reproducible panic in netisr

[Rick Macklem]

On Thu, 6 Aug 2009, Julian Elischer wrote:

> Rick Macklem wrote:
>> 
>> 
>> On Thu, 6 Aug 2009, Robert Watson wrote:
>> 
>>> other places where we have very strong alignment requirements on 
>>> i386/amd64, such as the td_ucred pointer that we check for change on 
>>> system calls/traps to see if we need to refresh the thread's credential 
>>> from the process credential.
>>> 
>> Does this imply that the krpc/nlm/nfs hack of:
>>     oldcred = td->td_ucred;
>>     td->td_ucred = "some other cred ptr, such as the mount one"
>>     ...
>>     td->td_ucred = oldcred;
>> 
>> could be dangerous?
>> 
>> Maybe it should be converted to code that replaces the contents instead
>> of replacing the *cred? (Variants of the above live in a bunch of places
>> in the krpc, nlm and nfs code, due to the fact that the socket functions
>> use td->td_ucred in various places.)
>
> no, creds are read-only .. you never change a cred.
> You alwasy make a new one ans use it, becasue you may be shareing your cred 
> with hundreds of other sibling threads or processes. (they are refcounted)
>
Righto, yes. So does that imply that the alignment provided by crget()
{ which uses malloc() } is sufficient for td->td_ucred or is td->td_ucred
a special case?

rick
ps: The above hack, which came up in a separate discussion yesterday,
     isn't gonna be easy to get rid of, imho. A whole bunch of network
     related functions use td->td_ucred and the only fix I can see would
     be to add "*cred" arguments to them all, so that the krpc/nlm/nfs
     code could pass the correct *cred in. (It is set to the one used at
     mount time for network reconnects, etc.)