On 6/18/22 17:10, Huacai Chen wrote:
Hi,
On Fri, Jun 17, 2022 at 11:35 PM hev <r@xxxxxx> wrote:
Hello,
On Fri, Jun 17, 2022 at 10:57 PM Huacai Chen <chenhuacai@xxxxxxxxxxx> wrote:
We test 20 million times of getcpu(), the real syscall version take 25
seconds, while the vsyscall version take only 2.4 seconds.
Signed-off-by: Huacai Chen <chenhuacai@xxxxxxxxxxx>
---
arch/loongarch/include/asm/vdso.h | 4 +++
arch/loongarch/include/asm/vdso/vdso.h | 10 +++++-
arch/loongarch/kernel/vdso.c | 23 +++++++++-----
arch/loongarch/vdso/Makefile | 3 +-
arch/loongarch/vdso/vdso.lds.S | 1 +
arch/loongarch/vdso/vgetcpu.c | 43 ++++++++++++++++++++++++++
6 files changed, 74 insertions(+), 10 deletions(-)
create mode 100644 arch/loongarch/vdso/vgetcpu.c
diff --git a/arch/loongarch/include/asm/vdso.h b/arch/loongarch/include/asm/vdso.h
index 8f8a0f9a4953..e76d5e37480d 100644
--- a/arch/loongarch/include/asm/vdso.h
+++ b/arch/loongarch/include/asm/vdso.h
@@ -12,6 +12,10 @@
#include <asm/barrier.h>
+typedef struct vdso_pcpu_data {
+ u32 node;
+} ____cacheline_aligned_in_smp vdso_pcpu_data;
+
/*
* struct loongarch_vdso_info - Details of a VDSO image.
* @vdso: Pointer to VDSO image (page-aligned).
diff --git a/arch/loongarch/include/asm/vdso/vdso.h b/arch/loongarch/include/asm/vdso/vdso.h
index 5a01643a65b3..94055f7c54b7 100644
--- a/arch/loongarch/include/asm/vdso/vdso.h
+++ b/arch/loongarch/include/asm/vdso/vdso.h
@@ -8,6 +8,13 @@
#include <asm/asm.h>
#include <asm/page.h>
+#include <asm/vdso.h>
+
+#if PAGE_SIZE < SZ_16K
+#define VDSO_DATA_SIZE SZ_16K
Whether we add members to the vdso data structure or extend
SMP_CACHE_BYTES/NR_CPUS, the static VDSO_DATA_SIZE may not match, and
there is no assertion checking to help us catch bugs early. So I
suggest defining VDSO_DATA_SIZE as ALIGN_UP(sizeof (struct vdso_data),
PAGE_SIZE).
VSYSCALL usage is very limited (you know, VSYSCALL appears for so many
years, but the number nearly doesn't increase until now), so I think
16KB is enough in the future.
I don't think omitting compile-time assertions for *correctness* is
worth the negligible improvement in brevity and ease of maintenance. In
fact, static checks for correctness actually *lightens* maintenance
burden, by explicitly calling out the assumptions so that newcomers
(i.e. me or some other random linux/arch developer refactoring code)
would find them very helpful.
So I'm in support for declaring the VDSO_DATA_SIZE explicitly in terms
of sizeof(struct vdso_data) and PAGE_SIZE.
