Re: 2.6.0-test6: more __init bugs

Re: 2.6.0-test6: more __init bugs

[Corey Minyard]

Robert T. Johnson wrote:

>Here are some cases where __init code or data is referenced by 
>non-__init code.
>
>Questions:
>- Is init_module allowed, required or forbidden to be __init?
>- Ditto for Scsi_Host_Template.detect()?
>- Ditto for net_device->set_config()?
>
>Thanks for looking at these potential bugs, and sorry if I've made 
>any mistakes.
>
>Best,
>Rob
>
>P.S. All these bugs were found with Cqual, the bug-finding tool
>developed by Jeff Foster, John Kodumal, and many others, and available
>at http://www.cs.umd.edu/~jfoster/cqual/, although the currently
>released version of cqual only has primitive support for 
>__init bug-finding.
>
>
>** Possible bug:
>** drivers/char/ipmi/ipmi_msghandler.c:ipmi_init_msghandler()         (__init)
>     called by numerous non-__init functions
>Note: ipmi_init_msghandler() is an alias for init_module
>Fix: declare ipmi_init_msghandler non-__init.
>
>  
>
This is not actually a bug, but it may be bad style (and thus could lead 
to a bug).  It is possible that something that uses IPMI can do some 
IPMI things before IPMI is initialized.  This can only happen during 
initialization, though.  Thus the check; once IPMI is initialized the 
function will never be called.

What's the opinion on this?  Should I just force IPMI users to 
initialize after IPMI?

Thanks,

-Corey

2.6.0-test6: more __init bugs

[Robert T. Johnson]

Here are some cases where __init code or data is referenced by 
non-__init code.

Questions:
- Is init_module allowed, required or forbidden to be __init?
- Ditto for Scsi_Host_Template.detect()?
- Ditto for net_device->set_config()?

Thanks for looking at these potential bugs, and sorry if I've made 
any mistakes.

Best,
Rob

P.S. All these bugs were found with Cqual, the bug-finding tool
developed by Jeff Foster, John Kodumal, and many others, and available
at http://www.cs.umd.edu/~jfoster/cqual/, although the currently
released version of cqual only has primitive support for 
__init bug-finding.


Linux 2.6.0-test6:

** Probably a bug:
** drivers/net/tokenring/ibmtr.c:channel_def                          (__init)
   referenced by drivers/net/tokenring/ibmtr.c:ibmtr_probe1()         (__devinit)
     called by drivers/net/tokenring/ibmtr.c:ibmtr_probe()            (__devinit)
     is stored in a dev_link_t->irq.Instance->init()
     returned by drivers/net/pcmcia/ibmtr_cs.c:ibmtr_attach()         (not __init)
Note: So it looks like ibmtr_probe() can be called any time
      a token ring pcmcia card is inserted, which may be after
      init-time
Fix: Make all this stuff non-__init when it's used for the pcmcia 
     version of the driver?

** Probably a bug?
** drivers/net/wan/sdla.c:valid_port                                  (__init)
   referenced by sdla_set_config()                                    (not __init)
Note: sdla_set_config() is stored as a net_device->set_config().
      Is such a function allowed to touch __init data?
Fix: declare valid_port as not __init.

** Possible bug:
** drivers/net/tokenring/3c359.c:xl_init()                            (__init)
     called by xl_probe()                                             (__devinit)
Fix: declare xl_init __devinit or declare xl_probe __init.
Note: xl_probe is used as a pci_driver->probe() field.

** Possible bug:
** drivers/char/ipmi/ipmi_msghandler.c:ipmi_init_msghandler()         (__init)
     called by numerous non-__init functions
Note: ipmi_init_msghandler() is an alias for init_module
Fix: declare ipmi_init_msghandler non-__init.

** Code can be declared __init
** drivers/net/acenic.c:probed                                        (__init)
   referenced by: acenic_probe()                                      (__devinit)
     only caller: ace_module_init()                                   (__init)
Fix: Make acenic_probe() __init?

** Probably not a bug?
** drivers/message/fusion/mptscsih.c:mptscsih_setup()                 (__init)
     called from drivers/message/fusion/mptscsih.c:mptscsih_detect()  (not __init)
Note: mptscsih_detect() is a Scsi_Host_Template.detect() function.
      Can detect() functions be __init?
Fix: either declare mptscsih_setup() non-__init OR
     declare mptscsih_detect() as __init

** Probably not a bug?
** drivers/scsi/qla1280.c:driver_setup                                (__init)
   referenced by qla1280_read_nvram()                                 (not __init)
     called by qla1280_initialize_adapter()                           (not __init)
       called by qla1280_do_device_init()                             (not __init)
         called by qla1280_detect()                                   (not __init)
Note: qla1280_detect is a Scsi_Host_Template->detect() routine.
Fix: make all this stuff __init?

Re: 2.6.0-test6: more __init bugs

[Robert T. Johnson]

On Tue, 2003-09-30 at 02:42, Corey Minyard wrote:
> This is not actually a bug, but it may be bad style (and thus could lead 
> to a bug).  It is possible that something that uses IPMI can do some 
> IPMI things before IPMI is initialized.  This can only happen during 
> initialization, though.  Thus the check; once IPMI is initialized the 
> function will never be called.
> 
> What's the opinion on this?  Should I just force IPMI users to 
> initialize after IPMI?

Thanks for looking at it.  Would it be reasonable to fail if a client
tries to use the ipmi interface before it is initialized?  That would be
a simple change, e.g.:

if (!initialized)
       return -ENODEV;

Best,
Rob