On Fri, Jul 15, 2022, Gavin Shan wrote:
Hi Paolo and Sean,
On 7/15/22 1:35 AM, Sean Christopherson wrote:
On Thu, Jul 14, 2022, Paolo Bonzini wrote:
Well, I don't think migration_worker() does correct thing, if I'm understanding correctly. The intention seems to force migration on 'main' thread by 'migration' thread? If that is the case, I don't think the following function call has correct parameters.
r = sched_setaffinity(0, sizeof(allowed_mask), &allowed_mask); it should be something like: r = sched_setaffinity(getpid(), sizeof(allowed_mask), &allowed_mask);If we're using sched_setaffinity(0, ...) in the 'migration' thread, the CPU affinity of 'main' thread won't be affected. It means 'main' thread can be migrated from one CPU to another at any time, even in the following point:
int main(...) { : /* * migration can happen immediately after sched_getcpu(). If * CPU affinity of 'main' thread is sticky to one particular * CPU, which 'migration' thread supposes to do, then there * should have no migration. */ cpu = sched_getcpu(); rseq_cpu = READ_ONCE(__rseq.cpu_id); : }So I think the correct fix is to have sched_setaffinity(getpid(), ...) ? Please refer to the manpage.
https://man7.org/linux/man-pages/man2/sched_setaffinity.2.html 'If pid is zero, then the calling thread is used'
Oof, and more explicitly the rest of that sentence clarifies that the result of getpid() will target the main thread (I assume "main" means thread group leader).
Specifying pid as 0 will set the attribute for the calling thread, and passing the value returned from a call to getpid(2) will set the attribute for the main thread of the thread group.
I'm guessing my test worked (in that it reproduced the bug) by virtue of the scheduler trying to colocate all threads in the process.
In my defense, the die.net copy of the manpages quite clearly uses "process"[1], but that was fixed in the manpages in 2013[2]!?!!? So I guess the takeaway is to use only the official manpages.
Anyways, for the code, my preference would be to snapshot gettid() in main() before spawning the migration worker. Same result, but I would rather the test explicitly target the thread doing rseq instead of relying on (a) getpid() targeting only the main thread and (b) the main thread always being the rseq thread. E.g. if for some reason a future patch moves the rseq code to its own worker thread, then getpid() would be incorrect.
Thanks for figuring this out!
[1] https://linux.die.net/man/2/sched_setaffinity [2] 6a7fcf3cc ("sched_setaffinity.2: Clarify that these system calls affect a per-thread attribute")