Loading...
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_sched.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 */
40static u64 prefetch_disable_bits;
41
42/*
43 * Major number assigned to and shared by all devices exposing
44 * pseudo-locked regions.
45 */
46static unsigned int pseudo_lock_major;
47static unsigned long pseudo_lock_minor_avail = GENMASK(MINORBITS, 0);
48static struct class *pseudo_lock_class;
49
50/**
51 * get_prefetch_disable_bits - prefetch disable bits of supported platforms
52 *
53 * Capture the list of platforms that have been validated to support
54 * pseudo-locking. This includes testing to ensure pseudo-locked regions
55 * with low cache miss rates can be created under variety of load conditions
56 * as well as that these pseudo-locked regions can maintain their low cache
57 * miss rates under variety of load conditions for significant lengths of time.
58 *
59 * After a platform has been validated to support pseudo-locking its
60 * hardware prefetch disable bits are included here as they are documented
61 * in the SDM.
62 *
63 * When adding a platform here also add support for its cache events to
64 * measure_cycles_perf_fn()
65 *
66 * Return:
67 * If platform is supported, the bits to disable hardware prefetchers, 0
68 * if platform is not supported.
69 */
70static u64 get_prefetch_disable_bits(void)
71{
72 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL ||
73 boot_cpu_data.x86 != 6)
74 return 0;
75
76 switch (boot_cpu_data.x86_model) {
77 case INTEL_FAM6_BROADWELL_X:
78 /*
79 * SDM defines bits of MSR_MISC_FEATURE_CONTROL register
80 * as:
81 * 0 L2 Hardware Prefetcher Disable (R/W)
82 * 1 L2 Adjacent Cache Line Prefetcher Disable (R/W)
83 * 2 DCU Hardware Prefetcher Disable (R/W)
84 * 3 DCU IP Prefetcher Disable (R/W)
85 * 63:4 Reserved
86 */
87 return 0xF;
88 case INTEL_FAM6_ATOM_GOLDMONT:
89 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
90 /*
91 * SDM defines bits of MSR_MISC_FEATURE_CONTROL register
92 * as:
93 * 0 L2 Hardware Prefetcher Disable (R/W)
94 * 1 Reserved
95 * 2 DCU Hardware Prefetcher Disable (R/W)
96 * 63:3 Reserved
97 */
98 return 0x5;
99 }
100
101 return 0;
102}
103
104/**
105 * pseudo_lock_minor_get - Obtain available minor number
106 * @minor: Pointer to where new minor number will be stored
107 *
108 * A bitmask is used to track available minor numbers. Here the next free
109 * minor number is marked as unavailable and returned.
110 *
111 * Return: 0 on success, <0 on failure.
112 */
113static int pseudo_lock_minor_get(unsigned int *minor)
114{
115 unsigned long first_bit;
116
117 first_bit = find_first_bit(&pseudo_lock_minor_avail, MINORBITS);
118
119 if (first_bit == MINORBITS)
120 return -ENOSPC;
121
122 __clear_bit(first_bit, &pseudo_lock_minor_avail);
123 *minor = first_bit;
124
125 return 0;
126}
127
128/**
129 * pseudo_lock_minor_release - Return minor number to available
130 * @minor: The minor number made available
131 */
132static void pseudo_lock_minor_release(unsigned int minor)
133{
134 __set_bit(minor, &pseudo_lock_minor_avail);
135}
136
137/**
138 * region_find_by_minor - Locate a pseudo-lock region by inode minor number
139 * @minor: The minor number of the device representing pseudo-locked region
140 *
141 * When the character device is accessed we need to determine which
142 * pseudo-locked region it belongs to. This is done by matching the minor
143 * number of the device to the pseudo-locked region it belongs.
144 *
145 * Minor numbers are assigned at the time a pseudo-locked region is associated
146 * with a cache instance.
147 *
148 * Return: On success return pointer to resource group owning the pseudo-locked
149 * region, NULL on failure.
150 */
151static struct rdtgroup *region_find_by_minor(unsigned int minor)
152{
153 struct rdtgroup *rdtgrp, *rdtgrp_match = NULL;
154
155 list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
156 if (rdtgrp->plr && rdtgrp->plr->minor == minor) {
157 rdtgrp_match = rdtgrp;
158 break;
159 }
160 }
161 return rdtgrp_match;
162}
163
164/**
165 * pseudo_lock_pm_req - A power management QoS request list entry
166 * @list: Entry within the @pm_reqs list for a pseudo-locked region
167 * @req: PM QoS request
168 */
169struct pseudo_lock_pm_req {
170 struct list_head list;
171 struct dev_pm_qos_request req;
172};
173
174static void pseudo_lock_cstates_relax(struct pseudo_lock_region *plr)
175{
176 struct pseudo_lock_pm_req *pm_req, *next;
177
178 list_for_each_entry_safe(pm_req, next, &plr->pm_reqs, list) {
179 dev_pm_qos_remove_request(&pm_req->req);
180 list_del(&pm_req->list);
181 kfree(pm_req);
182 }
183}
184
185/**
186 * pseudo_lock_cstates_constrain - Restrict cores from entering C6
187 *
188 * To prevent the cache from being affected by power management entering
189 * C6 has to be avoided. This is accomplished by requesting a latency
190 * requirement lower than lowest C6 exit latency of all supported
191 * platforms as found in the cpuidle state tables in the intel_idle driver.
192 * At this time it is possible to do so with a single latency requirement
193 * for all supported platforms.
194 *
195 * Since Goldmont is supported, which is affected by X86_BUG_MONITOR,
196 * the ACPI latencies need to be considered while keeping in mind that C2
197 * may be set to map to deeper sleep states. In this case the latency
198 * requirement needs to prevent entering C2 also.
199 */
200static int pseudo_lock_cstates_constrain(struct pseudo_lock_region *plr)
201{
202 struct pseudo_lock_pm_req *pm_req;
203 int cpu;
204 int ret;
205
206 for_each_cpu(cpu, &plr->d->cpu_mask) {
207 pm_req = kzalloc(sizeof(*pm_req), GFP_KERNEL);
208 if (!pm_req) {
209 rdt_last_cmd_puts("Failure to allocate memory for PM QoS\n");
210 ret = -ENOMEM;
211 goto out_err;
212 }
213 ret = dev_pm_qos_add_request(get_cpu_device(cpu),
214 &pm_req->req,
215 DEV_PM_QOS_RESUME_LATENCY,
216 30);
217 if (ret < 0) {
218 rdt_last_cmd_printf("Failed to add latency req CPU%d\n",
219 cpu);
220 kfree(pm_req);
221 ret = -1;
222 goto out_err;
223 }
224 list_add(&pm_req->list, &plr->pm_reqs);
225 }
226
227 return 0;
228
229out_err:
230 pseudo_lock_cstates_relax(plr);
231 return ret;
232}
233
234/**
235 * pseudo_lock_region_clear - Reset pseudo-lock region data
236 * @plr: pseudo-lock region
237 *
238 * All content of the pseudo-locked region is reset - any memory allocated
239 * freed.
240 *
241 * Return: void
242 */
243static void pseudo_lock_region_clear(struct pseudo_lock_region *plr)
244{
245 plr->size = 0;
246 plr->line_size = 0;
247 kfree(plr->kmem);
248 plr->kmem = NULL;
249 plr->r = NULL;
250 if (plr->d)
251 plr->d->plr = NULL;
252 plr->d = NULL;
253 plr->cbm = 0;
254 plr->debugfs_dir = NULL;
255}
256
257/**
258 * pseudo_lock_region_init - Initialize pseudo-lock region information
259 * @plr: pseudo-lock region
260 *
261 * Called after user provided a schemata to be pseudo-locked. From the
262 * schemata the &struct pseudo_lock_region is on entry already initialized
263 * with the resource, domain, and capacity bitmask. Here the information
264 * required for pseudo-locking is deduced from this data and &struct
265 * pseudo_lock_region initialized further. This information includes:
266 * - size in bytes of the region to be pseudo-locked
267 * - cache line size to know the stride with which data needs to be accessed
268 * to be pseudo-locked
269 * - a cpu associated with the cache instance on which the pseudo-locking
270 * flow can be executed
271 *
272 * Return: 0 on success, <0 on failure. Descriptive error will be written
273 * to last_cmd_status buffer.
274 */
275static int pseudo_lock_region_init(struct pseudo_lock_region *plr)
276{
277 struct cpu_cacheinfo *ci;
278 int ret;
279 int i;
280
281 /* Pick the first cpu we find that is associated with the cache. */
282 plr->cpu = cpumask_first(&plr->d->cpu_mask);
283
284 if (!cpu_online(plr->cpu)) {
285 rdt_last_cmd_printf("CPU %u associated with cache not online\n",
286 plr->cpu);
287 ret = -ENODEV;
288 goto out_region;
289 }
290
291 ci = get_cpu_cacheinfo(plr->cpu);
292
293 plr->size = rdtgroup_cbm_to_size(plr->r, plr->d, plr->cbm);
294
295 for (i = 0; i < ci->num_leaves; i++) {
296 if (ci->info_list[i].level == plr->r->cache_level) {
297 plr->line_size = ci->info_list[i].coherency_line_size;
298 return 0;
299 }
300 }
301
302 ret = -1;
303 rdt_last_cmd_puts("Unable to determine cache line size\n");
304out_region:
305 pseudo_lock_region_clear(plr);
306 return ret;
307}
308
309/**
310 * pseudo_lock_init - Initialize a pseudo-lock region
311 * @rdtgrp: resource group to which new pseudo-locked region will belong
312 *
313 * A pseudo-locked region is associated with a resource group. When this
314 * association is created the pseudo-locked region is initialized. The
315 * details of the pseudo-locked region are not known at this time so only
316 * allocation is done and association established.
317 *
318 * Return: 0 on success, <0 on failure
319 */
320static int pseudo_lock_init(struct rdtgroup *rdtgrp)
321{
322 struct pseudo_lock_region *plr;
323
324 plr = kzalloc(sizeof(*plr), GFP_KERNEL);
325 if (!plr)
326 return -ENOMEM;
327
328 init_waitqueue_head(&plr->lock_thread_wq);
329 INIT_LIST_HEAD(&plr->pm_reqs);
330 rdtgrp->plr = plr;
331 return 0;
332}
333
334/**
335 * pseudo_lock_region_alloc - Allocate kernel memory that will be pseudo-locked
336 * @plr: pseudo-lock region
337 *
338 * Initialize the details required to set up the pseudo-locked region and
339 * allocate the contiguous memory that will be pseudo-locked to the cache.
340 *
341 * Return: 0 on success, <0 on failure. Descriptive error will be written
342 * to last_cmd_status buffer.
343 */
344static int pseudo_lock_region_alloc(struct pseudo_lock_region *plr)
345{
346 int ret;
347
348 ret = pseudo_lock_region_init(plr);
349 if (ret < 0)
350 return ret;
351
352 /*
353 * We do not yet support contiguous regions larger than
354 * KMALLOC_MAX_SIZE.
355 */
356 if (plr->size > KMALLOC_MAX_SIZE) {
357 rdt_last_cmd_puts("Requested region exceeds maximum size\n");
358 ret = -E2BIG;
359 goto out_region;
360 }
361
362 plr->kmem = kzalloc(plr->size, GFP_KERNEL);
363 if (!plr->kmem) {
364 rdt_last_cmd_puts("Unable to allocate memory\n");
365 ret = -ENOMEM;
366 goto out_region;
367 }
368
369 ret = 0;
370 goto out;
371out_region:
372 pseudo_lock_region_clear(plr);
373out:
374 return ret;
375}
376
377/**
378 * pseudo_lock_free - Free a pseudo-locked region
379 * @rdtgrp: resource group to which pseudo-locked region belonged
380 *
381 * The pseudo-locked region's resources have already been released, or not
382 * yet created at this point. Now it can be freed and disassociated from the
383 * resource group.
384 *
385 * Return: void
386 */
387static void pseudo_lock_free(struct rdtgroup *rdtgrp)
388{
389 pseudo_lock_region_clear(rdtgrp->plr);
390 kfree(rdtgrp->plr);
391 rdtgrp->plr = NULL;
392}
393
394/**
395 * pseudo_lock_fn - Load kernel memory into cache
396 * @_rdtgrp: resource group to which pseudo-lock region belongs
397 *
398 * This is the core pseudo-locking flow.
399 *
400 * First we ensure that the kernel memory cannot be found in the cache.
401 * Then, while taking care that there will be as little interference as
402 * possible, the memory to be loaded is accessed while core is running
403 * with class of service set to the bitmask of the pseudo-locked region.
404 * After this is complete no future CAT allocations will be allowed to
405 * overlap with this bitmask.
406 *
407 * Local register variables are utilized to ensure that the memory region
408 * to be locked is the only memory access made during the critical locking
409 * loop.
410 *
411 * Return: 0. Waiter on waitqueue will be woken on completion.
412 */
413static int pseudo_lock_fn(void *_rdtgrp)
414{
415 struct rdtgroup *rdtgrp = _rdtgrp;
416 struct pseudo_lock_region *plr = rdtgrp->plr;
417 u32 rmid_p, closid_p;
418 unsigned long i;
419#ifdef CONFIG_KASAN
420 /*
421 * The registers used for local register variables are also used
422 * when KASAN is active. When KASAN is active we use a regular
423 * variable to ensure we always use a valid pointer, but the cost
424 * is that this variable will enter the cache through evicting the
425 * memory we are trying to lock into the cache. Thus expect lower
426 * pseudo-locking success rate when KASAN is active.
427 */
428 unsigned int line_size;
429 unsigned int size;
430 void *mem_r;
431#else
432 register unsigned int line_size asm("esi");
433 register unsigned int size asm("edi");
434 register void *mem_r asm(_ASM_BX);
435#endif /* CONFIG_KASAN */
436
437 /*
438 * Make sure none of the allocated memory is cached. If it is we
439 * will get a cache hit in below loop from outside of pseudo-locked
440 * region.
441 * wbinvd (as opposed to clflush/clflushopt) is required to
442 * increase likelihood that allocated cache portion will be filled
443 * with associated memory.
444 */
445 native_wbinvd();
446
447 /*
448 * Always called with interrupts enabled. By disabling interrupts
449 * ensure that we will not be preempted during this critical section.
450 */
451 local_irq_disable();
452
453 /*
454 * Call wrmsr and rdmsr as directly as possible to avoid tracing
455 * clobbering local register variables or affecting cache accesses.
456 *
457 * Disable the hardware prefetcher so that when the end of the memory
458 * being pseudo-locked is reached the hardware will not read beyond
459 * the buffer and evict pseudo-locked memory read earlier from the
460 * cache.
461 */
462 __wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
463 closid_p = this_cpu_read(pqr_state.cur_closid);
464 rmid_p = this_cpu_read(pqr_state.cur_rmid);
465 mem_r = plr->kmem;
466 size = plr->size;
467 line_size = plr->line_size;
468 /*
469 * Critical section begin: start by writing the closid associated
470 * with the capacity bitmask of the cache region being
471 * pseudo-locked followed by reading of kernel memory to load it
472 * into the cache.
473 */
474 __wrmsr(IA32_PQR_ASSOC, rmid_p, rdtgrp->closid);
475 /*
476 * Cache was flushed earlier. Now access kernel memory to read it
477 * into cache region associated with just activated plr->closid.
478 * Loop over data twice:
479 * - In first loop the cache region is shared with the page walker
480 * as it populates the paging structure caches (including TLB).
481 * - In the second loop the paging structure caches are used and
482 * cache region is populated with the memory being referenced.
483 */
484 for (i = 0; i < size; i += PAGE_SIZE) {
485 /*
486 * Add a barrier to prevent speculative execution of this
487 * loop reading beyond the end of the buffer.
488 */
489 rmb();
490 asm volatile("mov (%0,%1,1), %%eax\n\t"
491 :
492 : "r" (mem_r), "r" (i)
493 : "%eax", "memory");
494 }
495 for (i = 0; i < size; i += line_size) {
496 /*
497 * Add a barrier to prevent speculative execution of this
498 * loop reading beyond the end of the buffer.
499 */
500 rmb();
501 asm volatile("mov (%0,%1,1), %%eax\n\t"
502 :
503 : "r" (mem_r), "r" (i)
504 : "%eax", "memory");
505 }
506 /*
507 * Critical section end: restore closid with capacity bitmask that
508 * does not overlap with pseudo-locked region.
509 */
510 __wrmsr(IA32_PQR_ASSOC, rmid_p, closid_p);
511
512 /* Re-enable the hardware prefetcher(s) */
513 wrmsr(MSR_MISC_FEATURE_CONTROL, 0x0, 0x0);
514 local_irq_enable();
515
516 plr->thread_done = 1;
517 wake_up_interruptible(&plr->lock_thread_wq);
518 return 0;
519}
520
521/**
522 * rdtgroup_monitor_in_progress - Test if monitoring in progress
523 * @r: resource group being queried
524 *
525 * Return: 1 if monitor groups have been created for this resource
526 * group, 0 otherwise.
527 */
528static int rdtgroup_monitor_in_progress(struct rdtgroup *rdtgrp)
529{
530 return !list_empty(&rdtgrp->mon.crdtgrp_list);
531}
532
533/**
534 * rdtgroup_locksetup_user_restrict - Restrict user access to group
535 * @rdtgrp: resource group needing access restricted
536 *
537 * A resource group used for cache pseudo-locking cannot have cpus or tasks
538 * assigned to it. This is communicated to the user by restricting access
539 * to all the files that can be used to make such changes.
540 *
541 * Permissions restored with rdtgroup_locksetup_user_restore()
542 *
543 * Return: 0 on success, <0 on failure. If a failure occurs during the
544 * restriction of access an attempt will be made to restore permissions but
545 * the state of the mode of these files will be uncertain when a failure
546 * occurs.
547 */
548static int rdtgroup_locksetup_user_restrict(struct rdtgroup *rdtgrp)
549{
550 int ret;
551
552 ret = rdtgroup_kn_mode_restrict(rdtgrp, "tasks");
553 if (ret)
554 return ret;
555
556 ret = rdtgroup_kn_mode_restrict(rdtgrp, "cpus");
557 if (ret)
558 goto err_tasks;
559
560 ret = rdtgroup_kn_mode_restrict(rdtgrp, "cpus_list");
561 if (ret)
562 goto err_cpus;
563
564 if (rdt_mon_capable) {
565 ret = rdtgroup_kn_mode_restrict(rdtgrp, "mon_groups");
566 if (ret)
567 goto err_cpus_list;
568 }
569
570 ret = 0;
571 goto out;
572
573err_cpus_list:
574 rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0777);
575err_cpus:
576 rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0777);
577err_tasks:
578 rdtgroup_kn_mode_restore(rdtgrp, "tasks", 0777);
579out:
580 return ret;
581}
582
583/**
584 * rdtgroup_locksetup_user_restore - Restore user access to group
585 * @rdtgrp: resource group needing access restored
586 *
587 * Restore all file access previously removed using
588 * rdtgroup_locksetup_user_restrict()
589 *
590 * Return: 0 on success, <0 on failure. If a failure occurs during the
591 * restoration of access an attempt will be made to restrict permissions
592 * again but the state of the mode of these files will be uncertain when
593 * a failure occurs.
594 */
595static int rdtgroup_locksetup_user_restore(struct rdtgroup *rdtgrp)
596{
597 int ret;
598
599 ret = rdtgroup_kn_mode_restore(rdtgrp, "tasks", 0777);
600 if (ret)
601 return ret;
602
603 ret = rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0777);
604 if (ret)
605 goto err_tasks;
606
607 ret = rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0777);
608 if (ret)
609 goto err_cpus;
610
611 if (rdt_mon_capable) {
612 ret = rdtgroup_kn_mode_restore(rdtgrp, "mon_groups", 0777);
613 if (ret)
614 goto err_cpus_list;
615 }
616
617 ret = 0;
618 goto out;
619
620err_cpus_list:
621 rdtgroup_kn_mode_restrict(rdtgrp, "cpus_list");
622err_cpus:
623 rdtgroup_kn_mode_restrict(rdtgrp, "cpus");
624err_tasks:
625 rdtgroup_kn_mode_restrict(rdtgrp, "tasks");
626out:
627 return ret;
628}
629
630/**
631 * rdtgroup_locksetup_enter - Resource group enters locksetup mode
632 * @rdtgrp: resource group requested to enter locksetup mode
633 *
634 * A resource group enters locksetup mode to reflect that it would be used
635 * to represent a pseudo-locked region and is in the process of being set
636 * up to do so. A resource group used for a pseudo-locked region would
637 * lose the closid associated with it so we cannot allow it to have any
638 * tasks or cpus assigned nor permit tasks or cpus to be assigned in the
639 * future. Monitoring of a pseudo-locked region is not allowed either.
640 *
641 * The above and more restrictions on a pseudo-locked region are checked
642 * for and enforced before the resource group enters the locksetup mode.
643 *
644 * Returns: 0 if the resource group successfully entered locksetup mode, <0
645 * on failure. On failure the last_cmd_status buffer is updated with text to
646 * communicate details of failure to the user.
647 */
648int rdtgroup_locksetup_enter(struct rdtgroup *rdtgrp)
649{
650 int ret;
651
652 /*
653 * The default resource group can neither be removed nor lose the
654 * default closid associated with it.
655 */
656 if (rdtgrp == &rdtgroup_default) {
657 rdt_last_cmd_puts("Cannot pseudo-lock default group\n");
658 return -EINVAL;
659 }
660
661 /*
662 * Cache Pseudo-locking not supported when CDP is enabled.
663 *
664 * Some things to consider if you would like to enable this
665 * support (using L3 CDP as example):
666 * - When CDP is enabled two separate resources are exposed,
667 * L3DATA and L3CODE, but they are actually on the same cache.
668 * The implication for pseudo-locking is that if a
669 * pseudo-locked region is created on a domain of one
670 * resource (eg. L3CODE), then a pseudo-locked region cannot
671 * be created on that same domain of the other resource
672 * (eg. L3DATA). This is because the creation of a
673 * pseudo-locked region involves a call to wbinvd that will
674 * affect all cache allocations on particular domain.
675 * - Considering the previous, it may be possible to only
676 * expose one of the CDP resources to pseudo-locking and
677 * hide the other. For example, we could consider to only
678 * expose L3DATA and since the L3 cache is unified it is
679 * still possible to place instructions there are execute it.
680 * - If only one region is exposed to pseudo-locking we should
681 * still keep in mind that availability of a portion of cache
682 * for pseudo-locking should take into account both resources.
683 * Similarly, if a pseudo-locked region is created in one
684 * resource, the portion of cache used by it should be made
685 * unavailable to all future allocations from both resources.
686 */
687 if (rdt_resources_all[RDT_RESOURCE_L3DATA].alloc_enabled ||
688 rdt_resources_all[RDT_RESOURCE_L2DATA].alloc_enabled) {
689 rdt_last_cmd_puts("CDP enabled\n");
690 return -EINVAL;
691 }
692
693 /*
694 * Not knowing the bits to disable prefetching implies that this
695 * platform does not support Cache Pseudo-Locking.
696 */
697 prefetch_disable_bits = get_prefetch_disable_bits();
698 if (prefetch_disable_bits == 0) {
699 rdt_last_cmd_puts("Pseudo-locking not supported\n");
700 return -EINVAL;
701 }
702
703 if (rdtgroup_monitor_in_progress(rdtgrp)) {
704 rdt_last_cmd_puts("Monitoring in progress\n");
705 return -EINVAL;
706 }
707
708 if (rdtgroup_tasks_assigned(rdtgrp)) {
709 rdt_last_cmd_puts("Tasks assigned to resource group\n");
710 return -EINVAL;
711 }
712
713 if (!cpumask_empty(&rdtgrp->cpu_mask)) {
714 rdt_last_cmd_puts("CPUs assigned to resource group\n");
715 return -EINVAL;
716 }
717
718 if (rdtgroup_locksetup_user_restrict(rdtgrp)) {
719 rdt_last_cmd_puts("Unable to modify resctrl permissions\n");
720 return -EIO;
721 }
722
723 ret = pseudo_lock_init(rdtgrp);
724 if (ret) {
725 rdt_last_cmd_puts("Unable to init pseudo-lock region\n");
726 goto out_release;
727 }
728
729 /*
730 * If this system is capable of monitoring a rmid would have been
731 * allocated when the control group was created. This is not needed
732 * anymore when this group would be used for pseudo-locking. This
733 * is safe to call on platforms not capable of monitoring.
734 */
735 free_rmid(rdtgrp->mon.rmid);
736
737 ret = 0;
738 goto out;
739
740out_release:
741 rdtgroup_locksetup_user_restore(rdtgrp);
742out:
743 return ret;
744}
745
746/**
747 * rdtgroup_locksetup_exit - resource group exist locksetup mode
748 * @rdtgrp: resource group
749 *
750 * When a resource group exits locksetup mode the earlier restrictions are
751 * lifted.
752 *
753 * Return: 0 on success, <0 on failure
754 */
755int rdtgroup_locksetup_exit(struct rdtgroup *rdtgrp)
756{
757 int ret;
758
759 if (rdt_mon_capable) {
760 ret = alloc_rmid();
761 if (ret < 0) {
762 rdt_last_cmd_puts("Out of RMIDs\n");
763 return ret;
764 }
765 rdtgrp->mon.rmid = ret;
766 }
767
768 ret = rdtgroup_locksetup_user_restore(rdtgrp);
769 if (ret) {
770 free_rmid(rdtgrp->mon.rmid);
771 return ret;
772 }
773
774 pseudo_lock_free(rdtgrp);
775 return 0;
776}
777
778/**
779 * rdtgroup_cbm_overlaps_pseudo_locked - Test if CBM or portion is pseudo-locked
780 * @d: RDT domain
781 * @cbm: CBM to test
782 *
783 * @d represents a cache instance and @cbm a capacity bitmask that is
784 * considered for it. Determine if @cbm overlaps with any existing
785 * pseudo-locked region on @d.
786 *
787 * @cbm is unsigned long, even if only 32 bits are used, to make the
788 * bitmap functions work correctly.
789 *
790 * Return: true if @cbm overlaps with pseudo-locked region on @d, false
791 * otherwise.
792 */
793bool rdtgroup_cbm_overlaps_pseudo_locked(struct rdt_domain *d, unsigned long cbm)
794{
795 unsigned int cbm_len;
796 unsigned long cbm_b;
797
798 if (d->plr) {
799 cbm_len = d->plr->r->cache.cbm_len;
800 cbm_b = d->plr->cbm;
801 if (bitmap_intersects(&cbm, &cbm_b, cbm_len))
802 return true;
803 }
804 return false;
805}
806
807/**
808 * rdtgroup_pseudo_locked_in_hierarchy - Pseudo-locked region in cache hierarchy
809 * @d: RDT domain under test
810 *
811 * The setup of a pseudo-locked region affects all cache instances within
812 * the hierarchy of the region. It is thus essential to know if any
813 * pseudo-locked regions exist within a cache hierarchy to prevent any
814 * attempts to create new pseudo-locked regions in the same hierarchy.
815 *
816 * Return: true if a pseudo-locked region exists in the hierarchy of @d or
817 * if it is not possible to test due to memory allocation issue,
818 * false otherwise.
819 */
820bool rdtgroup_pseudo_locked_in_hierarchy(struct rdt_domain *d)
821{
822 cpumask_var_t cpu_with_psl;
823 struct rdt_resource *r;
824 struct rdt_domain *d_i;
825 bool ret = false;
826
827 if (!zalloc_cpumask_var(&cpu_with_psl, GFP_KERNEL))
828 return true;
829
830 /*
831 * First determine which cpus have pseudo-locked regions
832 * associated with them.
833 */
834 for_each_alloc_enabled_rdt_resource(r) {
835 list_for_each_entry(d_i, &r->domains, list) {
836 if (d_i->plr)
837 cpumask_or(cpu_with_psl, cpu_with_psl,
838 &d_i->cpu_mask);
839 }
840 }
841
842 /*
843 * Next test if new pseudo-locked region would intersect with
844 * existing region.
845 */
846 if (cpumask_intersects(&d->cpu_mask, cpu_with_psl))
847 ret = true;
848
849 free_cpumask_var(cpu_with_psl);
850 return ret;
851}
852
853/**
854 * measure_cycles_lat_fn - Measure cycle latency to read pseudo-locked memory
855 * @_plr: pseudo-lock region to measure
856 *
857 * There is no deterministic way to test if a memory region is cached. One
858 * way is to measure how long it takes to read the memory, the speed of
859 * access is a good way to learn how close to the cpu the data was. Even
860 * more, if the prefetcher is disabled and the memory is read at a stride
861 * of half the cache line, then a cache miss will be easy to spot since the
862 * read of the first half would be significantly slower than the read of
863 * the second half.
864 *
865 * Return: 0. Waiter on waitqueue will be woken on completion.
866 */
867static int measure_cycles_lat_fn(void *_plr)
868{
869 struct pseudo_lock_region *plr = _plr;
870 unsigned long i;
871 u64 start, end;
872 void *mem_r;
873
874 local_irq_disable();
875 /*
876 * Disable hardware prefetchers.
877 */
878 wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
879 mem_r = READ_ONCE(plr->kmem);
880 /*
881 * Dummy execute of the time measurement to load the needed
882 * instructions into the L1 instruction cache.
883 */
884 start = rdtsc_ordered();
885 for (i = 0; i < plr->size; i += 32) {
886 start = rdtsc_ordered();
887 asm volatile("mov (%0,%1,1), %%eax\n\t"
888 :
889 : "r" (mem_r), "r" (i)
890 : "%eax", "memory");
891 end = rdtsc_ordered();
892 trace_pseudo_lock_mem_latency((u32)(end - start));
893 }
894 wrmsr(MSR_MISC_FEATURE_CONTROL, 0x0, 0x0);
895 local_irq_enable();
896 plr->thread_done = 1;
897 wake_up_interruptible(&plr->lock_thread_wq);
898 return 0;
899}
900
901/*
902 * Create a perf_event_attr for the hit and miss perf events that will
903 * be used during the performance measurement. A perf_event maintains
904 * a pointer to its perf_event_attr so a unique attribute structure is
905 * created for each perf_event.
906 *
907 * The actual configuration of the event is set right before use in order
908 * to use the X86_CONFIG macro.
909 */
910static struct perf_event_attr perf_miss_attr = {
911 .type = PERF_TYPE_RAW,
912 .size = sizeof(struct perf_event_attr),
913 .pinned = 1,
914 .disabled = 0,
915 .exclude_user = 1,
916};
917
918static struct perf_event_attr perf_hit_attr = {
919 .type = PERF_TYPE_RAW,
920 .size = sizeof(struct perf_event_attr),
921 .pinned = 1,
922 .disabled = 0,
923 .exclude_user = 1,
924};
925
926struct residency_counts {
927 u64 miss_before, hits_before;
928 u64 miss_after, hits_after;
929};
930
931static int measure_residency_fn(struct perf_event_attr *miss_attr,
932 struct perf_event_attr *hit_attr,
933 struct pseudo_lock_region *plr,
934 struct residency_counts *counts)
935{
936 u64 hits_before = 0, hits_after = 0, miss_before = 0, miss_after = 0;
937 struct perf_event *miss_event, *hit_event;
938 int hit_pmcnum, miss_pmcnum;
939 unsigned int line_size;
940 unsigned int size;
941 unsigned long i;
942 void *mem_r;
943 u64 tmp;
944
945 miss_event = perf_event_create_kernel_counter(miss_attr, plr->cpu,
946 NULL, NULL, NULL);
947 if (IS_ERR(miss_event))
948 goto out;
949
950 hit_event = perf_event_create_kernel_counter(hit_attr, plr->cpu,
951 NULL, NULL, NULL);
952 if (IS_ERR(hit_event))
953 goto out_miss;
954
955 local_irq_disable();
956 /*
957 * Check any possible error state of events used by performing
958 * one local read.
959 */
960 if (perf_event_read_local(miss_event, &tmp, NULL, NULL)) {
961 local_irq_enable();
962 goto out_hit;
963 }
964 if (perf_event_read_local(hit_event, &tmp, NULL, NULL)) {
965 local_irq_enable();
966 goto out_hit;
967 }
968
969 /*
970 * Disable hardware prefetchers.
971 */
972 wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
973
974 /* Initialize rest of local variables */
975 /*
976 * Performance event has been validated right before this with
977 * interrupts disabled - it is thus safe to read the counter index.
978 */
979 miss_pmcnum = x86_perf_rdpmc_index(miss_event);
980 hit_pmcnum = x86_perf_rdpmc_index(hit_event);
981 line_size = READ_ONCE(plr->line_size);
982 mem_r = READ_ONCE(plr->kmem);
983 size = READ_ONCE(plr->size);
984
985 /*
986 * Read counter variables twice - first to load the instructions
987 * used in L1 cache, second to capture accurate value that does not
988 * include cache misses incurred because of instruction loads.
989 */
990 rdpmcl(hit_pmcnum, hits_before);
991 rdpmcl(miss_pmcnum, miss_before);
992 /*
993 * From SDM: Performing back-to-back fast reads are not guaranteed
994 * to be monotonic.
995 * Use LFENCE to ensure all previous instructions are retired
996 * before proceeding.
997 */
998 rmb();
999 rdpmcl(hit_pmcnum, hits_before);
1000 rdpmcl(miss_pmcnum, miss_before);
1001 /*
1002 * Use LFENCE to ensure all previous instructions are retired
1003 * before proceeding.
1004 */
1005 rmb();
1006 for (i = 0; i < size; i += line_size) {
1007 /*
1008 * Add a barrier to prevent speculative execution of this
1009 * loop reading beyond the end of the buffer.
1010 */
1011 rmb();
1012 asm volatile("mov (%0,%1,1), %%eax\n\t"
1013 :
1014 : "r" (mem_r), "r" (i)
1015 : "%eax", "memory");
1016 }
1017 /*
1018 * Use LFENCE to ensure all previous instructions are retired
1019 * before proceeding.
1020 */
1021 rmb();
1022 rdpmcl(hit_pmcnum, hits_after);
1023 rdpmcl(miss_pmcnum, miss_after);
1024 /*
1025 * Use LFENCE to ensure all previous instructions are retired
1026 * before proceeding.
1027 */
1028 rmb();
1029 /* Re-enable hardware prefetchers */
1030 wrmsr(MSR_MISC_FEATURE_CONTROL, 0x0, 0x0);
1031 local_irq_enable();
1032out_hit:
1033 perf_event_release_kernel(hit_event);
1034out_miss:
1035 perf_event_release_kernel(miss_event);
1036out:
1037 /*
1038 * All counts will be zero on failure.
1039 */
1040 counts->miss_before = miss_before;
1041 counts->hits_before = hits_before;
1042 counts->miss_after = miss_after;
1043 counts->hits_after = hits_after;
1044 return 0;
1045}
1046
1047static int measure_l2_residency(void *_plr)
1048{
1049 struct pseudo_lock_region *plr = _plr;
1050 struct residency_counts counts = {0};
1051
1052 /*
1053 * Non-architectural event for the Goldmont Microarchitecture
1054 * from Intel x86 Architecture Software Developer Manual (SDM):
1055 * MEM_LOAD_UOPS_RETIRED D1H (event number)
1056 * Umask values:
1057 * L2_HIT 02H
1058 * L2_MISS 10H
1059 */
1060 switch (boot_cpu_data.x86_model) {
1061 case INTEL_FAM6_ATOM_GOLDMONT:
1062 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
1063 perf_miss_attr.config = X86_CONFIG(.event = 0xd1,
1064 .umask = 0x10);
1065 perf_hit_attr.config = X86_CONFIG(.event = 0xd1,
1066 .umask = 0x2);
1067 break;
1068 default:
1069 goto out;
1070 }
1071
1072 measure_residency_fn(&perf_miss_attr, &perf_hit_attr, plr, &counts);
1073 /*
1074 * If a failure prevented the measurements from succeeding
1075 * tracepoints will still be written and all counts will be zero.
1076 */
1077 trace_pseudo_lock_l2(counts.hits_after - counts.hits_before,
1078 counts.miss_after - counts.miss_before);
1079out:
1080 plr->thread_done = 1;
1081 wake_up_interruptible(&plr->lock_thread_wq);
1082 return 0;
1083}
1084
1085static int measure_l3_residency(void *_plr)
1086{
1087 struct pseudo_lock_region *plr = _plr;
1088 struct residency_counts counts = {0};
1089
1090 /*
1091 * On Broadwell Microarchitecture the MEM_LOAD_UOPS_RETIRED event
1092 * has two "no fix" errata associated with it: BDM35 and BDM100. On
1093 * this platform the following events are used instead:
1094 * LONGEST_LAT_CACHE 2EH (Documented in SDM)
1095 * REFERENCE 4FH
1096 * MISS 41H
1097 */
1098
1099 switch (boot_cpu_data.x86_model) {
1100 case INTEL_FAM6_BROADWELL_X:
1101 /* On BDW the hit event counts references, not hits */
1102 perf_hit_attr.config = X86_CONFIG(.event = 0x2e,
1103 .umask = 0x4f);
1104 perf_miss_attr.config = X86_CONFIG(.event = 0x2e,
1105 .umask = 0x41);
1106 break;
1107 default:
1108 goto out;
1109 }
1110
1111 measure_residency_fn(&perf_miss_attr, &perf_hit_attr, plr, &counts);
1112 /*
1113 * If a failure prevented the measurements from succeeding
1114 * tracepoints will still be written and all counts will be zero.
1115 */
1116
1117 counts.miss_after -= counts.miss_before;
1118 if (boot_cpu_data.x86_model == INTEL_FAM6_BROADWELL_X) {
1119 /*
1120 * On BDW references and misses are counted, need to adjust.
1121 * Sometimes the "hits" counter is a bit more than the
1122 * references, for example, x references but x + 1 hits.
1123 * To not report invalid hit values in this case we treat
1124 * that as misses equal to references.
1125 */
1126 /* First compute the number of cache references measured */
1127 counts.hits_after -= counts.hits_before;
1128 /* Next convert references to cache hits */
1129 counts.hits_after -= min(counts.miss_after, counts.hits_after);
1130 } else {
1131 counts.hits_after -= counts.hits_before;
1132 }
1133
1134 trace_pseudo_lock_l3(counts.hits_after, counts.miss_after);
1135out:
1136 plr->thread_done = 1;
1137 wake_up_interruptible(&plr->lock_thread_wq);
1138 return 0;
1139}
1140
1141/**
1142 * pseudo_lock_measure_cycles - Trigger latency measure to pseudo-locked region
1143 *
1144 * The measurement of latency to access a pseudo-locked region should be
1145 * done from a cpu that is associated with that pseudo-locked region.
1146 * Determine which cpu is associated with this region and start a thread on
1147 * that cpu to perform the measurement, wait for that thread to complete.
1148 *
1149 * Return: 0 on success, <0 on failure
1150 */
1151static int pseudo_lock_measure_cycles(struct rdtgroup *rdtgrp, int sel)
1152{
1153 struct pseudo_lock_region *plr = rdtgrp->plr;
1154 struct task_struct *thread;
1155 unsigned int cpu;
1156 int ret = -1;
1157
1158 cpus_read_lock();
1159 mutex_lock(&rdtgroup_mutex);
1160
1161 if (rdtgrp->flags & RDT_DELETED) {
1162 ret = -ENODEV;
1163 goto out;
1164 }
1165
1166 if (!plr->d) {
1167 ret = -ENODEV;
1168 goto out;
1169 }
1170
1171 plr->thread_done = 0;
1172 cpu = cpumask_first(&plr->d->cpu_mask);
1173 if (!cpu_online(cpu)) {
1174 ret = -ENODEV;
1175 goto out;
1176 }
1177
1178 plr->cpu = cpu;
1179
1180 if (sel == 1)
1181 thread = kthread_create_on_node(measure_cycles_lat_fn, plr,
1182 cpu_to_node(cpu),
1183 "pseudo_lock_measure/%u",
1184 cpu);
1185 else if (sel == 2)
1186 thread = kthread_create_on_node(measure_l2_residency, plr,
1187 cpu_to_node(cpu),
1188 "pseudo_lock_measure/%u",
1189 cpu);
1190 else if (sel == 3)
1191 thread = kthread_create_on_node(measure_l3_residency, plr,
1192 cpu_to_node(cpu),
1193 "pseudo_lock_measure/%u",
1194 cpu);
1195 else
1196 goto out;
1197
1198 if (IS_ERR(thread)) {
1199 ret = PTR_ERR(thread);
1200 goto out;
1201 }
1202 kthread_bind(thread, cpu);
1203 wake_up_process(thread);
1204
1205 ret = wait_event_interruptible(plr->lock_thread_wq,
1206 plr->thread_done == 1);
1207 if (ret < 0)
1208 goto out;
1209
1210 ret = 0;
1211
1212out:
1213 mutex_unlock(&rdtgroup_mutex);
1214 cpus_read_unlock();
1215 return ret;
1216}
1217
1218static ssize_t pseudo_lock_measure_trigger(struct file *file,
1219 const char __user *user_buf,
1220 size_t count, loff_t *ppos)
1221{
1222 struct rdtgroup *rdtgrp = file->private_data;
1223 size_t buf_size;
1224 char buf[32];
1225 int ret;
1226 int sel;
1227
1228 buf_size = min(count, (sizeof(buf) - 1));
1229 if (copy_from_user(buf, user_buf, buf_size))
1230 return -EFAULT;
1231
1232 buf[buf_size] = '\0';
1233 ret = kstrtoint(buf, 10, &sel);
1234 if (ret == 0) {
1235 if (sel != 1 && sel != 2 && sel != 3)
1236 return -EINVAL;
1237 ret = debugfs_file_get(file->f_path.dentry);
1238 if (ret)
1239 return ret;
1240 ret = pseudo_lock_measure_cycles(rdtgrp, sel);
1241 if (ret == 0)
1242 ret = count;
1243 debugfs_file_put(file->f_path.dentry);
1244 }
1245
1246 return ret;
1247}
1248
1249static const struct file_operations pseudo_measure_fops = {
1250 .write = pseudo_lock_measure_trigger,
1251 .open = simple_open,
1252 .llseek = default_llseek,
1253};
1254
1255/**
1256 * rdtgroup_pseudo_lock_create - Create a pseudo-locked region
1257 * @rdtgrp: resource group to which pseudo-lock region belongs
1258 *
1259 * Called when a resource group in the pseudo-locksetup mode receives a
1260 * valid schemata that should be pseudo-locked. Since the resource group is
1261 * in pseudo-locksetup mode the &struct pseudo_lock_region has already been
1262 * allocated and initialized with the essential information. If a failure
1263 * occurs the resource group remains in the pseudo-locksetup mode with the
1264 * &struct pseudo_lock_region associated with it, but cleared from all
1265 * information and ready for the user to re-attempt pseudo-locking by
1266 * writing the schemata again.
1267 *
1268 * Return: 0 if the pseudo-locked region was successfully pseudo-locked, <0
1269 * on failure. Descriptive error will be written to last_cmd_status buffer.
1270 */
1271int rdtgroup_pseudo_lock_create(struct rdtgroup *rdtgrp)
1272{
1273 struct pseudo_lock_region *plr = rdtgrp->plr;
1274 struct task_struct *thread;
1275 unsigned int new_minor;
1276 struct device *dev;
1277 int ret;
1278
1279 ret = pseudo_lock_region_alloc(plr);
1280 if (ret < 0)
1281 return ret;
1282
1283 ret = pseudo_lock_cstates_constrain(plr);
1284 if (ret < 0) {
1285 ret = -EINVAL;
1286 goto out_region;
1287 }
1288
1289 plr->thread_done = 0;
1290
1291 thread = kthread_create_on_node(pseudo_lock_fn, rdtgrp,
1292 cpu_to_node(plr->cpu),
1293 "pseudo_lock/%u", plr->cpu);
1294 if (IS_ERR(thread)) {
1295 ret = PTR_ERR(thread);
1296 rdt_last_cmd_printf("Locking thread returned error %d\n", ret);
1297 goto out_cstates;
1298 }
1299
1300 kthread_bind(thread, plr->cpu);
1301 wake_up_process(thread);
1302
1303 ret = wait_event_interruptible(plr->lock_thread_wq,
1304 plr->thread_done == 1);
1305 if (ret < 0) {
1306 /*
1307 * If the thread does not get on the CPU for whatever
1308 * reason and the process which sets up the region is
1309 * interrupted then this will leave the thread in runnable
1310 * state and once it gets on the CPU it will derefence
1311 * the cleared, but not freed, plr struct resulting in an
1312 * empty pseudo-locking loop.
1313 */
1314 rdt_last_cmd_puts("Locking thread interrupted\n");
1315 goto out_cstates;
1316 }
1317
1318 ret = pseudo_lock_minor_get(&new_minor);
1319 if (ret < 0) {
1320 rdt_last_cmd_puts("Unable to obtain a new minor number\n");
1321 goto out_cstates;
1322 }
1323
1324 /*
1325 * Unlock access but do not release the reference. The
1326 * pseudo-locked region will still be here on return.
1327 *
1328 * The mutex has to be released temporarily to avoid a potential
1329 * deadlock with the mm->mmap_sem semaphore which is obtained in
1330 * the device_create() and debugfs_create_dir() callpath below
1331 * as well as before the mmap() callback is called.
1332 */
1333 mutex_unlock(&rdtgroup_mutex);
1334
1335 if (!IS_ERR_OR_NULL(debugfs_resctrl)) {
1336 plr->debugfs_dir = debugfs_create_dir(rdtgrp->kn->name,
1337 debugfs_resctrl);
1338 if (!IS_ERR_OR_NULL(plr->debugfs_dir))
1339 debugfs_create_file("pseudo_lock_measure", 0200,
1340 plr->debugfs_dir, rdtgrp,
1341 &pseudo_measure_fops);
1342 }
1343
1344 dev = device_create(pseudo_lock_class, NULL,
1345 MKDEV(pseudo_lock_major, new_minor),
1346 rdtgrp, "%s", rdtgrp->kn->name);
1347
1348 mutex_lock(&rdtgroup_mutex);
1349
1350 if (IS_ERR(dev)) {
1351 ret = PTR_ERR(dev);
1352 rdt_last_cmd_printf("Failed to create character device: %d\n",
1353 ret);
1354 goto out_debugfs;
1355 }
1356
1357 /* We released the mutex - check if group was removed while we did so */
1358 if (rdtgrp->flags & RDT_DELETED) {
1359 ret = -ENODEV;
1360 goto out_device;
1361 }
1362
1363 plr->minor = new_minor;
1364
1365 rdtgrp->mode = RDT_MODE_PSEUDO_LOCKED;
1366 closid_free(rdtgrp->closid);
1367 rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0444);
1368 rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0444);
1369
1370 ret = 0;
1371 goto out;
1372
1373out_device:
1374 device_destroy(pseudo_lock_class, MKDEV(pseudo_lock_major, new_minor));
1375out_debugfs:
1376 debugfs_remove_recursive(plr->debugfs_dir);
1377 pseudo_lock_minor_release(new_minor);
1378out_cstates:
1379 pseudo_lock_cstates_relax(plr);
1380out_region:
1381 pseudo_lock_region_clear(plr);
1382out:
1383 return ret;
1384}
1385
1386/**
1387 * rdtgroup_pseudo_lock_remove - Remove a pseudo-locked region
1388 * @rdtgrp: resource group to which the pseudo-locked region belongs
1389 *
1390 * The removal of a pseudo-locked region can be initiated when the resource
1391 * group is removed from user space via a "rmdir" from userspace or the
1392 * unmount of the resctrl filesystem. On removal the resource group does
1393 * not go back to pseudo-locksetup mode before it is removed, instead it is
1394 * removed directly. There is thus assymmetry with the creation where the
1395 * &struct pseudo_lock_region is removed here while it was not created in
1396 * rdtgroup_pseudo_lock_create().
1397 *
1398 * Return: void
1399 */
1400void rdtgroup_pseudo_lock_remove(struct rdtgroup *rdtgrp)
1401{
1402 struct pseudo_lock_region *plr = rdtgrp->plr;
1403
1404 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1405 /*
1406 * Default group cannot be a pseudo-locked region so we can
1407 * free closid here.
1408 */
1409 closid_free(rdtgrp->closid);
1410 goto free;
1411 }
1412
1413 pseudo_lock_cstates_relax(plr);
1414 debugfs_remove_recursive(rdtgrp->plr->debugfs_dir);
1415 device_destroy(pseudo_lock_class, MKDEV(pseudo_lock_major, plr->minor));
1416 pseudo_lock_minor_release(plr->minor);
1417
1418free:
1419 pseudo_lock_free(rdtgrp);
1420}
1421
1422static int pseudo_lock_dev_open(struct inode *inode, struct file *filp)
1423{
1424 struct rdtgroup *rdtgrp;
1425
1426 mutex_lock(&rdtgroup_mutex);
1427
1428 rdtgrp = region_find_by_minor(iminor(inode));
1429 if (!rdtgrp) {
1430 mutex_unlock(&rdtgroup_mutex);
1431 return -ENODEV;
1432 }
1433
1434 filp->private_data = rdtgrp;
1435 atomic_inc(&rdtgrp->waitcount);
1436 /* Perform a non-seekable open - llseek is not supported */
1437 filp->f_mode &= ~(FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE);
1438
1439 mutex_unlock(&rdtgroup_mutex);
1440
1441 return 0;
1442}
1443
1444static int pseudo_lock_dev_release(struct inode *inode, struct file *filp)
1445{
1446 struct rdtgroup *rdtgrp;
1447
1448 mutex_lock(&rdtgroup_mutex);
1449 rdtgrp = filp->private_data;
1450 WARN_ON(!rdtgrp);
1451 if (!rdtgrp) {
1452 mutex_unlock(&rdtgroup_mutex);
1453 return -ENODEV;
1454 }
1455 filp->private_data = NULL;
1456 atomic_dec(&rdtgrp->waitcount);
1457 mutex_unlock(&rdtgroup_mutex);
1458 return 0;
1459}
1460
1461static int pseudo_lock_dev_mremap(struct vm_area_struct *area)
1462{
1463 /* Not supported */
1464 return -EINVAL;
1465}
1466
1467static const struct vm_operations_struct pseudo_mmap_ops = {
1468 .mremap = pseudo_lock_dev_mremap,
1469};
1470
1471static int pseudo_lock_dev_mmap(struct file *filp, struct vm_area_struct *vma)
1472{
1473 unsigned long vsize = vma->vm_end - vma->vm_start;
1474 unsigned long off = vma->vm_pgoff << PAGE_SHIFT;
1475 struct pseudo_lock_region *plr;
1476 struct rdtgroup *rdtgrp;
1477 unsigned long physical;
1478 unsigned long psize;
1479
1480 mutex_lock(&rdtgroup_mutex);
1481
1482 rdtgrp = filp->private_data;
1483 WARN_ON(!rdtgrp);
1484 if (!rdtgrp) {
1485 mutex_unlock(&rdtgroup_mutex);
1486 return -ENODEV;
1487 }
1488
1489 plr = rdtgrp->plr;
1490
1491 if (!plr->d) {
1492 mutex_unlock(&rdtgroup_mutex);
1493 return -ENODEV;
1494 }
1495
1496 /*
1497 * Task is required to run with affinity to the cpus associated
1498 * with the pseudo-locked region. If this is not the case the task
1499 * may be scheduled elsewhere and invalidate entries in the
1500 * pseudo-locked region.
1501 */
1502 if (!cpumask_subset(current->cpus_ptr, &plr->d->cpu_mask)) {
1503 mutex_unlock(&rdtgroup_mutex);
1504 return -EINVAL;
1505 }
1506
1507 physical = __pa(plr->kmem) >> PAGE_SHIFT;
1508 psize = plr->size - off;
1509
1510 if (off > plr->size) {
1511 mutex_unlock(&rdtgroup_mutex);
1512 return -ENOSPC;
1513 }
1514
1515 /*
1516 * Ensure changes are carried directly to the memory being mapped,
1517 * do not allow copy-on-write mapping.
1518 */
1519 if (!(vma->vm_flags & VM_SHARED)) {
1520 mutex_unlock(&rdtgroup_mutex);
1521 return -EINVAL;
1522 }
1523
1524 if (vsize > psize) {
1525 mutex_unlock(&rdtgroup_mutex);
1526 return -ENOSPC;
1527 }
1528
1529 memset(plr->kmem + off, 0, vsize);
1530
1531 if (remap_pfn_range(vma, vma->vm_start, physical + vma->vm_pgoff,
1532 vsize, vma->vm_page_prot)) {
1533 mutex_unlock(&rdtgroup_mutex);
1534 return -EAGAIN;
1535 }
1536 vma->vm_ops = &pseudo_mmap_ops;
1537 mutex_unlock(&rdtgroup_mutex);
1538 return 0;
1539}
1540
1541static const struct file_operations pseudo_lock_dev_fops = {
1542 .owner = THIS_MODULE,
1543 .llseek = no_llseek,
1544 .read = NULL,
1545 .write = NULL,
1546 .open = pseudo_lock_dev_open,
1547 .release = pseudo_lock_dev_release,
1548 .mmap = pseudo_lock_dev_mmap,
1549};
1550
1551static char *pseudo_lock_devnode(struct device *dev, umode_t *mode)
1552{
1553 struct rdtgroup *rdtgrp;
1554
1555 rdtgrp = dev_get_drvdata(dev);
1556 if (mode)
1557 *mode = 0600;
1558 return kasprintf(GFP_KERNEL, "pseudo_lock/%s", rdtgrp->kn->name);
1559}
1560
1561int rdt_pseudo_lock_init(void)
1562{
1563 int ret;
1564
1565 ret = register_chrdev(0, "pseudo_lock", &pseudo_lock_dev_fops);
1566 if (ret < 0)
1567 return ret;
1568
1569 pseudo_lock_major = ret;
1570
1571 pseudo_lock_class = class_create(THIS_MODULE, "pseudo_lock");
1572 if (IS_ERR(pseudo_lock_class)) {
1573 ret = PTR_ERR(pseudo_lock_class);
1574 unregister_chrdev(pseudo_lock_major, "pseudo_lock");
1575 return ret;
1576 }
1577
1578 pseudo_lock_class->devnode = pseudo_lock_devnode;
1579 return 0;
1580}
1581
1582void rdt_pseudo_lock_release(void)
1583{
1584 class_destroy(pseudo_lock_class);
1585 pseudo_lock_class = NULL;
1586 unregister_chrdev(pseudo_lock_major, "pseudo_lock");
1587 pseudo_lock_major = 0;
1588}
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/cpu.h>
15#include <linux/cpumask.h>
16#include <linux/debugfs.h>
17#include <linux/kthread.h>
18#include <linux/mman.h>
19#include <linux/perf_event.h>
20#include <linux/pm_qos.h>
21#include <linux/slab.h>
22#include <linux/uaccess.h>
23
24#include <asm/cacheflush.h>
25#include <asm/cpu_device_id.h>
26#include <asm/resctrl.h>
27#include <asm/perf_event.h>
28
29#include "../../events/perf_event.h" /* For X86_CONFIG() */
30#include "internal.h"
31
32#define CREATE_TRACE_POINTS
33#include "trace.h"
34
35/*
36 * The bits needed to disable hardware prefetching varies based on the
37 * platform. During initialization we will discover which bits to use.
38 */
39static u64 prefetch_disable_bits;
40
41/*
42 * Major number assigned to and shared by all devices exposing
43 * pseudo-locked regions.
44 */
45static unsigned int pseudo_lock_major;
46static unsigned long pseudo_lock_minor_avail = GENMASK(MINORBITS, 0);
47
48static char *pseudo_lock_devnode(const struct device *dev, umode_t *mode)
49{
50 const struct rdtgroup *rdtgrp;
51
52 rdtgrp = dev_get_drvdata(dev);
53 if (mode)
54 *mode = 0600;
55 return kasprintf(GFP_KERNEL, "pseudo_lock/%s", rdtgrp->kn->name);
56}
57
58static const struct class pseudo_lock_class = {
59 .name = "pseudo_lock",
60 .devnode = pseudo_lock_devnode,
61};
62
63/**
64 * get_prefetch_disable_bits - prefetch disable bits of supported platforms
65 * @void: It takes no parameters.
66 *
67 * Capture the list of platforms that have been validated to support
68 * pseudo-locking. This includes testing to ensure pseudo-locked regions
69 * with low cache miss rates can be created under variety of load conditions
70 * as well as that these pseudo-locked regions can maintain their low cache
71 * miss rates under variety of load conditions for significant lengths of time.
72 *
73 * After a platform has been validated to support pseudo-locking its
74 * hardware prefetch disable bits are included here as they are documented
75 * in the SDM.
76 *
77 * When adding a platform here also add support for its cache events to
78 * measure_cycles_perf_fn()
79 *
80 * Return:
81 * If platform is supported, the bits to disable hardware prefetchers, 0
82 * if platform is not supported.
83 */
84static u64 get_prefetch_disable_bits(void)
85{
86 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL ||
87 boot_cpu_data.x86 != 6)
88 return 0;
89
90 switch (boot_cpu_data.x86_vfm) {
91 case INTEL_BROADWELL_X:
92 /*
93 * SDM defines bits of MSR_MISC_FEATURE_CONTROL register
94 * as:
95 * 0 L2 Hardware Prefetcher Disable (R/W)
96 * 1 L2 Adjacent Cache Line Prefetcher Disable (R/W)
97 * 2 DCU Hardware Prefetcher Disable (R/W)
98 * 3 DCU IP Prefetcher Disable (R/W)
99 * 63:4 Reserved
100 */
101 return 0xF;
102 case INTEL_ATOM_GOLDMONT:
103 case INTEL_ATOM_GOLDMONT_PLUS:
104 /*
105 * SDM defines bits of MSR_MISC_FEATURE_CONTROL register
106 * as:
107 * 0 L2 Hardware Prefetcher Disable (R/W)
108 * 1 Reserved
109 * 2 DCU Hardware Prefetcher Disable (R/W)
110 * 63:3 Reserved
111 */
112 return 0x5;
113 }
114
115 return 0;
116}
117
118/**
119 * pseudo_lock_minor_get - Obtain available minor number
120 * @minor: Pointer to where new minor number will be stored
121 *
122 * A bitmask is used to track available minor numbers. Here the next free
123 * minor number is marked as unavailable and returned.
124 *
125 * Return: 0 on success, <0 on failure.
126 */
127static int pseudo_lock_minor_get(unsigned int *minor)
128{
129 unsigned long first_bit;
130
131 first_bit = find_first_bit(&pseudo_lock_minor_avail, MINORBITS);
132
133 if (first_bit == MINORBITS)
134 return -ENOSPC;
135
136 __clear_bit(first_bit, &pseudo_lock_minor_avail);
137 *minor = first_bit;
138
139 return 0;
140}
141
142/**
143 * pseudo_lock_minor_release - Return minor number to available
144 * @minor: The minor number made available
145 */
146static void pseudo_lock_minor_release(unsigned int minor)
147{
148 __set_bit(minor, &pseudo_lock_minor_avail);
149}
150
151/**
152 * region_find_by_minor - Locate a pseudo-lock region by inode minor number
153 * @minor: The minor number of the device representing pseudo-locked region
154 *
155 * When the character device is accessed we need to determine which
156 * pseudo-locked region it belongs to. This is done by matching the minor
157 * number of the device to the pseudo-locked region it belongs.
158 *
159 * Minor numbers are assigned at the time a pseudo-locked region is associated
160 * with a cache instance.
161 *
162 * Return: On success return pointer to resource group owning the pseudo-locked
163 * region, NULL on failure.
164 */
165static struct rdtgroup *region_find_by_minor(unsigned int minor)
166{
167 struct rdtgroup *rdtgrp, *rdtgrp_match = NULL;
168
169 list_for_each_entry(rdtgrp, &rdt_all_groups, rdtgroup_list) {
170 if (rdtgrp->plr && rdtgrp->plr->minor == minor) {
171 rdtgrp_match = rdtgrp;
172 break;
173 }
174 }
175 return rdtgrp_match;
176}
177
178/**
179 * struct pseudo_lock_pm_req - A power management QoS request list entry
180 * @list: Entry within the @pm_reqs list for a pseudo-locked region
181 * @req: PM QoS request
182 */
183struct pseudo_lock_pm_req {
184 struct list_head list;
185 struct dev_pm_qos_request req;
186};
187
188static void pseudo_lock_cstates_relax(struct pseudo_lock_region *plr)
189{
190 struct pseudo_lock_pm_req *pm_req, *next;
191
192 list_for_each_entry_safe(pm_req, next, &plr->pm_reqs, list) {
193 dev_pm_qos_remove_request(&pm_req->req);
194 list_del(&pm_req->list);
195 kfree(pm_req);
196 }
197}
198
199/**
200 * pseudo_lock_cstates_constrain - Restrict cores from entering C6
201 * @plr: Pseudo-locked region
202 *
203 * To prevent the cache from being affected by power management entering
204 * C6 has to be avoided. This is accomplished by requesting a latency
205 * requirement lower than lowest C6 exit latency of all supported
206 * platforms as found in the cpuidle state tables in the intel_idle driver.
207 * At this time it is possible to do so with a single latency requirement
208 * for all supported platforms.
209 *
210 * Since Goldmont is supported, which is affected by X86_BUG_MONITOR,
211 * the ACPI latencies need to be considered while keeping in mind that C2
212 * may be set to map to deeper sleep states. In this case the latency
213 * requirement needs to prevent entering C2 also.
214 *
215 * Return: 0 on success, <0 on failure
216 */
217static int pseudo_lock_cstates_constrain(struct pseudo_lock_region *plr)
218{
219 struct pseudo_lock_pm_req *pm_req;
220 int cpu;
221 int ret;
222
223 for_each_cpu(cpu, &plr->d->hdr.cpu_mask) {
224 pm_req = kzalloc(sizeof(*pm_req), GFP_KERNEL);
225 if (!pm_req) {
226 rdt_last_cmd_puts("Failure to allocate memory for PM QoS\n");
227 ret = -ENOMEM;
228 goto out_err;
229 }
230 ret = dev_pm_qos_add_request(get_cpu_device(cpu),
231 &pm_req->req,
232 DEV_PM_QOS_RESUME_LATENCY,
233 30);
234 if (ret < 0) {
235 rdt_last_cmd_printf("Failed to add latency req CPU%d\n",
236 cpu);
237 kfree(pm_req);
238 ret = -1;
239 goto out_err;
240 }
241 list_add(&pm_req->list, &plr->pm_reqs);
242 }
243
244 return 0;
245
246out_err:
247 pseudo_lock_cstates_relax(plr);
248 return ret;
249}
250
251/**
252 * pseudo_lock_region_clear - Reset pseudo-lock region data
253 * @plr: pseudo-lock region
254 *
255 * All content of the pseudo-locked region is reset - any memory allocated
256 * freed.
257 *
258 * Return: void
259 */
260static void pseudo_lock_region_clear(struct pseudo_lock_region *plr)
261{
262 plr->size = 0;
263 plr->line_size = 0;
264 kfree(plr->kmem);
265 plr->kmem = NULL;
266 plr->s = NULL;
267 if (plr->d)
268 plr->d->plr = NULL;
269 plr->d = NULL;
270 plr->cbm = 0;
271 plr->debugfs_dir = NULL;
272}
273
274/**
275 * pseudo_lock_region_init - Initialize pseudo-lock region information
276 * @plr: pseudo-lock region
277 *
278 * Called after user provided a schemata to be pseudo-locked. From the
279 * schemata the &struct pseudo_lock_region is on entry already initialized
280 * with the resource, domain, and capacity bitmask. Here the information
281 * required for pseudo-locking is deduced from this data and &struct
282 * pseudo_lock_region initialized further. This information includes:
283 * - size in bytes of the region to be pseudo-locked
284 * - cache line size to know the stride with which data needs to be accessed
285 * to be pseudo-locked
286 * - a cpu associated with the cache instance on which the pseudo-locking
287 * flow can be executed
288 *
289 * Return: 0 on success, <0 on failure. Descriptive error will be written
290 * to last_cmd_status buffer.
291 */
292static int pseudo_lock_region_init(struct pseudo_lock_region *plr)
293{
294 enum resctrl_scope scope = plr->s->res->ctrl_scope;
295 struct cacheinfo *ci;
296 int ret;
297
298 if (WARN_ON_ONCE(scope != RESCTRL_L2_CACHE && scope != RESCTRL_L3_CACHE))
299 return -ENODEV;
300
301 /* Pick the first cpu we find that is associated with the cache. */
302 plr->cpu = cpumask_first(&plr->d->hdr.cpu_mask);
303
304 if (!cpu_online(plr->cpu)) {
305 rdt_last_cmd_printf("CPU %u associated with cache not online\n",
306 plr->cpu);
307 ret = -ENODEV;
308 goto out_region;
309 }
310
311 ci = get_cpu_cacheinfo_level(plr->cpu, scope);
312 if (ci) {
313 plr->line_size = ci->coherency_line_size;
314 plr->size = rdtgroup_cbm_to_size(plr->s->res, plr->d, plr->cbm);
315 return 0;
316 }
317
318 ret = -1;
319 rdt_last_cmd_puts("Unable to determine cache line size\n");
320out_region:
321 pseudo_lock_region_clear(plr);
322 return ret;
323}
324
325/**
326 * pseudo_lock_init - Initialize a pseudo-lock region
327 * @rdtgrp: resource group to which new pseudo-locked region will belong
328 *
329 * A pseudo-locked region is associated with a resource group. When this
330 * association is created the pseudo-locked region is initialized. The
331 * details of the pseudo-locked region are not known at this time so only
332 * allocation is done and association established.
333 *
334 * Return: 0 on success, <0 on failure
335 */
336static int pseudo_lock_init(struct rdtgroup *rdtgrp)
337{
338 struct pseudo_lock_region *plr;
339
340 plr = kzalloc(sizeof(*plr), GFP_KERNEL);
341 if (!plr)
342 return -ENOMEM;
343
344 init_waitqueue_head(&plr->lock_thread_wq);
345 INIT_LIST_HEAD(&plr->pm_reqs);
346 rdtgrp->plr = plr;
347 return 0;
348}
349
350/**
351 * pseudo_lock_region_alloc - Allocate kernel memory that will be pseudo-locked
352 * @plr: pseudo-lock region
353 *
354 * Initialize the details required to set up the pseudo-locked region and
355 * allocate the contiguous memory that will be pseudo-locked to the cache.
356 *
357 * Return: 0 on success, <0 on failure. Descriptive error will be written
358 * to last_cmd_status buffer.
359 */
360static int pseudo_lock_region_alloc(struct pseudo_lock_region *plr)
361{
362 int ret;
363
364 ret = pseudo_lock_region_init(plr);
365 if (ret < 0)
366 return ret;
367
368 /*
369 * We do not yet support contiguous regions larger than
370 * KMALLOC_MAX_SIZE.
371 */
372 if (plr->size > KMALLOC_MAX_SIZE) {
373 rdt_last_cmd_puts("Requested region exceeds maximum size\n");
374 ret = -E2BIG;
375 goto out_region;
376 }
377
378 plr->kmem = kzalloc(plr->size, GFP_KERNEL);
379 if (!plr->kmem) {
380 rdt_last_cmd_puts("Unable to allocate memory\n");
381 ret = -ENOMEM;
382 goto out_region;
383 }
384
385 ret = 0;
386 goto out;
387out_region:
388 pseudo_lock_region_clear(plr);
389out:
390 return ret;
391}
392
393/**
394 * pseudo_lock_free - Free a pseudo-locked region
395 * @rdtgrp: resource group to which pseudo-locked region belonged
396 *
397 * The pseudo-locked region's resources have already been released, or not
398 * yet created at this point. Now it can be freed and disassociated from the
399 * resource group.
400 *
401 * Return: void
402 */
403static void pseudo_lock_free(struct rdtgroup *rdtgrp)
404{
405 pseudo_lock_region_clear(rdtgrp->plr);
406 kfree(rdtgrp->plr);
407 rdtgrp->plr = NULL;
408}
409
410/**
411 * pseudo_lock_fn - Load kernel memory into cache
412 * @_rdtgrp: resource group to which pseudo-lock region belongs
413 *
414 * This is the core pseudo-locking flow.
415 *
416 * First we ensure that the kernel memory cannot be found in the cache.
417 * Then, while taking care that there will be as little interference as
418 * possible, the memory to be loaded is accessed while core is running
419 * with class of service set to the bitmask of the pseudo-locked region.
420 * After this is complete no future CAT allocations will be allowed to
421 * overlap with this bitmask.
422 *
423 * Local register variables are utilized to ensure that the memory region
424 * to be locked is the only memory access made during the critical locking
425 * loop.
426 *
427 * Return: 0. Waiter on waitqueue will be woken on completion.
428 */
429static int pseudo_lock_fn(void *_rdtgrp)
430{
431 struct rdtgroup *rdtgrp = _rdtgrp;
432 struct pseudo_lock_region *plr = rdtgrp->plr;
433 u32 rmid_p, closid_p;
434 unsigned long i;
435 u64 saved_msr;
436#ifdef CONFIG_KASAN
437 /*
438 * The registers used for local register variables are also used
439 * when KASAN is active. When KASAN is active we use a regular
440 * variable to ensure we always use a valid pointer, but the cost
441 * is that this variable will enter the cache through evicting the
442 * memory we are trying to lock into the cache. Thus expect lower
443 * pseudo-locking success rate when KASAN is active.
444 */
445 unsigned int line_size;
446 unsigned int size;
447 void *mem_r;
448#else
449 register unsigned int line_size asm("esi");
450 register unsigned int size asm("edi");
451 register void *mem_r asm(_ASM_BX);
452#endif /* CONFIG_KASAN */
453
454 /*
455 * Make sure none of the allocated memory is cached. If it is we
456 * will get a cache hit in below loop from outside of pseudo-locked
457 * region.
458 * wbinvd (as opposed to clflush/clflushopt) is required to
459 * increase likelihood that allocated cache portion will be filled
460 * with associated memory.
461 */
462 native_wbinvd();
463
464 /*
465 * Always called with interrupts enabled. By disabling interrupts
466 * ensure that we will not be preempted during this critical section.
467 */
468 local_irq_disable();
469
470 /*
471 * Call wrmsr and rdmsr as directly as possible to avoid tracing
472 * clobbering local register variables or affecting cache accesses.
473 *
474 * Disable the hardware prefetcher so that when the end of the memory
475 * being pseudo-locked is reached the hardware will not read beyond
476 * the buffer and evict pseudo-locked memory read earlier from the
477 * cache.
478 */
479 saved_msr = __rdmsr(MSR_MISC_FEATURE_CONTROL);
480 __wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
481 closid_p = this_cpu_read(pqr_state.cur_closid);
482 rmid_p = this_cpu_read(pqr_state.cur_rmid);
483 mem_r = plr->kmem;
484 size = plr->size;
485 line_size = plr->line_size;
486 /*
487 * Critical section begin: start by writing the closid associated
488 * with the capacity bitmask of the cache region being
489 * pseudo-locked followed by reading of kernel memory to load it
490 * into the cache.
491 */
492 __wrmsr(MSR_IA32_PQR_ASSOC, rmid_p, rdtgrp->closid);
493 /*
494 * Cache was flushed earlier. Now access kernel memory to read it
495 * into cache region associated with just activated plr->closid.
496 * Loop over data twice:
497 * - In first loop the cache region is shared with the page walker
498 * as it populates the paging structure caches (including TLB).
499 * - In the second loop the paging structure caches are used and
500 * cache region is populated with the memory being referenced.
501 */
502 for (i = 0; i < size; i += PAGE_SIZE) {
503 /*
504 * Add a barrier to prevent speculative execution of this
505 * loop reading beyond the end of the buffer.
506 */
507 rmb();
508 asm volatile("mov (%0,%1,1), %%eax\n\t"
509 :
510 : "r" (mem_r), "r" (i)
511 : "%eax", "memory");
512 }
513 for (i = 0; i < size; i += line_size) {
514 /*
515 * Add a barrier to prevent speculative execution of this
516 * loop reading beyond the end of the buffer.
517 */
518 rmb();
519 asm volatile("mov (%0,%1,1), %%eax\n\t"
520 :
521 : "r" (mem_r), "r" (i)
522 : "%eax", "memory");
523 }
524 /*
525 * Critical section end: restore closid with capacity bitmask that
526 * does not overlap with pseudo-locked region.
527 */
528 __wrmsr(MSR_IA32_PQR_ASSOC, rmid_p, closid_p);
529
530 /* Re-enable the hardware prefetcher(s) */
531 wrmsrl(MSR_MISC_FEATURE_CONTROL, saved_msr);
532 local_irq_enable();
533
534 plr->thread_done = 1;
535 wake_up_interruptible(&plr->lock_thread_wq);
536 return 0;
537}
538
539/**
540 * rdtgroup_monitor_in_progress - Test if monitoring in progress
541 * @rdtgrp: resource group being queried
542 *
543 * Return: 1 if monitor groups have been created for this resource
544 * group, 0 otherwise.
545 */
546static int rdtgroup_monitor_in_progress(struct rdtgroup *rdtgrp)
547{
548 return !list_empty(&rdtgrp->mon.crdtgrp_list);
549}
550
551/**
552 * rdtgroup_locksetup_user_restrict - Restrict user access to group
553 * @rdtgrp: resource group needing access restricted
554 *
555 * A resource group used for cache pseudo-locking cannot have cpus or tasks
556 * assigned to it. This is communicated to the user by restricting access
557 * to all the files that can be used to make such changes.
558 *
559 * Permissions restored with rdtgroup_locksetup_user_restore()
560 *
561 * Return: 0 on success, <0 on failure. If a failure occurs during the
562 * restriction of access an attempt will be made to restore permissions but
563 * the state of the mode of these files will be uncertain when a failure
564 * occurs.
565 */
566static int rdtgroup_locksetup_user_restrict(struct rdtgroup *rdtgrp)
567{
568 int ret;
569
570 ret = rdtgroup_kn_mode_restrict(rdtgrp, "tasks");
571 if (ret)
572 return ret;
573
574 ret = rdtgroup_kn_mode_restrict(rdtgrp, "cpus");
575 if (ret)
576 goto err_tasks;
577
578 ret = rdtgroup_kn_mode_restrict(rdtgrp, "cpus_list");
579 if (ret)
580 goto err_cpus;
581
582 if (resctrl_arch_mon_capable()) {
583 ret = rdtgroup_kn_mode_restrict(rdtgrp, "mon_groups");
584 if (ret)
585 goto err_cpus_list;
586 }
587
588 ret = 0;
589 goto out;
590
591err_cpus_list:
592 rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0777);
593err_cpus:
594 rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0777);
595err_tasks:
596 rdtgroup_kn_mode_restore(rdtgrp, "tasks", 0777);
597out:
598 return ret;
599}
600
601/**
602 * rdtgroup_locksetup_user_restore - Restore user access to group
603 * @rdtgrp: resource group needing access restored
604 *
605 * Restore all file access previously removed using
606 * rdtgroup_locksetup_user_restrict()
607 *
608 * Return: 0 on success, <0 on failure. If a failure occurs during the
609 * restoration of access an attempt will be made to restrict permissions
610 * again but the state of the mode of these files will be uncertain when
611 * a failure occurs.
612 */
613static int rdtgroup_locksetup_user_restore(struct rdtgroup *rdtgrp)
614{
615 int ret;
616
617 ret = rdtgroup_kn_mode_restore(rdtgrp, "tasks", 0777);
618 if (ret)
619 return ret;
620
621 ret = rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0777);
622 if (ret)
623 goto err_tasks;
624
625 ret = rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0777);
626 if (ret)
627 goto err_cpus;
628
629 if (resctrl_arch_mon_capable()) {
630 ret = rdtgroup_kn_mode_restore(rdtgrp, "mon_groups", 0777);
631 if (ret)
632 goto err_cpus_list;
633 }
634
635 ret = 0;
636 goto out;
637
638err_cpus_list:
639 rdtgroup_kn_mode_restrict(rdtgrp, "cpus_list");
640err_cpus:
641 rdtgroup_kn_mode_restrict(rdtgrp, "cpus");
642err_tasks:
643 rdtgroup_kn_mode_restrict(rdtgrp, "tasks");
644out:
645 return ret;
646}
647
648/**
649 * rdtgroup_locksetup_enter - Resource group enters locksetup mode
650 * @rdtgrp: resource group requested to enter locksetup mode
651 *
652 * A resource group enters locksetup mode to reflect that it would be used
653 * to represent a pseudo-locked region and is in the process of being set
654 * up to do so. A resource group used for a pseudo-locked region would
655 * lose the closid associated with it so we cannot allow it to have any
656 * tasks or cpus assigned nor permit tasks or cpus to be assigned in the
657 * future. Monitoring of a pseudo-locked region is not allowed either.
658 *
659 * The above and more restrictions on a pseudo-locked region are checked
660 * for and enforced before the resource group enters the locksetup mode.
661 *
662 * Returns: 0 if the resource group successfully entered locksetup mode, <0
663 * on failure. On failure the last_cmd_status buffer is updated with text to
664 * communicate details of failure to the user.
665 */
666int rdtgroup_locksetup_enter(struct rdtgroup *rdtgrp)
667{
668 int ret;
669
670 /*
671 * The default resource group can neither be removed nor lose the
672 * default closid associated with it.
673 */
674 if (rdtgrp == &rdtgroup_default) {
675 rdt_last_cmd_puts("Cannot pseudo-lock default group\n");
676 return -EINVAL;
677 }
678
679 /*
680 * Cache Pseudo-locking not supported when CDP is enabled.
681 *
682 * Some things to consider if you would like to enable this
683 * support (using L3 CDP as example):
684 * - When CDP is enabled two separate resources are exposed,
685 * L3DATA and L3CODE, but they are actually on the same cache.
686 * The implication for pseudo-locking is that if a
687 * pseudo-locked region is created on a domain of one
688 * resource (eg. L3CODE), then a pseudo-locked region cannot
689 * be created on that same domain of the other resource
690 * (eg. L3DATA). This is because the creation of a
691 * pseudo-locked region involves a call to wbinvd that will
692 * affect all cache allocations on particular domain.
693 * - Considering the previous, it may be possible to only
694 * expose one of the CDP resources to pseudo-locking and
695 * hide the other. For example, we could consider to only
696 * expose L3DATA and since the L3 cache is unified it is
697 * still possible to place instructions there are execute it.
698 * - If only one region is exposed to pseudo-locking we should
699 * still keep in mind that availability of a portion of cache
700 * for pseudo-locking should take into account both resources.
701 * Similarly, if a pseudo-locked region is created in one
702 * resource, the portion of cache used by it should be made
703 * unavailable to all future allocations from both resources.
704 */
705 if (resctrl_arch_get_cdp_enabled(RDT_RESOURCE_L3) ||
706 resctrl_arch_get_cdp_enabled(RDT_RESOURCE_L2)) {
707 rdt_last_cmd_puts("CDP enabled\n");
708 return -EINVAL;
709 }
710
711 /*
712 * Not knowing the bits to disable prefetching implies that this
713 * platform does not support Cache Pseudo-Locking.
714 */
715 prefetch_disable_bits = get_prefetch_disable_bits();
716 if (prefetch_disable_bits == 0) {
717 rdt_last_cmd_puts("Pseudo-locking not supported\n");
718 return -EINVAL;
719 }
720
721 if (rdtgroup_monitor_in_progress(rdtgrp)) {
722 rdt_last_cmd_puts("Monitoring in progress\n");
723 return -EINVAL;
724 }
725
726 if (rdtgroup_tasks_assigned(rdtgrp)) {
727 rdt_last_cmd_puts("Tasks assigned to resource group\n");
728 return -EINVAL;
729 }
730
731 if (!cpumask_empty(&rdtgrp->cpu_mask)) {
732 rdt_last_cmd_puts("CPUs assigned to resource group\n");
733 return -EINVAL;
734 }
735
736 if (rdtgroup_locksetup_user_restrict(rdtgrp)) {
737 rdt_last_cmd_puts("Unable to modify resctrl permissions\n");
738 return -EIO;
739 }
740
741 ret = pseudo_lock_init(rdtgrp);
742 if (ret) {
743 rdt_last_cmd_puts("Unable to init pseudo-lock region\n");
744 goto out_release;
745 }
746
747 /*
748 * If this system is capable of monitoring a rmid would have been
749 * allocated when the control group was created. This is not needed
750 * anymore when this group would be used for pseudo-locking. This
751 * is safe to call on platforms not capable of monitoring.
752 */
753 free_rmid(rdtgrp->closid, rdtgrp->mon.rmid);
754
755 ret = 0;
756 goto out;
757
758out_release:
759 rdtgroup_locksetup_user_restore(rdtgrp);
760out:
761 return ret;
762}
763
764/**
765 * rdtgroup_locksetup_exit - resource group exist locksetup mode
766 * @rdtgrp: resource group
767 *
768 * When a resource group exits locksetup mode the earlier restrictions are
769 * lifted.
770 *
771 * Return: 0 on success, <0 on failure
772 */
773int rdtgroup_locksetup_exit(struct rdtgroup *rdtgrp)
774{
775 int ret;
776
777 if (resctrl_arch_mon_capable()) {
778 ret = alloc_rmid(rdtgrp->closid);
779 if (ret < 0) {
780 rdt_last_cmd_puts("Out of RMIDs\n");
781 return ret;
782 }
783 rdtgrp->mon.rmid = ret;
784 }
785
786 ret = rdtgroup_locksetup_user_restore(rdtgrp);
787 if (ret) {
788 free_rmid(rdtgrp->closid, rdtgrp->mon.rmid);
789 return ret;
790 }
791
792 pseudo_lock_free(rdtgrp);
793 return 0;
794}
795
796/**
797 * rdtgroup_cbm_overlaps_pseudo_locked - Test if CBM or portion is pseudo-locked
798 * @d: RDT domain
799 * @cbm: CBM to test
800 *
801 * @d represents a cache instance and @cbm a capacity bitmask that is
802 * considered for it. Determine if @cbm overlaps with any existing
803 * pseudo-locked region on @d.
804 *
805 * @cbm is unsigned long, even if only 32 bits are used, to make the
806 * bitmap functions work correctly.
807 *
808 * Return: true if @cbm overlaps with pseudo-locked region on @d, false
809 * otherwise.
810 */
811bool rdtgroup_cbm_overlaps_pseudo_locked(struct rdt_ctrl_domain *d, unsigned long cbm)
812{
813 unsigned int cbm_len;
814 unsigned long cbm_b;
815
816 if (d->plr) {
817 cbm_len = d->plr->s->res->cache.cbm_len;
818 cbm_b = d->plr->cbm;
819 if (bitmap_intersects(&cbm, &cbm_b, cbm_len))
820 return true;
821 }
822 return false;
823}
824
825/**
826 * rdtgroup_pseudo_locked_in_hierarchy - Pseudo-locked region in cache hierarchy
827 * @d: RDT domain under test
828 *
829 * The setup of a pseudo-locked region affects all cache instances within
830 * the hierarchy of the region. It is thus essential to know if any
831 * pseudo-locked regions exist within a cache hierarchy to prevent any
832 * attempts to create new pseudo-locked regions in the same hierarchy.
833 *
834 * Return: true if a pseudo-locked region exists in the hierarchy of @d or
835 * if it is not possible to test due to memory allocation issue,
836 * false otherwise.
837 */
838bool rdtgroup_pseudo_locked_in_hierarchy(struct rdt_ctrl_domain *d)
839{
840 struct rdt_ctrl_domain *d_i;
841 cpumask_var_t cpu_with_psl;
842 struct rdt_resource *r;
843 bool ret = false;
844
845 /* Walking r->domains, ensure it can't race with cpuhp */
846 lockdep_assert_cpus_held();
847
848 if (!zalloc_cpumask_var(&cpu_with_psl, GFP_KERNEL))
849 return true;
850
851 /*
852 * First determine which cpus have pseudo-locked regions
853 * associated with them.
854 */
855 for_each_alloc_capable_rdt_resource(r) {
856 list_for_each_entry(d_i, &r->ctrl_domains, hdr.list) {
857 if (d_i->plr)
858 cpumask_or(cpu_with_psl, cpu_with_psl,
859 &d_i->hdr.cpu_mask);
860 }
861 }
862
863 /*
864 * Next test if new pseudo-locked region would intersect with
865 * existing region.
866 */
867 if (cpumask_intersects(&d->hdr.cpu_mask, cpu_with_psl))
868 ret = true;
869
870 free_cpumask_var(cpu_with_psl);
871 return ret;
872}
873
874/**
875 * measure_cycles_lat_fn - Measure cycle latency to read pseudo-locked memory
876 * @_plr: pseudo-lock region to measure
877 *
878 * There is no deterministic way to test if a memory region is cached. One
879 * way is to measure how long it takes to read the memory, the speed of
880 * access is a good way to learn how close to the cpu the data was. Even
881 * more, if the prefetcher is disabled and the memory is read at a stride
882 * of half the cache line, then a cache miss will be easy to spot since the
883 * read of the first half would be significantly slower than the read of
884 * the second half.
885 *
886 * Return: 0. Waiter on waitqueue will be woken on completion.
887 */
888static int measure_cycles_lat_fn(void *_plr)
889{
890 struct pseudo_lock_region *plr = _plr;
891 u32 saved_low, saved_high;
892 unsigned long i;
893 u64 start, end;
894 void *mem_r;
895
896 local_irq_disable();
897 /*
898 * Disable hardware prefetchers.
899 */
900 rdmsr(MSR_MISC_FEATURE_CONTROL, saved_low, saved_high);
901 wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
902 mem_r = READ_ONCE(plr->kmem);
903 /*
904 * Dummy execute of the time measurement to load the needed
905 * instructions into the L1 instruction cache.
906 */
907 start = rdtsc_ordered();
908 for (i = 0; i < plr->size; i += 32) {
909 start = rdtsc_ordered();
910 asm volatile("mov (%0,%1,1), %%eax\n\t"
911 :
912 : "r" (mem_r), "r" (i)
913 : "%eax", "memory");
914 end = rdtsc_ordered();
915 trace_pseudo_lock_mem_latency((u32)(end - start));
916 }
917 wrmsr(MSR_MISC_FEATURE_CONTROL, saved_low, saved_high);
918 local_irq_enable();
919 plr->thread_done = 1;
920 wake_up_interruptible(&plr->lock_thread_wq);
921 return 0;
922}
923
924/*
925 * Create a perf_event_attr for the hit and miss perf events that will
926 * be used during the performance measurement. A perf_event maintains
927 * a pointer to its perf_event_attr so a unique attribute structure is
928 * created for each perf_event.
929 *
930 * The actual configuration of the event is set right before use in order
931 * to use the X86_CONFIG macro.
932 */
933static struct perf_event_attr perf_miss_attr = {
934 .type = PERF_TYPE_RAW,
935 .size = sizeof(struct perf_event_attr),
936 .pinned = 1,
937 .disabled = 0,
938 .exclude_user = 1,
939};
940
941static struct perf_event_attr perf_hit_attr = {
942 .type = PERF_TYPE_RAW,
943 .size = sizeof(struct perf_event_attr),
944 .pinned = 1,
945 .disabled = 0,
946 .exclude_user = 1,
947};
948
949struct residency_counts {
950 u64 miss_before, hits_before;
951 u64 miss_after, hits_after;
952};
953
954static int measure_residency_fn(struct perf_event_attr *miss_attr,
955 struct perf_event_attr *hit_attr,
956 struct pseudo_lock_region *plr,
957 struct residency_counts *counts)
958{
959 u64 hits_before = 0, hits_after = 0, miss_before = 0, miss_after = 0;
960 struct perf_event *miss_event, *hit_event;
961 int hit_pmcnum, miss_pmcnum;
962 u32 saved_low, saved_high;
963 unsigned int line_size;
964 unsigned int size;
965 unsigned long i;
966 void *mem_r;
967 u64 tmp;
968
969 miss_event = perf_event_create_kernel_counter(miss_attr, plr->cpu,
970 NULL, NULL, NULL);
971 if (IS_ERR(miss_event))
972 goto out;
973
974 hit_event = perf_event_create_kernel_counter(hit_attr, plr->cpu,
975 NULL, NULL, NULL);
976 if (IS_ERR(hit_event))
977 goto out_miss;
978
979 local_irq_disable();
980 /*
981 * Check any possible error state of events used by performing
982 * one local read.
983 */
984 if (perf_event_read_local(miss_event, &tmp, NULL, NULL)) {
985 local_irq_enable();
986 goto out_hit;
987 }
988 if (perf_event_read_local(hit_event, &tmp, NULL, NULL)) {
989 local_irq_enable();
990 goto out_hit;
991 }
992
993 /*
994 * Disable hardware prefetchers.
995 */
996 rdmsr(MSR_MISC_FEATURE_CONTROL, saved_low, saved_high);
997 wrmsr(MSR_MISC_FEATURE_CONTROL, prefetch_disable_bits, 0x0);
998
999 /* Initialize rest of local variables */
1000 /*
1001 * Performance event has been validated right before this with
1002 * interrupts disabled - it is thus safe to read the counter index.
1003 */
1004 miss_pmcnum = x86_perf_rdpmc_index(miss_event);
1005 hit_pmcnum = x86_perf_rdpmc_index(hit_event);
1006 line_size = READ_ONCE(plr->line_size);
1007 mem_r = READ_ONCE(plr->kmem);
1008 size = READ_ONCE(plr->size);
1009
1010 /*
1011 * Read counter variables twice - first to load the instructions
1012 * used in L1 cache, second to capture accurate value that does not
1013 * include cache misses incurred because of instruction loads.
1014 */
1015 rdpmcl(hit_pmcnum, hits_before);
1016 rdpmcl(miss_pmcnum, miss_before);
1017 /*
1018 * From SDM: Performing back-to-back fast reads are not guaranteed
1019 * to be monotonic.
1020 * Use LFENCE to ensure all previous instructions are retired
1021 * before proceeding.
1022 */
1023 rmb();
1024 rdpmcl(hit_pmcnum, hits_before);
1025 rdpmcl(miss_pmcnum, miss_before);
1026 /*
1027 * Use LFENCE to ensure all previous instructions are retired
1028 * before proceeding.
1029 */
1030 rmb();
1031 for (i = 0; i < size; i += line_size) {
1032 /*
1033 * Add a barrier to prevent speculative execution of this
1034 * loop reading beyond the end of the buffer.
1035 */
1036 rmb();
1037 asm volatile("mov (%0,%1,1), %%eax\n\t"
1038 :
1039 : "r" (mem_r), "r" (i)
1040 : "%eax", "memory");
1041 }
1042 /*
1043 * Use LFENCE to ensure all previous instructions are retired
1044 * before proceeding.
1045 */
1046 rmb();
1047 rdpmcl(hit_pmcnum, hits_after);
1048 rdpmcl(miss_pmcnum, miss_after);
1049 /*
1050 * Use LFENCE to ensure all previous instructions are retired
1051 * before proceeding.
1052 */
1053 rmb();
1054 /* Re-enable hardware prefetchers */
1055 wrmsr(MSR_MISC_FEATURE_CONTROL, saved_low, saved_high);
1056 local_irq_enable();
1057out_hit:
1058 perf_event_release_kernel(hit_event);
1059out_miss:
1060 perf_event_release_kernel(miss_event);
1061out:
1062 /*
1063 * All counts will be zero on failure.
1064 */
1065 counts->miss_before = miss_before;
1066 counts->hits_before = hits_before;
1067 counts->miss_after = miss_after;
1068 counts->hits_after = hits_after;
1069 return 0;
1070}
1071
1072static int measure_l2_residency(void *_plr)
1073{
1074 struct pseudo_lock_region *plr = _plr;
1075 struct residency_counts counts = {0};
1076
1077 /*
1078 * Non-architectural event for the Goldmont Microarchitecture
1079 * from Intel x86 Architecture Software Developer Manual (SDM):
1080 * MEM_LOAD_UOPS_RETIRED D1H (event number)
1081 * Umask values:
1082 * L2_HIT 02H
1083 * L2_MISS 10H
1084 */
1085 switch (boot_cpu_data.x86_vfm) {
1086 case INTEL_ATOM_GOLDMONT:
1087 case INTEL_ATOM_GOLDMONT_PLUS:
1088 perf_miss_attr.config = X86_CONFIG(.event = 0xd1,
1089 .umask = 0x10);
1090 perf_hit_attr.config = X86_CONFIG(.event = 0xd1,
1091 .umask = 0x2);
1092 break;
1093 default:
1094 goto out;
1095 }
1096
1097 measure_residency_fn(&perf_miss_attr, &perf_hit_attr, plr, &counts);
1098 /*
1099 * If a failure prevented the measurements from succeeding
1100 * tracepoints will still be written and all counts will be zero.
1101 */
1102 trace_pseudo_lock_l2(counts.hits_after - counts.hits_before,
1103 counts.miss_after - counts.miss_before);
1104out:
1105 plr->thread_done = 1;
1106 wake_up_interruptible(&plr->lock_thread_wq);
1107 return 0;
1108}
1109
1110static int measure_l3_residency(void *_plr)
1111{
1112 struct pseudo_lock_region *plr = _plr;
1113 struct residency_counts counts = {0};
1114
1115 /*
1116 * On Broadwell Microarchitecture the MEM_LOAD_UOPS_RETIRED event
1117 * has two "no fix" errata associated with it: BDM35 and BDM100. On
1118 * this platform the following events are used instead:
1119 * LONGEST_LAT_CACHE 2EH (Documented in SDM)
1120 * REFERENCE 4FH
1121 * MISS 41H
1122 */
1123
1124 switch (boot_cpu_data.x86_vfm) {
1125 case INTEL_BROADWELL_X:
1126 /* On BDW the hit event counts references, not hits */
1127 perf_hit_attr.config = X86_CONFIG(.event = 0x2e,
1128 .umask = 0x4f);
1129 perf_miss_attr.config = X86_CONFIG(.event = 0x2e,
1130 .umask = 0x41);
1131 break;
1132 default:
1133 goto out;
1134 }
1135
1136 measure_residency_fn(&perf_miss_attr, &perf_hit_attr, plr, &counts);
1137 /*
1138 * If a failure prevented the measurements from succeeding
1139 * tracepoints will still be written and all counts will be zero.
1140 */
1141
1142 counts.miss_after -= counts.miss_before;
1143 if (boot_cpu_data.x86_vfm == INTEL_BROADWELL_X) {
1144 /*
1145 * On BDW references and misses are counted, need to adjust.
1146 * Sometimes the "hits" counter is a bit more than the
1147 * references, for example, x references but x + 1 hits.
1148 * To not report invalid hit values in this case we treat
1149 * that as misses equal to references.
1150 */
1151 /* First compute the number of cache references measured */
1152 counts.hits_after -= counts.hits_before;
1153 /* Next convert references to cache hits */
1154 counts.hits_after -= min(counts.miss_after, counts.hits_after);
1155 } else {
1156 counts.hits_after -= counts.hits_before;
1157 }
1158
1159 trace_pseudo_lock_l3(counts.hits_after, counts.miss_after);
1160out:
1161 plr->thread_done = 1;
1162 wake_up_interruptible(&plr->lock_thread_wq);
1163 return 0;
1164}
1165
1166/**
1167 * pseudo_lock_measure_cycles - Trigger latency measure to pseudo-locked region
1168 * @rdtgrp: Resource group to which the pseudo-locked region belongs.
1169 * @sel: Selector of which measurement to perform on a pseudo-locked region.
1170 *
1171 * The measurement of latency to access a pseudo-locked region should be
1172 * done from a cpu that is associated with that pseudo-locked region.
1173 * Determine which cpu is associated with this region and start a thread on
1174 * that cpu to perform the measurement, wait for that thread to complete.
1175 *
1176 * Return: 0 on success, <0 on failure
1177 */
1178static int pseudo_lock_measure_cycles(struct rdtgroup *rdtgrp, int sel)
1179{
1180 struct pseudo_lock_region *plr = rdtgrp->plr;
1181 struct task_struct *thread;
1182 unsigned int cpu;
1183 int ret = -1;
1184
1185 cpus_read_lock();
1186 mutex_lock(&rdtgroup_mutex);
1187
1188 if (rdtgrp->flags & RDT_DELETED) {
1189 ret = -ENODEV;
1190 goto out;
1191 }
1192
1193 if (!plr->d) {
1194 ret = -ENODEV;
1195 goto out;
1196 }
1197
1198 plr->thread_done = 0;
1199 cpu = cpumask_first(&plr->d->hdr.cpu_mask);
1200 if (!cpu_online(cpu)) {
1201 ret = -ENODEV;
1202 goto out;
1203 }
1204
1205 plr->cpu = cpu;
1206
1207 if (sel == 1)
1208 thread = kthread_create_on_node(measure_cycles_lat_fn, plr,
1209 cpu_to_node(cpu),
1210 "pseudo_lock_measure/%u",
1211 cpu);
1212 else if (sel == 2)
1213 thread = kthread_create_on_node(measure_l2_residency, plr,
1214 cpu_to_node(cpu),
1215 "pseudo_lock_measure/%u",
1216 cpu);
1217 else if (sel == 3)
1218 thread = kthread_create_on_node(measure_l3_residency, plr,
1219 cpu_to_node(cpu),
1220 "pseudo_lock_measure/%u",
1221 cpu);
1222 else
1223 goto out;
1224
1225 if (IS_ERR(thread)) {
1226 ret = PTR_ERR(thread);
1227 goto out;
1228 }
1229 kthread_bind(thread, cpu);
1230 wake_up_process(thread);
1231
1232 ret = wait_event_interruptible(plr->lock_thread_wq,
1233 plr->thread_done == 1);
1234 if (ret < 0)
1235 goto out;
1236
1237 ret = 0;
1238
1239out:
1240 mutex_unlock(&rdtgroup_mutex);
1241 cpus_read_unlock();
1242 return ret;
1243}
1244
1245static ssize_t pseudo_lock_measure_trigger(struct file *file,
1246 const char __user *user_buf,
1247 size_t count, loff_t *ppos)
1248{
1249 struct rdtgroup *rdtgrp = file->private_data;
1250 size_t buf_size;
1251 char buf[32];
1252 int ret;
1253 int sel;
1254
1255 buf_size = min(count, (sizeof(buf) - 1));
1256 if (copy_from_user(buf, user_buf, buf_size))
1257 return -EFAULT;
1258
1259 buf[buf_size] = '\0';
1260 ret = kstrtoint(buf, 10, &sel);
1261 if (ret == 0) {
1262 if (sel != 1 && sel != 2 && sel != 3)
1263 return -EINVAL;
1264 ret = debugfs_file_get(file->f_path.dentry);
1265 if (ret)
1266 return ret;
1267 ret = pseudo_lock_measure_cycles(rdtgrp, sel);
1268 if (ret == 0)
1269 ret = count;
1270 debugfs_file_put(file->f_path.dentry);
1271 }
1272
1273 return ret;
1274}
1275
1276static const struct file_operations pseudo_measure_fops = {
1277 .write = pseudo_lock_measure_trigger,
1278 .open = simple_open,
1279 .llseek = default_llseek,
1280};
1281
1282/**
1283 * rdtgroup_pseudo_lock_create - Create a pseudo-locked region
1284 * @rdtgrp: resource group to which pseudo-lock region belongs
1285 *
1286 * Called when a resource group in the pseudo-locksetup mode receives a
1287 * valid schemata that should be pseudo-locked. Since the resource group is
1288 * in pseudo-locksetup mode the &struct pseudo_lock_region has already been
1289 * allocated and initialized with the essential information. If a failure
1290 * occurs the resource group remains in the pseudo-locksetup mode with the
1291 * &struct pseudo_lock_region associated with it, but cleared from all
1292 * information and ready for the user to re-attempt pseudo-locking by
1293 * writing the schemata again.
1294 *
1295 * Return: 0 if the pseudo-locked region was successfully pseudo-locked, <0
1296 * on failure. Descriptive error will be written to last_cmd_status buffer.
1297 */
1298int rdtgroup_pseudo_lock_create(struct rdtgroup *rdtgrp)
1299{
1300 struct pseudo_lock_region *plr = rdtgrp->plr;
1301 struct task_struct *thread;
1302 unsigned int new_minor;
1303 struct device *dev;
1304 int ret;
1305
1306 ret = pseudo_lock_region_alloc(plr);
1307 if (ret < 0)
1308 return ret;
1309
1310 ret = pseudo_lock_cstates_constrain(plr);
1311 if (ret < 0) {
1312 ret = -EINVAL;
1313 goto out_region;
1314 }
1315
1316 plr->thread_done = 0;
1317
1318 thread = kthread_create_on_node(pseudo_lock_fn, rdtgrp,
1319 cpu_to_node(plr->cpu),
1320 "pseudo_lock/%u", plr->cpu);
1321 if (IS_ERR(thread)) {
1322 ret = PTR_ERR(thread);
1323 rdt_last_cmd_printf("Locking thread returned error %d\n", ret);
1324 goto out_cstates;
1325 }
1326
1327 kthread_bind(thread, plr->cpu);
1328 wake_up_process(thread);
1329
1330 ret = wait_event_interruptible(plr->lock_thread_wq,
1331 plr->thread_done == 1);
1332 if (ret < 0) {
1333 /*
1334 * If the thread does not get on the CPU for whatever
1335 * reason and the process which sets up the region is
1336 * interrupted then this will leave the thread in runnable
1337 * state and once it gets on the CPU it will dereference
1338 * the cleared, but not freed, plr struct resulting in an
1339 * empty pseudo-locking loop.
1340 */
1341 rdt_last_cmd_puts("Locking thread interrupted\n");
1342 goto out_cstates;
1343 }
1344
1345 ret = pseudo_lock_minor_get(&new_minor);
1346 if (ret < 0) {
1347 rdt_last_cmd_puts("Unable to obtain a new minor number\n");
1348 goto out_cstates;
1349 }
1350
1351 /*
1352 * Unlock access but do not release the reference. The
1353 * pseudo-locked region will still be here on return.
1354 *
1355 * The mutex has to be released temporarily to avoid a potential
1356 * deadlock with the mm->mmap_lock which is obtained in the
1357 * device_create() and debugfs_create_dir() callpath below as well as
1358 * before the mmap() callback is called.
1359 */
1360 mutex_unlock(&rdtgroup_mutex);
1361
1362 if (!IS_ERR_OR_NULL(debugfs_resctrl)) {
1363 plr->debugfs_dir = debugfs_create_dir(rdtgrp->kn->name,
1364 debugfs_resctrl);
1365 if (!IS_ERR_OR_NULL(plr->debugfs_dir))
1366 debugfs_create_file("pseudo_lock_measure", 0200,
1367 plr->debugfs_dir, rdtgrp,
1368 &pseudo_measure_fops);
1369 }
1370
1371 dev = device_create(&pseudo_lock_class, NULL,
1372 MKDEV(pseudo_lock_major, new_minor),
1373 rdtgrp, "%s", rdtgrp->kn->name);
1374
1375 mutex_lock(&rdtgroup_mutex);
1376
1377 if (IS_ERR(dev)) {
1378 ret = PTR_ERR(dev);
1379 rdt_last_cmd_printf("Failed to create character device: %d\n",
1380 ret);
1381 goto out_debugfs;
1382 }
1383
1384 /* We released the mutex - check if group was removed while we did so */
1385 if (rdtgrp->flags & RDT_DELETED) {
1386 ret = -ENODEV;
1387 goto out_device;
1388 }
1389
1390 plr->minor = new_minor;
1391
1392 rdtgrp->mode = RDT_MODE_PSEUDO_LOCKED;
1393 closid_free(rdtgrp->closid);
1394 rdtgroup_kn_mode_restore(rdtgrp, "cpus", 0444);
1395 rdtgroup_kn_mode_restore(rdtgrp, "cpus_list", 0444);
1396
1397 ret = 0;
1398 goto out;
1399
1400out_device:
1401 device_destroy(&pseudo_lock_class, MKDEV(pseudo_lock_major, new_minor));
1402out_debugfs:
1403 debugfs_remove_recursive(plr->debugfs_dir);
1404 pseudo_lock_minor_release(new_minor);
1405out_cstates:
1406 pseudo_lock_cstates_relax(plr);
1407out_region:
1408 pseudo_lock_region_clear(plr);
1409out:
1410 return ret;
1411}
1412
1413/**
1414 * rdtgroup_pseudo_lock_remove - Remove a pseudo-locked region
1415 * @rdtgrp: resource group to which the pseudo-locked region belongs
1416 *
1417 * The removal of a pseudo-locked region can be initiated when the resource
1418 * group is removed from user space via a "rmdir" from userspace or the
1419 * unmount of the resctrl filesystem. On removal the resource group does
1420 * not go back to pseudo-locksetup mode before it is removed, instead it is
1421 * removed directly. There is thus asymmetry with the creation where the
1422 * &struct pseudo_lock_region is removed here while it was not created in
1423 * rdtgroup_pseudo_lock_create().
1424 *
1425 * Return: void
1426 */
1427void rdtgroup_pseudo_lock_remove(struct rdtgroup *rdtgrp)
1428{
1429 struct pseudo_lock_region *plr = rdtgrp->plr;
1430
1431 if (rdtgrp->mode == RDT_MODE_PSEUDO_LOCKSETUP) {
1432 /*
1433 * Default group cannot be a pseudo-locked region so we can
1434 * free closid here.
1435 */
1436 closid_free(rdtgrp->closid);
1437 goto free;
1438 }
1439
1440 pseudo_lock_cstates_relax(plr);
1441 debugfs_remove_recursive(rdtgrp->plr->debugfs_dir);
1442 device_destroy(&pseudo_lock_class, MKDEV(pseudo_lock_major, plr->minor));
1443 pseudo_lock_minor_release(plr->minor);
1444
1445free:
1446 pseudo_lock_free(rdtgrp);
1447}
1448
1449static int pseudo_lock_dev_open(struct inode *inode, struct file *filp)
1450{
1451 struct rdtgroup *rdtgrp;
1452
1453 mutex_lock(&rdtgroup_mutex);
1454
1455 rdtgrp = region_find_by_minor(iminor(inode));
1456 if (!rdtgrp) {
1457 mutex_unlock(&rdtgroup_mutex);
1458 return -ENODEV;
1459 }
1460
1461 filp->private_data = rdtgrp;
1462 atomic_inc(&rdtgrp->waitcount);
1463 /* Perform a non-seekable open - llseek is not supported */
1464 filp->f_mode &= ~(FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE);
1465
1466 mutex_unlock(&rdtgroup_mutex);
1467
1468 return 0;
1469}
1470
1471static int pseudo_lock_dev_release(struct inode *inode, struct file *filp)
1472{
1473 struct rdtgroup *rdtgrp;
1474
1475 mutex_lock(&rdtgroup_mutex);
1476 rdtgrp = filp->private_data;
1477 WARN_ON(!rdtgrp);
1478 if (!rdtgrp) {
1479 mutex_unlock(&rdtgroup_mutex);
1480 return -ENODEV;
1481 }
1482 filp->private_data = NULL;
1483 atomic_dec(&rdtgrp->waitcount);
1484 mutex_unlock(&rdtgroup_mutex);
1485 return 0;
1486}
1487
1488static int pseudo_lock_dev_mremap(struct vm_area_struct *area)
1489{
1490 /* Not supported */
1491 return -EINVAL;
1492}
1493
1494static const struct vm_operations_struct pseudo_mmap_ops = {
1495 .mremap = pseudo_lock_dev_mremap,
1496};
1497
1498static int pseudo_lock_dev_mmap(struct file *filp, struct vm_area_struct *vma)
1499{
1500 unsigned long vsize = vma->vm_end - vma->vm_start;
1501 unsigned long off = vma->vm_pgoff << PAGE_SHIFT;
1502 struct pseudo_lock_region *plr;
1503 struct rdtgroup *rdtgrp;
1504 unsigned long physical;
1505 unsigned long psize;
1506
1507 mutex_lock(&rdtgroup_mutex);
1508
1509 rdtgrp = filp->private_data;
1510 WARN_ON(!rdtgrp);
1511 if (!rdtgrp) {
1512 mutex_unlock(&rdtgroup_mutex);
1513 return -ENODEV;
1514 }
1515
1516 plr = rdtgrp->plr;
1517
1518 if (!plr->d) {
1519 mutex_unlock(&rdtgroup_mutex);
1520 return -ENODEV;
1521 }
1522
1523 /*
1524 * Task is required to run with affinity to the cpus associated
1525 * with the pseudo-locked region. If this is not the case the task
1526 * may be scheduled elsewhere and invalidate entries in the
1527 * pseudo-locked region.
1528 */
1529 if (!cpumask_subset(current->cpus_ptr, &plr->d->hdr.cpu_mask)) {
1530 mutex_unlock(&rdtgroup_mutex);
1531 return -EINVAL;
1532 }
1533
1534 physical = __pa(plr->kmem) >> PAGE_SHIFT;
1535 psize = plr->size - off;
1536
1537 if (off > plr->size) {
1538 mutex_unlock(&rdtgroup_mutex);
1539 return -ENOSPC;
1540 }
1541
1542 /*
1543 * Ensure changes are carried directly to the memory being mapped,
1544 * do not allow copy-on-write mapping.
1545 */
1546 if (!(vma->vm_flags & VM_SHARED)) {
1547 mutex_unlock(&rdtgroup_mutex);
1548 return -EINVAL;
1549 }
1550
1551 if (vsize > psize) {
1552 mutex_unlock(&rdtgroup_mutex);
1553 return -ENOSPC;
1554 }
1555
1556 memset(plr->kmem + off, 0, vsize);
1557
1558 if (remap_pfn_range(vma, vma->vm_start, physical + vma->vm_pgoff,
1559 vsize, vma->vm_page_prot)) {
1560 mutex_unlock(&rdtgroup_mutex);
1561 return -EAGAIN;
1562 }
1563 vma->vm_ops = &pseudo_mmap_ops;
1564 mutex_unlock(&rdtgroup_mutex);
1565 return 0;
1566}
1567
1568static const struct file_operations pseudo_lock_dev_fops = {
1569 .owner = THIS_MODULE,
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
1577int 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
1596void 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}