net/core/xdp.c

   1 /* net/core/xdp.c
   2  *
   3  * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
   4  * Released under terms in GPL version 2.  See COPYING.
   5  */
   6 #include <linux/bpf.h>
   7 #include <linux/filter.h>
   8 #include <linux/types.h>
   9 #include <linux/mm.h>
  10 #include <linux/netdevice.h>
  11 #include <linux/slab.h>
  12 #include <linux/idr.h>
  13 #include <linux/rhashtable.h>
  14 #include <net/page_pool.h>
  15
  16 #include <net/xdp.h>
  17
  18 #define REG_STATE_NEW           0x0
  19 #define REG_STATE_REGISTERED    0x1
  20 #define REG_STATE_UNREGISTERED  0x2
  21 #define REG_STATE_UNUSED        0x3
  22
  23 static DEFINE_IDA(mem_id_pool);
  24 static DEFINE_MUTEX(mem_id_lock);
  25 #define MEM_ID_MAX 0xFFFE
  26 #define MEM_ID_MIN 1
  27 static int mem_id_next = MEM_ID_MIN;
  28
  29 static bool mem_id_init; /* false */
  30 static struct rhashtable *mem_id_ht;
  31
  32 struct xdp_mem_allocator {
  33         struct xdp_mem_info mem;
  34         union {
  35                 void *allocator;
  36                 struct page_pool *page_pool;
  37                 struct zero_copy_allocator *zc_alloc;
  38         };
  39         struct rhash_head node;
  40         struct rcu_head rcu;
  41 };
  42
  43 static u32 xdp_mem_id_hashfn(const void *data, u32 len, u32 seed)
  44 {
  45         const u32 *k = data;
  46         const u32 key = *k;
  47
  48         BUILD_BUG_ON(FIELD_SIZEOF(struct xdp_mem_allocator, mem.id)
  49                      != sizeof(u32));
  50
  51         /* Use cyclic increasing ID as direct hash key */
  52         return key;
  53 }
  54
  55 static int xdp_mem_id_cmp(struct rhashtable_compare_arg *arg,
  56                           const void *ptr)
  57 {
  58         const struct xdp_mem_allocator *xa = ptr;
  59         u32 mem_id = *(u32 *)arg->key;
  60
  61         return xa->mem.id != mem_id;
  62 }
  63
  64 static const struct rhashtable_params mem_id_rht_params = {
  65         .nelem_hint = 64,
  66         .head_offset = offsetof(struct xdp_mem_allocator, node),
  67         .key_offset  = offsetof(struct xdp_mem_allocator, mem.id),
  68         .key_len = FIELD_SIZEOF(struct xdp_mem_allocator, mem.id),
  69         .max_size = MEM_ID_MAX,
  70         .min_size = 8,
  71         .automatic_shrinking = true,
  72         .hashfn    = xdp_mem_id_hashfn,
  73         .obj_cmpfn = xdp_mem_id_cmp,
  74 };
  75
  76 static void __xdp_mem_allocator_rcu_free(struct rcu_head *rcu)
  77 {
  78         struct xdp_mem_allocator *xa;
  79
  80         xa = container_of(rcu, struct xdp_mem_allocator, rcu);
  81
  82         /* Allow this ID to be reused */
  83         ida_simple_remove(&mem_id_pool, xa->mem.id);
  84
  85         /* Notice, driver is expected to free the *allocator,
  86          * e.g. page_pool, and MUST also use RCU free.
  87          */
  88
  89         /* Poison memory */
  90         xa->mem.id = 0xFFFF;
  91         xa->mem.type = 0xF0F0;
  92         xa->allocator = (void *)0xDEAD9001;
  93
  94         kfree(xa);
  95 }
  96
  97 static void __xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
  98 {
  99         struct xdp_mem_allocator *xa;
 100         int id = xdp_rxq->mem.id;
 101
 102         if (id == 0)
 103                 return;
 104
 105         mutex_lock(&mem_id_lock);
 106
 107         xa = rhashtable_lookup_fast(mem_id_ht, &id, mem_id_rht_params);
 108         if (xa && !rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
 109                 call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
 110
 111         mutex_unlock(&mem_id_lock);
 112 }
 113
 114 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
 115 {
 116         /* Simplify driver cleanup code paths, allow unreg "unused" */
 117         if (xdp_rxq->reg_state == REG_STATE_UNUSED)
 118                 return;
 119
 120         WARN(!(xdp_rxq->reg_state == REG_STATE_REGISTERED), "Driver BUG");
 121
 122         __xdp_rxq_info_unreg_mem_model(xdp_rxq);
 123
 124         xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
 125         xdp_rxq->dev = NULL;
 126
 127         /* Reset mem info to defaults */
 128         xdp_rxq->mem.id = 0;
 129         xdp_rxq->mem.type = 0;
 130 }
 131 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
 132
 133 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
 134 {
 135         memset(xdp_rxq, 0, sizeof(*xdp_rxq));
 136 }
 137
 138 /* Returns 0 on success, negative on failure */
 139 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
 140                      struct net_device *dev, u32 queue_index)
 141 {
 142         if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
 143                 WARN(1, "Driver promised not to register this");
 144                 return -EINVAL;
 145         }
 146
 147         if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
 148                 WARN(1, "Missing unregister, handled but fix driver");
 149                 xdp_rxq_info_unreg(xdp_rxq);
 150         }
 151
 152         if (!dev) {
 153                 WARN(1, "Missing net_device from driver");
 154                 return -ENODEV;
 155         }
 156
 157         /* State either UNREGISTERED or NEW */
 158         xdp_rxq_info_init(xdp_rxq);
 159         xdp_rxq->dev = dev;
 160         xdp_rxq->queue_index = queue_index;
 161
 162         xdp_rxq->reg_state = REG_STATE_REGISTERED;
 163         return 0;
 164 }
 165 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg);
 166
 167 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
 168 {
 169         xdp_rxq->reg_state = REG_STATE_UNUSED;
 170 }
 171 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
 172
 173 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
 174 {
 175         return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
 176 }
 177 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
 178
 179 static int __mem_id_init_hash_table(void)
 180 {
 181         struct rhashtable *rht;
 182         int ret;
 183
 184         if (unlikely(mem_id_init))
 185                 return 0;
 186
 187         rht = kzalloc(sizeof(*rht), GFP_KERNEL);
 188         if (!rht)
 189                 return -ENOMEM;
 190
 191         ret = rhashtable_init(rht, &mem_id_rht_params);
 192         if (ret < 0) {
 193                 kfree(rht);
 194                 return ret;
 195         }
 196         mem_id_ht = rht;
 197         smp_mb(); /* mutex lock should provide enough pairing */
 198         mem_id_init = true;
 199
 200         return 0;
 201 }
 202
 203 /* Allocate a cyclic ID that maps to allocator pointer.
 204  * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
 205  *
 206  * Caller must lock mem_id_lock.
 207  */
 208 static int __mem_id_cyclic_get(gfp_t gfp)
 209 {
 210         int retries = 1;
 211         int id;
 212
 213 again:
 214         id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
 215         if (id < 0) {
 216                 if (id == -ENOSPC) {
 217                         /* Cyclic allocator, reset next id */
 218                         if (retries--) {
 219                                 mem_id_next = MEM_ID_MIN;
 220                                 goto again;
 221                         }
 222                 }
 223                 return id; /* errno */
 224         }
 225         mem_id_next = id + 1;
 226
 227         return id;
 228 }
 229
 230 static bool __is_supported_mem_type(enum xdp_mem_type type)
 231 {
 232         if (type == MEM_TYPE_PAGE_POOL)
 233                 return is_page_pool_compiled_in();
 234
 235         if (type >= MEM_TYPE_MAX)
 236                 return false;
 237
 238         return true;
 239 }
 240
 241 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
 242                                enum xdp_mem_type type, void *allocator)
 243 {
 244         struct xdp_mem_allocator *xdp_alloc;
 245         gfp_t gfp = GFP_KERNEL;
 246         int id, errno, ret;
 247         void *ptr;
 248
 249         if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
 250                 WARN(1, "Missing register, driver bug");
 251                 return -EFAULT;
 252         }
 253
 254         if (!__is_supported_mem_type(type))
 255                 return -EOPNOTSUPP;
 256
 257         xdp_rxq->mem.type = type;
 258
 259         if (!allocator) {
 260                 if (type == MEM_TYPE_PAGE_POOL || type == MEM_TYPE_ZERO_COPY)
 261                         return -EINVAL; /* Setup time check page_pool req */
 262                 return 0;
 263         }
 264
 265         /* Delay init of rhashtable to save memory if feature isn't used */
 266         if (!mem_id_init) {
 267                 mutex_lock(&mem_id_lock);
 268                 ret = __mem_id_init_hash_table();
 269                 mutex_unlock(&mem_id_lock);
 270                 if (ret < 0) {
 271                         WARN_ON(1);
 272                         return ret;
 273                 }
 274         }
 275
 276         xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
 277         if (!xdp_alloc)
 278                 return -ENOMEM;
 279
 280         mutex_lock(&mem_id_lock);
 281         id = __mem_id_cyclic_get(gfp);
 282         if (id < 0) {
 283                 errno = id;
 284                 goto err;
 285         }
 286         xdp_rxq->mem.id = id;
 287         xdp_alloc->mem  = xdp_rxq->mem;
 288         xdp_alloc->allocator = allocator;
 289
 290         /* Insert allocator into ID lookup table */
 291         ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
 292         if (IS_ERR(ptr)) {
 293                 errno = PTR_ERR(ptr);
 294                 goto err;
 295         }
 296
 297         mutex_unlock(&mem_id_lock);
 298
 299         return 0;
 300 err:
 301         mutex_unlock(&mem_id_lock);
 302         kfree(xdp_alloc);
 303         return errno;
 304 }
 305 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
 306
 307 /* XDP RX runs under NAPI protection, and in different delivery error
 308  * scenarios (e.g. queue full), it is possible to return the xdp_frame
 309  * while still leveraging this protection.  The @napi_direct boolian
 310  * is used for those calls sites.  Thus, allowing for faster recycling
 311  * of xdp_frames/pages in those cases.
 312  */
 313 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
 314                          unsigned long handle)
 315 {
 316         struct xdp_mem_allocator *xa;
 317         struct page *page;
 318
 319         switch (mem->type) {
 320         case MEM_TYPE_PAGE_POOL:
 321                 rcu_read_lock();
 322                 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
 323                 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 324                 page = virt_to_head_page(data);
 325                 if (xa) {
 326                         napi_direct &= !xdp_return_frame_no_direct();
 327                         page_pool_put_page(xa->page_pool, page, napi_direct);
 328                 } else {
 329                         put_page(page);
 330                 }
 331                 rcu_read_unlock();
 332                 break;
 333         case MEM_TYPE_PAGE_SHARED:
 334                 page_frag_free(data);
 335                 break;
 336         case MEM_TYPE_PAGE_ORDER0:
 337                 page = virt_to_page(data); /* Assumes order0 page*/
 338                 put_page(page);
 339                 break;
 340         case MEM_TYPE_ZERO_COPY:
 341                 /* NB! Only valid from an xdp_buff! */
 342                 rcu_read_lock();
 343                 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
 344                 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 345                 xa->zc_alloc->free(xa->zc_alloc, handle);
 346                 rcu_read_unlock();
 347         default:
 348                 /* Not possible, checked in xdp_rxq_info_reg_mem_model() */
 349                 break;
 350         }
 351 }
 352
 353 void xdp_return_frame(struct xdp_frame *xdpf)
 354 {
 355         __xdp_return(xdpf->data, &xdpf->mem, false, 0);
 356 }
 357 EXPORT_SYMBOL_GPL(xdp_return_frame);
 358
 359 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
 360 {
 361         __xdp_return(xdpf->data, &xdpf->mem, true, 0);
 362 }
 363 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
 364
 365 void xdp_return_buff(struct xdp_buff *xdp)
 366 {
 367         __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp->handle);
 368 }
 369 EXPORT_SYMBOL_GPL(xdp_return_buff);
 370
 371 int xdp_attachment_query(struct xdp_attachment_info *info,
 372                          struct netdev_bpf *bpf)
 373 {
 374         bpf->prog_id = info->prog ? info->prog->aux->id : 0;
 375         bpf->prog_flags = info->prog ? info->flags : 0;
 376         return 0;
 377 }
 378 EXPORT_SYMBOL_GPL(xdp_attachment_query);
 379
 380 bool xdp_attachment_flags_ok(struct xdp_attachment_info *info,
 381                              struct netdev_bpf *bpf)
 382 {
 383         if (info->prog && (bpf->flags ^ info->flags) & XDP_FLAGS_MODES) {
 384                 NL_SET_ERR_MSG(bpf->extack,
 385                                "program loaded with different flags");
 386                 return false;
 387         }
 388         return true;
 389 }
 390 EXPORT_SYMBOL_GPL(xdp_attachment_flags_ok);
 391
 392 void xdp_attachment_setup(struct xdp_attachment_info *info,
 393                           struct netdev_bpf *bpf)
 394 {
 395         if (info->prog)
 396                 bpf_prog_put(info->prog);
 397         info->prog = bpf->prog;
 398         info->flags = bpf->flags;
 399 }
 400 EXPORT_SYMBOL_GPL(xdp_attachment_setup);