, domain_string(domain));
1747 
1748 	if (!kfd_mem_is_attached(avm, mem)) {
1749 		ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue);
1750 		if (ret)
1751 			goto out;
1752 	}
1753 
1754 	ret = reserve_bo_and_vm(mem, avm, &ctx);
1755 	if (unlikely(ret))
1756 		goto out;
1757 
1758 	/* Userptr can be marked as "not invalid", but not actually be
1759 	 * validated yet (still in the system domain). In that case
1760 	 * the queues are still stopped and we can leave mapping for
1761 	 * the next restore worker
1762 	 */
1763 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
1764 	    bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
1765 		is_invalid_userptr = true;
1766 
1767 	ret = vm_validate_pt_pd_bos(avm);
1768 	if (unlikely(ret))
1769 		goto out_unreserve;
1770 
1771 	if (mem->mapped_to_gpu_memory == 0 &&
1772 	    !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1773 		/* Validate BO only once. The eviction fence gets added to BO
1774 		 * the first time it is mapped. Validate will wait for all
1775 		 * background evictions to complete.
1776 		 */
1777 		ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
1778 		if (ret) {
1779 			pr_debug("Validate failed\n");
1780 			goto out_unreserve;
1781 		}
1782 	}
1783 
1784 	list_for_each_entry(entry, &mem->attachments, list) {
1785 		if (entry->bo_va->base.vm != avm || entry->is_mapped)
1786 			continue;
1787 
1788 		pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
1789 			 entry->va, entry->va + bo_size, entry);
1790 
1791 		ret = map_bo_to_gpuvm(mem, entry, ctx.sync,
1792 				      is_invalid_userptr, table_freed);
1793 		if (ret) {
1794 			pr_err("Failed to map bo to gpuvm\n");
1795 			goto out_unreserve;
1796 		}
1797 
1798 		ret = vm_update_pds(avm, ctx.sync);
1799 		if (ret) {
1800 			pr_err("Failed to update page directories\n");
1801 			goto out_unreserve;
1802 		}
1803 
1804 		entry->is_mapped = true;
1805 		mem->mapped_to_gpu_memory++;
1806 		pr_debug("\t INC mapping count %d\n",
1807 			 mem->mapped_to_gpu_memory);
1808 	}
1809 
1810 	if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->tbo.pin_count)
1811 		amdgpu_bo_fence(bo,
1812 				&avm->process_info->eviction_fence->base,
1813 				true);
1814 	ret = unreserve_bo_and_vms(&ctx, false, false);
1815 
1816 	/* Only apply no TLB flush on Aldebaran to
1817 	 * workaround regressions on other Asics.
1818 	 */
1819 	if (table_freed && (adev->asic_type != CHIP_ALDEBARAN))
1820 		*table_freed = true;
1821 
1822 	goto out;
1823 
1824 out_unreserve:
1825 	unreserve_bo_and_vms(&ctx, false, false);
1826 out:
1827 	mutex_unlock(&mem->process_info->lock);
1828 	mutex_unlock(&mem->lock);
1829 	return ret;
1830 }
1831 
1832 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1833 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
1834 {
1835 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1836 	struct amdkfd_process_info *process_info = avm->process_info;
1837 	unsigned long bo_size = mem->bo->tbo.base.size;
1838 	struct kfd_mem_attachment *entry;
1839 	struct bo_vm_reservation_context ctx;
1840 	int ret;
1841 
1842 	mutex_lock(&mem->lock);
1843 
1844 	ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx);
1845 	if (unlikely(ret))
1846 		goto out;
1847 	/* If no VMs were reserved, it means the BO wasn't actually mapped */
1848 	if (ctx.n_vms == 0) {
1849 		ret = -EINVAL;
1850 		goto unreserve_out;
1851 	}
1852 
1853 	ret = vm_validate_pt_pd_bos(avm);
1854 	if (unlikely(ret))
1855 		goto unreserve_out;
1856 
1857 	pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
1858 		mem->va,
1859 		mem->va + bo_size * (1 + mem->aql_queue),
1860 		avm);
1861 
1862 	list_for_each_entry(entry, &mem->attachments, list) {
1863 		if (entry->bo_va->base.vm != avm || !entry->is_mapped)
1864 			continue;
1865 
1866 		pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
1867 			 entry->va, entry->va + bo_size, entry);
1868 
1869 		unmap_bo_from_gpuvm(mem, entry, ctx.sync);
1870 		entry->is_mapped = false;
1871 
1872 		mem->mapped_to_gpu_memory--;
1873 		pr_debug("\t DEC mapping count %d\n",
1874 			 mem->mapped_to_gpu_memory);
1875 	}
1876 
1877 	/* If BO is unmapped from all VMs, unfence it. It can be evicted if
1878 	 * required.
1879 	 */
1880 	if (mem->mapped_to_gpu_memory == 0 &&
1881 	    !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) &&
1882 	    !mem->bo->tbo.pin_count)
1883 		amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1884 						process_info->eviction_fence);
1885 
1886 unreserve_out:
1887 	unreserve_bo_and_vms(&ctx, false, false);
1888 out:
1889 	mutex_unlock(&mem->lock);
1890 	return ret;
1891 }
1892 
1893 int amdgpu_amdkfd_gpuvm_sync_memory(
1894 		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr)
1895 {
1896 	struct amdgpu_sync sync;
1897 	int ret;
1898 
1899 	amdgpu_sync_create(&sync);
1900 
1901 	mutex_lock(&mem->lock);
1902 	amdgpu_sync_clone(&mem->sync, &sync);
1903 	mutex_unlock(&mem->lock);
1904 
1905 	ret = amdgpu_sync_wait(&sync, intr);
1906 	amdgpu_sync_free(&sync);
1907 	return ret;
1908 }
1909 
1910 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct amdgpu_device *adev,
1911 		struct kgd_mem *mem, void **kptr, uint64_t *size)
1912 {
1913 	int ret;
1914 	struct amdgpu_bo *bo = mem->bo;
1915 
1916 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1917 		pr_err("userptr can't be mapped to kernel\n");
1918 		return -EINVAL;
1919 	}
1920 
1921 	/* delete kgd_mem from kfd_bo_list to avoid re-validating
1922 	 * this BO in BO's restoring after eviction.
1923 	 */
1924 	mutex_lock(&mem->process_info->lock);
1925 
1926 	ret = amdgpu_bo_reserve(bo, true);
1927 	if (ret) {
1928 		pr_err("Failed to reserve bo. ret %d\n", ret);
1929 		goto bo_reserve_failed;
1930 	}
1931 
1932 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
1933 	if (ret) {
1934 		pr_err("Failed to pin bo. ret %d\n", ret);
1935 		goto pin_failed;
1936 	}
1937 
1938 	ret = amdgpu_bo_kmap(bo, kptr);
1939 	if (ret) {
1940 		pr_err("Failed to map bo to kernel. ret %d\n", ret);
1941 		goto kmap_failed;
1942 	}
1943 
1944 	amdgpu_amdkfd_remove_eviction_fence(
1945 		bo, mem->process_info->eviction_fence);
1946 	list_del_init(&mem->validate_list.head);
1947 
1948 	if (size)
1949 		*size = amdgpu_bo_size(bo);
1950 
1951 	amdgpu_bo_unreserve(bo);
1952 
1953 	mutex_unlock(&mem->process_info->lock);
1954 	return 0;
1955 
1956 kmap_failed:
1957 	amdgpu_bo_unpin(bo);
1958 pin_failed:
1959 	amdgpu_bo_unreserve(bo);
1960 bo_reserve_failed:
1961 	mutex_unlock(&mem->process_info->lock);
1962 
1963 	return ret;
1964 }
1965 
1966 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct amdgpu_device *adev,
1967 						  struct kgd_mem *mem)
1968 {
1969 	struct amdgpu_bo *bo = mem->bo;
1970 
1971 	amdgpu_bo_reserve(bo, true);
1972 	amdgpu_bo_kunmap(bo);
1973 	amdgpu_bo_unpin(bo);
1974 	amdgpu_bo_unreserve(bo);
1975 }
1976 
1977 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
1978 					  struct kfd_vm_fault_info *mem)
1979 {
1980 	if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
1981 		*mem = *adev->gmc.vm_fault_info;
1982 		mb();
1983 		atomic_set(&adev->gmc.vm_fault_info_updated, 0);
1984 	}
1985 	return 0;
1986 }
1987 
1988 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
1989 				      struct dma_buf *dma_buf,
1990 				      uint64_t va, void *drm_priv,
1991 				      struct kgd_mem **mem, uint64_t *size,
1992 				      uint64_t *mmap_offset)
1993 {
1994 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1995 	struct drm_gem_object *obj;
1996 	struct amdgpu_bo *bo;
1997 	int ret;
1998 
1999 	if (dma_buf->ops != &amdgpu_dmabuf_ops)
2000 		/* Can't handle non-graphics buffers */
2001 		return -EINVAL;
2002 
2003 	obj = dma_buf->priv;
2004 	if (drm_to_adev(obj->dev) != adev)
2005 		/* Can't handle buffers from other devices */
2006 		return -EINVAL;
2007 
2008 	bo = gem_to_amdgpu_bo(obj);
2009 	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
2010 				    AMDGPU_GEM_DOMAIN_GTT)))
2011 		/* Only VRAM and GTT BOs are supported */
2012 		return -EINVAL;
2013 
2014 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2015 	if (!*mem)
2016 		return -ENOMEM;
2017 
2018 	ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
2019 	if (ret) {
2020 		kfree(mem);
2021 		return ret;
2022 	}
2023 
2024 	if (size)
2025 		*size = amdgpu_bo_size(bo);
2026 
2027 	if (mmap_offset)
2028 		*mmap_offset = amdgpu_bo_mmap_offset(bo);
2029 
2030 	INIT_LIST_HEAD(&(*mem)->attachments);
2031 	mutex_init(&(*mem)->lock);
2032 
2033 	(*mem)->alloc_flags =
2034 		((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2035 		KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
2036 		| KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
2037 		| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
2038 
2039 	drm_gem_object_get(&bo->tbo.base);
2040 	(*mem)->bo = bo;
2041 	(*mem)->va = va;
2042 	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2043 		AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
2044 	(*mem)->mapped_to_gpu_memory = 0;
2045 	(*mem)->process_info = avm->process_info;
2046 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
2047 	amdgpu_sync_create(&(*mem)->sync);
2048 	(*mem)->is_imported = true;
2049 
2050 	return 0;
2051 }
2052 
2053 /* Evict a userptr BO by stopping the queues if necessary
2054  *
2055  * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
2056  * cannot do any memory allocations, and cannot take any locks that
2057  * are held elsewhere while allocating memory. Therefore this is as
2058  * simple as possible, using atomic counters.
2059  *
2060  * It doesn't do anything to the BO itself. The real work happens in
2061  * restore, where we get updated page addresses. This function only
2062  * ensures that GPU access to the BO is stopped.
2063  */
2064 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem,
2065 				struct mm_struct *mm)
2066 {
2067 	struct amdkfd_process_info *process_info = mem->process_info;
2068 	int evicted_bos;
2069 	int r = 0;
2070 
2071 	atomic_inc(&mem->invalid);
2072 	evicted_bos = atomic_inc_return(&process_info->evicted_bos);
2073 	if (evicted_bos == 1) {
2074 		/* First eviction, stop the queues */
2075 		r = kgd2kfd_quiesce_mm(mm);
2076 		if (r)
2077 			pr_err("Failed to quiesce KFD\n");
2078 		schedule_delayed_work(&process_info->restore_userptr_work,
2079 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2080 	}
2081 
2082 	return r;
2083 }
2084 
2085 /* Update invalid userptr BOs
2086  *
2087  * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
2088  * userptr_inval_list and updates user pages for all BOs that have
2089  * been invalidated since their last update.
2090  */
2091 static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
2092 				     struct mm_struct *mm)
2093 {
2094 	struct kgd_mem *mem, *tmp_mem;
2095 	struct amdgpu_bo *bo;
2096 	struct ttm_operation_ctx ctx = { false, false };
2097 	int invalid, ret;
2098 
2099 	/* Move all invalidated BOs to the userptr_inval_list and
2100 	 * release their user pages by migration to the CPU domain
2101 	 */
2102 	list_for_each_entry_safe(mem, tmp_mem,
2103 				 &process_info->userptr_valid_list,
2104 				 validate_list.head) {
2105 		if (!atomic_read(&mem->invalid))
2106 			continue; /* BO is still valid */
2107 
2108 		bo = mem->bo;
2109 
2110 		if (amdgpu_bo_reserve(bo, true))
2111 			return -EAGAIN;
2112 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
2113 		ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2114 		amdgpu_bo_unreserve(bo);
2115 		if (ret) {
2116 			pr_err("%s: Failed to invalidate userptr BO\n",
2117 			       __func__);
2118 			return -EAGAIN;
2119 		}
2120 
2121 		list_move_tail(&mem->validate_list.head,
2122 			       &process_info->userptr_inval_list);
2123 	}
2124 
2125 	if (list_empty(&process_info->userptr_inval_list))
2126 		return 0; /* All evicted userptr BOs were freed */
2127 
2128 	/* Go through userptr_inval_list and update any invalid user_pages */
2129 	list_for_each_entry(mem, &process_info->userptr_inval_list,
2130 			    validate_list.head) {
2131 		invalid = atomic_read(&mem->invalid);
2132 		if (!invalid)
2133 			/* BO hasn't been invalidated since the last
2134 			 * revalidation attempt. Keep its BO list.
2135 			 */
2136 			continue;
2137 
2138 		bo = mem->bo;
2139 
2140 		/* Get updated user pages */
2141 		ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
2142 		if (ret) {
2143 			pr_debug("Failed %d to get user pages\n", ret);
2144 
2145 			/* Return -EFAULT bad address error as success. It will
2146 			 * fail later with a VM fault if the GPU tries to access
2147 			 * it. Better than hanging indefinitely with stalled
2148 			 * user mode queues.
2149 			 *
2150 			 * Return other error -EBUSY or -ENOMEM to retry restore
2151 			 */
2152 			if (ret != -EFAULT)
2153 				return ret;
2154 		} else {
2155 
2156 			/*
2157 			 * FIXME: Cannot ignore the return code, must hold
2158 			 * notifier_lock
2159 			 */
2160 			amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
2161 		}
2162 
2163 		/* Mark the BO as valid unless it was invalidated
2164 		 * again concurrently.
2165 		 */
2166 		if (atomic_cmpxchg(&mem->invalid, invalid, 0) != invalid)
2167 			return -EAGAIN;
2168 	}
2169 
2170 	return 0;
2171 }
2172 
2173 /* Validate invalid userptr BOs
2174  *
2175  * Validates BOs on the userptr_inval_list, and moves them back to the
2176  * userptr_valid_list. Also updates GPUVM page tables with new page
2177  * addresses and waits for the page table updates to complete.
2178  */
2179 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
2180 {
2181 	struct amdgpu_bo_list_entry *pd_bo_list_entries;
2182 	struct list_head resv_list, duplicates;
2183 	struct ww_acquire_ctx ticket;
2184 	struct amdgpu_sync sync;
2185 
2186 	struct amdgpu_vm *peer_vm;
2187 	struct kgd_mem *mem, *tmp_mem;
2188 	struct amdgpu_bo *bo;
2189 	struct ttm_operation_ctx ctx = { false, false };
2190 	int i, ret;
2191 
2192 	pd_bo_list_entries = kcalloc(process_info->n_vms,
2193 				     sizeof(struct amdgpu_bo_list_entry),
2194 				     GFP_KERNEL);
2195 	if (!pd_bo_list_entries) {
2196 		pr_err("%s: Failed to allocate PD BO list entries\n", __func__);
2197 		ret = -ENOMEM;
2198 		goto out_no_mem;
2199 	}
2200 
2201 	INIT_LIST_HEAD(&resv_list);
2202 	INIT_LIST_HEAD(&duplicates);
2203 
2204 	/* Get all the page directory BOs that need to be reserved */
2205 	i = 0;
2206 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2207 			    vm_list_node)
2208 		amdgpu_vm_get_pd_bo(peer_vm, &resv_list,
2209 				    &pd_bo_list_entries[i++]);
2210 	/* Add the userptr_inval_list entries to resv_list */
2211 	list_for_each_entry(mem, &process_info->userptr_inval_list,
2212 			    validate_list.head) {
2213 		list_add_tail(&mem->resv_list.head, &resv_list);
2214 		mem->resv_list.bo = mem->validate_list.bo;
2215 		mem->resv_list.num_shared = mem->validate_list.num_shared;
2216 	}
2217 
2218 	/* Reserve all BOs and page tables for validation */
2219 	ret = ttm_eu_reserve_buffers(&ticket, &resv_list, false, &duplicates);
2220 	WARN(!list_empty(&duplicates), "Duplicates should be empty");
2221 	if (ret)
2222 		goto out_free;
2223 
2224 	amdgpu_sync_create(&sync);
2225 
2226 	ret = process_validate_vms(process_info);
2227 	if (ret)
2228 		goto unreserve_out;
2229 
2230 	/* Validate BOs and update GPUVM page tables */
2231 	list_for_each_entry_safe(mem, tmp_mem,
2232 				 &process_info->userptr_inval_list,
2233 				 validate_list.head) {
2234 		struct kfd_mem_attachment *attachment;
2235 
2236 		bo = mem->bo;
2237 
2238 		/* Validate the BO if we got user pages */
2239 		if (bo->tbo.ttm->pages[0]) {
2240 			amdgpu_bo_placement_from_domain(bo, mem->domain);
2241 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2242 			if (ret) {
2243 				pr_err("%s: failed to validate BO\n", __func__);
2244 				goto unreserve_out;
2245 			}
2246 		}
2247 
2248 		list_move_tail(&mem->validate_list.head,
2249 			       &process_info->userptr_valid_list);
2250 
2251 		/* Update mapping. If the BO was not validated
2252 		 * (because we couldn't get user pages), this will
2253 		 * clear the page table entries, which will result in
2254 		 * VM faults if the GPU tries to access the invalid
2255 		 * memory.
2256 		 */
2257 		list_for_each_entry(attachment, &mem->attachments, list) {
2258 			if (!attachment->is_mapped)
2259 				continue;
2260 
2261 			kfd_mem_dmaunmap_attachment(mem, attachment);
2262 			ret = update_gpuvm_pte(mem, attachment, &sync, NULL);
2263 			if (ret) {
2264 				pr_err("%s: update PTE failed\n", __func__);
2265 				/* make sure this gets validated again */
2266 				atomic_inc(&mem->invalid);
2267 				goto unreserve_out;
2268 			}
2269 		}
2270 	}
2271 
2272 	/* Update page directories */
2273 	ret = process_update_pds(process_info, &sync);
2274 
2275 unreserve_out:
2276 	ttm_eu_backoff_reservation(&ticket, &resv_list);
2277 	amdgpu_sync_wait(&sync, false);
2278 	amdgpu_sync_free(&sync);
2279 out_free:
2280 	kfree(pd_bo_list_entries);
2281 out_no_mem:
2282 
2283 	return ret;
2284 }
2285 
2286 /* Worker callback to restore evicted userptr BOs
2287  *
2288  * Tries to update and validate all userptr BOs. If successful and no
2289  * concurrent evictions happened, the queues are restarted. Otherwise,
2290  * reschedule for another attempt later.
2291  */
2292 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
2293 {
2294 	struct delayed_work *dwork = to_delayed_work(work);
2295 	struct amdkfd_process_info *process_info =
2296 		container_of(dwork, struct amdkfd_process_info,
2297 			     restore_userptr_work);
2298 	struct task_struct *usertask;
2299 	struct mm_struct *mm;
2300 	int evicted_bos;
2301 
2302 	evicted_bos = atomic_read(&process_info->evicted_bos);
2303 	if (!evicted_bos)
2304 		return;
2305 
2306 	/* Reference task and mm in case of concurrent process termination */
2307 	usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
2308 	if (!usertask)
2309 		return;
2310 	mm = get_task_mm(usertask);
2311 	if (!mm) {
2312 		put_task_struct(usertask);
2313 		return;
2314 	}
2315 
2316 	mutex_lock(&process_info->lock);
2317 
2318 	if (update_invalid_user_pages(process_info, mm))
2319 		goto unlock_out;
2320 	/* userptr_inval_list can be empty if all evicted userptr BOs
2321 	 * have been freed. In that case there is nothing to validate
2322 	 * and we can just restart the queues.
2323 	 */
2324 	if (!list_empty(&process_info->userptr_inval_list)) {
2325 		if (atomic_read(&process_info->evicted_bos) != evicted_bos)
2326 			goto unlock_out; /* Concurrent eviction, try again */
2327 
2328 		if (validate_invalid_user_pages(process_info))
2329 			goto unlock_out;
2330 	}
2331 	/* Final check for concurrent evicton and atomic update. If
2332 	 * another eviction happens after successful update, it will
2333 	 * be a first eviction that calls quiesce_mm. The eviction
2334 	 * reference counting inside KFD will handle this case.
2335 	 */
2336 	if (atomic_cmpxchg(&process_info->evicted_bos, evicted_bos, 0) !=
2337 	    evicted_bos)
2338 		goto unlock_out;
2339 	evicted_bos = 0;
2340 	if (kgd2kfd_resume_mm(mm)) {
2341 		pr_err("%s: Failed to resume KFD\n", __func__);
2342 		/* No recovery from this failure. Probably the CP is
2343 		 * hanging. No point trying again.
2344 		 */
2345 	}
2346 
2347 unlock_out:
2348 	mutex_unlock(&process_info->lock);
2349 	mmput(mm);
2350 	put_task_struct(usertask);
2351 
2352 	/* If validation failed, reschedule another attempt */
2353 	if (evicted_bos)
2354 		schedule_delayed_work(&process_info->restore_userptr_work,
2355 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2356 }
2357 
2358 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2359  *   KFD process identified by process_info
2360  *
2361  * @process_info: amdkfd_process_info of the KFD process
2362  *
2363  * After memory eviction, restore thread calls this function. The function
2364  * should be called when the Process is still valid. BO restore involves -
2365  *
2366  * 1.  Release old eviction fence and create new one
2367  * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2368  * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2369  *     BOs that need to be reserved.
2370  * 4.  Reserve all the BOs
2371  * 5.  Validate of PD and PT BOs.
2372  * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2373  * 7.  Add fence to all PD and PT BOs.
2374  * 8.  Unreserve all BOs
2375  */
2376 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
2377 {
2378 	struct amdgpu_bo_list_entry *pd_bo_list;
2379 	struct amdkfd_process_info *process_info = info;
2380 	struct amdgpu_vm *peer_vm;
2381 	struct kgd_mem *mem;
2382 	struct bo_vm_reservation_context ctx;
2383 	struct amdgpu_amdkfd_fence *new_fence;
2384 	int ret = 0, i;
2385 	struct list_head duplicate_save;
2386 	struct amdgpu_sync sync_obj;
2387 	unsigned long failed_size = 0;
2388 	unsigned long total_size = 0;
2389 
2390 	INIT_LIST_HEAD(&duplicate_save);
2391 	INIT_LIST_HEAD(&ctx.list);
2392 	INIT_LIST_HEAD(&ctx.duplicates);
2393 
2394 	pd_bo_list = kcalloc(process_info->n_vms,
2395 			     sizeof(struct amdgpu_bo_list_entry),
2396 			     GFP_KERNEL);
2397 	if (!pd_bo_list)
2398 		return -ENOMEM;
2399 
2400 	i = 0;
2401 	mutex_lock(&process_info->lock);
2402 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2403 			vm_list_node)
2404 		amdgpu_vm_get_pd_bo(peer_vm, &ctx.list, &pd_bo_list[i++]);
2405 
2406 	/* Reserve all BOs and page tables/directory. Add all BOs from
2407 	 * kfd_bo_list to ctx.list
2408 	 */
2409 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2410 			    validate_list.head) {
2411 
2412 		list_add_tail(&mem->resv_list.head, &ctx.list);
2413 		mem->resv_list.bo = mem->validate_list.bo;
2414 		mem->resv_list.num_shared = mem->validate_list.num_shared;
2415 	}
2416 
2417 	ret = ttm_eu_reserve_buffers(&ctx.ticket, &ctx.list,
2418 				     false, &duplicate_save);
2419 	if (ret) {
2420 		pr_debug("Memory eviction: TTM Reserve Failed. Try again\n");
2421 		goto ttm_reserve_fail;
2422 	}
2423 
2424 	amdgpu_sync_create(&sync_obj);
2425 
2426 	/* Validate PDs and PTs */
2427 	ret = process_validate_vms(process_info);
2428 	if (ret)
2429 		goto validate_map_fail;
2430 
2431 	ret = process_sync_pds_resv(process_info, &sync_obj);
2432 	if (ret) {
2433 		pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2434 		goto validate_map_fail;
2435 	}
2436 
2437 	/* Validate BOs and map them to GPUVM (update VM page tables). */
2438 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2439 			    validate_list.head) {
2440 
2441 		struct amdgpu_bo *bo = mem->bo;
2442 		uint32_t domain = mem->domain;
2443 		struct kfd_mem_attachment *attachment;
2444 
2445 		total_size += amdgpu_bo_size(bo);
2446 
2447 		ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2448 		if (ret) {
2449 			pr_debug("Memory eviction: Validate BOs failed\n");
2450 			failed_size += amdgpu_bo_size(bo);
2451 			ret = amdgpu_amdkfd_bo_validate(bo,
2452 						AMDGPU_GEM_DOMAIN_GTT, false);
2453 			if (ret) {
2454 				pr_debug("Memory eviction: Try again\n");
2455 				goto validate_map_fail;
2456 			}
2457 		}
2458 		ret = amdgpu_sync_fence(&sync_obj, bo->tbo.moving);
2459 		if (ret) {
2460 			pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2461 			goto validate_map_fail;
2462 		}
2463 		list_for_each_entry(attachment, &mem->attachments, list) {
2464 			if (!attachment->is_mapped)
2465 				continue;
2466 
2467 			kfd_mem_dmaunmap_attachment(mem, attachment);
2468 			ret = update_gpuvm_pte(mem, attachment, &sync_obj, NULL);
2469 			if (ret) {
2470 				pr_debug("Memory eviction: update PTE failed. Try again\n");
2471 				goto validate_map_fail;
2472 			}
2473 		}
2474 	}
2475 
2476 	if (failed_size)
2477 		pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2478 
2479 	/* Update page directories */
2480 	ret = process_update_pds(process_info, &sync_obj);
2481 	if (ret) {
2482 		pr_debug("Memory eviction: update PDs failed. Try again\n");
2483 		goto validate_map_fail;
2484 	}
2485 
2486 	/* Wait for validate and PT updates to finish */
2487 	amdgpu_sync_wait(&sync_obj, false);
2488 
2489 	/* Release old eviction fence and create new one, because fence only
2490 	 * goes from unsignaled to signaled, fence cannot be reused.
2491 	 * Use context and mm from the old fence.
2492 	 */
2493 	new_fence = amdgpu_amdkfd_fence_create(
2494 				process_info->eviction_fence->base.context,
2495 				process_info->eviction_fence->mm,
2496 				NULL);
2497 	if (!new_fence) {
2498 		pr_err("Failed to create eviction fence\n");
2499 		ret = -ENOMEM;
2500 		goto validate_map_fail;
2501 	}
2502 	dma_fence_put(&process_info->eviction_fence->base);
2503 	process_info->eviction_fence = new_fence;
2504 	*ef = dma_fence_get(&new_fence->base);
2505 
2506 	/* Attach new eviction fence to all BOs */
2507 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2508 		validate_list.head)
2509 		amdgpu_bo_fence(mem->bo,
2510 			&process_info->eviction_fence->base, true);
2511 
2512 	/* Attach eviction fence to PD / PT BOs */
2513 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2514 			    vm_list_node) {
2515 		struct amdgpu_bo *bo = peer_vm->root.bo;
2516 
2517 		amdgpu_bo_fence(bo, &process_info->eviction_fence->base, true);
2518 	}
2519 
2520 validate_map_fail:
2521 	ttm_eu_backoff_reservation(&ctx.ticket, &ctx.list);
2522 	amdgpu_sync_free(&sync_obj);
2523 ttm_reserve_fail:
2524 	mutex_unlock(&process_info->lock);
2525 	kfree(pd_bo_list);
2526 	return ret;
2527 }
2528 
2529 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
2530 {
2531 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2532 	struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
2533 	int ret;
2534 
2535 	if (!info || !gws)
2536 		return -EINVAL;
2537 
2538 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2539 	if (!*mem)
2540 		return -ENOMEM;
2541 
2542 	mutex_init(&(*mem)->lock);
2543 	INIT_LIST_HEAD(&(*mem)->attachments);
2544 	(*mem)->bo = amdgpu_bo_ref(gws_bo);
2545 	(*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
2546 	(*mem)->process_info = process_info;
2547 	add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
2548 	amdgpu_sync_create(&(*mem)->sync);
2549 
2550 
2551 	/* Validate gws bo the first time it is added to process */
2552 	mutex_lock(&(*mem)->process_info->lock);
2553 	ret = amdgpu_bo_reserve(gws_bo, false);
2554 	if (unlikely(ret)) {
2555 		pr_err("Reserve gws bo failed %d\n", ret);
2556 		goto bo_reservation_failure;
2557 	}
2558 
2559 	ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
2560 	if (ret) {
2561 		pr_err("GWS BO validate failed %d\n", ret);
2562 		goto bo_validation_failure;
2563 	}
2564 	/* GWS resource is shared b/t amdgpu and amdkfd
2565 	 * Add process eviction fence to bo so they can
2566 	 * evict each other.
2567 	 */
2568 	ret = dma_resv_reserve_shared(gws_bo->tbo.base.resv, 1);
2569 	if (ret)
2570 		goto reserve_shared_fail;
2571 	amdgpu_bo_fence(gws_bo, &process_info->eviction_fence->base, true);
2572 	amdgpu_bo_unreserve(gws_bo);
2573 	mutex_unlock(&(*mem)->process_info->lock);
2574 
2575 	return ret;
2576 
2577 reserve_shared_fail:
2578 bo_validation_failure:
2579 	amdgpu_bo_unreserve(gws_bo);
2580 bo_reservation_failure:
2581 	mutex_unlock(&(*mem)->process_info->lock);
2582 	amdgpu_sync_free(&(*mem)->sync);
2583 	remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
2584 	amdgpu_bo_unref(&gws_bo);
2585 	mutex_destroy(&(*mem)->lock);
2586 	kfree(*mem);
2587 	*mem = NULL;
2588 	return ret;
2589 }
2590 
2591 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
2592 {
2593 	int ret;
2594 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2595 	struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
2596 	struct amdgpu_bo *gws_bo = kgd_mem->bo;
2597 
2598 	/* Remove BO from process's validate list so restore worker won't touch
2599 	 * it anymore
2600 	 */
2601 	remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
2602 
2603 	ret = amdgpu_bo_reserve(gws_bo, false);
2604 	if (unlikely(ret)) {
2605 		pr_err("Reserve gws bo failed %d\n", ret);
2606 		//TODO add BO back to validate_list?
2607 		return ret;
2608 	}
2609 	amdgpu_amdkfd_remove_eviction_fence(gws_bo,
2610 			process_info->eviction_fence);
2611 	amdgpu_bo_unreserve(gws_bo);
2612 	amdgpu_sync_free(&kgd_mem->sync);
2613 	amdgpu_bo_unref(&gws_bo);
2614 	mutex_destroy(&kgd_mem->lock);
2615 	kfree(mem);
2616 	return 0;
2617 }
2618 
2619 /* Returns GPU-specific tiling mode information */
2620 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
2621 				struct tile_config *config)
2622 {
2623 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
2624 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
2625 	config->num_tile_configs =
2626 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
2627 	config->macro_tile_config_ptr =
2628 			adev->gfx.config.macrotile_mode_array;
2629 	config->num_macro_tile_configs =
2630 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
2631 
2632 	/* Those values are not set from GFX9 onwards */
2633 	config->num_banks = adev->gfx.config.num_banks;
2634 	config->num_ranks = adev->gfx.config.num_ranks;
2635 
2636 	return 0;
2637 }
2638