From: Steven Price <email@example.com> To: Amanieu d'Antras <firstname.lastname@example.org> Cc: Arnd Bergmann <email@example.com>, Ryan Houdek <Houdek.Ryan@fex-emu.org>, Catalin Marinas <firstname.lastname@example.org>, Will Deacon <email@example.com>, Mark Rutland <firstname.lastname@example.org>, David Laight <David.Laight@aculab.com>, Mark Brown <email@example.com>, Linux ARM <firstname.lastname@example.org>, Linux Kernel Mailing List <email@example.com> Subject: Re: [RESEND PATCH v4 8/8] arm64: Allow 64-bit tasks to invoke compat syscalls Date: Mon, 24 May 2021 12:20:38 +0100 [thread overview] Message-ID: <firstname.lastname@example.org> (raw) In-Reply-To: <CA+y5pbSbky2kS+O5j9bn57nROdYaYeHcd2R-46X1cc388PKOvg@mail.gmail.com> On 21/05/2021 20:18, Amanieu d'Antras wrote: > On Fri, May 21, 2021 at 9:51 AM Steven Price <email@example.com> wrote: >>>> In those cases to correctly emulate seccomp, isn't Tango is going to >>>> have to implement the seccomp filter in user space? >>> >>> I have not implemented user-mode seccomp emulation because it can >>> trivially be bypassed by spawning a 64-bit child process which runs >>> outside Tango. Even when spawning another translated process, the >>> user-mode filter will not be preserved across an execve. >> >> Clearly if you have user-mode seccomp emulation then you'd hook execve >> and either install the real BPF filter (if spawning a 64 bit child >> outside Tango) or ensure that the user-mode emulation is passed on to >> the child (if running within Tango). > > Spawning another process is just an example. Fundamentally, Tango is > not intended or designed to be a sandbox around the 32-bit code. For > example, many of the newer ioctls use u64 instead of a pointer type to > avoid the need for a compat_ioctl handler. This means that such ioctls > could be abused to read/write any address in the process address > space, including the code that is performing the usermode seccomp > emulation. Indeed, I think it's almost impossible to fully preserve all the security aspects of the seccomp filter - whether that's because the 32 bit code can "attack" Tango itself, or whether it's because Tango requires white-listing ioctls that would normally be excluded by the seccomp filter. Even with the compat syscalls exposed, ioctls that take a u64 and don't do any special handling for a compat call would happily accept addresses >4GB even though they wouldn't if it was a real compat process. For the user space emulation of the compat syscalls, Tango does have the option of attempting to fix up such ioctls to restrict the address range - but it's obviously a very large set of ioctls to audit if you want to be complete. >> You already have a potential 'issue' here of a 64 bit process setting up >> a seccomp filter and then execve()ing a 32 bit (Tango'd) process. The >> set of syscalls needed for the system which supports AArch32 natively is >> going to be different from the syscalls needed for Tango. (Fundamentally >> this is a major limitation with the whole seccomp syscall filtering >> approach). > > The specific example I had in mind here is Android which installs a > global seccomp filter on the zygote process from which app processes > are forked from. This filter is designed for mixed arm32/arm64 systems > and therefore has syscall whitelists for both AArch32 and AArch64. > This filter allows 32-bit processes to spawn 64-bit processes and > vice-versa: for example, many 32-bit apps will invoke another 32-bit > executable via system() which uses a 64-bit /system/bin/sh. Given that Tango is going to be constrained by the 64 bit version of the seccomp filter, I don't really understand why the emulated 32-bit process needs to be separately constrained by the list of 32-bit calls in this case. The problem really comes if the controlling 64 bit process is written "knowing" that it is executing a 32 bit process and tailors the seccomp filter for that. I don't see any solution actually solving that (Tango requires extra 64 bit syscalls which won't be naturally included). >>>> I guess the question comes down to how big a hole is >>>> syscall_in_tango_whitelist() - if Tango only requires a small set of >>>> syscalls then there is still some security benefit, but otherwise this >>>> doesn't seem like a particularly big benefit considering you're already >>>> going to need the BPF infrastructure in user space. >>> >>> Currently Tango only whitelists ~50 syscalls, which is small enough to >>> provide security benefits and definitely better than not supporting >>> seccomp at all. >> >> Agreed, and I don't want to imply that this approach is necessarily >> wrong. But given that the approach of getting the kernel to do the >> compat syscall filtering is not perfect, I'm not sure in itself it's a >> great justification for needing the kernel to support all the compat >> syscalls. > > I feel that exposing all compat syscalls to 64-bit processes is better > than the alternative of only exposing a subset of them. Of the top of > my head I can think of quite a few compat syscalls that cannot be > fully emulated in userspace and would need to be exposed in the > kernel: > - mmap/mremap/shmat/io_setup: anything that allocates VM space needs > to return a pointer in the low 4GB. So a "generic" way of requested the kernel limit the address space for allocations would be potentially useful for other purposes. Adding a new syscall for this purpose would be sensible. We already have (at least) two "hacks" in mmap for controlling the address range that can be used: * MAP_32BIT - x86 only, and really "31 bit" * Providing a mmap() hint with the top bits set to opt-in to 52-bit VAs. A well defined mechanism for controlling the valid VA range for allocations would be much better than adding more hacks - and bonus points if it works for all the different types of allocation unlike the above. > - ioctl: too many variants to reasonably maintain a separate compat > layer in userspace. I do agree that it probably makes sense to expose a compat_ioctl() syscall, but this isn't a complete fix and user space is likely to need to do some ioctl translations. And x86->aarch64 is obviously going to require translation in many cases. > - getdents/lseek: ext4 uses 32-bit directory offsets for 32-bit processes. My guess is that a getdents32() call makes sense, but I haven't looked into this. > - get_robust_list/set_robust_list: different in-memory ABI for > 32/64-bit processes. Here some careful thought needs to go into how both 32 bit and 64 bit robust lists would coexist. I haven't looked into it, but I doubt it's as simple as just exposing the compat calls. > - open: don't force O_LARGEFILE for 32-bit processes. At first glance it seems reasonable to expose the compat version. > - io_uring_create: different in-memory ABI for 32/64-bit processes. I'm not familiar enough with io_uring to sensibly comment on this. > - (and possibly many others) > > Also consider the churn involved when adding a new syscall which > behaves differently in compat processes: rather than just using > in_compat_syscall() or wiring up a COMPAT_SYSCALL_DEFINE, a compat > variant of this syscall would also need to be added to the 64-bit > syscall table to support translation layers like Tango and FEX. Just exposing a new compat call to translated processes would also be potentially dangerous - Tango needs to actually consider whether the syscall is safe before exposing it. Yes there's some potential cost, but in general there's much more of an attempt to make the compat syscalls the same as native ones wherever possible. So I would expect that the cases needing special handling are reasonably rare. >> One other thought: I suspect in practise there aren't actually many >> variations in the BPF programs used with seccomp. It may well be quite >> possible to convert the 32-bit syscall filtering programs to filter the >> equivalent 64-bit syscalls that Tango would use. Sadly this would be >> fragile if a program used a BPF program which didn't follow the "normal" >> pattern. > > This might work for simple filters that only look at the syscall > number, but becomes much harder when the filter also inspects the > syscall arguments. Yes it heavily relies on being able to pattern match what specific programs/libraries do. My hunch is that most seccomp filters are set by a very small number of programs/libraries. But I've no data to back that up. Steve _______________________________________________ linux-arm-kernel mailing list firstname.lastname@example.org http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
next prev parent reply other threads:[~2021-05-24 17:51 UTC|newest] Thread overview: 20+ messages / expand[flat|nested] mbox.gz Atom feed top 2021-05-18 9:06 [RESEND PATCH v4 0/8] " Amanieu d'Antras 2021-05-18 9:06 ` [RESEND PATCH v4 1/8] mm: Add arch_get_mmap_base_topdown macro Amanieu d'Antras 2021-05-18 9:06 ` [RESEND PATCH v4 2/8] hugetlbfs: Use arch_get_mmap_* macros Amanieu d'Antras 2021-05-18 9:06 ` [RESEND PATCH v4 3/8] mm: Support mmap_compat_base with the generic layout Amanieu d'Antras 2021-05-18 9:06 ` [RESEND PATCH v4 4/8] arm64: Separate in_compat_syscall from is_compat_task Amanieu d'Antras 2021-05-18 9:06 ` [RESEND PATCH v4 5/8] arm64: mm: Use HAVE_ARCH_COMPAT_MMAP_BASES Amanieu d'Antras 2021-05-18 9:06 ` [RESEND PATCH v4 6/8] arm64: Add a compat syscall flag to thread_info Amanieu d'Antras 2021-05-18 9:06 ` [RESEND PATCH v4 7/8] arm64: Forbid calling compat sigreturn from 64-bit tasks Amanieu d'Antras 2021-05-18 9:06 ` [RESEND PATCH v4 8/8] arm64: Allow 64-bit tasks to invoke compat syscalls Amanieu d'Antras 2021-05-18 13:02 ` Arnd Bergmann 2021-05-18 20:26 ` David Laight 2021-05-18 22:41 ` Ryan Houdek 2021-05-18 23:51 ` Amanieu d'Antras 2021-05-19 15:30 ` Steven Price 2021-05-19 16:14 ` Amanieu d'Antras 2021-05-21 8:51 ` Steven Price 2021-05-21 19:18 ` Amanieu d'Antras 2021-05-24 11:20 ` Steven Price [this message] 2021-05-24 12:38 ` David Laight 2021-05-18 23:52 ` Ryan Houdek
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --email@example.com \ --firstname.lastname@example.org \ --cc=David.Laight@aculab.com \ --cc=Houdek.Ryan@fex-emu.org \ --email@example.com \ --firstname.lastname@example.org \ --email@example.com \ --firstname.lastname@example.org \ --email@example.com \ --firstname.lastname@example.org \ --email@example.com \ --firstname.lastname@example.org \ --subject='Re: [RESEND PATCH v4 8/8] arm64: Allow 64-bit tasks to invoke compat syscalls' \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: link
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for NNTP newsgroup(s).