Fix upload thread creation timing issue (s3ql#240)

If the block cache is initialised (i.e. associated threads created, with
references to trio channels) before daemonisation, those threads end up
running in a separate process to the main trio code and thus
sending/receiving on two unrelated/"disconnected" channels.

Fix this by moving initialisation back after daemonisation, as was the
case before the commit in fbd3116

Changes.txt

@@ -2,8 +2,11 @@ UNRELEASED CHANGES
 
   * The tcp-timeout backend option of the B2 Backend works now.
 
+  * mount.s3ql no longer crashes with "No Upload Threads available" when not running in
+    foreground.
 
-2021-01-03, S3Ql 3.7.0
+
+2021-01-03, S3QL 3.7.0
 
   * S3QL now requires Python 3.7 or newer.
 

src/s3ql/mount.py

@@ -204,7 +204,6 @@ async def main_async(options, stdout_log_handler):
       async with AsyncExitStack() as cm:
         block_cache = BlockCache(backend_pool, db, cachepath + '-cache',
                                  options.cachesize * 1024, options.max_cache_entries)
-        block_cache.init(options.threads)
         cm.push_async_callback(block_cache.destroy, options.keep_cache)
 
         operations = fs.Operations(block_cache, db, max_obj_size=param['max_obj_size'],
@@ -240,6 +239,8 @@ def unmount():
 
         mark_metadata_dirty(backend, cachepath, param)
 
+        block_cache.init(options.threads)
+
         nursery.start_soon(metadata_upload_task.run, name='metadata-upload-task')
         cm.callback(metadata_upload_task.stop)