[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
/*
|
|
|
|
* NET Generic infrastructure for Network protocols.
|
|
|
|
*
|
|
|
|
* Definitions for request_sock
|
|
|
|
*
|
|
|
|
* Authors: Arnaldo Carvalho de Melo <acme@conectiva.com.br>
|
|
|
|
*
|
|
|
|
* From code originally in include/net/tcp.h
|
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or
|
|
|
|
* modify it under the terms of the GNU General Public License
|
|
|
|
* as published by the Free Software Foundation; either version
|
|
|
|
* 2 of the License, or (at your option) any later version.
|
|
|
|
*/
|
|
|
|
#ifndef _REQUEST_SOCK_H
|
|
|
|
#define _REQUEST_SOCK_H
|
|
|
|
|
|
|
|
#include <linux/slab.h>
|
2005-06-19 12:47:59 +07:00
|
|
|
#include <linux/spinlock.h>
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
#include <linux/types.h>
|
2008-07-26 11:43:18 +07:00
|
|
|
#include <linux/bug.h>
|
2005-06-19 12:47:59 +07:00
|
|
|
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
#include <net/sock.h>
|
|
|
|
|
2005-06-19 12:47:21 +07:00
|
|
|
struct request_sock;
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
struct sk_buff;
|
|
|
|
struct dst_entry;
|
|
|
|
struct proto;
|
|
|
|
|
2005-06-19 12:47:21 +07:00
|
|
|
struct request_sock_ops {
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
int family;
|
|
|
|
int obj_size;
|
2006-12-07 11:33:20 +07:00
|
|
|
struct kmem_cache *slab;
|
2008-11-22 07:45:22 +07:00
|
|
|
char *slab_name;
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
int (*rtx_syn_ack)(struct sock *sk,
|
2013-03-17 15:23:34 +07:00
|
|
|
struct request_sock *req);
|
2008-08-07 13:50:04 +07:00
|
|
|
void (*send_ack)(struct sock *sk, struct sk_buff *skb,
|
2005-06-19 12:47:21 +07:00
|
|
|
struct request_sock *req);
|
2006-11-15 10:07:45 +07:00
|
|
|
void (*send_reset)(struct sock *sk,
|
|
|
|
struct sk_buff *skb);
|
2005-06-19 12:47:21 +07:00
|
|
|
void (*destructor)(struct request_sock *req);
|
2015-03-23 00:22:19 +07:00
|
|
|
void (*syn_ack_timeout)(const struct request_sock *req);
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
};
|
|
|
|
|
2013-09-23 00:32:20 +07:00
|
|
|
int inet_rtx_syn_ack(struct sock *parent, struct request_sock *req);
|
2012-10-28 06:16:46 +07:00
|
|
|
|
2005-06-19 12:47:21 +07:00
|
|
|
/* struct request_sock - mini sock to represent a connection request
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
*/
|
2005-06-19 12:47:21 +07:00
|
|
|
struct request_sock {
|
2013-10-10 05:21:29 +07:00
|
|
|
struct sock_common __req_common;
|
2015-03-13 06:44:06 +07:00
|
|
|
#define rsk_refcnt __req_common.skc_refcnt
|
2015-03-20 09:04:19 +07:00
|
|
|
#define rsk_hash __req_common.skc_hash
|
2015-03-13 06:44:06 +07:00
|
|
|
|
2013-04-20 04:29:25 +07:00
|
|
|
struct request_sock *dl_next;
|
2015-03-18 08:32:28 +07:00
|
|
|
struct sock *rsk_listener;
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
u16 mss;
|
2012-10-28 06:16:46 +07:00
|
|
|
u8 num_retrans; /* number of retransmits */
|
|
|
|
u8 cookie_ts:1; /* syncookie: encode tcpopts in timestamp */
|
|
|
|
u8 num_timeout:7; /* number of timeouts */
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
/* The following two fields can be easily recomputed I think -AK */
|
|
|
|
u32 window_clamp; /* window clamp at creation time */
|
|
|
|
u32 rcv_wnd; /* rcv_wnd offered first time */
|
|
|
|
u32 ts_recent;
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
struct timer_list rsk_timer;
|
2006-11-16 17:30:37 +07:00
|
|
|
const struct request_sock_ops *rsk_ops;
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
struct sock *sk;
|
2006-07-25 13:32:50 +07:00
|
|
|
u32 secid;
|
2006-11-09 06:04:09 +07:00
|
|
|
u32 peer_secid;
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
};
|
|
|
|
|
2015-03-18 08:32:28 +07:00
|
|
|
static inline struct request_sock *
|
|
|
|
reqsk_alloc(const struct request_sock_ops *ops, struct sock *sk_listener)
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
{
|
2006-12-07 11:33:16 +07:00
|
|
|
struct request_sock *req = kmem_cache_alloc(ops->slab, GFP_ATOMIC);
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
|
2015-03-18 08:32:28 +07:00
|
|
|
if (req) {
|
2005-06-19 12:47:21 +07:00
|
|
|
req->rsk_ops = ops;
|
2015-03-18 08:32:28 +07:00
|
|
|
sock_hold(sk_listener);
|
|
|
|
req->rsk_listener = sk_listener;
|
2015-03-18 08:32:31 +07:00
|
|
|
|
|
|
|
/* Following is temporary. It is coupled with debugging
|
|
|
|
* helpers in reqsk_put() & reqsk_free()
|
|
|
|
*/
|
|
|
|
atomic_set(&req->rsk_refcnt, 0);
|
2015-03-18 08:32:28 +07:00
|
|
|
}
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
return req;
|
|
|
|
}
|
|
|
|
|
2015-03-13 06:44:08 +07:00
|
|
|
static inline struct request_sock *inet_reqsk(struct sock *sk)
|
|
|
|
{
|
|
|
|
return (struct request_sock *)sk;
|
|
|
|
}
|
|
|
|
|
2015-03-19 04:05:38 +07:00
|
|
|
static inline struct sock *req_to_sk(struct request_sock *req)
|
|
|
|
{
|
|
|
|
return (struct sock *)req;
|
|
|
|
}
|
|
|
|
|
2005-06-19 12:47:21 +07:00
|
|
|
static inline void reqsk_free(struct request_sock *req)
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
{
|
2015-03-16 11:12:16 +07:00
|
|
|
/* temporary debugging */
|
|
|
|
WARN_ON_ONCE(atomic_read(&req->rsk_refcnt) != 0);
|
|
|
|
|
2005-06-19 12:47:21 +07:00
|
|
|
req->rsk_ops->destructor(req);
|
2015-03-18 08:32:28 +07:00
|
|
|
if (req->rsk_listener)
|
|
|
|
sock_put(req->rsk_listener);
|
2015-03-16 11:12:16 +07:00
|
|
|
kmem_cache_free(req->rsk_ops->slab, req);
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
}
|
|
|
|
|
2015-03-13 06:44:06 +07:00
|
|
|
static inline void reqsk_put(struct request_sock *req)
|
|
|
|
{
|
|
|
|
if (atomic_dec_and_test(&req->rsk_refcnt))
|
|
|
|
reqsk_free(req);
|
|
|
|
}
|
|
|
|
|
2005-06-19 12:47:59 +07:00
|
|
|
extern int sysctl_max_syn_backlog;
|
|
|
|
|
2005-06-19 12:48:55 +07:00
|
|
|
/** struct listen_sock - listen state
|
2005-06-19 12:47:59 +07:00
|
|
|
*
|
|
|
|
* @max_qlen_log - log_2 of maximal queued SYNs/REQUESTs
|
|
|
|
*/
|
2005-06-19 12:48:55 +07:00
|
|
|
struct listen_sock {
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
int qlen_inc; /* protected by listener lock */
|
|
|
|
int young_inc;/* protected by listener lock */
|
|
|
|
|
|
|
|
/* following fields can be updated by timer */
|
|
|
|
atomic_t qlen_dec; /* qlen = qlen_inc - qlen_dec */
|
|
|
|
atomic_t young_dec;
|
|
|
|
|
|
|
|
u8 max_qlen_log ____cacheline_aligned_in_smp;
|
2011-08-30 10:21:44 +07:00
|
|
|
u8 synflood_warned;
|
|
|
|
/* 2 bytes hole, try to use */
|
2005-06-19 12:47:59 +07:00
|
|
|
u32 hash_rnd;
|
2005-08-10 09:33:31 +07:00
|
|
|
u32 nr_table_entries;
|
2005-06-19 12:47:59 +07:00
|
|
|
struct request_sock *syn_table[0];
|
|
|
|
};
|
|
|
|
|
2012-08-31 19:29:11 +07:00
|
|
|
/*
|
|
|
|
* For a TCP Fast Open listener -
|
|
|
|
* lock - protects the access to all the reqsk, which is co-owned by
|
|
|
|
* the listener and the child socket.
|
|
|
|
* qlen - pending TFO requests (still in TCP_SYN_RECV).
|
|
|
|
* max_qlen - max TFO reqs allowed before TFO is disabled.
|
|
|
|
*
|
|
|
|
* XXX (TFO) - ideally these fields can be made as part of "listen_sock"
|
|
|
|
* structure above. But there is some implementation difficulty due to
|
|
|
|
* listen_sock being part of request_sock_queue hence will be freed when
|
|
|
|
* a listener is stopped. But TFO related fields may continue to be
|
|
|
|
* accessed even after a listener is closed, until its sk_refcnt drops
|
|
|
|
* to 0 implying no more outstanding TFO reqs. One solution is to keep
|
|
|
|
* listen_opt around until sk_refcnt drops to 0. But there is some other
|
|
|
|
* complexity that needs to be resolved. E.g., a listener can be disabled
|
|
|
|
* temporarily through shutdown()->tcp_disconnect(), and re-enabled later.
|
|
|
|
*/
|
|
|
|
struct fastopen_queue {
|
|
|
|
struct request_sock *rskq_rst_head; /* Keep track of past TFO */
|
|
|
|
struct request_sock *rskq_rst_tail; /* requests that caused RST.
|
|
|
|
* This is part of the defense
|
|
|
|
* against spoofing attack.
|
|
|
|
*/
|
|
|
|
spinlock_t lock;
|
|
|
|
int qlen; /* # of pending (TCP_SYN_RECV) reqs */
|
|
|
|
int max_qlen; /* != 0 iff TFO is currently enabled */
|
|
|
|
};
|
|
|
|
|
2005-06-19 12:47:59 +07:00
|
|
|
/** struct request_sock_queue - queue of request_socks
|
|
|
|
*
|
|
|
|
* @rskq_accept_head - FIFO head of established children
|
|
|
|
* @rskq_accept_tail - FIFO tail of established children
|
2005-08-10 10:11:56 +07:00
|
|
|
* @rskq_defer_accept - User waits for some data after accept()
|
2005-06-19 12:47:59 +07:00
|
|
|
* @syn_wait_lock - serializer
|
|
|
|
*
|
|
|
|
* %syn_wait_lock is necessary only to avoid proc interface having to grab the main
|
|
|
|
* lock sock while browsing the listening hash (otherwise it's deadlock prone).
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
struct request_sock_queue {
|
|
|
|
struct request_sock *rskq_accept_head;
|
|
|
|
struct request_sock *rskq_accept_tail;
|
tcp: Revert 'process defer accept as established' changes.
This reverts two changesets, ec3c0982a2dd1e671bad8e9d26c28dcba0039d87
("[TCP]: TCP_DEFER_ACCEPT updates - process as established") and
the follow-on bug fix 9ae27e0adbf471c7a6b80102e38e1d5a346b3b38
("tcp: Fix slab corruption with ipv6 and tcp6fuzz").
This change causes several problems, first reported by Ingo Molnar
as a distcc-over-loopback regression where connections were getting
stuck.
Ilpo Järvinen first spotted the locking problems. The new function
added by this code, tcp_defer_accept_check(), only has the
child socket locked, yet it is modifying state of the parent
listening socket.
Fixing that is non-trivial at best, because we can't simply just grab
the parent listening socket lock at this point, because it would
create an ABBA deadlock. The normal ordering is parent listening
socket --> child socket, but this code path would require the
reverse lock ordering.
Next is a problem noticed by Vitaliy Gusev, he noted:
----------------------------------------
>--- a/net/ipv4/tcp_timer.c
>+++ b/net/ipv4/tcp_timer.c
>@@ -481,6 +481,11 @@ static void tcp_keepalive_timer (unsigned long data)
> goto death;
> }
>
>+ if (tp->defer_tcp_accept.request && sk->sk_state == TCP_ESTABLISHED) {
>+ tcp_send_active_reset(sk, GFP_ATOMIC);
>+ goto death;
Here socket sk is not attached to listening socket's request queue. tcp_done()
will not call inet_csk_destroy_sock() (and tcp_v4_destroy_sock() which should
release this sk) as socket is not DEAD. Therefore socket sk will be lost for
freeing.
----------------------------------------
Finally, Alexey Kuznetsov argues that there might not even be any
real value or advantage to these new semantics even if we fix all
of the bugs:
----------------------------------------
Hiding from accept() sockets with only out-of-order data only
is the only thing which is impossible with old approach. Is this really
so valuable? My opinion: no, this is nothing but a new loophole
to consume memory without control.
----------------------------------------
So revert this thing for now.
Signed-off-by: David S. Miller <davem@davemloft.net>
2008-06-13 06:31:35 +07:00
|
|
|
u8 rskq_defer_accept;
|
2005-06-19 12:48:55 +07:00
|
|
|
struct listen_sock *listen_opt;
|
2012-08-31 19:29:11 +07:00
|
|
|
struct fastopen_queue *fastopenq; /* This is non-NULL iff TFO has been
|
|
|
|
* enabled on this listener. Check
|
|
|
|
* max_qlen != 0 in fastopen_queue
|
|
|
|
* to determine if TFO is enabled
|
|
|
|
* right at this moment.
|
|
|
|
*/
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
|
|
|
|
/* temporary alignment, our goal is to get rid of this lock */
|
2015-03-23 00:22:21 +07:00
|
|
|
spinlock_t syn_wait_lock ____cacheline_aligned_in_smp;
|
2005-06-19 12:47:59 +07:00
|
|
|
};
|
|
|
|
|
2013-09-23 00:32:20 +07:00
|
|
|
int reqsk_queue_alloc(struct request_sock_queue *queue,
|
|
|
|
unsigned int nr_table_entries);
|
2005-06-19 12:47:59 +07:00
|
|
|
|
2013-09-23 00:32:20 +07:00
|
|
|
void __reqsk_queue_destroy(struct request_sock_queue *queue);
|
|
|
|
void reqsk_queue_destroy(struct request_sock_queue *queue);
|
|
|
|
void reqsk_fastopen_remove(struct sock *sk, struct request_sock *req,
|
|
|
|
bool reset);
|
2005-08-10 09:33:31 +07:00
|
|
|
|
2005-06-19 12:47:59 +07:00
|
|
|
static inline struct request_sock *
|
|
|
|
reqsk_queue_yank_acceptq(struct request_sock_queue *queue)
|
|
|
|
{
|
|
|
|
struct request_sock *req = queue->rskq_accept_head;
|
|
|
|
|
2006-03-27 08:39:55 +07:00
|
|
|
queue->rskq_accept_head = NULL;
|
2005-06-19 12:47:59 +07:00
|
|
|
return req;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline int reqsk_queue_empty(struct request_sock_queue *queue)
|
|
|
|
{
|
|
|
|
return queue->rskq_accept_head == NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void reqsk_queue_add(struct request_sock_queue *queue,
|
|
|
|
struct request_sock *req,
|
|
|
|
struct sock *parent,
|
|
|
|
struct sock *child)
|
|
|
|
{
|
|
|
|
req->sk = child;
|
|
|
|
sk_acceptq_added(parent);
|
|
|
|
|
|
|
|
if (queue->rskq_accept_head == NULL)
|
|
|
|
queue->rskq_accept_head = req;
|
|
|
|
else
|
|
|
|
queue->rskq_accept_tail->dl_next = req;
|
|
|
|
|
|
|
|
queue->rskq_accept_tail = req;
|
|
|
|
req->dl_next = NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline struct request_sock *reqsk_queue_remove(struct request_sock_queue *queue)
|
|
|
|
{
|
|
|
|
struct request_sock *req = queue->rskq_accept_head;
|
|
|
|
|
2008-07-26 11:43:18 +07:00
|
|
|
WARN_ON(req == NULL);
|
2005-06-19 12:47:59 +07:00
|
|
|
|
|
|
|
queue->rskq_accept_head = req->dl_next;
|
|
|
|
if (queue->rskq_accept_head == NULL)
|
|
|
|
queue->rskq_accept_tail = NULL;
|
|
|
|
|
|
|
|
return req;
|
|
|
|
}
|
|
|
|
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
static inline void reqsk_queue_removed(struct request_sock_queue *queue,
|
|
|
|
const struct request_sock *req)
|
2005-06-19 12:47:59 +07:00
|
|
|
{
|
2005-06-19 12:48:55 +07:00
|
|
|
struct listen_sock *lopt = queue->listen_opt;
|
2005-06-19 12:47:59 +07:00
|
|
|
|
2012-10-28 06:16:46 +07:00
|
|
|
if (req->num_timeout == 0)
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
atomic_inc(&lopt->young_dec);
|
|
|
|
atomic_inc(&lopt->qlen_dec);
|
2005-06-19 12:47:59 +07:00
|
|
|
}
|
|
|
|
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
static inline void reqsk_queue_added(struct request_sock_queue *queue)
|
2005-06-19 12:47:59 +07:00
|
|
|
{
|
2005-06-19 12:48:55 +07:00
|
|
|
struct listen_sock *lopt = queue->listen_opt;
|
2005-06-19 12:47:59 +07:00
|
|
|
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
lopt->young_inc++;
|
|
|
|
lopt->qlen_inc++;
|
2005-06-19 12:47:59 +07:00
|
|
|
}
|
|
|
|
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
static inline int listen_sock_qlen(const struct listen_sock *lopt)
|
2005-06-19 12:47:59 +07:00
|
|
|
{
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
return lopt->qlen_inc - atomic_read(&lopt->qlen_dec);
|
2005-06-19 12:47:59 +07:00
|
|
|
}
|
|
|
|
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
static inline int listen_sock_young(const struct listen_sock *lopt)
|
2005-06-19 12:47:59 +07:00
|
|
|
{
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
return lopt->young_inc - atomic_read(&lopt->young_dec);
|
2005-06-19 12:47:59 +07:00
|
|
|
}
|
|
|
|
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
static inline int reqsk_queue_len(const struct request_sock_queue *queue)
|
2005-06-19 12:47:59 +07:00
|
|
|
{
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
const struct listen_sock *lopt = queue->listen_opt;
|
|
|
|
|
|
|
|
return lopt ? listen_sock_qlen(lopt) : 0;
|
2005-06-19 12:47:59 +07:00
|
|
|
}
|
|
|
|
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
static inline int reqsk_queue_len_young(const struct request_sock_queue *queue)
|
2005-06-19 12:47:59 +07:00
|
|
|
{
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
return listen_sock_young(queue->listen_opt);
|
|
|
|
}
|
2015-03-18 08:32:31 +07:00
|
|
|
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
static inline int reqsk_queue_is_full(const struct request_sock_queue *queue)
|
|
|
|
{
|
|
|
|
return reqsk_queue_len(queue) >> queue->listen_opt->max_qlen_log;
|
2005-06-19 12:47:59 +07:00
|
|
|
}
|
|
|
|
|
inet: get rid of central tcp/dccp listener timer
One of the major issue for TCP is the SYNACK rtx handling,
done by inet_csk_reqsk_queue_prune(), fired by the keepalive
timer of a TCP_LISTEN socket.
This function runs for awful long times, with socket lock held,
meaning that other cpus needing this lock have to spin for hundred of ms.
SYNACK are sent in huge bursts, likely to cause severe drops anyway.
This model was OK 15 years ago when memory was very tight.
We now can afford to have a timer per request sock.
Timer invocations no longer need to lock the listener,
and can be run from all cpus in parallel.
With following patch increasing somaxconn width to 32 bits,
I tested a listener with more than 4 million active request sockets,
and a steady SYNFLOOD of ~200,000 SYN per second.
Host was sending ~830,000 SYNACK per second.
This is ~100 times more what we could achieve before this patch.
Later, we will get rid of the listener hash and use ehash instead.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-03-20 09:04:20 +07:00
|
|
|
void reqsk_queue_hash_req(struct request_sock_queue *queue,
|
|
|
|
u32 hash, struct request_sock *req,
|
|
|
|
unsigned long timeout);
|
|
|
|
|
[NET] Generalise TCP's struct open_request minisock infrastructure
Kept this first changeset minimal, without changing existing names to
ease peer review.
Basicaly tcp_openreq_alloc now receives the or_calltable, that in turn
has two new members:
->slab, that replaces tcp_openreq_cachep
->obj_size, to inform the size of the openreq descendant for
a specific protocol
The protocol specific fields in struct open_request were moved to a
class hierarchy, with the things that are common to all connection
oriented PF_INET protocols in struct inet_request_sock, the TCP ones
in tcp_request_sock, that is an inet_request_sock, that is an
open_request.
I.e. this uses the same approach used for the struct sock class
hierarchy, with sk_prot indicating if the protocol wants to use the
open_request infrastructure by filling in sk_prot->rsk_prot with an
or_calltable.
Results? Performance is improved and TCP v4 now uses only 64 bytes per
open request minisock, down from 96 without this patch :-)
Next changeset will rename some of the structs, fields and functions
mentioned above, struct or_calltable is way unclear, better name it
struct request_sock_ops, s/struct open_request/struct request_sock/g,
etc.
Signed-off-by: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
2005-06-19 12:46:52 +07:00
|
|
|
#endif /* _REQUEST_SOCK_H */
|