Snapd is affected by several bugs related to content-layout interaction.
There's a very popular scheme where a snap has two interacting elements
that only work in one special edge case and break regularly:

- A content plug is used to receive content of another snap, typically
  under $SNAP/foo somewhere. Where foo exists and is not created with a
  mimic.
- A layout entry is used to "distribute" $SNAP/foo to /usr/share/foo or
  other locations.

This particular arrangement is used to share essential libraries from a
common "runtime" snap which is not a dedicated base. Several basic tests
explore the actual behavior in the following situations.

- The content is initially connected and a mount namespace is
  constructed. This is the only correct behavior show in this group of
  tests. This is known as test case 1.
- The content is initially connected and a mount namespace is
  constructed. The content is then disconnected. This shows several
  anomalies that are referenced in the code (ordering bug, propagation
  bug). This is known as test case 2.
- The content is initially connected and a mount namespace is
  constructed. The content is then disconnected and re-connected.
  This is known as test case 3.
- The content is initially disconnected and a mount namespace is
  constructed. The content is then connected.
  This is known as test case 4.
- The content interface is initially connected and a mount namespace is
  constructed. The content snap is then refreshed.
  This is known as test case 5.
- The content interface is initially connected and a mount namespace is
  constructed. The application snap is then refreshed.
  This is known as test case 6.

This is related to issues:
- https://warthogs.atlassian.net/browse/SNAPDENG-31644 (flip flop)
- https://warthogs.atlassian.net/browse/SNAPDENG-31645 (propagation)

A support makefile and a set of data files is provided to re-generate
the test data once snapd is improved.

Signed-off-by: Zygmunt Krynicki <[email protected]>

Signed-off-by: Zygmunt Krynicki <[email protected]>

Signed-off-by: Zygmunt Krynicki <[email protected]>

We recently realized that the order of saved mount profile, after
performing mount namespace update, was subtly incorrect. The order was
seemingly swapped after each operation. In reality, the order reflected
the order of performed changes, including the special change type of
"keep", which means that an entry is kept intact.

Fix the problem by storing "kept" changes in the original order, which
is back-to-front compared to the order of changes computed by the
algorithm. At the same time, keep track of mount changes in the order of
their execution.

This probably fixes a hell lot of edge cases and weird behaviors that
went undiagnosed for many years. Interestingly a test case had a TODO
that hinted at the problem, but it was not clear what the problem was at
the time. The TODO is now gone and the test case is now up-to-date.

Test data for cmd/snap-update-ns/testdata was re-generated.

Jira: https://warthogs.atlassian.net/browse/SNAPDENG-31644

Signed-off-by: Zygmunt Krynicki <[email protected]>

The new function takes the previously existing algorithm out of
executeMountProfileUpdate, as the original location makes it very tricky
to test due to the presence of all the system interactions through
change.Perform.

Add some simple tests that show how keep, mount and unmount changes are
preserved.

Signed-off-by: Zygmunt Krynicki <[email protected]>

Signed-off-by: Zygmunt Krynicki <[email protected]>