On 26.07.2012 12:33, Alex Rousskov wrote:
Hello,
The attached patch avoids memory corruption and segfaults during
ufs
cache store maintenance by removing being-deleted StoreEntry from ufs
cache replacement policies.
SMP Store changes added a SwapDir::disconnect() method that notified
entry's cache_dir that its entry is being destroyed. I added that
code
so that shared cache_dirs can update their index when idle
store_table
entries are destroyed on cleanup (completely avoiding idle entries
proved to be too difficult with store_table still around).
However, I did not realize that non-shared cache_dirs have the same
problem: Their StoreEntries may be deleted without cache_dirs knowing
that. UFS cache_dir now implement the SwapDir::disconnect() API to
stay
in sync. It is likely that the problem was introduced by SMP Store
changes, but I do not know whether there is a better fix for it.
I also added assertions to catch cases where a dirty StoreEntry or
MemObject (i.e., an entry or object that is still in some replacement
policy index) is being destroyed.
TODO: This patch breaks "make check". We need more STUBs somewhere!
Alex.
I've been thinking for some time that we need to implement a list of
being-deleted entries in each cache and add a unlinkd response handler
to erase them from that list after actual deletion which unregisters the
fileno for re-use.
Amos
