There are several cases where we make a goodput measurement and we are running
out of data when we decide to make the measurement. In reality we should not make
such a measurement if there is no chance we can have "enough" data. There is also
some corner case TLP's that end up not registering as a TLP like they should, we
fix this by pushing the doing_tlp setup to the actual timeout that knows it did
a TLP. This makes it so we always have the appropriate flag on the sendmap
indicating a TLP being done as well as count correctly so we make no more
that two TLP's.
In addressing the goodput lets also add a "quality" metric that can be viewed via
blackbox logs so that a casual observer does not have to figure out how good
of a measurement it is. This is needed due to the fact that we may still make
a measurement that is of a poorer quality as we run out of data but still have
a minimal amount of data to make a measurement.
Reviewed by: tuexen
Sponsored by: Netflix Inc.
Differential Revision: https://reviews.freebsd.org/D31076
Ifnet (inline) hw kTLS NICs typically keep state within
a TLS record, so that when transmitting in-order,
they can continue encryption on each segment sent without
DMA'ing extra state from the host.
This breaks down when transmits are out of order (eg,
TCP retransmits). In this case, the NIC must re-DMA
the entire TLS record up to and including the segment
being retransmitted. This means that when re-transmitting
the last 1448 byte segment of a TLS record, the NIC will
have to re-DMA the entire 16KB TLS record. This can lead
to the NIC running out of PCIe bus bandwidth well before
it saturates the network link if a lot of TCP connections have
a high retransmoit rate.
This change introduces a new sysctl (kern.ipc.tls.ifnet_max_rexmit_pct),
where TCP connections with higher retransmit rate will be
switched to SW kTLS so as to conserve PCIe bandwidth.
Reviewed by: hselasky, markj, rrs
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D30908
The kernel part of ipfw(8) does initialize LibAlias uncondistionally
with an zeroized port range (allowed ports from 0 to 0). During
restucturing of libalias, port ranges are used everytime and are
therefor initialized with different values than zero. The secondary
initialization from ipfw (and probably others) overrides the new
default values and leave the instance in an unfunctional state. The
obvious solution is to detect such reinitializations and use the new
default value instead.
MFC after: 3 days
The expiration time of direct address mappings is explicitly
uninitialized. Expire times are always compared during housekeeping.
Despite the uninitialized value does not harm, it's simpler to just
set it to a reasonable default. This was detected during valgrinding
the test suite.
MFC after: 3 days
Comparing elements in a tree requires transitiviy. If a < b and b < c
then a must be smaller than c. This way the tree elements are always
pairwise comparable.
Tristate comparsion functions returning values lower, equal, or
greater than zero, are usually implemented by a simple subtraction of
the operands. If the size of the operands are equal to the size of
the result, integer modular arithmetics kick in and violates the
transitivity.
Example:
Working on byte with 0, 120, and 240. Now computing the differences:
120 - 0 = 120
240 - 120 = 120
240 - 0 = -16
MFC after: 3 days
Some TCP stacks negotiate TS support, but do not send TS at all
or not for keep-alive segments. Since this includes modern widely
deployed stacks, tolerate the violation of RFC 7323 per default.
Reviewed by: rgrimes, rrs, rscheff
MFC after: 3 days
Differential Revision: https://reviews.freebsd.org/D30740
Sponsored by: Netflix, Inc.
Hardware TLS is now supported in some interface cards and it works well. Except that
when we have connections that retransmit a lot we get into trouble with all the retransmits.
This prep step makes way for change that Drew will be making so that we can "kick out" a
session from hardware TLS.
Reviewed by: mtuexen, gallatin
Sponsored by: Netflix Inc
Differential Revision: https://reviews.freebsd.org/D30895
There were two bugs that prevented V4 sockets from connecting to
a rack server running a V4/V6 socket. As well as a bug that stops the
mapped v4 in V6 address from working.
Reviewed by: mtuexen
Sponsored by: Netflix Inc
Differential Revision: https://reviews.freebsd.org/D30885
Compiling libalias results in warnings about unused functions.
Those warnings are caused by clang's heuristic to consider an inline
function as in use, iff the declaration is in a *.c file.
Declarations in *.h files do not emit those warnings.
Hence the declarations must be moved to an extra *.h file.
MFC after: 3 days
Differential Revision: https://reviews.freebsd.org/D30844
tcp_twcheck now expects a read lock on the inp for the SYN case
instead of a write lock.
Reviewed by: np
Fixes: 1db08fbe3f tcp_input: always request read-locking of PCB for any pure SYN segment.
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D30782
are unneccessary and used to be there before TFO as an invariant. With
TFO and after 8d5719aa74 the "so" value is still needed.
Reported & tested by: tuexen
Fixes: 8d5719aa74
Current data structure is using a hash of unordered lists. Those
unordered lists are quite efficient, because the least recently
inserted entries are most likely to be used again. In order to avoid
long search times in other cases, the lists are hashed into many
buckets. Unfortunatly a search for a miss needs an exhaustive
inspection and a careful definition of the hash.
Splay trees offer a similar feature: Almost O(1) for access of the
least recently used entries, and amortized O(ln(n)) for almost all
other cases. Get rid of the hash.
Now the data structure should able to quickly react to external
packets without eating CPU cycles for breakfast, preventing a DoS.
PR: 192888
Discussed with: Dimitry Luhtionov
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30536
Current data structure is using a hash of unordered lists. Those
unordered lists are quite efficient, because the least recently
inserted entries are most likely to be used again. In order to avoid
long search times in other cases, the lists are hashed into many
buckets. Unfortunatly a search for a miss needs an exhaustive
inspection and a careful definition of the hash.
Splay trees offer a similar feature - almost O(1) for access of the
least recently used entries), and amortized O(ln(n) - for almost all
other cases. Get rid of the hash.
Discussed with: Dimitry Luhtionov
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30516
The entry deleteAllLinks in the struct libalias is only used to signal
a state between internal calls. It's not used between API calls.
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30604
Get rid of PORT_BASE, replace by AliasRange. Simplify code.
Factor out the search for a new port. Improves the perfomance a bit.
Discussed with: Dimitry Luhtionov
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30581
Let PPTP use its own data structure.
Regroup and rename other lists, which are not PPTP.
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30580
Reorder incoming links by grouping of common search terms.
Significant performance improvement for incoming (missing) flows.
Remove LSNAT from outgoing search.
Slight speedup due to less comparsions in the loop.
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30574
Factor out the outgoing search function.
Preparation for a new data structure.
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30572
Separate the partially specified links into a separate data structure.
This would causes a major parformance impact, if there are many of
them. Use a (smaller) hash table to speed up the partially link
access.
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30570
This completes PRR cwnd reduction in all circumstances
for the base TCP stack (SACK loss recovery, ECN window reduction,
non-SACK loss recovery), preventing the arriving ACKs to
clock out new data at the old, too high rate. This
reduces the chance to induce additional losses while
recovering from loss (during congested network conditions).
For non-SACK loss recovery, each ACK is assumed to have
one MSS delivered. In order to prevent ACK-split attacks,
only one window worth of ACKs is considered to actually
have delivered new data.
MFC after: 6 weeks
Reviewed By: rrs, #transport
Sponsored by: NetApp, Inc.
Differential Revision: https://reviews.freebsd.org/D29441
Search fully specified links first. Some performance loss due to need
to revisit the db twice, if not found.
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30569
Factor out the common Out and In filter
Slightly better performance due to eager skip of search loop
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30568
Summary:
- Use LibAliasTime as a real global variable for central timekeeping.
- Reduce number of syscalls in user space considerably.
- Dynamically adjust the packet counters to match the second resolution.
- Only check the first few packets after a time increase for expiry.
Discussed with: hselasky
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30566
Stats counters are used as unsigned valued (i.e. printf("%u")) but are
defined as signed int. This causes trouble later, so fix it early.
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30587
Some code was using it already, but in many places we were testing
SO_ACCEPTCONN directly. As a small step towards fixing some bugs
involving synchronization with listen(2), make the kernel consistently
use SOLISTENING(). No functional change intended.
MFC after: 1 week
Sponsored by: The FreeBSD Foundation
Recently (Nov) we added logic that protects against a peer negotiating a timestamp, and
then not including a timestamp. This involved in the input path doing a goto done_with_input
label. Now I suspect the code was cribbed from one in Rack that has to do with the SYN.
This had a bug, i.e. it should have a m_freem(m) before going to the label (bbr had this
missing m_freem() but rack did not). This then caused the missing m_freem to show
up in both BBR and Rack. Also looking at the code referencing m->m_pkthdr.lro_nsegs
later (after processing) is not a good idea, even though its only for logging. Best to
copy that off before any frees can take place.
Reviewed by: mtuexen
Sponsored by: Netflix Inc
Differential Revision: https://reviews.freebsd.org/D30727
* Completely initialise the CC module specific data
* Use beta_ecn in case of an ECN event whenever ABE is enabled
or it is requested by the stack.
Reviewed by: rscheff, rrs
MFC after: 3 days
Sponsored by: Netflix, Inc.
Replace current expensive, but sparsly called housekeeping
by a single, repetive action.
This is part of a larger restructure of libalias in order to switch to
more efficient data structures. The whole restructure process is
split into 15 reviews to ease reviewing. All those steps will be
squashed into a single commit for MFC in order to hide the
intermediate states from production systems.
Reviewed by: hselasky
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30277
When running at NF the current Rack and BBR changes with the recent
commits from Richard that cause the socket buffer lock to be held over
the ip_output() call and then finally culminating in a call to tcp_handle_wakeup()
we get a lot of leaked mbufs. I don't think that this leak is actually caused
by holding the lock or what Richard has done, but is exposing some other
bug that has probably been lying dormant for a long time. I will continue to
look (using his changes) at what is going on to try to root cause out the issue.
In the meantime I can't leave the leaks out for everyone else. So this commit
will revert all of Richards changes and move both Rack and BBR back to just
doing the old sorwakeup_locked() calls after messing with the so_rcv buffer.
We may want to look at adding back in Richards changes after I have pinpointed
the root cause of the mbuf leak and fixed it.
Reviewed by: mtuexen,rscheff
Sponsored by: Netflix Inc
Differential Revision: https://reviews.freebsd.org/D30704
Recently we had a rewrite to tcp_lro.c that was tested but one subtle change
was the move to a less precise timestamp. This causes all kinds of chaos
in tcp's that do pacing and needs to be fixed to use the more precise
time that was there before.
Reviewed by: mtuexen, gallatin, hselasky
Sponsored by: Netflix Inc
Differential Revision: https://reviews.freebsd.org/D30695
So it turns out that my fix before was not correct. It ended with us failing
some of the "improved" SYN tests, since we are not in the correct states.
With more digging I have figured out the root of the problem is that when
we receive a SYN|FIN the reassembly code made it so we create a segq entry
to hold the FIN. In the established state where we were not in order this
would be correct i.e. a 0 len with a FIN would need to be accepted. But
if you are in a front state we need to strip the FIN so we correctly handle
the ACK but ignore the FIN. This gets us into the proper states
and avoids the previous ack war.
I back out some of the previous changes but then add a new change
here in tcp_reass() that fixes the root cause of the issue. We still
leave the rack panic fixes in place however.
Reviewed by: mtuexen
Sponsored by: Netflix Inc
Differential Revision: https://reviews.freebsd.org/D30627
Prior to commit f161d294b we only checked the sockaddr length, but now
we verify the address family as well. This breaks at least ttcp. Relax
the check to avoid breaking compatibility too much: permit AF_UNSPEC if
the address is INADDR_ANY.
Fixes: f161d294b
Reported by: Bakul Shah <bakul@iitbombay.org>
Reviewed by: tuexen
MFC after: 3 days
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D30539
The functionality to detect a newly created link after processing a
single packet is decoupled from the packet processing. Every new
packet is processed asynchronously and will reset the indicator, hence
the function is unusable. I made a Google search for third party code,
which uses the function, and failed to find one.
That's why the function should be removed: It unusable and unused.
A much simplified API/ABI will remain in anything below 14.
Discussed with: kp
Reviewed by: manpages (bcr)
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D30275
Approved by: mw
Obtained from: Semihalf
Sponsored by: Stormshield
Differential Revision: https://reviews.freebsd.org/D30354
Changes:
1. add spinlock to bw_meter
If two contexts read and modify bw_meter values
it might happen that these are corrupted.
Guard only code fragments which do read-and-modify.
Context which only do "reads" are not done inside
spinlock block. The only sideffect that can happen is
an 1-p;acket outdated value reported back to userspace.
2. replace all locks with a single RWLOCK
Multiple locks caused a performance issue in routing
hot path, when two of them had to be taken. All locks
were replaced with single RWLOCK which makes the hot
path able to take only shared access to lock most of
the times.
All configuration routines have to take exclusive lock
(as it was done before) but these operation are very rare
compared to packet routing.
3. redesign MFC expire and UPCALL expire
Use generic kthread and cv_wait/cv_signal for deferring
work. Previously, upcalls could be sent from two contexts
which complicated the design. All upcall sending is now
done in a kthread which allows hot path to work more
efficient in some rare cases.
4. replace mutex-guarded linked list with lock free buf_ring
All message and data is now passed over lockless buf_ring.
This allowed to remove some heavy locking when linked
lists were used.
The commit 189f8eea contains a refactorisation of a constant. During
later review D30283 the naming of the constant was improved and the
initialization became explicit. Put this into the tree, in order to
MFC the correct naming.
The last set of commits fixed both a panic (in rack) and an ACK-war (in freebsd and bbr).
However there was a missing case, i.e. where we get an out-of-order FIN by itself.
In such a case we don't want to leave the FIN bit set, otherwise we will do the
wrong thing and ack the FIN incorrectly. Instead we need to go through the
tcp_reasm() code and that way the FIN will be stripped and all will be well.
Reviewed by: mtuexen,rscheff
Sponsored by: Netflix Inc
Differential Revision: https://reviews.freebsd.org/D30497
Most CC algos do use local data, and when calling
newreno_cong_signal from there, the latter misinterprets
the data as its own struct, leading to incorrect behavior.
Reported by: chengc_netapp.com
Reviewed By: chengc_netapp.com, tuexen, #transport
MFC after: 3 days
Sponsored By: NetApp, Inc.
Differential Revision: https://reviews.freebsd.org/D30470
