xref: /linux/arch/x86/kernel/cpu/resctrl/core.c (revision 7fc2cd2e4b398c57c9cf961cfea05eadbf34c05c)
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->mon.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_set_io_alloc_capable(struct rdt_resource *r)
278 {
279 	r->cache.io_alloc_capable = true;
280 }
281 
282 static void rdt_get_cdp_l3_config(void)
283 {
284 	rdt_get_cdp_config(RDT_RESOURCE_L3);
285 }
286 
287 static void rdt_get_cdp_l2_config(void)
288 {
289 	rdt_get_cdp_config(RDT_RESOURCE_L2);
290 }
291 
292 static void mba_wrmsr_amd(struct msr_param *m)
293 {
294 	struct rdt_hw_ctrl_domain *hw_dom = resctrl_to_arch_ctrl_dom(m->dom);
295 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(m->res);
296 	unsigned int i;
297 
298 	for (i = m->low; i < m->high; i++)
299 		wrmsrq(hw_res->msr_base + i, hw_dom->ctrl_val[i]);
300 }
301 
302 /*
303  * Map the memory b/w percentage value to delay values
304  * that can be written to QOS_MSRs.
305  * There are currently no SKUs which support non linear delay values.
306  */
307 static u32 delay_bw_map(unsigned long bw, struct rdt_resource *r)
308 {
309 	if (r->membw.delay_linear)
310 		return MAX_MBA_BW - bw;
311 
312 	pr_warn_once("Non Linear delay-bw map not supported but queried\n");
313 	return MAX_MBA_BW;
314 }
315 
316 static void mba_wrmsr_intel(struct msr_param *m)
317 {
318 	struct rdt_hw_ctrl_domain *hw_dom = resctrl_to_arch_ctrl_dom(m->dom);
319 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(m->res);
320 	unsigned int i;
321 
322 	/*  Write the delay values for mba. */
323 	for (i = m->low; i < m->high; i++)
324 		wrmsrq(hw_res->msr_base + i, delay_bw_map(hw_dom->ctrl_val[i], m->res));
325 }
326 
327 static void cat_wrmsr(struct msr_param *m)
328 {
329 	struct rdt_hw_ctrl_domain *hw_dom = resctrl_to_arch_ctrl_dom(m->dom);
330 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(m->res);
331 	unsigned int i;
332 
333 	for (i = m->low; i < m->high; i++)
334 		wrmsrq(hw_res->msr_base + i, hw_dom->ctrl_val[i]);
335 }
336 
337 u32 resctrl_arch_get_num_closid(struct rdt_resource *r)
338 {
339 	return resctrl_to_arch_res(r)->num_closid;
340 }
341 
342 void rdt_ctrl_update(void *arg)
343 {
344 	struct rdt_hw_resource *hw_res;
345 	struct msr_param *m = arg;
346 
347 	hw_res = resctrl_to_arch_res(m->res);
348 	hw_res->msr_update(m);
349 }
350 
351 static void setup_default_ctrlval(struct rdt_resource *r, u32 *dc)
352 {
353 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
354 	int i;
355 
356 	/*
357 	 * Initialize the Control MSRs to having no control.
358 	 * For Cache Allocation: Set all bits in cbm
359 	 * For Memory Allocation: Set b/w requested to 100%
360 	 */
361 	for (i = 0; i < hw_res->num_closid; i++, dc++)
362 		*dc = resctrl_get_default_ctrl(r);
363 }
364 
365 static void ctrl_domain_free(struct rdt_hw_ctrl_domain *hw_dom)
366 {
367 	kfree(hw_dom->ctrl_val);
368 	kfree(hw_dom);
369 }
370 
371 static void mon_domain_free(struct rdt_hw_mon_domain *hw_dom)
372 {
373 	int idx;
374 
375 	for_each_mbm_idx(idx)
376 		kfree(hw_dom->arch_mbm_states[idx]);
377 	kfree(hw_dom);
378 }
379 
380 static int domain_setup_ctrlval(struct rdt_resource *r, struct rdt_ctrl_domain *d)
381 {
382 	struct rdt_hw_ctrl_domain *hw_dom = resctrl_to_arch_ctrl_dom(d);
383 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
384 	struct msr_param m;
385 	u32 *dc;
386 
387 	dc = kmalloc_array(hw_res->num_closid, sizeof(*hw_dom->ctrl_val),
388 			   GFP_KERNEL);
389 	if (!dc)
390 		return -ENOMEM;
391 
392 	hw_dom->ctrl_val = dc;
393 	setup_default_ctrlval(r, dc);
394 
395 	m.res = r;
396 	m.dom = d;
397 	m.low = 0;
398 	m.high = hw_res->num_closid;
399 	hw_res->msr_update(&m);
400 	return 0;
401 }
402 
403 /**
404  * arch_domain_mbm_alloc() - Allocate arch private storage for the MBM counters
405  * @num_rmid:	The size of the MBM counter array
406  * @hw_dom:	The domain that owns the allocated arrays
407  */
408 static int arch_domain_mbm_alloc(u32 num_rmid, struct rdt_hw_mon_domain *hw_dom)
409 {
410 	size_t tsize = sizeof(*hw_dom->arch_mbm_states[0]);
411 	enum resctrl_event_id eventid;
412 	int idx;
413 
414 	for_each_mbm_event_id(eventid) {
415 		if (!resctrl_is_mon_event_enabled(eventid))
416 			continue;
417 		idx = MBM_STATE_IDX(eventid);
418 		hw_dom->arch_mbm_states[idx] = kcalloc(num_rmid, tsize, GFP_KERNEL);
419 		if (!hw_dom->arch_mbm_states[idx])
420 			goto cleanup;
421 	}
422 
423 	return 0;
424 cleanup:
425 	for_each_mbm_idx(idx) {
426 		kfree(hw_dom->arch_mbm_states[idx]);
427 		hw_dom->arch_mbm_states[idx] = NULL;
428 	}
429 
430 	return -ENOMEM;
431 }
432 
433 static int get_domain_id_from_scope(int cpu, enum resctrl_scope scope)
434 {
435 	switch (scope) {
436 	case RESCTRL_L2_CACHE:
437 	case RESCTRL_L3_CACHE:
438 		return get_cpu_cacheinfo_id(cpu, scope);
439 	case RESCTRL_L3_NODE:
440 		return cpu_to_node(cpu);
441 	default:
442 		break;
443 	}
444 
445 	return -EINVAL;
446 }
447 
448 static void domain_add_cpu_ctrl(int cpu, struct rdt_resource *r)
449 {
450 	int id = get_domain_id_from_scope(cpu, r->ctrl_scope);
451 	struct rdt_hw_ctrl_domain *hw_dom;
452 	struct list_head *add_pos = NULL;
453 	struct rdt_domain_hdr *hdr;
454 	struct rdt_ctrl_domain *d;
455 	int err;
456 
457 	lockdep_assert_held(&domain_list_lock);
458 
459 	if (id < 0) {
460 		pr_warn_once("Can't find control domain id for CPU:%d scope:%d for resource %s\n",
461 			     cpu, r->ctrl_scope, r->name);
462 		return;
463 	}
464 
465 	hdr = resctrl_find_domain(&r->ctrl_domains, id, &add_pos);
466 	if (hdr) {
467 		if (WARN_ON_ONCE(hdr->type != RESCTRL_CTRL_DOMAIN))
468 			return;
469 		d = container_of(hdr, struct rdt_ctrl_domain, hdr);
470 
471 		cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
472 		if (r->cache.arch_has_per_cpu_cfg)
473 			rdt_domain_reconfigure_cdp(r);
474 		return;
475 	}
476 
477 	hw_dom = kzalloc_node(sizeof(*hw_dom), GFP_KERNEL, cpu_to_node(cpu));
478 	if (!hw_dom)
479 		return;
480 
481 	d = &hw_dom->d_resctrl;
482 	d->hdr.id = id;
483 	d->hdr.type = RESCTRL_CTRL_DOMAIN;
484 	cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
485 
486 	rdt_domain_reconfigure_cdp(r);
487 
488 	if (domain_setup_ctrlval(r, d)) {
489 		ctrl_domain_free(hw_dom);
490 		return;
491 	}
492 
493 	list_add_tail_rcu(&d->hdr.list, add_pos);
494 
495 	err = resctrl_online_ctrl_domain(r, d);
496 	if (err) {
497 		list_del_rcu(&d->hdr.list);
498 		synchronize_rcu();
499 		ctrl_domain_free(hw_dom);
500 	}
501 }
502 
503 static void domain_add_cpu_mon(int cpu, struct rdt_resource *r)
504 {
505 	int id = get_domain_id_from_scope(cpu, r->mon_scope);
506 	struct list_head *add_pos = NULL;
507 	struct rdt_hw_mon_domain *hw_dom;
508 	struct rdt_domain_hdr *hdr;
509 	struct rdt_mon_domain *d;
510 	struct cacheinfo *ci;
511 	int err;
512 
513 	lockdep_assert_held(&domain_list_lock);
514 
515 	if (id < 0) {
516 		pr_warn_once("Can't find monitor domain id for CPU:%d scope:%d for resource %s\n",
517 			     cpu, r->mon_scope, r->name);
518 		return;
519 	}
520 
521 	hdr = resctrl_find_domain(&r->mon_domains, id, &add_pos);
522 	if (hdr) {
523 		if (WARN_ON_ONCE(hdr->type != RESCTRL_MON_DOMAIN))
524 			return;
525 		d = container_of(hdr, struct rdt_mon_domain, hdr);
526 
527 		cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
528 		/* Update the mbm_assign_mode state for the CPU if supported */
529 		if (r->mon.mbm_cntr_assignable)
530 			resctrl_arch_mbm_cntr_assign_set_one(r);
531 		return;
532 	}
533 
534 	hw_dom = kzalloc_node(sizeof(*hw_dom), GFP_KERNEL, cpu_to_node(cpu));
535 	if (!hw_dom)
536 		return;
537 
538 	d = &hw_dom->d_resctrl;
539 	d->hdr.id = id;
540 	d->hdr.type = RESCTRL_MON_DOMAIN;
541 	ci = get_cpu_cacheinfo_level(cpu, RESCTRL_L3_CACHE);
542 	if (!ci) {
543 		pr_warn_once("Can't find L3 cache for CPU:%d resource %s\n", cpu, r->name);
544 		mon_domain_free(hw_dom);
545 		return;
546 	}
547 	d->ci_id = ci->id;
548 	cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
549 
550 	/* Update the mbm_assign_mode state for the CPU if supported */
551 	if (r->mon.mbm_cntr_assignable)
552 		resctrl_arch_mbm_cntr_assign_set_one(r);
553 
554 	arch_mon_domain_online(r, d);
555 
556 	if (arch_domain_mbm_alloc(r->mon.num_rmid, hw_dom)) {
557 		mon_domain_free(hw_dom);
558 		return;
559 	}
560 
561 	list_add_tail_rcu(&d->hdr.list, add_pos);
562 
563 	err = resctrl_online_mon_domain(r, d);
564 	if (err) {
565 		list_del_rcu(&d->hdr.list);
566 		synchronize_rcu();
567 		mon_domain_free(hw_dom);
568 	}
569 }
570 
571 static void domain_add_cpu(int cpu, struct rdt_resource *r)
572 {
573 	if (r->alloc_capable)
574 		domain_add_cpu_ctrl(cpu, r);
575 	if (r->mon_capable)
576 		domain_add_cpu_mon(cpu, r);
577 }
578 
579 static void domain_remove_cpu_ctrl(int cpu, struct rdt_resource *r)
580 {
581 	int id = get_domain_id_from_scope(cpu, r->ctrl_scope);
582 	struct rdt_hw_ctrl_domain *hw_dom;
583 	struct rdt_domain_hdr *hdr;
584 	struct rdt_ctrl_domain *d;
585 
586 	lockdep_assert_held(&domain_list_lock);
587 
588 	if (id < 0) {
589 		pr_warn_once("Can't find control domain id for CPU:%d scope:%d for resource %s\n",
590 			     cpu, r->ctrl_scope, r->name);
591 		return;
592 	}
593 
594 	hdr = resctrl_find_domain(&r->ctrl_domains, id, NULL);
595 	if (!hdr) {
596 		pr_warn("Can't find control domain for id=%d for CPU %d for resource %s\n",
597 			id, cpu, r->name);
598 		return;
599 	}
600 
601 	if (WARN_ON_ONCE(hdr->type != RESCTRL_CTRL_DOMAIN))
602 		return;
603 
604 	d = container_of(hdr, struct rdt_ctrl_domain, hdr);
605 	hw_dom = resctrl_to_arch_ctrl_dom(d);
606 
607 	cpumask_clear_cpu(cpu, &d->hdr.cpu_mask);
608 	if (cpumask_empty(&d->hdr.cpu_mask)) {
609 		resctrl_offline_ctrl_domain(r, d);
610 		list_del_rcu(&d->hdr.list);
611 		synchronize_rcu();
612 
613 		/*
614 		 * rdt_ctrl_domain "d" is going to be freed below, so clear
615 		 * its pointer from pseudo_lock_region struct.
616 		 */
617 		if (d->plr)
618 			d->plr->d = NULL;
619 		ctrl_domain_free(hw_dom);
620 
621 		return;
622 	}
623 }
624 
625 static void domain_remove_cpu_mon(int cpu, struct rdt_resource *r)
626 {
627 	int id = get_domain_id_from_scope(cpu, r->mon_scope);
628 	struct rdt_hw_mon_domain *hw_dom;
629 	struct rdt_domain_hdr *hdr;
630 	struct rdt_mon_domain *d;
631 
632 	lockdep_assert_held(&domain_list_lock);
633 
634 	if (id < 0) {
635 		pr_warn_once("Can't find monitor domain id for CPU:%d scope:%d for resource %s\n",
636 			     cpu, r->mon_scope, r->name);
637 		return;
638 	}
639 
640 	hdr = resctrl_find_domain(&r->mon_domains, id, NULL);
641 	if (!hdr) {
642 		pr_warn("Can't find monitor domain for id=%d for CPU %d for resource %s\n",
643 			id, cpu, r->name);
644 		return;
645 	}
646 
647 	if (WARN_ON_ONCE(hdr->type != RESCTRL_MON_DOMAIN))
648 		return;
649 
650 	d = container_of(hdr, struct rdt_mon_domain, hdr);
651 	hw_dom = resctrl_to_arch_mon_dom(d);
652 
653 	cpumask_clear_cpu(cpu, &d->hdr.cpu_mask);
654 	if (cpumask_empty(&d->hdr.cpu_mask)) {
655 		resctrl_offline_mon_domain(r, d);
656 		list_del_rcu(&d->hdr.list);
657 		synchronize_rcu();
658 		mon_domain_free(hw_dom);
659 
660 		return;
661 	}
662 }
663 
664 static void domain_remove_cpu(int cpu, struct rdt_resource *r)
665 {
666 	if (r->alloc_capable)
667 		domain_remove_cpu_ctrl(cpu, r);
668 	if (r->mon_capable)
669 		domain_remove_cpu_mon(cpu, r);
670 }
671 
672 static void clear_closid_rmid(int cpu)
673 {
674 	struct resctrl_pqr_state *state = this_cpu_ptr(&pqr_state);
675 
676 	state->default_closid = RESCTRL_RESERVED_CLOSID;
677 	state->default_rmid = RESCTRL_RESERVED_RMID;
678 	state->cur_closid = RESCTRL_RESERVED_CLOSID;
679 	state->cur_rmid = RESCTRL_RESERVED_RMID;
680 	wrmsr(MSR_IA32_PQR_ASSOC, RESCTRL_RESERVED_RMID,
681 	      RESCTRL_RESERVED_CLOSID);
682 }
683 
684 static int resctrl_arch_online_cpu(unsigned int cpu)
685 {
686 	struct rdt_resource *r;
687 
688 	mutex_lock(&domain_list_lock);
689 	for_each_capable_rdt_resource(r)
690 		domain_add_cpu(cpu, r);
691 	mutex_unlock(&domain_list_lock);
692 
693 	clear_closid_rmid(cpu);
694 	resctrl_online_cpu(cpu);
695 
696 	return 0;
697 }
698 
699 static int resctrl_arch_offline_cpu(unsigned int cpu)
700 {
701 	struct rdt_resource *r;
702 
703 	resctrl_offline_cpu(cpu);
704 
705 	mutex_lock(&domain_list_lock);
706 	for_each_capable_rdt_resource(r)
707 		domain_remove_cpu(cpu, r);
708 	mutex_unlock(&domain_list_lock);
709 
710 	clear_closid_rmid(cpu);
711 
712 	return 0;
713 }
714 
715 enum {
716 	RDT_FLAG_CMT,
717 	RDT_FLAG_MBM_TOTAL,
718 	RDT_FLAG_MBM_LOCAL,
719 	RDT_FLAG_L3_CAT,
720 	RDT_FLAG_L3_CDP,
721 	RDT_FLAG_L2_CAT,
722 	RDT_FLAG_L2_CDP,
723 	RDT_FLAG_MBA,
724 	RDT_FLAG_SMBA,
725 	RDT_FLAG_BMEC,
726 	RDT_FLAG_ABMC,
727 	RDT_FLAG_SDCIAE,
728 };
729 
730 #define RDT_OPT(idx, n, f)	\
731 [idx] = {			\
732 	.name = n,		\
733 	.flag = f		\
734 }
735 
736 struct rdt_options {
737 	char	*name;
738 	int	flag;
739 	bool	force_off, force_on;
740 };
741 
742 static struct rdt_options rdt_options[]  __ro_after_init = {
743 	RDT_OPT(RDT_FLAG_CMT,	    "cmt",	X86_FEATURE_CQM_OCCUP_LLC),
744 	RDT_OPT(RDT_FLAG_MBM_TOTAL, "mbmtotal", X86_FEATURE_CQM_MBM_TOTAL),
745 	RDT_OPT(RDT_FLAG_MBM_LOCAL, "mbmlocal", X86_FEATURE_CQM_MBM_LOCAL),
746 	RDT_OPT(RDT_FLAG_L3_CAT,    "l3cat",	X86_FEATURE_CAT_L3),
747 	RDT_OPT(RDT_FLAG_L3_CDP,    "l3cdp",	X86_FEATURE_CDP_L3),
748 	RDT_OPT(RDT_FLAG_L2_CAT,    "l2cat",	X86_FEATURE_CAT_L2),
749 	RDT_OPT(RDT_FLAG_L2_CDP,    "l2cdp",	X86_FEATURE_CDP_L2),
750 	RDT_OPT(RDT_FLAG_MBA,	    "mba",	X86_FEATURE_MBA),
751 	RDT_OPT(RDT_FLAG_SMBA,	    "smba",	X86_FEATURE_SMBA),
752 	RDT_OPT(RDT_FLAG_BMEC,	    "bmec",	X86_FEATURE_BMEC),
753 	RDT_OPT(RDT_FLAG_ABMC,	    "abmc",	X86_FEATURE_ABMC),
754 	RDT_OPT(RDT_FLAG_SDCIAE,    "sdciae",	X86_FEATURE_SDCIAE),
755 };
756 #define NUM_RDT_OPTIONS ARRAY_SIZE(rdt_options)
757 
758 static int __init set_rdt_options(char *str)
759 {
760 	struct rdt_options *o;
761 	bool force_off;
762 	char *tok;
763 
764 	if (*str == '=')
765 		str++;
766 	while ((tok = strsep(&str, ",")) != NULL) {
767 		force_off = *tok == '!';
768 		if (force_off)
769 			tok++;
770 		for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
771 			if (strcmp(tok, o->name) == 0) {
772 				if (force_off)
773 					o->force_off = true;
774 				else
775 					o->force_on = true;
776 				break;
777 			}
778 		}
779 	}
780 	return 1;
781 }
782 __setup("rdt", set_rdt_options);
783 
784 bool rdt_cpu_has(int flag)
785 {
786 	bool ret = boot_cpu_has(flag);
787 	struct rdt_options *o;
788 
789 	if (!ret)
790 		return ret;
791 
792 	for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
793 		if (flag == o->flag) {
794 			if (o->force_off)
795 				ret = false;
796 			if (o->force_on)
797 				ret = true;
798 			break;
799 		}
800 	}
801 	return ret;
802 }
803 
804 bool resctrl_arch_is_evt_configurable(enum resctrl_event_id evt)
805 {
806 	if (!rdt_cpu_has(X86_FEATURE_BMEC))
807 		return false;
808 
809 	switch (evt) {
810 	case QOS_L3_MBM_TOTAL_EVENT_ID:
811 		return rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL);
812 	case QOS_L3_MBM_LOCAL_EVENT_ID:
813 		return rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL);
814 	default:
815 		return false;
816 	}
817 }
818 
819 static __init bool get_mem_config(void)
820 {
821 	struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_MBA];
822 
823 	if (!rdt_cpu_has(X86_FEATURE_MBA))
824 		return false;
825 
826 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
827 		return __get_mem_config_intel(&hw_res->r_resctrl);
828 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
829 		return __rdt_get_mem_config_amd(&hw_res->r_resctrl);
830 
831 	return false;
832 }
833 
834 static __init bool get_slow_mem_config(void)
835 {
836 	struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_SMBA];
837 
838 	if (!rdt_cpu_has(X86_FEATURE_SMBA))
839 		return false;
840 
841 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
842 		return __rdt_get_mem_config_amd(&hw_res->r_resctrl);
843 
844 	return false;
845 }
846 
847 static __init bool get_rdt_alloc_resources(void)
848 {
849 	struct rdt_resource *r;
850 	bool ret = false;
851 
852 	if (rdt_alloc_capable)
853 		return true;
854 
855 	if (!boot_cpu_has(X86_FEATURE_RDT_A))
856 		return false;
857 
858 	if (rdt_cpu_has(X86_FEATURE_CAT_L3)) {
859 		r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
860 		rdt_get_cache_alloc_cfg(1, r);
861 		if (rdt_cpu_has(X86_FEATURE_CDP_L3))
862 			rdt_get_cdp_l3_config();
863 		if (rdt_cpu_has(X86_FEATURE_SDCIAE))
864 			rdt_set_io_alloc_capable(r);
865 		ret = true;
866 	}
867 	if (rdt_cpu_has(X86_FEATURE_CAT_L2)) {
868 		/* CPUID 0x10.2 fields are same format at 0x10.1 */
869 		r = &rdt_resources_all[RDT_RESOURCE_L2].r_resctrl;
870 		rdt_get_cache_alloc_cfg(2, r);
871 		if (rdt_cpu_has(X86_FEATURE_CDP_L2))
872 			rdt_get_cdp_l2_config();
873 		ret = true;
874 	}
875 
876 	if (get_mem_config())
877 		ret = true;
878 
879 	if (get_slow_mem_config())
880 		ret = true;
881 
882 	return ret;
883 }
884 
885 static __init bool get_rdt_mon_resources(void)
886 {
887 	struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
888 	bool ret = false;
889 
890 	if (rdt_cpu_has(X86_FEATURE_CQM_OCCUP_LLC)) {
891 		resctrl_enable_mon_event(QOS_L3_OCCUP_EVENT_ID);
892 		ret = true;
893 	}
894 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL)) {
895 		resctrl_enable_mon_event(QOS_L3_MBM_TOTAL_EVENT_ID);
896 		ret = true;
897 	}
898 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL)) {
899 		resctrl_enable_mon_event(QOS_L3_MBM_LOCAL_EVENT_ID);
900 		ret = true;
901 	}
902 	if (rdt_cpu_has(X86_FEATURE_ABMC))
903 		ret = true;
904 
905 	if (!ret)
906 		return false;
907 
908 	return !rdt_get_mon_l3_config(r);
909 }
910 
911 static __init void __check_quirks_intel(void)
912 {
913 	switch (boot_cpu_data.x86_vfm) {
914 	case INTEL_HASWELL_X:
915 		if (!rdt_options[RDT_FLAG_L3_CAT].force_off)
916 			cache_alloc_hsw_probe();
917 		break;
918 	case INTEL_SKYLAKE_X:
919 		if (boot_cpu_data.x86_stepping <= 4)
920 			set_rdt_options("!cmt,!mbmtotal,!mbmlocal,!l3cat");
921 		else
922 			set_rdt_options("!l3cat");
923 		fallthrough;
924 	case INTEL_BROADWELL_X:
925 		intel_rdt_mbm_apply_quirk();
926 		break;
927 	}
928 }
929 
930 static __init void check_quirks(void)
931 {
932 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
933 		__check_quirks_intel();
934 }
935 
936 static __init bool get_rdt_resources(void)
937 {
938 	rdt_alloc_capable = get_rdt_alloc_resources();
939 	rdt_mon_capable = get_rdt_mon_resources();
940 
941 	return (rdt_mon_capable || rdt_alloc_capable);
942 }
943 
944 static __init void rdt_init_res_defs_intel(void)
945 {
946 	struct rdt_hw_resource *hw_res;
947 	struct rdt_resource *r;
948 
949 	for_each_rdt_resource(r) {
950 		hw_res = resctrl_to_arch_res(r);
951 
952 		if (r->rid == RDT_RESOURCE_L3 ||
953 		    r->rid == RDT_RESOURCE_L2) {
954 			r->cache.arch_has_per_cpu_cfg = false;
955 			r->cache.min_cbm_bits = 1;
956 		} else if (r->rid == RDT_RESOURCE_MBA) {
957 			hw_res->msr_base = MSR_IA32_MBA_THRTL_BASE;
958 			hw_res->msr_update = mba_wrmsr_intel;
959 		}
960 	}
961 }
962 
963 static __init void rdt_init_res_defs_amd(void)
964 {
965 	struct rdt_hw_resource *hw_res;
966 	struct rdt_resource *r;
967 
968 	for_each_rdt_resource(r) {
969 		hw_res = resctrl_to_arch_res(r);
970 
971 		if (r->rid == RDT_RESOURCE_L3 ||
972 		    r->rid == RDT_RESOURCE_L2) {
973 			r->cache.arch_has_sparse_bitmasks = true;
974 			r->cache.arch_has_per_cpu_cfg = true;
975 			r->cache.min_cbm_bits = 0;
976 		} else if (r->rid == RDT_RESOURCE_MBA) {
977 			hw_res->msr_base = MSR_IA32_MBA_BW_BASE;
978 			hw_res->msr_update = mba_wrmsr_amd;
979 		} else if (r->rid == RDT_RESOURCE_SMBA) {
980 			hw_res->msr_base = MSR_IA32_SMBA_BW_BASE;
981 			hw_res->msr_update = mba_wrmsr_amd;
982 		}
983 	}
984 }
985 
986 static __init void rdt_init_res_defs(void)
987 {
988 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
989 		rdt_init_res_defs_intel();
990 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
991 		rdt_init_res_defs_amd();
992 }
993 
994 static enum cpuhp_state rdt_online;
995 
996 /* Runs once on the BSP during boot. */
997 void resctrl_cpu_detect(struct cpuinfo_x86 *c)
998 {
999 	if (!cpu_has(c, X86_FEATURE_CQM_LLC) && !cpu_has(c, X86_FEATURE_ABMC)) {
1000 		c->x86_cache_max_rmid  = -1;
1001 		c->x86_cache_occ_scale = -1;
1002 		c->x86_cache_mbm_width_offset = -1;
1003 		return;
1004 	}
1005 
1006 	/* will be overridden if occupancy monitoring exists */
1007 	c->x86_cache_max_rmid = cpuid_ebx(0xf);
1008 
1009 	if (cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC) ||
1010 	    cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL) ||
1011 	    cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL) ||
1012 	    cpu_has(c, X86_FEATURE_ABMC)) {
1013 		u32 eax, ebx, ecx, edx;
1014 
1015 		/* QoS sub-leaf, EAX=0Fh, ECX=1 */
1016 		cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx);
1017 
1018 		c->x86_cache_max_rmid  = ecx;
1019 		c->x86_cache_occ_scale = ebx;
1020 		c->x86_cache_mbm_width_offset = eax & 0xff;
1021 
1022 		if (c->x86_vendor == X86_VENDOR_AMD && !c->x86_cache_mbm_width_offset)
1023 			c->x86_cache_mbm_width_offset = MBM_CNTR_WIDTH_OFFSET_AMD;
1024 	}
1025 }
1026 
1027 static int __init resctrl_arch_late_init(void)
1028 {
1029 	struct rdt_resource *r;
1030 	int state, ret, i;
1031 
1032 	/* for_each_rdt_resource() requires all rid to be initialised. */
1033 	for (i = 0; i < RDT_NUM_RESOURCES; i++)
1034 		rdt_resources_all[i].r_resctrl.rid = i;
1035 
1036 	/*
1037 	 * Initialize functions(or definitions) that are different
1038 	 * between vendors here.
1039 	 */
1040 	rdt_init_res_defs();
1041 
1042 	check_quirks();
1043 
1044 	if (!get_rdt_resources())
1045 		return -ENODEV;
1046 
1047 	state = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
1048 				  "x86/resctrl/cat:online:",
1049 				  resctrl_arch_online_cpu,
1050 				  resctrl_arch_offline_cpu);
1051 	if (state < 0)
1052 		return state;
1053 
1054 	ret = resctrl_init();
1055 	if (ret) {
1056 		cpuhp_remove_state(state);
1057 		return ret;
1058 	}
1059 	rdt_online = state;
1060 
1061 	for_each_alloc_capable_rdt_resource(r)
1062 		pr_info("%s allocation detected\n", r->name);
1063 
1064 	for_each_mon_capable_rdt_resource(r)
1065 		pr_info("%s monitoring detected\n", r->name);
1066 
1067 	return 0;
1068 }
1069 
1070 late_initcall(resctrl_arch_late_init);
1071 
1072 static void __exit resctrl_arch_exit(void)
1073 {
1074 	cpuhp_remove_state(rdt_online);
1075 
1076 	resctrl_exit();
1077 }
1078 
1079 __exitcall(resctrl_arch_exit);
1080