xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c (revision 25396684b57f7d16306ca149c545db60b2d08dda)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright 2014-2018 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 #include <linux/dma-buf.h>
24 #include <linux/list.h>
25 #include <linux/pagemap.h>
26 #include <linux/sched/mm.h>
27 #include <linux/sched/task.h>
28 #include <drm/ttm/ttm_tt.h>
29 
30 #include <drm/drm_exec.h>
31 
32 #include "amdgpu_object.h"
33 #include "amdgpu_gem.h"
34 #include "amdgpu_vm.h"
35 #include "amdgpu_hmm.h"
36 #include "amdgpu_amdkfd.h"
37 #include "amdgpu_dma_buf.h"
38 #include <uapi/linux/kfd_ioctl.h>
39 #include "amdgpu_xgmi.h"
40 #include "kfd_priv.h"
41 #include "kfd_smi_events.h"
42 
43 /* Userptr restore delay, just long enough to allow consecutive VM
44  * changes to accumulate
45  */
46 #define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
47 
48 /*
49  * Align VRAM availability to 2MB to avoid fragmentation caused by 4K allocations in the tail 2MB
50  * BO chunk
51  */
52 #define VRAM_AVAILABLITY_ALIGN (1 << 21)
53 
54 /* Impose limit on how much memory KFD can use */
55 static struct {
56 	uint64_t max_system_mem_limit;
57 	uint64_t max_ttm_mem_limit;
58 	int64_t system_mem_used;
59 	int64_t ttm_mem_used;
60 	spinlock_t mem_limit_lock;
61 } kfd_mem_limit;
62 
63 static const char * const domain_bit_to_string[] = {
64 		"CPU",
65 		"GTT",
66 		"VRAM",
67 		"GDS",
68 		"GWS",
69 		"OA"
70 };
71 
72 #define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
73 
74 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work);
75 
76 static bool kfd_mem_is_attached(struct amdgpu_vm *avm,
77 		struct kgd_mem *mem)
78 {
79 	struct kfd_mem_attachment *entry;
80 
81 	list_for_each_entry(entry, &mem->attachments, list)
82 		if (entry->bo_va->base.vm == avm)
83 			return true;
84 
85 	return false;
86 }
87 
88 /**
89  * reuse_dmamap() - Check whether adev can share the original
90  * userptr BO
91  *
92  * If both adev and bo_adev are in direct mapping or
93  * in the same iommu group, they can share the original BO.
94  *
95  * @adev: Device to which can or cannot share the original BO
96  * @bo_adev: Device to which allocated BO belongs to
97  *
98  * Return: returns true if adev can share original userptr BO,
99  * false otherwise.
100  */
101 static bool reuse_dmamap(struct amdgpu_device *adev, struct amdgpu_device *bo_adev)
102 {
103 	return (adev->ram_is_direct_mapped && bo_adev->ram_is_direct_mapped) ||
104 			(adev->dev->iommu_group == bo_adev->dev->iommu_group);
105 }
106 
107 /* Set memory usage limits. Current, limits are
108  *  System (TTM + userptr) memory - 15/16th System RAM
109  *  TTM memory - 3/8th System RAM
110  */
111 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
112 {
113 	struct sysinfo si;
114 	uint64_t mem;
115 
116 	if (kfd_mem_limit.max_system_mem_limit)
117 		return;
118 
119 	si_meminfo(&si);
120 	mem = si.freeram - si.freehigh;
121 	mem *= si.mem_unit;
122 
123 	spin_lock_init(&kfd_mem_limit.mem_limit_lock);
124 	kfd_mem_limit.max_system_mem_limit = mem - (mem >> 4);
125 	kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() << PAGE_SHIFT;
126 	pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
127 		(kfd_mem_limit.max_system_mem_limit >> 20),
128 		(kfd_mem_limit.max_ttm_mem_limit >> 20));
129 }
130 
131 void amdgpu_amdkfd_reserve_system_mem(uint64_t size)
132 {
133 	kfd_mem_limit.system_mem_used += size;
134 }
135 
136 /* Estimate page table size needed to represent a given memory size
137  *
138  * With 4KB pages, we need one 8 byte PTE for each 4KB of memory
139  * (factor 512, >> 9). With 2MB pages, we need one 8 byte PTE for 2MB
140  * of memory (factor 256K, >> 18). ROCm user mode tries to optimize
141  * for 2MB pages for TLB efficiency. However, small allocations and
142  * fragmented system memory still need some 4KB pages. We choose a
143  * compromise that should work in most cases without reserving too
144  * much memory for page tables unnecessarily (factor 16K, >> 14).
145  */
146 
147 #define ESTIMATE_PT_SIZE(mem_size) max(((mem_size) >> 14), AMDGPU_VM_RESERVED_VRAM)
148 
149 /**
150  * amdgpu_amdkfd_reserve_mem_limit() - Decrease available memory by size
151  * of buffer.
152  *
153  * @adev: Device to which allocated BO belongs to
154  * @size: Size of buffer, in bytes, encapsulated by B0. This should be
155  * equivalent to amdgpu_bo_size(BO)
156  * @alloc_flag: Flag used in allocating a BO as noted above
157  * @xcp_id: xcp_id is used to get xcp from xcp manager, one xcp is
158  * managed as one compute node in driver for app
159  *
160  * Return:
161  *	returns -ENOMEM in case of error, ZERO otherwise
162  */
163 int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
164 		uint64_t size, u32 alloc_flag, int8_t xcp_id)
165 {
166 	uint64_t reserved_for_pt =
167 		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
168 	size_t system_mem_needed, ttm_mem_needed, vram_needed;
169 	int ret = 0;
170 	uint64_t vram_size = 0;
171 
172 	system_mem_needed = 0;
173 	ttm_mem_needed = 0;
174 	vram_needed = 0;
175 	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
176 		system_mem_needed = size;
177 		ttm_mem_needed = size;
178 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
179 		/*
180 		 * Conservatively round up the allocation requirement to 2 MB
181 		 * to avoid fragmentation caused by 4K allocations in the tail
182 		 * 2M BO chunk.
183 		 */
184 		vram_needed = size;
185 		/*
186 		 * For GFX 9.4.3, get the VRAM size from XCP structs
187 		 */
188 		if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
189 			return -EINVAL;
190 
191 		vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id);
192 		if (adev->gmc.is_app_apu) {
193 			system_mem_needed = size;
194 			ttm_mem_needed = size;
195 		}
196 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
197 		system_mem_needed = size;
198 	} else if (!(alloc_flag &
199 				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
200 				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
201 		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
202 		return -ENOMEM;
203 	}
204 
205 	spin_lock(&kfd_mem_limit.mem_limit_lock);
206 
207 	if (kfd_mem_limit.system_mem_used + system_mem_needed >
208 	    kfd_mem_limit.max_system_mem_limit)
209 		pr_debug("Set no_system_mem_limit=1 if using shared memory\n");
210 
211 	if ((kfd_mem_limit.system_mem_used + system_mem_needed >
212 	     kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) ||
213 	    (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
214 	     kfd_mem_limit.max_ttm_mem_limit) ||
215 	    (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] + vram_needed >
216 	     vram_size - reserved_for_pt)) {
217 		ret = -ENOMEM;
218 		goto release;
219 	}
220 
221 	/* Update memory accounting by decreasing available system
222 	 * memory, TTM memory and GPU memory as computed above
223 	 */
224 	WARN_ONCE(vram_needed && !adev,
225 		  "adev reference can't be null when vram is used");
226 	if (adev && xcp_id >= 0) {
227 		adev->kfd.vram_used[xcp_id] += vram_needed;
228 		adev->kfd.vram_used_aligned[xcp_id] += adev->gmc.is_app_apu ?
229 				vram_needed :
230 				ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
231 	}
232 	kfd_mem_limit.system_mem_used += system_mem_needed;
233 	kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
234 
235 release:
236 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
237 	return ret;
238 }
239 
240 void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
241 		uint64_t size, u32 alloc_flag, int8_t xcp_id)
242 {
243 	spin_lock(&kfd_mem_limit.mem_limit_lock);
244 
245 	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
246 		kfd_mem_limit.system_mem_used -= size;
247 		kfd_mem_limit.ttm_mem_used -= size;
248 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
249 		WARN_ONCE(!adev,
250 			  "adev reference can't be null when alloc mem flags vram is set");
251 		if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
252 			goto release;
253 
254 		if (adev) {
255 			adev->kfd.vram_used[xcp_id] -= size;
256 			if (adev->gmc.is_app_apu) {
257 				adev->kfd.vram_used_aligned[xcp_id] -= size;
258 				kfd_mem_limit.system_mem_used -= size;
259 				kfd_mem_limit.ttm_mem_used -= size;
260 			} else {
261 				adev->kfd.vram_used_aligned[xcp_id] -=
262 					ALIGN(size, VRAM_AVAILABLITY_ALIGN);
263 			}
264 		}
265 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
266 		kfd_mem_limit.system_mem_used -= size;
267 	} else if (!(alloc_flag &
268 				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
269 				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
270 		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
271 		goto release;
272 	}
273 	WARN_ONCE(adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] < 0,
274 		  "KFD VRAM memory accounting unbalanced for xcp: %d", xcp_id);
275 	WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
276 		  "KFD TTM memory accounting unbalanced");
277 	WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
278 		  "KFD system memory accounting unbalanced");
279 
280 release:
281 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
282 }
283 
284 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
285 {
286 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
287 	u32 alloc_flags = bo->kfd_bo->alloc_flags;
288 	u64 size = amdgpu_bo_size(bo);
289 
290 	amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags,
291 					  bo->xcp_id);
292 
293 	kfree(bo->kfd_bo);
294 }
295 
296 /**
297  * create_dmamap_sg_bo() - Creates a amdgpu_bo object to reflect information
298  * about USERPTR or DOOREBELL or MMIO BO.
299  *
300  * @adev: Device for which dmamap BO is being created
301  * @mem: BO of peer device that is being DMA mapped. Provides parameters
302  *	 in building the dmamap BO
303  * @bo_out: Output parameter updated with handle of dmamap BO
304  */
305 static int
306 create_dmamap_sg_bo(struct amdgpu_device *adev,
307 		 struct kgd_mem *mem, struct amdgpu_bo **bo_out)
308 {
309 	struct drm_gem_object *gem_obj;
310 	int ret;
311 	uint64_t flags = 0;
312 
313 	ret = amdgpu_bo_reserve(mem->bo, false);
314 	if (ret)
315 		return ret;
316 
317 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)
318 		flags |= mem->bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
319 					AMDGPU_GEM_CREATE_UNCACHED);
320 
321 	ret = amdgpu_gem_object_create(adev, mem->bo->tbo.base.size, 1,
322 			AMDGPU_GEM_DOMAIN_CPU, AMDGPU_GEM_CREATE_PREEMPTIBLE | flags,
323 			ttm_bo_type_sg, mem->bo->tbo.base.resv, &gem_obj, 0);
324 
325 	amdgpu_bo_unreserve(mem->bo);
326 
327 	if (ret) {
328 		pr_err("Error in creating DMA mappable SG BO on domain: %d\n", ret);
329 		return -EINVAL;
330 	}
331 
332 	*bo_out = gem_to_amdgpu_bo(gem_obj);
333 	(*bo_out)->parent = amdgpu_bo_ref(mem->bo);
334 	return ret;
335 }
336 
337 /* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
338  *  reservation object.
339  *
340  * @bo: [IN] Remove eviction fence(s) from this BO
341  * @ef: [IN] This eviction fence is removed if it
342  *  is present in the shared list.
343  *
344  * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
345  */
346 static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
347 					struct amdgpu_amdkfd_fence *ef)
348 {
349 	struct dma_fence *replacement;
350 
351 	if (!ef)
352 		return -EINVAL;
353 
354 	/* TODO: Instead of block before we should use the fence of the page
355 	 * table update and TLB flush here directly.
356 	 */
357 	replacement = dma_fence_get_stub();
358 	dma_resv_replace_fences(bo->tbo.base.resv, ef->base.context,
359 				replacement, DMA_RESV_USAGE_BOOKKEEP);
360 	dma_fence_put(replacement);
361 	return 0;
362 }
363 
364 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
365 {
366 	struct amdgpu_bo *root = bo;
367 	struct amdgpu_vm_bo_base *vm_bo;
368 	struct amdgpu_vm *vm;
369 	struct amdkfd_process_info *info;
370 	struct amdgpu_amdkfd_fence *ef;
371 	int ret;
372 
373 	/* we can always get vm_bo from root PD bo.*/
374 	while (root->parent)
375 		root = root->parent;
376 
377 	vm_bo = root->vm_bo;
378 	if (!vm_bo)
379 		return 0;
380 
381 	vm = vm_bo->vm;
382 	if (!vm)
383 		return 0;
384 
385 	info = vm->process_info;
386 	if (!info || !info->eviction_fence)
387 		return 0;
388 
389 	ef = container_of(dma_fence_get(&info->eviction_fence->base),
390 			struct amdgpu_amdkfd_fence, base);
391 
392 	BUG_ON(!dma_resv_trylock(bo->tbo.base.resv));
393 	ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef);
394 	dma_resv_unlock(bo->tbo.base.resv);
395 
396 	dma_fence_put(&ef->base);
397 	return ret;
398 }
399 
400 static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
401 				     bool wait)
402 {
403 	struct ttm_operation_ctx ctx = { false, false };
404 	int ret;
405 
406 	if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
407 		 "Called with userptr BO"))
408 		return -EINVAL;
409 
410 	amdgpu_bo_placement_from_domain(bo, domain);
411 
412 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
413 	if (ret)
414 		goto validate_fail;
415 	if (wait)
416 		amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
417 
418 validate_fail:
419 	return ret;
420 }
421 
422 static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo)
423 {
424 	return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false);
425 }
426 
427 /* vm_validate_pt_pd_bos - Validate page table and directory BOs
428  *
429  * Page directories are not updated here because huge page handling
430  * during page table updates can invalidate page directory entries
431  * again. Page directories are only updated after updating page
432  * tables.
433  */
434 static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm)
435 {
436 	struct amdgpu_bo *pd = vm->root.bo;
437 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
438 	int ret;
439 
440 	ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate_vm_bo, NULL);
441 	if (ret) {
442 		pr_err("failed to validate PT BOs\n");
443 		return ret;
444 	}
445 
446 	vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo);
447 
448 	return 0;
449 }
450 
451 static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
452 {
453 	struct amdgpu_bo *pd = vm->root.bo;
454 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
455 	int ret;
456 
457 	ret = amdgpu_vm_update_pdes(adev, vm, false);
458 	if (ret)
459 		return ret;
460 
461 	return amdgpu_sync_fence(sync, vm->last_update);
462 }
463 
464 static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem)
465 {
466 	uint32_t mapping_flags = AMDGPU_VM_PAGE_READABLE |
467 				 AMDGPU_VM_MTYPE_DEFAULT;
468 
469 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE)
470 		mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
471 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE)
472 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
473 
474 	return amdgpu_gem_va_map_flags(adev, mapping_flags);
475 }
476 
477 /**
478  * create_sg_table() - Create an sg_table for a contiguous DMA addr range
479  * @addr: The starting address to point to
480  * @size: Size of memory area in bytes being pointed to
481  *
482  * Allocates an instance of sg_table and initializes it to point to memory
483  * area specified by input parameters. The address used to build is assumed
484  * to be DMA mapped, if needed.
485  *
486  * DOORBELL or MMIO BOs use only one scatterlist node in their sg_table
487  * because they are physically contiguous.
488  *
489  * Return: Initialized instance of SG Table or NULL
490  */
491 static struct sg_table *create_sg_table(uint64_t addr, uint32_t size)
492 {
493 	struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
494 
495 	if (!sg)
496 		return NULL;
497 	if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
498 		kfree(sg);
499 		return NULL;
500 	}
501 	sg_dma_address(sg->sgl) = addr;
502 	sg->sgl->length = size;
503 #ifdef CONFIG_NEED_SG_DMA_LENGTH
504 	sg->sgl->dma_length = size;
505 #endif
506 	return sg;
507 }
508 
509 static int
510 kfd_mem_dmamap_userptr(struct kgd_mem *mem,
511 		       struct kfd_mem_attachment *attachment)
512 {
513 	enum dma_data_direction direction =
514 		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
515 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
516 	struct ttm_operation_ctx ctx = {.interruptible = true};
517 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
518 	struct amdgpu_device *adev = attachment->adev;
519 	struct ttm_tt *src_ttm = mem->bo->tbo.ttm;
520 	struct ttm_tt *ttm = bo->tbo.ttm;
521 	int ret;
522 
523 	if (WARN_ON(ttm->num_pages != src_ttm->num_pages))
524 		return -EINVAL;
525 
526 	ttm->sg = kmalloc(sizeof(*ttm->sg), GFP_KERNEL);
527 	if (unlikely(!ttm->sg))
528 		return -ENOMEM;
529 
530 	/* Same sequence as in amdgpu_ttm_tt_pin_userptr */
531 	ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages,
532 					ttm->num_pages, 0,
533 					(u64)ttm->num_pages << PAGE_SHIFT,
534 					GFP_KERNEL);
535 	if (unlikely(ret))
536 		goto free_sg;
537 
538 	ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
539 	if (unlikely(ret))
540 		goto release_sg;
541 
542 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
543 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
544 	if (ret)
545 		goto unmap_sg;
546 
547 	return 0;
548 
549 unmap_sg:
550 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
551 release_sg:
552 	pr_err("DMA map userptr failed: %d\n", ret);
553 	sg_free_table(ttm->sg);
554 free_sg:
555 	kfree(ttm->sg);
556 	ttm->sg = NULL;
557 	return ret;
558 }
559 
560 static int
561 kfd_mem_dmamap_dmabuf(struct kfd_mem_attachment *attachment)
562 {
563 	struct ttm_operation_ctx ctx = {.interruptible = true};
564 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
565 	int ret;
566 
567 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
568 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
569 	if (ret)
570 		return ret;
571 
572 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
573 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
574 }
575 
576 /**
577  * kfd_mem_dmamap_sg_bo() - Create DMA mapped sg_table to access DOORBELL or MMIO BO
578  * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
579  * @attachment: Virtual address attachment of the BO on accessing device
580  *
581  * An access request from the device that owns DOORBELL does not require DMA mapping.
582  * This is because the request doesn't go through PCIe root complex i.e. it instead
583  * loops back. The need to DMA map arises only when accessing peer device's DOORBELL
584  *
585  * In contrast, all access requests for MMIO need to be DMA mapped without regard to
586  * device ownership. This is because access requests for MMIO go through PCIe root
587  * complex.
588  *
589  * This is accomplished in two steps:
590  *   - Obtain DMA mapped address of DOORBELL or MMIO memory that could be used
591  *         in updating requesting device's page table
592  *   - Signal TTM to mark memory pointed to by requesting device's BO as GPU
593  *         accessible. This allows an update of requesting device's page table
594  *         with entries associated with DOOREBELL or MMIO memory
595  *
596  * This method is invoked in the following contexts:
597  *   - Mapping of DOORBELL or MMIO BO of same or peer device
598  *   - Validating an evicted DOOREBELL or MMIO BO on device seeking access
599  *
600  * Return: ZERO if successful, NON-ZERO otherwise
601  */
602 static int
603 kfd_mem_dmamap_sg_bo(struct kgd_mem *mem,
604 		     struct kfd_mem_attachment *attachment)
605 {
606 	struct ttm_operation_ctx ctx = {.interruptible = true};
607 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
608 	struct amdgpu_device *adev = attachment->adev;
609 	struct ttm_tt *ttm = bo->tbo.ttm;
610 	enum dma_data_direction dir;
611 	dma_addr_t dma_addr;
612 	bool mmio;
613 	int ret;
614 
615 	/* Expect SG Table of dmapmap BO to be NULL */
616 	mmio = (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP);
617 	if (unlikely(ttm->sg)) {
618 		pr_err("SG Table of %d BO for peer device is UNEXPECTEDLY NON-NULL", mmio);
619 		return -EINVAL;
620 	}
621 
622 	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
623 			DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
624 	dma_addr = mem->bo->tbo.sg->sgl->dma_address;
625 	pr_debug("%d BO size: %d\n", mmio, mem->bo->tbo.sg->sgl->length);
626 	pr_debug("%d BO address before DMA mapping: %llx\n", mmio, dma_addr);
627 	dma_addr = dma_map_resource(adev->dev, dma_addr,
628 			mem->bo->tbo.sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
629 	ret = dma_mapping_error(adev->dev, dma_addr);
630 	if (unlikely(ret))
631 		return ret;
632 	pr_debug("%d BO address after DMA mapping: %llx\n", mmio, dma_addr);
633 
634 	ttm->sg = create_sg_table(dma_addr, mem->bo->tbo.sg->sgl->length);
635 	if (unlikely(!ttm->sg)) {
636 		ret = -ENOMEM;
637 		goto unmap_sg;
638 	}
639 
640 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
641 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
642 	if (unlikely(ret))
643 		goto free_sg;
644 
645 	return ret;
646 
647 free_sg:
648 	sg_free_table(ttm->sg);
649 	kfree(ttm->sg);
650 	ttm->sg = NULL;
651 unmap_sg:
652 	dma_unmap_resource(adev->dev, dma_addr, mem->bo->tbo.sg->sgl->length,
653 			   dir, DMA_ATTR_SKIP_CPU_SYNC);
654 	return ret;
655 }
656 
657 static int
658 kfd_mem_dmamap_attachment(struct kgd_mem *mem,
659 			  struct kfd_mem_attachment *attachment)
660 {
661 	switch (attachment->type) {
662 	case KFD_MEM_ATT_SHARED:
663 		return 0;
664 	case KFD_MEM_ATT_USERPTR:
665 		return kfd_mem_dmamap_userptr(mem, attachment);
666 	case KFD_MEM_ATT_DMABUF:
667 		return kfd_mem_dmamap_dmabuf(attachment);
668 	case KFD_MEM_ATT_SG:
669 		return kfd_mem_dmamap_sg_bo(mem, attachment);
670 	default:
671 		WARN_ON_ONCE(1);
672 	}
673 	return -EINVAL;
674 }
675 
676 static void
677 kfd_mem_dmaunmap_userptr(struct kgd_mem *mem,
678 			 struct kfd_mem_attachment *attachment)
679 {
680 	enum dma_data_direction direction =
681 		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
682 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
683 	struct ttm_operation_ctx ctx = {.interruptible = false};
684 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
685 	struct amdgpu_device *adev = attachment->adev;
686 	struct ttm_tt *ttm = bo->tbo.ttm;
687 
688 	if (unlikely(!ttm->sg))
689 		return;
690 
691 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
692 	ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
693 
694 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
695 	sg_free_table(ttm->sg);
696 	kfree(ttm->sg);
697 	ttm->sg = NULL;
698 }
699 
700 static void
701 kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment)
702 {
703 	/* This is a no-op. We don't want to trigger eviction fences when
704 	 * unmapping DMABufs. Therefore the invalidation (moving to system
705 	 * domain) is done in kfd_mem_dmamap_dmabuf.
706 	 */
707 }
708 
709 /**
710  * kfd_mem_dmaunmap_sg_bo() - Free DMA mapped sg_table of DOORBELL or MMIO BO
711  * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
712  * @attachment: Virtual address attachment of the BO on accessing device
713  *
714  * The method performs following steps:
715  *   - Signal TTM to mark memory pointed to by BO as GPU inaccessible
716  *   - Free SG Table that is used to encapsulate DMA mapped memory of
717  *          peer device's DOORBELL or MMIO memory
718  *
719  * This method is invoked in the following contexts:
720  *     UNMapping of DOORBELL or MMIO BO on a device having access to its memory
721  *     Eviction of DOOREBELL or MMIO BO on device having access to its memory
722  *
723  * Return: void
724  */
725 static void
726 kfd_mem_dmaunmap_sg_bo(struct kgd_mem *mem,
727 		       struct kfd_mem_attachment *attachment)
728 {
729 	struct ttm_operation_ctx ctx = {.interruptible = true};
730 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
731 	struct amdgpu_device *adev = attachment->adev;
732 	struct ttm_tt *ttm = bo->tbo.ttm;
733 	enum dma_data_direction dir;
734 
735 	if (unlikely(!ttm->sg)) {
736 		pr_err("SG Table of BO is UNEXPECTEDLY NULL");
737 		return;
738 	}
739 
740 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
741 	ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
742 
743 	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
744 				DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
745 	dma_unmap_resource(adev->dev, ttm->sg->sgl->dma_address,
746 			ttm->sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
747 	sg_free_table(ttm->sg);
748 	kfree(ttm->sg);
749 	ttm->sg = NULL;
750 	bo->tbo.sg = NULL;
751 }
752 
753 static void
754 kfd_mem_dmaunmap_attachment(struct kgd_mem *mem,
755 			    struct kfd_mem_attachment *attachment)
756 {
757 	switch (attachment->type) {
758 	case KFD_MEM_ATT_SHARED:
759 		break;
760 	case KFD_MEM_ATT_USERPTR:
761 		kfd_mem_dmaunmap_userptr(mem, attachment);
762 		break;
763 	case KFD_MEM_ATT_DMABUF:
764 		kfd_mem_dmaunmap_dmabuf(attachment);
765 		break;
766 	case KFD_MEM_ATT_SG:
767 		kfd_mem_dmaunmap_sg_bo(mem, attachment);
768 		break;
769 	default:
770 		WARN_ON_ONCE(1);
771 	}
772 }
773 
774 static int kfd_mem_export_dmabuf(struct kgd_mem *mem)
775 {
776 	if (!mem->dmabuf) {
777 		struct dma_buf *ret = amdgpu_gem_prime_export(
778 			&mem->bo->tbo.base,
779 			mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
780 				DRM_RDWR : 0);
781 		if (IS_ERR(ret))
782 			return PTR_ERR(ret);
783 		mem->dmabuf = ret;
784 	}
785 
786 	return 0;
787 }
788 
789 static int
790 kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem,
791 		      struct amdgpu_bo **bo)
792 {
793 	struct drm_gem_object *gobj;
794 	int ret;
795 
796 	ret = kfd_mem_export_dmabuf(mem);
797 	if (ret)
798 		return ret;
799 
800 	gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf);
801 	if (IS_ERR(gobj))
802 		return PTR_ERR(gobj);
803 
804 	*bo = gem_to_amdgpu_bo(gobj);
805 	(*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE;
806 
807 	return 0;
808 }
809 
810 /* kfd_mem_attach - Add a BO to a VM
811  *
812  * Everything that needs to bo done only once when a BO is first added
813  * to a VM. It can later be mapped and unmapped many times without
814  * repeating these steps.
815  *
816  * 0. Create BO for DMA mapping, if needed
817  * 1. Allocate and initialize BO VA entry data structure
818  * 2. Add BO to the VM
819  * 3. Determine ASIC-specific PTE flags
820  * 4. Alloc page tables and directories if needed
821  * 4a.  Validate new page tables and directories
822  */
823 static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
824 		struct amdgpu_vm *vm, bool is_aql)
825 {
826 	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
827 	unsigned long bo_size = mem->bo->tbo.base.size;
828 	uint64_t va = mem->va;
829 	struct kfd_mem_attachment *attachment[2] = {NULL, NULL};
830 	struct amdgpu_bo *bo[2] = {NULL, NULL};
831 	bool same_hive = false;
832 	int i, ret;
833 
834 	if (!va) {
835 		pr_err("Invalid VA when adding BO to VM\n");
836 		return -EINVAL;
837 	}
838 
839 	/* Determine access to VRAM, MMIO and DOORBELL BOs of peer devices
840 	 *
841 	 * The access path of MMIO and DOORBELL BOs of is always over PCIe.
842 	 * In contrast the access path of VRAM BOs depens upon the type of
843 	 * link that connects the peer device. Access over PCIe is allowed
844 	 * if peer device has large BAR. In contrast, access over xGMI is
845 	 * allowed for both small and large BAR configurations of peer device
846 	 */
847 	if ((adev != bo_adev && !adev->gmc.is_app_apu) &&
848 	    ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) ||
849 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) ||
850 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
851 		if (mem->domain == AMDGPU_GEM_DOMAIN_VRAM)
852 			same_hive = amdgpu_xgmi_same_hive(adev, bo_adev);
853 		if (!same_hive && !amdgpu_device_is_peer_accessible(bo_adev, adev))
854 			return -EINVAL;
855 	}
856 
857 	for (i = 0; i <= is_aql; i++) {
858 		attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL);
859 		if (unlikely(!attachment[i])) {
860 			ret = -ENOMEM;
861 			goto unwind;
862 		}
863 
864 		pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
865 			 va + bo_size, vm);
866 
867 		if ((adev == bo_adev && !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) ||
868 		    (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && reuse_dmamap(adev, bo_adev)) ||
869 		    (mem->domain == AMDGPU_GEM_DOMAIN_GTT && reuse_dmamap(adev, bo_adev)) ||
870 		    same_hive) {
871 			/* Mappings on the local GPU, or VRAM mappings in the
872 			 * local hive, or userptr, or GTT mapping can reuse dma map
873 			 * address space share the original BO
874 			 */
875 			attachment[i]->type = KFD_MEM_ATT_SHARED;
876 			bo[i] = mem->bo;
877 			drm_gem_object_get(&bo[i]->tbo.base);
878 		} else if (i > 0) {
879 			/* Multiple mappings on the same GPU share the BO */
880 			attachment[i]->type = KFD_MEM_ATT_SHARED;
881 			bo[i] = bo[0];
882 			drm_gem_object_get(&bo[i]->tbo.base);
883 		} else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
884 			/* Create an SG BO to DMA-map userptrs on other GPUs */
885 			attachment[i]->type = KFD_MEM_ATT_USERPTR;
886 			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
887 			if (ret)
888 				goto unwind;
889 		/* Handle DOORBELL BOs of peer devices and MMIO BOs of local and peer devices */
890 		} else if (mem->bo->tbo.type == ttm_bo_type_sg) {
891 			WARN_ONCE(!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL ||
892 				    mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP),
893 				  "Handing invalid SG BO in ATTACH request");
894 			attachment[i]->type = KFD_MEM_ATT_SG;
895 			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
896 			if (ret)
897 				goto unwind;
898 		/* Enable acces to GTT and VRAM BOs of peer devices */
899 		} else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT ||
900 			   mem->domain == AMDGPU_GEM_DOMAIN_VRAM) {
901 			attachment[i]->type = KFD_MEM_ATT_DMABUF;
902 			ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]);
903 			if (ret)
904 				goto unwind;
905 			pr_debug("Employ DMABUF mechanism to enable peer GPU access\n");
906 		} else {
907 			WARN_ONCE(true, "Handling invalid ATTACH request");
908 			ret = -EINVAL;
909 			goto unwind;
910 		}
911 
912 		/* Add BO to VM internal data structures */
913 		ret = amdgpu_bo_reserve(bo[i], false);
914 		if (ret) {
915 			pr_debug("Unable to reserve BO during memory attach");
916 			goto unwind;
917 		}
918 		attachment[i]->bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]);
919 		amdgpu_bo_unreserve(bo[i]);
920 		if (unlikely(!attachment[i]->bo_va)) {
921 			ret = -ENOMEM;
922 			pr_err("Failed to add BO object to VM. ret == %d\n",
923 			       ret);
924 			goto unwind;
925 		}
926 		attachment[i]->va = va;
927 		attachment[i]->pte_flags = get_pte_flags(adev, mem);
928 		attachment[i]->adev = adev;
929 		list_add(&attachment[i]->list, &mem->attachments);
930 
931 		va += bo_size;
932 	}
933 
934 	return 0;
935 
936 unwind:
937 	for (; i >= 0; i--) {
938 		if (!attachment[i])
939 			continue;
940 		if (attachment[i]->bo_va) {
941 			amdgpu_bo_reserve(bo[i], true);
942 			amdgpu_vm_bo_del(adev, attachment[i]->bo_va);
943 			amdgpu_bo_unreserve(bo[i]);
944 			list_del(&attachment[i]->list);
945 		}
946 		if (bo[i])
947 			drm_gem_object_put(&bo[i]->tbo.base);
948 		kfree(attachment[i]);
949 	}
950 	return ret;
951 }
952 
953 static void kfd_mem_detach(struct kfd_mem_attachment *attachment)
954 {
955 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
956 
957 	pr_debug("\t remove VA 0x%llx in entry %p\n",
958 			attachment->va, attachment);
959 	amdgpu_vm_bo_del(attachment->adev, attachment->bo_va);
960 	drm_gem_object_put(&bo->tbo.base);
961 	list_del(&attachment->list);
962 	kfree(attachment);
963 }
964 
965 static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
966 				struct amdkfd_process_info *process_info,
967 				bool userptr)
968 {
969 	mutex_lock(&process_info->lock);
970 	if (userptr)
971 		list_add_tail(&mem->validate_list,
972 			      &process_info->userptr_valid_list);
973 	else
974 		list_add_tail(&mem->validate_list, &process_info->kfd_bo_list);
975 	mutex_unlock(&process_info->lock);
976 }
977 
978 static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
979 		struct amdkfd_process_info *process_info)
980 {
981 	mutex_lock(&process_info->lock);
982 	list_del(&mem->validate_list);
983 	mutex_unlock(&process_info->lock);
984 }
985 
986 /* Initializes user pages. It registers the MMU notifier and validates
987  * the userptr BO in the GTT domain.
988  *
989  * The BO must already be on the userptr_valid_list. Otherwise an
990  * eviction and restore may happen that leaves the new BO unmapped
991  * with the user mode queues running.
992  *
993  * Takes the process_info->lock to protect against concurrent restore
994  * workers.
995  *
996  * Returns 0 for success, negative errno for errors.
997  */
998 static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr,
999 			   bool criu_resume)
1000 {
1001 	struct amdkfd_process_info *process_info = mem->process_info;
1002 	struct amdgpu_bo *bo = mem->bo;
1003 	struct ttm_operation_ctx ctx = { true, false };
1004 	struct hmm_range *range;
1005 	int ret = 0;
1006 
1007 	mutex_lock(&process_info->lock);
1008 
1009 	ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
1010 	if (ret) {
1011 		pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
1012 		goto out;
1013 	}
1014 
1015 	ret = amdgpu_hmm_register(bo, user_addr);
1016 	if (ret) {
1017 		pr_err("%s: Failed to register MMU notifier: %d\n",
1018 		       __func__, ret);
1019 		goto out;
1020 	}
1021 
1022 	if (criu_resume) {
1023 		/*
1024 		 * During a CRIU restore operation, the userptr buffer objects
1025 		 * will be validated in the restore_userptr_work worker at a
1026 		 * later stage when it is scheduled by another ioctl called by
1027 		 * CRIU master process for the target pid for restore.
1028 		 */
1029 		mutex_lock(&process_info->notifier_lock);
1030 		mem->invalid++;
1031 		mutex_unlock(&process_info->notifier_lock);
1032 		mutex_unlock(&process_info->lock);
1033 		return 0;
1034 	}
1035 
1036 	ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, &range);
1037 	if (ret) {
1038 		pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
1039 		goto unregister_out;
1040 	}
1041 
1042 	ret = amdgpu_bo_reserve(bo, true);
1043 	if (ret) {
1044 		pr_err("%s: Failed to reserve BO\n", __func__);
1045 		goto release_out;
1046 	}
1047 	amdgpu_bo_placement_from_domain(bo, mem->domain);
1048 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1049 	if (ret)
1050 		pr_err("%s: failed to validate BO\n", __func__);
1051 	amdgpu_bo_unreserve(bo);
1052 
1053 release_out:
1054 	amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range);
1055 unregister_out:
1056 	if (ret)
1057 		amdgpu_hmm_unregister(bo);
1058 out:
1059 	mutex_unlock(&process_info->lock);
1060 	return ret;
1061 }
1062 
1063 /* Reserving a BO and its page table BOs must happen atomically to
1064  * avoid deadlocks. Some operations update multiple VMs at once. Track
1065  * all the reservation info in a context structure. Optionally a sync
1066  * object can track VM updates.
1067  */
1068 struct bo_vm_reservation_context {
1069 	/* DRM execution context for the reservation */
1070 	struct drm_exec exec;
1071 	/* Number of VMs reserved */
1072 	unsigned int n_vms;
1073 	/* Pointer to sync object */
1074 	struct amdgpu_sync *sync;
1075 };
1076 
1077 enum bo_vm_match {
1078 	BO_VM_NOT_MAPPED = 0,	/* Match VMs where a BO is not mapped */
1079 	BO_VM_MAPPED,		/* Match VMs where a BO is mapped     */
1080 	BO_VM_ALL,		/* Match all VMs a BO was added to    */
1081 };
1082 
1083 /**
1084  * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
1085  * @mem: KFD BO structure.
1086  * @vm: the VM to reserve.
1087  * @ctx: the struct that will be used in unreserve_bo_and_vms().
1088  */
1089 static int reserve_bo_and_vm(struct kgd_mem *mem,
1090 			      struct amdgpu_vm *vm,
1091 			      struct bo_vm_reservation_context *ctx)
1092 {
1093 	struct amdgpu_bo *bo = mem->bo;
1094 	int ret;
1095 
1096 	WARN_ON(!vm);
1097 
1098 	ctx->n_vms = 1;
1099 	ctx->sync = &mem->sync;
1100 	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT);
1101 	drm_exec_until_all_locked(&ctx->exec) {
1102 		ret = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
1103 		drm_exec_retry_on_contention(&ctx->exec);
1104 		if (unlikely(ret))
1105 			goto error;
1106 
1107 		ret = drm_exec_lock_obj(&ctx->exec, &bo->tbo.base);
1108 		drm_exec_retry_on_contention(&ctx->exec);
1109 		if (unlikely(ret))
1110 			goto error;
1111 	}
1112 	return 0;
1113 
1114 error:
1115 	pr_err("Failed to reserve buffers in ttm.\n");
1116 	drm_exec_fini(&ctx->exec);
1117 	return ret;
1118 }
1119 
1120 /**
1121  * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
1122  * @mem: KFD BO structure.
1123  * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
1124  * is used. Otherwise, a single VM associated with the BO.
1125  * @map_type: the mapping status that will be used to filter the VMs.
1126  * @ctx: the struct that will be used in unreserve_bo_and_vms().
1127  *
1128  * Returns 0 for success, negative for failure.
1129  */
1130 static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
1131 				struct amdgpu_vm *vm, enum bo_vm_match map_type,
1132 				struct bo_vm_reservation_context *ctx)
1133 {
1134 	struct kfd_mem_attachment *entry;
1135 	struct amdgpu_bo *bo = mem->bo;
1136 	int ret;
1137 
1138 	ctx->sync = &mem->sync;
1139 	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT);
1140 	drm_exec_until_all_locked(&ctx->exec) {
1141 		ctx->n_vms = 0;
1142 		list_for_each_entry(entry, &mem->attachments, list) {
1143 			if ((vm && vm != entry->bo_va->base.vm) ||
1144 				(entry->is_mapped != map_type
1145 				&& map_type != BO_VM_ALL))
1146 				continue;
1147 
1148 			ret = amdgpu_vm_lock_pd(entry->bo_va->base.vm,
1149 						&ctx->exec, 2);
1150 			drm_exec_retry_on_contention(&ctx->exec);
1151 			if (unlikely(ret))
1152 				goto error;
1153 			++ctx->n_vms;
1154 		}
1155 
1156 		ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
1157 		drm_exec_retry_on_contention(&ctx->exec);
1158 		if (unlikely(ret))
1159 			goto error;
1160 	}
1161 	return 0;
1162 
1163 error:
1164 	pr_err("Failed to reserve buffers in ttm.\n");
1165 	drm_exec_fini(&ctx->exec);
1166 	return ret;
1167 }
1168 
1169 /**
1170  * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1171  * @ctx: Reservation context to unreserve
1172  * @wait: Optionally wait for a sync object representing pending VM updates
1173  * @intr: Whether the wait is interruptible
1174  *
1175  * Also frees any resources allocated in
1176  * reserve_bo_and_(cond_)vm(s). Returns the status from
1177  * amdgpu_sync_wait.
1178  */
1179 static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
1180 				 bool wait, bool intr)
1181 {
1182 	int ret = 0;
1183 
1184 	if (wait)
1185 		ret = amdgpu_sync_wait(ctx->sync, intr);
1186 
1187 	drm_exec_fini(&ctx->exec);
1188 	ctx->sync = NULL;
1189 	return ret;
1190 }
1191 
1192 static void unmap_bo_from_gpuvm(struct kgd_mem *mem,
1193 				struct kfd_mem_attachment *entry,
1194 				struct amdgpu_sync *sync)
1195 {
1196 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1197 	struct amdgpu_device *adev = entry->adev;
1198 	struct amdgpu_vm *vm = bo_va->base.vm;
1199 
1200 	amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
1201 
1202 	amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
1203 
1204 	amdgpu_sync_fence(sync, bo_va->last_pt_update);
1205 
1206 	kfd_mem_dmaunmap_attachment(mem, entry);
1207 }
1208 
1209 static int update_gpuvm_pte(struct kgd_mem *mem,
1210 			    struct kfd_mem_attachment *entry,
1211 			    struct amdgpu_sync *sync)
1212 {
1213 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1214 	struct amdgpu_device *adev = entry->adev;
1215 	int ret;
1216 
1217 	ret = kfd_mem_dmamap_attachment(mem, entry);
1218 	if (ret)
1219 		return ret;
1220 
1221 	/* Update the page tables  */
1222 	ret = amdgpu_vm_bo_update(adev, bo_va, false);
1223 	if (ret) {
1224 		pr_err("amdgpu_vm_bo_update failed\n");
1225 		return ret;
1226 	}
1227 
1228 	return amdgpu_sync_fence(sync, bo_va->last_pt_update);
1229 }
1230 
1231 static int map_bo_to_gpuvm(struct kgd_mem *mem,
1232 			   struct kfd_mem_attachment *entry,
1233 			   struct amdgpu_sync *sync,
1234 			   bool no_update_pte)
1235 {
1236 	int ret;
1237 
1238 	/* Set virtual address for the allocation */
1239 	ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0,
1240 			       amdgpu_bo_size(entry->bo_va->base.bo),
1241 			       entry->pte_flags);
1242 	if (ret) {
1243 		pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
1244 				entry->va, ret);
1245 		return ret;
1246 	}
1247 
1248 	if (no_update_pte)
1249 		return 0;
1250 
1251 	ret = update_gpuvm_pte(mem, entry, sync);
1252 	if (ret) {
1253 		pr_err("update_gpuvm_pte() failed\n");
1254 		goto update_gpuvm_pte_failed;
1255 	}
1256 
1257 	return 0;
1258 
1259 update_gpuvm_pte_failed:
1260 	unmap_bo_from_gpuvm(mem, entry, sync);
1261 	return ret;
1262 }
1263 
1264 static int process_validate_vms(struct amdkfd_process_info *process_info)
1265 {
1266 	struct amdgpu_vm *peer_vm;
1267 	int ret;
1268 
1269 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1270 			    vm_list_node) {
1271 		ret = vm_validate_pt_pd_bos(peer_vm);
1272 		if (ret)
1273 			return ret;
1274 	}
1275 
1276 	return 0;
1277 }
1278 
1279 static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
1280 				 struct amdgpu_sync *sync)
1281 {
1282 	struct amdgpu_vm *peer_vm;
1283 	int ret;
1284 
1285 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1286 			    vm_list_node) {
1287 		struct amdgpu_bo *pd = peer_vm->root.bo;
1288 
1289 		ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
1290 				       AMDGPU_SYNC_NE_OWNER,
1291 				       AMDGPU_FENCE_OWNER_KFD);
1292 		if (ret)
1293 			return ret;
1294 	}
1295 
1296 	return 0;
1297 }
1298 
1299 static int process_update_pds(struct amdkfd_process_info *process_info,
1300 			      struct amdgpu_sync *sync)
1301 {
1302 	struct amdgpu_vm *peer_vm;
1303 	int ret;
1304 
1305 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1306 			    vm_list_node) {
1307 		ret = vm_update_pds(peer_vm, sync);
1308 		if (ret)
1309 			return ret;
1310 	}
1311 
1312 	return 0;
1313 }
1314 
1315 static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
1316 		       struct dma_fence **ef)
1317 {
1318 	struct amdkfd_process_info *info = NULL;
1319 	int ret;
1320 
1321 	if (!*process_info) {
1322 		info = kzalloc(sizeof(*info), GFP_KERNEL);
1323 		if (!info)
1324 			return -ENOMEM;
1325 
1326 		mutex_init(&info->lock);
1327 		mutex_init(&info->notifier_lock);
1328 		INIT_LIST_HEAD(&info->vm_list_head);
1329 		INIT_LIST_HEAD(&info->kfd_bo_list);
1330 		INIT_LIST_HEAD(&info->userptr_valid_list);
1331 		INIT_LIST_HEAD(&info->userptr_inval_list);
1332 
1333 		info->eviction_fence =
1334 			amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
1335 						   current->mm,
1336 						   NULL);
1337 		if (!info->eviction_fence) {
1338 			pr_err("Failed to create eviction fence\n");
1339 			ret = -ENOMEM;
1340 			goto create_evict_fence_fail;
1341 		}
1342 
1343 		info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
1344 		INIT_DELAYED_WORK(&info->restore_userptr_work,
1345 				  amdgpu_amdkfd_restore_userptr_worker);
1346 
1347 		*process_info = info;
1348 		*ef = dma_fence_get(&info->eviction_fence->base);
1349 	}
1350 
1351 	vm->process_info = *process_info;
1352 
1353 	/* Validate page directory and attach eviction fence */
1354 	ret = amdgpu_bo_reserve(vm->root.bo, true);
1355 	if (ret)
1356 		goto reserve_pd_fail;
1357 	ret = vm_validate_pt_pd_bos(vm);
1358 	if (ret) {
1359 		pr_err("validate_pt_pd_bos() failed\n");
1360 		goto validate_pd_fail;
1361 	}
1362 	ret = amdgpu_bo_sync_wait(vm->root.bo,
1363 				  AMDGPU_FENCE_OWNER_KFD, false);
1364 	if (ret)
1365 		goto wait_pd_fail;
1366 	ret = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1);
1367 	if (ret)
1368 		goto reserve_shared_fail;
1369 	dma_resv_add_fence(vm->root.bo->tbo.base.resv,
1370 			   &vm->process_info->eviction_fence->base,
1371 			   DMA_RESV_USAGE_BOOKKEEP);
1372 	amdgpu_bo_unreserve(vm->root.bo);
1373 
1374 	/* Update process info */
1375 	mutex_lock(&vm->process_info->lock);
1376 	list_add_tail(&vm->vm_list_node,
1377 			&(vm->process_info->vm_list_head));
1378 	vm->process_info->n_vms++;
1379 	mutex_unlock(&vm->process_info->lock);
1380 
1381 	return 0;
1382 
1383 reserve_shared_fail:
1384 wait_pd_fail:
1385 validate_pd_fail:
1386 	amdgpu_bo_unreserve(vm->root.bo);
1387 reserve_pd_fail:
1388 	vm->process_info = NULL;
1389 	if (info) {
1390 		/* Two fence references: one in info and one in *ef */
1391 		dma_fence_put(&info->eviction_fence->base);
1392 		dma_fence_put(*ef);
1393 		*ef = NULL;
1394 		*process_info = NULL;
1395 		put_pid(info->pid);
1396 create_evict_fence_fail:
1397 		mutex_destroy(&info->lock);
1398 		mutex_destroy(&info->notifier_lock);
1399 		kfree(info);
1400 	}
1401 	return ret;
1402 }
1403 
1404 /**
1405  * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
1406  * @bo: Handle of buffer object being pinned
1407  * @domain: Domain into which BO should be pinned
1408  *
1409  *   - USERPTR BOs are UNPINNABLE and will return error
1410  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1411  *     PIN count incremented. It is valid to PIN a BO multiple times
1412  *
1413  * Return: ZERO if successful in pinning, Non-Zero in case of error.
1414  */
1415 static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain)
1416 {
1417 	int ret = 0;
1418 
1419 	ret = amdgpu_bo_reserve(bo, false);
1420 	if (unlikely(ret))
1421 		return ret;
1422 
1423 	ret = amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1424 	if (ret)
1425 		pr_err("Error in Pinning BO to domain: %d\n", domain);
1426 
1427 	amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
1428 	amdgpu_bo_unreserve(bo);
1429 
1430 	return ret;
1431 }
1432 
1433 /**
1434  * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
1435  * @bo: Handle of buffer object being unpinned
1436  *
1437  *   - Is a illegal request for USERPTR BOs and is ignored
1438  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1439  *     PIN count decremented. Calls to UNPIN must balance calls to PIN
1440  */
1441 static void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo)
1442 {
1443 	int ret = 0;
1444 
1445 	ret = amdgpu_bo_reserve(bo, false);
1446 	if (unlikely(ret))
1447 		return;
1448 
1449 	amdgpu_bo_unpin(bo);
1450 	amdgpu_bo_unreserve(bo);
1451 }
1452 
1453 int amdgpu_amdkfd_gpuvm_set_vm_pasid(struct amdgpu_device *adev,
1454 				     struct amdgpu_vm *avm, u32 pasid)
1455 
1456 {
1457 	int ret;
1458 
1459 	/* Free the original amdgpu allocated pasid,
1460 	 * will be replaced with kfd allocated pasid.
1461 	 */
1462 	if (avm->pasid) {
1463 		amdgpu_pasid_free(avm->pasid);
1464 		amdgpu_vm_set_pasid(adev, avm, 0);
1465 	}
1466 
1467 	ret = amdgpu_vm_set_pasid(adev, avm, pasid);
1468 	if (ret)
1469 		return ret;
1470 
1471 	return 0;
1472 }
1473 
1474 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
1475 					   struct amdgpu_vm *avm,
1476 					   void **process_info,
1477 					   struct dma_fence **ef)
1478 {
1479 	int ret;
1480 
1481 	/* Already a compute VM? */
1482 	if (avm->process_info)
1483 		return -EINVAL;
1484 
1485 	/* Convert VM into a compute VM */
1486 	ret = amdgpu_vm_make_compute(adev, avm);
1487 	if (ret)
1488 		return ret;
1489 
1490 	/* Initialize KFD part of the VM and process info */
1491 	ret = init_kfd_vm(avm, process_info, ef);
1492 	if (ret)
1493 		return ret;
1494 
1495 	amdgpu_vm_set_task_info(avm);
1496 
1497 	return 0;
1498 }
1499 
1500 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1501 				    struct amdgpu_vm *vm)
1502 {
1503 	struct amdkfd_process_info *process_info = vm->process_info;
1504 
1505 	if (!process_info)
1506 		return;
1507 
1508 	/* Update process info */
1509 	mutex_lock(&process_info->lock);
1510 	process_info->n_vms--;
1511 	list_del(&vm->vm_list_node);
1512 	mutex_unlock(&process_info->lock);
1513 
1514 	vm->process_info = NULL;
1515 
1516 	/* Release per-process resources when last compute VM is destroyed */
1517 	if (!process_info->n_vms) {
1518 		WARN_ON(!list_empty(&process_info->kfd_bo_list));
1519 		WARN_ON(!list_empty(&process_info->userptr_valid_list));
1520 		WARN_ON(!list_empty(&process_info->userptr_inval_list));
1521 
1522 		dma_fence_put(&process_info->eviction_fence->base);
1523 		cancel_delayed_work_sync(&process_info->restore_userptr_work);
1524 		put_pid(process_info->pid);
1525 		mutex_destroy(&process_info->lock);
1526 		mutex_destroy(&process_info->notifier_lock);
1527 		kfree(process_info);
1528 	}
1529 }
1530 
1531 void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev,
1532 					    void *drm_priv)
1533 {
1534 	struct amdgpu_vm *avm;
1535 
1536 	if (WARN_ON(!adev || !drm_priv))
1537 		return;
1538 
1539 	avm = drm_priv_to_vm(drm_priv);
1540 
1541 	pr_debug("Releasing process vm %p\n", avm);
1542 
1543 	/* The original pasid of amdgpu vm has already been
1544 	 * released during making a amdgpu vm to a compute vm
1545 	 * The current pasid is managed by kfd and will be
1546 	 * released on kfd process destroy. Set amdgpu pasid
1547 	 * to 0 to avoid duplicate release.
1548 	 */
1549 	amdgpu_vm_release_compute(adev, avm);
1550 }
1551 
1552 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv)
1553 {
1554 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1555 	struct amdgpu_bo *pd = avm->root.bo;
1556 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1557 
1558 	if (adev->asic_type < CHIP_VEGA10)
1559 		return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1560 	return avm->pd_phys_addr;
1561 }
1562 
1563 void amdgpu_amdkfd_block_mmu_notifications(void *p)
1564 {
1565 	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1566 
1567 	mutex_lock(&pinfo->lock);
1568 	WRITE_ONCE(pinfo->block_mmu_notifications, true);
1569 	mutex_unlock(&pinfo->lock);
1570 }
1571 
1572 int amdgpu_amdkfd_criu_resume(void *p)
1573 {
1574 	int ret = 0;
1575 	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1576 
1577 	mutex_lock(&pinfo->lock);
1578 	pr_debug("scheduling work\n");
1579 	mutex_lock(&pinfo->notifier_lock);
1580 	pinfo->evicted_bos++;
1581 	mutex_unlock(&pinfo->notifier_lock);
1582 	if (!READ_ONCE(pinfo->block_mmu_notifications)) {
1583 		ret = -EINVAL;
1584 		goto out_unlock;
1585 	}
1586 	WRITE_ONCE(pinfo->block_mmu_notifications, false);
1587 	schedule_delayed_work(&pinfo->restore_userptr_work, 0);
1588 
1589 out_unlock:
1590 	mutex_unlock(&pinfo->lock);
1591 	return ret;
1592 }
1593 
1594 size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
1595 					  uint8_t xcp_id)
1596 {
1597 	uint64_t reserved_for_pt =
1598 		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
1599 	ssize_t available;
1600 	uint64_t vram_available, system_mem_available, ttm_mem_available;
1601 
1602 	spin_lock(&kfd_mem_limit.mem_limit_lock);
1603 	vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id)
1604 		- adev->kfd.vram_used_aligned[xcp_id]
1605 		- atomic64_read(&adev->vram_pin_size)
1606 		- reserved_for_pt;
1607 
1608 	if (adev->gmc.is_app_apu) {
1609 		system_mem_available = no_system_mem_limit ?
1610 					kfd_mem_limit.max_system_mem_limit :
1611 					kfd_mem_limit.max_system_mem_limit -
1612 					kfd_mem_limit.system_mem_used;
1613 
1614 		ttm_mem_available = kfd_mem_limit.max_ttm_mem_limit -
1615 				kfd_mem_limit.ttm_mem_used;
1616 
1617 		available = min3(system_mem_available, ttm_mem_available,
1618 				 vram_available);
1619 		available = ALIGN_DOWN(available, PAGE_SIZE);
1620 	} else {
1621 		available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN);
1622 	}
1623 
1624 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
1625 
1626 	if (available < 0)
1627 		available = 0;
1628 
1629 	return available;
1630 }
1631 
1632 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1633 		struct amdgpu_device *adev, uint64_t va, uint64_t size,
1634 		void *drm_priv, struct kgd_mem **mem,
1635 		uint64_t *offset, uint32_t flags, bool criu_resume)
1636 {
1637 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1638 	struct amdgpu_fpriv *fpriv = container_of(avm, struct amdgpu_fpriv, vm);
1639 	enum ttm_bo_type bo_type = ttm_bo_type_device;
1640 	struct sg_table *sg = NULL;
1641 	uint64_t user_addr = 0;
1642 	struct amdgpu_bo *bo;
1643 	struct drm_gem_object *gobj = NULL;
1644 	u32 domain, alloc_domain;
1645 	uint64_t aligned_size;
1646 	int8_t xcp_id = -1;
1647 	u64 alloc_flags;
1648 	int ret;
1649 
1650 	/*
1651 	 * Check on which domain to allocate BO
1652 	 */
1653 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1654 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1655 
1656 		if (adev->gmc.is_app_apu) {
1657 			domain = AMDGPU_GEM_DOMAIN_GTT;
1658 			alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1659 			alloc_flags = 0;
1660 		} else {
1661 			alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1662 			alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1663 			AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0;
1664 		}
1665 		xcp_id = fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION ?
1666 					0 : fpriv->xcp_id;
1667 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1668 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1669 		alloc_flags = 0;
1670 	} else {
1671 		domain = AMDGPU_GEM_DOMAIN_GTT;
1672 		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1673 		alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE;
1674 
1675 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1676 			if (!offset || !*offset)
1677 				return -EINVAL;
1678 			user_addr = untagged_addr(*offset);
1679 		} else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1680 				    KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1681 			bo_type = ttm_bo_type_sg;
1682 			if (size > UINT_MAX)
1683 				return -EINVAL;
1684 			sg = create_sg_table(*offset, size);
1685 			if (!sg)
1686 				return -ENOMEM;
1687 		} else {
1688 			return -EINVAL;
1689 		}
1690 	}
1691 
1692 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT)
1693 		alloc_flags |= AMDGPU_GEM_CREATE_COHERENT;
1694 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENT)
1695 		alloc_flags |= AMDGPU_GEM_CREATE_EXT_COHERENT;
1696 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED)
1697 		alloc_flags |= AMDGPU_GEM_CREATE_UNCACHED;
1698 
1699 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1700 	if (!*mem) {
1701 		ret = -ENOMEM;
1702 		goto err;
1703 	}
1704 	INIT_LIST_HEAD(&(*mem)->attachments);
1705 	mutex_init(&(*mem)->lock);
1706 	(*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1707 
1708 	/* Workaround for AQL queue wraparound bug. Map the same
1709 	 * memory twice. That means we only actually allocate half
1710 	 * the memory.
1711 	 */
1712 	if ((*mem)->aql_queue)
1713 		size >>= 1;
1714 	aligned_size = PAGE_ALIGN(size);
1715 
1716 	(*mem)->alloc_flags = flags;
1717 
1718 	amdgpu_sync_create(&(*mem)->sync);
1719 
1720 	ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags,
1721 					      xcp_id);
1722 	if (ret) {
1723 		pr_debug("Insufficient memory\n");
1724 		goto err_reserve_limit;
1725 	}
1726 
1727 	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s xcp_id %d\n",
1728 		 va, (*mem)->aql_queue ? size << 1 : size,
1729 		 domain_string(alloc_domain), xcp_id);
1730 
1731 	ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags,
1732 				       bo_type, NULL, &gobj, xcp_id + 1);
1733 	if (ret) {
1734 		pr_debug("Failed to create BO on domain %s. ret %d\n",
1735 			 domain_string(alloc_domain), ret);
1736 		goto err_bo_create;
1737 	}
1738 	ret = drm_vma_node_allow(&gobj->vma_node, drm_priv);
1739 	if (ret) {
1740 		pr_debug("Failed to allow vma node access. ret %d\n", ret);
1741 		goto err_node_allow;
1742 	}
1743 	bo = gem_to_amdgpu_bo(gobj);
1744 	if (bo_type == ttm_bo_type_sg) {
1745 		bo->tbo.sg = sg;
1746 		bo->tbo.ttm->sg = sg;
1747 	}
1748 	bo->kfd_bo = *mem;
1749 	(*mem)->bo = bo;
1750 	if (user_addr)
1751 		bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO;
1752 
1753 	(*mem)->va = va;
1754 	(*mem)->domain = domain;
1755 	(*mem)->mapped_to_gpu_memory = 0;
1756 	(*mem)->process_info = avm->process_info;
1757 
1758 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1759 
1760 	if (user_addr) {
1761 		pr_debug("creating userptr BO for user_addr = %llx\n", user_addr);
1762 		ret = init_user_pages(*mem, user_addr, criu_resume);
1763 		if (ret)
1764 			goto allocate_init_user_pages_failed;
1765 	} else  if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1766 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1767 		ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT);
1768 		if (ret) {
1769 			pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n");
1770 			goto err_pin_bo;
1771 		}
1772 		bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
1773 		bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
1774 	}
1775 
1776 	if (offset)
1777 		*offset = amdgpu_bo_mmap_offset(bo);
1778 
1779 	return 0;
1780 
1781 allocate_init_user_pages_failed:
1782 err_pin_bo:
1783 	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1784 	drm_vma_node_revoke(&gobj->vma_node, drm_priv);
1785 err_node_allow:
1786 	/* Don't unreserve system mem limit twice */
1787 	goto err_reserve_limit;
1788 err_bo_create:
1789 	amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags, xcp_id);
1790 err_reserve_limit:
1791 	mutex_destroy(&(*mem)->lock);
1792 	if (gobj)
1793 		drm_gem_object_put(gobj);
1794 	else
1795 		kfree(*mem);
1796 err:
1797 	if (sg) {
1798 		sg_free_table(sg);
1799 		kfree(sg);
1800 	}
1801 	return ret;
1802 }
1803 
1804 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1805 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
1806 		uint64_t *size)
1807 {
1808 	struct amdkfd_process_info *process_info = mem->process_info;
1809 	unsigned long bo_size = mem->bo->tbo.base.size;
1810 	bool use_release_notifier = (mem->bo->kfd_bo == mem);
1811 	struct kfd_mem_attachment *entry, *tmp;
1812 	struct bo_vm_reservation_context ctx;
1813 	unsigned int mapped_to_gpu_memory;
1814 	int ret;
1815 	bool is_imported = false;
1816 
1817 	mutex_lock(&mem->lock);
1818 
1819 	/* Unpin MMIO/DOORBELL BO's that were pinned during allocation */
1820 	if (mem->alloc_flags &
1821 	    (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1822 	     KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1823 		amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo);
1824 	}
1825 
1826 	mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1827 	is_imported = mem->is_imported;
1828 	mutex_unlock(&mem->lock);
1829 	/* lock is not needed after this, since mem is unused and will
1830 	 * be freed anyway
1831 	 */
1832 
1833 	if (mapped_to_gpu_memory > 0) {
1834 		pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1835 				mem->va, bo_size);
1836 		return -EBUSY;
1837 	}
1838 
1839 	/* Make sure restore workers don't access the BO any more */
1840 	mutex_lock(&process_info->lock);
1841 	list_del(&mem->validate_list);
1842 	mutex_unlock(&process_info->lock);
1843 
1844 	/* Cleanup user pages and MMU notifiers */
1845 	if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
1846 		amdgpu_hmm_unregister(mem->bo);
1847 		mutex_lock(&process_info->notifier_lock);
1848 		amdgpu_ttm_tt_discard_user_pages(mem->bo->tbo.ttm, mem->range);
1849 		mutex_unlock(&process_info->notifier_lock);
1850 	}
1851 
1852 	ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1853 	if (unlikely(ret))
1854 		return ret;
1855 
1856 	/* The eviction fence should be removed by the last unmap.
1857 	 * TODO: Log an error condition if the bo still has the eviction fence
1858 	 * attached
1859 	 */
1860 	amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1861 					process_info->eviction_fence);
1862 	pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1863 		mem->va + bo_size * (1 + mem->aql_queue));
1864 
1865 	/* Remove from VM internal data structures */
1866 	list_for_each_entry_safe(entry, tmp, &mem->attachments, list)
1867 		kfd_mem_detach(entry);
1868 
1869 	ret = unreserve_bo_and_vms(&ctx, false, false);
1870 
1871 	/* Free the sync object */
1872 	amdgpu_sync_free(&mem->sync);
1873 
1874 	/* If the SG is not NULL, it's one we created for a doorbell or mmio
1875 	 * remap BO. We need to free it.
1876 	 */
1877 	if (mem->bo->tbo.sg) {
1878 		sg_free_table(mem->bo->tbo.sg);
1879 		kfree(mem->bo->tbo.sg);
1880 	}
1881 
1882 	/* Update the size of the BO being freed if it was allocated from
1883 	 * VRAM and is not imported. For APP APU VRAM allocations are done
1884 	 * in GTT domain
1885 	 */
1886 	if (size) {
1887 		if (!is_imported &&
1888 		   (mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM ||
1889 		   (adev->gmc.is_app_apu &&
1890 		    mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT)))
1891 			*size = bo_size;
1892 		else
1893 			*size = 0;
1894 	}
1895 
1896 	/* Free the BO*/
1897 	drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
1898 	if (mem->dmabuf)
1899 		dma_buf_put(mem->dmabuf);
1900 	mutex_destroy(&mem->lock);
1901 
1902 	/* If this releases the last reference, it will end up calling
1903 	 * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why
1904 	 * this needs to be the last call here.
1905 	 */
1906 	drm_gem_object_put(&mem->bo->tbo.base);
1907 
1908 	/*
1909 	 * For kgd_mem allocated in amdgpu_amdkfd_gpuvm_import_dmabuf(),
1910 	 * explicitly free it here.
1911 	 */
1912 	if (!use_release_notifier)
1913 		kfree(mem);
1914 
1915 	return ret;
1916 }
1917 
1918 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1919 		struct amdgpu_device *adev, struct kgd_mem *mem,
1920 		void *drm_priv)
1921 {
1922 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1923 	int ret;
1924 	struct amdgpu_bo *bo;
1925 	uint32_t domain;
1926 	struct kfd_mem_attachment *entry;
1927 	struct bo_vm_reservation_context ctx;
1928 	unsigned long bo_size;
1929 	bool is_invalid_userptr = false;
1930 
1931 	bo = mem->bo;
1932 	if (!bo) {
1933 		pr_err("Invalid BO when mapping memory to GPU\n");
1934 		return -EINVAL;
1935 	}
1936 
1937 	/* Make sure restore is not running concurrently. Since we
1938 	 * don't map invalid userptr BOs, we rely on the next restore
1939 	 * worker to do the mapping
1940 	 */
1941 	mutex_lock(&mem->process_info->lock);
1942 
1943 	/* Lock notifier lock. If we find an invalid userptr BO, we can be
1944 	 * sure that the MMU notifier is no longer running
1945 	 * concurrently and the queues are actually stopped
1946 	 */
1947 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1948 		mutex_lock(&mem->process_info->notifier_lock);
1949 		is_invalid_userptr = !!mem->invalid;
1950 		mutex_unlock(&mem->process_info->notifier_lock);
1951 	}
1952 
1953 	mutex_lock(&mem->lock);
1954 
1955 	domain = mem->domain;
1956 	bo_size = bo->tbo.base.size;
1957 
1958 	pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
1959 			mem->va,
1960 			mem->va + bo_size * (1 + mem->aql_queue),
1961 			avm, domain_string(domain));
1962 
1963 	if (!kfd_mem_is_attached(avm, mem)) {
1964 		ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue);
1965 		if (ret)
1966 			goto out;
1967 	}
1968 
1969 	ret = reserve_bo_and_vm(mem, avm, &ctx);
1970 	if (unlikely(ret))
1971 		goto out;
1972 
1973 	/* Userptr can be marked as "not invalid", but not actually be
1974 	 * validated yet (still in the system domain). In that case
1975 	 * the queues are still stopped and we can leave mapping for
1976 	 * the next restore worker
1977 	 */
1978 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
1979 	    bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
1980 		is_invalid_userptr = true;
1981 
1982 	ret = vm_validate_pt_pd_bos(avm);
1983 	if (unlikely(ret))
1984 		goto out_unreserve;
1985 
1986 	if (mem->mapped_to_gpu_memory == 0 &&
1987 	    !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1988 		/* Validate BO only once. The eviction fence gets added to BO
1989 		 * the first time it is mapped. Validate will wait for all
1990 		 * background evictions to complete.
1991 		 */
1992 		ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
1993 		if (ret) {
1994 			pr_debug("Validate failed\n");
1995 			goto out_unreserve;
1996 		}
1997 	}
1998 
1999 	list_for_each_entry(entry, &mem->attachments, list) {
2000 		if (entry->bo_va->base.vm != avm || entry->is_mapped)
2001 			continue;
2002 
2003 		pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
2004 			 entry->va, entry->va + bo_size, entry);
2005 
2006 		ret = map_bo_to_gpuvm(mem, entry, ctx.sync,
2007 				      is_invalid_userptr);
2008 		if (ret) {
2009 			pr_err("Failed to map bo to gpuvm\n");
2010 			goto out_unreserve;
2011 		}
2012 
2013 		ret = vm_update_pds(avm, ctx.sync);
2014 		if (ret) {
2015 			pr_err("Failed to update page directories\n");
2016 			goto out_unreserve;
2017 		}
2018 
2019 		entry->is_mapped = true;
2020 		mem->mapped_to_gpu_memory++;
2021 		pr_debug("\t INC mapping count %d\n",
2022 			 mem->mapped_to_gpu_memory);
2023 	}
2024 
2025 	if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->tbo.pin_count)
2026 		dma_resv_add_fence(bo->tbo.base.resv,
2027 				   &avm->process_info->eviction_fence->base,
2028 				   DMA_RESV_USAGE_BOOKKEEP);
2029 	ret = unreserve_bo_and_vms(&ctx, false, false);
2030 
2031 	goto out;
2032 
2033 out_unreserve:
2034 	unreserve_bo_and_vms(&ctx, false, false);
2035 out:
2036 	mutex_unlock(&mem->process_info->lock);
2037 	mutex_unlock(&mem->lock);
2038 	return ret;
2039 }
2040 
2041 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
2042 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
2043 {
2044 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2045 	struct amdkfd_process_info *process_info = avm->process_info;
2046 	unsigned long bo_size = mem->bo->tbo.base.size;
2047 	struct kfd_mem_attachment *entry;
2048 	struct bo_vm_reservation_context ctx;
2049 	int ret;
2050 
2051 	mutex_lock(&mem->lock);
2052 
2053 	ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx);
2054 	if (unlikely(ret))
2055 		goto out;
2056 	/* If no VMs were reserved, it means the BO wasn't actually mapped */
2057 	if (ctx.n_vms == 0) {
2058 		ret = -EINVAL;
2059 		goto unreserve_out;
2060 	}
2061 
2062 	ret = vm_validate_pt_pd_bos(avm);
2063 	if (unlikely(ret))
2064 		goto unreserve_out;
2065 
2066 	pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
2067 		mem->va,
2068 		mem->va + bo_size * (1 + mem->aql_queue),
2069 		avm);
2070 
2071 	list_for_each_entry(entry, &mem->attachments, list) {
2072 		if (entry->bo_va->base.vm != avm || !entry->is_mapped)
2073 			continue;
2074 
2075 		pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
2076 			 entry->va, entry->va + bo_size, entry);
2077 
2078 		unmap_bo_from_gpuvm(mem, entry, ctx.sync);
2079 		entry->is_mapped = false;
2080 
2081 		mem->mapped_to_gpu_memory--;
2082 		pr_debug("\t DEC mapping count %d\n",
2083 			 mem->mapped_to_gpu_memory);
2084 	}
2085 
2086 	/* If BO is unmapped from all VMs, unfence it. It can be evicted if
2087 	 * required.
2088 	 */
2089 	if (mem->mapped_to_gpu_memory == 0 &&
2090 	    !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) &&
2091 	    !mem->bo->tbo.pin_count)
2092 		amdgpu_amdkfd_remove_eviction_fence(mem->bo,
2093 						process_info->eviction_fence);
2094 
2095 unreserve_out:
2096 	unreserve_bo_and_vms(&ctx, false, false);
2097 out:
2098 	mutex_unlock(&mem->lock);
2099 	return ret;
2100 }
2101 
2102 int amdgpu_amdkfd_gpuvm_sync_memory(
2103 		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr)
2104 {
2105 	struct amdgpu_sync sync;
2106 	int ret;
2107 
2108 	amdgpu_sync_create(&sync);
2109 
2110 	mutex_lock(&mem->lock);
2111 	amdgpu_sync_clone(&mem->sync, &sync);
2112 	mutex_unlock(&mem->lock);
2113 
2114 	ret = amdgpu_sync_wait(&sync, intr);
2115 	amdgpu_sync_free(&sync);
2116 	return ret;
2117 }
2118 
2119 /**
2120  * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count
2121  * @adev: Device to which allocated BO belongs
2122  * @bo: Buffer object to be mapped
2123  *
2124  * Before return, bo reference count is incremented. To release the reference and unpin/
2125  * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
2126  */
2127 int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo)
2128 {
2129 	int ret;
2130 
2131 	ret = amdgpu_bo_reserve(bo, true);
2132 	if (ret) {
2133 		pr_err("Failed to reserve bo. ret %d\n", ret);
2134 		goto err_reserve_bo_failed;
2135 	}
2136 
2137 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2138 	if (ret) {
2139 		pr_err("Failed to pin bo. ret %d\n", ret);
2140 		goto err_pin_bo_failed;
2141 	}
2142 
2143 	ret = amdgpu_ttm_alloc_gart(&bo->tbo);
2144 	if (ret) {
2145 		pr_err("Failed to bind bo to GART. ret %d\n", ret);
2146 		goto err_map_bo_gart_failed;
2147 	}
2148 
2149 	amdgpu_amdkfd_remove_eviction_fence(
2150 		bo, bo->vm_bo->vm->process_info->eviction_fence);
2151 
2152 	amdgpu_bo_unreserve(bo);
2153 
2154 	bo = amdgpu_bo_ref(bo);
2155 
2156 	return 0;
2157 
2158 err_map_bo_gart_failed:
2159 	amdgpu_bo_unpin(bo);
2160 err_pin_bo_failed:
2161 	amdgpu_bo_unreserve(bo);
2162 err_reserve_bo_failed:
2163 
2164 	return ret;
2165 }
2166 
2167 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Map a GTT BO for kernel CPU access
2168  *
2169  * @mem: Buffer object to be mapped for CPU access
2170  * @kptr[out]: pointer in kernel CPU address space
2171  * @size[out]: size of the buffer
2172  *
2173  * Pins the BO and maps it for kernel CPU access. The eviction fence is removed
2174  * from the BO, since pinned BOs cannot be evicted. The bo must remain on the
2175  * validate_list, so the GPU mapping can be restored after a page table was
2176  * evicted.
2177  *
2178  * Return: 0 on success, error code on failure
2179  */
2180 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,
2181 					     void **kptr, uint64_t *size)
2182 {
2183 	int ret;
2184 	struct amdgpu_bo *bo = mem->bo;
2185 
2186 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2187 		pr_err("userptr can't be mapped to kernel\n");
2188 		return -EINVAL;
2189 	}
2190 
2191 	mutex_lock(&mem->process_info->lock);
2192 
2193 	ret = amdgpu_bo_reserve(bo, true);
2194 	if (ret) {
2195 		pr_err("Failed to reserve bo. ret %d\n", ret);
2196 		goto bo_reserve_failed;
2197 	}
2198 
2199 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2200 	if (ret) {
2201 		pr_err("Failed to pin bo. ret %d\n", ret);
2202 		goto pin_failed;
2203 	}
2204 
2205 	ret = amdgpu_bo_kmap(bo, kptr);
2206 	if (ret) {
2207 		pr_err("Failed to map bo to kernel. ret %d\n", ret);
2208 		goto kmap_failed;
2209 	}
2210 
2211 	amdgpu_amdkfd_remove_eviction_fence(
2212 		bo, mem->process_info->eviction_fence);
2213 
2214 	if (size)
2215 		*size = amdgpu_bo_size(bo);
2216 
2217 	amdgpu_bo_unreserve(bo);
2218 
2219 	mutex_unlock(&mem->process_info->lock);
2220 	return 0;
2221 
2222 kmap_failed:
2223 	amdgpu_bo_unpin(bo);
2224 pin_failed:
2225 	amdgpu_bo_unreserve(bo);
2226 bo_reserve_failed:
2227 	mutex_unlock(&mem->process_info->lock);
2228 
2229 	return ret;
2230 }
2231 
2232 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Unmap a GTT BO for kernel CPU access
2233  *
2234  * @mem: Buffer object to be unmapped for CPU access
2235  *
2236  * Removes the kernel CPU mapping and unpins the BO. It does not restore the
2237  * eviction fence, so this function should only be used for cleanup before the
2238  * BO is destroyed.
2239  */
2240 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem)
2241 {
2242 	struct amdgpu_bo *bo = mem->bo;
2243 
2244 	amdgpu_bo_reserve(bo, true);
2245 	amdgpu_bo_kunmap(bo);
2246 	amdgpu_bo_unpin(bo);
2247 	amdgpu_bo_unreserve(bo);
2248 }
2249 
2250 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
2251 					  struct kfd_vm_fault_info *mem)
2252 {
2253 	if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
2254 		*mem = *adev->gmc.vm_fault_info;
2255 		mb(); /* make sure read happened */
2256 		atomic_set(&adev->gmc.vm_fault_info_updated, 0);
2257 	}
2258 	return 0;
2259 }
2260 
2261 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
2262 				      struct dma_buf *dma_buf,
2263 				      uint64_t va, void *drm_priv,
2264 				      struct kgd_mem **mem, uint64_t *size,
2265 				      uint64_t *mmap_offset)
2266 {
2267 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2268 	struct drm_gem_object *obj;
2269 	struct amdgpu_bo *bo;
2270 	int ret;
2271 
2272 	obj = amdgpu_gem_prime_import(adev_to_drm(adev), dma_buf);
2273 	if (IS_ERR(obj))
2274 		return PTR_ERR(obj);
2275 
2276 	bo = gem_to_amdgpu_bo(obj);
2277 	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
2278 				    AMDGPU_GEM_DOMAIN_GTT))) {
2279 		/* Only VRAM and GTT BOs are supported */
2280 		ret = -EINVAL;
2281 		goto err_put_obj;
2282 	}
2283 
2284 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2285 	if (!*mem) {
2286 		ret = -ENOMEM;
2287 		goto err_put_obj;
2288 	}
2289 
2290 	ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
2291 	if (ret)
2292 		goto err_free_mem;
2293 
2294 	if (size)
2295 		*size = amdgpu_bo_size(bo);
2296 
2297 	if (mmap_offset)
2298 		*mmap_offset = amdgpu_bo_mmap_offset(bo);
2299 
2300 	INIT_LIST_HEAD(&(*mem)->attachments);
2301 	mutex_init(&(*mem)->lock);
2302 
2303 	(*mem)->alloc_flags =
2304 		((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2305 		KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
2306 		| KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
2307 		| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
2308 
2309 	get_dma_buf(dma_buf);
2310 	(*mem)->dmabuf = dma_buf;
2311 	(*mem)->bo = bo;
2312 	(*mem)->va = va;
2313 	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) && !adev->gmc.is_app_apu ?
2314 		AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
2315 
2316 	(*mem)->mapped_to_gpu_memory = 0;
2317 	(*mem)->process_info = avm->process_info;
2318 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
2319 	amdgpu_sync_create(&(*mem)->sync);
2320 	(*mem)->is_imported = true;
2321 
2322 	return 0;
2323 
2324 err_free_mem:
2325 	kfree(*mem);
2326 err_put_obj:
2327 	drm_gem_object_put(obj);
2328 	return ret;
2329 }
2330 
2331 int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem,
2332 				      struct dma_buf **dma_buf)
2333 {
2334 	int ret;
2335 
2336 	mutex_lock(&mem->lock);
2337 	ret = kfd_mem_export_dmabuf(mem);
2338 	if (ret)
2339 		goto out;
2340 
2341 	get_dma_buf(mem->dmabuf);
2342 	*dma_buf = mem->dmabuf;
2343 out:
2344 	mutex_unlock(&mem->lock);
2345 	return ret;
2346 }
2347 
2348 /* Evict a userptr BO by stopping the queues if necessary
2349  *
2350  * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
2351  * cannot do any memory allocations, and cannot take any locks that
2352  * are held elsewhere while allocating memory.
2353  *
2354  * It doesn't do anything to the BO itself. The real work happens in
2355  * restore, where we get updated page addresses. This function only
2356  * ensures that GPU access to the BO is stopped.
2357  */
2358 int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
2359 				unsigned long cur_seq, struct kgd_mem *mem)
2360 {
2361 	struct amdkfd_process_info *process_info = mem->process_info;
2362 	int r = 0;
2363 
2364 	/* Do not process MMU notifications during CRIU restore until
2365 	 * KFD_CRIU_OP_RESUME IOCTL is received
2366 	 */
2367 	if (READ_ONCE(process_info->block_mmu_notifications))
2368 		return 0;
2369 
2370 	mutex_lock(&process_info->notifier_lock);
2371 	mmu_interval_set_seq(mni, cur_seq);
2372 
2373 	mem->invalid++;
2374 	if (++process_info->evicted_bos == 1) {
2375 		/* First eviction, stop the queues */
2376 		r = kgd2kfd_quiesce_mm(mni->mm,
2377 				       KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
2378 		if (r)
2379 			pr_err("Failed to quiesce KFD\n");
2380 		schedule_delayed_work(&process_info->restore_userptr_work,
2381 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2382 	}
2383 	mutex_unlock(&process_info->notifier_lock);
2384 
2385 	return r;
2386 }
2387 
2388 /* Update invalid userptr BOs
2389  *
2390  * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
2391  * userptr_inval_list and updates user pages for all BOs that have
2392  * been invalidated since their last update.
2393  */
2394 static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
2395 				     struct mm_struct *mm)
2396 {
2397 	struct kgd_mem *mem, *tmp_mem;
2398 	struct amdgpu_bo *bo;
2399 	struct ttm_operation_ctx ctx = { false, false };
2400 	uint32_t invalid;
2401 	int ret = 0;
2402 
2403 	mutex_lock(&process_info->notifier_lock);
2404 
2405 	/* Move all invalidated BOs to the userptr_inval_list */
2406 	list_for_each_entry_safe(mem, tmp_mem,
2407 				 &process_info->userptr_valid_list,
2408 				 validate_list)
2409 		if (mem->invalid)
2410 			list_move_tail(&mem->validate_list,
2411 				       &process_info->userptr_inval_list);
2412 
2413 	/* Go through userptr_inval_list and update any invalid user_pages */
2414 	list_for_each_entry(mem, &process_info->userptr_inval_list,
2415 			    validate_list) {
2416 		invalid = mem->invalid;
2417 		if (!invalid)
2418 			/* BO hasn't been invalidated since the last
2419 			 * revalidation attempt. Keep its page list.
2420 			 */
2421 			continue;
2422 
2423 		bo = mem->bo;
2424 
2425 		amdgpu_ttm_tt_discard_user_pages(bo->tbo.ttm, mem->range);
2426 		mem->range = NULL;
2427 
2428 		/* BO reservations and getting user pages (hmm_range_fault)
2429 		 * must happen outside the notifier lock
2430 		 */
2431 		mutex_unlock(&process_info->notifier_lock);
2432 
2433 		/* Move the BO to system (CPU) domain if necessary to unmap
2434 		 * and free the SG table
2435 		 */
2436 		if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) {
2437 			if (amdgpu_bo_reserve(bo, true))
2438 				return -EAGAIN;
2439 			amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
2440 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2441 			amdgpu_bo_unreserve(bo);
2442 			if (ret) {
2443 				pr_err("%s: Failed to invalidate userptr BO\n",
2444 				       __func__);
2445 				return -EAGAIN;
2446 			}
2447 		}
2448 
2449 		/* Get updated user pages */
2450 		ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages,
2451 						   &mem->range);
2452 		if (ret) {
2453 			pr_debug("Failed %d to get user pages\n", ret);
2454 
2455 			/* Return -EFAULT bad address error as success. It will
2456 			 * fail later with a VM fault if the GPU tries to access
2457 			 * it. Better than hanging indefinitely with stalled
2458 			 * user mode queues.
2459 			 *
2460 			 * Return other error -EBUSY or -ENOMEM to retry restore
2461 			 */
2462 			if (ret != -EFAULT)
2463 				return ret;
2464 
2465 			ret = 0;
2466 		}
2467 
2468 		mutex_lock(&process_info->notifier_lock);
2469 
2470 		/* Mark the BO as valid unless it was invalidated
2471 		 * again concurrently.
2472 		 */
2473 		if (mem->invalid != invalid) {
2474 			ret = -EAGAIN;
2475 			goto unlock_out;
2476 		}
2477 		 /* set mem valid if mem has hmm range associated */
2478 		if (mem->range)
2479 			mem->invalid = 0;
2480 	}
2481 
2482 unlock_out:
2483 	mutex_unlock(&process_info->notifier_lock);
2484 
2485 	return ret;
2486 }
2487 
2488 /* Validate invalid userptr BOs
2489  *
2490  * Validates BOs on the userptr_inval_list. Also updates GPUVM page tables
2491  * with new page addresses and waits for the page table updates to complete.
2492  */
2493 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
2494 {
2495 	struct ttm_operation_ctx ctx = { false, false };
2496 	struct amdgpu_sync sync;
2497 	struct drm_exec exec;
2498 
2499 	struct amdgpu_vm *peer_vm;
2500 	struct kgd_mem *mem, *tmp_mem;
2501 	struct amdgpu_bo *bo;
2502 	int ret;
2503 
2504 	amdgpu_sync_create(&sync);
2505 
2506 	drm_exec_init(&exec, 0);
2507 	/* Reserve all BOs and page tables for validation */
2508 	drm_exec_until_all_locked(&exec) {
2509 		/* Reserve all the page directories */
2510 		list_for_each_entry(peer_vm, &process_info->vm_list_head,
2511 				    vm_list_node) {
2512 			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2513 			drm_exec_retry_on_contention(&exec);
2514 			if (unlikely(ret))
2515 				goto unreserve_out;
2516 		}
2517 
2518 		/* Reserve the userptr_inval_list entries to resv_list */
2519 		list_for_each_entry(mem, &process_info->userptr_inval_list,
2520 				    validate_list) {
2521 			struct drm_gem_object *gobj;
2522 
2523 			gobj = &mem->bo->tbo.base;
2524 			ret = drm_exec_prepare_obj(&exec, gobj, 1);
2525 			drm_exec_retry_on_contention(&exec);
2526 			if (unlikely(ret))
2527 				goto unreserve_out;
2528 		}
2529 	}
2530 
2531 	ret = process_validate_vms(process_info);
2532 	if (ret)
2533 		goto unreserve_out;
2534 
2535 	/* Validate BOs and update GPUVM page tables */
2536 	list_for_each_entry_safe(mem, tmp_mem,
2537 				 &process_info->userptr_inval_list,
2538 				 validate_list) {
2539 		struct kfd_mem_attachment *attachment;
2540 
2541 		bo = mem->bo;
2542 
2543 		/* Validate the BO if we got user pages */
2544 		if (bo->tbo.ttm->pages[0]) {
2545 			amdgpu_bo_placement_from_domain(bo, mem->domain);
2546 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2547 			if (ret) {
2548 				pr_err("%s: failed to validate BO\n", __func__);
2549 				goto unreserve_out;
2550 			}
2551 		}
2552 
2553 		/* Update mapping. If the BO was not validated
2554 		 * (because we couldn't get user pages), this will
2555 		 * clear the page table entries, which will result in
2556 		 * VM faults if the GPU tries to access the invalid
2557 		 * memory.
2558 		 */
2559 		list_for_each_entry(attachment, &mem->attachments, list) {
2560 			if (!attachment->is_mapped)
2561 				continue;
2562 
2563 			kfd_mem_dmaunmap_attachment(mem, attachment);
2564 			ret = update_gpuvm_pte(mem, attachment, &sync);
2565 			if (ret) {
2566 				pr_err("%s: update PTE failed\n", __func__);
2567 				/* make sure this gets validated again */
2568 				mutex_lock(&process_info->notifier_lock);
2569 				mem->invalid++;
2570 				mutex_unlock(&process_info->notifier_lock);
2571 				goto unreserve_out;
2572 			}
2573 		}
2574 	}
2575 
2576 	/* Update page directories */
2577 	ret = process_update_pds(process_info, &sync);
2578 
2579 unreserve_out:
2580 	drm_exec_fini(&exec);
2581 	amdgpu_sync_wait(&sync, false);
2582 	amdgpu_sync_free(&sync);
2583 
2584 	return ret;
2585 }
2586 
2587 /* Confirm that all user pages are valid while holding the notifier lock
2588  *
2589  * Moves valid BOs from the userptr_inval_list back to userptr_val_list.
2590  */
2591 static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_info)
2592 {
2593 	struct kgd_mem *mem, *tmp_mem;
2594 	int ret = 0;
2595 
2596 	list_for_each_entry_safe(mem, tmp_mem,
2597 				 &process_info->userptr_inval_list,
2598 				 validate_list) {
2599 		bool valid;
2600 
2601 		/* keep mem without hmm range at userptr_inval_list */
2602 		if (!mem->range)
2603 			 continue;
2604 
2605 		/* Only check mem with hmm range associated */
2606 		valid = amdgpu_ttm_tt_get_user_pages_done(
2607 					mem->bo->tbo.ttm, mem->range);
2608 
2609 		mem->range = NULL;
2610 		if (!valid) {
2611 			WARN(!mem->invalid, "Invalid BO not marked invalid");
2612 			ret = -EAGAIN;
2613 			continue;
2614 		}
2615 
2616 		if (mem->invalid) {
2617 			WARN(1, "Valid BO is marked invalid");
2618 			ret = -EAGAIN;
2619 			continue;
2620 		}
2621 
2622 		list_move_tail(&mem->validate_list,
2623 			       &process_info->userptr_valid_list);
2624 	}
2625 
2626 	return ret;
2627 }
2628 
2629 /* Worker callback to restore evicted userptr BOs
2630  *
2631  * Tries to update and validate all userptr BOs. If successful and no
2632  * concurrent evictions happened, the queues are restarted. Otherwise,
2633  * reschedule for another attempt later.
2634  */
2635 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
2636 {
2637 	struct delayed_work *dwork = to_delayed_work(work);
2638 	struct amdkfd_process_info *process_info =
2639 		container_of(dwork, struct amdkfd_process_info,
2640 			     restore_userptr_work);
2641 	struct task_struct *usertask;
2642 	struct mm_struct *mm;
2643 	uint32_t evicted_bos;
2644 
2645 	mutex_lock(&process_info->notifier_lock);
2646 	evicted_bos = process_info->evicted_bos;
2647 	mutex_unlock(&process_info->notifier_lock);
2648 	if (!evicted_bos)
2649 		return;
2650 
2651 	/* Reference task and mm in case of concurrent process termination */
2652 	usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
2653 	if (!usertask)
2654 		return;
2655 	mm = get_task_mm(usertask);
2656 	if (!mm) {
2657 		put_task_struct(usertask);
2658 		return;
2659 	}
2660 
2661 	mutex_lock(&process_info->lock);
2662 
2663 	if (update_invalid_user_pages(process_info, mm))
2664 		goto unlock_out;
2665 	/* userptr_inval_list can be empty if all evicted userptr BOs
2666 	 * have been freed. In that case there is nothing to validate
2667 	 * and we can just restart the queues.
2668 	 */
2669 	if (!list_empty(&process_info->userptr_inval_list)) {
2670 		if (validate_invalid_user_pages(process_info))
2671 			goto unlock_out;
2672 	}
2673 	/* Final check for concurrent evicton and atomic update. If
2674 	 * another eviction happens after successful update, it will
2675 	 * be a first eviction that calls quiesce_mm. The eviction
2676 	 * reference counting inside KFD will handle this case.
2677 	 */
2678 	mutex_lock(&process_info->notifier_lock);
2679 	if (process_info->evicted_bos != evicted_bos)
2680 		goto unlock_notifier_out;
2681 
2682 	if (confirm_valid_user_pages_locked(process_info)) {
2683 		WARN(1, "User pages unexpectedly invalid");
2684 		goto unlock_notifier_out;
2685 	}
2686 
2687 	process_info->evicted_bos = evicted_bos = 0;
2688 
2689 	if (kgd2kfd_resume_mm(mm)) {
2690 		pr_err("%s: Failed to resume KFD\n", __func__);
2691 		/* No recovery from this failure. Probably the CP is
2692 		 * hanging. No point trying again.
2693 		 */
2694 	}
2695 
2696 unlock_notifier_out:
2697 	mutex_unlock(&process_info->notifier_lock);
2698 unlock_out:
2699 	mutex_unlock(&process_info->lock);
2700 
2701 	/* If validation failed, reschedule another attempt */
2702 	if (evicted_bos) {
2703 		schedule_delayed_work(&process_info->restore_userptr_work,
2704 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2705 
2706 		kfd_smi_event_queue_restore_rescheduled(mm);
2707 	}
2708 	mmput(mm);
2709 	put_task_struct(usertask);
2710 }
2711 
2712 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2713  *   KFD process identified by process_info
2714  *
2715  * @process_info: amdkfd_process_info of the KFD process
2716  *
2717  * After memory eviction, restore thread calls this function. The function
2718  * should be called when the Process is still valid. BO restore involves -
2719  *
2720  * 1.  Release old eviction fence and create new one
2721  * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2722  * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2723  *     BOs that need to be reserved.
2724  * 4.  Reserve all the BOs
2725  * 5.  Validate of PD and PT BOs.
2726  * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2727  * 7.  Add fence to all PD and PT BOs.
2728  * 8.  Unreserve all BOs
2729  */
2730 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
2731 {
2732 	struct amdkfd_process_info *process_info = info;
2733 	struct amdgpu_vm *peer_vm;
2734 	struct kgd_mem *mem;
2735 	struct amdgpu_amdkfd_fence *new_fence;
2736 	struct list_head duplicate_save;
2737 	struct amdgpu_sync sync_obj;
2738 	unsigned long failed_size = 0;
2739 	unsigned long total_size = 0;
2740 	struct drm_exec exec;
2741 	int ret;
2742 
2743 	INIT_LIST_HEAD(&duplicate_save);
2744 
2745 	mutex_lock(&process_info->lock);
2746 
2747 	drm_exec_init(&exec, 0);
2748 	drm_exec_until_all_locked(&exec) {
2749 		list_for_each_entry(peer_vm, &process_info->vm_list_head,
2750 				    vm_list_node) {
2751 			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2752 			drm_exec_retry_on_contention(&exec);
2753 			if (unlikely(ret))
2754 				goto ttm_reserve_fail;
2755 		}
2756 
2757 		/* Reserve all BOs and page tables/directory. Add all BOs from
2758 		 * kfd_bo_list to ctx.list
2759 		 */
2760 		list_for_each_entry(mem, &process_info->kfd_bo_list,
2761 				    validate_list) {
2762 			struct drm_gem_object *gobj;
2763 
2764 			gobj = &mem->bo->tbo.base;
2765 			ret = drm_exec_prepare_obj(&exec, gobj, 1);
2766 			drm_exec_retry_on_contention(&exec);
2767 			if (unlikely(ret))
2768 				goto ttm_reserve_fail;
2769 		}
2770 	}
2771 
2772 	amdgpu_sync_create(&sync_obj);
2773 
2774 	/* Validate PDs and PTs */
2775 	ret = process_validate_vms(process_info);
2776 	if (ret)
2777 		goto validate_map_fail;
2778 
2779 	ret = process_sync_pds_resv(process_info, &sync_obj);
2780 	if (ret) {
2781 		pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2782 		goto validate_map_fail;
2783 	}
2784 
2785 	/* Validate BOs and map them to GPUVM (update VM page tables). */
2786 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2787 			    validate_list) {
2788 
2789 		struct amdgpu_bo *bo = mem->bo;
2790 		uint32_t domain = mem->domain;
2791 		struct kfd_mem_attachment *attachment;
2792 		struct dma_resv_iter cursor;
2793 		struct dma_fence *fence;
2794 
2795 		total_size += amdgpu_bo_size(bo);
2796 
2797 		ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2798 		if (ret) {
2799 			pr_debug("Memory eviction: Validate BOs failed\n");
2800 			failed_size += amdgpu_bo_size(bo);
2801 			ret = amdgpu_amdkfd_bo_validate(bo,
2802 						AMDGPU_GEM_DOMAIN_GTT, false);
2803 			if (ret) {
2804 				pr_debug("Memory eviction: Try again\n");
2805 				goto validate_map_fail;
2806 			}
2807 		}
2808 		dma_resv_for_each_fence(&cursor, bo->tbo.base.resv,
2809 					DMA_RESV_USAGE_KERNEL, fence) {
2810 			ret = amdgpu_sync_fence(&sync_obj, fence);
2811 			if (ret) {
2812 				pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2813 				goto validate_map_fail;
2814 			}
2815 		}
2816 		list_for_each_entry(attachment, &mem->attachments, list) {
2817 			if (!attachment->is_mapped)
2818 				continue;
2819 
2820 			if (attachment->bo_va->base.bo->tbo.pin_count)
2821 				continue;
2822 
2823 			kfd_mem_dmaunmap_attachment(mem, attachment);
2824 			ret = update_gpuvm_pte(mem, attachment, &sync_obj);
2825 			if (ret) {
2826 				pr_debug("Memory eviction: update PTE failed. Try again\n");
2827 				goto validate_map_fail;
2828 			}
2829 		}
2830 	}
2831 
2832 	if (failed_size)
2833 		pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2834 
2835 	/* Update page directories */
2836 	ret = process_update_pds(process_info, &sync_obj);
2837 	if (ret) {
2838 		pr_debug("Memory eviction: update PDs failed. Try again\n");
2839 		goto validate_map_fail;
2840 	}
2841 
2842 	/* Wait for validate and PT updates to finish */
2843 	amdgpu_sync_wait(&sync_obj, false);
2844 
2845 	/* Release old eviction fence and create new one, because fence only
2846 	 * goes from unsignaled to signaled, fence cannot be reused.
2847 	 * Use context and mm from the old fence.
2848 	 */
2849 	new_fence = amdgpu_amdkfd_fence_create(
2850 				process_info->eviction_fence->base.context,
2851 				process_info->eviction_fence->mm,
2852 				NULL);
2853 	if (!new_fence) {
2854 		pr_err("Failed to create eviction fence\n");
2855 		ret = -ENOMEM;
2856 		goto validate_map_fail;
2857 	}
2858 	dma_fence_put(&process_info->eviction_fence->base);
2859 	process_info->eviction_fence = new_fence;
2860 	*ef = dma_fence_get(&new_fence->base);
2861 
2862 	/* Attach new eviction fence to all BOs except pinned ones */
2863 	list_for_each_entry(mem, &process_info->kfd_bo_list, validate_list) {
2864 		if (mem->bo->tbo.pin_count)
2865 			continue;
2866 
2867 		dma_resv_add_fence(mem->bo->tbo.base.resv,
2868 				   &process_info->eviction_fence->base,
2869 				   DMA_RESV_USAGE_BOOKKEEP);
2870 	}
2871 	/* Attach eviction fence to PD / PT BOs */
2872 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2873 			    vm_list_node) {
2874 		struct amdgpu_bo *bo = peer_vm->root.bo;
2875 
2876 		dma_resv_add_fence(bo->tbo.base.resv,
2877 				   &process_info->eviction_fence->base,
2878 				   DMA_RESV_USAGE_BOOKKEEP);
2879 	}
2880 
2881 validate_map_fail:
2882 	amdgpu_sync_free(&sync_obj);
2883 ttm_reserve_fail:
2884 	drm_exec_fini(&exec);
2885 	mutex_unlock(&process_info->lock);
2886 	return ret;
2887 }
2888 
2889 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
2890 {
2891 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2892 	struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
2893 	int ret;
2894 
2895 	if (!info || !gws)
2896 		return -EINVAL;
2897 
2898 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2899 	if (!*mem)
2900 		return -ENOMEM;
2901 
2902 	mutex_init(&(*mem)->lock);
2903 	INIT_LIST_HEAD(&(*mem)->attachments);
2904 	(*mem)->bo = amdgpu_bo_ref(gws_bo);
2905 	(*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
2906 	(*mem)->process_info = process_info;
2907 	add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
2908 	amdgpu_sync_create(&(*mem)->sync);
2909 
2910 
2911 	/* Validate gws bo the first time it is added to process */
2912 	mutex_lock(&(*mem)->process_info->lock);
2913 	ret = amdgpu_bo_reserve(gws_bo, false);
2914 	if (unlikely(ret)) {
2915 		pr_err("Reserve gws bo failed %d\n", ret);
2916 		goto bo_reservation_failure;
2917 	}
2918 
2919 	ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
2920 	if (ret) {
2921 		pr_err("GWS BO validate failed %d\n", ret);
2922 		goto bo_validation_failure;
2923 	}
2924 	/* GWS resource is shared b/t amdgpu and amdkfd
2925 	 * Add process eviction fence to bo so they can
2926 	 * evict each other.
2927 	 */
2928 	ret = dma_resv_reserve_fences(gws_bo->tbo.base.resv, 1);
2929 	if (ret)
2930 		goto reserve_shared_fail;
2931 	dma_resv_add_fence(gws_bo->tbo.base.resv,
2932 			   &process_info->eviction_fence->base,
2933 			   DMA_RESV_USAGE_BOOKKEEP);
2934 	amdgpu_bo_unreserve(gws_bo);
2935 	mutex_unlock(&(*mem)->process_info->lock);
2936 
2937 	return ret;
2938 
2939 reserve_shared_fail:
2940 bo_validation_failure:
2941 	amdgpu_bo_unreserve(gws_bo);
2942 bo_reservation_failure:
2943 	mutex_unlock(&(*mem)->process_info->lock);
2944 	amdgpu_sync_free(&(*mem)->sync);
2945 	remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
2946 	amdgpu_bo_unref(&gws_bo);
2947 	mutex_destroy(&(*mem)->lock);
2948 	kfree(*mem);
2949 	*mem = NULL;
2950 	return ret;
2951 }
2952 
2953 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
2954 {
2955 	int ret;
2956 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2957 	struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
2958 	struct amdgpu_bo *gws_bo = kgd_mem->bo;
2959 
2960 	/* Remove BO from process's validate list so restore worker won't touch
2961 	 * it anymore
2962 	 */
2963 	remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
2964 
2965 	ret = amdgpu_bo_reserve(gws_bo, false);
2966 	if (unlikely(ret)) {
2967 		pr_err("Reserve gws bo failed %d\n", ret);
2968 		//TODO add BO back to validate_list?
2969 		return ret;
2970 	}
2971 	amdgpu_amdkfd_remove_eviction_fence(gws_bo,
2972 			process_info->eviction_fence);
2973 	amdgpu_bo_unreserve(gws_bo);
2974 	amdgpu_sync_free(&kgd_mem->sync);
2975 	amdgpu_bo_unref(&gws_bo);
2976 	mutex_destroy(&kgd_mem->lock);
2977 	kfree(mem);
2978 	return 0;
2979 }
2980 
2981 /* Returns GPU-specific tiling mode information */
2982 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
2983 				struct tile_config *config)
2984 {
2985 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
2986 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
2987 	config->num_tile_configs =
2988 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
2989 	config->macro_tile_config_ptr =
2990 			adev->gfx.config.macrotile_mode_array;
2991 	config->num_macro_tile_configs =
2992 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
2993 
2994 	/* Those values are not set from GFX9 onwards */
2995 	config->num_banks = adev->gfx.config.num_banks;
2996 	config->num_ranks = adev->gfx.config.num_ranks;
2997 
2998 	return 0;
2999 }
3000 
3001 bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *mem)
3002 {
3003 	struct kfd_mem_attachment *entry;
3004 
3005 	list_for_each_entry(entry, &mem->attachments, list) {
3006 		if (entry->is_mapped && entry->adev == adev)
3007 			return true;
3008 	}
3009 	return false;
3010 }
3011 
3012 #if defined(CONFIG_DEBUG_FS)
3013 
3014 int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data)
3015 {
3016 
3017 	spin_lock(&kfd_mem_limit.mem_limit_lock);
3018 	seq_printf(m, "System mem used %lldM out of %lluM\n",
3019 		  (kfd_mem_limit.system_mem_used >> 20),
3020 		  (kfd_mem_limit.max_system_mem_limit >> 20));
3021 	seq_printf(m, "TTM mem used %lldM out of %lluM\n",
3022 		  (kfd_mem_limit.ttm_mem_used >> 20),
3023 		  (kfd_mem_limit.max_ttm_mem_limit >> 20));
3024 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
3025 
3026 	return 0;
3027 }
3028 
3029 #endif
3030