On Thu, May 9, 2019 at 4:30 PM Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > > On Thu, May 09, 2019 at 01:49:25PM +0200, Daniel Borkmann wrote: > > On 05/09/2019 12:58 PM, Eric Dumazet wrote: > > > On Thu, May 9, 2019 at 3:52 AM Eric Dumazet <edumazet@xxxxxxxxxx> wrote: > > >> On Wed, May 8, 2019 at 9:47 PM Alexei Starovoitov > > >> <alexei.starovoitov@xxxxxxxxx> wrote: > > >>> On Wed, May 08, 2019 at 04:17:29PM -0700, Eric Dumazet wrote: > > >>>> On Wed, May 8, 2019 at 4:09 PM Alexei Starovoitov > > >>>> <alexei.starovoitov@xxxxxxxxx> wrote: > > >>>>> On Wed, May 08, 2019 at 02:21:52PM -0700, Eric Dumazet wrote: > > >>>>>> Hi Alexei and Daniel > > >>>>>> > > >>>>>> I have a question about seccomp. > > >>>>>> > > >>>>>> It seems that after this patch, seccomp no longer needs a helper > > >>>>>> (seccomp_bpf_load()) > > >>>>>> > > >>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=bd4cf0ed331a275e9bf5a49e6d0fd55dffc551b8 > > >>>>>> > > >>>>>> Are we detecting that a particular JIT code needs to call at least one > > >>>>>> function from the kernel at all ? > > >>>>> > > >>>>> Currently we don't track such things and trying very hard to avoid > > >>>>> any special cases for classic vs extended. > > >>>>> > > >>>>>> If the filter contains self-contained code (no call, just inline > > >>>>>> code), then we could use any room in whole vmalloc space, > > >>>>>> not only from the modules (which is something like 2GB total on x86_64) > > >>>>> > > >>>>> I believe there was an effort to make bpf progs and other executable things > > >>>>> to be everywhere too, but I lost the track of it. > > >>>>> It's not that hard to tweak x64 jit to emit 64-bit calls to helpers > > >>>>> when delta between call insn and a helper is more than 32-bit that fits > > >>>>> into call insn. iirc there was even such patch floating around. > > >>>>> > > >>>>> but what motivated you question? do you see 2GB space being full?! > > >>>> > > >>>> A customer seems to hit the limit, with about 75,000 threads, > > >>>> each one having a seccomp filter with 6 pages (plus one guard page > > >>>> given by vmalloc) > > >>> > > >>> Since cbpf doesn't have "fd as a program" concept I suspect > > >>> the same program was loaded 75k times. What a waste of kernel memory. > > >>> And, no, we're not going to extend or fix cbpf for this. > > >>> cbpf is frozen. seccomp needs to start using ebpf. > > >>> It can have one program to secure all threads. > > >>> If necessary single program can be customized via bpf maps > > >>> for each thread. > > >> > > >> Yes, docker seems to have a very generic implementation and should > > >> probably be fixed > > >> ( https://github.com/moby/moby/blob/v17.03.2-ce/profiles/seccomp/seccomp.go ) > > > > > > Even if the seccomp program was optimized to a few bytes, it would > > > still consume at least 2 pages in module vmalloc space, > > > so the limit in number of concurrent programs would be around 262,144 > > > > > > We might ask seccomp guys to detect that the same program is used, by > > > maintaining a hash of already loaded ones. > > > ( I see struct seccomp_filter has a @usage refcount_t ) > > > > +1, that would indeed be worth to pursue as a short term solution. > > I'm not sure how that can work. seccomp's prctl accepts a list of insns. > There is no handle. > kernel can keep a hashtable of all progs ever loaded and do a search > in it before loading another one, but that's an ugly hack. I guess that if such a hack is doable and can save 2GB of memory, then it is an acceptable one. > Another alternative is to attach seccomp prog to parent task > instead of N childrens. seccomp filters are stacked, the parent(s) filter(s) might be very different.
