On 2023-02-21 13:13, Gleb Smirnoff wrote:
On Mon, Feb 20, 2023 at 09:12:41PM +0000, Rick Macklem wrote:
R> nfsd: Add VNET_SYSUNINIT() macros for vnet cleanup
R>
R> Commit ed03776ca7f4 enabled the vnet front end macros.
R> As such, for kernels built with the VIMAGE option will malloc
R> data and initialize locks on a per-vnet basis, typically
R> via a VNET_SYSINIT().
R>
R> This patch adds VNET_SYSUNINIT() macros to do the frees
R> of the per-vnet malloc'd data and destroys of per-vnet
R> locks. It also removes the mtx_lock/mtx_unlock calls
R> from nfsrvd_cleancache(), since they are not needed.
In the netinet/netinet6/TCP/UDP we came to a style where we avoid
using IS_DEFAULT_VNET(). Instead doing global and per-vnet init
in the same function:
static globalfoo;
VNET_DEFINE_STATIC(foobar);
static void
foo_vnet_init()
{
initialize(V_foobar);
if (IS_DEFAULT_VNET(curvnet))
initialize(globalfoo);
}
VNET_SYSINIT(foo_vnet_init, ....)
We can do a separate init of global state and separate of per-VNET:
static globalfoo;
static void
foo_init()
{
initialize(globalfoo);
}
SYSINIT(foo_init, ....)
VNET_DEFINE_STATIC(foobar);
static void
foo_vnet_init()
{
initialize(V_foobar);
}
This allows to:
* guarantee that global state is initialized earlier than per-vnet
* separate all global vars from all vnet vars, and keep them together,
easier to maintain
* makes it easier to write VNET_SYSUNINIT() that is complement to
VNET_SYSINIT()
* makes it easier to write SYSUNINIT(), if module is unloadable
* sometimes global SYSINIT cab be avoided, if a static initializer is
enough
What do you guys think?
I'm all for that, just on the first point alone.
- Jamie
