diff mbox series

[v7,8/8] Linux: Use rseq to accelerate getcpu

Message ID 20240201193648.584917-9-mjeanson@efficios.com
State Superseded
Headers
Series Extend rseq support |

Checks

Context Check Description
redhat-pt-bot/TryBot-apply_patch success Patch applied to master at the time it was sent
linaro-tcwg-bot/tcwg_glibc_build--master-arm success Testing passed
redhat-pt-bot/TryBot-32bit success Build for i686
linaro-tcwg-bot/tcwg_glibc_check--master-arm fail Testing failed
linaro-tcwg-bot/tcwg_glibc_build--master-aarch64 success Testing passed
linaro-tcwg-bot/tcwg_glibc_check--master-aarch64 success Testing passed

Commit Message

Michael Jeanson Feb. 1, 2024, 7:36 p.m. UTC 
  On architectures that implement rseq_load32_load32_relaxed() (and thus
define RSEQ_HAS_LOAD32_LOAD32_RELAXED), when the node_id feature is
available, use rseq to fetch the cpu_id and node_id atomically with
respect to preemption and signal delivery to speed up getcpu() compared
to a vsyscall or system call implementation.

Loading both cpu_id and node_id atomically with respect to preemption
is required to ensure consistency of the topology mapping between cpu_id
and node_id due to migration between both loads.

On an aarch64 system (Snapdragon 8cx Gen 3) which lacks a vDSO for
getcpu() we measured an improvement from 130 ns to 1 ns while on x86_64
(i7-8550U) which has a vDSO we measured a more modest improvement from
10 ns to 2 ns.

Co-authored-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
---
 sysdeps/unix/sysv/linux/getcpu.c | 32 ++++++++++++++++++++++++++++++--
 1 file changed, 30 insertions(+), 2 deletions(-)
diff mbox series

Patch

diff --git a/sysdeps/unix/sysv/linux/getcpu.c b/sysdeps/unix/sysv/linux/getcpu.c
index 0e7c3238c9..7e34d6d1eb 100644
--- a/sysdeps/unix/sysv/linux/getcpu.c
+++ b/sysdeps/unix/sysv/linux/getcpu.c
@@ -19,9 +19,10 @@ 
 #include <sched.h>
 #include <sysdep.h>
 #include <sysdep-vdso.h>
+#include <rseq-internal.h>
 
-int
-__getcpu (unsigned int *cpu, unsigned int *node)
+static int
+vsyscall_getcpu (unsigned int *cpu, unsigned int *node)
 {
 #ifdef HAVE_GETCPU_VSYSCALL
   return INLINE_VSYSCALL (getcpu, 3, cpu, node, NULL);
@@ -29,5 +30,32 @@  __getcpu (unsigned int *cpu, unsigned int *node)
   return INLINE_SYSCALL_CALL (getcpu, cpu, node, NULL);
 #endif
 }
+
+#if defined (RSEQ_SIG) && defined (RSEQ_HAS_LOAD32_LOAD32_RELAXED)
+int
+__getcpu (unsigned int *cpu, unsigned int *node)
+{
+  /* Check if rseq is registered and the node_id feature is available.  */
+  if (__glibc_likely (rseq_node_id_available()))
+  {
+    struct rseq_area *rseq_area = rseq_get_area();
+
+    if (rseq_load32_load32_relaxed(cpu, &rseq_area->cpu_id,
+                                   node, &rseq_area->node_id) == 0)
+    {
+      /* The critical section was not aborted, return 0.  */
+      return 0;
+    }
+  }
+
+  return vsyscall_getcpu (cpu, node);
+}
+#else
+int
+__getcpu (unsigned int *cpu, unsigned int *node)
+{
+  return vsyscall_getcpu (cpu, node);
+}
+#endif
 weak_alias (__getcpu, getcpu)
 libc_hidden_def (__getcpu)