I decided that I should do some microbenchmarking to see
if it really makes a difference or not. Since, we know that
lock contention is not an issue (rwlock with rare updates), I
thought I should concentrate on the cache line bouncing of
the per-bucket rwlock. I figured that if we have lots of
hits in the route cache, the corresponding rwlock cache lines
are going to bounce. So I wrote a test that
uses about 8 processes to send a large number of randomly sized packets to
a single destination. For 1 to 8 CPUs I used the test script
to send a fixed number of packets to a single destination address.
The results show that time needed for lookup continuously increases
for 2.5.3 wherease for rt_rcu-2.5.3, it remains constant.
http://lse.sourceforge.net/locking/rcu/rtcache/rttest/rttest.png
The results are in -
http://lse.sourceforge.net/locking/rcu/rtcache/rttest/
The measurements were done in an 8way PIII xeon with 1MB L2 cache
and 6GB memory.
Comments/suggestions are welcome.
Thanks
-- Dipankar Sarma <dipankar@in.ibm.com> http://lse.sourceforge.net Linux Technology Center, IBM Software Lab, Bangalore, India.
diff -urN linux-2.5.3/include/net/dst.h linux-2.5.3-rt_rcu/include/net/dst.h --- linux-2.5.3/include/net/dst.h Wed Jan 30 11:11:52 2002 +++ linux-2.5.3-rt_rcu/include/net/dst.h Fri Feb 8 11:45:15 2002 @@ -9,6 +9,7 @@ #define _NET_DST_H #include <linux/config.h> +#include <linux/rcupdate.h> #include <net/neighbour.h> /* @@ -62,6 +63,7 @@ #endif struct dst_ops *ops; + struct rcu_head rcu_head; char info[0]; }; diff -urN linux-2.5.3/net/ipv4/route.c linux-2.5.3-rt_rcu/net/ipv4/route.c --- linux-2.5.3/net/ipv4/route.c Wed Jan 16 00:26:35 2002 +++ linux-2.5.3-rt_rcu/net/ipv4/route.c Fri Feb 8 11:37:22 2002 @@ -85,6 +85,7 @@ #include <linux/mroute.h> #include <linux/netfilter_ipv4.h> #include <linux/random.h> +#include <linux/rcupdate.h> #include <net/protocol.h> #include <net/ip.h> #include <net/route.h> @@ -188,7 +189,7 @@ struct rt_hash_bucket { struct rtable *chain; - rwlock_t lock; + spinlock_t lock; } __attribute__((__aligned__(8))); static struct rt_hash_bucket *rt_hash_table; @@ -227,7 +228,6 @@ } for (i = rt_hash_mask; i >= 0; i--) { - read_lock_bh(&rt_hash_table[i].lock); for (r = rt_hash_table[i].chain; r; r = r->u.rt_next) { /* * Spin through entries until we are ready @@ -238,6 +238,8 @@ len = 0; continue; } + /* read_barrier_depends() here */ + rmb(); sprintf(temp, "%s\t%08lX\t%08lX\t%8X\t%d\t%u\t%d\t" "%08lX\t%d\t%u\t%u\t%02X\t%d\t%1d\t%08X", r->u.dst.dev ? r->u.dst.dev->name : "*", @@ -262,11 +264,9 @@ sprintf(buffer + len, "%-127s\n", temp); len += 128; if (pos >= offset+length) { - read_unlock_bh(&rt_hash_table[i].lock); goto done; } } - read_unlock_bh(&rt_hash_table[i].lock); } done: @@ -314,13 +314,13 @@ static __inline__ void rt_free(struct rtable *rt) { - dst_free(&rt->u.dst); + call_rcu(&rt->u.dst.rcu_head, (void (*)(void *))dst_free, &rt->u.dst); } static __inline__ void rt_drop(struct rtable *rt) { ip_rt_put(rt); - dst_free(&rt->u.dst); + call_rcu(&rt->u.dst.rcu_head, (void (*)(void *))dst_free, &rt->u.dst); } static __inline__ int rt_fast_clean(struct rtable *rth) @@ -373,7 +373,7 @@ i = (i + 1) & rt_hash_mask; rthp = &rt_hash_table[i].chain; - write_lock(&rt_hash_table[i].lock); + spin_lock(&rt_hash_table[i].lock); while ((rth = *rthp) != NULL) { if (rth->u.dst.expires) { /* Entry is expired even if it is in use */ @@ -392,7 +392,7 @@ *rthp = rth->u.rt_next; rt_free(rth); } - write_unlock(&rt_hash_table[i].lock); + spin_unlock(&rt_hash_table[i].lock); /* Fallback loop breaker. */ if ((jiffies - now) > 0) @@ -415,11 +415,11 @@ rt_deadline = 0; for (i = rt_hash_mask; i >= 0; i--) { - write_lock_bh(&rt_hash_table[i].lock); + spin_lock_bh(&rt_hash_table[i].lock); rth = rt_hash_table[i].chain; if (rth) rt_hash_table[i].chain = NULL; - write_unlock_bh(&rt_hash_table[i].lock); + spin_unlock_bh(&rt_hash_table[i].lock); for (; rth; rth = next) { next = rth->u.rt_next; @@ -538,7 +538,7 @@ k = (k + 1) & rt_hash_mask; rthp = &rt_hash_table[k].chain; - write_lock_bh(&rt_hash_table[k].lock); + spin_lock_bh(&rt_hash_table[k].lock); while ((rth = *rthp) != NULL) { if (!rt_may_expire(rth, tmo, expire)) { tmo >>= 1; @@ -549,7 +549,7 @@ rt_free(rth); goal--; } - write_unlock_bh(&rt_hash_table[k].lock); + spin_unlock_bh(&rt_hash_table[k].lock); if (goal <= 0) break; } @@ -607,18 +607,20 @@ restart: rthp = &rt_hash_table[hash].chain; - write_lock_bh(&rt_hash_table[hash].lock); + spin_lock_bh(&rt_hash_table[hash].lock); while ((rth = *rthp) != NULL) { if (memcmp(&rth->key, &rt->key, sizeof(rt->key)) == 0) { /* Put it first */ *rthp = rth->u.rt_next; + wmb(); rth->u.rt_next = rt_hash_table[hash].chain; + wmb(); rt_hash_table[hash].chain = rth; rth->u.dst.__use++; dst_hold(&rth->u.dst); rth->u.dst.lastuse = now; - write_unlock_bh(&rt_hash_table[hash].lock); + spin_unlock_bh(&rt_hash_table[hash].lock); rt_drop(rt); *rp = rth; @@ -634,7 +636,7 @@ if (rt->rt_type == RTN_UNICAST || rt->key.iif == 0) { int err = arp_bind_neighbour(&rt->u.dst); if (err) { - write_unlock_bh(&rt_hash_table[hash].lock); + spin_unlock_bh(&rt_hash_table[hash].lock); if (err != -ENOBUFS) { rt_drop(rt); @@ -675,7 +677,7 @@ } #endif rt_hash_table[hash].chain = rt; - write_unlock_bh(&rt_hash_table[hash].lock); + spin_unlock_bh(&rt_hash_table[hash].lock); *rp = rt; return 0; } @@ -742,7 +744,7 @@ { struct rtable **rthp; - write_lock_bh(&rt_hash_table[hash].lock); + spin_lock_bh(&rt_hash_table[hash].lock); ip_rt_put(rt); for (rthp = &rt_hash_table[hash].chain; *rthp; rthp = &(*rthp)->u.rt_next) @@ -751,7 +753,7 @@ rt_free(rt); break; } - write_unlock_bh(&rt_hash_table[hash].lock); + spin_unlock_bh(&rt_hash_table[hash].lock); } void ip_rt_redirect(u32 old_gw, u32 daddr, u32 new_gw, @@ -790,10 +792,10 @@ rthp=&rt_hash_table[hash].chain; - read_lock(&rt_hash_table[hash].lock); while ((rth = *rthp) != NULL) { struct rtable *rt; - + /* read_barrier_depends() here */ + rmb(); if (rth->key.dst != daddr || rth->key.src != skeys[i] || rth->key.tos != tos || @@ -811,7 +813,6 @@ break; dst_clone(&rth->u.dst); - read_unlock(&rt_hash_table[hash].lock); rt = dst_alloc(&ipv4_dst_ops); if (rt == NULL) { @@ -822,6 +823,8 @@ /* Copy all the information. */ *rt = *rth; + memset(&rt->u.dst.rcu_head, 0, + sizeof(struct rcu_head)); rt->u.dst.__use = 1; atomic_set(&rt->u.dst.__refcnt, 1); if (rt->u.dst.dev) @@ -857,7 +860,6 @@ ip_rt_put(rt); goto do_next; } - read_unlock(&rt_hash_table[hash].lock); do_next: ; } @@ -1037,9 +1039,10 @@ for (i = 0; i < 2; i++) { unsigned hash = rt_hash_code(daddr, skeys[i], tos); - read_lock(&rt_hash_table[hash].lock); for (rth = rt_hash_table[hash].chain; rth; rth = rth->u.rt_next) { + /* read_barrier_depends() here */ + rmb(); if (rth->key.dst == daddr && rth->key.src == skeys[i] && rth->rt_dst == daddr && @@ -1075,7 +1078,6 @@ } } } - read_unlock(&rt_hash_table[hash].lock); } return est_mtu ? : new_mtu; } @@ -1629,8 +1631,9 @@ tos &= IPTOS_RT_MASK; hash = rt_hash_code(daddr, saddr ^ (iif << 5), tos); - read_lock(&rt_hash_table[hash].lock); for (rth = rt_hash_table[hash].chain; rth; rth = rth->u.rt_next) { + /* read_barrier_depends() here */ + rmb(); if (rth->key.dst == daddr && rth->key.src == saddr && rth->key.iif == iif && @@ -1643,12 +1646,10 @@ dst_hold(&rth->u.dst); rth->u.dst.__use++; rt_cache_stat[smp_processor_id()].in_hit++; - read_unlock(&rt_hash_table[hash].lock); skb->dst = (struct dst_entry*)rth; return 0; } } - read_unlock(&rt_hash_table[hash].lock); /* Multicast recognition logic is moved from route cache to here. The problem was that too many Ethernet cards have broken/missing @@ -1988,8 +1989,9 @@ hash = rt_hash_code(key->dst, key->src ^ (key->oif << 5), key->tos); - read_lock_bh(&rt_hash_table[hash].lock); for (rth = rt_hash_table[hash].chain; rth; rth = rth->u.rt_next) { + /* read_barrier_depends() here */ + rmb(); if (rth->key.dst == key->dst && rth->key.src == key->src && rth->key.iif == 0 && @@ -2003,12 +2005,10 @@ dst_hold(&rth->u.dst); rth->u.dst.__use++; rt_cache_stat[smp_processor_id()].out_hit++; - read_unlock_bh(&rt_hash_table[hash].lock); *rp = rth; return 0; } } - read_unlock_bh(&rt_hash_table[hash].lock); return ip_route_output_slow(rp, key); } @@ -2194,9 +2194,10 @@ if (h < s_h) continue; if (h > s_h) s_idx = 0; - read_lock_bh(&rt_hash_table[h].lock); for (rt = rt_hash_table[h].chain, idx = 0; rt; rt = rt->u.rt_next, idx++) { + /* read_barrier_depends() here */ + rmb(); if (idx < s_idx) continue; skb->dst = dst_clone(&rt->u.dst); @@ -2204,12 +2205,10 @@ cb->nlh->nlmsg_seq, RTM_NEWROUTE, 1) <= 0) { dst_release(xchg(&skb->dst, NULL)); - read_unlock_bh(&rt_hash_table[h].lock); goto done; } dst_release(xchg(&skb->dst, NULL)); } - read_unlock_bh(&rt_hash_table[h].lock); } done: @@ -2496,7 +2495,7 @@ rt_hash_mask--; for (i = 0; i <= rt_hash_mask; i++) { - rt_hash_table[i].lock = RW_LOCK_UNLOCKED; + rt_hash_table[i].lock = SPIN_LOCK_UNLOCKED; rt_hash_table[i].chain = NULL; } - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
