svn commit: r276008 - in head/sys: kern sys

[Andriy Gapon]

On 23/12/2014 09:55, Rui Paulo wrote:
> On Dec 22, 2014, at 23:03, Andriy Gapon <avg at FreeBSD.org> wrote:
>>
>> On 23/12/2014 04:39, Rui Paulo wrote:
>>> On Dec 22, 2014, at 11:17, John Baldwin <jhb at freebsd.org> wrote:
>>>>
>>>> On Monday, December 22, 2014 1:29:38 pm Rui Paulo wrote:
>>>>> On Dec 22, 2014, at 06:40, John Baldwin <jhb at freebsd.org> wrote:
>>>>>> Is there something specific to core dumps that makes vn_fullpath() more
>>>>>> useful to have working before a process tries to open the core?  (As
>>>>>> compared to other newly-created files)
>>>>>
>>>>> Yes: the ability to provide the full path to userland when a core dump file 
>>>> is generated.
>>>>
>>>> Can you be more specific?  Are we printing the path on the console after
>>>> destroying the generated path?  Is it being written into a note in the core
>>>> itself (but only having the vnode of the core file available and not the 
>>>> generated path)?
>>>
>>> No.  I have some code that calls devctl_notify() when a core dump is generated which is useful for running an automated debugging session.  We use this at work and I'll see if I can upstream it.  What Konstantin fixed was the generation of the cache entry in the corefile_open() routine.  This lets me call vn_fullpath() after vn_close() with a high probability that it will work whereas, in the past, it was never in the cache, so vn_fullpath() would always fail.
>>
>> What is not entirely clear to me is why we need to recover the path from the
>> vnode if we, obviously, have the path even before we have the vnode.
> 
> Using the default setting for core files, it's based on the CWD of the process.  If you're using a different kern.corefile setting, it's different.  You will also need to account for the %N format string (check the code for indexpos).  Given that this is far from being a hot path, it's much easier to just do a vn_fullpath() on the vnode than to deal with all the other details. 


Hmm, I mean that given this code:

        flags = O_CREAT | FWRITE | O_NOFOLLOW;
        NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, name, td);
        error = vn_open_cred(&nd, &flags, cmode, oflags, td->td_ucred, NULL);

'name' is the name, right?  Can we keep and use it?

-- 
Andriy Gapon