xref: /openbmc/linux/arch/x86/kernel/cpu/resctrl/core.c (revision 547840bd)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Resource Director Technology(RDT)
4  * - Cache Allocation code.
5  *
6  * Copyright (C) 2016 Intel Corporation
7  *
8  * Authors:
9  *    Fenghua Yu <fenghua.yu@intel.com>
10  *    Tony Luck <tony.luck@intel.com>
11  *    Vikas Shivappa <vikas.shivappa@intel.com>
12  *
13  * More information about RDT be found in the Intel (R) x86 Architecture
14  * Software Developer Manual June 2016, volume 3, section 17.17.
15  */
16 
17 #define pr_fmt(fmt)	"resctrl: " fmt
18 
19 #include <linux/slab.h>
20 #include <linux/err.h>
21 #include <linux/cacheinfo.h>
22 #include <linux/cpuhotplug.h>
23 
24 #include <asm/intel-family.h>
25 #include <asm/resctrl.h>
26 #include "internal.h"
27 
28 /* Mutex to protect rdtgroup access. */
29 DEFINE_MUTEX(rdtgroup_mutex);
30 
31 /*
32  * The cached resctrl_pqr_state is strictly per CPU and can never be
33  * updated from a remote CPU. Functions which modify the state
34  * are called with interrupts disabled and no preemption, which
35  * is sufficient for the protection.
36  */
37 DEFINE_PER_CPU(struct resctrl_pqr_state, pqr_state);
38 
39 /*
40  * Used to store the max resource name width and max resource data width
41  * to display the schemata in a tabular format
42  */
43 int max_name_width, max_data_width;
44 
45 /*
46  * Global boolean for rdt_alloc which is true if any
47  * resource allocation is enabled.
48  */
49 bool rdt_alloc_capable;
50 
51 static void
52 mba_wrmsr_intel(struct rdt_domain *d, struct msr_param *m,
53 		struct rdt_resource *r);
54 static void
55 cat_wrmsr(struct rdt_domain *d, struct msr_param *m, struct rdt_resource *r);
56 static void
57 mba_wrmsr_amd(struct rdt_domain *d, struct msr_param *m,
58 	      struct rdt_resource *r);
59 
60 #define domain_init(id) LIST_HEAD_INIT(rdt_resources_all[id].domains)
61 
62 struct rdt_resource rdt_resources_all[] = {
63 	[RDT_RESOURCE_L3] =
64 	{
65 		.rid			= RDT_RESOURCE_L3,
66 		.name			= "L3",
67 		.domains		= domain_init(RDT_RESOURCE_L3),
68 		.msr_base		= MSR_IA32_L3_CBM_BASE,
69 		.msr_update		= cat_wrmsr,
70 		.cache_level		= 3,
71 		.cache = {
72 			.min_cbm_bits	= 1,
73 			.cbm_idx_mult	= 1,
74 			.cbm_idx_offset	= 0,
75 		},
76 		.parse_ctrlval		= parse_cbm,
77 		.format_str		= "%d=%0*x",
78 		.fflags			= RFTYPE_RES_CACHE,
79 	},
80 	[RDT_RESOURCE_L3DATA] =
81 	{
82 		.rid			= RDT_RESOURCE_L3DATA,
83 		.name			= "L3DATA",
84 		.domains		= domain_init(RDT_RESOURCE_L3DATA),
85 		.msr_base		= MSR_IA32_L3_CBM_BASE,
86 		.msr_update		= cat_wrmsr,
87 		.cache_level		= 3,
88 		.cache = {
89 			.min_cbm_bits	= 1,
90 			.cbm_idx_mult	= 2,
91 			.cbm_idx_offset	= 0,
92 		},
93 		.parse_ctrlval		= parse_cbm,
94 		.format_str		= "%d=%0*x",
95 		.fflags			= RFTYPE_RES_CACHE,
96 	},
97 	[RDT_RESOURCE_L3CODE] =
98 	{
99 		.rid			= RDT_RESOURCE_L3CODE,
100 		.name			= "L3CODE",
101 		.domains		= domain_init(RDT_RESOURCE_L3CODE),
102 		.msr_base		= MSR_IA32_L3_CBM_BASE,
103 		.msr_update		= cat_wrmsr,
104 		.cache_level		= 3,
105 		.cache = {
106 			.min_cbm_bits	= 1,
107 			.cbm_idx_mult	= 2,
108 			.cbm_idx_offset	= 1,
109 		},
110 		.parse_ctrlval		= parse_cbm,
111 		.format_str		= "%d=%0*x",
112 		.fflags			= RFTYPE_RES_CACHE,
113 	},
114 	[RDT_RESOURCE_L2] =
115 	{
116 		.rid			= RDT_RESOURCE_L2,
117 		.name			= "L2",
118 		.domains		= domain_init(RDT_RESOURCE_L2),
119 		.msr_base		= MSR_IA32_L2_CBM_BASE,
120 		.msr_update		= cat_wrmsr,
121 		.cache_level		= 2,
122 		.cache = {
123 			.min_cbm_bits	= 1,
124 			.cbm_idx_mult	= 1,
125 			.cbm_idx_offset	= 0,
126 		},
127 		.parse_ctrlval		= parse_cbm,
128 		.format_str		= "%d=%0*x",
129 		.fflags			= RFTYPE_RES_CACHE,
130 	},
131 	[RDT_RESOURCE_L2DATA] =
132 	{
133 		.rid			= RDT_RESOURCE_L2DATA,
134 		.name			= "L2DATA",
135 		.domains		= domain_init(RDT_RESOURCE_L2DATA),
136 		.msr_base		= MSR_IA32_L2_CBM_BASE,
137 		.msr_update		= cat_wrmsr,
138 		.cache_level		= 2,
139 		.cache = {
140 			.min_cbm_bits	= 1,
141 			.cbm_idx_mult	= 2,
142 			.cbm_idx_offset	= 0,
143 		},
144 		.parse_ctrlval		= parse_cbm,
145 		.format_str		= "%d=%0*x",
146 		.fflags			= RFTYPE_RES_CACHE,
147 	},
148 	[RDT_RESOURCE_L2CODE] =
149 	{
150 		.rid			= RDT_RESOURCE_L2CODE,
151 		.name			= "L2CODE",
152 		.domains		= domain_init(RDT_RESOURCE_L2CODE),
153 		.msr_base		= MSR_IA32_L2_CBM_BASE,
154 		.msr_update		= cat_wrmsr,
155 		.cache_level		= 2,
156 		.cache = {
157 			.min_cbm_bits	= 1,
158 			.cbm_idx_mult	= 2,
159 			.cbm_idx_offset	= 1,
160 		},
161 		.parse_ctrlval		= parse_cbm,
162 		.format_str		= "%d=%0*x",
163 		.fflags			= RFTYPE_RES_CACHE,
164 	},
165 	[RDT_RESOURCE_MBA] =
166 	{
167 		.rid			= RDT_RESOURCE_MBA,
168 		.name			= "MB",
169 		.domains		= domain_init(RDT_RESOURCE_MBA),
170 		.cache_level		= 3,
171 		.format_str		= "%d=%*u",
172 		.fflags			= RFTYPE_RES_MB,
173 	},
174 };
175 
176 static unsigned int cbm_idx(struct rdt_resource *r, unsigned int closid)
177 {
178 	return closid * r->cache.cbm_idx_mult + r->cache.cbm_idx_offset;
179 }
180 
181 /*
182  * cache_alloc_hsw_probe() - Have to probe for Intel haswell server CPUs
183  * as they do not have CPUID enumeration support for Cache allocation.
184  * The check for Vendor/Family/Model is not enough to guarantee that
185  * the MSRs won't #GP fault because only the following SKUs support
186  * CAT:
187  *	Intel(R) Xeon(R)  CPU E5-2658  v3  @  2.20GHz
188  *	Intel(R) Xeon(R)  CPU E5-2648L v3  @  1.80GHz
189  *	Intel(R) Xeon(R)  CPU E5-2628L v3  @  2.00GHz
190  *	Intel(R) Xeon(R)  CPU E5-2618L v3  @  2.30GHz
191  *	Intel(R) Xeon(R)  CPU E5-2608L v3  @  2.00GHz
192  *	Intel(R) Xeon(R)  CPU E5-2658A v3  @  2.20GHz
193  *
194  * Probe by trying to write the first of the L3 cach mask registers
195  * and checking that the bits stick. Max CLOSids is always 4 and max cbm length
196  * is always 20 on hsw server parts. The minimum cache bitmask length
197  * allowed for HSW server is always 2 bits. Hardcode all of them.
198  */
199 static inline void cache_alloc_hsw_probe(void)
200 {
201 	struct rdt_resource *r  = &rdt_resources_all[RDT_RESOURCE_L3];
202 	u32 l, h, max_cbm = BIT_MASK(20) - 1;
203 
204 	if (wrmsr_safe(MSR_IA32_L3_CBM_BASE, max_cbm, 0))
205 		return;
206 
207 	rdmsr(MSR_IA32_L3_CBM_BASE, l, h);
208 
209 	/* If all the bits were set in MSR, return success */
210 	if (l != max_cbm)
211 		return;
212 
213 	r->num_closid = 4;
214 	r->default_ctrl = max_cbm;
215 	r->cache.cbm_len = 20;
216 	r->cache.shareable_bits = 0xc0000;
217 	r->cache.min_cbm_bits = 2;
218 	r->alloc_capable = true;
219 	r->alloc_enabled = true;
220 
221 	rdt_alloc_capable = true;
222 }
223 
224 bool is_mba_sc(struct rdt_resource *r)
225 {
226 	if (!r)
227 		return rdt_resources_all[RDT_RESOURCE_MBA].membw.mba_sc;
228 
229 	return r->membw.mba_sc;
230 }
231 
232 /*
233  * rdt_get_mb_table() - get a mapping of bandwidth(b/w) percentage values
234  * exposed to user interface and the h/w understandable delay values.
235  *
236  * The non-linear delay values have the granularity of power of two
237  * and also the h/w does not guarantee a curve for configured delay
238  * values vs. actual b/w enforced.
239  * Hence we need a mapping that is pre calibrated so the user can
240  * express the memory b/w as a percentage value.
241  */
242 static inline bool rdt_get_mb_table(struct rdt_resource *r)
243 {
244 	/*
245 	 * There are no Intel SKUs as of now to support non-linear delay.
246 	 */
247 	pr_info("MBA b/w map not implemented for cpu:%d, model:%d",
248 		boot_cpu_data.x86, boot_cpu_data.x86_model);
249 
250 	return false;
251 }
252 
253 static bool __get_mem_config_intel(struct rdt_resource *r)
254 {
255 	union cpuid_0x10_3_eax eax;
256 	union cpuid_0x10_x_edx edx;
257 	u32 ebx, ecx;
258 
259 	cpuid_count(0x00000010, 3, &eax.full, &ebx, &ecx, &edx.full);
260 	r->num_closid = edx.split.cos_max + 1;
261 	r->membw.max_delay = eax.split.max_delay + 1;
262 	r->default_ctrl = MAX_MBA_BW;
263 	if (ecx & MBA_IS_LINEAR) {
264 		r->membw.delay_linear = true;
265 		r->membw.min_bw = MAX_MBA_BW - r->membw.max_delay;
266 		r->membw.bw_gran = MAX_MBA_BW - r->membw.max_delay;
267 	} else {
268 		if (!rdt_get_mb_table(r))
269 			return false;
270 	}
271 	r->data_width = 3;
272 
273 	r->alloc_capable = true;
274 	r->alloc_enabled = true;
275 
276 	return true;
277 }
278 
279 static bool __rdt_get_mem_config_amd(struct rdt_resource *r)
280 {
281 	union cpuid_0x10_3_eax eax;
282 	union cpuid_0x10_x_edx edx;
283 	u32 ebx, ecx;
284 
285 	cpuid_count(0x80000020, 1, &eax.full, &ebx, &ecx, &edx.full);
286 	r->num_closid = edx.split.cos_max + 1;
287 	r->default_ctrl = MAX_MBA_BW_AMD;
288 
289 	/* AMD does not use delay */
290 	r->membw.delay_linear = false;
291 
292 	r->membw.min_bw = 0;
293 	r->membw.bw_gran = 1;
294 	/* Max value is 2048, Data width should be 4 in decimal */
295 	r->data_width = 4;
296 
297 	r->alloc_capable = true;
298 	r->alloc_enabled = true;
299 
300 	return true;
301 }
302 
303 static void rdt_get_cache_alloc_cfg(int idx, struct rdt_resource *r)
304 {
305 	union cpuid_0x10_1_eax eax;
306 	union cpuid_0x10_x_edx edx;
307 	u32 ebx, ecx;
308 
309 	cpuid_count(0x00000010, idx, &eax.full, &ebx, &ecx, &edx.full);
310 	r->num_closid = edx.split.cos_max + 1;
311 	r->cache.cbm_len = eax.split.cbm_len + 1;
312 	r->default_ctrl = BIT_MASK(eax.split.cbm_len + 1) - 1;
313 	r->cache.shareable_bits = ebx & r->default_ctrl;
314 	r->data_width = (r->cache.cbm_len + 3) / 4;
315 	r->alloc_capable = true;
316 	r->alloc_enabled = true;
317 }
318 
319 static void rdt_get_cdp_config(int level, int type)
320 {
321 	struct rdt_resource *r_l = &rdt_resources_all[level];
322 	struct rdt_resource *r = &rdt_resources_all[type];
323 
324 	r->num_closid = r_l->num_closid / 2;
325 	r->cache.cbm_len = r_l->cache.cbm_len;
326 	r->default_ctrl = r_l->default_ctrl;
327 	r->cache.shareable_bits = r_l->cache.shareable_bits;
328 	r->data_width = (r->cache.cbm_len + 3) / 4;
329 	r->alloc_capable = true;
330 	/*
331 	 * By default, CDP is disabled. CDP can be enabled by mount parameter
332 	 * "cdp" during resctrl file system mount time.
333 	 */
334 	r->alloc_enabled = false;
335 }
336 
337 static void rdt_get_cdp_l3_config(void)
338 {
339 	rdt_get_cdp_config(RDT_RESOURCE_L3, RDT_RESOURCE_L3DATA);
340 	rdt_get_cdp_config(RDT_RESOURCE_L3, RDT_RESOURCE_L3CODE);
341 }
342 
343 static void rdt_get_cdp_l2_config(void)
344 {
345 	rdt_get_cdp_config(RDT_RESOURCE_L2, RDT_RESOURCE_L2DATA);
346 	rdt_get_cdp_config(RDT_RESOURCE_L2, RDT_RESOURCE_L2CODE);
347 }
348 
349 static int get_cache_id(int cpu, int level)
350 {
351 	struct cpu_cacheinfo *ci = get_cpu_cacheinfo(cpu);
352 	int i;
353 
354 	for (i = 0; i < ci->num_leaves; i++) {
355 		if (ci->info_list[i].level == level)
356 			return ci->info_list[i].id;
357 	}
358 
359 	return -1;
360 }
361 
362 static void
363 mba_wrmsr_amd(struct rdt_domain *d, struct msr_param *m, struct rdt_resource *r)
364 {
365 	unsigned int i;
366 
367 	for (i = m->low; i < m->high; i++)
368 		wrmsrl(r->msr_base + i, d->ctrl_val[i]);
369 }
370 
371 /*
372  * Map the memory b/w percentage value to delay values
373  * that can be written to QOS_MSRs.
374  * There are currently no SKUs which support non linear delay values.
375  */
376 u32 delay_bw_map(unsigned long bw, struct rdt_resource *r)
377 {
378 	if (r->membw.delay_linear)
379 		return MAX_MBA_BW - bw;
380 
381 	pr_warn_once("Non Linear delay-bw map not supported but queried\n");
382 	return r->default_ctrl;
383 }
384 
385 static void
386 mba_wrmsr_intel(struct rdt_domain *d, struct msr_param *m,
387 		struct rdt_resource *r)
388 {
389 	unsigned int i;
390 
391 	/*  Write the delay values for mba. */
392 	for (i = m->low; i < m->high; i++)
393 		wrmsrl(r->msr_base + i, delay_bw_map(d->ctrl_val[i], r));
394 }
395 
396 static void
397 cat_wrmsr(struct rdt_domain *d, struct msr_param *m, struct rdt_resource *r)
398 {
399 	unsigned int i;
400 
401 	for (i = m->low; i < m->high; i++)
402 		wrmsrl(r->msr_base + cbm_idx(r, i), d->ctrl_val[i]);
403 }
404 
405 struct rdt_domain *get_domain_from_cpu(int cpu, struct rdt_resource *r)
406 {
407 	struct rdt_domain *d;
408 
409 	list_for_each_entry(d, &r->domains, list) {
410 		/* Find the domain that contains this CPU */
411 		if (cpumask_test_cpu(cpu, &d->cpu_mask))
412 			return d;
413 	}
414 
415 	return NULL;
416 }
417 
418 void rdt_ctrl_update(void *arg)
419 {
420 	struct msr_param *m = arg;
421 	struct rdt_resource *r = m->res;
422 	int cpu = smp_processor_id();
423 	struct rdt_domain *d;
424 
425 	d = get_domain_from_cpu(cpu, r);
426 	if (d) {
427 		r->msr_update(d, m, r);
428 		return;
429 	}
430 	pr_warn_once("cpu %d not found in any domain for resource %s\n",
431 		     cpu, r->name);
432 }
433 
434 /*
435  * rdt_find_domain - Find a domain in a resource that matches input resource id
436  *
437  * Search resource r's domain list to find the resource id. If the resource
438  * id is found in a domain, return the domain. Otherwise, if requested by
439  * caller, return the first domain whose id is bigger than the input id.
440  * The domain list is sorted by id in ascending order.
441  */
442 struct rdt_domain *rdt_find_domain(struct rdt_resource *r, int id,
443 				   struct list_head **pos)
444 {
445 	struct rdt_domain *d;
446 	struct list_head *l;
447 
448 	if (id < 0)
449 		return ERR_PTR(-ENODEV);
450 
451 	list_for_each(l, &r->domains) {
452 		d = list_entry(l, struct rdt_domain, list);
453 		/* When id is found, return its domain. */
454 		if (id == d->id)
455 			return d;
456 		/* Stop searching when finding id's position in sorted list. */
457 		if (id < d->id)
458 			break;
459 	}
460 
461 	if (pos)
462 		*pos = l;
463 
464 	return NULL;
465 }
466 
467 void setup_default_ctrlval(struct rdt_resource *r, u32 *dc, u32 *dm)
468 {
469 	int i;
470 
471 	/*
472 	 * Initialize the Control MSRs to having no control.
473 	 * For Cache Allocation: Set all bits in cbm
474 	 * For Memory Allocation: Set b/w requested to 100%
475 	 * and the bandwidth in MBps to U32_MAX
476 	 */
477 	for (i = 0; i < r->num_closid; i++, dc++, dm++) {
478 		*dc = r->default_ctrl;
479 		*dm = MBA_MAX_MBPS;
480 	}
481 }
482 
483 static int domain_setup_ctrlval(struct rdt_resource *r, struct rdt_domain *d)
484 {
485 	struct msr_param m;
486 	u32 *dc, *dm;
487 
488 	dc = kmalloc_array(r->num_closid, sizeof(*d->ctrl_val), GFP_KERNEL);
489 	if (!dc)
490 		return -ENOMEM;
491 
492 	dm = kmalloc_array(r->num_closid, sizeof(*d->mbps_val), GFP_KERNEL);
493 	if (!dm) {
494 		kfree(dc);
495 		return -ENOMEM;
496 	}
497 
498 	d->ctrl_val = dc;
499 	d->mbps_val = dm;
500 	setup_default_ctrlval(r, dc, dm);
501 
502 	m.low = 0;
503 	m.high = r->num_closid;
504 	r->msr_update(d, &m, r);
505 	return 0;
506 }
507 
508 static int domain_setup_mon_state(struct rdt_resource *r, struct rdt_domain *d)
509 {
510 	size_t tsize;
511 
512 	if (is_llc_occupancy_enabled()) {
513 		d->rmid_busy_llc = bitmap_zalloc(r->num_rmid, GFP_KERNEL);
514 		if (!d->rmid_busy_llc)
515 			return -ENOMEM;
516 		INIT_DELAYED_WORK(&d->cqm_limbo, cqm_handle_limbo);
517 	}
518 	if (is_mbm_total_enabled()) {
519 		tsize = sizeof(*d->mbm_total);
520 		d->mbm_total = kcalloc(r->num_rmid, tsize, GFP_KERNEL);
521 		if (!d->mbm_total) {
522 			bitmap_free(d->rmid_busy_llc);
523 			return -ENOMEM;
524 		}
525 	}
526 	if (is_mbm_local_enabled()) {
527 		tsize = sizeof(*d->mbm_local);
528 		d->mbm_local = kcalloc(r->num_rmid, tsize, GFP_KERNEL);
529 		if (!d->mbm_local) {
530 			bitmap_free(d->rmid_busy_llc);
531 			kfree(d->mbm_total);
532 			return -ENOMEM;
533 		}
534 	}
535 
536 	if (is_mbm_enabled()) {
537 		INIT_DELAYED_WORK(&d->mbm_over, mbm_handle_overflow);
538 		mbm_setup_overflow_handler(d, MBM_OVERFLOW_INTERVAL);
539 	}
540 
541 	return 0;
542 }
543 
544 /*
545  * domain_add_cpu - Add a cpu to a resource's domain list.
546  *
547  * If an existing domain in the resource r's domain list matches the cpu's
548  * resource id, add the cpu in the domain.
549  *
550  * Otherwise, a new domain is allocated and inserted into the right position
551  * in the domain list sorted by id in ascending order.
552  *
553  * The order in the domain list is visible to users when we print entries
554  * in the schemata file and schemata input is validated to have the same order
555  * as this list.
556  */
557 static void domain_add_cpu(int cpu, struct rdt_resource *r)
558 {
559 	int id = get_cache_id(cpu, r->cache_level);
560 	struct list_head *add_pos = NULL;
561 	struct rdt_domain *d;
562 
563 	d = rdt_find_domain(r, id, &add_pos);
564 	if (IS_ERR(d)) {
565 		pr_warn("Could't find cache id for cpu %d\n", cpu);
566 		return;
567 	}
568 
569 	if (d) {
570 		cpumask_set_cpu(cpu, &d->cpu_mask);
571 		return;
572 	}
573 
574 	d = kzalloc_node(sizeof(*d), GFP_KERNEL, cpu_to_node(cpu));
575 	if (!d)
576 		return;
577 
578 	d->id = id;
579 	cpumask_set_cpu(cpu, &d->cpu_mask);
580 
581 	rdt_domain_reconfigure_cdp(r);
582 
583 	if (r->alloc_capable && domain_setup_ctrlval(r, d)) {
584 		kfree(d);
585 		return;
586 	}
587 
588 	if (r->mon_capable && domain_setup_mon_state(r, d)) {
589 		kfree(d);
590 		return;
591 	}
592 
593 	list_add_tail(&d->list, add_pos);
594 
595 	/*
596 	 * If resctrl is mounted, add
597 	 * per domain monitor data directories.
598 	 */
599 	if (static_branch_unlikely(&rdt_mon_enable_key))
600 		mkdir_mondata_subdir_allrdtgrp(r, d);
601 }
602 
603 static void domain_remove_cpu(int cpu, struct rdt_resource *r)
604 {
605 	int id = get_cache_id(cpu, r->cache_level);
606 	struct rdt_domain *d;
607 
608 	d = rdt_find_domain(r, id, NULL);
609 	if (IS_ERR_OR_NULL(d)) {
610 		pr_warn("Could't find cache id for cpu %d\n", cpu);
611 		return;
612 	}
613 
614 	cpumask_clear_cpu(cpu, &d->cpu_mask);
615 	if (cpumask_empty(&d->cpu_mask)) {
616 		/*
617 		 * If resctrl is mounted, remove all the
618 		 * per domain monitor data directories.
619 		 */
620 		if (static_branch_unlikely(&rdt_mon_enable_key))
621 			rmdir_mondata_subdir_allrdtgrp(r, d->id);
622 		list_del(&d->list);
623 		if (r->mon_capable && is_mbm_enabled())
624 			cancel_delayed_work(&d->mbm_over);
625 		if (is_llc_occupancy_enabled() &&  has_busy_rmid(r, d)) {
626 			/*
627 			 * When a package is going down, forcefully
628 			 * decrement rmid->ebusy. There is no way to know
629 			 * that the L3 was flushed and hence may lead to
630 			 * incorrect counts in rare scenarios, but leaving
631 			 * the RMID as busy creates RMID leaks if the
632 			 * package never comes back.
633 			 */
634 			__check_limbo(d, true);
635 			cancel_delayed_work(&d->cqm_limbo);
636 		}
637 
638 		/*
639 		 * rdt_domain "d" is going to be freed below, so clear
640 		 * its pointer from pseudo_lock_region struct.
641 		 */
642 		if (d->plr)
643 			d->plr->d = NULL;
644 
645 		kfree(d->ctrl_val);
646 		kfree(d->mbps_val);
647 		bitmap_free(d->rmid_busy_llc);
648 		kfree(d->mbm_total);
649 		kfree(d->mbm_local);
650 		kfree(d);
651 		return;
652 	}
653 
654 	if (r == &rdt_resources_all[RDT_RESOURCE_L3]) {
655 		if (is_mbm_enabled() && cpu == d->mbm_work_cpu) {
656 			cancel_delayed_work(&d->mbm_over);
657 			mbm_setup_overflow_handler(d, 0);
658 		}
659 		if (is_llc_occupancy_enabled() && cpu == d->cqm_work_cpu &&
660 		    has_busy_rmid(r, d)) {
661 			cancel_delayed_work(&d->cqm_limbo);
662 			cqm_setup_limbo_handler(d, 0);
663 		}
664 	}
665 }
666 
667 static void clear_closid_rmid(int cpu)
668 {
669 	struct resctrl_pqr_state *state = this_cpu_ptr(&pqr_state);
670 
671 	state->default_closid = 0;
672 	state->default_rmid = 0;
673 	state->cur_closid = 0;
674 	state->cur_rmid = 0;
675 	wrmsr(IA32_PQR_ASSOC, 0, 0);
676 }
677 
678 static int resctrl_online_cpu(unsigned int cpu)
679 {
680 	struct rdt_resource *r;
681 
682 	mutex_lock(&rdtgroup_mutex);
683 	for_each_capable_rdt_resource(r)
684 		domain_add_cpu(cpu, r);
685 	/* The cpu is set in default rdtgroup after online. */
686 	cpumask_set_cpu(cpu, &rdtgroup_default.cpu_mask);
687 	clear_closid_rmid(cpu);
688 	mutex_unlock(&rdtgroup_mutex);
689 
690 	return 0;
691 }
692 
693 static void clear_childcpus(struct rdtgroup *r, unsigned int cpu)
694 {
695 	struct rdtgroup *cr;
696 
697 	list_for_each_entry(cr, &r->mon.crdtgrp_list, mon.crdtgrp_list) {
698 		if (cpumask_test_and_clear_cpu(cpu, &cr->cpu_mask)) {
699 			break;
700 		}
701 	}
702 }
703 
704 static int resctrl_offline_cpu(unsigned int cpu)
705 {
706 	struct rdtgroup *rdtgrp;
707 	struct rdt_resource *r;
708 
709 	mutex_lock(&rdtgroup_mutex);
710 	for_each_capable_rdt_resource(r)
711 		domain_remove_cpu(cpu, r);
712 	list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
713 		if (cpumask_test_and_clear_cpu(cpu, &rdtgrp->cpu_mask)) {
714 			clear_childcpus(rdtgrp, cpu);
715 			break;
716 		}
717 	}
718 	clear_closid_rmid(cpu);
719 	mutex_unlock(&rdtgroup_mutex);
720 
721 	return 0;
722 }
723 
724 /*
725  * Choose a width for the resource name and resource data based on the
726  * resource that has widest name and cbm.
727  */
728 static __init void rdt_init_padding(void)
729 {
730 	struct rdt_resource *r;
731 	int cl;
732 
733 	for_each_alloc_capable_rdt_resource(r) {
734 		cl = strlen(r->name);
735 		if (cl > max_name_width)
736 			max_name_width = cl;
737 
738 		if (r->data_width > max_data_width)
739 			max_data_width = r->data_width;
740 	}
741 }
742 
743 enum {
744 	RDT_FLAG_CMT,
745 	RDT_FLAG_MBM_TOTAL,
746 	RDT_FLAG_MBM_LOCAL,
747 	RDT_FLAG_L3_CAT,
748 	RDT_FLAG_L3_CDP,
749 	RDT_FLAG_L2_CAT,
750 	RDT_FLAG_L2_CDP,
751 	RDT_FLAG_MBA,
752 };
753 
754 #define RDT_OPT(idx, n, f)	\
755 [idx] = {			\
756 	.name = n,		\
757 	.flag = f		\
758 }
759 
760 struct rdt_options {
761 	char	*name;
762 	int	flag;
763 	bool	force_off, force_on;
764 };
765 
766 static struct rdt_options rdt_options[]  __initdata = {
767 	RDT_OPT(RDT_FLAG_CMT,	    "cmt",	X86_FEATURE_CQM_OCCUP_LLC),
768 	RDT_OPT(RDT_FLAG_MBM_TOTAL, "mbmtotal", X86_FEATURE_CQM_MBM_TOTAL),
769 	RDT_OPT(RDT_FLAG_MBM_LOCAL, "mbmlocal", X86_FEATURE_CQM_MBM_LOCAL),
770 	RDT_OPT(RDT_FLAG_L3_CAT,    "l3cat",	X86_FEATURE_CAT_L3),
771 	RDT_OPT(RDT_FLAG_L3_CDP,    "l3cdp",	X86_FEATURE_CDP_L3),
772 	RDT_OPT(RDT_FLAG_L2_CAT,    "l2cat",	X86_FEATURE_CAT_L2),
773 	RDT_OPT(RDT_FLAG_L2_CDP,    "l2cdp",	X86_FEATURE_CDP_L2),
774 	RDT_OPT(RDT_FLAG_MBA,	    "mba",	X86_FEATURE_MBA),
775 };
776 #define NUM_RDT_OPTIONS ARRAY_SIZE(rdt_options)
777 
778 static int __init set_rdt_options(char *str)
779 {
780 	struct rdt_options *o;
781 	bool force_off;
782 	char *tok;
783 
784 	if (*str == '=')
785 		str++;
786 	while ((tok = strsep(&str, ",")) != NULL) {
787 		force_off = *tok == '!';
788 		if (force_off)
789 			tok++;
790 		for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
791 			if (strcmp(tok, o->name) == 0) {
792 				if (force_off)
793 					o->force_off = true;
794 				else
795 					o->force_on = true;
796 				break;
797 			}
798 		}
799 	}
800 	return 1;
801 }
802 __setup("rdt", set_rdt_options);
803 
804 static bool __init rdt_cpu_has(int flag)
805 {
806 	bool ret = boot_cpu_has(flag);
807 	struct rdt_options *o;
808 
809 	if (!ret)
810 		return ret;
811 
812 	for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
813 		if (flag == o->flag) {
814 			if (o->force_off)
815 				ret = false;
816 			if (o->force_on)
817 				ret = true;
818 			break;
819 		}
820 	}
821 	return ret;
822 }
823 
824 static __init bool get_mem_config(void)
825 {
826 	if (!rdt_cpu_has(X86_FEATURE_MBA))
827 		return false;
828 
829 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
830 		return __get_mem_config_intel(&rdt_resources_all[RDT_RESOURCE_MBA]);
831 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
832 		return __rdt_get_mem_config_amd(&rdt_resources_all[RDT_RESOURCE_MBA]);
833 
834 	return false;
835 }
836 
837 static __init bool get_rdt_alloc_resources(void)
838 {
839 	bool ret = false;
840 
841 	if (rdt_alloc_capable)
842 		return true;
843 
844 	if (!boot_cpu_has(X86_FEATURE_RDT_A))
845 		return false;
846 
847 	if (rdt_cpu_has(X86_FEATURE_CAT_L3)) {
848 		rdt_get_cache_alloc_cfg(1, &rdt_resources_all[RDT_RESOURCE_L3]);
849 		if (rdt_cpu_has(X86_FEATURE_CDP_L3))
850 			rdt_get_cdp_l3_config();
851 		ret = true;
852 	}
853 	if (rdt_cpu_has(X86_FEATURE_CAT_L2)) {
854 		/* CPUID 0x10.2 fields are same format at 0x10.1 */
855 		rdt_get_cache_alloc_cfg(2, &rdt_resources_all[RDT_RESOURCE_L2]);
856 		if (rdt_cpu_has(X86_FEATURE_CDP_L2))
857 			rdt_get_cdp_l2_config();
858 		ret = true;
859 	}
860 
861 	if (get_mem_config())
862 		ret = true;
863 
864 	return ret;
865 }
866 
867 static __init bool get_rdt_mon_resources(void)
868 {
869 	if (rdt_cpu_has(X86_FEATURE_CQM_OCCUP_LLC))
870 		rdt_mon_features |= (1 << QOS_L3_OCCUP_EVENT_ID);
871 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL))
872 		rdt_mon_features |= (1 << QOS_L3_MBM_TOTAL_EVENT_ID);
873 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL))
874 		rdt_mon_features |= (1 << QOS_L3_MBM_LOCAL_EVENT_ID);
875 
876 	if (!rdt_mon_features)
877 		return false;
878 
879 	return !rdt_get_mon_l3_config(&rdt_resources_all[RDT_RESOURCE_L3]);
880 }
881 
882 static __init void __check_quirks_intel(void)
883 {
884 	switch (boot_cpu_data.x86_model) {
885 	case INTEL_FAM6_HASWELL_X:
886 		if (!rdt_options[RDT_FLAG_L3_CAT].force_off)
887 			cache_alloc_hsw_probe();
888 		break;
889 	case INTEL_FAM6_SKYLAKE_X:
890 		if (boot_cpu_data.x86_stepping <= 4)
891 			set_rdt_options("!cmt,!mbmtotal,!mbmlocal,!l3cat");
892 		else
893 			set_rdt_options("!l3cat");
894 	}
895 }
896 
897 static __init void check_quirks(void)
898 {
899 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
900 		__check_quirks_intel();
901 }
902 
903 static __init bool get_rdt_resources(void)
904 {
905 	rdt_alloc_capable = get_rdt_alloc_resources();
906 	rdt_mon_capable = get_rdt_mon_resources();
907 
908 	return (rdt_mon_capable || rdt_alloc_capable);
909 }
910 
911 static __init void rdt_init_res_defs_intel(void)
912 {
913 	struct rdt_resource *r;
914 
915 	for_each_rdt_resource(r) {
916 		if (r->rid == RDT_RESOURCE_L3 ||
917 		    r->rid == RDT_RESOURCE_L3DATA ||
918 		    r->rid == RDT_RESOURCE_L3CODE ||
919 		    r->rid == RDT_RESOURCE_L2 ||
920 		    r->rid == RDT_RESOURCE_L2DATA ||
921 		    r->rid == RDT_RESOURCE_L2CODE)
922 			r->cbm_validate = cbm_validate_intel;
923 		else if (r->rid == RDT_RESOURCE_MBA) {
924 			r->msr_base = MSR_IA32_MBA_THRTL_BASE;
925 			r->msr_update = mba_wrmsr_intel;
926 			r->parse_ctrlval = parse_bw_intel;
927 		}
928 	}
929 }
930 
931 static __init void rdt_init_res_defs_amd(void)
932 {
933 	struct rdt_resource *r;
934 
935 	for_each_rdt_resource(r) {
936 		if (r->rid == RDT_RESOURCE_L3 ||
937 		    r->rid == RDT_RESOURCE_L3DATA ||
938 		    r->rid == RDT_RESOURCE_L3CODE ||
939 		    r->rid == RDT_RESOURCE_L2 ||
940 		    r->rid == RDT_RESOURCE_L2DATA ||
941 		    r->rid == RDT_RESOURCE_L2CODE)
942 			r->cbm_validate = cbm_validate_amd;
943 		else if (r->rid == RDT_RESOURCE_MBA) {
944 			r->msr_base = MSR_IA32_MBA_BW_BASE;
945 			r->msr_update = mba_wrmsr_amd;
946 			r->parse_ctrlval = parse_bw_amd;
947 		}
948 	}
949 }
950 
951 static __init void rdt_init_res_defs(void)
952 {
953 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
954 		rdt_init_res_defs_intel();
955 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
956 		rdt_init_res_defs_amd();
957 }
958 
959 static enum cpuhp_state rdt_online;
960 
961 /* Runs once on the BSP during boot. */
962 void resctrl_cpu_detect(struct cpuinfo_x86 *c)
963 {
964 	if (!cpu_has(c, X86_FEATURE_CQM_LLC)) {
965 		c->x86_cache_max_rmid  = -1;
966 		c->x86_cache_occ_scale = -1;
967 		c->x86_cache_mbm_width_offset = -1;
968 		return;
969 	}
970 
971 	/* will be overridden if occupancy monitoring exists */
972 	c->x86_cache_max_rmid = cpuid_ebx(0xf);
973 
974 	if (cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC) ||
975 	    cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL) ||
976 	    cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)) {
977 		u32 eax, ebx, ecx, edx;
978 
979 		/* QoS sub-leaf, EAX=0Fh, ECX=1 */
980 		cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx);
981 
982 		c->x86_cache_max_rmid  = ecx;
983 		c->x86_cache_occ_scale = ebx;
984 		if (c->x86_vendor == X86_VENDOR_INTEL)
985 			c->x86_cache_mbm_width_offset = eax & 0xff;
986 		else
987 			c->x86_cache_mbm_width_offset = -1;
988 	}
989 }
990 
991 static int __init resctrl_late_init(void)
992 {
993 	struct rdt_resource *r;
994 	int state, ret;
995 
996 	/*
997 	 * Initialize functions(or definitions) that are different
998 	 * between vendors here.
999 	 */
1000 	rdt_init_res_defs();
1001 
1002 	check_quirks();
1003 
1004 	if (!get_rdt_resources())
1005 		return -ENODEV;
1006 
1007 	rdt_init_padding();
1008 
1009 	state = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
1010 				  "x86/resctrl/cat:online:",
1011 				  resctrl_online_cpu, resctrl_offline_cpu);
1012 	if (state < 0)
1013 		return state;
1014 
1015 	ret = rdtgroup_init();
1016 	if (ret) {
1017 		cpuhp_remove_state(state);
1018 		return ret;
1019 	}
1020 	rdt_online = state;
1021 
1022 	for_each_alloc_capable_rdt_resource(r)
1023 		pr_info("%s allocation detected\n", r->name);
1024 
1025 	for_each_mon_capable_rdt_resource(r)
1026 		pr_info("%s monitoring detected\n", r->name);
1027 
1028 	return 0;
1029 }
1030 
1031 late_initcall(resctrl_late_init);
1032 
1033 static void __exit resctrl_exit(void)
1034 {
1035 	cpuhp_remove_state(rdt_online);
1036 	rdtgroup_exit();
1037 }
1038 
1039 __exitcall(resctrl_exit);
1040