xref: /openbmc/linux/drivers/infiniband/core/umem.c (revision b830f94f)
1 /*
2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Cisco Systems.  All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 
35 #include <linux/mm.h>
36 #include <linux/dma-mapping.h>
37 #include <linux/sched/signal.h>
38 #include <linux/sched/mm.h>
39 #include <linux/export.h>
40 #include <linux/slab.h>
41 #include <linux/pagemap.h>
42 #include <rdma/ib_umem_odp.h>
43 
44 #include "uverbs.h"
45 
46 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty)
47 {
48 	struct sg_page_iter sg_iter;
49 	struct page *page;
50 
51 	if (umem->nmap > 0)
52 		ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents,
53 				DMA_BIDIRECTIONAL);
54 
55 	for_each_sg_page(umem->sg_head.sgl, &sg_iter, umem->sg_nents, 0) {
56 		page = sg_page_iter_page(&sg_iter);
57 		if (umem->writable && dirty)
58 			put_user_pages_dirty_lock(&page, 1);
59 		else
60 			put_user_page(page);
61 	}
62 
63 	sg_free_table(&umem->sg_head);
64 }
65 
66 /* ib_umem_add_sg_table - Add N contiguous pages to scatter table
67  *
68  * sg: current scatterlist entry
69  * page_list: array of npage struct page pointers
70  * npages: number of pages in page_list
71  * max_seg_sz: maximum segment size in bytes
72  * nents: [out] number of entries in the scatterlist
73  *
74  * Return new end of scatterlist
75  */
76 static struct scatterlist *ib_umem_add_sg_table(struct scatterlist *sg,
77 						struct page **page_list,
78 						unsigned long npages,
79 						unsigned int max_seg_sz,
80 						int *nents)
81 {
82 	unsigned long first_pfn;
83 	unsigned long i = 0;
84 	bool update_cur_sg = false;
85 	bool first = !sg_page(sg);
86 
87 	/* Check if new page_list is contiguous with end of previous page_list.
88 	 * sg->length here is a multiple of PAGE_SIZE and sg->offset is 0.
89 	 */
90 	if (!first && (page_to_pfn(sg_page(sg)) + (sg->length >> PAGE_SHIFT) ==
91 		       page_to_pfn(page_list[0])))
92 		update_cur_sg = true;
93 
94 	while (i != npages) {
95 		unsigned long len;
96 		struct page *first_page = page_list[i];
97 
98 		first_pfn = page_to_pfn(first_page);
99 
100 		/* Compute the number of contiguous pages we have starting
101 		 * at i
102 		 */
103 		for (len = 0; i != npages &&
104 			      first_pfn + len == page_to_pfn(page_list[i]) &&
105 			      len < (max_seg_sz >> PAGE_SHIFT);
106 		     len++)
107 			i++;
108 
109 		/* Squash N contiguous pages from page_list into current sge */
110 		if (update_cur_sg) {
111 			if ((max_seg_sz - sg->length) >= (len << PAGE_SHIFT)) {
112 				sg_set_page(sg, sg_page(sg),
113 					    sg->length + (len << PAGE_SHIFT),
114 					    0);
115 				update_cur_sg = false;
116 				continue;
117 			}
118 			update_cur_sg = false;
119 		}
120 
121 		/* Squash N contiguous pages into next sge or first sge */
122 		if (!first)
123 			sg = sg_next(sg);
124 
125 		(*nents)++;
126 		sg_set_page(sg, first_page, len << PAGE_SHIFT, 0);
127 		first = false;
128 	}
129 
130 	return sg;
131 }
132 
133 /**
134  * ib_umem_find_best_pgsz - Find best HW page size to use for this MR
135  *
136  * @umem: umem struct
137  * @pgsz_bitmap: bitmap of HW supported page sizes
138  * @virt: IOVA
139  *
140  * This helper is intended for HW that support multiple page
141  * sizes but can do only a single page size in an MR.
142  *
143  * Returns 0 if the umem requires page sizes not supported by
144  * the driver to be mapped. Drivers always supporting PAGE_SIZE
145  * or smaller will never see a 0 result.
146  */
147 unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
148 				     unsigned long pgsz_bitmap,
149 				     unsigned long virt)
150 {
151 	struct scatterlist *sg;
152 	unsigned int best_pg_bit;
153 	unsigned long va, pgoff;
154 	dma_addr_t mask;
155 	int i;
156 
157 	/* At minimum, drivers must support PAGE_SIZE or smaller */
158 	if (WARN_ON(!(pgsz_bitmap & GENMASK(PAGE_SHIFT, 0))))
159 		return 0;
160 
161 	va = virt;
162 	/* max page size not to exceed MR length */
163 	mask = roundup_pow_of_two(umem->length);
164 	/* offset into first SGL */
165 	pgoff = umem->address & ~PAGE_MASK;
166 
167 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
168 		/* Walk SGL and reduce max page size if VA/PA bits differ
169 		 * for any address.
170 		 */
171 		mask |= (sg_dma_address(sg) + pgoff) ^ va;
172 		if (i && i != (umem->nmap - 1))
173 			/* restrict by length as well for interior SGEs */
174 			mask |= sg_dma_len(sg);
175 		va += sg_dma_len(sg) - pgoff;
176 		pgoff = 0;
177 	}
178 	best_pg_bit = rdma_find_pg_bit(mask, pgsz_bitmap);
179 
180 	return BIT_ULL(best_pg_bit);
181 }
182 EXPORT_SYMBOL(ib_umem_find_best_pgsz);
183 
184 /**
185  * ib_umem_get - Pin and DMA map userspace memory.
186  *
187  * If access flags indicate ODP memory, avoid pinning. Instead, stores
188  * the mm for future page fault handling in conjunction with MMU notifiers.
189  *
190  * @udata: userspace context to pin memory for
191  * @addr: userspace virtual address to start at
192  * @size: length of region to pin
193  * @access: IB_ACCESS_xxx flags for memory being pinned
194  * @dmasync: flush in-flight DMA when the memory region is written
195  */
196 struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr,
197 			    size_t size, int access, int dmasync)
198 {
199 	struct ib_ucontext *context;
200 	struct ib_umem *umem;
201 	struct page **page_list;
202 	unsigned long lock_limit;
203 	unsigned long new_pinned;
204 	unsigned long cur_base;
205 	struct mm_struct *mm;
206 	unsigned long npages;
207 	int ret;
208 	unsigned long dma_attrs = 0;
209 	struct scatterlist *sg;
210 	unsigned int gup_flags = FOLL_WRITE;
211 
212 	if (!udata)
213 		return ERR_PTR(-EIO);
214 
215 	context = container_of(udata, struct uverbs_attr_bundle, driver_udata)
216 			  ->context;
217 	if (!context)
218 		return ERR_PTR(-EIO);
219 
220 	if (dmasync)
221 		dma_attrs |= DMA_ATTR_WRITE_BARRIER;
222 
223 	/*
224 	 * If the combination of the addr and size requested for this memory
225 	 * region causes an integer overflow, return error.
226 	 */
227 	if (((addr + size) < addr) ||
228 	    PAGE_ALIGN(addr + size) < (addr + size))
229 		return ERR_PTR(-EINVAL);
230 
231 	if (!can_do_mlock())
232 		return ERR_PTR(-EPERM);
233 
234 	if (access & IB_ACCESS_ON_DEMAND) {
235 		umem = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL);
236 		if (!umem)
237 			return ERR_PTR(-ENOMEM);
238 		umem->is_odp = 1;
239 	} else {
240 		umem = kzalloc(sizeof(*umem), GFP_KERNEL);
241 		if (!umem)
242 			return ERR_PTR(-ENOMEM);
243 	}
244 
245 	umem->context    = context;
246 	umem->length     = size;
247 	umem->address    = addr;
248 	umem->writable   = ib_access_writable(access);
249 	umem->owning_mm = mm = current->mm;
250 	mmgrab(mm);
251 
252 	if (access & IB_ACCESS_ON_DEMAND) {
253 		if (WARN_ON_ONCE(!context->invalidate_range)) {
254 			ret = -EINVAL;
255 			goto umem_kfree;
256 		}
257 
258 		ret = ib_umem_odp_get(to_ib_umem_odp(umem), access);
259 		if (ret)
260 			goto umem_kfree;
261 		return umem;
262 	}
263 
264 	page_list = (struct page **) __get_free_page(GFP_KERNEL);
265 	if (!page_list) {
266 		ret = -ENOMEM;
267 		goto umem_kfree;
268 	}
269 
270 	npages = ib_umem_num_pages(umem);
271 	if (npages == 0 || npages > UINT_MAX) {
272 		ret = -EINVAL;
273 		goto out;
274 	}
275 
276 	lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
277 
278 	new_pinned = atomic64_add_return(npages, &mm->pinned_vm);
279 	if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) {
280 		atomic64_sub(npages, &mm->pinned_vm);
281 		ret = -ENOMEM;
282 		goto out;
283 	}
284 
285 	cur_base = addr & PAGE_MASK;
286 
287 	ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
288 	if (ret)
289 		goto vma;
290 
291 	if (!umem->writable)
292 		gup_flags |= FOLL_FORCE;
293 
294 	sg = umem->sg_head.sgl;
295 
296 	while (npages) {
297 		down_read(&mm->mmap_sem);
298 		ret = get_user_pages(cur_base,
299 				     min_t(unsigned long, npages,
300 					   PAGE_SIZE / sizeof (struct page *)),
301 				     gup_flags | FOLL_LONGTERM,
302 				     page_list, NULL);
303 		if (ret < 0) {
304 			up_read(&mm->mmap_sem);
305 			goto umem_release;
306 		}
307 
308 		cur_base += ret * PAGE_SIZE;
309 		npages   -= ret;
310 
311 		sg = ib_umem_add_sg_table(sg, page_list, ret,
312 			dma_get_max_seg_size(context->device->dma_device),
313 			&umem->sg_nents);
314 
315 		up_read(&mm->mmap_sem);
316 	}
317 
318 	sg_mark_end(sg);
319 
320 	umem->nmap = ib_dma_map_sg_attrs(context->device,
321 				  umem->sg_head.sgl,
322 				  umem->sg_nents,
323 				  DMA_BIDIRECTIONAL,
324 				  dma_attrs);
325 
326 	if (!umem->nmap) {
327 		ret = -ENOMEM;
328 		goto umem_release;
329 	}
330 
331 	ret = 0;
332 	goto out;
333 
334 umem_release:
335 	__ib_umem_release(context->device, umem, 0);
336 vma:
337 	atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
338 out:
339 	free_page((unsigned long) page_list);
340 umem_kfree:
341 	if (ret) {
342 		mmdrop(umem->owning_mm);
343 		kfree(umem);
344 	}
345 	return ret ? ERR_PTR(ret) : umem;
346 }
347 EXPORT_SYMBOL(ib_umem_get);
348 
349 static void __ib_umem_release_tail(struct ib_umem *umem)
350 {
351 	mmdrop(umem->owning_mm);
352 	if (umem->is_odp)
353 		kfree(to_ib_umem_odp(umem));
354 	else
355 		kfree(umem);
356 }
357 
358 /**
359  * ib_umem_release - release memory pinned with ib_umem_get
360  * @umem: umem struct to release
361  */
362 void ib_umem_release(struct ib_umem *umem)
363 {
364 	if (!umem)
365 		return;
366 
367 	if (umem->is_odp) {
368 		ib_umem_odp_release(to_ib_umem_odp(umem));
369 		__ib_umem_release_tail(umem);
370 		return;
371 	}
372 
373 	__ib_umem_release(umem->context->device, umem, 1);
374 
375 	atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
376 	__ib_umem_release_tail(umem);
377 }
378 EXPORT_SYMBOL(ib_umem_release);
379 
380 int ib_umem_page_count(struct ib_umem *umem)
381 {
382 	int i;
383 	int n;
384 	struct scatterlist *sg;
385 
386 	if (umem->is_odp)
387 		return ib_umem_num_pages(umem);
388 
389 	n = 0;
390 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
391 		n += sg_dma_len(sg) >> PAGE_SHIFT;
392 
393 	return n;
394 }
395 EXPORT_SYMBOL(ib_umem_page_count);
396 
397 /*
398  * Copy from the given ib_umem's pages to the given buffer.
399  *
400  * umem - the umem to copy from
401  * offset - offset to start copying from
402  * dst - destination buffer
403  * length - buffer length
404  *
405  * Returns 0 on success, or an error code.
406  */
407 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
408 		      size_t length)
409 {
410 	size_t end = offset + length;
411 	int ret;
412 
413 	if (offset > umem->length || length > umem->length - offset) {
414 		pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
415 		       offset, umem->length, end);
416 		return -EINVAL;
417 	}
418 
419 	ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length,
420 				 offset + ib_umem_offset(umem));
421 
422 	if (ret < 0)
423 		return ret;
424 	else if (ret != length)
425 		return -EINVAL;
426 	else
427 		return 0;
428 }
429 EXPORT_SYMBOL(ib_umem_copy_from);
430