On 6/17/21 7:37 AM, Gerd Hoffmann wrote: > Hi, > >>>> However, please correct me if I'm wrong, I understand that an accelerator >>>> as a >>>> module will add an overhead that some user won't be willing to pay. So, >>>> give >>>> them the option to have built-in accelerators seems a good idea. >>> >>> Modules add some overhead, yes, and there are surely use-cases where you >> >> Where do we expect the overhead to be, and of which nature? > > The dynamic linking needed then when loading the module adds a bit of > overhead (compared to static linked code). Increases qemu start time > a bit. > > On the other hand the start overhead can be reduced by modules, > specifically for the case that a module depends on shared libraries and > is *not* needed. With for example gtk being modular the gtk shared > libraries (plus indirect dependencies like pango, cairo etc) are only > loaded when you actually use gtk, whereas a non-modular build would load > them no matter what.
Interesting observation. > > The code reorganization needed for modularization can add some overhead > too, when using function pointers instead of direct calls for example > (see QemuSpiceOps). That overhead doesn't go away when you do a > non-modular build though ... > > take care, > Gerd > Do we need to be able to unload modules that we previously loaded? Or is this not a realistic requirement? Thanks, Claudio
