On 4 Feb 2019, at 14:20, Benjamin Coddington wrote: > On 4 Feb 2019, at 14:07, Benjamin Coddington wrote: > >> On 4 Feb 2019, at 14:03, Benjamin Coddington wrote: >> >>> On 4 Feb 2019, at 13:32, Trond Myklebust wrote: >>> >>>> On Mon, 2019-02-04 at 13:13 -0500, Benjamin Coddington wrote: >>>>> If nfs_page_async_flush() removes the page from the mapping, then we >>>>> can't use page_file_mapping() on it as nfs_updatepate() is wont to do >>>>> when receiving an error. Instead, simplify nfs_zap_mapping() to take >>>>> the inode. >>>>> >>>> >>>> Won't this break NFS swap? >>> >>> I guess it may, but then I wonder: is that a thing that anyone does? It >>> sounds like a terrible idea.. >>> >>> I'm curious enough though, and doing some research on it. >> >> There's enough to commits in the tree fixing problem or supporting swap >> on NFS, so I suppose I've just missed all that. > > So for diskless systems, ah. It makes sense. > > But does it make any sense to set NFS_INO_INVALID_DATA on a swap file? > Does anyone share them? > > Dazed and confused, but trying to continue.. So if the page is backing a swap file, nfs_zap_mapping() will fail to set NFS_INO_INVALID_DATA because i_mapping.nrpages will always be 0. But it looks like NFS has already decided to not care if the mapping needs revalidation for swap, since we assume /shared/ swap files are insane. So without testing swap on NFS at all, it seems safe. This is all new to me, though, so I need experienced revalidation of this reasoning myself. Perhaps it needs a comment to explain why it is safe to do.. any thoughts? Ben
