On Thu, Jan 18, 2018 at 08:58:08AM -0800, Dan Williams wrote: > On Thu, Jan 18, 2018 at 5:18 AM, Will Deacon <will.deacon@xxxxxxx> wrote: > > On Thu, Jan 11, 2018 at 05:41:08PM -0800, Dan Williams wrote: > >> On Thu, Jan 11, 2018 at 5:19 PM, Linus Torvalds > >> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > >> > On Thu, Jan 11, 2018 at 4:46 PM, Dan Williams <dan.j.williams@xxxxxxxxx> wrote: > >> >> > >> >> This series incorporates Mark Rutland's latest ARM changes and adds > >> >> the x86 specific implementation of 'ifence_array_ptr'. That ifence > >> >> based approach is provided as an opt-in fallback, but the default > >> >> mitigation, '__array_ptr', uses a 'mask' approach that removes > >> >> conditional branches instructions, and otherwise aims to redirect > >> >> speculation to use a NULL pointer rather than a user controlled value. > >> > > >> > Do you have any performance numbers and perhaps example code > >> > generation? Is this noticeable? Are there any microbenchmarks showing > >> > the difference between lfence use and the masking model? > >> > >> I don't have performance numbers, but here's a sample code generation > >> from __fcheck_files, where the 'and; lea; and' sequence is portion of > >> array_ptr() after the mask generation with 'sbb'. > >> > >> fdp = array_ptr(fdt->fd, fd, fdt->max_fds); > >> 8e7: 8b 02 mov (%rdx),%eax > >> 8e9: 48 39 c7 cmp %rax,%rdi > >> 8ec: 48 19 c9 sbb %rcx,%rcx > >> 8ef: 48 8b 42 08 mov 0x8(%rdx),%rax > >> 8f3: 48 89 fe mov %rdi,%rsi > >> 8f6: 48 21 ce and %rcx,%rsi > >> 8f9: 48 8d 04 f0 lea (%rax,%rsi,8),%rax > >> 8fd: 48 21 c8 and %rcx,%rax > >> > >> > >> > Having both seems good for testing, but wouldn't we want to pick one in the end? > >> > >> I was thinking we'd keep it as a 'just in case' sort of thing, at > >> least until the 'probably safe' assumption of the 'mask' approach has > >> more time to settle out. > > > > From the arm64 side, the only concern I have (and this actually applies to > > our CSDB sequence as well) is the calculation of the array size by the > > caller. As Linus mentioned at the end of [1], if the determination of the > > size argument is based on a conditional branch, then masking doesn't help > > because you bound within the wrong range under speculation. > > > > We ran into this when trying to use masking to protect our uaccess routines > > where the conditional bound is either KERNEL_DS or USER_DS. It's possible > > that a prior conditional set_fs(KERNEL_DS) could defeat the masking and so > > we'd need to throw some heavy barriers in set_fs to make it robust. > > At least in the conditional mask case near set_fs() usage the approach > we are taking is to use a barrier. I.e. the following guidance from > Linus: > > "Basically, the rule is trivial: find all 'stac' users, and use address > masking if those users already integrate the limit check, and lfence > they don't." > > ...which translates to narrow the pointer for get_user() and use a > barrier for __get_user(). Great, that matches my thinking re set_fs but I'm still worried about finding all the places where the bound is conditional for other users of the macro. Then again, finding the places that need this macro in the first place is tough enough so perhaps analysing the bound calculation doesn't make it much worse. Will
