Re: [PATCH 19/22] multipathd: make multipath devices manage their path check times

[Benjamin Marzinski <bmarzins@redhat.com>]

On Mon, Jul 15, 2024 at 10:49:52PM +0200, Martin Wilck wrote:
> On Sat, 2024-07-13 at 02:05 -0400, Benjamin Marzinski wrote:
> > multipathd's path checking can be very bursty, with all the paths
> > running their path checkers at the same time, and then all doing
> > nothing
> > while waiting for the next check time. Alternatively, the paths in a
> > multipath device might all run their path checkers at a different
> > time,
> > which can keep the multipath device from having a coherent view of
> > the
> > state of all of its paths.
> > 
> > This patch makes all the paths of a multipath device converge to
> > running
> > their checkers at the same time, and spreads out when this time is
> > for
> > the different multipath devices.
> > 
> > To do this, the checking time is divided into adjustment intervals
> > (conf->adjust_int), so that the checkers run at some index within
> > this
> > interval. conf->adjust_int is chosen so that it is at least twice as
> > long as conf->max_checkint, and is a multiple of all possible
> > pp->checkint values. This means that regardless of pp->checkint, the
> > path should always be checked on the same indexes, each adjustement
> > interval.
> > 
> > Each multipath device has a goal index. These are evenly spread out
> > between 0 and conf->max_checkint. Every conf->adjust_int seconds,
> > each
> > multipath device should try to check all of its paths on its goal
> > index.
> > If a path isn't checked on the goal index, then pp->tick for the next
> > check after the goal_idx will be decremented by one, to align it over
> > time.
> > 
> > In order for the path checkers to run every pp->checkint seconds,
> > multipathd needs to track how long a path check has been pending for,
> > and subtract that time from the number of ticks till the checker is
> > run
> > again. If the checker has been pending for more that pp->checkint,
> > the path will be rechecked on the next tick after the checker
> > returns.
> > 
> > Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>
> 
> Nice, thanks a lot. I have a few remarks  below.
> 
> 
> > ---
> >  libmultipath/config.c  | 13 ++++++
> >  libmultipath/config.h  |  1 +
> >  libmultipath/structs.c |  1 +
> >  libmultipath/structs.h |  1 +
> >  multipathd/main.c      | 90 ++++++++++++++++++++++++++++++++++------
> > --
> >  5 files changed, 89 insertions(+), 17 deletions(-)
> > 
> > diff --git a/libmultipath/config.c b/libmultipath/config.c
> > index 83fa7369..ca584372 100644
> > --- a/libmultipath/config.c
> > +++ b/libmultipath/config.c
> > @@ -982,6 +982,19 @@ int _init_config (const char *file, struct
> > config *conf)
> >  		conf->checkint = conf->max_checkint;
> >  	condlog(3, "polling interval: %d, max: %d",
> >  		conf->checkint, conf->max_checkint);
> > +	/*
> > +	 * make sure that that adjust_int is at least twice as large
> > +	 * as checkint and is a multiple of all possible values of
> > +	 * pp->checkint.
> > +	 */
> > +	if (conf->max_checkint % conf->checkint == 0) {
> > +		conf->adjust_int = 2 * conf->max_checkint;
> > +	} else {
> > +		conf->adjust_int = conf->checkint;
> > +		while (2 * conf->adjust_int < conf->max_checkint)
> > +			conf->adjust_int *= 2;
> > +		conf->adjust_int *= conf->max_checkint;
> > +	}
> >  
> >  	if (conf->blist_devnode == NULL) {
> >  		conf->blist_devnode = vector_alloc();
> > diff --git a/libmultipath/config.h b/libmultipath/config.h
> > index 384193ab..800c0ca9 100644
> > --- a/libmultipath/config.h
> > +++ b/libmultipath/config.h
> > @@ -147,6 +147,7 @@ struct config {
> >  	int minio_rq;
> >  	unsigned int checkint;
> >  	unsigned int max_checkint;
> > +	unsigned int adjust_int;
> >  	bool use_watchdog;
> >  	int pgfailback;
> >  	int rr_weight;
> > diff --git a/libmultipath/structs.c b/libmultipath/structs.c
> > index 1583e001..472e1001 100644
> > --- a/libmultipath/structs.c
> > +++ b/libmultipath/structs.c
> > @@ -148,6 +148,7 @@ uninitialize_path(struct path *pp)
> >  	pp->dmstate = PSTATE_UNDEF;
> >  	pp->uid_attribute = NULL;
> >  	pp->checker_timeout = 0;
> > +	pp->pending_ticks = 0;
> >  
> >  	if (checker_selected(&pp->checker))
> >  		checker_put(&pp->checker);
> > diff --git a/libmultipath/structs.h b/libmultipath/structs.h
> > index 002eeae1..457d7836 100644
> > --- a/libmultipath/structs.h
> > +++ b/libmultipath/structs.h
> > @@ -360,6 +360,7 @@ struct path {
> >  	unsigned long long size;
> >  	unsigned int checkint;
> >  	unsigned int tick;
> > +	unsigned int pending_ticks;
> >  	int bus;
> >  	int offline;
> >  	int state;
> > diff --git a/multipathd/main.c b/multipathd/main.c
> > index daf668eb..42ccb92f 100644
> > --- a/multipathd/main.c
> > +++ b/multipathd/main.c
> > @@ -2386,7 +2386,7 @@ check_mpp(struct vectors * vecs, struct
> > multipath *mpp, unsigned int ticks)
> >   * and '0' otherwise
> >   */
> >  static int
> > -check_path (struct vectors * vecs, struct path * pp, unsigned int
> > ticks)
> > +do_check_path (struct vectors * vecs, struct path * pp)
> >  {
> >  	int newstate;
> >  	int new_path_up = 0;
> > @@ -2398,14 +2398,6 @@ check_path (struct vectors * vecs, struct path
> > * pp, unsigned int ticks)
> >  	int marginal_pathgroups, marginal_changed = 0;
> >  	bool need_reload;
> >  
> > -	if (pp->initialized == INIT_REMOVED)
> > -		return 0;
> > -
> > -	if (pp->tick)
> > -		pp->tick -= (pp->tick > ticks) ? ticks : pp->tick;
> > -	if (pp->tick)
> > -		return 0; /* don't check this path yet */
> > -
> >  	conf = get_multipath_config();
> >  	checkint = conf->checkint;
> >  	max_checkint = conf->max_checkint;
> > @@ -2417,12 +2409,6 @@ check_path (struct vectors * vecs, struct path
> > * pp, unsigned int ticks)
> >  		pp->checkint = checkint;
> >  	};
> >  
> > -	/*
> > -	 * provision a next check soonest,
> > -	 * in case we exit abnormally from here
> > -	 */
> > -	pp->tick = checkint;
> > -
> >  	newstate = check_path_state(pp);
> >  	if (newstate == PATH_WILD || newstate == PATH_UNCHECKED)
> >  		return 0;
> > @@ -2584,7 +2570,6 @@ check_path (struct vectors * vecs, struct path
> > * pp, unsigned int ticks)
> >  				condlog(4, "%s: delay next check
> > %is",
> >  					pp->dev_t, pp->checkint);
> >  			}
> > -			pp->tick = pp->checkint;
> >  		}
> >  	}
> >  	else if (newstate != PATH_UP && newstate != PATH_GHOST) {
> > @@ -2637,6 +2622,76 @@ check_path (struct vectors * vecs, struct path
> > * pp, unsigned int ticks)
> >  	return 1;
> >  }
> >  
> > +static int
> > +check_path (struct vectors * vecs, struct path * pp, unsigned int
> > ticks,
> > +	    time_t start_secs)
> > +{
> > +	int r;
> > +	unsigned int adjust_int, max_checkint;
> > +	struct config *conf;
> > +	time_t next_idx, goal_idx;
> > +
> > +	if (pp->initialized == INIT_REMOVED)
> > +		return 0;
> > +
> > +	if (pp->tick)
> > +		pp->tick -= (pp->tick > ticks) ? ticks : pp->tick;
> > +	if (pp->tick)
> > +		return 0; /* don't check this path yet */
> > +
> > +	conf = get_multipath_config();
> > +	max_checkint = conf->max_checkint;
> > +	adjust_int = conf->adjust_int;
> > +	put_multipath_config(conf);
> > +
> > +	r = do_check_path(vecs, pp);
> > +
> > +	/*
> > +	 * do_check_path() removed or orphaned the path.
> > +	 */
> > +	if (r < 0 || !pp->mpp)
> > +		return r;
> > +
> > +	/*
> > +	 * the path checker is pending
> > +	 */
> > +	if (pp->tick != 0) {
> 
> I'm not getting this logic. Please explain. Why is pp->tick != 0 at
> this point equivalent to the checker being pending?

It's not actually just pending. This also happens for a specific case of
PATH_DELAY paths. The short answer is that we should always respect a
pp->tick value set directly in do_check_path(), since that means that we
are supposed to recheck as soon as possible. And as far as incrementing
pp->pending_ticks for PATH_DELAY devices, it seemed harmless. It will
just make the first recheck after a path stops being delayed happen
sooner.

But you are correct, and it's not hard to clear pp->pending_ticks when
the path is in the PATH_DELAY state.

Actually setting pp->ticks to 1 for PATH_DELAY paths seems wrong. I
believe the only reason to do it is to make sure that we re-enqueue the
path for io err checking on time. But we know exactly how long we need
to wait until we do that, so we should be setting pp->tick to the
correct time when we need to, and leaving it alone otherwise. And it
shouldn't have anything to do with whether or not the path is in the
marginal pathgroup.

> 
> > +		pp->pending_ticks++;
> > +		return r;
> > +	}
> > +
> > +	/* schedule the next check */
> > +	pp->tick = pp->checkint;
> > +	if (pp->pending_ticks >= pp->tick)
> > +		pp->tick = 1;
> > +	else
> > +		pp->tick -= pp->pending_ticks;
> > +	pp->pending_ticks = 0;
> > +
> > +	if (pp->tick == 1)
> > +		return r;
> > +
> > +	/*
> > +	 * every mpp has a goal_idx in the range of
> > +	 * 0 <= goal_idx < conf->max_checkint
> > +	 *
> > +	 * The next check has an index, next_idx, in the range of
> > +	 * 0 <= next_idx < conf->adjust_int
> > +	 *
> > +	 * Every adjust_int seconds, each multipath device should
> > try to
> > +	 * check all of its paths on its goal_idx. If a path isn't
> > checked
> > +	 * on the goal_idx, then pp->tick for the next check after
> > the
> > +	 * goal_idx will be decremented by one, to align it over
> > time.
> > +	 */
> > +	goal_idx = (find_slot(vecs->mpvec, pp->mpp)) *
> > +		   max_checkint / VECTOR_SIZE(vecs->mpvec);
> > +	next_idx = (start_secs + pp->tick) % adjust_int;
> > +	if (next_idx > goal_idx && next_idx - goal_idx < pp-
> > >checkint)
> > +		pp->tick--;
> 
> Hm. IMHO this calculation should be little simpler, as follows:
>         
> 	cur_idx = start_secs % adjust_int;
>         if ((goal_idx - cur_idx) % pp->checkint != 0)
>                 pp->tick--;

We need to use next_idx instead of cur_idx. cur_idx is the time when the
path checker completed, not when it started. If the checker returned
PATH_PENDING, these two times will be different. Since path checkers may
not always take the same amount of time to complete, we can't use their
completing time to line them up on the same index. Instead we track how
long the checker was pending for, and remove that from the length of
time till the next check, so that checkers always start every
pp->checkint seconds.

But you are correct that we could just use

((goal_idx - next_idx) % pp->checkint != 0)
 
> My reasoning goes like this: If (goal_idx - cur_idx) % pp->checkint is
> 0, we need no adjustment, because (assuming checkint remains constant)
> it will be checked at the next goal_idx. Otherwise, we check one tick
> earlier, and will do so at the next check again, etc. This is similar
> to your approach, but not exactly the same, AFAICS. In particular, I am
> unsure about your "if (next_idx - goal_idx < pp->checkint)" condition.

The idea was that we would only modify pp->tick once per adjust_int. But
in reflection, there's really no reason to do that. Your version makes
more sense.

> If checkint grows toward max_checkint, it may take longer until this
> algorithm converges, but sooner or later, it should.
> 
> Instead of stepping, we could also force the idx immediately, by doing
> 
>         pp->tick -= (goal_idx - cur_idx) % pp->checkint;
> 
> which would be even simpler (the check intervals would be less evenly
> spaced, but I don't think that would hurt us much).

I think you mean:
	pp->tick -= (cur_idx - goal_idx) % pp->checkint;

But at any rate, I think we should try to keep the intervals mostly the
same. Plus I feel like shortening the intervals makes the most sense,
while the above will sometimes shorten and sometimes lengthen them. 

-Ben

> Am I misguided here?
> 
> Regards
> Martin