| Message ID | 20190614232221.248392-4-edumazet@google.com |
|---|---|
| State | Accepted |
| Delegated to: | David Miller |
| Headers | show |
| Series | tcp: add three static keys | expand |
Hi Eric, On Fri, Jun 14, 2019 at 04:22:20PM -0700, Eric Dumazet wrote: > Feng Tang reported a performance regression after introduction > of per TCP socket tx/rx caches, for TCP over loopback (netperf) > > There is high chance the regression is caused by a change on > how well the 32 KB per-thread page (current->task_frag) can > be recycled, and lack of pcp caches for order-3 pages. Exactly! When I checked the regression, I did several experiments, and thought of the simliar idea to add the per-CPU orderX pcp list, the other idea is to add a order3 list in per-cpu softnet_data as local cache. Thanks, Feng > > I could not reproduce the regression myself, cpus all being > spinning on the mm spinlocks for page allocs/freeing, regardless > of enabling or disabling the per tcp socket caches. > > It seems best to disable the feature by default, and let > admins enabling it. > > MM layer either needs to provide scalable order-3 pages > allocations, or could attempt a trylock on zone->lock if > the caller only attempts to get a high-order page and is > able to fallback to order-0 ones in case of pressure. > > Tests run on a 56 cores host (112 hyper threads) > > - 35.49% netperf [kernel.vmlinux] [k] queued_spin_lock_slowpath > - 35.49% queued_spin_lock_slowpath > - 18.18% get_page_from_freelist > - __alloc_pages_nodemask > - 18.18% alloc_pages_current > skb_page_frag_refill > sk_page_frag_refill > tcp_sendmsg_locked > tcp_sendmsg > inet_sendmsg > sock_sendmsg > __sys_sendto > __x64_sys_sendto > do_syscall_64 > entry_SYSCALL_64_after_hwframe > __libc_send > + 17.31% __free_pages_ok > + 31.43% swapper [kernel.vmlinux] [k] intel_idle > + 9.12% netperf [kernel.vmlinux] [k] copy_user_enhanced_fast_string > + 6.53% netserver [kernel.vmlinux] [k] copy_user_enhanced_fast_string > + 0.69% netserver [kernel.vmlinux] [k] queued_spin_lock_slowpath > + 0.68% netperf [kernel.vmlinux] [k] skb_release_data > + 0.52% netperf [kernel.vmlinux] [k] tcp_sendmsg_locked > 0.46% netperf [kernel.vmlinux] [k] _raw_spin_lock_irqsave > > Fixes: 472c2e07eef0 ("tcp: add one skb cache for tx") > Signed-off-by: Eric Dumazet <edumazet@google.com> > Reported-by: Feng Tang <feng.tang@intel.com> > --- > include/net/sock.h | 4 +++- > net/ipv4/sysctl_net_ipv4.c | 8 ++++++++ > 2 files changed, 11 insertions(+), 1 deletion(-)
diff --git a/include/net/sock.h b/include/net/sock.h index b02645e2dfad722769c1455bcde76e46da9fc5ac..7d7f4ce63bb2aae7c87a9445d11339b6e6b19724 100644 --- a/include/net/sock.h +++ b/include/net/sock.h @@ -1463,12 +1463,14 @@ static inline void sk_mem_uncharge(struct sock *sk, int size) __sk_mem_reclaim(sk, 1 << 20); } +DECLARE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key); static inline void sk_wmem_free_skb(struct sock *sk, struct sk_buff *skb) { sock_set_flag(sk, SOCK_QUEUE_SHRUNK); sk->sk_wmem_queued -= skb->truesize; sk_mem_uncharge(sk, skb->truesize); - if (!sk->sk_tx_skb_cache && !skb_cloned(skb)) { + if (static_branch_unlikely(&tcp_tx_skb_cache_key) && + !sk->sk_tx_skb_cache && !skb_cloned(skb)) { skb_zcopy_clear(skb, true); sk->sk_tx_skb_cache = skb; return; diff --git a/net/ipv4/sysctl_net_ipv4.c b/net/ipv4/sysctl_net_ipv4.c index 886b58d31351df44725bdc34081e798bcb89ecf0..08a428a7b2749c4f2a03aa6352e44c053596ef75 100644 --- a/net/ipv4/sysctl_net_ipv4.c +++ b/net/ipv4/sysctl_net_ipv4.c @@ -54,6 +54,8 @@ static int one_day_secs = 24 * 3600; DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key); EXPORT_SYMBOL(tcp_rx_skb_cache_key); +DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key); + /* obsolete */ static int sysctl_tcp_low_latency __read_mostly; @@ -568,6 +570,12 @@ static struct ctl_table ipv4_table[] = { .mode = 0644, .proc_handler = proc_do_static_key, }, + { + .procname = "tcp_tx_skb_cache", + .data = &tcp_tx_skb_cache_key.key, + .mode = 0644, + .proc_handler = proc_do_static_key, + }, { } };
Feng Tang reported a performance regression after introduction of per TCP socket tx/rx caches, for TCP over loopback (netperf) There is high chance the regression is caused by a change on how well the 32 KB per-thread page (current->task_frag) can be recycled, and lack of pcp caches for order-3 pages. I could not reproduce the regression myself, cpus all being spinning on the mm spinlocks for page allocs/freeing, regardless of enabling or disabling the per tcp socket caches. It seems best to disable the feature by default, and let admins enabling it. MM layer either needs to provide scalable order-3 pages allocations, or could attempt a trylock on zone->lock if the caller only attempts to get a high-order page and is able to fallback to order-0 ones in case of pressure. Tests run on a 56 cores host (112 hyper threads) - 35.49% netperf [kernel.vmlinux] [k] queued_spin_lock_slowpath - 35.49% queued_spin_lock_slowpath - 18.18% get_page_from_freelist - __alloc_pages_nodemask - 18.18% alloc_pages_current skb_page_frag_refill sk_page_frag_refill tcp_sendmsg_locked tcp_sendmsg inet_sendmsg sock_sendmsg __sys_sendto __x64_sys_sendto do_syscall_64 entry_SYSCALL_64_after_hwframe __libc_send + 17.31% __free_pages_ok + 31.43% swapper [kernel.vmlinux] [k] intel_idle + 9.12% netperf [kernel.vmlinux] [k] copy_user_enhanced_fast_string + 6.53% netserver [kernel.vmlinux] [k] copy_user_enhanced_fast_string + 0.69% netserver [kernel.vmlinux] [k] queued_spin_lock_slowpath + 0.68% netperf [kernel.vmlinux] [k] skb_release_data + 0.52% netperf [kernel.vmlinux] [k] tcp_sendmsg_locked 0.46% netperf [kernel.vmlinux] [k] _raw_spin_lock_irqsave Fixes: 472c2e07eef0 ("tcp: add one skb cache for tx") Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: Feng Tang <feng.tang@intel.com> --- include/net/sock.h | 4 +++- net/ipv4/sysctl_net_ipv4.c | 8 ++++++++ 2 files changed, 11 insertions(+), 1 deletion(-)