1 // SPDX-License-Identifier: GPL-2.0-only
3 * Resource Director Technology(RDT)
6 * Copyright (C) 2017 Intel Corporation
11 * This replaces the cqm.c based on perf but we reuse a lot of
12 * code and datastructures originally from Peter Zijlstra and Matt Fleming.
14 * More information about RDT be found in the Intel (R) x86 Architecture
15 * Software Developer Manual June 2016, volume 3, section 17.17.
18 #include <linux/module.h>
19 #include <linux/sizes.h>
20 #include <linux/slab.h>
22 #include <asm/cpu_device_id.h>
23 #include <asm/resctrl.h>
30 struct list_head list;
34 * @rmid_free_lru A least recently used list of free RMIDs
35 * These RMIDs are guaranteed to have an occupancy less than the
38 static LIST_HEAD(rmid_free_lru);
41 * @rmid_limbo_count count of currently unused but (potentially)
43 * This counts RMIDs that no one is currently using but that
44 * may have a occupancy value > resctrl_rmid_realloc_threshold. User can
45 * change the threshold occupancy value.
47 static unsigned int rmid_limbo_count;
50 * @rmid_entry - The entry in the limbo and free lists.
52 static struct rmid_entry *rmid_ptrs;
55 * Global boolean for rdt_monitor which is true if any
56 * resource monitoring is enabled.
61 * Global to indicate which monitoring events are enabled.
63 unsigned int rdt_mon_features;
66 * This is the threshold cache occupancy in bytes at which we will consider an
67 * RMID available for re-allocation.
69 unsigned int resctrl_rmid_realloc_threshold;
72 * This is the maximum value for the reallocation threshold, in bytes.
74 unsigned int resctrl_rmid_realloc_limit;
76 #define CF(cf) ((unsigned long)(1048576 * (cf) + 0.5))
79 * The correction factor table is documented in Documentation/arch/x86/resctrl.rst.
80 * If rmid > rmid threshold, MBM total and local values should be multiplied
81 * by the correction factor.
83 * The original table is modified for better code:
85 * 1. The threshold 0 is changed to rmid count - 1 so don't do correction
87 * 2. MBM total and local correction table indexed by core counter which is
88 * equal to (x86_cache_max_rmid + 1) / 8 - 1 and is from 0 up to 27.
89 * 3. The correction factor is normalized to 2^20 (1048576) so it's faster
90 * to calculate corrected value by shifting:
91 * corrected_value = (original_value * correction_factor) >> 20
93 static const struct mbm_correction_factor_table {
96 } mbm_cf_table[] __initconst = {
127 static u32 mbm_cf_rmidthreshold __read_mostly = UINT_MAX;
128 static u64 mbm_cf __read_mostly;
130 static inline u64 get_corrected_mbm_count(u32 rmid, unsigned long val)
132 /* Correct MBM value. */
133 if (rmid > mbm_cf_rmidthreshold)
134 val = (val * mbm_cf) >> 20;
139 static inline struct rmid_entry *__rmid_entry(u32 rmid)
141 struct rmid_entry *entry;
143 entry = &rmid_ptrs[rmid];
144 WARN_ON(entry->rmid != rmid);
149 static int __rmid_read(u32 rmid, enum resctrl_event_id eventid, u64 *val)
154 * As per the SDM, when IA32_QM_EVTSEL.EvtID (bits 7:0) is configured
155 * with a valid event code for supported resource type and the bits
156 * IA32_QM_EVTSEL.RMID (bits 41:32) are configured with valid RMID,
157 * IA32_QM_CTR.data (bits 61:0) reports the monitored data.
158 * IA32_QM_CTR.Error (bit 63) and IA32_QM_CTR.Unavailable (bit 62)
161 wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid);
162 rdmsrl(MSR_IA32_QM_CTR, msr_val);
164 if (msr_val & RMID_VAL_ERROR)
166 if (msr_val & RMID_VAL_UNAVAIL)
173 static struct arch_mbm_state *get_arch_mbm_state(struct rdt_hw_domain *hw_dom,
175 enum resctrl_event_id eventid)
178 case QOS_L3_OCCUP_EVENT_ID:
180 case QOS_L3_MBM_TOTAL_EVENT_ID:
181 return &hw_dom->arch_mbm_total[rmid];
182 case QOS_L3_MBM_LOCAL_EVENT_ID:
183 return &hw_dom->arch_mbm_local[rmid];
186 /* Never expect to get here */
192 void resctrl_arch_reset_rmid(struct rdt_resource *r, struct rdt_domain *d,
193 u32 rmid, enum resctrl_event_id eventid)
195 struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
196 struct arch_mbm_state *am;
198 am = get_arch_mbm_state(hw_dom, rmid, eventid);
200 memset(am, 0, sizeof(*am));
202 /* Record any initial, non-zero count value. */
203 __rmid_read(rmid, eventid, &am->prev_msr);
208 * Assumes that hardware counters are also reset and thus that there is
209 * no need to record initial non-zero counts.
211 void resctrl_arch_reset_rmid_all(struct rdt_resource *r, struct rdt_domain *d)
213 struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
215 if (is_mbm_total_enabled())
216 memset(hw_dom->arch_mbm_total, 0,
217 sizeof(*hw_dom->arch_mbm_total) * r->num_rmid);
219 if (is_mbm_local_enabled())
220 memset(hw_dom->arch_mbm_local, 0,
221 sizeof(*hw_dom->arch_mbm_local) * r->num_rmid);
224 static u64 mbm_overflow_count(u64 prev_msr, u64 cur_msr, unsigned int width)
226 u64 shift = 64 - width, chunks;
228 chunks = (cur_msr << shift) - (prev_msr << shift);
229 return chunks >> shift;
232 int resctrl_arch_rmid_read(struct rdt_resource *r, struct rdt_domain *d,
233 u32 rmid, enum resctrl_event_id eventid, u64 *val)
235 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
236 struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
237 struct arch_mbm_state *am;
241 if (!cpumask_test_cpu(smp_processor_id(), &d->cpu_mask))
244 ret = __rmid_read(rmid, eventid, &msr_val);
248 am = get_arch_mbm_state(hw_dom, rmid, eventid);
250 am->chunks += mbm_overflow_count(am->prev_msr, msr_val,
252 chunks = get_corrected_mbm_count(rmid, am->chunks);
253 am->prev_msr = msr_val;
258 *val = chunks * hw_res->mon_scale;
264 * Check the RMIDs that are marked as busy for this domain. If the
265 * reported LLC occupancy is below the threshold clear the busy bit and
266 * decrement the count. If the busy count gets to zero on an RMID, we
269 void __check_limbo(struct rdt_domain *d, bool force_free)
271 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
272 struct rmid_entry *entry;
273 u32 crmid = 1, nrmid;
278 * Skip RMID 0 and start from RMID 1 and check all the RMIDs that
279 * are marked as busy for occupancy < threshold. If the occupancy
280 * is less than the threshold decrement the busy counter of the
281 * RMID and move it to the free list when the counter reaches 0.
284 nrmid = find_next_bit(d->rmid_busy_llc, r->num_rmid, crmid);
285 if (nrmid >= r->num_rmid)
288 entry = __rmid_entry(nrmid);
290 if (resctrl_arch_rmid_read(r, d, entry->rmid,
291 QOS_L3_OCCUP_EVENT_ID, &val)) {
294 rmid_dirty = (val >= resctrl_rmid_realloc_threshold);
297 if (force_free || !rmid_dirty) {
298 clear_bit(entry->rmid, d->rmid_busy_llc);
299 if (!--entry->busy) {
301 list_add_tail(&entry->list, &rmid_free_lru);
308 bool has_busy_rmid(struct rdt_resource *r, struct rdt_domain *d)
310 return find_first_bit(d->rmid_busy_llc, r->num_rmid) != r->num_rmid;
314 * As of now the RMIDs allocation is global.
315 * However we keep track of which packages the RMIDs
316 * are used to optimize the limbo list management.
320 struct rmid_entry *entry;
322 lockdep_assert_held(&rdtgroup_mutex);
324 if (list_empty(&rmid_free_lru))
325 return rmid_limbo_count ? -EBUSY : -ENOSPC;
327 entry = list_first_entry(&rmid_free_lru,
328 struct rmid_entry, list);
329 list_del(&entry->list);
334 static void add_rmid_to_limbo(struct rmid_entry *entry)
336 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
337 struct rdt_domain *d;
343 list_for_each_entry(d, &r->domains, list) {
344 if (cpumask_test_cpu(cpu, &d->cpu_mask)) {
345 err = resctrl_arch_rmid_read(r, d, entry->rmid,
346 QOS_L3_OCCUP_EVENT_ID,
348 if (err || val <= resctrl_rmid_realloc_threshold)
353 * For the first limbo RMID in the domain,
354 * setup up the limbo worker.
356 if (!has_busy_rmid(r, d))
357 cqm_setup_limbo_handler(d, CQM_LIMBOCHECK_INTERVAL);
358 set_bit(entry->rmid, d->rmid_busy_llc);
366 list_add_tail(&entry->list, &rmid_free_lru);
369 void free_rmid(u32 rmid)
371 struct rmid_entry *entry;
376 lockdep_assert_held(&rdtgroup_mutex);
378 entry = __rmid_entry(rmid);
380 if (is_llc_occupancy_enabled())
381 add_rmid_to_limbo(entry);
383 list_add_tail(&entry->list, &rmid_free_lru);
386 static struct mbm_state *get_mbm_state(struct rdt_domain *d, u32 rmid,
387 enum resctrl_event_id evtid)
390 case QOS_L3_MBM_TOTAL_EVENT_ID:
391 return &d->mbm_total[rmid];
392 case QOS_L3_MBM_LOCAL_EVENT_ID:
393 return &d->mbm_local[rmid];
399 static int __mon_event_count(u32 rmid, struct rmid_read *rr)
405 resctrl_arch_reset_rmid(rr->r, rr->d, rmid, rr->evtid);
406 m = get_mbm_state(rr->d, rmid, rr->evtid);
408 memset(m, 0, sizeof(struct mbm_state));
412 rr->err = resctrl_arch_rmid_read(rr->r, rr->d, rmid, rr->evtid, &tval);
422 * mbm_bw_count() - Update bw count from values previously read by
423 * __mon_event_count().
424 * @rmid: The rmid used to identify the cached mbm_state.
425 * @rr: The struct rmid_read populated by __mon_event_count().
427 * Supporting function to calculate the memory bandwidth
428 * and delta bandwidth in MBps. The chunks value previously read by
429 * __mon_event_count() is compared with the chunks value from the previous
430 * invocation. This must be called once per second to maintain values in MBps.
432 static void mbm_bw_count(u32 rmid, struct rmid_read *rr)
434 struct mbm_state *m = &rr->d->mbm_local[rmid];
435 u64 cur_bw, bytes, cur_bytes;
438 bytes = cur_bytes - m->prev_bw_bytes;
439 m->prev_bw_bytes = cur_bytes;
441 cur_bw = bytes / SZ_1M;
444 m->delta_bw = abs(cur_bw - m->prev_bw);
445 m->delta_comp = false;
450 * This is called via IPI to read the CQM/MBM counters
453 void mon_event_count(void *info)
455 struct rdtgroup *rdtgrp, *entry;
456 struct rmid_read *rr = info;
457 struct list_head *head;
462 ret = __mon_event_count(rdtgrp->mon.rmid, rr);
465 * For Ctrl groups read data from child monitor groups and
466 * add them together. Count events which are read successfully.
467 * Discard the rmid_read's reporting errors.
469 head = &rdtgrp->mon.crdtgrp_list;
471 if (rdtgrp->type == RDTCTRL_GROUP) {
472 list_for_each_entry(entry, head, mon.crdtgrp_list) {
473 if (__mon_event_count(entry->mon.rmid, rr) == 0)
479 * __mon_event_count() calls for newly created monitor groups may
480 * report -EINVAL/Unavailable if the monitor hasn't seen any traffic.
481 * Discard error if any of the monitor event reads succeeded.
488 * Feedback loop for MBA software controller (mba_sc)
490 * mba_sc is a feedback loop where we periodically read MBM counters and
491 * adjust the bandwidth percentage values via the IA32_MBA_THRTL_MSRs so
494 * current bandwidth(cur_bw) < user specified bandwidth(user_bw)
496 * This uses the MBM counters to measure the bandwidth and MBA throttle
497 * MSRs to control the bandwidth for a particular rdtgrp. It builds on the
498 * fact that resctrl rdtgroups have both monitoring and control.
500 * The frequency of the checks is 1s and we just tag along the MBM overflow
501 * timer. Having 1s interval makes the calculation of bandwidth simpler.
503 * Although MBA's goal is to restrict the bandwidth to a maximum, there may
504 * be a need to increase the bandwidth to avoid unnecessarily restricting
505 * the L2 <-> L3 traffic.
507 * Since MBA controls the L2 external bandwidth where as MBM measures the
508 * L3 external bandwidth the following sequence could lead to such a
511 * Consider an rdtgroup which had high L3 <-> memory traffic in initial
512 * phases -> mba_sc kicks in and reduced bandwidth percentage values -> but
513 * after some time rdtgroup has mostly L2 <-> L3 traffic.
515 * In this case we may restrict the rdtgroup's L2 <-> L3 traffic as its
516 * throttle MSRs already have low percentage values. To avoid
517 * unnecessarily restricting such rdtgroups, we also increase the bandwidth.
519 static void update_mba_bw(struct rdtgroup *rgrp, struct rdt_domain *dom_mbm)
521 u32 closid, rmid, cur_msr_val, new_msr_val;
522 struct mbm_state *pmbm_data, *cmbm_data;
523 u32 cur_bw, delta_bw, user_bw;
524 struct rdt_resource *r_mba;
525 struct rdt_domain *dom_mba;
526 struct list_head *head;
527 struct rdtgroup *entry;
529 if (!is_mbm_local_enabled())
532 r_mba = &rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl;
534 closid = rgrp->closid;
535 rmid = rgrp->mon.rmid;
536 pmbm_data = &dom_mbm->mbm_local[rmid];
538 dom_mba = get_domain_from_cpu(smp_processor_id(), r_mba);
540 pr_warn_once("Failure to get domain for MBA update\n");
544 cur_bw = pmbm_data->prev_bw;
545 user_bw = dom_mba->mbps_val[closid];
546 delta_bw = pmbm_data->delta_bw;
548 /* MBA resource doesn't support CDP */
549 cur_msr_val = resctrl_arch_get_config(r_mba, dom_mba, closid, CDP_NONE);
552 * For Ctrl groups read data from child monitor groups.
554 head = &rgrp->mon.crdtgrp_list;
555 list_for_each_entry(entry, head, mon.crdtgrp_list) {
556 cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
557 cur_bw += cmbm_data->prev_bw;
558 delta_bw += cmbm_data->delta_bw;
562 * Scale up/down the bandwidth linearly for the ctrl group. The
563 * bandwidth step is the bandwidth granularity specified by the
566 * The delta_bw is used when increasing the bandwidth so that we
567 * dont alternately increase and decrease the control values
570 * For ex: consider cur_bw = 90MBps, user_bw = 100MBps and if
571 * bandwidth step is 20MBps(> user_bw - cur_bw), we would keep
572 * switching between 90 and 110 continuously if we only check
575 if (cur_msr_val > r_mba->membw.min_bw && user_bw < cur_bw) {
576 new_msr_val = cur_msr_val - r_mba->membw.bw_gran;
577 } else if (cur_msr_val < MAX_MBA_BW &&
578 (user_bw > (cur_bw + delta_bw))) {
579 new_msr_val = cur_msr_val + r_mba->membw.bw_gran;
584 resctrl_arch_update_one(r_mba, dom_mba, closid, CDP_NONE, new_msr_val);
587 * Delta values are updated dynamically package wise for each
588 * rdtgrp every time the throttle MSR changes value.
590 * This is because (1)the increase in bandwidth is not perfectly
591 * linear and only "approximately" linear even when the hardware
592 * says it is linear.(2)Also since MBA is a core specific
593 * mechanism, the delta values vary based on number of cores used
596 pmbm_data->delta_comp = true;
597 list_for_each_entry(entry, head, mon.crdtgrp_list) {
598 cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
599 cmbm_data->delta_comp = true;
603 static void mbm_update(struct rdt_resource *r, struct rdt_domain *d, int rmid)
612 * This is protected from concurrent reads from user
613 * as both the user and we hold the global mutex.
615 if (is_mbm_total_enabled()) {
616 rr.evtid = QOS_L3_MBM_TOTAL_EVENT_ID;
618 __mon_event_count(rmid, &rr);
620 if (is_mbm_local_enabled()) {
621 rr.evtid = QOS_L3_MBM_LOCAL_EVENT_ID;
623 __mon_event_count(rmid, &rr);
626 * Call the MBA software controller only for the
627 * control groups and when user has enabled
628 * the software controller explicitly.
631 mbm_bw_count(rmid, &rr);
636 * Handler to scan the limbo list and move the RMIDs
637 * to free list whose occupancy < threshold_occupancy.
639 void cqm_handle_limbo(struct work_struct *work)
641 unsigned long delay = msecs_to_jiffies(CQM_LIMBOCHECK_INTERVAL);
642 int cpu = smp_processor_id();
643 struct rdt_resource *r;
644 struct rdt_domain *d;
646 mutex_lock(&rdtgroup_mutex);
648 r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
649 d = container_of(work, struct rdt_domain, cqm_limbo.work);
651 __check_limbo(d, false);
653 if (has_busy_rmid(r, d))
654 schedule_delayed_work_on(cpu, &d->cqm_limbo, delay);
656 mutex_unlock(&rdtgroup_mutex);
659 void cqm_setup_limbo_handler(struct rdt_domain *dom, unsigned long delay_ms)
661 unsigned long delay = msecs_to_jiffies(delay_ms);
664 cpu = cpumask_any(&dom->cpu_mask);
665 dom->cqm_work_cpu = cpu;
667 schedule_delayed_work_on(cpu, &dom->cqm_limbo, delay);
670 void mbm_handle_overflow(struct work_struct *work)
672 unsigned long delay = msecs_to_jiffies(MBM_OVERFLOW_INTERVAL);
673 struct rdtgroup *prgrp, *crgrp;
674 int cpu = smp_processor_id();
675 struct list_head *head;
676 struct rdt_resource *r;
677 struct rdt_domain *d;
679 mutex_lock(&rdtgroup_mutex);
681 if (!static_branch_likely(&rdt_mon_enable_key))
684 r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
685 d = container_of(work, struct rdt_domain, mbm_over.work);
687 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
688 mbm_update(r, d, prgrp->mon.rmid);
690 head = &prgrp->mon.crdtgrp_list;
691 list_for_each_entry(crgrp, head, mon.crdtgrp_list)
692 mbm_update(r, d, crgrp->mon.rmid);
695 update_mba_bw(prgrp, d);
698 schedule_delayed_work_on(cpu, &d->mbm_over, delay);
701 mutex_unlock(&rdtgroup_mutex);
704 void mbm_setup_overflow_handler(struct rdt_domain *dom, unsigned long delay_ms)
706 unsigned long delay = msecs_to_jiffies(delay_ms);
709 if (!static_branch_likely(&rdt_mon_enable_key))
711 cpu = cpumask_any(&dom->cpu_mask);
712 dom->mbm_work_cpu = cpu;
713 schedule_delayed_work_on(cpu, &dom->mbm_over, delay);
716 static int dom_data_init(struct rdt_resource *r)
718 struct rmid_entry *entry = NULL;
721 nr_rmids = r->num_rmid;
722 rmid_ptrs = kcalloc(nr_rmids, sizeof(struct rmid_entry), GFP_KERNEL);
726 for (i = 0; i < nr_rmids; i++) {
727 entry = &rmid_ptrs[i];
728 INIT_LIST_HEAD(&entry->list);
731 list_add_tail(&entry->list, &rmid_free_lru);
735 * RMID 0 is special and is always allocated. It's used for all
736 * tasks that are not monitored.
738 entry = __rmid_entry(0);
739 list_del(&entry->list);
744 static struct mon_evt llc_occupancy_event = {
745 .name = "llc_occupancy",
746 .evtid = QOS_L3_OCCUP_EVENT_ID,
749 static struct mon_evt mbm_total_event = {
750 .name = "mbm_total_bytes",
751 .evtid = QOS_L3_MBM_TOTAL_EVENT_ID,
754 static struct mon_evt mbm_local_event = {
755 .name = "mbm_local_bytes",
756 .evtid = QOS_L3_MBM_LOCAL_EVENT_ID,
760 * Initialize the event list for the resource.
762 * Note that MBM events are also part of RDT_RESOURCE_L3 resource
763 * because as per the SDM the total and local memory bandwidth
764 * are enumerated as part of L3 monitoring.
766 static void l3_mon_evt_init(struct rdt_resource *r)
768 INIT_LIST_HEAD(&r->evt_list);
770 if (is_llc_occupancy_enabled())
771 list_add_tail(&llc_occupancy_event.list, &r->evt_list);
772 if (is_mbm_total_enabled())
773 list_add_tail(&mbm_total_event.list, &r->evt_list);
774 if (is_mbm_local_enabled())
775 list_add_tail(&mbm_local_event.list, &r->evt_list);
778 int __init rdt_get_mon_l3_config(struct rdt_resource *r)
780 unsigned int mbm_offset = boot_cpu_data.x86_cache_mbm_width_offset;
781 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
782 unsigned int threshold;
785 resctrl_rmid_realloc_limit = boot_cpu_data.x86_cache_size * 1024;
786 hw_res->mon_scale = boot_cpu_data.x86_cache_occ_scale;
787 r->num_rmid = boot_cpu_data.x86_cache_max_rmid + 1;
788 hw_res->mbm_width = MBM_CNTR_WIDTH_BASE;
790 if (mbm_offset > 0 && mbm_offset <= MBM_CNTR_WIDTH_OFFSET_MAX)
791 hw_res->mbm_width += mbm_offset;
792 else if (mbm_offset > MBM_CNTR_WIDTH_OFFSET_MAX)
793 pr_warn("Ignoring impossible MBM counter offset\n");
796 * A reasonable upper limit on the max threshold is the number
797 * of lines tagged per RMID if all RMIDs have the same number of
798 * lines tagged in the LLC.
800 * For a 35MB LLC and 56 RMIDs, this is ~1.8% of the LLC.
802 threshold = resctrl_rmid_realloc_limit / r->num_rmid;
805 * Because num_rmid may not be a power of two, round the value
806 * to the nearest multiple of hw_res->mon_scale so it matches a
807 * value the hardware will measure. mon_scale may not be a power of 2.
809 resctrl_rmid_realloc_threshold = resctrl_arch_round_mon_val(threshold);
811 ret = dom_data_init(r);
815 if (rdt_cpu_has(X86_FEATURE_BMEC)) {
816 if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL)) {
817 mbm_total_event.configurable = true;
818 mbm_config_rftype_init("mbm_total_bytes_config");
820 if (rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL)) {
821 mbm_local_event.configurable = true;
822 mbm_config_rftype_init("mbm_local_bytes_config");
828 r->mon_capable = true;
833 void __init intel_rdt_mbm_apply_quirk(void)
837 cf_index = (boot_cpu_data.x86_cache_max_rmid + 1) / 8 - 1;
838 if (cf_index >= ARRAY_SIZE(mbm_cf_table)) {
839 pr_info("No MBM correction factor available\n");
843 mbm_cf_rmidthreshold = mbm_cf_table[cf_index].rmidthreshold;
844 mbm_cf = mbm_cf_table[cf_index].cf;