On Mon, Jan 16, 2023 at 11:08:09PM +0000, David Howells wrote: > IOCB_WRITE is set by aio, io_uring and cachefiles before submitting a write > operation to the VFS, but it isn't set by, say, the write() system call. > > Fix this by setting IOCB_WRITE unconditionally in call_write_iter(). Which does nothing for places that do not use call_write_iter()... __kernel_write_iter() is one such; for less obvious specimen see drivers/nvme/target/io-cmd-file.c:nvmet_file_submit_bvec() - there we have iocb coming from the caller and *not* fed to init_sync_kiocb(), so Christoph's suggestion doesn't work either. Sure, we could take care of that by adding ki_flags |= IOCB_WRITE in there, but... FWIW, call chains for ->write_iter() (as an explicit method call) are: ->write_iter() <- __kernel_write_iter() [init_sync_kiocb()] ->write_iter() <- call_write_iter() <- new_sync_write() [init_sync_kiocb()] ->write_iter() <- call_write_iter() <- do_iter_read_write() [init_sync_kiocb()] ->write_iter() <- call_write_iter() <- aio_write() [sets KIOCB_WRITE] ->write_iter() <- call_write_iter() <- io_write() [sets KIOCB_WRITE] ->write_iter() <- nvmet_file_submit_bvec() ->write_iter() <- call_write_iter() <- lo_rw_aio() ->write_iter() <- call_write_iter() <- fd_execute_rw_aio() ->write_iter() <- call_write_iter() <- vfs_iocb_iter_write() The last 4 neither set KIOCB_WRITE nor call init_sync_kiocb(). What's more, there are places that call instances (or their guts - look at btrfs_do_write_iter() callers) directly...
