Lines Matching +full:no +full:- +full:read +full:- +full:rollover
1 // SPDX-License-Identifier: GPL-2.0-or-later
7 * PACKET - implements raw packet sockets.
15 * Alan Cox : new skbuff lists, look ma no backlogs!
18 * added. Also fixed the peek/read crash
22 * Alan Cox : Re-commented the code.
27 * dubious gcc output. Can you read
30 * Alan Cox : New buffers. Use sk->mac.raw.
96 - If the device has no dev->header_ops->create, there is no LL header
103 needed_headroom to be (the real WiFi header length - the fake Ethernet
105 - packet socket receives packets with pulled ll header,
109 -----------
112 mac_header -> ll header
113 data -> data
116 mac_header -> ll header
117 data -> ll header
120 mac_header -> data
123 data -> data
126 mac_header -> data. ll header is invisible to us.
127 data -> data
135 ------------
137 dev->header_ops != NULL
138 mac_header -> ll header
139 data -> ll header
141 dev->header_ops == NULL (ll header is invisible to us)
142 mac_header -> data
143 data -> data
178 #define BLOCK_STATUS(x) ((x)->hdr.bh1.block_status)
179 #define BLOCK_NUM_PKTS(x) ((x)->hdr.bh1.num_pkts)
180 #define BLOCK_O2FP(x) ((x)->hdr.bh1.offset_to_first_pkt)
181 #define BLOCK_LEN(x) ((x)->hdr.bh1.blk_len)
182 #define BLOCK_SNUM(x) ((x)->hdr.bh1.seq_num)
183 #define BLOCK_O2PRIV(x) ((x)->offset_to_priv)
227 #define PACKET_SKB_CB(__skb) ((struct packet_skb_cb *)((__skb)->cb))
229 #define GET_PBDQC_FROM_RB(x) ((struct tpacket_kbdq_core *)(&(x)->prb_bdqc))
231 ((struct tpacket_block_desc *)((x)->pkbdq[(bid)].buffer))
233 ((struct tpacket_block_desc *)((x)->pkbdq[(x)->kactive_blk_num].buffer))
235 (((x)->kactive_blk_num < ((x)->knum_blocks-1)) ? \
236 ((x)->kactive_blk_num+1) : 0)
251 dev = rcu_dereference(po->cached_dev); in packet_cached_dev_get()
262 rcu_assign_pointer(po->cached_dev, dev); in packet_cached_dev_assign()
267 RCU_INIT_POINTER(po->cached_dev, NULL); in packet_cached_dev_reset()
273 return READ_ONCE(po->xmit) == packet_direct_xmit; in packet_use_direct_xmit()
278 struct net_device *dev = skb->dev; in packet_pick_tx_queue()
279 const struct net_device_ops *ops = dev->netdev_ops; in packet_pick_tx_queue()
284 skb->sender_cpu = cpu + 1; in packet_pick_tx_queue()
286 skb_record_rx_queue(skb, cpu % dev->real_num_tx_queues); in packet_pick_tx_queue()
287 if (ops->ndo_select_queue) { in packet_pick_tx_queue()
288 queue_index = ops->ndo_select_queue(dev, skb, NULL); in packet_pick_tx_queue()
305 if (!po->running) { in __register_prot_hook()
306 if (po->fanout) in __register_prot_hook()
309 dev_add_pack(&po->prot_hook); in __register_prot_hook()
312 po->running = 1; in __register_prot_hook()
318 lockdep_assert_held_once(&pkt_sk(sk)->bind_lock); in register_prot_hook()
323 * the po->bind_lock and do a synchronize_net to make sure no
325 * of po->prot_hook. If the sync parameter is false, it is the
332 lockdep_assert_held_once(&po->bind_lock); in __unregister_prot_hook()
334 po->running = 0; in __unregister_prot_hook()
336 if (po->fanout) in __unregister_prot_hook()
339 __dev_remove_pack(&po->prot_hook); in __unregister_prot_hook()
344 spin_unlock(&po->bind_lock); in __unregister_prot_hook()
346 spin_lock(&po->bind_lock); in __unregister_prot_hook()
354 if (po->running) in unregister_prot_hook()
372 switch (po->tp_version) { in __packet_set_status()
374 WRITE_ONCE(h.h1->tp_status, status); in __packet_set_status()
375 flush_dcache_page(pgv_to_page(&h.h1->tp_status)); in __packet_set_status()
378 WRITE_ONCE(h.h2->tp_status, status); in __packet_set_status()
379 flush_dcache_page(pgv_to_page(&h.h2->tp_status)); in __packet_set_status()
382 WRITE_ONCE(h.h3->tp_status, status); in __packet_set_status()
383 flush_dcache_page(pgv_to_page(&h.h3->tp_status)); in __packet_set_status()
402 switch (po->tp_version) { in __packet_get_status()
404 flush_dcache_page(pgv_to_page(&h.h1->tp_status)); in __packet_get_status()
405 return READ_ONCE(h.h1->tp_status); in __packet_get_status()
407 flush_dcache_page(pgv_to_page(&h.h2->tp_status)); in __packet_get_status()
408 return READ_ONCE(h.h2->tp_status); in __packet_get_status()
410 flush_dcache_page(pgv_to_page(&h.h3->tp_status)); in __packet_get_status()
411 return READ_ONCE(h.h3->tp_status); in __packet_get_status()
426 ktime_to_timespec64_cond(shhwtstamps->hwtstamp, ts)) in tpacket_get_timestamp()
430 ktime_to_timespec64_cond(skb->tstamp, ts)) in tpacket_get_timestamp()
443 if (!(ts_status = tpacket_get_timestamp(skb, &ts, po->tp_tstamp))) in __packet_set_timestamp()
449 * all store the seconds in a 32-bit unsigned integer. in __packet_set_timestamp()
450 * If we create a version 4, that should have a 64-bit timestamp, in __packet_set_timestamp()
451 * either 64-bit seconds + 32-bit nanoseconds, or just 64-bit in __packet_set_timestamp()
454 switch (po->tp_version) { in __packet_set_timestamp()
456 h.h1->tp_sec = ts.tv_sec; in __packet_set_timestamp()
457 h.h1->tp_usec = ts.tv_nsec / NSEC_PER_USEC; in __packet_set_timestamp()
460 h.h2->tp_sec = ts.tv_sec; in __packet_set_timestamp()
461 h.h2->tp_nsec = ts.tv_nsec; in __packet_set_timestamp()
464 h.h3->tp_sec = ts.tv_sec; in __packet_set_timestamp()
465 h.h3->tp_nsec = ts.tv_nsec; in __packet_set_timestamp()
473 flush_dcache_page(pgv_to_page(&h.h1->tp_sec)); in __packet_set_timestamp()
487 pg_vec_pos = position / rb->frames_per_block; in packet_lookup_frame()
488 frame_offset = position % rb->frames_per_block; in packet_lookup_frame()
490 h.raw = rb->pg_vec[pg_vec_pos].buffer + in packet_lookup_frame()
491 (frame_offset * rb->frame_size); in packet_lookup_frame()
503 return packet_lookup_frame(po, rb, rb->head, status); in packet_current_frame()
508 del_timer_sync(&pkc->retire_blk_timer); in prb_del_retire_blk_timer()
516 pkc = GET_PBDQC_FROM_RB(&po->rx_ring); in prb_shutdown_retire_blk_timer()
518 spin_lock_bh(&rb_queue->lock); in prb_shutdown_retire_blk_timer()
519 pkc->delete_blk_timer = 1; in prb_shutdown_retire_blk_timer()
520 spin_unlock_bh(&rb_queue->lock); in prb_shutdown_retire_blk_timer()
529 pkc = GET_PBDQC_FROM_RB(&po->rx_ring); in prb_setup_retire_blk_timer()
530 timer_setup(&pkc->retire_blk_timer, prb_retire_rx_blk_timer_expired, in prb_setup_retire_blk_timer()
532 pkc->retire_blk_timer.expires = jiffies; in prb_setup_retire_blk_timer()
544 dev = __dev_get_by_index(sock_net(&po->sk), po->ifindex); in prb_calc_retire_blk_tmo()
575 p1->feature_req_word = req_u->req3.tp_feature_req_word; in prb_init_ft_ops()
588 p1->knxt_seq_num = 1; in init_prb_bdqc()
589 p1->pkbdq = pg_vec; in init_prb_bdqc()
591 p1->pkblk_start = pg_vec[0].buffer; in init_prb_bdqc()
592 p1->kblk_size = req_u->req3.tp_block_size; in init_prb_bdqc()
593 p1->knum_blocks = req_u->req3.tp_block_nr; in init_prb_bdqc()
594 p1->hdrlen = po->tp_hdrlen; in init_prb_bdqc()
595 p1->version = po->tp_version; in init_prb_bdqc()
596 p1->last_kactive_blk_num = 0; in init_prb_bdqc()
597 po->stats.stats3.tp_freeze_q_cnt = 0; in init_prb_bdqc()
598 if (req_u->req3.tp_retire_blk_tov) in init_prb_bdqc()
599 p1->retire_blk_tov = req_u->req3.tp_retire_blk_tov; in init_prb_bdqc()
601 p1->retire_blk_tov = prb_calc_retire_blk_tmo(po, in init_prb_bdqc()
602 req_u->req3.tp_block_size); in init_prb_bdqc()
603 p1->tov_in_jiffies = msecs_to_jiffies(p1->retire_blk_tov); in init_prb_bdqc()
604 p1->blk_sizeof_priv = req_u->req3.tp_sizeof_priv; in init_prb_bdqc()
605 rwlock_init(&p1->blk_fill_in_prog_lock); in init_prb_bdqc()
607 p1->max_frame_len = p1->kblk_size - BLK_PLUS_PRIV(p1->blk_sizeof_priv); in init_prb_bdqc()
618 mod_timer(&pkc->retire_blk_timer, in _prb_refresh_rx_retire_blk_timer()
619 jiffies + pkc->tov_in_jiffies); in _prb_refresh_rx_retire_blk_timer()
620 pkc->last_kactive_blk_num = pkc->kactive_blk_num; in _prb_refresh_rx_retire_blk_timer()
627 * on packet-by-packet basis.
629 * With a 1MB block-size, on a 1Gbps line, it will take
642 * a) line-speed and b) block-size.
650 struct tpacket_kbdq_core *pkc = GET_PBDQC_FROM_RB(&po->rx_ring); in prb_retire_rx_blk_timer_expired()
654 spin_lock(&po->sk.sk_receive_queue.lock); in prb_retire_rx_blk_timer_expired()
659 if (unlikely(pkc->delete_blk_timer)) in prb_retire_rx_blk_timer_expired()
673 write_lock(&pkc->blk_fill_in_prog_lock); in prb_retire_rx_blk_timer_expired()
674 write_unlock(&pkc->blk_fill_in_prog_lock); in prb_retire_rx_blk_timer_expired()
677 if (pkc->last_kactive_blk_num == pkc->kactive_blk_num) { in prb_retire_rx_blk_timer_expired()
689 /* Case 1. Queue was frozen because user-space was in prb_retire_rx_blk_timer_expired()
694 * Ok, user-space is still behind. in prb_retire_rx_blk_timer_expired()
699 /* Case 2. queue was frozen,user-space caught up, in prb_retire_rx_blk_timer_expired()
704 * Thawing/timer-refresh is a side effect. in prb_retire_rx_blk_timer_expired()
716 spin_unlock(&po->sk.sk_receive_queue.lock); in prb_retire_rx_blk_timer_expired()
732 end = (u8 *)PAGE_ALIGN((unsigned long)pkc1->pkblk_end); in prb_flush_block()
769 struct tpacket_hdr_v1 *h1 = &pbd1->hdr.bh1; in prb_close_block()
770 struct sock *sk = &po->sk; in prb_close_block()
772 if (atomic_read(&po->tp_drops)) in prb_close_block()
775 last_pkt = (struct tpacket3_hdr *)pkc1->prev; in prb_close_block()
776 last_pkt->tp_next_offset = 0; in prb_close_block()
780 h1->ts_last_pkt.ts_sec = last_pkt->tp_sec; in prb_close_block()
781 h1->ts_last_pkt.ts_nsec = last_pkt->tp_nsec; in prb_close_block()
783 /* Ok, we tmo'd - so get the current time. in prb_close_block()
790 h1->ts_last_pkt.ts_sec = ts.tv_sec; in prb_close_block()
791 h1->ts_last_pkt.ts_nsec = ts.tv_nsec; in prb_close_block()
799 sk->sk_data_ready(sk); in prb_close_block()
801 pkc1->kactive_blk_num = GET_NEXT_PRB_BLK_NUM(pkc1); in prb_close_block()
806 pkc->reset_pending_on_curr_blk = 0; in prb_thaw_queue()
820 struct tpacket_hdr_v1 *h1 = &pbd1->hdr.bh1; in prb_open_block()
828 BLOCK_SNUM(pbd1) = pkc1->knxt_seq_num++; in prb_open_block()
830 BLOCK_LEN(pbd1) = BLK_PLUS_PRIV(pkc1->blk_sizeof_priv); in prb_open_block()
834 h1->ts_first_pkt.ts_sec = ts.tv_sec; in prb_open_block()
835 h1->ts_first_pkt.ts_nsec = ts.tv_nsec; in prb_open_block()
837 pkc1->pkblk_start = (char *)pbd1; in prb_open_block()
838 pkc1->nxt_offset = pkc1->pkblk_start + BLK_PLUS_PRIV(pkc1->blk_sizeof_priv); in prb_open_block()
840 BLOCK_O2FP(pbd1) = (__u32)BLK_PLUS_PRIV(pkc1->blk_sizeof_priv); in prb_open_block()
843 pbd1->version = pkc1->version; in prb_open_block()
844 pkc1->prev = pkc1->nxt_offset; in prb_open_block()
845 pkc1->pkblk_end = pkc1->pkblk_start + pkc1->kblk_size; in prb_open_block()
858 * 4) user-space is either sleeping or processing block '0'.
859 * 5) tpacket_rcv is currently filling block '7', since there is no space left,
860 * it will close block-7,loop around and try to fill block '0'.
861 * call-flow:
865 * |->(BLOCK_STATUS == USER) evaluates to true
866 * 5.1) Since block-0 is currently in-use, we just freeze the queue.
871 * re-open block-0 in near future.
873 * case and __packet_lookup_frame_in_block will check if block-0
874 * is free and can now be re-used.
879 pkc->reset_pending_on_curr_blk = 1; in prb_freeze_queue()
880 po->stats.stats3.tp_freeze_q_cnt++; in prb_freeze_queue()
913 return (void *)pkc->nxt_offset; in prb_dispatch_next_block()
925 * cpu-0 and tpacket_rcv() got invoked on cpu-1, didn't in prb_retire_current_block()
930 * the timer-handler already handled this case. in prb_retire_current_block()
934 write_lock(&pkc->blk_fill_in_prog_lock); in prb_retire_current_block()
935 write_unlock(&pkc->blk_fill_in_prog_lock); in prb_retire_current_block()
949 return pkc->reset_pending_on_curr_blk; in prb_queue_frozen()
953 __releases(&pkc->blk_fill_in_prog_lock) in prb_clear_blk_fill_status()
957 read_unlock(&pkc->blk_fill_in_prog_lock); in prb_clear_blk_fill_status()
963 ppd->hv1.tp_rxhash = skb_get_hash(pkc->skb); in prb_fill_rxhash()
969 ppd->hv1.tp_rxhash = 0; in prb_clear_rxhash()
975 if (skb_vlan_tag_present(pkc->skb)) { in prb_fill_vlan_info()
976 ppd->hv1.tp_vlan_tci = skb_vlan_tag_get(pkc->skb); in prb_fill_vlan_info()
977 ppd->hv1.tp_vlan_tpid = ntohs(pkc->skb->vlan_proto); in prb_fill_vlan_info()
978 ppd->tp_status = TP_STATUS_VLAN_VALID | TP_STATUS_VLAN_TPID_VALID; in prb_fill_vlan_info()
980 ppd->hv1.tp_vlan_tci = 0; in prb_fill_vlan_info()
981 ppd->hv1.tp_vlan_tpid = 0; in prb_fill_vlan_info()
982 ppd->tp_status = TP_STATUS_AVAILABLE; in prb_fill_vlan_info()
989 ppd->hv1.tp_padding = 0; in prb_run_all_ft_ops()
992 if (pkc->feature_req_word & TP_FT_REQ_FILL_RXHASH) in prb_run_all_ft_ops()
1002 __acquires(&pkc->blk_fill_in_prog_lock) in prb_fill_curr_block()
1007 ppd->tp_next_offset = TOTAL_PKT_LEN_INCL_ALIGN(len); in prb_fill_curr_block()
1008 pkc->prev = curr; in prb_fill_curr_block()
1009 pkc->nxt_offset += TOTAL_PKT_LEN_INCL_ALIGN(len); in prb_fill_curr_block()
1012 read_lock(&pkc->blk_fill_in_prog_lock); in prb_fill_curr_block()
1016 /* Assumes caller has the sk->rx_queue.lock */
1026 pkc = GET_PBDQC_FROM_RB(&po->rx_ring); in __packet_lookup_frame_in_block()
1033 * is still in_use by user-space. in __packet_lookup_frame_in_block()
1040 * Ok, the block was released by user-space. in __packet_lookup_frame_in_block()
1050 curr = pkc->nxt_offset; in __packet_lookup_frame_in_block()
1051 pkc->skb = skb; in __packet_lookup_frame_in_block()
1052 end = (char *)pbd + pkc->kblk_size; in __packet_lookup_frame_in_block()
1072 * No free blocks are available.user_space hasn't caught up yet. in __packet_lookup_frame_in_block()
1083 switch (po->tp_version) { in packet_current_rx_frame()
1086 curr = packet_lookup_frame(po, &po->rx_ring, in packet_current_rx_frame()
1087 po->rx_ring.head, status); in packet_current_rx_frame()
1114 if (rb->prb_bdqc.kactive_blk_num) in prb_previous_blk_num()
1115 prev = rb->prb_bdqc.kactive_blk_num-1; in prb_previous_blk_num()
1117 prev = rb->prb_bdqc.knum_blocks-1; in prb_previous_blk_num()
1134 if (po->tp_version <= TPACKET_V2) in packet_previous_rx_frame()
1143 switch (po->tp_version) { in packet_increment_rx_head()
1159 unsigned int previous = rb->head ? rb->head - 1 : rb->frame_max; in packet_previous_frame()
1165 buff->head = buff->head != buff->frame_max ? buff->head+1 : 0; in packet_increment_head()
1170 this_cpu_inc(*rb->pending_refcnt); in packet_inc_pending()
1175 this_cpu_dec(*rb->pending_refcnt); in packet_dec_pending()
1184 if (rb->pending_refcnt == NULL) in packet_read_pending()
1188 refcnt += *per_cpu_ptr(rb->pending_refcnt, cpu); in packet_read_pending()
1195 po->rx_ring.pending_refcnt = NULL; in packet_alloc_pending()
1197 po->tx_ring.pending_refcnt = alloc_percpu(unsigned int); in packet_alloc_pending()
1198 if (unlikely(po->tx_ring.pending_refcnt == NULL)) in packet_alloc_pending()
1199 return -ENOBUFS; in packet_alloc_pending()
1206 free_percpu(po->tx_ring.pending_refcnt); in packet_free_pending()
1218 len = READ_ONCE(po->rx_ring.frame_max) + 1; in __tpacket_has_room()
1219 idx = READ_ONCE(po->rx_ring.head); in __tpacket_has_room()
1223 idx -= len; in __tpacket_has_room()
1224 return packet_lookup_frame(po, &po->rx_ring, idx, TP_STATUS_KERNEL); in __tpacket_has_room()
1231 len = READ_ONCE(po->rx_ring.prb_bdqc.knum_blocks); in __tpacket_v3_has_room()
1232 idx = READ_ONCE(po->rx_ring.prb_bdqc.kactive_blk_num); in __tpacket_v3_has_room()
1236 idx -= len; in __tpacket_v3_has_room()
1237 return prb_lookup_block(po, &po->rx_ring, idx, TP_STATUS_KERNEL); in __tpacket_v3_has_room()
1243 const struct sock *sk = &po->sk; in __packet_rcv_has_room()
1246 if (po->prot_hook.func != tpacket_rcv) { in __packet_rcv_has_room()
1247 int rcvbuf = READ_ONCE(sk->sk_rcvbuf); in __packet_rcv_has_room()
1248 int avail = rcvbuf - atomic_read(&sk->sk_rmem_alloc) in __packet_rcv_has_room()
1249 - (skb ? skb->truesize : 0); in __packet_rcv_has_room()
1259 if (po->tp_version == TPACKET_V3) { in __packet_rcv_has_room()
1281 if (READ_ONCE(po->pressure) != pressure) in packet_rcv_has_room()
1282 WRITE_ONCE(po->pressure, pressure); in packet_rcv_has_room()
1289 if (READ_ONCE(po->pressure) && in packet_rcv_try_clear_pressure()
1291 WRITE_ONCE(po->pressure, 0); in packet_rcv_try_clear_pressure()
1296 skb_queue_purge(&sk->sk_error_queue); in packet_sock_destruct()
1298 WARN_ON(atomic_read(&sk->sk_rmem_alloc)); in packet_sock_destruct()
1299 WARN_ON(refcount_read(&sk->sk_wmem_alloc)); in packet_sock_destruct()
1311 u32 *history = po->rollover->history; in fanout_flow_is_huge()
1340 unsigned int val = atomic_inc_return(&f->rr_cur); in fanout_demux_lb()
1367 po = pkt_sk(rcu_dereference(f->arr[idx])); in fanout_demux_rollover()
1377 i = j = min_t(int, po->rollover->sock, num - 1); in fanout_demux_rollover()
1379 po_next = pkt_sk(rcu_dereference(f->arr[i])); in fanout_demux_rollover()
1380 if (po_next != po_skip && !READ_ONCE(po_next->pressure) && in fanout_demux_rollover()
1383 po->rollover->sock = i; in fanout_demux_rollover()
1384 atomic_long_inc(&po->rollover->num); in fanout_demux_rollover()
1386 atomic_long_inc(&po->rollover->num_huge); in fanout_demux_rollover()
1394 atomic_long_inc(&po->rollover->num_failed); in fanout_demux_rollover()
1413 prog = rcu_dereference(f->bpf_prog); in fanout_demux_bpf()
1423 return f->flags & (flag >> 8); in fanout_has_flag()
1429 struct packet_fanout *f = pt->af_packet_priv; in packet_rcv_fanout()
1430 unsigned int num = READ_ONCE(f->num_members); in packet_rcv_fanout()
1431 struct net *net = read_pnet(&f->net); in packet_rcv_fanout()
1445 switch (f->type) { in packet_rcv_fanout()
1474 po = pkt_sk(rcu_dereference(f->arr[idx])); in packet_rcv_fanout()
1475 return po->prot_hook.func(skb, dev, &po->prot_hook, orig_dev); in packet_rcv_fanout()
1485 struct packet_fanout *f = po->fanout; in __fanout_link()
1487 spin_lock(&f->lock); in __fanout_link()
1488 rcu_assign_pointer(f->arr[f->num_members], sk); in __fanout_link()
1490 f->num_members++; in __fanout_link()
1491 if (f->num_members == 1) in __fanout_link()
1492 dev_add_pack(&f->prot_hook); in __fanout_link()
1493 spin_unlock(&f->lock); in __fanout_link()
1498 struct packet_fanout *f = po->fanout; in __fanout_unlink()
1501 spin_lock(&f->lock); in __fanout_unlink()
1502 for (i = 0; i < f->num_members; i++) { in __fanout_unlink()
1503 if (rcu_dereference_protected(f->arr[i], in __fanout_unlink()
1504 lockdep_is_held(&f->lock)) == sk) in __fanout_unlink()
1507 BUG_ON(i >= f->num_members); in __fanout_unlink()
1508 rcu_assign_pointer(f->arr[i], in __fanout_unlink()
1509 rcu_dereference_protected(f->arr[f->num_members - 1], in __fanout_unlink()
1510 lockdep_is_held(&f->lock))); in __fanout_unlink()
1511 f->num_members--; in __fanout_unlink()
1512 if (f->num_members == 0) in __fanout_unlink()
1513 __dev_remove_pack(&f->prot_hook); in __fanout_unlink()
1514 spin_unlock(&f->lock); in __fanout_unlink()
1519 if (sk->sk_family != PF_PACKET) in match_fanout_group()
1522 return ptype->af_packet_priv == pkt_sk(sk)->fanout; in match_fanout_group()
1527 switch (f->type) { in fanout_init_data()
1529 atomic_set(&f->rr_cur, 0); in fanout_init_data()
1533 RCU_INIT_POINTER(f->bpf_prog, NULL); in fanout_init_data()
1542 spin_lock(&f->lock); in __fanout_set_data_bpf()
1543 old = rcu_dereference_protected(f->bpf_prog, lockdep_is_held(&f->lock)); in __fanout_set_data_bpf()
1544 rcu_assign_pointer(f->bpf_prog, new); in __fanout_set_data_bpf()
1545 spin_unlock(&f->lock); in __fanout_set_data_bpf()
1560 if (sock_flag(&po->sk, SOCK_FILTER_LOCKED)) in fanout_set_data_cbpf()
1561 return -EPERM; in fanout_set_data_cbpf()
1571 __fanout_set_data_bpf(po->fanout, new); in fanout_set_data_cbpf()
1581 if (sock_flag(&po->sk, SOCK_FILTER_LOCKED)) in fanout_set_data_ebpf()
1582 return -EPERM; in fanout_set_data_ebpf()
1584 return -EINVAL; in fanout_set_data_ebpf()
1586 return -EFAULT; in fanout_set_data_ebpf()
1592 __fanout_set_data_bpf(po->fanout, new); in fanout_set_data_ebpf()
1599 switch (po->fanout->type) { in fanout_set_data()
1605 return -EINVAL; in fanout_set_data()
1611 switch (f->type) { in fanout_release_data()
1623 if (f->id == candidate_id && in __fanout_id_is_free()
1624 read_pnet(&f->net) == sock_net(sk)) { in __fanout_id_is_free()
1650 struct packet_rollover *rollover = NULL; in fanout_add() local
1652 u16 type_flags = args->type_flags; in fanout_add()
1656 u16 id = args->id; in fanout_add()
1662 return -EINVAL; in fanout_add()
1672 return -EINVAL; in fanout_add()
1677 err = -EALREADY; in fanout_add()
1678 if (po->fanout) in fanout_add()
1683 err = -ENOMEM; in fanout_add()
1684 rollover = kzalloc(sizeof(*rollover), GFP_KERNEL); in fanout_add()
1685 if (!rollover) in fanout_add()
1687 atomic_long_set(&rollover->num, 0); in fanout_add()
1688 atomic_long_set(&rollover->num_huge, 0); in fanout_add()
1689 atomic_long_set(&rollover->num_failed, 0); in fanout_add()
1694 err = -EINVAL; in fanout_add()
1698 err = -ENOMEM; in fanout_add()
1707 if (f->id == id && in fanout_add()
1708 read_pnet(&f->net) == sock_net(sk)) { in fanout_add()
1713 err = -EINVAL; in fanout_add()
1715 if (match->flags != flags) in fanout_add()
1717 if (args->max_num_members && in fanout_add()
1718 args->max_num_members != match->max_num_members) in fanout_add()
1721 if (args->max_num_members > PACKET_FANOUT_MAX) in fanout_add()
1723 if (!args->max_num_members) in fanout_add()
1725 args->max_num_members = 256; in fanout_add()
1726 err = -ENOMEM; in fanout_add()
1727 match = kvzalloc(struct_size(match, arr, args->max_num_members), in fanout_add()
1731 write_pnet(&match->net, sock_net(sk)); in fanout_add()
1732 match->id = id; in fanout_add()
1733 match->type = type; in fanout_add()
1734 match->flags = flags; in fanout_add()
1735 INIT_LIST_HEAD(&match->list); in fanout_add()
1736 spin_lock_init(&match->lock); in fanout_add()
1737 refcount_set(&match->sk_ref, 0); in fanout_add()
1739 match->prot_hook.type = po->prot_hook.type; in fanout_add()
1740 match->prot_hook.dev = po->prot_hook.dev; in fanout_add()
1741 match->prot_hook.func = packet_rcv_fanout; in fanout_add()
1742 match->prot_hook.af_packet_priv = match; in fanout_add()
1743 match->prot_hook.af_packet_net = read_pnet(&match->net); in fanout_add()
1744 match->prot_hook.id_match = match_fanout_group; in fanout_add()
1745 match->max_num_members = args->max_num_members; in fanout_add()
1746 list_add(&match->list, &fanout_list); in fanout_add()
1748 err = -EINVAL; in fanout_add()
1750 spin_lock(&po->bind_lock); in fanout_add()
1751 if (po->running && in fanout_add()
1752 match->type == type && in fanout_add()
1753 match->prot_hook.type == po->prot_hook.type && in fanout_add()
1754 match->prot_hook.dev == po->prot_hook.dev) { in fanout_add()
1755 err = -ENOSPC; in fanout_add()
1756 if (refcount_read(&match->sk_ref) < match->max_num_members) { in fanout_add()
1757 __dev_remove_pack(&po->prot_hook); in fanout_add()
1760 WRITE_ONCE(po->fanout, match); in fanout_add()
1762 po->rollover = rollover; in fanout_add()
1763 rollover = NULL; in fanout_add()
1764 refcount_set(&match->sk_ref, refcount_read(&match->sk_ref) + 1); in fanout_add()
1769 spin_unlock(&po->bind_lock); in fanout_add()
1771 if (err && !refcount_read(&match->sk_ref)) { in fanout_add()
1772 list_del(&match->list); in fanout_add()
1777 kfree(rollover); in fanout_add()
1782 /* If pkt_sk(sk)->fanout->sk_ref is zero, this function removes
1783 * pkt_sk(sk)->fanout from fanout_list and returns pkt_sk(sk)->fanout.
1793 f = po->fanout; in fanout_release()
1795 po->fanout = NULL; in fanout_release()
1797 if (refcount_dec_and_test(&f->sk_ref)) in fanout_release()
1798 list_del(&f->list); in fanout_release()
1810 /* Earlier code assumed this would be a VLAN pkt, double-check in packet_extra_vlan_len_allowed()
1814 if (unlikely(dev->type != ARPHRD_ETHER)) in packet_extra_vlan_len_allowed()
1818 return likely(eth_hdr(skb)->h_proto == htons(ETH_P_8021Q)); in packet_extra_vlan_len_allowed()
1836 sk = pt->af_packet_priv; in packet_rcv_spkt()
1845 * For outgoing ones skb->data == skb_mac_header(skb) in packet_rcv_spkt()
1849 if (skb->pkt_type == PACKET_LOOPBACK) in packet_rcv_spkt()
1865 spkt = &PACKET_SKB_CB(skb)->sa.pkt; in packet_rcv_spkt()
1867 skb_push(skb, skb->data - skb_mac_header(skb)); in packet_rcv_spkt()
1873 spkt->spkt_family = dev->type; in packet_rcv_spkt()
1874 strlcpy(spkt->spkt_device, dev->name, sizeof(spkt->spkt_device)); in packet_rcv_spkt()
1875 spkt->spkt_protocol = skb->protocol; in packet_rcv_spkt()
1895 if ((!skb->protocol || skb->protocol == htons(ETH_P_ALL)) && in packet_parse_headers()
1896 sock->type == SOCK_RAW) { in packet_parse_headers()
1898 skb->protocol = dev_parse_header_protocol(skb); in packet_parse_headers()
1902 if (likely(skb->dev->type == ARPHRD_ETHER) && in packet_parse_headers()
1903 eth_type_vlan(skb->protocol) && in packet_parse_headers()
1904 vlan_get_protocol_and_depth(skb, skb->protocol, &depth) != 0) in packet_parse_headers()
1918 struct sock *sk = sock->sk; in packet_sendmsg_spkt()
1919 DECLARE_SOCKADDR(struct sockaddr_pkt *, saddr, msg->msg_name); in packet_sendmsg_spkt()
1932 if (msg->msg_namelen < sizeof(struct sockaddr)) in packet_sendmsg_spkt()
1933 return -EINVAL; in packet_sendmsg_spkt()
1934 if (msg->msg_namelen == sizeof(struct sockaddr_pkt)) in packet_sendmsg_spkt()
1935 proto = saddr->spkt_protocol; in packet_sendmsg_spkt()
1937 return -ENOTCONN; /* SOCK_PACKET must be sent giving an address */ in packet_sendmsg_spkt()
1943 saddr->spkt_device[sizeof(saddr->spkt_device) - 1] = 0; in packet_sendmsg_spkt()
1946 dev = dev_get_by_name_rcu(sock_net(sk), saddr->spkt_device); in packet_sendmsg_spkt()
1947 err = -ENODEV; in packet_sendmsg_spkt()
1951 err = -ENETDOWN; in packet_sendmsg_spkt()
1952 if (!(dev->flags & IFF_UP)) in packet_sendmsg_spkt()
1962 err = -EPROTONOSUPPORT; in packet_sendmsg_spkt()
1968 err = -EMSGSIZE; in packet_sendmsg_spkt()
1969 if (len > dev->mtu + dev->hard_header_len + VLAN_HLEN + extra_len) in packet_sendmsg_spkt()
1974 int tlen = dev->needed_tailroom; in packet_sendmsg_spkt()
1975 unsigned int hhlen = dev->header_ops ? dev->hard_header_len : 0; in packet_sendmsg_spkt()
1980 return -ENOBUFS; in packet_sendmsg_spkt()
1990 skb->data -= hhlen; in packet_sendmsg_spkt()
1991 skb->tail -= hhlen; in packet_sendmsg_spkt()
2001 if (!dev_validate_header(dev, skb->data, len) || !skb->len) { in packet_sendmsg_spkt()
2002 err = -EINVAL; in packet_sendmsg_spkt()
2005 if (len > (dev->mtu + dev->hard_header_len + extra_len) && in packet_sendmsg_spkt()
2007 err = -EMSGSIZE; in packet_sendmsg_spkt()
2012 if (msg->msg_controllen) { in packet_sendmsg_spkt()
2018 skb->protocol = proto; in packet_sendmsg_spkt()
2019 skb->dev = dev; in packet_sendmsg_spkt()
2020 skb->priority = sk->sk_priority; in packet_sendmsg_spkt()
2021 skb->mark = sk->sk_mark; in packet_sendmsg_spkt()
2022 skb->tstamp = sockc.transmit_time; in packet_sendmsg_spkt()
2027 skb->no_fcs = 1; in packet_sendmsg_spkt()
2049 filter = rcu_dereference(sk->sk_filter); in run_filter()
2051 res = bpf_prog_run_clear_cb(filter->prog, skb); in run_filter()
2063 return -EINVAL; in packet_rcv_vnet()
2064 *len -= sizeof(vnet_hdr); in packet_rcv_vnet()
2067 return -EINVAL; in packet_rcv_vnet()
2076 * Note tricky part: we DO mangle shared skb! skb->data, skb->len
2077 * and skb->cb are mangled. It works because (and until) packets
2090 u8 *skb_head = skb->data; in packet_rcv()
2091 int skb_len = skb->len; in packet_rcv()
2095 if (skb->pkt_type == PACKET_LOOPBACK) in packet_rcv()
2098 sk = pt->af_packet_priv; in packet_rcv()
2104 skb->dev = dev; in packet_rcv()
2114 if (sk->sk_type != SOCK_DGRAM) in packet_rcv()
2115 skb_push(skb, skb->data - skb_mac_header(skb)); in packet_rcv()
2116 else if (skb->pkt_type == PACKET_OUTGOING) { in packet_rcv()
2122 snaplen = skb->len; in packet_rcv()
2130 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) in packet_rcv()
2138 if (skb_head != skb->data) { in packet_rcv()
2139 skb->data = skb_head; in packet_rcv()
2140 skb->len = skb_len; in packet_rcv()
2146 sock_skb_cb_check_size(sizeof(*PACKET_SKB_CB(skb)) + MAX_ADDR_LEN - 8); in packet_rcv()
2148 sll = &PACKET_SKB_CB(skb)->sa.ll; in packet_rcv()
2149 sll->sll_hatype = dev->type; in packet_rcv()
2150 sll->sll_pkttype = skb->pkt_type; in packet_rcv()
2152 sll->sll_ifindex = orig_dev->ifindex; in packet_rcv()
2154 sll->sll_ifindex = dev->ifindex; in packet_rcv()
2156 sll->sll_halen = dev_parse_header(skb, sll->sll_addr); in packet_rcv()
2158 /* sll->sll_family and sll->sll_protocol are set in packet_recvmsg(). in packet_rcv()
2161 PACKET_SKB_CB(skb)->sa.origlen = skb->len; in packet_rcv()
2167 skb->dev = NULL; in packet_rcv()
2173 spin_lock(&sk->sk_receive_queue.lock); in packet_rcv()
2174 po->stats.stats1.tp_packets++; in packet_rcv()
2176 __skb_queue_tail(&sk->sk_receive_queue, skb); in packet_rcv()
2177 spin_unlock(&sk->sk_receive_queue.lock); in packet_rcv()
2178 sk->sk_data_ready(sk); in packet_rcv()
2183 atomic_inc(&po->tp_drops); in packet_rcv()
2184 atomic_inc(&sk->sk_drops); in packet_rcv()
2187 if (skb_head != skb->data && skb_shared(skb)) { in packet_rcv()
2188 skb->data = skb_head; in packet_rcv()
2189 skb->len = skb_len; in packet_rcv()
2206 u8 *skb_head = skb->data; in tpacket_rcv()
2207 int skb_len = skb->len; in tpacket_rcv()
2226 if (skb->pkt_type == PACKET_LOOPBACK) in tpacket_rcv()
2229 sk = pt->af_packet_priv; in tpacket_rcv()
2236 if (sk->sk_type != SOCK_DGRAM) in tpacket_rcv()
2237 skb_push(skb, skb->data - skb_mac_header(skb)); in tpacket_rcv()
2238 else if (skb->pkt_type == PACKET_OUTGOING) { in tpacket_rcv()
2244 snaplen = skb->len; in tpacket_rcv()
2252 atomic_inc(&po->tp_drops); in tpacket_rcv()
2256 if (skb->ip_summed == CHECKSUM_PARTIAL) in tpacket_rcv()
2258 else if (skb->pkt_type != PACKET_OUTGOING && in tpacket_rcv()
2265 if (sk->sk_type == SOCK_DGRAM) { in tpacket_rcv()
2266 macoff = netoff = TPACKET_ALIGN(po->tp_hdrlen) + 16 + in tpacket_rcv()
2267 po->tp_reserve; in tpacket_rcv()
2270 netoff = TPACKET_ALIGN(po->tp_hdrlen + in tpacket_rcv()
2272 po->tp_reserve; in tpacket_rcv()
2273 if (po->has_vnet_hdr) { in tpacket_rcv()
2277 macoff = netoff - maclen; in tpacket_rcv()
2280 atomic_inc(&po->tp_drops); in tpacket_rcv()
2283 if (po->tp_version <= TPACKET_V2) { in tpacket_rcv()
2284 if (macoff + snaplen > po->rx_ring.frame_size) { in tpacket_rcv()
2285 if (po->copy_thresh && in tpacket_rcv()
2286 atomic_read(&sk->sk_rmem_alloc) < sk->sk_rcvbuf) { in tpacket_rcv()
2291 skb_head = skb->data; in tpacket_rcv()
2294 memset(&PACKET_SKB_CB(copy_skb)->sa.ll, 0, in tpacket_rcv()
2295 sizeof(PACKET_SKB_CB(copy_skb)->sa.ll)); in tpacket_rcv()
2299 snaplen = po->rx_ring.frame_size - macoff; in tpacket_rcv()
2306 GET_PBDQC_FROM_RB(&po->rx_ring)->max_frame_len)) { in tpacket_rcv()
2309 nval = GET_PBDQC_FROM_RB(&po->rx_ring)->max_frame_len - macoff; in tpacket_rcv()
2315 macoff = GET_PBDQC_FROM_RB(&po->rx_ring)->max_frame_len; in tpacket_rcv()
2319 spin_lock(&sk->sk_receive_queue.lock); in tpacket_rcv()
2325 if (po->tp_version <= TPACKET_V2) { in tpacket_rcv()
2326 slot_id = po->rx_ring.head; in tpacket_rcv()
2327 if (test_bit(slot_id, po->rx_ring.rx_owner_map)) in tpacket_rcv()
2329 __set_bit(slot_id, po->rx_ring.rx_owner_map); in tpacket_rcv()
2333 virtio_net_hdr_from_skb(skb, h.raw + macoff - in tpacket_rcv()
2336 if (po->tp_version == TPACKET_V3) in tpacket_rcv()
2337 prb_clear_blk_fill_status(&po->rx_ring); in tpacket_rcv()
2341 if (po->tp_version <= TPACKET_V2) { in tpacket_rcv()
2342 packet_increment_rx_head(po, &po->rx_ring); in tpacket_rcv()
2344 * LOSING will be reported till you read the stats, in tpacket_rcv()
2345 * because it's COR - Clear On Read. in tpacket_rcv()
2349 if (atomic_read(&po->tp_drops)) in tpacket_rcv()
2353 po->stats.stats1.tp_packets++; in tpacket_rcv()
2356 __skb_queue_tail(&sk->sk_receive_queue, copy_skb); in tpacket_rcv()
2358 spin_unlock(&sk->sk_receive_queue.lock); in tpacket_rcv()
2366 po->tp_tstamp | SOF_TIMESTAMPING_SOFTWARE); in tpacket_rcv()
2372 switch (po->tp_version) { in tpacket_rcv()
2374 h.h1->tp_len = skb->len; in tpacket_rcv()
2375 h.h1->tp_snaplen = snaplen; in tpacket_rcv()
2376 h.h1->tp_mac = macoff; in tpacket_rcv()
2377 h.h1->tp_net = netoff; in tpacket_rcv()
2378 h.h1->tp_sec = ts.tv_sec; in tpacket_rcv()
2379 h.h1->tp_usec = ts.tv_nsec / NSEC_PER_USEC; in tpacket_rcv()
2383 h.h2->tp_len = skb->len; in tpacket_rcv()
2384 h.h2->tp_snaplen = snaplen; in tpacket_rcv()
2385 h.h2->tp_mac = macoff; in tpacket_rcv()
2386 h.h2->tp_net = netoff; in tpacket_rcv()
2387 h.h2->tp_sec = ts.tv_sec; in tpacket_rcv()
2388 h.h2->tp_nsec = ts.tv_nsec; in tpacket_rcv()
2390 h.h2->tp_vlan_tci = skb_vlan_tag_get(skb); in tpacket_rcv()
2391 h.h2->tp_vlan_tpid = ntohs(skb->vlan_proto); in tpacket_rcv()
2394 h.h2->tp_vlan_tci = 0; in tpacket_rcv()
2395 h.h2->tp_vlan_tpid = 0; in tpacket_rcv()
2397 memset(h.h2->tp_padding, 0, sizeof(h.h2->tp_padding)); in tpacket_rcv()
2404 h.h3->tp_status |= status; in tpacket_rcv()
2405 h.h3->tp_len = skb->len; in tpacket_rcv()
2406 h.h3->tp_snaplen = snaplen; in tpacket_rcv()
2407 h.h3->tp_mac = macoff; in tpacket_rcv()
2408 h.h3->tp_net = netoff; in tpacket_rcv()
2409 h.h3->tp_sec = ts.tv_sec; in tpacket_rcv()
2410 h.h3->tp_nsec = ts.tv_nsec; in tpacket_rcv()
2411 memset(h.h3->tp_padding, 0, sizeof(h.h3->tp_padding)); in tpacket_rcv()
2419 sll->sll_halen = dev_parse_header(skb, sll->sll_addr); in tpacket_rcv()
2420 sll->sll_family = AF_PACKET; in tpacket_rcv()
2421 sll->sll_hatype = dev->type; in tpacket_rcv()
2422 sll->sll_protocol = skb->protocol; in tpacket_rcv()
2423 sll->sll_pkttype = skb->pkt_type; in tpacket_rcv()
2425 sll->sll_ifindex = orig_dev->ifindex; in tpacket_rcv()
2427 sll->sll_ifindex = dev->ifindex; in tpacket_rcv()
2432 if (po->tp_version <= TPACKET_V2) { in tpacket_rcv()
2444 if (po->tp_version <= TPACKET_V2) { in tpacket_rcv()
2445 spin_lock(&sk->sk_receive_queue.lock); in tpacket_rcv()
2447 __clear_bit(slot_id, po->rx_ring.rx_owner_map); in tpacket_rcv()
2448 spin_unlock(&sk->sk_receive_queue.lock); in tpacket_rcv()
2449 sk->sk_data_ready(sk); in tpacket_rcv()
2450 } else if (po->tp_version == TPACKET_V3) { in tpacket_rcv()
2451 prb_clear_blk_fill_status(&po->rx_ring); in tpacket_rcv()
2455 if (skb_head != skb->data && skb_shared(skb)) { in tpacket_rcv()
2456 skb->data = skb_head; in tpacket_rcv()
2457 skb->len = skb_len; in tpacket_rcv()
2467 spin_unlock(&sk->sk_receive_queue.lock); in tpacket_rcv()
2468 atomic_inc(&po->tp_drops); in tpacket_rcv()
2471 sk->sk_data_ready(sk); in tpacket_rcv()
2478 struct packet_sock *po = pkt_sk(skb->sk); in tpacket_destruct_skb()
2480 if (likely(po->tx_ring.pg_vec)) { in tpacket_destruct_skb()
2485 packet_dec_pending(&po->tx_ring); in tpacket_destruct_skb()
2490 if (!packet_read_pending(&po->tx_ring)) in tpacket_destruct_skb()
2491 complete(&po->skb_completion); in tpacket_destruct_skb()
2499 if ((vnet_hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) && in __packet_snd_vnet_parse()
2500 (__virtio16_to_cpu(vio_le(), vnet_hdr->csum_start) + in __packet_snd_vnet_parse()
2501 __virtio16_to_cpu(vio_le(), vnet_hdr->csum_offset) + 2 > in __packet_snd_vnet_parse()
2502 __virtio16_to_cpu(vio_le(), vnet_hdr->hdr_len))) in __packet_snd_vnet_parse()
2503 vnet_hdr->hdr_len = __cpu_to_virtio16(vio_le(), in __packet_snd_vnet_parse()
2504 __virtio16_to_cpu(vio_le(), vnet_hdr->csum_start) + in __packet_snd_vnet_parse()
2505 __virtio16_to_cpu(vio_le(), vnet_hdr->csum_offset) + 2); in __packet_snd_vnet_parse()
2507 if (__virtio16_to_cpu(vio_le(), vnet_hdr->hdr_len) > len) in __packet_snd_vnet_parse()
2508 return -EINVAL; in __packet_snd_vnet_parse()
2517 return -EINVAL; in packet_snd_vnet_parse()
2518 *len -= sizeof(*vnet_hdr); in packet_snd_vnet_parse()
2520 if (!copy_from_iter_full(vnet_hdr, sizeof(*vnet_hdr), &msg->msg_iter)) in packet_snd_vnet_parse()
2521 return -EFAULT; in packet_snd_vnet_parse()
2533 struct socket *sock = po->sk.sk_socket; in tpacket_fill_skb()
2539 skb->protocol = proto; in tpacket_fill_skb()
2540 skb->dev = dev; in tpacket_fill_skb()
2541 skb->priority = po->sk.sk_priority; in tpacket_fill_skb()
2542 skb->mark = po->sk.sk_mark; in tpacket_fill_skb()
2543 skb->tstamp = sockc->transmit_time; in tpacket_fill_skb()
2544 skb_setup_tx_timestamp(skb, sockc->tsflags); in tpacket_fill_skb()
2552 if (sock->type == SOCK_DGRAM) { in tpacket_fill_skb()
2556 return -EINVAL; in tpacket_fill_skb()
2560 skb_push(skb, dev->hard_header_len); in tpacket_fill_skb()
2561 skb_put(skb, copylen - dev->hard_header_len); in tpacket_fill_skb()
2565 if (!dev_validate_header(dev, skb->data, hdrlen)) in tpacket_fill_skb()
2566 return -EINVAL; in tpacket_fill_skb()
2569 to_write -= hdrlen; in tpacket_fill_skb()
2573 len_max = PAGE_SIZE - offset; in tpacket_fill_skb()
2576 skb->data_len = to_write; in tpacket_fill_skb()
2577 skb->len += to_write; in tpacket_fill_skb()
2578 skb->truesize += to_write; in tpacket_fill_skb()
2579 refcount_add(to_write, &po->sk.sk_wmem_alloc); in tpacket_fill_skb()
2582 nr_frags = skb_shinfo(skb)->nr_frags; in tpacket_fill_skb()
2587 return -EFAULT; in tpacket_fill_skb()
2595 to_write -= len; in tpacket_fill_skb()
2614 switch (po->tp_version) { in tpacket_parse_header()
2616 if (ph.h3->tp_next_offset != 0) { in tpacket_parse_header()
2618 return -EINVAL; in tpacket_parse_header()
2620 tp_len = ph.h3->tp_len; in tpacket_parse_header()
2623 tp_len = ph.h2->tp_len; in tpacket_parse_header()
2626 tp_len = ph.h1->tp_len; in tpacket_parse_header()
2631 return -EMSGSIZE; in tpacket_parse_header()
2634 if (unlikely(po->tp_tx_has_off)) { in tpacket_parse_header()
2637 off_min = po->tp_hdrlen - sizeof(struct sockaddr_ll); in tpacket_parse_header()
2638 off_max = po->tx_ring.frame_size - tp_len; in tpacket_parse_header()
2639 if (po->sk.sk_type == SOCK_DGRAM) { in tpacket_parse_header()
2640 switch (po->tp_version) { in tpacket_parse_header()
2642 off = ph.h3->tp_net; in tpacket_parse_header()
2645 off = ph.h2->tp_net; in tpacket_parse_header()
2648 off = ph.h1->tp_net; in tpacket_parse_header()
2652 switch (po->tp_version) { in tpacket_parse_header()
2654 off = ph.h3->tp_mac; in tpacket_parse_header()
2657 off = ph.h2->tp_mac; in tpacket_parse_header()
2660 off = ph.h1->tp_mac; in tpacket_parse_header()
2665 return -EINVAL; in tpacket_parse_header()
2667 off = po->tp_hdrlen - sizeof(struct sockaddr_ll); in tpacket_parse_header()
2683 DECLARE_SOCKADDR(struct sockaddr_ll *, saddr, msg->msg_name); in tpacket_snd()
2684 bool need_wait = !(msg->msg_flags & MSG_DONTWAIT); in tpacket_snd()
2693 mutex_lock(&po->pg_vec_lock); in tpacket_snd()
2698 if (unlikely(!po->tx_ring.pg_vec)) { in tpacket_snd()
2699 err = -EBUSY; in tpacket_snd()
2704 proto = READ_ONCE(po->num); in tpacket_snd()
2706 err = -EINVAL; in tpacket_snd()
2707 if (msg->msg_namelen < sizeof(struct sockaddr_ll)) in tpacket_snd()
2709 if (msg->msg_namelen < (saddr->sll_halen in tpacket_snd()
2713 proto = saddr->sll_protocol; in tpacket_snd()
2714 dev = dev_get_by_index(sock_net(&po->sk), saddr->sll_ifindex); in tpacket_snd()
2715 if (po->sk.sk_socket->type == SOCK_DGRAM) { in tpacket_snd()
2716 if (dev && msg->msg_namelen < dev->addr_len + in tpacket_snd()
2719 addr = saddr->sll_addr; in tpacket_snd()
2723 err = -ENXIO; in tpacket_snd()
2726 err = -ENETDOWN; in tpacket_snd()
2727 if (unlikely(!(dev->flags & IFF_UP))) in tpacket_snd()
2730 sockcm_init(&sockc, &po->sk); in tpacket_snd()
2731 if (msg->msg_controllen) { in tpacket_snd()
2732 err = sock_cmsg_send(&po->sk, msg, &sockc); in tpacket_snd()
2737 if (po->sk.sk_socket->type == SOCK_RAW) in tpacket_snd()
2738 reserve = dev->hard_header_len; in tpacket_snd()
2739 size_max = po->tx_ring.frame_size in tpacket_snd()
2740 - (po->tp_hdrlen - sizeof(struct sockaddr_ll)); in tpacket_snd()
2742 if ((size_max > dev->mtu + reserve + VLAN_HLEN) && !po->has_vnet_hdr) in tpacket_snd()
2743 size_max = dev->mtu + reserve + VLAN_HLEN; in tpacket_snd()
2745 reinit_completion(&po->skb_completion); in tpacket_snd()
2748 ph = packet_current_frame(po, &po->tx_ring, in tpacket_snd()
2752 timeo = sock_sndtimeo(&po->sk, msg->msg_flags & MSG_DONTWAIT); in tpacket_snd()
2753 timeo = wait_for_completion_interruptible_timeout(&po->skb_completion, timeo); in tpacket_snd()
2755 err = !timeo ? -ETIMEDOUT : -ERESTARTSYS; in tpacket_snd()
2770 tlen = dev->needed_tailroom; in tpacket_snd()
2771 if (po->has_vnet_hdr) { in tpacket_snd()
2774 tp_len -= sizeof(*vnet_hdr); in tpacket_snd()
2777 tp_len = -EINVAL; in tpacket_snd()
2781 vnet_hdr->hdr_len); in tpacket_snd()
2783 copylen = max_t(int, copylen, dev->hard_header_len); in tpacket_snd()
2784 skb = sock_alloc_send_skb(&po->sk, in tpacket_snd()
2786 (copylen - dev->hard_header_len), in tpacket_snd()
2798 tp_len > dev->mtu + reserve && in tpacket_snd()
2799 !po->has_vnet_hdr && in tpacket_snd()
2801 tp_len = -EMSGSIZE; in tpacket_snd()
2805 if (po->tp_loss) { in tpacket_snd()
2808 packet_increment_head(&po->tx_ring); in tpacket_snd()
2818 if (po->has_vnet_hdr) { in tpacket_snd()
2820 tp_len = -EINVAL; in tpacket_snd()
2826 skb->destructor = tpacket_destruct_skb; in tpacket_snd()
2828 packet_inc_pending(&po->tx_ring); in tpacket_snd()
2832 err = READ_ONCE(po->xmit)(skb); in tpacket_snd()
2848 packet_increment_head(&po->tx_ring); in tpacket_snd()
2852 * to call it as it's per_cpu variable, but in fast-path in tpacket_snd()
2853 * we already short-circuit the loop with the first in tpacket_snd()
2857 (need_wait && packet_read_pending(&po->tx_ring)))); in tpacket_snd()
2868 mutex_unlock(&po->pg_vec_lock); in tpacket_snd()
2883 skb = sock_alloc_send_pskb(sk, prepad + linear, len - linear, noblock, in packet_alloc_skb()
2890 skb->data_len = len - linear; in packet_alloc_skb()
2891 skb->len += len - linear; in packet_alloc_skb()
2898 struct sock *sk = sock->sk; in packet_snd()
2899 DECLARE_SOCKADDR(struct sockaddr_ll *, saddr, msg->msg_name); in packet_snd()
2919 proto = READ_ONCE(po->num); in packet_snd()
2921 err = -EINVAL; in packet_snd()
2922 if (msg->msg_namelen < sizeof(struct sockaddr_ll)) in packet_snd()
2924 if (msg->msg_namelen < (saddr->sll_halen + offsetof(struct sockaddr_ll, sll_addr))) in packet_snd()
2926 proto = saddr->sll_protocol; in packet_snd()
2927 dev = dev_get_by_index(sock_net(sk), saddr->sll_ifindex); in packet_snd()
2928 if (sock->type == SOCK_DGRAM) { in packet_snd()
2929 if (dev && msg->msg_namelen < dev->addr_len + in packet_snd()
2932 addr = saddr->sll_addr; in packet_snd()
2936 err = -ENXIO; in packet_snd()
2939 err = -ENETDOWN; in packet_snd()
2940 if (unlikely(!(dev->flags & IFF_UP))) in packet_snd()
2944 sockc.mark = sk->sk_mark; in packet_snd()
2945 if (msg->msg_controllen) { in packet_snd()
2951 if (sock->type == SOCK_RAW) in packet_snd()
2952 reserve = dev->hard_header_len; in packet_snd()
2953 if (po->has_vnet_hdr) { in packet_snd()
2962 err = -EPROTONOSUPPORT; in packet_snd()
2968 err = -EMSGSIZE; in packet_snd()
2970 (len > dev->mtu + reserve + VLAN_HLEN + extra_len)) in packet_snd()
2973 err = -ENOBUFS; in packet_snd()
2975 tlen = dev->needed_tailroom; in packet_snd()
2977 linear = max(linear, min_t(int, len, dev->hard_header_len)); in packet_snd()
2979 msg->msg_flags & MSG_DONTWAIT, &err); in packet_snd()
2985 err = -EINVAL; in packet_snd()
2986 if (sock->type == SOCK_DGRAM) { in packet_snd()
2991 skb_reserve(skb, -reserve); in packet_snd()
2993 dev->min_header_len != dev->hard_header_len) in packet_snd()
2997 /* Returns -EFAULT on error */ in packet_snd()
2998 err = skb_copy_datagram_from_iter(skb, offset, &msg->msg_iter, len); in packet_snd()
3002 if ((sock->type == SOCK_RAW && in packet_snd()
3003 !dev_validate_header(dev, skb->data, len)) || !skb->len) { in packet_snd()
3004 err = -EINVAL; in packet_snd()
3010 if (!vnet_hdr.gso_type && (len > dev->mtu + reserve + extra_len) && in packet_snd()
3012 err = -EMSGSIZE; in packet_snd()
3016 skb->protocol = proto; in packet_snd()
3017 skb->dev = dev; in packet_snd()
3018 skb->priority = sk->sk_priority; in packet_snd()
3019 skb->mark = sockc.mark; in packet_snd()
3020 skb->tstamp = sockc.transmit_time; in packet_snd()
3023 skb->no_fcs = 1; in packet_snd()
3036 err = READ_ONCE(po->xmit)(skb); in packet_snd()
3059 struct sock *sk = sock->sk; in packet_sendmsg()
3065 if (data_race(po->tx_ring.pg_vec)) in packet_sendmsg()
3078 struct sock *sk = sock->sk; in packet_release()
3090 mutex_lock(&net->packet.sklist_lock); in packet_release()
3092 mutex_unlock(&net->packet.sklist_lock); in packet_release()
3095 sock_prot_inuse_add(net, sk->sk_prot, -1); in packet_release()
3098 spin_lock(&po->bind_lock); in packet_release()
3102 if (po->prot_hook.dev) { in packet_release()
3103 dev_put(po->prot_hook.dev); in packet_release()
3104 po->prot_hook.dev = NULL; in packet_release()
3106 spin_unlock(&po->bind_lock); in packet_release()
3111 if (po->rx_ring.pg_vec) { in packet_release()
3116 if (po->tx_ring.pg_vec) { in packet_release()
3126 kfree(po->rollover); in packet_release()
3132 * Now the socket is dead. No more input will appear. in packet_release()
3135 sock->sk = NULL; in packet_release()
3139 skb_queue_purge(&sk->sk_receive_queue); in packet_release()
3163 spin_lock(&po->bind_lock); in packet_do_bind()
3165 proto = po->num; in packet_do_bind()
3169 if (po->fanout) { in packet_do_bind()
3170 ret = -EINVAL; in packet_do_bind()
3177 ret = -ENODEV; in packet_do_bind()
3183 ret = -ENODEV; in packet_do_bind()
3191 proto_curr = po->prot_hook.type; in packet_do_bind()
3192 dev_curr = po->prot_hook.dev; in packet_do_bind()
3197 if (po->running) { in packet_do_bind()
3202 WRITE_ONCE(po->num, 0); in packet_do_bind()
3205 dev_curr = po->prot_hook.dev; in packet_do_bind()
3208 dev->ifindex); in packet_do_bind()
3211 BUG_ON(po->running); in packet_do_bind()
3212 WRITE_ONCE(po->num, proto); in packet_do_bind()
3213 po->prot_hook.type = proto; in packet_do_bind()
3217 po->prot_hook.dev = NULL; in packet_do_bind()
3218 WRITE_ONCE(po->ifindex, -1); in packet_do_bind()
3221 po->prot_hook.dev = dev; in packet_do_bind()
3222 WRITE_ONCE(po->ifindex, dev ? dev->ifindex : 0); in packet_do_bind()
3232 if (!unlisted && (!dev || (dev->flags & IFF_UP))) { in packet_do_bind()
3235 sk->sk_err = ENETDOWN; in packet_do_bind()
3237 sk->sk_error_report(sk); in packet_do_bind()
3242 spin_unlock(&po->bind_lock); in packet_do_bind()
3254 struct sock *sk = sock->sk; in packet_bind_spkt()
3255 char name[sizeof(uaddr->sa_data_min) + 1]; in packet_bind_spkt()
3262 return -EINVAL; in packet_bind_spkt()
3263 /* uaddr->sa_data comes from the userspace, it's not guaranteed to be in packet_bind_spkt()
3264 * zero-terminated. in packet_bind_spkt()
3266 memcpy(name, uaddr->sa_data, sizeof(uaddr->sa_data_min)); in packet_bind_spkt()
3267 name[sizeof(uaddr->sa_data_min)] = 0; in packet_bind_spkt()
3275 struct sock *sk = sock->sk; in packet_bind()
3282 return -EINVAL; in packet_bind()
3283 if (sll->sll_family != AF_PACKET) in packet_bind()
3284 return -EINVAL; in packet_bind()
3286 return packet_do_bind(sk, NULL, sll->sll_ifindex, sll->sll_protocol); in packet_bind()
3307 if (!ns_capable(net->user_ns, CAP_NET_RAW)) in packet_create()
3308 return -EPERM; in packet_create()
3309 if (sock->type != SOCK_DGRAM && sock->type != SOCK_RAW && in packet_create()
3310 sock->type != SOCK_PACKET) in packet_create()
3311 return -ESOCKTNOSUPPORT; in packet_create()
3313 sock->state = SS_UNCONNECTED; in packet_create()
3315 err = -ENOBUFS; in packet_create()
3320 sock->ops = &packet_ops; in packet_create()
3321 if (sock->type == SOCK_PACKET) in packet_create()
3322 sock->ops = &packet_ops_spkt; in packet_create()
3331 init_completion(&po->skb_completion); in packet_create()
3332 sk->sk_family = PF_PACKET; in packet_create()
3333 po->num = proto; in packet_create()
3334 po->xmit = dev_queue_xmit; in packet_create()
3338 sk->sk_destruct = packet_sock_destruct; in packet_create()
3345 spin_lock_init(&po->bind_lock); in packet_create()
3346 mutex_init(&po->pg_vec_lock); in packet_create()
3347 po->rollover = NULL; in packet_create()
3348 po->prot_hook.func = packet_rcv; in packet_create()
3350 if (sock->type == SOCK_PACKET) in packet_create()
3351 po->prot_hook.func = packet_rcv_spkt; in packet_create()
3353 po->prot_hook.af_packet_priv = sk; in packet_create()
3354 po->prot_hook.af_packet_net = sock_net(sk); in packet_create()
3357 po->prot_hook.type = proto; in packet_create()
3361 mutex_lock(&net->packet.sklist_lock); in packet_create()
3362 sk_add_node_tail_rcu(sk, &net->packet.sklist); in packet_create()
3363 mutex_unlock(&net->packet.sklist_lock); in packet_create()
3384 struct sock *sk = sock->sk; in packet_recvmsg()
3390 err = -EINVAL; in packet_recvmsg()
3396 if (pkt_sk(sk)->ifindex < 0) in packet_recvmsg()
3397 return -ENODEV; in packet_recvmsg()
3408 * of horrible races and re-entrancy so we can forget about it in packet_recvmsg()
3428 if (pkt_sk(sk)->has_vnet_hdr) { in packet_recvmsg()
3439 copied = skb->len; in packet_recvmsg()
3442 msg->msg_flags |= MSG_TRUNC; in packet_recvmsg()
3449 if (sock->type != SOCK_PACKET) { in packet_recvmsg()
3450 struct sockaddr_ll *sll = &PACKET_SKB_CB(skb)->sa.ll; in packet_recvmsg()
3453 origlen = PACKET_SKB_CB(skb)->sa.origlen; in packet_recvmsg()
3454 sll->sll_family = AF_PACKET; in packet_recvmsg()
3455 sll->sll_protocol = skb->protocol; in packet_recvmsg()
3460 if (msg->msg_name) { in packet_recvmsg()
3461 const size_t max_len = min(sizeof(skb->cb), in packet_recvmsg()
3468 if (sock->type == SOCK_PACKET) { in packet_recvmsg()
3470 msg->msg_namelen = sizeof(struct sockaddr_pkt); in packet_recvmsg()
3471 copy_len = msg->msg_namelen; in packet_recvmsg()
3473 struct sockaddr_ll *sll = &PACKET_SKB_CB(skb)->sa.ll; in packet_recvmsg()
3475 msg->msg_namelen = sll->sll_halen + in packet_recvmsg()
3477 copy_len = msg->msg_namelen; in packet_recvmsg()
3478 if (msg->msg_namelen < sizeof(struct sockaddr_ll)) { in packet_recvmsg()
3479 memset(msg->msg_name + in packet_recvmsg()
3481 0, sizeof(sll->sll_addr)); in packet_recvmsg()
3482 msg->msg_namelen = sizeof(struct sockaddr_ll); in packet_recvmsg()
3487 msg->msg_namelen = copy_len; in packet_recvmsg()
3489 memcpy(msg->msg_name, &PACKET_SKB_CB(skb)->sa, copy_len); in packet_recvmsg()
3496 if (skb->ip_summed == CHECKSUM_PARTIAL) in packet_recvmsg()
3498 else if (skb->pkt_type != PACKET_OUTGOING && in packet_recvmsg()
3503 aux.tp_snaplen = skb->len; in packet_recvmsg()
3508 aux.tp_vlan_tpid = ntohs(skb->vlan_proto); in packet_recvmsg()
3519 * hides all the races and re-entrancy issues from us. in packet_recvmsg()
3521 err = vnet_hdr_len + ((flags&MSG_TRUNC) ? skb->len : copied); in packet_recvmsg()
3533 struct sock *sk = sock->sk; in packet_getname_spkt()
3536 return -EOPNOTSUPP; in packet_getname_spkt()
3538 uaddr->sa_family = AF_PACKET; in packet_getname_spkt()
3539 memset(uaddr->sa_data, 0, sizeof(uaddr->sa_data_min)); in packet_getname_spkt()
3541 dev = dev_get_by_index_rcu(sock_net(sk), READ_ONCE(pkt_sk(sk)->ifindex)); in packet_getname_spkt()
3543 strscpy(uaddr->sa_data, dev->name, sizeof(uaddr->sa_data_min)); in packet_getname_spkt()
3553 struct sock *sk = sock->sk; in packet_getname()
3559 return -EOPNOTSUPP; in packet_getname()
3561 ifindex = READ_ONCE(po->ifindex); in packet_getname()
3562 sll->sll_family = AF_PACKET; in packet_getname()
3563 sll->sll_ifindex = ifindex; in packet_getname()
3564 sll->sll_protocol = READ_ONCE(po->num); in packet_getname()
3565 sll->sll_pkttype = 0; in packet_getname()
3569 sll->sll_hatype = dev->type; in packet_getname()
3570 sll->sll_halen = dev->addr_len; in packet_getname()
3571 memcpy(sll->sll_addr, dev->dev_addr, dev->addr_len); in packet_getname()
3573 sll->sll_hatype = 0; /* Bad: we have no ARPHRD_UNSPEC */ in packet_getname()
3574 sll->sll_halen = 0; in packet_getname()
3578 return offsetof(struct sockaddr_ll, sll_addr) + sll->sll_halen; in packet_getname()
3584 switch (i->type) { in packet_dev_mc()
3586 if (i->alen != dev->addr_len) in packet_dev_mc()
3587 return -EINVAL; in packet_dev_mc()
3589 return dev_mc_add(dev, i->addr); in packet_dev_mc()
3591 return dev_mc_del(dev, i->addr); in packet_dev_mc()
3598 if (i->alen != dev->addr_len) in packet_dev_mc()
3599 return -EINVAL; in packet_dev_mc()
3601 return dev_uc_add(dev, i->addr); in packet_dev_mc()
3603 return dev_uc_del(dev, i->addr); in packet_dev_mc()
3617 if (ml->ifindex == dev->ifindex) { in packet_dev_mclist_delete()
3618 packet_dev_mc(dev, ml, -1); in packet_dev_mclist_delete()
3619 *mlp = ml->next; in packet_dev_mclist_delete()
3622 mlp = &ml->next; in packet_dev_mclist_delete()
3635 err = -ENODEV; in packet_mc_add()
3636 dev = __dev_get_by_index(sock_net(sk), mreq->mr_ifindex); in packet_mc_add()
3640 err = -EINVAL; in packet_mc_add()
3641 if (mreq->mr_alen > dev->addr_len) in packet_mc_add()
3644 err = -ENOBUFS; in packet_mc_add()
3650 for (ml = po->mclist; ml; ml = ml->next) { in packet_mc_add()
3651 if (ml->ifindex == mreq->mr_ifindex && in packet_mc_add()
3652 ml->type == mreq->mr_type && in packet_mc_add()
3653 ml->alen == mreq->mr_alen && in packet_mc_add()
3654 memcmp(ml->addr, mreq->mr_address, ml->alen) == 0) { in packet_mc_add()
3655 ml->count++; in packet_mc_add()
3662 i->type = mreq->mr_type; in packet_mc_add()
3663 i->ifindex = mreq->mr_ifindex; in packet_mc_add()
3664 i->alen = mreq->mr_alen; in packet_mc_add()
3665 memcpy(i->addr, mreq->mr_address, i->alen); in packet_mc_add()
3666 memset(i->addr + i->alen, 0, sizeof(i->addr) - i->alen); in packet_mc_add()
3667 i->count = 1; in packet_mc_add()
3668 i->next = po->mclist; in packet_mc_add()
3669 po->mclist = i; in packet_mc_add()
3672 po->mclist = i->next; in packet_mc_add()
3687 for (mlp = &pkt_sk(sk)->mclist; (ml = *mlp) != NULL; mlp = &ml->next) { in packet_mc_drop()
3688 if (ml->ifindex == mreq->mr_ifindex && in packet_mc_drop()
3689 ml->type == mreq->mr_type && in packet_mc_drop()
3690 ml->alen == mreq->mr_alen && in packet_mc_drop()
3691 memcmp(ml->addr, mreq->mr_address, ml->alen) == 0) { in packet_mc_drop()
3692 if (--ml->count == 0) { in packet_mc_drop()
3694 *mlp = ml->next; in packet_mc_drop()
3695 dev = __dev_get_by_index(sock_net(sk), ml->ifindex); in packet_mc_drop()
3697 packet_dev_mc(dev, ml, -1); in packet_mc_drop()
3712 if (!po->mclist) in packet_flush_mclist()
3716 while ((ml = po->mclist) != NULL) { in packet_flush_mclist()
3719 po->mclist = ml->next; in packet_flush_mclist()
3720 dev = __dev_get_by_index(sock_net(sk), ml->ifindex); in packet_flush_mclist()
3722 packet_dev_mc(dev, ml, -1); in packet_flush_mclist()
3732 struct sock *sk = sock->sk; in packet_setsockopt()
3737 return -ENOPROTOOPT; in packet_setsockopt()
3747 return -EINVAL; in packet_setsockopt()
3751 return -EFAULT; in packet_setsockopt()
3753 return -EINVAL; in packet_setsockopt()
3768 switch (po->tp_version) { in packet_setsockopt()
3779 ret = -EINVAL; in packet_setsockopt()
3782 ret = -EFAULT; in packet_setsockopt()
3795 return -EINVAL; in packet_setsockopt()
3797 return -EFAULT; in packet_setsockopt()
3799 pkt_sk(sk)->copy_thresh = val; in packet_setsockopt()
3807 return -EINVAL; in packet_setsockopt()
3809 return -EFAULT; in packet_setsockopt()
3816 return -EINVAL; in packet_setsockopt()
3819 if (po->rx_ring.pg_vec || po->tx_ring.pg_vec) { in packet_setsockopt()
3820 ret = -EBUSY; in packet_setsockopt()
3822 po->tp_version = val; in packet_setsockopt()
3833 return -EINVAL; in packet_setsockopt()
3835 return -EFAULT; in packet_setsockopt()
3837 return -EINVAL; in packet_setsockopt()
3839 if (po->rx_ring.pg_vec || po->tx_ring.pg_vec) { in packet_setsockopt()
3840 ret = -EBUSY; in packet_setsockopt()
3842 po->tp_reserve = val; in packet_setsockopt()
3853 return -EINVAL; in packet_setsockopt()
3855 return -EFAULT; in packet_setsockopt()
3858 if (po->rx_ring.pg_vec || po->tx_ring.pg_vec) { in packet_setsockopt()
3859 ret = -EBUSY; in packet_setsockopt()
3861 po->tp_loss = !!val; in packet_setsockopt()
3872 return -EINVAL; in packet_setsockopt()
3874 return -EFAULT; in packet_setsockopt()
3884 return -EINVAL; in packet_setsockopt()
3886 return -EFAULT; in packet_setsockopt()
3895 if (sock->type != SOCK_RAW) in packet_setsockopt()
3896 return -EINVAL; in packet_setsockopt()
3898 return -EINVAL; in packet_setsockopt()
3900 return -EFAULT; in packet_setsockopt()
3903 if (po->rx_ring.pg_vec || po->tx_ring.pg_vec) { in packet_setsockopt()
3904 ret = -EBUSY; in packet_setsockopt()
3906 po->has_vnet_hdr = !!val; in packet_setsockopt()
3917 return -EINVAL; in packet_setsockopt()
3919 return -EFAULT; in packet_setsockopt()
3921 po->tp_tstamp = val; in packet_setsockopt()
3929 return -EINVAL; in packet_setsockopt()
3931 return -EFAULT; in packet_setsockopt()
3938 if (!READ_ONCE(po->fanout)) in packet_setsockopt()
3939 return -EINVAL; in packet_setsockopt()
3948 return -EINVAL; in packet_setsockopt()
3950 return -EFAULT; in packet_setsockopt()
3952 return -EINVAL; in packet_setsockopt()
3954 WRITE_ONCE(po->prot_hook.ignore_outgoing, !!val); in packet_setsockopt()
3962 return -EINVAL; in packet_setsockopt()
3964 return -EFAULT; in packet_setsockopt()
3967 if (po->rx_ring.pg_vec || po->tx_ring.pg_vec) { in packet_setsockopt()
3968 ret = -EBUSY; in packet_setsockopt()
3970 po->tp_tx_has_off = !!val; in packet_setsockopt()
3981 return -EINVAL; in packet_setsockopt()
3983 return -EFAULT; in packet_setsockopt()
3985 /* Paired with all lockless reads of po->xmit */ in packet_setsockopt()
3986 WRITE_ONCE(po->xmit, val ? packet_direct_xmit : dev_queue_xmit); in packet_setsockopt()
3990 return -ENOPROTOOPT; in packet_setsockopt()
3999 struct sock *sk = sock->sk; in packet_getsockopt()
4007 return -ENOPROTOOPT; in packet_getsockopt()
4010 return -EFAULT; in packet_getsockopt()
4013 return -EINVAL; in packet_getsockopt()
4017 spin_lock_bh(&sk->sk_receive_queue.lock); in packet_getsockopt()
4018 memcpy(&st, &po->stats, sizeof(st)); in packet_getsockopt()
4019 memset(&po->stats, 0, sizeof(po->stats)); in packet_getsockopt()
4020 spin_unlock_bh(&sk->sk_receive_queue.lock); in packet_getsockopt()
4021 drops = atomic_xchg(&po->tp_drops, 0); in packet_getsockopt()
4023 if (po->tp_version == TPACKET_V3) { in packet_getsockopt()
4043 val = po->has_vnet_hdr; in packet_getsockopt()
4046 val = po->tp_version; in packet_getsockopt()
4052 return -EINVAL; in packet_getsockopt()
4054 return -EFAULT; in packet_getsockopt()
4066 return -EINVAL; in packet_getsockopt()
4070 val = po->tp_reserve; in packet_getsockopt()
4073 val = po->tp_loss; in packet_getsockopt()
4076 val = po->tp_tstamp; in packet_getsockopt()
4079 val = (po->fanout ? in packet_getsockopt()
4080 ((u32)po->fanout->id | in packet_getsockopt()
4081 ((u32)po->fanout->type << 16) | in packet_getsockopt()
4082 ((u32)po->fanout->flags << 24)) : in packet_getsockopt()
4086 val = READ_ONCE(po->prot_hook.ignore_outgoing); in packet_getsockopt()
4089 if (!po->rollover) in packet_getsockopt()
4090 return -EINVAL; in packet_getsockopt()
4091 rstats.tp_all = atomic_long_read(&po->rollover->num); in packet_getsockopt()
4092 rstats.tp_huge = atomic_long_read(&po->rollover->num_huge); in packet_getsockopt()
4093 rstats.tp_failed = atomic_long_read(&po->rollover->num_failed); in packet_getsockopt()
4098 val = po->tp_tx_has_off; in packet_getsockopt()
4104 return -ENOPROTOOPT; in packet_getsockopt()
4110 return -EFAULT; in packet_getsockopt()
4112 return -EFAULT; in packet_getsockopt()
4124 sk_for_each_rcu(sk, &net->packet.sklist) { in packet_notifier()
4129 if (po->mclist) in packet_notifier()
4130 packet_dev_mclist_delete(dev, &po->mclist); in packet_notifier()
4134 if (dev->ifindex == po->ifindex) { in packet_notifier()
4135 spin_lock(&po->bind_lock); in packet_notifier()
4136 if (po->running) { in packet_notifier()
4138 sk->sk_err = ENETDOWN; in packet_notifier()
4140 sk->sk_error_report(sk); in packet_notifier()
4144 WRITE_ONCE(po->ifindex, -1); in packet_notifier()
4145 if (po->prot_hook.dev) in packet_notifier()
4146 dev_put(po->prot_hook.dev); in packet_notifier()
4147 po->prot_hook.dev = NULL; in packet_notifier()
4149 spin_unlock(&po->bind_lock); in packet_notifier()
4153 if (dev->ifindex == po->ifindex) { in packet_notifier()
4154 spin_lock(&po->bind_lock); in packet_notifier()
4155 if (po->num) in packet_notifier()
4157 spin_unlock(&po->bind_lock); in packet_notifier()
4170 struct sock *sk = sock->sk; in packet_ioctl()
4184 spin_lock_bh(&sk->sk_receive_queue.lock); in packet_ioctl()
4185 skb = skb_peek(&sk->sk_receive_queue); in packet_ioctl()
4187 amount = skb->len; in packet_ioctl()
4188 spin_unlock_bh(&sk->sk_receive_queue.lock); in packet_ioctl()
4210 return -ENOIOCTLCMD; in packet_ioctl()
4218 struct sock *sk = sock->sk; in packet_poll()
4222 spin_lock_bh(&sk->sk_receive_queue.lock); in packet_poll()
4223 if (po->rx_ring.pg_vec) { in packet_poll()
4224 if (!packet_previous_rx_frame(po, &po->rx_ring, in packet_poll()
4229 spin_unlock_bh(&sk->sk_receive_queue.lock); in packet_poll()
4230 spin_lock_bh(&sk->sk_write_queue.lock); in packet_poll()
4231 if (po->tx_ring.pg_vec) { in packet_poll()
4232 if (packet_current_frame(po, &po->tx_ring, TP_STATUS_AVAILABLE)) in packet_poll()
4235 spin_unlock_bh(&sk->sk_write_queue.lock); in packet_poll()
4246 struct file *file = vma->vm_file; in packet_mm_open()
4247 struct socket *sock = file->private_data; in packet_mm_open()
4248 struct sock *sk = sock->sk; in packet_mm_open()
4251 atomic_long_inc(&pkt_sk(sk)->mapped); in packet_mm_open()
4256 struct file *file = vma->vm_file; in packet_mm_close()
4257 struct socket *sock = file->private_data; in packet_mm_close()
4258 struct sock *sk = sock->sk; in packet_mm_close()
4261 atomic_long_dec(&pkt_sk(sk)->mapped); in packet_mm_close()
4314 unsigned int block_nr = req->tp_block_nr; in alloc_pg_vec()
4349 struct tpacket_req *req = &req_u->req; in packet_set_ring()
4351 rb = tx_ring ? &po->tx_ring : &po->rx_ring; in packet_set_ring()
4352 rb_queue = tx_ring ? &sk->sk_write_queue : &sk->sk_receive_queue; in packet_set_ring()
4354 err = -EBUSY; in packet_set_ring()
4356 if (atomic_long_read(&po->mapped)) in packet_set_ring()
4362 if (req->tp_block_nr) { in packet_set_ring()
4366 err = -EBUSY; in packet_set_ring()
4367 if (unlikely(rb->pg_vec)) in packet_set_ring()
4370 switch (po->tp_version) { in packet_set_ring()
4372 po->tp_hdrlen = TPACKET_HDRLEN; in packet_set_ring()
4375 po->tp_hdrlen = TPACKET2_HDRLEN; in packet_set_ring()
4378 po->tp_hdrlen = TPACKET3_HDRLEN; in packet_set_ring()
4382 err = -EINVAL; in packet_set_ring()
4383 if (unlikely((int)req->tp_block_size <= 0)) in packet_set_ring()
4385 if (unlikely(!PAGE_ALIGNED(req->tp_block_size))) in packet_set_ring()
4387 min_frame_size = po->tp_hdrlen + po->tp_reserve; in packet_set_ring()
4388 if (po->tp_version >= TPACKET_V3 && in packet_set_ring()
4389 req->tp_block_size < in packet_set_ring()
4390 BLK_PLUS_PRIV((u64)req_u->req3.tp_sizeof_priv) + min_frame_size) in packet_set_ring()
4392 if (unlikely(req->tp_frame_size < min_frame_size)) in packet_set_ring()
4394 if (unlikely(req->tp_frame_size & (TPACKET_ALIGNMENT - 1))) in packet_set_ring()
4397 rb->frames_per_block = req->tp_block_size / req->tp_frame_size; in packet_set_ring()
4398 if (unlikely(rb->frames_per_block == 0)) in packet_set_ring()
4400 if (unlikely(rb->frames_per_block > UINT_MAX / req->tp_block_nr)) in packet_set_ring()
4402 if (unlikely((rb->frames_per_block * req->tp_block_nr) != in packet_set_ring()
4403 req->tp_frame_nr)) in packet_set_ring()
4406 err = -ENOMEM; in packet_set_ring()
4407 order = get_order(req->tp_block_size); in packet_set_ring()
4411 switch (po->tp_version) { in packet_set_ring()
4417 struct tpacket_req3 *req3 = &req_u->req3; in packet_set_ring()
4419 if (req3->tp_retire_blk_tov || in packet_set_ring()
4420 req3->tp_sizeof_priv || in packet_set_ring()
4421 req3->tp_feature_req_word) { in packet_set_ring()
4422 err = -EINVAL; in packet_set_ring()
4429 rx_owner_map = bitmap_alloc(req->tp_frame_nr, in packet_set_ring()
4439 err = -EINVAL; in packet_set_ring()
4440 if (unlikely(req->tp_frame_nr)) in packet_set_ring()
4446 spin_lock(&po->bind_lock); in packet_set_ring()
4447 was_running = po->running; in packet_set_ring()
4448 num = po->num; in packet_set_ring()
4450 WRITE_ONCE(po->num, 0); in packet_set_ring()
4453 spin_unlock(&po->bind_lock); in packet_set_ring()
4457 err = -EBUSY; in packet_set_ring()
4458 mutex_lock(&po->pg_vec_lock); in packet_set_ring()
4459 if (closing || atomic_long_read(&po->mapped) == 0) { in packet_set_ring()
4461 spin_lock_bh(&rb_queue->lock); in packet_set_ring()
4462 swap(rb->pg_vec, pg_vec); in packet_set_ring()
4463 if (po->tp_version <= TPACKET_V2) in packet_set_ring()
4464 swap(rb->rx_owner_map, rx_owner_map); in packet_set_ring()
4465 rb->frame_max = (req->tp_frame_nr - 1); in packet_set_ring()
4466 rb->head = 0; in packet_set_ring()
4467 rb->frame_size = req->tp_frame_size; in packet_set_ring()
4468 spin_unlock_bh(&rb_queue->lock); in packet_set_ring()
4470 swap(rb->pg_vec_order, order); in packet_set_ring()
4471 swap(rb->pg_vec_len, req->tp_block_nr); in packet_set_ring()
4473 rb->pg_vec_pages = req->tp_block_size/PAGE_SIZE; in packet_set_ring()
4474 po->prot_hook.func = (po->rx_ring.pg_vec) ? in packet_set_ring()
4477 if (atomic_long_read(&po->mapped)) in packet_set_ring()
4479 atomic_long_read(&po->mapped)); in packet_set_ring()
4481 mutex_unlock(&po->pg_vec_lock); in packet_set_ring()
4483 spin_lock(&po->bind_lock); in packet_set_ring()
4485 WRITE_ONCE(po->num, num); in packet_set_ring()
4488 spin_unlock(&po->bind_lock); in packet_set_ring()
4489 if (pg_vec && (po->tp_version > TPACKET_V2)) { in packet_set_ring()
4490 /* Because we don't support block-based V3 on tx-ring */ in packet_set_ring()
4498 free_pg_vec(pg_vec, order, req->tp_block_nr); in packet_set_ring()
4507 struct sock *sk = sock->sk; in packet_mmap()
4512 int err = -EINVAL; in packet_mmap()
4515 if (vma->vm_pgoff) in packet_mmap()
4516 return -EINVAL; in packet_mmap()
4518 mutex_lock(&po->pg_vec_lock); in packet_mmap()
4521 for (rb = &po->rx_ring; rb <= &po->tx_ring; rb++) { in packet_mmap()
4522 if (rb->pg_vec) { in packet_mmap()
4523 expected_size += rb->pg_vec_len in packet_mmap()
4524 * rb->pg_vec_pages in packet_mmap()
4532 size = vma->vm_end - vma->vm_start; in packet_mmap()
4536 start = vma->vm_start; in packet_mmap()
4537 for (rb = &po->rx_ring; rb <= &po->tx_ring; rb++) { in packet_mmap()
4538 if (rb->pg_vec == NULL) in packet_mmap()
4541 for (i = 0; i < rb->pg_vec_len; i++) { in packet_mmap()
4543 void *kaddr = rb->pg_vec[i].buffer; in packet_mmap()
4546 for (pg_num = 0; pg_num < rb->pg_vec_pages; pg_num++) { in packet_mmap()
4557 atomic_long_inc(&po->mapped); in packet_mmap()
4558 vma->vm_ops = &packet_mmap_ops; in packet_mmap()
4562 mutex_unlock(&po->pg_vec_lock); in packet_mmap()
4626 return seq_hlist_start_head_rcu(&net->packet.sklist, *pos); in packet_seq_start()
4632 return seq_hlist_next_rcu(v, &net->packet.sklist, pos); in packet_seq_next()
4650 "%pK %-6d %-4d %04x %-5d %1d %-6u %-6u %-6lu\n", in packet_seq_show()
4652 refcount_read(&s->sk_refcnt), in packet_seq_show()
4653 s->sk_type, in packet_seq_show()
4654 ntohs(READ_ONCE(po->num)), in packet_seq_show()
4655 READ_ONCE(po->ifindex), in packet_seq_show()
4656 po->running, in packet_seq_show()
4657 atomic_read(&s->sk_rmem_alloc), in packet_seq_show()
4675 mutex_init(&net->packet.sklist_lock); in packet_net_init()
4676 INIT_HLIST_HEAD(&net->packet.sklist); in packet_net_init()
4679 if (!proc_create_net("packet", 0, net->proc_net, &packet_seq_ops, in packet_net_init()
4681 return -ENOMEM; in packet_net_init()
4689 remove_proc_entry("packet", net->proc_net); in packet_net_exit()
4690 WARN_ON_ONCE(!hlist_empty(&net->packet.sklist)); in packet_net_exit()