1 /*
2  * Copyright (c) 2014 Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
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/pagemap.h>
44 
45 #include <rdma/ib_verbs.h>
46 #include <rdma/ib_umem.h>
47 #include <rdma/ib_umem_odp.h>
48 
49 #include "uverbs.h"
50 
51 static inline int ib_init_umem_odp(struct ib_umem_odp *umem_odp,
52 				   const struct mmu_interval_notifier_ops *ops)
53 {
54 	int ret;
55 
56 	umem_odp->umem.is_odp = 1;
57 	mutex_init(&umem_odp->umem_mutex);
58 
59 	if (!umem_odp->is_implicit_odp) {
60 		size_t page_size = 1UL << umem_odp->page_shift;
61 		unsigned long start;
62 		unsigned long end;
63 		size_t pages;
64 
65 		start = ALIGN_DOWN(umem_odp->umem.address, page_size);
66 		if (check_add_overflow(umem_odp->umem.address,
67 				       (unsigned long)umem_odp->umem.length,
68 				       &end))
69 			return -EOVERFLOW;
70 		end = ALIGN(end, page_size);
71 		if (unlikely(end < page_size))
72 			return -EOVERFLOW;
73 
74 		pages = (end - start) >> umem_odp->page_shift;
75 		if (!pages)
76 			return -EINVAL;
77 
78 		umem_odp->page_list = kvcalloc(
79 			pages, sizeof(*umem_odp->page_list), GFP_KERNEL);
80 		if (!umem_odp->page_list)
81 			return -ENOMEM;
82 
83 		umem_odp->dma_list = kvcalloc(
84 			pages, sizeof(*umem_odp->dma_list), GFP_KERNEL);
85 		if (!umem_odp->dma_list) {
86 			ret = -ENOMEM;
87 			goto out_page_list;
88 		}
89 
90 		ret = mmu_interval_notifier_insert(&umem_odp->notifier,
91 						   umem_odp->umem.owning_mm,
92 						   start, end - start, ops);
93 		if (ret)
94 			goto out_dma_list;
95 	}
96 
97 	return 0;
98 
99 out_dma_list:
100 	kvfree(umem_odp->dma_list);
101 out_page_list:
102 	kvfree(umem_odp->page_list);
103 	return ret;
104 }
105 
106 /**
107  * ib_umem_odp_alloc_implicit - Allocate a parent implicit ODP umem
108  *
109  * Implicit ODP umems do not have a VA range and do not have any page lists.
110  * They exist only to hold the per_mm reference to help the driver create
111  * children umems.
112  *
113  * @device: IB device to create UMEM
114  * @access: ib_reg_mr access flags
115  */
116 struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
117 					       int access)
118 {
119 	struct ib_umem *umem;
120 	struct ib_umem_odp *umem_odp;
121 	int ret;
122 
123 	if (access & IB_ACCESS_HUGETLB)
124 		return ERR_PTR(-EINVAL);
125 
126 	umem_odp = kzalloc(sizeof(*umem_odp), GFP_KERNEL);
127 	if (!umem_odp)
128 		return ERR_PTR(-ENOMEM);
129 	umem = &umem_odp->umem;
130 	umem->ibdev = device;
131 	umem->writable = ib_access_writable(access);
132 	umem->owning_mm = current->mm;
133 	umem_odp->is_implicit_odp = 1;
134 	umem_odp->page_shift = PAGE_SHIFT;
135 
136 	umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
137 	ret = ib_init_umem_odp(umem_odp, NULL);
138 	if (ret) {
139 		put_pid(umem_odp->tgid);
140 		kfree(umem_odp);
141 		return ERR_PTR(ret);
142 	}
143 	return umem_odp;
144 }
145 EXPORT_SYMBOL(ib_umem_odp_alloc_implicit);
146 
147 /**
148  * ib_umem_odp_alloc_child - Allocate a child ODP umem under an implicit
149  *                           parent ODP umem
150  *
151  * @root: The parent umem enclosing the child. This must be allocated using
152  *        ib_alloc_implicit_odp_umem()
153  * @addr: The starting userspace VA
154  * @size: The length of the userspace VA
155  */
156 struct ib_umem_odp *
157 ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr,
158 			size_t size,
159 			const struct mmu_interval_notifier_ops *ops)
160 {
161 	/*
162 	 * Caller must ensure that root cannot be freed during the call to
163 	 * ib_alloc_odp_umem.
164 	 */
165 	struct ib_umem_odp *odp_data;
166 	struct ib_umem *umem;
167 	int ret;
168 
169 	if (WARN_ON(!root->is_implicit_odp))
170 		return ERR_PTR(-EINVAL);
171 
172 	odp_data = kzalloc(sizeof(*odp_data), GFP_KERNEL);
173 	if (!odp_data)
174 		return ERR_PTR(-ENOMEM);
175 	umem = &odp_data->umem;
176 	umem->ibdev = root->umem.ibdev;
177 	umem->length     = size;
178 	umem->address    = addr;
179 	umem->writable   = root->umem.writable;
180 	umem->owning_mm  = root->umem.owning_mm;
181 	odp_data->page_shift = PAGE_SHIFT;
182 	odp_data->notifier.ops = ops;
183 
184 	odp_data->tgid = get_pid(root->tgid);
185 	ret = ib_init_umem_odp(odp_data, ops);
186 	if (ret) {
187 		put_pid(odp_data->tgid);
188 		kfree(odp_data);
189 		return ERR_PTR(ret);
190 	}
191 	return odp_data;
192 }
193 EXPORT_SYMBOL(ib_umem_odp_alloc_child);
194 
195 /**
196  * ib_umem_odp_get - Create a umem_odp for a userspace va
197  *
198  * @device: IB device struct to get UMEM
199  * @addr: userspace virtual address to start at
200  * @size: length of region to pin
201  * @access: IB_ACCESS_xxx flags for memory being pinned
202  *
203  * The driver should use when the access flags indicate ODP memory. It avoids
204  * pinning, instead, stores the mm for future page fault handling in
205  * conjunction with MMU notifiers.
206  */
207 struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
208 				    unsigned long addr, size_t size, int access,
209 				    const struct mmu_interval_notifier_ops *ops)
210 {
211 	struct ib_umem_odp *umem_odp;
212 	struct mm_struct *mm;
213 	int ret;
214 
215 	if (WARN_ON_ONCE(!(access & IB_ACCESS_ON_DEMAND)))
216 		return ERR_PTR(-EINVAL);
217 
218 	umem_odp = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL);
219 	if (!umem_odp)
220 		return ERR_PTR(-ENOMEM);
221 
222 	umem_odp->umem.ibdev = device;
223 	umem_odp->umem.length = size;
224 	umem_odp->umem.address = addr;
225 	umem_odp->umem.writable = ib_access_writable(access);
226 	umem_odp->umem.owning_mm = mm = current->mm;
227 	umem_odp->notifier.ops = ops;
228 
229 	umem_odp->page_shift = PAGE_SHIFT;
230 #ifdef CONFIG_HUGETLB_PAGE
231 	if (access & IB_ACCESS_HUGETLB)
232 		umem_odp->page_shift = HPAGE_SHIFT;
233 #endif
234 
235 	umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
236 	ret = ib_init_umem_odp(umem_odp, ops);
237 	if (ret)
238 		goto err_put_pid;
239 	return umem_odp;
240 
241 err_put_pid:
242 	put_pid(umem_odp->tgid);
243 	kfree(umem_odp);
244 	return ERR_PTR(ret);
245 }
246 EXPORT_SYMBOL(ib_umem_odp_get);
247 
248 void ib_umem_odp_release(struct ib_umem_odp *umem_odp)
249 {
250 	/*
251 	 * Ensure that no more pages are mapped in the umem.
252 	 *
253 	 * It is the driver's responsibility to ensure, before calling us,
254 	 * that the hardware will not attempt to access the MR any more.
255 	 */
256 	if (!umem_odp->is_implicit_odp) {
257 		mutex_lock(&umem_odp->umem_mutex);
258 		ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
259 					    ib_umem_end(umem_odp));
260 		mutex_unlock(&umem_odp->umem_mutex);
261 		mmu_interval_notifier_remove(&umem_odp->notifier);
262 		kvfree(umem_odp->dma_list);
263 		kvfree(umem_odp->page_list);
264 		put_pid(umem_odp->tgid);
265 	}
266 	kfree(umem_odp);
267 }
268 EXPORT_SYMBOL(ib_umem_odp_release);
269 
270 /*
271  * Map for DMA and insert a single page into the on-demand paging page tables.
272  *
273  * @umem: the umem to insert the page to.
274  * @page_index: index in the umem to add the page to.
275  * @page: the page struct to map and add.
276  * @access_mask: access permissions needed for this page.
277  * @current_seq: sequence number for synchronization with invalidations.
278  *               the sequence number is taken from
279  *               umem_odp->notifiers_seq.
280  *
281  * The function returns -EFAULT if the DMA mapping operation fails. It returns
282  * -EAGAIN if a concurrent invalidation prevents us from updating the page.
283  *
284  * The page is released via put_page even if the operation failed. For on-demand
285  * pinning, the page is released whenever it isn't stored in the umem.
286  */
287 static int ib_umem_odp_map_dma_single_page(
288 		struct ib_umem_odp *umem_odp,
289 		unsigned int page_index,
290 		struct page *page,
291 		u64 access_mask,
292 		unsigned long current_seq)
293 {
294 	struct ib_device *dev = umem_odp->umem.ibdev;
295 	dma_addr_t dma_addr;
296 	int ret = 0;
297 
298 	if (mmu_interval_check_retry(&umem_odp->notifier, current_seq)) {
299 		ret = -EAGAIN;
300 		goto out;
301 	}
302 	if (!(umem_odp->dma_list[page_index])) {
303 		dma_addr =
304 			ib_dma_map_page(dev, page, 0, BIT(umem_odp->page_shift),
305 					DMA_BIDIRECTIONAL);
306 		if (ib_dma_mapping_error(dev, dma_addr)) {
307 			ret = -EFAULT;
308 			goto out;
309 		}
310 		umem_odp->dma_list[page_index] = dma_addr | access_mask;
311 		umem_odp->page_list[page_index] = page;
312 		umem_odp->npages++;
313 	} else if (umem_odp->page_list[page_index] == page) {
314 		umem_odp->dma_list[page_index] |= access_mask;
315 	} else {
316 		/*
317 		 * This is a race here where we could have done:
318 		 *
319 		 *         CPU0                             CPU1
320 		 *   get_user_pages()
321 		 *                                       invalidate()
322 		 *                                       page_fault()
323 		 *   mutex_lock(umem_mutex)
324 		 *    page from GUP != page in ODP
325 		 *
326 		 * It should be prevented by the retry test above as reading
327 		 * the seq number should be reliable under the
328 		 * umem_mutex. Thus something is really not working right if
329 		 * things get here.
330 		 */
331 		WARN(true,
332 		     "Got different pages in IB device and from get_user_pages. IB device page: %p, gup page: %p\n",
333 		     umem_odp->page_list[page_index], page);
334 		ret = -EAGAIN;
335 	}
336 
337 out:
338 	put_page(page);
339 	return ret;
340 }
341 
342 /**
343  * ib_umem_odp_map_dma_pages - Pin and DMA map userspace memory in an ODP MR.
344  *
345  * Pins the range of pages passed in the argument, and maps them to
346  * DMA addresses. The DMA addresses of the mapped pages is updated in
347  * umem_odp->dma_list.
348  *
349  * Returns the number of pages mapped in success, negative error code
350  * for failure.
351  * An -EAGAIN error code is returned when a concurrent mmu notifier prevents
352  * the function from completing its task.
353  * An -ENOENT error code indicates that userspace process is being terminated
354  * and mm was already destroyed.
355  * @umem_odp: the umem to map and pin
356  * @user_virt: the address from which we need to map.
357  * @bcnt: the minimal number of bytes to pin and map. The mapping might be
358  *        bigger due to alignment, and may also be smaller in case of an error
359  *        pinning or mapping a page. The actual pages mapped is returned in
360  *        the return value.
361  * @access_mask: bit mask of the requested access permissions for the given
362  *               range.
363  * @current_seq: the MMU notifiers sequance value for synchronization with
364  *               invalidations. the sequance number is read from
365  *               umem_odp->notifiers_seq before calling this function
366  */
367 int ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, u64 user_virt,
368 			      u64 bcnt, u64 access_mask,
369 			      unsigned long current_seq)
370 {
371 	struct task_struct *owning_process  = NULL;
372 	struct mm_struct *owning_mm = umem_odp->umem.owning_mm;
373 	struct page       **local_page_list = NULL;
374 	u64 page_mask, off;
375 	int j, k, ret = 0, start_idx, npages = 0;
376 	unsigned int flags = 0, page_shift;
377 	phys_addr_t p = 0;
378 
379 	if (access_mask == 0)
380 		return -EINVAL;
381 
382 	if (user_virt < ib_umem_start(umem_odp) ||
383 	    user_virt + bcnt > ib_umem_end(umem_odp))
384 		return -EFAULT;
385 
386 	local_page_list = (struct page **)__get_free_page(GFP_KERNEL);
387 	if (!local_page_list)
388 		return -ENOMEM;
389 
390 	page_shift = umem_odp->page_shift;
391 	page_mask = ~(BIT(page_shift) - 1);
392 	off = user_virt & (~page_mask);
393 	user_virt = user_virt & page_mask;
394 	bcnt += off; /* Charge for the first page offset as well. */
395 
396 	/*
397 	 * owning_process is allowed to be NULL, this means somehow the mm is
398 	 * existing beyond the lifetime of the originating process.. Presumably
399 	 * mmget_not_zero will fail in this case.
400 	 */
401 	owning_process = get_pid_task(umem_odp->tgid, PIDTYPE_PID);
402 	if (!owning_process || !mmget_not_zero(owning_mm)) {
403 		ret = -EINVAL;
404 		goto out_put_task;
405 	}
406 
407 	if (access_mask & ODP_WRITE_ALLOWED_BIT)
408 		flags |= FOLL_WRITE;
409 
410 	start_idx = (user_virt - ib_umem_start(umem_odp)) >> page_shift;
411 	k = start_idx;
412 
413 	while (bcnt > 0) {
414 		const size_t gup_num_pages = min_t(size_t,
415 				ALIGN(bcnt, PAGE_SIZE) / PAGE_SIZE,
416 				PAGE_SIZE / sizeof(struct page *));
417 
418 		down_read(&owning_mm->mmap_sem);
419 		/*
420 		 * Note: this might result in redundent page getting. We can
421 		 * avoid this by checking dma_list to be 0 before calling
422 		 * get_user_pages. However, this make the code much more
423 		 * complex (and doesn't gain us much performance in most use
424 		 * cases).
425 		 */
426 		npages = get_user_pages_remote(owning_process, owning_mm,
427 				user_virt, gup_num_pages,
428 				flags, local_page_list, NULL, NULL);
429 		up_read(&owning_mm->mmap_sem);
430 
431 		if (npages < 0) {
432 			if (npages != -EAGAIN)
433 				pr_warn("fail to get %zu user pages with error %d\n", gup_num_pages, npages);
434 			else
435 				pr_debug("fail to get %zu user pages with error %d\n", gup_num_pages, npages);
436 			break;
437 		}
438 
439 		bcnt -= min_t(size_t, npages << PAGE_SHIFT, bcnt);
440 		mutex_lock(&umem_odp->umem_mutex);
441 		for (j = 0; j < npages; j++, user_virt += PAGE_SIZE) {
442 			if (user_virt & ~page_mask) {
443 				p += PAGE_SIZE;
444 				if (page_to_phys(local_page_list[j]) != p) {
445 					ret = -EFAULT;
446 					break;
447 				}
448 				put_page(local_page_list[j]);
449 				continue;
450 			}
451 
452 			ret = ib_umem_odp_map_dma_single_page(
453 					umem_odp, k, local_page_list[j],
454 					access_mask, current_seq);
455 			if (ret < 0) {
456 				if (ret != -EAGAIN)
457 					pr_warn("ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
458 				else
459 					pr_debug("ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
460 				break;
461 			}
462 
463 			p = page_to_phys(local_page_list[j]);
464 			k++;
465 		}
466 		mutex_unlock(&umem_odp->umem_mutex);
467 
468 		if (ret < 0) {
469 			/*
470 			 * Release pages, remembering that the first page
471 			 * to hit an error was already released by
472 			 * ib_umem_odp_map_dma_single_page().
473 			 */
474 			if (npages - (j + 1) > 0)
475 				release_pages(&local_page_list[j+1],
476 					      npages - (j + 1));
477 			break;
478 		}
479 	}
480 
481 	if (ret >= 0) {
482 		if (npages < 0 && k == start_idx)
483 			ret = npages;
484 		else
485 			ret = k - start_idx;
486 	}
487 
488 	mmput(owning_mm);
489 out_put_task:
490 	if (owning_process)
491 		put_task_struct(owning_process);
492 	free_page((unsigned long)local_page_list);
493 	return ret;
494 }
495 EXPORT_SYMBOL(ib_umem_odp_map_dma_pages);
496 
497 void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt,
498 				 u64 bound)
499 {
500 	int idx;
501 	u64 addr;
502 	struct ib_device *dev = umem_odp->umem.ibdev;
503 
504 	lockdep_assert_held(&umem_odp->umem_mutex);
505 
506 	virt = max_t(u64, virt, ib_umem_start(umem_odp));
507 	bound = min_t(u64, bound, ib_umem_end(umem_odp));
508 	/* Note that during the run of this function, the
509 	 * notifiers_count of the MR is > 0, preventing any racing
510 	 * faults from completion. We might be racing with other
511 	 * invalidations, so we must make sure we free each page only
512 	 * once. */
513 	for (addr = virt; addr < bound; addr += BIT(umem_odp->page_shift)) {
514 		idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
515 		if (umem_odp->page_list[idx]) {
516 			struct page *page = umem_odp->page_list[idx];
517 			dma_addr_t dma = umem_odp->dma_list[idx];
518 			dma_addr_t dma_addr = dma & ODP_DMA_ADDR_MASK;
519 
520 			WARN_ON(!dma_addr);
521 
522 			ib_dma_unmap_page(dev, dma_addr,
523 					  BIT(umem_odp->page_shift),
524 					  DMA_BIDIRECTIONAL);
525 			if (dma & ODP_WRITE_ALLOWED_BIT) {
526 				struct page *head_page = compound_head(page);
527 				/*
528 				 * set_page_dirty prefers being called with
529 				 * the page lock. However, MMU notifiers are
530 				 * called sometimes with and sometimes without
531 				 * the lock. We rely on the umem_mutex instead
532 				 * to prevent other mmu notifiers from
533 				 * continuing and allowing the page mapping to
534 				 * be removed.
535 				 */
536 				set_page_dirty(head_page);
537 			}
538 			umem_odp->page_list[idx] = NULL;
539 			umem_odp->dma_list[idx] = 0;
540 			umem_odp->npages--;
541 		}
542 	}
543 }
544 EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);
545