The reallocated part of the enlarged association array was left
uninitialized which would have resulted in trying to free random
pointers.
This was a theoretical bug because it wasn't possible to register
more than 32 cache types since no netlink family supports that
many individual cache types.
Nevertheless this patch fixes the bug and also reduces the default
size of the allocation table and expandations a bit to reduce the
memory footprint slightly.
The requirement to have the caller provide the socket does not
make much sense. Automatically allocate the socket if none was
provided.
This may also avoid some future abuse of reusing request sockets
for handling notifications.
Also rename cm_handle to cm_sock for clarity (no API change)
Instead, clone it and modify a temporary copy. Although it is not
recommended to use the same socket for requests and to serve a
cache manager, this change might prevent some unwanted side effects
if done so.
I'm using libnl in a program which I give to valgrind
in order to track memory errors / leaks. When my program
exits, it complains about non-freed memory, allocated in
3 places in classid.c, at lines 280, 284 and 289.
It seems related to the module's constructor classid_init
which allocates resources, with no destructor to free it.
The attached patch tries to fix this issue by registering
a destructor which performs the tree liberation at exit.
I found a small bug in libnl, about extended table id ( above 256 ).
Signed-off-by: Romary Sonrier <romary@sonrier.com>
Signed-off-by: Thomas Graf <tgraf@redhat.com>
The function can be used to make a copy of an existing cache. It is very
similar to nl_cache_subset() except that it allows no filtering but
copies every object.
Signed-off-by: Thierry Reding <thierry.reding@avionic-design.de>
Signed-off-by: Thomas Graf <tgraf@redhat.com>
The plug qdisc supports two operations - plug and unplug. When the
qdisc receives a plug ("buffer") command via netlink request,
packets arriving henceforth are buffered until a corresponding unplug
command is received. Depending on the type of unplug ("release_one"
or "release_indefinite"), the queue can be unplugged indefinitely or
selectively.
The plug qdisc allows a user to implement network output buffering
(aka output commit), used commonly in checkpoint based fault tolerance
systems. It also supports a general purpose queue plug/unplug
functionality.
The associated kernel module is available in David Miller's net-next
tree, commit: c3059be16c9ef29c05f0876a9df5fea21f29724f
This patch introduces userspace tools and API, to control the qdisc
via netlink messages.
Signed-off-by: Shriram Rajagopalan <rshriram@cs.ubc.ca>
Signed-off-by: Thomas Graf <tgraf@redhat.com>
the neighbour parsing function was previously not accessible, so
custom callback functions had to handle the decoding itself.
rtnl_neigh_parse is introduced and implemented in much the same way
as rtnl_route_parse.
Signed-off-by: Brett Ciphery <brett.ciphery@windriver.com>
Signed-off-by: Thomas Graf <tgraf@redhat.com>
In a simple test program that queries the source IP for a given
destination address I get a crash in the call to rtnl_link_alloc_cache.
Here is the stack trace (created with version 3.2.4):
Program received signal SIGSEGV, Segmentation fault.
0xb7eb7553 in strlen () from /lib/libc.so.6
(gdb) bt
#0 0xb7eb7553 in strlen () from /lib/libc.so.6
#1 0xb7eb7285 in strdup () from /lib/libc.so.6
#2 0xb7fcc305 in nla_strdup (nla=0x0) at attr.c:1033
#3 0xb7f9c173 in link_msg_parser (ops=0xb7fc2940, who=0x804b330, n=0x804c3e8, pp=0xbffffbf8) at route/link.c:486
#4 0xb7fcd485 in nl_cache_parse (ops=0xb7fc2940, who=0x804b330, nlh=0x804c3e8, params=0xbffffbf8) at cache.c:724
#5 0xb7fcd547 in update_msg_parser (msg=0x804b328, arg=0xbffffbb8) at cache.c:531
#6 0xb7fd1f25 in nl_cb_call (cb=<optimized out>, msg=<optimized out>, type=<optimized out>) at ../include/netlink-local.h:126
#7 recvmsgs (cb=<optimized out>, sk=<optimized out>) at nl.c:729
#8 nl_recvmsgs (sk=0x804b2d0, cb=0x804b368) at nl.c:780
#9 0xb7fcd5fd in __cache_pickup (sk=0x804b2d0, cache=<optimized out>, param=0xbffffbf8) at cache.c:560
#10 0xb7fcd83f in nl_cache_pickup (sk=0x804b2d0, cache=0x804b308) at cache.c:593
#11 0xb7fcd8c8 in nl_cache_refill (sk=0x804b2d0, cache=0x804b308) at cache.c:780
#12 0xb7f9d1fc in rtnl_link_alloc_cache (sk=0x804b2d0, family=4, result=0xbffffcd4) at route/link.c:868
#13 0x08048fd0 in libnl_init (data=<optimized out>) at helper_route.c:60
#14 iproute_get_source (destination=0xbffffeff "127.0.0.1",
source=0xbffffd0f "\b\004c\370\267\364_\370\267\260\224\004\b8\375\377\277e\024\347\267\320\016\377\267\273\224\004\b\364_\370\267\260\224\004\b", source_size=17)
at helper_route.c:105
#15 0x08048e6a in main (argc=2, argv=0xbffffde4) at ip_route_get.c:25
The attached patch (against 3.2.4) solves the problem, fixing something that
looks like a typo. The bug is still present in current Git master.
store type kind in rtnl_link independently. That would allow to use this
value even if type_ops are not present. This allows for example to
create devices of type unknown to libnl.
Signed-off-by: Jiri Pirko <jpirko@redhat.com>
Adds rtnl_link_enslave() / rtnl_link_release() providing a genreic
link enslaving/release API for use with all link types which use
the IFLA_MASTER property.
Only use the MULTIPATH attribute when adding routes with more than one
next hop.
This solves issues with two scenarios:
1. Adding an IPv4 route to a kernel configured without
CONFIG_IP_ROUTE_MULTIPATH=y.
2. Adding an IPv6 route in general, since the MULTIPATH attribute is not
supported there.
Signed-off-by: Michael Altizer <xiche@verizon.net>
Although it has been possible to create bonding devices, enslave and
release using the regular link API. The added API simplifies usage
and hides some of the compatibility logic.
F.e. enslave() and release() will both verify that the master assignment
has in fact been changed and return -NLE_OPNOTSUPP if it did not.
Also the API will make sure to use RTM_NEWLINK or RTM_SETLINK depending
on what is availble.
Examples are provided in src/ as nl-link-enslave.c and nl-link-release.c
It has been a request that multiple libnl versions should be installabe
in parallel.
In order to achieve this, the basename of the library was changed to
libnl-3 which reflects the 3rd generation of libnl APIs. It also means
that release based library versioning is left behind and libtool
versioning is used instead.
Projects using pkgconfig will automatically link against the new library
basename and will not notice a difference.
The SO versioning is based on the glib model:
current := 100 * minor + micro - revision
revision := revision
age := age (number of backwards compatible versions)
As Jan Engelhardt pointed out, pktloc.c is currently licenses under the GPL
rather than the LGPL. This is a result of its previous existance in src/
which is GPL licensed. I missed to change its license when moving it to
lib/. Since I am the only contributor to the code, I am changing the license
hereby.
This patch fixes an unaligned access for IPv6. On systems with strict alignment requirements, the unaligned access will either result in garbage data or a crash.
Now the return value in case group id is not find is 0. Change it to
return -NLE_OBJ_NOTFOUND which makes more sense. This should not break
anything because genl_ctrl_grp_by_name is static and called only from
genl_ctrl_resolve_grp. genl_ctrl_resolve_grp already might return
-NLE_OBJ_NOTFOUND.
Signed-off-by: Jiri Pirko <jpirko@redhat.com>
I've found a bug in the following scenario (fragment of code):
while (1) {
struct nl_sock *sk = nl_socket_alloc();
if (sk == NULL) {
fprintf(stderr, "Failed to allocate nl socket\n");
break;
}
nl_socket_set_local_port(sk, 0);
nl_socket_free(sk);
}
The problem is that nl_socket_set_local_port(, 0) does not
release local port if it is allocated before.
This patch has been in Fedora and RHEL for a while. It adds a mutex
to protect the port map from concurrent thread accesses.
Original patch from Stefan Berger <stefanb@us.ibm.com>.
Modified to use configure.in to check for libpthread
stores rtnl_link object in address if cache is availble. Provide access
via rtnl_addr_get_link() and rtnl_addr_set_link().
Add rtnl_addr_get() which searches a address cache for an address
matching ifindex and local address.
OK i found the bug, is is the patch !!
Can someone push it into the git tree ?
Regards,
Romary Sonrier
=========================
My analysis:
I guess that the pay load is *struct fib_rule_hdr*
struct fib_rule_hdr {
__u8 family;
__u8 dst_len;
__u8 src_len;
__u8 tos;
__u8 table;
__u8 res1; /* reserved */
__u8 res2; /* reserved */
__u8 action;
__u32 flags;
};
> [PAYLOAD] 12 octets
> 02 00 00 00 0a 00 00 01 00 00 00 00 ............
so
family = 02 , fine
table = 10, fine
but
src_len =0 !!!! should be 32
Adds rtnl_tc_get_link() returning the link associated with a tc
object.
Don't clone link associated with tc object when cloning the tc
object, refer to same link instead and bump refcnt.
Fix refcnt leak when assigning link in msg parser.
- avoid unncessary name change requests
The kernel does not check if the specified IFNAME is different
from the current name. It assumes that if IFNAME and ifindex
are both specified, a name change is requested. Therefore avoid
specyfing IFNAME if ifindex is provided and original and new
name are identical.
- move link building to own function (to allow link add later on)
- error if immutable changes have been made
- better documentation
Provide nl_pickup() to pick up an answer from a netlink request and parse
it using the supplied parser.
Add rtnl_link_get_kernel() which sends an RTM_GETLINK to the kernel to
fetch a single link directly from the kernel. This can be faster than
dumping the whole table, especially if lots of links are configured.
- removed dead functions in header files
- deprecated rtnl_class_foreach_*() functions due to their missing
handling possibility of OOM situations
- improved API documentation
Deprecated the functions rtnl_qdisc_change() and
rtnl_qdisc_build_change_request() for their lack of being able to
specify flags. The new functions rtnl_qdisc_update() and
rtnl_qdisc_build_update_request() may be used instead. The old
functions are still available though. However, rtnl_qdisc_update()
no longer implies NLM_F_REPLACE, it has to specified implicitely
to allow updating a qdisc without risking to replace another qdisc.
Included detailed documentation of qdisc addition/update/deletion.
Introduced APPBUG() macro to let application developer know of
API abuse.
Removing non existent route (both ipv4 and ipv6) using
rtnl_route_delete return NLE_FAILURE generic error code. After
investigation, it seems the linux kernel return -ESRCH in that case
(see http://lxr.linux.no/linux+v2.6.38/net/ipv6/route.c#L1367). Same
behaviour apply for removing non existent address. The function
nl_syserr2nlerr (lib/error.c) which is responsible for translating
from kernel error to libnl error doesn't handle ESRCH.
Their usage is not completely safe, it is not possible to handle
the out of memory situation of the allocate filter. It is very
unlikely for this to cause any problem though.
The functions are still accessible but gcc will warn about their
deprecation.
This commit adds some missing files (some header files, the files below
/etc and the bison/flex files) to the distribution tarball to ensure
that libnl can be built from the tarballs created using "make dist".
It also adds some incantations to properly generate the flex and bison
output since the generated output is no longer shipped in the tarball.
This patch fixes out-of-tree build, that is, when building with
$builddir != $srcdir. In such a case, some generated header files are
not found because they live in a different tree and can be fixed by
adding the directories to the compiler's include search path.
Finally got rid of all the qdisc/class/cls code duplication in
the tc module API. The API takes care of allocation/freeing the
tc object specific data.
I hope I got it right this time.
Thomas Graf