Restartable sequences

I think the whole idea of the approach (using a memory region) is to avoid making syscalls, basically for performance. This way the scheduler can just compare the Instruction Pointer to detect if a restart is needed. Note that the critical section should run much more often than the scheduler for this to be a net win.

Maybe another (naive?) approach would be to share a memory pointer with the kernel where the thread can publish a restart address. Say something like

set_atomic_restart(&ptr);

with ptr some memory owned by the thread big enough for a pointer. With that, an application can do

ptr=&&some_label;

when entering a critical section, and ptr=NULL; when leaving. If the scheduler needs to pre-empt the thread and finds that ptr has a non-null value, execution would restart at ptr instead of the saved IP.