x86/intel_rdt/mba_sc: Feedback loop to dynamically update mem bandwidth

mba_sc is a feedback loop where we periodically read MBM counters and
try to restrict the bandwidth below a max value so the below is always
true:

  "current bandwidth(cur_bw) < user specified bandwidth(user_bw)"

The frequency of these checks is currently 1s and we just tag along the
MBM overflow timer to do the updates. Doing it once in a second also
makes the calculation of bandwidth easy. The steps of increase or
decrease of bandwidth is the minimum granularity specified by the
hardware.

Although the MBA's goal is to restrict the bandwidth below a maximum,
there may be a need to even increase the bandwidth. Since MBA controls
the L2 external bandwidth where as MBM measures the L3 external
bandwidth, we may end up restricting some rdtgroups unnecessarily. This
may happen in the sequence where rdtgroup (set of jobs) had high
"L3 <-> memory traffic" in initial phases -> mba_sc kicks in and reduced
bandwidth percentage values -> but after some it has mostly "L2 <-> L3"
traffic. In this scenario mba_sc increases the bandwidth percentage when
there is lesser memory traffic.

Signed-off-by: Vikas Shivappa <vikas.shivappa@linux.intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: ravi.v.shankar@intel.com
Cc: tony.luck@intel.com
Cc: fenghua.yu@intel.com
Cc: vikas.shivappa@intel.com
Cc: ak@linux.intel.com
Cc: hpa@zytor.com
Link: https://lkml.kernel.org/r/1524263781-14267-7-git-send-email-vikas.shivappa@linux.intel.com

3 files changed, 128 insertions, 3 deletions

diff --git a/arch/x86/kernel/cpu/intel_rdt.c b/arch/x86/kernel/cpu/intel_rdt.c
index ad03d975883e..24bfa63e86cf 100644
--- a/arch/x86/kernel/cpu/intel_rdt.c
+++ b/arch/x86/kernel/cpu/intel_rdt.c
@@ -33,7 +33,6 @@
 #include <asm/intel_rdt_sched.h>
 #include "intel_rdt.h"
 
-#define MAX_MBA_BW	100u
 #define MBA_IS_LINEAR	0x4
 #define MBA_MAX_MBPS	U32_MAX
 
@@ -350,7 +349,7 @@ static int get_cache_id(int cpu, int level)
  * that can be written to QOS_MSRs.
  * There are currently no SKUs which support non linear delay values.
  */
-static u32 delay_bw_map(unsigned long bw, struct rdt_resource *r)
+u32 delay_bw_map(unsigned long bw, struct rdt_resource *r)
 {
 	if (r->membw.delay_linear)
 		return MAX_MBA_BW - bw;
diff --git a/arch/x86/kernel/cpu/intel_rdt.h b/arch/x86/kernel/cpu/intel_rdt.h
index 66a0ba37a8a3..39752825e376 100644
--- a/arch/x86/kernel/cpu/intel_rdt.h
+++ b/arch/x86/kernel/cpu/intel_rdt.h
@@ -28,6 +28,7 @@
 
 #define MBM_CNTR_WIDTH			24
 #define MBM_OVERFLOW_INTERVAL		1000
+#define MAX_MBA_BW			100u
 
 #define RMID_VAL_ERROR			BIT_ULL(63)
 #define RMID_VAL_UNAVAIL		BIT_ULL(62)
@@ -461,6 +462,7 @@ void mbm_setup_overflow_handler(struct rdt_domain *dom,
 void mbm_handle_overflow(struct work_struct *work);
 bool is_mba_sc(struct rdt_resource *r);
 void setup_default_ctrlval(struct rdt_resource *r, u32 *dc, u32 *dm);
+u32 delay_bw_map(unsigned long bw, struct rdt_resource *r);
 void cqm_setup_limbo_handler(struct rdt_domain *dom, unsigned long delay_ms);
 void cqm_handle_limbo(struct work_struct *work);
 bool has_busy_rmid(struct rdt_resource *r, struct rdt_domain *d);
diff --git a/arch/x86/kernel/cpu/intel_rdt_monitor.c b/arch/x86/kernel/cpu/intel_rdt_monitor.c
index 7690402c42b7..b0f3aed76b75 100644
--- a/arch/x86/kernel/cpu/intel_rdt_monitor.c
+++ b/arch/x86/kernel/cpu/intel_rdt_monitor.c
@@ -329,6 +329,118 @@ void mon_event_count(void *info)
 	}
 }
 
+/*
+ * Feedback loop for MBA software controller (mba_sc)
+ *
+ * mba_sc is a feedback loop where we periodically read MBM counters and
+ * adjust the bandwidth percentage values via the IA32_MBA_THRTL_MSRs so
+ * that:
+ *
+ *   current bandwdith(cur_bw) < user specified bandwidth(user_bw)
+ *
+ * This uses the MBM counters to measure the bandwidth and MBA throttle
+ * MSRs to control the bandwidth for a particular rdtgrp. It builds on the
+ * fact that resctrl rdtgroups have both monitoring and control.
+ *
+ * The frequency of the checks is 1s and we just tag along the MBM overflow
+ * timer. Having 1s interval makes the calculation of bandwidth simpler.
+ *
+ * Although MBA's goal is to restrict the bandwidth to a maximum, there may
+ * be a need to increase the bandwidth to avoid uncecessarily restricting
+ * the L2 <-> L3 traffic.
+ *
+ * Since MBA controls the L2 external bandwidth where as MBM measures the
+ * L3 external bandwidth the following sequence could lead to such a
+ * situation.
+ *
+ * Consider an rdtgroup which had high L3 <-> memory traffic in initial
+ * phases -> mba_sc kicks in and reduced bandwidth percentage values -> but
+ * after some time rdtgroup has mostly L2 <-> L3 traffic.
+ *
+ * In this case we may restrict the rdtgroup's L2 <-> L3 traffic as its
+ * throttle MSRs already have low percentage values.  To avoid
+ * unnecessarily restricting such rdtgroups, we also increase the bandwidth.
+ */
+static void update_mba_bw(struct rdtgroup *rgrp, struct rdt_domain *dom_mbm)
+{
+	u32 closid, rmid, cur_msr, cur_msr_val, new_msr_val;
+	struct mbm_state *pmbm_data, *cmbm_data;
+	u32 cur_bw, delta_bw, user_bw;
+	struct rdt_resource *r_mba;
+	struct rdt_domain *dom_mba;
+	struct list_head *head;
+	struct rdtgroup *entry;
+
+	r_mba = &rdt_resources_all[RDT_RESOURCE_MBA];
+	closid = rgrp->closid;
+	rmid = rgrp->mon.rmid;
+	pmbm_data = &dom_mbm->mbm_local[rmid];
+
+	dom_mba = get_domain_from_cpu(smp_processor_id(), r_mba);
+	if (!dom_mba) {
+		pr_warn_once("Failure to get domain for MBA update\n");
+		return;
+	}
+
+	cur_bw = pmbm_data->prev_bw;
+	user_bw = dom_mba->mbps_val[closid];
+	delta_bw = pmbm_data->delta_bw;
+	cur_msr_val = dom_mba->ctrl_val[closid];
+
+	/*
+	 * For Ctrl groups read data from child monitor groups.
+	 */
+	head = &rgrp->mon.crdtgrp_list;
+	list_for_each_entry(entry, head, mon.crdtgrp_list) {
+		cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
+		cur_bw += cmbm_data->prev_bw;
+		delta_bw += cmbm_data->delta_bw;
+	}
+
+	/*
+	 * Scale up/down the bandwidth linearly for the ctrl group.  The
+	 * bandwidth step is the bandwidth granularity specified by the
+	 * hardware.
+	 *
+	 * The delta_bw is used when increasing the bandwidth so that we
+	 * dont alternately increase and decrease the control values
+	 * continuously.
+	 *
+	 * For ex: consider cur_bw = 90MBps, user_bw = 100MBps and if
+	 * bandwidth step is 20MBps(> user_bw - cur_bw), we would keep
+	 * switching between 90 and 110 continuously if we only check
+	 * cur_bw < user_bw.
+	 */
+	if (cur_msr_val > r_mba->membw.min_bw && user_bw < cur_bw) {
+		new_msr_val = cur_msr_val - r_mba->membw.bw_gran;
+	} else if (cur_msr_val < MAX_MBA_BW &&
+		   (user_bw > (cur_bw + delta_bw))) {
+		new_msr_val = cur_msr_val + r_mba->membw.bw_gran;
+	} else {
+		return;
+	}
+
+	cur_msr = r_mba->msr_base + closid;
+	wrmsrl(cur_msr, delay_bw_map(new_msr_val, r_mba));
+	dom_mba->ctrl_val[closid] = new_msr_val;
+
+	/*
+	 * Delta values are updated dynamically package wise for each
+	 * rdtgrp everytime the throttle MSR changes value.
+	 *
+	 * This is because (1)the increase in bandwidth is not perfectly
+	 * linear and only "approximately" linear even when the hardware
+	 * says it is linear.(2)Also since MBA is a core specific
+	 * mechanism, the delta values vary based on number of cores used
+	 * by the rdtgrp.
+	 */
+	pmbm_data->delta_comp = true;
+	list_for_each_entry(entry, head, mon.crdtgrp_list) {
+		cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
+		cmbm_data->delta_comp = true;
+	}
+}
+
 static void mbm_update(struct rdt_domain *d, int rmid)
 {
 	struct rmid_read rr;
@@ -346,7 +458,16 @@ static void mbm_update(struct rdt_domain *d, int rmid)
 	}
 	if (is_mbm_local_enabled()) {
 		rr.evtid = QOS_L3_MBM_LOCAL_EVENT_ID;
-		__mon_event_count(rmid, &rr);
+
+		/*
+		 * Call the MBA software controller only for the
+		 * control groups and when user has enabled
+		 * the software controller explicitly.
+		 */
+		if (!is_mba_sc(NULL))
+			__mon_event_count(rmid, &rr);
+		else
+			mbm_bw_count(rmid, &rr);
 	}
 }
 
@@ -417,6 +538,9 @@ void mbm_handle_overflow(struct work_struct *work)
 		head = &prgrp->mon.crdtgrp_list;
 		list_for_each_entry(crgrp, head, mon.crdtgrp_list)
 			mbm_update(d, crgrp->mon.rmid);
+
+		if (is_mba_sc(NULL))
+			update_mba_bw(prgrp, d);
 	}
 
 	schedule_delayed_work_on(cpu, &d->mbm_over, delay);