Remove the compatibility code that caught these exceptions and made the
functions return 1 instead of throwing the exception. Make bbstoreaccounts
catch these exceptions and exit with status 1 instead.

Enable the relevant compiler warning as an error, if supported
(-Werror=undefined-bool-conversion).

http://www.gotw.ca/conv/002.htm
http://stackoverflow.com/questions/2165078/a-reference-can-not-be-null-or-it-can-be-null

(cherry picked from commit f2911aca)

Should increase reliability of tests running on Travis.

ReleaseLock() could fail if it can't contact the S3/SimpleDB server, but
destructors aren't allowed to throw exceptions, so catch this case and log
a warning instead (because we failed to release the lock that we were
holding).

(cherry picked from commit 6226c014dd2988a6a29cd21f6cc29fa0e7816607)

If syslog can't log to the Windows event log for some reason, instead of
just printing the log message to the console, include some explanatory
text about what went wrong.

(cherry picked from commit c492c90062d2e88db1016f13bd912629477a8a46)

This extends the -L command-line option to allow overriding by
filename/category=level as well as the existing support for overriding by
filename=level.

(cherry picked from commit 4421c53ab59adb494308f9c9f881a6a57bfd571c)

(cherry picked from commit feddbef8052208508a2e542a1a70c23138f476b8)

(cherry picked from commit cc14b3b4b9ec509233000c7d0a995580afd97f15)

(cherry picked from commit ca39b88d4a73b64de68400dfa29270f203fd0f14)

(cherry picked from commit c50aac67ee6ce6af9d63d5b96290585f2c5b7c21)

(cherry picked from commit 24bc49f30b9ce54d5024777f9c34d70a476d9663)

error: '&&' within '||' [-Werror,-Wlogical-op-parentheses]

I'm not convinced that it does anything on Windows without an RTT estimate,
which we can't get without administrator access:
https://msdn.microsoft.com/en-us/library/windows/desktop/bb485738(v=vs.85).aspx

(cherry picked from commit b646d22f415bbc8f5b5b065ed7462166591404f2)

(cherry picked from commit 9df71c89)

(cherry picked from commit 3a8ca0c97e7346912889df72a1a87b80b430fe7a)

This extends test_server_commands() to the S3 backend implementation, using
test specialisation.

This tests that HTTPRequest::SendHeaders includes the "Expect: 100-continue"
header only when expected. A separate test calls HTTPRequest::SendWithStream
which sends these headers, and therefore checks that both sides correctly
handle such exchanges.

Needed to clear the cache in backupstore tests specialised for S3.

This was designed to stop untested code from running. Now we are ready to use
this code, with tests soon to come.

This enables the S3 bucket name to be included in the virtual hostname, for
example bucketname.s3.amazonaws.com, when connecting to a different hostname
such as localhost. This makes bucket names usable in tests with
BackupAccountControl.

"If the object is being created as part of a new expression, and an exception
is thrown, the object’s memory is deallocated by calling the appropriate
deallocation function. If the object is being created with a placement new
operator, the corresponding placement delete operator is called—that is, the
delete function that takes the same additional parameters as the placement new
operator. If no matching placement delete is found, no deallocation takes
place."

So to avoid memory leaks, we need to implement placement delete operators that
correspond to our placement new, which we use for leak detection (ironically)
in debug builds.

The class name was a problem, because the method that we want to call on it is
also called GetLine, which looks too much like a constructor in C++, so I
renamed it to LineBuffer.  I haven't yet renamed the child classes.

Also, it was impossible to distinguish between the different causes of
GetLine() returning false (timeout and signal) and EOF could only be checked by
calling a different function. Now all of these causes throw different
exceptions.

This really fixes the problem with attaching a debugger to a process waiting in
GetLine, which would previously cause the read() to return immediately with
EINTR, which we never retried because it looked like a timeout.

This allows protocol files to be regenerated automatically if
makeprotocol.pl.in changes, and similarly exception files and
makeexception.pl.in.

Previously it would truncate the expected refcount list at the first zero entry,
instead of just removing zero entries from the end.

Also move some test code to test_server_commands which should have been there,
since they were causing issues with the check for the refcount database in
test_server_housekeeping. It's hard to disentangle these two changes.

Don't allow an uncaught exception to kill the whole server. This makes it
consistent with multi-process servers, where death of a child worker does
not stop the main server process, so it continues to serve requests.

This makes the HTTPServer robust to exceptions (such as the child
disconnecting while the server is streaming a response) on Windows.

(cherry picked from commit 4a515b9ce8e616b357dc07bee27aa7b3692c3602)

This allows us to wrap an existing stream (not necessarily a file, e.g. a
streaming HTTPRequest) in a VerifyStream and pass that stream to a
S3PutFileCompleteTransaction, which isn't just going to CopyStreamTo()
somewhere, but do something more complex to it.

These should not have been part of the API, and are not needed since mrClient
has the same methods.

Should fix random test failures due to housekeeping process realising that the
master has asked it to terminate, and logging that it has, after the master has
already died and the test finished, confusing the test runner with its extra
output. Should also help make the occasionally-reported housekeeping crashes
more obvious and easier to debug.

Call ReleaseLock in the child destructor (not the parent) to allow the refcount
database to clean itself up (normally involving calling
BackupFileSystem::SaveRefCountDatabase) before the BackupFileSystem is too
destroyed to be able to fulfil the request.