: Defer InDataGrams increment until recvmsg() does checksum

PATCH Applied Wang Chen 8 participants 26 messages Nov 30, 2007 – Dec 20, 2007

Subsystems (4)

Networking Network drivers Documentation UDP

lore

Share & tools

lore

Wang Chen

author

2007-11-30 03:12 UTC #0 lore raw

PATCH 1/3 : Defer InDataGrams increment until recvmsg() does checksum

From: Wang Chen <[email protected]>

Thanks dave, herbert, gerrit, andi and other people for your discussion about this problem.

UdpInDatagrams can be confusing because it counts packets that might be dropped later.
Move UdpInDatagrams into recvmsg() as allowed by the RFC.

Signed-off-by Wang Chen <[email protected]>

---
 Documentation/networking/udplite.txt |    2 +-
 net/ipv4/udp.c                       |    7 +++----
 net/ipv6/udp.c                       |    4 +++-
 3 files changed, 7 insertions(+), 6 deletions(-)

Documentation/networking/udplite.txt +1 -1

diff -Nurp linux-2.6.24.rc3.org/Documentation/networking/udplite.txt linux-2.6.24.rc3/Documentation/networking/udplite.txt
@@ -236,7 +236,7 @@
 
   This displays UDP-Lite statistics variables, whose meaning is as follows.
 
-   InDatagrams:     Total number of received datagrams.
+   InDatagrams:     The total number of datagrams delivered to users.
 
    NoPorts:         Number of packets received to an unknown port.
                     These cases are counted separately (not as InErrors).

net/ipv4/udp.c +3 -4

diff -Nurp linux-2.6.24.rc3.org/net/ipv4/udp.c linux-2.6.24.rc3/net/ipv4/udp.c
@@ -873,6 +873,8 @@ try_again:
 	if (err)
 		goto out_free;
 
+	UDP_INC_STATS_BH(UDP_MIB_INDATAGRAMS, is_udplite);
+
 	sock_recv_timestamp(msg, sk, skb);
 
 	/* Copy the address. */

@@ -966,10 +968,8 @@ int udp_queue_rcv_skb(struct sock * sk, 
 			int ret;
 
 			ret = (*up->encap_rcv)(sk, skb);
-			if (ret <= 0) {
-				UDP_INC_STATS_BH(UDP_MIB_INDATAGRAMS, up->pcflag);
+			if (ret <= 0)
 				return -ret;
-			}
 		}
 
 		/* FALLTHROUGH -- it's a UDP Packet */

@@ -1023,7 +1023,6 @@ int udp_queue_rcv_skb(struct sock * sk, 
 		goto drop;
 	}
 
-	UDP_INC_STATS_BH(UDP_MIB_INDATAGRAMS, up->pcflag);
 	return 0;
 
 drop:

net/ipv6/udp.c +3 -1

diff -Nurp linux-2.6.24.rc3.org/net/ipv6/udp.c linux-2.6.24.rc3/net/ipv6/udp.c
@@ -164,6 +164,8 @@ try_again:
 	if (err)
 		goto out_free;
 
+	UDP6_INC_STATS_BH(UDP_MIB_INDATAGRAMS, is_udplite);
+
 	sock_recv_timestamp(msg, sk, skb);
 
 	/* Copy the address. */

@@ -292,7 +294,7 @@ int udpv6_queue_rcv_skb(struct sock * sk
 			UDP6_INC_STATS_BH(UDP_MIB_RCVBUFERRORS, up->pcflag);
 		goto drop;
 	}
-	UDP6_INC_STATS_BH(UDP_MIB_INDATAGRAMS, up->pcflag);
+
 	return 0;
 drop:
 	UDP6_INC_STATS_BH(UDP_MIB_INERRORS, up->pcflag);

Wang Chen

author

2007-11-30 03:18 UTC #1 lore raw

PATCH 2/3 : Clean up for IS_UDPLITE macro

(This patch base on "PATCH 1/3".)

Since we have macro IS_UDPLITE, we can use it.

Signed-off-by Wang Chen <[email protected]>

---
 ipv4/udp.c |   19 +++++++++++--------
 ipv6/udp.c |   14 ++++++++------
 2 files changed, 19 insertions(+), 14 deletions(-)

net/ipv4/udp.c +11 -8

diff -Nurp linux-2.6.24.rc3.org/net/ipv4/udp.c linux-2.6.24.rc3/net/ipv4/udp.c
@@ -471,6 +471,7 @@ static int udp_push_pending_frames(struc
 	struct sk_buff *skb;
 	struct udphdr *uh;
 	int err = 0;
+	int is_udplite = IS_UDPLITE(sk);
 	__wsum csum = 0;
 
 	/* Grab the skbuff where UDP header space exists. */

@@ -486,7 +487,7 @@ static int udp_push_pending_frames(struc
 	uh->len = htons(up->len);
 	uh->check = 0;
 
-	if (up->pcflag)  				 /*     UDP-Lite      */
+	if (is_udplite)  				 /*     UDP-Lite      */
 		csum  = udplite_csum_outgoing(sk, skb);
 
 	else if (sk->sk_no_check == UDP_CSUM_NOXMIT) {   /* UDP csum disabled */

@@ -514,7 +515,7 @@ out:
 	up->len = 0;
 	up->pending = 0;
 	if (!err)
-		UDP_INC_STATS_USER(UDP_MIB_OUTDATAGRAMS, up->pcflag);
+		UDP_INC_STATS_USER(UDP_MIB_OUTDATAGRAMS, is_udplite);
 	return err;
 }

@@ -531,7 +532,7 @@ int udp_sendmsg(struct kiocb *iocb, stru
 	__be32 daddr, faddr, saddr;
 	__be16 dport;
 	u8  tos;
-	int err, is_udplite = up->pcflag;
+	int err, is_udplite = IS_UDPLITE(sk);
 	int corkreq = up->corkflag || msg->msg_flags&MSG_MORE;
 	int (*getfrag)(void *, char *, int, int, int, struct sk_buff *);

@@ -942,6 +943,7 @@ int udp_queue_rcv_skb(struct sock * sk, 
 {
 	struct udp_sock *up = udp_sk(sk);
 	int rc;
+	int is_udplite = IS_UDPLITE(sk);
 
 	/*
 	 *	Charge it to the socket, dropping if the queue is full.

@@ -978,7 +980,7 @@ int udp_queue_rcv_skb(struct sock * sk, 
 	/*
 	 * 	UDP-Lite specific tests, ignored on UDP sockets
 	 */
-	if ((up->pcflag & UDPLITE_RECV_CC)  &&  UDP_SKB_CB(skb)->partial_cov) {
+	if ((is_udplite & UDPLITE_RECV_CC)  &&  UDP_SKB_CB(skb)->partial_cov) {
 
 		/*
 		 * MIB statistics other than incrementing the error count are

@@ -1019,14 +1021,14 @@ int udp_queue_rcv_skb(struct sock * sk, 
 	if ((rc = sock_queue_rcv_skb(sk,skb)) < 0) {
 		/* Note that an ENOMEM error is charged twice */
 		if (rc == -ENOMEM)
-			UDP_INC_STATS_BH(UDP_MIB_RCVBUFERRORS, up->pcflag);
+			UDP_INC_STATS_BH(UDP_MIB_RCVBUFERRORS, is_udplite);
 		goto drop;
 	}
 
 	return 0;
 
 drop:
-	UDP_INC_STATS_BH(UDP_MIB_INERRORS, up->pcflag);
+	UDP_INC_STATS_BH(UDP_MIB_INERRORS, is_udplite);
 	kfree_skb(skb);
 	return -1;
 }

@@ -1235,6 +1237,7 @@ int udp_lib_setsockopt(struct sock *sk, 
 	struct udp_sock *up = udp_sk(sk);
 	int val;
 	int err = 0;
+	int is_udplite = IS_UDPLITE(sk);
 
 	if (optlen<sizeof(int))
 		return -EINVAL;

@@ -1276,7 +1279,7 @@ int udp_lib_setsockopt(struct sock *sk, 
 	/* The sender sets actual checksum coverage length via this option.
 	 * The case coverage > packet length is handled by send module. */
 	case UDPLITE_SEND_CSCOV:
-		if (!up->pcflag)         /* Disable the option on UDP sockets */
+		if (!is_udplite)         /* Disable the option on UDP sockets */
 			return -ENOPROTOOPT;
 		if (val != 0 && val < 8) /* Illegal coverage: use default (8) */
 			val = 8;

@@ -1288,7 +1291,7 @@ int udp_lib_setsockopt(struct sock *sk, 
 	 * sense, this should be set to at least 8 (as done below). If zero is
 	 * used, this again means full checksum coverage.                     */
 	case UDPLITE_RECV_CSCOV:
-		if (!up->pcflag)         /* Disable the option on UDP sockets */
+		if (!is_udplite)         /* Disable the option on UDP sockets */
 			return -ENOPROTOOPT;
 		if (val != 0 && val < 8) /* Avoid silly minimal values.       */
 			val = 8;

net/ipv6/udp.c +8 -6

diff -Nurp linux-2.6.24.rc3.org/net/ipv6/udp.c linux-2.6.24.rc3/net/ipv6/udp.c
@@ -260,6 +260,7 @@ int udpv6_queue_rcv_skb(struct sock * sk
 {
 	struct udp_sock *up = udp_sk(sk);
 	int rc;
+	int is_udplite = IS_UDPLITE(sk);
 
 	if (!xfrm6_policy_check(sk, XFRM_POLICY_IN, skb))
 		goto drop;

@@ -267,7 +268,7 @@ int udpv6_queue_rcv_skb(struct sock * sk
 	/*
 	 * UDP-Lite specific tests, ignored on UDP sockets (see net/ipv4/udp.c).
 	 */
-	if ((up->pcflag & UDPLITE_RECV_CC)  &&  UDP_SKB_CB(skb)->partial_cov) {
+	if ((is_udplite & UDPLITE_RECV_CC)  &&  UDP_SKB_CB(skb)->partial_cov) {
 
 		if (up->pcrlen == 0) {          /* full coverage was set  */
 			LIMIT_NETDEBUG(KERN_WARNING "UDPLITE6: partial coverage"

@@ -291,13 +292,13 @@ int udpv6_queue_rcv_skb(struct sock * sk
 	if ((rc = sock_queue_rcv_skb(sk,skb)) < 0) {
 		/* Note that an ENOMEM error is charged twice */
 		if (rc == -ENOMEM)
-			UDP6_INC_STATS_BH(UDP_MIB_RCVBUFERRORS, up->pcflag);
+			UDP6_INC_STATS_BH(UDP_MIB_RCVBUFERRORS, is_udplite);
 		goto drop;
 	}
 
 	return 0;
 drop:
-	UDP6_INC_STATS_BH(UDP_MIB_INERRORS, up->pcflag);
+	UDP6_INC_STATS_BH(UDP_MIB_INERRORS, is_udplite);
 	kfree_skb(skb);
 	return -1;
 }

@@ -525,6 +526,7 @@ static int udp_v6_push_pending_frames(st
 	struct inet_sock *inet = inet_sk(sk);
 	struct flowi *fl = &inet->cork.fl;
 	int err = 0;
+	int is_udplite = IS_UDPLITE(sk);
 	__wsum csum = 0;
 
 	/* Grab the skbuff where UDP header space exists. */

@@ -540,7 +542,7 @@ static int udp_v6_push_pending_frames(st
 	uh->len = htons(up->len);
 	uh->check = 0;
 
-	if (up->pcflag)
+	if (is_udplite)
 		csum = udplite_csum_outgoing(sk, skb);
 	 else
 		csum = udp_csum_outgoing(sk, skb);

@@ -556,7 +558,7 @@ out:
 	up->len = 0;
 	up->pending = 0;
 	if (!err)
-		UDP6_INC_STATS_USER(UDP_MIB_OUTDATAGRAMS, up->pcflag);
+		UDP6_INC_STATS_USER(UDP_MIB_OUTDATAGRAMS, is_udplite);
 	return err;
 }

@@ -580,7 +582,7 @@ int udpv6_sendmsg(struct kiocb *iocb, st
 	int corkreq = up->corkflag || msg->msg_flags&MSG_MORE;
 	int err;
 	int connected = 0;
-	int is_udplite = up->pcflag;
+	int is_udplite = IS_UDPLITE(sk);
 	int (*getfrag)(void *, char *, int, int, int, struct sk_buff *);
 
 	/* destination address check */

Gerrit Renker

Acked-by on v1 2/3

2007-11-30 11:06 UTC #2 lore raw

Re: [PATCH 2/3] [UDP]: Clean up for IS_UDPLITE macro

| (This patch base on "PATCH 1/3".)
| 
| Since we have macro IS_UDPLITE, we can use it.
| 
| Signed-off-by: Wang Chen <[email protected]>
| ---

Acked-by Gerrit Renker <[email protected]>

Wang Chen

author

2007-11-30 03:26 UTC #3 lore raw

PATCH 3/3 : Counter increment on BH mode

In reply to Wang Chen

(This patch base on "PATCH 2/3".)

UDP_MIB_INERRORS increment is in different way for IPv4 and IPv6. For the consistence, change UDP6_INC_STATS_USER to UDP6_INC_STATS_BH.

Signed-off-by Wang Chen <[email protected]>

---
 udp.c |    2 +-
 1 files changed, 1 insertion(+), 1 deletion(-)

net/ipv6/udp.c +1 -1

@@ -207,7 +207,7 @@ out:
 	return err;
 
 csum_copy_err:
-	UDP6_INC_STATS_USER(UDP_MIB_INERRORS, is_udplite);
+	UDP6_INC_STATS_BH(UDP_MIB_INERRORS, is_udplite);
 	skb_kill_datagram(sk, skb, flags);
 
 	if (flags & MSG_DONTWAIT)

Gerrit Renker

2007-11-30 11:20 UTC #4 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Quoting Wang Chen:

| (This patch base on "PATCH 2/3".)
| 
| UDP_MIB_INERRORS increment is in different way for IPv4
| and IPv6. For the consistence, change UDP6_INC_STATS_USER
| to UDP6_INC_STATS_BH.
| 
| Signed-off-by: Wang Chen <[email protected]>
| ---
|  udp.c |    2 +-
|  1 files changed, 1 insertion(+), 1 deletion(-)
| 
| --- linux-2.6.24.rc3.org/net/ipv6/udp.c	2007-11-30 10:38:36.000000000 +0800
| +++ linux-2.6.24.rc3/net/ipv6/udp.c	2007-11-30 10:39:01.000000000 +0800
| @@ -207,7 +207,7 @@ out:
|  	return err;
|  
|  csum_copy_err:
| -	UDP6_INC_STATS_USER(UDP_MIB_INERRORS, is_udplite);
| +	UDP6_INC_STATS_BH(UDP_MIB_INERRORS, is_udplite);
|  	skb_kill_datagram(sk, skb, flags);
|  
|  	if (flags & MSG_DONTWAIT)
|

Is it not the other way round ?? :-

	struct proto udp{,v6}_prot = {
		// ...
		.recvmsg           = udp{,v6}_recvmsg,
	};

Herbert Xu

2007-12-01 01:54 UTC #5 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Quoted 9 lines · Gerrit Renker

|  csum_copy_err:
| -	UDP6_INC_STATS_USER(UDP_MIB_INERRORS, is_udplite);
| +	UDP6_INC_STATS_BH(UDP_MIB_INERRORS, is_udplite);
|  	skb_kill_datagram(sk, skb, flags);
|  
|  	if (flags & MSG_DONTWAIT)
|

Is it not the other way round ?? :-

I agree. Wang Chen, please change this and other appropriate BH calls to USER. Basically recvmsg should always be USER while rcv is BH.

Thanks,

Signature · 4 lines

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Wang Chen

author

2007-12-03 07:22 UTC #6 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Herbert Xu said the following on 2007-12-1 9:54:

Quoted 10 lines · Gerrit Renker

Quoted 8 lines

|  csum_copy_err:
| -	UDP6_INC_STATS_USER(UDP_MIB_INERRORS, is_udplite);
| +	UDP6_INC_STATS_BH(UDP_MIB_INERRORS, is_udplite);
|  	skb_kill_datagram(sk, skb, flags);
|  
|  	if (flags & MSG_DONTWAIT)
|

Is it not the other way round ?? :-

Quoted 4 lines

I agree. Wang Chen, please change this and other appropriate BH calls to USER. Basically recvmsg should always be USER while rcv is BH.

Resubmit the patch.

System calls should be USER. So change the BH to USER for UDP*_INC_STATS_BH().

Signed-off-by Wang Chen <[email protected]>

---
 ipv4/udp.c |    4 ++--
 ipv6/udp.c |    2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

net/ipv4/udp.c +2 -2

@@ -874,7 +874,7 @@ try_again:
 	if (err)
 		goto out_free;
 
-	UDP_INC_STATS_BH(UDP_MIB_INDATAGRAMS, is_udplite);
+	UDP_INC_STATS_USER(UDP_MIB_INDATAGRAMS, is_udplite);
 
 	sock_recv_timestamp(msg, sk, skb);

@@ -899,7 +899,7 @@ out:
 	return err;
 
 csum_copy_err:
-	UDP_INC_STATS_BH(UDP_MIB_INERRORS, is_udplite);
+	UDP_INC_STATS_USER(UDP_MIB_INERRORS, is_udplite);
 
 	skb_kill_datagram(sk, skb, flags);

net/ipv6/udp.c +1 -1

@@ -164,7 +164,7 @@ try_again:
 	if (err)
 		goto out_free;
 
-	UDP6_INC_STATS_BH(UDP_MIB_INDATAGRAMS, is_udplite);
+	UDP6_INC_STATS_USER(UDP_MIB_INDATAGRAMS, is_udplite);
 
 	sock_recv_timestamp(msg, sk, skb);

Herbert Xu

2007-12-03 11:39 UTC #7 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Quoted 5 lines · Wang Chen

System calls should be USER. So change the BH to USER for UDP*_INC_STATS_BH().
Signed-off-by Wang Chen <[email protected]>

I've rearragned it so that we the new INC_STATS call is USER in the first patch. Otherwise it's all in net-2.6.25 now.

BTW, thanks to Gerrit's note I took a look at the underlying code for the _BH/_USER stuff. It's in fact totally broken when PREEMPT is on. It relies on the fact that process-context kernel code does not get preempted by other process-context kernel code or for that matter migrate to other CPUs, neither of which is true with PREEMPT on.

So we need to fix this, and whatever the fix is will probably render the BH/USER distinction obsolete.

Any takers?

Thanks,

Signature · 4 lines

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Alexey Kuznetsov

2007-12-03 11:49 UTC #8 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Quoted 2 lines · Herbert Xu

So we need to fix this, and whatever the fix is will probably render the BH/USER distinction obsolete.

Hmm, I would think opposite. USER (or generic) is expensive variant, BH is lite. No?

Alexey

Herbert Xu

2007-12-03 11:54 UTC #9 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Quoted 6 lines · Alexey Kuznetsov

Quoted 2 lines · Herbert Xu
So we need to fix this, and whatever the fix is will probably render the BH/USER distinction obsolete.
Hmm, I would think opposite. USER (or generic) is expensive variant, BH is lite. No?

Yes that would certainly be the obvious fix. In other words, just use smp_processor_id instead of the raw version. I suppose the only issue would be that disabling/enabling preemption isn't exactly cost-free and we're going to be doing that for every single increment.

Hmm, wasn't someone else talking about a non-atomic version of atomic ops lately (i.e., atomic with respect to the local CPU only)? Perhaps this is the killer app for it :)

Thanks,

Signature · 4 lines

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Herbert Xu

2007-12-03 13:18 UTC #10 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Quoted 4 lines · Herbert Xu

Hmm, wasn't someone else talking about a non-atomic version of atomic ops lately (i.e., atomic with respect to the local CPU only)? Perhaps this is the killer app for it :)

Never mind, we already have that in local_t and as Alexey correctly points out, USER is still going to be the expensive variant with the preempt_disable (well until BH gets threaded). So how about this patch?

Thanks,

Signature · 5 lines

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
--

include/net/snmp.h +9 -2

index ea206bf..9444b54 100644
@@ -23,6 +23,7 @@
 
 #include <linux/cache.h>
 #include <linux/snmp.h>
+#include <linux/smp.h>
 
 /*
  * Mibs are stored in array of unsigned long.

@@ -137,7 +138,10 @@ struct linux_mib {
 #define SNMP_INC_STATS_OFFSET_BH(mib, field, offset)	\
 	(per_cpu_ptr(mib[0], raw_smp_processor_id())->mibs[field + (offset)]++)
 #define SNMP_INC_STATS_USER(mib, field) \
-	(per_cpu_ptr(mib[1], raw_smp_processor_id())->mibs[field]++)
+	do { \
+		per_cpu_ptr(mib[1], get_cpu())->mibs[field]++; \
+		put_cpu(); \
+	} while (0)
 #define SNMP_INC_STATS(mib, field) 	\
 	(per_cpu_ptr(mib[!in_softirq()], raw_smp_processor_id())->mibs[field]++)
 #define SNMP_DEC_STATS(mib, field) 	\

@@ -145,6 +149,9 @@ struct linux_mib {
 #define SNMP_ADD_STATS_BH(mib, field, addend) 	\
 	(per_cpu_ptr(mib[0], raw_smp_processor_id())->mibs[field] += addend)
 #define SNMP_ADD_STATS_USER(mib, field, addend) 	\
-	(per_cpu_ptr(mib[1], raw_smp_processor_id())->mibs[field] += addend)
+	do { \
+		per_cpu_ptr(mib[1], get_cpu())->mibs[field] += addend; \
+		put_cpu(); \
+	} while (0)
 
 #endif

Wang Chen

author

2007-12-04 03:53 UTC #11 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Herbert Xu said the following on 2007-12-3 21:17:

include/net/snmp.h line 26

Quoted 10 lines · Herbert Xu

include/net/snmp.h +1 -0

index ea206bf..9444b54 100644
@@ -23,6 +23,7 @@
 
 #include <linux/cache.h>
 #include <linux/snmp.h>
+#include <linux/smp.h>

It's no need to include smp.h?

--
WCN

Herbert Xu

2007-12-04 03:58 UTC #12 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Quoted 6 lines · Wang Chen

Quoted 4 lines
#include <linux/cache.h>
#include <linux/snmp.h>
+#include <linux/smp.h>
It's no need to include smp.h?

We need it for smp_processor_id. Well we needed it before too but there's probably some implicit inclusion which has made it work. It's better to declare these inclusions explicitly as otherwise they may break on less common architectures later.

Cheers,

Signature · 4 lines

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Herbert Xu

2007-12-15 13:59 UTC #13 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

In reply to Herbert Xu

Quoted 4 lines · Herbert Xu

Never mind, we already have that in local_t and as Alexey correctly points out, USER is still going to be the expensive variant with the preempt_disable (well until BH gets threaded). So how about this patch?

I didn't hear any objections so here is the patch again.

[SNMP]: Fix SNMP counters with PREEMPT

The SNMP macros use raw_smp_processor_id() in process context which is illegal because the process may be preempted and then migrated to another CPU.

This patch makes it use get_cpu/put_cpu to disable preemption.

Signed-off-by Herbert Xu <[email protected]>

Cheers,

Signature · 5 lines

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
--

include/net/snmp.h +9 -2

index ea206bf..9444b54 100644
@@ -23,6 +23,7 @@
 
 #include <linux/cache.h>
 #include <linux/snmp.h>
+#include <linux/smp.h>
 
 /*
  * Mibs are stored in array of unsigned long.

@@ -137,7 +138,10 @@ struct linux_mib {
 #define SNMP_INC_STATS_OFFSET_BH(mib, field, offset)	\
 	(per_cpu_ptr(mib[0], raw_smp_processor_id())->mibs[field + (offset)]++)
 #define SNMP_INC_STATS_USER(mib, field) \
-	(per_cpu_ptr(mib[1], raw_smp_processor_id())->mibs[field]++)
+	do { \
+		per_cpu_ptr(mib[1], get_cpu())->mibs[field]++; \
+		put_cpu(); \
+	} while (0)
 #define SNMP_INC_STATS(mib, field) 	\
 	(per_cpu_ptr(mib[!in_softirq()], raw_smp_processor_id())->mibs[field]++)
 #define SNMP_DEC_STATS(mib, field) 	\

@@ -145,6 +149,9 @@ struct linux_mib {
 #define SNMP_ADD_STATS_BH(mib, field, addend) 	\
 	(per_cpu_ptr(mib[0], raw_smp_processor_id())->mibs[field] += addend)
 #define SNMP_ADD_STATS_USER(mib, field, addend) 	\
-	(per_cpu_ptr(mib[1], raw_smp_processor_id())->mibs[field] += addend)
+	do { \
+		per_cpu_ptr(mib[1], get_cpu())->mibs[field] += addend; \
+		put_cpu(); \
+	} while (0)
 
 #endif

Eric Dumazet

2007-12-15 17:07 UTC #14 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Quoted 13 lines · Herbert Xu

Quoted 3 lines · Herbert Xu
Never mind, we already have that in local_t and as Alexey correctly points out, USER is still going to be the expensive variant with the preempt_disable (well until BH gets threaded). So how about this patch?
I didn't hear any objections so here is the patch again.
[SNMP]: Fix SNMP counters with PREEMPT
The SNMP macros use raw_smp_processor_id() in process context which is illegal because the process may be preempted and then migrated to another CPU.

Quoted 5 lines

This patch makes it use get_cpu/put_cpu to disable preemption.
Signed-off-by Herbert Xu <[email protected]>
Cheers,

#define SNMP_INC_STATS_USER(mib, field) \
- (per_cpu_ptr(mib[1], raw_smp_processor_id())->mibs[field]++)
+ do { \
+ per_cpu_ptr(mib[1], get_cpu())->mibs[field]++; \
+ put_cpu(); \
+ } while (0)
#define SNMP_INC_STATS(mib, field) \
(per_cpu_ptr(mib[!in_softirq()], raw_smp_processor_id())->mibs[field]++)

How come you change SNMP_INC_STATS_USER() but not SNMP_INC_STATS() ?

Herbert Xu

2007-12-16 02:30 UTC #15 lore raw

[SNMP]: Fix SNMP counters with PREEMPT

Quoted 2 lines · Eric Dumazet

How come you change SNMP_INC_STATS_USER() but not SNMP_INC_STATS() ?

Heh, my brain must have blocked me from seeing it because it's too hard :)

Let's fix it the stupid way first and I'll do a local_t conversion later.

[SNMP]: Fix SNMP counters with PREEMPT

The SNMP macros use raw_smp_processor_id() in process context which is illegal because the process may be preempted and then migrated to another CPU.

This patch makes it use get_cpu/put_cpu to disable preemption.

Signed-off-by Herbert Xu <[email protected]>

Thanks,

Signature · 5 lines

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
--

include/net/snmp.h +17 -4

index 9c5793d..fbb6666 100644
@@ -23,6 +23,7 @@
 
 #include <linux/cache.h>
 #include <linux/snmp.h>
+#include <linux/smp.h>
 
 /*
  * Mibs are stored in array of unsigned long.

@@ -135,14 +136,26 @@ struct linux_mib {
 #define SNMP_INC_STATS_BH(mib, field) 	\
 	(per_cpu_ptr(mib[0], raw_smp_processor_id())->mibs[field]++)
 #define SNMP_INC_STATS_USER(mib, field) \
-	(per_cpu_ptr(mib[1], raw_smp_processor_id())->mibs[field]++)
+	do { \
+		per_cpu_ptr(mib[1], get_cpu())->mibs[field]++; \
+		put_cpu(); \
+	} while (0)
 #define SNMP_INC_STATS(mib, field) 	\
-	(per_cpu_ptr(mib[!in_softirq()], raw_smp_processor_id())->mibs[field]++)
+	do { \
+		per_cpu_ptr(mib[!in_softirq()], get_cpu())->mibs[field]++; \
+		put_cpu(); \
+	} while (0)
 #define SNMP_DEC_STATS(mib, field) 	\
-	(per_cpu_ptr(mib[!in_softirq()], raw_smp_processor_id())->mibs[field]--)
+	do { \
+		per_cpu_ptr(mib[!in_softirq()], get_cpu())->mibs[field]--; \
+		put_cpu(); \
+	} while (0)
 #define SNMP_ADD_STATS_BH(mib, field, addend) 	\
 	(per_cpu_ptr(mib[0], raw_smp_processor_id())->mibs[field] += addend)
 #define SNMP_ADD_STATS_USER(mib, field, addend) 	\
-	(per_cpu_ptr(mib[1], raw_smp_processor_id())->mibs[field] += addend)
+	do { \
+		per_cpu_ptr(mib[1], get_cpu())->mibs[field] += addend; \
+		put_cpu(); \
+	} while (0)
 
 #endif

David Miller

maintainer

2007-12-20 12:13 UTC #16 lore raw

Re: [SNMP]: Fix SNMP counters with PREEMPT

From: Herbert Xu <[email protected]>
Date: Sun, 16 Dec 2007 10:30:25 +0800

Quoted 12 lines · Eric Dumazet

Quoted 2 lines
How come you change SNMP_INC_STATS_USER() but not SNMP_INC_STATS() ?
Heh, my brain must have blocked me from seeing it because it's too hard :)
Let's fix it the stupid way first and I'll do a local_t conversion later.
[SNMP]: Fix SNMP counters with PREEMPT

Quoted 7 lines

The SNMP macros use raw_smp_processor_id() in process context which is illegal because the process may be preempted and then migrated to another CPU.
This patch makes it use get_cpu/put_cpu to disable preemption.
Signed-off-by Herbert Xu <[email protected]>

I just noticed this and replaced the other SNMP fix patch with this one.

Ingo Molnar

2007-12-15 18:44 UTC #17 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

In reply to Herbert Xu

Quoted 6 lines · Herbert Xu

Quoted 4 lines · Herbert Xu
Never mind, we already have that in local_t and as Alexey correctly points out, USER is still going to be the expensive variant with the preempt_disable (well until BH gets threaded). So how about this patch?

Quoted 7 lines

I didn't hear any objections so here is the patch again.
[SNMP]: Fix SNMP counters with PREEMPT
The SNMP macros use raw_smp_processor_id() in process context which is illegal because the process may be preempted and then migrated to another CPU.

nit: please use 'invalid' instead of 'illegal'.

Quoted 5 lines

This patch makes it use get_cpu/put_cpu to disable preemption.
Signed-off-by Herbert Xu <[email protected]>
(quoted diff) +0 -1
-	(per_cpu_ptr(mib[1], raw_smp_processor_id())->mibs[field]++)

Review

Quoted 10 lines

(quoted diff) +8 -1

+	do { \
+		per_cpu_ptr(mib[1], get_cpu())->mibs[field]++; \
+		put_cpu(); \
+	} while (0)
 
-	(per_cpu_ptr(mib[1], raw_smp_processor_id())->mibs[field] += addend)
+	do { \
+		per_cpu_ptr(mib[1], get_cpu())->mibs[field] += addend; \
+		put_cpu(); \
+	} while (0)

we could perhaps introduce stat_smp_processor_id(), which signals that the CPU id is used for statistical purposes and does not have to be exact? In any case, your patch looks good too.

Ingo

Herbert Xu

2007-12-16 02:36 UTC #18 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Quoted 4 lines · Ingo Molnar

we could perhaps introduce stat_smp_processor_id(), which signals that the CPU id is used for statistical purposes and does not have to be exact? In any case, your patch looks good too.

Unfortunately that doesn't work because we can then have two CPUs trying to update the same counter which may corrupt it.

However, an optimisation is indeed possible with some work. If we can get the address of the per-cpu counter against some sort of a per-cpu base pointer, e.g., %gs on x86, then we can do

	incq	%gs:(%rax)

where %rax would be the offset with %gs as the base. This would obviate the need for the CPU ID and therefore avoid disabling preemption.

Hmm, wasn't Christoph working on something like that?

Cheers,

Signature · 4 lines

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Ingo Molnar

2007-12-16 08:58 UTC #19 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Quoted 8 lines · Herbert Xu

Quoted 4 lines · Ingo Molnar
we could perhaps introduce stat_smp_processor_id(), which signals that the CPU id is used for statistical purposes and does not have to be exact? In any case, your patch looks good too.
Unfortunately that doesn't work because we can then have two CPUs trying to update the same counter which may corrupt it.

ah, indeed. I missed that correctness aspect of your patch. Good catch!

Ingo

Christoph Lameter

2007-12-17 19:42 UTC #20 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

In reply to Herbert Xu

Quoted 11 lines · Herbert Xu

If we can get the address of the per-cpu counter against some sort of a per-cpu base pointer, e.g., %gs on x86, then we can do
	incq	%gs:(%rax)
where %rax would be the offset with %gs as the base. This would obviate the need for the CPU ID and therefore avoid disabling preemption.
Hmm, wasn't Christoph working on something like that?

Yes that is what the cpu alloc patchset implements.

Christoph Lameter

2007-12-17 19:40 UTC #21 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

In reply to Ingo Molnar

The cpu alloc patches also fix this issue one way (disabling preempt) or the other (atomic instruction that does not need disabling of preeemption).

Gerrit Renker

2007-12-18 11:44 UTC #22 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

In reply to Herbert Xu

| 
| I didn't hear any objections so here is the patch again.
| 
| [SNMP]: Fix SNMP counters with PREEMPT
|

I don't feel competent to say whether this is correct, but it seems that this is going a long way towards resolving older problems with the SNMP counters.

What I can say is that a year ago, when doing the UDP/-Lite work, I found that the counters were not accurate, which was due to the double-counting problem that Wang Chen brought up again.

Resolving this issue was stalled at that time due to the fact that a counter may be incremented on one CPU and then later decremented on another.

It looks as if this work is reaching the cause of the problem, which would be good since the problem propagates to all users of such counters (UDP, UDP-Lite, and similar MIB counters). So thank you for taking this issue up, I hope that this will lead to a patch set resolving the counter issues.

Best regards,
Gerrit

Herbert Xu

2007-12-18 12:49 UTC #23 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

Quoted 5 lines · Gerrit Renker

It looks as if this work is reaching the cause of the problem, which would be good since the problem propagates to all users of such counters (UDP, UDP-Lite, and similar MIB counters). So thank you for taking this issue up, I hope that this will lead to a patch set resolving the counter issues.

Heh the counter issue is what led to this patch. I think the UDP counters should be fixed already. Please have a look at net-2.6.25 and let us know if it's still broken.

As to the SNMP counter issue in general, Christoph Lameter's new percpu code will greatly simplify this and we won't need to disable preemption or do the fancy softirq/user splitting that we have right now. I eagerly await its acceptance into the Linux kernel.

Cheers,

Signature · 4 lines

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

David Miller

maintainer Applied

2007-12-20 12:12 UTC #24 lore raw

Re: [PATCH 3/3] [UDP6]: Counter increment on BH mode

In reply to Herbert Xu

From: Herbert Xu <[email protected]>
Date: Sat, 15 Dec 2007 21:58:52 +0800

Quoted 9 lines

[SNMP]: Fix SNMP counters with PREEMPT
The SNMP macros use raw_smp_processor_id() in process context which is illegal because the process may be preempted and then migrated to another CPU.
This patch makes it use get_cpu/put_cpu to disable preemption.
Signed-off-by Herbert Xu <[email protected]>

Applied to net-2.6.25, thanks.

Gerrit Renker

Acked-by on v1 1/3

2007-11-30 10:58 UTC #25 lore raw

In reply to Wang Chen

Thanks again for all the work.

| UdpInDatagrams can be confusing because it counts packets that 
| might be dropped later.
| Move UdpInDatagrams into recvmsg() as allowed by the RFC.
| 
| Signed-off-by: Wang Chen <[email protected]>
| ---

Acked-by Gerrit Renker <[email protected]>