Linux Audio

Check our new training course

Loading...
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0-only
   2/* Copyright (c) 2017 Covalent IO, Inc. http://covalent.io
   3 */
   4
   5/* Devmaps primary use is as a backend map for XDP BPF helper call
   6 * bpf_redirect_map(). Because XDP is mostly concerned with performance we
   7 * spent some effort to ensure the datapath with redirect maps does not use
   8 * any locking. This is a quick note on the details.
   9 *
  10 * We have three possible paths to get into the devmap control plane bpf
  11 * syscalls, bpf programs, and driver side xmit/flush operations. A bpf syscall
  12 * will invoke an update, delete, or lookup operation. To ensure updates and
  13 * deletes appear atomic from the datapath side xchg() is used to modify the
  14 * netdev_map array. Then because the datapath does a lookup into the netdev_map
  15 * array (read-only) from an RCU critical section we use call_rcu() to wait for
  16 * an rcu grace period before free'ing the old data structures. This ensures the
  17 * datapath always has a valid copy. However, the datapath does a "flush"
  18 * operation that pushes any pending packets in the driver outside the RCU
  19 * critical section. Each bpf_dtab_netdev tracks these pending operations using
  20 * a per-cpu flush list. The bpf_dtab_netdev object will not be destroyed  until
  21 * this list is empty, indicating outstanding flush operations have completed.
  22 *
  23 * BPF syscalls may race with BPF program calls on any of the update, delete
  24 * or lookup operations. As noted above the xchg() operation also keep the
  25 * netdev_map consistent in this case. From the devmap side BPF programs
  26 * calling into these operations are the same as multiple user space threads
  27 * making system calls.
  28 *
  29 * Finally, any of the above may race with a netdev_unregister notifier. The
  30 * unregister notifier must search for net devices in the map structure that
  31 * contain a reference to the net device and remove them. This is a two step
  32 * process (a) dereference the bpf_dtab_netdev object in netdev_map and (b)
  33 * check to see if the ifindex is the same as the net_device being removed.
  34 * When removing the dev a cmpxchg() is used to ensure the correct dev is
  35 * removed, in the case of a concurrent update or delete operation it is
  36 * possible that the initially referenced dev is no longer in the map. As the
  37 * notifier hook walks the map we know that new dev references can not be
  38 * added by the user because core infrastructure ensures dev_get_by_index()
  39 * calls will fail at this point.
  40 *
  41 * The devmap_hash type is a map type which interprets keys as ifindexes and
  42 * indexes these using a hashmap. This allows maps that use ifindex as key to be
  43 * densely packed instead of having holes in the lookup array for unused
  44 * ifindexes. The setup and packet enqueue/send code is shared between the two
  45 * types of devmap; only the lookup and insertion is different.
  46 */
  47#include <linux/bpf.h>
  48#include <net/xdp.h>
  49#include <linux/filter.h>
  50#include <trace/events/xdp.h>
  51#include <linux/btf_ids.h>
  52
  53#define DEV_CREATE_FLAG_MASK \
  54	(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
  55
  56struct xdp_dev_bulk_queue {
  57	struct xdp_frame *q[DEV_MAP_BULK_SIZE];
  58	struct list_head flush_node;
  59	struct net_device *dev;
  60	struct net_device *dev_rx;
  61	struct bpf_prog *xdp_prog;
  62	unsigned int count;
  63};
  64
  65struct bpf_dtab_netdev {
  66	struct net_device *dev; /* must be first member, due to tracepoint */
  67	struct hlist_node index_hlist;
 
  68	struct bpf_prog *xdp_prog;
  69	struct rcu_head rcu;
  70	unsigned int idx;
  71	struct bpf_devmap_val val;
  72};
  73
  74struct bpf_dtab {
  75	struct bpf_map map;
  76	struct bpf_dtab_netdev __rcu **netdev_map; /* DEVMAP type only */
  77	struct list_head list;
  78
  79	/* these are only used for DEVMAP_HASH type maps */
  80	struct hlist_head *dev_index_head;
  81	spinlock_t index_lock;
  82	unsigned int items;
  83	u32 n_buckets;
  84};
  85
 
  86static DEFINE_SPINLOCK(dev_map_lock);
  87static LIST_HEAD(dev_map_list);
  88
  89static struct hlist_head *dev_map_create_hash(unsigned int entries,
  90					      int numa_node)
  91{
  92	int i;
  93	struct hlist_head *hash;
  94
  95	hash = bpf_map_area_alloc((u64) entries * sizeof(*hash), numa_node);
  96	if (hash != NULL)
  97		for (i = 0; i < entries; i++)
  98			INIT_HLIST_HEAD(&hash[i]);
  99
 100	return hash;
 101}
 102
 103static inline struct hlist_head *dev_map_index_hash(struct bpf_dtab *dtab,
 104						    int idx)
 105{
 106	return &dtab->dev_index_head[idx & (dtab->n_buckets - 1)];
 107}
 108
 109static int dev_map_alloc_check(union bpf_attr *attr)
 110{
 111	u32 valsize = attr->value_size;
 112
 113	/* check sanity of attributes. 2 value sizes supported:
 114	 * 4 bytes: ifindex
 115	 * 8 bytes: ifindex + prog fd
 116	 */
 117	if (attr->max_entries == 0 || attr->key_size != 4 ||
 118	    (valsize != offsetofend(struct bpf_devmap_val, ifindex) &&
 119	     valsize != offsetofend(struct bpf_devmap_val, bpf_prog.fd)) ||
 120	    attr->map_flags & ~DEV_CREATE_FLAG_MASK)
 121		return -EINVAL;
 122
 123	if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 124		/* Hash table size must be power of 2; roundup_pow_of_two()
 125		 * can overflow into UB on 32-bit arches
 126		 */
 127		if (attr->max_entries > 1UL << 31)
 128			return -EINVAL;
 129	}
 130
 131	return 0;
 132}
 133
 134static int dev_map_init_map(struct bpf_dtab *dtab, union bpf_attr *attr)
 135{
 136	/* Lookup returns a pointer straight to dev->ifindex, so make sure the
 137	 * verifier prevents writes from the BPF side
 138	 */
 139	attr->map_flags |= BPF_F_RDONLY_PROG;
 
 
 140	bpf_map_init_from_attr(&dtab->map, attr);
 141
 142	if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 143		/* Hash table size must be power of 2 */
 144		dtab->n_buckets = roundup_pow_of_two(dtab->map.max_entries);
 
 
 
 
 
 
 145		dtab->dev_index_head = dev_map_create_hash(dtab->n_buckets,
 146							   dtab->map.numa_node);
 147		if (!dtab->dev_index_head)
 148			return -ENOMEM;
 149
 150		spin_lock_init(&dtab->index_lock);
 151	} else {
 152		dtab->netdev_map = bpf_map_area_alloc((u64) dtab->map.max_entries *
 153						      sizeof(struct bpf_dtab_netdev *),
 154						      dtab->map.numa_node);
 155		if (!dtab->netdev_map)
 156			return -ENOMEM;
 157	}
 158
 159	return 0;
 160}
 161
 162static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
 163{
 164	struct bpf_dtab *dtab;
 165	int err;
 166
 167	dtab = bpf_map_area_alloc(sizeof(*dtab), NUMA_NO_NODE);
 
 
 
 168	if (!dtab)
 169		return ERR_PTR(-ENOMEM);
 170
 171	err = dev_map_init_map(dtab, attr);
 172	if (err) {
 173		bpf_map_area_free(dtab);
 174		return ERR_PTR(err);
 175	}
 176
 177	spin_lock(&dev_map_lock);
 178	list_add_tail_rcu(&dtab->list, &dev_map_list);
 179	spin_unlock(&dev_map_lock);
 180
 181	return &dtab->map;
 182}
 183
 184static void dev_map_free(struct bpf_map *map)
 185{
 186	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 187	u32 i;
 188
 189	/* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 190	 * so the programs (can be more than one that used this map) were
 191	 * disconnected from events. The following synchronize_rcu() guarantees
 192	 * both rcu read critical sections complete and waits for
 193	 * preempt-disable regions (NAPI being the relevant context here) so we
 194	 * are certain there will be no further reads against the netdev_map and
 195	 * all flush operations are complete. Flush operations can only be done
 196	 * from NAPI context for this reason.
 197	 */
 198
 199	spin_lock(&dev_map_lock);
 200	list_del_rcu(&dtab->list);
 201	spin_unlock(&dev_map_lock);
 202
 203	/* bpf_redirect_info->map is assigned in __bpf_xdp_redirect_map()
 204	 * during NAPI callback and cleared after the XDP redirect. There is no
 205	 * explicit RCU read section which protects bpf_redirect_info->map but
 206	 * local_bh_disable() also marks the beginning an RCU section. This
 207	 * makes the complete softirq callback RCU protected. Thus after
 208	 * following synchronize_rcu() there no bpf_redirect_info->map == map
 209	 * assignment.
 210	 */
 211	synchronize_rcu();
 212
 213	/* Make sure prior __dev_map_entry_free() have completed. */
 214	rcu_barrier();
 215
 216	if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 217		for (i = 0; i < dtab->n_buckets; i++) {
 218			struct bpf_dtab_netdev *dev;
 219			struct hlist_head *head;
 220			struct hlist_node *next;
 221
 222			head = dev_map_index_hash(dtab, i);
 223
 224			hlist_for_each_entry_safe(dev, next, head, index_hlist) {
 225				hlist_del_rcu(&dev->index_hlist);
 226				if (dev->xdp_prog)
 227					bpf_prog_put(dev->xdp_prog);
 228				dev_put(dev->dev);
 229				kfree(dev);
 230			}
 231		}
 232
 233		bpf_map_area_free(dtab->dev_index_head);
 234	} else {
 235		for (i = 0; i < dtab->map.max_entries; i++) {
 236			struct bpf_dtab_netdev *dev;
 237
 238			dev = rcu_dereference_raw(dtab->netdev_map[i]);
 239			if (!dev)
 240				continue;
 241
 242			if (dev->xdp_prog)
 243				bpf_prog_put(dev->xdp_prog);
 244			dev_put(dev->dev);
 245			kfree(dev);
 246		}
 247
 248		bpf_map_area_free(dtab->netdev_map);
 249	}
 250
 251	bpf_map_area_free(dtab);
 252}
 253
 254static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 255{
 256	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 257	u32 index = key ? *(u32 *)key : U32_MAX;
 258	u32 *next = next_key;
 259
 260	if (index >= dtab->map.max_entries) {
 261		*next = 0;
 262		return 0;
 263	}
 264
 265	if (index == dtab->map.max_entries - 1)
 266		return -ENOENT;
 267	*next = index + 1;
 268	return 0;
 269}
 270
 271/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 272 * by local_bh_disable() (from XDP calls inside NAPI). The
 273 * rcu_read_lock_bh_held() below makes lockdep accept both.
 274 */
 275static void *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
 276{
 277	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 278	struct hlist_head *head = dev_map_index_hash(dtab, key);
 279	struct bpf_dtab_netdev *dev;
 280
 281	hlist_for_each_entry_rcu(dev, head, index_hlist,
 282				 lockdep_is_held(&dtab->index_lock))
 283		if (dev->idx == key)
 284			return dev;
 285
 286	return NULL;
 287}
 288
 289static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
 290				    void *next_key)
 291{
 292	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 293	u32 idx, *next = next_key;
 294	struct bpf_dtab_netdev *dev, *next_dev;
 295	struct hlist_head *head;
 296	int i = 0;
 297
 298	if (!key)
 299		goto find_first;
 300
 301	idx = *(u32 *)key;
 302
 303	dev = __dev_map_hash_lookup_elem(map, idx);
 304	if (!dev)
 305		goto find_first;
 306
 307	next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
 308				    struct bpf_dtab_netdev, index_hlist);
 309
 310	if (next_dev) {
 311		*next = next_dev->idx;
 312		return 0;
 313	}
 314
 315	i = idx & (dtab->n_buckets - 1);
 316	i++;
 317
 318 find_first:
 319	for (; i < dtab->n_buckets; i++) {
 320		head = dev_map_index_hash(dtab, i);
 321
 322		next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
 323					    struct bpf_dtab_netdev,
 324					    index_hlist);
 325		if (next_dev) {
 326			*next = next_dev->idx;
 327			return 0;
 328		}
 329	}
 330
 331	return -ENOENT;
 332}
 333
 
 
 
 
 
 
 
 
 
 
 334static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
 335				struct xdp_frame **frames, int n,
 336				struct net_device *tx_dev,
 337				struct net_device *rx_dev)
 338{
 339	struct xdp_txq_info txq = { .dev = tx_dev };
 340	struct xdp_rxq_info rxq = { .dev = rx_dev };
 341	struct xdp_buff xdp;
 342	int i, nframes = 0;
 343
 344	for (i = 0; i < n; i++) {
 345		struct xdp_frame *xdpf = frames[i];
 346		u32 act;
 347		int err;
 348
 349		xdp_convert_frame_to_buff(xdpf, &xdp);
 350		xdp.txq = &txq;
 351		xdp.rxq = &rxq;
 352
 353		act = bpf_prog_run_xdp(xdp_prog, &xdp);
 354		switch (act) {
 355		case XDP_PASS:
 356			err = xdp_update_frame_from_buff(&xdp, xdpf);
 357			if (unlikely(err < 0))
 358				xdp_return_frame_rx_napi(xdpf);
 359			else
 360				frames[nframes++] = xdpf;
 361			break;
 362		default:
 363			bpf_warn_invalid_xdp_action(NULL, xdp_prog, act);
 364			fallthrough;
 365		case XDP_ABORTED:
 366			trace_xdp_exception(tx_dev, xdp_prog, act);
 367			fallthrough;
 368		case XDP_DROP:
 369			xdp_return_frame_rx_napi(xdpf);
 370			break;
 371		}
 372	}
 373	return nframes; /* sent frames count */
 374}
 375
 376static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
 377{
 378	struct net_device *dev = bq->dev;
 379	unsigned int cnt = bq->count;
 380	int sent = 0, err = 0;
 381	int to_send = cnt;
 382	int i;
 383
 384	if (unlikely(!cnt))
 385		return;
 386
 387	for (i = 0; i < cnt; i++) {
 388		struct xdp_frame *xdpf = bq->q[i];
 389
 390		prefetch(xdpf);
 391	}
 392
 393	if (bq->xdp_prog) {
 394		to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev, bq->dev_rx);
 395		if (!to_send)
 396			goto out;
 397	}
 398
 399	sent = dev->netdev_ops->ndo_xdp_xmit(dev, to_send, bq->q, flags);
 400	if (sent < 0) {
 401		/* If ndo_xdp_xmit fails with an errno, no frames have
 402		 * been xmit'ed.
 403		 */
 404		err = sent;
 405		sent = 0;
 406	}
 407
 408	/* If not all frames have been transmitted, it is our
 409	 * responsibility to free them
 410	 */
 411	for (i = sent; unlikely(i < to_send); i++)
 412		xdp_return_frame_rx_napi(bq->q[i]);
 413
 414out:
 415	bq->count = 0;
 416	trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent, err);
 417}
 418
 419/* __dev_flush is called from xdp_do_flush() which _must_ be signalled from the
 420 * driver before returning from its napi->poll() routine. See the comment above
 421 * xdp_do_flush() in filter.c.
 422 */
 423void __dev_flush(struct list_head *flush_list)
 424{
 
 425	struct xdp_dev_bulk_queue *bq, *tmp;
 426
 427	list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
 428		bq_xmit_all(bq, XDP_XMIT_FLUSH);
 429		bq->dev_rx = NULL;
 430		bq->xdp_prog = NULL;
 431		__list_del_clearprev(&bq->flush_node);
 432	}
 433}
 434
 435/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 436 * by local_bh_disable() (from XDP calls inside NAPI). The
 437 * rcu_read_lock_bh_held() below makes lockdep accept both.
 438 */
 439static void *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
 440{
 441	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 442	struct bpf_dtab_netdev *obj;
 443
 444	if (key >= map->max_entries)
 445		return NULL;
 446
 447	obj = rcu_dereference_check(dtab->netdev_map[key],
 448				    rcu_read_lock_bh_held());
 449	return obj;
 450}
 451
 452/* Runs in NAPI, i.e., softirq under local_bh_disable(). Thus, safe percpu
 453 * variable access, and map elements stick around. See comment above
 454 * xdp_do_flush() in filter.c.
 455 */
 456static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 457		       struct net_device *dev_rx, struct bpf_prog *xdp_prog)
 458{
 
 459	struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
 460
 461	if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
 462		bq_xmit_all(bq, 0);
 463
 464	/* Ingress dev_rx will be the same for all xdp_frame's in
 465	 * bulk_queue, because bq stored per-CPU and must be flushed
 466	 * from net_device drivers NAPI func end.
 467	 *
 468	 * Do the same with xdp_prog and flush_list since these fields
 469	 * are only ever modified together.
 470	 */
 471	if (!bq->dev_rx) {
 472		struct list_head *flush_list = bpf_net_ctx_get_dev_flush_list();
 473
 474		bq->dev_rx = dev_rx;
 475		bq->xdp_prog = xdp_prog;
 476		list_add(&bq->flush_node, flush_list);
 477	}
 478
 479	bq->q[bq->count++] = xdpf;
 480}
 481
 482static inline int __xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 483				struct net_device *dev_rx,
 484				struct bpf_prog *xdp_prog)
 485{
 
 486	int err;
 487
 488	if (!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
 489		return -EOPNOTSUPP;
 490
 491	if (unlikely(!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
 492		     xdp_frame_has_frags(xdpf)))
 493		return -EOPNOTSUPP;
 494
 495	err = xdp_ok_fwd_dev(dev, xdp_get_frame_len(xdpf));
 496	if (unlikely(err))
 497		return err;
 498
 
 
 
 
 499	bq_enqueue(dev, xdpf, dev_rx, xdp_prog);
 500	return 0;
 501}
 502
 503static u32 dev_map_bpf_prog_run_skb(struct sk_buff *skb, struct bpf_dtab_netdev *dst)
 504{
 505	struct xdp_txq_info txq = { .dev = dst->dev };
 506	struct xdp_buff xdp;
 507	u32 act;
 508
 509	if (!dst->xdp_prog)
 510		return XDP_PASS;
 511
 512	__skb_pull(skb, skb->mac_len);
 513	xdp.txq = &txq;
 514
 515	act = bpf_prog_run_generic_xdp(skb, &xdp, dst->xdp_prog);
 516	switch (act) {
 517	case XDP_PASS:
 518		__skb_push(skb, skb->mac_len);
 519		break;
 520	default:
 521		bpf_warn_invalid_xdp_action(NULL, dst->xdp_prog, act);
 522		fallthrough;
 523	case XDP_ABORTED:
 524		trace_xdp_exception(dst->dev, dst->xdp_prog, act);
 525		fallthrough;
 526	case XDP_DROP:
 527		kfree_skb(skb);
 528		break;
 529	}
 530
 531	return act;
 532}
 533
 534int dev_xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 535		    struct net_device *dev_rx)
 536{
 537	return __xdp_enqueue(dev, xdpf, dev_rx, NULL);
 538}
 539
 540int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_frame *xdpf,
 541		    struct net_device *dev_rx)
 542{
 543	struct net_device *dev = dst->dev;
 544
 545	return __xdp_enqueue(dev, xdpf, dev_rx, dst->xdp_prog);
 546}
 547
 548static bool is_valid_dst(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf)
 
 549{
 550	if (!obj)
 551		return false;
 552
 553	if (!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
 554		return false;
 555
 556	if (unlikely(!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
 557		     xdp_frame_has_frags(xdpf)))
 558		return false;
 559
 560	if (xdp_ok_fwd_dev(obj->dev, xdp_get_frame_len(xdpf)))
 561		return false;
 562
 563	return true;
 564}
 565
 566static int dev_map_enqueue_clone(struct bpf_dtab_netdev *obj,
 567				 struct net_device *dev_rx,
 568				 struct xdp_frame *xdpf)
 569{
 570	struct xdp_frame *nxdpf;
 571
 572	nxdpf = xdpf_clone(xdpf);
 573	if (!nxdpf)
 574		return -ENOMEM;
 575
 576	bq_enqueue(obj->dev, nxdpf, dev_rx, obj->xdp_prog);
 577
 578	return 0;
 579}
 580
 581static inline bool is_ifindex_excluded(int *excluded, int num_excluded, int ifindex)
 582{
 583	while (num_excluded--) {
 584		if (ifindex == excluded[num_excluded])
 585			return true;
 586	}
 587	return false;
 588}
 589
 590/* Get ifindex of each upper device. 'indexes' must be able to hold at
 591 * least MAX_NEST_DEV elements.
 592 * Returns the number of ifindexes added.
 593 */
 594static int get_upper_ifindexes(struct net_device *dev, int *indexes)
 595{
 596	struct net_device *upper;
 597	struct list_head *iter;
 598	int n = 0;
 599
 600	netdev_for_each_upper_dev_rcu(dev, upper, iter) {
 601		indexes[n++] = upper->ifindex;
 602	}
 603	return n;
 604}
 605
 606int dev_map_enqueue_multi(struct xdp_frame *xdpf, struct net_device *dev_rx,
 607			  struct bpf_map *map, bool exclude_ingress)
 608{
 609	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 
 610	struct bpf_dtab_netdev *dst, *last_dst = NULL;
 611	int excluded_devices[1+MAX_NEST_DEV];
 612	struct hlist_head *head;
 613	int num_excluded = 0;
 614	unsigned int i;
 615	int err;
 616
 617	if (exclude_ingress) {
 618		num_excluded = get_upper_ifindexes(dev_rx, excluded_devices);
 619		excluded_devices[num_excluded++] = dev_rx->ifindex;
 620	}
 621
 622	if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
 623		for (i = 0; i < map->max_entries; i++) {
 624			dst = rcu_dereference_check(dtab->netdev_map[i],
 625						    rcu_read_lock_bh_held());
 626			if (!is_valid_dst(dst, xdpf))
 627				continue;
 628
 629			if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
 630				continue;
 631
 632			/* we only need n-1 clones; last_dst enqueued below */
 633			if (!last_dst) {
 634				last_dst = dst;
 635				continue;
 636			}
 637
 638			err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
 639			if (err)
 640				return err;
 641
 642			last_dst = dst;
 643		}
 644	} else { /* BPF_MAP_TYPE_DEVMAP_HASH */
 645		for (i = 0; i < dtab->n_buckets; i++) {
 646			head = dev_map_index_hash(dtab, i);
 647			hlist_for_each_entry_rcu(dst, head, index_hlist,
 648						 lockdep_is_held(&dtab->index_lock)) {
 649				if (!is_valid_dst(dst, xdpf))
 650					continue;
 651
 652				if (is_ifindex_excluded(excluded_devices, num_excluded,
 653							dst->dev->ifindex))
 654					continue;
 655
 656				/* we only need n-1 clones; last_dst enqueued below */
 657				if (!last_dst) {
 658					last_dst = dst;
 659					continue;
 660				}
 661
 662				err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
 663				if (err)
 664					return err;
 665
 666				last_dst = dst;
 667			}
 668		}
 669	}
 670
 671	/* consume the last copy of the frame */
 672	if (last_dst)
 673		bq_enqueue(last_dst->dev, xdpf, dev_rx, last_dst->xdp_prog);
 674	else
 675		xdp_return_frame_rx_napi(xdpf); /* dtab is empty */
 676
 677	return 0;
 678}
 679
 680int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
 681			     struct bpf_prog *xdp_prog)
 682{
 683	int err;
 684
 685	err = xdp_ok_fwd_dev(dst->dev, skb->len);
 686	if (unlikely(err))
 687		return err;
 688
 689	/* Redirect has already succeeded semantically at this point, so we just
 690	 * return 0 even if packet is dropped. Helper below takes care of
 691	 * freeing skb.
 692	 */
 693	if (dev_map_bpf_prog_run_skb(skb, dst) != XDP_PASS)
 694		return 0;
 695
 696	skb->dev = dst->dev;
 697	generic_xdp_tx(skb, xdp_prog);
 698
 699	return 0;
 700}
 701
 702static int dev_map_redirect_clone(struct bpf_dtab_netdev *dst,
 703				  struct sk_buff *skb,
 704				  struct bpf_prog *xdp_prog)
 705{
 706	struct sk_buff *nskb;
 707	int err;
 708
 709	nskb = skb_clone(skb, GFP_ATOMIC);
 710	if (!nskb)
 711		return -ENOMEM;
 712
 713	err = dev_map_generic_redirect(dst, nskb, xdp_prog);
 714	if (unlikely(err)) {
 715		consume_skb(nskb);
 716		return err;
 717	}
 718
 719	return 0;
 720}
 721
 722int dev_map_redirect_multi(struct net_device *dev, struct sk_buff *skb,
 723			   struct bpf_prog *xdp_prog, struct bpf_map *map,
 724			   bool exclude_ingress)
 725{
 726	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 
 727	struct bpf_dtab_netdev *dst, *last_dst = NULL;
 728	int excluded_devices[1+MAX_NEST_DEV];
 729	struct hlist_head *head;
 730	struct hlist_node *next;
 731	int num_excluded = 0;
 732	unsigned int i;
 733	int err;
 734
 735	if (exclude_ingress) {
 736		num_excluded = get_upper_ifindexes(dev, excluded_devices);
 737		excluded_devices[num_excluded++] = dev->ifindex;
 738	}
 739
 740	if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
 741		for (i = 0; i < map->max_entries; i++) {
 742			dst = rcu_dereference_check(dtab->netdev_map[i],
 743						    rcu_read_lock_bh_held());
 744			if (!dst)
 745				continue;
 746
 747			if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
 748				continue;
 749
 750			/* we only need n-1 clones; last_dst enqueued below */
 751			if (!last_dst) {
 752				last_dst = dst;
 753				continue;
 754			}
 755
 756			err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
 757			if (err)
 758				return err;
 759
 760			last_dst = dst;
 761
 762		}
 763	} else { /* BPF_MAP_TYPE_DEVMAP_HASH */
 764		for (i = 0; i < dtab->n_buckets; i++) {
 765			head = dev_map_index_hash(dtab, i);
 766			hlist_for_each_entry_safe(dst, next, head, index_hlist) {
 767				if (is_ifindex_excluded(excluded_devices, num_excluded,
 768							dst->dev->ifindex))
 769					continue;
 770
 771				/* we only need n-1 clones; last_dst enqueued below */
 772				if (!last_dst) {
 773					last_dst = dst;
 774					continue;
 775				}
 776
 777				err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
 778				if (err)
 779					return err;
 780
 781				last_dst = dst;
 782			}
 783		}
 784	}
 785
 786	/* consume the first skb and return */
 787	if (last_dst)
 788		return dev_map_generic_redirect(last_dst, skb, xdp_prog);
 789
 790	/* dtab is empty */
 791	consume_skb(skb);
 792	return 0;
 793}
 794
 795static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
 796{
 797	struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
 798
 799	return obj ? &obj->val : NULL;
 800}
 801
 802static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
 803{
 804	struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
 805								*(u32 *)key);
 806	return obj ? &obj->val : NULL;
 807}
 808
 809static void __dev_map_entry_free(struct rcu_head *rcu)
 810{
 811	struct bpf_dtab_netdev *dev;
 812
 813	dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
 814	if (dev->xdp_prog)
 815		bpf_prog_put(dev->xdp_prog);
 816	dev_put(dev->dev);
 817	kfree(dev);
 818}
 819
 820static long dev_map_delete_elem(struct bpf_map *map, void *key)
 821{
 822	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 823	struct bpf_dtab_netdev *old_dev;
 824	u32 k = *(u32 *)key;
 825
 826	if (k >= map->max_entries)
 827		return -EINVAL;
 828
 829	old_dev = unrcu_pointer(xchg(&dtab->netdev_map[k], NULL));
 830	if (old_dev) {
 831		call_rcu(&old_dev->rcu, __dev_map_entry_free);
 832		atomic_dec((atomic_t *)&dtab->items);
 833	}
 834	return 0;
 835}
 836
 837static long dev_map_hash_delete_elem(struct bpf_map *map, void *key)
 838{
 839	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 840	struct bpf_dtab_netdev *old_dev;
 841	u32 k = *(u32 *)key;
 842	unsigned long flags;
 843	int ret = -ENOENT;
 844
 845	spin_lock_irqsave(&dtab->index_lock, flags);
 846
 847	old_dev = __dev_map_hash_lookup_elem(map, k);
 848	if (old_dev) {
 849		dtab->items--;
 850		hlist_del_init_rcu(&old_dev->index_hlist);
 851		call_rcu(&old_dev->rcu, __dev_map_entry_free);
 852		ret = 0;
 853	}
 854	spin_unlock_irqrestore(&dtab->index_lock, flags);
 855
 856	return ret;
 857}
 858
 859static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
 860						    struct bpf_dtab *dtab,
 861						    struct bpf_devmap_val *val,
 862						    unsigned int idx)
 863{
 864	struct bpf_prog *prog = NULL;
 865	struct bpf_dtab_netdev *dev;
 866
 867	dev = bpf_map_kmalloc_node(&dtab->map, sizeof(*dev),
 868				   GFP_NOWAIT | __GFP_NOWARN,
 869				   dtab->map.numa_node);
 870	if (!dev)
 871		return ERR_PTR(-ENOMEM);
 872
 873	dev->dev = dev_get_by_index(net, val->ifindex);
 874	if (!dev->dev)
 875		goto err_out;
 876
 877	if (val->bpf_prog.fd > 0) {
 878		prog = bpf_prog_get_type_dev(val->bpf_prog.fd,
 879					     BPF_PROG_TYPE_XDP, false);
 880		if (IS_ERR(prog))
 881			goto err_put_dev;
 882		if (prog->expected_attach_type != BPF_XDP_DEVMAP ||
 883		    !bpf_prog_map_compatible(&dtab->map, prog))
 884			goto err_put_prog;
 885	}
 886
 887	dev->idx = idx;
 
 888	if (prog) {
 889		dev->xdp_prog = prog;
 890		dev->val.bpf_prog.id = prog->aux->id;
 891	} else {
 892		dev->xdp_prog = NULL;
 893		dev->val.bpf_prog.id = 0;
 894	}
 895	dev->val.ifindex = val->ifindex;
 896
 897	return dev;
 898err_put_prog:
 899	bpf_prog_put(prog);
 900err_put_dev:
 901	dev_put(dev->dev);
 902err_out:
 903	kfree(dev);
 904	return ERR_PTR(-EINVAL);
 905}
 906
 907static long __dev_map_update_elem(struct net *net, struct bpf_map *map,
 908				  void *key, void *value, u64 map_flags)
 909{
 910	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 911	struct bpf_dtab_netdev *dev, *old_dev;
 912	struct bpf_devmap_val val = {};
 913	u32 i = *(u32 *)key;
 914
 915	if (unlikely(map_flags > BPF_EXIST))
 916		return -EINVAL;
 917	if (unlikely(i >= dtab->map.max_entries))
 918		return -E2BIG;
 919	if (unlikely(map_flags == BPF_NOEXIST))
 920		return -EEXIST;
 921
 922	/* already verified value_size <= sizeof val */
 923	memcpy(&val, value, map->value_size);
 924
 925	if (!val.ifindex) {
 926		dev = NULL;
 927		/* can not specify fd if ifindex is 0 */
 928		if (val.bpf_prog.fd > 0)
 929			return -EINVAL;
 930	} else {
 931		dev = __dev_map_alloc_node(net, dtab, &val, i);
 932		if (IS_ERR(dev))
 933			return PTR_ERR(dev);
 934	}
 935
 936	/* Use call_rcu() here to ensure rcu critical sections have completed
 937	 * Remembering the driver side flush operation will happen before the
 938	 * net device is removed.
 939	 */
 940	old_dev = unrcu_pointer(xchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev)));
 941	if (old_dev)
 942		call_rcu(&old_dev->rcu, __dev_map_entry_free);
 943	else
 944		atomic_inc((atomic_t *)&dtab->items);
 945
 946	return 0;
 947}
 948
 949static long dev_map_update_elem(struct bpf_map *map, void *key, void *value,
 950				u64 map_flags)
 951{
 952	return __dev_map_update_elem(current->nsproxy->net_ns,
 953				     map, key, value, map_flags);
 954}
 955
 956static long __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
 957				       void *key, void *value, u64 map_flags)
 958{
 959	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 960	struct bpf_dtab_netdev *dev, *old_dev;
 961	struct bpf_devmap_val val = {};
 962	u32 idx = *(u32 *)key;
 963	unsigned long flags;
 964	int err = -EEXIST;
 965
 966	/* already verified value_size <= sizeof val */
 967	memcpy(&val, value, map->value_size);
 968
 969	if (unlikely(map_flags > BPF_EXIST || !val.ifindex))
 970		return -EINVAL;
 971
 972	spin_lock_irqsave(&dtab->index_lock, flags);
 973
 974	old_dev = __dev_map_hash_lookup_elem(map, idx);
 975	if (old_dev && (map_flags & BPF_NOEXIST))
 976		goto out_err;
 977
 978	dev = __dev_map_alloc_node(net, dtab, &val, idx);
 979	if (IS_ERR(dev)) {
 980		err = PTR_ERR(dev);
 981		goto out_err;
 982	}
 983
 984	if (old_dev) {
 985		hlist_del_rcu(&old_dev->index_hlist);
 986	} else {
 987		if (dtab->items >= dtab->map.max_entries) {
 988			spin_unlock_irqrestore(&dtab->index_lock, flags);
 989			call_rcu(&dev->rcu, __dev_map_entry_free);
 990			return -E2BIG;
 991		}
 992		dtab->items++;
 993	}
 994
 995	hlist_add_head_rcu(&dev->index_hlist,
 996			   dev_map_index_hash(dtab, idx));
 997	spin_unlock_irqrestore(&dtab->index_lock, flags);
 998
 999	if (old_dev)
1000		call_rcu(&old_dev->rcu, __dev_map_entry_free);
1001
1002	return 0;
1003
1004out_err:
1005	spin_unlock_irqrestore(&dtab->index_lock, flags);
1006	return err;
1007}
1008
1009static long dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
1010				     u64 map_flags)
1011{
1012	return __dev_map_hash_update_elem(current->nsproxy->net_ns,
1013					 map, key, value, map_flags);
1014}
1015
1016static long dev_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
1017{
1018	return __bpf_xdp_redirect_map(map, ifindex, flags,
1019				      BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1020				      __dev_map_lookup_elem);
1021}
1022
1023static long dev_hash_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
1024{
1025	return __bpf_xdp_redirect_map(map, ifindex, flags,
1026				      BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1027				      __dev_map_hash_lookup_elem);
1028}
1029
1030static u64 dev_map_mem_usage(const struct bpf_map *map)
1031{
1032	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
1033	u64 usage = sizeof(struct bpf_dtab);
1034
1035	if (map->map_type == BPF_MAP_TYPE_DEVMAP_HASH)
1036		usage += (u64)dtab->n_buckets * sizeof(struct hlist_head);
1037	else
1038		usage += (u64)map->max_entries * sizeof(struct bpf_dtab_netdev *);
1039	usage += atomic_read((atomic_t *)&dtab->items) *
1040			 (u64)sizeof(struct bpf_dtab_netdev);
1041	return usage;
1042}
1043
1044BTF_ID_LIST_SINGLE(dev_map_btf_ids, struct, bpf_dtab)
1045const struct bpf_map_ops dev_map_ops = {
1046	.map_meta_equal = bpf_map_meta_equal,
1047	.map_alloc_check = dev_map_alloc_check,
1048	.map_alloc = dev_map_alloc,
1049	.map_free = dev_map_free,
1050	.map_get_next_key = dev_map_get_next_key,
1051	.map_lookup_elem = dev_map_lookup_elem,
1052	.map_update_elem = dev_map_update_elem,
1053	.map_delete_elem = dev_map_delete_elem,
1054	.map_check_btf = map_check_no_btf,
1055	.map_mem_usage = dev_map_mem_usage,
1056	.map_btf_id = &dev_map_btf_ids[0],
1057	.map_redirect = dev_map_redirect,
1058};
1059
 
1060const struct bpf_map_ops dev_map_hash_ops = {
1061	.map_meta_equal = bpf_map_meta_equal,
1062	.map_alloc_check = dev_map_alloc_check,
1063	.map_alloc = dev_map_alloc,
1064	.map_free = dev_map_free,
1065	.map_get_next_key = dev_map_hash_get_next_key,
1066	.map_lookup_elem = dev_map_hash_lookup_elem,
1067	.map_update_elem = dev_map_hash_update_elem,
1068	.map_delete_elem = dev_map_hash_delete_elem,
1069	.map_check_btf = map_check_no_btf,
1070	.map_mem_usage = dev_map_mem_usage,
1071	.map_btf_id = &dev_map_btf_ids[0],
1072	.map_redirect = dev_hash_map_redirect,
1073};
1074
1075static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
1076				       struct net_device *netdev)
1077{
1078	unsigned long flags;
1079	u32 i;
1080
1081	spin_lock_irqsave(&dtab->index_lock, flags);
1082	for (i = 0; i < dtab->n_buckets; i++) {
1083		struct bpf_dtab_netdev *dev;
1084		struct hlist_head *head;
1085		struct hlist_node *next;
1086
1087		head = dev_map_index_hash(dtab, i);
1088
1089		hlist_for_each_entry_safe(dev, next, head, index_hlist) {
1090			if (netdev != dev->dev)
1091				continue;
1092
1093			dtab->items--;
1094			hlist_del_rcu(&dev->index_hlist);
1095			call_rcu(&dev->rcu, __dev_map_entry_free);
1096		}
1097	}
1098	spin_unlock_irqrestore(&dtab->index_lock, flags);
1099}
1100
1101static int dev_map_notification(struct notifier_block *notifier,
1102				ulong event, void *ptr)
1103{
1104	struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
1105	struct bpf_dtab *dtab;
1106	int i, cpu;
1107
1108	switch (event) {
1109	case NETDEV_REGISTER:
1110		if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
1111			break;
1112
1113		/* will be freed in free_netdev() */
1114		netdev->xdp_bulkq = alloc_percpu(struct xdp_dev_bulk_queue);
1115		if (!netdev->xdp_bulkq)
1116			return NOTIFY_BAD;
1117
1118		for_each_possible_cpu(cpu)
1119			per_cpu_ptr(netdev->xdp_bulkq, cpu)->dev = netdev;
1120		break;
1121	case NETDEV_UNREGISTER:
1122		/* This rcu_read_lock/unlock pair is needed because
1123		 * dev_map_list is an RCU list AND to ensure a delete
1124		 * operation does not free a netdev_map entry while we
1125		 * are comparing it against the netdev being unregistered.
1126		 */
1127		rcu_read_lock();
1128		list_for_each_entry_rcu(dtab, &dev_map_list, list) {
1129			if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
1130				dev_map_hash_remove_netdev(dtab, netdev);
1131				continue;
1132			}
1133
1134			for (i = 0; i < dtab->map.max_entries; i++) {
1135				struct bpf_dtab_netdev *dev, *odev;
1136
1137				dev = rcu_dereference(dtab->netdev_map[i]);
1138				if (!dev || netdev != dev->dev)
1139					continue;
1140				odev = unrcu_pointer(cmpxchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev), NULL));
1141				if (dev == odev) {
1142					call_rcu(&dev->rcu,
1143						 __dev_map_entry_free);
1144					atomic_dec((atomic_t *)&dtab->items);
1145				}
1146			}
1147		}
1148		rcu_read_unlock();
1149		break;
1150	default:
1151		break;
1152	}
1153	return NOTIFY_OK;
1154}
1155
1156static struct notifier_block dev_map_notifier = {
1157	.notifier_call = dev_map_notification,
1158};
1159
1160static int __init dev_map_init(void)
1161{
 
 
1162	/* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
1163	BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
1164		     offsetof(struct _bpf_dtab_netdev, dev));
1165	register_netdevice_notifier(&dev_map_notifier);
1166
 
 
1167	return 0;
1168}
1169
1170subsys_initcall(dev_map_init);
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-only
   2/* Copyright (c) 2017 Covalent IO, Inc. http://covalent.io
   3 */
   4
   5/* Devmaps primary use is as a backend map for XDP BPF helper call
   6 * bpf_redirect_map(). Because XDP is mostly concerned with performance we
   7 * spent some effort to ensure the datapath with redirect maps does not use
   8 * any locking. This is a quick note on the details.
   9 *
  10 * We have three possible paths to get into the devmap control plane bpf
  11 * syscalls, bpf programs, and driver side xmit/flush operations. A bpf syscall
  12 * will invoke an update, delete, or lookup operation. To ensure updates and
  13 * deletes appear atomic from the datapath side xchg() is used to modify the
  14 * netdev_map array. Then because the datapath does a lookup into the netdev_map
  15 * array (read-only) from an RCU critical section we use call_rcu() to wait for
  16 * an rcu grace period before free'ing the old data structures. This ensures the
  17 * datapath always has a valid copy. However, the datapath does a "flush"
  18 * operation that pushes any pending packets in the driver outside the RCU
  19 * critical section. Each bpf_dtab_netdev tracks these pending operations using
  20 * a per-cpu flush list. The bpf_dtab_netdev object will not be destroyed  until
  21 * this list is empty, indicating outstanding flush operations have completed.
  22 *
  23 * BPF syscalls may race with BPF program calls on any of the update, delete
  24 * or lookup operations. As noted above the xchg() operation also keep the
  25 * netdev_map consistent in this case. From the devmap side BPF programs
  26 * calling into these operations are the same as multiple user space threads
  27 * making system calls.
  28 *
  29 * Finally, any of the above may race with a netdev_unregister notifier. The
  30 * unregister notifier must search for net devices in the map structure that
  31 * contain a reference to the net device and remove them. This is a two step
  32 * process (a) dereference the bpf_dtab_netdev object in netdev_map and (b)
  33 * check to see if the ifindex is the same as the net_device being removed.
  34 * When removing the dev a cmpxchg() is used to ensure the correct dev is
  35 * removed, in the case of a concurrent update or delete operation it is
  36 * possible that the initially referenced dev is no longer in the map. As the
  37 * notifier hook walks the map we know that new dev references can not be
  38 * added by the user because core infrastructure ensures dev_get_by_index()
  39 * calls will fail at this point.
  40 *
  41 * The devmap_hash type is a map type which interprets keys as ifindexes and
  42 * indexes these using a hashmap. This allows maps that use ifindex as key to be
  43 * densely packed instead of having holes in the lookup array for unused
  44 * ifindexes. The setup and packet enqueue/send code is shared between the two
  45 * types of devmap; only the lookup and insertion is different.
  46 */
  47#include <linux/bpf.h>
  48#include <net/xdp.h>
  49#include <linux/filter.h>
  50#include <trace/events/xdp.h>
 
  51
  52#define DEV_CREATE_FLAG_MASK \
  53	(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
  54
  55struct xdp_dev_bulk_queue {
  56	struct xdp_frame *q[DEV_MAP_BULK_SIZE];
  57	struct list_head flush_node;
  58	struct net_device *dev;
  59	struct net_device *dev_rx;
  60	struct bpf_prog *xdp_prog;
  61	unsigned int count;
  62};
  63
  64struct bpf_dtab_netdev {
  65	struct net_device *dev; /* must be first member, due to tracepoint */
  66	struct hlist_node index_hlist;
  67	struct bpf_dtab *dtab;
  68	struct bpf_prog *xdp_prog;
  69	struct rcu_head rcu;
  70	unsigned int idx;
  71	struct bpf_devmap_val val;
  72};
  73
  74struct bpf_dtab {
  75	struct bpf_map map;
  76	struct bpf_dtab_netdev __rcu **netdev_map; /* DEVMAP type only */
  77	struct list_head list;
  78
  79	/* these are only used for DEVMAP_HASH type maps */
  80	struct hlist_head *dev_index_head;
  81	spinlock_t index_lock;
  82	unsigned int items;
  83	u32 n_buckets;
  84};
  85
  86static DEFINE_PER_CPU(struct list_head, dev_flush_list);
  87static DEFINE_SPINLOCK(dev_map_lock);
  88static LIST_HEAD(dev_map_list);
  89
  90static struct hlist_head *dev_map_create_hash(unsigned int entries,
  91					      int numa_node)
  92{
  93	int i;
  94	struct hlist_head *hash;
  95
  96	hash = bpf_map_area_alloc((u64) entries * sizeof(*hash), numa_node);
  97	if (hash != NULL)
  98		for (i = 0; i < entries; i++)
  99			INIT_HLIST_HEAD(&hash[i]);
 100
 101	return hash;
 102}
 103
 104static inline struct hlist_head *dev_map_index_hash(struct bpf_dtab *dtab,
 105						    int idx)
 106{
 107	return &dtab->dev_index_head[idx & (dtab->n_buckets - 1)];
 108}
 109
 110static int dev_map_init_map(struct bpf_dtab *dtab, union bpf_attr *attr)
 111{
 112	u32 valsize = attr->value_size;
 113
 114	/* check sanity of attributes. 2 value sizes supported:
 115	 * 4 bytes: ifindex
 116	 * 8 bytes: ifindex + prog fd
 117	 */
 118	if (attr->max_entries == 0 || attr->key_size != 4 ||
 119	    (valsize != offsetofend(struct bpf_devmap_val, ifindex) &&
 120	     valsize != offsetofend(struct bpf_devmap_val, bpf_prog.fd)) ||
 121	    attr->map_flags & ~DEV_CREATE_FLAG_MASK)
 122		return -EINVAL;
 123
 
 
 
 
 
 
 
 
 
 
 
 
 
 124	/* Lookup returns a pointer straight to dev->ifindex, so make sure the
 125	 * verifier prevents writes from the BPF side
 126	 */
 127	attr->map_flags |= BPF_F_RDONLY_PROG;
 128
 129
 130	bpf_map_init_from_attr(&dtab->map, attr);
 131
 132	if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 
 133		dtab->n_buckets = roundup_pow_of_two(dtab->map.max_entries);
 134
 135		if (!dtab->n_buckets) /* Overflow check */
 136			return -EINVAL;
 137	}
 138
 139	if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 140		dtab->dev_index_head = dev_map_create_hash(dtab->n_buckets,
 141							   dtab->map.numa_node);
 142		if (!dtab->dev_index_head)
 143			return -ENOMEM;
 144
 145		spin_lock_init(&dtab->index_lock);
 146	} else {
 147		dtab->netdev_map = bpf_map_area_alloc((u64) dtab->map.max_entries *
 148						      sizeof(struct bpf_dtab_netdev *),
 149						      dtab->map.numa_node);
 150		if (!dtab->netdev_map)
 151			return -ENOMEM;
 152	}
 153
 154	return 0;
 155}
 156
 157static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
 158{
 159	struct bpf_dtab *dtab;
 160	int err;
 161
 162	if (!capable(CAP_NET_ADMIN))
 163		return ERR_PTR(-EPERM);
 164
 165	dtab = kzalloc(sizeof(*dtab), GFP_USER | __GFP_ACCOUNT);
 166	if (!dtab)
 167		return ERR_PTR(-ENOMEM);
 168
 169	err = dev_map_init_map(dtab, attr);
 170	if (err) {
 171		kfree(dtab);
 172		return ERR_PTR(err);
 173	}
 174
 175	spin_lock(&dev_map_lock);
 176	list_add_tail_rcu(&dtab->list, &dev_map_list);
 177	spin_unlock(&dev_map_lock);
 178
 179	return &dtab->map;
 180}
 181
 182static void dev_map_free(struct bpf_map *map)
 183{
 184	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 185	int i;
 186
 187	/* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 188	 * so the programs (can be more than one that used this map) were
 189	 * disconnected from events. The following synchronize_rcu() guarantees
 190	 * both rcu read critical sections complete and waits for
 191	 * preempt-disable regions (NAPI being the relevant context here) so we
 192	 * are certain there will be no further reads against the netdev_map and
 193	 * all flush operations are complete. Flush operations can only be done
 194	 * from NAPI context for this reason.
 195	 */
 196
 197	spin_lock(&dev_map_lock);
 198	list_del_rcu(&dtab->list);
 199	spin_unlock(&dev_map_lock);
 200
 201	bpf_clear_redirect_map(map);
 
 
 
 
 
 
 
 202	synchronize_rcu();
 203
 204	/* Make sure prior __dev_map_entry_free() have completed. */
 205	rcu_barrier();
 206
 207	if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 208		for (i = 0; i < dtab->n_buckets; i++) {
 209			struct bpf_dtab_netdev *dev;
 210			struct hlist_head *head;
 211			struct hlist_node *next;
 212
 213			head = dev_map_index_hash(dtab, i);
 214
 215			hlist_for_each_entry_safe(dev, next, head, index_hlist) {
 216				hlist_del_rcu(&dev->index_hlist);
 217				if (dev->xdp_prog)
 218					bpf_prog_put(dev->xdp_prog);
 219				dev_put(dev->dev);
 220				kfree(dev);
 221			}
 222		}
 223
 224		bpf_map_area_free(dtab->dev_index_head);
 225	} else {
 226		for (i = 0; i < dtab->map.max_entries; i++) {
 227			struct bpf_dtab_netdev *dev;
 228
 229			dev = rcu_dereference_raw(dtab->netdev_map[i]);
 230			if (!dev)
 231				continue;
 232
 233			if (dev->xdp_prog)
 234				bpf_prog_put(dev->xdp_prog);
 235			dev_put(dev->dev);
 236			kfree(dev);
 237		}
 238
 239		bpf_map_area_free(dtab->netdev_map);
 240	}
 241
 242	kfree(dtab);
 243}
 244
 245static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 246{
 247	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 248	u32 index = key ? *(u32 *)key : U32_MAX;
 249	u32 *next = next_key;
 250
 251	if (index >= dtab->map.max_entries) {
 252		*next = 0;
 253		return 0;
 254	}
 255
 256	if (index == dtab->map.max_entries - 1)
 257		return -ENOENT;
 258	*next = index + 1;
 259	return 0;
 260}
 261
 262/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 263 * by local_bh_disable() (from XDP calls inside NAPI). The
 264 * rcu_read_lock_bh_held() below makes lockdep accept both.
 265 */
 266static void *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
 267{
 268	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 269	struct hlist_head *head = dev_map_index_hash(dtab, key);
 270	struct bpf_dtab_netdev *dev;
 271
 272	hlist_for_each_entry_rcu(dev, head, index_hlist,
 273				 lockdep_is_held(&dtab->index_lock))
 274		if (dev->idx == key)
 275			return dev;
 276
 277	return NULL;
 278}
 279
 280static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
 281				    void *next_key)
 282{
 283	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 284	u32 idx, *next = next_key;
 285	struct bpf_dtab_netdev *dev, *next_dev;
 286	struct hlist_head *head;
 287	int i = 0;
 288
 289	if (!key)
 290		goto find_first;
 291
 292	idx = *(u32 *)key;
 293
 294	dev = __dev_map_hash_lookup_elem(map, idx);
 295	if (!dev)
 296		goto find_first;
 297
 298	next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
 299				    struct bpf_dtab_netdev, index_hlist);
 300
 301	if (next_dev) {
 302		*next = next_dev->idx;
 303		return 0;
 304	}
 305
 306	i = idx & (dtab->n_buckets - 1);
 307	i++;
 308
 309 find_first:
 310	for (; i < dtab->n_buckets; i++) {
 311		head = dev_map_index_hash(dtab, i);
 312
 313		next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
 314					    struct bpf_dtab_netdev,
 315					    index_hlist);
 316		if (next_dev) {
 317			*next = next_dev->idx;
 318			return 0;
 319		}
 320	}
 321
 322	return -ENOENT;
 323}
 324
 325bool dev_map_can_have_prog(struct bpf_map *map)
 326{
 327	if ((map->map_type == BPF_MAP_TYPE_DEVMAP ||
 328	     map->map_type == BPF_MAP_TYPE_DEVMAP_HASH) &&
 329	    map->value_size != offsetofend(struct bpf_devmap_val, ifindex))
 330		return true;
 331
 332	return false;
 333}
 334
 335static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
 336				struct xdp_frame **frames, int n,
 337				struct net_device *dev)
 
 338{
 339	struct xdp_txq_info txq = { .dev = dev };
 
 340	struct xdp_buff xdp;
 341	int i, nframes = 0;
 342
 343	for (i = 0; i < n; i++) {
 344		struct xdp_frame *xdpf = frames[i];
 345		u32 act;
 346		int err;
 347
 348		xdp_convert_frame_to_buff(xdpf, &xdp);
 349		xdp.txq = &txq;
 
 350
 351		act = bpf_prog_run_xdp(xdp_prog, &xdp);
 352		switch (act) {
 353		case XDP_PASS:
 354			err = xdp_update_frame_from_buff(&xdp, xdpf);
 355			if (unlikely(err < 0))
 356				xdp_return_frame_rx_napi(xdpf);
 357			else
 358				frames[nframes++] = xdpf;
 359			break;
 360		default:
 361			bpf_warn_invalid_xdp_action(act);
 362			fallthrough;
 363		case XDP_ABORTED:
 364			trace_xdp_exception(dev, xdp_prog, act);
 365			fallthrough;
 366		case XDP_DROP:
 367			xdp_return_frame_rx_napi(xdpf);
 368			break;
 369		}
 370	}
 371	return nframes; /* sent frames count */
 372}
 373
 374static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
 375{
 376	struct net_device *dev = bq->dev;
 377	unsigned int cnt = bq->count;
 378	int sent = 0, err = 0;
 379	int to_send = cnt;
 380	int i;
 381
 382	if (unlikely(!cnt))
 383		return;
 384
 385	for (i = 0; i < cnt; i++) {
 386		struct xdp_frame *xdpf = bq->q[i];
 387
 388		prefetch(xdpf);
 389	}
 390
 391	if (bq->xdp_prog) {
 392		to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev);
 393		if (!to_send)
 394			goto out;
 395	}
 396
 397	sent = dev->netdev_ops->ndo_xdp_xmit(dev, to_send, bq->q, flags);
 398	if (sent < 0) {
 399		/* If ndo_xdp_xmit fails with an errno, no frames have
 400		 * been xmit'ed.
 401		 */
 402		err = sent;
 403		sent = 0;
 404	}
 405
 406	/* If not all frames have been transmitted, it is our
 407	 * responsibility to free them
 408	 */
 409	for (i = sent; unlikely(i < to_send); i++)
 410		xdp_return_frame_rx_napi(bq->q[i]);
 411
 412out:
 413	bq->count = 0;
 414	trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent, err);
 415}
 416
 417/* __dev_flush is called from xdp_do_flush() which _must_ be signalled from the
 418 * driver before returning from its napi->poll() routine. See the comment above
 419 * xdp_do_flush() in filter.c.
 420 */
 421void __dev_flush(void)
 422{
 423	struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
 424	struct xdp_dev_bulk_queue *bq, *tmp;
 425
 426	list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
 427		bq_xmit_all(bq, XDP_XMIT_FLUSH);
 428		bq->dev_rx = NULL;
 429		bq->xdp_prog = NULL;
 430		__list_del_clearprev(&bq->flush_node);
 431	}
 432}
 433
 434/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 435 * by local_bh_disable() (from XDP calls inside NAPI). The
 436 * rcu_read_lock_bh_held() below makes lockdep accept both.
 437 */
 438static void *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
 439{
 440	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 441	struct bpf_dtab_netdev *obj;
 442
 443	if (key >= map->max_entries)
 444		return NULL;
 445
 446	obj = rcu_dereference_check(dtab->netdev_map[key],
 447				    rcu_read_lock_bh_held());
 448	return obj;
 449}
 450
 451/* Runs in NAPI, i.e., softirq under local_bh_disable(). Thus, safe percpu
 452 * variable access, and map elements stick around. See comment above
 453 * xdp_do_flush() in filter.c.
 454 */
 455static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 456		       struct net_device *dev_rx, struct bpf_prog *xdp_prog)
 457{
 458	struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
 459	struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
 460
 461	if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
 462		bq_xmit_all(bq, 0);
 463
 464	/* Ingress dev_rx will be the same for all xdp_frame's in
 465	 * bulk_queue, because bq stored per-CPU and must be flushed
 466	 * from net_device drivers NAPI func end.
 467	 *
 468	 * Do the same with xdp_prog and flush_list since these fields
 469	 * are only ever modified together.
 470	 */
 471	if (!bq->dev_rx) {
 
 
 472		bq->dev_rx = dev_rx;
 473		bq->xdp_prog = xdp_prog;
 474		list_add(&bq->flush_node, flush_list);
 475	}
 476
 477	bq->q[bq->count++] = xdpf;
 478}
 479
 480static inline int __xdp_enqueue(struct net_device *dev, struct xdp_buff *xdp,
 481				struct net_device *dev_rx,
 482				struct bpf_prog *xdp_prog)
 483{
 484	struct xdp_frame *xdpf;
 485	int err;
 486
 487	if (!dev->netdev_ops->ndo_xdp_xmit)
 
 
 
 
 488		return -EOPNOTSUPP;
 489
 490	err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data);
 491	if (unlikely(err))
 492		return err;
 493
 494	xdpf = xdp_convert_buff_to_frame(xdp);
 495	if (unlikely(!xdpf))
 496		return -EOVERFLOW;
 497
 498	bq_enqueue(dev, xdpf, dev_rx, xdp_prog);
 499	return 0;
 500}
 501
 502int dev_xdp_enqueue(struct net_device *dev, struct xdp_buff *xdp,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 503		    struct net_device *dev_rx)
 504{
 505	return __xdp_enqueue(dev, xdp, dev_rx, NULL);
 506}
 507
 508int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
 509		    struct net_device *dev_rx)
 510{
 511	struct net_device *dev = dst->dev;
 512
 513	return __xdp_enqueue(dev, xdp, dev_rx, dst->xdp_prog);
 514}
 515
 516static bool is_valid_dst(struct bpf_dtab_netdev *obj, struct xdp_buff *xdp,
 517			 int exclude_ifindex)
 518{
 519	if (!obj || obj->dev->ifindex == exclude_ifindex ||
 520	    !obj->dev->netdev_ops->ndo_xdp_xmit)
 
 
 521		return false;
 522
 523	if (xdp_ok_fwd_dev(obj->dev, xdp->data_end - xdp->data))
 
 
 
 
 524		return false;
 525
 526	return true;
 527}
 528
 529static int dev_map_enqueue_clone(struct bpf_dtab_netdev *obj,
 530				 struct net_device *dev_rx,
 531				 struct xdp_frame *xdpf)
 532{
 533	struct xdp_frame *nxdpf;
 534
 535	nxdpf = xdpf_clone(xdpf);
 536	if (!nxdpf)
 537		return -ENOMEM;
 538
 539	bq_enqueue(obj->dev, nxdpf, dev_rx, obj->xdp_prog);
 540
 541	return 0;
 542}
 543
 544int dev_map_enqueue_multi(struct xdp_buff *xdp, struct net_device *dev_rx,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 545			  struct bpf_map *map, bool exclude_ingress)
 546{
 547	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 548	int exclude_ifindex = exclude_ingress ? dev_rx->ifindex : 0;
 549	struct bpf_dtab_netdev *dst, *last_dst = NULL;
 
 550	struct hlist_head *head;
 551	struct xdp_frame *xdpf;
 552	unsigned int i;
 553	int err;
 554
 555	xdpf = xdp_convert_buff_to_frame(xdp);
 556	if (unlikely(!xdpf))
 557		return -EOVERFLOW;
 
 558
 559	if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
 560		for (i = 0; i < map->max_entries; i++) {
 561			dst = rcu_dereference_check(dtab->netdev_map[i],
 562						    rcu_read_lock_bh_held());
 563			if (!is_valid_dst(dst, xdp, exclude_ifindex))
 
 
 
 564				continue;
 565
 566			/* we only need n-1 clones; last_dst enqueued below */
 567			if (!last_dst) {
 568				last_dst = dst;
 569				continue;
 570			}
 571
 572			err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
 573			if (err)
 574				return err;
 575
 576			last_dst = dst;
 577		}
 578	} else { /* BPF_MAP_TYPE_DEVMAP_HASH */
 579		for (i = 0; i < dtab->n_buckets; i++) {
 580			head = dev_map_index_hash(dtab, i);
 581			hlist_for_each_entry_rcu(dst, head, index_hlist,
 582						 lockdep_is_held(&dtab->index_lock)) {
 583				if (!is_valid_dst(dst, xdp, exclude_ifindex))
 
 
 
 
 584					continue;
 585
 586				/* we only need n-1 clones; last_dst enqueued below */
 587				if (!last_dst) {
 588					last_dst = dst;
 589					continue;
 590				}
 591
 592				err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
 593				if (err)
 594					return err;
 595
 596				last_dst = dst;
 597			}
 598		}
 599	}
 600
 601	/* consume the last copy of the frame */
 602	if (last_dst)
 603		bq_enqueue(last_dst->dev, xdpf, dev_rx, last_dst->xdp_prog);
 604	else
 605		xdp_return_frame_rx_napi(xdpf); /* dtab is empty */
 606
 607	return 0;
 608}
 609
 610int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
 611			     struct bpf_prog *xdp_prog)
 612{
 613	int err;
 614
 615	err = xdp_ok_fwd_dev(dst->dev, skb->len);
 616	if (unlikely(err))
 617		return err;
 
 
 
 
 
 
 
 
 618	skb->dev = dst->dev;
 619	generic_xdp_tx(skb, xdp_prog);
 620
 621	return 0;
 622}
 623
 624static int dev_map_redirect_clone(struct bpf_dtab_netdev *dst,
 625				  struct sk_buff *skb,
 626				  struct bpf_prog *xdp_prog)
 627{
 628	struct sk_buff *nskb;
 629	int err;
 630
 631	nskb = skb_clone(skb, GFP_ATOMIC);
 632	if (!nskb)
 633		return -ENOMEM;
 634
 635	err = dev_map_generic_redirect(dst, nskb, xdp_prog);
 636	if (unlikely(err)) {
 637		consume_skb(nskb);
 638		return err;
 639	}
 640
 641	return 0;
 642}
 643
 644int dev_map_redirect_multi(struct net_device *dev, struct sk_buff *skb,
 645			   struct bpf_prog *xdp_prog, struct bpf_map *map,
 646			   bool exclude_ingress)
 647{
 648	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 649	int exclude_ifindex = exclude_ingress ? dev->ifindex : 0;
 650	struct bpf_dtab_netdev *dst, *last_dst = NULL;
 
 651	struct hlist_head *head;
 652	struct hlist_node *next;
 
 653	unsigned int i;
 654	int err;
 655
 
 
 
 
 
 656	if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
 657		for (i = 0; i < map->max_entries; i++) {
 658			dst = rcu_dereference_check(dtab->netdev_map[i],
 659						    rcu_read_lock_bh_held());
 660			if (!dst || dst->dev->ifindex == exclude_ifindex)
 
 
 
 661				continue;
 662
 663			/* we only need n-1 clones; last_dst enqueued below */
 664			if (!last_dst) {
 665				last_dst = dst;
 666				continue;
 667			}
 668
 669			err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
 670			if (err)
 671				return err;
 672
 673			last_dst = dst;
 
 674		}
 675	} else { /* BPF_MAP_TYPE_DEVMAP_HASH */
 676		for (i = 0; i < dtab->n_buckets; i++) {
 677			head = dev_map_index_hash(dtab, i);
 678			hlist_for_each_entry_safe(dst, next, head, index_hlist) {
 679				if (!dst || dst->dev->ifindex == exclude_ifindex)
 
 680					continue;
 681
 682				/* we only need n-1 clones; last_dst enqueued below */
 683				if (!last_dst) {
 684					last_dst = dst;
 685					continue;
 686				}
 687
 688				err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
 689				if (err)
 690					return err;
 691
 692				last_dst = dst;
 693			}
 694		}
 695	}
 696
 697	/* consume the first skb and return */
 698	if (last_dst)
 699		return dev_map_generic_redirect(last_dst, skb, xdp_prog);
 700
 701	/* dtab is empty */
 702	consume_skb(skb);
 703	return 0;
 704}
 705
 706static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
 707{
 708	struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
 709
 710	return obj ? &obj->val : NULL;
 711}
 712
 713static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
 714{
 715	struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
 716								*(u32 *)key);
 717	return obj ? &obj->val : NULL;
 718}
 719
 720static void __dev_map_entry_free(struct rcu_head *rcu)
 721{
 722	struct bpf_dtab_netdev *dev;
 723
 724	dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
 725	if (dev->xdp_prog)
 726		bpf_prog_put(dev->xdp_prog);
 727	dev_put(dev->dev);
 728	kfree(dev);
 729}
 730
 731static int dev_map_delete_elem(struct bpf_map *map, void *key)
 732{
 733	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 734	struct bpf_dtab_netdev *old_dev;
 735	int k = *(u32 *)key;
 736
 737	if (k >= map->max_entries)
 738		return -EINVAL;
 739
 740	old_dev = unrcu_pointer(xchg(&dtab->netdev_map[k], NULL));
 741	if (old_dev)
 742		call_rcu(&old_dev->rcu, __dev_map_entry_free);
 
 
 743	return 0;
 744}
 745
 746static int dev_map_hash_delete_elem(struct bpf_map *map, void *key)
 747{
 748	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 749	struct bpf_dtab_netdev *old_dev;
 750	int k = *(u32 *)key;
 751	unsigned long flags;
 752	int ret = -ENOENT;
 753
 754	spin_lock_irqsave(&dtab->index_lock, flags);
 755
 756	old_dev = __dev_map_hash_lookup_elem(map, k);
 757	if (old_dev) {
 758		dtab->items--;
 759		hlist_del_init_rcu(&old_dev->index_hlist);
 760		call_rcu(&old_dev->rcu, __dev_map_entry_free);
 761		ret = 0;
 762	}
 763	spin_unlock_irqrestore(&dtab->index_lock, flags);
 764
 765	return ret;
 766}
 767
 768static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
 769						    struct bpf_dtab *dtab,
 770						    struct bpf_devmap_val *val,
 771						    unsigned int idx)
 772{
 773	struct bpf_prog *prog = NULL;
 774	struct bpf_dtab_netdev *dev;
 775
 776	dev = bpf_map_kmalloc_node(&dtab->map, sizeof(*dev),
 777				   GFP_ATOMIC | __GFP_NOWARN,
 778				   dtab->map.numa_node);
 779	if (!dev)
 780		return ERR_PTR(-ENOMEM);
 781
 782	dev->dev = dev_get_by_index(net, val->ifindex);
 783	if (!dev->dev)
 784		goto err_out;
 785
 786	if (val->bpf_prog.fd > 0) {
 787		prog = bpf_prog_get_type_dev(val->bpf_prog.fd,
 788					     BPF_PROG_TYPE_XDP, false);
 789		if (IS_ERR(prog))
 790			goto err_put_dev;
 791		if (prog->expected_attach_type != BPF_XDP_DEVMAP)
 
 792			goto err_put_prog;
 793	}
 794
 795	dev->idx = idx;
 796	dev->dtab = dtab;
 797	if (prog) {
 798		dev->xdp_prog = prog;
 799		dev->val.bpf_prog.id = prog->aux->id;
 800	} else {
 801		dev->xdp_prog = NULL;
 802		dev->val.bpf_prog.id = 0;
 803	}
 804	dev->val.ifindex = val->ifindex;
 805
 806	return dev;
 807err_put_prog:
 808	bpf_prog_put(prog);
 809err_put_dev:
 810	dev_put(dev->dev);
 811err_out:
 812	kfree(dev);
 813	return ERR_PTR(-EINVAL);
 814}
 815
 816static int __dev_map_update_elem(struct net *net, struct bpf_map *map,
 817				 void *key, void *value, u64 map_flags)
 818{
 819	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 820	struct bpf_dtab_netdev *dev, *old_dev;
 821	struct bpf_devmap_val val = {};
 822	u32 i = *(u32 *)key;
 823
 824	if (unlikely(map_flags > BPF_EXIST))
 825		return -EINVAL;
 826	if (unlikely(i >= dtab->map.max_entries))
 827		return -E2BIG;
 828	if (unlikely(map_flags == BPF_NOEXIST))
 829		return -EEXIST;
 830
 831	/* already verified value_size <= sizeof val */
 832	memcpy(&val, value, map->value_size);
 833
 834	if (!val.ifindex) {
 835		dev = NULL;
 836		/* can not specify fd if ifindex is 0 */
 837		if (val.bpf_prog.fd > 0)
 838			return -EINVAL;
 839	} else {
 840		dev = __dev_map_alloc_node(net, dtab, &val, i);
 841		if (IS_ERR(dev))
 842			return PTR_ERR(dev);
 843	}
 844
 845	/* Use call_rcu() here to ensure rcu critical sections have completed
 846	 * Remembering the driver side flush operation will happen before the
 847	 * net device is removed.
 848	 */
 849	old_dev = unrcu_pointer(xchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev)));
 850	if (old_dev)
 851		call_rcu(&old_dev->rcu, __dev_map_entry_free);
 
 
 852
 853	return 0;
 854}
 855
 856static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
 857			       u64 map_flags)
 858{
 859	return __dev_map_update_elem(current->nsproxy->net_ns,
 860				     map, key, value, map_flags);
 861}
 862
 863static int __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
 864				     void *key, void *value, u64 map_flags)
 865{
 866	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 867	struct bpf_dtab_netdev *dev, *old_dev;
 868	struct bpf_devmap_val val = {};
 869	u32 idx = *(u32 *)key;
 870	unsigned long flags;
 871	int err = -EEXIST;
 872
 873	/* already verified value_size <= sizeof val */
 874	memcpy(&val, value, map->value_size);
 875
 876	if (unlikely(map_flags > BPF_EXIST || !val.ifindex))
 877		return -EINVAL;
 878
 879	spin_lock_irqsave(&dtab->index_lock, flags);
 880
 881	old_dev = __dev_map_hash_lookup_elem(map, idx);
 882	if (old_dev && (map_flags & BPF_NOEXIST))
 883		goto out_err;
 884
 885	dev = __dev_map_alloc_node(net, dtab, &val, idx);
 886	if (IS_ERR(dev)) {
 887		err = PTR_ERR(dev);
 888		goto out_err;
 889	}
 890
 891	if (old_dev) {
 892		hlist_del_rcu(&old_dev->index_hlist);
 893	} else {
 894		if (dtab->items >= dtab->map.max_entries) {
 895			spin_unlock_irqrestore(&dtab->index_lock, flags);
 896			call_rcu(&dev->rcu, __dev_map_entry_free);
 897			return -E2BIG;
 898		}
 899		dtab->items++;
 900	}
 901
 902	hlist_add_head_rcu(&dev->index_hlist,
 903			   dev_map_index_hash(dtab, idx));
 904	spin_unlock_irqrestore(&dtab->index_lock, flags);
 905
 906	if (old_dev)
 907		call_rcu(&old_dev->rcu, __dev_map_entry_free);
 908
 909	return 0;
 910
 911out_err:
 912	spin_unlock_irqrestore(&dtab->index_lock, flags);
 913	return err;
 914}
 915
 916static int dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
 917				   u64 map_flags)
 918{
 919	return __dev_map_hash_update_elem(current->nsproxy->net_ns,
 920					 map, key, value, map_flags);
 921}
 922
 923static int dev_map_redirect(struct bpf_map *map, u32 ifindex, u64 flags)
 924{
 925	return __bpf_xdp_redirect_map(map, ifindex, flags,
 926				      BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
 927				      __dev_map_lookup_elem);
 928}
 929
 930static int dev_hash_map_redirect(struct bpf_map *map, u32 ifindex, u64 flags)
 931{
 932	return __bpf_xdp_redirect_map(map, ifindex, flags,
 933				      BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
 934				      __dev_map_hash_lookup_elem);
 935}
 936
 937static int dev_map_btf_id;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 938const struct bpf_map_ops dev_map_ops = {
 939	.map_meta_equal = bpf_map_meta_equal,
 
 940	.map_alloc = dev_map_alloc,
 941	.map_free = dev_map_free,
 942	.map_get_next_key = dev_map_get_next_key,
 943	.map_lookup_elem = dev_map_lookup_elem,
 944	.map_update_elem = dev_map_update_elem,
 945	.map_delete_elem = dev_map_delete_elem,
 946	.map_check_btf = map_check_no_btf,
 947	.map_btf_name = "bpf_dtab",
 948	.map_btf_id = &dev_map_btf_id,
 949	.map_redirect = dev_map_redirect,
 950};
 951
 952static int dev_map_hash_map_btf_id;
 953const struct bpf_map_ops dev_map_hash_ops = {
 954	.map_meta_equal = bpf_map_meta_equal,
 
 955	.map_alloc = dev_map_alloc,
 956	.map_free = dev_map_free,
 957	.map_get_next_key = dev_map_hash_get_next_key,
 958	.map_lookup_elem = dev_map_hash_lookup_elem,
 959	.map_update_elem = dev_map_hash_update_elem,
 960	.map_delete_elem = dev_map_hash_delete_elem,
 961	.map_check_btf = map_check_no_btf,
 962	.map_btf_name = "bpf_dtab",
 963	.map_btf_id = &dev_map_hash_map_btf_id,
 964	.map_redirect = dev_hash_map_redirect,
 965};
 966
 967static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
 968				       struct net_device *netdev)
 969{
 970	unsigned long flags;
 971	u32 i;
 972
 973	spin_lock_irqsave(&dtab->index_lock, flags);
 974	for (i = 0; i < dtab->n_buckets; i++) {
 975		struct bpf_dtab_netdev *dev;
 976		struct hlist_head *head;
 977		struct hlist_node *next;
 978
 979		head = dev_map_index_hash(dtab, i);
 980
 981		hlist_for_each_entry_safe(dev, next, head, index_hlist) {
 982			if (netdev != dev->dev)
 983				continue;
 984
 985			dtab->items--;
 986			hlist_del_rcu(&dev->index_hlist);
 987			call_rcu(&dev->rcu, __dev_map_entry_free);
 988		}
 989	}
 990	spin_unlock_irqrestore(&dtab->index_lock, flags);
 991}
 992
 993static int dev_map_notification(struct notifier_block *notifier,
 994				ulong event, void *ptr)
 995{
 996	struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
 997	struct bpf_dtab *dtab;
 998	int i, cpu;
 999
1000	switch (event) {
1001	case NETDEV_REGISTER:
1002		if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
1003			break;
1004
1005		/* will be freed in free_netdev() */
1006		netdev->xdp_bulkq = alloc_percpu(struct xdp_dev_bulk_queue);
1007		if (!netdev->xdp_bulkq)
1008			return NOTIFY_BAD;
1009
1010		for_each_possible_cpu(cpu)
1011			per_cpu_ptr(netdev->xdp_bulkq, cpu)->dev = netdev;
1012		break;
1013	case NETDEV_UNREGISTER:
1014		/* This rcu_read_lock/unlock pair is needed because
1015		 * dev_map_list is an RCU list AND to ensure a delete
1016		 * operation does not free a netdev_map entry while we
1017		 * are comparing it against the netdev being unregistered.
1018		 */
1019		rcu_read_lock();
1020		list_for_each_entry_rcu(dtab, &dev_map_list, list) {
1021			if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
1022				dev_map_hash_remove_netdev(dtab, netdev);
1023				continue;
1024			}
1025
1026			for (i = 0; i < dtab->map.max_entries; i++) {
1027				struct bpf_dtab_netdev *dev, *odev;
1028
1029				dev = rcu_dereference(dtab->netdev_map[i]);
1030				if (!dev || netdev != dev->dev)
1031					continue;
1032				odev = unrcu_pointer(cmpxchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev), NULL));
1033				if (dev == odev)
1034					call_rcu(&dev->rcu,
1035						 __dev_map_entry_free);
 
 
1036			}
1037		}
1038		rcu_read_unlock();
1039		break;
1040	default:
1041		break;
1042	}
1043	return NOTIFY_OK;
1044}
1045
1046static struct notifier_block dev_map_notifier = {
1047	.notifier_call = dev_map_notification,
1048};
1049
1050static int __init dev_map_init(void)
1051{
1052	int cpu;
1053
1054	/* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
1055	BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
1056		     offsetof(struct _bpf_dtab_netdev, dev));
1057	register_netdevice_notifier(&dev_map_notifier);
1058
1059	for_each_possible_cpu(cpu)
1060		INIT_LIST_HEAD(&per_cpu(dev_flush_list, cpu));
1061	return 0;
1062}
1063
1064subsys_initcall(dev_map_init);