On Fri, Mar 31, 2023 at 11:08 AM Dmitry Safonov 0x7f454c46@gmail.com wrote:
On Fri, 31 Mar 2023 at 17:52, Axel Rasmussen axelrasmussen@google.com wrote:
On Thu, Mar 30, 2023 at 3:27 PM Peter Xu peterx@redhat.com wrote:
On Thu, Mar 30, 2023 at 12:04:09PM -0700, Axel Rasmussen wrote:
On Thu, Mar 30, 2023 at 8:57 AM Peter Xu peterx@redhat.com wrote:
This is a proposal to revert commit 914eedcb9ba0ff53c33808.
I found this when writting a simple UFFDIO_API test to be the first unit test in this set. Two things breaks with the commit:
- UFFDIO_API check was lost and missing. According to man page, the
kernel should reject ioctl(UFFDIO_API) if uffdio_api.api != 0xaa. This check is needed if the api version will be extended in the future, or user app won't be able to identify which is a new kernel.
- Feature flags checks were removed, which means UFFDIO_API with a
feature that does not exist will also succeed. According to the man page, we should (and it makes sense) to reject ioctl(UFFDIO_API) if unknown features passed in.
If features/flags are not checked in kernel, and the kernel doesn't return an error on an unknown flag/error, that makes the syscall non-extendable, meaning that adding any new feature may break existing software, which doesn't sanitize them properly. https://lwn.net/Articles/588444/
I don't think the same problem applies here. In the case of syscalls, the problem is the only way the kernel can communicate is by the EINVAL return value. Without the check, if a call succeeds the caller can't tell: was the flag supported + applied, or unrecognized + ignored?
With UFFDIO_API (we aren't talking about userfaultfd(2) itself), when you pass in a set of flags, we return the subset of flags which were enabled, in addition to the return code. So via that mechanism, one is "able to check whether it is running on a kernel where [userfaultfd] supports [the feature]" as the article describes - the only difference is, the caller must check the returned set of features, instead of checking for an error code. I don't think it's exactly *how* userspace can check that's important, but rather *that* it can check.
Another important difference: I have a hard time imagining a case where adding a new feature could break userspace, even with my approach, but let's say for the sake of argument one arises in the future. Unlike normal syscalls, we have the UFFD_API version check, so we have the option of incrementing that to separate users relying on the old behavior, from users willing to deal with the new behavior.
(Syscalls can kind of replicate this by adding a new syscall, like clone() vs clone2(), but I think that's messier than the API version check being built-in to the API.)
See a bunch of painful exercises from syscalls with numbers in the end: https://lwn.net/Articles/792628/ To adding an additional setsockopt() because an old one didn't have sanity checks for flags: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?i... (not the best example, as the new setsockopt() didn't check flags for sanity as well (sic!), but that's near the code I work on now)
This is even documented nowadays: https://www.kernel.org/doc/html/latest/process/adding-syscalls.html#designin...
...and everyone knows what happens when you blame userspace for breaking by not doing what you would have expected it to do: https://lkml.org/lkml/2012/12/23/75
100% agreed. :)
[..]
There's one reason that we may consider keeping the behavior. IMHO it is when there're major softwares that uses the "wrong" ABI (let's say so; because it's not following the man pages). If you're aware any such major softwares (especially open sourced) will break due to this revert patch, please shoot.
Well, I did find one example, criu: https://github.com/checkpoint-restore/criu/blob/criu-dev/criu/uffd.c#L266
Mike can speak better than me about uffd, but AFAICS, CRIU correctly detects features with kerneldat/kdat: https://github.com/checkpoint-restore/criu/blob/criu-dev/criu/kerndat.c#L123...
Ah, right, this is the simplest case where no optional features are asked for. So, it's not a great example; this particular case would look the same regardless of what the kernel does.
So, doing a sane thing in kernel shouldn't break CRIU (at least here).
Thanks, Dmitry