xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c (revision 69050f8d6d075dc01af7a5f2f550a8067510366f)
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 		if (!no_system_mem_limit) {
219 			ret = -ENOMEM;
220 			goto release;
221 		}
222 	}
223 
224 	if (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
225 		kfd_mem_limit.max_ttm_mem_limit) {
226 		ret = -ENOMEM;
227 		goto release;
228 	}
229 
230 	/*if is_app_apu is false and apu_prefer_gtt is true, it is an APU with
231 	 * carve out < gtt. In that case, VRAM allocation will go to gtt domain, skip
232 	 * VRAM check since ttm_mem_limit check already cover this allocation
233 	 */
234 
235 	if (adev && xcp_id >= 0 && (!adev->apu_prefer_gtt || adev->gmc.is_app_apu)) {
236 		uint64_t vram_available =
237 			vram_size - reserved_for_pt - reserved_for_ras -
238 			atomic64_read(&adev->vram_pin_size);
239 		if (adev->kfd.vram_used[xcp_id] + vram_needed > vram_available) {
240 			ret = -ENOMEM;
241 			goto release;
242 		}
243 	}
244 
245 	/* Update memory accounting by decreasing available system
246 	 * memory, TTM memory and GPU memory as computed above
247 	 */
248 	WARN_ONCE(vram_needed && !adev,
249 		  "adev reference can't be null when vram is used");
250 	if (adev && xcp_id >= 0) {
251 		adev->kfd.vram_used[xcp_id] += vram_needed;
252 		adev->kfd.vram_used_aligned[xcp_id] +=
253 				adev->apu_prefer_gtt ?
254 				vram_needed :
255 				ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
256 	}
257 	kfd_mem_limit.system_mem_used += system_mem_needed;
258 	kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
259 
260 release:
261 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
262 	return ret;
263 }
264 
265 void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
266 		uint64_t size, u32 alloc_flag, int8_t xcp_id)
267 {
268 	spin_lock(&kfd_mem_limit.mem_limit_lock);
269 
270 	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
271 		kfd_mem_limit.system_mem_used -= size;
272 		kfd_mem_limit.ttm_mem_used -= size;
273 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
274 		WARN_ONCE(!adev,
275 			  "adev reference can't be null when alloc mem flags vram is set");
276 		if (WARN_ONCE(xcp_id < 0, "invalid XCP ID %d", xcp_id))
277 			goto release;
278 
279 		if (adev) {
280 			adev->kfd.vram_used[xcp_id] -= size;
281 			if (adev->apu_prefer_gtt) {
282 				adev->kfd.vram_used_aligned[xcp_id] -= size;
283 				kfd_mem_limit.system_mem_used -= size;
284 				kfd_mem_limit.ttm_mem_used -= size;
285 			} else {
286 				adev->kfd.vram_used_aligned[xcp_id] -=
287 					ALIGN(size, VRAM_AVAILABLITY_ALIGN);
288 			}
289 		}
290 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
291 		kfd_mem_limit.system_mem_used -= size;
292 	} else if (!(alloc_flag &
293 				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
294 				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
295 		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
296 		goto release;
297 	}
298 	WARN_ONCE(adev && xcp_id >= 0 && adev->kfd.vram_used[xcp_id] < 0,
299 		  "KFD VRAM memory accounting unbalanced for xcp: %d", xcp_id);
300 	WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
301 		  "KFD TTM memory accounting unbalanced");
302 	WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
303 		  "KFD system memory accounting unbalanced");
304 
305 release:
306 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
307 }
308 
309 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
310 {
311 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
312 	u32 alloc_flags = bo->kfd_bo->alloc_flags;
313 	u64 size = amdgpu_bo_size(bo);
314 
315 	amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags,
316 					  bo->xcp_id);
317 
318 	kfree(bo->kfd_bo);
319 }
320 
321 /**
322  * create_dmamap_sg_bo() - Creates a amdgpu_bo object to reflect information
323  * about USERPTR or DOOREBELL or MMIO BO.
324  *
325  * @adev: Device for which dmamap BO is being created
326  * @mem: BO of peer device that is being DMA mapped. Provides parameters
327  *	 in building the dmamap BO
328  * @bo_out: Output parameter updated with handle of dmamap BO
329  */
330 static int
331 create_dmamap_sg_bo(struct amdgpu_device *adev,
332 		 struct kgd_mem *mem, struct amdgpu_bo **bo_out)
333 {
334 	struct drm_gem_object *gem_obj;
335 	int ret;
336 	uint64_t flags = 0;
337 
338 	ret = amdgpu_bo_reserve(mem->bo, false);
339 	if (ret)
340 		return ret;
341 
342 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)
343 		flags |= mem->bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
344 					AMDGPU_GEM_CREATE_UNCACHED);
345 
346 	ret = amdgpu_gem_object_create(adev, mem->bo->tbo.base.size, 1,
347 			AMDGPU_GEM_DOMAIN_CPU, AMDGPU_GEM_CREATE_PREEMPTIBLE | flags,
348 			ttm_bo_type_sg, mem->bo->tbo.base.resv, &gem_obj, 0);
349 
350 	amdgpu_bo_unreserve(mem->bo);
351 
352 	if (ret) {
353 		pr_err("Error in creating DMA mappable SG BO on domain: %d\n", ret);
354 		return -EINVAL;
355 	}
356 
357 	*bo_out = gem_to_amdgpu_bo(gem_obj);
358 	(*bo_out)->parent = amdgpu_bo_ref(mem->bo);
359 	return ret;
360 }
361 
362 /* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
363  *  reservation object.
364  *
365  * @bo: [IN] Remove eviction fence(s) from this BO
366  * @ef: [IN] This eviction fence is removed if it
367  *  is present in the shared list.
368  *
369  * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
370  */
371 static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
372 					struct amdgpu_amdkfd_fence *ef)
373 {
374 	struct dma_fence *replacement;
375 
376 	if (!ef)
377 		return -EINVAL;
378 
379 	/* TODO: Instead of block before we should use the fence of the page
380 	 * table update and TLB flush here directly.
381 	 */
382 	replacement = dma_fence_get_stub();
383 	dma_resv_replace_fences(bo->tbo.base.resv, ef->base.context,
384 				replacement, DMA_RESV_USAGE_BOOKKEEP);
385 	dma_fence_put(replacement);
386 	return 0;
387 }
388 
389 /**
390  * amdgpu_amdkfd_remove_all_eviction_fences - Remove all eviction fences
391  * @bo: the BO where to remove the evictions fences from.
392  *
393  * This functions should only be used on release when all references to the BO
394  * are already dropped. We remove the eviction fence from the private copy of
395  * the dma_resv object here since that is what is used during release to
396  * determine of the BO is idle or not.
397  */
398 void amdgpu_amdkfd_remove_all_eviction_fences(struct amdgpu_bo *bo)
399 {
400 	struct dma_resv *resv = &bo->tbo.base._resv;
401 	struct dma_fence *fence, *stub;
402 	struct dma_resv_iter cursor;
403 
404 	dma_resv_assert_held(resv);
405 
406 	stub = dma_fence_get_stub();
407 	dma_resv_for_each_fence(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP, fence) {
408 		if (!to_amdgpu_amdkfd_fence(fence))
409 			continue;
410 
411 		dma_resv_replace_fences(resv, fence->context, stub,
412 					DMA_RESV_USAGE_BOOKKEEP);
413 	}
414 	dma_fence_put(stub);
415 }
416 
417 static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
418 				     bool wait)
419 {
420 	struct ttm_operation_ctx ctx = { false, false };
421 	int ret;
422 
423 	if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
424 		 "Called with userptr BO"))
425 		return -EINVAL;
426 
427 	/* bo has been pinned, not need validate it */
428 	if (bo->tbo.pin_count)
429 		return 0;
430 
431 	amdgpu_bo_placement_from_domain(bo, domain);
432 
433 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
434 	if (ret)
435 		goto validate_fail;
436 	if (wait)
437 		amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
438 
439 validate_fail:
440 	return ret;
441 }
442 
443 int amdgpu_amdkfd_bo_validate_and_fence(struct amdgpu_bo *bo,
444 					uint32_t domain,
445 					struct dma_fence *fence)
446 {
447 	int ret = amdgpu_bo_reserve(bo, false);
448 
449 	if (ret)
450 		return ret;
451 
452 	ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
453 	if (ret)
454 		goto unreserve_out;
455 
456 	ret = dma_resv_reserve_fences(bo->tbo.base.resv, 1);
457 	if (ret)
458 		goto unreserve_out;
459 
460 	dma_resv_add_fence(bo->tbo.base.resv, fence,
461 			   DMA_RESV_USAGE_BOOKKEEP);
462 
463 unreserve_out:
464 	amdgpu_bo_unreserve(bo);
465 
466 	return ret;
467 }
468 
469 static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo)
470 {
471 	return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false);
472 }
473 
474 /* vm_validate_pt_pd_bos - Validate page table and directory BOs
475  *
476  * Page directories are not updated here because huge page handling
477  * during page table updates can invalidate page directory entries
478  * again. Page directories are only updated after updating page
479  * tables.
480  */
481 static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm,
482 				 struct ww_acquire_ctx *ticket)
483 {
484 	struct amdgpu_bo *pd = vm->root.bo;
485 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
486 	int ret;
487 
488 	ret = amdgpu_vm_validate(adev, vm, ticket,
489 				 amdgpu_amdkfd_validate_vm_bo, NULL);
490 	if (ret) {
491 		pr_err("failed to validate PT BOs\n");
492 		return ret;
493 	}
494 
495 	vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo);
496 
497 	return 0;
498 }
499 
500 static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
501 {
502 	struct amdgpu_bo *pd = vm->root.bo;
503 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
504 	int ret;
505 
506 	ret = amdgpu_vm_update_pdes(adev, vm, false);
507 	if (ret)
508 		return ret;
509 
510 	return amdgpu_sync_fence(sync, vm->last_update, GFP_KERNEL);
511 }
512 
513 static uint64_t get_pte_flags(struct amdgpu_device *adev, struct amdgpu_vm *vm,
514 			      struct kgd_mem *mem)
515 {
516 	uint32_t mapping_flags = AMDGPU_VM_PAGE_READABLE |
517 				 AMDGPU_VM_MTYPE_DEFAULT;
518 
519 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE)
520 		mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
521 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE)
522 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
523 
524 	return mapping_flags;
525 }
526 
527 /**
528  * create_sg_table() - Create an sg_table for a contiguous DMA addr range
529  * @addr: The starting address to point to
530  * @size: Size of memory area in bytes being pointed to
531  *
532  * Allocates an instance of sg_table and initializes it to point to memory
533  * area specified by input parameters. The address used to build is assumed
534  * to be DMA mapped, if needed.
535  *
536  * DOORBELL or MMIO BOs use only one scatterlist node in their sg_table
537  * because they are physically contiguous.
538  *
539  * Return: Initialized instance of SG Table or NULL
540  */
541 static struct sg_table *create_sg_table(uint64_t addr, uint32_t size)
542 {
543 	struct sg_table *sg = kmalloc_obj(*sg, GFP_KERNEL);
544 
545 	if (!sg)
546 		return NULL;
547 	if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
548 		kfree(sg);
549 		return NULL;
550 	}
551 	sg_dma_address(sg->sgl) = addr;
552 	sg->sgl->length = size;
553 #ifdef CONFIG_NEED_SG_DMA_LENGTH
554 	sg->sgl->dma_length = size;
555 #endif
556 	return sg;
557 }
558 
559 static int
560 kfd_mem_dmamap_userptr(struct kgd_mem *mem,
561 		       struct kfd_mem_attachment *attachment)
562 {
563 	enum dma_data_direction direction =
564 		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
565 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
566 	struct ttm_operation_ctx ctx = {.interruptible = true};
567 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
568 	struct amdgpu_device *adev = attachment->adev;
569 	struct ttm_tt *src_ttm = mem->bo->tbo.ttm;
570 	struct ttm_tt *ttm = bo->tbo.ttm;
571 	int ret;
572 
573 	if (WARN_ON(ttm->num_pages != src_ttm->num_pages))
574 		return -EINVAL;
575 
576 	ttm->sg = kmalloc_obj(*ttm->sg, GFP_KERNEL);
577 	if (unlikely(!ttm->sg))
578 		return -ENOMEM;
579 
580 	/* Same sequence as in amdgpu_ttm_tt_pin_userptr */
581 	ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages,
582 					ttm->num_pages, 0,
583 					(u64)ttm->num_pages << PAGE_SHIFT,
584 					GFP_KERNEL);
585 	if (unlikely(ret))
586 		goto free_sg;
587 
588 	ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
589 	if (unlikely(ret))
590 		goto release_sg;
591 
592 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
593 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
594 	if (ret)
595 		goto unmap_sg;
596 
597 	return 0;
598 
599 unmap_sg:
600 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
601 release_sg:
602 	pr_err("DMA map userptr failed: %d\n", ret);
603 	sg_free_table(ttm->sg);
604 free_sg:
605 	kfree(ttm->sg);
606 	ttm->sg = NULL;
607 	return ret;
608 }
609 
610 static int
611 kfd_mem_dmamap_dmabuf(struct kfd_mem_attachment *attachment)
612 {
613 	struct ttm_operation_ctx ctx = {.interruptible = true};
614 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
615 
616 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
617 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
618 }
619 
620 /**
621  * kfd_mem_dmamap_sg_bo() - Create DMA mapped sg_table to access DOORBELL or MMIO BO
622  * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
623  * @attachment: Virtual address attachment of the BO on accessing device
624  *
625  * An access request from the device that owns DOORBELL does not require DMA mapping.
626  * This is because the request doesn't go through PCIe root complex i.e. it instead
627  * loops back. The need to DMA map arises only when accessing peer device's DOORBELL
628  *
629  * In contrast, all access requests for MMIO need to be DMA mapped without regard to
630  * device ownership. This is because access requests for MMIO go through PCIe root
631  * complex.
632  *
633  * This is accomplished in two steps:
634  *   - Obtain DMA mapped address of DOORBELL or MMIO memory that could be used
635  *         in updating requesting device's page table
636  *   - Signal TTM to mark memory pointed to by requesting device's BO as GPU
637  *         accessible. This allows an update of requesting device's page table
638  *         with entries associated with DOOREBELL or MMIO memory
639  *
640  * This method is invoked in the following contexts:
641  *   - Mapping of DOORBELL or MMIO BO of same or peer device
642  *   - Validating an evicted DOOREBELL or MMIO BO on device seeking access
643  *
644  * Return: ZERO if successful, NON-ZERO otherwise
645  */
646 static int
647 kfd_mem_dmamap_sg_bo(struct kgd_mem *mem,
648 		     struct kfd_mem_attachment *attachment)
649 {
650 	struct ttm_operation_ctx ctx = {.interruptible = true};
651 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
652 	struct amdgpu_device *adev = attachment->adev;
653 	struct ttm_tt *ttm = bo->tbo.ttm;
654 	enum dma_data_direction dir;
655 	dma_addr_t dma_addr;
656 	bool mmio;
657 	int ret;
658 
659 	/* Expect SG Table of dmapmap BO to be NULL */
660 	mmio = (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP);
661 	if (unlikely(ttm->sg)) {
662 		pr_err("SG Table of %d BO for peer device is UNEXPECTEDLY NON-NULL", mmio);
663 		return -EINVAL;
664 	}
665 
666 	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
667 			DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
668 	dma_addr = mem->bo->tbo.sg->sgl->dma_address;
669 	pr_debug("%d BO size: %d\n", mmio, mem->bo->tbo.sg->sgl->length);
670 	pr_debug("%d BO address before DMA mapping: %llx\n", mmio, dma_addr);
671 	dma_addr = dma_map_resource(adev->dev, dma_addr,
672 			mem->bo->tbo.sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
673 	ret = dma_mapping_error(adev->dev, dma_addr);
674 	if (unlikely(ret))
675 		return ret;
676 	pr_debug("%d BO address after DMA mapping: %llx\n", mmio, dma_addr);
677 
678 	ttm->sg = create_sg_table(dma_addr, mem->bo->tbo.sg->sgl->length);
679 	if (unlikely(!ttm->sg)) {
680 		ret = -ENOMEM;
681 		goto unmap_sg;
682 	}
683 
684 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
685 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
686 	if (unlikely(ret))
687 		goto free_sg;
688 
689 	return ret;
690 
691 free_sg:
692 	sg_free_table(ttm->sg);
693 	kfree(ttm->sg);
694 	ttm->sg = NULL;
695 unmap_sg:
696 	dma_unmap_resource(adev->dev, dma_addr, mem->bo->tbo.sg->sgl->length,
697 			   dir, DMA_ATTR_SKIP_CPU_SYNC);
698 	return ret;
699 }
700 
701 static int
702 kfd_mem_dmamap_attachment(struct kgd_mem *mem,
703 			  struct kfd_mem_attachment *attachment)
704 {
705 	switch (attachment->type) {
706 	case KFD_MEM_ATT_SHARED:
707 		return 0;
708 	case KFD_MEM_ATT_USERPTR:
709 		return kfd_mem_dmamap_userptr(mem, attachment);
710 	case KFD_MEM_ATT_DMABUF:
711 		return kfd_mem_dmamap_dmabuf(attachment);
712 	case KFD_MEM_ATT_SG:
713 		return kfd_mem_dmamap_sg_bo(mem, attachment);
714 	default:
715 		WARN_ON_ONCE(1);
716 	}
717 	return -EINVAL;
718 }
719 
720 static void
721 kfd_mem_dmaunmap_userptr(struct kgd_mem *mem,
722 			 struct kfd_mem_attachment *attachment)
723 {
724 	enum dma_data_direction direction =
725 		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
726 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
727 	struct ttm_operation_ctx ctx = {.interruptible = false};
728 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
729 	struct amdgpu_device *adev = attachment->adev;
730 	struct ttm_tt *ttm = bo->tbo.ttm;
731 
732 	if (unlikely(!ttm->sg))
733 		return;
734 
735 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
736 	(void)ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
737 
738 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
739 	sg_free_table(ttm->sg);
740 	kfree(ttm->sg);
741 	ttm->sg = NULL;
742 }
743 
744 static void
745 kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment)
746 {
747 	/* This is a no-op. We don't want to trigger eviction fences when
748 	 * unmapping DMABufs. Therefore the invalidation (moving to system
749 	 * domain) is done in kfd_mem_dmamap_dmabuf.
750 	 */
751 }
752 
753 /**
754  * kfd_mem_dmaunmap_sg_bo() - Free DMA mapped sg_table of DOORBELL or MMIO BO
755  * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
756  * @attachment: Virtual address attachment of the BO on accessing device
757  *
758  * The method performs following steps:
759  *   - Signal TTM to mark memory pointed to by BO as GPU inaccessible
760  *   - Free SG Table that is used to encapsulate DMA mapped memory of
761  *          peer device's DOORBELL or MMIO memory
762  *
763  * This method is invoked in the following contexts:
764  *     UNMapping of DOORBELL or MMIO BO on a device having access to its memory
765  *     Eviction of DOOREBELL or MMIO BO on device having access to its memory
766  *
767  * Return: void
768  */
769 static void
770 kfd_mem_dmaunmap_sg_bo(struct kgd_mem *mem,
771 		       struct kfd_mem_attachment *attachment)
772 {
773 	struct ttm_operation_ctx ctx = {.interruptible = true};
774 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
775 	struct amdgpu_device *adev = attachment->adev;
776 	struct ttm_tt *ttm = bo->tbo.ttm;
777 	enum dma_data_direction dir;
778 
779 	if (unlikely(!ttm->sg)) {
780 		pr_debug("SG Table of BO is NULL");
781 		return;
782 	}
783 
784 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
785 	(void)ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
786 
787 	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
788 				DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
789 	dma_unmap_resource(adev->dev, ttm->sg->sgl->dma_address,
790 			ttm->sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
791 	sg_free_table(ttm->sg);
792 	kfree(ttm->sg);
793 	ttm->sg = NULL;
794 	bo->tbo.sg = NULL;
795 }
796 
797 static void
798 kfd_mem_dmaunmap_attachment(struct kgd_mem *mem,
799 			    struct kfd_mem_attachment *attachment)
800 {
801 	switch (attachment->type) {
802 	case KFD_MEM_ATT_SHARED:
803 		break;
804 	case KFD_MEM_ATT_USERPTR:
805 		kfd_mem_dmaunmap_userptr(mem, attachment);
806 		break;
807 	case KFD_MEM_ATT_DMABUF:
808 		kfd_mem_dmaunmap_dmabuf(attachment);
809 		break;
810 	case KFD_MEM_ATT_SG:
811 		kfd_mem_dmaunmap_sg_bo(mem, attachment);
812 		break;
813 	default:
814 		WARN_ON_ONCE(1);
815 	}
816 }
817 
818 static int kfd_mem_export_dmabuf(struct kgd_mem *mem)
819 {
820 	if (!mem->dmabuf) {
821 		struct amdgpu_device *bo_adev;
822 		struct dma_buf *dmabuf;
823 
824 		bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
825 		dmabuf = drm_gem_prime_handle_to_dmabuf(&bo_adev->ddev, bo_adev->kfd.client.file,
826 					       mem->gem_handle,
827 			mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
828 					       DRM_RDWR : 0);
829 		if (IS_ERR(dmabuf))
830 			return PTR_ERR(dmabuf);
831 		mem->dmabuf = dmabuf;
832 	}
833 
834 	return 0;
835 }
836 
837 static int
838 kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem,
839 		      struct amdgpu_bo **bo)
840 {
841 	struct drm_gem_object *gobj;
842 	int ret;
843 
844 	ret = kfd_mem_export_dmabuf(mem);
845 	if (ret)
846 		return ret;
847 
848 	gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf);
849 	if (IS_ERR(gobj))
850 		return PTR_ERR(gobj);
851 
852 	*bo = gem_to_amdgpu_bo(gobj);
853 	(*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE;
854 
855 	return 0;
856 }
857 
858 /* kfd_mem_attach - Add a BO to a VM
859  *
860  * Everything that needs to bo done only once when a BO is first added
861  * to a VM. It can later be mapped and unmapped many times without
862  * repeating these steps.
863  *
864  * 0. Create BO for DMA mapping, if needed
865  * 1. Allocate and initialize BO VA entry data structure
866  * 2. Add BO to the VM
867  * 3. Determine ASIC-specific PTE flags
868  * 4. Alloc page tables and directories if needed
869  * 4a.  Validate new page tables and directories
870  */
871 static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
872 		struct amdgpu_vm *vm, bool is_aql)
873 {
874 	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
875 	unsigned long bo_size = mem->bo->tbo.base.size;
876 	uint64_t va = mem->va;
877 	struct kfd_mem_attachment *attachment[2] = {NULL, NULL};
878 	struct amdgpu_bo *bo[2] = {NULL, NULL};
879 	struct amdgpu_bo_va *bo_va;
880 	bool same_hive = false;
881 	struct drm_exec exec;
882 	int i, ret;
883 
884 	if (!va) {
885 		pr_err("Invalid VA when adding BO to VM\n");
886 		return -EINVAL;
887 	}
888 
889 	/* Determine access to VRAM, MMIO and DOORBELL BOs of peer devices
890 	 *
891 	 * The access path of MMIO and DOORBELL BOs of is always over PCIe.
892 	 * In contrast the access path of VRAM BOs depens upon the type of
893 	 * link that connects the peer device. Access over PCIe is allowed
894 	 * if peer device has large BAR. In contrast, access over xGMI is
895 	 * allowed for both small and large BAR configurations of peer device
896 	 */
897 	if ((adev != bo_adev && !adev->apu_prefer_gtt) &&
898 	    ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) ||
899 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) ||
900 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
901 		if (mem->domain == AMDGPU_GEM_DOMAIN_VRAM)
902 			same_hive = amdgpu_xgmi_same_hive(adev, bo_adev);
903 		if (!same_hive && !amdgpu_device_is_peer_accessible(bo_adev, adev))
904 			return -EINVAL;
905 	}
906 
907 	for (i = 0; i <= is_aql; i++) {
908 		attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL);
909 		if (unlikely(!attachment[i])) {
910 			ret = -ENOMEM;
911 			goto unwind;
912 		}
913 
914 		pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
915 			 va + bo_size, vm);
916 
917 		if ((adev == bo_adev && !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) ||
918 		    (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && reuse_dmamap(adev, bo_adev)) ||
919 		    (mem->domain == AMDGPU_GEM_DOMAIN_GTT && reuse_dmamap(adev, bo_adev)) ||
920 		    same_hive) {
921 			/* Mappings on the local GPU, or VRAM mappings in the
922 			 * local hive, or userptr, or GTT mapping can reuse dma map
923 			 * address space share the original BO
924 			 */
925 			attachment[i]->type = KFD_MEM_ATT_SHARED;
926 			bo[i] = mem->bo;
927 			drm_gem_object_get(&bo[i]->tbo.base);
928 		} else if (i > 0) {
929 			/* Multiple mappings on the same GPU share the BO */
930 			attachment[i]->type = KFD_MEM_ATT_SHARED;
931 			bo[i] = bo[0];
932 			drm_gem_object_get(&bo[i]->tbo.base);
933 		} else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
934 			/* Create an SG BO to DMA-map userptrs on other GPUs */
935 			attachment[i]->type = KFD_MEM_ATT_USERPTR;
936 			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
937 			if (ret)
938 				goto unwind;
939 		/* Handle DOORBELL BOs of peer devices and MMIO BOs of local and peer devices */
940 		} else if (mem->bo->tbo.type == ttm_bo_type_sg) {
941 			WARN_ONCE(!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL ||
942 				    mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP),
943 				  "Handing invalid SG BO in ATTACH request");
944 			attachment[i]->type = KFD_MEM_ATT_SG;
945 			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
946 			if (ret)
947 				goto unwind;
948 		/* Enable acces to GTT and VRAM BOs of peer devices */
949 		} else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT ||
950 			   mem->domain == AMDGPU_GEM_DOMAIN_VRAM) {
951 			attachment[i]->type = KFD_MEM_ATT_DMABUF;
952 			ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]);
953 			if (ret)
954 				goto unwind;
955 			pr_debug("Employ DMABUF mechanism to enable peer GPU access\n");
956 		} else {
957 			WARN_ONCE(true, "Handling invalid ATTACH request");
958 			ret = -EINVAL;
959 			goto unwind;
960 		}
961 
962 		drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
963 		drm_exec_until_all_locked(&exec) {
964 			ret = amdgpu_vm_lock_pd(vm, &exec, 0);
965 			drm_exec_retry_on_contention(&exec);
966 			if (unlikely(ret))
967 				goto unwind;
968 			ret = drm_exec_lock_obj(&exec, &bo[i]->tbo.base);
969 			drm_exec_retry_on_contention(&exec);
970 			if (unlikely(ret))
971 				goto unwind;
972 		}
973 
974 		bo_va = amdgpu_vm_bo_find(vm, bo[i]);
975 		if (!bo_va)
976 			bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]);
977 		else
978 			++bo_va->ref_count;
979 		attachment[i]->bo_va = bo_va;
980 		drm_exec_fini(&exec);
981 		if (unlikely(!attachment[i]->bo_va)) {
982 			ret = -ENOMEM;
983 			pr_err("Failed to add BO object to VM. ret == %d\n",
984 			       ret);
985 			goto unwind;
986 		}
987 		attachment[i]->va = va;
988 		attachment[i]->pte_flags = get_pte_flags(adev, vm, mem);
989 		attachment[i]->adev = adev;
990 		list_add(&attachment[i]->list, &mem->attachments);
991 
992 		va += bo_size;
993 	}
994 
995 	return 0;
996 
997 unwind:
998 	for (; i >= 0; i--) {
999 		if (!attachment[i])
1000 			continue;
1001 		if (attachment[i]->bo_va) {
1002 			(void)amdgpu_bo_reserve(bo[i], true);
1003 			if (--attachment[i]->bo_va->ref_count == 0)
1004 				amdgpu_vm_bo_del(adev, attachment[i]->bo_va);
1005 			amdgpu_bo_unreserve(bo[i]);
1006 			list_del(&attachment[i]->list);
1007 		}
1008 		if (bo[i])
1009 			drm_gem_object_put(&bo[i]->tbo.base);
1010 		kfree(attachment[i]);
1011 	}
1012 	return ret;
1013 }
1014 
1015 static void kfd_mem_detach(struct kfd_mem_attachment *attachment)
1016 {
1017 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
1018 
1019 	pr_debug("\t remove VA 0x%llx in entry %p\n",
1020 			attachment->va, attachment);
1021 	if (--attachment->bo_va->ref_count == 0)
1022 		amdgpu_vm_bo_del(attachment->adev, attachment->bo_va);
1023 	drm_gem_object_put(&bo->tbo.base);
1024 	list_del(&attachment->list);
1025 	kfree(attachment);
1026 }
1027 
1028 static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
1029 				struct amdkfd_process_info *process_info,
1030 				bool userptr)
1031 {
1032 	mutex_lock(&process_info->lock);
1033 	if (userptr)
1034 		list_add_tail(&mem->validate_list,
1035 			      &process_info->userptr_valid_list);
1036 	else
1037 		list_add_tail(&mem->validate_list, &process_info->kfd_bo_list);
1038 	mutex_unlock(&process_info->lock);
1039 }
1040 
1041 static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
1042 		struct amdkfd_process_info *process_info)
1043 {
1044 	mutex_lock(&process_info->lock);
1045 	list_del(&mem->validate_list);
1046 	mutex_unlock(&process_info->lock);
1047 }
1048 
1049 /* Initializes user pages. It registers the MMU notifier and validates
1050  * the userptr BO in the GTT domain.
1051  *
1052  * The BO must already be on the userptr_valid_list. Otherwise an
1053  * eviction and restore may happen that leaves the new BO unmapped
1054  * with the user mode queues running.
1055  *
1056  * Takes the process_info->lock to protect against concurrent restore
1057  * workers.
1058  *
1059  * Returns 0 for success, negative errno for errors.
1060  */
1061 static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr,
1062 			   bool criu_resume)
1063 {
1064 	struct amdkfd_process_info *process_info = mem->process_info;
1065 	struct amdgpu_bo *bo = mem->bo;
1066 	struct ttm_operation_ctx ctx = { true, false };
1067 	struct amdgpu_hmm_range *range;
1068 	int ret = 0;
1069 
1070 	mutex_lock(&process_info->lock);
1071 
1072 	ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
1073 	if (ret) {
1074 		pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
1075 		goto out;
1076 	}
1077 
1078 	ret = amdgpu_hmm_register(bo, user_addr);
1079 	if (ret) {
1080 		pr_err("%s: Failed to register MMU notifier: %d\n",
1081 		       __func__, ret);
1082 		goto out;
1083 	}
1084 
1085 	if (criu_resume) {
1086 		/*
1087 		 * During a CRIU restore operation, the userptr buffer objects
1088 		 * will be validated in the restore_userptr_work worker at a
1089 		 * later stage when it is scheduled by another ioctl called by
1090 		 * CRIU master process for the target pid for restore.
1091 		 */
1092 		mutex_lock(&process_info->notifier_lock);
1093 		mem->invalid++;
1094 		mutex_unlock(&process_info->notifier_lock);
1095 		mutex_unlock(&process_info->lock);
1096 		return 0;
1097 	}
1098 
1099 	range = amdgpu_hmm_range_alloc(NULL);
1100 	if (unlikely(!range)) {
1101 		ret = -ENOMEM;
1102 		goto unregister_out;
1103 	}
1104 
1105 	ret = amdgpu_ttm_tt_get_user_pages(bo, range);
1106 	if (ret) {
1107 		amdgpu_hmm_range_free(range);
1108 		if (ret == -EAGAIN)
1109 			pr_debug("Failed to get user pages, try again\n");
1110 		else
1111 			pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
1112 		goto unregister_out;
1113 	}
1114 
1115 	ret = amdgpu_bo_reserve(bo, true);
1116 	if (ret) {
1117 		pr_err("%s: Failed to reserve BO\n", __func__);
1118 		goto release_out;
1119 	}
1120 
1121 	amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm, range);
1122 
1123 	amdgpu_bo_placement_from_domain(bo, mem->domain);
1124 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1125 	if (ret)
1126 		pr_err("%s: failed to validate BO\n", __func__);
1127 	amdgpu_bo_unreserve(bo);
1128 
1129 release_out:
1130 	amdgpu_hmm_range_free(range);
1131 unregister_out:
1132 	if (ret)
1133 		amdgpu_hmm_unregister(bo);
1134 out:
1135 	mutex_unlock(&process_info->lock);
1136 	return ret;
1137 }
1138 
1139 /* Reserving a BO and its page table BOs must happen atomically to
1140  * avoid deadlocks. Some operations update multiple VMs at once. Track
1141  * all the reservation info in a context structure. Optionally a sync
1142  * object can track VM updates.
1143  */
1144 struct bo_vm_reservation_context {
1145 	/* DRM execution context for the reservation */
1146 	struct drm_exec exec;
1147 	/* Number of VMs reserved */
1148 	unsigned int n_vms;
1149 	/* Pointer to sync object */
1150 	struct amdgpu_sync *sync;
1151 };
1152 
1153 enum bo_vm_match {
1154 	BO_VM_NOT_MAPPED = 0,	/* Match VMs where a BO is not mapped */
1155 	BO_VM_MAPPED,		/* Match VMs where a BO is mapped     */
1156 	BO_VM_ALL,		/* Match all VMs a BO was added to    */
1157 };
1158 
1159 /**
1160  * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
1161  * @mem: KFD BO structure.
1162  * @vm: the VM to reserve.
1163  * @ctx: the struct that will be used in unreserve_bo_and_vms().
1164  */
1165 static int reserve_bo_and_vm(struct kgd_mem *mem,
1166 			      struct amdgpu_vm *vm,
1167 			      struct bo_vm_reservation_context *ctx)
1168 {
1169 	struct amdgpu_bo *bo = mem->bo;
1170 	int ret;
1171 
1172 	WARN_ON(!vm);
1173 
1174 	ctx->n_vms = 1;
1175 	ctx->sync = &mem->sync;
1176 	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
1177 	drm_exec_until_all_locked(&ctx->exec) {
1178 		ret = amdgpu_vm_lock_pd(vm, &ctx->exec, 2);
1179 		drm_exec_retry_on_contention(&ctx->exec);
1180 		if (unlikely(ret))
1181 			goto error;
1182 
1183 		ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
1184 		drm_exec_retry_on_contention(&ctx->exec);
1185 		if (unlikely(ret))
1186 			goto error;
1187 	}
1188 	return 0;
1189 
1190 error:
1191 	pr_err("Failed to reserve buffers in ttm.\n");
1192 	drm_exec_fini(&ctx->exec);
1193 	return ret;
1194 }
1195 
1196 /**
1197  * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
1198  * @mem: KFD BO structure.
1199  * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
1200  * is used. Otherwise, a single VM associated with the BO.
1201  * @map_type: the mapping status that will be used to filter the VMs.
1202  * @ctx: the struct that will be used in unreserve_bo_and_vms().
1203  *
1204  * Returns 0 for success, negative for failure.
1205  */
1206 static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
1207 				struct amdgpu_vm *vm, enum bo_vm_match map_type,
1208 				struct bo_vm_reservation_context *ctx)
1209 {
1210 	struct kfd_mem_attachment *entry;
1211 	struct amdgpu_bo *bo = mem->bo;
1212 	int ret;
1213 
1214 	ctx->sync = &mem->sync;
1215 	drm_exec_init(&ctx->exec, DRM_EXEC_INTERRUPTIBLE_WAIT |
1216 		      DRM_EXEC_IGNORE_DUPLICATES, 0);
1217 	drm_exec_until_all_locked(&ctx->exec) {
1218 		ctx->n_vms = 0;
1219 		list_for_each_entry(entry, &mem->attachments, list) {
1220 			if ((vm && vm != entry->bo_va->base.vm) ||
1221 				(entry->is_mapped != map_type
1222 				&& map_type != BO_VM_ALL))
1223 				continue;
1224 
1225 			ret = amdgpu_vm_lock_pd(entry->bo_va->base.vm,
1226 						&ctx->exec, 2);
1227 			drm_exec_retry_on_contention(&ctx->exec);
1228 			if (unlikely(ret))
1229 				goto error;
1230 			++ctx->n_vms;
1231 		}
1232 
1233 		ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
1234 		drm_exec_retry_on_contention(&ctx->exec);
1235 		if (unlikely(ret))
1236 			goto error;
1237 	}
1238 	return 0;
1239 
1240 error:
1241 	pr_err("Failed to reserve buffers in ttm.\n");
1242 	drm_exec_fini(&ctx->exec);
1243 	return ret;
1244 }
1245 
1246 /**
1247  * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1248  * @ctx: Reservation context to unreserve
1249  * @wait: Optionally wait for a sync object representing pending VM updates
1250  * @intr: Whether the wait is interruptible
1251  *
1252  * Also frees any resources allocated in
1253  * reserve_bo_and_(cond_)vm(s). Returns the status from
1254  * amdgpu_sync_wait.
1255  */
1256 static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
1257 				 bool wait, bool intr)
1258 {
1259 	int ret = 0;
1260 
1261 	if (wait)
1262 		ret = amdgpu_sync_wait(ctx->sync, intr);
1263 
1264 	drm_exec_fini(&ctx->exec);
1265 	ctx->sync = NULL;
1266 	return ret;
1267 }
1268 
1269 static int unmap_bo_from_gpuvm(struct kgd_mem *mem,
1270 				struct kfd_mem_attachment *entry,
1271 				struct amdgpu_sync *sync)
1272 {
1273 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1274 	struct amdgpu_device *adev = entry->adev;
1275 	struct amdgpu_vm *vm = bo_va->base.vm;
1276 
1277 	if (bo_va->queue_refcount) {
1278 		pr_debug("bo_va->queue_refcount %d\n", bo_va->queue_refcount);
1279 		return -EBUSY;
1280 	}
1281 
1282 	(void)amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
1283 
1284 	/* VM entity stopped if process killed, don't clear freed pt bo */
1285 	if (!amdgpu_vm_ready(vm))
1286 		return 0;
1287 
1288 	(void)amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
1289 
1290 	(void)amdgpu_sync_fence(sync, bo_va->last_pt_update, GFP_KERNEL);
1291 
1292 	return 0;
1293 }
1294 
1295 static int update_gpuvm_pte(struct kgd_mem *mem,
1296 			    struct kfd_mem_attachment *entry,
1297 			    struct amdgpu_sync *sync)
1298 {
1299 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1300 	struct amdgpu_device *adev = entry->adev;
1301 	int ret;
1302 
1303 	ret = kfd_mem_dmamap_attachment(mem, entry);
1304 	if (ret)
1305 		return ret;
1306 
1307 	/* Update the page tables  */
1308 	ret = amdgpu_vm_bo_update(adev, bo_va, false);
1309 	if (ret) {
1310 		pr_err("amdgpu_vm_bo_update failed\n");
1311 		return ret;
1312 	}
1313 
1314 	return amdgpu_sync_fence(sync, bo_va->last_pt_update, GFP_KERNEL);
1315 }
1316 
1317 static int map_bo_to_gpuvm(struct kgd_mem *mem,
1318 			   struct kfd_mem_attachment *entry,
1319 			   struct amdgpu_sync *sync,
1320 			   bool no_update_pte)
1321 {
1322 	int ret;
1323 
1324 	/* Set virtual address for the allocation */
1325 	ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0,
1326 			       amdgpu_bo_size(entry->bo_va->base.bo),
1327 			       entry->pte_flags);
1328 	if (ret) {
1329 		pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
1330 				entry->va, ret);
1331 		return ret;
1332 	}
1333 
1334 	if (no_update_pte)
1335 		return 0;
1336 
1337 	ret = update_gpuvm_pte(mem, entry, sync);
1338 	if (ret) {
1339 		pr_err("update_gpuvm_pte() failed\n");
1340 		goto update_gpuvm_pte_failed;
1341 	}
1342 
1343 	return 0;
1344 
1345 update_gpuvm_pte_failed:
1346 	unmap_bo_from_gpuvm(mem, entry, sync);
1347 	kfd_mem_dmaunmap_attachment(mem, entry);
1348 	return ret;
1349 }
1350 
1351 static int process_validate_vms(struct amdkfd_process_info *process_info,
1352 				struct ww_acquire_ctx *ticket)
1353 {
1354 	struct amdgpu_vm *peer_vm;
1355 	int ret;
1356 
1357 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1358 			    vm_list_node) {
1359 		ret = vm_validate_pt_pd_bos(peer_vm, ticket);
1360 		if (ret)
1361 			return ret;
1362 	}
1363 
1364 	return 0;
1365 }
1366 
1367 static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
1368 				 struct amdgpu_sync *sync)
1369 {
1370 	struct amdgpu_vm *peer_vm;
1371 	int ret;
1372 
1373 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1374 			    vm_list_node) {
1375 		struct amdgpu_bo *pd = peer_vm->root.bo;
1376 
1377 		ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
1378 				       AMDGPU_SYNC_NE_OWNER,
1379 				       AMDGPU_FENCE_OWNER_KFD);
1380 		if (ret)
1381 			return ret;
1382 	}
1383 
1384 	return 0;
1385 }
1386 
1387 static int process_update_pds(struct amdkfd_process_info *process_info,
1388 			      struct amdgpu_sync *sync)
1389 {
1390 	struct amdgpu_vm *peer_vm;
1391 	int ret;
1392 
1393 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1394 			    vm_list_node) {
1395 		ret = vm_update_pds(peer_vm, sync);
1396 		if (ret)
1397 			return ret;
1398 	}
1399 
1400 	return 0;
1401 }
1402 
1403 static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
1404 		       struct dma_fence **ef)
1405 {
1406 	struct amdkfd_process_info *info = NULL;
1407 	struct kfd_process *process = NULL;
1408 	int ret;
1409 
1410 	process = container_of(process_info, struct kfd_process, kgd_process_info);
1411 	if (!*process_info) {
1412 		info = kzalloc_obj(*info, GFP_KERNEL);
1413 		if (!info)
1414 			return -ENOMEM;
1415 
1416 		mutex_init(&info->lock);
1417 		mutex_init(&info->notifier_lock);
1418 		INIT_LIST_HEAD(&info->vm_list_head);
1419 		INIT_LIST_HEAD(&info->kfd_bo_list);
1420 		INIT_LIST_HEAD(&info->userptr_valid_list);
1421 		INIT_LIST_HEAD(&info->userptr_inval_list);
1422 
1423 		info->eviction_fence =
1424 			amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
1425 						   current->mm,
1426 						   NULL, process->context_id);
1427 		if (!info->eviction_fence) {
1428 			pr_err("Failed to create eviction fence\n");
1429 			ret = -ENOMEM;
1430 			goto create_evict_fence_fail;
1431 		}
1432 
1433 		info->pid = get_task_pid(current, PIDTYPE_TGID);
1434 		INIT_DELAYED_WORK(&info->restore_userptr_work,
1435 				  amdgpu_amdkfd_restore_userptr_worker);
1436 
1437 		info->context_id = process->context_id;
1438 
1439 		*process_info = info;
1440 	}
1441 
1442 	vm->process_info = *process_info;
1443 
1444 	/* Validate page directory and attach eviction fence */
1445 	ret = amdgpu_bo_reserve(vm->root.bo, true);
1446 	if (ret)
1447 		goto reserve_pd_fail;
1448 	ret = vm_validate_pt_pd_bos(vm, NULL);
1449 	if (ret) {
1450 		pr_err("validate_pt_pd_bos() failed\n");
1451 		goto validate_pd_fail;
1452 	}
1453 	ret = amdgpu_bo_sync_wait(vm->root.bo,
1454 				  AMDGPU_FENCE_OWNER_KFD, false);
1455 	if (ret)
1456 		goto wait_pd_fail;
1457 	ret = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1);
1458 	if (ret)
1459 		goto reserve_shared_fail;
1460 	dma_resv_add_fence(vm->root.bo->tbo.base.resv,
1461 			   &vm->process_info->eviction_fence->base,
1462 			   DMA_RESV_USAGE_BOOKKEEP);
1463 	amdgpu_bo_unreserve(vm->root.bo);
1464 
1465 	/* Update process info */
1466 	mutex_lock(&vm->process_info->lock);
1467 	list_add_tail(&vm->vm_list_node,
1468 			&(vm->process_info->vm_list_head));
1469 	vm->process_info->n_vms++;
1470 	if (ef)
1471 		*ef = dma_fence_get(&vm->process_info->eviction_fence->base);
1472 	mutex_unlock(&vm->process_info->lock);
1473 
1474 	return 0;
1475 
1476 reserve_shared_fail:
1477 wait_pd_fail:
1478 validate_pd_fail:
1479 	amdgpu_bo_unreserve(vm->root.bo);
1480 reserve_pd_fail:
1481 	vm->process_info = NULL;
1482 	if (info) {
1483 		dma_fence_put(&info->eviction_fence->base);
1484 		*process_info = NULL;
1485 		put_pid(info->pid);
1486 create_evict_fence_fail:
1487 		mutex_destroy(&info->lock);
1488 		mutex_destroy(&info->notifier_lock);
1489 		kfree(info);
1490 	}
1491 	return ret;
1492 }
1493 
1494 /**
1495  * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
1496  * @bo: Handle of buffer object being pinned
1497  * @domain: Domain into which BO should be pinned
1498  *
1499  *   - USERPTR BOs are UNPINNABLE and will return error
1500  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1501  *     PIN count incremented. It is valid to PIN a BO multiple times
1502  *
1503  * Return: ZERO if successful in pinning, Non-Zero in case of error.
1504  */
1505 static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain)
1506 {
1507 	int ret = 0;
1508 
1509 	ret = amdgpu_bo_reserve(bo, false);
1510 	if (unlikely(ret))
1511 		return ret;
1512 
1513 	if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) {
1514 		/*
1515 		 * If bo is not contiguous on VRAM, move to system memory first to ensure
1516 		 * we can get contiguous VRAM space after evicting other BOs.
1517 		 */
1518 		if (!(bo->tbo.resource->placement & TTM_PL_FLAG_CONTIGUOUS)) {
1519 			struct ttm_operation_ctx ctx = { true, false };
1520 
1521 			amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
1522 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1523 			if (unlikely(ret)) {
1524 				pr_debug("validate bo 0x%p to GTT failed %d\n", &bo->tbo, ret);
1525 				goto out;
1526 			}
1527 		}
1528 	}
1529 
1530 	ret = amdgpu_bo_pin(bo, domain);
1531 	if (ret)
1532 		pr_err("Error in Pinning BO to domain: %d\n", domain);
1533 
1534 	amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
1535 out:
1536 	amdgpu_bo_unreserve(bo);
1537 	return ret;
1538 }
1539 
1540 /**
1541  * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
1542  * @bo: Handle of buffer object being unpinned
1543  *
1544  *   - Is a illegal request for USERPTR BOs and is ignored
1545  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1546  *     PIN count decremented. Calls to UNPIN must balance calls to PIN
1547  */
1548 static void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo)
1549 {
1550 	int ret = 0;
1551 
1552 	ret = amdgpu_bo_reserve(bo, false);
1553 	if (unlikely(ret))
1554 		return;
1555 
1556 	amdgpu_bo_unpin(bo);
1557 	amdgpu_bo_unreserve(bo);
1558 }
1559 
1560 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
1561 					   struct amdgpu_vm *avm,
1562 					   void **process_info,
1563 					   struct dma_fence **ef)
1564 {
1565 	int ret;
1566 
1567 	/* Already a compute VM? */
1568 	if (avm->process_info)
1569 		return -EINVAL;
1570 
1571 	/* Convert VM into a compute VM */
1572 	ret = amdgpu_vm_make_compute(adev, avm);
1573 	if (ret)
1574 		return ret;
1575 
1576 	/* Initialize KFD part of the VM and process info */
1577 	ret = init_kfd_vm(avm, process_info, ef);
1578 	if (ret)
1579 		return ret;
1580 
1581 	amdgpu_vm_set_task_info(avm);
1582 
1583 	return 0;
1584 }
1585 
1586 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1587 				    struct amdgpu_vm *vm)
1588 {
1589 	struct amdkfd_process_info *process_info = vm->process_info;
1590 
1591 	if (!process_info)
1592 		return;
1593 
1594 	/* Update process info */
1595 	mutex_lock(&process_info->lock);
1596 	process_info->n_vms--;
1597 	list_del(&vm->vm_list_node);
1598 	mutex_unlock(&process_info->lock);
1599 
1600 	vm->process_info = NULL;
1601 
1602 	/* Release per-process resources when last compute VM is destroyed */
1603 	if (!process_info->n_vms) {
1604 		WARN_ON(!list_empty(&process_info->kfd_bo_list));
1605 		WARN_ON(!list_empty(&process_info->userptr_valid_list));
1606 		WARN_ON(!list_empty(&process_info->userptr_inval_list));
1607 
1608 		dma_fence_put(&process_info->eviction_fence->base);
1609 		cancel_delayed_work_sync(&process_info->restore_userptr_work);
1610 		put_pid(process_info->pid);
1611 		mutex_destroy(&process_info->lock);
1612 		mutex_destroy(&process_info->notifier_lock);
1613 		kfree(process_info);
1614 	}
1615 }
1616 
1617 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv)
1618 {
1619 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1620 	struct amdgpu_bo *pd = avm->root.bo;
1621 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1622 
1623 	if (adev->asic_type < CHIP_VEGA10)
1624 		return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1625 	return avm->pd_phys_addr;
1626 }
1627 
1628 void amdgpu_amdkfd_block_mmu_notifications(void *p)
1629 {
1630 	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1631 
1632 	mutex_lock(&pinfo->lock);
1633 	WRITE_ONCE(pinfo->block_mmu_notifications, true);
1634 	mutex_unlock(&pinfo->lock);
1635 }
1636 
1637 int amdgpu_amdkfd_criu_resume(void *p)
1638 {
1639 	int ret = 0;
1640 	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1641 
1642 	mutex_lock(&pinfo->lock);
1643 	pr_debug("scheduling work\n");
1644 	mutex_lock(&pinfo->notifier_lock);
1645 	pinfo->evicted_bos++;
1646 	mutex_unlock(&pinfo->notifier_lock);
1647 	if (!READ_ONCE(pinfo->block_mmu_notifications)) {
1648 		ret = -EINVAL;
1649 		goto out_unlock;
1650 	}
1651 	WRITE_ONCE(pinfo->block_mmu_notifications, false);
1652 	queue_delayed_work(system_freezable_wq,
1653 			   &pinfo->restore_userptr_work, 0);
1654 
1655 out_unlock:
1656 	mutex_unlock(&pinfo->lock);
1657 	return ret;
1658 }
1659 
1660 size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev,
1661 					  uint8_t xcp_id)
1662 {
1663 	uint64_t reserved_for_pt =
1664 		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
1665 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1666 	uint64_t reserved_for_ras = (con ? con->reserved_pages_in_bytes : 0);
1667 	ssize_t available;
1668 	uint64_t vram_available, system_mem_available, ttm_mem_available;
1669 
1670 	spin_lock(&kfd_mem_limit.mem_limit_lock);
1671 	if (adev->apu_prefer_gtt && !adev->gmc.is_app_apu)
1672 		vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id)
1673 			- adev->kfd.vram_used_aligned[xcp_id];
1674 	else
1675 		vram_available = KFD_XCP_MEMORY_SIZE(adev, xcp_id)
1676 			- adev->kfd.vram_used_aligned[xcp_id]
1677 			- atomic64_read(&adev->vram_pin_size)
1678 			- reserved_for_pt
1679 			- reserved_for_ras;
1680 
1681 	if (adev->apu_prefer_gtt) {
1682 		system_mem_available = no_system_mem_limit ?
1683 					kfd_mem_limit.max_system_mem_limit :
1684 					kfd_mem_limit.max_system_mem_limit -
1685 					kfd_mem_limit.system_mem_used;
1686 
1687 		ttm_mem_available = kfd_mem_limit.max_ttm_mem_limit -
1688 				kfd_mem_limit.ttm_mem_used;
1689 
1690 		available = min3(system_mem_available, ttm_mem_available,
1691 				 vram_available);
1692 		available = ALIGN_DOWN(available, PAGE_SIZE);
1693 	} else {
1694 		available = ALIGN_DOWN(vram_available, VRAM_AVAILABLITY_ALIGN);
1695 	}
1696 
1697 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
1698 
1699 	if (available < 0)
1700 		available = 0;
1701 
1702 	return available;
1703 }
1704 
1705 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1706 		struct amdgpu_device *adev, uint64_t va, uint64_t size,
1707 		void *drm_priv, struct kgd_mem **mem,
1708 		uint64_t *offset, uint32_t flags, bool criu_resume)
1709 {
1710 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1711 	struct amdgpu_fpriv *fpriv = container_of(avm, struct amdgpu_fpriv, vm);
1712 	enum ttm_bo_type bo_type = ttm_bo_type_device;
1713 	struct sg_table *sg = NULL;
1714 	uint64_t user_addr = 0;
1715 	struct amdgpu_bo *bo;
1716 	struct drm_gem_object *gobj = NULL;
1717 	u32 domain, alloc_domain;
1718 	uint64_t aligned_size;
1719 	int8_t xcp_id = -1;
1720 	u64 alloc_flags;
1721 	int ret;
1722 
1723 	/*
1724 	 * Check on which domain to allocate BO
1725 	 */
1726 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1727 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1728 
1729 		if (adev->apu_prefer_gtt) {
1730 			domain = AMDGPU_GEM_DOMAIN_GTT;
1731 			alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1732 			alloc_flags = 0;
1733 		} else {
1734 			alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1735 			alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1736 			AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0;
1737 
1738 			/* For contiguous VRAM allocation */
1739 			if (flags & KFD_IOC_ALLOC_MEM_FLAGS_CONTIGUOUS)
1740 				alloc_flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1741 		}
1742 		xcp_id = fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION ?
1743 					0 : fpriv->xcp_id;
1744 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1745 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1746 		alloc_flags = 0;
1747 	} else {
1748 		domain = AMDGPU_GEM_DOMAIN_GTT;
1749 		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1750 		alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE;
1751 
1752 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1753 			if (!offset || !*offset)
1754 				return -EINVAL;
1755 			user_addr = untagged_addr(*offset);
1756 		} else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1757 				    KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1758 			bo_type = ttm_bo_type_sg;
1759 			if (size > UINT_MAX)
1760 				return -EINVAL;
1761 			sg = create_sg_table(*offset, size);
1762 			if (!sg)
1763 				return -ENOMEM;
1764 		} else {
1765 			return -EINVAL;
1766 		}
1767 	}
1768 
1769 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT)
1770 		alloc_flags |= AMDGPU_GEM_CREATE_COHERENT;
1771 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_EXT_COHERENT)
1772 		alloc_flags |= AMDGPU_GEM_CREATE_EXT_COHERENT;
1773 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED)
1774 		alloc_flags |= AMDGPU_GEM_CREATE_UNCACHED;
1775 
1776 	*mem = kzalloc_obj(struct kgd_mem, GFP_KERNEL);
1777 	if (!*mem) {
1778 		ret = -ENOMEM;
1779 		goto err;
1780 	}
1781 	INIT_LIST_HEAD(&(*mem)->attachments);
1782 	mutex_init(&(*mem)->lock);
1783 	(*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1784 
1785 	/* Workaround for AQL queue wraparound bug. Map the same
1786 	 * memory twice. That means we only actually allocate half
1787 	 * the memory.
1788 	 */
1789 	if ((*mem)->aql_queue)
1790 		size >>= 1;
1791 	aligned_size = PAGE_ALIGN(size);
1792 
1793 	(*mem)->alloc_flags = flags;
1794 
1795 	amdgpu_sync_create(&(*mem)->sync);
1796 
1797 	ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags,
1798 					      xcp_id);
1799 	if (ret) {
1800 		pr_debug("Insufficient memory\n");
1801 		goto err_reserve_limit;
1802 	}
1803 
1804 	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s xcp_id %d\n",
1805 		 va, (*mem)->aql_queue ? size << 1 : size,
1806 		 domain_string(alloc_domain), xcp_id);
1807 
1808 	ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags,
1809 				       bo_type, NULL, &gobj, xcp_id + 1);
1810 	if (ret) {
1811 		pr_debug("Failed to create BO on domain %s. ret %d\n",
1812 			 domain_string(alloc_domain), ret);
1813 		goto err_bo_create;
1814 	}
1815 	ret = drm_vma_node_allow(&gobj->vma_node, drm_priv);
1816 	if (ret) {
1817 		pr_debug("Failed to allow vma node access. ret %d\n", ret);
1818 		goto err_node_allow;
1819 	}
1820 	ret = drm_gem_handle_create(adev->kfd.client.file, gobj, &(*mem)->gem_handle);
1821 	if (ret)
1822 		goto err_gem_handle_create;
1823 	bo = gem_to_amdgpu_bo(gobj);
1824 	if (bo_type == ttm_bo_type_sg) {
1825 		bo->tbo.sg = sg;
1826 		bo->tbo.ttm->sg = sg;
1827 	}
1828 	bo->kfd_bo = *mem;
1829 	(*mem)->bo = bo;
1830 	if (user_addr)
1831 		bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO;
1832 
1833 	(*mem)->va = va;
1834 	(*mem)->domain = domain;
1835 	(*mem)->mapped_to_gpu_memory = 0;
1836 	(*mem)->process_info = avm->process_info;
1837 
1838 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1839 
1840 	if (user_addr) {
1841 		pr_debug("creating userptr BO for user_addr = %llx\n", user_addr);
1842 		ret = init_user_pages(*mem, user_addr, criu_resume);
1843 		if (ret)
1844 			goto allocate_init_user_pages_failed;
1845 	} else  if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1846 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1847 		ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT);
1848 		if (ret) {
1849 			pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n");
1850 			goto err_pin_bo;
1851 		}
1852 		bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
1853 		bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
1854 	} else {
1855 		mutex_lock(&avm->process_info->lock);
1856 		if (avm->process_info->eviction_fence &&
1857 		    !dma_fence_is_signaled(&avm->process_info->eviction_fence->base))
1858 			ret = amdgpu_amdkfd_bo_validate_and_fence(bo, domain,
1859 				&avm->process_info->eviction_fence->base);
1860 		mutex_unlock(&avm->process_info->lock);
1861 		if (ret)
1862 			goto err_validate_bo;
1863 	}
1864 
1865 	if (offset)
1866 		*offset = amdgpu_bo_mmap_offset(bo);
1867 
1868 	return 0;
1869 
1870 allocate_init_user_pages_failed:
1871 err_pin_bo:
1872 err_validate_bo:
1873 	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1874 	drm_gem_handle_delete(adev->kfd.client.file, (*mem)->gem_handle);
1875 err_gem_handle_create:
1876 	drm_vma_node_revoke(&gobj->vma_node, drm_priv);
1877 err_node_allow:
1878 	/* Don't unreserve system mem limit twice */
1879 	goto err_reserve_limit;
1880 err_bo_create:
1881 	amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags, xcp_id);
1882 err_reserve_limit:
1883 	amdgpu_sync_free(&(*mem)->sync);
1884 	mutex_destroy(&(*mem)->lock);
1885 	if (gobj)
1886 		drm_gem_object_put(gobj);
1887 	else
1888 		kfree(*mem);
1889 err:
1890 	if (sg) {
1891 		sg_free_table(sg);
1892 		kfree(sg);
1893 	}
1894 	return ret;
1895 }
1896 
1897 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1898 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
1899 		uint64_t *size)
1900 {
1901 	struct amdkfd_process_info *process_info = mem->process_info;
1902 	unsigned long bo_size = mem->bo->tbo.base.size;
1903 	bool use_release_notifier = (mem->bo->kfd_bo == mem);
1904 	struct kfd_mem_attachment *entry, *tmp;
1905 	struct bo_vm_reservation_context ctx;
1906 	unsigned int mapped_to_gpu_memory;
1907 	int ret;
1908 	bool is_imported = false;
1909 
1910 	mutex_lock(&mem->lock);
1911 
1912 	/* Unpin MMIO/DOORBELL BO's that were pinned during allocation */
1913 	if (mem->alloc_flags &
1914 	    (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1915 	     KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1916 		amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo);
1917 	}
1918 
1919 	mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1920 	is_imported = mem->is_imported;
1921 	mutex_unlock(&mem->lock);
1922 	/* lock is not needed after this, since mem is unused and will
1923 	 * be freed anyway
1924 	 */
1925 
1926 	if (mapped_to_gpu_memory > 0) {
1927 		pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1928 				mem->va, bo_size);
1929 		return -EBUSY;
1930 	}
1931 
1932 	/* Make sure restore workers don't access the BO any more */
1933 	mutex_lock(&process_info->lock);
1934 	if (!list_empty(&mem->validate_list))
1935 		list_del_init(&mem->validate_list);
1936 	mutex_unlock(&process_info->lock);
1937 
1938 	ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1939 	if (unlikely(ret))
1940 		return ret;
1941 
1942 	/* Cleanup user pages and MMU notifiers */
1943 	if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
1944 		amdgpu_hmm_unregister(mem->bo);
1945 		amdgpu_hmm_range_free(mem->range);
1946 		mem->range = NULL;
1947 	}
1948 
1949 	amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1950 					process_info->eviction_fence);
1951 	pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1952 		mem->va + bo_size * (1 + mem->aql_queue));
1953 
1954 	/* Remove from VM internal data structures */
1955 	list_for_each_entry_safe(entry, tmp, &mem->attachments, list) {
1956 		kfd_mem_dmaunmap_attachment(mem, entry);
1957 		kfd_mem_detach(entry);
1958 	}
1959 
1960 	ret = unreserve_bo_and_vms(&ctx, false, false);
1961 
1962 	/* Free the sync object */
1963 	amdgpu_sync_free(&mem->sync);
1964 
1965 	/* If the SG is not NULL, it's one we created for a doorbell or mmio
1966 	 * remap BO. We need to free it.
1967 	 */
1968 	if (mem->bo->tbo.sg) {
1969 		sg_free_table(mem->bo->tbo.sg);
1970 		kfree(mem->bo->tbo.sg);
1971 	}
1972 
1973 	/* Update the size of the BO being freed if it was allocated from
1974 	 * VRAM and is not imported. For APP APU VRAM allocations are done
1975 	 * in GTT domain
1976 	 */
1977 	if (size) {
1978 		if (!is_imported &&
1979 		   mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)
1980 			*size = bo_size;
1981 		else
1982 			*size = 0;
1983 	}
1984 
1985 	/* Free the BO*/
1986 	drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
1987 	drm_gem_handle_delete(adev->kfd.client.file, mem->gem_handle);
1988 	if (mem->dmabuf) {
1989 		dma_buf_put(mem->dmabuf);
1990 		mem->dmabuf = NULL;
1991 	}
1992 	mutex_destroy(&mem->lock);
1993 
1994 	/* If this releases the last reference, it will end up calling
1995 	 * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why
1996 	 * this needs to be the last call here.
1997 	 */
1998 	drm_gem_object_put(&mem->bo->tbo.base);
1999 
2000 	/*
2001 	 * For kgd_mem allocated in import_obj_create() via
2002 	 * amdgpu_amdkfd_gpuvm_import_dmabuf_fd(),
2003 	 * explicitly free it here.
2004 	 */
2005 	if (!use_release_notifier)
2006 		kfree(mem);
2007 
2008 	return ret;
2009 }
2010 
2011 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
2012 		struct amdgpu_device *adev, struct kgd_mem *mem,
2013 		void *drm_priv)
2014 {
2015 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2016 	int ret;
2017 	struct amdgpu_bo *bo;
2018 	uint32_t domain;
2019 	struct kfd_mem_attachment *entry;
2020 	struct bo_vm_reservation_context ctx;
2021 	unsigned long bo_size;
2022 	bool is_invalid_userptr = false;
2023 
2024 	bo = mem->bo;
2025 	if (!bo) {
2026 		pr_err("Invalid BO when mapping memory to GPU\n");
2027 		return -EINVAL;
2028 	}
2029 
2030 	/* Make sure restore is not running concurrently. Since we
2031 	 * don't map invalid userptr BOs, we rely on the next restore
2032 	 * worker to do the mapping
2033 	 */
2034 	mutex_lock(&mem->process_info->lock);
2035 
2036 	/* Lock notifier lock. If we find an invalid userptr BO, we can be
2037 	 * sure that the MMU notifier is no longer running
2038 	 * concurrently and the queues are actually stopped
2039 	 */
2040 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2041 		mutex_lock(&mem->process_info->notifier_lock);
2042 		is_invalid_userptr = !!mem->invalid;
2043 		mutex_unlock(&mem->process_info->notifier_lock);
2044 	}
2045 
2046 	mutex_lock(&mem->lock);
2047 
2048 	domain = mem->domain;
2049 	bo_size = bo->tbo.base.size;
2050 
2051 	pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
2052 			mem->va,
2053 			mem->va + bo_size * (1 + mem->aql_queue),
2054 			avm, domain_string(domain));
2055 
2056 	if (!kfd_mem_is_attached(avm, mem)) {
2057 		ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue);
2058 		if (ret)
2059 			goto out;
2060 	}
2061 
2062 	ret = reserve_bo_and_vm(mem, avm, &ctx);
2063 	if (unlikely(ret))
2064 		goto out;
2065 
2066 	/* Userptr can be marked as "not invalid", but not actually be
2067 	 * validated yet (still in the system domain). In that case
2068 	 * the queues are still stopped and we can leave mapping for
2069 	 * the next restore worker
2070 	 */
2071 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
2072 	    bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
2073 		is_invalid_userptr = true;
2074 
2075 	ret = vm_validate_pt_pd_bos(avm, NULL);
2076 	if (unlikely(ret))
2077 		goto out_unreserve;
2078 
2079 	list_for_each_entry(entry, &mem->attachments, list) {
2080 		if (entry->bo_va->base.vm != avm || entry->is_mapped)
2081 			continue;
2082 
2083 		pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
2084 			 entry->va, entry->va + bo_size, entry);
2085 
2086 		ret = map_bo_to_gpuvm(mem, entry, ctx.sync,
2087 				      is_invalid_userptr);
2088 		if (ret) {
2089 			pr_err("Failed to map bo to gpuvm\n");
2090 			goto out_unreserve;
2091 		}
2092 
2093 		ret = vm_update_pds(avm, ctx.sync);
2094 		if (ret) {
2095 			pr_err("Failed to update page directories\n");
2096 			goto out_unreserve;
2097 		}
2098 
2099 		entry->is_mapped = true;
2100 		mem->mapped_to_gpu_memory++;
2101 		pr_debug("\t INC mapping count %d\n",
2102 			 mem->mapped_to_gpu_memory);
2103 	}
2104 
2105 	ret = unreserve_bo_and_vms(&ctx, false, false);
2106 
2107 	goto out;
2108 
2109 out_unreserve:
2110 	unreserve_bo_and_vms(&ctx, false, false);
2111 out:
2112 	mutex_unlock(&mem->process_info->lock);
2113 	mutex_unlock(&mem->lock);
2114 	return ret;
2115 }
2116 
2117 int amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem, void *drm_priv)
2118 {
2119 	struct kfd_mem_attachment *entry;
2120 	struct amdgpu_vm *vm;
2121 	int ret;
2122 
2123 	vm = drm_priv_to_vm(drm_priv);
2124 
2125 	mutex_lock(&mem->lock);
2126 
2127 	ret = amdgpu_bo_reserve(mem->bo, true);
2128 	if (ret)
2129 		goto out;
2130 
2131 	list_for_each_entry(entry, &mem->attachments, list) {
2132 		if (entry->bo_va->base.vm != vm)
2133 			continue;
2134 		if (entry->bo_va->base.bo->tbo.ttm &&
2135 		    !entry->bo_va->base.bo->tbo.ttm->sg)
2136 			continue;
2137 
2138 		kfd_mem_dmaunmap_attachment(mem, entry);
2139 	}
2140 
2141 	amdgpu_bo_unreserve(mem->bo);
2142 out:
2143 	mutex_unlock(&mem->lock);
2144 
2145 	return ret;
2146 }
2147 
2148 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
2149 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
2150 {
2151 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2152 	unsigned long bo_size = mem->bo->tbo.base.size;
2153 	struct kfd_mem_attachment *entry;
2154 	struct bo_vm_reservation_context ctx;
2155 	int ret;
2156 
2157 	mutex_lock(&mem->lock);
2158 
2159 	ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx);
2160 	if (unlikely(ret))
2161 		goto out;
2162 	/* If no VMs were reserved, it means the BO wasn't actually mapped */
2163 	if (ctx.n_vms == 0) {
2164 		ret = -EINVAL;
2165 		goto unreserve_out;
2166 	}
2167 
2168 	ret = vm_validate_pt_pd_bos(avm, NULL);
2169 	if (unlikely(ret))
2170 		goto unreserve_out;
2171 
2172 	pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
2173 		mem->va,
2174 		mem->va + bo_size * (1 + mem->aql_queue),
2175 		avm);
2176 
2177 	list_for_each_entry(entry, &mem->attachments, list) {
2178 		if (entry->bo_va->base.vm != avm || !entry->is_mapped)
2179 			continue;
2180 
2181 		pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
2182 			 entry->va, entry->va + bo_size, entry);
2183 
2184 		ret = unmap_bo_from_gpuvm(mem, entry, ctx.sync);
2185 		if (ret)
2186 			goto unreserve_out;
2187 
2188 		entry->is_mapped = false;
2189 
2190 		mem->mapped_to_gpu_memory--;
2191 		pr_debug("\t DEC mapping count %d\n",
2192 			 mem->mapped_to_gpu_memory);
2193 	}
2194 
2195 unreserve_out:
2196 	unreserve_bo_and_vms(&ctx, false, false);
2197 out:
2198 	mutex_unlock(&mem->lock);
2199 	return ret;
2200 }
2201 
2202 int amdgpu_amdkfd_gpuvm_sync_memory(
2203 		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr)
2204 {
2205 	struct amdgpu_sync sync;
2206 	int ret;
2207 
2208 	amdgpu_sync_create(&sync);
2209 
2210 	mutex_lock(&mem->lock);
2211 	amdgpu_sync_clone(&mem->sync, &sync);
2212 	mutex_unlock(&mem->lock);
2213 
2214 	ret = amdgpu_sync_wait(&sync, intr);
2215 	amdgpu_sync_free(&sync);
2216 	return ret;
2217 }
2218 
2219 /**
2220  * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count
2221  * @bo: Buffer object to be mapped
2222  * @bo_gart: Return bo reference
2223  *
2224  * Before return, bo reference count is incremented. To release the reference and unpin/
2225  * unmap the BO, call amdgpu_amdkfd_free_kernel_mem.
2226  */
2227 int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_bo *bo, struct amdgpu_bo **bo_gart)
2228 {
2229 	int ret;
2230 
2231 	ret = amdgpu_bo_reserve(bo, true);
2232 	if (ret) {
2233 		pr_err("Failed to reserve bo. ret %d\n", ret);
2234 		goto err_reserve_bo_failed;
2235 	}
2236 
2237 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2238 	if (ret) {
2239 		pr_err("Failed to pin bo. ret %d\n", ret);
2240 		goto err_pin_bo_failed;
2241 	}
2242 
2243 	ret = amdgpu_ttm_alloc_gart(&bo->tbo);
2244 	if (ret) {
2245 		pr_err("Failed to bind bo to GART. ret %d\n", ret);
2246 		goto err_map_bo_gart_failed;
2247 	}
2248 
2249 	amdgpu_amdkfd_remove_eviction_fence(
2250 		bo, bo->vm_bo->vm->process_info->eviction_fence);
2251 
2252 	amdgpu_bo_unreserve(bo);
2253 
2254 	*bo_gart = amdgpu_bo_ref(bo);
2255 
2256 	return 0;
2257 
2258 err_map_bo_gart_failed:
2259 	amdgpu_bo_unpin(bo);
2260 err_pin_bo_failed:
2261 	amdgpu_bo_unreserve(bo);
2262 err_reserve_bo_failed:
2263 
2264 	return ret;
2265 }
2266 
2267 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Map a GTT BO for kernel CPU access
2268  *
2269  * @mem: Buffer object to be mapped for CPU access
2270  * @kptr[out]: pointer in kernel CPU address space
2271  * @size[out]: size of the buffer
2272  *
2273  * Pins the BO and maps it for kernel CPU access. The eviction fence is removed
2274  * from the BO, since pinned BOs cannot be evicted. The bo must remain on the
2275  * validate_list, so the GPU mapping can be restored after a page table was
2276  * evicted.
2277  *
2278  * Return: 0 on success, error code on failure
2279  */
2280 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,
2281 					     void **kptr, uint64_t *size)
2282 {
2283 	int ret;
2284 	struct amdgpu_bo *bo = mem->bo;
2285 
2286 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2287 		pr_err("userptr can't be mapped to kernel\n");
2288 		return -EINVAL;
2289 	}
2290 
2291 	mutex_lock(&mem->process_info->lock);
2292 
2293 	ret = amdgpu_bo_reserve(bo, true);
2294 	if (ret) {
2295 		pr_err("Failed to reserve bo. ret %d\n", ret);
2296 		goto bo_reserve_failed;
2297 	}
2298 
2299 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2300 	if (ret) {
2301 		pr_err("Failed to pin bo. ret %d\n", ret);
2302 		goto pin_failed;
2303 	}
2304 
2305 	ret = amdgpu_bo_kmap(bo, kptr);
2306 	if (ret) {
2307 		pr_err("Failed to map bo to kernel. ret %d\n", ret);
2308 		goto kmap_failed;
2309 	}
2310 
2311 	amdgpu_amdkfd_remove_eviction_fence(
2312 		bo, mem->process_info->eviction_fence);
2313 
2314 	if (size)
2315 		*size = amdgpu_bo_size(bo);
2316 
2317 	amdgpu_bo_unreserve(bo);
2318 
2319 	mutex_unlock(&mem->process_info->lock);
2320 	return 0;
2321 
2322 kmap_failed:
2323 	amdgpu_bo_unpin(bo);
2324 pin_failed:
2325 	amdgpu_bo_unreserve(bo);
2326 bo_reserve_failed:
2327 	mutex_unlock(&mem->process_info->lock);
2328 
2329 	return ret;
2330 }
2331 
2332 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Unmap a GTT BO for kernel CPU access
2333  *
2334  * @mem: Buffer object to be unmapped for CPU access
2335  *
2336  * Removes the kernel CPU mapping and unpins the BO. It does not restore the
2337  * eviction fence, so this function should only be used for cleanup before the
2338  * BO is destroyed.
2339  */
2340 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem)
2341 {
2342 	struct amdgpu_bo *bo = mem->bo;
2343 
2344 	(void)amdgpu_bo_reserve(bo, true);
2345 	amdgpu_bo_kunmap(bo);
2346 	amdgpu_bo_unpin(bo);
2347 	amdgpu_bo_unreserve(bo);
2348 }
2349 
2350 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
2351 					  struct kfd_vm_fault_info *mem)
2352 {
2353 	if (atomic_read_acquire(&adev->gmc.vm_fault_info_updated) == 1) {
2354 		*mem = *adev->gmc.vm_fault_info;
2355 		atomic_set_release(&adev->gmc.vm_fault_info_updated, 0);
2356 	}
2357 	return 0;
2358 }
2359 
2360 static int import_obj_create(struct amdgpu_device *adev,
2361 			     struct dma_buf *dma_buf,
2362 			     struct drm_gem_object *obj,
2363 			     uint64_t va, void *drm_priv,
2364 			     struct kgd_mem **mem, uint64_t *size,
2365 			     uint64_t *mmap_offset)
2366 {
2367 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2368 	struct amdgpu_bo *bo;
2369 	int ret;
2370 
2371 	bo = gem_to_amdgpu_bo(obj);
2372 	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
2373 				    AMDGPU_GEM_DOMAIN_GTT)))
2374 		/* Only VRAM and GTT BOs are supported */
2375 		return -EINVAL;
2376 
2377 	*mem = kzalloc_obj(struct kgd_mem, GFP_KERNEL);
2378 	if (!*mem)
2379 		return -ENOMEM;
2380 
2381 	ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
2382 	if (ret)
2383 		goto err_free_mem;
2384 
2385 	if (size)
2386 		*size = amdgpu_bo_size(bo);
2387 
2388 	if (mmap_offset)
2389 		*mmap_offset = amdgpu_bo_mmap_offset(bo);
2390 
2391 	INIT_LIST_HEAD(&(*mem)->attachments);
2392 	mutex_init(&(*mem)->lock);
2393 
2394 	(*mem)->alloc_flags =
2395 		((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2396 		KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
2397 		| KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
2398 		| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
2399 
2400 	get_dma_buf(dma_buf);
2401 	(*mem)->dmabuf = dma_buf;
2402 	(*mem)->bo = bo;
2403 	(*mem)->va = va;
2404 	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) &&
2405 			 !adev->apu_prefer_gtt ?
2406 			 AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
2407 
2408 	(*mem)->mapped_to_gpu_memory = 0;
2409 	(*mem)->process_info = avm->process_info;
2410 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
2411 	amdgpu_sync_create(&(*mem)->sync);
2412 	(*mem)->is_imported = true;
2413 
2414 	mutex_lock(&avm->process_info->lock);
2415 	if (avm->process_info->eviction_fence &&
2416 	    !dma_fence_is_signaled(&avm->process_info->eviction_fence->base))
2417 		ret = amdgpu_amdkfd_bo_validate_and_fence(bo, (*mem)->domain,
2418 				&avm->process_info->eviction_fence->base);
2419 	mutex_unlock(&avm->process_info->lock);
2420 	if (ret)
2421 		goto err_remove_mem;
2422 
2423 	return 0;
2424 
2425 err_remove_mem:
2426 	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
2427 	drm_vma_node_revoke(&obj->vma_node, drm_priv);
2428 err_free_mem:
2429 	kfree(*mem);
2430 	return ret;
2431 }
2432 
2433 int amdgpu_amdkfd_gpuvm_import_dmabuf_fd(struct amdgpu_device *adev, int fd,
2434 					 uint64_t va, void *drm_priv,
2435 					 struct kgd_mem **mem, uint64_t *size,
2436 					 uint64_t *mmap_offset)
2437 {
2438 	struct drm_gem_object *obj;
2439 	uint32_t handle;
2440 	int ret;
2441 
2442 	ret = drm_gem_prime_fd_to_handle(&adev->ddev, adev->kfd.client.file, fd,
2443 					 &handle);
2444 	if (ret)
2445 		return ret;
2446 	obj = drm_gem_object_lookup(adev->kfd.client.file, handle);
2447 	if (!obj) {
2448 		ret = -EINVAL;
2449 		goto err_release_handle;
2450 	}
2451 
2452 	ret = import_obj_create(adev, obj->dma_buf, obj, va, drm_priv, mem, size,
2453 				mmap_offset);
2454 	if (ret)
2455 		goto err_put_obj;
2456 
2457 	(*mem)->gem_handle = handle;
2458 
2459 	return 0;
2460 
2461 err_put_obj:
2462 	drm_gem_object_put(obj);
2463 err_release_handle:
2464 	drm_gem_handle_delete(adev->kfd.client.file, handle);
2465 	return ret;
2466 }
2467 
2468 int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem,
2469 				      struct dma_buf **dma_buf)
2470 {
2471 	int ret;
2472 
2473 	mutex_lock(&mem->lock);
2474 	ret = kfd_mem_export_dmabuf(mem);
2475 	if (ret)
2476 		goto out;
2477 
2478 	get_dma_buf(mem->dmabuf);
2479 	*dma_buf = mem->dmabuf;
2480 out:
2481 	mutex_unlock(&mem->lock);
2482 	return ret;
2483 }
2484 
2485 /* Evict a userptr BO by stopping the queues if necessary
2486  *
2487  * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
2488  * cannot do any memory allocations, and cannot take any locks that
2489  * are held elsewhere while allocating memory.
2490  *
2491  * It doesn't do anything to the BO itself. The real work happens in
2492  * restore, where we get updated page addresses. This function only
2493  * ensures that GPU access to the BO is stopped.
2494  */
2495 int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
2496 				unsigned long cur_seq, struct kgd_mem *mem)
2497 {
2498 	struct amdkfd_process_info *process_info = mem->process_info;
2499 	int r = 0;
2500 
2501 	/* Do not process MMU notifications during CRIU restore until
2502 	 * KFD_CRIU_OP_RESUME IOCTL is received
2503 	 */
2504 	if (READ_ONCE(process_info->block_mmu_notifications))
2505 		return 0;
2506 
2507 	mutex_lock(&process_info->notifier_lock);
2508 	mmu_interval_set_seq(mni, cur_seq);
2509 
2510 	mem->invalid++;
2511 	if (++process_info->evicted_bos == 1) {
2512 		/* First eviction, stop the queues */
2513 		r = kgd2kfd_quiesce_mm(mni->mm,
2514 				       KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
2515 
2516 		if (r && r != -ESRCH)
2517 			pr_err("Failed to quiesce KFD\n");
2518 
2519 		if (r != -ESRCH)
2520 			queue_delayed_work(system_freezable_wq,
2521 				&process_info->restore_userptr_work,
2522 				msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2523 	}
2524 	mutex_unlock(&process_info->notifier_lock);
2525 
2526 	return r;
2527 }
2528 
2529 /* Update invalid userptr BOs
2530  *
2531  * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
2532  * userptr_inval_list and updates user pages for all BOs that have
2533  * been invalidated since their last update.
2534  */
2535 static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
2536 				     struct mm_struct *mm)
2537 {
2538 	struct kgd_mem *mem, *tmp_mem;
2539 	struct amdgpu_bo *bo;
2540 	struct ttm_operation_ctx ctx = { false, false };
2541 	uint32_t invalid;
2542 	int ret = 0;
2543 
2544 	mutex_lock(&process_info->notifier_lock);
2545 
2546 	/* Move all invalidated BOs to the userptr_inval_list */
2547 	list_for_each_entry_safe(mem, tmp_mem,
2548 				 &process_info->userptr_valid_list,
2549 				 validate_list)
2550 		if (mem->invalid)
2551 			list_move_tail(&mem->validate_list,
2552 				       &process_info->userptr_inval_list);
2553 
2554 	/* Go through userptr_inval_list and update any invalid user_pages */
2555 	list_for_each_entry(mem, &process_info->userptr_inval_list,
2556 			    validate_list) {
2557 		invalid = mem->invalid;
2558 		if (!invalid)
2559 			/* BO hasn't been invalidated since the last
2560 			 * revalidation attempt. Keep its page list.
2561 			 */
2562 			continue;
2563 
2564 		bo = mem->bo;
2565 
2566 		amdgpu_hmm_range_free(mem->range);
2567 		mem->range = NULL;
2568 
2569 		/* BO reservations and getting user pages (hmm_range_fault)
2570 		 * must happen outside the notifier lock
2571 		 */
2572 		mutex_unlock(&process_info->notifier_lock);
2573 
2574 		/* Move the BO to system (CPU) domain if necessary to unmap
2575 		 * and free the SG table
2576 		 */
2577 		if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) {
2578 			if (amdgpu_bo_reserve(bo, true))
2579 				return -EAGAIN;
2580 			amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
2581 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2582 			amdgpu_bo_unreserve(bo);
2583 			if (ret) {
2584 				pr_err("%s: Failed to invalidate userptr BO\n",
2585 				       __func__);
2586 				return -EAGAIN;
2587 			}
2588 		}
2589 
2590 		mem->range = amdgpu_hmm_range_alloc(NULL);
2591 		if (unlikely(!mem->range))
2592 			return -ENOMEM;
2593 		/* Get updated user pages */
2594 		ret = amdgpu_ttm_tt_get_user_pages(bo, mem->range);
2595 		if (ret) {
2596 			amdgpu_hmm_range_free(mem->range);
2597 			mem->range = NULL;
2598 			pr_debug("Failed %d to get user pages\n", ret);
2599 
2600 			/* Return -EFAULT bad address error as success. It will
2601 			 * fail later with a VM fault if the GPU tries to access
2602 			 * it. Better than hanging indefinitely with stalled
2603 			 * user mode queues.
2604 			 *
2605 			 * Return other error -EBUSY or -ENOMEM to retry restore
2606 			 */
2607 			if (ret != -EFAULT)
2608 				return ret;
2609 
2610 			/* If applications unmap memory before destroying the userptr
2611 			 * from the KFD, trigger a segmentation fault in VM debug mode.
2612 			 */
2613 			if (amdgpu_ttm_adev(bo->tbo.bdev)->debug_vm_userptr) {
2614 				struct kfd_process *p;
2615 
2616 				pr_err("Pid %d unmapped memory before destroying userptr at GPU addr 0x%llx\n",
2617 								pid_nr(process_info->pid), mem->va);
2618 
2619 				// Send GPU VM fault to user space
2620 				p = kfd_lookup_process_by_pid(process_info->pid);
2621 				if (p) {
2622 					kfd_signal_vm_fault_event_with_userptr(p, mem->va);
2623 					kfd_unref_process(p);
2624 				}
2625 			}
2626 
2627 			ret = 0;
2628 		}
2629 
2630 		amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm, mem->range);
2631 
2632 		mutex_lock(&process_info->notifier_lock);
2633 
2634 		/* Mark the BO as valid unless it was invalidated
2635 		 * again concurrently.
2636 		 */
2637 		if (mem->invalid != invalid) {
2638 			ret = -EAGAIN;
2639 			goto unlock_out;
2640 		}
2641 		 /* set mem valid if mem has hmm range associated */
2642 		if (mem->range)
2643 			mem->invalid = 0;
2644 	}
2645 
2646 unlock_out:
2647 	mutex_unlock(&process_info->notifier_lock);
2648 
2649 	return ret;
2650 }
2651 
2652 /* Validate invalid userptr BOs
2653  *
2654  * Validates BOs on the userptr_inval_list. Also updates GPUVM page tables
2655  * with new page addresses and waits for the page table updates to complete.
2656  */
2657 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
2658 {
2659 	struct ttm_operation_ctx ctx = { false, false };
2660 	struct amdgpu_sync sync;
2661 	struct drm_exec exec;
2662 
2663 	struct amdgpu_vm *peer_vm;
2664 	struct kgd_mem *mem, *tmp_mem;
2665 	struct amdgpu_bo *bo;
2666 	int ret;
2667 
2668 	amdgpu_sync_create(&sync);
2669 
2670 	drm_exec_init(&exec, 0, 0);
2671 	/* Reserve all BOs and page tables for validation */
2672 	drm_exec_until_all_locked(&exec) {
2673 		/* Reserve all the page directories */
2674 		list_for_each_entry(peer_vm, &process_info->vm_list_head,
2675 				    vm_list_node) {
2676 			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2677 			drm_exec_retry_on_contention(&exec);
2678 			if (unlikely(ret))
2679 				goto unreserve_out;
2680 		}
2681 
2682 		/* Reserve the userptr_inval_list entries to resv_list */
2683 		list_for_each_entry(mem, &process_info->userptr_inval_list,
2684 				    validate_list) {
2685 			struct drm_gem_object *gobj;
2686 
2687 			gobj = &mem->bo->tbo.base;
2688 			ret = drm_exec_prepare_obj(&exec, gobj, 1);
2689 			drm_exec_retry_on_contention(&exec);
2690 			if (unlikely(ret))
2691 				goto unreserve_out;
2692 		}
2693 	}
2694 
2695 	ret = process_validate_vms(process_info, NULL);
2696 	if (ret)
2697 		goto unreserve_out;
2698 
2699 	/* Validate BOs and update GPUVM page tables */
2700 	list_for_each_entry_safe(mem, tmp_mem,
2701 				 &process_info->userptr_inval_list,
2702 				 validate_list) {
2703 		struct kfd_mem_attachment *attachment;
2704 
2705 		bo = mem->bo;
2706 
2707 		/* Validate the BO if we got user pages */
2708 		if (bo->tbo.ttm->pages[0]) {
2709 			amdgpu_bo_placement_from_domain(bo, mem->domain);
2710 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2711 			if (ret) {
2712 				pr_err("%s: failed to validate BO\n", __func__);
2713 				goto unreserve_out;
2714 			}
2715 		}
2716 
2717 		/* Update mapping. If the BO was not validated
2718 		 * (because we couldn't get user pages), this will
2719 		 * clear the page table entries, which will result in
2720 		 * VM faults if the GPU tries to access the invalid
2721 		 * memory.
2722 		 */
2723 		list_for_each_entry(attachment, &mem->attachments, list) {
2724 			if (!attachment->is_mapped)
2725 				continue;
2726 
2727 			kfd_mem_dmaunmap_attachment(mem, attachment);
2728 			ret = update_gpuvm_pte(mem, attachment, &sync);
2729 			if (ret) {
2730 				pr_err("%s: update PTE failed\n", __func__);
2731 				/* make sure this gets validated again */
2732 				mutex_lock(&process_info->notifier_lock);
2733 				mem->invalid++;
2734 				mutex_unlock(&process_info->notifier_lock);
2735 				goto unreserve_out;
2736 			}
2737 		}
2738 	}
2739 
2740 	/* Update page directories */
2741 	ret = process_update_pds(process_info, &sync);
2742 
2743 unreserve_out:
2744 	drm_exec_fini(&exec);
2745 	amdgpu_sync_wait(&sync, false);
2746 	amdgpu_sync_free(&sync);
2747 
2748 	return ret;
2749 }
2750 
2751 /* Confirm that all user pages are valid while holding the notifier lock
2752  *
2753  * Moves valid BOs from the userptr_inval_list back to userptr_val_list.
2754  */
2755 static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_info)
2756 {
2757 	struct kgd_mem *mem, *tmp_mem;
2758 	int ret = 0;
2759 
2760 	list_for_each_entry_safe(mem, tmp_mem,
2761 				 &process_info->userptr_inval_list,
2762 				 validate_list) {
2763 		bool valid;
2764 
2765 		/* keep mem without hmm range at userptr_inval_list */
2766 		if (!mem->range)
2767 			continue;
2768 
2769 		/* Only check mem with hmm range associated */
2770 		valid = amdgpu_hmm_range_valid(mem->range);
2771 		amdgpu_hmm_range_free(mem->range);
2772 
2773 		mem->range = NULL;
2774 		if (!valid) {
2775 			WARN(!mem->invalid, "Invalid BO not marked invalid");
2776 			ret = -EAGAIN;
2777 			continue;
2778 		}
2779 
2780 		if (mem->invalid) {
2781 			WARN(1, "Valid BO is marked invalid");
2782 			ret = -EAGAIN;
2783 			continue;
2784 		}
2785 
2786 		list_move_tail(&mem->validate_list,
2787 			       &process_info->userptr_valid_list);
2788 	}
2789 
2790 	return ret;
2791 }
2792 
2793 /* Worker callback to restore evicted userptr BOs
2794  *
2795  * Tries to update and validate all userptr BOs. If successful and no
2796  * concurrent evictions happened, the queues are restarted. Otherwise,
2797  * reschedule for another attempt later.
2798  */
2799 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
2800 {
2801 	struct delayed_work *dwork = to_delayed_work(work);
2802 	struct amdkfd_process_info *process_info =
2803 		container_of(dwork, struct amdkfd_process_info,
2804 			     restore_userptr_work);
2805 	struct task_struct *usertask;
2806 	struct mm_struct *mm;
2807 	uint32_t evicted_bos;
2808 
2809 	mutex_lock(&process_info->notifier_lock);
2810 	evicted_bos = process_info->evicted_bos;
2811 	mutex_unlock(&process_info->notifier_lock);
2812 	if (!evicted_bos)
2813 		return;
2814 
2815 	/* Reference task and mm in case of concurrent process termination */
2816 	usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
2817 	if (!usertask)
2818 		return;
2819 	mm = get_task_mm(usertask);
2820 	if (!mm) {
2821 		put_task_struct(usertask);
2822 		return;
2823 	}
2824 
2825 	mutex_lock(&process_info->lock);
2826 
2827 	if (update_invalid_user_pages(process_info, mm))
2828 		goto unlock_out;
2829 	/* userptr_inval_list can be empty if all evicted userptr BOs
2830 	 * have been freed. In that case there is nothing to validate
2831 	 * and we can just restart the queues.
2832 	 */
2833 	if (!list_empty(&process_info->userptr_inval_list)) {
2834 		if (validate_invalid_user_pages(process_info))
2835 			goto unlock_out;
2836 	}
2837 	/* Final check for concurrent evicton and atomic update. If
2838 	 * another eviction happens after successful update, it will
2839 	 * be a first eviction that calls quiesce_mm. The eviction
2840 	 * reference counting inside KFD will handle this case.
2841 	 */
2842 	mutex_lock(&process_info->notifier_lock);
2843 	if (process_info->evicted_bos != evicted_bos)
2844 		goto unlock_notifier_out;
2845 
2846 	if (confirm_valid_user_pages_locked(process_info)) {
2847 		WARN(1, "User pages unexpectedly invalid");
2848 		goto unlock_notifier_out;
2849 	}
2850 
2851 	process_info->evicted_bos = evicted_bos = 0;
2852 
2853 	if (kgd2kfd_resume_mm(mm)) {
2854 		pr_err("%s: Failed to resume KFD\n", __func__);
2855 		/* No recovery from this failure. Probably the CP is
2856 		 * hanging. No point trying again.
2857 		 */
2858 	}
2859 
2860 unlock_notifier_out:
2861 	mutex_unlock(&process_info->notifier_lock);
2862 unlock_out:
2863 	mutex_unlock(&process_info->lock);
2864 
2865 	/* If validation failed, reschedule another attempt */
2866 	if (evicted_bos) {
2867 		queue_delayed_work(system_freezable_wq,
2868 			&process_info->restore_userptr_work,
2869 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2870 
2871 		kfd_smi_event_queue_restore_rescheduled(mm);
2872 	}
2873 	mmput(mm);
2874 	put_task_struct(usertask);
2875 }
2876 
2877 static void replace_eviction_fence(struct dma_fence __rcu **ef,
2878 				   struct dma_fence *new_ef)
2879 {
2880 	struct dma_fence *old_ef = rcu_replace_pointer(*ef, new_ef, true
2881 		/* protected by process_info->lock */);
2882 
2883 	/* If we're replacing an unsignaled eviction fence, that fence will
2884 	 * never be signaled, and if anyone is still waiting on that fence,
2885 	 * they will hang forever. This should never happen. We should only
2886 	 * replace the fence in restore_work that only gets scheduled after
2887 	 * eviction work signaled the fence.
2888 	 */
2889 	WARN_ONCE(!dma_fence_is_signaled(old_ef),
2890 		  "Replacing unsignaled eviction fence");
2891 	dma_fence_put(old_ef);
2892 }
2893 
2894 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2895  *   KFD process identified by process_info
2896  *
2897  * @process_info: amdkfd_process_info of the KFD process
2898  *
2899  * After memory eviction, restore thread calls this function. The function
2900  * should be called when the Process is still valid. BO restore involves -
2901  *
2902  * 1.  Release old eviction fence and create new one
2903  * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2904  * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2905  *     BOs that need to be reserved.
2906  * 4.  Reserve all the BOs
2907  * 5.  Validate of PD and PT BOs.
2908  * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2909  * 7.  Add fence to all PD and PT BOs.
2910  * 8.  Unreserve all BOs
2911  */
2912 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu **ef)
2913 {
2914 	struct amdkfd_process_info *process_info = info;
2915 	struct amdgpu_vm *peer_vm;
2916 	struct kgd_mem *mem;
2917 	struct list_head duplicate_save;
2918 	struct amdgpu_sync sync_obj;
2919 	unsigned long failed_size = 0;
2920 	unsigned long total_size = 0;
2921 	struct drm_exec exec;
2922 	int ret;
2923 
2924 	INIT_LIST_HEAD(&duplicate_save);
2925 
2926 	mutex_lock(&process_info->lock);
2927 
2928 	drm_exec_init(&exec, DRM_EXEC_IGNORE_DUPLICATES, 0);
2929 	drm_exec_until_all_locked(&exec) {
2930 		list_for_each_entry(peer_vm, &process_info->vm_list_head,
2931 				    vm_list_node) {
2932 			ret = amdgpu_vm_lock_pd(peer_vm, &exec, 2);
2933 			drm_exec_retry_on_contention(&exec);
2934 			if (unlikely(ret)) {
2935 				pr_err("Locking VM PD failed, ret: %d\n", ret);
2936 				goto ttm_reserve_fail;
2937 			}
2938 		}
2939 
2940 		/* Reserve all BOs and page tables/directory. Add all BOs from
2941 		 * kfd_bo_list to ctx.list
2942 		 */
2943 		list_for_each_entry(mem, &process_info->kfd_bo_list,
2944 				    validate_list) {
2945 			struct drm_gem_object *gobj;
2946 
2947 			gobj = &mem->bo->tbo.base;
2948 			ret = drm_exec_prepare_obj(&exec, gobj, 1);
2949 			drm_exec_retry_on_contention(&exec);
2950 			if (unlikely(ret)) {
2951 				pr_err("drm_exec_prepare_obj failed, ret: %d\n", ret);
2952 				goto ttm_reserve_fail;
2953 			}
2954 		}
2955 	}
2956 
2957 	amdgpu_sync_create(&sync_obj);
2958 
2959 	/* Validate BOs managed by KFD */
2960 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2961 			    validate_list) {
2962 
2963 		struct amdgpu_bo *bo = mem->bo;
2964 		uint32_t domain = mem->domain;
2965 		struct dma_resv_iter cursor;
2966 		struct dma_fence *fence;
2967 
2968 		total_size += amdgpu_bo_size(bo);
2969 
2970 		ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2971 		if (ret) {
2972 			pr_debug("Memory eviction: Validate BOs failed\n");
2973 			failed_size += amdgpu_bo_size(bo);
2974 			ret = amdgpu_amdkfd_bo_validate(bo,
2975 						AMDGPU_GEM_DOMAIN_GTT, false);
2976 			if (ret) {
2977 				pr_debug("Memory eviction: Try again\n");
2978 				goto validate_map_fail;
2979 			}
2980 		}
2981 		dma_resv_for_each_fence(&cursor, bo->tbo.base.resv,
2982 					DMA_RESV_USAGE_KERNEL, fence) {
2983 			ret = amdgpu_sync_fence(&sync_obj, fence, GFP_KERNEL);
2984 			if (ret) {
2985 				pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2986 				goto validate_map_fail;
2987 			}
2988 		}
2989 	}
2990 
2991 	if (failed_size)
2992 		pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2993 
2994 	/* Validate PDs, PTs and evicted DMABuf imports last. Otherwise BO
2995 	 * validations above would invalidate DMABuf imports again.
2996 	 */
2997 	ret = process_validate_vms(process_info, &exec.ticket);
2998 	if (ret) {
2999 		pr_debug("Validating VMs failed, ret: %d\n", ret);
3000 		goto validate_map_fail;
3001 	}
3002 
3003 	/* Update mappings managed by KFD. */
3004 	list_for_each_entry(mem, &process_info->kfd_bo_list,
3005 			    validate_list) {
3006 		struct kfd_mem_attachment *attachment;
3007 
3008 		list_for_each_entry(attachment, &mem->attachments, list) {
3009 			if (!attachment->is_mapped)
3010 				continue;
3011 
3012 			kfd_mem_dmaunmap_attachment(mem, attachment);
3013 			ret = update_gpuvm_pte(mem, attachment, &sync_obj);
3014 			if (ret) {
3015 				pr_debug("Memory eviction: update PTE failed. Try again\n");
3016 				goto validate_map_fail;
3017 			}
3018 		}
3019 	}
3020 
3021 	/* Update mappings not managed by KFD */
3022 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
3023 			vm_list_node) {
3024 		struct amdgpu_device *adev = amdgpu_ttm_adev(
3025 			peer_vm->root.bo->tbo.bdev);
3026 
3027 		struct amdgpu_fpriv *fpriv =
3028 			container_of(peer_vm, struct amdgpu_fpriv, vm);
3029 
3030 		ret = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
3031 		if (ret) {
3032 			dev_dbg(adev->dev,
3033 				"Memory eviction: handle PRT moved failed, pid %8d. Try again.\n",
3034 				pid_nr(process_info->pid));
3035 			goto validate_map_fail;
3036 		}
3037 
3038 		ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket);
3039 		if (ret) {
3040 			dev_dbg(adev->dev,
3041 				"Memory eviction: handle moved failed, pid %8d. Try again.\n",
3042 				pid_nr(process_info->pid));
3043 			goto validate_map_fail;
3044 		}
3045 	}
3046 
3047 	/* Update page directories */
3048 	ret = process_update_pds(process_info, &sync_obj);
3049 	if (ret) {
3050 		pr_debug("Memory eviction: update PDs failed. Try again\n");
3051 		goto validate_map_fail;
3052 	}
3053 
3054 	/* Sync with fences on all the page tables. They implicitly depend on any
3055 	 * move fences from amdgpu_vm_handle_moved above.
3056 	 */
3057 	ret = process_sync_pds_resv(process_info, &sync_obj);
3058 	if (ret) {
3059 		pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
3060 		goto validate_map_fail;
3061 	}
3062 
3063 	/* Wait for validate and PT updates to finish */
3064 	amdgpu_sync_wait(&sync_obj, false);
3065 
3066 	/* The old eviction fence may be unsignaled if restore happens
3067 	 * after a GPU reset or suspend/resume. Keep the old fence in that
3068 	 * case. Otherwise release the old eviction fence and create new
3069 	 * one, because fence only goes from unsignaled to signaled once
3070 	 * and cannot be reused. Use context and mm from the old fence.
3071 	 *
3072 	 * If an old eviction fence signals after this check, that's OK.
3073 	 * Anyone signaling an eviction fence must stop the queues first
3074 	 * and schedule another restore worker.
3075 	 */
3076 	if (dma_fence_is_signaled(&process_info->eviction_fence->base)) {
3077 		struct amdgpu_amdkfd_fence *new_fence =
3078 			amdgpu_amdkfd_fence_create(
3079 				process_info->eviction_fence->base.context,
3080 				process_info->eviction_fence->mm,
3081 				NULL, process_info->context_id);
3082 
3083 		if (!new_fence) {
3084 			pr_err("Failed to create eviction fence\n");
3085 			ret = -ENOMEM;
3086 			goto validate_map_fail;
3087 		}
3088 		dma_fence_put(&process_info->eviction_fence->base);
3089 		process_info->eviction_fence = new_fence;
3090 		replace_eviction_fence(ef, dma_fence_get(&new_fence->base));
3091 	} else {
3092 		WARN_ONCE(*ef != &process_info->eviction_fence->base,
3093 			  "KFD eviction fence doesn't match KGD process_info");
3094 	}
3095 
3096 	/* Attach new eviction fence to all BOs except pinned ones */
3097 	list_for_each_entry(mem, &process_info->kfd_bo_list, validate_list) {
3098 		if (mem->bo->tbo.pin_count)
3099 			continue;
3100 
3101 		dma_resv_add_fence(mem->bo->tbo.base.resv,
3102 				   &process_info->eviction_fence->base,
3103 				   DMA_RESV_USAGE_BOOKKEEP);
3104 	}
3105 	/* Attach eviction fence to PD / PT BOs and DMABuf imports */
3106 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
3107 			    vm_list_node) {
3108 		struct amdgpu_bo *bo = peer_vm->root.bo;
3109 
3110 		dma_resv_add_fence(bo->tbo.base.resv,
3111 				   &process_info->eviction_fence->base,
3112 				   DMA_RESV_USAGE_BOOKKEEP);
3113 	}
3114 
3115 validate_map_fail:
3116 	amdgpu_sync_free(&sync_obj);
3117 ttm_reserve_fail:
3118 	drm_exec_fini(&exec);
3119 	mutex_unlock(&process_info->lock);
3120 	return ret;
3121 }
3122 
3123 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
3124 {
3125 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
3126 	struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
3127 	int ret;
3128 
3129 	if (!info || !gws)
3130 		return -EINVAL;
3131 
3132 	*mem = kzalloc_obj(struct kgd_mem, GFP_KERNEL);
3133 	if (!*mem)
3134 		return -ENOMEM;
3135 
3136 	mutex_init(&(*mem)->lock);
3137 	INIT_LIST_HEAD(&(*mem)->attachments);
3138 	(*mem)->bo = amdgpu_bo_ref(gws_bo);
3139 	(*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
3140 	(*mem)->process_info = process_info;
3141 	add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
3142 	amdgpu_sync_create(&(*mem)->sync);
3143 
3144 
3145 	/* Validate gws bo the first time it is added to process */
3146 	mutex_lock(&(*mem)->process_info->lock);
3147 	ret = amdgpu_bo_reserve(gws_bo, false);
3148 	if (unlikely(ret)) {
3149 		pr_err("Reserve gws bo failed %d\n", ret);
3150 		goto bo_reservation_failure;
3151 	}
3152 
3153 	ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
3154 	if (ret) {
3155 		pr_err("GWS BO validate failed %d\n", ret);
3156 		goto bo_validation_failure;
3157 	}
3158 	/* GWS resource is shared b/t amdgpu and amdkfd
3159 	 * Add process eviction fence to bo so they can
3160 	 * evict each other.
3161 	 */
3162 	ret = dma_resv_reserve_fences(gws_bo->tbo.base.resv, 1);
3163 	if (ret)
3164 		goto reserve_shared_fail;
3165 	dma_resv_add_fence(gws_bo->tbo.base.resv,
3166 			   &process_info->eviction_fence->base,
3167 			   DMA_RESV_USAGE_BOOKKEEP);
3168 	amdgpu_bo_unreserve(gws_bo);
3169 	mutex_unlock(&(*mem)->process_info->lock);
3170 
3171 	return ret;
3172 
3173 reserve_shared_fail:
3174 bo_validation_failure:
3175 	amdgpu_bo_unreserve(gws_bo);
3176 bo_reservation_failure:
3177 	mutex_unlock(&(*mem)->process_info->lock);
3178 	amdgpu_sync_free(&(*mem)->sync);
3179 	remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
3180 	amdgpu_bo_unref(&gws_bo);
3181 	mutex_destroy(&(*mem)->lock);
3182 	kfree(*mem);
3183 	*mem = NULL;
3184 	return ret;
3185 }
3186 
3187 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
3188 {
3189 	int ret;
3190 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
3191 	struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
3192 	struct amdgpu_bo *gws_bo = kgd_mem->bo;
3193 
3194 	/* Remove BO from process's validate list so restore worker won't touch
3195 	 * it anymore
3196 	 */
3197 	remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
3198 
3199 	ret = amdgpu_bo_reserve(gws_bo, false);
3200 	if (unlikely(ret)) {
3201 		pr_err("Reserve gws bo failed %d\n", ret);
3202 		//TODO add BO back to validate_list?
3203 		return ret;
3204 	}
3205 	amdgpu_amdkfd_remove_eviction_fence(gws_bo,
3206 			process_info->eviction_fence);
3207 	amdgpu_bo_unreserve(gws_bo);
3208 	amdgpu_sync_free(&kgd_mem->sync);
3209 	amdgpu_bo_unref(&gws_bo);
3210 	mutex_destroy(&kgd_mem->lock);
3211 	kfree(mem);
3212 	return 0;
3213 }
3214 
3215 /* Returns GPU-specific tiling mode information */
3216 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
3217 				struct tile_config *config)
3218 {
3219 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
3220 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
3221 	config->num_tile_configs =
3222 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
3223 	config->macro_tile_config_ptr =
3224 			adev->gfx.config.macrotile_mode_array;
3225 	config->num_macro_tile_configs =
3226 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
3227 
3228 	/* Those values are not set from GFX9 onwards */
3229 	config->num_banks = adev->gfx.config.num_banks;
3230 	config->num_ranks = adev->gfx.config.num_ranks;
3231 
3232 	return 0;
3233 }
3234 
3235 bool amdgpu_amdkfd_bo_mapped_to_dev(void *drm_priv, struct kgd_mem *mem)
3236 {
3237 	struct amdgpu_vm *vm = drm_priv_to_vm(drm_priv);
3238 	struct kfd_mem_attachment *entry;
3239 
3240 	list_for_each_entry(entry, &mem->attachments, list) {
3241 		if (entry->is_mapped && entry->bo_va->base.vm == vm)
3242 			return true;
3243 	}
3244 	return false;
3245 }
3246 
3247 #if defined(CONFIG_DEBUG_FS)
3248 
3249 int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data)
3250 {
3251 
3252 	spin_lock(&kfd_mem_limit.mem_limit_lock);
3253 	seq_printf(m, "System mem used %lldM out of %lluM\n",
3254 		  (kfd_mem_limit.system_mem_used >> 20),
3255 		  (kfd_mem_limit.max_system_mem_limit >> 20));
3256 	seq_printf(m, "TTM mem used %lldM out of %lluM\n",
3257 		  (kfd_mem_limit.ttm_mem_used >> 20),
3258 		  (kfd_mem_limit.max_ttm_mem_limit >> 20));
3259 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
3260 
3261 	return 0;
3262 }
3263 
3264 #endif
3265