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/dma-attrs.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->nmap, 56 DMA_BIDIRECTIONAL); 57 58 for_each_sg(umem->sg_head.sgl, sg, umem->npages, i) { 59 60 page = sg_page(sg); 61 if (umem->writable && dirty) 62 set_page_dirty_lock(page); 63 put_page(page); 64 } 65 66 sg_free_table(&umem->sg_head); 67 return; 68 69 } 70 71 /** 72 * ib_umem_get - Pin and DMA map userspace memory. 73 * 74 * If access flags indicate ODP memory, avoid pinning. Instead, stores 75 * the mm for future page fault handling in conjunction with MMU notifiers. 76 * 77 * @context: userspace context to pin memory for 78 * @addr: userspace virtual address to start at 79 * @size: length of region to pin 80 * @access: IB_ACCESS_xxx flags for memory being pinned 81 * @dmasync: flush in-flight DMA when the memory region is written 82 */ 83 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr, 84 size_t size, int access, int dmasync) 85 { 86 struct ib_umem *umem; 87 struct page **page_list; 88 struct vm_area_struct **vma_list; 89 unsigned long locked; 90 unsigned long lock_limit; 91 unsigned long cur_base; 92 unsigned long npages; 93 int ret; 94 int i; 95 DEFINE_DMA_ATTRS(attrs); 96 struct scatterlist *sg, *sg_list_start; 97 int need_release = 0; 98 99 if (dmasync) 100 dma_set_attr(DMA_ATTR_WRITE_BARRIER, &attrs); 101 102 /* 103 * If the combination of the addr and size requested for this memory 104 * region causes an integer overflow, return error. 105 */ 106 if ((PAGE_ALIGN(addr + size) <= size) || 107 (PAGE_ALIGN(addr + size) <= addr)) 108 return ERR_PTR(-EINVAL); 109 110 if (!can_do_mlock()) 111 return ERR_PTR(-EPERM); 112 113 umem = kzalloc(sizeof *umem, GFP_KERNEL); 114 if (!umem) 115 return ERR_PTR(-ENOMEM); 116 117 umem->context = context; 118 umem->length = size; 119 umem->address = addr; 120 umem->page_size = PAGE_SIZE; 121 umem->pid = get_task_pid(current, PIDTYPE_PID); 122 /* 123 * We ask for writable memory if any of the following 124 * access flags are set. "Local write" and "remote write" 125 * obviously require write access. "Remote atomic" can do 126 * things like fetch and add, which will modify memory, and 127 * "MW bind" can change permissions by binding a window. 128 */ 129 umem->writable = !!(access & 130 (IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE | 131 IB_ACCESS_REMOTE_ATOMIC | IB_ACCESS_MW_BIND)); 132 133 if (access & IB_ACCESS_ON_DEMAND) { 134 ret = ib_umem_odp_get(context, umem); 135 if (ret) { 136 kfree(umem); 137 return ERR_PTR(ret); 138 } 139 return umem; 140 } 141 142 umem->odp_data = NULL; 143 144 /* We assume the memory is from hugetlb until proved otherwise */ 145 umem->hugetlb = 1; 146 147 page_list = (struct page **) __get_free_page(GFP_KERNEL); 148 if (!page_list) { 149 kfree(umem); 150 return ERR_PTR(-ENOMEM); 151 } 152 153 /* 154 * if we can't alloc the vma_list, it's not so bad; 155 * just assume the memory is not hugetlb memory 156 */ 157 vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL); 158 if (!vma_list) 159 umem->hugetlb = 0; 160 161 npages = ib_umem_num_pages(umem); 162 163 down_write(¤t->mm->mmap_sem); 164 165 locked = npages + current->mm->pinned_vm; 166 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 167 168 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) { 169 ret = -ENOMEM; 170 goto out; 171 } 172 173 cur_base = addr & PAGE_MASK; 174 175 if (npages == 0) { 176 ret = -EINVAL; 177 goto out; 178 } 179 180 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL); 181 if (ret) 182 goto out; 183 184 need_release = 1; 185 sg_list_start = umem->sg_head.sgl; 186 187 while (npages) { 188 ret = get_user_pages(current, current->mm, cur_base, 189 min_t(unsigned long, npages, 190 PAGE_SIZE / sizeof (struct page *)), 191 1, !umem->writable, page_list, vma_list); 192 193 if (ret < 0) 194 goto out; 195 196 umem->npages += ret; 197 cur_base += ret * PAGE_SIZE; 198 npages -= ret; 199 200 for_each_sg(sg_list_start, sg, ret, i) { 201 if (vma_list && !is_vm_hugetlb_page(vma_list[i])) 202 umem->hugetlb = 0; 203 204 sg_set_page(sg, page_list[i], PAGE_SIZE, 0); 205 } 206 207 /* preparing for next loop */ 208 sg_list_start = sg; 209 } 210 211 umem->nmap = ib_dma_map_sg_attrs(context->device, 212 umem->sg_head.sgl, 213 umem->npages, 214 DMA_BIDIRECTIONAL, 215 &attrs); 216 217 if (umem->nmap <= 0) { 218 ret = -ENOMEM; 219 goto out; 220 } 221 222 ret = 0; 223 224 out: 225 if (ret < 0) { 226 if (need_release) 227 __ib_umem_release(context->device, umem, 0); 228 put_pid(umem->pid); 229 kfree(umem); 230 } else 231 current->mm->pinned_vm = locked; 232 233 up_write(¤t->mm->mmap_sem); 234 if (vma_list) 235 free_page((unsigned long) vma_list); 236 free_page((unsigned long) page_list); 237 238 return ret < 0 ? ERR_PTR(ret) : umem; 239 } 240 EXPORT_SYMBOL(ib_umem_get); 241 242 static void ib_umem_account(struct work_struct *work) 243 { 244 struct ib_umem *umem = container_of(work, struct ib_umem, work); 245 246 down_write(&umem->mm->mmap_sem); 247 umem->mm->pinned_vm -= umem->diff; 248 up_write(&umem->mm->mmap_sem); 249 mmput(umem->mm); 250 kfree(umem); 251 } 252 253 /** 254 * ib_umem_release - release memory pinned with ib_umem_get 255 * @umem: umem struct to release 256 */ 257 void ib_umem_release(struct ib_umem *umem) 258 { 259 struct ib_ucontext *context = umem->context; 260 struct mm_struct *mm; 261 struct task_struct *task; 262 unsigned long diff; 263 264 if (umem->odp_data) { 265 ib_umem_odp_release(umem); 266 return; 267 } 268 269 __ib_umem_release(umem->context->device, umem, 1); 270 271 task = get_pid_task(umem->pid, PIDTYPE_PID); 272 put_pid(umem->pid); 273 if (!task) 274 goto out; 275 mm = get_task_mm(task); 276 put_task_struct(task); 277 if (!mm) 278 goto out; 279 280 diff = ib_umem_num_pages(umem); 281 282 /* 283 * We may be called with the mm's mmap_sem already held. This 284 * can happen when a userspace munmap() is the call that drops 285 * the last reference to our file and calls our release 286 * method. If there are memory regions to destroy, we'll end 287 * up here and not be able to take the mmap_sem. In that case 288 * we defer the vm_locked accounting to the system workqueue. 289 */ 290 if (context->closing) { 291 if (!down_write_trylock(&mm->mmap_sem)) { 292 INIT_WORK(&umem->work, ib_umem_account); 293 umem->mm = mm; 294 umem->diff = diff; 295 296 queue_work(ib_wq, &umem->work); 297 return; 298 } 299 } else 300 down_write(&mm->mmap_sem); 301 302 mm->pinned_vm -= diff; 303 up_write(&mm->mmap_sem); 304 mmput(mm); 305 out: 306 kfree(umem); 307 } 308 EXPORT_SYMBOL(ib_umem_release); 309 310 int ib_umem_page_count(struct ib_umem *umem) 311 { 312 int shift; 313 int i; 314 int n; 315 struct scatterlist *sg; 316 317 if (umem->odp_data) 318 return ib_umem_num_pages(umem); 319 320 shift = ilog2(umem->page_size); 321 322 n = 0; 323 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) 324 n += sg_dma_len(sg) >> shift; 325 326 return n; 327 } 328 EXPORT_SYMBOL(ib_umem_page_count); 329 330 /* 331 * Copy from the given ib_umem's pages to the given buffer. 332 * 333 * umem - the umem to copy from 334 * offset - offset to start copying from 335 * dst - destination buffer 336 * length - buffer length 337 * 338 * Returns 0 on success, or an error code. 339 */ 340 int ib_umem_copy_from(void *dst, struct ib_umem *umem, size_t offset, 341 size_t length) 342 { 343 size_t end = offset + length; 344 int ret; 345 346 if (offset > umem->length || length > umem->length - offset) { 347 pr_err("ib_umem_copy_from not in range. offset: %zd umem length: %zd end: %zd\n", 348 offset, umem->length, end); 349 return -EINVAL; 350 } 351 352 ret = sg_pcopy_to_buffer(umem->sg_head.sgl, umem->nmap, dst, length, 353 offset + ib_umem_offset(umem)); 354 355 if (ret < 0) 356 return ret; 357 else if (ret != length) 358 return -EINVAL; 359 else 360 return 0; 361 } 362 EXPORT_SYMBOL(ib_umem_copy_from); 363