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.
This means that Command.Move and Command.Get don't need to
manually set the stage, and is a lot cleaner conceptually.
Also, this makes Command.Sync.syncFile use the worker pool better.
In the scenario where it first downloads content and then uploads it to
some other remotes, it will start in TransferStage, then enter VerifyStage
and then go back to TransferStage for each transfer to the remotes.
Before, it entered CleanupStage after the download, and stayed in it for
the upload, so too many transfer jobs could run at the same time.
Note that, in Remote.Git, it uses runTransfer and also verifyKeyContent
inside onLocal. That has a Annex state for the remote, with no worker pool.
So the resulting calls to enteringStage won't block in there.
While Remote.Git.copyToRemote does do checksum verification, I
realized that should not use a verification slot in the WorkerPool
to do it. Because, it's reading back from eg, a removable disk to checksum.
That will contend with other writes to that disk. It's best to treat
that checksum verification as just part of the transer. So, removed the todo
item about that, as there's nothing needing to be done.
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.
get, move, copy, sync: When -J or annex.jobs has enabled concurrency,
checksum verification uses a separate job pool than is used for
downloads, to keep bandwidth saturated.
Not yet done for upload checksum verification, but that only affects
remotes on local disks.
I couldn't find a way to avoid the deadlock w/o rewriting it to clearly
not have one. I'm not quite sure what was the actual cause of the
deadlock.
This makes me unsure how I now know it clearly doesn't have a
deadlock. But, it was easy to reproduce before (just call it twice in a
row) and doesn't happen now.
Avoid a delay at startup when concurrency is enabled and there are
rsync or gcrypt special remotes, which was caused by git-annex
opening a ssh connection to the remote too early.
sshOptions makes a connection to the ssh server if one is not already open,
when concurrency is enabled. Avoid doing that at startup, when the remote
list is being built, but the remote may not be used at all.
Instead, rsync/gcrypt now runs sshOptions once per ssh connection to the
server. This should not be significant overhead since Remote.Git already
has the same overhead (as do Bup and Ddar).
Fixes the last wart in the StartMessage transition. A few commands
include other CommandStart actions that generate output, and
do not themselves need to display a start/end message.
Oh joyous day, this is probably git-annex's oldest implementation wart,
source of much unncessary bother.
Now that we have a StartMessage, showEndResult' can look at it to know
if it needs to display an end message or not.
This is also going to be faster, because it avoids an uncessary state
lookup for each file processed.
The hoped for optimisation of CommandStart with -J did not materialize.
In fact, not runnign CommandStart in parallel is slower than -J3.
So, CommandStart are still run in parallel.
(The actual bad performance I've been seeing with -J in my big repo
has to do with building the remoteList.)
But, this is still progress toward making -J faster, because it gets rid
of the onlyActionOn roadblock in the way of making CommandCleanup jobs
run separate from CommandPerform jobs.
Added OnlyActionOn constructor for ActionItem which fixes the
onlyActionOn breakage in the last commit.
Made CustomOutput include an ActionItem, so even things using it can
specify OnlyActionOn.
In Command.Move and Command.Sync, there were CommandStarts that used
includeCommandAction, so output messages, which is no longer allowed.
Fixed by using startingCustomOutput, but that's still not quite right,
since it prevents message display for the includeCommandAction run
inside it too.
The goal is to be able to run CommandStart in the main thread when -J is
used, rather than unncessarily passing it off to a worker thread, which
incurs overhead that is signficant when the CommandStart is going to
quickly decide to stop.
To do that, the message it displays needs to be displayed in the worker
thread, after the CommandStart has run.
Also, the change will mean that CommandStart will no longer necessarily
run with the same Annex state as CommandPerform. While its docs already
said it should avoid modifying Annex state, I audited all the
CommandStart code as part of the conversion. (Note that CommandSeek
already sometimes runs with a different Annex state, and that has not been
a source of any problems, so I am not too worried that this change will
lead to breakage going forward.)
The only modification of Annex state I found was it calling
allowMessages in some Commands that default to noMessages. Dealt with
that by adding a startCustomOutput and a startingUsualMessages.
This lets a command start with noMessages and then select the output it
wants for each CommandStart.
One bit of breakage: onlyActionOn has been removed from commands that used it.
The plan is that, since a StartMessage contains an ActionItem,
when a Key can be extracted from that, the parallel job runner can
run onlyActionOn' automatically. Then commands won't need to worry about
this detail. Future work.
Otherwise, this was a fairly straightforward process of making each
CommandStart compile again. Hopefully other behavior changes were mostly
avoided.
In a few cases, a command had a CommandStart that called a CommandPerform
that then called showStart multiple times. I have collapsed those
down to a single start action. The main command to perhaps suffer from it
is Command.Direct, which used to show a start for each file, and no
longer does.
Another minor behavior change is that some commands used showStart
before, but had an associated file and a Key available, so were changed
to ShowStart with an ActionItemAssociatedFile. That will not change the
normal output or behavior, but --json output will now include the key.
This should not break it for anyone using a real json parser.
