| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Add a new function ovs_ct_find_existing() to find an existing
conntrack entry for which this packet was already applied to. This is
only to be called when there is evidence that the packet was already
tracked and committed, but we lost the ct reference due to an
userspace upcall.
ovs_ct_find_existing() is called from skb_nfct_cached(), which can now
hide the fact that the ct reference may have been lost due to an
upcall. This allows ovs_ct_commit() to be simplified.
This patch is needed by later "openvswitch: Interface with NAT" patch,
as we need to be able to pass the packet through NAT using the
original ct reference also after the reference is lost after an
upcall.
Signed-off-by: Jarno Rajahalme <jarno@ovn.org>
Acked-by: Joe Stringer <joe@ovn.org>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Only a successful nf_conntrack_in() call can effect a connection state
change, so it suffices to update the key only after the
nf_conntrack_in() returns.
This change is needed for the later NAT patches.
Signed-off-by: Jarno Rajahalme <jarno@ovn.org>
Acked-by: Joe Stringer <joe@ovn.org>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| |
|
|
|
|
|
|
|
| |
This makes the code easier to understand and the following patches
more focused.
Signed-off-by: Jarno Rajahalme <jarno@ovn.org>
Acked-by: Joe Stringer <joe@ovn.org>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| |
|
|
|
|
|
|
|
|
|
| |
NAT checksum recalculation code assumes existence of skb_dst, which
becomes a problem for a later patch in the series ("openvswitch:
Interface with NAT."). Simplify this by removing the check on
skb_dst, as the checksum will be dealt with later in the stack.
Suggested-by: Pravin Shelar <pshelar@nicira.com>
Signed-off-by: Jarno Rajahalme <jarno@ovn.org>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| |
|
|
|
|
|
|
|
| |
Remove the definition of IP_CT_NEW_REPLY from the kernel as it does
not make sense. This allows the definition of IP_CT_NUMBER to be
simplified as well.
Signed-off-by: Jarno Rajahalme <jarno@ovn.org>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| |
|
|
|
|
|
|
|
|
|
| |
Ben Hawkes says:
integer overflow in xt_alloc_table_info, which on 32-bit systems can
lead to small structure allocation and a copy_from_user based heap
corruption.
Reported-by: Ben Hawkes <hawkes@google.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| |
|
|
|
|
|
|
|
| |
We copy according to ->target|matchsize, so check that the netlink attribute
(which can include padding and might be larger) contains enough data.
Reported-by: Julia Lawall <Julia.Lawall@lip6.fr>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| |\
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
https://git.kernel.org/pub/scm/linux/kernel/git/horms/ipvs
Simon Horman says:
====================
please consider these IPVS fixes for v4.5 or
if it is too late please consider them for v4.6.
* Arnd Bergman has corrected an error whereby the SIP persistence engine
may incorrectly access protocol fields
* Julian Anastasov has corrected a problem reported by Jiri Bohac with the
connection rescheduling mechanism added in 3.10 when new SYNs in
connection to dead real server can be redirected to another real server.
* Marco Angaroni resolved a problem in the SIP persistence engine
whereby the Call-ID could not be found if it was at the beginning of a
SIP message.
====================
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
The IPVS SIP persistence engine is not able to parse the SIP header
"Call-ID" when such header is inserted in the first positions of
the SIP message.
When IPVS is configured with "--pe sip" option, like for example:
ipvsadm -A -u 1.2.3.4:5060 -s rr --pe sip -p 120 -o
some particular messages (see below for details) do not create entries
in the connection template table, which can be listed with:
ipvsadm -Lcn --persistent-conn
Problematic SIP messages are SIP responses having "Call-ID" header
positioned just after message first line:
SIP/2.0 200 OK
[Call-ID header here]
[rest of the headers]
When "Call-ID" header is positioned down (after a few other headers)
it is correctly recognized.
This is due to the data offset used in get_callid function call inside
ip_vs_pe_sip.c file: since dptr already points to the start of the
SIP message, the value of dataoff should be initially 0.
Otherwise the header is searched starting from some bytes after the
first character of the SIP message.
Fixes: 758ff0338722 ("IPVS: sip persistence engine")
Signed-off-by: Marco Angaroni <marcoangaroni@gmail.com>
Acked-by: Julian Anastasov <ja@ssi.bg>
Signed-off-by: Simon Horman <horms@verge.net.au>
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
"RFC 5961, 4.2. Mitigation" describes a mechanism to request
client to confirm with RST the restart of TCP connection
before resending its SYN. As result, IPVS can see SYNs for
existing connection in CLOSE state. Add check to allow
rescheduling in this state.
Signed-off-by: Julian Anastasov <ja@ssi.bg>
Signed-off-by: Simon Horman <horms@verge.net.au>
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Jiri Bohac is reporting for a problem where the attempt
to reschedule existing connection to another real server
needs proper redirect for the conntrack used by the IPVS
connection. For example, when IPVS connection is created
to NAT-ed real server we alter the reply direction of
conntrack. If we later decide to select different real
server we can not alter again the conntrack. And if we
expire the old connection, the new connection is left
without conntrack.
So, the only way to redirect both the IPVS connection and
the Netfilter's conntrack is to drop the SYN packet that
hits existing connection, to wait for the next jiffie
to expire the old connection and its conntrack and to rely
on client's retransmission to create new connection as
usually.
Jiri Bohac provided a fix that drops all SYNs on rescheduling,
I extended his patch to do such drops only for connections
that use conntrack. Here is the original report from Jiri Bohac:
Since commit dc7b3eb900aa ("ipvs: Fix reuse connection if real server
is dead"), new connections to dead servers are redistributed
immediately to new servers. The old connection is expired using
ip_vs_conn_expire_now() which sets the connection timer to expire
immediately.
However, before the timer callback, ip_vs_conn_expire(), is run
to clean the connection's conntrack entry, the new redistributed
connection may already be established and its conntrack removed
instead.
Fix this by dropping the first packet of the new connection
instead, like we do when the destination server is not available.
The timer will have deleted the old conntrack entry long before
the first packet of the new connection is retransmitted.
Fixes: dc7b3eb900aa ("ipvs: Fix reuse connection if real server is dead")
Signed-off-by: Jiri Bohac <jbohac@suse.cz>
Signed-off-by: Julian Anastasov <ja@ssi.bg>
Signed-off-by: Simon Horman <horms@verge.net.au>
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
ip_vs_fill_iph_skb_off() may not find an IP header, and gcc has
determined that ip_vs_sip_fill_param() then incorrectly accesses
the protocol fields:
net/netfilter/ipvs/ip_vs_pe_sip.c: In function 'ip_vs_sip_fill_param':
net/netfilter/ipvs/ip_vs_pe_sip.c:76:5: error: 'iph.protocol' may be used uninitialized in this function [-Werror=maybe-uninitialized]
if (iph.protocol != IPPROTO_UDP)
^
net/netfilter/ipvs/ip_vs_pe_sip.c:81:10: error: 'iph.len' may be used uninitialized in this function [-Werror=maybe-uninitialized]
dataoff = iph.len + sizeof(struct udphdr);
^
This adds a check for the ip_vs_fill_iph_skb_off() return code
before looking at the ip header data returned from it.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Fixes: b0e010c527de ("ipvs: replace ip_vs_fill_ip4hdr with ip_vs_fill_iph_skb_off")
Acked-by: Julian Anastasov <ja@ssi.bg>
Signed-off-by: Simon Horman <horms@verge.net.au>
|
| | |
| |
| |
| |
| |
| |
| |
| |
| | |
nfacct_filter_alloc doesn't validate the NFACCT_FILTER_MASK and
NFACCT_FILTER_VALUE parameters which can trigger a NULL pointer
dereference. CAP_NET_ADMIN is required to trigger the bug.
Signed-off-by: Phil Turnbull <phil.turnbull@oracle.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| |\ \
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Jozsef Kadlecsik says:
====================
Please apply the next patch against the nf tree:
- Deniz Eren reported that parallel flush/dump of list:set type of sets
can lead to kernel crash. The bug was due to non-RCU compatible
flushing, listing in the set type, fixed by me.
- Julia Lawall pointed out that IPSET_ATTR_ETHER netlink attribute
length was not checked explicitly. The patch adds the missing
checkings.
====================
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| | | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Julia Lawall pointed out that IPSET_ATTR_ETHER netlink attribute length
was not checked explicitly, just for the maximum possible size. Malicious
netlink clients could send shorter attribute and thus resulting a kernel
read after the buffer.
The patch adds the explicit length checkings.
Reported-by: Julia Lawall <julia.lawall@lip6.fr>
Signed-off-by: Jozsef Kadlecsik <kadlec@blackhole.kfki.hu>
|
| | | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Flushing/listing entries was not RCU safe, so parallel flush/dump
could lead to kernel crash. Bug reported by Deniz Eren.
Fixes netfilter bugzilla id #1050.
Signed-off-by: Jozsef Kadlecsik <kadlec@blackhole.kfki.hu>
|
| |\ \ \
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Willem de Bruijn says:
====================
net: validate variable length ll headers
Allow device-specific validation of link layer headers. Existing
checks drop all packets shorter than hard_header_len. For variable
length protocols, such packets can be valid.
patch 1 adds header_ops.validate and dev_validate_header
patch 2 implements the protocol specific callback for AX25
patch 3 replaces ll_header_truncated with dev_validate_header
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Replace link layer header validation check ll_header_truncate with
more generic dev_validate_header.
Validation based on hard_header_len incorrectly drops valid packets
in variable length protocols, such as AX25. dev_validate_header
calls header_ops.validate for such protocols to ensure correctness
below hard_header_len.
See also http://comments.gmane.org/gmane.linux.network/401064
Fixes 9c7077622dd9 ("packet: make packet_snd fail on len smaller than l2 header")
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
As variable length protocol, AX25 fails link layer header validation
tests based on a minimum length. header_ops.validate allows protocols
to validate headers that are shorter than hard_header_len. Implement
this callback for AX25.
See also http://comments.gmane.org/gmane.linux.network/401064
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| |/ / /
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Netdevice parameter hard_header_len is variously interpreted both as
an upper and lower bound on link layer header length. The field is
used as upper bound when reserving room at allocation, as lower bound
when validating user input in PF_PACKET.
Clarify the definition to be maximum header length. For validation
of untrusted headers, add an optional validate member to header_ops.
Allow bypassing of validation by passing CAP_SYS_RAWIO, for instance
for deliberate testing of corrupt input. In this case, pad trailing
bytes, as some device drivers expect completely initialized headers.
See also http://comments.gmane.org/gmane.linux.network/401064
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| |\ \ \
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Tom Herbert says:
====================
kcm: Kernel Connection Multiplexor (KCM)
Kernel Connection Multiplexor (KCM) is a facility that provides a
message based interface over TCP for generic application protocols.
The motivation for this is based on the observation that although
TCP is byte stream transport protocol with no concept of message
boundaries, a common use case is to implement a framed application
layer protocol running over TCP. To date, most TCP stacks offer
byte stream API for applications, which places the burden of message
delineation, message I/O operation atomicity, and load balancing
in the application. With KCM an application can efficiently send
and receive application protocol messages over TCP using a
datagram interface.
In order to delineate message in a TCP stream for receive in KCM, the
kernel implements a message parser. For this we chose to employ BPF
which is applied to the TCP stream. BPF code parses application layer
messages and returns a message length. Nearly all binary application
protocols are parsable in this manner, so KCM should be applicable
across a wide range of applications. Other than message length
determination in receive, KCM does not require any other application
specific awareness. KCM does not implement any other application
protocol semantics-- these are are provided in userspace or could be
implemented in a kernel module layered above KCM.
KCM implements an NxM multiplexor in the kernel as diagrammed below:
+------------+ +------------+ +------------+ +------------+
| KCM socket | | KCM socket | | KCM socket | | KCM socket |
+------------+ +------------+ +------------+ +------------+
| | | |
+-----------+ | | +----------+
| | | |
+----------------------------------+
| Multiplexor |
+----------------------------------+
| | | | |
+---------+ | | | ------------+
| | | | |
+----------+ +----------+ +----------+ +----------+ +----------+
| Psock | | Psock | | Psock | | Psock | | Psock |
+----------+ +----------+ +----------+ +----------+ +----------+
| | | | |
+----------+ +----------+ +----------+ +----------+ +----------+
| TCP sock | | TCP sock | | TCP sock | | TCP sock | | TCP sock |
+----------+ +----------+ +----------+ +----------+ +----------+
The KCM sockets provide the datagram interface to applications,
Psocks are the state for each attached TCP connection (i.e. where
message delineation is performed on receive).
A description of the APIs and design can be found in the included
Documentation/networking/kcm.txt.
In this patch set:
- Add MSG_BATCH flag. This is used in sendmsg msg_hdr flags to
indicate that more messages will be sent on the socket. The stack
may batch messages up if it is beneficial for transmission.
- In sendmmsg, set MSG_BATCH in all sub messages except for the last
one.
- In order to allow sendmmsg to contain multiple messages with
SOCK_SEQPAKET we allow each msg_hdr in the sendmmsg to set MSG_EOR.
- Add KCM module
- This supports SOCK_DGRAM and SOCK_SEQPACKET.
- KCM documentation
v2:
- Added splice and page operations.
- Assemble receive messages in place on TCP socket (don't have a
separate assembly queue.
- Based on above, enforce maxmimum receive message to be the size
of the recceive socket buffer.
- Support message assembly timeout. Use the timeout value in
sk_rcvtimeo on the TCP socket.
- Tested some with a couple of other production applications,
see ~5% improvement in application latency.
Testing:
Dave Watson has integrated KCM into Thrift and we intend to put these
changes into open source. Example of this is in:
https://github.com/djwatson/fbthrift/commit/
dd7e0f9cf4e80912fdb90f6cd394db24e61a14cc
Some initial KCM Thrift benchmark numbers (comment from Dave)
Thrift by default ties a single connection to a single thread. KCM is
instead able to load balance multiple connections across multiple epoll
loops easily.
A test sending ~5k bytes of data to a kcm thrift server, dropping the
bytes on recv:
QPS Latency / std dev Latency
without KCM
70336 209/123
with KCM
70353 191/124
A test sending a small request, then doing work in the epoll thread,
before serving more requests:
QPS Latency / std dev Latency
without KCM
14282 559/602
with KCM
23192 344/234
At the high end, there's definitely some additional kernel overhead:
Cranking the pipelining way up, with lots of small requests
QPS Latency / std dev Latency
without KCM
1863429 127/119
with KCM
1337713 192/241
---
So for a "realistic" workload, KCM performs pretty well (second case).
Under extreme conditions of highest tps we still have some work to do.
In its nature a multiplexor will spread work between CPUs which is
logically good for load balancing but coan conflict with the goal
promoting affinity. Batching messages on both send and receive are
the means to recoup performance.
Future support:
- Integration with TLS (TLS-in-kernel is a separate initiative).
- Page operations/splice support
- Unconnected KCM sockets. Will be able to attach sockets to different
destinations, AF_KCM addresses with be used in sendmsg and recvmsg
to indicate destination
- Explore more utility in performing BPF inline with a TCP data stream
(setting SO_MARK, rxhash for messages being sent received on
KCM sockets).
- Performance work
- Diagnose performance issues under high message load
FAQ (Questions posted on LWN)
Q: Why do this in the kernel?
A: Because the kernel is good at scheduling threads and steering packets
to threads. KCM fits well into this model since it allows the unit
of work for scheduling and steering to be the application layer
messages themselves. KCM should be thought of as generic application
protocol acceleration. It to the philosophy that the kernel provides
generic and extensible interfaces.
Q: How can adding code in the path yield better performance?
A: It is true that for just sending receiving a single message there
would be some performance loss since the code path is longer (for
instance comparing netperf to KCM). But for real production
applications performance takes on many dynamics. Parallelism, context
switching, affinity, granularity of locking, and load balancing are
all relevant. The theory of KCM is that by an application-centric
interface, the kernel can provide better support for these
performance characteristics.
Q: Why not use an existing message-oriented protocol such as RUDP,
DCCP, SCTP, RDS, and others?
A: Because that would entail using a completely new transport protocol.
Deploying a new protocol at scale is either a huge undertaking or
fundamentally infeasible. This is true in either the Internet and in
the data center due in a large part to protocol ossification.
Besides, KCM we want KCM to work existing, well deployed application
protocols that we couldn't change even if we wanted to (e.g. http/2).
KCM simply defines a new interface method, it does not redefine any
aspect of the transport protocol nor application protocol, nor set
any new requirements on these. Neither does KCM attempt to implement
any application protocol logic other than message deliniation in the
stream. These are fundamental requirement of KCM.
Q: How does this affect TCP?
A: It doesn't, not in the slightest. The use of KCM can be one-sided,
KCM has no effect on the wire.
Q: Why force TCP into doing something it's not designed for?
A: TCP is defined as transport protocol and there is no standard that
says the API into TCP must be stream based sockets, or for that
matter sockets at all (or even that TCP needs to be implemented in a
kernel). KCM is not inconsistent with the design of TCP just because
to makes an message based interface over TCP, if it were then every
application protocol sending messages over TCP would also be! :-)
Q: What about the problem of a connections with very slow rate of
incoming data? As a result your application can get storms of very
short reads. And it actually happens a lot with connection from
mobile devices and it is a problem for servers handling a lot of
connections.
A: The storm of short reads will occur regardless of whether KCM is used
or not. KCM does have one advantage in this scenario though, it will
only wake up the application when a full message has been received,
not for each packet that makes up part of a bigger messages. If a
bunch of small messages are received, the application can receive
messages in batches using recvmmsg.
Q: Why not just use DPDK, or at least provide KCM like functionality in
DPDK?
A: DPDK, or more generally OS bypass presumably with a TCP stack in
userland, presents a different model of load balancing than that of
KCM (and the kernel). KCM implements load balancing of messages
across the threads of an application, whereas DPDK load balances
based on queues which are more static and coarse-grained since
multiple connections are bound to queues. DPDK works best when
processing of packets is silo'ed in a thread on the CPU processing
a queue, and packet processing (for both the stack and application)
is fairly uniform. KCM works well for applications where the amount
of work to process messages varies an application work is commonly
delegated to worker threads often on different CPUs.
The message based interface over TCP is something that could be
provide by a DPDK or OS bypass library.
Q: I'm not quite seeing this for HTTP. Maybe for HTTP/2, I guess, or web
sockets?
A: Yes. KCM is most appropriate for message based protocols over TCP
where is easy to deduce the message length (e.g. a length field)
and the protocol implements its own message ordering semantics.
Fortunately this encompasses many modern protocols.
Q: How is memory limited and controlled?
A: In v2 all data for messages is now kept in socket buffers, either
those for TCP or KCM, so socket buffer limits are applicable.
This includes receive messages assembly which is now done ont teh
TCP socket buffer instead of a separate queue-- this has the
consequence that the TCP socket buffer limit provides an
enforceable maxmimum message size.
Additionally, a timeout may be set for messages assembly. The
value used for this is taken from sk_rcvtimeo of the TCP socket.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Add kcm.txt to desribe KCM and interfaces.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
This patch adds receive timeout for message assembly on the attached TCP
sockets. The timeout is set when a new messages is started and the whole
message has not been received by TCP (not in the receive queue). If the
completely message is subsequently received the timer is cancelled, if the
timer expires the RX side is aborted.
The timeout value is taken from the socket timeout (SO_RCVTIMEO) that is
set on a TCP socket (i.e. set by get sockopt before attaching a TCP socket
to KCM.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Message assembly is performed on the TCP socket. This is logically
equivalent of an application that performs a peek on the socket to find
out how much memory is needed for a receive buffer. The receive socket
buffer also provides the maximum message size which is checked.
The receive algorithm is something like:
1) Receive the first skbuf for a message (or skbufs if multiple are
needed to determine message length).
2) Check the message length against the number of bytes in the TCP
receive queue (tcp_inq()).
- If all the bytes of the message are in the queue (incluing the
skbuf received), then proceed with message assembly (it should
complete with the tcp_read_sock)
- Else, mark the psock with the number of bytes needed to
complete the message.
3) In TCP data ready function, if the psock indicates that we are
waiting for the rest of the bytes of a messages, check the number
of queued bytes against that.
- If there are still not enough bytes for the message, just
return
- Else, clear the waiting bytes and proceed to receive the
skbufs. The message should now be received in one
tcp_read_sock
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Implement kcm_sendpage. Set in sendpage to kcm_sendpage in both
dgram and seqpacket ops.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Implement kcm_splice_read. This is supported only for seqpacket.
Add kcm_seqpacket_ops and set splice read to kcm_splice_read.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
This patch adds various counters for KCM. These include counters for
messages and bytes received or sent, as well as counters for number of
attached/unattached TCP sockets and other error or edge events.
The statistics are exposed via a proc interface. /proc/net/kcm provides
statistics per KCM socket and per psock (attached TCP sockets).
/proc/net/kcm_stats provides aggregate statistics.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
This module implements the Kernel Connection Multiplexor.
Kernel Connection Multiplexor (KCM) is a facility that provides a
message based interface over TCP for generic application protocols.
With KCM an application can efficiently send and receive application
protocol messages over TCP using datagram sockets.
For more information see the included Documentation/networking/kcm.txt
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Create a common kernel function to get the number of bytes available
on a TCP socket. This is based on code in INQ getsockopt and we now call
the function for that getsockopt.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Add walking of fragments in __skb_splice_bits.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Add a new msg flag called MSG_BATCH. This flag is used in sendmsg to
indicate that more messages will follow (i.e. a batch of messages is
being sent). This is similar to MSG_MORE except that the following
messages are not merged into one packet, they are sent individually.
sendmmsg is updated so that each contained message except for the
last one is marked as MSG_BATCH.
MSG_BATCH is a performance optimization in cases where a socket
implementation can benefit by transmitting packets in a batch.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
This patch allows setting MSG_EOR in each individual msghdr passed
in sendmmsg. This allows a sendmmsg to send multiple messages when
using SOCK_SEQPACKET.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Export it for cases where we want to create sockets by hand.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| |/ / /
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
This is a convenience function that returns the next entry in an RCU
list or NULL if at the end of the list.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
performance tests for hash map and per-cpu hash map
with and without pre-allocation
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
increase stress by also calling bpf_get_stackid() from
various *spin* functions
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
this test calls bpf programs from different contexts:
from inside of slub, from rcu, from pretty much everywhere,
since it kprobes all spin_lock functions.
It stresses the bpf hash and percpu map pre-allocation,
deallocation logic and call_rcu mechanisms.
User space part adding more stress by walking and deleting map elements.
Note that due to nature bpf_load.c the earlier kprobe+bpf programs are
already active while loader loads new programs, creates new kprobes and
attaches them.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
Helpers like ip_tunnel_info_opts_{get,set}() are only available if
CONFIG_INET is set, thus add an empty definition into the header for
the !CONFIG_INET case, where already other empty inline helpers are
defined.
This avoids ifdef kludge inside filter.c, but also vxlan and geneve
themself where this facility can only be used with, depend on INET
being set. For the !INET case TUNNEL_OPTIONS_PRESENT would never be
set in flags.
Fixes: 14ca0751c96f ("bpf: support for access to tunnel options")
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| |\ \ \
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Alexei Starovoitov says:
====================
bpf: map pre-alloc
v1->v2:
. fix few issues spotted by Daniel
. converted stackmap into pre-allocation as well
. added a workaround for lockdep false positive
. added pcpu_freelist_populate to be used by hashmap and stackmap
this path set switches bpf hash map to use pre-allocation by default
and introduces BPF_F_NO_PREALLOC flag to keep old behavior for cases
where full map pre-allocation is too memory expensive.
Some time back Daniel Wagner reported crashes when bpf hash map is
used to compute time intervals between preempt_disable->preempt_enable
and recently Tom Zanussi reported a dead lock in iovisor/bcc/funccount
tool if it's used to count the number of invocations of kernel
'*spin*' functions. Both problems are due to the recursive use of
slub and can only be solved by pre-allocating all map elements.
A lot of different solutions were considered. Many implemented,
but at the end pre-allocation seems to be the only feasible answer.
As far as pre-allocation goes it also was implemented 4 different ways:
- simple free-list with single lock
- percpu_ida with optimizations
- blk-mq-tag variant customized for bpf use case
- percpu_freelist
For bpf style of alloc/free patterns percpu_freelist is the best
and implemented in this patch set.
Detailed performance numbers in patch 3.
Patch 2 introduces percpu_freelist
Patch 1 fixes simple deadlocks due to missing recursion checks
Patch 5: converts stackmap to pre-allocation
Patches 6-9: prepare test infra
Patch 10: stress test for hash map infra. It attaches to spin_lock
functions and bpf_map_update/delete are called from different contexts
Patch 11: stress for bpf_get_stackid
Patch 12: map performance test
Reported-by: Daniel Wagner <daniel.wagner@bmw-carit.de>
Reported-by: Tom Zanussi <tom.zanussi@linux.intel.com>
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
extend test coveraged to include pre-allocated and run-time alloc maps
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
note old loader is compatible with new kernel.
map_flags are optional
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
move ksym search from offwaketime into library to be reused
in other tests
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
map creation is typically the first one to fail when rlimits are
too low, not enough memory, etc
Make this failure scenario more verbose
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
It was observed that calling bpf_get_stackid() from a kprobe inside
slub or from spin_unlock causes similar deadlock as with hashmap,
therefore convert stackmap to use pre-allocated memory.
The call_rcu is no longer feasible mechanism, since delayed freeing
causes bpf_get_stackid() to fail unpredictably when number of actual
stacks is significantly less than user requested max_entries.
Since elements are no longer freed into slub, we can push elements into
freelist immediately and let them be recycled.
However the very unlikley race between user space map_lookup() and
program-side recycling is possible:
cpu0 cpu1
---- ----
user does lookup(stackidX)
starts copying ips into buffer
delete(stackidX)
calls bpf_get_stackid()
which recyles the element and
overwrites with new stack trace
To avoid user space seeing a partial stack trace consisting of two
merged stack traces, do bucket = xchg(, NULL); copy; xchg(,bucket);
to preserve consistent stack trace delivery to user space.
Now we can move memset(,0) of left-over element value from critical
path of bpf_get_stackid() into slow-path of user space lookup.
Also disallow lookup() from bpf program, since it's useless and
program shouldn't be messing with collected stack trace.
Note that similar race between user space lookup and kernel side updates
is also present in hashmap, but it's not a new race. bpf programs were
always allowed to modify hash and array map elements while user space
is copying them.
Fixes: d5a3b1f69186 ("bpf: introduce BPF_MAP_TYPE_STACK_TRACE")
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Suggested-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
If kprobe is placed on spin_unlock then calling kmalloc/kfree from
bpf programs is not safe, since the following dead lock is possible:
kfree->spin_lock(kmem_cache_node->lock)...spin_unlock->kprobe->
bpf_prog->map_update->kmalloc->spin_lock(of the same kmem_cache_node->lock)
and deadlocks.
The following solutions were considered and some implemented, but
eventually discarded
- kmem_cache_create for every map
- add recursion check to slow-path of slub
- use reserved memory in bpf_map_update for in_irq or in preempt_disabled
- kmalloc via irq_work
At the end pre-allocation of all map elements turned out to be the simplest
solution and since the user is charged upfront for all the memory, such
pre-allocation doesn't affect the user space visible behavior.
Since it's impossible to tell whether kprobe is triggered in a safe
location from kmalloc point of view, use pre-allocation by default
and introduce new BPF_F_NO_PREALLOC flag.
While testing of per-cpu hash maps it was discovered
that alloc_percpu(GFP_ATOMIC) has odd corner cases and often
fails to allocate memory even when 90% of it is free.
The pre-allocation of per-cpu hash elements solves this problem as well.
Turned out that bpf_map_update() quickly followed by
bpf_map_lookup()+bpf_map_delete() is very common pattern used
in many of iovisor/bcc/tools, so there is additional benefit of
pre-allocation, since such use cases are must faster.
Since all hash map elements are now pre-allocated we can remove
atomic increment of htab->count and save few more cycles.
Also add bpf_map_precharge_memlock() to check rlimit_memlock early to avoid
large malloc/free done by users who don't have sufficient limits.
Pre-allocation is done with vmalloc and alloc/free is done
via percpu_freelist. Here are performance numbers for different
pre-allocation algorithms that were implemented, but discarded
in favor of percpu_freelist:
1 cpu:
pcpu_ida 2.1M
pcpu_ida nolock 2.3M
bt 2.4M
kmalloc 1.8M
hlist+spinlock 2.3M
pcpu_freelist 2.6M
4 cpu:
pcpu_ida 1.5M
pcpu_ida nolock 1.8M
bt w/smp_align 1.7M
bt no/smp_align 1.1M
kmalloc 0.7M
hlist+spinlock 0.2M
pcpu_freelist 2.0M
8 cpu:
pcpu_ida 0.7M
bt w/smp_align 0.8M
kmalloc 0.4M
pcpu_freelist 1.5M
32 cpu:
kmalloc 0.13M
pcpu_freelist 0.49M
pcpu_ida nolock is a modified percpu_ida algorithm without
percpu_ida_cpu locks and without cross-cpu tag stealing.
It's faster than existing percpu_ida, but not as fast as pcpu_freelist.
bt is a variant of block/blk-mq-tag.c simlified and customized
for bpf use case. bt w/smp_align is using cache line for every 'long'
(similar to blk-mq-tag). bt no/smp_align allocates 'long'
bitmasks continuously to save memory. It's comparable to percpu_ida
and in some cases faster, but slower than percpu_freelist
hlist+spinlock is the simplest free list with single spinlock.
As expeceted it has very bad scaling in SMP.
kmalloc is existing implementation which is still available via
BPF_F_NO_PREALLOC flag. It's significantly slower in single cpu and
in 8 cpu setup it's 3 times slower than pre-allocation with pcpu_freelist,
but saves memory, so in cases where map->max_entries can be large
and number of map update/delete per second is low, it may make
sense to use it.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Introduce simple percpu_freelist to keep single list of elements
spread across per-cpu singly linked lists.
/* push element into the list */
void pcpu_freelist_push(struct pcpu_freelist *, struct pcpu_freelist_node *);
/* pop element from the list */
struct pcpu_freelist_node *pcpu_freelist_pop(struct pcpu_freelist *);
The object is pushed to the current cpu list.
Pop first trying to get the object from the current cpu list,
if it's empty goes to the neigbour cpu list.
For bpf program usage pattern the collision rate is very low,
since programs push and pop the objects typically on the same cpu.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| |/ / /
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
if kprobe is placed within update or delete hash map helpers
that hold bucket spin lock and triggered bpf program is trying to
grab the spinlock for the same bucket on the same cpu, it will
deadlock.
Fix it by extending existing recursion prevention mechanism.
Note, map_lookup and other tracing helpers don't have this problem,
since they don't hold any locks and don't modify global data.
bpf_trace_printk has its own recursive check and ok as well.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| |\ \ \
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
Michal Kubecek says:
====================
ipv6: per netns FIB6 walkers and garbage collector
Commit 2ac3ac8f86f2 ("ipv6: prevent fib6_run_gc() contention") reduced
the risk of contention on FIB6 garbage collector lock on systems with
many CPUs. However, one of our customers can still observe heavy
contention on fib6_gc_lock which can even trigger the soft lockup
detector.
This is caused by garbage collector running in forced mode from a timer.
While there is one timer per network namespace, the instances of
fib6_run_gc() running from them are protected by one global spinlock so
that only one garbage collector can run at any moment and other
namespaces have to wait. As most relevant data structures are separated
per netns, there is little reason for garbage collectors blocking each
other.
Similar problem exists for walkers: changes in one tree do not need to
adjust (and block) walkers traversing FIB trees in other namespaces.
This series separates both the walkers infrastructure and garbage
collector so that they work independently in network namespaces.
v2: get rid of ifdef in ipv6_route_seq_setup_walk(), pass net from
callers instead
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
|
| | | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | |
| | | | |
One of our customers observed issues with FIB6 garbage collectors
running in different network namespaces blocking each other, resulting
in soft lockups (fib6_run_gc() initiated from timer runs always in
forced mode).
Now that FIB6 walkers are separated per namespace, there is no more need
for instances of fib6_run_gc() in different namespaces blocking each
other. There is still a call to icmp6_dst_gc() which operates on shared
data but this function is protected by its own shared lock.
Signed-off-by: Michal Kubecek <mkubecek@suse.cz>
Reviewed-by: Cong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
