2024-03-27 09:09:29

by zhangpeng (AS)

[permalink] [raw]
Subject: [PATCH] userfaultfd: early return in dup_userfaultfd()

From: ZhangPeng <[email protected]>

When vma->vm_userfaultfd_ctx.ctx is NULL, vma->vm_flags should have
cleared __VM_UFFD_FLAGS. Therefore, there is no need to down_write or
clear the flag, which will affect fork performance. Fix this by
returning early if octx is NULL in dup_userfaultfd().

By applying this patch we can get a 1.3% performance improvement for
lmbench fork_prot. Results are as follows:
base early return
Process fork+exit: 419.1106 413.4804

Signed-off-by: ZhangPeng <[email protected]>
---
fs/userfaultfd.c | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/fs/userfaultfd.c b/fs/userfaultfd.c
index 60dcfafdc11a..3e6ddda6f159 100644
--- a/fs/userfaultfd.c
+++ b/fs/userfaultfd.c
@@ -657,7 +657,10 @@ int dup_userfaultfd(struct vm_area_struct *vma, struct list_head *fcs)
struct userfaultfd_fork_ctx *fctx;

octx = vma->vm_userfaultfd_ctx.ctx;
- if (!octx || !(octx->features & UFFD_FEATURE_EVENT_FORK)) {
+ if (!octx)
+ return 0;
+
+ if (!(octx->features & UFFD_FEATURE_EVENT_FORK)) {
vma_start_write(vma);
vma->vm_userfaultfd_ctx = NULL_VM_UFFD_CTX;
userfaultfd_set_vm_flags(vma, vma->vm_flags & ~__VM_UFFD_FLAGS);
--
2.25.1



2024-03-27 18:58:42

by Peter Xu

[permalink] [raw]
Subject: Re: [PATCH] userfaultfd: early return in dup_userfaultfd()

On Wed, Mar 27, 2024 at 05:08:35PM +0800, Peng Zhang wrote:
> From: ZhangPeng <[email protected]>
>
> When vma->vm_userfaultfd_ctx.ctx is NULL, vma->vm_flags should have
> cleared __VM_UFFD_FLAGS. Therefore, there is no need to down_write or
> clear the flag, which will affect fork performance. Fix this by
> returning early if octx is NULL in dup_userfaultfd().
>
> By applying this patch we can get a 1.3% performance improvement for
> lmbench fork_prot. Results are as follows:
> base early return
> Process fork+exit: 419.1106 413.4804
>
> Signed-off-by: ZhangPeng <[email protected]>

Reviewed-by: Peter Xu <[email protected]>

--
Peter Xu