Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Resource Director Technology(RDT)
4 * - Monitoring code
5 *
6 * Copyright (C) 2017 Intel Corporation
7 *
8 * Author:
9 * Vikas Shivappa <vikas.shivappa@intel.com>
10 *
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.
13 *
14 * More information about RDT be found in the Intel (R) x86 Architecture
15 * Software Developer Manual June 2016, volume 3, section 17.17.
16 */
17
18#include <linux/module.h>
19#include <linux/slab.h>
20#include <asm/cpu_device_id.h>
21#include "internal.h"
22
23struct rmid_entry {
24 u32 rmid;
25 int busy;
26 struct list_head list;
27};
28
29/**
30 * @rmid_free_lru A least recently used list of free RMIDs
31 * These RMIDs are guaranteed to have an occupancy less than the
32 * threshold occupancy
33 */
34static LIST_HEAD(rmid_free_lru);
35
36/**
37 * @rmid_limbo_count count of currently unused but (potentially)
38 * dirty RMIDs.
39 * This counts RMIDs that no one is currently using but that
40 * may have a occupancy value > intel_cqm_threshold. User can change
41 * the threshold occupancy value.
42 */
43static unsigned int rmid_limbo_count;
44
45/**
46 * @rmid_entry - The entry in the limbo and free lists.
47 */
48static struct rmid_entry *rmid_ptrs;
49
50/*
51 * Global boolean for rdt_monitor which is true if any
52 * resource monitoring is enabled.
53 */
54bool rdt_mon_capable;
55
56/*
57 * Global to indicate which monitoring events are enabled.
58 */
59unsigned int rdt_mon_features;
60
61/*
62 * This is the threshold cache occupancy at which we will consider an
63 * RMID available for re-allocation.
64 */
65unsigned int resctrl_cqm_threshold;
66
67static inline struct rmid_entry *__rmid_entry(u32 rmid)
68{
69 struct rmid_entry *entry;
70
71 entry = &rmid_ptrs[rmid];
72 WARN_ON(entry->rmid != rmid);
73
74 return entry;
75}
76
77static u64 __rmid_read(u32 rmid, u32 eventid)
78{
79 u64 val;
80
81 /*
82 * As per the SDM, when IA32_QM_EVTSEL.EvtID (bits 7:0) is configured
83 * with a valid event code for supported resource type and the bits
84 * IA32_QM_EVTSEL.RMID (bits 41:32) are configured with valid RMID,
85 * IA32_QM_CTR.data (bits 61:0) reports the monitored data.
86 * IA32_QM_CTR.Error (bit 63) and IA32_QM_CTR.Unavailable (bit 62)
87 * are error bits.
88 */
89 wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid);
90 rdmsrl(MSR_IA32_QM_CTR, val);
91
92 return val;
93}
94
95static bool rmid_dirty(struct rmid_entry *entry)
96{
97 u64 val = __rmid_read(entry->rmid, QOS_L3_OCCUP_EVENT_ID);
98
99 return val >= resctrl_cqm_threshold;
100}
101
102/*
103 * Check the RMIDs that are marked as busy for this domain. If the
104 * reported LLC occupancy is below the threshold clear the busy bit and
105 * decrement the count. If the busy count gets to zero on an RMID, we
106 * free the RMID
107 */
108void __check_limbo(struct rdt_domain *d, bool force_free)
109{
110 struct rmid_entry *entry;
111 struct rdt_resource *r;
112 u32 crmid = 1, nrmid;
113
114 r = &rdt_resources_all[RDT_RESOURCE_L3];
115
116 /*
117 * Skip RMID 0 and start from RMID 1 and check all the RMIDs that
118 * are marked as busy for occupancy < threshold. If the occupancy
119 * is less than the threshold decrement the busy counter of the
120 * RMID and move it to the free list when the counter reaches 0.
121 */
122 for (;;) {
123 nrmid = find_next_bit(d->rmid_busy_llc, r->num_rmid, crmid);
124 if (nrmid >= r->num_rmid)
125 break;
126
127 entry = __rmid_entry(nrmid);
128 if (force_free || !rmid_dirty(entry)) {
129 clear_bit(entry->rmid, d->rmid_busy_llc);
130 if (!--entry->busy) {
131 rmid_limbo_count--;
132 list_add_tail(&entry->list, &rmid_free_lru);
133 }
134 }
135 crmid = nrmid + 1;
136 }
137}
138
139bool has_busy_rmid(struct rdt_resource *r, struct rdt_domain *d)
140{
141 return find_first_bit(d->rmid_busy_llc, r->num_rmid) != r->num_rmid;
142}
143
144/*
145 * As of now the RMIDs allocation is global.
146 * However we keep track of which packages the RMIDs
147 * are used to optimize the limbo list management.
148 */
149int alloc_rmid(void)
150{
151 struct rmid_entry *entry;
152
153 lockdep_assert_held(&rdtgroup_mutex);
154
155 if (list_empty(&rmid_free_lru))
156 return rmid_limbo_count ? -EBUSY : -ENOSPC;
157
158 entry = list_first_entry(&rmid_free_lru,
159 struct rmid_entry, list);
160 list_del(&entry->list);
161
162 return entry->rmid;
163}
164
165static void add_rmid_to_limbo(struct rmid_entry *entry)
166{
167 struct rdt_resource *r;
168 struct rdt_domain *d;
169 int cpu;
170 u64 val;
171
172 r = &rdt_resources_all[RDT_RESOURCE_L3];
173
174 entry->busy = 0;
175 cpu = get_cpu();
176 list_for_each_entry(d, &r->domains, list) {
177 if (cpumask_test_cpu(cpu, &d->cpu_mask)) {
178 val = __rmid_read(entry->rmid, QOS_L3_OCCUP_EVENT_ID);
179 if (val <= resctrl_cqm_threshold)
180 continue;
181 }
182
183 /*
184 * For the first limbo RMID in the domain,
185 * setup up the limbo worker.
186 */
187 if (!has_busy_rmid(r, d))
188 cqm_setup_limbo_handler(d, CQM_LIMBOCHECK_INTERVAL);
189 set_bit(entry->rmid, d->rmid_busy_llc);
190 entry->busy++;
191 }
192 put_cpu();
193
194 if (entry->busy)
195 rmid_limbo_count++;
196 else
197 list_add_tail(&entry->list, &rmid_free_lru);
198}
199
200void free_rmid(u32 rmid)
201{
202 struct rmid_entry *entry;
203
204 if (!rmid)
205 return;
206
207 lockdep_assert_held(&rdtgroup_mutex);
208
209 entry = __rmid_entry(rmid);
210
211 if (is_llc_occupancy_enabled())
212 add_rmid_to_limbo(entry);
213 else
214 list_add_tail(&entry->list, &rmid_free_lru);
215}
216
217static u64 mbm_overflow_count(u64 prev_msr, u64 cur_msr)
218{
219 u64 shift = 64 - MBM_CNTR_WIDTH, chunks;
220
221 chunks = (cur_msr << shift) - (prev_msr << shift);
222 return chunks >>= shift;
223}
224
225static int __mon_event_count(u32 rmid, struct rmid_read *rr)
226{
227 struct mbm_state *m;
228 u64 chunks, tval;
229
230 tval = __rmid_read(rmid, rr->evtid);
231 if (tval & (RMID_VAL_ERROR | RMID_VAL_UNAVAIL)) {
232 rr->val = tval;
233 return -EINVAL;
234 }
235 switch (rr->evtid) {
236 case QOS_L3_OCCUP_EVENT_ID:
237 rr->val += tval;
238 return 0;
239 case QOS_L3_MBM_TOTAL_EVENT_ID:
240 m = &rr->d->mbm_total[rmid];
241 break;
242 case QOS_L3_MBM_LOCAL_EVENT_ID:
243 m = &rr->d->mbm_local[rmid];
244 break;
245 default:
246 /*
247 * Code would never reach here because
248 * an invalid event id would fail the __rmid_read.
249 */
250 return -EINVAL;
251 }
252
253 if (rr->first) {
254 memset(m, 0, sizeof(struct mbm_state));
255 m->prev_bw_msr = m->prev_msr = tval;
256 return 0;
257 }
258
259 chunks = mbm_overflow_count(m->prev_msr, tval);
260 m->chunks += chunks;
261 m->prev_msr = tval;
262
263 rr->val += m->chunks;
264 return 0;
265}
266
267/*
268 * Supporting function to calculate the memory bandwidth
269 * and delta bandwidth in MBps.
270 */
271static void mbm_bw_count(u32 rmid, struct rmid_read *rr)
272{
273 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3];
274 struct mbm_state *m = &rr->d->mbm_local[rmid];
275 u64 tval, cur_bw, chunks;
276
277 tval = __rmid_read(rmid, rr->evtid);
278 if (tval & (RMID_VAL_ERROR | RMID_VAL_UNAVAIL))
279 return;
280
281 chunks = mbm_overflow_count(m->prev_bw_msr, tval);
282 m->chunks_bw += chunks;
283 m->chunks = m->chunks_bw;
284 cur_bw = (chunks * r->mon_scale) >> 20;
285
286 if (m->delta_comp)
287 m->delta_bw = abs(cur_bw - m->prev_bw);
288 m->delta_comp = false;
289 m->prev_bw = cur_bw;
290 m->prev_bw_msr = tval;
291}
292
293/*
294 * This is called via IPI to read the CQM/MBM counters
295 * on a domain.
296 */
297void mon_event_count(void *info)
298{
299 struct rdtgroup *rdtgrp, *entry;
300 struct rmid_read *rr = info;
301 struct list_head *head;
302
303 rdtgrp = rr->rgrp;
304
305 if (__mon_event_count(rdtgrp->mon.rmid, rr))
306 return;
307
308 /*
309 * For Ctrl groups read data from child monitor groups.
310 */
311 head = &rdtgrp->mon.crdtgrp_list;
312
313 if (rdtgrp->type == RDTCTRL_GROUP) {
314 list_for_each_entry(entry, head, mon.crdtgrp_list) {
315 if (__mon_event_count(entry->mon.rmid, rr))
316 return;
317 }
318 }
319}
320
321/*
322 * Feedback loop for MBA software controller (mba_sc)
323 *
324 * mba_sc is a feedback loop where we periodically read MBM counters and
325 * adjust the bandwidth percentage values via the IA32_MBA_THRTL_MSRs so
326 * that:
327 *
328 * current bandwdith(cur_bw) < user specified bandwidth(user_bw)
329 *
330 * This uses the MBM counters to measure the bandwidth and MBA throttle
331 * MSRs to control the bandwidth for a particular rdtgrp. It builds on the
332 * fact that resctrl rdtgroups have both monitoring and control.
333 *
334 * The frequency of the checks is 1s and we just tag along the MBM overflow
335 * timer. Having 1s interval makes the calculation of bandwidth simpler.
336 *
337 * Although MBA's goal is to restrict the bandwidth to a maximum, there may
338 * be a need to increase the bandwidth to avoid uncecessarily restricting
339 * the L2 <-> L3 traffic.
340 *
341 * Since MBA controls the L2 external bandwidth where as MBM measures the
342 * L3 external bandwidth the following sequence could lead to such a
343 * situation.
344 *
345 * Consider an rdtgroup which had high L3 <-> memory traffic in initial
346 * phases -> mba_sc kicks in and reduced bandwidth percentage values -> but
347 * after some time rdtgroup has mostly L2 <-> L3 traffic.
348 *
349 * In this case we may restrict the rdtgroup's L2 <-> L3 traffic as its
350 * throttle MSRs already have low percentage values. To avoid
351 * unnecessarily restricting such rdtgroups, we also increase the bandwidth.
352 */
353static void update_mba_bw(struct rdtgroup *rgrp, struct rdt_domain *dom_mbm)
354{
355 u32 closid, rmid, cur_msr, cur_msr_val, new_msr_val;
356 struct mbm_state *pmbm_data, *cmbm_data;
357 u32 cur_bw, delta_bw, user_bw;
358 struct rdt_resource *r_mba;
359 struct rdt_domain *dom_mba;
360 struct list_head *head;
361 struct rdtgroup *entry;
362
363 if (!is_mbm_local_enabled())
364 return;
365
366 r_mba = &rdt_resources_all[RDT_RESOURCE_MBA];
367 closid = rgrp->closid;
368 rmid = rgrp->mon.rmid;
369 pmbm_data = &dom_mbm->mbm_local[rmid];
370
371 dom_mba = get_domain_from_cpu(smp_processor_id(), r_mba);
372 if (!dom_mba) {
373 pr_warn_once("Failure to get domain for MBA update\n");
374 return;
375 }
376
377 cur_bw = pmbm_data->prev_bw;
378 user_bw = dom_mba->mbps_val[closid];
379 delta_bw = pmbm_data->delta_bw;
380 cur_msr_val = dom_mba->ctrl_val[closid];
381
382 /*
383 * For Ctrl groups read data from child monitor groups.
384 */
385 head = &rgrp->mon.crdtgrp_list;
386 list_for_each_entry(entry, head, mon.crdtgrp_list) {
387 cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
388 cur_bw += cmbm_data->prev_bw;
389 delta_bw += cmbm_data->delta_bw;
390 }
391
392 /*
393 * Scale up/down the bandwidth linearly for the ctrl group. The
394 * bandwidth step is the bandwidth granularity specified by the
395 * hardware.
396 *
397 * The delta_bw is used when increasing the bandwidth so that we
398 * dont alternately increase and decrease the control values
399 * continuously.
400 *
401 * For ex: consider cur_bw = 90MBps, user_bw = 100MBps and if
402 * bandwidth step is 20MBps(> user_bw - cur_bw), we would keep
403 * switching between 90 and 110 continuously if we only check
404 * cur_bw < user_bw.
405 */
406 if (cur_msr_val > r_mba->membw.min_bw && user_bw < cur_bw) {
407 new_msr_val = cur_msr_val - r_mba->membw.bw_gran;
408 } else if (cur_msr_val < MAX_MBA_BW &&
409 (user_bw > (cur_bw + delta_bw))) {
410 new_msr_val = cur_msr_val + r_mba->membw.bw_gran;
411 } else {
412 return;
413 }
414
415 cur_msr = r_mba->msr_base + closid;
416 wrmsrl(cur_msr, delay_bw_map(new_msr_val, r_mba));
417 dom_mba->ctrl_val[closid] = new_msr_val;
418
419 /*
420 * Delta values are updated dynamically package wise for each
421 * rdtgrp everytime the throttle MSR changes value.
422 *
423 * This is because (1)the increase in bandwidth is not perfectly
424 * linear and only "approximately" linear even when the hardware
425 * says it is linear.(2)Also since MBA is a core specific
426 * mechanism, the delta values vary based on number of cores used
427 * by the rdtgrp.
428 */
429 pmbm_data->delta_comp = true;
430 list_for_each_entry(entry, head, mon.crdtgrp_list) {
431 cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
432 cmbm_data->delta_comp = true;
433 }
434}
435
436static void mbm_update(struct rdt_domain *d, int rmid)
437{
438 struct rmid_read rr;
439
440 rr.first = false;
441 rr.d = d;
442
443 /*
444 * This is protected from concurrent reads from user
445 * as both the user and we hold the global mutex.
446 */
447 if (is_mbm_total_enabled()) {
448 rr.evtid = QOS_L3_MBM_TOTAL_EVENT_ID;
449 __mon_event_count(rmid, &rr);
450 }
451 if (is_mbm_local_enabled()) {
452 rr.evtid = QOS_L3_MBM_LOCAL_EVENT_ID;
453
454 /*
455 * Call the MBA software controller only for the
456 * control groups and when user has enabled
457 * the software controller explicitly.
458 */
459 if (!is_mba_sc(NULL))
460 __mon_event_count(rmid, &rr);
461 else
462 mbm_bw_count(rmid, &rr);
463 }
464}
465
466/*
467 * Handler to scan the limbo list and move the RMIDs
468 * to free list whose occupancy < threshold_occupancy.
469 */
470void cqm_handle_limbo(struct work_struct *work)
471{
472 unsigned long delay = msecs_to_jiffies(CQM_LIMBOCHECK_INTERVAL);
473 int cpu = smp_processor_id();
474 struct rdt_resource *r;
475 struct rdt_domain *d;
476
477 mutex_lock(&rdtgroup_mutex);
478
479 r = &rdt_resources_all[RDT_RESOURCE_L3];
480 d = get_domain_from_cpu(cpu, r);
481
482 if (!d) {
483 pr_warn_once("Failure to get domain for limbo worker\n");
484 goto out_unlock;
485 }
486
487 __check_limbo(d, false);
488
489 if (has_busy_rmid(r, d))
490 schedule_delayed_work_on(cpu, &d->cqm_limbo, delay);
491
492out_unlock:
493 mutex_unlock(&rdtgroup_mutex);
494}
495
496void cqm_setup_limbo_handler(struct rdt_domain *dom, unsigned long delay_ms)
497{
498 unsigned long delay = msecs_to_jiffies(delay_ms);
499 int cpu;
500
501 cpu = cpumask_any(&dom->cpu_mask);
502 dom->cqm_work_cpu = cpu;
503
504 schedule_delayed_work_on(cpu, &dom->cqm_limbo, delay);
505}
506
507void mbm_handle_overflow(struct work_struct *work)
508{
509 unsigned long delay = msecs_to_jiffies(MBM_OVERFLOW_INTERVAL);
510 struct rdtgroup *prgrp, *crgrp;
511 int cpu = smp_processor_id();
512 struct list_head *head;
513 struct rdt_domain *d;
514
515 mutex_lock(&rdtgroup_mutex);
516
517 if (!static_branch_likely(&rdt_enable_key))
518 goto out_unlock;
519
520 d = get_domain_from_cpu(cpu, &rdt_resources_all[RDT_RESOURCE_L3]);
521 if (!d)
522 goto out_unlock;
523
524 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
525 mbm_update(d, prgrp->mon.rmid);
526
527 head = &prgrp->mon.crdtgrp_list;
528 list_for_each_entry(crgrp, head, mon.crdtgrp_list)
529 mbm_update(d, crgrp->mon.rmid);
530
531 if (is_mba_sc(NULL))
532 update_mba_bw(prgrp, d);
533 }
534
535 schedule_delayed_work_on(cpu, &d->mbm_over, delay);
536
537out_unlock:
538 mutex_unlock(&rdtgroup_mutex);
539}
540
541void mbm_setup_overflow_handler(struct rdt_domain *dom, unsigned long delay_ms)
542{
543 unsigned long delay = msecs_to_jiffies(delay_ms);
544 int cpu;
545
546 if (!static_branch_likely(&rdt_enable_key))
547 return;
548 cpu = cpumask_any(&dom->cpu_mask);
549 dom->mbm_work_cpu = cpu;
550 schedule_delayed_work_on(cpu, &dom->mbm_over, delay);
551}
552
553static int dom_data_init(struct rdt_resource *r)
554{
555 struct rmid_entry *entry = NULL;
556 int i, nr_rmids;
557
558 nr_rmids = r->num_rmid;
559 rmid_ptrs = kcalloc(nr_rmids, sizeof(struct rmid_entry), GFP_KERNEL);
560 if (!rmid_ptrs)
561 return -ENOMEM;
562
563 for (i = 0; i < nr_rmids; i++) {
564 entry = &rmid_ptrs[i];
565 INIT_LIST_HEAD(&entry->list);
566
567 entry->rmid = i;
568 list_add_tail(&entry->list, &rmid_free_lru);
569 }
570
571 /*
572 * RMID 0 is special and is always allocated. It's used for all
573 * tasks that are not monitored.
574 */
575 entry = __rmid_entry(0);
576 list_del(&entry->list);
577
578 return 0;
579}
580
581static struct mon_evt llc_occupancy_event = {
582 .name = "llc_occupancy",
583 .evtid = QOS_L3_OCCUP_EVENT_ID,
584};
585
586static struct mon_evt mbm_total_event = {
587 .name = "mbm_total_bytes",
588 .evtid = QOS_L3_MBM_TOTAL_EVENT_ID,
589};
590
591static struct mon_evt mbm_local_event = {
592 .name = "mbm_local_bytes",
593 .evtid = QOS_L3_MBM_LOCAL_EVENT_ID,
594};
595
596/*
597 * Initialize the event list for the resource.
598 *
599 * Note that MBM events are also part of RDT_RESOURCE_L3 resource
600 * because as per the SDM the total and local memory bandwidth
601 * are enumerated as part of L3 monitoring.
602 */
603static void l3_mon_evt_init(struct rdt_resource *r)
604{
605 INIT_LIST_HEAD(&r->evt_list);
606
607 if (is_llc_occupancy_enabled())
608 list_add_tail(&llc_occupancy_event.list, &r->evt_list);
609 if (is_mbm_total_enabled())
610 list_add_tail(&mbm_total_event.list, &r->evt_list);
611 if (is_mbm_local_enabled())
612 list_add_tail(&mbm_local_event.list, &r->evt_list);
613}
614
615int rdt_get_mon_l3_config(struct rdt_resource *r)
616{
617 unsigned int cl_size = boot_cpu_data.x86_cache_size;
618 int ret;
619
620 r->mon_scale = boot_cpu_data.x86_cache_occ_scale;
621 r->num_rmid = boot_cpu_data.x86_cache_max_rmid + 1;
622
623 /*
624 * A reasonable upper limit on the max threshold is the number
625 * of lines tagged per RMID if all RMIDs have the same number of
626 * lines tagged in the LLC.
627 *
628 * For a 35MB LLC and 56 RMIDs, this is ~1.8% of the LLC.
629 */
630 resctrl_cqm_threshold = cl_size * 1024 / r->num_rmid;
631
632 /* h/w works in units of "boot_cpu_data.x86_cache_occ_scale" */
633 resctrl_cqm_threshold /= r->mon_scale;
634
635 ret = dom_data_init(r);
636 if (ret)
637 return ret;
638
639 l3_mon_evt_init(r);
640
641 r->mon_capable = true;
642 r->mon_enabled = true;
643
644 return 0;
645}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Resource Director Technology(RDT)
4 * - Monitoring code
5 *
6 * Copyright (C) 2017 Intel Corporation
7 *
8 * Author:
9 * Vikas Shivappa <vikas.shivappa@intel.com>
10 *
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.
13 *
14 * More information about RDT be found in the Intel (R) x86 Architecture
15 * Software Developer Manual June 2016, volume 3, section 17.17.
16 */
17
18#include <linux/module.h>
19#include <linux/slab.h>
20#include <asm/cpu_device_id.h>
21#include "internal.h"
22
23struct rmid_entry {
24 u32 rmid;
25 int busy;
26 struct list_head list;
27};
28
29/**
30 * @rmid_free_lru A least recently used list of free RMIDs
31 * These RMIDs are guaranteed to have an occupancy less than the
32 * threshold occupancy
33 */
34static LIST_HEAD(rmid_free_lru);
35
36/**
37 * @rmid_limbo_count count of currently unused but (potentially)
38 * dirty RMIDs.
39 * This counts RMIDs that no one is currently using but that
40 * may have a occupancy value > intel_cqm_threshold. User can change
41 * the threshold occupancy value.
42 */
43static unsigned int rmid_limbo_count;
44
45/**
46 * @rmid_entry - The entry in the limbo and free lists.
47 */
48static struct rmid_entry *rmid_ptrs;
49
50/*
51 * Global boolean for rdt_monitor which is true if any
52 * resource monitoring is enabled.
53 */
54bool rdt_mon_capable;
55
56/*
57 * Global to indicate which monitoring events are enabled.
58 */
59unsigned int rdt_mon_features;
60
61/*
62 * This is the threshold cache occupancy at which we will consider an
63 * RMID available for re-allocation.
64 */
65unsigned int resctrl_cqm_threshold;
66
67#define CF(cf) ((unsigned long)(1048576 * (cf) + 0.5))
68
69/*
70 * The correction factor table is documented in Documentation/x86/resctrl.rst.
71 * If rmid > rmid threshold, MBM total and local values should be multiplied
72 * by the correction factor.
73 *
74 * The original table is modified for better code:
75 *
76 * 1. The threshold 0 is changed to rmid count - 1 so don't do correction
77 * for the case.
78 * 2. MBM total and local correction table indexed by core counter which is
79 * equal to (x86_cache_max_rmid + 1) / 8 - 1 and is from 0 up to 27.
80 * 3. The correction factor is normalized to 2^20 (1048576) so it's faster
81 * to calculate corrected value by shifting:
82 * corrected_value = (original_value * correction_factor) >> 20
83 */
84static const struct mbm_correction_factor_table {
85 u32 rmidthreshold;
86 u64 cf;
87} mbm_cf_table[] __initconst = {
88 {7, CF(1.000000)},
89 {15, CF(1.000000)},
90 {15, CF(0.969650)},
91 {31, CF(1.000000)},
92 {31, CF(1.066667)},
93 {31, CF(0.969650)},
94 {47, CF(1.142857)},
95 {63, CF(1.000000)},
96 {63, CF(1.185115)},
97 {63, CF(1.066553)},
98 {79, CF(1.454545)},
99 {95, CF(1.000000)},
100 {95, CF(1.230769)},
101 {95, CF(1.142857)},
102 {95, CF(1.066667)},
103 {127, CF(1.000000)},
104 {127, CF(1.254863)},
105 {127, CF(1.185255)},
106 {151, CF(1.000000)},
107 {127, CF(1.066667)},
108 {167, CF(1.000000)},
109 {159, CF(1.454334)},
110 {183, CF(1.000000)},
111 {127, CF(0.969744)},
112 {191, CF(1.280246)},
113 {191, CF(1.230921)},
114 {215, CF(1.000000)},
115 {191, CF(1.143118)},
116};
117
118static u32 mbm_cf_rmidthreshold __read_mostly = UINT_MAX;
119static u64 mbm_cf __read_mostly;
120
121static inline u64 get_corrected_mbm_count(u32 rmid, unsigned long val)
122{
123 /* Correct MBM value. */
124 if (rmid > mbm_cf_rmidthreshold)
125 val = (val * mbm_cf) >> 20;
126
127 return val;
128}
129
130static inline struct rmid_entry *__rmid_entry(u32 rmid)
131{
132 struct rmid_entry *entry;
133
134 entry = &rmid_ptrs[rmid];
135 WARN_ON(entry->rmid != rmid);
136
137 return entry;
138}
139
140static u64 __rmid_read(u32 rmid, u32 eventid)
141{
142 u64 val;
143
144 /*
145 * As per the SDM, when IA32_QM_EVTSEL.EvtID (bits 7:0) is configured
146 * with a valid event code for supported resource type and the bits
147 * IA32_QM_EVTSEL.RMID (bits 41:32) are configured with valid RMID,
148 * IA32_QM_CTR.data (bits 61:0) reports the monitored data.
149 * IA32_QM_CTR.Error (bit 63) and IA32_QM_CTR.Unavailable (bit 62)
150 * are error bits.
151 */
152 wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid);
153 rdmsrl(MSR_IA32_QM_CTR, val);
154
155 return val;
156}
157
158static bool rmid_dirty(struct rmid_entry *entry)
159{
160 u64 val = __rmid_read(entry->rmid, QOS_L3_OCCUP_EVENT_ID);
161
162 return val >= resctrl_cqm_threshold;
163}
164
165/*
166 * Check the RMIDs that are marked as busy for this domain. If the
167 * reported LLC occupancy is below the threshold clear the busy bit and
168 * decrement the count. If the busy count gets to zero on an RMID, we
169 * free the RMID
170 */
171void __check_limbo(struct rdt_domain *d, bool force_free)
172{
173 struct rmid_entry *entry;
174 struct rdt_resource *r;
175 u32 crmid = 1, nrmid;
176
177 r = &rdt_resources_all[RDT_RESOURCE_L3];
178
179 /*
180 * Skip RMID 0 and start from RMID 1 and check all the RMIDs that
181 * are marked as busy for occupancy < threshold. If the occupancy
182 * is less than the threshold decrement the busy counter of the
183 * RMID and move it to the free list when the counter reaches 0.
184 */
185 for (;;) {
186 nrmid = find_next_bit(d->rmid_busy_llc, r->num_rmid, crmid);
187 if (nrmid >= r->num_rmid)
188 break;
189
190 entry = __rmid_entry(nrmid);
191 if (force_free || !rmid_dirty(entry)) {
192 clear_bit(entry->rmid, d->rmid_busy_llc);
193 if (!--entry->busy) {
194 rmid_limbo_count--;
195 list_add_tail(&entry->list, &rmid_free_lru);
196 }
197 }
198 crmid = nrmid + 1;
199 }
200}
201
202bool has_busy_rmid(struct rdt_resource *r, struct rdt_domain *d)
203{
204 return find_first_bit(d->rmid_busy_llc, r->num_rmid) != r->num_rmid;
205}
206
207/*
208 * As of now the RMIDs allocation is global.
209 * However we keep track of which packages the RMIDs
210 * are used to optimize the limbo list management.
211 */
212int alloc_rmid(void)
213{
214 struct rmid_entry *entry;
215
216 lockdep_assert_held(&rdtgroup_mutex);
217
218 if (list_empty(&rmid_free_lru))
219 return rmid_limbo_count ? -EBUSY : -ENOSPC;
220
221 entry = list_first_entry(&rmid_free_lru,
222 struct rmid_entry, list);
223 list_del(&entry->list);
224
225 return entry->rmid;
226}
227
228static void add_rmid_to_limbo(struct rmid_entry *entry)
229{
230 struct rdt_resource *r;
231 struct rdt_domain *d;
232 int cpu;
233 u64 val;
234
235 r = &rdt_resources_all[RDT_RESOURCE_L3];
236
237 entry->busy = 0;
238 cpu = get_cpu();
239 list_for_each_entry(d, &r->domains, list) {
240 if (cpumask_test_cpu(cpu, &d->cpu_mask)) {
241 val = __rmid_read(entry->rmid, QOS_L3_OCCUP_EVENT_ID);
242 if (val <= resctrl_cqm_threshold)
243 continue;
244 }
245
246 /*
247 * For the first limbo RMID in the domain,
248 * setup up the limbo worker.
249 */
250 if (!has_busy_rmid(r, d))
251 cqm_setup_limbo_handler(d, CQM_LIMBOCHECK_INTERVAL);
252 set_bit(entry->rmid, d->rmid_busy_llc);
253 entry->busy++;
254 }
255 put_cpu();
256
257 if (entry->busy)
258 rmid_limbo_count++;
259 else
260 list_add_tail(&entry->list, &rmid_free_lru);
261}
262
263void free_rmid(u32 rmid)
264{
265 struct rmid_entry *entry;
266
267 if (!rmid)
268 return;
269
270 lockdep_assert_held(&rdtgroup_mutex);
271
272 entry = __rmid_entry(rmid);
273
274 if (is_llc_occupancy_enabled())
275 add_rmid_to_limbo(entry);
276 else
277 list_add_tail(&entry->list, &rmid_free_lru);
278}
279
280static u64 mbm_overflow_count(u64 prev_msr, u64 cur_msr, unsigned int width)
281{
282 u64 shift = 64 - width, chunks;
283
284 chunks = (cur_msr << shift) - (prev_msr << shift);
285 return chunks >>= shift;
286}
287
288static u64 __mon_event_count(u32 rmid, struct rmid_read *rr)
289{
290 struct mbm_state *m;
291 u64 chunks, tval;
292
293 tval = __rmid_read(rmid, rr->evtid);
294 if (tval & (RMID_VAL_ERROR | RMID_VAL_UNAVAIL)) {
295 return tval;
296 }
297 switch (rr->evtid) {
298 case QOS_L3_OCCUP_EVENT_ID:
299 rr->val += tval;
300 return 0;
301 case QOS_L3_MBM_TOTAL_EVENT_ID:
302 m = &rr->d->mbm_total[rmid];
303 break;
304 case QOS_L3_MBM_LOCAL_EVENT_ID:
305 m = &rr->d->mbm_local[rmid];
306 break;
307 default:
308 /*
309 * Code would never reach here because an invalid
310 * event id would fail the __rmid_read.
311 */
312 return RMID_VAL_ERROR;
313 }
314
315 if (rr->first) {
316 memset(m, 0, sizeof(struct mbm_state));
317 m->prev_bw_msr = m->prev_msr = tval;
318 return 0;
319 }
320
321 chunks = mbm_overflow_count(m->prev_msr, tval, rr->r->mbm_width);
322 m->chunks += chunks;
323 m->prev_msr = tval;
324
325 rr->val += get_corrected_mbm_count(rmid, m->chunks);
326
327 return 0;
328}
329
330/*
331 * Supporting function to calculate the memory bandwidth
332 * and delta bandwidth in MBps.
333 */
334static void mbm_bw_count(u32 rmid, struct rmid_read *rr)
335{
336 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3];
337 struct mbm_state *m = &rr->d->mbm_local[rmid];
338 u64 tval, cur_bw, chunks;
339
340 tval = __rmid_read(rmid, rr->evtid);
341 if (tval & (RMID_VAL_ERROR | RMID_VAL_UNAVAIL))
342 return;
343
344 chunks = mbm_overflow_count(m->prev_bw_msr, tval, rr->r->mbm_width);
345 cur_bw = (get_corrected_mbm_count(rmid, chunks) * r->mon_scale) >> 20;
346
347 if (m->delta_comp)
348 m->delta_bw = abs(cur_bw - m->prev_bw);
349 m->delta_comp = false;
350 m->prev_bw = cur_bw;
351 m->prev_bw_msr = tval;
352}
353
354/*
355 * This is called via IPI to read the CQM/MBM counters
356 * on a domain.
357 */
358void mon_event_count(void *info)
359{
360 struct rdtgroup *rdtgrp, *entry;
361 struct rmid_read *rr = info;
362 struct list_head *head;
363 u64 ret_val;
364
365 rdtgrp = rr->rgrp;
366
367 ret_val = __mon_event_count(rdtgrp->mon.rmid, rr);
368
369 /*
370 * For Ctrl groups read data from child monitor groups and
371 * add them together. Count events which are read successfully.
372 * Discard the rmid_read's reporting errors.
373 */
374 head = &rdtgrp->mon.crdtgrp_list;
375
376 if (rdtgrp->type == RDTCTRL_GROUP) {
377 list_for_each_entry(entry, head, mon.crdtgrp_list) {
378 if (__mon_event_count(entry->mon.rmid, rr) == 0)
379 ret_val = 0;
380 }
381 }
382
383 /* Report error if none of rmid_reads are successful */
384 if (ret_val)
385 rr->val = ret_val;
386}
387
388/*
389 * Feedback loop for MBA software controller (mba_sc)
390 *
391 * mba_sc is a feedback loop where we periodically read MBM counters and
392 * adjust the bandwidth percentage values via the IA32_MBA_THRTL_MSRs so
393 * that:
394 *
395 * current bandwidth(cur_bw) < user specified bandwidth(user_bw)
396 *
397 * This uses the MBM counters to measure the bandwidth and MBA throttle
398 * MSRs to control the bandwidth for a particular rdtgrp. It builds on the
399 * fact that resctrl rdtgroups have both monitoring and control.
400 *
401 * The frequency of the checks is 1s and we just tag along the MBM overflow
402 * timer. Having 1s interval makes the calculation of bandwidth simpler.
403 *
404 * Although MBA's goal is to restrict the bandwidth to a maximum, there may
405 * be a need to increase the bandwidth to avoid unnecessarily restricting
406 * the L2 <-> L3 traffic.
407 *
408 * Since MBA controls the L2 external bandwidth where as MBM measures the
409 * L3 external bandwidth the following sequence could lead to such a
410 * situation.
411 *
412 * Consider an rdtgroup which had high L3 <-> memory traffic in initial
413 * phases -> mba_sc kicks in and reduced bandwidth percentage values -> but
414 * after some time rdtgroup has mostly L2 <-> L3 traffic.
415 *
416 * In this case we may restrict the rdtgroup's L2 <-> L3 traffic as its
417 * throttle MSRs already have low percentage values. To avoid
418 * unnecessarily restricting such rdtgroups, we also increase the bandwidth.
419 */
420static void update_mba_bw(struct rdtgroup *rgrp, struct rdt_domain *dom_mbm)
421{
422 u32 closid, rmid, cur_msr, cur_msr_val, new_msr_val;
423 struct mbm_state *pmbm_data, *cmbm_data;
424 u32 cur_bw, delta_bw, user_bw;
425 struct rdt_resource *r_mba;
426 struct rdt_domain *dom_mba;
427 struct list_head *head;
428 struct rdtgroup *entry;
429
430 if (!is_mbm_local_enabled())
431 return;
432
433 r_mba = &rdt_resources_all[RDT_RESOURCE_MBA];
434 closid = rgrp->closid;
435 rmid = rgrp->mon.rmid;
436 pmbm_data = &dom_mbm->mbm_local[rmid];
437
438 dom_mba = get_domain_from_cpu(smp_processor_id(), r_mba);
439 if (!dom_mba) {
440 pr_warn_once("Failure to get domain for MBA update\n");
441 return;
442 }
443
444 cur_bw = pmbm_data->prev_bw;
445 user_bw = dom_mba->mbps_val[closid];
446 delta_bw = pmbm_data->delta_bw;
447 cur_msr_val = dom_mba->ctrl_val[closid];
448
449 /*
450 * For Ctrl groups read data from child monitor groups.
451 */
452 head = &rgrp->mon.crdtgrp_list;
453 list_for_each_entry(entry, head, mon.crdtgrp_list) {
454 cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
455 cur_bw += cmbm_data->prev_bw;
456 delta_bw += cmbm_data->delta_bw;
457 }
458
459 /*
460 * Scale up/down the bandwidth linearly for the ctrl group. The
461 * bandwidth step is the bandwidth granularity specified by the
462 * hardware.
463 *
464 * The delta_bw is used when increasing the bandwidth so that we
465 * dont alternately increase and decrease the control values
466 * continuously.
467 *
468 * For ex: consider cur_bw = 90MBps, user_bw = 100MBps and if
469 * bandwidth step is 20MBps(> user_bw - cur_bw), we would keep
470 * switching between 90 and 110 continuously if we only check
471 * cur_bw < user_bw.
472 */
473 if (cur_msr_val > r_mba->membw.min_bw && user_bw < cur_bw) {
474 new_msr_val = cur_msr_val - r_mba->membw.bw_gran;
475 } else if (cur_msr_val < MAX_MBA_BW &&
476 (user_bw > (cur_bw + delta_bw))) {
477 new_msr_val = cur_msr_val + r_mba->membw.bw_gran;
478 } else {
479 return;
480 }
481
482 cur_msr = r_mba->msr_base + closid;
483 wrmsrl(cur_msr, delay_bw_map(new_msr_val, r_mba));
484 dom_mba->ctrl_val[closid] = new_msr_val;
485
486 /*
487 * Delta values are updated dynamically package wise for each
488 * rdtgrp every time the throttle MSR changes value.
489 *
490 * This is because (1)the increase in bandwidth is not perfectly
491 * linear and only "approximately" linear even when the hardware
492 * says it is linear.(2)Also since MBA is a core specific
493 * mechanism, the delta values vary based on number of cores used
494 * by the rdtgrp.
495 */
496 pmbm_data->delta_comp = true;
497 list_for_each_entry(entry, head, mon.crdtgrp_list) {
498 cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
499 cmbm_data->delta_comp = true;
500 }
501}
502
503static void mbm_update(struct rdt_resource *r, struct rdt_domain *d, int rmid)
504{
505 struct rmid_read rr;
506
507 rr.first = false;
508 rr.r = r;
509 rr.d = d;
510
511 /*
512 * This is protected from concurrent reads from user
513 * as both the user and we hold the global mutex.
514 */
515 if (is_mbm_total_enabled()) {
516 rr.evtid = QOS_L3_MBM_TOTAL_EVENT_ID;
517 __mon_event_count(rmid, &rr);
518 }
519 if (is_mbm_local_enabled()) {
520 rr.evtid = QOS_L3_MBM_LOCAL_EVENT_ID;
521 __mon_event_count(rmid, &rr);
522
523 /*
524 * Call the MBA software controller only for the
525 * control groups and when user has enabled
526 * the software controller explicitly.
527 */
528 if (is_mba_sc(NULL))
529 mbm_bw_count(rmid, &rr);
530 }
531}
532
533/*
534 * Handler to scan the limbo list and move the RMIDs
535 * to free list whose occupancy < threshold_occupancy.
536 */
537void cqm_handle_limbo(struct work_struct *work)
538{
539 unsigned long delay = msecs_to_jiffies(CQM_LIMBOCHECK_INTERVAL);
540 int cpu = smp_processor_id();
541 struct rdt_resource *r;
542 struct rdt_domain *d;
543
544 mutex_lock(&rdtgroup_mutex);
545
546 r = &rdt_resources_all[RDT_RESOURCE_L3];
547 d = container_of(work, struct rdt_domain, cqm_limbo.work);
548
549 __check_limbo(d, false);
550
551 if (has_busy_rmid(r, d))
552 schedule_delayed_work_on(cpu, &d->cqm_limbo, delay);
553
554 mutex_unlock(&rdtgroup_mutex);
555}
556
557void cqm_setup_limbo_handler(struct rdt_domain *dom, unsigned long delay_ms)
558{
559 unsigned long delay = msecs_to_jiffies(delay_ms);
560 int cpu;
561
562 cpu = cpumask_any(&dom->cpu_mask);
563 dom->cqm_work_cpu = cpu;
564
565 schedule_delayed_work_on(cpu, &dom->cqm_limbo, delay);
566}
567
568void mbm_handle_overflow(struct work_struct *work)
569{
570 unsigned long delay = msecs_to_jiffies(MBM_OVERFLOW_INTERVAL);
571 struct rdtgroup *prgrp, *crgrp;
572 int cpu = smp_processor_id();
573 struct list_head *head;
574 struct rdt_resource *r;
575 struct rdt_domain *d;
576
577 mutex_lock(&rdtgroup_mutex);
578
579 if (!static_branch_likely(&rdt_mon_enable_key))
580 goto out_unlock;
581
582 r = &rdt_resources_all[RDT_RESOURCE_L3];
583 d = container_of(work, struct rdt_domain, mbm_over.work);
584
585 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
586 mbm_update(r, d, prgrp->mon.rmid);
587
588 head = &prgrp->mon.crdtgrp_list;
589 list_for_each_entry(crgrp, head, mon.crdtgrp_list)
590 mbm_update(r, d, crgrp->mon.rmid);
591
592 if (is_mba_sc(NULL))
593 update_mba_bw(prgrp, d);
594 }
595
596 schedule_delayed_work_on(cpu, &d->mbm_over, delay);
597
598out_unlock:
599 mutex_unlock(&rdtgroup_mutex);
600}
601
602void mbm_setup_overflow_handler(struct rdt_domain *dom, unsigned long delay_ms)
603{
604 unsigned long delay = msecs_to_jiffies(delay_ms);
605 int cpu;
606
607 if (!static_branch_likely(&rdt_mon_enable_key))
608 return;
609 cpu = cpumask_any(&dom->cpu_mask);
610 dom->mbm_work_cpu = cpu;
611 schedule_delayed_work_on(cpu, &dom->mbm_over, delay);
612}
613
614static int dom_data_init(struct rdt_resource *r)
615{
616 struct rmid_entry *entry = NULL;
617 int i, nr_rmids;
618
619 nr_rmids = r->num_rmid;
620 rmid_ptrs = kcalloc(nr_rmids, sizeof(struct rmid_entry), GFP_KERNEL);
621 if (!rmid_ptrs)
622 return -ENOMEM;
623
624 for (i = 0; i < nr_rmids; i++) {
625 entry = &rmid_ptrs[i];
626 INIT_LIST_HEAD(&entry->list);
627
628 entry->rmid = i;
629 list_add_tail(&entry->list, &rmid_free_lru);
630 }
631
632 /*
633 * RMID 0 is special and is always allocated. It's used for all
634 * tasks that are not monitored.
635 */
636 entry = __rmid_entry(0);
637 list_del(&entry->list);
638
639 return 0;
640}
641
642static struct mon_evt llc_occupancy_event = {
643 .name = "llc_occupancy",
644 .evtid = QOS_L3_OCCUP_EVENT_ID,
645};
646
647static struct mon_evt mbm_total_event = {
648 .name = "mbm_total_bytes",
649 .evtid = QOS_L3_MBM_TOTAL_EVENT_ID,
650};
651
652static struct mon_evt mbm_local_event = {
653 .name = "mbm_local_bytes",
654 .evtid = QOS_L3_MBM_LOCAL_EVENT_ID,
655};
656
657/*
658 * Initialize the event list for the resource.
659 *
660 * Note that MBM events are also part of RDT_RESOURCE_L3 resource
661 * because as per the SDM the total and local memory bandwidth
662 * are enumerated as part of L3 monitoring.
663 */
664static void l3_mon_evt_init(struct rdt_resource *r)
665{
666 INIT_LIST_HEAD(&r->evt_list);
667
668 if (is_llc_occupancy_enabled())
669 list_add_tail(&llc_occupancy_event.list, &r->evt_list);
670 if (is_mbm_total_enabled())
671 list_add_tail(&mbm_total_event.list, &r->evt_list);
672 if (is_mbm_local_enabled())
673 list_add_tail(&mbm_local_event.list, &r->evt_list);
674}
675
676int rdt_get_mon_l3_config(struct rdt_resource *r)
677{
678 unsigned int mbm_offset = boot_cpu_data.x86_cache_mbm_width_offset;
679 unsigned int cl_size = boot_cpu_data.x86_cache_size;
680 int ret;
681
682 r->mon_scale = boot_cpu_data.x86_cache_occ_scale;
683 r->num_rmid = boot_cpu_data.x86_cache_max_rmid + 1;
684 r->mbm_width = MBM_CNTR_WIDTH_BASE;
685
686 if (mbm_offset > 0 && mbm_offset <= MBM_CNTR_WIDTH_OFFSET_MAX)
687 r->mbm_width += mbm_offset;
688 else if (mbm_offset > MBM_CNTR_WIDTH_OFFSET_MAX)
689 pr_warn("Ignoring impossible MBM counter offset\n");
690
691 /*
692 * A reasonable upper limit on the max threshold is the number
693 * of lines tagged per RMID if all RMIDs have the same number of
694 * lines tagged in the LLC.
695 *
696 * For a 35MB LLC and 56 RMIDs, this is ~1.8% of the LLC.
697 */
698 resctrl_cqm_threshold = cl_size * 1024 / r->num_rmid;
699
700 /* h/w works in units of "boot_cpu_data.x86_cache_occ_scale" */
701 resctrl_cqm_threshold /= r->mon_scale;
702
703 ret = dom_data_init(r);
704 if (ret)
705 return ret;
706
707 l3_mon_evt_init(r);
708
709 r->mon_capable = true;
710 r->mon_enabled = true;
711
712 return 0;
713}
714
715void __init intel_rdt_mbm_apply_quirk(void)
716{
717 int cf_index;
718
719 cf_index = (boot_cpu_data.x86_cache_max_rmid + 1) / 8 - 1;
720 if (cf_index >= ARRAY_SIZE(mbm_cf_table)) {
721 pr_info("No MBM correction factor available\n");
722 return;
723 }
724
725 mbm_cf_rmidthreshold = mbm_cf_table[cf_index].rmidthreshold;
726 mbm_cf = mbm_cf_table[cf_index].cf;
727}