xref: /linux/arch/x86/kernel/cpu/resctrl/core.c (revision 39208e73a40e0e81a5b12ddc11157c0a414df307)
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 l3_mon_domain_free(struct rdt_hw_l3_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  * l3_mon_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  * Return:	0 for success, or -ENOMEM.
409  */
410 static int l3_mon_domain_mbm_alloc(u32 num_rmid, struct rdt_hw_l3_mon_domain *hw_dom)
411 {
412 	size_t tsize = sizeof(*hw_dom->arch_mbm_states[0]);
413 	enum resctrl_event_id eventid;
414 	int idx;
415 
416 	for_each_mbm_event_id(eventid) {
417 		if (!resctrl_is_mon_event_enabled(eventid))
418 			continue;
419 		idx = MBM_STATE_IDX(eventid);
420 		hw_dom->arch_mbm_states[idx] = kcalloc(num_rmid, tsize, GFP_KERNEL);
421 		if (!hw_dom->arch_mbm_states[idx])
422 			goto cleanup;
423 	}
424 
425 	return 0;
426 cleanup:
427 	for_each_mbm_idx(idx) {
428 		kfree(hw_dom->arch_mbm_states[idx]);
429 		hw_dom->arch_mbm_states[idx] = NULL;
430 	}
431 
432 	return -ENOMEM;
433 }
434 
435 static int get_domain_id_from_scope(int cpu, enum resctrl_scope scope)
436 {
437 	switch (scope) {
438 	case RESCTRL_L2_CACHE:
439 	case RESCTRL_L3_CACHE:
440 		return get_cpu_cacheinfo_id(cpu, scope);
441 	case RESCTRL_L3_NODE:
442 		return cpu_to_node(cpu);
443 	default:
444 		break;
445 	}
446 
447 	return -EINVAL;
448 }
449 
450 static void domain_add_cpu_ctrl(int cpu, struct rdt_resource *r)
451 {
452 	int id = get_domain_id_from_scope(cpu, r->ctrl_scope);
453 	struct rdt_hw_ctrl_domain *hw_dom;
454 	struct list_head *add_pos = NULL;
455 	struct rdt_domain_hdr *hdr;
456 	struct rdt_ctrl_domain *d;
457 	int err;
458 
459 	lockdep_assert_held(&domain_list_lock);
460 
461 	if (id < 0) {
462 		pr_warn_once("Can't find control domain id for CPU:%d scope:%d for resource %s\n",
463 			     cpu, r->ctrl_scope, r->name);
464 		return;
465 	}
466 
467 	hdr = resctrl_find_domain(&r->ctrl_domains, id, &add_pos);
468 	if (hdr) {
469 		if (!domain_header_is_valid(hdr, RESCTRL_CTRL_DOMAIN, r->rid))
470 			return;
471 		d = container_of(hdr, struct rdt_ctrl_domain, hdr);
472 
473 		cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
474 		if (r->cache.arch_has_per_cpu_cfg)
475 			rdt_domain_reconfigure_cdp(r);
476 		return;
477 	}
478 
479 	hw_dom = kzalloc_node(sizeof(*hw_dom), GFP_KERNEL, cpu_to_node(cpu));
480 	if (!hw_dom)
481 		return;
482 
483 	d = &hw_dom->d_resctrl;
484 	d->hdr.id = id;
485 	d->hdr.type = RESCTRL_CTRL_DOMAIN;
486 	d->hdr.rid = r->rid;
487 	cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
488 
489 	rdt_domain_reconfigure_cdp(r);
490 
491 	if (domain_setup_ctrlval(r, d)) {
492 		ctrl_domain_free(hw_dom);
493 		return;
494 	}
495 
496 	list_add_tail_rcu(&d->hdr.list, add_pos);
497 
498 	err = resctrl_online_ctrl_domain(r, d);
499 	if (err) {
500 		list_del_rcu(&d->hdr.list);
501 		synchronize_rcu();
502 		ctrl_domain_free(hw_dom);
503 	}
504 }
505 
506 static void l3_mon_domain_setup(int cpu, int id, struct rdt_resource *r, struct list_head *add_pos)
507 {
508 	struct rdt_hw_l3_mon_domain *hw_dom;
509 	struct rdt_l3_mon_domain *d;
510 	struct cacheinfo *ci;
511 	int err;
512 
513 	hw_dom = kzalloc_node(sizeof(*hw_dom), GFP_KERNEL, cpu_to_node(cpu));
514 	if (!hw_dom)
515 		return;
516 
517 	d = &hw_dom->d_resctrl;
518 	d->hdr.id = id;
519 	d->hdr.type = RESCTRL_MON_DOMAIN;
520 	d->hdr.rid = RDT_RESOURCE_L3;
521 	ci = get_cpu_cacheinfo_level(cpu, RESCTRL_L3_CACHE);
522 	if (!ci) {
523 		pr_warn_once("Can't find L3 cache for CPU:%d resource %s\n", cpu, r->name);
524 		l3_mon_domain_free(hw_dom);
525 		return;
526 	}
527 	d->ci_id = ci->id;
528 	cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
529 
530 	arch_mon_domain_online(r, d);
531 
532 	if (l3_mon_domain_mbm_alloc(r->mon.num_rmid, hw_dom)) {
533 		l3_mon_domain_free(hw_dom);
534 		return;
535 	}
536 
537 	list_add_tail_rcu(&d->hdr.list, add_pos);
538 
539 	err = resctrl_online_mon_domain(r, &d->hdr);
540 	if (err) {
541 		list_del_rcu(&d->hdr.list);
542 		synchronize_rcu();
543 		l3_mon_domain_free(hw_dom);
544 	}
545 }
546 
547 static void domain_add_cpu_mon(int cpu, struct rdt_resource *r)
548 {
549 	int id = get_domain_id_from_scope(cpu, r->mon_scope);
550 	struct list_head *add_pos = NULL;
551 	struct rdt_domain_hdr *hdr;
552 
553 	lockdep_assert_held(&domain_list_lock);
554 
555 	if (id < 0) {
556 		pr_warn_once("Can't find monitor domain id for CPU:%d scope:%d for resource %s\n",
557 			     cpu, r->mon_scope, r->name);
558 		return;
559 	}
560 
561 	hdr = resctrl_find_domain(&r->mon_domains, id, &add_pos);
562 	if (hdr)
563 		cpumask_set_cpu(cpu, &hdr->cpu_mask);
564 
565 	switch (r->rid) {
566 	case RDT_RESOURCE_L3:
567 		/* Update the mbm_assign_mode state for the CPU if supported */
568 		if (r->mon.mbm_cntr_assignable)
569 			resctrl_arch_mbm_cntr_assign_set_one(r);
570 		if (!hdr)
571 			l3_mon_domain_setup(cpu, id, r, add_pos);
572 		break;
573 	default:
574 		pr_warn_once("Unknown resource rid=%d\n", r->rid);
575 		break;
576 	}
577 }
578 
579 static void domain_add_cpu(int cpu, struct rdt_resource *r)
580 {
581 	if (r->alloc_capable)
582 		domain_add_cpu_ctrl(cpu, r);
583 	if (r->mon_capable)
584 		domain_add_cpu_mon(cpu, r);
585 }
586 
587 static void domain_remove_cpu_ctrl(int cpu, struct rdt_resource *r)
588 {
589 	int id = get_domain_id_from_scope(cpu, r->ctrl_scope);
590 	struct rdt_hw_ctrl_domain *hw_dom;
591 	struct rdt_domain_hdr *hdr;
592 	struct rdt_ctrl_domain *d;
593 
594 	lockdep_assert_held(&domain_list_lock);
595 
596 	if (id < 0) {
597 		pr_warn_once("Can't find control domain id for CPU:%d scope:%d for resource %s\n",
598 			     cpu, r->ctrl_scope, r->name);
599 		return;
600 	}
601 
602 	hdr = resctrl_find_domain(&r->ctrl_domains, id, NULL);
603 	if (!hdr) {
604 		pr_warn("Can't find control domain for id=%d for CPU %d for resource %s\n",
605 			id, cpu, r->name);
606 		return;
607 	}
608 
609 	cpumask_clear_cpu(cpu, &hdr->cpu_mask);
610 	if (!cpumask_empty(&hdr->cpu_mask))
611 		return;
612 
613 	if (!domain_header_is_valid(hdr, RESCTRL_CTRL_DOMAIN, r->rid))
614 		return;
615 
616 	d = container_of(hdr, struct rdt_ctrl_domain, hdr);
617 	hw_dom = resctrl_to_arch_ctrl_dom(d);
618 
619 	resctrl_offline_ctrl_domain(r, d);
620 	list_del_rcu(&hdr->list);
621 	synchronize_rcu();
622 
623 	/*
624 	 * rdt_ctrl_domain "d" is going to be freed below, so clear
625 	 * its pointer from pseudo_lock_region struct.
626 	 */
627 	if (d->plr)
628 		d->plr->d = NULL;
629 	ctrl_domain_free(hw_dom);
630 }
631 
632 static void domain_remove_cpu_mon(int cpu, struct rdt_resource *r)
633 {
634 	int id = get_domain_id_from_scope(cpu, r->mon_scope);
635 	struct rdt_domain_hdr *hdr;
636 
637 	lockdep_assert_held(&domain_list_lock);
638 
639 	if (id < 0) {
640 		pr_warn_once("Can't find monitor domain id for CPU:%d scope:%d for resource %s\n",
641 			     cpu, r->mon_scope, r->name);
642 		return;
643 	}
644 
645 	hdr = resctrl_find_domain(&r->mon_domains, id, NULL);
646 	if (!hdr) {
647 		pr_warn("Can't find monitor domain for id=%d for CPU %d for resource %s\n",
648 			id, cpu, r->name);
649 		return;
650 	}
651 
652 	cpumask_clear_cpu(cpu, &hdr->cpu_mask);
653 	if (!cpumask_empty(&hdr->cpu_mask))
654 		return;
655 
656 	switch (r->rid) {
657 	case RDT_RESOURCE_L3: {
658 		struct rdt_hw_l3_mon_domain *hw_dom;
659 		struct rdt_l3_mon_domain *d;
660 
661 		if (!domain_header_is_valid(hdr, RESCTRL_MON_DOMAIN, RDT_RESOURCE_L3))
662 			return;
663 
664 		d = container_of(hdr, struct rdt_l3_mon_domain, hdr);
665 		hw_dom = resctrl_to_arch_mon_dom(d);
666 		resctrl_offline_mon_domain(r, hdr);
667 		list_del_rcu(&hdr->list);
668 		synchronize_rcu();
669 		l3_mon_domain_free(hw_dom);
670 		break;
671 	}
672 	default:
673 		pr_warn_once("Unknown resource rid=%d\n", r->rid);
674 		break;
675 	}
676 }
677 
678 static void domain_remove_cpu(int cpu, struct rdt_resource *r)
679 {
680 	if (r->alloc_capable)
681 		domain_remove_cpu_ctrl(cpu, r);
682 	if (r->mon_capable)
683 		domain_remove_cpu_mon(cpu, r);
684 }
685 
686 static void clear_closid_rmid(int cpu)
687 {
688 	struct resctrl_pqr_state *state = this_cpu_ptr(&pqr_state);
689 
690 	state->default_closid = RESCTRL_RESERVED_CLOSID;
691 	state->default_rmid = RESCTRL_RESERVED_RMID;
692 	state->cur_closid = RESCTRL_RESERVED_CLOSID;
693 	state->cur_rmid = RESCTRL_RESERVED_RMID;
694 	wrmsr(MSR_IA32_PQR_ASSOC, RESCTRL_RESERVED_RMID,
695 	      RESCTRL_RESERVED_CLOSID);
696 }
697 
698 static int resctrl_arch_online_cpu(unsigned int cpu)
699 {
700 	struct rdt_resource *r;
701 
702 	mutex_lock(&domain_list_lock);
703 	for_each_capable_rdt_resource(r)
704 		domain_add_cpu(cpu, r);
705 	mutex_unlock(&domain_list_lock);
706 
707 	clear_closid_rmid(cpu);
708 	resctrl_online_cpu(cpu);
709 
710 	return 0;
711 }
712 
713 static int resctrl_arch_offline_cpu(unsigned int cpu)
714 {
715 	struct rdt_resource *r;
716 
717 	resctrl_offline_cpu(cpu);
718 
719 	mutex_lock(&domain_list_lock);
720 	for_each_capable_rdt_resource(r)
721 		domain_remove_cpu(cpu, r);
722 	mutex_unlock(&domain_list_lock);
723 
724 	clear_closid_rmid(cpu);
725 
726 	return 0;
727 }
728 
729 void resctrl_arch_pre_mount(void)
730 {
731 }
732 
733 enum {
734 	RDT_FLAG_CMT,
735 	RDT_FLAG_MBM_TOTAL,
736 	RDT_FLAG_MBM_LOCAL,
737 	RDT_FLAG_L3_CAT,
738 	RDT_FLAG_L3_CDP,
739 	RDT_FLAG_L2_CAT,
740 	RDT_FLAG_L2_CDP,
741 	RDT_FLAG_MBA,
742 	RDT_FLAG_SMBA,
743 	RDT_FLAG_BMEC,
744 	RDT_FLAG_ABMC,
745 	RDT_FLAG_SDCIAE,
746 };
747 
748 #define RDT_OPT(idx, n, f)	\
749 [idx] = {			\
750 	.name = n,		\
751 	.flag = f		\
752 }
753 
754 struct rdt_options {
755 	char	*name;
756 	int	flag;
757 	bool	force_off, force_on;
758 };
759 
760 static struct rdt_options rdt_options[]  __ro_after_init = {
761 	RDT_OPT(RDT_FLAG_CMT,	    "cmt",	X86_FEATURE_CQM_OCCUP_LLC),
762 	RDT_OPT(RDT_FLAG_MBM_TOTAL, "mbmtotal", X86_FEATURE_CQM_MBM_TOTAL),
763 	RDT_OPT(RDT_FLAG_MBM_LOCAL, "mbmlocal", X86_FEATURE_CQM_MBM_LOCAL),
764 	RDT_OPT(RDT_FLAG_L3_CAT,    "l3cat",	X86_FEATURE_CAT_L3),
765 	RDT_OPT(RDT_FLAG_L3_CDP,    "l3cdp",	X86_FEATURE_CDP_L3),
766 	RDT_OPT(RDT_FLAG_L2_CAT,    "l2cat",	X86_FEATURE_CAT_L2),
767 	RDT_OPT(RDT_FLAG_L2_CDP,    "l2cdp",	X86_FEATURE_CDP_L2),
768 	RDT_OPT(RDT_FLAG_MBA,	    "mba",	X86_FEATURE_MBA),
769 	RDT_OPT(RDT_FLAG_SMBA,	    "smba",	X86_FEATURE_SMBA),
770 	RDT_OPT(RDT_FLAG_BMEC,	    "bmec",	X86_FEATURE_BMEC),
771 	RDT_OPT(RDT_FLAG_ABMC,	    "abmc",	X86_FEATURE_ABMC),
772 	RDT_OPT(RDT_FLAG_SDCIAE,    "sdciae",	X86_FEATURE_SDCIAE),
773 };
774 #define NUM_RDT_OPTIONS ARRAY_SIZE(rdt_options)
775 
776 static int __init set_rdt_options(char *str)
777 {
778 	struct rdt_options *o;
779 	bool force_off;
780 	char *tok;
781 
782 	if (*str == '=')
783 		str++;
784 	while ((tok = strsep(&str, ",")) != NULL) {
785 		force_off = *tok == '!';
786 		if (force_off)
787 			tok++;
788 		for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
789 			if (strcmp(tok, o->name) == 0) {
790 				if (force_off)
791 					o->force_off = true;
792 				else
793 					o->force_on = true;
794 				break;
795 			}
796 		}
797 	}
798 	return 1;
799 }
800 __setup("rdt", set_rdt_options);
801 
802 bool rdt_cpu_has(int flag)
803 {
804 	bool ret = boot_cpu_has(flag);
805 	struct rdt_options *o;
806 
807 	if (!ret)
808 		return ret;
809 
810 	for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
811 		if (flag == o->flag) {
812 			if (o->force_off)
813 				ret = false;
814 			if (o->force_on)
815 				ret = true;
816 			break;
817 		}
818 	}
819 	return ret;
820 }
821 
822 bool resctrl_arch_is_evt_configurable(enum resctrl_event_id evt)
823 {
824 	if (!rdt_cpu_has(X86_FEATURE_BMEC))
825 		return false;
826 
827 	switch (evt) {
828 	case QOS_L3_MBM_TOTAL_EVENT_ID:
829 		return rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL);
830 	case QOS_L3_MBM_LOCAL_EVENT_ID:
831 		return rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL);
832 	default:
833 		return false;
834 	}
835 }
836 
837 static __init bool get_mem_config(void)
838 {
839 	struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_MBA];
840 
841 	if (!rdt_cpu_has(X86_FEATURE_MBA))
842 		return false;
843 
844 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
845 		return __get_mem_config_intel(&hw_res->r_resctrl);
846 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
847 		return __rdt_get_mem_config_amd(&hw_res->r_resctrl);
848 
849 	return false;
850 }
851 
852 static __init bool get_slow_mem_config(void)
853 {
854 	struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_SMBA];
855 
856 	if (!rdt_cpu_has(X86_FEATURE_SMBA))
857 		return false;
858 
859 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
860 		return __rdt_get_mem_config_amd(&hw_res->r_resctrl);
861 
862 	return false;
863 }
864 
865 static __init bool get_rdt_alloc_resources(void)
866 {
867 	struct rdt_resource *r;
868 	bool ret = false;
869 
870 	if (rdt_alloc_capable)
871 		return true;
872 
873 	if (!boot_cpu_has(X86_FEATURE_RDT_A))
874 		return false;
875 
876 	if (rdt_cpu_has(X86_FEATURE_CAT_L3)) {
877 		r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
878 		rdt_get_cache_alloc_cfg(1, r);
879 		if (rdt_cpu_has(X86_FEATURE_CDP_L3))
880 			rdt_get_cdp_l3_config();
881 		if (rdt_cpu_has(X86_FEATURE_SDCIAE))
882 			rdt_set_io_alloc_capable(r);
883 		ret = true;
884 	}
885 	if (rdt_cpu_has(X86_FEATURE_CAT_L2)) {
886 		/* CPUID 0x10.2 fields are same format at 0x10.1 */
887 		r = &rdt_resources_all[RDT_RESOURCE_L2].r_resctrl;
888 		rdt_get_cache_alloc_cfg(2, r);
889 		if (rdt_cpu_has(X86_FEATURE_CDP_L2))
890 			rdt_get_cdp_l2_config();
891 		ret = true;
892 	}
893 
894 	if (get_mem_config())
895 		ret = true;
896 
897 	if (get_slow_mem_config())
898 		ret = true;
899 
900 	return ret;
901 }
902 
903 static __init bool get_rdt_mon_resources(void)
904 {
905 	struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
906 	bool ret = false;
907 
908 	if (rdt_cpu_has(X86_FEATURE_CQM_OCCUP_LLC)) {
909 		resctrl_enable_mon_event(QOS_L3_OCCUP_EVENT_ID, false, 0);
910 		ret = true;
911 	}
912 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL)) {
913 		resctrl_enable_mon_event(QOS_L3_MBM_TOTAL_EVENT_ID, false, 0);
914 		ret = true;
915 	}
916 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL)) {
917 		resctrl_enable_mon_event(QOS_L3_MBM_LOCAL_EVENT_ID, false, 0);
918 		ret = true;
919 	}
920 	if (rdt_cpu_has(X86_FEATURE_ABMC))
921 		ret = true;
922 
923 	if (!ret)
924 		return false;
925 
926 	return !rdt_get_l3_mon_config(r);
927 }
928 
929 static __init void __check_quirks_intel(void)
930 {
931 	switch (boot_cpu_data.x86_vfm) {
932 	case INTEL_HASWELL_X:
933 		if (!rdt_options[RDT_FLAG_L3_CAT].force_off)
934 			cache_alloc_hsw_probe();
935 		break;
936 	case INTEL_SKYLAKE_X:
937 		if (boot_cpu_data.x86_stepping <= 4)
938 			set_rdt_options("!cmt,!mbmtotal,!mbmlocal,!l3cat");
939 		else
940 			set_rdt_options("!l3cat");
941 		fallthrough;
942 	case INTEL_BROADWELL_X:
943 		intel_rdt_mbm_apply_quirk();
944 		break;
945 	}
946 }
947 
948 static __init void check_quirks(void)
949 {
950 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
951 		__check_quirks_intel();
952 }
953 
954 static __init bool get_rdt_resources(void)
955 {
956 	rdt_alloc_capable = get_rdt_alloc_resources();
957 	rdt_mon_capable = get_rdt_mon_resources();
958 
959 	return (rdt_mon_capable || rdt_alloc_capable);
960 }
961 
962 static __init void rdt_init_res_defs_intel(void)
963 {
964 	struct rdt_hw_resource *hw_res;
965 	struct rdt_resource *r;
966 
967 	for_each_rdt_resource(r) {
968 		hw_res = resctrl_to_arch_res(r);
969 
970 		if (r->rid == RDT_RESOURCE_L3 ||
971 		    r->rid == RDT_RESOURCE_L2) {
972 			r->cache.arch_has_per_cpu_cfg = false;
973 			r->cache.min_cbm_bits = 1;
974 		} else if (r->rid == RDT_RESOURCE_MBA) {
975 			hw_res->msr_base = MSR_IA32_MBA_THRTL_BASE;
976 			hw_res->msr_update = mba_wrmsr_intel;
977 		}
978 	}
979 }
980 
981 static __init void rdt_init_res_defs_amd(void)
982 {
983 	struct rdt_hw_resource *hw_res;
984 	struct rdt_resource *r;
985 
986 	for_each_rdt_resource(r) {
987 		hw_res = resctrl_to_arch_res(r);
988 
989 		if (r->rid == RDT_RESOURCE_L3 ||
990 		    r->rid == RDT_RESOURCE_L2) {
991 			r->cache.arch_has_sparse_bitmasks = true;
992 			r->cache.arch_has_per_cpu_cfg = true;
993 			r->cache.min_cbm_bits = 0;
994 		} else if (r->rid == RDT_RESOURCE_MBA) {
995 			hw_res->msr_base = MSR_IA32_MBA_BW_BASE;
996 			hw_res->msr_update = mba_wrmsr_amd;
997 		} else if (r->rid == RDT_RESOURCE_SMBA) {
998 			hw_res->msr_base = MSR_IA32_SMBA_BW_BASE;
999 			hw_res->msr_update = mba_wrmsr_amd;
1000 		}
1001 	}
1002 }
1003 
1004 static __init void rdt_init_res_defs(void)
1005 {
1006 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
1007 		rdt_init_res_defs_intel();
1008 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
1009 		rdt_init_res_defs_amd();
1010 }
1011 
1012 static enum cpuhp_state rdt_online;
1013 
1014 /* Runs once on the BSP during boot. */
1015 void resctrl_cpu_detect(struct cpuinfo_x86 *c)
1016 {
1017 	if (!cpu_has(c, X86_FEATURE_CQM_LLC) && !cpu_has(c, X86_FEATURE_ABMC)) {
1018 		c->x86_cache_max_rmid  = -1;
1019 		c->x86_cache_occ_scale = -1;
1020 		c->x86_cache_mbm_width_offset = -1;
1021 		return;
1022 	}
1023 
1024 	/* will be overridden if occupancy monitoring exists */
1025 	c->x86_cache_max_rmid = cpuid_ebx(0xf);
1026 
1027 	if (cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC) ||
1028 	    cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL) ||
1029 	    cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL) ||
1030 	    cpu_has(c, X86_FEATURE_ABMC)) {
1031 		u32 eax, ebx, ecx, edx;
1032 
1033 		/* QoS sub-leaf, EAX=0Fh, ECX=1 */
1034 		cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx);
1035 
1036 		c->x86_cache_max_rmid  = ecx;
1037 		c->x86_cache_occ_scale = ebx;
1038 		c->x86_cache_mbm_width_offset = eax & 0xff;
1039 
1040 		if (c->x86_vendor == X86_VENDOR_AMD && !c->x86_cache_mbm_width_offset)
1041 			c->x86_cache_mbm_width_offset = MBM_CNTR_WIDTH_OFFSET_AMD;
1042 	}
1043 }
1044 
1045 static int __init resctrl_arch_late_init(void)
1046 {
1047 	struct rdt_resource *r;
1048 	int state, ret, i;
1049 
1050 	/* for_each_rdt_resource() requires all rid to be initialised. */
1051 	for (i = 0; i < RDT_NUM_RESOURCES; i++)
1052 		rdt_resources_all[i].r_resctrl.rid = i;
1053 
1054 	/*
1055 	 * Initialize functions(or definitions) that are different
1056 	 * between vendors here.
1057 	 */
1058 	rdt_init_res_defs();
1059 
1060 	check_quirks();
1061 
1062 	if (!get_rdt_resources())
1063 		return -ENODEV;
1064 
1065 	state = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
1066 				  "x86/resctrl/cat:online:",
1067 				  resctrl_arch_online_cpu,
1068 				  resctrl_arch_offline_cpu);
1069 	if (state < 0)
1070 		return state;
1071 
1072 	ret = resctrl_init();
1073 	if (ret) {
1074 		cpuhp_remove_state(state);
1075 		return ret;
1076 	}
1077 	rdt_online = state;
1078 
1079 	for_each_alloc_capable_rdt_resource(r)
1080 		pr_info("%s allocation detected\n", r->name);
1081 
1082 	for_each_mon_capable_rdt_resource(r)
1083 		pr_info("%s monitoring detected\n", r->name);
1084 
1085 	return 0;
1086 }
1087 
1088 late_initcall(resctrl_arch_late_init);
1089 
1090 static void __exit resctrl_arch_exit(void)
1091 {
1092 	cpuhp_remove_state(rdt_online);
1093 
1094 	resctrl_exit();
1095 }
1096 
1097 __exitcall(resctrl_arch_exit);
1098