On Fri, Nov 01, 2024 at 03:58:07PM -0500, Eric W. Biederman wrote:
Roman Gushchin roman.gushchin@linux.dev writes:
On Fri, Nov 01, 2024 at 02:51:00PM -0500, Eric W. Biederman wrote:
Roman Gushchin roman.gushchin@linux.dev writes:
Prior to commit d64696905554 ("Reimplement RLIMIT_SIGPENDING on top of ucounts") UCOUNT_RLIMIT_SIGPENDING rlimit was not enforced for a class of signals. However now it's enforced unconditionally, even if override_rlimit is set.
Not true.
It added a limit on the number of siginfo structures that a container may allocate. Have you tried not limiting your container?
This behavior change caused production issues.
For example, if the limit is reached and a process receives a SIGSEGV signal, sigqueue_alloc fails to allocate the necessary resources for the signal delivery, preventing the signal from being delivered with siginfo. This prevents the process from correctly identifying the fault address and handling the error. From the user-space perspective, applications are unaware that the limit has been reached and that the siginfo is effectively 'corrupted'. This can lead to unpredictable behavior and crashes, as we observed with java applications.
Note. There are always conditions when the allocation may fail. The structure is allocated with __GFP_ATOMIC so it is much more likely to fail than a typical kernel memory allocation.
But I agree it does look like there is a quality of implementation issue here.
Fix this by passing override_rlimit into inc_rlimit_get_ucounts() and skip the comparison to max there if override_rlimit is set. This effectively restores the old behavior.
Instead please just give the container and unlimited number of siginfo structures it can play with.
Well, personally I'd not use this limit too, but I don't think "it's broken, userspace shouldn't use it" argument is valid.
I said if you don't want the limit don't use it.
A version of "Doctor it hurts when I do this". To which the doctor replies "Don't do that then".
I was also asking that you test with the limit disabled (at user namespace creation time) so that you can verify that is problem.
The maximum for rlimit(RLIM_SIGPENDING) is the rlimit(RLIM_SIGPENDING) value when the user namespace is created.
Given that it took 3 and half years to report this. I am going to say this really looks like a userspace bug.
The trick here is another bug fixed by https://lkml.org/lkml/2024/10/31/185. Basically it's a leak of the rlimit value. If a limit is set and reached in the reality, all following signals will not have a siginfo attached, causing applications which depend on handling SIGSEGV to crash.
I will take a deeper look at the patch you are referring to.
Beyond that your patch is actually buggy, and should not be applied.
If we want to change the semantics and ignore the maximum number of pending signals in a container (when override_rlimit is set) then the code should change the computation of the max value (pegging it at LONG_MAX) and not ignore it.
Hm, isn't the unconditional (new < 0) enough to capture the overflow? Actually I'm not sure I understand how "long new" can be "> LONG_MAX" anyway.
Agreed "new < 0" should catch that, but still splitting the logic between the calculation of max and the test of max is quite confusing. It makes much more sense to put the logic into the calculate of max.
You mean something like this?
diff --git a/kernel/ucount.c b/kernel/ucount.c index 046b3d57ebb4..49fcec41e5b4 100644 --- a/kernel/ucount.c +++ b/kernel/ucount.c @@ -317,11 +317,12 @@ long inc_rlimit_get_ucounts(struct ucounts *ucounts, enum rlimit_type type,
for (iter = ucounts; iter; iter = iter->ns->ucounts) { long new = atomic_long_add_return(1, &iter->rlimit[type]); - if (new < 0 || (!override_rlimit && (new > max))) + if (new < 0 || new > max) goto unwind; if (iter == ucounts) ret = new; - max = get_userns_rlimit_max(iter->ns, type); + if (!override_rlimit) + max = get_userns_rlimit_max(iter->ns, type); /* * Grab an extra ucount reference for the caller when * the rlimit count was previously 0.
--
If you strongly prefer this version, I can send a v2. I like my original version slightly better, but not a strong preference. Please, let me know.
Thanks!