| 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 | #include <linux/vmalloc.h> |
| 18 | |
| 19 | #include "xdp_umem.h" |
| 20 | #include "xsk_queue.h" |
| 21 | |
| 22 | #define XDP_UMEM_MIN_CHUNK_SIZE 2048 |
| 23 | |
| 24 | static DEFINE_IDA(umem_ida); |
| 25 | |
| 26 | void xdp_add_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs) |
| 27 | { |
| 28 | unsigned long flags; |
| 29 | |
| 30 | if (!xs->tx) |
| 31 | return; |
| 32 | |
| 33 | spin_lock_irqsave(&umem->xsk_tx_list_lock, flags); |
| 34 | list_add_rcu(&xs->list, &umem->xsk_tx_list); |
| 35 | spin_unlock_irqrestore(&umem->xsk_tx_list_lock, flags); |
| 36 | } |
| 37 | |
| 38 | void xdp_del_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs) |
| 39 | { |
| 40 | unsigned long flags; |
| 41 | |
| 42 | if (!xs->tx) |
| 43 | return; |
| 44 | |
| 45 | spin_lock_irqsave(&umem->xsk_tx_list_lock, flags); |
| 46 | list_del_rcu(&xs->list); |
| 47 | spin_unlock_irqrestore(&umem->xsk_tx_list_lock, flags); |
| 48 | } |
| 49 | |
| 50 | /* The umem is stored both in the _rx struct and the _tx struct as we do |
| 51 | * not know if the device has more tx queues than rx, or the opposite. |
| 52 | * This might also change during run time. |
| 53 | */ |
| 54 | static int xdp_reg_umem_at_qid(struct net_device *dev, struct xdp_umem *umem, |
| 55 | u16 queue_id) |
| 56 | { |
| 57 | if (queue_id >= max_t(unsigned int, |
| 58 | dev->real_num_rx_queues, |
| 59 | dev->real_num_tx_queues)) |
| 60 | return -EINVAL; |
| 61 | |
| 62 | if (queue_id < dev->real_num_rx_queues) |
| 63 | dev->_rx[queue_id].umem = umem; |
| 64 | if (queue_id < dev->real_num_tx_queues) |
| 65 | dev->_tx[queue_id].umem = umem; |
| 66 | |
| 67 | return 0; |
| 68 | } |
| 69 | |
| 70 | struct xdp_umem *xdp_get_umem_from_qid(struct net_device *dev, |
| 71 | u16 queue_id) |
| 72 | { |
| 73 | if (queue_id < dev->real_num_rx_queues) |
| 74 | return dev->_rx[queue_id].umem; |
| 75 | if (queue_id < dev->real_num_tx_queues) |
| 76 | return dev->_tx[queue_id].umem; |
| 77 | |
| 78 | return NULL; |
| 79 | } |
| 80 | EXPORT_SYMBOL(xdp_get_umem_from_qid); |
| 81 | |
| 82 | static void xdp_clear_umem_at_qid(struct net_device *dev, u16 queue_id) |
| 83 | { |
| 84 | if (queue_id < dev->real_num_rx_queues) |
| 85 | dev->_rx[queue_id].umem = NULL; |
| 86 | if (queue_id < dev->real_num_tx_queues) |
| 87 | dev->_tx[queue_id].umem = NULL; |
| 88 | } |
| 89 | |
| 90 | int xdp_umem_assign_dev(struct xdp_umem *umem, struct net_device *dev, |
| 91 | u16 queue_id, u16 flags) |
| 92 | { |
| 93 | bool force_zc, force_copy; |
| 94 | struct netdev_bpf bpf; |
| 95 | int err = 0; |
| 96 | |
| 97 | ASSERT_RTNL(); |
| 98 | |
| 99 | force_zc = flags & XDP_ZEROCOPY; |
| 100 | force_copy = flags & XDP_COPY; |
| 101 | |
| 102 | if (force_zc && force_copy) |
| 103 | return -EINVAL; |
| 104 | |
| 105 | if (xdp_get_umem_from_qid(dev, queue_id)) |
| 106 | return -EBUSY; |
| 107 | |
| 108 | err = xdp_reg_umem_at_qid(dev, umem, queue_id); |
| 109 | if (err) |
| 110 | return err; |
| 111 | |
| 112 | umem->dev = dev; |
| 113 | umem->queue_id = queue_id; |
| 114 | |
| 115 | if (flags & XDP_USE_NEED_WAKEUP) { |
| 116 | umem->flags |= XDP_UMEM_USES_NEED_WAKEUP; |
| 117 | /* Tx needs to be explicitly woken up the first time. |
| 118 | * Also for supporting drivers that do not implement this |
| 119 | * feature. They will always have to call sendto(). |
| 120 | */ |
| 121 | xsk_set_tx_need_wakeup(umem); |
| 122 | } |
| 123 | |
| 124 | dev_hold(dev); |
| 125 | |
| 126 | if (force_copy) |
| 127 | /* For copy-mode, we are done. */ |
| 128 | return 0; |
| 129 | |
| 130 | if (!dev->netdev_ops->ndo_bpf || !dev->netdev_ops->ndo_xsk_wakeup) { |
| 131 | err = -EOPNOTSUPP; |
| 132 | goto err_unreg_umem; |
| 133 | } |
| 134 | |
| 135 | bpf.command = XDP_SETUP_XSK_UMEM; |
| 136 | bpf.xsk.umem = umem; |
| 137 | bpf.xsk.queue_id = queue_id; |
| 138 | |
| 139 | err = dev->netdev_ops->ndo_bpf(dev, &bpf); |
| 140 | if (err) |
| 141 | goto err_unreg_umem; |
| 142 | |
| 143 | umem->zc = true; |
| 144 | return 0; |
| 145 | |
| 146 | err_unreg_umem: |
| 147 | if (!force_zc) |
| 148 | err = 0; /* fallback to copy mode */ |
| 149 | if (err) |
| 150 | xdp_clear_umem_at_qid(dev, queue_id); |
| 151 | return err; |
| 152 | } |
| 153 | |
| 154 | void xdp_umem_clear_dev(struct xdp_umem *umem) |
| 155 | { |
| 156 | struct netdev_bpf bpf; |
| 157 | int err; |
| 158 | |
| 159 | ASSERT_RTNL(); |
| 160 | |
| 161 | if (!umem->dev) |
| 162 | return; |
| 163 | |
| 164 | if (umem->zc) { |
| 165 | bpf.command = XDP_SETUP_XSK_UMEM; |
| 166 | bpf.xsk.umem = NULL; |
| 167 | bpf.xsk.queue_id = umem->queue_id; |
| 168 | |
| 169 | err = umem->dev->netdev_ops->ndo_bpf(umem->dev, &bpf); |
| 170 | |
| 171 | if (err) |
| 172 | WARN(1, "failed to disable umem!\n"); |
| 173 | } |
| 174 | |
| 175 | xdp_clear_umem_at_qid(umem->dev, umem->queue_id); |
| 176 | |
| 177 | dev_put(umem->dev); |
| 178 | umem->dev = NULL; |
| 179 | umem->zc = false; |
| 180 | } |
| 181 | |
| 182 | static void xdp_umem_unpin_pages(struct xdp_umem *umem) |
| 183 | { |
| 184 | unpin_user_pages_dirty_lock(umem->pgs, umem->npgs, true); |
| 185 | |
| 186 | kfree(umem->pgs); |
| 187 | umem->pgs = NULL; |
| 188 | } |
| 189 | |
| 190 | static void xdp_umem_unaccount_pages(struct xdp_umem *umem) |
| 191 | { |
| 192 | if (umem->user) { |
| 193 | atomic_long_sub(umem->npgs, &umem->user->locked_vm); |
| 194 | free_uid(umem->user); |
| 195 | } |
| 196 | } |
| 197 | |
| 198 | static void xdp_umem_release(struct xdp_umem *umem) |
| 199 | { |
| 200 | rtnl_lock(); |
| 201 | xdp_umem_clear_dev(umem); |
| 202 | rtnl_unlock(); |
| 203 | |
| 204 | ida_simple_remove(&umem_ida, umem->id); |
| 205 | |
| 206 | if (umem->fq) { |
| 207 | xskq_destroy(umem->fq); |
| 208 | umem->fq = NULL; |
| 209 | } |
| 210 | |
| 211 | if (umem->cq) { |
| 212 | xskq_destroy(umem->cq); |
| 213 | umem->cq = NULL; |
| 214 | } |
| 215 | |
| 216 | xp_destroy(umem->pool); |
| 217 | xdp_umem_unpin_pages(umem); |
| 218 | |
| 219 | xdp_umem_unaccount_pages(umem); |
| 220 | kfree(umem); |
| 221 | } |
| 222 | |
| 223 | static void xdp_umem_release_deferred(struct work_struct *work) |
| 224 | { |
| 225 | struct xdp_umem *umem = container_of(work, struct xdp_umem, work); |
| 226 | |
| 227 | xdp_umem_release(umem); |
| 228 | } |
| 229 | |
| 230 | void xdp_get_umem(struct xdp_umem *umem) |
| 231 | { |
| 232 | refcount_inc(&umem->users); |
| 233 | } |
| 234 | |
| 235 | void xdp_put_umem(struct xdp_umem *umem) |
| 236 | { |
| 237 | if (!umem) |
| 238 | return; |
| 239 | |
| 240 | if (refcount_dec_and_test(&umem->users)) { |
| 241 | INIT_WORK(&umem->work, xdp_umem_release_deferred); |
| 242 | schedule_work(&umem->work); |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | static int xdp_umem_pin_pages(struct xdp_umem *umem, unsigned long address) |
| 247 | { |
| 248 | unsigned int gup_flags = FOLL_WRITE; |
| 249 | long npgs; |
| 250 | int err; |
| 251 | |
| 252 | umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs), |
| 253 | GFP_KERNEL | __GFP_NOWARN); |
| 254 | if (!umem->pgs) |
| 255 | return -ENOMEM; |
| 256 | |
| 257 | mmap_read_lock(current->mm); |
| 258 | npgs = pin_user_pages(address, umem->npgs, |
| 259 | gup_flags | FOLL_LONGTERM, &umem->pgs[0], NULL); |
| 260 | mmap_read_unlock(current->mm); |
| 261 | |
| 262 | if (npgs != umem->npgs) { |
| 263 | if (npgs >= 0) { |
| 264 | umem->npgs = npgs; |
| 265 | err = -ENOMEM; |
| 266 | goto out_pin; |
| 267 | } |
| 268 | err = npgs; |
| 269 | goto out_pgs; |
| 270 | } |
| 271 | return 0; |
| 272 | |
| 273 | out_pin: |
| 274 | xdp_umem_unpin_pages(umem); |
| 275 | out_pgs: |
| 276 | kfree(umem->pgs); |
| 277 | umem->pgs = NULL; |
| 278 | return err; |
| 279 | } |
| 280 | |
| 281 | static int xdp_umem_account_pages(struct xdp_umem *umem) |
| 282 | { |
| 283 | unsigned long lock_limit, new_npgs, old_npgs; |
| 284 | |
| 285 | if (capable(CAP_IPC_LOCK)) |
| 286 | return 0; |
| 287 | |
| 288 | lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
| 289 | umem->user = get_uid(current_user()); |
| 290 | |
| 291 | do { |
| 292 | old_npgs = atomic_long_read(&umem->user->locked_vm); |
| 293 | new_npgs = old_npgs + umem->npgs; |
| 294 | if (new_npgs > lock_limit) { |
| 295 | free_uid(umem->user); |
| 296 | umem->user = NULL; |
| 297 | return -ENOBUFS; |
| 298 | } |
| 299 | } while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs, |
| 300 | new_npgs) != old_npgs); |
| 301 | return 0; |
| 302 | } |
| 303 | |
| 304 | static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr) |
| 305 | { |
| 306 | bool unaligned_chunks = mr->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG; |
| 307 | u32 chunk_size = mr->chunk_size, headroom = mr->headroom; |
| 308 | u64 npgs, addr = mr->addr, size = mr->len; |
| 309 | unsigned int chunks, chunks_per_page; |
| 310 | int err; |
| 311 | |
| 312 | if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) { |
| 313 | /* Strictly speaking we could support this, if: |
| 314 | * - huge pages, or* |
| 315 | * - using an IOMMU, or |
| 316 | * - making sure the memory area is consecutive |
| 317 | * but for now, we simply say "computer says no". |
| 318 | */ |
| 319 | return -EINVAL; |
| 320 | } |
| 321 | |
| 322 | if (mr->flags & ~(XDP_UMEM_UNALIGNED_CHUNK_FLAG | |
| 323 | XDP_UMEM_USES_NEED_WAKEUP)) |
| 324 | return -EINVAL; |
| 325 | |
| 326 | if (!unaligned_chunks && !is_power_of_2(chunk_size)) |
| 327 | return -EINVAL; |
| 328 | |
| 329 | if (!PAGE_ALIGNED(addr)) { |
| 330 | /* Memory area has to be page size aligned. For |
| 331 | * simplicity, this might change. |
| 332 | */ |
| 333 | return -EINVAL; |
| 334 | } |
| 335 | |
| 336 | if ((addr + size) < addr) |
| 337 | return -EINVAL; |
| 338 | |
| 339 | npgs = size >> PAGE_SHIFT; |
| 340 | if (npgs > U32_MAX) |
| 341 | return -EINVAL; |
| 342 | |
| 343 | chunks = (unsigned int)div_u64(size, chunk_size); |
| 344 | if (chunks == 0) |
| 345 | return -EINVAL; |
| 346 | |
| 347 | if (!unaligned_chunks) { |
| 348 | chunks_per_page = PAGE_SIZE / chunk_size; |
| 349 | if (chunks < chunks_per_page || chunks % chunks_per_page) |
| 350 | return -EINVAL; |
| 351 | } |
| 352 | |
| 353 | if (headroom >= chunk_size - XDP_PACKET_HEADROOM) |
| 354 | return -EINVAL; |
| 355 | |
| 356 | umem->size = size; |
| 357 | umem->headroom = headroom; |
| 358 | umem->chunk_size = chunk_size; |
| 359 | umem->npgs = (u32)npgs; |
| 360 | umem->pgs = NULL; |
| 361 | umem->user = NULL; |
| 362 | umem->flags = mr->flags; |
| 363 | INIT_LIST_HEAD(&umem->xsk_tx_list); |
| 364 | spin_lock_init(&umem->xsk_tx_list_lock); |
| 365 | |
| 366 | refcount_set(&umem->users, 1); |
| 367 | |
| 368 | err = xdp_umem_account_pages(umem); |
| 369 | if (err) |
| 370 | return err; |
| 371 | |
| 372 | err = xdp_umem_pin_pages(umem, (unsigned long)addr); |
| 373 | if (err) |
| 374 | goto out_account; |
| 375 | |
| 376 | umem->pool = xp_create(umem->pgs, umem->npgs, chunks, chunk_size, |
| 377 | headroom, size, unaligned_chunks); |
| 378 | if (!umem->pool) { |
| 379 | err = -ENOMEM; |
| 380 | goto out_pin; |
| 381 | } |
| 382 | return 0; |
| 383 | |
| 384 | out_pin: |
| 385 | xdp_umem_unpin_pages(umem); |
| 386 | out_account: |
| 387 | xdp_umem_unaccount_pages(umem); |
| 388 | return err; |
| 389 | } |
| 390 | |
| 391 | struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr) |
| 392 | { |
| 393 | struct xdp_umem *umem; |
| 394 | int err; |
| 395 | |
| 396 | umem = kzalloc(sizeof(*umem), GFP_KERNEL); |
| 397 | if (!umem) |
| 398 | return ERR_PTR(-ENOMEM); |
| 399 | |
| 400 | err = ida_simple_get(&umem_ida, 0, 0, GFP_KERNEL); |
| 401 | if (err < 0) { |
| 402 | kfree(umem); |
| 403 | return ERR_PTR(err); |
| 404 | } |
| 405 | umem->id = err; |
| 406 | |
| 407 | err = xdp_umem_reg(umem, mr); |
| 408 | if (err) { |
| 409 | ida_simple_remove(&umem_ida, umem->id); |
| 410 | kfree(umem); |
| 411 | return ERR_PTR(err); |
| 412 | } |
| 413 | |
| 414 | return umem; |
| 415 | } |
| 416 | |
| 417 | bool xdp_umem_validate_queues(struct xdp_umem *umem) |
| 418 | { |
| 419 | return umem->fq && umem->cq; |
| 420 | } |