xref: /openbmc/linux/drivers/infiniband/core/umem.c (revision 4e95bc26)
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 (!PageDirty(page) && umem->writable && dirty)
58 			set_page_dirty_lock(page);
59 		put_page(page);
60 	}
61 
62 	sg_free_table(&umem->sg_head);
63 }
64 
65 /* ib_umem_add_sg_table - Add N contiguous pages to scatter table
66  *
67  * sg: current scatterlist entry
68  * page_list: array of npage struct page pointers
69  * npages: number of pages in page_list
70  * max_seg_sz: maximum segment size in bytes
71  * nents: [out] number of entries in the scatterlist
72  *
73  * Return new end of scatterlist
74  */
75 static struct scatterlist *ib_umem_add_sg_table(struct scatterlist *sg,
76 						struct page **page_list,
77 						unsigned long npages,
78 						unsigned int max_seg_sz,
79 						int *nents)
80 {
81 	unsigned long first_pfn;
82 	unsigned long i = 0;
83 	bool update_cur_sg = false;
84 	bool first = !sg_page(sg);
85 
86 	/* Check if new page_list is contiguous with end of previous page_list.
87 	 * sg->length here is a multiple of PAGE_SIZE and sg->offset is 0.
88 	 */
89 	if (!first && (page_to_pfn(sg_page(sg)) + (sg->length >> PAGE_SHIFT) ==
90 		       page_to_pfn(page_list[0])))
91 		update_cur_sg = true;
92 
93 	while (i != npages) {
94 		unsigned long len;
95 		struct page *first_page = page_list[i];
96 
97 		first_pfn = page_to_pfn(first_page);
98 
99 		/* Compute the number of contiguous pages we have starting
100 		 * at i
101 		 */
102 		for (len = 0; i != npages &&
103 			      first_pfn + len == page_to_pfn(page_list[i]) &&
104 			      len < (max_seg_sz >> PAGE_SHIFT);
105 		     len++)
106 			i++;
107 
108 		/* Squash N contiguous pages from page_list into current sge */
109 		if (update_cur_sg) {
110 			if ((max_seg_sz - sg->length) >= (len << PAGE_SHIFT)) {
111 				sg_set_page(sg, sg_page(sg),
112 					    sg->length + (len << PAGE_SHIFT),
113 					    0);
114 				update_cur_sg = false;
115 				continue;
116 			}
117 			update_cur_sg = false;
118 		}
119 
120 		/* Squash N contiguous pages into next sge or first sge */
121 		if (!first)
122 			sg = sg_next(sg);
123 
124 		(*nents)++;
125 		sg_set_page(sg, first_page, len << PAGE_SHIFT, 0);
126 		first = false;
127 	}
128 
129 	return sg;
130 }
131 
132 /**
133  * ib_umem_find_best_pgsz - Find best HW page size to use for this MR
134  *
135  * @umem: umem struct
136  * @pgsz_bitmap: bitmap of HW supported page sizes
137  * @virt: IOVA
138  *
139  * This helper is intended for HW that support multiple page
140  * sizes but can do only a single page size in an MR.
141  *
142  * Returns 0 if the umem requires page sizes not supported by
143  * the driver to be mapped. Drivers always supporting PAGE_SIZE
144  * or smaller will never see a 0 result.
145  */
146 unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
147 				     unsigned long pgsz_bitmap,
148 				     unsigned long virt)
149 {
150 	struct scatterlist *sg;
151 	unsigned int best_pg_bit;
152 	unsigned long va, pgoff;
153 	dma_addr_t mask;
154 	int i;
155 
156 	/* At minimum, drivers must support PAGE_SIZE or smaller */
157 	if (WARN_ON(!(pgsz_bitmap & GENMASK(PAGE_SHIFT, 0))))
158 		return 0;
159 
160 	va = virt;
161 	/* max page size not to exceed MR length */
162 	mask = roundup_pow_of_two(umem->length);
163 	/* offset into first SGL */
164 	pgoff = umem->address & ~PAGE_MASK;
165 
166 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) {
167 		/* Walk SGL and reduce max page size if VA/PA bits differ
168 		 * for any address.
169 		 */
170 		mask |= (sg_dma_address(sg) + pgoff) ^ va;
171 		if (i && i != (umem->nmap - 1))
172 			/* restrict by length as well for interior SGEs */
173 			mask |= sg_dma_len(sg);
174 		va += sg_dma_len(sg) - pgoff;
175 		pgoff = 0;
176 	}
177 	best_pg_bit = rdma_find_pg_bit(mask, pgsz_bitmap);
178 
179 	return BIT_ULL(best_pg_bit);
180 }
181 EXPORT_SYMBOL(ib_umem_find_best_pgsz);
182 
183 /**
184  * ib_umem_get - Pin and DMA map userspace memory.
185  *
186  * If access flags indicate ODP memory, avoid pinning. Instead, stores
187  * the mm for future page fault handling in conjunction with MMU notifiers.
188  *
189  * @udata: userspace context to pin memory for
190  * @addr: userspace virtual address to start at
191  * @size: length of region to pin
192  * @access: IB_ACCESS_xxx flags for memory being pinned
193  * @dmasync: flush in-flight DMA when the memory region is written
194  */
195 struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr,
196 			    size_t size, int access, int dmasync)
197 {
198 	struct ib_ucontext *context;
199 	struct ib_umem *umem;
200 	struct page **page_list;
201 	unsigned long lock_limit;
202 	unsigned long new_pinned;
203 	unsigned long cur_base;
204 	struct mm_struct *mm;
205 	unsigned long npages;
206 	int ret;
207 	unsigned long dma_attrs = 0;
208 	struct scatterlist *sg;
209 	unsigned int gup_flags = FOLL_WRITE;
210 
211 	if (!udata)
212 		return ERR_PTR(-EIO);
213 
214 	context = container_of(udata, struct uverbs_attr_bundle, driver_udata)
215 			  ->context;
216 	if (!context)
217 		return ERR_PTR(-EIO);
218 
219 	if (dmasync)
220 		dma_attrs |= DMA_ATTR_WRITE_BARRIER;
221 
222 	/*
223 	 * If the combination of the addr and size requested for this memory
224 	 * region causes an integer overflow, return error.
225 	 */
226 	if (((addr + size) < addr) ||
227 	    PAGE_ALIGN(addr + size) < (addr + size))
228 		return ERR_PTR(-EINVAL);
229 
230 	if (!can_do_mlock())
231 		return ERR_PTR(-EPERM);
232 
233 	if (access & IB_ACCESS_ON_DEMAND) {
234 		umem = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL);
235 		if (!umem)
236 			return ERR_PTR(-ENOMEM);
237 		umem->is_odp = 1;
238 	} else {
239 		umem = kzalloc(sizeof(*umem), GFP_KERNEL);
240 		if (!umem)
241 			return ERR_PTR(-ENOMEM);
242 	}
243 
244 	umem->context    = context;
245 	umem->length     = size;
246 	umem->address    = addr;
247 	umem->page_shift = PAGE_SHIFT;
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->is_odp) {
365 		ib_umem_odp_release(to_ib_umem_odp(umem));
366 		__ib_umem_release_tail(umem);
367 		return;
368 	}
369 
370 	__ib_umem_release(umem->context->device, umem, 1);
371 
372 	atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm);
373 	__ib_umem_release_tail(umem);
374 }
375 EXPORT_SYMBOL(ib_umem_release);
376 
377 int ib_umem_page_count(struct ib_umem *umem)
378 {
379 	int i;
380 	int n;
381 	struct scatterlist *sg;
382 
383 	if (umem->is_odp)
384 		return ib_umem_num_pages(umem);
385 
386 	n = 0;
387 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i)
388 		n += sg_dma_len(sg) >> umem->page_shift;
389 
390 	return n;
391 }
392 EXPORT_SYMBOL(ib_umem_page_count);
393 
394 /*
395  * Copy from the given ib_umem's pages to the given buffer.
396  *
397  * umem - the umem to copy from
398  * offset - offset to start copying from
399  * dst - destination buffer
400  * length - buffer length
401  *
402  * Returns 0 on success, or an error code.
403  */
404 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset,
405 		      size_t length)
406 {
407 	size_t end = offset + length;
408 	int ret;
409 
410 	if (offset > umem->length || length > umem->length - offset) {
411 		pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n",
412 		       offset, umem->length, end);
413 		return -EINVAL;
414 	}
415 
416 	ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length,
417 				 offset + ib_umem_offset(umem));
418 
419 	if (ret < 0)
420 		return ret;
421 	else if (ret != length)
422 		return -EINVAL;
423 	else
424 		return 0;
425 }
426 EXPORT_SYMBOL(ib_umem_copy_from);
427