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 17 #include "xdp_umem.h" 18 #include "xsk_queue.h" 19 20 #define XDP_UMEM_MIN_CHUNK_SIZE 2048 21 22 void xdp_add_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs) 23 { 24 unsigned long flags; 25 26 spin_lock_irqsave(&umem->xsk_list_lock, flags); 27 list_add_rcu(&xs->list, &umem->xsk_list); 28 spin_unlock_irqrestore(&umem->xsk_list_lock, flags); 29 } 30 31 void xdp_del_sk_umem(struct xdp_umem *umem, struct xdp_sock *xs) 32 { 33 unsigned long flags; 34 35 if (xs->dev) { 36 spin_lock_irqsave(&umem->xsk_list_lock, flags); 37 list_del_rcu(&xs->list); 38 spin_unlock_irqrestore(&umem->xsk_list_lock, flags); 39 40 if (umem->zc) 41 synchronize_net(); 42 } 43 } 44 45 int xdp_umem_query(struct net_device *dev, u16 queue_id) 46 { 47 struct netdev_bpf bpf; 48 49 ASSERT_RTNL(); 50 51 memset(&bpf, 0, sizeof(bpf)); 52 bpf.command = XDP_QUERY_XSK_UMEM; 53 bpf.xsk.queue_id = queue_id; 54 55 if (!dev->netdev_ops->ndo_bpf) 56 return 0; 57 return dev->netdev_ops->ndo_bpf(dev, &bpf) ?: !!bpf.xsk.umem; 58 } 59 60 int xdp_umem_assign_dev(struct xdp_umem *umem, struct net_device *dev, 61 u32 queue_id, u16 flags) 62 { 63 bool force_zc, force_copy; 64 struct netdev_bpf bpf; 65 int err; 66 67 force_zc = flags & XDP_ZEROCOPY; 68 force_copy = flags & XDP_COPY; 69 70 if (force_zc && force_copy) 71 return -EINVAL; 72 73 if (force_copy) 74 return 0; 75 76 if (!dev->netdev_ops->ndo_bpf || !dev->netdev_ops->ndo_xsk_async_xmit) 77 return force_zc ? -EOPNOTSUPP : 0; /* fail or fallback */ 78 79 rtnl_lock(); 80 err = xdp_umem_query(dev, queue_id); 81 if (err) { 82 err = err < 0 ? -EOPNOTSUPP : -EBUSY; 83 goto err_rtnl_unlock; 84 } 85 86 bpf.command = XDP_SETUP_XSK_UMEM; 87 bpf.xsk.umem = umem; 88 bpf.xsk.queue_id = queue_id; 89 90 err = dev->netdev_ops->ndo_bpf(dev, &bpf); 91 if (err) 92 goto err_rtnl_unlock; 93 rtnl_unlock(); 94 95 dev_hold(dev); 96 umem->dev = dev; 97 umem->queue_id = queue_id; 98 umem->zc = true; 99 return 0; 100 101 err_rtnl_unlock: 102 rtnl_unlock(); 103 return force_zc ? err : 0; /* fail or fallback */ 104 } 105 106 static void xdp_umem_clear_dev(struct xdp_umem *umem) 107 { 108 struct netdev_bpf bpf; 109 int err; 110 111 if (umem->dev) { 112 bpf.command = XDP_SETUP_XSK_UMEM; 113 bpf.xsk.umem = NULL; 114 bpf.xsk.queue_id = umem->queue_id; 115 116 rtnl_lock(); 117 err = umem->dev->netdev_ops->ndo_bpf(umem->dev, &bpf); 118 rtnl_unlock(); 119 120 if (err) 121 WARN(1, "failed to disable umem!\n"); 122 123 dev_put(umem->dev); 124 umem->dev = NULL; 125 } 126 } 127 128 static void xdp_umem_unpin_pages(struct xdp_umem *umem) 129 { 130 unsigned int i; 131 132 for (i = 0; i < umem->npgs; i++) { 133 struct page *page = umem->pgs[i]; 134 135 set_page_dirty_lock(page); 136 put_page(page); 137 } 138 139 kfree(umem->pgs); 140 umem->pgs = NULL; 141 } 142 143 static void xdp_umem_unaccount_pages(struct xdp_umem *umem) 144 { 145 if (umem->user) { 146 atomic_long_sub(umem->npgs, &umem->user->locked_vm); 147 free_uid(umem->user); 148 } 149 } 150 151 static void xdp_umem_release(struct xdp_umem *umem) 152 { 153 struct task_struct *task; 154 struct mm_struct *mm; 155 156 xdp_umem_clear_dev(umem); 157 158 if (umem->fq) { 159 xskq_destroy(umem->fq); 160 umem->fq = NULL; 161 } 162 163 if (umem->cq) { 164 xskq_destroy(umem->cq); 165 umem->cq = NULL; 166 } 167 168 xdp_umem_unpin_pages(umem); 169 170 task = get_pid_task(umem->pid, PIDTYPE_PID); 171 put_pid(umem->pid); 172 if (!task) 173 goto out; 174 mm = get_task_mm(task); 175 put_task_struct(task); 176 if (!mm) 177 goto out; 178 179 mmput(mm); 180 kfree(umem->pages); 181 umem->pages = NULL; 182 183 xdp_umem_unaccount_pages(umem); 184 out: 185 kfree(umem); 186 } 187 188 static void xdp_umem_release_deferred(struct work_struct *work) 189 { 190 struct xdp_umem *umem = container_of(work, struct xdp_umem, work); 191 192 xdp_umem_release(umem); 193 } 194 195 void xdp_get_umem(struct xdp_umem *umem) 196 { 197 refcount_inc(&umem->users); 198 } 199 200 void xdp_put_umem(struct xdp_umem *umem) 201 { 202 if (!umem) 203 return; 204 205 if (refcount_dec_and_test(&umem->users)) { 206 INIT_WORK(&umem->work, xdp_umem_release_deferred); 207 schedule_work(&umem->work); 208 } 209 } 210 211 static int xdp_umem_pin_pages(struct xdp_umem *umem) 212 { 213 unsigned int gup_flags = FOLL_WRITE; 214 long npgs; 215 int err; 216 217 umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs), 218 GFP_KERNEL | __GFP_NOWARN); 219 if (!umem->pgs) 220 return -ENOMEM; 221 222 down_write(¤t->mm->mmap_sem); 223 npgs = get_user_pages(umem->address, umem->npgs, 224 gup_flags, &umem->pgs[0], NULL); 225 up_write(¤t->mm->mmap_sem); 226 227 if (npgs != umem->npgs) { 228 if (npgs >= 0) { 229 umem->npgs = npgs; 230 err = -ENOMEM; 231 goto out_pin; 232 } 233 err = npgs; 234 goto out_pgs; 235 } 236 return 0; 237 238 out_pin: 239 xdp_umem_unpin_pages(umem); 240 out_pgs: 241 kfree(umem->pgs); 242 umem->pgs = NULL; 243 return err; 244 } 245 246 static int xdp_umem_account_pages(struct xdp_umem *umem) 247 { 248 unsigned long lock_limit, new_npgs, old_npgs; 249 250 if (capable(CAP_IPC_LOCK)) 251 return 0; 252 253 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 254 umem->user = get_uid(current_user()); 255 256 do { 257 old_npgs = atomic_long_read(&umem->user->locked_vm); 258 new_npgs = old_npgs + umem->npgs; 259 if (new_npgs > lock_limit) { 260 free_uid(umem->user); 261 umem->user = NULL; 262 return -ENOBUFS; 263 } 264 } while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs, 265 new_npgs) != old_npgs); 266 return 0; 267 } 268 269 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr) 270 { 271 u32 chunk_size = mr->chunk_size, headroom = mr->headroom; 272 unsigned int chunks, chunks_per_page; 273 u64 addr = mr->addr, size = mr->len; 274 int size_chk, err, i; 275 276 if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) { 277 /* Strictly speaking we could support this, if: 278 * - huge pages, or* 279 * - using an IOMMU, or 280 * - making sure the memory area is consecutive 281 * but for now, we simply say "computer says no". 282 */ 283 return -EINVAL; 284 } 285 286 if (!is_power_of_2(chunk_size)) 287 return -EINVAL; 288 289 if (!PAGE_ALIGNED(addr)) { 290 /* Memory area has to be page size aligned. For 291 * simplicity, this might change. 292 */ 293 return -EINVAL; 294 } 295 296 if ((addr + size) < addr) 297 return -EINVAL; 298 299 chunks = (unsigned int)div_u64(size, chunk_size); 300 if (chunks == 0) 301 return -EINVAL; 302 303 chunks_per_page = PAGE_SIZE / chunk_size; 304 if (chunks < chunks_per_page || chunks % chunks_per_page) 305 return -EINVAL; 306 307 headroom = ALIGN(headroom, 64); 308 309 size_chk = chunk_size - headroom - XDP_PACKET_HEADROOM; 310 if (size_chk < 0) 311 return -EINVAL; 312 313 umem->pid = get_task_pid(current, PIDTYPE_PID); 314 umem->address = (unsigned long)addr; 315 umem->chunk_mask = ~((u64)chunk_size - 1); 316 umem->size = size; 317 umem->headroom = headroom; 318 umem->chunk_size_nohr = chunk_size - headroom; 319 umem->npgs = size / PAGE_SIZE; 320 umem->pgs = NULL; 321 umem->user = NULL; 322 INIT_LIST_HEAD(&umem->xsk_list); 323 spin_lock_init(&umem->xsk_list_lock); 324 325 refcount_set(&umem->users, 1); 326 327 err = xdp_umem_account_pages(umem); 328 if (err) 329 goto out; 330 331 err = xdp_umem_pin_pages(umem); 332 if (err) 333 goto out_account; 334 335 umem->pages = kcalloc(umem->npgs, sizeof(*umem->pages), GFP_KERNEL); 336 if (!umem->pages) { 337 err = -ENOMEM; 338 goto out_account; 339 } 340 341 for (i = 0; i < umem->npgs; i++) 342 umem->pages[i].addr = page_address(umem->pgs[i]); 343 344 return 0; 345 346 out_account: 347 xdp_umem_unaccount_pages(umem); 348 out: 349 put_pid(umem->pid); 350 return err; 351 } 352 353 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr) 354 { 355 struct xdp_umem *umem; 356 int err; 357 358 umem = kzalloc(sizeof(*umem), GFP_KERNEL); 359 if (!umem) 360 return ERR_PTR(-ENOMEM); 361 362 err = xdp_umem_reg(umem, mr); 363 if (err) { 364 kfree(umem); 365 return ERR_PTR(err); 366 } 367 368 return umem; 369 } 370 371 bool xdp_umem_validate_queues(struct xdp_umem *umem) 372 { 373 return umem->fq && umem->cq; 374 } 375