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 * @udata: userspace context to pin memory for 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 * @dmasync: flush in-flight DMA when the memory region is written 192 */ 193 struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr, 194 size_t size, int access, int dmasync) 195 { 196 struct ib_ucontext *context; 197 struct ib_umem *umem; 198 struct page **page_list; 199 unsigned long lock_limit; 200 unsigned long new_pinned; 201 unsigned long cur_base; 202 struct mm_struct *mm; 203 unsigned long npages; 204 int ret; 205 unsigned long dma_attrs = 0; 206 struct scatterlist *sg; 207 unsigned int gup_flags = FOLL_WRITE; 208 209 if (!udata) 210 return ERR_PTR(-EIO); 211 212 context = container_of(udata, struct uverbs_attr_bundle, driver_udata) 213 ->context; 214 if (!context) 215 return ERR_PTR(-EIO); 216 217 if (dmasync) 218 dma_attrs |= DMA_ATTR_WRITE_BARRIER; 219 220 /* 221 * If the combination of the addr and size requested for this memory 222 * region causes an integer overflow, return error. 223 */ 224 if (((addr + size) < addr) || 225 PAGE_ALIGN(addr + size) < (addr + size)) 226 return ERR_PTR(-EINVAL); 227 228 if (!can_do_mlock()) 229 return ERR_PTR(-EPERM); 230 231 if (access & IB_ACCESS_ON_DEMAND) 232 return ERR_PTR(-EOPNOTSUPP); 233 234 umem = kzalloc(sizeof(*umem), GFP_KERNEL); 235 if (!umem) 236 return ERR_PTR(-ENOMEM); 237 umem->ibdev = context->device; 238 umem->length = size; 239 umem->address = addr; 240 umem->writable = ib_access_writable(access); 241 umem->owning_mm = mm = current->mm; 242 mmgrab(mm); 243 244 page_list = (struct page **) __get_free_page(GFP_KERNEL); 245 if (!page_list) { 246 ret = -ENOMEM; 247 goto umem_kfree; 248 } 249 250 npages = ib_umem_num_pages(umem); 251 if (npages == 0 || npages > UINT_MAX) { 252 ret = -EINVAL; 253 goto out; 254 } 255 256 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 257 258 new_pinned = atomic64_add_return(npages, &mm->pinned_vm); 259 if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) { 260 atomic64_sub(npages, &mm->pinned_vm); 261 ret = -ENOMEM; 262 goto out; 263 } 264 265 cur_base = addr & PAGE_MASK; 266 267 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL); 268 if (ret) 269 goto vma; 270 271 if (!umem->writable) 272 gup_flags |= FOLL_FORCE; 273 274 sg = umem->sg_head.sgl; 275 276 while (npages) { 277 down_read(&mm->mmap_sem); 278 ret = get_user_pages(cur_base, 279 min_t(unsigned long, npages, 280 PAGE_SIZE / sizeof (struct page *)), 281 gup_flags | FOLL_LONGTERM, 282 page_list, NULL); 283 if (ret < 0) { 284 up_read(&mm->mmap_sem); 285 goto umem_release; 286 } 287 288 cur_base += ret * PAGE_SIZE; 289 npages -= ret; 290 291 sg = ib_umem_add_sg_table(sg, page_list, ret, 292 dma_get_max_seg_size(context->device->dma_device), 293 &umem->sg_nents); 294 295 up_read(&mm->mmap_sem); 296 } 297 298 sg_mark_end(sg); 299 300 umem->nmap = ib_dma_map_sg_attrs(context->device, 301 umem->sg_head.sgl, 302 umem->sg_nents, 303 DMA_BIDIRECTIONAL, 304 dma_attrs); 305 306 if (!umem->nmap) { 307 ret = -ENOMEM; 308 goto umem_release; 309 } 310 311 ret = 0; 312 goto out; 313 314 umem_release: 315 __ib_umem_release(context->device, umem, 0); 316 vma: 317 atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm); 318 out: 319 free_page((unsigned long) page_list); 320 umem_kfree: 321 if (ret) { 322 mmdrop(umem->owning_mm); 323 kfree(umem); 324 } 325 return ret ? ERR_PTR(ret) : umem; 326 } 327 EXPORT_SYMBOL(ib_umem_get); 328 329 /** 330 * ib_umem_release - release memory pinned with ib_umem_get 331 * @umem: umem struct to release 332 */ 333 void ib_umem_release(struct ib_umem *umem) 334 { 335 if (!umem) 336 return; 337 if (umem->is_odp) 338 return ib_umem_odp_release(to_ib_umem_odp(umem)); 339 340 __ib_umem_release(umem->ibdev, umem, 1); 341 342 atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm); 343 mmdrop(umem->owning_mm); 344 kfree(umem); 345 } 346 EXPORT_SYMBOL(ib_umem_release); 347 348 int ib_umem_page_count(struct ib_umem *umem) 349 { 350 int i, n = 0; 351 struct scatterlist *sg; 352 353 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) 354 n += sg_dma_len(sg) >> PAGE_SHIFT; 355 356 return n; 357 } 358 EXPORT_SYMBOL(ib_umem_page_count); 359 360 /* 361 * Copy from the given ib_umem's pages to the given buffer. 362 * 363 * umem - the umem to copy from 364 * offset - offset to start copying from 365 * dst - destination buffer 366 * length - buffer length 367 * 368 * Returns 0 on success, or an error code. 369 */ 370 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset, 371 size_t length) 372 { 373 size_t end = offset + length; 374 int ret; 375 376 if (offset > umem->length || length > umem->length - offset) { 377 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n", 378 offset, umem->length, end); 379 return -EINVAL; 380 } 381 382 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length, 383 offset + ib_umem_offset(umem)); 384 385 if (ret < 0) 386 return ret; 387 else if (ret != length) 388 return -EINVAL; 389 else 390 return 0; 391 } 392 EXPORT_SYMBOL(ib_umem_copy_from); 393