[PATCH 6/5] cpuset: rebind numa vma mempolicy fix

[PATCH 6/5] cpuset: rebind numa vma mempolicy fix

[Paul Jackson]

The patch:
  [PATCH 2/5] cpuset: rebind numa vma mempolicy
requires this fix to avoid a deadlock situation
in certain cpuset removal cases.

It's ok to not complete the refresh_mems() operation
if we notice we are already holding mmap_sem.  We
will try again, next time we allocate memory.

Signed-off-by: Paul Jackson <pj@sgi.com>

---

 (This fix was included in my testing, but in another
  patch that has not been sent in yet.)

 kernel/cpuset.c |    5 +++++
 1 files changed, 5 insertions(+)

--- 2.6.14-rc5-mm1-cpuset-patches.orig/kernel/cpuset.c	2005-11-03 21:18:26.783391082 -0800
+++ 2.6.14-rc5-mm1-cpuset-patches/kernel/cpuset.c	2005-11-03 23:11:15.480042373 -0800
@@ -656,7 +656,12 @@ static void refresh_mems(void)
 	if (current->cpuset_mems_generation != my_cpusets_mem_gen) {
 		struct cpuset *cs;
 		nodemask_t oldmem = current->mems_allowed;
+		struct mm_struct *mm = current->mm;
 
+		/* numa_policy_rebind() needs mmap_sem - don't nest */
+		if (!mm || !down_write_trylock(&mm->mmap_sem))
+			return;
+		up_write(&mm->mmap_sem);
 		down(&callback_sem);
 		task_lock(current);
 		cs = current->cpuset;

-- 
                          I won't rest till it's the best ...
                          Programmer, Linux Scalability
                          Paul Jackson <pj@sgi.com> 1.650.933.1373

Re: [PATCH 6/5] cpuset: rebind numa vma mempolicy fix

[Andrew Morton]

Paul Jackson <pj@sgi.com> wrote:
>
> It's ok to not complete the refresh_mems() operation
>  if we notice we are already holding mmap_sem.  We
>  will try again, next time we allocate memory.

What's the call path? alloc_pages_current() and alloc_page_vma()?

> 
>  --- 2.6.14-rc5-mm1-cpuset-patches.orig/kernel/cpuset.c	2005-11-03 21:18:26.783391082 -0800
>  +++ 2.6.14-rc5-mm1-cpuset-patches/kernel/cpuset.c	2005-11-03 23:11:15.480042373 -0800
>  @@ -656,7 +656,12 @@ static void refresh_mems(void)
>   	if (current->cpuset_mems_generation != my_cpusets_mem_gen) {
>   		struct cpuset *cs;
>   		nodemask_t oldmem = current->mems_allowed;
>  +		struct mm_struct *mm = current->mm;
>   
>  +		/* numa_policy_rebind() needs mmap_sem - don't nest */
>  +		if (!mm || !down_write_trylock(&mm->mmap_sem))
>  +			return;
>  +		up_write(&mm->mmap_sem);

What happens if the task is doing _all_ its allocation under
down_read(mmap_sem)?  Like, memset(malloc(lots))?  Does all that memory end
up in the wrong place, or what?

Something less hacky^W^W more deterministic would be nice.