You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Data race on LOGGER.buf_console between set_buf_console and the logger #807
awkernel_lib::logger::set_buf_console writes LOGGER.buf_console without synchronization while other CPUs read it through log::info! and the other logging macros. This is a data race, which is undefined behavior in Rust.
Writer: set_buf_console runs (*ptr).buf_console = Some(logger) with no lock and no atomic operation. The buffered logger task (applications/awkernel_services/src/buffered_logger.rs) calls it, so it runs after all CPUs have started to execute tasks.
Readers: Logger::enabled, Logger::log, and Logger::flush read self.buf_console on every CPU each time a logging macro is called, also with no synchronization.
Possible effect
Option<&'static dyn Log> is two words: a data pointer and a vtable pointer. The write can be two stores. A CPU that reads between them can get the new data pointer with the old vtable pointer, and log.log(record) then calls through an invalid vtable.
The window is a few instructions and the write happens once at boot, so this is unlikely to occur in practice. I have not reproduced it.
Proposed fix
buf_console is written once and never changed afterward, so a separate AtomicBool that publishes it is enough:
set_buf_console writes the pointer, and then sets the flag with Ordering::Release.
Logger loads the flag with Ordering::Acquire, and reads buf_console only when the flag is set.
With this, no CPU reads the pointer while it is written.
Related
set_raw_console writes raw_console in the same way. It is called from console::register_console, which appears to run during early boot before the other CPUs start. Whether it can run concurrently with logging on another CPU should be checked as part of this issue.
Activity