Bugfix: When -J was enabled, getting files leaked a ever-growing number of
git cat-file processes.
(Since commit dd39e9e255)
The leak happened when mergeState called stopNonConcurrentSafeCoProcesses.
While stopNonConcurrentSafeCoProcesses usually manages to stop everything,
there was a race condition where cat-file processes were leaked. Because
catFileStop modifies Annex.catfilehandles in a non-concurrency safe way,
and could clobber modifications made in between. Which should have been ok,
since originally catFileStop was only used at shutdown.
Note the comment on catFileStop saying it should only be used when nothing
else is using the handles. It would be possible to make catFileStop
race-safe, but it should just not be used in a situation where a race is
possible. So I didn't bother.
Instead, the fix is just not to stop any processes in mergeState. Because
in order for mergeState to be called, dupState must have been run, and it
enables concurrency mode, stops any non-concurrent processes, and so all
processes that are running are concurrency safea. So there is no need to
stop them when merging state. Indeed, stopping them would be extra work,
even if there was not this bug.
Sponsored-by: Dartmouth College's Datalad project
When non-concurrent git coprocesses have been started, setConcurrency
used to not stop them, and so could leak processes when enabling
concurrency, eg when forkState is called.
I do not think that ever actually happened, given where setConcurrency
is called. And it probably would only leak one of each process, since it
never downgrades from concurrent to non-concurrent.
Clear visible progress bar first.
Removed showSideActionAfter because it can't be used in reconcileStaged
(import loop). Instead, it counts the number of files it
processes and displays it after it's seen a sufficient to know it's
taking a while.
Sponsored-by: Dartmouth College's Datalad project
Avoids users thinking this scan is a big deal, when it's not in the
majority of repos.
showSideActionAfter has some ugly caveats, since it has to display in
the background of another action. I could not see a better way to do it
and it works fine in this particular case. It also doesn't really belong
in Annex.Concurrent, but cannot go in Messages due to an import loop.
Sponsored-by: Dartmouth College's Datalad project
Values in AnnexRead can be read more efficiently, without MVar overhead.
Only a few things have been moved into there, and the performance
increase so far is not likely to be noticable.
This is groundwork for putting more stuff in there, particularly a value
that indicates if debugging is enabled.
The obvious next step is to change option parsing to not run in the
Annex monad to set values in AnnexState, and instead return a pure value
that gets stored in AnnexRead.
When annex.stalldetection is not enabled, and a likely stall is detected,
display a suggestion to enable it.
Note that the progress meter display is not taken down when displaying
the message, so it will display like this:
0% 8 B 0 B/s
Transfer seems to have stalled. To handle stalling transfers, configure annex.stalldetection
0% 10 B 0 B/s
Although of course if it's really stalled, it will never update
again after the message. Taking down the progress meter and starting
a new one doesn't seem too necessary given how unusual this is,
also this does help show the state it was at when it stalled.
Use of uninterruptibleCancel here is ok, the thread it's canceling
only does STM transactions and sleeps. The annex thread that gets
forked off is separate to avoid it being canceled, so that it
can be joined back at the end.
A module cycle required moving from dupState the precaching of the
remote list. Doing it at startConcurrency should cover all the cases
where the remote list is used in concurrent actions.
This commit was sponsored by Kevin Mueller on Patreon.
Done on unix, could not implement it on windows quite.
The signal library gets part of the way needed for windows.
But I had to open https://github.com/pmlodawski/signal/issues/1 because
it lacks raiseSignal.
Also, I don't know what the equivilant of getProcessGroupIDOf is on
windows. And System.Process does not provide a way to send any signal to
a process group except for SIGINT.
This commit was sponsored by Boyd Stephen Smith Jr. on Patreon.
Doing this at shutdown is not very important at all, but I do like to
make sure that when git-annex allocates a resource, it later cleans it
up.
More importantly, stopCoProcesses is used in eg, Remote.Git in a
situation where it needs to stop long-running processes like these.
Limited to min of -JN or number of CPU cores, because it will often be
CPU bound, once it's read the gitignore file for a directory.
In some situations it's more disk bound, but in any case it's unlikely
to be the main bottleneck that -J is used to avoid. Eg, when dropping,
this is used for numcopies checks, but the main bottleneck will be
accessing the remotes to verify presence. So the user might decide to
-J32 that, but having 32 check-attr processes would just waste however
many filehandles they open, and probably worsen their performance due to
CPU contention.
Note that, I first tried just letting up to the -JN be started. However,
even when it's no bottleneck at all, that still results in all of them
being started. Why? Well, all the worker threads start up nearly
simulantaneously, so there's a thundering herd..
Avoid running a large number of git cat-file child processes when run with
a large -J value.
This implementation takes care to avoid adding any overhead to git-annex
when run without -J. When run with -J, there is a small bit of added
overhead, to manipulate the resource pool. That optimisation added a
fair bit of complexity.
This does mean that RemoteDaemon.Transport.Tor's call runs it, otherwise
no change, but this is groundwork for doing more such expensive actions
in dupState.
See the comment for a trace of the deadlock.
Added a new StartStage. New worker threads begin in the StartStage.
Once a thread is ready to do work, it moves away from the StartStage,
and no thread will ever transition back to it.
A thread that blocks waiting on another thread that is processing
the same key will block while in the StartStage. That other thread
will never switch back to the StartStage, and so the deadlock is avoided.
Avoid an unncessary STM transaction. This will happen when the worker
pool is not completely full of the new stage, which is the common case.
In the uncommon case, this adds only a tiny bit of overhead for the
extra traversal of the worker pool. And the thread is going to block
for some time anyway.
When all worker threads are running and enteringStage is called,
it waits for an idle slot. If all off the other threads then call it in
turn, a deadlock occurrs.
This is the same problem I didn't actually fix in
5a9842d7ed.
Fixed by doing two separate STM transactions, the first replaces its
active thread with an idle thread, and the second waits for another idle
thread. That guarantees there will eventually be an idle thread to find.
The changes to WorkerPool were necessary because it can't add an idle
thread containing the Annex state and go on to run an action using that
same state, so I had to remove the Annex state from IdleWorker.
Rather than limiting it to PerformStage and CleanupStage, this opens it
up so any number of stages can be added as needed by commands.
Each concurrent command has a set of stages that it uses, and only
transitions between those can block waiting for a free slot in the
worker pool. Calling enteringStage for some other stage does not block,
and has very little overhead.
Note that while before the Annex state was duplicated on the first call
to commandAction, this now happens earlier, in startConcurrency.
That means that seek stage actions should that use startConcurrency
and then modify Annex state won't modify the state of worker threads
they then start. I audited all of them, and only Command.Seek
did so; prepMerge changes the working directory and so has to come
before startConcurrency.
Also, the remote list is built before duplicating the state, which means
that it gets built earlier now than it used to. This would only have an
effect of making commands that end up not needing to perform any actions
unncessary build the remote list (only when they're run with concurrency
enable), but that's a minor overhead compared to commands seeking
through the work tree and determining they don't need to do anything.
When running multiple concurrent actions, the cleanup phase is run in a
separate queue than the main action queue. This can make some commands
faster, because less time is spent on bookkeeping in between each file
transfer.
But as far as I can see, nothing will be sped up much by this yet, because
all the existing cleanup actions are very light-weight. This is just groundwork
for deferring checksum verification to cleanup time.
This change does mean that if the user expects -J2 will mean that they see no
more than 2 jobs running at a time, they may be surprised to see 4 in some
cases (if the cleanup actions are slow enough to notice).
It might also make sense to enable background cleanup without the -J,
for at least one cleanup action. Indeed, that's the behavior that -J1
has now. At some point in the future, it make make sense to make the
behavior with no -J the same as -J1. The only reason it's not currently
is that git-annex can build w/o concurrent-output, and also any bugs
in concurrent-output (such as perhaps misbehaving on non-VT100 compatible
terminals) are avoided by default by only using it when -J is used.
This does not change the overall license of the git-annex program, which
was already AGPL due to a number of sources files being AGPL already.
Legally speaking, I'm adding a new license under which these files are
now available; I already released their current contents under the GPL
license. Now they're dual licensed GPL and AGPL. However, I intend
for all my future changes to these files to only be released under the
AGPL license, and I won't be tracking the dual licensing status, so I'm
simply changing the license statement to say it's AGPL.
(In some cases, others wrote parts of the code of a file and released it
under the GPL; but in all cases I have contributed a significant portion
of the code in each file and it's that code that is getting the AGPL
license; the GPL license of other contributors allows combining with
AGPL code.)
Probably not noticed until now because the queue is large enough that two
threads each filling theirs at the same time and flushing is unlikely to
happen.
Also made explicit that each worker thread gets its own queue.
I think that was the case before, but if something was put in the queue
before worker threads were forked off, they could have each inherited the
same queue.
Could have gone with a single shared queue, but per-worker queues is more
efficient, because a worker can add lots of stuff to its own queue without
any locking.
This commit was sponsored by Ole-Morten Duesund on Patreon.
Fix process and file descriptor leak that was exposed when git-annex was
built with ghc 8.2.1. Apparently ghc has changed its behavior of GC
of open file handles that are pipes to running processes. That
broke git-annex test on OSX due to running out of FDs.
Audited for all uses of Annex.new and made stopCoProcesses be called
once it's done with the state. Fixed several places that might have
leaked in other situations than running the test suite.
This commit was sponsored by Ewen McNeill.
