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/sizes.h>
20 #include <linux/slab.h>
21
22 #include <asm/cpu_device_id.h>
23 #include <asm/resctrl.h>
24
25 #include "internal.h"
26
27 struct rmid_entry {
28 u32 rmid;
29 int busy;
30 struct list_head list;
31 };
32
33 /**
34 * @rmid_free_lru A least recently used list of free RMIDs
35 * These RMIDs are guaranteed to have an occupancy less than the
36 * threshold occupancy
37 */
38 static LIST_HEAD(rmid_free_lru);
39
40 /**
41 * @rmid_limbo_count count of currently unused but (potentially)
42 * dirty RMIDs.
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.
46 */
47 static unsigned int rmid_limbo_count;
48
49 /**
50 * @rmid_entry - The entry in the limbo and free lists.
51 */
52 static struct rmid_entry *rmid_ptrs;
53
54 /*
55 * Global boolean for rdt_monitor which is true if any
56 * resource monitoring is enabled.
57 */
58 bool rdt_mon_capable;
59
60 /*
61 * Global to indicate which monitoring events are enabled.
62 */
63 unsigned int rdt_mon_features;
64
65 /*
66 * This is the threshold cache occupancy in bytes at which we will consider an
67 * RMID available for re-allocation.
68 */
69 unsigned int resctrl_rmid_realloc_threshold;
70
71 /*
72 * This is the maximum value for the reallocation threshold, in bytes.
73 */
74 unsigned int resctrl_rmid_realloc_limit;
75
76 #define CF(cf) ((unsigned long)(1048576 * (cf) + 0.5))
77
78 /*
79 * The correction factor table is documented in Documentation/x86/resctrl.rst.
80 * If rmid > rmid threshold, MBM total and local values should be multiplied
81 * by the correction factor.
82 *
83 * The original table is modified for better code:
84 *
85 * 1. The threshold 0 is changed to rmid count - 1 so don't do correction
86 * for the case.
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
92 */
93 static const struct mbm_correction_factor_table {
94 u32 rmidthreshold;
95 u64 cf;
96 } mbm_cf_table[] __initconst = {
97 {7, CF(1.000000)},
98 {15, CF(1.000000)},
99 {15, CF(0.969650)},
100 {31, CF(1.000000)},
101 {31, CF(1.066667)},
102 {31, CF(0.969650)},
103 {47, CF(1.142857)},
104 {63, CF(1.000000)},
105 {63, CF(1.185115)},
106 {63, CF(1.066553)},
107 {79, CF(1.454545)},
108 {95, CF(1.000000)},
109 {95, CF(1.230769)},
110 {95, CF(1.142857)},
111 {95, CF(1.066667)},
112 {127, CF(1.000000)},
113 {127, CF(1.254863)},
114 {127, CF(1.185255)},
115 {151, CF(1.000000)},
116 {127, CF(1.066667)},
117 {167, CF(1.000000)},
118 {159, CF(1.454334)},
119 {183, CF(1.000000)},
120 {127, CF(0.969744)},
121 {191, CF(1.280246)},
122 {191, CF(1.230921)},
123 {215, CF(1.000000)},
124 {191, CF(1.143118)},
125 };
126
127 static u32 mbm_cf_rmidthreshold __read_mostly = UINT_MAX;
128 static u64 mbm_cf __read_mostly;
129
get_corrected_mbm_count(u32 rmid,unsigned long val)130 static inline u64 get_corrected_mbm_count(u32 rmid, unsigned long val)
131 {
132 /* Correct MBM value. */
133 if (rmid > mbm_cf_rmidthreshold)
134 val = (val * mbm_cf) >> 20;
135
136 return val;
137 }
138
__rmid_entry(u32 rmid)139 static inline struct rmid_entry *__rmid_entry(u32 rmid)
140 {
141 struct rmid_entry *entry;
142
143 entry = &rmid_ptrs[rmid];
144 WARN_ON(entry->rmid != rmid);
145
146 return entry;
147 }
148
get_arch_mbm_state(struct rdt_hw_domain * hw_dom,u32 rmid,enum resctrl_event_id eventid)149 static struct arch_mbm_state *get_arch_mbm_state(struct rdt_hw_domain *hw_dom,
150 u32 rmid,
151 enum resctrl_event_id eventid)
152 {
153 switch (eventid) {
154 case QOS_L3_OCCUP_EVENT_ID:
155 return NULL;
156 case QOS_L3_MBM_TOTAL_EVENT_ID:
157 return &hw_dom->arch_mbm_total[rmid];
158 case QOS_L3_MBM_LOCAL_EVENT_ID:
159 return &hw_dom->arch_mbm_local[rmid];
160 }
161
162 /* Never expect to get here */
163 WARN_ON_ONCE(1);
164
165 return NULL;
166 }
167
resctrl_arch_reset_rmid(struct rdt_resource * r,struct rdt_domain * d,u32 rmid,enum resctrl_event_id eventid)168 void resctrl_arch_reset_rmid(struct rdt_resource *r, struct rdt_domain *d,
169 u32 rmid, enum resctrl_event_id eventid)
170 {
171 struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
172 struct arch_mbm_state *am;
173
174 am = get_arch_mbm_state(hw_dom, rmid, eventid);
175 if (am)
176 memset(am, 0, sizeof(*am));
177 }
178
mbm_overflow_count(u64 prev_msr,u64 cur_msr,unsigned int width)179 static u64 mbm_overflow_count(u64 prev_msr, u64 cur_msr, unsigned int width)
180 {
181 u64 shift = 64 - width, chunks;
182
183 chunks = (cur_msr << shift) - (prev_msr << shift);
184 return chunks >> shift;
185 }
186
resctrl_arch_rmid_read(struct rdt_resource * r,struct rdt_domain * d,u32 rmid,enum resctrl_event_id eventid,u64 * val)187 int resctrl_arch_rmid_read(struct rdt_resource *r, struct rdt_domain *d,
188 u32 rmid, enum resctrl_event_id eventid, u64 *val)
189 {
190 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
191 struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
192 struct arch_mbm_state *am;
193 u64 msr_val, chunks;
194
195 if (!cpumask_test_cpu(smp_processor_id(), &d->cpu_mask))
196 return -EINVAL;
197
198 /*
199 * As per the SDM, when IA32_QM_EVTSEL.EvtID (bits 7:0) is configured
200 * with a valid event code for supported resource type and the bits
201 * IA32_QM_EVTSEL.RMID (bits 41:32) are configured with valid RMID,
202 * IA32_QM_CTR.data (bits 61:0) reports the monitored data.
203 * IA32_QM_CTR.Error (bit 63) and IA32_QM_CTR.Unavailable (bit 62)
204 * are error bits.
205 */
206 wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid);
207 rdmsrl(MSR_IA32_QM_CTR, msr_val);
208
209 if (msr_val & RMID_VAL_ERROR)
210 return -EIO;
211 if (msr_val & RMID_VAL_UNAVAIL)
212 return -EINVAL;
213
214 am = get_arch_mbm_state(hw_dom, rmid, eventid);
215 if (am) {
216 am->chunks += mbm_overflow_count(am->prev_msr, msr_val,
217 hw_res->mbm_width);
218 chunks = get_corrected_mbm_count(rmid, am->chunks);
219 am->prev_msr = msr_val;
220 } else {
221 chunks = msr_val;
222 }
223
224 *val = chunks * hw_res->mon_scale;
225
226 return 0;
227 }
228
229 /*
230 * Check the RMIDs that are marked as busy for this domain. If the
231 * reported LLC occupancy is below the threshold clear the busy bit and
232 * decrement the count. If the busy count gets to zero on an RMID, we
233 * free the RMID
234 */
__check_limbo(struct rdt_domain * d,bool force_free)235 void __check_limbo(struct rdt_domain *d, bool force_free)
236 {
237 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
238 struct rmid_entry *entry;
239 u32 crmid = 1, nrmid;
240 bool rmid_dirty;
241 u64 val = 0;
242
243 /*
244 * Skip RMID 0 and start from RMID 1 and check all the RMIDs that
245 * are marked as busy for occupancy < threshold. If the occupancy
246 * is less than the threshold decrement the busy counter of the
247 * RMID and move it to the free list when the counter reaches 0.
248 */
249 for (;;) {
250 nrmid = find_next_bit(d->rmid_busy_llc, r->num_rmid, crmid);
251 if (nrmid >= r->num_rmid)
252 break;
253
254 entry = __rmid_entry(nrmid);
255
256 if (resctrl_arch_rmid_read(r, d, entry->rmid,
257 QOS_L3_OCCUP_EVENT_ID, &val)) {
258 rmid_dirty = true;
259 } else {
260 rmid_dirty = (val >= resctrl_rmid_realloc_threshold);
261 }
262
263 if (force_free || !rmid_dirty) {
264 clear_bit(entry->rmid, d->rmid_busy_llc);
265 if (!--entry->busy) {
266 rmid_limbo_count--;
267 list_add_tail(&entry->list, &rmid_free_lru);
268 }
269 }
270 crmid = nrmid + 1;
271 }
272 }
273
has_busy_rmid(struct rdt_resource * r,struct rdt_domain * d)274 bool has_busy_rmid(struct rdt_resource *r, struct rdt_domain *d)
275 {
276 return find_first_bit(d->rmid_busy_llc, r->num_rmid) != r->num_rmid;
277 }
278
279 /*
280 * As of now the RMIDs allocation is global.
281 * However we keep track of which packages the RMIDs
282 * are used to optimize the limbo list management.
283 */
alloc_rmid(void)284 int alloc_rmid(void)
285 {
286 struct rmid_entry *entry;
287
288 lockdep_assert_held(&rdtgroup_mutex);
289
290 if (list_empty(&rmid_free_lru))
291 return rmid_limbo_count ? -EBUSY : -ENOSPC;
292
293 entry = list_first_entry(&rmid_free_lru,
294 struct rmid_entry, list);
295 list_del(&entry->list);
296
297 return entry->rmid;
298 }
299
add_rmid_to_limbo(struct rmid_entry * entry)300 static void add_rmid_to_limbo(struct rmid_entry *entry)
301 {
302 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
303 struct rdt_domain *d;
304 int cpu, err;
305 u64 val = 0;
306
307 entry->busy = 0;
308 cpu = get_cpu();
309 list_for_each_entry(d, &r->domains, list) {
310 if (cpumask_test_cpu(cpu, &d->cpu_mask)) {
311 err = resctrl_arch_rmid_read(r, d, entry->rmid,
312 QOS_L3_OCCUP_EVENT_ID,
313 &val);
314 if (err || val <= resctrl_rmid_realloc_threshold)
315 continue;
316 }
317
318 /*
319 * For the first limbo RMID in the domain,
320 * setup up the limbo worker.
321 */
322 if (!has_busy_rmid(r, d))
323 cqm_setup_limbo_handler(d, CQM_LIMBOCHECK_INTERVAL);
324 set_bit(entry->rmid, d->rmid_busy_llc);
325 entry->busy++;
326 }
327 put_cpu();
328
329 if (entry->busy)
330 rmid_limbo_count++;
331 else
332 list_add_tail(&entry->list, &rmid_free_lru);
333 }
334
free_rmid(u32 rmid)335 void free_rmid(u32 rmid)
336 {
337 struct rmid_entry *entry;
338
339 if (!rmid)
340 return;
341
342 lockdep_assert_held(&rdtgroup_mutex);
343
344 entry = __rmid_entry(rmid);
345
346 if (is_llc_occupancy_enabled())
347 add_rmid_to_limbo(entry);
348 else
349 list_add_tail(&entry->list, &rmid_free_lru);
350 }
351
__mon_event_count(u32 rmid,struct rmid_read * rr)352 static int __mon_event_count(u32 rmid, struct rmid_read *rr)
353 {
354 struct mbm_state *m;
355 u64 tval = 0;
356
357 if (rr->first)
358 resctrl_arch_reset_rmid(rr->r, rr->d, rmid, rr->evtid);
359
360 rr->err = resctrl_arch_rmid_read(rr->r, rr->d, rmid, rr->evtid, &tval);
361 if (rr->err)
362 return rr->err;
363
364 switch (rr->evtid) {
365 case QOS_L3_OCCUP_EVENT_ID:
366 rr->val += tval;
367 return 0;
368 case QOS_L3_MBM_TOTAL_EVENT_ID:
369 m = &rr->d->mbm_total[rmid];
370 break;
371 case QOS_L3_MBM_LOCAL_EVENT_ID:
372 m = &rr->d->mbm_local[rmid];
373 break;
374 default:
375 /*
376 * Code would never reach here because an invalid
377 * event id would fail in resctrl_arch_rmid_read().
378 */
379 return -EINVAL;
380 }
381
382 if (rr->first) {
383 memset(m, 0, sizeof(struct mbm_state));
384 return 0;
385 }
386
387 rr->val += tval;
388
389 return 0;
390 }
391
392 /*
393 * mbm_bw_count() - Update bw count from values previously read by
394 * __mon_event_count().
395 * @rmid: The rmid used to identify the cached mbm_state.
396 * @rr: The struct rmid_read populated by __mon_event_count().
397 *
398 * Supporting function to calculate the memory bandwidth
399 * and delta bandwidth in MBps. The chunks value previously read by
400 * __mon_event_count() is compared with the chunks value from the previous
401 * invocation. This must be called once per second to maintain values in MBps.
402 */
mbm_bw_count(u32 rmid,struct rmid_read * rr)403 static void mbm_bw_count(u32 rmid, struct rmid_read *rr)
404 {
405 struct mbm_state *m = &rr->d->mbm_local[rmid];
406 u64 cur_bw, bytes, cur_bytes;
407
408 cur_bytes = rr->val;
409 bytes = cur_bytes - m->prev_bw_bytes;
410 m->prev_bw_bytes = cur_bytes;
411
412 cur_bw = bytes / SZ_1M;
413
414 if (m->delta_comp)
415 m->delta_bw = abs(cur_bw - m->prev_bw);
416 m->delta_comp = false;
417 m->prev_bw = cur_bw;
418 }
419
420 /*
421 * This is called via IPI to read the CQM/MBM counters
422 * on a domain.
423 */
mon_event_count(void * info)424 void mon_event_count(void *info)
425 {
426 struct rdtgroup *rdtgrp, *entry;
427 struct rmid_read *rr = info;
428 struct list_head *head;
429 int ret;
430
431 rdtgrp = rr->rgrp;
432
433 ret = __mon_event_count(rdtgrp->mon.rmid, rr);
434
435 /*
436 * For Ctrl groups read data from child monitor groups and
437 * add them together. Count events which are read successfully.
438 * Discard the rmid_read's reporting errors.
439 */
440 head = &rdtgrp->mon.crdtgrp_list;
441
442 if (rdtgrp->type == RDTCTRL_GROUP) {
443 list_for_each_entry(entry, head, mon.crdtgrp_list) {
444 if (__mon_event_count(entry->mon.rmid, rr) == 0)
445 ret = 0;
446 }
447 }
448
449 /*
450 * __mon_event_count() calls for newly created monitor groups may
451 * report -EINVAL/Unavailable if the monitor hasn't seen any traffic.
452 * Discard error if any of the monitor event reads succeeded.
453 */
454 if (ret == 0)
455 rr->err = 0;
456 }
457
458 /*
459 * Feedback loop for MBA software controller (mba_sc)
460 *
461 * mba_sc is a feedback loop where we periodically read MBM counters and
462 * adjust the bandwidth percentage values via the IA32_MBA_THRTL_MSRs so
463 * that:
464 *
465 * current bandwidth(cur_bw) < user specified bandwidth(user_bw)
466 *
467 * This uses the MBM counters to measure the bandwidth and MBA throttle
468 * MSRs to control the bandwidth for a particular rdtgrp. It builds on the
469 * fact that resctrl rdtgroups have both monitoring and control.
470 *
471 * The frequency of the checks is 1s and we just tag along the MBM overflow
472 * timer. Having 1s interval makes the calculation of bandwidth simpler.
473 *
474 * Although MBA's goal is to restrict the bandwidth to a maximum, there may
475 * be a need to increase the bandwidth to avoid unnecessarily restricting
476 * the L2 <-> L3 traffic.
477 *
478 * Since MBA controls the L2 external bandwidth where as MBM measures the
479 * L3 external bandwidth the following sequence could lead to such a
480 * situation.
481 *
482 * Consider an rdtgroup which had high L3 <-> memory traffic in initial
483 * phases -> mba_sc kicks in and reduced bandwidth percentage values -> but
484 * after some time rdtgroup has mostly L2 <-> L3 traffic.
485 *
486 * In this case we may restrict the rdtgroup's L2 <-> L3 traffic as its
487 * throttle MSRs already have low percentage values. To avoid
488 * unnecessarily restricting such rdtgroups, we also increase the bandwidth.
489 */
update_mba_bw(struct rdtgroup * rgrp,struct rdt_domain * dom_mbm)490 static void update_mba_bw(struct rdtgroup *rgrp, struct rdt_domain *dom_mbm)
491 {
492 u32 closid, rmid, cur_msr_val, new_msr_val;
493 struct mbm_state *pmbm_data, *cmbm_data;
494 u32 cur_bw, delta_bw, user_bw;
495 struct rdt_resource *r_mba;
496 struct rdt_domain *dom_mba;
497 struct list_head *head;
498 struct rdtgroup *entry;
499
500 if (!is_mbm_local_enabled())
501 return;
502
503 r_mba = &rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl;
504
505 closid = rgrp->closid;
506 rmid = rgrp->mon.rmid;
507 pmbm_data = &dom_mbm->mbm_local[rmid];
508
509 dom_mba = get_domain_from_cpu(smp_processor_id(), r_mba);
510 if (!dom_mba) {
511 pr_warn_once("Failure to get domain for MBA update\n");
512 return;
513 }
514
515 cur_bw = pmbm_data->prev_bw;
516 user_bw = dom_mba->mbps_val[closid];
517 delta_bw = pmbm_data->delta_bw;
518
519 /* MBA resource doesn't support CDP */
520 cur_msr_val = resctrl_arch_get_config(r_mba, dom_mba, closid, CDP_NONE);
521
522 /*
523 * For Ctrl groups read data from child monitor groups.
524 */
525 head = &rgrp->mon.crdtgrp_list;
526 list_for_each_entry(entry, head, mon.crdtgrp_list) {
527 cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
528 cur_bw += cmbm_data->prev_bw;
529 delta_bw += cmbm_data->delta_bw;
530 }
531
532 /*
533 * Scale up/down the bandwidth linearly for the ctrl group. The
534 * bandwidth step is the bandwidth granularity specified by the
535 * hardware.
536 *
537 * The delta_bw is used when increasing the bandwidth so that we
538 * dont alternately increase and decrease the control values
539 * continuously.
540 *
541 * For ex: consider cur_bw = 90MBps, user_bw = 100MBps and if
542 * bandwidth step is 20MBps(> user_bw - cur_bw), we would keep
543 * switching between 90 and 110 continuously if we only check
544 * cur_bw < user_bw.
545 */
546 if (cur_msr_val > r_mba->membw.min_bw && user_bw < cur_bw) {
547 new_msr_val = cur_msr_val - r_mba->membw.bw_gran;
548 } else if (cur_msr_val < MAX_MBA_BW &&
549 (user_bw > (cur_bw + delta_bw))) {
550 new_msr_val = cur_msr_val + r_mba->membw.bw_gran;
551 } else {
552 return;
553 }
554
555 resctrl_arch_update_one(r_mba, dom_mba, closid, CDP_NONE, new_msr_val);
556
557 /*
558 * Delta values are updated dynamically package wise for each
559 * rdtgrp every time the throttle MSR changes value.
560 *
561 * This is because (1)the increase in bandwidth is not perfectly
562 * linear and only "approximately" linear even when the hardware
563 * says it is linear.(2)Also since MBA is a core specific
564 * mechanism, the delta values vary based on number of cores used
565 * by the rdtgrp.
566 */
567 pmbm_data->delta_comp = true;
568 list_for_each_entry(entry, head, mon.crdtgrp_list) {
569 cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
570 cmbm_data->delta_comp = true;
571 }
572 }
573
mbm_update(struct rdt_resource * r,struct rdt_domain * d,int rmid)574 static void mbm_update(struct rdt_resource *r, struct rdt_domain *d, int rmid)
575 {
576 struct rmid_read rr;
577
578 rr.first = false;
579 rr.r = r;
580 rr.d = d;
581
582 /*
583 * This is protected from concurrent reads from user
584 * as both the user and we hold the global mutex.
585 */
586 if (is_mbm_total_enabled()) {
587 rr.evtid = QOS_L3_MBM_TOTAL_EVENT_ID;
588 rr.val = 0;
589 __mon_event_count(rmid, &rr);
590 }
591 if (is_mbm_local_enabled()) {
592 rr.evtid = QOS_L3_MBM_LOCAL_EVENT_ID;
593 rr.val = 0;
594 __mon_event_count(rmid, &rr);
595
596 /*
597 * Call the MBA software controller only for the
598 * control groups and when user has enabled
599 * the software controller explicitly.
600 */
601 if (is_mba_sc(NULL))
602 mbm_bw_count(rmid, &rr);
603 }
604 }
605
606 /*
607 * Handler to scan the limbo list and move the RMIDs
608 * to free list whose occupancy < threshold_occupancy.
609 */
cqm_handle_limbo(struct work_struct * work)610 void cqm_handle_limbo(struct work_struct *work)
611 {
612 unsigned long delay = msecs_to_jiffies(CQM_LIMBOCHECK_INTERVAL);
613 int cpu = smp_processor_id();
614 struct rdt_resource *r;
615 struct rdt_domain *d;
616
617 mutex_lock(&rdtgroup_mutex);
618
619 r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
620 d = container_of(work, struct rdt_domain, cqm_limbo.work);
621
622 __check_limbo(d, false);
623
624 if (has_busy_rmid(r, d))
625 schedule_delayed_work_on(cpu, &d->cqm_limbo, delay);
626
627 mutex_unlock(&rdtgroup_mutex);
628 }
629
cqm_setup_limbo_handler(struct rdt_domain * dom,unsigned long delay_ms)630 void cqm_setup_limbo_handler(struct rdt_domain *dom, unsigned long delay_ms)
631 {
632 unsigned long delay = msecs_to_jiffies(delay_ms);
633 int cpu;
634
635 cpu = cpumask_any(&dom->cpu_mask);
636 dom->cqm_work_cpu = cpu;
637
638 schedule_delayed_work_on(cpu, &dom->cqm_limbo, delay);
639 }
640
mbm_handle_overflow(struct work_struct * work)641 void mbm_handle_overflow(struct work_struct *work)
642 {
643 unsigned long delay = msecs_to_jiffies(MBM_OVERFLOW_INTERVAL);
644 struct rdtgroup *prgrp, *crgrp;
645 int cpu = smp_processor_id();
646 struct list_head *head;
647 struct rdt_resource *r;
648 struct rdt_domain *d;
649
650 mutex_lock(&rdtgroup_mutex);
651
652 if (!static_branch_likely(&rdt_mon_enable_key))
653 goto out_unlock;
654
655 r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
656 d = container_of(work, struct rdt_domain, mbm_over.work);
657
658 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
659 mbm_update(r, d, prgrp->mon.rmid);
660
661 head = &prgrp->mon.crdtgrp_list;
662 list_for_each_entry(crgrp, head, mon.crdtgrp_list)
663 mbm_update(r, d, crgrp->mon.rmid);
664
665 if (is_mba_sc(NULL))
666 update_mba_bw(prgrp, d);
667 }
668
669 schedule_delayed_work_on(cpu, &d->mbm_over, delay);
670
671 out_unlock:
672 mutex_unlock(&rdtgroup_mutex);
673 }
674
mbm_setup_overflow_handler(struct rdt_domain * dom,unsigned long delay_ms)675 void mbm_setup_overflow_handler(struct rdt_domain *dom, unsigned long delay_ms)
676 {
677 unsigned long delay = msecs_to_jiffies(delay_ms);
678 int cpu;
679
680 if (!static_branch_likely(&rdt_mon_enable_key))
681 return;
682 cpu = cpumask_any(&dom->cpu_mask);
683 dom->mbm_work_cpu = cpu;
684 schedule_delayed_work_on(cpu, &dom->mbm_over, delay);
685 }
686
dom_data_init(struct rdt_resource * r)687 static int dom_data_init(struct rdt_resource *r)
688 {
689 struct rmid_entry *entry = NULL;
690 int i, nr_rmids;
691
692 nr_rmids = r->num_rmid;
693 rmid_ptrs = kcalloc(nr_rmids, sizeof(struct rmid_entry), GFP_KERNEL);
694 if (!rmid_ptrs)
695 return -ENOMEM;
696
697 for (i = 0; i < nr_rmids; i++) {
698 entry = &rmid_ptrs[i];
699 INIT_LIST_HEAD(&entry->list);
700
701 entry->rmid = i;
702 list_add_tail(&entry->list, &rmid_free_lru);
703 }
704
705 /*
706 * RMID 0 is special and is always allocated. It's used for all
707 * tasks that are not monitored.
708 */
709 entry = __rmid_entry(0);
710 list_del(&entry->list);
711
712 return 0;
713 }
714
715 static struct mon_evt llc_occupancy_event = {
716 .name = "llc_occupancy",
717 .evtid = QOS_L3_OCCUP_EVENT_ID,
718 };
719
720 static struct mon_evt mbm_total_event = {
721 .name = "mbm_total_bytes",
722 .evtid = QOS_L3_MBM_TOTAL_EVENT_ID,
723 };
724
725 static struct mon_evt mbm_local_event = {
726 .name = "mbm_local_bytes",
727 .evtid = QOS_L3_MBM_LOCAL_EVENT_ID,
728 };
729
730 /*
731 * Initialize the event list for the resource.
732 *
733 * Note that MBM events are also part of RDT_RESOURCE_L3 resource
734 * because as per the SDM the total and local memory bandwidth
735 * are enumerated as part of L3 monitoring.
736 */
l3_mon_evt_init(struct rdt_resource * r)737 static void l3_mon_evt_init(struct rdt_resource *r)
738 {
739 INIT_LIST_HEAD(&r->evt_list);
740
741 if (is_llc_occupancy_enabled())
742 list_add_tail(&llc_occupancy_event.list, &r->evt_list);
743 if (is_mbm_total_enabled())
744 list_add_tail(&mbm_total_event.list, &r->evt_list);
745 if (is_mbm_local_enabled())
746 list_add_tail(&mbm_local_event.list, &r->evt_list);
747 }
748
rdt_get_mon_l3_config(struct rdt_resource * r)749 int rdt_get_mon_l3_config(struct rdt_resource *r)
750 {
751 unsigned int mbm_offset = boot_cpu_data.x86_cache_mbm_width_offset;
752 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
753 unsigned int threshold;
754 int ret;
755
756 resctrl_rmid_realloc_limit = boot_cpu_data.x86_cache_size * 1024;
757 hw_res->mon_scale = boot_cpu_data.x86_cache_occ_scale;
758 r->num_rmid = boot_cpu_data.x86_cache_max_rmid + 1;
759 hw_res->mbm_width = MBM_CNTR_WIDTH_BASE;
760
761 if (mbm_offset > 0 && mbm_offset <= MBM_CNTR_WIDTH_OFFSET_MAX)
762 hw_res->mbm_width += mbm_offset;
763 else if (mbm_offset > MBM_CNTR_WIDTH_OFFSET_MAX)
764 pr_warn("Ignoring impossible MBM counter offset\n");
765
766 /*
767 * A reasonable upper limit on the max threshold is the number
768 * of lines tagged per RMID if all RMIDs have the same number of
769 * lines tagged in the LLC.
770 *
771 * For a 35MB LLC and 56 RMIDs, this is ~1.8% of the LLC.
772 */
773 threshold = resctrl_rmid_realloc_limit / r->num_rmid;
774
775 /*
776 * Because num_rmid may not be a power of two, round the value
777 * to the nearest multiple of hw_res->mon_scale so it matches a
778 * value the hardware will measure. mon_scale may not be a power of 2.
779 */
780 resctrl_rmid_realloc_threshold = resctrl_arch_round_mon_val(threshold);
781
782 ret = dom_data_init(r);
783 if (ret)
784 return ret;
785
786 l3_mon_evt_init(r);
787
788 r->mon_capable = true;
789
790 return 0;
791 }
792
intel_rdt_mbm_apply_quirk(void)793 void __init intel_rdt_mbm_apply_quirk(void)
794 {
795 int cf_index;
796
797 cf_index = (boot_cpu_data.x86_cache_max_rmid + 1) / 8 - 1;
798 if (cf_index >= ARRAY_SIZE(mbm_cf_table)) {
799 pr_info("No MBM correction factor available\n");
800 return;
801 }
802
803 mbm_cf_rmidthreshold = mbm_cf_table[cf_index].rmidthreshold;
804 mbm_cf = mbm_cf_table[cf_index].cf;
805 }
806
