This version of the signal-free qemu_cpu_kick patches is, ehm, much
better. Variable are accessed either with Java-style volatiles or
protected by memory barriers, and the cleanups go further by removing
qemu/tls.h and C volatiles.
The logic is relatively simple. The I/O thread does (letters in
parentheses indicates the synchronizes-with edges):
run_on_cpu or similar
...
seq_cst write 1 to exit_request (C)
seq_cst read tcg_current_cpu to cpu (B)
if not NULL
write 1 to cpu->exit_request
release barrier (A)
write 1 to cpu->tcg_exit_req
The CPU thread does either this:
(in generated code) read cpu->tcg_exit_req
acquire barrier (A)
read cpu->exit_request
exit from cpu_exec
seq_cst write 0 to exit_request
...
flush_queued_work or similar
or this:
seq_cst write to tcg_current_cpu (B)
seq_cst read from exit_request (C)
exit from cpu_exec
seq_cst write 0 to exit_request
...
flush_queued_work or similar
The non-TLS tcg_current_cpu will go away with multi-threaded TCG.
Paolo
Paolo Bonzini (9):
i8257: rewrite DMA_schedule to avoid hooking into the CPU loop
i8257: remove cpu_request_exit irq
tcg: introduce tcg_current_cpu
remove qemu/tls.h
tcg: assign cpu->current_tb in a simpler place
tcg: synchronize cpu->exit_request and cpu->tcg_exit_req accesses
tcg: synchronize exit_request and tcg_current_cpu accesses
use qemu_cpu_kick instead of cpu_exit or qemu_cpu_kick_thread
tcg: signal-free qemu_cpu_kick