xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c (revision e619ac419174fdb6093b9e78b41bb5d0a97de9dd)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright 2014-2018 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 #include <linux/dma-buf.h>
24 #include <linux/list.h>
25 #include <linux/pagemap.h>
26 #include <linux/sched/mm.h>
27 #include <linux/sched/task.h>
28 #include <drm/ttm/ttm_tt.h>
29 
30 #include <drm/drm_exec.h>
31 
32 #include "amdgpu_object.h"
33 #include "amdgpu_gem.h"
34 #include "amdgpu_vm.h"
35 #include "amdgpu_hmm.h"
36 #include "amdgpu_amdkfd.h"
37 #include "amdgpu_dma_buf.h"
38 #include <uapi/linux/kfd_ioctl.h>
39 #include "amdgpu_xgmi.h"
40 #include "kfd_priv.h"
41 #include "kfd_smi_events.h"
42 
43 /* Userptr restore delay, just long enough to allow consecutive VM
44  * changes to accumulate
45  */
46 #define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
47 #define AMDGPU_RESERVE_MEM_LIMIT			(3UL << 29)
48 
49 /*
50  * Align VRAM availability to 2MB to avoid fragmentation caused by 4K allocations in the tail 2MB
51  * BO chunk
52  */
53 #define VRAM_AVAILABLITY_ALIGN (1 << 21)
54 
55 /* Impose limit on how much memory KFD can use */
56 static struct {
57 	uint64_t max_system_mem_limit;
58 	uint64_t max_ttm_mem_limit;
59 	int64_t system_mem_used;
60 	int64_t ttm_mem_used;
61 	spinlock_t mem_limit_lock;
62 } kfd_mem_limit;
63 
64 static const char * const domain_bit_to_string[] = {
65 		"CPU",
66 		"GTT",
67 		"VRAM",
68 		"GDS",
69 		"GWS",
70 		"OA"
71 };
72 
73 #define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
74 
75 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work);
76 
77 static bool kfd_mem_is_attached(struct amdgpu_vm *avm,
78 		struct kgd_mem *mem)
79 {
80 	struct kfd_mem_attachment *entry;
81 
82 	list_for_each_entry(entry, &mem->attachments, list)
83 		if (entry->bo_va->base.vm == avm)
84 			return true;
85 
86 	return false;
87 }
88 
89 /**
90  * reuse_dmamap() - Check whether adev can share the original
91  * userptr BO
92  *
93  * If both adev and bo_adev are in direct mapping or
94  * in the same iommu group, they can share the original BO.
95  *
96  * @adev: Device to which can or cannot share the original BO
97  * @bo_adev: Device to which allocated BO belongs to
98  *
99  * Return: returns true if adev can share original userptr BO,
100  * false otherwise.
101  */
102 static bool reuse_dmamap(struct amdgpu_device *adev, struct amdgpu_device *bo_adev)
103 {
104 	return (adev->ram_is_direct_mapped && bo_adev->ram_is_direct_mapped) ||
105 			(adev->dev->iommu_group == bo_adev->dev->iommu_group);
106 }
107 
108 /* Set memory usage limits. Current, limits are
109  *  System (TTM + userptr) memory - 15/16th System RAM
110  *  TTM memory - 3/8th System RAM
111  */
112 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
113 {
114 	struct sysinfo si;
115 	uint64_t mem;
116 
117 	if (kfd_mem_limit.max_system_mem_limit)
118 		return;
119 
120 	si_meminfo(&si);
121 	mem = si.totalram - si.totalhigh;
122 	mem *= si.mem_unit;
123 
124 	spin_lock_init(&kfd_mem_limit.mem_limit_lock);
125 	kfd_mem_limit.max_system_mem_limit = mem - (mem >> 6);
126 	if (kfd_mem_limit.max_system_mem_limit < 2 * AMDGPU_RESERVE_MEM_LIMIT)
127 		kfd_mem_limit.max_system_mem_limit >>= 1;
128 	else
129 		kfd_mem_limit.max_system_mem_limit -= AMDGPU_RESERVE_MEM_LIMIT;
130 
131 	kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() << PAGE_SHIFT;
132 	pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
133 		(kfd_mem_limit.max_system_mem_limit >> 20),
134 		(kfd_mem_limit.max_ttm_mem_limit >> 20));
135 }
136 
137 void amdgpu_amdkfd_reserve_system_mem(uint64_t size)
138 {
139 	kfd_mem_limit.system_mem_used += size;
140 }
141 
142 /* Estimate page table size needed to represent a given memory size
143  *
144  * With 4KB pages, we need one 8 byte PTE for each 4KB of memory
145  * (factor 512, >> 9). With 2MB pages, we need one 8 byte PTE for 2MB
146  * of memory (factor 256K, >> 18). ROCm user mode tries to optimize
147  * for 2MB pages for TLB efficiency. However, small allocations and
148  * fragmented system memory still need some 4KB pages. We choose a
149  * compromise that should work in most cases without reserving too
150  * much memory for page tables unnecessarily (factor 16K, >> 14).
151  */
152 
153 #define ESTIMATE_PT_SIZE(mem_size) max(((mem_size) >> 14), AMDGPU_VM_RESERVED_VRAM)
154 
155 /**
156  * amdgpu_amdkfd_reserve_mem_limit() - Decrease available memory by size
157  * of buffer.
158  *
159  * @adev: Device to which allocated BO belongs to
160  * @size: Size of buffer, in bytes, encapsulated by B0. This should be
161  * equivalent to amdgpu_bo_size(BO)
162  * @alloc_flag: Flag used in allocating a BO as noted above
163  * @xcp_id: xcp_id is used to get xcp from xcp manager, one xcp is
164  * managed as one compute node in driver for app
165  *
166  * Return:
167  *	returns -ENOMEM in case of error, ZERO otherwise
168  */
169 int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
170 		uint64_t size, u32 alloc_flag, int8_t xcp_id)
171 {
172 	uint64_t reserved_for_pt =
173 		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
174 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
175 	uint64_t reserved_for_ras = (con ? con->reserved_pages_in_bytes : 0);
176 	size_t system_mem_needed, ttm_mem_needed, vram_needed;
177 	int ret = 0;
178 	uint64_t vram_size = 0;
179 
180 	system_mem_needed = 0;
181 	ttm_mem_needed = 0;
182 	vram_needed = 0;
183 	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
184 		system_mem_needed = size;
185 		ttm_mem_needed = size;
186 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
187 		/*
188 		 * Conservatively round up the allocation requirement to 2 MB
189 		 * to avoid fragmentation caused by 4K allocations in the tail
190 		 * 2M BO chunk.
191 		 */
192 		vram_needed = size;
193 		/*
194 		 * For GFX 9.4.3, get the VRAM size from XCP structs
195 		 */
196 		if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
197 			return -EINVAL;
198 
199 		vram_size = KFD_XCP_MEMORY_SIZE(adev, xcp_id);
200 		if (adev->apu_prefer_gtt) {
201 			system_mem_needed = size;
202 			ttm_mem_needed = size;
203 		}
204 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
205 		system_mem_needed = size;
206 	} else if (!(alloc_flag &
207 				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
208 				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
209 		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
210 		return -ENOMEM;
211 	}
212 
213 	spin_lock(&kfd_mem_limit.mem_limit_lock);
214 
215 	if (kfd_mem_limit.system_mem_used + system_mem_needed >
216 	    kfd_mem_limit.max_system_mem_limit)
217 		pr_debug("Set no_system_mem_limit=1 if using shared memory\n");
218 
219 	if ((kfd_mem_limit.system_mem_used + system_mem_needed >
220 	     kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) ||
221 	    (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
222 	     kfd_mem_limit.max_ttm_mem_limit) ||
223 	    (adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] + vram_needed >
224 	     vram_size - reserved_for_pt - reserved_for_ras - atomic64_read(&adev->vram_pin_size))) {
225 		ret = -ENOMEM;
226 		goto release;
227 	}
228 
229 	/* Update memory accounting by decreasing available system
230 	 * memory, TTM memory and GPU memory as computed above
231 	 */
232 	WARN_ONCE(vram_needed && !adev,
233 		  "adev reference can't be null when vram is used");
234 	if (adev && xcp_id >= 0) {
235 		adev->kfd.vram_used[xcp_id] += vram_needed;
236 		adev->kfd.vram_used_aligned[xcp_id] +=
237 				adev->apu_prefer_gtt ?
238 				vram_needed :
239 				ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
240 	}
241 	kfd_mem_limit.system_mem_used += system_mem_needed;
242 	kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
243 
244 release:
245 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
246 	return ret;
247 }
248 
249 void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
250 		uint64_t size, u32 alloc_flag, int8_t xcp_id)
251 {
252 	spin_lock(&kfd_mem_limit.mem_limit_lock);
253 
254 	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
255 		kfd_mem_limit.system_mem_used -= size;
256 		kfd_mem_limit.ttm_mem_used -= size;
257 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
258 		WARN_ONCE(!adev,
259 			  "adev reference can't be null when alloc mem flags vram is set");
260 		if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
261 			goto release;
262 
263 		if (adev) {
264 			adev->kfd.vram_used[xcp_id] -= size;
265 			if (adev->apu_prefer_gtt) {
266 				adev->kfd.vram_used_aligned[xcp_id] -= size;
267 				kfd_mem_limit.system_mem_used -= size;
268 				kfd_mem_limit.ttm_mem_used -= size;
269 			} else {
270 				adev->kfd.vram_used_aligned[xcp_id] -=
271 					ALIGN(size, VRAM_AVAILABLITY_ALIGN);
272 			}
273 		}
274 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
275 		kfd_mem_limit.system_mem_used -= size;
276 	} else if (!(alloc_flag &
277 				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
278 				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
279 		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
280 		goto release;
281 	}
282 	WARN_ONCE(adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] < 0,
283 		  "KFD VRAM memory accounting unbalanced for xcp: %d", xcp_id);
284 	WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
285 		  "KFD TTM memory accounting unbalanced");
286 	WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
287 		  "KFD system memory accounting unbalanced");
288 
289 release:
290 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
291 }
292 
293 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
294 {
295 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
296 	u32 alloc_flags = bo->kfd_bo->alloc_flags;
297 	u64 size = amdgpu_bo_size(bo);
298 
299 	amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags,
300 					  bo->xcp_id);
301 
302 	kfree(bo->kfd_bo);
303 }
304 
305 /**
306  * create_dmamap_sg_bo() - Creates a amdgpu_bo object to reflect information
307  * about USERPTR or DOOREBELL or MMIO BO.
308  *
309  * @adev: Device for which dmamap BO is being created
310  * @mem: BO of peer device that is being DMA mapped. Provides parameters
311  *	 in building the dmamap BO
312  * @bo_out: Output parameter updated with handle of dmamap BO
313  */
314 static int
315 create_dmamap_sg_bo(struct amdgpu_device *adev,
316 		 struct kgd_mem *mem, struct amdgpu_bo **bo_out)
317 {
318 	struct drm_gem_object *gem_obj;
319 	int ret;
320 	uint64_t flags = 0;
321 
322 	ret = amdgpu_bo_reserve(mem->bo, false);
323 	if (ret)
324 		return ret;
325 
326 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)
327 		flags |= mem->bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
328 					AMDGPU_GEM_CREATE_UNCACHED);
329 
330 	ret = amdgpu_gem_object_create(adev, mem->bo->tbo.base.size, 1,
331 			AMDGPU_GEM_DOMAIN_CPU, AMDGPU_GEM_CREATE_PREEMPTIBLE | flags,
332 			ttm_bo_type_sg, mem->bo->tbo.base.resv, &gem_obj, 0);
333 
334 	amdgpu_bo_unreserve(mem->bo);
335 
336 	if (ret) {
337 		pr_err("Error in creating DMA mappable SG BO on domain: %d\n", ret);
338 		return -EINVAL;
339 	}
340 
341 	*bo_out = gem_to_amdgpu_bo(gem_obj);
342 	(*bo_out)->parent = amdgpu_bo_ref(mem->bo);
343 	return ret;
344 }
345 
346 /* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
347  *  reservation object.
348  *
349  * @bo: [IN] Remove eviction fence(s) from this BO
350  * @ef: [IN] This eviction fence is removed if it
351  *  is present in the shared list.
352  *
353  * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
354  */
355 static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
356 					struct amdgpu_amdkfd_fence *ef)
357 {
358 	struct dma_fence *replacement;
359 
360 	if (!ef)
361 		return -EINVAL;
362 
363 	/* TODO: Instead of block before we should use the fence of the page
364 	 * table update and TLB flush here directly.
365 	 */
366 	replacement = dma_fence_get_stub();
367 	dma_resv_replace_fences(bo->tbo.base.resv, ef->base.context,
368 				replacement, DMA_RESV_USAGE_BOOKKEEP);
369 	dma_fence_put(replacement);
370 	return 0;
371 }
372 
373 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
374 {
375 	struct amdgpu_bo *root = bo;
376 	struct amdgpu_vm_bo_base *vm_bo;
377 	struct amdgpu_vm *vm;
378 	struct amdkfd_process_info *info;
379 	struct amdgpu_amdkfd_fence *ef;
380 	int ret;
381 
382 	/* we can always get vm_bo from root PD bo.*/
383 	while (root->parent)
384 		root = root->parent;
385 
386 	vm_bo = root->vm_bo;
387 	if (!vm_bo)
388 		return 0;
389 
390 	vm = vm_bo->vm;
391 	if (!vm)
392 		return 0;
393 
394 	info = vm->process_info;
395 	if (!info || !info->eviction_fence)
396 		return 0;
397 
398 	ef = container_of(dma_fence_get(&info->eviction_fence->base),
399 			struct amdgpu_amdkfd_fence, base);
400 
401 	BUG_ON(!dma_resv_trylock(bo->tbo.base.resv));
402 	ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef);
403 	dma_resv_unlock(bo->tbo.base.resv);
404 
405 	dma_fence_put(&ef->base);
406 	return ret;
407 }
408 
409 static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
410 				     bool wait)
411 {
412 	struct ttm_operation_ctx ctx = { false, false };
413 	int ret;
414 
415 	if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
416 		 "Called with userptr BO"))
417 		return -EINVAL;
418 
419 	/* bo has been pinned, not need validate it */
420 	if (bo->tbo.pin_count)
421 		return 0;
422 
423 	amdgpu_bo_placement_from_domain(bo, domain);
424 
425 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
426 	if (ret)
427 		goto validate_fail;
428 	if (wait)
429 		amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
430 
431 validate_fail:
432 	return ret;
433 }
434 
435 int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo,
436 					uint32_t domain,
437 					struct dma_fence *fence)
438 {
439 	int ret = amdgpu_bo_reserve(bo, false);
440 
441 	if (ret)
442 		return ret;
443 
444 	ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
445 	if (ret)
446 		goto unreserve_out;
447 
448 	ret = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
449 	if (ret)
450 		goto unreserve_out;
451 
452 	dma_resv_add_fence(bo->tbo.base.resv, fence,
453 			   DMA_RESV_USAGE_BOOKKEEP);
454 
455 unreserve_out:
456 	amdgpu_bo_unreserve(bo);
457 
458 	return ret;
459 }
460 
461 static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo)
462 {
463 	return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false);
464 }
465 
466 /* vm_validate_pt_pd_bos - Validate page table and directory BOs
467  *
468  * Page directories are not updated here because huge page handling
469  * during page table updates can invalidate page directory entries
470  * again. Page directories are only updated after updating page
471  * tables.
472  */
473 static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm,
474 				 struct ww_acquire_ctx *ticket)
475 {
476 	struct amdgpu_bo *pd = vm->root.bo;
477 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
478 	int ret;
479 
480 	ret = amdgpu_vm_validate(adev, vm, ticket,
481 				 amdgpu_amdkfd_validate_vm_bo, NULL);
482 	if (ret) {
483 		pr_err("failed to validate PT BOs\n");
484 		return ret;
485 	}
486 
487 	vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo);
488 
489 	return 0;
490 }
491 
492 static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
493 {
494 	struct amdgpu_bo *pd = vm->root.bo;
495 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
496 	int ret;
497 
498 	ret = amdgpu_vm_update_pdes(adev, vm, false);
499 	if (ret)
500 		return ret;
501 
502 	return amdgpu_sync_fence(sync, vm->last_update);
503 }
504 
505 static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem)
506 {
507 	uint32_t mapping_flags = AMDGPU_VM_PAGE_READABLE |
508 				 AMDGPU_VM_MTYPE_DEFAULT;
509 
510 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE)
511 		mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
512 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE)
513 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
514 
515 	return amdgpu_gem_va_map_flags(adev, mapping_flags);
516 }
517 
518 /**
519  * create_sg_table() - Create an sg_table for a contiguous DMA addr range
520  * @addr: The starting address to point to
521  * @size: Size of memory area in bytes being pointed to
522  *
523  * Allocates an instance of sg_table and initializes it to point to memory
524  * area specified by input parameters. The address used to build is assumed
525  * to be DMA mapped, if needed.
526  *
527  * DOORBELL or MMIO BOs use only one scatterlist node in their sg_table
528  * because they are physically contiguous.
529  *
530  * Return: Initialized instance of SG Table or NULL
531  */
532 static struct sg_table *create_sg_table(uint64_t addr, uint32_t size)
533 {
534 	struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
535 
536 	if (!sg)
537 		return NULL;
538 	if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
539 		kfree(sg);
540 		return NULL;
541 	}
542 	sg_dma_address(sg->sgl) = addr;
543 	sg->sgl->length = size;
544 #ifdef CONFIG_NEED_SG_DMA_LENGTH
545 	sg->sgl->dma_length = size;
546 #endif
547 	return sg;
548 }
549 
550 static int
551 kfd_mem_dmamap_userptr(struct kgd_mem *mem,
552 		       struct kfd_mem_attachment *attachment)
553 {
554 	enum dma_data_direction direction =
555 		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
556 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
557 	struct ttm_operation_ctx ctx = {.interruptible = true};
558 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
559 	struct amdgpu_device *adev = attachment->adev;
560 	struct ttm_tt *src_ttm = mem->bo->tbo.ttm;
561 	struct ttm_tt *ttm = bo->tbo.ttm;
562 	int ret;
563 
564 	if (WARN_ON(ttm->num_pages != src_ttm->num_pages))
565 		return -EINVAL;
566 
567 	ttm->sg = kmalloc(sizeof(*ttm->sg), GFP_KERNEL);
568 	if (unlikely(!ttm->sg))
569 		return -ENOMEM;
570 
571 	/* Same sequence as in amdgpu_ttm_tt_pin_userptr */
572 	ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages,
573 					ttm->num_pages, 0,
574 					(u64)ttm->num_pages << PAGE_SHIFT,
575 					GFP_KERNEL);
576 	if (unlikely(ret))
577 		goto free_sg;
578 
579 	ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
580 	if (unlikely(ret))
581 		goto release_sg;
582 
583 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
584 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
585 	if (ret)
586 		goto unmap_sg;
587 
588 	return 0;
589 
590 unmap_sg:
591 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
592 release_sg:
593 	pr_err("DMA map userptr failed: %d\n", ret);
594 	sg_free_table(ttm->sg);
595 free_sg:
596 	kfree(ttm->sg);
597 	ttm->sg = NULL;
598 	return ret;
599 }
600 
601 static int
602 kfd_mem_dmamap_dmabuf(struct kfd_mem_attachment *attachment)
603 {
604 	struct ttm_operation_ctx ctx = {.interruptible = true};
605 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
606 	int ret;
607 
608 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
609 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
610 	if (ret)
611 		return ret;
612 
613 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
614 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
615 }
616 
617 /**
618  * kfd_mem_dmamap_sg_bo() - Create DMA mapped sg_table to access DOORBELL or MMIO BO
619  * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
620  * @attachment: Virtual address attachment of the BO on accessing device
621  *
622  * An access request from the device that owns DOORBELL does not require DMA mapping.
623  * This is because the request doesn't go through PCIe root complex i.e. it instead
624  * loops back. The need to DMA map arises only when accessing peer device's DOORBELL
625  *
626  * In contrast, all access requests for MMIO need to be DMA mapped without regard to
627  * device ownership. This is because access requests for MMIO go through PCIe root
628  * complex.
629  *
630  * This is accomplished in two steps:
631  *   - Obtain DMA mapped address of DOORBELL or MMIO memory that could be used
632  *         in updating requesting device's page table
633  *   - Signal TTM to mark memory pointed to by requesting device's BO as GPU
634  *         accessible. This allows an update of requesting device's page table
635  *         with entries associated with DOOREBELL or MMIO memory
636  *
637  * This method is invoked in the following contexts:
638  *   - Mapping of DOORBELL or MMIO BO of same or peer device
639  *   - Validating an evicted DOOREBELL or MMIO BO on device seeking access
640  *
641  * Return: ZERO if successful, NON-ZERO otherwise
642  */
643 static int
644 kfd_mem_dmamap_sg_bo(struct kgd_mem *mem,
645 		     struct kfd_mem_attachment *attachment)
646 {
647 	struct ttm_operation_ctx ctx = {.interruptible = true};
648 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
649 	struct amdgpu_device *adev = attachment->adev;
650 	struct ttm_tt *ttm = bo->tbo.ttm;
651 	enum dma_data_direction dir;
652 	dma_addr_t dma_addr;
653 	bool mmio;
654 	int ret;
655 
656 	/* Expect SG Table of dmapmap BO to be NULL */
657 	mmio = (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP);
658 	if (unlikely(ttm->sg)) {
659 		pr_err("SG Table of %d BO for peer device is UNEXPECTEDLY NON-NULL", mmio);
660 		return -EINVAL;
661 	}
662 
663 	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
664 			DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
665 	dma_addr = mem->bo->tbo.sg->sgl->dma_address;
666 	pr_debug("%d BO size: %d\n", mmio, mem->bo->tbo.sg->sgl->length);
667 	pr_debug("%d BO address before DMA mapping: %llx\n", mmio, dma_addr);
668 	dma_addr = dma_map_resource(adev->dev, dma_addr,
669 			mem->bo->tbo.sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
670 	ret = dma_mapping_error(adev->dev, dma_addr);
671 	if (unlikely(ret))
672 		return ret;
673 	pr_debug("%d BO address after DMA mapping: %llx\n", mmio, dma_addr);
674 
675 	ttm->sg = create_sg_table(dma_addr, mem->bo->tbo.sg->sgl->length);
676 	if (unlikely(!ttm->sg)) {
677 		ret = -ENOMEM;
678 		goto unmap_sg;
679 	}
680 
681 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
682 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
683 	if (unlikely(ret))
684 		goto free_sg;
685 
686 	return ret;
687 
688 free_sg:
689 	sg_free_table(ttm->sg);
690 	kfree(ttm->sg);
691 	ttm->sg = NULL;
692 unmap_sg:
693 	dma_unmap_resource(adev->dev, dma_addr, mem->bo->tbo.sg->sgl->length,
694 			   dir, DMA_ATTR_SKIP_CPU_SYNC);
695 	return ret;
696 }
697 
698 static int
699 kfd_mem_dmamap_attachment(struct kgd_mem *mem,
700 			  struct kfd_mem_attachment *attachment)
701 {
702 	switch (attachment->type) {
703 	case KFD_MEM_ATT_SHARED:
704 		return 0;
705 	case KFD_MEM_ATT_USERPTR:
706 		return kfd_mem_dmamap_userptr(mem, attachment);
707 	case KFD_MEM_ATT_DMABUF:
708 		return kfd_mem_dmamap_dmabuf(attachment);
709 	case KFD_MEM_ATT_SG:
710 		return kfd_mem_dmamap_sg_bo(mem, attachment);
711 	default:
712 		WARN_ON_ONCE(1);
713 	}
714 	return -EINVAL;
715 }
716 
717 static void
718 kfd_mem_dmaunmap_userptr(struct kgd_mem *mem,
719 			 struct kfd_mem_attachment *attachment)
720 {
721 	enum dma_data_direction direction =
722 		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
723 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
724 	struct ttm_operation_ctx ctx = {.interruptible = false};
725 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
726 	struct amdgpu_device *adev = attachment->adev;
727 	struct ttm_tt *ttm = bo->tbo.ttm;
728 
729 	if (unlikely(!ttm->sg))
730 		return;
731 
732 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
733 	(void)ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
734 
735 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
736 	sg_free_table(ttm->sg);
737 	kfree(ttm->sg);
738 	ttm->sg = NULL;
739 }
740 
741 static void
742 kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment)
743 {
744 	/* This is a no-op. We don't want to trigger eviction fences when
745 	 * unmapping DMABufs. Therefore the invalidation (moving to system
746 	 * domain) is done in kfd_mem_dmamap_dmabuf.
747 	 */
748 }
749 
750 /**
751  * kfd_mem_dmaunmap_sg_bo() - Free DMA mapped sg_table of DOORBELL or MMIO BO
752  * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
753  * @attachment: Virtual address attachment of the BO on accessing device
754  *
755  * The method performs following steps:
756  *   - Signal TTM to mark memory pointed to by BO as GPU inaccessible
757  *   - Free SG Table that is used to encapsulate DMA mapped memory of
758  *          peer device's DOORBELL or MMIO memory
759  *
760  * This method is invoked in the following contexts:
761  *     UNMapping of DOORBELL or MMIO BO on a device having access to its memory
762  *     Eviction of DOOREBELL or MMIO BO on device having access to its memory
763  *
764  * Return: void
765  */
766 static void
767 kfd_mem_dmaunmap_sg_bo(struct kgd_mem *mem,
768 		       struct kfd_mem_attachment *attachment)
769 {
770 	struct ttm_operation_ctx ctx = {.interruptible = true};
771 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
772 	struct amdgpu_device *adev = attachment->adev;
773 	struct ttm_tt *ttm = bo->tbo.ttm;
774 	enum dma_data_direction dir;
775 
776 	if (unlikely(!ttm->sg)) {
777 		pr_debug("SG Table of BO is NULL");
778 		return;
779 	}
780 
781 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
782 	(void)ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
783 
784 	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
785 				DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
786 	dma_unmap_resource(adev->dev, ttm->sg->sgl->dma_address,
787 			ttm->sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
788 	sg_free_table(ttm->sg);
789 	kfree(ttm->sg);
790 	ttm->sg = NULL;
791 	bo->tbo.sg = NULL;
792 }
793 
794 static void
795 kfd_mem_dmaunmap_attachment(struct kgd_mem *mem,
796 			    struct kfd_mem_attachment *attachment)
797 {
798 	switch (attachment->type) {
799 	case KFD_MEM_ATT_SHARED:
800 		break;
801 	case KFD_MEM_ATT_USERPTR:
802 		kfd_mem_dmaunmap_userptr(mem, attachment);
803 		break;
804 	case KFD_MEM_ATT_DMABUF:
805 		kfd_mem_dmaunmap_dmabuf(attachment);
806 		break;
807 	case KFD_MEM_ATT_SG:
808 		kfd_mem_dmaunmap_sg_bo(mem, attachment);
809 		break;
810 	default:
811 		WARN_ON_ONCE(1);
812 	}
813 }
814 
815 static int kfd_mem_export_dmabuf(struct kgd_mem *mem)
816 {
817 	if (!mem->dmabuf) {
818 		struct amdgpu_device *bo_adev;
819 		struct dma_buf *dmabuf;
820 
821 		bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
822 		dmabuf = drm_gem_prime_handle_to_dmabuf(&bo_adev->ddev, bo_adev->kfd.client.file,
823 					       mem->gem_handle,
824 			mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
825 					       DRM_RDWR : 0);
826 		if (IS_ERR(dmabuf))
827 			return PTR_ERR(dmabuf);
828 		mem->dmabuf = dmabuf;
829 	}
830 
831 	return 0;
832 }
833 
834 static int
835 kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem,
836 		      struct amdgpu_bo **bo)
837 {
838 	struct drm_gem_object *gobj;
839 	int ret;
840 
841 	ret = kfd_mem_export_dmabuf(mem);
842 	if (ret)
843 		return ret;
844 
845 	gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf);
846 	if (IS_ERR(gobj))
847 		return PTR_ERR(gobj);
848 
849 	*bo = gem_to_amdgpu_bo(gobj);
850 	(*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE;
851 
852 	return 0;
853 }
854 
855 /* kfd_mem_attach - Add a BO to a VM
856  *
857  * Everything that needs to bo done only once when a BO is first added
858  * to a VM. It can later be mapped and unmapped many times without
859  * repeating these steps.
860  *
861  * 0. Create BO for DMA mapping, if needed
862  * 1. Allocate and initialize BO VA entry data structure
863  * 2. Add BO to the VM
864  * 3. Determine ASIC-specific PTE flags
865  * 4. Alloc page tables and directories if needed
866  * 4a.  Validate new page tables and directories
867  */
868 static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
869 		struct amdgpu_vm *vm, bool is_aql)
870 {
871 	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
872 	unsigned long bo_size = mem->bo->tbo.base.size;
873 	uint64_t va = mem->va;
874 	struct kfd_mem_attachment *attachment[2] = {NULL, NULL};
875 	struct amdgpu_bo *bo[2] = {NULL, NULL};
876 	struct amdgpu_bo_va *bo_va;
877 	bool same_hive = false;
878 	int i, ret;
879 
880 	if (!va) {
881 		pr_err("Invalid VA when adding BO to VM\n");
882 		return -EINVAL;
883 	}
884 
885 	/* Determine access to VRAM, MMIO and DOORBELL BOs of peer devices
886 	 *
887 	 * The access path of MMIO and DOORBELL BOs of is always over PCIe.
888 	 * In contrast the access path of VRAM BOs depens upon the type of
889 	 * link that connects the peer device. Access over PCIe is allowed
890 	 * if peer device has large BAR. In contrast, access over xGMI is
891 	 * allowed for both small and large BAR configurations of peer device
892 	 */
893 	if ((adev != bo_adev && !adev->apu_prefer_gtt) &&
894 	    ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) ||
895 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) ||
896 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
897 		if (mem->domain == AMDGPU_GEM_DOMAIN_VRAM)
898 			same_hive = amdgpu_xgmi_same_hive(adev, bo_adev);
899 		if (!same_hive && !amdgpu_device_is_peer_accessible(bo_adev, adev))
900 			return -EINVAL;
901 	}
902 
903 	for (i = 0; i <= is_aql; i++) {
904 		attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL);
905 		if (unlikely(!attachment[i])) {
906 			ret = -ENOMEM;
907 			goto unwind;
908 		}
909 
910 		pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
911 			 va + bo_size, vm);
912 
913 		if ((adev == bo_adev && !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) ||
914 		    (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && reuse_dmamap(adev, bo_adev)) ||
915 		    (mem->domain == AMDGPU_GEM_DOMAIN_GTT && reuse_dmamap(adev, bo_adev)) ||
916 		    same_hive) {
917 			/* Mappings on the local GPU, or VRAM mappings in the
918 			 * local hive, or userptr, or GTT mapping can reuse dma map
919 			 * address space share the original BO
920 			 */
921 			attachment[i]->type = KFD_MEM_ATT_SHARED;
922 			bo[i] = mem->bo;
923 			drm_gem_object_get(&bo[i]->tbo.base);
924 		} else if (i > 0) {
925 			/* Multiple mappings on the same GPU share the BO */
926 			attachment[i]->type = KFD_MEM_ATT_SHARED;
927 			bo[i] = bo[0];
928 			drm_gem_object_get(&bo[i]->tbo.base);
929 		} else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
930 			/* Create an SG BO to DMA-map userptrs on other GPUs */
931 			attachment[i]->type = KFD_MEM_ATT_USERPTR;
932 			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
933 			if (ret)
934 				goto unwind;
935 		/* Handle DOORBELL BOs of peer devices and MMIO BOs of local and peer devices */
936 		} else if (mem->bo->tbo.type == ttm_bo_type_sg) {
937 			WARN_ONCE(!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL ||
938 				    mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP),
939 				  "Handing invalid SG BO in ATTACH request");
940 			attachment[i]->type = KFD_MEM_ATT_SG;
941 			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
942 			if (ret)
943 				goto unwind;
944 		/* Enable acces to GTT and VRAM BOs of peer devices */
945 		} else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT ||
946 			   mem->domain == AMDGPU_GEM_DOMAIN_VRAM) {
947 			attachment[i]->type = KFD_MEM_ATT_DMABUF;
948 			ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]);
949 			if (ret)
950 				goto unwind;
951 			pr_debug("Employ DMABUF mechanism to enable peer GPU access\n");
952 		} else {
953 			WARN_ONCE(true, "Handling invalid ATTACH request");
954 			ret = -EINVAL;
955 			goto unwind;
956 		}
957 
958 		/* Add BO to VM internal data structures */
959 		ret = amdgpu_bo_reserve(bo[i], false);
960 		if (ret) {
961 			pr_debug("Unable to reserve BO during memory attach");
962 			goto unwind;
963 		}
964 		bo_va = amdgpu_vm_bo_find(vm, bo[i]);
965 		if (!bo_va)
966 			bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]);
967 		else
968 			++bo_va->ref_count;
969 		attachment[i]->bo_va = bo_va;
970 		amdgpu_bo_unreserve(bo[i]);
971 		if (unlikely(!attachment[i]->bo_va)) {
972 			ret = -ENOMEM;
973 			pr_err("Failed to add BO object to VM. ret == %d\n",
974 			       ret);
975 			goto unwind;
976 		}
977 		attachment[i]->va = va;
978 		attachment[i]->pte_flags = get_pte_flags(adev, mem);
979 		attachment[i]->adev = adev;
980 		list_add(&attachment[i]->list, &mem->attachments);
981 
982 		va += bo_size;
983 	}
984 
985 	return 0;
986 
987 unwind:
988 	for (; i >= 0; i--) {
989 		if (!attachment[i])
990 			continue;
991 		if (attachment[i]->bo_va) {
992 			(void)amdgpu_bo_reserve(bo[i], true);
993 			if (--attachment[i]->bo_va->ref_count == 0)
994 				amdgpu_vm_bo_del(adev, attachment[i]->bo_va);
995 			amdgpu_bo_unreserve(bo[i]);
996 			list_del(&attachment[i]->list);
997 		}
998 		if (bo[i])
999 			drm_gem_object_put(&bo[i]->tbo.base);
1000 		kfree(attachment[i]);
1001 	}
1002 	return ret;
1003 }
1004 
1005 static void kfd_mem_detach(struct kfd_mem_attachment *attachment)
1006 {
1007 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
1008 
1009 	pr_debug("\t remove VA 0x%llx in entry %p\n",
1010 			attachment->va, attachment);
1011 	if (--attachment->bo_va->ref_count == 0)
1012 		amdgpu_vm_bo_del(attachment->adev, attachment->bo_va);
1013 	drm_gem_object_put(&bo->tbo.base);
1014 	list_del(&attachment->list);
1015 	kfree(attachment);
1016 }
1017 
1018 static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
1019 				struct amdkfd_process_info *process_info,
1020 				bool userptr)
1021 {
1022 	mutex_lock(&process_info->lock);
1023 	if (userptr)
1024 		list_add_tail(&mem->validate_list,
1025 			      &process_info->userptr_valid_list);
1026 	else
1027 		list_add_tail(&mem->validate_list, &process_info->kfd_bo_list);
1028 	mutex_unlock(&process_info->lock);
1029 }
1030 
1031 static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
1032 		struct amdkfd_process_info *process_info)
1033 {
1034 	mutex_lock(&process_info->lock);
1035 	list_del(&mem->validate_list);
1036 	mutex_unlock(&process_info->lock);
1037 }
1038 
1039 /* Initializes user pages. It registers the MMU notifier and validates
1040  * the userptr BO in the GTT domain.
1041  *
1042  * The BO must already be on the userptr_valid_list. Otherwise an
1043  * eviction and restore may happen that leaves the new BO unmapped
1044  * with the user mode queues running.
1045  *
1046  * Takes the process_info->lock to protect against concurrent restore
1047  * workers.
1048  *
1049  * Returns 0 for success, negative errno for errors.
1050  */
1051 static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr,
1052 			   bool criu_resume)
1053 {
1054 	struct amdkfd_process_info *process_info = mem->process_info;
1055 	struct amdgpu_bo *bo = mem->bo;
1056 	struct ttm_operation_ctx ctx = { true, false };
1057 	struct hmm_range *range;
1058 	int ret = 0;
1059 
1060 	mutex_lock(&process_info->lock);
1061 
1062 	ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
1063 	if (ret) {
1064 		pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
1065 		goto out;
1066 	}
1067 
1068 	ret = amdgpu_hmm_register(bo, user_addr);
1069 	if (ret) {
1070 		pr_err("%s: Failed to register MMU notifier: %d\n",
1071 		       __func__, ret);
1072 		goto out;
1073 	}
1074 
1075 	if (criu_resume) {
1076 		/*
1077 		 * During a CRIU restore operation, the userptr buffer objects
1078 		 * will be validated in the restore_userptr_work worker at a
1079 		 * later stage when it is scheduled by another ioctl called by
1080 		 * CRIU master process for the target pid for restore.
1081 		 */
1082 		mutex_lock(&process_info->notifier_lock);
1083 		mem->invalid++;
1084 		mutex_unlock(&process_info->notifier_lock);
1085 		mutex_unlock(&process_info->lock);
1086 		return 0;
1087 	}
1088 
1089 	ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, &range);
1090 	if (ret) {
1091 		if (ret == -EAGAIN)
1092 			pr_debug("Failed to get user pages, try again\n");
1093 		else
1094 			pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
1095 		goto unregister_out;
1096 	}
1097 
1098 	ret = amdgpu_bo_reserve(bo, true);
1099 	if (ret) {
1100 		pr_err("%s: Failed to reserve BO\n", __func__);
1101 		goto release_out;
1102 	}
1103 	amdgpu_bo_placement_from_domain(bo, mem->domain);
1104 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1105 	if (ret)
1106 		pr_err("%s: failed to validate BO\n", __func__);
1107 	amdgpu_bo_unreserve(bo);
1108 
1109 release_out:
1110 	amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range);
1111 unregister_out:
1112 	if (ret)
1113 		amdgpu_hmm_unregister(bo);
1114 out:
1115 	mutex_unlock(&process_info->lock);
1116 	return ret;
1117 }
1118 
1119 /* Reserving a BO and its page table BOs must happen atomically to
1120  * avoid deadlocks. Some operations update multiple VMs at once. Track
1121  * all the reservation info in a context structure. Optionally a sync
1122  * object can track VM updates.
1123  */
1124 struct bo_vm_reservation_context {
1125 	/* DRM execution context for the reservation */
1126 	struct drm_exec exec;
1127 	/* Number of VMs reserved */
1128 	unsigned int n_vms;
1129 	/* Pointer to sync object */
1130 	struct amdgpu_sync *sync;
1131 };
1132 
1133 enum bo_vm_match {
1134 	BO_VM_NOT_MAPPED = 0,	/* Match VMs where a BO is not mapped */
1135 	BO_VM_MAPPED,		/* Match VMs where a BO is mapped     */
1136 	BO_VM_ALL,		/* Match all VMs a BO was added to    */
1137 };
1138 
1139 /**
1140  * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
1141  * @mem: KFD BO structure.
1142  * @vm: the VM to reserve.
1143  * @ctx: the struct that will be used in unreserve_bo_and_vms().
1144  */
1145 static int reserve_bo_and_vm(struct kgd_mem *mem,
1146 			      struct amdgpu_vm *vm,
1147 			      struct bo_vm_reservation_context *ctx)
1148 {
1149 	struct amdgpu_bo *bo = mem->bo;
1150 	int ret;
1151 
1152 	WARN_ON(!vm);
1153 
1154 	ctx->n_vms = 1;
1155 	ctx->sync = &mem->sync;
1156 	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
1157 	drm_exec_until_all_locked(&ctx->exec) {
1158 		ret = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
1159 		drm_exec_retry_on_contention(&ctx->exec);
1160 		if (unlikely(ret))
1161 			goto error;
1162 
1163 		ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
1164 		drm_exec_retry_on_contention(&ctx->exec);
1165 		if (unlikely(ret))
1166 			goto error;
1167 	}
1168 	return 0;
1169 
1170 error:
1171 	pr_err("Failed to reserve buffers in ttm.\n");
1172 	drm_exec_fini(&ctx->exec);
1173 	return ret;
1174 }
1175 
1176 /**
1177  * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
1178  * @mem: KFD BO structure.
1179  * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
1180  * is used. Otherwise, a single VM associated with the BO.
1181  * @map_type: the mapping status that will be used to filter the VMs.
1182  * @ctx: the struct that will be used in unreserve_bo_and_vms().
1183  *
1184  * Returns 0 for success, negative for failure.
1185  */
1186 static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
1187 				struct amdgpu_vm *vm, enum bo_vm_match map_type,
1188 				struct bo_vm_reservation_context *ctx)
1189 {
1190 	struct kfd_mem_attachment *entry;
1191 	struct amdgpu_bo *bo = mem->bo;
1192 	int ret;
1193 
1194 	ctx->sync = &mem->sync;
1195 	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT |
1196 		      DRM_EXEC_IGNORE_DUPLICATES, 0);
1197 	drm_exec_until_all_locked(&ctx->exec) {
1198 		ctx->n_vms = 0;
1199 		list_for_each_entry(entry, &mem->attachments, list) {
1200 			if ((vm && vm != entry->bo_va->base.vm) ||
1201 				(entry->is_mapped != map_type
1202 				&& map_type != BO_VM_ALL))
1203 				continue;
1204 
1205 			ret = amdgpu_vm_lock_pd(entry->bo_va->base.vm,
1206 						&ctx->exec, 2);
1207 			drm_exec_retry_on_contention(&ctx->exec);
1208 			if (unlikely(ret))
1209 				goto error;
1210 			++ctx->n_vms;
1211 		}
1212 
1213 		ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
1214 		drm_exec_retry_on_contention(&ctx->exec);
1215 		if (unlikely(ret))
1216 			goto error;
1217 	}
1218 	return 0;
1219 
1220 error:
1221 	pr_err("Failed to reserve buffers in ttm.\n");
1222 	drm_exec_fini(&ctx->exec);
1223 	return ret;
1224 }
1225 
1226 /**
1227  * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1228  * @ctx: Reservation context to unreserve
1229  * @wait: Optionally wait for a sync object representing pending VM updates
1230  * @intr: Whether the wait is interruptible
1231  *
1232  * Also frees any resources allocated in
1233  * reserve_bo_and_(cond_)vm(s). Returns the status from
1234  * amdgpu_sync_wait.
1235  */
1236 static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
1237 				 bool wait, bool intr)
1238 {
1239 	int ret = 0;
1240 
1241 	if (wait)
1242 		ret = amdgpu_sync_wait(ctx->sync, intr);
1243 
1244 	drm_exec_fini(&ctx->exec);
1245 	ctx->sync = NULL;
1246 	return ret;
1247 }
1248 
1249 static int unmap_bo_from_gpuvm(struct kgd_mem *mem,
1250 				struct kfd_mem_attachment *entry,
1251 				struct amdgpu_sync *sync)
1252 {
1253 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1254 	struct amdgpu_device *adev = entry->adev;
1255 	struct amdgpu_vm *vm = bo_va->base.vm;
1256 
1257 	if (bo_va->queue_refcount) {
1258 		pr_debug("bo_va->queue_refcount %d\n", bo_va->queue_refcount);
1259 		return -EBUSY;
1260 	}
1261 
1262 	(void)amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
1263 
1264 	(void)amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
1265 
1266 	(void)amdgpu_sync_fence(sync, bo_va->last_pt_update);
1267 
1268 	return 0;
1269 }
1270 
1271 static int update_gpuvm_pte(struct kgd_mem *mem,
1272 			    struct kfd_mem_attachment *entry,
1273 			    struct amdgpu_sync *sync)
1274 {
1275 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1276 	struct amdgpu_device *adev = entry->adev;
1277 	int ret;
1278 
1279 	ret = kfd_mem_dmamap_attachment(mem, entry);
1280 	if (ret)
1281 		return ret;
1282 
1283 	/* Update the page tables  */
1284 	ret = amdgpu_vm_bo_update(adev, bo_va, false);
1285 	if (ret) {
1286 		pr_err("amdgpu_vm_bo_update failed\n");
1287 		return ret;
1288 	}
1289 
1290 	return amdgpu_sync_fence(sync, bo_va->last_pt_update);
1291 }
1292 
1293 static int map_bo_to_gpuvm(struct kgd_mem *mem,
1294 			   struct kfd_mem_attachment *entry,
1295 			   struct amdgpu_sync *sync,
1296 			   bool no_update_pte)
1297 {
1298 	int ret;
1299 
1300 	/* Set virtual address for the allocation */
1301 	ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0,
1302 			       amdgpu_bo_size(entry->bo_va->base.bo),
1303 			       entry->pte_flags);
1304 	if (ret) {
1305 		pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
1306 				entry->va, ret);
1307 		return ret;
1308 	}
1309 
1310 	if (no_update_pte)
1311 		return 0;
1312 
1313 	ret = update_gpuvm_pte(mem, entry, sync);
1314 	if (ret) {
1315 		pr_err("update_gpuvm_pte() failed\n");
1316 		goto update_gpuvm_pte_failed;
1317 	}
1318 
1319 	return 0;
1320 
1321 update_gpuvm_pte_failed:
1322 	unmap_bo_from_gpuvm(mem, entry, sync);
1323 	kfd_mem_dmaunmap_attachment(mem, entry);
1324 	return ret;
1325 }
1326 
1327 static int process_validate_vms(struct amdkfd_process_info *process_info,
1328 				struct ww_acquire_ctx *ticket)
1329 {
1330 	struct amdgpu_vm *peer_vm;
1331 	int ret;
1332 
1333 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1334 			    vm_list_node) {
1335 		ret = vm_validate_pt_pd_bos(peer_vm, ticket);
1336 		if (ret)
1337 			return ret;
1338 	}
1339 
1340 	return 0;
1341 }
1342 
1343 static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
1344 				 struct amdgpu_sync *sync)
1345 {
1346 	struct amdgpu_vm *peer_vm;
1347 	int ret;
1348 
1349 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1350 			    vm_list_node) {
1351 		struct amdgpu_bo *pd = peer_vm->root.bo;
1352 
1353 		ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
1354 				       AMDGPU_SYNC_NE_OWNER,
1355 				       AMDGPU_FENCE_OWNER_KFD);
1356 		if (ret)
1357 			return ret;
1358 	}
1359 
1360 	return 0;
1361 }
1362 
1363 static int process_update_pds(struct amdkfd_process_info *process_info,
1364 			      struct amdgpu_sync *sync)
1365 {
1366 	struct amdgpu_vm *peer_vm;
1367 	int ret;
1368 
1369 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1370 			    vm_list_node) {
1371 		ret = vm_update_pds(peer_vm, sync);
1372 		if (ret)
1373 			return ret;
1374 	}
1375 
1376 	return 0;
1377 }
1378 
1379 static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
1380 		       struct dma_fence **ef)
1381 {
1382 	struct amdkfd_process_info *info = NULL;
1383 	int ret;
1384 
1385 	if (!*process_info) {
1386 		info = kzalloc(sizeof(*info), GFP_KERNEL);
1387 		if (!info)
1388 			return -ENOMEM;
1389 
1390 		mutex_init(&info->lock);
1391 		mutex_init(&info->notifier_lock);
1392 		INIT_LIST_HEAD(&info->vm_list_head);
1393 		INIT_LIST_HEAD(&info->kfd_bo_list);
1394 		INIT_LIST_HEAD(&info->userptr_valid_list);
1395 		INIT_LIST_HEAD(&info->userptr_inval_list);
1396 
1397 		info->eviction_fence =
1398 			amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
1399 						   current->mm,
1400 						   NULL);
1401 		if (!info->eviction_fence) {
1402 			pr_err("Failed to create eviction fence\n");
1403 			ret = -ENOMEM;
1404 			goto create_evict_fence_fail;
1405 		}
1406 
1407 		info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
1408 		INIT_DELAYED_WORK(&info->restore_userptr_work,
1409 				  amdgpu_amdkfd_restore_userptr_worker);
1410 
1411 		*process_info = info;
1412 	}
1413 
1414 	vm->process_info = *process_info;
1415 
1416 	/* Validate page directory and attach eviction fence */
1417 	ret = amdgpu_bo_reserve(vm->root.bo, true);
1418 	if (ret)
1419 		goto reserve_pd_fail;
1420 	ret = vm_validate_pt_pd_bos(vm, NULL);
1421 	if (ret) {
1422 		pr_err("validate_pt_pd_bos() failed\n");
1423 		goto validate_pd_fail;
1424 	}
1425 	ret = amdgpu_bo_sync_wait(vm->root.bo,
1426 				  AMDGPU_FENCE_OWNER_KFD, false);
1427 	if (ret)
1428 		goto wait_pd_fail;
1429 	ret = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1);
1430 	if (ret)
1431 		goto reserve_shared_fail;
1432 	dma_resv_add_fence(vm->root.bo->tbo.base.resv,
1433 			   &vm->process_info->eviction_fence->base,
1434 			   DMA_RESV_USAGE_BOOKKEEP);
1435 	amdgpu_bo_unreserve(vm->root.bo);
1436 
1437 	/* Update process info */
1438 	mutex_lock(&vm->process_info->lock);
1439 	list_add_tail(&vm->vm_list_node,
1440 			&(vm->process_info->vm_list_head));
1441 	vm->process_info->n_vms++;
1442 	if (ef)
1443 		*ef = dma_fence_get(&vm->process_info->eviction_fence->base);
1444 	mutex_unlock(&vm->process_info->lock);
1445 
1446 	return 0;
1447 
1448 reserve_shared_fail:
1449 wait_pd_fail:
1450 validate_pd_fail:
1451 	amdgpu_bo_unreserve(vm->root.bo);
1452 reserve_pd_fail:
1453 	vm->process_info = NULL;
1454 	if (info) {
1455 		dma_fence_put(&info->eviction_fence->base);
1456 		*process_info = NULL;
1457 		put_pid(info->pid);
1458 create_evict_fence_fail:
1459 		mutex_destroy(&info->lock);
1460 		mutex_destroy(&info->notifier_lock);
1461 		kfree(info);
1462 	}
1463 	return ret;
1464 }
1465 
1466 /**
1467  * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
1468  * @bo: Handle of buffer object being pinned
1469  * @domain: Domain into which BO should be pinned
1470  *
1471  *   - USERPTR BOs are UNPINNABLE and will return error
1472  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1473  *     PIN count incremented. It is valid to PIN a BO multiple times
1474  *
1475  * Return: ZERO if successful in pinning, Non-Zero in case of error.
1476  */
1477 static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain)
1478 {
1479 	int ret = 0;
1480 
1481 	ret = amdgpu_bo_reserve(bo, false);
1482 	if (unlikely(ret))
1483 		return ret;
1484 
1485 	if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) {
1486 		/*
1487 		 * If bo is not contiguous on VRAM, move to system memory first to ensure
1488 		 * we can get contiguous VRAM space after evicting other BOs.
1489 		 */
1490 		if (!(bo->tbo.resource->placement & TTM_PL_FLAG_CONTIGUOUS)) {
1491 			struct ttm_operation_ctx ctx = { true, false };
1492 
1493 			amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
1494 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1495 			if (unlikely(ret)) {
1496 				pr_debug("validate bo 0x%p to GTT failed %d\n", &bo->tbo, ret);
1497 				goto out;
1498 			}
1499 		}
1500 	}
1501 
1502 	ret = amdgpu_bo_pin(bo, domain);
1503 	if (ret)
1504 		pr_err("Error in Pinning BO to domain: %d\n", domain);
1505 
1506 	amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
1507 out:
1508 	amdgpu_bo_unreserve(bo);
1509 	return ret;
1510 }
1511 
1512 /**
1513  * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
1514  * @bo: Handle of buffer object being unpinned
1515  *
1516  *   - Is a illegal request for USERPTR BOs and is ignored
1517  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1518  *     PIN count decremented. Calls to UNPIN must balance calls to PIN
1519  */
1520 static void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo)
1521 {
1522 	int ret = 0;
1523 
1524 	ret = amdgpu_bo_reserve(bo, false);
1525 	if (unlikely(ret))
1526 		return;
1527 
1528 	amdgpu_bo_unpin(bo);
1529 	amdgpu_bo_unreserve(bo);
1530 }
1531 
1532 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
1533 					   struct amdgpu_vm *avm,
1534 					   void **process_info,
1535 					   struct dma_fence **ef)
1536 {
1537 	int ret;
1538 
1539 	/* Already a compute VM? */
1540 	if (avm->process_info)
1541 		return -EINVAL;
1542 
1543 	/* Convert VM into a compute VM */
1544 	ret = amdgpu_vm_make_compute(adev, avm);
1545 	if (ret)
1546 		return ret;
1547 
1548 	/* Initialize KFD part of the VM and process info */
1549 	ret = init_kfd_vm(avm, process_info, ef);
1550 	if (ret)
1551 		return ret;
1552 
1553 	amdgpu_vm_set_task_info(avm);
1554 
1555 	return 0;
1556 }
1557 
1558 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1559 				    struct amdgpu_vm *vm)
1560 {
1561 	struct amdkfd_process_info *process_info = vm->process_info;
1562 
1563 	if (!process_info)
1564 		return;
1565 
1566 	/* Update process info */
1567 	mutex_lock(&process_info->lock);
1568 	process_info->n_vms--;
1569 	list_del(&vm->vm_list_node);
1570 	mutex_unlock(&process_info->lock);
1571 
1572 	vm->process_info = NULL;
1573 
1574 	/* Release per-process resources when last compute VM is destroyed */
1575 	if (!process_info->n_vms) {
1576 		WARN_ON(!list_empty(&process_info->kfd_bo_list));
1577 		WARN_ON(!list_empty(&process_info->userptr_valid_list));
1578 		WARN_ON(!list_empty(&process_info->userptr_inval_list));
1579 
1580 		dma_fence_put(&process_info->eviction_fence->base);
1581 		cancel_delayed_work_sync(&process_info->restore_userptr_work);
1582 		put_pid(process_info->pid);
1583 		mutex_destroy(&process_info->lock);
1584 		mutex_destroy(&process_info->notifier_lock);
1585 		kfree(process_info);
1586 	}
1587 }
1588 
1589 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv)
1590 {
1591 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1592 	struct amdgpu_bo *pd = avm->root.bo;
1593 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1594 
1595 	if (adev->asic_type < CHIP_VEGA10)
1596 		return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1597 	return avm->pd_phys_addr;
1598 }
1599 
1600 void amdgpu_amdkfd_block_mmu_notifications(void *p)
1601 {
1602 	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1603 
1604 	mutex_lock(&pinfo->lock);
1605 	WRITE_ONCE(pinfo->block_mmu_notifications, true);
1606 	mutex_unlock(&pinfo->lock);
1607 }
1608 
1609 int amdgpu_amdkfd_criu_resume(void *p)
1610 {
1611 	int ret = 0;
1612 	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1613 
1614 	mutex_lock(&pinfo->lock);
1615 	pr_debug("scheduling work\n");
1616 	mutex_lock(&pinfo->notifier_lock);
1617 	pinfo->evicted_bos++;
1618 	mutex_unlock(&pinfo->notifier_lock);
1619 	if (!READ_ONCE(pinfo->block_mmu_notifications)) {
1620 		ret = -EINVAL;
1621 		goto out_unlock;
1622 	}
1623 	WRITE_ONCE(pinfo->block_mmu_notifications, false);
1624 	queue_delayed_work(system_freezable_wq,
1625 			   &pinfo->restore_userptr_work, 0);
1626 
1627 out_unlock:
1628 	mutex_unlock(&pinfo->lock);
1629 	return ret;
1630 }
1631 
1632 size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
1633 					  uint8_t xcp_id)
1634 {
1635 	uint64_t reserved_for_pt =
1636 		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
1637 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1638 	uint64_t reserved_for_ras = (con ? con->reserved_pages_in_bytes : 0);
1639 	ssize_t available;
1640 	uint64_t vram_available, system_mem_available, ttm_mem_available;
1641 
1642 	spin_lock(&kfd_mem_limit.mem_limit_lock);
1643 	vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id)
1644 		- adev->kfd.vram_used_aligned[xcp_id]
1645 		- atomic64_read(&adev->vram_pin_size)
1646 		- reserved_for_pt
1647 		- reserved_for_ras;
1648 
1649 	if (adev->apu_prefer_gtt) {
1650 		system_mem_available = no_system_mem_limit ?
1651 					kfd_mem_limit.max_system_mem_limit :
1652 					kfd_mem_limit.max_system_mem_limit -
1653 					kfd_mem_limit.system_mem_used;
1654 
1655 		ttm_mem_available = kfd_mem_limit.max_ttm_mem_limit -
1656 				kfd_mem_limit.ttm_mem_used;
1657 
1658 		available = min3(system_mem_available, ttm_mem_available,
1659 				 vram_available);
1660 		available = ALIGN_DOWN(available, PAGE_SIZE);
1661 	} else {
1662 		available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN);
1663 	}
1664 
1665 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
1666 
1667 	if (available < 0)
1668 		available = 0;
1669 
1670 	return available;
1671 }
1672 
1673 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1674 		struct amdgpu_device *adev, uint64_t va, uint64_t size,
1675 		void *drm_priv, struct kgd_mem **mem,
1676 		uint64_t *offset, uint32_t flags, bool criu_resume)
1677 {
1678 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1679 	struct amdgpu_fpriv *fpriv = container_of(avm, struct amdgpu_fpriv, vm);
1680 	enum ttm_bo_type bo_type = ttm_bo_type_device;
1681 	struct sg_table *sg = NULL;
1682 	uint64_t user_addr = 0;
1683 	struct amdgpu_bo *bo;
1684 	struct drm_gem_object *gobj = NULL;
1685 	u32 domain, alloc_domain;
1686 	uint64_t aligned_size;
1687 	int8_t xcp_id = -1;
1688 	u64 alloc_flags;
1689 	int ret;
1690 
1691 	/*
1692 	 * Check on which domain to allocate BO
1693 	 */
1694 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1695 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1696 
1697 		if (adev->apu_prefer_gtt) {
1698 			domain = AMDGPU_GEM_DOMAIN_GTT;
1699 			alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1700 			alloc_flags = 0;
1701 		} else {
1702 			alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1703 			alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1704 			AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0;
1705 
1706 			/* For contiguous VRAM allocation */
1707 			if (flags & KFD_IOC_ALLOC_MEM_FLAGS_CONTIGUOUS)
1708 				alloc_flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1709 		}
1710 		xcp_id = fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION ?
1711 					0 : fpriv->xcp_id;
1712 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1713 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1714 		alloc_flags = 0;
1715 	} else {
1716 		domain = AMDGPU_GEM_DOMAIN_GTT;
1717 		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1718 		alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE;
1719 
1720 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1721 			if (!offset || !*offset)
1722 				return -EINVAL;
1723 			user_addr = untagged_addr(*offset);
1724 		} else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1725 				    KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1726 			bo_type = ttm_bo_type_sg;
1727 			if (size > UINT_MAX)
1728 				return -EINVAL;
1729 			sg = create_sg_table(*offset, size);
1730 			if (!sg)
1731 				return -ENOMEM;
1732 		} else {
1733 			return -EINVAL;
1734 		}
1735 	}
1736 
1737 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT)
1738 		alloc_flags |= AMDGPU_GEM_CREATE_COHERENT;
1739 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENT)
1740 		alloc_flags |= AMDGPU_GEM_CREATE_EXT_COHERENT;
1741 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED)
1742 		alloc_flags |= AMDGPU_GEM_CREATE_UNCACHED;
1743 
1744 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1745 	if (!*mem) {
1746 		ret = -ENOMEM;
1747 		goto err;
1748 	}
1749 	INIT_LIST_HEAD(&(*mem)->attachments);
1750 	mutex_init(&(*mem)->lock);
1751 	(*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1752 
1753 	/* Workaround for AQL queue wraparound bug. Map the same
1754 	 * memory twice. That means we only actually allocate half
1755 	 * the memory.
1756 	 */
1757 	if ((*mem)->aql_queue)
1758 		size >>= 1;
1759 	aligned_size = PAGE_ALIGN(size);
1760 
1761 	(*mem)->alloc_flags = flags;
1762 
1763 	amdgpu_sync_create(&(*mem)->sync);
1764 
1765 	ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags,
1766 					      xcp_id);
1767 	if (ret) {
1768 		pr_debug("Insufficient memory\n");
1769 		goto err_reserve_limit;
1770 	}
1771 
1772 	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s xcp_id %d\n",
1773 		 va, (*mem)->aql_queue ? size << 1 : size,
1774 		 domain_string(alloc_domain), xcp_id);
1775 
1776 	ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags,
1777 				       bo_type, NULL, &gobj, xcp_id + 1);
1778 	if (ret) {
1779 		pr_debug("Failed to create BO on domain %s. ret %d\n",
1780 			 domain_string(alloc_domain), ret);
1781 		goto err_bo_create;
1782 	}
1783 	ret = drm_vma_node_allow(&gobj->vma_node, drm_priv);
1784 	if (ret) {
1785 		pr_debug("Failed to allow vma node access. ret %d\n", ret);
1786 		goto err_node_allow;
1787 	}
1788 	ret = drm_gem_handle_create(adev->kfd.client.file, gobj, &(*mem)->gem_handle);
1789 	if (ret)
1790 		goto err_gem_handle_create;
1791 	bo = gem_to_amdgpu_bo(gobj);
1792 	if (bo_type == ttm_bo_type_sg) {
1793 		bo->tbo.sg = sg;
1794 		bo->tbo.ttm->sg = sg;
1795 	}
1796 	bo->kfd_bo = *mem;
1797 	(*mem)->bo = bo;
1798 	if (user_addr)
1799 		bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO;
1800 
1801 	(*mem)->va = va;
1802 	(*mem)->domain = domain;
1803 	(*mem)->mapped_to_gpu_memory = 0;
1804 	(*mem)->process_info = avm->process_info;
1805 
1806 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1807 
1808 	if (user_addr) {
1809 		pr_debug("creating userptr BO for user_addr = %llx\n", user_addr);
1810 		ret = init_user_pages(*mem, user_addr, criu_resume);
1811 		if (ret)
1812 			goto allocate_init_user_pages_failed;
1813 	} else  if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1814 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1815 		ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT);
1816 		if (ret) {
1817 			pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n");
1818 			goto err_pin_bo;
1819 		}
1820 		bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
1821 		bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
1822 	} else {
1823 		mutex_lock(&avm->process_info->lock);
1824 		if (avm->process_info->eviction_fence &&
1825 		    !dma_fence_is_signaled(&avm->process_info->eviction_fence->base))
1826 			ret = amdgpu_amdkfd_bo_validate_and_fence(bo, domain,
1827 				&avm->process_info->eviction_fence->base);
1828 		mutex_unlock(&avm->process_info->lock);
1829 		if (ret)
1830 			goto err_validate_bo;
1831 	}
1832 
1833 	if (offset)
1834 		*offset = amdgpu_bo_mmap_offset(bo);
1835 
1836 	return 0;
1837 
1838 allocate_init_user_pages_failed:
1839 err_pin_bo:
1840 err_validate_bo:
1841 	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1842 	drm_gem_handle_delete(adev->kfd.client.file, (*mem)->gem_handle);
1843 err_gem_handle_create:
1844 	drm_vma_node_revoke(&gobj->vma_node, drm_priv);
1845 err_node_allow:
1846 	/* Don't unreserve system mem limit twice */
1847 	goto err_reserve_limit;
1848 err_bo_create:
1849 	amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags, xcp_id);
1850 err_reserve_limit:
1851 	amdgpu_sync_free(&(*mem)->sync);
1852 	mutex_destroy(&(*mem)->lock);
1853 	if (gobj)
1854 		drm_gem_object_put(gobj);
1855 	else
1856 		kfree(*mem);
1857 err:
1858 	if (sg) {
1859 		sg_free_table(sg);
1860 		kfree(sg);
1861 	}
1862 	return ret;
1863 }
1864 
1865 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1866 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
1867 		uint64_t *size)
1868 {
1869 	struct amdkfd_process_info *process_info = mem->process_info;
1870 	unsigned long bo_size = mem->bo->tbo.base.size;
1871 	bool use_release_notifier = (mem->bo->kfd_bo == mem);
1872 	struct kfd_mem_attachment *entry, *tmp;
1873 	struct bo_vm_reservation_context ctx;
1874 	unsigned int mapped_to_gpu_memory;
1875 	int ret;
1876 	bool is_imported = false;
1877 
1878 	mutex_lock(&mem->lock);
1879 
1880 	/* Unpin MMIO/DOORBELL BO's that were pinned during allocation */
1881 	if (mem->alloc_flags &
1882 	    (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1883 	     KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1884 		amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo);
1885 	}
1886 
1887 	mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1888 	is_imported = mem->is_imported;
1889 	mutex_unlock(&mem->lock);
1890 	/* lock is not needed after this, since mem is unused and will
1891 	 * be freed anyway
1892 	 */
1893 
1894 	if (mapped_to_gpu_memory > 0) {
1895 		pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1896 				mem->va, bo_size);
1897 		return -EBUSY;
1898 	}
1899 
1900 	/* Make sure restore workers don't access the BO any more */
1901 	mutex_lock(&process_info->lock);
1902 	list_del(&mem->validate_list);
1903 	mutex_unlock(&process_info->lock);
1904 
1905 	/* Cleanup user pages and MMU notifiers */
1906 	if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
1907 		amdgpu_hmm_unregister(mem->bo);
1908 		mutex_lock(&process_info->notifier_lock);
1909 		amdgpu_ttm_tt_discard_user_pages(mem->bo->tbo.ttm, mem->range);
1910 		mutex_unlock(&process_info->notifier_lock);
1911 	}
1912 
1913 	ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1914 	if (unlikely(ret))
1915 		return ret;
1916 
1917 	amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1918 					process_info->eviction_fence);
1919 	pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1920 		mem->va + bo_size * (1 + mem->aql_queue));
1921 
1922 	/* Remove from VM internal data structures */
1923 	list_for_each_entry_safe(entry, tmp, &mem->attachments, list) {
1924 		kfd_mem_dmaunmap_attachment(mem, entry);
1925 		kfd_mem_detach(entry);
1926 	}
1927 
1928 	ret = unreserve_bo_and_vms(&ctx, false, false);
1929 
1930 	/* Free the sync object */
1931 	amdgpu_sync_free(&mem->sync);
1932 
1933 	/* If the SG is not NULL, it's one we created for a doorbell or mmio
1934 	 * remap BO. We need to free it.
1935 	 */
1936 	if (mem->bo->tbo.sg) {
1937 		sg_free_table(mem->bo->tbo.sg);
1938 		kfree(mem->bo->tbo.sg);
1939 	}
1940 
1941 	/* Update the size of the BO being freed if it was allocated from
1942 	 * VRAM and is not imported. For APP APU VRAM allocations are done
1943 	 * in GTT domain
1944 	 */
1945 	if (size) {
1946 		if (!is_imported &&
1947 		   (mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM ||
1948 		   (adev->apu_prefer_gtt &&
1949 		    mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_GTT)))
1950 			*size = bo_size;
1951 		else
1952 			*size = 0;
1953 	}
1954 
1955 	/* Free the BO*/
1956 	drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
1957 	drm_gem_handle_delete(adev->kfd.client.file, mem->gem_handle);
1958 	if (mem->dmabuf) {
1959 		dma_buf_put(mem->dmabuf);
1960 		mem->dmabuf = NULL;
1961 	}
1962 	mutex_destroy(&mem->lock);
1963 
1964 	/* If this releases the last reference, it will end up calling
1965 	 * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why
1966 	 * this needs to be the last call here.
1967 	 */
1968 	drm_gem_object_put(&mem->bo->tbo.base);
1969 
1970 	/*
1971 	 * For kgd_mem allocated in amdgpu_amdkfd_gpuvm_import_dmabuf(),
1972 	 * explicitly free it here.
1973 	 */
1974 	if (!use_release_notifier)
1975 		kfree(mem);
1976 
1977 	return ret;
1978 }
1979 
1980 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1981 		struct amdgpu_device *adev, struct kgd_mem *mem,
1982 		void *drm_priv)
1983 {
1984 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1985 	int ret;
1986 	struct amdgpu_bo *bo;
1987 	uint32_t domain;
1988 	struct kfd_mem_attachment *entry;
1989 	struct bo_vm_reservation_context ctx;
1990 	unsigned long bo_size;
1991 	bool is_invalid_userptr = false;
1992 
1993 	bo = mem->bo;
1994 	if (!bo) {
1995 		pr_err("Invalid BO when mapping memory to GPU\n");
1996 		return -EINVAL;
1997 	}
1998 
1999 	/* Make sure restore is not running concurrently. Since we
2000 	 * don't map invalid userptr BOs, we rely on the next restore
2001 	 * worker to do the mapping
2002 	 */
2003 	mutex_lock(&mem->process_info->lock);
2004 
2005 	/* Lock notifier lock. If we find an invalid userptr BO, we can be
2006 	 * sure that the MMU notifier is no longer running
2007 	 * concurrently and the queues are actually stopped
2008 	 */
2009 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2010 		mutex_lock(&mem->process_info->notifier_lock);
2011 		is_invalid_userptr = !!mem->invalid;
2012 		mutex_unlock(&mem->process_info->notifier_lock);
2013 	}
2014 
2015 	mutex_lock(&mem->lock);
2016 
2017 	domain = mem->domain;
2018 	bo_size = bo->tbo.base.size;
2019 
2020 	pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
2021 			mem->va,
2022 			mem->va + bo_size * (1 + mem->aql_queue),
2023 			avm, domain_string(domain));
2024 
2025 	if (!kfd_mem_is_attached(avm, mem)) {
2026 		ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue);
2027 		if (ret)
2028 			goto out;
2029 	}
2030 
2031 	ret = reserve_bo_and_vm(mem, avm, &ctx);
2032 	if (unlikely(ret))
2033 		goto out;
2034 
2035 	/* Userptr can be marked as "not invalid", but not actually be
2036 	 * validated yet (still in the system domain). In that case
2037 	 * the queues are still stopped and we can leave mapping for
2038 	 * the next restore worker
2039 	 */
2040 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
2041 	    bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
2042 		is_invalid_userptr = true;
2043 
2044 	ret = vm_validate_pt_pd_bos(avm, NULL);
2045 	if (unlikely(ret))
2046 		goto out_unreserve;
2047 
2048 	list_for_each_entry(entry, &mem->attachments, list) {
2049 		if (entry->bo_va->base.vm != avm || entry->is_mapped)
2050 			continue;
2051 
2052 		pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
2053 			 entry->va, entry->va + bo_size, entry);
2054 
2055 		ret = map_bo_to_gpuvm(mem, entry, ctx.sync,
2056 				      is_invalid_userptr);
2057 		if (ret) {
2058 			pr_err("Failed to map bo to gpuvm\n");
2059 			goto out_unreserve;
2060 		}
2061 
2062 		ret = vm_update_pds(avm, ctx.sync);
2063 		if (ret) {
2064 			pr_err("Failed to update page directories\n");
2065 			goto out_unreserve;
2066 		}
2067 
2068 		entry->is_mapped = true;
2069 		mem->mapped_to_gpu_memory++;
2070 		pr_debug("\t INC mapping count %d\n",
2071 			 mem->mapped_to_gpu_memory);
2072 	}
2073 
2074 	ret = unreserve_bo_and_vms(&ctx, false, false);
2075 
2076 	goto out;
2077 
2078 out_unreserve:
2079 	unreserve_bo_and_vms(&ctx, false, false);
2080 out:
2081 	mutex_unlock(&mem->process_info->lock);
2082 	mutex_unlock(&mem->lock);
2083 	return ret;
2084 }
2085 
2086 int amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem, void *drm_priv)
2087 {
2088 	struct kfd_mem_attachment *entry;
2089 	struct amdgpu_vm *vm;
2090 	int ret;
2091 
2092 	vm = drm_priv_to_vm(drm_priv);
2093 
2094 	mutex_lock(&mem->lock);
2095 
2096 	ret = amdgpu_bo_reserve(mem->bo, true);
2097 	if (ret)
2098 		goto out;
2099 
2100 	list_for_each_entry(entry, &mem->attachments, list) {
2101 		if (entry->bo_va->base.vm != vm)
2102 			continue;
2103 		if (entry->bo_va->base.bo->tbo.ttm &&
2104 		    !entry->bo_va->base.bo->tbo.ttm->sg)
2105 			continue;
2106 
2107 		kfd_mem_dmaunmap_attachment(mem, entry);
2108 	}
2109 
2110 	amdgpu_bo_unreserve(mem->bo);
2111 out:
2112 	mutex_unlock(&mem->lock);
2113 
2114 	return ret;
2115 }
2116 
2117 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
2118 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
2119 {
2120 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2121 	unsigned long bo_size = mem->bo->tbo.base.size;
2122 	struct kfd_mem_attachment *entry;
2123 	struct bo_vm_reservation_context ctx;
2124 	int ret;
2125 
2126 	mutex_lock(&mem->lock);
2127 
2128 	ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx);
2129 	if (unlikely(ret))
2130 		goto out;
2131 	/* If no VMs were reserved, it means the BO wasn't actually mapped */
2132 	if (ctx.n_vms == 0) {
2133 		ret = -EINVAL;
2134 		goto unreserve_out;
2135 	}
2136 
2137 	ret = vm_validate_pt_pd_bos(avm, NULL);
2138 	if (unlikely(ret))
2139 		goto unreserve_out;
2140 
2141 	pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
2142 		mem->va,
2143 		mem->va + bo_size * (1 + mem->aql_queue),
2144 		avm);
2145 
2146 	list_for_each_entry(entry, &mem->attachments, list) {
2147 		if (entry->bo_va->base.vm != avm || !entry->is_mapped)
2148 			continue;
2149 
2150 		pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
2151 			 entry->va, entry->va + bo_size, entry);
2152 
2153 		ret = unmap_bo_from_gpuvm(mem, entry, ctx.sync);
2154 		if (ret)
2155 			goto unreserve_out;
2156 
2157 		entry->is_mapped = false;
2158 
2159 		mem->mapped_to_gpu_memory--;
2160 		pr_debug("\t DEC mapping count %d\n",
2161 			 mem->mapped_to_gpu_memory);
2162 	}
2163 
2164 unreserve_out:
2165 	unreserve_bo_and_vms(&ctx, false, false);
2166 out:
2167 	mutex_unlock(&mem->lock);
2168 	return ret;
2169 }
2170 
2171 int amdgpu_amdkfd_gpuvm_sync_memory(
2172 		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr)
2173 {
2174 	struct amdgpu_sync sync;
2175 	int ret;
2176 
2177 	amdgpu_sync_create(&sync);
2178 
2179 	mutex_lock(&mem->lock);
2180 	amdgpu_sync_clone(&mem->sync, &sync);
2181 	mutex_unlock(&mem->lock);
2182 
2183 	ret = amdgpu_sync_wait(&sync, intr);
2184 	amdgpu_sync_free(&sync);
2185 	return ret;
2186 }
2187 
2188 /**
2189  * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count
2190  * @bo: Buffer object to be mapped
2191  * @bo_gart: Return bo reference
2192  *
2193  * Before return, bo reference count is incremented. To release the reference and unpin/
2194  * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
2195  */
2196 int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_bo *bo, struct amdgpu_bo **bo_gart)
2197 {
2198 	int ret;
2199 
2200 	ret = amdgpu_bo_reserve(bo, true);
2201 	if (ret) {
2202 		pr_err("Failed to reserve bo. ret %d\n", ret);
2203 		goto err_reserve_bo_failed;
2204 	}
2205 
2206 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2207 	if (ret) {
2208 		pr_err("Failed to pin bo. ret %d\n", ret);
2209 		goto err_pin_bo_failed;
2210 	}
2211 
2212 	ret = amdgpu_ttm_alloc_gart(&bo->tbo);
2213 	if (ret) {
2214 		pr_err("Failed to bind bo to GART. ret %d\n", ret);
2215 		goto err_map_bo_gart_failed;
2216 	}
2217 
2218 	amdgpu_amdkfd_remove_eviction_fence(
2219 		bo, bo->vm_bo->vm->process_info->eviction_fence);
2220 
2221 	amdgpu_bo_unreserve(bo);
2222 
2223 	*bo_gart = amdgpu_bo_ref(bo);
2224 
2225 	return 0;
2226 
2227 err_map_bo_gart_failed:
2228 	amdgpu_bo_unpin(bo);
2229 err_pin_bo_failed:
2230 	amdgpu_bo_unreserve(bo);
2231 err_reserve_bo_failed:
2232 
2233 	return ret;
2234 }
2235 
2236 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Map a GTT BO for kernel CPU access
2237  *
2238  * @mem: Buffer object to be mapped for CPU access
2239  * @kptr[out]: pointer in kernel CPU address space
2240  * @size[out]: size of the buffer
2241  *
2242  * Pins the BO and maps it for kernel CPU access. The eviction fence is removed
2243  * from the BO, since pinned BOs cannot be evicted. The bo must remain on the
2244  * validate_list, so the GPU mapping can be restored after a page table was
2245  * evicted.
2246  *
2247  * Return: 0 on success, error code on failure
2248  */
2249 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,
2250 					     void **kptr, uint64_t *size)
2251 {
2252 	int ret;
2253 	struct amdgpu_bo *bo = mem->bo;
2254 
2255 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2256 		pr_err("userptr can't be mapped to kernel\n");
2257 		return -EINVAL;
2258 	}
2259 
2260 	mutex_lock(&mem->process_info->lock);
2261 
2262 	ret = amdgpu_bo_reserve(bo, true);
2263 	if (ret) {
2264 		pr_err("Failed to reserve bo. ret %d\n", ret);
2265 		goto bo_reserve_failed;
2266 	}
2267 
2268 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2269 	if (ret) {
2270 		pr_err("Failed to pin bo. ret %d\n", ret);
2271 		goto pin_failed;
2272 	}
2273 
2274 	ret = amdgpu_bo_kmap(bo, kptr);
2275 	if (ret) {
2276 		pr_err("Failed to map bo to kernel. ret %d\n", ret);
2277 		goto kmap_failed;
2278 	}
2279 
2280 	amdgpu_amdkfd_remove_eviction_fence(
2281 		bo, mem->process_info->eviction_fence);
2282 
2283 	if (size)
2284 		*size = amdgpu_bo_size(bo);
2285 
2286 	amdgpu_bo_unreserve(bo);
2287 
2288 	mutex_unlock(&mem->process_info->lock);
2289 	return 0;
2290 
2291 kmap_failed:
2292 	amdgpu_bo_unpin(bo);
2293 pin_failed:
2294 	amdgpu_bo_unreserve(bo);
2295 bo_reserve_failed:
2296 	mutex_unlock(&mem->process_info->lock);
2297 
2298 	return ret;
2299 }
2300 
2301 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Unmap a GTT BO for kernel CPU access
2302  *
2303  * @mem: Buffer object to be unmapped for CPU access
2304  *
2305  * Removes the kernel CPU mapping and unpins the BO. It does not restore the
2306  * eviction fence, so this function should only be used for cleanup before the
2307  * BO is destroyed.
2308  */
2309 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem)
2310 {
2311 	struct amdgpu_bo *bo = mem->bo;
2312 
2313 	(void)amdgpu_bo_reserve(bo, true);
2314 	amdgpu_bo_kunmap(bo);
2315 	amdgpu_bo_unpin(bo);
2316 	amdgpu_bo_unreserve(bo);
2317 }
2318 
2319 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
2320 					  struct kfd_vm_fault_info *mem)
2321 {
2322 	if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
2323 		*mem = *adev->gmc.vm_fault_info;
2324 		mb(); /* make sure read happened */
2325 		atomic_set(&adev->gmc.vm_fault_info_updated, 0);
2326 	}
2327 	return 0;
2328 }
2329 
2330 static int import_obj_create(struct amdgpu_device *adev,
2331 			     struct dma_buf *dma_buf,
2332 			     struct drm_gem_object *obj,
2333 			     uint64_t va, void *drm_priv,
2334 			     struct kgd_mem **mem, uint64_t *size,
2335 			     uint64_t *mmap_offset)
2336 {
2337 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2338 	struct amdgpu_bo *bo;
2339 	int ret;
2340 
2341 	bo = gem_to_amdgpu_bo(obj);
2342 	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
2343 				    AMDGPU_GEM_DOMAIN_GTT)))
2344 		/* Only VRAM and GTT BOs are supported */
2345 		return -EINVAL;
2346 
2347 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2348 	if (!*mem)
2349 		return -ENOMEM;
2350 
2351 	ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
2352 	if (ret)
2353 		goto err_free_mem;
2354 
2355 	if (size)
2356 		*size = amdgpu_bo_size(bo);
2357 
2358 	if (mmap_offset)
2359 		*mmap_offset = amdgpu_bo_mmap_offset(bo);
2360 
2361 	INIT_LIST_HEAD(&(*mem)->attachments);
2362 	mutex_init(&(*mem)->lock);
2363 
2364 	(*mem)->alloc_flags =
2365 		((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2366 		KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
2367 		| KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
2368 		| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
2369 
2370 	get_dma_buf(dma_buf);
2371 	(*mem)->dmabuf = dma_buf;
2372 	(*mem)->bo = bo;
2373 	(*mem)->va = va;
2374 	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) &&
2375 			 !adev->apu_prefer_gtt ?
2376 			 AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
2377 
2378 	(*mem)->mapped_to_gpu_memory = 0;
2379 	(*mem)->process_info = avm->process_info;
2380 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
2381 	amdgpu_sync_create(&(*mem)->sync);
2382 	(*mem)->is_imported = true;
2383 
2384 	mutex_lock(&avm->process_info->lock);
2385 	if (avm->process_info->eviction_fence &&
2386 	    !dma_fence_is_signaled(&avm->process_info->eviction_fence->base))
2387 		ret = amdgpu_amdkfd_bo_validate_and_fence(bo, (*mem)->domain,
2388 				&avm->process_info->eviction_fence->base);
2389 	mutex_unlock(&avm->process_info->lock);
2390 	if (ret)
2391 		goto err_remove_mem;
2392 
2393 	return 0;
2394 
2395 err_remove_mem:
2396 	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
2397 	drm_vma_node_revoke(&obj->vma_node, drm_priv);
2398 err_free_mem:
2399 	kfree(*mem);
2400 	return ret;
2401 }
2402 
2403 int amdgpu_amdkfd_gpuvm_import_dmabuf_fd(struct amdgpu_device *adev, int fd,
2404 					 uint64_t va, void *drm_priv,
2405 					 struct kgd_mem **mem, uint64_t *size,
2406 					 uint64_t *mmap_offset)
2407 {
2408 	struct drm_gem_object *obj;
2409 	uint32_t handle;
2410 	int ret;
2411 
2412 	ret = drm_gem_prime_fd_to_handle(&adev->ddev, adev->kfd.client.file, fd,
2413 					 &handle);
2414 	if (ret)
2415 		return ret;
2416 	obj = drm_gem_object_lookup(adev->kfd.client.file, handle);
2417 	if (!obj) {
2418 		ret = -EINVAL;
2419 		goto err_release_handle;
2420 	}
2421 
2422 	ret = import_obj_create(adev, obj->dma_buf, obj, va, drm_priv, mem, size,
2423 				mmap_offset);
2424 	if (ret)
2425 		goto err_put_obj;
2426 
2427 	(*mem)->gem_handle = handle;
2428 
2429 	return 0;
2430 
2431 err_put_obj:
2432 	drm_gem_object_put(obj);
2433 err_release_handle:
2434 	drm_gem_handle_delete(adev->kfd.client.file, handle);
2435 	return ret;
2436 }
2437 
2438 int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem,
2439 				      struct dma_buf **dma_buf)
2440 {
2441 	int ret;
2442 
2443 	mutex_lock(&mem->lock);
2444 	ret = kfd_mem_export_dmabuf(mem);
2445 	if (ret)
2446 		goto out;
2447 
2448 	get_dma_buf(mem->dmabuf);
2449 	*dma_buf = mem->dmabuf;
2450 out:
2451 	mutex_unlock(&mem->lock);
2452 	return ret;
2453 }
2454 
2455 /* Evict a userptr BO by stopping the queues if necessary
2456  *
2457  * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
2458  * cannot do any memory allocations, and cannot take any locks that
2459  * are held elsewhere while allocating memory.
2460  *
2461  * It doesn't do anything to the BO itself. The real work happens in
2462  * restore, where we get updated page addresses. This function only
2463  * ensures that GPU access to the BO is stopped.
2464  */
2465 int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
2466 				unsigned long cur_seq, struct kgd_mem *mem)
2467 {
2468 	struct amdkfd_process_info *process_info = mem->process_info;
2469 	int r = 0;
2470 
2471 	/* Do not process MMU notifications during CRIU restore until
2472 	 * KFD_CRIU_OP_RESUME IOCTL is received
2473 	 */
2474 	if (READ_ONCE(process_info->block_mmu_notifications))
2475 		return 0;
2476 
2477 	mutex_lock(&process_info->notifier_lock);
2478 	mmu_interval_set_seq(mni, cur_seq);
2479 
2480 	mem->invalid++;
2481 	if (++process_info->evicted_bos == 1) {
2482 		/* First eviction, stop the queues */
2483 		r = kgd2kfd_quiesce_mm(mni->mm,
2484 				       KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
2485 
2486 		if (r && r != -ESRCH)
2487 			pr_err("Failed to quiesce KFD\n");
2488 
2489 		if (r != -ESRCH)
2490 			queue_delayed_work(system_freezable_wq,
2491 				&process_info->restore_userptr_work,
2492 				msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2493 	}
2494 	mutex_unlock(&process_info->notifier_lock);
2495 
2496 	return r;
2497 }
2498 
2499 /* Update invalid userptr BOs
2500  *
2501  * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
2502  * userptr_inval_list and updates user pages for all BOs that have
2503  * been invalidated since their last update.
2504  */
2505 static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
2506 				     struct mm_struct *mm)
2507 {
2508 	struct kgd_mem *mem, *tmp_mem;
2509 	struct amdgpu_bo *bo;
2510 	struct ttm_operation_ctx ctx = { false, false };
2511 	uint32_t invalid;
2512 	int ret = 0;
2513 
2514 	mutex_lock(&process_info->notifier_lock);
2515 
2516 	/* Move all invalidated BOs to the userptr_inval_list */
2517 	list_for_each_entry_safe(mem, tmp_mem,
2518 				 &process_info->userptr_valid_list,
2519 				 validate_list)
2520 		if (mem->invalid)
2521 			list_move_tail(&mem->validate_list,
2522 				       &process_info->userptr_inval_list);
2523 
2524 	/* Go through userptr_inval_list and update any invalid user_pages */
2525 	list_for_each_entry(mem, &process_info->userptr_inval_list,
2526 			    validate_list) {
2527 		invalid = mem->invalid;
2528 		if (!invalid)
2529 			/* BO hasn't been invalidated since the last
2530 			 * revalidation attempt. Keep its page list.
2531 			 */
2532 			continue;
2533 
2534 		bo = mem->bo;
2535 
2536 		amdgpu_ttm_tt_discard_user_pages(bo->tbo.ttm, mem->range);
2537 		mem->range = NULL;
2538 
2539 		/* BO reservations and getting user pages (hmm_range_fault)
2540 		 * must happen outside the notifier lock
2541 		 */
2542 		mutex_unlock(&process_info->notifier_lock);
2543 
2544 		/* Move the BO to system (CPU) domain if necessary to unmap
2545 		 * and free the SG table
2546 		 */
2547 		if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) {
2548 			if (amdgpu_bo_reserve(bo, true))
2549 				return -EAGAIN;
2550 			amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
2551 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2552 			amdgpu_bo_unreserve(bo);
2553 			if (ret) {
2554 				pr_err("%s: Failed to invalidate userptr BO\n",
2555 				       __func__);
2556 				return -EAGAIN;
2557 			}
2558 		}
2559 
2560 		/* Get updated user pages */
2561 		ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages,
2562 						   &mem->range);
2563 		if (ret) {
2564 			pr_debug("Failed %d to get user pages\n", ret);
2565 
2566 			/* Return -EFAULT bad address error as success. It will
2567 			 * fail later with a VM fault if the GPU tries to access
2568 			 * it. Better than hanging indefinitely with stalled
2569 			 * user mode queues.
2570 			 *
2571 			 * Return other error -EBUSY or -ENOMEM to retry restore
2572 			 */
2573 			if (ret != -EFAULT)
2574 				return ret;
2575 
2576 			ret = 0;
2577 		}
2578 
2579 		mutex_lock(&process_info->notifier_lock);
2580 
2581 		/* Mark the BO as valid unless it was invalidated
2582 		 * again concurrently.
2583 		 */
2584 		if (mem->invalid != invalid) {
2585 			ret = -EAGAIN;
2586 			goto unlock_out;
2587 		}
2588 		 /* set mem valid if mem has hmm range associated */
2589 		if (mem->range)
2590 			mem->invalid = 0;
2591 	}
2592 
2593 unlock_out:
2594 	mutex_unlock(&process_info->notifier_lock);
2595 
2596 	return ret;
2597 }
2598 
2599 /* Validate invalid userptr BOs
2600  *
2601  * Validates BOs on the userptr_inval_list. Also updates GPUVM page tables
2602  * with new page addresses and waits for the page table updates to complete.
2603  */
2604 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
2605 {
2606 	struct ttm_operation_ctx ctx = { false, false };
2607 	struct amdgpu_sync sync;
2608 	struct drm_exec exec;
2609 
2610 	struct amdgpu_vm *peer_vm;
2611 	struct kgd_mem *mem, *tmp_mem;
2612 	struct amdgpu_bo *bo;
2613 	int ret;
2614 
2615 	amdgpu_sync_create(&sync);
2616 
2617 	drm_exec_init(&exec, 0, 0);
2618 	/* Reserve all BOs and page tables for validation */
2619 	drm_exec_until_all_locked(&exec) {
2620 		/* Reserve all the page directories */
2621 		list_for_each_entry(peer_vm, &process_info->vm_list_head,
2622 				    vm_list_node) {
2623 			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2624 			drm_exec_retry_on_contention(&exec);
2625 			if (unlikely(ret))
2626 				goto unreserve_out;
2627 		}
2628 
2629 		/* Reserve the userptr_inval_list entries to resv_list */
2630 		list_for_each_entry(mem, &process_info->userptr_inval_list,
2631 				    validate_list) {
2632 			struct drm_gem_object *gobj;
2633 
2634 			gobj = &mem->bo->tbo.base;
2635 			ret = drm_exec_prepare_obj(&exec, gobj, 1);
2636 			drm_exec_retry_on_contention(&exec);
2637 			if (unlikely(ret))
2638 				goto unreserve_out;
2639 		}
2640 	}
2641 
2642 	ret = process_validate_vms(process_info, NULL);
2643 	if (ret)
2644 		goto unreserve_out;
2645 
2646 	/* Validate BOs and update GPUVM page tables */
2647 	list_for_each_entry_safe(mem, tmp_mem,
2648 				 &process_info->userptr_inval_list,
2649 				 validate_list) {
2650 		struct kfd_mem_attachment *attachment;
2651 
2652 		bo = mem->bo;
2653 
2654 		/* Validate the BO if we got user pages */
2655 		if (bo->tbo.ttm->pages[0]) {
2656 			amdgpu_bo_placement_from_domain(bo, mem->domain);
2657 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2658 			if (ret) {
2659 				pr_err("%s: failed to validate BO\n", __func__);
2660 				goto unreserve_out;
2661 			}
2662 		}
2663 
2664 		/* Update mapping. If the BO was not validated
2665 		 * (because we couldn't get user pages), this will
2666 		 * clear the page table entries, which will result in
2667 		 * VM faults if the GPU tries to access the invalid
2668 		 * memory.
2669 		 */
2670 		list_for_each_entry(attachment, &mem->attachments, list) {
2671 			if (!attachment->is_mapped)
2672 				continue;
2673 
2674 			kfd_mem_dmaunmap_attachment(mem, attachment);
2675 			ret = update_gpuvm_pte(mem, attachment, &sync);
2676 			if (ret) {
2677 				pr_err("%s: update PTE failed\n", __func__);
2678 				/* make sure this gets validated again */
2679 				mutex_lock(&process_info->notifier_lock);
2680 				mem->invalid++;
2681 				mutex_unlock(&process_info->notifier_lock);
2682 				goto unreserve_out;
2683 			}
2684 		}
2685 	}
2686 
2687 	/* Update page directories */
2688 	ret = process_update_pds(process_info, &sync);
2689 
2690 unreserve_out:
2691 	drm_exec_fini(&exec);
2692 	amdgpu_sync_wait(&sync, false);
2693 	amdgpu_sync_free(&sync);
2694 
2695 	return ret;
2696 }
2697 
2698 /* Confirm that all user pages are valid while holding the notifier lock
2699  *
2700  * Moves valid BOs from the userptr_inval_list back to userptr_val_list.
2701  */
2702 static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_info)
2703 {
2704 	struct kgd_mem *mem, *tmp_mem;
2705 	int ret = 0;
2706 
2707 	list_for_each_entry_safe(mem, tmp_mem,
2708 				 &process_info->userptr_inval_list,
2709 				 validate_list) {
2710 		bool valid;
2711 
2712 		/* keep mem without hmm range at userptr_inval_list */
2713 		if (!mem->range)
2714 			continue;
2715 
2716 		/* Only check mem with hmm range associated */
2717 		valid = amdgpu_ttm_tt_get_user_pages_done(
2718 					mem->bo->tbo.ttm, mem->range);
2719 
2720 		mem->range = NULL;
2721 		if (!valid) {
2722 			WARN(!mem->invalid, "Invalid BO not marked invalid");
2723 			ret = -EAGAIN;
2724 			continue;
2725 		}
2726 
2727 		if (mem->invalid) {
2728 			WARN(1, "Valid BO is marked invalid");
2729 			ret = -EAGAIN;
2730 			continue;
2731 		}
2732 
2733 		list_move_tail(&mem->validate_list,
2734 			       &process_info->userptr_valid_list);
2735 	}
2736 
2737 	return ret;
2738 }
2739 
2740 /* Worker callback to restore evicted userptr BOs
2741  *
2742  * Tries to update and validate all userptr BOs. If successful and no
2743  * concurrent evictions happened, the queues are restarted. Otherwise,
2744  * reschedule for another attempt later.
2745  */
2746 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
2747 {
2748 	struct delayed_work *dwork = to_delayed_work(work);
2749 	struct amdkfd_process_info *process_info =
2750 		container_of(dwork, struct amdkfd_process_info,
2751 			     restore_userptr_work);
2752 	struct task_struct *usertask;
2753 	struct mm_struct *mm;
2754 	uint32_t evicted_bos;
2755 
2756 	mutex_lock(&process_info->notifier_lock);
2757 	evicted_bos = process_info->evicted_bos;
2758 	mutex_unlock(&process_info->notifier_lock);
2759 	if (!evicted_bos)
2760 		return;
2761 
2762 	/* Reference task and mm in case of concurrent process termination */
2763 	usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
2764 	if (!usertask)
2765 		return;
2766 	mm = get_task_mm(usertask);
2767 	if (!mm) {
2768 		put_task_struct(usertask);
2769 		return;
2770 	}
2771 
2772 	mutex_lock(&process_info->lock);
2773 
2774 	if (update_invalid_user_pages(process_info, mm))
2775 		goto unlock_out;
2776 	/* userptr_inval_list can be empty if all evicted userptr BOs
2777 	 * have been freed. In that case there is nothing to validate
2778 	 * and we can just restart the queues.
2779 	 */
2780 	if (!list_empty(&process_info->userptr_inval_list)) {
2781 		if (validate_invalid_user_pages(process_info))
2782 			goto unlock_out;
2783 	}
2784 	/* Final check for concurrent evicton and atomic update. If
2785 	 * another eviction happens after successful update, it will
2786 	 * be a first eviction that calls quiesce_mm. The eviction
2787 	 * reference counting inside KFD will handle this case.
2788 	 */
2789 	mutex_lock(&process_info->notifier_lock);
2790 	if (process_info->evicted_bos != evicted_bos)
2791 		goto unlock_notifier_out;
2792 
2793 	if (confirm_valid_user_pages_locked(process_info)) {
2794 		WARN(1, "User pages unexpectedly invalid");
2795 		goto unlock_notifier_out;
2796 	}
2797 
2798 	process_info->evicted_bos = evicted_bos = 0;
2799 
2800 	if (kgd2kfd_resume_mm(mm)) {
2801 		pr_err("%s: Failed to resume KFD\n", __func__);
2802 		/* No recovery from this failure. Probably the CP is
2803 		 * hanging. No point trying again.
2804 		 */
2805 	}
2806 
2807 unlock_notifier_out:
2808 	mutex_unlock(&process_info->notifier_lock);
2809 unlock_out:
2810 	mutex_unlock(&process_info->lock);
2811 
2812 	/* If validation failed, reschedule another attempt */
2813 	if (evicted_bos) {
2814 		queue_delayed_work(system_freezable_wq,
2815 			&process_info->restore_userptr_work,
2816 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2817 
2818 		kfd_smi_event_queue_restore_rescheduled(mm);
2819 	}
2820 	mmput(mm);
2821 	put_task_struct(usertask);
2822 }
2823 
2824 static void replace_eviction_fence(struct dma_fence __rcu **ef,
2825 				   struct dma_fence *new_ef)
2826 {
2827 	struct dma_fence *old_ef = rcu_replace_pointer(*ef, new_ef, true
2828 		/* protected by process_info->lock */);
2829 
2830 	/* If we're replacing an unsignaled eviction fence, that fence will
2831 	 * never be signaled, and if anyone is still waiting on that fence,
2832 	 * they will hang forever. This should never happen. We should only
2833 	 * replace the fence in restore_work that only gets scheduled after
2834 	 * eviction work signaled the fence.
2835 	 */
2836 	WARN_ONCE(!dma_fence_is_signaled(old_ef),
2837 		  "Replacing unsignaled eviction fence");
2838 	dma_fence_put(old_ef);
2839 }
2840 
2841 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2842  *   KFD process identified by process_info
2843  *
2844  * @process_info: amdkfd_process_info of the KFD process
2845  *
2846  * After memory eviction, restore thread calls this function. The function
2847  * should be called when the Process is still valid. BO restore involves -
2848  *
2849  * 1.  Release old eviction fence and create new one
2850  * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2851  * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2852  *     BOs that need to be reserved.
2853  * 4.  Reserve all the BOs
2854  * 5.  Validate of PD and PT BOs.
2855  * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2856  * 7.  Add fence to all PD and PT BOs.
2857  * 8.  Unreserve all BOs
2858  */
2859 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu **ef)
2860 {
2861 	struct amdkfd_process_info *process_info = info;
2862 	struct amdgpu_vm *peer_vm;
2863 	struct kgd_mem *mem;
2864 	struct list_head duplicate_save;
2865 	struct amdgpu_sync sync_obj;
2866 	unsigned long failed_size = 0;
2867 	unsigned long total_size = 0;
2868 	struct drm_exec exec;
2869 	int ret;
2870 
2871 	INIT_LIST_HEAD(&duplicate_save);
2872 
2873 	mutex_lock(&process_info->lock);
2874 
2875 	drm_exec_init(&exec, DRM_EXEC_IGNORE_DUPLICATES, 0);
2876 	drm_exec_until_all_locked(&exec) {
2877 		list_for_each_entry(peer_vm, &process_info->vm_list_head,
2878 				    vm_list_node) {
2879 			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2880 			drm_exec_retry_on_contention(&exec);
2881 			if (unlikely(ret)) {
2882 				pr_err("Locking VM PD failed, ret: %d\n", ret);
2883 				goto ttm_reserve_fail;
2884 			}
2885 		}
2886 
2887 		/* Reserve all BOs and page tables/directory. Add all BOs from
2888 		 * kfd_bo_list to ctx.list
2889 		 */
2890 		list_for_each_entry(mem, &process_info->kfd_bo_list,
2891 				    validate_list) {
2892 			struct drm_gem_object *gobj;
2893 
2894 			gobj = &mem->bo->tbo.base;
2895 			ret = drm_exec_prepare_obj(&exec, gobj, 1);
2896 			drm_exec_retry_on_contention(&exec);
2897 			if (unlikely(ret)) {
2898 				pr_err("drm_exec_prepare_obj failed, ret: %d\n", ret);
2899 				goto ttm_reserve_fail;
2900 			}
2901 		}
2902 	}
2903 
2904 	amdgpu_sync_create(&sync_obj);
2905 
2906 	/* Validate BOs managed by KFD */
2907 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2908 			    validate_list) {
2909 
2910 		struct amdgpu_bo *bo = mem->bo;
2911 		uint32_t domain = mem->domain;
2912 		struct dma_resv_iter cursor;
2913 		struct dma_fence *fence;
2914 
2915 		total_size += amdgpu_bo_size(bo);
2916 
2917 		ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2918 		if (ret) {
2919 			pr_debug("Memory eviction: Validate BOs failed\n");
2920 			failed_size += amdgpu_bo_size(bo);
2921 			ret = amdgpu_amdkfd_bo_validate(bo,
2922 						AMDGPU_GEM_DOMAIN_GTT, false);
2923 			if (ret) {
2924 				pr_debug("Memory eviction: Try again\n");
2925 				goto validate_map_fail;
2926 			}
2927 		}
2928 		dma_resv_for_each_fence(&cursor, bo->tbo.base.resv,
2929 					DMA_RESV_USAGE_KERNEL, fence) {
2930 			ret = amdgpu_sync_fence(&sync_obj, fence);
2931 			if (ret) {
2932 				pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2933 				goto validate_map_fail;
2934 			}
2935 		}
2936 	}
2937 
2938 	if (failed_size)
2939 		pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2940 
2941 	/* Validate PDs, PTs and evicted DMABuf imports last. Otherwise BO
2942 	 * validations above would invalidate DMABuf imports again.
2943 	 */
2944 	ret = process_validate_vms(process_info, &exec.ticket);
2945 	if (ret) {
2946 		pr_debug("Validating VMs failed, ret: %d\n", ret);
2947 		goto validate_map_fail;
2948 	}
2949 
2950 	/* Update mappings managed by KFD. */
2951 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2952 			    validate_list) {
2953 		struct kfd_mem_attachment *attachment;
2954 
2955 		list_for_each_entry(attachment, &mem->attachments, list) {
2956 			if (!attachment->is_mapped)
2957 				continue;
2958 
2959 			kfd_mem_dmaunmap_attachment(mem, attachment);
2960 			ret = update_gpuvm_pte(mem, attachment, &sync_obj);
2961 			if (ret) {
2962 				pr_debug("Memory eviction: update PTE failed. Try again\n");
2963 				goto validate_map_fail;
2964 			}
2965 		}
2966 	}
2967 
2968 	/* Update mappings not managed by KFD */
2969 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2970 			vm_list_node) {
2971 		struct amdgpu_device *adev = amdgpu_ttm_adev(
2972 			peer_vm->root.bo->tbo.bdev);
2973 
2974 		ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket);
2975 		if (ret) {
2976 			pr_debug("Memory eviction: handle moved failed. Try again\n");
2977 			goto validate_map_fail;
2978 		}
2979 	}
2980 
2981 	/* Update page directories */
2982 	ret = process_update_pds(process_info, &sync_obj);
2983 	if (ret) {
2984 		pr_debug("Memory eviction: update PDs failed. Try again\n");
2985 		goto validate_map_fail;
2986 	}
2987 
2988 	/* Sync with fences on all the page tables. They implicitly depend on any
2989 	 * move fences from amdgpu_vm_handle_moved above.
2990 	 */
2991 	ret = process_sync_pds_resv(process_info, &sync_obj);
2992 	if (ret) {
2993 		pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2994 		goto validate_map_fail;
2995 	}
2996 
2997 	/* Wait for validate and PT updates to finish */
2998 	amdgpu_sync_wait(&sync_obj, false);
2999 
3000 	/* The old eviction fence may be unsignaled if restore happens
3001 	 * after a GPU reset or suspend/resume. Keep the old fence in that
3002 	 * case. Otherwise release the old eviction fence and create new
3003 	 * one, because fence only goes from unsignaled to signaled once
3004 	 * and cannot be reused. Use context and mm from the old fence.
3005 	 *
3006 	 * If an old eviction fence signals after this check, that's OK.
3007 	 * Anyone signaling an eviction fence must stop the queues first
3008 	 * and schedule another restore worker.
3009 	 */
3010 	if (dma_fence_is_signaled(&process_info->eviction_fence->base)) {
3011 		struct amdgpu_amdkfd_fence *new_fence =
3012 			amdgpu_amdkfd_fence_create(
3013 				process_info->eviction_fence->base.context,
3014 				process_info->eviction_fence->mm,
3015 				NULL);
3016 
3017 		if (!new_fence) {
3018 			pr_err("Failed to create eviction fence\n");
3019 			ret = -ENOMEM;
3020 			goto validate_map_fail;
3021 		}
3022 		dma_fence_put(&process_info->eviction_fence->base);
3023 		process_info->eviction_fence = new_fence;
3024 		replace_eviction_fence(ef, dma_fence_get(&new_fence->base));
3025 	} else {
3026 		WARN_ONCE(*ef != &process_info->eviction_fence->base,
3027 			  "KFD eviction fence doesn't match KGD process_info");
3028 	}
3029 
3030 	/* Attach new eviction fence to all BOs except pinned ones */
3031 	list_for_each_entry(mem, &process_info->kfd_bo_list, validate_list) {
3032 		if (mem->bo->tbo.pin_count)
3033 			continue;
3034 
3035 		dma_resv_add_fence(mem->bo->tbo.base.resv,
3036 				   &process_info->eviction_fence->base,
3037 				   DMA_RESV_USAGE_BOOKKEEP);
3038 	}
3039 	/* Attach eviction fence to PD / PT BOs and DMABuf imports */
3040 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
3041 			    vm_list_node) {
3042 		struct amdgpu_bo *bo = peer_vm->root.bo;
3043 
3044 		dma_resv_add_fence(bo->tbo.base.resv,
3045 				   &process_info->eviction_fence->base,
3046 				   DMA_RESV_USAGE_BOOKKEEP);
3047 	}
3048 
3049 validate_map_fail:
3050 	amdgpu_sync_free(&sync_obj);
3051 ttm_reserve_fail:
3052 	drm_exec_fini(&exec);
3053 	mutex_unlock(&process_info->lock);
3054 	return ret;
3055 }
3056 
3057 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
3058 {
3059 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
3060 	struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
3061 	int ret;
3062 
3063 	if (!info || !gws)
3064 		return -EINVAL;
3065 
3066 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
3067 	if (!*mem)
3068 		return -ENOMEM;
3069 
3070 	mutex_init(&(*mem)->lock);
3071 	INIT_LIST_HEAD(&(*mem)->attachments);
3072 	(*mem)->bo = amdgpu_bo_ref(gws_bo);
3073 	(*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
3074 	(*mem)->process_info = process_info;
3075 	add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
3076 	amdgpu_sync_create(&(*mem)->sync);
3077 
3078 
3079 	/* Validate gws bo the first time it is added to process */
3080 	mutex_lock(&(*mem)->process_info->lock);
3081 	ret = amdgpu_bo_reserve(gws_bo, false);
3082 	if (unlikely(ret)) {
3083 		pr_err("Reserve gws bo failed %d\n", ret);
3084 		goto bo_reservation_failure;
3085 	}
3086 
3087 	ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
3088 	if (ret) {
3089 		pr_err("GWS BO validate failed %d\n", ret);
3090 		goto bo_validation_failure;
3091 	}
3092 	/* GWS resource is shared b/t amdgpu and amdkfd
3093 	 * Add process eviction fence to bo so they can
3094 	 * evict each other.
3095 	 */
3096 	ret = dma_resv_reserve_fences(gws_bo->tbo.base.resv, 1);
3097 	if (ret)
3098 		goto reserve_shared_fail;
3099 	dma_resv_add_fence(gws_bo->tbo.base.resv,
3100 			   &process_info->eviction_fence->base,
3101 			   DMA_RESV_USAGE_BOOKKEEP);
3102 	amdgpu_bo_unreserve(gws_bo);
3103 	mutex_unlock(&(*mem)->process_info->lock);
3104 
3105 	return ret;
3106 
3107 reserve_shared_fail:
3108 bo_validation_failure:
3109 	amdgpu_bo_unreserve(gws_bo);
3110 bo_reservation_failure:
3111 	mutex_unlock(&(*mem)->process_info->lock);
3112 	amdgpu_sync_free(&(*mem)->sync);
3113 	remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
3114 	amdgpu_bo_unref(&gws_bo);
3115 	mutex_destroy(&(*mem)->lock);
3116 	kfree(*mem);
3117 	*mem = NULL;
3118 	return ret;
3119 }
3120 
3121 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
3122 {
3123 	int ret;
3124 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
3125 	struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
3126 	struct amdgpu_bo *gws_bo = kgd_mem->bo;
3127 
3128 	/* Remove BO from process's validate list so restore worker won't touch
3129 	 * it anymore
3130 	 */
3131 	remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
3132 
3133 	ret = amdgpu_bo_reserve(gws_bo, false);
3134 	if (unlikely(ret)) {
3135 		pr_err("Reserve gws bo failed %d\n", ret);
3136 		//TODO add BO back to validate_list?
3137 		return ret;
3138 	}
3139 	amdgpu_amdkfd_remove_eviction_fence(gws_bo,
3140 			process_info->eviction_fence);
3141 	amdgpu_bo_unreserve(gws_bo);
3142 	amdgpu_sync_free(&kgd_mem->sync);
3143 	amdgpu_bo_unref(&gws_bo);
3144 	mutex_destroy(&kgd_mem->lock);
3145 	kfree(mem);
3146 	return 0;
3147 }
3148 
3149 /* Returns GPU-specific tiling mode information */
3150 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
3151 				struct tile_config *config)
3152 {
3153 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
3154 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
3155 	config->num_tile_configs =
3156 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
3157 	config->macro_tile_config_ptr =
3158 			adev->gfx.config.macrotile_mode_array;
3159 	config->num_macro_tile_configs =
3160 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
3161 
3162 	/* Those values are not set from GFX9 onwards */
3163 	config->num_banks = adev->gfx.config.num_banks;
3164 	config->num_ranks = adev->gfx.config.num_ranks;
3165 
3166 	return 0;
3167 }
3168 
3169 bool amdgpu_amdkfd_bo_mapped_to_dev(void *drm_priv, struct kgd_mem *mem)
3170 {
3171 	struct amdgpu_vm *vm = drm_priv_to_vm(drm_priv);
3172 	struct kfd_mem_attachment *entry;
3173 
3174 	list_for_each_entry(entry, &mem->attachments, list) {
3175 		if (entry->is_mapped && entry->bo_va->base.vm == vm)
3176 			return true;
3177 	}
3178 	return false;
3179 }
3180 
3181 #if defined(CONFIG_DEBUG_FS)
3182 
3183 int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data)
3184 {
3185 
3186 	spin_lock(&kfd_mem_limit.mem_limit_lock);
3187 	seq_printf(m, "System mem used %lldM out of %lluM\n",
3188 		  (kfd_mem_limit.system_mem_used >> 20),
3189 		  (kfd_mem_limit.max_system_mem_limit >> 20));
3190 	seq_printf(m, "TTM mem used %lldM out of %lluM\n",
3191 		  (kfd_mem_limit.ttm_mem_used >> 20),
3192 		  (kfd_mem_limit.max_ttm_mem_limit >> 20));
3193 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
3194 
3195 	return 0;
3196 }
3197 
3198 #endif
3199