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 "internal.h"
23
24 /*
25 * Check whether MBA bandwidth percentage value is correct. The value is
26 * checked against the minimum and max bandwidth values specified by the
27 * hardware. The allocated bandwidth percentage is rounded to the next
28 * control step available on the hardware.
29 */
bw_validate(char * buf,unsigned long * data,struct rdt_resource * r)30 static bool bw_validate(char *buf, unsigned long *data, struct rdt_resource *r)
31 {
32 unsigned long bw;
33 int ret;
34
35 /*
36 * Only linear delay values is supported for current Intel SKUs.
37 */
38 if (!r->membw.delay_linear && r->membw.arch_needs_linear) {
39 rdt_last_cmd_puts("No support for non-linear MB domains\n");
40 return false;
41 }
42
43 ret = kstrtoul(buf, 10, &bw);
44 if (ret) {
45 rdt_last_cmd_printf("Non-decimal digit in MB value %s\n", buf);
46 return false;
47 }
48
49 if ((bw < r->membw.min_bw || bw > r->default_ctrl) &&
50 !is_mba_sc(r)) {
51 rdt_last_cmd_printf("MB value %ld out of range [%d,%d]\n", bw,
52 r->membw.min_bw, r->default_ctrl);
53 return false;
54 }
55
56 *data = roundup(bw, (unsigned long)r->membw.bw_gran);
57 return true;
58 }
59
parse_bw(struct rdt_parse_data * data,struct resctrl_schema * s,struct rdt_domain * d)60 int parse_bw(struct rdt_parse_data *data, struct resctrl_schema *s,
61 struct rdt_domain *d)
62 {
63 struct resctrl_staged_config *cfg;
64 u32 closid = data->rdtgrp->closid;
65 struct rdt_resource *r = s->res;
66 unsigned long bw_val;
67
68 cfg = &d->staged_config[s->conf_type];
69 if (cfg->have_new_ctrl) {
70 rdt_last_cmd_printf("Duplicate domain %d\n", d->id);
71 return -EINVAL;
72 }
73
74 if (!bw_validate(data->buf, &bw_val, r))
75 return -EINVAL;
76
77 if (is_mba_sc(r)) {
78 d->mbps_val[closid] = bw_val;
79 return 0;
80 }
81
82 cfg->new_ctrl = bw_val;
83 cfg->have_new_ctrl = true;
84
85 return 0;
86 }
87
88 /*
89 * Check whether a cache bit mask is valid.
90 * For Intel the SDM says:
91 * Please note that all (and only) contiguous '1' combinations
92 * are allowed (e.g. FFFFH, 0FF0H, 003CH, etc.).
93 * Additionally Haswell requires at least two bits set.
94 * AMD allows non-contiguous bitmasks.
95 */
cbm_validate(char * buf,u32 * data,struct rdt_resource * r)96 static bool cbm_validate(char *buf, u32 *data, struct rdt_resource *r)
97 {
98 unsigned long first_bit, zero_bit, val;
99 unsigned int cbm_len = r->cache.cbm_len;
100 int ret;
101
102 ret = kstrtoul(buf, 16, &val);
103 if (ret) {
104 rdt_last_cmd_printf("Non-hex character in the mask %s\n", buf);
105 return false;
106 }
107
108 if ((r->cache.min_cbm_bits > 0 && val == 0) || val > r->default_ctrl) {
109 rdt_last_cmd_puts("Mask out of range\n");
110 return false;
111 }
112
113 first_bit = find_first_bit(&val, cbm_len);
114 zero_bit = find_next_zero_bit(&val, cbm_len, first_bit);
115
116 /* Are non-contiguous bitmaps allowed? */
117 if (!r->cache.arch_has_sparse_bitmaps &&
118 (find_next_bit(&val, cbm_len, zero_bit) < cbm_len)) {
119 rdt_last_cmd_printf("The mask %lx has non-consecutive 1-bits\n", val);
120 return false;
121 }
122
123 if ((zero_bit - first_bit) < r->cache.min_cbm_bits) {
124 rdt_last_cmd_printf("Need at least %d bits in the mask\n",
125 r->cache.min_cbm_bits);
126 return false;
127 }
128
129 *data = val;
130 return true;
131 }
132
133 /*
134 * Read one cache bit mask (hex). Check that it is valid for the current
135 * resource type.
136 */
parse_cbm(struct rdt_parse_data * data,struct resctrl_schema * s,struct rdt_domain * d)137 int parse_cbm(struct rdt_parse_data *data, struct resctrl_schema *s,
138 struct rdt_domain *d)
139 {
140 struct rdtgroup *rdtgrp = data->rdtgrp;
141 struct resctrl_staged_config *cfg;
142 struct rdt_resource *r = s->res;
143 u32 cbm_val;
144
145 cfg = &d->staged_config[s->conf_type];
146 if (cfg->have_new_ctrl) {
147 rdt_last_cmd_printf("Duplicate domain %d\n", d->id);
148 return -EINVAL;
149 }
150
151 /*
152 * Cannot set up more than one pseudo-locked region in a cache
153 * hierarchy.
154 */
155 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP &&
156 rdtgroup_pseudo_locked_in_hierarchy(d)) {
157 rdt_last_cmd_puts("Pseudo-locked region in hierarchy\n");
158 return -EINVAL;
159 }
160
161 if (!cbm_validate(data->buf, &cbm_val, r))
162 return -EINVAL;
163
164 if ((rdtgrp->mode == RDT_MODE_EXCLUSIVE ||
165 rdtgrp->mode == RDT_MODE_SHAREABLE) &&
166 rdtgroup_cbm_overlaps_pseudo_locked(d, cbm_val)) {
167 rdt_last_cmd_puts("CBM overlaps with pseudo-locked region\n");
168 return -EINVAL;
169 }
170
171 /*
172 * The CBM may not overlap with the CBM of another closid if
173 * either is exclusive.
174 */
175 if (rdtgroup_cbm_overlaps(s, d, cbm_val, rdtgrp->closid, true)) {
176 rdt_last_cmd_puts("Overlaps with exclusive group\n");
177 return -EINVAL;
178 }
179
180 if (rdtgroup_cbm_overlaps(s, d, cbm_val, rdtgrp->closid, false)) {
181 if (rdtgrp->mode == RDT_MODE_EXCLUSIVE ||
182 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
183 rdt_last_cmd_puts("Overlaps with other group\n");
184 return -EINVAL;
185 }
186 }
187
188 cfg->new_ctrl = cbm_val;
189 cfg->have_new_ctrl = true;
190
191 return 0;
192 }
193
194 /*
195 * For each domain in this resource we expect to find a series of:
196 * id=mask
197 * separated by ";". The "id" is in decimal, and must match one of
198 * the "id"s for this resource.
199 */
parse_line(char * line,struct resctrl_schema * s,struct rdtgroup * rdtgrp)200 static int parse_line(char *line, struct resctrl_schema *s,
201 struct rdtgroup *rdtgrp)
202 {
203 enum resctrl_conf_type t = s->conf_type;
204 struct resctrl_staged_config *cfg;
205 struct rdt_resource *r = s->res;
206 struct rdt_parse_data data;
207 char *dom = NULL, *id;
208 struct rdt_domain *d;
209 unsigned long dom_id;
210
211 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP &&
212 (r->rid == RDT_RESOURCE_MBA || r->rid == RDT_RESOURCE_SMBA)) {
213 rdt_last_cmd_puts("Cannot pseudo-lock MBA resource\n");
214 return -EINVAL;
215 }
216
217 next:
218 if (!line || line[0] == '\0')
219 return 0;
220 dom = strsep(&line, ";");
221 id = strsep(&dom, "=");
222 if (!dom || kstrtoul(id, 10, &dom_id)) {
223 rdt_last_cmd_puts("Missing '=' or non-numeric domain\n");
224 return -EINVAL;
225 }
226 dom = strim(dom);
227 list_for_each_entry(d, &r->domains, list) {
228 if (d->id == dom_id) {
229 data.buf = dom;
230 data.rdtgrp = rdtgrp;
231 if (r->parse_ctrlval(&data, s, d))
232 return -EINVAL;
233 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
234 cfg = &d->staged_config[t];
235 /*
236 * In pseudo-locking setup mode and just
237 * parsed a valid CBM that should be
238 * pseudo-locked. Only one locked region per
239 * resource group and domain so just do
240 * the required initialization for single
241 * region and return.
242 */
243 rdtgrp->plr->s = s;
244 rdtgrp->plr->d = d;
245 rdtgrp->plr->cbm = cfg->new_ctrl;
246 d->plr = rdtgrp->plr;
247 return 0;
248 }
249 goto next;
250 }
251 }
252 return -EINVAL;
253 }
254
get_config_index(u32 closid,enum resctrl_conf_type type)255 static u32 get_config_index(u32 closid, enum resctrl_conf_type type)
256 {
257 switch (type) {
258 default:
259 case CDP_NONE:
260 return closid;
261 case CDP_CODE:
262 return closid * 2 + 1;
263 case CDP_DATA:
264 return closid * 2;
265 }
266 }
267
apply_config(struct rdt_hw_domain * hw_dom,struct resctrl_staged_config * cfg,u32 idx,cpumask_var_t cpu_mask)268 static bool apply_config(struct rdt_hw_domain *hw_dom,
269 struct resctrl_staged_config *cfg, u32 idx,
270 cpumask_var_t cpu_mask)
271 {
272 struct rdt_domain *dom = &hw_dom->d_resctrl;
273
274 if (cfg->new_ctrl != hw_dom->ctrl_val[idx]) {
275 cpumask_set_cpu(cpumask_any(&dom->cpu_mask), cpu_mask);
276 hw_dom->ctrl_val[idx] = cfg->new_ctrl;
277
278 return true;
279 }
280
281 return false;
282 }
283
resctrl_arch_update_one(struct rdt_resource * r,struct rdt_domain * d,u32 closid,enum resctrl_conf_type t,u32 cfg_val)284 int resctrl_arch_update_one(struct rdt_resource *r, struct rdt_domain *d,
285 u32 closid, enum resctrl_conf_type t, u32 cfg_val)
286 {
287 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
288 struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
289 u32 idx = get_config_index(closid, t);
290 struct msr_param msr_param;
291
292 if (!cpumask_test_cpu(smp_processor_id(), &d->cpu_mask))
293 return -EINVAL;
294
295 hw_dom->ctrl_val[idx] = cfg_val;
296
297 msr_param.res = r;
298 msr_param.low = idx;
299 msr_param.high = idx + 1;
300 hw_res->msr_update(d, &msr_param, r);
301
302 return 0;
303 }
304
resctrl_arch_update_domains(struct rdt_resource * r,u32 closid)305 int resctrl_arch_update_domains(struct rdt_resource *r, u32 closid)
306 {
307 struct resctrl_staged_config *cfg;
308 struct rdt_hw_domain *hw_dom;
309 struct msr_param msr_param;
310 enum resctrl_conf_type t;
311 cpumask_var_t cpu_mask;
312 struct rdt_domain *d;
313 u32 idx;
314
315 if (!zalloc_cpumask_var(&cpu_mask, GFP_KERNEL))
316 return -ENOMEM;
317
318 msr_param.res = NULL;
319 list_for_each_entry(d, &r->domains, list) {
320 hw_dom = resctrl_to_arch_dom(d);
321 for (t = 0; t < CDP_NUM_TYPES; t++) {
322 cfg = &hw_dom->d_resctrl.staged_config[t];
323 if (!cfg->have_new_ctrl)
324 continue;
325
326 idx = get_config_index(closid, t);
327 if (!apply_config(hw_dom, cfg, idx, cpu_mask))
328 continue;
329
330 if (!msr_param.res) {
331 msr_param.low = idx;
332 msr_param.high = msr_param.low + 1;
333 msr_param.res = r;
334 } else {
335 msr_param.low = min(msr_param.low, idx);
336 msr_param.high = max(msr_param.high, idx + 1);
337 }
338 }
339 }
340
341 if (cpumask_empty(cpu_mask))
342 goto done;
343
344 /* Update resource control msr on all the CPUs. */
345 on_each_cpu_mask(cpu_mask, rdt_ctrl_update, &msr_param, 1);
346
347 done:
348 free_cpumask_var(cpu_mask);
349
350 return 0;
351 }
352
rdtgroup_parse_resource(char * resname,char * tok,struct rdtgroup * rdtgrp)353 static int rdtgroup_parse_resource(char *resname, char *tok,
354 struct rdtgroup *rdtgrp)
355 {
356 struct resctrl_schema *s;
357
358 list_for_each_entry(s, &resctrl_schema_all, list) {
359 if (!strcmp(resname, s->name) && rdtgrp->closid < s->num_closid)
360 return parse_line(tok, s, rdtgrp);
361 }
362 rdt_last_cmd_printf("Unknown or unsupported resource name '%s'\n", resname);
363 return -EINVAL;
364 }
365
rdtgroup_schemata_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)366 ssize_t rdtgroup_schemata_write(struct kernfs_open_file *of,
367 char *buf, size_t nbytes, loff_t off)
368 {
369 struct resctrl_schema *s;
370 struct rdtgroup *rdtgrp;
371 struct rdt_resource *r;
372 char *tok, *resname;
373 int ret = 0;
374
375 /* Valid input requires a trailing newline */
376 if (nbytes == 0 || buf[nbytes - 1] != '\n')
377 return -EINVAL;
378 buf[nbytes - 1] = '\0';
379
380 cpus_read_lock();
381 rdtgrp = rdtgroup_kn_lock_live(of->kn);
382 if (!rdtgrp) {
383 rdtgroup_kn_unlock(of->kn);
384 cpus_read_unlock();
385 return -ENOENT;
386 }
387 rdt_last_cmd_clear();
388
389 /*
390 * No changes to pseudo-locked region allowed. It has to be removed
391 * and re-created instead.
392 */
393 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
394 ret = -EINVAL;
395 rdt_last_cmd_puts("Resource group is pseudo-locked\n");
396 goto out;
397 }
398
399 rdt_staged_configs_clear();
400
401 while ((tok = strsep(&buf, "\n")) != NULL) {
402 resname = strim(strsep(&tok, ":"));
403 if (!tok) {
404 rdt_last_cmd_puts("Missing ':'\n");
405 ret = -EINVAL;
406 goto out;
407 }
408 if (tok[0] == '\0') {
409 rdt_last_cmd_printf("Missing '%s' value\n", resname);
410 ret = -EINVAL;
411 goto out;
412 }
413 ret = rdtgroup_parse_resource(resname, tok, rdtgrp);
414 if (ret)
415 goto out;
416 }
417
418 list_for_each_entry(s, &resctrl_schema_all, list) {
419 r = s->res;
420
421 /*
422 * Writes to mba_sc resources update the software controller,
423 * not the control MSR.
424 */
425 if (is_mba_sc(r))
426 continue;
427
428 ret = resctrl_arch_update_domains(r, rdtgrp->closid);
429 if (ret)
430 goto out;
431 }
432
433 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
434 /*
435 * If pseudo-locking fails we keep the resource group in
436 * mode RDT_MODE_PSEUDO_LOCKSETUP with its class of service
437 * active and updated for just the domain the pseudo-locked
438 * region was requested for.
439 */
440 ret = rdtgroup_pseudo_lock_create(rdtgrp);
441 }
442
443 out:
444 rdt_staged_configs_clear();
445 rdtgroup_kn_unlock(of->kn);
446 cpus_read_unlock();
447 return ret ?: nbytes;
448 }
449
resctrl_arch_get_config(struct rdt_resource * r,struct rdt_domain * d,u32 closid,enum resctrl_conf_type type)450 u32 resctrl_arch_get_config(struct rdt_resource *r, struct rdt_domain *d,
451 u32 closid, enum resctrl_conf_type type)
452 {
453 struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
454 u32 idx = get_config_index(closid, type);
455
456 return hw_dom->ctrl_val[idx];
457 }
458
show_doms(struct seq_file * s,struct resctrl_schema * schema,int closid)459 static void show_doms(struct seq_file *s, struct resctrl_schema *schema, int closid)
460 {
461 struct rdt_resource *r = schema->res;
462 struct rdt_domain *dom;
463 bool sep = false;
464 u32 ctrl_val;
465
466 seq_printf(s, "%*s:", max_name_width, schema->name);
467 list_for_each_entry(dom, &r->domains, list) {
468 if (sep)
469 seq_puts(s, ";");
470
471 if (is_mba_sc(r))
472 ctrl_val = dom->mbps_val[closid];
473 else
474 ctrl_val = resctrl_arch_get_config(r, dom, closid,
475 schema->conf_type);
476
477 seq_printf(s, r->format_str, dom->id, max_data_width,
478 ctrl_val);
479 sep = true;
480 }
481 seq_puts(s, "\n");
482 }
483
rdtgroup_schemata_show(struct kernfs_open_file * of,struct seq_file * s,void * v)484 int rdtgroup_schemata_show(struct kernfs_open_file *of,
485 struct seq_file *s, void *v)
486 {
487 struct resctrl_schema *schema;
488 struct rdtgroup *rdtgrp;
489 int ret = 0;
490 u32 closid;
491
492 rdtgrp = rdtgroup_kn_lock_live(of->kn);
493 if (rdtgrp) {
494 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
495 list_for_each_entry(schema, &resctrl_schema_all, list) {
496 seq_printf(s, "%s:uninitialized\n", schema->name);
497 }
498 } else if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
499 if (!rdtgrp->plr->d) {
500 rdt_last_cmd_clear();
501 rdt_last_cmd_puts("Cache domain offline\n");
502 ret = -ENODEV;
503 } else {
504 seq_printf(s, "%s:%d=%x\n",
505 rdtgrp->plr->s->res->name,
506 rdtgrp->plr->d->id,
507 rdtgrp->plr->cbm);
508 }
509 } else {
510 closid = rdtgrp->closid;
511 list_for_each_entry(schema, &resctrl_schema_all, list) {
512 if (closid < schema->num_closid)
513 show_doms(s, schema, closid);
514 }
515 }
516 } else {
517 ret = -ENOENT;
518 }
519 rdtgroup_kn_unlock(of->kn);
520 return ret;
521 }
522
mon_event_read(struct rmid_read * rr,struct rdt_resource * r,struct rdt_domain * d,struct rdtgroup * rdtgrp,int evtid,int first)523 void mon_event_read(struct rmid_read *rr, struct rdt_resource *r,
524 struct rdt_domain *d, struct rdtgroup *rdtgrp,
525 int evtid, int first)
526 {
527 /*
528 * setup the parameters to send to the IPI to read the data.
529 */
530 rr->rgrp = rdtgrp;
531 rr->evtid = evtid;
532 rr->r = r;
533 rr->d = d;
534 rr->val = 0;
535 rr->first = first;
536
537 smp_call_function_any(&d->cpu_mask, mon_event_count, rr, 1);
538 }
539
rdtgroup_mondata_show(struct seq_file * m,void * arg)540 int rdtgroup_mondata_show(struct seq_file *m, void *arg)
541 {
542 struct kernfs_open_file *of = m->private;
543 u32 resid, evtid, domid;
544 struct rdtgroup *rdtgrp;
545 struct rdt_resource *r;
546 union mon_data_bits md;
547 struct rdt_domain *d;
548 struct rmid_read rr;
549 int ret = 0;
550
551 rdtgrp = rdtgroup_kn_lock_live(of->kn);
552 if (!rdtgrp) {
553 ret = -ENOENT;
554 goto out;
555 }
556
557 md.priv = of->kn->priv;
558 resid = md.u.rid;
559 domid = md.u.domid;
560 evtid = md.u.evtid;
561
562 r = &rdt_resources_all[resid].r_resctrl;
563 d = rdt_find_domain(r, domid, NULL);
564 if (IS_ERR_OR_NULL(d)) {
565 ret = -ENOENT;
566 goto out;
567 }
568
569 mon_event_read(&rr, r, d, rdtgrp, evtid, false);
570
571 if (rr.err == -EIO)
572 seq_puts(m, "Error\n");
573 else if (rr.err == -EINVAL)
574 seq_puts(m, "Unavailable\n");
575 else
576 seq_printf(m, "%llu\n", rr.val);
577
578 out:
579 rdtgroup_kn_unlock(of->kn);
580 return ret;
581 }
582