On Fri, Dec 30, 2016 at 10:17:53AM -0800, Davidlohr Bueso wrote: > Secondly for a higher overview, an unlink microbenchmark was used, > which pounds on a single file with open, close,unlink combos with > increasing thread counts (up to 4x ncpus). While the workload is > quite unrealistic, it does contend a lot on the inode mutex or now > rwsem. With the archs I had access to, the differences are as follows: > > == 1. arm64 == > > 0000000000002784 <set_task_state>: > 2784: f9000c1f str xzr, [x0,#24] > > 0000000000002790 <set_current_state>: > 2790: d5384100 mrs x0, sp_el0 > 2794: f9000c1f str xzr, [x0,#24] > > Avg runtime set_task_state(): 2648 msecs > Avg runtime set_current_state(): 2686 msecs
> Unsurprisingly, the big looser is arm64, due to the masking of sp_el0. > otoh, x86-64 (known to be fast for get_current()/this_cpu_read_stable() > caching) and ppc64 (with paca) show similar improvements in the unlink > microbenches. x86's write latencies delta is similar to the opposite of > arm64: 50ms vs -40ms, respectively. The small delta for ppc64 (2ms), does > not represent the gains on the unlink runs. In the case of x86, there was > a decent amount of variation in the latency runs, but always within a 20 > to 50ms increase), ppc was more constant. > > So, do we want to get rid of the interface (and improve performance on > other archs) at the expense of arm64? Can arm64 do better? We can defineitely do better; the asm constraints in read_sysreg() are overly pessimistic for get_current(). Does the below help? Thanks, Mark. ---->8---- diff --git a/arch/arm64/include/asm/current.h b/arch/arm64/include/asm/current.h index f2bcbe2..c9ba5ac 100644 --- a/arch/arm64/include/asm/current.h +++ b/arch/arm64/include/asm/current.h @@ -11,7 +11,11 @@ static __always_inline struct task_struct *get_current(void) { - return (struct task_struct *)read_sysreg(sp_el0); + struct task_struct *tsk; + + asm ("mrs %0, sp_el0" : "=r" (tsk)); + + return tsk; } #define current get_current()