From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from mail.ozlabs.org (mail.ozlabs.org [IPv6:2404:9400:2221:ea00::3]) by passt.top (Postfix) with ESMTPS id BB7665A0308 for ; Wed, 05 Jun 2024 03:39:10 +0200 (CEST) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gibson.dropbear.id.au; s=202312; t=1717551545; bh=34Ub5FeTIodgekbHCpX8NKz/7uBbW2HjIrUqFUVNqCo=; h=From:To:Cc:Subject:Date:In-Reply-To:References:From; b=XRftxOXNlbtNGozp2XJn8cCz8Lup+Bq0K8KMKCco4Wet92koeqjRJTdi7yWd7vIKr 3rvg6D9p8Z9XQsgBUSAArirXa9a1fDvxaZJJDvH+ocB8tWH9kkQCzC4NYqMpYR2bOX tUqS9pxGkYpBW7eEUa9t02V/POIkEhVq9Q9keHyp05WcRoGi0N3+vqeudnmUzR9x8e 9HCaoo4uc2SdbGCyXp/pBizwdKgUdfNpX8Jd8Pv3qhVuodefGJS2Q8HdPymDCM8A2k i6KsVSp6SmNb3bq9YQlCxTu7r8AYhhDJ19PquhHbH0J1W3vMdvSEt+rD6nlzpF4R9e PMjfM5/OapMog== Received: by gandalf.ozlabs.org (Postfix, from userid 1007) id 4Vv99P71N6z4x2M; Wed, 5 Jun 2024 11:39:05 +1000 (AEST) From: David Gibson To: Stefano Brivio , passt-dev@passt.top Subject: [PATCH 3/4] udp: Rework how we divide queued datagrams between sending methods Date: Wed, 5 Jun 2024 11:39:02 +1000 Message-ID: <20240605013903.3694452-4-david@gibson.dropbear.id.au> X-Mailer: git-send-email 2.45.1 In-Reply-To: <20240605013903.3694452-1-david@gibson.dropbear.id.au> References: <20240605013903.3694452-1-david@gibson.dropbear.id.au> MIME-Version: 1.0 Content-Transfer-Encoding: 8bit Message-ID-Hash: 33AXG6XEGXQKP23OS2TPNM65AV5MCU46 X-Message-ID-Hash: 33AXG6XEGXQKP23OS2TPNM65AV5MCU46 X-MailFrom: dgibson@gandalf.ozlabs.org X-Mailman-Rule-Misses: dmarc-mitigation; no-senders; approved; emergency; loop; banned-address; member-moderation; nonmember-moderation; administrivia; implicit-dest; max-recipients; max-size; news-moderation; no-subject; digests; suspicious-header CC: David Gibson X-Mailman-Version: 3.3.8 Precedence: list List-Id: Development discussion and patches for passt Archived-At: Archived-At: List-Archive: List-Archive: List-Help: List-Owner: List-Post: List-Subscribe: List-Unsubscribe: udp_sock_handler() takes a number of datagrams from sockets that depending on their addresses could be forwarded either to the L2 interface ("tap") or to another socket ("spliced"). In the latter case we can also only send packets together if they have the same source port, and therefore are sent via the same socket. To reduce the total number of system calls we gather contiguous batches of datagrams with the same destination interface and socket where applicable. The determination of what the target is is made by udp_mmh_splice_port(). It returns the source port for splice packets and -1 for "tap" packets. We find batches by looking ahead in our queue until we find a datagram whose "splicefrom" port doesn't match the first in our current batch. udp_mmh_splice_port() is moderately expensive, since it must examine IPv6 addresses. But unfortunately we can call it twice on the same datagram: once as the (last + 1) entry in one batch (showing that it's not in that match, then again as the first entry in the next batch. Avoid this by keeping track of the "splice port" in the metadata structure, and filling it in one entry ahead of the one we're currently considering. This is a bit subtle, but not that hard. It will also generalise better when we have more complex possibilities based on the flow table. Signed-off-by: David Gibson --- udp.c | 147 ++++++++++++++++++++++++++++++++++------------------------ 1 file changed, 86 insertions(+), 61 deletions(-) diff --git a/udp.c b/udp.c index 7487d2b2..757c10ab 100644 --- a/udp.c +++ b/udp.c @@ -198,6 +198,7 @@ static struct ethhdr udp6_eth_hdr; * @ip4h: Pre-filled IPv4 header (except for tot_len and saddr) * @taph: Tap backend specific header * @s_in: Source socket address, filled in by recvmmsg() + * @splicesrc: Source port for splicing, or -1 if not spliceable */ static struct udp_meta_t { struct ipv6hdr ip6h; @@ -205,6 +206,7 @@ static struct udp_meta_t { struct tap_hdr taph; union sockaddr_inany s_in; + int splicesrc; } #ifdef __AVX2__ __attribute__ ((aligned(32))) @@ -492,28 +494,32 @@ static int udp_mmh_splice_port(union udp_epoll_ref uref, } /** - * udp_splice_sendfrom() - Send datagrams from given port to given port + * udp_splice_send() - Send datagrams from socket to socket * @c: Execution context * @start: Index of first datagram in udp[46]_l2_buf - * @n: Number of datagrams to send - * @src: Datagrams will be sent from this port (on origin side) - * @dst: Datagrams will be send to this port (on destination side) - * @from_pif: pif from which the packet originated - * @v6: Send as IPv6? - * @allow_new: If true create sending socket if needed, if false discard - * if no sending socket is available + * @n: Total number of datagrams in udp[46]_l2_buf pool + * @dst: Datagrams will be sent to this port (on destination side) + * @uref: UDP epoll reference for origin socket * @now: Timestamp + * + * This consumes as many frames as are sendable via a single socket. It + * requires that udp_meta[@start].splicesrc is initialised, and will initialise + * udp_meta[].splicesrc for each frame it consumes *and one more* (if present). + * + * Return: Number of frames sent */ -static void udp_splice_sendfrom(const struct ctx *c, unsigned start, unsigned n, - in_port_t src, in_port_t dst, uint8_t from_pif, - bool v6, bool allow_new, +static unsigned udp_splice_send(const struct ctx *c, size_t start, size_t n, + in_port_t dst, union udp_epoll_ref uref, const struct timespec *now) { + in_port_t src = udp_meta[start].splicesrc; struct mmsghdr *mmh_recv, *mmh_send; - unsigned int i; + unsigned int i = start; int s; - if (v6) { + ASSERT(udp_meta[start].splicesrc >= 0); + + if (uref.v6) { mmh_recv = udp6_l2_mh_sock; mmh_send = udp6_mh_splice; } else { @@ -521,40 +527,48 @@ static void udp_splice_sendfrom(const struct ctx *c, unsigned start, unsigned n, mmh_send = udp4_mh_splice; } - if (from_pif == PIF_SPLICE) { + do { + mmh_send[i].msg_hdr.msg_iov->iov_len = mmh_recv[i].msg_len; + + if (++i >= n) + break; + + udp_meta[i].splicesrc = udp_mmh_splice_port(uref, &mmh_recv[i]); + } while (udp_meta[i].splicesrc == src); + + if (uref.pif == PIF_SPLICE) { src += c->udp.fwd_in.rdelta[src]; - s = udp_splice_init[v6][src].sock; - if (s < 0 && allow_new) - s = udp_splice_new(c, v6, src, false); + s = udp_splice_init[uref.v6][src].sock; + if (s < 0 && uref.orig) + s = udp_splice_new(c, uref.v6, src, false); if (s < 0) - return; + goto out; - udp_splice_ns[v6][dst].ts = now->tv_sec; - udp_splice_init[v6][src].ts = now->tv_sec; + udp_splice_ns[uref.v6][dst].ts = now->tv_sec; + udp_splice_init[uref.v6][src].ts = now->tv_sec; } else { - ASSERT(from_pif == PIF_HOST); + ASSERT(uref.pif == PIF_HOST); src += c->udp.fwd_out.rdelta[src]; - s = udp_splice_ns[v6][src].sock; - if (s < 0 && allow_new) { + s = udp_splice_ns[uref.v6][src].sock; + if (s < 0 && uref.orig) { struct udp_splice_new_ns_arg arg = { - c, v6, src, -1, + c, uref.v6, src, -1, }; NS_CALL(udp_splice_new_ns, &arg); s = arg.s; } if (s < 0) - return; + goto out; - udp_splice_init[v6][dst].ts = now->tv_sec; - udp_splice_ns[v6][src].ts = now->tv_sec; + udp_splice_init[uref.v6][dst].ts = now->tv_sec; + udp_splice_ns[uref.v6][src].ts = now->tv_sec; } - for (i = start; i < start + n; i++) - mmh_send[i].msg_hdr.msg_iov->iov_len = mmh_recv[i].msg_len; - - sendmmsg(s, mmh_send + start, n, MSG_NOSIGNAL); + sendmmsg(s, mmh_send + start, i - start, MSG_NOSIGNAL); +out: + return i - start; } /** @@ -687,31 +701,41 @@ static size_t udp_update_hdr6(const struct ctx *c, * udp_tap_send() - Prepare UDP datagrams and send to tap interface * @c: Execution context * @start: Index of first datagram in udp[46]_l2_buf pool - * @n: Number of datagrams to send - * @dstport: Destination port number - * @v6: True if using IPv6 + * @n: Total number of datagrams in udp[46]_l2_buf pool + * @dstport: Destination port number on destination side + * @uref: UDP epoll reference for origin socket * @now: Current timestamp * - * Return: size of tap frame with headers + * This consumes as many frames as are sendable via tap. It requires that + * udp_meta[@start].splicesrc is initialised, and will initialise + * udp_meta[].splicesrc for each frame it consumes *and one more* (if present). + * + * Return: Number of frames sent via tap */ -static void udp_tap_send(const struct ctx *c, - unsigned int start, unsigned int n, - in_port_t dstport, bool v6, const struct timespec *now) +static unsigned udp_tap_send(const struct ctx *c, size_t start, size_t n, + in_port_t dstport, union udp_epoll_ref uref, + const struct timespec *now) { struct iovec (*tap_iov)[UDP_NUM_IOVS]; - size_t i; + struct mmsghdr *mmh_recv; + size_t i = start; - if (v6) + ASSERT(udp_meta[start].splicesrc == -1); + + if (uref.v6) { tap_iov = udp6_l2_iov_tap; - else + mmh_recv = udp6_l2_mh_sock; + } else { + mmh_recv = udp4_l2_mh_sock; tap_iov = udp4_l2_iov_tap; + } - for (i = start; i < start + n; i++) { + do { struct udp_payload_t *bp = &udp_payload[i]; struct udp_meta_t *bm = &udp_meta[i]; size_t l4len; - if (v6) { + if (uref.v6) { l4len = udp_update_hdr6(c, bm, bp, dstport, udp6_l2_mh_sock[i].msg_len, now); } else { @@ -719,9 +743,15 @@ static void udp_tap_send(const struct ctx *c, udp4_l2_mh_sock[i].msg_len, now); } tap_iov[i][UDP_IOV_PAYLOAD].iov_len = l4len; - } - tap_send_frames(c, &tap_iov[start][0], UDP_NUM_IOVS, n); + if (++i >= n) + break; + + udp_meta[i].splicesrc = udp_mmh_splice_port(uref, &mmh_recv[i]); + } while (udp_meta[i].splicesrc == -1); + + tap_send_frames(c, &tap_iov[start][0], UDP_NUM_IOVS, i - start); + return i - start; } /** @@ -770,24 +800,19 @@ void udp_sock_handler(const struct ctx *c, union epoll_ref ref, uint32_t events, if (n <= 0) return; + /* We divide things into batches based on how we need to send them, + * determined by udp_meta[i].splicesrc. To avoid either two passes + * through the array, or recalculating splicesrc for a single entry, we + * have to populate it one entry *ahead* of the loop counter (if + * present). So we fill in entry 0 before the loop, then udp_*_send() + * populate one entry past where they consume. + */ + udp_meta[0].splicesrc = udp_mmh_splice_port(ref.udp, mmh_recv); for (i = 0; i < n; i += m) { - int splicefrom = -1; - - splicefrom = udp_mmh_splice_port(ref.udp, mmh_recv + i); - - for (m = 1; i + m < n; m++) { - int p; - - p = udp_mmh_splice_port(ref.udp, mmh_recv + i + m); - if (p != splicefrom) - break; - } - - if (splicefrom >= 0) - udp_splice_sendfrom(c, i, m, splicefrom, dstport, - ref.udp.pif, v6, ref.udp.orig, now); + if (udp_meta[i].splicesrc >= 0) + m = udp_splice_send(c, i, n, dstport, ref.udp, now); else - udp_tap_send(c, i, m, dstport, v6, now); + m = udp_tap_send(c, i, n, dstport, ref.udp, now); } } -- 2.45.1