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
__rmid_read(u32 rmid,enum resctrl_event_id eventid,u64 * val)149 static int __rmid_read(u32 rmid, enum resctrl_event_id eventid, u64 *val)
150 {
151 u64 msr_val;
152
153 /*
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)
159 * are error bits.
160 */
161 wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid);
162 rdmsrl(MSR_IA32_QM_CTR, msr_val);
163
164 if (msr_val & RMID_VAL_ERROR)
165 return -EIO;
166 if (msr_val & RMID_VAL_UNAVAIL)
167 return -EINVAL;
168
169 *val = msr_val;
170 return 0;
171 }
172
get_arch_mbm_state(struct rdt_hw_domain * hw_dom,u32 rmid,enum resctrl_event_id eventid)173 static struct arch_mbm_state *get_arch_mbm_state(struct rdt_hw_domain *hw_dom,
174 u32 rmid,
175 enum resctrl_event_id eventid)
176 {
177 switch (eventid) {
178 case QOS_L3_OCCUP_EVENT_ID:
179 return NULL;
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];
184 }
185
186 /* Never expect to get here */
187 WARN_ON_ONCE(1);
188
189 return NULL;
190 }
191
resctrl_arch_reset_rmid(struct rdt_resource * r,struct rdt_domain * d,u32 rmid,enum resctrl_event_id eventid)192 void resctrl_arch_reset_rmid(struct rdt_resource *r, struct rdt_domain *d,
193 u32 rmid, enum resctrl_event_id eventid)
194 {
195 struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
196 struct arch_mbm_state *am;
197
198 am = get_arch_mbm_state(hw_dom, rmid, eventid);
199 if (am) {
200 memset(am, 0, sizeof(*am));
201
202 /* Record any initial, non-zero count value. */
203 __rmid_read(rmid, eventid, &am->prev_msr);
204 }
205 }
206
207 /*
208 * Assumes that hardware counters are also reset and thus that there is
209 * no need to record initial non-zero counts.
210 */
resctrl_arch_reset_rmid_all(struct rdt_resource * r,struct rdt_domain * d)211 void resctrl_arch_reset_rmid_all(struct rdt_resource *r, struct rdt_domain *d)
212 {
213 struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
214
215 if (is_mbm_total_enabled())
216 memset(hw_dom->arch_mbm_total, 0,
217 sizeof(*hw_dom->arch_mbm_total) * r->num_rmid);
218
219 if (is_mbm_local_enabled())
220 memset(hw_dom->arch_mbm_local, 0,
221 sizeof(*hw_dom->arch_mbm_local) * r->num_rmid);
222 }
223
mbm_overflow_count(u64 prev_msr,u64 cur_msr,unsigned int width)224 static u64 mbm_overflow_count(u64 prev_msr, u64 cur_msr, unsigned int width)
225 {
226 u64 shift = 64 - width, chunks;
227
228 chunks = (cur_msr << shift) - (prev_msr << shift);
229 return chunks >> shift;
230 }
231
resctrl_arch_rmid_read(struct rdt_resource * r,struct rdt_domain * d,u32 rmid,enum resctrl_event_id eventid,u64 * val)232 int resctrl_arch_rmid_read(struct rdt_resource *r, struct rdt_domain *d,
233 u32 rmid, enum resctrl_event_id eventid, u64 *val)
234 {
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;
238 u64 msr_val, chunks;
239 int ret;
240
241 if (!cpumask_test_cpu(smp_processor_id(), &d->cpu_mask))
242 return -EINVAL;
243
244 ret = __rmid_read(rmid, eventid, &msr_val);
245 if (ret)
246 return ret;
247
248 am = get_arch_mbm_state(hw_dom, rmid, eventid);
249 if (am) {
250 am->chunks += mbm_overflow_count(am->prev_msr, msr_val,
251 hw_res->mbm_width);
252 chunks = get_corrected_mbm_count(rmid, am->chunks);
253 am->prev_msr = msr_val;
254 } else {
255 chunks = msr_val;
256 }
257
258 *val = chunks * hw_res->mon_scale;
259
260 return 0;
261 }
262
263 /*
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
267 * free the RMID
268 */
__check_limbo(struct rdt_domain * d,bool force_free)269 void __check_limbo(struct rdt_domain *d, bool force_free)
270 {
271 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
272 struct rmid_entry *entry;
273 u32 crmid = 1, nrmid;
274 bool rmid_dirty;
275 u64 val = 0;
276
277 /*
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.
282 */
283 for (;;) {
284 nrmid = find_next_bit(d->rmid_busy_llc, r->num_rmid, crmid);
285 if (nrmid >= r->num_rmid)
286 break;
287
288 entry = __rmid_entry(nrmid);
289
290 if (resctrl_arch_rmid_read(r, d, entry->rmid,
291 QOS_L3_OCCUP_EVENT_ID, &val)) {
292 rmid_dirty = true;
293 } else {
294 rmid_dirty = (val >= resctrl_rmid_realloc_threshold);
295 }
296
297 if (force_free || !rmid_dirty) {
298 clear_bit(entry->rmid, d->rmid_busy_llc);
299 if (!--entry->busy) {
300 rmid_limbo_count--;
301 list_add_tail(&entry->list, &rmid_free_lru);
302 }
303 }
304 crmid = nrmid + 1;
305 }
306 }
307
has_busy_rmid(struct rdt_resource * r,struct rdt_domain * d)308 bool has_busy_rmid(struct rdt_resource *r, struct rdt_domain *d)
309 {
310 return find_first_bit(d->rmid_busy_llc, r->num_rmid) != r->num_rmid;
311 }
312
313 /*
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.
317 */
alloc_rmid(void)318 int alloc_rmid(void)
319 {
320 struct rmid_entry *entry;
321
322 lockdep_assert_held(&rdtgroup_mutex);
323
324 if (list_empty(&rmid_free_lru))
325 return rmid_limbo_count ? -EBUSY : -ENOSPC;
326
327 entry = list_first_entry(&rmid_free_lru,
328 struct rmid_entry, list);
329 list_del(&entry->list);
330
331 return entry->rmid;
332 }
333
add_rmid_to_limbo(struct rmid_entry * entry)334 static void add_rmid_to_limbo(struct rmid_entry *entry)
335 {
336 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
337 struct rdt_domain *d;
338 int cpu, err;
339 u64 val = 0;
340
341 entry->busy = 0;
342 cpu = get_cpu();
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,
347 &val);
348 if (err || val <= resctrl_rmid_realloc_threshold)
349 continue;
350 }
351
352 /*
353 * For the first limbo RMID in the domain,
354 * setup up the limbo worker.
355 */
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);
359 entry->busy++;
360 }
361 put_cpu();
362
363 if (entry->busy)
364 rmid_limbo_count++;
365 else
366 list_add_tail(&entry->list, &rmid_free_lru);
367 }
368
free_rmid(u32 rmid)369 void free_rmid(u32 rmid)
370 {
371 struct rmid_entry *entry;
372
373 if (!rmid)
374 return;
375
376 lockdep_assert_held(&rdtgroup_mutex);
377
378 entry = __rmid_entry(rmid);
379
380 if (is_llc_occupancy_enabled())
381 add_rmid_to_limbo(entry);
382 else
383 list_add_tail(&entry->list, &rmid_free_lru);
384 }
385
__mon_event_count(u32 rmid,struct rmid_read * rr)386 static int __mon_event_count(u32 rmid, struct rmid_read *rr)
387 {
388 struct mbm_state *m;
389 u64 tval = 0;
390
391 if (rr->first)
392 resctrl_arch_reset_rmid(rr->r, rr->d, rmid, rr->evtid);
393
394 rr->err = resctrl_arch_rmid_read(rr->r, rr->d, rmid, rr->evtid, &tval);
395 if (rr->err)
396 return rr->err;
397
398 switch (rr->evtid) {
399 case QOS_L3_OCCUP_EVENT_ID:
400 rr->val += tval;
401 return 0;
402 case QOS_L3_MBM_TOTAL_EVENT_ID:
403 m = &rr->d->mbm_total[rmid];
404 break;
405 case QOS_L3_MBM_LOCAL_EVENT_ID:
406 m = &rr->d->mbm_local[rmid];
407 break;
408 default:
409 /*
410 * Code would never reach here because an invalid
411 * event id would fail in resctrl_arch_rmid_read().
412 */
413 return -EINVAL;
414 }
415
416 if (rr->first) {
417 memset(m, 0, sizeof(struct mbm_state));
418 return 0;
419 }
420
421 rr->val += tval;
422
423 return 0;
424 }
425
426 /*
427 * mbm_bw_count() - Update bw count from values previously read by
428 * __mon_event_count().
429 * @rmid: The rmid used to identify the cached mbm_state.
430 * @rr: The struct rmid_read populated by __mon_event_count().
431 *
432 * Supporting function to calculate the memory bandwidth
433 * and delta bandwidth in MBps. The chunks value previously read by
434 * __mon_event_count() is compared with the chunks value from the previous
435 * invocation. This must be called once per second to maintain values in MBps.
436 */
mbm_bw_count(u32 rmid,struct rmid_read * rr)437 static void mbm_bw_count(u32 rmid, struct rmid_read *rr)
438 {
439 struct mbm_state *m = &rr->d->mbm_local[rmid];
440 u64 cur_bw, bytes, cur_bytes;
441
442 cur_bytes = rr->val;
443 bytes = cur_bytes - m->prev_bw_bytes;
444 m->prev_bw_bytes = cur_bytes;
445
446 cur_bw = bytes / SZ_1M;
447
448 if (m->delta_comp)
449 m->delta_bw = abs(cur_bw - m->prev_bw);
450 m->delta_comp = false;
451 m->prev_bw = cur_bw;
452 }
453
454 /*
455 * This is called via IPI to read the CQM/MBM counters
456 * on a domain.
457 */
mon_event_count(void * info)458 void mon_event_count(void *info)
459 {
460 struct rdtgroup *rdtgrp, *entry;
461 struct rmid_read *rr = info;
462 struct list_head *head;
463 int ret;
464
465 rdtgrp = rr->rgrp;
466
467 ret = __mon_event_count(rdtgrp->mon.rmid, rr);
468
469 /*
470 * For Ctrl groups read data from child monitor groups and
471 * add them together. Count events which are read successfully.
472 * Discard the rmid_read's reporting errors.
473 */
474 head = &rdtgrp->mon.crdtgrp_list;
475
476 if (rdtgrp->type == RDTCTRL_GROUP) {
477 list_for_each_entry(entry, head, mon.crdtgrp_list) {
478 if (__mon_event_count(entry->mon.rmid, rr) == 0)
479 ret = 0;
480 }
481 }
482
483 /*
484 * __mon_event_count() calls for newly created monitor groups may
485 * report -EINVAL/Unavailable if the monitor hasn't seen any traffic.
486 * Discard error if any of the monitor event reads succeeded.
487 */
488 if (ret == 0)
489 rr->err = 0;
490 }
491
492 /*
493 * Feedback loop for MBA software controller (mba_sc)
494 *
495 * mba_sc is a feedback loop where we periodically read MBM counters and
496 * adjust the bandwidth percentage values via the IA32_MBA_THRTL_MSRs so
497 * that:
498 *
499 * current bandwidth(cur_bw) < user specified bandwidth(user_bw)
500 *
501 * This uses the MBM counters to measure the bandwidth and MBA throttle
502 * MSRs to control the bandwidth for a particular rdtgrp. It builds on the
503 * fact that resctrl rdtgroups have both monitoring and control.
504 *
505 * The frequency of the checks is 1s and we just tag along the MBM overflow
506 * timer. Having 1s interval makes the calculation of bandwidth simpler.
507 *
508 * Although MBA's goal is to restrict the bandwidth to a maximum, there may
509 * be a need to increase the bandwidth to avoid unnecessarily restricting
510 * the L2 <-> L3 traffic.
511 *
512 * Since MBA controls the L2 external bandwidth where as MBM measures the
513 * L3 external bandwidth the following sequence could lead to such a
514 * situation.
515 *
516 * Consider an rdtgroup which had high L3 <-> memory traffic in initial
517 * phases -> mba_sc kicks in and reduced bandwidth percentage values -> but
518 * after some time rdtgroup has mostly L2 <-> L3 traffic.
519 *
520 * In this case we may restrict the rdtgroup's L2 <-> L3 traffic as its
521 * throttle MSRs already have low percentage values. To avoid
522 * unnecessarily restricting such rdtgroups, we also increase the bandwidth.
523 */
update_mba_bw(struct rdtgroup * rgrp,struct rdt_domain * dom_mbm)524 static void update_mba_bw(struct rdtgroup *rgrp, struct rdt_domain *dom_mbm)
525 {
526 u32 closid, rmid, cur_msr_val, new_msr_val;
527 struct mbm_state *pmbm_data, *cmbm_data;
528 u32 cur_bw, delta_bw, user_bw;
529 struct rdt_resource *r_mba;
530 struct rdt_domain *dom_mba;
531 struct list_head *head;
532 struct rdtgroup *entry;
533
534 if (!is_mbm_local_enabled())
535 return;
536
537 r_mba = &rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl;
538
539 closid = rgrp->closid;
540 rmid = rgrp->mon.rmid;
541 pmbm_data = &dom_mbm->mbm_local[rmid];
542
543 dom_mba = get_domain_from_cpu(smp_processor_id(), r_mba);
544 if (!dom_mba) {
545 pr_warn_once("Failure to get domain for MBA update\n");
546 return;
547 }
548
549 cur_bw = pmbm_data->prev_bw;
550 user_bw = dom_mba->mbps_val[closid];
551 delta_bw = pmbm_data->delta_bw;
552
553 /* MBA resource doesn't support CDP */
554 cur_msr_val = resctrl_arch_get_config(r_mba, dom_mba, closid, CDP_NONE);
555
556 /*
557 * For Ctrl groups read data from child monitor groups.
558 */
559 head = &rgrp->mon.crdtgrp_list;
560 list_for_each_entry(entry, head, mon.crdtgrp_list) {
561 cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
562 cur_bw += cmbm_data->prev_bw;
563 delta_bw += cmbm_data->delta_bw;
564 }
565
566 /*
567 * Scale up/down the bandwidth linearly for the ctrl group. The
568 * bandwidth step is the bandwidth granularity specified by the
569 * hardware.
570 *
571 * The delta_bw is used when increasing the bandwidth so that we
572 * dont alternately increase and decrease the control values
573 * continuously.
574 *
575 * For ex: consider cur_bw = 90MBps, user_bw = 100MBps and if
576 * bandwidth step is 20MBps(> user_bw - cur_bw), we would keep
577 * switching between 90 and 110 continuously if we only check
578 * cur_bw < user_bw.
579 */
580 if (cur_msr_val > r_mba->membw.min_bw && user_bw < cur_bw) {
581 new_msr_val = cur_msr_val - r_mba->membw.bw_gran;
582 } else if (cur_msr_val < MAX_MBA_BW &&
583 (user_bw > (cur_bw + delta_bw))) {
584 new_msr_val = cur_msr_val + r_mba->membw.bw_gran;
585 } else {
586 return;
587 }
588
589 resctrl_arch_update_one(r_mba, dom_mba, closid, CDP_NONE, new_msr_val);
590
591 /*
592 * Delta values are updated dynamically package wise for each
593 * rdtgrp every time the throttle MSR changes value.
594 *
595 * This is because (1)the increase in bandwidth is not perfectly
596 * linear and only "approximately" linear even when the hardware
597 * says it is linear.(2)Also since MBA is a core specific
598 * mechanism, the delta values vary based on number of cores used
599 * by the rdtgrp.
600 */
601 pmbm_data->delta_comp = true;
602 list_for_each_entry(entry, head, mon.crdtgrp_list) {
603 cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
604 cmbm_data->delta_comp = true;
605 }
606 }
607
mbm_update(struct rdt_resource * r,struct rdt_domain * d,int rmid)608 static void mbm_update(struct rdt_resource *r, struct rdt_domain *d, int rmid)
609 {
610 struct rmid_read rr;
611
612 rr.first = false;
613 rr.r = r;
614 rr.d = d;
615
616 /*
617 * This is protected from concurrent reads from user
618 * as both the user and we hold the global mutex.
619 */
620 if (is_mbm_total_enabled()) {
621 rr.evtid = QOS_L3_MBM_TOTAL_EVENT_ID;
622 rr.val = 0;
623 __mon_event_count(rmid, &rr);
624 }
625 if (is_mbm_local_enabled()) {
626 rr.evtid = QOS_L3_MBM_LOCAL_EVENT_ID;
627 rr.val = 0;
628 __mon_event_count(rmid, &rr);
629
630 /*
631 * Call the MBA software controller only for the
632 * control groups and when user has enabled
633 * the software controller explicitly.
634 */
635 if (is_mba_sc(NULL))
636 mbm_bw_count(rmid, &rr);
637 }
638 }
639
640 /*
641 * Handler to scan the limbo list and move the RMIDs
642 * to free list whose occupancy < threshold_occupancy.
643 */
cqm_handle_limbo(struct work_struct * work)644 void cqm_handle_limbo(struct work_struct *work)
645 {
646 unsigned long delay = msecs_to_jiffies(CQM_LIMBOCHECK_INTERVAL);
647 int cpu = smp_processor_id();
648 struct rdt_resource *r;
649 struct rdt_domain *d;
650
651 mutex_lock(&rdtgroup_mutex);
652
653 r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
654 d = container_of(work, struct rdt_domain, cqm_limbo.work);
655
656 __check_limbo(d, false);
657
658 if (has_busy_rmid(r, d))
659 schedule_delayed_work_on(cpu, &d->cqm_limbo, delay);
660
661 mutex_unlock(&rdtgroup_mutex);
662 }
663
cqm_setup_limbo_handler(struct rdt_domain * dom,unsigned long delay_ms)664 void cqm_setup_limbo_handler(struct rdt_domain *dom, unsigned long delay_ms)
665 {
666 unsigned long delay = msecs_to_jiffies(delay_ms);
667 int cpu;
668
669 cpu = cpumask_any(&dom->cpu_mask);
670 dom->cqm_work_cpu = cpu;
671
672 schedule_delayed_work_on(cpu, &dom->cqm_limbo, delay);
673 }
674
mbm_handle_overflow(struct work_struct * work)675 void mbm_handle_overflow(struct work_struct *work)
676 {
677 unsigned long delay = msecs_to_jiffies(MBM_OVERFLOW_INTERVAL);
678 struct rdtgroup *prgrp, *crgrp;
679 int cpu = smp_processor_id();
680 struct list_head *head;
681 struct rdt_resource *r;
682 struct rdt_domain *d;
683
684 mutex_lock(&rdtgroup_mutex);
685
686 if (!static_branch_likely(&rdt_mon_enable_key))
687 goto out_unlock;
688
689 r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
690 d = container_of(work, struct rdt_domain, mbm_over.work);
691
692 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
693 mbm_update(r, d, prgrp->mon.rmid);
694
695 head = &prgrp->mon.crdtgrp_list;
696 list_for_each_entry(crgrp, head, mon.crdtgrp_list)
697 mbm_update(r, d, crgrp->mon.rmid);
698
699 if (is_mba_sc(NULL))
700 update_mba_bw(prgrp, d);
701 }
702
703 schedule_delayed_work_on(cpu, &d->mbm_over, delay);
704
705 out_unlock:
706 mutex_unlock(&rdtgroup_mutex);
707 }
708
mbm_setup_overflow_handler(struct rdt_domain * dom,unsigned long delay_ms)709 void mbm_setup_overflow_handler(struct rdt_domain *dom, unsigned long delay_ms)
710 {
711 unsigned long delay = msecs_to_jiffies(delay_ms);
712 int cpu;
713
714 if (!static_branch_likely(&rdt_mon_enable_key))
715 return;
716 cpu = cpumask_any(&dom->cpu_mask);
717 dom->mbm_work_cpu = cpu;
718 schedule_delayed_work_on(cpu, &dom->mbm_over, delay);
719 }
720
dom_data_init(struct rdt_resource * r)721 static int dom_data_init(struct rdt_resource *r)
722 {
723 struct rmid_entry *entry = NULL;
724 int i, nr_rmids;
725
726 nr_rmids = r->num_rmid;
727 rmid_ptrs = kcalloc(nr_rmids, sizeof(struct rmid_entry), GFP_KERNEL);
728 if (!rmid_ptrs)
729 return -ENOMEM;
730
731 for (i = 0; i < nr_rmids; i++) {
732 entry = &rmid_ptrs[i];
733 INIT_LIST_HEAD(&entry->list);
734
735 entry->rmid = i;
736 list_add_tail(&entry->list, &rmid_free_lru);
737 }
738
739 /*
740 * RMID 0 is special and is always allocated. It's used for all
741 * tasks that are not monitored.
742 */
743 entry = __rmid_entry(0);
744 list_del(&entry->list);
745
746 return 0;
747 }
748
749 static struct mon_evt llc_occupancy_event = {
750 .name = "llc_occupancy",
751 .evtid = QOS_L3_OCCUP_EVENT_ID,
752 };
753
754 static struct mon_evt mbm_total_event = {
755 .name = "mbm_total_bytes",
756 .evtid = QOS_L3_MBM_TOTAL_EVENT_ID,
757 };
758
759 static struct mon_evt mbm_local_event = {
760 .name = "mbm_local_bytes",
761 .evtid = QOS_L3_MBM_LOCAL_EVENT_ID,
762 };
763
764 /*
765 * Initialize the event list for the resource.
766 *
767 * Note that MBM events are also part of RDT_RESOURCE_L3 resource
768 * because as per the SDM the total and local memory bandwidth
769 * are enumerated as part of L3 monitoring.
770 */
l3_mon_evt_init(struct rdt_resource * r)771 static void l3_mon_evt_init(struct rdt_resource *r)
772 {
773 INIT_LIST_HEAD(&r->evt_list);
774
775 if (is_llc_occupancy_enabled())
776 list_add_tail(&llc_occupancy_event.list, &r->evt_list);
777 if (is_mbm_total_enabled())
778 list_add_tail(&mbm_total_event.list, &r->evt_list);
779 if (is_mbm_local_enabled())
780 list_add_tail(&mbm_local_event.list, &r->evt_list);
781 }
782
rdt_get_mon_l3_config(struct rdt_resource * r)783 int __init rdt_get_mon_l3_config(struct rdt_resource *r)
784 {
785 unsigned int mbm_offset = boot_cpu_data.x86_cache_mbm_width_offset;
786 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
787 unsigned int threshold;
788 int ret;
789
790 resctrl_rmid_realloc_limit = boot_cpu_data.x86_cache_size * 1024;
791 hw_res->mon_scale = boot_cpu_data.x86_cache_occ_scale;
792 r->num_rmid = boot_cpu_data.x86_cache_max_rmid + 1;
793 hw_res->mbm_width = MBM_CNTR_WIDTH_BASE;
794
795 if (mbm_offset > 0 && mbm_offset <= MBM_CNTR_WIDTH_OFFSET_MAX)
796 hw_res->mbm_width += mbm_offset;
797 else if (mbm_offset > MBM_CNTR_WIDTH_OFFSET_MAX)
798 pr_warn("Ignoring impossible MBM counter offset\n");
799
800 /*
801 * A reasonable upper limit on the max threshold is the number
802 * of lines tagged per RMID if all RMIDs have the same number of
803 * lines tagged in the LLC.
804 *
805 * For a 35MB LLC and 56 RMIDs, this is ~1.8% of the LLC.
806 */
807 threshold = resctrl_rmid_realloc_limit / r->num_rmid;
808
809 /*
810 * Because num_rmid may not be a power of two, round the value
811 * to the nearest multiple of hw_res->mon_scale so it matches a
812 * value the hardware will measure. mon_scale may not be a power of 2.
813 */
814 resctrl_rmid_realloc_threshold = resctrl_arch_round_mon_val(threshold);
815
816 ret = dom_data_init(r);
817 if (ret)
818 return ret;
819
820 if (rdt_cpu_has(X86_FEATURE_BMEC)) {
821 if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL)) {
822 mbm_total_event.configurable = true;
823 mbm_config_rftype_init("mbm_total_bytes_config");
824 }
825 if (rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL)) {
826 mbm_local_event.configurable = true;
827 mbm_config_rftype_init("mbm_local_bytes_config");
828 }
829 }
830
831 l3_mon_evt_init(r);
832
833 r->mon_capable = true;
834
835 return 0;
836 }
837
intel_rdt_mbm_apply_quirk(void)838 void __init intel_rdt_mbm_apply_quirk(void)
839 {
840 int cf_index;
841
842 cf_index = (boot_cpu_data.x86_cache_max_rmid + 1) / 8 - 1;
843 if (cf_index >= ARRAY_SIZE(mbm_cf_table)) {
844 pr_info("No MBM correction factor available\n");
845 return;
846 }
847
848 mbm_cf_rmidthreshold = mbm_cf_table[cf_index].rmidthreshold;
849 mbm_cf = mbm_cf_table[cf_index].cf;
850 }
851