smp.c 19 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776
  1. /*
  2. * Xen SMP support
  3. *
  4. * This file implements the Xen versions of smp_ops. SMP under Xen is
  5. * very straightforward. Bringing a CPU up is simply a matter of
  6. * loading its initial context and setting it running.
  7. *
  8. * IPIs are handled through the Xen event mechanism.
  9. *
  10. * Because virtual CPUs can be scheduled onto any real CPU, there's no
  11. * useful topology information for the kernel to make use of. As a
  12. * result, all CPUs are treated as if they're single-core and
  13. * single-threaded.
  14. */
  15. #include <linux/sched.h>
  16. #include <linux/err.h>
  17. #include <linux/slab.h>
  18. #include <linux/smp.h>
  19. #include <linux/irq_work.h>
  20. #include <linux/tick.h>
  21. #include <asm/paravirt.h>
  22. #include <asm/desc.h>
  23. #include <asm/pgtable.h>
  24. #include <asm/cpu.h>
  25. #include <xen/interface/xen.h>
  26. #include <xen/interface/vcpu.h>
  27. #include <xen/interface/xenpmu.h>
  28. #include <asm/xen/interface.h>
  29. #include <asm/xen/hypercall.h>
  30. #include <xen/xen.h>
  31. #include <xen/page.h>
  32. #include <xen/events.h>
  33. #include <xen/hvc-console.h>
  34. #include "xen-ops.h"
  35. #include "mmu.h"
  36. #include "smp.h"
  37. #include "pmu.h"
  38. cpumask_var_t xen_cpu_initialized_map;
  39. struct xen_common_irq {
  40. int irq;
  41. char *name;
  42. };
  43. static DEFINE_PER_CPU(struct xen_common_irq, xen_resched_irq) = { .irq = -1 };
  44. static DEFINE_PER_CPU(struct xen_common_irq, xen_callfunc_irq) = { .irq = -1 };
  45. static DEFINE_PER_CPU(struct xen_common_irq, xen_callfuncsingle_irq) = { .irq = -1 };
  46. static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
  47. static DEFINE_PER_CPU(struct xen_common_irq, xen_debug_irq) = { .irq = -1 };
  48. static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
  49. static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
  50. static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id);
  51. static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
  52. /*
  53. * Reschedule call back.
  54. */
  55. static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
  56. {
  57. inc_irq_stat(irq_resched_count);
  58. scheduler_ipi();
  59. return IRQ_HANDLED;
  60. }
  61. static void cpu_bringup(void)
  62. {
  63. int cpu;
  64. cpu_init();
  65. touch_softlockup_watchdog();
  66. preempt_disable();
  67. /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
  68. if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
  69. xen_enable_sysenter();
  70. xen_enable_syscall();
  71. }
  72. cpu = smp_processor_id();
  73. smp_store_cpu_info(cpu);
  74. cpu_data(cpu).x86_max_cores = 1;
  75. set_cpu_sibling_map(cpu);
  76. xen_setup_cpu_clockevents();
  77. notify_cpu_starting(cpu);
  78. set_cpu_online(cpu, true);
  79. cpu_set_state_online(cpu); /* Implies full memory barrier. */
  80. /* We can take interrupts now: we're officially "up". */
  81. local_irq_enable();
  82. }
  83. /*
  84. * Note: cpu parameter is only relevant for PVH. The reason for passing it
  85. * is we can't do smp_processor_id until the percpu segments are loaded, for
  86. * which we need the cpu number! So we pass it in rdi as first parameter.
  87. */
  88. asmlinkage __visible void cpu_bringup_and_idle(int cpu)
  89. {
  90. #ifdef CONFIG_XEN_PVH
  91. if (xen_feature(XENFEAT_auto_translated_physmap) &&
  92. xen_feature(XENFEAT_supervisor_mode_kernel))
  93. xen_pvh_secondary_vcpu_init(cpu);
  94. #endif
  95. cpu_bringup();
  96. cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
  97. }
  98. void xen_smp_intr_free(unsigned int cpu)
  99. {
  100. if (per_cpu(xen_resched_irq, cpu).irq >= 0) {
  101. unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu).irq, NULL);
  102. per_cpu(xen_resched_irq, cpu).irq = -1;
  103. kfree(per_cpu(xen_resched_irq, cpu).name);
  104. per_cpu(xen_resched_irq, cpu).name = NULL;
  105. }
  106. if (per_cpu(xen_callfunc_irq, cpu).irq >= 0) {
  107. unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu).irq, NULL);
  108. per_cpu(xen_callfunc_irq, cpu).irq = -1;
  109. kfree(per_cpu(xen_callfunc_irq, cpu).name);
  110. per_cpu(xen_callfunc_irq, cpu).name = NULL;
  111. }
  112. if (per_cpu(xen_debug_irq, cpu).irq >= 0) {
  113. unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu).irq, NULL);
  114. per_cpu(xen_debug_irq, cpu).irq = -1;
  115. kfree(per_cpu(xen_debug_irq, cpu).name);
  116. per_cpu(xen_debug_irq, cpu).name = NULL;
  117. }
  118. if (per_cpu(xen_callfuncsingle_irq, cpu).irq >= 0) {
  119. unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu).irq,
  120. NULL);
  121. per_cpu(xen_callfuncsingle_irq, cpu).irq = -1;
  122. kfree(per_cpu(xen_callfuncsingle_irq, cpu).name);
  123. per_cpu(xen_callfuncsingle_irq, cpu).name = NULL;
  124. }
  125. if (xen_hvm_domain())
  126. return;
  127. if (per_cpu(xen_irq_work, cpu).irq >= 0) {
  128. unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
  129. per_cpu(xen_irq_work, cpu).irq = -1;
  130. kfree(per_cpu(xen_irq_work, cpu).name);
  131. per_cpu(xen_irq_work, cpu).name = NULL;
  132. }
  133. if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
  134. unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
  135. per_cpu(xen_pmu_irq, cpu).irq = -1;
  136. kfree(per_cpu(xen_pmu_irq, cpu).name);
  137. per_cpu(xen_pmu_irq, cpu).name = NULL;
  138. }
  139. };
  140. int xen_smp_intr_init(unsigned int cpu)
  141. {
  142. int rc;
  143. char *resched_name, *callfunc_name, *debug_name, *pmu_name;
  144. resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
  145. rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
  146. cpu,
  147. xen_reschedule_interrupt,
  148. IRQF_PERCPU|IRQF_NOBALANCING,
  149. resched_name,
  150. NULL);
  151. if (rc < 0)
  152. goto fail;
  153. per_cpu(xen_resched_irq, cpu).irq = rc;
  154. per_cpu(xen_resched_irq, cpu).name = resched_name;
  155. callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
  156. rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
  157. cpu,
  158. xen_call_function_interrupt,
  159. IRQF_PERCPU|IRQF_NOBALANCING,
  160. callfunc_name,
  161. NULL);
  162. if (rc < 0)
  163. goto fail;
  164. per_cpu(xen_callfunc_irq, cpu).irq = rc;
  165. per_cpu(xen_callfunc_irq, cpu).name = callfunc_name;
  166. debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
  167. rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
  168. IRQF_PERCPU | IRQF_NOBALANCING,
  169. debug_name, NULL);
  170. if (rc < 0)
  171. goto fail;
  172. per_cpu(xen_debug_irq, cpu).irq = rc;
  173. per_cpu(xen_debug_irq, cpu).name = debug_name;
  174. callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
  175. rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
  176. cpu,
  177. xen_call_function_single_interrupt,
  178. IRQF_PERCPU|IRQF_NOBALANCING,
  179. callfunc_name,
  180. NULL);
  181. if (rc < 0)
  182. goto fail;
  183. per_cpu(xen_callfuncsingle_irq, cpu).irq = rc;
  184. per_cpu(xen_callfuncsingle_irq, cpu).name = callfunc_name;
  185. /*
  186. * The IRQ worker on PVHVM goes through the native path and uses the
  187. * IPI mechanism.
  188. */
  189. if (xen_hvm_domain())
  190. return 0;
  191. callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
  192. rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
  193. cpu,
  194. xen_irq_work_interrupt,
  195. IRQF_PERCPU|IRQF_NOBALANCING,
  196. callfunc_name,
  197. NULL);
  198. if (rc < 0)
  199. goto fail;
  200. per_cpu(xen_irq_work, cpu).irq = rc;
  201. per_cpu(xen_irq_work, cpu).name = callfunc_name;
  202. if (is_xen_pmu(cpu)) {
  203. pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
  204. rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
  205. xen_pmu_irq_handler,
  206. IRQF_PERCPU|IRQF_NOBALANCING,
  207. pmu_name, NULL);
  208. if (rc < 0)
  209. goto fail;
  210. per_cpu(xen_pmu_irq, cpu).irq = rc;
  211. per_cpu(xen_pmu_irq, cpu).name = pmu_name;
  212. }
  213. return 0;
  214. fail:
  215. xen_smp_intr_free(cpu);
  216. return rc;
  217. }
  218. static void __init xen_fill_possible_map(void)
  219. {
  220. int i, rc;
  221. if (xen_initial_domain())
  222. return;
  223. for (i = 0; i < nr_cpu_ids; i++) {
  224. rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
  225. if (rc >= 0) {
  226. num_processors++;
  227. set_cpu_possible(i, true);
  228. }
  229. }
  230. }
  231. static void __init xen_filter_cpu_maps(void)
  232. {
  233. int i, rc;
  234. unsigned int subtract = 0;
  235. if (!xen_initial_domain())
  236. return;
  237. num_processors = 0;
  238. disabled_cpus = 0;
  239. for (i = 0; i < nr_cpu_ids; i++) {
  240. rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
  241. if (rc >= 0) {
  242. num_processors++;
  243. set_cpu_possible(i, true);
  244. } else {
  245. set_cpu_possible(i, false);
  246. set_cpu_present(i, false);
  247. subtract++;
  248. }
  249. }
  250. #ifdef CONFIG_HOTPLUG_CPU
  251. /* This is akin to using 'nr_cpus' on the Linux command line.
  252. * Which is OK as when we use 'dom0_max_vcpus=X' we can only
  253. * have up to X, while nr_cpu_ids is greater than X. This
  254. * normally is not a problem, except when CPU hotplugging
  255. * is involved and then there might be more than X CPUs
  256. * in the guest - which will not work as there is no
  257. * hypercall to expand the max number of VCPUs an already
  258. * running guest has. So cap it up to X. */
  259. if (subtract)
  260. nr_cpu_ids = nr_cpu_ids - subtract;
  261. #endif
  262. }
  263. static void __init xen_smp_prepare_boot_cpu(void)
  264. {
  265. BUG_ON(smp_processor_id() != 0);
  266. native_smp_prepare_boot_cpu();
  267. if (xen_pv_domain()) {
  268. if (!xen_feature(XENFEAT_writable_page_tables))
  269. /* We've switched to the "real" per-cpu gdt, so make
  270. * sure the old memory can be recycled. */
  271. make_lowmem_page_readwrite(xen_initial_gdt);
  272. #ifdef CONFIG_X86_32
  273. /*
  274. * Xen starts us with XEN_FLAT_RING1_DS, but linux code
  275. * expects __USER_DS
  276. */
  277. loadsegment(ds, __USER_DS);
  278. loadsegment(es, __USER_DS);
  279. #endif
  280. xen_filter_cpu_maps();
  281. xen_setup_vcpu_info_placement();
  282. }
  283. /*
  284. * Setup vcpu_info for boot CPU.
  285. */
  286. if (xen_hvm_domain())
  287. xen_vcpu_setup(0);
  288. /*
  289. * The alternative logic (which patches the unlock/lock) runs before
  290. * the smp bootup up code is activated. Hence we need to set this up
  291. * the core kernel is being patched. Otherwise we will have only
  292. * modules patched but not core code.
  293. */
  294. xen_init_spinlocks();
  295. }
  296. static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
  297. {
  298. unsigned cpu;
  299. unsigned int i;
  300. if (skip_ioapic_setup) {
  301. char *m = (max_cpus == 0) ?
  302. "The nosmp parameter is incompatible with Xen; " \
  303. "use Xen dom0_max_vcpus=1 parameter" :
  304. "The noapic parameter is incompatible with Xen";
  305. xen_raw_printk(m);
  306. panic(m);
  307. }
  308. xen_init_lock_cpu(0);
  309. smp_store_boot_cpu_info();
  310. cpu_data(0).x86_max_cores = 1;
  311. for_each_possible_cpu(i) {
  312. zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
  313. zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
  314. zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
  315. }
  316. set_cpu_sibling_map(0);
  317. xen_pmu_init(0);
  318. if (xen_smp_intr_init(0))
  319. BUG();
  320. if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
  321. panic("could not allocate xen_cpu_initialized_map\n");
  322. cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
  323. /* Restrict the possible_map according to max_cpus. */
  324. while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
  325. for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
  326. continue;
  327. set_cpu_possible(cpu, false);
  328. }
  329. for_each_possible_cpu(cpu)
  330. set_cpu_present(cpu, true);
  331. }
  332. static int
  333. cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
  334. {
  335. struct vcpu_guest_context *ctxt;
  336. struct desc_struct *gdt;
  337. unsigned long gdt_mfn;
  338. /* used to tell cpu_init() that it can proceed with initialization */
  339. cpumask_set_cpu(cpu, cpu_callout_mask);
  340. if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
  341. return 0;
  342. ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
  343. if (ctxt == NULL)
  344. return -ENOMEM;
  345. gdt = get_cpu_gdt_table(cpu);
  346. #ifdef CONFIG_X86_32
  347. /* Note: PVH is not yet supported on x86_32. */
  348. ctxt->user_regs.fs = __KERNEL_PERCPU;
  349. ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
  350. #endif
  351. memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
  352. if (!xen_feature(XENFEAT_auto_translated_physmap)) {
  353. ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
  354. ctxt->flags = VGCF_IN_KERNEL;
  355. ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
  356. ctxt->user_regs.ds = __USER_DS;
  357. ctxt->user_regs.es = __USER_DS;
  358. ctxt->user_regs.ss = __KERNEL_DS;
  359. xen_copy_trap_info(ctxt->trap_ctxt);
  360. ctxt->ldt_ents = 0;
  361. BUG_ON((unsigned long)gdt & ~PAGE_MASK);
  362. gdt_mfn = arbitrary_virt_to_mfn(gdt);
  363. make_lowmem_page_readonly(gdt);
  364. make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
  365. ctxt->gdt_frames[0] = gdt_mfn;
  366. ctxt->gdt_ents = GDT_ENTRIES;
  367. ctxt->kernel_ss = __KERNEL_DS;
  368. ctxt->kernel_sp = idle->thread.sp0;
  369. #ifdef CONFIG_X86_32
  370. ctxt->event_callback_cs = __KERNEL_CS;
  371. ctxt->failsafe_callback_cs = __KERNEL_CS;
  372. #else
  373. ctxt->gs_base_kernel = per_cpu_offset(cpu);
  374. #endif
  375. ctxt->event_callback_eip =
  376. (unsigned long)xen_hypervisor_callback;
  377. ctxt->failsafe_callback_eip =
  378. (unsigned long)xen_failsafe_callback;
  379. ctxt->user_regs.cs = __KERNEL_CS;
  380. per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
  381. }
  382. #ifdef CONFIG_XEN_PVH
  383. else {
  384. /*
  385. * The vcpu comes on kernel page tables which have the NX pte
  386. * bit set. This means before DS/SS is touched, NX in
  387. * EFER must be set. Hence the following assembly glue code.
  388. */
  389. ctxt->user_regs.eip = (unsigned long)xen_pvh_early_cpu_init;
  390. ctxt->user_regs.rdi = cpu;
  391. ctxt->user_regs.rsi = true; /* entry == true */
  392. }
  393. #endif
  394. ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
  395. ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
  396. if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
  397. BUG();
  398. kfree(ctxt);
  399. return 0;
  400. }
  401. static int xen_cpu_up(unsigned int cpu, struct task_struct *idle)
  402. {
  403. int rc;
  404. common_cpu_up(cpu, idle);
  405. xen_setup_runstate_info(cpu);
  406. /*
  407. * PV VCPUs are always successfully taken down (see 'while' loop
  408. * in xen_cpu_die()), so -EBUSY is an error.
  409. */
  410. rc = cpu_check_up_prepare(cpu);
  411. if (rc)
  412. return rc;
  413. /* make sure interrupts start blocked */
  414. per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
  415. rc = cpu_initialize_context(cpu, idle);
  416. if (rc)
  417. return rc;
  418. xen_pmu_init(cpu);
  419. rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
  420. BUG_ON(rc);
  421. while (cpu_report_state(cpu) != CPU_ONLINE)
  422. HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
  423. return 0;
  424. }
  425. static void xen_smp_cpus_done(unsigned int max_cpus)
  426. {
  427. }
  428. #ifdef CONFIG_HOTPLUG_CPU
  429. static int xen_cpu_disable(void)
  430. {
  431. unsigned int cpu = smp_processor_id();
  432. if (cpu == 0)
  433. return -EBUSY;
  434. cpu_disable_common();
  435. load_cr3(swapper_pg_dir);
  436. return 0;
  437. }
  438. static void xen_cpu_die(unsigned int cpu)
  439. {
  440. while (xen_pv_domain() && HYPERVISOR_vcpu_op(VCPUOP_is_up,
  441. xen_vcpu_nr(cpu), NULL)) {
  442. __set_current_state(TASK_UNINTERRUPTIBLE);
  443. schedule_timeout(HZ/10);
  444. }
  445. if (common_cpu_die(cpu) == 0) {
  446. xen_smp_intr_free(cpu);
  447. xen_uninit_lock_cpu(cpu);
  448. xen_teardown_timer(cpu);
  449. xen_pmu_finish(cpu);
  450. }
  451. }
  452. static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
  453. {
  454. play_dead_common();
  455. HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
  456. cpu_bringup();
  457. /*
  458. * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
  459. * clears certain data that the cpu_idle loop (which called us
  460. * and that we return from) expects. The only way to get that
  461. * data back is to call:
  462. */
  463. tick_nohz_idle_enter();
  464. cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
  465. }
  466. #else /* !CONFIG_HOTPLUG_CPU */
  467. static int xen_cpu_disable(void)
  468. {
  469. return -ENOSYS;
  470. }
  471. static void xen_cpu_die(unsigned int cpu)
  472. {
  473. BUG();
  474. }
  475. static void xen_play_dead(void)
  476. {
  477. BUG();
  478. }
  479. #endif
  480. static void stop_self(void *v)
  481. {
  482. int cpu = smp_processor_id();
  483. /* make sure we're not pinning something down */
  484. load_cr3(swapper_pg_dir);
  485. /* should set up a minimal gdt */
  486. set_cpu_online(cpu, false);
  487. HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
  488. BUG();
  489. }
  490. static void xen_stop_other_cpus(int wait)
  491. {
  492. smp_call_function(stop_self, NULL, wait);
  493. }
  494. static void xen_smp_send_reschedule(int cpu)
  495. {
  496. xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
  497. }
  498. static void __xen_send_IPI_mask(const struct cpumask *mask,
  499. int vector)
  500. {
  501. unsigned cpu;
  502. for_each_cpu_and(cpu, mask, cpu_online_mask)
  503. xen_send_IPI_one(cpu, vector);
  504. }
  505. static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
  506. {
  507. int cpu;
  508. __xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
  509. /* Make sure other vcpus get a chance to run if they need to. */
  510. for_each_cpu(cpu, mask) {
  511. if (xen_vcpu_stolen(cpu)) {
  512. HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
  513. break;
  514. }
  515. }
  516. }
  517. static void xen_smp_send_call_function_single_ipi(int cpu)
  518. {
  519. __xen_send_IPI_mask(cpumask_of(cpu),
  520. XEN_CALL_FUNCTION_SINGLE_VECTOR);
  521. }
  522. static inline int xen_map_vector(int vector)
  523. {
  524. int xen_vector;
  525. switch (vector) {
  526. case RESCHEDULE_VECTOR:
  527. xen_vector = XEN_RESCHEDULE_VECTOR;
  528. break;
  529. case CALL_FUNCTION_VECTOR:
  530. xen_vector = XEN_CALL_FUNCTION_VECTOR;
  531. break;
  532. case CALL_FUNCTION_SINGLE_VECTOR:
  533. xen_vector = XEN_CALL_FUNCTION_SINGLE_VECTOR;
  534. break;
  535. case IRQ_WORK_VECTOR:
  536. xen_vector = XEN_IRQ_WORK_VECTOR;
  537. break;
  538. #ifdef CONFIG_X86_64
  539. case NMI_VECTOR:
  540. case APIC_DM_NMI: /* Some use that instead of NMI_VECTOR */
  541. xen_vector = XEN_NMI_VECTOR;
  542. break;
  543. #endif
  544. default:
  545. xen_vector = -1;
  546. printk(KERN_ERR "xen: vector 0x%x is not implemented\n",
  547. vector);
  548. }
  549. return xen_vector;
  550. }
  551. void xen_send_IPI_mask(const struct cpumask *mask,
  552. int vector)
  553. {
  554. int xen_vector = xen_map_vector(vector);
  555. if (xen_vector >= 0)
  556. __xen_send_IPI_mask(mask, xen_vector);
  557. }
  558. void xen_send_IPI_all(int vector)
  559. {
  560. int xen_vector = xen_map_vector(vector);
  561. if (xen_vector >= 0)
  562. __xen_send_IPI_mask(cpu_online_mask, xen_vector);
  563. }
  564. void xen_send_IPI_self(int vector)
  565. {
  566. int xen_vector = xen_map_vector(vector);
  567. if (xen_vector >= 0)
  568. xen_send_IPI_one(smp_processor_id(), xen_vector);
  569. }
  570. void xen_send_IPI_mask_allbutself(const struct cpumask *mask,
  571. int vector)
  572. {
  573. unsigned cpu;
  574. unsigned int this_cpu = smp_processor_id();
  575. int xen_vector = xen_map_vector(vector);
  576. if (!(num_online_cpus() > 1) || (xen_vector < 0))
  577. return;
  578. for_each_cpu_and(cpu, mask, cpu_online_mask) {
  579. if (this_cpu == cpu)
  580. continue;
  581. xen_send_IPI_one(cpu, xen_vector);
  582. }
  583. }
  584. void xen_send_IPI_allbutself(int vector)
  585. {
  586. xen_send_IPI_mask_allbutself(cpu_online_mask, vector);
  587. }
  588. static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
  589. {
  590. irq_enter();
  591. generic_smp_call_function_interrupt();
  592. inc_irq_stat(irq_call_count);
  593. irq_exit();
  594. return IRQ_HANDLED;
  595. }
  596. static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
  597. {
  598. irq_enter();
  599. generic_smp_call_function_single_interrupt();
  600. inc_irq_stat(irq_call_count);
  601. irq_exit();
  602. return IRQ_HANDLED;
  603. }
  604. static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
  605. {
  606. irq_enter();
  607. irq_work_run();
  608. inc_irq_stat(apic_irq_work_irqs);
  609. irq_exit();
  610. return IRQ_HANDLED;
  611. }
  612. static const struct smp_ops xen_smp_ops __initconst = {
  613. .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
  614. .smp_prepare_cpus = xen_smp_prepare_cpus,
  615. .smp_cpus_done = xen_smp_cpus_done,
  616. .cpu_up = xen_cpu_up,
  617. .cpu_die = xen_cpu_die,
  618. .cpu_disable = xen_cpu_disable,
  619. .play_dead = xen_play_dead,
  620. .stop_other_cpus = xen_stop_other_cpus,
  621. .smp_send_reschedule = xen_smp_send_reschedule,
  622. .send_call_func_ipi = xen_smp_send_call_function_ipi,
  623. .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
  624. };
  625. void __init xen_smp_init(void)
  626. {
  627. smp_ops = xen_smp_ops;
  628. xen_fill_possible_map();
  629. }
  630. static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
  631. {
  632. native_smp_prepare_cpus(max_cpus);
  633. WARN_ON(xen_smp_intr_init(0));
  634. xen_init_lock_cpu(0);
  635. }
  636. void __init xen_hvm_smp_init(void)
  637. {
  638. if (!xen_have_vector_callback)
  639. return;
  640. smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus;
  641. smp_ops.smp_send_reschedule = xen_smp_send_reschedule;
  642. smp_ops.cpu_die = xen_cpu_die;
  643. smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi;
  644. smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi;
  645. smp_ops.smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu;
  646. }