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 bool make_dirty = umem->writable && dirty; 51 struct scatterlist *sg; 52 unsigned int i; 53 54 if (dirty) 55 ib_dma_unmap_sgtable_attrs(dev, &umem->sgt_append.sgt, 56 DMA_BIDIRECTIONAL, 0); 57 58 for_each_sgtable_sg(&umem->sgt_append.sgt, sg, i) 59 unpin_user_page_range_dirty_lock(sg_page(sg), 60 DIV_ROUND_UP(sg->length, PAGE_SIZE), make_dirty); 61 62 sg_free_append_table(&umem->sgt_append); 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 unsigned long curr_len = 0; 84 dma_addr_t curr_base = ~0; 85 unsigned long va, pgoff; 86 struct scatterlist *sg; 87 dma_addr_t mask; 88 dma_addr_t end; 89 int i; 90 91 umem->iova = va = virt; 92 93 if (umem->is_odp) { 94 unsigned int page_size = BIT(to_ib_umem_odp(umem)->page_shift); 95 96 /* ODP must always be self consistent. */ 97 if (!(pgsz_bitmap & page_size)) 98 return 0; 99 return page_size; 100 } 101 102 /* The best result is the smallest page size that results in the minimum 103 * number of required pages. Compute the largest page size that could 104 * work based on VA address bits that don't change. 105 */ 106 mask = pgsz_bitmap & 107 GENMASK(BITS_PER_LONG - 1, 108 bits_per((umem->length - 1 + virt) ^ virt)); 109 /* offset into first SGL */ 110 pgoff = umem->address & ~PAGE_MASK; 111 112 for_each_sgtable_dma_sg(&umem->sgt_append.sgt, sg, i) { 113 /* If the current entry is physically contiguous with the previous 114 * one, no need to take its start addresses into consideration. 115 */ 116 if (check_add_overflow(curr_base, curr_len, &end) || 117 end != sg_dma_address(sg)) { 118 119 curr_base = sg_dma_address(sg); 120 curr_len = 0; 121 122 /* Reduce max page size if VA/PA bits differ */ 123 mask |= (curr_base + pgoff) ^ va; 124 125 /* The alignment of any VA matching a discontinuity point 126 * in the physical memory sets the maximum possible page 127 * size as this must be a starting point of a new page that 128 * needs to be aligned. 129 */ 130 if (i != 0) 131 mask |= va; 132 } 133 134 curr_len += sg_dma_len(sg); 135 va += sg_dma_len(sg) - pgoff; 136 137 pgoff = 0; 138 } 139 140 /* The mask accumulates 1's in each position where the VA and physical 141 * address differ, thus the length of trailing 0 is the largest page 142 * size that can pass the VA through to the physical. 143 */ 144 if (mask) 145 pgsz_bitmap &= GENMASK(count_trailing_zeros(mask), 0); 146 return pgsz_bitmap ? rounddown_pow_of_two(pgsz_bitmap) : 0; 147 } 148 EXPORT_SYMBOL(ib_umem_find_best_pgsz); 149 150 /** 151 * ib_umem_get - Pin and DMA map userspace memory. 152 * 153 * @device: IB device to connect UMEM 154 * @addr: userspace virtual address to start at 155 * @size: length of region to pin 156 * @access: IB_ACCESS_xxx flags for memory being pinned 157 */ 158 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr, 159 size_t size, int access) 160 { 161 struct ib_umem *umem; 162 struct page **page_list; 163 unsigned long lock_limit; 164 unsigned long new_pinned; 165 unsigned long cur_base; 166 unsigned long dma_attr = 0; 167 struct mm_struct *mm; 168 unsigned long npages; 169 int pinned, ret; 170 unsigned int gup_flags = FOLL_LONGTERM; 171 172 /* 173 * If the combination of the addr and size requested for this memory 174 * region causes an integer overflow, return error. 175 */ 176 if (((addr + size) < addr) || 177 PAGE_ALIGN(addr + size) < (addr + size)) 178 return ERR_PTR(-EINVAL); 179 180 if (!can_do_mlock()) 181 return ERR_PTR(-EPERM); 182 183 if (access & IB_ACCESS_ON_DEMAND) 184 return ERR_PTR(-EOPNOTSUPP); 185 186 umem = kzalloc(sizeof(*umem), GFP_KERNEL); 187 if (!umem) 188 return ERR_PTR(-ENOMEM); 189 umem->ibdev = device; 190 umem->length = size; 191 umem->address = addr; 192 /* 193 * Drivers should call ib_umem_find_best_pgsz() to set the iova 194 * correctly. 195 */ 196 umem->iova = addr; 197 umem->writable = ib_access_writable(access); 198 umem->owning_mm = mm = current->mm; 199 mmgrab(mm); 200 201 page_list = (struct page **) __get_free_page(GFP_KERNEL); 202 if (!page_list) { 203 ret = -ENOMEM; 204 goto umem_kfree; 205 } 206 207 npages = ib_umem_num_pages(umem); 208 if (npages == 0 || npages > UINT_MAX) { 209 ret = -EINVAL; 210 goto out; 211 } 212 213 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 214 215 new_pinned = atomic64_add_return(npages, &mm->pinned_vm); 216 if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) { 217 atomic64_sub(npages, &mm->pinned_vm); 218 ret = -ENOMEM; 219 goto out; 220 } 221 222 cur_base = addr & PAGE_MASK; 223 224 if (umem->writable) 225 gup_flags |= FOLL_WRITE; 226 227 while (npages) { 228 cond_resched(); 229 pinned = pin_user_pages_fast(cur_base, 230 min_t(unsigned long, npages, 231 PAGE_SIZE / 232 sizeof(struct page *)), 233 gup_flags, page_list); 234 if (pinned < 0) { 235 ret = pinned; 236 goto umem_release; 237 } 238 239 cur_base += pinned * PAGE_SIZE; 240 npages -= pinned; 241 ret = sg_alloc_append_table_from_pages( 242 &umem->sgt_append, page_list, pinned, 0, 243 pinned << PAGE_SHIFT, ib_dma_max_seg_size(device), 244 npages, GFP_KERNEL); 245 if (ret) { 246 unpin_user_pages_dirty_lock(page_list, pinned, 0); 247 goto umem_release; 248 } 249 } 250 251 if (access & IB_ACCESS_RELAXED_ORDERING) 252 dma_attr |= DMA_ATTR_WEAK_ORDERING; 253 254 ret = ib_dma_map_sgtable_attrs(device, &umem->sgt_append.sgt, 255 DMA_BIDIRECTIONAL, dma_attr); 256 if (ret) 257 goto umem_release; 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("%s not in range. offset: %zd umem length: %zd end: %zd\n", 313 __func__, offset, umem->length, end); 314 return -EINVAL; 315 } 316 317 ret = sg_pcopy_to_buffer(umem->sgt_append.sgt.sgl, 318 umem->sgt_append.sgt.orig_nents, dst, length, 319 offset + ib_umem_offset(umem)); 320 321 if (ret < 0) 322 return ret; 323 else if (ret != length) 324 return -EINVAL; 325 else 326 return 0; 327 } 328 EXPORT_SYMBOL(ib_umem_copy_from); 329