[lustre-discuss] [PATCH 32/37] staging/lustre: use 64-bit times for exp_last_request_time

[Drokin, Oleg]

On Sep 24, 2015, at 11:18 AM, Arnd Bergmann wrote:

> On Thursday 24 September 2015 03:55:20 Drokin, Oleg wrote:
>> On Sep 23, 2015, at 3:13 PM, Arnd Bergmann wrote:
>> 
>>> The last request time is stored as an 'unsigned long', which is
>>> good enough until 2106, but it is then converted to 'long' in
>>> some places, which overflows in 2038.
>>> 
>>> This changes the type to time64_t to avoid those problems.
>> 
>> Hm…
>> All this code is actually only making sense on server and is unused otherwise,
>> so it's probably best to drop ping_evictor_start, ping_evictor_main, exp_expired,
>> class_disconnect_export_list (and two places where it's called from) functions
>> and exp_last_request_time field.
>> And with ping evictor gone, we should also drop ptlrpc_update_export_timer.
>> 
>> While clients do retain the request handling code (to process server-originated
>> requests like lock callbacks), they are not going to evict the servers because
>> the server have not talked to us in a while or anything of the sort.
> 
> I tried doing this, but could not figure out how to get rid of
> class_disconnect_exports().

It's only called from class_cleanup like this:

        /* The three references that should be remaining are the
         * obd_self_export and the attach and setup references. */
        if (atomic_read(&obd->obd_refcount) > 3) {
                /* refcount - 3 might be the number of real exports
                   (excluding self export). But class_incref is called
                   by other things as well, so don't count on it. */
                CDEBUG(D_IOCTL, "%s: forcing exports to disconnect: %d\n",
                       obd->obd_name, atomic_read(&obd->obd_refcount) - 3);
                dump_exports(obd, 0);
                class_disconnect_exports(obd);
        }


This is only true on the servers so we can replace it with a corresponding LASSERT,
I imagine.

> However, I started removing dead code and ended up with a huge patch
> that would not make it to mailing list servers and that I therefor
> pasted on http://pastebin.com/uncZQNh7 for reference.

Wow, this is a large patch indeed.
Some parts of it I was contemplating on my own like whole dt_object.[ch] removal
with all the ties in llog code, though it was somewhat tricky as some bits in mgc
seems to be using that somehow.

I have not finished a detailed runthrough yet, but on the surface, why did you remove
suppress_pings parameter 0 that is still valid on clients, to let them not ping servers.
Also ptlrpc_ping_import_soon and friends - all that code is actually needed.

Only "ping evictor" is not needed on the client as it's only servers that are going to
kick out clients that are silent for too long. Clients are still expected to
send their "keep alive" pings in periodically (unless suppress_pings option is enabled).

Hm…. I now see it is not fully implemented, that's why you are removing it as it's really not
called from anywhere.


Anyway this does remove a lot of stuff that we don't really need in the client,
I'll try to get it built and tested just to make sure it does not really break anything
(unfortunately it does not seem to apply cleanly to the tip of staging-next tree).

Thank you!

Bye,
    Oleg