On Wed, Nov 13, 2024 at 02:25:53PM -0800, Reinette Chatre wrote:
> Hi Tony,
>
> On 10/29/24 10:28 AM, Tony Luck wrote:
> > Computing memory bandwidth for all enabled events resulted in
> > identical code blocks for total and local bandwidth in mbm_update().
> >
> > Refactor with a helper function to eliminate code duplication.
> >
> > No functional change.
> >
> > Signed-off-by: Tony Luck <tony.luck@xxxxxxxxx>
> > ---
> > arch/x86/kernel/cpu/resctrl/monitor.c | 69 ++++++++++-----------------
> > 1 file changed, 24 insertions(+), 45 deletions(-)
> >
> > diff --git a/arch/x86/kernel/cpu/resctrl/monitor.c b/arch/x86/kernel/cpu/resctrl/monitor.c
> > index 3ef339e405c2..1b6cb3bbc008 100644
> > --- a/arch/x86/kernel/cpu/resctrl/monitor.c
> > +++ b/arch/x86/kernel/cpu/resctrl/monitor.c
> > @@ -829,62 +829,41 @@ static void update_mba_bw(struct rdtgroup *rgrp, struct rdt_mon_domain *dom_mbm)
> > resctrl_arch_update_one(r_mba, dom_mba, closid, CDP_NONE, new_msr_val);
> > }
> >
> > -static void mbm_update(struct rdt_resource *r, struct rdt_mon_domain *d,
> > - u32 closid, u32 rmid)
> > +static void mbm_update_one_event(struct rdt_resource *r, struct rdt_mon_domain *d,
> > + u32 closid, u32 rmid, enum resctrl_event_id evtid)
> > {
> > struct rmid_read rr = {0};
> >
> > rr.r = r;
> > rr.d = d;
> > + rr.evtid = evtid;
> > + rr.arch_mon_ctx = resctrl_arch_mon_ctx_alloc(rr.r, rr.evtid);
> > + if (IS_ERR(rr.arch_mon_ctx)) {
> > + pr_warn_ratelimited("Failed to allocate monitor context: %ld",
> > + PTR_ERR(rr.arch_mon_ctx));
> > + return;
> > + }
> > +
> > + __mon_event_count(closid, rmid, &rr);
> > +
> > + if (is_mba_sc(NULL))
> > + mbm_bw_count(closid, rmid, &rr);
> > +
>
> As I am staring at this more there seems to be an existing issue here ... note how
> __mon_event_count()'s return value is not checked before mbm_bw_count() is called.
> This means that mbm_bw_count() may run with rr.val of 0 that results in wraparound
> inside it resulting in some unexpected bandwidth numbers. Since a counter read can fail
> with a "Unavailable"/"Error" from hardware it is not deterministic how frequently this
> issue can be encountered.
>
> Skipping mbm_bw_count() if rr.val is 0 is one option ... that would keep the bandwidth
> measurement static at whatever was the last successful read and thus not cause dramatic
> changes by the software controller ... setting bandwidth to 0 if rr.val is 0 is another
> option to reflect that bandwidth data is unavailable, but then the software controller should
> perhaps get signal to not make adjustments? I expect there are better options? What do
> you think?
Skipping mbm_bw_count() is also undesirable. If some later
__mon_event_count() does succeed the bandwidth will be computed
based on the last and current values as if they were one second
apart, when actually some longer interval elapsed.
I don't think this is a big issue for current Intel CPU RDT
implementations because I don't think they will return the
bit 62 unavailable value in the IA32_QM_CTR MSR. I'll ask
around to check.
But it does mean that implementing the "summary bandwidth"
file discussed in the other e-mail thread[1] may be more
complex on systems that can return that a counter is
unavailable. We'd have to keep track that two succesful
counter reads occured, with a measure of the interval
between them before reporting a value in the summary file.
-Tony
[1] https://lore.kernel.org/all/CALPaoCjCWZ4ZYfwooFEzMn15jJM7s9Rfq83YhorOGUD=1GdSyw@xxxxxxxxxxxxxx/
