On Fri, Mar 3, 2023 at 8:31 AM szabolcs.nagy@xxxxxxx <szabolcs.nagy@xxxxxxx> wrote: > > The 03/02/2023 21:17, Edgecombe, Rick P wrote: > > Is the idea that shadow stack would be forced on regardless of if the > > linked libraries support it? In which case it could be allowed to crash > > if they do not? > > execute a binary > - with shstk enabled and locked (only if marked?). > - with shstk disabled and locked. > could be managed in userspace, but it is libc dependent then. > > > > > > - I think it's better to have a new limit specifically for shadow > > > > > stack size (which by default can be RLIMIT_STACK) so userspace > > > > > can adjust it if needed (another reason is that stack size is > > > > > not always a good indicator of max call depth). > > > > Looking at this again, I'm not sure why a new rlimit is needed. It > > seems many of those points were just formulations of that the clone3 > > stack size was not used, but it actually is and just not documented. If > > you disagree perhaps you could elaborate on what the requirements are > > and we can see if it seems tricky to do in a follow up. > > - tiny thread stack and deep signal stack. > (note that this does not really work with glibc because it has > implementation internal signals that don't run on alt stack, > cannot be masked and don't fit on a tiny thread stack, but > with other runtimes this can be a valid use-case, e.g. musl > allows tiny thread stacks, < pagesize.) > > - thread runtimes with clone (glibc uses clone3 but some dont). > > - huge stacks but small call depth (problem if some va limit > is hit or memory overcommit is disabled). > > > > "sigaltshstk() is separate from sigaltstack(). You can have one > > > without the other, neither or both together. Because the shadow > > > stack specific state is pushed to the shadow stack, the two > > > features don’t need to know about each other." > ... > > > i don't see why automatic alt shadow stack allocation would > > > not work (kernel manages it transparently when an alt stack > > > is installed or disabled). > > > > Ah, I think I see where maybe I can fill you in. Andy Luto had > > discounted this idea out of hand originally, but I didn't see it at > > first. sigaltstack lets you set, retrieve, or disable the shadow stack, > > right... But this doesn't allocate anything, it just sets where the > > next signal will be handled. This is different than things like threads > > where there is a new resources being allocated and it makes coming up > > with logic to guess when to de-allocate the alt shadow stack difficult. > > You probably already know... > > > > But because of this there can be some modes where the shadow stack is > > changed while on it. For one example, SS_AUTODISARM will disable the > > alt shadow stack while switching to it and restore when sigreturning. > > At which point a new altstack can be set. In the non-shadow stack case > > this is nice because future signals won't clobber the alt stack if you > > switch away from it (swapcontext(), etc). But it also means you can > > "change" the alt stack while on it ("change" sort of, the auto disarm > > results in the kernel forgetting it temporarily). > > the problem with swapcontext is that it may unmask signals > that run on the alt stack, which means the code cannot jump > back after another signal clobbered the alt stack. > > the non-standard SS_AUTODISARM aims to solve this by disabling > alt stack settings on signal entry until the handler returns. > > so this use case is not about supporting swapcontext out, but > about jumping back. however that does not work reliably with > this patchset: if swapcontext goes to the thread stack (and > not to another stack e.g. used by makecontext), then jump back > fails. (and if there is a sigaltshstk installed then even jump > out fails.) > > assuming > - jump out from alt shadow stack can be made to work. > - alt shadow stack management can be automatic. > then this can be improved so jump back works reliably. > > > I hear where you are coming from with the desire to have it "just work" > > with existing code, but I think the resulting ABI around the alt shadow > > stack allocation lifecycle would be way too complicated even if it > > could be made to work. Hence making a new interface. But also, the idea > > was that the x86 signal ABI should support handling alt shadow stacks, > > which is what we have done with this series. If a different interface > > for configuring it is better than the one from the POC, I'm not seeing > > a problem jump out. Is there any specific concern about backwards > > compatibility here? > > sigaltstack syscall behaviour may be hard to change later > and currently > - shadow stack overflow cannot be recovered from. > - longjmp out of signal handler fails (with sigaltshstk). > - SS_AUTODISARM does not work (jump back can fail). > > > > "Since shadow alt stacks are a new feature, longjmp()ing from an > > > alt shadow stack will simply not be supported. If a libc want’s > > > to support this it will need to enable WRSS and write it’s own > > > restore token." > > > > > > i think longjmp should work without enabling writes to the shadow > > > stack in the libc. this can also affect unwinding across signal > > > handlers (not for c++ but e.g. glibc thread cancellation). > > > > glibc today does not support longjmp()ing from a different stack (for > > example even today after a swapcontext()) when shadow stack is used. If > > glibc used wrss it could be supported maybe, but otherwise I don't see > > how the HW can support it. > > > > HJ and I were actually just discussing this the other day. Are you > > looking at this series with respect to the arm shadow stack feature by > > any chance? I would love if glibc/tools would document what the shadow > > stack limitations are. If the all the arch's have the same or similar > > limitations perhaps this could be one developer guide. For the most > > part though, the limitations I've encountered are in glibc and the > > kernel is more the building blocks. > > well we hope that shadow stack behaviour and limitations can > be similar across targets. > > longjmp to different stack should work: it can do the same as > setcontext/swapcontext: scan for the pivot token. then only > longjmp out of alt shadow stack fails. (this is non-conforming > longjmp use, but e.g. qemu relies on it.) Restore token may not be used with longjmp. Unlike setcontext/swapcontext, longjmp is optional. If longjmp isn't called, there will be an extra token on shadow stack and RET will fail. > for longjmp out of alt shadow stack, the target shadow stack > needs a pivot token, which implies the kernel needs to push that > on signal entry, which can overflow. but i suspect that can be > handled the same way as stackoverflow on signal entry is handled. > > > A general comment. Not sure if you are aware, but this shadow stack > > enabling effort is quite old at this point and there have been many > > discussions on these topics stretching back years. The latest > > conversation was around getting this series into linux-next soon to get > > some testing on the MM pieces. I really appreciate getting this ABI > > feedback as it is always tricky to get right, but at this stage I would > > hope to be focusing mostly on concrete problems. > > > > I also expect to have some amount of ABI growth going forward with all > > the normal things that entails. Shadow stack is not special in that it > > can come fully finalized without the need for the real world usage > > iterative feedback process. At some point we need to move forward with > > something, and we have quite a bit of initial changes at this point. > > > > So I would like to minimize the initial implementation unless anyone > > sees any likely problems with future growth. Can you be clear if you > > see any concrete problems at this point or are more looking to evaluate > > the design reasoning? I'm under the assumption there is nothing that > > would prohibit linux-next testing while any ABI shakedown happens > > concurrently at least? > > understood. > > the points that i think are worth raising: > > - shadow stack size logic may need to change later. > (it can be too big, or too small in practice.) > - shadow stack overflow is not recoverable and the > possible fix for that (sigaltshstk) breaks longjmp > out of signal handlers. > - jump back after SS_AUTODISARM swapcontext cannot be > reliable if alt signal uses thread shadow stack. > - the above two concerns may be mitigated by different > sigaltstack behaviour which may be hard to add later. > - end token for backtrace may be useful, if added > later it can be hard to check. > > thanks. -- H.J.