xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c (revision 8e65320d91cdc3b241d4b94855c88459b91abf66)
1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 /*
27  * Authors:
28  *    Christian König <christian.koenig@amd.com>
29  */
30 
31 /**
32  * DOC: MMU Notifier
33  *
34  * For coherent userptr handling registers an MMU notifier to inform the driver
35  * about updates on the page tables of a process.
36  *
37  * When somebody tries to invalidate the page tables we block the update until
38  * all operations on the pages in question are completed, then those pages are
39  * marked as accessed and also dirty if it wasn't a read only access.
40  *
41  * New command submissions using the userptrs in question are delayed until all
42  * page table invalidation are completed and we once more see a coherent process
43  * address space.
44  */
45 
46 #include <linux/firmware.h>
47 #include <linux/module.h>
48 #include <drm/drm.h>
49 
50 #include "amdgpu.h"
51 #include "amdgpu_amdkfd.h"
52 #include "amdgpu_hmm.h"
53 
54 /**
55  * amdgpu_hmm_invalidate_gfx - callback to notify about mm change
56  *
57  * @mni: the range (mm) is about to update
58  * @range: details on the invalidation
59  * @cur_seq: Value to pass to mmu_interval_set_seq()
60  *
61  * Block for operations on BOs to finish and mark pages as accessed and
62  * potentially dirty.
63  */
amdgpu_hmm_invalidate_gfx(struct mmu_interval_notifier * mni,const struct mmu_notifier_range * range,unsigned long cur_seq)64 static bool amdgpu_hmm_invalidate_gfx(struct mmu_interval_notifier *mni,
65 				      const struct mmu_notifier_range *range,
66 				      unsigned long cur_seq)
67 {
68 	struct amdgpu_bo *bo = container_of(mni, struct amdgpu_bo, notifier);
69 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
70 	struct amdgpu_bo *vm_root = bo->vm_bo->vm->root.bo;
71 	long r;
72 
73 	if (!mmu_notifier_range_blockable(range))
74 		return false;
75 
76 	mutex_lock(&adev->notifier_lock);
77 
78 	mmu_interval_set_seq(mni, cur_seq);
79 
80 	amdgpu_vm_bo_invalidate(bo, false);
81 	r = dma_resv_wait_timeout(vm_root->tbo.base.resv,
82 				  DMA_RESV_USAGE_BOOKKEEP, false,
83 				  MAX_SCHEDULE_TIMEOUT);
84 	mutex_unlock(&adev->notifier_lock);
85 	if (r <= 0)
86 		DRM_ERROR("(%ld) failed to wait for user bo\n", r);
87 	return true;
88 }
89 
90 static const struct mmu_interval_notifier_ops amdgpu_hmm_gfx_ops = {
91 	.invalidate = amdgpu_hmm_invalidate_gfx,
92 };
93 
94 /**
95  * amdgpu_hmm_invalidate_hsa - callback to notify about mm change
96  *
97  * @mni: the range (mm) is about to update
98  * @range: details on the invalidation
99  * @cur_seq: Value to pass to mmu_interval_set_seq()
100  *
101  * We temporarily evict the BO attached to this range. This necessitates
102  * evicting all user-mode queues of the process.
103  */
amdgpu_hmm_invalidate_hsa(struct mmu_interval_notifier * mni,const struct mmu_notifier_range * range,unsigned long cur_seq)104 static bool amdgpu_hmm_invalidate_hsa(struct mmu_interval_notifier *mni,
105 				      const struct mmu_notifier_range *range,
106 				      unsigned long cur_seq)
107 {
108 	struct amdgpu_bo *bo = container_of(mni, struct amdgpu_bo, notifier);
109 
110 	if (!mmu_notifier_range_blockable(range))
111 		return false;
112 
113 	amdgpu_amdkfd_evict_userptr(mni, cur_seq, bo->kfd_bo);
114 
115 	return true;
116 }
117 
118 static const struct mmu_interval_notifier_ops amdgpu_hmm_hsa_ops = {
119 	.invalidate = amdgpu_hmm_invalidate_hsa,
120 };
121 
122 /**
123  * amdgpu_hmm_register - register a BO for notifier updates
124  *
125  * @bo: amdgpu buffer object
126  * @addr: userptr addr we should monitor
127  *
128  * Registers a mmu_notifier for the given BO at the specified address.
129  * Returns 0 on success, -ERRNO if anything goes wrong.
130  */
amdgpu_hmm_register(struct amdgpu_bo * bo,unsigned long addr)131 int amdgpu_hmm_register(struct amdgpu_bo *bo, unsigned long addr)
132 {
133 	int r;
134 
135 	if (bo->kfd_bo)
136 		r = mmu_interval_notifier_insert(&bo->notifier, current->mm,
137 						    addr, amdgpu_bo_size(bo),
138 						    &amdgpu_hmm_hsa_ops);
139 	else
140 		r = mmu_interval_notifier_insert(&bo->notifier, current->mm, addr,
141 							amdgpu_bo_size(bo),
142 							&amdgpu_hmm_gfx_ops);
143 	if (r)
144 		/*
145 		 * Make sure amdgpu_hmm_unregister() doesn't call
146 		 * mmu_interval_notifier_remove() when the notifier isn't properly
147 		 * initialized.
148 		 */
149 		bo->notifier.mm = NULL;
150 
151 	return r;
152 }
153 
154 /**
155  * amdgpu_hmm_unregister - unregister a BO for notifier updates
156  *
157  * @bo: amdgpu buffer object
158  *
159  * Remove any registration of mmu notifier updates from the buffer object.
160  */
amdgpu_hmm_unregister(struct amdgpu_bo * bo)161 void amdgpu_hmm_unregister(struct amdgpu_bo *bo)
162 {
163 	if (!bo->notifier.mm)
164 		return;
165 	mmu_interval_notifier_remove(&bo->notifier);
166 	bo->notifier.mm = NULL;
167 }
168 
amdgpu_hmm_range_get_pages(struct mmu_interval_notifier * notifier,uint64_t start,uint64_t npages,bool readonly,void * owner,struct amdgpu_hmm_range * range)169 int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier,
170 			       uint64_t start, uint64_t npages, bool readonly,
171 			       void *owner,
172 			       struct amdgpu_hmm_range *range)
173 {
174 	const u64 max_bytes = SZ_2G;
175 
176 	struct hmm_range *hmm_range = &range->hmm_range;
177 	unsigned long timeout;
178 	unsigned long *pfns;
179 	unsigned long end;
180 	int r;
181 
182 	pfns = kvmalloc_array(npages, sizeof(*pfns), GFP_KERNEL);
183 	if (unlikely(!pfns)) {
184 		r = -ENOMEM;
185 		goto out_free_range;
186 	}
187 
188 	hmm_range->notifier = notifier;
189 	hmm_range->default_flags = HMM_PFN_REQ_FAULT;
190 	if (!readonly)
191 		hmm_range->default_flags |= HMM_PFN_REQ_WRITE;
192 	hmm_range->hmm_pfns = pfns;
193 	hmm_range->start = start;
194 	end = start + npages * PAGE_SIZE;
195 	hmm_range->dev_private_owner = owner;
196 
197 	hmm_range->notifier_seq = mmu_interval_read_begin(notifier);
198 	do {
199 		hmm_range->end = min(hmm_range->start + max_bytes, end);
200 
201 		pr_debug("hmm range: start = 0x%lx, end = 0x%lx",
202 			hmm_range->start, hmm_range->end);
203 
204 		timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
205 
206 retry:
207 		r = hmm_range_fault(hmm_range);
208 		if (unlikely(r)) {
209 			if (r == -EBUSY && !time_after(jiffies, timeout))
210 				goto retry;
211 			goto out_free_pfns;
212 		}
213 
214 		if (hmm_range->end == end)
215 			break;
216 		hmm_range->hmm_pfns += max_bytes >> PAGE_SHIFT;
217 		hmm_range->start = hmm_range->end;
218 	} while (hmm_range->end < end);
219 
220 	hmm_range->start = start;
221 	hmm_range->hmm_pfns = pfns;
222 
223 	return 0;
224 
225 out_free_pfns:
226 	kvfree(pfns);
227 	hmm_range->hmm_pfns = NULL;
228 out_free_range:
229 	if (r == -EBUSY)
230 		r = -EAGAIN;
231 	return r;
232 }
233 
234 /**
235  * amdgpu_hmm_range_valid - check if an HMM range is still valid
236  * @range: pointer to the &struct amdgpu_hmm_range to validate
237  *
238  * Determines whether the given HMM range @range is still valid by
239  * checking for invalidations via the MMU notifier sequence. This is
240  * typically used to verify that the range has not been invalidated
241  * by concurrent address space updates before it is accessed.
242  *
243  * Return:
244  * * true if @range is valid and can be used safely
245  * * false if @range is NULL or has been invalidated
246  */
amdgpu_hmm_range_valid(struct amdgpu_hmm_range * range)247 bool amdgpu_hmm_range_valid(struct amdgpu_hmm_range *range)
248 {
249 	if (!range)
250 		return false;
251 
252 	return !mmu_interval_read_retry(range->hmm_range.notifier,
253 					range->hmm_range.notifier_seq);
254 }
255 
256 /**
257  * amdgpu_hmm_range_alloc - allocate and initialize an AMDGPU HMM range
258  * @bo: optional buffer object to associate with this HMM range
259  *
260  * Allocates memory for amdgpu_hmm_range and associates it with the @bo passed.
261  * The reference count of the @bo is incremented.
262  *
263  * Return:
264  * Pointer to a newly allocated struct amdgpu_hmm_range on success,
265  * or NULL if memory allocation fails.
266  */
amdgpu_hmm_range_alloc(struct amdgpu_bo * bo)267 struct amdgpu_hmm_range *amdgpu_hmm_range_alloc(struct amdgpu_bo *bo)
268 {
269 	struct amdgpu_hmm_range *range;
270 
271 	range = kzalloc_obj(*range);
272 	if (!range)
273 		return NULL;
274 
275 	range->bo = amdgpu_bo_ref(bo);
276 	return range;
277 }
278 
279 /**
280  * amdgpu_hmm_range_free - release an AMDGPU HMM range
281  * @range: pointer to the range object to free
282  *
283  * Releases all resources held by @range, including the associated
284  * hmm_pfns and the dropping reference of associated bo if any.
285  *
286  * Return: void
287  */
amdgpu_hmm_range_free(struct amdgpu_hmm_range * range)288 void amdgpu_hmm_range_free(struct amdgpu_hmm_range *range)
289 {
290 	if (!range)
291 		return;
292 
293 	kvfree(range->hmm_range.hmm_pfns);
294 	amdgpu_bo_unref(&range->bo);
295 	kfree(range);
296 }
297