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 *
12 * More information about RDT be found in the Intel (R) x86 Architecture
13 * Software Developer Manual June 2016, volume 3, section 17.17.
14 */
15
16 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17
18 #include <linux/cpu.h>
19 #include <linux/kernfs.h>
20 #include <linux/seq_file.h>
21 #include <linux/slab.h>
22 #include <linux/tick.h>
23
24 #include "internal.h"
25
26 /*
27 * Check whether MBA bandwidth percentage value is correct. The value is
28 * checked against the minimum and max bandwidth values specified by the
29 * hardware. The allocated bandwidth percentage is rounded to the next
30 * control step available on the hardware.
31 */
bw_validate(char * buf,unsigned long * data,struct rdt_resource * r)32 static bool bw_validate(char *buf, unsigned long *data, struct rdt_resource *r)
33 {
34 unsigned long bw;
35 int ret;
36
37 /*
38 * Only linear delay values is supported for current Intel SKUs.
39 */
40 if (!r->membw.delay_linear && r->membw.arch_needs_linear) {
41 rdt_last_cmd_puts("No support for non-linear MB domains\n");
42 return false;
43 }
44
45 ret = kstrtoul(buf, 10, &bw);
46 if (ret) {
47 rdt_last_cmd_printf("Non-decimal digit in MB value %s\n", buf);
48 return false;
49 }
50
51 if ((bw < r->membw.min_bw || bw > r->default_ctrl) &&
52 !is_mba_sc(r)) {
53 rdt_last_cmd_printf("MB value %ld out of range [%d,%d]\n", bw,
54 r->membw.min_bw, r->default_ctrl);
55 return false;
56 }
57
58 *data = roundup(bw, (unsigned long)r->membw.bw_gran);
59 return true;
60 }
61
parse_bw(struct rdt_parse_data * data,struct resctrl_schema * s,struct rdt_domain * d)62 int parse_bw(struct rdt_parse_data *data, struct resctrl_schema *s,
63 struct rdt_domain *d)
64 {
65 struct resctrl_staged_config *cfg;
66 u32 closid = data->rdtgrp->closid;
67 struct rdt_resource *r = s->res;
68 unsigned long bw_val;
69
70 cfg = &d->staged_config[s->conf_type];
71 if (cfg->have_new_ctrl) {
72 rdt_last_cmd_printf("Duplicate domain %d\n", d->id);
73 return -EINVAL;
74 }
75
76 if (!bw_validate(data->buf, &bw_val, r))
77 return -EINVAL;
78
79 if (is_mba_sc(r)) {
80 d->mbps_val[closid] = bw_val;
81 return 0;
82 }
83
84 cfg->new_ctrl = bw_val;
85 cfg->have_new_ctrl = true;
86
87 return 0;
88 }
89
90 /*
91 * Check whether a cache bit mask is valid.
92 * On Intel CPUs, non-contiguous 1s value support is indicated by CPUID:
93 * - CPUID.0x10.1:ECX[3]: L3 non-contiguous 1s value supported if 1
94 * - CPUID.0x10.2:ECX[3]: L2 non-contiguous 1s value supported if 1
95 *
96 * Haswell does not support a non-contiguous 1s value and additionally
97 * requires at least two bits set.
98 * AMD allows non-contiguous bitmasks.
99 */
cbm_validate(char * buf,u32 * data,struct rdt_resource * r)100 static bool cbm_validate(char *buf, u32 *data, struct rdt_resource *r)
101 {
102 unsigned long first_bit, zero_bit, val;
103 unsigned int cbm_len = r->cache.cbm_len;
104 int ret;
105
106 ret = kstrtoul(buf, 16, &val);
107 if (ret) {
108 rdt_last_cmd_printf("Non-hex character in the mask %s\n", buf);
109 return false;
110 }
111
112 if ((r->cache.min_cbm_bits > 0 && val == 0) || val > r->default_ctrl) {
113 rdt_last_cmd_puts("Mask out of range\n");
114 return false;
115 }
116
117 first_bit = find_first_bit(&val, cbm_len);
118 zero_bit = find_next_zero_bit(&val, cbm_len, first_bit);
119
120 /* Are non-contiguous bitmasks allowed? */
121 if (!r->cache.arch_has_sparse_bitmasks &&
122 (find_next_bit(&val, cbm_len, zero_bit) < cbm_len)) {
123 rdt_last_cmd_printf("The mask %lx has non-consecutive 1-bits\n", val);
124 return false;
125 }
126
127 if ((zero_bit - first_bit) < r->cache.min_cbm_bits) {
128 rdt_last_cmd_printf("Need at least %d bits in the mask\n",
129 r->cache.min_cbm_bits);
130 return false;
131 }
132
133 *data = val;
134 return true;
135 }
136
137 /*
138 * Read one cache bit mask (hex). Check that it is valid for the current
139 * resource type.
140 */
parse_cbm(struct rdt_parse_data * data,struct resctrl_schema * s,struct rdt_domain * d)141 int parse_cbm(struct rdt_parse_data *data, struct resctrl_schema *s,
142 struct rdt_domain *d)
143 {
144 struct rdtgroup *rdtgrp = data->rdtgrp;
145 struct resctrl_staged_config *cfg;
146 struct rdt_resource *r = s->res;
147 u32 cbm_val;
148
149 cfg = &d->staged_config[s->conf_type];
150 if (cfg->have_new_ctrl) {
151 rdt_last_cmd_printf("Duplicate domain %d\n", d->id);
152 return -EINVAL;
153 }
154
155 /*
156 * Cannot set up more than one pseudo-locked region in a cache
157 * hierarchy.
158 */
159 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP &&
160 rdtgroup_pseudo_locked_in_hierarchy(d)) {
161 rdt_last_cmd_puts("Pseudo-locked region in hierarchy\n");
162 return -EINVAL;
163 }
164
165 if (!cbm_validate(data->buf, &cbm_val, r))
166 return -EINVAL;
167
168 if ((rdtgrp->mode == RDT_MODE_EXCLUSIVE ||
169 rdtgrp->mode == RDT_MODE_SHAREABLE) &&
170 rdtgroup_cbm_overlaps_pseudo_locked(d, cbm_val)) {
171 rdt_last_cmd_puts("CBM overlaps with pseudo-locked region\n");
172 return -EINVAL;
173 }
174
175 /*
176 * The CBM may not overlap with the CBM of another closid if
177 * either is exclusive.
178 */
179 if (rdtgroup_cbm_overlaps(s, d, cbm_val, rdtgrp->closid, true)) {
180 rdt_last_cmd_puts("Overlaps with exclusive group\n");
181 return -EINVAL;
182 }
183
184 if (rdtgroup_cbm_overlaps(s, d, cbm_val, rdtgrp->closid, false)) {
185 if (rdtgrp->mode == RDT_MODE_EXCLUSIVE ||
186 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
187 rdt_last_cmd_puts("Overlaps with other group\n");
188 return -EINVAL;
189 }
190 }
191
192 cfg->new_ctrl = cbm_val;
193 cfg->have_new_ctrl = true;
194
195 return 0;
196 }
197
198 /*
199 * For each domain in this resource we expect to find a series of:
200 * id=mask
201 * separated by ";". The "id" is in decimal, and must match one of
202 * the "id"s for this resource.
203 */
parse_line(char * line,struct resctrl_schema * s,struct rdtgroup * rdtgrp)204 static int parse_line(char *line, struct resctrl_schema *s,
205 struct rdtgroup *rdtgrp)
206 {
207 enum resctrl_conf_type t = s->conf_type;
208 struct resctrl_staged_config *cfg;
209 struct rdt_resource *r = s->res;
210 struct rdt_parse_data data;
211 char *dom = NULL, *id;
212 struct rdt_domain *d;
213 unsigned long dom_id;
214
215 /* Walking r->domains, ensure it can't race with cpuhp */
216 lockdep_assert_cpus_held();
217
218 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP &&
219 (r->rid == RDT_RESOURCE_MBA || r->rid == RDT_RESOURCE_SMBA)) {
220 rdt_last_cmd_puts("Cannot pseudo-lock MBA resource\n");
221 return -EINVAL;
222 }
223
224 next:
225 if (!line || line[0] == '\0')
226 return 0;
227 dom = strsep(&line, ";");
228 id = strsep(&dom, "=");
229 if (!dom || kstrtoul(id, 10, &dom_id)) {
230 rdt_last_cmd_puts("Missing '=' or non-numeric domain\n");
231 return -EINVAL;
232 }
233 dom = strim(dom);
234 list_for_each_entry(d, &r->domains, list) {
235 if (d->id == dom_id) {
236 data.buf = dom;
237 data.rdtgrp = rdtgrp;
238 if (r->parse_ctrlval(&data, s, d))
239 return -EINVAL;
240 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
241 cfg = &d->staged_config[t];
242 /*
243 * In pseudo-locking setup mode and just
244 * parsed a valid CBM that should be
245 * pseudo-locked. Only one locked region per
246 * resource group and domain so just do
247 * the required initialization for single
248 * region and return.
249 */
250 rdtgrp->plr->s = s;
251 rdtgrp->plr->d = d;
252 rdtgrp->plr->cbm = cfg->new_ctrl;
253 d->plr = rdtgrp->plr;
254 return 0;
255 }
256 goto next;
257 }
258 }
259 return -EINVAL;
260 }
261
get_config_index(u32 closid,enum resctrl_conf_type type)262 static u32 get_config_index(u32 closid, enum resctrl_conf_type type)
263 {
264 switch (type) {
265 default:
266 case CDP_NONE:
267 return closid;
268 case CDP_CODE:
269 return closid * 2 + 1;
270 case CDP_DATA:
271 return closid * 2;
272 }
273 }
274
resctrl_arch_update_one(struct rdt_resource * r,struct rdt_domain * d,u32 closid,enum resctrl_conf_type t,u32 cfg_val)275 int resctrl_arch_update_one(struct rdt_resource *r, struct rdt_domain *d,
276 u32 closid, enum resctrl_conf_type t, u32 cfg_val)
277 {
278 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
279 struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
280 u32 idx = get_config_index(closid, t);
281 struct msr_param msr_param;
282
283 if (!cpumask_test_cpu(smp_processor_id(), &d->cpu_mask))
284 return -EINVAL;
285
286 hw_dom->ctrl_val[idx] = cfg_val;
287
288 msr_param.res = r;
289 msr_param.dom = d;
290 msr_param.low = idx;
291 msr_param.high = idx + 1;
292 hw_res->msr_update(&msr_param);
293
294 return 0;
295 }
296
resctrl_arch_update_domains(struct rdt_resource * r,u32 closid)297 int resctrl_arch_update_domains(struct rdt_resource *r, u32 closid)
298 {
299 struct resctrl_staged_config *cfg;
300 struct rdt_hw_domain *hw_dom;
301 struct msr_param msr_param;
302 enum resctrl_conf_type t;
303 struct rdt_domain *d;
304 u32 idx;
305
306 /* Walking r->domains, ensure it can't race with cpuhp */
307 lockdep_assert_cpus_held();
308
309 list_for_each_entry(d, &r->domains, list) {
310 hw_dom = resctrl_to_arch_dom(d);
311 msr_param.res = NULL;
312 for (t = 0; t < CDP_NUM_TYPES; t++) {
313 cfg = &hw_dom->d_resctrl.staged_config[t];
314 if (!cfg->have_new_ctrl)
315 continue;
316
317 idx = get_config_index(closid, t);
318 if (cfg->new_ctrl == hw_dom->ctrl_val[idx])
319 continue;
320 hw_dom->ctrl_val[idx] = cfg->new_ctrl;
321
322 if (!msr_param.res) {
323 msr_param.low = idx;
324 msr_param.high = msr_param.low + 1;
325 msr_param.res = r;
326 msr_param.dom = d;
327 } else {
328 msr_param.low = min(msr_param.low, idx);
329 msr_param.high = max(msr_param.high, idx + 1);
330 }
331 }
332 if (msr_param.res)
333 smp_call_function_any(&d->cpu_mask, rdt_ctrl_update, &msr_param, 1);
334 }
335
336 return 0;
337 }
338
rdtgroup_parse_resource(char * resname,char * tok,struct rdtgroup * rdtgrp)339 static int rdtgroup_parse_resource(char *resname, char *tok,
340 struct rdtgroup *rdtgrp)
341 {
342 struct resctrl_schema *s;
343
344 list_for_each_entry(s, &resctrl_schema_all, list) {
345 if (!strcmp(resname, s->name) && rdtgrp->closid < s->num_closid)
346 return parse_line(tok, s, rdtgrp);
347 }
348 rdt_last_cmd_printf("Unknown or unsupported resource name '%s'\n", resname);
349 return -EINVAL;
350 }
351
rdtgroup_schemata_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)352 ssize_t rdtgroup_schemata_write(struct kernfs_open_file *of,
353 char *buf, size_t nbytes, loff_t off)
354 {
355 struct resctrl_schema *s;
356 struct rdtgroup *rdtgrp;
357 struct rdt_resource *r;
358 char *tok, *resname;
359 int ret = 0;
360
361 /* Valid input requires a trailing newline */
362 if (nbytes == 0 || buf[nbytes - 1] != '\n')
363 return -EINVAL;
364 buf[nbytes - 1] = '\0';
365
366 rdtgrp = rdtgroup_kn_lock_live(of->kn);
367 if (!rdtgrp) {
368 rdtgroup_kn_unlock(of->kn);
369 return -ENOENT;
370 }
371 rdt_last_cmd_clear();
372
373 /*
374 * No changes to pseudo-locked region allowed. It has to be removed
375 * and re-created instead.
376 */
377 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
378 ret = -EINVAL;
379 rdt_last_cmd_puts("Resource group is pseudo-locked\n");
380 goto out;
381 }
382
383 rdt_staged_configs_clear();
384
385 while ((tok = strsep(&buf, "\n")) != NULL) {
386 resname = strim(strsep(&tok, ":"));
387 if (!tok) {
388 rdt_last_cmd_puts("Missing ':'\n");
389 ret = -EINVAL;
390 goto out;
391 }
392 if (tok[0] == '\0') {
393 rdt_last_cmd_printf("Missing '%s' value\n", resname);
394 ret = -EINVAL;
395 goto out;
396 }
397 ret = rdtgroup_parse_resource(resname, tok, rdtgrp);
398 if (ret)
399 goto out;
400 }
401
402 list_for_each_entry(s, &resctrl_schema_all, list) {
403 r = s->res;
404
405 /*
406 * Writes to mba_sc resources update the software controller,
407 * not the control MSR.
408 */
409 if (is_mba_sc(r))
410 continue;
411
412 ret = resctrl_arch_update_domains(r, rdtgrp->closid);
413 if (ret)
414 goto out;
415 }
416
417 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
418 /*
419 * If pseudo-locking fails we keep the resource group in
420 * mode RDT_MODE_PSEUDO_LOCKSETUP with its class of service
421 * active and updated for just the domain the pseudo-locked
422 * region was requested for.
423 */
424 ret = rdtgroup_pseudo_lock_create(rdtgrp);
425 }
426
427 out:
428 rdt_staged_configs_clear();
429 rdtgroup_kn_unlock(of->kn);
430 return ret ?: nbytes;
431 }
432
resctrl_arch_get_config(struct rdt_resource * r,struct rdt_domain * d,u32 closid,enum resctrl_conf_type type)433 u32 resctrl_arch_get_config(struct rdt_resource *r, struct rdt_domain *d,
434 u32 closid, enum resctrl_conf_type type)
435 {
436 struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
437 u32 idx = get_config_index(closid, type);
438
439 return hw_dom->ctrl_val[idx];
440 }
441
show_doms(struct seq_file * s,struct resctrl_schema * schema,int closid)442 static void show_doms(struct seq_file *s, struct resctrl_schema *schema, int closid)
443 {
444 struct rdt_resource *r = schema->res;
445 struct rdt_domain *dom;
446 bool sep = false;
447 u32 ctrl_val;
448
449 /* Walking r->domains, ensure it can't race with cpuhp */
450 lockdep_assert_cpus_held();
451
452 seq_printf(s, "%*s:", max_name_width, schema->name);
453 list_for_each_entry(dom, &r->domains, list) {
454 if (sep)
455 seq_puts(s, ";");
456
457 if (is_mba_sc(r))
458 ctrl_val = dom->mbps_val[closid];
459 else
460 ctrl_val = resctrl_arch_get_config(r, dom, closid,
461 schema->conf_type);
462
463 seq_printf(s, r->format_str, dom->id, max_data_width,
464 ctrl_val);
465 sep = true;
466 }
467 seq_puts(s, "\n");
468 }
469
rdtgroup_schemata_show(struct kernfs_open_file * of,struct seq_file * s,void * v)470 int rdtgroup_schemata_show(struct kernfs_open_file *of,
471 struct seq_file *s, void *v)
472 {
473 struct resctrl_schema *schema;
474 struct rdtgroup *rdtgrp;
475 int ret = 0;
476 u32 closid;
477
478 rdtgrp = rdtgroup_kn_lock_live(of->kn);
479 if (rdtgrp) {
480 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
481 list_for_each_entry(schema, &resctrl_schema_all, list) {
482 seq_printf(s, "%s:uninitialized\n", schema->name);
483 }
484 } else if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
485 if (!rdtgrp->plr->d) {
486 rdt_last_cmd_clear();
487 rdt_last_cmd_puts("Cache domain offline\n");
488 ret = -ENODEV;
489 } else {
490 seq_printf(s, "%s:%d=%x\n",
491 rdtgrp->plr->s->res->name,
492 rdtgrp->plr->d->id,
493 rdtgrp->plr->cbm);
494 }
495 } else {
496 closid = rdtgrp->closid;
497 list_for_each_entry(schema, &resctrl_schema_all, list) {
498 if (closid < schema->num_closid)
499 show_doms(s, schema, closid);
500 }
501 }
502 } else {
503 ret = -ENOENT;
504 }
505 rdtgroup_kn_unlock(of->kn);
506 return ret;
507 }
508
smp_mon_event_count(void * arg)509 static int smp_mon_event_count(void *arg)
510 {
511 mon_event_count(arg);
512
513 return 0;
514 }
515
mon_event_read(struct rmid_read * rr,struct rdt_resource * r,struct rdt_domain * d,struct rdtgroup * rdtgrp,int evtid,int first)516 void mon_event_read(struct rmid_read *rr, struct rdt_resource *r,
517 struct rdt_domain *d, struct rdtgroup *rdtgrp,
518 int evtid, int first)
519 {
520 int cpu;
521
522 /* When picking a CPU from cpu_mask, ensure it can't race with cpuhp */
523 lockdep_assert_cpus_held();
524
525 /*
526 * Setup the parameters to pass to mon_event_count() to read the data.
527 */
528 rr->rgrp = rdtgrp;
529 rr->evtid = evtid;
530 rr->r = r;
531 rr->d = d;
532 rr->val = 0;
533 rr->first = first;
534 rr->arch_mon_ctx = resctrl_arch_mon_ctx_alloc(r, evtid);
535 if (IS_ERR(rr->arch_mon_ctx)) {
536 rr->err = -EINVAL;
537 return;
538 }
539
540 cpu = cpumask_any_housekeeping(&d->cpu_mask, RESCTRL_PICK_ANY_CPU);
541
542 /*
543 * cpumask_any_housekeeping() prefers housekeeping CPUs, but
544 * are all the CPUs nohz_full? If yes, pick a CPU to IPI.
545 * MPAM's resctrl_arch_rmid_read() is unable to read the
546 * counters on some platforms if its called in IRQ context.
547 */
548 if (tick_nohz_full_cpu(cpu))
549 smp_call_function_any(&d->cpu_mask, mon_event_count, rr, 1);
550 else
551 smp_call_on_cpu(cpu, smp_mon_event_count, rr, false);
552
553 resctrl_arch_mon_ctx_free(r, evtid, rr->arch_mon_ctx);
554 }
555
rdtgroup_mondata_show(struct seq_file * m,void * arg)556 int rdtgroup_mondata_show(struct seq_file *m, void *arg)
557 {
558 struct kernfs_open_file *of = m->private;
559 u32 resid, evtid, domid;
560 struct rdtgroup *rdtgrp;
561 struct rdt_resource *r;
562 union mon_data_bits md;
563 struct rdt_domain *d;
564 struct rmid_read rr;
565 int ret = 0;
566
567 rdtgrp = rdtgroup_kn_lock_live(of->kn);
568 if (!rdtgrp) {
569 ret = -ENOENT;
570 goto out;
571 }
572
573 md.priv = of->kn->priv;
574 resid = md.u.rid;
575 domid = md.u.domid;
576 evtid = md.u.evtid;
577
578 r = &rdt_resources_all[resid].r_resctrl;
579 d = rdt_find_domain(r, domid, NULL);
580 if (IS_ERR_OR_NULL(d)) {
581 ret = -ENOENT;
582 goto out;
583 }
584
585 mon_event_read(&rr, r, d, rdtgrp, evtid, false);
586
587 if (rr.err == -EIO)
588 seq_puts(m, "Error\n");
589 else if (rr.err == -EINVAL)
590 seq_puts(m, "Unavailable\n");
591 else
592 seq_printf(m, "%llu\n", rr.val);
593
594 out:
595 rdtgroup_kn_unlock(of->kn);
596 return ret;
597 }
598