Am Freitag, 14. Dezember 2018, 13:26:28 CET schrieb Gao Xiang: > Hi Richard, > > On 2018/12/14 19:25, Richard Weinberger wrote: > > This is the third place which needs this workaround. > > UBIFS, F2FS, and now iomap. > > > > I agree with Dave that nobody can assume that PG_private implies an additional > > page reference. > > But page migration does that. Including parts of the write back code. > > It seems that it's clearly documented in > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/linux/mm.h#n780 > > * A pagecache page contains an opaque `private' member, which belongs to the > * page's address_space. Usually, this is the address of a circular list of > * the page's disk buffers. PG_private must be set to tell the VM to call > * into the filesystem to release these pages. > * > * A page may belong to an inode's memory mapping. In this case, page->mapping > * is the pointer to the inode, and page->index is the file offset of the page, > * in units of PAGE_SIZE. > * > * If pagecache pages are not associated with an inode, they are said to be > * anonymous pages. These may become associated with the swapcache, and in that > * case PG_swapcache is set, and page->private is an offset into the swapcache. > * > * In either case (swapcache or inode backed), the pagecache itself holds one > * reference to the page. Setting PG_private should also increment the > * refcount. The each user mapping also has a reference to the page. > > and when I looked into that, I found > https://lore.kernel.org/lkml/3CB3CA93.D141680B@xxxxxxxxxx/ Hmm, in case of UBIFS it seems easy. We can add a get/put_page() around setting/clearing the flag. I did that now and so far none of my tests exploded. Artem, do you remember why UBIFS never raised the page counter when setting PG_private? Thanks, //richard
