Re: [Qemu-devel] [PATCH 0/3] Dynamic module loading for block drivers

[Fam Zheng <famz@redhat.com>]

On Wed, 06/22 17:35, Colin Lord wrote:
> This is v2 of the series I sent out last week. These are the changes I
> made based on the feedback I got:
> - Fixed typo and Marc's email address in the python script
> - Moved registration of iscsi_opts into vl.c
> 
> What I didn't do:
> - Remove copy-pasted loops
> 
> There was a bit of discussion about how to remove the need for the copy-
> paste loops that are in block.c. I attempted to solve it by using
> g_module_sym to load the BlockDriver struct directly at the time the
> module gets loaded and returning it so that the loops were not
> necessary. I accomplished this by adding a field to the struct which,
> for a given format/protocol configuration, had the name of the
> corresponding BlockDriver struct. Having the name allowed me to load the
> symbol right out of the loaded module. However, it turns out that, at
> least as far as I can tell, g_module_sym can't load the BlockDriver
> structs in this way because they are declared static.
> 
> I tested the attempt to remove the copy-pasted loops by using the
> qemu-iotests on it with ssh (which is modularized). The errors I got
> were along the lines of:
> 
> can't open device ssh://[address removed]: Unknown protocol 'ssh'
> Failed to find driver in module
> 
> To test my theory (I haven't had much luck finding reliable
> documentation about this) that it was because they were static, I
> changed the definition of the bdrv_ssh BlockDriver to not be static.
> Unfortunately I still got errors, but I believe the drivers got loaded.
> The errors were not the same, rather these ones were complaining about
> the host key not matching the one in known_hosts. I've had this issue
> while trying to set up ssh with qemu in the past, so I'm not quite as
> worried about it (although I'd love to hear a fix), and more importantly
> there aren't any messages about the driver not being found.

Maybe it's not a big deal to export the symbols.

For testing you can also test with iscsi.

Fam

> 
> That hopefully explains some of the issues, and why I'm submitting this
> with the duplicated loops still intact. If there are other ideas for
> dealing with this my ears are open, but this is what I have for now.
> 
> Colin Lord (1):
>   blockdev: prepare iSCSI block driver for dynamic loading
> 
> Marc Mari (2):
>   blockdev: Add dynamic generation of module_block.h
>   blockdev: Add dynamic module loading for block drivers
> 
>  .gitignore                      |   1 +
>  Makefile                        |  11 +++-
>  block.c                         |  86 +++++++++++++++++++++++---
>  block/iscsi.c                   |  36 -----------
>  include/qemu/module.h           |   3 +
>  scripts/modules/module_block.py | 134 ++++++++++++++++++++++++++++++++++++++++
>  util/module.c                   |  37 +++--------
>  vl.c                            |  36 +++++++++++
>  8 files changed, 269 insertions(+), 75 deletions(-)
>  create mode 100644 scripts/modules/module_block.py
> 
> -- 
> 2.5.5
> 
>