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.
101 - If the device has no dev->header_ops->create, there is no LL header
108 needed_headroom to be (the real WiFi header length - the fake Ethernet
110 - packet socket receives packets with pulled ll header,
114 -----------
117 mac_header -> ll header
118 data -> data
121 mac_header -> ll header
122 data -> ll header
125 mac_header -> data
128 data -> data
131 mac_header -> data. ll header is invisible to us.
132 data -> data
140 ------------
143 mac_header -> ll header
144 data -> ll header
147 mac_header -> data
148 data -> data
183 #define BLOCK_STATUS(x) ((x)->hdr.bh1.block_status)
184 #define BLOCK_NUM_PKTS(x) ((x)->hdr.bh1.num_pkts)
185 #define BLOCK_O2FP(x) ((x)->hdr.bh1.offset_to_first_pkt)
186 #define BLOCK_LEN(x) ((x)->hdr.bh1.blk_len)
187 #define BLOCK_SNUM(x) ((x)->hdr.bh1.seq_num)
188 #define BLOCK_O2PRIV(x) ((x)->offset_to_priv)
232 #define PACKET_SKB_CB(__skb) ((struct packet_skb_cb *)((__skb)->cb))
234 #define GET_PBDQC_FROM_RB(x) ((struct tpacket_kbdq_core *)(&(x)->prb_bdqc))
236 ((struct tpacket_block_desc *)((x)->pkbdq[(bid)].buffer))
238 ((struct tpacket_block_desc *)((x)->pkbdq[(x)->kactive_blk_num].buffer))
240 (((x)->kactive_blk_num < ((x)->knum_blocks-1)) ? \
241 ((x)->kactive_blk_num+1) : 0)
254 next = skb->next; in nf_hook_direct_egress()
257 if (!nf_hook_egress(skb, &rc, skb->dev)) in nf_hook_direct_egress()
263 tail->next = skb; in nf_hook_direct_egress()
293 dev = rcu_dereference(po->cached_dev); in packet_cached_dev_get()
303 rcu_assign_pointer(po->cached_dev, dev); in packet_cached_dev_assign()
308 RCU_INIT_POINTER(po->cached_dev, NULL); in packet_cached_dev_reset()
313 struct net_device *dev = skb->dev; in packet_pick_tx_queue()
314 const struct net_device_ops *ops = dev->netdev_ops; in packet_pick_tx_queue()
319 skb->sender_cpu = cpu + 1; in packet_pick_tx_queue()
321 skb_record_rx_queue(skb, cpu % dev->real_num_tx_queues); in packet_pick_tx_queue()
322 if (ops->ndo_select_queue) { in packet_pick_tx_queue()
323 queue_index = ops->ndo_select_queue(dev, skb, NULL); in packet_pick_tx_queue()
341 if (po->fanout) in __register_prot_hook()
344 dev_add_pack(&po->prot_hook); in __register_prot_hook()
353 lockdep_assert_held_once(&pkt_sk(sk)->bind_lock); in register_prot_hook()
358 * the po->bind_lock and do a synchronize_net to make sure no
360 * of po->prot_hook. If the sync parameter is false, it is the
367 lockdep_assert_held_once(&po->bind_lock); in __unregister_prot_hook()
371 if (po->fanout) in __unregister_prot_hook()
374 __dev_remove_pack(&po->prot_hook); in __unregister_prot_hook()
379 spin_unlock(&po->bind_lock); in __unregister_prot_hook()
381 spin_lock(&po->bind_lock); in __unregister_prot_hook()
407 switch (po->tp_version) { in __packet_set_status()
409 WRITE_ONCE(h.h1->tp_status, status); in __packet_set_status()
410 flush_dcache_page(pgv_to_page(&h.h1->tp_status)); in __packet_set_status()
413 WRITE_ONCE(h.h2->tp_status, status); in __packet_set_status()
414 flush_dcache_page(pgv_to_page(&h.h2->tp_status)); in __packet_set_status()
417 WRITE_ONCE(h.h3->tp_status, status); in __packet_set_status()
418 flush_dcache_page(pgv_to_page(&h.h3->tp_status)); in __packet_set_status()
437 switch (po->tp_version) { in __packet_get_status()
439 flush_dcache_page(pgv_to_page(&h.h1->tp_status)); in __packet_get_status()
440 return READ_ONCE(h.h1->tp_status); in __packet_get_status()
442 flush_dcache_page(pgv_to_page(&h.h2->tp_status)); in __packet_get_status()
443 return READ_ONCE(h.h2->tp_status); in __packet_get_status()
445 flush_dcache_page(pgv_to_page(&h.h3->tp_status)); in __packet_get_status()
446 return READ_ONCE(h.h3->tp_status); in __packet_get_status()
461 ktime_to_timespec64_cond(shhwtstamps->hwtstamp, ts)) in tpacket_get_timestamp()
478 if (!(ts_status = tpacket_get_timestamp(skb, &ts, READ_ONCE(po->tp_tstamp)))) in __packet_set_timestamp()
484 * all store the seconds in a 32-bit unsigned integer. in __packet_set_timestamp()
485 * If we create a version 4, that should have a 64-bit timestamp, in __packet_set_timestamp()
486 * either 64-bit seconds + 32-bit nanoseconds, or just 64-bit in __packet_set_timestamp()
489 switch (po->tp_version) { in __packet_set_timestamp()
491 h.h1->tp_sec = ts.tv_sec; in __packet_set_timestamp()
492 h.h1->tp_usec = ts.tv_nsec / NSEC_PER_USEC; in __packet_set_timestamp()
495 h.h2->tp_sec = ts.tv_sec; in __packet_set_timestamp()
496 h.h2->tp_nsec = ts.tv_nsec; in __packet_set_timestamp()
499 h.h3->tp_sec = ts.tv_sec; in __packet_set_timestamp()
500 h.h3->tp_nsec = ts.tv_nsec; in __packet_set_timestamp()
508 flush_dcache_page(pgv_to_page(&h.h1->tp_sec)); in __packet_set_timestamp()
522 pg_vec_pos = position / rb->frames_per_block; in packet_lookup_frame()
523 frame_offset = position % rb->frames_per_block; in packet_lookup_frame()
525 h.raw = rb->pg_vec[pg_vec_pos].buffer + in packet_lookup_frame()
526 (frame_offset * rb->frame_size); in packet_lookup_frame()
538 return packet_lookup_frame(po, rb, rb->head, status); in packet_current_frame()
543 del_timer_sync(&pkc->retire_blk_timer); in prb_del_retire_blk_timer()
551 pkc = GET_PBDQC_FROM_RB(&po->rx_ring); in prb_shutdown_retire_blk_timer()
553 spin_lock_bh(&rb_queue->lock); in prb_shutdown_retire_blk_timer()
554 pkc->delete_blk_timer = 1; in prb_shutdown_retire_blk_timer()
555 spin_unlock_bh(&rb_queue->lock); in prb_shutdown_retire_blk_timer()
564 pkc = GET_PBDQC_FROM_RB(&po->rx_ring); in prb_setup_retire_blk_timer()
565 timer_setup(&pkc->retire_blk_timer, prb_retire_rx_blk_timer_expired, in prb_setup_retire_blk_timer()
567 pkc->retire_blk_timer.expires = jiffies; in prb_setup_retire_blk_timer()
579 dev = __dev_get_by_index(sock_net(&po->sk), po->ifindex); in prb_calc_retire_blk_tmo()
610 p1->feature_req_word = req_u->req3.tp_feature_req_word; in prb_init_ft_ops()
623 p1->knxt_seq_num = 1; in init_prb_bdqc()
624 p1->pkbdq = pg_vec; in init_prb_bdqc()
626 p1->pkblk_start = pg_vec[0].buffer; in init_prb_bdqc()
627 p1->kblk_size = req_u->req3.tp_block_size; in init_prb_bdqc()
628 p1->knum_blocks = req_u->req3.tp_block_nr; in init_prb_bdqc()
629 p1->hdrlen = po->tp_hdrlen; in init_prb_bdqc()
630 p1->version = po->tp_version; in init_prb_bdqc()
631 p1->last_kactive_blk_num = 0; in init_prb_bdqc()
632 po->stats.stats3.tp_freeze_q_cnt = 0; in init_prb_bdqc()
633 if (req_u->req3.tp_retire_blk_tov) in init_prb_bdqc()
634 p1->retire_blk_tov = req_u->req3.tp_retire_blk_tov; in init_prb_bdqc()
636 p1->retire_blk_tov = prb_calc_retire_blk_tmo(po, in init_prb_bdqc()
637 req_u->req3.tp_block_size); in init_prb_bdqc()
638 p1->tov_in_jiffies = msecs_to_jiffies(p1->retire_blk_tov); in init_prb_bdqc()
639 p1->blk_sizeof_priv = req_u->req3.tp_sizeof_priv; in init_prb_bdqc()
640 rwlock_init(&p1->blk_fill_in_prog_lock); in init_prb_bdqc()
642 p1->max_frame_len = p1->kblk_size - BLK_PLUS_PRIV(p1->blk_sizeof_priv); in init_prb_bdqc()
653 mod_timer(&pkc->retire_blk_timer, in _prb_refresh_rx_retire_blk_timer()
654 jiffies + pkc->tov_in_jiffies); in _prb_refresh_rx_retire_blk_timer()
655 pkc->last_kactive_blk_num = pkc->kactive_blk_num; in _prb_refresh_rx_retire_blk_timer()
662 * on packet-by-packet basis.
664 * With a 1MB block-size, on a 1Gbps line, it will take
677 * a) line-speed and b) block-size.
685 struct tpacket_kbdq_core *pkc = GET_PBDQC_FROM_RB(&po->rx_ring); in prb_retire_rx_blk_timer_expired()
689 spin_lock(&po->sk.sk_receive_queue.lock); in prb_retire_rx_blk_timer_expired()
694 if (unlikely(pkc->delete_blk_timer)) in prb_retire_rx_blk_timer_expired()
708 write_lock(&pkc->blk_fill_in_prog_lock); in prb_retire_rx_blk_timer_expired()
709 write_unlock(&pkc->blk_fill_in_prog_lock); in prb_retire_rx_blk_timer_expired()
712 if (pkc->last_kactive_blk_num == pkc->kactive_blk_num) { in prb_retire_rx_blk_timer_expired()
724 /* Case 1. Queue was frozen because user-space was in prb_retire_rx_blk_timer_expired()
729 * Ok, user-space is still behind. in prb_retire_rx_blk_timer_expired()
734 /* Case 2. queue was frozen,user-space caught up, in prb_retire_rx_blk_timer_expired()
739 * Thawing/timer-refresh is a side effect. in prb_retire_rx_blk_timer_expired()
751 spin_unlock(&po->sk.sk_receive_queue.lock); in prb_retire_rx_blk_timer_expired()
767 end = (u8 *)PAGE_ALIGN((unsigned long)pkc1->pkblk_end); in prb_flush_block()
804 struct tpacket_hdr_v1 *h1 = &pbd1->hdr.bh1; in prb_close_block()
805 struct sock *sk = &po->sk; in prb_close_block()
807 if (atomic_read(&po->tp_drops)) in prb_close_block()
810 last_pkt = (struct tpacket3_hdr *)pkc1->prev; in prb_close_block()
811 last_pkt->tp_next_offset = 0; in prb_close_block()
815 h1->ts_last_pkt.ts_sec = last_pkt->tp_sec; in prb_close_block()
816 h1->ts_last_pkt.ts_nsec = last_pkt->tp_nsec; in prb_close_block()
818 /* Ok, we tmo'd - so get the current time. in prb_close_block()
825 h1->ts_last_pkt.ts_sec = ts.tv_sec; in prb_close_block()
826 h1->ts_last_pkt.ts_nsec = ts.tv_nsec; in prb_close_block()
834 sk->sk_data_ready(sk); in prb_close_block()
836 pkc1->kactive_blk_num = GET_NEXT_PRB_BLK_NUM(pkc1); in prb_close_block()
841 pkc->reset_pending_on_curr_blk = 0; in prb_thaw_queue()
855 struct tpacket_hdr_v1 *h1 = &pbd1->hdr.bh1; in prb_open_block()
863 BLOCK_SNUM(pbd1) = pkc1->knxt_seq_num++; in prb_open_block()
865 BLOCK_LEN(pbd1) = BLK_PLUS_PRIV(pkc1->blk_sizeof_priv); in prb_open_block()
869 h1->ts_first_pkt.ts_sec = ts.tv_sec; in prb_open_block()
870 h1->ts_first_pkt.ts_nsec = ts.tv_nsec; in prb_open_block()
872 pkc1->pkblk_start = (char *)pbd1; in prb_open_block()
873 pkc1->nxt_offset = pkc1->pkblk_start + BLK_PLUS_PRIV(pkc1->blk_sizeof_priv); in prb_open_block()
875 BLOCK_O2FP(pbd1) = (__u32)BLK_PLUS_PRIV(pkc1->blk_sizeof_priv); in prb_open_block()
878 pbd1->version = pkc1->version; in prb_open_block()
879 pkc1->prev = pkc1->nxt_offset; in prb_open_block()
880 pkc1->pkblk_end = pkc1->pkblk_start + pkc1->kblk_size; in prb_open_block()
893 * 4) user-space is either sleeping or processing block '0'.
894 * 5) tpacket_rcv is currently filling block '7', since there is no space left,
895 * it will close block-7,loop around and try to fill block '0'.
896 * call-flow:
900 * |->(BLOCK_STATUS == USER) evaluates to true
901 * 5.1) Since block-0 is currently in-use, we just freeze the queue.
906 * re-open block-0 in near future.
908 * case and __packet_lookup_frame_in_block will check if block-0
909 * is free and can now be re-used.
914 pkc->reset_pending_on_curr_blk = 1; in prb_freeze_queue()
915 po->stats.stats3.tp_freeze_q_cnt++; in prb_freeze_queue()
948 return (void *)pkc->nxt_offset; in prb_dispatch_next_block()
960 * cpu-0 and tpacket_rcv() got invoked on cpu-1, didn't in prb_retire_current_block()
965 * the timer-handler already handled this case. in prb_retire_current_block()
969 write_lock(&pkc->blk_fill_in_prog_lock); in prb_retire_current_block()
970 write_unlock(&pkc->blk_fill_in_prog_lock); in prb_retire_current_block()
984 return pkc->reset_pending_on_curr_blk; in prb_queue_frozen()
988 __releases(&pkc->blk_fill_in_prog_lock) in prb_clear_blk_fill_status()
992 read_unlock(&pkc->blk_fill_in_prog_lock); in prb_clear_blk_fill_status()
998 ppd->hv1.tp_rxhash = skb_get_hash(pkc->skb); in prb_fill_rxhash()
1004 ppd->hv1.tp_rxhash = 0; in prb_clear_rxhash()
1010 if (skb_vlan_tag_present(pkc->skb)) { in prb_fill_vlan_info()
1011 ppd->hv1.tp_vlan_tci = skb_vlan_tag_get(pkc->skb); in prb_fill_vlan_info()
1012 ppd->hv1.tp_vlan_tpid = ntohs(pkc->skb->vlan_proto); in prb_fill_vlan_info()
1013 ppd->tp_status = TP_STATUS_VLAN_VALID | TP_STATUS_VLAN_TPID_VALID; in prb_fill_vlan_info()
1015 ppd->hv1.tp_vlan_tci = 0; in prb_fill_vlan_info()
1016 ppd->hv1.tp_vlan_tpid = 0; in prb_fill_vlan_info()
1017 ppd->tp_status = TP_STATUS_AVAILABLE; in prb_fill_vlan_info()
1024 ppd->hv1.tp_padding = 0; in prb_run_all_ft_ops()
1027 if (pkc->feature_req_word & TP_FT_REQ_FILL_RXHASH) in prb_run_all_ft_ops()
1037 __acquires(&pkc->blk_fill_in_prog_lock) in prb_fill_curr_block()
1042 ppd->tp_next_offset = TOTAL_PKT_LEN_INCL_ALIGN(len); in prb_fill_curr_block()
1043 pkc->prev = curr; in prb_fill_curr_block()
1044 pkc->nxt_offset += TOTAL_PKT_LEN_INCL_ALIGN(len); in prb_fill_curr_block()
1047 read_lock(&pkc->blk_fill_in_prog_lock); in prb_fill_curr_block()
1051 /* Assumes caller has the sk->rx_queue.lock */
1061 pkc = GET_PBDQC_FROM_RB(&po->rx_ring); in __packet_lookup_frame_in_block()
1068 * is still in_use by user-space. in __packet_lookup_frame_in_block()
1075 * Ok, the block was released by user-space. in __packet_lookup_frame_in_block()
1085 curr = pkc->nxt_offset; in __packet_lookup_frame_in_block()
1086 pkc->skb = skb; in __packet_lookup_frame_in_block()
1087 end = (char *)pbd + pkc->kblk_size; in __packet_lookup_frame_in_block()
1107 * No free blocks are available.user_space hasn't caught up yet. in __packet_lookup_frame_in_block()
1118 switch (po->tp_version) { in packet_current_rx_frame()
1121 curr = packet_lookup_frame(po, &po->rx_ring, in packet_current_rx_frame()
1122 po->rx_ring.head, status); in packet_current_rx_frame()
1149 if (rb->prb_bdqc.kactive_blk_num) in prb_previous_blk_num()
1150 prev = rb->prb_bdqc.kactive_blk_num-1; in prb_previous_blk_num()
1152 prev = rb->prb_bdqc.knum_blocks-1; in prb_previous_blk_num()
1169 if (po->tp_version <= TPACKET_V2) in packet_previous_rx_frame()
1178 switch (po->tp_version) { in packet_increment_rx_head()
1194 unsigned int previous = rb->head ? rb->head - 1 : rb->frame_max; in packet_previous_frame()
1200 buff->head = buff->head != buff->frame_max ? buff->head+1 : 0; in packet_increment_head()
1205 this_cpu_inc(*rb->pending_refcnt); in packet_inc_pending()
1210 this_cpu_dec(*rb->pending_refcnt); in packet_dec_pending()
1219 if (rb->pending_refcnt == NULL) in packet_read_pending()
1223 refcnt += *per_cpu_ptr(rb->pending_refcnt, cpu); in packet_read_pending()
1230 po->rx_ring.pending_refcnt = NULL; in packet_alloc_pending()
1232 po->tx_ring.pending_refcnt = alloc_percpu(unsigned int); in packet_alloc_pending()
1233 if (unlikely(po->tx_ring.pending_refcnt == NULL)) in packet_alloc_pending()
1234 return -ENOBUFS; in packet_alloc_pending()
1241 free_percpu(po->tx_ring.pending_refcnt); in packet_free_pending()
1253 len = READ_ONCE(po->rx_ring.frame_max) + 1; in __tpacket_has_room()
1254 idx = READ_ONCE(po->rx_ring.head); in __tpacket_has_room()
1258 idx -= len; in __tpacket_has_room()
1259 return packet_lookup_frame(po, &po->rx_ring, idx, TP_STATUS_KERNEL); in __tpacket_has_room()
1266 len = READ_ONCE(po->rx_ring.prb_bdqc.knum_blocks); in __tpacket_v3_has_room()
1267 idx = READ_ONCE(po->rx_ring.prb_bdqc.kactive_blk_num); in __tpacket_v3_has_room()
1271 idx -= len; in __tpacket_v3_has_room()
1272 return prb_lookup_block(po, &po->rx_ring, idx, TP_STATUS_KERNEL); in __tpacket_v3_has_room()
1278 const struct sock *sk = &po->sk; in __packet_rcv_has_room()
1281 if (po->prot_hook.func != tpacket_rcv) { in __packet_rcv_has_room()
1282 int rcvbuf = READ_ONCE(sk->sk_rcvbuf); in __packet_rcv_has_room()
1283 int avail = rcvbuf - atomic_read(&sk->sk_rmem_alloc) in __packet_rcv_has_room()
1284 - (skb ? skb->truesize : 0); in __packet_rcv_has_room()
1294 if (po->tp_version == TPACKET_V3) { in __packet_rcv_has_room()
1332 skb_queue_purge(&sk->sk_error_queue); in packet_sock_destruct()
1334 WARN_ON(atomic_read(&sk->sk_rmem_alloc)); in packet_sock_destruct()
1335 WARN_ON(refcount_read(&sk->sk_wmem_alloc)); in packet_sock_destruct()
1345 u32 *history = po->rollover->history; in fanout_flow_is_huge()
1374 unsigned int val = atomic_inc_return(&f->rr_cur); in fanout_demux_lb()
1401 po = pkt_sk(rcu_dereference(f->arr[idx])); in fanout_demux_rollover()
1411 i = j = min_t(int, po->rollover->sock, num - 1); in fanout_demux_rollover()
1413 po_next = pkt_sk(rcu_dereference(f->arr[i])); in fanout_demux_rollover()
1418 po->rollover->sock = i; in fanout_demux_rollover()
1419 atomic_long_inc(&po->rollover->num); in fanout_demux_rollover()
1421 atomic_long_inc(&po->rollover->num_huge); in fanout_demux_rollover()
1429 atomic_long_inc(&po->rollover->num_failed); in fanout_demux_rollover()
1448 prog = rcu_dereference(f->bpf_prog); in fanout_demux_bpf()
1458 return f->flags & (flag >> 8); in fanout_has_flag()
1464 struct packet_fanout *f = pt->af_packet_priv; in packet_rcv_fanout()
1465 unsigned int num = READ_ONCE(f->num_members); in packet_rcv_fanout()
1466 struct net *net = read_pnet(&f->net); in packet_rcv_fanout()
1480 switch (f->type) { in packet_rcv_fanout()
1509 po = pkt_sk(rcu_dereference(f->arr[idx])); in packet_rcv_fanout()
1510 return po->prot_hook.func(skb, dev, &po->prot_hook, orig_dev); in packet_rcv_fanout()
1520 struct packet_fanout *f = po->fanout; in __fanout_link()
1522 spin_lock(&f->lock); in __fanout_link()
1523 rcu_assign_pointer(f->arr[f->num_members], sk); in __fanout_link()
1525 f->num_members++; in __fanout_link()
1526 if (f->num_members == 1) in __fanout_link()
1527 dev_add_pack(&f->prot_hook); in __fanout_link()
1528 spin_unlock(&f->lock); in __fanout_link()
1533 struct packet_fanout *f = po->fanout; in __fanout_unlink()
1536 spin_lock(&f->lock); in __fanout_unlink()
1537 for (i = 0; i < f->num_members; i++) { in __fanout_unlink()
1538 if (rcu_dereference_protected(f->arr[i], in __fanout_unlink()
1539 lockdep_is_held(&f->lock)) == sk) in __fanout_unlink()
1542 BUG_ON(i >= f->num_members); in __fanout_unlink()
1543 rcu_assign_pointer(f->arr[i], in __fanout_unlink()
1544 rcu_dereference_protected(f->arr[f->num_members - 1], in __fanout_unlink()
1545 lockdep_is_held(&f->lock))); in __fanout_unlink()
1546 f->num_members--; in __fanout_unlink()
1547 if (f->num_members == 0) in __fanout_unlink()
1548 __dev_remove_pack(&f->prot_hook); in __fanout_unlink()
1549 spin_unlock(&f->lock); in __fanout_unlink()
1554 if (sk->sk_family != PF_PACKET) in match_fanout_group()
1557 return ptype->af_packet_priv == pkt_sk(sk)->fanout; in match_fanout_group()
1562 switch (f->type) { in fanout_init_data()
1564 atomic_set(&f->rr_cur, 0); in fanout_init_data()
1568 RCU_INIT_POINTER(f->bpf_prog, NULL); in fanout_init_data()
1577 spin_lock(&f->lock); in __fanout_set_data_bpf()
1578 old = rcu_dereference_protected(f->bpf_prog, lockdep_is_held(&f->lock)); in __fanout_set_data_bpf()
1579 rcu_assign_pointer(f->bpf_prog, new); in __fanout_set_data_bpf()
1580 spin_unlock(&f->lock); in __fanout_set_data_bpf()
1595 if (sock_flag(&po->sk, SOCK_FILTER_LOCKED)) in fanout_set_data_cbpf()
1596 return -EPERM; in fanout_set_data_cbpf()
1606 __fanout_set_data_bpf(po->fanout, new); in fanout_set_data_cbpf()
1616 if (sock_flag(&po->sk, SOCK_FILTER_LOCKED)) in fanout_set_data_ebpf()
1617 return -EPERM; in fanout_set_data_ebpf()
1619 return -EINVAL; in fanout_set_data_ebpf()
1621 return -EFAULT; in fanout_set_data_ebpf()
1627 __fanout_set_data_bpf(po->fanout, new); in fanout_set_data_ebpf()
1634 switch (po->fanout->type) { in fanout_set_data()
1640 return -EINVAL; in fanout_set_data()
1646 switch (f->type) { in fanout_release_data()
1658 if (f->id == candidate_id && in __fanout_id_is_free()
1659 read_pnet(&f->net) == sock_net(sk)) { in __fanout_id_is_free()
1685 struct packet_rollover *rollover = NULL; in fanout_add() local
1687 u16 type_flags = args->type_flags; in fanout_add()
1691 u16 id = args->id; in fanout_add()
1697 return -EINVAL; in fanout_add()
1708 return -EINVAL; in fanout_add()
1713 err = -EALREADY; in fanout_add()
1714 if (po->fanout) in fanout_add()
1719 err = -ENOMEM; in fanout_add()
1720 rollover = kzalloc(sizeof(*rollover), GFP_KERNEL); in fanout_add()
1721 if (!rollover) in fanout_add()
1723 atomic_long_set(&rollover->num, 0); in fanout_add()
1724 atomic_long_set(&rollover->num_huge, 0); in fanout_add()
1725 atomic_long_set(&rollover->num_failed, 0); in fanout_add()
1730 err = -EINVAL; in fanout_add()
1734 err = -ENOMEM; in fanout_add()
1743 if (f->id == id && in fanout_add()
1744 read_pnet(&f->net) == sock_net(sk)) { in fanout_add()
1749 err = -EINVAL; in fanout_add()
1751 if (match->flags != flags) in fanout_add()
1753 if (args->max_num_members && in fanout_add()
1754 args->max_num_members != match->max_num_members) in fanout_add()
1757 if (args->max_num_members > PACKET_FANOUT_MAX) in fanout_add()
1759 if (!args->max_num_members) in fanout_add()
1761 args->max_num_members = 256; in fanout_add()
1762 err = -ENOMEM; in fanout_add()
1763 match = kvzalloc(struct_size(match, arr, args->max_num_members), in fanout_add()
1767 write_pnet(&match->net, sock_net(sk)); in fanout_add()
1768 match->id = id; in fanout_add()
1769 match->type = type; in fanout_add()
1770 match->flags = flags; in fanout_add()
1771 INIT_LIST_HEAD(&match->list); in fanout_add()
1772 spin_lock_init(&match->lock); in fanout_add()
1773 refcount_set(&match->sk_ref, 0); in fanout_add()
1775 match->prot_hook.type = po->prot_hook.type; in fanout_add()
1776 match->prot_hook.dev = po->prot_hook.dev; in fanout_add()
1777 match->prot_hook.func = packet_rcv_fanout; in fanout_add()
1778 match->prot_hook.af_packet_priv = match; in fanout_add()
1779 match->prot_hook.af_packet_net = read_pnet(&match->net); in fanout_add()
1780 match->prot_hook.id_match = match_fanout_group; in fanout_add()
1781 match->max_num_members = args->max_num_members; in fanout_add()
1782 match->prot_hook.ignore_outgoing = type_flags & PACKET_FANOUT_FLAG_IGNORE_OUTGOING; in fanout_add()
1783 list_add(&match->list, &fanout_list); in fanout_add()
1785 err = -EINVAL; in fanout_add()
1787 spin_lock(&po->bind_lock); in fanout_add()
1789 match->type == type && in fanout_add()
1790 match->prot_hook.type == po->prot_hook.type && in fanout_add()
1791 match->prot_hook.dev == po->prot_hook.dev) { in fanout_add()
1792 err = -ENOSPC; in fanout_add()
1793 if (refcount_read(&match->sk_ref) < match->max_num_members) { in fanout_add()
1794 __dev_remove_pack(&po->prot_hook); in fanout_add()
1797 WRITE_ONCE(po->fanout, match); in fanout_add()
1799 po->rollover = rollover; in fanout_add()
1800 rollover = NULL; in fanout_add()
1801 refcount_set(&match->sk_ref, refcount_read(&match->sk_ref) + 1); in fanout_add()
1806 spin_unlock(&po->bind_lock); in fanout_add()
1808 if (err && !refcount_read(&match->sk_ref)) { in fanout_add()
1809 list_del(&match->list); in fanout_add()
1814 kfree(rollover); in fanout_add()
1819 /* If pkt_sk(sk)->fanout->sk_ref is zero, this function removes
1820 * pkt_sk(sk)->fanout from fanout_list and returns pkt_sk(sk)->fanout.
1830 f = po->fanout; in fanout_release()
1832 po->fanout = NULL; in fanout_release()
1834 if (refcount_dec_and_test(&f->sk_ref)) in fanout_release()
1835 list_del(&f->list); in fanout_release()
1847 /* Earlier code assumed this would be a VLAN pkt, double-check in packet_extra_vlan_len_allowed()
1851 if (unlikely(dev->type != ARPHRD_ETHER)) in packet_extra_vlan_len_allowed()
1855 return likely(eth_hdr(skb)->h_proto == htons(ETH_P_8021Q)); in packet_extra_vlan_len_allowed()
1873 sk = pt->af_packet_priv; in packet_rcv_spkt()
1882 * For outgoing ones skb->data == skb_mac_header(skb) in packet_rcv_spkt()
1886 if (skb->pkt_type == PACKET_LOOPBACK) in packet_rcv_spkt()
1902 spkt = &PACKET_SKB_CB(skb)->sa.pkt; in packet_rcv_spkt()
1904 skb_push(skb, skb->data - skb_mac_header(skb)); in packet_rcv_spkt()
1910 spkt->spkt_family = dev->type; in packet_rcv_spkt()
1911 strscpy(spkt->spkt_device, dev->name, sizeof(spkt->spkt_device)); in packet_rcv_spkt()
1912 spkt->spkt_protocol = skb->protocol; in packet_rcv_spkt()
1932 if ((!skb->protocol || skb->protocol == htons(ETH_P_ALL)) && in packet_parse_headers()
1933 sock->type == SOCK_RAW) { in packet_parse_headers()
1935 skb->protocol = dev_parse_header_protocol(skb); in packet_parse_headers()
1939 if (likely(skb->dev->type == ARPHRD_ETHER) && in packet_parse_headers()
1940 eth_type_vlan(skb->protocol) && in packet_parse_headers()
1941 vlan_get_protocol_and_depth(skb, skb->protocol, &depth) != 0) in packet_parse_headers()
1955 struct sock *sk = sock->sk; in packet_sendmsg_spkt()
1956 DECLARE_SOCKADDR(struct sockaddr_pkt *, saddr, msg->msg_name); in packet_sendmsg_spkt()
1969 if (msg->msg_namelen < sizeof(struct sockaddr)) in packet_sendmsg_spkt()
1970 return -EINVAL; in packet_sendmsg_spkt()
1971 if (msg->msg_namelen == sizeof(struct sockaddr_pkt)) in packet_sendmsg_spkt()
1972 proto = saddr->spkt_protocol; in packet_sendmsg_spkt()
1974 return -ENOTCONN; /* SOCK_PACKET must be sent giving an address */ in packet_sendmsg_spkt()
1980 saddr->spkt_device[sizeof(saddr->spkt_device) - 1] = 0; in packet_sendmsg_spkt()
1983 dev = dev_get_by_name_rcu(sock_net(sk), saddr->spkt_device); in packet_sendmsg_spkt()
1984 err = -ENODEV; in packet_sendmsg_spkt()
1988 err = -ENETDOWN; in packet_sendmsg_spkt()
1989 if (!(dev->flags & IFF_UP)) in packet_sendmsg_spkt()
1999 err = -EPROTONOSUPPORT; in packet_sendmsg_spkt()
2005 err = -EMSGSIZE; in packet_sendmsg_spkt()
2006 if (len > dev->mtu + dev->hard_header_len + VLAN_HLEN + extra_len) in packet_sendmsg_spkt()
2011 int tlen = dev->needed_tailroom; in packet_sendmsg_spkt()
2012 unsigned int hhlen = dev->header_ops ? dev->hard_header_len : 0; in packet_sendmsg_spkt()
2017 return -ENOBUFS; in packet_sendmsg_spkt()
2027 skb->data -= hhlen; in packet_sendmsg_spkt()
2028 skb->tail -= hhlen; in packet_sendmsg_spkt()
2038 if (!dev_validate_header(dev, skb->data, len) || !skb->len) { in packet_sendmsg_spkt()
2039 err = -EINVAL; in packet_sendmsg_spkt()
2042 if (len > (dev->mtu + dev->hard_header_len + extra_len) && in packet_sendmsg_spkt()
2044 err = -EMSGSIZE; in packet_sendmsg_spkt()
2049 if (msg->msg_controllen) { in packet_sendmsg_spkt()
2055 skb->protocol = proto; in packet_sendmsg_spkt()
2056 skb->dev = dev; in packet_sendmsg_spkt()
2057 skb->priority = READ_ONCE(sk->sk_priority); in packet_sendmsg_spkt()
2058 skb->mark = READ_ONCE(sk->sk_mark); in packet_sendmsg_spkt()
2059 skb->tstamp = sockc.transmit_time; in packet_sendmsg_spkt()
2064 skb->no_fcs = 1; in packet_sendmsg_spkt()
2086 filter = rcu_dereference(sk->sk_filter); in run_filter()
2088 res = bpf_prog_run_clear_cb(filter->prog, skb); in run_filter()
2100 return -EINVAL; in packet_rcv_vnet()
2101 *len -= vnet_hdr_sz; in packet_rcv_vnet()
2104 return -EINVAL; in packet_rcv_vnet()
2113 * Note tricky part: we DO mangle shared skb! skb->data, skb->len
2114 * and skb->cb are mangled. It works because (and until) packets
2128 u8 *skb_head = skb->data; in packet_rcv()
2129 int skb_len = skb->len; in packet_rcv()
2132 if (skb->pkt_type == PACKET_LOOPBACK) in packet_rcv()
2135 sk = pt->af_packet_priv; in packet_rcv()
2141 skb->dev = dev; in packet_rcv()
2151 if (sk->sk_type != SOCK_DGRAM) in packet_rcv()
2152 skb_push(skb, skb->data - skb_mac_header(skb)); in packet_rcv()
2153 else if (skb->pkt_type == PACKET_OUTGOING) { in packet_rcv()
2159 snaplen = skb->len; in packet_rcv()
2167 if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) in packet_rcv()
2175 if (skb_head != skb->data) { in packet_rcv()
2176 skb->data = skb_head; in packet_rcv()
2177 skb->len = skb_len; in packet_rcv()
2183 sock_skb_cb_check_size(sizeof(*PACKET_SKB_CB(skb)) + MAX_ADDR_LEN - 8); in packet_rcv()
2185 sll = &PACKET_SKB_CB(skb)->sa.ll; in packet_rcv()
2186 sll->sll_hatype = dev->type; in packet_rcv()
2187 sll->sll_pkttype = skb->pkt_type; in packet_rcv()
2189 sll->sll_ifindex = orig_dev->ifindex; in packet_rcv()
2191 sll->sll_ifindex = dev->ifindex; in packet_rcv()
2193 sll->sll_halen = dev_parse_header(skb, sll->sll_addr); in packet_rcv()
2195 /* sll->sll_family and sll->sll_protocol are set in packet_recvmsg(). in packet_rcv()
2198 PACKET_SKB_CB(skb)->sa.origlen = skb->len; in packet_rcv()
2204 skb->dev = NULL; in packet_rcv()
2210 spin_lock(&sk->sk_receive_queue.lock); in packet_rcv()
2211 po->stats.stats1.tp_packets++; in packet_rcv()
2214 __skb_queue_tail(&sk->sk_receive_queue, skb); in packet_rcv()
2215 spin_unlock(&sk->sk_receive_queue.lock); in packet_rcv()
2216 sk->sk_data_ready(sk); in packet_rcv()
2220 atomic_inc(&po->tp_drops); in packet_rcv()
2221 atomic_inc(&sk->sk_drops); in packet_rcv()
2225 if (skb_head != skb->data && skb_shared(skb)) { in packet_rcv()
2226 skb->data = skb_head; in packet_rcv()
2227 skb->len = skb_len; in packet_rcv()
2242 u8 *skb_head = skb->data; in tpacket_rcv()
2243 int skb_len = skb->len; in tpacket_rcv()
2261 if (skb->pkt_type == PACKET_LOOPBACK) in tpacket_rcv()
2264 sk = pt->af_packet_priv; in tpacket_rcv()
2271 if (sk->sk_type != SOCK_DGRAM) in tpacket_rcv()
2272 skb_push(skb, skb->data - skb_mac_header(skb)); in tpacket_rcv()
2273 else if (skb->pkt_type == PACKET_OUTGOING) { in tpacket_rcv()
2279 snaplen = skb->len; in tpacket_rcv()
2287 atomic_inc(&po->tp_drops); in tpacket_rcv()
2291 if (skb->ip_summed == CHECKSUM_PARTIAL) in tpacket_rcv()
2293 else if (skb->pkt_type != PACKET_OUTGOING && in tpacket_rcv()
2302 if (sk->sk_type == SOCK_DGRAM) { in tpacket_rcv()
2303 macoff = netoff = TPACKET_ALIGN(po->tp_hdrlen) + 16 + in tpacket_rcv()
2304 po->tp_reserve; in tpacket_rcv()
2307 netoff = TPACKET_ALIGN(po->tp_hdrlen + in tpacket_rcv()
2309 po->tp_reserve; in tpacket_rcv()
2310 vnet_hdr_sz = READ_ONCE(po->vnet_hdr_sz); in tpacket_rcv()
2313 macoff = netoff - maclen; in tpacket_rcv()
2316 atomic_inc(&po->tp_drops); in tpacket_rcv()
2319 if (po->tp_version <= TPACKET_V2) { in tpacket_rcv()
2320 if (macoff + snaplen > po->rx_ring.frame_size) { in tpacket_rcv()
2321 if (po->copy_thresh && in tpacket_rcv()
2322 atomic_read(&sk->sk_rmem_alloc) < sk->sk_rcvbuf) { in tpacket_rcv()
2327 skb_head = skb->data; in tpacket_rcv()
2330 memset(&PACKET_SKB_CB(copy_skb)->sa.ll, 0, in tpacket_rcv()
2331 sizeof(PACKET_SKB_CB(copy_skb)->sa.ll)); in tpacket_rcv()
2335 snaplen = po->rx_ring.frame_size - macoff; in tpacket_rcv()
2342 GET_PBDQC_FROM_RB(&po->rx_ring)->max_frame_len)) { in tpacket_rcv()
2345 nval = GET_PBDQC_FROM_RB(&po->rx_ring)->max_frame_len - macoff; in tpacket_rcv()
2351 macoff = GET_PBDQC_FROM_RB(&po->rx_ring)->max_frame_len; in tpacket_rcv()
2355 spin_lock(&sk->sk_receive_queue.lock); in tpacket_rcv()
2361 if (po->tp_version <= TPACKET_V2) { in tpacket_rcv()
2362 slot_id = po->rx_ring.head; in tpacket_rcv()
2363 if (test_bit(slot_id, po->rx_ring.rx_owner_map)) in tpacket_rcv()
2365 __set_bit(slot_id, po->rx_ring.rx_owner_map); in tpacket_rcv()
2369 virtio_net_hdr_from_skb(skb, h.raw + macoff - in tpacket_rcv()
2372 if (po->tp_version == TPACKET_V3) in tpacket_rcv()
2373 prb_clear_blk_fill_status(&po->rx_ring); in tpacket_rcv()
2377 if (po->tp_version <= TPACKET_V2) { in tpacket_rcv()
2378 packet_increment_rx_head(po, &po->rx_ring); in tpacket_rcv()
2380 * LOSING will be reported till you read the stats, in tpacket_rcv()
2381 * because it's COR - Clear On Read. in tpacket_rcv()
2385 if (atomic_read(&po->tp_drops)) in tpacket_rcv()
2389 po->stats.stats1.tp_packets++; in tpacket_rcv()
2393 __skb_queue_tail(&sk->sk_receive_queue, copy_skb); in tpacket_rcv()
2395 spin_unlock(&sk->sk_receive_queue.lock); in tpacket_rcv()
2403 READ_ONCE(po->tp_tstamp) | in tpacket_rcv()
2410 switch (po->tp_version) { in tpacket_rcv()
2412 h.h1->tp_len = skb->len; in tpacket_rcv()
2413 h.h1->tp_snaplen = snaplen; in tpacket_rcv()
2414 h.h1->tp_mac = macoff; in tpacket_rcv()
2415 h.h1->tp_net = netoff; in tpacket_rcv()
2416 h.h1->tp_sec = ts.tv_sec; in tpacket_rcv()
2417 h.h1->tp_usec = ts.tv_nsec / NSEC_PER_USEC; in tpacket_rcv()
2421 h.h2->tp_len = skb->len; in tpacket_rcv()
2422 h.h2->tp_snaplen = snaplen; in tpacket_rcv()
2423 h.h2->tp_mac = macoff; in tpacket_rcv()
2424 h.h2->tp_net = netoff; in tpacket_rcv()
2425 h.h2->tp_sec = ts.tv_sec; in tpacket_rcv()
2426 h.h2->tp_nsec = ts.tv_nsec; in tpacket_rcv()
2428 h.h2->tp_vlan_tci = skb_vlan_tag_get(skb); in tpacket_rcv()
2429 h.h2->tp_vlan_tpid = ntohs(skb->vlan_proto); in tpacket_rcv()
2432 h.h2->tp_vlan_tci = 0; in tpacket_rcv()
2433 h.h2->tp_vlan_tpid = 0; in tpacket_rcv()
2435 memset(h.h2->tp_padding, 0, sizeof(h.h2->tp_padding)); in tpacket_rcv()
2442 h.h3->tp_status |= status; in tpacket_rcv()
2443 h.h3->tp_len = skb->len; in tpacket_rcv()
2444 h.h3->tp_snaplen = snaplen; in tpacket_rcv()
2445 h.h3->tp_mac = macoff; in tpacket_rcv()
2446 h.h3->tp_net = netoff; in tpacket_rcv()
2447 h.h3->tp_sec = ts.tv_sec; in tpacket_rcv()
2448 h.h3->tp_nsec = ts.tv_nsec; in tpacket_rcv()
2449 memset(h.h3->tp_padding, 0, sizeof(h.h3->tp_padding)); in tpacket_rcv()
2457 sll->sll_halen = dev_parse_header(skb, sll->sll_addr); in tpacket_rcv()
2458 sll->sll_family = AF_PACKET; in tpacket_rcv()
2459 sll->sll_hatype = dev->type; in tpacket_rcv()
2460 sll->sll_protocol = skb->protocol; in tpacket_rcv()
2461 sll->sll_pkttype = skb->pkt_type; in tpacket_rcv()
2463 sll->sll_ifindex = orig_dev->ifindex; in tpacket_rcv()
2465 sll->sll_ifindex = dev->ifindex; in tpacket_rcv()
2470 if (po->tp_version <= TPACKET_V2) { in tpacket_rcv()
2482 if (po->tp_version <= TPACKET_V2) { in tpacket_rcv()
2483 spin_lock(&sk->sk_receive_queue.lock); in tpacket_rcv()
2485 __clear_bit(slot_id, po->rx_ring.rx_owner_map); in tpacket_rcv()
2486 spin_unlock(&sk->sk_receive_queue.lock); in tpacket_rcv()
2487 sk->sk_data_ready(sk); in tpacket_rcv()
2488 } else if (po->tp_version == TPACKET_V3) { in tpacket_rcv()
2489 prb_clear_blk_fill_status(&po->rx_ring); in tpacket_rcv()
2493 if (skb_head != skb->data && skb_shared(skb)) { in tpacket_rcv()
2494 skb->data = skb_head; in tpacket_rcv()
2495 skb->len = skb_len; in tpacket_rcv()
2502 spin_unlock(&sk->sk_receive_queue.lock); in tpacket_rcv()
2503 atomic_inc(&po->tp_drops); in tpacket_rcv()
2506 sk->sk_data_ready(sk); in tpacket_rcv()
2513 struct packet_sock *po = pkt_sk(skb->sk); in tpacket_destruct_skb()
2515 if (likely(po->tx_ring.pg_vec)) { in tpacket_destruct_skb()
2520 packet_dec_pending(&po->tx_ring); in tpacket_destruct_skb()
2525 if (!packet_read_pending(&po->tx_ring)) in tpacket_destruct_skb()
2526 complete(&po->skb_completion); in tpacket_destruct_skb()
2534 if ((vnet_hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) && in __packet_snd_vnet_parse()
2535 (__virtio16_to_cpu(vio_le(), vnet_hdr->csum_start) + in __packet_snd_vnet_parse()
2536 __virtio16_to_cpu(vio_le(), vnet_hdr->csum_offset) + 2 > in __packet_snd_vnet_parse()
2537 __virtio16_to_cpu(vio_le(), vnet_hdr->hdr_len))) in __packet_snd_vnet_parse()
2538 vnet_hdr->hdr_len = __cpu_to_virtio16(vio_le(), in __packet_snd_vnet_parse()
2539 __virtio16_to_cpu(vio_le(), vnet_hdr->csum_start) + in __packet_snd_vnet_parse()
2540 __virtio16_to_cpu(vio_le(), vnet_hdr->csum_offset) + 2); in __packet_snd_vnet_parse()
2542 if (__virtio16_to_cpu(vio_le(), vnet_hdr->hdr_len) > len) in __packet_snd_vnet_parse()
2543 return -EINVAL; in __packet_snd_vnet_parse()
2554 return -EINVAL; in packet_snd_vnet_parse()
2555 *len -= vnet_hdr_sz; in packet_snd_vnet_parse()
2557 if (!copy_from_iter_full(vnet_hdr, sizeof(*vnet_hdr), &msg->msg_iter)) in packet_snd_vnet_parse()
2558 return -EFAULT; in packet_snd_vnet_parse()
2566 iov_iter_advance(&msg->msg_iter, vnet_hdr_sz - sizeof(struct virtio_net_hdr)); in packet_snd_vnet_parse()
2578 struct socket *sock = po->sk.sk_socket; in tpacket_fill_skb()
2584 skb->protocol = proto; in tpacket_fill_skb()
2585 skb->dev = dev; in tpacket_fill_skb()
2586 skb->priority = READ_ONCE(po->sk.sk_priority); in tpacket_fill_skb()
2587 skb->mark = READ_ONCE(po->sk.sk_mark); in tpacket_fill_skb()
2588 skb->tstamp = sockc->transmit_time; in tpacket_fill_skb()
2589 skb_setup_tx_timestamp(skb, sockc->tsflags); in tpacket_fill_skb()
2597 if (sock->type == SOCK_DGRAM) { in tpacket_fill_skb()
2601 return -EINVAL; in tpacket_fill_skb()
2605 skb_push(skb, dev->hard_header_len); in tpacket_fill_skb()
2606 skb_put(skb, copylen - dev->hard_header_len); in tpacket_fill_skb()
2610 if (!dev_validate_header(dev, skb->data, hdrlen)) in tpacket_fill_skb()
2611 return -EINVAL; in tpacket_fill_skb()
2614 to_write -= hdrlen; in tpacket_fill_skb()
2618 len_max = PAGE_SIZE - offset; in tpacket_fill_skb()
2621 skb->data_len = to_write; in tpacket_fill_skb()
2622 skb->len += to_write; in tpacket_fill_skb()
2623 skb->truesize += to_write; in tpacket_fill_skb()
2624 refcount_add(to_write, &po->sk.sk_wmem_alloc); in tpacket_fill_skb()
2627 nr_frags = skb_shinfo(skb)->nr_frags; in tpacket_fill_skb()
2632 return -EFAULT; in tpacket_fill_skb()
2640 to_write -= len; in tpacket_fill_skb()
2659 switch (po->tp_version) { in tpacket_parse_header()
2661 if (ph.h3->tp_next_offset != 0) { in tpacket_parse_header()
2663 return -EINVAL; in tpacket_parse_header()
2665 tp_len = ph.h3->tp_len; in tpacket_parse_header()
2668 tp_len = ph.h2->tp_len; in tpacket_parse_header()
2671 tp_len = ph.h1->tp_len; in tpacket_parse_header()
2676 return -EMSGSIZE; in tpacket_parse_header()
2682 off_min = po->tp_hdrlen - sizeof(struct sockaddr_ll); in tpacket_parse_header()
2683 off_max = po->tx_ring.frame_size - tp_len; in tpacket_parse_header()
2684 if (po->sk.sk_type == SOCK_DGRAM) { in tpacket_parse_header()
2685 switch (po->tp_version) { in tpacket_parse_header()
2687 off = ph.h3->tp_net; in tpacket_parse_header()
2690 off = ph.h2->tp_net; in tpacket_parse_header()
2693 off = ph.h1->tp_net; in tpacket_parse_header()
2697 switch (po->tp_version) { in tpacket_parse_header()
2699 off = ph.h3->tp_mac; in tpacket_parse_header()
2702 off = ph.h2->tp_mac; in tpacket_parse_header()
2705 off = ph.h1->tp_mac; in tpacket_parse_header()
2710 return -EINVAL; in tpacket_parse_header()
2712 off = po->tp_hdrlen - sizeof(struct sockaddr_ll); in tpacket_parse_header()
2728 DECLARE_SOCKADDR(struct sockaddr_ll *, saddr, msg->msg_name); in tpacket_snd()
2729 bool need_wait = !(msg->msg_flags & MSG_DONTWAIT); in tpacket_snd()
2730 int vnet_hdr_sz = READ_ONCE(po->vnet_hdr_sz); in tpacket_snd()
2739 mutex_lock(&po->pg_vec_lock); in tpacket_snd()
2744 if (unlikely(!po->tx_ring.pg_vec)) { in tpacket_snd()
2745 err = -EBUSY; in tpacket_snd()
2750 proto = READ_ONCE(po->num); in tpacket_snd()
2752 err = -EINVAL; in tpacket_snd()
2753 if (msg->msg_namelen < sizeof(struct sockaddr_ll)) in tpacket_snd()
2755 if (msg->msg_namelen < (saddr->sll_halen in tpacket_snd()
2759 proto = saddr->sll_protocol; in tpacket_snd()
2760 dev = dev_get_by_index(sock_net(&po->sk), saddr->sll_ifindex); in tpacket_snd()
2761 if (po->sk.sk_socket->type == SOCK_DGRAM) { in tpacket_snd()
2762 if (dev && msg->msg_namelen < dev->addr_len + in tpacket_snd()
2765 addr = saddr->sll_addr; in tpacket_snd()
2769 err = -ENXIO; in tpacket_snd()
2772 err = -ENETDOWN; in tpacket_snd()
2773 if (unlikely(!(dev->flags & IFF_UP))) in tpacket_snd()
2776 sockcm_init(&sockc, &po->sk); in tpacket_snd()
2777 if (msg->msg_controllen) { in tpacket_snd()
2778 err = sock_cmsg_send(&po->sk, msg, &sockc); in tpacket_snd()
2783 if (po->sk.sk_socket->type == SOCK_RAW) in tpacket_snd()
2784 reserve = dev->hard_header_len; in tpacket_snd()
2785 size_max = po->tx_ring.frame_size in tpacket_snd()
2786 - (po->tp_hdrlen - sizeof(struct sockaddr_ll)); in tpacket_snd()
2788 if ((size_max > dev->mtu + reserve + VLAN_HLEN) && !vnet_hdr_sz) in tpacket_snd()
2789 size_max = dev->mtu + reserve + VLAN_HLEN; in tpacket_snd()
2791 reinit_completion(&po->skb_completion); in tpacket_snd()
2794 ph = packet_current_frame(po, &po->tx_ring, in tpacket_snd()
2798 timeo = sock_sndtimeo(&po->sk, msg->msg_flags & MSG_DONTWAIT); in tpacket_snd()
2799 timeo = wait_for_completion_interruptible_timeout(&po->skb_completion, timeo); in tpacket_snd()
2801 err = !timeo ? -ETIMEDOUT : -ERESTARTSYS; in tpacket_snd()
2816 tlen = dev->needed_tailroom; in tpacket_snd()
2820 tp_len -= vnet_hdr_sz; in tpacket_snd()
2823 tp_len = -EINVAL; in tpacket_snd()
2827 vnet_hdr->hdr_len); in tpacket_snd()
2829 copylen = max_t(int, copylen, dev->hard_header_len); in tpacket_snd()
2830 skb = sock_alloc_send_skb(&po->sk, in tpacket_snd()
2832 (copylen - dev->hard_header_len), in tpacket_snd()
2844 tp_len > dev->mtu + reserve && in tpacket_snd()
2847 tp_len = -EMSGSIZE; in tpacket_snd()
2854 packet_increment_head(&po->tx_ring); in tpacket_snd()
2866 tp_len = -EINVAL; in tpacket_snd()
2872 skb->destructor = tpacket_destruct_skb; in tpacket_snd()
2874 packet_inc_pending(&po->tx_ring); in tpacket_snd()
2893 packet_increment_head(&po->tx_ring); in tpacket_snd()
2897 * to call it as it's per_cpu variable, but in fast-path in tpacket_snd()
2898 * we already short-circuit the loop with the first in tpacket_snd()
2902 (need_wait && packet_read_pending(&po->tx_ring)))); in tpacket_snd()
2913 mutex_unlock(&po->pg_vec_lock); in tpacket_snd()
2928 if (len - linear > MAX_SKB_FRAGS * (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) in packet_alloc_skb()
2929 linear = len - MAX_SKB_FRAGS * (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER); in packet_alloc_skb()
2930 skb = sock_alloc_send_pskb(sk, prepad + linear, len - linear, noblock, in packet_alloc_skb()
2937 skb->data_len = len - linear; in packet_alloc_skb()
2938 skb->len += len - linear; in packet_alloc_skb()
2945 struct sock *sk = sock->sk; in packet_snd()
2946 DECLARE_SOCKADDR(struct sockaddr_ll *, saddr, msg->msg_name); in packet_snd()
2956 int vnet_hdr_sz = READ_ONCE(po->vnet_hdr_sz); in packet_snd()
2966 proto = READ_ONCE(po->num); in packet_snd()
2968 err = -EINVAL; in packet_snd()
2969 if (msg->msg_namelen < sizeof(struct sockaddr_ll)) in packet_snd()
2971 if (msg->msg_namelen < (saddr->sll_halen + offsetof(struct sockaddr_ll, sll_addr))) in packet_snd()
2973 proto = saddr->sll_protocol; in packet_snd()
2974 dev = dev_get_by_index(sock_net(sk), saddr->sll_ifindex); in packet_snd()
2975 if (sock->type == SOCK_DGRAM) { in packet_snd()
2976 if (dev && msg->msg_namelen < dev->addr_len + in packet_snd()
2979 addr = saddr->sll_addr; in packet_snd()
2983 err = -ENXIO; in packet_snd()
2986 err = -ENETDOWN; in packet_snd()
2987 if (unlikely(!(dev->flags & IFF_UP))) in packet_snd()
2991 sockc.mark = READ_ONCE(sk->sk_mark); in packet_snd()
2992 if (msg->msg_controllen) { in packet_snd()
2998 if (sock->type == SOCK_RAW) in packet_snd()
2999 reserve = dev->hard_header_len; in packet_snd()
3008 err = -EPROTONOSUPPORT; in packet_snd()
3014 err = -EMSGSIZE; in packet_snd()
3016 (len > dev->mtu + reserve + VLAN_HLEN + extra_len)) in packet_snd()
3019 err = -ENOBUFS; in packet_snd()
3021 tlen = dev->needed_tailroom; in packet_snd()
3023 linear = max(linear, min_t(int, len, dev->hard_header_len)); in packet_snd()
3025 msg->msg_flags & MSG_DONTWAIT, &err); in packet_snd()
3031 err = -EINVAL; in packet_snd()
3032 if (sock->type == SOCK_DGRAM) { in packet_snd()
3037 skb_reserve(skb, -reserve); in packet_snd()
3039 dev->min_header_len != dev->hard_header_len) in packet_snd()
3043 /* Returns -EFAULT on error */ in packet_snd()
3044 err = skb_copy_datagram_from_iter(skb, offset, &msg->msg_iter, len); in packet_snd()
3048 if ((sock->type == SOCK_RAW && in packet_snd()
3049 !dev_validate_header(dev, skb->data, len)) || !skb->len) { in packet_snd()
3050 err = -EINVAL; in packet_snd()
3056 if (!vnet_hdr.gso_type && (len > dev->mtu + reserve + extra_len) && in packet_snd()
3058 err = -EMSGSIZE; in packet_snd()
3062 skb->protocol = proto; in packet_snd()
3063 skb->dev = dev; in packet_snd()
3064 skb->priority = READ_ONCE(sk->sk_priority); in packet_snd()
3065 skb->mark = sockc.mark; in packet_snd()
3066 skb->tstamp = sockc.transmit_time; in packet_snd()
3069 skb->no_fcs = 1; in packet_snd()
3104 struct sock *sk = sock->sk; in packet_sendmsg()
3110 if (data_race(po->tx_ring.pg_vec)) in packet_sendmsg()
3123 struct sock *sk = sock->sk; in packet_release()
3135 mutex_lock(&net->packet.sklist_lock); in packet_release()
3137 mutex_unlock(&net->packet.sklist_lock); in packet_release()
3139 sock_prot_inuse_add(net, sk->sk_prot, -1); in packet_release()
3141 spin_lock(&po->bind_lock); in packet_release()
3145 if (po->prot_hook.dev) { in packet_release()
3146 netdev_put(po->prot_hook.dev, &po->prot_hook.dev_tracker); in packet_release()
3147 po->prot_hook.dev = NULL; in packet_release()
3149 spin_unlock(&po->bind_lock); in packet_release()
3154 if (po->rx_ring.pg_vec) { in packet_release()
3159 if (po->tx_ring.pg_vec) { in packet_release()
3169 kfree(po->rollover); in packet_release()
3175 * Now the socket is dead. No more input will appear. in packet_release()
3178 sock->sk = NULL; in packet_release()
3182 skb_queue_purge(&sk->sk_receive_queue); in packet_release()
3203 spin_lock(&po->bind_lock); in packet_do_bind()
3205 proto = po->num; in packet_do_bind()
3209 if (po->fanout) { in packet_do_bind()
3210 ret = -EINVAL; in packet_do_bind()
3217 ret = -ENODEV; in packet_do_bind()
3223 ret = -ENODEV; in packet_do_bind()
3228 need_rehook = po->prot_hook.type != proto || po->prot_hook.dev != dev; in packet_do_bind()
3237 WRITE_ONCE(po->num, 0); in packet_do_bind()
3242 dev->ifindex); in packet_do_bind()
3246 WRITE_ONCE(po->num, proto); in packet_do_bind()
3247 po->prot_hook.type = proto; in packet_do_bind()
3249 netdev_put(po->prot_hook.dev, &po->prot_hook.dev_tracker); in packet_do_bind()
3252 po->prot_hook.dev = NULL; in packet_do_bind()
3253 WRITE_ONCE(po->ifindex, -1); in packet_do_bind()
3256 netdev_hold(dev, &po->prot_hook.dev_tracker, in packet_do_bind()
3258 po->prot_hook.dev = dev; in packet_do_bind()
3259 WRITE_ONCE(po->ifindex, dev ? dev->ifindex : 0); in packet_do_bind()
3268 if (!unlisted && (!dev || (dev->flags & IFF_UP))) { in packet_do_bind()
3271 sk->sk_err = ENETDOWN; in packet_do_bind()
3278 spin_unlock(&po->bind_lock); in packet_do_bind()
3290 struct sock *sk = sock->sk; in packet_bind_spkt()
3291 char name[sizeof(uaddr->sa_data_min) + 1]; in packet_bind_spkt()
3298 return -EINVAL; in packet_bind_spkt()
3299 /* uaddr->sa_data comes from the userspace, it's not guaranteed to be in packet_bind_spkt()
3300 * zero-terminated. in packet_bind_spkt()
3302 memcpy(name, uaddr->sa_data, sizeof(uaddr->sa_data_min)); in packet_bind_spkt()
3303 name[sizeof(uaddr->sa_data_min)] = 0; in packet_bind_spkt()
3311 struct sock *sk = sock->sk; in packet_bind()
3318 return -EINVAL; in packet_bind()
3319 if (sll->sll_family != AF_PACKET) in packet_bind()
3320 return -EINVAL; in packet_bind()
3322 return packet_do_bind(sk, NULL, sll->sll_ifindex, sll->sll_protocol); in packet_bind()
3343 if (!ns_capable(net->user_ns, CAP_NET_RAW)) in packet_create()
3344 return -EPERM; in packet_create()
3345 if (sock->type != SOCK_DGRAM && sock->type != SOCK_RAW && in packet_create()
3346 sock->type != SOCK_PACKET) in packet_create()
3347 return -ESOCKTNOSUPPORT; in packet_create()
3349 sock->state = SS_UNCONNECTED; in packet_create()
3351 err = -ENOBUFS; in packet_create()
3356 sock->ops = &packet_ops; in packet_create()
3357 if (sock->type == SOCK_PACKET) in packet_create()
3358 sock->ops = &packet_ops_spkt; in packet_create()
3363 init_completion(&po->skb_completion); in packet_create()
3364 sk->sk_family = PF_PACKET; in packet_create()
3365 po->num = proto; in packet_create()
3373 sk->sk_destruct = packet_sock_destruct; in packet_create()
3379 spin_lock_init(&po->bind_lock); in packet_create()
3380 mutex_init(&po->pg_vec_lock); in packet_create()
3381 po->rollover = NULL; in packet_create()
3382 po->prot_hook.func = packet_rcv; in packet_create()
3384 if (sock->type == SOCK_PACKET) in packet_create()
3385 po->prot_hook.func = packet_rcv_spkt; in packet_create()
3387 po->prot_hook.af_packet_priv = sk; in packet_create()
3388 po->prot_hook.af_packet_net = sock_net(sk); in packet_create()
3391 po->prot_hook.type = proto; in packet_create()
3395 mutex_lock(&net->packet.sklist_lock); in packet_create()
3396 sk_add_node_tail_rcu(sk, &net->packet.sklist); in packet_create()
3397 mutex_unlock(&net->packet.sklist_lock); in packet_create()
3416 struct sock *sk = sock->sk; in packet_recvmsg()
3419 int vnet_hdr_len = READ_ONCE(pkt_sk(sk)->vnet_hdr_sz); in packet_recvmsg()
3422 err = -EINVAL; in packet_recvmsg()
3428 if (pkt_sk(sk)->ifindex < 0) in packet_recvmsg()
3429 return -ENODEV; in packet_recvmsg()
3440 * of horrible races and re-entrancy so we can forget about it in packet_recvmsg()
3470 copied = skb->len; in packet_recvmsg()
3473 msg->msg_flags |= MSG_TRUNC; in packet_recvmsg()
3480 if (sock->type != SOCK_PACKET) { in packet_recvmsg()
3481 struct sockaddr_ll *sll = &PACKET_SKB_CB(skb)->sa.ll; in packet_recvmsg()
3484 origlen = PACKET_SKB_CB(skb)->sa.origlen; in packet_recvmsg()
3485 sll->sll_family = AF_PACKET; in packet_recvmsg()
3486 sll->sll_protocol = skb->protocol; in packet_recvmsg()
3491 if (msg->msg_name) { in packet_recvmsg()
3492 const size_t max_len = min(sizeof(skb->cb), in packet_recvmsg()
3499 if (sock->type == SOCK_PACKET) { in packet_recvmsg()
3501 msg->msg_namelen = sizeof(struct sockaddr_pkt); in packet_recvmsg()
3502 copy_len = msg->msg_namelen; in packet_recvmsg()
3504 struct sockaddr_ll *sll = &PACKET_SKB_CB(skb)->sa.ll; in packet_recvmsg()
3506 msg->msg_namelen = sll->sll_halen + in packet_recvmsg()
3508 copy_len = msg->msg_namelen; in packet_recvmsg()
3509 if (msg->msg_namelen < sizeof(struct sockaddr_ll)) { in packet_recvmsg()
3510 memset(msg->msg_name + in packet_recvmsg()
3512 0, sizeof(sll->sll_addr)); in packet_recvmsg()
3513 msg->msg_namelen = sizeof(struct sockaddr_ll); in packet_recvmsg()
3518 msg->msg_namelen = copy_len; in packet_recvmsg()
3520 memcpy(msg->msg_name, &PACKET_SKB_CB(skb)->sa, copy_len); in packet_recvmsg()
3527 if (skb->ip_summed == CHECKSUM_PARTIAL) in packet_recvmsg()
3529 else if (skb->pkt_type != PACKET_OUTGOING && in packet_recvmsg()
3536 aux.tp_snaplen = skb->len; in packet_recvmsg()
3541 aux.tp_vlan_tpid = ntohs(skb->vlan_proto); in packet_recvmsg()
3552 * hides all the races and re-entrancy issues from us. in packet_recvmsg()
3554 err = vnet_hdr_len + ((flags&MSG_TRUNC) ? skb->len : copied); in packet_recvmsg()
3566 struct sock *sk = sock->sk; in packet_getname_spkt()
3569 return -EOPNOTSUPP; in packet_getname_spkt()
3571 uaddr->sa_family = AF_PACKET; in packet_getname_spkt()
3572 memset(uaddr->sa_data, 0, sizeof(uaddr->sa_data_min)); in packet_getname_spkt()
3574 dev = dev_get_by_index_rcu(sock_net(sk), READ_ONCE(pkt_sk(sk)->ifindex)); in packet_getname_spkt()
3576 strscpy(uaddr->sa_data, dev->name, sizeof(uaddr->sa_data_min)); in packet_getname_spkt()
3586 struct sock *sk = sock->sk; in packet_getname()
3592 return -EOPNOTSUPP; in packet_getname()
3594 ifindex = READ_ONCE(po->ifindex); in packet_getname()
3595 sll->sll_family = AF_PACKET; in packet_getname()
3596 sll->sll_ifindex = ifindex; in packet_getname()
3597 sll->sll_protocol = READ_ONCE(po->num); in packet_getname()
3598 sll->sll_pkttype = 0; in packet_getname()
3602 sll->sll_hatype = dev->type; in packet_getname()
3603 sll->sll_halen = dev->addr_len; in packet_getname()
3606 memcpy(((struct sockaddr_storage *)sll)->__data + in packet_getname()
3607 offsetof(struct sockaddr_ll, sll_addr) - in packet_getname()
3609 dev->dev_addr, dev->addr_len); in packet_getname()
3611 sll->sll_hatype = 0; /* Bad: we have no ARPHRD_UNSPEC */ in packet_getname()
3612 sll->sll_halen = 0; in packet_getname()
3616 return offsetof(struct sockaddr_ll, sll_addr) + sll->sll_halen; in packet_getname()
3622 switch (i->type) { in packet_dev_mc()
3624 if (i->alen != dev->addr_len) in packet_dev_mc()
3625 return -EINVAL; in packet_dev_mc()
3627 return dev_mc_add(dev, i->addr); in packet_dev_mc()
3629 return dev_mc_del(dev, i->addr); in packet_dev_mc()
3636 if (i->alen != dev->addr_len) in packet_dev_mc()
3637 return -EINVAL; in packet_dev_mc()
3639 return dev_uc_add(dev, i->addr); in packet_dev_mc()
3641 return dev_uc_del(dev, i->addr); in packet_dev_mc()
3655 if (ml->ifindex == dev->ifindex) { in packet_dev_mclist_delete()
3656 packet_dev_mc(dev, ml, -1); in packet_dev_mclist_delete()
3657 *mlp = ml->next; in packet_dev_mclist_delete()
3660 mlp = &ml->next; in packet_dev_mclist_delete()
3673 err = -ENODEV; in packet_mc_add()
3674 dev = __dev_get_by_index(sock_net(sk), mreq->mr_ifindex); in packet_mc_add()
3678 err = -EINVAL; in packet_mc_add()
3679 if (mreq->mr_alen > dev->addr_len) in packet_mc_add()
3682 err = -ENOBUFS; in packet_mc_add()
3688 for (ml = po->mclist; ml; ml = ml->next) { in packet_mc_add()
3689 if (ml->ifindex == mreq->mr_ifindex && in packet_mc_add()
3690 ml->type == mreq->mr_type && in packet_mc_add()
3691 ml->alen == mreq->mr_alen && in packet_mc_add()
3692 memcmp(ml->addr, mreq->mr_address, ml->alen) == 0) { in packet_mc_add()
3693 ml->count++; in packet_mc_add()
3700 i->type = mreq->mr_type; in packet_mc_add()
3701 i->ifindex = mreq->mr_ifindex; in packet_mc_add()
3702 i->alen = mreq->mr_alen; in packet_mc_add()
3703 memcpy(i->addr, mreq->mr_address, i->alen); in packet_mc_add()
3704 memset(i->addr + i->alen, 0, sizeof(i->addr) - i->alen); in packet_mc_add()
3705 i->count = 1; in packet_mc_add()
3706 i->next = po->mclist; in packet_mc_add()
3707 po->mclist = i; in packet_mc_add()
3710 po->mclist = i->next; in packet_mc_add()
3725 for (mlp = &pkt_sk(sk)->mclist; (ml = *mlp) != NULL; mlp = &ml->next) { in packet_mc_drop()
3726 if (ml->ifindex == mreq->mr_ifindex && in packet_mc_drop()
3727 ml->type == mreq->mr_type && in packet_mc_drop()
3728 ml->alen == mreq->mr_alen && in packet_mc_drop()
3729 memcmp(ml->addr, mreq->mr_address, ml->alen) == 0) { in packet_mc_drop()
3730 if (--ml->count == 0) { in packet_mc_drop()
3732 *mlp = ml->next; in packet_mc_drop()
3733 dev = __dev_get_by_index(sock_net(sk), ml->ifindex); in packet_mc_drop()
3735 packet_dev_mc(dev, ml, -1); in packet_mc_drop()
3750 if (!po->mclist) in packet_flush_mclist()
3754 while ((ml = po->mclist) != NULL) { in packet_flush_mclist()
3757 po->mclist = ml->next; in packet_flush_mclist()
3758 dev = __dev_get_by_index(sock_net(sk), ml->ifindex); in packet_flush_mclist()
3760 packet_dev_mc(dev, ml, -1); in packet_flush_mclist()
3770 struct sock *sk = sock->sk; in packet_setsockopt()
3775 return -ENOPROTOOPT; in packet_setsockopt()
3785 return -EINVAL; in packet_setsockopt()
3789 return -EFAULT; in packet_setsockopt()
3791 return -EINVAL; in packet_setsockopt()
3806 switch (po->tp_version) { in packet_setsockopt()
3817 ret = -EINVAL; in packet_setsockopt()
3820 ret = -EFAULT; in packet_setsockopt()
3833 return -EINVAL; in packet_setsockopt()
3835 return -EFAULT; in packet_setsockopt()
3837 pkt_sk(sk)->copy_thresh = val; in packet_setsockopt()
3845 return -EINVAL; in packet_setsockopt()
3847 return -EFAULT; in packet_setsockopt()
3854 return -EINVAL; in packet_setsockopt()
3857 if (po->rx_ring.pg_vec || po->tx_ring.pg_vec) { in packet_setsockopt()
3858 ret = -EBUSY; in packet_setsockopt()
3860 po->tp_version = val; in packet_setsockopt()
3871 return -EINVAL; in packet_setsockopt()
3873 return -EFAULT; in packet_setsockopt()
3875 return -EINVAL; in packet_setsockopt()
3877 if (po->rx_ring.pg_vec || po->tx_ring.pg_vec) { in packet_setsockopt()
3878 ret = -EBUSY; in packet_setsockopt()
3880 po->tp_reserve = val; in packet_setsockopt()
3891 return -EINVAL; in packet_setsockopt()
3893 return -EFAULT; in packet_setsockopt()
3896 if (po->rx_ring.pg_vec || po->tx_ring.pg_vec) { in packet_setsockopt()
3897 ret = -EBUSY; in packet_setsockopt()
3910 return -EINVAL; in packet_setsockopt()
3912 return -EFAULT; in packet_setsockopt()
3922 return -EINVAL; in packet_setsockopt()
3924 return -EFAULT; in packet_setsockopt()
3934 if (sock->type != SOCK_RAW) in packet_setsockopt()
3935 return -EINVAL; in packet_setsockopt()
3937 return -EINVAL; in packet_setsockopt()
3939 return -EFAULT; in packet_setsockopt()
3944 return -EINVAL; in packet_setsockopt()
3950 if (po->rx_ring.pg_vec || po->tx_ring.pg_vec) { in packet_setsockopt()
3951 ret = -EBUSY; in packet_setsockopt()
3953 WRITE_ONCE(po->vnet_hdr_sz, hdr_len); in packet_setsockopt()
3964 return -EINVAL; in packet_setsockopt()
3966 return -EFAULT; in packet_setsockopt()
3968 WRITE_ONCE(po->tp_tstamp, val); in packet_setsockopt()
3976 return -EINVAL; in packet_setsockopt()
3978 return -EFAULT; in packet_setsockopt()
3985 if (!READ_ONCE(po->fanout)) in packet_setsockopt()
3986 return -EINVAL; in packet_setsockopt()
3995 return -EINVAL; in packet_setsockopt()
3997 return -EFAULT; in packet_setsockopt()
3999 return -EINVAL; in packet_setsockopt()
4001 po->prot_hook.ignore_outgoing = !!val; in packet_setsockopt()
4009 return -EINVAL; in packet_setsockopt()
4011 return -EFAULT; in packet_setsockopt()
4014 if (!po->rx_ring.pg_vec && !po->tx_ring.pg_vec) in packet_setsockopt()
4025 return -EINVAL; in packet_setsockopt()
4027 return -EFAULT; in packet_setsockopt()
4033 return -ENOPROTOOPT; in packet_setsockopt()
4042 struct sock *sk = sock->sk; in packet_getsockopt()
4050 return -ENOPROTOOPT; in packet_getsockopt()
4053 return -EFAULT; in packet_getsockopt()
4056 return -EINVAL; in packet_getsockopt()
4060 spin_lock_bh(&sk->sk_receive_queue.lock); in packet_getsockopt()
4061 memcpy(&st, &po->stats, sizeof(st)); in packet_getsockopt()
4062 memset(&po->stats, 0, sizeof(po->stats)); in packet_getsockopt()
4063 spin_unlock_bh(&sk->sk_receive_queue.lock); in packet_getsockopt()
4064 drops = atomic_xchg(&po->tp_drops, 0); in packet_getsockopt()
4066 if (po->tp_version == TPACKET_V3) { in packet_getsockopt()
4086 val = !!READ_ONCE(po->vnet_hdr_sz); in packet_getsockopt()
4089 val = READ_ONCE(po->vnet_hdr_sz); in packet_getsockopt()
4092 val = po->tp_version; in packet_getsockopt()
4098 return -EINVAL; in packet_getsockopt()
4100 return -EFAULT; in packet_getsockopt()
4112 return -EINVAL; in packet_getsockopt()
4116 val = po->tp_reserve; in packet_getsockopt()
4122 val = READ_ONCE(po->tp_tstamp); in packet_getsockopt()
4125 val = (po->fanout ? in packet_getsockopt()
4126 ((u32)po->fanout->id | in packet_getsockopt()
4127 ((u32)po->fanout->type << 16) | in packet_getsockopt()
4128 ((u32)po->fanout->flags << 24)) : in packet_getsockopt()
4132 val = po->prot_hook.ignore_outgoing; in packet_getsockopt()
4135 if (!po->rollover) in packet_getsockopt()
4136 return -EINVAL; in packet_getsockopt()
4137 rstats.tp_all = atomic_long_read(&po->rollover->num); in packet_getsockopt()
4138 rstats.tp_huge = atomic_long_read(&po->rollover->num_huge); in packet_getsockopt()
4139 rstats.tp_failed = atomic_long_read(&po->rollover->num_failed); in packet_getsockopt()
4150 return -ENOPROTOOPT; in packet_getsockopt()
4156 return -EFAULT; in packet_getsockopt()
4158 return -EFAULT; in packet_getsockopt()
4170 sk_for_each_rcu(sk, &net->packet.sklist) { in packet_notifier()
4175 if (po->mclist) in packet_notifier()
4176 packet_dev_mclist_delete(dev, &po->mclist); in packet_notifier()
4180 if (dev->ifindex == po->ifindex) { in packet_notifier()
4181 spin_lock(&po->bind_lock); in packet_notifier()
4184 sk->sk_err = ENETDOWN; in packet_notifier()
4190 WRITE_ONCE(po->ifindex, -1); in packet_notifier()
4191 netdev_put(po->prot_hook.dev, in packet_notifier()
4192 &po->prot_hook.dev_tracker); in packet_notifier()
4193 po->prot_hook.dev = NULL; in packet_notifier()
4195 spin_unlock(&po->bind_lock); in packet_notifier()
4199 if (dev->ifindex == po->ifindex) { in packet_notifier()
4200 spin_lock(&po->bind_lock); in packet_notifier()
4201 if (po->num) in packet_notifier()
4203 spin_unlock(&po->bind_lock); in packet_notifier()
4216 struct sock *sk = sock->sk; in packet_ioctl()
4230 spin_lock_bh(&sk->sk_receive_queue.lock); in packet_ioctl()
4231 skb = skb_peek(&sk->sk_receive_queue); in packet_ioctl()
4233 amount = skb->len; in packet_ioctl()
4234 spin_unlock_bh(&sk->sk_receive_queue.lock); in packet_ioctl()
4256 return -ENOIOCTLCMD; in packet_ioctl()
4264 struct sock *sk = sock->sk; in packet_poll()
4268 spin_lock_bh(&sk->sk_receive_queue.lock); in packet_poll()
4269 if (po->rx_ring.pg_vec) { in packet_poll()
4270 if (!packet_previous_rx_frame(po, &po->rx_ring, in packet_poll()
4275 spin_unlock_bh(&sk->sk_receive_queue.lock); in packet_poll()
4276 spin_lock_bh(&sk->sk_write_queue.lock); in packet_poll()
4277 if (po->tx_ring.pg_vec) { in packet_poll()
4278 if (packet_current_frame(po, &po->tx_ring, TP_STATUS_AVAILABLE)) in packet_poll()
4281 spin_unlock_bh(&sk->sk_write_queue.lock); in packet_poll()
4292 struct file *file = vma->vm_file; in packet_mm_open()
4293 struct socket *sock = file->private_data; in packet_mm_open()
4294 struct sock *sk = sock->sk; in packet_mm_open()
4297 atomic_long_inc(&pkt_sk(sk)->mapped); in packet_mm_open()
4302 struct file *file = vma->vm_file; in packet_mm_close()
4303 struct socket *sock = file->private_data; in packet_mm_close()
4304 struct sock *sk = sock->sk; in packet_mm_close()
4307 atomic_long_dec(&pkt_sk(sk)->mapped); in packet_mm_close()
4360 unsigned int block_nr = req->tp_block_nr; in alloc_pg_vec()
4395 struct tpacket_req *req = &req_u->req; in packet_set_ring()
4397 rb = tx_ring ? &po->tx_ring : &po->rx_ring; in packet_set_ring()
4398 rb_queue = tx_ring ? &sk->sk_write_queue : &sk->sk_receive_queue; in packet_set_ring()
4400 err = -EBUSY; in packet_set_ring()
4402 if (atomic_long_read(&po->mapped)) in packet_set_ring()
4408 if (req->tp_block_nr) { in packet_set_ring()
4412 err = -EBUSY; in packet_set_ring()
4413 if (unlikely(rb->pg_vec)) in packet_set_ring()
4416 switch (po->tp_version) { in packet_set_ring()
4418 po->tp_hdrlen = TPACKET_HDRLEN; in packet_set_ring()
4421 po->tp_hdrlen = TPACKET2_HDRLEN; in packet_set_ring()
4424 po->tp_hdrlen = TPACKET3_HDRLEN; in packet_set_ring()
4428 err = -EINVAL; in packet_set_ring()
4429 if (unlikely((int)req->tp_block_size <= 0)) in packet_set_ring()
4431 if (unlikely(!PAGE_ALIGNED(req->tp_block_size))) in packet_set_ring()
4433 min_frame_size = po->tp_hdrlen + po->tp_reserve; in packet_set_ring()
4434 if (po->tp_version >= TPACKET_V3 && in packet_set_ring()
4435 req->tp_block_size < in packet_set_ring()
4436 BLK_PLUS_PRIV((u64)req_u->req3.tp_sizeof_priv) + min_frame_size) in packet_set_ring()
4438 if (unlikely(req->tp_frame_size < min_frame_size)) in packet_set_ring()
4440 if (unlikely(req->tp_frame_size & (TPACKET_ALIGNMENT - 1))) in packet_set_ring()
4443 rb->frames_per_block = req->tp_block_size / req->tp_frame_size; in packet_set_ring()
4444 if (unlikely(rb->frames_per_block == 0)) in packet_set_ring()
4446 if (unlikely(rb->frames_per_block > UINT_MAX / req->tp_block_nr)) in packet_set_ring()
4448 if (unlikely((rb->frames_per_block * req->tp_block_nr) != in packet_set_ring()
4449 req->tp_frame_nr)) in packet_set_ring()
4452 err = -ENOMEM; in packet_set_ring()
4453 order = get_order(req->tp_block_size); in packet_set_ring()
4457 switch (po->tp_version) { in packet_set_ring()
4463 struct tpacket_req3 *req3 = &req_u->req3; in packet_set_ring()
4465 if (req3->tp_retire_blk_tov || in packet_set_ring()
4466 req3->tp_sizeof_priv || in packet_set_ring()
4467 req3->tp_feature_req_word) { in packet_set_ring()
4468 err = -EINVAL; in packet_set_ring()
4475 rx_owner_map = bitmap_alloc(req->tp_frame_nr, in packet_set_ring()
4485 err = -EINVAL; in packet_set_ring()
4486 if (unlikely(req->tp_frame_nr)) in packet_set_ring()
4492 spin_lock(&po->bind_lock); in packet_set_ring()
4494 num = po->num; in packet_set_ring()
4496 WRITE_ONCE(po->num, 0); in packet_set_ring()
4499 spin_unlock(&po->bind_lock); in packet_set_ring()
4503 err = -EBUSY; in packet_set_ring()
4504 mutex_lock(&po->pg_vec_lock); in packet_set_ring()
4505 if (closing || atomic_long_read(&po->mapped) == 0) { in packet_set_ring()
4507 spin_lock_bh(&rb_queue->lock); in packet_set_ring()
4508 swap(rb->pg_vec, pg_vec); in packet_set_ring()
4509 if (po->tp_version <= TPACKET_V2) in packet_set_ring()
4510 swap(rb->rx_owner_map, rx_owner_map); in packet_set_ring()
4511 rb->frame_max = (req->tp_frame_nr - 1); in packet_set_ring()
4512 rb->head = 0; in packet_set_ring()
4513 rb->frame_size = req->tp_frame_size; in packet_set_ring()
4514 spin_unlock_bh(&rb_queue->lock); in packet_set_ring()
4516 swap(rb->pg_vec_order, order); in packet_set_ring()
4517 swap(rb->pg_vec_len, req->tp_block_nr); in packet_set_ring()
4519 rb->pg_vec_pages = req->tp_block_size/PAGE_SIZE; in packet_set_ring()
4520 po->prot_hook.func = (po->rx_ring.pg_vec) ? in packet_set_ring()
4523 if (atomic_long_read(&po->mapped)) in packet_set_ring()
4525 atomic_long_read(&po->mapped)); in packet_set_ring()
4527 mutex_unlock(&po->pg_vec_lock); in packet_set_ring()
4529 spin_lock(&po->bind_lock); in packet_set_ring()
4531 WRITE_ONCE(po->num, num); in packet_set_ring()
4534 spin_unlock(&po->bind_lock); in packet_set_ring()
4535 if (pg_vec && (po->tp_version > TPACKET_V2)) { in packet_set_ring()
4536 /* Because we don't support block-based V3 on tx-ring */ in packet_set_ring()
4544 free_pg_vec(pg_vec, order, req->tp_block_nr); in packet_set_ring()
4553 struct sock *sk = sock->sk; in packet_mmap()
4558 int err = -EINVAL; in packet_mmap()
4561 if (vma->vm_pgoff) in packet_mmap()
4562 return -EINVAL; in packet_mmap()
4564 mutex_lock(&po->pg_vec_lock); in packet_mmap()
4567 for (rb = &po->rx_ring; rb <= &po->tx_ring; rb++) { in packet_mmap()
4568 if (rb->pg_vec) { in packet_mmap()
4569 expected_size += rb->pg_vec_len in packet_mmap()
4570 * rb->pg_vec_pages in packet_mmap()
4578 size = vma->vm_end - vma->vm_start; in packet_mmap()
4582 start = vma->vm_start; in packet_mmap()
4583 for (rb = &po->rx_ring; rb <= &po->tx_ring; rb++) { in packet_mmap()
4584 if (rb->pg_vec == NULL) in packet_mmap()
4587 for (i = 0; i < rb->pg_vec_len; i++) { in packet_mmap()
4589 void *kaddr = rb->pg_vec[i].buffer; in packet_mmap()
4592 for (pg_num = 0; pg_num < rb->pg_vec_pages; pg_num++) { in packet_mmap()
4603 atomic_long_inc(&po->mapped); in packet_mmap()
4604 vma->vm_ops = &packet_mmap_ops; in packet_mmap()
4608 mutex_unlock(&po->pg_vec_lock); in packet_mmap()
4670 return seq_hlist_start_head_rcu(&net->packet.sklist, *pos); in packet_seq_start()
4676 return seq_hlist_next_rcu(v, &net->packet.sklist, pos); in packet_seq_next()
4690 IS_ENABLED(CONFIG_64BIT) ? -17 : -9, "sk"); in packet_seq_show()
4696 "%pK %-6d %-4d %04x %-5d %1d %-6u %-6u %-6lu\n", in packet_seq_show()
4698 refcount_read(&s->sk_refcnt), in packet_seq_show()
4699 s->sk_type, in packet_seq_show()
4700 ntohs(READ_ONCE(po->num)), in packet_seq_show()
4701 READ_ONCE(po->ifindex), in packet_seq_show()
4703 atomic_read(&s->sk_rmem_alloc), in packet_seq_show()
4721 mutex_init(&net->packet.sklist_lock); in packet_net_init()
4722 INIT_HLIST_HEAD(&net->packet.sklist); in packet_net_init()
4725 if (!proc_create_net("packet", 0, net->proc_net, &packet_seq_ops, in packet_net_init()
4727 return -ENOMEM; in packet_net_init()
4735 remove_proc_entry("packet", net->proc_net); in packet_net_exit()
4736 WARN_ON_ONCE(!hlist_empty(&net->packet.sklist)); in packet_net_exit()