Messages in this thread | | | Date | Tue, 21 Jan 2014 23:00:44 -0800 (PST) | From | Hugh Dickins <> | Subject | Re: [PATCH] swap: do not skip lowest_bit in scan_swap_map() scan loop |
| |
On Tue, 21 Jan 2014, Jamie Liu wrote:
> In the second half of scan_swap_map()'s scan loop, offset is set to > si->lowest_bit and then incremented before entering the loop for the > first time, causing si->swap_map[si->lowest_bit] to be skipped. > > Signed-off-by: Jamie Liu <jamieliu@google.com>
Acked-by: Hugh Dickins <hughd@google.com>
Good catch. At first I was puzzled that this off-by-one could have gone unnoticed for so long (ever since 2.6.29); but now I think that almost always we have a good amount of slack, in those pages duplicated between swap and swapcache, which can be reclaimed at the vm_swap_full() check, and so conceal this loss of a single slot.
> --- > mm/swapfile.c | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/mm/swapfile.c b/mm/swapfile.c > index 612a7c9..6635081 100644 > --- a/mm/swapfile.c > +++ b/mm/swapfile.c > @@ -616,7 +616,7 @@ scan: > } > } > offset = si->lowest_bit; > - while (++offset < scan_base) { > + while (offset < scan_base) { > if (!si->swap_map[offset]) { > spin_lock(&si->lock); > goto checks; > @@ -629,6 +629,7 @@ scan: > cond_resched(); > latency_ration = LATENCY_LIMIT; > } > + offset++; > } > spin_lock(&si->lock); > > -- > 1.8.5.3 > >
| |