- 17 Jan, 2005 19 commits
-
-
Rusty Russell authored
Anton points out that cacheline aligning conntrack entries is a wank. He's right: there's lots of them, and they're currently ~200 bytes. Same with the cargo-cult programming in ipt_hashlimit. Signed-off-by:
Rusty Russell <rusty@rustcorp.com.au> Signed-off-by:
David S. Miller <davem@davemloft.net>
-
Rusty Russell authored
Fairly simple patch to move the 'initialized' NAT bitfield to bits in the 'status' word. This saves the size of a pointer from the connection tracking structure. Signed-off-by:
-
Rusty Russell authored
Rather than calling the sequence adjustment code on every connection which has a helper, we can set a status bit on the conntrack when we change the length of a TCP packet, and use that to indicate that we should call the routine. Signed-off-by:
-
Rusty Russell authored
We keep a pointer from the hash table entry into the connection tracking entry it's a part of. However, there's a spare byte in the hash entry anyway, which we can use to indicate which of the two tuples it is, and the simply use container_of() to access the conntrack. This saves two pointers per connection tracking entry. Signed-off-by:
-
Rusty Russell authored
Original patch and multo bugfixes by Krisztian Kovacs. Now NAT has been simplified, there is only one place to NAT each packet. That means we can intuit what to do by looking at the difference between this packet and the reply we expect, getting rid of the manips[] array in the connection tracking structure, which is 72 bytes. Rework NAT to be based on 'change this packet to make src/dst look like this tuple'. 1) Each protocol's manip_pkt takes a 'struct ip_conntrack_manip', which is half (the source half) of a tuple. Hand the whole desired tuple to the NAT code and have it use the 'maniptype' arg to decide what part to copy. 2) Krisztian points out that we don't need the NAT lock to read the NAT information (or the tuples) as they never change once set, and while being set we have exclusive access. A lock is only needed to deal with only remaining NAT list: the bysource hash. 3) We don't need to rehash for the bysource hash: it depends on the incoming packet, which we can't change. 4) Many NAT functions only need the maniptype they are to perform, not the actual hook, which makes the code clearer. 5) New status bits to indicate what NAT needs to be done. We can always figure it out by inverting the tuple we expect in the other direction and comparing it, but this is faster. 6) Rename 'do_bindings' to 'nat_packet'. 7) ICMP handing is vastly simplified: we unconditionally change to look the way we want. Signed-off-by:
-
Rusty Russell authored
Adrian Bunk's cleanup patch, updated for after all the Rusty patches. The ip_nat_protocol_register/unregister EXPORT_SYMBOLs() stay, as they are used by future patches. Signed-off-by:
Adrian Bunk <bunk@stusta.de> Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> (modified) Signed-off-by:
-
Rusty Russell authored
From: KOVACS Krisztian <hidden@sch.bme.hu> Hi Rusty, Your recent patch which removed the byipsproto hash left some unused code around. The following patch cleans up that. I'm not sure it's correct, but please take a look at it. Signed-off-by: -
Rusty Russell authored
This patch simplifies the code by always having expectation timeouts. Signed-off-by:
-
Rusty Russell authored
Now we've changed expect handling, we can simplify it significantly. 1) struct ip_conntrack_expect only exists until the connection matching it is created. Now NAT is done directly at the time the expectation is matched, we don't need to keep this information around. 2) The term 'master' is used everywhere to mean the connection that expected this connection. The 'master' field in the new connection points straight to the master connection, and holds a reference. 3) There is no direct link from the connection to the expectations it has created: we walk the global list to find them if we need to clean them up. Each expectation holds a reference. 4) The ip_conntrack_expect_tuple_lock is now a proper subset of ip_conntrack_lock, so we can eliminate it. 5) Remove flags from helper: the policy of evicting the oldest expectation seems to be appropriate for everyone. 6) ip_conntrack_expect_find_get() and ip_conntrack_expect_put() are no longer required. 7) Remove reference count from expectations, and don't free when we fail ip_conntrack_expect_related(): have user call ip_conntrack_expect_free(). Signed-off-by:
-
Rusty Russell authored
Fixes up the other helpers for direct conntrack->NAT helper calling. SNMP doesn't really need a conntrack helper, but under this new model, the NAT helper will register at that point anyway: NAT helpers themselves are removed. Signed-off-by:
-
Rusty Russell authored
Currently connection tracking and NAT helper modules for a protocol interact only indirectly (the conntrack module places information in the conntrack structure, which the NAT module pulls out). This leads to several issues: 1) Both modules must know what port to watch, and must match. 2) Identifying the particular packet which created the connection is cumbersome (TCP) or impossible (UDP). 3) The connection tracking code sets up an expectation which the NAT code then has to change. 4) The lack of direct symbol dependencies means we have to contrive one, since they are functionally dependent. Here is the current code flow: FTP CONTROL PACKET: NF_IP_PRE_ROUTING: ip_conntrack_in resolve_normal_ct init_conntrack: sets ct->helper to ip_conntrack_ftp.c:help() ct->help(): if PORT/PASV command: Sets exp->help.exp_ftp_info to tcp seq number of data. ip_conntrack_expect(): expects the connection ip_nat_setup_info: sets ct->nat.info->helper to ip_nat_ftp.c:help() ip_nat_fn: proto->exp_matches_pkt: if packet matches expectation ct->nat.info->helper(): If packet going client->server, and packet data is one in ct_ftp_info: ftp_data_fixup(): ip_conntrack_change_expect(): change the expectation Modify packet contents with new address. NF_IP_POST_ROUTING: ip_nat_fn ct->nat.info->helper(): If packet going server->client, and packet data is one in ct_ftp_info: ftp_data_fixup(): ip_conntrack_change_expect(): change the expectation Modify packet contents with new address. FTP DATA (EXPECTED) CONNECTION FIRST PACKET: NF_IP_PRE_ROUTING: ip_conntrack_in resolve_normal_ct init_conntrack: set ct->master. ip_nat_fn: master->nat.info.helper->expect() Set up source NAT mapping to match FTP control connection. NF_IP_PRE_ROUTING: ip_nat_fn: master->nat.info.helper->expect() Set up dest NAT mapping to match FTP control connection. The new flow looks like this: FTP CONTROL PACKET: NF_IP_PRE_ROUTING: ip_conntrack_in resolve_normal_ct init_conntrack: sets ct->helper to ip_conntrack_ftp.c:help() NF_IP_POST_ROUTING: ip_confirm: ct->helper->help: If !ip_nat_ftp_hook: ip_conntrack_expect(). ip_nat_ftp: set exp->oldproto to old port. ip_conntrack_change_expect(): change the expectation set exp->expectfn to ftp_nat_expected. Modify packet contents with new address. FTP DATA (EXPECTED) CONNECTION FIRST PACKET: NF_IP_PRE_ROUTING: ip_conntrack_in resolve_normal_ct init_conntrack: set ct->master. call exp->expectfn (ftp_nat_expected): call ip_nat_follow_master(). The big changes are that the ip_nat_ftp module sets ip_conntrack_ftp's ip_nat_ftp_hook when it initializes, so it calls the NAT code directly when a packet containing the expect information is found by the conntrack helper: and this interface can carry all the information these two want to share. Also, that conntrack helper is called as the packet leaves the box, so there are no issues with expectations being set up before the packet has been filtered. The NAT helper doesn't need to register and duplicate the conntrack ports. The other trick is ip_nat_follow_master(), which does the NAT setup all at once (source and destination NAT as required) such that the expected connection is NATed the same way the master connection was. We also call ip_conntrack_tcp_update() (which I incidentally neatened) after mangling a TCP packet; ip_nat_seq_adjust() does this, but now mangling is done at the last possible moment, after ip_nat_seq_adjust() was already called. Signed-off-by: -
Rusty Russell authored
Change kmem_cache_free() calls in ip_conntrack_expect_related() to ip_conntrack_expect_put(): they should be equivalent but allows a hack in next patch (caller can keep expect). More importantly, a previous expectation should only be refreshed and return EEXIST if it's owned by the same connection (nfsim found this bug). Signed-off-by:
-
David S. Miller authored
Signed-off-by: -
Arthur Kepner authored
Signed-off-by:
Arthur Kepner <akepner@sgi.com> Signed-off-by:
-
Christoph Hellwig authored
- put a dev_id field in struct net_device, so that it uses space that would be wasted by padding otherwise. - if this fields is non-null let ipv6_generate_eui64 use the algorithm from the QETH code to generate an EUI that's different for each OS instance. See code comments for details. Signed-off-by: -
Thomas Graf authored
Signed-off-by:
Thomas Graf <tgraf@suug.ch> Signed-off-by:
-
Herbert Xu authored
This makes the skb->truesize modifications always OK. Signed-off-by:
Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by:
-
bk://kernel.bkbits.net/acme/connection_sock-2.6David S. Miller authored
into nuts.davemloft.net:/disk1/BK/net-2.6
-
bk://bk.skbuff.net:20611/linux-2.6-inet6David S. Miller authored
into nuts.davemloft.net:/disk1/BK/net-2.6
-
- 16 Jan, 2005 21 commits
-
-
Trond Myklebust authored
Signed-off-by:
Trond Myklebust <Trond.Myklebust@netapp.com> Signed-off-by:
Linus Torvalds <torvalds@osdl.org>
-
Trond Myklebust authored
RPC: Fix a module refcount leak in RPCSEC_GSS Signed-off-by:
-
Trond Myklebust authored
Signed-off-by:
-
Trond Myklebust authored
Signed-off-by:
-
bk://bk.arm.linux.org.uk/linux-2.6-smpLinus Torvalds authored
into ppc970.osdl.org:/home/torvalds/v2.6/linux
-
Russell King authored
profile_pc() used thread_saved_fp() with the current task. However, thread_saved_fp() only returns sane values when called for threads which are presently sleeping, so this caused an oops. Instead, use regs->ARM_fp, which correspond with the frame pointer.
-
Pawel Sikora authored
The build is clean now.
-
Hideaki Yoshifuji authored
As NAs do not create new entries (RFC2461 7.2.5), NA should not change state of FAILED entries. Signed-off-by:Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
-
Hideaki Yoshifuji authored
Signed-off-by: -
Roland McGrath authored
God invented symbolic names to help you. Repeating magic constants by hand is begging to lose, especially when you get them wrong. Don't be a loser. [ Editor's hint: 0xfffe000 vs 0xffffe000 ] Signed-off-by:
Roland McGrath <roland@redhat.com> Signed-off-by:
-
Matthew Wilcox authored
Remove unreferenced file Signed-off-by:
Domen Puncer <domen@coderock.org> Signed-off-by:
Grant Grundler <grundler@parisc-linux.org> Signed-off-by:
Matthew Wilcox <willy@parisc-linux.org> Signed-off-by:
-
Matthew Wilcox authored
- sparse annotations for ioremap/iounmap (Randolph Chung) - Turn gsc_readb, __raw_readb and readb functions into static inline functions (Matthew Wilcox) - Document the difference between the gsc_readb, __raw_readb and readb families of functions (Matthew Wilcox) - Add a debugging option to determine when they are being used incorrectly (Matthew Wilcox) - Make memcpy_fromio's second argument const (Matthew Wilcox) Signed-off-by:
-
Matthew Wilcox authored
Fix _syscallN wrappers (Mike Frysinger) Signed-off-by:
-
Matthew Wilcox authored
Fix ptrace(SINGLESTEP) through system call Signed-off-by:
Randolph Chung <tausq@parisc-linux.org> Signed-off-by:
-
Matthew Wilcox authored
Remove some unused definitions Signed-off-by:
-
Matthew Wilcox authored
Add PDC Stable Storage wrappers (Thibaut Varene) Rewrite PDC Initiator (Matthew Wilcox) Signed-off-by:
-
Matthew Wilcox authored
Implement the iomap interfaces on PA-RISC Signed-off-by:
-
Matthew Wilcox authored
- Remove declaration of sys_setpgid - __user annotations - Rewrite hpux_statfs - Add hpux_fstatfs Signed-off-by:
-
Matthew Wilcox authored
- Removes the hardware path from /proc/interrupts for Dino to make it consistant with the rest of /proc/interrupts Signed-off-by: Ryan Bradetich - Remove iomem related warnings from dino.c Signed-off-by:
Kyle McMartin <kyle@parisc-linux.org> - Convert SPIN_LOCK_UNLOCKED to spin_lock_init (Thomas Gleixner) Signed-off-by:
Thomas Gleixner <tglx@linutronix.de> Acked-by:
Ingo Molnar <mingo@elte.hu> Signed-off-by: Matthew Wilcox Signed-off-by:
-
Matthew Wilcox authored
Defconfig updates from Grant Grundler and Paul Bame Signed-off-by:
-
Matthew Wilcox authored
- Remove parent/child/sibling links from parisc_device in favour of the ones in the embedded struct device. - Display irq and device IDs through sysfs - Translate a PA-RISC firmware path into a struct device (Thibaut Varene) Signed-off-by:
-
