Re: [Qemu-devel] [PATCH v2 0/2] Dynamic module support for block drivers

[Marc Marí]

On Mon, 21 Sep 2015 16:43:27 +0300
"Denis V. Lunev" <address@hidden> wrote:

> On 09/08/2015 04:53 PM, Marc Marí wrote:
> > The current module infrastructure has been improved to enable
> > dynamic module loading.
> >
> > This reduces the load time for very simple guests. For the following
> > configuration (very loaded)
> >
> > ./configure --enable-sdl --enable-gtk --enable-vte --enable-curses \
> >      --enable-vnc --enable-vnc-{jpeg,tls,sasl,png} --enable-virtfs \
> >      --enable-brlapi --enable-curl --enable-fdt --enable-bluez \
> >      --enable-kvm --enable-rdma --enable-uuid --enable-vde \
> >      --enable-linux-aio --enable-cap-ng --enable-attr
> > --enable-vhost-net \ --enable-vhost-scsi --enable-spice
> > --enable-rbd --enable-libiscsi \ --enable-smartcard-nss
> > --enable-guest-agent --enable-libusb \ --enable-usb-redir
> > --enable-lzo --enable-snappy --enable-bzip2 \ --enable-seccomp
> > --enable-coroutine-pool --enable-glusterfs \ --enable-tpm
> > --enable-libssh2 --enable-vhdx --enable-numa \ --enable-tcmalloc
> > --target-list=x86_64-softmmu
> >
> > With modules disabled, there are 142 libraries loaded at startup.
> > Time is the following:
> >   LD time: 0.065 seconds
> >   QEMU time: 0.02 seconds
> >   Total time: 0.085 seconds
> >
> > With this patch series and modules enabled, there are 128 libraries
> > loaded at startup. Time is the following:
> >   LD time: 0.02 seconds
> >   QEMU time: 0.02 seconds
> >   Total time: 0.04 seconds
> >
> > Where LD time is the time between the program startup and the jump
> > to main, and QEMU time is the time between the start of main and
> > the first kvm_entry.
> >
> > These results are just with a few block drivers, that were already
> > a module. Adding more modules (block or not block) should be easy,
> > and will reduce the load time even more.
> >
> > Marc Marí (2):
> >    Add dynamic module loading for block drivers
> >    Add dynamic generation of module_block.h
> >
> >   .gitignore                      |   1 +
> >   Makefile                        |  10 ++-
> >   block.c                         |  70 +++++++++++++++++++++
> >   configure                       |   2 +-
> >   include/qemu/module.h           |   3 +
> >   scripts/modules/module_block.py | 134
> > ++++++++++++++++++++++++++++++++++++++++
> > util/module.c                   |  38 ++++-------- 7 files changed,
> > 227 insertions(+), 31 deletions(-) create mode 100755
> > scripts/modules/module_block.py
> >
> 
>  From my point of view the design looks a bit complex.
> The approach should be quite similar to one used
> in Linux kernel.
> 
> If the block driver is configured as a module, block_init
> should register proper hooks and create generic lists.
> C code parsing does not look like a good approach
> to me.
> 
> If block_init is a bad name, we could use something like
> module_init.
> 

I think applying the kind of modules of the Linux kernel here would
mean reworking an important part of the drivers that want to be
converted into modules.

If that's the case, I think it's better to have a less good solution
that is still good enough, and doesn't mean reworking half of the QEMU
codebase. Correct me if I'm wrong.

But of course, this is not the only valid approach.

Thanks
Marc