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/hugetlb.h> 41 #include <linux/slab.h> 42 #include <rdma/ib_umem_odp.h> 43 44 #include "uverbs.h" 45 46 47 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty) 48 { 49 struct scatterlist *sg; 50 struct page *page; 51 int i; 52 53 if (umem->nmap > 0) 54 ib_dma_unmap_sg(dev, umem->sg_head.sgl, 55 umem->npages, 56 DMA_BIDIRECTIONAL); 57 58 for_each_sg(umem->sg_head.sgl, sg, umem->npages, i) { 59 60 page = sg_page(sg); 61 if (!PageDirty(page) && umem->writable && dirty) 62 set_page_dirty_lock(page); 63 put_page(page); 64 } 65 66 sg_free_table(&umem->sg_head); 67 } 68 69 /** 70 * ib_umem_get - Pin and DMA map userspace memory. 71 * 72 * If access flags indicate ODP memory, avoid pinning. Instead, stores 73 * the mm for future page fault handling in conjunction with MMU notifiers. 74 * 75 * @udata: userspace context to pin memory for 76 * @addr: userspace virtual address to start at 77 * @size: length of region to pin 78 * @access: IB_ACCESS_xxx flags for memory being pinned 79 * @dmasync: flush in-flight DMA when the memory region is written 80 */ 81 struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr, 82 size_t size, int access, int dmasync) 83 { 84 struct ib_ucontext *context; 85 struct ib_umem *umem; 86 struct page **page_list; 87 struct vm_area_struct **vma_list; 88 unsigned long lock_limit; 89 unsigned long new_pinned; 90 unsigned long cur_base; 91 struct mm_struct *mm; 92 unsigned long npages; 93 int ret; 94 int i; 95 unsigned long dma_attrs = 0; 96 struct scatterlist *sg, *sg_list_start; 97 unsigned int gup_flags = FOLL_WRITE; 98 99 if (!udata) 100 return ERR_PTR(-EIO); 101 102 context = container_of(udata, struct uverbs_attr_bundle, driver_udata) 103 ->context; 104 if (!context) 105 return ERR_PTR(-EIO); 106 107 if (dmasync) 108 dma_attrs |= DMA_ATTR_WRITE_BARRIER; 109 110 /* 111 * If the combination of the addr and size requested for this memory 112 * region causes an integer overflow, return error. 113 */ 114 if (((addr + size) < addr) || 115 PAGE_ALIGN(addr + size) < (addr + size)) 116 return ERR_PTR(-EINVAL); 117 118 if (!can_do_mlock()) 119 return ERR_PTR(-EPERM); 120 121 if (access & IB_ACCESS_ON_DEMAND) { 122 umem = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL); 123 if (!umem) 124 return ERR_PTR(-ENOMEM); 125 umem->is_odp = 1; 126 } else { 127 umem = kzalloc(sizeof(*umem), GFP_KERNEL); 128 if (!umem) 129 return ERR_PTR(-ENOMEM); 130 } 131 132 umem->context = context; 133 umem->length = size; 134 umem->address = addr; 135 umem->page_shift = PAGE_SHIFT; 136 umem->writable = ib_access_writable(access); 137 umem->owning_mm = mm = current->mm; 138 mmgrab(mm); 139 140 if (access & IB_ACCESS_ON_DEMAND) { 141 ret = ib_umem_odp_get(to_ib_umem_odp(umem), access); 142 if (ret) 143 goto umem_kfree; 144 return umem; 145 } 146 147 /* We assume the memory is from hugetlb until proved otherwise */ 148 umem->hugetlb = 1; 149 150 page_list = (struct page **) __get_free_page(GFP_KERNEL); 151 if (!page_list) { 152 ret = -ENOMEM; 153 goto umem_kfree; 154 } 155 156 /* 157 * if we can't alloc the vma_list, it's not so bad; 158 * just assume the memory is not hugetlb memory 159 */ 160 vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL); 161 if (!vma_list) 162 umem->hugetlb = 0; 163 164 npages = ib_umem_num_pages(umem); 165 if (npages == 0 || npages > UINT_MAX) { 166 ret = -EINVAL; 167 goto out; 168 } 169 170 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 171 172 new_pinned = atomic64_add_return(npages, &mm->pinned_vm); 173 if (new_pinned > lock_limit && !capable(CAP_IPC_LOCK)) { 174 atomic64_sub(npages, &mm->pinned_vm); 175 ret = -ENOMEM; 176 goto out; 177 } 178 179 cur_base = addr & PAGE_MASK; 180 181 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL); 182 if (ret) 183 goto vma; 184 185 if (!umem->writable) 186 gup_flags |= FOLL_FORCE; 187 188 sg_list_start = umem->sg_head.sgl; 189 190 while (npages) { 191 down_read(&mm->mmap_sem); 192 ret = get_user_pages_longterm(cur_base, 193 min_t(unsigned long, npages, 194 PAGE_SIZE / sizeof (struct page *)), 195 gup_flags, page_list, vma_list); 196 if (ret < 0) { 197 up_read(&mm->mmap_sem); 198 goto umem_release; 199 } 200 201 umem->npages += ret; 202 cur_base += ret * PAGE_SIZE; 203 npages -= ret; 204 205 /* Continue to hold the mmap_sem as vma_list access 206 * needs to be protected. 207 */ 208 for_each_sg(sg_list_start, sg, ret, i) { 209 if (vma_list && !is_vm_hugetlb_page(vma_list[i])) 210 umem->hugetlb = 0; 211 212 sg_set_page(sg, page_list[i], PAGE_SIZE, 0); 213 } 214 up_read(&mm->mmap_sem); 215 216 /* preparing for next loop */ 217 sg_list_start = sg; 218 } 219 220 umem->nmap = ib_dma_map_sg_attrs(context->device, 221 umem->sg_head.sgl, 222 umem->npages, 223 DMA_BIDIRECTIONAL, 224 dma_attrs); 225 226 if (!umem->nmap) { 227 ret = -ENOMEM; 228 goto umem_release; 229 } 230 231 ret = 0; 232 goto out; 233 234 umem_release: 235 __ib_umem_release(context->device, umem, 0); 236 vma: 237 atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm); 238 out: 239 if (vma_list) 240 free_page((unsigned long) vma_list); 241 free_page((unsigned long) page_list); 242 umem_kfree: 243 if (ret) { 244 mmdrop(umem->owning_mm); 245 kfree(umem); 246 } 247 return ret ? ERR_PTR(ret) : umem; 248 } 249 EXPORT_SYMBOL(ib_umem_get); 250 251 static void __ib_umem_release_tail(struct ib_umem *umem) 252 { 253 mmdrop(umem->owning_mm); 254 if (umem->is_odp) 255 kfree(to_ib_umem_odp(umem)); 256 else 257 kfree(umem); 258 } 259 260 /** 261 * ib_umem_release - release memory pinned with ib_umem_get 262 * @umem: umem struct to release 263 */ 264 void ib_umem_release(struct ib_umem *umem) 265 { 266 if (umem->is_odp) { 267 ib_umem_odp_release(to_ib_umem_odp(umem)); 268 __ib_umem_release_tail(umem); 269 return; 270 } 271 272 __ib_umem_release(umem->context->device, umem, 1); 273 274 atomic64_sub(ib_umem_num_pages(umem), &umem->owning_mm->pinned_vm); 275 __ib_umem_release_tail(umem); 276 } 277 EXPORT_SYMBOL(ib_umem_release); 278 279 int ib_umem_page_count(struct ib_umem *umem) 280 { 281 int i; 282 int n; 283 struct scatterlist *sg; 284 285 if (umem->is_odp) 286 return ib_umem_num_pages(umem); 287 288 n = 0; 289 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) 290 n += sg_dma_len(sg) >> umem->page_shift; 291 292 return n; 293 } 294 EXPORT_SYMBOL(ib_umem_page_count); 295 296 /* 297 * Copy from the given ib_umem's pages to the given buffer. 298 * 299 * umem - the umem to copy from 300 * offset - offset to start copying from 301 * dst - destination buffer 302 * length - buffer length 303 * 304 * Returns 0 on success, or an error code. 305 */ 306 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset, 307 size_t length) 308 { 309 size_t end = offset + length; 310 int ret; 311 312 if (offset > umem->length || length > umem->length - offset) { 313 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n", 314 offset, umem->length, end); 315 return -EINVAL; 316 } 317 318 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->npages, 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