1 // SPDX-License-Identifier: GPL-2.0-only
3 * User interface for Resource Allocation in Resource Director Technology(RDT)
5 * Copyright (C) 2016 Intel Corporation
9 * More information about RDT be found in the Intel (R) x86 Architecture
10 * Software Developer Manual.
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 #include <linux/cpu.h>
16 #include <linux/debugfs.h>
18 #include <linux/fs_parser.h>
19 #include <linux/sysfs.h>
20 #include <linux/kernfs.h>
21 #include <linux/seq_buf.h>
22 #include <linux/seq_file.h>
23 #include <linux/sched/signal.h>
24 #include <linux/sched/task.h>
25 #include <linux/slab.h>
26 #include <linux/task_work.h>
27 #include <linux/user_namespace.h>
29 #include <uapi/linux/magic.h>
31 #include <asm/resctrl.h>
34 DEFINE_STATIC_KEY_FALSE(rdt_enable_key);
35 DEFINE_STATIC_KEY_FALSE(rdt_mon_enable_key);
36 DEFINE_STATIC_KEY_FALSE(rdt_alloc_enable_key);
38 /* Mutex to protect rdtgroup access. */
39 DEFINE_MUTEX(rdtgroup_mutex);
41 static struct kernfs_root *rdt_root;
42 struct rdtgroup rdtgroup_default;
43 LIST_HEAD(rdt_all_groups);
45 /* list of entries for the schemata file */
46 LIST_HEAD(resctrl_schema_all);
48 /* The filesystem can only be mounted once. */
51 /* Kernel fs node for "info" directory under root */
52 static struct kernfs_node *kn_info;
54 /* Kernel fs node for "mon_groups" directory under root */
55 static struct kernfs_node *kn_mongrp;
57 /* Kernel fs node for "mon_data" directory under root */
58 static struct kernfs_node *kn_mondata;
60 static struct seq_buf last_cmd_status;
61 static char last_cmd_status_buf[512];
63 static int rdtgroup_setup_root(struct rdt_fs_context *ctx);
64 static void rdtgroup_destroy_root(void);
66 struct dentry *debugfs_resctrl;
68 static bool resctrl_debug;
70 void rdt_last_cmd_clear(void)
72 lockdep_assert_held(&rdtgroup_mutex);
73 seq_buf_clear(&last_cmd_status);
76 void rdt_last_cmd_puts(const char *s)
78 lockdep_assert_held(&rdtgroup_mutex);
79 seq_buf_puts(&last_cmd_status, s);
82 void rdt_last_cmd_printf(const char *fmt, ...)
87 lockdep_assert_held(&rdtgroup_mutex);
88 seq_buf_vprintf(&last_cmd_status, fmt, ap);
92 void rdt_staged_configs_clear(void)
94 struct rdt_ctrl_domain *dom;
95 struct rdt_resource *r;
97 lockdep_assert_held(&rdtgroup_mutex);
99 for_each_alloc_capable_rdt_resource(r) {
100 list_for_each_entry(dom, &r->ctrl_domains, hdr.list)
101 memset(dom->staged_config, 0, sizeof(dom->staged_config));
106 * Trivial allocator for CLOSIDs. Since h/w only supports a small number,
107 * we can keep a bitmap of free CLOSIDs in a single integer.
109 * Using a global CLOSID across all resources has some advantages and
111 * + We can simply set current's closid to assign a task to a resource
113 * + Context switch code can avoid extra memory references deciding which
114 * CLOSID to load into the PQR_ASSOC MSR
115 * - We give up some options in configuring resource groups across multi-socket
117 * - Our choices on how to configure each resource become progressively more
118 * limited as the number of resources grows.
120 static unsigned long closid_free_map;
121 static int closid_free_map_len;
123 int closids_supported(void)
125 return closid_free_map_len;
128 static void closid_init(void)
130 struct resctrl_schema *s;
131 u32 rdt_min_closid = 32;
133 /* Compute rdt_min_closid across all resources */
134 list_for_each_entry(s, &resctrl_schema_all, list)
135 rdt_min_closid = min(rdt_min_closid, s->num_closid);
137 closid_free_map = BIT_MASK(rdt_min_closid) - 1;
139 /* RESCTRL_RESERVED_CLOSID is always reserved for the default group */
140 __clear_bit(RESCTRL_RESERVED_CLOSID, &closid_free_map);
141 closid_free_map_len = rdt_min_closid;
144 static int closid_alloc(void)
149 lockdep_assert_held(&rdtgroup_mutex);
151 if (IS_ENABLED(CONFIG_RESCTRL_RMID_DEPENDS_ON_CLOSID)) {
152 cleanest_closid = resctrl_find_cleanest_closid();
153 if (cleanest_closid < 0)
154 return cleanest_closid;
155 closid = cleanest_closid;
157 closid = ffs(closid_free_map);
162 __clear_bit(closid, &closid_free_map);
167 void closid_free(int closid)
169 lockdep_assert_held(&rdtgroup_mutex);
171 __set_bit(closid, &closid_free_map);
175 * closid_allocated - test if provided closid is in use
176 * @closid: closid to be tested
178 * Return: true if @closid is currently associated with a resource group,
179 * false if @closid is free
181 bool closid_allocated(unsigned int closid)
183 lockdep_assert_held(&rdtgroup_mutex);
185 return !test_bit(closid, &closid_free_map);
189 * rdtgroup_mode_by_closid - Return mode of resource group with closid
190 * @closid: closid if the resource group
192 * Each resource group is associated with a @closid. Here the mode
193 * of a resource group can be queried by searching for it using its closid.
195 * Return: mode as &enum rdtgrp_mode of resource group with closid @closid
197 enum rdtgrp_mode rdtgroup_mode_by_closid(int closid)
199 struct rdtgroup *rdtgrp;
201 list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
202 if (rdtgrp->closid == closid)
206 return RDT_NUM_MODES;
209 static const char * const rdt_mode_str[] = {
210 [RDT_MODE_SHAREABLE] = "shareable",
211 [RDT_MODE_EXCLUSIVE] = "exclusive",
212 [RDT_MODE_PSEUDO_LOCKSETUP] = "pseudo-locksetup",
213 [RDT_MODE_PSEUDO_LOCKED] = "pseudo-locked",
217 * rdtgroup_mode_str - Return the string representation of mode
218 * @mode: the resource group mode as &enum rdtgroup_mode
220 * Return: string representation of valid mode, "unknown" otherwise
222 static const char *rdtgroup_mode_str(enum rdtgrp_mode mode)
224 if (mode < RDT_MODE_SHAREABLE || mode >= RDT_NUM_MODES)
227 return rdt_mode_str[mode];
230 /* set uid and gid of rdtgroup dirs and files to that of the creator */
231 static int rdtgroup_kn_set_ugid(struct kernfs_node *kn)
233 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
234 .ia_uid = current_fsuid(),
235 .ia_gid = current_fsgid(), };
237 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
238 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
241 return kernfs_setattr(kn, &iattr);
244 static int rdtgroup_add_file(struct kernfs_node *parent_kn, struct rftype *rft)
246 struct kernfs_node *kn;
249 kn = __kernfs_create_file(parent_kn, rft->name, rft->mode,
250 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
251 0, rft->kf_ops, rft, NULL, NULL);
255 ret = rdtgroup_kn_set_ugid(kn);
264 static int rdtgroup_seqfile_show(struct seq_file *m, void *arg)
266 struct kernfs_open_file *of = m->private;
267 struct rftype *rft = of->kn->priv;
270 return rft->seq_show(of, m, arg);
274 static ssize_t rdtgroup_file_write(struct kernfs_open_file *of, char *buf,
275 size_t nbytes, loff_t off)
277 struct rftype *rft = of->kn->priv;
280 return rft->write(of, buf, nbytes, off);
285 static const struct kernfs_ops rdtgroup_kf_single_ops = {
286 .atomic_write_len = PAGE_SIZE,
287 .write = rdtgroup_file_write,
288 .seq_show = rdtgroup_seqfile_show,
291 static const struct kernfs_ops kf_mondata_ops = {
292 .atomic_write_len = PAGE_SIZE,
293 .seq_show = rdtgroup_mondata_show,
296 static bool is_cpu_list(struct kernfs_open_file *of)
298 struct rftype *rft = of->kn->priv;
300 return rft->flags & RFTYPE_FLAGS_CPUS_LIST;
303 static int rdtgroup_cpus_show(struct kernfs_open_file *of,
304 struct seq_file *s, void *v)
306 struct rdtgroup *rdtgrp;
307 struct cpumask *mask;
310 rdtgrp = rdtgroup_kn_lock_live(of->kn);
313 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
314 if (!rdtgrp->plr->d) {
315 rdt_last_cmd_clear();
316 rdt_last_cmd_puts("Cache domain offline\n");
319 mask = &rdtgrp->plr->d->hdr.cpu_mask;
320 seq_printf(s, is_cpu_list(of) ?
321 "%*pbl\n" : "%*pb\n",
322 cpumask_pr_args(mask));
325 seq_printf(s, is_cpu_list(of) ? "%*pbl\n" : "%*pb\n",
326 cpumask_pr_args(&rdtgrp->cpu_mask));
331 rdtgroup_kn_unlock(of->kn);
337 * This is safe against resctrl_sched_in() called from __switch_to()
338 * because __switch_to() is executed with interrupts disabled. A local call
339 * from update_closid_rmid() is protected against __switch_to() because
340 * preemption is disabled.
342 static void update_cpu_closid_rmid(void *info)
344 struct rdtgroup *r = info;
347 this_cpu_write(pqr_state.default_closid, r->closid);
348 this_cpu_write(pqr_state.default_rmid, r->mon.rmid);
352 * We cannot unconditionally write the MSR because the current
353 * executing task might have its own closid selected. Just reuse
354 * the context switch code.
356 resctrl_sched_in(current);
360 * Update the PGR_ASSOC MSR on all cpus in @cpu_mask,
362 * Per task closids/rmids must have been set up before calling this function.
365 update_closid_rmid(const struct cpumask *cpu_mask, struct rdtgroup *r)
367 on_each_cpu_mask(cpu_mask, update_cpu_closid_rmid, r, 1);
370 static int cpus_mon_write(struct rdtgroup *rdtgrp, cpumask_var_t newmask,
371 cpumask_var_t tmpmask)
373 struct rdtgroup *prgrp = rdtgrp->mon.parent, *crgrp;
374 struct list_head *head;
376 /* Check whether cpus belong to parent ctrl group */
377 cpumask_andnot(tmpmask, newmask, &prgrp->cpu_mask);
378 if (!cpumask_empty(tmpmask)) {
379 rdt_last_cmd_puts("Can only add CPUs to mongroup that belong to parent\n");
383 /* Check whether cpus are dropped from this group */
384 cpumask_andnot(tmpmask, &rdtgrp->cpu_mask, newmask);
385 if (!cpumask_empty(tmpmask)) {
386 /* Give any dropped cpus to parent rdtgroup */
387 cpumask_or(&prgrp->cpu_mask, &prgrp->cpu_mask, tmpmask);
388 update_closid_rmid(tmpmask, prgrp);
392 * If we added cpus, remove them from previous group that owned them
393 * and update per-cpu rmid
395 cpumask_andnot(tmpmask, newmask, &rdtgrp->cpu_mask);
396 if (!cpumask_empty(tmpmask)) {
397 head = &prgrp->mon.crdtgrp_list;
398 list_for_each_entry(crgrp, head, mon.crdtgrp_list) {
401 cpumask_andnot(&crgrp->cpu_mask, &crgrp->cpu_mask,
404 update_closid_rmid(tmpmask, rdtgrp);
407 /* Done pushing/pulling - update this group with new mask */
408 cpumask_copy(&rdtgrp->cpu_mask, newmask);
413 static void cpumask_rdtgrp_clear(struct rdtgroup *r, struct cpumask *m)
415 struct rdtgroup *crgrp;
417 cpumask_andnot(&r->cpu_mask, &r->cpu_mask, m);
418 /* update the child mon group masks as well*/
419 list_for_each_entry(crgrp, &r->mon.crdtgrp_list, mon.crdtgrp_list)
420 cpumask_and(&crgrp->cpu_mask, &r->cpu_mask, &crgrp->cpu_mask);
423 static int cpus_ctrl_write(struct rdtgroup *rdtgrp, cpumask_var_t newmask,
424 cpumask_var_t tmpmask, cpumask_var_t tmpmask1)
426 struct rdtgroup *r, *crgrp;
427 struct list_head *head;
429 /* Check whether cpus are dropped from this group */
430 cpumask_andnot(tmpmask, &rdtgrp->cpu_mask, newmask);
431 if (!cpumask_empty(tmpmask)) {
432 /* Can't drop from default group */
433 if (rdtgrp == &rdtgroup_default) {
434 rdt_last_cmd_puts("Can't drop CPUs from default group\n");
438 /* Give any dropped cpus to rdtgroup_default */
439 cpumask_or(&rdtgroup_default.cpu_mask,
440 &rdtgroup_default.cpu_mask, tmpmask);
441 update_closid_rmid(tmpmask, &rdtgroup_default);
445 * If we added cpus, remove them from previous group and
446 * the prev group's child groups that owned them
447 * and update per-cpu closid/rmid.
449 cpumask_andnot(tmpmask, newmask, &rdtgrp->cpu_mask);
450 if (!cpumask_empty(tmpmask)) {
451 list_for_each_entry(r, &rdt_all_groups, rdtgroup_list) {
454 cpumask_and(tmpmask1, &r->cpu_mask, tmpmask);
455 if (!cpumask_empty(tmpmask1))
456 cpumask_rdtgrp_clear(r, tmpmask1);
458 update_closid_rmid(tmpmask, rdtgrp);
461 /* Done pushing/pulling - update this group with new mask */
462 cpumask_copy(&rdtgrp->cpu_mask, newmask);
465 * Clear child mon group masks since there is a new parent mask
466 * now and update the rmid for the cpus the child lost.
468 head = &rdtgrp->mon.crdtgrp_list;
469 list_for_each_entry(crgrp, head, mon.crdtgrp_list) {
470 cpumask_and(tmpmask, &rdtgrp->cpu_mask, &crgrp->cpu_mask);
471 update_closid_rmid(tmpmask, rdtgrp);
472 cpumask_clear(&crgrp->cpu_mask);
478 static ssize_t rdtgroup_cpus_write(struct kernfs_open_file *of,
479 char *buf, size_t nbytes, loff_t off)
481 cpumask_var_t tmpmask, newmask, tmpmask1;
482 struct rdtgroup *rdtgrp;
488 if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
490 if (!zalloc_cpumask_var(&newmask, GFP_KERNEL)) {
491 free_cpumask_var(tmpmask);
494 if (!zalloc_cpumask_var(&tmpmask1, GFP_KERNEL)) {
495 free_cpumask_var(tmpmask);
496 free_cpumask_var(newmask);
500 rdtgrp = rdtgroup_kn_lock_live(of->kn);
506 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED ||
507 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
509 rdt_last_cmd_puts("Pseudo-locking in progress\n");
514 ret = cpulist_parse(buf, newmask);
516 ret = cpumask_parse(buf, newmask);
519 rdt_last_cmd_puts("Bad CPU list/mask\n");
523 /* check that user didn't specify any offline cpus */
524 cpumask_andnot(tmpmask, newmask, cpu_online_mask);
525 if (!cpumask_empty(tmpmask)) {
527 rdt_last_cmd_puts("Can only assign online CPUs\n");
531 if (rdtgrp->type == RDTCTRL_GROUP)
532 ret = cpus_ctrl_write(rdtgrp, newmask, tmpmask, tmpmask1);
533 else if (rdtgrp->type == RDTMON_GROUP)
534 ret = cpus_mon_write(rdtgrp, newmask, tmpmask);
539 rdtgroup_kn_unlock(of->kn);
540 free_cpumask_var(tmpmask);
541 free_cpumask_var(newmask);
542 free_cpumask_var(tmpmask1);
544 return ret ?: nbytes;
548 * rdtgroup_remove - the helper to remove resource group safely
549 * @rdtgrp: resource group to remove
551 * On resource group creation via a mkdir, an extra kernfs_node reference is
552 * taken to ensure that the rdtgroup structure remains accessible for the
553 * rdtgroup_kn_unlock() calls where it is removed.
555 * Drop the extra reference here, then free the rdtgroup structure.
559 static void rdtgroup_remove(struct rdtgroup *rdtgrp)
561 kernfs_put(rdtgrp->kn);
565 static void _update_task_closid_rmid(void *task)
568 * If the task is still current on this CPU, update PQR_ASSOC MSR.
569 * Otherwise, the MSR is updated when the task is scheduled in.
572 resctrl_sched_in(task);
575 static void update_task_closid_rmid(struct task_struct *t)
577 if (IS_ENABLED(CONFIG_SMP) && task_curr(t))
578 smp_call_function_single(task_cpu(t), _update_task_closid_rmid, t, 1);
580 _update_task_closid_rmid(t);
583 static bool task_in_rdtgroup(struct task_struct *tsk, struct rdtgroup *rdtgrp)
585 u32 closid, rmid = rdtgrp->mon.rmid;
587 if (rdtgrp->type == RDTCTRL_GROUP)
588 closid = rdtgrp->closid;
589 else if (rdtgrp->type == RDTMON_GROUP)
590 closid = rdtgrp->mon.parent->closid;
594 return resctrl_arch_match_closid(tsk, closid) &&
595 resctrl_arch_match_rmid(tsk, closid, rmid);
598 static int __rdtgroup_move_task(struct task_struct *tsk,
599 struct rdtgroup *rdtgrp)
601 /* If the task is already in rdtgrp, no need to move the task. */
602 if (task_in_rdtgroup(tsk, rdtgrp))
606 * Set the task's closid/rmid before the PQR_ASSOC MSR can be
609 * For ctrl_mon groups, move both closid and rmid.
610 * For monitor groups, can move the tasks only from
611 * their parent CTRL group.
613 if (rdtgrp->type == RDTMON_GROUP &&
614 !resctrl_arch_match_closid(tsk, rdtgrp->mon.parent->closid)) {
615 rdt_last_cmd_puts("Can't move task to different control group\n");
619 if (rdtgrp->type == RDTMON_GROUP)
620 resctrl_arch_set_closid_rmid(tsk, rdtgrp->mon.parent->closid,
623 resctrl_arch_set_closid_rmid(tsk, rdtgrp->closid,
627 * Ensure the task's closid and rmid are written before determining if
628 * the task is current that will decide if it will be interrupted.
629 * This pairs with the full barrier between the rq->curr update and
630 * resctrl_sched_in() during context switch.
635 * By now, the task's closid and rmid are set. If the task is current
636 * on a CPU, the PQR_ASSOC MSR needs to be updated to make the resource
637 * group go into effect. If the task is not current, the MSR will be
638 * updated when the task is scheduled in.
640 update_task_closid_rmid(tsk);
645 static bool is_closid_match(struct task_struct *t, struct rdtgroup *r)
647 return (resctrl_arch_alloc_capable() && (r->type == RDTCTRL_GROUP) &&
648 resctrl_arch_match_closid(t, r->closid));
651 static bool is_rmid_match(struct task_struct *t, struct rdtgroup *r)
653 return (resctrl_arch_mon_capable() && (r->type == RDTMON_GROUP) &&
654 resctrl_arch_match_rmid(t, r->mon.parent->closid,
659 * rdtgroup_tasks_assigned - Test if tasks have been assigned to resource group
662 * Return: 1 if tasks have been assigned to @r, 0 otherwise
664 int rdtgroup_tasks_assigned(struct rdtgroup *r)
666 struct task_struct *p, *t;
669 lockdep_assert_held(&rdtgroup_mutex);
672 for_each_process_thread(p, t) {
673 if (is_closid_match(t, r) || is_rmid_match(t, r)) {
683 static int rdtgroup_task_write_permission(struct task_struct *task,
684 struct kernfs_open_file *of)
686 const struct cred *tcred = get_task_cred(task);
687 const struct cred *cred = current_cred();
691 * Even if we're attaching all tasks in the thread group, we only
692 * need to check permissions on one of them.
694 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
695 !uid_eq(cred->euid, tcred->uid) &&
696 !uid_eq(cred->euid, tcred->suid)) {
697 rdt_last_cmd_printf("No permission to move task %d\n", task->pid);
705 static int rdtgroup_move_task(pid_t pid, struct rdtgroup *rdtgrp,
706 struct kernfs_open_file *of)
708 struct task_struct *tsk;
713 tsk = find_task_by_vpid(pid);
716 rdt_last_cmd_printf("No task %d\n", pid);
723 get_task_struct(tsk);
726 ret = rdtgroup_task_write_permission(tsk, of);
728 ret = __rdtgroup_move_task(tsk, rdtgrp);
730 put_task_struct(tsk);
734 static ssize_t rdtgroup_tasks_write(struct kernfs_open_file *of,
735 char *buf, size_t nbytes, loff_t off)
737 struct rdtgroup *rdtgrp;
742 rdtgrp = rdtgroup_kn_lock_live(of->kn);
744 rdtgroup_kn_unlock(of->kn);
747 rdt_last_cmd_clear();
749 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED ||
750 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
752 rdt_last_cmd_puts("Pseudo-locking in progress\n");
756 while (buf && buf[0] != '\0' && buf[0] != '\n') {
757 pid_str = strim(strsep(&buf, ","));
759 if (kstrtoint(pid_str, 0, &pid)) {
760 rdt_last_cmd_printf("Task list parsing error pid %s\n", pid_str);
766 rdt_last_cmd_printf("Invalid pid %d\n", pid);
771 ret = rdtgroup_move_task(pid, rdtgrp, of);
773 rdt_last_cmd_printf("Error while processing task %d\n", pid);
779 rdtgroup_kn_unlock(of->kn);
781 return ret ?: nbytes;
784 static void show_rdt_tasks(struct rdtgroup *r, struct seq_file *s)
786 struct task_struct *p, *t;
790 for_each_process_thread(p, t) {
791 if (is_closid_match(t, r) || is_rmid_match(t, r)) {
792 pid = task_pid_vnr(t);
794 seq_printf(s, "%d\n", pid);
800 static int rdtgroup_tasks_show(struct kernfs_open_file *of,
801 struct seq_file *s, void *v)
803 struct rdtgroup *rdtgrp;
806 rdtgrp = rdtgroup_kn_lock_live(of->kn);
808 show_rdt_tasks(rdtgrp, s);
811 rdtgroup_kn_unlock(of->kn);
816 static int rdtgroup_closid_show(struct kernfs_open_file *of,
817 struct seq_file *s, void *v)
819 struct rdtgroup *rdtgrp;
822 rdtgrp = rdtgroup_kn_lock_live(of->kn);
824 seq_printf(s, "%u\n", rdtgrp->closid);
827 rdtgroup_kn_unlock(of->kn);
832 static int rdtgroup_rmid_show(struct kernfs_open_file *of,
833 struct seq_file *s, void *v)
835 struct rdtgroup *rdtgrp;
838 rdtgrp = rdtgroup_kn_lock_live(of->kn);
840 seq_printf(s, "%u\n", rdtgrp->mon.rmid);
843 rdtgroup_kn_unlock(of->kn);
848 #ifdef CONFIG_PROC_CPU_RESCTRL
851 * A task can only be part of one resctrl control group and of one monitor
852 * group which is associated to that control group.
857 * resctrl is not available.
862 * Task is part of the root resctrl control group, and it is not associated
863 * to any monitor group.
868 * Task is part of the root resctrl control group and monitor group mon0.
873 * Task is part of resctrl control group group0, and it is not associated
874 * to any monitor group.
879 * Task is part of resctrl control group group0 and monitor group mon1.
881 int proc_resctrl_show(struct seq_file *s, struct pid_namespace *ns,
882 struct pid *pid, struct task_struct *tsk)
884 struct rdtgroup *rdtg;
887 mutex_lock(&rdtgroup_mutex);
889 /* Return empty if resctrl has not been mounted. */
890 if (!resctrl_mounted) {
891 seq_puts(s, "res:\nmon:\n");
895 list_for_each_entry(rdtg, &rdt_all_groups, rdtgroup_list) {
896 struct rdtgroup *crg;
899 * Task information is only relevant for shareable
900 * and exclusive groups.
902 if (rdtg->mode != RDT_MODE_SHAREABLE &&
903 rdtg->mode != RDT_MODE_EXCLUSIVE)
906 if (!resctrl_arch_match_closid(tsk, rdtg->closid))
909 seq_printf(s, "res:%s%s\n", (rdtg == &rdtgroup_default) ? "/" : "",
912 list_for_each_entry(crg, &rdtg->mon.crdtgrp_list,
914 if (!resctrl_arch_match_rmid(tsk, crg->mon.parent->closid,
917 seq_printf(s, "%s", crg->kn->name);
924 * The above search should succeed. Otherwise return
929 mutex_unlock(&rdtgroup_mutex);
935 static int rdt_last_cmd_status_show(struct kernfs_open_file *of,
936 struct seq_file *seq, void *v)
940 mutex_lock(&rdtgroup_mutex);
941 len = seq_buf_used(&last_cmd_status);
943 seq_printf(seq, "%.*s", len, last_cmd_status_buf);
945 seq_puts(seq, "ok\n");
946 mutex_unlock(&rdtgroup_mutex);
950 static int rdt_num_closids_show(struct kernfs_open_file *of,
951 struct seq_file *seq, void *v)
953 struct resctrl_schema *s = of->kn->parent->priv;
955 seq_printf(seq, "%u\n", s->num_closid);
959 static int rdt_default_ctrl_show(struct kernfs_open_file *of,
960 struct seq_file *seq, void *v)
962 struct resctrl_schema *s = of->kn->parent->priv;
963 struct rdt_resource *r = s->res;
965 seq_printf(seq, "%x\n", r->default_ctrl);
969 static int rdt_min_cbm_bits_show(struct kernfs_open_file *of,
970 struct seq_file *seq, void *v)
972 struct resctrl_schema *s = of->kn->parent->priv;
973 struct rdt_resource *r = s->res;
975 seq_printf(seq, "%u\n", r->cache.min_cbm_bits);
979 static int rdt_shareable_bits_show(struct kernfs_open_file *of,
980 struct seq_file *seq, void *v)
982 struct resctrl_schema *s = of->kn->parent->priv;
983 struct rdt_resource *r = s->res;
985 seq_printf(seq, "%x\n", r->cache.shareable_bits);
990 * rdt_bit_usage_show - Display current usage of resources
992 * A domain is a shared resource that can now be allocated differently. Here
993 * we display the current regions of the domain as an annotated bitmask.
994 * For each domain of this resource its allocation bitmask
995 * is annotated as below to indicate the current usage of the corresponding bit:
996 * 0 - currently unused
997 * X - currently available for sharing and used by software and hardware
998 * H - currently used by hardware only but available for software use
999 * S - currently used and shareable by software only
1000 * E - currently used exclusively by one resource group
1001 * P - currently pseudo-locked by one resource group
1003 static int rdt_bit_usage_show(struct kernfs_open_file *of,
1004 struct seq_file *seq, void *v)
1006 struct resctrl_schema *s = of->kn->parent->priv;
1008 * Use unsigned long even though only 32 bits are used to ensure
1009 * test_bit() is used safely.
1011 unsigned long sw_shareable = 0, hw_shareable = 0;
1012 unsigned long exclusive = 0, pseudo_locked = 0;
1013 struct rdt_resource *r = s->res;
1014 struct rdt_ctrl_domain *dom;
1015 int i, hwb, swb, excl, psl;
1016 enum rdtgrp_mode mode;
1021 mutex_lock(&rdtgroup_mutex);
1022 hw_shareable = r->cache.shareable_bits;
1023 list_for_each_entry(dom, &r->ctrl_domains, hdr.list) {
1028 seq_printf(seq, "%d=", dom->hdr.id);
1029 for (i = 0; i < closids_supported(); i++) {
1030 if (!closid_allocated(i))
1032 ctrl_val = resctrl_arch_get_config(r, dom, i,
1034 mode = rdtgroup_mode_by_closid(i);
1036 case RDT_MODE_SHAREABLE:
1037 sw_shareable |= ctrl_val;
1039 case RDT_MODE_EXCLUSIVE:
1040 exclusive |= ctrl_val;
1042 case RDT_MODE_PSEUDO_LOCKSETUP:
1044 * RDT_MODE_PSEUDO_LOCKSETUP is possible
1045 * here but not included since the CBM
1046 * associated with this CLOSID in this mode
1047 * is not initialized and no task or cpu can be
1048 * assigned this CLOSID.
1051 case RDT_MODE_PSEUDO_LOCKED:
1054 "invalid mode for closid %d\n", i);
1058 for (i = r->cache.cbm_len - 1; i >= 0; i--) {
1059 pseudo_locked = dom->plr ? dom->plr->cbm : 0;
1060 hwb = test_bit(i, &hw_shareable);
1061 swb = test_bit(i, &sw_shareable);
1062 excl = test_bit(i, &exclusive);
1063 psl = test_bit(i, &pseudo_locked);
1066 else if (hwb && !swb)
1068 else if (!hwb && swb)
1074 else /* Unused bits remain */
1079 seq_putc(seq, '\n');
1080 mutex_unlock(&rdtgroup_mutex);
1085 static int rdt_min_bw_show(struct kernfs_open_file *of,
1086 struct seq_file *seq, void *v)
1088 struct resctrl_schema *s = of->kn->parent->priv;
1089 struct rdt_resource *r = s->res;
1091 seq_printf(seq, "%u\n", r->membw.min_bw);
1095 static int rdt_num_rmids_show(struct kernfs_open_file *of,
1096 struct seq_file *seq, void *v)
1098 struct rdt_resource *r = of->kn->parent->priv;
1100 seq_printf(seq, "%d\n", r->num_rmid);
1105 static int rdt_mon_features_show(struct kernfs_open_file *of,
1106 struct seq_file *seq, void *v)
1108 struct rdt_resource *r = of->kn->parent->priv;
1109 struct mon_evt *mevt;
1111 list_for_each_entry(mevt, &r->evt_list, list) {
1112 seq_printf(seq, "%s\n", mevt->name);
1113 if (mevt->configurable)
1114 seq_printf(seq, "%s_config\n", mevt->name);
1120 static int rdt_bw_gran_show(struct kernfs_open_file *of,
1121 struct seq_file *seq, void *v)
1123 struct resctrl_schema *s = of->kn->parent->priv;
1124 struct rdt_resource *r = s->res;
1126 seq_printf(seq, "%u\n", r->membw.bw_gran);
1130 static int rdt_delay_linear_show(struct kernfs_open_file *of,
1131 struct seq_file *seq, void *v)
1133 struct resctrl_schema *s = of->kn->parent->priv;
1134 struct rdt_resource *r = s->res;
1136 seq_printf(seq, "%u\n", r->membw.delay_linear);
1140 static int max_threshold_occ_show(struct kernfs_open_file *of,
1141 struct seq_file *seq, void *v)
1143 seq_printf(seq, "%u\n", resctrl_rmid_realloc_threshold);
1148 static int rdt_thread_throttle_mode_show(struct kernfs_open_file *of,
1149 struct seq_file *seq, void *v)
1151 struct resctrl_schema *s = of->kn->parent->priv;
1152 struct rdt_resource *r = s->res;
1154 if (r->membw.throttle_mode == THREAD_THROTTLE_PER_THREAD)
1155 seq_puts(seq, "per-thread\n");
1157 seq_puts(seq, "max\n");
1162 static ssize_t max_threshold_occ_write(struct kernfs_open_file *of,
1163 char *buf, size_t nbytes, loff_t off)
1168 ret = kstrtouint(buf, 0, &bytes);
1172 if (bytes > resctrl_rmid_realloc_limit)
1175 resctrl_rmid_realloc_threshold = resctrl_arch_round_mon_val(bytes);
1181 * rdtgroup_mode_show - Display mode of this resource group
1183 static int rdtgroup_mode_show(struct kernfs_open_file *of,
1184 struct seq_file *s, void *v)
1186 struct rdtgroup *rdtgrp;
1188 rdtgrp = rdtgroup_kn_lock_live(of->kn);
1190 rdtgroup_kn_unlock(of->kn);
1194 seq_printf(s, "%s\n", rdtgroup_mode_str(rdtgrp->mode));
1196 rdtgroup_kn_unlock(of->kn);
1200 static enum resctrl_conf_type resctrl_peer_type(enum resctrl_conf_type my_type)
1213 static int rdt_has_sparse_bitmasks_show(struct kernfs_open_file *of,
1214 struct seq_file *seq, void *v)
1216 struct resctrl_schema *s = of->kn->parent->priv;
1217 struct rdt_resource *r = s->res;
1219 seq_printf(seq, "%u\n", r->cache.arch_has_sparse_bitmasks);
1225 * __rdtgroup_cbm_overlaps - Does CBM for intended closid overlap with other
1226 * @r: Resource to which domain instance @d belongs.
1227 * @d: The domain instance for which @closid is being tested.
1228 * @cbm: Capacity bitmask being tested.
1229 * @closid: Intended closid for @cbm.
1230 * @type: CDP type of @r.
1231 * @exclusive: Only check if overlaps with exclusive resource groups
1233 * Checks if provided @cbm intended to be used for @closid on domain
1234 * @d overlaps with any other closids or other hardware usage associated
1235 * with this domain. If @exclusive is true then only overlaps with
1236 * resource groups in exclusive mode will be considered. If @exclusive
1237 * is false then overlaps with any resource group or hardware entities
1238 * will be considered.
1240 * @cbm is unsigned long, even if only 32 bits are used, to make the
1241 * bitmap functions work correctly.
1243 * Return: false if CBM does not overlap, true if it does.
1245 static bool __rdtgroup_cbm_overlaps(struct rdt_resource *r, struct rdt_ctrl_domain *d,
1246 unsigned long cbm, int closid,
1247 enum resctrl_conf_type type, bool exclusive)
1249 enum rdtgrp_mode mode;
1250 unsigned long ctrl_b;
1253 /* Check for any overlap with regions used by hardware directly */
1255 ctrl_b = r->cache.shareable_bits;
1256 if (bitmap_intersects(&cbm, &ctrl_b, r->cache.cbm_len))
1260 /* Check for overlap with other resource groups */
1261 for (i = 0; i < closids_supported(); i++) {
1262 ctrl_b = resctrl_arch_get_config(r, d, i, type);
1263 mode = rdtgroup_mode_by_closid(i);
1264 if (closid_allocated(i) && i != closid &&
1265 mode != RDT_MODE_PSEUDO_LOCKSETUP) {
1266 if (bitmap_intersects(&cbm, &ctrl_b, r->cache.cbm_len)) {
1268 if (mode == RDT_MODE_EXCLUSIVE)
1281 * rdtgroup_cbm_overlaps - Does CBM overlap with other use of hardware
1282 * @s: Schema for the resource to which domain instance @d belongs.
1283 * @d: The domain instance for which @closid is being tested.
1284 * @cbm: Capacity bitmask being tested.
1285 * @closid: Intended closid for @cbm.
1286 * @exclusive: Only check if overlaps with exclusive resource groups
1288 * Resources that can be allocated using a CBM can use the CBM to control
1289 * the overlap of these allocations. rdtgroup_cmb_overlaps() is the test
1290 * for overlap. Overlap test is not limited to the specific resource for
1291 * which the CBM is intended though - when dealing with CDP resources that
1292 * share the underlying hardware the overlap check should be performed on
1293 * the CDP resource sharing the hardware also.
1295 * Refer to description of __rdtgroup_cbm_overlaps() for the details of the
1298 * Return: true if CBM overlap detected, false if there is no overlap
1300 bool rdtgroup_cbm_overlaps(struct resctrl_schema *s, struct rdt_ctrl_domain *d,
1301 unsigned long cbm, int closid, bool exclusive)
1303 enum resctrl_conf_type peer_type = resctrl_peer_type(s->conf_type);
1304 struct rdt_resource *r = s->res;
1306 if (__rdtgroup_cbm_overlaps(r, d, cbm, closid, s->conf_type,
1310 if (!resctrl_arch_get_cdp_enabled(r->rid))
1312 return __rdtgroup_cbm_overlaps(r, d, cbm, closid, peer_type, exclusive);
1316 * rdtgroup_mode_test_exclusive - Test if this resource group can be exclusive
1317 * @rdtgrp: Resource group identified through its closid.
1319 * An exclusive resource group implies that there should be no sharing of
1320 * its allocated resources. At the time this group is considered to be
1321 * exclusive this test can determine if its current schemata supports this
1322 * setting by testing for overlap with all other resource groups.
1324 * Return: true if resource group can be exclusive, false if there is overlap
1325 * with allocations of other resource groups and thus this resource group
1326 * cannot be exclusive.
1328 static bool rdtgroup_mode_test_exclusive(struct rdtgroup *rdtgrp)
1330 int closid = rdtgrp->closid;
1331 struct rdt_ctrl_domain *d;
1332 struct resctrl_schema *s;
1333 struct rdt_resource *r;
1334 bool has_cache = false;
1337 /* Walking r->domains, ensure it can't race with cpuhp */
1338 lockdep_assert_cpus_held();
1340 list_for_each_entry(s, &resctrl_schema_all, list) {
1342 if (r->rid == RDT_RESOURCE_MBA || r->rid == RDT_RESOURCE_SMBA)
1345 list_for_each_entry(d, &r->ctrl_domains, hdr.list) {
1346 ctrl = resctrl_arch_get_config(r, d, closid,
1348 if (rdtgroup_cbm_overlaps(s, d, ctrl, closid, false)) {
1349 rdt_last_cmd_puts("Schemata overlaps\n");
1356 rdt_last_cmd_puts("Cannot be exclusive without CAT/CDP\n");
1364 * rdtgroup_mode_write - Modify the resource group's mode
1366 static ssize_t rdtgroup_mode_write(struct kernfs_open_file *of,
1367 char *buf, size_t nbytes, loff_t off)
1369 struct rdtgroup *rdtgrp;
1370 enum rdtgrp_mode mode;
1373 /* Valid input requires a trailing newline */
1374 if (nbytes == 0 || buf[nbytes - 1] != '\n')
1376 buf[nbytes - 1] = '\0';
1378 rdtgrp = rdtgroup_kn_lock_live(of->kn);
1380 rdtgroup_kn_unlock(of->kn);
1384 rdt_last_cmd_clear();
1386 mode = rdtgrp->mode;
1388 if ((!strcmp(buf, "shareable") && mode == RDT_MODE_SHAREABLE) ||
1389 (!strcmp(buf, "exclusive") && mode == RDT_MODE_EXCLUSIVE) ||
1390 (!strcmp(buf, "pseudo-locksetup") &&
1391 mode == RDT_MODE_PSEUDO_LOCKSETUP) ||
1392 (!strcmp(buf, "pseudo-locked") && mode == RDT_MODE_PSEUDO_LOCKED))
1395 if (mode == RDT_MODE_PSEUDO_LOCKED) {
1396 rdt_last_cmd_puts("Cannot change pseudo-locked group\n");
1401 if (!strcmp(buf, "shareable")) {
1402 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1403 ret = rdtgroup_locksetup_exit(rdtgrp);
1407 rdtgrp->mode = RDT_MODE_SHAREABLE;
1408 } else if (!strcmp(buf, "exclusive")) {
1409 if (!rdtgroup_mode_test_exclusive(rdtgrp)) {
1413 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1414 ret = rdtgroup_locksetup_exit(rdtgrp);
1418 rdtgrp->mode = RDT_MODE_EXCLUSIVE;
1419 } else if (!strcmp(buf, "pseudo-locksetup")) {
1420 ret = rdtgroup_locksetup_enter(rdtgrp);
1423 rdtgrp->mode = RDT_MODE_PSEUDO_LOCKSETUP;
1425 rdt_last_cmd_puts("Unknown or unsupported mode\n");
1430 rdtgroup_kn_unlock(of->kn);
1431 return ret ?: nbytes;
1435 * rdtgroup_cbm_to_size - Translate CBM to size in bytes
1436 * @r: RDT resource to which @d belongs.
1437 * @d: RDT domain instance.
1438 * @cbm: bitmask for which the size should be computed.
1440 * The bitmask provided associated with the RDT domain instance @d will be
1441 * translated into how many bytes it represents. The size in bytes is
1442 * computed by first dividing the total cache size by the CBM length to
1443 * determine how many bytes each bit in the bitmask represents. The result
1444 * is multiplied with the number of bits set in the bitmask.
1446 * @cbm is unsigned long, even if only 32 bits are used to make the
1447 * bitmap functions work correctly.
1449 unsigned int rdtgroup_cbm_to_size(struct rdt_resource *r,
1450 struct rdt_ctrl_domain *d, unsigned long cbm)
1452 unsigned int size = 0;
1453 struct cacheinfo *ci;
1456 if (WARN_ON_ONCE(r->ctrl_scope != RESCTRL_L2_CACHE && r->ctrl_scope != RESCTRL_L3_CACHE))
1459 num_b = bitmap_weight(&cbm, r->cache.cbm_len);
1460 ci = get_cpu_cacheinfo_level(cpumask_any(&d->hdr.cpu_mask), r->ctrl_scope);
1462 size = ci->size / r->cache.cbm_len * num_b;
1468 * rdtgroup_size_show - Display size in bytes of allocated regions
1470 * The "size" file mirrors the layout of the "schemata" file, printing the
1471 * size in bytes of each region instead of the capacity bitmask.
1473 static int rdtgroup_size_show(struct kernfs_open_file *of,
1474 struct seq_file *s, void *v)
1476 struct resctrl_schema *schema;
1477 enum resctrl_conf_type type;
1478 struct rdt_ctrl_domain *d;
1479 struct rdtgroup *rdtgrp;
1480 struct rdt_resource *r;
1487 rdtgrp = rdtgroup_kn_lock_live(of->kn);
1489 rdtgroup_kn_unlock(of->kn);
1493 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
1494 if (!rdtgrp->plr->d) {
1495 rdt_last_cmd_clear();
1496 rdt_last_cmd_puts("Cache domain offline\n");
1499 seq_printf(s, "%*s:", max_name_width,
1500 rdtgrp->plr->s->name);
1501 size = rdtgroup_cbm_to_size(rdtgrp->plr->s->res,
1504 seq_printf(s, "%d=%u\n", rdtgrp->plr->d->hdr.id, size);
1509 closid = rdtgrp->closid;
1511 list_for_each_entry(schema, &resctrl_schema_all, list) {
1513 type = schema->conf_type;
1515 seq_printf(s, "%*s:", max_name_width, schema->name);
1516 list_for_each_entry(d, &r->ctrl_domains, hdr.list) {
1519 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1523 ctrl = d->mbps_val[closid];
1525 ctrl = resctrl_arch_get_config(r, d,
1528 if (r->rid == RDT_RESOURCE_MBA ||
1529 r->rid == RDT_RESOURCE_SMBA)
1532 size = rdtgroup_cbm_to_size(r, d, ctrl);
1534 seq_printf(s, "%d=%u", d->hdr.id, size);
1541 rdtgroup_kn_unlock(of->kn);
1546 struct mon_config_info {
1551 #define INVALID_CONFIG_INDEX UINT_MAX
1554 * mon_event_config_index_get - get the hardware index for the
1555 * configurable event
1558 * Return: 0 for evtid == QOS_L3_MBM_TOTAL_EVENT_ID
1559 * 1 for evtid == QOS_L3_MBM_LOCAL_EVENT_ID
1560 * INVALID_CONFIG_INDEX for invalid evtid
1562 static inline unsigned int mon_event_config_index_get(u32 evtid)
1565 case QOS_L3_MBM_TOTAL_EVENT_ID:
1567 case QOS_L3_MBM_LOCAL_EVENT_ID:
1570 /* Should never reach here */
1571 return INVALID_CONFIG_INDEX;
1575 static void mon_event_config_read(void *info)
1577 struct mon_config_info *mon_info = info;
1581 index = mon_event_config_index_get(mon_info->evtid);
1582 if (index == INVALID_CONFIG_INDEX) {
1583 pr_warn_once("Invalid event id %d\n", mon_info->evtid);
1586 rdmsrl(MSR_IA32_EVT_CFG_BASE + index, msrval);
1588 /* Report only the valid event configuration bits */
1589 mon_info->mon_config = msrval & MAX_EVT_CONFIG_BITS;
1592 static void mondata_config_read(struct rdt_mon_domain *d, struct mon_config_info *mon_info)
1594 smp_call_function_any(&d->hdr.cpu_mask, mon_event_config_read, mon_info, 1);
1597 static int mbm_config_show(struct seq_file *s, struct rdt_resource *r, u32 evtid)
1599 struct mon_config_info mon_info;
1600 struct rdt_mon_domain *dom;
1604 mutex_lock(&rdtgroup_mutex);
1606 list_for_each_entry(dom, &r->mon_domains, hdr.list) {
1610 memset(&mon_info, 0, sizeof(struct mon_config_info));
1611 mon_info.evtid = evtid;
1612 mondata_config_read(dom, &mon_info);
1614 seq_printf(s, "%d=0x%02x", dom->hdr.id, mon_info.mon_config);
1619 mutex_unlock(&rdtgroup_mutex);
1625 static int mbm_total_bytes_config_show(struct kernfs_open_file *of,
1626 struct seq_file *seq, void *v)
1628 struct rdt_resource *r = of->kn->parent->priv;
1630 mbm_config_show(seq, r, QOS_L3_MBM_TOTAL_EVENT_ID);
1635 static int mbm_local_bytes_config_show(struct kernfs_open_file *of,
1636 struct seq_file *seq, void *v)
1638 struct rdt_resource *r = of->kn->parent->priv;
1640 mbm_config_show(seq, r, QOS_L3_MBM_LOCAL_EVENT_ID);
1645 static void mon_event_config_write(void *info)
1647 struct mon_config_info *mon_info = info;
1650 index = mon_event_config_index_get(mon_info->evtid);
1651 if (index == INVALID_CONFIG_INDEX) {
1652 pr_warn_once("Invalid event id %d\n", mon_info->evtid);
1655 wrmsr(MSR_IA32_EVT_CFG_BASE + index, mon_info->mon_config, 0);
1658 static void mbm_config_write_domain(struct rdt_resource *r,
1659 struct rdt_mon_domain *d, u32 evtid, u32 val)
1661 struct mon_config_info mon_info = {0};
1664 * Read the current config value first. If both are the same then
1665 * no need to write it again.
1667 mon_info.evtid = evtid;
1668 mondata_config_read(d, &mon_info);
1669 if (mon_info.mon_config == val)
1672 mon_info.mon_config = val;
1675 * Update MSR_IA32_EVT_CFG_BASE MSR on one of the CPUs in the
1676 * domain. The MSRs offset from MSR MSR_IA32_EVT_CFG_BASE
1677 * are scoped at the domain level. Writing any of these MSRs
1678 * on one CPU is observed by all the CPUs in the domain.
1680 smp_call_function_any(&d->hdr.cpu_mask, mon_event_config_write,
1684 * When an Event Configuration is changed, the bandwidth counters
1685 * for all RMIDs and Events will be cleared by the hardware. The
1686 * hardware also sets MSR_IA32_QM_CTR.Unavailable (bit 62) for
1687 * every RMID on the next read to any event for every RMID.
1688 * Subsequent reads will have MSR_IA32_QM_CTR.Unavailable (bit 62)
1689 * cleared while it is tracked by the hardware. Clear the
1690 * mbm_local and mbm_total counts for all the RMIDs.
1692 resctrl_arch_reset_rmid_all(r, d);
1695 static int mon_config_write(struct rdt_resource *r, char *tok, u32 evtid)
1697 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
1698 char *dom_str = NULL, *id_str;
1699 unsigned long dom_id, val;
1700 struct rdt_mon_domain *d;
1702 /* Walking r->domains, ensure it can't race with cpuhp */
1703 lockdep_assert_cpus_held();
1706 if (!tok || tok[0] == '\0')
1709 /* Start processing the strings for each domain */
1710 dom_str = strim(strsep(&tok, ";"));
1711 id_str = strsep(&dom_str, "=");
1713 if (!id_str || kstrtoul(id_str, 10, &dom_id)) {
1714 rdt_last_cmd_puts("Missing '=' or non-numeric domain id\n");
1718 if (!dom_str || kstrtoul(dom_str, 16, &val)) {
1719 rdt_last_cmd_puts("Non-numeric event configuration value\n");
1723 /* Value from user cannot be more than the supported set of events */
1724 if ((val & hw_res->mbm_cfg_mask) != val) {
1725 rdt_last_cmd_printf("Invalid event configuration: max valid mask is 0x%02x\n",
1726 hw_res->mbm_cfg_mask);
1730 list_for_each_entry(d, &r->mon_domains, hdr.list) {
1731 if (d->hdr.id == dom_id) {
1732 mbm_config_write_domain(r, d, evtid, val);
1740 static ssize_t mbm_total_bytes_config_write(struct kernfs_open_file *of,
1741 char *buf, size_t nbytes,
1744 struct rdt_resource *r = of->kn->parent->priv;
1747 /* Valid input requires a trailing newline */
1748 if (nbytes == 0 || buf[nbytes - 1] != '\n')
1752 mutex_lock(&rdtgroup_mutex);
1754 rdt_last_cmd_clear();
1756 buf[nbytes - 1] = '\0';
1758 ret = mon_config_write(r, buf, QOS_L3_MBM_TOTAL_EVENT_ID);
1760 mutex_unlock(&rdtgroup_mutex);
1763 return ret ?: nbytes;
1766 static ssize_t mbm_local_bytes_config_write(struct kernfs_open_file *of,
1767 char *buf, size_t nbytes,
1770 struct rdt_resource *r = of->kn->parent->priv;
1773 /* Valid input requires a trailing newline */
1774 if (nbytes == 0 || buf[nbytes - 1] != '\n')
1778 mutex_lock(&rdtgroup_mutex);
1780 rdt_last_cmd_clear();
1782 buf[nbytes - 1] = '\0';
1784 ret = mon_config_write(r, buf, QOS_L3_MBM_LOCAL_EVENT_ID);
1786 mutex_unlock(&rdtgroup_mutex);
1789 return ret ?: nbytes;
1792 /* rdtgroup information files for one cache resource. */
1793 static struct rftype res_common_files[] = {
1795 .name = "last_cmd_status",
1797 .kf_ops = &rdtgroup_kf_single_ops,
1798 .seq_show = rdt_last_cmd_status_show,
1799 .fflags = RFTYPE_TOP_INFO,
1802 .name = "num_closids",
1804 .kf_ops = &rdtgroup_kf_single_ops,
1805 .seq_show = rdt_num_closids_show,
1806 .fflags = RFTYPE_CTRL_INFO,
1809 .name = "mon_features",
1811 .kf_ops = &rdtgroup_kf_single_ops,
1812 .seq_show = rdt_mon_features_show,
1813 .fflags = RFTYPE_MON_INFO,
1816 .name = "num_rmids",
1818 .kf_ops = &rdtgroup_kf_single_ops,
1819 .seq_show = rdt_num_rmids_show,
1820 .fflags = RFTYPE_MON_INFO,
1825 .kf_ops = &rdtgroup_kf_single_ops,
1826 .seq_show = rdt_default_ctrl_show,
1827 .fflags = RFTYPE_CTRL_INFO | RFTYPE_RES_CACHE,
1830 .name = "min_cbm_bits",
1832 .kf_ops = &rdtgroup_kf_single_ops,
1833 .seq_show = rdt_min_cbm_bits_show,
1834 .fflags = RFTYPE_CTRL_INFO | RFTYPE_RES_CACHE,
1837 .name = "shareable_bits",
1839 .kf_ops = &rdtgroup_kf_single_ops,
1840 .seq_show = rdt_shareable_bits_show,
1841 .fflags = RFTYPE_CTRL_INFO | RFTYPE_RES_CACHE,
1844 .name = "bit_usage",
1846 .kf_ops = &rdtgroup_kf_single_ops,
1847 .seq_show = rdt_bit_usage_show,
1848 .fflags = RFTYPE_CTRL_INFO | RFTYPE_RES_CACHE,
1851 .name = "min_bandwidth",
1853 .kf_ops = &rdtgroup_kf_single_ops,
1854 .seq_show = rdt_min_bw_show,
1855 .fflags = RFTYPE_CTRL_INFO | RFTYPE_RES_MB,
1858 .name = "bandwidth_gran",
1860 .kf_ops = &rdtgroup_kf_single_ops,
1861 .seq_show = rdt_bw_gran_show,
1862 .fflags = RFTYPE_CTRL_INFO | RFTYPE_RES_MB,
1865 .name = "delay_linear",
1867 .kf_ops = &rdtgroup_kf_single_ops,
1868 .seq_show = rdt_delay_linear_show,
1869 .fflags = RFTYPE_CTRL_INFO | RFTYPE_RES_MB,
1872 * Platform specific which (if any) capabilities are provided by
1873 * thread_throttle_mode. Defer "fflags" initialization to platform
1877 .name = "thread_throttle_mode",
1879 .kf_ops = &rdtgroup_kf_single_ops,
1880 .seq_show = rdt_thread_throttle_mode_show,
1883 .name = "max_threshold_occupancy",
1885 .kf_ops = &rdtgroup_kf_single_ops,
1886 .write = max_threshold_occ_write,
1887 .seq_show = max_threshold_occ_show,
1888 .fflags = RFTYPE_MON_INFO | RFTYPE_RES_CACHE,
1891 .name = "mbm_total_bytes_config",
1893 .kf_ops = &rdtgroup_kf_single_ops,
1894 .seq_show = mbm_total_bytes_config_show,
1895 .write = mbm_total_bytes_config_write,
1898 .name = "mbm_local_bytes_config",
1900 .kf_ops = &rdtgroup_kf_single_ops,
1901 .seq_show = mbm_local_bytes_config_show,
1902 .write = mbm_local_bytes_config_write,
1907 .kf_ops = &rdtgroup_kf_single_ops,
1908 .write = rdtgroup_cpus_write,
1909 .seq_show = rdtgroup_cpus_show,
1910 .fflags = RFTYPE_BASE,
1913 .name = "cpus_list",
1915 .kf_ops = &rdtgroup_kf_single_ops,
1916 .write = rdtgroup_cpus_write,
1917 .seq_show = rdtgroup_cpus_show,
1918 .flags = RFTYPE_FLAGS_CPUS_LIST,
1919 .fflags = RFTYPE_BASE,
1924 .kf_ops = &rdtgroup_kf_single_ops,
1925 .write = rdtgroup_tasks_write,
1926 .seq_show = rdtgroup_tasks_show,
1927 .fflags = RFTYPE_BASE,
1930 .name = "mon_hw_id",
1932 .kf_ops = &rdtgroup_kf_single_ops,
1933 .seq_show = rdtgroup_rmid_show,
1934 .fflags = RFTYPE_MON_BASE | RFTYPE_DEBUG,
1939 .kf_ops = &rdtgroup_kf_single_ops,
1940 .write = rdtgroup_schemata_write,
1941 .seq_show = rdtgroup_schemata_show,
1942 .fflags = RFTYPE_CTRL_BASE,
1947 .kf_ops = &rdtgroup_kf_single_ops,
1948 .write = rdtgroup_mode_write,
1949 .seq_show = rdtgroup_mode_show,
1950 .fflags = RFTYPE_CTRL_BASE,
1955 .kf_ops = &rdtgroup_kf_single_ops,
1956 .seq_show = rdtgroup_size_show,
1957 .fflags = RFTYPE_CTRL_BASE,
1960 .name = "sparse_masks",
1962 .kf_ops = &rdtgroup_kf_single_ops,
1963 .seq_show = rdt_has_sparse_bitmasks_show,
1964 .fflags = RFTYPE_CTRL_INFO | RFTYPE_RES_CACHE,
1967 .name = "ctrl_hw_id",
1969 .kf_ops = &rdtgroup_kf_single_ops,
1970 .seq_show = rdtgroup_closid_show,
1971 .fflags = RFTYPE_CTRL_BASE | RFTYPE_DEBUG,
1976 static int rdtgroup_add_files(struct kernfs_node *kn, unsigned long fflags)
1978 struct rftype *rfts, *rft;
1981 rfts = res_common_files;
1982 len = ARRAY_SIZE(res_common_files);
1984 lockdep_assert_held(&rdtgroup_mutex);
1987 fflags |= RFTYPE_DEBUG;
1989 for (rft = rfts; rft < rfts + len; rft++) {
1990 if (rft->fflags && ((fflags & rft->fflags) == rft->fflags)) {
1991 ret = rdtgroup_add_file(kn, rft);
1999 pr_warn("Failed to add %s, err=%d\n", rft->name, ret);
2000 while (--rft >= rfts) {
2001 if ((fflags & rft->fflags) == rft->fflags)
2002 kernfs_remove_by_name(kn, rft->name);
2007 static struct rftype *rdtgroup_get_rftype_by_name(const char *name)
2009 struct rftype *rfts, *rft;
2012 rfts = res_common_files;
2013 len = ARRAY_SIZE(res_common_files);
2015 for (rft = rfts; rft < rfts + len; rft++) {
2016 if (!strcmp(rft->name, name))
2023 void __init thread_throttle_mode_init(void)
2027 rft = rdtgroup_get_rftype_by_name("thread_throttle_mode");
2031 rft->fflags = RFTYPE_CTRL_INFO | RFTYPE_RES_MB;
2034 void __init mbm_config_rftype_init(const char *config)
2038 rft = rdtgroup_get_rftype_by_name(config);
2040 rft->fflags = RFTYPE_MON_INFO | RFTYPE_RES_CACHE;
2044 * rdtgroup_kn_mode_restrict - Restrict user access to named resctrl file
2045 * @r: The resource group with which the file is associated.
2046 * @name: Name of the file
2048 * The permissions of named resctrl file, directory, or link are modified
2049 * to not allow read, write, or execute by any user.
2051 * WARNING: This function is intended to communicate to the user that the
2052 * resctrl file has been locked down - that it is not relevant to the
2053 * particular state the system finds itself in. It should not be relied
2054 * on to protect from user access because after the file's permissions
2055 * are restricted the user can still change the permissions using chmod
2056 * from the command line.
2058 * Return: 0 on success, <0 on failure.
2060 int rdtgroup_kn_mode_restrict(struct rdtgroup *r, const char *name)
2062 struct iattr iattr = {.ia_valid = ATTR_MODE,};
2063 struct kernfs_node *kn;
2066 kn = kernfs_find_and_get_ns(r->kn, name, NULL);
2070 switch (kernfs_type(kn)) {
2072 iattr.ia_mode = S_IFDIR;
2075 iattr.ia_mode = S_IFREG;
2078 iattr.ia_mode = S_IFLNK;
2082 ret = kernfs_setattr(kn, &iattr);
2088 * rdtgroup_kn_mode_restore - Restore user access to named resctrl file
2089 * @r: The resource group with which the file is associated.
2090 * @name: Name of the file
2091 * @mask: Mask of permissions that should be restored
2093 * Restore the permissions of the named file. If @name is a directory the
2094 * permissions of its parent will be used.
2096 * Return: 0 on success, <0 on failure.
2098 int rdtgroup_kn_mode_restore(struct rdtgroup *r, const char *name,
2101 struct iattr iattr = {.ia_valid = ATTR_MODE,};
2102 struct kernfs_node *kn, *parent;
2103 struct rftype *rfts, *rft;
2106 rfts = res_common_files;
2107 len = ARRAY_SIZE(res_common_files);
2109 for (rft = rfts; rft < rfts + len; rft++) {
2110 if (!strcmp(rft->name, name))
2111 iattr.ia_mode = rft->mode & mask;
2114 kn = kernfs_find_and_get_ns(r->kn, name, NULL);
2118 switch (kernfs_type(kn)) {
2120 parent = kernfs_get_parent(kn);
2122 iattr.ia_mode |= parent->mode;
2125 iattr.ia_mode |= S_IFDIR;
2128 iattr.ia_mode |= S_IFREG;
2131 iattr.ia_mode |= S_IFLNK;
2135 ret = kernfs_setattr(kn, &iattr);
2140 static int rdtgroup_mkdir_info_resdir(void *priv, char *name,
2141 unsigned long fflags)
2143 struct kernfs_node *kn_subdir;
2146 kn_subdir = kernfs_create_dir(kn_info, name,
2147 kn_info->mode, priv);
2148 if (IS_ERR(kn_subdir))
2149 return PTR_ERR(kn_subdir);
2151 ret = rdtgroup_kn_set_ugid(kn_subdir);
2155 ret = rdtgroup_add_files(kn_subdir, fflags);
2157 kernfs_activate(kn_subdir);
2162 static int rdtgroup_create_info_dir(struct kernfs_node *parent_kn)
2164 struct resctrl_schema *s;
2165 struct rdt_resource *r;
2166 unsigned long fflags;
2170 /* create the directory */
2171 kn_info = kernfs_create_dir(parent_kn, "info", parent_kn->mode, NULL);
2172 if (IS_ERR(kn_info))
2173 return PTR_ERR(kn_info);
2175 ret = rdtgroup_add_files(kn_info, RFTYPE_TOP_INFO);
2179 /* loop over enabled controls, these are all alloc_capable */
2180 list_for_each_entry(s, &resctrl_schema_all, list) {
2182 fflags = r->fflags | RFTYPE_CTRL_INFO;
2183 ret = rdtgroup_mkdir_info_resdir(s, s->name, fflags);
2188 for_each_mon_capable_rdt_resource(r) {
2189 fflags = r->fflags | RFTYPE_MON_INFO;
2190 sprintf(name, "%s_MON", r->name);
2191 ret = rdtgroup_mkdir_info_resdir(r, name, fflags);
2196 ret = rdtgroup_kn_set_ugid(kn_info);
2200 kernfs_activate(kn_info);
2205 kernfs_remove(kn_info);
2210 mongroup_create_dir(struct kernfs_node *parent_kn, struct rdtgroup *prgrp,
2211 char *name, struct kernfs_node **dest_kn)
2213 struct kernfs_node *kn;
2216 /* create the directory */
2217 kn = kernfs_create_dir(parent_kn, name, parent_kn->mode, prgrp);
2224 ret = rdtgroup_kn_set_ugid(kn);
2228 kernfs_activate(kn);
2237 static void l3_qos_cfg_update(void *arg)
2241 wrmsrl(MSR_IA32_L3_QOS_CFG, *enable ? L3_QOS_CDP_ENABLE : 0ULL);
2244 static void l2_qos_cfg_update(void *arg)
2248 wrmsrl(MSR_IA32_L2_QOS_CFG, *enable ? L2_QOS_CDP_ENABLE : 0ULL);
2251 static inline bool is_mba_linear(void)
2253 return rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl.membw.delay_linear;
2256 static int set_cache_qos_cfg(int level, bool enable)
2258 void (*update)(void *arg);
2259 struct rdt_ctrl_domain *d;
2260 struct rdt_resource *r_l;
2261 cpumask_var_t cpu_mask;
2264 /* Walking r->domains, ensure it can't race with cpuhp */
2265 lockdep_assert_cpus_held();
2267 if (level == RDT_RESOURCE_L3)
2268 update = l3_qos_cfg_update;
2269 else if (level == RDT_RESOURCE_L2)
2270 update = l2_qos_cfg_update;
2274 if (!zalloc_cpumask_var(&cpu_mask, GFP_KERNEL))
2277 r_l = &rdt_resources_all[level].r_resctrl;
2278 list_for_each_entry(d, &r_l->ctrl_domains, hdr.list) {
2279 if (r_l->cache.arch_has_per_cpu_cfg)
2280 /* Pick all the CPUs in the domain instance */
2281 for_each_cpu(cpu, &d->hdr.cpu_mask)
2282 cpumask_set_cpu(cpu, cpu_mask);
2284 /* Pick one CPU from each domain instance to update MSR */
2285 cpumask_set_cpu(cpumask_any(&d->hdr.cpu_mask), cpu_mask);
2288 /* Update QOS_CFG MSR on all the CPUs in cpu_mask */
2289 on_each_cpu_mask(cpu_mask, update, &enable, 1);
2291 free_cpumask_var(cpu_mask);
2296 /* Restore the qos cfg state when a domain comes online */
2297 void rdt_domain_reconfigure_cdp(struct rdt_resource *r)
2299 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
2301 if (!r->cdp_capable)
2304 if (r->rid == RDT_RESOURCE_L2)
2305 l2_qos_cfg_update(&hw_res->cdp_enabled);
2307 if (r->rid == RDT_RESOURCE_L3)
2308 l3_qos_cfg_update(&hw_res->cdp_enabled);
2311 static int mba_sc_domain_allocate(struct rdt_resource *r, struct rdt_ctrl_domain *d)
2313 u32 num_closid = resctrl_arch_get_num_closid(r);
2314 int cpu = cpumask_any(&d->hdr.cpu_mask);
2317 d->mbps_val = kcalloc_node(num_closid, sizeof(*d->mbps_val),
2318 GFP_KERNEL, cpu_to_node(cpu));
2322 for (i = 0; i < num_closid; i++)
2323 d->mbps_val[i] = MBA_MAX_MBPS;
2328 static void mba_sc_domain_destroy(struct rdt_resource *r,
2329 struct rdt_ctrl_domain *d)
2336 * MBA software controller is supported only if
2337 * MBM is supported and MBA is in linear scale,
2338 * and the MBM monitor scope is the same as MBA
2341 static bool supports_mba_mbps(void)
2343 struct rdt_resource *rmbm = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
2344 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl;
2346 return (is_mbm_local_enabled() &&
2347 r->alloc_capable && is_mba_linear() &&
2348 r->ctrl_scope == rmbm->mon_scope);
2352 * Enable or disable the MBA software controller
2353 * which helps user specify bandwidth in MBps.
2355 static int set_mba_sc(bool mba_sc)
2357 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl;
2358 u32 num_closid = resctrl_arch_get_num_closid(r);
2359 struct rdt_ctrl_domain *d;
2362 if (!supports_mba_mbps() || mba_sc == is_mba_sc(r))
2365 r->membw.mba_sc = mba_sc;
2367 list_for_each_entry(d, &r->ctrl_domains, hdr.list) {
2368 for (i = 0; i < num_closid; i++)
2369 d->mbps_val[i] = MBA_MAX_MBPS;
2375 static int cdp_enable(int level)
2377 struct rdt_resource *r_l = &rdt_resources_all[level].r_resctrl;
2380 if (!r_l->alloc_capable)
2383 ret = set_cache_qos_cfg(level, true);
2385 rdt_resources_all[level].cdp_enabled = true;
2390 static void cdp_disable(int level)
2392 struct rdt_hw_resource *r_hw = &rdt_resources_all[level];
2394 if (r_hw->cdp_enabled) {
2395 set_cache_qos_cfg(level, false);
2396 r_hw->cdp_enabled = false;
2400 int resctrl_arch_set_cdp_enabled(enum resctrl_res_level l, bool enable)
2402 struct rdt_hw_resource *hw_res = &rdt_resources_all[l];
2404 if (!hw_res->r_resctrl.cdp_capable)
2408 return cdp_enable(l);
2416 * We don't allow rdtgroup directories to be created anywhere
2417 * except the root directory. Thus when looking for the rdtgroup
2418 * structure for a kernfs node we are either looking at a directory,
2419 * in which case the rdtgroup structure is pointed at by the "priv"
2420 * field, otherwise we have a file, and need only look to the parent
2421 * to find the rdtgroup.
2423 static struct rdtgroup *kernfs_to_rdtgroup(struct kernfs_node *kn)
2425 if (kernfs_type(kn) == KERNFS_DIR) {
2427 * All the resource directories use "kn->priv"
2428 * to point to the "struct rdtgroup" for the
2429 * resource. "info" and its subdirectories don't
2430 * have rdtgroup structures, so return NULL here.
2432 if (kn == kn_info || kn->parent == kn_info)
2437 return kn->parent->priv;
2441 static void rdtgroup_kn_get(struct rdtgroup *rdtgrp, struct kernfs_node *kn)
2443 atomic_inc(&rdtgrp->waitcount);
2444 kernfs_break_active_protection(kn);
2447 static void rdtgroup_kn_put(struct rdtgroup *rdtgrp, struct kernfs_node *kn)
2449 if (atomic_dec_and_test(&rdtgrp->waitcount) &&
2450 (rdtgrp->flags & RDT_DELETED)) {
2451 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP ||
2452 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED)
2453 rdtgroup_pseudo_lock_remove(rdtgrp);
2454 kernfs_unbreak_active_protection(kn);
2455 rdtgroup_remove(rdtgrp);
2457 kernfs_unbreak_active_protection(kn);
2461 struct rdtgroup *rdtgroup_kn_lock_live(struct kernfs_node *kn)
2463 struct rdtgroup *rdtgrp = kernfs_to_rdtgroup(kn);
2468 rdtgroup_kn_get(rdtgrp, kn);
2471 mutex_lock(&rdtgroup_mutex);
2473 /* Was this group deleted while we waited? */
2474 if (rdtgrp->flags & RDT_DELETED)
2480 void rdtgroup_kn_unlock(struct kernfs_node *kn)
2482 struct rdtgroup *rdtgrp = kernfs_to_rdtgroup(kn);
2487 mutex_unlock(&rdtgroup_mutex);
2490 rdtgroup_kn_put(rdtgrp, kn);
2493 static int mkdir_mondata_all(struct kernfs_node *parent_kn,
2494 struct rdtgroup *prgrp,
2495 struct kernfs_node **mon_data_kn);
2497 static void rdt_disable_ctx(void)
2499 resctrl_arch_set_cdp_enabled(RDT_RESOURCE_L3, false);
2500 resctrl_arch_set_cdp_enabled(RDT_RESOURCE_L2, false);
2503 resctrl_debug = false;
2506 static int rdt_enable_ctx(struct rdt_fs_context *ctx)
2510 if (ctx->enable_cdpl2) {
2511 ret = resctrl_arch_set_cdp_enabled(RDT_RESOURCE_L2, true);
2516 if (ctx->enable_cdpl3) {
2517 ret = resctrl_arch_set_cdp_enabled(RDT_RESOURCE_L3, true);
2522 if (ctx->enable_mba_mbps) {
2523 ret = set_mba_sc(true);
2528 if (ctx->enable_debug)
2529 resctrl_debug = true;
2534 resctrl_arch_set_cdp_enabled(RDT_RESOURCE_L3, false);
2536 resctrl_arch_set_cdp_enabled(RDT_RESOURCE_L2, false);
2541 static int schemata_list_add(struct rdt_resource *r, enum resctrl_conf_type type)
2543 struct resctrl_schema *s;
2544 const char *suffix = "";
2547 s = kzalloc(sizeof(*s), GFP_KERNEL);
2552 s->num_closid = resctrl_arch_get_num_closid(r);
2553 if (resctrl_arch_get_cdp_enabled(r->rid))
2556 s->conf_type = type;
2569 ret = snprintf(s->name, sizeof(s->name), "%s%s", r->name, suffix);
2570 if (ret >= sizeof(s->name)) {
2575 cl = strlen(s->name);
2578 * If CDP is supported by this resource, but not enabled,
2579 * include the suffix. This ensures the tabular format of the
2580 * schemata file does not change between mounts of the filesystem.
2582 if (r->cdp_capable && !resctrl_arch_get_cdp_enabled(r->rid))
2585 if (cl > max_name_width)
2586 max_name_width = cl;
2588 INIT_LIST_HEAD(&s->list);
2589 list_add(&s->list, &resctrl_schema_all);
2594 static int schemata_list_create(void)
2596 struct rdt_resource *r;
2599 for_each_alloc_capable_rdt_resource(r) {
2600 if (resctrl_arch_get_cdp_enabled(r->rid)) {
2601 ret = schemata_list_add(r, CDP_CODE);
2605 ret = schemata_list_add(r, CDP_DATA);
2607 ret = schemata_list_add(r, CDP_NONE);
2617 static void schemata_list_destroy(void)
2619 struct resctrl_schema *s, *tmp;
2621 list_for_each_entry_safe(s, tmp, &resctrl_schema_all, list) {
2627 static int rdt_get_tree(struct fs_context *fc)
2629 struct rdt_fs_context *ctx = rdt_fc2context(fc);
2630 unsigned long flags = RFTYPE_CTRL_BASE;
2631 struct rdt_mon_domain *dom;
2632 struct rdt_resource *r;
2636 mutex_lock(&rdtgroup_mutex);
2638 * resctrl file system can only be mounted once.
2640 if (resctrl_mounted) {
2645 ret = rdtgroup_setup_root(ctx);
2649 ret = rdt_enable_ctx(ctx);
2653 ret = schemata_list_create();
2655 schemata_list_destroy();
2661 if (resctrl_arch_mon_capable())
2662 flags |= RFTYPE_MON;
2664 ret = rdtgroup_add_files(rdtgroup_default.kn, flags);
2666 goto out_schemata_free;
2668 kernfs_activate(rdtgroup_default.kn);
2670 ret = rdtgroup_create_info_dir(rdtgroup_default.kn);
2672 goto out_schemata_free;
2674 if (resctrl_arch_mon_capable()) {
2675 ret = mongroup_create_dir(rdtgroup_default.kn,
2676 &rdtgroup_default, "mon_groups",
2681 ret = mkdir_mondata_all(rdtgroup_default.kn,
2682 &rdtgroup_default, &kn_mondata);
2685 rdtgroup_default.mon.mon_data_kn = kn_mondata;
2688 ret = rdt_pseudo_lock_init();
2692 ret = kernfs_get_tree(fc);
2696 if (resctrl_arch_alloc_capable())
2697 resctrl_arch_enable_alloc();
2698 if (resctrl_arch_mon_capable())
2699 resctrl_arch_enable_mon();
2701 if (resctrl_arch_alloc_capable() || resctrl_arch_mon_capable())
2702 resctrl_mounted = true;
2704 if (is_mbm_enabled()) {
2705 r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
2706 list_for_each_entry(dom, &r->mon_domains, hdr.list)
2707 mbm_setup_overflow_handler(dom, MBM_OVERFLOW_INTERVAL,
2708 RESCTRL_PICK_ANY_CPU);
2714 rdt_pseudo_lock_release();
2716 if (resctrl_arch_mon_capable())
2717 kernfs_remove(kn_mondata);
2719 if (resctrl_arch_mon_capable())
2720 kernfs_remove(kn_mongrp);
2722 kernfs_remove(kn_info);
2724 schemata_list_destroy();
2728 rdtgroup_destroy_root();
2730 rdt_last_cmd_clear();
2731 mutex_unlock(&rdtgroup_mutex);
2744 static const struct fs_parameter_spec rdt_fs_parameters[] = {
2745 fsparam_flag("cdp", Opt_cdp),
2746 fsparam_flag("cdpl2", Opt_cdpl2),
2747 fsparam_flag("mba_MBps", Opt_mba_mbps),
2748 fsparam_flag("debug", Opt_debug),
2752 static int rdt_parse_param(struct fs_context *fc, struct fs_parameter *param)
2754 struct rdt_fs_context *ctx = rdt_fc2context(fc);
2755 struct fs_parse_result result;
2759 opt = fs_parse(fc, rdt_fs_parameters, param, &result);
2765 ctx->enable_cdpl3 = true;
2768 ctx->enable_cdpl2 = true;
2771 msg = "mba_MBps requires local MBM and linear scale MBA at L3 scope";
2772 if (!supports_mba_mbps())
2773 return invalfc(fc, msg);
2774 ctx->enable_mba_mbps = true;
2777 ctx->enable_debug = true;
2784 static void rdt_fs_context_free(struct fs_context *fc)
2786 struct rdt_fs_context *ctx = rdt_fc2context(fc);
2788 kernfs_free_fs_context(fc);
2792 static const struct fs_context_operations rdt_fs_context_ops = {
2793 .free = rdt_fs_context_free,
2794 .parse_param = rdt_parse_param,
2795 .get_tree = rdt_get_tree,
2798 static int rdt_init_fs_context(struct fs_context *fc)
2800 struct rdt_fs_context *ctx;
2802 ctx = kzalloc(sizeof(struct rdt_fs_context), GFP_KERNEL);
2806 ctx->kfc.magic = RDTGROUP_SUPER_MAGIC;
2807 fc->fs_private = &ctx->kfc;
2808 fc->ops = &rdt_fs_context_ops;
2809 put_user_ns(fc->user_ns);
2810 fc->user_ns = get_user_ns(&init_user_ns);
2815 static int reset_all_ctrls(struct rdt_resource *r)
2817 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
2818 struct rdt_hw_ctrl_domain *hw_dom;
2819 struct msr_param msr_param;
2820 struct rdt_ctrl_domain *d;
2823 /* Walking r->domains, ensure it can't race with cpuhp */
2824 lockdep_assert_cpus_held();
2828 msr_param.high = hw_res->num_closid;
2831 * Disable resource control for this resource by setting all
2832 * CBMs in all ctrl_domains to the maximum mask value. Pick one CPU
2833 * from each domain to update the MSRs below.
2835 list_for_each_entry(d, &r->ctrl_domains, hdr.list) {
2836 hw_dom = resctrl_to_arch_ctrl_dom(d);
2838 for (i = 0; i < hw_res->num_closid; i++)
2839 hw_dom->ctrl_val[i] = r->default_ctrl;
2841 smp_call_function_any(&d->hdr.cpu_mask, rdt_ctrl_update, &msr_param, 1);
2848 * Move tasks from one to the other group. If @from is NULL, then all tasks
2849 * in the systems are moved unconditionally (used for teardown).
2851 * If @mask is not NULL the cpus on which moved tasks are running are set
2852 * in that mask so the update smp function call is restricted to affected
2855 static void rdt_move_group_tasks(struct rdtgroup *from, struct rdtgroup *to,
2856 struct cpumask *mask)
2858 struct task_struct *p, *t;
2860 read_lock(&tasklist_lock);
2861 for_each_process_thread(p, t) {
2862 if (!from || is_closid_match(t, from) ||
2863 is_rmid_match(t, from)) {
2864 resctrl_arch_set_closid_rmid(t, to->closid,
2868 * Order the closid/rmid stores above before the loads
2869 * in task_curr(). This pairs with the full barrier
2870 * between the rq->curr update and resctrl_sched_in()
2871 * during context switch.
2876 * If the task is on a CPU, set the CPU in the mask.
2877 * The detection is inaccurate as tasks might move or
2878 * schedule before the smp function call takes place.
2879 * In such a case the function call is pointless, but
2880 * there is no other side effect.
2882 if (IS_ENABLED(CONFIG_SMP) && mask && task_curr(t))
2883 cpumask_set_cpu(task_cpu(t), mask);
2886 read_unlock(&tasklist_lock);
2889 static void free_all_child_rdtgrp(struct rdtgroup *rdtgrp)
2891 struct rdtgroup *sentry, *stmp;
2892 struct list_head *head;
2894 head = &rdtgrp->mon.crdtgrp_list;
2895 list_for_each_entry_safe(sentry, stmp, head, mon.crdtgrp_list) {
2896 free_rmid(sentry->closid, sentry->mon.rmid);
2897 list_del(&sentry->mon.crdtgrp_list);
2899 if (atomic_read(&sentry->waitcount) != 0)
2900 sentry->flags = RDT_DELETED;
2902 rdtgroup_remove(sentry);
2907 * Forcibly remove all of subdirectories under root.
2909 static void rmdir_all_sub(void)
2911 struct rdtgroup *rdtgrp, *tmp;
2913 /* Move all tasks to the default resource group */
2914 rdt_move_group_tasks(NULL, &rdtgroup_default, NULL);
2916 list_for_each_entry_safe(rdtgrp, tmp, &rdt_all_groups, rdtgroup_list) {
2917 /* Free any child rmids */
2918 free_all_child_rdtgrp(rdtgrp);
2920 /* Remove each rdtgroup other than root */
2921 if (rdtgrp == &rdtgroup_default)
2924 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP ||
2925 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED)
2926 rdtgroup_pseudo_lock_remove(rdtgrp);
2929 * Give any CPUs back to the default group. We cannot copy
2930 * cpu_online_mask because a CPU might have executed the
2931 * offline callback already, but is still marked online.
2933 cpumask_or(&rdtgroup_default.cpu_mask,
2934 &rdtgroup_default.cpu_mask, &rdtgrp->cpu_mask);
2936 free_rmid(rdtgrp->closid, rdtgrp->mon.rmid);
2938 kernfs_remove(rdtgrp->kn);
2939 list_del(&rdtgrp->rdtgroup_list);
2941 if (atomic_read(&rdtgrp->waitcount) != 0)
2942 rdtgrp->flags = RDT_DELETED;
2944 rdtgroup_remove(rdtgrp);
2946 /* Notify online CPUs to update per cpu storage and PQR_ASSOC MSR */
2947 update_closid_rmid(cpu_online_mask, &rdtgroup_default);
2949 kernfs_remove(kn_info);
2950 kernfs_remove(kn_mongrp);
2951 kernfs_remove(kn_mondata);
2954 static void rdt_kill_sb(struct super_block *sb)
2956 struct rdt_resource *r;
2959 mutex_lock(&rdtgroup_mutex);
2963 /*Put everything back to default values. */
2964 for_each_alloc_capable_rdt_resource(r)
2967 rdt_pseudo_lock_release();
2968 rdtgroup_default.mode = RDT_MODE_SHAREABLE;
2969 schemata_list_destroy();
2970 rdtgroup_destroy_root();
2971 if (resctrl_arch_alloc_capable())
2972 resctrl_arch_disable_alloc();
2973 if (resctrl_arch_mon_capable())
2974 resctrl_arch_disable_mon();
2975 resctrl_mounted = false;
2977 mutex_unlock(&rdtgroup_mutex);
2981 static struct file_system_type rdt_fs_type = {
2983 .init_fs_context = rdt_init_fs_context,
2984 .parameters = rdt_fs_parameters,
2985 .kill_sb = rdt_kill_sb,
2988 static int mon_addfile(struct kernfs_node *parent_kn, const char *name,
2991 struct kernfs_node *kn;
2994 kn = __kernfs_create_file(parent_kn, name, 0444,
2995 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID, 0,
2996 &kf_mondata_ops, priv, NULL, NULL);
3000 ret = rdtgroup_kn_set_ugid(kn);
3009 static void mon_rmdir_one_subdir(struct kernfs_node *pkn, char *name, char *subname)
3011 struct kernfs_node *kn;
3013 kn = kernfs_find_and_get(pkn, name);
3018 if (kn->dir.subdirs <= 1)
3021 kernfs_remove_by_name(kn, subname);
3025 * Remove all subdirectories of mon_data of ctrl_mon groups
3026 * and monitor groups for the given domain.
3027 * Remove files and directories containing "sum" of domain data
3028 * when last domain being summed is removed.
3030 static void rmdir_mondata_subdir_allrdtgrp(struct rdt_resource *r,
3031 struct rdt_mon_domain *d)
3033 struct rdtgroup *prgrp, *crgrp;
3038 snc_mode = r->mon_scope == RESCTRL_L3_NODE;
3039 sprintf(name, "mon_%s_%02d", r->name, snc_mode ? d->ci->id : d->hdr.id);
3041 sprintf(subname, "mon_sub_%s_%02d", r->name, d->hdr.id);
3043 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
3044 mon_rmdir_one_subdir(prgrp->mon.mon_data_kn, name, subname);
3046 list_for_each_entry(crgrp, &prgrp->mon.crdtgrp_list, mon.crdtgrp_list)
3047 mon_rmdir_one_subdir(crgrp->mon.mon_data_kn, name, subname);
3051 static int mon_add_all_files(struct kernfs_node *kn, struct rdt_mon_domain *d,
3052 struct rdt_resource *r, struct rdtgroup *prgrp,
3055 struct rmid_read rr = {0};
3056 union mon_data_bits priv;
3057 struct mon_evt *mevt;
3060 if (WARN_ON(list_empty(&r->evt_list)))
3063 priv.u.rid = r->rid;
3064 priv.u.domid = do_sum ? d->ci->id : d->hdr.id;
3065 priv.u.sum = do_sum;
3066 list_for_each_entry(mevt, &r->evt_list, list) {
3067 priv.u.evtid = mevt->evtid;
3068 ret = mon_addfile(kn, mevt->name, priv.priv);
3072 if (!do_sum && is_mbm_event(mevt->evtid))
3073 mon_event_read(&rr, r, d, prgrp, &d->hdr.cpu_mask, mevt->evtid, true);
3079 static int mkdir_mondata_subdir(struct kernfs_node *parent_kn,
3080 struct rdt_mon_domain *d,
3081 struct rdt_resource *r, struct rdtgroup *prgrp)
3083 struct kernfs_node *kn, *ckn;
3088 lockdep_assert_held(&rdtgroup_mutex);
3090 snc_mode = r->mon_scope == RESCTRL_L3_NODE;
3091 sprintf(name, "mon_%s_%02d", r->name, snc_mode ? d->ci->id : d->hdr.id);
3092 kn = kernfs_find_and_get(parent_kn, name);
3095 * rdtgroup_mutex will prevent this directory from being
3096 * removed. No need to keep this hold.
3100 kn = kernfs_create_dir(parent_kn, name, parent_kn->mode, prgrp);
3104 ret = rdtgroup_kn_set_ugid(kn);
3107 ret = mon_add_all_files(kn, d, r, prgrp, snc_mode);
3113 sprintf(name, "mon_sub_%s_%02d", r->name, d->hdr.id);
3114 ckn = kernfs_create_dir(kn, name, parent_kn->mode, prgrp);
3120 ret = rdtgroup_kn_set_ugid(ckn);
3124 ret = mon_add_all_files(ckn, d, r, prgrp, false);
3129 kernfs_activate(kn);
3138 * Add all subdirectories of mon_data for "ctrl_mon" groups
3139 * and "monitor" groups with given domain id.
3141 static void mkdir_mondata_subdir_allrdtgrp(struct rdt_resource *r,
3142 struct rdt_mon_domain *d)
3144 struct kernfs_node *parent_kn;
3145 struct rdtgroup *prgrp, *crgrp;
3146 struct list_head *head;
3148 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
3149 parent_kn = prgrp->mon.mon_data_kn;
3150 mkdir_mondata_subdir(parent_kn, d, r, prgrp);
3152 head = &prgrp->mon.crdtgrp_list;
3153 list_for_each_entry(crgrp, head, mon.crdtgrp_list) {
3154 parent_kn = crgrp->mon.mon_data_kn;
3155 mkdir_mondata_subdir(parent_kn, d, r, crgrp);
3160 static int mkdir_mondata_subdir_alldom(struct kernfs_node *parent_kn,
3161 struct rdt_resource *r,
3162 struct rdtgroup *prgrp)
3164 struct rdt_mon_domain *dom;
3167 /* Walking r->domains, ensure it can't race with cpuhp */
3168 lockdep_assert_cpus_held();
3170 list_for_each_entry(dom, &r->mon_domains, hdr.list) {
3171 ret = mkdir_mondata_subdir(parent_kn, dom, r, prgrp);
3180 * This creates a directory mon_data which contains the monitored data.
3182 * mon_data has one directory for each domain which are named
3183 * in the format mon_<domain_name>_<domain_id>. For ex: A mon_data
3184 * with L3 domain looks as below:
3191 * Each domain directory has one file per event:
3196 static int mkdir_mondata_all(struct kernfs_node *parent_kn,
3197 struct rdtgroup *prgrp,
3198 struct kernfs_node **dest_kn)
3200 struct rdt_resource *r;
3201 struct kernfs_node *kn;
3205 * Create the mon_data directory first.
3207 ret = mongroup_create_dir(parent_kn, prgrp, "mon_data", &kn);
3215 * Create the subdirectories for each domain. Note that all events
3216 * in a domain like L3 are grouped into a resource whose domain is L3
3218 for_each_mon_capable_rdt_resource(r) {
3219 ret = mkdir_mondata_subdir_alldom(kn, r, prgrp);
3232 * cbm_ensure_valid - Enforce validity on provided CBM
3233 * @_val: Candidate CBM
3234 * @r: RDT resource to which the CBM belongs
3236 * The provided CBM represents all cache portions available for use. This
3237 * may be represented by a bitmap that does not consist of contiguous ones
3238 * and thus be an invalid CBM.
3239 * Here the provided CBM is forced to be a valid CBM by only considering
3240 * the first set of contiguous bits as valid and clearing all bits.
3241 * The intention here is to provide a valid default CBM with which a new
3242 * resource group is initialized. The user can follow this with a
3243 * modification to the CBM if the default does not satisfy the
3246 static u32 cbm_ensure_valid(u32 _val, struct rdt_resource *r)
3248 unsigned int cbm_len = r->cache.cbm_len;
3249 unsigned long first_bit, zero_bit;
3250 unsigned long val = _val;
3255 first_bit = find_first_bit(&val, cbm_len);
3256 zero_bit = find_next_zero_bit(&val, cbm_len, first_bit);
3258 /* Clear any remaining bits to ensure contiguous region */
3259 bitmap_clear(&val, zero_bit, cbm_len - zero_bit);
3264 * Initialize cache resources per RDT domain
3266 * Set the RDT domain up to start off with all usable allocations. That is,
3267 * all shareable and unused bits. All-zero CBM is invalid.
3269 static int __init_one_rdt_domain(struct rdt_ctrl_domain *d, struct resctrl_schema *s,
3272 enum resctrl_conf_type peer_type = resctrl_peer_type(s->conf_type);
3273 enum resctrl_conf_type t = s->conf_type;
3274 struct resctrl_staged_config *cfg;
3275 struct rdt_resource *r = s->res;
3276 u32 used_b = 0, unused_b = 0;
3277 unsigned long tmp_cbm;
3278 enum rdtgrp_mode mode;
3279 u32 peer_ctl, ctrl_val;
3282 cfg = &d->staged_config[t];
3283 cfg->have_new_ctrl = false;
3284 cfg->new_ctrl = r->cache.shareable_bits;
3285 used_b = r->cache.shareable_bits;
3286 for (i = 0; i < closids_supported(); i++) {
3287 if (closid_allocated(i) && i != closid) {
3288 mode = rdtgroup_mode_by_closid(i);
3289 if (mode == RDT_MODE_PSEUDO_LOCKSETUP)
3291 * ctrl values for locksetup aren't relevant
3292 * until the schemata is written, and the mode
3293 * becomes RDT_MODE_PSEUDO_LOCKED.
3297 * If CDP is active include peer domain's
3298 * usage to ensure there is no overlap
3299 * with an exclusive group.
3301 if (resctrl_arch_get_cdp_enabled(r->rid))
3302 peer_ctl = resctrl_arch_get_config(r, d, i,
3306 ctrl_val = resctrl_arch_get_config(r, d, i,
3308 used_b |= ctrl_val | peer_ctl;
3309 if (mode == RDT_MODE_SHAREABLE)
3310 cfg->new_ctrl |= ctrl_val | peer_ctl;
3313 if (d->plr && d->plr->cbm > 0)
3314 used_b |= d->plr->cbm;
3315 unused_b = used_b ^ (BIT_MASK(r->cache.cbm_len) - 1);
3316 unused_b &= BIT_MASK(r->cache.cbm_len) - 1;
3317 cfg->new_ctrl |= unused_b;
3319 * Force the initial CBM to be valid, user can
3320 * modify the CBM based on system availability.
3322 cfg->new_ctrl = cbm_ensure_valid(cfg->new_ctrl, r);
3324 * Assign the u32 CBM to an unsigned long to ensure that
3325 * bitmap_weight() does not access out-of-bound memory.
3327 tmp_cbm = cfg->new_ctrl;
3328 if (bitmap_weight(&tmp_cbm, r->cache.cbm_len) < r->cache.min_cbm_bits) {
3329 rdt_last_cmd_printf("No space on %s:%d\n", s->name, d->hdr.id);
3332 cfg->have_new_ctrl = true;
3338 * Initialize cache resources with default values.
3340 * A new RDT group is being created on an allocation capable (CAT)
3341 * supporting system. Set this group up to start off with all usable
3344 * If there are no more shareable bits available on any domain then
3345 * the entire allocation will fail.
3347 static int rdtgroup_init_cat(struct resctrl_schema *s, u32 closid)
3349 struct rdt_ctrl_domain *d;
3352 list_for_each_entry(d, &s->res->ctrl_domains, hdr.list) {
3353 ret = __init_one_rdt_domain(d, s, closid);
3361 /* Initialize MBA resource with default values. */
3362 static void rdtgroup_init_mba(struct rdt_resource *r, u32 closid)
3364 struct resctrl_staged_config *cfg;
3365 struct rdt_ctrl_domain *d;
3367 list_for_each_entry(d, &r->ctrl_domains, hdr.list) {
3369 d->mbps_val[closid] = MBA_MAX_MBPS;
3373 cfg = &d->staged_config[CDP_NONE];
3374 cfg->new_ctrl = r->default_ctrl;
3375 cfg->have_new_ctrl = true;
3379 /* Initialize the RDT group's allocations. */
3380 static int rdtgroup_init_alloc(struct rdtgroup *rdtgrp)
3382 struct resctrl_schema *s;
3383 struct rdt_resource *r;
3386 rdt_staged_configs_clear();
3388 list_for_each_entry(s, &resctrl_schema_all, list) {
3390 if (r->rid == RDT_RESOURCE_MBA ||
3391 r->rid == RDT_RESOURCE_SMBA) {
3392 rdtgroup_init_mba(r, rdtgrp->closid);
3396 ret = rdtgroup_init_cat(s, rdtgrp->closid);
3401 ret = resctrl_arch_update_domains(r, rdtgrp->closid);
3403 rdt_last_cmd_puts("Failed to initialize allocations\n");
3409 rdtgrp->mode = RDT_MODE_SHAREABLE;
3412 rdt_staged_configs_clear();
3416 static int mkdir_rdt_prepare_rmid_alloc(struct rdtgroup *rdtgrp)
3420 if (!resctrl_arch_mon_capable())
3423 ret = alloc_rmid(rdtgrp->closid);
3425 rdt_last_cmd_puts("Out of RMIDs\n");
3428 rdtgrp->mon.rmid = ret;
3430 ret = mkdir_mondata_all(rdtgrp->kn, rdtgrp, &rdtgrp->mon.mon_data_kn);
3432 rdt_last_cmd_puts("kernfs subdir error\n");
3433 free_rmid(rdtgrp->closid, rdtgrp->mon.rmid);
3440 static void mkdir_rdt_prepare_rmid_free(struct rdtgroup *rgrp)
3442 if (resctrl_arch_mon_capable())
3443 free_rmid(rgrp->closid, rgrp->mon.rmid);
3446 static int mkdir_rdt_prepare(struct kernfs_node *parent_kn,
3447 const char *name, umode_t mode,
3448 enum rdt_group_type rtype, struct rdtgroup **r)
3450 struct rdtgroup *prdtgrp, *rdtgrp;
3451 unsigned long files = 0;
3452 struct kernfs_node *kn;
3455 prdtgrp = rdtgroup_kn_lock_live(parent_kn);
3461 if (rtype == RDTMON_GROUP &&
3462 (prdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP ||
3463 prdtgrp->mode == RDT_MODE_PSEUDO_LOCKED)) {
3465 rdt_last_cmd_puts("Pseudo-locking in progress\n");
3469 /* allocate the rdtgroup. */
3470 rdtgrp = kzalloc(sizeof(*rdtgrp), GFP_KERNEL);
3473 rdt_last_cmd_puts("Kernel out of memory\n");
3477 rdtgrp->mon.parent = prdtgrp;
3478 rdtgrp->type = rtype;
3479 INIT_LIST_HEAD(&rdtgrp->mon.crdtgrp_list);
3481 /* kernfs creates the directory for rdtgrp */
3482 kn = kernfs_create_dir(parent_kn, name, mode, rdtgrp);
3485 rdt_last_cmd_puts("kernfs create error\n");
3491 * kernfs_remove() will drop the reference count on "kn" which
3492 * will free it. But we still need it to stick around for the
3493 * rdtgroup_kn_unlock(kn) call. Take one extra reference here,
3494 * which will be dropped by kernfs_put() in rdtgroup_remove().
3498 ret = rdtgroup_kn_set_ugid(kn);
3500 rdt_last_cmd_puts("kernfs perm error\n");
3504 if (rtype == RDTCTRL_GROUP) {
3505 files = RFTYPE_BASE | RFTYPE_CTRL;
3506 if (resctrl_arch_mon_capable())
3507 files |= RFTYPE_MON;
3509 files = RFTYPE_BASE | RFTYPE_MON;
3512 ret = rdtgroup_add_files(kn, files);
3514 rdt_last_cmd_puts("kernfs fill error\n");
3519 * The caller unlocks the parent_kn upon success.
3524 kernfs_put(rdtgrp->kn);
3525 kernfs_remove(rdtgrp->kn);
3529 rdtgroup_kn_unlock(parent_kn);
3533 static void mkdir_rdt_prepare_clean(struct rdtgroup *rgrp)
3535 kernfs_remove(rgrp->kn);
3536 rdtgroup_remove(rgrp);
3540 * Create a monitor group under "mon_groups" directory of a control
3541 * and monitor group(ctrl_mon). This is a resource group
3542 * to monitor a subset of tasks and cpus in its parent ctrl_mon group.
3544 static int rdtgroup_mkdir_mon(struct kernfs_node *parent_kn,
3545 const char *name, umode_t mode)
3547 struct rdtgroup *rdtgrp, *prgrp;
3550 ret = mkdir_rdt_prepare(parent_kn, name, mode, RDTMON_GROUP, &rdtgrp);
3554 prgrp = rdtgrp->mon.parent;
3555 rdtgrp->closid = prgrp->closid;
3557 ret = mkdir_rdt_prepare_rmid_alloc(rdtgrp);
3559 mkdir_rdt_prepare_clean(rdtgrp);
3563 kernfs_activate(rdtgrp->kn);
3566 * Add the rdtgrp to the list of rdtgrps the parent
3567 * ctrl_mon group has to track.
3569 list_add_tail(&rdtgrp->mon.crdtgrp_list, &prgrp->mon.crdtgrp_list);
3572 rdtgroup_kn_unlock(parent_kn);
3577 * These are rdtgroups created under the root directory. Can be used
3578 * to allocate and monitor resources.
3580 static int rdtgroup_mkdir_ctrl_mon(struct kernfs_node *parent_kn,
3581 const char *name, umode_t mode)
3583 struct rdtgroup *rdtgrp;
3584 struct kernfs_node *kn;
3588 ret = mkdir_rdt_prepare(parent_kn, name, mode, RDTCTRL_GROUP, &rdtgrp);
3593 ret = closid_alloc();
3595 rdt_last_cmd_puts("Out of CLOSIDs\n");
3596 goto out_common_fail;
3601 rdtgrp->closid = closid;
3603 ret = mkdir_rdt_prepare_rmid_alloc(rdtgrp);
3605 goto out_closid_free;
3607 kernfs_activate(rdtgrp->kn);
3609 ret = rdtgroup_init_alloc(rdtgrp);
3613 list_add(&rdtgrp->rdtgroup_list, &rdt_all_groups);
3615 if (resctrl_arch_mon_capable()) {
3617 * Create an empty mon_groups directory to hold the subset
3618 * of tasks and cpus to monitor.
3620 ret = mongroup_create_dir(kn, rdtgrp, "mon_groups", NULL);
3622 rdt_last_cmd_puts("kernfs subdir error\n");
3630 list_del(&rdtgrp->rdtgroup_list);
3632 mkdir_rdt_prepare_rmid_free(rdtgrp);
3634 closid_free(closid);
3636 mkdir_rdt_prepare_clean(rdtgrp);
3638 rdtgroup_kn_unlock(parent_kn);
3643 * We allow creating mon groups only with in a directory called "mon_groups"
3644 * which is present in every ctrl_mon group. Check if this is a valid
3645 * "mon_groups" directory.
3647 * 1. The directory should be named "mon_groups".
3648 * 2. The mon group itself should "not" be named "mon_groups".
3649 * This makes sure "mon_groups" directory always has a ctrl_mon group
3652 static bool is_mon_groups(struct kernfs_node *kn, const char *name)
3654 return (!strcmp(kn->name, "mon_groups") &&
3655 strcmp(name, "mon_groups"));
3658 static int rdtgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
3661 /* Do not accept '\n' to avoid unparsable situation. */
3662 if (strchr(name, '\n'))
3666 * If the parent directory is the root directory and RDT
3667 * allocation is supported, add a control and monitoring
3670 if (resctrl_arch_alloc_capable() && parent_kn == rdtgroup_default.kn)
3671 return rdtgroup_mkdir_ctrl_mon(parent_kn, name, mode);
3674 * If RDT monitoring is supported and the parent directory is a valid
3675 * "mon_groups" directory, add a monitoring subdirectory.
3677 if (resctrl_arch_mon_capable() && is_mon_groups(parent_kn, name))
3678 return rdtgroup_mkdir_mon(parent_kn, name, mode);
3683 static int rdtgroup_rmdir_mon(struct rdtgroup *rdtgrp, cpumask_var_t tmpmask)
3685 struct rdtgroup *prdtgrp = rdtgrp->mon.parent;
3688 /* Give any tasks back to the parent group */
3689 rdt_move_group_tasks(rdtgrp, prdtgrp, tmpmask);
3691 /* Update per cpu rmid of the moved CPUs first */
3692 for_each_cpu(cpu, &rdtgrp->cpu_mask)
3693 per_cpu(pqr_state.default_rmid, cpu) = prdtgrp->mon.rmid;
3695 * Update the MSR on moved CPUs and CPUs which have moved
3696 * task running on them.
3698 cpumask_or(tmpmask, tmpmask, &rdtgrp->cpu_mask);
3699 update_closid_rmid(tmpmask, NULL);
3701 rdtgrp->flags = RDT_DELETED;
3702 free_rmid(rdtgrp->closid, rdtgrp->mon.rmid);
3705 * Remove the rdtgrp from the parent ctrl_mon group's list
3707 WARN_ON(list_empty(&prdtgrp->mon.crdtgrp_list));
3708 list_del(&rdtgrp->mon.crdtgrp_list);
3710 kernfs_remove(rdtgrp->kn);
3715 static int rdtgroup_ctrl_remove(struct rdtgroup *rdtgrp)
3717 rdtgrp->flags = RDT_DELETED;
3718 list_del(&rdtgrp->rdtgroup_list);
3720 kernfs_remove(rdtgrp->kn);
3724 static int rdtgroup_rmdir_ctrl(struct rdtgroup *rdtgrp, cpumask_var_t tmpmask)
3728 /* Give any tasks back to the default group */
3729 rdt_move_group_tasks(rdtgrp, &rdtgroup_default, tmpmask);
3731 /* Give any CPUs back to the default group */
3732 cpumask_or(&rdtgroup_default.cpu_mask,
3733 &rdtgroup_default.cpu_mask, &rdtgrp->cpu_mask);
3735 /* Update per cpu closid and rmid of the moved CPUs first */
3736 for_each_cpu(cpu, &rdtgrp->cpu_mask) {
3737 per_cpu(pqr_state.default_closid, cpu) = rdtgroup_default.closid;
3738 per_cpu(pqr_state.default_rmid, cpu) = rdtgroup_default.mon.rmid;
3742 * Update the MSR on moved CPUs and CPUs which have moved
3743 * task running on them.
3745 cpumask_or(tmpmask, tmpmask, &rdtgrp->cpu_mask);
3746 update_closid_rmid(tmpmask, NULL);
3748 free_rmid(rdtgrp->closid, rdtgrp->mon.rmid);
3749 closid_free(rdtgrp->closid);
3751 rdtgroup_ctrl_remove(rdtgrp);
3754 * Free all the child monitor group rmids.
3756 free_all_child_rdtgrp(rdtgrp);
3761 static int rdtgroup_rmdir(struct kernfs_node *kn)
3763 struct kernfs_node *parent_kn = kn->parent;
3764 struct rdtgroup *rdtgrp;
3765 cpumask_var_t tmpmask;
3768 if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
3771 rdtgrp = rdtgroup_kn_lock_live(kn);
3778 * If the rdtgroup is a ctrl_mon group and parent directory
3779 * is the root directory, remove the ctrl_mon group.
3781 * If the rdtgroup is a mon group and parent directory
3782 * is a valid "mon_groups" directory, remove the mon group.
3784 if (rdtgrp->type == RDTCTRL_GROUP && parent_kn == rdtgroup_default.kn &&
3785 rdtgrp != &rdtgroup_default) {
3786 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP ||
3787 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
3788 ret = rdtgroup_ctrl_remove(rdtgrp);
3790 ret = rdtgroup_rmdir_ctrl(rdtgrp, tmpmask);
3792 } else if (rdtgrp->type == RDTMON_GROUP &&
3793 is_mon_groups(parent_kn, kn->name)) {
3794 ret = rdtgroup_rmdir_mon(rdtgrp, tmpmask);
3800 rdtgroup_kn_unlock(kn);
3801 free_cpumask_var(tmpmask);
3806 * mongrp_reparent() - replace parent CTRL_MON group of a MON group
3807 * @rdtgrp: the MON group whose parent should be replaced
3808 * @new_prdtgrp: replacement parent CTRL_MON group for @rdtgrp
3809 * @cpus: cpumask provided by the caller for use during this call
3811 * Replaces the parent CTRL_MON group for a MON group, resulting in all member
3812 * tasks' CLOSID immediately changing to that of the new parent group.
3813 * Monitoring data for the group is unaffected by this operation.
3815 static void mongrp_reparent(struct rdtgroup *rdtgrp,
3816 struct rdtgroup *new_prdtgrp,
3819 struct rdtgroup *prdtgrp = rdtgrp->mon.parent;
3821 WARN_ON(rdtgrp->type != RDTMON_GROUP);
3822 WARN_ON(new_prdtgrp->type != RDTCTRL_GROUP);
3824 /* Nothing to do when simply renaming a MON group. */
3825 if (prdtgrp == new_prdtgrp)
3828 WARN_ON(list_empty(&prdtgrp->mon.crdtgrp_list));
3829 list_move_tail(&rdtgrp->mon.crdtgrp_list,
3830 &new_prdtgrp->mon.crdtgrp_list);
3832 rdtgrp->mon.parent = new_prdtgrp;
3833 rdtgrp->closid = new_prdtgrp->closid;
3835 /* Propagate updated closid to all tasks in this group. */
3836 rdt_move_group_tasks(rdtgrp, rdtgrp, cpus);
3838 update_closid_rmid(cpus, NULL);
3841 static int rdtgroup_rename(struct kernfs_node *kn,
3842 struct kernfs_node *new_parent, const char *new_name)
3844 struct rdtgroup *new_prdtgrp;
3845 struct rdtgroup *rdtgrp;
3846 cpumask_var_t tmpmask;
3849 rdtgrp = kernfs_to_rdtgroup(kn);
3850 new_prdtgrp = kernfs_to_rdtgroup(new_parent);
3851 if (!rdtgrp || !new_prdtgrp)
3854 /* Release both kernfs active_refs before obtaining rdtgroup mutex. */
3855 rdtgroup_kn_get(rdtgrp, kn);
3856 rdtgroup_kn_get(new_prdtgrp, new_parent);
3858 mutex_lock(&rdtgroup_mutex);
3860 rdt_last_cmd_clear();
3863 * Don't allow kernfs_to_rdtgroup() to return a parent rdtgroup if
3864 * either kernfs_node is a file.
3866 if (kernfs_type(kn) != KERNFS_DIR ||
3867 kernfs_type(new_parent) != KERNFS_DIR) {
3868 rdt_last_cmd_puts("Source and destination must be directories");
3873 if ((rdtgrp->flags & RDT_DELETED) || (new_prdtgrp->flags & RDT_DELETED)) {
3878 if (rdtgrp->type != RDTMON_GROUP || !kn->parent ||
3879 !is_mon_groups(kn->parent, kn->name)) {
3880 rdt_last_cmd_puts("Source must be a MON group\n");
3885 if (!is_mon_groups(new_parent, new_name)) {
3886 rdt_last_cmd_puts("Destination must be a mon_groups subdirectory\n");
3892 * If the MON group is monitoring CPUs, the CPUs must be assigned to the
3893 * current parent CTRL_MON group and therefore cannot be assigned to
3894 * the new parent, making the move illegal.
3896 if (!cpumask_empty(&rdtgrp->cpu_mask) &&
3897 rdtgrp->mon.parent != new_prdtgrp) {
3898 rdt_last_cmd_puts("Cannot move a MON group that monitors CPUs\n");
3904 * Allocate the cpumask for use in mongrp_reparent() to avoid the
3905 * possibility of failing to allocate it after kernfs_rename() has
3908 if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL)) {
3914 * Perform all input validation and allocations needed to ensure
3915 * mongrp_reparent() will succeed before calling kernfs_rename(),
3916 * otherwise it would be necessary to revert this call if
3917 * mongrp_reparent() failed.
3919 ret = kernfs_rename(kn, new_parent, new_name);
3921 mongrp_reparent(rdtgrp, new_prdtgrp, tmpmask);
3923 free_cpumask_var(tmpmask);
3926 mutex_unlock(&rdtgroup_mutex);
3927 rdtgroup_kn_put(rdtgrp, kn);
3928 rdtgroup_kn_put(new_prdtgrp, new_parent);
3932 static int rdtgroup_show_options(struct seq_file *seq, struct kernfs_root *kf)
3934 if (resctrl_arch_get_cdp_enabled(RDT_RESOURCE_L3))
3935 seq_puts(seq, ",cdp");
3937 if (resctrl_arch_get_cdp_enabled(RDT_RESOURCE_L2))
3938 seq_puts(seq, ",cdpl2");
3940 if (is_mba_sc(&rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl))
3941 seq_puts(seq, ",mba_MBps");
3944 seq_puts(seq, ",debug");
3949 static struct kernfs_syscall_ops rdtgroup_kf_syscall_ops = {
3950 .mkdir = rdtgroup_mkdir,
3951 .rmdir = rdtgroup_rmdir,
3952 .rename = rdtgroup_rename,
3953 .show_options = rdtgroup_show_options,
3956 static int rdtgroup_setup_root(struct rdt_fs_context *ctx)
3958 rdt_root = kernfs_create_root(&rdtgroup_kf_syscall_ops,
3959 KERNFS_ROOT_CREATE_DEACTIVATED |
3960 KERNFS_ROOT_EXTRA_OPEN_PERM_CHECK,
3962 if (IS_ERR(rdt_root))
3963 return PTR_ERR(rdt_root);
3965 ctx->kfc.root = rdt_root;
3966 rdtgroup_default.kn = kernfs_root_to_node(rdt_root);
3971 static void rdtgroup_destroy_root(void)
3973 kernfs_destroy_root(rdt_root);
3974 rdtgroup_default.kn = NULL;
3977 static void __init rdtgroup_setup_default(void)
3979 mutex_lock(&rdtgroup_mutex);
3981 rdtgroup_default.closid = RESCTRL_RESERVED_CLOSID;
3982 rdtgroup_default.mon.rmid = RESCTRL_RESERVED_RMID;
3983 rdtgroup_default.type = RDTCTRL_GROUP;
3984 INIT_LIST_HEAD(&rdtgroup_default.mon.crdtgrp_list);
3986 list_add(&rdtgroup_default.rdtgroup_list, &rdt_all_groups);
3988 mutex_unlock(&rdtgroup_mutex);
3991 static void domain_destroy_mon_state(struct rdt_mon_domain *d)
3993 bitmap_free(d->rmid_busy_llc);
3994 kfree(d->mbm_total);
3995 kfree(d->mbm_local);
3998 void resctrl_offline_ctrl_domain(struct rdt_resource *r, struct rdt_ctrl_domain *d)
4000 mutex_lock(&rdtgroup_mutex);
4002 if (supports_mba_mbps() && r->rid == RDT_RESOURCE_MBA)
4003 mba_sc_domain_destroy(r, d);
4005 mutex_unlock(&rdtgroup_mutex);
4008 void resctrl_offline_mon_domain(struct rdt_resource *r, struct rdt_mon_domain *d)
4010 mutex_lock(&rdtgroup_mutex);
4013 * If resctrl is mounted, remove all the
4014 * per domain monitor data directories.
4016 if (resctrl_mounted && resctrl_arch_mon_capable())
4017 rmdir_mondata_subdir_allrdtgrp(r, d);
4019 if (is_mbm_enabled())
4020 cancel_delayed_work(&d->mbm_over);
4021 if (is_llc_occupancy_enabled() && has_busy_rmid(d)) {
4023 * When a package is going down, forcefully
4024 * decrement rmid->ebusy. There is no way to know
4025 * that the L3 was flushed and hence may lead to
4026 * incorrect counts in rare scenarios, but leaving
4027 * the RMID as busy creates RMID leaks if the
4028 * package never comes back.
4030 __check_limbo(d, true);
4031 cancel_delayed_work(&d->cqm_limbo);
4034 domain_destroy_mon_state(d);
4036 mutex_unlock(&rdtgroup_mutex);
4039 static int domain_setup_mon_state(struct rdt_resource *r, struct rdt_mon_domain *d)
4041 u32 idx_limit = resctrl_arch_system_num_rmid_idx();
4044 if (is_llc_occupancy_enabled()) {
4045 d->rmid_busy_llc = bitmap_zalloc(idx_limit, GFP_KERNEL);
4046 if (!d->rmid_busy_llc)
4049 if (is_mbm_total_enabled()) {
4050 tsize = sizeof(*d->mbm_total);
4051 d->mbm_total = kcalloc(idx_limit, tsize, GFP_KERNEL);
4052 if (!d->mbm_total) {
4053 bitmap_free(d->rmid_busy_llc);
4057 if (is_mbm_local_enabled()) {
4058 tsize = sizeof(*d->mbm_local);
4059 d->mbm_local = kcalloc(idx_limit, tsize, GFP_KERNEL);
4060 if (!d->mbm_local) {
4061 bitmap_free(d->rmid_busy_llc);
4062 kfree(d->mbm_total);
4070 int resctrl_online_ctrl_domain(struct rdt_resource *r, struct rdt_ctrl_domain *d)
4074 mutex_lock(&rdtgroup_mutex);
4076 if (supports_mba_mbps() && r->rid == RDT_RESOURCE_MBA) {
4077 /* RDT_RESOURCE_MBA is never mon_capable */
4078 err = mba_sc_domain_allocate(r, d);
4081 mutex_unlock(&rdtgroup_mutex);
4086 int resctrl_online_mon_domain(struct rdt_resource *r, struct rdt_mon_domain *d)
4090 mutex_lock(&rdtgroup_mutex);
4092 err = domain_setup_mon_state(r, d);
4096 if (is_mbm_enabled()) {
4097 INIT_DELAYED_WORK(&d->mbm_over, mbm_handle_overflow);
4098 mbm_setup_overflow_handler(d, MBM_OVERFLOW_INTERVAL,
4099 RESCTRL_PICK_ANY_CPU);
4102 if (is_llc_occupancy_enabled())
4103 INIT_DELAYED_WORK(&d->cqm_limbo, cqm_handle_limbo);
4106 * If the filesystem is not mounted then only the default resource group
4107 * exists. Creation of its directories is deferred until mount time
4108 * by rdt_get_tree() calling mkdir_mondata_all().
4109 * If resctrl is mounted, add per domain monitor data directories.
4111 if (resctrl_mounted && resctrl_arch_mon_capable())
4112 mkdir_mondata_subdir_allrdtgrp(r, d);
4115 mutex_unlock(&rdtgroup_mutex);
4120 void resctrl_online_cpu(unsigned int cpu)
4122 mutex_lock(&rdtgroup_mutex);
4123 /* The CPU is set in default rdtgroup after online. */
4124 cpumask_set_cpu(cpu, &rdtgroup_default.cpu_mask);
4125 mutex_unlock(&rdtgroup_mutex);
4128 static void clear_childcpus(struct rdtgroup *r, unsigned int cpu)
4130 struct rdtgroup *cr;
4132 list_for_each_entry(cr, &r->mon.crdtgrp_list, mon.crdtgrp_list) {
4133 if (cpumask_test_and_clear_cpu(cpu, &cr->cpu_mask))
4138 void resctrl_offline_cpu(unsigned int cpu)
4140 struct rdt_resource *l3 = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
4141 struct rdt_mon_domain *d;
4142 struct rdtgroup *rdtgrp;
4144 mutex_lock(&rdtgroup_mutex);
4145 list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
4146 if (cpumask_test_and_clear_cpu(cpu, &rdtgrp->cpu_mask)) {
4147 clear_childcpus(rdtgrp, cpu);
4152 if (!l3->mon_capable)
4155 d = get_mon_domain_from_cpu(cpu, l3);
4157 if (is_mbm_enabled() && cpu == d->mbm_work_cpu) {
4158 cancel_delayed_work(&d->mbm_over);
4159 mbm_setup_overflow_handler(d, 0, cpu);
4161 if (is_llc_occupancy_enabled() && cpu == d->cqm_work_cpu &&
4163 cancel_delayed_work(&d->cqm_limbo);
4164 cqm_setup_limbo_handler(d, 0, cpu);
4169 mutex_unlock(&rdtgroup_mutex);
4173 * rdtgroup_init - rdtgroup initialization
4175 * Setup resctrl file system including set up root, create mount point,
4176 * register rdtgroup filesystem, and initialize files under root directory.
4178 * Return: 0 on success or -errno
4180 int __init rdtgroup_init(void)
4184 seq_buf_init(&last_cmd_status, last_cmd_status_buf,
4185 sizeof(last_cmd_status_buf));
4187 rdtgroup_setup_default();
4189 ret = sysfs_create_mount_point(fs_kobj, "resctrl");
4193 ret = register_filesystem(&rdt_fs_type);
4195 goto cleanup_mountpoint;
4198 * Adding the resctrl debugfs directory here may not be ideal since
4199 * it would let the resctrl debugfs directory appear on the debugfs
4200 * filesystem before the resctrl filesystem is mounted.
4201 * It may also be ok since that would enable debugging of RDT before
4202 * resctrl is mounted.
4203 * The reason why the debugfs directory is created here and not in
4204 * rdt_get_tree() is because rdt_get_tree() takes rdtgroup_mutex and
4205 * during the debugfs directory creation also &sb->s_type->i_mutex_key
4206 * (the lockdep class of inode->i_rwsem). Other filesystem
4207 * interactions (eg. SyS_getdents) have the lock ordering:
4208 * &sb->s_type->i_mutex_key --> &mm->mmap_lock
4209 * During mmap(), called with &mm->mmap_lock, the rdtgroup_mutex
4210 * is taken, thus creating dependency:
4211 * &mm->mmap_lock --> rdtgroup_mutex for the latter that can cause
4212 * issues considering the other two lock dependencies.
4213 * By creating the debugfs directory here we avoid a dependency
4214 * that may cause deadlock (even though file operations cannot
4215 * occur until the filesystem is mounted, but I do not know how to
4216 * tell lockdep that).
4218 debugfs_resctrl = debugfs_create_dir("resctrl", NULL);
4223 sysfs_remove_mount_point(fs_kobj, "resctrl");
4228 void __exit rdtgroup_exit(void)
4230 debugfs_remove_recursive(debugfs_resctrl);
4231 unregister_filesystem(&rdt_fs_type);
4232 sysfs_remove_mount_point(fs_kobj, "resctrl");