Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * User interface for Resource Alloction in Resource Director Technology(RDT)
4 *
5 * Copyright (C) 2016 Intel Corporation
6 *
7 * Author: Fenghua Yu <fenghua.yu@intel.com>
8 *
9 * More information about RDT be found in the Intel (R) x86 Architecture
10 * Software Developer Manual.
11 */
12
13#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14
15#include <linux/cacheinfo.h>
16#include <linux/cpu.h>
17#include <linux/debugfs.h>
18#include <linux/fs.h>
19#include <linux/fs_parser.h>
20#include <linux/sysfs.h>
21#include <linux/kernfs.h>
22#include <linux/seq_buf.h>
23#include <linux/seq_file.h>
24#include <linux/sched/signal.h>
25#include <linux/sched/task.h>
26#include <linux/slab.h>
27#include <linux/task_work.h>
28#include <linux/user_namespace.h>
29
30#include <uapi/linux/magic.h>
31
32#include <asm/resctrl_sched.h>
33#include "internal.h"
34
35DEFINE_STATIC_KEY_FALSE(rdt_enable_key);
36DEFINE_STATIC_KEY_FALSE(rdt_mon_enable_key);
37DEFINE_STATIC_KEY_FALSE(rdt_alloc_enable_key);
38static struct kernfs_root *rdt_root;
39struct rdtgroup rdtgroup_default;
40LIST_HEAD(rdt_all_groups);
41
42/* Kernel fs node for "info" directory under root */
43static struct kernfs_node *kn_info;
44
45/* Kernel fs node for "mon_groups" directory under root */
46static struct kernfs_node *kn_mongrp;
47
48/* Kernel fs node for "mon_data" directory under root */
49static struct kernfs_node *kn_mondata;
50
51static struct seq_buf last_cmd_status;
52static char last_cmd_status_buf[512];
53
54struct dentry *debugfs_resctrl;
55
56void rdt_last_cmd_clear(void)
57{
58 lockdep_assert_held(&rdtgroup_mutex);
59 seq_buf_clear(&last_cmd_status);
60}
61
62void rdt_last_cmd_puts(const char *s)
63{
64 lockdep_assert_held(&rdtgroup_mutex);
65 seq_buf_puts(&last_cmd_status, s);
66}
67
68void rdt_last_cmd_printf(const char *fmt, ...)
69{
70 va_list ap;
71
72 va_start(ap, fmt);
73 lockdep_assert_held(&rdtgroup_mutex);
74 seq_buf_vprintf(&last_cmd_status, fmt, ap);
75 va_end(ap);
76}
77
78/*
79 * Trivial allocator for CLOSIDs. Since h/w only supports a small number,
80 * we can keep a bitmap of free CLOSIDs in a single integer.
81 *
82 * Using a global CLOSID across all resources has some advantages and
83 * some drawbacks:
84 * + We can simply set "current->closid" to assign a task to a resource
85 * group.
86 * + Context switch code can avoid extra memory references deciding which
87 * CLOSID to load into the PQR_ASSOC MSR
88 * - We give up some options in configuring resource groups across multi-socket
89 * systems.
90 * - Our choices on how to configure each resource become progressively more
91 * limited as the number of resources grows.
92 */
93static int closid_free_map;
94static int closid_free_map_len;
95
96int closids_supported(void)
97{
98 return closid_free_map_len;
99}
100
101static void closid_init(void)
102{
103 struct rdt_resource *r;
104 int rdt_min_closid = 32;
105
106 /* Compute rdt_min_closid across all resources */
107 for_each_alloc_enabled_rdt_resource(r)
108 rdt_min_closid = min(rdt_min_closid, r->num_closid);
109
110 closid_free_map = BIT_MASK(rdt_min_closid) - 1;
111
112 /* CLOSID 0 is always reserved for the default group */
113 closid_free_map &= ~1;
114 closid_free_map_len = rdt_min_closid;
115}
116
117static int closid_alloc(void)
118{
119 u32 closid = ffs(closid_free_map);
120
121 if (closid == 0)
122 return -ENOSPC;
123 closid--;
124 closid_free_map &= ~(1 << closid);
125
126 return closid;
127}
128
129void closid_free(int closid)
130{
131 closid_free_map |= 1 << closid;
132}
133
134/**
135 * closid_allocated - test if provided closid is in use
136 * @closid: closid to be tested
137 *
138 * Return: true if @closid is currently associated with a resource group,
139 * false if @closid is free
140 */
141static bool closid_allocated(unsigned int closid)
142{
143 return (closid_free_map & (1 << closid)) == 0;
144}
145
146/**
147 * rdtgroup_mode_by_closid - Return mode of resource group with closid
148 * @closid: closid if the resource group
149 *
150 * Each resource group is associated with a @closid. Here the mode
151 * of a resource group can be queried by searching for it using its closid.
152 *
153 * Return: mode as &enum rdtgrp_mode of resource group with closid @closid
154 */
155enum rdtgrp_mode rdtgroup_mode_by_closid(int closid)
156{
157 struct rdtgroup *rdtgrp;
158
159 list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
160 if (rdtgrp->closid == closid)
161 return rdtgrp->mode;
162 }
163
164 return RDT_NUM_MODES;
165}
166
167static const char * const rdt_mode_str[] = {
168 [RDT_MODE_SHAREABLE] = "shareable",
169 [RDT_MODE_EXCLUSIVE] = "exclusive",
170 [RDT_MODE_PSEUDO_LOCKSETUP] = "pseudo-locksetup",
171 [RDT_MODE_PSEUDO_LOCKED] = "pseudo-locked",
172};
173
174/**
175 * rdtgroup_mode_str - Return the string representation of mode
176 * @mode: the resource group mode as &enum rdtgroup_mode
177 *
178 * Return: string representation of valid mode, "unknown" otherwise
179 */
180static const char *rdtgroup_mode_str(enum rdtgrp_mode mode)
181{
182 if (mode < RDT_MODE_SHAREABLE || mode >= RDT_NUM_MODES)
183 return "unknown";
184
185 return rdt_mode_str[mode];
186}
187
188/* set uid and gid of rdtgroup dirs and files to that of the creator */
189static int rdtgroup_kn_set_ugid(struct kernfs_node *kn)
190{
191 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
192 .ia_uid = current_fsuid(),
193 .ia_gid = current_fsgid(), };
194
195 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
196 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
197 return 0;
198
199 return kernfs_setattr(kn, &iattr);
200}
201
202static int rdtgroup_add_file(struct kernfs_node *parent_kn, struct rftype *rft)
203{
204 struct kernfs_node *kn;
205 int ret;
206
207 kn = __kernfs_create_file(parent_kn, rft->name, rft->mode,
208 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
209 0, rft->kf_ops, rft, NULL, NULL);
210 if (IS_ERR(kn))
211 return PTR_ERR(kn);
212
213 ret = rdtgroup_kn_set_ugid(kn);
214 if (ret) {
215 kernfs_remove(kn);
216 return ret;
217 }
218
219 return 0;
220}
221
222static int rdtgroup_seqfile_show(struct seq_file *m, void *arg)
223{
224 struct kernfs_open_file *of = m->private;
225 struct rftype *rft = of->kn->priv;
226
227 if (rft->seq_show)
228 return rft->seq_show(of, m, arg);
229 return 0;
230}
231
232static ssize_t rdtgroup_file_write(struct kernfs_open_file *of, char *buf,
233 size_t nbytes, loff_t off)
234{
235 struct rftype *rft = of->kn->priv;
236
237 if (rft->write)
238 return rft->write(of, buf, nbytes, off);
239
240 return -EINVAL;
241}
242
243static struct kernfs_ops rdtgroup_kf_single_ops = {
244 .atomic_write_len = PAGE_SIZE,
245 .write = rdtgroup_file_write,
246 .seq_show = rdtgroup_seqfile_show,
247};
248
249static struct kernfs_ops kf_mondata_ops = {
250 .atomic_write_len = PAGE_SIZE,
251 .seq_show = rdtgroup_mondata_show,
252};
253
254static bool is_cpu_list(struct kernfs_open_file *of)
255{
256 struct rftype *rft = of->kn->priv;
257
258 return rft->flags & RFTYPE_FLAGS_CPUS_LIST;
259}
260
261static int rdtgroup_cpus_show(struct kernfs_open_file *of,
262 struct seq_file *s, void *v)
263{
264 struct rdtgroup *rdtgrp;
265 struct cpumask *mask;
266 int ret = 0;
267
268 rdtgrp = rdtgroup_kn_lock_live(of->kn);
269
270 if (rdtgrp) {
271 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
272 if (!rdtgrp->plr->d) {
273 rdt_last_cmd_clear();
274 rdt_last_cmd_puts("Cache domain offline\n");
275 ret = -ENODEV;
276 } else {
277 mask = &rdtgrp->plr->d->cpu_mask;
278 seq_printf(s, is_cpu_list(of) ?
279 "%*pbl\n" : "%*pb\n",
280 cpumask_pr_args(mask));
281 }
282 } else {
283 seq_printf(s, is_cpu_list(of) ? "%*pbl\n" : "%*pb\n",
284 cpumask_pr_args(&rdtgrp->cpu_mask));
285 }
286 } else {
287 ret = -ENOENT;
288 }
289 rdtgroup_kn_unlock(of->kn);
290
291 return ret;
292}
293
294/*
295 * This is safe against resctrl_sched_in() called from __switch_to()
296 * because __switch_to() is executed with interrupts disabled. A local call
297 * from update_closid_rmid() is proteced against __switch_to() because
298 * preemption is disabled.
299 */
300static void update_cpu_closid_rmid(void *info)
301{
302 struct rdtgroup *r = info;
303
304 if (r) {
305 this_cpu_write(pqr_state.default_closid, r->closid);
306 this_cpu_write(pqr_state.default_rmid, r->mon.rmid);
307 }
308
309 /*
310 * We cannot unconditionally write the MSR because the current
311 * executing task might have its own closid selected. Just reuse
312 * the context switch code.
313 */
314 resctrl_sched_in();
315}
316
317/*
318 * Update the PGR_ASSOC MSR on all cpus in @cpu_mask,
319 *
320 * Per task closids/rmids must have been set up before calling this function.
321 */
322static void
323update_closid_rmid(const struct cpumask *cpu_mask, struct rdtgroup *r)
324{
325 int cpu = get_cpu();
326
327 if (cpumask_test_cpu(cpu, cpu_mask))
328 update_cpu_closid_rmid(r);
329 smp_call_function_many(cpu_mask, update_cpu_closid_rmid, r, 1);
330 put_cpu();
331}
332
333static int cpus_mon_write(struct rdtgroup *rdtgrp, cpumask_var_t newmask,
334 cpumask_var_t tmpmask)
335{
336 struct rdtgroup *prgrp = rdtgrp->mon.parent, *crgrp;
337 struct list_head *head;
338
339 /* Check whether cpus belong to parent ctrl group */
340 cpumask_andnot(tmpmask, newmask, &prgrp->cpu_mask);
341 if (cpumask_weight(tmpmask)) {
342 rdt_last_cmd_puts("Can only add CPUs to mongroup that belong to parent\n");
343 return -EINVAL;
344 }
345
346 /* Check whether cpus are dropped from this group */
347 cpumask_andnot(tmpmask, &rdtgrp->cpu_mask, newmask);
348 if (cpumask_weight(tmpmask)) {
349 /* Give any dropped cpus to parent rdtgroup */
350 cpumask_or(&prgrp->cpu_mask, &prgrp->cpu_mask, tmpmask);
351 update_closid_rmid(tmpmask, prgrp);
352 }
353
354 /*
355 * If we added cpus, remove them from previous group that owned them
356 * and update per-cpu rmid
357 */
358 cpumask_andnot(tmpmask, newmask, &rdtgrp->cpu_mask);
359 if (cpumask_weight(tmpmask)) {
360 head = &prgrp->mon.crdtgrp_list;
361 list_for_each_entry(crgrp, head, mon.crdtgrp_list) {
362 if (crgrp == rdtgrp)
363 continue;
364 cpumask_andnot(&crgrp->cpu_mask, &crgrp->cpu_mask,
365 tmpmask);
366 }
367 update_closid_rmid(tmpmask, rdtgrp);
368 }
369
370 /* Done pushing/pulling - update this group with new mask */
371 cpumask_copy(&rdtgrp->cpu_mask, newmask);
372
373 return 0;
374}
375
376static void cpumask_rdtgrp_clear(struct rdtgroup *r, struct cpumask *m)
377{
378 struct rdtgroup *crgrp;
379
380 cpumask_andnot(&r->cpu_mask, &r->cpu_mask, m);
381 /* update the child mon group masks as well*/
382 list_for_each_entry(crgrp, &r->mon.crdtgrp_list, mon.crdtgrp_list)
383 cpumask_and(&crgrp->cpu_mask, &r->cpu_mask, &crgrp->cpu_mask);
384}
385
386static int cpus_ctrl_write(struct rdtgroup *rdtgrp, cpumask_var_t newmask,
387 cpumask_var_t tmpmask, cpumask_var_t tmpmask1)
388{
389 struct rdtgroup *r, *crgrp;
390 struct list_head *head;
391
392 /* Check whether cpus are dropped from this group */
393 cpumask_andnot(tmpmask, &rdtgrp->cpu_mask, newmask);
394 if (cpumask_weight(tmpmask)) {
395 /* Can't drop from default group */
396 if (rdtgrp == &rdtgroup_default) {
397 rdt_last_cmd_puts("Can't drop CPUs from default group\n");
398 return -EINVAL;
399 }
400
401 /* Give any dropped cpus to rdtgroup_default */
402 cpumask_or(&rdtgroup_default.cpu_mask,
403 &rdtgroup_default.cpu_mask, tmpmask);
404 update_closid_rmid(tmpmask, &rdtgroup_default);
405 }
406
407 /*
408 * If we added cpus, remove them from previous group and
409 * the prev group's child groups that owned them
410 * and update per-cpu closid/rmid.
411 */
412 cpumask_andnot(tmpmask, newmask, &rdtgrp->cpu_mask);
413 if (cpumask_weight(tmpmask)) {
414 list_for_each_entry(r, &rdt_all_groups, rdtgroup_list) {
415 if (r == rdtgrp)
416 continue;
417 cpumask_and(tmpmask1, &r->cpu_mask, tmpmask);
418 if (cpumask_weight(tmpmask1))
419 cpumask_rdtgrp_clear(r, tmpmask1);
420 }
421 update_closid_rmid(tmpmask, rdtgrp);
422 }
423
424 /* Done pushing/pulling - update this group with new mask */
425 cpumask_copy(&rdtgrp->cpu_mask, newmask);
426
427 /*
428 * Clear child mon group masks since there is a new parent mask
429 * now and update the rmid for the cpus the child lost.
430 */
431 head = &rdtgrp->mon.crdtgrp_list;
432 list_for_each_entry(crgrp, head, mon.crdtgrp_list) {
433 cpumask_and(tmpmask, &rdtgrp->cpu_mask, &crgrp->cpu_mask);
434 update_closid_rmid(tmpmask, rdtgrp);
435 cpumask_clear(&crgrp->cpu_mask);
436 }
437
438 return 0;
439}
440
441static ssize_t rdtgroup_cpus_write(struct kernfs_open_file *of,
442 char *buf, size_t nbytes, loff_t off)
443{
444 cpumask_var_t tmpmask, newmask, tmpmask1;
445 struct rdtgroup *rdtgrp;
446 int ret;
447
448 if (!buf)
449 return -EINVAL;
450
451 if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
452 return -ENOMEM;
453 if (!zalloc_cpumask_var(&newmask, GFP_KERNEL)) {
454 free_cpumask_var(tmpmask);
455 return -ENOMEM;
456 }
457 if (!zalloc_cpumask_var(&tmpmask1, GFP_KERNEL)) {
458 free_cpumask_var(tmpmask);
459 free_cpumask_var(newmask);
460 return -ENOMEM;
461 }
462
463 rdtgrp = rdtgroup_kn_lock_live(of->kn);
464 if (!rdtgrp) {
465 ret = -ENOENT;
466 goto unlock;
467 }
468
469 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED ||
470 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
471 ret = -EINVAL;
472 rdt_last_cmd_puts("Pseudo-locking in progress\n");
473 goto unlock;
474 }
475
476 if (is_cpu_list(of))
477 ret = cpulist_parse(buf, newmask);
478 else
479 ret = cpumask_parse(buf, newmask);
480
481 if (ret) {
482 rdt_last_cmd_puts("Bad CPU list/mask\n");
483 goto unlock;
484 }
485
486 /* check that user didn't specify any offline cpus */
487 cpumask_andnot(tmpmask, newmask, cpu_online_mask);
488 if (cpumask_weight(tmpmask)) {
489 ret = -EINVAL;
490 rdt_last_cmd_puts("Can only assign online CPUs\n");
491 goto unlock;
492 }
493
494 if (rdtgrp->type == RDTCTRL_GROUP)
495 ret = cpus_ctrl_write(rdtgrp, newmask, tmpmask, tmpmask1);
496 else if (rdtgrp->type == RDTMON_GROUP)
497 ret = cpus_mon_write(rdtgrp, newmask, tmpmask);
498 else
499 ret = -EINVAL;
500
501unlock:
502 rdtgroup_kn_unlock(of->kn);
503 free_cpumask_var(tmpmask);
504 free_cpumask_var(newmask);
505 free_cpumask_var(tmpmask1);
506
507 return ret ?: nbytes;
508}
509
510struct task_move_callback {
511 struct callback_head work;
512 struct rdtgroup *rdtgrp;
513};
514
515static void move_myself(struct callback_head *head)
516{
517 struct task_move_callback *callback;
518 struct rdtgroup *rdtgrp;
519
520 callback = container_of(head, struct task_move_callback, work);
521 rdtgrp = callback->rdtgrp;
522
523 /*
524 * If resource group was deleted before this task work callback
525 * was invoked, then assign the task to root group and free the
526 * resource group.
527 */
528 if (atomic_dec_and_test(&rdtgrp->waitcount) &&
529 (rdtgrp->flags & RDT_DELETED)) {
530 current->closid = 0;
531 current->rmid = 0;
532 kfree(rdtgrp);
533 }
534
535 preempt_disable();
536 /* update PQR_ASSOC MSR to make resource group go into effect */
537 resctrl_sched_in();
538 preempt_enable();
539
540 kfree(callback);
541}
542
543static int __rdtgroup_move_task(struct task_struct *tsk,
544 struct rdtgroup *rdtgrp)
545{
546 struct task_move_callback *callback;
547 int ret;
548
549 callback = kzalloc(sizeof(*callback), GFP_KERNEL);
550 if (!callback)
551 return -ENOMEM;
552 callback->work.func = move_myself;
553 callback->rdtgrp = rdtgrp;
554
555 /*
556 * Take a refcount, so rdtgrp cannot be freed before the
557 * callback has been invoked.
558 */
559 atomic_inc(&rdtgrp->waitcount);
560 ret = task_work_add(tsk, &callback->work, true);
561 if (ret) {
562 /*
563 * Task is exiting. Drop the refcount and free the callback.
564 * No need to check the refcount as the group cannot be
565 * deleted before the write function unlocks rdtgroup_mutex.
566 */
567 atomic_dec(&rdtgrp->waitcount);
568 kfree(callback);
569 rdt_last_cmd_puts("Task exited\n");
570 } else {
571 /*
572 * For ctrl_mon groups move both closid and rmid.
573 * For monitor groups, can move the tasks only from
574 * their parent CTRL group.
575 */
576 if (rdtgrp->type == RDTCTRL_GROUP) {
577 tsk->closid = rdtgrp->closid;
578 tsk->rmid = rdtgrp->mon.rmid;
579 } else if (rdtgrp->type == RDTMON_GROUP) {
580 if (rdtgrp->mon.parent->closid == tsk->closid) {
581 tsk->rmid = rdtgrp->mon.rmid;
582 } else {
583 rdt_last_cmd_puts("Can't move task to different control group\n");
584 ret = -EINVAL;
585 }
586 }
587 }
588 return ret;
589}
590
591/**
592 * rdtgroup_tasks_assigned - Test if tasks have been assigned to resource group
593 * @r: Resource group
594 *
595 * Return: 1 if tasks have been assigned to @r, 0 otherwise
596 */
597int rdtgroup_tasks_assigned(struct rdtgroup *r)
598{
599 struct task_struct *p, *t;
600 int ret = 0;
601
602 lockdep_assert_held(&rdtgroup_mutex);
603
604 rcu_read_lock();
605 for_each_process_thread(p, t) {
606 if ((r->type == RDTCTRL_GROUP && t->closid == r->closid) ||
607 (r->type == RDTMON_GROUP && t->rmid == r->mon.rmid)) {
608 ret = 1;
609 break;
610 }
611 }
612 rcu_read_unlock();
613
614 return ret;
615}
616
617static int rdtgroup_task_write_permission(struct task_struct *task,
618 struct kernfs_open_file *of)
619{
620 const struct cred *tcred = get_task_cred(task);
621 const struct cred *cred = current_cred();
622 int ret = 0;
623
624 /*
625 * Even if we're attaching all tasks in the thread group, we only
626 * need to check permissions on one of them.
627 */
628 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
629 !uid_eq(cred->euid, tcred->uid) &&
630 !uid_eq(cred->euid, tcred->suid)) {
631 rdt_last_cmd_printf("No permission to move task %d\n", task->pid);
632 ret = -EPERM;
633 }
634
635 put_cred(tcred);
636 return ret;
637}
638
639static int rdtgroup_move_task(pid_t pid, struct rdtgroup *rdtgrp,
640 struct kernfs_open_file *of)
641{
642 struct task_struct *tsk;
643 int ret;
644
645 rcu_read_lock();
646 if (pid) {
647 tsk = find_task_by_vpid(pid);
648 if (!tsk) {
649 rcu_read_unlock();
650 rdt_last_cmd_printf("No task %d\n", pid);
651 return -ESRCH;
652 }
653 } else {
654 tsk = current;
655 }
656
657 get_task_struct(tsk);
658 rcu_read_unlock();
659
660 ret = rdtgroup_task_write_permission(tsk, of);
661 if (!ret)
662 ret = __rdtgroup_move_task(tsk, rdtgrp);
663
664 put_task_struct(tsk);
665 return ret;
666}
667
668static ssize_t rdtgroup_tasks_write(struct kernfs_open_file *of,
669 char *buf, size_t nbytes, loff_t off)
670{
671 struct rdtgroup *rdtgrp;
672 int ret = 0;
673 pid_t pid;
674
675 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
676 return -EINVAL;
677 rdtgrp = rdtgroup_kn_lock_live(of->kn);
678 if (!rdtgrp) {
679 rdtgroup_kn_unlock(of->kn);
680 return -ENOENT;
681 }
682 rdt_last_cmd_clear();
683
684 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED ||
685 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
686 ret = -EINVAL;
687 rdt_last_cmd_puts("Pseudo-locking in progress\n");
688 goto unlock;
689 }
690
691 ret = rdtgroup_move_task(pid, rdtgrp, of);
692
693unlock:
694 rdtgroup_kn_unlock(of->kn);
695
696 return ret ?: nbytes;
697}
698
699static void show_rdt_tasks(struct rdtgroup *r, struct seq_file *s)
700{
701 struct task_struct *p, *t;
702
703 rcu_read_lock();
704 for_each_process_thread(p, t) {
705 if ((r->type == RDTCTRL_GROUP && t->closid == r->closid) ||
706 (r->type == RDTMON_GROUP && t->rmid == r->mon.rmid))
707 seq_printf(s, "%d\n", t->pid);
708 }
709 rcu_read_unlock();
710}
711
712static int rdtgroup_tasks_show(struct kernfs_open_file *of,
713 struct seq_file *s, void *v)
714{
715 struct rdtgroup *rdtgrp;
716 int ret = 0;
717
718 rdtgrp = rdtgroup_kn_lock_live(of->kn);
719 if (rdtgrp)
720 show_rdt_tasks(rdtgrp, s);
721 else
722 ret = -ENOENT;
723 rdtgroup_kn_unlock(of->kn);
724
725 return ret;
726}
727
728static int rdt_last_cmd_status_show(struct kernfs_open_file *of,
729 struct seq_file *seq, void *v)
730{
731 int len;
732
733 mutex_lock(&rdtgroup_mutex);
734 len = seq_buf_used(&last_cmd_status);
735 if (len)
736 seq_printf(seq, "%.*s", len, last_cmd_status_buf);
737 else
738 seq_puts(seq, "ok\n");
739 mutex_unlock(&rdtgroup_mutex);
740 return 0;
741}
742
743static int rdt_num_closids_show(struct kernfs_open_file *of,
744 struct seq_file *seq, void *v)
745{
746 struct rdt_resource *r = of->kn->parent->priv;
747
748 seq_printf(seq, "%d\n", r->num_closid);
749 return 0;
750}
751
752static int rdt_default_ctrl_show(struct kernfs_open_file *of,
753 struct seq_file *seq, void *v)
754{
755 struct rdt_resource *r = of->kn->parent->priv;
756
757 seq_printf(seq, "%x\n", r->default_ctrl);
758 return 0;
759}
760
761static int rdt_min_cbm_bits_show(struct kernfs_open_file *of,
762 struct seq_file *seq, void *v)
763{
764 struct rdt_resource *r = of->kn->parent->priv;
765
766 seq_printf(seq, "%u\n", r->cache.min_cbm_bits);
767 return 0;
768}
769
770static int rdt_shareable_bits_show(struct kernfs_open_file *of,
771 struct seq_file *seq, void *v)
772{
773 struct rdt_resource *r = of->kn->parent->priv;
774
775 seq_printf(seq, "%x\n", r->cache.shareable_bits);
776 return 0;
777}
778
779/**
780 * rdt_bit_usage_show - Display current usage of resources
781 *
782 * A domain is a shared resource that can now be allocated differently. Here
783 * we display the current regions of the domain as an annotated bitmask.
784 * For each domain of this resource its allocation bitmask
785 * is annotated as below to indicate the current usage of the corresponding bit:
786 * 0 - currently unused
787 * X - currently available for sharing and used by software and hardware
788 * H - currently used by hardware only but available for software use
789 * S - currently used and shareable by software only
790 * E - currently used exclusively by one resource group
791 * P - currently pseudo-locked by one resource group
792 */
793static int rdt_bit_usage_show(struct kernfs_open_file *of,
794 struct seq_file *seq, void *v)
795{
796 struct rdt_resource *r = of->kn->parent->priv;
797 /*
798 * Use unsigned long even though only 32 bits are used to ensure
799 * test_bit() is used safely.
800 */
801 unsigned long sw_shareable = 0, hw_shareable = 0;
802 unsigned long exclusive = 0, pseudo_locked = 0;
803 struct rdt_domain *dom;
804 int i, hwb, swb, excl, psl;
805 enum rdtgrp_mode mode;
806 bool sep = false;
807 u32 *ctrl;
808
809 mutex_lock(&rdtgroup_mutex);
810 hw_shareable = r->cache.shareable_bits;
811 list_for_each_entry(dom, &r->domains, list) {
812 if (sep)
813 seq_putc(seq, ';');
814 ctrl = dom->ctrl_val;
815 sw_shareable = 0;
816 exclusive = 0;
817 seq_printf(seq, "%d=", dom->id);
818 for (i = 0; i < closids_supported(); i++, ctrl++) {
819 if (!closid_allocated(i))
820 continue;
821 mode = rdtgroup_mode_by_closid(i);
822 switch (mode) {
823 case RDT_MODE_SHAREABLE:
824 sw_shareable |= *ctrl;
825 break;
826 case RDT_MODE_EXCLUSIVE:
827 exclusive |= *ctrl;
828 break;
829 case RDT_MODE_PSEUDO_LOCKSETUP:
830 /*
831 * RDT_MODE_PSEUDO_LOCKSETUP is possible
832 * here but not included since the CBM
833 * associated with this CLOSID in this mode
834 * is not initialized and no task or cpu can be
835 * assigned this CLOSID.
836 */
837 break;
838 case RDT_MODE_PSEUDO_LOCKED:
839 case RDT_NUM_MODES:
840 WARN(1,
841 "invalid mode for closid %d\n", i);
842 break;
843 }
844 }
845 for (i = r->cache.cbm_len - 1; i >= 0; i--) {
846 pseudo_locked = dom->plr ? dom->plr->cbm : 0;
847 hwb = test_bit(i, &hw_shareable);
848 swb = test_bit(i, &sw_shareable);
849 excl = test_bit(i, &exclusive);
850 psl = test_bit(i, &pseudo_locked);
851 if (hwb && swb)
852 seq_putc(seq, 'X');
853 else if (hwb && !swb)
854 seq_putc(seq, 'H');
855 else if (!hwb && swb)
856 seq_putc(seq, 'S');
857 else if (excl)
858 seq_putc(seq, 'E');
859 else if (psl)
860 seq_putc(seq, 'P');
861 else /* Unused bits remain */
862 seq_putc(seq, '0');
863 }
864 sep = true;
865 }
866 seq_putc(seq, '\n');
867 mutex_unlock(&rdtgroup_mutex);
868 return 0;
869}
870
871static int rdt_min_bw_show(struct kernfs_open_file *of,
872 struct seq_file *seq, void *v)
873{
874 struct rdt_resource *r = of->kn->parent->priv;
875
876 seq_printf(seq, "%u\n", r->membw.min_bw);
877 return 0;
878}
879
880static int rdt_num_rmids_show(struct kernfs_open_file *of,
881 struct seq_file *seq, void *v)
882{
883 struct rdt_resource *r = of->kn->parent->priv;
884
885 seq_printf(seq, "%d\n", r->num_rmid);
886
887 return 0;
888}
889
890static int rdt_mon_features_show(struct kernfs_open_file *of,
891 struct seq_file *seq, void *v)
892{
893 struct rdt_resource *r = of->kn->parent->priv;
894 struct mon_evt *mevt;
895
896 list_for_each_entry(mevt, &r->evt_list, list)
897 seq_printf(seq, "%s\n", mevt->name);
898
899 return 0;
900}
901
902static int rdt_bw_gran_show(struct kernfs_open_file *of,
903 struct seq_file *seq, void *v)
904{
905 struct rdt_resource *r = of->kn->parent->priv;
906
907 seq_printf(seq, "%u\n", r->membw.bw_gran);
908 return 0;
909}
910
911static int rdt_delay_linear_show(struct kernfs_open_file *of,
912 struct seq_file *seq, void *v)
913{
914 struct rdt_resource *r = of->kn->parent->priv;
915
916 seq_printf(seq, "%u\n", r->membw.delay_linear);
917 return 0;
918}
919
920static int max_threshold_occ_show(struct kernfs_open_file *of,
921 struct seq_file *seq, void *v)
922{
923 struct rdt_resource *r = of->kn->parent->priv;
924
925 seq_printf(seq, "%u\n", resctrl_cqm_threshold * r->mon_scale);
926
927 return 0;
928}
929
930static ssize_t max_threshold_occ_write(struct kernfs_open_file *of,
931 char *buf, size_t nbytes, loff_t off)
932{
933 struct rdt_resource *r = of->kn->parent->priv;
934 unsigned int bytes;
935 int ret;
936
937 ret = kstrtouint(buf, 0, &bytes);
938 if (ret)
939 return ret;
940
941 if (bytes > (boot_cpu_data.x86_cache_size * 1024))
942 return -EINVAL;
943
944 resctrl_cqm_threshold = bytes / r->mon_scale;
945
946 return nbytes;
947}
948
949/*
950 * rdtgroup_mode_show - Display mode of this resource group
951 */
952static int rdtgroup_mode_show(struct kernfs_open_file *of,
953 struct seq_file *s, void *v)
954{
955 struct rdtgroup *rdtgrp;
956
957 rdtgrp = rdtgroup_kn_lock_live(of->kn);
958 if (!rdtgrp) {
959 rdtgroup_kn_unlock(of->kn);
960 return -ENOENT;
961 }
962
963 seq_printf(s, "%s\n", rdtgroup_mode_str(rdtgrp->mode));
964
965 rdtgroup_kn_unlock(of->kn);
966 return 0;
967}
968
969/**
970 * rdt_cdp_peer_get - Retrieve CDP peer if it exists
971 * @r: RDT resource to which RDT domain @d belongs
972 * @d: Cache instance for which a CDP peer is requested
973 * @r_cdp: RDT resource that shares hardware with @r (RDT resource peer)
974 * Used to return the result.
975 * @d_cdp: RDT domain that shares hardware with @d (RDT domain peer)
976 * Used to return the result.
977 *
978 * RDT resources are managed independently and by extension the RDT domains
979 * (RDT resource instances) are managed independently also. The Code and
980 * Data Prioritization (CDP) RDT resources, while managed independently,
981 * could refer to the same underlying hardware. For example,
982 * RDT_RESOURCE_L2CODE and RDT_RESOURCE_L2DATA both refer to the L2 cache.
983 *
984 * When provided with an RDT resource @r and an instance of that RDT
985 * resource @d rdt_cdp_peer_get() will return if there is a peer RDT
986 * resource and the exact instance that shares the same hardware.
987 *
988 * Return: 0 if a CDP peer was found, <0 on error or if no CDP peer exists.
989 * If a CDP peer was found, @r_cdp will point to the peer RDT resource
990 * and @d_cdp will point to the peer RDT domain.
991 */
992static int rdt_cdp_peer_get(struct rdt_resource *r, struct rdt_domain *d,
993 struct rdt_resource **r_cdp,
994 struct rdt_domain **d_cdp)
995{
996 struct rdt_resource *_r_cdp = NULL;
997 struct rdt_domain *_d_cdp = NULL;
998 int ret = 0;
999
1000 switch (r->rid) {
1001 case RDT_RESOURCE_L3DATA:
1002 _r_cdp = &rdt_resources_all[RDT_RESOURCE_L3CODE];
1003 break;
1004 case RDT_RESOURCE_L3CODE:
1005 _r_cdp = &rdt_resources_all[RDT_RESOURCE_L3DATA];
1006 break;
1007 case RDT_RESOURCE_L2DATA:
1008 _r_cdp = &rdt_resources_all[RDT_RESOURCE_L2CODE];
1009 break;
1010 case RDT_RESOURCE_L2CODE:
1011 _r_cdp = &rdt_resources_all[RDT_RESOURCE_L2DATA];
1012 break;
1013 default:
1014 ret = -ENOENT;
1015 goto out;
1016 }
1017
1018 /*
1019 * When a new CPU comes online and CDP is enabled then the new
1020 * RDT domains (if any) associated with both CDP RDT resources
1021 * are added in the same CPU online routine while the
1022 * rdtgroup_mutex is held. It should thus not happen for one
1023 * RDT domain to exist and be associated with its RDT CDP
1024 * resource but there is no RDT domain associated with the
1025 * peer RDT CDP resource. Hence the WARN.
1026 */
1027 _d_cdp = rdt_find_domain(_r_cdp, d->id, NULL);
1028 if (WARN_ON(IS_ERR_OR_NULL(_d_cdp))) {
1029 _r_cdp = NULL;
1030 ret = -EINVAL;
1031 }
1032
1033out:
1034 *r_cdp = _r_cdp;
1035 *d_cdp = _d_cdp;
1036
1037 return ret;
1038}
1039
1040/**
1041 * __rdtgroup_cbm_overlaps - Does CBM for intended closid overlap with other
1042 * @r: Resource to which domain instance @d belongs.
1043 * @d: The domain instance for which @closid is being tested.
1044 * @cbm: Capacity bitmask being tested.
1045 * @closid: Intended closid for @cbm.
1046 * @exclusive: Only check if overlaps with exclusive resource groups
1047 *
1048 * Checks if provided @cbm intended to be used for @closid on domain
1049 * @d overlaps with any other closids or other hardware usage associated
1050 * with this domain. If @exclusive is true then only overlaps with
1051 * resource groups in exclusive mode will be considered. If @exclusive
1052 * is false then overlaps with any resource group or hardware entities
1053 * will be considered.
1054 *
1055 * @cbm is unsigned long, even if only 32 bits are used, to make the
1056 * bitmap functions work correctly.
1057 *
1058 * Return: false if CBM does not overlap, true if it does.
1059 */
1060static bool __rdtgroup_cbm_overlaps(struct rdt_resource *r, struct rdt_domain *d,
1061 unsigned long cbm, int closid, bool exclusive)
1062{
1063 enum rdtgrp_mode mode;
1064 unsigned long ctrl_b;
1065 u32 *ctrl;
1066 int i;
1067
1068 /* Check for any overlap with regions used by hardware directly */
1069 if (!exclusive) {
1070 ctrl_b = r->cache.shareable_bits;
1071 if (bitmap_intersects(&cbm, &ctrl_b, r->cache.cbm_len))
1072 return true;
1073 }
1074
1075 /* Check for overlap with other resource groups */
1076 ctrl = d->ctrl_val;
1077 for (i = 0; i < closids_supported(); i++, ctrl++) {
1078 ctrl_b = *ctrl;
1079 mode = rdtgroup_mode_by_closid(i);
1080 if (closid_allocated(i) && i != closid &&
1081 mode != RDT_MODE_PSEUDO_LOCKSETUP) {
1082 if (bitmap_intersects(&cbm, &ctrl_b, r->cache.cbm_len)) {
1083 if (exclusive) {
1084 if (mode == RDT_MODE_EXCLUSIVE)
1085 return true;
1086 continue;
1087 }
1088 return true;
1089 }
1090 }
1091 }
1092
1093 return false;
1094}
1095
1096/**
1097 * rdtgroup_cbm_overlaps - Does CBM overlap with other use of hardware
1098 * @r: Resource to which domain instance @d belongs.
1099 * @d: The domain instance for which @closid is being tested.
1100 * @cbm: Capacity bitmask being tested.
1101 * @closid: Intended closid for @cbm.
1102 * @exclusive: Only check if overlaps with exclusive resource groups
1103 *
1104 * Resources that can be allocated using a CBM can use the CBM to control
1105 * the overlap of these allocations. rdtgroup_cmb_overlaps() is the test
1106 * for overlap. Overlap test is not limited to the specific resource for
1107 * which the CBM is intended though - when dealing with CDP resources that
1108 * share the underlying hardware the overlap check should be performed on
1109 * the CDP resource sharing the hardware also.
1110 *
1111 * Refer to description of __rdtgroup_cbm_overlaps() for the details of the
1112 * overlap test.
1113 *
1114 * Return: true if CBM overlap detected, false if there is no overlap
1115 */
1116bool rdtgroup_cbm_overlaps(struct rdt_resource *r, struct rdt_domain *d,
1117 unsigned long cbm, int closid, bool exclusive)
1118{
1119 struct rdt_resource *r_cdp;
1120 struct rdt_domain *d_cdp;
1121
1122 if (__rdtgroup_cbm_overlaps(r, d, cbm, closid, exclusive))
1123 return true;
1124
1125 if (rdt_cdp_peer_get(r, d, &r_cdp, &d_cdp) < 0)
1126 return false;
1127
1128 return __rdtgroup_cbm_overlaps(r_cdp, d_cdp, cbm, closid, exclusive);
1129}
1130
1131/**
1132 * rdtgroup_mode_test_exclusive - Test if this resource group can be exclusive
1133 *
1134 * An exclusive resource group implies that there should be no sharing of
1135 * its allocated resources. At the time this group is considered to be
1136 * exclusive this test can determine if its current schemata supports this
1137 * setting by testing for overlap with all other resource groups.
1138 *
1139 * Return: true if resource group can be exclusive, false if there is overlap
1140 * with allocations of other resource groups and thus this resource group
1141 * cannot be exclusive.
1142 */
1143static bool rdtgroup_mode_test_exclusive(struct rdtgroup *rdtgrp)
1144{
1145 int closid = rdtgrp->closid;
1146 struct rdt_resource *r;
1147 bool has_cache = false;
1148 struct rdt_domain *d;
1149
1150 for_each_alloc_enabled_rdt_resource(r) {
1151 if (r->rid == RDT_RESOURCE_MBA)
1152 continue;
1153 has_cache = true;
1154 list_for_each_entry(d, &r->domains, list) {
1155 if (rdtgroup_cbm_overlaps(r, d, d->ctrl_val[closid],
1156 rdtgrp->closid, false)) {
1157 rdt_last_cmd_puts("Schemata overlaps\n");
1158 return false;
1159 }
1160 }
1161 }
1162
1163 if (!has_cache) {
1164 rdt_last_cmd_puts("Cannot be exclusive without CAT/CDP\n");
1165 return false;
1166 }
1167
1168 return true;
1169}
1170
1171/**
1172 * rdtgroup_mode_write - Modify the resource group's mode
1173 *
1174 */
1175static ssize_t rdtgroup_mode_write(struct kernfs_open_file *of,
1176 char *buf, size_t nbytes, loff_t off)
1177{
1178 struct rdtgroup *rdtgrp;
1179 enum rdtgrp_mode mode;
1180 int ret = 0;
1181
1182 /* Valid input requires a trailing newline */
1183 if (nbytes == 0 || buf[nbytes - 1] != '\n')
1184 return -EINVAL;
1185 buf[nbytes - 1] = '\0';
1186
1187 rdtgrp = rdtgroup_kn_lock_live(of->kn);
1188 if (!rdtgrp) {
1189 rdtgroup_kn_unlock(of->kn);
1190 return -ENOENT;
1191 }
1192
1193 rdt_last_cmd_clear();
1194
1195 mode = rdtgrp->mode;
1196
1197 if ((!strcmp(buf, "shareable") && mode == RDT_MODE_SHAREABLE) ||
1198 (!strcmp(buf, "exclusive") && mode == RDT_MODE_EXCLUSIVE) ||
1199 (!strcmp(buf, "pseudo-locksetup") &&
1200 mode == RDT_MODE_PSEUDO_LOCKSETUP) ||
1201 (!strcmp(buf, "pseudo-locked") && mode == RDT_MODE_PSEUDO_LOCKED))
1202 goto out;
1203
1204 if (mode == RDT_MODE_PSEUDO_LOCKED) {
1205 rdt_last_cmd_puts("Cannot change pseudo-locked group\n");
1206 ret = -EINVAL;
1207 goto out;
1208 }
1209
1210 if (!strcmp(buf, "shareable")) {
1211 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1212 ret = rdtgroup_locksetup_exit(rdtgrp);
1213 if (ret)
1214 goto out;
1215 }
1216 rdtgrp->mode = RDT_MODE_SHAREABLE;
1217 } else if (!strcmp(buf, "exclusive")) {
1218 if (!rdtgroup_mode_test_exclusive(rdtgrp)) {
1219 ret = -EINVAL;
1220 goto out;
1221 }
1222 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1223 ret = rdtgroup_locksetup_exit(rdtgrp);
1224 if (ret)
1225 goto out;
1226 }
1227 rdtgrp->mode = RDT_MODE_EXCLUSIVE;
1228 } else if (!strcmp(buf, "pseudo-locksetup")) {
1229 ret = rdtgroup_locksetup_enter(rdtgrp);
1230 if (ret)
1231 goto out;
1232 rdtgrp->mode = RDT_MODE_PSEUDO_LOCKSETUP;
1233 } else {
1234 rdt_last_cmd_puts("Unknown or unsupported mode\n");
1235 ret = -EINVAL;
1236 }
1237
1238out:
1239 rdtgroup_kn_unlock(of->kn);
1240 return ret ?: nbytes;
1241}
1242
1243/**
1244 * rdtgroup_cbm_to_size - Translate CBM to size in bytes
1245 * @r: RDT resource to which @d belongs.
1246 * @d: RDT domain instance.
1247 * @cbm: bitmask for which the size should be computed.
1248 *
1249 * The bitmask provided associated with the RDT domain instance @d will be
1250 * translated into how many bytes it represents. The size in bytes is
1251 * computed by first dividing the total cache size by the CBM length to
1252 * determine how many bytes each bit in the bitmask represents. The result
1253 * is multiplied with the number of bits set in the bitmask.
1254 *
1255 * @cbm is unsigned long, even if only 32 bits are used to make the
1256 * bitmap functions work correctly.
1257 */
1258unsigned int rdtgroup_cbm_to_size(struct rdt_resource *r,
1259 struct rdt_domain *d, unsigned long cbm)
1260{
1261 struct cpu_cacheinfo *ci;
1262 unsigned int size = 0;
1263 int num_b, i;
1264
1265 num_b = bitmap_weight(&cbm, r->cache.cbm_len);
1266 ci = get_cpu_cacheinfo(cpumask_any(&d->cpu_mask));
1267 for (i = 0; i < ci->num_leaves; i++) {
1268 if (ci->info_list[i].level == r->cache_level) {
1269 size = ci->info_list[i].size / r->cache.cbm_len * num_b;
1270 break;
1271 }
1272 }
1273
1274 return size;
1275}
1276
1277/**
1278 * rdtgroup_size_show - Display size in bytes of allocated regions
1279 *
1280 * The "size" file mirrors the layout of the "schemata" file, printing the
1281 * size in bytes of each region instead of the capacity bitmask.
1282 *
1283 */
1284static int rdtgroup_size_show(struct kernfs_open_file *of,
1285 struct seq_file *s, void *v)
1286{
1287 struct rdtgroup *rdtgrp;
1288 struct rdt_resource *r;
1289 struct rdt_domain *d;
1290 unsigned int size;
1291 int ret = 0;
1292 bool sep;
1293 u32 ctrl;
1294
1295 rdtgrp = rdtgroup_kn_lock_live(of->kn);
1296 if (!rdtgrp) {
1297 rdtgroup_kn_unlock(of->kn);
1298 return -ENOENT;
1299 }
1300
1301 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
1302 if (!rdtgrp->plr->d) {
1303 rdt_last_cmd_clear();
1304 rdt_last_cmd_puts("Cache domain offline\n");
1305 ret = -ENODEV;
1306 } else {
1307 seq_printf(s, "%*s:", max_name_width,
1308 rdtgrp->plr->r->name);
1309 size = rdtgroup_cbm_to_size(rdtgrp->plr->r,
1310 rdtgrp->plr->d,
1311 rdtgrp->plr->cbm);
1312 seq_printf(s, "%d=%u\n", rdtgrp->plr->d->id, size);
1313 }
1314 goto out;
1315 }
1316
1317 for_each_alloc_enabled_rdt_resource(r) {
1318 sep = false;
1319 seq_printf(s, "%*s:", max_name_width, r->name);
1320 list_for_each_entry(d, &r->domains, list) {
1321 if (sep)
1322 seq_putc(s, ';');
1323 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1324 size = 0;
1325 } else {
1326 ctrl = (!is_mba_sc(r) ?
1327 d->ctrl_val[rdtgrp->closid] :
1328 d->mbps_val[rdtgrp->closid]);
1329 if (r->rid == RDT_RESOURCE_MBA)
1330 size = ctrl;
1331 else
1332 size = rdtgroup_cbm_to_size(r, d, ctrl);
1333 }
1334 seq_printf(s, "%d=%u", d->id, size);
1335 sep = true;
1336 }
1337 seq_putc(s, '\n');
1338 }
1339
1340out:
1341 rdtgroup_kn_unlock(of->kn);
1342
1343 return ret;
1344}
1345
1346/* rdtgroup information files for one cache resource. */
1347static struct rftype res_common_files[] = {
1348 {
1349 .name = "last_cmd_status",
1350 .mode = 0444,
1351 .kf_ops = &rdtgroup_kf_single_ops,
1352 .seq_show = rdt_last_cmd_status_show,
1353 .fflags = RF_TOP_INFO,
1354 },
1355 {
1356 .name = "num_closids",
1357 .mode = 0444,
1358 .kf_ops = &rdtgroup_kf_single_ops,
1359 .seq_show = rdt_num_closids_show,
1360 .fflags = RF_CTRL_INFO,
1361 },
1362 {
1363 .name = "mon_features",
1364 .mode = 0444,
1365 .kf_ops = &rdtgroup_kf_single_ops,
1366 .seq_show = rdt_mon_features_show,
1367 .fflags = RF_MON_INFO,
1368 },
1369 {
1370 .name = "num_rmids",
1371 .mode = 0444,
1372 .kf_ops = &rdtgroup_kf_single_ops,
1373 .seq_show = rdt_num_rmids_show,
1374 .fflags = RF_MON_INFO,
1375 },
1376 {
1377 .name = "cbm_mask",
1378 .mode = 0444,
1379 .kf_ops = &rdtgroup_kf_single_ops,
1380 .seq_show = rdt_default_ctrl_show,
1381 .fflags = RF_CTRL_INFO | RFTYPE_RES_CACHE,
1382 },
1383 {
1384 .name = "min_cbm_bits",
1385 .mode = 0444,
1386 .kf_ops = &rdtgroup_kf_single_ops,
1387 .seq_show = rdt_min_cbm_bits_show,
1388 .fflags = RF_CTRL_INFO | RFTYPE_RES_CACHE,
1389 },
1390 {
1391 .name = "shareable_bits",
1392 .mode = 0444,
1393 .kf_ops = &rdtgroup_kf_single_ops,
1394 .seq_show = rdt_shareable_bits_show,
1395 .fflags = RF_CTRL_INFO | RFTYPE_RES_CACHE,
1396 },
1397 {
1398 .name = "bit_usage",
1399 .mode = 0444,
1400 .kf_ops = &rdtgroup_kf_single_ops,
1401 .seq_show = rdt_bit_usage_show,
1402 .fflags = RF_CTRL_INFO | RFTYPE_RES_CACHE,
1403 },
1404 {
1405 .name = "min_bandwidth",
1406 .mode = 0444,
1407 .kf_ops = &rdtgroup_kf_single_ops,
1408 .seq_show = rdt_min_bw_show,
1409 .fflags = RF_CTRL_INFO | RFTYPE_RES_MB,
1410 },
1411 {
1412 .name = "bandwidth_gran",
1413 .mode = 0444,
1414 .kf_ops = &rdtgroup_kf_single_ops,
1415 .seq_show = rdt_bw_gran_show,
1416 .fflags = RF_CTRL_INFO | RFTYPE_RES_MB,
1417 },
1418 {
1419 .name = "delay_linear",
1420 .mode = 0444,
1421 .kf_ops = &rdtgroup_kf_single_ops,
1422 .seq_show = rdt_delay_linear_show,
1423 .fflags = RF_CTRL_INFO | RFTYPE_RES_MB,
1424 },
1425 {
1426 .name = "max_threshold_occupancy",
1427 .mode = 0644,
1428 .kf_ops = &rdtgroup_kf_single_ops,
1429 .write = max_threshold_occ_write,
1430 .seq_show = max_threshold_occ_show,
1431 .fflags = RF_MON_INFO | RFTYPE_RES_CACHE,
1432 },
1433 {
1434 .name = "cpus",
1435 .mode = 0644,
1436 .kf_ops = &rdtgroup_kf_single_ops,
1437 .write = rdtgroup_cpus_write,
1438 .seq_show = rdtgroup_cpus_show,
1439 .fflags = RFTYPE_BASE,
1440 },
1441 {
1442 .name = "cpus_list",
1443 .mode = 0644,
1444 .kf_ops = &rdtgroup_kf_single_ops,
1445 .write = rdtgroup_cpus_write,
1446 .seq_show = rdtgroup_cpus_show,
1447 .flags = RFTYPE_FLAGS_CPUS_LIST,
1448 .fflags = RFTYPE_BASE,
1449 },
1450 {
1451 .name = "tasks",
1452 .mode = 0644,
1453 .kf_ops = &rdtgroup_kf_single_ops,
1454 .write = rdtgroup_tasks_write,
1455 .seq_show = rdtgroup_tasks_show,
1456 .fflags = RFTYPE_BASE,
1457 },
1458 {
1459 .name = "schemata",
1460 .mode = 0644,
1461 .kf_ops = &rdtgroup_kf_single_ops,
1462 .write = rdtgroup_schemata_write,
1463 .seq_show = rdtgroup_schemata_show,
1464 .fflags = RF_CTRL_BASE,
1465 },
1466 {
1467 .name = "mode",
1468 .mode = 0644,
1469 .kf_ops = &rdtgroup_kf_single_ops,
1470 .write = rdtgroup_mode_write,
1471 .seq_show = rdtgroup_mode_show,
1472 .fflags = RF_CTRL_BASE,
1473 },
1474 {
1475 .name = "size",
1476 .mode = 0444,
1477 .kf_ops = &rdtgroup_kf_single_ops,
1478 .seq_show = rdtgroup_size_show,
1479 .fflags = RF_CTRL_BASE,
1480 },
1481
1482};
1483
1484static int rdtgroup_add_files(struct kernfs_node *kn, unsigned long fflags)
1485{
1486 struct rftype *rfts, *rft;
1487 int ret, len;
1488
1489 rfts = res_common_files;
1490 len = ARRAY_SIZE(res_common_files);
1491
1492 lockdep_assert_held(&rdtgroup_mutex);
1493
1494 for (rft = rfts; rft < rfts + len; rft++) {
1495 if ((fflags & rft->fflags) == rft->fflags) {
1496 ret = rdtgroup_add_file(kn, rft);
1497 if (ret)
1498 goto error;
1499 }
1500 }
1501
1502 return 0;
1503error:
1504 pr_warn("Failed to add %s, err=%d\n", rft->name, ret);
1505 while (--rft >= rfts) {
1506 if ((fflags & rft->fflags) == rft->fflags)
1507 kernfs_remove_by_name(kn, rft->name);
1508 }
1509 return ret;
1510}
1511
1512/**
1513 * rdtgroup_kn_mode_restrict - Restrict user access to named resctrl file
1514 * @r: The resource group with which the file is associated.
1515 * @name: Name of the file
1516 *
1517 * The permissions of named resctrl file, directory, or link are modified
1518 * to not allow read, write, or execute by any user.
1519 *
1520 * WARNING: This function is intended to communicate to the user that the
1521 * resctrl file has been locked down - that it is not relevant to the
1522 * particular state the system finds itself in. It should not be relied
1523 * on to protect from user access because after the file's permissions
1524 * are restricted the user can still change the permissions using chmod
1525 * from the command line.
1526 *
1527 * Return: 0 on success, <0 on failure.
1528 */
1529int rdtgroup_kn_mode_restrict(struct rdtgroup *r, const char *name)
1530{
1531 struct iattr iattr = {.ia_valid = ATTR_MODE,};
1532 struct kernfs_node *kn;
1533 int ret = 0;
1534
1535 kn = kernfs_find_and_get_ns(r->kn, name, NULL);
1536 if (!kn)
1537 return -ENOENT;
1538
1539 switch (kernfs_type(kn)) {
1540 case KERNFS_DIR:
1541 iattr.ia_mode = S_IFDIR;
1542 break;
1543 case KERNFS_FILE:
1544 iattr.ia_mode = S_IFREG;
1545 break;
1546 case KERNFS_LINK:
1547 iattr.ia_mode = S_IFLNK;
1548 break;
1549 }
1550
1551 ret = kernfs_setattr(kn, &iattr);
1552 kernfs_put(kn);
1553 return ret;
1554}
1555
1556/**
1557 * rdtgroup_kn_mode_restore - Restore user access to named resctrl file
1558 * @r: The resource group with which the file is associated.
1559 * @name: Name of the file
1560 * @mask: Mask of permissions that should be restored
1561 *
1562 * Restore the permissions of the named file. If @name is a directory the
1563 * permissions of its parent will be used.
1564 *
1565 * Return: 0 on success, <0 on failure.
1566 */
1567int rdtgroup_kn_mode_restore(struct rdtgroup *r, const char *name,
1568 umode_t mask)
1569{
1570 struct iattr iattr = {.ia_valid = ATTR_MODE,};
1571 struct kernfs_node *kn, *parent;
1572 struct rftype *rfts, *rft;
1573 int ret, len;
1574
1575 rfts = res_common_files;
1576 len = ARRAY_SIZE(res_common_files);
1577
1578 for (rft = rfts; rft < rfts + len; rft++) {
1579 if (!strcmp(rft->name, name))
1580 iattr.ia_mode = rft->mode & mask;
1581 }
1582
1583 kn = kernfs_find_and_get_ns(r->kn, name, NULL);
1584 if (!kn)
1585 return -ENOENT;
1586
1587 switch (kernfs_type(kn)) {
1588 case KERNFS_DIR:
1589 parent = kernfs_get_parent(kn);
1590 if (parent) {
1591 iattr.ia_mode |= parent->mode;
1592 kernfs_put(parent);
1593 }
1594 iattr.ia_mode |= S_IFDIR;
1595 break;
1596 case KERNFS_FILE:
1597 iattr.ia_mode |= S_IFREG;
1598 break;
1599 case KERNFS_LINK:
1600 iattr.ia_mode |= S_IFLNK;
1601 break;
1602 }
1603
1604 ret = kernfs_setattr(kn, &iattr);
1605 kernfs_put(kn);
1606 return ret;
1607}
1608
1609static int rdtgroup_mkdir_info_resdir(struct rdt_resource *r, char *name,
1610 unsigned long fflags)
1611{
1612 struct kernfs_node *kn_subdir;
1613 int ret;
1614
1615 kn_subdir = kernfs_create_dir(kn_info, name,
1616 kn_info->mode, r);
1617 if (IS_ERR(kn_subdir))
1618 return PTR_ERR(kn_subdir);
1619
1620 kernfs_get(kn_subdir);
1621 ret = rdtgroup_kn_set_ugid(kn_subdir);
1622 if (ret)
1623 return ret;
1624
1625 ret = rdtgroup_add_files(kn_subdir, fflags);
1626 if (!ret)
1627 kernfs_activate(kn_subdir);
1628
1629 return ret;
1630}
1631
1632static int rdtgroup_create_info_dir(struct kernfs_node *parent_kn)
1633{
1634 struct rdt_resource *r;
1635 unsigned long fflags;
1636 char name[32];
1637 int ret;
1638
1639 /* create the directory */
1640 kn_info = kernfs_create_dir(parent_kn, "info", parent_kn->mode, NULL);
1641 if (IS_ERR(kn_info))
1642 return PTR_ERR(kn_info);
1643 kernfs_get(kn_info);
1644
1645 ret = rdtgroup_add_files(kn_info, RF_TOP_INFO);
1646 if (ret)
1647 goto out_destroy;
1648
1649 for_each_alloc_enabled_rdt_resource(r) {
1650 fflags = r->fflags | RF_CTRL_INFO;
1651 ret = rdtgroup_mkdir_info_resdir(r, r->name, fflags);
1652 if (ret)
1653 goto out_destroy;
1654 }
1655
1656 for_each_mon_enabled_rdt_resource(r) {
1657 fflags = r->fflags | RF_MON_INFO;
1658 sprintf(name, "%s_MON", r->name);
1659 ret = rdtgroup_mkdir_info_resdir(r, name, fflags);
1660 if (ret)
1661 goto out_destroy;
1662 }
1663
1664 /*
1665 * This extra ref will be put in kernfs_remove() and guarantees
1666 * that @rdtgrp->kn is always accessible.
1667 */
1668 kernfs_get(kn_info);
1669
1670 ret = rdtgroup_kn_set_ugid(kn_info);
1671 if (ret)
1672 goto out_destroy;
1673
1674 kernfs_activate(kn_info);
1675
1676 return 0;
1677
1678out_destroy:
1679 kernfs_remove(kn_info);
1680 return ret;
1681}
1682
1683static int
1684mongroup_create_dir(struct kernfs_node *parent_kn, struct rdtgroup *prgrp,
1685 char *name, struct kernfs_node **dest_kn)
1686{
1687 struct kernfs_node *kn;
1688 int ret;
1689
1690 /* create the directory */
1691 kn = kernfs_create_dir(parent_kn, name, parent_kn->mode, prgrp);
1692 if (IS_ERR(kn))
1693 return PTR_ERR(kn);
1694
1695 if (dest_kn)
1696 *dest_kn = kn;
1697
1698 /*
1699 * This extra ref will be put in kernfs_remove() and guarantees
1700 * that @rdtgrp->kn is always accessible.
1701 */
1702 kernfs_get(kn);
1703
1704 ret = rdtgroup_kn_set_ugid(kn);
1705 if (ret)
1706 goto out_destroy;
1707
1708 kernfs_activate(kn);
1709
1710 return 0;
1711
1712out_destroy:
1713 kernfs_remove(kn);
1714 return ret;
1715}
1716
1717static void l3_qos_cfg_update(void *arg)
1718{
1719 bool *enable = arg;
1720
1721 wrmsrl(MSR_IA32_L3_QOS_CFG, *enable ? L3_QOS_CDP_ENABLE : 0ULL);
1722}
1723
1724static void l2_qos_cfg_update(void *arg)
1725{
1726 bool *enable = arg;
1727
1728 wrmsrl(MSR_IA32_L2_QOS_CFG, *enable ? L2_QOS_CDP_ENABLE : 0ULL);
1729}
1730
1731static inline bool is_mba_linear(void)
1732{
1733 return rdt_resources_all[RDT_RESOURCE_MBA].membw.delay_linear;
1734}
1735
1736static int set_cache_qos_cfg(int level, bool enable)
1737{
1738 void (*update)(void *arg);
1739 struct rdt_resource *r_l;
1740 cpumask_var_t cpu_mask;
1741 struct rdt_domain *d;
1742 int cpu;
1743
1744 if (!zalloc_cpumask_var(&cpu_mask, GFP_KERNEL))
1745 return -ENOMEM;
1746
1747 if (level == RDT_RESOURCE_L3)
1748 update = l3_qos_cfg_update;
1749 else if (level == RDT_RESOURCE_L2)
1750 update = l2_qos_cfg_update;
1751 else
1752 return -EINVAL;
1753
1754 r_l = &rdt_resources_all[level];
1755 list_for_each_entry(d, &r_l->domains, list) {
1756 /* Pick one CPU from each domain instance to update MSR */
1757 cpumask_set_cpu(cpumask_any(&d->cpu_mask), cpu_mask);
1758 }
1759 cpu = get_cpu();
1760 /* Update QOS_CFG MSR on this cpu if it's in cpu_mask. */
1761 if (cpumask_test_cpu(cpu, cpu_mask))
1762 update(&enable);
1763 /* Update QOS_CFG MSR on all other cpus in cpu_mask. */
1764 smp_call_function_many(cpu_mask, update, &enable, 1);
1765 put_cpu();
1766
1767 free_cpumask_var(cpu_mask);
1768
1769 return 0;
1770}
1771
1772/*
1773 * Enable or disable the MBA software controller
1774 * which helps user specify bandwidth in MBps.
1775 * MBA software controller is supported only if
1776 * MBM is supported and MBA is in linear scale.
1777 */
1778static int set_mba_sc(bool mba_sc)
1779{
1780 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_MBA];
1781 struct rdt_domain *d;
1782
1783 if (!is_mbm_enabled() || !is_mba_linear() ||
1784 mba_sc == is_mba_sc(r))
1785 return -EINVAL;
1786
1787 r->membw.mba_sc = mba_sc;
1788 list_for_each_entry(d, &r->domains, list)
1789 setup_default_ctrlval(r, d->ctrl_val, d->mbps_val);
1790
1791 return 0;
1792}
1793
1794static int cdp_enable(int level, int data_type, int code_type)
1795{
1796 struct rdt_resource *r_ldata = &rdt_resources_all[data_type];
1797 struct rdt_resource *r_lcode = &rdt_resources_all[code_type];
1798 struct rdt_resource *r_l = &rdt_resources_all[level];
1799 int ret;
1800
1801 if (!r_l->alloc_capable || !r_ldata->alloc_capable ||
1802 !r_lcode->alloc_capable)
1803 return -EINVAL;
1804
1805 ret = set_cache_qos_cfg(level, true);
1806 if (!ret) {
1807 r_l->alloc_enabled = false;
1808 r_ldata->alloc_enabled = true;
1809 r_lcode->alloc_enabled = true;
1810 }
1811 return ret;
1812}
1813
1814static int cdpl3_enable(void)
1815{
1816 return cdp_enable(RDT_RESOURCE_L3, RDT_RESOURCE_L3DATA,
1817 RDT_RESOURCE_L3CODE);
1818}
1819
1820static int cdpl2_enable(void)
1821{
1822 return cdp_enable(RDT_RESOURCE_L2, RDT_RESOURCE_L2DATA,
1823 RDT_RESOURCE_L2CODE);
1824}
1825
1826static void cdp_disable(int level, int data_type, int code_type)
1827{
1828 struct rdt_resource *r = &rdt_resources_all[level];
1829
1830 r->alloc_enabled = r->alloc_capable;
1831
1832 if (rdt_resources_all[data_type].alloc_enabled) {
1833 rdt_resources_all[data_type].alloc_enabled = false;
1834 rdt_resources_all[code_type].alloc_enabled = false;
1835 set_cache_qos_cfg(level, false);
1836 }
1837}
1838
1839static void cdpl3_disable(void)
1840{
1841 cdp_disable(RDT_RESOURCE_L3, RDT_RESOURCE_L3DATA, RDT_RESOURCE_L3CODE);
1842}
1843
1844static void cdpl2_disable(void)
1845{
1846 cdp_disable(RDT_RESOURCE_L2, RDT_RESOURCE_L2DATA, RDT_RESOURCE_L2CODE);
1847}
1848
1849static void cdp_disable_all(void)
1850{
1851 if (rdt_resources_all[RDT_RESOURCE_L3DATA].alloc_enabled)
1852 cdpl3_disable();
1853 if (rdt_resources_all[RDT_RESOURCE_L2DATA].alloc_enabled)
1854 cdpl2_disable();
1855}
1856
1857/*
1858 * We don't allow rdtgroup directories to be created anywhere
1859 * except the root directory. Thus when looking for the rdtgroup
1860 * structure for a kernfs node we are either looking at a directory,
1861 * in which case the rdtgroup structure is pointed at by the "priv"
1862 * field, otherwise we have a file, and need only look to the parent
1863 * to find the rdtgroup.
1864 */
1865static struct rdtgroup *kernfs_to_rdtgroup(struct kernfs_node *kn)
1866{
1867 if (kernfs_type(kn) == KERNFS_DIR) {
1868 /*
1869 * All the resource directories use "kn->priv"
1870 * to point to the "struct rdtgroup" for the
1871 * resource. "info" and its subdirectories don't
1872 * have rdtgroup structures, so return NULL here.
1873 */
1874 if (kn == kn_info || kn->parent == kn_info)
1875 return NULL;
1876 else
1877 return kn->priv;
1878 } else {
1879 return kn->parent->priv;
1880 }
1881}
1882
1883struct rdtgroup *rdtgroup_kn_lock_live(struct kernfs_node *kn)
1884{
1885 struct rdtgroup *rdtgrp = kernfs_to_rdtgroup(kn);
1886
1887 if (!rdtgrp)
1888 return NULL;
1889
1890 atomic_inc(&rdtgrp->waitcount);
1891 kernfs_break_active_protection(kn);
1892
1893 mutex_lock(&rdtgroup_mutex);
1894
1895 /* Was this group deleted while we waited? */
1896 if (rdtgrp->flags & RDT_DELETED)
1897 return NULL;
1898
1899 return rdtgrp;
1900}
1901
1902void rdtgroup_kn_unlock(struct kernfs_node *kn)
1903{
1904 struct rdtgroup *rdtgrp = kernfs_to_rdtgroup(kn);
1905
1906 if (!rdtgrp)
1907 return;
1908
1909 mutex_unlock(&rdtgroup_mutex);
1910
1911 if (atomic_dec_and_test(&rdtgrp->waitcount) &&
1912 (rdtgrp->flags & RDT_DELETED)) {
1913 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP ||
1914 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED)
1915 rdtgroup_pseudo_lock_remove(rdtgrp);
1916 kernfs_unbreak_active_protection(kn);
1917 kernfs_put(rdtgrp->kn);
1918 kfree(rdtgrp);
1919 } else {
1920 kernfs_unbreak_active_protection(kn);
1921 }
1922}
1923
1924static int mkdir_mondata_all(struct kernfs_node *parent_kn,
1925 struct rdtgroup *prgrp,
1926 struct kernfs_node **mon_data_kn);
1927
1928static int rdt_enable_ctx(struct rdt_fs_context *ctx)
1929{
1930 int ret = 0;
1931
1932 if (ctx->enable_cdpl2)
1933 ret = cdpl2_enable();
1934
1935 if (!ret && ctx->enable_cdpl3)
1936 ret = cdpl3_enable();
1937
1938 if (!ret && ctx->enable_mba_mbps)
1939 ret = set_mba_sc(true);
1940
1941 return ret;
1942}
1943
1944static int rdt_get_tree(struct fs_context *fc)
1945{
1946 struct rdt_fs_context *ctx = rdt_fc2context(fc);
1947 struct rdt_domain *dom;
1948 struct rdt_resource *r;
1949 int ret;
1950
1951 cpus_read_lock();
1952 mutex_lock(&rdtgroup_mutex);
1953 /*
1954 * resctrl file system can only be mounted once.
1955 */
1956 if (static_branch_unlikely(&rdt_enable_key)) {
1957 ret = -EBUSY;
1958 goto out;
1959 }
1960
1961 ret = rdt_enable_ctx(ctx);
1962 if (ret < 0)
1963 goto out_cdp;
1964
1965 closid_init();
1966
1967 ret = rdtgroup_create_info_dir(rdtgroup_default.kn);
1968 if (ret < 0)
1969 goto out_mba;
1970
1971 if (rdt_mon_capable) {
1972 ret = mongroup_create_dir(rdtgroup_default.kn,
1973 NULL, "mon_groups",
1974 &kn_mongrp);
1975 if (ret < 0)
1976 goto out_info;
1977 kernfs_get(kn_mongrp);
1978
1979 ret = mkdir_mondata_all(rdtgroup_default.kn,
1980 &rdtgroup_default, &kn_mondata);
1981 if (ret < 0)
1982 goto out_mongrp;
1983 kernfs_get(kn_mondata);
1984 rdtgroup_default.mon.mon_data_kn = kn_mondata;
1985 }
1986
1987 ret = rdt_pseudo_lock_init();
1988 if (ret)
1989 goto out_mondata;
1990
1991 ret = kernfs_get_tree(fc);
1992 if (ret < 0)
1993 goto out_psl;
1994
1995 if (rdt_alloc_capable)
1996 static_branch_enable_cpuslocked(&rdt_alloc_enable_key);
1997 if (rdt_mon_capable)
1998 static_branch_enable_cpuslocked(&rdt_mon_enable_key);
1999
2000 if (rdt_alloc_capable || rdt_mon_capable)
2001 static_branch_enable_cpuslocked(&rdt_enable_key);
2002
2003 if (is_mbm_enabled()) {
2004 r = &rdt_resources_all[RDT_RESOURCE_L3];
2005 list_for_each_entry(dom, &r->domains, list)
2006 mbm_setup_overflow_handler(dom, MBM_OVERFLOW_INTERVAL);
2007 }
2008
2009 goto out;
2010
2011out_psl:
2012 rdt_pseudo_lock_release();
2013out_mondata:
2014 if (rdt_mon_capable)
2015 kernfs_remove(kn_mondata);
2016out_mongrp:
2017 if (rdt_mon_capable)
2018 kernfs_remove(kn_mongrp);
2019out_info:
2020 kernfs_remove(kn_info);
2021out_mba:
2022 if (ctx->enable_mba_mbps)
2023 set_mba_sc(false);
2024out_cdp:
2025 cdp_disable_all();
2026out:
2027 rdt_last_cmd_clear();
2028 mutex_unlock(&rdtgroup_mutex);
2029 cpus_read_unlock();
2030 return ret;
2031}
2032
2033enum rdt_param {
2034 Opt_cdp,
2035 Opt_cdpl2,
2036 Opt_mba_mbps,
2037 nr__rdt_params
2038};
2039
2040static const struct fs_parameter_spec rdt_param_specs[] = {
2041 fsparam_flag("cdp", Opt_cdp),
2042 fsparam_flag("cdpl2", Opt_cdpl2),
2043 fsparam_flag("mba_MBps", Opt_mba_mbps),
2044 {}
2045};
2046
2047static const struct fs_parameter_description rdt_fs_parameters = {
2048 .name = "rdt",
2049 .specs = rdt_param_specs,
2050};
2051
2052static int rdt_parse_param(struct fs_context *fc, struct fs_parameter *param)
2053{
2054 struct rdt_fs_context *ctx = rdt_fc2context(fc);
2055 struct fs_parse_result result;
2056 int opt;
2057
2058 opt = fs_parse(fc, &rdt_fs_parameters, param, &result);
2059 if (opt < 0)
2060 return opt;
2061
2062 switch (opt) {
2063 case Opt_cdp:
2064 ctx->enable_cdpl3 = true;
2065 return 0;
2066 case Opt_cdpl2:
2067 ctx->enable_cdpl2 = true;
2068 return 0;
2069 case Opt_mba_mbps:
2070 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
2071 return -EINVAL;
2072 ctx->enable_mba_mbps = true;
2073 return 0;
2074 }
2075
2076 return -EINVAL;
2077}
2078
2079static void rdt_fs_context_free(struct fs_context *fc)
2080{
2081 struct rdt_fs_context *ctx = rdt_fc2context(fc);
2082
2083 kernfs_free_fs_context(fc);
2084 kfree(ctx);
2085}
2086
2087static const struct fs_context_operations rdt_fs_context_ops = {
2088 .free = rdt_fs_context_free,
2089 .parse_param = rdt_parse_param,
2090 .get_tree = rdt_get_tree,
2091};
2092
2093static int rdt_init_fs_context(struct fs_context *fc)
2094{
2095 struct rdt_fs_context *ctx;
2096
2097 ctx = kzalloc(sizeof(struct rdt_fs_context), GFP_KERNEL);
2098 if (!ctx)
2099 return -ENOMEM;
2100
2101 ctx->kfc.root = rdt_root;
2102 ctx->kfc.magic = RDTGROUP_SUPER_MAGIC;
2103 fc->fs_private = &ctx->kfc;
2104 fc->ops = &rdt_fs_context_ops;
2105 put_user_ns(fc->user_ns);
2106 fc->user_ns = get_user_ns(&init_user_ns);
2107 fc->global = true;
2108 return 0;
2109}
2110
2111static int reset_all_ctrls(struct rdt_resource *r)
2112{
2113 struct msr_param msr_param;
2114 cpumask_var_t cpu_mask;
2115 struct rdt_domain *d;
2116 int i, cpu;
2117
2118 if (!zalloc_cpumask_var(&cpu_mask, GFP_KERNEL))
2119 return -ENOMEM;
2120
2121 msr_param.res = r;
2122 msr_param.low = 0;
2123 msr_param.high = r->num_closid;
2124
2125 /*
2126 * Disable resource control for this resource by setting all
2127 * CBMs in all domains to the maximum mask value. Pick one CPU
2128 * from each domain to update the MSRs below.
2129 */
2130 list_for_each_entry(d, &r->domains, list) {
2131 cpumask_set_cpu(cpumask_any(&d->cpu_mask), cpu_mask);
2132
2133 for (i = 0; i < r->num_closid; i++)
2134 d->ctrl_val[i] = r->default_ctrl;
2135 }
2136 cpu = get_cpu();
2137 /* Update CBM on this cpu if it's in cpu_mask. */
2138 if (cpumask_test_cpu(cpu, cpu_mask))
2139 rdt_ctrl_update(&msr_param);
2140 /* Update CBM on all other cpus in cpu_mask. */
2141 smp_call_function_many(cpu_mask, rdt_ctrl_update, &msr_param, 1);
2142 put_cpu();
2143
2144 free_cpumask_var(cpu_mask);
2145
2146 return 0;
2147}
2148
2149static bool is_closid_match(struct task_struct *t, struct rdtgroup *r)
2150{
2151 return (rdt_alloc_capable &&
2152 (r->type == RDTCTRL_GROUP) && (t->closid == r->closid));
2153}
2154
2155static bool is_rmid_match(struct task_struct *t, struct rdtgroup *r)
2156{
2157 return (rdt_mon_capable &&
2158 (r->type == RDTMON_GROUP) && (t->rmid == r->mon.rmid));
2159}
2160
2161/*
2162 * Move tasks from one to the other group. If @from is NULL, then all tasks
2163 * in the systems are moved unconditionally (used for teardown).
2164 *
2165 * If @mask is not NULL the cpus on which moved tasks are running are set
2166 * in that mask so the update smp function call is restricted to affected
2167 * cpus.
2168 */
2169static void rdt_move_group_tasks(struct rdtgroup *from, struct rdtgroup *to,
2170 struct cpumask *mask)
2171{
2172 struct task_struct *p, *t;
2173
2174 read_lock(&tasklist_lock);
2175 for_each_process_thread(p, t) {
2176 if (!from || is_closid_match(t, from) ||
2177 is_rmid_match(t, from)) {
2178 t->closid = to->closid;
2179 t->rmid = to->mon.rmid;
2180
2181#ifdef CONFIG_SMP
2182 /*
2183 * This is safe on x86 w/o barriers as the ordering
2184 * of writing to task_cpu() and t->on_cpu is
2185 * reverse to the reading here. The detection is
2186 * inaccurate as tasks might move or schedule
2187 * before the smp function call takes place. In
2188 * such a case the function call is pointless, but
2189 * there is no other side effect.
2190 */
2191 if (mask && t->on_cpu)
2192 cpumask_set_cpu(task_cpu(t), mask);
2193#endif
2194 }
2195 }
2196 read_unlock(&tasklist_lock);
2197}
2198
2199static void free_all_child_rdtgrp(struct rdtgroup *rdtgrp)
2200{
2201 struct rdtgroup *sentry, *stmp;
2202 struct list_head *head;
2203
2204 head = &rdtgrp->mon.crdtgrp_list;
2205 list_for_each_entry_safe(sentry, stmp, head, mon.crdtgrp_list) {
2206 free_rmid(sentry->mon.rmid);
2207 list_del(&sentry->mon.crdtgrp_list);
2208 kfree(sentry);
2209 }
2210}
2211
2212/*
2213 * Forcibly remove all of subdirectories under root.
2214 */
2215static void rmdir_all_sub(void)
2216{
2217 struct rdtgroup *rdtgrp, *tmp;
2218
2219 /* Move all tasks to the default resource group */
2220 rdt_move_group_tasks(NULL, &rdtgroup_default, NULL);
2221
2222 list_for_each_entry_safe(rdtgrp, tmp, &rdt_all_groups, rdtgroup_list) {
2223 /* Free any child rmids */
2224 free_all_child_rdtgrp(rdtgrp);
2225
2226 /* Remove each rdtgroup other than root */
2227 if (rdtgrp == &rdtgroup_default)
2228 continue;
2229
2230 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP ||
2231 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED)
2232 rdtgroup_pseudo_lock_remove(rdtgrp);
2233
2234 /*
2235 * Give any CPUs back to the default group. We cannot copy
2236 * cpu_online_mask because a CPU might have executed the
2237 * offline callback already, but is still marked online.
2238 */
2239 cpumask_or(&rdtgroup_default.cpu_mask,
2240 &rdtgroup_default.cpu_mask, &rdtgrp->cpu_mask);
2241
2242 free_rmid(rdtgrp->mon.rmid);
2243
2244 kernfs_remove(rdtgrp->kn);
2245 list_del(&rdtgrp->rdtgroup_list);
2246 kfree(rdtgrp);
2247 }
2248 /* Notify online CPUs to update per cpu storage and PQR_ASSOC MSR */
2249 update_closid_rmid(cpu_online_mask, &rdtgroup_default);
2250
2251 kernfs_remove(kn_info);
2252 kernfs_remove(kn_mongrp);
2253 kernfs_remove(kn_mondata);
2254}
2255
2256static void rdt_kill_sb(struct super_block *sb)
2257{
2258 struct rdt_resource *r;
2259
2260 cpus_read_lock();
2261 mutex_lock(&rdtgroup_mutex);
2262
2263 set_mba_sc(false);
2264
2265 /*Put everything back to default values. */
2266 for_each_alloc_enabled_rdt_resource(r)
2267 reset_all_ctrls(r);
2268 cdp_disable_all();
2269 rmdir_all_sub();
2270 rdt_pseudo_lock_release();
2271 rdtgroup_default.mode = RDT_MODE_SHAREABLE;
2272 static_branch_disable_cpuslocked(&rdt_alloc_enable_key);
2273 static_branch_disable_cpuslocked(&rdt_mon_enable_key);
2274 static_branch_disable_cpuslocked(&rdt_enable_key);
2275 kernfs_kill_sb(sb);
2276 mutex_unlock(&rdtgroup_mutex);
2277 cpus_read_unlock();
2278}
2279
2280static struct file_system_type rdt_fs_type = {
2281 .name = "resctrl",
2282 .init_fs_context = rdt_init_fs_context,
2283 .parameters = &rdt_fs_parameters,
2284 .kill_sb = rdt_kill_sb,
2285};
2286
2287static int mon_addfile(struct kernfs_node *parent_kn, const char *name,
2288 void *priv)
2289{
2290 struct kernfs_node *kn;
2291 int ret = 0;
2292
2293 kn = __kernfs_create_file(parent_kn, name, 0444,
2294 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID, 0,
2295 &kf_mondata_ops, priv, NULL, NULL);
2296 if (IS_ERR(kn))
2297 return PTR_ERR(kn);
2298
2299 ret = rdtgroup_kn_set_ugid(kn);
2300 if (ret) {
2301 kernfs_remove(kn);
2302 return ret;
2303 }
2304
2305 return ret;
2306}
2307
2308/*
2309 * Remove all subdirectories of mon_data of ctrl_mon groups
2310 * and monitor groups with given domain id.
2311 */
2312void rmdir_mondata_subdir_allrdtgrp(struct rdt_resource *r, unsigned int dom_id)
2313{
2314 struct rdtgroup *prgrp, *crgrp;
2315 char name[32];
2316
2317 if (!r->mon_enabled)
2318 return;
2319
2320 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
2321 sprintf(name, "mon_%s_%02d", r->name, dom_id);
2322 kernfs_remove_by_name(prgrp->mon.mon_data_kn, name);
2323
2324 list_for_each_entry(crgrp, &prgrp->mon.crdtgrp_list, mon.crdtgrp_list)
2325 kernfs_remove_by_name(crgrp->mon.mon_data_kn, name);
2326 }
2327}
2328
2329static int mkdir_mondata_subdir(struct kernfs_node *parent_kn,
2330 struct rdt_domain *d,
2331 struct rdt_resource *r, struct rdtgroup *prgrp)
2332{
2333 union mon_data_bits priv;
2334 struct kernfs_node *kn;
2335 struct mon_evt *mevt;
2336 struct rmid_read rr;
2337 char name[32];
2338 int ret;
2339
2340 sprintf(name, "mon_%s_%02d", r->name, d->id);
2341 /* create the directory */
2342 kn = kernfs_create_dir(parent_kn, name, parent_kn->mode, prgrp);
2343 if (IS_ERR(kn))
2344 return PTR_ERR(kn);
2345
2346 /*
2347 * This extra ref will be put in kernfs_remove() and guarantees
2348 * that kn is always accessible.
2349 */
2350 kernfs_get(kn);
2351 ret = rdtgroup_kn_set_ugid(kn);
2352 if (ret)
2353 goto out_destroy;
2354
2355 if (WARN_ON(list_empty(&r->evt_list))) {
2356 ret = -EPERM;
2357 goto out_destroy;
2358 }
2359
2360 priv.u.rid = r->rid;
2361 priv.u.domid = d->id;
2362 list_for_each_entry(mevt, &r->evt_list, list) {
2363 priv.u.evtid = mevt->evtid;
2364 ret = mon_addfile(kn, mevt->name, priv.priv);
2365 if (ret)
2366 goto out_destroy;
2367
2368 if (is_mbm_event(mevt->evtid))
2369 mon_event_read(&rr, d, prgrp, mevt->evtid, true);
2370 }
2371 kernfs_activate(kn);
2372 return 0;
2373
2374out_destroy:
2375 kernfs_remove(kn);
2376 return ret;
2377}
2378
2379/*
2380 * Add all subdirectories of mon_data for "ctrl_mon" groups
2381 * and "monitor" groups with given domain id.
2382 */
2383void mkdir_mondata_subdir_allrdtgrp(struct rdt_resource *r,
2384 struct rdt_domain *d)
2385{
2386 struct kernfs_node *parent_kn;
2387 struct rdtgroup *prgrp, *crgrp;
2388 struct list_head *head;
2389
2390 if (!r->mon_enabled)
2391 return;
2392
2393 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
2394 parent_kn = prgrp->mon.mon_data_kn;
2395 mkdir_mondata_subdir(parent_kn, d, r, prgrp);
2396
2397 head = &prgrp->mon.crdtgrp_list;
2398 list_for_each_entry(crgrp, head, mon.crdtgrp_list) {
2399 parent_kn = crgrp->mon.mon_data_kn;
2400 mkdir_mondata_subdir(parent_kn, d, r, crgrp);
2401 }
2402 }
2403}
2404
2405static int mkdir_mondata_subdir_alldom(struct kernfs_node *parent_kn,
2406 struct rdt_resource *r,
2407 struct rdtgroup *prgrp)
2408{
2409 struct rdt_domain *dom;
2410 int ret;
2411
2412 list_for_each_entry(dom, &r->domains, list) {
2413 ret = mkdir_mondata_subdir(parent_kn, dom, r, prgrp);
2414 if (ret)
2415 return ret;
2416 }
2417
2418 return 0;
2419}
2420
2421/*
2422 * This creates a directory mon_data which contains the monitored data.
2423 *
2424 * mon_data has one directory for each domain whic are named
2425 * in the format mon_<domain_name>_<domain_id>. For ex: A mon_data
2426 * with L3 domain looks as below:
2427 * ./mon_data:
2428 * mon_L3_00
2429 * mon_L3_01
2430 * mon_L3_02
2431 * ...
2432 *
2433 * Each domain directory has one file per event:
2434 * ./mon_L3_00/:
2435 * llc_occupancy
2436 *
2437 */
2438static int mkdir_mondata_all(struct kernfs_node *parent_kn,
2439 struct rdtgroup *prgrp,
2440 struct kernfs_node **dest_kn)
2441{
2442 struct rdt_resource *r;
2443 struct kernfs_node *kn;
2444 int ret;
2445
2446 /*
2447 * Create the mon_data directory first.
2448 */
2449 ret = mongroup_create_dir(parent_kn, NULL, "mon_data", &kn);
2450 if (ret)
2451 return ret;
2452
2453 if (dest_kn)
2454 *dest_kn = kn;
2455
2456 /*
2457 * Create the subdirectories for each domain. Note that all events
2458 * in a domain like L3 are grouped into a resource whose domain is L3
2459 */
2460 for_each_mon_enabled_rdt_resource(r) {
2461 ret = mkdir_mondata_subdir_alldom(kn, r, prgrp);
2462 if (ret)
2463 goto out_destroy;
2464 }
2465
2466 return 0;
2467
2468out_destroy:
2469 kernfs_remove(kn);
2470 return ret;
2471}
2472
2473/**
2474 * cbm_ensure_valid - Enforce validity on provided CBM
2475 * @_val: Candidate CBM
2476 * @r: RDT resource to which the CBM belongs
2477 *
2478 * The provided CBM represents all cache portions available for use. This
2479 * may be represented by a bitmap that does not consist of contiguous ones
2480 * and thus be an invalid CBM.
2481 * Here the provided CBM is forced to be a valid CBM by only considering
2482 * the first set of contiguous bits as valid and clearing all bits.
2483 * The intention here is to provide a valid default CBM with which a new
2484 * resource group is initialized. The user can follow this with a
2485 * modification to the CBM if the default does not satisfy the
2486 * requirements.
2487 */
2488static u32 cbm_ensure_valid(u32 _val, struct rdt_resource *r)
2489{
2490 unsigned int cbm_len = r->cache.cbm_len;
2491 unsigned long first_bit, zero_bit;
2492 unsigned long val = _val;
2493
2494 if (!val)
2495 return 0;
2496
2497 first_bit = find_first_bit(&val, cbm_len);
2498 zero_bit = find_next_zero_bit(&val, cbm_len, first_bit);
2499
2500 /* Clear any remaining bits to ensure contiguous region */
2501 bitmap_clear(&val, zero_bit, cbm_len - zero_bit);
2502 return (u32)val;
2503}
2504
2505/*
2506 * Initialize cache resources per RDT domain
2507 *
2508 * Set the RDT domain up to start off with all usable allocations. That is,
2509 * all shareable and unused bits. All-zero CBM is invalid.
2510 */
2511static int __init_one_rdt_domain(struct rdt_domain *d, struct rdt_resource *r,
2512 u32 closid)
2513{
2514 struct rdt_resource *r_cdp = NULL;
2515 struct rdt_domain *d_cdp = NULL;
2516 u32 used_b = 0, unused_b = 0;
2517 unsigned long tmp_cbm;
2518 enum rdtgrp_mode mode;
2519 u32 peer_ctl, *ctrl;
2520 int i;
2521
2522 rdt_cdp_peer_get(r, d, &r_cdp, &d_cdp);
2523 d->have_new_ctrl = false;
2524 d->new_ctrl = r->cache.shareable_bits;
2525 used_b = r->cache.shareable_bits;
2526 ctrl = d->ctrl_val;
2527 for (i = 0; i < closids_supported(); i++, ctrl++) {
2528 if (closid_allocated(i) && i != closid) {
2529 mode = rdtgroup_mode_by_closid(i);
2530 if (mode == RDT_MODE_PSEUDO_LOCKSETUP)
2531 /*
2532 * ctrl values for locksetup aren't relevant
2533 * until the schemata is written, and the mode
2534 * becomes RDT_MODE_PSEUDO_LOCKED.
2535 */
2536 continue;
2537 /*
2538 * If CDP is active include peer domain's
2539 * usage to ensure there is no overlap
2540 * with an exclusive group.
2541 */
2542 if (d_cdp)
2543 peer_ctl = d_cdp->ctrl_val[i];
2544 else
2545 peer_ctl = 0;
2546 used_b |= *ctrl | peer_ctl;
2547 if (mode == RDT_MODE_SHAREABLE)
2548 d->new_ctrl |= *ctrl | peer_ctl;
2549 }
2550 }
2551 if (d->plr && d->plr->cbm > 0)
2552 used_b |= d->plr->cbm;
2553 unused_b = used_b ^ (BIT_MASK(r->cache.cbm_len) - 1);
2554 unused_b &= BIT_MASK(r->cache.cbm_len) - 1;
2555 d->new_ctrl |= unused_b;
2556 /*
2557 * Force the initial CBM to be valid, user can
2558 * modify the CBM based on system availability.
2559 */
2560 d->new_ctrl = cbm_ensure_valid(d->new_ctrl, r);
2561 /*
2562 * Assign the u32 CBM to an unsigned long to ensure that
2563 * bitmap_weight() does not access out-of-bound memory.
2564 */
2565 tmp_cbm = d->new_ctrl;
2566 if (bitmap_weight(&tmp_cbm, r->cache.cbm_len) < r->cache.min_cbm_bits) {
2567 rdt_last_cmd_printf("No space on %s:%d\n", r->name, d->id);
2568 return -ENOSPC;
2569 }
2570 d->have_new_ctrl = true;
2571
2572 return 0;
2573}
2574
2575/*
2576 * Initialize cache resources with default values.
2577 *
2578 * A new RDT group is being created on an allocation capable (CAT)
2579 * supporting system. Set this group up to start off with all usable
2580 * allocations.
2581 *
2582 * If there are no more shareable bits available on any domain then
2583 * the entire allocation will fail.
2584 */
2585static int rdtgroup_init_cat(struct rdt_resource *r, u32 closid)
2586{
2587 struct rdt_domain *d;
2588 int ret;
2589
2590 list_for_each_entry(d, &r->domains, list) {
2591 ret = __init_one_rdt_domain(d, r, closid);
2592 if (ret < 0)
2593 return ret;
2594 }
2595
2596 return 0;
2597}
2598
2599/* Initialize MBA resource with default values. */
2600static void rdtgroup_init_mba(struct rdt_resource *r)
2601{
2602 struct rdt_domain *d;
2603
2604 list_for_each_entry(d, &r->domains, list) {
2605 d->new_ctrl = is_mba_sc(r) ? MBA_MAX_MBPS : r->default_ctrl;
2606 d->have_new_ctrl = true;
2607 }
2608}
2609
2610/* Initialize the RDT group's allocations. */
2611static int rdtgroup_init_alloc(struct rdtgroup *rdtgrp)
2612{
2613 struct rdt_resource *r;
2614 int ret;
2615
2616 for_each_alloc_enabled_rdt_resource(r) {
2617 if (r->rid == RDT_RESOURCE_MBA) {
2618 rdtgroup_init_mba(r);
2619 } else {
2620 ret = rdtgroup_init_cat(r, rdtgrp->closid);
2621 if (ret < 0)
2622 return ret;
2623 }
2624
2625 ret = update_domains(r, rdtgrp->closid);
2626 if (ret < 0) {
2627 rdt_last_cmd_puts("Failed to initialize allocations\n");
2628 return ret;
2629 }
2630
2631 }
2632
2633 rdtgrp->mode = RDT_MODE_SHAREABLE;
2634
2635 return 0;
2636}
2637
2638static int mkdir_rdt_prepare(struct kernfs_node *parent_kn,
2639 struct kernfs_node *prgrp_kn,
2640 const char *name, umode_t mode,
2641 enum rdt_group_type rtype, struct rdtgroup **r)
2642{
2643 struct rdtgroup *prdtgrp, *rdtgrp;
2644 struct kernfs_node *kn;
2645 uint files = 0;
2646 int ret;
2647
2648 prdtgrp = rdtgroup_kn_lock_live(prgrp_kn);
2649 if (!prdtgrp) {
2650 ret = -ENODEV;
2651 goto out_unlock;
2652 }
2653
2654 if (rtype == RDTMON_GROUP &&
2655 (prdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP ||
2656 prdtgrp->mode == RDT_MODE_PSEUDO_LOCKED)) {
2657 ret = -EINVAL;
2658 rdt_last_cmd_puts("Pseudo-locking in progress\n");
2659 goto out_unlock;
2660 }
2661
2662 /* allocate the rdtgroup. */
2663 rdtgrp = kzalloc(sizeof(*rdtgrp), GFP_KERNEL);
2664 if (!rdtgrp) {
2665 ret = -ENOSPC;
2666 rdt_last_cmd_puts("Kernel out of memory\n");
2667 goto out_unlock;
2668 }
2669 *r = rdtgrp;
2670 rdtgrp->mon.parent = prdtgrp;
2671 rdtgrp->type = rtype;
2672 INIT_LIST_HEAD(&rdtgrp->mon.crdtgrp_list);
2673
2674 /* kernfs creates the directory for rdtgrp */
2675 kn = kernfs_create_dir(parent_kn, name, mode, rdtgrp);
2676 if (IS_ERR(kn)) {
2677 ret = PTR_ERR(kn);
2678 rdt_last_cmd_puts("kernfs create error\n");
2679 goto out_free_rgrp;
2680 }
2681 rdtgrp->kn = kn;
2682
2683 /*
2684 * kernfs_remove() will drop the reference count on "kn" which
2685 * will free it. But we still need it to stick around for the
2686 * rdtgroup_kn_unlock(kn} call below. Take one extra reference
2687 * here, which will be dropped inside rdtgroup_kn_unlock().
2688 */
2689 kernfs_get(kn);
2690
2691 ret = rdtgroup_kn_set_ugid(kn);
2692 if (ret) {
2693 rdt_last_cmd_puts("kernfs perm error\n");
2694 goto out_destroy;
2695 }
2696
2697 files = RFTYPE_BASE | BIT(RF_CTRLSHIFT + rtype);
2698 ret = rdtgroup_add_files(kn, files);
2699 if (ret) {
2700 rdt_last_cmd_puts("kernfs fill error\n");
2701 goto out_destroy;
2702 }
2703
2704 if (rdt_mon_capable) {
2705 ret = alloc_rmid();
2706 if (ret < 0) {
2707 rdt_last_cmd_puts("Out of RMIDs\n");
2708 goto out_destroy;
2709 }
2710 rdtgrp->mon.rmid = ret;
2711
2712 ret = mkdir_mondata_all(kn, rdtgrp, &rdtgrp->mon.mon_data_kn);
2713 if (ret) {
2714 rdt_last_cmd_puts("kernfs subdir error\n");
2715 goto out_idfree;
2716 }
2717 }
2718 kernfs_activate(kn);
2719
2720 /*
2721 * The caller unlocks the prgrp_kn upon success.
2722 */
2723 return 0;
2724
2725out_idfree:
2726 free_rmid(rdtgrp->mon.rmid);
2727out_destroy:
2728 kernfs_remove(rdtgrp->kn);
2729out_free_rgrp:
2730 kfree(rdtgrp);
2731out_unlock:
2732 rdtgroup_kn_unlock(prgrp_kn);
2733 return ret;
2734}
2735
2736static void mkdir_rdt_prepare_clean(struct rdtgroup *rgrp)
2737{
2738 kernfs_remove(rgrp->kn);
2739 free_rmid(rgrp->mon.rmid);
2740 kfree(rgrp);
2741}
2742
2743/*
2744 * Create a monitor group under "mon_groups" directory of a control
2745 * and monitor group(ctrl_mon). This is a resource group
2746 * to monitor a subset of tasks and cpus in its parent ctrl_mon group.
2747 */
2748static int rdtgroup_mkdir_mon(struct kernfs_node *parent_kn,
2749 struct kernfs_node *prgrp_kn,
2750 const char *name,
2751 umode_t mode)
2752{
2753 struct rdtgroup *rdtgrp, *prgrp;
2754 int ret;
2755
2756 ret = mkdir_rdt_prepare(parent_kn, prgrp_kn, name, mode, RDTMON_GROUP,
2757 &rdtgrp);
2758 if (ret)
2759 return ret;
2760
2761 prgrp = rdtgrp->mon.parent;
2762 rdtgrp->closid = prgrp->closid;
2763
2764 /*
2765 * Add the rdtgrp to the list of rdtgrps the parent
2766 * ctrl_mon group has to track.
2767 */
2768 list_add_tail(&rdtgrp->mon.crdtgrp_list, &prgrp->mon.crdtgrp_list);
2769
2770 rdtgroup_kn_unlock(prgrp_kn);
2771 return ret;
2772}
2773
2774/*
2775 * These are rdtgroups created under the root directory. Can be used
2776 * to allocate and monitor resources.
2777 */
2778static int rdtgroup_mkdir_ctrl_mon(struct kernfs_node *parent_kn,
2779 struct kernfs_node *prgrp_kn,
2780 const char *name, umode_t mode)
2781{
2782 struct rdtgroup *rdtgrp;
2783 struct kernfs_node *kn;
2784 u32 closid;
2785 int ret;
2786
2787 ret = mkdir_rdt_prepare(parent_kn, prgrp_kn, name, mode, RDTCTRL_GROUP,
2788 &rdtgrp);
2789 if (ret)
2790 return ret;
2791
2792 kn = rdtgrp->kn;
2793 ret = closid_alloc();
2794 if (ret < 0) {
2795 rdt_last_cmd_puts("Out of CLOSIDs\n");
2796 goto out_common_fail;
2797 }
2798 closid = ret;
2799 ret = 0;
2800
2801 rdtgrp->closid = closid;
2802 ret = rdtgroup_init_alloc(rdtgrp);
2803 if (ret < 0)
2804 goto out_id_free;
2805
2806 list_add(&rdtgrp->rdtgroup_list, &rdt_all_groups);
2807
2808 if (rdt_mon_capable) {
2809 /*
2810 * Create an empty mon_groups directory to hold the subset
2811 * of tasks and cpus to monitor.
2812 */
2813 ret = mongroup_create_dir(kn, NULL, "mon_groups", NULL);
2814 if (ret) {
2815 rdt_last_cmd_puts("kernfs subdir error\n");
2816 goto out_del_list;
2817 }
2818 }
2819
2820 goto out_unlock;
2821
2822out_del_list:
2823 list_del(&rdtgrp->rdtgroup_list);
2824out_id_free:
2825 closid_free(closid);
2826out_common_fail:
2827 mkdir_rdt_prepare_clean(rdtgrp);
2828out_unlock:
2829 rdtgroup_kn_unlock(prgrp_kn);
2830 return ret;
2831}
2832
2833/*
2834 * We allow creating mon groups only with in a directory called "mon_groups"
2835 * which is present in every ctrl_mon group. Check if this is a valid
2836 * "mon_groups" directory.
2837 *
2838 * 1. The directory should be named "mon_groups".
2839 * 2. The mon group itself should "not" be named "mon_groups".
2840 * This makes sure "mon_groups" directory always has a ctrl_mon group
2841 * as parent.
2842 */
2843static bool is_mon_groups(struct kernfs_node *kn, const char *name)
2844{
2845 return (!strcmp(kn->name, "mon_groups") &&
2846 strcmp(name, "mon_groups"));
2847}
2848
2849static int rdtgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
2850 umode_t mode)
2851{
2852 /* Do not accept '\n' to avoid unparsable situation. */
2853 if (strchr(name, '\n'))
2854 return -EINVAL;
2855
2856 /*
2857 * If the parent directory is the root directory and RDT
2858 * allocation is supported, add a control and monitoring
2859 * subdirectory
2860 */
2861 if (rdt_alloc_capable && parent_kn == rdtgroup_default.kn)
2862 return rdtgroup_mkdir_ctrl_mon(parent_kn, parent_kn, name, mode);
2863
2864 /*
2865 * If RDT monitoring is supported and the parent directory is a valid
2866 * "mon_groups" directory, add a monitoring subdirectory.
2867 */
2868 if (rdt_mon_capable && is_mon_groups(parent_kn, name))
2869 return rdtgroup_mkdir_mon(parent_kn, parent_kn->parent, name, mode);
2870
2871 return -EPERM;
2872}
2873
2874static int rdtgroup_rmdir_mon(struct kernfs_node *kn, struct rdtgroup *rdtgrp,
2875 cpumask_var_t tmpmask)
2876{
2877 struct rdtgroup *prdtgrp = rdtgrp->mon.parent;
2878 int cpu;
2879
2880 /* Give any tasks back to the parent group */
2881 rdt_move_group_tasks(rdtgrp, prdtgrp, tmpmask);
2882
2883 /* Update per cpu rmid of the moved CPUs first */
2884 for_each_cpu(cpu, &rdtgrp->cpu_mask)
2885 per_cpu(pqr_state.default_rmid, cpu) = prdtgrp->mon.rmid;
2886 /*
2887 * Update the MSR on moved CPUs and CPUs which have moved
2888 * task running on them.
2889 */
2890 cpumask_or(tmpmask, tmpmask, &rdtgrp->cpu_mask);
2891 update_closid_rmid(tmpmask, NULL);
2892
2893 rdtgrp->flags = RDT_DELETED;
2894 free_rmid(rdtgrp->mon.rmid);
2895
2896 /*
2897 * Remove the rdtgrp from the parent ctrl_mon group's list
2898 */
2899 WARN_ON(list_empty(&prdtgrp->mon.crdtgrp_list));
2900 list_del(&rdtgrp->mon.crdtgrp_list);
2901
2902 /*
2903 * one extra hold on this, will drop when we kfree(rdtgrp)
2904 * in rdtgroup_kn_unlock()
2905 */
2906 kernfs_get(kn);
2907 kernfs_remove(rdtgrp->kn);
2908
2909 return 0;
2910}
2911
2912static int rdtgroup_ctrl_remove(struct kernfs_node *kn,
2913 struct rdtgroup *rdtgrp)
2914{
2915 rdtgrp->flags = RDT_DELETED;
2916 list_del(&rdtgrp->rdtgroup_list);
2917
2918 /*
2919 * one extra hold on this, will drop when we kfree(rdtgrp)
2920 * in rdtgroup_kn_unlock()
2921 */
2922 kernfs_get(kn);
2923 kernfs_remove(rdtgrp->kn);
2924 return 0;
2925}
2926
2927static int rdtgroup_rmdir_ctrl(struct kernfs_node *kn, struct rdtgroup *rdtgrp,
2928 cpumask_var_t tmpmask)
2929{
2930 int cpu;
2931
2932 /* Give any tasks back to the default group */
2933 rdt_move_group_tasks(rdtgrp, &rdtgroup_default, tmpmask);
2934
2935 /* Give any CPUs back to the default group */
2936 cpumask_or(&rdtgroup_default.cpu_mask,
2937 &rdtgroup_default.cpu_mask, &rdtgrp->cpu_mask);
2938
2939 /* Update per cpu closid and rmid of the moved CPUs first */
2940 for_each_cpu(cpu, &rdtgrp->cpu_mask) {
2941 per_cpu(pqr_state.default_closid, cpu) = rdtgroup_default.closid;
2942 per_cpu(pqr_state.default_rmid, cpu) = rdtgroup_default.mon.rmid;
2943 }
2944
2945 /*
2946 * Update the MSR on moved CPUs and CPUs which have moved
2947 * task running on them.
2948 */
2949 cpumask_or(tmpmask, tmpmask, &rdtgrp->cpu_mask);
2950 update_closid_rmid(tmpmask, NULL);
2951
2952 closid_free(rdtgrp->closid);
2953 free_rmid(rdtgrp->mon.rmid);
2954
2955 /*
2956 * Free all the child monitor group rmids.
2957 */
2958 free_all_child_rdtgrp(rdtgrp);
2959
2960 rdtgroup_ctrl_remove(kn, rdtgrp);
2961
2962 return 0;
2963}
2964
2965static int rdtgroup_rmdir(struct kernfs_node *kn)
2966{
2967 struct kernfs_node *parent_kn = kn->parent;
2968 struct rdtgroup *rdtgrp;
2969 cpumask_var_t tmpmask;
2970 int ret = 0;
2971
2972 if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
2973 return -ENOMEM;
2974
2975 rdtgrp = rdtgroup_kn_lock_live(kn);
2976 if (!rdtgrp) {
2977 ret = -EPERM;
2978 goto out;
2979 }
2980
2981 /*
2982 * If the rdtgroup is a ctrl_mon group and parent directory
2983 * is the root directory, remove the ctrl_mon group.
2984 *
2985 * If the rdtgroup is a mon group and parent directory
2986 * is a valid "mon_groups" directory, remove the mon group.
2987 */
2988 if (rdtgrp->type == RDTCTRL_GROUP && parent_kn == rdtgroup_default.kn) {
2989 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP ||
2990 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
2991 ret = rdtgroup_ctrl_remove(kn, rdtgrp);
2992 } else {
2993 ret = rdtgroup_rmdir_ctrl(kn, rdtgrp, tmpmask);
2994 }
2995 } else if (rdtgrp->type == RDTMON_GROUP &&
2996 is_mon_groups(parent_kn, kn->name)) {
2997 ret = rdtgroup_rmdir_mon(kn, rdtgrp, tmpmask);
2998 } else {
2999 ret = -EPERM;
3000 }
3001
3002out:
3003 rdtgroup_kn_unlock(kn);
3004 free_cpumask_var(tmpmask);
3005 return ret;
3006}
3007
3008static int rdtgroup_show_options(struct seq_file *seq, struct kernfs_root *kf)
3009{
3010 if (rdt_resources_all[RDT_RESOURCE_L3DATA].alloc_enabled)
3011 seq_puts(seq, ",cdp");
3012
3013 if (rdt_resources_all[RDT_RESOURCE_L2DATA].alloc_enabled)
3014 seq_puts(seq, ",cdpl2");
3015
3016 if (is_mba_sc(&rdt_resources_all[RDT_RESOURCE_MBA]))
3017 seq_puts(seq, ",mba_MBps");
3018
3019 return 0;
3020}
3021
3022static struct kernfs_syscall_ops rdtgroup_kf_syscall_ops = {
3023 .mkdir = rdtgroup_mkdir,
3024 .rmdir = rdtgroup_rmdir,
3025 .show_options = rdtgroup_show_options,
3026};
3027
3028static int __init rdtgroup_setup_root(void)
3029{
3030 int ret;
3031
3032 rdt_root = kernfs_create_root(&rdtgroup_kf_syscall_ops,
3033 KERNFS_ROOT_CREATE_DEACTIVATED |
3034 KERNFS_ROOT_EXTRA_OPEN_PERM_CHECK,
3035 &rdtgroup_default);
3036 if (IS_ERR(rdt_root))
3037 return PTR_ERR(rdt_root);
3038
3039 mutex_lock(&rdtgroup_mutex);
3040
3041 rdtgroup_default.closid = 0;
3042 rdtgroup_default.mon.rmid = 0;
3043 rdtgroup_default.type = RDTCTRL_GROUP;
3044 INIT_LIST_HEAD(&rdtgroup_default.mon.crdtgrp_list);
3045
3046 list_add(&rdtgroup_default.rdtgroup_list, &rdt_all_groups);
3047
3048 ret = rdtgroup_add_files(rdt_root->kn, RF_CTRL_BASE);
3049 if (ret) {
3050 kernfs_destroy_root(rdt_root);
3051 goto out;
3052 }
3053
3054 rdtgroup_default.kn = rdt_root->kn;
3055 kernfs_activate(rdtgroup_default.kn);
3056
3057out:
3058 mutex_unlock(&rdtgroup_mutex);
3059
3060 return ret;
3061}
3062
3063/*
3064 * rdtgroup_init - rdtgroup initialization
3065 *
3066 * Setup resctrl file system including set up root, create mount point,
3067 * register rdtgroup filesystem, and initialize files under root directory.
3068 *
3069 * Return: 0 on success or -errno
3070 */
3071int __init rdtgroup_init(void)
3072{
3073 int ret = 0;
3074
3075 seq_buf_init(&last_cmd_status, last_cmd_status_buf,
3076 sizeof(last_cmd_status_buf));
3077
3078 ret = rdtgroup_setup_root();
3079 if (ret)
3080 return ret;
3081
3082 ret = sysfs_create_mount_point(fs_kobj, "resctrl");
3083 if (ret)
3084 goto cleanup_root;
3085
3086 ret = register_filesystem(&rdt_fs_type);
3087 if (ret)
3088 goto cleanup_mountpoint;
3089
3090 /*
3091 * Adding the resctrl debugfs directory here may not be ideal since
3092 * it would let the resctrl debugfs directory appear on the debugfs
3093 * filesystem before the resctrl filesystem is mounted.
3094 * It may also be ok since that would enable debugging of RDT before
3095 * resctrl is mounted.
3096 * The reason why the debugfs directory is created here and not in
3097 * rdt_mount() is because rdt_mount() takes rdtgroup_mutex and
3098 * during the debugfs directory creation also &sb->s_type->i_mutex_key
3099 * (the lockdep class of inode->i_rwsem). Other filesystem
3100 * interactions (eg. SyS_getdents) have the lock ordering:
3101 * &sb->s_type->i_mutex_key --> &mm->mmap_sem
3102 * During mmap(), called with &mm->mmap_sem, the rdtgroup_mutex
3103 * is taken, thus creating dependency:
3104 * &mm->mmap_sem --> rdtgroup_mutex for the latter that can cause
3105 * issues considering the other two lock dependencies.
3106 * By creating the debugfs directory here we avoid a dependency
3107 * that may cause deadlock (even though file operations cannot
3108 * occur until the filesystem is mounted, but I do not know how to
3109 * tell lockdep that).
3110 */
3111 debugfs_resctrl = debugfs_create_dir("resctrl", NULL);
3112
3113 return 0;
3114
3115cleanup_mountpoint:
3116 sysfs_remove_mount_point(fs_kobj, "resctrl");
3117cleanup_root:
3118 kernfs_destroy_root(rdt_root);
3119
3120 return ret;
3121}
3122
3123void __exit rdtgroup_exit(void)
3124{
3125 debugfs_remove_recursive(debugfs_resctrl);
3126 unregister_filesystem(&rdt_fs_type);
3127 sysfs_remove_mount_point(fs_kobj, "resctrl");
3128 kernfs_destroy_root(rdt_root);
3129}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * User interface for Resource Allocation in Resource Director Technology(RDT)
4 *
5 * Copyright (C) 2016 Intel Corporation
6 *
7 * Author: Fenghua Yu <fenghua.yu@intel.com>
8 *
9 * More information about RDT be found in the Intel (R) x86 Architecture
10 * Software Developer Manual.
11 */
12
13#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14
15#include <linux/cacheinfo.h>
16#include <linux/cpu.h>
17#include <linux/debugfs.h>
18#include <linux/fs.h>
19#include <linux/fs_parser.h>
20#include <linux/sysfs.h>
21#include <linux/kernfs.h>
22#include <linux/seq_buf.h>
23#include <linux/seq_file.h>
24#include <linux/sched/signal.h>
25#include <linux/sched/task.h>
26#include <linux/slab.h>
27#include <linux/task_work.h>
28#include <linux/user_namespace.h>
29
30#include <uapi/linux/magic.h>
31
32#include <asm/resctrl.h>
33#include "internal.h"
34
35DEFINE_STATIC_KEY_FALSE(rdt_enable_key);
36DEFINE_STATIC_KEY_FALSE(rdt_mon_enable_key);
37DEFINE_STATIC_KEY_FALSE(rdt_alloc_enable_key);
38static struct kernfs_root *rdt_root;
39struct rdtgroup rdtgroup_default;
40LIST_HEAD(rdt_all_groups);
41
42/* list of entries for the schemata file */
43LIST_HEAD(resctrl_schema_all);
44
45/* Kernel fs node for "info" directory under root */
46static struct kernfs_node *kn_info;
47
48/* Kernel fs node for "mon_groups" directory under root */
49static struct kernfs_node *kn_mongrp;
50
51/* Kernel fs node for "mon_data" directory under root */
52static struct kernfs_node *kn_mondata;
53
54static struct seq_buf last_cmd_status;
55static char last_cmd_status_buf[512];
56
57struct dentry *debugfs_resctrl;
58
59void rdt_last_cmd_clear(void)
60{
61 lockdep_assert_held(&rdtgroup_mutex);
62 seq_buf_clear(&last_cmd_status);
63}
64
65void rdt_last_cmd_puts(const char *s)
66{
67 lockdep_assert_held(&rdtgroup_mutex);
68 seq_buf_puts(&last_cmd_status, s);
69}
70
71void rdt_last_cmd_printf(const char *fmt, ...)
72{
73 va_list ap;
74
75 va_start(ap, fmt);
76 lockdep_assert_held(&rdtgroup_mutex);
77 seq_buf_vprintf(&last_cmd_status, fmt, ap);
78 va_end(ap);
79}
80
81/*
82 * Trivial allocator for CLOSIDs. Since h/w only supports a small number,
83 * we can keep a bitmap of free CLOSIDs in a single integer.
84 *
85 * Using a global CLOSID across all resources has some advantages and
86 * some drawbacks:
87 * + We can simply set "current->closid" to assign a task to a resource
88 * group.
89 * + Context switch code can avoid extra memory references deciding which
90 * CLOSID to load into the PQR_ASSOC MSR
91 * - We give up some options in configuring resource groups across multi-socket
92 * systems.
93 * - Our choices on how to configure each resource become progressively more
94 * limited as the number of resources grows.
95 */
96static int closid_free_map;
97static int closid_free_map_len;
98
99int closids_supported(void)
100{
101 return closid_free_map_len;
102}
103
104static void closid_init(void)
105{
106 struct resctrl_schema *s;
107 u32 rdt_min_closid = 32;
108
109 /* Compute rdt_min_closid across all resources */
110 list_for_each_entry(s, &resctrl_schema_all, list)
111 rdt_min_closid = min(rdt_min_closid, s->num_closid);
112
113 closid_free_map = BIT_MASK(rdt_min_closid) - 1;
114
115 /* CLOSID 0 is always reserved for the default group */
116 closid_free_map &= ~1;
117 closid_free_map_len = rdt_min_closid;
118}
119
120static int closid_alloc(void)
121{
122 u32 closid = ffs(closid_free_map);
123
124 if (closid == 0)
125 return -ENOSPC;
126 closid--;
127 closid_free_map &= ~(1 << closid);
128
129 return closid;
130}
131
132void closid_free(int closid)
133{
134 closid_free_map |= 1 << closid;
135}
136
137/**
138 * closid_allocated - test if provided closid is in use
139 * @closid: closid to be tested
140 *
141 * Return: true if @closid is currently associated with a resource group,
142 * false if @closid is free
143 */
144static bool closid_allocated(unsigned int closid)
145{
146 return (closid_free_map & (1 << closid)) == 0;
147}
148
149/**
150 * rdtgroup_mode_by_closid - Return mode of resource group with closid
151 * @closid: closid if the resource group
152 *
153 * Each resource group is associated with a @closid. Here the mode
154 * of a resource group can be queried by searching for it using its closid.
155 *
156 * Return: mode as &enum rdtgrp_mode of resource group with closid @closid
157 */
158enum rdtgrp_mode rdtgroup_mode_by_closid(int closid)
159{
160 struct rdtgroup *rdtgrp;
161
162 list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
163 if (rdtgrp->closid == closid)
164 return rdtgrp->mode;
165 }
166
167 return RDT_NUM_MODES;
168}
169
170static const char * const rdt_mode_str[] = {
171 [RDT_MODE_SHAREABLE] = "shareable",
172 [RDT_MODE_EXCLUSIVE] = "exclusive",
173 [RDT_MODE_PSEUDO_LOCKSETUP] = "pseudo-locksetup",
174 [RDT_MODE_PSEUDO_LOCKED] = "pseudo-locked",
175};
176
177/**
178 * rdtgroup_mode_str - Return the string representation of mode
179 * @mode: the resource group mode as &enum rdtgroup_mode
180 *
181 * Return: string representation of valid mode, "unknown" otherwise
182 */
183static const char *rdtgroup_mode_str(enum rdtgrp_mode mode)
184{
185 if (mode < RDT_MODE_SHAREABLE || mode >= RDT_NUM_MODES)
186 return "unknown";
187
188 return rdt_mode_str[mode];
189}
190
191/* set uid and gid of rdtgroup dirs and files to that of the creator */
192static int rdtgroup_kn_set_ugid(struct kernfs_node *kn)
193{
194 struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
195 .ia_uid = current_fsuid(),
196 .ia_gid = current_fsgid(), };
197
198 if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
199 gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
200 return 0;
201
202 return kernfs_setattr(kn, &iattr);
203}
204
205static int rdtgroup_add_file(struct kernfs_node *parent_kn, struct rftype *rft)
206{
207 struct kernfs_node *kn;
208 int ret;
209
210 kn = __kernfs_create_file(parent_kn, rft->name, rft->mode,
211 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
212 0, rft->kf_ops, rft, NULL, NULL);
213 if (IS_ERR(kn))
214 return PTR_ERR(kn);
215
216 ret = rdtgroup_kn_set_ugid(kn);
217 if (ret) {
218 kernfs_remove(kn);
219 return ret;
220 }
221
222 return 0;
223}
224
225static int rdtgroup_seqfile_show(struct seq_file *m, void *arg)
226{
227 struct kernfs_open_file *of = m->private;
228 struct rftype *rft = of->kn->priv;
229
230 if (rft->seq_show)
231 return rft->seq_show(of, m, arg);
232 return 0;
233}
234
235static ssize_t rdtgroup_file_write(struct kernfs_open_file *of, char *buf,
236 size_t nbytes, loff_t off)
237{
238 struct rftype *rft = of->kn->priv;
239
240 if (rft->write)
241 return rft->write(of, buf, nbytes, off);
242
243 return -EINVAL;
244}
245
246static const struct kernfs_ops rdtgroup_kf_single_ops = {
247 .atomic_write_len = PAGE_SIZE,
248 .write = rdtgroup_file_write,
249 .seq_show = rdtgroup_seqfile_show,
250};
251
252static const struct kernfs_ops kf_mondata_ops = {
253 .atomic_write_len = PAGE_SIZE,
254 .seq_show = rdtgroup_mondata_show,
255};
256
257static bool is_cpu_list(struct kernfs_open_file *of)
258{
259 struct rftype *rft = of->kn->priv;
260
261 return rft->flags & RFTYPE_FLAGS_CPUS_LIST;
262}
263
264static int rdtgroup_cpus_show(struct kernfs_open_file *of,
265 struct seq_file *s, void *v)
266{
267 struct rdtgroup *rdtgrp;
268 struct cpumask *mask;
269 int ret = 0;
270
271 rdtgrp = rdtgroup_kn_lock_live(of->kn);
272
273 if (rdtgrp) {
274 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
275 if (!rdtgrp->plr->d) {
276 rdt_last_cmd_clear();
277 rdt_last_cmd_puts("Cache domain offline\n");
278 ret = -ENODEV;
279 } else {
280 mask = &rdtgrp->plr->d->cpu_mask;
281 seq_printf(s, is_cpu_list(of) ?
282 "%*pbl\n" : "%*pb\n",
283 cpumask_pr_args(mask));
284 }
285 } else {
286 seq_printf(s, is_cpu_list(of) ? "%*pbl\n" : "%*pb\n",
287 cpumask_pr_args(&rdtgrp->cpu_mask));
288 }
289 } else {
290 ret = -ENOENT;
291 }
292 rdtgroup_kn_unlock(of->kn);
293
294 return ret;
295}
296
297/*
298 * This is safe against resctrl_sched_in() called from __switch_to()
299 * because __switch_to() is executed with interrupts disabled. A local call
300 * from update_closid_rmid() is protected against __switch_to() because
301 * preemption is disabled.
302 */
303static void update_cpu_closid_rmid(void *info)
304{
305 struct rdtgroup *r = info;
306
307 if (r) {
308 this_cpu_write(pqr_state.default_closid, r->closid);
309 this_cpu_write(pqr_state.default_rmid, r->mon.rmid);
310 }
311
312 /*
313 * We cannot unconditionally write the MSR because the current
314 * executing task might have its own closid selected. Just reuse
315 * the context switch code.
316 */
317 resctrl_sched_in();
318}
319
320/*
321 * Update the PGR_ASSOC MSR on all cpus in @cpu_mask,
322 *
323 * Per task closids/rmids must have been set up before calling this function.
324 */
325static void
326update_closid_rmid(const struct cpumask *cpu_mask, struct rdtgroup *r)
327{
328 int cpu = get_cpu();
329
330 if (cpumask_test_cpu(cpu, cpu_mask))
331 update_cpu_closid_rmid(r);
332 smp_call_function_many(cpu_mask, update_cpu_closid_rmid, r, 1);
333 put_cpu();
334}
335
336static int cpus_mon_write(struct rdtgroup *rdtgrp, cpumask_var_t newmask,
337 cpumask_var_t tmpmask)
338{
339 struct rdtgroup *prgrp = rdtgrp->mon.parent, *crgrp;
340 struct list_head *head;
341
342 /* Check whether cpus belong to parent ctrl group */
343 cpumask_andnot(tmpmask, newmask, &prgrp->cpu_mask);
344 if (!cpumask_empty(tmpmask)) {
345 rdt_last_cmd_puts("Can only add CPUs to mongroup that belong to parent\n");
346 return -EINVAL;
347 }
348
349 /* Check whether cpus are dropped from this group */
350 cpumask_andnot(tmpmask, &rdtgrp->cpu_mask, newmask);
351 if (!cpumask_empty(tmpmask)) {
352 /* Give any dropped cpus to parent rdtgroup */
353 cpumask_or(&prgrp->cpu_mask, &prgrp->cpu_mask, tmpmask);
354 update_closid_rmid(tmpmask, prgrp);
355 }
356
357 /*
358 * If we added cpus, remove them from previous group that owned them
359 * and update per-cpu rmid
360 */
361 cpumask_andnot(tmpmask, newmask, &rdtgrp->cpu_mask);
362 if (!cpumask_empty(tmpmask)) {
363 head = &prgrp->mon.crdtgrp_list;
364 list_for_each_entry(crgrp, head, mon.crdtgrp_list) {
365 if (crgrp == rdtgrp)
366 continue;
367 cpumask_andnot(&crgrp->cpu_mask, &crgrp->cpu_mask,
368 tmpmask);
369 }
370 update_closid_rmid(tmpmask, rdtgrp);
371 }
372
373 /* Done pushing/pulling - update this group with new mask */
374 cpumask_copy(&rdtgrp->cpu_mask, newmask);
375
376 return 0;
377}
378
379static void cpumask_rdtgrp_clear(struct rdtgroup *r, struct cpumask *m)
380{
381 struct rdtgroup *crgrp;
382
383 cpumask_andnot(&r->cpu_mask, &r->cpu_mask, m);
384 /* update the child mon group masks as well*/
385 list_for_each_entry(crgrp, &r->mon.crdtgrp_list, mon.crdtgrp_list)
386 cpumask_and(&crgrp->cpu_mask, &r->cpu_mask, &crgrp->cpu_mask);
387}
388
389static int cpus_ctrl_write(struct rdtgroup *rdtgrp, cpumask_var_t newmask,
390 cpumask_var_t tmpmask, cpumask_var_t tmpmask1)
391{
392 struct rdtgroup *r, *crgrp;
393 struct list_head *head;
394
395 /* Check whether cpus are dropped from this group */
396 cpumask_andnot(tmpmask, &rdtgrp->cpu_mask, newmask);
397 if (!cpumask_empty(tmpmask)) {
398 /* Can't drop from default group */
399 if (rdtgrp == &rdtgroup_default) {
400 rdt_last_cmd_puts("Can't drop CPUs from default group\n");
401 return -EINVAL;
402 }
403
404 /* Give any dropped cpus to rdtgroup_default */
405 cpumask_or(&rdtgroup_default.cpu_mask,
406 &rdtgroup_default.cpu_mask, tmpmask);
407 update_closid_rmid(tmpmask, &rdtgroup_default);
408 }
409
410 /*
411 * If we added cpus, remove them from previous group and
412 * the prev group's child groups that owned them
413 * and update per-cpu closid/rmid.
414 */
415 cpumask_andnot(tmpmask, newmask, &rdtgrp->cpu_mask);
416 if (!cpumask_empty(tmpmask)) {
417 list_for_each_entry(r, &rdt_all_groups, rdtgroup_list) {
418 if (r == rdtgrp)
419 continue;
420 cpumask_and(tmpmask1, &r->cpu_mask, tmpmask);
421 if (!cpumask_empty(tmpmask1))
422 cpumask_rdtgrp_clear(r, tmpmask1);
423 }
424 update_closid_rmid(tmpmask, rdtgrp);
425 }
426
427 /* Done pushing/pulling - update this group with new mask */
428 cpumask_copy(&rdtgrp->cpu_mask, newmask);
429
430 /*
431 * Clear child mon group masks since there is a new parent mask
432 * now and update the rmid for the cpus the child lost.
433 */
434 head = &rdtgrp->mon.crdtgrp_list;
435 list_for_each_entry(crgrp, head, mon.crdtgrp_list) {
436 cpumask_and(tmpmask, &rdtgrp->cpu_mask, &crgrp->cpu_mask);
437 update_closid_rmid(tmpmask, rdtgrp);
438 cpumask_clear(&crgrp->cpu_mask);
439 }
440
441 return 0;
442}
443
444static ssize_t rdtgroup_cpus_write(struct kernfs_open_file *of,
445 char *buf, size_t nbytes, loff_t off)
446{
447 cpumask_var_t tmpmask, newmask, tmpmask1;
448 struct rdtgroup *rdtgrp;
449 int ret;
450
451 if (!buf)
452 return -EINVAL;
453
454 if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
455 return -ENOMEM;
456 if (!zalloc_cpumask_var(&newmask, GFP_KERNEL)) {
457 free_cpumask_var(tmpmask);
458 return -ENOMEM;
459 }
460 if (!zalloc_cpumask_var(&tmpmask1, GFP_KERNEL)) {
461 free_cpumask_var(tmpmask);
462 free_cpumask_var(newmask);
463 return -ENOMEM;
464 }
465
466 rdtgrp = rdtgroup_kn_lock_live(of->kn);
467 if (!rdtgrp) {
468 ret = -ENOENT;
469 goto unlock;
470 }
471
472 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED ||
473 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
474 ret = -EINVAL;
475 rdt_last_cmd_puts("Pseudo-locking in progress\n");
476 goto unlock;
477 }
478
479 if (is_cpu_list(of))
480 ret = cpulist_parse(buf, newmask);
481 else
482 ret = cpumask_parse(buf, newmask);
483
484 if (ret) {
485 rdt_last_cmd_puts("Bad CPU list/mask\n");
486 goto unlock;
487 }
488
489 /* check that user didn't specify any offline cpus */
490 cpumask_andnot(tmpmask, newmask, cpu_online_mask);
491 if (!cpumask_empty(tmpmask)) {
492 ret = -EINVAL;
493 rdt_last_cmd_puts("Can only assign online CPUs\n");
494 goto unlock;
495 }
496
497 if (rdtgrp->type == RDTCTRL_GROUP)
498 ret = cpus_ctrl_write(rdtgrp, newmask, tmpmask, tmpmask1);
499 else if (rdtgrp->type == RDTMON_GROUP)
500 ret = cpus_mon_write(rdtgrp, newmask, tmpmask);
501 else
502 ret = -EINVAL;
503
504unlock:
505 rdtgroup_kn_unlock(of->kn);
506 free_cpumask_var(tmpmask);
507 free_cpumask_var(newmask);
508 free_cpumask_var(tmpmask1);
509
510 return ret ?: nbytes;
511}
512
513/**
514 * rdtgroup_remove - the helper to remove resource group safely
515 * @rdtgrp: resource group to remove
516 *
517 * On resource group creation via a mkdir, an extra kernfs_node reference is
518 * taken to ensure that the rdtgroup structure remains accessible for the
519 * rdtgroup_kn_unlock() calls where it is removed.
520 *
521 * Drop the extra reference here, then free the rdtgroup structure.
522 *
523 * Return: void
524 */
525static void rdtgroup_remove(struct rdtgroup *rdtgrp)
526{
527 kernfs_put(rdtgrp->kn);
528 kfree(rdtgrp);
529}
530
531static void _update_task_closid_rmid(void *task)
532{
533 /*
534 * If the task is still current on this CPU, update PQR_ASSOC MSR.
535 * Otherwise, the MSR is updated when the task is scheduled in.
536 */
537 if (task == current)
538 resctrl_sched_in();
539}
540
541static void update_task_closid_rmid(struct task_struct *t)
542{
543 if (IS_ENABLED(CONFIG_SMP) && task_curr(t))
544 smp_call_function_single(task_cpu(t), _update_task_closid_rmid, t, 1);
545 else
546 _update_task_closid_rmid(t);
547}
548
549static int __rdtgroup_move_task(struct task_struct *tsk,
550 struct rdtgroup *rdtgrp)
551{
552 /* If the task is already in rdtgrp, no need to move the task. */
553 if ((rdtgrp->type == RDTCTRL_GROUP && tsk->closid == rdtgrp->closid &&
554 tsk->rmid == rdtgrp->mon.rmid) ||
555 (rdtgrp->type == RDTMON_GROUP && tsk->rmid == rdtgrp->mon.rmid &&
556 tsk->closid == rdtgrp->mon.parent->closid))
557 return 0;
558
559 /*
560 * Set the task's closid/rmid before the PQR_ASSOC MSR can be
561 * updated by them.
562 *
563 * For ctrl_mon groups, move both closid and rmid.
564 * For monitor groups, can move the tasks only from
565 * their parent CTRL group.
566 */
567
568 if (rdtgrp->type == RDTCTRL_GROUP) {
569 WRITE_ONCE(tsk->closid, rdtgrp->closid);
570 WRITE_ONCE(tsk->rmid, rdtgrp->mon.rmid);
571 } else if (rdtgrp->type == RDTMON_GROUP) {
572 if (rdtgrp->mon.parent->closid == tsk->closid) {
573 WRITE_ONCE(tsk->rmid, rdtgrp->mon.rmid);
574 } else {
575 rdt_last_cmd_puts("Can't move task to different control group\n");
576 return -EINVAL;
577 }
578 }
579
580 /*
581 * Ensure the task's closid and rmid are written before determining if
582 * the task is current that will decide if it will be interrupted.
583 * This pairs with the full barrier between the rq->curr update and
584 * resctrl_sched_in() during context switch.
585 */
586 smp_mb();
587
588 /*
589 * By now, the task's closid and rmid are set. If the task is current
590 * on a CPU, the PQR_ASSOC MSR needs to be updated to make the resource
591 * group go into effect. If the task is not current, the MSR will be
592 * updated when the task is scheduled in.
593 */
594 update_task_closid_rmid(tsk);
595
596 return 0;
597}
598
599static bool is_closid_match(struct task_struct *t, struct rdtgroup *r)
600{
601 return (rdt_alloc_capable &&
602 (r->type == RDTCTRL_GROUP) && (t->closid == r->closid));
603}
604
605static bool is_rmid_match(struct task_struct *t, struct rdtgroup *r)
606{
607 return (rdt_mon_capable &&
608 (r->type == RDTMON_GROUP) && (t->rmid == r->mon.rmid));
609}
610
611/**
612 * rdtgroup_tasks_assigned - Test if tasks have been assigned to resource group
613 * @r: Resource group
614 *
615 * Return: 1 if tasks have been assigned to @r, 0 otherwise
616 */
617int rdtgroup_tasks_assigned(struct rdtgroup *r)
618{
619 struct task_struct *p, *t;
620 int ret = 0;
621
622 lockdep_assert_held(&rdtgroup_mutex);
623
624 rcu_read_lock();
625 for_each_process_thread(p, t) {
626 if (is_closid_match(t, r) || is_rmid_match(t, r)) {
627 ret = 1;
628 break;
629 }
630 }
631 rcu_read_unlock();
632
633 return ret;
634}
635
636static int rdtgroup_task_write_permission(struct task_struct *task,
637 struct kernfs_open_file *of)
638{
639 const struct cred *tcred = get_task_cred(task);
640 const struct cred *cred = current_cred();
641 int ret = 0;
642
643 /*
644 * Even if we're attaching all tasks in the thread group, we only
645 * need to check permissions on one of them.
646 */
647 if (!uid_eq(cred->euid, GLOBAL_ROOT_UID) &&
648 !uid_eq(cred->euid, tcred->uid) &&
649 !uid_eq(cred->euid, tcred->suid)) {
650 rdt_last_cmd_printf("No permission to move task %d\n", task->pid);
651 ret = -EPERM;
652 }
653
654 put_cred(tcred);
655 return ret;
656}
657
658static int rdtgroup_move_task(pid_t pid, struct rdtgroup *rdtgrp,
659 struct kernfs_open_file *of)
660{
661 struct task_struct *tsk;
662 int ret;
663
664 rcu_read_lock();
665 if (pid) {
666 tsk = find_task_by_vpid(pid);
667 if (!tsk) {
668 rcu_read_unlock();
669 rdt_last_cmd_printf("No task %d\n", pid);
670 return -ESRCH;
671 }
672 } else {
673 tsk = current;
674 }
675
676 get_task_struct(tsk);
677 rcu_read_unlock();
678
679 ret = rdtgroup_task_write_permission(tsk, of);
680 if (!ret)
681 ret = __rdtgroup_move_task(tsk, rdtgrp);
682
683 put_task_struct(tsk);
684 return ret;
685}
686
687static ssize_t rdtgroup_tasks_write(struct kernfs_open_file *of,
688 char *buf, size_t nbytes, loff_t off)
689{
690 struct rdtgroup *rdtgrp;
691 int ret = 0;
692 pid_t pid;
693
694 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
695 return -EINVAL;
696 rdtgrp = rdtgroup_kn_lock_live(of->kn);
697 if (!rdtgrp) {
698 rdtgroup_kn_unlock(of->kn);
699 return -ENOENT;
700 }
701 rdt_last_cmd_clear();
702
703 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED ||
704 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
705 ret = -EINVAL;
706 rdt_last_cmd_puts("Pseudo-locking in progress\n");
707 goto unlock;
708 }
709
710 ret = rdtgroup_move_task(pid, rdtgrp, of);
711
712unlock:
713 rdtgroup_kn_unlock(of->kn);
714
715 return ret ?: nbytes;
716}
717
718static void show_rdt_tasks(struct rdtgroup *r, struct seq_file *s)
719{
720 struct task_struct *p, *t;
721
722 rcu_read_lock();
723 for_each_process_thread(p, t) {
724 if (is_closid_match(t, r) || is_rmid_match(t, r))
725 seq_printf(s, "%d\n", t->pid);
726 }
727 rcu_read_unlock();
728}
729
730static int rdtgroup_tasks_show(struct kernfs_open_file *of,
731 struct seq_file *s, void *v)
732{
733 struct rdtgroup *rdtgrp;
734 int ret = 0;
735
736 rdtgrp = rdtgroup_kn_lock_live(of->kn);
737 if (rdtgrp)
738 show_rdt_tasks(rdtgrp, s);
739 else
740 ret = -ENOENT;
741 rdtgroup_kn_unlock(of->kn);
742
743 return ret;
744}
745
746#ifdef CONFIG_PROC_CPU_RESCTRL
747
748/*
749 * A task can only be part of one resctrl control group and of one monitor
750 * group which is associated to that control group.
751 *
752 * 1) res:
753 * mon:
754 *
755 * resctrl is not available.
756 *
757 * 2) res:/
758 * mon:
759 *
760 * Task is part of the root resctrl control group, and it is not associated
761 * to any monitor group.
762 *
763 * 3) res:/
764 * mon:mon0
765 *
766 * Task is part of the root resctrl control group and monitor group mon0.
767 *
768 * 4) res:group0
769 * mon:
770 *
771 * Task is part of resctrl control group group0, and it is not associated
772 * to any monitor group.
773 *
774 * 5) res:group0
775 * mon:mon1
776 *
777 * Task is part of resctrl control group group0 and monitor group mon1.
778 */
779int proc_resctrl_show(struct seq_file *s, struct pid_namespace *ns,
780 struct pid *pid, struct task_struct *tsk)
781{
782 struct rdtgroup *rdtg;
783 int ret = 0;
784
785 mutex_lock(&rdtgroup_mutex);
786
787 /* Return empty if resctrl has not been mounted. */
788 if (!static_branch_unlikely(&rdt_enable_key)) {
789 seq_puts(s, "res:\nmon:\n");
790 goto unlock;
791 }
792
793 list_for_each_entry(rdtg, &rdt_all_groups, rdtgroup_list) {
794 struct rdtgroup *crg;
795
796 /*
797 * Task information is only relevant for shareable
798 * and exclusive groups.
799 */
800 if (rdtg->mode != RDT_MODE_SHAREABLE &&
801 rdtg->mode != RDT_MODE_EXCLUSIVE)
802 continue;
803
804 if (rdtg->closid != tsk->closid)
805 continue;
806
807 seq_printf(s, "res:%s%s\n", (rdtg == &rdtgroup_default) ? "/" : "",
808 rdtg->kn->name);
809 seq_puts(s, "mon:");
810 list_for_each_entry(crg, &rdtg->mon.crdtgrp_list,
811 mon.crdtgrp_list) {
812 if (tsk->rmid != crg->mon.rmid)
813 continue;
814 seq_printf(s, "%s", crg->kn->name);
815 break;
816 }
817 seq_putc(s, '\n');
818 goto unlock;
819 }
820 /*
821 * The above search should succeed. Otherwise return
822 * with an error.
823 */
824 ret = -ENOENT;
825unlock:
826 mutex_unlock(&rdtgroup_mutex);
827
828 return ret;
829}
830#endif
831
832static int rdt_last_cmd_status_show(struct kernfs_open_file *of,
833 struct seq_file *seq, void *v)
834{
835 int len;
836
837 mutex_lock(&rdtgroup_mutex);
838 len = seq_buf_used(&last_cmd_status);
839 if (len)
840 seq_printf(seq, "%.*s", len, last_cmd_status_buf);
841 else
842 seq_puts(seq, "ok\n");
843 mutex_unlock(&rdtgroup_mutex);
844 return 0;
845}
846
847static int rdt_num_closids_show(struct kernfs_open_file *of,
848 struct seq_file *seq, void *v)
849{
850 struct resctrl_schema *s = of->kn->parent->priv;
851
852 seq_printf(seq, "%u\n", s->num_closid);
853 return 0;
854}
855
856static int rdt_default_ctrl_show(struct kernfs_open_file *of,
857 struct seq_file *seq, void *v)
858{
859 struct resctrl_schema *s = of->kn->parent->priv;
860 struct rdt_resource *r = s->res;
861
862 seq_printf(seq, "%x\n", r->default_ctrl);
863 return 0;
864}
865
866static int rdt_min_cbm_bits_show(struct kernfs_open_file *of,
867 struct seq_file *seq, void *v)
868{
869 struct resctrl_schema *s = of->kn->parent->priv;
870 struct rdt_resource *r = s->res;
871
872 seq_printf(seq, "%u\n", r->cache.min_cbm_bits);
873 return 0;
874}
875
876static int rdt_shareable_bits_show(struct kernfs_open_file *of,
877 struct seq_file *seq, void *v)
878{
879 struct resctrl_schema *s = of->kn->parent->priv;
880 struct rdt_resource *r = s->res;
881
882 seq_printf(seq, "%x\n", r->cache.shareable_bits);
883 return 0;
884}
885
886/**
887 * rdt_bit_usage_show - Display current usage of resources
888 *
889 * A domain is a shared resource that can now be allocated differently. Here
890 * we display the current regions of the domain as an annotated bitmask.
891 * For each domain of this resource its allocation bitmask
892 * is annotated as below to indicate the current usage of the corresponding bit:
893 * 0 - currently unused
894 * X - currently available for sharing and used by software and hardware
895 * H - currently used by hardware only but available for software use
896 * S - currently used and shareable by software only
897 * E - currently used exclusively by one resource group
898 * P - currently pseudo-locked by one resource group
899 */
900static int rdt_bit_usage_show(struct kernfs_open_file *of,
901 struct seq_file *seq, void *v)
902{
903 struct resctrl_schema *s = of->kn->parent->priv;
904 /*
905 * Use unsigned long even though only 32 bits are used to ensure
906 * test_bit() is used safely.
907 */
908 unsigned long sw_shareable = 0, hw_shareable = 0;
909 unsigned long exclusive = 0, pseudo_locked = 0;
910 struct rdt_resource *r = s->res;
911 struct rdt_domain *dom;
912 int i, hwb, swb, excl, psl;
913 enum rdtgrp_mode mode;
914 bool sep = false;
915 u32 ctrl_val;
916
917 mutex_lock(&rdtgroup_mutex);
918 hw_shareable = r->cache.shareable_bits;
919 list_for_each_entry(dom, &r->domains, list) {
920 if (sep)
921 seq_putc(seq, ';');
922 sw_shareable = 0;
923 exclusive = 0;
924 seq_printf(seq, "%d=", dom->id);
925 for (i = 0; i < closids_supported(); i++) {
926 if (!closid_allocated(i))
927 continue;
928 ctrl_val = resctrl_arch_get_config(r, dom, i,
929 s->conf_type);
930 mode = rdtgroup_mode_by_closid(i);
931 switch (mode) {
932 case RDT_MODE_SHAREABLE:
933 sw_shareable |= ctrl_val;
934 break;
935 case RDT_MODE_EXCLUSIVE:
936 exclusive |= ctrl_val;
937 break;
938 case RDT_MODE_PSEUDO_LOCKSETUP:
939 /*
940 * RDT_MODE_PSEUDO_LOCKSETUP is possible
941 * here but not included since the CBM
942 * associated with this CLOSID in this mode
943 * is not initialized and no task or cpu can be
944 * assigned this CLOSID.
945 */
946 break;
947 case RDT_MODE_PSEUDO_LOCKED:
948 case RDT_NUM_MODES:
949 WARN(1,
950 "invalid mode for closid %d\n", i);
951 break;
952 }
953 }
954 for (i = r->cache.cbm_len - 1; i >= 0; i--) {
955 pseudo_locked = dom->plr ? dom->plr->cbm : 0;
956 hwb = test_bit(i, &hw_shareable);
957 swb = test_bit(i, &sw_shareable);
958 excl = test_bit(i, &exclusive);
959 psl = test_bit(i, &pseudo_locked);
960 if (hwb && swb)
961 seq_putc(seq, 'X');
962 else if (hwb && !swb)
963 seq_putc(seq, 'H');
964 else if (!hwb && swb)
965 seq_putc(seq, 'S');
966 else if (excl)
967 seq_putc(seq, 'E');
968 else if (psl)
969 seq_putc(seq, 'P');
970 else /* Unused bits remain */
971 seq_putc(seq, '0');
972 }
973 sep = true;
974 }
975 seq_putc(seq, '\n');
976 mutex_unlock(&rdtgroup_mutex);
977 return 0;
978}
979
980static int rdt_min_bw_show(struct kernfs_open_file *of,
981 struct seq_file *seq, void *v)
982{
983 struct resctrl_schema *s = of->kn->parent->priv;
984 struct rdt_resource *r = s->res;
985
986 seq_printf(seq, "%u\n", r->membw.min_bw);
987 return 0;
988}
989
990static int rdt_num_rmids_show(struct kernfs_open_file *of,
991 struct seq_file *seq, void *v)
992{
993 struct rdt_resource *r = of->kn->parent->priv;
994
995 seq_printf(seq, "%d\n", r->num_rmid);
996
997 return 0;
998}
999
1000static int rdt_mon_features_show(struct kernfs_open_file *of,
1001 struct seq_file *seq, void *v)
1002{
1003 struct rdt_resource *r = of->kn->parent->priv;
1004 struct mon_evt *mevt;
1005
1006 list_for_each_entry(mevt, &r->evt_list, list)
1007 seq_printf(seq, "%s\n", mevt->name);
1008
1009 return 0;
1010}
1011
1012static int rdt_bw_gran_show(struct kernfs_open_file *of,
1013 struct seq_file *seq, void *v)
1014{
1015 struct resctrl_schema *s = of->kn->parent->priv;
1016 struct rdt_resource *r = s->res;
1017
1018 seq_printf(seq, "%u\n", r->membw.bw_gran);
1019 return 0;
1020}
1021
1022static int rdt_delay_linear_show(struct kernfs_open_file *of,
1023 struct seq_file *seq, void *v)
1024{
1025 struct resctrl_schema *s = of->kn->parent->priv;
1026 struct rdt_resource *r = s->res;
1027
1028 seq_printf(seq, "%u\n", r->membw.delay_linear);
1029 return 0;
1030}
1031
1032static int max_threshold_occ_show(struct kernfs_open_file *of,
1033 struct seq_file *seq, void *v)
1034{
1035 seq_printf(seq, "%u\n", resctrl_rmid_realloc_threshold);
1036
1037 return 0;
1038}
1039
1040static int rdt_thread_throttle_mode_show(struct kernfs_open_file *of,
1041 struct seq_file *seq, void *v)
1042{
1043 struct resctrl_schema *s = of->kn->parent->priv;
1044 struct rdt_resource *r = s->res;
1045
1046 if (r->membw.throttle_mode == THREAD_THROTTLE_PER_THREAD)
1047 seq_puts(seq, "per-thread\n");
1048 else
1049 seq_puts(seq, "max\n");
1050
1051 return 0;
1052}
1053
1054static ssize_t max_threshold_occ_write(struct kernfs_open_file *of,
1055 char *buf, size_t nbytes, loff_t off)
1056{
1057 unsigned int bytes;
1058 int ret;
1059
1060 ret = kstrtouint(buf, 0, &bytes);
1061 if (ret)
1062 return ret;
1063
1064 if (bytes > resctrl_rmid_realloc_limit)
1065 return -EINVAL;
1066
1067 resctrl_rmid_realloc_threshold = resctrl_arch_round_mon_val(bytes);
1068
1069 return nbytes;
1070}
1071
1072/*
1073 * rdtgroup_mode_show - Display mode of this resource group
1074 */
1075static int rdtgroup_mode_show(struct kernfs_open_file *of,
1076 struct seq_file *s, void *v)
1077{
1078 struct rdtgroup *rdtgrp;
1079
1080 rdtgrp = rdtgroup_kn_lock_live(of->kn);
1081 if (!rdtgrp) {
1082 rdtgroup_kn_unlock(of->kn);
1083 return -ENOENT;
1084 }
1085
1086 seq_printf(s, "%s\n", rdtgroup_mode_str(rdtgrp->mode));
1087
1088 rdtgroup_kn_unlock(of->kn);
1089 return 0;
1090}
1091
1092static enum resctrl_conf_type resctrl_peer_type(enum resctrl_conf_type my_type)
1093{
1094 switch (my_type) {
1095 case CDP_CODE:
1096 return CDP_DATA;
1097 case CDP_DATA:
1098 return CDP_CODE;
1099 default:
1100 case CDP_NONE:
1101 return CDP_NONE;
1102 }
1103}
1104
1105/**
1106 * __rdtgroup_cbm_overlaps - Does CBM for intended closid overlap with other
1107 * @r: Resource to which domain instance @d belongs.
1108 * @d: The domain instance for which @closid is being tested.
1109 * @cbm: Capacity bitmask being tested.
1110 * @closid: Intended closid for @cbm.
1111 * @exclusive: Only check if overlaps with exclusive resource groups
1112 *
1113 * Checks if provided @cbm intended to be used for @closid on domain
1114 * @d overlaps with any other closids or other hardware usage associated
1115 * with this domain. If @exclusive is true then only overlaps with
1116 * resource groups in exclusive mode will be considered. If @exclusive
1117 * is false then overlaps with any resource group or hardware entities
1118 * will be considered.
1119 *
1120 * @cbm is unsigned long, even if only 32 bits are used, to make the
1121 * bitmap functions work correctly.
1122 *
1123 * Return: false if CBM does not overlap, true if it does.
1124 */
1125static bool __rdtgroup_cbm_overlaps(struct rdt_resource *r, struct rdt_domain *d,
1126 unsigned long cbm, int closid,
1127 enum resctrl_conf_type type, bool exclusive)
1128{
1129 enum rdtgrp_mode mode;
1130 unsigned long ctrl_b;
1131 int i;
1132
1133 /* Check for any overlap with regions used by hardware directly */
1134 if (!exclusive) {
1135 ctrl_b = r->cache.shareable_bits;
1136 if (bitmap_intersects(&cbm, &ctrl_b, r->cache.cbm_len))
1137 return true;
1138 }
1139
1140 /* Check for overlap with other resource groups */
1141 for (i = 0; i < closids_supported(); i++) {
1142 ctrl_b = resctrl_arch_get_config(r, d, i, type);
1143 mode = rdtgroup_mode_by_closid(i);
1144 if (closid_allocated(i) && i != closid &&
1145 mode != RDT_MODE_PSEUDO_LOCKSETUP) {
1146 if (bitmap_intersects(&cbm, &ctrl_b, r->cache.cbm_len)) {
1147 if (exclusive) {
1148 if (mode == RDT_MODE_EXCLUSIVE)
1149 return true;
1150 continue;
1151 }
1152 return true;
1153 }
1154 }
1155 }
1156
1157 return false;
1158}
1159
1160/**
1161 * rdtgroup_cbm_overlaps - Does CBM overlap with other use of hardware
1162 * @s: Schema for the resource to which domain instance @d belongs.
1163 * @d: The domain instance for which @closid is being tested.
1164 * @cbm: Capacity bitmask being tested.
1165 * @closid: Intended closid for @cbm.
1166 * @exclusive: Only check if overlaps with exclusive resource groups
1167 *
1168 * Resources that can be allocated using a CBM can use the CBM to control
1169 * the overlap of these allocations. rdtgroup_cmb_overlaps() is the test
1170 * for overlap. Overlap test is not limited to the specific resource for
1171 * which the CBM is intended though - when dealing with CDP resources that
1172 * share the underlying hardware the overlap check should be performed on
1173 * the CDP resource sharing the hardware also.
1174 *
1175 * Refer to description of __rdtgroup_cbm_overlaps() for the details of the
1176 * overlap test.
1177 *
1178 * Return: true if CBM overlap detected, false if there is no overlap
1179 */
1180bool rdtgroup_cbm_overlaps(struct resctrl_schema *s, struct rdt_domain *d,
1181 unsigned long cbm, int closid, bool exclusive)
1182{
1183 enum resctrl_conf_type peer_type = resctrl_peer_type(s->conf_type);
1184 struct rdt_resource *r = s->res;
1185
1186 if (__rdtgroup_cbm_overlaps(r, d, cbm, closid, s->conf_type,
1187 exclusive))
1188 return true;
1189
1190 if (!resctrl_arch_get_cdp_enabled(r->rid))
1191 return false;
1192 return __rdtgroup_cbm_overlaps(r, d, cbm, closid, peer_type, exclusive);
1193}
1194
1195/**
1196 * rdtgroup_mode_test_exclusive - Test if this resource group can be exclusive
1197 *
1198 * An exclusive resource group implies that there should be no sharing of
1199 * its allocated resources. At the time this group is considered to be
1200 * exclusive this test can determine if its current schemata supports this
1201 * setting by testing for overlap with all other resource groups.
1202 *
1203 * Return: true if resource group can be exclusive, false if there is overlap
1204 * with allocations of other resource groups and thus this resource group
1205 * cannot be exclusive.
1206 */
1207static bool rdtgroup_mode_test_exclusive(struct rdtgroup *rdtgrp)
1208{
1209 int closid = rdtgrp->closid;
1210 struct resctrl_schema *s;
1211 struct rdt_resource *r;
1212 bool has_cache = false;
1213 struct rdt_domain *d;
1214 u32 ctrl;
1215
1216 list_for_each_entry(s, &resctrl_schema_all, list) {
1217 r = s->res;
1218 if (r->rid == RDT_RESOURCE_MBA)
1219 continue;
1220 has_cache = true;
1221 list_for_each_entry(d, &r->domains, list) {
1222 ctrl = resctrl_arch_get_config(r, d, closid,
1223 s->conf_type);
1224 if (rdtgroup_cbm_overlaps(s, d, ctrl, closid, false)) {
1225 rdt_last_cmd_puts("Schemata overlaps\n");
1226 return false;
1227 }
1228 }
1229 }
1230
1231 if (!has_cache) {
1232 rdt_last_cmd_puts("Cannot be exclusive without CAT/CDP\n");
1233 return false;
1234 }
1235
1236 return true;
1237}
1238
1239/**
1240 * rdtgroup_mode_write - Modify the resource group's mode
1241 *
1242 */
1243static ssize_t rdtgroup_mode_write(struct kernfs_open_file *of,
1244 char *buf, size_t nbytes, loff_t off)
1245{
1246 struct rdtgroup *rdtgrp;
1247 enum rdtgrp_mode mode;
1248 int ret = 0;
1249
1250 /* Valid input requires a trailing newline */
1251 if (nbytes == 0 || buf[nbytes - 1] != '\n')
1252 return -EINVAL;
1253 buf[nbytes - 1] = '\0';
1254
1255 rdtgrp = rdtgroup_kn_lock_live(of->kn);
1256 if (!rdtgrp) {
1257 rdtgroup_kn_unlock(of->kn);
1258 return -ENOENT;
1259 }
1260
1261 rdt_last_cmd_clear();
1262
1263 mode = rdtgrp->mode;
1264
1265 if ((!strcmp(buf, "shareable") && mode == RDT_MODE_SHAREABLE) ||
1266 (!strcmp(buf, "exclusive") && mode == RDT_MODE_EXCLUSIVE) ||
1267 (!strcmp(buf, "pseudo-locksetup") &&
1268 mode == RDT_MODE_PSEUDO_LOCKSETUP) ||
1269 (!strcmp(buf, "pseudo-locked") && mode == RDT_MODE_PSEUDO_LOCKED))
1270 goto out;
1271
1272 if (mode == RDT_MODE_PSEUDO_LOCKED) {
1273 rdt_last_cmd_puts("Cannot change pseudo-locked group\n");
1274 ret = -EINVAL;
1275 goto out;
1276 }
1277
1278 if (!strcmp(buf, "shareable")) {
1279 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1280 ret = rdtgroup_locksetup_exit(rdtgrp);
1281 if (ret)
1282 goto out;
1283 }
1284 rdtgrp->mode = RDT_MODE_SHAREABLE;
1285 } else if (!strcmp(buf, "exclusive")) {
1286 if (!rdtgroup_mode_test_exclusive(rdtgrp)) {
1287 ret = -EINVAL;
1288 goto out;
1289 }
1290 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1291 ret = rdtgroup_locksetup_exit(rdtgrp);
1292 if (ret)
1293 goto out;
1294 }
1295 rdtgrp->mode = RDT_MODE_EXCLUSIVE;
1296 } else if (!strcmp(buf, "pseudo-locksetup")) {
1297 ret = rdtgroup_locksetup_enter(rdtgrp);
1298 if (ret)
1299 goto out;
1300 rdtgrp->mode = RDT_MODE_PSEUDO_LOCKSETUP;
1301 } else {
1302 rdt_last_cmd_puts("Unknown or unsupported mode\n");
1303 ret = -EINVAL;
1304 }
1305
1306out:
1307 rdtgroup_kn_unlock(of->kn);
1308 return ret ?: nbytes;
1309}
1310
1311/**
1312 * rdtgroup_cbm_to_size - Translate CBM to size in bytes
1313 * @r: RDT resource to which @d belongs.
1314 * @d: RDT domain instance.
1315 * @cbm: bitmask for which the size should be computed.
1316 *
1317 * The bitmask provided associated with the RDT domain instance @d will be
1318 * translated into how many bytes it represents. The size in bytes is
1319 * computed by first dividing the total cache size by the CBM length to
1320 * determine how many bytes each bit in the bitmask represents. The result
1321 * is multiplied with the number of bits set in the bitmask.
1322 *
1323 * @cbm is unsigned long, even if only 32 bits are used to make the
1324 * bitmap functions work correctly.
1325 */
1326unsigned int rdtgroup_cbm_to_size(struct rdt_resource *r,
1327 struct rdt_domain *d, unsigned long cbm)
1328{
1329 struct cpu_cacheinfo *ci;
1330 unsigned int size = 0;
1331 int num_b, i;
1332
1333 num_b = bitmap_weight(&cbm, r->cache.cbm_len);
1334 ci = get_cpu_cacheinfo(cpumask_any(&d->cpu_mask));
1335 for (i = 0; i < ci->num_leaves; i++) {
1336 if (ci->info_list[i].level == r->cache_level) {
1337 size = ci->info_list[i].size / r->cache.cbm_len * num_b;
1338 break;
1339 }
1340 }
1341
1342 return size;
1343}
1344
1345/**
1346 * rdtgroup_size_show - Display size in bytes of allocated regions
1347 *
1348 * The "size" file mirrors the layout of the "schemata" file, printing the
1349 * size in bytes of each region instead of the capacity bitmask.
1350 *
1351 */
1352static int rdtgroup_size_show(struct kernfs_open_file *of,
1353 struct seq_file *s, void *v)
1354{
1355 struct resctrl_schema *schema;
1356 enum resctrl_conf_type type;
1357 struct rdtgroup *rdtgrp;
1358 struct rdt_resource *r;
1359 struct rdt_domain *d;
1360 unsigned int size;
1361 int ret = 0;
1362 u32 closid;
1363 bool sep;
1364 u32 ctrl;
1365
1366 rdtgrp = rdtgroup_kn_lock_live(of->kn);
1367 if (!rdtgrp) {
1368 rdtgroup_kn_unlock(of->kn);
1369 return -ENOENT;
1370 }
1371
1372 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
1373 if (!rdtgrp->plr->d) {
1374 rdt_last_cmd_clear();
1375 rdt_last_cmd_puts("Cache domain offline\n");
1376 ret = -ENODEV;
1377 } else {
1378 seq_printf(s, "%*s:", max_name_width,
1379 rdtgrp->plr->s->name);
1380 size = rdtgroup_cbm_to_size(rdtgrp->plr->s->res,
1381 rdtgrp->plr->d,
1382 rdtgrp->plr->cbm);
1383 seq_printf(s, "%d=%u\n", rdtgrp->plr->d->id, size);
1384 }
1385 goto out;
1386 }
1387
1388 closid = rdtgrp->closid;
1389
1390 list_for_each_entry(schema, &resctrl_schema_all, list) {
1391 r = schema->res;
1392 type = schema->conf_type;
1393 sep = false;
1394 seq_printf(s, "%*s:", max_name_width, schema->name);
1395 list_for_each_entry(d, &r->domains, list) {
1396 if (sep)
1397 seq_putc(s, ';');
1398 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1399 size = 0;
1400 } else {
1401 if (is_mba_sc(r))
1402 ctrl = d->mbps_val[closid];
1403 else
1404 ctrl = resctrl_arch_get_config(r, d,
1405 closid,
1406 type);
1407 if (r->rid == RDT_RESOURCE_MBA)
1408 size = ctrl;
1409 else
1410 size = rdtgroup_cbm_to_size(r, d, ctrl);
1411 }
1412 seq_printf(s, "%d=%u", d->id, size);
1413 sep = true;
1414 }
1415 seq_putc(s, '\n');
1416 }
1417
1418out:
1419 rdtgroup_kn_unlock(of->kn);
1420
1421 return ret;
1422}
1423
1424/* rdtgroup information files for one cache resource. */
1425static struct rftype res_common_files[] = {
1426 {
1427 .name = "last_cmd_status",
1428 .mode = 0444,
1429 .kf_ops = &rdtgroup_kf_single_ops,
1430 .seq_show = rdt_last_cmd_status_show,
1431 .fflags = RF_TOP_INFO,
1432 },
1433 {
1434 .name = "num_closids",
1435 .mode = 0444,
1436 .kf_ops = &rdtgroup_kf_single_ops,
1437 .seq_show = rdt_num_closids_show,
1438 .fflags = RF_CTRL_INFO,
1439 },
1440 {
1441 .name = "mon_features",
1442 .mode = 0444,
1443 .kf_ops = &rdtgroup_kf_single_ops,
1444 .seq_show = rdt_mon_features_show,
1445 .fflags = RF_MON_INFO,
1446 },
1447 {
1448 .name = "num_rmids",
1449 .mode = 0444,
1450 .kf_ops = &rdtgroup_kf_single_ops,
1451 .seq_show = rdt_num_rmids_show,
1452 .fflags = RF_MON_INFO,
1453 },
1454 {
1455 .name = "cbm_mask",
1456 .mode = 0444,
1457 .kf_ops = &rdtgroup_kf_single_ops,
1458 .seq_show = rdt_default_ctrl_show,
1459 .fflags = RF_CTRL_INFO | RFTYPE_RES_CACHE,
1460 },
1461 {
1462 .name = "min_cbm_bits",
1463 .mode = 0444,
1464 .kf_ops = &rdtgroup_kf_single_ops,
1465 .seq_show = rdt_min_cbm_bits_show,
1466 .fflags = RF_CTRL_INFO | RFTYPE_RES_CACHE,
1467 },
1468 {
1469 .name = "shareable_bits",
1470 .mode = 0444,
1471 .kf_ops = &rdtgroup_kf_single_ops,
1472 .seq_show = rdt_shareable_bits_show,
1473 .fflags = RF_CTRL_INFO | RFTYPE_RES_CACHE,
1474 },
1475 {
1476 .name = "bit_usage",
1477 .mode = 0444,
1478 .kf_ops = &rdtgroup_kf_single_ops,
1479 .seq_show = rdt_bit_usage_show,
1480 .fflags = RF_CTRL_INFO | RFTYPE_RES_CACHE,
1481 },
1482 {
1483 .name = "min_bandwidth",
1484 .mode = 0444,
1485 .kf_ops = &rdtgroup_kf_single_ops,
1486 .seq_show = rdt_min_bw_show,
1487 .fflags = RF_CTRL_INFO | RFTYPE_RES_MB,
1488 },
1489 {
1490 .name = "bandwidth_gran",
1491 .mode = 0444,
1492 .kf_ops = &rdtgroup_kf_single_ops,
1493 .seq_show = rdt_bw_gran_show,
1494 .fflags = RF_CTRL_INFO | RFTYPE_RES_MB,
1495 },
1496 {
1497 .name = "delay_linear",
1498 .mode = 0444,
1499 .kf_ops = &rdtgroup_kf_single_ops,
1500 .seq_show = rdt_delay_linear_show,
1501 .fflags = RF_CTRL_INFO | RFTYPE_RES_MB,
1502 },
1503 /*
1504 * Platform specific which (if any) capabilities are provided by
1505 * thread_throttle_mode. Defer "fflags" initialization to platform
1506 * discovery.
1507 */
1508 {
1509 .name = "thread_throttle_mode",
1510 .mode = 0444,
1511 .kf_ops = &rdtgroup_kf_single_ops,
1512 .seq_show = rdt_thread_throttle_mode_show,
1513 },
1514 {
1515 .name = "max_threshold_occupancy",
1516 .mode = 0644,
1517 .kf_ops = &rdtgroup_kf_single_ops,
1518 .write = max_threshold_occ_write,
1519 .seq_show = max_threshold_occ_show,
1520 .fflags = RF_MON_INFO | RFTYPE_RES_CACHE,
1521 },
1522 {
1523 .name = "cpus",
1524 .mode = 0644,
1525 .kf_ops = &rdtgroup_kf_single_ops,
1526 .write = rdtgroup_cpus_write,
1527 .seq_show = rdtgroup_cpus_show,
1528 .fflags = RFTYPE_BASE,
1529 },
1530 {
1531 .name = "cpus_list",
1532 .mode = 0644,
1533 .kf_ops = &rdtgroup_kf_single_ops,
1534 .write = rdtgroup_cpus_write,
1535 .seq_show = rdtgroup_cpus_show,
1536 .flags = RFTYPE_FLAGS_CPUS_LIST,
1537 .fflags = RFTYPE_BASE,
1538 },
1539 {
1540 .name = "tasks",
1541 .mode = 0644,
1542 .kf_ops = &rdtgroup_kf_single_ops,
1543 .write = rdtgroup_tasks_write,
1544 .seq_show = rdtgroup_tasks_show,
1545 .fflags = RFTYPE_BASE,
1546 },
1547 {
1548 .name = "schemata",
1549 .mode = 0644,
1550 .kf_ops = &rdtgroup_kf_single_ops,
1551 .write = rdtgroup_schemata_write,
1552 .seq_show = rdtgroup_schemata_show,
1553 .fflags = RF_CTRL_BASE,
1554 },
1555 {
1556 .name = "mode",
1557 .mode = 0644,
1558 .kf_ops = &rdtgroup_kf_single_ops,
1559 .write = rdtgroup_mode_write,
1560 .seq_show = rdtgroup_mode_show,
1561 .fflags = RF_CTRL_BASE,
1562 },
1563 {
1564 .name = "size",
1565 .mode = 0444,
1566 .kf_ops = &rdtgroup_kf_single_ops,
1567 .seq_show = rdtgroup_size_show,
1568 .fflags = RF_CTRL_BASE,
1569 },
1570
1571};
1572
1573static int rdtgroup_add_files(struct kernfs_node *kn, unsigned long fflags)
1574{
1575 struct rftype *rfts, *rft;
1576 int ret, len;
1577
1578 rfts = res_common_files;
1579 len = ARRAY_SIZE(res_common_files);
1580
1581 lockdep_assert_held(&rdtgroup_mutex);
1582
1583 for (rft = rfts; rft < rfts + len; rft++) {
1584 if (rft->fflags && ((fflags & rft->fflags) == rft->fflags)) {
1585 ret = rdtgroup_add_file(kn, rft);
1586 if (ret)
1587 goto error;
1588 }
1589 }
1590
1591 return 0;
1592error:
1593 pr_warn("Failed to add %s, err=%d\n", rft->name, ret);
1594 while (--rft >= rfts) {
1595 if ((fflags & rft->fflags) == rft->fflags)
1596 kernfs_remove_by_name(kn, rft->name);
1597 }
1598 return ret;
1599}
1600
1601static struct rftype *rdtgroup_get_rftype_by_name(const char *name)
1602{
1603 struct rftype *rfts, *rft;
1604 int len;
1605
1606 rfts = res_common_files;
1607 len = ARRAY_SIZE(res_common_files);
1608
1609 for (rft = rfts; rft < rfts + len; rft++) {
1610 if (!strcmp(rft->name, name))
1611 return rft;
1612 }
1613
1614 return NULL;
1615}
1616
1617void __init thread_throttle_mode_init(void)
1618{
1619 struct rftype *rft;
1620
1621 rft = rdtgroup_get_rftype_by_name("thread_throttle_mode");
1622 if (!rft)
1623 return;
1624
1625 rft->fflags = RF_CTRL_INFO | RFTYPE_RES_MB;
1626}
1627
1628/**
1629 * rdtgroup_kn_mode_restrict - Restrict user access to named resctrl file
1630 * @r: The resource group with which the file is associated.
1631 * @name: Name of the file
1632 *
1633 * The permissions of named resctrl file, directory, or link are modified
1634 * to not allow read, write, or execute by any user.
1635 *
1636 * WARNING: This function is intended to communicate to the user that the
1637 * resctrl file has been locked down - that it is not relevant to the
1638 * particular state the system finds itself in. It should not be relied
1639 * on to protect from user access because after the file's permissions
1640 * are restricted the user can still change the permissions using chmod
1641 * from the command line.
1642 *
1643 * Return: 0 on success, <0 on failure.
1644 */
1645int rdtgroup_kn_mode_restrict(struct rdtgroup *r, const char *name)
1646{
1647 struct iattr iattr = {.ia_valid = ATTR_MODE,};
1648 struct kernfs_node *kn;
1649 int ret = 0;
1650
1651 kn = kernfs_find_and_get_ns(r->kn, name, NULL);
1652 if (!kn)
1653 return -ENOENT;
1654
1655 switch (kernfs_type(kn)) {
1656 case KERNFS_DIR:
1657 iattr.ia_mode = S_IFDIR;
1658 break;
1659 case KERNFS_FILE:
1660 iattr.ia_mode = S_IFREG;
1661 break;
1662 case KERNFS_LINK:
1663 iattr.ia_mode = S_IFLNK;
1664 break;
1665 }
1666
1667 ret = kernfs_setattr(kn, &iattr);
1668 kernfs_put(kn);
1669 return ret;
1670}
1671
1672/**
1673 * rdtgroup_kn_mode_restore - Restore user access to named resctrl file
1674 * @r: The resource group with which the file is associated.
1675 * @name: Name of the file
1676 * @mask: Mask of permissions that should be restored
1677 *
1678 * Restore the permissions of the named file. If @name is a directory the
1679 * permissions of its parent will be used.
1680 *
1681 * Return: 0 on success, <0 on failure.
1682 */
1683int rdtgroup_kn_mode_restore(struct rdtgroup *r, const char *name,
1684 umode_t mask)
1685{
1686 struct iattr iattr = {.ia_valid = ATTR_MODE,};
1687 struct kernfs_node *kn, *parent;
1688 struct rftype *rfts, *rft;
1689 int ret, len;
1690
1691 rfts = res_common_files;
1692 len = ARRAY_SIZE(res_common_files);
1693
1694 for (rft = rfts; rft < rfts + len; rft++) {
1695 if (!strcmp(rft->name, name))
1696 iattr.ia_mode = rft->mode & mask;
1697 }
1698
1699 kn = kernfs_find_and_get_ns(r->kn, name, NULL);
1700 if (!kn)
1701 return -ENOENT;
1702
1703 switch (kernfs_type(kn)) {
1704 case KERNFS_DIR:
1705 parent = kernfs_get_parent(kn);
1706 if (parent) {
1707 iattr.ia_mode |= parent->mode;
1708 kernfs_put(parent);
1709 }
1710 iattr.ia_mode |= S_IFDIR;
1711 break;
1712 case KERNFS_FILE:
1713 iattr.ia_mode |= S_IFREG;
1714 break;
1715 case KERNFS_LINK:
1716 iattr.ia_mode |= S_IFLNK;
1717 break;
1718 }
1719
1720 ret = kernfs_setattr(kn, &iattr);
1721 kernfs_put(kn);
1722 return ret;
1723}
1724
1725static int rdtgroup_mkdir_info_resdir(void *priv, char *name,
1726 unsigned long fflags)
1727{
1728 struct kernfs_node *kn_subdir;
1729 int ret;
1730
1731 kn_subdir = kernfs_create_dir(kn_info, name,
1732 kn_info->mode, priv);
1733 if (IS_ERR(kn_subdir))
1734 return PTR_ERR(kn_subdir);
1735
1736 ret = rdtgroup_kn_set_ugid(kn_subdir);
1737 if (ret)
1738 return ret;
1739
1740 ret = rdtgroup_add_files(kn_subdir, fflags);
1741 if (!ret)
1742 kernfs_activate(kn_subdir);
1743
1744 return ret;
1745}
1746
1747static int rdtgroup_create_info_dir(struct kernfs_node *parent_kn)
1748{
1749 struct resctrl_schema *s;
1750 struct rdt_resource *r;
1751 unsigned long fflags;
1752 char name[32];
1753 int ret;
1754
1755 /* create the directory */
1756 kn_info = kernfs_create_dir(parent_kn, "info", parent_kn->mode, NULL);
1757 if (IS_ERR(kn_info))
1758 return PTR_ERR(kn_info);
1759
1760 ret = rdtgroup_add_files(kn_info, RF_TOP_INFO);
1761 if (ret)
1762 goto out_destroy;
1763
1764 /* loop over enabled controls, these are all alloc_capable */
1765 list_for_each_entry(s, &resctrl_schema_all, list) {
1766 r = s->res;
1767 fflags = r->fflags | RF_CTRL_INFO;
1768 ret = rdtgroup_mkdir_info_resdir(s, s->name, fflags);
1769 if (ret)
1770 goto out_destroy;
1771 }
1772
1773 for_each_mon_capable_rdt_resource(r) {
1774 fflags = r->fflags | RF_MON_INFO;
1775 sprintf(name, "%s_MON", r->name);
1776 ret = rdtgroup_mkdir_info_resdir(r, name, fflags);
1777 if (ret)
1778 goto out_destroy;
1779 }
1780
1781 ret = rdtgroup_kn_set_ugid(kn_info);
1782 if (ret)
1783 goto out_destroy;
1784
1785 kernfs_activate(kn_info);
1786
1787 return 0;
1788
1789out_destroy:
1790 kernfs_remove(kn_info);
1791 return ret;
1792}
1793
1794static int
1795mongroup_create_dir(struct kernfs_node *parent_kn, struct rdtgroup *prgrp,
1796 char *name, struct kernfs_node **dest_kn)
1797{
1798 struct kernfs_node *kn;
1799 int ret;
1800
1801 /* create the directory */
1802 kn = kernfs_create_dir(parent_kn, name, parent_kn->mode, prgrp);
1803 if (IS_ERR(kn))
1804 return PTR_ERR(kn);
1805
1806 if (dest_kn)
1807 *dest_kn = kn;
1808
1809 ret = rdtgroup_kn_set_ugid(kn);
1810 if (ret)
1811 goto out_destroy;
1812
1813 kernfs_activate(kn);
1814
1815 return 0;
1816
1817out_destroy:
1818 kernfs_remove(kn);
1819 return ret;
1820}
1821
1822static void l3_qos_cfg_update(void *arg)
1823{
1824 bool *enable = arg;
1825
1826 wrmsrl(MSR_IA32_L3_QOS_CFG, *enable ? L3_QOS_CDP_ENABLE : 0ULL);
1827}
1828
1829static void l2_qos_cfg_update(void *arg)
1830{
1831 bool *enable = arg;
1832
1833 wrmsrl(MSR_IA32_L2_QOS_CFG, *enable ? L2_QOS_CDP_ENABLE : 0ULL);
1834}
1835
1836static inline bool is_mba_linear(void)
1837{
1838 return rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl.membw.delay_linear;
1839}
1840
1841static int set_cache_qos_cfg(int level, bool enable)
1842{
1843 void (*update)(void *arg);
1844 struct rdt_resource *r_l;
1845 cpumask_var_t cpu_mask;
1846 struct rdt_domain *d;
1847 int cpu;
1848
1849 if (level == RDT_RESOURCE_L3)
1850 update = l3_qos_cfg_update;
1851 else if (level == RDT_RESOURCE_L2)
1852 update = l2_qos_cfg_update;
1853 else
1854 return -EINVAL;
1855
1856 if (!zalloc_cpumask_var(&cpu_mask, GFP_KERNEL))
1857 return -ENOMEM;
1858
1859 r_l = &rdt_resources_all[level].r_resctrl;
1860 list_for_each_entry(d, &r_l->domains, list) {
1861 if (r_l->cache.arch_has_per_cpu_cfg)
1862 /* Pick all the CPUs in the domain instance */
1863 for_each_cpu(cpu, &d->cpu_mask)
1864 cpumask_set_cpu(cpu, cpu_mask);
1865 else
1866 /* Pick one CPU from each domain instance to update MSR */
1867 cpumask_set_cpu(cpumask_any(&d->cpu_mask), cpu_mask);
1868 }
1869 cpu = get_cpu();
1870 /* Update QOS_CFG MSR on this cpu if it's in cpu_mask. */
1871 if (cpumask_test_cpu(cpu, cpu_mask))
1872 update(&enable);
1873 /* Update QOS_CFG MSR on all other cpus in cpu_mask. */
1874 smp_call_function_many(cpu_mask, update, &enable, 1);
1875 put_cpu();
1876
1877 free_cpumask_var(cpu_mask);
1878
1879 return 0;
1880}
1881
1882/* Restore the qos cfg state when a domain comes online */
1883void rdt_domain_reconfigure_cdp(struct rdt_resource *r)
1884{
1885 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
1886
1887 if (!r->cdp_capable)
1888 return;
1889
1890 if (r->rid == RDT_RESOURCE_L2)
1891 l2_qos_cfg_update(&hw_res->cdp_enabled);
1892
1893 if (r->rid == RDT_RESOURCE_L3)
1894 l3_qos_cfg_update(&hw_res->cdp_enabled);
1895}
1896
1897static int mba_sc_domain_allocate(struct rdt_resource *r, struct rdt_domain *d)
1898{
1899 u32 num_closid = resctrl_arch_get_num_closid(r);
1900 int cpu = cpumask_any(&d->cpu_mask);
1901 int i;
1902
1903 d->mbps_val = kcalloc_node(num_closid, sizeof(*d->mbps_val),
1904 GFP_KERNEL, cpu_to_node(cpu));
1905 if (!d->mbps_val)
1906 return -ENOMEM;
1907
1908 for (i = 0; i < num_closid; i++)
1909 d->mbps_val[i] = MBA_MAX_MBPS;
1910
1911 return 0;
1912}
1913
1914static void mba_sc_domain_destroy(struct rdt_resource *r,
1915 struct rdt_domain *d)
1916{
1917 kfree(d->mbps_val);
1918 d->mbps_val = NULL;
1919}
1920
1921/*
1922 * MBA software controller is supported only if
1923 * MBM is supported and MBA is in linear scale.
1924 */
1925static bool supports_mba_mbps(void)
1926{
1927 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl;
1928
1929 return (is_mbm_local_enabled() &&
1930 r->alloc_capable && is_mba_linear());
1931}
1932
1933/*
1934 * Enable or disable the MBA software controller
1935 * which helps user specify bandwidth in MBps.
1936 */
1937static int set_mba_sc(bool mba_sc)
1938{
1939 struct rdt_resource *r = &rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl;
1940 u32 num_closid = resctrl_arch_get_num_closid(r);
1941 struct rdt_domain *d;
1942 int i;
1943
1944 if (!supports_mba_mbps() || mba_sc == is_mba_sc(r))
1945 return -EINVAL;
1946
1947 r->membw.mba_sc = mba_sc;
1948
1949 list_for_each_entry(d, &r->domains, list) {
1950 for (i = 0; i < num_closid; i++)
1951 d->mbps_val[i] = MBA_MAX_MBPS;
1952 }
1953
1954 return 0;
1955}
1956
1957static int cdp_enable(int level)
1958{
1959 struct rdt_resource *r_l = &rdt_resources_all[level].r_resctrl;
1960 int ret;
1961
1962 if (!r_l->alloc_capable)
1963 return -EINVAL;
1964
1965 ret = set_cache_qos_cfg(level, true);
1966 if (!ret)
1967 rdt_resources_all[level].cdp_enabled = true;
1968
1969 return ret;
1970}
1971
1972static void cdp_disable(int level)
1973{
1974 struct rdt_hw_resource *r_hw = &rdt_resources_all[level];
1975
1976 if (r_hw->cdp_enabled) {
1977 set_cache_qos_cfg(level, false);
1978 r_hw->cdp_enabled = false;
1979 }
1980}
1981
1982int resctrl_arch_set_cdp_enabled(enum resctrl_res_level l, bool enable)
1983{
1984 struct rdt_hw_resource *hw_res = &rdt_resources_all[l];
1985
1986 if (!hw_res->r_resctrl.cdp_capable)
1987 return -EINVAL;
1988
1989 if (enable)
1990 return cdp_enable(l);
1991
1992 cdp_disable(l);
1993
1994 return 0;
1995}
1996
1997static void cdp_disable_all(void)
1998{
1999 if (resctrl_arch_get_cdp_enabled(RDT_RESOURCE_L3))
2000 resctrl_arch_set_cdp_enabled(RDT_RESOURCE_L3, false);
2001 if (resctrl_arch_get_cdp_enabled(RDT_RESOURCE_L2))
2002 resctrl_arch_set_cdp_enabled(RDT_RESOURCE_L2, false);
2003}
2004
2005/*
2006 * We don't allow rdtgroup directories to be created anywhere
2007 * except the root directory. Thus when looking for the rdtgroup
2008 * structure for a kernfs node we are either looking at a directory,
2009 * in which case the rdtgroup structure is pointed at by the "priv"
2010 * field, otherwise we have a file, and need only look to the parent
2011 * to find the rdtgroup.
2012 */
2013static struct rdtgroup *kernfs_to_rdtgroup(struct kernfs_node *kn)
2014{
2015 if (kernfs_type(kn) == KERNFS_DIR) {
2016 /*
2017 * All the resource directories use "kn->priv"
2018 * to point to the "struct rdtgroup" for the
2019 * resource. "info" and its subdirectories don't
2020 * have rdtgroup structures, so return NULL here.
2021 */
2022 if (kn == kn_info || kn->parent == kn_info)
2023 return NULL;
2024 else
2025 return kn->priv;
2026 } else {
2027 return kn->parent->priv;
2028 }
2029}
2030
2031struct rdtgroup *rdtgroup_kn_lock_live(struct kernfs_node *kn)
2032{
2033 struct rdtgroup *rdtgrp = kernfs_to_rdtgroup(kn);
2034
2035 if (!rdtgrp)
2036 return NULL;
2037
2038 atomic_inc(&rdtgrp->waitcount);
2039 kernfs_break_active_protection(kn);
2040
2041 mutex_lock(&rdtgroup_mutex);
2042
2043 /* Was this group deleted while we waited? */
2044 if (rdtgrp->flags & RDT_DELETED)
2045 return NULL;
2046
2047 return rdtgrp;
2048}
2049
2050void rdtgroup_kn_unlock(struct kernfs_node *kn)
2051{
2052 struct rdtgroup *rdtgrp = kernfs_to_rdtgroup(kn);
2053
2054 if (!rdtgrp)
2055 return;
2056
2057 mutex_unlock(&rdtgroup_mutex);
2058
2059 if (atomic_dec_and_test(&rdtgrp->waitcount) &&
2060 (rdtgrp->flags & RDT_DELETED)) {
2061 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP ||
2062 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED)
2063 rdtgroup_pseudo_lock_remove(rdtgrp);
2064 kernfs_unbreak_active_protection(kn);
2065 rdtgroup_remove(rdtgrp);
2066 } else {
2067 kernfs_unbreak_active_protection(kn);
2068 }
2069}
2070
2071static int mkdir_mondata_all(struct kernfs_node *parent_kn,
2072 struct rdtgroup *prgrp,
2073 struct kernfs_node **mon_data_kn);
2074
2075static int rdt_enable_ctx(struct rdt_fs_context *ctx)
2076{
2077 int ret = 0;
2078
2079 if (ctx->enable_cdpl2)
2080 ret = resctrl_arch_set_cdp_enabled(RDT_RESOURCE_L2, true);
2081
2082 if (!ret && ctx->enable_cdpl3)
2083 ret = resctrl_arch_set_cdp_enabled(RDT_RESOURCE_L3, true);
2084
2085 if (!ret && ctx->enable_mba_mbps)
2086 ret = set_mba_sc(true);
2087
2088 return ret;
2089}
2090
2091static int schemata_list_add(struct rdt_resource *r, enum resctrl_conf_type type)
2092{
2093 struct resctrl_schema *s;
2094 const char *suffix = "";
2095 int ret, cl;
2096
2097 s = kzalloc(sizeof(*s), GFP_KERNEL);
2098 if (!s)
2099 return -ENOMEM;
2100
2101 s->res = r;
2102 s->num_closid = resctrl_arch_get_num_closid(r);
2103 if (resctrl_arch_get_cdp_enabled(r->rid))
2104 s->num_closid /= 2;
2105
2106 s->conf_type = type;
2107 switch (type) {
2108 case CDP_CODE:
2109 suffix = "CODE";
2110 break;
2111 case CDP_DATA:
2112 suffix = "DATA";
2113 break;
2114 case CDP_NONE:
2115 suffix = "";
2116 break;
2117 }
2118
2119 ret = snprintf(s->name, sizeof(s->name), "%s%s", r->name, suffix);
2120 if (ret >= sizeof(s->name)) {
2121 kfree(s);
2122 return -EINVAL;
2123 }
2124
2125 cl = strlen(s->name);
2126
2127 /*
2128 * If CDP is supported by this resource, but not enabled,
2129 * include the suffix. This ensures the tabular format of the
2130 * schemata file does not change between mounts of the filesystem.
2131 */
2132 if (r->cdp_capable && !resctrl_arch_get_cdp_enabled(r->rid))
2133 cl += 4;
2134
2135 if (cl > max_name_width)
2136 max_name_width = cl;
2137
2138 INIT_LIST_HEAD(&s->list);
2139 list_add(&s->list, &resctrl_schema_all);
2140
2141 return 0;
2142}
2143
2144static int schemata_list_create(void)
2145{
2146 struct rdt_resource *r;
2147 int ret = 0;
2148
2149 for_each_alloc_capable_rdt_resource(r) {
2150 if (resctrl_arch_get_cdp_enabled(r->rid)) {
2151 ret = schemata_list_add(r, CDP_CODE);
2152 if (ret)
2153 break;
2154
2155 ret = schemata_list_add(r, CDP_DATA);
2156 } else {
2157 ret = schemata_list_add(r, CDP_NONE);
2158 }
2159
2160 if (ret)
2161 break;
2162 }
2163
2164 return ret;
2165}
2166
2167static void schemata_list_destroy(void)
2168{
2169 struct resctrl_schema *s, *tmp;
2170
2171 list_for_each_entry_safe(s, tmp, &resctrl_schema_all, list) {
2172 list_del(&s->list);
2173 kfree(s);
2174 }
2175}
2176
2177static int rdt_get_tree(struct fs_context *fc)
2178{
2179 struct rdt_fs_context *ctx = rdt_fc2context(fc);
2180 struct rdt_domain *dom;
2181 struct rdt_resource *r;
2182 int ret;
2183
2184 cpus_read_lock();
2185 mutex_lock(&rdtgroup_mutex);
2186 /*
2187 * resctrl file system can only be mounted once.
2188 */
2189 if (static_branch_unlikely(&rdt_enable_key)) {
2190 ret = -EBUSY;
2191 goto out;
2192 }
2193
2194 ret = rdt_enable_ctx(ctx);
2195 if (ret < 0)
2196 goto out_cdp;
2197
2198 ret = schemata_list_create();
2199 if (ret) {
2200 schemata_list_destroy();
2201 goto out_mba;
2202 }
2203
2204 closid_init();
2205
2206 ret = rdtgroup_create_info_dir(rdtgroup_default.kn);
2207 if (ret < 0)
2208 goto out_schemata_free;
2209
2210 if (rdt_mon_capable) {
2211 ret = mongroup_create_dir(rdtgroup_default.kn,
2212 &rdtgroup_default, "mon_groups",
2213 &kn_mongrp);
2214 if (ret < 0)
2215 goto out_info;
2216
2217 ret = mkdir_mondata_all(rdtgroup_default.kn,
2218 &rdtgroup_default, &kn_mondata);
2219 if (ret < 0)
2220 goto out_mongrp;
2221 rdtgroup_default.mon.mon_data_kn = kn_mondata;
2222 }
2223
2224 ret = rdt_pseudo_lock_init();
2225 if (ret)
2226 goto out_mondata;
2227
2228 ret = kernfs_get_tree(fc);
2229 if (ret < 0)
2230 goto out_psl;
2231
2232 if (rdt_alloc_capable)
2233 static_branch_enable_cpuslocked(&rdt_alloc_enable_key);
2234 if (rdt_mon_capable)
2235 static_branch_enable_cpuslocked(&rdt_mon_enable_key);
2236
2237 if (rdt_alloc_capable || rdt_mon_capable)
2238 static_branch_enable_cpuslocked(&rdt_enable_key);
2239
2240 if (is_mbm_enabled()) {
2241 r = &rdt_resources_all[RDT_RESOURCE_L3].r_resctrl;
2242 list_for_each_entry(dom, &r->domains, list)
2243 mbm_setup_overflow_handler(dom, MBM_OVERFLOW_INTERVAL);
2244 }
2245
2246 goto out;
2247
2248out_psl:
2249 rdt_pseudo_lock_release();
2250out_mondata:
2251 if (rdt_mon_capable)
2252 kernfs_remove(kn_mondata);
2253out_mongrp:
2254 if (rdt_mon_capable)
2255 kernfs_remove(kn_mongrp);
2256out_info:
2257 kernfs_remove(kn_info);
2258out_schemata_free:
2259 schemata_list_destroy();
2260out_mba:
2261 if (ctx->enable_mba_mbps)
2262 set_mba_sc(false);
2263out_cdp:
2264 cdp_disable_all();
2265out:
2266 rdt_last_cmd_clear();
2267 mutex_unlock(&rdtgroup_mutex);
2268 cpus_read_unlock();
2269 return ret;
2270}
2271
2272enum rdt_param {
2273 Opt_cdp,
2274 Opt_cdpl2,
2275 Opt_mba_mbps,
2276 nr__rdt_params
2277};
2278
2279static const struct fs_parameter_spec rdt_fs_parameters[] = {
2280 fsparam_flag("cdp", Opt_cdp),
2281 fsparam_flag("cdpl2", Opt_cdpl2),
2282 fsparam_flag("mba_MBps", Opt_mba_mbps),
2283 {}
2284};
2285
2286static int rdt_parse_param(struct fs_context *fc, struct fs_parameter *param)
2287{
2288 struct rdt_fs_context *ctx = rdt_fc2context(fc);
2289 struct fs_parse_result result;
2290 int opt;
2291
2292 opt = fs_parse(fc, rdt_fs_parameters, param, &result);
2293 if (opt < 0)
2294 return opt;
2295
2296 switch (opt) {
2297 case Opt_cdp:
2298 ctx->enable_cdpl3 = true;
2299 return 0;
2300 case Opt_cdpl2:
2301 ctx->enable_cdpl2 = true;
2302 return 0;
2303 case Opt_mba_mbps:
2304 if (!supports_mba_mbps())
2305 return -EINVAL;
2306 ctx->enable_mba_mbps = true;
2307 return 0;
2308 }
2309
2310 return -EINVAL;
2311}
2312
2313static void rdt_fs_context_free(struct fs_context *fc)
2314{
2315 struct rdt_fs_context *ctx = rdt_fc2context(fc);
2316
2317 kernfs_free_fs_context(fc);
2318 kfree(ctx);
2319}
2320
2321static const struct fs_context_operations rdt_fs_context_ops = {
2322 .free = rdt_fs_context_free,
2323 .parse_param = rdt_parse_param,
2324 .get_tree = rdt_get_tree,
2325};
2326
2327static int rdt_init_fs_context(struct fs_context *fc)
2328{
2329 struct rdt_fs_context *ctx;
2330
2331 ctx = kzalloc(sizeof(struct rdt_fs_context), GFP_KERNEL);
2332 if (!ctx)
2333 return -ENOMEM;
2334
2335 ctx->kfc.root = rdt_root;
2336 ctx->kfc.magic = RDTGROUP_SUPER_MAGIC;
2337 fc->fs_private = &ctx->kfc;
2338 fc->ops = &rdt_fs_context_ops;
2339 put_user_ns(fc->user_ns);
2340 fc->user_ns = get_user_ns(&init_user_ns);
2341 fc->global = true;
2342 return 0;
2343}
2344
2345static int reset_all_ctrls(struct rdt_resource *r)
2346{
2347 struct rdt_hw_resource *hw_res = resctrl_to_arch_res(r);
2348 struct rdt_hw_domain *hw_dom;
2349 struct msr_param msr_param;
2350 cpumask_var_t cpu_mask;
2351 struct rdt_domain *d;
2352 int i, cpu;
2353
2354 if (!zalloc_cpumask_var(&cpu_mask, GFP_KERNEL))
2355 return -ENOMEM;
2356
2357 msr_param.res = r;
2358 msr_param.low = 0;
2359 msr_param.high = hw_res->num_closid;
2360
2361 /*
2362 * Disable resource control for this resource by setting all
2363 * CBMs in all domains to the maximum mask value. Pick one CPU
2364 * from each domain to update the MSRs below.
2365 */
2366 list_for_each_entry(d, &r->domains, list) {
2367 hw_dom = resctrl_to_arch_dom(d);
2368 cpumask_set_cpu(cpumask_any(&d->cpu_mask), cpu_mask);
2369
2370 for (i = 0; i < hw_res->num_closid; i++)
2371 hw_dom->ctrl_val[i] = r->default_ctrl;
2372 }
2373 cpu = get_cpu();
2374 /* Update CBM on this cpu if it's in cpu_mask. */
2375 if (cpumask_test_cpu(cpu, cpu_mask))
2376 rdt_ctrl_update(&msr_param);
2377 /* Update CBM on all other cpus in cpu_mask. */
2378 smp_call_function_many(cpu_mask, rdt_ctrl_update, &msr_param, 1);
2379 put_cpu();
2380
2381 free_cpumask_var(cpu_mask);
2382
2383 return 0;
2384}
2385
2386/*
2387 * Move tasks from one to the other group. If @from is NULL, then all tasks
2388 * in the systems are moved unconditionally (used for teardown).
2389 *
2390 * If @mask is not NULL the cpus on which moved tasks are running are set
2391 * in that mask so the update smp function call is restricted to affected
2392 * cpus.
2393 */
2394static void rdt_move_group_tasks(struct rdtgroup *from, struct rdtgroup *to,
2395 struct cpumask *mask)
2396{
2397 struct task_struct *p, *t;
2398
2399 read_lock(&tasklist_lock);
2400 for_each_process_thread(p, t) {
2401 if (!from || is_closid_match(t, from) ||
2402 is_rmid_match(t, from)) {
2403 WRITE_ONCE(t->closid, to->closid);
2404 WRITE_ONCE(t->rmid, to->mon.rmid);
2405
2406 /*
2407 * Order the closid/rmid stores above before the loads
2408 * in task_curr(). This pairs with the full barrier
2409 * between the rq->curr update and resctrl_sched_in()
2410 * during context switch.
2411 */
2412 smp_mb();
2413
2414 /*
2415 * If the task is on a CPU, set the CPU in the mask.
2416 * The detection is inaccurate as tasks might move or
2417 * schedule before the smp function call takes place.
2418 * In such a case the function call is pointless, but
2419 * there is no other side effect.
2420 */
2421 if (IS_ENABLED(CONFIG_SMP) && mask && task_curr(t))
2422 cpumask_set_cpu(task_cpu(t), mask);
2423 }
2424 }
2425 read_unlock(&tasklist_lock);
2426}
2427
2428static void free_all_child_rdtgrp(struct rdtgroup *rdtgrp)
2429{
2430 struct rdtgroup *sentry, *stmp;
2431 struct list_head *head;
2432
2433 head = &rdtgrp->mon.crdtgrp_list;
2434 list_for_each_entry_safe(sentry, stmp, head, mon.crdtgrp_list) {
2435 free_rmid(sentry->mon.rmid);
2436 list_del(&sentry->mon.crdtgrp_list);
2437
2438 if (atomic_read(&sentry->waitcount) != 0)
2439 sentry->flags = RDT_DELETED;
2440 else
2441 rdtgroup_remove(sentry);
2442 }
2443}
2444
2445/*
2446 * Forcibly remove all of subdirectories under root.
2447 */
2448static void rmdir_all_sub(void)
2449{
2450 struct rdtgroup *rdtgrp, *tmp;
2451
2452 /* Move all tasks to the default resource group */
2453 rdt_move_group_tasks(NULL, &rdtgroup_default, NULL);
2454
2455 list_for_each_entry_safe(rdtgrp, tmp, &rdt_all_groups, rdtgroup_list) {
2456 /* Free any child rmids */
2457 free_all_child_rdtgrp(rdtgrp);
2458
2459 /* Remove each rdtgroup other than root */
2460 if (rdtgrp == &rdtgroup_default)
2461 continue;
2462
2463 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP ||
2464 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED)
2465 rdtgroup_pseudo_lock_remove(rdtgrp);
2466
2467 /*
2468 * Give any CPUs back to the default group. We cannot copy
2469 * cpu_online_mask because a CPU might have executed the
2470 * offline callback already, but is still marked online.
2471 */
2472 cpumask_or(&rdtgroup_default.cpu_mask,
2473 &rdtgroup_default.cpu_mask, &rdtgrp->cpu_mask);
2474
2475 free_rmid(rdtgrp->mon.rmid);
2476
2477 kernfs_remove(rdtgrp->kn);
2478 list_del(&rdtgrp->rdtgroup_list);
2479
2480 if (atomic_read(&rdtgrp->waitcount) != 0)
2481 rdtgrp->flags = RDT_DELETED;
2482 else
2483 rdtgroup_remove(rdtgrp);
2484 }
2485 /* Notify online CPUs to update per cpu storage and PQR_ASSOC MSR */
2486 update_closid_rmid(cpu_online_mask, &rdtgroup_default);
2487
2488 kernfs_remove(kn_info);
2489 kernfs_remove(kn_mongrp);
2490 kernfs_remove(kn_mondata);
2491}
2492
2493static void rdt_kill_sb(struct super_block *sb)
2494{
2495 struct rdt_resource *r;
2496
2497 cpus_read_lock();
2498 mutex_lock(&rdtgroup_mutex);
2499
2500 set_mba_sc(false);
2501
2502 /*Put everything back to default values. */
2503 for_each_alloc_capable_rdt_resource(r)
2504 reset_all_ctrls(r);
2505 cdp_disable_all();
2506 rmdir_all_sub();
2507 rdt_pseudo_lock_release();
2508 rdtgroup_default.mode = RDT_MODE_SHAREABLE;
2509 schemata_list_destroy();
2510 static_branch_disable_cpuslocked(&rdt_alloc_enable_key);
2511 static_branch_disable_cpuslocked(&rdt_mon_enable_key);
2512 static_branch_disable_cpuslocked(&rdt_enable_key);
2513 kernfs_kill_sb(sb);
2514 mutex_unlock(&rdtgroup_mutex);
2515 cpus_read_unlock();
2516}
2517
2518static struct file_system_type rdt_fs_type = {
2519 .name = "resctrl",
2520 .init_fs_context = rdt_init_fs_context,
2521 .parameters = rdt_fs_parameters,
2522 .kill_sb = rdt_kill_sb,
2523};
2524
2525static int mon_addfile(struct kernfs_node *parent_kn, const char *name,
2526 void *priv)
2527{
2528 struct kernfs_node *kn;
2529 int ret = 0;
2530
2531 kn = __kernfs_create_file(parent_kn, name, 0444,
2532 GLOBAL_ROOT_UID, GLOBAL_ROOT_GID, 0,
2533 &kf_mondata_ops, priv, NULL, NULL);
2534 if (IS_ERR(kn))
2535 return PTR_ERR(kn);
2536
2537 ret = rdtgroup_kn_set_ugid(kn);
2538 if (ret) {
2539 kernfs_remove(kn);
2540 return ret;
2541 }
2542
2543 return ret;
2544}
2545
2546/*
2547 * Remove all subdirectories of mon_data of ctrl_mon groups
2548 * and monitor groups with given domain id.
2549 */
2550static void rmdir_mondata_subdir_allrdtgrp(struct rdt_resource *r,
2551 unsigned int dom_id)
2552{
2553 struct rdtgroup *prgrp, *crgrp;
2554 char name[32];
2555
2556 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
2557 sprintf(name, "mon_%s_%02d", r->name, dom_id);
2558 kernfs_remove_by_name(prgrp->mon.mon_data_kn, name);
2559
2560 list_for_each_entry(crgrp, &prgrp->mon.crdtgrp_list, mon.crdtgrp_list)
2561 kernfs_remove_by_name(crgrp->mon.mon_data_kn, name);
2562 }
2563}
2564
2565static int mkdir_mondata_subdir(struct kernfs_node *parent_kn,
2566 struct rdt_domain *d,
2567 struct rdt_resource *r, struct rdtgroup *prgrp)
2568{
2569 union mon_data_bits priv;
2570 struct kernfs_node *kn;
2571 struct mon_evt *mevt;
2572 struct rmid_read rr;
2573 char name[32];
2574 int ret;
2575
2576 sprintf(name, "mon_%s_%02d", r->name, d->id);
2577 /* create the directory */
2578 kn = kernfs_create_dir(parent_kn, name, parent_kn->mode, prgrp);
2579 if (IS_ERR(kn))
2580 return PTR_ERR(kn);
2581
2582 ret = rdtgroup_kn_set_ugid(kn);
2583 if (ret)
2584 goto out_destroy;
2585
2586 if (WARN_ON(list_empty(&r->evt_list))) {
2587 ret = -EPERM;
2588 goto out_destroy;
2589 }
2590
2591 priv.u.rid = r->rid;
2592 priv.u.domid = d->id;
2593 list_for_each_entry(mevt, &r->evt_list, list) {
2594 priv.u.evtid = mevt->evtid;
2595 ret = mon_addfile(kn, mevt->name, priv.priv);
2596 if (ret)
2597 goto out_destroy;
2598
2599 if (is_mbm_event(mevt->evtid))
2600 mon_event_read(&rr, r, d, prgrp, mevt->evtid, true);
2601 }
2602 kernfs_activate(kn);
2603 return 0;
2604
2605out_destroy:
2606 kernfs_remove(kn);
2607 return ret;
2608}
2609
2610/*
2611 * Add all subdirectories of mon_data for "ctrl_mon" groups
2612 * and "monitor" groups with given domain id.
2613 */
2614static void mkdir_mondata_subdir_allrdtgrp(struct rdt_resource *r,
2615 struct rdt_domain *d)
2616{
2617 struct kernfs_node *parent_kn;
2618 struct rdtgroup *prgrp, *crgrp;
2619 struct list_head *head;
2620
2621 list_for_each_entry(prgrp, &rdt_all_groups, rdtgroup_list) {
2622 parent_kn = prgrp->mon.mon_data_kn;
2623 mkdir_mondata_subdir(parent_kn, d, r, prgrp);
2624
2625 head = &prgrp->mon.crdtgrp_list;
2626 list_for_each_entry(crgrp, head, mon.crdtgrp_list) {
2627 parent_kn = crgrp->mon.mon_data_kn;
2628 mkdir_mondata_subdir(parent_kn, d, r, crgrp);
2629 }
2630 }
2631}
2632
2633static int mkdir_mondata_subdir_alldom(struct kernfs_node *parent_kn,
2634 struct rdt_resource *r,
2635 struct rdtgroup *prgrp)
2636{
2637 struct rdt_domain *dom;
2638 int ret;
2639
2640 list_for_each_entry(dom, &r->domains, list) {
2641 ret = mkdir_mondata_subdir(parent_kn, dom, r, prgrp);
2642 if (ret)
2643 return ret;
2644 }
2645
2646 return 0;
2647}
2648
2649/*
2650 * This creates a directory mon_data which contains the monitored data.
2651 *
2652 * mon_data has one directory for each domain which are named
2653 * in the format mon_<domain_name>_<domain_id>. For ex: A mon_data
2654 * with L3 domain looks as below:
2655 * ./mon_data:
2656 * mon_L3_00
2657 * mon_L3_01
2658 * mon_L3_02
2659 * ...
2660 *
2661 * Each domain directory has one file per event:
2662 * ./mon_L3_00/:
2663 * llc_occupancy
2664 *
2665 */
2666static int mkdir_mondata_all(struct kernfs_node *parent_kn,
2667 struct rdtgroup *prgrp,
2668 struct kernfs_node **dest_kn)
2669{
2670 struct rdt_resource *r;
2671 struct kernfs_node *kn;
2672 int ret;
2673
2674 /*
2675 * Create the mon_data directory first.
2676 */
2677 ret = mongroup_create_dir(parent_kn, prgrp, "mon_data", &kn);
2678 if (ret)
2679 return ret;
2680
2681 if (dest_kn)
2682 *dest_kn = kn;
2683
2684 /*
2685 * Create the subdirectories for each domain. Note that all events
2686 * in a domain like L3 are grouped into a resource whose domain is L3
2687 */
2688 for_each_mon_capable_rdt_resource(r) {
2689 ret = mkdir_mondata_subdir_alldom(kn, r, prgrp);
2690 if (ret)
2691 goto out_destroy;
2692 }
2693
2694 return 0;
2695
2696out_destroy:
2697 kernfs_remove(kn);
2698 return ret;
2699}
2700
2701/**
2702 * cbm_ensure_valid - Enforce validity on provided CBM
2703 * @_val: Candidate CBM
2704 * @r: RDT resource to which the CBM belongs
2705 *
2706 * The provided CBM represents all cache portions available for use. This
2707 * may be represented by a bitmap that does not consist of contiguous ones
2708 * and thus be an invalid CBM.
2709 * Here the provided CBM is forced to be a valid CBM by only considering
2710 * the first set of contiguous bits as valid and clearing all bits.
2711 * The intention here is to provide a valid default CBM with which a new
2712 * resource group is initialized. The user can follow this with a
2713 * modification to the CBM if the default does not satisfy the
2714 * requirements.
2715 */
2716static u32 cbm_ensure_valid(u32 _val, struct rdt_resource *r)
2717{
2718 unsigned int cbm_len = r->cache.cbm_len;
2719 unsigned long first_bit, zero_bit;
2720 unsigned long val = _val;
2721
2722 if (!val)
2723 return 0;
2724
2725 first_bit = find_first_bit(&val, cbm_len);
2726 zero_bit = find_next_zero_bit(&val, cbm_len, first_bit);
2727
2728 /* Clear any remaining bits to ensure contiguous region */
2729 bitmap_clear(&val, zero_bit, cbm_len - zero_bit);
2730 return (u32)val;
2731}
2732
2733/*
2734 * Initialize cache resources per RDT domain
2735 *
2736 * Set the RDT domain up to start off with all usable allocations. That is,
2737 * all shareable and unused bits. All-zero CBM is invalid.
2738 */
2739static int __init_one_rdt_domain(struct rdt_domain *d, struct resctrl_schema *s,
2740 u32 closid)
2741{
2742 enum resctrl_conf_type peer_type = resctrl_peer_type(s->conf_type);
2743 enum resctrl_conf_type t = s->conf_type;
2744 struct resctrl_staged_config *cfg;
2745 struct rdt_resource *r = s->res;
2746 u32 used_b = 0, unused_b = 0;
2747 unsigned long tmp_cbm;
2748 enum rdtgrp_mode mode;
2749 u32 peer_ctl, ctrl_val;
2750 int i;
2751
2752 cfg = &d->staged_config[t];
2753 cfg->have_new_ctrl = false;
2754 cfg->new_ctrl = r->cache.shareable_bits;
2755 used_b = r->cache.shareable_bits;
2756 for (i = 0; i < closids_supported(); i++) {
2757 if (closid_allocated(i) && i != closid) {
2758 mode = rdtgroup_mode_by_closid(i);
2759 if (mode == RDT_MODE_PSEUDO_LOCKSETUP)
2760 /*
2761 * ctrl values for locksetup aren't relevant
2762 * until the schemata is written, and the mode
2763 * becomes RDT_MODE_PSEUDO_LOCKED.
2764 */
2765 continue;
2766 /*
2767 * If CDP is active include peer domain's
2768 * usage to ensure there is no overlap
2769 * with an exclusive group.
2770 */
2771 if (resctrl_arch_get_cdp_enabled(r->rid))
2772 peer_ctl = resctrl_arch_get_config(r, d, i,
2773 peer_type);
2774 else
2775 peer_ctl = 0;
2776 ctrl_val = resctrl_arch_get_config(r, d, i,
2777 s->conf_type);
2778 used_b |= ctrl_val | peer_ctl;
2779 if (mode == RDT_MODE_SHAREABLE)
2780 cfg->new_ctrl |= ctrl_val | peer_ctl;
2781 }
2782 }
2783 if (d->plr && d->plr->cbm > 0)
2784 used_b |= d->plr->cbm;
2785 unused_b = used_b ^ (BIT_MASK(r->cache.cbm_len) - 1);
2786 unused_b &= BIT_MASK(r->cache.cbm_len) - 1;
2787 cfg->new_ctrl |= unused_b;
2788 /*
2789 * Force the initial CBM to be valid, user can
2790 * modify the CBM based on system availability.
2791 */
2792 cfg->new_ctrl = cbm_ensure_valid(cfg->new_ctrl, r);
2793 /*
2794 * Assign the u32 CBM to an unsigned long to ensure that
2795 * bitmap_weight() does not access out-of-bound memory.
2796 */
2797 tmp_cbm = cfg->new_ctrl;
2798 if (bitmap_weight(&tmp_cbm, r->cache.cbm_len) < r->cache.min_cbm_bits) {
2799 rdt_last_cmd_printf("No space on %s:%d\n", s->name, d->id);
2800 return -ENOSPC;
2801 }
2802 cfg->have_new_ctrl = true;
2803
2804 return 0;
2805}
2806
2807/*
2808 * Initialize cache resources with default values.
2809 *
2810 * A new RDT group is being created on an allocation capable (CAT)
2811 * supporting system. Set this group up to start off with all usable
2812 * allocations.
2813 *
2814 * If there are no more shareable bits available on any domain then
2815 * the entire allocation will fail.
2816 */
2817static int rdtgroup_init_cat(struct resctrl_schema *s, u32 closid)
2818{
2819 struct rdt_domain *d;
2820 int ret;
2821
2822 list_for_each_entry(d, &s->res->domains, list) {
2823 ret = __init_one_rdt_domain(d, s, closid);
2824 if (ret < 0)
2825 return ret;
2826 }
2827
2828 return 0;
2829}
2830
2831/* Initialize MBA resource with default values. */
2832static void rdtgroup_init_mba(struct rdt_resource *r, u32 closid)
2833{
2834 struct resctrl_staged_config *cfg;
2835 struct rdt_domain *d;
2836
2837 list_for_each_entry(d, &r->domains, list) {
2838 if (is_mba_sc(r)) {
2839 d->mbps_val[closid] = MBA_MAX_MBPS;
2840 continue;
2841 }
2842
2843 cfg = &d->staged_config[CDP_NONE];
2844 cfg->new_ctrl = r->default_ctrl;
2845 cfg->have_new_ctrl = true;
2846 }
2847}
2848
2849/* Initialize the RDT group's allocations. */
2850static int rdtgroup_init_alloc(struct rdtgroup *rdtgrp)
2851{
2852 struct resctrl_schema *s;
2853 struct rdt_resource *r;
2854 int ret;
2855
2856 list_for_each_entry(s, &resctrl_schema_all, list) {
2857 r = s->res;
2858 if (r->rid == RDT_RESOURCE_MBA) {
2859 rdtgroup_init_mba(r, rdtgrp->closid);
2860 if (is_mba_sc(r))
2861 continue;
2862 } else {
2863 ret = rdtgroup_init_cat(s, rdtgrp->closid);
2864 if (ret < 0)
2865 return ret;
2866 }
2867
2868 ret = resctrl_arch_update_domains(r, rdtgrp->closid);
2869 if (ret < 0) {
2870 rdt_last_cmd_puts("Failed to initialize allocations\n");
2871 return ret;
2872 }
2873
2874 }
2875
2876 rdtgrp->mode = RDT_MODE_SHAREABLE;
2877
2878 return 0;
2879}
2880
2881static int mkdir_rdt_prepare(struct kernfs_node *parent_kn,
2882 const char *name, umode_t mode,
2883 enum rdt_group_type rtype, struct rdtgroup **r)
2884{
2885 struct rdtgroup *prdtgrp, *rdtgrp;
2886 struct kernfs_node *kn;
2887 uint files = 0;
2888 int ret;
2889
2890 prdtgrp = rdtgroup_kn_lock_live(parent_kn);
2891 if (!prdtgrp) {
2892 ret = -ENODEV;
2893 goto out_unlock;
2894 }
2895
2896 if (rtype == RDTMON_GROUP &&
2897 (prdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP ||
2898 prdtgrp->mode == RDT_MODE_PSEUDO_LOCKED)) {
2899 ret = -EINVAL;
2900 rdt_last_cmd_puts("Pseudo-locking in progress\n");
2901 goto out_unlock;
2902 }
2903
2904 /* allocate the rdtgroup. */
2905 rdtgrp = kzalloc(sizeof(*rdtgrp), GFP_KERNEL);
2906 if (!rdtgrp) {
2907 ret = -ENOSPC;
2908 rdt_last_cmd_puts("Kernel out of memory\n");
2909 goto out_unlock;
2910 }
2911 *r = rdtgrp;
2912 rdtgrp->mon.parent = prdtgrp;
2913 rdtgrp->type = rtype;
2914 INIT_LIST_HEAD(&rdtgrp->mon.crdtgrp_list);
2915
2916 /* kernfs creates the directory for rdtgrp */
2917 kn = kernfs_create_dir(parent_kn, name, mode, rdtgrp);
2918 if (IS_ERR(kn)) {
2919 ret = PTR_ERR(kn);
2920 rdt_last_cmd_puts("kernfs create error\n");
2921 goto out_free_rgrp;
2922 }
2923 rdtgrp->kn = kn;
2924
2925 /*
2926 * kernfs_remove() will drop the reference count on "kn" which
2927 * will free it. But we still need it to stick around for the
2928 * rdtgroup_kn_unlock(kn) call. Take one extra reference here,
2929 * which will be dropped by kernfs_put() in rdtgroup_remove().
2930 */
2931 kernfs_get(kn);
2932
2933 ret = rdtgroup_kn_set_ugid(kn);
2934 if (ret) {
2935 rdt_last_cmd_puts("kernfs perm error\n");
2936 goto out_destroy;
2937 }
2938
2939 files = RFTYPE_BASE | BIT(RF_CTRLSHIFT + rtype);
2940 ret = rdtgroup_add_files(kn, files);
2941 if (ret) {
2942 rdt_last_cmd_puts("kernfs fill error\n");
2943 goto out_destroy;
2944 }
2945
2946 if (rdt_mon_capable) {
2947 ret = alloc_rmid();
2948 if (ret < 0) {
2949 rdt_last_cmd_puts("Out of RMIDs\n");
2950 goto out_destroy;
2951 }
2952 rdtgrp->mon.rmid = ret;
2953
2954 ret = mkdir_mondata_all(kn, rdtgrp, &rdtgrp->mon.mon_data_kn);
2955 if (ret) {
2956 rdt_last_cmd_puts("kernfs subdir error\n");
2957 goto out_idfree;
2958 }
2959 }
2960 kernfs_activate(kn);
2961
2962 /*
2963 * The caller unlocks the parent_kn upon success.
2964 */
2965 return 0;
2966
2967out_idfree:
2968 free_rmid(rdtgrp->mon.rmid);
2969out_destroy:
2970 kernfs_put(rdtgrp->kn);
2971 kernfs_remove(rdtgrp->kn);
2972out_free_rgrp:
2973 kfree(rdtgrp);
2974out_unlock:
2975 rdtgroup_kn_unlock(parent_kn);
2976 return ret;
2977}
2978
2979static void mkdir_rdt_prepare_clean(struct rdtgroup *rgrp)
2980{
2981 kernfs_remove(rgrp->kn);
2982 free_rmid(rgrp->mon.rmid);
2983 rdtgroup_remove(rgrp);
2984}
2985
2986/*
2987 * Create a monitor group under "mon_groups" directory of a control
2988 * and monitor group(ctrl_mon). This is a resource group
2989 * to monitor a subset of tasks and cpus in its parent ctrl_mon group.
2990 */
2991static int rdtgroup_mkdir_mon(struct kernfs_node *parent_kn,
2992 const char *name, umode_t mode)
2993{
2994 struct rdtgroup *rdtgrp, *prgrp;
2995 int ret;
2996
2997 ret = mkdir_rdt_prepare(parent_kn, name, mode, RDTMON_GROUP, &rdtgrp);
2998 if (ret)
2999 return ret;
3000
3001 prgrp = rdtgrp->mon.parent;
3002 rdtgrp->closid = prgrp->closid;
3003
3004 /*
3005 * Add the rdtgrp to the list of rdtgrps the parent
3006 * ctrl_mon group has to track.
3007 */
3008 list_add_tail(&rdtgrp->mon.crdtgrp_list, &prgrp->mon.crdtgrp_list);
3009
3010 rdtgroup_kn_unlock(parent_kn);
3011 return ret;
3012}
3013
3014/*
3015 * These are rdtgroups created under the root directory. Can be used
3016 * to allocate and monitor resources.
3017 */
3018static int rdtgroup_mkdir_ctrl_mon(struct kernfs_node *parent_kn,
3019 const char *name, umode_t mode)
3020{
3021 struct rdtgroup *rdtgrp;
3022 struct kernfs_node *kn;
3023 u32 closid;
3024 int ret;
3025
3026 ret = mkdir_rdt_prepare(parent_kn, name, mode, RDTCTRL_GROUP, &rdtgrp);
3027 if (ret)
3028 return ret;
3029
3030 kn = rdtgrp->kn;
3031 ret = closid_alloc();
3032 if (ret < 0) {
3033 rdt_last_cmd_puts("Out of CLOSIDs\n");
3034 goto out_common_fail;
3035 }
3036 closid = ret;
3037 ret = 0;
3038
3039 rdtgrp->closid = closid;
3040 ret = rdtgroup_init_alloc(rdtgrp);
3041 if (ret < 0)
3042 goto out_id_free;
3043
3044 list_add(&rdtgrp->rdtgroup_list, &rdt_all_groups);
3045
3046 if (rdt_mon_capable) {
3047 /*
3048 * Create an empty mon_groups directory to hold the subset
3049 * of tasks and cpus to monitor.
3050 */
3051 ret = mongroup_create_dir(kn, rdtgrp, "mon_groups", NULL);
3052 if (ret) {
3053 rdt_last_cmd_puts("kernfs subdir error\n");
3054 goto out_del_list;
3055 }
3056 }
3057
3058 goto out_unlock;
3059
3060out_del_list:
3061 list_del(&rdtgrp->rdtgroup_list);
3062out_id_free:
3063 closid_free(closid);
3064out_common_fail:
3065 mkdir_rdt_prepare_clean(rdtgrp);
3066out_unlock:
3067 rdtgroup_kn_unlock(parent_kn);
3068 return ret;
3069}
3070
3071/*
3072 * We allow creating mon groups only with in a directory called "mon_groups"
3073 * which is present in every ctrl_mon group. Check if this is a valid
3074 * "mon_groups" directory.
3075 *
3076 * 1. The directory should be named "mon_groups".
3077 * 2. The mon group itself should "not" be named "mon_groups".
3078 * This makes sure "mon_groups" directory always has a ctrl_mon group
3079 * as parent.
3080 */
3081static bool is_mon_groups(struct kernfs_node *kn, const char *name)
3082{
3083 return (!strcmp(kn->name, "mon_groups") &&
3084 strcmp(name, "mon_groups"));
3085}
3086
3087static int rdtgroup_mkdir(struct kernfs_node *parent_kn, const char *name,
3088 umode_t mode)
3089{
3090 /* Do not accept '\n' to avoid unparsable situation. */
3091 if (strchr(name, '\n'))
3092 return -EINVAL;
3093
3094 /*
3095 * If the parent directory is the root directory and RDT
3096 * allocation is supported, add a control and monitoring
3097 * subdirectory
3098 */
3099 if (rdt_alloc_capable && parent_kn == rdtgroup_default.kn)
3100 return rdtgroup_mkdir_ctrl_mon(parent_kn, name, mode);
3101
3102 /*
3103 * If RDT monitoring is supported and the parent directory is a valid
3104 * "mon_groups" directory, add a monitoring subdirectory.
3105 */
3106 if (rdt_mon_capable && is_mon_groups(parent_kn, name))
3107 return rdtgroup_mkdir_mon(parent_kn, name, mode);
3108
3109 return -EPERM;
3110}
3111
3112static int rdtgroup_rmdir_mon(struct rdtgroup *rdtgrp, cpumask_var_t tmpmask)
3113{
3114 struct rdtgroup *prdtgrp = rdtgrp->mon.parent;
3115 int cpu;
3116
3117 /* Give any tasks back to the parent group */
3118 rdt_move_group_tasks(rdtgrp, prdtgrp, tmpmask);
3119
3120 /* Update per cpu rmid of the moved CPUs first */
3121 for_each_cpu(cpu, &rdtgrp->cpu_mask)
3122 per_cpu(pqr_state.default_rmid, cpu) = prdtgrp->mon.rmid;
3123 /*
3124 * Update the MSR on moved CPUs and CPUs which have moved
3125 * task running on them.
3126 */
3127 cpumask_or(tmpmask, tmpmask, &rdtgrp->cpu_mask);
3128 update_closid_rmid(tmpmask, NULL);
3129
3130 rdtgrp->flags = RDT_DELETED;
3131 free_rmid(rdtgrp->mon.rmid);
3132
3133 /*
3134 * Remove the rdtgrp from the parent ctrl_mon group's list
3135 */
3136 WARN_ON(list_empty(&prdtgrp->mon.crdtgrp_list));
3137 list_del(&rdtgrp->mon.crdtgrp_list);
3138
3139 kernfs_remove(rdtgrp->kn);
3140
3141 return 0;
3142}
3143
3144static int rdtgroup_ctrl_remove(struct rdtgroup *rdtgrp)
3145{
3146 rdtgrp->flags = RDT_DELETED;
3147 list_del(&rdtgrp->rdtgroup_list);
3148
3149 kernfs_remove(rdtgrp->kn);
3150 return 0;
3151}
3152
3153static int rdtgroup_rmdir_ctrl(struct rdtgroup *rdtgrp, cpumask_var_t tmpmask)
3154{
3155 int cpu;
3156
3157 /* Give any tasks back to the default group */
3158 rdt_move_group_tasks(rdtgrp, &rdtgroup_default, tmpmask);
3159
3160 /* Give any CPUs back to the default group */
3161 cpumask_or(&rdtgroup_default.cpu_mask,
3162 &rdtgroup_default.cpu_mask, &rdtgrp->cpu_mask);
3163
3164 /* Update per cpu closid and rmid of the moved CPUs first */
3165 for_each_cpu(cpu, &rdtgrp->cpu_mask) {
3166 per_cpu(pqr_state.default_closid, cpu) = rdtgroup_default.closid;
3167 per_cpu(pqr_state.default_rmid, cpu) = rdtgroup_default.mon.rmid;
3168 }
3169
3170 /*
3171 * Update the MSR on moved CPUs and CPUs which have moved
3172 * task running on them.
3173 */
3174 cpumask_or(tmpmask, tmpmask, &rdtgrp->cpu_mask);
3175 update_closid_rmid(tmpmask, NULL);
3176
3177 closid_free(rdtgrp->closid);
3178 free_rmid(rdtgrp->mon.rmid);
3179
3180 rdtgroup_ctrl_remove(rdtgrp);
3181
3182 /*
3183 * Free all the child monitor group rmids.
3184 */
3185 free_all_child_rdtgrp(rdtgrp);
3186
3187 return 0;
3188}
3189
3190static int rdtgroup_rmdir(struct kernfs_node *kn)
3191{
3192 struct kernfs_node *parent_kn = kn->parent;
3193 struct rdtgroup *rdtgrp;
3194 cpumask_var_t tmpmask;
3195 int ret = 0;
3196
3197 if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL))
3198 return -ENOMEM;
3199
3200 rdtgrp = rdtgroup_kn_lock_live(kn);
3201 if (!rdtgrp) {
3202 ret = -EPERM;
3203 goto out;
3204 }
3205
3206 /*
3207 * If the rdtgroup is a ctrl_mon group and parent directory
3208 * is the root directory, remove the ctrl_mon group.
3209 *
3210 * If the rdtgroup is a mon group and parent directory
3211 * is a valid "mon_groups" directory, remove the mon group.
3212 */
3213 if (rdtgrp->type == RDTCTRL_GROUP && parent_kn == rdtgroup_default.kn &&
3214 rdtgrp != &rdtgroup_default) {
3215 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP ||
3216 rdtgrp->mode == RDT_MODE_PSEUDO_LOCKED) {
3217 ret = rdtgroup_ctrl_remove(rdtgrp);
3218 } else {
3219 ret = rdtgroup_rmdir_ctrl(rdtgrp, tmpmask);
3220 }
3221 } else if (rdtgrp->type == RDTMON_GROUP &&
3222 is_mon_groups(parent_kn, kn->name)) {
3223 ret = rdtgroup_rmdir_mon(rdtgrp, tmpmask);
3224 } else {
3225 ret = -EPERM;
3226 }
3227
3228out:
3229 rdtgroup_kn_unlock(kn);
3230 free_cpumask_var(tmpmask);
3231 return ret;
3232}
3233
3234static int rdtgroup_show_options(struct seq_file *seq, struct kernfs_root *kf)
3235{
3236 if (resctrl_arch_get_cdp_enabled(RDT_RESOURCE_L3))
3237 seq_puts(seq, ",cdp");
3238
3239 if (resctrl_arch_get_cdp_enabled(RDT_RESOURCE_L2))
3240 seq_puts(seq, ",cdpl2");
3241
3242 if (is_mba_sc(&rdt_resources_all[RDT_RESOURCE_MBA].r_resctrl))
3243 seq_puts(seq, ",mba_MBps");
3244
3245 return 0;
3246}
3247
3248static struct kernfs_syscall_ops rdtgroup_kf_syscall_ops = {
3249 .mkdir = rdtgroup_mkdir,
3250 .rmdir = rdtgroup_rmdir,
3251 .show_options = rdtgroup_show_options,
3252};
3253
3254static int __init rdtgroup_setup_root(void)
3255{
3256 int ret;
3257
3258 rdt_root = kernfs_create_root(&rdtgroup_kf_syscall_ops,
3259 KERNFS_ROOT_CREATE_DEACTIVATED |
3260 KERNFS_ROOT_EXTRA_OPEN_PERM_CHECK,
3261 &rdtgroup_default);
3262 if (IS_ERR(rdt_root))
3263 return PTR_ERR(rdt_root);
3264
3265 mutex_lock(&rdtgroup_mutex);
3266
3267 rdtgroup_default.closid = 0;
3268 rdtgroup_default.mon.rmid = 0;
3269 rdtgroup_default.type = RDTCTRL_GROUP;
3270 INIT_LIST_HEAD(&rdtgroup_default.mon.crdtgrp_list);
3271
3272 list_add(&rdtgroup_default.rdtgroup_list, &rdt_all_groups);
3273
3274 ret = rdtgroup_add_files(kernfs_root_to_node(rdt_root), RF_CTRL_BASE);
3275 if (ret) {
3276 kernfs_destroy_root(rdt_root);
3277 goto out;
3278 }
3279
3280 rdtgroup_default.kn = kernfs_root_to_node(rdt_root);
3281 kernfs_activate(rdtgroup_default.kn);
3282
3283out:
3284 mutex_unlock(&rdtgroup_mutex);
3285
3286 return ret;
3287}
3288
3289static void domain_destroy_mon_state(struct rdt_domain *d)
3290{
3291 bitmap_free(d->rmid_busy_llc);
3292 kfree(d->mbm_total);
3293 kfree(d->mbm_local);
3294}
3295
3296void resctrl_offline_domain(struct rdt_resource *r, struct rdt_domain *d)
3297{
3298 lockdep_assert_held(&rdtgroup_mutex);
3299
3300 if (supports_mba_mbps() && r->rid == RDT_RESOURCE_MBA)
3301 mba_sc_domain_destroy(r, d);
3302
3303 if (!r->mon_capable)
3304 return;
3305
3306 /*
3307 * If resctrl is mounted, remove all the
3308 * per domain monitor data directories.
3309 */
3310 if (static_branch_unlikely(&rdt_mon_enable_key))
3311 rmdir_mondata_subdir_allrdtgrp(r, d->id);
3312
3313 if (is_mbm_enabled())
3314 cancel_delayed_work(&d->mbm_over);
3315 if (is_llc_occupancy_enabled() && has_busy_rmid(r, d)) {
3316 /*
3317 * When a package is going down, forcefully
3318 * decrement rmid->ebusy. There is no way to know
3319 * that the L3 was flushed and hence may lead to
3320 * incorrect counts in rare scenarios, but leaving
3321 * the RMID as busy creates RMID leaks if the
3322 * package never comes back.
3323 */
3324 __check_limbo(d, true);
3325 cancel_delayed_work(&d->cqm_limbo);
3326 }
3327
3328 domain_destroy_mon_state(d);
3329}
3330
3331static int domain_setup_mon_state(struct rdt_resource *r, struct rdt_domain *d)
3332{
3333 size_t tsize;
3334
3335 if (is_llc_occupancy_enabled()) {
3336 d->rmid_busy_llc = bitmap_zalloc(r->num_rmid, GFP_KERNEL);
3337 if (!d->rmid_busy_llc)
3338 return -ENOMEM;
3339 }
3340 if (is_mbm_total_enabled()) {
3341 tsize = sizeof(*d->mbm_total);
3342 d->mbm_total = kcalloc(r->num_rmid, tsize, GFP_KERNEL);
3343 if (!d->mbm_total) {
3344 bitmap_free(d->rmid_busy_llc);
3345 return -ENOMEM;
3346 }
3347 }
3348 if (is_mbm_local_enabled()) {
3349 tsize = sizeof(*d->mbm_local);
3350 d->mbm_local = kcalloc(r->num_rmid, tsize, GFP_KERNEL);
3351 if (!d->mbm_local) {
3352 bitmap_free(d->rmid_busy_llc);
3353 kfree(d->mbm_total);
3354 return -ENOMEM;
3355 }
3356 }
3357
3358 return 0;
3359}
3360
3361int resctrl_online_domain(struct rdt_resource *r, struct rdt_domain *d)
3362{
3363 int err;
3364
3365 lockdep_assert_held(&rdtgroup_mutex);
3366
3367 if (supports_mba_mbps() && r->rid == RDT_RESOURCE_MBA)
3368 /* RDT_RESOURCE_MBA is never mon_capable */
3369 return mba_sc_domain_allocate(r, d);
3370
3371 if (!r->mon_capable)
3372 return 0;
3373
3374 err = domain_setup_mon_state(r, d);
3375 if (err)
3376 return err;
3377
3378 if (is_mbm_enabled()) {
3379 INIT_DELAYED_WORK(&d->mbm_over, mbm_handle_overflow);
3380 mbm_setup_overflow_handler(d, MBM_OVERFLOW_INTERVAL);
3381 }
3382
3383 if (is_llc_occupancy_enabled())
3384 INIT_DELAYED_WORK(&d->cqm_limbo, cqm_handle_limbo);
3385
3386 /* If resctrl is mounted, add per domain monitor data directories. */
3387 if (static_branch_unlikely(&rdt_mon_enable_key))
3388 mkdir_mondata_subdir_allrdtgrp(r, d);
3389
3390 return 0;
3391}
3392
3393/*
3394 * rdtgroup_init - rdtgroup initialization
3395 *
3396 * Setup resctrl file system including set up root, create mount point,
3397 * register rdtgroup filesystem, and initialize files under root directory.
3398 *
3399 * Return: 0 on success or -errno
3400 */
3401int __init rdtgroup_init(void)
3402{
3403 int ret = 0;
3404
3405 seq_buf_init(&last_cmd_status, last_cmd_status_buf,
3406 sizeof(last_cmd_status_buf));
3407
3408 ret = rdtgroup_setup_root();
3409 if (ret)
3410 return ret;
3411
3412 ret = sysfs_create_mount_point(fs_kobj, "resctrl");
3413 if (ret)
3414 goto cleanup_root;
3415
3416 ret = register_filesystem(&rdt_fs_type);
3417 if (ret)
3418 goto cleanup_mountpoint;
3419
3420 /*
3421 * Adding the resctrl debugfs directory here may not be ideal since
3422 * it would let the resctrl debugfs directory appear on the debugfs
3423 * filesystem before the resctrl filesystem is mounted.
3424 * It may also be ok since that would enable debugging of RDT before
3425 * resctrl is mounted.
3426 * The reason why the debugfs directory is created here and not in
3427 * rdt_get_tree() is because rdt_get_tree() takes rdtgroup_mutex and
3428 * during the debugfs directory creation also &sb->s_type->i_mutex_key
3429 * (the lockdep class of inode->i_rwsem). Other filesystem
3430 * interactions (eg. SyS_getdents) have the lock ordering:
3431 * &sb->s_type->i_mutex_key --> &mm->mmap_lock
3432 * During mmap(), called with &mm->mmap_lock, the rdtgroup_mutex
3433 * is taken, thus creating dependency:
3434 * &mm->mmap_lock --> rdtgroup_mutex for the latter that can cause
3435 * issues considering the other two lock dependencies.
3436 * By creating the debugfs directory here we avoid a dependency
3437 * that may cause deadlock (even though file operations cannot
3438 * occur until the filesystem is mounted, but I do not know how to
3439 * tell lockdep that).
3440 */
3441 debugfs_resctrl = debugfs_create_dir("resctrl", NULL);
3442
3443 return 0;
3444
3445cleanup_mountpoint:
3446 sysfs_remove_mount_point(fs_kobj, "resctrl");
3447cleanup_root:
3448 kernfs_destroy_root(rdt_root);
3449
3450 return ret;
3451}
3452
3453void __exit rdtgroup_exit(void)
3454{
3455 debugfs_remove_recursive(debugfs_resctrl);
3456 unregister_filesystem(&rdt_fs_type);
3457 sysfs_remove_mount_point(fs_kobj, "resctrl");
3458 kernfs_destroy_root(rdt_root);
3459}