Minki/linux
1
0
Fork 1
mirror of https://github.com/torvalds/linux.git synced 2024-11-30 08:01:59 +00:00
Code Issues Packages Projects Releases Wiki Activity
2d2c5ea242
linux/include/net/inet_hashtables.h
Kuniyuki Iwashima d1e5e6408b tcp: Introduce optional per-netns ehash. 
The more sockets we have in the hash table, the longer we spend looking
up the socket.  While running a number of small workloads on the same
host, they penalise each other and cause performance degradation.

The root cause might be a single workload that consumes much more
resources than the others.  It often happens on a cloud service where
different workloads share the same computing resource.

On EC2 c5.24xlarge instance (196 GiB memory and 524288 (1Mi / 2) ehash
entries), after running iperf3 in different netns, creating 24Mi sockets
without data transfer in the root netns causes about 10% performance
regression for the iperf3's connection.

 thash_entries		sockets		length		Gbps
	524288		      1		     1		50.7
			   24Mi		    48		45.1

It is basically related to the length of the list of each hash bucket.
For testing purposes to see how performance drops along the length,
I set 131072 (1Mi / 8) to thash_entries, and here's the result.

 thash_entries		sockets		length		Gbps
        131072		      1		     1		50.7
			    1Mi		     8		49.9
			    2Mi		    16		48.9
			    4Mi		    32		47.3
			    8Mi		    64		44.6
			   16Mi		   128		40.6
			   24Mi		   192		36.3
			   32Mi		   256		32.5
			   40Mi		   320		27.0
			   48Mi		   384		25.0

To resolve the socket lookup degradation, we introduce an optional
per-netns hash table for TCP, but it's just ehash, and we still share
the global bhash, bhash2 and lhash2.

With a smaller ehash, we can look up non-listener sockets faster and
isolate such noisy neighbours.  In addition, we can reduce lock contention.

We can control the ehash size by a new sysctl knob.  However, depending
on workloads, it will require very sensitive tuning, so we disable the
feature by default (net.ipv4.tcp_child_ehash_entries == 0).  Moreover,
we can fall back to using the global ehash in case we fail to allocate
enough memory for a new ehash.  The maximum size is 16Mi, which is large
enough that even if we have 48Mi sockets, the average list length is 3,
and regression would be less than 1%.

We can check the current ehash size by another read-only sysctl knob,
net.ipv4.tcp_ehash_entries.  A negative value means the netns shares
the global ehash (per-netns ehash is disabled or failed to allocate
memory).

  # dmesg | cut -d ' ' -f 5- | grep "established hash"
  TCP established hash table entries: 524288 (order: 10, 4194304 bytes, vmalloc hugepage)

  # sysctl net.ipv4.tcp_ehash_entries
  net.ipv4.tcp_ehash_entries = 524288  # can be changed by thash_entries

  # sysctl net.ipv4.tcp_child_ehash_entries
  net.ipv4.tcp_child_ehash_entries = 0  # disabled by default

  # ip netns add test1
  # ip netns exec test1 sysctl net.ipv4.tcp_ehash_entries
  net.ipv4.tcp_ehash_entries = -524288  # share the global ehash

  # sysctl -w net.ipv4.tcp_child_ehash_entries=100
  net.ipv4.tcp_child_ehash_entries = 100

  # ip netns add test2
  # ip netns exec test2 sysctl net.ipv4.tcp_ehash_entries
  net.ipv4.tcp_ehash_entries = 128  # own a per-netns ehash with 2^n buckets

When more than two processes in the same netns create per-netns ehash
concurrently with different sizes, we need to guarantee the size in
one of the following ways:

  1) Share the global ehash and create per-netns ehash

  First, unshare() with tcp_child_ehash_entries==0.  It creates dedicated
  netns sysctl knobs where we can safely change tcp_child_ehash_entries
  and clone()/unshare() to create a per-netns ehash.

  2) Control write on sysctl by BPF

  We can use BPF_PROG_TYPE_CGROUP_SYSCTL to allow/deny read/write on
  sysctl knobs.

Note that the global ehash allocated at the boot time is spread over
available NUMA nodes, but inet_pernet_hashinfo_alloc() will allocate
pages for each per-netns ehash depending on the current process's NUMA
policy.  By default, the allocation is done in the local node only, so
the per-netns hash table could fully reside on a random node.  Thus,
depending on the NUMA policy the netns is created with and the CPU the
current thread is running on, we could see some performance differences
for highly optimised networking applications.

Note also that the default values of two sysctl knobs depend on the ehash
size and should be tuned carefully:

  tcp_max_tw_buckets  : tcp_child_ehash_entries / 2
  tcp_max_syn_backlog : max(128, tcp_child_ehash_entries / 128)

As a bonus, we can dismantle netns faster.  Currently, while destroying
netns, we call inet_twsk_purge(), which walks through the global ehash.
It can be potentially big because it can have many sockets other than
TIME_WAIT in all netns.  Splitting ehash changes that situation, where
it's only necessary for inet_twsk_purge() to clean up TIME_WAIT sockets
in each netns.

With regard to this, we do not free the per-netns ehash in inet_twsk_kill()
to avoid UAF while iterating the per-netns ehash in inet_twsk_purge().
Instead, we do it in tcp_sk_exit_batch() after calling tcp_twsk_purge() to
keep it protocol-family-independent.

In the future, we could optimise ehash lookup/iteration further by removing
netns comparison for the per-netns ehash.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2022-09-20 10:21:50 -07:00

472 lines
15 KiB
C
Raw Blame History

/* SPDX-License-Identifier: GPL-2.0-or-later */
/*
* INET An implementation of the TCP/IP protocol suite for the LINUX
* operating system. INET is implemented using the BSD Socket
* interface as the means of communication with the user level.
*
* Authors: Lotsa people, from code originally in tcp
*/
#ifndef _INET_HASHTABLES_H
#define _INET_HASHTABLES_H
#include <linux/interrupt.h>
#include <linux/ip.h>
#include <linux/ipv6.h>
#include <linux/list.h>
#include <linux/slab.h>
#include <linux/socket.h>
#include <linux/spinlock.h>
#include <linux/types.h>
#include <linux/wait.h>
#include <net/inet_connection_sock.h>
#include <net/inet_sock.h>
#include <net/ip.h>
#include <net/sock.h>
#include <net/route.h>
#include <net/tcp_states.h>
#include <net/netns/hash.h>
#include <linux/refcount.h>
#include <asm/byteorder.h>
/* This is for all connections with a full identity, no wildcards.
* The 'e' prefix stands for Establish, but we really put all sockets
* but LISTEN ones.
*/
struct inet_ehash_bucket {
struct hlist_nulls_head chain;
};
/* There are a few simple rules, which allow for local port reuse by
* an application. In essence:
*
* 1) Sockets bound to different interfaces may share a local port.
* Failing that, goto test 2.
* 2) If all sockets have sk->sk_reuse set, and none of them are in
* TCP_LISTEN state, the port may be shared.
* Failing that, goto test 3.
* 3) If all sockets are bound to a specific inet_sk(sk)->rcv_saddr local
* address, and none of them are the same, the port may be
* shared.
* Failing this, the port cannot be shared.
*
* The interesting point, is test #2. This is what an FTP server does
* all day. To optimize this case we use a specific flag bit defined
* below. As we add sockets to a bind bucket list, we perform a
* check of: (newsk->sk_reuse && (newsk->sk_state != TCP_LISTEN))
* As long as all sockets added to a bind bucket pass this test,
* the flag bit will be set.
* The resulting situation is that tcp_v[46]_verify_bind() can just check
* for this flag bit, if it is set and the socket trying to bind has
* sk->sk_reuse set, we don't even have to walk the owners list at all,
* we return that it is ok to bind this socket to the requested local port.
*
* Sounds like a lot of work, but it is worth it. In a more naive
* implementation (ie. current FreeBSD etc.) the entire list of ports
* must be walked for each data port opened by an ftp server. Needless
* to say, this does not scale at all. With a couple thousand FTP
* users logged onto your box, isn't it nice to know that new data
* ports are created in O(1) time? I thought so. ;-) -DaveM
*/
#define FASTREUSEPORT_ANY 1
#define FASTREUSEPORT_STRICT 2
struct inet_bind_bucket {
possible_net_t ib_net;
int l3mdev;
unsigned short port;
signed char fastreuse;
signed char fastreuseport;
kuid_t fastuid;
#if IS_ENABLED(CONFIG_IPV6)
struct in6_addr fast_v6_rcv_saddr;
#endif
__be32 fast_rcv_saddr;
unsigned short fast_sk_family;
bool fast_ipv6_only;
struct hlist_node node;
struct hlist_head owners;
};
struct inet_bind2_bucket {
possible_net_t ib_net;
int l3mdev;
unsigned short port;
union {
#if IS_ENABLED(CONFIG_IPV6)
struct in6_addr v6_rcv_saddr;
#endif
__be32 rcv_saddr;
};
/* Node in the bhash2 inet_bind_hashbucket chain */
struct hlist_node node;
/* List of sockets hashed to this bucket */
struct hlist_head owners;
};
static inline struct net *ib_net(const struct inet_bind_bucket *ib)
{
return read_pnet(&ib->ib_net);
}
static inline struct net *ib2_net(const struct inet_bind2_bucket *ib)
{
return read_pnet(&ib->ib_net);
}
#define inet_bind_bucket_for_each(tb, head) \
hlist_for_each_entry(tb, head, node)
struct inet_bind_hashbucket {
spinlock_t lock;
struct hlist_head chain;
};
/* Sockets can be hashed in established or listening table.
* We must use different 'nulls' end-of-chain value for all hash buckets :
* A socket might transition from ESTABLISH to LISTEN state without
* RCU grace period. A lookup in ehash table needs to handle this case.
*/
#define LISTENING_NULLS_BASE (1U << 29)
struct inet_listen_hashbucket {
spinlock_t lock;
struct hlist_nulls_head nulls_head;
};
/* This is for listening sockets, thus all sockets which possess wildcards. */
#define INET_LHTABLE_SIZE 32 /* Yes, really, this is all you need. */
struct inet_hashinfo {
/* This is for sockets with full identity only. Sockets here will
* always be without wildcards and will have the following invariant:
*
* TCP_ESTABLISHED <= sk->sk_state < TCP_CLOSE
*
*/
struct inet_ehash_bucket *ehash;
spinlock_t *ehash_locks;
unsigned int ehash_mask;
unsigned int ehash_locks_mask;
/* Ok, let's try this, I give up, we do need a local binding
* TCP hash as well as the others for fast bind/connect.
*/
struct kmem_cache *bind_bucket_cachep;
/* This bind table is hashed by local port */
struct inet_bind_hashbucket *bhash;
struct kmem_cache *bind2_bucket_cachep;
/* This bind table is hashed by local port and sk->sk_rcv_saddr (ipv4)
* or sk->sk_v6_rcv_saddr (ipv6). This 2nd bind table is used
* primarily for expediting bind conflict resolution.
*/
struct inet_bind_hashbucket *bhash2;
unsigned int bhash_size;
/* The 2nd listener table hashed by local port and address */
unsigned int lhash2_mask;
struct inet_listen_hashbucket *lhash2;
bool pernet;
};
static inline struct inet_hashinfo *tcp_or_dccp_get_hashinfo(const struct sock *sk)
{
#if IS_ENABLED(CONFIG_IP_DCCP)
return sk->sk_prot->h.hashinfo ? :
sock_net(sk)->ipv4.tcp_death_row.hashinfo;
#else
return sock_net(sk)->ipv4.tcp_death_row.hashinfo;
#endif
}
static inline struct inet_listen_hashbucket *
inet_lhash2_bucket(struct inet_hashinfo *h, u32 hash)
{
return &h->lhash2[hash & h->lhash2_mask];
}
static inline struct inet_ehash_bucket *inet_ehash_bucket(
struct inet_hashinfo *hashinfo,
unsigned int hash)
{
return &hashinfo->ehash[hash & hashinfo->ehash_mask];
}
static inline spinlock_t *inet_ehash_lockp(
struct inet_hashinfo *hashinfo,
unsigned int hash)
{
return &hashinfo->ehash_locks[hash & hashinfo->ehash_locks_mask];
}
int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo);
static inline void inet_hashinfo2_free_mod(struct inet_hashinfo *h)
{
kfree(h->lhash2);
h->lhash2 = NULL;
}
static inline void inet_ehash_locks_free(struct inet_hashinfo *hashinfo)
{
kvfree(hashinfo->ehash_locks);
hashinfo->ehash_locks = NULL;
}
struct inet_hashinfo *inet_pernet_hashinfo_alloc(struct inet_hashinfo *hashinfo,
unsigned int ehash_entries);
void inet_pernet_hashinfo_free(struct inet_hashinfo *hashinfo);
struct inet_bind_bucket *
inet_bind_bucket_create(struct kmem_cache *cachep, struct net *net,
struct inet_bind_hashbucket *head,
const unsigned short snum, int l3mdev);
void inet_bind_bucket_destroy(struct kmem_cache *cachep,
struct inet_bind_bucket *tb);
bool inet_bind_bucket_match(const struct inet_bind_bucket *tb,
const struct net *net, unsigned short port,
int l3mdev);
struct inet_bind2_bucket *
inet_bind2_bucket_create(struct kmem_cache *cachep, struct net *net,
struct inet_bind_hashbucket *head,
unsigned short port, int l3mdev,
const struct sock *sk);
void inet_bind2_bucket_destroy(struct kmem_cache *cachep,
struct inet_bind2_bucket *tb);
struct inet_bind2_bucket *
inet_bind2_bucket_find(const struct inet_bind_hashbucket *head,
const struct net *net,
unsigned short port, int l3mdev,
const struct sock *sk);
bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb,
const struct net *net, unsigned short port,
int l3mdev, const struct sock *sk);
static inline u32 inet_bhashfn(const struct net *net, const __u16 lport,
const u32 bhash_size)
{
return (lport + net_hash_mix(net)) & (bhash_size - 1);
}
static inline struct inet_bind_hashbucket *
inet_bhashfn_portaddr(const struct inet_hashinfo *hinfo, const struct sock *sk,
const struct net *net, unsigned short port)
{
u32 hash;
#if IS_ENABLED(CONFIG_IPV6)
if (sk->sk_family == AF_INET6)
hash = ipv6_portaddr_hash(net, &sk->sk_v6_rcv_saddr, port);
else
#endif
hash = ipv4_portaddr_hash(net, sk->sk_rcv_saddr, port);
return &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
}
struct inet_bind_hashbucket *
inet_bhash2_addr_any_hashbucket(const struct sock *sk, const struct net *net, int port);
/* This should be called whenever a socket's sk_rcv_saddr (ipv4) or
* sk_v6_rcv_saddr (ipv6) changes after it has been binded. The socket's
* rcv_saddr field should already have been updated when this is called.
*/
int inet_bhash2_update_saddr(struct inet_bind_hashbucket *prev_saddr, struct sock *sk);
void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
struct inet_bind2_bucket *tb2, unsigned short port);
/* Caller must disable local BH processing. */
int __inet_inherit_port(const struct sock *sk, struct sock *child);
void inet_put_port(struct sock *sk);
void inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
unsigned long numentries, int scale,
unsigned long low_limit,
unsigned long high_limit);
int inet_hashinfo2_init_mod(struct inet_hashinfo *h);
bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk);
bool inet_ehash_nolisten(struct sock *sk, struct sock *osk,
bool *found_dup_sk);
int __inet_hash(struct sock *sk, struct sock *osk);
int inet_hash(struct sock *sk);
void inet_unhash(struct sock *sk);
struct sock *__inet_lookup_listener(struct net *net,
struct inet_hashinfo *hashinfo,
struct sk_buff *skb, int doff,
const __be32 saddr, const __be16 sport,
const __be32 daddr,
const unsigned short hnum,
const int dif, const int sdif);
static inline struct sock *inet_lookup_listener(struct net *net,
struct inet_hashinfo *hashinfo,
struct sk_buff *skb, int doff,
__be32 saddr, __be16 sport,
__be32 daddr, __be16 dport, int dif, int sdif)
{
return __inet_lookup_listener(net, hashinfo, skb, doff, saddr, sport,
daddr, ntohs(dport), dif, sdif);
}
/* Socket demux engine toys. */
/* What happens here is ugly; there's a pair of adjacent fields in
struct inet_sock; __be16 dport followed by __u16 num. We want to
search by pair, so we combine the keys into a single 32bit value
and compare with 32bit value read from &...->dport. Let's at least
make sure that it's not mixed with anything else...
On 64bit targets we combine comparisons with pair of adjacent __be32
fields in the same way.
*/
#ifdef __BIG_ENDIAN
#define INET_COMBINED_PORTS(__sport, __dport) \
((__force __portpair)(((__force __u32)(__be16)(__sport) << 16) | (__u32)(__dport)))
#else /* __LITTLE_ENDIAN */
#define INET_COMBINED_PORTS(__sport, __dport) \
((__force __portpair)(((__u32)(__dport) << 16) | (__force __u32)(__be16)(__sport)))
#endif
#ifdef __BIG_ENDIAN
#define INET_ADDR_COOKIE(__name, __saddr, __daddr) \
const __addrpair __name = (__force __addrpair) ( \
(((__force __u64)(__be32)(__saddr)) << 32) | \
((__force __u64)(__be32)(__daddr)))
#else /* __LITTLE_ENDIAN */
#define INET_ADDR_COOKIE(__name, __saddr, __daddr) \
const __addrpair __name = (__force __addrpair) ( \
(((__force __u64)(__be32)(__daddr)) << 32) | \
((__force __u64)(__be32)(__saddr)))
#endif /* __BIG_ENDIAN */
static inline bool inet_match(struct net *net, const struct sock *sk,
const __addrpair cookie, const __portpair ports,
int dif, int sdif)
{
if (!net_eq(sock_net(sk), net) ||
sk->sk_portpair != ports ||
sk->sk_addrpair != cookie)
return false;
/* READ_ONCE() paired with WRITE_ONCE() in sock_bindtoindex_locked() */
return inet_sk_bound_dev_eq(net, READ_ONCE(sk->sk_bound_dev_if), dif,
sdif);
}
/* Sockets in TCP_CLOSE state are _always_ taken out of the hash, so we need
* not check it for lookups anymore, thanks Alexey. -DaveM
*/
struct sock *__inet_lookup_established(struct net *net,
struct inet_hashinfo *hashinfo,
const __be32 saddr, const __be16 sport,
const __be32 daddr, const u16 hnum,
const int dif, const int sdif);
static inline struct sock *
inet_lookup_established(struct net *net, struct inet_hashinfo *hashinfo,
const __be32 saddr, const __be16 sport,
const __be32 daddr, const __be16 dport,
const int dif)
{
return __inet_lookup_established(net, hashinfo, saddr, sport, daddr,
ntohs(dport), dif, 0);
}
static inline struct sock *__inet_lookup(struct net *net,
struct inet_hashinfo *hashinfo,
struct sk_buff *skb, int doff,
const __be32 saddr, const __be16 sport,
const __be32 daddr, const __be16 dport,
const int dif, const int sdif,
bool *refcounted)
{
u16 hnum = ntohs(dport);
struct sock *sk;
sk = __inet_lookup_established(net, hashinfo, saddr, sport,
daddr, hnum, dif, sdif);
*refcounted = true;
if (sk)
return sk;
*refcounted = false;
return __inet_lookup_listener(net, hashinfo, skb, doff, saddr,
sport, daddr, hnum, dif, sdif);
}
static inline struct sock *inet_lookup(struct net *net,
struct inet_hashinfo *hashinfo,
struct sk_buff *skb, int doff,
const __be32 saddr, const __be16 sport,
const __be32 daddr, const __be16 dport,
const int dif)
{
struct sock *sk;
bool refcounted;
sk = __inet_lookup(net, hashinfo, skb, doff, saddr, sport, daddr,
dport, dif, 0, &refcounted);
if (sk && !refcounted && !refcount_inc_not_zero(&sk->sk_refcnt))
sk = NULL;
return sk;
}
static inline struct sock *__inet_lookup_skb(struct inet_hashinfo *hashinfo,
struct sk_buff *skb,
int doff,
const __be16 sport,
const __be16 dport,
const int sdif,
bool *refcounted)
{
struct sock *sk = skb_steal_sock(skb, refcounted);
const struct iphdr *iph = ip_hdr(skb);
if (sk)
return sk;
return __inet_lookup(dev_net(skb_dst(skb)->dev), hashinfo, skb,
doff, iph->saddr, sport,
iph->daddr, dport, inet_iif(skb), sdif,
refcounted);
}
u32 inet6_ehashfn(const struct net *net,
const struct in6_addr *laddr, const u16 lport,
const struct in6_addr *faddr, const __be16 fport);
static inline void sk_daddr_set(struct sock *sk, __be32 addr)
{
sk->sk_daddr = addr; /* alias of inet_daddr */
#if IS_ENABLED(CONFIG_IPV6)
ipv6_addr_set_v4mapped(addr, &sk->sk_v6_daddr);
#endif
}
static inline void sk_rcv_saddr_set(struct sock *sk, __be32 addr)
{
sk->sk_rcv_saddr = addr; /* alias of inet_rcv_saddr */
#if IS_ENABLED(CONFIG_IPV6)
ipv6_addr_set_v4mapped(addr, &sk->sk_v6_rcv_saddr);
#endif
}
int __inet_hash_connect(struct inet_timewait_death_row *death_row,
struct sock *sk, u64 port_offset,
int (*check_established)(struct inet_timewait_death_row *,
struct sock *, __u16,
struct inet_timewait_sock **));
int inet_hash_connect(struct inet_timewait_death_row *death_row,
struct sock *sk);
#endif /* _INET_HASHTABLES_H */
Reference in New Issue View Git Blame Copy Permalink