After IPv6 support has recently been added to metadata dst and related
encaps, add support for populating/reading it from an eBPF program.
Commit d3aa45ce6b ("bpf: add helpers to access tunnel metadata") started
with initial IPv4-only support back then (due to IPv6 metadata support
not being available yet).
To stay compatible with older programs, we need to test for the passed
structure size. Also TOS and TTL support from the ip_tunnel_info key has
been added. Tested with vxlan devs in collect meta data mode with IPv4,
IPv6 and in compat mode over different network namespaces.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Export flags used by eBPF helper functions through UAPI, so they can be
used by programs (instead of them redefining all flags each time or just
using the hard-coded values). It also gives a better overview what flags
are used where and we can further get rid of the extra macros defined in
filter.c. Moreover, reject invalid flags.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
This work adds a generalization of the ingress qdisc as a qdisc holding
only classifiers. The clsact qdisc works on ingress, but also on egress.
In both cases, it's execution happens without taking the qdisc lock, and
the main difference for the egress part compared to prior version of [1]
is that this can be applied with _any_ underlying real egress qdisc (also
classless ones).
Besides solving the use-case of [1], that is, allowing for more programmability
on assigning skb->priority for the mqprio case that is supported by most
popular 10G+ NICs, it also opens up a lot more flexibility for other tc
applications. The main work on classification can already be done at clsact
egress time if the use-case allows and state stored for later retrieval
f.e. again in skb->priority with major/minors (which is checked by most
classful qdiscs before consulting tc_classify()) and/or in other skb fields
like skb->tc_index for some light-weight post-processing to get to the
eventual classid in case of a classful qdisc. Another use case is that
the clsact egress part allows to have a central egress counterpart to
the ingress classifiers, so that classifiers can easily share state (e.g.
in cls_bpf via eBPF maps) for ingress and egress.
Currently, default setups like mq + pfifo_fast would require for this to
use, for example, prio qdisc instead (to get a tc_classify() run) and to
duplicate the egress classifier for each queue. With clsact, it allows
for leaving the setup as is, it can additionally assign skb->priority to
put the skb in one of pfifo_fast's bands and it can share state with maps.
Moreover, we can access the skb's dst entry (f.e. to retrieve tclassid)
w/o the need to perform a skb_dst_force() to hold on to it any longer. In
lwt case, we can also use this facility to setup dst metadata via cls_bpf
(bpf_skb_set_tunnel_key()) without needing a real egress qdisc just for
that (case of IFF_NO_QUEUE devices, for example).
The realization can be done without any changes to the scheduler core
framework. All it takes is that we have two a-priori defined minors/child
classes, where we can mux between ingress and egress classifier list
(dev->ingress_cl_list and dev->egress_cl_list, latter stored close to
dev->_tx to avoid extra cacheline miss for moderate loads). The egress
part is a bit similar modelled to handle_ing() and patched to a noop in
case the functionality is not used. Both handlers are now called
sch_handle_ingress() and sch_handle_egress(), code sharing among the two
doesn't seem practical as there are various minor differences in both
paths, so that making them conditional in a single handler would rather
slow things down.
Full compatibility to ingress qdisc is provided as well. Since both
piggyback on TC_H_CLSACT, only one of them (ingress/clsact) can exist
per netdevice, and thus ingress qdisc specific behaviour can be retained
for user space. This means, either a user does 'tc qdisc add dev foo ingress'
and configures ingress qdisc as usual, or the 'tc qdisc add dev foo clsact'
alternative, where both, ingress and egress classifier can be configured
as in the below example. ingress qdisc supports attaching classifier to any
minor number whereas clsact has two fixed minors for muxing between the
lists, therefore to not break user space setups, they are better done as
two separate qdiscs.
I decided to extend the sch_ingress module with clsact functionality so
that commonly used code can be reused, the module is being aliased with
sch_clsact so that it can be auto-loaded properly. Alternative would have been
to add a flag when initializing ingress to alter its behaviour plus aliasing
to a different name (as it's more than just ingress). However, the first would
end up, based on the flag, choosing the new/old behaviour by calling different
function implementations to handle each anyway, the latter would require to
register ingress qdisc once again under different alias. So, this really begs
to provide a minimal, cleaner approach to have Qdisc_ops and Qdisc_class_ops
by its own that share callbacks used by both.
Example, adding qdisc:
# tc qdisc add dev foo clsact
# tc qdisc show dev foo
qdisc mq 0: root
qdisc pfifo_fast 0: parent :1 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
qdisc pfifo_fast 0: parent :2 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
qdisc pfifo_fast 0: parent :3 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
qdisc pfifo_fast 0: parent :4 bands 3 priomap 1 2 2 2 1 2 0 0 1 1 1 1 1 1 1 1
qdisc clsact ffff: parent ffff:fff1
Adding filters (deleting, etc works analogous by specifying ingress/egress):
# tc filter add dev foo ingress bpf da obj bar.o sec ingress
# tc filter add dev foo egress bpf da obj bar.o sec egress
# tc filter show dev foo ingress
filter protocol all pref 49152 bpf
filter protocol all pref 49152 bpf handle 0x1 bar.o:[ingress] direct-action
# tc filter show dev foo egress
filter protocol all pref 49152 bpf
filter protocol all pref 49152 bpf handle 0x1 bar.o:[egress] direct-action
A 'tc filter show dev foo' or 'tc filter show dev foo parent ffff:' will
show an empty list for clsact. Either using the parent names (ingress/egress)
or specifying the full major/minor will then show the related filter lists.
Prior work on a mqprio prequeue() facility [1] was done mainly by John Fastabend.
[1] http://patchwork.ozlabs.org/patch/512949/
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: John Fastabend <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Add a small helper skb_postpush_rcsum() and fix up redirect locations
that need CHECKSUM_COMPLETE fixups on ingress. dev_forward_skb() expects
a proper csum that covers also Ethernet header, f.e. since 2c26d34bbc
("net/core: Handle csum for CHECKSUM_COMPLETE VXLAN forwarding"), we
also do skb_postpull_rcsum() after pulling Ethernet header off via
eth_type_trans().
When using eBPF in a netns setup f.e. with vxlan in collect metadata mode,
I can trigger the following csum issue with an IPv6 setup:
[ 505.144065] dummy1: hw csum failure
[...]
[ 505.144108] Call Trace:
[ 505.144112] <IRQ> [<ffffffff81372f08>] dump_stack+0x44/0x5c
[ 505.144134] [<ffffffff81607cea>] netdev_rx_csum_fault+0x3a/0x40
[ 505.144142] [<ffffffff815fee3f>] __skb_checksum_complete+0xcf/0xe0
[ 505.144149] [<ffffffff816f0902>] nf_ip6_checksum+0xb2/0x120
[ 505.144161] [<ffffffffa08c0e0e>] icmpv6_error+0x17e/0x328 [nf_conntrack_ipv6]
[ 505.144170] [<ffffffffa0898eca>] ? ip6t_do_table+0x2fa/0x645 [ip6_tables]
[ 505.144177] [<ffffffffa08c0725>] ? ipv6_get_l4proto+0x65/0xd0 [nf_conntrack_ipv6]
[ 505.144189] [<ffffffffa06c9a12>] nf_conntrack_in+0xc2/0x5a0 [nf_conntrack]
[ 505.144196] [<ffffffffa08c039c>] ipv6_conntrack_in+0x1c/0x20 [nf_conntrack_ipv6]
[ 505.144204] [<ffffffff8164385d>] nf_iterate+0x5d/0x70
[ 505.144210] [<ffffffff816438d6>] nf_hook_slow+0x66/0xc0
[ 505.144218] [<ffffffff816bd302>] ipv6_rcv+0x3f2/0x4f0
[ 505.144225] [<ffffffff816bca40>] ? ip6_make_skb+0x1b0/0x1b0
[ 505.144232] [<ffffffff8160b77b>] __netif_receive_skb_core+0x36b/0x9a0
[ 505.144239] [<ffffffff8160bdc8>] ? __netif_receive_skb+0x18/0x60
[ 505.144245] [<ffffffff8160bdc8>] __netif_receive_skb+0x18/0x60
[ 505.144252] [<ffffffff8160ccff>] process_backlog+0x9f/0x140
[ 505.144259] [<ffffffff8160c4a5>] net_rx_action+0x145/0x320
[...]
What happens is that on ingress, we push Ethernet header back in, either
from cls_bpf or right before skb_do_redirect(), but without updating csum.
The "hw csum failure" can be fixed by using the new skb_postpush_rcsum()
helper for the dev_forward_skb() case to correct the csum diff again.
Thanks to Hannes Frederic Sowa for the csum_partial() idea!
Fixes: 3896d655f4 ("bpf: introduce bpf_clone_redirect() helper")
Fixes: 27b29f6305 ("bpf: add bpf_redirect() helper")
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Add a skb_at_tc_ingress() as this will be needed elsewhere as well and
can hide the ugly ifdef.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Move the details behind the cb[] access into a small helper to decouple
and make them generic for bpf_prog_run_save_cb()/bpf_prog_run_clear_cb()
that was introduced via commit ff936a04e5 ("bpf: fix cb access in socket
filter programs"). Also add a comment to better clarify what is done in
bpf_skb_cb().
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Add couple of test cases for interpreter but also JITs, f.e. to test that
when imm32 moves are being done, upper 32bits of the regs are being zero
extended.
Without JIT:
[...]
[ 1114.129301] test_bpf: #43 MOV REG64 jited:0 128 PASS
[ 1114.130626] test_bpf: #44 MOV REG32 jited:0 139 PASS
[ 1114.132055] test_bpf: #45 LD IMM64 jited:0 124 PASS
[...]
With JIT (generated code can as usual be nicely verified with the help of
bpf_jit_disasm tool):
[...]
[ 1062.726782] test_bpf: #43 MOV REG64 jited:1 6 PASS
[ 1062.726890] test_bpf: #44 MOV REG32 jited:1 6 PASS
[ 1062.726993] test_bpf: #45 LD IMM64 jited:1 6 PASS
[...]
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
When sometimes structs or variables need to be initialized/'memset' to 0 in
an eBPF C program, the x86 BPF JIT converts this to use immediates. We can
however save a couple of bytes (f.e. even up to 7 bytes on a single emmission
of BPF_LD | BPF_IMM | BPF_DW) in the image by detecting such case and use xor
on the dst register instead.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Comment says "User BPF's register A is mapped to our BPF register 6",
which is actually wrong as the mapping is on register 0. This can
already be inferred from the code itself. So just remove it before
someone makes assumptions based on that. Only code tells truth. ;)
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Back in the days where eBPF (or back then "internal BPF" ;->) was not
exposed to user space, and only the classic BPF programs internally
translated into eBPF programs, we missed the fact that for classic BPF
A and X needed to be cleared. It was fixed back then via 83d5b7ef99
("net: filter: initialize A and X registers"), and thus classic BPF
specifics were added to the eBPF interpreter core to work around it.
This added some confusion for JIT developers later on that take the
eBPF interpreter code as an example for deriving their JIT. F.e. in
f75298f5c3 ("s390/bpf: clear correct BPF accumulator register"), at
least X could leak stack memory. Furthermore, since this is only needed
for classic BPF translations and not for eBPF (verifier takes care
that read access to regs cannot be done uninitialized), more complexity
is added to JITs as they need to determine whether they deal with
migrations or native eBPF where they can just omit clearing A/X in
their prologue and thus reduce image size a bit, see f.e. cde66c2d88
("s390/bpf: Only clear A and X for converted BPF programs"). In other
cases (x86, arm64), A and X is being cleared in the prologue also for
eBPF case, which is unnecessary.
Lets move this into the BPF migration in bpf_convert_filter() where it
actually belongs as long as the number of eBPF JITs are still few. It
can thus be done generically; allowing us to remove the quirk from
__bpf_prog_run() and to slightly reduce JIT image size in case of eBPF,
while reducing code duplication on this matter in current(/future) eBPF
JITs.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Reviewed-by: Michael Holzheu <[email protected]>
Tested-by: Michael Holzheu <[email protected]>
Cc: Zi Shen Lim <[email protected]>
Cc: Yang Shi <[email protected]>
Acked-by: Yang Shi <[email protected]>
Acked-by: Zi Shen Lim <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
When hacking tc programs with eBPF, one of the issues that come up
from time to time is to load addresses from headers. In eBPF as in
classic BPF, we have BPF_LD | BPF_ABS | BPF_{B,H,W} instructions that
extract a byte, half-word or word out of the skb data though helpers
such as bpf_load_pointer() (interpreter case).
F.e. extracting a whole IPv6 address could possibly look like ...
union v6addr {
struct {
__u32 p1;
__u32 p2;
__u32 p3;
__u32 p4;
};
__u8 addr[16];
};
[...]
a.p1 = htonl(load_word(skb, off));
a.p2 = htonl(load_word(skb, off + 4));
a.p3 = htonl(load_word(skb, off + 8));
a.p4 = htonl(load_word(skb, off + 12));
[...]
/* access to a.addr[...] */
This work adds a complementary helper bpf_skb_load_bytes() (we also
have bpf_skb_store_bytes()) as an alternative where the same call
would look like from an eBPF program:
ret = bpf_skb_load_bytes(skb, off, addr, sizeof(addr));
Same verifier restrictions apply as in ffeedafbf0 ("bpf: introduce
current->pid, tgid, uid, gid, comm accessors") case, where stack memory
access needs to be statically verified and thus guaranteed to be
initialized in first use (otherwise verifier cannot tell whether a
subsequent access to it is valid or not as it's runtime dependent).
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Add support for renaming and hard links to the fs. Most of this can be
implemented by using simple library operations under the same constraints
that we don't use a reserved name like elsewhere. Linking can be useful
to share/manage things like maps across subsystem users. It works within
the file system boundary, but is not allowed for directories.
Symbolic links are explicitly not implemented here, as it can be better
done already by doing bind mounts inside bpf fs to set up shared directories
f.e. useful when using volumes in docker containers that map a private
working directory into /sys/fs/bpf/ which contains itself a bind mounted
path from the host's /sys/fs/bpf/ mount that is shared among multiple
containers. For single maps instead of whole directory, hard links can
be easily used to do the same.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
During own review but also reported by Dmitry's syzkaller [1] it has been
noticed that we trigger a heap out-of-bounds access on eBPF array maps
when updating elements. This happens with each map whose map->value_size
(specified during map creation time) is not multiple of 8 bytes.
In array_map_alloc(), elem_size is round_up(attr->value_size, 8) and
used to align array map slots for faster access. However, in function
array_map_update_elem(), we update the element as ...
memcpy(array->value + array->elem_size * index, value, array->elem_size);
... where we access 'value' out-of-bounds, since it was allocated from
map_update_elem() from syscall side as kmalloc(map->value_size, GFP_USER)
and later on copied through copy_from_user(value, uvalue, map->value_size).
Thus, up to 7 bytes, we can access out-of-bounds.
Same could happen from within an eBPF program, where in worst case we
access beyond an eBPF program's designated stack.
Since 1be7f75d16 ("bpf: enable non-root eBPF programs") didn't hit an
official release yet, it only affects priviledged users.
In case of array_map_lookup_elem(), the verifier prevents eBPF programs
from accessing beyond map->value_size through check_map_access(). Also
from syscall side map_lookup_elem() only copies map->value_size back to
user, so nothing could leak.
[1] http://github.com/google/syzkaller
Fixes: 28fbcfa08d ("bpf: add array type of eBPF maps")
Reported-by: Dmitry Vyukov <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Currently, when having map file descriptors pointing to program arrays,
there's still the issue that we unconditionally flush program array
contents via bpf_fd_array_map_clear() in bpf_map_release(). This happens
when such a file descriptor is released and is independent of the map's
refcount.
Having this flush independent of the refcount is for a reason: there
can be arbitrary complex dependency chains among tail calls, also circular
ones (direct or indirect, nesting limit determined during runtime), and
we need to make sure that the map drops all references to eBPF programs
it holds, so that the map's refcount can eventually drop to zero and
initiate its freeing. Btw, a walk of the whole dependency graph would
not be possible for various reasons, one being complexity and another
one inconsistency, i.e. new programs can be added to parts of the graph
at any time, so there's no guaranteed consistent state for the time of
such a walk.
Now, the program array pinning itself works, but the issue is that each
derived file descriptor on close would nevertheless call unconditionally
into bpf_fd_array_map_clear(). Instead, keep track of users and postpone
this flush until the last reference to a user is dropped. As this only
concerns a subset of references (f.e. a prog array could hold a program
that itself has reference on the prog array holding it, etc), we need to
track them separately.
Short analysis on the refcounting: on map creation time usercnt will be
one, so there's no change in behaviour for bpf_map_release(), if unpinned.
If we already fail in map_create(), we are immediately freed, and no
file descriptor has been made public yet. In bpf_obj_pin_user(), we need
to probe for a possible map in bpf_fd_probe_obj() already with a usercnt
reference, so before we drop the reference on the fd with fdput().
Therefore, if actual pinning fails, we need to drop that reference again
in bpf_any_put(), otherwise we keep holding it. When last reference
drops on the inode, the bpf_any_put() in bpf_evict_inode() will take
care of dropping the usercnt again. In the bpf_obj_get_user() case, the
bpf_any_get() will grab a reference on the usercnt, still at a time when
we have the reference on the path. Should we later on fail to grab a new
file descriptor, bpf_any_put() will drop it, otherwise we hold it until
bpf_map_release() time.
Joint work with Alexei.
Fixes: b2197755b2 ("bpf: add support for persistent maps/progs")
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
David and HacKurx reported a following/similar size overflow triggered
in a grsecurity kernel, thanks to PaX's gcc size overflow plugin:
(Already fixed in later grsecurity versions by Brad and PaX Team.)
[ 1002.296137] PAX: size overflow detected in function scm_detach_fds net/core/scm.c:314
cicus.202_127 min, count: 4, decl: msg_controllen; num: 0; context: msghdr;
[ 1002.296145] CPU: 0 PID: 3685 Comm: scm_rights_recv Not tainted 4.2.3-grsec+ #7
[ 1002.296149] Hardware name: Apple Inc. MacBookAir5,1/Mac-66F35F19FE2A0D05, [...]
[ 1002.296153] ffffffff81c27366 0000000000000000 ffffffff81c27375 ffffc90007843aa8
[ 1002.296162] ffffffff818129ba 0000000000000000 ffffffff81c27366 ffffc90007843ad8
[ 1002.296169] ffffffff8121f838 fffffffffffffffc fffffffffffffffc ffffc90007843e60
[ 1002.296176] Call Trace:
[ 1002.296190] [<ffffffff818129ba>] dump_stack+0x45/0x57
[ 1002.296200] [<ffffffff8121f838>] report_size_overflow+0x38/0x60
[ 1002.296209] [<ffffffff816a979e>] scm_detach_fds+0x2ce/0x300
[ 1002.296220] [<ffffffff81791899>] unix_stream_read_generic+0x609/0x930
[ 1002.296228] [<ffffffff81791c9f>] unix_stream_recvmsg+0x4f/0x60
[ 1002.296236] [<ffffffff8178dc00>] ? unix_set_peek_off+0x50/0x50
[ 1002.296243] [<ffffffff8168fac7>] sock_recvmsg+0x47/0x60
[ 1002.296248] [<ffffffff81691522>] ___sys_recvmsg+0xe2/0x1e0
[ 1002.296257] [<ffffffff81693496>] __sys_recvmsg+0x46/0x80
[ 1002.296263] [<ffffffff816934fc>] SyS_recvmsg+0x2c/0x40
[ 1002.296271] [<ffffffff8181a3ab>] entry_SYSCALL_64_fastpath+0x12/0x85
Further investigation showed that this can happen when an *odd* number of
fds are being passed over AF_UNIX sockets.
In these cases CMSG_LEN(i * sizeof(int)) and CMSG_SPACE(i * sizeof(int)),
where i is the number of successfully passed fds, differ by 4 bytes due
to the extra CMSG_ALIGN() padding in CMSG_SPACE() to an 8 byte boundary
on 64 bit. The padding is used to align subsequent cmsg headers in the
control buffer.
When the control buffer passed in from the receiver side *lacks* these 4
bytes (e.g. due to buggy/wrong API usage), then msg->msg_controllen will
overflow in scm_detach_fds():
int cmlen = CMSG_LEN(i * sizeof(int)); <--- cmlen w/o tail-padding
err = put_user(SOL_SOCKET, &cm->cmsg_level);
if (!err)
err = put_user(SCM_RIGHTS, &cm->cmsg_type);
if (!err)
err = put_user(cmlen, &cm->cmsg_len);
if (!err) {
cmlen = CMSG_SPACE(i * sizeof(int)); <--- cmlen w/ 4 byte extra tail-padding
msg->msg_control += cmlen;
msg->msg_controllen -= cmlen; <--- iff no tail-padding space here ...
} ... wrap-around
F.e. it will wrap to a length of 18446744073709551612 bytes in case the
receiver passed in msg->msg_controllen of 20 bytes, and the sender
properly transferred 1 fd to the receiver, so that its CMSG_LEN results
in 20 bytes and CMSG_SPACE in 24 bytes.
In case of MSG_CMSG_COMPAT (scm_detach_fds_compat()), I haven't seen an
issue in my tests as alignment seems always on 4 byte boundary. Same
should be in case of native 32 bit, where we end up with 4 byte boundaries
as well.
In practice, passing msg->msg_controllen of 20 to recvmsg() while receiving
a single fd would mean that on successful return, msg->msg_controllen is
being set by the kernel to 24 bytes instead, thus more than the input
buffer advertised. It could f.e. become an issue if such application later
on zeroes or copies the control buffer based on the returned msg->msg_controllen
elsewhere.
Maximum number of fds we can send is a hard upper limit SCM_MAX_FD (253).
Going over the code, it seems like msg->msg_controllen is not being read
after scm_detach_fds() in scm_recv() anymore by the kernel, good!
Relevant recvmsg() handler are unix_dgram_recvmsg() (unix_seqpacket_recvmsg())
and unix_stream_recvmsg(). Both return back to their recvmsg() caller,
and ___sys_recvmsg() places the updated length, that is, new msg_control -
old msg_control pointer into msg->msg_controllen (hence the 24 bytes seen
in the example).
Long time ago, Wei Yongjun fixed something related in commit 1ac70e7ad2
("[NET]: Fix function put_cmsg() which may cause usr application memory
overflow").
RFC3542, section 20.2. says:
The fields shown as "XX" are possible padding, between the cmsghdr
structure and the data, and between the data and the next cmsghdr
structure, if required by the implementation. While sending an
application may or may not include padding at the end of last
ancillary data in msg_controllen and implementations must accept both
as valid. On receiving a portable application must provide space for
padding at the end of the last ancillary data as implementations may
copy out the padding at the end of the control message buffer and
include it in the received msg_controllen. When recvmsg() is called
if msg_controllen is too small for all the ancillary data items
including any trailing padding after the last item an implementation
may set MSG_CTRUNC.
Since we didn't place MSG_CTRUNC for already quite a long time, just do
the same as in 1ac70e7ad2 to avoid an overflow.
Btw, even man-page author got this wrong :/ See db939c9b26e9 ("cmsg.3: Fix
error in SCM_RIGHTS code sample"). Some people must have copied this (?),
thus it got triggered in the wild (reported several times during boot by
David and HacKurx).
No Fixes tag this time as pre 2002 (that is, pre history tree).
Reported-by: David Sterba <[email protected]>
Reported-by: HacKurx <[email protected]>
Cc: PaX Team <[email protected]>
Cc: Emese Revfy <[email protected]>
Cc: Brad Spengler <[email protected]>
Cc: Wei Yongjun <[email protected]>
Cc: Eric Dumazet <[email protected]>
Reviewed-by: Hannes Frederic Sowa <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Add a handler for show_fdinfo() to be used by the anon-inodes
backend for eBPF maps, and dump the map specification there. Not
only useful for admins, but also it provides a minimal way to
compare specs from ELF vs pinned object.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Acked-by: Hannes Frederic Sowa <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
While recently going over ARM64's BPF code, I noticed that the icache
range we're flushing should start at header already and not at ctx.image.
Reason is that after b569c1c622 ("net: bpf: arm64: address randomize
and write protect JIT code"), we also want to make sure to flush the
random-sized trap in front of the start of the actual program (analogous
to x86). No operational differences from user side.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Zi Shen Lim <[email protected]>
Cc: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
During review I noticed that the icache range we're flushing should
start at header already and not at ctx.image.
Reason is that after 55309dd3d4 ("net: bpf: arm: address randomize
and write protect JIT code"), we also want to make sure to flush the
random-sized trap in front of the start of the actual program (analogous
to x86). No operational differences from user side.
Signed-off-by: Daniel Borkmann <[email protected]>
Tested-by: Nicolas Schichan <[email protected]>
Cc: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Since it's introduction in commit 69e3c75f4d ("net: TX_RING and
packet mmap"), TX_RING could be used from SOCK_DGRAM and SOCK_RAW
side. When used with SOCK_DGRAM only, the size_max > dev->mtu +
reserve check should have reserve as 0, but currently, this is
unconditionally set (in it's original form as dev->hard_header_len).
I think this is not correct since tpacket_fill_skb() would then
take dev->mtu and dev->hard_header_len into account for SOCK_DGRAM,
the extra VLAN_HLEN could be possible in both cases. Presumably, the
reserve code was copied from packet_snd(), but later on missed the
check. Make it similar as we have it in packet_snd().
Fixes: 69e3c75f4d ("net: TX_RING and packet mmap")
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Willem de Bruijn <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
In case no struct sockaddr_ll has been passed to packet
socket's sendmsg() when doing a TX_RING flush run, then
skb->protocol is set to po->num instead, which is the protocol
passed via socket(2)/bind(2).
Applications only xmitting can go the path of allocating the
socket as socket(PF_PACKET, <mode>, 0) and do a bind(2) on the
TX_RING with sll_protocol of 0. That way, register_prot_hook()
is neither called on creation nor on bind time, which saves
cycles when there's no interest in capturing anyway.
That leaves us however with po->num 0 instead and therefore
the TX_RING flush run sets skb->protocol to 0 as well. Eric
reported that this leads to problems when using tools like
trafgen over bonding device. I.e. the bonding's hash function
could invoke the kernel's flow dissector, which depends on
skb->protocol being properly set. In the current situation, all
the traffic is then directed to a single slave.
Fix it up by inferring skb->protocol from the Ethernet header
when not set and we have ARPHRD_ETHER device type. This is only
done in case of SOCK_RAW and where we have a dev->hard_header_len
length. In case of ARPHRD_ETHER devices, this is guaranteed to
cover ETH_HLEN, and therefore being accessed on the skb after
the skb_store_bits().
Reported-by: Eric Dumazet <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Willem de Bruijn <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Packet sockets can be used by various net devices and are not
really restricted to ARPHRD_ETHER device types. However, when
currently checking for the extra 4 bytes that can be transmitted
in VLAN case, our assumption is that we generally probe on
ARPHRD_ETHER devices. Therefore, before looking into Ethernet
header, check the device type first.
This also fixes the issue where non-ARPHRD_ETHER devices could
have no dev->hard_header_len in TX_RING SOCK_RAW case, and thus
the check would test unfilled linear part of the skb (instead
of non-linear).
Fixes: 57f89bfa21 ("network: Allow af_packet to transmit +4 bytes for VLAN packets.")
Fixes: 52f1454f62 ("packet: allow to transmit +4 byte in TX_RING slot for VLAN case")
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Willem de Bruijn <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
We concluded that the skb_probe_transport_header() should better be
called unconditionally. Avoiding the call into the flow dissector has
also not really much to do with the direct xmit mode.
While it seems that only virtio_net code makes use of GSO from non
RX/TX ring packet socket paths, we should probe for a transport header
nevertheless before they hit devices.
Reference: http://thread.gmane.org/gmane.linux.network/386173/
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Jason Wang <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
In tpacket_fill_skb() commit c1aad275b0 ("packet: set transport
header before doing xmit") and later on 40893fd0fd ("net: switch
to use skb_probe_transport_header()") was probing for a transport
header on the skb from a ring buffer slot, but at a time, where
the skb has _not even_ been filled with data yet. So that call into
the flow dissector is pretty useless. Lets do it after we've set
up the skb frags.
Fixes: c1aad275b0 ("packet: set transport header before doing xmit")
Reported-by: Eric Dumazet <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Jason Wang <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
In debugfs' start_creating(), we pin the file system to safely access
its root. When we failed to create a file, we unpin the file system via
failed_creating() to release the mount count and eventually the reference
of the vfsmount.
However, when we run into an error during lookup_one_len() when still
in start_creating(), we only release the parent's mutex but not so the
reference on the mount. Looks like it was done in the past, but after
splitting portions of __create_file() into start_creating() and
end_creating() via 190afd81e4 ("debugfs: split the beginning and the
end of __create_file() off"), this seemed missed. Noticed during code
review.
Fixes: 190afd81e4 ("debugfs: split the beginning and the end of __create_file() off")
Cc: [email protected] # v4.0+
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: Al Viro <[email protected]>
In tracefs' start_creating(), we pin the file system to safely access
its root. When we failed to create a file, we unpin the file system via
failed_creating() to release the mount count and eventually the reference
of the singleton vfsmount.
However, when we run into an error during lookup_one_len() when still
in start_creating(), we only release the parent's mutex but not so the
reference on the mount.
F.e., in securityfs_create_file(), after doing simple_pin_fs() when
lookup_one_len() fails there, we infact do simple_release_fs(). This
seems necessary here as well.
Same issue seen in debugfs due to 190afd81e4 ("debugfs: split the
beginning and the end of __create_file() off"), which seemed to got
carried over into tracefs, too. Noticed during code review.
Link: http://lkml.kernel.org/r/68efa86101b778cf7517ed7c6ad573bd69f60ec6.1446672850.git.daniel@iogearbox.net
Fixes: 4282d60689 ("tracefs: Add new tracefs file system")
Cc: [email protected] # 4.1+
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: Steven Rostedt <[email protected]>
The verbose() printer dumps the verifier state to user space, so let gcc
take care to check calls to verbose() for (future) errors. make with W=1
correctly suggests: function might be possible candidate for 'gnu_printf'
format attribute [-Wsuggest-attribute=format].
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
This work adds support for "persistent" eBPF maps/programs. The term
"persistent" is to be understood that maps/programs have a facility
that lets them survive process termination. This is desired by various
eBPF subsystem users.
Just to name one example: tc classifier/action. Whenever tc parses
the ELF object, extracts and loads maps/progs into the kernel, these
file descriptors will be out of reach after the tc instance exits.
So a subsequent tc invocation won't be able to access/relocate on this
resource, and therefore maps cannot easily be shared, f.e. between the
ingress and egress networking data path.
The current workaround is that Unix domain sockets (UDS) need to be
instrumented in order to pass the created eBPF map/program file
descriptors to a third party management daemon through UDS' socket
passing facility. This makes it a bit complicated to deploy shared
eBPF maps or programs (programs f.e. for tail calls) among various
processes.
We've been brainstorming on how we could tackle this issue and various
approches have been tried out so far, which can be read up further in
the below reference.
The architecture we eventually ended up with is a minimal file system
that can hold map/prog objects. The file system is a per mount namespace
singleton, and the default mount point is /sys/fs/bpf/. Any subsequent
mounts within a given namespace will point to the same instance. The
file system allows for creating a user-defined directory structure.
The objects for maps/progs are created/fetched through bpf(2) with
two new commands (BPF_OBJ_PIN/BPF_OBJ_GET). I.e. a bpf file descriptor
along with a pathname is being passed to bpf(2) that in turn creates
(we call it eBPF object pinning) the file system nodes. Only the pathname
is being passed to bpf(2) for getting a new BPF file descriptor to an
existing node. The user can use that to access maps and progs later on,
through bpf(2). Removal of file system nodes is being managed through
normal VFS functions such as unlink(2), etc. The file system code is
kept to a very minimum and can be further extended later on.
The next step I'm working on is to add dump eBPF map/prog commands
to bpf(2), so that a specification from a given file descriptor can
be retrieved. This can be used by things like CRIU but also applications
can inspect the meta data after calling BPF_OBJ_GET.
Big thanks also to Alexei and Hannes who significantly contributed
in the design discussion that eventually let us end up with this
architecture here.
Reference: https://lkml.org/lkml/2015/10/15/925
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: Alexei Starovoitov <[email protected]>
Signed-off-by: Hannes Frederic Sowa <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
We currently have duplicated cleanup code in bpf_prog_put() and
bpf_prog_put_rcu() cleanup paths. Back then we decided that it was
not worth it to make it a common helper called by both, but with
the recent addition of resource charging, we could have avoided
the fix in commit ac00737f4e ("bpf: Need to call bpf_prog_uncharge_memlock
from bpf_prog_put") if we would have had only a single, common path.
We can simplify it further by assigning aux->prog only once during
allocation time.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Add a bpf_map_get() function that we're going to use later on and
align/clean the remaining helpers a bit so that we have them a bit
more consistent:
- __bpf_map_get() and __bpf_prog_get() that both work on the fd
struct, check whether the descriptor is eBPF and return the
pointer to the map/prog stored in the private data.
Also, we can return f.file->private_data directly, the function
signature is enough of a documentation already.
- bpf_map_get() and bpf_prog_get() that both work on u32 user fd,
call their respective __bpf_map_get()/__bpf_prog_get() variants,
and take a reference.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Since we're going to use anon_inode_getfd() invocations in more than just
the current places, make a helper function for both, so that we only need
to pass a map/prog pointer to the helper itself in order to get a fd. The
new helpers are called bpf_map_new_fd() and bpf_prog_new_fd().
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
While recently arguing on a seccomp discussion that raw prandom_u32()
access shouldn't be exposed to unpriviledged user space, I forgot the
fact that SKF_AD_RANDOM extension actually already does it for some time
in cBPF via commit 4cd3675ebf ("filter: added BPF random opcode").
Since prandom_u32() is being used in a lot of critical networking code,
lets be more conservative and split their states. Furthermore, consolidate
eBPF and cBPF prandom handlers to use the new internal PRNG. For eBPF,
bpf_get_prandom_u32() was only accessible for priviledged users, but
should that change one day, we also don't want to leak raw sequences
through things like eBPF maps.
One thought was also to have own per bpf_prog states, but due to ABI
reasons this is not easily possible, i.e. the program code currently
cannot access bpf_prog itself, and copying the rnd_state to/from the
stack scratch space whenever a program uses the prng seems not really
worth the trouble and seems too hacky. If needed, taus113 could in such
cases be implemented within eBPF using a map entry to keep the state
space, or get_random_bytes() could become a second helper in cases where
performance would not be critical.
Both sides can trigger a one-time late init via prandom_init_once() on
the shared state. Performance-wise, there should even be a tiny gain
as bpf_user_rnd_u32() saves one function call. The PRNG needs to live
inside the BPF core since kernels could have a NET-less config as well.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Hannes Frederic Sowa <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Cc: Chema Gonzalez <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Add a prandom_init_once() facility that works on the rnd_state, so that
users that are keeping their own state independent from prandom_u32() can
initialize their taus113 per cpu states.
The motivation here is similar to net_get_random_once(): initialize the
state as late as possible in the hope that enough entropy has been
collected for the seeding. prandom_init_once() makes use of the recently
introduced prandom_seed_full_state() helper and is generic enough so that
it could also be used on fast-paths due to the DO_ONCE().
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Hannes Frederic Sowa <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Factor out the full reseed handling code that populates the state
through get_random_bytes() and runs prandom_warmup(). The resulting
prandom_seed_full_state() will be used later on in more than the
current __prandom_reseed() user. Fix also two minor whitespace
issues along the way.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Hannes Frederic Sowa <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Similar to commit c29390c6df ("xps: must clear sender_cpu before
forwarding"), we also need to clear the skb->sender_cpu when moving
from RX to TX via skb_do_redirect() due to the shared location of
napi_id (used on RX) and sender_cpu (used on TX).
Fixes: 27b29f6305 ("bpf: add bpf_redirect() helper")
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Commit ea317b267e ("bpf: Add new bpf map type to store the pointer
to struct perf_event") added perf_event.h to the main eBPF header, so
it gets included for all users. perf_event.h is actually only needed
from array map side, so lets sanitize this a bit.
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Kaixu Xia <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
The current ongoing effort to dump existing cBPF seccomp filters back
to user space requires to hold the pre-transformed instructions like
we do in case of socket filters from sk_attach_filter() side, so they
can be reloaded in original form at a later point in time by utilities
such as criu.
To prepare for this, simply extend the bpf_prog_create_from_user()
API to hold a flag that tells whether we should store the original
or not. Also, fanout filters could make use of that in future for
things like diag. While fanout filters already use bpf_prog_destroy(),
move seccomp over to them as well to handle original programs when
present.
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Tycho Andersen <[email protected]>
Cc: Pavel Emelyanov <[email protected]>
Cc: Kees Cook <[email protected]>
Cc: Andy Lutomirski <[email protected]>
Cc: Alexei Starovoitov <[email protected]>
Tested-by: Tycho Andersen <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
When sockets have a native eBPF program attached through
setsockopt(sk, SOL_SOCKET, SO_ATTACH_BPF, ...), and then try to
dump these over getsockopt(sk, SOL_SOCKET, SO_GET_FILTER, ...),
the following panic appears:
[49904.178642] BUG: unable to handle kernel NULL pointer dereference at (null)
[49904.178762] IP: [<ffffffff81610fd9>] sk_get_filter+0x39/0x90
[49904.182000] PGD 86fc9067 PUD 531a1067 PMD 0
[49904.185196] Oops: 0000 [#1] SMP
[...]
[49904.224677] Call Trace:
[49904.226090] [<ffffffff815e3d49>] sock_getsockopt+0x319/0x740
[49904.227535] [<ffffffff812f59e3>] ? sock_has_perm+0x63/0x70
[49904.228953] [<ffffffff815e2fc8>] ? release_sock+0x108/0x150
[49904.230380] [<ffffffff812f5a43>] ? selinux_socket_getsockopt+0x23/0x30
[49904.231788] [<ffffffff815dff36>] SyS_getsockopt+0xa6/0xc0
[49904.233267] [<ffffffff8171b9ae>] entry_SYSCALL_64_fastpath+0x12/0x71
The underlying issue is the very same as in commit b382c08656
("sock, diag: fix panic in sock_diag_put_filterinfo"), that is,
native eBPF programs don't store an original program since this
is only needed in cBPF ones.
However, sk_get_filter() wasn't updated to test for this at the
time when eBPF could be attached. Just throw an error to the user
to indicate that eBPF cannot be dumped over this interface.
That way, it can also be known that a program _is_ attached (as
opposed to just return 0), and a different (future) method needs
to be consulted for a dump.
Fixes: 89aa075832 ("net: sock: allow eBPF programs to be attached to sockets")
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
{cls,act}_bpf can now set the skb->priority from an eBPF program based
on various critera, so that for example classful qdiscs like multiq can
update the skb's priority during enqueue time and further push it down
into subsequent qdiscs.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Using routing realms as part of the classifier is quite useful, it
can be viewed as a tag for one or multiple routing entries (think of
an analogy to net_cls cgroup for processes), set by user space routing
daemons or via iproute2 as an indicator for traffic classifiers and
later on processed in the eBPF program.
Unlike actions, the classifier can inspect device flags and enable
netif_keep_dst() if necessary. tc actions don't have that possibility,
but in case people know what they are doing, it can be used from there
as well (e.g. via devs that must keep dsts by design anyway).
If a realm is set, the handler returns the non-zero realm. User space
can set the full 32bit realm for the dst.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
As we need to add further flags to the bpf_prog structure, lets migrate
both bools to a bitfield representation. The size of the base structure
(excluding insns) remains unchanged at 40 bytes.
Add also tags for the kmemchecker, so that it doesn't throw false
positives. Even in case gcc would generate suboptimal code, it's not
being accessed in performance critical paths.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Jamal suggested to further limit the currently allowed subset of opcodes
that may be used by a direct action return code as the intention is not
to replace the full action engine, but rather to have a minimal set that
can be used in the fast-path on things like ingress for some features
that cls_bpf supports.
Classifiers can, of course, still be chained together that have direct
action mode with those that have a full exec pass. For more complex
scenarios that go beyond this minimal set here, the full tcf_exts_exec()
path must be used.
Suggested-by: Jamal Hadi Salim <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
The binding to a particular classid was so far always mandatory for
cls_bpf, but it doesn't need to be. Therefore, lift this restriction
as similarly done in other classifiers.
Only a couple of qdiscs make use of class from the tcf_result, others
don't strictly care, so let the user choose his needs (those that read
out class can handle situations where it could be NULL).
An explicit check for tcf_unbind_filter() is also not needed here, as
the previous r->class was 0, so the xchg() will return that and
therefore a callback to the qdisc's unbind_tcf() is skipped.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
In commit 43388da42a49 ("cls_bpf: introduce integrated actions") we
have added TCA_BPF_FLAGS. We can also retrieve this information from
the prog, dump it back to user space as well. It's useful in tc when
displaying/dumping filter info.
Also, remove tp from cls_bpf_prog_from_efd(), came in as a conflict
from a rebase and it's unused here (later work may add it along with
a real user).
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Similarly as already the case in bpf_redirect()/skb_do_redirect()
pair, let the stack deal with devs that are !IFF_UP.
dev_forward_skb() as well as dev_queue_xmit() will free the skb
and increment drop counter internally in such cases, so we can
spare the condition in bpf_clone_redirect().
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Often cls_bpf classifier is used with single action drop attached.
Optimize this use case and let cls_bpf return both classid and action.
For backwards compatibility reasons enable this feature under
TCA_BPF_FLAG_ACT_DIRECT flag.
Then more interesting programs like the following are easier to write:
int cls_bpf_prog(struct __sk_buff *skb)
{
/* classify arp, ip, ipv6 into different traffic classes
* and drop all other packets
*/
switch (skb->protocol) {
case htons(ETH_P_ARP):
skb->tc_classid = 1;
break;
case htons(ETH_P_IP):
skb->tc_classid = 2;
break;
case htons(ETH_P_IPV6):
skb->tc_classid = 3;
break;
default:
return TC_ACT_SHOT;
}
return TC_ACT_OK;
}
Joint work with Daniel Borkmann.
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Ken-ichirou reported that running netlink in mmap mode for receive in
combination with nlmon will throw a NULL pointer dereference in
__kfree_skb() on nlmon_xmit(), in my case I can also trigger an "unable
to handle kernel paging request". The problem is the skb_clone() in
__netlink_deliver_tap_skb() for skbs that are mmaped.
I.e. the cloned skb doesn't have a destructor, whereas the mmap netlink
skb has it pointed to netlink_skb_destructor(), set in the handler
netlink_ring_setup_skb(). There, skb->head is being set to NULL, so
that in such cases, __kfree_skb() doesn't perform a skb_release_data()
via skb_release_all(), where skb->head is possibly being freed through
kfree(head) into slab allocator, although netlink mmap skb->head points
to the mmap buffer. Similarly, the same has to be done also for large
netlink skbs where the data area is vmalloced. Therefore, as discussed,
make a copy for these rather rare cases for now. This fixes the issue
on my and Ken-ichirou's test-cases.
Reference: http://thread.gmane.org/gmane.linux.network/371129
Fixes: bcbde0d449 ("net: netlink: virtual tap device management")
Reported-by: Ken-ichirou MATSUZAWA <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Tested-by: Ken-ichirou MATSUZAWA <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
When netlink mmap on receive side is the consumer of nf queue data,
it can happen that in some edge cases, we write skb shared info into
the user space mmap buffer:
Assume a possible rx ring frame size of only 4096, and the network skb,
which is being zero-copied into the netlink skb, contains page frags
with an overall skb->len larger than the linear part of the netlink
skb.
skb_zerocopy(), which is generic and thus not aware of the fact that
shared info cannot be accessed for such skbs then tries to write and
fill frags, thus leaking kernel data/pointers and in some corner cases
possibly writing out of bounds of the mmap area (when filling the
last slot in the ring buffer this way).
I.e. the ring buffer slot is then of status NL_MMAP_STATUS_VALID, has
an advertised length larger than 4096, where the linear part is visible
at the slot beginning, and the leaked sizeof(struct skb_shared_info)
has been written to the beginning of the next slot (also corrupting
the struct nl_mmap_hdr slot header incl. status etc), since skb->end
points to skb->data + ring->frame_size - NL_MMAP_HDRLEN.
The fix adds and lets __netlink_alloc_skb() take the actual needed
linear room for the network skb + meta data into account. It's completely
irrelevant for non-mmaped netlink sockets, but in case mmap sockets
are used, it can be decided whether the available skb_tailroom() is
really large enough for the buffer, or whether it needs to internally
fallback to a normal alloc_skb().
>From nf queue side, the information whether the destination port is
an mmap RX ring is not really available without extra port-to-socket
lookup, thus it can only be determined in lower layers i.e. when
__netlink_alloc_skb() is called that checks internally for this. I
chose to add the extra ldiff parameter as mmap will then still work:
We have data_len and hlen in nfqnl_build_packet_message(), data_len
is the full length (capped at queue->copy_range) for skb_zerocopy()
and hlen some possible part of data_len that needs to be copied; the
rem_len variable indicates the needed remaining linear mmap space.
The only other workaround in nf queue internally would be after
allocation time by f.e. cap'ing the data_len to the skb_tailroom()
iff we deal with an mmap skb, but that would 1) expose the fact that
we use a mmap skb to upper layers, and 2) trim the skb where we
otherwise could just have moved the full skb into the normal receive
queue.
After the patch, in my test case the ring slot doesn't fit and therefore
shows NL_MMAP_STATUS_COPY, where a full skb carries all the data and
thus needs to be picked up via recv().
Fixes: 3ab1f683bf ("nfnetlink: add support for memory mapped netlink")
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
In case of netlink mmap, there can be situations where received frames
have to be placed into the normal receive queue. The ring buffer indicates
this through NL_MMAP_STATUS_COPY, so the user is asked to pick them up
via recvmsg(2) syscall, and to put the slot back to NL_MMAP_STATUS_UNUSED.
Commit 0ef707700f ("netlink: rx mmap: fix POLLIN condition") changed
polling, so that we walk in the worst case the whole ring through the
new netlink_has_valid_frame(), for example, when the ring would have no
NL_MMAP_STATUS_VALID, but at least one NL_MMAP_STATUS_COPY frame.
Since we do a datagram_poll() already earlier to pick up a mask that could
possibly contain POLLIN | POLLRDNORM already (due to NL_MMAP_STATUS_COPY),
we can skip checking the rx ring entirely.
In case the kernel is compiled with !CONFIG_NETLINK_MMAP, then all this is
irrelevant anyway as netlink_poll() is just defined as datagram_poll().
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
We may already have gotten a proper fd struct through fdget(), so
whenever we return at the end of an map operation, we need to call
fdput(). However, each map operation from syscall side first probes
CHECK_ATTR() to verify that unused fields in the bpf_attr union are
zero.
In case of malformed input, we return with error, but the lookup to
the map_fd was already performed at that time, so that we return
without an corresponding fdput(). Fix it by performing an fdget()
only right before bpf_map_get(). The fdget() invocation on maps in
the verifier is not affected.
Fixes: db20fd2b01 ("bpf: add lookup/update/delete/iterate methods to BPF maps")
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
While debugging a networking issue, I hit a condition that triggered an
object to be freed into the wrong kmem cache, and thus triggered the
warning in cache_from_obj().
The arguments in the error message are in wrong order: the location
of the object's kmem cache is in cachep, not s.
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Christoph Lameter <[email protected]>
Cc: Pekka Enberg <[email protected]>
Cc: David Rientjes <[email protected]>
Cc: Joonsoo Kim <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Fengguang reported, that some randconfig generated the following linker
issue with nf_ct_zone_dflt object involved:
[...]
CC init/version.o
LD init/built-in.o
net/built-in.o: In function `ipv4_conntrack_defrag':
nf_defrag_ipv4.c:(.text+0x93e95): undefined reference to `nf_ct_zone_dflt'
net/built-in.o: In function `ipv6_defrag':
nf_defrag_ipv6_hooks.c:(.text+0xe3ffe): undefined reference to `nf_ct_zone_dflt'
make: *** [vmlinux] Error 1
Given that configurations exist where we have a built-in part, which is
accessing nf_ct_zone_dflt such as the two handlers nf_ct_defrag_user()
and nf_ct6_defrag_user(), and a part that configures nf_conntrack as a
module, we must move nf_ct_zone_dflt into a fixed, guaranteed built-in
area when netfilter is configured in general.
Therefore, split the more generic parts into a common header under
include/linux/netfilter/ and move nf_ct_zone_dflt into the built-in
section that already holds parts related to CONFIG_NF_CONNTRACK in the
netfilter core. This fixes the issue on my side.
Fixes: 308ac9143e ("netfilter: nf_conntrack: push zone object into functions")
Reported-by: Fengguang Wu <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
While testing various Kconfig options on another issue, I found that
the following one triggers as well on allmodconfig and nf_conntrack
disabled:
net/ipv4/netfilter/nf_dup_ipv4.c: In function ‘nf_dup_ipv4’:
net/ipv4/netfilter/nf_dup_ipv4.c:72:20: error: ‘nf_skb_duplicated’ undeclared (first use in this function)
if (this_cpu_read(nf_skb_duplicated))
[...]
net/ipv6/netfilter/nf_dup_ipv6.c: In function ‘nf_dup_ipv6’:
net/ipv6/netfilter/nf_dup_ipv6.c:66:20: error: ‘nf_skb_duplicated’ undeclared (first use in this function)
if (this_cpu_read(nf_skb_duplicated))
Fix it by including directly the header where it is defined.
Fixes: bbde9fc182 ("netfilter: factor out packet duplication for IPv4/IPv6")
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
We previously register IPPROTO_ROUTING offload under inet6_add_offload(),
but in error path, we try to unregister it with inet_del_offload(). This
doesn't seem correct, it should actually be inet6_del_offload(), also
ipv6_exthdrs_offload_exit() from that commit seems rather incorrect (it
also uses rthdr_offload twice), but it got removed entirely later on.
Fixes: 3336288a9f ("ipv6: Switch to using new offload infrastructure.")
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
diag socket's sock_diag_put_filterinfo() dumps classic BPF programs
upon request to user space (ss -0 -b). However, native eBPF programs
attached to sockets (SO_ATTACH_BPF) cannot be dumped with this method:
Their orig_prog is always NULL. However, sock_diag_put_filterinfo()
unconditionally tries to access its filter length resp. wants to copy
the filter insns from there. Internal cBPF to eBPF transformations
attached to sockets don't have this issue, as orig_prog state is kept.
It's currently only used by packet sockets. If we would want to add
native eBPF support in the future, this needs to be done through
a different attribute than PACKET_DIAG_FILTER to not confuse possible
user space disassemblers that work on diag data.
Fixes: 89aa075832 ("net: sock: allow eBPF programs to be attached to sockets")
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Nicolas Dichtel <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Commit 0838aa7fcf ("netfilter: fix netns dependencies with conntrack
templates") migrated templates to the new allocator api, but forgot to
update error paths for them in CT and synproxy to use nf_ct_tmpl_free()
instead of nf_conntrack_free().
Due to that, memory is being freed into the wrong kmemcache, but also
we drop the per net reference count of ct objects causing an imbalance.
In Brad's case, this leads to a wrap-around of net->ct.count and thus
lets __nf_conntrack_alloc() refuse to create a new ct object:
[ 10.340913] xt_addrtype: ipv6 does not support BROADCAST matching
[ 10.810168] nf_conntrack: table full, dropping packet
[ 11.917416] r8169 0000:07:00.0 eth0: link up
[ 11.917438] IPv6: ADDRCONF(NETDEV_CHANGE): eth0: link becomes ready
[ 12.815902] nf_conntrack: table full, dropping packet
[ 15.688561] nf_conntrack: table full, dropping packet
[ 15.689365] nf_conntrack: table full, dropping packet
[ 15.690169] nf_conntrack: table full, dropping packet
[ 15.690967] nf_conntrack: table full, dropping packet
[...]
With slab debugging, it also reports the wrong kmemcache (kmalloc-512 vs.
nf_conntrack_ffffffff81ce75c0) and reports poison overwrites, etc. Thus,
to fix the problem, export and use nf_ct_tmpl_free() instead.
Fixes: 0838aa7fcf ("netfilter: fix netns dependencies with conntrack templates")
Reported-by: Brad Jackson <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: Pablo Neira Ayuso <[email protected]>
Currently, the following case doesn't use DCTCP, even if it should:
A responder has f.e. Cubic as system wide default, but for a specific
route to the initiating host, DCTCP is being set in RTAX_CC_ALGO. The
initiating host then uses DCTCP as congestion control, but since the
initiator sets ECT(0), tcp_ecn_create_request() doesn't set ecn_ok,
and we have to fall back to Reno after 3WHS completes.
We were thinking on how to solve this in a minimal, non-intrusive
way without bloating tcp_ecn_create_request() needlessly: lets cache
the CA ecn option flag in RTAX_FEATURES. In other words, when ECT(0)
is set on the SYN packet, set ecn_ok=1 iff route RTAX_FEATURES
contains the unexposed (internal-only) DST_FEATURE_ECN_CA. This allows
to only do a single metric feature lookup inside tcp_ecn_create_request().
Joint work with Florian Westphal.
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: Florian Westphal <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Feature bits that are invalid should not be accepted by the kernel,
only the lower 4 bits may be configured, but not the remaining ones.
Even from these 4, 2 of them are unused.
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Reduce the identation a bit, there's no need to artificically have
it increased.
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Just some minor noise follow-up to address some stylistic issues of
commit 3b3ae88026 ("net: sched: consolidate tc_classify{,_compat}").
Accidentally v1 instead of v2 of that commit got applied, so this
patch adds the relative diff.
Suggested-by: Alexei Starovoitov <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
For classifiers getting invoked via tc_classify(), we always need an
extra function call into tc_classify_compat(), as both are being
exported as symbols and tc_classify() itself doesn't do much except
handling of reclassifications when tp->classify() returned with
TC_ACT_RECLASSIFY.
CBQ and ATM are the only qdiscs that directly call into tc_classify_compat(),
all others use tc_classify(). When tc actions are being configured
out in the kernel, tc_classify() effectively does nothing besides
delegating.
We could spare this layer and consolidate both functions. pktgen on
single CPU constantly pushing skbs directly into the netif_receive_skb()
path with a dummy classifier on ingress qdisc attached, improves
slightly from 22.3Mpps to 23.1Mpps.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
This work adds the possibility of deriving the zone id from the skb->mark
field in a scalable manner. This allows for having only a single template
serving hundreds/thousands of different zones, for example, instead of the
need to have one match for each zone as an extra CT jump target.
Note that we'd need to have this information attached to the template as at
the time when we're trying to lookup a possible ct object, we already need
to know zone information for a possible match when going into
__nf_conntrack_find_get(). This work provides a minimal implementation for
a possible mapping.
In order to not add/expose an extra ct->status bit, the zone structure has
been extended to carry a flag for deriving the mark.
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: Pablo Neira Ayuso <[email protected]>
This work adds a direction parameter to netfilter zones, so identity
separation can be performed only in original/reply or both directions
(default). This basically opens up the possibility of doing NAT with
conflicting IP address/port tuples from multiple, isolated tenants
on a host (e.g. from a netns) without requiring each tenant to NAT
twice resp. to use its own dedicated IP address to SNAT to, meaning
overlapping tuples can be made unique with the zone identifier in
original direction, where the NAT engine will then allocate a unique
tuple in the commonly shared default zone for the reply direction.
In some restricted, local DNAT cases, also port redirection could be
used for making the reply traffic unique w/o requiring SNAT.
The consensus we've reached and discussed at NFWS and since the initial
implementation [1] was to directly integrate the direction meta data
into the existing zones infrastructure, as opposed to the ct->mark
approach we proposed initially.
As we pass the nf_conntrack_zone object directly around, we don't have
to touch all call-sites, but only those, that contain equality checks
of zones. Thus, based on the current direction (original or reply),
we either return the actual id, or the default NF_CT_DEFAULT_ZONE_ID.
CT expectations are direction-agnostic entities when expectations are
being compared among themselves, so we can only use the identifier
in this case.
Note that zone identifiers can not be included into the hash mix
anymore as they don't contain a "stable" value that would be equal
for both directions at all times, f.e. if only zone->id would
unconditionally be xor'ed into the table slot hash, then replies won't
find the corresponding conntracking entry anymore.
If no particular direction is specified when configuring zones, the
behaviour is exactly as we expect currently (both directions).
Support has been added for the CT netlink interface as well as the
x_tables raw CT target, which both already offer existing interfaces
to user space for the configuration of zones.
Below a minimal, simplified collision example (script in [2]) with
netperf sessions:
+--- tenant-1 ---+ mark := 1
| netperf |--+
+----------------+ | CT zone := mark [ORIGINAL]
[ip,sport] := X +--------------+ +--- gateway ---+
| mark routing |--| SNAT |-- ... +
+--------------+ +---------------+ |
+--- tenant-2 ---+ | ~~~|~~~
| netperf |--+ +-----------+ |
+----------------+ mark := 2 | netserver |------ ... +
[ip,sport] := X +-----------+
[ip,port] := Y
On the gateway netns, example:
iptables -t raw -A PREROUTING -j CT --zone mark --zone-dir ORIGINAL
iptables -t nat -A POSTROUTING -o <dev> -j SNAT --to-source <ip> --random-fully
iptables -t mangle -A PREROUTING -m conntrack --ctdir ORIGINAL -j CONNMARK --save-mark
iptables -t mangle -A POSTROUTING -m conntrack --ctdir REPLY -j CONNMARK --restore-mark
conntrack dump from gateway netns:
netperf -H 10.1.1.2 -t TCP_STREAM -l60 -p12865,5555 from each tenant netns
tcp 6 431995 ESTABLISHED src=40.1.1.1 dst=10.1.1.2 sport=5555 dport=12865 zone-orig=1
src=10.1.1.2 dst=10.1.1.1 sport=12865 dport=1024
[ASSURED] mark=1 secctx=system_u:object_r:unlabeled_t:s0 use=1
tcp 6 431994 ESTABLISHED src=40.1.1.1 dst=10.1.1.2 sport=5555 dport=12865 zone-orig=2
src=10.1.1.2 dst=10.1.1.1 sport=12865 dport=5555
[ASSURED] mark=2 secctx=system_u:object_r:unlabeled_t:s0 use=1
tcp 6 299 ESTABLISHED src=40.1.1.1 dst=10.1.1.2 sport=39438 dport=33768 zone-orig=1
src=10.1.1.2 dst=10.1.1.1 sport=33768 dport=39438
[ASSURED] mark=1 secctx=system_u:object_r:unlabeled_t:s0 use=1
tcp 6 300 ESTABLISHED src=40.1.1.1 dst=10.1.1.2 sport=32889 dport=40206 zone-orig=2
src=10.1.1.2 dst=10.1.1.1 sport=40206 dport=32889
[ASSURED] mark=2 secctx=system_u:object_r:unlabeled_t:s0 use=2
Taking this further, test script in [2] creates 200 tenants and runs
original-tuple colliding netperf sessions each. A conntrack -L dump in
the gateway netns also confirms 200 overlapping entries, all in ESTABLISHED
state as expected.
I also did run various other tests with some permutations of the script,
to mention some: SNAT in random/random-fully/persistent mode, no zones (no
overlaps), static zones (original, reply, both directions), etc.
[1] http://thread.gmane.org/gmane.comp.security.firewalls.netfilter.devel/57412/
[2] https://paste.fedoraproject.org/242835/65657871/
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: Pablo Neira Ayuso <[email protected]>
This patch replaces the zone id which is pushed down into functions
with the actual zone object. It's a bigger one-time change, but
needed for later on extending zones with a direction parameter, and
thus decoupling this additional information from all call-sites.
No functional changes in this patch.
The default zone becomes a global const object, namely nf_ct_zone_dflt
and will be returned directly in various cases, one being, when there's
f.e. no zoning support.
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: Pablo Neira Ayuso <[email protected]>
Linus reports the following deadlock on rtnl_mutex; triggered only
once so far (extract):
[12236.694209] NetworkManager D 0000000000013b80 0 1047 1 0x00000000
[12236.694218] ffff88003f902640 0000000000000000 ffffffff815d15a9 0000000000000018
[12236.694224] ffff880119538000 ffff88003f902640 ffffffff81a8ff84 00000000ffffffff
[12236.694230] ffffffff81a8ff88 ffff880119c47f00 ffffffff815d133a ffffffff81a8ff80
[12236.694235] Call Trace:
[12236.694250] [<ffffffff815d15a9>] ? schedule_preempt_disabled+0x9/0x10
[12236.694257] [<ffffffff815d133a>] ? schedule+0x2a/0x70
[12236.694263] [<ffffffff815d15a9>] ? schedule_preempt_disabled+0x9/0x10
[12236.694271] [<ffffffff815d2c3f>] ? __mutex_lock_slowpath+0x7f/0xf0
[12236.694280] [<ffffffff815d2cc6>] ? mutex_lock+0x16/0x30
[12236.694291] [<ffffffff814f1f90>] ? rtnetlink_rcv+0x10/0x30
[12236.694299] [<ffffffff8150ce3b>] ? netlink_unicast+0xfb/0x180
[12236.694309] [<ffffffff814f5ad3>] ? rtnl_getlink+0x113/0x190
[12236.694319] [<ffffffff814f202a>] ? rtnetlink_rcv_msg+0x7a/0x210
[12236.694331] [<ffffffff8124565c>] ? sock_has_perm+0x5c/0x70
[12236.694339] [<ffffffff814f1fb0>] ? rtnetlink_rcv+0x30/0x30
[12236.694346] [<ffffffff8150d62c>] ? netlink_rcv_skb+0x9c/0xc0
[12236.694354] [<ffffffff814f1f9f>] ? rtnetlink_rcv+0x1f/0x30
[12236.694360] [<ffffffff8150ce3b>] ? netlink_unicast+0xfb/0x180
[12236.694367] [<ffffffff8150d344>] ? netlink_sendmsg+0x484/0x5d0
[12236.694376] [<ffffffff810a236f>] ? __wake_up+0x2f/0x50
[12236.694387] [<ffffffff814cad23>] ? sock_sendmsg+0x33/0x40
[12236.694396] [<ffffffff814cb05e>] ? ___sys_sendmsg+0x22e/0x240
[12236.694405] [<ffffffff814cab75>] ? ___sys_recvmsg+0x135/0x1a0
[12236.694415] [<ffffffff811a9d12>] ? eventfd_write+0x82/0x210
[12236.694423] [<ffffffff811a0f9e>] ? fsnotify+0x32e/0x4c0
[12236.694429] [<ffffffff8108cb70>] ? wake_up_q+0x60/0x60
[12236.694434] [<ffffffff814cba09>] ? __sys_sendmsg+0x39/0x70
[12236.694440] [<ffffffff815d4797>] ? entry_SYSCALL_64_fastpath+0x12/0x6a
It seems so far plausible that the recursive call into rtnetlink_rcv()
looks suspicious. One way, where this could trigger is that the senders
NETLINK_CB(skb).portid was wrongly 0 (which is rtnetlink socket), so
the rtnl_getlink() request's answer would be sent to the kernel instead
to the actual user process, thus grabbing rtnl_mutex() twice.
One theory would be that netlink_autobind() triggered via netlink_sendmsg()
internally overwrites the -EBUSY error to 0, but where it is wrongly
originating from __netlink_insert() instead. That would reset the
socket's portid to 0, which is then filled into NETLINK_CB(skb).portid
later on. As commit d470e3b483 ("[NETLINK]: Fix two socket hashing bugs.")
also puts it, -EBUSY should not be propagated from netlink_insert().
It looks like it's very unlikely to reproduce. We need to trigger the
rhashtable_insert_rehash() handler under a situation where rehashing
currently occurs (one /rare/ way would be to hit ht->elasticity limits
while not filled enough to expand the hashtable, but that would rather
require a specifically crafted bind() sequence with knowledge about
destination slots, seems unlikely). It probably makes sense to guard
__netlink_insert() in any case and remap that error. It was suggested
that EOVERFLOW might be better than an already overloaded ENOMEM.
Reference: http://thread.gmane.org/gmane.linux.network/372676
Reported-by: Linus Torvalds <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Herbert Xu <[email protected]>
Acked-by: Thomas Graf <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Since the introduction of the BPF action in d23b8ad8ab ("tc: add BPF
based action"), late binding was not working as expected. I.e. setting
the action part for a classifier only via 'bpf index <num>', where <num>
is the index of an existing action, is being rejected by the kernel due
to other missing parameters.
It doesn't make sense to require these parameters such as BPF opcodes
etc, as they are not going to be used anyway: in this case, they're just
allocated/parsed and then freed again w/o doing anything meaningful.
Instead, parse and verify the remaining parameters *after* the test on
tcf_hash_check(), when we really know that we're dealing with creation
of a new action or replacement of an existing one and where late binding
is thus irrelevant.
After patch, test case is now working:
FOO="1,6 0 0 4294967295,"
tc actions add action bpf bytecode "$FOO"
tc filter add dev foo parent 1: bpf bytecode "$FOO" flowid 1:1 action bpf index 1
tc actions show action bpf
action order 0: bpf bytecode '1,6 0 0 4294967295' default-action pipe
index 1 ref 2 bind 1
tc filter show dev foo
filter protocol all pref 49152 bpf
filter protocol all pref 49152 bpf handle 0x1 flowid 1:1 bytecode '1,6 0 0 4294967295'
action order 1: bpf bytecode '1,6 0 0 4294967295' default-action pipe
index 1 ref 2 bind 1
Late binding of a BPF action can be useful for preloading maps (e.g. before
they hit traffic) in case of eBPF programs, or to share a single eBPF action
with multiple classifiers.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Add skb->hash to the __sk_buff offset map, so it can be accessed from
an eBPF program. We currently already do this for classic BPF filters,
but not yet on eBPF, it might be useful as a demuxer in combination with
helpers like bpf_clone_redirect(), toy example:
__section("cls-lb") int ingress_main(struct __sk_buff *skb)
{
unsigned int which = 3 + (skb->hash & 7);
/* bpf_skb_store_bytes(skb, ...); */
/* bpf_l{3,4}_csum_replace(skb, ...); */
bpf_clone_redirect(skb, which, 0);
return -1;
}
I was thinking whether to add skb_get_hash(), but then concluded the
raw skb->hash seems fine in this case: we can directly access the hash
w/o extra eBPF helper function call, it's filled out by many NICs on
ingress, and in case the entropy level would not be sufficient, people
can still implement their own specific sw fallback hash mix anyway.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
This patch adds support to read the dmesg BPF JIT dump also from a
file instead of the klog buffer. I found this quite useful when going
through some 'before/after patch' logs. It also fixes a regex leak
found by valgrind when no image dump was found.
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Since commit 55334a5db5 ("net_sched: act: refuse to remove bound action
outside"), we end up with a wrong reference count for a tc action.
Test case 1:
FOO="1,6 0 0 4294967295,"
BAR="1,6 0 0 4294967294,"
tc filter add dev foo parent 1: bpf bytecode "$FOO" flowid 1:1 \
action bpf bytecode "$FOO"
tc actions show action bpf
action order 0: bpf bytecode '1,6 0 0 4294967295' default-action pipe
index 1 ref 1 bind 1
tc actions replace action bpf bytecode "$BAR" index 1
tc actions show action bpf
action order 0: bpf bytecode '1,6 0 0 4294967294' default-action pipe
index 1 ref 2 bind 1
tc actions replace action bpf bytecode "$FOO" index 1
tc actions show action bpf
action order 0: bpf bytecode '1,6 0 0 4294967295' default-action pipe
index 1 ref 3 bind 1
Test case 2:
FOO="1,6 0 0 4294967295,"
tc filter add dev foo parent 1: bpf bytecode "$FOO" flowid 1:1 action ok
tc actions show action gact
action order 0: gact action pass
random type none pass val 0
index 1 ref 1 bind 1
tc actions add action drop index 1
RTNETLINK answers: File exists [...]
tc actions show action gact
action order 0: gact action pass
random type none pass val 0
index 1 ref 2 bind 1
tc actions add action drop index 1
RTNETLINK answers: File exists [...]
tc actions show action gact
action order 0: gact action pass
random type none pass val 0
index 1 ref 3 bind 1
What happens is that in tcf_hash_check(), we check tcf_common for a given
index and increase tcfc_refcnt and conditionally tcfc_bindcnt when we've
found an existing action. Now there are the following cases:
1) We do a late binding of an action. In that case, we leave the
tcfc_refcnt/tcfc_bindcnt increased and are done with the ->init()
handler. This is correctly handeled.
2) We replace the given action, or we try to add one without replacing
and find out that the action at a specific index already exists
(thus, we go out with error in that case).
In case of 2), we have to undo the reference count increase from
tcf_hash_check() in the tcf_hash_check() function. Currently, we fail to
do so because of the 'tcfc_bindcnt > 0' check which bails out early with
an -EPERM error.
Now, while commit 55334a5db5 prevents 'tc actions del action ...' on an
already classifier-bound action to drop the reference count (which could
then become negative, wrap around etc), this restriction only accounts for
invocations outside a specific action's ->init() handler.
One possible solution would be to add a flag thus we possibly trigger
the -EPERM ony in situations where it is indeed relevant.
After the patch, above test cases have correct reference count again.
Fixes: 55334a5db5 ("net_sched: act: refuse to remove bound action outside")
Signed-off-by: Daniel Borkmann <[email protected]>
Reviewed-by: Cong Wang <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
It can be useful for testing to see the actual process/pid who is loading
a given filter. I was running some BPF test program and noticed unusual
filter loads from time to time, triggered by some other application in the
background. bpf_jit_disasm is still working after this change.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
When bpf_jit_compile() got split into two functions via commit
f3c2af7ba1 ("net: filter: x86: split bpf_jit_compile()"), bpf_jit_dump()
was changed to always show 0 as number of compiler passes. Change it to
dump the actual number. Also on sparc, we count passes starting from 0,
so add 1 for the debug dump as well.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
During recent discussions we had with Michael, we found that it would
be useful to have an indicator that tells the JIT that an eBPF program
had been migrated from classic instructions into eBPF instructions, as
only in that case A and X need to be cleared in the prologue. Such eBPF
programs do not set a particular type, but all have BPF_PROG_TYPE_UNSPEC.
Thus, introduce a small helper for cde66c2d88 ("s390/bpf: Only clear
A and X for converted BPF programs") and possibly others in future.
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Michael Holzheu <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
As JITs start to perform optimizations whether to clear A and X on eBPF
programs in the prologue, we should actually assign a program type to the
native eBPF test cases. It doesn't really matter which program type, as
these instructions don't go through the verifier, but it needs to be a
type != BPF_PROG_TYPE_UNSPEC. This reflects eBPF programs loaded via bpf(2)
system call (!= type unspec) vs. classic BPF to eBPF migrations (== type
unspec).
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Michael Holzheu <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Back then when we added support for SCTP_SNDINFO/SCTP_RCVINFO from
RFC6458 5.3.4/5.3.5, we decided to add a deprecation warning for the
(as per RFC deprecated) SCTP_SNDRCV via commit bbbea41d5e ("net:
sctp: deprecate rfc6458, 5.3.2. SCTP_SNDRCV support"), see [1].
Imho, it was not a good idea, and we should just revert that message
for a couple of reasons:
1) It's uapi and therefore set in stone forever.
2) To be able to run on older and newer kernels, an SCTP application
would need to probe for both, SCTP_SNDRCV, but also SCTP_SNDINFO/
SCTP_RCVINFO support, so that on older kernels, it can make use
of SCTP_SNDRCV, and on newer kernels SCTP_SNDINFO/SCTP_RCVINFO.
In my (limited) experience, a lot of SCTP appliances are migrating
to newer kernels only ve(ee)ry slowly.
3) Some people don't have the chance to change their applications,
f.e. due to proprietary legacy stuff. So, they'll hit this warning
in fast path and are stuck with older kernels.
But i.e. due to point 1) I really fail to see the benefit of a warning.
So just revert that for now, the issue was reported up Jamal.
[1] http://thread.gmane.org/gmane.linux.network/321960/
Reported-by: Jamal Hadi Salim <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Michael Tuexen <[email protected]>
Acked-by: Jamal Hadi Salim <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
The following test case causes a NULL pointer dereference in cls_flow:
tc filter add dev foo parent 1: handle 0x1 flow hash keys dst action ok
tc filter replace dev foo parent 1: pref 49152 handle 0x1 \
flow hash keys mark action drop
To be more precise, actually two different panics are fixed, the first
occurs because tcf_exts_init() is not called on the newly allocated
filter when we do a replace. And the second panic uncovered after that
happens since the arguments of list_replace_rcu() are swapped, the old
element needs to be the first argument and the new element the second.
Fixes: 70da9f0bf9 ("net: sched: cls_flow use RCU")
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: John Fastabend <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
The following test case causes a NULL pointer dereference in cls_flower:
tc filter add dev foo parent 1: flower eth_type ipv4 action ok flowid 1:1
tc filter replace dev foo parent 1: pref 49152 handle 0x1 \
flower eth_type ipv6 action ok flowid 1:1
The problem is that commit 77b9900ef5 ("tc: introduce Flower classifier")
accidentally swapped the arguments of list_replace_rcu(), the old
element needs to be the first argument and the new element the second.
Fixes: 77b9900ef5 ("tc: introduce Flower classifier")
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Jiri Pirko <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
The following test case causes a NULL pointer dereference in cls_bpf:
FOO="1,6 0 0 4294967295,"
tc filter add dev foo parent 1: bpf bytecode "$FOO" flowid 1:1 action ok
tc filter replace dev foo parent 1: pref 49152 handle 0x1 \
bpf bytecode "$FOO" flowid 1:1 action drop
The problem is that commit 1f947bf151 ("net: sched: rcu'ify cls_bpf")
accidentally swapped the arguments of list_replace_rcu(), the old
element needs to be the first argument and the new element the second.
Fixes: 1f947bf151 ("net: sched: rcu'ify cls_bpf")
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: John Fastabend <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
It would be very useful to retrieve the net_cls's classid from an eBPF
program to allow for a more fine-grained classification, it could be
directly used or in conjunction with additional policies. I.e. docker,
but also tooling such as cgexec, can easily run applications via net_cls
cgroups:
cgcreate -g net_cls:/foo
echo 42 > foo/net_cls.classid
cgexec -g net_cls:foo <prog>
Thus, their respecitve classid cookie of foo can then be looked up on
the egress path to apply further policies. The helper is desigend such
that a non-zero value returns the cgroup id.
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Thomas Graf <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Split out retrieving the cgroups net_cls classid retrieval into its
own function, so that it can be reused later on from other parts of
the traffic control subsystem. If there's no skb->sk, then the small
helper returns 0 as well, which in cls_cgroup terms means 'could not
classify'.
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Thomas Graf <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Similarly as in commit 4f7d2cdfdd ("rtnetlink: verify IFLA_VF_INFO
attributes before passing them to driver"), we have a double nesting
of netlink attributes, i.e. IFLA_VF_PORTS only contains IFLA_VF_PORT
that is nested itself. While IFLA_VF_PORTS is a verified attribute
from ifla_policy[], we only check if the IFLA_VF_PORTS container has
IFLA_VF_PORT attributes and then pass the attribute's content itself
via nla_parse_nested(). It would be more correct to reject inner types
other than IFLA_VF_PORT instead of continuing parsing and also similarly
as in commit 4f7d2cdfdd, to check for a minimum of NLA_HDRLEN.
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Roopa Prabhu <[email protected]>
Cc: Scott Feldman <[email protected]>
Cc: Jason Gunthorpe <[email protected]>
Acked-by: Roopa Prabhu <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
ARG1 = BPF_R1 as it stands, evaluates to regs[BPF_REG_1] = regs[BPF_REG_1]
and thus has no effect. Add a comment instead, explaining what happens and
why it's okay to just remove it. Since from user space side, a tail call is
invoked as a pseudo helper function via bpf_tail_call_proto, the verifier
checks the arguments just like with any other helper function and makes
sure that the first argument (regs[BPF_REG_1])'s type is ARG_PTR_TO_CTX.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Jason Gunthorpe reported that since commit c02db8c629 ("rtnetlink: make
SR-IOV VF interface symmetric"), we don't verify IFLA_VF_INFO attributes
anymore with respect to their policy, that is, ifla_vfinfo_policy[].
Before, they were part of ifla_policy[], but they have been nested since
placed under IFLA_VFINFO_LIST, that contains the attribute IFLA_VF_INFO,
which is another nested attribute for the actual VF attributes such as
IFLA_VF_MAC, IFLA_VF_VLAN, etc.
Despite the policy being split out from ifla_policy[] in this commit,
it's never applied anywhere. nla_for_each_nested() only does basic nla_ok()
testing for struct nlattr, but it doesn't know about the data context and
their requirements.
Fix, on top of Jason's initial work, does 1) parsing of the attributes
with the right policy, and 2) using the resulting parsed attribute table
from 1) instead of the nla_for_each_nested() loop (just like we used to
do when still part of ifla_policy[]).
Reference: http://thread.gmane.org/gmane.linux.network/368913
Fixes: c02db8c629 ("rtnetlink: make SR-IOV VF interface symmetric")
Reported-by: Jason Gunthorpe <[email protected]>
Cc: Chris Wright <[email protected]>
Cc: Sucheta Chakraborty <[email protected]>
Cc: Greg Rose <[email protected]>
Cc: Jeff Kirsher <[email protected]>
Cc: Rony Efraim <[email protected]>
Cc: Vlad Zolotarov <[email protected]>
Cc: Nicolas Dichtel <[email protected]>
Cc: Thomas Graf <[email protected]>
Signed-off-by: Jason Gunthorpe <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Vlad Zolotarov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Cleanup commit 73679e5082 ("compiler-intel.h: Remove duplicate
definition") removed the double definition of __memory_barrier()
intrinsics.
However, in doing so, it also removed the preceding #undef barrier by
accident, meaning, the actual barrier() macro from compiler-gcc.h with
inline asm is still in place as __GNUC__ is provided.
Subsequently, barrier() can never be defined as __memory_barrier() from
compiler.h since it already has a definition in place and if we trust
the comment in compiler-intel.h, ecc doesn't support gcc specific asm
statements.
I don't have an ecc at hand (unsure if that's still used in the field?)
and only found this by accident during code review, a revert of that
cleanup would be simplest option.
Fixes: 73679e5082 ("compiler-intel.h: Remove duplicate definition")
Signed-off-by: Daniel Borkmann <[email protected]>
Reviewed-by: Pranith Kumar <[email protected]>
Cc: Pranith Kumar <[email protected]>
Cc: H. Peter Anvin <[email protected]>
Cc: mancha security <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Besides others, move bpf_tail_call_proto to the remaining definitions
of other protos, improve comments a bit (i.e. remove some obvious ones,
where the code is already self-documenting, add objectives for others),
simplify bpf_prog_array_compatible() a bit.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
As this is already exported from tracing side via commit d9847d310a
("tracing: Allow BPF programs to call bpf_ktime_get_ns()"), we might
as well want to move it to the core, so also networking users can make
use of it, e.g. to measure diffs for certain flows from ingress/egress.
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Alexei Starovoitov <[email protected]>
Cc: Ingo Molnar <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
While 3b52960266 ("test_bpf: add more eBPF jump torture cases")
added the int3 bug test case only for eBPF, which needs exactly 11
passes to converge, here's a version for classic BPF with 11 passes,
and one that would need 70 passes on x86_64 to actually converge for
being successfully JITed. Effectively, all jumps are being optimized
out resulting in a JIT image of just 89 bytes (from originally max
BPF insns), only returning K.
Might be useful as a receipe for folks wanting to craft a test case
when backporting the fix in commit 3f7352bf21 ("x86: bpf_jit: fix
compilation of large bpf programs") while not having eBPF. The 2nd
one is delegated to the interpreter as the last pass still results
in shrinking, in other words, this one won't be JITed on x86_64.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
With recent debugging, I noticed that bpf_jit_disasm segfaults when
there's no debugging output from the JIT compiler to the kernel log.
Reason is that when regexec(3) doesn't match on anything, start/end
offsets are not being filled out and contain some uninitialized garbage
from stack. Thus, we need zero out offsets first.
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Add two more eBPF test cases for JITs, i.e. the second one revealed a
bug in the x86_64 JIT compiler, where only an int3 filled image from
the allocator was emitted and later wrongly set by the compiler as the
bpf_func program code since optimization pass boundary was surpassed
w/o actually emitting opcodes.
Interpreter:
[ 45.782892] test_bpf: #242 BPF_MAXINSNS: Very long jump backwards jited:0 11 PASS
[ 45.783062] test_bpf: #243 BPF_MAXINSNS: Edge hopping nuthouse jited:0 14705 PASS
After x86_64 JIT (fixed):
[ 80.495638] test_bpf: #242 BPF_MAXINSNS: Very long jump backwards jited:1 6 PASS
[ 80.495957] test_bpf: #243 BPF_MAXINSNS: Edge hopping nuthouse jited:1 17157 PASS
Reference: http://thread.gmane.org/gmane.linux.network/364729
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Vijay reported that a loop as simple as ...
while true; do
tc qdisc add dev foo root handle 1: prio
tc filter add dev foo parent 1: u32 match u32 0 0 flowid 1
tc qdisc del dev foo root
rmmod cls_u32
done
... will panic the kernel. Moreover, he bisected the change
apparently introducing it to 78fd1d0ab0 ("netlink: Re-add
locking to netlink_lookup() and seq walker").
The removal of synchronize_net() from the netlink socket
triggering the qdisc to be removed, seems to have uncovered
an RCU resp. module reference count race from the tc API.
Given that RCU conversion was done after e341694e3e ("netlink:
Convert netlink_lookup() to use RCU protected hash table")
which added the synchronize_net() originally, occasion of
hitting the bug was less likely (not impossible though):
When qdiscs that i) support attaching classifiers and,
ii) have at least one of them attached, get deleted, they
invoke tcf_destroy_chain(), and thus call into ->destroy()
handler from a classifier module.
After RCU conversion, all classifier that have an internal
prio list, unlink them and initiate freeing via call_rcu()
deferral.
Meanhile, tcf_destroy() releases already reference to the
tp->ops->owner module before the queued RCU callback handler
has been invoked.
Subsequent rmmod on the classifier module is then not prevented
since all module references are already dropped.
By the time, the kernel invokes the RCU callback handler from
the module, that function address is then invalid.
One way to fix it would be to add an rcu_barrier() to
unregister_tcf_proto_ops() to wait for all pending call_rcu()s
to complete.
synchronize_rcu() is not appropriate as under heavy RCU
callback load, registered call_rcu()s could be deferred
longer than a grace period. In case we don't have any pending
call_rcu()s, the barrier is allowed to return immediately.
Since we came here via unregister_tcf_proto_ops(), there
are no users of a given classifier anymore. Further nested
call_rcu()s pointing into the module space are not being
done anywhere.
Only cls_bpf_delete_prog() may schedule a work item, to
unlock pages eventually, but that is not in the range/context
of cls_bpf anymore.
Fixes: 25d8c0d55f ("net: rcu-ify tcf_proto")
Fixes: 9888faefe1 ("net: sched: cls_basic use RCU")
Reported-by: Vijay Subramanian <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: John Fastabend <[email protected]>
Cc: Eric Dumazet <[email protected]>
Cc: Thomas Graf <[email protected]>
Cc: Jamal Hadi Salim <[email protected]>
Cc: Alexei Starovoitov <[email protected]>
Tested-by: Vijay Subramanian <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Acked-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Reduce ifdef pollution slightly, no functional change. We can simply
remove the extra alternative definition of handle_ing() and nf_ingress().
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Pablo Neira Ayuso <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
This work as a follow-up of commit f7b3bec6f5 ("net: allow setting ecn
via routing table") and adds RFC3168 section 6.1.1.1. fallback for outgoing
ECN connections. In other words, this work adds a retry with a non-ECN
setup SYN packet, as suggested from the RFC on the first timeout:
[...] A host that receives no reply to an ECN-setup SYN within the
normal SYN retransmission timeout interval MAY resend the SYN and
any subsequent SYN retransmissions with CWR and ECE cleared. [...]
Schematic client-side view when assuming the server is in tcp_ecn=2 mode,
that is, Linux default since 2009 via commit 255cac91c3 ("tcp: extend
ECN sysctl to allow server-side only ECN"):
1) Normal ECN-capable path:
SYN ECE CWR ----->
<----- SYN ACK ECE
ACK ----->
2) Path with broken middlebox, when client has fallback:
SYN ECE CWR ----X crappy middlebox drops packet
(timeout, rtx)
SYN ----->
<----- SYN ACK
ACK ----->
In case we would not have the fallback implemented, the middlebox drop
point would basically end up as:
SYN ECE CWR ----X crappy middlebox drops packet
(timeout, rtx)
SYN ECE CWR ----X crappy middlebox drops packet
(timeout, rtx)
SYN ECE CWR ----X crappy middlebox drops packet
(timeout, rtx)
In any case, it's rather a smaller percentage of sites where there would
occur such additional setup latency: it was found in end of 2014 that ~56%
of IPv4 and 65% of IPv6 servers of Alexa 1 million list would negotiate
ECN (aka tcp_ecn=2 default), 0.42% of these webservers will fail to connect
when trying to negotiate with ECN (tcp_ecn=1) due to timeouts, which the
fallback would mitigate with a slight latency trade-off. Recent related
paper on this topic:
Brian Trammell, Mirja Kühlewind, Damiano Boppart, Iain Learmonth,
Gorry Fairhurst, and Richard Scheffenegger:
"Enabling Internet-Wide Deployment of Explicit Congestion Notification."
Proc. PAM 2015, New York.
http://ecn.ethz.ch/ecn-pam15.pdf
Thus, when net.ipv4.tcp_ecn=1 is being set, the patch will perform RFC3168,
section 6.1.1.1. fallback on timeout. For users explicitly not wanting this
which can be in DC use case, we add a net.ipv4.tcp_ecn_fallback knob that
allows for disabling the fallback.
tp->ecn_flags are not being cleared in tcp_ecn_clear_syn() on output, but
rather we let tcp_ecn_rcv_synack() take that over on input path in case a
SYN ACK ECE was delayed. Thus a spurious SYN retransmission will not prevent
ECN being negotiated eventually in that case.
Reference: https://www.ietf.org/proceedings/92/slides/slides-92-iccrg-1.pdf
Reference: https://www.ietf.org/proceedings/89/slides/slides-89-tsvarea-1.pdf
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: Florian Westphal <[email protected]>
Signed-off-by: Mirja Kühlewind <[email protected]>
Signed-off-by: Brian Trammell <[email protected]>
Cc: Eric Dumazet <[email protected]>
Cc: Dave That <[email protected]>
Acked-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Couple of torture test cases related to the bug fixed in 0b59d8806a
("ARM: net: delegate filter to kernel interpreter when imm_offset()
return value can't fit into 12bits.").
I've added a helper to allocate and fill the insn space. Output on
x86_64 from my laptop:
test_bpf: #233 BPF_MAXINSNS: Maximum possible literals jited:0 7 PASS
test_bpf: #234 BPF_MAXINSNS: Single literal jited:0 8 PASS
test_bpf: #235 BPF_MAXINSNS: Run/add until end jited:0 11553 PASS
test_bpf: #236 BPF_MAXINSNS: Too many instructions PASS
test_bpf: #237 BPF_MAXINSNS: Very long jump jited:0 9 PASS
test_bpf: #238 BPF_MAXINSNS: Ctx heavy transformations jited:0 20329 20398 PASS
test_bpf: #239 BPF_MAXINSNS: Call heavy transformations jited:0 32178 32475 PASS
test_bpf: #240 BPF_MAXINSNS: Jump heavy test jited:0 10518 PASS
test_bpf: #233 BPF_MAXINSNS: Maximum possible literals jited:1 4 PASS
test_bpf: #234 BPF_MAXINSNS: Single literal jited:1 4 PASS
test_bpf: #235 BPF_MAXINSNS: Run/add until end jited:1 1625 PASS
test_bpf: #236 BPF_MAXINSNS: Too many instructions PASS
test_bpf: #237 BPF_MAXINSNS: Very long jump jited:1 8 PASS
test_bpf: #238 BPF_MAXINSNS: Ctx heavy transformations jited:1 3301 3174 PASS
test_bpf: #239 BPF_MAXINSNS: Call heavy transformations jited:1 24107 23491 PASS
test_bpf: #240 BPF_MAXINSNS: Jump heavy test jited:1 8651 PASS
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Alexei Starovoitov <[email protected]>
Cc: Nicolas Schichan <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
This should have been #ifdef not #if.
Reported-by: Fengguang Wu <[email protected]>
Fixes: d2788d3488 ("net: sched: further simplify handle_ing")
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Ingress qdisc has no other purpose than calling into tc_classify()
that executes attached classifier(s) and action(s).
It has a 1:1 relationship to dev->ingress_queue. After having commit
087c1a601a ("net: sched: run ingress qdisc without locks") removed
the central ingress lock, one major contention point is gone.
The extra indirection layers however, are not necessary for calling
into ingress qdisc. pktgen calling locally into netif_receive_skb()
with a dummy u32, single CPU result on a Supermicro X10SLM-F, Xeon
E3-1240: before ~21,1 Mpps, after patch ~22,9 Mpps.
We can redirect the private classifier list to the netdev directly,
without changing any classifier API bits (!) and execute on that from
handle_ing() side. The __QDISC_STATE_DEACTIVATE test can be removed,
ingress qdisc doesn't have a queue and thus dev_deactivate_queue()
is also not applicable, ingress_cl_list provides similar behaviour.
In other words, ingress qdisc acts like TCQ_F_BUILTIN qdisc.
One next possible step is the removal of the dev's ingress (dummy)
netdev_queue, and to only have the list member in the netdevice
itself.
Note, the filter chain is RCU protected and individual filter elements
are being kfree'd by sched subsystem after RCU grace period. RCU read
lock is being held by __netif_receive_skb_core().
Joint work with Alexei Starovoitov.
Signed-off-by: Daniel Borkmann <[email protected]>
Signed-off-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Given quite some code has been removed from ing_filter(), we can just
consolidate that function into handle_ing() and get rid of a few
instructions at the same time.
Signed-off-by: Daniel Borkmann <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Seccomp has always been a special candidate when it comes to preparation
of its filters in seccomp_prepare_filter(). Due to the extra checks and
filter rewrite it partially duplicates code and has BPF internals exposed.
This patch adds a generic API inside the BPF code code that seccomp can use
and thus keep it's filter preparation code minimal and better maintainable.
The other side-effect is that now classic JITs can add seccomp support as
well by only providing a BPF_LDX | BPF_W | BPF_ABS translation.
Tested with seccomp and BPF test suites.
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Nicolas Schichan <[email protected]>
Cc: Alexei Starovoitov <[email protected]>
Cc: Kees Cook <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
When seccomp BPF was added, it was discussed to add __GFP_NOWARN
flag for their configuration path as f.e. up to 32K allocations are
more prone to fail under stress. As we're going to reuse BPF API,
add __GFP_NOWARN flags where larger kmalloc() and friends allocations
could fail.
It doesn't make much sense to pass around __GFP_NOWARN everywhere as
an extra argument only for seccomp while we just as well could run
into similar issues for socket filters, where it's not desired to
have a user application throw a WARN() due to allocation failure.
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Nicolas Schichan <[email protected]>
Cc: Alexei Starovoitov <[email protected]>
Cc: Kees Cook <[email protected]>
Acked-by: Alexei Starovoitov <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
In commit 0b053c9518 ("lib: memzero_explicit: use barrier instead
of OPTIMIZER_HIDE_VAR"), we made memzero_explicit() more robust in
case LTO would decide to inline memzero_explicit() and eventually
find out it could be elimiated as dead store.
While using barrier() works well for the case of gcc, recent efforts
from LLVMLinux people suggest to use llvm as an alternative to gcc,
and there, Stephan found in a simple stand-alone user space example
that llvm could nevertheless optimize and thus elimitate the memset().
A similar issue has been observed in the referenced llvm bug report,
which is regarded as not-a-bug.
Based on some experiments, icc is a bit special on its own, while it
doesn't seem to eliminate the memset(), it could do so with an own
implementation, and then result in similar findings as with llvm.
The fix in this patch now works for all three compilers (also tested
with more aggressive optimization levels). Arguably, in the current
kernel tree it's more of a theoretical issue, but imho, it's better
to be pedantic about it.
It's clearly visible with gcc/llvm though, with the below code: if we
would have used barrier() only here, llvm would have omitted clearing,
not so with barrier_data() variant:
static inline void memzero_explicit(void *s, size_t count)
{
memset(s, 0, count);
barrier_data(s);
}
int main(void)
{
char buff[20];
memzero_explicit(buff, sizeof(buff));
return 0;
}
$ gcc -O2 test.c
$ gdb a.out
(gdb) disassemble main
Dump of assembler code for function main:
0x0000000000400400 <+0>: lea -0x28(%rsp),%rax
0x0000000000400405 <+5>: movq $0x0,-0x28(%rsp)
0x000000000040040e <+14>: movq $0x0,-0x20(%rsp)
0x0000000000400417 <+23>: movl $0x0,-0x18(%rsp)
0x000000000040041f <+31>: xor %eax,%eax
0x0000000000400421 <+33>: retq
End of assembler dump.
$ clang -O2 test.c
$ gdb a.out
(gdb) disassemble main
Dump of assembler code for function main:
0x00000000004004f0 <+0>: xorps %xmm0,%xmm0
0x00000000004004f3 <+3>: movaps %xmm0,-0x18(%rsp)
0x00000000004004f8 <+8>: movl $0x0,-0x8(%rsp)
0x0000000000400500 <+16>: lea -0x18(%rsp),%rax
0x0000000000400505 <+21>: xor %eax,%eax
0x0000000000400507 <+23>: retq
End of assembler dump.
As gcc, clang, but also icc defines __GNUC__, it's sufficient to define
this in compiler-gcc.h only to be picked up. For a fallback or otherwise
unsupported compiler, we define it as a barrier. Similarly, for ecc which
does not support gcc inline asm.
Reference: https://llvm.org/bugs/show_bug.cgi?id=15495
Reported-by: Stephan Mueller <[email protected]>
Tested-by: Stephan Mueller <[email protected]>
Signed-off-by: Daniel Borkmann <[email protected]>
Cc: Theodore Ts'o <[email protected]>
Cc: Stephan Mueller <[email protected]>
Cc: Hannes Frederic Sowa <[email protected]>
Cc: mancha security <[email protected]>
Cc: Mark Charlebois <[email protected]>
Cc: Behan Webster <[email protected]>
Signed-off-by: Herbert Xu <[email protected]>
Daniel Borkmann