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 * Copyright (c) 2020 Intel Corporation. All rights reserved. 6 * 7 * This software is available to you under a choice of one of two 8 * licenses. You may choose to be licensed under the terms of the GNU 9 * General Public License (GPL) Version 2, available from the file 10 * COPYING in the main directory of this source tree, or the 11 * OpenIB.org BSD license below: 12 * 13 * Redistribution and use in source and binary forms, with or 14 * without modification, are permitted provided that the following 15 * conditions are met: 16 * 17 * - Redistributions of source code must retain the above 18 * copyright notice, this list of conditions and the following 19 * disclaimer. 20 * 21 * - Redistributions in binary form must reproduce the above 22 * copyright notice, this list of conditions and the following 23 * disclaimer in the documentation and/or other materials 24 * provided with the distribution. 25 * 26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 33 * SOFTWARE. 34 */ 35 36 #include <linux/mm.h> 37 #include <linux/dma-mapping.h> 38 #include <linux/sched/signal.h> 39 #include <linux/sched/mm.h> 40 #include <linux/export.h> 41 #include <linux/slab.h> 42 #include <linux/pagemap.h> 43 #include <linux/count_zeros.h> 44 #include <rdma/ib_umem_odp.h> 45 46 #include "uverbs.h" 47 48 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty) 49 { 50 struct sg_page_iter sg_iter; 51 struct page *page; 52 53 if (umem->nmap > 0) 54 ib_dma_unmap_sg(dev, umem->sg_head.sgl, umem->sg_nents, 55 DMA_BIDIRECTIONAL); 56 57 for_each_sg_page(umem->sg_head.sgl, &sg_iter, umem->sg_nents, 0) { 58 page = sg_page_iter_page(&sg_iter); 59 unpin_user_pages_dirty_lock(&page, 1, umem->writable && dirty); 60 } 61 62 sg_free_table(&umem->sg_head); 63 } 64 65 /** 66 * ib_umem_find_best_pgsz - Find best HW page size to use for this MR 67 * 68 * @umem: umem struct 69 * @pgsz_bitmap: bitmap of HW supported page sizes 70 * @virt: IOVA 71 * 72 * This helper is intended for HW that support multiple page 73 * sizes but can do only a single page size in an MR. 74 * 75 * Returns 0 if the umem requires page sizes not supported by 76 * the driver to be mapped. Drivers always supporting PAGE_SIZE 77 * or smaller will never see a 0 result. 78 */ 79 unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem, 80 unsigned long pgsz_bitmap, 81 unsigned long virt) 82 { 83 struct scatterlist *sg; 84 unsigned long va, pgoff; 85 dma_addr_t mask; 86 int i; 87 88 if (umem->is_odp) { 89 unsigned int page_size = BIT(to_ib_umem_odp(umem)->page_shift); 90 91 /* ODP must always be self consistent. */ 92 if (!(pgsz_bitmap & page_size)) 93 return 0; 94 return page_size; 95 } 96 97 /* rdma_for_each_block() has a bug if the page size is smaller than the 98 * page size used to build the umem. For now prevent smaller page sizes 99 * from being returned. 100 */ 101 pgsz_bitmap &= GENMASK(BITS_PER_LONG - 1, PAGE_SHIFT); 102 103 /* At minimum, drivers must support PAGE_SIZE or smaller */ 104 if (WARN_ON(!(pgsz_bitmap & GENMASK(PAGE_SHIFT, 0)))) 105 return 0; 106 107 umem->iova = va = virt; 108 /* The best result is the smallest page size that results in the minimum 109 * number of required pages. Compute the largest page size that could 110 * work based on VA address bits that don't change. 111 */ 112 mask = pgsz_bitmap & 113 GENMASK(BITS_PER_LONG - 1, 114 bits_per((umem->length - 1 + virt) ^ virt)); 115 /* offset into first SGL */ 116 pgoff = umem->address & ~PAGE_MASK; 117 118 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) { 119 /* Walk SGL and reduce max page size if VA/PA bits differ 120 * for any address. 121 */ 122 mask |= (sg_dma_address(sg) + pgoff) ^ va; 123 va += sg_dma_len(sg) - pgoff; 124 /* Except for the last entry, the ending iova alignment sets 125 * the maximum possible page size as the low bits of the iova 126 * must be zero when starting the next chunk. 127 */ 128 if (i != (umem->nmap - 1)) 129 mask |= va; 130 pgoff = 0; 131 } 132 133 /* The mask accumulates 1's in each position where the VA and physical 134 * address differ, thus the length of trailing 0 is the largest page 135 * size that can pass the VA through to the physical. 136 */ 137 if (mask) 138 pgsz_bitmap &= GENMASK(count_trailing_zeros(mask), 0); 139 return pgsz_bitmap ? rounddown_pow_of_two(pgsz_bitmap) : 0; 140 } 141 EXPORT_SYMBOL(ib_umem_find_best_pgsz); 142 143 /** 144 * ib_umem_get - Pin and DMA map userspace memory. 145 * 146 * @device: IB device to connect UMEM 147 * @addr: userspace virtual address to start at 148 * @size: length of region to pin 149 * @access: IB_ACCESS_xxx flags for memory being pinned 150 */ 151 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr, 152 size_t size, int access) 153 { 154 struct ib_umem *umem; 155 struct page **page_list; 156 unsigned long lock_limit; 157 unsigned long new_pinned; 158 unsigned long cur_base; 159 unsigned long dma_attr = 0; 160 struct mm_struct *mm; 161 unsigned long npages; 162 int ret; 163 struct scatterlist *sg = NULL; 164 unsigned int gup_flags = FOLL_WRITE; 165 166 /* 167 * If the combination of the addr and size requested for this memory 168 * region causes an integer overflow, return error. 169 */ 170 if (((addr + size) < addr) || 171 PAGE_ALIGN(addr + size) < (addr + size)) 172 return ERR_PTR(-EINVAL); 173 174 if (!can_do_mlock()) 175 return ERR_PTR(-EPERM); 176 177 if (access & IB_ACCESS_ON_DEMAND) 178 return ERR_PTR(-EOPNOTSUPP); 179 180 umem = kzalloc(sizeof(*umem), GFP_KERNEL); 181 if (!umem) 182 return ERR_PTR(-ENOMEM); 183 umem->ibdev = device; 184 umem->length = size; 185 umem->address = addr; 186 /* 187 * Drivers should call ib_umem_find_best_pgsz() to set the iova 188 * correctly. 189 */ 190 umem->iova = addr; 191 umem->writable = ib_access_writable(access); 192 umem->owning_mm = mm = current->mm; 193 mmgrab(mm); 194 195 page_list = (struct page **) __get_free_page(GFP_KERNEL); 196 if (!page_list) { 197 ret = -ENOMEM; 198 goto umem_kfree; 199 } 200 201 npages = ib_umem_num_pages(umem); 202 if (npages == 0 || npages > UINT_MAX) { 203 ret = -EINVAL; 204 goto out; 205 } 206 207 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 208 209 new_pinned = atomic64_add_return(npages, &mm->pinned_vm); 210 if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) { 211 atomic64_sub(npages, &mm->pinned_vm); 212 ret = -ENOMEM; 213 goto out; 214 } 215 216 cur_base = addr & PAGE_MASK; 217 218 if (!umem->writable) 219 gup_flags |= FOLL_FORCE; 220 221 while (npages) { 222 cond_resched(); 223 ret = pin_user_pages_fast(cur_base, 224 min_t(unsigned long, npages, 225 PAGE_SIZE / 226 sizeof(struct page *)), 227 gup_flags | FOLL_LONGTERM, page_list); 228 if (ret < 0) 229 goto umem_release; 230 231 cur_base += ret * PAGE_SIZE; 232 npages -= ret; 233 sg = __sg_alloc_table_from_pages(&umem->sg_head, page_list, ret, 234 0, ret << PAGE_SHIFT, 235 ib_dma_max_seg_size(device), sg, npages, 236 GFP_KERNEL); 237 umem->sg_nents = umem->sg_head.nents; 238 if (IS_ERR(sg)) { 239 unpin_user_pages_dirty_lock(page_list, ret, 0); 240 ret = PTR_ERR(sg); 241 goto umem_release; 242 } 243 } 244 245 if (access & IB_ACCESS_RELAXED_ORDERING) 246 dma_attr |= DMA_ATTR_WEAK_ORDERING; 247 248 umem->nmap = 249 ib_dma_map_sg_attrs(device, umem->sg_head.sgl, umem->sg_nents, 250 DMA_BIDIRECTIONAL, dma_attr); 251 252 if (!umem->nmap) { 253 ret = -ENOMEM; 254 goto umem_release; 255 } 256 257 ret = 0; 258 goto out; 259 260 umem_release: 261 __ib_umem_release(device, umem, 0); 262 atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm); 263 out: 264 free_page((unsigned long) page_list); 265 umem_kfree: 266 if (ret) { 267 mmdrop(umem->owning_mm); 268 kfree(umem); 269 } 270 return ret ? ERR_PTR(ret) : umem; 271 } 272 EXPORT_SYMBOL(ib_umem_get); 273 274 /** 275 * ib_umem_release - release memory pinned with ib_umem_get 276 * @umem: umem struct to release 277 */ 278 void ib_umem_release(struct ib_umem *umem) 279 { 280 if (!umem) 281 return; 282 if (umem->is_dmabuf) 283 return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem)); 284 if (umem->is_odp) 285 return ib_umem_odp_release(to_ib_umem_odp(umem)); 286 287 __ib_umem_release(umem->ibdev, umem, 1); 288 289 atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm); 290 mmdrop(umem->owning_mm); 291 kfree(umem); 292 } 293 EXPORT_SYMBOL(ib_umem_release); 294 295 /* 296 * Copy from the given ib_umem's pages to the given buffer. 297 * 298 * umem - the umem to copy from 299 * offset - offset to start copying from 300 * dst - destination buffer 301 * length - buffer length 302 * 303 * Returns 0 on success, or an error code. 304 */ 305 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset, 306 size_t length) 307 { 308 size_t end = offset + length; 309 int ret; 310 311 if (offset > umem->length || length > umem->length - offset) { 312 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n", 313 offset, umem->length, end); 314 return -EINVAL; 315 } 316 317 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->sg_nents, dst, length, 318 offset + ib_umem_offset(umem)); 319 320 if (ret < 0) 321 return ret; 322 else if (ret != length) 323 return -EINVAL; 324 else 325 return 0; 326 } 327 EXPORT_SYMBOL(ib_umem_copy_from); 328