xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c (revision db5b5c679e6cad2bb147337af6c378d278231b45)
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 	/* Import takes an extra reference on the dmabuf. Drop it now to
667 	 * avoid leaking it. We only need the one reference in
668 	 * kgd_mem->dmabuf.
669 	 */
670 	dma_buf_put(mem->dmabuf);
671 
672 	*bo = gem_to_amdgpu_bo(gobj);
673 	(*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE;
674 	(*bo)->parent = amdgpu_bo_ref(mem->bo);
675 
676 	return 0;
677 }
678 
679 /* kfd_mem_attach - Add a BO to a VM
680  *
681  * Everything that needs to bo done only once when a BO is first added
682  * to a VM. It can later be mapped and unmapped many times without
683  * repeating these steps.
684  *
685  * 0. Create BO for DMA mapping, if needed
686  * 1. Allocate and initialize BO VA entry data structure
687  * 2. Add BO to the VM
688  * 3. Determine ASIC-specific PTE flags
689  * 4. Alloc page tables and directories if needed
690  * 4a.  Validate new page tables and directories
691  */
692 static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
693 		struct amdgpu_vm *vm, bool is_aql)
694 {
695 	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
696 	unsigned long bo_size = mem->bo->tbo.base.size;
697 	uint64_t va = mem->va;
698 	struct kfd_mem_attachment *attachment[2] = {NULL, NULL};
699 	struct amdgpu_bo *bo[2] = {NULL, NULL};
700 	int i, ret;
701 
702 	if (!va) {
703 		pr_err("Invalid VA when adding BO to VM\n");
704 		return -EINVAL;
705 	}
706 
707 	for (i = 0; i <= is_aql; i++) {
708 		attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL);
709 		if (unlikely(!attachment[i])) {
710 			ret = -ENOMEM;
711 			goto unwind;
712 		}
713 
714 		pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
715 			 va + bo_size, vm);
716 
717 		if (adev == bo_adev || (mem->domain == AMDGPU_GEM_DOMAIN_VRAM &&
718 					amdgpu_xgmi_same_hive(adev, bo_adev))) {
719 			/* Mappings on the local GPU and VRAM mappings in the
720 			 * local hive share the original BO
721 			 */
722 			attachment[i]->type = KFD_MEM_ATT_SHARED;
723 			bo[i] = mem->bo;
724 			drm_gem_object_get(&bo[i]->tbo.base);
725 		} else if (i > 0) {
726 			/* Multiple mappings on the same GPU share the BO */
727 			attachment[i]->type = KFD_MEM_ATT_SHARED;
728 			bo[i] = bo[0];
729 			drm_gem_object_get(&bo[i]->tbo.base);
730 		} else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
731 			/* Create an SG BO to DMA-map userptrs on other GPUs */
732 			attachment[i]->type = KFD_MEM_ATT_USERPTR;
733 			ret = kfd_mem_attach_userptr(adev, mem, &bo[i]);
734 			if (ret)
735 				goto unwind;
736 		} else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT &&
737 			   mem->bo->tbo.type != ttm_bo_type_sg) {
738 			/* GTT BOs use DMA-mapping ability of dynamic-attach
739 			 * DMA bufs. TODO: The same should work for VRAM on
740 			 * large-BAR GPUs.
741 			 */
742 			attachment[i]->type = KFD_MEM_ATT_DMABUF;
743 			ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]);
744 			if (ret)
745 				goto unwind;
746 		} else {
747 			/* FIXME: Need to DMA-map other BO types:
748 			 * large-BAR VRAM, doorbells, MMIO remap
749 			 */
750 			attachment[i]->type = KFD_MEM_ATT_SHARED;
751 			bo[i] = mem->bo;
752 			drm_gem_object_get(&bo[i]->tbo.base);
753 		}
754 
755 		/* Add BO to VM internal data structures */
756 		ret = amdgpu_bo_reserve(bo[i], false);
757 		if (ret) {
758 			pr_debug("Unable to reserve BO during memory attach");
759 			goto unwind;
760 		}
761 		attachment[i]->bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]);
762 		amdgpu_bo_unreserve(bo[i]);
763 		if (unlikely(!attachment[i]->bo_va)) {
764 			ret = -ENOMEM;
765 			pr_err("Failed to add BO object to VM. ret == %d\n",
766 			       ret);
767 			goto unwind;
768 		}
769 		attachment[i]->va = va;
770 		attachment[i]->pte_flags = get_pte_flags(adev, mem);
771 		attachment[i]->adev = adev;
772 		list_add(&attachment[i]->list, &mem->attachments);
773 
774 		va += bo_size;
775 	}
776 
777 	return 0;
778 
779 unwind:
780 	for (; i >= 0; i--) {
781 		if (!attachment[i])
782 			continue;
783 		if (attachment[i]->bo_va) {
784 			amdgpu_bo_reserve(bo[i], true);
785 			amdgpu_vm_bo_rmv(adev, attachment[i]->bo_va);
786 			amdgpu_bo_unreserve(bo[i]);
787 			list_del(&attachment[i]->list);
788 		}
789 		if (bo[i])
790 			drm_gem_object_put(&bo[i]->tbo.base);
791 		kfree(attachment[i]);
792 	}
793 	return ret;
794 }
795 
796 static void kfd_mem_detach(struct kfd_mem_attachment *attachment)
797 {
798 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
799 
800 	pr_debug("\t remove VA 0x%llx in entry %p\n",
801 			attachment->va, attachment);
802 	amdgpu_vm_bo_rmv(attachment->adev, attachment->bo_va);
803 	drm_gem_object_put(&bo->tbo.base);
804 	list_del(&attachment->list);
805 	kfree(attachment);
806 }
807 
808 static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
809 				struct amdkfd_process_info *process_info,
810 				bool userptr)
811 {
812 	struct ttm_validate_buffer *entry = &mem->validate_list;
813 	struct amdgpu_bo *bo = mem->bo;
814 
815 	INIT_LIST_HEAD(&entry->head);
816 	entry->num_shared = 1;
817 	entry->bo = &bo->tbo;
818 	mutex_lock(&process_info->lock);
819 	if (userptr)
820 		list_add_tail(&entry->head, &process_info->userptr_valid_list);
821 	else
822 		list_add_tail(&entry->head, &process_info->kfd_bo_list);
823 	mutex_unlock(&process_info->lock);
824 }
825 
826 static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
827 		struct amdkfd_process_info *process_info)
828 {
829 	struct ttm_validate_buffer *bo_list_entry;
830 
831 	bo_list_entry = &mem->validate_list;
832 	mutex_lock(&process_info->lock);
833 	list_del(&bo_list_entry->head);
834 	mutex_unlock(&process_info->lock);
835 }
836 
837 /* Initializes user pages. It registers the MMU notifier and validates
838  * the userptr BO in the GTT domain.
839  *
840  * The BO must already be on the userptr_valid_list. Otherwise an
841  * eviction and restore may happen that leaves the new BO unmapped
842  * with the user mode queues running.
843  *
844  * Takes the process_info->lock to protect against concurrent restore
845  * workers.
846  *
847  * Returns 0 for success, negative errno for errors.
848  */
849 static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr)
850 {
851 	struct amdkfd_process_info *process_info = mem->process_info;
852 	struct amdgpu_bo *bo = mem->bo;
853 	struct ttm_operation_ctx ctx = { true, false };
854 	int ret = 0;
855 
856 	mutex_lock(&process_info->lock);
857 
858 	ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
859 	if (ret) {
860 		pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
861 		goto out;
862 	}
863 
864 	ret = amdgpu_mn_register(bo, user_addr);
865 	if (ret) {
866 		pr_err("%s: Failed to register MMU notifier: %d\n",
867 		       __func__, ret);
868 		goto out;
869 	}
870 
871 	ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
872 	if (ret) {
873 		pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
874 		goto unregister_out;
875 	}
876 
877 	ret = amdgpu_bo_reserve(bo, true);
878 	if (ret) {
879 		pr_err("%s: Failed to reserve BO\n", __func__);
880 		goto release_out;
881 	}
882 	amdgpu_bo_placement_from_domain(bo, mem->domain);
883 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
884 	if (ret)
885 		pr_err("%s: failed to validate BO\n", __func__);
886 	amdgpu_bo_unreserve(bo);
887 
888 release_out:
889 	amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
890 unregister_out:
891 	if (ret)
892 		amdgpu_mn_unregister(bo);
893 out:
894 	mutex_unlock(&process_info->lock);
895 	return ret;
896 }
897 
898 /* Reserving a BO and its page table BOs must happen atomically to
899  * avoid deadlocks. Some operations update multiple VMs at once. Track
900  * all the reservation info in a context structure. Optionally a sync
901  * object can track VM updates.
902  */
903 struct bo_vm_reservation_context {
904 	struct amdgpu_bo_list_entry kfd_bo; /* BO list entry for the KFD BO */
905 	unsigned int n_vms;		    /* Number of VMs reserved	    */
906 	struct amdgpu_bo_list_entry *vm_pd; /* Array of VM BO list entries  */
907 	struct ww_acquire_ctx ticket;	    /* Reservation ticket	    */
908 	struct list_head list, duplicates;  /* BO lists			    */
909 	struct amdgpu_sync *sync;	    /* Pointer to sync object	    */
910 	bool reserved;			    /* Whether BOs are reserved	    */
911 };
912 
913 enum bo_vm_match {
914 	BO_VM_NOT_MAPPED = 0,	/* Match VMs where a BO is not mapped */
915 	BO_VM_MAPPED,		/* Match VMs where a BO is mapped     */
916 	BO_VM_ALL,		/* Match all VMs a BO was added to    */
917 };
918 
919 /**
920  * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
921  * @mem: KFD BO structure.
922  * @vm: the VM to reserve.
923  * @ctx: the struct that will be used in unreserve_bo_and_vms().
924  */
925 static int reserve_bo_and_vm(struct kgd_mem *mem,
926 			      struct amdgpu_vm *vm,
927 			      struct bo_vm_reservation_context *ctx)
928 {
929 	struct amdgpu_bo *bo = mem->bo;
930 	int ret;
931 
932 	WARN_ON(!vm);
933 
934 	ctx->reserved = false;
935 	ctx->n_vms = 1;
936 	ctx->sync = &mem->sync;
937 
938 	INIT_LIST_HEAD(&ctx->list);
939 	INIT_LIST_HEAD(&ctx->duplicates);
940 
941 	ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), GFP_KERNEL);
942 	if (!ctx->vm_pd)
943 		return -ENOMEM;
944 
945 	ctx->kfd_bo.priority = 0;
946 	ctx->kfd_bo.tv.bo = &bo->tbo;
947 	ctx->kfd_bo.tv.num_shared = 1;
948 	list_add(&ctx->kfd_bo.tv.head, &ctx->list);
949 
950 	amdgpu_vm_get_pd_bo(vm, &ctx->list, &ctx->vm_pd[0]);
951 
952 	ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
953 				     false, &ctx->duplicates);
954 	if (ret) {
955 		pr_err("Failed to reserve buffers in ttm.\n");
956 		kfree(ctx->vm_pd);
957 		ctx->vm_pd = NULL;
958 		return ret;
959 	}
960 
961 	ctx->reserved = true;
962 	return 0;
963 }
964 
965 /**
966  * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
967  * @mem: KFD BO structure.
968  * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
969  * is used. Otherwise, a single VM associated with the BO.
970  * @map_type: the mapping status that will be used to filter the VMs.
971  * @ctx: the struct that will be used in unreserve_bo_and_vms().
972  *
973  * Returns 0 for success, negative for failure.
974  */
975 static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
976 				struct amdgpu_vm *vm, enum bo_vm_match map_type,
977 				struct bo_vm_reservation_context *ctx)
978 {
979 	struct amdgpu_bo *bo = mem->bo;
980 	struct kfd_mem_attachment *entry;
981 	unsigned int i;
982 	int ret;
983 
984 	ctx->reserved = false;
985 	ctx->n_vms = 0;
986 	ctx->vm_pd = NULL;
987 	ctx->sync = &mem->sync;
988 
989 	INIT_LIST_HEAD(&ctx->list);
990 	INIT_LIST_HEAD(&ctx->duplicates);
991 
992 	list_for_each_entry(entry, &mem->attachments, list) {
993 		if ((vm && vm != entry->bo_va->base.vm) ||
994 			(entry->is_mapped != map_type
995 			&& map_type != BO_VM_ALL))
996 			continue;
997 
998 		ctx->n_vms++;
999 	}
1000 
1001 	if (ctx->n_vms != 0) {
1002 		ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd),
1003 				     GFP_KERNEL);
1004 		if (!ctx->vm_pd)
1005 			return -ENOMEM;
1006 	}
1007 
1008 	ctx->kfd_bo.priority = 0;
1009 	ctx->kfd_bo.tv.bo = &bo->tbo;
1010 	ctx->kfd_bo.tv.num_shared = 1;
1011 	list_add(&ctx->kfd_bo.tv.head, &ctx->list);
1012 
1013 	i = 0;
1014 	list_for_each_entry(entry, &mem->attachments, list) {
1015 		if ((vm && vm != entry->bo_va->base.vm) ||
1016 			(entry->is_mapped != map_type
1017 			&& map_type != BO_VM_ALL))
1018 			continue;
1019 
1020 		amdgpu_vm_get_pd_bo(entry->bo_va->base.vm, &ctx->list,
1021 				&ctx->vm_pd[i]);
1022 		i++;
1023 	}
1024 
1025 	ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
1026 				     false, &ctx->duplicates);
1027 	if (ret) {
1028 		pr_err("Failed to reserve buffers in ttm.\n");
1029 		kfree(ctx->vm_pd);
1030 		ctx->vm_pd = NULL;
1031 		return ret;
1032 	}
1033 
1034 	ctx->reserved = true;
1035 	return 0;
1036 }
1037 
1038 /**
1039  * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1040  * @ctx: Reservation context to unreserve
1041  * @wait: Optionally wait for a sync object representing pending VM updates
1042  * @intr: Whether the wait is interruptible
1043  *
1044  * Also frees any resources allocated in
1045  * reserve_bo_and_(cond_)vm(s). Returns the status from
1046  * amdgpu_sync_wait.
1047  */
1048 static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
1049 				 bool wait, bool intr)
1050 {
1051 	int ret = 0;
1052 
1053 	if (wait)
1054 		ret = amdgpu_sync_wait(ctx->sync, intr);
1055 
1056 	if (ctx->reserved)
1057 		ttm_eu_backoff_reservation(&ctx->ticket, &ctx->list);
1058 	kfree(ctx->vm_pd);
1059 
1060 	ctx->sync = NULL;
1061 
1062 	ctx->reserved = false;
1063 	ctx->vm_pd = NULL;
1064 
1065 	return ret;
1066 }
1067 
1068 static void unmap_bo_from_gpuvm(struct kgd_mem *mem,
1069 				struct kfd_mem_attachment *entry,
1070 				struct amdgpu_sync *sync)
1071 {
1072 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1073 	struct amdgpu_device *adev = entry->adev;
1074 	struct amdgpu_vm *vm = bo_va->base.vm;
1075 
1076 	amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
1077 
1078 	amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
1079 
1080 	amdgpu_sync_fence(sync, bo_va->last_pt_update);
1081 
1082 	kfd_mem_dmaunmap_attachment(mem, entry);
1083 }
1084 
1085 static int update_gpuvm_pte(struct kgd_mem *mem,
1086 			    struct kfd_mem_attachment *entry,
1087 			    struct amdgpu_sync *sync,
1088 			    bool *table_freed)
1089 {
1090 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1091 	struct amdgpu_device *adev = entry->adev;
1092 	int ret;
1093 
1094 	ret = kfd_mem_dmamap_attachment(mem, entry);
1095 	if (ret)
1096 		return ret;
1097 
1098 	/* Update the page tables  */
1099 	ret = amdgpu_vm_bo_update(adev, bo_va, false, table_freed);
1100 	if (ret) {
1101 		pr_err("amdgpu_vm_bo_update failed\n");
1102 		return ret;
1103 	}
1104 
1105 	return amdgpu_sync_fence(sync, bo_va->last_pt_update);
1106 }
1107 
1108 static int map_bo_to_gpuvm(struct kgd_mem *mem,
1109 			   struct kfd_mem_attachment *entry,
1110 			   struct amdgpu_sync *sync,
1111 			   bool no_update_pte,
1112 			   bool *table_freed)
1113 {
1114 	int ret;
1115 
1116 	/* Set virtual address for the allocation */
1117 	ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0,
1118 			       amdgpu_bo_size(entry->bo_va->base.bo),
1119 			       entry->pte_flags);
1120 	if (ret) {
1121 		pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
1122 				entry->va, ret);
1123 		return ret;
1124 	}
1125 
1126 	if (no_update_pte)
1127 		return 0;
1128 
1129 	ret = update_gpuvm_pte(mem, entry, sync, table_freed);
1130 	if (ret) {
1131 		pr_err("update_gpuvm_pte() failed\n");
1132 		goto update_gpuvm_pte_failed;
1133 	}
1134 
1135 	return 0;
1136 
1137 update_gpuvm_pte_failed:
1138 	unmap_bo_from_gpuvm(mem, entry, sync);
1139 	return ret;
1140 }
1141 
1142 static struct sg_table *create_doorbell_sg(uint64_t addr, uint32_t size)
1143 {
1144 	struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
1145 
1146 	if (!sg)
1147 		return NULL;
1148 	if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
1149 		kfree(sg);
1150 		return NULL;
1151 	}
1152 	sg->sgl->dma_address = addr;
1153 	sg->sgl->length = size;
1154 #ifdef CONFIG_NEED_SG_DMA_LENGTH
1155 	sg->sgl->dma_length = size;
1156 #endif
1157 	return sg;
1158 }
1159 
1160 static int process_validate_vms(struct amdkfd_process_info *process_info)
1161 {
1162 	struct amdgpu_vm *peer_vm;
1163 	int ret;
1164 
1165 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1166 			    vm_list_node) {
1167 		ret = vm_validate_pt_pd_bos(peer_vm);
1168 		if (ret)
1169 			return ret;
1170 	}
1171 
1172 	return 0;
1173 }
1174 
1175 static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
1176 				 struct amdgpu_sync *sync)
1177 {
1178 	struct amdgpu_vm *peer_vm;
1179 	int ret;
1180 
1181 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1182 			    vm_list_node) {
1183 		struct amdgpu_bo *pd = peer_vm->root.bo;
1184 
1185 		ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
1186 				       AMDGPU_SYNC_NE_OWNER,
1187 				       AMDGPU_FENCE_OWNER_KFD);
1188 		if (ret)
1189 			return ret;
1190 	}
1191 
1192 	return 0;
1193 }
1194 
1195 static int process_update_pds(struct amdkfd_process_info *process_info,
1196 			      struct amdgpu_sync *sync)
1197 {
1198 	struct amdgpu_vm *peer_vm;
1199 	int ret;
1200 
1201 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1202 			    vm_list_node) {
1203 		ret = vm_update_pds(peer_vm, sync);
1204 		if (ret)
1205 			return ret;
1206 	}
1207 
1208 	return 0;
1209 }
1210 
1211 static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
1212 		       struct dma_fence **ef)
1213 {
1214 	struct amdkfd_process_info *info = NULL;
1215 	int ret;
1216 
1217 	if (!*process_info) {
1218 		info = kzalloc(sizeof(*info), GFP_KERNEL);
1219 		if (!info)
1220 			return -ENOMEM;
1221 
1222 		mutex_init(&info->lock);
1223 		INIT_LIST_HEAD(&info->vm_list_head);
1224 		INIT_LIST_HEAD(&info->kfd_bo_list);
1225 		INIT_LIST_HEAD(&info->userptr_valid_list);
1226 		INIT_LIST_HEAD(&info->userptr_inval_list);
1227 
1228 		info->eviction_fence =
1229 			amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
1230 						   current->mm,
1231 						   NULL);
1232 		if (!info->eviction_fence) {
1233 			pr_err("Failed to create eviction fence\n");
1234 			ret = -ENOMEM;
1235 			goto create_evict_fence_fail;
1236 		}
1237 
1238 		info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
1239 		atomic_set(&info->evicted_bos, 0);
1240 		INIT_DELAYED_WORK(&info->restore_userptr_work,
1241 				  amdgpu_amdkfd_restore_userptr_worker);
1242 
1243 		*process_info = info;
1244 		*ef = dma_fence_get(&info->eviction_fence->base);
1245 	}
1246 
1247 	vm->process_info = *process_info;
1248 
1249 	/* Validate page directory and attach eviction fence */
1250 	ret = amdgpu_bo_reserve(vm->root.bo, true);
1251 	if (ret)
1252 		goto reserve_pd_fail;
1253 	ret = vm_validate_pt_pd_bos(vm);
1254 	if (ret) {
1255 		pr_err("validate_pt_pd_bos() failed\n");
1256 		goto validate_pd_fail;
1257 	}
1258 	ret = amdgpu_bo_sync_wait(vm->root.bo,
1259 				  AMDGPU_FENCE_OWNER_KFD, false);
1260 	if (ret)
1261 		goto wait_pd_fail;
1262 	ret = dma_resv_reserve_shared(vm->root.bo->tbo.base.resv, 1);
1263 	if (ret)
1264 		goto reserve_shared_fail;
1265 	amdgpu_bo_fence(vm->root.bo,
1266 			&vm->process_info->eviction_fence->base, true);
1267 	amdgpu_bo_unreserve(vm->root.bo);
1268 
1269 	/* Update process info */
1270 	mutex_lock(&vm->process_info->lock);
1271 	list_add_tail(&vm->vm_list_node,
1272 			&(vm->process_info->vm_list_head));
1273 	vm->process_info->n_vms++;
1274 	mutex_unlock(&vm->process_info->lock);
1275 
1276 	return 0;
1277 
1278 reserve_shared_fail:
1279 wait_pd_fail:
1280 validate_pd_fail:
1281 	amdgpu_bo_unreserve(vm->root.bo);
1282 reserve_pd_fail:
1283 	vm->process_info = NULL;
1284 	if (info) {
1285 		/* Two fence references: one in info and one in *ef */
1286 		dma_fence_put(&info->eviction_fence->base);
1287 		dma_fence_put(*ef);
1288 		*ef = NULL;
1289 		*process_info = NULL;
1290 		put_pid(info->pid);
1291 create_evict_fence_fail:
1292 		mutex_destroy(&info->lock);
1293 		kfree(info);
1294 	}
1295 	return ret;
1296 }
1297 
1298 /**
1299  * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
1300  * @bo: Handle of buffer object being pinned
1301  * @domain: Domain into which BO should be pinned
1302  *
1303  *   - USERPTR BOs are UNPINNABLE and will return error
1304  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1305  *     PIN count incremented. It is valid to PIN a BO multiple times
1306  *
1307  * Return: ZERO if successful in pinning, Non-Zero in case of error.
1308  */
1309 static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain)
1310 {
1311 	int ret = 0;
1312 
1313 	ret = amdgpu_bo_reserve(bo, false);
1314 	if (unlikely(ret))
1315 		return ret;
1316 
1317 	ret = amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1318 	if (ret)
1319 		pr_err("Error in Pinning BO to domain: %d\n", domain);
1320 
1321 	amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
1322 	amdgpu_bo_unreserve(bo);
1323 
1324 	return ret;
1325 }
1326 
1327 /**
1328  * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
1329  * @bo: Handle of buffer object being unpinned
1330  *
1331  *   - Is a illegal request for USERPTR BOs and is ignored
1332  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1333  *     PIN count decremented. Calls to UNPIN must balance calls to PIN
1334  */
1335 void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo)
1336 {
1337 	int ret = 0;
1338 
1339 	ret = amdgpu_bo_reserve(bo, false);
1340 	if (unlikely(ret))
1341 		return;
1342 
1343 	amdgpu_bo_unpin(bo);
1344 	amdgpu_bo_unreserve(bo);
1345 }
1346 
1347 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
1348 					   struct file *filp, u32 pasid,
1349 					   void **process_info,
1350 					   struct dma_fence **ef)
1351 {
1352 	struct amdgpu_fpriv *drv_priv;
1353 	struct amdgpu_vm *avm;
1354 	int ret;
1355 
1356 	ret = amdgpu_file_to_fpriv(filp, &drv_priv);
1357 	if (ret)
1358 		return ret;
1359 	avm = &drv_priv->vm;
1360 
1361 	/* Already a compute VM? */
1362 	if (avm->process_info)
1363 		return -EINVAL;
1364 
1365 	/* Free the original amdgpu allocated pasid,
1366 	 * will be replaced with kfd allocated pasid.
1367 	 */
1368 	if (avm->pasid) {
1369 		amdgpu_pasid_free(avm->pasid);
1370 		amdgpu_vm_set_pasid(adev, avm, 0);
1371 	}
1372 
1373 	/* Convert VM into a compute VM */
1374 	ret = amdgpu_vm_make_compute(adev, avm);
1375 	if (ret)
1376 		return ret;
1377 
1378 	ret = amdgpu_vm_set_pasid(adev, avm, pasid);
1379 	if (ret)
1380 		return ret;
1381 	/* Initialize KFD part of the VM and process info */
1382 	ret = init_kfd_vm(avm, process_info, ef);
1383 	if (ret)
1384 		return ret;
1385 
1386 	amdgpu_vm_set_task_info(avm);
1387 
1388 	return 0;
1389 }
1390 
1391 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1392 				    struct amdgpu_vm *vm)
1393 {
1394 	struct amdkfd_process_info *process_info = vm->process_info;
1395 	struct amdgpu_bo *pd = vm->root.bo;
1396 
1397 	if (!process_info)
1398 		return;
1399 
1400 	/* Release eviction fence from PD */
1401 	amdgpu_bo_reserve(pd, false);
1402 	amdgpu_bo_fence(pd, NULL, false);
1403 	amdgpu_bo_unreserve(pd);
1404 
1405 	/* Update process info */
1406 	mutex_lock(&process_info->lock);
1407 	process_info->n_vms--;
1408 	list_del(&vm->vm_list_node);
1409 	mutex_unlock(&process_info->lock);
1410 
1411 	vm->process_info = NULL;
1412 
1413 	/* Release per-process resources when last compute VM is destroyed */
1414 	if (!process_info->n_vms) {
1415 		WARN_ON(!list_empty(&process_info->kfd_bo_list));
1416 		WARN_ON(!list_empty(&process_info->userptr_valid_list));
1417 		WARN_ON(!list_empty(&process_info->userptr_inval_list));
1418 
1419 		dma_fence_put(&process_info->eviction_fence->base);
1420 		cancel_delayed_work_sync(&process_info->restore_userptr_work);
1421 		put_pid(process_info->pid);
1422 		mutex_destroy(&process_info->lock);
1423 		kfree(process_info);
1424 	}
1425 }
1426 
1427 void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev,
1428 					    void *drm_priv)
1429 {
1430 	struct amdgpu_vm *avm;
1431 
1432 	if (WARN_ON(!adev || !drm_priv))
1433 		return;
1434 
1435 	avm = drm_priv_to_vm(drm_priv);
1436 
1437 	pr_debug("Releasing process vm %p\n", avm);
1438 
1439 	/* The original pasid of amdgpu vm has already been
1440 	 * released during making a amdgpu vm to a compute vm
1441 	 * The current pasid is managed by kfd and will be
1442 	 * released on kfd process destroy. Set amdgpu pasid
1443 	 * to 0 to avoid duplicate release.
1444 	 */
1445 	amdgpu_vm_release_compute(adev, avm);
1446 }
1447 
1448 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv)
1449 {
1450 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1451 	struct amdgpu_bo *pd = avm->root.bo;
1452 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1453 
1454 	if (adev->asic_type < CHIP_VEGA10)
1455 		return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1456 	return avm->pd_phys_addr;
1457 }
1458 
1459 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1460 		struct amdgpu_device *adev, uint64_t va, uint64_t size,
1461 		void *drm_priv, struct kgd_mem **mem,
1462 		uint64_t *offset, uint32_t flags)
1463 {
1464 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1465 	enum ttm_bo_type bo_type = ttm_bo_type_device;
1466 	struct sg_table *sg = NULL;
1467 	uint64_t user_addr = 0;
1468 	struct amdgpu_bo *bo;
1469 	struct drm_gem_object *gobj;
1470 	u32 domain, alloc_domain;
1471 	u64 alloc_flags;
1472 	int ret;
1473 
1474 	/*
1475 	 * Check on which domain to allocate BO
1476 	 */
1477 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1478 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1479 		alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1480 		alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1481 			AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0;
1482 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1483 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1484 		alloc_flags = 0;
1485 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1486 		domain = AMDGPU_GEM_DOMAIN_GTT;
1487 		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1488 		alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE;
1489 		if (!offset || !*offset)
1490 			return -EINVAL;
1491 		user_addr = untagged_addr(*offset);
1492 	} else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1493 			KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1494 		domain = AMDGPU_GEM_DOMAIN_GTT;
1495 		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1496 		bo_type = ttm_bo_type_sg;
1497 		alloc_flags = 0;
1498 		if (size > UINT_MAX)
1499 			return -EINVAL;
1500 		sg = create_doorbell_sg(*offset, size);
1501 		if (!sg)
1502 			return -ENOMEM;
1503 	} else {
1504 		return -EINVAL;
1505 	}
1506 
1507 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1508 	if (!*mem) {
1509 		ret = -ENOMEM;
1510 		goto err;
1511 	}
1512 	INIT_LIST_HEAD(&(*mem)->attachments);
1513 	mutex_init(&(*mem)->lock);
1514 	(*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1515 
1516 	/* Workaround for AQL queue wraparound bug. Map the same
1517 	 * memory twice. That means we only actually allocate half
1518 	 * the memory.
1519 	 */
1520 	if ((*mem)->aql_queue)
1521 		size = size >> 1;
1522 
1523 	(*mem)->alloc_flags = flags;
1524 
1525 	amdgpu_sync_create(&(*mem)->sync);
1526 
1527 	ret = amdgpu_amdkfd_reserve_mem_limit(adev, size, flags);
1528 	if (ret) {
1529 		pr_debug("Insufficient memory\n");
1530 		goto err_reserve_limit;
1531 	}
1532 
1533 	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n",
1534 			va, size, domain_string(alloc_domain));
1535 
1536 	ret = amdgpu_gem_object_create(adev, size, 1, alloc_domain, alloc_flags,
1537 				       bo_type, NULL, &gobj);
1538 	if (ret) {
1539 		pr_debug("Failed to create BO on domain %s. ret %d\n",
1540 			 domain_string(alloc_domain), ret);
1541 		goto err_bo_create;
1542 	}
1543 	ret = drm_vma_node_allow(&gobj->vma_node, drm_priv);
1544 	if (ret) {
1545 		pr_debug("Failed to allow vma node access. ret %d\n", ret);
1546 		goto err_node_allow;
1547 	}
1548 	bo = gem_to_amdgpu_bo(gobj);
1549 	if (bo_type == ttm_bo_type_sg) {
1550 		bo->tbo.sg = sg;
1551 		bo->tbo.ttm->sg = sg;
1552 	}
1553 	bo->kfd_bo = *mem;
1554 	(*mem)->bo = bo;
1555 	if (user_addr)
1556 		bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO;
1557 
1558 	(*mem)->va = va;
1559 	(*mem)->domain = domain;
1560 	(*mem)->mapped_to_gpu_memory = 0;
1561 	(*mem)->process_info = avm->process_info;
1562 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1563 
1564 	if (user_addr) {
1565 		ret = init_user_pages(*mem, user_addr);
1566 		if (ret)
1567 			goto allocate_init_user_pages_failed;
1568 	}
1569 
1570 	if (offset)
1571 		*offset = amdgpu_bo_mmap_offset(bo);
1572 
1573 	if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1574 			KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1575 		ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT);
1576 		if (ret) {
1577 			pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n");
1578 			goto err_pin_bo;
1579 		}
1580 		bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
1581 		bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
1582 	}
1583 
1584 	return 0;
1585 
1586 allocate_init_user_pages_failed:
1587 	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1588 err_pin_bo:
1589 	drm_vma_node_revoke(&gobj->vma_node, drm_priv);
1590 err_node_allow:
1591 	drm_gem_object_put(gobj);
1592 	/* Don't unreserve system mem limit twice */
1593 	goto err_reserve_limit;
1594 err_bo_create:
1595 	unreserve_mem_limit(adev, size, flags);
1596 err_reserve_limit:
1597 	mutex_destroy(&(*mem)->lock);
1598 	kfree(*mem);
1599 err:
1600 	if (sg) {
1601 		sg_free_table(sg);
1602 		kfree(sg);
1603 	}
1604 	return ret;
1605 }
1606 
1607 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1608 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
1609 		uint64_t *size)
1610 {
1611 	struct amdkfd_process_info *process_info = mem->process_info;
1612 	unsigned long bo_size = mem->bo->tbo.base.size;
1613 	struct kfd_mem_attachment *entry, *tmp;
1614 	struct bo_vm_reservation_context ctx;
1615 	struct ttm_validate_buffer *bo_list_entry;
1616 	unsigned int mapped_to_gpu_memory;
1617 	int ret;
1618 	bool is_imported = false;
1619 
1620 	mutex_lock(&mem->lock);
1621 
1622 	/* Unpin MMIO/DOORBELL BO's that were pinnned during allocation */
1623 	if (mem->alloc_flags &
1624 	    (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1625 	     KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1626 		amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo);
1627 	}
1628 
1629 	mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1630 	is_imported = mem->is_imported;
1631 	mutex_unlock(&mem->lock);
1632 	/* lock is not needed after this, since mem is unused and will
1633 	 * be freed anyway
1634 	 */
1635 
1636 	if (mapped_to_gpu_memory > 0) {
1637 		pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1638 				mem->va, bo_size);
1639 		return -EBUSY;
1640 	}
1641 
1642 	/* Make sure restore workers don't access the BO any more */
1643 	bo_list_entry = &mem->validate_list;
1644 	mutex_lock(&process_info->lock);
1645 	list_del(&bo_list_entry->head);
1646 	mutex_unlock(&process_info->lock);
1647 
1648 	/* No more MMU notifiers */
1649 	amdgpu_mn_unregister(mem->bo);
1650 
1651 	ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1652 	if (unlikely(ret))
1653 		return ret;
1654 
1655 	/* The eviction fence should be removed by the last unmap.
1656 	 * TODO: Log an error condition if the bo still has the eviction fence
1657 	 * attached
1658 	 */
1659 	amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1660 					process_info->eviction_fence);
1661 	pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1662 		mem->va + bo_size * (1 + mem->aql_queue));
1663 
1664 	/* Remove from VM internal data structures */
1665 	list_for_each_entry_safe(entry, tmp, &mem->attachments, list)
1666 		kfd_mem_detach(entry);
1667 
1668 	ret = unreserve_bo_and_vms(&ctx, false, false);
1669 
1670 	/* Free the sync object */
1671 	amdgpu_sync_free(&mem->sync);
1672 
1673 	/* If the SG is not NULL, it's one we created for a doorbell or mmio
1674 	 * remap BO. We need to free it.
1675 	 */
1676 	if (mem->bo->tbo.sg) {
1677 		sg_free_table(mem->bo->tbo.sg);
1678 		kfree(mem->bo->tbo.sg);
1679 	}
1680 
1681 	/* Update the size of the BO being freed if it was allocated from
1682 	 * VRAM and is not imported.
1683 	 */
1684 	if (size) {
1685 		if ((mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM) &&
1686 		    (!is_imported))
1687 			*size = bo_size;
1688 		else
1689 			*size = 0;
1690 	}
1691 
1692 	/* Free the BO*/
1693 	drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
1694 	if (mem->dmabuf)
1695 		dma_buf_put(mem->dmabuf);
1696 	mutex_destroy(&mem->lock);
1697 
1698 	/* If this releases the last reference, it will end up calling
1699 	 * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why
1700 	 * this needs to be the last call here.
1701 	 */
1702 	drm_gem_object_put(&mem->bo->tbo.base);
1703 
1704 	return ret;
1705 }
1706 
1707 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1708 		struct amdgpu_device *adev, struct kgd_mem *mem,
1709 		void *drm_priv, bool *table_freed)
1710 {
1711 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1712 	int ret;
1713 	struct amdgpu_bo *bo;
1714 	uint32_t domain;
1715 	struct kfd_mem_attachment *entry;
1716 	struct bo_vm_reservation_context ctx;
1717 	unsigned long bo_size;
1718 	bool is_invalid_userptr = false;
1719 
1720 	bo = mem->bo;
1721 	if (!bo) {
1722 		pr_err("Invalid BO when mapping memory to GPU\n");
1723 		return -EINVAL;
1724 	}
1725 
1726 	/* Make sure restore is not running concurrently. Since we
1727 	 * don't map invalid userptr BOs, we rely on the next restore
1728 	 * worker to do the mapping
1729 	 */
1730 	mutex_lock(&mem->process_info->lock);
1731 
1732 	/* Lock mmap-sem. If we find an invalid userptr BO, we can be
1733 	 * sure that the MMU notifier is no longer running
1734 	 * concurrently and the queues are actually stopped
1735 	 */
1736 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1737 		mmap_write_lock(current->mm);
1738 		is_invalid_userptr = atomic_read(&mem->invalid);
1739 		mmap_write_unlock(current->mm);
1740 	}
1741 
1742 	mutex_lock(&mem->lock);
1743 
1744 	domain = mem->domain;
1745 	bo_size = bo->tbo.base.size;
1746 
1747 	pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
1748 			mem->va,
1749 			mem->va + bo_size * (1 + mem->aql_queue),
1750 			avm, domain_string(domain));
1751 
1752 	if (!kfd_mem_is_attached(avm, mem)) {
1753 		ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue);
1754 		if (ret)
1755 			goto out;
1756 	}
1757 
1758 	ret = reserve_bo_and_vm(mem, avm, &ctx);
1759 	if (unlikely(ret))
1760 		goto out;
1761 
1762 	/* Userptr can be marked as "not invalid", but not actually be
1763 	 * validated yet (still in the system domain). In that case
1764 	 * the queues are still stopped and we can leave mapping for
1765 	 * the next restore worker
1766 	 */
1767 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
1768 	    bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
1769 		is_invalid_userptr = true;
1770 
1771 	ret = vm_validate_pt_pd_bos(avm);
1772 	if (unlikely(ret))
1773 		goto out_unreserve;
1774 
1775 	if (mem->mapped_to_gpu_memory == 0 &&
1776 	    !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1777 		/* Validate BO only once. The eviction fence gets added to BO
1778 		 * the first time it is mapped. Validate will wait for all
1779 		 * background evictions to complete.
1780 		 */
1781 		ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
1782 		if (ret) {
1783 			pr_debug("Validate failed\n");
1784 			goto out_unreserve;
1785 		}
1786 	}
1787 
1788 	list_for_each_entry(entry, &mem->attachments, list) {
1789 		if (entry->bo_va->base.vm != avm || entry->is_mapped)
1790 			continue;
1791 
1792 		pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
1793 			 entry->va, entry->va + bo_size, entry);
1794 
1795 		ret = map_bo_to_gpuvm(mem, entry, ctx.sync,
1796 				      is_invalid_userptr, table_freed);
1797 		if (ret) {
1798 			pr_err("Failed to map bo to gpuvm\n");
1799 			goto out_unreserve;
1800 		}
1801 
1802 		ret = vm_update_pds(avm, ctx.sync);
1803 		if (ret) {
1804 			pr_err("Failed to update page directories\n");
1805 			goto out_unreserve;
1806 		}
1807 
1808 		entry->is_mapped = true;
1809 		mem->mapped_to_gpu_memory++;
1810 		pr_debug("\t INC mapping count %d\n",
1811 			 mem->mapped_to_gpu_memory);
1812 	}
1813 
1814 	if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->tbo.pin_count)
1815 		amdgpu_bo_fence(bo,
1816 				&avm->process_info->eviction_fence->base,
1817 				true);
1818 	ret = unreserve_bo_and_vms(&ctx, false, false);
1819 
1820 	/* Only apply no TLB flush on Aldebaran to
1821 	 * workaround regressions on other Asics.
1822 	 */
1823 	if (table_freed && (adev->asic_type != CHIP_ALDEBARAN))
1824 		*table_freed = true;
1825 
1826 	goto out;
1827 
1828 out_unreserve:
1829 	unreserve_bo_and_vms(&ctx, false, false);
1830 out:
1831 	mutex_unlock(&mem->process_info->lock);
1832 	mutex_unlock(&mem->lock);
1833 	return ret;
1834 }
1835 
1836 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1837 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
1838 {
1839 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1840 	struct amdkfd_process_info *process_info = avm->process_info;
1841 	unsigned long bo_size = mem->bo->tbo.base.size;
1842 	struct kfd_mem_attachment *entry;
1843 	struct bo_vm_reservation_context ctx;
1844 	int ret;
1845 
1846 	mutex_lock(&mem->lock);
1847 
1848 	ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx);
1849 	if (unlikely(ret))
1850 		goto out;
1851 	/* If no VMs were reserved, it means the BO wasn't actually mapped */
1852 	if (ctx.n_vms == 0) {
1853 		ret = -EINVAL;
1854 		goto unreserve_out;
1855 	}
1856 
1857 	ret = vm_validate_pt_pd_bos(avm);
1858 	if (unlikely(ret))
1859 		goto unreserve_out;
1860 
1861 	pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
1862 		mem->va,
1863 		mem->va + bo_size * (1 + mem->aql_queue),
1864 		avm);
1865 
1866 	list_for_each_entry(entry, &mem->attachments, list) {
1867 		if (entry->bo_va->base.vm != avm || !entry->is_mapped)
1868 			continue;
1869 
1870 		pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
1871 			 entry->va, entry->va + bo_size, entry);
1872 
1873 		unmap_bo_from_gpuvm(mem, entry, ctx.sync);
1874 		entry->is_mapped = false;
1875 
1876 		mem->mapped_to_gpu_memory--;
1877 		pr_debug("\t DEC mapping count %d\n",
1878 			 mem->mapped_to_gpu_memory);
1879 	}
1880 
1881 	/* If BO is unmapped from all VMs, unfence it. It can be evicted if
1882 	 * required.
1883 	 */
1884 	if (mem->mapped_to_gpu_memory == 0 &&
1885 	    !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) &&
1886 	    !mem->bo->tbo.pin_count)
1887 		amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1888 						process_info->eviction_fence);
1889 
1890 unreserve_out:
1891 	unreserve_bo_and_vms(&ctx, false, false);
1892 out:
1893 	mutex_unlock(&mem->lock);
1894 	return ret;
1895 }
1896 
1897 int amdgpu_amdkfd_gpuvm_sync_memory(
1898 		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr)
1899 {
1900 	struct amdgpu_sync sync;
1901 	int ret;
1902 
1903 	amdgpu_sync_create(&sync);
1904 
1905 	mutex_lock(&mem->lock);
1906 	amdgpu_sync_clone(&mem->sync, &sync);
1907 	mutex_unlock(&mem->lock);
1908 
1909 	ret = amdgpu_sync_wait(&sync, intr);
1910 	amdgpu_sync_free(&sync);
1911 	return ret;
1912 }
1913 
1914 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct amdgpu_device *adev,
1915 		struct kgd_mem *mem, void **kptr, uint64_t *size)
1916 {
1917 	int ret;
1918 	struct amdgpu_bo *bo = mem->bo;
1919 
1920 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1921 		pr_err("userptr can't be mapped to kernel\n");
1922 		return -EINVAL;
1923 	}
1924 
1925 	/* delete kgd_mem from kfd_bo_list to avoid re-validating
1926 	 * this BO in BO's restoring after eviction.
1927 	 */
1928 	mutex_lock(&mem->process_info->lock);
1929 
1930 	ret = amdgpu_bo_reserve(bo, true);
1931 	if (ret) {
1932 		pr_err("Failed to reserve bo. ret %d\n", ret);
1933 		goto bo_reserve_failed;
1934 	}
1935 
1936 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
1937 	if (ret) {
1938 		pr_err("Failed to pin bo. ret %d\n", ret);
1939 		goto pin_failed;
1940 	}
1941 
1942 	ret = amdgpu_bo_kmap(bo, kptr);
1943 	if (ret) {
1944 		pr_err("Failed to map bo to kernel. ret %d\n", ret);
1945 		goto kmap_failed;
1946 	}
1947 
1948 	amdgpu_amdkfd_remove_eviction_fence(
1949 		bo, mem->process_info->eviction_fence);
1950 	list_del_init(&mem->validate_list.head);
1951 
1952 	if (size)
1953 		*size = amdgpu_bo_size(bo);
1954 
1955 	amdgpu_bo_unreserve(bo);
1956 
1957 	mutex_unlock(&mem->process_info->lock);
1958 	return 0;
1959 
1960 kmap_failed:
1961 	amdgpu_bo_unpin(bo);
1962 pin_failed:
1963 	amdgpu_bo_unreserve(bo);
1964 bo_reserve_failed:
1965 	mutex_unlock(&mem->process_info->lock);
1966 
1967 	return ret;
1968 }
1969 
1970 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct amdgpu_device *adev,
1971 						  struct kgd_mem *mem)
1972 {
1973 	struct amdgpu_bo *bo = mem->bo;
1974 
1975 	amdgpu_bo_reserve(bo, true);
1976 	amdgpu_bo_kunmap(bo);
1977 	amdgpu_bo_unpin(bo);
1978 	amdgpu_bo_unreserve(bo);
1979 }
1980 
1981 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
1982 					  struct kfd_vm_fault_info *mem)
1983 {
1984 	if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
1985 		*mem = *adev->gmc.vm_fault_info;
1986 		mb();
1987 		atomic_set(&adev->gmc.vm_fault_info_updated, 0);
1988 	}
1989 	return 0;
1990 }
1991 
1992 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
1993 				      struct dma_buf *dma_buf,
1994 				      uint64_t va, void *drm_priv,
1995 				      struct kgd_mem **mem, uint64_t *size,
1996 				      uint64_t *mmap_offset)
1997 {
1998 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1999 	struct drm_gem_object *obj;
2000 	struct amdgpu_bo *bo;
2001 	int ret;
2002 
2003 	if (dma_buf->ops != &amdgpu_dmabuf_ops)
2004 		/* Can't handle non-graphics buffers */
2005 		return -EINVAL;
2006 
2007 	obj = dma_buf->priv;
2008 	if (drm_to_adev(obj->dev) != adev)
2009 		/* Can't handle buffers from other devices */
2010 		return -EINVAL;
2011 
2012 	bo = gem_to_amdgpu_bo(obj);
2013 	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
2014 				    AMDGPU_GEM_DOMAIN_GTT)))
2015 		/* Only VRAM and GTT BOs are supported */
2016 		return -EINVAL;
2017 
2018 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2019 	if (!*mem)
2020 		return -ENOMEM;
2021 
2022 	ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
2023 	if (ret) {
2024 		kfree(mem);
2025 		return ret;
2026 	}
2027 
2028 	if (size)
2029 		*size = amdgpu_bo_size(bo);
2030 
2031 	if (mmap_offset)
2032 		*mmap_offset = amdgpu_bo_mmap_offset(bo);
2033 
2034 	INIT_LIST_HEAD(&(*mem)->attachments);
2035 	mutex_init(&(*mem)->lock);
2036 
2037 	(*mem)->alloc_flags =
2038 		((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2039 		KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
2040 		| KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
2041 		| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
2042 
2043 	drm_gem_object_get(&bo->tbo.base);
2044 	(*mem)->bo = bo;
2045 	(*mem)->va = va;
2046 	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2047 		AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
2048 	(*mem)->mapped_to_gpu_memory = 0;
2049 	(*mem)->process_info = avm->process_info;
2050 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
2051 	amdgpu_sync_create(&(*mem)->sync);
2052 	(*mem)->is_imported = true;
2053 
2054 	return 0;
2055 }
2056 
2057 /* Evict a userptr BO by stopping the queues if necessary
2058  *
2059  * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
2060  * cannot do any memory allocations, and cannot take any locks that
2061  * are held elsewhere while allocating memory. Therefore this is as
2062  * simple as possible, using atomic counters.
2063  *
2064  * It doesn't do anything to the BO itself. The real work happens in
2065  * restore, where we get updated page addresses. This function only
2066  * ensures that GPU access to the BO is stopped.
2067  */
2068 int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem,
2069 				struct mm_struct *mm)
2070 {
2071 	struct amdkfd_process_info *process_info = mem->process_info;
2072 	int evicted_bos;
2073 	int r = 0;
2074 
2075 	atomic_inc(&mem->invalid);
2076 	evicted_bos = atomic_inc_return(&process_info->evicted_bos);
2077 	if (evicted_bos == 1) {
2078 		/* First eviction, stop the queues */
2079 		r = kgd2kfd_quiesce_mm(mm);
2080 		if (r)
2081 			pr_err("Failed to quiesce KFD\n");
2082 		schedule_delayed_work(&process_info->restore_userptr_work,
2083 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2084 	}
2085 
2086 	return r;
2087 }
2088 
2089 /* Update invalid userptr BOs
2090  *
2091  * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
2092  * userptr_inval_list and updates user pages for all BOs that have
2093  * been invalidated since their last update.
2094  */
2095 static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
2096 				     struct mm_struct *mm)
2097 {
2098 	struct kgd_mem *mem, *tmp_mem;
2099 	struct amdgpu_bo *bo;
2100 	struct ttm_operation_ctx ctx = { false, false };
2101 	int invalid, ret;
2102 
2103 	/* Move all invalidated BOs to the userptr_inval_list and
2104 	 * release their user pages by migration to the CPU domain
2105 	 */
2106 	list_for_each_entry_safe(mem, tmp_mem,
2107 				 &process_info->userptr_valid_list,
2108 				 validate_list.head) {
2109 		if (!atomic_read(&mem->invalid))
2110 			continue; /* BO is still valid */
2111 
2112 		bo = mem->bo;
2113 
2114 		if (amdgpu_bo_reserve(bo, true))
2115 			return -EAGAIN;
2116 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
2117 		ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2118 		amdgpu_bo_unreserve(bo);
2119 		if (ret) {
2120 			pr_err("%s: Failed to invalidate userptr BO\n",
2121 			       __func__);
2122 			return -EAGAIN;
2123 		}
2124 
2125 		list_move_tail(&mem->validate_list.head,
2126 			       &process_info->userptr_inval_list);
2127 	}
2128 
2129 	if (list_empty(&process_info->userptr_inval_list))
2130 		return 0; /* All evicted userptr BOs were freed */
2131 
2132 	/* Go through userptr_inval_list and update any invalid user_pages */
2133 	list_for_each_entry(mem, &process_info->userptr_inval_list,
2134 			    validate_list.head) {
2135 		invalid = atomic_read(&mem->invalid);
2136 		if (!invalid)
2137 			/* BO hasn't been invalidated since the last
2138 			 * revalidation attempt. Keep its BO list.
2139 			 */
2140 			continue;
2141 
2142 		bo = mem->bo;
2143 
2144 		/* Get updated user pages */
2145 		ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
2146 		if (ret) {
2147 			pr_debug("Failed %d to get user pages\n", ret);
2148 
2149 			/* Return -EFAULT bad address error as success. It will
2150 			 * fail later with a VM fault if the GPU tries to access
2151 			 * it. Better than hanging indefinitely with stalled
2152 			 * user mode queues.
2153 			 *
2154 			 * Return other error -EBUSY or -ENOMEM to retry restore
2155 			 */
2156 			if (ret != -EFAULT)
2157 				return ret;
2158 		} else {
2159 
2160 			/*
2161 			 * FIXME: Cannot ignore the return code, must hold
2162 			 * notifier_lock
2163 			 */
2164 			amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
2165 		}
2166 
2167 		/* Mark the BO as valid unless it was invalidated
2168 		 * again concurrently.
2169 		 */
2170 		if (atomic_cmpxchg(&mem->invalid, invalid, 0) != invalid)
2171 			return -EAGAIN;
2172 	}
2173 
2174 	return 0;
2175 }
2176 
2177 /* Validate invalid userptr BOs
2178  *
2179  * Validates BOs on the userptr_inval_list, and moves them back to the
2180  * userptr_valid_list. Also updates GPUVM page tables with new page
2181  * addresses and waits for the page table updates to complete.
2182  */
2183 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
2184 {
2185 	struct amdgpu_bo_list_entry *pd_bo_list_entries;
2186 	struct list_head resv_list, duplicates;
2187 	struct ww_acquire_ctx ticket;
2188 	struct amdgpu_sync sync;
2189 
2190 	struct amdgpu_vm *peer_vm;
2191 	struct kgd_mem *mem, *tmp_mem;
2192 	struct amdgpu_bo *bo;
2193 	struct ttm_operation_ctx ctx = { false, false };
2194 	int i, ret;
2195 
2196 	pd_bo_list_entries = kcalloc(process_info->n_vms,
2197 				     sizeof(struct amdgpu_bo_list_entry),
2198 				     GFP_KERNEL);
2199 	if (!pd_bo_list_entries) {
2200 		pr_err("%s: Failed to allocate PD BO list entries\n", __func__);
2201 		ret = -ENOMEM;
2202 		goto out_no_mem;
2203 	}
2204 
2205 	INIT_LIST_HEAD(&resv_list);
2206 	INIT_LIST_HEAD(&duplicates);
2207 
2208 	/* Get all the page directory BOs that need to be reserved */
2209 	i = 0;
2210 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2211 			    vm_list_node)
2212 		amdgpu_vm_get_pd_bo(peer_vm, &resv_list,
2213 				    &pd_bo_list_entries[i++]);
2214 	/* Add the userptr_inval_list entries to resv_list */
2215 	list_for_each_entry(mem, &process_info->userptr_inval_list,
2216 			    validate_list.head) {
2217 		list_add_tail(&mem->resv_list.head, &resv_list);
2218 		mem->resv_list.bo = mem->validate_list.bo;
2219 		mem->resv_list.num_shared = mem->validate_list.num_shared;
2220 	}
2221 
2222 	/* Reserve all BOs and page tables for validation */
2223 	ret = ttm_eu_reserve_buffers(&ticket, &resv_list, false, &duplicates);
2224 	WARN(!list_empty(&duplicates), "Duplicates should be empty");
2225 	if (ret)
2226 		goto out_free;
2227 
2228 	amdgpu_sync_create(&sync);
2229 
2230 	ret = process_validate_vms(process_info);
2231 	if (ret)
2232 		goto unreserve_out;
2233 
2234 	/* Validate BOs and update GPUVM page tables */
2235 	list_for_each_entry_safe(mem, tmp_mem,
2236 				 &process_info->userptr_inval_list,
2237 				 validate_list.head) {
2238 		struct kfd_mem_attachment *attachment;
2239 
2240 		bo = mem->bo;
2241 
2242 		/* Validate the BO if we got user pages */
2243 		if (bo->tbo.ttm->pages[0]) {
2244 			amdgpu_bo_placement_from_domain(bo, mem->domain);
2245 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2246 			if (ret) {
2247 				pr_err("%s: failed to validate BO\n", __func__);
2248 				goto unreserve_out;
2249 			}
2250 		}
2251 
2252 		list_move_tail(&mem->validate_list.head,
2253 			       &process_info->userptr_valid_list);
2254 
2255 		/* Update mapping. If the BO was not validated
2256 		 * (because we couldn't get user pages), this will
2257 		 * clear the page table entries, which will result in
2258 		 * VM faults if the GPU tries to access the invalid
2259 		 * memory.
2260 		 */
2261 		list_for_each_entry(attachment, &mem->attachments, list) {
2262 			if (!attachment->is_mapped)
2263 				continue;
2264 
2265 			kfd_mem_dmaunmap_attachment(mem, attachment);
2266 			ret = update_gpuvm_pte(mem, attachment, &sync, NULL);
2267 			if (ret) {
2268 				pr_err("%s: update PTE failed\n", __func__);
2269 				/* make sure this gets validated again */
2270 				atomic_inc(&mem->invalid);
2271 				goto unreserve_out;
2272 			}
2273 		}
2274 	}
2275 
2276 	/* Update page directories */
2277 	ret = process_update_pds(process_info, &sync);
2278 
2279 unreserve_out:
2280 	ttm_eu_backoff_reservation(&ticket, &resv_list);
2281 	amdgpu_sync_wait(&sync, false);
2282 	amdgpu_sync_free(&sync);
2283 out_free:
2284 	kfree(pd_bo_list_entries);
2285 out_no_mem:
2286 
2287 	return ret;
2288 }
2289 
2290 /* Worker callback to restore evicted userptr BOs
2291  *
2292  * Tries to update and validate all userptr BOs. If successful and no
2293  * concurrent evictions happened, the queues are restarted. Otherwise,
2294  * reschedule for another attempt later.
2295  */
2296 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
2297 {
2298 	struct delayed_work *dwork = to_delayed_work(work);
2299 	struct amdkfd_process_info *process_info =
2300 		container_of(dwork, struct amdkfd_process_info,
2301 			     restore_userptr_work);
2302 	struct task_struct *usertask;
2303 	struct mm_struct *mm;
2304 	int evicted_bos;
2305 
2306 	evicted_bos = atomic_read(&process_info->evicted_bos);
2307 	if (!evicted_bos)
2308 		return;
2309 
2310 	/* Reference task and mm in case of concurrent process termination */
2311 	usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
2312 	if (!usertask)
2313 		return;
2314 	mm = get_task_mm(usertask);
2315 	if (!mm) {
2316 		put_task_struct(usertask);
2317 		return;
2318 	}
2319 
2320 	mutex_lock(&process_info->lock);
2321 
2322 	if (update_invalid_user_pages(process_info, mm))
2323 		goto unlock_out;
2324 	/* userptr_inval_list can be empty if all evicted userptr BOs
2325 	 * have been freed. In that case there is nothing to validate
2326 	 * and we can just restart the queues.
2327 	 */
2328 	if (!list_empty(&process_info->userptr_inval_list)) {
2329 		if (atomic_read(&process_info->evicted_bos) != evicted_bos)
2330 			goto unlock_out; /* Concurrent eviction, try again */
2331 
2332 		if (validate_invalid_user_pages(process_info))
2333 			goto unlock_out;
2334 	}
2335 	/* Final check for concurrent evicton and atomic update. If
2336 	 * another eviction happens after successful update, it will
2337 	 * be a first eviction that calls quiesce_mm. The eviction
2338 	 * reference counting inside KFD will handle this case.
2339 	 */
2340 	if (atomic_cmpxchg(&process_info->evicted_bos, evicted_bos, 0) !=
2341 	    evicted_bos)
2342 		goto unlock_out;
2343 	evicted_bos = 0;
2344 	if (kgd2kfd_resume_mm(mm)) {
2345 		pr_err("%s: Failed to resume KFD\n", __func__);
2346 		/* No recovery from this failure. Probably the CP is
2347 		 * hanging. No point trying again.
2348 		 */
2349 	}
2350 
2351 unlock_out:
2352 	mutex_unlock(&process_info->lock);
2353 	mmput(mm);
2354 	put_task_struct(usertask);
2355 
2356 	/* If validation failed, reschedule another attempt */
2357 	if (evicted_bos)
2358 		schedule_delayed_work(&process_info->restore_userptr_work,
2359 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2360 }
2361 
2362 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2363  *   KFD process identified by process_info
2364  *
2365  * @process_info: amdkfd_process_info of the KFD process
2366  *
2367  * After memory eviction, restore thread calls this function. The function
2368  * should be called when the Process is still valid. BO restore involves -
2369  *
2370  * 1.  Release old eviction fence and create new one
2371  * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2372  * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2373  *     BOs that need to be reserved.
2374  * 4.  Reserve all the BOs
2375  * 5.  Validate of PD and PT BOs.
2376  * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2377  * 7.  Add fence to all PD and PT BOs.
2378  * 8.  Unreserve all BOs
2379  */
2380 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
2381 {
2382 	struct amdgpu_bo_list_entry *pd_bo_list;
2383 	struct amdkfd_process_info *process_info = info;
2384 	struct amdgpu_vm *peer_vm;
2385 	struct kgd_mem *mem;
2386 	struct bo_vm_reservation_context ctx;
2387 	struct amdgpu_amdkfd_fence *new_fence;
2388 	int ret = 0, i;
2389 	struct list_head duplicate_save;
2390 	struct amdgpu_sync sync_obj;
2391 	unsigned long failed_size = 0;
2392 	unsigned long total_size = 0;
2393 
2394 	INIT_LIST_HEAD(&duplicate_save);
2395 	INIT_LIST_HEAD(&ctx.list);
2396 	INIT_LIST_HEAD(&ctx.duplicates);
2397 
2398 	pd_bo_list = kcalloc(process_info->n_vms,
2399 			     sizeof(struct amdgpu_bo_list_entry),
2400 			     GFP_KERNEL);
2401 	if (!pd_bo_list)
2402 		return -ENOMEM;
2403 
2404 	i = 0;
2405 	mutex_lock(&process_info->lock);
2406 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2407 			vm_list_node)
2408 		amdgpu_vm_get_pd_bo(peer_vm, &ctx.list, &pd_bo_list[i++]);
2409 
2410 	/* Reserve all BOs and page tables/directory. Add all BOs from
2411 	 * kfd_bo_list to ctx.list
2412 	 */
2413 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2414 			    validate_list.head) {
2415 
2416 		list_add_tail(&mem->resv_list.head, &ctx.list);
2417 		mem->resv_list.bo = mem->validate_list.bo;
2418 		mem->resv_list.num_shared = mem->validate_list.num_shared;
2419 	}
2420 
2421 	ret = ttm_eu_reserve_buffers(&ctx.ticket, &ctx.list,
2422 				     false, &duplicate_save);
2423 	if (ret) {
2424 		pr_debug("Memory eviction: TTM Reserve Failed. Try again\n");
2425 		goto ttm_reserve_fail;
2426 	}
2427 
2428 	amdgpu_sync_create(&sync_obj);
2429 
2430 	/* Validate PDs and PTs */
2431 	ret = process_validate_vms(process_info);
2432 	if (ret)
2433 		goto validate_map_fail;
2434 
2435 	ret = process_sync_pds_resv(process_info, &sync_obj);
2436 	if (ret) {
2437 		pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2438 		goto validate_map_fail;
2439 	}
2440 
2441 	/* Validate BOs and map them to GPUVM (update VM page tables). */
2442 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2443 			    validate_list.head) {
2444 
2445 		struct amdgpu_bo *bo = mem->bo;
2446 		uint32_t domain = mem->domain;
2447 		struct kfd_mem_attachment *attachment;
2448 
2449 		total_size += amdgpu_bo_size(bo);
2450 
2451 		ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2452 		if (ret) {
2453 			pr_debug("Memory eviction: Validate BOs failed\n");
2454 			failed_size += amdgpu_bo_size(bo);
2455 			ret = amdgpu_amdkfd_bo_validate(bo,
2456 						AMDGPU_GEM_DOMAIN_GTT, false);
2457 			if (ret) {
2458 				pr_debug("Memory eviction: Try again\n");
2459 				goto validate_map_fail;
2460 			}
2461 		}
2462 		ret = amdgpu_sync_fence(&sync_obj, bo->tbo.moving);
2463 		if (ret) {
2464 			pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2465 			goto validate_map_fail;
2466 		}
2467 		list_for_each_entry(attachment, &mem->attachments, list) {
2468 			if (!attachment->is_mapped)
2469 				continue;
2470 
2471 			kfd_mem_dmaunmap_attachment(mem, attachment);
2472 			ret = update_gpuvm_pte(mem, attachment, &sync_obj, NULL);
2473 			if (ret) {
2474 				pr_debug("Memory eviction: update PTE failed. Try again\n");
2475 				goto validate_map_fail;
2476 			}
2477 		}
2478 	}
2479 
2480 	if (failed_size)
2481 		pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2482 
2483 	/* Update page directories */
2484 	ret = process_update_pds(process_info, &sync_obj);
2485 	if (ret) {
2486 		pr_debug("Memory eviction: update PDs failed. Try again\n");
2487 		goto validate_map_fail;
2488 	}
2489 
2490 	/* Wait for validate and PT updates to finish */
2491 	amdgpu_sync_wait(&sync_obj, false);
2492 
2493 	/* Release old eviction fence and create new one, because fence only
2494 	 * goes from unsignaled to signaled, fence cannot be reused.
2495 	 * Use context and mm from the old fence.
2496 	 */
2497 	new_fence = amdgpu_amdkfd_fence_create(
2498 				process_info->eviction_fence->base.context,
2499 				process_info->eviction_fence->mm,
2500 				NULL);
2501 	if (!new_fence) {
2502 		pr_err("Failed to create eviction fence\n");
2503 		ret = -ENOMEM;
2504 		goto validate_map_fail;
2505 	}
2506 	dma_fence_put(&process_info->eviction_fence->base);
2507 	process_info->eviction_fence = new_fence;
2508 	*ef = dma_fence_get(&new_fence->base);
2509 
2510 	/* Attach new eviction fence to all BOs */
2511 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2512 		validate_list.head)
2513 		amdgpu_bo_fence(mem->bo,
2514 			&process_info->eviction_fence->base, true);
2515 
2516 	/* Attach eviction fence to PD / PT BOs */
2517 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2518 			    vm_list_node) {
2519 		struct amdgpu_bo *bo = peer_vm->root.bo;
2520 
2521 		amdgpu_bo_fence(bo, &process_info->eviction_fence->base, true);
2522 	}
2523 
2524 validate_map_fail:
2525 	ttm_eu_backoff_reservation(&ctx.ticket, &ctx.list);
2526 	amdgpu_sync_free(&sync_obj);
2527 ttm_reserve_fail:
2528 	mutex_unlock(&process_info->lock);
2529 	kfree(pd_bo_list);
2530 	return ret;
2531 }
2532 
2533 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
2534 {
2535 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2536 	struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
2537 	int ret;
2538 
2539 	if (!info || !gws)
2540 		return -EINVAL;
2541 
2542 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2543 	if (!*mem)
2544 		return -ENOMEM;
2545 
2546 	mutex_init(&(*mem)->lock);
2547 	INIT_LIST_HEAD(&(*mem)->attachments);
2548 	(*mem)->bo = amdgpu_bo_ref(gws_bo);
2549 	(*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
2550 	(*mem)->process_info = process_info;
2551 	add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
2552 	amdgpu_sync_create(&(*mem)->sync);
2553 
2554 
2555 	/* Validate gws bo the first time it is added to process */
2556 	mutex_lock(&(*mem)->process_info->lock);
2557 	ret = amdgpu_bo_reserve(gws_bo, false);
2558 	if (unlikely(ret)) {
2559 		pr_err("Reserve gws bo failed %d\n", ret);
2560 		goto bo_reservation_failure;
2561 	}
2562 
2563 	ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
2564 	if (ret) {
2565 		pr_err("GWS BO validate failed %d\n", ret);
2566 		goto bo_validation_failure;
2567 	}
2568 	/* GWS resource is shared b/t amdgpu and amdkfd
2569 	 * Add process eviction fence to bo so they can
2570 	 * evict each other.
2571 	 */
2572 	ret = dma_resv_reserve_shared(gws_bo->tbo.base.resv, 1);
2573 	if (ret)
2574 		goto reserve_shared_fail;
2575 	amdgpu_bo_fence(gws_bo, &process_info->eviction_fence->base, true);
2576 	amdgpu_bo_unreserve(gws_bo);
2577 	mutex_unlock(&(*mem)->process_info->lock);
2578 
2579 	return ret;
2580 
2581 reserve_shared_fail:
2582 bo_validation_failure:
2583 	amdgpu_bo_unreserve(gws_bo);
2584 bo_reservation_failure:
2585 	mutex_unlock(&(*mem)->process_info->lock);
2586 	amdgpu_sync_free(&(*mem)->sync);
2587 	remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
2588 	amdgpu_bo_unref(&gws_bo);
2589 	mutex_destroy(&(*mem)->lock);
2590 	kfree(*mem);
2591 	*mem = NULL;
2592 	return ret;
2593 }
2594 
2595 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
2596 {
2597 	int ret;
2598 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2599 	struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
2600 	struct amdgpu_bo *gws_bo = kgd_mem->bo;
2601 
2602 	/* Remove BO from process's validate list so restore worker won't touch
2603 	 * it anymore
2604 	 */
2605 	remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
2606 
2607 	ret = amdgpu_bo_reserve(gws_bo, false);
2608 	if (unlikely(ret)) {
2609 		pr_err("Reserve gws bo failed %d\n", ret);
2610 		//TODO add BO back to validate_list?
2611 		return ret;
2612 	}
2613 	amdgpu_amdkfd_remove_eviction_fence(gws_bo,
2614 			process_info->eviction_fence);
2615 	amdgpu_bo_unreserve(gws_bo);
2616 	amdgpu_sync_free(&kgd_mem->sync);
2617 	amdgpu_bo_unref(&gws_bo);
2618 	mutex_destroy(&kgd_mem->lock);
2619 	kfree(mem);
2620 	return 0;
2621 }
2622 
2623 /* Returns GPU-specific tiling mode information */
2624 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
2625 				struct tile_config *config)
2626 {
2627 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
2628 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
2629 	config->num_tile_configs =
2630 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
2631 	config->macro_tile_config_ptr =
2632 			adev->gfx.config.macrotile_mode_array;
2633 	config->num_macro_tile_configs =
2634 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
2635 
2636 	/* Those values are not set from GFX9 onwards */
2637 	config->num_banks = adev->gfx.config.num_banks;
2638 	config->num_ranks = adev->gfx.config.num_ranks;
2639 
2640 	return 0;
2641 }
2642