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 */
66static 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(bdev: 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(interval_sub: mni, cur_seq);
80
81 r = dma_resv_wait_timeout(obj: bo->tbo.base.resv, usage: DMA_RESV_USAGE_BOOKKEEP,
82 intr: false, MAX_SCHEDULE_TIMEOUT);
83 mutex_unlock(lock: &adev->notifier_lock);
84 if (r <= 0)
85 DRM_ERROR("(%ld) failed to wait for user bo\n", r);
86 return true;
87}
88
89static 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 */
103static 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, mem: bo->kfd_bo);
113
114 return true;
115}
116
117static 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 */
130int amdgpu_hmm_register(struct amdgpu_bo *bo, unsigned long addr)
131{
132 int r;
133
134 if (bo->kfd_bo)
135 r = mmu_interval_notifier_insert(interval_sub: &bo->notifier, current->mm,
136 start: addr, length: amdgpu_bo_size(bo),
137 ops: &amdgpu_hmm_hsa_ops);
138 else
139 r = mmu_interval_notifier_insert(interval_sub: &bo->notifier, current->mm, start: addr,
140 length: amdgpu_bo_size(bo),
141 ops: &amdgpu_hmm_gfx_ops);
142 if (r)
143 /*
144 * Make sure amdgpu_hmm_unregister() doesn't call
145 * mmu_interval_notifier_remove() when the notifier isn't properly
146 * initialized.
147 */
148 bo->notifier.mm = NULL;
149
150 return r;
151}
152
153/**
154 * amdgpu_hmm_unregister - unregister a BO for notifier updates
155 *
156 * @bo: amdgpu buffer object
157 *
158 * Remove any registration of mmu notifier updates from the buffer object.
159 */
160void amdgpu_hmm_unregister(struct amdgpu_bo *bo)
161{
162 if (!bo->notifier.mm)
163 return;
164 mmu_interval_notifier_remove(interval_sub: &bo->notifier);
165 bo->notifier.mm = NULL;
166}
167
168int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier,
169 uint64_t start, uint64_t npages, bool readonly,
170 void *owner, struct page **pages,
171 struct hmm_range **phmm_range)
172{
173 struct hmm_range *hmm_range;
174 unsigned long end;
175 unsigned long timeout;
176 unsigned long i;
177 unsigned long *pfns;
178 int r = 0;
179
180 hmm_range = kzalloc(size: sizeof(*hmm_range), GFP_KERNEL);
181 if (unlikely(!hmm_range))
182 return -ENOMEM;
183
184 pfns = kvmalloc_array(n: npages, size: sizeof(*pfns), GFP_KERNEL);
185 if (unlikely(!pfns)) {
186 r = -ENOMEM;
187 goto out_free_range;
188 }
189
190 hmm_range->notifier = notifier;
191 hmm_range->default_flags = HMM_PFN_REQ_FAULT;
192 if (!readonly)
193 hmm_range->default_flags |= HMM_PFN_REQ_WRITE;
194 hmm_range->hmm_pfns = pfns;
195 hmm_range->start = start;
196 end = start + npages * PAGE_SIZE;
197 hmm_range->dev_private_owner = owner;
198
199 do {
200 hmm_range->end = min(hmm_range->start + MAX_WALK_BYTE, end);
201
202 pr_debug("hmm range: start = 0x%lx, end = 0x%lx",
203 hmm_range->start, hmm_range->end);
204
205 /* Assuming 64MB takes maximum 1 second to fault page address */
206 timeout = max((hmm_range->end - hmm_range->start) >> 26, 1UL);
207 timeout *= HMM_RANGE_DEFAULT_TIMEOUT;
208 timeout = jiffies + msecs_to_jiffies(m: timeout);
209
210retry:
211 hmm_range->notifier_seq = mmu_interval_read_begin(interval_sub: notifier);
212 r = hmm_range_fault(range: hmm_range);
213 if (unlikely(r)) {
214 schedule();
215 /*
216 * FIXME: This timeout should encompass the retry from
217 * mmu_interval_read_retry() as well.
218 */
219 if (r == -EBUSY && !time_after(jiffies, timeout))
220 goto retry;
221 goto out_free_pfns;
222 }
223
224 if (hmm_range->end == end)
225 break;
226 hmm_range->hmm_pfns += MAX_WALK_BYTE >> PAGE_SHIFT;
227 hmm_range->start = hmm_range->end;
228 } while (hmm_range->end < end);
229
230 hmm_range->start = start;
231 hmm_range->hmm_pfns = pfns;
232
233 /*
234 * Due to default_flags, all pages are HMM_PFN_VALID or
235 * hmm_range_fault() fails. FIXME: The pages cannot be touched outside
236 * the notifier_lock, and mmu_interval_read_retry() must be done first.
237 */
238 for (i = 0; pages && i < npages; i++)
239 pages[i] = hmm_pfn_to_page(hmm_pfn: pfns[i]);
240
241 *phmm_range = hmm_range;
242
243 return 0;
244
245out_free_pfns:
246 kvfree(addr: pfns);
247out_free_range:
248 kfree(objp: hmm_range);
249
250 return r;
251}
252
253bool amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range)
254{
255 bool r;
256
257 r = mmu_interval_read_retry(interval_sub: hmm_range->notifier,
258 seq: hmm_range->notifier_seq);
259 kvfree(addr: hmm_range->hmm_pfns);
260 kfree(objp: hmm_range);
261
262 return r;
263}
264

source code of linux/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c