2 * Resource Director Technology(RDT)
5 * Copyright (C) 2017 Intel Corporation
8 * Vikas Shivappa <vikas.shivappa@intel.com>
10 * This replaces the cqm.c based on perf but we reuse a lot of
11 * code and datastructures originally from Peter Zijlstra and Matt Fleming.
13 * This program is free software; you can redistribute it and/or modify it
14 * under the terms and conditions of the GNU General Public License,
15 * version 2, as published by the Free Software Foundation.
17 * This program is distributed in the hope it will be useful, but WITHOUT
18 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
19 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
22 * More information about RDT be found in the Intel (R) x86 Architecture
23 * Software Developer Manual June 2016, volume 3, section 17.17.
26 #include <linux/module.h>
27 #include <linux/slab.h>
28 #include <asm/cpu_device_id.h>
29 #include "intel_rdt.h"
31 #define MSR_IA32_QM_CTR 0x0c8e
32 #define MSR_IA32_QM_EVTSEL 0x0c8d
37 struct list_head list;
41 * @rmid_free_lru A least recently used list of free RMIDs
42 * These RMIDs are guaranteed to have an occupancy less than the
45 static LIST_HEAD(rmid_free_lru);
48 * @rmid_limbo_count count of currently unused but (potentially)
50 * This counts RMIDs that no one is currently using but that
51 * may have a occupancy value > intel_cqm_threshold. User can change
52 * the threshold occupancy value.
54 unsigned int rmid_limbo_count;
57 * @rmid_entry - The entry in the limbo and free lists.
59 static struct rmid_entry *rmid_ptrs;
62 * Global boolean for rdt_monitor which is true if any
63 * resource monitoring is enabled.
68 * Global to indicate which monitoring events are enabled.
70 unsigned int rdt_mon_features;
73 * This is the threshold cache occupancy at which we will consider an
74 * RMID available for re-allocation.
76 unsigned int intel_cqm_threshold;
78 static inline struct rmid_entry *__rmid_entry(u32 rmid)
80 struct rmid_entry *entry;
82 entry = &rmid_ptrs[rmid];
83 WARN_ON(entry->rmid != rmid);
88 static u64 __rmid_read(u32 rmid, u32 eventid)
93 * As per the SDM, when IA32_QM_EVTSEL.EvtID (bits 7:0) is configured
94 * with a valid event code for supported resource type and the bits
95 * IA32_QM_EVTSEL.RMID (bits 41:32) are configured with valid RMID,
96 * IA32_QM_CTR.data (bits 61:0) reports the monitored data.
97 * IA32_QM_CTR.Error (bit 63) and IA32_QM_CTR.Unavailable (bit 62)
100 wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid);
101 rdmsrl(MSR_IA32_QM_CTR, val);
106 static bool rmid_dirty(struct rmid_entry *entry)
108 u64 val = __rmid_read(entry->rmid, QOS_L3_OCCUP_EVENT_ID);
110 return val >= intel_cqm_threshold;
114 * Check the RMIDs that are marked as busy for this domain. If the
115 * reported LLC occupancy is below the threshold clear the busy bit and
116 * decrement the count. If the busy count gets to zero on an RMID, we
119 void __check_limbo(struct rdt_domain *d, bool force_free)
121 struct rmid_entry *entry;
122 struct rdt_resource *r;
123 u32 crmid = 1, nrmid;
125 r = &rdt_resources_all[RDT_RESOURCE_L3];
128 * Skip RMID 0 and start from RMID 1 and check all the RMIDs that
129 * are marked as busy for occupancy < threshold. If the occupancy
130 * is less than the threshold decrement the busy counter of the
131 * RMID and move it to the free list when the counter reaches 0.
134 nrmid = find_next_bit(d->rmid_busy_llc, r->num_rmid, crmid);
135 if (nrmid >= r->num_rmid)
138 entry = __rmid_entry(nrmid);
139 if (force_free || !rmid_dirty(entry)) {
140 clear_bit(entry->rmid, d->rmid_busy_llc);
141 if (!--entry->busy) {
143 list_add_tail(&entry->list, &rmid_free_lru);
150 bool has_busy_rmid(struct rdt_resource *r, struct rdt_domain *d)
152 return find_first_bit(d->rmid_busy_llc, r->num_rmid) != r->num_rmid;
156 * As of now the RMIDs allocation is global.
157 * However we keep track of which packages the RMIDs
158 * are used to optimize the limbo list management.
162 struct rmid_entry *entry;
164 lockdep_assert_held(&rdtgroup_mutex);
166 if (list_empty(&rmid_free_lru))
167 return rmid_limbo_count ? -EBUSY : -ENOSPC;
169 entry = list_first_entry(&rmid_free_lru,
170 struct rmid_entry, list);
171 list_del(&entry->list);
176 static void add_rmid_to_limbo(struct rmid_entry *entry)
178 struct rdt_resource *r;
179 struct rdt_domain *d;
183 r = &rdt_resources_all[RDT_RESOURCE_L3];
187 list_for_each_entry(d, &r->domains, list) {
188 if (cpumask_test_cpu(cpu, &d->cpu_mask)) {
189 val = __rmid_read(entry->rmid, QOS_L3_OCCUP_EVENT_ID);
190 if (val <= intel_cqm_threshold)
195 * For the first limbo RMID in the domain,
196 * setup up the limbo worker.
198 if (!has_busy_rmid(r, d))
199 cqm_setup_limbo_handler(d, CQM_LIMBOCHECK_INTERVAL);
200 set_bit(entry->rmid, d->rmid_busy_llc);
208 list_add_tail(&entry->list, &rmid_free_lru);
211 void free_rmid(u32 rmid)
213 struct rmid_entry *entry;
218 lockdep_assert_held(&rdtgroup_mutex);
220 entry = __rmid_entry(rmid);
222 if (is_llc_occupancy_enabled())
223 add_rmid_to_limbo(entry);
225 list_add_tail(&entry->list, &rmid_free_lru);
228 static u64 __mon_event_count(u32 rmid, struct rmid_read *rr)
230 u64 chunks, shift, tval;
233 tval = __rmid_read(rmid, rr->evtid);
234 if (tval & (RMID_VAL_ERROR | RMID_VAL_UNAVAIL)) {
238 case QOS_L3_OCCUP_EVENT_ID:
241 case QOS_L3_MBM_TOTAL_EVENT_ID:
242 m = &rr->d->mbm_total[rmid];
244 case QOS_L3_MBM_LOCAL_EVENT_ID:
245 m = &rr->d->mbm_local[rmid];
249 * Code would never reach here because an invalid
250 * event id would fail the __rmid_read.
252 return RMID_VAL_ERROR;
261 shift = 64 - MBM_CNTR_WIDTH;
262 chunks = (tval << shift) - (m->prev_msr << shift);
267 rr->val += m->chunks;
272 * This is called via IPI to read the CQM/MBM counters
275 void mon_event_count(void *info)
277 struct rdtgroup *rdtgrp, *entry;
278 struct rmid_read *rr = info;
279 struct list_head *head;
284 ret_val = __mon_event_count(rdtgrp->mon.rmid, rr);
287 * For Ctrl groups read data from child monitor groups and
288 * add them together. Count events which are read successfully.
289 * Discard the rmid_read's reporting errors.
291 head = &rdtgrp->mon.crdtgrp_list;
293 if (rdtgrp->type == RDTCTRL_GROUP) {
294 list_for_each_entry(entry, head, mon.crdtgrp_list) {
295 if (__mon_event_count(entry->mon.rmid, rr) == 0)
300 /* Report error if none of rmid_reads are successful */
305 static void mbm_update(struct rdt_domain *d, int rmid)
313 * This is protected from concurrent reads from user
314 * as both the user and we hold the global mutex.
316 if (is_mbm_total_enabled()) {
317 rr.evtid = QOS_L3_MBM_TOTAL_EVENT_ID;
318 __mon_event_count(rmid, &rr);
320 if (is_mbm_local_enabled()) {
321 rr.evtid = QOS_L3_MBM_LOCAL_EVENT_ID;
322 __mon_event_count(rmid, &rr);
327 * Handler to scan the limbo list and move the RMIDs
328 * to free list whose occupancy < threshold_occupancy.
330 void cqm_handle_limbo(struct work_struct *work)
332 unsigned long delay = msecs_to_jiffies(CQM_LIMBOCHECK_INTERVAL);
333 int cpu = smp_processor_id();
334 struct rdt_resource *r;
335 struct rdt_domain *d;
337 mutex_lock(&rdtgroup_mutex);
339 r = &rdt_resources_all[RDT_RESOURCE_L3];
340 d = get_domain_from_cpu(cpu, r);
343 pr_warn_once("Failure to get domain for limbo worker\n");
347 __check_limbo(d, false);
349 if (has_busy_rmid(r, d))
350 schedule_delayed_work_on(cpu, &d->cqm_limbo, delay);
353 mutex_unlock(&rdtgroup_mutex);
356 void cqm_setup_limbo_handler(struct rdt_domain *dom, unsigned long delay_ms)
358 unsigned long delay = msecs_to_jiffies(delay_ms);
359 struct rdt_resource *r;
362 r = &rdt_resources_all[RDT_RESOURCE_L3];
364 cpu = cpumask_any(&dom->cpu_mask);
365 dom->cqm_work_cpu = cpu;
367 schedule_delayed_work_on(cpu, &dom->cqm_limbo, delay);
370 void mbm_handle_overflow(struct work_struct *work)
372 unsigned long delay = msecs_to_jiffies(MBM_OVERFLOW_INTERVAL);
373 struct rdtgroup *prgrp, *crgrp;
374 int cpu = smp_processor_id();
375 struct list_head *head;
376 struct rdt_domain *d;
378 mutex_lock(&rdtgroup_mutex);
380 if (!static_branch_likely(&rdt_enable_key))
383 d = get_domain_from_cpu(cpu, &rdt_resources_all[RDT_RESOURCE_L3]);
387 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
388 mbm_update(d, prgrp->mon.rmid);
390 head = &prgrp->mon.crdtgrp_list;
391 list_for_each_entry(crgrp, head, mon.crdtgrp_list)
392 mbm_update(d, crgrp->mon.rmid);
395 schedule_delayed_work_on(cpu, &d->mbm_over, delay);
398 mutex_unlock(&rdtgroup_mutex);
401 void mbm_setup_overflow_handler(struct rdt_domain *dom, unsigned long delay_ms)
403 unsigned long delay = msecs_to_jiffies(delay_ms);
406 if (!static_branch_likely(&rdt_enable_key))
408 cpu = cpumask_any(&dom->cpu_mask);
409 dom->mbm_work_cpu = cpu;
410 schedule_delayed_work_on(cpu, &dom->mbm_over, delay);
413 static int dom_data_init(struct rdt_resource *r)
415 struct rmid_entry *entry = NULL;
418 nr_rmids = r->num_rmid;
419 rmid_ptrs = kcalloc(nr_rmids, sizeof(struct rmid_entry), GFP_KERNEL);
423 for (i = 0; i < nr_rmids; i++) {
424 entry = &rmid_ptrs[i];
425 INIT_LIST_HEAD(&entry->list);
428 list_add_tail(&entry->list, &rmid_free_lru);
432 * RMID 0 is special and is always allocated. It's used for all
433 * tasks that are not monitored.
435 entry = __rmid_entry(0);
436 list_del(&entry->list);
441 static struct mon_evt llc_occupancy_event = {
442 .name = "llc_occupancy",
443 .evtid = QOS_L3_OCCUP_EVENT_ID,
446 static struct mon_evt mbm_total_event = {
447 .name = "mbm_total_bytes",
448 .evtid = QOS_L3_MBM_TOTAL_EVENT_ID,
451 static struct mon_evt mbm_local_event = {
452 .name = "mbm_local_bytes",
453 .evtid = QOS_L3_MBM_LOCAL_EVENT_ID,
457 * Initialize the event list for the resource.
459 * Note that MBM events are also part of RDT_RESOURCE_L3 resource
460 * because as per the SDM the total and local memory bandwidth
461 * are enumerated as part of L3 monitoring.
463 static void l3_mon_evt_init(struct rdt_resource *r)
465 INIT_LIST_HEAD(&r->evt_list);
467 if (is_llc_occupancy_enabled())
468 list_add_tail(&llc_occupancy_event.list, &r->evt_list);
469 if (is_mbm_total_enabled())
470 list_add_tail(&mbm_total_event.list, &r->evt_list);
471 if (is_mbm_local_enabled())
472 list_add_tail(&mbm_local_event.list, &r->evt_list);
475 int rdt_get_mon_l3_config(struct rdt_resource *r)
479 r->mon_scale = boot_cpu_data.x86_cache_occ_scale;
480 r->num_rmid = boot_cpu_data.x86_cache_max_rmid + 1;
483 * A reasonable upper limit on the max threshold is the number
484 * of lines tagged per RMID if all RMIDs have the same number of
485 * lines tagged in the LLC.
487 * For a 35MB LLC and 56 RMIDs, this is ~1.8% of the LLC.
489 intel_cqm_threshold = boot_cpu_data.x86_cache_size * 1024 / r->num_rmid;
491 /* h/w works in units of "boot_cpu_data.x86_cache_occ_scale" */
492 intel_cqm_threshold /= r->mon_scale;
494 ret = dom_data_init(r);
500 r->mon_capable = true;
501 r->mon_enabled = true;