xref: /openbmc/linux/arch/x86/kernel/cpu/resctrl/core.c (revision 2bdd5238)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Resource Director Technology(RDT)
4  * - Cache Allocation code.
5  *
6  * Copyright (C) 2016 Intel Corporation
7  *
8  * Authors:
9  *    Fenghua Yu <fenghua.yu@intel.com>
10  *    Tony Luck <tony.luck@intel.com>
11  *    Vikas Shivappa <vikas.shivappa@intel.com>
12  *
13  * More information about RDT be found in the Intel (R) x86 Architecture
14  * Software Developer Manual June 2016, volume 3, section 17.17.
15  */
16 
17 #define pr_fmt(fmt)	"resctrl: " fmt
18 
19 #include <linux/slab.h>
20 #include <linux/err.h>
21 #include <linux/cacheinfo.h>
22 #include <linux/cpuhotplug.h>
23 
24 #include <asm/intel-family.h>
25 #include <asm/resctrl.h>
26 #include "internal.h"
27 
28 /* Mutex to protect rdtgroup access. */
29 DEFINE_MUTEX(rdtgroup_mutex);
30 
31 /*
32  * The cached resctrl_pqr_state is strictly per CPU and can never be
33  * updated from a remote CPU. Functions which modify the state
34  * are called with interrupts disabled and no preemption, which
35  * is sufficient for the protection.
36  */
37 DEFINE_PER_CPU(struct resctrl_pqr_state, pqr_state);
38 
39 /*
40  * Used to store the max resource name width and max resource data width
41  * to display the schemata in a tabular format
42  */
43 int max_name_width, max_data_width;
44 
45 /*
46  * Global boolean for rdt_alloc which is true if any
47  * resource allocation is enabled.
48  */
49 bool rdt_alloc_capable;
50 
51 static void
52 mba_wrmsr_intel(struct rdt_domain *d, struct msr_param *m,
53 		struct rdt_resource *r);
54 static void
55 cat_wrmsr(struct rdt_domain *d, struct msr_param *m, struct rdt_resource *r);
56 static void
57 mba_wrmsr_amd(struct rdt_domain *d, struct msr_param *m,
58 	      struct rdt_resource *r);
59 
60 #define domain_init(id) LIST_HEAD_INIT(rdt_resources_all[id].r_resctrl.domains)
61 
62 struct rdt_hw_resource rdt_resources_all[] = {
63 	[RDT_RESOURCE_L3] =
64 	{
65 		.r_resctrl = {
66 			.rid			= RDT_RESOURCE_L3,
67 			.name			= "L3",
68 			.cache_level		= 3,
69 			.cache = {
70 				.min_cbm_bits	= 1,
71 			},
72 			.domains		= domain_init(RDT_RESOURCE_L3),
73 			.parse_ctrlval		= parse_cbm,
74 			.format_str		= "%d=%0*x",
75 			.fflags			= RFTYPE_RES_CACHE,
76 		},
77 		.msr_base		= MSR_IA32_L3_CBM_BASE,
78 		.msr_update		= cat_wrmsr,
79 	},
80 	[RDT_RESOURCE_L2] =
81 	{
82 		.r_resctrl = {
83 			.rid			= RDT_RESOURCE_L2,
84 			.name			= "L2",
85 			.cache_level		= 2,
86 			.cache = {
87 				.min_cbm_bits	= 1,
88 			},
89 			.domains		= domain_init(RDT_RESOURCE_L2),
90 			.parse_ctrlval		= parse_cbm,
91 			.format_str		= "%d=%0*x",
92 			.fflags			= RFTYPE_RES_CACHE,
93 		},
94 		.msr_base		= MSR_IA32_L2_CBM_BASE,
95 		.msr_update		= cat_wrmsr,
96 	},
97 	[RDT_RESOURCE_MBA] =
98 	{
99 		.r_resctrl = {
100 			.rid			= RDT_RESOURCE_MBA,
101 			.name			= "MB",
102 			.cache_level		= 3,
103 			.domains		= domain_init(RDT_RESOURCE_MBA),
104 			.parse_ctrlval		= parse_bw,
105 			.format_str		= "%d=%*u",
106 			.fflags			= RFTYPE_RES_MB,
107 		},
108 	},
109 };
110 
111 /*
112  * cache_alloc_hsw_probe() - Have to probe for Intel haswell server CPUs
113  * as they do not have CPUID enumeration support for Cache allocation.
114  * The check for Vendor/Family/Model is not enough to guarantee that
115  * the MSRs won't #GP fault because only the following SKUs support
116  * CAT:
117  *	Intel(R) Xeon(R)  CPU E5-2658  v3  @  2.20GHz
118  *	Intel(R) Xeon(R)  CPU E5-2648L v3  @  1.80GHz
119  *	Intel(R) Xeon(R)  CPU E5-2628L v3  @  2.00GHz
120  *	Intel(R) Xeon(R)  CPU E5-2618L v3  @  2.30GHz
121  *	Intel(R) Xeon(R)  CPU E5-2608L v3  @  2.00GHz
122  *	Intel(R) Xeon(R)  CPU E5-2658A v3  @  2.20GHz
123  *
124  * Probe by trying to write the first of the L3 cache mask registers
125  * and checking that the bits stick. Max CLOSids is always 4 and max cbm length
126  * is always 20 on hsw server parts. The minimum cache bitmask length
127  * allowed for HSW server is always 2 bits. Hardcode all of them.
128  */
129 static inline void cache_alloc_hsw_probe(void)
130 {
131 	struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_L3];
132 	struct rdt_resource *r  = &hw_res->r_resctrl;
133 	u32 l, h, max_cbm = BIT_MASK(20) - 1;
134 
135 	if (wrmsr_safe(MSR_IA32_L3_CBM_BASE, max_cbm, 0))
136 		return;
137 
138 	rdmsr(MSR_IA32_L3_CBM_BASE, l, h);
139 
140 	/* If all the bits were set in MSR, return success */
141 	if (l != max_cbm)
142 		return;
143 
144 	hw_res->num_closid = 4;
145 	r->default_ctrl = max_cbm;
146 	r->cache.cbm_len = 20;
147 	r->cache.shareable_bits = 0xc0000;
148 	r->cache.min_cbm_bits = 2;
149 	r->alloc_capable = true;
150 	r->alloc_enabled = true;
151 
152 	rdt_alloc_capable = true;
153 }
154 
155 bool is_mba_sc(struct rdt_resource *r)
156 {
157 	if (!r)
158 		return rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl.membw.mba_sc;
159 
160 	return r->membw.mba_sc;
161 }
162 
163 /*
164  * rdt_get_mb_table() - get a mapping of bandwidth(b/w) percentage values
165  * exposed to user interface and the h/w understandable delay values.
166  *
167  * The non-linear delay values have the granularity of power of two
168  * and also the h/w does not guarantee a curve for configured delay
169  * values vs. actual b/w enforced.
170  * Hence we need a mapping that is pre calibrated so the user can
171  * express the memory b/w as a percentage value.
172  */
173 static inline bool rdt_get_mb_table(struct rdt_resource *r)
174 {
175 	/*
176 	 * There are no Intel SKUs as of now to support non-linear delay.
177 	 */
178 	pr_info("MBA b/w map not implemented for cpu:%d, model:%d",
179 		boot_cpu_data.x86, boot_cpu_data.x86_model);
180 
181 	return false;
182 }
183 
184 static bool __get_mem_config_intel(struct rdt_resource *r)
185 {
186 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
187 	union cpuid_0x10_3_eax eax;
188 	union cpuid_0x10_x_edx edx;
189 	u32 ebx, ecx, max_delay;
190 
191 	cpuid_count(0x00000010, 3, &eax.full, &ebx, &ecx, &edx.full);
192 	hw_res->num_closid = edx.split.cos_max + 1;
193 	max_delay = eax.split.max_delay + 1;
194 	r->default_ctrl = MAX_MBA_BW;
195 	r->membw.arch_needs_linear = true;
196 	if (ecx & MBA_IS_LINEAR) {
197 		r->membw.delay_linear = true;
198 		r->membw.min_bw = MAX_MBA_BW - max_delay;
199 		r->membw.bw_gran = MAX_MBA_BW - max_delay;
200 	} else {
201 		if (!rdt_get_mb_table(r))
202 			return false;
203 		r->membw.arch_needs_linear = false;
204 	}
205 	r->data_width = 3;
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 	thread_throttle_mode_init();
212 
213 	r->alloc_capable = true;
214 	r->alloc_enabled = true;
215 
216 	return true;
217 }
218 
219 static bool __rdt_get_mem_config_amd(struct rdt_resource *r)
220 {
221 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
222 	union cpuid_0x10_3_eax eax;
223 	union cpuid_0x10_x_edx edx;
224 	u32 ebx, ecx;
225 
226 	cpuid_count(0x80000020, 1, &eax.full, &ebx, &ecx, &edx.full);
227 	hw_res->num_closid = edx.split.cos_max + 1;
228 	r->default_ctrl = MAX_MBA_BW_AMD;
229 
230 	/* AMD does not use delay */
231 	r->membw.delay_linear = false;
232 	r->membw.arch_needs_linear = false;
233 
234 	/*
235 	 * AMD does not use memory delay throttle model to control
236 	 * the allocation like Intel does.
237 	 */
238 	r->membw.throttle_mode = THREAD_THROTTLE_UNDEFINED;
239 	r->membw.min_bw = 0;
240 	r->membw.bw_gran = 1;
241 	/* Max value is 2048, Data width should be 4 in decimal */
242 	r->data_width = 4;
243 
244 	r->alloc_capable = true;
245 	r->alloc_enabled = true;
246 
247 	return true;
248 }
249 
250 static void rdt_get_cache_alloc_cfg(int idx, struct rdt_resource *r)
251 {
252 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
253 	union cpuid_0x10_1_eax eax;
254 	union cpuid_0x10_x_edx edx;
255 	u32 ebx, ecx;
256 
257 	cpuid_count(0x00000010, idx, &eax.full, &ebx, &ecx, &edx.full);
258 	hw_res->num_closid = edx.split.cos_max + 1;
259 	r->cache.cbm_len = eax.split.cbm_len + 1;
260 	r->default_ctrl = BIT_MASK(eax.split.cbm_len + 1) - 1;
261 	r->cache.shareable_bits = ebx & r->default_ctrl;
262 	r->data_width = (r->cache.cbm_len + 3) / 4;
263 	r->alloc_capable = true;
264 	r->alloc_enabled = 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
288 mba_wrmsr_amd(struct rdt_domain *d, struct msr_param *m, struct rdt_resource *r)
289 {
290 	unsigned int i;
291 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
292 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
293 
294 	for (i = m->low; i < m->high; i++)
295 		wrmsrl(hw_res->msr_base + i, hw_dom->ctrl_val[i]);
296 }
297 
298 /*
299  * Map the memory b/w percentage value to delay values
300  * that can be written to QOS_MSRs.
301  * There are currently no SKUs which support non linear delay values.
302  */
303 u32 delay_bw_map(unsigned long bw, struct rdt_resource *r)
304 {
305 	if (r->membw.delay_linear)
306 		return MAX_MBA_BW - bw;
307 
308 	pr_warn_once("Non Linear delay-bw map not supported but queried\n");
309 	return r->default_ctrl;
310 }
311 
312 static void
313 mba_wrmsr_intel(struct rdt_domain *d, struct msr_param *m,
314 		struct rdt_resource *r)
315 {
316 	unsigned int i;
317 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
318 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
319 
320 	/*  Write the delay values for mba. */
321 	for (i = m->low; i < m->high; i++)
322 		wrmsrl(hw_res->msr_base + i, delay_bw_map(hw_dom->ctrl_val[i], r));
323 }
324 
325 static void
326 cat_wrmsr(struct rdt_domain *d, struct msr_param *m, struct rdt_resource *r)
327 {
328 	unsigned int i;
329 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
330 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
331 
332 	for (i = m->low; i < m->high; i++)
333 		wrmsrl(hw_res->msr_base + i, hw_dom->ctrl_val[i]);
334 }
335 
336 struct rdt_domain *get_domain_from_cpu(int cpu, struct rdt_resource *r)
337 {
338 	struct rdt_domain *d;
339 
340 	list_for_each_entry(d, &r->domains, list) {
341 		/* Find the domain that contains this CPU */
342 		if (cpumask_test_cpu(cpu, &d->cpu_mask))
343 			return d;
344 	}
345 
346 	return NULL;
347 }
348 
349 u32 resctrl_arch_get_num_closid(struct rdt_resource *r)
350 {
351 	return resctrl_to_arch_res(r)->num_closid;
352 }
353 
354 void rdt_ctrl_update(void *arg)
355 {
356 	struct msr_param *m = arg;
357 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(m->res);
358 	struct rdt_resource *r = m->res;
359 	int cpu = smp_processor_id();
360 	struct rdt_domain *d;
361 
362 	d = get_domain_from_cpu(cpu, r);
363 	if (d) {
364 		hw_res->msr_update(d, m, r);
365 		return;
366 	}
367 	pr_warn_once("cpu %d not found in any domain for resource %s\n",
368 		     cpu, r->name);
369 }
370 
371 /*
372  * rdt_find_domain - Find a domain in a resource that matches input resource id
373  *
374  * Search resource r's domain list to find the resource id. If the resource
375  * id is found in a domain, return the domain. Otherwise, if requested by
376  * caller, return the first domain whose id is bigger than the input id.
377  * The domain list is sorted by id in ascending order.
378  */
379 struct rdt_domain *rdt_find_domain(struct rdt_resource *r, int id,
380 				   struct list_head **pos)
381 {
382 	struct rdt_domain *d;
383 	struct list_head *l;
384 
385 	if (id < 0)
386 		return ERR_PTR(-ENODEV);
387 
388 	list_for_each(l, &r->domains) {
389 		d = list_entry(l, struct rdt_domain, list);
390 		/* When id is found, return its domain. */
391 		if (id == d->id)
392 			return d;
393 		/* Stop searching when finding id's position in sorted list. */
394 		if (id < d->id)
395 			break;
396 	}
397 
398 	if (pos)
399 		*pos = l;
400 
401 	return NULL;
402 }
403 
404 void setup_default_ctrlval(struct rdt_resource *r, u32 *dc, u32 *dm)
405 {
406 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
407 	int i;
408 
409 	/*
410 	 * Initialize the Control MSRs to having no control.
411 	 * For Cache Allocation: Set all bits in cbm
412 	 * For Memory Allocation: Set b/w requested to 100%
413 	 * and the bandwidth in MBps to U32_MAX
414 	 */
415 	for (i = 0; i < hw_res->num_closid; i++, dc++, dm++) {
416 		*dc = r->default_ctrl;
417 		*dm = MBA_MAX_MBPS;
418 	}
419 }
420 
421 static int domain_setup_ctrlval(struct rdt_resource *r, struct rdt_domain *d)
422 {
423 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
424 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
425 	struct msr_param m;
426 	u32 *dc, *dm;
427 
428 	dc = kmalloc_array(hw_res->num_closid, sizeof(*hw_dom->ctrl_val),
429 			   GFP_KERNEL);
430 	if (!dc)
431 		return -ENOMEM;
432 
433 	dm = kmalloc_array(hw_res->num_closid, sizeof(*hw_dom->mbps_val),
434 			   GFP_KERNEL);
435 	if (!dm) {
436 		kfree(dc);
437 		return -ENOMEM;
438 	}
439 
440 	hw_dom->ctrl_val = dc;
441 	hw_dom->mbps_val = dm;
442 	setup_default_ctrlval(r, dc, dm);
443 
444 	m.low = 0;
445 	m.high = hw_res->num_closid;
446 	hw_res->msr_update(d, &m, r);
447 	return 0;
448 }
449 
450 static int domain_setup_mon_state(struct rdt_resource *r, struct rdt_domain *d)
451 {
452 	size_t tsize;
453 
454 	if (is_llc_occupancy_enabled()) {
455 		d->rmid_busy_llc = bitmap_zalloc(r->num_rmid, GFP_KERNEL);
456 		if (!d->rmid_busy_llc)
457 			return -ENOMEM;
458 		INIT_DELAYED_WORK(&d->cqm_limbo, cqm_handle_limbo);
459 	}
460 	if (is_mbm_total_enabled()) {
461 		tsize = sizeof(*d->mbm_total);
462 		d->mbm_total = kcalloc(r->num_rmid, tsize, GFP_KERNEL);
463 		if (!d->mbm_total) {
464 			bitmap_free(d->rmid_busy_llc);
465 			return -ENOMEM;
466 		}
467 	}
468 	if (is_mbm_local_enabled()) {
469 		tsize = sizeof(*d->mbm_local);
470 		d->mbm_local = kcalloc(r->num_rmid, tsize, GFP_KERNEL);
471 		if (!d->mbm_local) {
472 			bitmap_free(d->rmid_busy_llc);
473 			kfree(d->mbm_total);
474 			return -ENOMEM;
475 		}
476 	}
477 
478 	if (is_mbm_enabled()) {
479 		INIT_DELAYED_WORK(&d->mbm_over, mbm_handle_overflow);
480 		mbm_setup_overflow_handler(d, MBM_OVERFLOW_INTERVAL);
481 	}
482 
483 	return 0;
484 }
485 
486 /*
487  * domain_add_cpu - Add a cpu to a resource's domain list.
488  *
489  * If an existing domain in the resource r's domain list matches the cpu's
490  * resource id, add the cpu in the domain.
491  *
492  * Otherwise, a new domain is allocated and inserted into the right position
493  * in the domain list sorted by id in ascending order.
494  *
495  * The order in the domain list is visible to users when we print entries
496  * in the schemata file and schemata input is validated to have the same order
497  * as this list.
498  */
499 static void domain_add_cpu(int cpu, struct rdt_resource *r)
500 {
501 	int id = get_cpu_cacheinfo_id(cpu, r->cache_level);
502 	struct list_head *add_pos = NULL;
503 	struct rdt_hw_domain *hw_dom;
504 	struct rdt_domain *d;
505 
506 	d = rdt_find_domain(r, id, &add_pos);
507 	if (IS_ERR(d)) {
508 		pr_warn("Couldn't find cache id for CPU %d\n", cpu);
509 		return;
510 	}
511 
512 	if (d) {
513 		cpumask_set_cpu(cpu, &d->cpu_mask);
514 		if (r->cache.arch_has_per_cpu_cfg)
515 			rdt_domain_reconfigure_cdp(r);
516 		return;
517 	}
518 
519 	hw_dom = kzalloc_node(sizeof(*hw_dom), GFP_KERNEL, cpu_to_node(cpu));
520 	if (!hw_dom)
521 		return;
522 
523 	d = &hw_dom->d_resctrl;
524 	d->id = id;
525 	cpumask_set_cpu(cpu, &d->cpu_mask);
526 
527 	rdt_domain_reconfigure_cdp(r);
528 
529 	if (r->alloc_capable && domain_setup_ctrlval(r, d)) {
530 		kfree(d);
531 		return;
532 	}
533 
534 	if (r->mon_capable && domain_setup_mon_state(r, d)) {
535 		kfree(d);
536 		return;
537 	}
538 
539 	list_add_tail(&d->list, add_pos);
540 
541 	/*
542 	 * If resctrl is mounted, add
543 	 * per domain monitor data directories.
544 	 */
545 	if (static_branch_unlikely(&rdt_mon_enable_key))
546 		mkdir_mondata_subdir_allrdtgrp(r, d);
547 }
548 
549 static void domain_remove_cpu(int cpu, struct rdt_resource *r)
550 {
551 	int id = get_cpu_cacheinfo_id(cpu, r->cache_level);
552 	struct rdt_hw_domain *hw_dom;
553 	struct rdt_domain *d;
554 
555 	d = rdt_find_domain(r, id, NULL);
556 	if (IS_ERR_OR_NULL(d)) {
557 		pr_warn("Couldn't find cache id for CPU %d\n", cpu);
558 		return;
559 	}
560 	hw_dom = resctrl_to_arch_dom(d);
561 
562 	cpumask_clear_cpu(cpu, &d->cpu_mask);
563 	if (cpumask_empty(&d->cpu_mask)) {
564 		/*
565 		 * If resctrl is mounted, remove all the
566 		 * per domain monitor data directories.
567 		 */
568 		if (static_branch_unlikely(&rdt_mon_enable_key))
569 			rmdir_mondata_subdir_allrdtgrp(r, d->id);
570 		list_del(&d->list);
571 		if (r->mon_capable && is_mbm_enabled())
572 			cancel_delayed_work(&d->mbm_over);
573 		if (is_llc_occupancy_enabled() &&  has_busy_rmid(r, d)) {
574 			/*
575 			 * When a package is going down, forcefully
576 			 * decrement rmid->ebusy. There is no way to know
577 			 * that the L3 was flushed and hence may lead to
578 			 * incorrect counts in rare scenarios, but leaving
579 			 * the RMID as busy creates RMID leaks if the
580 			 * package never comes back.
581 			 */
582 			__check_limbo(d, true);
583 			cancel_delayed_work(&d->cqm_limbo);
584 		}
585 
586 		/*
587 		 * rdt_domain "d" is going to be freed below, so clear
588 		 * its pointer from pseudo_lock_region struct.
589 		 */
590 		if (d->plr)
591 			d->plr->d = NULL;
592 
593 		kfree(hw_dom->ctrl_val);
594 		kfree(hw_dom->mbps_val);
595 		bitmap_free(d->rmid_busy_llc);
596 		kfree(d->mbm_total);
597 		kfree(d->mbm_local);
598 		kfree(hw_dom);
599 		return;
600 	}
601 
602 	if (r == &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl) {
603 		if (is_mbm_enabled() && cpu == d->mbm_work_cpu) {
604 			cancel_delayed_work(&d->mbm_over);
605 			mbm_setup_overflow_handler(d, 0);
606 		}
607 		if (is_llc_occupancy_enabled() && cpu == d->cqm_work_cpu &&
608 		    has_busy_rmid(r, d)) {
609 			cancel_delayed_work(&d->cqm_limbo);
610 			cqm_setup_limbo_handler(d, 0);
611 		}
612 	}
613 }
614 
615 static void clear_closid_rmid(int cpu)
616 {
617 	struct resctrl_pqr_state *state = this_cpu_ptr(&pqr_state);
618 
619 	state->default_closid = 0;
620 	state->default_rmid = 0;
621 	state->cur_closid = 0;
622 	state->cur_rmid = 0;
623 	wrmsr(IA32_PQR_ASSOC, 0, 0);
624 }
625 
626 static int resctrl_online_cpu(unsigned int cpu)
627 {
628 	struct rdt_resource *r;
629 
630 	mutex_lock(&rdtgroup_mutex);
631 	for_each_capable_rdt_resource(r)
632 		domain_add_cpu(cpu, r);
633 	/* The cpu is set in default rdtgroup after online. */
634 	cpumask_set_cpu(cpu, &rdtgroup_default.cpu_mask);
635 	clear_closid_rmid(cpu);
636 	mutex_unlock(&rdtgroup_mutex);
637 
638 	return 0;
639 }
640 
641 static void clear_childcpus(struct rdtgroup *r, unsigned int cpu)
642 {
643 	struct rdtgroup *cr;
644 
645 	list_for_each_entry(cr, &r->mon.crdtgrp_list, mon.crdtgrp_list) {
646 		if (cpumask_test_and_clear_cpu(cpu, &cr->cpu_mask)) {
647 			break;
648 		}
649 	}
650 }
651 
652 static int resctrl_offline_cpu(unsigned int cpu)
653 {
654 	struct rdtgroup *rdtgrp;
655 	struct rdt_resource *r;
656 
657 	mutex_lock(&rdtgroup_mutex);
658 	for_each_capable_rdt_resource(r)
659 		domain_remove_cpu(cpu, r);
660 	list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
661 		if (cpumask_test_and_clear_cpu(cpu, &rdtgrp->cpu_mask)) {
662 			clear_childcpus(rdtgrp, cpu);
663 			break;
664 		}
665 	}
666 	clear_closid_rmid(cpu);
667 	mutex_unlock(&rdtgroup_mutex);
668 
669 	return 0;
670 }
671 
672 /*
673  * Choose a width for the resource name and resource data based on the
674  * resource that has widest name and cbm.
675  */
676 static __init void rdt_init_padding(void)
677 {
678 	struct rdt_resource *r;
679 
680 	for_each_alloc_capable_rdt_resource(r) {
681 		if (r->data_width > max_data_width)
682 			max_data_width = r->data_width;
683 	}
684 }
685 
686 enum {
687 	RDT_FLAG_CMT,
688 	RDT_FLAG_MBM_TOTAL,
689 	RDT_FLAG_MBM_LOCAL,
690 	RDT_FLAG_L3_CAT,
691 	RDT_FLAG_L3_CDP,
692 	RDT_FLAG_L2_CAT,
693 	RDT_FLAG_L2_CDP,
694 	RDT_FLAG_MBA,
695 };
696 
697 #define RDT_OPT(idx, n, f)	\
698 [idx] = {			\
699 	.name = n,		\
700 	.flag = f		\
701 }
702 
703 struct rdt_options {
704 	char	*name;
705 	int	flag;
706 	bool	force_off, force_on;
707 };
708 
709 static struct rdt_options rdt_options[]  __initdata = {
710 	RDT_OPT(RDT_FLAG_CMT,	    "cmt",	X86_FEATURE_CQM_OCCUP_LLC),
711 	RDT_OPT(RDT_FLAG_MBM_TOTAL, "mbmtotal", X86_FEATURE_CQM_MBM_TOTAL),
712 	RDT_OPT(RDT_FLAG_MBM_LOCAL, "mbmlocal", X86_FEATURE_CQM_MBM_LOCAL),
713 	RDT_OPT(RDT_FLAG_L3_CAT,    "l3cat",	X86_FEATURE_CAT_L3),
714 	RDT_OPT(RDT_FLAG_L3_CDP,    "l3cdp",	X86_FEATURE_CDP_L3),
715 	RDT_OPT(RDT_FLAG_L2_CAT,    "l2cat",	X86_FEATURE_CAT_L2),
716 	RDT_OPT(RDT_FLAG_L2_CDP,    "l2cdp",	X86_FEATURE_CDP_L2),
717 	RDT_OPT(RDT_FLAG_MBA,	    "mba",	X86_FEATURE_MBA),
718 };
719 #define NUM_RDT_OPTIONS ARRAY_SIZE(rdt_options)
720 
721 static int __init set_rdt_options(char *str)
722 {
723 	struct rdt_options *o;
724 	bool force_off;
725 	char *tok;
726 
727 	if (*str == '=')
728 		str++;
729 	while ((tok = strsep(&str, ",")) != NULL) {
730 		force_off = *tok == '!';
731 		if (force_off)
732 			tok++;
733 		for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
734 			if (strcmp(tok, o->name) == 0) {
735 				if (force_off)
736 					o->force_off = true;
737 				else
738 					o->force_on = true;
739 				break;
740 			}
741 		}
742 	}
743 	return 1;
744 }
745 __setup("rdt", set_rdt_options);
746 
747 static bool __init rdt_cpu_has(int flag)
748 {
749 	bool ret = boot_cpu_has(flag);
750 	struct rdt_options *o;
751 
752 	if (!ret)
753 		return ret;
754 
755 	for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
756 		if (flag == o->flag) {
757 			if (o->force_off)
758 				ret = false;
759 			if (o->force_on)
760 				ret = true;
761 			break;
762 		}
763 	}
764 	return ret;
765 }
766 
767 static __init bool get_mem_config(void)
768 {
769 	struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_MBA];
770 
771 	if (!rdt_cpu_has(X86_FEATURE_MBA))
772 		return false;
773 
774 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
775 		return __get_mem_config_intel(&hw_res->r_resctrl);
776 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
777 		return __rdt_get_mem_config_amd(&hw_res->r_resctrl);
778 
779 	return false;
780 }
781 
782 static __init bool get_rdt_alloc_resources(void)
783 {
784 	struct rdt_resource *r;
785 	bool ret = false;
786 
787 	if (rdt_alloc_capable)
788 		return true;
789 
790 	if (!boot_cpu_has(X86_FEATURE_RDT_A))
791 		return false;
792 
793 	if (rdt_cpu_has(X86_FEATURE_CAT_L3)) {
794 		r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
795 		rdt_get_cache_alloc_cfg(1, r);
796 		if (rdt_cpu_has(X86_FEATURE_CDP_L3))
797 			rdt_get_cdp_l3_config();
798 		ret = true;
799 	}
800 	if (rdt_cpu_has(X86_FEATURE_CAT_L2)) {
801 		/* CPUID 0x10.2 fields are same format at 0x10.1 */
802 		r = &rdt_resources_all[RDT_RESOURCE_L2].r_resctrl;
803 		rdt_get_cache_alloc_cfg(2, r);
804 		if (rdt_cpu_has(X86_FEATURE_CDP_L2))
805 			rdt_get_cdp_l2_config();
806 		ret = true;
807 	}
808 
809 	if (get_mem_config())
810 		ret = true;
811 
812 	return ret;
813 }
814 
815 static __init bool get_rdt_mon_resources(void)
816 {
817 	struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
818 
819 	if (rdt_cpu_has(X86_FEATURE_CQM_OCCUP_LLC))
820 		rdt_mon_features |= (1 << QOS_L3_OCCUP_EVENT_ID);
821 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL))
822 		rdt_mon_features |= (1 << QOS_L3_MBM_TOTAL_EVENT_ID);
823 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL))
824 		rdt_mon_features |= (1 << QOS_L3_MBM_LOCAL_EVENT_ID);
825 
826 	if (!rdt_mon_features)
827 		return false;
828 
829 	return !rdt_get_mon_l3_config(r);
830 }
831 
832 static __init void __check_quirks_intel(void)
833 {
834 	switch (boot_cpu_data.x86_model) {
835 	case INTEL_FAM6_HASWELL_X:
836 		if (!rdt_options[RDT_FLAG_L3_CAT].force_off)
837 			cache_alloc_hsw_probe();
838 		break;
839 	case INTEL_FAM6_SKYLAKE_X:
840 		if (boot_cpu_data.x86_stepping <= 4)
841 			set_rdt_options("!cmt,!mbmtotal,!mbmlocal,!l3cat");
842 		else
843 			set_rdt_options("!l3cat");
844 		fallthrough;
845 	case INTEL_FAM6_BROADWELL_X:
846 		intel_rdt_mbm_apply_quirk();
847 		break;
848 	}
849 }
850 
851 static __init void check_quirks(void)
852 {
853 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
854 		__check_quirks_intel();
855 }
856 
857 static __init bool get_rdt_resources(void)
858 {
859 	rdt_alloc_capable = get_rdt_alloc_resources();
860 	rdt_mon_capable = get_rdt_mon_resources();
861 
862 	return (rdt_mon_capable || rdt_alloc_capable);
863 }
864 
865 static __init void rdt_init_res_defs_intel(void)
866 {
867 	struct rdt_hw_resource *hw_res;
868 	struct rdt_resource *r;
869 
870 	for_each_rdt_resource(r) {
871 		hw_res = resctrl_to_arch_res(r);
872 
873 		if (r->rid == RDT_RESOURCE_L3 ||
874 		    r->rid == RDT_RESOURCE_L2) {
875 			r->cache.arch_has_sparse_bitmaps = false;
876 			r->cache.arch_has_empty_bitmaps = false;
877 			r->cache.arch_has_per_cpu_cfg = false;
878 		} else if (r->rid == RDT_RESOURCE_MBA) {
879 			hw_res->msr_base = MSR_IA32_MBA_THRTL_BASE;
880 			hw_res->msr_update = mba_wrmsr_intel;
881 		}
882 	}
883 }
884 
885 static __init void rdt_init_res_defs_amd(void)
886 {
887 	struct rdt_hw_resource *hw_res;
888 	struct rdt_resource *r;
889 
890 	for_each_rdt_resource(r) {
891 		hw_res = resctrl_to_arch_res(r);
892 
893 		if (r->rid == RDT_RESOURCE_L3 ||
894 		    r->rid == RDT_RESOURCE_L2) {
895 			r->cache.arch_has_sparse_bitmaps = true;
896 			r->cache.arch_has_empty_bitmaps = true;
897 			r->cache.arch_has_per_cpu_cfg = true;
898 		} else if (r->rid == RDT_RESOURCE_MBA) {
899 			hw_res->msr_base = MSR_IA32_MBA_BW_BASE;
900 			hw_res->msr_update = mba_wrmsr_amd;
901 		}
902 	}
903 }
904 
905 static __init void rdt_init_res_defs(void)
906 {
907 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
908 		rdt_init_res_defs_intel();
909 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
910 		rdt_init_res_defs_amd();
911 }
912 
913 static enum cpuhp_state rdt_online;
914 
915 /* Runs once on the BSP during boot. */
916 void resctrl_cpu_detect(struct cpuinfo_x86 *c)
917 {
918 	if (!cpu_has(c, X86_FEATURE_CQM_LLC)) {
919 		c->x86_cache_max_rmid  = -1;
920 		c->x86_cache_occ_scale = -1;
921 		c->x86_cache_mbm_width_offset = -1;
922 		return;
923 	}
924 
925 	/* will be overridden if occupancy monitoring exists */
926 	c->x86_cache_max_rmid = cpuid_ebx(0xf);
927 
928 	if (cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC) ||
929 	    cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL) ||
930 	    cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)) {
931 		u32 eax, ebx, ecx, edx;
932 
933 		/* QoS sub-leaf, EAX=0Fh, ECX=1 */
934 		cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx);
935 
936 		c->x86_cache_max_rmid  = ecx;
937 		c->x86_cache_occ_scale = ebx;
938 		c->x86_cache_mbm_width_offset = eax & 0xff;
939 
940 		if (c->x86_vendor == X86_VENDOR_AMD && !c->x86_cache_mbm_width_offset)
941 			c->x86_cache_mbm_width_offset = MBM_CNTR_WIDTH_OFFSET_AMD;
942 	}
943 }
944 
945 static int __init resctrl_late_init(void)
946 {
947 	struct rdt_resource *r;
948 	int state, ret;
949 
950 	/*
951 	 * Initialize functions(or definitions) that are different
952 	 * between vendors here.
953 	 */
954 	rdt_init_res_defs();
955 
956 	check_quirks();
957 
958 	if (!get_rdt_resources())
959 		return -ENODEV;
960 
961 	rdt_init_padding();
962 
963 	state = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
964 				  "x86/resctrl/cat:online:",
965 				  resctrl_online_cpu, resctrl_offline_cpu);
966 	if (state < 0)
967 		return state;
968 
969 	ret = rdtgroup_init();
970 	if (ret) {
971 		cpuhp_remove_state(state);
972 		return ret;
973 	}
974 	rdt_online = state;
975 
976 	for_each_alloc_capable_rdt_resource(r)
977 		pr_info("%s allocation detected\n", r->name);
978 
979 	for_each_mon_capable_rdt_resource(r)
980 		pr_info("%s monitoring detected\n", r->name);
981 
982 	return 0;
983 }
984 
985 late_initcall(resctrl_late_init);
986 
987 static void __exit resctrl_exit(void)
988 {
989 	cpuhp_remove_state(rdt_online);
990 	rdtgroup_exit();
991 }
992 
993 __exitcall(resctrl_exit);
994