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.h> 38 #include <linux/export.h> 39 #include <linux/hugetlb.h> 40 #include <linux/slab.h> 41 #include <rdma/ib_umem_odp.h> 42 43 #include "uverbs.h" 44 45 46 static void __ib_umem_release(struct ib_device *dev, struct ib_umem *umem, int dirty) 47 { 48 struct scatterlist *sg; 49 struct page *page; 50 int i; 51 52 if (umem->nmap > 0) 53 ib_dma_unmap_sg(dev, umem->sg_head.sgl, 54 umem->npages, 55 DMA_BIDIRECTIONAL); 56 57 for_each_sg(umem->sg_head.sgl, sg, umem->npages, i) { 58 59 page = sg_page(sg); 60 if (umem->writable && dirty) 61 set_page_dirty_lock(page); 62 put_page(page); 63 } 64 65 sg_free_table(&umem->sg_head); 66 return; 67 68 } 69 70 /** 71 * ib_umem_get - Pin and DMA map userspace memory. 72 * 73 * If access flags indicate ODP memory, avoid pinning. Instead, stores 74 * the mm for future page fault handling in conjunction with MMU notifiers. 75 * 76 * @context: userspace context to pin memory for 77 * @addr: userspace virtual address to start at 78 * @size: length of region to pin 79 * @access: IB_ACCESS_xxx flags for memory being pinned 80 * @dmasync: flush in-flight DMA when the memory region is written 81 */ 82 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr, 83 size_t size, int access, int dmasync) 84 { 85 struct ib_umem *umem; 86 struct page **page_list; 87 struct vm_area_struct **vma_list; 88 unsigned long locked; 89 unsigned long lock_limit; 90 unsigned long cur_base; 91 unsigned long npages; 92 int ret; 93 int i; 94 unsigned long dma_attrs = 0; 95 struct scatterlist *sg, *sg_list_start; 96 int need_release = 0; 97 unsigned int gup_flags = FOLL_WRITE; 98 99 if (dmasync) 100 dma_attrs |= DMA_ATTR_WRITE_BARRIER; 101 102 if (!size) 103 return ERR_PTR(-EINVAL); 104 105 /* 106 * If the combination of the addr and size requested for this memory 107 * region causes an integer overflow, return error. 108 */ 109 if (((addr + size) < addr) || 110 PAGE_ALIGN(addr + size) < (addr + size)) 111 return ERR_PTR(-EINVAL); 112 113 if (!can_do_mlock()) 114 return ERR_PTR(-EPERM); 115 116 umem = kzalloc(sizeof *umem, GFP_KERNEL); 117 if (!umem) 118 return ERR_PTR(-ENOMEM); 119 120 umem->context = context; 121 umem->length = size; 122 umem->address = addr; 123 umem->page_size = PAGE_SIZE; 124 umem->pid = get_task_pid(current, PIDTYPE_PID); 125 /* 126 * We ask for writable memory if any of the following 127 * access flags are set. "Local write" and "remote write" 128 * obviously require write access. "Remote atomic" can do 129 * things like fetch and add, which will modify memory, and 130 * "MW bind" can change permissions by binding a window. 131 */ 132 umem->writable = !!(access & 133 (IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE | 134 IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_MW_BIND)); 135 136 if (access & IB_ACCESS_ON_DEMAND) { 137 put_pid(umem->pid); 138 ret = ib_umem_odp_get(context, umem); 139 if (ret) { 140 kfree(umem); 141 return ERR_PTR(ret); 142 } 143 return umem; 144 } 145 146 umem->odp_data = NULL; 147 148 /* We assume the memory is from hugetlb until proved otherwise */ 149 umem->hugetlb = 1; 150 151 page_list = (struct page **) __get_free_page(GFP_KERNEL); 152 if (!page_list) { 153 put_pid(umem->pid); 154 kfree(umem); 155 return ERR_PTR(-ENOMEM); 156 } 157 158 /* 159 * if we can't alloc the vma_list, it's not so bad; 160 * just assume the memory is not hugetlb memory 161 */ 162 vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL); 163 if (!vma_list) 164 umem->hugetlb = 0; 165 166 npages = ib_umem_num_pages(umem); 167 168 down_write(¤t->mm->mmap_sem); 169 170 locked = npages + current->mm->pinned_vm; 171 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 172 173 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) { 174 ret = -ENOMEM; 175 goto out; 176 } 177 178 cur_base = addr & PAGE_MASK; 179 180 if (npages == 0 || npages > UINT_MAX) { 181 ret = -EINVAL; 182 goto out; 183 } 184 185 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL); 186 if (ret) 187 goto out; 188 189 if (!umem->writable) 190 gup_flags |= FOLL_FORCE; 191 192 need_release = 1; 193 sg_list_start = umem->sg_head.sgl; 194 195 while (npages) { 196 ret = get_user_pages(cur_base, 197 min_t(unsigned long, npages, 198 PAGE_SIZE / sizeof (struct page *)), 199 gup_flags, page_list, vma_list); 200 201 if (ret < 0) 202 goto out; 203 204 umem->npages += ret; 205 cur_base += ret * PAGE_SIZE; 206 npages -= ret; 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 } 214 215 /* preparing for next loop */ 216 sg_list_start = sg; 217 } 218 219 umem->nmap = ib_dma_map_sg_attrs(context->device, 220 umem->sg_head.sgl, 221 umem->npages, 222 DMA_BIDIRECTIONAL, 223 dma_attrs); 224 225 if (umem->nmap <= 0) { 226 ret = -ENOMEM; 227 goto out; 228 } 229 230 ret = 0; 231 232 out: 233 if (ret < 0) { 234 if (need_release) 235 __ib_umem_release(context->device, umem, 0); 236 put_pid(umem->pid); 237 kfree(umem); 238 } else 239 current->mm->pinned_vm = locked; 240 241 up_write(¤t->mm->mmap_sem); 242 if (vma_list) 243 free_page((unsigned long) vma_list); 244 free_page((unsigned long) page_list); 245 246 return ret < 0 ? ERR_PTR(ret) : umem; 247 } 248 EXPORT_SYMBOL(ib_umem_get); 249 250 static void ib_umem_account(struct work_struct *work) 251 { 252 struct ib_umem *umem = container_of(work, struct ib_umem, work); 253 254 down_write(&umem->mm->mmap_sem); 255 umem->mm->pinned_vm -= umem->diff; 256 up_write(&umem->mm->mmap_sem); 257 mmput(umem->mm); 258 kfree(umem); 259 } 260 261 /** 262 * ib_umem_release - release memory pinned with ib_umem_get 263 * @umem: umem struct to release 264 */ 265 void ib_umem_release(struct ib_umem *umem) 266 { 267 struct ib_ucontext *context = umem->context; 268 struct mm_struct *mm; 269 struct task_struct *task; 270 unsigned long diff; 271 272 if (umem->odp_data) { 273 ib_umem_odp_release(umem); 274 return; 275 } 276 277 __ib_umem_release(umem->context->device, umem, 1); 278 279 task = get_pid_task(umem->pid, PIDTYPE_PID); 280 put_pid(umem->pid); 281 if (!task) 282 goto out; 283 mm = get_task_mm(task); 284 put_task_struct(task); 285 if (!mm) 286 goto out; 287 288 diff = ib_umem_num_pages(umem); 289 290 /* 291 * We may be called with the mm's mmap_sem already held. This 292 * can happen when a userspace munmap() is the call that drops 293 * the last reference to our file and calls our release 294 * method. If there are memory regions to destroy, we'll end 295 * up here and not be able to take the mmap_sem. In that case 296 * we defer the vm_locked accounting to the system workqueue. 297 */ 298 if (context->closing) { 299 if (!down_write_trylock(&mm->mmap_sem)) { 300 INIT_WORK(&umem->work, ib_umem_account); 301 umem->mm = mm; 302 umem->diff = diff; 303 304 queue_work(ib_wq, &umem->work); 305 return; 306 } 307 } else 308 down_write(&mm->mmap_sem); 309 310 mm->pinned_vm -= diff; 311 up_write(&mm->mmap_sem); 312 mmput(mm); 313 out: 314 kfree(umem); 315 } 316 EXPORT_SYMBOL(ib_umem_release); 317 318 int ib_umem_page_count(struct ib_umem *umem) 319 { 320 int shift; 321 int i; 322 int n; 323 struct scatterlist *sg; 324 325 if (umem->odp_data) 326 return ib_umem_num_pages(umem); 327 328 shift = ilog2(umem->page_size); 329 330 n = 0; 331 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) 332 n += sg_dma_len(sg) >> shift; 333 334 return n; 335 } 336 EXPORT_SYMBOL(ib_umem_page_count); 337 338 /* 339 * Copy from the given ib_umem's pages to the given buffer. 340 * 341 * umem - the umem to copy from 342 * offset - offset to start copying from 343 * dst - destination buffer 344 * length - buffer length 345 * 346 * Returns 0 on success, or an error code. 347 */ 348 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset, 349 size_t length) 350 { 351 size_t end = offset + length; 352 int ret; 353 354 if (offset > umem->length || length > umem->length - offset) { 355 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n", 356 offset, umem->length, end); 357 return -EINVAL; 358 } 359 360 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->nmap, dst, length, 361 offset + ib_umem_offset(umem)); 362 363 if (ret < 0) 364 return ret; 365 else if (ret != length) 366 return -EINVAL; 367 else 368 return 0; 369 } 370 EXPORT_SYMBOL(ib_umem_copy_from); 371