Solves a race condition when deleting a stream, as the fd might get
closed just as another thread tries to read from it, leading to EBADF
read error.
Change-Id: I8ea3ff72c2788ce3051a86a0bbd1650b03965376
(cherry picked from commit 64eba103e3)
(cherry picked from commit 1d97823318)
These are potentially computed from inside each subdir, and in addition
due to what appears to be a regression in GNU make 4.4, where it is
reevaluating variables that contain $(shell) functions, many times (in
the order of thousands, this was slowing down the build, were on the
Debian amd64 build daemons it went from 5m with GNU make 4.3 to 2h40m
with GNU make 4.4. Although the bulk of the slow down has been fixed
with previous commits, the remaining optimizations are only to avoid
this potentially happening again in the future, and to reduce useless
duplicate work.
Instead of trying to cache the values from within make itself, where
programming this there is extremely painful, and does not seem to be
able to greatly reduce the number of calls, because the build system
is going to be called multiple times for different targets. Simply
externalize the generation into several shell scripts, that we call
to generate a make fragment that then we include from the various
Makefiles.
For a Debian build with GNU make 4.3, this reduces the amount of total
pkg-config calls from around ~1600 to 128, for dpkg-buildflags from
~1100 down to 6, and for dpkg-parsechangelog from ~56 to 17, but the
slow down is not as significant there anyway.
For a Debian build with GNU make 4.4, this reduces the amount of total
pkg-config calls from around ~2600 to 128, for dpkg-buildflags from
~2800 down to 6, and for dpkg-parsechangelog from ~350 to 21.
For a Debian build with GNU make 4.4, this reduces the build time
on this system from 2m10s to ~ 1m30s.
Change-Id: I427d0ea5106dc6ed1ff9e664ccdba2fa0725b7d0
(cherry picked from commit 7dbe24e4e5)
(cherry picked from commit 1e1075bf38)
Starting with GNU make 4.4, build time have massively regressed
where before they would take 5m on amd64 now can take 2h40m. While this
seems clearly broken, the release notes are filled with notices for
breaking changes, and in particular the one for passing all make
variables down to the invoked programs executed via the «shell» GNU make
function, so it is not clear what is expected breakage and what is not.
This has been reported in Debian, but not yet upstream, and while it
seems like a clear regression, it's not clear what will be the upstream
take on it. For now apply workarounds that do not change semantics, and
which do not regress with older GNU make versions.
Use the GNU make «origin» function instead of «?=» which defaults to
defining a variable as a recursive one. Coerce already defined variables
into simple ones to avoid GNU make re-evaluating these variables for
each «shell» function invocation.
Ref: https://bugs.debian.org/1092051
Change-Id: I076fc05dd616918473a22e7e942fecfdc9851d47
(cherry picked from commit 887fb40f3f)
(cherry picked from commit 67a2b222c7)
To support asynchronous pollers which may hold references on underlying
sockets, let the poller close the socket after it has released its
references. This prevents cases of file descriptor re-use while an
underlying socket is still open.
Add reset_socket() to be used in place of close_socket() which does the
same thing except the actual closing of the socket.
Add poller_del_item_callback() for cases where more action than just
closing the file descriptor is needed.
Change-Id: Iefda1487ecb89263729120ecb964436dd79b2a0e
The poller-per-thread feature was broken with a division by zero. Take
the opportunity to rework it and eliminate the poller_map object. Use a
simple array of pollers for media sockets, plus one global poller for
control sockets. In the regular case only one poller is created and
everything points to that poller. In the poller-per-thread case, one
poller per thread is created, plus one poller (also with its own single
thread) for control connections. All control sockets use the single
control poller, while all media sockets get assigned one poller from the
pool in a round-robin fashion.
closes#1801
Change-Id: Iae91a3e10b7206455c6df33b1a472254c700ce21
Richard Fuchs
Guillem Jover