xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c (revision 3ba2a0bfd8cf94eb225e1c60dff16e5c35bde1da)
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 <linux/fdtable.h>
29 #include <drm/ttm/ttm_tt.h>
30 
31 #include <drm/drm_exec.h>
32 
33 #include "amdgpu_object.h"
34 #include "amdgpu_gem.h"
35 #include "amdgpu_vm.h"
36 #include "amdgpu_hmm.h"
37 #include "amdgpu_amdkfd.h"
38 #include "amdgpu_dma_buf.h"
39 #include <uapi/linux/kfd_ioctl.h>
40 #include "amdgpu_xgmi.h"
41 #include "kfd_priv.h"
42 #include "kfd_smi_events.h"
43 
44 /* Userptr restore delay, just long enough to allow consecutive VM
45  * changes to accumulate
46  */
47 #define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
48 #define AMDGPU_RESERVE_MEM_LIMIT			(3UL << 29)
49 
50 /*
51  * Align VRAM availability to 2MB to avoid fragmentation caused by 4K allocations in the tail 2MB
52  * BO chunk
53  */
54 #define VRAM_AVAILABLITY_ALIGN (1 << 21)
55 
56 /* Impose limit on how much memory KFD can use */
57 static struct {
58 	uint64_t max_system_mem_limit;
59 	uint64_t max_ttm_mem_limit;
60 	int64_t system_mem_used;
61 	int64_t ttm_mem_used;
62 	spinlock_t mem_limit_lock;
63 } kfd_mem_limit;
64 
65 static const char * const domain_bit_to_string[] = {
66 		"CPU",
67 		"GTT",
68 		"VRAM",
69 		"GDS",
70 		"GWS",
71 		"OA"
72 };
73 
74 #define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
75 
76 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work);
77 
78 static bool kfd_mem_is_attached(struct amdgpu_vm *avm,
79 		struct kgd_mem *mem)
80 {
81 	struct kfd_mem_attachment *entry;
82 
83 	list_for_each_entry(entry, &mem->attachments, list)
84 		if (entry->bo_va->base.vm == avm)
85 			return true;
86 
87 	return false;
88 }
89 
90 /**
91  * reuse_dmamap() - Check whether adev can share the original
92  * userptr BO
93  *
94  * If both adev and bo_adev are in direct mapping or
95  * in the same iommu group, they can share the original BO.
96  *
97  * @adev: Device to which can or cannot share the original BO
98  * @bo_adev: Device to which allocated BO belongs to
99  *
100  * Return: returns true if adev can share original userptr BO,
101  * false otherwise.
102  */
103 static bool reuse_dmamap(struct amdgpu_device *adev, struct amdgpu_device *bo_adev)
104 {
105 	return (adev->ram_is_direct_mapped && bo_adev->ram_is_direct_mapped) ||
106 			(adev->dev->iommu_group == bo_adev->dev->iommu_group);
107 }
108 
109 /* Set memory usage limits. Current, limits are
110  *  System (TTM + userptr) memory - 15/16th System RAM
111  *  TTM memory - 3/8th System RAM
112  */
113 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
114 {
115 	struct sysinfo si;
116 	uint64_t mem;
117 
118 	if (kfd_mem_limit.max_system_mem_limit)
119 		return;
120 
121 	si_meminfo(&si);
122 	mem = si.totalram - si.totalhigh;
123 	mem *= si.mem_unit;
124 
125 	spin_lock_init(&kfd_mem_limit.mem_limit_lock);
126 	kfd_mem_limit.max_system_mem_limit = mem - (mem >> 6);
127 	if (kfd_mem_limit.max_system_mem_limit < 2 * AMDGPU_RESERVE_MEM_LIMIT)
128 		kfd_mem_limit.max_system_mem_limit >>= 1;
129 	else
130 		kfd_mem_limit.max_system_mem_limit -= AMDGPU_RESERVE_MEM_LIMIT;
131 
132 	kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() << PAGE_SHIFT;
133 	pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
134 		(kfd_mem_limit.max_system_mem_limit >> 20),
135 		(kfd_mem_limit.max_ttm_mem_limit >> 20));
136 }
137 
138 void amdgpu_amdkfd_reserve_system_mem(uint64_t size)
139 {
140 	kfd_mem_limit.system_mem_used += size;
141 }
142 
143 /* Estimate page table size needed to represent a given memory size
144  *
145  * With 4KB pages, we need one 8 byte PTE for each 4KB of memory
146  * (factor 512, >> 9). With 2MB pages, we need one 8 byte PTE for 2MB
147  * of memory (factor 256K, >> 18). ROCm user mode tries to optimize
148  * for 2MB pages for TLB efficiency. However, small allocations and
149  * fragmented system memory still need some 4KB pages. We choose a
150  * compromise that should work in most cases without reserving too
151  * much memory for page tables unnecessarily (factor 16K, >> 14).
152  */
153 
154 #define ESTIMATE_PT_SIZE(mem_size) max(((mem_size) >> 14), AMDGPU_VM_RESERVED_VRAM)
155 
156 /**
157  * amdgpu_amdkfd_reserve_mem_limit() - Decrease available memory by size
158  * of buffer.
159  *
160  * @adev: Device to which allocated BO belongs to
161  * @size: Size of buffer, in bytes, encapsulated by B0. This should be
162  * equivalent to amdgpu_bo_size(BO)
163  * @alloc_flag: Flag used in allocating a BO as noted above
164  * @xcp_id: xcp_id is used to get xcp from xcp manager, one xcp is
165  * managed as one compute node in driver for app
166  *
167  * Return:
168  *	returns -ENOMEM in case of error, ZERO otherwise
169  */
170 int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
171 		uint64_t size, u32 alloc_flag, int8_t xcp_id)
172 {
173 	uint64_t reserved_for_pt =
174 		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
175 	size_t system_mem_needed, ttm_mem_needed, vram_needed;
176 	int ret = 0;
177 	uint64_t vram_size = 0;
178 
179 	system_mem_needed = 0;
180 	ttm_mem_needed = 0;
181 	vram_needed = 0;
182 	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
183 		system_mem_needed = size;
184 		ttm_mem_needed = size;
185 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
186 		/*
187 		 * Conservatively round up the allocation requirement to 2 MB
188 		 * to avoid fragmentation caused by 4K allocations in the tail
189 		 * 2M BO chunk.
190 		 */
191 		vram_needed = size;
192 		/*
193 		 * For GFX 9.4.3, get the VRAM size from XCP structs
194 		 */
195 		if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
196 			return -EINVAL;
197 
198 		vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id);
199 		if (adev->gmc.is_app_apu) {
200 			system_mem_needed = size;
201 			ttm_mem_needed = size;
202 		}
203 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
204 		system_mem_needed = size;
205 	} else if (!(alloc_flag &
206 				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
207 				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
208 		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
209 		return -ENOMEM;
210 	}
211 
212 	spin_lock(&kfd_mem_limit.mem_limit_lock);
213 
214 	if (kfd_mem_limit.system_mem_used + system_mem_needed >
215 	    kfd_mem_limit.max_system_mem_limit)
216 		pr_debug("Set no_system_mem_limit=1 if using shared memory\n");
217 
218 	if ((kfd_mem_limit.system_mem_used + system_mem_needed >
219 	     kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) ||
220 	    (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
221 	     kfd_mem_limit.max_ttm_mem_limit) ||
222 	    (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] + vram_needed >
223 	     vram_size - reserved_for_pt)) {
224 		ret = -ENOMEM;
225 		goto release;
226 	}
227 
228 	/* Update memory accounting by decreasing available system
229 	 * memory, TTM memory and GPU memory as computed above
230 	 */
231 	WARN_ONCE(vram_needed && !adev,
232 		  "adev reference can't be null when vram is used");
233 	if (adev && xcp_id >= 0) {
234 		adev->kfd.vram_used[xcp_id] += vram_needed;
235 		adev->kfd.vram_used_aligned[xcp_id] += adev->gmc.is_app_apu ?
236 				vram_needed :
237 				ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
238 	}
239 	kfd_mem_limit.system_mem_used += system_mem_needed;
240 	kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
241 
242 release:
243 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
244 	return ret;
245 }
246 
247 void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
248 		uint64_t size, u32 alloc_flag, int8_t xcp_id)
249 {
250 	spin_lock(&kfd_mem_limit.mem_limit_lock);
251 
252 	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
253 		kfd_mem_limit.system_mem_used -= size;
254 		kfd_mem_limit.ttm_mem_used -= size;
255 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
256 		WARN_ONCE(!adev,
257 			  "adev reference can't be null when alloc mem flags vram is set");
258 		if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
259 			goto release;
260 
261 		if (adev) {
262 			adev->kfd.vram_used[xcp_id] -= size;
263 			if (adev->gmc.is_app_apu) {
264 				adev->kfd.vram_used_aligned[xcp_id] -= size;
265 				kfd_mem_limit.system_mem_used -= size;
266 				kfd_mem_limit.ttm_mem_used -= size;
267 			} else {
268 				adev->kfd.vram_used_aligned[xcp_id] -=
269 					ALIGN(size, VRAM_AVAILABLITY_ALIGN);
270 			}
271 		}
272 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
273 		kfd_mem_limit.system_mem_used -= size;
274 	} else if (!(alloc_flag &
275 				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
276 				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
277 		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
278 		goto release;
279 	}
280 	WARN_ONCE(adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] < 0,
281 		  "KFD VRAM memory accounting unbalanced for xcp: %d", xcp_id);
282 	WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
283 		  "KFD TTM memory accounting unbalanced");
284 	WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
285 		  "KFD system memory accounting unbalanced");
286 
287 release:
288 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
289 }
290 
291 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
292 {
293 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
294 	u32 alloc_flags = bo->kfd_bo->alloc_flags;
295 	u64 size = amdgpu_bo_size(bo);
296 
297 	amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags,
298 					  bo->xcp_id);
299 
300 	kfree(bo->kfd_bo);
301 }
302 
303 /**
304  * create_dmamap_sg_bo() - Creates a amdgpu_bo object to reflect information
305  * about USERPTR or DOOREBELL or MMIO BO.
306  *
307  * @adev: Device for which dmamap BO is being created
308  * @mem: BO of peer device that is being DMA mapped. Provides parameters
309  *	 in building the dmamap BO
310  * @bo_out: Output parameter updated with handle of dmamap BO
311  */
312 static int
313 create_dmamap_sg_bo(struct amdgpu_device *adev,
314 		 struct kgd_mem *mem, struct amdgpu_bo **bo_out)
315 {
316 	struct drm_gem_object *gem_obj;
317 	int ret;
318 	uint64_t flags = 0;
319 
320 	ret = amdgpu_bo_reserve(mem->bo, false);
321 	if (ret)
322 		return ret;
323 
324 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)
325 		flags |= mem->bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
326 					AMDGPU_GEM_CREATE_UNCACHED);
327 
328 	ret = amdgpu_gem_object_create(adev, mem->bo->tbo.base.size, 1,
329 			AMDGPU_GEM_DOMAIN_CPU, AMDGPU_GEM_CREATE_PREEMPTIBLE | flags,
330 			ttm_bo_type_sg, mem->bo->tbo.base.resv, &gem_obj, 0);
331 
332 	amdgpu_bo_unreserve(mem->bo);
333 
334 	if (ret) {
335 		pr_err("Error in creating DMA mappable SG BO on domain: %d\n", ret);
336 		return -EINVAL;
337 	}
338 
339 	*bo_out = gem_to_amdgpu_bo(gem_obj);
340 	(*bo_out)->parent = amdgpu_bo_ref(mem->bo);
341 	return ret;
342 }
343 
344 /* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
345  *  reservation object.
346  *
347  * @bo: [IN] Remove eviction fence(s) from this BO
348  * @ef: [IN] This eviction fence is removed if it
349  *  is present in the shared list.
350  *
351  * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
352  */
353 static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
354 					struct amdgpu_amdkfd_fence *ef)
355 {
356 	struct dma_fence *replacement;
357 
358 	if (!ef)
359 		return -EINVAL;
360 
361 	/* TODO: Instead of block before we should use the fence of the page
362 	 * table update and TLB flush here directly.
363 	 */
364 	replacement = dma_fence_get_stub();
365 	dma_resv_replace_fences(bo->tbo.base.resv, ef->base.context,
366 				replacement, DMA_RESV_USAGE_BOOKKEEP);
367 	dma_fence_put(replacement);
368 	return 0;
369 }
370 
371 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
372 {
373 	struct amdgpu_bo *root = bo;
374 	struct amdgpu_vm_bo_base *vm_bo;
375 	struct amdgpu_vm *vm;
376 	struct amdkfd_process_info *info;
377 	struct amdgpu_amdkfd_fence *ef;
378 	int ret;
379 
380 	/* we can always get vm_bo from root PD bo.*/
381 	while (root->parent)
382 		root = root->parent;
383 
384 	vm_bo = root->vm_bo;
385 	if (!vm_bo)
386 		return 0;
387 
388 	vm = vm_bo->vm;
389 	if (!vm)
390 		return 0;
391 
392 	info = vm->process_info;
393 	if (!info || !info->eviction_fence)
394 		return 0;
395 
396 	ef = container_of(dma_fence_get(&info->eviction_fence->base),
397 			struct amdgpu_amdkfd_fence, base);
398 
399 	BUG_ON(!dma_resv_trylock(bo->tbo.base.resv));
400 	ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef);
401 	dma_resv_unlock(bo->tbo.base.resv);
402 
403 	dma_fence_put(&ef->base);
404 	return ret;
405 }
406 
407 static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
408 				     bool wait)
409 {
410 	struct ttm_operation_ctx ctx = { false, false };
411 	int ret;
412 
413 	if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
414 		 "Called with userptr BO"))
415 		return -EINVAL;
416 
417 	amdgpu_bo_placement_from_domain(bo, domain);
418 
419 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
420 	if (ret)
421 		goto validate_fail;
422 	if (wait)
423 		amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
424 
425 validate_fail:
426 	return ret;
427 }
428 
429 static int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo,
430 					       uint32_t domain,
431 					       struct dma_fence *fence)
432 {
433 	int ret = amdgpu_bo_reserve(bo, false);
434 
435 	if (ret)
436 		return ret;
437 
438 	ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
439 	if (ret)
440 		goto unreserve_out;
441 
442 	ret = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
443 	if (ret)
444 		goto unreserve_out;
445 
446 	dma_resv_add_fence(bo->tbo.base.resv, fence,
447 			   DMA_RESV_USAGE_BOOKKEEP);
448 
449 unreserve_out:
450 	amdgpu_bo_unreserve(bo);
451 
452 	return ret;
453 }
454 
455 static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo)
456 {
457 	return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false);
458 }
459 
460 /* vm_validate_pt_pd_bos - Validate page table and directory BOs
461  *
462  * Page directories are not updated here because huge page handling
463  * during page table updates can invalidate page directory entries
464  * again. Page directories are only updated after updating page
465  * tables.
466  */
467 static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm)
468 {
469 	struct amdgpu_bo *pd = vm->root.bo;
470 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
471 	int ret;
472 
473 	ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate_vm_bo, NULL);
474 	if (ret) {
475 		pr_err("failed to validate PT BOs\n");
476 		return ret;
477 	}
478 
479 	vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo);
480 
481 	return 0;
482 }
483 
484 static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
485 {
486 	struct amdgpu_bo *pd = vm->root.bo;
487 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
488 	int ret;
489 
490 	ret = amdgpu_vm_update_pdes(adev, vm, false);
491 	if (ret)
492 		return ret;
493 
494 	return amdgpu_sync_fence(sync, vm->last_update);
495 }
496 
497 static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem)
498 {
499 	uint32_t mapping_flags = AMDGPU_VM_PAGE_READABLE |
500 				 AMDGPU_VM_MTYPE_DEFAULT;
501 
502 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE)
503 		mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
504 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE)
505 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
506 
507 	return amdgpu_gem_va_map_flags(adev, mapping_flags);
508 }
509 
510 /**
511  * create_sg_table() - Create an sg_table for a contiguous DMA addr range
512  * @addr: The starting address to point to
513  * @size: Size of memory area in bytes being pointed to
514  *
515  * Allocates an instance of sg_table and initializes it to point to memory
516  * area specified by input parameters. The address used to build is assumed
517  * to be DMA mapped, if needed.
518  *
519  * DOORBELL or MMIO BOs use only one scatterlist node in their sg_table
520  * because they are physically contiguous.
521  *
522  * Return: Initialized instance of SG Table or NULL
523  */
524 static struct sg_table *create_sg_table(uint64_t addr, uint32_t size)
525 {
526 	struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
527 
528 	if (!sg)
529 		return NULL;
530 	if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
531 		kfree(sg);
532 		return NULL;
533 	}
534 	sg_dma_address(sg->sgl) = addr;
535 	sg->sgl->length = size;
536 #ifdef CONFIG_NEED_SG_DMA_LENGTH
537 	sg->sgl->dma_length = size;
538 #endif
539 	return sg;
540 }
541 
542 static int
543 kfd_mem_dmamap_userptr(struct kgd_mem *mem,
544 		       struct kfd_mem_attachment *attachment)
545 {
546 	enum dma_data_direction direction =
547 		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
548 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
549 	struct ttm_operation_ctx ctx = {.interruptible = true};
550 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
551 	struct amdgpu_device *adev = attachment->adev;
552 	struct ttm_tt *src_ttm = mem->bo->tbo.ttm;
553 	struct ttm_tt *ttm = bo->tbo.ttm;
554 	int ret;
555 
556 	if (WARN_ON(ttm->num_pages != src_ttm->num_pages))
557 		return -EINVAL;
558 
559 	ttm->sg = kmalloc(sizeof(*ttm->sg), GFP_KERNEL);
560 	if (unlikely(!ttm->sg))
561 		return -ENOMEM;
562 
563 	/* Same sequence as in amdgpu_ttm_tt_pin_userptr */
564 	ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages,
565 					ttm->num_pages, 0,
566 					(u64)ttm->num_pages << PAGE_SHIFT,
567 					GFP_KERNEL);
568 	if (unlikely(ret))
569 		goto free_sg;
570 
571 	ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
572 	if (unlikely(ret))
573 		goto release_sg;
574 
575 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
576 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
577 	if (ret)
578 		goto unmap_sg;
579 
580 	return 0;
581 
582 unmap_sg:
583 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
584 release_sg:
585 	pr_err("DMA map userptr failed: %d\n", ret);
586 	sg_free_table(ttm->sg);
587 free_sg:
588 	kfree(ttm->sg);
589 	ttm->sg = NULL;
590 	return ret;
591 }
592 
593 static int
594 kfd_mem_dmamap_dmabuf(struct kfd_mem_attachment *attachment)
595 {
596 	struct ttm_operation_ctx ctx = {.interruptible = true};
597 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
598 	int ret;
599 
600 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
601 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
602 	if (ret)
603 		return ret;
604 
605 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
606 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
607 }
608 
609 /**
610  * kfd_mem_dmamap_sg_bo() - Create DMA mapped sg_table to access DOORBELL or MMIO BO
611  * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
612  * @attachment: Virtual address attachment of the BO on accessing device
613  *
614  * An access request from the device that owns DOORBELL does not require DMA mapping.
615  * This is because the request doesn't go through PCIe root complex i.e. it instead
616  * loops back. The need to DMA map arises only when accessing peer device's DOORBELL
617  *
618  * In contrast, all access requests for MMIO need to be DMA mapped without regard to
619  * device ownership. This is because access requests for MMIO go through PCIe root
620  * complex.
621  *
622  * This is accomplished in two steps:
623  *   - Obtain DMA mapped address of DOORBELL or MMIO memory that could be used
624  *         in updating requesting device's page table
625  *   - Signal TTM to mark memory pointed to by requesting device's BO as GPU
626  *         accessible. This allows an update of requesting device's page table
627  *         with entries associated with DOOREBELL or MMIO memory
628  *
629  * This method is invoked in the following contexts:
630  *   - Mapping of DOORBELL or MMIO BO of same or peer device
631  *   - Validating an evicted DOOREBELL or MMIO BO on device seeking access
632  *
633  * Return: ZERO if successful, NON-ZERO otherwise
634  */
635 static int
636 kfd_mem_dmamap_sg_bo(struct kgd_mem *mem,
637 		     struct kfd_mem_attachment *attachment)
638 {
639 	struct ttm_operation_ctx ctx = {.interruptible = true};
640 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
641 	struct amdgpu_device *adev = attachment->adev;
642 	struct ttm_tt *ttm = bo->tbo.ttm;
643 	enum dma_data_direction dir;
644 	dma_addr_t dma_addr;
645 	bool mmio;
646 	int ret;
647 
648 	/* Expect SG Table of dmapmap BO to be NULL */
649 	mmio = (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP);
650 	if (unlikely(ttm->sg)) {
651 		pr_err("SG Table of %d BO for peer device is UNEXPECTEDLY NON-NULL", mmio);
652 		return -EINVAL;
653 	}
654 
655 	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
656 			DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
657 	dma_addr = mem->bo->tbo.sg->sgl->dma_address;
658 	pr_debug("%d BO size: %d\n", mmio, mem->bo->tbo.sg->sgl->length);
659 	pr_debug("%d BO address before DMA mapping: %llx\n", mmio, dma_addr);
660 	dma_addr = dma_map_resource(adev->dev, dma_addr,
661 			mem->bo->tbo.sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
662 	ret = dma_mapping_error(adev->dev, dma_addr);
663 	if (unlikely(ret))
664 		return ret;
665 	pr_debug("%d BO address after DMA mapping: %llx\n", mmio, dma_addr);
666 
667 	ttm->sg = create_sg_table(dma_addr, mem->bo->tbo.sg->sgl->length);
668 	if (unlikely(!ttm->sg)) {
669 		ret = -ENOMEM;
670 		goto unmap_sg;
671 	}
672 
673 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
674 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
675 	if (unlikely(ret))
676 		goto free_sg;
677 
678 	return ret;
679 
680 free_sg:
681 	sg_free_table(ttm->sg);
682 	kfree(ttm->sg);
683 	ttm->sg = NULL;
684 unmap_sg:
685 	dma_unmap_resource(adev->dev, dma_addr, mem->bo->tbo.sg->sgl->length,
686 			   dir, DMA_ATTR_SKIP_CPU_SYNC);
687 	return ret;
688 }
689 
690 static int
691 kfd_mem_dmamap_attachment(struct kgd_mem *mem,
692 			  struct kfd_mem_attachment *attachment)
693 {
694 	switch (attachment->type) {
695 	case KFD_MEM_ATT_SHARED:
696 		return 0;
697 	case KFD_MEM_ATT_USERPTR:
698 		return kfd_mem_dmamap_userptr(mem, attachment);
699 	case KFD_MEM_ATT_DMABUF:
700 		return kfd_mem_dmamap_dmabuf(attachment);
701 	case KFD_MEM_ATT_SG:
702 		return kfd_mem_dmamap_sg_bo(mem, attachment);
703 	default:
704 		WARN_ON_ONCE(1);
705 	}
706 	return -EINVAL;
707 }
708 
709 static void
710 kfd_mem_dmaunmap_userptr(struct kgd_mem *mem,
711 			 struct kfd_mem_attachment *attachment)
712 {
713 	enum dma_data_direction direction =
714 		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
715 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
716 	struct ttm_operation_ctx ctx = {.interruptible = false};
717 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
718 	struct amdgpu_device *adev = attachment->adev;
719 	struct ttm_tt *ttm = bo->tbo.ttm;
720 
721 	if (unlikely(!ttm->sg))
722 		return;
723 
724 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
725 	ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
726 
727 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
728 	sg_free_table(ttm->sg);
729 	kfree(ttm->sg);
730 	ttm->sg = NULL;
731 }
732 
733 static void
734 kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment)
735 {
736 	/* This is a no-op. We don't want to trigger eviction fences when
737 	 * unmapping DMABufs. Therefore the invalidation (moving to system
738 	 * domain) is done in kfd_mem_dmamap_dmabuf.
739 	 */
740 }
741 
742 /**
743  * kfd_mem_dmaunmap_sg_bo() - Free DMA mapped sg_table of DOORBELL or MMIO BO
744  * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
745  * @attachment: Virtual address attachment of the BO on accessing device
746  *
747  * The method performs following steps:
748  *   - Signal TTM to mark memory pointed to by BO as GPU inaccessible
749  *   - Free SG Table that is used to encapsulate DMA mapped memory of
750  *          peer device's DOORBELL or MMIO memory
751  *
752  * This method is invoked in the following contexts:
753  *     UNMapping of DOORBELL or MMIO BO on a device having access to its memory
754  *     Eviction of DOOREBELL or MMIO BO on device having access to its memory
755  *
756  * Return: void
757  */
758 static void
759 kfd_mem_dmaunmap_sg_bo(struct kgd_mem *mem,
760 		       struct kfd_mem_attachment *attachment)
761 {
762 	struct ttm_operation_ctx ctx = {.interruptible = true};
763 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
764 	struct amdgpu_device *adev = attachment->adev;
765 	struct ttm_tt *ttm = bo->tbo.ttm;
766 	enum dma_data_direction dir;
767 
768 	if (unlikely(!ttm->sg)) {
769 		pr_debug("SG Table of BO is NULL");
770 		return;
771 	}
772 
773 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
774 	ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
775 
776 	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
777 				DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
778 	dma_unmap_resource(adev->dev, ttm->sg->sgl->dma_address,
779 			ttm->sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
780 	sg_free_table(ttm->sg);
781 	kfree(ttm->sg);
782 	ttm->sg = NULL;
783 	bo->tbo.sg = NULL;
784 }
785 
786 static void
787 kfd_mem_dmaunmap_attachment(struct kgd_mem *mem,
788 			    struct kfd_mem_attachment *attachment)
789 {
790 	switch (attachment->type) {
791 	case KFD_MEM_ATT_SHARED:
792 		break;
793 	case KFD_MEM_ATT_USERPTR:
794 		kfd_mem_dmaunmap_userptr(mem, attachment);
795 		break;
796 	case KFD_MEM_ATT_DMABUF:
797 		kfd_mem_dmaunmap_dmabuf(attachment);
798 		break;
799 	case KFD_MEM_ATT_SG:
800 		kfd_mem_dmaunmap_sg_bo(mem, attachment);
801 		break;
802 	default:
803 		WARN_ON_ONCE(1);
804 	}
805 }
806 
807 static int kfd_mem_export_dmabuf(struct kgd_mem *mem)
808 {
809 	if (!mem->dmabuf) {
810 		struct amdgpu_device *bo_adev;
811 		struct dma_buf *dmabuf;
812 		int r, fd;
813 
814 		bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
815 		r = drm_gem_prime_handle_to_fd(&bo_adev->ddev, bo_adev->kfd.client.file,
816 					       mem->gem_handle,
817 			mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
818 					       DRM_RDWR : 0, &fd);
819 		if (r)
820 			return r;
821 		dmabuf = dma_buf_get(fd);
822 		close_fd(fd);
823 		if (WARN_ON_ONCE(IS_ERR(dmabuf)))
824 			return PTR_ERR(dmabuf);
825 		mem->dmabuf = dmabuf;
826 	}
827 
828 	return 0;
829 }
830 
831 static int
832 kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem,
833 		      struct amdgpu_bo **bo)
834 {
835 	struct drm_gem_object *gobj;
836 	int ret;
837 
838 	ret = kfd_mem_export_dmabuf(mem);
839 	if (ret)
840 		return ret;
841 
842 	gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf);
843 	if (IS_ERR(gobj))
844 		return PTR_ERR(gobj);
845 
846 	*bo = gem_to_amdgpu_bo(gobj);
847 	(*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE;
848 
849 	return 0;
850 }
851 
852 /* kfd_mem_attach - Add a BO to a VM
853  *
854  * Everything that needs to bo done only once when a BO is first added
855  * to a VM. It can later be mapped and unmapped many times without
856  * repeating these steps.
857  *
858  * 0. Create BO for DMA mapping, if needed
859  * 1. Allocate and initialize BO VA entry data structure
860  * 2. Add BO to the VM
861  * 3. Determine ASIC-specific PTE flags
862  * 4. Alloc page tables and directories if needed
863  * 4a.  Validate new page tables and directories
864  */
865 static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
866 		struct amdgpu_vm *vm, bool is_aql)
867 {
868 	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
869 	unsigned long bo_size = mem->bo->tbo.base.size;
870 	uint64_t va = mem->va;
871 	struct kfd_mem_attachment *attachment[2] = {NULL, NULL};
872 	struct amdgpu_bo *bo[2] = {NULL, NULL};
873 	struct amdgpu_bo_va *bo_va;
874 	bool same_hive = false;
875 	int i, ret;
876 
877 	if (!va) {
878 		pr_err("Invalid VA when adding BO to VM\n");
879 		return -EINVAL;
880 	}
881 
882 	/* Determine access to VRAM, MMIO and DOORBELL BOs of peer devices
883 	 *
884 	 * The access path of MMIO and DOORBELL BOs of is always over PCIe.
885 	 * In contrast the access path of VRAM BOs depens upon the type of
886 	 * link that connects the peer device. Access over PCIe is allowed
887 	 * if peer device has large BAR. In contrast, access over xGMI is
888 	 * allowed for both small and large BAR configurations of peer device
889 	 */
890 	if ((adev != bo_adev && !adev->gmc.is_app_apu) &&
891 	    ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) ||
892 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) ||
893 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
894 		if (mem->domain == AMDGPU_GEM_DOMAIN_VRAM)
895 			same_hive = amdgpu_xgmi_same_hive(adev, bo_adev);
896 		if (!same_hive && !amdgpu_device_is_peer_accessible(bo_adev, adev))
897 			return -EINVAL;
898 	}
899 
900 	for (i = 0; i <= is_aql; i++) {
901 		attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL);
902 		if (unlikely(!attachment[i])) {
903 			ret = -ENOMEM;
904 			goto unwind;
905 		}
906 
907 		pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
908 			 va + bo_size, vm);
909 
910 		if ((adev == bo_adev && !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) ||
911 		    (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && reuse_dmamap(adev, bo_adev)) ||
912 		    (mem->domain == AMDGPU_GEM_DOMAIN_GTT && reuse_dmamap(adev, bo_adev)) ||
913 		    same_hive) {
914 			/* Mappings on the local GPU, or VRAM mappings in the
915 			 * local hive, or userptr, or GTT mapping can reuse dma map
916 			 * address space share the original BO
917 			 */
918 			attachment[i]->type = KFD_MEM_ATT_SHARED;
919 			bo[i] = mem->bo;
920 			drm_gem_object_get(&bo[i]->tbo.base);
921 		} else if (i > 0) {
922 			/* Multiple mappings on the same GPU share the BO */
923 			attachment[i]->type = KFD_MEM_ATT_SHARED;
924 			bo[i] = bo[0];
925 			drm_gem_object_get(&bo[i]->tbo.base);
926 		} else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
927 			/* Create an SG BO to DMA-map userptrs on other GPUs */
928 			attachment[i]->type = KFD_MEM_ATT_USERPTR;
929 			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
930 			if (ret)
931 				goto unwind;
932 		/* Handle DOORBELL BOs of peer devices and MMIO BOs of local and peer devices */
933 		} else if (mem->bo->tbo.type == ttm_bo_type_sg) {
934 			WARN_ONCE(!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL ||
935 				    mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP),
936 				  "Handing invalid SG BO in ATTACH request");
937 			attachment[i]->type = KFD_MEM_ATT_SG;
938 			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
939 			if (ret)
940 				goto unwind;
941 		/* Enable acces to GTT and VRAM BOs of peer devices */
942 		} else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT ||
943 			   mem->domain == AMDGPU_GEM_DOMAIN_VRAM) {
944 			attachment[i]->type = KFD_MEM_ATT_DMABUF;
945 			ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]);
946 			if (ret)
947 				goto unwind;
948 			pr_debug("Employ DMABUF mechanism to enable peer GPU access\n");
949 		} else {
950 			WARN_ONCE(true, "Handling invalid ATTACH request");
951 			ret = -EINVAL;
952 			goto unwind;
953 		}
954 
955 		/* Add BO to VM internal data structures */
956 		ret = amdgpu_bo_reserve(bo[i], false);
957 		if (ret) {
958 			pr_debug("Unable to reserve BO during memory attach");
959 			goto unwind;
960 		}
961 		bo_va = amdgpu_vm_bo_find(vm, bo[i]);
962 		if (!bo_va)
963 			bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]);
964 		else
965 			++bo_va->ref_count;
966 		attachment[i]->bo_va = bo_va;
967 		amdgpu_bo_unreserve(bo[i]);
968 		if (unlikely(!attachment[i]->bo_va)) {
969 			ret = -ENOMEM;
970 			pr_err("Failed to add BO object to VM. ret == %d\n",
971 			       ret);
972 			goto unwind;
973 		}
974 		attachment[i]->va = va;
975 		attachment[i]->pte_flags = get_pte_flags(adev, mem);
976 		attachment[i]->adev = adev;
977 		list_add(&attachment[i]->list, &mem->attachments);
978 
979 		va += bo_size;
980 	}
981 
982 	return 0;
983 
984 unwind:
985 	for (; i >= 0; i--) {
986 		if (!attachment[i])
987 			continue;
988 		if (attachment[i]->bo_va) {
989 			amdgpu_bo_reserve(bo[i], true);
990 			if (--attachment[i]->bo_va->ref_count == 0)
991 				amdgpu_vm_bo_del(adev, attachment[i]->bo_va);
992 			amdgpu_bo_unreserve(bo[i]);
993 			list_del(&attachment[i]->list);
994 		}
995 		if (bo[i])
996 			drm_gem_object_put(&bo[i]->tbo.base);
997 		kfree(attachment[i]);
998 	}
999 	return ret;
1000 }
1001 
1002 static void kfd_mem_detach(struct kfd_mem_attachment *attachment)
1003 {
1004 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
1005 
1006 	pr_debug("\t remove VA 0x%llx in entry %p\n",
1007 			attachment->va, attachment);
1008 	if (--attachment->bo_va->ref_count == 0)
1009 		amdgpu_vm_bo_del(attachment->adev, attachment->bo_va);
1010 	drm_gem_object_put(&bo->tbo.base);
1011 	list_del(&attachment->list);
1012 	kfree(attachment);
1013 }
1014 
1015 static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
1016 				struct amdkfd_process_info *process_info,
1017 				bool userptr)
1018 {
1019 	mutex_lock(&process_info->lock);
1020 	if (userptr)
1021 		list_add_tail(&mem->validate_list,
1022 			      &process_info->userptr_valid_list);
1023 	else
1024 		list_add_tail(&mem->validate_list, &process_info->kfd_bo_list);
1025 	mutex_unlock(&process_info->lock);
1026 }
1027 
1028 static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
1029 		struct amdkfd_process_info *process_info)
1030 {
1031 	mutex_lock(&process_info->lock);
1032 	list_del(&mem->validate_list);
1033 	mutex_unlock(&process_info->lock);
1034 }
1035 
1036 /* Initializes user pages. It registers the MMU notifier and validates
1037  * the userptr BO in the GTT domain.
1038  *
1039  * The BO must already be on the userptr_valid_list. Otherwise an
1040  * eviction and restore may happen that leaves the new BO unmapped
1041  * with the user mode queues running.
1042  *
1043  * Takes the process_info->lock to protect against concurrent restore
1044  * workers.
1045  *
1046  * Returns 0 for success, negative errno for errors.
1047  */
1048 static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr,
1049 			   bool criu_resume)
1050 {
1051 	struct amdkfd_process_info *process_info = mem->process_info;
1052 	struct amdgpu_bo *bo = mem->bo;
1053 	struct ttm_operation_ctx ctx = { true, false };
1054 	struct hmm_range *range;
1055 	int ret = 0;
1056 
1057 	mutex_lock(&process_info->lock);
1058 
1059 	ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
1060 	if (ret) {
1061 		pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
1062 		goto out;
1063 	}
1064 
1065 	ret = amdgpu_hmm_register(bo, user_addr);
1066 	if (ret) {
1067 		pr_err("%s: Failed to register MMU notifier: %d\n",
1068 		       __func__, ret);
1069 		goto out;
1070 	}
1071 
1072 	if (criu_resume) {
1073 		/*
1074 		 * During a CRIU restore operation, the userptr buffer objects
1075 		 * will be validated in the restore_userptr_work worker at a
1076 		 * later stage when it is scheduled by another ioctl called by
1077 		 * CRIU master process for the target pid for restore.
1078 		 */
1079 		mutex_lock(&process_info->notifier_lock);
1080 		mem->invalid++;
1081 		mutex_unlock(&process_info->notifier_lock);
1082 		mutex_unlock(&process_info->lock);
1083 		return 0;
1084 	}
1085 
1086 	ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, &range);
1087 	if (ret) {
1088 		pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
1089 		goto unregister_out;
1090 	}
1091 
1092 	ret = amdgpu_bo_reserve(bo, true);
1093 	if (ret) {
1094 		pr_err("%s: Failed to reserve BO\n", __func__);
1095 		goto release_out;
1096 	}
1097 	amdgpu_bo_placement_from_domain(bo, mem->domain);
1098 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1099 	if (ret)
1100 		pr_err("%s: failed to validate BO\n", __func__);
1101 	amdgpu_bo_unreserve(bo);
1102 
1103 release_out:
1104 	amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range);
1105 unregister_out:
1106 	if (ret)
1107 		amdgpu_hmm_unregister(bo);
1108 out:
1109 	mutex_unlock(&process_info->lock);
1110 	return ret;
1111 }
1112 
1113 /* Reserving a BO and its page table BOs must happen atomically to
1114  * avoid deadlocks. Some operations update multiple VMs at once. Track
1115  * all the reservation info in a context structure. Optionally a sync
1116  * object can track VM updates.
1117  */
1118 struct bo_vm_reservation_context {
1119 	/* DRM execution context for the reservation */
1120 	struct drm_exec exec;
1121 	/* Number of VMs reserved */
1122 	unsigned int n_vms;
1123 	/* Pointer to sync object */
1124 	struct amdgpu_sync *sync;
1125 };
1126 
1127 enum bo_vm_match {
1128 	BO_VM_NOT_MAPPED = 0,	/* Match VMs where a BO is not mapped */
1129 	BO_VM_MAPPED,		/* Match VMs where a BO is mapped     */
1130 	BO_VM_ALL,		/* Match all VMs a BO was added to    */
1131 };
1132 
1133 /**
1134  * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
1135  * @mem: KFD BO structure.
1136  * @vm: the VM to reserve.
1137  * @ctx: the struct that will be used in unreserve_bo_and_vms().
1138  */
1139 static int reserve_bo_and_vm(struct kgd_mem *mem,
1140 			      struct amdgpu_vm *vm,
1141 			      struct bo_vm_reservation_context *ctx)
1142 {
1143 	struct amdgpu_bo *bo = mem->bo;
1144 	int ret;
1145 
1146 	WARN_ON(!vm);
1147 
1148 	ctx->n_vms = 1;
1149 	ctx->sync = &mem->sync;
1150 	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
1151 	drm_exec_until_all_locked(&ctx->exec) {
1152 		ret = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
1153 		drm_exec_retry_on_contention(&ctx->exec);
1154 		if (unlikely(ret))
1155 			goto error;
1156 
1157 		ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
1158 		drm_exec_retry_on_contention(&ctx->exec);
1159 		if (unlikely(ret))
1160 			goto error;
1161 	}
1162 	return 0;
1163 
1164 error:
1165 	pr_err("Failed to reserve buffers in ttm.\n");
1166 	drm_exec_fini(&ctx->exec);
1167 	return ret;
1168 }
1169 
1170 /**
1171  * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
1172  * @mem: KFD BO structure.
1173  * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
1174  * is used. Otherwise, a single VM associated with the BO.
1175  * @map_type: the mapping status that will be used to filter the VMs.
1176  * @ctx: the struct that will be used in unreserve_bo_and_vms().
1177  *
1178  * Returns 0 for success, negative for failure.
1179  */
1180 static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
1181 				struct amdgpu_vm *vm, enum bo_vm_match map_type,
1182 				struct bo_vm_reservation_context *ctx)
1183 {
1184 	struct kfd_mem_attachment *entry;
1185 	struct amdgpu_bo *bo = mem->bo;
1186 	int ret;
1187 
1188 	ctx->sync = &mem->sync;
1189 	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
1190 	drm_exec_until_all_locked(&ctx->exec) {
1191 		ctx->n_vms = 0;
1192 		list_for_each_entry(entry, &mem->attachments, list) {
1193 			if ((vm && vm != entry->bo_va->base.vm) ||
1194 				(entry->is_mapped != map_type
1195 				&& map_type != BO_VM_ALL))
1196 				continue;
1197 
1198 			ret = amdgpu_vm_lock_pd(entry->bo_va->base.vm,
1199 						&ctx->exec, 2);
1200 			drm_exec_retry_on_contention(&ctx->exec);
1201 			if (unlikely(ret))
1202 				goto error;
1203 			++ctx->n_vms;
1204 		}
1205 
1206 		ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
1207 		drm_exec_retry_on_contention(&ctx->exec);
1208 		if (unlikely(ret))
1209 			goto error;
1210 	}
1211 	return 0;
1212 
1213 error:
1214 	pr_err("Failed to reserve buffers in ttm.\n");
1215 	drm_exec_fini(&ctx->exec);
1216 	return ret;
1217 }
1218 
1219 /**
1220  * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1221  * @ctx: Reservation context to unreserve
1222  * @wait: Optionally wait for a sync object representing pending VM updates
1223  * @intr: Whether the wait is interruptible
1224  *
1225  * Also frees any resources allocated in
1226  * reserve_bo_and_(cond_)vm(s). Returns the status from
1227  * amdgpu_sync_wait.
1228  */
1229 static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
1230 				 bool wait, bool intr)
1231 {
1232 	int ret = 0;
1233 
1234 	if (wait)
1235 		ret = amdgpu_sync_wait(ctx->sync, intr);
1236 
1237 	drm_exec_fini(&ctx->exec);
1238 	ctx->sync = NULL;
1239 	return ret;
1240 }
1241 
1242 static void unmap_bo_from_gpuvm(struct kgd_mem *mem,
1243 				struct kfd_mem_attachment *entry,
1244 				struct amdgpu_sync *sync)
1245 {
1246 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1247 	struct amdgpu_device *adev = entry->adev;
1248 	struct amdgpu_vm *vm = bo_va->base.vm;
1249 
1250 	amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
1251 
1252 	amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
1253 
1254 	amdgpu_sync_fence(sync, bo_va->last_pt_update);
1255 }
1256 
1257 static int update_gpuvm_pte(struct kgd_mem *mem,
1258 			    struct kfd_mem_attachment *entry,
1259 			    struct amdgpu_sync *sync)
1260 {
1261 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1262 	struct amdgpu_device *adev = entry->adev;
1263 	int ret;
1264 
1265 	ret = kfd_mem_dmamap_attachment(mem, entry);
1266 	if (ret)
1267 		return ret;
1268 
1269 	/* Update the page tables  */
1270 	ret = amdgpu_vm_bo_update(adev, bo_va, false);
1271 	if (ret) {
1272 		pr_err("amdgpu_vm_bo_update failed\n");
1273 		return ret;
1274 	}
1275 
1276 	return amdgpu_sync_fence(sync, bo_va->last_pt_update);
1277 }
1278 
1279 static int map_bo_to_gpuvm(struct kgd_mem *mem,
1280 			   struct kfd_mem_attachment *entry,
1281 			   struct amdgpu_sync *sync,
1282 			   bool no_update_pte)
1283 {
1284 	int ret;
1285 
1286 	/* Set virtual address for the allocation */
1287 	ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0,
1288 			       amdgpu_bo_size(entry->bo_va->base.bo),
1289 			       entry->pte_flags);
1290 	if (ret) {
1291 		pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
1292 				entry->va, ret);
1293 		return ret;
1294 	}
1295 
1296 	if (no_update_pte)
1297 		return 0;
1298 
1299 	ret = update_gpuvm_pte(mem, entry, sync);
1300 	if (ret) {
1301 		pr_err("update_gpuvm_pte() failed\n");
1302 		goto update_gpuvm_pte_failed;
1303 	}
1304 
1305 	return 0;
1306 
1307 update_gpuvm_pte_failed:
1308 	unmap_bo_from_gpuvm(mem, entry, sync);
1309 	kfd_mem_dmaunmap_attachment(mem, entry);
1310 	return ret;
1311 }
1312 
1313 static int process_validate_vms(struct amdkfd_process_info *process_info)
1314 {
1315 	struct amdgpu_vm *peer_vm;
1316 	int ret;
1317 
1318 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1319 			    vm_list_node) {
1320 		ret = vm_validate_pt_pd_bos(peer_vm);
1321 		if (ret)
1322 			return ret;
1323 	}
1324 
1325 	return 0;
1326 }
1327 
1328 static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
1329 				 struct amdgpu_sync *sync)
1330 {
1331 	struct amdgpu_vm *peer_vm;
1332 	int ret;
1333 
1334 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1335 			    vm_list_node) {
1336 		struct amdgpu_bo *pd = peer_vm->root.bo;
1337 
1338 		ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
1339 				       AMDGPU_SYNC_NE_OWNER,
1340 				       AMDGPU_FENCE_OWNER_KFD);
1341 		if (ret)
1342 			return ret;
1343 	}
1344 
1345 	return 0;
1346 }
1347 
1348 static int process_update_pds(struct amdkfd_process_info *process_info,
1349 			      struct amdgpu_sync *sync)
1350 {
1351 	struct amdgpu_vm *peer_vm;
1352 	int ret;
1353 
1354 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1355 			    vm_list_node) {
1356 		ret = vm_update_pds(peer_vm, sync);
1357 		if (ret)
1358 			return ret;
1359 	}
1360 
1361 	return 0;
1362 }
1363 
1364 static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
1365 		       struct dma_fence **ef)
1366 {
1367 	struct amdkfd_process_info *info = NULL;
1368 	int ret;
1369 
1370 	if (!*process_info) {
1371 		info = kzalloc(sizeof(*info), GFP_KERNEL);
1372 		if (!info)
1373 			return -ENOMEM;
1374 
1375 		mutex_init(&info->lock);
1376 		mutex_init(&info->notifier_lock);
1377 		INIT_LIST_HEAD(&info->vm_list_head);
1378 		INIT_LIST_HEAD(&info->kfd_bo_list);
1379 		INIT_LIST_HEAD(&info->userptr_valid_list);
1380 		INIT_LIST_HEAD(&info->userptr_inval_list);
1381 
1382 		info->eviction_fence =
1383 			amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
1384 						   current->mm,
1385 						   NULL);
1386 		if (!info->eviction_fence) {
1387 			pr_err("Failed to create eviction fence\n");
1388 			ret = -ENOMEM;
1389 			goto create_evict_fence_fail;
1390 		}
1391 
1392 		info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
1393 		INIT_DELAYED_WORK(&info->restore_userptr_work,
1394 				  amdgpu_amdkfd_restore_userptr_worker);
1395 
1396 		*process_info = info;
1397 	}
1398 
1399 	vm->process_info = *process_info;
1400 
1401 	/* Validate page directory and attach eviction fence */
1402 	ret = amdgpu_bo_reserve(vm->root.bo, true);
1403 	if (ret)
1404 		goto reserve_pd_fail;
1405 	ret = vm_validate_pt_pd_bos(vm);
1406 	if (ret) {
1407 		pr_err("validate_pt_pd_bos() failed\n");
1408 		goto validate_pd_fail;
1409 	}
1410 	ret = amdgpu_bo_sync_wait(vm->root.bo,
1411 				  AMDGPU_FENCE_OWNER_KFD, false);
1412 	if (ret)
1413 		goto wait_pd_fail;
1414 	ret = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1);
1415 	if (ret)
1416 		goto reserve_shared_fail;
1417 	dma_resv_add_fence(vm->root.bo->tbo.base.resv,
1418 			   &vm->process_info->eviction_fence->base,
1419 			   DMA_RESV_USAGE_BOOKKEEP);
1420 	amdgpu_bo_unreserve(vm->root.bo);
1421 
1422 	/* Update process info */
1423 	mutex_lock(&vm->process_info->lock);
1424 	list_add_tail(&vm->vm_list_node,
1425 			&(vm->process_info->vm_list_head));
1426 	vm->process_info->n_vms++;
1427 
1428 	*ef = dma_fence_get(&vm->process_info->eviction_fence->base);
1429 	mutex_unlock(&vm->process_info->lock);
1430 
1431 	return 0;
1432 
1433 reserve_shared_fail:
1434 wait_pd_fail:
1435 validate_pd_fail:
1436 	amdgpu_bo_unreserve(vm->root.bo);
1437 reserve_pd_fail:
1438 	vm->process_info = NULL;
1439 	if (info) {
1440 		dma_fence_put(&info->eviction_fence->base);
1441 		*process_info = NULL;
1442 		put_pid(info->pid);
1443 create_evict_fence_fail:
1444 		mutex_destroy(&info->lock);
1445 		mutex_destroy(&info->notifier_lock);
1446 		kfree(info);
1447 	}
1448 	return ret;
1449 }
1450 
1451 /**
1452  * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
1453  * @bo: Handle of buffer object being pinned
1454  * @domain: Domain into which BO should be pinned
1455  *
1456  *   - USERPTR BOs are UNPINNABLE and will return error
1457  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1458  *     PIN count incremented. It is valid to PIN a BO multiple times
1459  *
1460  * Return: ZERO if successful in pinning, Non-Zero in case of error.
1461  */
1462 static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain)
1463 {
1464 	int ret = 0;
1465 
1466 	ret = amdgpu_bo_reserve(bo, false);
1467 	if (unlikely(ret))
1468 		return ret;
1469 
1470 	ret = amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1471 	if (ret)
1472 		pr_err("Error in Pinning BO to domain: %d\n", domain);
1473 
1474 	amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
1475 	amdgpu_bo_unreserve(bo);
1476 
1477 	return ret;
1478 }
1479 
1480 /**
1481  * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
1482  * @bo: Handle of buffer object being unpinned
1483  *
1484  *   - Is a illegal request for USERPTR BOs and is ignored
1485  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1486  *     PIN count decremented. Calls to UNPIN must balance calls to PIN
1487  */
1488 static void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo)
1489 {
1490 	int ret = 0;
1491 
1492 	ret = amdgpu_bo_reserve(bo, false);
1493 	if (unlikely(ret))
1494 		return;
1495 
1496 	amdgpu_bo_unpin(bo);
1497 	amdgpu_bo_unreserve(bo);
1498 }
1499 
1500 int amdgpu_amdkfd_gpuvm_set_vm_pasid(struct amdgpu_device *adev,
1501 				     struct amdgpu_vm *avm, u32 pasid)
1502 
1503 {
1504 	int ret;
1505 
1506 	/* Free the original amdgpu allocated pasid,
1507 	 * will be replaced with kfd allocated pasid.
1508 	 */
1509 	if (avm->pasid) {
1510 		amdgpu_pasid_free(avm->pasid);
1511 		amdgpu_vm_set_pasid(adev, avm, 0);
1512 	}
1513 
1514 	ret = amdgpu_vm_set_pasid(adev, avm, pasid);
1515 	if (ret)
1516 		return ret;
1517 
1518 	return 0;
1519 }
1520 
1521 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
1522 					   struct amdgpu_vm *avm,
1523 					   void **process_info,
1524 					   struct dma_fence **ef)
1525 {
1526 	int ret;
1527 
1528 	/* Already a compute VM? */
1529 	if (avm->process_info)
1530 		return -EINVAL;
1531 
1532 	/* Convert VM into a compute VM */
1533 	ret = amdgpu_vm_make_compute(adev, avm);
1534 	if (ret)
1535 		return ret;
1536 
1537 	/* Initialize KFD part of the VM and process info */
1538 	ret = init_kfd_vm(avm, process_info, ef);
1539 	if (ret)
1540 		return ret;
1541 
1542 	amdgpu_vm_set_task_info(avm);
1543 
1544 	return 0;
1545 }
1546 
1547 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1548 				    struct amdgpu_vm *vm)
1549 {
1550 	struct amdkfd_process_info *process_info = vm->process_info;
1551 
1552 	if (!process_info)
1553 		return;
1554 
1555 	/* Update process info */
1556 	mutex_lock(&process_info->lock);
1557 	process_info->n_vms--;
1558 	list_del(&vm->vm_list_node);
1559 	mutex_unlock(&process_info->lock);
1560 
1561 	vm->process_info = NULL;
1562 
1563 	/* Release per-process resources when last compute VM is destroyed */
1564 	if (!process_info->n_vms) {
1565 		WARN_ON(!list_empty(&process_info->kfd_bo_list));
1566 		WARN_ON(!list_empty(&process_info->userptr_valid_list));
1567 		WARN_ON(!list_empty(&process_info->userptr_inval_list));
1568 
1569 		dma_fence_put(&process_info->eviction_fence->base);
1570 		cancel_delayed_work_sync(&process_info->restore_userptr_work);
1571 		put_pid(process_info->pid);
1572 		mutex_destroy(&process_info->lock);
1573 		mutex_destroy(&process_info->notifier_lock);
1574 		kfree(process_info);
1575 	}
1576 }
1577 
1578 void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev,
1579 					    void *drm_priv)
1580 {
1581 	struct amdgpu_vm *avm;
1582 
1583 	if (WARN_ON(!adev || !drm_priv))
1584 		return;
1585 
1586 	avm = drm_priv_to_vm(drm_priv);
1587 
1588 	pr_debug("Releasing process vm %p\n", avm);
1589 
1590 	/* The original pasid of amdgpu vm has already been
1591 	 * released during making a amdgpu vm to a compute vm
1592 	 * The current pasid is managed by kfd and will be
1593 	 * released on kfd process destroy. Set amdgpu pasid
1594 	 * to 0 to avoid duplicate release.
1595 	 */
1596 	amdgpu_vm_release_compute(adev, avm);
1597 }
1598 
1599 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv)
1600 {
1601 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1602 	struct amdgpu_bo *pd = avm->root.bo;
1603 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1604 
1605 	if (adev->asic_type < CHIP_VEGA10)
1606 		return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1607 	return avm->pd_phys_addr;
1608 }
1609 
1610 void amdgpu_amdkfd_block_mmu_notifications(void *p)
1611 {
1612 	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1613 
1614 	mutex_lock(&pinfo->lock);
1615 	WRITE_ONCE(pinfo->block_mmu_notifications, true);
1616 	mutex_unlock(&pinfo->lock);
1617 }
1618 
1619 int amdgpu_amdkfd_criu_resume(void *p)
1620 {
1621 	int ret = 0;
1622 	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1623 
1624 	mutex_lock(&pinfo->lock);
1625 	pr_debug("scheduling work\n");
1626 	mutex_lock(&pinfo->notifier_lock);
1627 	pinfo->evicted_bos++;
1628 	mutex_unlock(&pinfo->notifier_lock);
1629 	if (!READ_ONCE(pinfo->block_mmu_notifications)) {
1630 		ret = -EINVAL;
1631 		goto out_unlock;
1632 	}
1633 	WRITE_ONCE(pinfo->block_mmu_notifications, false);
1634 	queue_delayed_work(system_freezable_wq,
1635 			   &pinfo->restore_userptr_work, 0);
1636 
1637 out_unlock:
1638 	mutex_unlock(&pinfo->lock);
1639 	return ret;
1640 }
1641 
1642 size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
1643 					  uint8_t xcp_id)
1644 {
1645 	uint64_t reserved_for_pt =
1646 		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
1647 	ssize_t available;
1648 	uint64_t vram_available, system_mem_available, ttm_mem_available;
1649 
1650 	spin_lock(&kfd_mem_limit.mem_limit_lock);
1651 	vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id)
1652 		- adev->kfd.vram_used_aligned[xcp_id]
1653 		- atomic64_read(&adev->vram_pin_size)
1654 		- reserved_for_pt;
1655 
1656 	if (adev->gmc.is_app_apu) {
1657 		system_mem_available = no_system_mem_limit ?
1658 					kfd_mem_limit.max_system_mem_limit :
1659 					kfd_mem_limit.max_system_mem_limit -
1660 					kfd_mem_limit.system_mem_used;
1661 
1662 		ttm_mem_available = kfd_mem_limit.max_ttm_mem_limit -
1663 				kfd_mem_limit.ttm_mem_used;
1664 
1665 		available = min3(system_mem_available, ttm_mem_available,
1666 				 vram_available);
1667 		available = ALIGN_DOWN(available, PAGE_SIZE);
1668 	} else {
1669 		available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN);
1670 	}
1671 
1672 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
1673 
1674 	if (available < 0)
1675 		available = 0;
1676 
1677 	return available;
1678 }
1679 
1680 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1681 		struct amdgpu_device *adev, uint64_t va, uint64_t size,
1682 		void *drm_priv, struct kgd_mem **mem,
1683 		uint64_t *offset, uint32_t flags, bool criu_resume)
1684 {
1685 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1686 	struct amdgpu_fpriv *fpriv = container_of(avm, struct amdgpu_fpriv, vm);
1687 	enum ttm_bo_type bo_type = ttm_bo_type_device;
1688 	struct sg_table *sg = NULL;
1689 	uint64_t user_addr = 0;
1690 	struct amdgpu_bo *bo;
1691 	struct drm_gem_object *gobj = NULL;
1692 	u32 domain, alloc_domain;
1693 	uint64_t aligned_size;
1694 	int8_t xcp_id = -1;
1695 	u64 alloc_flags;
1696 	int ret;
1697 
1698 	/*
1699 	 * Check on which domain to allocate BO
1700 	 */
1701 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1702 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1703 
1704 		if (adev->gmc.is_app_apu) {
1705 			domain = AMDGPU_GEM_DOMAIN_GTT;
1706 			alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1707 			alloc_flags = 0;
1708 		} else {
1709 			alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1710 			alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1711 			AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0;
1712 		}
1713 		xcp_id = fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION ?
1714 					0 : fpriv->xcp_id;
1715 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1716 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1717 		alloc_flags = 0;
1718 	} else {
1719 		domain = AMDGPU_GEM_DOMAIN_GTT;
1720 		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1721 		alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE;
1722 
1723 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1724 			if (!offset || !*offset)
1725 				return -EINVAL;
1726 			user_addr = untagged_addr(*offset);
1727 		} else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1728 				    KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1729 			bo_type = ttm_bo_type_sg;
1730 			if (size > UINT_MAX)
1731 				return -EINVAL;
1732 			sg = create_sg_table(*offset, size);
1733 			if (!sg)
1734 				return -ENOMEM;
1735 		} else {
1736 			return -EINVAL;
1737 		}
1738 	}
1739 
1740 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT)
1741 		alloc_flags |= AMDGPU_GEM_CREATE_COHERENT;
1742 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENT)
1743 		alloc_flags |= AMDGPU_GEM_CREATE_EXT_COHERENT;
1744 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED)
1745 		alloc_flags |= AMDGPU_GEM_CREATE_UNCACHED;
1746 
1747 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1748 	if (!*mem) {
1749 		ret = -ENOMEM;
1750 		goto err;
1751 	}
1752 	INIT_LIST_HEAD(&(*mem)->attachments);
1753 	mutex_init(&(*mem)->lock);
1754 	(*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1755 
1756 	/* Workaround for AQL queue wraparound bug. Map the same
1757 	 * memory twice. That means we only actually allocate half
1758 	 * the memory.
1759 	 */
1760 	if ((*mem)->aql_queue)
1761 		size >>= 1;
1762 	aligned_size = PAGE_ALIGN(size);
1763 
1764 	(*mem)->alloc_flags = flags;
1765 
1766 	amdgpu_sync_create(&(*mem)->sync);
1767 
1768 	ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags,
1769 					      xcp_id);
1770 	if (ret) {
1771 		pr_debug("Insufficient memory\n");
1772 		goto err_reserve_limit;
1773 	}
1774 
1775 	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s xcp_id %d\n",
1776 		 va, (*mem)->aql_queue ? size << 1 : size,
1777 		 domain_string(alloc_domain), xcp_id);
1778 
1779 	ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags,
1780 				       bo_type, NULL, &gobj, xcp_id + 1);
1781 	if (ret) {
1782 		pr_debug("Failed to create BO on domain %s. ret %d\n",
1783 			 domain_string(alloc_domain), ret);
1784 		goto err_bo_create;
1785 	}
1786 	ret = drm_vma_node_allow(&gobj->vma_node, drm_priv);
1787 	if (ret) {
1788 		pr_debug("Failed to allow vma node access. ret %d\n", ret);
1789 		goto err_node_allow;
1790 	}
1791 	ret = drm_gem_handle_create(adev->kfd.client.file, gobj, &(*mem)->gem_handle);
1792 	if (ret)
1793 		goto err_gem_handle_create;
1794 	bo = gem_to_amdgpu_bo(gobj);
1795 	if (bo_type == ttm_bo_type_sg) {
1796 		bo->tbo.sg = sg;
1797 		bo->tbo.ttm->sg = sg;
1798 	}
1799 	bo->kfd_bo = *mem;
1800 	(*mem)->bo = bo;
1801 	if (user_addr)
1802 		bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO;
1803 
1804 	(*mem)->va = va;
1805 	(*mem)->domain = domain;
1806 	(*mem)->mapped_to_gpu_memory = 0;
1807 	(*mem)->process_info = avm->process_info;
1808 
1809 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1810 
1811 	if (user_addr) {
1812 		pr_debug("creating userptr BO for user_addr = %llx\n", user_addr);
1813 		ret = init_user_pages(*mem, user_addr, criu_resume);
1814 		if (ret)
1815 			goto allocate_init_user_pages_failed;
1816 	} else  if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1817 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1818 		ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT);
1819 		if (ret) {
1820 			pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n");
1821 			goto err_pin_bo;
1822 		}
1823 		bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
1824 		bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
1825 	} else {
1826 		mutex_lock(&avm->process_info->lock);
1827 		if (avm->process_info->eviction_fence &&
1828 		    !dma_fence_is_signaled(&avm->process_info->eviction_fence->base))
1829 			ret = amdgpu_amdkfd_bo_validate_and_fence(bo, domain,
1830 				&avm->process_info->eviction_fence->base);
1831 		mutex_unlock(&avm->process_info->lock);
1832 		if (ret)
1833 			goto err_validate_bo;
1834 	}
1835 
1836 	if (offset)
1837 		*offset = amdgpu_bo_mmap_offset(bo);
1838 
1839 	return 0;
1840 
1841 allocate_init_user_pages_failed:
1842 err_pin_bo:
1843 err_validate_bo:
1844 	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1845 	drm_gem_handle_delete(adev->kfd.client.file, (*mem)->gem_handle);
1846 err_gem_handle_create:
1847 	drm_vma_node_revoke(&gobj->vma_node, drm_priv);
1848 err_node_allow:
1849 	/* Don't unreserve system mem limit twice */
1850 	goto err_reserve_limit;
1851 err_bo_create:
1852 	amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags, xcp_id);
1853 err_reserve_limit:
1854 	mutex_destroy(&(*mem)->lock);
1855 	if (gobj)
1856 		drm_gem_object_put(gobj);
1857 	else
1858 		kfree(*mem);
1859 err:
1860 	if (sg) {
1861 		sg_free_table(sg);
1862 		kfree(sg);
1863 	}
1864 	return ret;
1865 }
1866 
1867 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1868 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
1869 		uint64_t *size)
1870 {
1871 	struct amdkfd_process_info *process_info = mem->process_info;
1872 	unsigned long bo_size = mem->bo->tbo.base.size;
1873 	bool use_release_notifier = (mem->bo->kfd_bo == mem);
1874 	struct kfd_mem_attachment *entry, *tmp;
1875 	struct bo_vm_reservation_context ctx;
1876 	unsigned int mapped_to_gpu_memory;
1877 	int ret;
1878 	bool is_imported = false;
1879 
1880 	mutex_lock(&mem->lock);
1881 
1882 	/* Unpin MMIO/DOORBELL BO's that were pinned during allocation */
1883 	if (mem->alloc_flags &
1884 	    (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1885 	     KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1886 		amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo);
1887 	}
1888 
1889 	mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1890 	is_imported = mem->is_imported;
1891 	mutex_unlock(&mem->lock);
1892 	/* lock is not needed after this, since mem is unused and will
1893 	 * be freed anyway
1894 	 */
1895 
1896 	if (mapped_to_gpu_memory > 0) {
1897 		pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1898 				mem->va, bo_size);
1899 		return -EBUSY;
1900 	}
1901 
1902 	/* Make sure restore workers don't access the BO any more */
1903 	mutex_lock(&process_info->lock);
1904 	list_del(&mem->validate_list);
1905 	mutex_unlock(&process_info->lock);
1906 
1907 	/* Cleanup user pages and MMU notifiers */
1908 	if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
1909 		amdgpu_hmm_unregister(mem->bo);
1910 		mutex_lock(&process_info->notifier_lock);
1911 		amdgpu_ttm_tt_discard_user_pages(mem->bo->tbo.ttm, mem->range);
1912 		mutex_unlock(&process_info->notifier_lock);
1913 	}
1914 
1915 	ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1916 	if (unlikely(ret))
1917 		return ret;
1918 
1919 	amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1920 					process_info->eviction_fence);
1921 	pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1922 		mem->va + bo_size * (1 + mem->aql_queue));
1923 
1924 	/* Remove from VM internal data structures */
1925 	list_for_each_entry_safe(entry, tmp, &mem->attachments, list) {
1926 		kfd_mem_dmaunmap_attachment(mem, entry);
1927 		kfd_mem_detach(entry);
1928 	}
1929 
1930 	ret = unreserve_bo_and_vms(&ctx, false, false);
1931 
1932 	/* Free the sync object */
1933 	amdgpu_sync_free(&mem->sync);
1934 
1935 	/* If the SG is not NULL, it's one we created for a doorbell or mmio
1936 	 * remap BO. We need to free it.
1937 	 */
1938 	if (mem->bo->tbo.sg) {
1939 		sg_free_table(mem->bo->tbo.sg);
1940 		kfree(mem->bo->tbo.sg);
1941 	}
1942 
1943 	/* Update the size of the BO being freed if it was allocated from
1944 	 * VRAM and is not imported. For APP APU VRAM allocations are done
1945 	 * in GTT domain
1946 	 */
1947 	if (size) {
1948 		if (!is_imported &&
1949 		   (mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM ||
1950 		   (adev->gmc.is_app_apu &&
1951 		    mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT)))
1952 			*size = bo_size;
1953 		else
1954 			*size = 0;
1955 	}
1956 
1957 	/* Free the BO*/
1958 	drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
1959 	drm_gem_handle_delete(adev->kfd.client.file, mem->gem_handle);
1960 	if (mem->dmabuf) {
1961 		dma_buf_put(mem->dmabuf);
1962 		mem->dmabuf = NULL;
1963 	}
1964 	mutex_destroy(&mem->lock);
1965 
1966 	/* If this releases the last reference, it will end up calling
1967 	 * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why
1968 	 * this needs to be the last call here.
1969 	 */
1970 	drm_gem_object_put(&mem->bo->tbo.base);
1971 
1972 	/*
1973 	 * For kgd_mem allocated in amdgpu_amdkfd_gpuvm_import_dmabuf(),
1974 	 * explicitly free it here.
1975 	 */
1976 	if (!use_release_notifier)
1977 		kfree(mem);
1978 
1979 	return ret;
1980 }
1981 
1982 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1983 		struct amdgpu_device *adev, struct kgd_mem *mem,
1984 		void *drm_priv)
1985 {
1986 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1987 	int ret;
1988 	struct amdgpu_bo *bo;
1989 	uint32_t domain;
1990 	struct kfd_mem_attachment *entry;
1991 	struct bo_vm_reservation_context ctx;
1992 	unsigned long bo_size;
1993 	bool is_invalid_userptr = false;
1994 
1995 	bo = mem->bo;
1996 	if (!bo) {
1997 		pr_err("Invalid BO when mapping memory to GPU\n");
1998 		return -EINVAL;
1999 	}
2000 
2001 	/* Make sure restore is not running concurrently. Since we
2002 	 * don't map invalid userptr BOs, we rely on the next restore
2003 	 * worker to do the mapping
2004 	 */
2005 	mutex_lock(&mem->process_info->lock);
2006 
2007 	/* Lock notifier lock. If we find an invalid userptr BO, we can be
2008 	 * sure that the MMU notifier is no longer running
2009 	 * concurrently and the queues are actually stopped
2010 	 */
2011 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2012 		mutex_lock(&mem->process_info->notifier_lock);
2013 		is_invalid_userptr = !!mem->invalid;
2014 		mutex_unlock(&mem->process_info->notifier_lock);
2015 	}
2016 
2017 	mutex_lock(&mem->lock);
2018 
2019 	domain = mem->domain;
2020 	bo_size = bo->tbo.base.size;
2021 
2022 	pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
2023 			mem->va,
2024 			mem->va + bo_size * (1 + mem->aql_queue),
2025 			avm, domain_string(domain));
2026 
2027 	if (!kfd_mem_is_attached(avm, mem)) {
2028 		ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue);
2029 		if (ret)
2030 			goto out;
2031 	}
2032 
2033 	ret = reserve_bo_and_vm(mem, avm, &ctx);
2034 	if (unlikely(ret))
2035 		goto out;
2036 
2037 	/* Userptr can be marked as "not invalid", but not actually be
2038 	 * validated yet (still in the system domain). In that case
2039 	 * the queues are still stopped and we can leave mapping for
2040 	 * the next restore worker
2041 	 */
2042 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
2043 	    bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
2044 		is_invalid_userptr = true;
2045 
2046 	ret = vm_validate_pt_pd_bos(avm);
2047 	if (unlikely(ret))
2048 		goto out_unreserve;
2049 
2050 	list_for_each_entry(entry, &mem->attachments, list) {
2051 		if (entry->bo_va->base.vm != avm || entry->is_mapped)
2052 			continue;
2053 
2054 		pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
2055 			 entry->va, entry->va + bo_size, entry);
2056 
2057 		ret = map_bo_to_gpuvm(mem, entry, ctx.sync,
2058 				      is_invalid_userptr);
2059 		if (ret) {
2060 			pr_err("Failed to map bo to gpuvm\n");
2061 			goto out_unreserve;
2062 		}
2063 
2064 		ret = vm_update_pds(avm, ctx.sync);
2065 		if (ret) {
2066 			pr_err("Failed to update page directories\n");
2067 			goto out_unreserve;
2068 		}
2069 
2070 		entry->is_mapped = true;
2071 		mem->mapped_to_gpu_memory++;
2072 		pr_debug("\t INC mapping count %d\n",
2073 			 mem->mapped_to_gpu_memory);
2074 	}
2075 
2076 	ret = unreserve_bo_and_vms(&ctx, false, false);
2077 
2078 	goto out;
2079 
2080 out_unreserve:
2081 	unreserve_bo_and_vms(&ctx, false, false);
2082 out:
2083 	mutex_unlock(&mem->process_info->lock);
2084 	mutex_unlock(&mem->lock);
2085 	return ret;
2086 }
2087 
2088 void amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem, void *drm_priv)
2089 {
2090 	struct kfd_mem_attachment *entry;
2091 	struct amdgpu_vm *vm;
2092 
2093 	vm = drm_priv_to_vm(drm_priv);
2094 
2095 	mutex_lock(&mem->lock);
2096 
2097 	list_for_each_entry(entry, &mem->attachments, list) {
2098 		if (entry->bo_va->base.vm == vm)
2099 			kfd_mem_dmaunmap_attachment(mem, entry);
2100 	}
2101 
2102 	mutex_unlock(&mem->lock);
2103 }
2104 
2105 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
2106 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
2107 {
2108 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2109 	unsigned long bo_size = mem->bo->tbo.base.size;
2110 	struct kfd_mem_attachment *entry;
2111 	struct bo_vm_reservation_context ctx;
2112 	int ret;
2113 
2114 	mutex_lock(&mem->lock);
2115 
2116 	ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx);
2117 	if (unlikely(ret))
2118 		goto out;
2119 	/* If no VMs were reserved, it means the BO wasn't actually mapped */
2120 	if (ctx.n_vms == 0) {
2121 		ret = -EINVAL;
2122 		goto unreserve_out;
2123 	}
2124 
2125 	ret = vm_validate_pt_pd_bos(avm);
2126 	if (unlikely(ret))
2127 		goto unreserve_out;
2128 
2129 	pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
2130 		mem->va,
2131 		mem->va + bo_size * (1 + mem->aql_queue),
2132 		avm);
2133 
2134 	list_for_each_entry(entry, &mem->attachments, list) {
2135 		if (entry->bo_va->base.vm != avm || !entry->is_mapped)
2136 			continue;
2137 
2138 		pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
2139 			 entry->va, entry->va + bo_size, entry);
2140 
2141 		unmap_bo_from_gpuvm(mem, entry, ctx.sync);
2142 		entry->is_mapped = false;
2143 
2144 		mem->mapped_to_gpu_memory--;
2145 		pr_debug("\t DEC mapping count %d\n",
2146 			 mem->mapped_to_gpu_memory);
2147 	}
2148 
2149 unreserve_out:
2150 	unreserve_bo_and_vms(&ctx, false, false);
2151 out:
2152 	mutex_unlock(&mem->lock);
2153 	return ret;
2154 }
2155 
2156 int amdgpu_amdkfd_gpuvm_sync_memory(
2157 		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr)
2158 {
2159 	struct amdgpu_sync sync;
2160 	int ret;
2161 
2162 	amdgpu_sync_create(&sync);
2163 
2164 	mutex_lock(&mem->lock);
2165 	amdgpu_sync_clone(&mem->sync, &sync);
2166 	mutex_unlock(&mem->lock);
2167 
2168 	ret = amdgpu_sync_wait(&sync, intr);
2169 	amdgpu_sync_free(&sync);
2170 	return ret;
2171 }
2172 
2173 /**
2174  * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count
2175  * @adev: Device to which allocated BO belongs
2176  * @bo: Buffer object to be mapped
2177  *
2178  * Before return, bo reference count is incremented. To release the reference and unpin/
2179  * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
2180  */
2181 int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo)
2182 {
2183 	int ret;
2184 
2185 	ret = amdgpu_bo_reserve(bo, true);
2186 	if (ret) {
2187 		pr_err("Failed to reserve bo. ret %d\n", ret);
2188 		goto err_reserve_bo_failed;
2189 	}
2190 
2191 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2192 	if (ret) {
2193 		pr_err("Failed to pin bo. ret %d\n", ret);
2194 		goto err_pin_bo_failed;
2195 	}
2196 
2197 	ret = amdgpu_ttm_alloc_gart(&bo->tbo);
2198 	if (ret) {
2199 		pr_err("Failed to bind bo to GART. ret %d\n", ret);
2200 		goto err_map_bo_gart_failed;
2201 	}
2202 
2203 	amdgpu_amdkfd_remove_eviction_fence(
2204 		bo, bo->vm_bo->vm->process_info->eviction_fence);
2205 
2206 	amdgpu_bo_unreserve(bo);
2207 
2208 	bo = amdgpu_bo_ref(bo);
2209 
2210 	return 0;
2211 
2212 err_map_bo_gart_failed:
2213 	amdgpu_bo_unpin(bo);
2214 err_pin_bo_failed:
2215 	amdgpu_bo_unreserve(bo);
2216 err_reserve_bo_failed:
2217 
2218 	return ret;
2219 }
2220 
2221 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Map a GTT BO for kernel CPU access
2222  *
2223  * @mem: Buffer object to be mapped for CPU access
2224  * @kptr[out]: pointer in kernel CPU address space
2225  * @size[out]: size of the buffer
2226  *
2227  * Pins the BO and maps it for kernel CPU access. The eviction fence is removed
2228  * from the BO, since pinned BOs cannot be evicted. The bo must remain on the
2229  * validate_list, so the GPU mapping can be restored after a page table was
2230  * evicted.
2231  *
2232  * Return: 0 on success, error code on failure
2233  */
2234 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,
2235 					     void **kptr, uint64_t *size)
2236 {
2237 	int ret;
2238 	struct amdgpu_bo *bo = mem->bo;
2239 
2240 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2241 		pr_err("userptr can't be mapped to kernel\n");
2242 		return -EINVAL;
2243 	}
2244 
2245 	mutex_lock(&mem->process_info->lock);
2246 
2247 	ret = amdgpu_bo_reserve(bo, true);
2248 	if (ret) {
2249 		pr_err("Failed to reserve bo. ret %d\n", ret);
2250 		goto bo_reserve_failed;
2251 	}
2252 
2253 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2254 	if (ret) {
2255 		pr_err("Failed to pin bo. ret %d\n", ret);
2256 		goto pin_failed;
2257 	}
2258 
2259 	ret = amdgpu_bo_kmap(bo, kptr);
2260 	if (ret) {
2261 		pr_err("Failed to map bo to kernel. ret %d\n", ret);
2262 		goto kmap_failed;
2263 	}
2264 
2265 	amdgpu_amdkfd_remove_eviction_fence(
2266 		bo, mem->process_info->eviction_fence);
2267 
2268 	if (size)
2269 		*size = amdgpu_bo_size(bo);
2270 
2271 	amdgpu_bo_unreserve(bo);
2272 
2273 	mutex_unlock(&mem->process_info->lock);
2274 	return 0;
2275 
2276 kmap_failed:
2277 	amdgpu_bo_unpin(bo);
2278 pin_failed:
2279 	amdgpu_bo_unreserve(bo);
2280 bo_reserve_failed:
2281 	mutex_unlock(&mem->process_info->lock);
2282 
2283 	return ret;
2284 }
2285 
2286 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Unmap a GTT BO for kernel CPU access
2287  *
2288  * @mem: Buffer object to be unmapped for CPU access
2289  *
2290  * Removes the kernel CPU mapping and unpins the BO. It does not restore the
2291  * eviction fence, so this function should only be used for cleanup before the
2292  * BO is destroyed.
2293  */
2294 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem)
2295 {
2296 	struct amdgpu_bo *bo = mem->bo;
2297 
2298 	amdgpu_bo_reserve(bo, true);
2299 	amdgpu_bo_kunmap(bo);
2300 	amdgpu_bo_unpin(bo);
2301 	amdgpu_bo_unreserve(bo);
2302 }
2303 
2304 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
2305 					  struct kfd_vm_fault_info *mem)
2306 {
2307 	if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
2308 		*mem = *adev->gmc.vm_fault_info;
2309 		mb(); /* make sure read happened */
2310 		atomic_set(&adev->gmc.vm_fault_info_updated, 0);
2311 	}
2312 	return 0;
2313 }
2314 
2315 static int import_obj_create(struct amdgpu_device *adev,
2316 			     struct dma_buf *dma_buf,
2317 			     struct drm_gem_object *obj,
2318 			     uint64_t va, void *drm_priv,
2319 			     struct kgd_mem **mem, uint64_t *size,
2320 			     uint64_t *mmap_offset)
2321 {
2322 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2323 	struct amdgpu_bo *bo;
2324 	int ret;
2325 
2326 	bo = gem_to_amdgpu_bo(obj);
2327 	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
2328 				    AMDGPU_GEM_DOMAIN_GTT)))
2329 		/* Only VRAM and GTT BOs are supported */
2330 		return -EINVAL;
2331 
2332 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2333 	if (!*mem)
2334 		return -ENOMEM;
2335 
2336 	ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
2337 	if (ret)
2338 		goto err_free_mem;
2339 
2340 	if (size)
2341 		*size = amdgpu_bo_size(bo);
2342 
2343 	if (mmap_offset)
2344 		*mmap_offset = amdgpu_bo_mmap_offset(bo);
2345 
2346 	INIT_LIST_HEAD(&(*mem)->attachments);
2347 	mutex_init(&(*mem)->lock);
2348 
2349 	(*mem)->alloc_flags =
2350 		((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2351 		KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
2352 		| KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
2353 		| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
2354 
2355 	get_dma_buf(dma_buf);
2356 	(*mem)->dmabuf = dma_buf;
2357 	(*mem)->bo = bo;
2358 	(*mem)->va = va;
2359 	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) && !adev->gmc.is_app_apu ?
2360 		AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
2361 
2362 	(*mem)->mapped_to_gpu_memory = 0;
2363 	(*mem)->process_info = avm->process_info;
2364 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
2365 	amdgpu_sync_create(&(*mem)->sync);
2366 	(*mem)->is_imported = true;
2367 
2368 	mutex_lock(&avm->process_info->lock);
2369 	if (avm->process_info->eviction_fence &&
2370 	    !dma_fence_is_signaled(&avm->process_info->eviction_fence->base))
2371 		ret = amdgpu_amdkfd_bo_validate_and_fence(bo, (*mem)->domain,
2372 				&avm->process_info->eviction_fence->base);
2373 	mutex_unlock(&avm->process_info->lock);
2374 	if (ret)
2375 		goto err_remove_mem;
2376 
2377 	return 0;
2378 
2379 err_remove_mem:
2380 	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
2381 	drm_vma_node_revoke(&obj->vma_node, drm_priv);
2382 err_free_mem:
2383 	kfree(*mem);
2384 	return ret;
2385 }
2386 
2387 int amdgpu_amdkfd_gpuvm_import_dmabuf_fd(struct amdgpu_device *adev, int fd,
2388 					 uint64_t va, void *drm_priv,
2389 					 struct kgd_mem **mem, uint64_t *size,
2390 					 uint64_t *mmap_offset)
2391 {
2392 	struct drm_gem_object *obj;
2393 	uint32_t handle;
2394 	int ret;
2395 
2396 	ret = drm_gem_prime_fd_to_handle(&adev->ddev, adev->kfd.client.file, fd,
2397 					 &handle);
2398 	if (ret)
2399 		return ret;
2400 	obj = drm_gem_object_lookup(adev->kfd.client.file, handle);
2401 	if (!obj) {
2402 		ret = -EINVAL;
2403 		goto err_release_handle;
2404 	}
2405 
2406 	ret = import_obj_create(adev, obj->dma_buf, obj, va, drm_priv, mem, size,
2407 				mmap_offset);
2408 	if (ret)
2409 		goto err_put_obj;
2410 
2411 	(*mem)->gem_handle = handle;
2412 
2413 	return 0;
2414 
2415 err_put_obj:
2416 	drm_gem_object_put(obj);
2417 err_release_handle:
2418 	drm_gem_handle_delete(adev->kfd.client.file, handle);
2419 	return ret;
2420 }
2421 
2422 int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem,
2423 				      struct dma_buf **dma_buf)
2424 {
2425 	int ret;
2426 
2427 	mutex_lock(&mem->lock);
2428 	ret = kfd_mem_export_dmabuf(mem);
2429 	if (ret)
2430 		goto out;
2431 
2432 	get_dma_buf(mem->dmabuf);
2433 	*dma_buf = mem->dmabuf;
2434 out:
2435 	mutex_unlock(&mem->lock);
2436 	return ret;
2437 }
2438 
2439 /* Evict a userptr BO by stopping the queues if necessary
2440  *
2441  * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
2442  * cannot do any memory allocations, and cannot take any locks that
2443  * are held elsewhere while allocating memory.
2444  *
2445  * It doesn't do anything to the BO itself. The real work happens in
2446  * restore, where we get updated page addresses. This function only
2447  * ensures that GPU access to the BO is stopped.
2448  */
2449 int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
2450 				unsigned long cur_seq, struct kgd_mem *mem)
2451 {
2452 	struct amdkfd_process_info *process_info = mem->process_info;
2453 	int r = 0;
2454 
2455 	/* Do not process MMU notifications during CRIU restore until
2456 	 * KFD_CRIU_OP_RESUME IOCTL is received
2457 	 */
2458 	if (READ_ONCE(process_info->block_mmu_notifications))
2459 		return 0;
2460 
2461 	mutex_lock(&process_info->notifier_lock);
2462 	mmu_interval_set_seq(mni, cur_seq);
2463 
2464 	mem->invalid++;
2465 	if (++process_info->evicted_bos == 1) {
2466 		/* First eviction, stop the queues */
2467 		r = kgd2kfd_quiesce_mm(mni->mm,
2468 				       KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
2469 		if (r)
2470 			pr_err("Failed to quiesce KFD\n");
2471 		queue_delayed_work(system_freezable_wq,
2472 			&process_info->restore_userptr_work,
2473 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2474 	}
2475 	mutex_unlock(&process_info->notifier_lock);
2476 
2477 	return r;
2478 }
2479 
2480 /* Update invalid userptr BOs
2481  *
2482  * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
2483  * userptr_inval_list and updates user pages for all BOs that have
2484  * been invalidated since their last update.
2485  */
2486 static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
2487 				     struct mm_struct *mm)
2488 {
2489 	struct kgd_mem *mem, *tmp_mem;
2490 	struct amdgpu_bo *bo;
2491 	struct ttm_operation_ctx ctx = { false, false };
2492 	uint32_t invalid;
2493 	int ret = 0;
2494 
2495 	mutex_lock(&process_info->notifier_lock);
2496 
2497 	/* Move all invalidated BOs to the userptr_inval_list */
2498 	list_for_each_entry_safe(mem, tmp_mem,
2499 				 &process_info->userptr_valid_list,
2500 				 validate_list)
2501 		if (mem->invalid)
2502 			list_move_tail(&mem->validate_list,
2503 				       &process_info->userptr_inval_list);
2504 
2505 	/* Go through userptr_inval_list and update any invalid user_pages */
2506 	list_for_each_entry(mem, &process_info->userptr_inval_list,
2507 			    validate_list) {
2508 		invalid = mem->invalid;
2509 		if (!invalid)
2510 			/* BO hasn't been invalidated since the last
2511 			 * revalidation attempt. Keep its page list.
2512 			 */
2513 			continue;
2514 
2515 		bo = mem->bo;
2516 
2517 		amdgpu_ttm_tt_discard_user_pages(bo->tbo.ttm, mem->range);
2518 		mem->range = NULL;
2519 
2520 		/* BO reservations and getting user pages (hmm_range_fault)
2521 		 * must happen outside the notifier lock
2522 		 */
2523 		mutex_unlock(&process_info->notifier_lock);
2524 
2525 		/* Move the BO to system (CPU) domain if necessary to unmap
2526 		 * and free the SG table
2527 		 */
2528 		if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) {
2529 			if (amdgpu_bo_reserve(bo, true))
2530 				return -EAGAIN;
2531 			amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
2532 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2533 			amdgpu_bo_unreserve(bo);
2534 			if (ret) {
2535 				pr_err("%s: Failed to invalidate userptr BO\n",
2536 				       __func__);
2537 				return -EAGAIN;
2538 			}
2539 		}
2540 
2541 		/* Get updated user pages */
2542 		ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages,
2543 						   &mem->range);
2544 		if (ret) {
2545 			pr_debug("Failed %d to get user pages\n", ret);
2546 
2547 			/* Return -EFAULT bad address error as success. It will
2548 			 * fail later with a VM fault if the GPU tries to access
2549 			 * it. Better than hanging indefinitely with stalled
2550 			 * user mode queues.
2551 			 *
2552 			 * Return other error -EBUSY or -ENOMEM to retry restore
2553 			 */
2554 			if (ret != -EFAULT)
2555 				return ret;
2556 
2557 			ret = 0;
2558 		}
2559 
2560 		mutex_lock(&process_info->notifier_lock);
2561 
2562 		/* Mark the BO as valid unless it was invalidated
2563 		 * again concurrently.
2564 		 */
2565 		if (mem->invalid != invalid) {
2566 			ret = -EAGAIN;
2567 			goto unlock_out;
2568 		}
2569 		 /* set mem valid if mem has hmm range associated */
2570 		if (mem->range)
2571 			mem->invalid = 0;
2572 	}
2573 
2574 unlock_out:
2575 	mutex_unlock(&process_info->notifier_lock);
2576 
2577 	return ret;
2578 }
2579 
2580 /* Validate invalid userptr BOs
2581  *
2582  * Validates BOs on the userptr_inval_list. Also updates GPUVM page tables
2583  * with new page addresses and waits for the page table updates to complete.
2584  */
2585 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
2586 {
2587 	struct ttm_operation_ctx ctx = { false, false };
2588 	struct amdgpu_sync sync;
2589 	struct drm_exec exec;
2590 
2591 	struct amdgpu_vm *peer_vm;
2592 	struct kgd_mem *mem, *tmp_mem;
2593 	struct amdgpu_bo *bo;
2594 	int ret;
2595 
2596 	amdgpu_sync_create(&sync);
2597 
2598 	drm_exec_init(&exec, 0, 0);
2599 	/* Reserve all BOs and page tables for validation */
2600 	drm_exec_until_all_locked(&exec) {
2601 		/* Reserve all the page directories */
2602 		list_for_each_entry(peer_vm, &process_info->vm_list_head,
2603 				    vm_list_node) {
2604 			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2605 			drm_exec_retry_on_contention(&exec);
2606 			if (unlikely(ret))
2607 				goto unreserve_out;
2608 		}
2609 
2610 		/* Reserve the userptr_inval_list entries to resv_list */
2611 		list_for_each_entry(mem, &process_info->userptr_inval_list,
2612 				    validate_list) {
2613 			struct drm_gem_object *gobj;
2614 
2615 			gobj = &mem->bo->tbo.base;
2616 			ret = drm_exec_prepare_obj(&exec, gobj, 1);
2617 			drm_exec_retry_on_contention(&exec);
2618 			if (unlikely(ret))
2619 				goto unreserve_out;
2620 		}
2621 	}
2622 
2623 	ret = process_validate_vms(process_info);
2624 	if (ret)
2625 		goto unreserve_out;
2626 
2627 	/* Validate BOs and update GPUVM page tables */
2628 	list_for_each_entry_safe(mem, tmp_mem,
2629 				 &process_info->userptr_inval_list,
2630 				 validate_list) {
2631 		struct kfd_mem_attachment *attachment;
2632 
2633 		bo = mem->bo;
2634 
2635 		/* Validate the BO if we got user pages */
2636 		if (bo->tbo.ttm->pages[0]) {
2637 			amdgpu_bo_placement_from_domain(bo, mem->domain);
2638 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2639 			if (ret) {
2640 				pr_err("%s: failed to validate BO\n", __func__);
2641 				goto unreserve_out;
2642 			}
2643 		}
2644 
2645 		/* Update mapping. If the BO was not validated
2646 		 * (because we couldn't get user pages), this will
2647 		 * clear the page table entries, which will result in
2648 		 * VM faults if the GPU tries to access the invalid
2649 		 * memory.
2650 		 */
2651 		list_for_each_entry(attachment, &mem->attachments, list) {
2652 			if (!attachment->is_mapped)
2653 				continue;
2654 
2655 			kfd_mem_dmaunmap_attachment(mem, attachment);
2656 			ret = update_gpuvm_pte(mem, attachment, &sync);
2657 			if (ret) {
2658 				pr_err("%s: update PTE failed\n", __func__);
2659 				/* make sure this gets validated again */
2660 				mutex_lock(&process_info->notifier_lock);
2661 				mem->invalid++;
2662 				mutex_unlock(&process_info->notifier_lock);
2663 				goto unreserve_out;
2664 			}
2665 		}
2666 	}
2667 
2668 	/* Update page directories */
2669 	ret = process_update_pds(process_info, &sync);
2670 
2671 unreserve_out:
2672 	drm_exec_fini(&exec);
2673 	amdgpu_sync_wait(&sync, false);
2674 	amdgpu_sync_free(&sync);
2675 
2676 	return ret;
2677 }
2678 
2679 /* Confirm that all user pages are valid while holding the notifier lock
2680  *
2681  * Moves valid BOs from the userptr_inval_list back to userptr_val_list.
2682  */
2683 static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_info)
2684 {
2685 	struct kgd_mem *mem, *tmp_mem;
2686 	int ret = 0;
2687 
2688 	list_for_each_entry_safe(mem, tmp_mem,
2689 				 &process_info->userptr_inval_list,
2690 				 validate_list) {
2691 		bool valid;
2692 
2693 		/* keep mem without hmm range at userptr_inval_list */
2694 		if (!mem->range)
2695 			 continue;
2696 
2697 		/* Only check mem with hmm range associated */
2698 		valid = amdgpu_ttm_tt_get_user_pages_done(
2699 					mem->bo->tbo.ttm, mem->range);
2700 
2701 		mem->range = NULL;
2702 		if (!valid) {
2703 			WARN(!mem->invalid, "Invalid BO not marked invalid");
2704 			ret = -EAGAIN;
2705 			continue;
2706 		}
2707 
2708 		if (mem->invalid) {
2709 			WARN(1, "Valid BO is marked invalid");
2710 			ret = -EAGAIN;
2711 			continue;
2712 		}
2713 
2714 		list_move_tail(&mem->validate_list,
2715 			       &process_info->userptr_valid_list);
2716 	}
2717 
2718 	return ret;
2719 }
2720 
2721 /* Worker callback to restore evicted userptr BOs
2722  *
2723  * Tries to update and validate all userptr BOs. If successful and no
2724  * concurrent evictions happened, the queues are restarted. Otherwise,
2725  * reschedule for another attempt later.
2726  */
2727 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
2728 {
2729 	struct delayed_work *dwork = to_delayed_work(work);
2730 	struct amdkfd_process_info *process_info =
2731 		container_of(dwork, struct amdkfd_process_info,
2732 			     restore_userptr_work);
2733 	struct task_struct *usertask;
2734 	struct mm_struct *mm;
2735 	uint32_t evicted_bos;
2736 
2737 	mutex_lock(&process_info->notifier_lock);
2738 	evicted_bos = process_info->evicted_bos;
2739 	mutex_unlock(&process_info->notifier_lock);
2740 	if (!evicted_bos)
2741 		return;
2742 
2743 	/* Reference task and mm in case of concurrent process termination */
2744 	usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
2745 	if (!usertask)
2746 		return;
2747 	mm = get_task_mm(usertask);
2748 	if (!mm) {
2749 		put_task_struct(usertask);
2750 		return;
2751 	}
2752 
2753 	mutex_lock(&process_info->lock);
2754 
2755 	if (update_invalid_user_pages(process_info, mm))
2756 		goto unlock_out;
2757 	/* userptr_inval_list can be empty if all evicted userptr BOs
2758 	 * have been freed. In that case there is nothing to validate
2759 	 * and we can just restart the queues.
2760 	 */
2761 	if (!list_empty(&process_info->userptr_inval_list)) {
2762 		if (validate_invalid_user_pages(process_info))
2763 			goto unlock_out;
2764 	}
2765 	/* Final check for concurrent evicton and atomic update. If
2766 	 * another eviction happens after successful update, it will
2767 	 * be a first eviction that calls quiesce_mm. The eviction
2768 	 * reference counting inside KFD will handle this case.
2769 	 */
2770 	mutex_lock(&process_info->notifier_lock);
2771 	if (process_info->evicted_bos != evicted_bos)
2772 		goto unlock_notifier_out;
2773 
2774 	if (confirm_valid_user_pages_locked(process_info)) {
2775 		WARN(1, "User pages unexpectedly invalid");
2776 		goto unlock_notifier_out;
2777 	}
2778 
2779 	process_info->evicted_bos = evicted_bos = 0;
2780 
2781 	if (kgd2kfd_resume_mm(mm)) {
2782 		pr_err("%s: Failed to resume KFD\n", __func__);
2783 		/* No recovery from this failure. Probably the CP is
2784 		 * hanging. No point trying again.
2785 		 */
2786 	}
2787 
2788 unlock_notifier_out:
2789 	mutex_unlock(&process_info->notifier_lock);
2790 unlock_out:
2791 	mutex_unlock(&process_info->lock);
2792 
2793 	/* If validation failed, reschedule another attempt */
2794 	if (evicted_bos) {
2795 		queue_delayed_work(system_freezable_wq,
2796 			&process_info->restore_userptr_work,
2797 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2798 
2799 		kfd_smi_event_queue_restore_rescheduled(mm);
2800 	}
2801 	mmput(mm);
2802 	put_task_struct(usertask);
2803 }
2804 
2805 static void replace_eviction_fence(struct dma_fence __rcu **ef,
2806 				   struct dma_fence *new_ef)
2807 {
2808 	struct dma_fence *old_ef = rcu_replace_pointer(*ef, new_ef, true
2809 		/* protected by process_info->lock */);
2810 
2811 	/* If we're replacing an unsignaled eviction fence, that fence will
2812 	 * never be signaled, and if anyone is still waiting on that fence,
2813 	 * they will hang forever. This should never happen. We should only
2814 	 * replace the fence in restore_work that only gets scheduled after
2815 	 * eviction work signaled the fence.
2816 	 */
2817 	WARN_ONCE(!dma_fence_is_signaled(old_ef),
2818 		  "Replacing unsignaled eviction fence");
2819 	dma_fence_put(old_ef);
2820 }
2821 
2822 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2823  *   KFD process identified by process_info
2824  *
2825  * @process_info: amdkfd_process_info of the KFD process
2826  *
2827  * After memory eviction, restore thread calls this function. The function
2828  * should be called when the Process is still valid. BO restore involves -
2829  *
2830  * 1.  Release old eviction fence and create new one
2831  * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2832  * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2833  *     BOs that need to be reserved.
2834  * 4.  Reserve all the BOs
2835  * 5.  Validate of PD and PT BOs.
2836  * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2837  * 7.  Add fence to all PD and PT BOs.
2838  * 8.  Unreserve all BOs
2839  */
2840 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu **ef)
2841 {
2842 	struct amdkfd_process_info *process_info = info;
2843 	struct amdgpu_vm *peer_vm;
2844 	struct kgd_mem *mem;
2845 	struct list_head duplicate_save;
2846 	struct amdgpu_sync sync_obj;
2847 	unsigned long failed_size = 0;
2848 	unsigned long total_size = 0;
2849 	struct drm_exec exec;
2850 	int ret;
2851 
2852 	INIT_LIST_HEAD(&duplicate_save);
2853 
2854 	mutex_lock(&process_info->lock);
2855 
2856 	drm_exec_init(&exec, 0, 0);
2857 	drm_exec_until_all_locked(&exec) {
2858 		list_for_each_entry(peer_vm, &process_info->vm_list_head,
2859 				    vm_list_node) {
2860 			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2861 			drm_exec_retry_on_contention(&exec);
2862 			if (unlikely(ret))
2863 				goto ttm_reserve_fail;
2864 		}
2865 
2866 		/* Reserve all BOs and page tables/directory. Add all BOs from
2867 		 * kfd_bo_list to ctx.list
2868 		 */
2869 		list_for_each_entry(mem, &process_info->kfd_bo_list,
2870 				    validate_list) {
2871 			struct drm_gem_object *gobj;
2872 
2873 			gobj = &mem->bo->tbo.base;
2874 			ret = drm_exec_prepare_obj(&exec, gobj, 1);
2875 			drm_exec_retry_on_contention(&exec);
2876 			if (unlikely(ret))
2877 				goto ttm_reserve_fail;
2878 		}
2879 	}
2880 
2881 	amdgpu_sync_create(&sync_obj);
2882 
2883 	/* Validate PDs and PTs */
2884 	ret = process_validate_vms(process_info);
2885 	if (ret)
2886 		goto validate_map_fail;
2887 
2888 	/* Validate BOs and map them to GPUVM (update VM page tables). */
2889 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2890 			    validate_list) {
2891 
2892 		struct amdgpu_bo *bo = mem->bo;
2893 		uint32_t domain = mem->domain;
2894 		struct kfd_mem_attachment *attachment;
2895 		struct dma_resv_iter cursor;
2896 		struct dma_fence *fence;
2897 
2898 		total_size += amdgpu_bo_size(bo);
2899 
2900 		ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2901 		if (ret) {
2902 			pr_debug("Memory eviction: Validate BOs failed\n");
2903 			failed_size += amdgpu_bo_size(bo);
2904 			ret = amdgpu_amdkfd_bo_validate(bo,
2905 						AMDGPU_GEM_DOMAIN_GTT, false);
2906 			if (ret) {
2907 				pr_debug("Memory eviction: Try again\n");
2908 				goto validate_map_fail;
2909 			}
2910 		}
2911 		dma_resv_for_each_fence(&cursor, bo->tbo.base.resv,
2912 					DMA_RESV_USAGE_KERNEL, fence) {
2913 			ret = amdgpu_sync_fence(&sync_obj, fence);
2914 			if (ret) {
2915 				pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2916 				goto validate_map_fail;
2917 			}
2918 		}
2919 		list_for_each_entry(attachment, &mem->attachments, list) {
2920 			if (!attachment->is_mapped)
2921 				continue;
2922 
2923 			if (attachment->bo_va->base.bo->tbo.pin_count)
2924 				continue;
2925 
2926 			kfd_mem_dmaunmap_attachment(mem, attachment);
2927 			ret = update_gpuvm_pte(mem, attachment, &sync_obj);
2928 			if (ret) {
2929 				pr_debug("Memory eviction: update PTE failed. Try again\n");
2930 				goto validate_map_fail;
2931 			}
2932 		}
2933 	}
2934 
2935 	if (failed_size)
2936 		pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2937 
2938 	/* Update mappings not managed by KFD */
2939 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2940 			vm_list_node) {
2941 		struct amdgpu_device *adev = amdgpu_ttm_adev(
2942 			peer_vm->root.bo->tbo.bdev);
2943 
2944 		ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket);
2945 		if (ret) {
2946 			pr_debug("Memory eviction: handle moved failed. Try again\n");
2947 			goto validate_map_fail;
2948 		}
2949 	}
2950 
2951 	/* Update page directories */
2952 	ret = process_update_pds(process_info, &sync_obj);
2953 	if (ret) {
2954 		pr_debug("Memory eviction: update PDs failed. Try again\n");
2955 		goto validate_map_fail;
2956 	}
2957 
2958 	/* Sync with fences on all the page tables. They implicitly depend on any
2959 	 * move fences from amdgpu_vm_handle_moved above.
2960 	 */
2961 	ret = process_sync_pds_resv(process_info, &sync_obj);
2962 	if (ret) {
2963 		pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2964 		goto validate_map_fail;
2965 	}
2966 
2967 	/* Wait for validate and PT updates to finish */
2968 	amdgpu_sync_wait(&sync_obj, false);
2969 
2970 	/* The old eviction fence may be unsignaled if restore happens
2971 	 * after a GPU reset or suspend/resume. Keep the old fence in that
2972 	 * case. Otherwise release the old eviction fence and create new
2973 	 * one, because fence only goes from unsignaled to signaled once
2974 	 * and cannot be reused. Use context and mm from the old fence.
2975 	 *
2976 	 * If an old eviction fence signals after this check, that's OK.
2977 	 * Anyone signaling an eviction fence must stop the queues first
2978 	 * and schedule another restore worker.
2979 	 */
2980 	if (dma_fence_is_signaled(&process_info->eviction_fence->base)) {
2981 		struct amdgpu_amdkfd_fence *new_fence =
2982 			amdgpu_amdkfd_fence_create(
2983 				process_info->eviction_fence->base.context,
2984 				process_info->eviction_fence->mm,
2985 				NULL);
2986 
2987 		if (!new_fence) {
2988 			pr_err("Failed to create eviction fence\n");
2989 			ret = -ENOMEM;
2990 			goto validate_map_fail;
2991 		}
2992 		dma_fence_put(&process_info->eviction_fence->base);
2993 		process_info->eviction_fence = new_fence;
2994 		replace_eviction_fence(ef, dma_fence_get(&new_fence->base));
2995 	} else {
2996 		WARN_ONCE(*ef != &process_info->eviction_fence->base,
2997 			  "KFD eviction fence doesn't match KGD process_info");
2998 	}
2999 
3000 	/* Attach new eviction fence to all BOs except pinned ones */
3001 	list_for_each_entry(mem, &process_info->kfd_bo_list, validate_list) {
3002 		if (mem->bo->tbo.pin_count)
3003 			continue;
3004 
3005 		dma_resv_add_fence(mem->bo->tbo.base.resv,
3006 				   &process_info->eviction_fence->base,
3007 				   DMA_RESV_USAGE_BOOKKEEP);
3008 	}
3009 	/* Attach eviction fence to PD / PT BOs */
3010 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
3011 			    vm_list_node) {
3012 		struct amdgpu_bo *bo = peer_vm->root.bo;
3013 
3014 		dma_resv_add_fence(bo->tbo.base.resv,
3015 				   &process_info->eviction_fence->base,
3016 				   DMA_RESV_USAGE_BOOKKEEP);
3017 	}
3018 
3019 validate_map_fail:
3020 	amdgpu_sync_free(&sync_obj);
3021 ttm_reserve_fail:
3022 	drm_exec_fini(&exec);
3023 	mutex_unlock(&process_info->lock);
3024 	return ret;
3025 }
3026 
3027 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
3028 {
3029 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
3030 	struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
3031 	int ret;
3032 
3033 	if (!info || !gws)
3034 		return -EINVAL;
3035 
3036 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
3037 	if (!*mem)
3038 		return -ENOMEM;
3039 
3040 	mutex_init(&(*mem)->lock);
3041 	INIT_LIST_HEAD(&(*mem)->attachments);
3042 	(*mem)->bo = amdgpu_bo_ref(gws_bo);
3043 	(*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
3044 	(*mem)->process_info = process_info;
3045 	add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
3046 	amdgpu_sync_create(&(*mem)->sync);
3047 
3048 
3049 	/* Validate gws bo the first time it is added to process */
3050 	mutex_lock(&(*mem)->process_info->lock);
3051 	ret = amdgpu_bo_reserve(gws_bo, false);
3052 	if (unlikely(ret)) {
3053 		pr_err("Reserve gws bo failed %d\n", ret);
3054 		goto bo_reservation_failure;
3055 	}
3056 
3057 	ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
3058 	if (ret) {
3059 		pr_err("GWS BO validate failed %d\n", ret);
3060 		goto bo_validation_failure;
3061 	}
3062 	/* GWS resource is shared b/t amdgpu and amdkfd
3063 	 * Add process eviction fence to bo so they can
3064 	 * evict each other.
3065 	 */
3066 	ret = dma_resv_reserve_fences(gws_bo->tbo.base.resv, 1);
3067 	if (ret)
3068 		goto reserve_shared_fail;
3069 	dma_resv_add_fence(gws_bo->tbo.base.resv,
3070 			   &process_info->eviction_fence->base,
3071 			   DMA_RESV_USAGE_BOOKKEEP);
3072 	amdgpu_bo_unreserve(gws_bo);
3073 	mutex_unlock(&(*mem)->process_info->lock);
3074 
3075 	return ret;
3076 
3077 reserve_shared_fail:
3078 bo_validation_failure:
3079 	amdgpu_bo_unreserve(gws_bo);
3080 bo_reservation_failure:
3081 	mutex_unlock(&(*mem)->process_info->lock);
3082 	amdgpu_sync_free(&(*mem)->sync);
3083 	remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
3084 	amdgpu_bo_unref(&gws_bo);
3085 	mutex_destroy(&(*mem)->lock);
3086 	kfree(*mem);
3087 	*mem = NULL;
3088 	return ret;
3089 }
3090 
3091 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
3092 {
3093 	int ret;
3094 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
3095 	struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
3096 	struct amdgpu_bo *gws_bo = kgd_mem->bo;
3097 
3098 	/* Remove BO from process's validate list so restore worker won't touch
3099 	 * it anymore
3100 	 */
3101 	remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
3102 
3103 	ret = amdgpu_bo_reserve(gws_bo, false);
3104 	if (unlikely(ret)) {
3105 		pr_err("Reserve gws bo failed %d\n", ret);
3106 		//TODO add BO back to validate_list?
3107 		return ret;
3108 	}
3109 	amdgpu_amdkfd_remove_eviction_fence(gws_bo,
3110 			process_info->eviction_fence);
3111 	amdgpu_bo_unreserve(gws_bo);
3112 	amdgpu_sync_free(&kgd_mem->sync);
3113 	amdgpu_bo_unref(&gws_bo);
3114 	mutex_destroy(&kgd_mem->lock);
3115 	kfree(mem);
3116 	return 0;
3117 }
3118 
3119 /* Returns GPU-specific tiling mode information */
3120 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
3121 				struct tile_config *config)
3122 {
3123 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
3124 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
3125 	config->num_tile_configs =
3126 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
3127 	config->macro_tile_config_ptr =
3128 			adev->gfx.config.macrotile_mode_array;
3129 	config->num_macro_tile_configs =
3130 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
3131 
3132 	/* Those values are not set from GFX9 onwards */
3133 	config->num_banks = adev->gfx.config.num_banks;
3134 	config->num_ranks = adev->gfx.config.num_ranks;
3135 
3136 	return 0;
3137 }
3138 
3139 bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *mem)
3140 {
3141 	struct kfd_mem_attachment *entry;
3142 
3143 	list_for_each_entry(entry, &mem->attachments, list) {
3144 		if (entry->is_mapped && entry->adev == adev)
3145 			return true;
3146 	}
3147 	return false;
3148 }
3149 
3150 #if defined(CONFIG_DEBUG_FS)
3151 
3152 int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data)
3153 {
3154 
3155 	spin_lock(&kfd_mem_limit.mem_limit_lock);
3156 	seq_printf(m, "System mem used %lldM out of %lluM\n",
3157 		  (kfd_mem_limit.system_mem_used >> 20),
3158 		  (kfd_mem_limit.max_system_mem_limit >> 20));
3159 	seq_printf(m, "TTM mem used %lldM out of %lluM\n",
3160 		  (kfd_mem_limit.ttm_mem_used >> 20),
3161 		  (kfd_mem_limit.max_ttm_mem_limit >> 20));
3162 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
3163 
3164 	return 0;
3165 }
3166 
3167 #endif
3168