Re: [PATCH 2/3] uprobes: change build_map_info() to try kmalloc(GFP_NOWAIT) first

[Peter Zijlstra <peterz@infradead.org>]

On Mon, 2012-06-04 at 16:53 +0200, Oleg Nesterov wrote:
> build_map_info() doesn't allocate the memory under i_mmap_mutex
> to avoid the deadlock with page reclaim. But it can try GFP_NOWAIT
> first, it should work in the likely case and thus we almost never
> need the pre-alloc-and-retry path.
> 
> Signed-off-by: Oleg Nesterov <oleg@redhat.com>
> ---
>  kernel/events/uprobes.c |    9 +++++++--
>  1 files changed, 7 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel/events/uprobes.c b/kernel/events/uprobes.c
> index a1bbcab..c9836ae 100644
> --- a/kernel/events/uprobes.c
> +++ b/kernel/events/uprobes.c
> @@ -772,8 +772,13 @@ build_map_info(struct address_space *mapping, loff_t offset, bool is_register)
>  			continue;
>  
>  		if (!prev) {
> -			more++;
> -			continue;


Should we add something like this:

			/*
			 * Needs GFP_NOWAIT to avoid i_mmap_mutex recursion through reclaim.
			 * This is optimistic, no harm done if it fails.
			 */


> +			prev = kmalloc(sizeof(struct map_info),
> +					GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
> +			if (!prev) {
> +				more++;
> +				continue;
> +			}
> +			prev->next = NULL;
>  		}
>  
>  		if (!atomic_inc_not_zero(&vma->vm_mm->mm_users))