1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33#include <linux/types.h>
34#include <linux/sched.h>
35#include <linux/sched/mm.h>
36#include <linux/sched/task.h>
37#include <linux/pid.h>
38#include <linux/slab.h>
39#include <linux/export.h>
40#include <linux/vmalloc.h>
41#include <linux/hugetlb.h>
42#include <linux/interval_tree.h>
43#include <linux/hmm.h>
44#include <linux/pagemap.h>
45
46#include <rdma/ib_verbs.h>
47#include <rdma/ib_umem.h>
48#include <rdma/ib_umem_odp.h>
49
50#include "uverbs.h"
51
52static inline int ib_init_umem_odp(struct ib_umem_odp *umem_odp,
53 const struct mmu_interval_notifier_ops *ops)
54{
55 int ret;
56
57 umem_odp->umem.is_odp = 1;
58 mutex_init(&umem_odp->umem_mutex);
59
60 if (!umem_odp->is_implicit_odp) {
61 size_t page_size = 1UL << umem_odp->page_shift;
62 unsigned long start;
63 unsigned long end;
64 size_t ndmas, npfns;
65
66 start = ALIGN_DOWN(umem_odp->umem.address, page_size);
67 if (check_add_overflow(umem_odp->umem.address,
68 (unsigned long)umem_odp->umem.length,
69 &end))
70 return -EOVERFLOW;
71 end = ALIGN(end, page_size);
72 if (unlikely(end < page_size))
73 return -EOVERFLOW;
74
75 ndmas = (end - start) >> umem_odp->page_shift;
76 if (!ndmas)
77 return -EINVAL;
78
79 npfns = (end - start) >> PAGE_SHIFT;
80 umem_odp->pfn_list = kvcalloc(
81 npfns, sizeof(*umem_odp->pfn_list), GFP_KERNEL);
82 if (!umem_odp->pfn_list)
83 return -ENOMEM;
84
85 umem_odp->dma_list = kvcalloc(
86 ndmas, sizeof(*umem_odp->dma_list), GFP_KERNEL);
87 if (!umem_odp->dma_list) {
88 ret = -ENOMEM;
89 goto out_pfn_list;
90 }
91
92 ret = mmu_interval_notifier_insert(&umem_odp->notifier,
93 umem_odp->umem.owning_mm,
94 start, end - start, ops);
95 if (ret)
96 goto out_dma_list;
97 }
98
99 return 0;
100
101out_dma_list:
102 kvfree(umem_odp->dma_list);
103out_pfn_list:
104 kvfree(umem_odp->pfn_list);
105 return ret;
106}
107
108
109
110
111
112
113
114
115
116
117
118struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
119 int access)
120{
121 struct ib_umem *umem;
122 struct ib_umem_odp *umem_odp;
123 int ret;
124
125 if (access & IB_ACCESS_HUGETLB)
126 return ERR_PTR(-EINVAL);
127
128 umem_odp = kzalloc(sizeof(*umem_odp), GFP_KERNEL);
129 if (!umem_odp)
130 return ERR_PTR(-ENOMEM);
131 umem = &umem_odp->umem;
132 umem->ibdev = device;
133 umem->writable = ib_access_writable(access);
134 umem->owning_mm = current->mm;
135 umem_odp->is_implicit_odp = 1;
136 umem_odp->page_shift = PAGE_SHIFT;
137
138 umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
139 ret = ib_init_umem_odp(umem_odp, NULL);
140 if (ret) {
141 put_pid(umem_odp->tgid);
142 kfree(umem_odp);
143 return ERR_PTR(ret);
144 }
145 return umem_odp;
146}
147EXPORT_SYMBOL(ib_umem_odp_alloc_implicit);
148
149
150
151
152
153
154
155
156
157
158
159struct ib_umem_odp *
160ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
161 size_t size,
162 const struct mmu_interval_notifier_ops *ops)
163{
164
165
166
167
168 struct ib_umem_odp *odp_data;
169 struct ib_umem *umem;
170 int ret;
171
172 if (WARN_ON(!root->is_implicit_odp))
173 return ERR_PTR(-EINVAL);
174
175 odp_data = kzalloc(sizeof(*odp_data), GFP_KERNEL);
176 if (!odp_data)
177 return ERR_PTR(-ENOMEM);
178 umem = &odp_data->umem;
179 umem->ibdev = root->umem.ibdev;
180 umem->length = size;
181 umem->address = addr;
182 umem->writable = root->umem.writable;
183 umem->owning_mm = root->umem.owning_mm;
184 odp_data->page_shift = PAGE_SHIFT;
185 odp_data->notifier.ops = ops;
186
187
188
189
190
191 if (!mmget_not_zero(umem->owning_mm)) {
192 ret = -EFAULT;
193 goto out_free;
194 }
195
196 odp_data->tgid = get_pid(root->tgid);
197 ret = ib_init_umem_odp(odp_data, ops);
198 if (ret)
199 goto out_tgid;
200 mmput(umem->owning_mm);
201 return odp_data;
202
203out_tgid:
204 put_pid(odp_data->tgid);
205 mmput(umem->owning_mm);
206out_free:
207 kfree(odp_data);
208 return ERR_PTR(ret);
209}
210EXPORT_SYMBOL(ib_umem_odp_alloc_child);
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
226 unsigned long addr, size_t size, int access,
227 const struct mmu_interval_notifier_ops *ops)
228{
229 struct ib_umem_odp *umem_odp;
230 struct mm_struct *mm;
231 int ret;
232
233 if (WARN_ON_ONCE(!(access & IB_ACCESS_ON_DEMAND)))
234 return ERR_PTR(-EINVAL);
235
236 umem_odp = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL);
237 if (!umem_odp)
238 return ERR_PTR(-ENOMEM);
239
240 umem_odp->umem.ibdev = device;
241 umem_odp->umem.length = size;
242 umem_odp->umem.address = addr;
243 umem_odp->umem.writable = ib_access_writable(access);
244 umem_odp->umem.owning_mm = mm = current->mm;
245 umem_odp->notifier.ops = ops;
246
247 umem_odp->page_shift = PAGE_SHIFT;
248#ifdef CONFIG_HUGETLB_PAGE
249 if (access & IB_ACCESS_HUGETLB)
250 umem_odp->page_shift = HPAGE_SHIFT;
251#endif
252
253 umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
254 ret = ib_init_umem_odp(umem_odp, ops);
255 if (ret)
256 goto err_put_pid;
257 return umem_odp;
258
259err_put_pid:
260 put_pid(umem_odp->tgid);
261 kfree(umem_odp);
262 return ERR_PTR(ret);
263}
264EXPORT_SYMBOL(ib_umem_odp_get);
265
266void ib_umem_odp_release(struct ib_umem_odp *umem_odp)
267{
268
269
270
271
272
273
274 if (!umem_odp->is_implicit_odp) {
275 mutex_lock(&umem_odp->umem_mutex);
276 ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
277 ib_umem_end(umem_odp));
278 mutex_unlock(&umem_odp->umem_mutex);
279 mmu_interval_notifier_remove(&umem_odp->notifier);
280 kvfree(umem_odp->dma_list);
281 kvfree(umem_odp->pfn_list);
282 }
283 put_pid(umem_odp->tgid);
284 kfree(umem_odp);
285}
286EXPORT_SYMBOL(ib_umem_odp_release);
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302static int ib_umem_odp_map_dma_single_page(
303 struct ib_umem_odp *umem_odp,
304 unsigned int dma_index,
305 struct page *page,
306 u64 access_mask)
307{
308 struct ib_device *dev = umem_odp->umem.ibdev;
309 dma_addr_t *dma_addr = &umem_odp->dma_list[dma_index];
310
311 if (*dma_addr) {
312
313
314
315
316
317 *dma_addr = (*dma_addr & ODP_DMA_ADDR_MASK) | access_mask;
318 return 0;
319 }
320
321 *dma_addr = ib_dma_map_page(dev, page, 0, 1 << umem_odp->page_shift,
322 DMA_BIDIRECTIONAL);
323 if (ib_dma_mapping_error(dev, *dma_addr)) {
324 *dma_addr = 0;
325 return -EFAULT;
326 }
327 umem_odp->npages++;
328 *dma_addr |= access_mask;
329 return 0;
330}
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352int ib_umem_odp_map_dma_and_lock(struct ib_umem_odp *umem_odp, u64 user_virt,
353 u64 bcnt, u64 access_mask, bool fault)
354 __acquires(&umem_odp->umem_mutex)
355{
356 struct task_struct *owning_process = NULL;
357 struct mm_struct *owning_mm = umem_odp->umem.owning_mm;
358 int pfn_index, dma_index, ret = 0, start_idx;
359 unsigned int page_shift, hmm_order, pfn_start_idx;
360 unsigned long num_pfns, current_seq;
361 struct hmm_range range = {};
362 unsigned long timeout;
363
364 if (access_mask == 0)
365 return -EINVAL;
366
367 if (user_virt < ib_umem_start(umem_odp) ||
368 user_virt + bcnt > ib_umem_end(umem_odp))
369 return -EFAULT;
370
371 page_shift = umem_odp->page_shift;
372
373
374
375
376
377
378 owning_process = get_pid_task(umem_odp->tgid, PIDTYPE_PID);
379 if (!owning_process || !mmget_not_zero(owning_mm)) {
380 ret = -EINVAL;
381 goto out_put_task;
382 }
383
384 range.notifier = &umem_odp->notifier;
385 range.start = ALIGN_DOWN(user_virt, 1UL << page_shift);
386 range.end = ALIGN(user_virt + bcnt, 1UL << page_shift);
387 pfn_start_idx = (range.start - ib_umem_start(umem_odp)) >> PAGE_SHIFT;
388 num_pfns = (range.end - range.start) >> PAGE_SHIFT;
389 if (fault) {
390 range.default_flags = HMM_PFN_REQ_FAULT;
391
392 if (access_mask & ODP_WRITE_ALLOWED_BIT)
393 range.default_flags |= HMM_PFN_REQ_WRITE;
394 }
395
396 range.hmm_pfns = &(umem_odp->pfn_list[pfn_start_idx]);
397 timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
398
399retry:
400 current_seq = range.notifier_seq =
401 mmu_interval_read_begin(&umem_odp->notifier);
402
403 mmap_read_lock(owning_mm);
404 ret = hmm_range_fault(&range);
405 mmap_read_unlock(owning_mm);
406 if (unlikely(ret)) {
407 if (ret == -EBUSY && !time_after(jiffies, timeout))
408 goto retry;
409 goto out_put_mm;
410 }
411
412 start_idx = (range.start - ib_umem_start(umem_odp)) >> page_shift;
413 dma_index = start_idx;
414
415 mutex_lock(&umem_odp->umem_mutex);
416 if (mmu_interval_read_retry(&umem_odp->notifier, current_seq)) {
417 mutex_unlock(&umem_odp->umem_mutex);
418 goto retry;
419 }
420
421 for (pfn_index = 0; pfn_index < num_pfns;
422 pfn_index += 1 << (page_shift - PAGE_SHIFT), dma_index++) {
423
424 if (fault) {
425
426
427
428
429 WARN_ON(range.hmm_pfns[pfn_index] & HMM_PFN_ERROR);
430 WARN_ON(!(range.hmm_pfns[pfn_index] & HMM_PFN_VALID));
431 } else {
432 if (!(range.hmm_pfns[pfn_index] & HMM_PFN_VALID)) {
433 WARN_ON(umem_odp->dma_list[dma_index]);
434 continue;
435 }
436 access_mask = ODP_READ_ALLOWED_BIT;
437 if (range.hmm_pfns[pfn_index] & HMM_PFN_WRITE)
438 access_mask |= ODP_WRITE_ALLOWED_BIT;
439 }
440
441 hmm_order = hmm_pfn_to_map_order(range.hmm_pfns[pfn_index]);
442
443
444
445 if (hmm_order + PAGE_SHIFT < page_shift) {
446 ret = -EINVAL;
447 ibdev_dbg(umem_odp->umem.ibdev,
448 "%s: un-expected hmm_order %u, page_shift %u\n",
449 __func__, hmm_order, page_shift);
450 break;
451 }
452
453 ret = ib_umem_odp_map_dma_single_page(
454 umem_odp, dma_index, hmm_pfn_to_page(range.hmm_pfns[pfn_index]),
455 access_mask);
456 if (ret < 0) {
457 ibdev_dbg(umem_odp->umem.ibdev,
458 "ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
459 break;
460 }
461 }
462
463 if (!ret)
464 ret = dma_index - start_idx;
465 else
466 mutex_unlock(&umem_odp->umem_mutex);
467
468out_put_mm:
469 mmput(owning_mm);
470out_put_task:
471 if (owning_process)
472 put_task_struct(owning_process);
473 return ret;
474}
475EXPORT_SYMBOL(ib_umem_odp_map_dma_and_lock);
476
477void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt,
478 u64 bound)
479{
480 dma_addr_t dma_addr;
481 dma_addr_t dma;
482 int idx;
483 u64 addr;
484 struct ib_device *dev = umem_odp->umem.ibdev;
485
486 lockdep_assert_held(&umem_odp->umem_mutex);
487
488 virt = max_t(u64, virt, ib_umem_start(umem_odp));
489 bound = min_t(u64, bound, ib_umem_end(umem_odp));
490 for (addr = virt; addr < bound; addr += BIT(umem_odp->page_shift)) {
491 idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
492 dma = umem_odp->dma_list[idx];
493
494
495 if (dma) {
496 unsigned long pfn_idx = (addr - ib_umem_start(umem_odp)) >> PAGE_SHIFT;
497 struct page *page = hmm_pfn_to_page(umem_odp->pfn_list[pfn_idx]);
498
499 dma_addr = dma & ODP_DMA_ADDR_MASK;
500 ib_dma_unmap_page(dev, dma_addr,
501 BIT(umem_odp->page_shift),
502 DMA_BIDIRECTIONAL);
503 if (dma & ODP_WRITE_ALLOWED_BIT) {
504 struct page *head_page = compound_head(page);
505
506
507
508
509
510
511
512
513
514 set_page_dirty(head_page);
515 }
516 umem_odp->dma_list[idx] = 0;
517 umem_odp->npages--;
518 }
519 }
520}
521EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);
522