net/xdp/xdp_umem.c

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