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         kvfree(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 = kvcalloc(umem->npgs, sizeof(*umem->pgs), GFP_KERNEL | __GFP_NOWARN);
 103         if (!umem->pgs)
 104                 return -ENOMEM;
 105
 106         mmap_read_lock(current->mm);
 107         npgs = pin_user_pages(address, umem->npgs,
 108                               gup_flags | FOLL_LONGTERM, &umem->pgs[0], NULL);
 109         mmap_read_unlock(current->mm);
 110
 111         if (npgs != umem->npgs) {
 112                 if (npgs >= 0) {
 113                         umem->npgs = npgs;
 114                         err = -ENOMEM;
 115                         goto out_pin;
 116                 }
 117                 err = npgs;
 118                 goto out_pgs;
 119         }
 120         return 0;
 121
 122 out_pin:
 123         xdp_umem_unpin_pages(umem);
 124 out_pgs:
 125         kvfree(umem->pgs);
 126         umem->pgs = NULL;
 127         return err;
 128 }
 129
 130 static int xdp_umem_account_pages(struct xdp_umem *umem)
 131 {
 132         unsigned long lock_limit, new_npgs, old_npgs;
 133
 134         if (capable(CAP_IPC_LOCK))
 135                 return 0;
 136
 137         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 138         umem->user = get_uid(current_user());
 139
 140         do {
 141                 old_npgs = atomic_long_read(&umem->user->locked_vm);
 142                 new_npgs = old_npgs + umem->npgs;
 143                 if (new_npgs > lock_limit) {
 144                         free_uid(umem->user);
 145                         umem->user = NULL;
 146                         return -ENOBUFS;
 147                 }
 148         } while (atomic_long_cmpxchg(&umem->user->locked_vm, old_npgs,
 149                                      new_npgs) != old_npgs);
 150         return 0;
 151 }
 152
 153 static int xdp_umem_reg(struct xdp_umem *umem, struct xdp_umem_reg *mr)
 154 {
 155         u32 npgs_rem, chunk_size = mr->chunk_size, headroom = mr->headroom;
 156         bool unaligned_chunks = mr->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
 157         u64 npgs, addr = mr->addr, size = mr->len;
 158         unsigned int chunks, chunks_rem;
 159         int err;
 160
 161         if (chunk_size < XDP_UMEM_MIN_CHUNK_SIZE || chunk_size > PAGE_SIZE) {
 162                 /* Strictly speaking we could support this, if:
 163                  * - huge pages, or*
 164                  * - using an IOMMU, or
 165                  * - making sure the memory area is consecutive
 166                  * but for now, we simply say "computer says no".
 167                  */
 168                 return -EINVAL;
 169         }
 170
 171         if (mr->flags & ~XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 172                 return -EINVAL;
 173
 174         if (!unaligned_chunks && !is_power_of_2(chunk_size))
 175                 return -EINVAL;
 176
 177         if (!PAGE_ALIGNED(addr)) {
 178                 /* Memory area has to be page size aligned. For
 179                  * simplicity, this might change.
 180                  */
 181                 return -EINVAL;
 182         }
 183
 184         if ((addr + size) < addr)
 185                 return -EINVAL;
 186
 187         npgs = div_u64_rem(size, PAGE_SIZE, &npgs_rem);
 188         if (npgs_rem)
 189                 npgs++;
 190         if (npgs > U32_MAX)
 191                 return -EINVAL;
 192
 193         chunks = (unsigned int)div_u64_rem(size, chunk_size, &chunks_rem);
 194         if (chunks == 0)
 195                 return -EINVAL;
 196
 197         if (!unaligned_chunks && chunks_rem)
 198                 return -EINVAL;
 199
 200         if (headroom >= chunk_size - XDP_PACKET_HEADROOM)
 201                 return -EINVAL;
 202
 203         umem->size = size;
 204         umem->headroom = headroom;
 205         umem->chunk_size = chunk_size;
 206         umem->chunks = chunks;
 207         umem->npgs = (u32)npgs;
 208         umem->pgs = NULL;
 209         umem->user = NULL;
 210         umem->flags = mr->flags;
 211
 212         INIT_LIST_HEAD(&umem->xsk_dma_list);
 213         refcount_set(&umem->users, 1);
 214
 215         err = xdp_umem_account_pages(umem);
 216         if (err)
 217                 return err;
 218
 219         err = xdp_umem_pin_pages(umem, (unsigned long)addr);
 220         if (err)
 221                 goto out_account;
 222
 223         err = xdp_umem_addr_map(umem, umem->pgs, umem->npgs);
 224         if (err)
 225                 goto out_unpin;
 226
 227         return 0;
 228
 229 out_unpin:
 230         xdp_umem_unpin_pages(umem);
 231 out_account:
 232         xdp_umem_unaccount_pages(umem);
 233         return err;
 234 }
 235
 236 struct xdp_umem *xdp_umem_create(struct xdp_umem_reg *mr)
 237 {
 238         struct xdp_umem *umem;
 239         int err;
 240
 241         umem = kzalloc(sizeof(*umem), GFP_KERNEL);
 242         if (!umem)
 243                 return ERR_PTR(-ENOMEM);
 244
 245         err = ida_simple_get(&umem_ida, 0, 0, GFP_KERNEL);
 246         if (err < 0) {
 247                 kfree(umem);
 248                 return ERR_PTR(err);
 249         }
 250         umem->id = err;
 251
 252         err = xdp_umem_reg(umem, mr);
 253         if (err) {
 254                 ida_simple_remove(&umem_ida, umem->id);
 255                 kfree(umem);
 256                 return ERR_PTR(err);
 257         }
 258
 259         return umem;
 260 }