Patchwork /proc/net/tcp, overhead removed

login
register
mail settings
Submitter Yakov Lerner
Date Sept. 28, 2009, 11:01 p.m.
Message ID <1254178906-5293-1-git-send-email-iler.ml@gmail.com>
Download mbox | patch
Permalink /patch/34405/
State Changes Requested
Delegated to: David Miller
Headers show

Comments

Yakov Lerner - Sept. 28, 2009, 11:01 p.m.
Take 2. 

"Sharp improvement in performance of /proc/net/tcp when number of 
sockets is large and hashsize is large. 
O(numsock * hashsize) time becomes O(numsock + hashsize). On slow
processors, speed difference can be x100 and more."

I must say that I'm not fully satisfied with my choice of "st->sbucket" 
for the new preserved index. The better name would be "st->snum". 
Re-using "st->sbucket" saves 4 bytes, and keeps the patch to one sourcefile.
But "st->sbucket" has different meaning in OPENREQ and LISTEN states;
this can be confusing. 
Maybe better add "snum" member to struct tcp_iter_state ?

Shall I change subject when sending "take N+1", or keep the old subject ?

Signed-off-by: Yakov Lerner <iler.ml@gmail.com>
---
 net/ipv4/tcp_ipv4.c |   35 +++++++++++++++++++++++++++++++++--
 1 files changed, 33 insertions(+), 2 deletions(-)
Eric Dumazet - Sept. 29, 2009, 4:39 a.m.
Yakov Lerner a écrit :
> Take 2. 
> 
> "Sharp improvement in performance of /proc/net/tcp when number of 
> sockets is large and hashsize is large. 
> O(numsock * hashsize) time becomes O(numsock + hashsize). On slow
> processors, speed difference can be x100 and more."
> 
> I must say that I'm not fully satisfied with my choice of "st->sbucket" 
> for the new preserved index. The better name would be "st->snum". 
> Re-using "st->sbucket" saves 4 bytes, and keeps the patch to one sourcefile.
> But "st->sbucket" has different meaning in OPENREQ and LISTEN states;
> this can be confusing. 
> Maybe better add "snum" member to struct tcp_iter_state ?

You can add more fields to tcp_iter_state if it makes code more easy to read
and faster.

This structure is allocated once at open("/proc/net/tcp") time and could
be any reasonable size. You can add 10 longs in it, it is not a big deal.

> 
> Shall I change subject when sending "take N+1", or keep the old subject ?

Not a big deal, but keeping old subject is probably the common way.

[PATCH v2] tcp: Remove /proc/net/tcp O(N*H) overhead

> 
> Signed-off-by: Yakov Lerner <iler.ml@gmail.com>
> ---
>  net/ipv4/tcp_ipv4.c |   35 +++++++++++++++++++++++++++++++++--
>  1 files changed, 33 insertions(+), 2 deletions(-)
> 
> diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
> index 7cda24b..e4c4f19 100644
> --- a/net/ipv4/tcp_ipv4.c
> +++ b/net/ipv4/tcp_ipv4.c
> @@ -1994,13 +1994,14 @@ static inline int empty_bucket(struct tcp_iter_state *st)
>  		hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].twchain);
>  }
>  
> -static void *established_get_first(struct seq_file *seq)
> +static void *established_get_first_after(struct seq_file *seq, int bucket)
>  {
>  	struct tcp_iter_state *st = seq->private;
>  	struct net *net = seq_file_net(seq);
>  	void *rc = NULL;
>  
> -	for (st->bucket = 0; st->bucket < tcp_hashinfo.ehash_size; ++st->bucket) {
> +	for (st->bucket = bucket; st->bucket < tcp_hashinfo.ehash_size;
> +	     ++st->bucket) {
>  		struct sock *sk;
>  		struct hlist_nulls_node *node;
>  		struct inet_timewait_sock *tw;
> @@ -2010,6 +2011,8 @@ static void *established_get_first(struct seq_file *seq)
>  		if (empty_bucket(st))
>  			continue;
>  

> +		st->sbucket = st->num;
> +

oh this is ugly...

Check tcp_seq_stop() to see why st->sbucket should not change after getting
lock. Any reader of this will have a heart attack :)

>  		spin_lock_bh(lock);
>  		sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[st->bucket].chain) {
>  			if (sk->sk_family != st->family ||
> @@ -2036,6 +2039,11 @@ out:
>  	return rc;
>  }
>  
> +static void *established_get_first(struct seq_file *seq)
> +{
> +	return established_get_first_after(seq, 0);
> +}
> +
>  static void *established_get_next(struct seq_file *seq, void *cur)
>  {
>  	struct sock *sk = cur;
> @@ -2064,6 +2072,9 @@ get_tw:
>  		while (++st->bucket < tcp_hashinfo.ehash_size &&
>  				empty_bucket(st))
>  			;
> +
> +		st->sbucket = st->num;

same here, this is ugly, even if it happens to work.

> +
>  		if (st->bucket >= tcp_hashinfo.ehash_size)
>  			return NULL;
>  
> @@ -2107,6 +2118,7 @@ static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
>  
>  	if (!rc) {
>  		st->state = TCP_SEQ_STATE_ESTABLISHED;
> +		st->sbucket = 0;
>  		rc	  = established_get_idx(seq, pos);
>  	}
>  
> @@ -2116,6 +2128,25 @@ static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
>  static void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
>  {
>  	struct tcp_iter_state *st = seq->private;
> +
> +	if (*pos && *pos >= st->sbucket &&
> +	    (st->state == TCP_SEQ_STATE_ESTABLISHED ||
> +	     st->state == TCP_SEQ_STATE_TIME_WAIT)) {
> +		void *cur;
> +		int nskip;
> +
> +		/* for states estab and tw, st->sbucket is index (*pos) */
> +		/* corresponding to the beginning of bucket st->bucket */
> +
> +		st->num = st->sbucket;
ugly...
> +		/* jump to st->bucket, then skip (*pos - st->sbucket) items */
> +		st->state = TCP_SEQ_STATE_ESTABLISHED;
> +		cur = established_get_first_after(seq, st->bucket);
> +		for (nskip = *pos - st->num; cur && nskip > 0; --nskip)
> +			cur = established_get_next(seq, cur);
> +		return cur;
> +	}
> +

I dont think you need this chunk in tcp_get_start(), and its also probably buggy,
even if its hard to prove this claim, we'll need some prog to get TIME_WAIT sockets
in a reproducable form.

Jumping to the right hash slot is more than enough to avoid the O(N*H) problem.

You should try to optimize both established/listening algos, so that
code is readable and maintenable. On pathological cases, we can also have 10000
sockets in LISTENING/OPENREQ state.

Maybe we need a first patch to cleanup code, since its a really complex one,
then a patch to optimize it ?

IMHO the /proc/net/tcp file suffers from bugs, before a performance problem.

Currently, we can miss to output some live sockets in the dump, if :

Thread A gets a block from /proc/net/tcp and stops in hash slot N, socket X.
Thread B deletes sockets X, before socket Y in hash chain, or any socket
in previous hash slots.
Thread A gets 'next block', missing socket Y and possibly Y+1, Y+2....

-> Thread A doesnt see socket Y as an established/timewait socket.

So I believe being able to store the hash slot could really help both performance and
avoid skiping lot of sockets in case a thread B destroys sockets 'before our cursor'

The remaining window would be small, as only deleting sockets in our hash slot could
make us skip live sockets. (And closing this hole is really tricky, inet_diag has
same problem I believe)

Following program to establish 10000 sockets in listening state, and 2*10000 in
established state. Non random ports so that we can compare before/after patches.

#include <sys/types.h>
#include <sys/socket.h>
#include <netinet/in.h>
#include <string.h>
#include <unistd.h>
#include <stdio.h>

int fdlisten[10000];
#define PORT 2222
int main(int argc, char *argv[])
{
        int i;
        struct sockaddr_in sockaddr, locaddr;

        for (i = 0; i < 10000; i++) {
                fdlisten[i] = socket(AF_INET, SOCK_STREAM, 0);
                memset(&sockaddr, 0, sizeof(sockaddr));
                sockaddr.sin_family = AF_INET;
                sockaddr.sin_port = htons(PORT);
                sockaddr.sin_addr.s_addr = htonl(0x7f000001 + i);
                if (bind(fdlisten[i], (struct sockaddr *)&sockaddr, sizeof(sockaddr))== -1) {
                        perror("bind");
                        return 1;
                }
                if (listen(fdlisten[i], 1)== -1) {
                        perror("listen");
                        return 1;
                }
        }
        if (fork() == 0) {
                i = 0;
                while (1) {
                        socklen_t len = sizeof(sockaddr);
                        int newfd = accept(fdlisten[i++], (struct sockaddr *)&sockaddr, &len);

                        if (newfd == -1)
                                perror("accept");
                        if (i == 10000)
                                i = 0;
                }
        }
        for (i = 0 ; i < 10000; i++) {
                int fd;

                close(fdlisten[i]);
                fd = socket(AF_INET, SOCK_STREAM, 0);
                if (fd == -1) {
                        perror("socket");
                        break;
                        }
                memset(&locaddr, 0, sizeof(locaddr));
                locaddr.sin_family = AF_INET;
                locaddr.sin_port = htons(i + 20000);
                locaddr.sin_addr.s_addr = htonl(0x7f000001 + i);
                bind(fd, (struct sockaddr *)&locaddr, sizeof(locaddr));

                memset(&sockaddr, 0, sizeof(sockaddr));
                sockaddr.sin_family = AF_INET;
                sockaddr.sin_port = htons(PORT);
                sockaddr.sin_addr.s_addr = htonl(0x7f000001 + i);
                connect(fd, (struct sockaddr *)&sockaddr, sizeof(sockaddr));
        }
        pause();
        return 0;
}
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Eric Dumazet - Sept. 29, 2009, 7:56 a.m.
Yakov Lerner a écrit :
> Take 2. 
> 
> "Sharp improvement in performance of /proc/net/tcp when number of 
> sockets is large and hashsize is large. 
> O(numsock * hashsize) time becomes O(numsock + hashsize). On slow
> processors, speed difference can be x100 and more."
> 
> I must say that I'm not fully satisfied with my choice of "st->sbucket" 
> for the new preserved index. The better name would be "st->snum". 
> Re-using "st->sbucket" saves 4 bytes, and keeps the patch to one sourcefile.
> But "st->sbucket" has different meaning in OPENREQ and LISTEN states;
> this can be confusing. 
> Maybe better add "snum" member to struct tcp_iter_state ?
> 
> Shall I change subject when sending "take N+1", or keep the old subject ?
> 
> Signed-off-by: Yakov Lerner <iler.ml@gmail.com>
> ---
>  net/ipv4/tcp_ipv4.c |   35 +++++++++++++++++++++++++++++++++--
>  1 files changed, 33 insertions(+), 2 deletions(-)
> 
> diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
> index 7cda24b..e4c4f19 100644
> --- a/net/ipv4/tcp_ipv4.c
> +++ b/net/ipv4/tcp_ipv4.c
> @@ -1994,13 +1994,14 @@ static inline int empty_bucket(struct tcp_iter_state *st)
>  		hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].twchain);
>  }
>  
> -static void *established_get_first(struct seq_file *seq)
> +static void *established_get_first_after(struct seq_file *seq, int bucket)
>  {
>  	struct tcp_iter_state *st = seq->private;
>  	struct net *net = seq_file_net(seq);
>  	void *rc = NULL;
>  
> -	for (st->bucket = 0; st->bucket < tcp_hashinfo.ehash_size; ++st->bucket) {
> +	for (st->bucket = bucket; st->bucket < tcp_hashinfo.ehash_size;
> +	     ++st->bucket) {
>  		struct sock *sk;
>  		struct hlist_nulls_node *node;
>  		struct inet_timewait_sock *tw;
> @@ -2010,6 +2011,8 @@ static void *established_get_first(struct seq_file *seq)
>  		if (empty_bucket(st))
>  			continue;
>  
> +		st->sbucket = st->num;
> +
>  		spin_lock_bh(lock);
>  		sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[st->bucket].chain) {
>  			if (sk->sk_family != st->family ||
> @@ -2036,6 +2039,11 @@ out:
>  	return rc;
>  }
>  
> +static void *established_get_first(struct seq_file *seq)
> +{
> +	return established_get_first_after(seq, 0);
> +}
> +
>  static void *established_get_next(struct seq_file *seq, void *cur)
>  {
>  	struct sock *sk = cur;
> @@ -2064,6 +2072,9 @@ get_tw:
>  		while (++st->bucket < tcp_hashinfo.ehash_size &&
>  				empty_bucket(st))
>  			;
> +
> +		st->sbucket = st->num;
> +
>  		if (st->bucket >= tcp_hashinfo.ehash_size)
>  			return NULL;
>  
> @@ -2107,6 +2118,7 @@ static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
>  
>  	if (!rc) {
>  		st->state = TCP_SEQ_STATE_ESTABLISHED;
> +		st->sbucket = 0;
>  		rc	  = established_get_idx(seq, pos);
>  	}
>  
> @@ -2116,6 +2128,25 @@ static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
>  static void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
>  {
>  	struct tcp_iter_state *st = seq->private;
> +
> +	if (*pos && *pos >= st->sbucket &&
> +	    (st->state == TCP_SEQ_STATE_ESTABLISHED ||
> +	     st->state == TCP_SEQ_STATE_TIME_WAIT)) {
> +		void *cur;
> +		int nskip;
> +
> +		/* for states estab and tw, st->sbucket is index (*pos) */
> +		/* corresponding to the beginning of bucket st->bucket */
> +
> +		st->num = st->sbucket;
> +		/* jump to st->bucket, then skip (*pos - st->sbucket) items */
> +		st->state = TCP_SEQ_STATE_ESTABLISHED;
> +		cur = established_get_first_after(seq, st->bucket);
> +		for (nskip = *pos - st->num; cur && nskip > 0; --nskip)
> +			cur = established_get_next(seq, cur);
> +		return cur;
> +	}
> +
>  	st->state = TCP_SEQ_STATE_LISTENING;
>  	st->num = 0;
>  	return *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;

Just in case you are working on "take 3" of the patch, there is a fondamental problem.

All the scalability problems come from the fact that tcp_seq_start()
*has* to rescan all the tables from the begining, because of lseek() capability
on /proc/net/tcp file 

We probably could disable llseek() (on other positions than start of the file),
and rely only on internal state (listening/established hashtable, hash bucket, position in chain)

I cannot imagine how an application could rely on lseek() on >0 position in this file.


--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Yakov Lerner - Sept. 29, 2009, 8:55 a.m.
On Tue, Sep 29, 2009 at 10:56, Eric Dumazet <eric.dumazet@gmail.com> wrote:
>
> Yakov Lerner a écrit :
> > Take 2.
> >
> > "Sharp improvement in performance of /proc/net/tcp when number of
> > sockets is large and hashsize is large.
> > O(numsock * hashsize) time becomes O(numsock + hashsize). On slow
> > processors, speed difference can be x100 and more."
> >
> > I must say that I'm not fully satisfied with my choice of "st->sbucket"
> > for the new preserved index. The better name would be "st->snum".
> > Re-using "st->sbucket" saves 4 bytes, and keeps the patch to one sourcefile.
> > But "st->sbucket" has different meaning in OPENREQ and LISTEN states;
> > this can be confusing.
> > Maybe better add "snum" member to struct tcp_iter_state ?
> >
> > Shall I change subject when sending "take N+1", or keep the old subject ?
> >
> > Signed-off-by: Yakov Lerner <iler.ml@gmail.com>
> > ---
> >  net/ipv4/tcp_ipv4.c |   35 +++++++++++++++++++++++++++++++++--
> >  1 files changed, 33 insertions(+), 2 deletions(-)
> >
> > diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
> > index 7cda24b..e4c4f19 100644
> > --- a/net/ipv4/tcp_ipv4.c
> > +++ b/net/ipv4/tcp_ipv4.c
> > @@ -1994,13 +1994,14 @@ static inline int empty_bucket(struct tcp_iter_state *st)
> >               hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].twchain);
> >  }
> >
> > -static void *established_get_first(struct seq_file *seq)
> > +static void *established_get_first_after(struct seq_file *seq, int bucket)
> >  {
> >       struct tcp_iter_state *st = seq->private;
> >       struct net *net = seq_file_net(seq);
> >       void *rc = NULL;
> >
> > -     for (st->bucket = 0; st->bucket < tcp_hashinfo.ehash_size; ++st->bucket) {
> > +     for (st->bucket = bucket; st->bucket < tcp_hashinfo.ehash_size;
> > +          ++st->bucket) {
> >               struct sock *sk;
> >               struct hlist_nulls_node *node;
> >               struct inet_timewait_sock *tw;
> > @@ -2010,6 +2011,8 @@ static void *established_get_first(struct seq_file *seq)
> >               if (empty_bucket(st))
> >                       continue;
> >
> > +             st->sbucket = st->num;
> > +
> >               spin_lock_bh(lock);
> >               sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[st->bucket].chain) {
> >                       if (sk->sk_family != st->family ||
> > @@ -2036,6 +2039,11 @@ out:
> >       return rc;
> >  }
> >
> > +static void *established_get_first(struct seq_file *seq)
> > +{
> > +     return established_get_first_after(seq, 0);
> > +}
> > +
> >  static void *established_get_next(struct seq_file *seq, void *cur)
> >  {
> >       struct sock *sk = cur;
> > @@ -2064,6 +2072,9 @@ get_tw:
> >               while (++st->bucket < tcp_hashinfo.ehash_size &&
> >                               empty_bucket(st))
> >                       ;
> > +
> > +             st->sbucket = st->num;
> > +
> >               if (st->bucket >= tcp_hashinfo.ehash_size)
> >                       return NULL;
> >
> > @@ -2107,6 +2118,7 @@ static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
> >
> >       if (!rc) {
> >               st->state = TCP_SEQ_STATE_ESTABLISHED;
> > +             st->sbucket = 0;
> >               rc        = established_get_idx(seq, pos);
> >       }
> >
> > @@ -2116,6 +2128,25 @@ static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
> >  static void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
> >  {
> >       struct tcp_iter_state *st = seq->private;
> > +
> > +     if (*pos && *pos >= st->sbucket &&
> > +         (st->state == TCP_SEQ_STATE_ESTABLISHED ||
> > +          st->state == TCP_SEQ_STATE_TIME_WAIT)) {
> > +             void *cur;
> > +             int nskip;
> > +
> > +             /* for states estab and tw, st->sbucket is index (*pos) */
> > +             /* corresponding to the beginning of bucket st->bucket */
> > +
> > +             st->num = st->sbucket;
> > +             /* jump to st->bucket, then skip (*pos - st->sbucket) items */
> > +             st->state = TCP_SEQ_STATE_ESTABLISHED;
> > +             cur = established_get_first_after(seq, st->bucket);
> > +             for (nskip = *pos - st->num; cur && nskip > 0; --nskip)
> > +                     cur = established_get_next(seq, cur);
> > +             return cur;
> > +     }
> > +
> >       st->state = TCP_SEQ_STATE_LISTENING;
> >       st->num = 0;
> >       return *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
>
> Just in case you are working on "take 3" of the patch, there is a fondamental problem.
>
> All the scalability problems come from the fact that tcp_seq_start()
> *has* to rescan all the tables from the begining, because of lseek() capability
> on /proc/net/tcp file
>
> We probably could disable llseek() (on other positions than start of the file),
> and rely only on internal state (listening/established hashtable, hash bucket, position in chain)
>
> I cannot imagine how an application could rely on lseek() on >0 position in this file.


I thought  /proc/net/tcp  can  both  be fast and allow lseek;
(1) when no lseek was issued since last read
(we can detect this), /proc/net/tcp can jump to the
last known bucket (common case), vs
(2) switch to slow mode (scan from the beginning of hash)
when lseek was used , no ?
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
stephen hemminger - Sept. 29, 2009, 3:45 p.m.
On Tue, 29 Sep 2009 11:55:18 +0300
Yakov Lerner <iler.ml@gmail.com> wrote:

> On Tue, Sep 29, 2009 at 10:56, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> >
> > Yakov Lerner a écrit :
> > > Take 2.
> > >
> > > "Sharp improvement in performance of /proc/net/tcp when number of
> > > sockets is large and hashsize is large.
> > > O(numsock * hashsize) time becomes O(numsock + hashsize). On slow
> > > processors, speed difference can be x100 and more."
> > >
> > > I must say that I'm not fully satisfied with my choice of "st->sbucket"
> > > for the new preserved index. The better name would be "st->snum".
> > > Re-using "st->sbucket" saves 4 bytes, and keeps the patch to one sourcefile.
> > > But "st->sbucket" has different meaning in OPENREQ and LISTEN states;
> > > this can be confusing.
> > > Maybe better add "snum" member to struct tcp_iter_state ?
> > >
> > > Shall I change subject when sending "take N+1", or keep the old subject ?
> > >
> > > Signed-off-by: Yakov Lerner <iler.ml@gmail.com>
> > > ---
> > >  net/ipv4/tcp_ipv4.c |   35 +++++++++++++++++++++++++++++++++--
> > >  1 files changed, 33 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
> > > index 7cda24b..e4c4f19 100644
> > > --- a/net/ipv4/tcp_ipv4.c
> > > +++ b/net/ipv4/tcp_ipv4.c
> > > @@ -1994,13 +1994,14 @@ static inline int empty_bucket(struct tcp_iter_state *st)
> > >               hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].twchain);
> > >  }
> > >
> > > -static void *established_get_first(struct seq_file *seq)
> > > +static void *established_get_first_after(struct seq_file *seq, int bucket)
> > >  {
> > >       struct tcp_iter_state *st = seq->private;
> > >       struct net *net = seq_file_net(seq);
> > >       void *rc = NULL;
> > >
> > > -     for (st->bucket = 0; st->bucket < tcp_hashinfo.ehash_size; ++st->bucket) {
> > > +     for (st->bucket = bucket; st->bucket < tcp_hashinfo.ehash_size;
> > > +          ++st->bucket) {
> > >               struct sock *sk;
> > >               struct hlist_nulls_node *node;
> > >               struct inet_timewait_sock *tw;
> > > @@ -2010,6 +2011,8 @@ static void *established_get_first(struct seq_file *seq)
> > >               if (empty_bucket(st))
> > >                       continue;
> > >
> > > +             st->sbucket = st->num;
> > > +
> > >               spin_lock_bh(lock);
> > >               sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[st->bucket].chain) {
> > >                       if (sk->sk_family != st->family ||
> > > @@ -2036,6 +2039,11 @@ out:
> > >       return rc;
> > >  }
> > >
> > > +static void *established_get_first(struct seq_file *seq)
> > > +{
> > > +     return established_get_first_after(seq, 0);
> > > +}
> > > +
> > >  static void *established_get_next(struct seq_file *seq, void *cur)
> > >  {
> > >       struct sock *sk = cur;
> > > @@ -2064,6 +2072,9 @@ get_tw:
> > >               while (++st->bucket < tcp_hashinfo.ehash_size &&
> > >                               empty_bucket(st))
> > >                       ;
> > > +
> > > +             st->sbucket = st->num;
> > > +
> > >               if (st->bucket >= tcp_hashinfo.ehash_size)
> > >                       return NULL;
> > >
> > > @@ -2107,6 +2118,7 @@ static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
> > >
> > >       if (!rc) {
> > >               st->state = TCP_SEQ_STATE_ESTABLISHED;
> > > +             st->sbucket = 0;
> > >               rc        = established_get_idx(seq, pos);
> > >       }
> > >
> > > @@ -2116,6 +2128,25 @@ static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
> > >  static void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
> > >  {
> > >       struct tcp_iter_state *st = seq->private;
> > > +
> > > +     if (*pos && *pos >= st->sbucket &&
> > > +         (st->state == TCP_SEQ_STATE_ESTABLISHED ||
> > > +          st->state == TCP_SEQ_STATE_TIME_WAIT)) {
> > > +             void *cur;
> > > +             int nskip;
> > > +
> > > +             /* for states estab and tw, st->sbucket is index (*pos) */
> > > +             /* corresponding to the beginning of bucket st->bucket */
> > > +
> > > +             st->num = st->sbucket;
> > > +             /* jump to st->bucket, then skip (*pos - st->sbucket) items */
> > > +             st->state = TCP_SEQ_STATE_ESTABLISHED;
> > > +             cur = established_get_first_after(seq, st->bucket);
> > > +             for (nskip = *pos - st->num; cur && nskip > 0; --nskip)
> > > +                     cur = established_get_next(seq, cur);
> > > +             return cur;
> > > +     }
> > > +
> > >       st->state = TCP_SEQ_STATE_LISTENING;
> > >       st->num = 0;
> > >       return *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
> >
> > Just in case you are working on "take 3" of the patch, there is a fondamental problem.
> >
> > All the scalability problems come from the fact that tcp_seq_start()
> > *has* to rescan all the tables from the begining, because of lseek() capability
> > on /proc/net/tcp file
> >
> > We probably could disable llseek() (on other positions than start of the file),
> > and rely only on internal state (listening/established hashtable, hash bucket, position in chain)
> >
> > I cannot imagine how an application could rely on lseek() on >0 position in this file.
> 
> 
> I thought  /proc/net/tcp  can  both  be fast and allow lseek;
> (1) when no lseek was issued since last read
> (we can detect this), /proc/net/tcp can jump to the
> last known bucket (common case), vs
> (2) switch to slow mode (scan from the beginning of hash)
> when lseek was used , no ?

If you look at fib_hash and fib_trie, they already do the same thing.
  * fib_hash records last hash chain to avoid overhead of rescan.
  * fib_trie records last route and does fast lookup to restart from there.
Yakov Lerner - Sept. 29, 2009, 5:34 p.m.
On Tue, Sep 29, 2009 at 18:45, Stephen Hemminger <shemminger@vyatta.com> wrote:
> On Tue, 29 Sep 2009 11:55:18 +0300
> Yakov Lerner <iler.ml@gmail.com> wrote:
>
>> On Tue, Sep 29, 2009 at 10:56, Eric Dumazet <eric.dumazet@gmail.com> wrote:
>> >
>> > Yakov Lerner a écrit :
>> > > Take 2.
>> > >
>> > > "Sharp improvement in performance of /proc/net/tcp when number of
>> > > sockets is large and hashsize is large.
>> > > O(numsock * hashsize) time becomes O(numsock + hashsize). On slow
>> > > processors, speed difference can be x100 and more."
>> > >
>> > > I must say that I'm not fully satisfied with my choice of "st->sbucket"
>> > > for the new preserved index. The better name would be "st->snum".
>> > > Re-using "st->sbucket" saves 4 bytes, and keeps the patch to one sourcefile.
>> > > But "st->sbucket" has different meaning in OPENREQ and LISTEN states;
>> > > this can be confusing.
>> > > Maybe better add "snum" member to struct tcp_iter_state ?
>> > >
>> > > Shall I change subject when sending "take N+1", or keep the old subject ?
>> > >
>> > > Signed-off-by: Yakov Lerner <iler.ml@gmail.com>
>> > > ---
>> > >  net/ipv4/tcp_ipv4.c |   35 +++++++++++++++++++++++++++++++++--
>> > >  1 files changed, 33 insertions(+), 2 deletions(-)
>> > >
>> > > diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
>> > > index 7cda24b..e4c4f19 100644
>> > > --- a/net/ipv4/tcp_ipv4.c
>> > > +++ b/net/ipv4/tcp_ipv4.c
>> > > @@ -1994,13 +1994,14 @@ static inline int empty_bucket(struct tcp_iter_state *st)
>> > >               hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].twchain);
>> > >  }
>> > >
>> > > -static void *established_get_first(struct seq_file *seq)
>> > > +static void *established_get_first_after(struct seq_file *seq, int bucket)
>> > >  {
>> > >       struct tcp_iter_state *st = seq->private;
>> > >       struct net *net = seq_file_net(seq);
>> > >       void *rc = NULL;
>> > >
>> > > -     for (st->bucket = 0; st->bucket < tcp_hashinfo.ehash_size; ++st->bucket) {
>> > > +     for (st->bucket = bucket; st->bucket < tcp_hashinfo.ehash_size;
>> > > +          ++st->bucket) {
>> > >               struct sock *sk;
>> > >               struct hlist_nulls_node *node;
>> > >               struct inet_timewait_sock *tw;
>> > > @@ -2010,6 +2011,8 @@ static void *established_get_first(struct seq_file *seq)
>> > >               if (empty_bucket(st))
>> > >                       continue;
>> > >
>> > > +             st->sbucket = st->num;
>> > > +
>> > >               spin_lock_bh(lock);
>> > >               sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[st->bucket].chain) {
>> > >                       if (sk->sk_family != st->family ||
>> > > @@ -2036,6 +2039,11 @@ out:
>> > >       return rc;
>> > >  }
>> > >
>> > > +static void *established_get_first(struct seq_file *seq)
>> > > +{
>> > > +     return established_get_first_after(seq, 0);
>> > > +}
>> > > +
>> > >  static void *established_get_next(struct seq_file *seq, void *cur)
>> > >  {
>> > >       struct sock *sk = cur;
>> > > @@ -2064,6 +2072,9 @@ get_tw:
>> > >               while (++st->bucket < tcp_hashinfo.ehash_size &&
>> > >                               empty_bucket(st))
>> > >                       ;
>> > > +
>> > > +             st->sbucket = st->num;
>> > > +
>> > >               if (st->bucket >= tcp_hashinfo.ehash_size)
>> > >                       return NULL;
>> > >
>> > > @@ -2107,6 +2118,7 @@ static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
>> > >
>> > >       if (!rc) {
>> > >               st->state = TCP_SEQ_STATE_ESTABLISHED;
>> > > +             st->sbucket = 0;
>> > >               rc        = established_get_idx(seq, pos);
>> > >       }
>> > >
>> > > @@ -2116,6 +2128,25 @@ static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
>> > >  static void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
>> > >  {
>> > >       struct tcp_iter_state *st = seq->private;
>> > > +
>> > > +     if (*pos && *pos >= st->sbucket &&
>> > > +         (st->state == TCP_SEQ_STATE_ESTABLISHED ||
>> > > +          st->state == TCP_SEQ_STATE_TIME_WAIT)) {
>> > > +             void *cur;
>> > > +             int nskip;
>> > > +
>> > > +             /* for states estab and tw, st->sbucket is index (*pos) */
>> > > +             /* corresponding to the beginning of bucket st->bucket */
>> > > +
>> > > +             st->num = st->sbucket;
>> > > +             /* jump to st->bucket, then skip (*pos - st->sbucket) items */
>> > > +             st->state = TCP_SEQ_STATE_ESTABLISHED;
>> > > +             cur = established_get_first_after(seq, st->bucket);
>> > > +             for (nskip = *pos - st->num; cur && nskip > 0; --nskip)
>> > > +                     cur = established_get_next(seq, cur);
>> > > +             return cur;
>> > > +     }
>> > > +
>> > >       st->state = TCP_SEQ_STATE_LISTENING;
>> > >       st->num = 0;
>> > >       return *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
>> >
>> > Just in case you are working on "take 3" of the patch, there is a fondamental problem.
>> >
>> > All the scalability problems come from the fact that tcp_seq_start()
>> > *has* to rescan all the tables from the begining, because of lseek() capability
>> > on /proc/net/tcp file
>> >
>> > We probably could disable llseek() (on other positions than start of the file),
>> > and rely only on internal state (listening/established hashtable, hash bucket, position in chain)
>> >
>> > I cannot imagine how an application could rely on lseek() on >0 position in this file.
>>
>>
>> I thought  /proc/net/tcp  can  both  be fast and allow lseek;
>> (1) when no lseek was issued since last read
>> (we can detect this), /proc/net/tcp can jump to the
>> last known bucket (common case), vs
>> (2) switch to slow mode (scan from the beginning of hash)
>> when lseek was used , no ?
>
> If you look at fib_hash and fib_trie, they already do the same thing.
>  * fib_hash records last hash chain to avoid overhead of rescan.
>  * fib_trie records last route and does fast lookup to restart from there.

Thanks for the pointer.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
index 7cda24b..e4c4f19 100644
--- a/net/ipv4/tcp_ipv4.c
+++ b/net/ipv4/tcp_ipv4.c
@@ -1994,13 +1994,14 @@  static inline int empty_bucket(struct tcp_iter_state *st)
 		hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].twchain);
 }
 
-static void *established_get_first(struct seq_file *seq)
+static void *established_get_first_after(struct seq_file *seq, int bucket)
 {
 	struct tcp_iter_state *st = seq->private;
 	struct net *net = seq_file_net(seq);
 	void *rc = NULL;
 
-	for (st->bucket = 0; st->bucket < tcp_hashinfo.ehash_size; ++st->bucket) {
+	for (st->bucket = bucket; st->bucket < tcp_hashinfo.ehash_size;
+	     ++st->bucket) {
 		struct sock *sk;
 		struct hlist_nulls_node *node;
 		struct inet_timewait_sock *tw;
@@ -2010,6 +2011,8 @@  static void *established_get_first(struct seq_file *seq)
 		if (empty_bucket(st))
 			continue;
 
+		st->sbucket = st->num;
+
 		spin_lock_bh(lock);
 		sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[st->bucket].chain) {
 			if (sk->sk_family != st->family ||
@@ -2036,6 +2039,11 @@  out:
 	return rc;
 }
 
+static void *established_get_first(struct seq_file *seq)
+{
+	return established_get_first_after(seq, 0);
+}
+
 static void *established_get_next(struct seq_file *seq, void *cur)
 {
 	struct sock *sk = cur;
@@ -2064,6 +2072,9 @@  get_tw:
 		while (++st->bucket < tcp_hashinfo.ehash_size &&
 				empty_bucket(st))
 			;
+
+		st->sbucket = st->num;
+
 		if (st->bucket >= tcp_hashinfo.ehash_size)
 			return NULL;
 
@@ -2107,6 +2118,7 @@  static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
 
 	if (!rc) {
 		st->state = TCP_SEQ_STATE_ESTABLISHED;
+		st->sbucket = 0;
 		rc	  = established_get_idx(seq, pos);
 	}
 
@@ -2116,6 +2128,25 @@  static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
 static void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
 {
 	struct tcp_iter_state *st = seq->private;
+
+	if (*pos && *pos >= st->sbucket &&
+	    (st->state == TCP_SEQ_STATE_ESTABLISHED ||
+	     st->state == TCP_SEQ_STATE_TIME_WAIT)) {
+		void *cur;
+		int nskip;
+
+		/* for states estab and tw, st->sbucket is index (*pos) */
+		/* corresponding to the beginning of bucket st->bucket */
+
+		st->num = st->sbucket;
+		/* jump to st->bucket, then skip (*pos - st->sbucket) items */
+		st->state = TCP_SEQ_STATE_ESTABLISHED;
+		cur = established_get_first_after(seq, st->bucket);
+		for (nskip = *pos - st->num; cur && nskip > 0; --nskip)
+			cur = established_get_next(seq, cur);
+		return cur;
+	}
+
 	st->state = TCP_SEQ_STATE_LISTENING;
 	st->num = 0;
 	return *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;