Re: [PATCH 7/9] rhashtable: Per bucket locks & deferred expansion/shrinking

[Pablo Neira Ayuso <pablo@netfilter.org>]

On Fri, Jan 16, 2015 at 07:35:57PM +0000, Patrick McHardy wrote:
> On 16.01, Thomas Graf wrote:
> > On 01/16/15 at 06:36pm, Patrick McHardy wrote:
> > > On 16.01, Thomas Graf wrote:
> > > > On 01/16/15 at 04:43pm, David Laight wrote:
> > > > > The walker is unlikely to see items that get inserted early in the hash
> > > > > table even without a resize.
> > > > 
> > > > I don't follow, you have to explain this statement.
> > > > 
> > > > Walkers which don't want to see duplicates or miss entries should
> > > > just take the mutex.
> > > 
> > > Well, we do have a problem with interrupted dumps. As you know once
> > > the netlink message buffer is full, we return to userspace and
> > > continue dumping during the next read. Expanding obviously changes
> > > the order since we rehash from bucket N to N and 2N, so this will
> > > indeed cause duplicate (doesn't matter) and missed entries.
> > 
> > Right,but that's a Netlink dump issue and not specific to rhashtable.
> 
> Well, rhashtable (or generally resizing) will make it a lot worse.
> Usually we at worst miss entries which were added during the dump,
> which is made up by the notifications.
> 
> With resizing we might miss anything, its completely undeterministic.
> 
> > Putting the sequence number check in place should be sufficient
> > for sets, right?
> 
> I don't see how. The problem is that the ordering of the hash changes
> and it will skip different entries than those that have already been
> dumped.

I think the generation counter should catch up this sort of problems.
The resizing is triggered by a new/deletion element, which bumps it
once the transaction is handled.