On Mon 19-05-14 13:44:25, David Herrmann wrote: > Hi > > On Thu, May 15, 2014 at 12:35 AM, Hugh Dickins <hughd@xxxxxxxxxx> wrote: > > The aspect which really worries me is this: the maintenance burden. > > This approach would add some peculiar new code, introducing a rare > > special case: which we might get right today, but will very easily > > forget tomorrow when making some other changes to mm. If we compile > > a list of danger areas in mm, this would surely belong on that list. > > I tried doing the page-replacement in the last 4 days, but honestly, > it's far more complex than I thought. So if no-one more experienced > with mm/ comes up with a simple implementation, I'll have to delay > this for some more weeks. > > However, I still wonder why we try to fix this as part of this > patchset. Using FUSE, a DIRECT-IO call can be delayed for an arbitrary > amount of time. Same is true for network block-devices, NFS, iscsi, > maybe loop-devices, ... This means, _any_ once mapped page can be > written to after an arbitrary delay. This can break any feature that > makes FS objects read-only (remounting read-only, setting S_IMMUTABLE, > sealing, ..). > > Shouldn't we try to fix the _cause_ of this? > > Isn't there a simple way to lock/mark/.. affected vmas in > get_user_pages(_fast)() and release them once done? We could increase > i_mmap_writable on all affected address_space and decrease it on > release. This would at least prevent sealing and could be check on > other operations, too (like setting S_IMMUTABLE). > This should be as easy as checking page_mapping(page) != NULL and then > adjusting ->i_mmap_writable in > get_writable_user_pages/put_writable_user_pages, right? Doing this would be quite a bit of work. Currently references returned by get_user_pages() are page references like any other and thus are released by put_page() or similar. Now you would make them special and they need special releasing and there are lots of places in kernel where get_user_pages() is used that would need changing. Another aspect is that it could have performance implications - if there are several processes using get_user_pages[_fast]() on a file, they would start contending on modifying i_mmap_writeable. One somewhat crazy idea I have is that maybe we could delay unmapping of a page if this was last VMA referencing it until all extra page references of pages in there are dropped. That would make i_mmap_writeable reliable for you and it would also close those races with remount. Hugh, do you think this might be viable? Honza -- Jan Kara <jack@xxxxxxx> SUSE Labs, CR -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
