cpu/resctrl/ctrlmondata.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * Resource Director Technology(RDT)
0004  * - Cache Allocation code.
0005  *
0006  * Copyright (C) 2016 Intel Corporation
0007  *
0008  * Authors:
0009  *    Fenghua Yu <fenghua.yu@intel.com>
0010  *    Tony Luck <tony.luck@intel.com>
0011  *
0012  * More information about RDT be found in the Intel (R) x86 Architecture
0013  * Software Developer Manual June 2016, volume 3, section 17.17.
0014  */
0015
0016 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
0017
0018 #include <linux/cpu.h>
0019 #include <linux/kernfs.h>
0020 #include <linux/seq_file.h>
0021 #include <linux/slab.h>
0022 #include "internal.h"
0023
0024 /*
0025  * Check whether MBA bandwidth percentage value is correct. The value is
0026  * checked against the minimum and max bandwidth values specified by the
0027  * hardware. The allocated bandwidth percentage is rounded to the next
0028  * control step available on the hardware.
0029  */
0030 static bool bw_validate(char *buf, unsigned long *data, struct rdt_resource *r)
0031 {
0032     unsigned long bw;
0033     int ret;
0034
0035     /*
0036      * Only linear delay values is supported for current Intel SKUs.
0037      */
0038     if (!r->membw.delay_linear && r->membw.arch_needs_linear) {
0039         rdt_last_cmd_puts("No support for non-linear MB domains\n");
0040         return false;
0041     }
0042
0043     ret = kstrtoul(buf, 10, &bw);
0044     if (ret) {
0045         rdt_last_cmd_printf("Non-decimal digit in MB value %s\n", buf);
0046         return false;
0047     }
0048
0049     if ((bw < r->membw.min_bw || bw > r->default_ctrl) &&
0050         !is_mba_sc(r)) {
0051         rdt_last_cmd_printf("MB value %ld out of range [%d,%d]\n", bw,
0052                     r->membw.min_bw, r->default_ctrl);
0053         return false;
0054     }
0055
0056     *data = roundup(bw, (unsigned long)r->membw.bw_gran);
0057     return true;
0058 }
0059
0060 int parse_bw(struct rdt_parse_data *data, struct resctrl_schema *s,
0061          struct rdt_domain *d)
0062 {
0063     struct resctrl_staged_config *cfg;
0064     struct rdt_resource *r = s->res;
0065     unsigned long bw_val;
0066
0067     cfg = &d->staged_config[s->conf_type];
0068     if (cfg->have_new_ctrl) {
0069         rdt_last_cmd_printf("Duplicate domain %d\n", d->id);
0070         return -EINVAL;
0071     }
0072
0073     if (!bw_validate(data->buf, &bw_val, r))
0074         return -EINVAL;
0075     cfg->new_ctrl = bw_val;
0076     cfg->have_new_ctrl = true;
0077
0078     return 0;
0079 }
0080
0081 /*
0082  * Check whether a cache bit mask is valid.
0083  * For Intel the SDM says:
0084  *  Please note that all (and only) contiguous '1' combinations
0085  *  are allowed (e.g. FFFFH, 0FF0H, 003CH, etc.).
0086  * Additionally Haswell requires at least two bits set.
0087  * AMD allows non-contiguous bitmasks.
0088  */
0089 static bool cbm_validate(char *buf, u32 *data, struct rdt_resource *r)
0090 {
0091     unsigned long first_bit, zero_bit, val;
0092     unsigned int cbm_len = r->cache.cbm_len;
0093     int ret;
0094
0095     ret = kstrtoul(buf, 16, &val);
0096     if (ret) {
0097         rdt_last_cmd_printf("Non-hex character in the mask %s\n", buf);
0098         return false;
0099     }
0100
0101     if ((!r->cache.arch_has_empty_bitmaps && val == 0) ||
0102         val > r->default_ctrl) {
0103         rdt_last_cmd_puts("Mask out of range\n");
0104         return false;
0105     }
0106
0107     first_bit = find_first_bit(&val, cbm_len);
0108     zero_bit = find_next_zero_bit(&val, cbm_len, first_bit);
0109
0110     /* Are non-contiguous bitmaps allowed? */
0111     if (!r->cache.arch_has_sparse_bitmaps &&
0112         (find_next_bit(&val, cbm_len, zero_bit) < cbm_len)) {
0113         rdt_last_cmd_printf("The mask %lx has non-consecutive 1-bits\n", val);
0114         return false;
0115     }
0116
0117     if ((zero_bit - first_bit) < r->cache.min_cbm_bits) {
0118         rdt_last_cmd_printf("Need at least %d bits in the mask\n",
0119                     r->cache.min_cbm_bits);
0120         return false;
0121     }
0122
0123     *data = val;
0124     return true;
0125 }
0126
0127 /*
0128  * Read one cache bit mask (hex). Check that it is valid for the current
0129  * resource type.
0130  */
0131 int parse_cbm(struct rdt_parse_data *data, struct resctrl_schema *s,
0132           struct rdt_domain *d)
0133 {
0134     struct rdtgroup *rdtgrp = data->rdtgrp;
0135     struct resctrl_staged_config *cfg;
0136     struct rdt_resource *r = s->res;
0137     u32 cbm_val;
0138
0139     cfg = &d->staged_config[s->conf_type];
0140     if (cfg->have_new_ctrl) {
0141         rdt_last_cmd_printf("Duplicate domain %d\n", d->id);
0142         return -EINVAL;
0143     }
0144
0145     /*
0146      * Cannot set up more than one pseudo-locked region in a cache
0147      * hierarchy.
0148      */
0149     if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP &&
0150         rdtgroup_pseudo_locked_in_hierarchy(d)) {
0151         rdt_last_cmd_puts("Pseudo-locked region in hierarchy\n");
0152         return -EINVAL;
0153     }
0154
0155     if (!cbm_validate(data->buf, &cbm_val, r))
0156         return -EINVAL;
0157
0158     if ((rdtgrp->mode == RDT_MODE_EXCLUSIVE ||
0159          rdtgrp->mode == RDT_MODE_SHAREABLE) &&
0160         rdtgroup_cbm_overlaps_pseudo_locked(d, cbm_val)) {
0161         rdt_last_cmd_puts("CBM overlaps with pseudo-locked region\n");
0162         return -EINVAL;
0163     }
0164
0165     /*
0166      * The CBM may not overlap with the CBM of another closid if
0167      * either is exclusive.
0168      */
0169     if (rdtgroup_cbm_overlaps(s, d, cbm_val, rdtgrp->closid, true)) {
0170         rdt_last_cmd_puts("Overlaps with exclusive group\n");
0171         return -EINVAL;
0172     }
0173
0174     if (rdtgroup_cbm_overlaps(s, d, cbm_val, rdtgrp->closid, false)) {
0175         if (rdtgrp->mode == RDT_MODE_EXCLUSIVE ||
0176             rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
0177             rdt_last_cmd_puts("Overlaps with other group\n");
0178             return -EINVAL;
0179         }
0180     }
0181
0182     cfg->new_ctrl = cbm_val;
0183     cfg->have_new_ctrl = true;
0184
0185     return 0;
0186 }
0187
0188 /*
0189  * For each domain in this resource we expect to find a series of:
0190  *  id=mask
0191  * separated by ";". The "id" is in decimal, and must match one of
0192  * the "id"s for this resource.
0193  */
0194 static int parse_line(char *line, struct resctrl_schema *s,
0195               struct rdtgroup *rdtgrp)
0196 {
0197     enum resctrl_conf_type t = s->conf_type;
0198     struct resctrl_staged_config *cfg;
0199     struct rdt_resource *r = s->res;
0200     struct rdt_parse_data data;
0201     char *dom = NULL, *id;
0202     struct rdt_domain *d;
0203     unsigned long dom_id;
0204
0205     if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP &&
0206         r->rid == RDT_RESOURCE_MBA) {
0207         rdt_last_cmd_puts("Cannot pseudo-lock MBA resource\n");
0208         return -EINVAL;
0209     }
0210
0211 next:
0212     if (!line || line[0] == '\0')
0213         return 0;
0214     dom = strsep(&line, ";");
0215     id = strsep(&dom, "=");
0216     if (!dom || kstrtoul(id, 10, &dom_id)) {
0217         rdt_last_cmd_puts("Missing '=' or non-numeric domain\n");
0218         return -EINVAL;
0219     }
0220     dom = strim(dom);
0221     list_for_each_entry(d, &r->domains, list) {
0222         if (d->id == dom_id) {
0223             data.buf = dom;
0224             data.rdtgrp = rdtgrp;
0225             if (r->parse_ctrlval(&data, s, d))
0226                 return -EINVAL;
0227             if (rdtgrp->mode ==  RDT_MODE_PSEUDO_LOCKSETUP) {
0228                 cfg = &d->staged_config[t];
0229                 /*
0230                  * In pseudo-locking setup mode and just
0231                  * parsed a valid CBM that should be
0232                  * pseudo-locked. Only one locked region per
0233                  * resource group and domain so just do
0234                  * the required initialization for single
0235                  * region and return.
0236                  */
0237                 rdtgrp->plr->s = s;
0238                 rdtgrp->plr->d = d;
0239                 rdtgrp->plr->cbm = cfg->new_ctrl;
0240                 d->plr = rdtgrp->plr;
0241                 return 0;
0242             }
0243             goto next;
0244         }
0245     }
0246     return -EINVAL;
0247 }
0248
0249 static u32 get_config_index(u32 closid, enum resctrl_conf_type type)
0250 {
0251     switch (type) {
0252     default:
0253     case CDP_NONE:
0254         return closid;
0255     case CDP_CODE:
0256         return closid * 2 + 1;
0257     case CDP_DATA:
0258         return closid * 2;
0259     }
0260 }
0261
0262 static bool apply_config(struct rdt_hw_domain *hw_dom,
0263              struct resctrl_staged_config *cfg, u32 idx,
0264              cpumask_var_t cpu_mask, bool mba_sc)
0265 {
0266     struct rdt_domain *dom = &hw_dom->d_resctrl;
0267     u32 *dc = !mba_sc ? hw_dom->ctrl_val : hw_dom->mbps_val;
0268
0269     if (cfg->new_ctrl != dc[idx]) {
0270         cpumask_set_cpu(cpumask_any(&dom->cpu_mask), cpu_mask);
0271         dc[idx] = cfg->new_ctrl;
0272
0273         return true;
0274     }
0275
0276     return false;
0277 }
0278
0279 int resctrl_arch_update_domains(struct rdt_resource *r, u32 closid)
0280 {
0281     struct resctrl_staged_config *cfg;
0282     struct rdt_hw_domain *hw_dom;
0283     struct msr_param msr_param;
0284     enum resctrl_conf_type t;
0285     cpumask_var_t cpu_mask;
0286     struct rdt_domain *d;
0287     bool mba_sc;
0288     int cpu;
0289     u32 idx;
0290
0291     if (!zalloc_cpumask_var(&cpu_mask, GFP_KERNEL))
0292         return -ENOMEM;
0293
0294     mba_sc = is_mba_sc(r);
0295     msr_param.res = NULL;
0296     list_for_each_entry(d, &r->domains, list) {
0297         hw_dom = resctrl_to_arch_dom(d);
0298         for (t = 0; t < CDP_NUM_TYPES; t++) {
0299             cfg = &hw_dom->d_resctrl.staged_config[t];
0300             if (!cfg->have_new_ctrl)
0301                 continue;
0302
0303             idx = get_config_index(closid, t);
0304             if (!apply_config(hw_dom, cfg, idx, cpu_mask, mba_sc))
0305                 continue;
0306
0307             if (!msr_param.res) {
0308                 msr_param.low = idx;
0309                 msr_param.high = msr_param.low + 1;
0310                 msr_param.res = r;
0311             } else {
0312                 msr_param.low = min(msr_param.low, idx);
0313                 msr_param.high = max(msr_param.high, idx + 1);
0314             }
0315         }
0316     }
0317
0318     /*
0319      * Avoid writing the control msr with control values when
0320      * MBA software controller is enabled
0321      */
0322     if (cpumask_empty(cpu_mask) || mba_sc)
0323         goto done;
0324     cpu = get_cpu();
0325     /* Update resource control msr on this CPU if it's in cpu_mask. */
0326     if (cpumask_test_cpu(cpu, cpu_mask))
0327         rdt_ctrl_update(&msr_param);
0328     /* Update resource control msr on other CPUs. */
0329     smp_call_function_many(cpu_mask, rdt_ctrl_update, &msr_param, 1);
0330     put_cpu();
0331
0332 done:
0333     free_cpumask_var(cpu_mask);
0334
0335     return 0;
0336 }
0337
0338 static int rdtgroup_parse_resource(char *resname, char *tok,
0339                    struct rdtgroup *rdtgrp)
0340 {
0341     struct resctrl_schema *s;
0342
0343     list_for_each_entry(s, &resctrl_schema_all, list) {
0344         if (!strcmp(resname, s->name) && rdtgrp->closid < s->num_closid)
0345             return parse_line(tok, s, rdtgrp);
0346     }
0347     rdt_last_cmd_printf("Unknown or unsupported resource name '%s'\n", resname);
0348     return -EINVAL;
0349 }
0350
0351 ssize_t rdtgroup_schemata_write(struct kernfs_open_file *of,
0352                 char *buf, size_t nbytes, loff_t off)
0353 {
0354     struct resctrl_schema *s;
0355     struct rdtgroup *rdtgrp;
0356     struct rdt_domain *dom;
0357     struct rdt_resource *r;
0358     char *tok, *resname;
0359     int ret = 0;
0360
0361     /* Valid input requires a trailing newline */
0362     if (nbytes == 0 || buf[nbytes - 1] != '\n')
0363         return -EINVAL;
0364     buf[nbytes - 1] = '\0';
0365
0366     cpus_read_lock();
0367     rdtgrp = rdtgroup_kn_lock_live(of->kn);
0368     if (!rdtgrp) {
0369         rdtgroup_kn_unlock(of->kn);
0370         cpus_read_unlock();
0371         return -ENOENT;
0372     }
0373     rdt_last_cmd_clear();
0374
0375     /*
0376      * No changes to pseudo-locked region allowed. It has to be removed
0377      * and re-created instead.
0378      */
0379     if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
0380         ret = -EINVAL;
0381         rdt_last_cmd_puts("Resource group is pseudo-locked\n");
0382         goto out;
0383     }
0384
0385     list_for_each_entry(s, &resctrl_schema_all, list) {
0386         list_for_each_entry(dom, &s->res->domains, list)
0387             memset(dom->staged_config, 0, sizeof(dom->staged_config));
0388     }
0389
0390     while ((tok = strsep(&buf, "\n")) != NULL) {
0391         resname = strim(strsep(&tok, ":"));
0392         if (!tok) {
0393             rdt_last_cmd_puts("Missing ':'\n");
0394             ret = -EINVAL;
0395             goto out;
0396         }
0397         if (tok[0] == '\0') {
0398             rdt_last_cmd_printf("Missing '%s' value\n", resname);
0399             ret = -EINVAL;
0400             goto out;
0401         }
0402         ret = rdtgroup_parse_resource(resname, tok, rdtgrp);
0403         if (ret)
0404             goto out;
0405     }
0406
0407     list_for_each_entry(s, &resctrl_schema_all, list) {
0408         r = s->res;
0409         ret = resctrl_arch_update_domains(r, rdtgrp->closid);
0410         if (ret)
0411             goto out;
0412     }
0413
0414     if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
0415         /*
0416          * If pseudo-locking fails we keep the resource group in
0417          * mode RDT_MODE_PSEUDO_LOCKSETUP with its class of service
0418          * active and updated for just the domain the pseudo-locked
0419          * region was requested for.
0420          */
0421         ret = rdtgroup_pseudo_lock_create(rdtgrp);
0422     }
0423
0424 out:
0425     rdtgroup_kn_unlock(of->kn);
0426     cpus_read_unlock();
0427     return ret ?: nbytes;
0428 }
0429
0430 u32 resctrl_arch_get_config(struct rdt_resource *r, struct rdt_domain *d,
0431                 u32 closid, enum resctrl_conf_type type)
0432 {
0433     struct rdt_hw_domain *hw_dom = resctrl_to_arch_dom(d);
0434     u32 idx = get_config_index(closid, type);
0435
0436     if (!is_mba_sc(r))
0437         return hw_dom->ctrl_val[idx];
0438     return hw_dom->mbps_val[idx];
0439 }
0440
0441 static void show_doms(struct seq_file *s, struct resctrl_schema *schema, int closid)
0442 {
0443     struct rdt_resource *r = schema->res;
0444     struct rdt_domain *dom;
0445     bool sep = false;
0446     u32 ctrl_val;
0447
0448     seq_printf(s, "%*s:", max_name_width, schema->name);
0449     list_for_each_entry(dom, &r->domains, list) {
0450         if (sep)
0451             seq_puts(s, ";");
0452
0453         ctrl_val = resctrl_arch_get_config(r, dom, closid,
0454                            schema->conf_type);
0455         seq_printf(s, r->format_str, dom->id, max_data_width,
0456                ctrl_val);
0457         sep = true;
0458     }
0459     seq_puts(s, "\n");
0460 }
0461
0462 int rdtgroup_schemata_show(struct kernfs_open_file *of,
0463                struct seq_file *s, void *v)
0464 {
0465     struct resctrl_schema *schema;
0466     struct rdtgroup *rdtgrp;
0467     int ret = 0;
0468     u32 closid;
0469
0470     rdtgrp = rdtgroup_kn_lock_live(of->kn);
0471     if (rdtgrp) {
0472         if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
0473             list_for_each_entry(schema, &resctrl_schema_all, list) {
0474                 seq_printf(s, "%s:uninitialized\n", schema->name);
0475             }
0476         } else if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
0477             if (!rdtgrp->plr->d) {
0478                 rdt_last_cmd_clear();
0479                 rdt_last_cmd_puts("Cache domain offline\n");
0480                 ret = -ENODEV;
0481             } else {
0482                 seq_printf(s, "%s:%d=%x\n",
0483                        rdtgrp->plr->s->res->name,
0484                        rdtgrp->plr->d->id,
0485                        rdtgrp->plr->cbm);
0486             }
0487         } else {
0488             closid = rdtgrp->closid;
0489             list_for_each_entry(schema, &resctrl_schema_all, list) {
0490                 if (closid < schema->num_closid)
0491                     show_doms(s, schema, closid);
0492             }
0493         }
0494     } else {
0495         ret = -ENOENT;
0496     }
0497     rdtgroup_kn_unlock(of->kn);
0498     return ret;
0499 }
0500
0501 void mon_event_read(struct rmid_read *rr, struct rdt_resource *r,
0502             struct rdt_domain *d, struct rdtgroup *rdtgrp,
0503             int evtid, int first)
0504 {
0505     /*
0506      * setup the parameters to send to the IPI to read the data.
0507      */
0508     rr->rgrp = rdtgrp;
0509     rr->evtid = evtid;
0510     rr->r = r;
0511     rr->d = d;
0512     rr->val = 0;
0513     rr->first = first;
0514
0515     smp_call_function_any(&d->cpu_mask, mon_event_count, rr, 1);
0516 }
0517
0518 int rdtgroup_mondata_show(struct seq_file *m, void *arg)
0519 {
0520     struct kernfs_open_file *of = m->private;
0521     struct rdt_hw_resource *hw_res;
0522     u32 resid, evtid, domid;
0523     struct rdtgroup *rdtgrp;
0524     struct rdt_resource *r;
0525     union mon_data_bits md;
0526     struct rdt_domain *d;
0527     struct rmid_read rr;
0528     int ret = 0;
0529
0530     rdtgrp = rdtgroup_kn_lock_live(of->kn);
0531     if (!rdtgrp) {
0532         ret = -ENOENT;
0533         goto out;
0534     }
0535
0536     md.priv = of->kn->priv;
0537     resid = md.u.rid;
0538     domid = md.u.domid;
0539     evtid = md.u.evtid;
0540
0541     hw_res = &rdt_resources_all[resid];
0542     r = &hw_res->r_resctrl;
0543     d = rdt_find_domain(r, domid, NULL);
0544     if (IS_ERR_OR_NULL(d)) {
0545         ret = -ENOENT;
0546         goto out;
0547     }
0548
0549     mon_event_read(&rr, r, d, rdtgrp, evtid, false);
0550
0551     if (rr.val & RMID_VAL_ERROR)
0552         seq_puts(m, "Error\n");
0553     else if (rr.val & RMID_VAL_UNAVAIL)
0554         seq_puts(m, "Unavailable\n");
0555     else
0556         seq_printf(m, "%llu\n", rr.val * hw_res->mon_scale);
0557
0558 out:
0559     rdtgroup_kn_unlock(of->kn);
0560     return ret;
0561 }