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/arch/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
get_mbm_state(struct rdt_domain * d,u32 rmid,enum resctrl_event_id evtid)386 static struct mbm_state *get_mbm_state(struct rdt_domain *d, u32 rmid,
387 enum resctrl_event_id evtid)
388 {
389 switch (evtid) {
390 case QOS_L3_MBM_TOTAL_EVENT_ID:
391 return &d->mbm_total[rmid];
392 case QOS_L3_MBM_LOCAL_EVENT_ID:
393 return &d->mbm_local[rmid];
394 default:
395 return NULL;
396 }
397 }
398
__mon_event_count(u32 rmid,struct rmid_read * rr)399 static int __mon_event_count(u32 rmid, struct rmid_read *rr)
400 {
401 struct mbm_state *m;
402 u64 tval = 0;
403
404 if (rr->first) {
405 resctrl_arch_reset_rmid(rr->r, rr->d, rmid, rr->evtid);
406 m = get_mbm_state(rr->d, rmid, rr->evtid);
407 if (m)
408 memset(m, 0, sizeof(struct mbm_state));
409 return 0;
410 }
411
412 rr->err = resctrl_arch_rmid_read(rr->r, rr->d, rmid, rr->evtid, &tval);
413 if (rr->err)
414 return rr->err;
415
416 rr->val += tval;
417
418 return 0;
419 }
420
421 /*
422 * mbm_bw_count() - Update bw count from values previously read by
423 * __mon_event_count().
424 * @rmid: The rmid used to identify the cached mbm_state.
425 * @rr: The struct rmid_read populated by __mon_event_count().
426 *
427 * Supporting function to calculate the memory bandwidth
428 * and delta bandwidth in MBps. The chunks value previously read by
429 * __mon_event_count() is compared with the chunks value from the previous
430 * invocation. This must be called once per second to maintain values in MBps.
431 */
mbm_bw_count(u32 rmid,struct rmid_read * rr)432 static void mbm_bw_count(u32 rmid, struct rmid_read *rr)
433 {
434 struct mbm_state *m = &rr->d->mbm_local[rmid];
435 u64 cur_bw, bytes, cur_bytes;
436
437 cur_bytes = rr->val;
438 bytes = cur_bytes - m->prev_bw_bytes;
439 m->prev_bw_bytes = cur_bytes;
440
441 cur_bw = bytes / SZ_1M;
442
443 m->prev_bw = cur_bw;
444 }
445
446 /*
447 * This is called via IPI to read the CQM/MBM counters
448 * on a domain.
449 */
mon_event_count(void * info)450 void mon_event_count(void *info)
451 {
452 struct rdtgroup *rdtgrp, *entry;
453 struct rmid_read *rr = info;
454 struct list_head *head;
455 int ret;
456
457 rdtgrp = rr->rgrp;
458
459 ret = __mon_event_count(rdtgrp->mon.rmid, rr);
460
461 /*
462 * For Ctrl groups read data from child monitor groups and
463 * add them together. Count events which are read successfully.
464 * Discard the rmid_read's reporting errors.
465 */
466 head = &rdtgrp->mon.crdtgrp_list;
467
468 if (rdtgrp->type == RDTCTRL_GROUP) {
469 list_for_each_entry(entry, head, mon.crdtgrp_list) {
470 if (__mon_event_count(entry->mon.rmid, rr) == 0)
471 ret = 0;
472 }
473 }
474
475 /*
476 * __mon_event_count() calls for newly created monitor groups may
477 * report -EINVAL/Unavailable if the monitor hasn't seen any traffic.
478 * Discard error if any of the monitor event reads succeeded.
479 */
480 if (ret == 0)
481 rr->err = 0;
482 }
483
484 /*
485 * Feedback loop for MBA software controller (mba_sc)
486 *
487 * mba_sc is a feedback loop where we periodically read MBM counters and
488 * adjust the bandwidth percentage values via the IA32_MBA_THRTL_MSRs so
489 * that:
490 *
491 * current bandwidth(cur_bw) < user specified bandwidth(user_bw)
492 *
493 * This uses the MBM counters to measure the bandwidth and MBA throttle
494 * MSRs to control the bandwidth for a particular rdtgrp. It builds on the
495 * fact that resctrl rdtgroups have both monitoring and control.
496 *
497 * The frequency of the checks is 1s and we just tag along the MBM overflow
498 * timer. Having 1s interval makes the calculation of bandwidth simpler.
499 *
500 * Although MBA's goal is to restrict the bandwidth to a maximum, there may
501 * be a need to increase the bandwidth to avoid unnecessarily restricting
502 * the L2 <-> L3 traffic.
503 *
504 * Since MBA controls the L2 external bandwidth where as MBM measures the
505 * L3 external bandwidth the following sequence could lead to such a
506 * situation.
507 *
508 * Consider an rdtgroup which had high L3 <-> memory traffic in initial
509 * phases -> mba_sc kicks in and reduced bandwidth percentage values -> but
510 * after some time rdtgroup has mostly L2 <-> L3 traffic.
511 *
512 * In this case we may restrict the rdtgroup's L2 <-> L3 traffic as its
513 * throttle MSRs already have low percentage values. To avoid
514 * unnecessarily restricting such rdtgroups, we also increase the bandwidth.
515 */
update_mba_bw(struct rdtgroup * rgrp,struct rdt_domain * dom_mbm)516 static void update_mba_bw(struct rdtgroup *rgrp, struct rdt_domain *dom_mbm)
517 {
518 u32 closid, rmid, cur_msr_val, new_msr_val;
519 struct mbm_state *pmbm_data, *cmbm_data;
520 struct rdt_resource *r_mba;
521 struct rdt_domain *dom_mba;
522 struct list_head *head;
523 struct rdtgroup *entry;
524 u32 cur_bw, user_bw;
525
526 if (!is_mbm_local_enabled())
527 return;
528
529 r_mba = &rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl;
530
531 closid = rgrp->closid;
532 rmid = rgrp->mon.rmid;
533 pmbm_data = &dom_mbm->mbm_local[rmid];
534
535 dom_mba = get_domain_from_cpu(smp_processor_id(), r_mba);
536 if (!dom_mba) {
537 pr_warn_once("Failure to get domain for MBA update\n");
538 return;
539 }
540
541 cur_bw = pmbm_data->prev_bw;
542 user_bw = dom_mba->mbps_val[closid];
543
544 /* MBA resource doesn't support CDP */
545 cur_msr_val = resctrl_arch_get_config(r_mba, dom_mba, closid, CDP_NONE);
546
547 /*
548 * For Ctrl groups read data from child monitor groups.
549 */
550 head = &rgrp->mon.crdtgrp_list;
551 list_for_each_entry(entry, head, mon.crdtgrp_list) {
552 cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
553 cur_bw += cmbm_data->prev_bw;
554 }
555
556 /*
557 * Scale up/down the bandwidth linearly for the ctrl group. The
558 * bandwidth step is the bandwidth granularity specified by the
559 * hardware.
560 * Always increase throttling if current bandwidth is above the
561 * target set by user.
562 * But avoid thrashing up and down on every poll by checking
563 * whether a decrease in throttling is likely to push the group
564 * back over target. E.g. if currently throttling to 30% of bandwidth
565 * on a system with 10% granularity steps, check whether moving to
566 * 40% would go past the limit by multiplying current bandwidth by
567 * "(30 + 10) / 30".
568 */
569 if (cur_msr_val > r_mba->membw.min_bw && user_bw < cur_bw) {
570 new_msr_val = cur_msr_val - r_mba->membw.bw_gran;
571 } else if (cur_msr_val < MAX_MBA_BW &&
572 (user_bw > (cur_bw * (cur_msr_val + r_mba->membw.min_bw) / cur_msr_val))) {
573 new_msr_val = cur_msr_val + r_mba->membw.bw_gran;
574 } else {
575 return;
576 }
577
578 resctrl_arch_update_one(r_mba, dom_mba, closid, CDP_NONE, new_msr_val);
579 }
580
mbm_update(struct rdt_resource * r,struct rdt_domain * d,int rmid)581 static void mbm_update(struct rdt_resource *r, struct rdt_domain *d, int rmid)
582 {
583 struct rmid_read rr;
584
585 rr.first = false;
586 rr.r = r;
587 rr.d = d;
588
589 /*
590 * This is protected from concurrent reads from user
591 * as both the user and we hold the global mutex.
592 */
593 if (is_mbm_total_enabled()) {
594 rr.evtid = QOS_L3_MBM_TOTAL_EVENT_ID;
595 rr.val = 0;
596 __mon_event_count(rmid, &rr);
597 }
598 if (is_mbm_local_enabled()) {
599 rr.evtid = QOS_L3_MBM_LOCAL_EVENT_ID;
600 rr.val = 0;
601 __mon_event_count(rmid, &rr);
602
603 /*
604 * Call the MBA software controller only for the
605 * control groups and when user has enabled
606 * the software controller explicitly.
607 */
608 if (is_mba_sc(NULL))
609 mbm_bw_count(rmid, &rr);
610 }
611 }
612
613 /*
614 * Handler to scan the limbo list and move the RMIDs
615 * to free list whose occupancy < threshold_occupancy.
616 */
cqm_handle_limbo(struct work_struct * work)617 void cqm_handle_limbo(struct work_struct *work)
618 {
619 unsigned long delay = msecs_to_jiffies(CQM_LIMBOCHECK_INTERVAL);
620 int cpu = smp_processor_id();
621 struct rdt_resource *r;
622 struct rdt_domain *d;
623
624 mutex_lock(&rdtgroup_mutex);
625
626 r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
627 d = container_of(work, struct rdt_domain, cqm_limbo.work);
628
629 __check_limbo(d, false);
630
631 if (has_busy_rmid(r, d))
632 schedule_delayed_work_on(cpu, &d->cqm_limbo, delay);
633
634 mutex_unlock(&rdtgroup_mutex);
635 }
636
cqm_setup_limbo_handler(struct rdt_domain * dom,unsigned long delay_ms)637 void cqm_setup_limbo_handler(struct rdt_domain *dom, unsigned long delay_ms)
638 {
639 unsigned long delay = msecs_to_jiffies(delay_ms);
640 int cpu;
641
642 cpu = cpumask_any(&dom->cpu_mask);
643 dom->cqm_work_cpu = cpu;
644
645 schedule_delayed_work_on(cpu, &dom->cqm_limbo, delay);
646 }
647
mbm_handle_overflow(struct work_struct * work)648 void mbm_handle_overflow(struct work_struct *work)
649 {
650 unsigned long delay = msecs_to_jiffies(MBM_OVERFLOW_INTERVAL);
651 struct rdtgroup *prgrp, *crgrp;
652 int cpu = smp_processor_id();
653 struct list_head *head;
654 struct rdt_resource *r;
655 struct rdt_domain *d;
656
657 mutex_lock(&rdtgroup_mutex);
658
659 if (!static_branch_likely(&rdt_mon_enable_key))
660 goto out_unlock;
661
662 r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
663 d = container_of(work, struct rdt_domain, mbm_over.work);
664
665 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
666 mbm_update(r, d, prgrp->mon.rmid);
667
668 head = &prgrp->mon.crdtgrp_list;
669 list_for_each_entry(crgrp, head, mon.crdtgrp_list)
670 mbm_update(r, d, crgrp->mon.rmid);
671
672 if (is_mba_sc(NULL))
673 update_mba_bw(prgrp, d);
674 }
675
676 schedule_delayed_work_on(cpu, &d->mbm_over, delay);
677
678 out_unlock:
679 mutex_unlock(&rdtgroup_mutex);
680 }
681
mbm_setup_overflow_handler(struct rdt_domain * dom,unsigned long delay_ms)682 void mbm_setup_overflow_handler(struct rdt_domain *dom, unsigned long delay_ms)
683 {
684 unsigned long delay = msecs_to_jiffies(delay_ms);
685 int cpu;
686
687 if (!static_branch_likely(&rdt_mon_enable_key))
688 return;
689 cpu = cpumask_any(&dom->cpu_mask);
690 dom->mbm_work_cpu = cpu;
691 schedule_delayed_work_on(cpu, &dom->mbm_over, delay);
692 }
693
dom_data_init(struct rdt_resource * r)694 static int dom_data_init(struct rdt_resource *r)
695 {
696 struct rmid_entry *entry = NULL;
697 int i, nr_rmids;
698
699 nr_rmids = r->num_rmid;
700 rmid_ptrs = kcalloc(nr_rmids, sizeof(struct rmid_entry), GFP_KERNEL);
701 if (!rmid_ptrs)
702 return -ENOMEM;
703
704 for (i = 0; i < nr_rmids; i++) {
705 entry = &rmid_ptrs[i];
706 INIT_LIST_HEAD(&entry->list);
707
708 entry->rmid = i;
709 list_add_tail(&entry->list, &rmid_free_lru);
710 }
711
712 /*
713 * RMID 0 is special and is always allocated. It's used for all
714 * tasks that are not monitored.
715 */
716 entry = __rmid_entry(0);
717 list_del(&entry->list);
718
719 return 0;
720 }
721
722 static struct mon_evt llc_occupancy_event = {
723 .name = "llc_occupancy",
724 .evtid = QOS_L3_OCCUP_EVENT_ID,
725 };
726
727 static struct mon_evt mbm_total_event = {
728 .name = "mbm_total_bytes",
729 .evtid = QOS_L3_MBM_TOTAL_EVENT_ID,
730 };
731
732 static struct mon_evt mbm_local_event = {
733 .name = "mbm_local_bytes",
734 .evtid = QOS_L3_MBM_LOCAL_EVENT_ID,
735 };
736
737 /*
738 * Initialize the event list for the resource.
739 *
740 * Note that MBM events are also part of RDT_RESOURCE_L3 resource
741 * because as per the SDM the total and local memory bandwidth
742 * are enumerated as part of L3 monitoring.
743 */
l3_mon_evt_init(struct rdt_resource * r)744 static void l3_mon_evt_init(struct rdt_resource *r)
745 {
746 INIT_LIST_HEAD(&r->evt_list);
747
748 if (is_llc_occupancy_enabled())
749 list_add_tail(&llc_occupancy_event.list, &r->evt_list);
750 if (is_mbm_total_enabled())
751 list_add_tail(&mbm_total_event.list, &r->evt_list);
752 if (is_mbm_local_enabled())
753 list_add_tail(&mbm_local_event.list, &r->evt_list);
754 }
755
rdt_get_mon_l3_config(struct rdt_resource * r)756 int __init rdt_get_mon_l3_config(struct rdt_resource *r)
757 {
758 unsigned int mbm_offset = boot_cpu_data.x86_cache_mbm_width_offset;
759 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
760 unsigned int threshold;
761 int ret;
762
763 resctrl_rmid_realloc_limit = boot_cpu_data.x86_cache_size * 1024;
764 hw_res->mon_scale = boot_cpu_data.x86_cache_occ_scale;
765 r->num_rmid = boot_cpu_data.x86_cache_max_rmid + 1;
766 hw_res->mbm_width = MBM_CNTR_WIDTH_BASE;
767
768 if (mbm_offset > 0 && mbm_offset <= MBM_CNTR_WIDTH_OFFSET_MAX)
769 hw_res->mbm_width += mbm_offset;
770 else if (mbm_offset > MBM_CNTR_WIDTH_OFFSET_MAX)
771 pr_warn("Ignoring impossible MBM counter offset\n");
772
773 /*
774 * A reasonable upper limit on the max threshold is the number
775 * of lines tagged per RMID if all RMIDs have the same number of
776 * lines tagged in the LLC.
777 *
778 * For a 35MB LLC and 56 RMIDs, this is ~1.8% of the LLC.
779 */
780 threshold = resctrl_rmid_realloc_limit / r->num_rmid;
781
782 /*
783 * Because num_rmid may not be a power of two, round the value
784 * to the nearest multiple of hw_res->mon_scale so it matches a
785 * value the hardware will measure. mon_scale may not be a power of 2.
786 */
787 resctrl_rmid_realloc_threshold = resctrl_arch_round_mon_val(threshold);
788
789 ret = dom_data_init(r);
790 if (ret)
791 return ret;
792
793 if (rdt_cpu_has(X86_FEATURE_BMEC)) {
794 u32 eax, ebx, ecx, edx;
795
796 /* Detect list of bandwidth sources that can be tracked */
797 cpuid_count(0x80000020, 3, &eax, &ebx, &ecx, &edx);
798 hw_res->mbm_cfg_mask = ecx & MAX_EVT_CONFIG_BITS;
799
800 if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL)) {
801 mbm_total_event.configurable = true;
802 mbm_config_rftype_init("mbm_total_bytes_config");
803 }
804 if (rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL)) {
805 mbm_local_event.configurable = true;
806 mbm_config_rftype_init("mbm_local_bytes_config");
807 }
808 }
809
810 l3_mon_evt_init(r);
811
812 r->mon_capable = true;
813
814 return 0;
815 }
816
intel_rdt_mbm_apply_quirk(void)817 void __init intel_rdt_mbm_apply_quirk(void)
818 {
819 int cf_index;
820
821 cf_index = (boot_cpu_data.x86_cache_max_rmid + 1) / 8 - 1;
822 if (cf_index >= ARRAY_SIZE(mbm_cf_table)) {
823 pr_info("No MBM correction factor available\n");
824 return;
825 }
826
827 mbm_cf_rmidthreshold = mbm_cf_table[cf_index].rmidthreshold;
828 mbm_cf = mbm_cf_table[cf_index].cf;
829 }
830