The rollback manager was using a saved MountTable rather than the
current table, causing it to attempt to rollback unmounted mounts, and
never rollback new mounts.
In fixing this, it became clear that bad things could happen to the
mount table...the table itself could be locked, but the table pointer
(which is what the rollback manager needs) could be modified at any time
without locking. This commit therefore also returns locking to a mutex
outside the table instead of inside, and plumbs RLock/RUnlock through to
the various places that are reading the table but not holding a write
lock.
Both unit tests and race detection pass.
Fixes#771
In order to implement this efficiently, I have introduced the concept of
"singleton" backends -- currently, 'sys' and 'cubbyhole'. There isn't
much reason to allow sys to be mounted at multiple places, and there
isn't much reason you'd need multiple per-token storage areas. By
restricting it to just one, I can store that particular mount instead of
iterating through them in order to call the appropriate revoke function.
Additionally, because revocation on the backend needs to be triggered by
the token store, the token store's salt is kept in the router and
client tokens going to the cubbyhole backend are double-salted by the
router. This allows the token store to drive when revocation happens
using its salted tokens.
up-to-date information. This allows remount to be implemented with the
same source and dest, allowing mount options to be changed on the fly.
If/when Vault gains the ability to HUP its configuration, this should
just work for the global values as well.
Need specific unit tests for this functionality.
specify more concrete error cases to make their way back up the stack.
Over time there is probably a cleaner way of doing this, but that's
looking like a more massive rewrite and this solves some issues in
the meantime.
Use a CodedError to return a more concrete HTTP return code for
operations you want to do so. Returning a regular error leaves
the existing behavior in place.
/cc @armon - So I know the conversation we had related to this about
auth, but I think we still need to export these and do auth only at the
external API layer. If you're writing to the internal API, then all bets
are off.
The reason is simply that if you have access to the code, you can
already work around it anyways (you can disable auth or w/e), so a
compromised Vault source/binary is already a failure, and that is the
only thing that our previous unexported methods were protecting against.
If you write an external tool to access a Vault, it still needs to be
unsealed so _that_ is the primary security mechanism from an API
perspective. Once it is unsealed then the core API has full access to
the Vault, and identity/auth is only done at the external API layer, not
at the internal API layer.
The benefits of this approach is that it lets us still treat the "sys"
mount specially but at least have sys adopt helper/backend and use that
machinery and it can still be the only backend which actually has a
reference to *vault.Core to do core things (a key difference). So, an
AWS backend still will never be able to muck with things it can't, but
we're explicitly giving Sys (via struct initialization in Go itself)
a reference to *vault.Core.
