1 // SPDX-License-Identifier: GPL-2.0 2 /* XDP user-space packet buffer 3 * Copyright(c) 2018 Intel Corporation. 4 */ 5 6 #include <linux/init.h> 7 #include <linux/sched/mm.h> 8 #include <linux/sched/signal.h> 9 #include <linux/sched/task.h> 10 #include <linux/uaccess.h> 11 #include <linux/slab.h> 12 #include <linux/bpf.h> 13 #include <linux/mm.h> 14 15 #include "xdp_umem.h" 16 #include "xsk_queue.h" 17 18 #define XDP_UMEM_MIN_CHUNK_SIZE 2048 19 20 void xdp_add_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs) 21 { 22 unsigned long flags; 23 24 spin_lock_irqsave(&umem->xsk_list_lock, flags); 25 list_add_rcu(&xs->list, &umem->xsk_list); 26 spin_unlock_irqrestore(&umem->xsk_list_lock, flags); 27 } 28 29 void xdp_del_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs) 30 { 31 unsigned long flags; 32 33 if (xs->dev) { 34 spin_lock_irqsave(&umem->xsk_list_lock, flags); 35 list_del_rcu(&xs->list); 36 spin_unlock_irqrestore(&umem->xsk_list_lock, flags); 37 38 if (umem->zc) 39 synchronize_net(); 40 } 41 } 42 43 int xdp_umem_assign_dev(struct xdp_umem *umem, struct net_device *dev, 44 u32 queue_id, u16 flags) 45 { 46 bool force_zc, force_copy; 47 struct netdev_bpf bpf; 48 int err; 49 50 force_zc = flags & XDP_ZEROCOPY; 51 force_copy = flags & XDP_COPY; 52 53 if (force_zc && force_copy) 54 return -EINVAL; 55 56 if (force_copy) 57 return 0; 58 59 dev_hold(dev); 60 61 if (dev->netdev_ops->ndo_bpf && dev->netdev_ops->ndo_xsk_async_xmit) { 62 bpf.command = XDP_QUERY_XSK_UMEM; 63 64 rtnl_lock(); 65 err = dev->netdev_ops->ndo_bpf(dev, &bpf); 66 rtnl_unlock(); 67 68 if (err) { 69 dev_put(dev); 70 return force_zc ? -ENOTSUPP : 0; 71 } 72 73 bpf.command = XDP_SETUP_XSK_UMEM; 74 bpf.xsk.umem = umem; 75 bpf.xsk.queue_id = queue_id; 76 77 rtnl_lock(); 78 err = dev->netdev_ops->ndo_bpf(dev, &bpf); 79 rtnl_unlock(); 80 81 if (err) { 82 dev_put(dev); 83 return force_zc ? err : 0; /* fail or fallback */ 84 } 85 86 umem->dev = dev; 87 umem->queue_id = queue_id; 88 umem->zc = true; 89 return 0; 90 } 91 92 dev_put(dev); 93 return force_zc ? -ENOTSUPP : 0; /* fail or fallback */ 94 } 95 96 static void xdp_umem_clear_dev(struct xdp_umem *umem) 97 { 98 struct netdev_bpf bpf; 99 int err; 100 101 if (umem->dev) { 102 bpf.command = XDP_SETUP_XSK_UMEM; 103 bpf.xsk.umem = NULL; 104 bpf.xsk.queue_id = umem->queue_id; 105 106 rtnl_lock(); 107 err = umem->dev->netdev_ops->ndo_bpf(umem->dev, &bpf); 108 rtnl_unlock(); 109 110 if (err) 111 WARN(1, "failed to disable umem!\n"); 112 113 dev_put(umem->dev); 114 umem->dev = NULL; 115 } 116 } 117 118 static void xdp_umem_unpin_pages(struct xdp_umem *umem) 119 { 120 unsigned int i; 121 122 for (i = 0; i < umem->npgs; i++) { 123 struct page *page = umem->pgs[i]; 124 125 set_page_dirty_lock(page); 126 put_page(page); 127 } 128 129 kfree(umem->pgs); 130 umem->pgs = NULL; 131 } 132 133 static void xdp_umem_unaccount_pages(struct xdp_umem *umem) 134 { 135 atomic_long_sub(umem->npgs, &umem->user->locked_vm); 136 free_uid(umem->user); 137 } 138 139 static void xdp_umem_release(struct xdp_umem *umem) 140 { 141 struct task_struct *task; 142 struct mm_struct *mm; 143 144 xdp_umem_clear_dev(umem); 145 146 if (umem->fq) { 147 xskq_destroy(umem->fq); 148 umem->fq = NULL; 149 } 150 151 if (umem->cq) { 152 xskq_destroy(umem->cq); 153 umem->cq = NULL; 154 } 155 156 xdp_umem_unpin_pages(umem); 157 158 task = get_pid_task(umem->pid, PIDTYPE_PID); 159 put_pid(umem->pid); 160 if (!task) 161 goto out; 162 mm = get_task_mm(task); 163 put_task_struct(task); 164 if (!mm) 165 goto out; 166 167 mmput(mm); 168 kfree(umem->pages); 169 umem->pages = NULL; 170 171 xdp_umem_unaccount_pages(umem); 172 out: 173 kfree(umem); 174 } 175 176 static void xdp_umem_release_deferred(struct work_struct *work) 177 { 178 struct xdp_umem *umem = container_of(work, struct xdp_umem, work); 179 180 xdp_umem_release(umem); 181 } 182 183 void xdp_get_umem(struct xdp_umem *umem) 184 { 185 refcount_inc(&umem->users); 186 } 187 188 void xdp_put_umem(struct xdp_umem *umem) 189 { 190 if (!umem) 191 return; 192 193 if (refcount_dec_and_test(&umem->users)) { 194 INIT_WORK(&umem->work, xdp_umem_release_deferred); 195 schedule_work(&umem->work); 196 } 197 } 198 199 static int xdp_umem_pin_pages(struct xdp_umem *umem) 200 { 201 unsigned int gup_flags = FOLL_WRITE; 202 long npgs; 203 int err; 204 205 umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs), GFP_KERNEL); 206 if (!umem->pgs) 207 return -ENOMEM; 208 209 down_write(¤t->mm->mmap_sem); 210 npgs = get_user_pages(umem->address, umem->npgs, 211 gup_flags, &umem->pgs[0], NULL); 212 up_write(¤t->mm->mmap_sem); 213 214 if (npgs != umem->npgs) { 215 if (npgs >= 0) { 216 umem->npgs = npgs; 217 err = -ENOMEM; 218 goto out_pin; 219 } 220 err = npgs; 221 goto out_pgs; 222 } 223 return 0; 224 225 out_pin: 226 xdp_umem_unpin_pages(umem); 227 out_pgs: 228 kfree(umem->pgs); 229 umem->pgs = NULL; 230 return err; 231 } 232 233 static int xdp_umem_account_pages(struct xdp_umem *umem) 234 { 235 unsigned long lock_limit, new_npgs, old_npgs; 236 237 if (capable(CAP_IPC_LOCK)) 238 return 0; 239 240 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 241 umem->user = get_uid(current_user()); 242 243 do { 244 old_npgs = atomic_long_read(&umem->user->locked_vm); 245 new_npgs = old_npgs + umem->npgs; 246 if (new_npgs > lock_limit) { 247 free_uid(umem->user); 248 umem->user = NULL; 249 return -ENOBUFS; 250 } 251 } while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs, 252 new_npgs) != old_npgs); 253 return 0; 254 } 255 256 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr) 257 { 258 u32 chunk_size = mr->chunk_size, headroom = mr->headroom; 259 unsigned int chunks, chunks_per_page; 260 u64 addr = mr->addr, size = mr->len; 261 int size_chk, err, i; 262 263 if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) { 264 /* Strictly speaking we could support this, if: 265 * - huge pages, or* 266 * - using an IOMMU, or 267 * - making sure the memory area is consecutive 268 * but for now, we simply say "computer says no". 269 */ 270 return -EINVAL; 271 } 272 273 if (!is_power_of_2(chunk_size)) 274 return -EINVAL; 275 276 if (!PAGE_ALIGNED(addr)) { 277 /* Memory area has to be page size aligned. For 278 * simplicity, this might change. 279 */ 280 return -EINVAL; 281 } 282 283 if ((addr + size) < addr) 284 return -EINVAL; 285 286 chunks = (unsigned int)div_u64(size, chunk_size); 287 if (chunks == 0) 288 return -EINVAL; 289 290 chunks_per_page = PAGE_SIZE / chunk_size; 291 if (chunks < chunks_per_page || chunks % chunks_per_page) 292 return -EINVAL; 293 294 headroom = ALIGN(headroom, 64); 295 296 size_chk = chunk_size - headroom - XDP_PACKET_HEADROOM; 297 if (size_chk < 0) 298 return -EINVAL; 299 300 umem->pid = get_task_pid(current, PIDTYPE_PID); 301 umem->address = (unsigned long)addr; 302 umem->props.chunk_mask = ~((u64)chunk_size - 1); 303 umem->props.size = size; 304 umem->headroom = headroom; 305 umem->chunk_size_nohr = chunk_size - headroom; 306 umem->npgs = size / PAGE_SIZE; 307 umem->pgs = NULL; 308 umem->user = NULL; 309 INIT_LIST_HEAD(&umem->xsk_list); 310 spin_lock_init(&umem->xsk_list_lock); 311 312 refcount_set(&umem->users, 1); 313 314 err = xdp_umem_account_pages(umem); 315 if (err) 316 goto out; 317 318 err = xdp_umem_pin_pages(umem); 319 if (err) 320 goto out_account; 321 322 umem->pages = kcalloc(umem->npgs, sizeof(*umem->pages), GFP_KERNEL); 323 if (!umem->pages) { 324 err = -ENOMEM; 325 goto out_account; 326 } 327 328 for (i = 0; i < umem->npgs; i++) 329 umem->pages[i].addr = page_address(umem->pgs[i]); 330 331 return 0; 332 333 out_account: 334 xdp_umem_unaccount_pages(umem); 335 out: 336 put_pid(umem->pid); 337 return err; 338 } 339 340 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr) 341 { 342 struct xdp_umem *umem; 343 int err; 344 345 umem = kzalloc(sizeof(*umem), GFP_KERNEL); 346 if (!umem) 347 return ERR_PTR(-ENOMEM); 348 349 err = xdp_umem_reg(umem, mr); 350 if (err) { 351 kfree(umem); 352 return ERR_PTR(err); 353 } 354 355 return umem; 356 } 357 358 bool xdp_umem_validate_queues(struct xdp_umem *umem) 359 { 360 return umem->fq && umem->cq; 361 } 362