kvm.c 19 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837
  1. /*
  2. * KVM paravirt_ops implementation
  3. *
  4. * This program is free software; you can redistribute it and/or modify
  5. * it under the terms of the GNU General Public License as published by
  6. * the Free Software Foundation; either version 2 of the License, or
  7. * (at your option) any later version.
  8. *
  9. * This program is distributed in the hope that it will be useful,
  10. * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
  12. * GNU General Public License for more details.
  13. *
  14. * You should have received a copy of the GNU General Public License
  15. * along with this program; if not, write to the Free Software
  16. * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  17. *
  18. * Copyright (C) 2007, Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
  19. * Copyright IBM Corporation, 2007
  20. * Authors: Anthony Liguori <aliguori@us.ibm.com>
  21. */
  22. #include <linux/context_tracking.h>
  23. #include <linux/module.h>
  24. #include <linux/kernel.h>
  25. #include <linux/kvm_para.h>
  26. #include <linux/cpu.h>
  27. #include <linux/mm.h>
  28. #include <linux/highmem.h>
  29. #include <linux/hardirq.h>
  30. #include <linux/notifier.h>
  31. #include <linux/reboot.h>
  32. #include <linux/hash.h>
  33. #include <linux/sched.h>
  34. #include <linux/slab.h>
  35. #include <linux/kprobes.h>
  36. #include <linux/debugfs.h>
  37. #include <linux/nmi.h>
  38. #include <asm/timer.h>
  39. #include <asm/cpu.h>
  40. #include <asm/traps.h>
  41. #include <asm/desc.h>
  42. #include <asm/tlbflush.h>
  43. #include <asm/idle.h>
  44. #include <asm/apic.h>
  45. #include <asm/apicdef.h>
  46. #include <asm/hypervisor.h>
  47. #include <asm/kvm_guest.h>
  48. static int kvmapf = 1;
  49. static int parse_no_kvmapf(char *arg)
  50. {
  51. kvmapf = 0;
  52. return 0;
  53. }
  54. early_param("no-kvmapf", parse_no_kvmapf);
  55. static int steal_acc = 1;
  56. static int parse_no_stealacc(char *arg)
  57. {
  58. steal_acc = 0;
  59. return 0;
  60. }
  61. early_param("no-steal-acc", parse_no_stealacc);
  62. static int kvmclock_vsyscall = 1;
  63. static int parse_no_kvmclock_vsyscall(char *arg)
  64. {
  65. kvmclock_vsyscall = 0;
  66. return 0;
  67. }
  68. early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
  69. static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  70. static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
  71. static int has_steal_clock = 0;
  72. /*
  73. * No need for any "IO delay" on KVM
  74. */
  75. static void kvm_io_delay(void)
  76. {
  77. }
  78. #define KVM_TASK_SLEEP_HASHBITS 8
  79. #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  80. struct kvm_task_sleep_node {
  81. struct hlist_node link;
  82. wait_queue_head_t wq;
  83. u32 token;
  84. int cpu;
  85. bool halted;
  86. };
  87. static struct kvm_task_sleep_head {
  88. spinlock_t lock;
  89. struct hlist_head list;
  90. } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
  91. static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
  92. u32 token)
  93. {
  94. struct hlist_node *p;
  95. hlist_for_each(p, &b->list) {
  96. struct kvm_task_sleep_node *n =
  97. hlist_entry(p, typeof(*n), link);
  98. if (n->token == token)
  99. return n;
  100. }
  101. return NULL;
  102. }
  103. void kvm_async_pf_task_wait(u32 token)
  104. {
  105. u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
  106. struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
  107. struct kvm_task_sleep_node n, *e;
  108. DEFINE_WAIT(wait);
  109. rcu_irq_enter();
  110. spin_lock(&b->lock);
  111. e = _find_apf_task(b, token);
  112. if (e) {
  113. /* dummy entry exist -> wake up was delivered ahead of PF */
  114. hlist_del(&e->link);
  115. kfree(e);
  116. spin_unlock(&b->lock);
  117. rcu_irq_exit();
  118. return;
  119. }
  120. n.token = token;
  121. n.cpu = smp_processor_id();
  122. n.halted = is_idle_task(current) || preempt_count() > 1;
  123. init_waitqueue_head(&n.wq);
  124. hlist_add_head(&n.link, &b->list);
  125. spin_unlock(&b->lock);
  126. for (;;) {
  127. if (!n.halted)
  128. prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
  129. if (hlist_unhashed(&n.link))
  130. break;
  131. if (!n.halted) {
  132. local_irq_enable();
  133. schedule();
  134. local_irq_disable();
  135. } else {
  136. /*
  137. * We cannot reschedule. So halt.
  138. */
  139. rcu_irq_exit();
  140. native_safe_halt();
  141. rcu_irq_enter();
  142. local_irq_disable();
  143. }
  144. }
  145. if (!n.halted)
  146. finish_wait(&n.wq, &wait);
  147. rcu_irq_exit();
  148. return;
  149. }
  150. EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
  151. static void apf_task_wake_one(struct kvm_task_sleep_node *n)
  152. {
  153. hlist_del_init(&n->link);
  154. if (n->halted)
  155. smp_send_reschedule(n->cpu);
  156. else if (waitqueue_active(&n->wq))
  157. wake_up(&n->wq);
  158. }
  159. static void apf_task_wake_all(void)
  160. {
  161. int i;
  162. for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
  163. struct hlist_node *p, *next;
  164. struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
  165. spin_lock(&b->lock);
  166. hlist_for_each_safe(p, next, &b->list) {
  167. struct kvm_task_sleep_node *n =
  168. hlist_entry(p, typeof(*n), link);
  169. if (n->cpu == smp_processor_id())
  170. apf_task_wake_one(n);
  171. }
  172. spin_unlock(&b->lock);
  173. }
  174. }
  175. void kvm_async_pf_task_wake(u32 token)
  176. {
  177. u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
  178. struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
  179. struct kvm_task_sleep_node *n;
  180. if (token == ~0) {
  181. apf_task_wake_all();
  182. return;
  183. }
  184. again:
  185. spin_lock(&b->lock);
  186. n = _find_apf_task(b, token);
  187. if (!n) {
  188. /*
  189. * async PF was not yet handled.
  190. * Add dummy entry for the token.
  191. */
  192. n = kzalloc(sizeof(*n), GFP_ATOMIC);
  193. if (!n) {
  194. /*
  195. * Allocation failed! Busy wait while other cpu
  196. * handles async PF.
  197. */
  198. spin_unlock(&b->lock);
  199. cpu_relax();
  200. goto again;
  201. }
  202. n->token = token;
  203. n->cpu = smp_processor_id();
  204. init_waitqueue_head(&n->wq);
  205. hlist_add_head(&n->link, &b->list);
  206. } else
  207. apf_task_wake_one(n);
  208. spin_unlock(&b->lock);
  209. return;
  210. }
  211. EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
  212. u32 kvm_read_and_reset_pf_reason(void)
  213. {
  214. u32 reason = 0;
  215. if (__this_cpu_read(apf_reason.enabled)) {
  216. reason = __this_cpu_read(apf_reason.reason);
  217. __this_cpu_write(apf_reason.reason, 0);
  218. }
  219. return reason;
  220. }
  221. EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
  222. NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
  223. dotraplinkage void
  224. do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
  225. {
  226. enum ctx_state prev_state;
  227. switch (kvm_read_and_reset_pf_reason()) {
  228. default:
  229. trace_do_page_fault(regs, error_code);
  230. break;
  231. case KVM_PV_REASON_PAGE_NOT_PRESENT:
  232. /* page is swapped out by the host. */
  233. prev_state = exception_enter();
  234. exit_idle();
  235. kvm_async_pf_task_wait((u32)read_cr2());
  236. exception_exit(prev_state);
  237. break;
  238. case KVM_PV_REASON_PAGE_READY:
  239. rcu_irq_enter();
  240. exit_idle();
  241. kvm_async_pf_task_wake((u32)read_cr2());
  242. rcu_irq_exit();
  243. break;
  244. }
  245. }
  246. NOKPROBE_SYMBOL(do_async_page_fault);
  247. static void __init paravirt_ops_setup(void)
  248. {
  249. pv_info.name = "KVM";
  250. pv_info.paravirt_enabled = 1;
  251. if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
  252. pv_cpu_ops.io_delay = kvm_io_delay;
  253. #ifdef CONFIG_X86_IO_APIC
  254. no_timer_check = 1;
  255. #endif
  256. }
  257. static void kvm_register_steal_time(void)
  258. {
  259. int cpu = smp_processor_id();
  260. struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
  261. if (!has_steal_clock)
  262. return;
  263. memset(st, 0, sizeof(*st));
  264. wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
  265. pr_info("kvm-stealtime: cpu %d, msr %llx\n",
  266. cpu, (unsigned long long) slow_virt_to_phys(st));
  267. }
  268. static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
  269. static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
  270. {
  271. /**
  272. * This relies on __test_and_clear_bit to modify the memory
  273. * in a way that is atomic with respect to the local CPU.
  274. * The hypervisor only accesses this memory from the local CPU so
  275. * there's no need for lock or memory barriers.
  276. * An optimization barrier is implied in apic write.
  277. */
  278. if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
  279. return;
  280. apic_write(APIC_EOI, APIC_EOI_ACK);
  281. }
  282. void kvm_guest_cpu_init(void)
  283. {
  284. if (!kvm_para_available())
  285. return;
  286. if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
  287. u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
  288. #ifdef CONFIG_PREEMPT
  289. pa |= KVM_ASYNC_PF_SEND_ALWAYS;
  290. #endif
  291. wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
  292. __this_cpu_write(apf_reason.enabled, 1);
  293. printk(KERN_INFO"KVM setup async PF for cpu %d\n",
  294. smp_processor_id());
  295. }
  296. if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
  297. unsigned long pa;
  298. /* Size alignment is implied but just to make it explicit. */
  299. BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
  300. __this_cpu_write(kvm_apic_eoi, 0);
  301. pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
  302. | KVM_MSR_ENABLED;
  303. wrmsrl(MSR_KVM_PV_EOI_EN, pa);
  304. }
  305. if (has_steal_clock)
  306. kvm_register_steal_time();
  307. }
  308. static void kvm_pv_disable_apf(void)
  309. {
  310. if (!__this_cpu_read(apf_reason.enabled))
  311. return;
  312. wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
  313. __this_cpu_write(apf_reason.enabled, 0);
  314. printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
  315. smp_processor_id());
  316. }
  317. static void kvm_pv_guest_cpu_reboot(void *unused)
  318. {
  319. /*
  320. * We disable PV EOI before we load a new kernel by kexec,
  321. * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
  322. * New kernel can re-enable when it boots.
  323. */
  324. if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
  325. wrmsrl(MSR_KVM_PV_EOI_EN, 0);
  326. kvm_pv_disable_apf();
  327. kvm_disable_steal_time();
  328. }
  329. static int kvm_pv_reboot_notify(struct notifier_block *nb,
  330. unsigned long code, void *unused)
  331. {
  332. if (code == SYS_RESTART)
  333. on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
  334. return NOTIFY_DONE;
  335. }
  336. static struct notifier_block kvm_pv_reboot_nb = {
  337. .notifier_call = kvm_pv_reboot_notify,
  338. };
  339. static u64 kvm_steal_clock(int cpu)
  340. {
  341. u64 steal;
  342. struct kvm_steal_time *src;
  343. int version;
  344. src = &per_cpu(steal_time, cpu);
  345. do {
  346. version = src->version;
  347. rmb();
  348. steal = src->steal;
  349. rmb();
  350. } while ((version & 1) || (version != src->version));
  351. return steal;
  352. }
  353. void kvm_disable_steal_time(void)
  354. {
  355. if (!has_steal_clock)
  356. return;
  357. wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
  358. }
  359. #ifdef CONFIG_SMP
  360. static void __init kvm_smp_prepare_boot_cpu(void)
  361. {
  362. kvm_guest_cpu_init();
  363. native_smp_prepare_boot_cpu();
  364. kvm_spinlock_init();
  365. }
  366. static void kvm_guest_cpu_online(void *dummy)
  367. {
  368. kvm_guest_cpu_init();
  369. }
  370. static void kvm_guest_cpu_offline(void *dummy)
  371. {
  372. kvm_disable_steal_time();
  373. if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
  374. wrmsrl(MSR_KVM_PV_EOI_EN, 0);
  375. kvm_pv_disable_apf();
  376. apf_task_wake_all();
  377. }
  378. static int kvm_cpu_notify(struct notifier_block *self, unsigned long action,
  379. void *hcpu)
  380. {
  381. int cpu = (unsigned long)hcpu;
  382. switch (action) {
  383. case CPU_ONLINE:
  384. case CPU_DOWN_FAILED:
  385. case CPU_ONLINE_FROZEN:
  386. smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
  387. break;
  388. case CPU_DOWN_PREPARE:
  389. case CPU_DOWN_PREPARE_FROZEN:
  390. smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
  391. break;
  392. default:
  393. break;
  394. }
  395. return NOTIFY_OK;
  396. }
  397. static struct notifier_block kvm_cpu_notifier = {
  398. .notifier_call = kvm_cpu_notify,
  399. };
  400. #endif
  401. static void __init kvm_apf_trap_init(void)
  402. {
  403. set_intr_gate(14, async_page_fault);
  404. }
  405. void __init kvm_guest_init(void)
  406. {
  407. int i;
  408. if (!kvm_para_available())
  409. return;
  410. paravirt_ops_setup();
  411. register_reboot_notifier(&kvm_pv_reboot_nb);
  412. for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
  413. spin_lock_init(&async_pf_sleepers[i].lock);
  414. if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
  415. x86_init.irqs.trap_init = kvm_apf_trap_init;
  416. if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
  417. has_steal_clock = 1;
  418. pv_time_ops.steal_clock = kvm_steal_clock;
  419. }
  420. if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
  421. apic_set_eoi_write(kvm_guest_apic_eoi_write);
  422. if (kvmclock_vsyscall)
  423. kvm_setup_vsyscall_timeinfo();
  424. #ifdef CONFIG_SMP
  425. smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
  426. register_cpu_notifier(&kvm_cpu_notifier);
  427. #else
  428. kvm_guest_cpu_init();
  429. #endif
  430. /*
  431. * Hard lockup detection is enabled by default. Disable it, as guests
  432. * can get false positives too easily, for example if the host is
  433. * overcommitted.
  434. */
  435. watchdog_enable_hardlockup_detector(false);
  436. }
  437. static noinline uint32_t __kvm_cpuid_base(void)
  438. {
  439. if (boot_cpu_data.cpuid_level < 0)
  440. return 0; /* So we don't blow up on old processors */
  441. if (cpu_has_hypervisor)
  442. return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
  443. return 0;
  444. }
  445. static inline uint32_t kvm_cpuid_base(void)
  446. {
  447. static int kvm_cpuid_base = -1;
  448. if (kvm_cpuid_base == -1)
  449. kvm_cpuid_base = __kvm_cpuid_base();
  450. return kvm_cpuid_base;
  451. }
  452. bool kvm_para_available(void)
  453. {
  454. return kvm_cpuid_base() != 0;
  455. }
  456. EXPORT_SYMBOL_GPL(kvm_para_available);
  457. unsigned int kvm_arch_para_features(void)
  458. {
  459. return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
  460. }
  461. static uint32_t __init kvm_detect(void)
  462. {
  463. return kvm_cpuid_base();
  464. }
  465. const struct hypervisor_x86 x86_hyper_kvm __refconst = {
  466. .name = "KVM",
  467. .detect = kvm_detect,
  468. .x2apic_available = kvm_para_available,
  469. };
  470. EXPORT_SYMBOL_GPL(x86_hyper_kvm);
  471. static __init int activate_jump_labels(void)
  472. {
  473. if (has_steal_clock) {
  474. static_key_slow_inc(&paravirt_steal_enabled);
  475. if (steal_acc)
  476. static_key_slow_inc(&paravirt_steal_rq_enabled);
  477. }
  478. return 0;
  479. }
  480. arch_initcall(activate_jump_labels);
  481. #ifdef CONFIG_PARAVIRT_SPINLOCKS
  482. /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
  483. static void kvm_kick_cpu(int cpu)
  484. {
  485. int apicid;
  486. unsigned long flags = 0;
  487. apicid = per_cpu(x86_cpu_to_apicid, cpu);
  488. kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
  489. }
  490. enum kvm_contention_stat {
  491. TAKEN_SLOW,
  492. TAKEN_SLOW_PICKUP,
  493. RELEASED_SLOW,
  494. RELEASED_SLOW_KICKED,
  495. NR_CONTENTION_STATS
  496. };
  497. #ifdef CONFIG_KVM_DEBUG_FS
  498. #define HISTO_BUCKETS 30
  499. static struct kvm_spinlock_stats
  500. {
  501. u32 contention_stats[NR_CONTENTION_STATS];
  502. u32 histo_spin_blocked[HISTO_BUCKETS+1];
  503. u64 time_blocked;
  504. } spinlock_stats;
  505. static u8 zero_stats;
  506. static inline void check_zero(void)
  507. {
  508. u8 ret;
  509. u8 old;
  510. old = ACCESS_ONCE(zero_stats);
  511. if (unlikely(old)) {
  512. ret = cmpxchg(&zero_stats, old, 0);
  513. /* This ensures only one fellow resets the stat */
  514. if (ret == old)
  515. memset(&spinlock_stats, 0, sizeof(spinlock_stats));
  516. }
  517. }
  518. static inline void add_stats(enum kvm_contention_stat var, u32 val)
  519. {
  520. check_zero();
  521. spinlock_stats.contention_stats[var] += val;
  522. }
  523. static inline u64 spin_time_start(void)
  524. {
  525. return sched_clock();
  526. }
  527. static void __spin_time_accum(u64 delta, u32 *array)
  528. {
  529. unsigned index;
  530. index = ilog2(delta);
  531. check_zero();
  532. if (index < HISTO_BUCKETS)
  533. array[index]++;
  534. else
  535. array[HISTO_BUCKETS]++;
  536. }
  537. static inline void spin_time_accum_blocked(u64 start)
  538. {
  539. u32 delta;
  540. delta = sched_clock() - start;
  541. __spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
  542. spinlock_stats.time_blocked += delta;
  543. }
  544. static struct dentry *d_spin_debug;
  545. static struct dentry *d_kvm_debug;
  546. struct dentry *kvm_init_debugfs(void)
  547. {
  548. d_kvm_debug = debugfs_create_dir("kvm-guest", NULL);
  549. if (!d_kvm_debug)
  550. printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n");
  551. return d_kvm_debug;
  552. }
  553. static int __init kvm_spinlock_debugfs(void)
  554. {
  555. struct dentry *d_kvm;
  556. d_kvm = kvm_init_debugfs();
  557. if (d_kvm == NULL)
  558. return -ENOMEM;
  559. d_spin_debug = debugfs_create_dir("spinlocks", d_kvm);
  560. debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
  561. debugfs_create_u32("taken_slow", 0444, d_spin_debug,
  562. &spinlock_stats.contention_stats[TAKEN_SLOW]);
  563. debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
  564. &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
  565. debugfs_create_u32("released_slow", 0444, d_spin_debug,
  566. &spinlock_stats.contention_stats[RELEASED_SLOW]);
  567. debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
  568. &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
  569. debugfs_create_u64("time_blocked", 0444, d_spin_debug,
  570. &spinlock_stats.time_blocked);
  571. debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
  572. spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
  573. return 0;
  574. }
  575. fs_initcall(kvm_spinlock_debugfs);
  576. #else /* !CONFIG_KVM_DEBUG_FS */
  577. static inline void add_stats(enum kvm_contention_stat var, u32 val)
  578. {
  579. }
  580. static inline u64 spin_time_start(void)
  581. {
  582. return 0;
  583. }
  584. static inline void spin_time_accum_blocked(u64 start)
  585. {
  586. }
  587. #endif /* CONFIG_KVM_DEBUG_FS */
  588. struct kvm_lock_waiting {
  589. struct arch_spinlock *lock;
  590. __ticket_t want;
  591. };
  592. /* cpus 'waiting' on a spinlock to become available */
  593. static cpumask_t waiting_cpus;
  594. /* Track spinlock on which a cpu is waiting */
  595. static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting);
  596. __visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
  597. {
  598. struct kvm_lock_waiting *w;
  599. int cpu;
  600. u64 start;
  601. unsigned long flags;
  602. if (in_nmi())
  603. return;
  604. w = this_cpu_ptr(&klock_waiting);
  605. cpu = smp_processor_id();
  606. start = spin_time_start();
  607. /*
  608. * Make sure an interrupt handler can't upset things in a
  609. * partially setup state.
  610. */
  611. local_irq_save(flags);
  612. /*
  613. * The ordering protocol on this is that the "lock" pointer
  614. * may only be set non-NULL if the "want" ticket is correct.
  615. * If we're updating "want", we must first clear "lock".
  616. */
  617. w->lock = NULL;
  618. smp_wmb();
  619. w->want = want;
  620. smp_wmb();
  621. w->lock = lock;
  622. add_stats(TAKEN_SLOW, 1);
  623. /*
  624. * This uses set_bit, which is atomic but we should not rely on its
  625. * reordering gurantees. So barrier is needed after this call.
  626. */
  627. cpumask_set_cpu(cpu, &waiting_cpus);
  628. barrier();
  629. /*
  630. * Mark entry to slowpath before doing the pickup test to make
  631. * sure we don't deadlock with an unlocker.
  632. */
  633. __ticket_enter_slowpath(lock);
  634. /*
  635. * check again make sure it didn't become free while
  636. * we weren't looking.
  637. */
  638. if (ACCESS_ONCE(lock->tickets.head) == want) {
  639. add_stats(TAKEN_SLOW_PICKUP, 1);
  640. goto out;
  641. }
  642. /*
  643. * halt until it's our turn and kicked. Note that we do safe halt
  644. * for irq enabled case to avoid hang when lock info is overwritten
  645. * in irq spinlock slowpath and no spurious interrupt occur to save us.
  646. */
  647. if (arch_irqs_disabled_flags(flags))
  648. halt();
  649. else
  650. safe_halt();
  651. out:
  652. cpumask_clear_cpu(cpu, &waiting_cpus);
  653. w->lock = NULL;
  654. local_irq_restore(flags);
  655. spin_time_accum_blocked(start);
  656. }
  657. PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning);
  658. /* Kick vcpu waiting on @lock->head to reach value @ticket */
  659. static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket)
  660. {
  661. int cpu;
  662. add_stats(RELEASED_SLOW, 1);
  663. for_each_cpu(cpu, &waiting_cpus) {
  664. const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu);
  665. if (ACCESS_ONCE(w->lock) == lock &&
  666. ACCESS_ONCE(w->want) == ticket) {
  667. add_stats(RELEASED_SLOW_KICKED, 1);
  668. kvm_kick_cpu(cpu);
  669. break;
  670. }
  671. }
  672. }
  673. /*
  674. * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
  675. */
  676. void __init kvm_spinlock_init(void)
  677. {
  678. if (!kvm_para_available())
  679. return;
  680. /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
  681. if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
  682. return;
  683. pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
  684. pv_lock_ops.unlock_kick = kvm_unlock_kick;
  685. }
  686. static __init int kvm_spinlock_init_jump(void)
  687. {
  688. if (!kvm_para_available())
  689. return 0;
  690. if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
  691. return 0;
  692. static_key_slow_inc(&paravirt_ticketlocks_enabled);
  693. printk(KERN_INFO "KVM setup paravirtual spinlock\n");
  694. return 0;
  695. }
  696. early_initcall(kvm_spinlock_init_jump);
  697. #endif /* CONFIG_PARAVIRT_SPINLOCKS */