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 if (umem->user) { 136 atomic_long_sub(umem->npgs, &umem->user->locked_vm); 137 free_uid(umem->user); 138 } 139 } 140 141 static void xdp_umem_release(struct xdp_umem *umem) 142 { 143 struct task_struct *task; 144 struct mm_struct *mm; 145 146 xdp_umem_clear_dev(umem); 147 148 if (umem->fq) { 149 xskq_destroy(umem->fq); 150 umem->fq = NULL; 151 } 152 153 if (umem->cq) { 154 xskq_destroy(umem->cq); 155 umem->cq = NULL; 156 } 157 158 xdp_umem_unpin_pages(umem); 159 160 task = get_pid_task(umem->pid, PIDTYPE_PID); 161 put_pid(umem->pid); 162 if (!task) 163 goto out; 164 mm = get_task_mm(task); 165 put_task_struct(task); 166 if (!mm) 167 goto out; 168 169 mmput(mm); 170 kfree(umem->pages); 171 umem->pages = NULL; 172 173 xdp_umem_unaccount_pages(umem); 174 out: 175 kfree(umem); 176 } 177 178 static void xdp_umem_release_deferred(struct work_struct *work) 179 { 180 struct xdp_umem *umem = container_of(work, struct xdp_umem, work); 181 182 xdp_umem_release(umem); 183 } 184 185 void xdp_get_umem(struct xdp_umem *umem) 186 { 187 refcount_inc(&umem->users); 188 } 189 190 void xdp_put_umem(struct xdp_umem *umem) 191 { 192 if (!umem) 193 return; 194 195 if (refcount_dec_and_test(&umem->users)) { 196 INIT_WORK(&umem->work, xdp_umem_release_deferred); 197 schedule_work(&umem->work); 198 } 199 } 200 201 static int xdp_umem_pin_pages(struct xdp_umem *umem) 202 { 203 unsigned int gup_flags = FOLL_WRITE; 204 long npgs; 205 int err; 206 207 umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs), 208 GFP_KERNEL | __GFP_NOWARN); 209 if (!umem->pgs) 210 return -ENOMEM; 211 212 down_write(¤t->mm->mmap_sem); 213 npgs = get_user_pages(umem->address, umem->npgs, 214 gup_flags, &umem->pgs[0], NULL); 215 up_write(¤t->mm->mmap_sem); 216 217 if (npgs != umem->npgs) { 218 if (npgs >= 0) { 219 umem->npgs = npgs; 220 err = -ENOMEM; 221 goto out_pin; 222 } 223 err = npgs; 224 goto out_pgs; 225 } 226 return 0; 227 228 out_pin: 229 xdp_umem_unpin_pages(umem); 230 out_pgs: 231 kfree(umem->pgs); 232 umem->pgs = NULL; 233 return err; 234 } 235 236 static int xdp_umem_account_pages(struct xdp_umem *umem) 237 { 238 unsigned long lock_limit, new_npgs, old_npgs; 239 240 if (capable(CAP_IPC_LOCK)) 241 return 0; 242 243 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 244 umem->user = get_uid(current_user()); 245 246 do { 247 old_npgs = atomic_long_read(&umem->user->locked_vm); 248 new_npgs = old_npgs + umem->npgs; 249 if (new_npgs > lock_limit) { 250 free_uid(umem->user); 251 umem->user = NULL; 252 return -ENOBUFS; 253 } 254 } while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs, 255 new_npgs) != old_npgs); 256 return 0; 257 } 258 259 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr) 260 { 261 u32 chunk_size = mr->chunk_size, headroom = mr->headroom; 262 unsigned int chunks, chunks_per_page; 263 u64 addr = mr->addr, size = mr->len; 264 int size_chk, err, i; 265 266 if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) { 267 /* Strictly speaking we could support this, if: 268 * - huge pages, or* 269 * - using an IOMMU, or 270 * - making sure the memory area is consecutive 271 * but for now, we simply say "computer says no". 272 */ 273 return -EINVAL; 274 } 275 276 if (!is_power_of_2(chunk_size)) 277 return -EINVAL; 278 279 if (!PAGE_ALIGNED(addr)) { 280 /* Memory area has to be page size aligned. For 281 * simplicity, this might change. 282 */ 283 return -EINVAL; 284 } 285 286 if ((addr + size) < addr) 287 return -EINVAL; 288 289 chunks = (unsigned int)div_u64(size, chunk_size); 290 if (chunks == 0) 291 return -EINVAL; 292 293 chunks_per_page = PAGE_SIZE / chunk_size; 294 if (chunks < chunks_per_page || chunks % chunks_per_page) 295 return -EINVAL; 296 297 headroom = ALIGN(headroom, 64); 298 299 size_chk = chunk_size - headroom - XDP_PACKET_HEADROOM; 300 if (size_chk < 0) 301 return -EINVAL; 302 303 umem->pid = get_task_pid(current, PIDTYPE_PID); 304 umem->address = (unsigned long)addr; 305 umem->props.chunk_mask = ~((u64)chunk_size - 1); 306 umem->props.size = size; 307 umem->headroom = headroom; 308 umem->chunk_size_nohr = chunk_size - headroom; 309 umem->npgs = size / PAGE_SIZE; 310 umem->pgs = NULL; 311 umem->user = NULL; 312 INIT_LIST_HEAD(&umem->xsk_list); 313 spin_lock_init(&umem->xsk_list_lock); 314 315 refcount_set(&umem->users, 1); 316 317 err = xdp_umem_account_pages(umem); 318 if (err) 319 goto out; 320 321 err = xdp_umem_pin_pages(umem); 322 if (err) 323 goto out_account; 324 325 umem->pages = kcalloc(umem->npgs, sizeof(*umem->pages), GFP_KERNEL); 326 if (!umem->pages) { 327 err = -ENOMEM; 328 goto out_account; 329 } 330 331 for (i = 0; i < umem->npgs; i++) 332 umem->pages[i].addr = page_address(umem->pgs[i]); 333 334 return 0; 335 336 out_account: 337 xdp_umem_unaccount_pages(umem); 338 out: 339 put_pid(umem->pid); 340 return err; 341 } 342 343 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr) 344 { 345 struct xdp_umem *umem; 346 int err; 347 348 umem = kzalloc(sizeof(*umem), GFP_KERNEL); 349 if (!umem) 350 return ERR_PTR(-ENOMEM); 351 352 err = xdp_umem_reg(umem, mr); 353 if (err) { 354 kfree(umem); 355 return ERR_PTR(err); 356 } 357 358 return umem; 359 } 360 361 bool xdp_umem_validate_queues(struct xdp_umem *umem) 362 { 363 return umem->fq && umem->cq; 364 } 365