On 17/01/25 16:52, Jann Horn wrote: > On Fri, Jan 17, 2025 at 4:25 PM Valentin Schneider <vschneid@xxxxxxxxxx> wrote: >> On 14/01/25 19:16, Jann Horn wrote: >> > On Tue, Jan 14, 2025 at 6:51 PM Valentin Schneider <vschneid@xxxxxxxxxx> wrote: >> >> vunmap()'s issued from housekeeping CPUs are a relatively common source of >> >> interference for isolated NOHZ_FULL CPUs, as they are hit by the >> >> flush_tlb_kernel_range() IPIs. >> >> >> >> Given that CPUs executing in userspace do not access data in the vmalloc >> >> range, these IPIs could be deferred until their next kernel entry. >> >> >> >> Deferral vs early entry danger zone >> >> =================================== >> >> >> >> This requires a guarantee that nothing in the vmalloc range can be vunmap'd >> >> and then accessed in early entry code. >> > >> > In other words, it needs a guarantee that no vmalloc allocations that >> > have been created in the vmalloc region while the CPU was idle can >> > then be accessed during early entry, right? >> >> I'm not sure if that would be a problem (not an mm expert, please do >> correct me) - looking at vmap_pages_range(), flush_cache_vmap() isn't >> deferred anyway. > > flush_cache_vmap() is about stuff like flushing data caches on > architectures with virtually indexed caches; that doesn't do TLB > maintenance. When you look for its definition on x86 or arm64, you'll > see that they use the generic implementation which is simply an empty > inline function. > >> So after vmapping something, I wouldn't expect isolated CPUs to have >> invalid TLB entries for the newly vmapped page. >> >> However, upon vunmap'ing something, the TLB flush is deferred, and thus >> stale TLB entries can and will remain on isolated CPUs, up until they >> execute the deferred flush themselves (IOW for the entire duration of the >> "danger zone"). >> >> Does that make sense? > > The design idea wrt TLB flushes in the vmap code is that you don't do > TLB flushes when you unmap stuff or when you map stuff, because doing > TLB flushes across the entire system on every vmap/vunmap would be a > bit costly; instead you just do batched TLB flushes in between, in > __purge_vmap_area_lazy(). > > In other words, the basic idea is that you can keep calling vmap() and > vunmap() a bunch of times without ever doing TLB flushes until you run > out of virtual memory in the vmap region; then you do one big TLB > flush, and afterwards you can reuse the free virtual address space for > new allocations again. > > So if you "defer" that batched TLB flush for CPUs that are not > currently running in the kernel, I think the consequence is that those > CPUs may end up with incoherent TLB state after a reallocation of the > virtual address space. > > Actually, I think this would mean that your optimization is disallowed > at least on arm64 - I'm not sure about the exact wording, but arm64 > has a "break before make" rule that forbids conflicting writable > address translations or something like that. > > (I said "until you run out of virtual memory in the vmap region", but > that's not actually true - see the comment above lazy_max_pages() for > an explanation of the actual heuristic. You might be able to tune that > a bit if you'd be significantly happier with less frequent > interruptions, or something along those lines.) I've been thinking some more (this is your cue to grab a brown paper bag)... Experimentation (unmap the whole VMALLOC range upon return to userspace, see what explodes upon entry into the kernel) suggests that the early entry "danger zone" should only access the vmaped stack, which itself isn't an issue. That is obviously just a test on one system configuration, and the problem I'm facing is trying put in place /some/ form of instrumentation that would at the very least cause a warning for any future patch that would introduce a vmap'd access in early entry code. That, or a complete mitigation that prevents those accesses altogether. What if isolated CPUs unconditionally did a TLBi as late as possible in the stack right before returning to userspace? This would mean that upon re-entering the kernel, an isolated CPU's TLB wouldn't contain any kernel range translation - with the exception of whatever lies between the last-minute flush and the actual userspace entry, which should be feasible to vet? Then AFAICT there wouldn't be any work/flush to defer, the IPI could be entirely silenced if it targets an isolated CPU.
