Re: Prevent inconsistent CPU state after sequence of dlclose/dlopen

[Adhemerval Zanella Netto]

On 10/01/25 14:02, Mathieu Desnoyers wrote:
> On 2025-01-10 11:54, Peter Zijlstra wrote:
>> On Fri, Jan 10, 2025 at 10:55:36AM -0500, Mathieu Desnoyers wrote:
>>> Hi,
>>>
>>> I was discussing with Mark Rutland recently, and he pointed out that a
>>> sequence of dlclose/dlopen mapping new code at the same addresses in
>>> multithreaded environments is an issue on ARM, and possibly on Intel/AMD
>>> with the newer TLB broadcast maintenance.
>>
>> What is the exact race? Should not munmap() invalidate the TLBs before
>> it allows overlapping mmap() to complete?
> 
> The race Mark mentioned (on ARM) is AFAIU the following scenario:
> 
> CPU 0                     CPU 1
> 
> - dlopen()
>   - mmap PROT_EXEC @addr
>                           - fetch insn @addr, CPU state expects unchanged insn.
>                           - execute unrelated code

It is not clear to me from userland/libc perspective how this would happen,
since to dlopen get the same address you will need to either dlclose or call
munmap.

Either you have UB where some thread dclose a library while is being used
by a different thread, or the thread will ended up executing a potentially
different code it is intended to do.

Do you have a realworld case where current glibc code show this issue?

> - dlclose(addr)
>   - munmap @addr
> - dlopen()
>   - mmap PROT_EXEC @addr
>                           - fetch new insn @addr. Incoherent CPU state.
> 
>>
>> Any concurrent access after munmap() / before mmap() completes is UB
>> anyway, no?
> 
> The problematic access happens after the second mmap. The issue is
> stale CPU state.
> 
>>
>>> I maintain the membarrier(2) system call, which provides a
>>> MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE command for this
>>> purpose. It's been there since Linux 4.16. It can be configured
>>> out (CONFIG_MEMBARRIER=n), but it's enabled by default.
>>>
>>> Calling this after dlclose() in glibc would prevent this issue.
>>>
>>> Is it handled in some other way, or should we open a bugzilla
>>> entry to track this ?
>>
>> The problem is that the membarrier() call has significant cost, and is
>> only really needed if dlopen() is called right after (in the same
>> location).
> 
> Or if it has any overlapping executable range.
> 
>>
>> Unconditionally adding that barrier, just in case, might regress things,
>> no?
> 
> Or perhaps we could add this barrier within mprotect(2) and munmap(2) in the
> following cases:
> 
> - mprotect removes PROT_EXEC from a mapping,
> - munmap unmaps a PROT_EXEC mapping.
> 
> Else userspace has to explicitly invoke membarrier sync-core from dlclose.
> 
> Thoughts ?
> 
> Thanks,
> 
> Mathieu
> 
>