Re: [f2fs-dev] [PATCH v2] f2fs: add F2FS_IOC_SEC_TRIM_FILE ioctl

[Eric Biggers]

On Thu, Jun 11, 2020 at 08:04:06PM +0900, Daeho Jeong wrote:
> > > +static int f2fs_sec_trim_file(struct file *filp, unsigned long arg)
> > > +{
> > > +     struct inode *inode = file_inode(filp);
> > > +     struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
> > > +     struct address_space *mapping = inode->i_mapping;
> > > +     struct block_device *prev_bdev = NULL;
> > > +     pgoff_t index, pg_start = 0, pg_end;
> > > +     block_t prev_block = 0, len = 0;
> > > +     u32 flags;
> > > +     int ret = 0;
> > > +
> > > +     if (!(filp->f_mode & FMODE_WRITE))
> > > +             return -EBADF;
> > > +
> > > +     if (get_user(flags, (u32 __user *)arg))
> > > +             return -EFAULT;
> > > +     if (flags == 0 || (flags & ~F2FS_TRIM_FILE_MASK))
> > > +             return -EINVAL;
> > > +
> > > +     if ((flags & F2FS_TRIM_FILE_DISCARD) && !f2fs_hw_support_discard(sbi))
> > > +             return -EOPNOTSUPP;
> > > +
> > > +     file_start_write(filp);
> >
> > Now, I'm a little confused about when we need to call __mnt_want_write_file(),
> > you know, vfs_write() still will call this function when updating time.
> > - __generic_file_write_iter
> >  - file_update_time
> >   - __mnt_want_write_file
> >
> > And previously, f2fs ioctl uses mnt_{want,drop}_write_file() whenever there is
> > any updates on fs/file, if Eric is correct, we need to clean up most of ioctl
> > interface as well.
> 
> I also saw most filesytem codes use just mnt_{want,drop}_write_file()
> and actually it doesn't affect code working. It's a matter of doing a
> redundant job or not.
> AFAIUI, if the file is not open for writing (FMODE_WRITE), we have to
> call mnt_want_write_file() to increase mnt_writers.
> In this case, we already checked it has FMODE_WRITE flag.

If the fd isn't writable (or may not be writable), mnt_want_write_file() is
needed.  That includes all ioctls that operate (or may operate) on directories,
since directories can't be opened for writing.

But when the fd is guaranteed to be writable, incrementing mnt_writers is
pointless.  I'm trying to clean this up in the VFS:
https://lkml.kernel.org/r/20200611160534.55042-1-ebiggers@xxxxxxxxxxx

mnt_want_write_file() still does the freeze protection, which file_start_write()
achieves more directly.

The only other thing that mnt_want_write_file() does is the check for emergency
remount r/o, which I doubt is very important.  It's racy, so the filesystem
needs to detect it in other places too.

I'm not sure why file_update_time() uses __mnt_want_write_file().  Either it
assumes the fd might not be writable, or it just wants the check for emergency
remount r/o, or it's just a mistake.  Note also that mtime isn't always updated,
so just because file_update_time() calls __mnt_want_write_file() doesn't mean
that write() always calls __mnt_want_write_file().

- Eric