This series introduces a per-TLB lock. This removes existing UB
(e.g. memset racing with cmpxchg on another thread while flushing),
and in my opinion makes the TLB code simpler to understand.
I had a bit of trouble finding the best place to initialize the
mutex, since it has to be called before tlb_flush, and tlb_flush
is called quite early during cpu initialization. I settled on
cpu_exec_realizefn, since then cpu->env_ptr has been set
but tlb_flush hasn't yet been called.
Perf-wise this change does have a small impact (~2% slowdown for
the aarch64 bootup+shutdown test; 1.2% comes from using atomic_read
consistently), but I think this is a fair price for avoiding UB.
Numbers below.
Initially I tried using atomics instead of memset for flushing (i.e.
no mutex), and the slowdown is close to 2X due to the repeated
(full) memory barriers. That's when I turned to using a lock.
Host: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz
- Before this series:
Performance counter stats for 'taskset -c 0 ../img/aarch64/die.sh' (10 runs):
7464.797838 task-clock (msec) # 0.998 CPUs utilized
( +- 0.14% )
31,473,652,436 cycles # 4.216 GHz
( +- 0.14% )
57,032,288,549 instructions # 1.81 insns per cycle
( +- 0.08% )
10,239,975,873 branches # 1371.769 M/sec
( +- 0.07% )
172,150,358 branch-misses # 1.68% of all branches
( +- 0.12% )
7.482009203 seconds time elapsed
( +- 0.18% )
- After:
Performance counter stats for 'taskset -c 0 ../img/aarch64/die.sh' (10 runs):
7621.625434 task-clock (msec) # 0.999 CPUs utilized
( +- 0.10% )
32,149,898,976 cycles # 4.218 GHz
( +- 0.10% )
58,168,454,452 instructions # 1.81 insns per cycle
( +- 0.10% )
10,486,183,612 branches # 1375.846 M/sec
( +- 0.10% )
173,900,633 branch-misses # 1.66% of all branches
( +- 0.11% )
7.632067213 seconds time elapsed
( +- 0.10% )
This series is checkpatch-clean. You can fetch the code from:
https://github.com/cota/qemu/tree/tlb-lock
Thanks,
Emilio
