xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c (revision 1d8355ad922423c9f765a644ed04526a6273d9ee)
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 "amdgpu_object.h"
31 #include "amdgpu_gem.h"
32 #include "amdgpu_vm.h"
33 #include "amdgpu_hmm.h"
34 #include "amdgpu_amdkfd.h"
35 #include "amdgpu_dma_buf.h"
36 #include <uapi/linux/kfd_ioctl.h>
37 #include "amdgpu_xgmi.h"
38 #include "kfd_smi_events.h"
39 #include <drm/ttm/ttm_tt.h>
40 
41 /* Userptr restore delay, just long enough to allow consecutive VM
42  * changes to accumulate
43  */
44 #define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
45 
46 /*
47  * Align VRAM availability to 2MB to avoid fragmentation caused by 4K allocations in the tail 2MB
48  * BO chunk
49  */
50 #define VRAM_AVAILABLITY_ALIGN (1 << 21)
51 
52 /* Impose limit on how much memory KFD can use */
53 static struct {
54 	uint64_t max_system_mem_limit;
55 	uint64_t max_ttm_mem_limit;
56 	int64_t system_mem_used;
57 	int64_t ttm_mem_used;
58 	spinlock_t mem_limit_lock;
59 } kfd_mem_limit;
60 
61 static const char * const domain_bit_to_string[] = {
62 		"CPU",
63 		"GTT",
64 		"VRAM",
65 		"GDS",
66 		"GWS",
67 		"OA"
68 };
69 
70 #define domain_string(domain) domain_bit_to_string[ffs(domain)-1]
71 
72 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work);
73 
74 static bool kfd_mem_is_attached(struct amdgpu_vm *avm,
75 		struct kgd_mem *mem)
76 {
77 	struct kfd_mem_attachment *entry;
78 
79 	list_for_each_entry(entry, &mem->attachments, list)
80 		if (entry->bo_va->base.vm == avm)
81 			return true;
82 
83 	return false;
84 }
85 
86 /**
87  * reuse_dmamap() - Check whether adev can share the original
88  * userptr BO
89  *
90  * If both adev and bo_adev are in direct mapping or
91  * in the same iommu group, they can share the original BO.
92  *
93  * @adev: Device to which can or cannot share the original BO
94  * @bo_adev: Device to which allocated BO belongs to
95  *
96  * Return: returns true if adev can share original userptr BO,
97  * false otherwise.
98  */
99 static bool reuse_dmamap(struct amdgpu_device *adev, struct amdgpu_device *bo_adev)
100 {
101 	return (adev->ram_is_direct_mapped && bo_adev->ram_is_direct_mapped) ||
102 			(adev->dev->iommu_group == bo_adev->dev->iommu_group);
103 }
104 
105 /* Set memory usage limits. Current, limits are
106  *  System (TTM + userptr) memory - 15/16th System RAM
107  *  TTM memory - 3/8th System RAM
108  */
109 void amdgpu_amdkfd_gpuvm_init_mem_limits(void)
110 {
111 	struct sysinfo si;
112 	uint64_t mem;
113 
114 	if (kfd_mem_limit.max_system_mem_limit)
115 		return;
116 
117 	si_meminfo(&si);
118 	mem = si.freeram - si.freehigh;
119 	mem *= si.mem_unit;
120 
121 	spin_lock_init(&kfd_mem_limit.mem_limit_lock);
122 	kfd_mem_limit.max_system_mem_limit = mem - (mem >> 4);
123 	kfd_mem_limit.max_ttm_mem_limit = ttm_tt_pages_limit() << PAGE_SHIFT;
124 	pr_debug("Kernel memory limit %lluM, TTM limit %lluM\n",
125 		(kfd_mem_limit.max_system_mem_limit >> 20),
126 		(kfd_mem_limit.max_ttm_mem_limit >> 20));
127 }
128 
129 void amdgpu_amdkfd_reserve_system_mem(uint64_t size)
130 {
131 	kfd_mem_limit.system_mem_used += size;
132 }
133 
134 /* Estimate page table size needed to represent a given memory size
135  *
136  * With 4KB pages, we need one 8 byte PTE for each 4KB of memory
137  * (factor 512, >> 9). With 2MB pages, we need one 8 byte PTE for 2MB
138  * of memory (factor 256K, >> 18). ROCm user mode tries to optimize
139  * for 2MB pages for TLB efficiency. However, small allocations and
140  * fragmented system memory still need some 4KB pages. We choose a
141  * compromise that should work in most cases without reserving too
142  * much memory for page tables unnecessarily (factor 16K, >> 14).
143  */
144 
145 #define ESTIMATE_PT_SIZE(mem_size) max(((mem_size) >> 14), AMDGPU_VM_RESERVED_VRAM)
146 
147 /**
148  * amdgpu_amdkfd_reserve_mem_limit() - Decrease available memory by size
149  * of buffer.
150  *
151  * @adev: Device to which allocated BO belongs to
152  * @size: Size of buffer, in bytes, encapsulated by B0. This should be
153  * equivalent to amdgpu_bo_size(BO)
154  * @alloc_flag: Flag used in allocating a BO as noted above
155  *
156  * Return: returns -ENOMEM in case of error, ZERO otherwise
157  */
158 int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
159 		uint64_t size, u32 alloc_flag)
160 {
161 	uint64_t reserved_for_pt =
162 		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
163 	size_t system_mem_needed, ttm_mem_needed, vram_needed;
164 	int ret = 0;
165 
166 	system_mem_needed = 0;
167 	ttm_mem_needed = 0;
168 	vram_needed = 0;
169 	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
170 		system_mem_needed = size;
171 		ttm_mem_needed = size;
172 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
173 		/*
174 		 * Conservatively round up the allocation requirement to 2 MB
175 		 * to avoid fragmentation caused by 4K allocations in the tail
176 		 * 2M BO chunk.
177 		 */
178 		vram_needed = size;
179 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
180 		system_mem_needed = size;
181 	} else if (!(alloc_flag &
182 				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
183 				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
184 		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
185 		return -ENOMEM;
186 	}
187 
188 	spin_lock(&kfd_mem_limit.mem_limit_lock);
189 
190 	if (kfd_mem_limit.system_mem_used + system_mem_needed >
191 	    kfd_mem_limit.max_system_mem_limit)
192 		pr_debug("Set no_system_mem_limit=1 if using shared memory\n");
193 
194 	if ((kfd_mem_limit.system_mem_used + system_mem_needed >
195 	     kfd_mem_limit.max_system_mem_limit && !no_system_mem_limit) ||
196 	    (kfd_mem_limit.ttm_mem_used + ttm_mem_needed >
197 	     kfd_mem_limit.max_ttm_mem_limit) ||
198 	    (adev && adev->kfd.vram_used + vram_needed >
199 	     adev->gmc.real_vram_size - reserved_for_pt)) {
200 		ret = -ENOMEM;
201 		goto release;
202 	}
203 
204 	/* Update memory accounting by decreasing available system
205 	 * memory, TTM memory and GPU memory as computed above
206 	 */
207 	WARN_ONCE(vram_needed && !adev,
208 		  "adev reference can't be null when vram is used");
209 	if (adev) {
210 		adev->kfd.vram_used += vram_needed;
211 		adev->kfd.vram_used_aligned += ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
212 	}
213 	kfd_mem_limit.system_mem_used += system_mem_needed;
214 	kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
215 
216 release:
217 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
218 	return ret;
219 }
220 
221 void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev,
222 		uint64_t size, u32 alloc_flag)
223 {
224 	spin_lock(&kfd_mem_limit.mem_limit_lock);
225 
226 	if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
227 		kfd_mem_limit.system_mem_used -= size;
228 		kfd_mem_limit.ttm_mem_used -= size;
229 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
230 		WARN_ONCE(!adev,
231 			  "adev reference can't be null when alloc mem flags vram is set");
232 		if (adev) {
233 			adev->kfd.vram_used -= size;
234 			adev->kfd.vram_used_aligned -= ALIGN(size, VRAM_AVAILABLITY_ALIGN);
235 		}
236 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
237 		kfd_mem_limit.system_mem_used -= size;
238 	} else if (!(alloc_flag &
239 				(KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
240 				 KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
241 		pr_err("%s: Invalid BO type %#x\n", __func__, alloc_flag);
242 		goto release;
243 	}
244 	WARN_ONCE(adev && adev->kfd.vram_used < 0,
245 		  "KFD VRAM memory accounting unbalanced");
246 	WARN_ONCE(kfd_mem_limit.ttm_mem_used < 0,
247 		  "KFD TTM memory accounting unbalanced");
248 	WARN_ONCE(kfd_mem_limit.system_mem_used < 0,
249 		  "KFD system memory accounting unbalanced");
250 
251 release:
252 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
253 }
254 
255 void amdgpu_amdkfd_release_notify(struct amdgpu_bo *bo)
256 {
257 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
258 	u32 alloc_flags = bo->kfd_bo->alloc_flags;
259 	u64 size = amdgpu_bo_size(bo);
260 
261 	amdgpu_amdkfd_unreserve_mem_limit(adev, size, alloc_flags);
262 
263 	kfree(bo->kfd_bo);
264 }
265 
266 /**
267  * @create_dmamap_sg_bo: Creates a amdgpu_bo object to reflect information
268  * about USERPTR or DOOREBELL or MMIO BO.
269  * @adev: Device for which dmamap BO is being created
270  * @mem: BO of peer device that is being DMA mapped. Provides parameters
271  *	 in building the dmamap BO
272  * @bo_out: Output parameter updated with handle of dmamap BO
273  */
274 static int
275 create_dmamap_sg_bo(struct amdgpu_device *adev,
276 		 struct kgd_mem *mem, struct amdgpu_bo **bo_out)
277 {
278 	struct drm_gem_object *gem_obj;
279 	int ret;
280 	uint64_t flags = 0;
281 
282 	ret = amdgpu_bo_reserve(mem->bo, false);
283 	if (ret)
284 		return ret;
285 
286 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR)
287 		flags |= mem->bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
288 					AMDGPU_GEM_CREATE_UNCACHED);
289 
290 	ret = amdgpu_gem_object_create(adev, mem->bo->tbo.base.size, 1,
291 			AMDGPU_GEM_DOMAIN_CPU, AMDGPU_GEM_CREATE_PREEMPTIBLE | flags,
292 			ttm_bo_type_sg, mem->bo->tbo.base.resv, &gem_obj);
293 
294 	amdgpu_bo_unreserve(mem->bo);
295 
296 	if (ret) {
297 		pr_err("Error in creating DMA mappable SG BO on domain: %d\n", ret);
298 		return -EINVAL;
299 	}
300 
301 	*bo_out = gem_to_amdgpu_bo(gem_obj);
302 	(*bo_out)->parent = amdgpu_bo_ref(mem->bo);
303 	return ret;
304 }
305 
306 /* amdgpu_amdkfd_remove_eviction_fence - Removes eviction fence from BO's
307  *  reservation object.
308  *
309  * @bo: [IN] Remove eviction fence(s) from this BO
310  * @ef: [IN] This eviction fence is removed if it
311  *  is present in the shared list.
312  *
313  * NOTE: Must be called with BO reserved i.e. bo->tbo.resv->lock held.
314  */
315 static int amdgpu_amdkfd_remove_eviction_fence(struct amdgpu_bo *bo,
316 					struct amdgpu_amdkfd_fence *ef)
317 {
318 	struct dma_fence *replacement;
319 
320 	if (!ef)
321 		return -EINVAL;
322 
323 	/* TODO: Instead of block before we should use the fence of the page
324 	 * table update and TLB flush here directly.
325 	 */
326 	replacement = dma_fence_get_stub();
327 	dma_resv_replace_fences(bo->tbo.base.resv, ef->base.context,
328 				replacement, DMA_RESV_USAGE_BOOKKEEP);
329 	dma_fence_put(replacement);
330 	return 0;
331 }
332 
333 int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
334 {
335 	struct amdgpu_bo *root = bo;
336 	struct amdgpu_vm_bo_base *vm_bo;
337 	struct amdgpu_vm *vm;
338 	struct amdkfd_process_info *info;
339 	struct amdgpu_amdkfd_fence *ef;
340 	int ret;
341 
342 	/* we can always get vm_bo from root PD bo.*/
343 	while (root->parent)
344 		root = root->parent;
345 
346 	vm_bo = root->vm_bo;
347 	if (!vm_bo)
348 		return 0;
349 
350 	vm = vm_bo->vm;
351 	if (!vm)
352 		return 0;
353 
354 	info = vm->process_info;
355 	if (!info || !info->eviction_fence)
356 		return 0;
357 
358 	ef = container_of(dma_fence_get(&info->eviction_fence->base),
359 			struct amdgpu_amdkfd_fence, base);
360 
361 	BUG_ON(!dma_resv_trylock(bo->tbo.base.resv));
362 	ret = amdgpu_amdkfd_remove_eviction_fence(bo, ef);
363 	dma_resv_unlock(bo->tbo.base.resv);
364 
365 	dma_fence_put(&ef->base);
366 	return ret;
367 }
368 
369 static int amdgpu_amdkfd_bo_validate(struct amdgpu_bo *bo, uint32_t domain,
370 				     bool wait)
371 {
372 	struct ttm_operation_ctx ctx = { false, false };
373 	int ret;
374 
375 	if (WARN(amdgpu_ttm_tt_get_usermm(bo->tbo.ttm),
376 		 "Called with userptr BO"))
377 		return -EINVAL;
378 
379 	amdgpu_bo_placement_from_domain(bo, domain);
380 
381 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
382 	if (ret)
383 		goto validate_fail;
384 	if (wait)
385 		amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
386 
387 validate_fail:
388 	return ret;
389 }
390 
391 static int amdgpu_amdkfd_validate_vm_bo(void *_unused, struct amdgpu_bo *bo)
392 {
393 	return amdgpu_amdkfd_bo_validate(bo, bo->allowed_domains, false);
394 }
395 
396 /* vm_validate_pt_pd_bos - Validate page table and directory BOs
397  *
398  * Page directories are not updated here because huge page handling
399  * during page table updates can invalidate page directory entries
400  * again. Page directories are only updated after updating page
401  * tables.
402  */
403 static int vm_validate_pt_pd_bos(struct amdgpu_vm *vm)
404 {
405 	struct amdgpu_bo *pd = vm->root.bo;
406 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
407 	int ret;
408 
409 	ret = amdgpu_vm_validate_pt_bos(adev, vm, amdgpu_amdkfd_validate_vm_bo, NULL);
410 	if (ret) {
411 		pr_err("failed to validate PT BOs\n");
412 		return ret;
413 	}
414 
415 	vm->pd_phys_addr = amdgpu_gmc_pd_addr(vm->root.bo);
416 
417 	return 0;
418 }
419 
420 static int vm_update_pds(struct amdgpu_vm *vm, struct amdgpu_sync *sync)
421 {
422 	struct amdgpu_bo *pd = vm->root.bo;
423 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
424 	int ret;
425 
426 	ret = amdgpu_vm_update_pdes(adev, vm, false);
427 	if (ret)
428 		return ret;
429 
430 	return amdgpu_sync_fence(sync, vm->last_update);
431 }
432 
433 static uint64_t get_pte_flags(struct amdgpu_device *adev, struct kgd_mem *mem)
434 {
435 	uint32_t mapping_flags = AMDGPU_VM_PAGE_READABLE |
436 				 AMDGPU_VM_MTYPE_DEFAULT;
437 
438 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE)
439 		mapping_flags |= AMDGPU_VM_PAGE_WRITEABLE;
440 	if (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE)
441 		mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
442 
443 	return amdgpu_gem_va_map_flags(adev, mapping_flags);
444 }
445 
446 /**
447  * create_sg_table() - Create an sg_table for a contiguous DMA addr range
448  * @addr: The starting address to point to
449  * @size: Size of memory area in bytes being pointed to
450  *
451  * Allocates an instance of sg_table and initializes it to point to memory
452  * area specified by input parameters. The address used to build is assumed
453  * to be DMA mapped, if needed.
454  *
455  * DOORBELL or MMIO BOs use only one scatterlist node in their sg_table
456  * because they are physically contiguous.
457  *
458  * Return: Initialized instance of SG Table or NULL
459  */
460 static struct sg_table *create_sg_table(uint64_t addr, uint32_t size)
461 {
462 	struct sg_table *sg = kmalloc(sizeof(*sg), GFP_KERNEL);
463 
464 	if (!sg)
465 		return NULL;
466 	if (sg_alloc_table(sg, 1, GFP_KERNEL)) {
467 		kfree(sg);
468 		return NULL;
469 	}
470 	sg_dma_address(sg->sgl) = addr;
471 	sg->sgl->length = size;
472 #ifdef CONFIG_NEED_SG_DMA_LENGTH
473 	sg->sgl->dma_length = size;
474 #endif
475 	return sg;
476 }
477 
478 static int
479 kfd_mem_dmamap_userptr(struct kgd_mem *mem,
480 		       struct kfd_mem_attachment *attachment)
481 {
482 	enum dma_data_direction direction =
483 		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
484 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
485 	struct ttm_operation_ctx ctx = {.interruptible = true};
486 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
487 	struct amdgpu_device *adev = attachment->adev;
488 	struct ttm_tt *src_ttm = mem->bo->tbo.ttm;
489 	struct ttm_tt *ttm = bo->tbo.ttm;
490 	int ret;
491 
492 	if (WARN_ON(ttm->num_pages != src_ttm->num_pages))
493 		return -EINVAL;
494 
495 	ttm->sg = kmalloc(sizeof(*ttm->sg), GFP_KERNEL);
496 	if (unlikely(!ttm->sg))
497 		return -ENOMEM;
498 
499 	/* Same sequence as in amdgpu_ttm_tt_pin_userptr */
500 	ret = sg_alloc_table_from_pages(ttm->sg, src_ttm->pages,
501 					ttm->num_pages, 0,
502 					(u64)ttm->num_pages << PAGE_SHIFT,
503 					GFP_KERNEL);
504 	if (unlikely(ret))
505 		goto free_sg;
506 
507 	ret = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
508 	if (unlikely(ret))
509 		goto release_sg;
510 
511 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
512 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
513 	if (ret)
514 		goto unmap_sg;
515 
516 	return 0;
517 
518 unmap_sg:
519 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
520 release_sg:
521 	pr_err("DMA map userptr failed: %d\n", ret);
522 	sg_free_table(ttm->sg);
523 free_sg:
524 	kfree(ttm->sg);
525 	ttm->sg = NULL;
526 	return ret;
527 }
528 
529 static int
530 kfd_mem_dmamap_dmabuf(struct kfd_mem_attachment *attachment)
531 {
532 	struct ttm_operation_ctx ctx = {.interruptible = true};
533 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
534 	int ret;
535 
536 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
537 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
538 	if (ret)
539 		return ret;
540 
541 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
542 	return ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
543 }
544 
545 /**
546  * kfd_mem_dmamap_sg_bo() - Create DMA mapped sg_table to access DOORBELL or MMIO BO
547  * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
548  * @attachment: Virtual address attachment of the BO on accessing device
549  *
550  * An access request from the device that owns DOORBELL does not require DMA mapping.
551  * This is because the request doesn't go through PCIe root complex i.e. it instead
552  * loops back. The need to DMA map arises only when accessing peer device's DOORBELL
553  *
554  * In contrast, all access requests for MMIO need to be DMA mapped without regard to
555  * device ownership. This is because access requests for MMIO go through PCIe root
556  * complex.
557  *
558  * This is accomplished in two steps:
559  *   - Obtain DMA mapped address of DOORBELL or MMIO memory that could be used
560  *         in updating requesting device's page table
561  *   - Signal TTM to mark memory pointed to by requesting device's BO as GPU
562  *         accessible. This allows an update of requesting device's page table
563  *         with entries associated with DOOREBELL or MMIO memory
564  *
565  * This method is invoked in the following contexts:
566  *   - Mapping of DOORBELL or MMIO BO of same or peer device
567  *   - Validating an evicted DOOREBELL or MMIO BO on device seeking access
568  *
569  * Return: ZERO if successful, NON-ZERO otherwise
570  */
571 static int
572 kfd_mem_dmamap_sg_bo(struct kgd_mem *mem,
573 		     struct kfd_mem_attachment *attachment)
574 {
575 	struct ttm_operation_ctx ctx = {.interruptible = true};
576 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
577 	struct amdgpu_device *adev = attachment->adev;
578 	struct ttm_tt *ttm = bo->tbo.ttm;
579 	enum dma_data_direction dir;
580 	dma_addr_t dma_addr;
581 	bool mmio;
582 	int ret;
583 
584 	/* Expect SG Table of dmapmap BO to be NULL */
585 	mmio = (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP);
586 	if (unlikely(ttm->sg)) {
587 		pr_err("SG Table of %d BO for peer device is UNEXPECTEDLY NON-NULL", mmio);
588 		return -EINVAL;
589 	}
590 
591 	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
592 			DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
593 	dma_addr = mem->bo->tbo.sg->sgl->dma_address;
594 	pr_debug("%d BO size: %d\n", mmio, mem->bo->tbo.sg->sgl->length);
595 	pr_debug("%d BO address before DMA mapping: %llx\n", mmio, dma_addr);
596 	dma_addr = dma_map_resource(adev->dev, dma_addr,
597 			mem->bo->tbo.sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
598 	ret = dma_mapping_error(adev->dev, dma_addr);
599 	if (unlikely(ret))
600 		return ret;
601 	pr_debug("%d BO address after DMA mapping: %llx\n", mmio, dma_addr);
602 
603 	ttm->sg = create_sg_table(dma_addr, mem->bo->tbo.sg->sgl->length);
604 	if (unlikely(!ttm->sg)) {
605 		ret = -ENOMEM;
606 		goto unmap_sg;
607 	}
608 
609 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
610 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
611 	if (unlikely(ret))
612 		goto free_sg;
613 
614 	return ret;
615 
616 free_sg:
617 	sg_free_table(ttm->sg);
618 	kfree(ttm->sg);
619 	ttm->sg = NULL;
620 unmap_sg:
621 	dma_unmap_resource(adev->dev, dma_addr, mem->bo->tbo.sg->sgl->length,
622 			   dir, DMA_ATTR_SKIP_CPU_SYNC);
623 	return ret;
624 }
625 
626 static int
627 kfd_mem_dmamap_attachment(struct kgd_mem *mem,
628 			  struct kfd_mem_attachment *attachment)
629 {
630 	switch (attachment->type) {
631 	case KFD_MEM_ATT_SHARED:
632 		return 0;
633 	case KFD_MEM_ATT_USERPTR:
634 		return kfd_mem_dmamap_userptr(mem, attachment);
635 	case KFD_MEM_ATT_DMABUF:
636 		return kfd_mem_dmamap_dmabuf(attachment);
637 	case KFD_MEM_ATT_SG:
638 		return kfd_mem_dmamap_sg_bo(mem, attachment);
639 	default:
640 		WARN_ON_ONCE(1);
641 	}
642 	return -EINVAL;
643 }
644 
645 static void
646 kfd_mem_dmaunmap_userptr(struct kgd_mem *mem,
647 			 struct kfd_mem_attachment *attachment)
648 {
649 	enum dma_data_direction direction =
650 		mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
651 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
652 	struct ttm_operation_ctx ctx = {.interruptible = false};
653 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
654 	struct amdgpu_device *adev = attachment->adev;
655 	struct ttm_tt *ttm = bo->tbo.ttm;
656 
657 	if (unlikely(!ttm->sg))
658 		return;
659 
660 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
661 	ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
662 
663 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
664 	sg_free_table(ttm->sg);
665 	kfree(ttm->sg);
666 	ttm->sg = NULL;
667 }
668 
669 static void
670 kfd_mem_dmaunmap_dmabuf(struct kfd_mem_attachment *attachment)
671 {
672 	/* This is a no-op. We don't want to trigger eviction fences when
673 	 * unmapping DMABufs. Therefore the invalidation (moving to system
674 	 * domain) is done in kfd_mem_dmamap_dmabuf.
675 	 */
676 }
677 
678 /**
679  * kfd_mem_dmaunmap_sg_bo() - Free DMA mapped sg_table of DOORBELL or MMIO BO
680  * @mem: SG BO of the DOORBELL or MMIO resource on the owning device
681  * @attachment: Virtual address attachment of the BO on accessing device
682  *
683  * The method performs following steps:
684  *   - Signal TTM to mark memory pointed to by BO as GPU inaccessible
685  *   - Free SG Table that is used to encapsulate DMA mapped memory of
686  *          peer device's DOORBELL or MMIO memory
687  *
688  * This method is invoked in the following contexts:
689  *     UNMapping of DOORBELL or MMIO BO on a device having access to its memory
690  *     Eviction of DOOREBELL or MMIO BO on device having access to its memory
691  *
692  * Return: void
693  */
694 static void
695 kfd_mem_dmaunmap_sg_bo(struct kgd_mem *mem,
696 		       struct kfd_mem_attachment *attachment)
697 {
698 	struct ttm_operation_ctx ctx = {.interruptible = true};
699 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
700 	struct amdgpu_device *adev = attachment->adev;
701 	struct ttm_tt *ttm = bo->tbo.ttm;
702 	enum dma_data_direction dir;
703 
704 	if (unlikely(!ttm->sg)) {
705 		pr_err("SG Table of BO is UNEXPECTEDLY NULL");
706 		return;
707 	}
708 
709 	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
710 	ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
711 
712 	dir = mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
713 				DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
714 	dma_unmap_resource(adev->dev, ttm->sg->sgl->dma_address,
715 			ttm->sg->sgl->length, dir, DMA_ATTR_SKIP_CPU_SYNC);
716 	sg_free_table(ttm->sg);
717 	kfree(ttm->sg);
718 	ttm->sg = NULL;
719 	bo->tbo.sg = NULL;
720 }
721 
722 static void
723 kfd_mem_dmaunmap_attachment(struct kgd_mem *mem,
724 			    struct kfd_mem_attachment *attachment)
725 {
726 	switch (attachment->type) {
727 	case KFD_MEM_ATT_SHARED:
728 		break;
729 	case KFD_MEM_ATT_USERPTR:
730 		kfd_mem_dmaunmap_userptr(mem, attachment);
731 		break;
732 	case KFD_MEM_ATT_DMABUF:
733 		kfd_mem_dmaunmap_dmabuf(attachment);
734 		break;
735 	case KFD_MEM_ATT_SG:
736 		kfd_mem_dmaunmap_sg_bo(mem, attachment);
737 		break;
738 	default:
739 		WARN_ON_ONCE(1);
740 	}
741 }
742 
743 static int kfd_mem_export_dmabuf(struct kgd_mem *mem)
744 {
745 	if (!mem->dmabuf) {
746 		struct dma_buf *ret = amdgpu_gem_prime_export(
747 			&mem->bo->tbo.base,
748 			mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE ?
749 				DRM_RDWR : 0);
750 		if (IS_ERR(ret))
751 			return PTR_ERR(ret);
752 		mem->dmabuf = ret;
753 	}
754 
755 	return 0;
756 }
757 
758 static int
759 kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem,
760 		      struct amdgpu_bo **bo)
761 {
762 	struct drm_gem_object *gobj;
763 	int ret;
764 
765 	ret = kfd_mem_export_dmabuf(mem);
766 	if (ret)
767 		return ret;
768 
769 	gobj = amdgpu_gem_prime_import(adev_to_drm(adev), mem->dmabuf);
770 	if (IS_ERR(gobj))
771 		return PTR_ERR(gobj);
772 
773 	*bo = gem_to_amdgpu_bo(gobj);
774 	(*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE;
775 
776 	return 0;
777 }
778 
779 /* kfd_mem_attach - Add a BO to a VM
780  *
781  * Everything that needs to bo done only once when a BO is first added
782  * to a VM. It can later be mapped and unmapped many times without
783  * repeating these steps.
784  *
785  * 0. Create BO for DMA mapping, if needed
786  * 1. Allocate and initialize BO VA entry data structure
787  * 2. Add BO to the VM
788  * 3. Determine ASIC-specific PTE flags
789  * 4. Alloc page tables and directories if needed
790  * 4a.  Validate new page tables and directories
791  */
792 static int kfd_mem_attach(struct amdgpu_device *adev, struct kgd_mem *mem,
793 		struct amdgpu_vm *vm, bool is_aql)
794 {
795 	struct amdgpu_device *bo_adev = amdgpu_ttm_adev(mem->bo->tbo.bdev);
796 	unsigned long bo_size = mem->bo->tbo.base.size;
797 	uint64_t va = mem->va;
798 	struct kfd_mem_attachment *attachment[2] = {NULL, NULL};
799 	struct amdgpu_bo *bo[2] = {NULL, NULL};
800 	bool same_hive = false;
801 	int i, ret;
802 
803 	if (!va) {
804 		pr_err("Invalid VA when adding BO to VM\n");
805 		return -EINVAL;
806 	}
807 
808 	/* Determine access to VRAM, MMIO and DOORBELL BOs of peer devices
809 	 *
810 	 * The access path of MMIO and DOORBELL BOs of is always over PCIe.
811 	 * In contrast the access path of VRAM BOs depens upon the type of
812 	 * link that connects the peer device. Access over PCIe is allowed
813 	 * if peer device has large BAR. In contrast, access over xGMI is
814 	 * allowed for both small and large BAR configurations of peer device
815 	 */
816 	if ((adev != bo_adev) &&
817 	    ((mem->domain == AMDGPU_GEM_DOMAIN_VRAM) ||
818 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL) ||
819 	     (mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))) {
820 		if (mem->domain == AMDGPU_GEM_DOMAIN_VRAM)
821 			same_hive = amdgpu_xgmi_same_hive(adev, bo_adev);
822 		if (!same_hive && !amdgpu_device_is_peer_accessible(bo_adev, adev))
823 			return -EINVAL;
824 	}
825 
826 	for (i = 0; i <= is_aql; i++) {
827 		attachment[i] = kzalloc(sizeof(*attachment[i]), GFP_KERNEL);
828 		if (unlikely(!attachment[i])) {
829 			ret = -ENOMEM;
830 			goto unwind;
831 		}
832 
833 		pr_debug("\t add VA 0x%llx - 0x%llx to vm %p\n", va,
834 			 va + bo_size, vm);
835 
836 		if ((adev == bo_adev && !(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) ||
837 		    (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) && reuse_dmamap(adev, bo_adev)) ||
838 			same_hive) {
839 			/* Mappings on the local GPU, or VRAM mappings in the
840 			 * local hive, or userptr mapping can reuse dma map
841 			 * address space share the original BO
842 			 */
843 			attachment[i]->type = KFD_MEM_ATT_SHARED;
844 			bo[i] = mem->bo;
845 			drm_gem_object_get(&bo[i]->tbo.base);
846 		} else if (i > 0) {
847 			/* Multiple mappings on the same GPU share the BO */
848 			attachment[i]->type = KFD_MEM_ATT_SHARED;
849 			bo[i] = bo[0];
850 			drm_gem_object_get(&bo[i]->tbo.base);
851 		} else if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
852 			/* Create an SG BO to DMA-map userptrs on other GPUs */
853 			attachment[i]->type = KFD_MEM_ATT_USERPTR;
854 			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
855 			if (ret)
856 				goto unwind;
857 		/* Handle DOORBELL BOs of peer devices and MMIO BOs of local and peer devices */
858 		} else if (mem->bo->tbo.type == ttm_bo_type_sg) {
859 			WARN_ONCE(!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL ||
860 				    mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP),
861 				  "Handing invalid SG BO in ATTACH request");
862 			attachment[i]->type = KFD_MEM_ATT_SG;
863 			ret = create_dmamap_sg_bo(adev, mem, &bo[i]);
864 			if (ret)
865 				goto unwind;
866 		/* Enable acces to GTT and VRAM BOs of peer devices */
867 		} else if (mem->domain == AMDGPU_GEM_DOMAIN_GTT ||
868 			   mem->domain == AMDGPU_GEM_DOMAIN_VRAM) {
869 			attachment[i]->type = KFD_MEM_ATT_DMABUF;
870 			ret = kfd_mem_attach_dmabuf(adev, mem, &bo[i]);
871 			if (ret)
872 				goto unwind;
873 			pr_debug("Employ DMABUF mechanism to enable peer GPU access\n");
874 		} else {
875 			WARN_ONCE(true, "Handling invalid ATTACH request");
876 			ret = -EINVAL;
877 			goto unwind;
878 		}
879 
880 		/* Add BO to VM internal data structures */
881 		ret = amdgpu_bo_reserve(bo[i], false);
882 		if (ret) {
883 			pr_debug("Unable to reserve BO during memory attach");
884 			goto unwind;
885 		}
886 		attachment[i]->bo_va = amdgpu_vm_bo_add(adev, vm, bo[i]);
887 		amdgpu_bo_unreserve(bo[i]);
888 		if (unlikely(!attachment[i]->bo_va)) {
889 			ret = -ENOMEM;
890 			pr_err("Failed to add BO object to VM. ret == %d\n",
891 			       ret);
892 			goto unwind;
893 		}
894 		attachment[i]->va = va;
895 		attachment[i]->pte_flags = get_pte_flags(adev, mem);
896 		attachment[i]->adev = adev;
897 		list_add(&attachment[i]->list, &mem->attachments);
898 
899 		va += bo_size;
900 	}
901 
902 	return 0;
903 
904 unwind:
905 	for (; i >= 0; i--) {
906 		if (!attachment[i])
907 			continue;
908 		if (attachment[i]->bo_va) {
909 			amdgpu_bo_reserve(bo[i], true);
910 			amdgpu_vm_bo_del(adev, attachment[i]->bo_va);
911 			amdgpu_bo_unreserve(bo[i]);
912 			list_del(&attachment[i]->list);
913 		}
914 		if (bo[i])
915 			drm_gem_object_put(&bo[i]->tbo.base);
916 		kfree(attachment[i]);
917 	}
918 	return ret;
919 }
920 
921 static void kfd_mem_detach(struct kfd_mem_attachment *attachment)
922 {
923 	struct amdgpu_bo *bo = attachment->bo_va->base.bo;
924 
925 	pr_debug("\t remove VA 0x%llx in entry %p\n",
926 			attachment->va, attachment);
927 	amdgpu_vm_bo_del(attachment->adev, attachment->bo_va);
928 	drm_gem_object_put(&bo->tbo.base);
929 	list_del(&attachment->list);
930 	kfree(attachment);
931 }
932 
933 static void add_kgd_mem_to_kfd_bo_list(struct kgd_mem *mem,
934 				struct amdkfd_process_info *process_info,
935 				bool userptr)
936 {
937 	struct ttm_validate_buffer *entry = &mem->validate_list;
938 	struct amdgpu_bo *bo = mem->bo;
939 
940 	INIT_LIST_HEAD(&entry->head);
941 	entry->num_shared = 1;
942 	entry->bo = &bo->tbo;
943 	mutex_lock(&process_info->lock);
944 	if (userptr)
945 		list_add_tail(&entry->head, &process_info->userptr_valid_list);
946 	else
947 		list_add_tail(&entry->head, &process_info->kfd_bo_list);
948 	mutex_unlock(&process_info->lock);
949 }
950 
951 static void remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
952 		struct amdkfd_process_info *process_info)
953 {
954 	struct ttm_validate_buffer *bo_list_entry;
955 
956 	bo_list_entry = &mem->validate_list;
957 	mutex_lock(&process_info->lock);
958 	list_del(&bo_list_entry->head);
959 	mutex_unlock(&process_info->lock);
960 }
961 
962 /* Initializes user pages. It registers the MMU notifier and validates
963  * the userptr BO in the GTT domain.
964  *
965  * The BO must already be on the userptr_valid_list. Otherwise an
966  * eviction and restore may happen that leaves the new BO unmapped
967  * with the user mode queues running.
968  *
969  * Takes the process_info->lock to protect against concurrent restore
970  * workers.
971  *
972  * Returns 0 for success, negative errno for errors.
973  */
974 static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr,
975 			   bool criu_resume)
976 {
977 	struct amdkfd_process_info *process_info = mem->process_info;
978 	struct amdgpu_bo *bo = mem->bo;
979 	struct ttm_operation_ctx ctx = { true, false };
980 	struct hmm_range *range;
981 	int ret = 0;
982 
983 	mutex_lock(&process_info->lock);
984 
985 	ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
986 	if (ret) {
987 		pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
988 		goto out;
989 	}
990 
991 	ret = amdgpu_hmm_register(bo, user_addr);
992 	if (ret) {
993 		pr_err("%s: Failed to register MMU notifier: %d\n",
994 		       __func__, ret);
995 		goto out;
996 	}
997 
998 	if (criu_resume) {
999 		/*
1000 		 * During a CRIU restore operation, the userptr buffer objects
1001 		 * will be validated in the restore_userptr_work worker at a
1002 		 * later stage when it is scheduled by another ioctl called by
1003 		 * CRIU master process for the target pid for restore.
1004 		 */
1005 		mutex_lock(&process_info->notifier_lock);
1006 		mem->invalid++;
1007 		mutex_unlock(&process_info->notifier_lock);
1008 		mutex_unlock(&process_info->lock);
1009 		return 0;
1010 	}
1011 
1012 	ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, &range);
1013 	if (ret) {
1014 		pr_err("%s: Failed to get user pages: %d\n", __func__, ret);
1015 		goto unregister_out;
1016 	}
1017 
1018 	ret = amdgpu_bo_reserve(bo, true);
1019 	if (ret) {
1020 		pr_err("%s: Failed to reserve BO\n", __func__);
1021 		goto release_out;
1022 	}
1023 	amdgpu_bo_placement_from_domain(bo, mem->domain);
1024 	ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1025 	if (ret)
1026 		pr_err("%s: failed to validate BO\n", __func__);
1027 	amdgpu_bo_unreserve(bo);
1028 
1029 release_out:
1030 	amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range);
1031 unregister_out:
1032 	if (ret)
1033 		amdgpu_hmm_unregister(bo);
1034 out:
1035 	mutex_unlock(&process_info->lock);
1036 	return ret;
1037 }
1038 
1039 /* Reserving a BO and its page table BOs must happen atomically to
1040  * avoid deadlocks. Some operations update multiple VMs at once. Track
1041  * all the reservation info in a context structure. Optionally a sync
1042  * object can track VM updates.
1043  */
1044 struct bo_vm_reservation_context {
1045 	struct amdgpu_bo_list_entry kfd_bo; /* BO list entry for the KFD BO */
1046 	unsigned int n_vms;		    /* Number of VMs reserved	    */
1047 	struct amdgpu_bo_list_entry *vm_pd; /* Array of VM BO list entries  */
1048 	struct ww_acquire_ctx ticket;	    /* Reservation ticket	    */
1049 	struct list_head list, duplicates;  /* BO lists			    */
1050 	struct amdgpu_sync *sync;	    /* Pointer to sync object	    */
1051 	bool reserved;			    /* Whether BOs are reserved	    */
1052 };
1053 
1054 enum bo_vm_match {
1055 	BO_VM_NOT_MAPPED = 0,	/* Match VMs where a BO is not mapped */
1056 	BO_VM_MAPPED,		/* Match VMs where a BO is mapped     */
1057 	BO_VM_ALL,		/* Match all VMs a BO was added to    */
1058 };
1059 
1060 /**
1061  * reserve_bo_and_vm - reserve a BO and a VM unconditionally.
1062  * @mem: KFD BO structure.
1063  * @vm: the VM to reserve.
1064  * @ctx: the struct that will be used in unreserve_bo_and_vms().
1065  */
1066 static int reserve_bo_and_vm(struct kgd_mem *mem,
1067 			      struct amdgpu_vm *vm,
1068 			      struct bo_vm_reservation_context *ctx)
1069 {
1070 	struct amdgpu_bo *bo = mem->bo;
1071 	int ret;
1072 
1073 	WARN_ON(!vm);
1074 
1075 	ctx->reserved = false;
1076 	ctx->n_vms = 1;
1077 	ctx->sync = &mem->sync;
1078 
1079 	INIT_LIST_HEAD(&ctx->list);
1080 	INIT_LIST_HEAD(&ctx->duplicates);
1081 
1082 	ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd), GFP_KERNEL);
1083 	if (!ctx->vm_pd)
1084 		return -ENOMEM;
1085 
1086 	ctx->kfd_bo.priority = 0;
1087 	ctx->kfd_bo.tv.bo = &bo->tbo;
1088 	ctx->kfd_bo.tv.num_shared = 1;
1089 	list_add(&ctx->kfd_bo.tv.head, &ctx->list);
1090 
1091 	amdgpu_vm_get_pd_bo(vm, &ctx->list, &ctx->vm_pd[0]);
1092 
1093 	ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
1094 				     false, &ctx->duplicates);
1095 	if (ret) {
1096 		pr_err("Failed to reserve buffers in ttm.\n");
1097 		kfree(ctx->vm_pd);
1098 		ctx->vm_pd = NULL;
1099 		return ret;
1100 	}
1101 
1102 	ctx->reserved = true;
1103 	return 0;
1104 }
1105 
1106 /**
1107  * reserve_bo_and_cond_vms - reserve a BO and some VMs conditionally
1108  * @mem: KFD BO structure.
1109  * @vm: the VM to reserve. If NULL, then all VMs associated with the BO
1110  * is used. Otherwise, a single VM associated with the BO.
1111  * @map_type: the mapping status that will be used to filter the VMs.
1112  * @ctx: the struct that will be used in unreserve_bo_and_vms().
1113  *
1114  * Returns 0 for success, negative for failure.
1115  */
1116 static int reserve_bo_and_cond_vms(struct kgd_mem *mem,
1117 				struct amdgpu_vm *vm, enum bo_vm_match map_type,
1118 				struct bo_vm_reservation_context *ctx)
1119 {
1120 	struct amdgpu_bo *bo = mem->bo;
1121 	struct kfd_mem_attachment *entry;
1122 	unsigned int i;
1123 	int ret;
1124 
1125 	ctx->reserved = false;
1126 	ctx->n_vms = 0;
1127 	ctx->vm_pd = NULL;
1128 	ctx->sync = &mem->sync;
1129 
1130 	INIT_LIST_HEAD(&ctx->list);
1131 	INIT_LIST_HEAD(&ctx->duplicates);
1132 
1133 	list_for_each_entry(entry, &mem->attachments, list) {
1134 		if ((vm && vm != entry->bo_va->base.vm) ||
1135 			(entry->is_mapped != map_type
1136 			&& map_type != BO_VM_ALL))
1137 			continue;
1138 
1139 		ctx->n_vms++;
1140 	}
1141 
1142 	if (ctx->n_vms != 0) {
1143 		ctx->vm_pd = kcalloc(ctx->n_vms, sizeof(*ctx->vm_pd),
1144 				     GFP_KERNEL);
1145 		if (!ctx->vm_pd)
1146 			return -ENOMEM;
1147 	}
1148 
1149 	ctx->kfd_bo.priority = 0;
1150 	ctx->kfd_bo.tv.bo = &bo->tbo;
1151 	ctx->kfd_bo.tv.num_shared = 1;
1152 	list_add(&ctx->kfd_bo.tv.head, &ctx->list);
1153 
1154 	i = 0;
1155 	list_for_each_entry(entry, &mem->attachments, list) {
1156 		if ((vm && vm != entry->bo_va->base.vm) ||
1157 			(entry->is_mapped != map_type
1158 			&& map_type != BO_VM_ALL))
1159 			continue;
1160 
1161 		amdgpu_vm_get_pd_bo(entry->bo_va->base.vm, &ctx->list,
1162 				&ctx->vm_pd[i]);
1163 		i++;
1164 	}
1165 
1166 	ret = ttm_eu_reserve_buffers(&ctx->ticket, &ctx->list,
1167 				     false, &ctx->duplicates);
1168 	if (ret) {
1169 		pr_err("Failed to reserve buffers in ttm.\n");
1170 		kfree(ctx->vm_pd);
1171 		ctx->vm_pd = NULL;
1172 		return ret;
1173 	}
1174 
1175 	ctx->reserved = true;
1176 	return 0;
1177 }
1178 
1179 /**
1180  * unreserve_bo_and_vms - Unreserve BO and VMs from a reservation context
1181  * @ctx: Reservation context to unreserve
1182  * @wait: Optionally wait for a sync object representing pending VM updates
1183  * @intr: Whether the wait is interruptible
1184  *
1185  * Also frees any resources allocated in
1186  * reserve_bo_and_(cond_)vm(s). Returns the status from
1187  * amdgpu_sync_wait.
1188  */
1189 static int unreserve_bo_and_vms(struct bo_vm_reservation_context *ctx,
1190 				 bool wait, bool intr)
1191 {
1192 	int ret = 0;
1193 
1194 	if (wait)
1195 		ret = amdgpu_sync_wait(ctx->sync, intr);
1196 
1197 	if (ctx->reserved)
1198 		ttm_eu_backoff_reservation(&ctx->ticket, &ctx->list);
1199 	kfree(ctx->vm_pd);
1200 
1201 	ctx->sync = NULL;
1202 
1203 	ctx->reserved = false;
1204 	ctx->vm_pd = NULL;
1205 
1206 	return ret;
1207 }
1208 
1209 static void unmap_bo_from_gpuvm(struct kgd_mem *mem,
1210 				struct kfd_mem_attachment *entry,
1211 				struct amdgpu_sync *sync)
1212 {
1213 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1214 	struct amdgpu_device *adev = entry->adev;
1215 	struct amdgpu_vm *vm = bo_va->base.vm;
1216 
1217 	amdgpu_vm_bo_unmap(adev, bo_va, entry->va);
1218 
1219 	amdgpu_vm_clear_freed(adev, vm, &bo_va->last_pt_update);
1220 
1221 	amdgpu_sync_fence(sync, bo_va->last_pt_update);
1222 
1223 	kfd_mem_dmaunmap_attachment(mem, entry);
1224 }
1225 
1226 static int update_gpuvm_pte(struct kgd_mem *mem,
1227 			    struct kfd_mem_attachment *entry,
1228 			    struct amdgpu_sync *sync)
1229 {
1230 	struct amdgpu_bo_va *bo_va = entry->bo_va;
1231 	struct amdgpu_device *adev = entry->adev;
1232 	int ret;
1233 
1234 	ret = kfd_mem_dmamap_attachment(mem, entry);
1235 	if (ret)
1236 		return ret;
1237 
1238 	/* Update the page tables  */
1239 	ret = amdgpu_vm_bo_update(adev, bo_va, false);
1240 	if (ret) {
1241 		pr_err("amdgpu_vm_bo_update failed\n");
1242 		return ret;
1243 	}
1244 
1245 	return amdgpu_sync_fence(sync, bo_va->last_pt_update);
1246 }
1247 
1248 static int map_bo_to_gpuvm(struct kgd_mem *mem,
1249 			   struct kfd_mem_attachment *entry,
1250 			   struct amdgpu_sync *sync,
1251 			   bool no_update_pte)
1252 {
1253 	int ret;
1254 
1255 	/* Set virtual address for the allocation */
1256 	ret = amdgpu_vm_bo_map(entry->adev, entry->bo_va, entry->va, 0,
1257 			       amdgpu_bo_size(entry->bo_va->base.bo),
1258 			       entry->pte_flags);
1259 	if (ret) {
1260 		pr_err("Failed to map VA 0x%llx in vm. ret %d\n",
1261 				entry->va, ret);
1262 		return ret;
1263 	}
1264 
1265 	if (no_update_pte)
1266 		return 0;
1267 
1268 	ret = update_gpuvm_pte(mem, entry, sync);
1269 	if (ret) {
1270 		pr_err("update_gpuvm_pte() failed\n");
1271 		goto update_gpuvm_pte_failed;
1272 	}
1273 
1274 	return 0;
1275 
1276 update_gpuvm_pte_failed:
1277 	unmap_bo_from_gpuvm(mem, entry, sync);
1278 	return ret;
1279 }
1280 
1281 static int process_validate_vms(struct amdkfd_process_info *process_info)
1282 {
1283 	struct amdgpu_vm *peer_vm;
1284 	int ret;
1285 
1286 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1287 			    vm_list_node) {
1288 		ret = vm_validate_pt_pd_bos(peer_vm);
1289 		if (ret)
1290 			return ret;
1291 	}
1292 
1293 	return 0;
1294 }
1295 
1296 static int process_sync_pds_resv(struct amdkfd_process_info *process_info,
1297 				 struct amdgpu_sync *sync)
1298 {
1299 	struct amdgpu_vm *peer_vm;
1300 	int ret;
1301 
1302 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1303 			    vm_list_node) {
1304 		struct amdgpu_bo *pd = peer_vm->root.bo;
1305 
1306 		ret = amdgpu_sync_resv(NULL, sync, pd->tbo.base.resv,
1307 				       AMDGPU_SYNC_NE_OWNER,
1308 				       AMDGPU_FENCE_OWNER_KFD);
1309 		if (ret)
1310 			return ret;
1311 	}
1312 
1313 	return 0;
1314 }
1315 
1316 static int process_update_pds(struct amdkfd_process_info *process_info,
1317 			      struct amdgpu_sync *sync)
1318 {
1319 	struct amdgpu_vm *peer_vm;
1320 	int ret;
1321 
1322 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
1323 			    vm_list_node) {
1324 		ret = vm_update_pds(peer_vm, sync);
1325 		if (ret)
1326 			return ret;
1327 	}
1328 
1329 	return 0;
1330 }
1331 
1332 static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
1333 		       struct dma_fence **ef)
1334 {
1335 	struct amdkfd_process_info *info = NULL;
1336 	int ret;
1337 
1338 	if (!*process_info) {
1339 		info = kzalloc(sizeof(*info), GFP_KERNEL);
1340 		if (!info)
1341 			return -ENOMEM;
1342 
1343 		mutex_init(&info->lock);
1344 		mutex_init(&info->notifier_lock);
1345 		INIT_LIST_HEAD(&info->vm_list_head);
1346 		INIT_LIST_HEAD(&info->kfd_bo_list);
1347 		INIT_LIST_HEAD(&info->userptr_valid_list);
1348 		INIT_LIST_HEAD(&info->userptr_inval_list);
1349 
1350 		info->eviction_fence =
1351 			amdgpu_amdkfd_fence_create(dma_fence_context_alloc(1),
1352 						   current->mm,
1353 						   NULL);
1354 		if (!info->eviction_fence) {
1355 			pr_err("Failed to create eviction fence\n");
1356 			ret = -ENOMEM;
1357 			goto create_evict_fence_fail;
1358 		}
1359 
1360 		info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
1361 		INIT_DELAYED_WORK(&info->restore_userptr_work,
1362 				  amdgpu_amdkfd_restore_userptr_worker);
1363 
1364 		*process_info = info;
1365 		*ef = dma_fence_get(&info->eviction_fence->base);
1366 	}
1367 
1368 	vm->process_info = *process_info;
1369 
1370 	/* Validate page directory and attach eviction fence */
1371 	ret = amdgpu_bo_reserve(vm->root.bo, true);
1372 	if (ret)
1373 		goto reserve_pd_fail;
1374 	ret = vm_validate_pt_pd_bos(vm);
1375 	if (ret) {
1376 		pr_err("validate_pt_pd_bos() failed\n");
1377 		goto validate_pd_fail;
1378 	}
1379 	ret = amdgpu_bo_sync_wait(vm->root.bo,
1380 				  AMDGPU_FENCE_OWNER_KFD, false);
1381 	if (ret)
1382 		goto wait_pd_fail;
1383 	ret = dma_resv_reserve_fences(vm->root.bo->tbo.base.resv, 1);
1384 	if (ret)
1385 		goto reserve_shared_fail;
1386 	dma_resv_add_fence(vm->root.bo->tbo.base.resv,
1387 			   &vm->process_info->eviction_fence->base,
1388 			   DMA_RESV_USAGE_BOOKKEEP);
1389 	amdgpu_bo_unreserve(vm->root.bo);
1390 
1391 	/* Update process info */
1392 	mutex_lock(&vm->process_info->lock);
1393 	list_add_tail(&vm->vm_list_node,
1394 			&(vm->process_info->vm_list_head));
1395 	vm->process_info->n_vms++;
1396 	mutex_unlock(&vm->process_info->lock);
1397 
1398 	return 0;
1399 
1400 reserve_shared_fail:
1401 wait_pd_fail:
1402 validate_pd_fail:
1403 	amdgpu_bo_unreserve(vm->root.bo);
1404 reserve_pd_fail:
1405 	vm->process_info = NULL;
1406 	if (info) {
1407 		/* Two fence references: one in info and one in *ef */
1408 		dma_fence_put(&info->eviction_fence->base);
1409 		dma_fence_put(*ef);
1410 		*ef = NULL;
1411 		*process_info = NULL;
1412 		put_pid(info->pid);
1413 create_evict_fence_fail:
1414 		mutex_destroy(&info->lock);
1415 		mutex_destroy(&info->notifier_lock);
1416 		kfree(info);
1417 	}
1418 	return ret;
1419 }
1420 
1421 /**
1422  * amdgpu_amdkfd_gpuvm_pin_bo() - Pins a BO using following criteria
1423  * @bo: Handle of buffer object being pinned
1424  * @domain: Domain into which BO should be pinned
1425  *
1426  *   - USERPTR BOs are UNPINNABLE and will return error
1427  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1428  *     PIN count incremented. It is valid to PIN a BO multiple times
1429  *
1430  * Return: ZERO if successful in pinning, Non-Zero in case of error.
1431  */
1432 static int amdgpu_amdkfd_gpuvm_pin_bo(struct amdgpu_bo *bo, u32 domain)
1433 {
1434 	int ret = 0;
1435 
1436 	ret = amdgpu_bo_reserve(bo, false);
1437 	if (unlikely(ret))
1438 		return ret;
1439 
1440 	ret = amdgpu_bo_pin_restricted(bo, domain, 0, 0);
1441 	if (ret)
1442 		pr_err("Error in Pinning BO to domain: %d\n", domain);
1443 
1444 	amdgpu_bo_sync_wait(bo, AMDGPU_FENCE_OWNER_KFD, false);
1445 	amdgpu_bo_unreserve(bo);
1446 
1447 	return ret;
1448 }
1449 
1450 /**
1451  * amdgpu_amdkfd_gpuvm_unpin_bo() - Unpins BO using following criteria
1452  * @bo: Handle of buffer object being unpinned
1453  *
1454  *   - Is a illegal request for USERPTR BOs and is ignored
1455  *   - All other BO types (GTT, VRAM, MMIO and DOORBELL) will have their
1456  *     PIN count decremented. Calls to UNPIN must balance calls to PIN
1457  */
1458 static void amdgpu_amdkfd_gpuvm_unpin_bo(struct amdgpu_bo *bo)
1459 {
1460 	int ret = 0;
1461 
1462 	ret = amdgpu_bo_reserve(bo, false);
1463 	if (unlikely(ret))
1464 		return;
1465 
1466 	amdgpu_bo_unpin(bo);
1467 	amdgpu_bo_unreserve(bo);
1468 }
1469 
1470 int amdgpu_amdkfd_gpuvm_set_vm_pasid(struct amdgpu_device *adev,
1471 				     struct amdgpu_vm *avm, u32 pasid)
1472 
1473 {
1474 	int ret;
1475 
1476 	/* Free the original amdgpu allocated pasid,
1477 	 * will be replaced with kfd allocated pasid.
1478 	 */
1479 	if (avm->pasid) {
1480 		amdgpu_pasid_free(avm->pasid);
1481 		amdgpu_vm_set_pasid(adev, avm, 0);
1482 	}
1483 
1484 	ret = amdgpu_vm_set_pasid(adev, avm, pasid);
1485 	if (ret)
1486 		return ret;
1487 
1488 	return 0;
1489 }
1490 
1491 int amdgpu_amdkfd_gpuvm_acquire_process_vm(struct amdgpu_device *adev,
1492 					   struct amdgpu_vm *avm,
1493 					   void **process_info,
1494 					   struct dma_fence **ef)
1495 {
1496 	int ret;
1497 
1498 	/* Already a compute VM? */
1499 	if (avm->process_info)
1500 		return -EINVAL;
1501 
1502 	/* Convert VM into a compute VM */
1503 	ret = amdgpu_vm_make_compute(adev, avm);
1504 	if (ret)
1505 		return ret;
1506 
1507 	/* Initialize KFD part of the VM and process info */
1508 	ret = init_kfd_vm(avm, process_info, ef);
1509 	if (ret)
1510 		return ret;
1511 
1512 	amdgpu_vm_set_task_info(avm);
1513 
1514 	return 0;
1515 }
1516 
1517 void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
1518 				    struct amdgpu_vm *vm)
1519 {
1520 	struct amdkfd_process_info *process_info = vm->process_info;
1521 
1522 	if (!process_info)
1523 		return;
1524 
1525 	/* Update process info */
1526 	mutex_lock(&process_info->lock);
1527 	process_info->n_vms--;
1528 	list_del(&vm->vm_list_node);
1529 	mutex_unlock(&process_info->lock);
1530 
1531 	vm->process_info = NULL;
1532 
1533 	/* Release per-process resources when last compute VM is destroyed */
1534 	if (!process_info->n_vms) {
1535 		WARN_ON(!list_empty(&process_info->kfd_bo_list));
1536 		WARN_ON(!list_empty(&process_info->userptr_valid_list));
1537 		WARN_ON(!list_empty(&process_info->userptr_inval_list));
1538 
1539 		dma_fence_put(&process_info->eviction_fence->base);
1540 		cancel_delayed_work_sync(&process_info->restore_userptr_work);
1541 		put_pid(process_info->pid);
1542 		mutex_destroy(&process_info->lock);
1543 		mutex_destroy(&process_info->notifier_lock);
1544 		kfree(process_info);
1545 	}
1546 }
1547 
1548 void amdgpu_amdkfd_gpuvm_release_process_vm(struct amdgpu_device *adev,
1549 					    void *drm_priv)
1550 {
1551 	struct amdgpu_vm *avm;
1552 
1553 	if (WARN_ON(!adev || !drm_priv))
1554 		return;
1555 
1556 	avm = drm_priv_to_vm(drm_priv);
1557 
1558 	pr_debug("Releasing process vm %p\n", avm);
1559 
1560 	/* The original pasid of amdgpu vm has already been
1561 	 * released during making a amdgpu vm to a compute vm
1562 	 * The current pasid is managed by kfd and will be
1563 	 * released on kfd process destroy. Set amdgpu pasid
1564 	 * to 0 to avoid duplicate release.
1565 	 */
1566 	amdgpu_vm_release_compute(adev, avm);
1567 }
1568 
1569 uint64_t amdgpu_amdkfd_gpuvm_get_process_page_dir(void *drm_priv)
1570 {
1571 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1572 	struct amdgpu_bo *pd = avm->root.bo;
1573 	struct amdgpu_device *adev = amdgpu_ttm_adev(pd->tbo.bdev);
1574 
1575 	if (adev->asic_type < CHIP_VEGA10)
1576 		return avm->pd_phys_addr >> AMDGPU_GPU_PAGE_SHIFT;
1577 	return avm->pd_phys_addr;
1578 }
1579 
1580 void amdgpu_amdkfd_block_mmu_notifications(void *p)
1581 {
1582 	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1583 
1584 	mutex_lock(&pinfo->lock);
1585 	WRITE_ONCE(pinfo->block_mmu_notifications, true);
1586 	mutex_unlock(&pinfo->lock);
1587 }
1588 
1589 int amdgpu_amdkfd_criu_resume(void *p)
1590 {
1591 	int ret = 0;
1592 	struct amdkfd_process_info *pinfo = (struct amdkfd_process_info *)p;
1593 
1594 	mutex_lock(&pinfo->lock);
1595 	pr_debug("scheduling work\n");
1596 	mutex_lock(&pinfo->notifier_lock);
1597 	pinfo->evicted_bos++;
1598 	mutex_unlock(&pinfo->notifier_lock);
1599 	if (!READ_ONCE(pinfo->block_mmu_notifications)) {
1600 		ret = -EINVAL;
1601 		goto out_unlock;
1602 	}
1603 	WRITE_ONCE(pinfo->block_mmu_notifications, false);
1604 	schedule_delayed_work(&pinfo->restore_userptr_work, 0);
1605 
1606 out_unlock:
1607 	mutex_unlock(&pinfo->lock);
1608 	return ret;
1609 }
1610 
1611 size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev)
1612 {
1613 	uint64_t reserved_for_pt =
1614 		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
1615 	ssize_t available;
1616 
1617 	spin_lock(&kfd_mem_limit.mem_limit_lock);
1618 	available = adev->gmc.real_vram_size
1619 		- adev->kfd.vram_used_aligned
1620 		- atomic64_read(&adev->vram_pin_size)
1621 		- reserved_for_pt;
1622 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
1623 
1624 	if (available < 0)
1625 		available = 0;
1626 
1627 	return ALIGN_DOWN(available, VRAM_AVAILABLITY_ALIGN);
1628 }
1629 
1630 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
1631 		struct amdgpu_device *adev, uint64_t va, uint64_t size,
1632 		void *drm_priv, struct kgd_mem **mem,
1633 		uint64_t *offset, uint32_t flags, bool criu_resume)
1634 {
1635 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1636 	enum ttm_bo_type bo_type = ttm_bo_type_device;
1637 	struct sg_table *sg = NULL;
1638 	uint64_t user_addr = 0;
1639 	struct amdgpu_bo *bo;
1640 	struct drm_gem_object *gobj = NULL;
1641 	u32 domain, alloc_domain;
1642 	uint64_t aligned_size;
1643 	u64 alloc_flags;
1644 	int ret;
1645 
1646 	/*
1647 	 * Check on which domain to allocate BO
1648 	 */
1649 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
1650 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_VRAM;
1651 		alloc_flags = AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
1652 		alloc_flags |= (flags & KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC) ?
1653 			AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : 0;
1654 	} else if (flags & KFD_IOC_ALLOC_MEM_FLAGS_GTT) {
1655 		domain = alloc_domain = AMDGPU_GEM_DOMAIN_GTT;
1656 		alloc_flags = 0;
1657 	} else {
1658 		domain = AMDGPU_GEM_DOMAIN_GTT;
1659 		alloc_domain = AMDGPU_GEM_DOMAIN_CPU;
1660 		alloc_flags = AMDGPU_GEM_CREATE_PREEMPTIBLE;
1661 
1662 		if (flags & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
1663 			if (!offset || !*offset)
1664 				return -EINVAL;
1665 			user_addr = untagged_addr(*offset);
1666 		} else if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1667 				    KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1668 			bo_type = ttm_bo_type_sg;
1669 			if (size > UINT_MAX)
1670 				return -EINVAL;
1671 			sg = create_sg_table(*offset, size);
1672 			if (!sg)
1673 				return -ENOMEM;
1674 		} else {
1675 			return -EINVAL;
1676 		}
1677 	}
1678 
1679 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_COHERENT)
1680 		alloc_flags |= AMDGPU_GEM_CREATE_COHERENT;
1681 	if (flags & KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED)
1682 		alloc_flags |= AMDGPU_GEM_CREATE_UNCACHED;
1683 
1684 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
1685 	if (!*mem) {
1686 		ret = -ENOMEM;
1687 		goto err;
1688 	}
1689 	INIT_LIST_HEAD(&(*mem)->attachments);
1690 	mutex_init(&(*mem)->lock);
1691 	(*mem)->aql_queue = !!(flags & KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM);
1692 
1693 	/* Workaround for AQL queue wraparound bug. Map the same
1694 	 * memory twice. That means we only actually allocate half
1695 	 * the memory.
1696 	 */
1697 	if ((*mem)->aql_queue)
1698 		size >>= 1;
1699 	aligned_size = PAGE_ALIGN(size);
1700 
1701 	(*mem)->alloc_flags = flags;
1702 
1703 	amdgpu_sync_create(&(*mem)->sync);
1704 
1705 	ret = amdgpu_amdkfd_reserve_mem_limit(adev, aligned_size, flags);
1706 	if (ret) {
1707 		pr_debug("Insufficient memory\n");
1708 		goto err_reserve_limit;
1709 	}
1710 
1711 	pr_debug("\tcreate BO VA 0x%llx size 0x%llx domain %s\n",
1712 			va, (*mem)->aql_queue ? size << 1 : size, domain_string(alloc_domain));
1713 
1714 	ret = amdgpu_gem_object_create(adev, aligned_size, 1, alloc_domain, alloc_flags,
1715 				       bo_type, NULL, &gobj);
1716 	if (ret) {
1717 		pr_debug("Failed to create BO on domain %s. ret %d\n",
1718 			 domain_string(alloc_domain), ret);
1719 		goto err_bo_create;
1720 	}
1721 	ret = drm_vma_node_allow(&gobj->vma_node, drm_priv);
1722 	if (ret) {
1723 		pr_debug("Failed to allow vma node access. ret %d\n", ret);
1724 		goto err_node_allow;
1725 	}
1726 	bo = gem_to_amdgpu_bo(gobj);
1727 	if (bo_type == ttm_bo_type_sg) {
1728 		bo->tbo.sg = sg;
1729 		bo->tbo.ttm->sg = sg;
1730 	}
1731 	bo->kfd_bo = *mem;
1732 	(*mem)->bo = bo;
1733 	if (user_addr)
1734 		bo->flags |= AMDGPU_AMDKFD_CREATE_USERPTR_BO;
1735 
1736 	(*mem)->va = va;
1737 	(*mem)->domain = domain;
1738 	(*mem)->mapped_to_gpu_memory = 0;
1739 	(*mem)->process_info = avm->process_info;
1740 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, user_addr);
1741 
1742 	if (user_addr) {
1743 		pr_debug("creating userptr BO for user_addr = %llx\n", user_addr);
1744 		ret = init_user_pages(*mem, user_addr, criu_resume);
1745 		if (ret)
1746 			goto allocate_init_user_pages_failed;
1747 	} else  if (flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1748 				KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1749 		ret = amdgpu_amdkfd_gpuvm_pin_bo(bo, AMDGPU_GEM_DOMAIN_GTT);
1750 		if (ret) {
1751 			pr_err("Pinning MMIO/DOORBELL BO during ALLOC FAILED\n");
1752 			goto err_pin_bo;
1753 		}
1754 		bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
1755 		bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
1756 	}
1757 
1758 	if (offset)
1759 		*offset = amdgpu_bo_mmap_offset(bo);
1760 
1761 	return 0;
1762 
1763 allocate_init_user_pages_failed:
1764 err_pin_bo:
1765 	remove_kgd_mem_from_kfd_bo_list(*mem, avm->process_info);
1766 	drm_vma_node_revoke(&gobj->vma_node, drm_priv);
1767 err_node_allow:
1768 	/* Don't unreserve system mem limit twice */
1769 	goto err_reserve_limit;
1770 err_bo_create:
1771 	amdgpu_amdkfd_unreserve_mem_limit(adev, aligned_size, flags);
1772 err_reserve_limit:
1773 	mutex_destroy(&(*mem)->lock);
1774 	if (gobj)
1775 		drm_gem_object_put(gobj);
1776 	else
1777 		kfree(*mem);
1778 err:
1779 	if (sg) {
1780 		sg_free_table(sg);
1781 		kfree(sg);
1782 	}
1783 	return ret;
1784 }
1785 
1786 int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
1787 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv,
1788 		uint64_t *size)
1789 {
1790 	struct amdkfd_process_info *process_info = mem->process_info;
1791 	unsigned long bo_size = mem->bo->tbo.base.size;
1792 	bool use_release_notifier = (mem->bo->kfd_bo == mem);
1793 	struct kfd_mem_attachment *entry, *tmp;
1794 	struct bo_vm_reservation_context ctx;
1795 	struct ttm_validate_buffer *bo_list_entry;
1796 	unsigned int mapped_to_gpu_memory;
1797 	int ret;
1798 	bool is_imported = false;
1799 
1800 	mutex_lock(&mem->lock);
1801 
1802 	/* Unpin MMIO/DOORBELL BO's that were pinned during allocation */
1803 	if (mem->alloc_flags &
1804 	    (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
1805 	     KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP)) {
1806 		amdgpu_amdkfd_gpuvm_unpin_bo(mem->bo);
1807 	}
1808 
1809 	mapped_to_gpu_memory = mem->mapped_to_gpu_memory;
1810 	is_imported = mem->is_imported;
1811 	mutex_unlock(&mem->lock);
1812 	/* lock is not needed after this, since mem is unused and will
1813 	 * be freed anyway
1814 	 */
1815 
1816 	if (mapped_to_gpu_memory > 0) {
1817 		pr_debug("BO VA 0x%llx size 0x%lx is still mapped.\n",
1818 				mem->va, bo_size);
1819 		return -EBUSY;
1820 	}
1821 
1822 	/* Make sure restore workers don't access the BO any more */
1823 	bo_list_entry = &mem->validate_list;
1824 	mutex_lock(&process_info->lock);
1825 	list_del(&bo_list_entry->head);
1826 	mutex_unlock(&process_info->lock);
1827 
1828 	/* Cleanup user pages and MMU notifiers */
1829 	if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
1830 		amdgpu_hmm_unregister(mem->bo);
1831 		mutex_lock(&process_info->notifier_lock);
1832 		amdgpu_ttm_tt_discard_user_pages(mem->bo->tbo.ttm, mem->range);
1833 		mutex_unlock(&process_info->notifier_lock);
1834 	}
1835 
1836 	ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
1837 	if (unlikely(ret))
1838 		return ret;
1839 
1840 	/* The eviction fence should be removed by the last unmap.
1841 	 * TODO: Log an error condition if the bo still has the eviction fence
1842 	 * attached
1843 	 */
1844 	amdgpu_amdkfd_remove_eviction_fence(mem->bo,
1845 					process_info->eviction_fence);
1846 	pr_debug("Release VA 0x%llx - 0x%llx\n", mem->va,
1847 		mem->va + bo_size * (1 + mem->aql_queue));
1848 
1849 	/* Remove from VM internal data structures */
1850 	list_for_each_entry_safe(entry, tmp, &mem->attachments, list)
1851 		kfd_mem_detach(entry);
1852 
1853 	ret = unreserve_bo_and_vms(&ctx, false, false);
1854 
1855 	/* Free the sync object */
1856 	amdgpu_sync_free(&mem->sync);
1857 
1858 	/* If the SG is not NULL, it's one we created for a doorbell or mmio
1859 	 * remap BO. We need to free it.
1860 	 */
1861 	if (mem->bo->tbo.sg) {
1862 		sg_free_table(mem->bo->tbo.sg);
1863 		kfree(mem->bo->tbo.sg);
1864 	}
1865 
1866 	/* Update the size of the BO being freed if it was allocated from
1867 	 * VRAM and is not imported.
1868 	 */
1869 	if (size) {
1870 		if ((mem->bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM) &&
1871 		    (!is_imported))
1872 			*size = bo_size;
1873 		else
1874 			*size = 0;
1875 	}
1876 
1877 	/* Free the BO*/
1878 	drm_vma_node_revoke(&mem->bo->tbo.base.vma_node, drm_priv);
1879 	if (mem->dmabuf)
1880 		dma_buf_put(mem->dmabuf);
1881 	mutex_destroy(&mem->lock);
1882 
1883 	/* If this releases the last reference, it will end up calling
1884 	 * amdgpu_amdkfd_release_notify and kfree the mem struct. That's why
1885 	 * this needs to be the last call here.
1886 	 */
1887 	drm_gem_object_put(&mem->bo->tbo.base);
1888 
1889 	/*
1890 	 * For kgd_mem allocated in amdgpu_amdkfd_gpuvm_import_dmabuf(),
1891 	 * explicitly free it here.
1892 	 */
1893 	if (!use_release_notifier)
1894 		kfree(mem);
1895 
1896 	return ret;
1897 }
1898 
1899 int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
1900 		struct amdgpu_device *adev, struct kgd_mem *mem,
1901 		void *drm_priv)
1902 {
1903 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
1904 	int ret;
1905 	struct amdgpu_bo *bo;
1906 	uint32_t domain;
1907 	struct kfd_mem_attachment *entry;
1908 	struct bo_vm_reservation_context ctx;
1909 	unsigned long bo_size;
1910 	bool is_invalid_userptr = false;
1911 
1912 	bo = mem->bo;
1913 	if (!bo) {
1914 		pr_err("Invalid BO when mapping memory to GPU\n");
1915 		return -EINVAL;
1916 	}
1917 
1918 	/* Make sure restore is not running concurrently. Since we
1919 	 * don't map invalid userptr BOs, we rely on the next restore
1920 	 * worker to do the mapping
1921 	 */
1922 	mutex_lock(&mem->process_info->lock);
1923 
1924 	/* Lock notifier lock. If we find an invalid userptr BO, we can be
1925 	 * sure that the MMU notifier is no longer running
1926 	 * concurrently and the queues are actually stopped
1927 	 */
1928 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1929 		mutex_lock(&mem->process_info->notifier_lock);
1930 		is_invalid_userptr = !!mem->invalid;
1931 		mutex_unlock(&mem->process_info->notifier_lock);
1932 	}
1933 
1934 	mutex_lock(&mem->lock);
1935 
1936 	domain = mem->domain;
1937 	bo_size = bo->tbo.base.size;
1938 
1939 	pr_debug("Map VA 0x%llx - 0x%llx to vm %p domain %s\n",
1940 			mem->va,
1941 			mem->va + bo_size * (1 + mem->aql_queue),
1942 			avm, domain_string(domain));
1943 
1944 	if (!kfd_mem_is_attached(avm, mem)) {
1945 		ret = kfd_mem_attach(adev, mem, avm, mem->aql_queue);
1946 		if (ret)
1947 			goto out;
1948 	}
1949 
1950 	ret = reserve_bo_and_vm(mem, avm, &ctx);
1951 	if (unlikely(ret))
1952 		goto out;
1953 
1954 	/* Userptr can be marked as "not invalid", but not actually be
1955 	 * validated yet (still in the system domain). In that case
1956 	 * the queues are still stopped and we can leave mapping for
1957 	 * the next restore worker
1958 	 */
1959 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) &&
1960 	    bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
1961 		is_invalid_userptr = true;
1962 
1963 	ret = vm_validate_pt_pd_bos(avm);
1964 	if (unlikely(ret))
1965 		goto out_unreserve;
1966 
1967 	if (mem->mapped_to_gpu_memory == 0 &&
1968 	    !amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
1969 		/* Validate BO only once. The eviction fence gets added to BO
1970 		 * the first time it is mapped. Validate will wait for all
1971 		 * background evictions to complete.
1972 		 */
1973 		ret = amdgpu_amdkfd_bo_validate(bo, domain, true);
1974 		if (ret) {
1975 			pr_debug("Validate failed\n");
1976 			goto out_unreserve;
1977 		}
1978 	}
1979 
1980 	list_for_each_entry(entry, &mem->attachments, list) {
1981 		if (entry->bo_va->base.vm != avm || entry->is_mapped)
1982 			continue;
1983 
1984 		pr_debug("\t map VA 0x%llx - 0x%llx in entry %p\n",
1985 			 entry->va, entry->va + bo_size, entry);
1986 
1987 		ret = map_bo_to_gpuvm(mem, entry, ctx.sync,
1988 				      is_invalid_userptr);
1989 		if (ret) {
1990 			pr_err("Failed to map bo to gpuvm\n");
1991 			goto out_unreserve;
1992 		}
1993 
1994 		ret = vm_update_pds(avm, ctx.sync);
1995 		if (ret) {
1996 			pr_err("Failed to update page directories\n");
1997 			goto out_unreserve;
1998 		}
1999 
2000 		entry->is_mapped = true;
2001 		mem->mapped_to_gpu_memory++;
2002 		pr_debug("\t INC mapping count %d\n",
2003 			 mem->mapped_to_gpu_memory);
2004 	}
2005 
2006 	if (!amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) && !bo->tbo.pin_count)
2007 		dma_resv_add_fence(bo->tbo.base.resv,
2008 				   &avm->process_info->eviction_fence->base,
2009 				   DMA_RESV_USAGE_BOOKKEEP);
2010 	ret = unreserve_bo_and_vms(&ctx, false, false);
2011 
2012 	goto out;
2013 
2014 out_unreserve:
2015 	unreserve_bo_and_vms(&ctx, false, false);
2016 out:
2017 	mutex_unlock(&mem->process_info->lock);
2018 	mutex_unlock(&mem->lock);
2019 	return ret;
2020 }
2021 
2022 int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
2023 		struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv)
2024 {
2025 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2026 	struct amdkfd_process_info *process_info = avm->process_info;
2027 	unsigned long bo_size = mem->bo->tbo.base.size;
2028 	struct kfd_mem_attachment *entry;
2029 	struct bo_vm_reservation_context ctx;
2030 	int ret;
2031 
2032 	mutex_lock(&mem->lock);
2033 
2034 	ret = reserve_bo_and_cond_vms(mem, avm, BO_VM_MAPPED, &ctx);
2035 	if (unlikely(ret))
2036 		goto out;
2037 	/* If no VMs were reserved, it means the BO wasn't actually mapped */
2038 	if (ctx.n_vms == 0) {
2039 		ret = -EINVAL;
2040 		goto unreserve_out;
2041 	}
2042 
2043 	ret = vm_validate_pt_pd_bos(avm);
2044 	if (unlikely(ret))
2045 		goto unreserve_out;
2046 
2047 	pr_debug("Unmap VA 0x%llx - 0x%llx from vm %p\n",
2048 		mem->va,
2049 		mem->va + bo_size * (1 + mem->aql_queue),
2050 		avm);
2051 
2052 	list_for_each_entry(entry, &mem->attachments, list) {
2053 		if (entry->bo_va->base.vm != avm || !entry->is_mapped)
2054 			continue;
2055 
2056 		pr_debug("\t unmap VA 0x%llx - 0x%llx from entry %p\n",
2057 			 entry->va, entry->va + bo_size, entry);
2058 
2059 		unmap_bo_from_gpuvm(mem, entry, ctx.sync);
2060 		entry->is_mapped = false;
2061 
2062 		mem->mapped_to_gpu_memory--;
2063 		pr_debug("\t DEC mapping count %d\n",
2064 			 mem->mapped_to_gpu_memory);
2065 	}
2066 
2067 	/* If BO is unmapped from all VMs, unfence it. It can be evicted if
2068 	 * required.
2069 	 */
2070 	if (mem->mapped_to_gpu_memory == 0 &&
2071 	    !amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm) &&
2072 	    !mem->bo->tbo.pin_count)
2073 		amdgpu_amdkfd_remove_eviction_fence(mem->bo,
2074 						process_info->eviction_fence);
2075 
2076 unreserve_out:
2077 	unreserve_bo_and_vms(&ctx, false, false);
2078 out:
2079 	mutex_unlock(&mem->lock);
2080 	return ret;
2081 }
2082 
2083 int amdgpu_amdkfd_gpuvm_sync_memory(
2084 		struct amdgpu_device *adev, struct kgd_mem *mem, bool intr)
2085 {
2086 	struct amdgpu_sync sync;
2087 	int ret;
2088 
2089 	amdgpu_sync_create(&sync);
2090 
2091 	mutex_lock(&mem->lock);
2092 	amdgpu_sync_clone(&mem->sync, &sync);
2093 	mutex_unlock(&mem->lock);
2094 
2095 	ret = amdgpu_sync_wait(&sync, intr);
2096 	amdgpu_sync_free(&sync);
2097 	return ret;
2098 }
2099 
2100 /**
2101  * amdgpu_amdkfd_map_gtt_bo_to_gart - Map BO to GART and increment reference count
2102  * @adev: Device to which allocated BO belongs
2103  * @bo: Buffer object to be mapped
2104  *
2105  * Before return, bo reference count is incremented. To release the reference and unpin/
2106  * unmap the BO, call amdgpu_amdkfd_free_gtt_mem.
2107  */
2108 int amdgpu_amdkfd_map_gtt_bo_to_gart(struct amdgpu_device *adev, struct amdgpu_bo *bo)
2109 {
2110 	int ret;
2111 
2112 	ret = amdgpu_bo_reserve(bo, true);
2113 	if (ret) {
2114 		pr_err("Failed to reserve bo. ret %d\n", ret);
2115 		goto err_reserve_bo_failed;
2116 	}
2117 
2118 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2119 	if (ret) {
2120 		pr_err("Failed to pin bo. ret %d\n", ret);
2121 		goto err_pin_bo_failed;
2122 	}
2123 
2124 	ret = amdgpu_ttm_alloc_gart(&bo->tbo);
2125 	if (ret) {
2126 		pr_err("Failed to bind bo to GART. ret %d\n", ret);
2127 		goto err_map_bo_gart_failed;
2128 	}
2129 
2130 	amdgpu_amdkfd_remove_eviction_fence(
2131 		bo, bo->vm_bo->vm->process_info->eviction_fence);
2132 
2133 	amdgpu_bo_unreserve(bo);
2134 
2135 	bo = amdgpu_bo_ref(bo);
2136 
2137 	return 0;
2138 
2139 err_map_bo_gart_failed:
2140 	amdgpu_bo_unpin(bo);
2141 err_pin_bo_failed:
2142 	amdgpu_bo_unreserve(bo);
2143 err_reserve_bo_failed:
2144 
2145 	return ret;
2146 }
2147 
2148 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Map a GTT BO for kernel CPU access
2149  *
2150  * @mem: Buffer object to be mapped for CPU access
2151  * @kptr[out]: pointer in kernel CPU address space
2152  * @size[out]: size of the buffer
2153  *
2154  * Pins the BO and maps it for kernel CPU access. The eviction fence is removed
2155  * from the BO, since pinned BOs cannot be evicted. The bo must remain on the
2156  * validate_list, so the GPU mapping can be restored after a page table was
2157  * evicted.
2158  *
2159  * Return: 0 on success, error code on failure
2160  */
2161 int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,
2162 					     void **kptr, uint64_t *size)
2163 {
2164 	int ret;
2165 	struct amdgpu_bo *bo = mem->bo;
2166 
2167 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
2168 		pr_err("userptr can't be mapped to kernel\n");
2169 		return -EINVAL;
2170 	}
2171 
2172 	mutex_lock(&mem->process_info->lock);
2173 
2174 	ret = amdgpu_bo_reserve(bo, true);
2175 	if (ret) {
2176 		pr_err("Failed to reserve bo. ret %d\n", ret);
2177 		goto bo_reserve_failed;
2178 	}
2179 
2180 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
2181 	if (ret) {
2182 		pr_err("Failed to pin bo. ret %d\n", ret);
2183 		goto pin_failed;
2184 	}
2185 
2186 	ret = amdgpu_bo_kmap(bo, kptr);
2187 	if (ret) {
2188 		pr_err("Failed to map bo to kernel. ret %d\n", ret);
2189 		goto kmap_failed;
2190 	}
2191 
2192 	amdgpu_amdkfd_remove_eviction_fence(
2193 		bo, mem->process_info->eviction_fence);
2194 
2195 	if (size)
2196 		*size = amdgpu_bo_size(bo);
2197 
2198 	amdgpu_bo_unreserve(bo);
2199 
2200 	mutex_unlock(&mem->process_info->lock);
2201 	return 0;
2202 
2203 kmap_failed:
2204 	amdgpu_bo_unpin(bo);
2205 pin_failed:
2206 	amdgpu_bo_unreserve(bo);
2207 bo_reserve_failed:
2208 	mutex_unlock(&mem->process_info->lock);
2209 
2210 	return ret;
2211 }
2212 
2213 /** amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel() - Unmap a GTT BO for kernel CPU access
2214  *
2215  * @mem: Buffer object to be unmapped for CPU access
2216  *
2217  * Removes the kernel CPU mapping and unpins the BO. It does not restore the
2218  * eviction fence, so this function should only be used for cleanup before the
2219  * BO is destroyed.
2220  */
2221 void amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(struct kgd_mem *mem)
2222 {
2223 	struct amdgpu_bo *bo = mem->bo;
2224 
2225 	amdgpu_bo_reserve(bo, true);
2226 	amdgpu_bo_kunmap(bo);
2227 	amdgpu_bo_unpin(bo);
2228 	amdgpu_bo_unreserve(bo);
2229 }
2230 
2231 int amdgpu_amdkfd_gpuvm_get_vm_fault_info(struct amdgpu_device *adev,
2232 					  struct kfd_vm_fault_info *mem)
2233 {
2234 	if (atomic_read(&adev->gmc.vm_fault_info_updated) == 1) {
2235 		*mem = *adev->gmc.vm_fault_info;
2236 		mb(); /* make sure read happened */
2237 		atomic_set(&adev->gmc.vm_fault_info_updated, 0);
2238 	}
2239 	return 0;
2240 }
2241 
2242 int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
2243 				      struct dma_buf *dma_buf,
2244 				      uint64_t va, void *drm_priv,
2245 				      struct kgd_mem **mem, uint64_t *size,
2246 				      uint64_t *mmap_offset)
2247 {
2248 	struct amdgpu_vm *avm = drm_priv_to_vm(drm_priv);
2249 	struct drm_gem_object *obj;
2250 	struct amdgpu_bo *bo;
2251 	int ret;
2252 
2253 	obj = amdgpu_gem_prime_import(adev_to_drm(adev), dma_buf);
2254 	if (IS_ERR(obj))
2255 		return PTR_ERR(obj);
2256 
2257 	bo = gem_to_amdgpu_bo(obj);
2258 	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
2259 				    AMDGPU_GEM_DOMAIN_GTT))) {
2260 		/* Only VRAM and GTT BOs are supported */
2261 		ret = -EINVAL;
2262 		goto err_put_obj;
2263 	}
2264 
2265 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2266 	if (!*mem) {
2267 		ret = -ENOMEM;
2268 		goto err_put_obj;
2269 	}
2270 
2271 	ret = drm_vma_node_allow(&obj->vma_node, drm_priv);
2272 	if (ret)
2273 		goto err_free_mem;
2274 
2275 	if (size)
2276 		*size = amdgpu_bo_size(bo);
2277 
2278 	if (mmap_offset)
2279 		*mmap_offset = amdgpu_bo_mmap_offset(bo);
2280 
2281 	INIT_LIST_HEAD(&(*mem)->attachments);
2282 	mutex_init(&(*mem)->lock);
2283 
2284 	(*mem)->alloc_flags =
2285 		((bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2286 		KFD_IOC_ALLOC_MEM_FLAGS_VRAM : KFD_IOC_ALLOC_MEM_FLAGS_GTT)
2287 		| KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE
2288 		| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
2289 
2290 	get_dma_buf(dma_buf);
2291 	(*mem)->dmabuf = dma_buf;
2292 	(*mem)->bo = bo;
2293 	(*mem)->va = va;
2294 	(*mem)->domain = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
2295 		AMDGPU_GEM_DOMAIN_VRAM : AMDGPU_GEM_DOMAIN_GTT;
2296 	(*mem)->mapped_to_gpu_memory = 0;
2297 	(*mem)->process_info = avm->process_info;
2298 	add_kgd_mem_to_kfd_bo_list(*mem, avm->process_info, false);
2299 	amdgpu_sync_create(&(*mem)->sync);
2300 	(*mem)->is_imported = true;
2301 
2302 	return 0;
2303 
2304 err_free_mem:
2305 	kfree(*mem);
2306 err_put_obj:
2307 	drm_gem_object_put(obj);
2308 	return ret;
2309 }
2310 
2311 int amdgpu_amdkfd_gpuvm_export_dmabuf(struct kgd_mem *mem,
2312 				      struct dma_buf **dma_buf)
2313 {
2314 	int ret;
2315 
2316 	mutex_lock(&mem->lock);
2317 	ret = kfd_mem_export_dmabuf(mem);
2318 	if (ret)
2319 		goto out;
2320 
2321 	get_dma_buf(mem->dmabuf);
2322 	*dma_buf = mem->dmabuf;
2323 out:
2324 	mutex_unlock(&mem->lock);
2325 	return ret;
2326 }
2327 
2328 /* Evict a userptr BO by stopping the queues if necessary
2329  *
2330  * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
2331  * cannot do any memory allocations, and cannot take any locks that
2332  * are held elsewhere while allocating memory.
2333  *
2334  * It doesn't do anything to the BO itself. The real work happens in
2335  * restore, where we get updated page addresses. This function only
2336  * ensures that GPU access to the BO is stopped.
2337  */
2338 int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
2339 				unsigned long cur_seq, struct kgd_mem *mem)
2340 {
2341 	struct amdkfd_process_info *process_info = mem->process_info;
2342 	int r = 0;
2343 
2344 	/* Do not process MMU notifications during CRIU restore until
2345 	 * KFD_CRIU_OP_RESUME IOCTL is received
2346 	 */
2347 	if (READ_ONCE(process_info->block_mmu_notifications))
2348 		return 0;
2349 
2350 	mutex_lock(&process_info->notifier_lock);
2351 	mmu_interval_set_seq(mni, cur_seq);
2352 
2353 	mem->invalid++;
2354 	if (++process_info->evicted_bos == 1) {
2355 		/* First eviction, stop the queues */
2356 		r = kgd2kfd_quiesce_mm(mni->mm,
2357 				       KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
2358 		if (r)
2359 			pr_err("Failed to quiesce KFD\n");
2360 		schedule_delayed_work(&process_info->restore_userptr_work,
2361 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2362 	}
2363 	mutex_unlock(&process_info->notifier_lock);
2364 
2365 	return r;
2366 }
2367 
2368 /* Update invalid userptr BOs
2369  *
2370  * Moves invalidated (evicted) userptr BOs from userptr_valid_list to
2371  * userptr_inval_list and updates user pages for all BOs that have
2372  * been invalidated since their last update.
2373  */
2374 static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
2375 				     struct mm_struct *mm)
2376 {
2377 	struct kgd_mem *mem, *tmp_mem;
2378 	struct amdgpu_bo *bo;
2379 	struct ttm_operation_ctx ctx = { false, false };
2380 	uint32_t invalid;
2381 	int ret = 0;
2382 
2383 	mutex_lock(&process_info->notifier_lock);
2384 
2385 	/* Move all invalidated BOs to the userptr_inval_list */
2386 	list_for_each_entry_safe(mem, tmp_mem,
2387 				 &process_info->userptr_valid_list,
2388 				 validate_list.head)
2389 		if (mem->invalid)
2390 			list_move_tail(&mem->validate_list.head,
2391 				       &process_info->userptr_inval_list);
2392 
2393 	/* Go through userptr_inval_list and update any invalid user_pages */
2394 	list_for_each_entry(mem, &process_info->userptr_inval_list,
2395 			    validate_list.head) {
2396 		invalid = mem->invalid;
2397 		if (!invalid)
2398 			/* BO hasn't been invalidated since the last
2399 			 * revalidation attempt. Keep its page list.
2400 			 */
2401 			continue;
2402 
2403 		bo = mem->bo;
2404 
2405 		amdgpu_ttm_tt_discard_user_pages(bo->tbo.ttm, mem->range);
2406 		mem->range = NULL;
2407 
2408 		/* BO reservations and getting user pages (hmm_range_fault)
2409 		 * must happen outside the notifier lock
2410 		 */
2411 		mutex_unlock(&process_info->notifier_lock);
2412 
2413 		/* Move the BO to system (CPU) domain if necessary to unmap
2414 		 * and free the SG table
2415 		 */
2416 		if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) {
2417 			if (amdgpu_bo_reserve(bo, true))
2418 				return -EAGAIN;
2419 			amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
2420 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2421 			amdgpu_bo_unreserve(bo);
2422 			if (ret) {
2423 				pr_err("%s: Failed to invalidate userptr BO\n",
2424 				       __func__);
2425 				return -EAGAIN;
2426 			}
2427 		}
2428 
2429 		/* Get updated user pages */
2430 		ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages,
2431 						   &mem->range);
2432 		if (ret) {
2433 			pr_debug("Failed %d to get user pages\n", ret);
2434 
2435 			/* Return -EFAULT bad address error as success. It will
2436 			 * fail later with a VM fault if the GPU tries to access
2437 			 * it. Better than hanging indefinitely with stalled
2438 			 * user mode queues.
2439 			 *
2440 			 * Return other error -EBUSY or -ENOMEM to retry restore
2441 			 */
2442 			if (ret != -EFAULT)
2443 				return ret;
2444 
2445 			ret = 0;
2446 		}
2447 
2448 		mutex_lock(&process_info->notifier_lock);
2449 
2450 		/* Mark the BO as valid unless it was invalidated
2451 		 * again concurrently.
2452 		 */
2453 		if (mem->invalid != invalid) {
2454 			ret = -EAGAIN;
2455 			goto unlock_out;
2456 		}
2457 		 /* set mem valid if mem has hmm range associated */
2458 		if (mem->range)
2459 			mem->invalid = 0;
2460 	}
2461 
2462 unlock_out:
2463 	mutex_unlock(&process_info->notifier_lock);
2464 
2465 	return ret;
2466 }
2467 
2468 /* Validate invalid userptr BOs
2469  *
2470  * Validates BOs on the userptr_inval_list. Also updates GPUVM page tables
2471  * with new page addresses and waits for the page table updates to complete.
2472  */
2473 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
2474 {
2475 	struct amdgpu_bo_list_entry *pd_bo_list_entries;
2476 	struct list_head resv_list, duplicates;
2477 	struct ww_acquire_ctx ticket;
2478 	struct amdgpu_sync sync;
2479 
2480 	struct amdgpu_vm *peer_vm;
2481 	struct kgd_mem *mem, *tmp_mem;
2482 	struct amdgpu_bo *bo;
2483 	struct ttm_operation_ctx ctx = { false, false };
2484 	int i, ret;
2485 
2486 	pd_bo_list_entries = kcalloc(process_info->n_vms,
2487 				     sizeof(struct amdgpu_bo_list_entry),
2488 				     GFP_KERNEL);
2489 	if (!pd_bo_list_entries) {
2490 		pr_err("%s: Failed to allocate PD BO list entries\n", __func__);
2491 		ret = -ENOMEM;
2492 		goto out_no_mem;
2493 	}
2494 
2495 	INIT_LIST_HEAD(&resv_list);
2496 	INIT_LIST_HEAD(&duplicates);
2497 
2498 	/* Get all the page directory BOs that need to be reserved */
2499 	i = 0;
2500 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2501 			    vm_list_node)
2502 		amdgpu_vm_get_pd_bo(peer_vm, &resv_list,
2503 				    &pd_bo_list_entries[i++]);
2504 	/* Add the userptr_inval_list entries to resv_list */
2505 	list_for_each_entry(mem, &process_info->userptr_inval_list,
2506 			    validate_list.head) {
2507 		list_add_tail(&mem->resv_list.head, &resv_list);
2508 		mem->resv_list.bo = mem->validate_list.bo;
2509 		mem->resv_list.num_shared = mem->validate_list.num_shared;
2510 	}
2511 
2512 	/* Reserve all BOs and page tables for validation */
2513 	ret = ttm_eu_reserve_buffers(&ticket, &resv_list, false, &duplicates);
2514 	WARN(!list_empty(&duplicates), "Duplicates should be empty");
2515 	if (ret)
2516 		goto out_free;
2517 
2518 	amdgpu_sync_create(&sync);
2519 
2520 	ret = process_validate_vms(process_info);
2521 	if (ret)
2522 		goto unreserve_out;
2523 
2524 	/* Validate BOs and update GPUVM page tables */
2525 	list_for_each_entry_safe(mem, tmp_mem,
2526 				 &process_info->userptr_inval_list,
2527 				 validate_list.head) {
2528 		struct kfd_mem_attachment *attachment;
2529 
2530 		bo = mem->bo;
2531 
2532 		/* Validate the BO if we got user pages */
2533 		if (bo->tbo.ttm->pages[0]) {
2534 			amdgpu_bo_placement_from_domain(bo, mem->domain);
2535 			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
2536 			if (ret) {
2537 				pr_err("%s: failed to validate BO\n", __func__);
2538 				goto unreserve_out;
2539 			}
2540 		}
2541 
2542 		/* Update mapping. If the BO was not validated
2543 		 * (because we couldn't get user pages), this will
2544 		 * clear the page table entries, which will result in
2545 		 * VM faults if the GPU tries to access the invalid
2546 		 * memory.
2547 		 */
2548 		list_for_each_entry(attachment, &mem->attachments, list) {
2549 			if (!attachment->is_mapped)
2550 				continue;
2551 
2552 			kfd_mem_dmaunmap_attachment(mem, attachment);
2553 			ret = update_gpuvm_pte(mem, attachment, &sync);
2554 			if (ret) {
2555 				pr_err("%s: update PTE failed\n", __func__);
2556 				/* make sure this gets validated again */
2557 				mutex_lock(&process_info->notifier_lock);
2558 				mem->invalid++;
2559 				mutex_unlock(&process_info->notifier_lock);
2560 				goto unreserve_out;
2561 			}
2562 		}
2563 	}
2564 
2565 	/* Update page directories */
2566 	ret = process_update_pds(process_info, &sync);
2567 
2568 unreserve_out:
2569 	ttm_eu_backoff_reservation(&ticket, &resv_list);
2570 	amdgpu_sync_wait(&sync, false);
2571 	amdgpu_sync_free(&sync);
2572 out_free:
2573 	kfree(pd_bo_list_entries);
2574 out_no_mem:
2575 
2576 	return ret;
2577 }
2578 
2579 /* Confirm that all user pages are valid while holding the notifier lock
2580  *
2581  * Moves valid BOs from the userptr_inval_list back to userptr_val_list.
2582  */
2583 static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_info)
2584 {
2585 	struct kgd_mem *mem, *tmp_mem;
2586 	int ret = 0;
2587 
2588 	list_for_each_entry_safe(mem, tmp_mem,
2589 				 &process_info->userptr_inval_list,
2590 				 validate_list.head) {
2591 		bool valid;
2592 
2593 		/* keep mem without hmm range at userptr_inval_list */
2594 		if (!mem->range)
2595 			 continue;
2596 
2597 		/* Only check mem with hmm range associated */
2598 		valid = amdgpu_ttm_tt_get_user_pages_done(
2599 					mem->bo->tbo.ttm, mem->range);
2600 
2601 		mem->range = NULL;
2602 		if (!valid) {
2603 			WARN(!mem->invalid, "Invalid BO not marked invalid");
2604 			ret = -EAGAIN;
2605 			continue;
2606 		}
2607 
2608 		if (mem->invalid) {
2609 			WARN(1, "Valid BO is marked invalid");
2610 			ret = -EAGAIN;
2611 			continue;
2612 		}
2613 
2614 		list_move_tail(&mem->validate_list.head,
2615 			       &process_info->userptr_valid_list);
2616 	}
2617 
2618 	return ret;
2619 }
2620 
2621 /* Worker callback to restore evicted userptr BOs
2622  *
2623  * Tries to update and validate all userptr BOs. If successful and no
2624  * concurrent evictions happened, the queues are restarted. Otherwise,
2625  * reschedule for another attempt later.
2626  */
2627 static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
2628 {
2629 	struct delayed_work *dwork = to_delayed_work(work);
2630 	struct amdkfd_process_info *process_info =
2631 		container_of(dwork, struct amdkfd_process_info,
2632 			     restore_userptr_work);
2633 	struct task_struct *usertask;
2634 	struct mm_struct *mm;
2635 	uint32_t evicted_bos;
2636 
2637 	mutex_lock(&process_info->notifier_lock);
2638 	evicted_bos = process_info->evicted_bos;
2639 	mutex_unlock(&process_info->notifier_lock);
2640 	if (!evicted_bos)
2641 		return;
2642 
2643 	/* Reference task and mm in case of concurrent process termination */
2644 	usertask = get_pid_task(process_info->pid, PIDTYPE_PID);
2645 	if (!usertask)
2646 		return;
2647 	mm = get_task_mm(usertask);
2648 	if (!mm) {
2649 		put_task_struct(usertask);
2650 		return;
2651 	}
2652 
2653 	mutex_lock(&process_info->lock);
2654 
2655 	if (update_invalid_user_pages(process_info, mm))
2656 		goto unlock_out;
2657 	/* userptr_inval_list can be empty if all evicted userptr BOs
2658 	 * have been freed. In that case there is nothing to validate
2659 	 * and we can just restart the queues.
2660 	 */
2661 	if (!list_empty(&process_info->userptr_inval_list)) {
2662 		if (validate_invalid_user_pages(process_info))
2663 			goto unlock_out;
2664 	}
2665 	/* Final check for concurrent evicton and atomic update. If
2666 	 * another eviction happens after successful update, it will
2667 	 * be a first eviction that calls quiesce_mm. The eviction
2668 	 * reference counting inside KFD will handle this case.
2669 	 */
2670 	mutex_lock(&process_info->notifier_lock);
2671 	if (process_info->evicted_bos != evicted_bos)
2672 		goto unlock_notifier_out;
2673 
2674 	if (confirm_valid_user_pages_locked(process_info)) {
2675 		WARN(1, "User pages unexpectedly invalid");
2676 		goto unlock_notifier_out;
2677 	}
2678 
2679 	process_info->evicted_bos = evicted_bos = 0;
2680 
2681 	if (kgd2kfd_resume_mm(mm)) {
2682 		pr_err("%s: Failed to resume KFD\n", __func__);
2683 		/* No recovery from this failure. Probably the CP is
2684 		 * hanging. No point trying again.
2685 		 */
2686 	}
2687 
2688 unlock_notifier_out:
2689 	mutex_unlock(&process_info->notifier_lock);
2690 unlock_out:
2691 	mutex_unlock(&process_info->lock);
2692 
2693 	/* If validation failed, reschedule another attempt */
2694 	if (evicted_bos) {
2695 		schedule_delayed_work(&process_info->restore_userptr_work,
2696 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
2697 
2698 		kfd_smi_event_queue_restore_rescheduled(mm);
2699 	}
2700 	mmput(mm);
2701 	put_task_struct(usertask);
2702 }
2703 
2704 /** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
2705  *   KFD process identified by process_info
2706  *
2707  * @process_info: amdkfd_process_info of the KFD process
2708  *
2709  * After memory eviction, restore thread calls this function. The function
2710  * should be called when the Process is still valid. BO restore involves -
2711  *
2712  * 1.  Release old eviction fence and create new one
2713  * 2.  Get two copies of PD BO list from all the VMs. Keep one copy as pd_list.
2714  * 3   Use the second PD list and kfd_bo_list to create a list (ctx.list) of
2715  *     BOs that need to be reserved.
2716  * 4.  Reserve all the BOs
2717  * 5.  Validate of PD and PT BOs.
2718  * 6.  Validate all KFD BOs using kfd_bo_list and Map them and add new fence
2719  * 7.  Add fence to all PD and PT BOs.
2720  * 8.  Unreserve all BOs
2721  */
2722 int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence **ef)
2723 {
2724 	struct amdgpu_bo_list_entry *pd_bo_list;
2725 	struct amdkfd_process_info *process_info = info;
2726 	struct amdgpu_vm *peer_vm;
2727 	struct kgd_mem *mem;
2728 	struct bo_vm_reservation_context ctx;
2729 	struct amdgpu_amdkfd_fence *new_fence;
2730 	int ret = 0, i;
2731 	struct list_head duplicate_save;
2732 	struct amdgpu_sync sync_obj;
2733 	unsigned long failed_size = 0;
2734 	unsigned long total_size = 0;
2735 
2736 	INIT_LIST_HEAD(&duplicate_save);
2737 	INIT_LIST_HEAD(&ctx.list);
2738 	INIT_LIST_HEAD(&ctx.duplicates);
2739 
2740 	pd_bo_list = kcalloc(process_info->n_vms,
2741 			     sizeof(struct amdgpu_bo_list_entry),
2742 			     GFP_KERNEL);
2743 	if (!pd_bo_list)
2744 		return -ENOMEM;
2745 
2746 	i = 0;
2747 	mutex_lock(&process_info->lock);
2748 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2749 			vm_list_node)
2750 		amdgpu_vm_get_pd_bo(peer_vm, &ctx.list, &pd_bo_list[i++]);
2751 
2752 	/* Reserve all BOs and page tables/directory. Add all BOs from
2753 	 * kfd_bo_list to ctx.list
2754 	 */
2755 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2756 			    validate_list.head) {
2757 
2758 		list_add_tail(&mem->resv_list.head, &ctx.list);
2759 		mem->resv_list.bo = mem->validate_list.bo;
2760 		mem->resv_list.num_shared = mem->validate_list.num_shared;
2761 	}
2762 
2763 	ret = ttm_eu_reserve_buffers(&ctx.ticket, &ctx.list,
2764 				     false, &duplicate_save);
2765 	if (ret) {
2766 		pr_debug("Memory eviction: TTM Reserve Failed. Try again\n");
2767 		goto ttm_reserve_fail;
2768 	}
2769 
2770 	amdgpu_sync_create(&sync_obj);
2771 
2772 	/* Validate PDs and PTs */
2773 	ret = process_validate_vms(process_info);
2774 	if (ret)
2775 		goto validate_map_fail;
2776 
2777 	ret = process_sync_pds_resv(process_info, &sync_obj);
2778 	if (ret) {
2779 		pr_debug("Memory eviction: Failed to sync to PD BO moving fence. Try again\n");
2780 		goto validate_map_fail;
2781 	}
2782 
2783 	/* Validate BOs and map them to GPUVM (update VM page tables). */
2784 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2785 			    validate_list.head) {
2786 
2787 		struct amdgpu_bo *bo = mem->bo;
2788 		uint32_t domain = mem->domain;
2789 		struct kfd_mem_attachment *attachment;
2790 		struct dma_resv_iter cursor;
2791 		struct dma_fence *fence;
2792 
2793 		total_size += amdgpu_bo_size(bo);
2794 
2795 		ret = amdgpu_amdkfd_bo_validate(bo, domain, false);
2796 		if (ret) {
2797 			pr_debug("Memory eviction: Validate BOs failed\n");
2798 			failed_size += amdgpu_bo_size(bo);
2799 			ret = amdgpu_amdkfd_bo_validate(bo,
2800 						AMDGPU_GEM_DOMAIN_GTT, false);
2801 			if (ret) {
2802 				pr_debug("Memory eviction: Try again\n");
2803 				goto validate_map_fail;
2804 			}
2805 		}
2806 		dma_resv_for_each_fence(&cursor, bo->tbo.base.resv,
2807 					DMA_RESV_USAGE_KERNEL, fence) {
2808 			ret = amdgpu_sync_fence(&sync_obj, fence);
2809 			if (ret) {
2810 				pr_debug("Memory eviction: Sync BO fence failed. Try again\n");
2811 				goto validate_map_fail;
2812 			}
2813 		}
2814 		list_for_each_entry(attachment, &mem->attachments, list) {
2815 			if (!attachment->is_mapped)
2816 				continue;
2817 
2818 			kfd_mem_dmaunmap_attachment(mem, attachment);
2819 			ret = update_gpuvm_pte(mem, attachment, &sync_obj);
2820 			if (ret) {
2821 				pr_debug("Memory eviction: update PTE failed. Try again\n");
2822 				goto validate_map_fail;
2823 			}
2824 		}
2825 	}
2826 
2827 	if (failed_size)
2828 		pr_debug("0x%lx/0x%lx in system\n", failed_size, total_size);
2829 
2830 	/* Update page directories */
2831 	ret = process_update_pds(process_info, &sync_obj);
2832 	if (ret) {
2833 		pr_debug("Memory eviction: update PDs failed. Try again\n");
2834 		goto validate_map_fail;
2835 	}
2836 
2837 	/* Wait for validate and PT updates to finish */
2838 	amdgpu_sync_wait(&sync_obj, false);
2839 
2840 	/* Release old eviction fence and create new one, because fence only
2841 	 * goes from unsignaled to signaled, fence cannot be reused.
2842 	 * Use context and mm from the old fence.
2843 	 */
2844 	new_fence = amdgpu_amdkfd_fence_create(
2845 				process_info->eviction_fence->base.context,
2846 				process_info->eviction_fence->mm,
2847 				NULL);
2848 	if (!new_fence) {
2849 		pr_err("Failed to create eviction fence\n");
2850 		ret = -ENOMEM;
2851 		goto validate_map_fail;
2852 	}
2853 	dma_fence_put(&process_info->eviction_fence->base);
2854 	process_info->eviction_fence = new_fence;
2855 	*ef = dma_fence_get(&new_fence->base);
2856 
2857 	/* Attach new eviction fence to all BOs except pinned ones */
2858 	list_for_each_entry(mem, &process_info->kfd_bo_list,
2859 		validate_list.head) {
2860 		if (mem->bo->tbo.pin_count)
2861 			continue;
2862 
2863 		dma_resv_add_fence(mem->bo->tbo.base.resv,
2864 				   &process_info->eviction_fence->base,
2865 				   DMA_RESV_USAGE_BOOKKEEP);
2866 	}
2867 	/* Attach eviction fence to PD / PT BOs */
2868 	list_for_each_entry(peer_vm, &process_info->vm_list_head,
2869 			    vm_list_node) {
2870 		struct amdgpu_bo *bo = peer_vm->root.bo;
2871 
2872 		dma_resv_add_fence(bo->tbo.base.resv,
2873 				   &process_info->eviction_fence->base,
2874 				   DMA_RESV_USAGE_BOOKKEEP);
2875 	}
2876 
2877 validate_map_fail:
2878 	ttm_eu_backoff_reservation(&ctx.ticket, &ctx.list);
2879 	amdgpu_sync_free(&sync_obj);
2880 ttm_reserve_fail:
2881 	mutex_unlock(&process_info->lock);
2882 	kfree(pd_bo_list);
2883 	return ret;
2884 }
2885 
2886 int amdgpu_amdkfd_add_gws_to_process(void *info, void *gws, struct kgd_mem **mem)
2887 {
2888 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2889 	struct amdgpu_bo *gws_bo = (struct amdgpu_bo *)gws;
2890 	int ret;
2891 
2892 	if (!info || !gws)
2893 		return -EINVAL;
2894 
2895 	*mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
2896 	if (!*mem)
2897 		return -ENOMEM;
2898 
2899 	mutex_init(&(*mem)->lock);
2900 	INIT_LIST_HEAD(&(*mem)->attachments);
2901 	(*mem)->bo = amdgpu_bo_ref(gws_bo);
2902 	(*mem)->domain = AMDGPU_GEM_DOMAIN_GWS;
2903 	(*mem)->process_info = process_info;
2904 	add_kgd_mem_to_kfd_bo_list(*mem, process_info, false);
2905 	amdgpu_sync_create(&(*mem)->sync);
2906 
2907 
2908 	/* Validate gws bo the first time it is added to process */
2909 	mutex_lock(&(*mem)->process_info->lock);
2910 	ret = amdgpu_bo_reserve(gws_bo, false);
2911 	if (unlikely(ret)) {
2912 		pr_err("Reserve gws bo failed %d\n", ret);
2913 		goto bo_reservation_failure;
2914 	}
2915 
2916 	ret = amdgpu_amdkfd_bo_validate(gws_bo, AMDGPU_GEM_DOMAIN_GWS, true);
2917 	if (ret) {
2918 		pr_err("GWS BO validate failed %d\n", ret);
2919 		goto bo_validation_failure;
2920 	}
2921 	/* GWS resource is shared b/t amdgpu and amdkfd
2922 	 * Add process eviction fence to bo so they can
2923 	 * evict each other.
2924 	 */
2925 	ret = dma_resv_reserve_fences(gws_bo->tbo.base.resv, 1);
2926 	if (ret)
2927 		goto reserve_shared_fail;
2928 	dma_resv_add_fence(gws_bo->tbo.base.resv,
2929 			   &process_info->eviction_fence->base,
2930 			   DMA_RESV_USAGE_BOOKKEEP);
2931 	amdgpu_bo_unreserve(gws_bo);
2932 	mutex_unlock(&(*mem)->process_info->lock);
2933 
2934 	return ret;
2935 
2936 reserve_shared_fail:
2937 bo_validation_failure:
2938 	amdgpu_bo_unreserve(gws_bo);
2939 bo_reservation_failure:
2940 	mutex_unlock(&(*mem)->process_info->lock);
2941 	amdgpu_sync_free(&(*mem)->sync);
2942 	remove_kgd_mem_from_kfd_bo_list(*mem, process_info);
2943 	amdgpu_bo_unref(&gws_bo);
2944 	mutex_destroy(&(*mem)->lock);
2945 	kfree(*mem);
2946 	*mem = NULL;
2947 	return ret;
2948 }
2949 
2950 int amdgpu_amdkfd_remove_gws_from_process(void *info, void *mem)
2951 {
2952 	int ret;
2953 	struct amdkfd_process_info *process_info = (struct amdkfd_process_info *)info;
2954 	struct kgd_mem *kgd_mem = (struct kgd_mem *)mem;
2955 	struct amdgpu_bo *gws_bo = kgd_mem->bo;
2956 
2957 	/* Remove BO from process's validate list so restore worker won't touch
2958 	 * it anymore
2959 	 */
2960 	remove_kgd_mem_from_kfd_bo_list(kgd_mem, process_info);
2961 
2962 	ret = amdgpu_bo_reserve(gws_bo, false);
2963 	if (unlikely(ret)) {
2964 		pr_err("Reserve gws bo failed %d\n", ret);
2965 		//TODO add BO back to validate_list?
2966 		return ret;
2967 	}
2968 	amdgpu_amdkfd_remove_eviction_fence(gws_bo,
2969 			process_info->eviction_fence);
2970 	amdgpu_bo_unreserve(gws_bo);
2971 	amdgpu_sync_free(&kgd_mem->sync);
2972 	amdgpu_bo_unref(&gws_bo);
2973 	mutex_destroy(&kgd_mem->lock);
2974 	kfree(mem);
2975 	return 0;
2976 }
2977 
2978 /* Returns GPU-specific tiling mode information */
2979 int amdgpu_amdkfd_get_tile_config(struct amdgpu_device *adev,
2980 				struct tile_config *config)
2981 {
2982 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
2983 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
2984 	config->num_tile_configs =
2985 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
2986 	config->macro_tile_config_ptr =
2987 			adev->gfx.config.macrotile_mode_array;
2988 	config->num_macro_tile_configs =
2989 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
2990 
2991 	/* Those values are not set from GFX9 onwards */
2992 	config->num_banks = adev->gfx.config.num_banks;
2993 	config->num_ranks = adev->gfx.config.num_ranks;
2994 
2995 	return 0;
2996 }
2997 
2998 bool amdgpu_amdkfd_bo_mapped_to_dev(struct amdgpu_device *adev, struct kgd_mem *mem)
2999 {
3000 	struct kfd_mem_attachment *entry;
3001 
3002 	list_for_each_entry(entry, &mem->attachments, list) {
3003 		if (entry->is_mapped && entry->adev == adev)
3004 			return true;
3005 	}
3006 	return false;
3007 }
3008 
3009 #if defined(CONFIG_DEBUG_FS)
3010 
3011 int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data)
3012 {
3013 
3014 	spin_lock(&kfd_mem_limit.mem_limit_lock);
3015 	seq_printf(m, "System mem used %lldM out of %lluM\n",
3016 		  (kfd_mem_limit.system_mem_used >> 20),
3017 		  (kfd_mem_limit.max_system_mem_limit >> 20));
3018 	seq_printf(m, "TTM mem used %lldM out of %lluM\n",
3019 		  (kfd_mem_limit.ttm_mem_used >> 20),
3020 		  (kfd_mem_limit.max_ttm_mem_limit >> 20));
3021 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
3022 
3023 	return 0;
3024 }
3025 
3026 #endif
3027