2018年11月21日(水) 13:54 Al Viro <v...@zeniv.linux.org.uk>: > > On Wed, Nov 21, 2018 at 11:43:56AM +0900, Eiichi Tsukata wrote: > > Some file systems (including ext4, xfs, ramfs ...) have the following > > problem as I've described in the commit message of the 1/4 patch. > > > > The commit ef3d0fd27e90 ("vfs: do (nearly) lockless generic_file_llseek") > > removed almost all locks in llseek() including SEEK_END. It based on the > > idea that write() updates size atomically. But in fact, write() can be > > divided into two or more parts in generic_perform_write() when pos > > straddles over the PAGE_SIZE, which results in updating size multiple > > times in one write(). It means that llseek() can see the size being > > updated during write(). > > And? Who has ever promised anything that insane? write(2) can take an > arbitrary > amount of time; another process doing lseek() on independently opened > descriptor > is *not* going to wait for that (e.g. page-in of the buffer being written, > which > just happens to be mmapped from a file on NFS over RFC1149 link).
Thanks. The lock I added in NFS was nothing but slow down lseek() because a file size is updated atomically. Even `spin_lock(&inode->i_lock)` is unnecessary. I'll fix the commit message which only refers to specific local file systems that use generic_perform_write() and remove unnecessary locks in some distributed file systems (e.g. nfs, cifs, or more) by replacing generic_file_llseek() with generic_file_llseek_unlocked() so that `tail` don't have to wait for avian carriers.