On Tue, Jun 27, 2023 at 7:55 PM Ahelenia Ziemiańska <nabijaczleweli@xxxxxxxxxxxxxxxxxx> wrote: > > In 1/3 I've applied if/else if/else tree like you said, > and expounded a bit in the message. > > This is less pretty now, however, since it turns out that If my advice turns out to be bad, then please drop it. > iter_file_splice_write() already marks the out fd as written because it > writes to it via vfs_iter_write(), and that sent a double notification. > > $ git grep -F .splice_write | grep -v iter_file_splice_write > drivers/char/mem.c: .splice_write = splice_write_null, > drivers/char/virtio_console.c: .splice_write = port_fops_splice_write, > fs/fuse/dev.c: .splice_write = fuse_dev_splice_write, > fs/gfs2/file.c: .splice_write = gfs2_file_splice_write, > fs/gfs2/file.c: .splice_write = gfs2_file_splice_write, > fs/overlayfs/file.c: .splice_write = ovl_splice_write, > net/socket.c: .splice_write = generic_splice_sendpage, > scripts/coccinelle/api/stream_open.cocci: .splice_write = splice_write_f, > > Of these, splice_write_null() doesn't mark out as written > (but it's for /dev/null so I think this is expected), > and I haven't been able to visually confirm whether > port_fops_splice_write() and generic_splice_sendpage() do. > > All the others delegate to iter_file_splice_write(). > All this is very troubling to me. It translates to a mental model that I cannot remember and cannot maintain for fixes whose value are still questionable. IIUC, the only thing you need to change in do_splice() for making your use case work is to add fsnotify_modify() for the splice_pipe_to_pipe() case. Right? So either make the change that you need, or all the changes that are simple to follow without trying to make the world consistent - these pipe iterators business is really messy. I don't know if avoiding a double event (which is likely not visible) is worth the complicated code that is hard to understand. > In 2/3 I fixed the vmsplice notification placement > (access from pipe, modify to pipe). > > I'm following this up with an LTP patch, where only sendfile_file_to_pipe > passes on 6.1.27-1 and all tests pass on v6.4 + this patchset. > Were these tests able to detect the double event? Maybe it's not visible because double consequent events get merged. > Ahelenia Ziemiańska (3): > splice: always fsnotify_access(in), fsnotify_modify(out) on success > splice: fsnotify_access(fd)/fsnotify_modify(fd) in vmsplice > splice: fsnotify_access(in), fsnotify_modify(out) on success in tee > > fs/splice.c | 43 +++++++++++++++++++++++++------------------ > 1 file changed, 25 insertions(+), 18 deletions(-) > > > Interdiff against v2: > diff --git a/fs/splice.c b/fs/splice.c > index 3234aaa6e957..0427f0a91c7d 100644 > --- a/fs/splice.c > +++ b/fs/splice.c > @@ -1155,10 +1155,7 @@ long do_splice(struct file *in, loff_t *off_in, struct file *out, > flags |= SPLICE_F_NONBLOCK; > > ret = splice_pipe_to_pipe(ipipe, opipe, len, flags); > - goto notify; > - } > - > - if (ipipe) { > + } else if (ipipe) { > if (off_in) > return -ESPIPE; > if (off_out) { > @@ -1188,10 +1185,10 @@ long do_splice(struct file *in, loff_t *off_in, struct file *out, > else > *off_out = offset; > > - goto notify; > - } > - > - if (opipe) { > + // ->splice_write already marked out > + // as modified via vfs_iter_write() > + goto noaccessout; That's too ugly IMO. Are you claiming that the code in master is wrong? Because in master there is fsnotify_modify(out) for (ipipe) case. > + } else if (opipe) { > if (off_out) > return -ESPIPE; > if (off_in) { > @@ -1211,17 +1208,14 @@ long do_splice(struct file *in, loff_t *off_in, struct file *out, > in->f_pos = offset; > else > *off_in = offset; > + } else > + return -EINVAL; > > - goto notify; > - } > - > - return -EINVAL; > - > -notify: > - if (ret > 0) { > - fsnotify_access(in); > + if (ret > 0) > fsnotify_modify(out); > - } > +noaccessout: > + if (ret > 0) > + fsnotify_access(in); > Not to mention that it should be nomodifyout, but I dislike this "common" code that it not common at all, so either just handle the pipe_to_pipe case to fix your use case or leave this code completely common ignoring the possible double events. Thanks, Amir.