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 #define RESCHED_LOOP_CNT_THRESHOLD 0x1000 49 50 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty) 51 { 52 bool make_dirty = umem->writable && dirty; 53 struct scatterlist *sg; 54 unsigned int i; 55 56 if (dirty) 57 ib_dma_unmap_sgtable_attrs(dev, &umem->sgt_append.sgt, 58 DMA_BIDIRECTIONAL, 59 DMA_ATTR_REQUIRE_COHERENT); 60 61 for_each_sgtable_sg(&umem->sgt_append.sgt, sg, i) { 62 unpin_user_page_range_dirty_lock(sg_page(sg), 63 DIV_ROUND_UP(sg->length, PAGE_SIZE), make_dirty); 64 65 if (i && !(i % RESCHED_LOOP_CNT_THRESHOLD)) 66 cond_resched(); 67 } 68 69 sg_free_append_table(&umem->sgt_append); 70 } 71 72 /** 73 * ib_umem_find_best_pgsz - Find best HW page size to use for this MR 74 * 75 * @umem: umem struct 76 * @pgsz_bitmap: bitmap of HW supported page sizes 77 * @virt: IOVA 78 * 79 * This helper is intended for HW that support multiple page 80 * sizes but can do only a single page size in an MR. 81 * 82 * Returns 0 if the umem requires page sizes not supported by 83 * the driver to be mapped. Drivers always supporting PAGE_SIZE 84 * or smaller will never see a 0 result. 85 */ 86 unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem, 87 unsigned long pgsz_bitmap, 88 unsigned long virt) 89 { 90 unsigned long curr_len = 0; 91 dma_addr_t curr_base = ~0; 92 unsigned long va, pgoff; 93 struct scatterlist *sg; 94 dma_addr_t mask; 95 dma_addr_t end; 96 int i; 97 98 umem->iova = va = virt; 99 100 if (umem->is_odp) { 101 unsigned int page_size = BIT(to_ib_umem_odp(umem)->page_shift); 102 103 /* ODP must always be self consistent. */ 104 if (!(pgsz_bitmap & page_size)) 105 return 0; 106 return page_size; 107 } 108 109 /* The best result is the smallest page size that results in the minimum 110 * number of required pages. Compute the largest page size that could 111 * work based on VA address bits that don't change. 112 */ 113 mask = pgsz_bitmap & 114 GENMASK(BITS_PER_LONG - 1, 115 bits_per((umem->length - 1 + virt) ^ virt)); 116 /* offset into first SGL */ 117 pgoff = umem->address & ~PAGE_MASK; 118 119 for_each_sgtable_dma_sg(&umem->sgt_append.sgt, sg, i) { 120 /* If the current entry is physically contiguous with the previous 121 * one, no need to take its start addresses into consideration. 122 */ 123 if (check_add_overflow(curr_base, curr_len, &end) || 124 end != sg_dma_address(sg)) { 125 126 curr_base = sg_dma_address(sg); 127 curr_len = 0; 128 129 /* Reduce max page size if VA/PA bits differ */ 130 mask |= (curr_base + pgoff) ^ va; 131 132 /* The alignment of any VA matching a discontinuity point 133 * in the physical memory sets the maximum possible page 134 * size as this must be a starting point of a new page that 135 * needs to be aligned. 136 */ 137 if (i != 0) 138 mask |= va; 139 } 140 141 curr_len += sg_dma_len(sg); 142 va += sg_dma_len(sg) - pgoff; 143 144 pgoff = 0; 145 } 146 147 /* The mask accumulates 1's in each position where the VA and physical 148 * address differ, thus the length of trailing 0 is the largest page 149 * size that can pass the VA through to the physical. 150 */ 151 if (mask) 152 pgsz_bitmap &= GENMASK(count_trailing_zeros(mask), 0); 153 return pgsz_bitmap ? rounddown_pow_of_two(pgsz_bitmap) : 0; 154 } 155 EXPORT_SYMBOL(ib_umem_find_best_pgsz); 156 157 /** 158 * ib_umem_get - Pin and DMA map userspace memory. 159 * 160 * @device: IB device to connect UMEM 161 * @addr: userspace virtual address to start at 162 * @size: length of region to pin 163 * @access: IB_ACCESS_xxx flags for memory being pinned 164 */ 165 struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr, 166 size_t size, int access) 167 { 168 struct ib_umem *umem; 169 struct page **page_list; 170 unsigned long lock_limit; 171 unsigned long new_pinned; 172 unsigned long cur_base; 173 unsigned long dma_attr = DMA_ATTR_REQUIRE_COHERENT; 174 struct mm_struct *mm; 175 unsigned long npages; 176 int pinned, ret; 177 unsigned int gup_flags = FOLL_LONGTERM; 178 179 /* 180 * If the combination of the addr and size requested for this memory 181 * region causes an integer overflow, return error. 182 */ 183 if (((addr + size) < addr) || 184 PAGE_ALIGN(addr + size) < (addr + size)) 185 return ERR_PTR(-EINVAL); 186 187 if (!can_do_mlock()) 188 return ERR_PTR(-EPERM); 189 190 if (access & IB_ACCESS_ON_DEMAND) 191 return ERR_PTR(-EOPNOTSUPP); 192 193 umem = kzalloc_obj(*umem); 194 if (!umem) 195 return ERR_PTR(-ENOMEM); 196 umem->ibdev = device; 197 umem->length = size; 198 umem->address = addr; 199 /* 200 * Drivers should call ib_umem_find_best_pgsz() to set the iova 201 * correctly. 202 */ 203 umem->iova = addr; 204 umem->writable = ib_access_writable(access); 205 umem->owning_mm = mm = current->mm; 206 mmgrab(mm); 207 208 page_list = (struct page **) __get_free_page(GFP_KERNEL); 209 if (!page_list) { 210 ret = -ENOMEM; 211 goto umem_kfree; 212 } 213 214 npages = ib_umem_num_pages(umem); 215 if (npages == 0 || npages > UINT_MAX) { 216 ret = -EINVAL; 217 goto out; 218 } 219 220 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 221 222 new_pinned = atomic64_add_return(npages, &mm->pinned_vm); 223 if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) { 224 atomic64_sub(npages, &mm->pinned_vm); 225 ret = -ENOMEM; 226 goto out; 227 } 228 229 cur_base = addr & PAGE_MASK; 230 231 if (umem->writable) 232 gup_flags |= FOLL_WRITE; 233 234 while (npages) { 235 cond_resched(); 236 pinned = pin_user_pages_fast(cur_base, 237 min_t(unsigned long, npages, 238 PAGE_SIZE / 239 sizeof(struct page *)), 240 gup_flags, page_list); 241 if (pinned < 0) { 242 ret = pinned; 243 goto umem_release; 244 } 245 246 cur_base += pinned * PAGE_SIZE; 247 npages -= pinned; 248 ret = sg_alloc_append_table_from_pages( 249 &umem->sgt_append, page_list, pinned, 0, 250 pinned << PAGE_SHIFT, ib_dma_max_seg_size(device), 251 npages, GFP_KERNEL); 252 if (ret) { 253 unpin_user_pages_dirty_lock(page_list, pinned, 0); 254 goto umem_release; 255 } 256 } 257 258 if (access & IB_ACCESS_RELAXED_ORDERING) 259 dma_attr |= DMA_ATTR_WEAK_ORDERING; 260 261 ret = ib_dma_map_sgtable_attrs(device, &umem->sgt_append.sgt, 262 DMA_BIDIRECTIONAL, dma_attr); 263 if (ret) 264 goto umem_release; 265 goto out; 266 267 umem_release: 268 __ib_umem_release(device, umem, 0); 269 atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm); 270 out: 271 free_page((unsigned long) page_list); 272 umem_kfree: 273 if (ret) { 274 mmdrop(umem->owning_mm); 275 kfree(umem); 276 } 277 return ret ? ERR_PTR(ret) : umem; 278 } 279 EXPORT_SYMBOL(ib_umem_get); 280 281 /** 282 * ib_umem_release - release memory pinned with ib_umem_get 283 * @umem: umem struct to release 284 */ 285 void ib_umem_release(struct ib_umem *umem) 286 { 287 if (!umem) 288 return; 289 if (umem->is_dmabuf) 290 return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem)); 291 if (umem->is_odp) 292 return ib_umem_odp_release(to_ib_umem_odp(umem)); 293 294 __ib_umem_release(umem->ibdev, umem, 1); 295 296 atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm); 297 mmdrop(umem->owning_mm); 298 kfree(umem); 299 } 300 EXPORT_SYMBOL(ib_umem_release); 301 302 /* 303 * Copy from the given ib_umem's pages to the given buffer. 304 * 305 * umem - the umem to copy from 306 * offset - offset to start copying from 307 * dst - destination buffer 308 * length - buffer length 309 * 310 * Returns 0 on success, or an error code. 311 */ 312 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset, 313 size_t length) 314 { 315 size_t end = offset + length; 316 int ret; 317 318 if (offset > umem->length || length > umem->length - offset) { 319 pr_err("%s not in range. offset: %zd umem length: %zd end: %zd\n", 320 __func__, offset, umem->length, end); 321 return -EINVAL; 322 } 323 324 ret = sg_pcopy_to_buffer(umem->sgt_append.sgt.sgl, 325 umem->sgt_append.sgt.orig_nents, dst, length, 326 offset + ib_umem_offset(umem)); 327 328 if (ret < 0) 329 return ret; 330 else if (ret != length) 331 return -EINVAL; 332 else 333 return 0; 334 } 335 EXPORT_SYMBOL(ib_umem_copy_from); 336