On 3/7/22 18:18, Warner Losh wrote:
I have a question related to the user-mode emulation and atomics. I asked on IRC, but
thinking about it, I think it may be too complex to discuss in that medium...
In FreeBSD we have a system call that uses host atomic operations to interact memory that
userland also interacts with using atomic operations.
In bsd-user we call the kernel with a special flag for dealing with 32-bit processes
running on a 64-bit kernel. In this case, we use 32-bit-sized atomics to set variables in
the address space of the bsd-user guest. This is used when running armv7 binaries on amd64
hosts.
First question: Is this expected to work? I know I'm a bit vague, so as a followup
question: If there's restrictions on this, what might they be? Do some classes of atomic
operations work, while others may fail or need additional cooperation? Are there any
conformance tests I could compile for FreeBSD/armv7 to test the hypothesis that atomic
operations are misbehaving?
Yes, qatomic_foo is expected to work. It's what we use across threads, and it is expected
to work "in kernel mode", i.e. within cpu_loop().
There are compile-time restrictions on the set of atomic operations, mostly based on what
the host supports. But anything that actually compiles is expected to work (there are a
set of ifdefs if you need something more than the default).
Beyond that, there is start_exclusive() / end_exclusive() which will stop-the-world and
make sure that the current thread is the only one running.
Thanks for any help you might be able to give.
Show the code in question?
r~
