When we're running with bitbake -b, BB_TASKDEPDATA is incorrect and limited.
We really need a way to know this from the metadata and this new variable
provides this in worker context. This means existing code can stop having
to guess.
(Bitbake rev: 05763bc886024dcce2ce6b3060fb00abf79a9402)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
For the purposes BB_SETSCENE_ENFORCE is designed for (in OE, it is used
by the installation process for the extensible SDK), we don't actually
need the whitelisted real tasks to execute - we just need to have them
in the dependency tree so that we get all of the setscene tasks they
depend on to run. Therefore we can actually dry-run those real tasks
i.e. they won't be run (and thus we won't waste a significant amount of
time doing so) and won't be stamped as having run either. We do already
have a dry-run mode in BitBake (activated by the -n or --dry-run command
line option), but it dry-runs the setscene tasks as well which we don't
want here.
Note that this has no effect on the checking we are doing with
BB_SETSCENE_ENFORCE to ensure that only whitelisted real tasks are
scheduled to run - that's handled separately.
Fixes [YOCTO #10369].
(Bitbake rev: 58f084291beb3a87d8d9fdb36dfe7eff911fa36b)
Signed-off-by: Paul Eggleton <paul.eggleton@linux.intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Allow the client to set variables with the setVariable command and have
those changes take effect when running tasks. This is accomplished by
collecting changes made by setVariable separately and pass these to the
worker so it can be applied on top of the datastore it creates.
(Bitbake rev: 69a3cd790da35c3898a8f50c284ad1a4677682a4)
Signed-off-by: Paul Eggleton <paul.eggleton@linux.intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Looking further at the CPU loads on systems running large numbers of tasks,
the following things helps performance:
* Loop on waitpid until there are no processes still waiting
* Using select to wait for the cooker pipe to be writable before writing
avoiding pointless 100% cpu usage
* Only reading from worker pipes that select highlights are readable
(Bitbake rev: 9375349e27b08b4d1cfe4825c042d4c82120e00b)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
getVarFlag() now defaults to expanding by default, thus remove the
True option from getVarFlag() calls with a regex search and
replace.
Search made with the following regex:
getVarFlag ?\(( ?[^,()]*, ?[^,()]*), True\)
(Bitbake rev: c19baa8c19ea8ab9b9b64fd30298d8764c6fd2cd)
Signed-off-by: Joshua Lock <joshua.g.lock@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
getVar() now defaults to expanding by default, thus remove the True
option from getVar() calls with a regex search and replace.
Search made with the following regex: getVar ?\(( ?[^,()]*), True\)
(Bitbake rev: 3b45c479de8640f92dd1d9f147b02e1eecfaadc8)
Signed-off-by: Joshua Lock <joshua.g.lock@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
I noiced builds where tasks seemed to be taking a surprisingly long time.
When I looked at the output of top/pstree, these tasks were no longer
running despite being listed in knotty. Some were in D/Z state waiting for
their exit code to be collected, others were simply not present at all.
strace showed communication problems between the worker and cooker, each
was trying to write to the other and nearly deadlocking. Eventually, timeouts
would allow them to echange 64kb of data but this was only happening every
few seconds.
Whilst this particularly affected builds on machines with large numbers
of cores (and hence highly parallal task execution) and in cases where
I had a lot of debug enabled, this situation is clearly bad in general.
This patch introduces a thread to the worker which is used to write data
back to cooker. This means that the deadlock can't occur and data flows
much more freely and effectively.
(Bitbake rev: 3cb0d1c78b4c2e4f251a59b86c8da583828ad08b)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Print full traceback instead of just the exception message in the
child() function inside fork_off_task(). This makes debugging a lot
easier as the function catches a generic "Exception" and the exception
message alone might not give much information.
[YOCTO #10393]
(Bitbake rev: 9c7cc981408c9b4bbbff98ae93ff22199f6a8219)
Signed-off-by: Markus Lehtonen <markus.lehtonen@linux.intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
This patch adds the notion of supporting multiple configurations within
a single build. To enable it, set a line in local.conf like:
BBMULTICONFIG = "configA configB configC"
This would tell bitbake that before it parses the base configuration,
it should load conf/configA.conf and so on for each different
configuration. These would contain lines like:
MACHINE = "A"
or other variables which can be set which can be built in the same
build directory (or change TMPDIR not to conflict).
One downside I've already discovered is that if we want to inherit this
file right at the start of parsing, the only place you can put the
configurations is in "cwd", since BBPATH isn't constructed until the
layers are parsed and therefore using it as a preconf file isn't
possible unless its located there.
Execution of these targets takes the form "bitbake
multiconfig:configA:core-image-minimal core-image-sato" so similar to
our virtclass approach for native/nativesdk/multilib using BBCLASSEXTEND.
Implementation wise, the implication is that instead of tasks being
uniquely referenced with "recipename/fn:task" it now needs to be
"configuration:recipename:task".
We already started using "virtual" filenames for recipes when we
implemented BBCLASSEXTEND and this patch adds a new prefix to
these, "multiconfig:<configname>:" and hence avoid changes to a large
part of the codebase thanks to this. databuilder has an internal array
of data stores and uses the right one depending on the supplied virtual
filename.
That trick allows us to use the existing parsing code including the
multithreading mostly unchanged as well as most of the cache code.
For recipecache, we end up with a dict of these accessed by
multiconfig (mc). taskdata and runqueue can only cope with one recipecache
so for taskdata, we pass in each recipecache and have it compute the result
and end up with an array of taskdatas. We can only have one runqueue so there
extensive changes there.
This initial implementation has some drawbacks:
a) There are no inter-multi-configuration dependencies as yet
b) There are no sstate optimisations. This means if the build uses the
same object twice in say two different TMPDIRs, it will either load from
an existing sstate cache at the start or build it twice. We can then in
due course look at ways in which it would only build it once and then
reuse it. This will likely need significant changes to the way sstate
currently works to make that possible.
(Bitbake rev: 5287991691578825c847bac2368e9b51c0ede3f0)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Rather than passing in a datastore to build on top of, use the data builder
object in the cache and base the parsed recipe from this. This turns
things into proper objects building from one another rather than messy
mixes of static and class functions.
This sets things up so we can support parsing and building multiple
configurations.
(Bitbake rev: fef18b445c0cb6b266cd939b9c78d7cbce38663f)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Worker needs input stream in binary mode as it reads binary content
from it. Current code does it by detaching a buffer from sys.stdin
and assigning it back to sys.stdin.
Detached buffer is io.BufferedReader in binary mode. This operation
is implicit as its purpose is not easily understandable from the code.
Replacing it with fdopen(sys.stdin.fileno(), 'rb') should make the
code more understandable.
Assigning the buffer to sys.stdin is not needed as worker doesn't
use sys.stdin. Moreover, it leads to difficult to debug issues down
the stack. For example, devpyshell doesn't work without reopening
sys.stdin in text mode. This is not needed anymore after this fix as
sys.stdin is not changed in worker code and remains in text mode.
(Bitbake rev: b26bcff4c4d72775f1def7e769015464953b955c)
Signed-off-by: Ed Bartosh <ed.bartosh@linux.intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Various misc changes to convert bitbake to python3 which don't warrant
separation into separate commits.
(Bitbake rev: d0f904d407f57998419bd9c305ce53e5eaa36b24)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The logging domain specified to bb.msg.fatal was invalid. Replace with
a logger.critical() call instead.
(Bitbake rev: 1ffd8737e065a3cd634c74cd67e634d785ea93a5)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
python deprecated logger.warn() in favour of logger.warning(). This is only
used in bitbake code so we may as well just translate everything to avoid
warnings under python 3. Its safe for python 2.7.
(Bitbake rev: 676a5f592e8507e81b8f748d58acfea7572f8796)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
I before E, except after C...
(Bitbake rev: 14c9593265f7469cb8a205a46f845ac7491246df)
Signed-off-by: Phil Blundell <pb@pbcl.net>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
This means that when you view the process tree, the processes
have meaningful names, aiding debugging:
$ pstree -p 30021
bash(30021)───KnottyUI(115579)───Cooker(115590)─┬─PRServ(115592)───{PRServ Handler}(115593)
├─Worker(115630)───bash:sleep(115631)───run.do_sleep.11(115633)───sleep(115634)
└─{ProcessEQueue}(115591)
$ pstree -p 30021
bash(30021)───KnottyUI(117319)───Cooker(117330)─┬─Cooker(117335)
├─PRServ(117332)───{PRServ Handler}(117333)
├─Parser-1:2(117336)
└─{ProcessEQueue}(117331)
Applies to parse threads, PR Server, cooker, the workers and execution
threads, working within the 16 character limit as best we can.
Needed to tweak the bitbake-worker magic values to tell the
workers apart.
(Bitbake rev: 539726a3b2202249a3f148d99e08909cb61902a5)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
In the forked child, we may use multiprocessing. There is only one event
pipe to the worker controlling process and if we're unlucky, multiple
processes can write to it at once corrupting the data by intermixing it.
We don't see this often but when we do, its quite puzzling. I suspect it
only happens in tasks which use multiprocessng (do_rootfs, do_package)
and is much more likely to happen when we have long messages, usually many
times PAGE_SIZE since PAGE_SIZE writes are atomic. This makes it much more
likely within do_roofs, when for example a subprocess lists the contents
of a rootfs.
To fix this, we give each child a Lock() object and use this to serialise
writes to the controlling worker.
(Bitbake rev: 3cb55bdf06148943960e438291f9a562857340a3)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
If the pipe is closed, we want to ensure that we kill any child processes
by triggering the sigterm handler before we exit. This code does that,
hopefully avoiding the remaining process left behind issues on the autobuilder.
(Bitbake rev: 60f6c2818f38c4d9c2d9aaa42acf3071636f4a3b)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
If we SIGKILL cooker (the parent process), ensure the worker notices
and shuts down gracefully. To do this:
* trigger the sigterm handler if the parent exits
* ensure broken pipe writes don't trigger backtraces which interfer with
other exit work
* notice if our command pipe is broken due to EOF and sigterm if so
(Bitbake rev: c43d6a8d711db8d3bd9a1976b9f8e3efdb4cb4ae)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
start/end are unused here and we can improve the code conditional blocks.
(Bitbake rev: 68f53dd77fe0bbfa044bd037a9484e0e1c9088b4)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
I noticed that I was seeing loss of the log files when hitting
control-c while debugging a function in bitbake. In fact if you
take a recipe and replace its compile function as shown below let
it run for a few seconds and hit control-c, you will see first
hand that log data is not there.
do_compile () {
while [ 1 ] ; do
echo -n "Output date: "
date
sleep 1
done
}
It turns out there was a regression introduced by commit:
d0f0e5d9e6 which created the bitbake worker. Since the bitbake
worker is started in its own process space, it needs the exact
same code added from commit: 88429f018b where the problem was
fixed the first time around.
(Bitbake rev: 8d1748f75763b4a66516cc46d5457ee6404b1b68)
Signed-off-by: Jason Wessel <jason.wessel@windriver.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The bug has a long discussion of this. Basically, in some environments,
the exact details of which aren't understood, a Ctrl+C signal to the
UI is being transmitted to all the process children. Looking at the output
of "ps ax -O tpgid", its clear the main process is still the terminal
owner of these processes.
stty -a on a problematic system shows: "-ignbrk brkint"
and on a working system shows: "-ignbrk -brkint"
The description of brkint would suggest this is the problem, setting up
that terminal environment wasn't able to reproduce the problem though.
It was confirmed that using setsid() caused the problem to be resolved
and is probably the right thing to be doing anyway, so lets do it.
[YOCTO #6949]
(Bitbake rev: 461aa73fff0ab616032d28c4fd0322eb88838be6)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Fixed:
1) Run "bitbake recipe" in the terminal
2) Close the terminal while building
3) $ ps aux | grep bitbake-worker
There will be many processes, and they will keep the resources (e.g.,
memory), and won't exit unless kill or kill -9.
(Bitbake rev: 40d2ae0723de2bf5fee343faafb4afda40546839)
Signed-off-by: Robert Yang <liezhi.yang@windriver.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Without this you see:
File "bitbake/bin/bitbake-worker", line 201, in fork_off_task
os._exit(child())
TypeError: an integer is required
(Bitbake rev: cd477b5e77ab0373248b8a8fa30e1c7b8ea984fd)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
We need to transfer some of the siggen data from the core/cooker into
the worker instances. There was a partial API created for this but
its ugly and its not possible to extend it from the siggen class.
This patch completes the interface/abstraction for the data and
means the class can extend/customise it in any siggen class.
(Bitbake rev: cf2d642052979d236185c5b8ca2c5478c06e62ae)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Currently we get no profiling oversight into either the main bitbake worker
process, or the overall parsing before task execution. This adds in extra
profiling hooks so we can truly capture all parts of bitbake's execution
into the profile data.
To do this we modify the 'magic' value passed to bitbake-worker to trigger
the profiling, before the configuration data is sent over to the worker.
(Bitbake rev: 446e490bf485b712e5cee733dab5805254cdcad0)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
When processes terminate, we really want all of the child processes to
terminate too. This was not happening for worker processes which spawned their
own multiprocessing pools, leading to build hangs. This change ensures any
sigterm gets passed to the whole process group. In local tests, this resolved
some hanging process workloads I could generate. It does rely on signals
being delivered in a timely fashion and there is a multiprocessing bug we have
to work around there.
(Bitbake rev: 96f8ea07ace1379380fab2d78eb592fa40c867d4)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Iterating through and calling setVar on this number of variables has significant
overhead in the profiling data. By not setting this, we save 3,000 calls
to setVar which gives a noticeable improvement to the speed of task execution.
The BBHASH variables have since been replaced by accessing that data through
the siggen code and going forward, that is the preferred way work with it.
(Bitbake rev: 92526eadd09d19938762290e0492076174367583)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
The children of the worker should have the default SIGTERM handler,
else they'll try and do cleanup which should only happen in the
parent leading to all kinds of bizarre build failures.
(Bitbake rev: a53c8d1f846d94082aa459996c4114f10970b8ef)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Currently if bitbake-worker handles a SIGTERM, it leaves the child
processes to complete or hang. It shouldn't do this so hook the SIGTERM
event and gracefully shutdown any children.
(Bitbake rev: 551406f3f9ee94de09d2da6e16fea054c6dbfdb7)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Currently tasks have no knowledge of which other tasks they depend
upon. This makes it impossible to do at least two things which would be
desirable/interesting:
a) Have the ability to create per recipe sysroots
b) Allow the aclocal files to be present only for the entries in
DEPENDS (directly and indirectly)
By exporting task data through this new variable, tasks can inspect
their dependencies and then take actions based upon this.
(Bitbake rev: 84f1dde717dac22435005b79d03ee0b80a3e8e62)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
When using the dry run option (-n), bitbake would still try and fire
a specific fakeroot worker. This is doomed to failure since it might
well not have been built.
Add in some checks to prevent the failures.
[YOCTO #5367]
(Bitbake rev: f34d0606f87ce9dacadeb78bac35879b74f10559)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Due to the worker split the ${DATE} and ${TIME} variables could end up
with different values for different workers.
E.g., a task like do_rootfs that is run within a fakeroot environment
had a slightly different view of the time than another task that was not
fakerooted which made it impossible to correctly refer to the image
generated by do_rootfs from the other task.
(Bitbake rev: 756cc69ebf8bfe8455d0c90f288dd51be2499773)
Signed-off-by: Peter Kjellerstedt <peter.kjellerstedt@axis.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
BUILDNAME is set from cooker by default, so since the worker split it
will not be set when executing functions. In OpenEmbedded this results
in /etc/version (which is populated from BUILDNAME) not having any
content. Pass this variable value through to the worker explicitly to
fix the issue.
Fixes [YOCTO #4818].
(Bitbake rev: 92940b0427d9b2b3f95e27c230ec1e36638a34bc)
Signed-off-by: Paul Eggleton <paul.eggleton@linux.intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
bitbake-worker makes use of the signal module
but it doesn't import it. This patch fixes the issue.
[YOCTO #4750]
(Bitbake rev: c2ed639690f135994199eb24d964e37f57259e3a)
Signed-off-by: Valentin Popa <valentin.popa@intel.com>
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
With the change to bitbake-worker we need to ensure the workers know
how to contact the PR service, the magic 0 port and singleton is
no longer enough.
(Bitbake rev: c761751e259bb8e940552a28794b45887b5a72d9)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
This is a pretty fundamental change to the way bitbake operates. It
splits out the task execution part of runqueue into a completely
separately exec'd process called bitbake-worker.
This means that the separate process has to build its own datastore and
that configuration needs to be passed from the cooker over to the
bitbake worker process.
Known issues:
* Hob is broken with this patch since it writes to the configuration
and that configuration isn't preserved in bitbake-worker.
* We create a worker for setscene, then a new worker for the main task
execution. This is wasteful but shouldn't be hard to fix.
* We probably send too much data over to bitbake-worker, need to
see if we can streamline it.
These are issues which will be followed up in subsequent patches.
This patch sets the groundwork for the removal of the double bitbake
execution for psuedo which will be in a follow on patch.
(Bitbake rev: b2e26f1db28d74f2dd9df8ab4ed3b472503b9a5c)
Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Richard Purdie
Paul Eggleton
Joshua Lock
Markus Lehtonen
Ed Bartosh
Christopher Larson
Phil Blundell
Jason Wessel
Robert Yang
Peter Kjellerstedt
Valentin Popa