On 06/21/2018 08:59 PM, Edgecombe, Rick P wrote: > On Thu, 2018-06-21 at 15:37 +0200, Jann Horn wrote: >> On Thu, Jun 21, 2018 at 12:34 AM Kees Cook <keescook@xxxxxxxxxxxx> >> wrote: >>> And most systems have <200 modules, really. I have 113 on a desktop >>> right now, 63 on a server. So this looks like a trivial win. >> But note that the eBPF JIT also uses module_alloc(). Every time a BPF >> program (this includes seccomp filters!) is JIT-compiled by the >> kernel, another module_alloc() allocation is made. For example, on my >> desktop machine, I have a bunch of seccomp-sandboxed processes thanks >> to Chrome. If I enable the net.core.bpf_jit_enable sysctl and open a >> few Chrome tabs, BPF JIT allocations start showing up between >> modules: >> >> # grep -C1 bpf_jit_binary_alloc /proc/vmallocinfo | cut -d' ' -f 2- >> 20480 load_module+0x1326/0x2ab0 pages=4 vmalloc N0=4 >> 12288 bpf_jit_binary_alloc+0x32/0x90 pages=2 vmalloc N0=2 >> 20480 load_module+0x1326/0x2ab0 pages=4 vmalloc N0=4 >> -- >> 20480 load_module+0x1326/0x2ab0 pages=4 vmalloc N0=4 >> 12288 bpf_jit_binary_alloc+0x32/0x90 pages=2 vmalloc N0=2 >> 36864 load_module+0x1326/0x2ab0 pages=8 vmalloc N0=8 >> -- >> 20480 load_module+0x1326/0x2ab0 pages=4 vmalloc N0=4 >> 12288 bpf_jit_binary_alloc+0x32/0x90 pages=2 vmalloc N0=2 >> 40960 load_module+0x1326/0x2ab0 pages=9 vmalloc N0=9 >> -- >> 20480 load_module+0x1326/0x2ab0 pages=4 vmalloc N0=4 >> 12288 bpf_jit_binary_alloc+0x32/0x90 pages=2 vmalloc N0=2 >> 253952 load_module+0x1326/0x2ab0 pages=61 vmalloc N0=61 >> >> If you use Chrome with Site Isolation, you have a few dozen open >> tabs, >> and the BPF JIT is enabled, reaching a few hundred allocations might >> not be that hard. >> >> Also: What's the impact on memory usage? Is this going to increase >> the >> number of pagetables that need to be allocated by the kernel per >> module_alloc() by 4K or 8K or so? > Thanks, it seems it might require some extra memory. I'll look into it > to find out exactly how much. > > I didn't include eBFP modules in the randomization estimates, but it > looks like they are usually smaller than a page. So with the slight > leap that the larger normal modules based estimate is the worst case, > you should still get ~800 modules at 18 bits. After that it will start > to go down to 10 bits and so in either case it at least won't regress > the randomness of the existing algorithm. Assume typically complex (real) programs at around 2.5k BPF insns today. In our case it's max a handful per net device, thus approx per netns (veth) which can be few hundreds. Worst case is 4k that BPF allows and then JITs. There's a BPF kselftest suite you could also run to check on worst case upper bounds.
