[PATCH 1/2] cooker: Try and avoid parseing hangs

[PATCH 1/2] cooker: Try and avoid parseing hangs

[Richard Purdie]

We sometimes see hangs in parsing during automated testing. It appears that
SIGINT was sent to the underlying processes which see KeyboardInterrupt but
they're stuck trying to write into the results pipe. The SIGINT was probably
from some kind of parsing failure which doens't happen often, hence the hang
being rare (in the incompatible license selftests from OE).

This patch:
  * sets a flag to indicate exit upon SIGINT so the exit is more graceful
    and a defined exit path
  * empties the results queue after we send the quit event
  * empties the results queue after the SIGINT for good measure
  * increases the 0.5s timeout to 2s since we now have some very slow to
    parse recipes due to class extensions (ptests)

This should hopefully make the parsing failure codepaths more robust.

Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
---
 lib/bb/cooker.py | 24 +++++++++++++++---------
 1 file changed, 15 insertions(+), 9 deletions(-)

diff --git a/lib/bb/cooker.py b/lib/bb/cooker.py
index 1810bcc6049..91e3ee025ea 100644
--- a/lib/bb/cooker.py
+++ b/lib/bb/cooker.py
@@ -2009,6 +2009,7 @@ class Parser(multiprocessing.Process):
         self.queue_signals = False
         self.signal_received = []
         self.signal_threadlock = threading.Lock()
+        self.exit = False
 
     def catch_sig(self, signum, frame):
         if self.queue_signals:
@@ -2021,7 +2022,7 @@ class Parser(multiprocessing.Process):
             signal.signal(signal.SIGTERM, signal.SIG_DFL)
             os.kill(os.getpid(), signal.SIGTERM)
         elif signum == signal.SIGINT:
-            signal.default_int_handler(signum, frame)
+            self.exit = True
 
     def run(self):
 
@@ -2059,7 +2060,7 @@ class Parser(multiprocessing.Process):
         pending = []
         havejobs = True
         try:
-            while havejobs or pending:
+            while (havejobs or pending) and not self.exit:
                 if self.quit.is_set():
                     break
 
@@ -2196,11 +2197,12 @@ class CookerParser(object):
 
         # Cleanup the queue before call process.join(), otherwise there might be
         # deadlocks.
-        while True:
-            try:
-               self.result_queue.get(timeout=0.25)
-            except queue.Empty:
-                break
+        def read_results():
+            while True:
+                try:
+                   self.result_queue.get(timeout=0.25)
+                except queue.Empty:
+                    break
 
         def sync_caches():
             for c in self.bb_caches.values():
@@ -2212,15 +2214,19 @@ class CookerParser(object):
 
         self.parser_quit.set()
 
+        read_results()
+
         for process in self.processes:
-            process.join(0.5)
+            process.join(2)
 
         for process in self.processes:
             if process.exitcode is None:
                 os.kill(process.pid, signal.SIGINT)
 
+        read_results()
+
         for process in self.processes:
-            process.join(0.5)
+            process.join(2)
 
         for process in self.processes:
             if process.exitcode is None:

[PATCH 2/2] cooker: Use a queue to feed parsing jobs

[Richard Purdie]

Curerntly, recipes to parse are split into equal groups and passed to
each parse thread at the start of parsing. We can replace this with
a queue and collect a new job as each parsing process becomes idle
to better spread load in the case of slow parsing jobs.

Some of the data we need has to be passed in at fork time since it
can't be pickled, so the job to parse is only referenced as an index
in that list.

This should better spread load for slow to parse recipes such as those
with many class extensions.

Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
---
 lib/bb/cooker.py | 29 ++++++++++++++++++-----------
 1 file changed, 18 insertions(+), 11 deletions(-)

diff --git a/lib/bb/cooker.py b/lib/bb/cooker.py
index 91e3ee025ea..e88ad24cf61 100644
--- a/lib/bb/cooker.py
+++ b/lib/bb/cooker.py
@@ -1998,8 +1998,9 @@ class ParsingFailure(Exception):
         Exception.__init__(self, realexception, recipe)
 
 class Parser(multiprocessing.Process):
-    def __init__(self, jobs, results, quit, profile):
+    def __init__(self, jobs, jobid_queue, results, quit, profile):
         self.jobs = jobs
+        self.jobid_queue = jobid_queue
         self.results = results
         self.quit = quit
         multiprocessing.Process.__init__(self)
@@ -2064,12 +2065,14 @@ class Parser(multiprocessing.Process):
                 if self.quit.is_set():
                     break
 
-                job = None
+                jobid = None
                 try:
-                    job = self.jobs.pop()
-                except IndexError:
+                    jobid = self.jobid_queue.get(True, 0.5)
+                except (ValueError, OSError):
                     havejobs = False
-                if job:
+
+                if jobid is not None:
+                    job = self.jobs[jobid]
                     result = self.parse(*job)
                     # Clear the siggen cache after parsing to control memory usage, its huge
                     bb.parse.siggen.postparsing_clean_cache()
@@ -2082,6 +2085,7 @@ class Parser(multiprocessing.Process):
                     except queue.Full:
                         pending.append(result)
         finally:
+            self.jobs.close()
             self.results.close()
             self.results.join_thread()
 
@@ -2134,13 +2138,13 @@ class CookerParser(object):
 
         self.bb_caches = bb.cache.MulticonfigCache(self.cfgbuilder, self.cfghash, cooker.caches_array)
         self.fromcache = set()
-        self.willparse = set()
+        self.willparse = []
         for mc in self.cooker.multiconfigs:
             for filename in self.mcfilelist[mc]:
                 appends = self.cooker.collections[mc].get_file_appends(filename)
                 layername = self.cooker.collections[mc].calc_bbfile_priority(filename)[2]
                 if not self.bb_caches[mc].cacheValid(filename, appends):
-                    self.willparse.add((mc, self.bb_caches[mc], filename, appends, layername))
+                    self.willparse.append((mc, self.bb_caches[mc], filename, appends, layername))
                 else:
                     self.fromcache.add((mc, self.bb_caches[mc], filename, appends, layername))
 
@@ -2159,22 +2163,25 @@ class CookerParser(object):
     def start(self):
         self.results = self.load_cached()
         self.processes = []
+
         if self.toparse:
             bb.event.fire(bb.event.ParseStarted(self.toparse), self.cfgdata)
 
+            self.toparse_queue = multiprocessing.Queue(len(self.willparse))
             self.parser_quit = multiprocessing.Event()
             self.result_queue = multiprocessing.Queue()
 
-            def chunkify(lst,n):
-                return [lst[i::n] for i in range(n)]
-            self.jobs = chunkify(list(self.willparse), self.num_processes)
+            for jobid in range(len(self.willparse)):
+                self.toparse_queue.put(jobid)
 
             for i in range(0, self.num_processes):
-                parser = Parser(self.jobs[i], self.result_queue, self.parser_quit, self.cooker.configuration.profile)
+                parser = Parser(self.willparse, self.toparse_queue, self.result_queue, self.parser_quit, self.cooker.configuration.profile)
                 parser.start()
                 self.process_names.append(parser.name)
                 self.processes.append(parser)
 
+            self.toparse_queue.close()
+
             self.results = itertools.chain(self.results, self.parse_generator())
 
     def shutdown(self, clean=True, eventmsg="Parsing halted due to errors"):

Re: [bitbake-devel] [PATCH 2/2] cooker: Use a queue to feed parsing jobs

[Joshua Watt]

On Wed, Jul 2, 2025 at 4:24 PM Richard Purdie via
lists.openembedded.org
<richard.purdie=linuxfoundation.org@lists.openembedded.org> wrote:
>
> Curerntly, recipes to parse are split into equal groups and passed to
> each parse thread at the start of parsing. We can replace this with
> a queue and collect a new job as each parsing process becomes idle
> to better spread load in the case of slow parsing jobs.
>
> Some of the data we need has to be passed in at fork time since it
> can't be pickled, so the job to parse is only referenced as an index
> in that list.
>
> This should better spread load for slow to parse recipes such as those
> with many class extensions.
>
> Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
> ---
>  lib/bb/cooker.py | 29 ++++++++++++++++++-----------
>  1 file changed, 18 insertions(+), 11 deletions(-)
>
> diff --git a/lib/bb/cooker.py b/lib/bb/cooker.py
> index 91e3ee025ea..e88ad24cf61 100644
> --- a/lib/bb/cooker.py
> +++ b/lib/bb/cooker.py
> @@ -1998,8 +1998,9 @@ class ParsingFailure(Exception):
>          Exception.__init__(self, realexception, recipe)
>
>  class Parser(multiprocessing.Process):
> -    def __init__(self, jobs, results, quit, profile):
> +    def __init__(self, jobs, jobid_queue, results, quit, profile):
>          self.jobs = jobs
> +        self.jobid_queue = jobid_queue
>          self.results = results
>          self.quit = quit
>          multiprocessing.Process.__init__(self)
> @@ -2064,12 +2065,14 @@ class Parser(multiprocessing.Process):
>                  if self.quit.is_set():
>                      break
>
> -                job = None
> +                jobid = None
>                  try:
> -                    job = self.jobs.pop()
> -                except IndexError:
> +                    jobid = self.jobid_queue.get(True, 0.5)
> +                except (ValueError, OSError):
>                      havejobs = False
> -                if job:
> +
> +                if jobid is not None:
> +                    job = self.jobs[jobid]
>                      result = self.parse(*job)
>                      # Clear the siggen cache after parsing to control memory usage, its huge
>                      bb.parse.siggen.postparsing_clean_cache()
> @@ -2082,6 +2085,7 @@ class Parser(multiprocessing.Process):
>                      except queue.Full:
>                          pending.append(result)
>          finally:
> +            self.jobs.close()
>              self.results.close()
>              self.results.join_thread()
>
> @@ -2134,13 +2138,13 @@ class CookerParser(object):
>
>          self.bb_caches = bb.cache.MulticonfigCache(self.cfgbuilder, self.cfghash, cooker.caches_array)
>          self.fromcache = set()
> -        self.willparse = set()
> +        self.willparse = []
>          for mc in self.cooker.multiconfigs:
>              for filename in self.mcfilelist[mc]:
>                  appends = self.cooker.collections[mc].get_file_appends(filename)
>                  layername = self.cooker.collections[mc].calc_bbfile_priority(filename)[2]
>                  if not self.bb_caches[mc].cacheValid(filename, appends):
> -                    self.willparse.add((mc, self.bb_caches[mc], filename, appends, layername))
> +                    self.willparse.append((mc, self.bb_caches[mc], filename, appends, layername))
>                  else:
>                      self.fromcache.add((mc, self.bb_caches[mc], filename, appends, layername))
>
> @@ -2159,22 +2163,25 @@ class CookerParser(object):
>      def start(self):
>          self.results = self.load_cached()
>          self.processes = []
> +
>          if self.toparse:
>              bb.event.fire(bb.event.ParseStarted(self.toparse), self.cfgdata)
>
> +            self.toparse_queue = multiprocessing.Queue(len(self.willparse))
>              self.parser_quit = multiprocessing.Event()
>              self.result_queue = multiprocessing.Queue()
>
> -            def chunkify(lst,n):
> -                return [lst[i::n] for i in range(n)]
> -            self.jobs = chunkify(list(self.willparse), self.num_processes)
> +            for jobid in range(len(self.willparse)):
> +                self.toparse_queue.put(jobid)

It's generally not a good idea to assume you can push all the items on
the queue before there are any consumers; if the queue fills, this
will deadlock. It would be better to push the items to the queue after
creating the processes. If you do that, I also don't see any reason to
push the indexes instead of the jobs directly; basically instead of
"chunkifying" and pre-determining the assignment of jobs to processes,
have them all pull from one queue of jobs.

>
>              for i in range(0, self.num_processes):
> -                parser = Parser(self.jobs[i], self.result_queue, self.parser_quit, self.cooker.configuration.profile)
> +                parser = Parser(self.willparse, self.toparse_queue, self.result_queue, self.parser_quit, self.cooker.configuration.profile)
>                  parser.start()
>                  self.process_names.append(parser.name)
>                  self.processes.append(parser)
>
> +            self.toparse_queue.close()
> +
>              self.results = itertools.chain(self.results, self.parse_generator())
>
>      def shutdown(self, clean=True, eventmsg="Parsing halted due to errors"):
>
> -=-=-=-=-=-=-=-=-=-=-=-
> Links: You receive all messages sent to this group.
> View/Reply Online (#17738): https://lists.openembedded.org/g/bitbake-devel/message/17738
> Mute This Topic: https://lists.openembedded.org/mt/113957083/3616693
> Group Owner: bitbake-devel+owner@lists.openembedded.org
> Unsubscribe: https://lists.openembedded.org/g/bitbake-devel/unsub [JPEWhacker@gmail.com]
> -=-=-=-=-=-=-=-=-=-=-=-
>

Re: [bitbake-devel] [PATCH 2/2] cooker: Use a queue to feed parsing jobs

[Richard Purdie]

On Thu, 2025-07-03 at 08:27 -0600, Joshua Watt wrote:
> On Wed, Jul 2, 2025 at 4:24 PM Richard Purdie via
> lists.openembedded.org
> <richard.purdie=linuxfoundation.org@lists.openembedded.org> wrote:
> > 
> > Curerntly, recipes to parse are split into equal groups and passed to
> > each parse thread at the start of parsing. We can replace this with
> > a queue and collect a new job as each parsing process becomes idle
> > to better spread load in the case of slow parsing jobs.
> > 
> > Some of the data we need has to be passed in at fork time since it
> > can't be pickled, so the job to parse is only referenced as an index
> > in that list.
> > 
> > This should better spread load for slow to parse recipes such as those
> > with many class extensions.
> > 
> > Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
> > ---
> >  lib/bb/cooker.py | 29 ++++++++++++++++++-----------
> >  1 file changed, 18 insertions(+), 11 deletions(-)
> > 
> > diff --git a/lib/bb/cooker.py b/lib/bb/cooker.py
> > index 91e3ee025ea..e88ad24cf61 100644
> > --- a/lib/bb/cooker.py
> > +++ b/lib/bb/cooker.py
> > @@ -1998,8 +1998,9 @@ class ParsingFailure(Exception):
> >          Exception.__init__(self, realexception, recipe)
> > 
> >  class Parser(multiprocessing.Process):
> > -    def __init__(self, jobs, results, quit, profile):
> > +    def __init__(self, jobs, jobid_queue, results, quit, profile):
> >          self.jobs = jobs
> > +        self.jobid_queue = jobid_queue
> >          self.results = results
> >          self.quit = quit
> >          multiprocessing.Process.__init__(self)
> > @@ -2064,12 +2065,14 @@ class Parser(multiprocessing.Process):
> >                  if self.quit.is_set():
> >                      break
> > 
> > -                job = None
> > +                jobid = None
> >                  try:
> > -                    job = self.jobs.pop()
> > -                except IndexError:
> > +                    jobid = self.jobid_queue.get(True, 0.5)
> > +                except (ValueError, OSError):
> >                      havejobs = False
> > -                if job:
> > +
> > +                if jobid is not None:
> > +                    job = self.jobs[jobid]
> >                      result = self.parse(*job)
> >                      # Clear the siggen cache after parsing to control memory usage, its huge
> >                      bb.parse.siggen.postparsing_clean_cache()
> > @@ -2082,6 +2085,7 @@ class Parser(multiprocessing.Process):
> >                      except queue.Full:
> >                          pending.append(result)
> >          finally:
> > +            self.jobs.close()
> >              self.results.close()
> >              self.results.join_thread()
> > 
> > @@ -2134,13 +2138,13 @@ class CookerParser(object):
> > 
> >          self.bb_caches = bb.cache.MulticonfigCache(self.cfgbuilder, self.cfghash, cooker.caches_array)
> >          self.fromcache = set()
> > -        self.willparse = set()
> > +        self.willparse = []
> >          for mc in self.cooker.multiconfigs:
> >              for filename in self.mcfilelist[mc]:
> >                  appends = self.cooker.collections[mc].get_file_appends(filename)
> >                  layername = self.cooker.collections[mc].calc_bbfile_priority(filename)[2]
> >                  if not self.bb_caches[mc].cacheValid(filename, appends):
> > -                    self.willparse.add((mc, self.bb_caches[mc], filename, appends, layername))
> > +                    self.willparse.append((mc, self.bb_caches[mc], filename, appends, layername))
> >                  else:
> >                      self.fromcache.add((mc, self.bb_caches[mc], filename, appends, layername))
> > 
> > @@ -2159,22 +2163,25 @@ class CookerParser(object):
> >      def start(self):
> >          self.results = self.load_cached()
> >          self.processes = []
> > +
> >          if self.toparse:
> >              bb.event.fire(bb.event.ParseStarted(self.toparse), self.cfgdata)
> > 
> > +            self.toparse_queue = multiprocessing.Queue(len(self.willparse))
> >              self.parser_quit = multiprocessing.Event()
> >              self.result_queue = multiprocessing.Queue()
> > 
> > -            def chunkify(lst,n):
> > -                return [lst[i::n] for i in range(n)]
> > -            self.jobs = chunkify(list(self.willparse), self.num_processes)
> > +            for jobid in range(len(self.willparse)):
> > +                self.toparse_queue.put(jobid)
> 
> It's generally not a good idea to assume you can push all the items on
> the queue before there are any consumers; if the queue fills, this
> will deadlock. It would be better to push the items to the queue after
> creating the processes.

True. I did make sure the queue was large enough above however it would
depend on the size of the job entries.

>  If you do that, I also don't see any reason to
> push the indexes instead of the jobs directly; basically instead of
> "chunkifying" and pre-determining the assignment of jobs to processes,
> have them all pull from one queue of jobs.

I added this comment to the patch in master next:
 
# Have to pass in willparse at fork time so all parsing processes have the unpickleable data
# then access it by index from the parse queue.


As if you don't do that, it can't rebuild the data as the cache objects
aren't pickleable :(

Cheers,

Richard

Re: [bitbake-devel] [PATCH 2/2] cooker: Use a queue to feed parsing jobs

[Joshua Watt]

On Thu, Jul 3, 2025 at 8:30 AM Richard Purdie
<richard.purdie@linuxfoundation.org> wrote:
>
> On Thu, 2025-07-03 at 08:27 -0600, Joshua Watt wrote:
> > On Wed, Jul 2, 2025 at 4:24 PM Richard Purdie via
> > lists.openembedded.org
> > <richard.purdie=linuxfoundation.org@lists.openembedded.org> wrote:
> > >
> > > Curerntly, recipes to parse are split into equal groups and passed to
> > > each parse thread at the start of parsing. We can replace this with
> > > a queue and collect a new job as each parsing process becomes idle
> > > to better spread load in the case of slow parsing jobs.
> > >
> > > Some of the data we need has to be passed in at fork time since it
> > > can't be pickled, so the job to parse is only referenced as an index
> > > in that list.
> > >
> > > This should better spread load for slow to parse recipes such as those
> > > with many class extensions.
> > >
> > > Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
> > > ---
> > >  lib/bb/cooker.py | 29 ++++++++++++++++++-----------
> > >  1 file changed, 18 insertions(+), 11 deletions(-)
> > >
> > > diff --git a/lib/bb/cooker.py b/lib/bb/cooker.py
> > > index 91e3ee025ea..e88ad24cf61 100644
> > > --- a/lib/bb/cooker.py
> > > +++ b/lib/bb/cooker.py
> > > @@ -1998,8 +1998,9 @@ class ParsingFailure(Exception):
> > >          Exception.__init__(self, realexception, recipe)
> > >
> > >  class Parser(multiprocessing.Process):
> > > -    def __init__(self, jobs, results, quit, profile):
> > > +    def __init__(self, jobs, jobid_queue, results, quit, profile):
> > >          self.jobs = jobs
> > > +        self.jobid_queue = jobid_queue
> > >          self.results = results
> > >          self.quit = quit
> > >          multiprocessing.Process.__init__(self)
> > > @@ -2064,12 +2065,14 @@ class Parser(multiprocessing.Process):
> > >                  if self.quit.is_set():
> > >                      break
> > >
> > > -                job = None
> > > +                jobid = None
> > >                  try:
> > > -                    job = self.jobs.pop()
> > > -                except IndexError:
> > > +                    jobid = self.jobid_queue.get(True, 0.5)
> > > +                except (ValueError, OSError):
> > >                      havejobs = False
> > > -                if job:
> > > +
> > > +                if jobid is not None:
> > > +                    job = self.jobs[jobid]
> > >                      result = self.parse(*job)
> > >                      # Clear the siggen cache after parsing to control memory usage, its huge
> > >                      bb.parse.siggen.postparsing_clean_cache()
> > > @@ -2082,6 +2085,7 @@ class Parser(multiprocessing.Process):
> > >                      except queue.Full:
> > >                          pending.append(result)
> > >          finally:
> > > +            self.jobs.close()
> > >              self.results.close()
> > >              self.results.join_thread()
> > >
> > > @@ -2134,13 +2138,13 @@ class CookerParser(object):
> > >
> > >          self.bb_caches = bb.cache.MulticonfigCache(self.cfgbuilder, self.cfghash, cooker.caches_array)
> > >          self.fromcache = set()
> > > -        self.willparse = set()
> > > +        self.willparse = []
> > >          for mc in self.cooker.multiconfigs:
> > >              for filename in self.mcfilelist[mc]:
> > >                  appends = self.cooker.collections[mc].get_file_appends(filename)
> > >                  layername = self.cooker.collections[mc].calc_bbfile_priority(filename)[2]
> > >                  if not self.bb_caches[mc].cacheValid(filename, appends):
> > > -                    self.willparse.add((mc, self.bb_caches[mc], filename, appends, layername))
> > > +                    self.willparse.append((mc, self.bb_caches[mc], filename, appends, layername))
> > >                  else:
> > >                      self.fromcache.add((mc, self.bb_caches[mc], filename, appends, layername))
> > >
> > > @@ -2159,22 +2163,25 @@ class CookerParser(object):
> > >      def start(self):
> > >          self.results = self.load_cached()
> > >          self.processes = []
> > > +
> > >          if self.toparse:
> > >              bb.event.fire(bb.event.ParseStarted(self.toparse), self.cfgdata)
> > >
> > > +            self.toparse_queue = multiprocessing.Queue(len(self.willparse))
> > >              self.parser_quit = multiprocessing.Event()
> > >              self.result_queue = multiprocessing.Queue()
> > >
> > > -            def chunkify(lst,n):
> > > -                return [lst[i::n] for i in range(n)]
> > > -            self.jobs = chunkify(list(self.willparse), self.num_processes)
> > > +            for jobid in range(len(self.willparse)):
> > > +                self.toparse_queue.put(jobid)
> >
> > It's generally not a good idea to assume you can push all the items on
> > the queue before there are any consumers; if the queue fills, this
> > will deadlock. It would be better to push the items to the queue after
> > creating the processes.
>
> True. I did make sure the queue was large enough above however it would
> depend on the size of the job entries.
>
> >  If you do that, I also don't see any reason to
> > push the indexes instead of the jobs directly; basically instead of
> > "chunkifying" and pre-determining the assignment of jobs to processes,
> > have them all pull from one queue of jobs.
>
> I added this comment to the patch in master next:
>
> # Have to pass in willparse at fork time so all parsing processes have the unpickleable data
> # then access it by index from the parse queue.
>
>
> As if you don't do that, it can't rebuild the data as the cache objects
> aren't pickleable :(

Ah, OK. We must be relying on the multiprocess being implemented with
fork() then to pass the data. That's reasonable and the indexes make
sense (but we should still push them after creating the process to
avoid the possibility of deadlock)

>
> Cheers,
>
> Richard

Re: [bitbake-devel] [PATCH 2/2] cooker: Use a queue to feed parsing jobs

[Richard Purdie]

On Thu, 2025-07-03 at 08:50 -0600, Joshua Watt wrote:
> On Thu, Jul 3, 2025 at 8:30 AM Richard Purdie
> <richard.purdie@linuxfoundation.org> wrote:
> > 
> > On Thu, 2025-07-03 at 08:27 -0600, Joshua Watt wrote:
> > > On Wed, Jul 2, 2025 at 4:24 PM Richard Purdie via
> > > lists.openembedded.org
> > > <richard.purdie=linuxfoundation.org@lists.openembedded.org>
> > > wrote:
> > > > 
> > > > Curerntly, recipes to parse are split into equal groups and
> > > > passed to
> > > > each parse thread at the start of parsing. We can replace this
> > > > with
> > > > a queue and collect a new job as each parsing process becomes
> > > > idle
> > > > to better spread load in the case of slow parsing jobs.
> > > > 
> > > > Some of the data we need has to be passed in at fork time since
> > > > it
> > > > can't be pickled, so the job to parse is only referenced as an
> > > > index
> > > > in that list.
> > > > 
> > > > This should better spread load for slow to parse recipes such
> > > > as those
> > > > with many class extensions.
> > > > 
> > > > Signed-off-by: Richard Purdie
> > > > <richard.purdie@linuxfoundation.org>
> > > > ---
> > > >  lib/bb/cooker.py | 29 ++++++++++++++++++-----------
> > > >  1 file changed, 18 insertions(+), 11 deletions(-)
> > > > 
> > > > diff --git a/lib/bb/cooker.py b/lib/bb/cooker.py
> > > > index 91e3ee025ea..e88ad24cf61 100644
> > > > --- a/lib/bb/cooker.py
> > > > +++ b/lib/bb/cooker.py
> > > > @@ -1998,8 +1998,9 @@ class ParsingFailure(Exception):
> > > >          Exception.__init__(self, realexception, recipe)
> > > > 
> > > >  class Parser(multiprocessing.Process):
> > > > -    def __init__(self, jobs, results, quit, profile):
> > > > +    def __init__(self, jobs, jobid_queue, results, quit,
> > > > profile):
> > > >          self.jobs = jobs
> > > > +        self.jobid_queue = jobid_queue
> > > >          self.results = results
> > > >          self.quit = quit
> > > >          multiprocessing.Process.__init__(self)
> > > > @@ -2064,12 +2065,14 @@ class Parser(multiprocessing.Process):
> > > >                  if self.quit.is_set():
> > > >                      break
> > > > 
> > > > -                job = None
> > > > +                jobid = None
> > > >                  try:
> > > > -                    job = self.jobs.pop()
> > > > -                except IndexError:
> > > > +                    jobid = self.jobid_queue.get(True, 0.5)
> > > > +                except (ValueError, OSError):
> > > >                      havejobs = False
> > > > -                if job:
> > > > +
> > > > +                if jobid is not None:
> > > > +                    job = self.jobs[jobid]
> > > >                      result = self.parse(*job)
> > > >                      # Clear the siggen cache after parsing to
> > > > control memory usage, its huge
> > > >                      bb.parse.siggen.postparsing_clean_cache()
> > > > @@ -2082,6 +2085,7 @@ class Parser(multiprocessing.Process):
> > > >                      except queue.Full:
> > > >                          pending.append(result)
> > > >          finally:
> > > > +            self.jobs.close()
> > > >              self.results.close()
> > > >              self.results.join_thread()
> > > > 
> > > > @@ -2134,13 +2138,13 @@ class CookerParser(object):
> > > > 
> > > >          self.bb_caches =
> > > > bb.cache.MulticonfigCache(self.cfgbuilder, self.cfghash,
> > > > cooker.caches_array)
> > > >          self.fromcache = set()
> > > > -        self.willparse = set()
> > > > +        self.willparse = []
> > > >          for mc in self.cooker.multiconfigs:
> > > >              for filename in self.mcfilelist[mc]:
> > > >                  appends =
> > > > self.cooker.collections[mc].get_file_appends(filename)
> > > >                  layername =
> > > > self.cooker.collections[mc].calc_bbfile_priority(filename)[2]
> > > >                  if not self.bb_caches[mc].cacheValid(filename,
> > > > appends):
> > > > -                    self.willparse.add((mc,
> > > > self.bb_caches[mc], filename, appends, layername))
> > > > +                    self.willparse.append((mc,
> > > > self.bb_caches[mc], filename, appends, layername))
> > > >                  else:
> > > >                      self.fromcache.add((mc,
> > > > self.bb_caches[mc], filename, appends, layername))
> > > > 
> > > > @@ -2159,22 +2163,25 @@ class CookerParser(object):
> > > >      def start(self):
> > > >          self.results = self.load_cached()
> > > >          self.processes = []
> > > > +
> > > >          if self.toparse:
> > > >              bb.event.fire(bb.event.ParseStarted(self.toparse),
> > > > self.cfgdata)
> > > > 
> > > > +            self.toparse_queue =
> > > > multiprocessing.Queue(len(self.willparse))
> > > >              self.parser_quit = multiprocessing.Event()
> > > >              self.result_queue = multiprocessing.Queue()
> > > > 
> > > > -            def chunkify(lst,n):
> > > > -                return [lst[i::n] for i in range(n)]
> > > > -            self.jobs = chunkify(list(self.willparse),
> > > > self.num_processes)
> > > > +            for jobid in range(len(self.willparse)):
> > > > +                self.toparse_queue.put(jobid)
> > > 
> > > It's generally not a good idea to assume you can push all the
> > > items on
> > > the queue before there are any consumers; if the queue fills,
> > > this
> > > will deadlock. It would be better to push the items to the queue
> > > after
> > > creating the processes.
> > 
> > True. I did make sure the queue was large enough above however it
> > would
> > depend on the size of the job entries.
> > 
> > >  If you do that, I also don't see any reason to
> > > push the indexes instead of the jobs directly; basically instead
> > > of
> > > "chunkifying" and pre-determining the assignment of jobs to
> > > processes,
> > > have them all pull from one queue of jobs.
> > 
> > I added this comment to the patch in master next:
> > 
> > # Have to pass in willparse at fork time so all parsing processes
> > have the unpickleable data
> > # then access it by index from the parse queue.
> > 
> > 
> > As if you don't do that, it can't rebuild the data as the cache
> > objects
> > aren't pickleable :(
> 
> Ah, OK. We must be relying on the multiprocess being implemented with
> fork() then to pass the data. That's reasonable and the indexes make
> sense (but we should still push them after creating the process to
> avoid the possibility of deadlock)

I changed it to queue after starting the parser processes however I
think they exit early due to no work and it causes occasional build
failures that way around so the code is going to need more extensive
changes.

Cheers,

Richard

Re: [bitbake-devel] [PATCH 2/2] cooker: Use a queue to feed parsing jobs

[Richard Purdie]

On Fri, 2025-07-04 at 22:42 +0100, Richard Purdie via lists.openembedded.org wrote:
> On Thu, 2025-07-03 at 08:50 -0600, Joshua Watt wrote:
> > On Thu, Jul 3, 2025 at 8:30 AM Richard Purdie
> > <richard.purdie@linuxfoundation.org> wrote:
> > > 
> > > On Thu, 2025-07-03 at 08:27 -0600, Joshua Watt wrote:
> > > > On Wed, Jul 2, 2025 at 4:24 PM Richard Purdie via
> > > > lists.openembedded.org
> > > > <richard.purdie=linuxfoundation.org@lists.openembedded.org>
> > > > wrote:
> > > > > 
> > > > > Curerntly, recipes to parse are split into equal groups and
> > > > > passed to
> > > > > each parse thread at the start of parsing. We can replace this
> > > > > with
> > > > > a queue and collect a new job as each parsing process becomes
> > > > > idle
> > > > > to better spread load in the case of slow parsing jobs.
> > > > > 
> > > > > Some of the data we need has to be passed in at fork time since
> > > > > it
> > > > > can't be pickled, so the job to parse is only referenced as an
> > > > > index
> > > > > in that list.
> > > > > 
> > > > > This should better spread load for slow to parse recipes such
> > > > > as those
> > > > > with many class extensions.
> > > > > 
> > > > > Signed-off-by: Richard Purdie
> > > > > <richard.purdie@linuxfoundation.org>
> > > > > ---
> > > > >  lib/bb/cooker.py | 29 ++++++++++++++++++-----------
> > > > >  1 file changed, 18 insertions(+), 11 deletions(-)
> > > > > 
> > > > > diff --git a/lib/bb/cooker.py b/lib/bb/cooker.py
> > > > > index 91e3ee025ea..e88ad24cf61 100644
> > > > > --- a/lib/bb/cooker.py
> > > > > +++ b/lib/bb/cooker.py
> > > > > @@ -1998,8 +1998,9 @@ class ParsingFailure(Exception):
> > > > >          Exception.__init__(self, realexception, recipe)
> > > > > 
> > > > >  class Parser(multiprocessing.Process):
> > > > > -    def __init__(self, jobs, results, quit, profile):
> > > > > +    def __init__(self, jobs, jobid_queue, results, quit,
> > > > > profile):
> > > > >          self.jobs = jobs
> > > > > +        self.jobid_queue = jobid_queue
> > > > >          self.results = results
> > > > >          self.quit = quit
> > > > >          multiprocessing.Process.__init__(self)
> > > > > @@ -2064,12 +2065,14 @@ class Parser(multiprocessing.Process):
> > > > >                  if self.quit.is_set():
> > > > >                      break
> > > > > 
> > > > > -                job = None
> > > > > +                jobid = None
> > > > >                  try:
> > > > > -                    job = self.jobs.pop()
> > > > > -                except IndexError:
> > > > > +                    jobid = self.jobid_queue.get(True, 0.5)
> > > > > +                except (ValueError, OSError):
> > > > >                      havejobs = False
> > > > > -                if job:
> > > > > +
> > > > > +                if jobid is not None:
> > > > > +                    job = self.jobs[jobid]
> > > > >                      result = self.parse(*job)
> > > > >                      # Clear the siggen cache after parsing to
> > > > > control memory usage, its huge
> > > > >                      bb.parse.siggen.postparsing_clean_cache()
> > > > > @@ -2082,6 +2085,7 @@ class Parser(multiprocessing.Process):
> > > > >                      except queue.Full:
> > > > >                          pending.append(result)
> > > > >          finally:
> > > > > +            self.jobs.close()
> > > > >              self.results.close()
> > > > >              self.results.join_thread()
> > > > > 
> > > > > @@ -2134,13 +2138,13 @@ class CookerParser(object):
> > > > > 
> > > > >          self.bb_caches =
> > > > > bb.cache.MulticonfigCache(self.cfgbuilder, self.cfghash,
> > > > > cooker.caches_array)
> > > > >          self.fromcache = set()
> > > > > -        self.willparse = set()
> > > > > +        self.willparse = []
> > > > >          for mc in self.cooker.multiconfigs:
> > > > >              for filename in self.mcfilelist[mc]:
> > > > >                  appends =
> > > > > self.cooker.collections[mc].get_file_appends(filename)
> > > > >                  layername =
> > > > > self.cooker.collections[mc].calc_bbfile_priority(filename)[2]
> > > > >                  if not self.bb_caches[mc].cacheValid(filename,
> > > > > appends):
> > > > > -                    self.willparse.add((mc,
> > > > > self.bb_caches[mc], filename, appends, layername))
> > > > > +                    self.willparse.append((mc,
> > > > > self.bb_caches[mc], filename, appends, layername))
> > > > >                  else:
> > > > >                      self.fromcache.add((mc,
> > > > > self.bb_caches[mc], filename, appends, layername))
> > > > > 
> > > > > @@ -2159,22 +2163,25 @@ class CookerParser(object):
> > > > >      def start(self):
> > > > >          self.results = self.load_cached()
> > > > >          self.processes = []
> > > > > +
> > > > >          if self.toparse:
> > > > >              bb.event.fire(bb.event.ParseStarted(self.toparse),
> > > > > self.cfgdata)
> > > > > 
> > > > > +            self.toparse_queue =
> > > > > multiprocessing.Queue(len(self.willparse))
> > > > >              self.parser_quit = multiprocessing.Event()
> > > > >              self.result_queue = multiprocessing.Queue()
> > > > > 
> > > > > -            def chunkify(lst,n):
> > > > > -                return [lst[i::n] for i in range(n)]
> > > > > -            self.jobs = chunkify(list(self.willparse),
> > > > > self.num_processes)
> > > > > +            for jobid in range(len(self.willparse)):
> > > > > +                self.toparse_queue.put(jobid)
> > > > 
> > > > It's generally not a good idea to assume you can push all the
> > > > items on
> > > > the queue before there are any consumers; if the queue fills,
> > > > this
> > > > will deadlock. It would be better to push the items to the queue
> > > > after
> > > > creating the processes.
> > > 
> > > True. I did make sure the queue was large enough above however it
> > > would
> > > depend on the size of the job entries.
> > > 
> > > >  If you do that, I also don't see any reason to
> > > > push the indexes instead of the jobs directly; basically instead
> > > > of
> > > > "chunkifying" and pre-determining the assignment of jobs to
> > > > processes,
> > > > have them all pull from one queue of jobs.
> > > 
> > > I added this comment to the patch in master next:
> > > 
> > > # Have to pass in willparse at fork time so all parsing processes
> > > have the unpickleable data
> > > # then access it by index from the parse queue.
> > > 
> > > 
> > > As if you don't do that, it can't rebuild the data as the cache
> > > objects
> > > aren't pickleable :(
> > 
> > Ah, OK. We must be relying on the multiprocess being implemented with
> > fork() then to pass the data. That's reasonable and the indexes make
> > sense (but we should still push them after creating the process to
> > avoid the possibility of deadlock)
> 
> I changed it to queue after starting the parser processes however I
> think they exit early due to no work and it causes occasional build
> failures that way around so the code is going to need more extensive
> changes.

I found that issue in the exception handling and have sent a v2 of the
patch with that fixed.

Cheers,

Richard

Re: [bitbake-devel] [PATCH 1/2] cooker: Try and avoid parseing hangs

[Joshua Watt]

LGTM

Reviewed-by: Joshua Watt <JPEWhacker@gmail.com>

On Wed, Jul 2, 2025 at 4:24 PM Richard Purdie via
lists.openembedded.org
<richard.purdie=linuxfoundation.org@lists.openembedded.org> wrote:
>
> We sometimes see hangs in parsing during automated testing. It appears that
> SIGINT was sent to the underlying processes which see KeyboardInterrupt but
> they're stuck trying to write into the results pipe. The SIGINT was probably
> from some kind of parsing failure which doens't happen often, hence the hang
> being rare (in the incompatible license selftests from OE).
>
> This patch:
>   * sets a flag to indicate exit upon SIGINT so the exit is more graceful
>     and a defined exit path
>   * empties the results queue after we send the quit event
>   * empties the results queue after the SIGINT for good measure
>   * increases the 0.5s timeout to 2s since we now have some very slow to
>     parse recipes due to class extensions (ptests)
>
> This should hopefully make the parsing failure codepaths more robust.
>
> Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
> ---
>  lib/bb/cooker.py | 24 +++++++++++++++---------
>  1 file changed, 15 insertions(+), 9 deletions(-)
>
> diff --git a/lib/bb/cooker.py b/lib/bb/cooker.py
> index 1810bcc6049..91e3ee025ea 100644
> --- a/lib/bb/cooker.py
> +++ b/lib/bb/cooker.py
> @@ -2009,6 +2009,7 @@ class Parser(multiprocessing.Process):
>          self.queue_signals = False
>          self.signal_received = []
>          self.signal_threadlock = threading.Lock()
> +        self.exit = False
>
>      def catch_sig(self, signum, frame):
>          if self.queue_signals:
> @@ -2021,7 +2022,7 @@ class Parser(multiprocessing.Process):
>              signal.signal(signal.SIGTERM, signal.SIG_DFL)
>              os.kill(os.getpid(), signal.SIGTERM)
>          elif signum == signal.SIGINT:
> -            signal.default_int_handler(signum, frame)
> +            self.exit = True
>
>      def run(self):
>
> @@ -2059,7 +2060,7 @@ class Parser(multiprocessing.Process):
>          pending = []
>          havejobs = True
>          try:
> -            while havejobs or pending:
> +            while (havejobs or pending) and not self.exit:
>                  if self.quit.is_set():
>                      break
>
> @@ -2196,11 +2197,12 @@ class CookerParser(object):
>
>          # Cleanup the queue before call process.join(), otherwise there might be
>          # deadlocks.
> -        while True:
> -            try:
> -               self.result_queue.get(timeout=0.25)
> -            except queue.Empty:
> -                break
> +        def read_results():
> +            while True:
> +                try:
> +                   self.result_queue.get(timeout=0.25)
> +                except queue.Empty:
> +                    break
>
>          def sync_caches():
>              for c in self.bb_caches.values():
> @@ -2212,15 +2214,19 @@ class CookerParser(object):
>
>          self.parser_quit.set()
>
> +        read_results()
> +
>          for process in self.processes:
> -            process.join(0.5)
> +            process.join(2)
>
>          for process in self.processes:
>              if process.exitcode is None:
>                  os.kill(process.pid, signal.SIGINT)
>
> +        read_results()
> +
>          for process in self.processes:
> -            process.join(0.5)
> +            process.join(2)
>
>          for process in self.processes:
>              if process.exitcode is None:
>
> -=-=-=-=-=-=-=-=-=-=-=-
> Links: You receive all messages sent to this group.
> View/Reply Online (#17737): https://lists.openembedded.org/g/bitbake-devel/message/17737
> Mute This Topic: https://lists.openembedded.org/mt/113957082/3616693
> Group Owner: bitbake-devel+owner@lists.openembedded.org
> Unsubscribe: https://lists.openembedded.org/g/bitbake-devel/unsub [JPEWhacker@gmail.com]
> -=-=-=-=-=-=-=-=-=-=-=-
>