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 #define MAX_WALK_BYTE	(2UL << 30)
55 
56 /**
57  * amdgpu_hmm_invalidate_gfx - callback to notify about mm change
58  *
59  * @mni: the range (mm) is about to update
60  * @range: details on the invalidation
61  * @cur_seq: Value to pass to mmu_interval_set_seq()
62  *
63  * Block for operations on BOs to finish and mark pages as accessed and
64  * potentially dirty.
65  */
66 static bool amdgpu_hmm_invalidate_gfx(struct mmu_interval_notifier *mni,
67 				      const struct mmu_notifier_range *range,
68 				      unsigned long cur_seq)
69 {
70 	struct amdgpu_bo *bo = container_of(mni, struct amdgpu_bo, notifier);
71 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
72 	long r;
73 
74 	if (!mmu_notifier_range_blockable(range))
75 		return false;
76 
77 	mutex_lock(&adev->notifier_lock);
78 
79 	mmu_interval_set_seq(mni, cur_seq);
80 
81 	r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP,
82 				  false, MAX_SCHEDULE_TIMEOUT);
83 	mutex_unlock(&adev->notifier_lock);
84 	if (r <= 0)
85 		DRM_ERROR("(%ld) failed to wait for user bo\n", r);
86 	return true;
87 }
88 
89 static const struct mmu_interval_notifier_ops amdgpu_hmm_gfx_ops = {
90 	.invalidate = amdgpu_hmm_invalidate_gfx,
91 };
92 
93 /**
94  * amdgpu_hmm_invalidate_hsa - callback to notify about mm change
95  *
96  * @mni: the range (mm) is about to update
97  * @range: details on the invalidation
98  * @cur_seq: Value to pass to mmu_interval_set_seq()
99  *
100  * We temporarily evict the BO attached to this range. This necessitates
101  * evicting all user-mode queues of the process.
102  */
103 static bool amdgpu_hmm_invalidate_hsa(struct mmu_interval_notifier *mni,
104 				      const struct mmu_notifier_range *range,
105 				      unsigned long cur_seq)
106 {
107 	struct amdgpu_bo *bo = container_of(mni, struct amdgpu_bo, notifier);
108 
109 	if (!mmu_notifier_range_blockable(range))
110 		return false;
111 
112 	amdgpu_amdkfd_evict_userptr(mni, cur_seq, bo->kfd_bo);
113 
114 	return true;
115 }
116 
117 static const struct mmu_interval_notifier_ops amdgpu_hmm_hsa_ops = {
118 	.invalidate = amdgpu_hmm_invalidate_hsa,
119 };
120 
121 /**
122  * amdgpu_hmm_register - register a BO for notifier updates
123  *
124  * @bo: amdgpu buffer object
125  * @addr: userptr addr we should monitor
126  *
127  * Registers a mmu_notifier for the given BO at the specified address.
128  * Returns 0 on success, -ERRNO if anything goes wrong.
129  */
130 int amdgpu_hmm_register(struct amdgpu_bo *bo, unsigned long addr)
131 {
132 	if (bo->kfd_bo)
133 		return mmu_interval_notifier_insert(&bo->notifier, current->mm,
134 						    addr, amdgpu_bo_size(bo),
135 						    &amdgpu_hmm_hsa_ops);
136 	return mmu_interval_notifier_insert(&bo->notifier, current->mm, addr,
137 					    amdgpu_bo_size(bo),
138 					    &amdgpu_hmm_gfx_ops);
139 }
140 
141 /**
142  * amdgpu_hmm_unregister - unregister a BO for notifier updates
143  *
144  * @bo: amdgpu buffer object
145  *
146  * Remove any registration of mmu notifier updates from the buffer object.
147  */
148 void amdgpu_hmm_unregister(struct amdgpu_bo *bo)
149 {
150 	if (!bo->notifier.mm)
151 		return;
152 	mmu_interval_notifier_remove(&bo->notifier);
153 	bo->notifier.mm = NULL;
154 }
155 
156 int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier,
157 			       uint64_t start, uint64_t npages, bool readonly,
158 			       void *owner, struct page **pages,
159 			       struct hmm_range **phmm_range)
160 {
161 	struct hmm_range *hmm_range;
162 	unsigned long end;
163 	unsigned long timeout;
164 	unsigned long i;
165 	unsigned long *pfns;
166 	int r = 0;
167 
168 	hmm_range = kzalloc(sizeof(*hmm_range), GFP_KERNEL);
169 	if (unlikely(!hmm_range))
170 		return -ENOMEM;
171 
172 	pfns = kvmalloc_array(npages, sizeof(*pfns), GFP_KERNEL);
173 	if (unlikely(!pfns)) {
174 		r = -ENOMEM;
175 		goto out_free_range;
176 	}
177 
178 	hmm_range->notifier = notifier;
179 	hmm_range->default_flags = HMM_PFN_REQ_FAULT;
180 	if (!readonly)
181 		hmm_range->default_flags |= HMM_PFN_REQ_WRITE;
182 	hmm_range->hmm_pfns = pfns;
183 	hmm_range->start = start;
184 	end = start + npages * PAGE_SIZE;
185 	hmm_range->dev_private_owner = owner;
186 
187 	do {
188 		hmm_range->end = min(hmm_range->start + MAX_WALK_BYTE, end);
189 
190 		pr_debug("hmm range: start = 0x%lx, end = 0x%lx",
191 			hmm_range->start, hmm_range->end);
192 
193 		/* Assuming 512MB takes maxmium 1 second to fault page address */
194 		timeout = max((hmm_range->end - hmm_range->start) >> 29, 1UL);
195 		timeout *= HMM_RANGE_DEFAULT_TIMEOUT;
196 		timeout = jiffies + msecs_to_jiffies(timeout);
197 
198 retry:
199 		hmm_range->notifier_seq = mmu_interval_read_begin(notifier);
200 		r = hmm_range_fault(hmm_range);
201 		if (unlikely(r)) {
202 			/*
203 			 * FIXME: This timeout should encompass the retry from
204 			 * mmu_interval_read_retry() as well.
205 			 */
206 			if (r == -EBUSY && !time_after(jiffies, timeout))
207 				goto retry;
208 			goto out_free_pfns;
209 		}
210 
211 		if (hmm_range->end == end)
212 			break;
213 		hmm_range->hmm_pfns += MAX_WALK_BYTE >> PAGE_SHIFT;
214 		hmm_range->start = hmm_range->end;
215 		schedule();
216 	} while (hmm_range->end < end);
217 
218 	hmm_range->start = start;
219 	hmm_range->hmm_pfns = pfns;
220 
221 	/*
222 	 * Due to default_flags, all pages are HMM_PFN_VALID or
223 	 * hmm_range_fault() fails. FIXME: The pages cannot be touched outside
224 	 * the notifier_lock, and mmu_interval_read_retry() must be done first.
225 	 */
226 	for (i = 0; pages && i < npages; i++)
227 		pages[i] = hmm_pfn_to_page(pfns[i]);
228 
229 	*phmm_range = hmm_range;
230 
231 	return 0;
232 
233 out_free_pfns:
234 	kvfree(pfns);
235 out_free_range:
236 	kfree(hmm_range);
237 
238 	return r;
239 }
240 
241 bool amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range)
242 {
243 	bool r;
244 
245 	r = mmu_interval_read_retry(hmm_range->notifier,
246 				    hmm_range->notifier_seq);
247 	kvfree(hmm_range->hmm_pfns);
248 	kfree(hmm_range);
249 
250 	return r;
251 }
252