Re: Filepaths in VM map for tmpfs files

On 02/05/2015 07:25 AM, John Baldwin wrote:
> On Thursday, February 05, 2015 10:37:55 AM Konstantin Belousov wrote:
>> On Wed, Feb 04, 2015 at 10:15:04AM -0500, John Baldwin wrote:
>>> On Tuesday, February 03, 2015 10:33:36 PM Konstantin Belousov wrote:
>>>> On Mon, Feb 02, 2015 at 09:50:22PM -0600, Eric Badger wrote:
>>>>> On 02/02/2015 03:30 AM, Konstantin Belousov wrote:
>>>>>> On Sun, Feb 01, 2015 at 08:38:29PM -0600, Eric Badger wrote:
>>>>>>> On 01/31/2015 09:36 AM, Konstantin Belousov wrote:
>>>>>>>> First, shouldn't the kve_type changed to KVME_TYPE_VNODE as well ?
>>>>>>> My thinking is no, because KVME_TYPE_SWAP is in fact the correct
>>>>>>> type;
>>>>>>> I'd opine that it is better to be transparent than make it look
>>>>>>> like
>>>>>>> there is an OBJT_VNODE object there. It may be that some programs
>>>>>>> would
>>>>>>> be confused by VNODE info returned on a SWAP type mapping, though I
>>>>>>> know
>>>>>>> that dtrace handles it OK.
>>>>>> kve_vn_* and kve_path fields are defined only for KVME_TYPE_VNODE
>>>>>> kve_type.
>>>>>> So this is in fact a bug in whatever used the API to access kve_path
>>>>>> for KVE_TYPE_SWAP.
>>>>> Hmm, is that documented anywhere? I think it's fair to assume that
>>>>> kve_vn* applies only to the VNODE type,
>>>>> but I know there are several in-tree users that reference kve_path
>>>>> regardless of type (ostensibly relying
>>>>> on the default of an empty string). Maybe one could determine the
>>>>> validity of the kve_vn* fields by
>>>>> inspecting the kve_vn_type (not sure of all the consequences of that)?
>>>>> Or change it to KVME_TYPE_VNODE
>>>>> and deal with the below problem...
>>>> There is no useful documentation for the kern.proc. sysctls.
>>>> My word (and statements from other involved developers) could be
>>>> considered as close to the truth as it can be.
>>>> Somebody taking the efforts to document the stuff would make very
>>>> valuable contribution.
>>> I think that kve_path should be valid for all types (e.g. shm_open()
>>> is not a vnode but has a pathname, and that should be fixed to display
>>> if possible). In the equivalent for files (kinfo_file), the pathname
>>> is type-independent and always valid.
>> Well, this means that it should be valid for vnodes and shm.  My point
>> is that kvme_vn_path should be used only after the check for type.
>> We can and do set it to nul string, but using the path unconditionally
>> is a bug in the user code.
> The problem is that shm's can have different types (DEFAULT vs SWAP vs PHYS).
> :)  For kinfo_file, tools like fstat always print kf_path regardless of type.
> I do think it would be more consistent if the path in a kvme worked the same
> way.  Then you don't have to update all the tools each time a type starts
> populating the path.

Re: the kve_vn* fields, isn't setting kve_status = KF_ATTR_VALID the way 
to mark them as valid (irrespective of kve_type)? As for path name, I'd 
agree that there's no inherent need to restrict it by type. The field is 
somewhat self-validating (if something other than an empty string was 
returned in the path name field, this field is obviously valid).

>
>>> That said, I think tmpfs nodes should be exposed as files. It is an
>>> implementation detail of tmpfs that they are swap-backed, but from a
>>> user's perspective these are files, and if you want to expose other
>>> vnode-specific fields than just the path, KVME_TYPE_VNODE would be
>>> more correct.
>> I agree, but doing it is not easy, since there might be no vnode
>> to get the required information from.  We do know that this swap
>> object is for tmpfs node, but currently we only store pointer to
>> object in the node, not pointer to node from the object.  When the
>> vnode exists, pointer to vnode is stored in the object.
>>
>> To fix the issue, we should store pointer to node. Code was not done
>> this way, because VM code  which handles special-case for OBJT_TMPFS,
>> would need to know tmpfs internals. Right now, code knows about vnodes
>> anyway, so object->vnode does not bring tmpfs internals into vm.
> I'm more arguing in support of your original proposal.  Doing a best effort if
> the vnode exists would certainly be an improvement over what we have now.
>

I'll make one more brief case for returning tmpfs vm objects as 
KVME_TYPE_SWAP. Isn't the purpose of this sysctl for debugging, or to 
help a user understand what is going on internally? I can imagine 
scenarios where knowing that a mapped file is swap backed is relevant 
information, and returning it as KVME_TYPE_VNODE would hide this 
information.

I'd put forth a vote for return vnode info on a best-effort basis, at 
least for now.

Eric