On Mon, Oct 21, 2024 at 3:31 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > > On Fri, Oct 18, 2024 at 11:22:00AM -0700, Andrii Nakryiko wrote: > > On Fri, Oct 18, 2024 at 1:26 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > > > > > > On Mon, Oct 07, 2024 at 05:25:55PM -0700, Andrii Nakryiko wrote: > > > > Currently put_uprobe() might trigger mutex_lock()/mutex_unlock(), which > > > > makes it unsuitable to be called from more restricted context like softirq. > > > > > > This is delayed_uprobe_lock, right? > > > > Not just delated_uprobe_lock, there is also uprobes_treelock (I forgot > > to update the commit message to mention that). Oleg had concerns (see > > [0]) with that being taken from the timer thread, so I just moved all > > of the locking into deferred work callback. > > > > [0] https://lore.kernel.org/linux-trace-kernel/20240915144910.GA27726@xxxxxxxxxx/ > > Right, but at least that's not a sleeping lock. He's right about it > needing to become a softirq-safe lock though. And yeah, unfortunate > that. > > > > So can't we do something like so instead? > > > > I'll need to look at this more thoroughly (and hopefully Oleg will get > > a chance as well), dropping lock from delayed_ref_ctr_inc() is a bit > > scary, but might be ok. > > So I figured that update_ref_ctr() is already doing the > __update_ref_ctr() thing without holding the lock, so that lock really > is only there to manage the list. > > And that list is super offensive... That really wants to be a per-mm > rb-tree or somesuch. Probably hard to justify to add that to mm_struct, tbh, given that uprobe+refcnt case (which is USDT with semaphore) isn't all that frequent, and even then it will be active on a very small subset of processes in the system, most probably. But, even if (see below), probably should be a separate change. > > AFAICT the only reason it is a mutex, is because doing unbouded list > iteration under a spinlock is a really bad idea. > > > But generally speaking, what's your concern with doing deferred work > > in put_uprobe()? It's not a hot path by any means, worst case we'll > > have maybe thousands of uprobes attached/detached. > > Mostly I got offended by the level of crap in that code, and working > around crap instead of fixing crap just ain't right. > Ok, so where are we at? Do you insist on the delayed_ref_ctr_inc() rework, switching uprobe_treelock to be softirq-safe and leaving put_uprobe() mostly as is? Or is it ok, to do a quick deferred work change for put_uprobe() to unblock uretprobe+SRCU and land it sooner? What if we split this work into two independent patch sets, go with deferred work for uretprobe + SRCU, and then work with Oleg and you on simplifying and improving delayed_uprobe_lock-related stuff? After all, neither deferred work nor delayed_ref_ctr_inc() change has much practical bearing on real-world performance. WDYT?
