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 struct rdt_parse_data {
27 	struct rdtgroup		*rdtgrp;
28 	char			*buf;
29 };
30 
31 typedef int (ctrlval_parser_t)(struct rdt_parse_data *data,
32 			       struct resctrl_schema *s,
33 			       struct rdt_ctrl_domain *d);
34 
35 /*
36  * Check whether MBA bandwidth percentage value is correct. The value is
37  * checked against the minimum and max bandwidth values specified by the
38  * hardware. The allocated bandwidth percentage is rounded to the next
39  * control step available on the hardware.
40  */
41 static bool bw_validate(char *buf, u32 *data, struct rdt_resource *r)
42 {
43 	int ret;
44 	u32 bw;
45 
46 	/*
47 	 * Only linear delay values is supported for current Intel SKUs.
48 	 */
49 	if (!r->membw.delay_linear && r->membw.arch_needs_linear) {
50 		rdt_last_cmd_puts("No support for non-linear MB domains\n");
51 		return false;
52 	}
53 
54 	ret = kstrtou32(buf, 10, &bw);
55 	if (ret) {
56 		rdt_last_cmd_printf("Invalid MB value %s\n", buf);
57 		return false;
58 	}
59 
60 	/* Nothing else to do if software controller is enabled. */
61 	if (is_mba_sc(r)) {
62 		*data = bw;
63 		return true;
64 	}
65 
66 	if (bw < r->membw.min_bw || bw > r->membw.max_bw) {
67 		rdt_last_cmd_printf("MB value %u out of range [%d,%d]\n",
68 				    bw, r->membw.min_bw, r->membw.max_bw);
69 		return false;
70 	}
71 
72 	*data = roundup(bw, (unsigned long)r->membw.bw_gran);
73 	return true;
74 }
75 
76 static int parse_bw(struct rdt_parse_data *data, struct resctrl_schema *s,
77 		    struct rdt_ctrl_domain *d)
78 {
79 	struct resctrl_staged_config *cfg;
80 	u32 closid = data->rdtgrp->closid;
81 	struct rdt_resource *r = s->res;
82 	u32 bw_val;
83 
84 	cfg = &d->staged_config[s->conf_type];
85 	if (cfg->have_new_ctrl) {
86 		rdt_last_cmd_printf("Duplicate domain %d\n", d->hdr.id);
87 		return -EINVAL;
88 	}
89 
90 	if (!bw_validate(data->buf, &bw_val, r))
91 		return -EINVAL;
92 
93 	if (is_mba_sc(r)) {
94 		d->mbps_val[closid] = bw_val;
95 		return 0;
96 	}
97 
98 	cfg->new_ctrl = bw_val;
99 	cfg->have_new_ctrl = true;
100 
101 	return 0;
102 }
103 
104 /*
105  * Check whether a cache bit mask is valid.
106  * On Intel CPUs, non-contiguous 1s value support is indicated by CPUID:
107  *   - CPUID.0x10.1:ECX[3]: L3 non-contiguous 1s value supported if 1
108  *   - CPUID.0x10.2:ECX[3]: L2 non-contiguous 1s value supported if 1
109  *
110  * Haswell does not support a non-contiguous 1s value and additionally
111  * requires at least two bits set.
112  * AMD allows non-contiguous bitmasks.
113  */
114 static bool cbm_validate(char *buf, u32 *data, struct rdt_resource *r)
115 {
116 	u32 supported_bits = BIT_MASK(r->cache.cbm_len) - 1;
117 	unsigned int cbm_len = r->cache.cbm_len;
118 	unsigned long first_bit, zero_bit, val;
119 	int ret;
120 
121 	ret = kstrtoul(buf, 16, &val);
122 	if (ret) {
123 		rdt_last_cmd_printf("Non-hex character in the mask %s\n", buf);
124 		return false;
125 	}
126 
127 	if ((r->cache.min_cbm_bits > 0 && val == 0) || val > supported_bits) {
128 		rdt_last_cmd_puts("Mask out of range\n");
129 		return false;
130 	}
131 
132 	first_bit = find_first_bit(&val, cbm_len);
133 	zero_bit = find_next_zero_bit(&val, cbm_len, first_bit);
134 
135 	/* Are non-contiguous bitmasks allowed? */
136 	if (!r->cache.arch_has_sparse_bitmasks &&
137 	    (find_next_bit(&val, cbm_len, zero_bit) < cbm_len)) {
138 		rdt_last_cmd_printf("The mask %lx has non-consecutive 1-bits\n", val);
139 		return false;
140 	}
141 
142 	if ((zero_bit - first_bit) < r->cache.min_cbm_bits) {
143 		rdt_last_cmd_printf("Need at least %d bits in the mask\n",
144 				    r->cache.min_cbm_bits);
145 		return false;
146 	}
147 
148 	*data = val;
149 	return true;
150 }
151 
152 /*
153  * Read one cache bit mask (hex). Check that it is valid for the current
154  * resource type.
155  */
156 static int parse_cbm(struct rdt_parse_data *data, struct resctrl_schema *s,
157 		     struct rdt_ctrl_domain *d)
158 {
159 	struct rdtgroup *rdtgrp = data->rdtgrp;
160 	struct resctrl_staged_config *cfg;
161 	struct rdt_resource *r = s->res;
162 	u32 cbm_val;
163 
164 	cfg = &d->staged_config[s->conf_type];
165 	if (cfg->have_new_ctrl) {
166 		rdt_last_cmd_printf("Duplicate domain %d\n", d->hdr.id);
167 		return -EINVAL;
168 	}
169 
170 	/*
171 	 * Cannot set up more than one pseudo-locked region in a cache
172 	 * hierarchy.
173 	 */
174 	if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP &&
175 	    rdtgroup_pseudo_locked_in_hierarchy(d)) {
176 		rdt_last_cmd_puts("Pseudo-locked region in hierarchy\n");
177 		return -EINVAL;
178 	}
179 
180 	if (!cbm_validate(data->buf, &cbm_val, r))
181 		return -EINVAL;
182 
183 	if ((rdtgrp->mode == RDT_MODE_EXCLUSIVE ||
184 	     rdtgrp->mode == RDT_MODE_SHAREABLE) &&
185 	    rdtgroup_cbm_overlaps_pseudo_locked(d, cbm_val)) {
186 		rdt_last_cmd_puts("CBM overlaps with pseudo-locked region\n");
187 		return -EINVAL;
188 	}
189 
190 	/*
191 	 * The CBM may not overlap with the CBM of another closid if
192 	 * either is exclusive.
193 	 */
194 	if (rdtgroup_cbm_overlaps(s, d, cbm_val, rdtgrp->closid, true)) {
195 		rdt_last_cmd_puts("Overlaps with exclusive group\n");
196 		return -EINVAL;
197 	}
198 
199 	if (rdtgroup_cbm_overlaps(s, d, cbm_val, rdtgrp->closid, false)) {
200 		if (rdtgrp->mode == RDT_MODE_EXCLUSIVE ||
201 		    rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
202 			rdt_last_cmd_puts("Overlaps with other group\n");
203 			return -EINVAL;
204 		}
205 	}
206 
207 	cfg->new_ctrl = cbm_val;
208 	cfg->have_new_ctrl = true;
209 
210 	return 0;
211 }
212 
213 /*
214  * For each domain in this resource we expect to find a series of:
215  *	id=mask
216  * separated by ";". The "id" is in decimal, and must match one of
217  * the "id"s for this resource.
218  */
219 static int parse_line(char *line, struct resctrl_schema *s,
220 		      struct rdtgroup *rdtgrp)
221 {
222 	enum resctrl_conf_type t = s->conf_type;
223 	ctrlval_parser_t *parse_ctrlval = NULL;
224 	struct resctrl_staged_config *cfg;
225 	struct rdt_resource *r = s->res;
226 	struct rdt_parse_data data;
227 	struct rdt_ctrl_domain *d;
228 	char *dom = NULL, *id;
229 	unsigned long dom_id;
230 
231 	/* Walking r->domains, ensure it can't race with cpuhp */
232 	lockdep_assert_cpus_held();
233 
234 	switch (r->schema_fmt) {
235 	case RESCTRL_SCHEMA_BITMAP:
236 		parse_ctrlval = &parse_cbm;
237 		break;
238 	case RESCTRL_SCHEMA_RANGE:
239 		parse_ctrlval = &parse_bw;
240 		break;
241 	}
242 
243 	if (WARN_ON_ONCE(!parse_ctrlval))
244 		return -EINVAL;
245 
246 	if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP &&
247 	    (r->rid == RDT_RESOURCE_MBA || r->rid == RDT_RESOURCE_SMBA)) {
248 		rdt_last_cmd_puts("Cannot pseudo-lock MBA resource\n");
249 		return -EINVAL;
250 	}
251 
252 next:
253 	if (!line || line[0] == '\0')
254 		return 0;
255 	dom = strsep(&line, ";");
256 	id = strsep(&dom, "=");
257 	if (!dom || kstrtoul(id, 10, &dom_id)) {
258 		rdt_last_cmd_puts("Missing '=' or non-numeric domain\n");
259 		return -EINVAL;
260 	}
261 	dom = strim(dom);
262 	list_for_each_entry(d, &r->ctrl_domains, hdr.list) {
263 		if (d->hdr.id == dom_id) {
264 			data.buf = dom;
265 			data.rdtgrp = rdtgrp;
266 			if (parse_ctrlval(&data, s, d))
267 				return -EINVAL;
268 			if (rdtgrp->mode ==  RDT_MODE_PSEUDO_LOCKSETUP) {
269 				cfg = &d->staged_config[t];
270 				/*
271 				 * In pseudo-locking setup mode and just
272 				 * parsed a valid CBM that should be
273 				 * pseudo-locked. Only one locked region per
274 				 * resource group and domain so just do
275 				 * the required initialization for single
276 				 * region and return.
277 				 */
278 				rdtgrp->plr->s = s;
279 				rdtgrp->plr->d = d;
280 				rdtgrp->plr->cbm = cfg->new_ctrl;
281 				d->plr = rdtgrp->plr;
282 				return 0;
283 			}
284 			goto next;
285 		}
286 	}
287 	return -EINVAL;
288 }
289 
290 static int rdtgroup_parse_resource(char *resname, char *tok,
291 				   struct rdtgroup *rdtgrp)
292 {
293 	struct resctrl_schema *s;
294 
295 	list_for_each_entry(s, &resctrl_schema_all, list) {
296 		if (!strcmp(resname, s->name) && rdtgrp->closid < s->num_closid)
297 			return parse_line(tok, s, rdtgrp);
298 	}
299 	rdt_last_cmd_printf("Unknown or unsupported resource name '%s'\n", resname);
300 	return -EINVAL;
301 }
302 
303 ssize_t rdtgroup_schemata_write(struct kernfs_open_file *of,
304 				char *buf, size_t nbytes, loff_t off)
305 {
306 	struct resctrl_schema *s;
307 	struct rdtgroup *rdtgrp;
308 	struct rdt_resource *r;
309 	char *tok, *resname;
310 	int ret = 0;
311 
312 	/* Valid input requires a trailing newline */
313 	if (nbytes == 0 || buf[nbytes - 1] != '\n')
314 		return -EINVAL;
315 	buf[nbytes - 1] = '\0';
316 
317 	rdtgrp = rdtgroup_kn_lock_live(of->kn);
318 	if (!rdtgrp) {
319 		rdtgroup_kn_unlock(of->kn);
320 		return -ENOENT;
321 	}
322 	rdt_last_cmd_clear();
323 
324 	/*
325 	 * No changes to pseudo-locked region allowed. It has to be removed
326 	 * and re-created instead.
327 	 */
328 	if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
329 		ret = -EINVAL;
330 		rdt_last_cmd_puts("Resource group is pseudo-locked\n");
331 		goto out;
332 	}
333 
334 	rdt_staged_configs_clear();
335 
336 	while ((tok = strsep(&buf, "\n")) != NULL) {
337 		resname = strim(strsep(&tok, ":"));
338 		if (!tok) {
339 			rdt_last_cmd_puts("Missing ':'\n");
340 			ret = -EINVAL;
341 			goto out;
342 		}
343 		if (tok[0] == '\0') {
344 			rdt_last_cmd_printf("Missing '%s' value\n", resname);
345 			ret = -EINVAL;
346 			goto out;
347 		}
348 		ret = rdtgroup_parse_resource(resname, tok, rdtgrp);
349 		if (ret)
350 			goto out;
351 	}
352 
353 	list_for_each_entry(s, &resctrl_schema_all, list) {
354 		r = s->res;
355 
356 		/*
357 		 * Writes to mba_sc resources update the software controller,
358 		 * not the control MSR.
359 		 */
360 		if (is_mba_sc(r))
361 			continue;
362 
363 		ret = resctrl_arch_update_domains(r, rdtgrp->closid);
364 		if (ret)
365 			goto out;
366 	}
367 
368 	if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
369 		/*
370 		 * If pseudo-locking fails we keep the resource group in
371 		 * mode RDT_MODE_PSEUDO_LOCKSETUP with its class of service
372 		 * active and updated for just the domain the pseudo-locked
373 		 * region was requested for.
374 		 */
375 		ret = rdtgroup_pseudo_lock_create(rdtgrp);
376 	}
377 
378 out:
379 	rdt_staged_configs_clear();
380 	rdtgroup_kn_unlock(of->kn);
381 	return ret ?: nbytes;
382 }
383 
384 static void show_doms(struct seq_file *s, struct resctrl_schema *schema, int closid)
385 {
386 	struct rdt_resource *r = schema->res;
387 	struct rdt_ctrl_domain *dom;
388 	bool sep = false;
389 	u32 ctrl_val;
390 
391 	/* Walking r->domains, ensure it can't race with cpuhp */
392 	lockdep_assert_cpus_held();
393 
394 	seq_printf(s, "%*s:", max_name_width, schema->name);
395 	list_for_each_entry(dom, &r->ctrl_domains, hdr.list) {
396 		if (sep)
397 			seq_puts(s, ";");
398 
399 		if (is_mba_sc(r))
400 			ctrl_val = dom->mbps_val[closid];
401 		else
402 			ctrl_val = resctrl_arch_get_config(r, dom, closid,
403 							   schema->conf_type);
404 
405 		seq_printf(s, schema->fmt_str, dom->hdr.id, ctrl_val);
406 		sep = true;
407 	}
408 	seq_puts(s, "\n");
409 }
410 
411 int rdtgroup_schemata_show(struct kernfs_open_file *of,
412 			   struct seq_file *s, void *v)
413 {
414 	struct resctrl_schema *schema;
415 	struct rdtgroup *rdtgrp;
416 	int ret = 0;
417 	u32 closid;
418 
419 	rdtgrp = rdtgroup_kn_lock_live(of->kn);
420 	if (rdtgrp) {
421 		if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
422 			list_for_each_entry(schema, &resctrl_schema_all, list) {
423 				seq_printf(s, "%s:uninitialized\n", schema->name);
424 			}
425 		} else if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
426 			if (!rdtgrp->plr->d) {
427 				rdt_last_cmd_clear();
428 				rdt_last_cmd_puts("Cache domain offline\n");
429 				ret = -ENODEV;
430 			} else {
431 				seq_printf(s, "%s:%d=%x\n",
432 					   rdtgrp->plr->s->res->name,
433 					   rdtgrp->plr->d->hdr.id,
434 					   rdtgrp->plr->cbm);
435 			}
436 		} else {
437 			closid = rdtgrp->closid;
438 			list_for_each_entry(schema, &resctrl_schema_all, list) {
439 				if (closid < schema->num_closid)
440 					show_doms(s, schema, closid);
441 			}
442 		}
443 	} else {
444 		ret = -ENOENT;
445 	}
446 	rdtgroup_kn_unlock(of->kn);
447 	return ret;
448 }
449 
450 static int smp_mon_event_count(void *arg)
451 {
452 	mon_event_count(arg);
453 
454 	return 0;
455 }
456 
457 ssize_t rdtgroup_mba_mbps_event_write(struct kernfs_open_file *of,
458 				      char *buf, size_t nbytes, loff_t off)
459 {
460 	struct rdtgroup *rdtgrp;
461 	int ret = 0;
462 
463 	/* Valid input requires a trailing newline */
464 	if (nbytes == 0 || buf[nbytes - 1] != '\n')
465 		return -EINVAL;
466 	buf[nbytes - 1] = '\0';
467 
468 	rdtgrp = rdtgroup_kn_lock_live(of->kn);
469 	if (!rdtgrp) {
470 		rdtgroup_kn_unlock(of->kn);
471 		return -ENOENT;
472 	}
473 	rdt_last_cmd_clear();
474 
475 	if (!strcmp(buf, "mbm_local_bytes")) {
476 		if (resctrl_arch_is_mbm_local_enabled())
477 			rdtgrp->mba_mbps_event = QOS_L3_MBM_LOCAL_EVENT_ID;
478 		else
479 			ret = -EINVAL;
480 	} else if (!strcmp(buf, "mbm_total_bytes")) {
481 		if (resctrl_arch_is_mbm_total_enabled())
482 			rdtgrp->mba_mbps_event = QOS_L3_MBM_TOTAL_EVENT_ID;
483 		else
484 			ret = -EINVAL;
485 	} else {
486 		ret = -EINVAL;
487 	}
488 
489 	if (ret)
490 		rdt_last_cmd_printf("Unsupported event id '%s'\n", buf);
491 
492 	rdtgroup_kn_unlock(of->kn);
493 
494 	return ret ?: nbytes;
495 }
496 
497 int rdtgroup_mba_mbps_event_show(struct kernfs_open_file *of,
498 				 struct seq_file *s, void *v)
499 {
500 	struct rdtgroup *rdtgrp;
501 	int ret = 0;
502 
503 	rdtgrp = rdtgroup_kn_lock_live(of->kn);
504 
505 	if (rdtgrp) {
506 		switch (rdtgrp->mba_mbps_event) {
507 		case QOS_L3_MBM_LOCAL_EVENT_ID:
508 			seq_puts(s, "mbm_local_bytes\n");
509 			break;
510 		case QOS_L3_MBM_TOTAL_EVENT_ID:
511 			seq_puts(s, "mbm_total_bytes\n");
512 			break;
513 		default:
514 			pr_warn_once("Bad event %d\n", rdtgrp->mba_mbps_event);
515 			ret = -EINVAL;
516 			break;
517 		}
518 	} else {
519 		ret = -ENOENT;
520 	}
521 
522 	rdtgroup_kn_unlock(of->kn);
523 
524 	return ret;
525 }
526 
527 struct rdt_domain_hdr *resctrl_find_domain(struct list_head *h, int id,
528 					   struct list_head **pos)
529 {
530 	struct rdt_domain_hdr *d;
531 	struct list_head *l;
532 
533 	list_for_each(l, h) {
534 		d = list_entry(l, struct rdt_domain_hdr, list);
535 		/* When id is found, return its domain. */
536 		if (id == d->id)
537 			return d;
538 		/* Stop searching when finding id's position in sorted list. */
539 		if (id < d->id)
540 			break;
541 	}
542 
543 	if (pos)
544 		*pos = l;
545 
546 	return NULL;
547 }
548 
549 void mon_event_read(struct rmid_read *rr, struct rdt_resource *r,
550 		    struct rdt_mon_domain *d, struct rdtgroup *rdtgrp,
551 		    cpumask_t *cpumask, int evtid, int first)
552 {
553 	int cpu;
554 
555 	/* When picking a CPU from cpu_mask, ensure it can't race with cpuhp */
556 	lockdep_assert_cpus_held();
557 
558 	/*
559 	 * Setup the parameters to pass to mon_event_count() to read the data.
560 	 */
561 	rr->rgrp = rdtgrp;
562 	rr->evtid = evtid;
563 	rr->r = r;
564 	rr->d = d;
565 	rr->first = first;
566 	rr->arch_mon_ctx = resctrl_arch_mon_ctx_alloc(r, evtid);
567 	if (IS_ERR(rr->arch_mon_ctx)) {
568 		rr->err = -EINVAL;
569 		return;
570 	}
571 
572 	cpu = cpumask_any_housekeeping(cpumask, RESCTRL_PICK_ANY_CPU);
573 
574 	/*
575 	 * cpumask_any_housekeeping() prefers housekeeping CPUs, but
576 	 * are all the CPUs nohz_full? If yes, pick a CPU to IPI.
577 	 * MPAM's resctrl_arch_rmid_read() is unable to read the
578 	 * counters on some platforms if its called in IRQ context.
579 	 */
580 	if (tick_nohz_full_cpu(cpu))
581 		smp_call_function_any(cpumask, mon_event_count, rr, 1);
582 	else
583 		smp_call_on_cpu(cpu, smp_mon_event_count, rr, false);
584 
585 	resctrl_arch_mon_ctx_free(r, evtid, rr->arch_mon_ctx);
586 }
587 
588 int rdtgroup_mondata_show(struct seq_file *m, void *arg)
589 {
590 	struct kernfs_open_file *of = m->private;
591 	enum resctrl_res_level resid;
592 	enum resctrl_event_id evtid;
593 	struct rdt_domain_hdr *hdr;
594 	struct rmid_read rr = {0};
595 	struct rdt_mon_domain *d;
596 	struct rdtgroup *rdtgrp;
597 	struct rdt_resource *r;
598 	struct mon_data *md;
599 	int domid, ret = 0;
600 
601 	rdtgrp = rdtgroup_kn_lock_live(of->kn);
602 	if (!rdtgrp) {
603 		ret = -ENOENT;
604 		goto out;
605 	}
606 
607 	md = of->kn->priv;
608 	if (WARN_ON_ONCE(!md)) {
609 		ret = -EIO;
610 		goto out;
611 	}
612 
613 	resid = md->rid;
614 	domid = md->domid;
615 	evtid = md->evtid;
616 	r = resctrl_arch_get_resource(resid);
617 
618 	if (md->sum) {
619 		/*
620 		 * This file requires summing across all domains that share
621 		 * the L3 cache id that was provided in the "domid" field of the
622 		 * struct mon_data. Search all domains in the resource for
623 		 * one that matches this cache id.
624 		 */
625 		list_for_each_entry(d, &r->mon_domains, hdr.list) {
626 			if (d->ci->id == domid) {
627 				rr.ci = d->ci;
628 				mon_event_read(&rr, r, NULL, rdtgrp,
629 					       &d->ci->shared_cpu_map, evtid, false);
630 				goto checkresult;
631 			}
632 		}
633 		ret = -ENOENT;
634 		goto out;
635 	} else {
636 		/*
637 		 * This file provides data from a single domain. Search
638 		 * the resource to find the domain with "domid".
639 		 */
640 		hdr = resctrl_find_domain(&r->mon_domains, domid, NULL);
641 		if (!hdr || WARN_ON_ONCE(hdr->type != RESCTRL_MON_DOMAIN)) {
642 			ret = -ENOENT;
643 			goto out;
644 		}
645 		d = container_of(hdr, struct rdt_mon_domain, hdr);
646 		mon_event_read(&rr, r, d, rdtgrp, &d->hdr.cpu_mask, evtid, false);
647 	}
648 
649 checkresult:
650 
651 	if (rr.err == -EIO)
652 		seq_puts(m, "Error\n");
653 	else if (rr.err == -EINVAL)
654 		seq_puts(m, "Unavailable\n");
655 	else
656 		seq_printf(m, "%llu\n", rr.val);
657 
658 out:
659 	rdtgroup_kn_unlock(of->kn);
660 	return ret;
661 }
662