On Mon, May 04, 2015 at 06:28:25PM +0200, David Herrmann wrote: > I'm not sure how to write a benchmark for this. I mean, I can easily > craft one that causes the IDR to degenerate, but it requires us to > keep huge numbers of files open. > But this fact makes IDR rather suboptimal for cyclic allocations, so I > switched to idr_alloc() now. This guarantees tight/packed ID ranges > and user-space cannot degenerate the layers, anymore. That is, unless > you open more than 256 FDs on a device in parallel, we're fine with a > single IDR layer; always. This should work fine, right? That pretty much circumvents my only worry! If there is a client leak, the system will eventually keel under the load, and that we have a huge number of magics open is insignificant. As far as test coverage I would focus on (a) authenticating up to vfs-file-max fds (i.e. check that we hit the system limits without authmagic failing) (b) churn through a small number of clients for a few minutes, just basically checking for anomalous behaviour and that allocation times every minute or so remain constant. (c) just check that we can authenticate! always useful for patches that touch the authmagic system I was thinking of a few more, but they basically serve to show the holes in the authmagic scheme. -Chris -- Chris Wilson, Intel Open Source Technology Centre _______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx http://lists.freedesktop.org/mailman/listinfo/dri-devel
