Hi Norbert,
thanks for also having an eye on this - I am looking for the failure
reports on ci.libreoffice.org currently, too.
Last is from http://ci.libreoffice.org/job/lo_tb_master_linux_dbg/7195/,
so last is from Friday, 13th (uhhh...)
Have you seen such or similar stacks anywhere else? In the meantime I
tried ChartTest massively locally on Linux and Win, but could never
locally reproduce.
The SolarMutex thing is sure not good, but only a symtom showing up I
would guess. There are tests and codes in SC e.g. that also use massive
parallelism, not limited to an upper core count. The basic problem is
that the MainThread always holds the SolarMutex, so also during calling
waitUntilEmpty(). The consequence is that no WorkerThread is allowed to
get the SolarMutex, limiting multithreaded actions to this.
I knew that and made sure that the multithreaded 3DRenderer
WorkerThreads do not need the SolarMutex for their work. I did not know
yet that the memory fail handler tries to get the SolarMutex, too, but
is logic when it wants to bring up a dialog in some form.
But the deeper problem is that allocation - here extending a vector of
pointers to a helper class from 1 to 2 entries - fails. Sometimes. And
that only on many cores on that machine (up to now).
I checked all involved classes, their refcounting and that the used
o3tl::cow_wrapper uses the ThreadSafeRefCountingPolicy, looks good so
far. It is also not the case that the WorkerThreads need massive amounts
of own memory, so I doublt that limiting to e.g. 8 thredads would change
this, except maybe making it less probable to happen. I looked at
o3tl::cow_wrapper itself, and the basic B2D/B3DPrimitive implementations
which internally use a comphelper::OBaseMutex e.g. for creating buffered
decompositions.
I found no concrete reason until now, any tipps/help much appreciated.
I keep watching this - at least it did not happen in all the builds
since 13th and on no other machine, so the thread now is to somehow nail
it to get it reproducable. If someone has other traces, please send
them! I would hate to take this back, esp. because we will need
multithreading more and more since Moore's law is tilting.
Sincerely,
Armin
Am 17.05.2016 um 14:35 schrieb Norbert Thiebaud:
On Tue, May 17, 2016 at 6:44 AM, Thorsten Behrens <thb@libreoffice.org> wrote:
Norbert Thiebaud wrote:
The threaded work then raise() due to some memory problem and out
signal handler try to acquire the solar mutex ->deadlock
Eek, that's ugly. Then again, at the core is the OOM condition, which
needs solving independently. Per chance, is that happening on a box
with massive amounts of CPU threads?
it is on the ci builder, so yeah 32 thread or so.
but I disagree that it is _at the core_
at the core this exhibit 2 things:
1/ we do a lot of thing that is verboten in a signal handler.
2/ taking a lock that rely on other thread to move forward while
holding the solarmutex is begging for deadlock.
Norbert
--
--
ALG (PGP Key: EE1C 4B3F E751 D8BC C485 DEC1 3C59 F953 D81C F4A2)
Context
Privacy Policy |
Impressum (Legal Info) |
Copyright information: Unless otherwise specified, all text and images
on this website are licensed under the
Creative Commons Attribution-Share Alike 3.0 License.
This does not include the source code of LibreOffice, which is
licensed under the Mozilla Public License (
MPLv2).
"LibreOffice" and "The Document Foundation" are
registered trademarks of their corresponding registered owners or are
in actual use as trademarks in one or more countries. Their respective
logos and icons are also subject to international copyright laws. Use
thereof is explained in our
trademark policy.