arch/x86/kernel/cpu/resctrl/ctrlmondata.c

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