Command Deadlines: HP's Modest Proposal

Ballard, Curtis C (HP Storage) curtis.ballard at hp.com
Fri Sep 5 12:16:58 PDT 2014


Formatted message: <a href="http://www.t10.org/cgi-bin/ac.pl?t=r&f=r1409050_f.htm">HTML-formatted message</a>

I doubt we would object to using more than one bit in the CONTROL byte to
allow selection of one of a list of timeouts and potentially a simpler method
of specifying timeouts.  There are only 3 reserved bits in the CONTROL byte
though and 2 obsolete bits with the NACA bit in between those so using 4
seems like a stretch.

There are some issues to getting to the CONTROL byte but in many cases the
raw CDB is accessible so it is possible to get to the CONTROL byte.

Curtis Ballard

Hewlett-Packard Company

+1 970 898 3013 / Tel

Curtis.Ballard at hp.com / Email

Fort Collins, CO

From: owner-t10 at t10.org [mailto:owner-t10 at t10.org] On Behalf Of Gerry Houlder

Sent: Thursday, September 4, 2014 3:20 PM

To: T10 Reflector

Subject: Re: Command Deadlines: HP's Modest Proposal

OK, i understand the reluctance to use the Priority field because existing
stacks can't use it. That is why the original proposal was trying to use the
group field in the CDB instead of the Priority field. A location in the CDB
also has advantages because it is not protocol dependent, allowing the
feature to potentially be used over an intervening FC protocol (for
instance).

Since you are willing to use the CONTROL field of the CDB (i have heard
similar complaints that stacks can't get to this byte either, so that is a
different issue), why not use 4 bits in that field so that 16 different
values are possible. This can use the 16 entry mode page that is described in
the current proposal. Up to now, we have been thinking that 16 values would
satisfy a bunch of different initiators especially since the applications run
by the different initiators might have similar deadline requirements. This
allows the target to have a shared mode page for all initiators and
simplifying a lot of things, compared to supporting a separate mode page for
each initiator.

On Thu, Sep 4, 2014 at 11:29 AM, Ballard, Curtis C (HP Storage)
<curtis.ballard at hp.com> wrote:

“per opcode” isn’t defined.  The simple path would be to say each
opcode/service_action pair but we could define a mapping so a single timeout
value maps to all READ commands regardless of size.

The mode page policy is an issue but which is more difficult, dealing with a
per I_T nexus mode page or getting the priority through the storage stack? 
There are use models where a shared mode page would still work but it would
be quite a bit less flexible.  In at least one use case we’re considering
there would only ever be one initiator so the mode page policy really
wouldn’t matter.  I don’t think the question of what values to use when a
new initiator first makes itself known is a problem in this case, a new
initiator wouldn’t have any timeout values set until it established them. 
I think we would have to define the model for a device server running out of
command timeout slots.	The model could allow for the device server to have a
pool of timeouts that is shared across all initiators and timeouts are
assigned to specific initiators following a MODE SELECT specifying timeouts,
then whenever that pool is expired attempts to set a new timeout would error
with a well-defined response.

The current proposal has the mode page policy as ‘should’ be per I_T
nexus but does allow for shared.

Curtis Ballard

Hewlett-Packard Company

+1 970 898 3013 / Tel

Curtis.Ballard at hp.com / Email

Fort Collins, CO

From: owner-t10 at t10.org<mailto:owner-t10 at t10.org>
[mailto:owner-t10 at t10.org<mailto:owner-t10 at t10.org>] On Behalf Of Gerry
Houlder

Sent: Thursday, September 4, 2014 10:02 AM

To: T10 Reflector

Subject: Re: Command Deadlines: HP's Modest Proposal

So when you say "per op code", this means one value for READ(10), one for
READ(16), and one for READ(32)? And three more for the comparable write
commands? And a requirement to save a separate copy of this new mode page
(that contains 6 timeout values) for every initiator? How many initiators
would you expect a target device to handle? Most target devices use shared
mode page policy for all mode pages to avoid that last issue -- a separate
mode page for each initiator brings a lot of questions, such as what values
to use when a new initiator first makes itself known and what to do when the
target has reached its limit of the number of initiators it can handle.

i would prefer an implementation that works OK when the mode page is shared
across all initiators.

On Wed, Sep 3, 2014 at 7:15 PM, Ballard, Curtis C (HP Storage)
<curtis.ballard at hp.com> wrote:

The total number of timeout values shouldn’t need to be any more than the
current proposal which defines 30 timeout values per I_T nexus.  The device
server could define a set of commands where timers are supported and could
have a maximum number of timers.  Instead of timers being associated with a
priority they would be associated with specific commands.  Most applications
will have a small set of commands where they want to use deadlines.

As we discussed in the last conference call, these timeout values don’t
have to map to timers.	The model is taking the command arrival time and
adding the timeout to generate a command deadline and the device server
checks the current time against the deadline rather than running a timer.

This model doesn’t provide as much flexibility for the application as being
able to pick from a table of timers for every command but can work with the
existing CDB’s without needing to use the transport specific command
priority.  If more flexibility is needed 2 bits in the control byte could
allow selecting from one of 4 timeout values.

Curtis Ballard

Hewlett-Packard

From: owner-t10 at t10.org<mailto:owner-t10 at t10.org> [mailto:owner-t10 at t10.org]
On Behalf Of George Penokie

Sent: Wednesday, September 3, 2014 4:19 PM

To: Paul Suhler

Cc: T10 E-mail Reflector (t10 at t10.org<mailto:t10 at t10.org>)

Subject: Re: Command Deadlines: HP's Modest Proposal

:Paul,

One timeout per opcode per I_T_L nexus sounds like one heck of a lot of
timers to me.

Bye for now,

George Penokie

Avago Technologies

Attn: George Penokie

4109 Manor View Dr NW

Rochester , MN 55901

507-328-9017

george.penokie at avagotech.com

On Wed, Sep 3, 2014 at 4:30 PM, Paul Suhler
<Paul.Suhler at hgst.com> wrote:

Hi, all.

Curtis Ballard has made an interesting suggestion for a major change to last
week’s T10/14-107r2:

Instead of providing two sets of fifteen timeouts, one timeout would be
provided per-opcode-per-I_T_L nexus.  That timeout would be used if a
newly-allocated bit in the CDB’s Control byte were set.

Advantage:

No need to wait for the Command Priority attribute to be added to transport
implementations and to driver stacks, which could take years.

Disadvantage:

In cases where an initiator wants to change the timeout for  a particular
command (e.g., READ(16)) on a particular LU, it would have to send a MODE
SELECT.

Questions:

1)	Does anyone feel that one timeout per opcode per I_T_L nexus is too
much of a constraint?  HP doesn’t.

2)	Does anyone object to using mode pages, with the requirement that the
data length on MODE SELECT must match the data length on MODE SENSE?  (The
alternative is yet another command.)

Thanks for your feedback,

Paul

Paul A. Suhler, PhD

Research Staff Member

HGST Research

paul.suhler at hgst.com

o: 949-476-1180 x275448<tel:949-476-1180%20x275448>

m: 949-241-6443<tel:949-241-6443>

3001 Daimler St.

Santa Ana, CA 92705-5812

www.hgst.com<<a href="https://urldefense.proofpoint.com/v2/url?u=http-3A__www.hgst.com">https://urldefense.proofpoint.com/v2/url?u=http-3A__www.hgst.com
_&d=AAMGaQ&c=IGDlg0lD0b-nebmJJ0Kp8A&r=TxI1DC4HavpWBdSmUqvdNvSwgOklhaW328zLt5A
OpPM&m=LGQXGaZphTKJi8d6PpOs1l17I1-8bUt7RIBmjCfdzmA&s=ya2S3vh6OSl_-B9k521cbVFK
PQZ2BOKOPdI35OYSTEU&e=>




More information about the T10 mailing list