1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Resource Director Technology (RDT)
4  *
5  * Pseudo-locking support built on top of Cache Allocation Technology (CAT)
6  *
7  * Copyright (C) 2018 Intel Corporation
8  *
9  * Author: Reinette Chatre <reinette.chatre@intel.com>
10  */
11 
12 #define pr_fmt(fmt)	KBUILD_MODNAME ": " fmt
13 
14 #include <linux/cacheinfo.h>
15 #include <linux/cpu.h>
16 #include <linux/cpumask.h>
17 #include <linux/debugfs.h>
18 #include <linux/kthread.h>
19 #include <linux/mman.h>
20 #include <linux/perf_event.h>
21 #include <linux/pm_qos.h>
22 #include <linux/slab.h>
23 #include <linux/uaccess.h>
24 
25 #include <asm/cacheflush.h>
26 #include <asm/intel-family.h>
27 #include <asm/resctrl.h>
28 #include <asm/perf_event.h>
29 
30 #include "../../events/perf_event.h" /* For X86_CONFIG() */
31 #include "internal.h"
32 
33 #define CREATE_TRACE_POINTS
34 #include "pseudo_lock_event.h"
35 
36 /*
37  * The bits needed to disable hardware prefetching varies based on the
38  * platform. During initialization we will discover which bits to use.
39  */
40 static u64 prefetch_disable_bits;
41 
42 /*
43  * Major number assigned to and shared by all devices exposing
44  * pseudo-locked regions.
45  */
46 static unsigned int pseudo_lock_major;
47 static unsigned long pseudo_lock_minor_avail = GENMASK(MINORBITS, 0);
48 
49 static char *pseudo_lock_devnode(const struct device *dev, umode_t *mode)
50 {
51 	const struct rdtgroup *rdtgrp;
52 
53 	rdtgrp = dev_get_drvdata(dev);
54 	if (mode)
55 		*mode = 0600;
56 	return kasprintf(GFP_KERNEL, "pseudo_lock/%s", rdtgrp->kn->name);
57 }
58 
59 static const struct class pseudo_lock_class = {
60 	.name = "pseudo_lock",
61 	.devnode = pseudo_lock_devnode,
62 };
63 
64 /**
65  * get_prefetch_disable_bits - prefetch disable bits of supported platforms
66  * @void: It takes no parameters.
67  *
68  * Capture the list of platforms that have been validated to support
69  * pseudo-locking. This includes testing to ensure pseudo-locked regions
70  * with low cache miss rates can be created under variety of load conditions
71  * as well as that these pseudo-locked regions can maintain their low cache
72  * miss rates under variety of load conditions for significant lengths of time.
73  *
74  * After a platform has been validated to support pseudo-locking its
75  * hardware prefetch disable bits are included here as they are documented
76  * in the SDM.
77  *
78  * When adding a platform here also add support for its cache events to
79  * measure_cycles_perf_fn()
80  *
81  * Return:
82  * If platform is supported, the bits to disable hardware prefetchers, 0
83  * if platform is not supported.
84  */
85 static u64 get_prefetch_disable_bits(void)
86 {
87 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL ||
88 	    boot_cpu_data.x86 != 6)
89 		return 0;
90 
91 	switch (boot_cpu_data.x86_model) {
92 	case INTEL_FAM6_BROADWELL_X:
93 		/*
94 		 * SDM defines bits of MSR_MISC_FEATURE_CONTROL register
95 		 * as:
96 		 * 0    L2 Hardware Prefetcher Disable (R/W)
97 		 * 1    L2 Adjacent Cache Line Prefetcher Disable (R/W)
98 		 * 2    DCU Hardware Prefetcher Disable (R/W)
99 		 * 3    DCU IP Prefetcher Disable (R/W)
100 		 * 63:4 Reserved
101 		 */
102 		return 0xF;
103 	case INTEL_FAM6_ATOM_GOLDMONT:
104 	case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
105 		/*
106 		 * SDM defines bits of MSR_MISC_FEATURE_CONTROL register
107 		 * as:
108 		 * 0     L2 Hardware Prefetcher Disable (R/W)
109 		 * 1     Reserved
110 		 * 2     DCU Hardware Prefetcher Disable (R/W)
111 		 * 63:3  Reserved
112 		 */
113 		return 0x5;
114 	}
115 
116 	return 0;
117 }
118 
119 /**
120  * pseudo_lock_minor_get - Obtain available minor number
121  * @minor: Pointer to where new minor number will be stored
122  *
123  * A bitmask is used to track available minor numbers. Here the next free
124  * minor number is marked as unavailable and returned.
125  *
126  * Return: 0 on success, <0 on failure.
127  */
128 static int pseudo_lock_minor_get(unsigned int *minor)
129 {
130 	unsigned long first_bit;
131 
132 	first_bit = find_first_bit(&pseudo_lock_minor_avail, MINORBITS);
133 
134 	if (first_bit == MINORBITS)
135 		return -ENOSPC;
136 
137 	__clear_bit(first_bit, &pseudo_lock_minor_avail);
138 	*minor = first_bit;
139 
140 	return 0;
141 }
142 
143 /**
144  * pseudo_lock_minor_release - Return minor number to available
145  * @minor: The minor number made available
146  */
147 static void pseudo_lock_minor_release(unsigned int minor)
148 {
149 	__set_bit(minor, &pseudo_lock_minor_avail);
150 }
151 
152 /**
153  * region_find_by_minor - Locate a pseudo-lock region by inode minor number
154  * @minor: The minor number of the device representing pseudo-locked region
155  *
156  * When the character device is accessed we need to determine which
157  * pseudo-locked region it belongs to. This is done by matching the minor
158  * number of the device to the pseudo-locked region it belongs.
159  *
160  * Minor numbers are assigned at the time a pseudo-locked region is associated
161  * with a cache instance.
162  *
163  * Return: On success return pointer to resource group owning the pseudo-locked
164  *         region, NULL on failure.
165  */
166 static struct rdtgroup *region_find_by_minor(unsigned int minor)
167 {
168 	struct rdtgroup *rdtgrp, *rdtgrp_match = NULL;
169 
170 	list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
171 		if (rdtgrp->plr && rdtgrp->plr->minor == minor) {
172 			rdtgrp_match = rdtgrp;
173 			break;
174 		}
175 	}
176 	return rdtgrp_match;
177 }
178 
179 /**
180  * struct pseudo_lock_pm_req - A power management QoS request list entry
181  * @list:	Entry within the @pm_reqs list for a pseudo-locked region
182  * @req:	PM QoS request
183  */
184 struct pseudo_lock_pm_req {
185 	struct list_head list;
186 	struct dev_pm_qos_request req;
187 };
188 
189 static void pseudo_lock_cstates_relax(struct pseudo_lock_region *plr)
190 {
191 	struct pseudo_lock_pm_req *pm_req, *next;
192 
193 	list_for_each_entry_safe(pm_req, next, &plr->pm_reqs, list) {
194 		dev_pm_qos_remove_request(&pm_req->req);
195 		list_del(&pm_req->list);
196 		kfree(pm_req);
197 	}
198 }
199 
200 /**
201  * pseudo_lock_cstates_constrain - Restrict cores from entering C6
202  * @plr: Pseudo-locked region
203  *
204  * To prevent the cache from being affected by power management entering
205  * C6 has to be avoided. This is accomplished by requesting a latency
206  * requirement lower than lowest C6 exit latency of all supported
207  * platforms as found in the cpuidle state tables in the intel_idle driver.
208  * At this time it is possible to do so with a single latency requirement
209  * for all supported platforms.
210  *
211  * Since Goldmont is supported, which is affected by X86_BUG_MONITOR,
212  * the ACPI latencies need to be considered while keeping in mind that C2
213  * may be set to map to deeper sleep states. In this case the latency
214  * requirement needs to prevent entering C2 also.
215  *
216  * Return: 0 on success, <0 on failure
217  */
218 static int pseudo_lock_cstates_constrain(struct pseudo_lock_region *plr)
219 {
220 	struct pseudo_lock_pm_req *pm_req;
221 	int cpu;
222 	int ret;
223 
224 	for_each_cpu(cpu, &plr->d->cpu_mask) {
225 		pm_req = kzalloc(sizeof(*pm_req), GFP_KERNEL);
226 		if (!pm_req) {
227 			rdt_last_cmd_puts("Failure to allocate memory for PM QoS\n");
228 			ret = -ENOMEM;
229 			goto out_err;
230 		}
231 		ret = dev_pm_qos_add_request(get_cpu_device(cpu),
232 					     &pm_req->req,
233 					     DEV_PM_QOS_RESUME_LATENCY,
234 					     30);
235 		if (ret < 0) {
236 			rdt_last_cmd_printf("Failed to add latency req CPU%d\n",
237 					    cpu);
238 			kfree(pm_req);
239 			ret = -1;
240 			goto out_err;
241 		}
242 		list_add(&pm_req->list, &plr->pm_reqs);
243 	}
244 
245 	return 0;
246 
247 out_err:
248 	pseudo_lock_cstates_relax(plr);
249 	return ret;
250 }
251 
252 /**
253  * pseudo_lock_region_clear - Reset pseudo-lock region data
254  * @plr: pseudo-lock region
255  *
256  * All content of the pseudo-locked region is reset - any memory allocated
257  * freed.
258  *
259  * Return: void
260  */
261 static void pseudo_lock_region_clear(struct pseudo_lock_region *plr)
262 {
263 	plr->size = 0;
264 	plr->line_size = 0;
265 	kfree(plr->kmem);
266 	plr->kmem = NULL;
267 	plr->s = NULL;
268 	if (plr->d)
269 		plr->d->plr = NULL;
270 	plr->d = NULL;
271 	plr->cbm = 0;
272 	plr->debugfs_dir = NULL;
273 }
274 
275 /**
276  * pseudo_lock_region_init - Initialize pseudo-lock region information
277  * @plr: pseudo-lock region
278  *
279  * Called after user provided a schemata to be pseudo-locked. From the
280  * schemata the &struct pseudo_lock_region is on entry already initialized
281  * with the resource, domain, and capacity bitmask. Here the information
282  * required for pseudo-locking is deduced from this data and &struct
283  * pseudo_lock_region initialized further. This information includes:
284  * - size in bytes of the region to be pseudo-locked
285  * - cache line size to know the stride with which data needs to be accessed
286  *   to be pseudo-locked
287  * - a cpu associated with the cache instance on which the pseudo-locking
288  *   flow can be executed
289  *
290  * Return: 0 on success, <0 on failure. Descriptive error will be written
291  * to last_cmd_status buffer.
292  */
293 static int pseudo_lock_region_init(struct pseudo_lock_region *plr)
294 {
295 	struct cpu_cacheinfo *ci;
296 	int ret;
297 	int i;
298 
299 	/* Pick the first cpu we find that is associated with the cache. */
300 	plr->cpu = cpumask_first(&plr->d->cpu_mask);
301 
302 	if (!cpu_online(plr->cpu)) {
303 		rdt_last_cmd_printf("CPU %u associated with cache not online\n",
304 				    plr->cpu);
305 		ret = -ENODEV;
306 		goto out_region;
307 	}
308 
309 	ci = get_cpu_cacheinfo(plr->cpu);
310 
311 	plr->size = rdtgroup_cbm_to_size(plr->s->res, plr->d, plr->cbm);
312 
313 	for (i = 0; i < ci->num_leaves; i++) {
314 		if (ci->info_list[i].level == plr->s->res->cache_level) {
315 			plr->line_size = ci->info_list[i].coherency_line_size;
316 			return 0;
317 		}
318 	}
319 
320 	ret = -1;
321 	rdt_last_cmd_puts("Unable to determine cache line size\n");
322 out_region:
323 	pseudo_lock_region_clear(plr);
324 	return ret;
325 }
326 
327 /**
328  * pseudo_lock_init - Initialize a pseudo-lock region
329  * @rdtgrp: resource group to which new pseudo-locked region will belong
330  *
331  * A pseudo-locked region is associated with a resource group. When this
332  * association is created the pseudo-locked region is initialized. The
333  * details of the pseudo-locked region are not known at this time so only
334  * allocation is done and association established.
335  *
336  * Return: 0 on success, <0 on failure
337  */
338 static int pseudo_lock_init(struct rdtgroup *rdtgrp)
339 {
340 	struct pseudo_lock_region *plr;
341 
342 	plr = kzalloc(sizeof(*plr), GFP_KERNEL);
343 	if (!plr)
344 		return -ENOMEM;
345 
346 	init_waitqueue_head(&plr->lock_thread_wq);
347 	INIT_LIST_HEAD(&plr->pm_reqs);
348 	rdtgrp->plr = plr;
349 	return 0;
350 }
351 
352 /**
353  * pseudo_lock_region_alloc - Allocate kernel memory that will be pseudo-locked
354  * @plr: pseudo-lock region
355  *
356  * Initialize the details required to set up the pseudo-locked region and
357  * allocate the contiguous memory that will be pseudo-locked to the cache.
358  *
359  * Return: 0 on success, <0 on failure.  Descriptive error will be written
360  * to last_cmd_status buffer.
361  */
362 static int pseudo_lock_region_alloc(struct pseudo_lock_region *plr)
363 {
364 	int ret;
365 
366 	ret = pseudo_lock_region_init(plr);
367 	if (ret < 0)
368 		return ret;
369 
370 	/*
371 	 * We do not yet support contiguous regions larger than
372 	 * KMALLOC_MAX_SIZE.
373 	 */
374 	if (plr->size > KMALLOC_MAX_SIZE) {
375 		rdt_last_cmd_puts("Requested region exceeds maximum size\n");
376 		ret = -E2BIG;
377 		goto out_region;
378 	}
379 
380 	plr->kmem = kzalloc(plr->size, GFP_KERNEL);
381 	if (!plr->kmem) {
382 		rdt_last_cmd_puts("Unable to allocate memory\n");
383 		ret = -ENOMEM;
384 		goto out_region;
385 	}
386 
387 	ret = 0;
388 	goto out;
389 out_region:
390 	pseudo_lock_region_clear(plr);
391 out:
392 	return ret;
393 }
394 
395 /**
396  * pseudo_lock_free - Free a pseudo-locked region
397  * @rdtgrp: resource group to which pseudo-locked region belonged
398  *
399  * The pseudo-locked region's resources have already been released, or not
400  * yet created at this point. Now it can be freed and disassociated from the
401  * resource group.
402  *
403  * Return: void
404  */
405 static void pseudo_lock_free(struct rdtgroup *rdtgrp)
406 {
407 	pseudo_lock_region_clear(rdtgrp->plr);
408 	kfree(rdtgrp->plr);
409 	rdtgrp->plr = NULL;
410 }
411 
412 /**
413  * pseudo_lock_fn - Load kernel memory into cache
414  * @_rdtgrp: resource group to which pseudo-lock region belongs
415  *
416  * This is the core pseudo-locking flow.
417  *
418  * First we ensure that the kernel memory cannot be found in the cache.
419  * Then, while taking care that there will be as little interference as
420  * possible, the memory to be loaded is accessed while core is running
421  * with class of service set to the bitmask of the pseudo-locked region.
422  * After this is complete no future CAT allocations will be allowed to
423  * overlap with this bitmask.
424  *
425  * Local register variables are utilized to ensure that the memory region
426  * to be locked is the only memory access made during the critical locking
427  * loop.
428  *
429  * Return: 0. Waiter on waitqueue will be woken on completion.
430  */
431 static int pseudo_lock_fn(void *_rdtgrp)
432 {
433 	struct rdtgroup *rdtgrp = _rdtgrp;
434 	struct pseudo_lock_region *plr = rdtgrp->plr;
435 	u32 rmid_p, closid_p;
436 	unsigned long i;
437 	u64 saved_msr;
438 #ifdef CONFIG_KASAN
439 	/*
440 	 * The registers used for local register variables are also used
441 	 * when KASAN is active. When KASAN is active we use a regular
442 	 * variable to ensure we always use a valid pointer, but the cost
443 	 * is that this variable will enter the cache through evicting the
444 	 * memory we are trying to lock into the cache. Thus expect lower
445 	 * pseudo-locking success rate when KASAN is active.
446 	 */
447 	unsigned int line_size;
448 	unsigned int size;
449 	void *mem_r;
450 #else
451 	register unsigned int line_size asm("esi");
452 	register unsigned int size asm("edi");
453 	register void *mem_r asm(_ASM_BX);
454 #endif /* CONFIG_KASAN */
455 
456 	/*
457 	 * Make sure none of the allocated memory is cached. If it is we
458 	 * will get a cache hit in below loop from outside of pseudo-locked
459 	 * region.
460 	 * wbinvd (as opposed to clflush/clflushopt) is required to
461 	 * increase likelihood that allocated cache portion will be filled
462 	 * with associated memory.
463 	 */
464 	native_wbinvd();
465 
466 	/*
467 	 * Always called with interrupts enabled. By disabling interrupts
468 	 * ensure that we will not be preempted during this critical section.
469 	 */
470 	local_irq_disable();
471 
472 	/*
473 	 * Call wrmsr and rdmsr as directly as possible to avoid tracing
474 	 * clobbering local register variables or affecting cache accesses.
475 	 *
476 	 * Disable the hardware prefetcher so that when the end of the memory
477 	 * being pseudo-locked is reached the hardware will not read beyond
478 	 * the buffer and evict pseudo-locked memory read earlier from the
479 	 * cache.
480 	 */
481 	saved_msr = __rdmsr(MSR_MISC_FEATURE_CONTROL);
482 	__wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
483 	closid_p = this_cpu_read(pqr_state.cur_closid);
484 	rmid_p = this_cpu_read(pqr_state.cur_rmid);
485 	mem_r = plr->kmem;
486 	size = plr->size;
487 	line_size = plr->line_size;
488 	/*
489 	 * Critical section begin: start by writing the closid associated
490 	 * with the capacity bitmask of the cache region being
491 	 * pseudo-locked followed by reading of kernel memory to load it
492 	 * into the cache.
493 	 */
494 	__wrmsr(MSR_IA32_PQR_ASSOC, rmid_p, rdtgrp->closid);
495 	/*
496 	 * Cache was flushed earlier. Now access kernel memory to read it
497 	 * into cache region associated with just activated plr->closid.
498 	 * Loop over data twice:
499 	 * - In first loop the cache region is shared with the page walker
500 	 *   as it populates the paging structure caches (including TLB).
501 	 * - In the second loop the paging structure caches are used and
502 	 *   cache region is populated with the memory being referenced.
503 	 */
504 	for (i = 0; i < size; i += PAGE_SIZE) {
505 		/*
506 		 * Add a barrier to prevent speculative execution of this
507 		 * loop reading beyond the end of the buffer.
508 		 */
509 		rmb();
510 		asm volatile("mov (%0,%1,1), %%eax\n\t"
511 			:
512 			: "r" (mem_r), "r" (i)
513 			: "%eax", "memory");
514 	}
515 	for (i = 0; i < size; i += line_size) {
516 		/*
517 		 * Add a barrier to prevent speculative execution of this
518 		 * loop reading beyond the end of the buffer.
519 		 */
520 		rmb();
521 		asm volatile("mov (%0,%1,1), %%eax\n\t"
522 			:
523 			: "r" (mem_r), "r" (i)
524 			: "%eax", "memory");
525 	}
526 	/*
527 	 * Critical section end: restore closid with capacity bitmask that
528 	 * does not overlap with pseudo-locked region.
529 	 */
530 	__wrmsr(MSR_IA32_PQR_ASSOC, rmid_p, closid_p);
531 
532 	/* Re-enable the hardware prefetcher(s) */
533 	wrmsrl(MSR_MISC_FEATURE_CONTROL, saved_msr);
534 	local_irq_enable();
535 
536 	plr->thread_done = 1;
537 	wake_up_interruptible(&plr->lock_thread_wq);
538 	return 0;
539 }
540 
541 /**
542  * rdtgroup_monitor_in_progress - Test if monitoring in progress
543  * @rdtgrp: resource group being queried
544  *
545  * Return: 1 if monitor groups have been created for this resource
546  * group, 0 otherwise.
547  */
548 static int rdtgroup_monitor_in_progress(struct rdtgroup *rdtgrp)
549 {
550 	return !list_empty(&rdtgrp->mon.crdtgrp_list);
551 }
552 
553 /**
554  * rdtgroup_locksetup_user_restrict - Restrict user access to group
555  * @rdtgrp: resource group needing access restricted
556  *
557  * A resource group used for cache pseudo-locking cannot have cpus or tasks
558  * assigned to it. This is communicated to the user by restricting access
559  * to all the files that can be used to make such changes.
560  *
561  * Permissions restored with rdtgroup_locksetup_user_restore()
562  *
563  * Return: 0 on success, <0 on failure. If a failure occurs during the
564  * restriction of access an attempt will be made to restore permissions but
565  * the state of the mode of these files will be uncertain when a failure
566  * occurs.
567  */
568 static int rdtgroup_locksetup_user_restrict(struct rdtgroup *rdtgrp)
569 {
570 	int ret;
571 
572 	ret = rdtgroup_kn_mode_restrict(rdtgrp, "tasks");
573 	if (ret)
574 		return ret;
575 
576 	ret = rdtgroup_kn_mode_restrict(rdtgrp, "cpus");
577 	if (ret)
578 		goto err_tasks;
579 
580 	ret = rdtgroup_kn_mode_restrict(rdtgrp, "cpus_list");
581 	if (ret)
582 		goto err_cpus;
583 
584 	if (rdt_mon_capable) {
585 		ret = rdtgroup_kn_mode_restrict(rdtgrp, "mon_groups");
586 		if (ret)
587 			goto err_cpus_list;
588 	}
589 
590 	ret = 0;
591 	goto out;
592 
593 err_cpus_list:
594 	rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0777);
595 err_cpus:
596 	rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0777);
597 err_tasks:
598 	rdtgroup_kn_mode_restore(rdtgrp, "tasks", 0777);
599 out:
600 	return ret;
601 }
602 
603 /**
604  * rdtgroup_locksetup_user_restore - Restore user access to group
605  * @rdtgrp: resource group needing access restored
606  *
607  * Restore all file access previously removed using
608  * rdtgroup_locksetup_user_restrict()
609  *
610  * Return: 0 on success, <0 on failure.  If a failure occurs during the
611  * restoration of access an attempt will be made to restrict permissions
612  * again but the state of the mode of these files will be uncertain when
613  * a failure occurs.
614  */
615 static int rdtgroup_locksetup_user_restore(struct rdtgroup *rdtgrp)
616 {
617 	int ret;
618 
619 	ret = rdtgroup_kn_mode_restore(rdtgrp, "tasks", 0777);
620 	if (ret)
621 		return ret;
622 
623 	ret = rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0777);
624 	if (ret)
625 		goto err_tasks;
626 
627 	ret = rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0777);
628 	if (ret)
629 		goto err_cpus;
630 
631 	if (rdt_mon_capable) {
632 		ret = rdtgroup_kn_mode_restore(rdtgrp, "mon_groups", 0777);
633 		if (ret)
634 			goto err_cpus_list;
635 	}
636 
637 	ret = 0;
638 	goto out;
639 
640 err_cpus_list:
641 	rdtgroup_kn_mode_restrict(rdtgrp, "cpus_list");
642 err_cpus:
643 	rdtgroup_kn_mode_restrict(rdtgrp, "cpus");
644 err_tasks:
645 	rdtgroup_kn_mode_restrict(rdtgrp, "tasks");
646 out:
647 	return ret;
648 }
649 
650 /**
651  * rdtgroup_locksetup_enter - Resource group enters locksetup mode
652  * @rdtgrp: resource group requested to enter locksetup mode
653  *
654  * A resource group enters locksetup mode to reflect that it would be used
655  * to represent a pseudo-locked region and is in the process of being set
656  * up to do so. A resource group used for a pseudo-locked region would
657  * lose the closid associated with it so we cannot allow it to have any
658  * tasks or cpus assigned nor permit tasks or cpus to be assigned in the
659  * future. Monitoring of a pseudo-locked region is not allowed either.
660  *
661  * The above and more restrictions on a pseudo-locked region are checked
662  * for and enforced before the resource group enters the locksetup mode.
663  *
664  * Returns: 0 if the resource group successfully entered locksetup mode, <0
665  * on failure. On failure the last_cmd_status buffer is updated with text to
666  * communicate details of failure to the user.
667  */
668 int rdtgroup_locksetup_enter(struct rdtgroup *rdtgrp)
669 {
670 	int ret;
671 
672 	/*
673 	 * The default resource group can neither be removed nor lose the
674 	 * default closid associated with it.
675 	 */
676 	if (rdtgrp == &rdtgroup_default) {
677 		rdt_last_cmd_puts("Cannot pseudo-lock default group\n");
678 		return -EINVAL;
679 	}
680 
681 	/*
682 	 * Cache Pseudo-locking not supported when CDP is enabled.
683 	 *
684 	 * Some things to consider if you would like to enable this
685 	 * support (using L3 CDP as example):
686 	 * - When CDP is enabled two separate resources are exposed,
687 	 *   L3DATA and L3CODE, but they are actually on the same cache.
688 	 *   The implication for pseudo-locking is that if a
689 	 *   pseudo-locked region is created on a domain of one
690 	 *   resource (eg. L3CODE), then a pseudo-locked region cannot
691 	 *   be created on that same domain of the other resource
692 	 *   (eg. L3DATA). This is because the creation of a
693 	 *   pseudo-locked region involves a call to wbinvd that will
694 	 *   affect all cache allocations on particular domain.
695 	 * - Considering the previous, it may be possible to only
696 	 *   expose one of the CDP resources to pseudo-locking and
697 	 *   hide the other. For example, we could consider to only
698 	 *   expose L3DATA and since the L3 cache is unified it is
699 	 *   still possible to place instructions there are execute it.
700 	 * - If only one region is exposed to pseudo-locking we should
701 	 *   still keep in mind that availability of a portion of cache
702 	 *   for pseudo-locking should take into account both resources.
703 	 *   Similarly, if a pseudo-locked region is created in one
704 	 *   resource, the portion of cache used by it should be made
705 	 *   unavailable to all future allocations from both resources.
706 	 */
707 	if (resctrl_arch_get_cdp_enabled(RDT_RESOURCE_L3) ||
708 	    resctrl_arch_get_cdp_enabled(RDT_RESOURCE_L2)) {
709 		rdt_last_cmd_puts("CDP enabled\n");
710 		return -EINVAL;
711 	}
712 
713 	/*
714 	 * Not knowing the bits to disable prefetching implies that this
715 	 * platform does not support Cache Pseudo-Locking.
716 	 */
717 	prefetch_disable_bits = get_prefetch_disable_bits();
718 	if (prefetch_disable_bits == 0) {
719 		rdt_last_cmd_puts("Pseudo-locking not supported\n");
720 		return -EINVAL;
721 	}
722 
723 	if (rdtgroup_monitor_in_progress(rdtgrp)) {
724 		rdt_last_cmd_puts("Monitoring in progress\n");
725 		return -EINVAL;
726 	}
727 
728 	if (rdtgroup_tasks_assigned(rdtgrp)) {
729 		rdt_last_cmd_puts("Tasks assigned to resource group\n");
730 		return -EINVAL;
731 	}
732 
733 	if (!cpumask_empty(&rdtgrp->cpu_mask)) {
734 		rdt_last_cmd_puts("CPUs assigned to resource group\n");
735 		return -EINVAL;
736 	}
737 
738 	if (rdtgroup_locksetup_user_restrict(rdtgrp)) {
739 		rdt_last_cmd_puts("Unable to modify resctrl permissions\n");
740 		return -EIO;
741 	}
742 
743 	ret = pseudo_lock_init(rdtgrp);
744 	if (ret) {
745 		rdt_last_cmd_puts("Unable to init pseudo-lock region\n");
746 		goto out_release;
747 	}
748 
749 	/*
750 	 * If this system is capable of monitoring a rmid would have been
751 	 * allocated when the control group was created. This is not needed
752 	 * anymore when this group would be used for pseudo-locking. This
753 	 * is safe to call on platforms not capable of monitoring.
754 	 */
755 	free_rmid(rdtgrp->mon.rmid);
756 
757 	ret = 0;
758 	goto out;
759 
760 out_release:
761 	rdtgroup_locksetup_user_restore(rdtgrp);
762 out:
763 	return ret;
764 }
765 
766 /**
767  * rdtgroup_locksetup_exit - resource group exist locksetup mode
768  * @rdtgrp: resource group
769  *
770  * When a resource group exits locksetup mode the earlier restrictions are
771  * lifted.
772  *
773  * Return: 0 on success, <0 on failure
774  */
775 int rdtgroup_locksetup_exit(struct rdtgroup *rdtgrp)
776 {
777 	int ret;
778 
779 	if (rdt_mon_capable) {
780 		ret = alloc_rmid();
781 		if (ret < 0) {
782 			rdt_last_cmd_puts("Out of RMIDs\n");
783 			return ret;
784 		}
785 		rdtgrp->mon.rmid = ret;
786 	}
787 
788 	ret = rdtgroup_locksetup_user_restore(rdtgrp);
789 	if (ret) {
790 		free_rmid(rdtgrp->mon.rmid);
791 		return ret;
792 	}
793 
794 	pseudo_lock_free(rdtgrp);
795 	return 0;
796 }
797 
798 /**
799  * rdtgroup_cbm_overlaps_pseudo_locked - Test if CBM or portion is pseudo-locked
800  * @d: RDT domain
801  * @cbm: CBM to test
802  *
803  * @d represents a cache instance and @cbm a capacity bitmask that is
804  * considered for it. Determine if @cbm overlaps with any existing
805  * pseudo-locked region on @d.
806  *
807  * @cbm is unsigned long, even if only 32 bits are used, to make the
808  * bitmap functions work correctly.
809  *
810  * Return: true if @cbm overlaps with pseudo-locked region on @d, false
811  * otherwise.
812  */
813 bool rdtgroup_cbm_overlaps_pseudo_locked(struct rdt_domain *d, unsigned long cbm)
814 {
815 	unsigned int cbm_len;
816 	unsigned long cbm_b;
817 
818 	if (d->plr) {
819 		cbm_len = d->plr->s->res->cache.cbm_len;
820 		cbm_b = d->plr->cbm;
821 		if (bitmap_intersects(&cbm, &cbm_b, cbm_len))
822 			return true;
823 	}
824 	return false;
825 }
826 
827 /**
828  * rdtgroup_pseudo_locked_in_hierarchy - Pseudo-locked region in cache hierarchy
829  * @d: RDT domain under test
830  *
831  * The setup of a pseudo-locked region affects all cache instances within
832  * the hierarchy of the region. It is thus essential to know if any
833  * pseudo-locked regions exist within a cache hierarchy to prevent any
834  * attempts to create new pseudo-locked regions in the same hierarchy.
835  *
836  * Return: true if a pseudo-locked region exists in the hierarchy of @d or
837  *         if it is not possible to test due to memory allocation issue,
838  *         false otherwise.
839  */
840 bool rdtgroup_pseudo_locked_in_hierarchy(struct rdt_domain *d)
841 {
842 	cpumask_var_t cpu_with_psl;
843 	struct rdt_resource *r;
844 	struct rdt_domain *d_i;
845 	bool ret = false;
846 
847 	if (!zalloc_cpumask_var(&cpu_with_psl, GFP_KERNEL))
848 		return true;
849 
850 	/*
851 	 * First determine which cpus have pseudo-locked regions
852 	 * associated with them.
853 	 */
854 	for_each_alloc_capable_rdt_resource(r) {
855 		list_for_each_entry(d_i, &r->domains, list) {
856 			if (d_i->plr)
857 				cpumask_or(cpu_with_psl, cpu_with_psl,
858 					   &d_i->cpu_mask);
859 		}
860 	}
861 
862 	/*
863 	 * Next test if new pseudo-locked region would intersect with
864 	 * existing region.
865 	 */
866 	if (cpumask_intersects(&d->cpu_mask, cpu_with_psl))
867 		ret = true;
868 
869 	free_cpumask_var(cpu_with_psl);
870 	return ret;
871 }
872 
873 /**
874  * measure_cycles_lat_fn - Measure cycle latency to read pseudo-locked memory
875  * @_plr: pseudo-lock region to measure
876  *
877  * There is no deterministic way to test if a memory region is cached. One
878  * way is to measure how long it takes to read the memory, the speed of
879  * access is a good way to learn how close to the cpu the data was. Even
880  * more, if the prefetcher is disabled and the memory is read at a stride
881  * of half the cache line, then a cache miss will be easy to spot since the
882  * read of the first half would be significantly slower than the read of
883  * the second half.
884  *
885  * Return: 0. Waiter on waitqueue will be woken on completion.
886  */
887 static int measure_cycles_lat_fn(void *_plr)
888 {
889 	struct pseudo_lock_region *plr = _plr;
890 	u32 saved_low, saved_high;
891 	unsigned long i;
892 	u64 start, end;
893 	void *mem_r;
894 
895 	local_irq_disable();
896 	/*
897 	 * Disable hardware prefetchers.
898 	 */
899 	rdmsr(MSR_MISC_FEATURE_CONTROL, saved_low, saved_high);
900 	wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
901 	mem_r = READ_ONCE(plr->kmem);
902 	/*
903 	 * Dummy execute of the time measurement to load the needed
904 	 * instructions into the L1 instruction cache.
905 	 */
906 	start = rdtsc_ordered();
907 	for (i = 0; i < plr->size; i += 32) {
908 		start = rdtsc_ordered();
909 		asm volatile("mov (%0,%1,1), %%eax\n\t"
910 			     :
911 			     : "r" (mem_r), "r" (i)
912 			     : "%eax", "memory");
913 		end = rdtsc_ordered();
914 		trace_pseudo_lock_mem_latency((u32)(end - start));
915 	}
916 	wrmsr(MSR_MISC_FEATURE_CONTROL, saved_low, saved_high);
917 	local_irq_enable();
918 	plr->thread_done = 1;
919 	wake_up_interruptible(&plr->lock_thread_wq);
920 	return 0;
921 }
922 
923 /*
924  * Create a perf_event_attr for the hit and miss perf events that will
925  * be used during the performance measurement. A perf_event maintains
926  * a pointer to its perf_event_attr so a unique attribute structure is
927  * created for each perf_event.
928  *
929  * The actual configuration of the event is set right before use in order
930  * to use the X86_CONFIG macro.
931  */
932 static struct perf_event_attr perf_miss_attr = {
933 	.type		= PERF_TYPE_RAW,
934 	.size		= sizeof(struct perf_event_attr),
935 	.pinned		= 1,
936 	.disabled	= 0,
937 	.exclude_user	= 1,
938 };
939 
940 static struct perf_event_attr perf_hit_attr = {
941 	.type		= PERF_TYPE_RAW,
942 	.size		= sizeof(struct perf_event_attr),
943 	.pinned		= 1,
944 	.disabled	= 0,
945 	.exclude_user	= 1,
946 };
947 
948 struct residency_counts {
949 	u64 miss_before, hits_before;
950 	u64 miss_after,  hits_after;
951 };
952 
953 static int measure_residency_fn(struct perf_event_attr *miss_attr,
954 				struct perf_event_attr *hit_attr,
955 				struct pseudo_lock_region *plr,
956 				struct residency_counts *counts)
957 {
958 	u64 hits_before = 0, hits_after = 0, miss_before = 0, miss_after = 0;
959 	struct perf_event *miss_event, *hit_event;
960 	int hit_pmcnum, miss_pmcnum;
961 	u32 saved_low, saved_high;
962 	unsigned int line_size;
963 	unsigned int size;
964 	unsigned long i;
965 	void *mem_r;
966 	u64 tmp;
967 
968 	miss_event = perf_event_create_kernel_counter(miss_attr, plr->cpu,
969 						      NULL, NULL, NULL);
970 	if (IS_ERR(miss_event))
971 		goto out;
972 
973 	hit_event = perf_event_create_kernel_counter(hit_attr, plr->cpu,
974 						     NULL, NULL, NULL);
975 	if (IS_ERR(hit_event))
976 		goto out_miss;
977 
978 	local_irq_disable();
979 	/*
980 	 * Check any possible error state of events used by performing
981 	 * one local read.
982 	 */
983 	if (perf_event_read_local(miss_event, &tmp, NULL, NULL)) {
984 		local_irq_enable();
985 		goto out_hit;
986 	}
987 	if (perf_event_read_local(hit_event, &tmp, NULL, NULL)) {
988 		local_irq_enable();
989 		goto out_hit;
990 	}
991 
992 	/*
993 	 * Disable hardware prefetchers.
994 	 */
995 	rdmsr(MSR_MISC_FEATURE_CONTROL, saved_low, saved_high);
996 	wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
997 
998 	/* Initialize rest of local variables */
999 	/*
1000 	 * Performance event has been validated right before this with
1001 	 * interrupts disabled - it is thus safe to read the counter index.
1002 	 */
1003 	miss_pmcnum = x86_perf_rdpmc_index(miss_event);
1004 	hit_pmcnum = x86_perf_rdpmc_index(hit_event);
1005 	line_size = READ_ONCE(plr->line_size);
1006 	mem_r = READ_ONCE(plr->kmem);
1007 	size = READ_ONCE(plr->size);
1008 
1009 	/*
1010 	 * Read counter variables twice - first to load the instructions
1011 	 * used in L1 cache, second to capture accurate value that does not
1012 	 * include cache misses incurred because of instruction loads.
1013 	 */
1014 	rdpmcl(hit_pmcnum, hits_before);
1015 	rdpmcl(miss_pmcnum, miss_before);
1016 	/*
1017 	 * From SDM: Performing back-to-back fast reads are not guaranteed
1018 	 * to be monotonic.
1019 	 * Use LFENCE to ensure all previous instructions are retired
1020 	 * before proceeding.
1021 	 */
1022 	rmb();
1023 	rdpmcl(hit_pmcnum, hits_before);
1024 	rdpmcl(miss_pmcnum, miss_before);
1025 	/*
1026 	 * Use LFENCE to ensure all previous instructions are retired
1027 	 * before proceeding.
1028 	 */
1029 	rmb();
1030 	for (i = 0; i < size; i += line_size) {
1031 		/*
1032 		 * Add a barrier to prevent speculative execution of this
1033 		 * loop reading beyond the end of the buffer.
1034 		 */
1035 		rmb();
1036 		asm volatile("mov (%0,%1,1), %%eax\n\t"
1037 			     :
1038 			     : "r" (mem_r), "r" (i)
1039 			     : "%eax", "memory");
1040 	}
1041 	/*
1042 	 * Use LFENCE to ensure all previous instructions are retired
1043 	 * before proceeding.
1044 	 */
1045 	rmb();
1046 	rdpmcl(hit_pmcnum, hits_after);
1047 	rdpmcl(miss_pmcnum, miss_after);
1048 	/*
1049 	 * Use LFENCE to ensure all previous instructions are retired
1050 	 * before proceeding.
1051 	 */
1052 	rmb();
1053 	/* Re-enable hardware prefetchers */
1054 	wrmsr(MSR_MISC_FEATURE_CONTROL, saved_low, saved_high);
1055 	local_irq_enable();
1056 out_hit:
1057 	perf_event_release_kernel(hit_event);
1058 out_miss:
1059 	perf_event_release_kernel(miss_event);
1060 out:
1061 	/*
1062 	 * All counts will be zero on failure.
1063 	 */
1064 	counts->miss_before = miss_before;
1065 	counts->hits_before = hits_before;
1066 	counts->miss_after  = miss_after;
1067 	counts->hits_after  = hits_after;
1068 	return 0;
1069 }
1070 
1071 static int measure_l2_residency(void *_plr)
1072 {
1073 	struct pseudo_lock_region *plr = _plr;
1074 	struct residency_counts counts = {0};
1075 
1076 	/*
1077 	 * Non-architectural event for the Goldmont Microarchitecture
1078 	 * from Intel x86 Architecture Software Developer Manual (SDM):
1079 	 * MEM_LOAD_UOPS_RETIRED D1H (event number)
1080 	 * Umask values:
1081 	 *     L2_HIT   02H
1082 	 *     L2_MISS  10H
1083 	 */
1084 	switch (boot_cpu_data.x86_model) {
1085 	case INTEL_FAM6_ATOM_GOLDMONT:
1086 	case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
1087 		perf_miss_attr.config = X86_CONFIG(.event = 0xd1,
1088 						   .umask = 0x10);
1089 		perf_hit_attr.config = X86_CONFIG(.event = 0xd1,
1090 						  .umask = 0x2);
1091 		break;
1092 	default:
1093 		goto out;
1094 	}
1095 
1096 	measure_residency_fn(&perf_miss_attr, &perf_hit_attr, plr, &counts);
1097 	/*
1098 	 * If a failure prevented the measurements from succeeding
1099 	 * tracepoints will still be written and all counts will be zero.
1100 	 */
1101 	trace_pseudo_lock_l2(counts.hits_after - counts.hits_before,
1102 			     counts.miss_after - counts.miss_before);
1103 out:
1104 	plr->thread_done = 1;
1105 	wake_up_interruptible(&plr->lock_thread_wq);
1106 	return 0;
1107 }
1108 
1109 static int measure_l3_residency(void *_plr)
1110 {
1111 	struct pseudo_lock_region *plr = _plr;
1112 	struct residency_counts counts = {0};
1113 
1114 	/*
1115 	 * On Broadwell Microarchitecture the MEM_LOAD_UOPS_RETIRED event
1116 	 * has two "no fix" errata associated with it: BDM35 and BDM100. On
1117 	 * this platform the following events are used instead:
1118 	 * LONGEST_LAT_CACHE 2EH (Documented in SDM)
1119 	 *       REFERENCE 4FH
1120 	 *       MISS      41H
1121 	 */
1122 
1123 	switch (boot_cpu_data.x86_model) {
1124 	case INTEL_FAM6_BROADWELL_X:
1125 		/* On BDW the hit event counts references, not hits */
1126 		perf_hit_attr.config = X86_CONFIG(.event = 0x2e,
1127 						  .umask = 0x4f);
1128 		perf_miss_attr.config = X86_CONFIG(.event = 0x2e,
1129 						   .umask = 0x41);
1130 		break;
1131 	default:
1132 		goto out;
1133 	}
1134 
1135 	measure_residency_fn(&perf_miss_attr, &perf_hit_attr, plr, &counts);
1136 	/*
1137 	 * If a failure prevented the measurements from succeeding
1138 	 * tracepoints will still be written and all counts will be zero.
1139 	 */
1140 
1141 	counts.miss_after -= counts.miss_before;
1142 	if (boot_cpu_data.x86_model == INTEL_FAM6_BROADWELL_X) {
1143 		/*
1144 		 * On BDW references and misses are counted, need to adjust.
1145 		 * Sometimes the "hits" counter is a bit more than the
1146 		 * references, for example, x references but x + 1 hits.
1147 		 * To not report invalid hit values in this case we treat
1148 		 * that as misses equal to references.
1149 		 */
1150 		/* First compute the number of cache references measured */
1151 		counts.hits_after -= counts.hits_before;
1152 		/* Next convert references to cache hits */
1153 		counts.hits_after -= min(counts.miss_after, counts.hits_after);
1154 	} else {
1155 		counts.hits_after -= counts.hits_before;
1156 	}
1157 
1158 	trace_pseudo_lock_l3(counts.hits_after, counts.miss_after);
1159 out:
1160 	plr->thread_done = 1;
1161 	wake_up_interruptible(&plr->lock_thread_wq);
1162 	return 0;
1163 }
1164 
1165 /**
1166  * pseudo_lock_measure_cycles - Trigger latency measure to pseudo-locked region
1167  * @rdtgrp: Resource group to which the pseudo-locked region belongs.
1168  * @sel: Selector of which measurement to perform on a pseudo-locked region.
1169  *
1170  * The measurement of latency to access a pseudo-locked region should be
1171  * done from a cpu that is associated with that pseudo-locked region.
1172  * Determine which cpu is associated with this region and start a thread on
1173  * that cpu to perform the measurement, wait for that thread to complete.
1174  *
1175  * Return: 0 on success, <0 on failure
1176  */
1177 static int pseudo_lock_measure_cycles(struct rdtgroup *rdtgrp, int sel)
1178 {
1179 	struct pseudo_lock_region *plr = rdtgrp->plr;
1180 	struct task_struct *thread;
1181 	unsigned int cpu;
1182 	int ret = -1;
1183 
1184 	cpus_read_lock();
1185 	mutex_lock(&rdtgroup_mutex);
1186 
1187 	if (rdtgrp->flags & RDT_DELETED) {
1188 		ret = -ENODEV;
1189 		goto out;
1190 	}
1191 
1192 	if (!plr->d) {
1193 		ret = -ENODEV;
1194 		goto out;
1195 	}
1196 
1197 	plr->thread_done = 0;
1198 	cpu = cpumask_first(&plr->d->cpu_mask);
1199 	if (!cpu_online(cpu)) {
1200 		ret = -ENODEV;
1201 		goto out;
1202 	}
1203 
1204 	plr->cpu = cpu;
1205 
1206 	if (sel == 1)
1207 		thread = kthread_create_on_node(measure_cycles_lat_fn, plr,
1208 						cpu_to_node(cpu),
1209 						"pseudo_lock_measure/%u",
1210 						cpu);
1211 	else if (sel == 2)
1212 		thread = kthread_create_on_node(measure_l2_residency, plr,
1213 						cpu_to_node(cpu),
1214 						"pseudo_lock_measure/%u",
1215 						cpu);
1216 	else if (sel == 3)
1217 		thread = kthread_create_on_node(measure_l3_residency, plr,
1218 						cpu_to_node(cpu),
1219 						"pseudo_lock_measure/%u",
1220 						cpu);
1221 	else
1222 		goto out;
1223 
1224 	if (IS_ERR(thread)) {
1225 		ret = PTR_ERR(thread);
1226 		goto out;
1227 	}
1228 	kthread_bind(thread, cpu);
1229 	wake_up_process(thread);
1230 
1231 	ret = wait_event_interruptible(plr->lock_thread_wq,
1232 				       plr->thread_done == 1);
1233 	if (ret < 0)
1234 		goto out;
1235 
1236 	ret = 0;
1237 
1238 out:
1239 	mutex_unlock(&rdtgroup_mutex);
1240 	cpus_read_unlock();
1241 	return ret;
1242 }
1243 
1244 static ssize_t pseudo_lock_measure_trigger(struct file *file,
1245 					   const char __user *user_buf,
1246 					   size_t count, loff_t *ppos)
1247 {
1248 	struct rdtgroup *rdtgrp = file->private_data;
1249 	size_t buf_size;
1250 	char buf[32];
1251 	int ret;
1252 	int sel;
1253 
1254 	buf_size = min(count, (sizeof(buf) - 1));
1255 	if (copy_from_user(buf, user_buf, buf_size))
1256 		return -EFAULT;
1257 
1258 	buf[buf_size] = '\0';
1259 	ret = kstrtoint(buf, 10, &sel);
1260 	if (ret == 0) {
1261 		if (sel != 1 && sel != 2 && sel != 3)
1262 			return -EINVAL;
1263 		ret = debugfs_file_get(file->f_path.dentry);
1264 		if (ret)
1265 			return ret;
1266 		ret = pseudo_lock_measure_cycles(rdtgrp, sel);
1267 		if (ret == 0)
1268 			ret = count;
1269 		debugfs_file_put(file->f_path.dentry);
1270 	}
1271 
1272 	return ret;
1273 }
1274 
1275 static const struct file_operations pseudo_measure_fops = {
1276 	.write = pseudo_lock_measure_trigger,
1277 	.open = simple_open,
1278 	.llseek = default_llseek,
1279 };
1280 
1281 /**
1282  * rdtgroup_pseudo_lock_create - Create a pseudo-locked region
1283  * @rdtgrp: resource group to which pseudo-lock region belongs
1284  *
1285  * Called when a resource group in the pseudo-locksetup mode receives a
1286  * valid schemata that should be pseudo-locked. Since the resource group is
1287  * in pseudo-locksetup mode the &struct pseudo_lock_region has already been
1288  * allocated and initialized with the essential information. If a failure
1289  * occurs the resource group remains in the pseudo-locksetup mode with the
1290  * &struct pseudo_lock_region associated with it, but cleared from all
1291  * information and ready for the user to re-attempt pseudo-locking by
1292  * writing the schemata again.
1293  *
1294  * Return: 0 if the pseudo-locked region was successfully pseudo-locked, <0
1295  * on failure. Descriptive error will be written to last_cmd_status buffer.
1296  */
1297 int rdtgroup_pseudo_lock_create(struct rdtgroup *rdtgrp)
1298 {
1299 	struct pseudo_lock_region *plr = rdtgrp->plr;
1300 	struct task_struct *thread;
1301 	unsigned int new_minor;
1302 	struct device *dev;
1303 	int ret;
1304 
1305 	ret = pseudo_lock_region_alloc(plr);
1306 	if (ret < 0)
1307 		return ret;
1308 
1309 	ret = pseudo_lock_cstates_constrain(plr);
1310 	if (ret < 0) {
1311 		ret = -EINVAL;
1312 		goto out_region;
1313 	}
1314 
1315 	plr->thread_done = 0;
1316 
1317 	thread = kthread_create_on_node(pseudo_lock_fn, rdtgrp,
1318 					cpu_to_node(plr->cpu),
1319 					"pseudo_lock/%u", plr->cpu);
1320 	if (IS_ERR(thread)) {
1321 		ret = PTR_ERR(thread);
1322 		rdt_last_cmd_printf("Locking thread returned error %d\n", ret);
1323 		goto out_cstates;
1324 	}
1325 
1326 	kthread_bind(thread, plr->cpu);
1327 	wake_up_process(thread);
1328 
1329 	ret = wait_event_interruptible(plr->lock_thread_wq,
1330 				       plr->thread_done == 1);
1331 	if (ret < 0) {
1332 		/*
1333 		 * If the thread does not get on the CPU for whatever
1334 		 * reason and the process which sets up the region is
1335 		 * interrupted then this will leave the thread in runnable
1336 		 * state and once it gets on the CPU it will dereference
1337 		 * the cleared, but not freed, plr struct resulting in an
1338 		 * empty pseudo-locking loop.
1339 		 */
1340 		rdt_last_cmd_puts("Locking thread interrupted\n");
1341 		goto out_cstates;
1342 	}
1343 
1344 	ret = pseudo_lock_minor_get(&new_minor);
1345 	if (ret < 0) {
1346 		rdt_last_cmd_puts("Unable to obtain a new minor number\n");
1347 		goto out_cstates;
1348 	}
1349 
1350 	/*
1351 	 * Unlock access but do not release the reference. The
1352 	 * pseudo-locked region will still be here on return.
1353 	 *
1354 	 * The mutex has to be released temporarily to avoid a potential
1355 	 * deadlock with the mm->mmap_lock which is obtained in the
1356 	 * device_create() and debugfs_create_dir() callpath below as well as
1357 	 * before the mmap() callback is called.
1358 	 */
1359 	mutex_unlock(&rdtgroup_mutex);
1360 
1361 	if (!IS_ERR_OR_NULL(debugfs_resctrl)) {
1362 		plr->debugfs_dir = debugfs_create_dir(rdtgrp->kn->name,
1363 						      debugfs_resctrl);
1364 		if (!IS_ERR_OR_NULL(plr->debugfs_dir))
1365 			debugfs_create_file("pseudo_lock_measure", 0200,
1366 					    plr->debugfs_dir, rdtgrp,
1367 					    &pseudo_measure_fops);
1368 	}
1369 
1370 	dev = device_create(&pseudo_lock_class, NULL,
1371 			    MKDEV(pseudo_lock_major, new_minor),
1372 			    rdtgrp, "%s", rdtgrp->kn->name);
1373 
1374 	mutex_lock(&rdtgroup_mutex);
1375 
1376 	if (IS_ERR(dev)) {
1377 		ret = PTR_ERR(dev);
1378 		rdt_last_cmd_printf("Failed to create character device: %d\n",
1379 				    ret);
1380 		goto out_debugfs;
1381 	}
1382 
1383 	/* We released the mutex - check if group was removed while we did so */
1384 	if (rdtgrp->flags & RDT_DELETED) {
1385 		ret = -ENODEV;
1386 		goto out_device;
1387 	}
1388 
1389 	plr->minor = new_minor;
1390 
1391 	rdtgrp->mode = RDT_MODE_PSEUDO_LOCKED;
1392 	closid_free(rdtgrp->closid);
1393 	rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0444);
1394 	rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0444);
1395 
1396 	ret = 0;
1397 	goto out;
1398 
1399 out_device:
1400 	device_destroy(&pseudo_lock_class, MKDEV(pseudo_lock_major, new_minor));
1401 out_debugfs:
1402 	debugfs_remove_recursive(plr->debugfs_dir);
1403 	pseudo_lock_minor_release(new_minor);
1404 out_cstates:
1405 	pseudo_lock_cstates_relax(plr);
1406 out_region:
1407 	pseudo_lock_region_clear(plr);
1408 out:
1409 	return ret;
1410 }
1411 
1412 /**
1413  * rdtgroup_pseudo_lock_remove - Remove a pseudo-locked region
1414  * @rdtgrp: resource group to which the pseudo-locked region belongs
1415  *
1416  * The removal of a pseudo-locked region can be initiated when the resource
1417  * group is removed from user space via a "rmdir" from userspace or the
1418  * unmount of the resctrl filesystem. On removal the resource group does
1419  * not go back to pseudo-locksetup mode before it is removed, instead it is
1420  * removed directly. There is thus asymmetry with the creation where the
1421  * &struct pseudo_lock_region is removed here while it was not created in
1422  * rdtgroup_pseudo_lock_create().
1423  *
1424  * Return: void
1425  */
1426 void rdtgroup_pseudo_lock_remove(struct rdtgroup *rdtgrp)
1427 {
1428 	struct pseudo_lock_region *plr = rdtgrp->plr;
1429 
1430 	if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1431 		/*
1432 		 * Default group cannot be a pseudo-locked region so we can
1433 		 * free closid here.
1434 		 */
1435 		closid_free(rdtgrp->closid);
1436 		goto free;
1437 	}
1438 
1439 	pseudo_lock_cstates_relax(plr);
1440 	debugfs_remove_recursive(rdtgrp->plr->debugfs_dir);
1441 	device_destroy(&pseudo_lock_class, MKDEV(pseudo_lock_major, plr->minor));
1442 	pseudo_lock_minor_release(plr->minor);
1443 
1444 free:
1445 	pseudo_lock_free(rdtgrp);
1446 }
1447 
1448 static int pseudo_lock_dev_open(struct inode *inode, struct file *filp)
1449 {
1450 	struct rdtgroup *rdtgrp;
1451 
1452 	mutex_lock(&rdtgroup_mutex);
1453 
1454 	rdtgrp = region_find_by_minor(iminor(inode));
1455 	if (!rdtgrp) {
1456 		mutex_unlock(&rdtgroup_mutex);
1457 		return -ENODEV;
1458 	}
1459 
1460 	filp->private_data = rdtgrp;
1461 	atomic_inc(&rdtgrp->waitcount);
1462 	/* Perform a non-seekable open - llseek is not supported */
1463 	filp->f_mode &= ~(FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE);
1464 
1465 	mutex_unlock(&rdtgroup_mutex);
1466 
1467 	return 0;
1468 }
1469 
1470 static int pseudo_lock_dev_release(struct inode *inode, struct file *filp)
1471 {
1472 	struct rdtgroup *rdtgrp;
1473 
1474 	mutex_lock(&rdtgroup_mutex);
1475 	rdtgrp = filp->private_data;
1476 	WARN_ON(!rdtgrp);
1477 	if (!rdtgrp) {
1478 		mutex_unlock(&rdtgroup_mutex);
1479 		return -ENODEV;
1480 	}
1481 	filp->private_data = NULL;
1482 	atomic_dec(&rdtgrp->waitcount);
1483 	mutex_unlock(&rdtgroup_mutex);
1484 	return 0;
1485 }
1486 
1487 static int pseudo_lock_dev_mremap(struct vm_area_struct *area)
1488 {
1489 	/* Not supported */
1490 	return -EINVAL;
1491 }
1492 
1493 static const struct vm_operations_struct pseudo_mmap_ops = {
1494 	.mremap = pseudo_lock_dev_mremap,
1495 };
1496 
1497 static int pseudo_lock_dev_mmap(struct file *filp, struct vm_area_struct *vma)
1498 {
1499 	unsigned long vsize = vma->vm_end - vma->vm_start;
1500 	unsigned long off = vma->vm_pgoff << PAGE_SHIFT;
1501 	struct pseudo_lock_region *plr;
1502 	struct rdtgroup *rdtgrp;
1503 	unsigned long physical;
1504 	unsigned long psize;
1505 
1506 	mutex_lock(&rdtgroup_mutex);
1507 
1508 	rdtgrp = filp->private_data;
1509 	WARN_ON(!rdtgrp);
1510 	if (!rdtgrp) {
1511 		mutex_unlock(&rdtgroup_mutex);
1512 		return -ENODEV;
1513 	}
1514 
1515 	plr = rdtgrp->plr;
1516 
1517 	if (!plr->d) {
1518 		mutex_unlock(&rdtgroup_mutex);
1519 		return -ENODEV;
1520 	}
1521 
1522 	/*
1523 	 * Task is required to run with affinity to the cpus associated
1524 	 * with the pseudo-locked region. If this is not the case the task
1525 	 * may be scheduled elsewhere and invalidate entries in the
1526 	 * pseudo-locked region.
1527 	 */
1528 	if (!cpumask_subset(current->cpus_ptr, &plr->d->cpu_mask)) {
1529 		mutex_unlock(&rdtgroup_mutex);
1530 		return -EINVAL;
1531 	}
1532 
1533 	physical = __pa(plr->kmem) >> PAGE_SHIFT;
1534 	psize = plr->size - off;
1535 
1536 	if (off > plr->size) {
1537 		mutex_unlock(&rdtgroup_mutex);
1538 		return -ENOSPC;
1539 	}
1540 
1541 	/*
1542 	 * Ensure changes are carried directly to the memory being mapped,
1543 	 * do not allow copy-on-write mapping.
1544 	 */
1545 	if (!(vma->vm_flags & VM_SHARED)) {
1546 		mutex_unlock(&rdtgroup_mutex);
1547 		return -EINVAL;
1548 	}
1549 
1550 	if (vsize > psize) {
1551 		mutex_unlock(&rdtgroup_mutex);
1552 		return -ENOSPC;
1553 	}
1554 
1555 	memset(plr->kmem + off, 0, vsize);
1556 
1557 	if (remap_pfn_range(vma, vma->vm_start, physical + vma->vm_pgoff,
1558 			    vsize, vma->vm_page_prot)) {
1559 		mutex_unlock(&rdtgroup_mutex);
1560 		return -EAGAIN;
1561 	}
1562 	vma->vm_ops = &pseudo_mmap_ops;
1563 	mutex_unlock(&rdtgroup_mutex);
1564 	return 0;
1565 }
1566 
1567 static const struct file_operations pseudo_lock_dev_fops = {
1568 	.owner =	THIS_MODULE,
1569 	.llseek =	no_llseek,
1570 	.read =		NULL,
1571 	.write =	NULL,
1572 	.open =		pseudo_lock_dev_open,
1573 	.release =	pseudo_lock_dev_release,
1574 	.mmap =		pseudo_lock_dev_mmap,
1575 };
1576 
1577 int rdt_pseudo_lock_init(void)
1578 {
1579 	int ret;
1580 
1581 	ret = register_chrdev(0, "pseudo_lock", &pseudo_lock_dev_fops);
1582 	if (ret < 0)
1583 		return ret;
1584 
1585 	pseudo_lock_major = ret;
1586 
1587 	ret = class_register(&pseudo_lock_class);
1588 	if (ret) {
1589 		unregister_chrdev(pseudo_lock_major, "pseudo_lock");
1590 		return ret;
1591 	}
1592 
1593 	return 0;
1594 }
1595 
1596 void rdt_pseudo_lock_release(void)
1597 {
1598 	class_unregister(&pseudo_lock_class);
1599 	unregister_chrdev(pseudo_lock_major, "pseudo_lock");
1600 	pseudo_lock_major = 0;
1601 }
1602