xref: /linux/arch/x86/kernel/cpu/resctrl/core.c (revision fcab107abe1ab5be9dbe874baa722372da8f4f73)
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/cpu.h>
20 #include <linux/slab.h>
21 #include <linux/err.h>
22 #include <linux/cpuhotplug.h>
23 
24 #include <asm/cpu_device_id.h>
25 #include <asm/msr.h>
26 #include <asm/resctrl.h>
27 #include "internal.h"
28 
29 /*
30  * rdt_domain structures are kfree()d when their last CPU goes offline,
31  * and allocated when the first CPU in a new domain comes online.
32  * The rdt_resource's domain list is updated when this happens. Readers of
33  * the domain list must either take cpus_read_lock(), or rely on an RCU
34  * read-side critical section, to avoid observing concurrent modification.
35  * All writers take this mutex:
36  */
37 static DEFINE_MUTEX(domain_list_lock);
38 
39 /*
40  * The cached resctrl_pqr_state is strictly per CPU and can never be
41  * updated from a remote CPU. Functions which modify the state
42  * are called with interrupts disabled and no preemption, which
43  * is sufficient for the protection.
44  */
45 DEFINE_PER_CPU(struct resctrl_pqr_state, pqr_state);
46 
47 /*
48  * Global boolean for rdt_alloc which is true if any
49  * resource allocation is enabled.
50  */
51 bool rdt_alloc_capable;
52 
53 static void mba_wrmsr_intel(struct msr_param *m);
54 static void cat_wrmsr(struct msr_param *m);
55 static void mba_wrmsr_amd(struct msr_param *m);
56 
57 #define ctrl_domain_init(id) LIST_HEAD_INIT(rdt_resources_all[id].r_resctrl.ctrl_domains)
58 #define mon_domain_init(id) LIST_HEAD_INIT(rdt_resources_all[id].r_resctrl.mon_domains)
59 
60 struct rdt_hw_resource rdt_resources_all[RDT_NUM_RESOURCES] = {
61 	[RDT_RESOURCE_L3] =
62 	{
63 		.r_resctrl = {
64 			.name			= "L3",
65 			.ctrl_scope		= RESCTRL_L3_CACHE,
66 			.mon_scope		= RESCTRL_L3_CACHE,
67 			.ctrl_domains		= ctrl_domain_init(RDT_RESOURCE_L3),
68 			.mon_domains		= mon_domain_init(RDT_RESOURCE_L3),
69 			.schema_fmt		= RESCTRL_SCHEMA_BITMAP,
70 		},
71 		.msr_base		= MSR_IA32_L3_CBM_BASE,
72 		.msr_update		= cat_wrmsr,
73 	},
74 	[RDT_RESOURCE_L2] =
75 	{
76 		.r_resctrl = {
77 			.name			= "L2",
78 			.ctrl_scope		= RESCTRL_L2_CACHE,
79 			.ctrl_domains		= ctrl_domain_init(RDT_RESOURCE_L2),
80 			.schema_fmt		= RESCTRL_SCHEMA_BITMAP,
81 		},
82 		.msr_base		= MSR_IA32_L2_CBM_BASE,
83 		.msr_update		= cat_wrmsr,
84 	},
85 	[RDT_RESOURCE_MBA] =
86 	{
87 		.r_resctrl = {
88 			.name			= "MB",
89 			.ctrl_scope		= RESCTRL_L3_CACHE,
90 			.ctrl_domains		= ctrl_domain_init(RDT_RESOURCE_MBA),
91 			.schema_fmt		= RESCTRL_SCHEMA_RANGE,
92 		},
93 	},
94 	[RDT_RESOURCE_SMBA] =
95 	{
96 		.r_resctrl = {
97 			.name			= "SMBA",
98 			.ctrl_scope		= RESCTRL_L3_CACHE,
99 			.ctrl_domains		= ctrl_domain_init(RDT_RESOURCE_SMBA),
100 			.schema_fmt		= RESCTRL_SCHEMA_RANGE,
101 		},
102 	},
103 };
104 
105 u32 resctrl_arch_system_num_rmid_idx(void)
106 {
107 	struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
108 
109 	/* RMID are independent numbers for x86. num_rmid_idx == num_rmid */
110 	return r->num_rmid;
111 }
112 
113 struct rdt_resource *resctrl_arch_get_resource(enum resctrl_res_level l)
114 {
115 	if (l >= RDT_NUM_RESOURCES)
116 		return NULL;
117 
118 	return &rdt_resources_all[l].r_resctrl;
119 }
120 
121 /*
122  * cache_alloc_hsw_probe() - Have to probe for Intel haswell server CPUs
123  * as they do not have CPUID enumeration support for Cache allocation.
124  * The check for Vendor/Family/Model is not enough to guarantee that
125  * the MSRs won't #GP fault because only the following SKUs support
126  * CAT:
127  *	Intel(R) Xeon(R)  CPU E5-2658  v3  @  2.20GHz
128  *	Intel(R) Xeon(R)  CPU E5-2648L v3  @  1.80GHz
129  *	Intel(R) Xeon(R)  CPU E5-2628L v3  @  2.00GHz
130  *	Intel(R) Xeon(R)  CPU E5-2618L v3  @  2.30GHz
131  *	Intel(R) Xeon(R)  CPU E5-2608L v3  @  2.00GHz
132  *	Intel(R) Xeon(R)  CPU E5-2658A v3  @  2.20GHz
133  *
134  * Probe by trying to write the first of the L3 cache mask registers
135  * and checking that the bits stick. Max CLOSids is always 4 and max cbm length
136  * is always 20 on hsw server parts. The minimum cache bitmask length
137  * allowed for HSW server is always 2 bits. Hardcode all of them.
138  */
139 static inline void cache_alloc_hsw_probe(void)
140 {
141 	struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_L3];
142 	struct rdt_resource *r  = &hw_res->r_resctrl;
143 	u64 max_cbm = BIT_ULL_MASK(20) - 1, l3_cbm_0;
144 
145 	if (wrmsrq_safe(MSR_IA32_L3_CBM_BASE, max_cbm))
146 		return;
147 
148 	rdmsrq(MSR_IA32_L3_CBM_BASE, l3_cbm_0);
149 
150 	/* If all the bits were set in MSR, return success */
151 	if (l3_cbm_0 != max_cbm)
152 		return;
153 
154 	hw_res->num_closid = 4;
155 	r->cache.cbm_len = 20;
156 	r->cache.shareable_bits = 0xc0000;
157 	r->cache.min_cbm_bits = 2;
158 	r->cache.arch_has_sparse_bitmasks = false;
159 	r->alloc_capable = true;
160 
161 	rdt_alloc_capable = true;
162 }
163 
164 /*
165  * rdt_get_mb_table() - get a mapping of bandwidth(b/w) percentage values
166  * exposed to user interface and the h/w understandable delay values.
167  *
168  * The non-linear delay values have the granularity of power of two
169  * and also the h/w does not guarantee a curve for configured delay
170  * values vs. actual b/w enforced.
171  * Hence we need a mapping that is pre calibrated so the user can
172  * express the memory b/w as a percentage value.
173  */
174 static inline bool rdt_get_mb_table(struct rdt_resource *r)
175 {
176 	/*
177 	 * There are no Intel SKUs as of now to support non-linear delay.
178 	 */
179 	pr_info("MBA b/w map not implemented for cpu:%d, model:%d",
180 		boot_cpu_data.x86, boot_cpu_data.x86_model);
181 
182 	return false;
183 }
184 
185 static __init bool __get_mem_config_intel(struct rdt_resource *r)
186 {
187 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
188 	union cpuid_0x10_3_eax eax;
189 	union cpuid_0x10_x_edx edx;
190 	u32 ebx, ecx, max_delay;
191 
192 	cpuid_count(0x00000010, 3, &eax.full, &ebx, &ecx, &edx.full);
193 	hw_res->num_closid = edx.split.cos_max + 1;
194 	max_delay = eax.split.max_delay + 1;
195 	r->membw.max_bw = MAX_MBA_BW;
196 	r->membw.arch_needs_linear = true;
197 	if (ecx & MBA_IS_LINEAR) {
198 		r->membw.delay_linear = true;
199 		r->membw.min_bw = MAX_MBA_BW - max_delay;
200 		r->membw.bw_gran = MAX_MBA_BW - max_delay;
201 	} else {
202 		if (!rdt_get_mb_table(r))
203 			return false;
204 		r->membw.arch_needs_linear = false;
205 	}
206 
207 	if (boot_cpu_has(X86_FEATURE_PER_THREAD_MBA))
208 		r->membw.throttle_mode = THREAD_THROTTLE_PER_THREAD;
209 	else
210 		r->membw.throttle_mode = THREAD_THROTTLE_MAX;
211 
212 	r->alloc_capable = true;
213 
214 	return true;
215 }
216 
217 static __init bool __rdt_get_mem_config_amd(struct rdt_resource *r)
218 {
219 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
220 	u32 eax, ebx, ecx, edx, subleaf;
221 
222 	/*
223 	 * Query CPUID_Fn80000020_EDX_x01 for MBA and
224 	 * CPUID_Fn80000020_EDX_x02 for SMBA
225 	 */
226 	subleaf = (r->rid == RDT_RESOURCE_SMBA) ? 2 :  1;
227 
228 	cpuid_count(0x80000020, subleaf, &eax, &ebx, &ecx, &edx);
229 	hw_res->num_closid = edx + 1;
230 	r->membw.max_bw = 1 << eax;
231 
232 	/* AMD does not use delay */
233 	r->membw.delay_linear = false;
234 	r->membw.arch_needs_linear = false;
235 
236 	/*
237 	 * AMD does not use memory delay throttle model to control
238 	 * the allocation like Intel does.
239 	 */
240 	r->membw.throttle_mode = THREAD_THROTTLE_UNDEFINED;
241 	r->membw.min_bw = 0;
242 	r->membw.bw_gran = 1;
243 
244 	r->alloc_capable = true;
245 
246 	return true;
247 }
248 
249 static void rdt_get_cache_alloc_cfg(int idx, struct rdt_resource *r)
250 {
251 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
252 	union cpuid_0x10_1_eax eax;
253 	union cpuid_0x10_x_ecx ecx;
254 	union cpuid_0x10_x_edx edx;
255 	u32 ebx, default_ctrl;
256 
257 	cpuid_count(0x00000010, idx, &eax.full, &ebx, &ecx.full, &edx.full);
258 	hw_res->num_closid = edx.split.cos_max + 1;
259 	r->cache.cbm_len = eax.split.cbm_len + 1;
260 	default_ctrl = BIT_MASK(eax.split.cbm_len + 1) - 1;
261 	r->cache.shareable_bits = ebx & default_ctrl;
262 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
263 		r->cache.arch_has_sparse_bitmasks = ecx.split.noncont;
264 	r->alloc_capable = true;
265 }
266 
267 static void rdt_get_cdp_config(int level)
268 {
269 	/*
270 	 * By default, CDP is disabled. CDP can be enabled by mount parameter
271 	 * "cdp" during resctrl file system mount time.
272 	 */
273 	rdt_resources_all[level].cdp_enabled = false;
274 	rdt_resources_all[level].r_resctrl.cdp_capable = true;
275 }
276 
277 static void rdt_get_cdp_l3_config(void)
278 {
279 	rdt_get_cdp_config(RDT_RESOURCE_L3);
280 }
281 
282 static void rdt_get_cdp_l2_config(void)
283 {
284 	rdt_get_cdp_config(RDT_RESOURCE_L2);
285 }
286 
287 static void mba_wrmsr_amd(struct msr_param *m)
288 {
289 	struct rdt_hw_ctrl_domain *hw_dom = resctrl_to_arch_ctrl_dom(m->dom);
290 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(m->res);
291 	unsigned int i;
292 
293 	for (i = m->low; i < m->high; i++)
294 		wrmsrq(hw_res->msr_base + i, hw_dom->ctrl_val[i]);
295 }
296 
297 /*
298  * Map the memory b/w percentage value to delay values
299  * that can be written to QOS_MSRs.
300  * There are currently no SKUs which support non linear delay values.
301  */
302 static u32 delay_bw_map(unsigned long bw, struct rdt_resource *r)
303 {
304 	if (r->membw.delay_linear)
305 		return MAX_MBA_BW - bw;
306 
307 	pr_warn_once("Non Linear delay-bw map not supported but queried\n");
308 	return MAX_MBA_BW;
309 }
310 
311 static void mba_wrmsr_intel(struct msr_param *m)
312 {
313 	struct rdt_hw_ctrl_domain *hw_dom = resctrl_to_arch_ctrl_dom(m->dom);
314 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(m->res);
315 	unsigned int i;
316 
317 	/*  Write the delay values for mba. */
318 	for (i = m->low; i < m->high; i++)
319 		wrmsrq(hw_res->msr_base + i, delay_bw_map(hw_dom->ctrl_val[i], m->res));
320 }
321 
322 static void cat_wrmsr(struct msr_param *m)
323 {
324 	struct rdt_hw_ctrl_domain *hw_dom = resctrl_to_arch_ctrl_dom(m->dom);
325 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(m->res);
326 	unsigned int i;
327 
328 	for (i = m->low; i < m->high; i++)
329 		wrmsrq(hw_res->msr_base + i, hw_dom->ctrl_val[i]);
330 }
331 
332 u32 resctrl_arch_get_num_closid(struct rdt_resource *r)
333 {
334 	return resctrl_to_arch_res(r)->num_closid;
335 }
336 
337 void rdt_ctrl_update(void *arg)
338 {
339 	struct rdt_hw_resource *hw_res;
340 	struct msr_param *m = arg;
341 
342 	hw_res = resctrl_to_arch_res(m->res);
343 	hw_res->msr_update(m);
344 }
345 
346 static void setup_default_ctrlval(struct rdt_resource *r, u32 *dc)
347 {
348 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
349 	int i;
350 
351 	/*
352 	 * Initialize the Control MSRs to having no control.
353 	 * For Cache Allocation: Set all bits in cbm
354 	 * For Memory Allocation: Set b/w requested to 100%
355 	 */
356 	for (i = 0; i < hw_res->num_closid; i++, dc++)
357 		*dc = resctrl_get_default_ctrl(r);
358 }
359 
360 static void ctrl_domain_free(struct rdt_hw_ctrl_domain *hw_dom)
361 {
362 	kfree(hw_dom->ctrl_val);
363 	kfree(hw_dom);
364 }
365 
366 static void mon_domain_free(struct rdt_hw_mon_domain *hw_dom)
367 {
368 	kfree(hw_dom->arch_mbm_total);
369 	kfree(hw_dom->arch_mbm_local);
370 	kfree(hw_dom);
371 }
372 
373 static int domain_setup_ctrlval(struct rdt_resource *r, struct rdt_ctrl_domain *d)
374 {
375 	struct rdt_hw_ctrl_domain *hw_dom = resctrl_to_arch_ctrl_dom(d);
376 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
377 	struct msr_param m;
378 	u32 *dc;
379 
380 	dc = kmalloc_array(hw_res->num_closid, sizeof(*hw_dom->ctrl_val),
381 			   GFP_KERNEL);
382 	if (!dc)
383 		return -ENOMEM;
384 
385 	hw_dom->ctrl_val = dc;
386 	setup_default_ctrlval(r, dc);
387 
388 	m.res = r;
389 	m.dom = d;
390 	m.low = 0;
391 	m.high = hw_res->num_closid;
392 	hw_res->msr_update(&m);
393 	return 0;
394 }
395 
396 /**
397  * arch_domain_mbm_alloc() - Allocate arch private storage for the MBM counters
398  * @num_rmid:	The size of the MBM counter array
399  * @hw_dom:	The domain that owns the allocated arrays
400  */
401 static int arch_domain_mbm_alloc(u32 num_rmid, struct rdt_hw_mon_domain *hw_dom)
402 {
403 	size_t tsize;
404 
405 	if (resctrl_arch_is_mbm_total_enabled()) {
406 		tsize = sizeof(*hw_dom->arch_mbm_total);
407 		hw_dom->arch_mbm_total = kcalloc(num_rmid, tsize, GFP_KERNEL);
408 		if (!hw_dom->arch_mbm_total)
409 			return -ENOMEM;
410 	}
411 	if (resctrl_arch_is_mbm_local_enabled()) {
412 		tsize = sizeof(*hw_dom->arch_mbm_local);
413 		hw_dom->arch_mbm_local = kcalloc(num_rmid, tsize, GFP_KERNEL);
414 		if (!hw_dom->arch_mbm_local) {
415 			kfree(hw_dom->arch_mbm_total);
416 			hw_dom->arch_mbm_total = NULL;
417 			return -ENOMEM;
418 		}
419 	}
420 
421 	return 0;
422 }
423 
424 static int get_domain_id_from_scope(int cpu, enum resctrl_scope scope)
425 {
426 	switch (scope) {
427 	case RESCTRL_L2_CACHE:
428 	case RESCTRL_L3_CACHE:
429 		return get_cpu_cacheinfo_id(cpu, scope);
430 	case RESCTRL_L3_NODE:
431 		return cpu_to_node(cpu);
432 	default:
433 		break;
434 	}
435 
436 	return -EINVAL;
437 }
438 
439 static void domain_add_cpu_ctrl(int cpu, struct rdt_resource *r)
440 {
441 	int id = get_domain_id_from_scope(cpu, r->ctrl_scope);
442 	struct rdt_hw_ctrl_domain *hw_dom;
443 	struct list_head *add_pos = NULL;
444 	struct rdt_domain_hdr *hdr;
445 	struct rdt_ctrl_domain *d;
446 	int err;
447 
448 	lockdep_assert_held(&domain_list_lock);
449 
450 	if (id < 0) {
451 		pr_warn_once("Can't find control domain id for CPU:%d scope:%d for resource %s\n",
452 			     cpu, r->ctrl_scope, r->name);
453 		return;
454 	}
455 
456 	hdr = resctrl_find_domain(&r->ctrl_domains, id, &add_pos);
457 	if (hdr) {
458 		if (WARN_ON_ONCE(hdr->type != RESCTRL_CTRL_DOMAIN))
459 			return;
460 		d = container_of(hdr, struct rdt_ctrl_domain, hdr);
461 
462 		cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
463 		if (r->cache.arch_has_per_cpu_cfg)
464 			rdt_domain_reconfigure_cdp(r);
465 		return;
466 	}
467 
468 	hw_dom = kzalloc_node(sizeof(*hw_dom), GFP_KERNEL, cpu_to_node(cpu));
469 	if (!hw_dom)
470 		return;
471 
472 	d = &hw_dom->d_resctrl;
473 	d->hdr.id = id;
474 	d->hdr.type = RESCTRL_CTRL_DOMAIN;
475 	cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
476 
477 	rdt_domain_reconfigure_cdp(r);
478 
479 	if (domain_setup_ctrlval(r, d)) {
480 		ctrl_domain_free(hw_dom);
481 		return;
482 	}
483 
484 	list_add_tail_rcu(&d->hdr.list, add_pos);
485 
486 	err = resctrl_online_ctrl_domain(r, d);
487 	if (err) {
488 		list_del_rcu(&d->hdr.list);
489 		synchronize_rcu();
490 		ctrl_domain_free(hw_dom);
491 	}
492 }
493 
494 static void domain_add_cpu_mon(int cpu, struct rdt_resource *r)
495 {
496 	int id = get_domain_id_from_scope(cpu, r->mon_scope);
497 	struct list_head *add_pos = NULL;
498 	struct rdt_hw_mon_domain *hw_dom;
499 	struct rdt_domain_hdr *hdr;
500 	struct rdt_mon_domain *d;
501 	int err;
502 
503 	lockdep_assert_held(&domain_list_lock);
504 
505 	if (id < 0) {
506 		pr_warn_once("Can't find monitor domain id for CPU:%d scope:%d for resource %s\n",
507 			     cpu, r->mon_scope, r->name);
508 		return;
509 	}
510 
511 	hdr = resctrl_find_domain(&r->mon_domains, id, &add_pos);
512 	if (hdr) {
513 		if (WARN_ON_ONCE(hdr->type != RESCTRL_MON_DOMAIN))
514 			return;
515 		d = container_of(hdr, struct rdt_mon_domain, hdr);
516 
517 		cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
518 		return;
519 	}
520 
521 	hw_dom = kzalloc_node(sizeof(*hw_dom), GFP_KERNEL, cpu_to_node(cpu));
522 	if (!hw_dom)
523 		return;
524 
525 	d = &hw_dom->d_resctrl;
526 	d->hdr.id = id;
527 	d->hdr.type = RESCTRL_MON_DOMAIN;
528 	d->ci = get_cpu_cacheinfo_level(cpu, RESCTRL_L3_CACHE);
529 	if (!d->ci) {
530 		pr_warn_once("Can't find L3 cache for CPU:%d resource %s\n", cpu, r->name);
531 		mon_domain_free(hw_dom);
532 		return;
533 	}
534 	cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
535 
536 	arch_mon_domain_online(r, d);
537 
538 	if (arch_domain_mbm_alloc(r->num_rmid, hw_dom)) {
539 		mon_domain_free(hw_dom);
540 		return;
541 	}
542 
543 	list_add_tail_rcu(&d->hdr.list, add_pos);
544 
545 	err = resctrl_online_mon_domain(r, d);
546 	if (err) {
547 		list_del_rcu(&d->hdr.list);
548 		synchronize_rcu();
549 		mon_domain_free(hw_dom);
550 	}
551 }
552 
553 static void domain_add_cpu(int cpu, struct rdt_resource *r)
554 {
555 	if (r->alloc_capable)
556 		domain_add_cpu_ctrl(cpu, r);
557 	if (r->mon_capable)
558 		domain_add_cpu_mon(cpu, r);
559 }
560 
561 static void domain_remove_cpu_ctrl(int cpu, struct rdt_resource *r)
562 {
563 	int id = get_domain_id_from_scope(cpu, r->ctrl_scope);
564 	struct rdt_hw_ctrl_domain *hw_dom;
565 	struct rdt_domain_hdr *hdr;
566 	struct rdt_ctrl_domain *d;
567 
568 	lockdep_assert_held(&domain_list_lock);
569 
570 	if (id < 0) {
571 		pr_warn_once("Can't find control domain id for CPU:%d scope:%d for resource %s\n",
572 			     cpu, r->ctrl_scope, r->name);
573 		return;
574 	}
575 
576 	hdr = resctrl_find_domain(&r->ctrl_domains, id, NULL);
577 	if (!hdr) {
578 		pr_warn("Can't find control domain for id=%d for CPU %d for resource %s\n",
579 			id, cpu, r->name);
580 		return;
581 	}
582 
583 	if (WARN_ON_ONCE(hdr->type != RESCTRL_CTRL_DOMAIN))
584 		return;
585 
586 	d = container_of(hdr, struct rdt_ctrl_domain, hdr);
587 	hw_dom = resctrl_to_arch_ctrl_dom(d);
588 
589 	cpumask_clear_cpu(cpu, &d->hdr.cpu_mask);
590 	if (cpumask_empty(&d->hdr.cpu_mask)) {
591 		resctrl_offline_ctrl_domain(r, d);
592 		list_del_rcu(&d->hdr.list);
593 		synchronize_rcu();
594 
595 		/*
596 		 * rdt_ctrl_domain "d" is going to be freed below, so clear
597 		 * its pointer from pseudo_lock_region struct.
598 		 */
599 		if (d->plr)
600 			d->plr->d = NULL;
601 		ctrl_domain_free(hw_dom);
602 
603 		return;
604 	}
605 }
606 
607 static void domain_remove_cpu_mon(int cpu, struct rdt_resource *r)
608 {
609 	int id = get_domain_id_from_scope(cpu, r->mon_scope);
610 	struct rdt_hw_mon_domain *hw_dom;
611 	struct rdt_domain_hdr *hdr;
612 	struct rdt_mon_domain *d;
613 
614 	lockdep_assert_held(&domain_list_lock);
615 
616 	if (id < 0) {
617 		pr_warn_once("Can't find monitor domain id for CPU:%d scope:%d for resource %s\n",
618 			     cpu, r->mon_scope, r->name);
619 		return;
620 	}
621 
622 	hdr = resctrl_find_domain(&r->mon_domains, id, NULL);
623 	if (!hdr) {
624 		pr_warn("Can't find monitor domain for id=%d for CPU %d for resource %s\n",
625 			id, cpu, r->name);
626 		return;
627 	}
628 
629 	if (WARN_ON_ONCE(hdr->type != RESCTRL_MON_DOMAIN))
630 		return;
631 
632 	d = container_of(hdr, struct rdt_mon_domain, hdr);
633 	hw_dom = resctrl_to_arch_mon_dom(d);
634 
635 	cpumask_clear_cpu(cpu, &d->hdr.cpu_mask);
636 	if (cpumask_empty(&d->hdr.cpu_mask)) {
637 		resctrl_offline_mon_domain(r, d);
638 		list_del_rcu(&d->hdr.list);
639 		synchronize_rcu();
640 		mon_domain_free(hw_dom);
641 
642 		return;
643 	}
644 }
645 
646 static void domain_remove_cpu(int cpu, struct rdt_resource *r)
647 {
648 	if (r->alloc_capable)
649 		domain_remove_cpu_ctrl(cpu, r);
650 	if (r->mon_capable)
651 		domain_remove_cpu_mon(cpu, r);
652 }
653 
654 static void clear_closid_rmid(int cpu)
655 {
656 	struct resctrl_pqr_state *state = this_cpu_ptr(&pqr_state);
657 
658 	state->default_closid = RESCTRL_RESERVED_CLOSID;
659 	state->default_rmid = RESCTRL_RESERVED_RMID;
660 	state->cur_closid = RESCTRL_RESERVED_CLOSID;
661 	state->cur_rmid = RESCTRL_RESERVED_RMID;
662 	wrmsr(MSR_IA32_PQR_ASSOC, RESCTRL_RESERVED_RMID,
663 	      RESCTRL_RESERVED_CLOSID);
664 }
665 
666 static int resctrl_arch_online_cpu(unsigned int cpu)
667 {
668 	struct rdt_resource *r;
669 
670 	mutex_lock(&domain_list_lock);
671 	for_each_capable_rdt_resource(r)
672 		domain_add_cpu(cpu, r);
673 	mutex_unlock(&domain_list_lock);
674 
675 	clear_closid_rmid(cpu);
676 	resctrl_online_cpu(cpu);
677 
678 	return 0;
679 }
680 
681 static int resctrl_arch_offline_cpu(unsigned int cpu)
682 {
683 	struct rdt_resource *r;
684 
685 	resctrl_offline_cpu(cpu);
686 
687 	mutex_lock(&domain_list_lock);
688 	for_each_capable_rdt_resource(r)
689 		domain_remove_cpu(cpu, r);
690 	mutex_unlock(&domain_list_lock);
691 
692 	clear_closid_rmid(cpu);
693 
694 	return 0;
695 }
696 
697 enum {
698 	RDT_FLAG_CMT,
699 	RDT_FLAG_MBM_TOTAL,
700 	RDT_FLAG_MBM_LOCAL,
701 	RDT_FLAG_L3_CAT,
702 	RDT_FLAG_L3_CDP,
703 	RDT_FLAG_L2_CAT,
704 	RDT_FLAG_L2_CDP,
705 	RDT_FLAG_MBA,
706 	RDT_FLAG_SMBA,
707 	RDT_FLAG_BMEC,
708 };
709 
710 #define RDT_OPT(idx, n, f)	\
711 [idx] = {			\
712 	.name = n,		\
713 	.flag = f		\
714 }
715 
716 struct rdt_options {
717 	char	*name;
718 	int	flag;
719 	bool	force_off, force_on;
720 };
721 
722 static struct rdt_options rdt_options[]  __ro_after_init = {
723 	RDT_OPT(RDT_FLAG_CMT,	    "cmt",	X86_FEATURE_CQM_OCCUP_LLC),
724 	RDT_OPT(RDT_FLAG_MBM_TOTAL, "mbmtotal", X86_FEATURE_CQM_MBM_TOTAL),
725 	RDT_OPT(RDT_FLAG_MBM_LOCAL, "mbmlocal", X86_FEATURE_CQM_MBM_LOCAL),
726 	RDT_OPT(RDT_FLAG_L3_CAT,    "l3cat",	X86_FEATURE_CAT_L3),
727 	RDT_OPT(RDT_FLAG_L3_CDP,    "l3cdp",	X86_FEATURE_CDP_L3),
728 	RDT_OPT(RDT_FLAG_L2_CAT,    "l2cat",	X86_FEATURE_CAT_L2),
729 	RDT_OPT(RDT_FLAG_L2_CDP,    "l2cdp",	X86_FEATURE_CDP_L2),
730 	RDT_OPT(RDT_FLAG_MBA,	    "mba",	X86_FEATURE_MBA),
731 	RDT_OPT(RDT_FLAG_SMBA,	    "smba",	X86_FEATURE_SMBA),
732 	RDT_OPT(RDT_FLAG_BMEC,	    "bmec",	X86_FEATURE_BMEC),
733 };
734 #define NUM_RDT_OPTIONS ARRAY_SIZE(rdt_options)
735 
736 static int __init set_rdt_options(char *str)
737 {
738 	struct rdt_options *o;
739 	bool force_off;
740 	char *tok;
741 
742 	if (*str == '=')
743 		str++;
744 	while ((tok = strsep(&str, ",")) != NULL) {
745 		force_off = *tok == '!';
746 		if (force_off)
747 			tok++;
748 		for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
749 			if (strcmp(tok, o->name) == 0) {
750 				if (force_off)
751 					o->force_off = true;
752 				else
753 					o->force_on = true;
754 				break;
755 			}
756 		}
757 	}
758 	return 1;
759 }
760 __setup("rdt", set_rdt_options);
761 
762 bool rdt_cpu_has(int flag)
763 {
764 	bool ret = boot_cpu_has(flag);
765 	struct rdt_options *o;
766 
767 	if (!ret)
768 		return ret;
769 
770 	for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
771 		if (flag == o->flag) {
772 			if (o->force_off)
773 				ret = false;
774 			if (o->force_on)
775 				ret = true;
776 			break;
777 		}
778 	}
779 	return ret;
780 }
781 
782 bool resctrl_arch_is_evt_configurable(enum resctrl_event_id evt)
783 {
784 	if (!rdt_cpu_has(X86_FEATURE_BMEC))
785 		return false;
786 
787 	switch (evt) {
788 	case QOS_L3_MBM_TOTAL_EVENT_ID:
789 		return rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL);
790 	case QOS_L3_MBM_LOCAL_EVENT_ID:
791 		return rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL);
792 	default:
793 		return false;
794 	}
795 }
796 
797 static __init bool get_mem_config(void)
798 {
799 	struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_MBA];
800 
801 	if (!rdt_cpu_has(X86_FEATURE_MBA))
802 		return false;
803 
804 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
805 		return __get_mem_config_intel(&hw_res->r_resctrl);
806 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
807 		return __rdt_get_mem_config_amd(&hw_res->r_resctrl);
808 
809 	return false;
810 }
811 
812 static __init bool get_slow_mem_config(void)
813 {
814 	struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_SMBA];
815 
816 	if (!rdt_cpu_has(X86_FEATURE_SMBA))
817 		return false;
818 
819 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
820 		return __rdt_get_mem_config_amd(&hw_res->r_resctrl);
821 
822 	return false;
823 }
824 
825 static __init bool get_rdt_alloc_resources(void)
826 {
827 	struct rdt_resource *r;
828 	bool ret = false;
829 
830 	if (rdt_alloc_capable)
831 		return true;
832 
833 	if (!boot_cpu_has(X86_FEATURE_RDT_A))
834 		return false;
835 
836 	if (rdt_cpu_has(X86_FEATURE_CAT_L3)) {
837 		r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
838 		rdt_get_cache_alloc_cfg(1, r);
839 		if (rdt_cpu_has(X86_FEATURE_CDP_L3))
840 			rdt_get_cdp_l3_config();
841 		ret = true;
842 	}
843 	if (rdt_cpu_has(X86_FEATURE_CAT_L2)) {
844 		/* CPUID 0x10.2 fields are same format at 0x10.1 */
845 		r = &rdt_resources_all[RDT_RESOURCE_L2].r_resctrl;
846 		rdt_get_cache_alloc_cfg(2, r);
847 		if (rdt_cpu_has(X86_FEATURE_CDP_L2))
848 			rdt_get_cdp_l2_config();
849 		ret = true;
850 	}
851 
852 	if (get_mem_config())
853 		ret = true;
854 
855 	if (get_slow_mem_config())
856 		ret = true;
857 
858 	return ret;
859 }
860 
861 static __init bool get_rdt_mon_resources(void)
862 {
863 	struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
864 
865 	if (rdt_cpu_has(X86_FEATURE_CQM_OCCUP_LLC))
866 		rdt_mon_features |= (1 << QOS_L3_OCCUP_EVENT_ID);
867 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL))
868 		rdt_mon_features |= (1 << QOS_L3_MBM_TOTAL_EVENT_ID);
869 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL))
870 		rdt_mon_features |= (1 << QOS_L3_MBM_LOCAL_EVENT_ID);
871 
872 	if (!rdt_mon_features)
873 		return false;
874 
875 	return !rdt_get_mon_l3_config(r);
876 }
877 
878 static __init void __check_quirks_intel(void)
879 {
880 	switch (boot_cpu_data.x86_vfm) {
881 	case INTEL_HASWELL_X:
882 		if (!rdt_options[RDT_FLAG_L3_CAT].force_off)
883 			cache_alloc_hsw_probe();
884 		break;
885 	case INTEL_SKYLAKE_X:
886 		if (boot_cpu_data.x86_stepping <= 4)
887 			set_rdt_options("!cmt,!mbmtotal,!mbmlocal,!l3cat");
888 		else
889 			set_rdt_options("!l3cat");
890 		fallthrough;
891 	case INTEL_BROADWELL_X:
892 		intel_rdt_mbm_apply_quirk();
893 		break;
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_hw_resource *hw_res;
914 	struct rdt_resource *r;
915 
916 	for_each_rdt_resource(r) {
917 		hw_res = resctrl_to_arch_res(r);
918 
919 		if (r->rid == RDT_RESOURCE_L3 ||
920 		    r->rid == RDT_RESOURCE_L2) {
921 			r->cache.arch_has_per_cpu_cfg = false;
922 			r->cache.min_cbm_bits = 1;
923 		} else if (r->rid == RDT_RESOURCE_MBA) {
924 			hw_res->msr_base = MSR_IA32_MBA_THRTL_BASE;
925 			hw_res->msr_update = mba_wrmsr_intel;
926 		}
927 	}
928 }
929 
930 static __init void rdt_init_res_defs_amd(void)
931 {
932 	struct rdt_hw_resource *hw_res;
933 	struct rdt_resource *r;
934 
935 	for_each_rdt_resource(r) {
936 		hw_res = resctrl_to_arch_res(r);
937 
938 		if (r->rid == RDT_RESOURCE_L3 ||
939 		    r->rid == RDT_RESOURCE_L2) {
940 			r->cache.arch_has_sparse_bitmasks = true;
941 			r->cache.arch_has_per_cpu_cfg = true;
942 			r->cache.min_cbm_bits = 0;
943 		} else if (r->rid == RDT_RESOURCE_MBA) {
944 			hw_res->msr_base = MSR_IA32_MBA_BW_BASE;
945 			hw_res->msr_update = mba_wrmsr_amd;
946 		} else if (r->rid == RDT_RESOURCE_SMBA) {
947 			hw_res->msr_base = MSR_IA32_SMBA_BW_BASE;
948 			hw_res->msr_update = mba_wrmsr_amd;
949 		}
950 	}
951 }
952 
953 static __init void rdt_init_res_defs(void)
954 {
955 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
956 		rdt_init_res_defs_intel();
957 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
958 		rdt_init_res_defs_amd();
959 }
960 
961 static enum cpuhp_state rdt_online;
962 
963 /* Runs once on the BSP during boot. */
964 void resctrl_cpu_detect(struct cpuinfo_x86 *c)
965 {
966 	if (!cpu_has(c, X86_FEATURE_CQM_LLC)) {
967 		c->x86_cache_max_rmid  = -1;
968 		c->x86_cache_occ_scale = -1;
969 		c->x86_cache_mbm_width_offset = -1;
970 		return;
971 	}
972 
973 	/* will be overridden if occupancy monitoring exists */
974 	c->x86_cache_max_rmid = cpuid_ebx(0xf);
975 
976 	if (cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC) ||
977 	    cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL) ||
978 	    cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)) {
979 		u32 eax, ebx, ecx, edx;
980 
981 		/* QoS sub-leaf, EAX=0Fh, ECX=1 */
982 		cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx);
983 
984 		c->x86_cache_max_rmid  = ecx;
985 		c->x86_cache_occ_scale = ebx;
986 		c->x86_cache_mbm_width_offset = eax & 0xff;
987 
988 		if (c->x86_vendor == X86_VENDOR_AMD && !c->x86_cache_mbm_width_offset)
989 			c->x86_cache_mbm_width_offset = MBM_CNTR_WIDTH_OFFSET_AMD;
990 	}
991 }
992 
993 static int __init resctrl_arch_late_init(void)
994 {
995 	struct rdt_resource *r;
996 	int state, ret, i;
997 
998 	/* for_each_rdt_resource() requires all rid to be initialised. */
999 	for (i = 0; i < RDT_NUM_RESOURCES; i++)
1000 		rdt_resources_all[i].r_resctrl.rid = i;
1001 
1002 	/*
1003 	 * Initialize functions(or definitions) that are different
1004 	 * between vendors here.
1005 	 */
1006 	rdt_init_res_defs();
1007 
1008 	check_quirks();
1009 
1010 	if (!get_rdt_resources())
1011 		return -ENODEV;
1012 
1013 	state = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
1014 				  "x86/resctrl/cat:online:",
1015 				  resctrl_arch_online_cpu,
1016 				  resctrl_arch_offline_cpu);
1017 	if (state < 0)
1018 		return state;
1019 
1020 	ret = resctrl_init();
1021 	if (ret) {
1022 		cpuhp_remove_state(state);
1023 		return ret;
1024 	}
1025 	rdt_online = state;
1026 
1027 	for_each_alloc_capable_rdt_resource(r)
1028 		pr_info("%s allocation detected\n", r->name);
1029 
1030 	for_each_mon_capable_rdt_resource(r)
1031 		pr_info("%s monitoring detected\n", r->name);
1032 
1033 	return 0;
1034 }
1035 
1036 late_initcall(resctrl_arch_late_init);
1037 
1038 static void __exit resctrl_arch_exit(void)
1039 {
1040 	cpuhp_remove_state(rdt_online);
1041 
1042 	resctrl_exit();
1043 }
1044 
1045 __exitcall(resctrl_arch_exit);
1046