Currently we deal with hash collisions by letting a hash bucket contain
multiple entries, forming a linked list using an index in the connection
structure.
That's a pretty standard and simple approach, but in our case we can use
an even simpler one: linear probing. Here if a hash bucket is occupied
we just move onto the next one until we find a feww one. This slightly
simplifies lookup and more importantly saves some precious bytes in the
connection structure by removing the need for a link. It does require some
additional complexity for hash removal.
This approach can perform poorly with hash table load is high. However, we
already size our hash table of pointers larger than the connection table,
which puts an upper bound on the load. It's relatively cheap to decrease
that bound if we find we need to.
I adapted the linear probing operations from Knuth's The Art of Computer
Programming, Volume 3, 2nd Edition. Specifically Algorithm L and Algorithm
R in Section 6.4. Note that there is an error in Algorithm R as printed,
see errata at [0].
[0] https://www-cs-faculty.stanford.edu/~knuth/all3-prepre.ps.gz
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
tcp_hash_lookup() expects the port numbers in host order, but the TCP
header, of course, has them in network order, so we need to switch them.
However we call htons() (host to network) instead of ntohs() (network to
host). This works because those do the same thing in practice (they only
wouldn't on very strange theoretical platforms which are neither big nor
little endian).
But, having this the "wrong" way around is misleading, so switch it around.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
On x32, glibc defines time_t and suseconds_t (the latter, also known as
__syscall_slong_t) as unsigned long long, whereas "everywhere else",
including x86_64 and i686, those are unsigned long.
See also https://sourceware.org/bugzilla/show_bug.cgi?id=16437 for
all the gory details.
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
When a TCP connection is closed, we mark it by setting events to CLOSED,
then some time later we do final cleanups: closing sockets, removing from
the hash table and so forth.
This does mean that when making a hash lookup we need to exclude any
apparent matches that are CLOSED, since they represent a stale connection.
This can happen in practice if one connection closes and a new one with the
same endpoints is started shortly afterward.
Checking for CLOSED is quite specific to TCP however, and won't work when
we extend the hash table to more general flows. So, alter the code to
immediately remove the connection from the hash table when CLOSED, although
we still defer closing sockets and other cleanup.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
The TCP state structure includes a 128-bit hash_secret which we use for
SipHash calculations to mitigate attacks on the TCP hash table and initial
sequence number.
We have plans to use SipHash in places that aren't TCP related, and there's
no particular reason they'd need their own secret. So move the hash_secret
to the general context structure.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Currently TCP uses the 'flow' epoll_ref field for both connected
sockets and timers, which consists of just the index of the relevant
flow (connection).
This is just fine for timers, for while it obviously works, it's
subtly incomplete for sockets on spliced connections. In that case we
want to know which side of the connection the event is occurring on as
well as which connection. At present, we deduce that information by
looking at the actual fd, and comparing it to the fds of the sockets
on each side.
When we use the flow table for more things, we expect more cases where
something will need to know a specific side of a specific flow for an
event, but nothing more.
Therefore add a new 'flowside' epoll_ref field, with exactly that
information. We use it for TCP connected sockets. This allows us to
directly know the side for spliced connections. For "tap"
connections, it's pretty meaningless, since the side is always the
socket side. It still makes logical sense though, and it may become
important for future flow table work.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
TCP uses three different epoll object types: one for connected sockets, one
for timers and one for listening sockets. Listening sockets really need
information that's specific to TCP, so need their own epoll_ref field.
Timers and connected sockets, however, only need the connection (flow)
they're associated with. As we expand the use of the flow table, we expect
that to be true for more epoll fds. So, rename the "TCP" epoll_ref field
to be a "flow" epoll_ref field that can be used both for TCP and for other
future cases.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
In tcp_timer_handler() we use conn_at_idx() to interpret the flow index
from the epoll reference. However, this will never be NULL - we always
put a valid index into the epoll_ref. Simplify slightly based on this.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Most of the messages logged by the TCP code (be they errors, debug or
trace messages) are related to a specific connection / flow. We're fairly
consistent about prefixing these with the type of connection and the
connection / flow index. However there are a few places where we put the
index later in the message or omit it entirely. The template with the
prefix is also a little bulky to carry around for every message,
particularly for spliced connections.
To help keep this consistent, introduce some helpers to log messages
linked to a specific flow. It takes the flow as a parameter and adds a
uniform prefix to each message. This makes things slightly neater now, but
more importantly will help keep formatting consistent as we add more things
to the flow table.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
tcp_table_compact() will move entries in the connection/flow table to keep
it compact when other entries are removed. The moved entries need not have
the same type as the flow removed, so it needs to be able to handle moving
any type of flow. Therefore, move it to flow.c rather than being
purportedly TCP specific.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Both tcp.c and tcp_splice.c define CONN_IDX() variants to find the index
of their connection structures in the connection table, now become the
unified flow table. We can easily combine these into a common helper.
While we're there, add some trickery for some additional type safety.
They also define their own CONN() versions, which aren't so easily combined
since they need to return different types, but we can have them use a
common helper.
In the process, we standardise on always using an unsigned type to store
the connection / flow index, which makes more sense. tcp.c's conn_at_idx()
remains for now, but we change its parameter to unsigned to match. That in
turn means we can remove a check for negative values from it.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
We want to generalise "connection" tracking to things other than true TCP
connections. Continue implenenting this by renaming the TCP connection
table to the "flow table" and moving it to flow.c. The definitions are
split between flow.h and flow_table.h - we need this separation to avoid
circular dependencies: the definitions in flow.h will be needed by many
headers using the flow mechanism, but flow_table.h needs all those protocol
specific headers in order to define the full flow table entry.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Currently TCP connections use a 1-bit selector, 'spliced', to determine the
rest of the contents of the structure. We want to generalise the TCP
connection table to other types of flows in other protocols. Make a start
on this by replacing the tcp_conn_common structure with a new flow_common
structure with an enum rather than a simple boolean indicating the type of
flow.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Types size_t and ssize_t are not necessarily long, it depends on the
architecture.
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
valgrind complains if we pass a NULL buffer to recv(), even if we use
MSG_TRUNC, in which case it's actually safe. For a long time we've had
a valgrind suppression for this. It singles out the recv() in
tcp_sock_consume(), the only place we use MSG_TRUNC.
However, tcp_sock_consume() only has a single caller, which makes it a
prime candidate for inlining. If inlined, it won't appear on the stack and
valgrind won't match the correct suppression.
It appears that certain compiler versions (for example gcc-13.2.1 in
Fedora 39) will inline this function even with the -O0 we use for valgrind
builds. This breaks the suppression leading to a spurious failure in the
tests.
There's not really any way to adjust the suppression itself without making
it overly broad (we don't want to match other recv() calls). So, as a hack
explicitly prevent inlining of this function when we're making a valgrind
build. To accomplish this add an explicit -DVALGRIND when making such a
build.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
tcp_port_rebind() is desgined to be called from NS_CALL() and has two
disjoint cases: one where it enters the namespace (outbound forwards) and
one where it doesn't (inbound forwards).
We only actually need the NS_CALL() framing for the outbound case, for
inbound we can just call tcp_port_do_rebind() directly. So simplify
tcp_port_rebind() to tcp_port_rebind_outbound(), allowing us to eliminate
an awkward parameters structure.
With that done we can safely rename tcp_port_do_rebind() to
tcp_port_rebind() for brevity.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
tcp_port_rebind() has two cases with almost but not quite identical code.
Simplify things a bit by factoring this out into a single parameterised
helper, tcp_port_do_rebind().
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
On the L2 tap side, we see TCP headers and know the TCP window that the
ultimate receiver is advertising. In order to avoid unnecessary buffering
within passt/pasta (or by the kernel on passt/pasta's behalf) we attempt
to advertise that window back to the original sock-side sender using
TCP_WINDOW_CLAMP.
However, TCP_WINDOW_CLAMP just doesn't work like this. Prior to kernel
commit 3aa7857fe1d7 ("tcp: enable mid stream window clamp"), it simply
had no effect on established sockets. After that commit, it does affect
established sockets but doesn't behave the way we need:
* It appears to be designed only to shrink the window, not to allow it to
re-expand.
* More importantly, that commit has a serious bug where if the
setsockopt() is made when the existing kernel advertised window for the
socket happens to be zero, it will now become locked at zero, stopping
any further data from being received on the socket.
Since this has never worked as intended, simply remove it. It might be
possible to re-implement the intended behaviour by manipulating SO_RCVBUF,
so we leave a comment to that effect.
This kernel bug is the underlying cause of both the linked passt bug and
the linked podman bug. We attempted to fix this before with passt commit
d3192f67 ("tcp: Force TCP_WINDOW_CLAMP before resetting STALLED flag").
However while that commit masked the bug for some cases, it didn't really
address the problem.
Fixes: d3192f67c4 ("tcp: Force TCP_WINDOW_CLAMP before resetting STALLED flag")
Link: https://github.com/containers/podman/issues/20170
Link: https://bugs.passt.top/show_bug.cgi?id=74
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
tcp_clamp_window() is _mostly_ about using TCP_WINDOW_CLAMP to control the
sock side advertised window, but it is also responsible for actually
updating the conn->wnd_from_tap value.
Rename to tcp_tap_window_update() to reflect that broader purpose, and pull
the logic that's not TCP_WINDOW_CLAMP related out to the front.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
logmsg() takes printf like arguments, but because it's not a built in, the
compiler won't generate warnings if the format string and parameters don't
match. Enable those by using the format attribute.
Strictly speaking this is a gcc extension, but I believe it is also
supported by some other common compilers. We already use some other
attributes in various places. For now, just use it and we can worry about
compilers that don't support it if it comes up.
This exposes some warnings from existing callers, both in gcc and in
clang-tidy:
- Some are straight out bugs, which we correct
- It's occasionally useful to invoke the logging functions with an empty
string, which gcc objects to, so disable that specific warning in the
Makefile
- Strictly speaking the C standard requires that the parameter for a %p
be a (void *), not some other pointer type. That's only likely to cause
problems in practice on weird architectures with different sized
representations for pointers to different types. Nonetheless add the
casts to make it happy.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Use of tcp_l2_mh has been removed in commit 38fbfdbcb9,
but its declaration and initialization are always in the code.
Remove them as they are useless.
Fixes: 38fbfdbcb9 ("tcp: Get rid of iov with cached MSS, drop sendmmsg(), add deferred flush")
Signed-off-by: Laurent Vivier <lvivier@redhat.com>
Acked-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
For now, packets passed to the various *_tap_handler() functions always
come from the single "tap" interface. We want to allow the possibility to
broaden that in future. As preparation for that, have the code in tap.c
pass the pif id of the originating interface to each of those handler
functions.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
For certain socket types, we record in the epoll ref whether they're
sockets in the namespace, or on the host. We now have the notion of "pif"
to indicate what "place" a socket is associated with, so generalise the
simple one-bit 'ns' to a pif id.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
get_bound_ports_*() now only use their context and ns parameters to
determine which forwarding maps they're operating on. Each function needs
the map they're actually updating, as well as the map for the other
direction, to avoid creating forwarding loops. The UDP function also
requires the corresponding TCP map, to implement the behaviour where we
forward UDP ports of the same number as bound TCP ports for tools like
iperf3.
Passing those maps directly as parameters simplifies the code without
making the callers life harder, because those already know the relevant
maps. IMO, invoking these functions in terms of where they're looking for
updated forwarding also makes more logical sense than in terms of where
they're looking for bound ports. Given that new way of looking at the
functions, also rename them to port_fwd_scan_*().
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Currently get_bound_ports() takes a parameter to determine if it scans for
UDP or TCP bound ports, but in fact there's almost nothing in common
between those two paths. The parameter appears primarily to have been
a convenience for when we needed to invoke this function via NS_CALL().
Now that we don't need that, split it into separate TCP and UDP versions.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
When we want to scan for bound ports in the namespace we use NS_CALL() to
run get_bound_ports() in the namespace. However, the only thing it
actually needed to be in the namespace for was to open the /proc/net file
it was scanning. Since we now always pre-open those, we no longer need
to switch to the namespace for the actual get_bound_ports() calls.
That in turn means that tcp_port_detect() doesn't need to run in the ns
either, and we can just replace it with inline calls to get_bound_ports().
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
The implementation of scanning /proc files to do automatic port forwarding
is a bit awkwardly split between procfs_scan_listen() in util.c,
get_bound_ports() and related functions in conf.c and the initial setup
code in conf().
Consolidate all of this into port_fwd.h, which already has some related
definitions, and a new port_fwd.c.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
...so that we'll retry sending them, instead of more-or-less silently
dropping them. This happens quite frequently if our sending buffer on
the UNIX domain socket is heavily constrained (for instance, by the
208 KiB default memory limit).
It might be argued that dropping frames is part of the expected TCP
flow: we don't dequeue those from the socket anyway, so we'll
eventually retransmit them.
But we don't need the receiver to tell us (by the way of duplicate or
missing ACKs) that we couldn't send them: we already know as
sendmsg() reports that. This seems to considerably increase
throughput stability and throughput itself for TCP connections with
default wmem_max values.
Unfortunately, the 16 bits left as padding in the frame descriptors
we use internally aren't enough to uniquely identify for which
connection we should update sequence numbers: create a parallel
array of pointers to sequence numbers and L4 lengths, of
TCP_FRAMES_MEM size, and go through it after calling sendmsg().
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
It looks like we need it as workaround for this situation, readily
reproducible at least with a 6.5 Linux kernel, with default rmem_max
and wmem_max values:
- an iperf3 client on the host sends about 160 KiB, typically
segmented into five frames by passt. We read this data using
MSG_PEEK
- the iperf3 server on the guest starts receiving
- meanwhile, the host kernel advertised a zero-sized window to the
sender, as expected
- eventually, the guest acknowledges all the data sent so far, and
we drop it from the buffer, courtesy of tcp_sock_consume(), using
recv() with MSG_TRUNC
- the client, however, doesn't get an updated window value, and
even keepalive packets are answered with zero-window segments,
until the connection is closed
It looks like dropping data from a socket using MSG_TRUNC doesn't
cause a recalculation of the window, which would be expected as a
result of any receiving operation that invalidates data on a buffer
(that is, not with MSG_PEEK).
Strangely enough, setting TCP_WINDOW_CLAMP via setsockopt(), even to
the previous value we clamped to, forces a recalculation of the
window which is advertised to the sender.
I couldn't quite confirm this issue by following all the possible
code paths in the kernel, yet. If confirmed, this should be fixed in
the kernel, but meanwhile this workaround looks robust to me (and it
will be needed for backward compatibility anyway).
Reported-by: Matej Hrica <mhrica@redhat.com>
Link: https://bugs.passt.top/show_bug.cgi?id=74
Analysed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
cppcheck 2.12.0 (and maybe some other versions) things this if condition
is always true, which is demonstrably not true. Work around the bug for
now.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Newer versions of cppcheck (as of 2.12.0, at least) added a warning for
pointers which could be declared to point at const data, but aren't.
Based on that, make many pointers throughout the codebase const.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
We have a bunch of variants of the siphash functions for different data
sizes. The callers, in tcp.c, need to pack the various values they want to
hash into a temporary structure, then call the appropriate version. We can
avoid the copy into the temporary by directly using the incremental
siphash functions.
The length specific hash functions also have an undocumented constraint
that the data pointer they take must, in fact, be aligned to avoid
unaligned accesses, which may cause crashes on some architectures.
So, prefer the incremental approach and remove the length-specific
functions.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Some of the siphas_*b() functions return 64-bit results, others 32-bit
results, with no obvious pattern. siphash_32b() also appears to do this
incorrectly - taking the 64-bit hash value and simply returning it
truncated, rather than folding the two halves together.
Since SipHash proper is defined to give a 64-bit hash, make all of them
return 64-bit results. In the one caller which needs a 32-bit value,
tcp_seq_init() do the fold down to 32-bits ourselves.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
We have several workarounds for a clang-tidy bug where the checker doesn't
recognize that a number of system calls write to - and therefore initialise
- a socket address. We can't neatly use a suppression, because the bogus
warning shows up some time after the actual system call, when we access
a field of the socket address which clang-tidy erroneously thinks is
uninitialised.
Consolidate these workarounds into one place by using macros to implement
wrappers around affected system calls which add a memset() of the sockaddr
to silence clang-tidy. This removes the need for the individual memset()
workarounds at the callers - and the somewhat longwinded explanatory
comments.
We can then use a #define to not include the hack in "real" builds, but
only consider it for clang-tidy.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
A classic gotcha of the standard C library is that its unwise to call any
variable 'index' because it will shadow the standard string library
function index(3). This can cause warnings from cppcheck amongst others,
and it also means that if the variable is removed you tend to get confusing
type errors (or sometimes nothing at all) instead of a nice simple "name is
not defined" error.
Strictly speaking this only occurs if <string.h> is included, but that
is so common that as a rule it's best to just avoid it always. We
have a number of places which hit this trap, so rename variables and
parameters to avoid it.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
The reporter is running a SMTP server behind pasta, and the client
waits for the server's banner before sending any data. In turn, the
server waits for our ACK after sending SYN,ACK, which never comes.
If we use the ACK_IF_NEEDED indication to tcp_send_flag(), given that
there's no pending data, we delay sending the ACK segment at the end
of the three-way handshake until we have some data to send to the
server.
This was actually intended, as I thought we would lower the latency
for new connections, but we can't assume that the client will start
sending data first (SMTP is the typical example where this doesn't
happen).
And, trying out this patch with SSH (where the client starts sending
data first), the reporter actually noticed we have a lower latency
by forcing an ACK right away. Comparing a capture before the patch:
13:07:14.007704 IP 10.1.2.1.42056 > 10.1.2.140.1234: Flags [S], seq 1797034836, win 65535, options [mss 4096,nop,wscale 7], length 0
13:07:14.007769 IP 10.1.2.140.1234 > 10.1.2.1.42056: Flags [S.], seq 2297052481, ack 1797034837, win 65480, options [mss 65480,nop,wscale 7], length 0
13:07:14.008462 IP 10.1.2.1.42056 > 10.1.2.140.1234: Flags [.], seq 1:22, ack 1, win 65535, length 21
13:07:14.008496 IP 10.1.2.140.1234 > 10.1.2.1.42056: Flags [.], ack 22, win 512, length 0
13:07:14.011799 IP 10.1.2.140.1234 > 10.1.2.1.42056: Flags [P.], seq 1:515, ack 22, win 512, length 514
and after:
13:10:26.165364 IP 10.1.2.1.59508 > 10.1.2.140.1234: Flags [S], seq 4165939595, win 65535, options [mss 4096,nop,wscale 7], length 0
13:10:26.165391 IP 10.1.2.140.1234 > 10.1.2.1.59508: Flags [S.], seq 985607380, ack 4165939596, win 65480, options [mss 65480,nop,wscale 7], length 0
13:10:26.165418 IP 10.1.2.1.59508 > 10.1.2.140.1234: Flags [.], ack 1, win 512, length 0
13:10:26.165683 IP 10.1.2.1.59508 > 10.1.2.140.1234: Flags [.], seq 1:22, ack 1, win 512, length 21
13:10:26.165698 IP 10.1.2.140.1234 > 10.1.2.1.59508: Flags [.], ack 22, win 512, length 0
13:10:26.167107 IP 10.1.2.140.1234 > 10.1.2.1.59508: Flags [P.], seq 1:515, ack 22, win 512, length 514
the latency between the initial SYN segment and the first data
transmission actually decreases from 792µs to 334µs. This is not
statistically relevant as we have a single measurement, but it can't
be that bad, either.
Reported-by: cr3bs (from IRC)
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
When the guest tries to establish a connection, it could give up on it by
sending a FIN,ACK instead of a plain ACK to our SYN,ACK. It could then
make a new attempt to establish a connection with the same addresses and
ports with a new SYN.
Although it's unlikely, it could send the 2nd SYN very shortly after the
FIN,ACK resulting in both being received in the same batch of packets from
the tap interface.
Currently, we don't handle that correctly, when we receive a FIN,ACK on a
not fully established connection we discard the remaining packets in the
batch, and so will never process the 2nd SYN. Correct this by returning
1 from tcp_tap_handler() in this case, so we'll just consume the FIN,ACK
and continue to process the rest of the batch.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
There are a number of conditions where we will issue a TCP RST in response
to something unexpected we received from the tap interface. These occur in
both tcp_data_from_tap() and tcp_tap_handler(). In tcp_tap_handler() use
a 'goto out of line' technique to consolidate all these paths into one
place. For the tcp_data_from_tap() cases use a negative return code and
direct that to the same path in tcp_tap_handler(), its caller.
In this case we want to discard all remaining packets in the batch we have
received: even if they're otherwise good, they'll be invalidated when the
guest receives the RST we're sending. This is subtly different from the
case where we *receive* an RST, where we could in theory get a new SYN
immediately afterwards. Clarify that with a common on the now common
reset path.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Although it's unlikely in practice, the guest could theoretically
reset one TCP connection then immediately start a new one with the
same addressses and ports, such that we get an RST then a SYN in the
same batch of received packets in tcp_tap_handler().
We don't correctly handle that unlikely case, because when we receive
the RST, we discard any remaining packets in the batch so we'd never
see the SYN. This could happen in either tcp_tap_handler() or
tcp_data_from_tap(). Correct that by returning 1, so that the caller
will continue calling tcp_tap_handler() on subsequent packets allowing
us to process any subsequent SYN.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Currently tcp_data_from_tap() is assumed to consume all packets remaining
in the packet pool it is given. However there are some edge cases where
that's not correct. In preparation for fixing those, change it to return
a count of packets consumed and use that in its caller.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
>From a practical point of view, when a TCP connection ends, whether by
FIN or by RST, we set the CLOSED event, then some time later we remove the
connection from the hash table and clean it up. However, from a protocol
point of view, once it's closed, it's gone, and any new packets with
matching addresses and ports are either forming a new connection, or are
invalid packets to discard.
Enforce these semantics in the TCP hash logic by never hash matching closed
connections.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Both tcp_data_from_tap() and tcp_tap_handler() call packet_get() to get
the entire L4 packet length, then immediately call it again to check that
the packet is long enough to include a TCP header. The features of
packet_get() let us easily combine these together, we just need to adjust
the length slightly, because we want the value to include the TCP header
length.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
In both tap4_handler() and tap6_handler(), once we've sorted incoming l3
packets into "sequences", we then step through all the packets in each TCP
sequence calling tcp_tap_handler(). Or so it appears.
In fact, tcp_tap_handler() doesn't take an index and always looks at packet
0 of the sequence, except when it calls tcp_data_from_tap() to process
data packets. It appears to be written with the idea that the struct pool
is a queue, from which it consumes packets as it processes them, but that's
not how the pool data structure works - they are more like an array of
packets.
We only get away with this, because setup packets for TCP tend to come in
separate batches (because we need to reply in between) and so we only get
a bunch of packets for the same connection together when they're data
packets (tcp_data_from_tap() has its own loop through packets).
Correct this by adding an index parameter to tcp_tap_handler() and altering
the loops in tap.c to step through the pool properly.
Link: https://bugs.passt.top/show_bug.cgi?id=68
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
tcp_defer_handler() performs a potentially expensive linear scan of the
connection table. So, to mitigate the cost of that we skip if if we're not
under at least moderate pressure: either 30% of available connections or
30% (estimated) of available fds used.
But, the calculation for this has been broken since it was introduced: we
calculate "max_conns" based on c->tcp.conn_count, not TCP_MAX_CONNS,
meaning we only exit early if conn_count is less than 30% of itself, i.e.
never.
If that calculation is "corrected" to be based on TCP_MAX_CONNS, it
completely tanks the TCP CRR times for passt - from ~60ms to >1000ms on my
laptop. My guess is that this is because in the case of many short lived
connections, we're letting the table become much fuller before compacting
it. That means that other places which perform a table scan now have to
do much, much more.
For the time being, simply remove the tests, since they're not doing
anything useful. We can reintroduce them more carefully if we see a need
for them.
This also removes the only user of c->tcp.splice_conn_count, so that can
be removed as well.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
The in_epoll boolean is one of only two fields (currently) in the common
structure shared between tap and spliced connections. It seems like it
belongs there, because both tap and spliced connections use it, and it has
roughly the same meaning.
Roughly, however, isn't exactly: which fds this flag says are in the epoll
varies between the two connection types, and are in type specific fields.
So, it's only possible to meaningfully use this value locally in type
specific code anyway.
This common field is going to get in the way of more widespread
generalisation of connection / flow tracking, so move it to separate fields
in the tap and splice specific structures.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Because packets sent on the tap interface will always be going to the
guest/namespace, we more-or-less know what address they'll be going to. So
we pre-fill this destination address in our header buffers for IPv4. We
can't do the same for IPv6 because we could need either the global or
link-local address for the guest. In future we're going to want more
flexibility for the destination address, so this pre-filling will get in
the way.
Change the flow so we always fill in the IPv4 destination address for each
packet, rather than prefilling it from proto_update_l2_buf(). In fact for
TCP we already redundantly filled the destination for each packet anyway.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
We partially prepopulate IP and TCP header structures including, amongst
other things the destination address, which for IPv4 is always the known
address of the guest/namespace. We partially precompute both the IPv4
header checksum and the TCP checksum based on this.
In future we're going to want more flexibility with controlling the
destination for IPv4 (as we already do for IPv6), so this precomputed value
gets in the way. Therefore remove the IPv4 destination from the
precomputed checksum and fold it into the checksum update when we actually
send a packet.
Doing this means we no longer need to recompute those partial sums when
the destination address changes ({tcp,udp}_update_l2_buf()) and instead
the computation can be moved to compile time. This means while we perform
slightly more computations on each packet, we slightly reduce the amount of
memory we need to access.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
In tcp_seq_init() the meaning of "src" and "dst" isn't really clear since
it's used for connections in both directions. However, these values are
just feeding a hash, so as long as we're consistent and include all the
information we want, it doesn't really matter.
Oddly, for the "src" side we supply the (tap side) forwarding address but
the (tap side) endpoint port. This again doesn't really matter, but it's
confusing. So swap this with dstport, so "src" is always forwarding
and "dst" is always endpoint.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
