xref: /openbmc/linux/drivers/infiniband/core/umem.c (revision dc6a81c3)
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 		unpin_user_pages_dirty_lock(&page, 1, umem->writable && dirty);
58 	}
59 
60 	sg_free_table(&umem->sg_head);
61 }
62 
63 /* ib_umem_add_sg_table - Add N contiguous pages to scatter table
64  *
65  * sg: current scatterlist entry
66  * page_list: array of npage struct page pointers
67  * npages: number of pages in page_list
68  * max_seg_sz: maximum segment size in bytes
69  * nents: [out] number of entries in the scatterlist
70  *
71  * Return new end of scatterlist
72  */
73 static struct scatterlist *ib_umem_add_sg_table(struct scatterlist *sg,
74 						struct page **page_list,
75 						unsigned long npages,
76 						unsigned int max_seg_sz,
77 						int *nents)
78 {
79 	unsigned long first_pfn;
80 	unsigned long i = 0;
81 	bool update_cur_sg = false;
82 	bool first = !sg_page(sg);
83 
84 	/* Check if new page_list is contiguous with end of previous page_list.
85 	 * sg->length here is a multiple of PAGE_SIZE and sg->offset is 0.
86 	 */
87 	if (!first && (page_to_pfn(sg_page(sg)) + (sg->length >> PAGE_SHIFT) ==
88 		       page_to_pfn(page_list[0])))
89 		update_cur_sg = true;
90 
91 	while (i != npages) {
92 		unsigned long len;
93 		struct page *first_page = page_list[i];
94 
95 		first_pfn = page_to_pfn(first_page);
96 
97 		/* Compute the number of contiguous pages we have starting
98 		 * at i
99 		 */
100 		for (len = 0; i != npages &&
101 			      first_pfn + len == page_to_pfn(page_list[i]) &&
102 			      len < (max_seg_sz >> PAGE_SHIFT);
103 		     len++)
104 			i++;
105 
106 		/* Squash N contiguous pages from page_list into current sge */
107 		if (update_cur_sg) {
108 			if ((max_seg_sz - sg->length) >= (len << PAGE_SHIFT)) {
109 				sg_set_page(sg, sg_page(sg),
110 					    sg->length + (len << PAGE_SHIFT),
111 					    0);
112 				update_cur_sg = false;
113 				continue;
114 			}
115 			update_cur_sg = false;
116 		}
117 
118 		/* Squash N contiguous pages into next sge or first sge */
119 		if (!first)
120 			sg = sg_next(sg);
121 
122 		(*nents)++;
123 		sg_set_page(sg, first_page, len << PAGE_SHIFT, 0);
124 		first = false;
125 	}
126 
127 	return sg;
128 }
129 
130 /**
131  * ib_umem_find_best_pgsz - Find best HW page size to use for this MR
132  *
133  * @umem: umem struct
134  * @pgsz_bitmap: bitmap of HW supported page sizes
135  * @virt: IOVA
136  *
137  * This helper is intended for HW that support multiple page
138  * sizes but can do only a single page size in an MR.
139  *
140  * Returns 0 if the umem requires page sizes not supported by
141  * the driver to be mapped. Drivers always supporting PAGE_SIZE
142  * or smaller will never see a 0 result.
143  */
144 unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
145 				     unsigned long pgsz_bitmap,
146 				     unsigned long virt)
147 {
148 	struct scatterlist *sg;
149 	unsigned int best_pg_bit;
150 	unsigned long va, pgoff;
151 	dma_addr_t mask;
152 	int i;
153 
154 	/* At minimum, drivers must support PAGE_SIZE or smaller */
155 	if (WARN_ON(!(pgsz_bitmap & GENMASK(PAGE_SHIFT, 0))))
156 		return 0;
157 
158 	va = virt;
159 	/* max page size not to exceed MR length */
160 	mask = roundup_pow_of_two(umem->length);
161 	/* offset into first SGL */
162 	pgoff = umem->address & ~PAGE_MASK;
163 
164 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
165 		/* Walk SGL and reduce max page size if VA/PA bits differ
166 		 * for any address.
167 		 */
168 		mask |= (sg_dma_address(sg) + pgoff) ^ va;
169 		va += sg_dma_len(sg) - pgoff;
170 		/* Except for the last entry, the ending iova alignment sets
171 		 * the maximum possible page size as the low bits of the iova
172 		 * must be zero when starting the next chunk.
173 		 */
174 		if (i != (umem->nmap - 1))
175 			mask |= va;
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  * @device: IB device to connect UMEM
188  * @addr: userspace virtual address to start at
189  * @size: length of region to pin
190  * @access: IB_ACCESS_xxx flags for memory being pinned
191  */
192 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
193 			    size_t size, int access)
194 {
195 	struct ib_umem *umem;
196 	struct page **page_list;
197 	unsigned long lock_limit;
198 	unsigned long new_pinned;
199 	unsigned long cur_base;
200 	struct mm_struct *mm;
201 	unsigned long npages;
202 	int ret;
203 	struct scatterlist *sg;
204 	unsigned int gup_flags = FOLL_WRITE;
205 
206 	/*
207 	 * If the combination of the addr and size requested for this memory
208 	 * region causes an integer overflow, return error.
209 	 */
210 	if (((addr + size) < addr) ||
211 	    PAGE_ALIGN(addr + size) < (addr + size))
212 		return ERR_PTR(-EINVAL);
213 
214 	if (!can_do_mlock())
215 		return ERR_PTR(-EPERM);
216 
217 	if (access & IB_ACCESS_ON_DEMAND)
218 		return ERR_PTR(-EOPNOTSUPP);
219 
220 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
221 	if (!umem)
222 		return ERR_PTR(-ENOMEM);
223 	umem->ibdev      = device;
224 	umem->length     = size;
225 	umem->address    = addr;
226 	umem->writable   = ib_access_writable(access);
227 	umem->owning_mm = mm = current->mm;
228 	mmgrab(mm);
229 
230 	page_list = (struct page **) __get_free_page(GFP_KERNEL);
231 	if (!page_list) {
232 		ret = -ENOMEM;
233 		goto umem_kfree;
234 	}
235 
236 	npages = ib_umem_num_pages(umem);
237 	if (npages == 0 || npages > UINT_MAX) {
238 		ret = -EINVAL;
239 		goto out;
240 	}
241 
242 	lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
243 
244 	new_pinned = atomic64_add_return(npages, &mm->pinned_vm);
245 	if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) {
246 		atomic64_sub(npages, &mm->pinned_vm);
247 		ret = -ENOMEM;
248 		goto out;
249 	}
250 
251 	cur_base = addr & PAGE_MASK;
252 
253 	ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
254 	if (ret)
255 		goto vma;
256 
257 	if (!umem->writable)
258 		gup_flags |= FOLL_FORCE;
259 
260 	sg = umem->sg_head.sgl;
261 
262 	while (npages) {
263 		ret = pin_user_pages_fast(cur_base,
264 					  min_t(unsigned long, npages,
265 						PAGE_SIZE /
266 						sizeof(struct page *)),
267 					  gup_flags | FOLL_LONGTERM, page_list);
268 		if (ret < 0)
269 			goto umem_release;
270 
271 		cur_base += ret * PAGE_SIZE;
272 		npages   -= ret;
273 
274 		sg = ib_umem_add_sg_table(sg, page_list, ret,
275 			dma_get_max_seg_size(device->dma_device),
276 			&umem->sg_nents);
277 	}
278 
279 	sg_mark_end(sg);
280 
281 	umem->nmap = ib_dma_map_sg(device,
282 				   umem->sg_head.sgl,
283 				   umem->sg_nents,
284 				   DMA_BIDIRECTIONAL);
285 
286 	if (!umem->nmap) {
287 		ret = -ENOMEM;
288 		goto umem_release;
289 	}
290 
291 	ret = 0;
292 	goto out;
293 
294 umem_release:
295 	__ib_umem_release(device, umem, 0);
296 vma:
297 	atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
298 out:
299 	free_page((unsigned long) page_list);
300 umem_kfree:
301 	if (ret) {
302 		mmdrop(umem->owning_mm);
303 		kfree(umem);
304 	}
305 	return ret ? ERR_PTR(ret) : umem;
306 }
307 EXPORT_SYMBOL(ib_umem_get);
308 
309 /**
310  * ib_umem_release - release memory pinned with ib_umem_get
311  * @umem: umem struct to release
312  */
313 void ib_umem_release(struct ib_umem *umem)
314 {
315 	if (!umem)
316 		return;
317 	if (umem->is_odp)
318 		return ib_umem_odp_release(to_ib_umem_odp(umem));
319 
320 	__ib_umem_release(umem->ibdev, umem, 1);
321 
322 	atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
323 	mmdrop(umem->owning_mm);
324 	kfree(umem);
325 }
326 EXPORT_SYMBOL(ib_umem_release);
327 
328 int ib_umem_page_count(struct ib_umem *umem)
329 {
330 	int i, n = 0;
331 	struct scatterlist *sg;
332 
333 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
334 		n += sg_dma_len(sg) >> PAGE_SHIFT;
335 
336 	return n;
337 }
338 EXPORT_SYMBOL(ib_umem_page_count);
339 
340 /*
341  * Copy from the given ib_umem's pages to the given buffer.
342  *
343  * umem - the umem to copy from
344  * offset - offset to start copying from
345  * dst - destination buffer
346  * length - buffer length
347  *
348  * Returns 0 on success, or an error code.
349  */
350 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
351 		      size_t length)
352 {
353 	size_t end = offset + length;
354 	int ret;
355 
356 	if (offset > umem->length || length > umem->length - offset) {
357 		pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
358 		       offset, umem->length, end);
359 		return -EINVAL;
360 	}
361 
362 	ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length,
363 				 offset + ib_umem_offset(umem));
364 
365 	if (ret < 0)
366 		return ret;
367 	else if (ret != length)
368 		return -EINVAL;
369 	else
370 		return 0;
371 }
372 EXPORT_SYMBOL(ib_umem_copy_from);
373