Oops. Stacked bugs: first, the desired user handle was not properly
being passed from the call() entry point to the database operations;
then on top of that, the client-side cache management was still
looking in the local user's cache for the data, so a request to read
a different user's settings would return the local user's instead if
that key was already known to the local user's cache.
Reads and writes of a different user's settings are now uncached,
so they're relatively much slower. They're rare, however, so this
is not something to worry about unless we encounter a real world
case where it's a significant factor.
This CL also adds a bit of cross-user settings read/write testing
to the existing provider suite. These new tests caught both the
known wrong-user-write bug and discovered the client-side cache
bug, so yay.
Finally, the existing wholesale mutual-exclusion approach would
deadlock in certain circumstances due to the fact that the
settings database creation code might have to call out to the
Package Manager while populating the bookmark/shortcut table,
and the Package Manager would then call back into the settings
provider in the course of handling that request. The synchronization
regime has been significantly tightened up now: now the database
code [which is known to deal with concurrency itself] is allowed
to cope with multiple parallel openers of the same db; this
allows the settings provider to avoid calling out to other parts
of the system even implicitly while its internal lock is held.
Change-Id: Ib77d445b4a2ec658cc5c210830f6977c981f87ed
78d2a66ac1