xref: /openbmc/linux/arch/x86/kernel/cpu/resctrl/core.c (revision 49c23519)
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].r_resctrl.domains)
61 
62 struct rdt_hw_resource rdt_resources_all[] = {
63 	[RDT_RESOURCE_L3] =
64 	{
65 		.r_resctrl = {
66 			.rid			= RDT_RESOURCE_L3,
67 			.name			= "L3",
68 			.cache_level		= 3,
69 			.domains		= domain_init(RDT_RESOURCE_L3),
70 			.parse_ctrlval		= parse_cbm,
71 			.format_str		= "%d=%0*x",
72 			.fflags			= RFTYPE_RES_CACHE,
73 		},
74 		.msr_base		= MSR_IA32_L3_CBM_BASE,
75 		.msr_update		= cat_wrmsr,
76 	},
77 	[RDT_RESOURCE_L2] =
78 	{
79 		.r_resctrl = {
80 			.rid			= RDT_RESOURCE_L2,
81 			.name			= "L2",
82 			.cache_level		= 2,
83 			.domains		= domain_init(RDT_RESOURCE_L2),
84 			.parse_ctrlval		= parse_cbm,
85 			.format_str		= "%d=%0*x",
86 			.fflags			= RFTYPE_RES_CACHE,
87 		},
88 		.msr_base		= MSR_IA32_L2_CBM_BASE,
89 		.msr_update		= cat_wrmsr,
90 	},
91 	[RDT_RESOURCE_MBA] =
92 	{
93 		.r_resctrl = {
94 			.rid			= RDT_RESOURCE_MBA,
95 			.name			= "MB",
96 			.cache_level		= 3,
97 			.domains		= domain_init(RDT_RESOURCE_MBA),
98 			.parse_ctrlval		= parse_bw,
99 			.format_str		= "%d=%*u",
100 			.fflags			= RFTYPE_RES_MB,
101 		},
102 	},
103 	[RDT_RESOURCE_SMBA] =
104 	{
105 		.r_resctrl = {
106 			.rid			= RDT_RESOURCE_SMBA,
107 			.name			= "SMBA",
108 			.cache_level		= 3,
109 			.domains		= domain_init(RDT_RESOURCE_SMBA),
110 			.parse_ctrlval		= parse_bw,
111 			.format_str		= "%d=%*u",
112 			.fflags			= RFTYPE_RES_MB,
113 		},
114 	},
115 };
116 
117 /*
118  * cache_alloc_hsw_probe() - Have to probe for Intel haswell server CPUs
119  * as they do not have CPUID enumeration support for Cache allocation.
120  * The check for Vendor/Family/Model is not enough to guarantee that
121  * the MSRs won't #GP fault because only the following SKUs support
122  * CAT:
123  *	Intel(R) Xeon(R)  CPU E5-2658  v3  @  2.20GHz
124  *	Intel(R) Xeon(R)  CPU E5-2648L v3  @  1.80GHz
125  *	Intel(R) Xeon(R)  CPU E5-2628L v3  @  2.00GHz
126  *	Intel(R) Xeon(R)  CPU E5-2618L v3  @  2.30GHz
127  *	Intel(R) Xeon(R)  CPU E5-2608L v3  @  2.00GHz
128  *	Intel(R) Xeon(R)  CPU E5-2658A v3  @  2.20GHz
129  *
130  * Probe by trying to write the first of the L3 cache mask registers
131  * and checking that the bits stick. Max CLOSids is always 4 and max cbm length
132  * is always 20 on hsw server parts. The minimum cache bitmask length
133  * allowed for HSW server is always 2 bits. Hardcode all of them.
134  */
135 static inline void cache_alloc_hsw_probe(void)
136 {
137 	struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_L3];
138 	struct rdt_resource *r  = &hw_res->r_resctrl;
139 	u32 l, h, max_cbm = BIT_MASK(20) - 1;
140 
141 	if (wrmsr_safe(MSR_IA32_L3_CBM_BASE, max_cbm, 0))
142 		return;
143 
144 	rdmsr(MSR_IA32_L3_CBM_BASE, l, h);
145 
146 	/* If all the bits were set in MSR, return success */
147 	if (l != max_cbm)
148 		return;
149 
150 	hw_res->num_closid = 4;
151 	r->default_ctrl = max_cbm;
152 	r->cache.cbm_len = 20;
153 	r->cache.shareable_bits = 0xc0000;
154 	r->cache.min_cbm_bits = 2;
155 	r->alloc_capable = true;
156 
157 	rdt_alloc_capable = true;
158 }
159 
160 bool is_mba_sc(struct rdt_resource *r)
161 {
162 	if (!r)
163 		return rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl.membw.mba_sc;
164 
165 	/*
166 	 * The software controller support is only applicable to MBA resource.
167 	 * Make sure to check for resource type.
168 	 */
169 	if (r->rid != RDT_RESOURCE_MBA)
170 		return false;
171 
172 	return r->membw.mba_sc;
173 }
174 
175 /*
176  * rdt_get_mb_table() - get a mapping of bandwidth(b/w) percentage values
177  * exposed to user interface and the h/w understandable delay values.
178  *
179  * The non-linear delay values have the granularity of power of two
180  * and also the h/w does not guarantee a curve for configured delay
181  * values vs. actual b/w enforced.
182  * Hence we need a mapping that is pre calibrated so the user can
183  * express the memory b/w as a percentage value.
184  */
185 static inline bool rdt_get_mb_table(struct rdt_resource *r)
186 {
187 	/*
188 	 * There are no Intel SKUs as of now to support non-linear delay.
189 	 */
190 	pr_info("MBA b/w map not implemented for cpu:%d, model:%d",
191 		boot_cpu_data.x86, boot_cpu_data.x86_model);
192 
193 	return false;
194 }
195 
196 static bool __get_mem_config_intel(struct rdt_resource *r)
197 {
198 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
199 	union cpuid_0x10_3_eax eax;
200 	union cpuid_0x10_x_edx edx;
201 	u32 ebx, ecx, max_delay;
202 
203 	cpuid_count(0x00000010, 3, &eax.full, &ebx, &ecx, &edx.full);
204 	hw_res->num_closid = edx.split.cos_max + 1;
205 	max_delay = eax.split.max_delay + 1;
206 	r->default_ctrl = MAX_MBA_BW;
207 	r->membw.arch_needs_linear = true;
208 	if (ecx & MBA_IS_LINEAR) {
209 		r->membw.delay_linear = true;
210 		r->membw.min_bw = MAX_MBA_BW - max_delay;
211 		r->membw.bw_gran = MAX_MBA_BW - max_delay;
212 	} else {
213 		if (!rdt_get_mb_table(r))
214 			return false;
215 		r->membw.arch_needs_linear = false;
216 	}
217 	r->data_width = 3;
218 
219 	if (boot_cpu_has(X86_FEATURE_PER_THREAD_MBA))
220 		r->membw.throttle_mode = THREAD_THROTTLE_PER_THREAD;
221 	else
222 		r->membw.throttle_mode = THREAD_THROTTLE_MAX;
223 	thread_throttle_mode_init();
224 
225 	r->alloc_capable = true;
226 
227 	return true;
228 }
229 
230 static bool __rdt_get_mem_config_amd(struct rdt_resource *r)
231 {
232 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
233 	u32 eax, ebx, ecx, edx, subleaf;
234 
235 	/*
236 	 * Query CPUID_Fn80000020_EDX_x01 for MBA and
237 	 * CPUID_Fn80000020_EDX_x02 for SMBA
238 	 */
239 	subleaf = (r->rid == RDT_RESOURCE_SMBA) ? 2 :  1;
240 
241 	cpuid_count(0x80000020, subleaf, &eax, &ebx, &ecx, &edx);
242 	hw_res->num_closid = edx + 1;
243 	r->default_ctrl = 1 << eax;
244 
245 	/* AMD does not use delay */
246 	r->membw.delay_linear = false;
247 	r->membw.arch_needs_linear = false;
248 
249 	/*
250 	 * AMD does not use memory delay throttle model to control
251 	 * the allocation like Intel does.
252 	 */
253 	r->membw.throttle_mode = THREAD_THROTTLE_UNDEFINED;
254 	r->membw.min_bw = 0;
255 	r->membw.bw_gran = 1;
256 	/* Max value is 2048, Data width should be 4 in decimal */
257 	r->data_width = 4;
258 
259 	r->alloc_capable = true;
260 
261 	return true;
262 }
263 
264 static void rdt_get_cache_alloc_cfg(int idx, struct rdt_resource *r)
265 {
266 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
267 	union cpuid_0x10_1_eax eax;
268 	union cpuid_0x10_x_edx edx;
269 	u32 ebx, ecx;
270 
271 	cpuid_count(0x00000010, idx, &eax.full, &ebx, &ecx, &edx.full);
272 	hw_res->num_closid = edx.split.cos_max + 1;
273 	r->cache.cbm_len = eax.split.cbm_len + 1;
274 	r->default_ctrl = BIT_MASK(eax.split.cbm_len + 1) - 1;
275 	r->cache.shareable_bits = ebx & r->default_ctrl;
276 	r->data_width = (r->cache.cbm_len + 3) / 4;
277 	r->alloc_capable = true;
278 }
279 
280 static void rdt_get_cdp_config(int level)
281 {
282 	/*
283 	 * By default, CDP is disabled. CDP can be enabled by mount parameter
284 	 * "cdp" during resctrl file system mount time.
285 	 */
286 	rdt_resources_all[level].cdp_enabled = false;
287 	rdt_resources_all[level].r_resctrl.cdp_capable = true;
288 }
289 
290 static void rdt_get_cdp_l3_config(void)
291 {
292 	rdt_get_cdp_config(RDT_RESOURCE_L3);
293 }
294 
295 static void rdt_get_cdp_l2_config(void)
296 {
297 	rdt_get_cdp_config(RDT_RESOURCE_L2);
298 }
299 
300 static void
301 mba_wrmsr_amd(struct rdt_domain *d, struct msr_param *m, struct rdt_resource *r)
302 {
303 	unsigned int i;
304 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
305 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
306 
307 	for (i = m->low; i < m->high; i++)
308 		wrmsrl(hw_res->msr_base + i, hw_dom->ctrl_val[i]);
309 }
310 
311 /*
312  * Map the memory b/w percentage value to delay values
313  * that can be written to QOS_MSRs.
314  * There are currently no SKUs which support non linear delay values.
315  */
316 static u32 delay_bw_map(unsigned long bw, struct rdt_resource *r)
317 {
318 	if (r->membw.delay_linear)
319 		return MAX_MBA_BW - bw;
320 
321 	pr_warn_once("Non Linear delay-bw map not supported but queried\n");
322 	return r->default_ctrl;
323 }
324 
325 static void
326 mba_wrmsr_intel(struct rdt_domain *d, struct msr_param *m,
327 		struct rdt_resource *r)
328 {
329 	unsigned int i;
330 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
331 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
332 
333 	/*  Write the delay values for mba. */
334 	for (i = m->low; i < m->high; i++)
335 		wrmsrl(hw_res->msr_base + i, delay_bw_map(hw_dom->ctrl_val[i], r));
336 }
337 
338 static void
339 cat_wrmsr(struct rdt_domain *d, struct msr_param *m, struct rdt_resource *r)
340 {
341 	unsigned int i;
342 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
343 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
344 
345 	for (i = m->low; i < m->high; i++)
346 		wrmsrl(hw_res->msr_base + i, hw_dom->ctrl_val[i]);
347 }
348 
349 struct rdt_domain *get_domain_from_cpu(int cpu, struct rdt_resource *r)
350 {
351 	struct rdt_domain *d;
352 
353 	list_for_each_entry(d, &r->domains, list) {
354 		/* Find the domain that contains this CPU */
355 		if (cpumask_test_cpu(cpu, &d->cpu_mask))
356 			return d;
357 	}
358 
359 	return NULL;
360 }
361 
362 u32 resctrl_arch_get_num_closid(struct rdt_resource *r)
363 {
364 	return resctrl_to_arch_res(r)->num_closid;
365 }
366 
367 void rdt_ctrl_update(void *arg)
368 {
369 	struct msr_param *m = arg;
370 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(m->res);
371 	struct rdt_resource *r = m->res;
372 	int cpu = smp_processor_id();
373 	struct rdt_domain *d;
374 
375 	d = get_domain_from_cpu(cpu, r);
376 	if (d) {
377 		hw_res->msr_update(d, m, r);
378 		return;
379 	}
380 	pr_warn_once("cpu %d not found in any domain for resource %s\n",
381 		     cpu, r->name);
382 }
383 
384 /*
385  * rdt_find_domain - Find a domain in a resource that matches input resource id
386  *
387  * Search resource r's domain list to find the resource id. If the resource
388  * id is found in a domain, return the domain. Otherwise, if requested by
389  * caller, return the first domain whose id is bigger than the input id.
390  * The domain list is sorted by id in ascending order.
391  */
392 struct rdt_domain *rdt_find_domain(struct rdt_resource *r, int id,
393 				   struct list_head **pos)
394 {
395 	struct rdt_domain *d;
396 	struct list_head *l;
397 
398 	if (id < 0)
399 		return ERR_PTR(-ENODEV);
400 
401 	list_for_each(l, &r->domains) {
402 		d = list_entry(l, struct rdt_domain, list);
403 		/* When id is found, return its domain. */
404 		if (id == d->id)
405 			return d;
406 		/* Stop searching when finding id's position in sorted list. */
407 		if (id < d->id)
408 			break;
409 	}
410 
411 	if (pos)
412 		*pos = l;
413 
414 	return NULL;
415 }
416 
417 static void setup_default_ctrlval(struct rdt_resource *r, u32 *dc)
418 {
419 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
420 	int i;
421 
422 	/*
423 	 * Initialize the Control MSRs to having no control.
424 	 * For Cache Allocation: Set all bits in cbm
425 	 * For Memory Allocation: Set b/w requested to 100%
426 	 */
427 	for (i = 0; i < hw_res->num_closid; i++, dc++)
428 		*dc = r->default_ctrl;
429 }
430 
431 static void domain_free(struct rdt_hw_domain *hw_dom)
432 {
433 	kfree(hw_dom->arch_mbm_total);
434 	kfree(hw_dom->arch_mbm_local);
435 	kfree(hw_dom->ctrl_val);
436 	kfree(hw_dom);
437 }
438 
439 static int domain_setup_ctrlval(struct rdt_resource *r, struct rdt_domain *d)
440 {
441 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
442 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
443 	struct msr_param m;
444 	u32 *dc;
445 
446 	dc = kmalloc_array(hw_res->num_closid, sizeof(*hw_dom->ctrl_val),
447 			   GFP_KERNEL);
448 	if (!dc)
449 		return -ENOMEM;
450 
451 	hw_dom->ctrl_val = dc;
452 	setup_default_ctrlval(r, dc);
453 
454 	m.low = 0;
455 	m.high = hw_res->num_closid;
456 	hw_res->msr_update(d, &m, r);
457 	return 0;
458 }
459 
460 /**
461  * arch_domain_mbm_alloc() - Allocate arch private storage for the MBM counters
462  * @num_rmid:	The size of the MBM counter array
463  * @hw_dom:	The domain that owns the allocated arrays
464  */
465 static int arch_domain_mbm_alloc(u32 num_rmid, struct rdt_hw_domain *hw_dom)
466 {
467 	size_t tsize;
468 
469 	if (is_mbm_total_enabled()) {
470 		tsize = sizeof(*hw_dom->arch_mbm_total);
471 		hw_dom->arch_mbm_total = kcalloc(num_rmid, tsize, GFP_KERNEL);
472 		if (!hw_dom->arch_mbm_total)
473 			return -ENOMEM;
474 	}
475 	if (is_mbm_local_enabled()) {
476 		tsize = sizeof(*hw_dom->arch_mbm_local);
477 		hw_dom->arch_mbm_local = kcalloc(num_rmid, tsize, GFP_KERNEL);
478 		if (!hw_dom->arch_mbm_local) {
479 			kfree(hw_dom->arch_mbm_total);
480 			hw_dom->arch_mbm_total = NULL;
481 			return -ENOMEM;
482 		}
483 	}
484 
485 	return 0;
486 }
487 
488 /*
489  * domain_add_cpu - Add a cpu to a resource's domain list.
490  *
491  * If an existing domain in the resource r's domain list matches the cpu's
492  * resource id, add the cpu in the domain.
493  *
494  * Otherwise, a new domain is allocated and inserted into the right position
495  * in the domain list sorted by id in ascending order.
496  *
497  * The order in the domain list is visible to users when we print entries
498  * in the schemata file and schemata input is validated to have the same order
499  * as this list.
500  */
501 static void domain_add_cpu(int cpu, struct rdt_resource *r)
502 {
503 	int id = get_cpu_cacheinfo_id(cpu, r->cache_level);
504 	struct list_head *add_pos = NULL;
505 	struct rdt_hw_domain *hw_dom;
506 	struct rdt_domain *d;
507 	int err;
508 
509 	d = rdt_find_domain(r, id, &add_pos);
510 	if (IS_ERR(d)) {
511 		pr_warn("Couldn't find cache id for CPU %d\n", cpu);
512 		return;
513 	}
514 
515 	if (d) {
516 		cpumask_set_cpu(cpu, &d->cpu_mask);
517 		if (r->cache.arch_has_per_cpu_cfg)
518 			rdt_domain_reconfigure_cdp(r);
519 		return;
520 	}
521 
522 	hw_dom = kzalloc_node(sizeof(*hw_dom), GFP_KERNEL, cpu_to_node(cpu));
523 	if (!hw_dom)
524 		return;
525 
526 	d = &hw_dom->d_resctrl;
527 	d->id = id;
528 	cpumask_set_cpu(cpu, &d->cpu_mask);
529 
530 	rdt_domain_reconfigure_cdp(r);
531 
532 	if (r->alloc_capable && domain_setup_ctrlval(r, d)) {
533 		domain_free(hw_dom);
534 		return;
535 	}
536 
537 	if (r->mon_capable && arch_domain_mbm_alloc(r->num_rmid, hw_dom)) {
538 		domain_free(hw_dom);
539 		return;
540 	}
541 
542 	list_add_tail(&d->list, add_pos);
543 
544 	err = resctrl_online_domain(r, d);
545 	if (err) {
546 		list_del(&d->list);
547 		domain_free(hw_dom);
548 	}
549 }
550 
551 static void domain_remove_cpu(int cpu, struct rdt_resource *r)
552 {
553 	int id = get_cpu_cacheinfo_id(cpu, r->cache_level);
554 	struct rdt_hw_domain *hw_dom;
555 	struct rdt_domain *d;
556 
557 	d = rdt_find_domain(r, id, NULL);
558 	if (IS_ERR_OR_NULL(d)) {
559 		pr_warn("Couldn't find cache id for CPU %d\n", cpu);
560 		return;
561 	}
562 	hw_dom = resctrl_to_arch_dom(d);
563 
564 	cpumask_clear_cpu(cpu, &d->cpu_mask);
565 	if (cpumask_empty(&d->cpu_mask)) {
566 		resctrl_offline_domain(r, d);
567 		list_del(&d->list);
568 
569 		/*
570 		 * rdt_domain "d" is going to be freed below, so clear
571 		 * its pointer from pseudo_lock_region struct.
572 		 */
573 		if (d->plr)
574 			d->plr->d = NULL;
575 		domain_free(hw_dom);
576 
577 		return;
578 	}
579 
580 	if (r == &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl) {
581 		if (is_mbm_enabled() && cpu == d->mbm_work_cpu) {
582 			cancel_delayed_work(&d->mbm_over);
583 			mbm_setup_overflow_handler(d, 0);
584 		}
585 		if (is_llc_occupancy_enabled() && cpu == d->cqm_work_cpu &&
586 		    has_busy_rmid(r, d)) {
587 			cancel_delayed_work(&d->cqm_limbo);
588 			cqm_setup_limbo_handler(d, 0);
589 		}
590 	}
591 }
592 
593 static void clear_closid_rmid(int cpu)
594 {
595 	struct resctrl_pqr_state *state = this_cpu_ptr(&pqr_state);
596 
597 	state->default_closid = 0;
598 	state->default_rmid = 0;
599 	state->cur_closid = 0;
600 	state->cur_rmid = 0;
601 	wrmsr(MSR_IA32_PQR_ASSOC, 0, 0);
602 }
603 
604 static int resctrl_online_cpu(unsigned int cpu)
605 {
606 	struct rdt_resource *r;
607 
608 	mutex_lock(&rdtgroup_mutex);
609 	for_each_capable_rdt_resource(r)
610 		domain_add_cpu(cpu, r);
611 	/* The cpu is set in default rdtgroup after online. */
612 	cpumask_set_cpu(cpu, &rdtgroup_default.cpu_mask);
613 	clear_closid_rmid(cpu);
614 	mutex_unlock(&rdtgroup_mutex);
615 
616 	return 0;
617 }
618 
619 static void clear_childcpus(struct rdtgroup *r, unsigned int cpu)
620 {
621 	struct rdtgroup *cr;
622 
623 	list_for_each_entry(cr, &r->mon.crdtgrp_list, mon.crdtgrp_list) {
624 		if (cpumask_test_and_clear_cpu(cpu, &cr->cpu_mask)) {
625 			break;
626 		}
627 	}
628 }
629 
630 static int resctrl_offline_cpu(unsigned int cpu)
631 {
632 	struct rdtgroup *rdtgrp;
633 	struct rdt_resource *r;
634 
635 	mutex_lock(&rdtgroup_mutex);
636 	for_each_capable_rdt_resource(r)
637 		domain_remove_cpu(cpu, r);
638 	list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
639 		if (cpumask_test_and_clear_cpu(cpu, &rdtgrp->cpu_mask)) {
640 			clear_childcpus(rdtgrp, cpu);
641 			break;
642 		}
643 	}
644 	clear_closid_rmid(cpu);
645 	mutex_unlock(&rdtgroup_mutex);
646 
647 	return 0;
648 }
649 
650 /*
651  * Choose a width for the resource name and resource data based on the
652  * resource that has widest name and cbm.
653  */
654 static __init void rdt_init_padding(void)
655 {
656 	struct rdt_resource *r;
657 
658 	for_each_alloc_capable_rdt_resource(r) {
659 		if (r->data_width > max_data_width)
660 			max_data_width = r->data_width;
661 	}
662 }
663 
664 enum {
665 	RDT_FLAG_CMT,
666 	RDT_FLAG_MBM_TOTAL,
667 	RDT_FLAG_MBM_LOCAL,
668 	RDT_FLAG_L3_CAT,
669 	RDT_FLAG_L3_CDP,
670 	RDT_FLAG_L2_CAT,
671 	RDT_FLAG_L2_CDP,
672 	RDT_FLAG_MBA,
673 	RDT_FLAG_SMBA,
674 	RDT_FLAG_BMEC,
675 };
676 
677 #define RDT_OPT(idx, n, f)	\
678 [idx] = {			\
679 	.name = n,		\
680 	.flag = f		\
681 }
682 
683 struct rdt_options {
684 	char	*name;
685 	int	flag;
686 	bool	force_off, force_on;
687 };
688 
689 static struct rdt_options rdt_options[]  __initdata = {
690 	RDT_OPT(RDT_FLAG_CMT,	    "cmt",	X86_FEATURE_CQM_OCCUP_LLC),
691 	RDT_OPT(RDT_FLAG_MBM_TOTAL, "mbmtotal", X86_FEATURE_CQM_MBM_TOTAL),
692 	RDT_OPT(RDT_FLAG_MBM_LOCAL, "mbmlocal", X86_FEATURE_CQM_MBM_LOCAL),
693 	RDT_OPT(RDT_FLAG_L3_CAT,    "l3cat",	X86_FEATURE_CAT_L3),
694 	RDT_OPT(RDT_FLAG_L3_CDP,    "l3cdp",	X86_FEATURE_CDP_L3),
695 	RDT_OPT(RDT_FLAG_L2_CAT,    "l2cat",	X86_FEATURE_CAT_L2),
696 	RDT_OPT(RDT_FLAG_L2_CDP,    "l2cdp",	X86_FEATURE_CDP_L2),
697 	RDT_OPT(RDT_FLAG_MBA,	    "mba",	X86_FEATURE_MBA),
698 	RDT_OPT(RDT_FLAG_SMBA,	    "smba",	X86_FEATURE_SMBA),
699 	RDT_OPT(RDT_FLAG_BMEC,	    "bmec",	X86_FEATURE_BMEC),
700 };
701 #define NUM_RDT_OPTIONS ARRAY_SIZE(rdt_options)
702 
703 static int __init set_rdt_options(char *str)
704 {
705 	struct rdt_options *o;
706 	bool force_off;
707 	char *tok;
708 
709 	if (*str == '=')
710 		str++;
711 	while ((tok = strsep(&str, ",")) != NULL) {
712 		force_off = *tok == '!';
713 		if (force_off)
714 			tok++;
715 		for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
716 			if (strcmp(tok, o->name) == 0) {
717 				if (force_off)
718 					o->force_off = true;
719 				else
720 					o->force_on = true;
721 				break;
722 			}
723 		}
724 	}
725 	return 1;
726 }
727 __setup("rdt", set_rdt_options);
728 
729 bool __init rdt_cpu_has(int flag)
730 {
731 	bool ret = boot_cpu_has(flag);
732 	struct rdt_options *o;
733 
734 	if (!ret)
735 		return ret;
736 
737 	for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
738 		if (flag == o->flag) {
739 			if (o->force_off)
740 				ret = false;
741 			if (o->force_on)
742 				ret = true;
743 			break;
744 		}
745 	}
746 	return ret;
747 }
748 
749 static __init bool get_mem_config(void)
750 {
751 	struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_MBA];
752 
753 	if (!rdt_cpu_has(X86_FEATURE_MBA))
754 		return false;
755 
756 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
757 		return __get_mem_config_intel(&hw_res->r_resctrl);
758 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
759 		return __rdt_get_mem_config_amd(&hw_res->r_resctrl);
760 
761 	return false;
762 }
763 
764 static __init bool get_slow_mem_config(void)
765 {
766 	struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_SMBA];
767 
768 	if (!rdt_cpu_has(X86_FEATURE_SMBA))
769 		return false;
770 
771 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
772 		return __rdt_get_mem_config_amd(&hw_res->r_resctrl);
773 
774 	return false;
775 }
776 
777 static __init bool get_rdt_alloc_resources(void)
778 {
779 	struct rdt_resource *r;
780 	bool ret = false;
781 
782 	if (rdt_alloc_capable)
783 		return true;
784 
785 	if (!boot_cpu_has(X86_FEATURE_RDT_A))
786 		return false;
787 
788 	if (rdt_cpu_has(X86_FEATURE_CAT_L3)) {
789 		r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
790 		rdt_get_cache_alloc_cfg(1, r);
791 		if (rdt_cpu_has(X86_FEATURE_CDP_L3))
792 			rdt_get_cdp_l3_config();
793 		ret = true;
794 	}
795 	if (rdt_cpu_has(X86_FEATURE_CAT_L2)) {
796 		/* CPUID 0x10.2 fields are same format at 0x10.1 */
797 		r = &rdt_resources_all[RDT_RESOURCE_L2].r_resctrl;
798 		rdt_get_cache_alloc_cfg(2, r);
799 		if (rdt_cpu_has(X86_FEATURE_CDP_L2))
800 			rdt_get_cdp_l2_config();
801 		ret = true;
802 	}
803 
804 	if (get_mem_config())
805 		ret = true;
806 
807 	if (get_slow_mem_config())
808 		ret = true;
809 
810 	return ret;
811 }
812 
813 static __init bool get_rdt_mon_resources(void)
814 {
815 	struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
816 
817 	if (rdt_cpu_has(X86_FEATURE_CQM_OCCUP_LLC))
818 		rdt_mon_features |= (1 << QOS_L3_OCCUP_EVENT_ID);
819 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL))
820 		rdt_mon_features |= (1 << QOS_L3_MBM_TOTAL_EVENT_ID);
821 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL))
822 		rdt_mon_features |= (1 << QOS_L3_MBM_LOCAL_EVENT_ID);
823 
824 	if (!rdt_mon_features)
825 		return false;
826 
827 	return !rdt_get_mon_l3_config(r);
828 }
829 
830 static __init void __check_quirks_intel(void)
831 {
832 	switch (boot_cpu_data.x86_model) {
833 	case INTEL_FAM6_HASWELL_X:
834 		if (!rdt_options[RDT_FLAG_L3_CAT].force_off)
835 			cache_alloc_hsw_probe();
836 		break;
837 	case INTEL_FAM6_SKYLAKE_X:
838 		if (boot_cpu_data.x86_stepping <= 4)
839 			set_rdt_options("!cmt,!mbmtotal,!mbmlocal,!l3cat");
840 		else
841 			set_rdt_options("!l3cat");
842 		fallthrough;
843 	case INTEL_FAM6_BROADWELL_X:
844 		intel_rdt_mbm_apply_quirk();
845 		break;
846 	}
847 }
848 
849 static __init void check_quirks(void)
850 {
851 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
852 		__check_quirks_intel();
853 }
854 
855 static __init bool get_rdt_resources(void)
856 {
857 	rdt_alloc_capable = get_rdt_alloc_resources();
858 	rdt_mon_capable = get_rdt_mon_resources();
859 
860 	return (rdt_mon_capable || rdt_alloc_capable);
861 }
862 
863 static __init void rdt_init_res_defs_intel(void)
864 {
865 	struct rdt_hw_resource *hw_res;
866 	struct rdt_resource *r;
867 
868 	for_each_rdt_resource(r) {
869 		hw_res = resctrl_to_arch_res(r);
870 
871 		if (r->rid == RDT_RESOURCE_L3 ||
872 		    r->rid == RDT_RESOURCE_L2) {
873 			r->cache.arch_has_sparse_bitmaps = false;
874 			r->cache.arch_has_per_cpu_cfg = false;
875 			r->cache.min_cbm_bits = 1;
876 		} else if (r->rid == RDT_RESOURCE_MBA) {
877 			hw_res->msr_base = MSR_IA32_MBA_THRTL_BASE;
878 			hw_res->msr_update = mba_wrmsr_intel;
879 		}
880 	}
881 }
882 
883 static __init void rdt_init_res_defs_amd(void)
884 {
885 	struct rdt_hw_resource *hw_res;
886 	struct rdt_resource *r;
887 
888 	for_each_rdt_resource(r) {
889 		hw_res = resctrl_to_arch_res(r);
890 
891 		if (r->rid == RDT_RESOURCE_L3 ||
892 		    r->rid == RDT_RESOURCE_L2) {
893 			r->cache.arch_has_sparse_bitmaps = true;
894 			r->cache.arch_has_per_cpu_cfg = true;
895 			r->cache.min_cbm_bits = 0;
896 		} else if (r->rid == RDT_RESOURCE_MBA) {
897 			hw_res->msr_base = MSR_IA32_MBA_BW_BASE;
898 			hw_res->msr_update = mba_wrmsr_amd;
899 		} else if (r->rid == RDT_RESOURCE_SMBA) {
900 			hw_res->msr_base = MSR_IA32_SMBA_BW_BASE;
901 			hw_res->msr_update = mba_wrmsr_amd;
902 		}
903 	}
904 }
905 
906 static __init void rdt_init_res_defs(void)
907 {
908 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
909 		rdt_init_res_defs_intel();
910 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
911 		rdt_init_res_defs_amd();
912 }
913 
914 static enum cpuhp_state rdt_online;
915 
916 /* Runs once on the BSP during boot. */
917 void resctrl_cpu_detect(struct cpuinfo_x86 *c)
918 {
919 	if (!cpu_has(c, X86_FEATURE_CQM_LLC)) {
920 		c->x86_cache_max_rmid  = -1;
921 		c->x86_cache_occ_scale = -1;
922 		c->x86_cache_mbm_width_offset = -1;
923 		return;
924 	}
925 
926 	/* will be overridden if occupancy monitoring exists */
927 	c->x86_cache_max_rmid = cpuid_ebx(0xf);
928 
929 	if (cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC) ||
930 	    cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL) ||
931 	    cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)) {
932 		u32 eax, ebx, ecx, edx;
933 
934 		/* QoS sub-leaf, EAX=0Fh, ECX=1 */
935 		cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx);
936 
937 		c->x86_cache_max_rmid  = ecx;
938 		c->x86_cache_occ_scale = ebx;
939 		c->x86_cache_mbm_width_offset = eax & 0xff;
940 
941 		if (c->x86_vendor == X86_VENDOR_AMD && !c->x86_cache_mbm_width_offset)
942 			c->x86_cache_mbm_width_offset = MBM_CNTR_WIDTH_OFFSET_AMD;
943 	}
944 }
945 
946 static int __init resctrl_late_init(void)
947 {
948 	struct rdt_resource *r;
949 	int state, ret;
950 
951 	/*
952 	 * Initialize functions(or definitions) that are different
953 	 * between vendors here.
954 	 */
955 	rdt_init_res_defs();
956 
957 	check_quirks();
958 
959 	if (!get_rdt_resources())
960 		return -ENODEV;
961 
962 	rdt_init_padding();
963 
964 	state = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
965 				  "x86/resctrl/cat:online:",
966 				  resctrl_online_cpu, resctrl_offline_cpu);
967 	if (state < 0)
968 		return state;
969 
970 	ret = rdtgroup_init();
971 	if (ret) {
972 		cpuhp_remove_state(state);
973 		return ret;
974 	}
975 	rdt_online = state;
976 
977 	for_each_alloc_capable_rdt_resource(r)
978 		pr_info("%s allocation detected\n", r->name);
979 
980 	for_each_mon_capable_rdt_resource(r)
981 		pr_info("%s monitoring detected\n", r->name);
982 
983 	return 0;
984 }
985 
986 late_initcall(resctrl_late_init);
987 
988 static void __exit resctrl_exit(void)
989 {
990 	cpuhp_remove_state(rdt_online);
991 	rdtgroup_exit();
992 }
993 
994 __exitcall(resctrl_exit);
995