umem.c (9062B)
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 48static 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 */ 79unsigned 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 umem->iova = va = virt; 104 /* The best result is the smallest page size that results in the minimum 105 * number of required pages. Compute the largest page size that could 106 * work based on VA address bits that don't change. 107 */ 108 mask = pgsz_bitmap & 109 GENMASK(BITS_PER_LONG - 1, 110 bits_per((umem->length - 1 + virt) ^ virt)); 111 /* offset into first SGL */ 112 pgoff = umem->address & ~PAGE_MASK; 113 114 for_each_sgtable_dma_sg(&umem->sgt_append.sgt, sg, i) { 115 /* Walk SGL and reduce max page size if VA/PA bits differ 116 * for any address. 117 */ 118 mask |= (sg_dma_address(sg) + pgoff) ^ va; 119 va += sg_dma_len(sg) - pgoff; 120 /* Except for the last entry, the ending iova alignment sets 121 * the maximum possible page size as the low bits of the iova 122 * must be zero when starting the next chunk. 123 */ 124 if (i != (umem->sgt_append.sgt.nents - 1)) 125 mask |= va; 126 pgoff = 0; 127 } 128 129 /* The mask accumulates 1's in each position where the VA and physical 130 * address differ, thus the length of trailing 0 is the largest page 131 * size that can pass the VA through to the physical. 132 */ 133 if (mask) 134 pgsz_bitmap &= GENMASK(count_trailing_zeros(mask), 0); 135 return pgsz_bitmap ? rounddown_pow_of_two(pgsz_bitmap) : 0; 136} 137EXPORT_SYMBOL(ib_umem_find_best_pgsz); 138 139/** 140 * ib_umem_get - Pin and DMA map userspace memory. 141 * 142 * @device: IB device to connect UMEM 143 * @addr: userspace virtual address to start at 144 * @size: length of region to pin 145 * @access: IB_ACCESS_xxx flags for memory being pinned 146 */ 147struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr, 148 size_t size, int access) 149{ 150 struct ib_umem *umem; 151 struct page **page_list; 152 unsigned long lock_limit; 153 unsigned long new_pinned; 154 unsigned long cur_base; 155 unsigned long dma_attr = 0; 156 struct mm_struct *mm; 157 unsigned long npages; 158 int pinned, ret; 159 unsigned int gup_flags = FOLL_WRITE; 160 161 /* 162 * If the combination of the addr and size requested for this memory 163 * region causes an integer overflow, return error. 164 */ 165 if (((addr + size) < addr) || 166 PAGE_ALIGN(addr + size) < (addr + size)) 167 return ERR_PTR(-EINVAL); 168 169 if (!can_do_mlock()) 170 return ERR_PTR(-EPERM); 171 172 if (access & IB_ACCESS_ON_DEMAND) 173 return ERR_PTR(-EOPNOTSUPP); 174 175 umem = kzalloc(sizeof(*umem), GFP_KERNEL); 176 if (!umem) 177 return ERR_PTR(-ENOMEM); 178 umem->ibdev = device; 179 umem->length = size; 180 umem->address = addr; 181 /* 182 * Drivers should call ib_umem_find_best_pgsz() to set the iova 183 * correctly. 184 */ 185 umem->iova = addr; 186 umem->writable = ib_access_writable(access); 187 umem->owning_mm = mm = current->mm; 188 mmgrab(mm); 189 190 page_list = (struct page **) __get_free_page(GFP_KERNEL); 191 if (!page_list) { 192 ret = -ENOMEM; 193 goto umem_kfree; 194 } 195 196 npages = ib_umem_num_pages(umem); 197 if (npages == 0 || npages > UINT_MAX) { 198 ret = -EINVAL; 199 goto out; 200 } 201 202 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 203 204 new_pinned = atomic64_add_return(npages, &mm->pinned_vm); 205 if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) { 206 atomic64_sub(npages, &mm->pinned_vm); 207 ret = -ENOMEM; 208 goto out; 209 } 210 211 cur_base = addr & PAGE_MASK; 212 213 if (!umem->writable) 214 gup_flags |= FOLL_FORCE; 215 216 while (npages) { 217 cond_resched(); 218 pinned = pin_user_pages_fast(cur_base, 219 min_t(unsigned long, npages, 220 PAGE_SIZE / 221 sizeof(struct page *)), 222 gup_flags | FOLL_LONGTERM, page_list); 223 if (pinned < 0) { 224 ret = pinned; 225 goto umem_release; 226 } 227 228 cur_base += pinned * PAGE_SIZE; 229 npages -= pinned; 230 ret = sg_alloc_append_table_from_pages( 231 &umem->sgt_append, page_list, pinned, 0, 232 pinned << PAGE_SHIFT, ib_dma_max_seg_size(device), 233 npages, GFP_KERNEL); 234 if (ret) { 235 unpin_user_pages_dirty_lock(page_list, pinned, 0); 236 goto umem_release; 237 } 238 } 239 240 if (access & IB_ACCESS_RELAXED_ORDERING) 241 dma_attr |= DMA_ATTR_WEAK_ORDERING; 242 243 ret = ib_dma_map_sgtable_attrs(device, &umem->sgt_append.sgt, 244 DMA_BIDIRECTIONAL, dma_attr); 245 if (ret) 246 goto umem_release; 247 goto out; 248 249umem_release: 250 __ib_umem_release(device, umem, 0); 251 atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm); 252out: 253 free_page((unsigned long) page_list); 254umem_kfree: 255 if (ret) { 256 mmdrop(umem->owning_mm); 257 kfree(umem); 258 } 259 return ret ? ERR_PTR(ret) : umem; 260} 261EXPORT_SYMBOL(ib_umem_get); 262 263/** 264 * ib_umem_release - release memory pinned with ib_umem_get 265 * @umem: umem struct to release 266 */ 267void ib_umem_release(struct ib_umem *umem) 268{ 269 if (!umem) 270 return; 271 if (umem->is_dmabuf) 272 return ib_umem_dmabuf_release(to_ib_umem_dmabuf(umem)); 273 if (umem->is_odp) 274 return ib_umem_odp_release(to_ib_umem_odp(umem)); 275 276 __ib_umem_release(umem->ibdev, umem, 1); 277 278 atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm); 279 mmdrop(umem->owning_mm); 280 kfree(umem); 281} 282EXPORT_SYMBOL(ib_umem_release); 283 284/* 285 * Copy from the given ib_umem's pages to the given buffer. 286 * 287 * umem - the umem to copy from 288 * offset - offset to start copying from 289 * dst - destination buffer 290 * length - buffer length 291 * 292 * Returns 0 on success, or an error code. 293 */ 294int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset, 295 size_t length) 296{ 297 size_t end = offset + length; 298 int ret; 299 300 if (offset > umem->length || length > umem->length - offset) { 301 pr_err("%s not in range. offset: %zd umem length: %zd end: %zd\n", 302 __func__, offset, umem->length, end); 303 return -EINVAL; 304 } 305 306 ret = sg_pcopy_to_buffer(umem->sgt_append.sgt.sgl, 307 umem->sgt_append.sgt.orig_nents, dst, length, 308 offset + ib_umem_offset(umem)); 309 310 if (ret < 0) 311 return ret; 312 else if (ret != length) 313 return -EINVAL; 314 else 315 return 0; 316} 317EXPORT_SYMBOL(ib_umem_copy_from);