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 #include <linux/netdevice.h> 15 #include <linux/rtnetlink.h> 16 #include <linux/idr.h> 17 18 #include "xdp_umem.h" 19 #include "xsk_queue.h" 20 21 #define XDP_UMEM_MIN_CHUNK_SIZE 2048 22 23 static DEFINE_IDA(umem_ida); 24 25 void xdp_add_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs) 26 { 27 unsigned long flags; 28 29 spin_lock_irqsave(&umem->xsk_list_lock, flags); 30 list_add_rcu(&xs->list, &umem->xsk_list); 31 spin_unlock_irqrestore(&umem->xsk_list_lock, flags); 32 } 33 34 void xdp_del_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs) 35 { 36 unsigned long flags; 37 38 spin_lock_irqsave(&umem->xsk_list_lock, flags); 39 list_del_rcu(&xs->list); 40 spin_unlock_irqrestore(&umem->xsk_list_lock, flags); 41 } 42 43 /* The umem is stored both in the _rx struct and the _tx struct as we do 44 * not know if the device has more tx queues than rx, or the opposite. 45 * This might also change during run time. 46 */ 47 static int xdp_reg_umem_at_qid(struct net_device *dev, struct xdp_umem *umem, 48 u16 queue_id) 49 { 50 if (queue_id >= max_t(unsigned int, 51 dev->real_num_rx_queues, 52 dev->real_num_tx_queues)) 53 return -EINVAL; 54 55 if (queue_id < dev->real_num_rx_queues) 56 dev->_rx[queue_id].umem = umem; 57 if (queue_id < dev->real_num_tx_queues) 58 dev->_tx[queue_id].umem = umem; 59 60 return 0; 61 } 62 63 struct xdp_umem *xdp_get_umem_from_qid(struct net_device *dev, 64 u16 queue_id) 65 { 66 if (queue_id < dev->real_num_rx_queues) 67 return dev->_rx[queue_id].umem; 68 if (queue_id < dev->real_num_tx_queues) 69 return dev->_tx[queue_id].umem; 70 71 return NULL; 72 } 73 EXPORT_SYMBOL(xdp_get_umem_from_qid); 74 75 static void xdp_clear_umem_at_qid(struct net_device *dev, u16 queue_id) 76 { 77 if (queue_id < dev->real_num_rx_queues) 78 dev->_rx[queue_id].umem = NULL; 79 if (queue_id < dev->real_num_tx_queues) 80 dev->_tx[queue_id].umem = NULL; 81 } 82 83 int xdp_umem_assign_dev(struct xdp_umem *umem, struct net_device *dev, 84 u16 queue_id, u16 flags) 85 { 86 bool force_zc, force_copy; 87 struct netdev_bpf bpf; 88 int err = 0; 89 90 force_zc = flags & XDP_ZEROCOPY; 91 force_copy = flags & XDP_COPY; 92 93 if (force_zc && force_copy) 94 return -EINVAL; 95 96 rtnl_lock(); 97 if (xdp_get_umem_from_qid(dev, queue_id)) { 98 err = -EBUSY; 99 goto out_rtnl_unlock; 100 } 101 102 err = xdp_reg_umem_at_qid(dev, umem, queue_id); 103 if (err) 104 goto out_rtnl_unlock; 105 106 umem->dev = dev; 107 umem->queue_id = queue_id; 108 if (force_copy) 109 /* For copy-mode, we are done. */ 110 goto out_rtnl_unlock; 111 112 if (!dev->netdev_ops->ndo_bpf || 113 !dev->netdev_ops->ndo_xsk_async_xmit) { 114 err = -EOPNOTSUPP; 115 goto err_unreg_umem; 116 } 117 118 bpf.command = XDP_SETUP_XSK_UMEM; 119 bpf.xsk.umem = umem; 120 bpf.xsk.queue_id = queue_id; 121 122 err = dev->netdev_ops->ndo_bpf(dev, &bpf); 123 if (err) 124 goto err_unreg_umem; 125 rtnl_unlock(); 126 127 dev_hold(dev); 128 umem->zc = true; 129 return 0; 130 131 err_unreg_umem: 132 if (!force_zc) 133 err = 0; /* fallback to copy mode */ 134 if (err) 135 xdp_clear_umem_at_qid(dev, queue_id); 136 out_rtnl_unlock: 137 rtnl_unlock(); 138 return err; 139 } 140 141 static void xdp_umem_clear_dev(struct xdp_umem *umem) 142 { 143 struct netdev_bpf bpf; 144 int err; 145 146 if (!umem->dev) 147 return; 148 149 if (umem->zc) { 150 bpf.command = XDP_SETUP_XSK_UMEM; 151 bpf.xsk.umem = NULL; 152 bpf.xsk.queue_id = umem->queue_id; 153 154 rtnl_lock(); 155 err = umem->dev->netdev_ops->ndo_bpf(umem->dev, &bpf); 156 rtnl_unlock(); 157 158 if (err) 159 WARN(1, "failed to disable umem!\n"); 160 } 161 162 rtnl_lock(); 163 xdp_clear_umem_at_qid(umem->dev, umem->queue_id); 164 rtnl_unlock(); 165 166 if (umem->zc) { 167 dev_put(umem->dev); 168 umem->zc = false; 169 } 170 } 171 172 static void xdp_umem_unpin_pages(struct xdp_umem *umem) 173 { 174 unsigned int i; 175 176 for (i = 0; i < umem->npgs; i++) { 177 struct page *page = umem->pgs[i]; 178 179 set_page_dirty_lock(page); 180 put_page(page); 181 } 182 183 kfree(umem->pgs); 184 umem->pgs = NULL; 185 } 186 187 static void xdp_umem_unaccount_pages(struct xdp_umem *umem) 188 { 189 if (umem->user) { 190 atomic_long_sub(umem->npgs, &umem->user->locked_vm); 191 free_uid(umem->user); 192 } 193 } 194 195 static void xdp_umem_release(struct xdp_umem *umem) 196 { 197 xdp_umem_clear_dev(umem); 198 199 ida_simple_remove(&umem_ida, umem->id); 200 201 if (umem->fq) { 202 xskq_destroy(umem->fq); 203 umem->fq = NULL; 204 } 205 206 if (umem->cq) { 207 xskq_destroy(umem->cq); 208 umem->cq = NULL; 209 } 210 211 xsk_reuseq_destroy(umem); 212 213 xdp_umem_unpin_pages(umem); 214 215 kfree(umem->pages); 216 umem->pages = NULL; 217 218 xdp_umem_unaccount_pages(umem); 219 kfree(umem); 220 } 221 222 static void xdp_umem_release_deferred(struct work_struct *work) 223 { 224 struct xdp_umem *umem = container_of(work, struct xdp_umem, work); 225 226 xdp_umem_release(umem); 227 } 228 229 void xdp_get_umem(struct xdp_umem *umem) 230 { 231 refcount_inc(&umem->users); 232 } 233 234 void xdp_put_umem(struct xdp_umem *umem) 235 { 236 if (!umem) 237 return; 238 239 if (refcount_dec_and_test(&umem->users)) { 240 INIT_WORK(&umem->work, xdp_umem_release_deferred); 241 schedule_work(&umem->work); 242 } 243 } 244 245 static int xdp_umem_pin_pages(struct xdp_umem *umem) 246 { 247 unsigned int gup_flags = FOLL_WRITE; 248 long npgs; 249 int err; 250 251 umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs), 252 GFP_KERNEL | __GFP_NOWARN); 253 if (!umem->pgs) 254 return -ENOMEM; 255 256 down_read(¤t->mm->mmap_sem); 257 npgs = get_user_pages(umem->address, umem->npgs, 258 gup_flags | FOLL_LONGTERM, &umem->pgs[0], NULL); 259 up_read(¤t->mm->mmap_sem); 260 261 if (npgs != umem->npgs) { 262 if (npgs >= 0) { 263 umem->npgs = npgs; 264 err = -ENOMEM; 265 goto out_pin; 266 } 267 err = npgs; 268 goto out_pgs; 269 } 270 return 0; 271 272 out_pin: 273 xdp_umem_unpin_pages(umem); 274 out_pgs: 275 kfree(umem->pgs); 276 umem->pgs = NULL; 277 return err; 278 } 279 280 static int xdp_umem_account_pages(struct xdp_umem *umem) 281 { 282 unsigned long lock_limit, new_npgs, old_npgs; 283 284 if (capable(CAP_IPC_LOCK)) 285 return 0; 286 287 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 288 umem->user = get_uid(current_user()); 289 290 do { 291 old_npgs = atomic_long_read(&umem->user->locked_vm); 292 new_npgs = old_npgs + umem->npgs; 293 if (new_npgs > lock_limit) { 294 free_uid(umem->user); 295 umem->user = NULL; 296 return -ENOBUFS; 297 } 298 } while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs, 299 new_npgs) != old_npgs); 300 return 0; 301 } 302 303 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr) 304 { 305 u32 chunk_size = mr->chunk_size, headroom = mr->headroom; 306 unsigned int chunks, chunks_per_page; 307 u64 addr = mr->addr, size = mr->len; 308 int size_chk, err, i; 309 310 if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) { 311 /* Strictly speaking we could support this, if: 312 * - huge pages, or* 313 * - using an IOMMU, or 314 * - making sure the memory area is consecutive 315 * but for now, we simply say "computer says no". 316 */ 317 return -EINVAL; 318 } 319 320 if (!is_power_of_2(chunk_size)) 321 return -EINVAL; 322 323 if (!PAGE_ALIGNED(addr)) { 324 /* Memory area has to be page size aligned. For 325 * simplicity, this might change. 326 */ 327 return -EINVAL; 328 } 329 330 if ((addr + size) < addr) 331 return -EINVAL; 332 333 chunks = (unsigned int)div_u64(size, chunk_size); 334 if (chunks == 0) 335 return -EINVAL; 336 337 chunks_per_page = PAGE_SIZE / chunk_size; 338 if (chunks < chunks_per_page || chunks % chunks_per_page) 339 return -EINVAL; 340 341 headroom = ALIGN(headroom, 64); 342 343 size_chk = chunk_size - headroom - XDP_PACKET_HEADROOM; 344 if (size_chk < 0) 345 return -EINVAL; 346 347 umem->address = (unsigned long)addr; 348 umem->chunk_mask = ~((u64)chunk_size - 1); 349 umem->size = size; 350 umem->headroom = headroom; 351 umem->chunk_size_nohr = chunk_size - headroom; 352 umem->npgs = size / PAGE_SIZE; 353 umem->pgs = NULL; 354 umem->user = NULL; 355 INIT_LIST_HEAD(&umem->xsk_list); 356 spin_lock_init(&umem->xsk_list_lock); 357 358 refcount_set(&umem->users, 1); 359 360 err = xdp_umem_account_pages(umem); 361 if (err) 362 return err; 363 364 err = xdp_umem_pin_pages(umem); 365 if (err) 366 goto out_account; 367 368 umem->pages = kcalloc(umem->npgs, sizeof(*umem->pages), GFP_KERNEL); 369 if (!umem->pages) { 370 err = -ENOMEM; 371 goto out_account; 372 } 373 374 for (i = 0; i < umem->npgs; i++) 375 umem->pages[i].addr = page_address(umem->pgs[i]); 376 377 return 0; 378 379 out_account: 380 xdp_umem_unaccount_pages(umem); 381 return err; 382 } 383 384 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr) 385 { 386 struct xdp_umem *umem; 387 int err; 388 389 umem = kzalloc(sizeof(*umem), GFP_KERNEL); 390 if (!umem) 391 return ERR_PTR(-ENOMEM); 392 393 err = ida_simple_get(&umem_ida, 0, 0, GFP_KERNEL); 394 if (err < 0) { 395 kfree(umem); 396 return ERR_PTR(err); 397 } 398 umem->id = err; 399 400 err = xdp_umem_reg(umem, mr); 401 if (err) { 402 ida_simple_remove(&umem_ida, umem->id); 403 kfree(umem); 404 return ERR_PTR(err); 405 } 406 407 return umem; 408 } 409 410 bool xdp_umem_validate_queues(struct xdp_umem *umem) 411 { 412 return umem->fq && umem->cq; 413 } 414