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 
53 /**
54  * amdgpu_mn_invalidate_gfx - callback to notify about mm change
55  *
56  * @mni: the range (mm) is about to update
57  * @range: details on the invalidation
58  * @cur_seq: Value to pass to mmu_interval_set_seq()
59  *
60  * Block for operations on BOs to finish and mark pages as accessed and
61  * potentially dirty.
62  */
amdgpu_mn_invalidate_gfx(struct mmu_interval_notifier * mni,const struct mmu_notifier_range * range,unsigned long cur_seq)63 static bool amdgpu_mn_invalidate_gfx(struct mmu_interval_notifier *mni,
64 				     const struct mmu_notifier_range *range,
65 				     unsigned long cur_seq)
66 {
67 	struct amdgpu_bo *bo = container_of(mni, struct amdgpu_bo, notifier);
68 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
69 	long r;
70 
71 	if (!mmu_notifier_range_blockable(range))
72 		return false;
73 
74 	mutex_lock(&adev->notifier_lock);
75 
76 	mmu_interval_set_seq(mni, cur_seq);
77 
78 	r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP,
79 				  false, MAX_SCHEDULE_TIMEOUT);
80 	mutex_unlock(&adev->notifier_lock);
81 	if (r <= 0)
82 		DRM_ERROR("(%ld) failed to wait for user bo\n", r);
83 	return true;
84 }
85 
86 static const struct mmu_interval_notifier_ops amdgpu_mn_gfx_ops = {
87 	.invalidate = amdgpu_mn_invalidate_gfx,
88 };
89 
90 /**
91  * amdgpu_mn_invalidate_hsa - callback to notify about mm change
92  *
93  * @mni: the range (mm) is about to update
94  * @range: details on the invalidation
95  * @cur_seq: Value to pass to mmu_interval_set_seq()
96  *
97  * We temporarily evict the BO attached to this range. This necessitates
98  * evicting all user-mode queues of the process.
99  */
amdgpu_mn_invalidate_hsa(struct mmu_interval_notifier * mni,const struct mmu_notifier_range * range,unsigned long cur_seq)100 static bool amdgpu_mn_invalidate_hsa(struct mmu_interval_notifier *mni,
101 				     const struct mmu_notifier_range *range,
102 				     unsigned long cur_seq)
103 {
104 	struct amdgpu_bo *bo = container_of(mni, struct amdgpu_bo, notifier);
105 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
106 
107 	if (!mmu_notifier_range_blockable(range))
108 		return false;
109 
110 	mutex_lock(&adev->notifier_lock);
111 
112 	mmu_interval_set_seq(mni, cur_seq);
113 
114 	amdgpu_amdkfd_evict_userptr(bo->kfd_bo, bo->notifier.mm);
115 	mutex_unlock(&adev->notifier_lock);
116 
117 	return true;
118 }
119 
120 static const struct mmu_interval_notifier_ops amdgpu_mn_hsa_ops = {
121 	.invalidate = amdgpu_mn_invalidate_hsa,
122 };
123 
124 /**
125  * amdgpu_mn_register - register a BO for notifier updates
126  *
127  * @bo: amdgpu buffer object
128  * @addr: userptr addr we should monitor
129  *
130  * Registers a mmu_notifier for the given BO at the specified address.
131  * Returns 0 on success, -ERRNO if anything goes wrong.
132  */
amdgpu_mn_register(struct amdgpu_bo * bo,unsigned long addr)133 int amdgpu_mn_register(struct amdgpu_bo *bo, unsigned long addr)
134 {
135 	if (bo->kfd_bo)
136 		return mmu_interval_notifier_insert(&bo->notifier, current->mm,
137 						    addr, amdgpu_bo_size(bo),
138 						    &amdgpu_mn_hsa_ops);
139 	return mmu_interval_notifier_insert(&bo->notifier, current->mm, addr,
140 					    amdgpu_bo_size(bo),
141 					    &amdgpu_mn_gfx_ops);
142 }
143 
144 /**
145  * amdgpu_mn_unregister - unregister a BO for notifier updates
146  *
147  * @bo: amdgpu buffer object
148  *
149  * Remove any registration of mmu notifier updates from the buffer object.
150  */
amdgpu_mn_unregister(struct amdgpu_bo * bo)151 void amdgpu_mn_unregister(struct amdgpu_bo *bo)
152 {
153 	if (!bo->notifier.mm)
154 		return;
155 	mmu_interval_notifier_remove(&bo->notifier);
156 	bo->notifier.mm = NULL;
157 }
158 
amdgpu_hmm_range_get_pages(struct mmu_interval_notifier * notifier,struct mm_struct * mm,struct page ** pages,uint64_t start,uint64_t npages,struct hmm_range ** phmm_range,bool readonly,bool mmap_locked,void * owner)159 int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier,
160 			       struct mm_struct *mm, struct page **pages,
161 			       uint64_t start, uint64_t npages,
162 			       struct hmm_range **phmm_range, bool readonly,
163 			       bool mmap_locked, void *owner)
164 {
165 	struct hmm_range *hmm_range;
166 	unsigned long timeout;
167 	unsigned long i;
168 	unsigned long *pfns;
169 	int r = 0;
170 
171 	hmm_range = kzalloc(sizeof(*hmm_range), GFP_KERNEL);
172 	if (unlikely(!hmm_range))
173 		return -ENOMEM;
174 
175 	pfns = kvmalloc_array(npages, sizeof(*pfns), GFP_KERNEL);
176 	if (unlikely(!pfns)) {
177 		r = -ENOMEM;
178 		goto out_free_range;
179 	}
180 
181 	hmm_range->notifier = notifier;
182 	hmm_range->default_flags = HMM_PFN_REQ_FAULT;
183 	if (!readonly)
184 		hmm_range->default_flags |= HMM_PFN_REQ_WRITE;
185 	hmm_range->hmm_pfns = pfns;
186 	hmm_range->start = start;
187 	hmm_range->end = start + npages * PAGE_SIZE;
188 	hmm_range->dev_private_owner = owner;
189 
190 	/* Assuming 512MB takes maxmium 1 second to fault page address */
191 	timeout = max(npages >> 17, 1ULL) * HMM_RANGE_DEFAULT_TIMEOUT;
192 	timeout = jiffies + msecs_to_jiffies(timeout);
193 
194 retry:
195 	hmm_range->notifier_seq = mmu_interval_read_begin(notifier);
196 
197 	if (likely(!mmap_locked))
198 		mmap_read_lock(mm);
199 
200 	r = hmm_range_fault(hmm_range);
201 
202 	if (likely(!mmap_locked))
203 		mmap_read_unlock(mm);
204 	if (unlikely(r)) {
205 		/*
206 		 * FIXME: This timeout should encompass the retry from
207 		 * mmu_interval_read_retry() as well.
208 		 */
209 		if (r == -EBUSY && !time_after(jiffies, timeout))
210 			goto retry;
211 		goto out_free_pfns;
212 	}
213 
214 	/*
215 	 * Due to default_flags, all pages are HMM_PFN_VALID or
216 	 * hmm_range_fault() fails. FIXME: The pages cannot be touched outside
217 	 * the notifier_lock, and mmu_interval_read_retry() must be done first.
218 	 */
219 	for (i = 0; pages && i < npages; i++)
220 		pages[i] = hmm_pfn_to_page(pfns[i]);
221 
222 	*phmm_range = hmm_range;
223 
224 	return 0;
225 
226 out_free_pfns:
227 	kvfree(pfns);
228 out_free_range:
229 	kfree(hmm_range);
230 
231 	return r;
232 }
233 
amdgpu_hmm_range_get_pages_done(struct hmm_range * hmm_range)234 int amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range)
235 {
236 	int r;
237 
238 	r = mmu_interval_read_retry(hmm_range->notifier,
239 				    hmm_range->notifier_seq);
240 	kvfree(hmm_range->hmm_pfns);
241 	kfree(hmm_range);
242 
243 	return r;
244 }
245