On Tue, 2014-08-12 at 21:56 -0700, Roland Dreier wrote:
> On Tue, Aug 12, 2014 at 4:20 PM, Doug Ledford <dledford@xxxxxxxxxx> wrote:
> > - ret = ib_find_pkey(target->srp_host->srp_dev->dev,
> > - target->srp_host->port,
> > - be16_to_cpu(target->path.pkey),
> > - &attr->pkey_index);
> > + ret = ib_find_cached_pkey(target->srp_host->srp_dev->dev,
> > + target->srp_host->port,
> > + be16_to_cpu(target->path.pkey),
> > + &attr->pkey_index);
>
> I consider the "cached" operations to be a deprecated interface,
I disagree.
> and
> I'd rather remove uses than add new ones.
This isn't practical. For the ib_get_* cache operations, you could
theoretically do away with the cache and not suffer too bad. But this
is one of the ib_find_* operations. These functions can not be done
away with.
> I have a vague dream of
> fixing everything up so the normal interfaces (ib_find_pkey() etc)
> don't sleep and are usable everywhere,
Sleeping isn't the problem. And that's probably part of why you are
still dreaming about this instead of accepting reality. The problem is
the latency required to query the firmware on a card. Some IB adapters
might keep their tables in kernel memory and can access them fast, but
some cards use firmware and the tables are there and when you query the
driver you are actually doing a query to the card via a firmware and a
mailbox and you have to wait for a return. *That's* the problem. And
specifically as it relates to the ib_find_* functions as this function
is, when the P_Key you're searching for isn't one of the first found,
the difference between the cache and direct to the card is very
significant.
> although that dream will
> probably never come true.
No offense Roland, but not if I have anything to say about it ;-) Not
because I don't want you to have your dreams of course, but this
particular dream is more nightmare for the rest of us than the unicorns
and rainbows raining skittles on the world that you envision it being.
> But is there any practical benefit to this change?
Absolutely yes!
> I don't think
> saving a few milliseconds at log in time counts
You can't think that way Roland....it breaks real world usage in very
unexpected ways. Ways that took me, 3 support engineers from two
companies, and top level admins at a joint customer's sight over a month
to finally figure out. So, let me be clear, this particular dream of
yours has wasted well over 4 man months of support/engineering time.
Email me off list if you want more details of how this nightmare dream
of yours did it.
> -- I would think it's
> lost in the noise of how long it takes to actually log in, scan for
> LUNs, etc.
No, it's really not. The simple case of finding the right element in
the first lookup or two is OK. But this is the ib_find_* function, and
if it doesn't find the right element at first, it searches all available
elements until it does. There are 128 slots (like there are for GID
indexes), so it's a few ms times 128 in the worst case scenario. That's
not noise. It's so far from noise that it kills real world
applications. Things start timing out and resetting when you have more
than just a few items search the table for an element that isn't there.
It really can't be thought of in the terms you are thinking of it, which
is the common/best case scenario. You *have* to consider worst case
scenario, and in that scenario, the cache functions are absolutely
essential to scalability. Especially the ib_find_* variants. And since
they are essential, they *must* be maintained and kept working properly,
and so there is no reason not to use them when they can be used.
--
Doug Ledford <dledford@xxxxxxxxxx>
GPG KeyID: 0E572FDD
Attachment:
signature.asc
Description: This is a digitally signed message part
