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