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