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 #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 static void xdp_umem_unpin_pages(struct xdp_umem *umem)
  27 {
  28         unpin_user_pages_dirty_lock(umem->pgs, umem->npgs, true);
  29
  30         kfree(umem->pgs);
  31         umem->pgs = NULL;
  32 }
  33
  34 static void xdp_umem_unaccount_pages(struct xdp_umem *umem)
  35 {
  36         if (umem->user) {
  37                 atomic_long_sub(umem->npgs, &umem->user->locked_vm);
  38                 free_uid(umem->user);
  39         }
  40 }
  41
  42 static void xdp_umem_addr_unmap(struct xdp_umem *umem)
  43 {
  44         vunmap(umem->addrs);
  45         umem->addrs = NULL;
  46 }
  47
  48 static int xdp_umem_addr_map(struct xdp_umem *umem, struct page **pages,
  49                              u32 nr_pages)
  50 {
  51         umem->addrs = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
  52         if (!umem->addrs)
  53                 return -ENOMEM;
  54         return 0;
  55 }
  56
  57 static void xdp_umem_release(struct xdp_umem *umem)
  58 {
  59         umem->zc = false;
  60         ida_simple_remove(&umem_ida, umem->id);
  61
  62         xdp_umem_addr_unmap(umem);
  63         xdp_umem_unpin_pages(umem);
  64
  65         xdp_umem_unaccount_pages(umem);
  66         kfree(umem);
  67 }
  68
  69 static void xdp_umem_release_deferred(struct work_struct *work)
  70 {
  71         struct xdp_umem *umem = container_of(work, struct xdp_umem, work);
  72
  73         xdp_umem_release(umem);
  74 }
  75
  76 void xdp_get_umem(struct xdp_umem *umem)
  77 {
  78         refcount_inc(&umem->users);
  79 }
  80
  81 void xdp_put_umem(struct xdp_umem *umem, bool defer_cleanup)
  82 {
  83         if (!umem)
  84                 return;
  85
  86         if (refcount_dec_and_test(&umem->users)) {
  87                 if (defer_cleanup) {
  88                         INIT_WORK(&umem->work, xdp_umem_release_deferred);
  89                         schedule_work(&umem->work);
  90                 } else {
  91                         xdp_umem_release(umem);
  92                 }
  93         }
  94 }
  95
  96 static int xdp_umem_pin_pages(struct xdp_umem *umem, unsigned long address)
  97 {
  98         unsigned int gup_flags = FOLL_WRITE;
  99         long npgs;
 100         int err;
 101
 102         umem->pgs = kcalloc(umem->npgs, sizeof(*umem->pgs),
 103                             GFP_KERNEL | __GFP_NOWARN);
 104         if (!umem->pgs)
 105                 return -ENOMEM;
 106
 107         mmap_read_lock(current->mm);
 108         npgs = pin_user_pages(address, umem->npgs,
 109                               gup_flags | FOLL_LONGTERM, &umem->pgs[0], NULL);
 110         mmap_read_unlock(current->mm);
 111
 112         if (npgs != umem->npgs) {
 113                 if (npgs >= 0) {
 114                         umem->npgs = npgs;
 115                         err = -ENOMEM;
 116                         goto out_pin;
 117                 }
 118                 err = npgs;
 119                 goto out_pgs;
 120         }
 121         return 0;
 122
 123 out_pin:
 124         xdp_umem_unpin_pages(umem);
 125 out_pgs:
 126         kfree(umem->pgs);
 127         umem->pgs = NULL;
 128         return err;
 129 }
 130
 131 static int xdp_umem_account_pages(struct xdp_umem *umem)
 132 {
 133         unsigned long lock_limit, new_npgs, old_npgs;
 134
 135         if (capable(CAP_IPC_LOCK))
 136                 return 0;
 137
 138         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 139         umem->user = get_uid(current_user());
 140
 141         do {
 142                 old_npgs = atomic_long_read(&umem->user->locked_vm);
 143                 new_npgs = old_npgs + umem->npgs;
 144                 if (new_npgs > lock_limit) {
 145                         free_uid(umem->user);
 146                         umem->user = NULL;
 147                         return -ENOBUFS;
 148                 }
 149         } while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
 150                                      new_npgs) != old_npgs);
 151         return 0;
 152 }
 153
 154 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
 155 {
 156         u32 npgs_rem, chunk_size = mr->chunk_size, headroom = mr->headroom;
 157         bool unaligned_chunks = mr->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
 158         u64 npgs, addr = mr->addr, size = mr->len;
 159         unsigned int chunks, chunks_rem;
 160         int err;
 161
 162         if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
 163                 /* Strictly speaking we could support this, if:
 164                  * - huge pages, or*
 165                  * - using an IOMMU, or
 166                  * - making sure the memory area is consecutive
 167                  * but for now, we simply say "computer says no".
 168                  */
 169                 return -EINVAL;
 170         }
 171
 172         if (mr->flags & ~XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 173                 return -EINVAL;
 174
 175         if (!unaligned_chunks && !is_power_of_2(chunk_size))
 176                 return -EINVAL;
 177
 178         if (!PAGE_ALIGNED(addr)) {
 179                 /* Memory area has to be page size aligned. For
 180                  * simplicity, this might change.
 181                  */
 182                 return -EINVAL;
 183         }
 184
 185         if ((addr + size) < addr)
 186                 return -EINVAL;
 187
 188         npgs = div_u64_rem(size, PAGE_SIZE, &npgs_rem);
 189         if (npgs_rem)
 190                 npgs++;
 191         if (npgs > U32_MAX)
 192                 return -EINVAL;
 193
 194         chunks = (unsigned int)div_u64_rem(size, chunk_size, &chunks_rem);
 195         if (chunks == 0)
 196                 return -EINVAL;
 197
 198         if (!unaligned_chunks && chunks_rem)
 199                 return -EINVAL;
 200
 201         if (headroom >= chunk_size - XDP_PACKET_HEADROOM)
 202                 return -EINVAL;
 203
 204         umem->size = size;
 205         umem->headroom = headroom;
 206         umem->chunk_size = chunk_size;
 207         umem->chunks = chunks;
 208         umem->npgs = (u32)npgs;
 209         umem->pgs = NULL;
 210         umem->user = NULL;
 211         umem->flags = mr->flags;
 212
 213         INIT_LIST_HEAD(&umem->xsk_dma_list);
 214         refcount_set(&umem->users, 1);
 215
 216         err = xdp_umem_account_pages(umem);
 217         if (err)
 218                 return err;
 219
 220         err = xdp_umem_pin_pages(umem, (unsigned long)addr);
 221         if (err)
 222                 goto out_account;
 223
 224         err = xdp_umem_addr_map(umem, umem->pgs, umem->npgs);
 225         if (err)
 226                 goto out_unpin;
 227
 228         return 0;
 229
 230 out_unpin:
 231         xdp_umem_unpin_pages(umem);
 232 out_account:
 233         xdp_umem_unaccount_pages(umem);
 234         return err;
 235 }
 236
 237 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr)
 238 {
 239         struct xdp_umem *umem;
 240         int err;
 241
 242         umem = kzalloc(sizeof(*umem), GFP_KERNEL);
 243         if (!umem)
 244                 return ERR_PTR(-ENOMEM);
 245
 246         err = ida_simple_get(&umem_ida, 0, 0, GFP_KERNEL);
 247         if (err < 0) {
 248                 kfree(umem);
 249                 return ERR_PTR(err);
 250         }
 251         umem->id = err;
 252
 253         err = xdp_umem_reg(umem, mr);
 254         if (err) {
 255                 ida_simple_remove(&umem_ida, umem->id);
 256                 kfree(umem);
 257                 return ERR_PTR(err);
 258         }
 259
 260         return umem;
 261 }