Re: Module variable initialization

> On Dec 12, 2024, at 10:44 AM, Rick Macklem <rick.macklem@gmail.com> wrote:
> 
> Hi,
> 
> Bugzilla pr#282156 reports a crash that appears to be caused by
> a NFS client variable (nfscbd_pool) not being initialized when a
> NFS mount is done.
> 
> Now, the NFS client module (nfscl.ko) is weird in that it has
> two definitions for the module. There is a VFS_SET() one for
> the file system and a separate DECLARE_MODULE() for nfscl.
> (The latter exists so that the module can refuse to unload and
> define dependencies on other modules.)
> 
> The variable (nfscbd_pool) is initialized in the modevent() function
> for nfscl in the MOD_LOAD section.
> 
> Does anyone know if this can somehow result in the variable not
> being initialized when an NFS mount occurs?

I'm not familiar with NFS. From a quick look of the source code I think
`nfscbd_pool` is correctly initialized.

I do not know the exact version pr#282156, so I guess and tried 14.1-p1,
```
$ addr2line -fip -e /.zfs/snapshot/14.1-p1/usr/lib/debug/boot/kernel/kernel.debug 0xffffffff80e1c558
svc_run at /usr/src/sys/rpc/svc.c:1414
```

https://cgit.freebsd.org/src/tree/sys/rpc/svc.c?h=releng/14.1&id=0892dff104440867956a53e78c12d66090fec36b#n1414

If `nfscbd_pool` is NULL, then I expect the panic should happens earlier. Say line 1405 or event earlier line 1389 .

Maybe `svc_run_internal()` is to be blamed ?

> 
> And, if the above is possible, would doing the initialization in the
> vfs_init function for VFS_SET() be guaranteed to happen before
> a mount is done?

The order of modules seems right to me. nfscl module has order  SI_ORDER_FIRST
and VFS_SET(... nfs ... ) has SI_ORDER_MIDDLE.

> 
> Thanks for any help with this, rick
> 

Best regards,
Zhenlei