Message ID | 49790C02.90800@cosmosbay.com |
---|---|
State | Superseded, archived |
Delegated to: | David Miller |
Headers | show |
At Fri, 23 Jan 2009 01:14:58 +0100, Eric Dumazet wrote: > Could you try following patch ? > > Thank you > > [PATCH] udp: optimize bind(0) if many ports are in use > > commit 9088c5609584684149f3fb5b065aa7f18dcb03ff > (udp: Improve port randomization) introduced a regression for UDP bind() syscall > to null port (getting a random port) in case lot of ports are already in use. > > This is because we do about 28000 scans of very long chains (220 sockets per chain), > with many spin_lock_bh()/spin_unlock_bh() calls. > > Fix this using a bitmap (64 bytes for current value of UDP_HTABLE_SIZE) > so that we scan chains at most once. > > Instead of 250 ms per bind() call, we get after patch a time of 2.9 ms It's much better, thanks. FPS in glxgears now drops down only 2x harder if compare with 2.6.28. However, this again kills randomness :) Now number distribution is k*x^2 with x-axis zero in the (high - low) / 2. Try this program, it produces input file for Octave + Gnuplot. #include <stdio.h> #include <errno.h> #include <string.h> #include <sys/types.h> #include <sys/socket.h> #include <netinet/in.h> #define PORTS 65536 int main() { int s, err, i, j; char buf[256]; struct sockaddr_in sa; int optval = 1, port; unsigned int p[PORTS] = { 0 }; for (i = 0; i < PORTS * 100; ++i) { s = socket(PF_INET, SOCK_DGRAM, IPPROTO_UDP); memset(&sa, 0, sizeof(sa)); sa.sin_addr.s_addr = htonl(INADDR_ANY); sa.sin_family = AF_INET; sa.sin_port = 0; setsockopt(s, IPPROTO_UDP, SO_REUSEADDR, &optval, sizeof(optval)); err = bind(s, (const struct sockaddr*)&sa, sizeof(sa)); getsockname(s, (struct sockaddr*)&sa, &j); port = ntohs(sa.sin_port); p[port]++; close(s); } printf("x = 32766:1:65535;\ny = [-100; "); for (i = 32767; i < PORTS; i++) printf("%d%s", p[i], (i + 1 < PORTS ? "; " : "")); printf("];\nplot(x,y,'.');pause;"); } I was thinking about bitmap also, but in a bit different approach. It is also uses bias (delta) value instead of exact port number. When we get next random port value (from rng or in the next iteration), we can calculate byte offset in that bitmap: A B C D 76543210 76543210 7654321 076543210 11110111 11011110 1001111 011110111 ^ We here land in the byte B in the marked bit position, but it is already busy. If there're any free bits in this byte B, we can stop further iterations and use any free bit. I don't think it can kill randomness too much, because average bias will be small. May be it only needs some more complicated logic for searching free bit in the byte, because it's not good to do scanning always from the beginning. -- wbr, Vitaly -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Vitaly Mayatskikh a écrit : > At Fri, 23 Jan 2009 01:14:58 +0100, Eric Dumazet wrote: > >> Could you try following patch ? >> >> Thank you >> >> [PATCH] udp: optimize bind(0) if many ports are in use >> >> commit 9088c5609584684149f3fb5b065aa7f18dcb03ff >> (udp: Improve port randomization) introduced a regression for UDP bind() syscall >> to null port (getting a random port) in case lot of ports are already in use. >> >> This is because we do about 28000 scans of very long chains (220 sockets per chain), >> with many spin_lock_bh()/spin_unlock_bh() calls. >> >> Fix this using a bitmap (64 bytes for current value of UDP_HTABLE_SIZE) >> so that we scan chains at most once. >> >> Instead of 250 ms per bind() call, we get after patch a time of 2.9 ms > > It's much better, thanks. FPS in glxgears now drops down only 2x > harder if compare with 2.6.28. However, this again kills randomness :) > Now number distribution is k*x^2 with x-axis zero in the (high - low) > / 2. Try this program, it produces input file for Octave + Gnuplot. > > #include <stdio.h> > #include <errno.h> > #include <string.h> > #include <sys/types.h> > #include <sys/socket.h> > #include <netinet/in.h> > > #define PORTS 65536 > > int main() > { > int s, err, i, j; > char buf[256]; > struct sockaddr_in sa; > int optval = 1, port; > unsigned int p[PORTS] = { 0 }; > > for (i = 0; i < PORTS * 100; ++i) { > s = socket(PF_INET, SOCK_DGRAM, IPPROTO_UDP); > memset(&sa, 0, sizeof(sa)); > sa.sin_addr.s_addr = htonl(INADDR_ANY); > sa.sin_family = AF_INET; > sa.sin_port = 0; > setsockopt(s, IPPROTO_UDP, SO_REUSEADDR, &optval, sizeof(optval)); > err = bind(s, (const struct sockaddr*)&sa, sizeof(sa)); > getsockname(s, (struct sockaddr*)&sa, &j); > port = ntohs(sa.sin_port); > p[port]++; > close(s); > } > printf("x = 32766:1:65535;\ny = [-100; "); > for (i = 32767; i < PORTS; i++) > printf("%d%s", p[i], (i + 1 < PORTS ? "; " : "")); > printf("];\nplot(x,y,'.');pause;"); > } > > I was thinking about bitmap also, but in a bit different approach. It > is also uses bias (delta) value instead of exact port number. When we > get next random port value (from rng or in the next iteration), we can > calculate byte offset in that bitmap: > > A B C D > 76543210 76543210 7654321 076543210 > 11110111 11011110 1001111 011110111 > ^ > > We here land in the byte B in the marked bit position, but it is > already busy. If there're any free bits in this byte B, we can stop > further iterations and use any free bit. I don't think it can kill > randomness too much, because average bias will be small. May be it > only needs some more complicated logic for searching free bit in the > byte, because it's not good to do scanning always from the beginning. > Interesting... Please note I dont search in the bitmap from its begining, but from a random point. Maybe we should study lib/random32.c and discover it has said distribution :) Since my algo uses net_random() (random32() to get 32 bits number, that we split in two "16 bits numbers" (bias & rand). One (bias) to select the starting chain and starting slot in chain (so it really should be random) first = bias + 0 (slotn=0) One (rand, forced to be odd) to select next slot in chain in case current slot is already in use. I feel this is the problem, because when we hit a slot outside of ip_local_port_range, it seems we escape from this range with the distribution you got. maybe we should get rand depending on ip_local_port_range -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c index cf5ab05..adbdbd8 100644 --- a/net/ipv4/udp.c +++ b/net/ipv4/udp.c @@ -120,8 +120,11 @@ EXPORT_SYMBOL(sysctl_udp_wmem_min); atomic_t udp_memory_allocated; EXPORT_SYMBOL(udp_memory_allocated); +#define PORTS_PER_CHAIN (65536 / UDP_HTABLE_SIZE) + static int udp_lib_lport_inuse(struct net *net, __u16 num, const struct udp_hslot *hslot, + unsigned long *bitmap, struct sock *sk, int (*saddr_comp)(const struct sock *sk1, const struct sock *sk2)) @@ -132,12 +135,16 @@ static int udp_lib_lport_inuse(struct net *net, __u16 num, sk_nulls_for_each(sk2, node, &hslot->head) if (net_eq(sock_net(sk2), net) && sk2 != sk && - sk2->sk_hash == num && + (bitmap || sk2->sk_hash == num) && (!sk2->sk_reuse || !sk->sk_reuse) && (!sk2->sk_bound_dev_if || !sk->sk_bound_dev_if || sk2->sk_bound_dev_if == sk->sk_bound_dev_if) && - (*saddr_comp)(sk, sk2)) - return 1; + (*saddr_comp)(sk, sk2)) { + if (bitmap) + __set_bit(sk2->sk_hash / UDP_HTABLE_SIZE, bitmap); + else + return 1; + } return 0; } @@ -158,34 +165,44 @@ int udp_lib_get_port(struct sock *sk, unsigned short snum, struct net *net = sock_net(sk); if (!snum) { - int low, high, remaining; - unsigned rand; + int low, high; + unsigned rand, slotn, bias; unsigned short first; + DECLARE_BITMAP(bitmap, PORTS_PER_CHAIN); inet_get_local_port_range(&low, &high); - remaining = (high - low) + 1; rand = net_random(); - snum = first = rand % remaining + low; - rand |= 1; - for (;;) { - hslot = &udptable->hash[udp_hashfn(net, snum)]; + bias = rand; + rand = ((rand >> 16) | 1) * UDP_HTABLE_SIZE; + for (slotn = 0; slotn < UDP_HTABLE_SIZE; slotn++) { + first = slotn + bias; + hslot = &udptable->hash[udp_hashfn(net, first)]; + bitmap_zero(bitmap, PORTS_PER_CHAIN); spin_lock_bh(&hslot->lock); - if (!udp_lib_lport_inuse(net, snum, hslot, sk, saddr_comp)) - break; - spin_unlock_bh(&hslot->lock); + udp_lib_lport_inuse(net, snum, hslot, bitmap, sk, saddr_comp); + + snum = first; + /* + * PORTS_PER_CHAIN loops, because snum is unsigned short + * and we add an odd multiple of UDP_HTABLE_SIZE + */ do { - snum = snum + rand; - } while (snum < low || snum > high); - if (snum == first) - goto fail; + if (low <= snum && snum <= high && + !test_bit(snum / UDP_HTABLE_SIZE, bitmap)) + goto found; + snum += rand; + } while (snum != first); + spin_unlock_bh(&hslot->lock); } + goto fail; } else { hslot = &udptable->hash[udp_hashfn(net, snum)]; spin_lock_bh(&hslot->lock); - if (udp_lib_lport_inuse(net, snum, hslot, sk, saddr_comp)) + if (udp_lib_lport_inuse(net, snum, hslot, NULL, sk, saddr_comp)) goto fail_unlock; } +found: inet_sk(sk)->num = snum; sk->sk_hash = snum; if (sk_unhashed(sk)) {