Re: [Qemu-block] [PATCH v2 0/7] Refactor DMG driver to have chunk size independence

[Stefan Hajnoczi]

On Wed, Aug 30, 2017 at 06:32:52PM +0530, Ashijeet Acharya wrote:
> On Tue, Aug 29, 2017 at 8:55 PM, Stefan Hajnoczi <address@hidden> wrote:
> 
> > On Sun, Aug 20, 2017 at 1:47 PM, Ashijeet Acharya
> > <address@hidden> wrote:
> > > On Fri, May 5, 2017 at 7:29 PM, Stefan Hajnoczi <address@hidden>
> > wrote:
> > >>
> > >> On Thu, Apr 27, 2017 at 01:36:30PM +0530, Ashijeet Acharya wrote:
> > >> > This series helps to provide chunk size independence for DMG driver to
> > >> > prevent
> > >> > denial-of-service in cases where untrusted files are being accessed by
> > >> > the user.
> > >>
> > >> The core of the chunk size dependence problem are these lines:
> > >>
> > >>     s->compressed_chunk = qemu_try_blockalign(bs->file->bs,
> > >>                                               ds.max_compressed_size +
> > 1);
> > >>     s->uncompressed_chunk = qemu_try_blockalign(bs->file->bs,
> > >>                                                 512 *
> > >> ds.max_sectors_per_chunk);
> > >>
> > >> The refactoring needs to eliminate these buffers because their size is
> > >> controlled by the untrusted input file.
> > >
> > >
> > > Oh okay, I understand now. But wouldn't I still need to allocate some
> > memory
> > > for these buffers to be able to use them for the compressed chunks case
> > you
> > > mentioned below. Instead of letting the DMG images control the size of
> > these
> > > buffers, maybe I can hard-code the size of these buffers instead?
> > >
> > >>
> > >>
> > >> After applying your patches these lines remain unchanged and we still
> > >> cannot use input files that have a 250 MB chunk size, for example.  So
> > >> I'm not sure how this series is supposed to work.
> > >>
> > >> Here is the approach I would take:
> > >>
> > >> In order to achieve this dmg_read_chunk() needs to be scrapped.  It is
> > >> designed to read a full chunk.  The new model does not read full chunks
> > >> anymore.
> > >>
> > >> Uncompressed reads or zeroes should operate directly on qiov, not
> > >> s->uncompressed_chunk.  This code will be dropped:
> > >>
> > >>     data = s->uncompressed_chunk + sector_offset_in_chunk * 512;
> > >>     qemu_iovec_from_buf(qiov, i * 512, data, 512);
> > >
> > >
> > > I have never worked with qiov before, are there any places where I can
> > refer
> > > to inside other drivers to get the idea of how to use it directly (I am
> > > searching by myself in the meantime...)?
> >
> > A QEMUIOVector is a utility type for struct iovec iov[] processing.
> > See util/iov.c.  This is called "vectored" or "scatter-gather" I/O.
> >
> > Instead of transferring data to/from a single <buffer, length> tuple,
> > they take an array [<buffer, length>].  For example, the buffer "Hello
> > world" could be split into two elements:
> > [{"Hello ", strlen("Hello ")},
> >  {"world", strlen("world")}]
> >
> > Vectored I/O is often used because it eliminates memory copies.  Say
> > you have a network packet header struct and also a data payload array.
> > Traditionally you would have to allocate a new buffer large enough for
> > both header and payload, copy the header and payload into the buffer,
> > and finally give this temporary buffer to the I/O function.  This is
> > inefficient.  With vectored I/O you can create a vector with two
> > elements, the header and the payload, and the I/O function can process
> > them without needing a temporary buffer copy.
> >
> 
> Thanks for the detailed explanation, I think I understood the concept now
> and how to use qiov efficiently.
> Correct me if I am wrong here. In order to use qiov directly for
> uncompressed chunks:
> 
> 1. Declare a new local_qiov inside dmg_co_preadv (let's say)

No, the operation should use qiov directly if the chunk is uncompressed.

A temporary buffer is only needed if the data needs to be uncompressed.

> 2. Initialize it with qemu_iovec_init()
> 3. Reset it with qemu_iovec_reset() (this is because we will perform this
> action in a loop and thus need to reset it before every loop?)
> 4. Declare a buffer "uncompressed_buf" and allocate it with
> qemu_try_blockalign()
> 5. Add this buffer to our local_qiov using qemu_iovec_add()
> 6. Read data from file directly into local_qiov using bdrv_co_preadv()
> 7. On success concatenate it with the qiov passed into the main
> dmg_co_preadv() function.
> 
> I think this method only works for uncompressed chunks. For the compressed
> ones, I believe we will still need to do it in the existing way, i.e. read
> chunk from file -> decompress into output buffer -> use
> qemu_iovec_from_buf() because we cannot read directly since data is in
> compressed form. Makes sense?
> 
> 
> > > I got clearly what you are trying
> > > to say, but don't know how to implement it. I think, don't we already do
> > > that for the zeroed chunks in DMG in dmg_co_preadv()?
> >
> > Yes, dmg_co_preadv() directly zeroes the qiov.  It doesn't use
> > s->compressed_chunk/s->uncompressed_chunk.
> >
> > >
> > >>
> > >>
> > >> Compressed reads still buffers.  I suggest the following buffers:
> > >>
> > >> 1. compressed_buf - compressed data is read into this buffer from file
> > >> 2. uncompressed_buf - a place to discard decompressed data while
> > >>                       simulating a seek operation
> > >
> > >
> > > Yes, these are the buffers whose size I can hard-code as discussed above?
> > > You can suggest the preferred size to me.
> >
> > Try starting with 256 KB for both buffers.
> >
> 
> Okay, I will do that. But I think we cannot use these buffer sizes for bz2
> chunks (see below)
> 
> 
> > >> Data is read from compressed chunks by reading a reasonable amount
> > >> (64k?) into compressed_buf.  If the user wishes to read at an offset
> > >> into this chunk then a loop decompresses data we are seeking over into
> > >> uncompressed_buf (and refills compressed_buf if it becomes empty) until
> > >> the desired offset is reached.  Then decompression can continue
> > >> directly into the user's qiov and uncompressed_buf isn't used to
> > >> decompress the data requested by the user.
> > >
> > >
> > > Yes, this series does exactly that but keeps using the "uncompressed"
> > buffer
> > > once we reach the desired offset. Once, I understand to use qiov
> > directly,
> > > we can do this. Also, Kevin did suggest me (as I remember vaguely) that
> > in
> > > reality we never actually get the read request at a particular offset
> > > because DMG driver is generally used with "qemu-img convert", which means
> > > all read requests are from the top.
> >
> > For performance it's fine to assume a sequential access pattern.  The
> > block driver should still support random access I/O though.
> >
> 
> Yes, I agree. But I don't think we can add random access for the bz2 chunks
> because they need to be decompressed as a whole and not in parts. I tried
> to explain it in my cover letter as an important note (
> https://lists.gnu.org/archive/html/qemu-devel/2017-04/msg05327.html).

Here is what you said:

  "bz2 compressed streams do not allow random access midway through a
  chunk/block as the BZ2_bzDecompress() API in bzlib seeks for the
  magic key "BZh" before starting decompression.[2] This magic key is
  present at the start of every chunk/block only and since our cached
  random access points need not necessarily point to the start of a
  chunk/block, BZ2_bzDecompress() fails with an error value
  BZ_DATA_ERROR_MAGIC"

The block driver can simulate random access.  Take a look at the
BZ2_bzDecompress() API docs:

  "You may provide and remove as little or as much data as you like on
  each call of BZ2_bzDecompress. In the limit, it is acceptable to
  supply and remove data one byte at a time, although this would be
  terribly inefficient. You should always ensure that at least one byte
  of output space is available at each call.

In other words, bz2 supports streaming.  Therefore you can use the
buffer sizes I suggested in a loop instead of reading the whole bz2
block upfront.

If you keep the bz_stream between dmg_uncompress_bz2_do() calls then
sequential reads can be optimized.  They do not need to reread from the
beginning of the bz2 block.  That's important because sequential I/O is
the main access pattern expected by this block driver.

Stefan