On Thu, Jan 10, 2019 at 1:10 PM Khalid Aziz <khalid.aziz@xxxxxxxxxx> wrote: > I implemented a solution to reduce performance penalty and > that has had large impact. When XPFO code flushes stale TLB entries, > it does so for all CPUs on the system which may include CPUs that > may not have any matching TLB entries or may never be scheduled to > run the userspace task causing TLB flush. Problem is made worse by > the fact that if number of entries being flushed exceeds > tlb_single_page_flush_ceiling, it results in a full TLB flush on > every CPU. A rogue process can launch a ret2dir attack only from a > CPU that has dual mapping for its pages in physmap in its TLB. We > can hence defer TLB flush on a CPU until a process that would have > caused a TLB flush is scheduled on that CPU. I have added a cpumask > to task_struct which is then used to post pending TLB flush on CPUs > other than the one a process is running on. This cpumask is checked > when a process migrates to a new CPU and TLB is flushed at that > time. I measured system time for parallel make with unmodified 4.20 > kernel, 4.20 with XPFO patches before this optimization and then > again after applying this optimization. Here are the results: > > Hardware: 96-core Intel Xeon Platinum 8160 CPU @ 2.10GHz, 768 GB RAM > make -j60 all > > 4.20 915.183s > 4.20+XPFO 24129.354s 26.366x > 4.20+XPFO+Deferred flush 1216.987s 1.330xx > > > Hardware: 4-core Intel Core i5-3550 CPU @ 3.30GHz, 8G RAM > make -j4 all > > 4.20 607.671s > 4.20+XPFO 1588.646s 2.614x > 4.20+XPFO+Deferred flush 794.473s 1.307xx Well that's an impressive improvement! Nice work. :) (Are the cpumask improvements possible to be extended to other TLB flushing needs? i.e. could there be other performance gains with that code even for a non-XPFO system?) > 30+% overhead is still very high and there is room for improvement. > Dave Hansen had suggested batch updating TLB entries and Tycho had > created an initial implementation but I have not been able to get > that to work correctly. I am still working on it and I suspect we > will see a noticeable improvement in performance with that. In the > code I added, I post a pending full TLB flush to all other CPUs even > when number of TLB entries being flushed on current CPU does not > exceed tlb_single_page_flush_ceiling. There has to be a better way > to do this. I just haven't found an efficient way to implemented > delayed limited TLB flush on other CPUs. > > I am not entirely sure if switch_mm_irqs_off() is indeed the right > place to perform the pending TLB flush for a CPU. Any feedback on > that will be very helpful. Delaying full TLB flushes on other CPUs > seems to help tremendously, so if there is a better way to implement > the same thing than what I have done in patch 16, I am open to > ideas. Dave, Andy, Ingo, Thomas, does anyone have time to look this over? > Performance with this patch set is good enough to use these as > starting point for further refinement before we merge it into main > kernel, hence RFC. > > Since not flushing stale TLB entries creates a false sense of > security, I would recommend making TLB flush mandatory and eliminate > the "xpfotlbflush" kernel parameter (patch "mm, x86: omit TLB > flushing by default for XPFO page table modifications"). At this point, yes, that does seem to make sense. > What remains to be done beyond this patch series: > > 1. Performance improvements > 2. Remove xpfotlbflush parameter > 3. Re-evaluate the patch "arm64/mm: Add support for XPFO to swiotlb" > from Juerg. I dropped it for now since swiotlb code for ARM has > changed a lot in 4.20. > 4. Extend the patch "xpfo, mm: Defer TLB flushes for non-current > CPUs" to other architectures besides x86. This seems like a good plan. I've put this series in one of my tree so that 0day will find it and grind tests... https://git.kernel.org/pub/scm/linux/kernel/git/kees/linux.git/log/?h=kspp/xpfo/v7 Thanks! -- Kees Cook
