xref: /openbmc/linux/drivers/infiniband/core/umem.c (revision a44e4f3a)
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  * @dmasync: flush in-flight DMA when the memory region is written
189  */
190 struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr,
191 			    size_t size, int access, int dmasync)
192 {
193 	struct ib_ucontext *context;
194 	struct ib_umem *umem;
195 	struct page **page_list;
196 	unsigned long lock_limit;
197 	unsigned long new_pinned;
198 	unsigned long cur_base;
199 	struct mm_struct *mm;
200 	unsigned long npages;
201 	int ret;
202 	unsigned long dma_attrs = 0;
203 	struct scatterlist *sg;
204 	unsigned int gup_flags = FOLL_WRITE;
205 
206 	if (!udata)
207 		return ERR_PTR(-EIO);
208 
209 	context = container_of(udata, struct uverbs_attr_bundle, driver_udata)
210 			  ->context;
211 	if (!context)
212 		return ERR_PTR(-EIO);
213 
214 	if (dmasync)
215 		dma_attrs |= DMA_ATTR_WRITE_BARRIER;
216 
217 	/*
218 	 * If the combination of the addr and size requested for this memory
219 	 * region causes an integer overflow, return error.
220 	 */
221 	if (((addr + size) < addr) ||
222 	    PAGE_ALIGN(addr + size) < (addr + size))
223 		return ERR_PTR(-EINVAL);
224 
225 	if (!can_do_mlock())
226 		return ERR_PTR(-EPERM);
227 
228 	if (access & IB_ACCESS_ON_DEMAND)
229 		return ERR_PTR(-EOPNOTSUPP);
230 
231 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
232 	if (!umem)
233 		return ERR_PTR(-ENOMEM);
234 	umem->ibdev = context->device;
235 	umem->length     = size;
236 	umem->address    = addr;
237 	umem->writable   = ib_access_writable(access);
238 	umem->owning_mm = mm = current->mm;
239 	mmgrab(mm);
240 
241 	page_list = (struct page **) __get_free_page(GFP_KERNEL);
242 	if (!page_list) {
243 		ret = -ENOMEM;
244 		goto umem_kfree;
245 	}
246 
247 	npages = ib_umem_num_pages(umem);
248 	if (npages == 0 || npages > UINT_MAX) {
249 		ret = -EINVAL;
250 		goto out;
251 	}
252 
253 	lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
254 
255 	new_pinned = atomic64_add_return(npages, &mm->pinned_vm);
256 	if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) {
257 		atomic64_sub(npages, &mm->pinned_vm);
258 		ret = -ENOMEM;
259 		goto out;
260 	}
261 
262 	cur_base = addr & PAGE_MASK;
263 
264 	ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL);
265 	if (ret)
266 		goto vma;
267 
268 	if (!umem->writable)
269 		gup_flags |= FOLL_FORCE;
270 
271 	sg = umem->sg_head.sgl;
272 
273 	while (npages) {
274 		down_read(&mm->mmap_sem);
275 		ret = get_user_pages(cur_base,
276 				     min_t(unsigned long, npages,
277 					   PAGE_SIZE / sizeof (struct page *)),
278 				     gup_flags | FOLL_LONGTERM,
279 				     page_list, NULL);
280 		if (ret < 0) {
281 			up_read(&mm->mmap_sem);
282 			goto umem_release;
283 		}
284 
285 		cur_base += ret * PAGE_SIZE;
286 		npages   -= ret;
287 
288 		sg = ib_umem_add_sg_table(sg, page_list, ret,
289 			dma_get_max_seg_size(context->device->dma_device),
290 			&umem->sg_nents);
291 
292 		up_read(&mm->mmap_sem);
293 	}
294 
295 	sg_mark_end(sg);
296 
297 	umem->nmap = ib_dma_map_sg_attrs(context->device,
298 				  umem->sg_head.sgl,
299 				  umem->sg_nents,
300 				  DMA_BIDIRECTIONAL,
301 				  dma_attrs);
302 
303 	if (!umem->nmap) {
304 		ret = -ENOMEM;
305 		goto umem_release;
306 	}
307 
308 	ret = 0;
309 	goto out;
310 
311 umem_release:
312 	__ib_umem_release(context->device, umem, 0);
313 vma:
314 	atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
315 out:
316 	free_page((unsigned long) page_list);
317 umem_kfree:
318 	if (ret) {
319 		mmdrop(umem->owning_mm);
320 		kfree(umem);
321 	}
322 	return ret ? ERR_PTR(ret) : umem;
323 }
324 EXPORT_SYMBOL(ib_umem_get);
325 
326 /**
327  * ib_umem_release - release memory pinned with ib_umem_get
328  * @umem: umem struct to release
329  */
330 void ib_umem_release(struct ib_umem *umem)
331 {
332 	if (!umem)
333 		return;
334 	if (umem->is_odp)
335 		return ib_umem_odp_release(to_ib_umem_odp(umem));
336 
337 	__ib_umem_release(umem->ibdev, umem, 1);
338 
339 	atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
340 	mmdrop(umem->owning_mm);
341 	kfree(umem);
342 }
343 EXPORT_SYMBOL(ib_umem_release);
344 
345 int ib_umem_page_count(struct ib_umem *umem)
346 {
347 	int i, n = 0;
348 	struct scatterlist *sg;
349 
350 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
351 		n += sg_dma_len(sg) >> PAGE_SHIFT;
352 
353 	return n;
354 }
355 EXPORT_SYMBOL(ib_umem_page_count);
356 
357 /*
358  * Copy from the given ib_umem's pages to the given buffer.
359  *
360  * umem - the umem to copy from
361  * offset - offset to start copying from
362  * dst - destination buffer
363  * length - buffer length
364  *
365  * Returns 0 on success, or an error code.
366  */
367 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
368 		      size_t length)
369 {
370 	size_t end = offset + length;
371 	int ret;
372 
373 	if (offset > umem->length || length > umem->length - offset) {
374 		pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
375 		       offset, umem->length, end);
376 		return -EINVAL;
377 	}
378 
379 	ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length,
380 				 offset + ib_umem_offset(umem));
381 
382 	if (ret < 0)
383 		return ret;
384 	else if (ret != length)
385 		return -EINVAL;
386 	else
387 		return 0;
388 }
389 EXPORT_SYMBOL(ib_umem_copy_from);
390