diff mbox series

[libnetfilter_queue,v2,1/1] Eliminate packet copy when constructing struct pkt_buff

Message ID 20210518030848.17694-2-duncan_roe@optusnet.com.au
State Superseded, archived
Delegated to: Pablo Neira
Headers show
Series Speed-up | expand

Commit Message

Duncan Roe May 18, 2021, 3:08 a.m. UTC 
To avoid a copy, the new code takes advantage of the fact that the netfilter
netlink queue never returns multipart messages.
This means that the buffer space following that callback data is available for
packet expansion when mangling.

nfq_cb_run() is a new nfq-specific callback runqueue for netlink messages.
The principal function of nfq_cb_run() is to pass to the called function what is
the length of free space after the packet.
As a side benefit, nfq_cb_run() also gives the called functio a pointer to a
zeroised struct pkt_buff, avoiding the malloc / free that was previously needed.

nfq_cb_t is a new typedef for the function called by nfq_cb_run()
[c.f. mnl_cb_t / mnl_cb_run].

No doxygen documentation update yet -
will do once function prototypes are agreed.
In the meantime, examples/nf-queue.c is updated to demonstrate the new API.

Signed-off-by: Duncan Roe <duncan_roe@optusnet.com.au>
---
 examples/nf-queue.c                    | 22 ++++++-
 include/libnetfilter_queue/Makefile.am |  1 +
 include/libnetfilter_queue/callback.h  | 11 ++++
 include/libnetfilter_queue/pktbuff.h   |  2 +
 src/Makefile.am                        |  1 +
 src/extra/callback.c                   | 60 +++++++++++++++++++
 src/extra/pktbuff.c                    | 80 ++++++++++++++++++--------
 7 files changed, 149 insertions(+), 28 deletions(-)
 create mode 100644 include/libnetfilter_queue/callback.h
 create mode 100644 src/extra/callback.c

Comments

Pablo Neira Ayuso May 27, 2021, 8:23 p.m. UTC | #1 
On Tue, May 18, 2021 at 01:08:48PM +1000, Duncan Roe wrote:
> To avoid a copy, the new code takes advantage of the fact that the netfilter
> netlink queue never returns multipart messages.
> This means that the buffer space following that callback data is available for
> packet expansion when mangling.
> 
> nfq_cb_run() is a new nfq-specific callback runqueue for netlink messages.
> The principal function of nfq_cb_run() is to pass to the called function what is
> the length of free space after the packet.
> As a side benefit, nfq_cb_run() also gives the called functio a pointer to a
> zeroised struct pkt_buff, avoiding the malloc / free that was previously needed.
> 
> nfq_cb_t is a new typedef for the function called by nfq_cb_run()
> [c.f. mnl_cb_t / mnl_cb_run].

Interesting idea: let me get back to you with a proposal based on this
patch.

Meanwhile, I have pushed out the __pktb_setup() function which is
going to be needed:

http://git.netfilter.org/libnetfilter_queue/commit/?id=710f891c8a6116f520948f5cf448489947fb7d78

Thanks.
Duncan Roe July 18, 2021, 4:28 a.m. UTC | #2 
On Thu, May 27, 2021 at 10:23:15PM +0200, Pablo Neira Ayuso wrote:
> On Tue, May 18, 2021 at 01:08:48PM +1000, Duncan Roe wrote:
> > To avoid a copy, the new code takes advantage of the fact that the netfilter
> > netlink queue never returns multipart messages.
> > This means that the buffer space following that callback data is available for
> > packet expansion when mangling.
> >
> > nfq_cb_run() is a new nfq-specific callback runqueue for netlink messages.
> > The principal function of nfq_cb_run() is to pass to the called function what is
> > the length of free space after the packet.
> > As a side benefit, nfq_cb_run() also gives the called functio a pointer to a
> > zeroised struct pkt_buff, avoiding the malloc / free that was previously needed.
> >
> > nfq_cb_t is a new typedef for the function called by nfq_cb_run()
> > [c.f. mnl_cb_t / mnl_cb_run].
>
> Interesting idea: let me get back to you with a proposal based on this
> patch.
>
[...]

It occurred to me there is no real need to use a callback any more.

However, mnl_cb_run() does some checks before and after invoking the cb.
Some of these checks may still be valid, so leave it as_is?

This patch has been on the table for a while, any idea when you might find time
to respond?

Cheers ... Duncan.
Duncan Roe Aug. 4, 2021, 1:38 a.m. UTC | #3 
Hi Pablo,

On Thu, May 27, 2021 at 10:23:15PM +0200, Pablo Neira Ayuso wrote:
> On Tue, May 18, 2021 at 01:08:48PM +1000, Duncan Roe wrote:
> > To avoid a copy, the new code takes advantage of the fact that the netfilter
> > netlink queue never returns multipart messages.
[...]
>
> Interesting idea: let me get back to you with a proposal based on this
> patch.
>
> Meanwhile, I have pushed out the __pktb_setup() function which is
> going to be needed:
>
> http://git.netfilter.org/libnetfilter_queue/commit/?id=710f891c8a6116f520948f5cf448489947fb7d78
>
> Thanks.

It also occurred to me to wonder what is the benefit of having struct pkt_buff
be opaque? It's never going to have a buffer tacked on the end of it any more,
so can simply be declared to be sizeof(struct pkt_buff).

Users could read the values of struct members directly rather than having to
learn and use the current procedural interface. That would have to use less
instructions to achieve, but I have yet to benchmark to see if the improvement
is measureable.

We could document when (if ever) the structure may be written to directly but
even if developers break the rules, what damage can they do? This is a userspace
program: they're not going to crash the kernel.

I sidestepped this question in the code by passing down pktb_instance from
local_cb().

Regardless of the above, do you think you might have a proposal for me some time
soon?

Cheers ... Duncan.
diff mbox series

Patch

diff --git a/examples/nf-queue.c b/examples/nf-queue.c
index 3da2c24..15c0d77 100644
--- a/examples/nf-queue.c
+++ b/examples/nf-queue.c
@@ -13,6 +13,8 @@ 
 #include <linux/types.h>
 #include <linux/netfilter/nfnetlink_queue.h>
 
+#include <libnetfilter_queue/pktbuff.h>
+#include <libnetfilter_queue/callback.h>
 #include <libnetfilter_queue/libnetfilter_queue.h>
 
 /* only for NFQA_CT, not needed otherwise: */
@@ -46,13 +48,17 @@  nfq_send_verdict(int queue_num, uint32_t id)
 	}
 }
 
-static int queue_cb(const struct nlmsghdr *nlh, void *data)
+static int queue_cb(const struct nlmsghdr *nlh, void *data,
+		    struct pkt_buff *supplied_pktb, size_t supplied_extra)
 {
 	struct nfqnl_msg_packet_hdr *ph = NULL;
 	struct nlattr *attr[NFQA_MAX+1] = {};
+	struct pkt_buff *pktb;
 	uint32_t id = 0, skbinfo;
 	struct nfgenmsg *nfg;
+	uint8_t *payload;
 	uint16_t plen;
+	int i;
 
 	if (nfq_nlmsg_parse(nlh, attr) < 0) {
 		perror("problems parsing");
@@ -69,7 +75,10 @@  static int queue_cb(const struct nlmsghdr *nlh, void *data)
 	ph = mnl_attr_get_payload(attr[NFQA_PACKET_HDR]);
 
 	plen = mnl_attr_get_payload_len(attr[NFQA_PAYLOAD]);
-	/* void *payload = mnl_attr_get_payload(attr[NFQA_PAYLOAD]); */
+	payload = mnl_attr_get_payload(attr[NFQA_PAYLOAD]);
+
+	pktb = pktb_populate(supplied_pktb, AF_INET, payload, plen,
+			     supplied_extra);
 
 	skbinfo = attr[NFQA_SKB_INFO] ? ntohl(mnl_attr_get_u32(attr[NFQA_SKB_INFO])) : 0;
 
@@ -97,6 +106,13 @@  static int queue_cb(const struct nlmsghdr *nlh, void *data)
 		printf(", checksum not ready");
 	puts(")");
 
+	printf("payload (len=%d) [", plen);
+
+	for (i = 0; i < pktb_len(pktb); i++)
+		printf("%02x", payload[i]);
+
+	printf("]\n");
+
 	nfq_send_verdict(ntohs(nfg->res_id), id);
 
 	return MNL_CB_OK;
@@ -168,7 +184,7 @@  int main(int argc, char *argv[])
 			exit(EXIT_FAILURE);
 		}
 
-		ret = mnl_cb_run(buf, ret, 0, portid, queue_cb, NULL);
+		ret = nfq_cb_run(buf, ret, sizeof_buf, portid, queue_cb, NULL);
 		if (ret < 0){
 			perror("mnl_cb_run");
 			exit(EXIT_FAILURE);
diff --git a/include/libnetfilter_queue/Makefile.am b/include/libnetfilter_queue/Makefile.am
index e436bab..ed6524b 100644
--- a/include/libnetfilter_queue/Makefile.am
+++ b/include/libnetfilter_queue/Makefile.am
@@ -5,4 +5,5 @@  pkginclude_HEADERS = libnetfilter_queue.h	\
 		     libnetfilter_queue_ipv6.h	\
 		     libnetfilter_queue_tcp.h	\
 		     libnetfilter_queue_udp.h	\
+		     callback.h			\
 		     pktbuff.h
diff --git a/include/libnetfilter_queue/callback.h b/include/libnetfilter_queue/callback.h
new file mode 100644
index 0000000..27bfe31
--- /dev/null
+++ b/include/libnetfilter_queue/callback.h
@@ -0,0 +1,11 @@ 
+#ifndef _LIBNETFILTER_QUEUE_CALLBACK_H_
+#define _LIBNETFILTER_QUEUE_CALLBACK_H_
+
+struct nlattr;
+struct pkt_buff;
+
+typedef int (*nfq_cb_t)(const struct nlmsghdr *nlh, void *data, struct pkt_buff *pktb, size_t extra);
+
+int nfq_cb_run(const void *buf, size_t buflen, size_t bufcap, unsigned int portid, nfq_cb_t cb_data, void *data);
+
+#endif
diff --git a/include/libnetfilter_queue/pktbuff.h b/include/libnetfilter_queue/pktbuff.h
index 42bc153..33829cc 100644
--- a/include/libnetfilter_queue/pktbuff.h
+++ b/include/libnetfilter_queue/pktbuff.h
@@ -6,6 +6,8 @@  struct pkt_buff;
 struct pkt_buff *pktb_alloc(int family, void *data, size_t len, size_t extra);
 void pktb_free(struct pkt_buff *pktb);
 
+struct pkt_buff *pktb_populate(struct pkt_buff *pktb, int family, void *data, size_t len, size_t extra);
+
 uint8_t *pktb_data(struct pkt_buff *pktb);
 uint32_t pktb_len(struct pkt_buff *pktb);
 
diff --git a/src/Makefile.am b/src/Makefile.am
index 079853e..376bc45 100644
--- a/src/Makefile.am
+++ b/src/Makefile.am
@@ -31,6 +31,7 @@  libnetfilter_queue_la_LDFLAGS = -Wc,-nostartfiles \
 libnetfilter_queue_la_SOURCES = libnetfilter_queue.c	\
 				nlmsg.c			\
 				extra/checksum.c	\
+				extra/callback.c	\
 				extra/icmp.c		\
 				extra/ipv6.c		\
 				extra/tcp.c		\
diff --git a/src/extra/callback.c b/src/extra/callback.c
new file mode 100644
index 0000000..2d67848
--- /dev/null
+++ b/src/extra/callback.c
@@ -0,0 +1,60 @@ 
+/*
+ * (C) 2021 by Duncan Roe <duncan_roe@optusnet.com.au>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU Lesser General Public License as published
+ * by the Free Software Foundation; either version 2.1 of the License, or
+ * (at your option) any later version.
+ */
+
+#include <errno.h>
+#include <stdlib.h>
+#include <libmnl/libmnl.h>
+#include <linux/netlink.h>
+#include <libnetfilter_queue/callback.h>
+#include <libnetfilter_queue/libnetfilter_queue.h>
+#include "internal.h"
+
+/* ---------------------------------------------------------------------- */
+/* It would be less code to have local_cb() declared within nfq_cb_run(). */
+/* gcc is fine with that; unfortunately clang 8.0.1 is not.               */
+/* Lexical scoping would have given access to the 3 outer stack variables */
+/* needed by local_cb(); absent that, pass them down through the data arg */
+/* ---------------------------------------------------------------------- */
+
+typedef struct qwerty
+{
+	nfq_cb_t cb_func;
+	size_t buflen;
+	size_t bufcap;
+	void *data;
+} werty;
+
+static int local_cb(const struct nlmsghdr *nlh, void *data)
+{
+	struct pkt_buff pktb_instance = { };
+	werty *w = (werty *)data;
+
+	return w->cb_func(nlh, w->data, &pktb_instance, w->bufcap - w->buflen);
+}
+
+EXPORT_SYMBOL int nfq_cb_run(const void *buf, size_t buflen, size_t bufcap,
+			     unsigned int portid, nfq_cb_t cb_func, void *data)
+{
+	const struct nlmsghdr *nlh = buf;
+	werty wert;
+
+	wert.cb_func = cb_func;
+	wert.bufcap = bufcap;
+	wert.buflen = buflen;
+	wert.data = data;
+
+	/* Verify not multi-part */
+	if (nlh->nlmsg_flags & NLM_F_MULTI) {
+		errno = E2BIG;
+		return MNL_CB_ERROR;
+	}
+
+
+	return mnl_cb_run(buf, buflen, 0, portid, local_cb, &wert);
+}
diff --git a/src/extra/pktbuff.c b/src/extra/pktbuff.c
index 6dd0ca9..df23335 100644
--- a/src/extra/pktbuff.c
+++ b/src/extra/pktbuff.c
@@ -19,6 +19,8 @@ 
 #include <netinet/tcp.h>
 
 #include "internal.h"
+#include <libnetfilter_queue/pktbuff.h> /* I.e. local copy */
+					/* (to verify prototypes) */
 
 /**
  * \defgroup pktbuff User-space network packet buffer
@@ -29,6 +31,44 @@ 
  * @{
  */
 
+static int __pktb_setup(int family, struct pkt_buff *pktb)
+{
+	struct ethhdr *ethhdr;
+
+	switch (family) {
+	case AF_INET:
+	case AF_INET6:
+		pktb->network_header = pktb->data;
+		break;
+	case AF_BRIDGE:
+		ethhdr = (struct ethhdr *)pktb->data;
+		pktb->mac_header = pktb->data;
+
+		switch(ethhdr->h_proto) {
+		case ETH_P_IP:
+		case ETH_P_IPV6:
+			pktb->network_header = pktb->data + ETH_HLEN;
+			break;
+		default:
+			/* This protocol is unsupported. */
+			errno = EPROTONOSUPPORT;
+			return -1;
+		}
+		break;
+	}
+
+	return 0;
+}
+
+static void pktb_setup_metadata(struct pkt_buff *pktb, void *pkt_data,
+				size_t len, size_t extra)
+{
+	pktb->len = len;
+	pktb->data_len = len + extra;
+
+	pktb->data = pkt_data;
+}
+
 /**
  * pktb_alloc - allocate a new packet buffer
  * \param family Indicate what family. Currently supported families are
@@ -52,7 +92,6 @@  EXPORT_SYMBOL
 struct pkt_buff *pktb_alloc(int family, void *data, size_t len, size_t extra)
 {
 	struct pkt_buff *pktb;
-	struct ethhdr *ethhdr;
 	void *pkt_data;
 
 	pktb = calloc(1, sizeof(struct pkt_buff) + len + extra);
@@ -63,36 +102,27 @@  struct pkt_buff *pktb_alloc(int family, void *data, size_t len, size_t extra)
 	pkt_data = (uint8_t *)pktb + sizeof(struct pkt_buff);
 	memcpy(pkt_data, data, len);
 
-	pktb->len = len;
-	pktb->data_len = len + extra;
+	pktb_setup_metadata(pktb, pkt_data, len, extra);
 
-	pktb->data = pkt_data;
+	if (__pktb_setup(family, pktb) < 0) {
+		free(pktb);
+		return NULL;
+	}
 
-	switch(family) {
-	case AF_INET:
-	case AF_INET6:
-		pktb->network_header = pktb->data;
-		break;
-	case AF_BRIDGE:
-		ethhdr = (struct ethhdr *)pktb->data;
-		pktb->mac_header = pktb->data;
+	return pktb;
+}
 
-		switch(ethhdr->h_proto) {
-		case ETH_P_IP:
-		case ETH_P_IPV6:
-			pktb->network_header = pktb->data + ETH_HLEN;
-			break;
-		default:
-			/* This protocol is unsupported. */
-			errno = EPROTONOSUPPORT;
-			free(pktb);
-			return NULL;
-		}
-		break;
-	}
+EXPORT_SYMBOL
+struct pkt_buff *pktb_populate(struct pkt_buff *pktb, int family, void *data,
+			       size_t len, size_t extra)
+{
+	pktb_setup_metadata(pktb, data, len, extra);
+	if (__pktb_setup(family, pktb) < 0)
+		pktb = NULL;
 	return pktb;
 }
 
+
 /**
  * pktb_data - get pointer to network packet
  * \param pktb Pointer to userspace packet buffer