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 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->nmap, 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 * @context: userspace context to pin memory for 73 * @addr: userspace virtual address to start at 74 * @size: length of region to pin 75 * @access: IB_ACCESS_xxx flags for memory being pinned 76 * @dmasync: flush in-flight DMA when the memory region is written 77 */ 78 struct ib_umem *ib_umem_get(struct ib_ucontext *context, unsigned long addr, 79 size_t size, int access, int dmasync) 80 { 81 struct ib_umem *umem; 82 struct page **page_list; 83 struct vm_area_struct **vma_list; 84 unsigned long locked; 85 unsigned long lock_limit; 86 unsigned long cur_base; 87 unsigned long npages; 88 int ret; 89 int i; 90 DEFINE_DMA_ATTRS(attrs); 91 struct scatterlist *sg, *sg_list_start; 92 int need_release = 0; 93 94 if (dmasync) 95 dma_set_attr(DMA_ATTR_WRITE_BARRIER, &attrs); 96 97 if (!can_do_mlock()) 98 return ERR_PTR(-EPERM); 99 100 umem = kzalloc(sizeof *umem, GFP_KERNEL); 101 if (!umem) 102 return ERR_PTR(-ENOMEM); 103 104 umem->context = context; 105 umem->length = size; 106 umem->offset = addr & ~PAGE_MASK; 107 umem->page_size = PAGE_SIZE; 108 umem->pid = get_task_pid(current, PIDTYPE_PID); 109 /* 110 * We ask for writable memory if any access flags other than 111 * "remote read" are set. "Local write" and "remote write" 112 * obviously require write access. "Remote atomic" can do 113 * things like fetch and add, which will modify memory, and 114 * "MW bind" can change permissions by binding a window. 115 */ 116 umem->writable = !!(access & ~IB_ACCESS_REMOTE_READ); 117 118 /* We assume the memory is from hugetlb until proved otherwise */ 119 umem->hugetlb = 1; 120 121 page_list = (struct page **) __get_free_page(GFP_KERNEL); 122 if (!page_list) { 123 kfree(umem); 124 return ERR_PTR(-ENOMEM); 125 } 126 127 /* 128 * if we can't alloc the vma_list, it's not so bad; 129 * just assume the memory is not hugetlb memory 130 */ 131 vma_list = (struct vm_area_struct **) __get_free_page(GFP_KERNEL); 132 if (!vma_list) 133 umem->hugetlb = 0; 134 135 npages = PAGE_ALIGN(size + umem->offset) >> PAGE_SHIFT; 136 137 down_write(¤t->mm->mmap_sem); 138 139 locked = npages + current->mm->pinned_vm; 140 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 141 142 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) { 143 ret = -ENOMEM; 144 goto out; 145 } 146 147 cur_base = addr & PAGE_MASK; 148 149 if (npages == 0) { 150 ret = -EINVAL; 151 goto out; 152 } 153 154 ret = sg_alloc_table(&umem->sg_head, npages, GFP_KERNEL); 155 if (ret) 156 goto out; 157 158 need_release = 1; 159 sg_list_start = umem->sg_head.sgl; 160 161 while (npages) { 162 ret = get_user_pages(current, current->mm, cur_base, 163 min_t(unsigned long, npages, 164 PAGE_SIZE / sizeof (struct page *)), 165 1, !umem->writable, page_list, vma_list); 166 167 if (ret < 0) 168 goto out; 169 170 umem->npages += ret; 171 cur_base += ret * PAGE_SIZE; 172 npages -= ret; 173 174 for_each_sg(sg_list_start, sg, ret, i) { 175 if (vma_list && !is_vm_hugetlb_page(vma_list[i])) 176 umem->hugetlb = 0; 177 178 sg_set_page(sg, page_list[i], PAGE_SIZE, 0); 179 } 180 181 /* preparing for next loop */ 182 sg_list_start = sg; 183 } 184 185 umem->nmap = ib_dma_map_sg_attrs(context->device, 186 umem->sg_head.sgl, 187 umem->npages, 188 DMA_BIDIRECTIONAL, 189 &attrs); 190 191 if (umem->nmap <= 0) { 192 ret = -ENOMEM; 193 goto out; 194 } 195 196 ret = 0; 197 198 out: 199 if (ret < 0) { 200 if (need_release) 201 __ib_umem_release(context->device, umem, 0); 202 put_pid(umem->pid); 203 kfree(umem); 204 } else 205 current->mm->pinned_vm = locked; 206 207 up_write(¤t->mm->mmap_sem); 208 if (vma_list) 209 free_page((unsigned long) vma_list); 210 free_page((unsigned long) page_list); 211 212 return ret < 0 ? ERR_PTR(ret) : umem; 213 } 214 EXPORT_SYMBOL(ib_umem_get); 215 216 static void ib_umem_account(struct work_struct *work) 217 { 218 struct ib_umem *umem = container_of(work, struct ib_umem, work); 219 220 down_write(&umem->mm->mmap_sem); 221 umem->mm->pinned_vm -= umem->diff; 222 up_write(&umem->mm->mmap_sem); 223 mmput(umem->mm); 224 kfree(umem); 225 } 226 227 /** 228 * ib_umem_release - release memory pinned with ib_umem_get 229 * @umem: umem struct to release 230 */ 231 void ib_umem_release(struct ib_umem *umem) 232 { 233 struct ib_ucontext *context = umem->context; 234 struct mm_struct *mm; 235 struct task_struct *task; 236 unsigned long diff; 237 238 __ib_umem_release(umem->context->device, umem, 1); 239 240 task = get_pid_task(umem->pid, PIDTYPE_PID); 241 put_pid(umem->pid); 242 if (!task) 243 goto out; 244 mm = get_task_mm(task); 245 put_task_struct(task); 246 if (!mm) 247 goto out; 248 249 diff = PAGE_ALIGN(umem->length + umem->offset) >> PAGE_SHIFT; 250 251 /* 252 * We may be called with the mm's mmap_sem already held. This 253 * can happen when a userspace munmap() is the call that drops 254 * the last reference to our file and calls our release 255 * method. If there are memory regions to destroy, we'll end 256 * up here and not be able to take the mmap_sem. In that case 257 * we defer the vm_locked accounting to the system workqueue. 258 */ 259 if (context->closing) { 260 if (!down_write_trylock(&mm->mmap_sem)) { 261 INIT_WORK(&umem->work, ib_umem_account); 262 umem->mm = mm; 263 umem->diff = diff; 264 265 queue_work(ib_wq, &umem->work); 266 return; 267 } 268 } else 269 down_write(&mm->mmap_sem); 270 271 mm->pinned_vm -= diff; 272 up_write(&mm->mmap_sem); 273 mmput(mm); 274 out: 275 kfree(umem); 276 } 277 EXPORT_SYMBOL(ib_umem_release); 278 279 int ib_umem_page_count(struct ib_umem *umem) 280 { 281 int shift; 282 int i; 283 int n; 284 struct scatterlist *sg; 285 286 shift = ilog2(umem->page_size); 287 288 n = 0; 289 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, i) 290 n += sg_dma_len(sg) >> shift; 291 292 return n; 293 } 294 EXPORT_SYMBOL(ib_umem_page_count); 295