When we migrated away from all the locks, there was a regression that
was not caught immediately. The tsc::get_asset() would attempt to modify
the snapshot, but the problem was that the snapshot was a .clone() of
the inner language server's assets, which meant that modifications to
that where lost. When we then attempted to do a hover on those assets,
the inner language servers assets didn't have the retrieved asset, and
therefore would throw an error.
Commit 2828690fc ("fix(lsp): fix deadlocks, use one big mutex") from
last month introduced a regression in asset cache lookups where results
of lazy caching were lost due to operating on a copy of the asset cache.
This commit fixes that by copying the asset from the copy to the real
cache.
The mutex was used to hide the fact that the Sources object mutates
itself when it's queried. Be honest about that and mark everything that
directly or indirectly mutates it as `mut`.
This is a follow-up to commit 2828690fc7
from last month ("fix(lsp): fix deadlocks, use one big mutex (#9271)")
The LSP code had numerous places where competing threads could take out
out locks in different orders, making it very prone to deadlocks.
This commit sidesteps the entire issue by switching to a single lock.
The above is a little white lie: the Sources struct still uses a mutex
internally to avoid having to boil the ocean (because being honest about
what it does involves changing all its methods to `&mut self` but that
ripples out extensively...) I'll save that one for another day.
