xref: /linux/arch/x86/kernel/cpu/resctrl/monitor.c (revision 2697b79a469b68e3ad3640f55284359c1396278d)
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/cpu.h>
19 #include <linux/module.h>
20 #include <linux/sizes.h>
21 #include <linux/slab.h>
22 
23 #include <asm/cpu_device_id.h>
24 #include <asm/resctrl.h>
25 
26 #include "internal.h"
27 #include "trace.h"
28 
29 /**
30  * struct rmid_entry - dirty tracking for all RMID.
31  * @closid:	The CLOSID for this entry.
32  * @rmid:	The RMID for this entry.
33  * @busy:	The number of domains with cached data using this RMID.
34  * @list:	Member of the rmid_free_lru list when busy == 0.
35  *
36  * Depending on the architecture the correct monitor is accessed using
37  * both @closid and @rmid, or @rmid only.
38  *
39  * Take the rdtgroup_mutex when accessing.
40  */
41 struct rmid_entry {
42 	u32				closid;
43 	u32				rmid;
44 	int				busy;
45 	struct list_head		list;
46 };
47 
48 /*
49  * @rmid_free_lru - A least recently used list of free RMIDs
50  *     These RMIDs are guaranteed to have an occupancy less than the
51  *     threshold occupancy
52  */
53 static LIST_HEAD(rmid_free_lru);
54 
55 /*
56  * @closid_num_dirty_rmid    The number of dirty RMID each CLOSID has.
57  *     Only allocated when CONFIG_RESCTRL_RMID_DEPENDS_ON_CLOSID is defined.
58  *     Indexed by CLOSID. Protected by rdtgroup_mutex.
59  */
60 static u32 *closid_num_dirty_rmid;
61 
62 /*
63  * @rmid_limbo_count - count of currently unused but (potentially)
64  *     dirty RMIDs.
65  *     This counts RMIDs that no one is currently using but that
66  *     may have a occupancy value > resctrl_rmid_realloc_threshold. User can
67  *     change the threshold occupancy value.
68  */
69 static unsigned int rmid_limbo_count;
70 
71 /*
72  * @rmid_entry - The entry in the limbo and free lists.
73  */
74 static struct rmid_entry	*rmid_ptrs;
75 
76 /*
77  * Global boolean for rdt_monitor which is true if any
78  * resource monitoring is enabled.
79  */
80 bool rdt_mon_capable;
81 
82 /*
83  * Global to indicate which monitoring events are enabled.
84  */
85 unsigned int rdt_mon_features;
86 
87 /*
88  * This is the threshold cache occupancy in bytes at which we will consider an
89  * RMID available for re-allocation.
90  */
91 unsigned int resctrl_rmid_realloc_threshold;
92 
93 /*
94  * This is the maximum value for the reallocation threshold, in bytes.
95  */
96 unsigned int resctrl_rmid_realloc_limit;
97 
98 #define CF(cf)	((unsigned long)(1048576 * (cf) + 0.5))
99 
100 /*
101  * The correction factor table is documented in Documentation/arch/x86/resctrl.rst.
102  * If rmid > rmid threshold, MBM total and local values should be multiplied
103  * by the correction factor.
104  *
105  * The original table is modified for better code:
106  *
107  * 1. The threshold 0 is changed to rmid count - 1 so don't do correction
108  *    for the case.
109  * 2. MBM total and local correction table indexed by core counter which is
110  *    equal to (x86_cache_max_rmid + 1) / 8 - 1 and is from 0 up to 27.
111  * 3. The correction factor is normalized to 2^20 (1048576) so it's faster
112  *    to calculate corrected value by shifting:
113  *    corrected_value = (original_value * correction_factor) >> 20
114  */
115 static const struct mbm_correction_factor_table {
116 	u32 rmidthreshold;
117 	u64 cf;
118 } mbm_cf_table[] __initconst = {
119 	{7,	CF(1.000000)},
120 	{15,	CF(1.000000)},
121 	{15,	CF(0.969650)},
122 	{31,	CF(1.000000)},
123 	{31,	CF(1.066667)},
124 	{31,	CF(0.969650)},
125 	{47,	CF(1.142857)},
126 	{63,	CF(1.000000)},
127 	{63,	CF(1.185115)},
128 	{63,	CF(1.066553)},
129 	{79,	CF(1.454545)},
130 	{95,	CF(1.000000)},
131 	{95,	CF(1.230769)},
132 	{95,	CF(1.142857)},
133 	{95,	CF(1.066667)},
134 	{127,	CF(1.000000)},
135 	{127,	CF(1.254863)},
136 	{127,	CF(1.185255)},
137 	{151,	CF(1.000000)},
138 	{127,	CF(1.066667)},
139 	{167,	CF(1.000000)},
140 	{159,	CF(1.454334)},
141 	{183,	CF(1.000000)},
142 	{127,	CF(0.969744)},
143 	{191,	CF(1.280246)},
144 	{191,	CF(1.230921)},
145 	{215,	CF(1.000000)},
146 	{191,	CF(1.143118)},
147 };
148 
149 static u32 mbm_cf_rmidthreshold __read_mostly = UINT_MAX;
150 static u64 mbm_cf __read_mostly;
151 
152 static inline u64 get_corrected_mbm_count(u32 rmid, unsigned long val)
153 {
154 	/* Correct MBM value. */
155 	if (rmid > mbm_cf_rmidthreshold)
156 		val = (val * mbm_cf) >> 20;
157 
158 	return val;
159 }
160 
161 /*
162  * x86 and arm64 differ in their handling of monitoring.
163  * x86's RMID are independent numbers, there is only one source of traffic
164  * with an RMID value of '1'.
165  * arm64's PMG extends the PARTID/CLOSID space, there are multiple sources of
166  * traffic with a PMG value of '1', one for each CLOSID, meaning the RMID
167  * value is no longer unique.
168  * To account for this, resctrl uses an index. On x86 this is just the RMID,
169  * on arm64 it encodes the CLOSID and RMID. This gives a unique number.
170  *
171  * The domain's rmid_busy_llc and rmid_ptrs[] are sized by index. The arch code
172  * must accept an attempt to read every index.
173  */
174 static inline struct rmid_entry *__rmid_entry(u32 idx)
175 {
176 	struct rmid_entry *entry;
177 	u32 closid, rmid;
178 
179 	entry = &rmid_ptrs[idx];
180 	resctrl_arch_rmid_idx_decode(idx, &closid, &rmid);
181 
182 	WARN_ON_ONCE(entry->closid != closid);
183 	WARN_ON_ONCE(entry->rmid != rmid);
184 
185 	return entry;
186 }
187 
188 static int __rmid_read(u32 rmid, enum resctrl_event_id eventid, u64 *val)
189 {
190 	u64 msr_val;
191 
192 	/*
193 	 * As per the SDM, when IA32_QM_EVTSEL.EvtID (bits 7:0) is configured
194 	 * with a valid event code for supported resource type and the bits
195 	 * IA32_QM_EVTSEL.RMID (bits 41:32) are configured with valid RMID,
196 	 * IA32_QM_CTR.data (bits 61:0) reports the monitored data.
197 	 * IA32_QM_CTR.Error (bit 63) and IA32_QM_CTR.Unavailable (bit 62)
198 	 * are error bits.
199 	 */
200 	wrmsr(MSR_IA32_QM_EVTSEL, eventid, rmid);
201 	rdmsrl(MSR_IA32_QM_CTR, msr_val);
202 
203 	if (msr_val & RMID_VAL_ERROR)
204 		return -EIO;
205 	if (msr_val & RMID_VAL_UNAVAIL)
206 		return -EINVAL;
207 
208 	*val = msr_val;
209 	return 0;
210 }
211 
212 static struct arch_mbm_state *get_arch_mbm_state(struct rdt_hw_domain *hw_dom,
213 						 u32 rmid,
214 						 enum resctrl_event_id eventid)
215 {
216 	switch (eventid) {
217 	case QOS_L3_OCCUP_EVENT_ID:
218 		return NULL;
219 	case QOS_L3_MBM_TOTAL_EVENT_ID:
220 		return &hw_dom->arch_mbm_total[rmid];
221 	case QOS_L3_MBM_LOCAL_EVENT_ID:
222 		return &hw_dom->arch_mbm_local[rmid];
223 	}
224 
225 	/* Never expect to get here */
226 	WARN_ON_ONCE(1);
227 
228 	return NULL;
229 }
230 
231 void resctrl_arch_reset_rmid(struct rdt_resource *r, struct rdt_domain *d,
232 			     u32 unused, u32 rmid,
233 			     enum resctrl_event_id eventid)
234 {
235 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
236 	struct arch_mbm_state *am;
237 
238 	am = get_arch_mbm_state(hw_dom, rmid, eventid);
239 	if (am) {
240 		memset(am, 0, sizeof(*am));
241 
242 		/* Record any initial, non-zero count value. */
243 		__rmid_read(rmid, eventid, &am->prev_msr);
244 	}
245 }
246 
247 /*
248  * Assumes that hardware counters are also reset and thus that there is
249  * no need to record initial non-zero counts.
250  */
251 void resctrl_arch_reset_rmid_all(struct rdt_resource *r, struct rdt_domain *d)
252 {
253 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
254 
255 	if (is_mbm_total_enabled())
256 		memset(hw_dom->arch_mbm_total, 0,
257 		       sizeof(*hw_dom->arch_mbm_total) * r->num_rmid);
258 
259 	if (is_mbm_local_enabled())
260 		memset(hw_dom->arch_mbm_local, 0,
261 		       sizeof(*hw_dom->arch_mbm_local) * r->num_rmid);
262 }
263 
264 static u64 mbm_overflow_count(u64 prev_msr, u64 cur_msr, unsigned int width)
265 {
266 	u64 shift = 64 - width, chunks;
267 
268 	chunks = (cur_msr << shift) - (prev_msr << shift);
269 	return chunks >> shift;
270 }
271 
272 int resctrl_arch_rmid_read(struct rdt_resource *r, struct rdt_domain *d,
273 			   u32 unused, u32 rmid, enum resctrl_event_id eventid,
274 			   u64 *val, void *ignored)
275 {
276 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
277 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
278 	struct arch_mbm_state *am;
279 	u64 msr_val, chunks;
280 	int ret;
281 
282 	resctrl_arch_rmid_read_context_check();
283 
284 	if (!cpumask_test_cpu(smp_processor_id(), &d->cpu_mask))
285 		return -EINVAL;
286 
287 	ret = __rmid_read(rmid, eventid, &msr_val);
288 	if (ret)
289 		return ret;
290 
291 	am = get_arch_mbm_state(hw_dom, rmid, eventid);
292 	if (am) {
293 		am->chunks += mbm_overflow_count(am->prev_msr, msr_val,
294 						 hw_res->mbm_width);
295 		chunks = get_corrected_mbm_count(rmid, am->chunks);
296 		am->prev_msr = msr_val;
297 	} else {
298 		chunks = msr_val;
299 	}
300 
301 	*val = chunks * hw_res->mon_scale;
302 
303 	return 0;
304 }
305 
306 static void limbo_release_entry(struct rmid_entry *entry)
307 {
308 	lockdep_assert_held(&rdtgroup_mutex);
309 
310 	rmid_limbo_count--;
311 	list_add_tail(&entry->list, &rmid_free_lru);
312 
313 	if (IS_ENABLED(CONFIG_RESCTRL_RMID_DEPENDS_ON_CLOSID))
314 		closid_num_dirty_rmid[entry->closid]--;
315 }
316 
317 /*
318  * Check the RMIDs that are marked as busy for this domain. If the
319  * reported LLC occupancy is below the threshold clear the busy bit and
320  * decrement the count. If the busy count gets to zero on an RMID, we
321  * free the RMID
322  */
323 void __check_limbo(struct rdt_domain *d, bool force_free)
324 {
325 	struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
326 	u32 idx_limit = resctrl_arch_system_num_rmid_idx();
327 	struct rmid_entry *entry;
328 	u32 idx, cur_idx = 1;
329 	void *arch_mon_ctx;
330 	bool rmid_dirty;
331 	u64 val = 0;
332 
333 	arch_mon_ctx = resctrl_arch_mon_ctx_alloc(r, QOS_L3_OCCUP_EVENT_ID);
334 	if (IS_ERR(arch_mon_ctx)) {
335 		pr_warn_ratelimited("Failed to allocate monitor context: %ld",
336 				    PTR_ERR(arch_mon_ctx));
337 		return;
338 	}
339 
340 	/*
341 	 * Skip RMID 0 and start from RMID 1 and check all the RMIDs that
342 	 * are marked as busy for occupancy < threshold. If the occupancy
343 	 * is less than the threshold decrement the busy counter of the
344 	 * RMID and move it to the free list when the counter reaches 0.
345 	 */
346 	for (;;) {
347 		idx = find_next_bit(d->rmid_busy_llc, idx_limit, cur_idx);
348 		if (idx >= idx_limit)
349 			break;
350 
351 		entry = __rmid_entry(idx);
352 		if (resctrl_arch_rmid_read(r, d, entry->closid, entry->rmid,
353 					   QOS_L3_OCCUP_EVENT_ID, &val,
354 					   arch_mon_ctx)) {
355 			rmid_dirty = true;
356 		} else {
357 			rmid_dirty = (val >= resctrl_rmid_realloc_threshold);
358 
359 			/*
360 			 * x86's CLOSID and RMID are independent numbers, so the entry's
361 			 * CLOSID is an empty CLOSID (X86_RESCTRL_EMPTY_CLOSID). On Arm the
362 			 * RMID (PMG) extends the CLOSID (PARTID) space with bits that aren't
363 			 * used to select the configuration. It is thus necessary to track both
364 			 * CLOSID and RMID because there may be dependencies between them
365 			 * on some architectures.
366 			 */
367 			trace_mon_llc_occupancy_limbo(entry->closid, entry->rmid, d->id, val);
368 		}
369 
370 		if (force_free || !rmid_dirty) {
371 			clear_bit(idx, d->rmid_busy_llc);
372 			if (!--entry->busy)
373 				limbo_release_entry(entry);
374 		}
375 		cur_idx = idx + 1;
376 	}
377 
378 	resctrl_arch_mon_ctx_free(r, QOS_L3_OCCUP_EVENT_ID, arch_mon_ctx);
379 }
380 
381 bool has_busy_rmid(struct rdt_domain *d)
382 {
383 	u32 idx_limit = resctrl_arch_system_num_rmid_idx();
384 
385 	return find_first_bit(d->rmid_busy_llc, idx_limit) != idx_limit;
386 }
387 
388 static struct rmid_entry *resctrl_find_free_rmid(u32 closid)
389 {
390 	struct rmid_entry *itr;
391 	u32 itr_idx, cmp_idx;
392 
393 	if (list_empty(&rmid_free_lru))
394 		return rmid_limbo_count ? ERR_PTR(-EBUSY) : ERR_PTR(-ENOSPC);
395 
396 	list_for_each_entry(itr, &rmid_free_lru, list) {
397 		/*
398 		 * Get the index of this free RMID, and the index it would need
399 		 * to be if it were used with this CLOSID.
400 		 * If the CLOSID is irrelevant on this architecture, the two
401 		 * index values are always the same on every entry and thus the
402 		 * very first entry will be returned.
403 		 */
404 		itr_idx = resctrl_arch_rmid_idx_encode(itr->closid, itr->rmid);
405 		cmp_idx = resctrl_arch_rmid_idx_encode(closid, itr->rmid);
406 
407 		if (itr_idx == cmp_idx)
408 			return itr;
409 	}
410 
411 	return ERR_PTR(-ENOSPC);
412 }
413 
414 /**
415  * resctrl_find_cleanest_closid() - Find a CLOSID where all the associated
416  *                                  RMID are clean, or the CLOSID that has
417  *                                  the most clean RMID.
418  *
419  * MPAM's equivalent of RMID are per-CLOSID, meaning a freshly allocated CLOSID
420  * may not be able to allocate clean RMID. To avoid this the allocator will
421  * choose the CLOSID with the most clean RMID.
422  *
423  * When the CLOSID and RMID are independent numbers, the first free CLOSID will
424  * be returned.
425  */
426 int resctrl_find_cleanest_closid(void)
427 {
428 	u32 cleanest_closid = ~0;
429 	int i = 0;
430 
431 	lockdep_assert_held(&rdtgroup_mutex);
432 
433 	if (!IS_ENABLED(CONFIG_RESCTRL_RMID_DEPENDS_ON_CLOSID))
434 		return -EIO;
435 
436 	for (i = 0; i < closids_supported(); i++) {
437 		int num_dirty;
438 
439 		if (closid_allocated(i))
440 			continue;
441 
442 		num_dirty = closid_num_dirty_rmid[i];
443 		if (num_dirty == 0)
444 			return i;
445 
446 		if (cleanest_closid == ~0)
447 			cleanest_closid = i;
448 
449 		if (num_dirty < closid_num_dirty_rmid[cleanest_closid])
450 			cleanest_closid = i;
451 	}
452 
453 	if (cleanest_closid == ~0)
454 		return -ENOSPC;
455 
456 	return cleanest_closid;
457 }
458 
459 /*
460  * For MPAM the RMID value is not unique, and has to be considered with
461  * the CLOSID. The (CLOSID, RMID) pair is allocated on all domains, which
462  * allows all domains to be managed by a single free list.
463  * Each domain also has a rmid_busy_llc to reduce the work of the limbo handler.
464  */
465 int alloc_rmid(u32 closid)
466 {
467 	struct rmid_entry *entry;
468 
469 	lockdep_assert_held(&rdtgroup_mutex);
470 
471 	entry = resctrl_find_free_rmid(closid);
472 	if (IS_ERR(entry))
473 		return PTR_ERR(entry);
474 
475 	list_del(&entry->list);
476 	return entry->rmid;
477 }
478 
479 static void add_rmid_to_limbo(struct rmid_entry *entry)
480 {
481 	struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
482 	struct rdt_domain *d;
483 	u32 idx;
484 
485 	lockdep_assert_held(&rdtgroup_mutex);
486 
487 	/* Walking r->domains, ensure it can't race with cpuhp */
488 	lockdep_assert_cpus_held();
489 
490 	idx = resctrl_arch_rmid_idx_encode(entry->closid, entry->rmid);
491 
492 	entry->busy = 0;
493 	list_for_each_entry(d, &r->domains, list) {
494 		/*
495 		 * For the first limbo RMID in the domain,
496 		 * setup up the limbo worker.
497 		 */
498 		if (!has_busy_rmid(d))
499 			cqm_setup_limbo_handler(d, CQM_LIMBOCHECK_INTERVAL,
500 						RESCTRL_PICK_ANY_CPU);
501 		set_bit(idx, d->rmid_busy_llc);
502 		entry->busy++;
503 	}
504 
505 	rmid_limbo_count++;
506 	if (IS_ENABLED(CONFIG_RESCTRL_RMID_DEPENDS_ON_CLOSID))
507 		closid_num_dirty_rmid[entry->closid]++;
508 }
509 
510 void free_rmid(u32 closid, u32 rmid)
511 {
512 	u32 idx = resctrl_arch_rmid_idx_encode(closid, rmid);
513 	struct rmid_entry *entry;
514 
515 	lockdep_assert_held(&rdtgroup_mutex);
516 
517 	/*
518 	 * Do not allow the default rmid to be free'd. Comparing by index
519 	 * allows architectures that ignore the closid parameter to avoid an
520 	 * unnecessary check.
521 	 */
522 	if (idx == resctrl_arch_rmid_idx_encode(RESCTRL_RESERVED_CLOSID,
523 						RESCTRL_RESERVED_RMID))
524 		return;
525 
526 	entry = __rmid_entry(idx);
527 
528 	if (is_llc_occupancy_enabled())
529 		add_rmid_to_limbo(entry);
530 	else
531 		list_add_tail(&entry->list, &rmid_free_lru);
532 }
533 
534 static struct mbm_state *get_mbm_state(struct rdt_domain *d, u32 closid,
535 				       u32 rmid, enum resctrl_event_id evtid)
536 {
537 	u32 idx = resctrl_arch_rmid_idx_encode(closid, rmid);
538 
539 	switch (evtid) {
540 	case QOS_L3_MBM_TOTAL_EVENT_ID:
541 		return &d->mbm_total[idx];
542 	case QOS_L3_MBM_LOCAL_EVENT_ID:
543 		return &d->mbm_local[idx];
544 	default:
545 		return NULL;
546 	}
547 }
548 
549 static int __mon_event_count(u32 closid, u32 rmid, struct rmid_read *rr)
550 {
551 	struct mbm_state *m;
552 	u64 tval = 0;
553 
554 	if (rr->first) {
555 		resctrl_arch_reset_rmid(rr->r, rr->d, closid, rmid, rr->evtid);
556 		m = get_mbm_state(rr->d, closid, rmid, rr->evtid);
557 		if (m)
558 			memset(m, 0, sizeof(struct mbm_state));
559 		return 0;
560 	}
561 
562 	rr->err = resctrl_arch_rmid_read(rr->r, rr->d, closid, rmid, rr->evtid,
563 					 &tval, rr->arch_mon_ctx);
564 	if (rr->err)
565 		return rr->err;
566 
567 	rr->val += tval;
568 
569 	return 0;
570 }
571 
572 /*
573  * mbm_bw_count() - Update bw count from values previously read by
574  *		    __mon_event_count().
575  * @closid:	The closid used to identify the cached mbm_state.
576  * @rmid:	The rmid used to identify the cached mbm_state.
577  * @rr:		The struct rmid_read populated by __mon_event_count().
578  *
579  * Supporting function to calculate the memory bandwidth
580  * and delta bandwidth in MBps. The chunks value previously read by
581  * __mon_event_count() is compared with the chunks value from the previous
582  * invocation. This must be called once per second to maintain values in MBps.
583  */
584 static void mbm_bw_count(u32 closid, u32 rmid, struct rmid_read *rr)
585 {
586 	u32 idx = resctrl_arch_rmid_idx_encode(closid, rmid);
587 	struct mbm_state *m = &rr->d->mbm_local[idx];
588 	u64 cur_bw, bytes, cur_bytes;
589 
590 	cur_bytes = rr->val;
591 	bytes = cur_bytes - m->prev_bw_bytes;
592 	m->prev_bw_bytes = cur_bytes;
593 
594 	cur_bw = bytes / SZ_1M;
595 
596 	m->prev_bw = cur_bw;
597 }
598 
599 /*
600  * This is scheduled by mon_event_read() to read the CQM/MBM counters
601  * on a domain.
602  */
603 void mon_event_count(void *info)
604 {
605 	struct rdtgroup *rdtgrp, *entry;
606 	struct rmid_read *rr = info;
607 	struct list_head *head;
608 	int ret;
609 
610 	rdtgrp = rr->rgrp;
611 
612 	ret = __mon_event_count(rdtgrp->closid, rdtgrp->mon.rmid, rr);
613 
614 	/*
615 	 * For Ctrl groups read data from child monitor groups and
616 	 * add them together. Count events which are read successfully.
617 	 * Discard the rmid_read's reporting errors.
618 	 */
619 	head = &rdtgrp->mon.crdtgrp_list;
620 
621 	if (rdtgrp->type == RDTCTRL_GROUP) {
622 		list_for_each_entry(entry, head, mon.crdtgrp_list) {
623 			if (__mon_event_count(entry->closid, entry->mon.rmid,
624 					      rr) == 0)
625 				ret = 0;
626 		}
627 	}
628 
629 	/*
630 	 * __mon_event_count() calls for newly created monitor groups may
631 	 * report -EINVAL/Unavailable if the monitor hasn't seen any traffic.
632 	 * Discard error if any of the monitor event reads succeeded.
633 	 */
634 	if (ret == 0)
635 		rr->err = 0;
636 }
637 
638 /*
639  * Feedback loop for MBA software controller (mba_sc)
640  *
641  * mba_sc is a feedback loop where we periodically read MBM counters and
642  * adjust the bandwidth percentage values via the IA32_MBA_THRTL_MSRs so
643  * that:
644  *
645  *   current bandwidth(cur_bw) < user specified bandwidth(user_bw)
646  *
647  * This uses the MBM counters to measure the bandwidth and MBA throttle
648  * MSRs to control the bandwidth for a particular rdtgrp. It builds on the
649  * fact that resctrl rdtgroups have both monitoring and control.
650  *
651  * The frequency of the checks is 1s and we just tag along the MBM overflow
652  * timer. Having 1s interval makes the calculation of bandwidth simpler.
653  *
654  * Although MBA's goal is to restrict the bandwidth to a maximum, there may
655  * be a need to increase the bandwidth to avoid unnecessarily restricting
656  * the L2 <-> L3 traffic.
657  *
658  * Since MBA controls the L2 external bandwidth where as MBM measures the
659  * L3 external bandwidth the following sequence could lead to such a
660  * situation.
661  *
662  * Consider an rdtgroup which had high L3 <-> memory traffic in initial
663  * phases -> mba_sc kicks in and reduced bandwidth percentage values -> but
664  * after some time rdtgroup has mostly L2 <-> L3 traffic.
665  *
666  * In this case we may restrict the rdtgroup's L2 <-> L3 traffic as its
667  * throttle MSRs already have low percentage values.  To avoid
668  * unnecessarily restricting such rdtgroups, we also increase the bandwidth.
669  */
670 static void update_mba_bw(struct rdtgroup *rgrp, struct rdt_domain *dom_mbm)
671 {
672 	u32 closid, rmid, cur_msr_val, new_msr_val;
673 	struct mbm_state *pmbm_data, *cmbm_data;
674 	struct rdt_resource *r_mba;
675 	struct rdt_domain *dom_mba;
676 	u32 cur_bw, user_bw, idx;
677 	struct list_head *head;
678 	struct rdtgroup *entry;
679 
680 	if (!is_mbm_local_enabled())
681 		return;
682 
683 	r_mba = &rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl;
684 
685 	closid = rgrp->closid;
686 	rmid = rgrp->mon.rmid;
687 	idx = resctrl_arch_rmid_idx_encode(closid, rmid);
688 	pmbm_data = &dom_mbm->mbm_local[idx];
689 
690 	dom_mba = get_domain_from_cpu(smp_processor_id(), r_mba);
691 	if (!dom_mba) {
692 		pr_warn_once("Failure to get domain for MBA update\n");
693 		return;
694 	}
695 
696 	cur_bw = pmbm_data->prev_bw;
697 	user_bw = dom_mba->mbps_val[closid];
698 
699 	/* MBA resource doesn't support CDP */
700 	cur_msr_val = resctrl_arch_get_config(r_mba, dom_mba, closid, CDP_NONE);
701 
702 	/*
703 	 * For Ctrl groups read data from child monitor groups.
704 	 */
705 	head = &rgrp->mon.crdtgrp_list;
706 	list_for_each_entry(entry, head, mon.crdtgrp_list) {
707 		cmbm_data = &dom_mbm->mbm_local[entry->mon.rmid];
708 		cur_bw += cmbm_data->prev_bw;
709 	}
710 
711 	/*
712 	 * Scale up/down the bandwidth linearly for the ctrl group.  The
713 	 * bandwidth step is the bandwidth granularity specified by the
714 	 * hardware.
715 	 * Always increase throttling if current bandwidth is above the
716 	 * target set by user.
717 	 * But avoid thrashing up and down on every poll by checking
718 	 * whether a decrease in throttling is likely to push the group
719 	 * back over target. E.g. if currently throttling to 30% of bandwidth
720 	 * on a system with 10% granularity steps, check whether moving to
721 	 * 40% would go past the limit by multiplying current bandwidth by
722 	 * "(30 + 10) / 30".
723 	 */
724 	if (cur_msr_val > r_mba->membw.min_bw && user_bw < cur_bw) {
725 		new_msr_val = cur_msr_val - r_mba->membw.bw_gran;
726 	} else if (cur_msr_val < MAX_MBA_BW &&
727 		   (user_bw > (cur_bw * (cur_msr_val + r_mba->membw.min_bw) / cur_msr_val))) {
728 		new_msr_val = cur_msr_val + r_mba->membw.bw_gran;
729 	} else {
730 		return;
731 	}
732 
733 	resctrl_arch_update_one(r_mba, dom_mba, closid, CDP_NONE, new_msr_val);
734 }
735 
736 static void mbm_update(struct rdt_resource *r, struct rdt_domain *d,
737 		       u32 closid, u32 rmid)
738 {
739 	struct rmid_read rr;
740 
741 	rr.first = false;
742 	rr.r = r;
743 	rr.d = d;
744 
745 	/*
746 	 * This is protected from concurrent reads from user
747 	 * as both the user and we hold the global mutex.
748 	 */
749 	if (is_mbm_total_enabled()) {
750 		rr.evtid = QOS_L3_MBM_TOTAL_EVENT_ID;
751 		rr.val = 0;
752 		rr.arch_mon_ctx = resctrl_arch_mon_ctx_alloc(rr.r, rr.evtid);
753 		if (IS_ERR(rr.arch_mon_ctx)) {
754 			pr_warn_ratelimited("Failed to allocate monitor context: %ld",
755 					    PTR_ERR(rr.arch_mon_ctx));
756 			return;
757 		}
758 
759 		__mon_event_count(closid, rmid, &rr);
760 
761 		resctrl_arch_mon_ctx_free(rr.r, rr.evtid, rr.arch_mon_ctx);
762 	}
763 	if (is_mbm_local_enabled()) {
764 		rr.evtid = QOS_L3_MBM_LOCAL_EVENT_ID;
765 		rr.val = 0;
766 		rr.arch_mon_ctx = resctrl_arch_mon_ctx_alloc(rr.r, rr.evtid);
767 		if (IS_ERR(rr.arch_mon_ctx)) {
768 			pr_warn_ratelimited("Failed to allocate monitor context: %ld",
769 					    PTR_ERR(rr.arch_mon_ctx));
770 			return;
771 		}
772 
773 		__mon_event_count(closid, rmid, &rr);
774 
775 		/*
776 		 * Call the MBA software controller only for the
777 		 * control groups and when user has enabled
778 		 * the software controller explicitly.
779 		 */
780 		if (is_mba_sc(NULL))
781 			mbm_bw_count(closid, rmid, &rr);
782 
783 		resctrl_arch_mon_ctx_free(rr.r, rr.evtid, rr.arch_mon_ctx);
784 	}
785 }
786 
787 /*
788  * Handler to scan the limbo list and move the RMIDs
789  * to free list whose occupancy < threshold_occupancy.
790  */
791 void cqm_handle_limbo(struct work_struct *work)
792 {
793 	unsigned long delay = msecs_to_jiffies(CQM_LIMBOCHECK_INTERVAL);
794 	struct rdt_domain *d;
795 
796 	cpus_read_lock();
797 	mutex_lock(&rdtgroup_mutex);
798 
799 	d = container_of(work, struct rdt_domain, cqm_limbo.work);
800 
801 	__check_limbo(d, false);
802 
803 	if (has_busy_rmid(d)) {
804 		d->cqm_work_cpu = cpumask_any_housekeeping(&d->cpu_mask,
805 							   RESCTRL_PICK_ANY_CPU);
806 		schedule_delayed_work_on(d->cqm_work_cpu, &d->cqm_limbo,
807 					 delay);
808 	}
809 
810 	mutex_unlock(&rdtgroup_mutex);
811 	cpus_read_unlock();
812 }
813 
814 /**
815  * cqm_setup_limbo_handler() - Schedule the limbo handler to run for this
816  *                             domain.
817  * @dom:           The domain the limbo handler should run for.
818  * @delay_ms:      How far in the future the handler should run.
819  * @exclude_cpu:   Which CPU the handler should not run on,
820  *		   RESCTRL_PICK_ANY_CPU to pick any CPU.
821  */
822 void cqm_setup_limbo_handler(struct rdt_domain *dom, unsigned long delay_ms,
823 			     int exclude_cpu)
824 {
825 	unsigned long delay = msecs_to_jiffies(delay_ms);
826 	int cpu;
827 
828 	cpu = cpumask_any_housekeeping(&dom->cpu_mask, exclude_cpu);
829 	dom->cqm_work_cpu = cpu;
830 
831 	if (cpu < nr_cpu_ids)
832 		schedule_delayed_work_on(cpu, &dom->cqm_limbo, delay);
833 }
834 
835 void mbm_handle_overflow(struct work_struct *work)
836 {
837 	unsigned long delay = msecs_to_jiffies(MBM_OVERFLOW_INTERVAL);
838 	struct rdtgroup *prgrp, *crgrp;
839 	struct list_head *head;
840 	struct rdt_resource *r;
841 	struct rdt_domain *d;
842 
843 	cpus_read_lock();
844 	mutex_lock(&rdtgroup_mutex);
845 
846 	/*
847 	 * If the filesystem has been unmounted this work no longer needs to
848 	 * run.
849 	 */
850 	if (!resctrl_mounted || !resctrl_arch_mon_capable())
851 		goto out_unlock;
852 
853 	r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
854 	d = container_of(work, struct rdt_domain, mbm_over.work);
855 
856 	list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
857 		mbm_update(r, d, prgrp->closid, prgrp->mon.rmid);
858 
859 		head = &prgrp->mon.crdtgrp_list;
860 		list_for_each_entry(crgrp, head, mon.crdtgrp_list)
861 			mbm_update(r, d, crgrp->closid, crgrp->mon.rmid);
862 
863 		if (is_mba_sc(NULL))
864 			update_mba_bw(prgrp, d);
865 	}
866 
867 	/*
868 	 * Re-check for housekeeping CPUs. This allows the overflow handler to
869 	 * move off a nohz_full CPU quickly.
870 	 */
871 	d->mbm_work_cpu = cpumask_any_housekeeping(&d->cpu_mask,
872 						   RESCTRL_PICK_ANY_CPU);
873 	schedule_delayed_work_on(d->mbm_work_cpu, &d->mbm_over, delay);
874 
875 out_unlock:
876 	mutex_unlock(&rdtgroup_mutex);
877 	cpus_read_unlock();
878 }
879 
880 /**
881  * mbm_setup_overflow_handler() - Schedule the overflow handler to run for this
882  *                                domain.
883  * @dom:           The domain the overflow handler should run for.
884  * @delay_ms:      How far in the future the handler should run.
885  * @exclude_cpu:   Which CPU the handler should not run on,
886  *		   RESCTRL_PICK_ANY_CPU to pick any CPU.
887  */
888 void mbm_setup_overflow_handler(struct rdt_domain *dom, unsigned long delay_ms,
889 				int exclude_cpu)
890 {
891 	unsigned long delay = msecs_to_jiffies(delay_ms);
892 	int cpu;
893 
894 	/*
895 	 * When a domain comes online there is no guarantee the filesystem is
896 	 * mounted. If not, there is no need to catch counter overflow.
897 	 */
898 	if (!resctrl_mounted || !resctrl_arch_mon_capable())
899 		return;
900 	cpu = cpumask_any_housekeeping(&dom->cpu_mask, exclude_cpu);
901 	dom->mbm_work_cpu = cpu;
902 
903 	if (cpu < nr_cpu_ids)
904 		schedule_delayed_work_on(cpu, &dom->mbm_over, delay);
905 }
906 
907 static int dom_data_init(struct rdt_resource *r)
908 {
909 	u32 idx_limit = resctrl_arch_system_num_rmid_idx();
910 	u32 num_closid = resctrl_arch_get_num_closid(r);
911 	struct rmid_entry *entry = NULL;
912 	int err = 0, i;
913 	u32 idx;
914 
915 	mutex_lock(&rdtgroup_mutex);
916 	if (IS_ENABLED(CONFIG_RESCTRL_RMID_DEPENDS_ON_CLOSID)) {
917 		u32 *tmp;
918 
919 		/*
920 		 * If the architecture hasn't provided a sanitised value here,
921 		 * this may result in larger arrays than necessary. Resctrl will
922 		 * use a smaller system wide value based on the resources in
923 		 * use.
924 		 */
925 		tmp = kcalloc(num_closid, sizeof(*tmp), GFP_KERNEL);
926 		if (!tmp) {
927 			err = -ENOMEM;
928 			goto out_unlock;
929 		}
930 
931 		closid_num_dirty_rmid = tmp;
932 	}
933 
934 	rmid_ptrs = kcalloc(idx_limit, sizeof(struct rmid_entry), GFP_KERNEL);
935 	if (!rmid_ptrs) {
936 		if (IS_ENABLED(CONFIG_RESCTRL_RMID_DEPENDS_ON_CLOSID)) {
937 			kfree(closid_num_dirty_rmid);
938 			closid_num_dirty_rmid = NULL;
939 		}
940 		err = -ENOMEM;
941 		goto out_unlock;
942 	}
943 
944 	for (i = 0; i < idx_limit; i++) {
945 		entry = &rmid_ptrs[i];
946 		INIT_LIST_HEAD(&entry->list);
947 
948 		resctrl_arch_rmid_idx_decode(i, &entry->closid, &entry->rmid);
949 		list_add_tail(&entry->list, &rmid_free_lru);
950 	}
951 
952 	/*
953 	 * RESCTRL_RESERVED_CLOSID and RESCTRL_RESERVED_RMID are special and
954 	 * are always allocated. These are used for the rdtgroup_default
955 	 * control group, which will be setup later in rdtgroup_init().
956 	 */
957 	idx = resctrl_arch_rmid_idx_encode(RESCTRL_RESERVED_CLOSID,
958 					   RESCTRL_RESERVED_RMID);
959 	entry = __rmid_entry(idx);
960 	list_del(&entry->list);
961 
962 out_unlock:
963 	mutex_unlock(&rdtgroup_mutex);
964 
965 	return err;
966 }
967 
968 static void __exit dom_data_exit(void)
969 {
970 	mutex_lock(&rdtgroup_mutex);
971 
972 	if (IS_ENABLED(CONFIG_RESCTRL_RMID_DEPENDS_ON_CLOSID)) {
973 		kfree(closid_num_dirty_rmid);
974 		closid_num_dirty_rmid = NULL;
975 	}
976 
977 	kfree(rmid_ptrs);
978 	rmid_ptrs = NULL;
979 
980 	mutex_unlock(&rdtgroup_mutex);
981 }
982 
983 static struct mon_evt llc_occupancy_event = {
984 	.name		= "llc_occupancy",
985 	.evtid		= QOS_L3_OCCUP_EVENT_ID,
986 };
987 
988 static struct mon_evt mbm_total_event = {
989 	.name		= "mbm_total_bytes",
990 	.evtid		= QOS_L3_MBM_TOTAL_EVENT_ID,
991 };
992 
993 static struct mon_evt mbm_local_event = {
994 	.name		= "mbm_local_bytes",
995 	.evtid		= QOS_L3_MBM_LOCAL_EVENT_ID,
996 };
997 
998 /*
999  * Initialize the event list for the resource.
1000  *
1001  * Note that MBM events are also part of RDT_RESOURCE_L3 resource
1002  * because as per the SDM the total and local memory bandwidth
1003  * are enumerated as part of L3 monitoring.
1004  */
1005 static void l3_mon_evt_init(struct rdt_resource *r)
1006 {
1007 	INIT_LIST_HEAD(&r->evt_list);
1008 
1009 	if (is_llc_occupancy_enabled())
1010 		list_add_tail(&llc_occupancy_event.list, &r->evt_list);
1011 	if (is_mbm_total_enabled())
1012 		list_add_tail(&mbm_total_event.list, &r->evt_list);
1013 	if (is_mbm_local_enabled())
1014 		list_add_tail(&mbm_local_event.list, &r->evt_list);
1015 }
1016 
1017 int __init rdt_get_mon_l3_config(struct rdt_resource *r)
1018 {
1019 	unsigned int mbm_offset = boot_cpu_data.x86_cache_mbm_width_offset;
1020 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
1021 	unsigned int threshold;
1022 	int ret;
1023 
1024 	resctrl_rmid_realloc_limit = boot_cpu_data.x86_cache_size * 1024;
1025 	hw_res->mon_scale = boot_cpu_data.x86_cache_occ_scale;
1026 	r->num_rmid = boot_cpu_data.x86_cache_max_rmid + 1;
1027 	hw_res->mbm_width = MBM_CNTR_WIDTH_BASE;
1028 
1029 	if (mbm_offset > 0 && mbm_offset <= MBM_CNTR_WIDTH_OFFSET_MAX)
1030 		hw_res->mbm_width += mbm_offset;
1031 	else if (mbm_offset > MBM_CNTR_WIDTH_OFFSET_MAX)
1032 		pr_warn("Ignoring impossible MBM counter offset\n");
1033 
1034 	/*
1035 	 * A reasonable upper limit on the max threshold is the number
1036 	 * of lines tagged per RMID if all RMIDs have the same number of
1037 	 * lines tagged in the LLC.
1038 	 *
1039 	 * For a 35MB LLC and 56 RMIDs, this is ~1.8% of the LLC.
1040 	 */
1041 	threshold = resctrl_rmid_realloc_limit / r->num_rmid;
1042 
1043 	/*
1044 	 * Because num_rmid may not be a power of two, round the value
1045 	 * to the nearest multiple of hw_res->mon_scale so it matches a
1046 	 * value the hardware will measure. mon_scale may not be a power of 2.
1047 	 */
1048 	resctrl_rmid_realloc_threshold = resctrl_arch_round_mon_val(threshold);
1049 
1050 	ret = dom_data_init(r);
1051 	if (ret)
1052 		return ret;
1053 
1054 	if (rdt_cpu_has(X86_FEATURE_BMEC)) {
1055 		u32 eax, ebx, ecx, edx;
1056 
1057 		/* Detect list of bandwidth sources that can be tracked */
1058 		cpuid_count(0x80000020, 3, &eax, &ebx, &ecx, &edx);
1059 		hw_res->mbm_cfg_mask = ecx & MAX_EVT_CONFIG_BITS;
1060 
1061 		if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL)) {
1062 			mbm_total_event.configurable = true;
1063 			mbm_config_rftype_init("mbm_total_bytes_config");
1064 		}
1065 		if (rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL)) {
1066 			mbm_local_event.configurable = true;
1067 			mbm_config_rftype_init("mbm_local_bytes_config");
1068 		}
1069 	}
1070 
1071 	l3_mon_evt_init(r);
1072 
1073 	r->mon_capable = true;
1074 
1075 	return 0;
1076 }
1077 
1078 void __exit rdt_put_mon_l3_config(void)
1079 {
1080 	dom_data_exit();
1081 }
1082 
1083 void __init intel_rdt_mbm_apply_quirk(void)
1084 {
1085 	int cf_index;
1086 
1087 	cf_index = (boot_cpu_data.x86_cache_max_rmid + 1) / 8 - 1;
1088 	if (cf_index >= ARRAY_SIZE(mbm_cf_table)) {
1089 		pr_info("No MBM correction factor available\n");
1090 		return;
1091 	}
1092 
1093 	mbm_cf_rmidthreshold = mbm_cf_table[cf_index].rmidthreshold;
1094 	mbm_cf = mbm_cf_table[cf_index].cf;
1095 }
1096