Re: increased vmap_area_lock contentions on "n_tty: Move buffers into n_tty_data"

[Lin Ming <minggr@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 1981 bytes --]

On Wed, 2013-09-25 at 07:30 -0400, Peter Hurley wrote:
> On 09/25/2013 05:04 AM, Lin Ming wrote:
> > On Wed, Sep 18, 2013 at 8:22 AM, Peter Hurley <peter@hurleysoftware.com> wrote:
> > [snip]
> >>
> >> Looking over vmalloc.c, the critical section footprint of the vmap_area_lock
> >> could definitely be reduced (even nearly eliminated), but that's a project
> >> for
> >> another day :)
> >
> > Hi Peter,
> >
> > I also looked over vmallo.c, but didn't find obvious way to reduce the
> > critical section footprint.
> > Could you share some hints how to do this?
> 
> vmap_area_list is an RCU list.
> get_vmalloc_info() doesn't need to take the vmap_area_lock at all.
> Look at __purge_vmap_area_lazy() for a howto.

Would you like below patch?

From: Lin Ming <minggr@gmail.com>
Date: Wed, 25 Sep 2013 23:48:19 +0800
Subject: [PATCH] mm/vmalloc.c: eliminate vmap_area_lock in get_vmalloc_info()

vmap_area_list is an RCU list.
get_vmalloc_info() doesn't need to take the vmap_area_lock at all.
Use RCU to protect list scan.

Signed-off-by: Lin Ming <minggr@gmail.com>
---
 mm/vmalloc.c |    6 +++---
 1 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 13a5495..4523c9c 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -2684,14 +2684,14 @@ void get_vmalloc_info(struct vmalloc_info *vmi)
 
 	prev_end = VMALLOC_START;
 
-	spin_lock(&vmap_area_lock);
+	rcu_read_lock();
 
 	if (list_empty(&vmap_area_list)) {
 		vmi->largest_chunk = VMALLOC_TOTAL;
 		goto out;
 	}
 
-	list_for_each_entry(va, &vmap_area_list, list) {
+	list_for_each_entry_rcu(va, &vmap_area_list, list) {
 		unsigned long addr = va->va_start;
 
 		/*
@@ -2718,7 +2718,7 @@ void get_vmalloc_info(struct vmalloc_info *vmi)
 		vmi->largest_chunk = VMALLOC_END - prev_end;
 
 out:
-	spin_unlock(&vmap_area_lock);
+	rcu_read_unlock();
 }
 #endif
 
-- 
1.7.2.5



> 
> Regards,
> Peter Hurley
>