This makes it possible to use a different RHC implementations for
different readers and removes the need for the RHC interface to be part
of the global state.
Signed-off-by: Erik Boasson <eb@ilities.com>
Commit 3afce30c37 introduced an error in
the calculation of the size of these submessages, making them (and
requiring them to be) larger than correct and putting the "count" field
at the wrong offset, breaking interoperability.
Signed-off-by: Erik Boasson <eb@ilities.com>
* Change default behaviour with no arguments to print help text;
"ddsperf sanity" now gives the behaviour that it used to give when run
with no arguments;
* Include context switch rate in output;
* Allow suffixing frequencies and sizes with standard units (so "size
1kB" is now allowed);
* Add missing option to help text, extend help text with some additional
informationr.
Signed-off-by: Erik Boasson <eb@ilities.com>
The main change is a longer sleep in the main thread while waiting for
the select call to timeout. Still not perfect, but more practical than
checking the stacktrace of the thread calling ddsrt_select() to see
whether it has spent too much time in the call.
Signed-off-by: Erik Boasson <eb@ilities.com>
Caused by the changes in a652ecb78e; the
sample that matters is the first in what may now be a chain of samples,
which requires some overlooked adjustments.
Signed-off-by: Erik Boasson <eb@ilities.com>
Some Linux versions sometimes return this (undocumented) error,
presumably because of firewalling. Better to ignore it.
Signed-off-by: Erik Boasson <eb@ilities.com>
* the scripted throughput test originally used for the throughput graph
in the README now uses ddsperf;
* a scripted latency test has been added;
* updated the README with the results of these tests (and so now gives
easy access not only to throughput, but also to latency and memory
usage, as well as to latency over GbE.
Signed-off-by: Erik Boasson <eb@ilities.com>
* per-thread CPU usage (only those threads where the load is over 0.5%,
if the sum of threads below that threshold exceeds 0.5%, it prints an
aggregate for those threads);
* also report RSS;
* network load (only on request, as percentage of specified network
bandwidth and actual bytes in/out, with the output suppressed if it is
0%);
* publish CPU usage so a ddsperf instance can display CPU loads for
its peers;
* handle SIGXFSZ (file size exceeded) by displaying one last line of
statistics before killing itself; this simply a debugging tool to make
it easier to get a trace covering a high sample-rate start-up issue;
* default topic changed to "KS" because that allows all the options to
be used, this has a negative impact on performance (both latency and
small-sample throughput) but it should be less surprising to users;
* specifying a size is now done by appending "size N" (where N is the
size in bytes) after a "ping" or "pub" command, rather than it having
to set it via a command-line option;
Note that some of this is platform-dependent -- SIGXFSZ is currently
only on Linux and macOS, and CPU and network load reporting is currently
only on Linux, macOS and Windows.
Signed-off-by: Erik Boasson <eb@ilities.com>
* Move the project top-level CMakeLists.txt to the root of the project;
this allows building Cyclone as part of ROS2 without any special
tricks;
* Clean up the build options:
ENABLE_SSL: whether to check for and include OpenSSL support if a
library can be found (default = ON); this used to be
called DDSC_ENABLE_OPENSSL, the old name is deprecated
but still works
BUILD_DOCS: whether to build docs (default = OFF)
BUILD_TESTING: whether to build test (default = OFF)
* Collect all documentation into top-level "docs" directory;
* Move the examples to the top-level directory;
* Remove the unused and somewhat misleading pseudo-default
cyclonedds.xml;
* Remove unused cmake files
Signed-off-by: Erik Boasson <eb@ilities.com>
* use multicast only for participant discovery if using a WiFi network
* default to using unicast for retransmits
Signed-off-by: Erik Boasson <eb@ilities.com>
It is an excellent platform for catching bugs: big-endian, slow enough
that a context switch in the middle of an operation becomes a regular
occurrence, and all that on a SMP box. Or: I just wanted to see if it
would work.
Signed-off-by: Erik Boasson <eb@ilities.com>
Creating a reader/writer in a listener for a built-in topic (as ddsperf
does) recursively ends up in reader/writer matching, recursively
read-locking qoslock. As nobody ever takes a write-lock, it is a
non-issue provided the rwlock really is an rwlock. On Solaris 2.6 those
don't exist, and mapping it onto a mutex deadlocks.
This commit removes the thing in its entirety, the fact that it is
currently only ever locked for reading is hint that perhaps it is not
that valuable a thing. The way it was used in the code would in any
case not have helped with re-matching on QoS changes (save for
duplicating all the matching code), and it is doubtful that serializing
the matching in that case would be necessary in the first place.
Signed-off-by: Erik Boasson <eb@ilities.com>
The two do essentially the same think, and ddsrt_strtok_r was only used
in one place. (Triggered by Solaris 2.6 not providing strtok_r.)
Signed-off-by: Erik Boasson <eb@ilities.com>
On 32-bit machines uintmax_t is likely to be slower than uintptr_t, and
for that reason, using an uintmax_t for a thread id seems highly
unlikely. For the current platforms, uintptr_t works.
Signed-off-by: Erik Boasson <eb@ilities.com>
The parameter list table indices ought to be a small as possible to
avoid wasting space, and that means the index size is dependent on
whether or not DDSI_INCLUDE_SSM is set.
Signed-off-by: Erik Boasson <eb@ilities.com>
The payload in a struct serdata_default is assumed to be at a 64-bit
offset for conversion to/from a dds_{i,o}stream_t and getting padding
calculations in the serialised representation correct. The definition
did not guarantee this and got it wrong on a 32-bit release build.
This commit computes the required padding at compile time and at
verifies the assumption holds where it matters.
Signed-off-by: Erik Boasson <eb@ilities.com>
Signed-off-by: Thijs Sassen <thijs.sassen@adlinktech.com>
Adjusted the close methode not to expand by the lwip close macro and added a check for DDSI_INCLUDE_SSM to match the correct pid table size.
Signed-off-by: Thijs Sassen <thijs.sassen@adlinktech.com>
Currently each DDSC (not DDSI) writer has its own "xpack" for packing
submessages into larger messages, but that is a bit wasteful, especially
when a lot of samples are being generated that never need to go onto the
wire. Lazily allocating them and only pushing message into them when
they have a destination address saves memory and improves speed for
local communications.
Signed-off-by: Erik Boasson <eb@ilities.com>
Multiple writers for a single instance is pretty rare, so it makes sense
to lazily allocate the tables for keeping track of them. The more
elegant solution would be to have a single lock-free table.
Signed-off-by: Erik Boasson <eb@ilities.com>
Rather than allocate a HH_HOP_RANGE large array of buckets, allocate
just 1 if the initial size is 1, then jump to HH_HOP_RANGE as soon as a
second element is added to the table. There are quite a few cases where
hash tables are created where there never be more than 1 (or even 0)
elements in the table (e.g., a writer without readers, a reader for a
keyless topic).
Signed-off-by: Erik Boasson <eb@ilities.com>
There were inconsistencies in the order in which entity locks were taken
when multiple entities needed to be locked at the same time. In most
cases, the order was first locking entity X, then locking the parent
entity of X. However, in some cases the order was reversed, a likely
cause of deadlocks.
This commit sorts these problems, and in particular propagating
operations into children. The entity refcount is now part of the handle
administration so that it is no longer necessary to lock an entity to
determine whether it is still allowed to be used (previously it had to
check the CLOSED flag afterward). This allows recursing into the
children while holding handles and the underlying objects alive, but
without violating lock order.
Attendant changes that would warrant there own commits but are too hard
to split off:
* Children are now no longer in a singly linked list, but in an AVL
tree; this was necessary at some intermediate stage to allow unlocking
an entity and restarting iteration over all children at the "next"
child (all thanks to the eternally unique instance handle);
* Waitsets shifted to using arrays of attached entities instead of
linked lists; this was a consequence of dealing with some locking
issues in reading triggers and considering which operations on the
"triggered" and "observed" sets are actually needed.
* Entity status flags and waitset/condition trigger counts are now
handled using atomic operations. Entities are now classified as
having a "status" with a corresponding mask, or as having a "trigger
count" (conditions). As there are fewer than 16 status bits, the
status and its mask can squeeze into the same 32-bits as the trigger
count. These atomic updates avoid the need for a separate lock just
for the trigger/status values and results in a significant speedup
with waitsets.
* Create topic now has a more rational behaviour when multiple
participants attempt to create the same topic: each participant now
gets its own topic definition, but the underlying type representation
is shared.
Signed-off-by: Erik Boasson <eb@ilities.com>
Add the instance handle to the DDSC entity type, initialize it properly
for all types, and remove the per-type handling of
dds_get_instance_handle. Those entities that have a DDSI variant take
the instance handle from DDSI (which plays tricks to get the instance
handles of the entities matching the built-in topics). For those that
do not have a DDSI variant, just generate a unique identifier using the
same generate that DDSI uses.
Signed-off-by: Erik Boasson <eb@ilities.com>
Lease_renew is the key one, and that one only ever shifts the lease
expiry to the future, providing the lease hasn't expired already. All
other operations work within leaseheap_lock.
Other updates to lease end time are in set_expiry (which is used in some
special cases). So ... the number of ways this can go wrong is rather
limited.
Tracking pings and expected number of pongs was done without holding the
correct locks. Terminate flag was also not a ddsrt_atomic... and hence
flagged by thread sanitizer as a race condition.
Signed-off-by: Erik Boasson <eb@ilities.com>
Thread sanitizer warns about reads and writes of variables that are
meant to be read without holding a lock:
* Global "keep_going" is now a ddsrt_atomic_uint32_t
* Thread "vtime" is now a ddsrt_atomic_uint32_t
Previously the code relied on the assumption that a 32-bit int would be
treated as atomic, now that is all wrapped in ddsrt_atomic_{ld,st}32.
These being inline functions doing exactly the same thing, there is no
functional change, but it does allow annotating the loads and stores for
via function attributes on the ddsrt_atomic_{ld,st}X.
The concurrent hashtable implementation is replaced by a locked version
of the non-concurrent implementation if thread sanitizer is used. This
changes eliminates the scores of problems signalled by thread sanitizer
in the GUID-to-entity translation and the key-to-instance id lookups.
Other than that, this replaces a flag used in a waitset test case to be
a ddsrt_atomic_uint32_t.
Signed-off-by: Erik Boasson <eb@ilities.com>
* calling ddsrt_memdup, ddsrt_strdup with a null pointer (they handle it
gracefully but forbid it in the interface ...)
* replacement of all pre-C99 flexible arrays (i.e., declaring as
array[1], then mallocing and using as if array[N]) by C99 flexible
arrays.
* also add a missing null-pointer test in dds_dispose_ts, and fix the
test cases that pass a null pointer and a non-writer handle to it to
instead pass an invalid adress
Signed-off-by: Erik Boasson <eb@ilities.com>
