xref: /openbmc/linux/arch/x86/kernel/cpu/resctrl/core.c (revision 79e790ff)
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].domains)
61 
62 struct rdt_resource rdt_resources_all[] = {
63 	[RDT_RESOURCE_L3] =
64 	{
65 		.rid			= RDT_RESOURCE_L3,
66 		.name			= "L3",
67 		.domains		= domain_init(RDT_RESOURCE_L3),
68 		.msr_base		= MSR_IA32_L3_CBM_BASE,
69 		.msr_update		= cat_wrmsr,
70 		.cache_level		= 3,
71 		.cache = {
72 			.min_cbm_bits	= 1,
73 			.cbm_idx_mult	= 1,
74 			.cbm_idx_offset	= 0,
75 		},
76 		.parse_ctrlval		= parse_cbm,
77 		.format_str		= "%d=%0*x",
78 		.fflags			= RFTYPE_RES_CACHE,
79 	},
80 	[RDT_RESOURCE_L3DATA] =
81 	{
82 		.rid			= RDT_RESOURCE_L3DATA,
83 		.name			= "L3DATA",
84 		.domains		= domain_init(RDT_RESOURCE_L3DATA),
85 		.msr_base		= MSR_IA32_L3_CBM_BASE,
86 		.msr_update		= cat_wrmsr,
87 		.cache_level		= 3,
88 		.cache = {
89 			.min_cbm_bits	= 1,
90 			.cbm_idx_mult	= 2,
91 			.cbm_idx_offset	= 0,
92 		},
93 		.parse_ctrlval		= parse_cbm,
94 		.format_str		= "%d=%0*x",
95 		.fflags			= RFTYPE_RES_CACHE,
96 	},
97 	[RDT_RESOURCE_L3CODE] =
98 	{
99 		.rid			= RDT_RESOURCE_L3CODE,
100 		.name			= "L3CODE",
101 		.domains		= domain_init(RDT_RESOURCE_L3CODE),
102 		.msr_base		= MSR_IA32_L3_CBM_BASE,
103 		.msr_update		= cat_wrmsr,
104 		.cache_level		= 3,
105 		.cache = {
106 			.min_cbm_bits	= 1,
107 			.cbm_idx_mult	= 2,
108 			.cbm_idx_offset	= 1,
109 		},
110 		.parse_ctrlval		= parse_cbm,
111 		.format_str		= "%d=%0*x",
112 		.fflags			= RFTYPE_RES_CACHE,
113 	},
114 	[RDT_RESOURCE_L2] =
115 	{
116 		.rid			= RDT_RESOURCE_L2,
117 		.name			= "L2",
118 		.domains		= domain_init(RDT_RESOURCE_L2),
119 		.msr_base		= MSR_IA32_L2_CBM_BASE,
120 		.msr_update		= cat_wrmsr,
121 		.cache_level		= 2,
122 		.cache = {
123 			.min_cbm_bits	= 1,
124 			.cbm_idx_mult	= 1,
125 			.cbm_idx_offset	= 0,
126 		},
127 		.parse_ctrlval		= parse_cbm,
128 		.format_str		= "%d=%0*x",
129 		.fflags			= RFTYPE_RES_CACHE,
130 	},
131 	[RDT_RESOURCE_L2DATA] =
132 	{
133 		.rid			= RDT_RESOURCE_L2DATA,
134 		.name			= "L2DATA",
135 		.domains		= domain_init(RDT_RESOURCE_L2DATA),
136 		.msr_base		= MSR_IA32_L2_CBM_BASE,
137 		.msr_update		= cat_wrmsr,
138 		.cache_level		= 2,
139 		.cache = {
140 			.min_cbm_bits	= 1,
141 			.cbm_idx_mult	= 2,
142 			.cbm_idx_offset	= 0,
143 		},
144 		.parse_ctrlval		= parse_cbm,
145 		.format_str		= "%d=%0*x",
146 		.fflags			= RFTYPE_RES_CACHE,
147 	},
148 	[RDT_RESOURCE_L2CODE] =
149 	{
150 		.rid			= RDT_RESOURCE_L2CODE,
151 		.name			= "L2CODE",
152 		.domains		= domain_init(RDT_RESOURCE_L2CODE),
153 		.msr_base		= MSR_IA32_L2_CBM_BASE,
154 		.msr_update		= cat_wrmsr,
155 		.cache_level		= 2,
156 		.cache = {
157 			.min_cbm_bits	= 1,
158 			.cbm_idx_mult	= 2,
159 			.cbm_idx_offset	= 1,
160 		},
161 		.parse_ctrlval		= parse_cbm,
162 		.format_str		= "%d=%0*x",
163 		.fflags			= RFTYPE_RES_CACHE,
164 	},
165 	[RDT_RESOURCE_MBA] =
166 	{
167 		.rid			= RDT_RESOURCE_MBA,
168 		.name			= "MB",
169 		.domains		= domain_init(RDT_RESOURCE_MBA),
170 		.cache_level		= 3,
171 		.parse_ctrlval		= parse_bw,
172 		.format_str		= "%d=%*u",
173 		.fflags			= RFTYPE_RES_MB,
174 	},
175 };
176 
177 static unsigned int cbm_idx(struct rdt_resource *r, unsigned int closid)
178 {
179 	return closid * r->cache.cbm_idx_mult + r->cache.cbm_idx_offset;
180 }
181 
182 /*
183  * cache_alloc_hsw_probe() - Have to probe for Intel haswell server CPUs
184  * as they do not have CPUID enumeration support for Cache allocation.
185  * The check for Vendor/Family/Model is not enough to guarantee that
186  * the MSRs won't #GP fault because only the following SKUs support
187  * CAT:
188  *	Intel(R) Xeon(R)  CPU E5-2658  v3  @  2.20GHz
189  *	Intel(R) Xeon(R)  CPU E5-2648L v3  @  1.80GHz
190  *	Intel(R) Xeon(R)  CPU E5-2628L v3  @  2.00GHz
191  *	Intel(R) Xeon(R)  CPU E5-2618L v3  @  2.30GHz
192  *	Intel(R) Xeon(R)  CPU E5-2608L v3  @  2.00GHz
193  *	Intel(R) Xeon(R)  CPU E5-2658A v3  @  2.20GHz
194  *
195  * Probe by trying to write the first of the L3 cache mask registers
196  * and checking that the bits stick. Max CLOSids is always 4 and max cbm length
197  * is always 20 on hsw server parts. The minimum cache bitmask length
198  * allowed for HSW server is always 2 bits. Hardcode all of them.
199  */
200 static inline void cache_alloc_hsw_probe(void)
201 {
202 	struct rdt_resource *r  = &rdt_resources_all[RDT_RESOURCE_L3];
203 	u32 l, h, max_cbm = BIT_MASK(20) - 1;
204 
205 	if (wrmsr_safe(MSR_IA32_L3_CBM_BASE, max_cbm, 0))
206 		return;
207 
208 	rdmsr(MSR_IA32_L3_CBM_BASE, l, h);
209 
210 	/* If all the bits were set in MSR, return success */
211 	if (l != max_cbm)
212 		return;
213 
214 	r->num_closid = 4;
215 	r->default_ctrl = max_cbm;
216 	r->cache.cbm_len = 20;
217 	r->cache.shareable_bits = 0xc0000;
218 	r->cache.min_cbm_bits = 2;
219 	r->alloc_capable = true;
220 	r->alloc_enabled = true;
221 
222 	rdt_alloc_capable = true;
223 }
224 
225 bool is_mba_sc(struct rdt_resource *r)
226 {
227 	if (!r)
228 		return rdt_resources_all[RDT_RESOURCE_MBA].membw.mba_sc;
229 
230 	return r->membw.mba_sc;
231 }
232 
233 /*
234  * rdt_get_mb_table() - get a mapping of bandwidth(b/w) percentage values
235  * exposed to user interface and the h/w understandable delay values.
236  *
237  * The non-linear delay values have the granularity of power of two
238  * and also the h/w does not guarantee a curve for configured delay
239  * values vs. actual b/w enforced.
240  * Hence we need a mapping that is pre calibrated so the user can
241  * express the memory b/w as a percentage value.
242  */
243 static inline bool rdt_get_mb_table(struct rdt_resource *r)
244 {
245 	/*
246 	 * There are no Intel SKUs as of now to support non-linear delay.
247 	 */
248 	pr_info("MBA b/w map not implemented for cpu:%d, model:%d",
249 		boot_cpu_data.x86, boot_cpu_data.x86_model);
250 
251 	return false;
252 }
253 
254 static bool __get_mem_config_intel(struct rdt_resource *r)
255 {
256 	union cpuid_0x10_3_eax eax;
257 	union cpuid_0x10_x_edx edx;
258 	u32 ebx, ecx, max_delay;
259 
260 	cpuid_count(0x00000010, 3, &eax.full, &ebx, &ecx, &edx.full);
261 	r->num_closid = edx.split.cos_max + 1;
262 	max_delay = eax.split.max_delay + 1;
263 	r->default_ctrl = MAX_MBA_BW;
264 	r->membw.arch_needs_linear = true;
265 	if (ecx & MBA_IS_LINEAR) {
266 		r->membw.delay_linear = true;
267 		r->membw.min_bw = MAX_MBA_BW - max_delay;
268 		r->membw.bw_gran = MAX_MBA_BW - max_delay;
269 	} else {
270 		if (!rdt_get_mb_table(r))
271 			return false;
272 		r->membw.arch_needs_linear = false;
273 	}
274 	r->data_width = 3;
275 
276 	if (boot_cpu_has(X86_FEATURE_PER_THREAD_MBA))
277 		r->membw.throttle_mode = THREAD_THROTTLE_PER_THREAD;
278 	else
279 		r->membw.throttle_mode = THREAD_THROTTLE_MAX;
280 	thread_throttle_mode_init();
281 
282 	r->alloc_capable = true;
283 	r->alloc_enabled = true;
284 
285 	return true;
286 }
287 
288 static bool __rdt_get_mem_config_amd(struct rdt_resource *r)
289 {
290 	union cpuid_0x10_3_eax eax;
291 	union cpuid_0x10_x_edx edx;
292 	u32 ebx, ecx;
293 
294 	cpuid_count(0x80000020, 1, &eax.full, &ebx, &ecx, &edx.full);
295 	r->num_closid = edx.split.cos_max + 1;
296 	r->default_ctrl = MAX_MBA_BW_AMD;
297 
298 	/* AMD does not use delay */
299 	r->membw.delay_linear = false;
300 	r->membw.arch_needs_linear = false;
301 
302 	/*
303 	 * AMD does not use memory delay throttle model to control
304 	 * the allocation like Intel does.
305 	 */
306 	r->membw.throttle_mode = THREAD_THROTTLE_UNDEFINED;
307 	r->membw.min_bw = 0;
308 	r->membw.bw_gran = 1;
309 	/* Max value is 2048, Data width should be 4 in decimal */
310 	r->data_width = 4;
311 
312 	r->alloc_capable = true;
313 	r->alloc_enabled = true;
314 
315 	return true;
316 }
317 
318 static void rdt_get_cache_alloc_cfg(int idx, struct rdt_resource *r)
319 {
320 	union cpuid_0x10_1_eax eax;
321 	union cpuid_0x10_x_edx edx;
322 	u32 ebx, ecx;
323 
324 	cpuid_count(0x00000010, idx, &eax.full, &ebx, &ecx, &edx.full);
325 	r->num_closid = edx.split.cos_max + 1;
326 	r->cache.cbm_len = eax.split.cbm_len + 1;
327 	r->default_ctrl = BIT_MASK(eax.split.cbm_len + 1) - 1;
328 	r->cache.shareable_bits = ebx & r->default_ctrl;
329 	r->data_width = (r->cache.cbm_len + 3) / 4;
330 	r->alloc_capable = true;
331 	r->alloc_enabled = true;
332 }
333 
334 static void rdt_get_cdp_config(int level, int type)
335 {
336 	struct rdt_resource *r_l = &rdt_resources_all[level];
337 	struct rdt_resource *r = &rdt_resources_all[type];
338 
339 	r->num_closid = r_l->num_closid / 2;
340 	r->cache.cbm_len = r_l->cache.cbm_len;
341 	r->default_ctrl = r_l->default_ctrl;
342 	r->cache.shareable_bits = r_l->cache.shareable_bits;
343 	r->data_width = (r->cache.cbm_len + 3) / 4;
344 	r->alloc_capable = true;
345 	/*
346 	 * By default, CDP is disabled. CDP can be enabled by mount parameter
347 	 * "cdp" during resctrl file system mount time.
348 	 */
349 	r->alloc_enabled = false;
350 }
351 
352 static void rdt_get_cdp_l3_config(void)
353 {
354 	rdt_get_cdp_config(RDT_RESOURCE_L3, RDT_RESOURCE_L3DATA);
355 	rdt_get_cdp_config(RDT_RESOURCE_L3, RDT_RESOURCE_L3CODE);
356 }
357 
358 static void rdt_get_cdp_l2_config(void)
359 {
360 	rdt_get_cdp_config(RDT_RESOURCE_L2, RDT_RESOURCE_L2DATA);
361 	rdt_get_cdp_config(RDT_RESOURCE_L2, RDT_RESOURCE_L2CODE);
362 }
363 
364 static void
365 mba_wrmsr_amd(struct rdt_domain *d, struct msr_param *m, struct rdt_resource *r)
366 {
367 	unsigned int i;
368 
369 	for (i = m->low; i < m->high; i++)
370 		wrmsrl(r->msr_base + i, d->ctrl_val[i]);
371 }
372 
373 /*
374  * Map the memory b/w percentage value to delay values
375  * that can be written to QOS_MSRs.
376  * There are currently no SKUs which support non linear delay values.
377  */
378 u32 delay_bw_map(unsigned long bw, struct rdt_resource *r)
379 {
380 	if (r->membw.delay_linear)
381 		return MAX_MBA_BW - bw;
382 
383 	pr_warn_once("Non Linear delay-bw map not supported but queried\n");
384 	return r->default_ctrl;
385 }
386 
387 static void
388 mba_wrmsr_intel(struct rdt_domain *d, struct msr_param *m,
389 		struct rdt_resource *r)
390 {
391 	unsigned int i;
392 
393 	/*  Write the delay values for mba. */
394 	for (i = m->low; i < m->high; i++)
395 		wrmsrl(r->msr_base + i, delay_bw_map(d->ctrl_val[i], r));
396 }
397 
398 static void
399 cat_wrmsr(struct rdt_domain *d, struct msr_param *m, struct rdt_resource *r)
400 {
401 	unsigned int i;
402 
403 	for (i = m->low; i < m->high; i++)
404 		wrmsrl(r->msr_base + cbm_idx(r, i), d->ctrl_val[i]);
405 }
406 
407 struct rdt_domain *get_domain_from_cpu(int cpu, struct rdt_resource *r)
408 {
409 	struct rdt_domain *d;
410 
411 	list_for_each_entry(d, &r->domains, list) {
412 		/* Find the domain that contains this CPU */
413 		if (cpumask_test_cpu(cpu, &d->cpu_mask))
414 			return d;
415 	}
416 
417 	return NULL;
418 }
419 
420 void rdt_ctrl_update(void *arg)
421 {
422 	struct msr_param *m = arg;
423 	struct rdt_resource *r = m->res;
424 	int cpu = smp_processor_id();
425 	struct rdt_domain *d;
426 
427 	d = get_domain_from_cpu(cpu, r);
428 	if (d) {
429 		r->msr_update(d, m, r);
430 		return;
431 	}
432 	pr_warn_once("cpu %d not found in any domain for resource %s\n",
433 		     cpu, r->name);
434 }
435 
436 /*
437  * rdt_find_domain - Find a domain in a resource that matches input resource id
438  *
439  * Search resource r's domain list to find the resource id. If the resource
440  * id is found in a domain, return the domain. Otherwise, if requested by
441  * caller, return the first domain whose id is bigger than the input id.
442  * The domain list is sorted by id in ascending order.
443  */
444 struct rdt_domain *rdt_find_domain(struct rdt_resource *r, int id,
445 				   struct list_head **pos)
446 {
447 	struct rdt_domain *d;
448 	struct list_head *l;
449 
450 	if (id < 0)
451 		return ERR_PTR(-ENODEV);
452 
453 	list_for_each(l, &r->domains) {
454 		d = list_entry(l, struct rdt_domain, list);
455 		/* When id is found, return its domain. */
456 		if (id == d->id)
457 			return d;
458 		/* Stop searching when finding id's position in sorted list. */
459 		if (id < d->id)
460 			break;
461 	}
462 
463 	if (pos)
464 		*pos = l;
465 
466 	return NULL;
467 }
468 
469 void setup_default_ctrlval(struct rdt_resource *r, u32 *dc, u32 *dm)
470 {
471 	int i;
472 
473 	/*
474 	 * Initialize the Control MSRs to having no control.
475 	 * For Cache Allocation: Set all bits in cbm
476 	 * For Memory Allocation: Set b/w requested to 100%
477 	 * and the bandwidth in MBps to U32_MAX
478 	 */
479 	for (i = 0; i < r->num_closid; i++, dc++, dm++) {
480 		*dc = r->default_ctrl;
481 		*dm = MBA_MAX_MBPS;
482 	}
483 }
484 
485 static int domain_setup_ctrlval(struct rdt_resource *r, struct rdt_domain *d)
486 {
487 	struct msr_param m;
488 	u32 *dc, *dm;
489 
490 	dc = kmalloc_array(r->num_closid, sizeof(*d->ctrl_val), GFP_KERNEL);
491 	if (!dc)
492 		return -ENOMEM;
493 
494 	dm = kmalloc_array(r->num_closid, sizeof(*d->mbps_val), GFP_KERNEL);
495 	if (!dm) {
496 		kfree(dc);
497 		return -ENOMEM;
498 	}
499 
500 	d->ctrl_val = dc;
501 	d->mbps_val = dm;
502 	setup_default_ctrlval(r, dc, dm);
503 
504 	m.low = 0;
505 	m.high = r->num_closid;
506 	r->msr_update(d, &m, r);
507 	return 0;
508 }
509 
510 static int domain_setup_mon_state(struct rdt_resource *r, struct rdt_domain *d)
511 {
512 	size_t tsize;
513 
514 	if (is_llc_occupancy_enabled()) {
515 		d->rmid_busy_llc = bitmap_zalloc(r->num_rmid, GFP_KERNEL);
516 		if (!d->rmid_busy_llc)
517 			return -ENOMEM;
518 		INIT_DELAYED_WORK(&d->cqm_limbo, cqm_handle_limbo);
519 	}
520 	if (is_mbm_total_enabled()) {
521 		tsize = sizeof(*d->mbm_total);
522 		d->mbm_total = kcalloc(r->num_rmid, tsize, GFP_KERNEL);
523 		if (!d->mbm_total) {
524 			bitmap_free(d->rmid_busy_llc);
525 			return -ENOMEM;
526 		}
527 	}
528 	if (is_mbm_local_enabled()) {
529 		tsize = sizeof(*d->mbm_local);
530 		d->mbm_local = kcalloc(r->num_rmid, tsize, GFP_KERNEL);
531 		if (!d->mbm_local) {
532 			bitmap_free(d->rmid_busy_llc);
533 			kfree(d->mbm_total);
534 			return -ENOMEM;
535 		}
536 	}
537 
538 	if (is_mbm_enabled()) {
539 		INIT_DELAYED_WORK(&d->mbm_over, mbm_handle_overflow);
540 		mbm_setup_overflow_handler(d, MBM_OVERFLOW_INTERVAL);
541 	}
542 
543 	return 0;
544 }
545 
546 /*
547  * domain_add_cpu - Add a cpu to a resource's domain list.
548  *
549  * If an existing domain in the resource r's domain list matches the cpu's
550  * resource id, add the cpu in the domain.
551  *
552  * Otherwise, a new domain is allocated and inserted into the right position
553  * in the domain list sorted by id in ascending order.
554  *
555  * The order in the domain list is visible to users when we print entries
556  * in the schemata file and schemata input is validated to have the same order
557  * as this list.
558  */
559 static void domain_add_cpu(int cpu, struct rdt_resource *r)
560 {
561 	int id = get_cpu_cacheinfo_id(cpu, r->cache_level);
562 	struct list_head *add_pos = NULL;
563 	struct rdt_domain *d;
564 
565 	d = rdt_find_domain(r, id, &add_pos);
566 	if (IS_ERR(d)) {
567 		pr_warn("Couldn't find cache id for CPU %d\n", cpu);
568 		return;
569 	}
570 
571 	if (d) {
572 		cpumask_set_cpu(cpu, &d->cpu_mask);
573 		if (r->cache.arch_has_per_cpu_cfg)
574 			rdt_domain_reconfigure_cdp(r);
575 		return;
576 	}
577 
578 	d = kzalloc_node(sizeof(*d), GFP_KERNEL, cpu_to_node(cpu));
579 	if (!d)
580 		return;
581 
582 	d->id = id;
583 	cpumask_set_cpu(cpu, &d->cpu_mask);
584 
585 	rdt_domain_reconfigure_cdp(r);
586 
587 	if (r->alloc_capable && domain_setup_ctrlval(r, d)) {
588 		kfree(d);
589 		return;
590 	}
591 
592 	if (r->mon_capable && domain_setup_mon_state(r, d)) {
593 		kfree(d);
594 		return;
595 	}
596 
597 	list_add_tail(&d->list, add_pos);
598 
599 	/*
600 	 * If resctrl is mounted, add
601 	 * per domain monitor data directories.
602 	 */
603 	if (static_branch_unlikely(&rdt_mon_enable_key))
604 		mkdir_mondata_subdir_allrdtgrp(r, d);
605 }
606 
607 static void domain_remove_cpu(int cpu, struct rdt_resource *r)
608 {
609 	int id = get_cpu_cacheinfo_id(cpu, r->cache_level);
610 	struct rdt_domain *d;
611 
612 	d = rdt_find_domain(r, id, NULL);
613 	if (IS_ERR_OR_NULL(d)) {
614 		pr_warn("Couldn't find cache id for CPU %d\n", cpu);
615 		return;
616 	}
617 
618 	cpumask_clear_cpu(cpu, &d->cpu_mask);
619 	if (cpumask_empty(&d->cpu_mask)) {
620 		/*
621 		 * If resctrl is mounted, remove all the
622 		 * per domain monitor data directories.
623 		 */
624 		if (static_branch_unlikely(&rdt_mon_enable_key))
625 			rmdir_mondata_subdir_allrdtgrp(r, d->id);
626 		list_del(&d->list);
627 		if (r->mon_capable && is_mbm_enabled())
628 			cancel_delayed_work(&d->mbm_over);
629 		if (is_llc_occupancy_enabled() &&  has_busy_rmid(r, d)) {
630 			/*
631 			 * When a package is going down, forcefully
632 			 * decrement rmid->ebusy. There is no way to know
633 			 * that the L3 was flushed and hence may lead to
634 			 * incorrect counts in rare scenarios, but leaving
635 			 * the RMID as busy creates RMID leaks if the
636 			 * package never comes back.
637 			 */
638 			__check_limbo(d, true);
639 			cancel_delayed_work(&d->cqm_limbo);
640 		}
641 
642 		/*
643 		 * rdt_domain "d" is going to be freed below, so clear
644 		 * its pointer from pseudo_lock_region struct.
645 		 */
646 		if (d->plr)
647 			d->plr->d = NULL;
648 
649 		kfree(d->ctrl_val);
650 		kfree(d->mbps_val);
651 		bitmap_free(d->rmid_busy_llc);
652 		kfree(d->mbm_total);
653 		kfree(d->mbm_local);
654 		kfree(d);
655 		return;
656 	}
657 
658 	if (r == &rdt_resources_all[RDT_RESOURCE_L3]) {
659 		if (is_mbm_enabled() && cpu == d->mbm_work_cpu) {
660 			cancel_delayed_work(&d->mbm_over);
661 			mbm_setup_overflow_handler(d, 0);
662 		}
663 		if (is_llc_occupancy_enabled() && cpu == d->cqm_work_cpu &&
664 		    has_busy_rmid(r, d)) {
665 			cancel_delayed_work(&d->cqm_limbo);
666 			cqm_setup_limbo_handler(d, 0);
667 		}
668 	}
669 }
670 
671 static void clear_closid_rmid(int cpu)
672 {
673 	struct resctrl_pqr_state *state = this_cpu_ptr(&pqr_state);
674 
675 	state->default_closid = 0;
676 	state->default_rmid = 0;
677 	state->cur_closid = 0;
678 	state->cur_rmid = 0;
679 	wrmsr(IA32_PQR_ASSOC, 0, 0);
680 }
681 
682 static int resctrl_online_cpu(unsigned int cpu)
683 {
684 	struct rdt_resource *r;
685 
686 	mutex_lock(&rdtgroup_mutex);
687 	for_each_capable_rdt_resource(r)
688 		domain_add_cpu(cpu, r);
689 	/* The cpu is set in default rdtgroup after online. */
690 	cpumask_set_cpu(cpu, &rdtgroup_default.cpu_mask);
691 	clear_closid_rmid(cpu);
692 	mutex_unlock(&rdtgroup_mutex);
693 
694 	return 0;
695 }
696 
697 static void clear_childcpus(struct rdtgroup *r, unsigned int cpu)
698 {
699 	struct rdtgroup *cr;
700 
701 	list_for_each_entry(cr, &r->mon.crdtgrp_list, mon.crdtgrp_list) {
702 		if (cpumask_test_and_clear_cpu(cpu, &cr->cpu_mask)) {
703 			break;
704 		}
705 	}
706 }
707 
708 static int resctrl_offline_cpu(unsigned int cpu)
709 {
710 	struct rdtgroup *rdtgrp;
711 	struct rdt_resource *r;
712 
713 	mutex_lock(&rdtgroup_mutex);
714 	for_each_capable_rdt_resource(r)
715 		domain_remove_cpu(cpu, r);
716 	list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
717 		if (cpumask_test_and_clear_cpu(cpu, &rdtgrp->cpu_mask)) {
718 			clear_childcpus(rdtgrp, cpu);
719 			break;
720 		}
721 	}
722 	clear_closid_rmid(cpu);
723 	mutex_unlock(&rdtgroup_mutex);
724 
725 	return 0;
726 }
727 
728 /*
729  * Choose a width for the resource name and resource data based on the
730  * resource that has widest name and cbm.
731  */
732 static __init void rdt_init_padding(void)
733 {
734 	struct rdt_resource *r;
735 	int cl;
736 
737 	for_each_alloc_capable_rdt_resource(r) {
738 		cl = strlen(r->name);
739 		if (cl > max_name_width)
740 			max_name_width = cl;
741 
742 		if (r->data_width > max_data_width)
743 			max_data_width = r->data_width;
744 	}
745 }
746 
747 enum {
748 	RDT_FLAG_CMT,
749 	RDT_FLAG_MBM_TOTAL,
750 	RDT_FLAG_MBM_LOCAL,
751 	RDT_FLAG_L3_CAT,
752 	RDT_FLAG_L3_CDP,
753 	RDT_FLAG_L2_CAT,
754 	RDT_FLAG_L2_CDP,
755 	RDT_FLAG_MBA,
756 };
757 
758 #define RDT_OPT(idx, n, f)	\
759 [idx] = {			\
760 	.name = n,		\
761 	.flag = f		\
762 }
763 
764 struct rdt_options {
765 	char	*name;
766 	int	flag;
767 	bool	force_off, force_on;
768 };
769 
770 static struct rdt_options rdt_options[]  __initdata = {
771 	RDT_OPT(RDT_FLAG_CMT,	    "cmt",	X86_FEATURE_CQM_OCCUP_LLC),
772 	RDT_OPT(RDT_FLAG_MBM_TOTAL, "mbmtotal", X86_FEATURE_CQM_MBM_TOTAL),
773 	RDT_OPT(RDT_FLAG_MBM_LOCAL, "mbmlocal", X86_FEATURE_CQM_MBM_LOCAL),
774 	RDT_OPT(RDT_FLAG_L3_CAT,    "l3cat",	X86_FEATURE_CAT_L3),
775 	RDT_OPT(RDT_FLAG_L3_CDP,    "l3cdp",	X86_FEATURE_CDP_L3),
776 	RDT_OPT(RDT_FLAG_L2_CAT,    "l2cat",	X86_FEATURE_CAT_L2),
777 	RDT_OPT(RDT_FLAG_L2_CDP,    "l2cdp",	X86_FEATURE_CDP_L2),
778 	RDT_OPT(RDT_FLAG_MBA,	    "mba",	X86_FEATURE_MBA),
779 };
780 #define NUM_RDT_OPTIONS ARRAY_SIZE(rdt_options)
781 
782 static int __init set_rdt_options(char *str)
783 {
784 	struct rdt_options *o;
785 	bool force_off;
786 	char *tok;
787 
788 	if (*str == '=')
789 		str++;
790 	while ((tok = strsep(&str, ",")) != NULL) {
791 		force_off = *tok == '!';
792 		if (force_off)
793 			tok++;
794 		for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
795 			if (strcmp(tok, o->name) == 0) {
796 				if (force_off)
797 					o->force_off = true;
798 				else
799 					o->force_on = true;
800 				break;
801 			}
802 		}
803 	}
804 	return 1;
805 }
806 __setup("rdt", set_rdt_options);
807 
808 static bool __init rdt_cpu_has(int flag)
809 {
810 	bool ret = boot_cpu_has(flag);
811 	struct rdt_options *o;
812 
813 	if (!ret)
814 		return ret;
815 
816 	for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
817 		if (flag == o->flag) {
818 			if (o->force_off)
819 				ret = false;
820 			if (o->force_on)
821 				ret = true;
822 			break;
823 		}
824 	}
825 	return ret;
826 }
827 
828 static __init bool get_mem_config(void)
829 {
830 	if (!rdt_cpu_has(X86_FEATURE_MBA))
831 		return false;
832 
833 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
834 		return __get_mem_config_intel(&rdt_resources_all[RDT_RESOURCE_MBA]);
835 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
836 		return __rdt_get_mem_config_amd(&rdt_resources_all[RDT_RESOURCE_MBA]);
837 
838 	return false;
839 }
840 
841 static __init bool get_rdt_alloc_resources(void)
842 {
843 	bool ret = false;
844 
845 	if (rdt_alloc_capable)
846 		return true;
847 
848 	if (!boot_cpu_has(X86_FEATURE_RDT_A))
849 		return false;
850 
851 	if (rdt_cpu_has(X86_FEATURE_CAT_L3)) {
852 		rdt_get_cache_alloc_cfg(1, &rdt_resources_all[RDT_RESOURCE_L3]);
853 		if (rdt_cpu_has(X86_FEATURE_CDP_L3))
854 			rdt_get_cdp_l3_config();
855 		ret = true;
856 	}
857 	if (rdt_cpu_has(X86_FEATURE_CAT_L2)) {
858 		/* CPUID 0x10.2 fields are same format at 0x10.1 */
859 		rdt_get_cache_alloc_cfg(2, &rdt_resources_all[RDT_RESOURCE_L2]);
860 		if (rdt_cpu_has(X86_FEATURE_CDP_L2))
861 			rdt_get_cdp_l2_config();
862 		ret = true;
863 	}
864 
865 	if (get_mem_config())
866 		ret = true;
867 
868 	return ret;
869 }
870 
871 static __init bool get_rdt_mon_resources(void)
872 {
873 	if (rdt_cpu_has(X86_FEATURE_CQM_OCCUP_LLC))
874 		rdt_mon_features |= (1 << QOS_L3_OCCUP_EVENT_ID);
875 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL))
876 		rdt_mon_features |= (1 << QOS_L3_MBM_TOTAL_EVENT_ID);
877 	if (rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL))
878 		rdt_mon_features |= (1 << QOS_L3_MBM_LOCAL_EVENT_ID);
879 
880 	if (!rdt_mon_features)
881 		return false;
882 
883 	return !rdt_get_mon_l3_config(&rdt_resources_all[RDT_RESOURCE_L3]);
884 }
885 
886 static __init void __check_quirks_intel(void)
887 {
888 	switch (boot_cpu_data.x86_model) {
889 	case INTEL_FAM6_HASWELL_X:
890 		if (!rdt_options[RDT_FLAG_L3_CAT].force_off)
891 			cache_alloc_hsw_probe();
892 		break;
893 	case INTEL_FAM6_SKYLAKE_X:
894 		if (boot_cpu_data.x86_stepping <= 4)
895 			set_rdt_options("!cmt,!mbmtotal,!mbmlocal,!l3cat");
896 		else
897 			set_rdt_options("!l3cat");
898 		fallthrough;
899 	case INTEL_FAM6_BROADWELL_X:
900 		intel_rdt_mbm_apply_quirk();
901 		break;
902 	}
903 }
904 
905 static __init void check_quirks(void)
906 {
907 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
908 		__check_quirks_intel();
909 }
910 
911 static __init bool get_rdt_resources(void)
912 {
913 	rdt_alloc_capable = get_rdt_alloc_resources();
914 	rdt_mon_capable = get_rdt_mon_resources();
915 
916 	return (rdt_mon_capable || rdt_alloc_capable);
917 }
918 
919 static __init void rdt_init_res_defs_intel(void)
920 {
921 	struct rdt_resource *r;
922 
923 	for_each_rdt_resource(r) {
924 		if (r->rid == RDT_RESOURCE_L3 ||
925 		    r->rid == RDT_RESOURCE_L3DATA ||
926 		    r->rid == RDT_RESOURCE_L3CODE ||
927 		    r->rid == RDT_RESOURCE_L2 ||
928 		    r->rid == RDT_RESOURCE_L2DATA ||
929 		    r->rid == RDT_RESOURCE_L2CODE) {
930 			r->cache.arch_has_sparse_bitmaps = false;
931 			r->cache.arch_has_empty_bitmaps = false;
932 			r->cache.arch_has_per_cpu_cfg = false;
933 		} else if (r->rid == RDT_RESOURCE_MBA) {
934 			r->msr_base = MSR_IA32_MBA_THRTL_BASE;
935 			r->msr_update = mba_wrmsr_intel;
936 		}
937 	}
938 }
939 
940 static __init void rdt_init_res_defs_amd(void)
941 {
942 	struct rdt_resource *r;
943 
944 	for_each_rdt_resource(r) {
945 		if (r->rid == RDT_RESOURCE_L3 ||
946 		    r->rid == RDT_RESOURCE_L3DATA ||
947 		    r->rid == RDT_RESOURCE_L3CODE ||
948 		    r->rid == RDT_RESOURCE_L2 ||
949 		    r->rid == RDT_RESOURCE_L2DATA ||
950 		    r->rid == RDT_RESOURCE_L2CODE) {
951 			r->cache.arch_has_sparse_bitmaps = true;
952 			r->cache.arch_has_empty_bitmaps = true;
953 			r->cache.arch_has_per_cpu_cfg = true;
954 		} else if (r->rid == RDT_RESOURCE_MBA) {
955 			r->msr_base = MSR_IA32_MBA_BW_BASE;
956 			r->msr_update = mba_wrmsr_amd;
957 		}
958 	}
959 }
960 
961 static __init void rdt_init_res_defs(void)
962 {
963 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
964 		rdt_init_res_defs_intel();
965 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
966 		rdt_init_res_defs_amd();
967 }
968 
969 static enum cpuhp_state rdt_online;
970 
971 /* Runs once on the BSP during boot. */
972 void resctrl_cpu_detect(struct cpuinfo_x86 *c)
973 {
974 	if (!cpu_has(c, X86_FEATURE_CQM_LLC)) {
975 		c->x86_cache_max_rmid  = -1;
976 		c->x86_cache_occ_scale = -1;
977 		c->x86_cache_mbm_width_offset = -1;
978 		return;
979 	}
980 
981 	/* will be overridden if occupancy monitoring exists */
982 	c->x86_cache_max_rmid = cpuid_ebx(0xf);
983 
984 	if (cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC) ||
985 	    cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL) ||
986 	    cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)) {
987 		u32 eax, ebx, ecx, edx;
988 
989 		/* QoS sub-leaf, EAX=0Fh, ECX=1 */
990 		cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx);
991 
992 		c->x86_cache_max_rmid  = ecx;
993 		c->x86_cache_occ_scale = ebx;
994 		c->x86_cache_mbm_width_offset = eax & 0xff;
995 
996 		if (c->x86_vendor == X86_VENDOR_AMD && !c->x86_cache_mbm_width_offset)
997 			c->x86_cache_mbm_width_offset = MBM_CNTR_WIDTH_OFFSET_AMD;
998 	}
999 }
1000 
1001 static int __init resctrl_late_init(void)
1002 {
1003 	struct rdt_resource *r;
1004 	int state, ret;
1005 
1006 	/*
1007 	 * Initialize functions(or definitions) that are different
1008 	 * between vendors here.
1009 	 */
1010 	rdt_init_res_defs();
1011 
1012 	check_quirks();
1013 
1014 	if (!get_rdt_resources())
1015 		return -ENODEV;
1016 
1017 	rdt_init_padding();
1018 
1019 	state = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
1020 				  "x86/resctrl/cat:online:",
1021 				  resctrl_online_cpu, resctrl_offline_cpu);
1022 	if (state < 0)
1023 		return state;
1024 
1025 	ret = rdtgroup_init();
1026 	if (ret) {
1027 		cpuhp_remove_state(state);
1028 		return ret;
1029 	}
1030 	rdt_online = state;
1031 
1032 	for_each_alloc_capable_rdt_resource(r)
1033 		pr_info("%s allocation detected\n", r->name);
1034 
1035 	for_each_mon_capable_rdt_resource(r)
1036 		pr_info("%s monitoring detected\n", r->name);
1037 
1038 	return 0;
1039 }
1040 
1041 late_initcall(resctrl_late_init);
1042 
1043 static void __exit resctrl_exit(void)
1044 {
1045 	cpuhp_remove_state(rdt_online);
1046 	rdtgroup_exit();
1047 }
1048 
1049 __exitcall(resctrl_exit);
1050