Re: [Xen-devel] [PATCH v2 16/18] libxl: introduce libxl_userdata_unlink

[Wei Liu <wei.liu2@xxxxxxxxxx>]

For the record, here is summary of the face-2-face discussion Ian and I
had.

The lock proposed in patch 2 should be renamed to "domain-data-lock" to
avoid making the false impression that it's related to libxl-json data.

For any specific application, it should hold a lock to protect its own
userdata if it cares, to protect against r-m-w by other threads.

With the introduction of libxl-json data, libxl becomes an application
of libxl itself, in the sense that libxl now does r-m-w to libxl-json
data internally. Hence it needs a lock (LockA).

The said lock, proposed in patch 2, is used to lock all types of user
data related to a specified domain. It protects userdata against domain
destruction (LockB).

I reused the lock in patch 2 as the application data lock for libxl-json
data.  The lock serves two purposes:
  1. to protect against r-m-w to libxl-json data
  2. to protect against domain destruction

This double-duty lock is what confuses most. In theory LockA and LockB
can be different locks but there isn't much benefit in doing so in terms
of protecting libxl-json data.

In general, locking hierarchy should be
  application userdata lock (if any)
  ---- (inside libxl)
  libxl domain data lock
  ...
  libxl domain data unlock
  ----
  application userdata unlock

In the case of protecting libxl-json data, the application userdata lock
is the same one as libxl domain data lock, so the locking hierarchy is
preserved.

Wei.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel