When readContent got Nothing from prepSendAnnex, it did not run its
callback, and the callback is what sends the DATA reply.
sendContent checks with contentSize that the object file is present, but
that doesn't really guarantee that prepSendAnnex won't return Nothing.
So, it was possible for a P2P protocol GET to not receive a response,
and appear to hang. When what it's really doing is waiting for the next
protocol command.
This seems most likely to happen when the annex is in direct mode, and the
file being requested has been modified. It could also happen in an indirect
mode repository if genInodeCache somehow failed. Perhaps due to a race
with a drop of the content file.
Fixed by making readContent behave the way its spec said it should,
and run the callback with L.empty in this case.
Note that, it's finee for readContent to send any amount of data
to the callback, including L.empty. sendBytes deals with that
by making sure it sends exactly the specified number of bytes,
aborting the protocol if it's too short. So, when L.empty is sent,
the protocol will end up aborting.
This work is supported by the NIH-funded NICEMAN (ReproNim TR&D3) project.
Instead remove enough data from it that this regression test tests what
it needs to.
Moving the database was the last thing that made the test suite
unstable, including sometimes hanging completely. It seems that, despite
closeDb being called before the move, sqlite was not quite done with it,
and somehow this causes other sqlite handles to become unstable. Not
good.
With this change, the test suite has successfully run 100+ times without
any issues.
Should be redundant, but test suite is ending up with
a lot of extra sqlite connections before unused keys database handles
get garbage collected.
While running the test suite, I often saw 2-4+ open fds to the same
repo's keys database. After this change, it seems to mostly have 1,
occasionally 2.
And that might explain some of the strange sqlite failures in the test suite.
Especially the failures of test_lock_v7_force, where the keys database
gets renamed to a new directory out from under sqlite.
I suspect this may be due to SQLITE_IOERR_SHORT_READ, but have not
verified.
I was able to reproduce it on Linux after running the test suite in a loop
for 1-3 hours until it failed.
The WAL mode entry change in 3963c5fcf5
may have hidden the problem I was seeing; I have not seen an ErrorIO
since then.
Cache high-resolution mtimes for improved detection of modified files in v7
(and direct mode).
Including on Windows.
With back-compat support so old low-res mtimes won't break anything, and
so the new information also won't break old versions of git-annex.
