On Wed, Apr 10, 2024 at 04:43:51PM -0400, Peter Xu wrote: > On Wed, Apr 10, 2024 at 09:26:45PM +0100, Matthew Wilcox wrote: > > On Wed, Apr 10, 2024 at 01:06:21PM -0400, Peter Xu wrote: > > > anon_vma is a tricky object in the context of per-vma lock, because it's > > > racy to modify it in that context and mmap lock is needed if it's not > > > stable yet. > > > > I object to this commit message. First, it's not a "sanity check". It's > > a check to see if we already have an anon VMA. Second, it's not "racy > > to modify it" at all. The problem is that we need to look at other > > VMAs, for which we do not hold the lock. > > For that "do not hold locks" part, isn't that "racy"? No. > > > - We may always use mmap lock for the initial READs on a private file > > > mappings, while before this patch it _can_ (only when no WRITE ever > > > happened... but it doesn't make much sense for a MAP_PRIVATE..) do the > > > read fault with per-vma lock. > > > > But that's a super common path! Look at 'cat /proc/self/maps'. All > > your program text (including libraries) is mapped PRIVATE, and never > > written to (except by ptrace, I guess). > > > > NAK this patch. > > We're talking about any vma that will first benefit from a per-vma lock > here, right? > > I think it should be only relevant to some major VMA or bunch of VMAs that > an userspace maps explicitly, then iiuc the goal is we want to reduce the > cache bouncing of the lock when it used to be per-mm, by replacing it with > a finer lock. It doesn't sound right that these libraries even fall into > this category as they should just get loaded soon enough when the program > starts. > > IOW, my understanding is that per-vma lock doesn't benefit from such normal > vmas or simple programs that much; we take either per-vma read lock, or > mmap read lock, and I would expect similar performance when such cache > bouncing isn't heavy. > > I can do some tests later today or tomorrow. Any suggestion you have on > amplifying such effect that you have concern with? 8 socket NUMA system, 800MB text segment, 10,000 threads. No, I'm not joking, that's a real customer workload.
