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/resctrl.h>
26 #include "internal.h"
27
28 /*
29 * rdt_domain structures are kfree()d when their last CPU goes offline,
30 * and allocated when the first CPU in a new domain comes online.
31 * The rdt_resource's domain list is updated when this happens. Readers of
32 * the domain list must either take cpus_read_lock(), or rely on an RCU
33 * read-side critical section, to avoid observing concurrent modification.
34 * All writers take this mutex:
35 */
36 static DEFINE_MUTEX(domain_list_lock);
37
38 /*
39 * The cached resctrl_pqr_state is strictly per CPU and can never be
40 * updated from a remote CPU. Functions which modify the state
41 * are called with interrupts disabled and no preemption, which
42 * is sufficient for the protection.
43 */
44 DEFINE_PER_CPU(struct resctrl_pqr_state, pqr_state);
45
46 /*
47 * Used to store the max resource name width and max resource data width
48 * to display the schemata in a tabular format
49 */
50 int max_name_width, max_data_width;
51
52 /*
53 * Global boolean for rdt_alloc which is true if any
54 * resource allocation is enabled.
55 */
56 bool rdt_alloc_capable;
57
58 static void mba_wrmsr_intel(struct msr_param *m);
59 static void cat_wrmsr(struct msr_param *m);
60 static void mba_wrmsr_amd(struct msr_param *m);
61
62 #define ctrl_domain_init(id) LIST_HEAD_INIT(rdt_resources_all[id].r_resctrl.ctrl_domains)
63 #define mon_domain_init(id) LIST_HEAD_INIT(rdt_resources_all[id].r_resctrl.mon_domains)
64
65 struct rdt_hw_resource rdt_resources_all[] = {
66 [RDT_RESOURCE_L3] =
67 {
68 .r_resctrl = {
69 .rid = RDT_RESOURCE_L3,
70 .name = "L3",
71 .ctrl_scope = RESCTRL_L3_CACHE,
72 .mon_scope = RESCTRL_L3_CACHE,
73 .ctrl_domains = ctrl_domain_init(RDT_RESOURCE_L3),
74 .mon_domains = mon_domain_init(RDT_RESOURCE_L3),
75 .parse_ctrlval = parse_cbm,
76 .format_str = "%d=%0*x",
77 .fflags = RFTYPE_RES_CACHE,
78 },
79 .msr_base = MSR_IA32_L3_CBM_BASE,
80 .msr_update = cat_wrmsr,
81 },
82 [RDT_RESOURCE_L2] =
83 {
84 .r_resctrl = {
85 .rid = RDT_RESOURCE_L2,
86 .name = "L2",
87 .ctrl_scope = RESCTRL_L2_CACHE,
88 .ctrl_domains = ctrl_domain_init(RDT_RESOURCE_L2),
89 .parse_ctrlval = parse_cbm,
90 .format_str = "%d=%0*x",
91 .fflags = RFTYPE_RES_CACHE,
92 },
93 .msr_base = MSR_IA32_L2_CBM_BASE,
94 .msr_update = cat_wrmsr,
95 },
96 [RDT_RESOURCE_MBA] =
97 {
98 .r_resctrl = {
99 .rid = RDT_RESOURCE_MBA,
100 .name = "MB",
101 .ctrl_scope = RESCTRL_L3_CACHE,
102 .ctrl_domains = ctrl_domain_init(RDT_RESOURCE_MBA),
103 .parse_ctrlval = parse_bw,
104 .format_str = "%d=%*u",
105 .fflags = RFTYPE_RES_MB,
106 },
107 },
108 [RDT_RESOURCE_SMBA] =
109 {
110 .r_resctrl = {
111 .rid = RDT_RESOURCE_SMBA,
112 .name = "SMBA",
113 .ctrl_scope = RESCTRL_L3_CACHE,
114 .ctrl_domains = ctrl_domain_init(RDT_RESOURCE_SMBA),
115 .parse_ctrlval = parse_bw,
116 .format_str = "%d=%*u",
117 .fflags = RFTYPE_RES_MB,
118 },
119 },
120 };
121
resctrl_arch_system_num_rmid_idx(void)122 u32 resctrl_arch_system_num_rmid_idx(void)
123 {
124 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
125
126 /* RMID are independent numbers for x86. num_rmid_idx == num_rmid */
127 return r->num_rmid;
128 }
129
130 /*
131 * cache_alloc_hsw_probe() - Have to probe for Intel haswell server CPUs
132 * as they do not have CPUID enumeration support for Cache allocation.
133 * The check for Vendor/Family/Model is not enough to guarantee that
134 * the MSRs won't #GP fault because only the following SKUs support
135 * CAT:
136 * Intel(R) Xeon(R) CPU E5-2658 v3 @ 2.20GHz
137 * Intel(R) Xeon(R) CPU E5-2648L v3 @ 1.80GHz
138 * Intel(R) Xeon(R) CPU E5-2628L v3 @ 2.00GHz
139 * Intel(R) Xeon(R) CPU E5-2618L v3 @ 2.30GHz
140 * Intel(R) Xeon(R) CPU E5-2608L v3 @ 2.00GHz
141 * Intel(R) Xeon(R) CPU E5-2658A v3 @ 2.20GHz
142 *
143 * Probe by trying to write the first of the L3 cache mask registers
144 * and checking that the bits stick. Max CLOSids is always 4 and max cbm length
145 * is always 20 on hsw server parts. The minimum cache bitmask length
146 * allowed for HSW server is always 2 bits. Hardcode all of them.
147 */
cache_alloc_hsw_probe(void)148 static inline void cache_alloc_hsw_probe(void)
149 {
150 struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_L3];
151 struct rdt_resource *r = &hw_res->r_resctrl;
152 u64 max_cbm = BIT_ULL_MASK(20) - 1, l3_cbm_0;
153
154 if (wrmsrl_safe(MSR_IA32_L3_CBM_BASE, max_cbm))
155 return;
156
157 rdmsrl(MSR_IA32_L3_CBM_BASE, l3_cbm_0);
158
159 /* If all the bits were set in MSR, return success */
160 if (l3_cbm_0 != max_cbm)
161 return;
162
163 hw_res->num_closid = 4;
164 r->default_ctrl = max_cbm;
165 r->cache.cbm_len = 20;
166 r->cache.shareable_bits = 0xc0000;
167 r->cache.min_cbm_bits = 2;
168 r->cache.arch_has_sparse_bitmasks = false;
169 r->alloc_capable = true;
170
171 rdt_alloc_capable = true;
172 }
173
is_mba_sc(struct rdt_resource * r)174 bool is_mba_sc(struct rdt_resource *r)
175 {
176 if (!r)
177 return rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl.membw.mba_sc;
178
179 /*
180 * The software controller support is only applicable to MBA resource.
181 * Make sure to check for resource type.
182 */
183 if (r->rid != RDT_RESOURCE_MBA)
184 return false;
185
186 return r->membw.mba_sc;
187 }
188
189 /*
190 * rdt_get_mb_table() - get a mapping of bandwidth(b/w) percentage values
191 * exposed to user interface and the h/w understandable delay values.
192 *
193 * The non-linear delay values have the granularity of power of two
194 * and also the h/w does not guarantee a curve for configured delay
195 * values vs. actual b/w enforced.
196 * Hence we need a mapping that is pre calibrated so the user can
197 * express the memory b/w as a percentage value.
198 */
rdt_get_mb_table(struct rdt_resource * r)199 static inline bool rdt_get_mb_table(struct rdt_resource *r)
200 {
201 /*
202 * There are no Intel SKUs as of now to support non-linear delay.
203 */
204 pr_info("MBA b/w map not implemented for cpu:%d, model:%d",
205 boot_cpu_data.x86, boot_cpu_data.x86_model);
206
207 return false;
208 }
209
__get_mem_config_intel(struct rdt_resource * r)210 static __init bool __get_mem_config_intel(struct rdt_resource *r)
211 {
212 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
213 union cpuid_0x10_3_eax eax;
214 union cpuid_0x10_x_edx edx;
215 u32 ebx, ecx, max_delay;
216
217 cpuid_count(0x00000010, 3, &eax.full, &ebx, &ecx, &edx.full);
218 hw_res->num_closid = edx.split.cos_max + 1;
219 max_delay = eax.split.max_delay + 1;
220 r->default_ctrl = MAX_MBA_BW;
221 r->membw.arch_needs_linear = true;
222 if (ecx & MBA_IS_LINEAR) {
223 r->membw.delay_linear = true;
224 r->membw.min_bw = MAX_MBA_BW - max_delay;
225 r->membw.bw_gran = MAX_MBA_BW - max_delay;
226 } else {
227 if (!rdt_get_mb_table(r))
228 return false;
229 r->membw.arch_needs_linear = false;
230 }
231 r->data_width = 3;
232
233 if (boot_cpu_has(X86_FEATURE_PER_THREAD_MBA))
234 r->membw.throttle_mode = THREAD_THROTTLE_PER_THREAD;
235 else
236 r->membw.throttle_mode = THREAD_THROTTLE_MAX;
237 thread_throttle_mode_init();
238
239 r->alloc_capable = true;
240
241 return true;
242 }
243
__rdt_get_mem_config_amd(struct rdt_resource * r)244 static __init bool __rdt_get_mem_config_amd(struct rdt_resource *r)
245 {
246 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
247 u32 eax, ebx, ecx, edx, subleaf;
248
249 /*
250 * Query CPUID_Fn80000020_EDX_x01 for MBA and
251 * CPUID_Fn80000020_EDX_x02 for SMBA
252 */
253 subleaf = (r->rid == RDT_RESOURCE_SMBA) ? 2 : 1;
254
255 cpuid_count(0x80000020, subleaf, &eax, &ebx, &ecx, &edx);
256 hw_res->num_closid = edx + 1;
257 r->default_ctrl = 1 << eax;
258
259 /* AMD does not use delay */
260 r->membw.delay_linear = false;
261 r->membw.arch_needs_linear = false;
262
263 /*
264 * AMD does not use memory delay throttle model to control
265 * the allocation like Intel does.
266 */
267 r->membw.throttle_mode = THREAD_THROTTLE_UNDEFINED;
268 r->membw.min_bw = 0;
269 r->membw.bw_gran = 1;
270 /* Max value is 2048, Data width should be 4 in decimal */
271 r->data_width = 4;
272
273 r->alloc_capable = true;
274
275 return true;
276 }
277
rdt_get_cache_alloc_cfg(int idx,struct rdt_resource * r)278 static void rdt_get_cache_alloc_cfg(int idx, struct rdt_resource *r)
279 {
280 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
281 union cpuid_0x10_1_eax eax;
282 union cpuid_0x10_x_ecx ecx;
283 union cpuid_0x10_x_edx edx;
284 u32 ebx;
285
286 cpuid_count(0x00000010, idx, &eax.full, &ebx, &ecx.full, &edx.full);
287 hw_res->num_closid = edx.split.cos_max + 1;
288 r->cache.cbm_len = eax.split.cbm_len + 1;
289 r->default_ctrl = BIT_MASK(eax.split.cbm_len + 1) - 1;
290 r->cache.shareable_bits = ebx & r->default_ctrl;
291 r->data_width = (r->cache.cbm_len + 3) / 4;
292 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
293 r->cache.arch_has_sparse_bitmasks = ecx.split.noncont;
294 r->alloc_capable = true;
295 }
296
rdt_get_cdp_config(int level)297 static void rdt_get_cdp_config(int level)
298 {
299 /*
300 * By default, CDP is disabled. CDP can be enabled by mount parameter
301 * "cdp" during resctrl file system mount time.
302 */
303 rdt_resources_all[level].cdp_enabled = false;
304 rdt_resources_all[level].r_resctrl.cdp_capable = true;
305 }
306
rdt_get_cdp_l3_config(void)307 static void rdt_get_cdp_l3_config(void)
308 {
309 rdt_get_cdp_config(RDT_RESOURCE_L3);
310 }
311
rdt_get_cdp_l2_config(void)312 static void rdt_get_cdp_l2_config(void)
313 {
314 rdt_get_cdp_config(RDT_RESOURCE_L2);
315 }
316
mba_wrmsr_amd(struct msr_param * m)317 static void mba_wrmsr_amd(struct msr_param *m)
318 {
319 struct rdt_hw_ctrl_domain *hw_dom = resctrl_to_arch_ctrl_dom(m->dom);
320 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(m->res);
321 unsigned int i;
322
323 for (i = m->low; i < m->high; i++)
324 wrmsrl(hw_res->msr_base + i, hw_dom->ctrl_val[i]);
325 }
326
327 /*
328 * Map the memory b/w percentage value to delay values
329 * that can be written to QOS_MSRs.
330 * There are currently no SKUs which support non linear delay values.
331 */
delay_bw_map(unsigned long bw,struct rdt_resource * r)332 static u32 delay_bw_map(unsigned long bw, struct rdt_resource *r)
333 {
334 if (r->membw.delay_linear)
335 return MAX_MBA_BW - bw;
336
337 pr_warn_once("Non Linear delay-bw map not supported but queried\n");
338 return r->default_ctrl;
339 }
340
mba_wrmsr_intel(struct msr_param * m)341 static void mba_wrmsr_intel(struct msr_param *m)
342 {
343 struct rdt_hw_ctrl_domain *hw_dom = resctrl_to_arch_ctrl_dom(m->dom);
344 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(m->res);
345 unsigned int i;
346
347 /* Write the delay values for mba. */
348 for (i = m->low; i < m->high; i++)
349 wrmsrl(hw_res->msr_base + i, delay_bw_map(hw_dom->ctrl_val[i], m->res));
350 }
351
cat_wrmsr(struct msr_param * m)352 static void cat_wrmsr(struct msr_param *m)
353 {
354 struct rdt_hw_ctrl_domain *hw_dom = resctrl_to_arch_ctrl_dom(m->dom);
355 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(m->res);
356 unsigned int i;
357
358 for (i = m->low; i < m->high; i++)
359 wrmsrl(hw_res->msr_base + i, hw_dom->ctrl_val[i]);
360 }
361
get_ctrl_domain_from_cpu(int cpu,struct rdt_resource * r)362 struct rdt_ctrl_domain *get_ctrl_domain_from_cpu(int cpu, struct rdt_resource *r)
363 {
364 struct rdt_ctrl_domain *d;
365
366 lockdep_assert_cpus_held();
367
368 list_for_each_entry(d, &r->ctrl_domains, hdr.list) {
369 /* Find the domain that contains this CPU */
370 if (cpumask_test_cpu(cpu, &d->hdr.cpu_mask))
371 return d;
372 }
373
374 return NULL;
375 }
376
get_mon_domain_from_cpu(int cpu,struct rdt_resource * r)377 struct rdt_mon_domain *get_mon_domain_from_cpu(int cpu, struct rdt_resource *r)
378 {
379 struct rdt_mon_domain *d;
380
381 lockdep_assert_cpus_held();
382
383 list_for_each_entry(d, &r->mon_domains, hdr.list) {
384 /* Find the domain that contains this CPU */
385 if (cpumask_test_cpu(cpu, &d->hdr.cpu_mask))
386 return d;
387 }
388
389 return NULL;
390 }
391
resctrl_arch_get_num_closid(struct rdt_resource * r)392 u32 resctrl_arch_get_num_closid(struct rdt_resource *r)
393 {
394 return resctrl_to_arch_res(r)->num_closid;
395 }
396
rdt_ctrl_update(void * arg)397 void rdt_ctrl_update(void *arg)
398 {
399 struct rdt_hw_resource *hw_res;
400 struct msr_param *m = arg;
401
402 hw_res = resctrl_to_arch_res(m->res);
403 hw_res->msr_update(m);
404 }
405
406 /*
407 * rdt_find_domain - Search for a domain id in a resource domain list.
408 *
409 * Search the domain list to find the domain id. If the domain id is
410 * found, return the domain. NULL otherwise. If the domain id is not
411 * found (and NULL returned) then the first domain with id bigger than
412 * the input id can be returned to the caller via @pos.
413 */
rdt_find_domain(struct list_head * h,int id,struct list_head ** pos)414 struct rdt_domain_hdr *rdt_find_domain(struct list_head *h, int id,
415 struct list_head **pos)
416 {
417 struct rdt_domain_hdr *d;
418 struct list_head *l;
419
420 list_for_each(l, h) {
421 d = list_entry(l, struct rdt_domain_hdr, list);
422 /* When id is found, return its domain. */
423 if (id == d->id)
424 return d;
425 /* Stop searching when finding id's position in sorted list. */
426 if (id < d->id)
427 break;
428 }
429
430 if (pos)
431 *pos = l;
432
433 return NULL;
434 }
435
setup_default_ctrlval(struct rdt_resource * r,u32 * dc)436 static void setup_default_ctrlval(struct rdt_resource *r, u32 *dc)
437 {
438 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
439 int i;
440
441 /*
442 * Initialize the Control MSRs to having no control.
443 * For Cache Allocation: Set all bits in cbm
444 * For Memory Allocation: Set b/w requested to 100%
445 */
446 for (i = 0; i < hw_res->num_closid; i++, dc++)
447 *dc = r->default_ctrl;
448 }
449
ctrl_domain_free(struct rdt_hw_ctrl_domain * hw_dom)450 static void ctrl_domain_free(struct rdt_hw_ctrl_domain *hw_dom)
451 {
452 kfree(hw_dom->ctrl_val);
453 kfree(hw_dom);
454 }
455
mon_domain_free(struct rdt_hw_mon_domain * hw_dom)456 static void mon_domain_free(struct rdt_hw_mon_domain *hw_dom)
457 {
458 kfree(hw_dom->arch_mbm_total);
459 kfree(hw_dom->arch_mbm_local);
460 kfree(hw_dom);
461 }
462
domain_setup_ctrlval(struct rdt_resource * r,struct rdt_ctrl_domain * d)463 static int domain_setup_ctrlval(struct rdt_resource *r, struct rdt_ctrl_domain *d)
464 {
465 struct rdt_hw_ctrl_domain *hw_dom = resctrl_to_arch_ctrl_dom(d);
466 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
467 struct msr_param m;
468 u32 *dc;
469
470 dc = kmalloc_array(hw_res->num_closid, sizeof(*hw_dom->ctrl_val),
471 GFP_KERNEL);
472 if (!dc)
473 return -ENOMEM;
474
475 hw_dom->ctrl_val = dc;
476 setup_default_ctrlval(r, dc);
477
478 m.res = r;
479 m.dom = d;
480 m.low = 0;
481 m.high = hw_res->num_closid;
482 hw_res->msr_update(&m);
483 return 0;
484 }
485
486 /**
487 * arch_domain_mbm_alloc() - Allocate arch private storage for the MBM counters
488 * @num_rmid: The size of the MBM counter array
489 * @hw_dom: The domain that owns the allocated arrays
490 */
arch_domain_mbm_alloc(u32 num_rmid,struct rdt_hw_mon_domain * hw_dom)491 static int arch_domain_mbm_alloc(u32 num_rmid, struct rdt_hw_mon_domain *hw_dom)
492 {
493 size_t tsize;
494
495 if (is_mbm_total_enabled()) {
496 tsize = sizeof(*hw_dom->arch_mbm_total);
497 hw_dom->arch_mbm_total = kcalloc(num_rmid, tsize, GFP_KERNEL);
498 if (!hw_dom->arch_mbm_total)
499 return -ENOMEM;
500 }
501 if (is_mbm_local_enabled()) {
502 tsize = sizeof(*hw_dom->arch_mbm_local);
503 hw_dom->arch_mbm_local = kcalloc(num_rmid, tsize, GFP_KERNEL);
504 if (!hw_dom->arch_mbm_local) {
505 kfree(hw_dom->arch_mbm_total);
506 hw_dom->arch_mbm_total = NULL;
507 return -ENOMEM;
508 }
509 }
510
511 return 0;
512 }
513
get_domain_id_from_scope(int cpu,enum resctrl_scope scope)514 static int get_domain_id_from_scope(int cpu, enum resctrl_scope scope)
515 {
516 switch (scope) {
517 case RESCTRL_L2_CACHE:
518 case RESCTRL_L3_CACHE:
519 return get_cpu_cacheinfo_id(cpu, scope);
520 case RESCTRL_L3_NODE:
521 return cpu_to_node(cpu);
522 default:
523 break;
524 }
525
526 return -EINVAL;
527 }
528
domain_add_cpu_ctrl(int cpu,struct rdt_resource * r)529 static void domain_add_cpu_ctrl(int cpu, struct rdt_resource *r)
530 {
531 int id = get_domain_id_from_scope(cpu, r->ctrl_scope);
532 struct rdt_hw_ctrl_domain *hw_dom;
533 struct list_head *add_pos = NULL;
534 struct rdt_domain_hdr *hdr;
535 struct rdt_ctrl_domain *d;
536 int err;
537
538 lockdep_assert_held(&domain_list_lock);
539
540 if (id < 0) {
541 pr_warn_once("Can't find control domain id for CPU:%d scope:%d for resource %s\n",
542 cpu, r->ctrl_scope, r->name);
543 return;
544 }
545
546 hdr = rdt_find_domain(&r->ctrl_domains, id, &add_pos);
547 if (hdr) {
548 if (WARN_ON_ONCE(hdr->type != RESCTRL_CTRL_DOMAIN))
549 return;
550 d = container_of(hdr, struct rdt_ctrl_domain, hdr);
551
552 cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
553 if (r->cache.arch_has_per_cpu_cfg)
554 rdt_domain_reconfigure_cdp(r);
555 return;
556 }
557
558 hw_dom = kzalloc_node(sizeof(*hw_dom), GFP_KERNEL, cpu_to_node(cpu));
559 if (!hw_dom)
560 return;
561
562 d = &hw_dom->d_resctrl;
563 d->hdr.id = id;
564 d->hdr.type = RESCTRL_CTRL_DOMAIN;
565 cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
566
567 rdt_domain_reconfigure_cdp(r);
568
569 if (domain_setup_ctrlval(r, d)) {
570 ctrl_domain_free(hw_dom);
571 return;
572 }
573
574 list_add_tail_rcu(&d->hdr.list, add_pos);
575
576 err = resctrl_online_ctrl_domain(r, d);
577 if (err) {
578 list_del_rcu(&d->hdr.list);
579 synchronize_rcu();
580 ctrl_domain_free(hw_dom);
581 }
582 }
583
domain_add_cpu_mon(int cpu,struct rdt_resource * r)584 static void domain_add_cpu_mon(int cpu, struct rdt_resource *r)
585 {
586 int id = get_domain_id_from_scope(cpu, r->mon_scope);
587 struct list_head *add_pos = NULL;
588 struct rdt_hw_mon_domain *hw_dom;
589 struct rdt_domain_hdr *hdr;
590 struct rdt_mon_domain *d;
591 int err;
592
593 lockdep_assert_held(&domain_list_lock);
594
595 if (id < 0) {
596 pr_warn_once("Can't find monitor domain id for CPU:%d scope:%d for resource %s\n",
597 cpu, r->mon_scope, r->name);
598 return;
599 }
600
601 hdr = rdt_find_domain(&r->mon_domains, id, &add_pos);
602 if (hdr) {
603 if (WARN_ON_ONCE(hdr->type != RESCTRL_MON_DOMAIN))
604 return;
605 d = container_of(hdr, struct rdt_mon_domain, hdr);
606
607 cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
608 return;
609 }
610
611 hw_dom = kzalloc_node(sizeof(*hw_dom), GFP_KERNEL, cpu_to_node(cpu));
612 if (!hw_dom)
613 return;
614
615 d = &hw_dom->d_resctrl;
616 d->hdr.id = id;
617 d->hdr.type = RESCTRL_MON_DOMAIN;
618 d->ci = get_cpu_cacheinfo_level(cpu, RESCTRL_L3_CACHE);
619 if (!d->ci) {
620 pr_warn_once("Can't find L3 cache for CPU:%d resource %s\n", cpu, r->name);
621 mon_domain_free(hw_dom);
622 return;
623 }
624 cpumask_set_cpu(cpu, &d->hdr.cpu_mask);
625
626 arch_mon_domain_online(r, d);
627
628 if (arch_domain_mbm_alloc(r->num_rmid, hw_dom)) {
629 mon_domain_free(hw_dom);
630 return;
631 }
632
633 list_add_tail_rcu(&d->hdr.list, add_pos);
634
635 err = resctrl_online_mon_domain(r, d);
636 if (err) {
637 list_del_rcu(&d->hdr.list);
638 synchronize_rcu();
639 mon_domain_free(hw_dom);
640 }
641 }
642
domain_add_cpu(int cpu,struct rdt_resource * r)643 static void domain_add_cpu(int cpu, struct rdt_resource *r)
644 {
645 if (r->alloc_capable)
646 domain_add_cpu_ctrl(cpu, r);
647 if (r->mon_capable)
648 domain_add_cpu_mon(cpu, r);
649 }
650
domain_remove_cpu_ctrl(int cpu,struct rdt_resource * r)651 static void domain_remove_cpu_ctrl(int cpu, struct rdt_resource *r)
652 {
653 int id = get_domain_id_from_scope(cpu, r->ctrl_scope);
654 struct rdt_hw_ctrl_domain *hw_dom;
655 struct rdt_domain_hdr *hdr;
656 struct rdt_ctrl_domain *d;
657
658 lockdep_assert_held(&domain_list_lock);
659
660 if (id < 0) {
661 pr_warn_once("Can't find control domain id for CPU:%d scope:%d for resource %s\n",
662 cpu, r->ctrl_scope, r->name);
663 return;
664 }
665
666 hdr = rdt_find_domain(&r->ctrl_domains, id, NULL);
667 if (!hdr) {
668 pr_warn("Can't find control domain for id=%d for CPU %d for resource %s\n",
669 id, cpu, r->name);
670 return;
671 }
672
673 if (WARN_ON_ONCE(hdr->type != RESCTRL_CTRL_DOMAIN))
674 return;
675
676 d = container_of(hdr, struct rdt_ctrl_domain, hdr);
677 hw_dom = resctrl_to_arch_ctrl_dom(d);
678
679 cpumask_clear_cpu(cpu, &d->hdr.cpu_mask);
680 if (cpumask_empty(&d->hdr.cpu_mask)) {
681 resctrl_offline_ctrl_domain(r, d);
682 list_del_rcu(&d->hdr.list);
683 synchronize_rcu();
684
685 /*
686 * rdt_ctrl_domain "d" is going to be freed below, so clear
687 * its pointer from pseudo_lock_region struct.
688 */
689 if (d->plr)
690 d->plr->d = NULL;
691 ctrl_domain_free(hw_dom);
692
693 return;
694 }
695 }
696
domain_remove_cpu_mon(int cpu,struct rdt_resource * r)697 static void domain_remove_cpu_mon(int cpu, struct rdt_resource *r)
698 {
699 int id = get_domain_id_from_scope(cpu, r->mon_scope);
700 struct rdt_hw_mon_domain *hw_dom;
701 struct rdt_domain_hdr *hdr;
702 struct rdt_mon_domain *d;
703
704 lockdep_assert_held(&domain_list_lock);
705
706 if (id < 0) {
707 pr_warn_once("Can't find monitor domain id for CPU:%d scope:%d for resource %s\n",
708 cpu, r->mon_scope, r->name);
709 return;
710 }
711
712 hdr = rdt_find_domain(&r->mon_domains, id, NULL);
713 if (!hdr) {
714 pr_warn("Can't find monitor domain for id=%d for CPU %d for resource %s\n",
715 id, cpu, r->name);
716 return;
717 }
718
719 if (WARN_ON_ONCE(hdr->type != RESCTRL_MON_DOMAIN))
720 return;
721
722 d = container_of(hdr, struct rdt_mon_domain, hdr);
723 hw_dom = resctrl_to_arch_mon_dom(d);
724
725 cpumask_clear_cpu(cpu, &d->hdr.cpu_mask);
726 if (cpumask_empty(&d->hdr.cpu_mask)) {
727 resctrl_offline_mon_domain(r, d);
728 list_del_rcu(&d->hdr.list);
729 synchronize_rcu();
730 mon_domain_free(hw_dom);
731
732 return;
733 }
734 }
735
domain_remove_cpu(int cpu,struct rdt_resource * r)736 static void domain_remove_cpu(int cpu, struct rdt_resource *r)
737 {
738 if (r->alloc_capable)
739 domain_remove_cpu_ctrl(cpu, r);
740 if (r->mon_capable)
741 domain_remove_cpu_mon(cpu, r);
742 }
743
clear_closid_rmid(int cpu)744 static void clear_closid_rmid(int cpu)
745 {
746 struct resctrl_pqr_state *state = this_cpu_ptr(&pqr_state);
747
748 state->default_closid = RESCTRL_RESERVED_CLOSID;
749 state->default_rmid = RESCTRL_RESERVED_RMID;
750 state->cur_closid = RESCTRL_RESERVED_CLOSID;
751 state->cur_rmid = RESCTRL_RESERVED_RMID;
752 wrmsr(MSR_IA32_PQR_ASSOC, RESCTRL_RESERVED_RMID,
753 RESCTRL_RESERVED_CLOSID);
754 }
755
resctrl_arch_online_cpu(unsigned int cpu)756 static int resctrl_arch_online_cpu(unsigned int cpu)
757 {
758 struct rdt_resource *r;
759
760 mutex_lock(&domain_list_lock);
761 for_each_capable_rdt_resource(r)
762 domain_add_cpu(cpu, r);
763 mutex_unlock(&domain_list_lock);
764
765 clear_closid_rmid(cpu);
766 resctrl_online_cpu(cpu);
767
768 return 0;
769 }
770
resctrl_arch_offline_cpu(unsigned int cpu)771 static int resctrl_arch_offline_cpu(unsigned int cpu)
772 {
773 struct rdt_resource *r;
774
775 resctrl_offline_cpu(cpu);
776
777 mutex_lock(&domain_list_lock);
778 for_each_capable_rdt_resource(r)
779 domain_remove_cpu(cpu, r);
780 mutex_unlock(&domain_list_lock);
781
782 clear_closid_rmid(cpu);
783
784 return 0;
785 }
786
787 /*
788 * Choose a width for the resource name and resource data based on the
789 * resource that has widest name and cbm.
790 */
rdt_init_padding(void)791 static __init void rdt_init_padding(void)
792 {
793 struct rdt_resource *r;
794
795 for_each_alloc_capable_rdt_resource(r) {
796 if (r->data_width > max_data_width)
797 max_data_width = r->data_width;
798 }
799 }
800
801 enum {
802 RDT_FLAG_CMT,
803 RDT_FLAG_MBM_TOTAL,
804 RDT_FLAG_MBM_LOCAL,
805 RDT_FLAG_L3_CAT,
806 RDT_FLAG_L3_CDP,
807 RDT_FLAG_L2_CAT,
808 RDT_FLAG_L2_CDP,
809 RDT_FLAG_MBA,
810 RDT_FLAG_SMBA,
811 RDT_FLAG_BMEC,
812 };
813
814 #define RDT_OPT(idx, n, f) \
815 [idx] = { \
816 .name = n, \
817 .flag = f \
818 }
819
820 struct rdt_options {
821 char *name;
822 int flag;
823 bool force_off, force_on;
824 };
825
826 static struct rdt_options rdt_options[] __initdata = {
827 RDT_OPT(RDT_FLAG_CMT, "cmt", X86_FEATURE_CQM_OCCUP_LLC),
828 RDT_OPT(RDT_FLAG_MBM_TOTAL, "mbmtotal", X86_FEATURE_CQM_MBM_TOTAL),
829 RDT_OPT(RDT_FLAG_MBM_LOCAL, "mbmlocal", X86_FEATURE_CQM_MBM_LOCAL),
830 RDT_OPT(RDT_FLAG_L3_CAT, "l3cat", X86_FEATURE_CAT_L3),
831 RDT_OPT(RDT_FLAG_L3_CDP, "l3cdp", X86_FEATURE_CDP_L3),
832 RDT_OPT(RDT_FLAG_L2_CAT, "l2cat", X86_FEATURE_CAT_L2),
833 RDT_OPT(RDT_FLAG_L2_CDP, "l2cdp", X86_FEATURE_CDP_L2),
834 RDT_OPT(RDT_FLAG_MBA, "mba", X86_FEATURE_MBA),
835 RDT_OPT(RDT_FLAG_SMBA, "smba", X86_FEATURE_SMBA),
836 RDT_OPT(RDT_FLAG_BMEC, "bmec", X86_FEATURE_BMEC),
837 };
838 #define NUM_RDT_OPTIONS ARRAY_SIZE(rdt_options)
839
set_rdt_options(char * str)840 static int __init set_rdt_options(char *str)
841 {
842 struct rdt_options *o;
843 bool force_off;
844 char *tok;
845
846 if (*str == '=')
847 str++;
848 while ((tok = strsep(&str, ",")) != NULL) {
849 force_off = *tok == '!';
850 if (force_off)
851 tok++;
852 for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
853 if (strcmp(tok, o->name) == 0) {
854 if (force_off)
855 o->force_off = true;
856 else
857 o->force_on = true;
858 break;
859 }
860 }
861 }
862 return 1;
863 }
864 __setup("rdt", set_rdt_options);
865
rdt_cpu_has(int flag)866 bool __init rdt_cpu_has(int flag)
867 {
868 bool ret = boot_cpu_has(flag);
869 struct rdt_options *o;
870
871 if (!ret)
872 return ret;
873
874 for (o = rdt_options; o < &rdt_options[NUM_RDT_OPTIONS]; o++) {
875 if (flag == o->flag) {
876 if (o->force_off)
877 ret = false;
878 if (o->force_on)
879 ret = true;
880 break;
881 }
882 }
883 return ret;
884 }
885
get_mem_config(void)886 static __init bool get_mem_config(void)
887 {
888 struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_MBA];
889
890 if (!rdt_cpu_has(X86_FEATURE_MBA))
891 return false;
892
893 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
894 return __get_mem_config_intel(&hw_res->r_resctrl);
895 else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
896 return __rdt_get_mem_config_amd(&hw_res->r_resctrl);
897
898 return false;
899 }
900
get_slow_mem_config(void)901 static __init bool get_slow_mem_config(void)
902 {
903 struct rdt_hw_resource *hw_res = &rdt_resources_all[RDT_RESOURCE_SMBA];
904
905 if (!rdt_cpu_has(X86_FEATURE_SMBA))
906 return false;
907
908 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
909 return __rdt_get_mem_config_amd(&hw_res->r_resctrl);
910
911 return false;
912 }
913
get_rdt_alloc_resources(void)914 static __init bool get_rdt_alloc_resources(void)
915 {
916 struct rdt_resource *r;
917 bool ret = false;
918
919 if (rdt_alloc_capable)
920 return true;
921
922 if (!boot_cpu_has(X86_FEATURE_RDT_A))
923 return false;
924
925 if (rdt_cpu_has(X86_FEATURE_CAT_L3)) {
926 r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
927 rdt_get_cache_alloc_cfg(1, r);
928 if (rdt_cpu_has(X86_FEATURE_CDP_L3))
929 rdt_get_cdp_l3_config();
930 ret = true;
931 }
932 if (rdt_cpu_has(X86_FEATURE_CAT_L2)) {
933 /* CPUID 0x10.2 fields are same format at 0x10.1 */
934 r = &rdt_resources_all[RDT_RESOURCE_L2].r_resctrl;
935 rdt_get_cache_alloc_cfg(2, r);
936 if (rdt_cpu_has(X86_FEATURE_CDP_L2))
937 rdt_get_cdp_l2_config();
938 ret = true;
939 }
940
941 if (get_mem_config())
942 ret = true;
943
944 if (get_slow_mem_config())
945 ret = true;
946
947 return ret;
948 }
949
get_rdt_mon_resources(void)950 static __init bool get_rdt_mon_resources(void)
951 {
952 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
953
954 if (rdt_cpu_has(X86_FEATURE_CQM_OCCUP_LLC))
955 rdt_mon_features |= (1 << QOS_L3_OCCUP_EVENT_ID);
956 if (rdt_cpu_has(X86_FEATURE_CQM_MBM_TOTAL))
957 rdt_mon_features |= (1 << QOS_L3_MBM_TOTAL_EVENT_ID);
958 if (rdt_cpu_has(X86_FEATURE_CQM_MBM_LOCAL))
959 rdt_mon_features |= (1 << QOS_L3_MBM_LOCAL_EVENT_ID);
960
961 if (!rdt_mon_features)
962 return false;
963
964 return !rdt_get_mon_l3_config(r);
965 }
966
__check_quirks_intel(void)967 static __init void __check_quirks_intel(void)
968 {
969 switch (boot_cpu_data.x86_vfm) {
970 case INTEL_HASWELL_X:
971 if (!rdt_options[RDT_FLAG_L3_CAT].force_off)
972 cache_alloc_hsw_probe();
973 break;
974 case INTEL_SKYLAKE_X:
975 if (boot_cpu_data.x86_stepping <= 4)
976 set_rdt_options("!cmt,!mbmtotal,!mbmlocal,!l3cat");
977 else
978 set_rdt_options("!l3cat");
979 fallthrough;
980 case INTEL_BROADWELL_X:
981 intel_rdt_mbm_apply_quirk();
982 break;
983 }
984 }
985
check_quirks(void)986 static __init void check_quirks(void)
987 {
988 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
989 __check_quirks_intel();
990 }
991
get_rdt_resources(void)992 static __init bool get_rdt_resources(void)
993 {
994 rdt_alloc_capable = get_rdt_alloc_resources();
995 rdt_mon_capable = get_rdt_mon_resources();
996
997 return (rdt_mon_capable || rdt_alloc_capable);
998 }
999
rdt_init_res_defs_intel(void)1000 static __init void rdt_init_res_defs_intel(void)
1001 {
1002 struct rdt_hw_resource *hw_res;
1003 struct rdt_resource *r;
1004
1005 for_each_rdt_resource(r) {
1006 hw_res = resctrl_to_arch_res(r);
1007
1008 if (r->rid == RDT_RESOURCE_L3 ||
1009 r->rid == RDT_RESOURCE_L2) {
1010 r->cache.arch_has_per_cpu_cfg = false;
1011 r->cache.min_cbm_bits = 1;
1012 } else if (r->rid == RDT_RESOURCE_MBA) {
1013 hw_res->msr_base = MSR_IA32_MBA_THRTL_BASE;
1014 hw_res->msr_update = mba_wrmsr_intel;
1015 }
1016 }
1017 }
1018
rdt_init_res_defs_amd(void)1019 static __init void rdt_init_res_defs_amd(void)
1020 {
1021 struct rdt_hw_resource *hw_res;
1022 struct rdt_resource *r;
1023
1024 for_each_rdt_resource(r) {
1025 hw_res = resctrl_to_arch_res(r);
1026
1027 if (r->rid == RDT_RESOURCE_L3 ||
1028 r->rid == RDT_RESOURCE_L2) {
1029 r->cache.arch_has_sparse_bitmasks = true;
1030 r->cache.arch_has_per_cpu_cfg = true;
1031 r->cache.min_cbm_bits = 0;
1032 } else if (r->rid == RDT_RESOURCE_MBA) {
1033 hw_res->msr_base = MSR_IA32_MBA_BW_BASE;
1034 hw_res->msr_update = mba_wrmsr_amd;
1035 } else if (r->rid == RDT_RESOURCE_SMBA) {
1036 hw_res->msr_base = MSR_IA32_SMBA_BW_BASE;
1037 hw_res->msr_update = mba_wrmsr_amd;
1038 }
1039 }
1040 }
1041
rdt_init_res_defs(void)1042 static __init void rdt_init_res_defs(void)
1043 {
1044 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
1045 rdt_init_res_defs_intel();
1046 else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
1047 rdt_init_res_defs_amd();
1048 }
1049
1050 static enum cpuhp_state rdt_online;
1051
1052 /* Runs once on the BSP during boot. */
resctrl_cpu_detect(struct cpuinfo_x86 * c)1053 void resctrl_cpu_detect(struct cpuinfo_x86 *c)
1054 {
1055 if (!cpu_has(c, X86_FEATURE_CQM_LLC)) {
1056 c->x86_cache_max_rmid = -1;
1057 c->x86_cache_occ_scale = -1;
1058 c->x86_cache_mbm_width_offset = -1;
1059 return;
1060 }
1061
1062 /* will be overridden if occupancy monitoring exists */
1063 c->x86_cache_max_rmid = cpuid_ebx(0xf);
1064
1065 if (cpu_has(c, X86_FEATURE_CQM_OCCUP_LLC) ||
1066 cpu_has(c, X86_FEATURE_CQM_MBM_TOTAL) ||
1067 cpu_has(c, X86_FEATURE_CQM_MBM_LOCAL)) {
1068 u32 eax, ebx, ecx, edx;
1069
1070 /* QoS sub-leaf, EAX=0Fh, ECX=1 */
1071 cpuid_count(0xf, 1, &eax, &ebx, &ecx, &edx);
1072
1073 c->x86_cache_max_rmid = ecx;
1074 c->x86_cache_occ_scale = ebx;
1075 c->x86_cache_mbm_width_offset = eax & 0xff;
1076
1077 if (c->x86_vendor == X86_VENDOR_AMD && !c->x86_cache_mbm_width_offset)
1078 c->x86_cache_mbm_width_offset = MBM_CNTR_WIDTH_OFFSET_AMD;
1079 }
1080 }
1081
resctrl_late_init(void)1082 static int __init resctrl_late_init(void)
1083 {
1084 struct rdt_resource *r;
1085 int state, ret;
1086
1087 /*
1088 * Initialize functions(or definitions) that are different
1089 * between vendors here.
1090 */
1091 rdt_init_res_defs();
1092
1093 check_quirks();
1094
1095 if (!get_rdt_resources())
1096 return -ENODEV;
1097
1098 rdt_init_padding();
1099
1100 state = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN,
1101 "x86/resctrl/cat:online:",
1102 resctrl_arch_online_cpu,
1103 resctrl_arch_offline_cpu);
1104 if (state < 0)
1105 return state;
1106
1107 ret = rdtgroup_init();
1108 if (ret) {
1109 cpuhp_remove_state(state);
1110 return ret;
1111 }
1112 rdt_online = state;
1113
1114 for_each_alloc_capable_rdt_resource(r)
1115 pr_info("%s allocation detected\n", r->name);
1116
1117 for_each_mon_capable_rdt_resource(r)
1118 pr_info("%s monitoring detected\n", r->name);
1119
1120 return 0;
1121 }
1122
1123 late_initcall(resctrl_late_init);
1124
resctrl_exit(void)1125 static void __exit resctrl_exit(void)
1126 {
1127 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
1128
1129 cpuhp_remove_state(rdt_online);
1130
1131 rdtgroup_exit();
1132
1133 if (r->mon_capable)
1134 rdt_put_mon_l3_config();
1135 }
1136
1137 __exitcall(resctrl_exit);
1138