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/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 
52 static 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 
101 out_dma_list:
102 	kvfree(umem_odp->dma_list);
103 out_pfn_list:
104 	kvfree(umem_odp->pfn_list);
105 	return ret;
106 }
107 
108 /**
109  * ib_umem_odp_alloc_implicit - Allocate a parent implicit ODP umem
110  *
111  * Implicit ODP umems do not have a VA range and do not have any page lists.
112  * They exist only to hold the per_mm reference to help the driver create
113  * children umems.
114  *
115  * @device: IB device to create UMEM
116  * @access: ib_reg_mr access flags
117  */
118 struct 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 }
147 EXPORT_SYMBOL(ib_umem_odp_alloc_implicit);
148 
149 /**
150  * ib_umem_odp_alloc_child - Allocate a child ODP umem under an implicit
151  *                           parent ODP umem
152  *
153  * @root: The parent umem enclosing the child. This must be allocated using
154  *        ib_alloc_implicit_odp_umem()
155  * @addr: The starting userspace VA
156  * @size: The length of the userspace VA
157  * @ops: MMU interval ops, currently only @invalidate
158  */
159 struct ib_umem_odp *
160 ib_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 	 * Caller must ensure that root cannot be freed during the call to
166 	 * ib_alloc_odp_umem.
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 	 * A mmget must be held when registering a notifier, the owming_mm only
189 	 * has a mm_grab at this point.
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 
203 out_tgid:
204 	put_pid(odp_data->tgid);
205 	mmput(umem->owning_mm);
206 out_free:
207 	kfree(odp_data);
208 	return ERR_PTR(ret);
209 }
210 EXPORT_SYMBOL(ib_umem_odp_alloc_child);
211 
212 /**
213  * ib_umem_odp_get - Create a umem_odp for a userspace va
214  *
215  * @device: IB device struct to get UMEM
216  * @addr: userspace virtual address to start at
217  * @size: length of region to pin
218  * @access: IB_ACCESS_xxx flags for memory being pinned
219  * @ops: MMU interval ops, currently only @invalidate
220  *
221  * The driver should use when the access flags indicate ODP memory. It avoids
222  * pinning, instead, stores the mm for future page fault handling in
223  * conjunction with MMU notifiers.
224  */
225 struct 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 	int ret;
231 
232 	if (WARN_ON_ONCE(!(access & IB_ACCESS_ON_DEMAND)))
233 		return ERR_PTR(-EINVAL);
234 
235 	umem_odp = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL);
236 	if (!umem_odp)
237 		return ERR_PTR(-ENOMEM);
238 
239 	umem_odp->umem.ibdev = device;
240 	umem_odp->umem.length = size;
241 	umem_odp->umem.address = addr;
242 	umem_odp->umem.writable = ib_access_writable(access);
243 	umem_odp->umem.owning_mm = current->mm;
244 	umem_odp->notifier.ops = ops;
245 
246 	umem_odp->page_shift = PAGE_SHIFT;
247 #ifdef CONFIG_HUGETLB_PAGE
248 	if (access & IB_ACCESS_HUGETLB)
249 		umem_odp->page_shift = HPAGE_SHIFT;
250 #endif
251 
252 	umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
253 	ret = ib_init_umem_odp(umem_odp, ops);
254 	if (ret)
255 		goto err_put_pid;
256 	return umem_odp;
257 
258 err_put_pid:
259 	put_pid(umem_odp->tgid);
260 	kfree(umem_odp);
261 	return ERR_PTR(ret);
262 }
263 EXPORT_SYMBOL(ib_umem_odp_get);
264 
265 void ib_umem_odp_release(struct ib_umem_odp *umem_odp)
266 {
267 	/*
268 	 * Ensure that no more pages are mapped in the umem.
269 	 *
270 	 * It is the driver's responsibility to ensure, before calling us,
271 	 * that the hardware will not attempt to access the MR any more.
272 	 */
273 	if (!umem_odp->is_implicit_odp) {
274 		mutex_lock(&umem_odp->umem_mutex);
275 		ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp),
276 					    ib_umem_end(umem_odp));
277 		mutex_unlock(&umem_odp->umem_mutex);
278 		mmu_interval_notifier_remove(&umem_odp->notifier);
279 		kvfree(umem_odp->dma_list);
280 		kvfree(umem_odp->pfn_list);
281 	}
282 	put_pid(umem_odp->tgid);
283 	kfree(umem_odp);
284 }
285 EXPORT_SYMBOL(ib_umem_odp_release);
286 
287 /*
288  * Map for DMA and insert a single page into the on-demand paging page tables.
289  *
290  * @umem: the umem to insert the page to.
291  * @dma_index: index in the umem to add the dma to.
292  * @page: the page struct to map and add.
293  * @access_mask: access permissions needed for this page.
294  *
295  * The function returns -EFAULT if the DMA mapping operation fails.
296  *
297  */
298 static int ib_umem_odp_map_dma_single_page(
299 		struct ib_umem_odp *umem_odp,
300 		unsigned int dma_index,
301 		struct page *page,
302 		u64 access_mask)
303 {
304 	struct ib_device *dev = umem_odp->umem.ibdev;
305 	dma_addr_t *dma_addr = &umem_odp->dma_list[dma_index];
306 
307 	if (*dma_addr) {
308 		/*
309 		 * If the page is already dma mapped it means it went through
310 		 * a non-invalidating trasition, like read-only to writable.
311 		 * Resync the flags.
312 		 */
313 		*dma_addr = (*dma_addr & ODP_DMA_ADDR_MASK) | access_mask;
314 		return 0;
315 	}
316 
317 	*dma_addr = ib_dma_map_page(dev, page, 0, 1 << umem_odp->page_shift,
318 				    DMA_BIDIRECTIONAL);
319 	if (ib_dma_mapping_error(dev, *dma_addr)) {
320 		*dma_addr = 0;
321 		return -EFAULT;
322 	}
323 	umem_odp->npages++;
324 	*dma_addr |= access_mask;
325 	return 0;
326 }
327 
328 /**
329  * ib_umem_odp_map_dma_and_lock - DMA map userspace memory in an ODP MR and lock it.
330  *
331  * Maps the range passed in the argument to DMA addresses.
332  * The DMA addresses of the mapped pages is updated in umem_odp->dma_list.
333  * Upon success the ODP MR will be locked to let caller complete its device
334  * page table update.
335  *
336  * Returns the number of pages mapped in success, negative error code
337  * for failure.
338  * @umem_odp: the umem to map and pin
339  * @user_virt: the address from which we need to map.
340  * @bcnt: the minimal number of bytes to pin and map. The mapping might be
341  *        bigger due to alignment, and may also be smaller in case of an error
342  *        pinning or mapping a page. The actual pages mapped is returned in
343  *        the return value.
344  * @access_mask: bit mask of the requested access permissions for the given
345  *               range.
346  * @fault: is faulting required for the given range
347  */
348 int ib_umem_odp_map_dma_and_lock(struct ib_umem_odp *umem_odp, u64 user_virt,
349 				 u64 bcnt, u64 access_mask, bool fault)
350 			__acquires(&umem_odp->umem_mutex)
351 {
352 	struct task_struct *owning_process  = NULL;
353 	struct mm_struct *owning_mm = umem_odp->umem.owning_mm;
354 	int pfn_index, dma_index, ret = 0, start_idx;
355 	unsigned int page_shift, hmm_order, pfn_start_idx;
356 	unsigned long num_pfns, current_seq;
357 	struct hmm_range range = {};
358 	unsigned long timeout;
359 
360 	if (access_mask == 0)
361 		return -EINVAL;
362 
363 	if (user_virt < ib_umem_start(umem_odp) ||
364 	    user_virt + bcnt > ib_umem_end(umem_odp))
365 		return -EFAULT;
366 
367 	page_shift = umem_odp->page_shift;
368 
369 	/*
370 	 * owning_process is allowed to be NULL, this means somehow the mm is
371 	 * existing beyond the lifetime of the originating process.. Presumably
372 	 * mmget_not_zero will fail in this case.
373 	 */
374 	owning_process = get_pid_task(umem_odp->tgid, PIDTYPE_PID);
375 	if (!owning_process || !mmget_not_zero(owning_mm)) {
376 		ret = -EINVAL;
377 		goto out_put_task;
378 	}
379 
380 	range.notifier = &umem_odp->notifier;
381 	range.start = ALIGN_DOWN(user_virt, 1UL << page_shift);
382 	range.end = ALIGN(user_virt + bcnt, 1UL << page_shift);
383 	pfn_start_idx = (range.start - ib_umem_start(umem_odp)) >> PAGE_SHIFT;
384 	num_pfns = (range.end - range.start) >> PAGE_SHIFT;
385 	if (fault) {
386 		range.default_flags = HMM_PFN_REQ_FAULT;
387 
388 		if (access_mask & ODP_WRITE_ALLOWED_BIT)
389 			range.default_flags |= HMM_PFN_REQ_WRITE;
390 	}
391 
392 	range.hmm_pfns = &(umem_odp->pfn_list[pfn_start_idx]);
393 	timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
394 
395 retry:
396 	current_seq = range.notifier_seq =
397 		mmu_interval_read_begin(&umem_odp->notifier);
398 
399 	mmap_read_lock(owning_mm);
400 	ret = hmm_range_fault(&range);
401 	mmap_read_unlock(owning_mm);
402 	if (unlikely(ret)) {
403 		if (ret == -EBUSY && !time_after(jiffies, timeout))
404 			goto retry;
405 		goto out_put_mm;
406 	}
407 
408 	start_idx = (range.start - ib_umem_start(umem_odp)) >> page_shift;
409 	dma_index = start_idx;
410 
411 	mutex_lock(&umem_odp->umem_mutex);
412 	if (mmu_interval_read_retry(&umem_odp->notifier, current_seq)) {
413 		mutex_unlock(&umem_odp->umem_mutex);
414 		goto retry;
415 	}
416 
417 	for (pfn_index = 0; pfn_index < num_pfns;
418 		pfn_index += 1 << (page_shift - PAGE_SHIFT), dma_index++) {
419 
420 		if (fault) {
421 			/*
422 			 * Since we asked for hmm_range_fault() to populate
423 			 * pages it shouldn't return an error entry on success.
424 			 */
425 			WARN_ON(range.hmm_pfns[pfn_index] & HMM_PFN_ERROR);
426 			WARN_ON(!(range.hmm_pfns[pfn_index] & HMM_PFN_VALID));
427 		} else {
428 			if (!(range.hmm_pfns[pfn_index] & HMM_PFN_VALID)) {
429 				WARN_ON(umem_odp->dma_list[dma_index]);
430 				continue;
431 			}
432 			access_mask = ODP_READ_ALLOWED_BIT;
433 			if (range.hmm_pfns[pfn_index] & HMM_PFN_WRITE)
434 				access_mask |= ODP_WRITE_ALLOWED_BIT;
435 		}
436 
437 		hmm_order = hmm_pfn_to_map_order(range.hmm_pfns[pfn_index]);
438 		/* If a hugepage was detected and ODP wasn't set for, the umem
439 		 * page_shift will be used, the opposite case is an error.
440 		 */
441 		if (hmm_order + PAGE_SHIFT < page_shift) {
442 			ret = -EINVAL;
443 			ibdev_dbg(umem_odp->umem.ibdev,
444 				  "%s: un-expected hmm_order %u, page_shift %u\n",
445 				  __func__, hmm_order, page_shift);
446 			break;
447 		}
448 
449 		ret = ib_umem_odp_map_dma_single_page(
450 				umem_odp, dma_index, hmm_pfn_to_page(range.hmm_pfns[pfn_index]),
451 				access_mask);
452 		if (ret < 0) {
453 			ibdev_dbg(umem_odp->umem.ibdev,
454 				  "ib_umem_odp_map_dma_single_page failed with error %d\n", ret);
455 			break;
456 		}
457 	}
458 	/* upon sucesss lock should stay on hold for the callee */
459 	if (!ret)
460 		ret = dma_index - start_idx;
461 	else
462 		mutex_unlock(&umem_odp->umem_mutex);
463 
464 out_put_mm:
465 	mmput(owning_mm);
466 out_put_task:
467 	if (owning_process)
468 		put_task_struct(owning_process);
469 	return ret;
470 }
471 EXPORT_SYMBOL(ib_umem_odp_map_dma_and_lock);
472 
473 void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt,
474 				 u64 bound)
475 {
476 	dma_addr_t dma_addr;
477 	dma_addr_t dma;
478 	int idx;
479 	u64 addr;
480 	struct ib_device *dev = umem_odp->umem.ibdev;
481 
482 	lockdep_assert_held(&umem_odp->umem_mutex);
483 
484 	virt = max_t(u64, virt, ib_umem_start(umem_odp));
485 	bound = min_t(u64, bound, ib_umem_end(umem_odp));
486 	for (addr = virt; addr < bound; addr += BIT(umem_odp->page_shift)) {
487 		idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
488 		dma = umem_odp->dma_list[idx];
489 
490 		/* The access flags guaranteed a valid DMA address in case was NULL */
491 		if (dma) {
492 			unsigned long pfn_idx = (addr - ib_umem_start(umem_odp)) >> PAGE_SHIFT;
493 			struct page *page = hmm_pfn_to_page(umem_odp->pfn_list[pfn_idx]);
494 
495 			dma_addr = dma & ODP_DMA_ADDR_MASK;
496 			ib_dma_unmap_page(dev, dma_addr,
497 					  BIT(umem_odp->page_shift),
498 					  DMA_BIDIRECTIONAL);
499 			if (dma & ODP_WRITE_ALLOWED_BIT) {
500 				struct page *head_page = compound_head(page);
501 				/*
502 				 * set_page_dirty prefers being called with
503 				 * the page lock. However, MMU notifiers are
504 				 * called sometimes with and sometimes without
505 				 * the lock. We rely on the umem_mutex instead
506 				 * to prevent other mmu notifiers from
507 				 * continuing and allowing the page mapping to
508 				 * be removed.
509 				 */
510 				set_page_dirty(head_page);
511 			}
512 			umem_odp->dma_list[idx] = 0;
513 			umem_odp->npages--;
514 		}
515 	}
516 }
517 EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages);
518