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  */
30 static bool bw_validate(char *buf, u32 *data, struct rdt_resource *r)
31 {
32 	int ret;
33 	u32 bw;
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 = kstrtou32(buf, 10, &bw);
44 	if (ret) {
45 		rdt_last_cmd_printf("Invalid MB value %s\n", buf);
46 		return false;
47 	}
48 
49 	/* Nothing else to do if software controller is enabled. */
50 	if (is_mba_sc(r)) {
51 		*data = bw;
52 		return true;
53 	}
54 
55 	if (bw < r->membw.min_bw || bw > r->default_ctrl) {
56 		rdt_last_cmd_printf("MB value %u out of range [%d,%d]\n",
57 				    bw, r->membw.min_bw, r->default_ctrl);
58 		return false;
59 	}
60 
61 	*data = roundup(bw, (unsigned long)r->membw.bw_gran);
62 	return true;
63 }
64 
65 int parse_bw(struct rdt_parse_data *data, struct resctrl_schema *s,
66 	     struct rdt_domain *d)
67 {
68 	struct resctrl_staged_config *cfg;
69 	u32 closid = data->rdtgrp->closid;
70 	struct rdt_resource *r = s->res;
71 	u32 bw_val;
72 
73 	cfg = &d->staged_config[s->conf_type];
74 	if (cfg->have_new_ctrl) {
75 		rdt_last_cmd_printf("Duplicate domain %d\n", d->id);
76 		return -EINVAL;
77 	}
78 
79 	if (!bw_validate(data->buf, &bw_val, r))
80 		return -EINVAL;
81 
82 	if (is_mba_sc(r)) {
83 		d->mbps_val[closid] = bw_val;
84 		return 0;
85 	}
86 
87 	cfg->new_ctrl = bw_val;
88 	cfg->have_new_ctrl = true;
89 
90 	return 0;
91 }
92 
93 /*
94  * Check whether a cache bit mask is valid.
95  * For Intel the SDM says:
96  *	Please note that all (and only) contiguous '1' combinations
97  *	are allowed (e.g. FFFFH, 0FF0H, 003CH, etc.).
98  * Additionally Haswell requires at least two bits set.
99  * AMD allows non-contiguous bitmasks.
100  */
101 static bool cbm_validate(char *buf, u32 *data, struct rdt_resource *r)
102 {
103 	unsigned long first_bit, zero_bit, val;
104 	unsigned int cbm_len = r->cache.cbm_len;
105 	int ret;
106 
107 	ret = kstrtoul(buf, 16, &val);
108 	if (ret) {
109 		rdt_last_cmd_printf("Non-hex character in the mask %s\n", buf);
110 		return false;
111 	}
112 
113 	if ((r->cache.min_cbm_bits > 0 && val == 0) || val > r->default_ctrl) {
114 		rdt_last_cmd_puts("Mask out of range\n");
115 		return false;
116 	}
117 
118 	first_bit = find_first_bit(&val, cbm_len);
119 	zero_bit = find_next_zero_bit(&val, cbm_len, first_bit);
120 
121 	/* Are non-contiguous bitmaps allowed? */
122 	if (!r->cache.arch_has_sparse_bitmaps &&
123 	    (find_next_bit(&val, cbm_len, zero_bit) < cbm_len)) {
124 		rdt_last_cmd_printf("The mask %lx has non-consecutive 1-bits\n", val);
125 		return false;
126 	}
127 
128 	if ((zero_bit - first_bit) < r->cache.min_cbm_bits) {
129 		rdt_last_cmd_printf("Need at least %d bits in the mask\n",
130 				    r->cache.min_cbm_bits);
131 		return false;
132 	}
133 
134 	*data = val;
135 	return true;
136 }
137 
138 /*
139  * Read one cache bit mask (hex). Check that it is valid for the current
140  * resource type.
141  */
142 int parse_cbm(struct rdt_parse_data *data, struct resctrl_schema *s,
143 	      struct rdt_domain *d)
144 {
145 	struct rdtgroup *rdtgrp = data->rdtgrp;
146 	struct resctrl_staged_config *cfg;
147 	struct rdt_resource *r = s->res;
148 	u32 cbm_val;
149 
150 	cfg = &d->staged_config[s->conf_type];
151 	if (cfg->have_new_ctrl) {
152 		rdt_last_cmd_printf("Duplicate domain %d\n", d->id);
153 		return -EINVAL;
154 	}
155 
156 	/*
157 	 * Cannot set up more than one pseudo-locked region in a cache
158 	 * hierarchy.
159 	 */
160 	if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP &&
161 	    rdtgroup_pseudo_locked_in_hierarchy(d)) {
162 		rdt_last_cmd_puts("Pseudo-locked region in hierarchy\n");
163 		return -EINVAL;
164 	}
165 
166 	if (!cbm_validate(data->buf, &cbm_val, r))
167 		return -EINVAL;
168 
169 	if ((rdtgrp->mode == RDT_MODE_EXCLUSIVE ||
170 	     rdtgrp->mode == RDT_MODE_SHAREABLE) &&
171 	    rdtgroup_cbm_overlaps_pseudo_locked(d, cbm_val)) {
172 		rdt_last_cmd_puts("CBM overlaps with pseudo-locked region\n");
173 		return -EINVAL;
174 	}
175 
176 	/*
177 	 * The CBM may not overlap with the CBM of another closid if
178 	 * either is exclusive.
179 	 */
180 	if (rdtgroup_cbm_overlaps(s, d, cbm_val, rdtgrp->closid, true)) {
181 		rdt_last_cmd_puts("Overlaps with exclusive group\n");
182 		return -EINVAL;
183 	}
184 
185 	if (rdtgroup_cbm_overlaps(s, d, cbm_val, rdtgrp->closid, false)) {
186 		if (rdtgrp->mode == RDT_MODE_EXCLUSIVE ||
187 		    rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
188 			rdt_last_cmd_puts("Overlaps with other group\n");
189 			return -EINVAL;
190 		}
191 	}
192 
193 	cfg->new_ctrl = cbm_val;
194 	cfg->have_new_ctrl = true;
195 
196 	return 0;
197 }
198 
199 /*
200  * For each domain in this resource we expect to find a series of:
201  *	id=mask
202  * separated by ";". The "id" is in decimal, and must match one of
203  * the "id"s for this resource.
204  */
205 static int parse_line(char *line, struct resctrl_schema *s,
206 		      struct rdtgroup *rdtgrp)
207 {
208 	enum resctrl_conf_type t = s->conf_type;
209 	struct resctrl_staged_config *cfg;
210 	struct rdt_resource *r = s->res;
211 	struct rdt_parse_data data;
212 	char *dom = NULL, *id;
213 	struct rdt_domain *d;
214 	unsigned long dom_id;
215 
216 	if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP &&
217 	    (r->rid == RDT_RESOURCE_MBA || r->rid == RDT_RESOURCE_SMBA)) {
218 		rdt_last_cmd_puts("Cannot pseudo-lock MBA resource\n");
219 		return -EINVAL;
220 	}
221 
222 next:
223 	if (!line || line[0] == '\0')
224 		return 0;
225 	dom = strsep(&line, ";");
226 	id = strsep(&dom, "=");
227 	if (!dom || kstrtoul(id, 10, &dom_id)) {
228 		rdt_last_cmd_puts("Missing '=' or non-numeric domain\n");
229 		return -EINVAL;
230 	}
231 	dom = strim(dom);
232 	list_for_each_entry(d, &r->domains, list) {
233 		if (d->id == dom_id) {
234 			data.buf = dom;
235 			data.rdtgrp = rdtgrp;
236 			if (r->parse_ctrlval(&data, s, d))
237 				return -EINVAL;
238 			if (rdtgrp->mode ==  RDT_MODE_PSEUDO_LOCKSETUP) {
239 				cfg = &d->staged_config[t];
240 				/*
241 				 * In pseudo-locking setup mode and just
242 				 * parsed a valid CBM that should be
243 				 * pseudo-locked. Only one locked region per
244 				 * resource group and domain so just do
245 				 * the required initialization for single
246 				 * region and return.
247 				 */
248 				rdtgrp->plr->s = s;
249 				rdtgrp->plr->d = d;
250 				rdtgrp->plr->cbm = cfg->new_ctrl;
251 				d->plr = rdtgrp->plr;
252 				return 0;
253 			}
254 			goto next;
255 		}
256 	}
257 	return -EINVAL;
258 }
259 
260 static u32 get_config_index(u32 closid, enum resctrl_conf_type type)
261 {
262 	switch (type) {
263 	default:
264 	case CDP_NONE:
265 		return closid;
266 	case CDP_CODE:
267 		return closid * 2 + 1;
268 	case CDP_DATA:
269 		return closid * 2;
270 	}
271 }
272 
273 static bool apply_config(struct rdt_hw_domain *hw_dom,
274 			 struct resctrl_staged_config *cfg, u32 idx,
275 			 cpumask_var_t cpu_mask)
276 {
277 	struct rdt_domain *dom = &hw_dom->d_resctrl;
278 
279 	if (cfg->new_ctrl != hw_dom->ctrl_val[idx]) {
280 		cpumask_set_cpu(cpumask_any(&dom->cpu_mask), cpu_mask);
281 		hw_dom->ctrl_val[idx] = cfg->new_ctrl;
282 
283 		return true;
284 	}
285 
286 	return false;
287 }
288 
289 int resctrl_arch_update_one(struct rdt_resource *r, struct rdt_domain *d,
290 			    u32 closid, enum resctrl_conf_type t, u32 cfg_val)
291 {
292 	struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
293 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
294 	u32 idx = get_config_index(closid, t);
295 	struct msr_param msr_param;
296 
297 	if (!cpumask_test_cpu(smp_processor_id(), &d->cpu_mask))
298 		return -EINVAL;
299 
300 	hw_dom->ctrl_val[idx] = cfg_val;
301 
302 	msr_param.res = r;
303 	msr_param.low = idx;
304 	msr_param.high = idx + 1;
305 	hw_res->msr_update(d, &msr_param, r);
306 
307 	return 0;
308 }
309 
310 int resctrl_arch_update_domains(struct rdt_resource *r, u32 closid)
311 {
312 	struct resctrl_staged_config *cfg;
313 	struct rdt_hw_domain *hw_dom;
314 	struct msr_param msr_param;
315 	enum resctrl_conf_type t;
316 	cpumask_var_t cpu_mask;
317 	struct rdt_domain *d;
318 	u32 idx;
319 
320 	if (!zalloc_cpumask_var(&cpu_mask, GFP_KERNEL))
321 		return -ENOMEM;
322 
323 	msr_param.res = NULL;
324 	list_for_each_entry(d, &r->domains, list) {
325 		hw_dom = resctrl_to_arch_dom(d);
326 		for (t = 0; t < CDP_NUM_TYPES; t++) {
327 			cfg = &hw_dom->d_resctrl.staged_config[t];
328 			if (!cfg->have_new_ctrl)
329 				continue;
330 
331 			idx = get_config_index(closid, t);
332 			if (!apply_config(hw_dom, cfg, idx, cpu_mask))
333 				continue;
334 
335 			if (!msr_param.res) {
336 				msr_param.low = idx;
337 				msr_param.high = msr_param.low + 1;
338 				msr_param.res = r;
339 			} else {
340 				msr_param.low = min(msr_param.low, idx);
341 				msr_param.high = max(msr_param.high, idx + 1);
342 			}
343 		}
344 	}
345 
346 	if (cpumask_empty(cpu_mask))
347 		goto done;
348 
349 	/* Update resource control msr on all the CPUs. */
350 	on_each_cpu_mask(cpu_mask, rdt_ctrl_update, &msr_param, 1);
351 
352 done:
353 	free_cpumask_var(cpu_mask);
354 
355 	return 0;
356 }
357 
358 static int rdtgroup_parse_resource(char *resname, char *tok,
359 				   struct rdtgroup *rdtgrp)
360 {
361 	struct resctrl_schema *s;
362 
363 	list_for_each_entry(s, &resctrl_schema_all, list) {
364 		if (!strcmp(resname, s->name) && rdtgrp->closid < s->num_closid)
365 			return parse_line(tok, s, rdtgrp);
366 	}
367 	rdt_last_cmd_printf("Unknown or unsupported resource name '%s'\n", resname);
368 	return -EINVAL;
369 }
370 
371 ssize_t rdtgroup_schemata_write(struct kernfs_open_file *of,
372 				char *buf, size_t nbytes, loff_t off)
373 {
374 	struct resctrl_schema *s;
375 	struct rdtgroup *rdtgrp;
376 	struct rdt_resource *r;
377 	char *tok, *resname;
378 	int ret = 0;
379 
380 	/* Valid input requires a trailing newline */
381 	if (nbytes == 0 || buf[nbytes - 1] != '\n')
382 		return -EINVAL;
383 	buf[nbytes - 1] = '\0';
384 
385 	cpus_read_lock();
386 	rdtgrp = rdtgroup_kn_lock_live(of->kn);
387 	if (!rdtgrp) {
388 		rdtgroup_kn_unlock(of->kn);
389 		cpus_read_unlock();
390 		return -ENOENT;
391 	}
392 	rdt_last_cmd_clear();
393 
394 	/*
395 	 * No changes to pseudo-locked region allowed. It has to be removed
396 	 * and re-created instead.
397 	 */
398 	if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
399 		ret = -EINVAL;
400 		rdt_last_cmd_puts("Resource group is pseudo-locked\n");
401 		goto out;
402 	}
403 
404 	rdt_staged_configs_clear();
405 
406 	while ((tok = strsep(&buf, "\n")) != NULL) {
407 		resname = strim(strsep(&tok, ":"));
408 		if (!tok) {
409 			rdt_last_cmd_puts("Missing ':'\n");
410 			ret = -EINVAL;
411 			goto out;
412 		}
413 		if (tok[0] == '\0') {
414 			rdt_last_cmd_printf("Missing '%s' value\n", resname);
415 			ret = -EINVAL;
416 			goto out;
417 		}
418 		ret = rdtgroup_parse_resource(resname, tok, rdtgrp);
419 		if (ret)
420 			goto out;
421 	}
422 
423 	list_for_each_entry(s, &resctrl_schema_all, list) {
424 		r = s->res;
425 
426 		/*
427 		 * Writes to mba_sc resources update the software controller,
428 		 * not the control MSR.
429 		 */
430 		if (is_mba_sc(r))
431 			continue;
432 
433 		ret = resctrl_arch_update_domains(r, rdtgrp->closid);
434 		if (ret)
435 			goto out;
436 	}
437 
438 	if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
439 		/*
440 		 * If pseudo-locking fails we keep the resource group in
441 		 * mode RDT_MODE_PSEUDO_LOCKSETUP with its class of service
442 		 * active and updated for just the domain the pseudo-locked
443 		 * region was requested for.
444 		 */
445 		ret = rdtgroup_pseudo_lock_create(rdtgrp);
446 	}
447 
448 out:
449 	rdt_staged_configs_clear();
450 	rdtgroup_kn_unlock(of->kn);
451 	cpus_read_unlock();
452 	return ret ?: nbytes;
453 }
454 
455 u32 resctrl_arch_get_config(struct rdt_resource *r, struct rdt_domain *d,
456 			    u32 closid, enum resctrl_conf_type type)
457 {
458 	struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
459 	u32 idx = get_config_index(closid, type);
460 
461 	return hw_dom->ctrl_val[idx];
462 }
463 
464 static void show_doms(struct seq_file *s, struct resctrl_schema *schema, int closid)
465 {
466 	struct rdt_resource *r = schema->res;
467 	struct rdt_domain *dom;
468 	bool sep = false;
469 	u32 ctrl_val;
470 
471 	seq_printf(s, "%*s:", max_name_width, schema->name);
472 	list_for_each_entry(dom, &r->domains, list) {
473 		if (sep)
474 			seq_puts(s, ";");
475 
476 		if (is_mba_sc(r))
477 			ctrl_val = dom->mbps_val[closid];
478 		else
479 			ctrl_val = resctrl_arch_get_config(r, dom, closid,
480 							   schema->conf_type);
481 
482 		seq_printf(s, r->format_str, dom->id, max_data_width,
483 			   ctrl_val);
484 		sep = true;
485 	}
486 	seq_puts(s, "\n");
487 }
488 
489 int rdtgroup_schemata_show(struct kernfs_open_file *of,
490 			   struct seq_file *s, void *v)
491 {
492 	struct resctrl_schema *schema;
493 	struct rdtgroup *rdtgrp;
494 	int ret = 0;
495 	u32 closid;
496 
497 	rdtgrp = rdtgroup_kn_lock_live(of->kn);
498 	if (rdtgrp) {
499 		if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
500 			list_for_each_entry(schema, &resctrl_schema_all, list) {
501 				seq_printf(s, "%s:uninitialized\n", schema->name);
502 			}
503 		} else if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
504 			if (!rdtgrp->plr->d) {
505 				rdt_last_cmd_clear();
506 				rdt_last_cmd_puts("Cache domain offline\n");
507 				ret = -ENODEV;
508 			} else {
509 				seq_printf(s, "%s:%d=%x\n",
510 					   rdtgrp->plr->s->res->name,
511 					   rdtgrp->plr->d->id,
512 					   rdtgrp->plr->cbm);
513 			}
514 		} else {
515 			closid = rdtgrp->closid;
516 			list_for_each_entry(schema, &resctrl_schema_all, list) {
517 				if (closid < schema->num_closid)
518 					show_doms(s, schema, closid);
519 			}
520 		}
521 	} else {
522 		ret = -ENOENT;
523 	}
524 	rdtgroup_kn_unlock(of->kn);
525 	return ret;
526 }
527 
528 void mon_event_read(struct rmid_read *rr, struct rdt_resource *r,
529 		    struct rdt_domain *d, struct rdtgroup *rdtgrp,
530 		    int evtid, int first)
531 {
532 	/*
533 	 * setup the parameters to send to the IPI to read the data.
534 	 */
535 	rr->rgrp = rdtgrp;
536 	rr->evtid = evtid;
537 	rr->r = r;
538 	rr->d = d;
539 	rr->val = 0;
540 	rr->first = first;
541 
542 	smp_call_function_any(&d->cpu_mask, mon_event_count, rr, 1);
543 }
544 
545 int rdtgroup_mondata_show(struct seq_file *m, void *arg)
546 {
547 	struct kernfs_open_file *of = m->private;
548 	u32 resid, evtid, domid;
549 	struct rdtgroup *rdtgrp;
550 	struct rdt_resource *r;
551 	union mon_data_bits md;
552 	struct rdt_domain *d;
553 	struct rmid_read rr;
554 	int ret = 0;
555 
556 	rdtgrp = rdtgroup_kn_lock_live(of->kn);
557 	if (!rdtgrp) {
558 		ret = -ENOENT;
559 		goto out;
560 	}
561 
562 	md.priv = of->kn->priv;
563 	resid = md.u.rid;
564 	domid = md.u.domid;
565 	evtid = md.u.evtid;
566 
567 	r = &rdt_resources_all[resid].r_resctrl;
568 	d = rdt_find_domain(r, domid, NULL);
569 	if (IS_ERR_OR_NULL(d)) {
570 		ret = -ENOENT;
571 		goto out;
572 	}
573 
574 	mon_event_read(&rr, r, d, rdtgrp, evtid, false);
575 
576 	if (rr.err == -EIO)
577 		seq_puts(m, "Error\n");
578 	else if (rr.err == -EINVAL)
579 		seq_puts(m, "Unavailable\n");
580 	else
581 		seq_printf(m, "%llu\n", rr.val);
582 
583 out:
584 	rdtgroup_kn_unlock(of->kn);
585 	return ret;
586 }
587