From f249be4d2c275fe2b98e389f471af75f758e5a59 Mon Sep 17 00:00:00 2001 From: Jason Wang Date: Sat, 24 Feb 2018 11:32:24 +0800 Subject: Revert "tuntap: add missing xdp flush" This reverts commit 762c330d670e3d4b795cf7a8d761866fdd1eef49. The reason is we try to batch packets for devmap which causes calling xdp_do_flush() in the process context. Simply disabling preemption may not work since process may move among processors which lead xdp_do_flush() to miss some flushes on some processors. So simply revert the patch, a follow-up patch will add the xdp flush correctly. Reported-by: Christoffer Dall Fixes: 762c330d670e ("tuntap: add missing xdp flush") Signed-off-by: Jason Wang Signed-off-by: David S. Miller --- drivers/net/tun.c | 15 --------------- 1 file changed, 15 deletions(-) (limited to 'drivers/net/tun.c') diff --git a/drivers/net/tun.c b/drivers/net/tun.c index b52258c..2823a4a 100644 --- a/drivers/net/tun.c +++ b/drivers/net/tun.c @@ -181,7 +181,6 @@ struct tun_file { struct tun_struct *detached; struct ptr_ring tx_ring; struct xdp_rxq_info xdp_rxq; - int xdp_pending_pkts; }; struct tun_flow_entry { @@ -1662,7 +1661,6 @@ static struct sk_buff *tun_build_skb(struct tun_struct *tun, case XDP_REDIRECT: get_page(alloc_frag->page); alloc_frag->offset += buflen; - ++tfile->xdp_pending_pkts; err = xdp_do_redirect(tun->dev, &xdp, xdp_prog); if (err) goto err_redirect; @@ -1984,11 +1982,6 @@ static ssize_t tun_chr_write_iter(struct kiocb *iocb, struct iov_iter *from) result = tun_get_user(tun, tfile, NULL, from, file->f_flags & O_NONBLOCK, false); - if (tfile->xdp_pending_pkts) { - tfile->xdp_pending_pkts = 0; - xdp_do_flush_map(); - } - tun_put(tun); return result; } @@ -2325,13 +2318,6 @@ static int tun_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len) ret = tun_get_user(tun, tfile, m->msg_control, &m->msg_iter, m->msg_flags & MSG_DONTWAIT, m->msg_flags & MSG_MORE); - - if (tfile->xdp_pending_pkts >= NAPI_POLL_WEIGHT || - !(m->msg_flags & MSG_MORE)) { - tfile->xdp_pending_pkts = 0; - xdp_do_flush_map(); - } - tun_put(tun); return ret; } @@ -3163,7 +3149,6 @@ static int tun_chr_open(struct inode *inode, struct file * file) sock_set_flag(&tfile->sk, SOCK_ZEROCOPY); memset(&tfile->tx_ring, 0, sizeof(tfile->tx_ring)); - tfile->xdp_pending_pkts = 0; return 0; } -- cgit v1.1 From 23e43f07f896f8578318cfcc9466f1e8b8ab21b6 Mon Sep 17 00:00:00 2001 From: Jason Wang Date: Sat, 24 Feb 2018 11:32:25 +0800 Subject: tuntap: disable preemption during XDP processing Except for tuntap, all other drivers' XDP was implemented at NAPI poll() routine in a bh. This guarantees all XDP operation were done at the same CPU which is required by e.g BFP_MAP_TYPE_PERCPU_ARRAY. But for tuntap, we do it in process context and we try to protect XDP processing by RCU reader lock. This is insufficient since CONFIG_PREEMPT_RCU can preempt the RCU reader critical section which breaks the assumption that all XDP were processed in the same CPU. Fixing this by simply disabling preemption during XDP processing. Fixes: 761876c857cb ("tap: XDP support") Signed-off-by: Jason Wang Signed-off-by: David S. Miller --- drivers/net/tun.c | 6 ++++++ 1 file changed, 6 insertions(+) (limited to 'drivers/net/tun.c') diff --git a/drivers/net/tun.c b/drivers/net/tun.c index 2823a4a..63d39fe6 100644 --- a/drivers/net/tun.c +++ b/drivers/net/tun.c @@ -1642,6 +1642,7 @@ static struct sk_buff *tun_build_skb(struct tun_struct *tun, else *skb_xdp = 0; + preempt_disable(); rcu_read_lock(); xdp_prog = rcu_dereference(tun->xdp_prog); if (xdp_prog && !*skb_xdp) { @@ -1665,6 +1666,7 @@ static struct sk_buff *tun_build_skb(struct tun_struct *tun, if (err) goto err_redirect; rcu_read_unlock(); + preempt_enable(); return NULL; case XDP_TX: xdp_xmit = true; @@ -1686,6 +1688,7 @@ static struct sk_buff *tun_build_skb(struct tun_struct *tun, skb = build_skb(buf, buflen); if (!skb) { rcu_read_unlock(); + preempt_enable(); return ERR_PTR(-ENOMEM); } @@ -1698,10 +1701,12 @@ static struct sk_buff *tun_build_skb(struct tun_struct *tun, skb->dev = tun->dev; generic_xdp_tx(skb, xdp_prog); rcu_read_unlock(); + preempt_enable(); return NULL; } rcu_read_unlock(); + preempt_enable(); return skb; @@ -1709,6 +1714,7 @@ err_redirect: put_page(alloc_frag->page); err_xdp: rcu_read_unlock(); + preempt_enable(); this_cpu_inc(tun->pcpu_stats->rx_dropped); return NULL; } -- cgit v1.1 From 1bb4f2e868a2891ab8bc668b8173d6ccb8c4ce6f Mon Sep 17 00:00:00 2001 From: Jason Wang Date: Sat, 24 Feb 2018 11:32:26 +0800 Subject: tuntap: correctly add the missing XDP flush We don't flush batched XDP packets through xdp_do_flush_map(), this will cause packets stall at TX queue. Consider we don't do XDP on NAPI poll(), the only possible fix is to call xdp_do_flush_map() immediately after xdp_do_redirect(). Note, this in fact won't try to batch packets through devmap, we could address in the future. Reported-by: Christoffer Dall Fixes: 761876c857cb ("tap: XDP support") Signed-off-by: Jason Wang Signed-off-by: David S. Miller --- drivers/net/tun.c | 1 + 1 file changed, 1 insertion(+) (limited to 'drivers/net/tun.c') diff --git a/drivers/net/tun.c b/drivers/net/tun.c index 63d39fe6..7433bb2 100644 --- a/drivers/net/tun.c +++ b/drivers/net/tun.c @@ -1663,6 +1663,7 @@ static struct sk_buff *tun_build_skb(struct tun_struct *tun, get_page(alloc_frag->page); alloc_frag->offset += buflen; err = xdp_do_redirect(tun->dev, &xdp, xdp_prog); + xdp_do_flush_map(); if (err) goto err_redirect; rcu_read_unlock(); -- cgit v1.1