Re: [Ipmitool-devel] ipmitool v1.8.10 and watchdog patch

[Al C. <ch...@ll...>]

Hey Carol,

> As you may recall, during the
> watchdog discussion several months back, we'd decided that full watchdog
> "set" command support is too dangerous and we subsequently had some
> patches pulled from the cvs tree which allowed that functionality.

I missed the original thread and may not be getting all of it by reading
the archived messages in the improper-thread-order.  Why is this deemed
too dangerous?  Is it b/c users should be using the ipmi kernel driver
watchdog?  While a kernel driver watchdog is best, it may not be
available to many users.  Especially those that may not be using linux
(not sure of the driver's porting), but more likely critical to those
that control an appliance w/ ipmitool.  i.e. they can manage the
appliance with IPMI, but they can't control the OS on it.

I would say rather than eliminate watchdog support, put something in the
manpage and help that says, "don't use this unless you really know what
you're doing."

That's my 2 cents.

Al

On Thu, 2008-05-01 at 13:49 -0700, Carol Hebert wrote:
> Hi,
> 
> First, please let me apologize for not having the 1.8.10
> patch-and-outstanding-issues list completed yet.  I had a series of
> crises to deal with since I posted my intent to get started on this and
> so was delayed.  However, all that is behind me now (knock on wood ;-)
> and I'm now back to working on the preliminary 1.8.10 full-time.  I
> expect to have the list of outstanding patches/issues posted to this
> mailing list soon for discussion.
> 
> I also mentioned some time back that I wanted to add support to ipmitool
> for turning off a running watchdog timer.  As you may recall, during the
> watchdog discussion several months back, we'd decided that full watchdog
> "set" command support is too dangerous and we subsequently had some
> patches pulled from the cvs tree which allowed that functionality.
> However, using the "set" command to turn off a running watchdog timer
> would be very useful.  (Of course any of the watchdog set commands can
> still be sent via the raw command interface.)
> 
> While browsing around the cvs tree, I noticed that a (second) set of
> watchdog commands is present under the mc command and having been added
> since 1.8.9 was released.  This seemed to me to be an ok place to have
> some watchdog support so I've gone ahead and made some modifications to
> the existing watchdog commands there.  I've expanded the "get" command
> results displayed, made "reset" safe (see below), and changed "set" to
> do one job only -- turn off a running timer.  I've also added man page
> support for this new functionality.  I am attaching a patch for review
> and comments.  Any and all (constructive ;-) comments are much
> appreciated.
> 
> For folks concerned about even having a reset command at all (valid
> concerns imo), I want to add a few notes on why I believe the "reset"
> I've implemented here is safe.  I ran some tests of this reset command
> in various scenarios and saw the behavior outlined below.  I only ran
> these tests on system and one (2.6.18) kernel so more extensive testing
> should be done to be sure it works ok on other systems before we roll to
> 1.8.10.  (And I intend to do more testing myself, too.)  If issues are
> found that we aren't able to fully address, we can ifdef or pull the
> reset code.
> 
> First, reset run in-band:
> 1) if the ipmi watchdog driver is not yet started, "reset" starts the
> timer but the action is "No action" so when the countdown ends, it does
> nothing.
> 
> 2) if the ipmi watchdog driver is started with start_now=0, it does the
> same as #1 -- timer started but action is "No action".  
> 
> 3) if the ipmi watchdog driver is started with start_now=1, "reset"
> resets the timer back to 5 minutes (a semi-arbitrary time I chose so it
> would be similar to the default time setting on many BMCs).  Other
> values set by the watchdog driver are unchanged.  Note:  this is where I
> see the most value for the reset command -- it gives more time if/when
> occasionally needed.
> 
> 4) After using "ipmitool mc watchdog off" to turn off a running timer,
> the action is set to "No action".  A reset command will reset the timer,
> but since it retains the former action setting of "No action", nothing
> happens when the countdown ends.
> 
> 5)  After sending '-V' to the ipmi watchdog driver for a graceful
> shutdown:  The watchdog driver doesn't appear to shut off the timer, but
> sets the action to "No action" -- the countdown ends and nothing happens
> (similar to #1, #2, #4 above).  A subsequent "restart" command will
> reset the timer, but there will be "No action" when the countdown ends.
> 
> Second, reset run out-of-band:
> 1) if the ipmi watchdog driver is not present/started, the behavior of
> reset is "No action" so the timer starts and then counts down, but
> nothing happens when the countdown completes.
> 
> 2) if the ipmi watchdog driver is present and start_now=1 is set,
> "reset" will restart the countdown but will not change anything else set
> by the ipmi watchdog driver (e.g., if the driver's action is to reset
> the system, the reset will still be triggered after the new countdown
> period ends).
> 
> --
> Again, all comments, concerns, and feedback are much appreciated.
> 
> Thanks,
> 
> Carol Hebert
> 
> 
> -------------------------------------------------------------------------
> This SF.net email is sponsored by the 2008 JavaOne(SM) Conference 
> Don't miss this year's exciting event. There's still time to save $100. 
> Use priority code J8TL2D2. 
> http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone
> _______________________________________________ Ipmitool-devel mailing list Ipm...@li... https://lists.sourceforge.net/lists/listinfo/ipmitool-devel
-- 
Albert Chu
ch...@ll...
925-422-5311
Computer Scientist
High Performance Systems Division
Lawrence Livermore National Laboratory