On Tue, Jul 09, 2024 at 04:29:43PM +0200, Peter Zijlstra wrote: > On Tue, Jul 09, 2024 at 07:11:23AM -0700, Paul E. McKenney wrote: > > On Tue, Jul 09, 2024 at 11:01:53AM +0200, Peter Zijlstra wrote: > > > On Mon, Jul 08, 2024 at 05:25:14PM -0700, Andrii Nakryiko wrote: > > > > > > > Quick profiling for the 8-threaded benchmark shows that we spend >20% > > > > in mmap_read_lock/mmap_read_unlock in find_active_uprobe. I think > > > > that's what would prevent uprobes from scaling linearly. If you have > > > > some good ideas on how to get rid of that, I think it would be > > > > extremely beneficial. > > > > > > That's find_vma() and friends. I started RCU-ifying that a *long* time > > > ago when I started the speculative page fault patches. I sorta lost > > > track of that effort, Willy where are we with that? > > > > > > Specifically, how feasible would it be to get a simple RCU based > > > find_vma() version sorted these days? > > > > Liam's and Willy's Maple Tree work, combined with Suren's per-VMA locking > > combined with some of Vlastimil's slab work is pushing in that direction. > > I believe that things are getting pretty close. > > So I fundamentally do not believe in per-VMA locking. Specifically for > this case that would be trading one hot line for another. I tried > telling people that, but it doesn't seem to stick :/ > > Per VMA refcounts or per VMA locks are a complete fail IMO. Not even to allow concurrent updates of the address space by different threads of a process? For me, per-VMA locking's need to RCU-protect the VMA is a good step towards permitting RCU-protected scans of the Maple Tree, which then gets lockless lookup. > I suppose I should go dig out the latest versions of those patches to > see where they're at :/ It would not be a bad thing to get another set of eyes on it. Thanx, Paul
