2020-05-01 02:16:39

by Mathieu Desnoyers

Subject: [PATCH glibc 2/3] glibc: sched_getcpu(): use rseq cpu_id TLS on Linux (v7)

When available, use the cpu_id field from __rseq_abi on Linux to
implement sched_getcpu(). Fall-back on the vgetcpu vDSO if unavailable.

Benchmarks:

x86-64: Intel E5-2630 [email protected], 16-core, hyperthreading

glibc sched_getcpu(): 13.7 ns (baseline)
glibc sched_getcpu() using rseq: 2.5 ns (speedup: 5.5x)
inline load cpuid from __rseq_abi TLS: 0.8 ns (speedup: 17.1x)

CC: Carlos O'Donell <[email protected]>
CC: Florian Weimer <[email protected]>
CC: Joseph Myers <[email protected]>
CC: Szabolcs Nagy <[email protected]>
CC: Thomas Gleixner <[email protected]>
CC: Ben Maurer <[email protected]>
CC: Peter Zijlstra <[email protected]>
CC: "Paul E. McKenney" <[email protected]>
CC: Boqun Feng <[email protected]>
CC: Will Deacon <[email protected]>
CC: Paul Turner <[email protected]>
CC: [email protected]
CC: [email protected]
CC: [email protected]
---
Changes since v1:
- rseq is only used if both __NR_rseq and RSEQ_SIG are defined.

Changes since v2:
- remove duplicated __rseq_abi extern declaration.

Changes since v3:
- update ChangeLog.

Changes since v4:
- Use atomic_load_relaxed to load the __rseq_abi.cpu_id field, a
consequence of the fact that __rseq_abi is not volatile anymore.
- Include atomic.h which provides atomic_load_relaxed.

Changes since v5:
- Use __ASSUME_RSEQ to detect rseq availability.

Changes since v6:
- Remove use of __ASSUME_RSEQ.
---
sysdeps/unix/sysv/linux/sched_getcpu.c | 27 ++++++++++++++++++++++++--
1 file changed, 25 insertions(+), 2 deletions(-)

diff --git a/sysdeps/unix/sysv/linux/sched_getcpu.c b/sysdeps/unix/sysv/linux/sched_getcpu.c
index c019cfb3cf..2269c4f2bd 100644
--- a/sysdeps/unix/sysv/linux/sched_getcpu.c
+++ b/sysdeps/unix/sysv/linux/sched_getcpu.c
@@ -18,10 +18,15 @@
#include <errno.h>
#include <sched.h>
#include <sysdep.h>
+#include <atomic.h>
#include <sysdep-vdso.h>

-int
-sched_getcpu (void)
+#ifdef HAVE_GETCPU_VSYSCALL
+# define HAVE_VSYSCALL
+#endif
+
+static int
+vsyscall_sched_getcpu (void)
{
unsigned int cpu;
int r = -1;
@@ -32,3 +37,21 @@ sched_getcpu (void)
#endif
return r == -1 ? r : cpu;
}
+
+#include <sys/rseq.h>
+
+#ifdef RSEQ_SIG
+int
+sched_getcpu (void)
+{
+ int cpu_id = atomic_load_relaxed (&__rseq_abi.cpu_id);
+
+ return cpu_id >= 0 ? cpu_id : vsyscall_sched_getcpu ();
+}
+#else
+int
+sched_getcpu (void)
+{
+ return vsyscall_sched_getcpu ();
+}
+#endif
--
2.17.1

2020-05-20 10:16:46

by Florian Weimer

Subject: Re: [PATCH glibc 2/3] glibc: sched_getcpu(): use rseq cpu_id TLS on Linux (v7)

* Mathieu Desnoyers via Libc-alpha:

> diff --git a/sysdeps/unix/sysv/linux/sched_getcpu.c b/sysdeps/unix/sysv/linux/sched_getcpu.c
> index c019cfb3cf..2269c4f2bd 100644
> --- a/sysdeps/unix/sysv/linux/sched_getcpu.c
> +++ b/sysdeps/unix/sysv/linux/sched_getcpu.c
> @@ -18,10 +18,15 @@
> #include <errno.h>
> #include <sched.h>
> #include <sysdep.h>
> +#include <atomic.h>
> #include <sysdep-vdso.h>
>
> -int
> -sched_getcpu (void)
> +#ifdef HAVE_GETCPU_VSYSCALL
> +# define HAVE_VSYSCALL
> +#endif

I think the #ifdef is a result of an incorrect merge of commit
d0def09ff6bbf1537beec305fdfe96a21174fb31 ("linux: Fix vDSO macros build
with time64 interfaces") and it should be removed.

The commit subject should probably say “Linux: Use rseq in sched_getcpu
if available”.

Thanks,
Florian