xref: /openbmc/linux/drivers/infiniband/core/umem.c (revision be122522)
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 		put_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 		if (i && i != (umem->nmap - 1))
170 			/* restrict by length as well for interior SGEs */
171 			mask |= sg_dma_len(sg);
172 		va += sg_dma_len(sg) - pgoff;
173 		pgoff = 0;
174 	}
175 	best_pg_bit = rdma_find_pg_bit(mask, pgsz_bitmap);
176 
177 	return BIT_ULL(best_pg_bit);
178 }
179 EXPORT_SYMBOL(ib_umem_find_best_pgsz);
180 
181 /**
182  * ib_umem_get - Pin and DMA map userspace memory.
183  *
184  * @udata: userspace context to pin memory for
185  * @addr: userspace virtual address to start at
186  * @size: length of region to pin
187  * @access: IB_ACCESS_xxx flags for memory being pinned
188  */
189 struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr,
190 			    size_t size, int access)
191 {
192 	struct ib_ucontext *context;
193 	struct ib_umem *umem;
194 	struct page **page_list;
195 	unsigned long lock_limit;
196 	unsigned long new_pinned;
197 	unsigned long cur_base;
198 	struct mm_struct *mm;
199 	unsigned long npages;
200 	int ret;
201 	struct scatterlist *sg;
202 	unsigned int gup_flags = FOLL_WRITE;
203 
204 	if (!udata)
205 		return ERR_PTR(-EIO);
206 
207 	context = container_of(udata, struct uverbs_attr_bundle, driver_udata)
208 			  ->context;
209 	if (!context)
210 		return ERR_PTR(-EIO);
211 
212 	/*
213 	 * If the combination of the addr and size requested for this memory
214 	 * region causes an integer overflow, return error.
215 	 */
216 	if (((addr + size) < addr) ||
217 	    PAGE_ALIGN(addr + size) < (addr + size))
218 		return ERR_PTR(-EINVAL);
219 
220 	if (!can_do_mlock())
221 		return ERR_PTR(-EPERM);
222 
223 	if (access & IB_ACCESS_ON_DEMAND)
224 		return ERR_PTR(-EOPNOTSUPP);
225 
226 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
227 	if (!umem)
228 		return ERR_PTR(-ENOMEM);
229 	umem->ibdev = context->device;
230 	umem->length     = size;
231 	umem->address    = addr;
232 	umem->writable   = ib_access_writable(access);
233 	umem->owning_mm = mm = current->mm;
234 	mmgrab(mm);
235 
236 	page_list = (struct page **) __get_free_page(GFP_KERNEL);
237 	if (!page_list) {
238 		ret = -ENOMEM;
239 		goto umem_kfree;
240 	}
241 
242 	npages = ib_umem_num_pages(umem);
243 	if (npages == 0 || npages > UINT_MAX) {
244 		ret = -EINVAL;
245 		goto out;
246 	}
247 
248 	lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
249 
250 	new_pinned = atomic64_add_return(npages, &mm->pinned_vm);
251 	if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) {
252 		atomic64_sub(npages, &mm->pinned_vm);
253 		ret = -ENOMEM;
254 		goto out;
255 	}
256 
257 	cur_base = addr & PAGE_MASK;
258 
259 	ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
260 	if (ret)
261 		goto vma;
262 
263 	if (!umem->writable)
264 		gup_flags |= FOLL_FORCE;
265 
266 	sg = umem->sg_head.sgl;
267 
268 	while (npages) {
269 		down_read(&mm->mmap_sem);
270 		ret = get_user_pages(cur_base,
271 				     min_t(unsigned long, npages,
272 					   PAGE_SIZE / sizeof (struct page *)),
273 				     gup_flags | FOLL_LONGTERM,
274 				     page_list, NULL);
275 		if (ret < 0) {
276 			up_read(&mm->mmap_sem);
277 			goto umem_release;
278 		}
279 
280 		cur_base += ret * PAGE_SIZE;
281 		npages   -= ret;
282 
283 		sg = ib_umem_add_sg_table(sg, page_list, ret,
284 			dma_get_max_seg_size(context->device->dma_device),
285 			&umem->sg_nents);
286 
287 		up_read(&mm->mmap_sem);
288 	}
289 
290 	sg_mark_end(sg);
291 
292 	umem->nmap = ib_dma_map_sg(context->device,
293 				  umem->sg_head.sgl,
294 				  umem->sg_nents,
295 				  DMA_BIDIRECTIONAL);
296 
297 	if (!umem->nmap) {
298 		ret = -ENOMEM;
299 		goto umem_release;
300 	}
301 
302 	ret = 0;
303 	goto out;
304 
305 umem_release:
306 	__ib_umem_release(context->device, umem, 0);
307 vma:
308 	atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
309 out:
310 	free_page((unsigned long) page_list);
311 umem_kfree:
312 	if (ret) {
313 		mmdrop(umem->owning_mm);
314 		kfree(umem);
315 	}
316 	return ret ? ERR_PTR(ret) : umem;
317 }
318 EXPORT_SYMBOL(ib_umem_get);
319 
320 /**
321  * ib_umem_release - release memory pinned with ib_umem_get
322  * @umem: umem struct to release
323  */
324 void ib_umem_release(struct ib_umem *umem)
325 {
326 	if (!umem)
327 		return;
328 	if (umem->is_odp)
329 		return ib_umem_odp_release(to_ib_umem_odp(umem));
330 
331 	__ib_umem_release(umem->ibdev, umem, 1);
332 
333 	atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
334 	mmdrop(umem->owning_mm);
335 	kfree(umem);
336 }
337 EXPORT_SYMBOL(ib_umem_release);
338 
339 int ib_umem_page_count(struct ib_umem *umem)
340 {
341 	int i, n = 0;
342 	struct scatterlist *sg;
343 
344 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
345 		n += sg_dma_len(sg) >> PAGE_SHIFT;
346 
347 	return n;
348 }
349 EXPORT_SYMBOL(ib_umem_page_count);
350 
351 /*
352  * Copy from the given ib_umem's pages to the given buffer.
353  *
354  * umem - the umem to copy from
355  * offset - offset to start copying from
356  * dst - destination buffer
357  * length - buffer length
358  *
359  * Returns 0 on success, or an error code.
360  */
361 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
362 		      size_t length)
363 {
364 	size_t end = offset + length;
365 	int ret;
366 
367 	if (offset > umem->length || length > umem->length - offset) {
368 		pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
369 		       offset, umem->length, end);
370 		return -EINVAL;
371 	}
372 
373 	ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length,
374 				 offset + ib_umem_offset(umem));
375 
376 	if (ret < 0)
377 		return ret;
378 	else if (ret != length)
379 		return -EINVAL;
380 	else
381 		return 0;
382 }
383 EXPORT_SYMBOL(ib_umem_copy_from);
384