Re: [PATCH v5 3/9] fs: add RWF_ENCODED for reading/writing compressed data

[Omar Sandoval <osandov@osandov.com>]

On Fri, Aug 21, 2020 at 11:47:54AM +0300, Amir Goldstein wrote:
> On Fri, Aug 21, 2020 at 10:38 AM Omar Sandoval <osandov@osandov.com> wrote:
> >
> > From: Omar Sandoval <osandov@fb.com>
> >
> > Btrfs supports transparent compression: data written by the user can be
> > compressed when written to disk and decompressed when read back.
> > However, we'd like to add an interface to write pre-compressed data
> > directly to the filesystem, and the matching interface to read
> > compressed data without decompressing it. This adds support for
> > so-called "encoded I/O" via preadv2() and pwritev2().
> >
> > A new RWF_ENCODED flags indicates that a read or write is "encoded". If
> > this flag is set, iov[0].iov_base points to a struct encoded_iov which
> > is used for metadata: namely, the compression algorithm, unencoded
> > (i.e., decompressed) length, and what subrange of the unencoded data
> > should be used (needed for truncated or hole-punched extents and when
> > reading in the middle of an extent). For reads, the filesystem returns
> > this information; for writes, the caller provides it to the filesystem.
> > iov[0].iov_len must be set to sizeof(struct encoded_iov), which can be
> > used to extend the interface in the future a la copy_struct_from_user().
> > The remaining iovecs contain the encoded extent.
> >
> > This adds the VFS helpers for supporting encoded I/O and documentation
> > for filesystem support.
> >
> > Signed-off-by: Omar Sandoval <osandov@fb.com>
> > ---
> >  Documentation/filesystems/encoded_io.rst |  74 ++++++++++
> >  Documentation/filesystems/index.rst      |   1 +
> >  include/linux/fs.h                       |  16 +++
> >  include/uapi/linux/fs.h                  |  33 ++++-
> >  mm/filemap.c                             | 166 +++++++++++++++++++++--
> >  5 files changed, 276 insertions(+), 14 deletions(-)
> >  create mode 100644 Documentation/filesystems/encoded_io.rst
> >
> > diff --git a/Documentation/filesystems/encoded_io.rst b/Documentation/filesystems/encoded_io.rst
> > new file mode 100644
> > index 000000000000..50405276d866
> > --- /dev/null
> > +++ b/Documentation/filesystems/encoded_io.rst
> > @@ -0,0 +1,74 @@
> > +===========
> > +Encoded I/O
> > +===========
> > +
> > +Encoded I/O is a mechanism for reading and writing encoded (e.g., compressed
> > +and/or encrypted) data directly from/to the filesystem. The userspace interface
> > +is thoroughly described in the :manpage:`encoded_io(7)` man page; this document
> > +describes the requirements for filesystem support.
> > +
> > +First of all, a filesystem supporting encoded I/O must indicate this by setting
> > +the ``FMODE_ENCODED_IO`` flag in its ``file_open`` file operation::
> > +
> > +    static int foo_file_open(struct inode *inode, struct file *filp)
> > +    {
> > +            ...
> > +            filep->f_mode |= FMODE_ENCODED_IO;
> > +            ...
> > +    }
> > +
> > +Encoded I/O goes through ``read_iter`` and ``write_iter``, designated by the
> > +``IOCB_ENCODED`` flag in ``kiocb->ki_flags``.
> > +
> > +Reads
> > +=====
> > +
> > +Encoded ``read_iter`` should:
> > +
> > +1. Call ``generic_encoded_read_checks()`` to validate the file and buffers
> > +   provided by userspace.
> > +2. Initialize the ``encoded_iov`` appropriately.
> > +3. Copy it to the user with ``copy_encoded_iov_to_iter()``.
> > +4. Copy the encoded data to the user.
> > +5. Advance ``kiocb->ki_pos`` by ``encoded_iov->len``.
> > +6. Return the size of the encoded data read, not including the ``encoded_iov``.
> > +
> > +There are a few details to be aware of:
> > +
> > +* Encoded ``read_iter`` should support reading unencoded data if the extent is
> > +  not encoded.
> > +* If the buffers provided by the user are not large enough to contain an entire
> > +  encoded extent, then ``read_iter`` should return ``-ENOBUFS``. This is to
> > +  avoid confusing userspace with truncated data that cannot be properly
> > +  decoded.
> > +* Reads in the middle of an encoded extent can be returned by setting
> > +  ``encoded_iov->unencoded_offset`` to non-zero.
> > +* Truncated unencoded data (e.g., because the file does not end on a block
> > +  boundary) may be returned by setting ``encoded_iov->len`` to a value smaller
> > +  value than ``encoded_iov->unencoded_len - encoded_iov->unencoded_offset``.
> > +
> > +Writes
> > +======
> > +
> > +Encoded ``write_iter`` should (in addition to the usual accounting/checks done
> > +by ``write_iter``):
> > +
> > +1. Call ``copy_encoded_iov_from_iter()`` to get and validate the
> > +   ``encoded_iov``.
> > +2. Call ``generic_encoded_write_checks()`` instead of
> > +   ``generic_write_checks()``.
> > +3. Check that the provided encoding in ``encoded_iov`` is supported.
> > +4. Advance ``kiocb->ki_pos`` by ``encoded_iov->len``.
> > +5. Return the size of the encoded data written.
> > +
> > +Again, there are a few details:
> > +
> > +* Encoded ``write_iter`` doesn't need to support writing unencoded data.
> > +* ``write_iter`` should either write all of the encoded data or none of it; it
> > +  must not do partial writes.
> > +* ``write_iter`` doesn't need to validate the encoded data; a subsequent read
> > +  may return, e.g., ``-EIO`` if the data is not valid.
> > +* The user may lie about the unencoded size of the data; a subsequent read
> > +  should truncate or zero-extend the unencoded data rather than returning an
> > +  error.
> > +* Be careful of page cache coherency.
> 
> Haha that rings in my head like the "Smoking kills!" warnings...
> 
> I find it a bit odd that you mix page cache at all when reading
> unencoded extents.
> Feels like a file with FMODE_ENCODED_IO should stick to direct IO in all cases.
> I don't know how btrfs deals with mixing direct IO and page cache IO normally,
> but surely the rules could be made even stricter for an inode accessed with this
> new API?
> 
> Is there something I am misunderstanding?
> 
> Thanks,
> Amir.

I'm not completely following here, are you suggesting that if a file is
open with O_ALLOW_ENCODED, buffered I/O to that file should return an
error? Btrfs at least does the necessary range locking and page cache
invalidation to ensure that direct I/O gets along with buffered I/O (and
now encoded I/O).