On Mon, Feb 04, 2019 at 01:37:00PM -0800, Hugh Dickins wrote: > On Thu, 31 Jan 2019, Andrew Morton wrote: > > On Thu, 31 Jan 2019 10:48:29 +0800 "Huang\, Ying" <ying.huang@xxxxxxxxx> wrote: > > > Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> writes: > > > > mm-swap-fix-race-between-swapoff-and-some-swap-operations.patch is very > > > > stuck so can you please redo this against mainline? > > > > I have no evidence that it has been reviewed, for a start. I've asked > > Hugh to look at it. > > I tried at the weekend. Usual story: I don't like it at all, the > ever-increasing complexity there, but certainly understand the need > for that fix, and have not managed to think up anything better - > and now I need to switch away, sorry. > > The multiple dynamically allocated and freed swapper address spaces > have indeed broken what used to make it safe. If those imaginary > address spaces did not have to be virtually contiguous, I'd say > cache them and reuse them, instead of freeing. But I don't see > how to do that as it stands. > > find_get_page(swapper_address_space(entry), swp_offset(entry)) has > become an unsafe construct, where it used to be safe against corrupted > page tables. Maybe we don't care so much about crashing on corrupted > page tables nowadays (I haven't heard recent complaints), and I think > Huang is correct that lookup_swap_cache() and __read_swap_cache_async() > happen to be the only instances that need to be guarded against swapoff > (the others are working with page table locked). > > The array of arrays of swapper spaces is all just to get a separate > lock for separate extents of the swapfile: I wonder whether Matthew has > anything in mind for that in XArray (I think Peter once got it working > in radix-tree, but the overhead not so good). Hi Hugh, thanks for putting me on the cc. I've certainly noticed what's been going on with the swapper code, but I've generally had a lack of tuits (round or otherwise) to really dig in and figure out what's going on. I've had some ideas about embedding a spinlock in each leaf node (giving one lock per 64 slots), but I know I've got about 800 things that I've actually promised to do ahead of looking at doing that. I have a suspicion that the swapper code could probably be replaced with an allocating XArray (like the IDR) and it doesn't really need to be a full on address_space, but I'm probably wrong because I haven't studied the swap code in depth.
