1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Xen SMP support
4 *
5 * This file implements the Xen versions of smp_ops. SMP under Xen is
6 * very straightforward. Bringing a CPU up is simply a matter of
7 * loading its initial context and setting it running.
8 *
9 * IPIs are handled through the Xen event mechanism.
10 *
11 * Because virtual CPUs can be scheduled onto any real CPU, there's no
12 * useful topology information for the kernel to make use of. As a
13 * result, all CPUs are treated as if they're single-core and
14 * single-threaded.
15 */
16 #include <linux/sched.h>
17 #include <linux/sched/task_stack.h>
18 #include <linux/err.h>
19 #include <linux/slab.h>
20 #include <linux/smp.h>
21 #include <linux/irq_work.h>
22 #include <linux/tick.h>
23 #include <linux/nmi.h>
24 #include <linux/cpuhotplug.h>
25
26 #include <asm/paravirt.h>
27 #include <asm/desc.h>
28 #include <asm/pgtable.h>
29 #include <asm/cpu.h>
30
31 #include <xen/interface/xen.h>
32 #include <xen/interface/vcpu.h>
33 #include <xen/interface/xenpmu.h>
34
35 #include <asm/spec-ctrl.h>
36 #include <asm/xen/interface.h>
37 #include <asm/xen/hypercall.h>
38
39 #include <xen/xen.h>
40 #include <xen/page.h>
41 #include <xen/events.h>
42
43 #include <xen/hvc-console.h>
44 #include "xen-ops.h"
45 #include "mmu.h"
46 #include "smp.h"
47 #include "pmu.h"
48
49 cpumask_var_t xen_cpu_initialized_map;
50
51 static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
52 static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
53
54 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
55
cpu_bringup(void)56 static void cpu_bringup(void)
57 {
58 int cpu;
59
60 cpu_init();
61 touch_softlockup_watchdog();
62 preempt_disable();
63
64 /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
65 if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
66 xen_enable_sysenter();
67 xen_enable_syscall();
68 }
69 cpu = smp_processor_id();
70 smp_store_cpu_info(cpu);
71 cpu_data(cpu).x86_max_cores = 1;
72 set_cpu_sibling_map(cpu);
73
74 speculative_store_bypass_ht_init();
75
76 xen_setup_cpu_clockevents();
77
78 notify_cpu_starting(cpu);
79
80 set_cpu_online(cpu, true);
81
82 cpu_set_state_online(cpu); /* Implies full memory barrier. */
83
84 /* We can take interrupts now: we're officially "up". */
85 local_irq_enable();
86 }
87
cpu_bringup_and_idle(void)88 asmlinkage __visible void cpu_bringup_and_idle(void)
89 {
90 cpu_bringup();
91 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
92 }
93
xen_smp_intr_free_pv(unsigned int cpu)94 void xen_smp_intr_free_pv(unsigned int cpu)
95 {
96 if (per_cpu(xen_irq_work, cpu).irq >= 0) {
97 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
98 per_cpu(xen_irq_work, cpu).irq = -1;
99 kfree(per_cpu(xen_irq_work, cpu).name);
100 per_cpu(xen_irq_work, cpu).name = NULL;
101 }
102
103 if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
104 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
105 per_cpu(xen_pmu_irq, cpu).irq = -1;
106 kfree(per_cpu(xen_pmu_irq, cpu).name);
107 per_cpu(xen_pmu_irq, cpu).name = NULL;
108 }
109 }
110
xen_smp_intr_init_pv(unsigned int cpu)111 int xen_smp_intr_init_pv(unsigned int cpu)
112 {
113 int rc;
114 char *callfunc_name, *pmu_name;
115
116 callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
117 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
118 cpu,
119 xen_irq_work_interrupt,
120 IRQF_PERCPU|IRQF_NOBALANCING,
121 callfunc_name,
122 NULL);
123 if (rc < 0)
124 goto fail;
125 per_cpu(xen_irq_work, cpu).irq = rc;
126 per_cpu(xen_irq_work, cpu).name = callfunc_name;
127
128 if (is_xen_pmu(cpu)) {
129 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
130 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
131 xen_pmu_irq_handler,
132 IRQF_PERCPU|IRQF_NOBALANCING,
133 pmu_name, NULL);
134 if (rc < 0)
135 goto fail;
136 per_cpu(xen_pmu_irq, cpu).irq = rc;
137 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
138 }
139
140 return 0;
141
142 fail:
143 xen_smp_intr_free_pv(cpu);
144 return rc;
145 }
146
xen_fill_possible_map(void)147 static void __init xen_fill_possible_map(void)
148 {
149 int i, rc;
150
151 if (xen_initial_domain())
152 return;
153
154 for (i = 0; i < nr_cpu_ids; i++) {
155 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
156 if (rc >= 0) {
157 num_processors++;
158 set_cpu_possible(i, true);
159 }
160 }
161 }
162
xen_filter_cpu_maps(void)163 static void __init xen_filter_cpu_maps(void)
164 {
165 int i, rc;
166 unsigned int subtract = 0;
167
168 if (!xen_initial_domain())
169 return;
170
171 num_processors = 0;
172 disabled_cpus = 0;
173 for (i = 0; i < nr_cpu_ids; i++) {
174 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
175 if (rc >= 0) {
176 num_processors++;
177 set_cpu_possible(i, true);
178 } else {
179 set_cpu_possible(i, false);
180 set_cpu_present(i, false);
181 subtract++;
182 }
183 }
184 #ifdef CONFIG_HOTPLUG_CPU
185 /* This is akin to using 'nr_cpus' on the Linux command line.
186 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
187 * have up to X, while nr_cpu_ids is greater than X. This
188 * normally is not a problem, except when CPU hotplugging
189 * is involved and then there might be more than X CPUs
190 * in the guest - which will not work as there is no
191 * hypercall to expand the max number of VCPUs an already
192 * running guest has. So cap it up to X. */
193 if (subtract)
194 nr_cpu_ids = nr_cpu_ids - subtract;
195 #endif
196
197 }
198
xen_pv_smp_prepare_boot_cpu(void)199 static void __init xen_pv_smp_prepare_boot_cpu(void)
200 {
201 BUG_ON(smp_processor_id() != 0);
202 native_smp_prepare_boot_cpu();
203
204 if (!xen_feature(XENFEAT_writable_page_tables))
205 /* We've switched to the "real" per-cpu gdt, so make
206 * sure the old memory can be recycled. */
207 make_lowmem_page_readwrite(xen_initial_gdt);
208
209 #ifdef CONFIG_X86_32
210 /*
211 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
212 * expects __USER_DS
213 */
214 loadsegment(ds, __USER_DS);
215 loadsegment(es, __USER_DS);
216 #endif
217
218 xen_filter_cpu_maps();
219 xen_setup_vcpu_info_placement();
220
221 /*
222 * The alternative logic (which patches the unlock/lock) runs before
223 * the smp bootup up code is activated. Hence we need to set this up
224 * the core kernel is being patched. Otherwise we will have only
225 * modules patched but not core code.
226 */
227 xen_init_spinlocks();
228 }
229
xen_pv_smp_prepare_cpus(unsigned int max_cpus)230 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
231 {
232 unsigned cpu;
233 unsigned int i;
234
235 if (skip_ioapic_setup) {
236 char *m = (max_cpus == 0) ?
237 "The nosmp parameter is incompatible with Xen; " \
238 "use Xen dom0_max_vcpus=1 parameter" :
239 "The noapic parameter is incompatible with Xen";
240
241 xen_raw_printk(m);
242 panic(m);
243 }
244 xen_init_lock_cpu(0);
245
246 smp_store_boot_cpu_info();
247 cpu_data(0).x86_max_cores = 1;
248
249 for_each_possible_cpu(i) {
250 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
251 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
252 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
253 }
254 set_cpu_sibling_map(0);
255
256 speculative_store_bypass_ht_init();
257
258 xen_pmu_init(0);
259
260 if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
261 BUG();
262
263 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
264 panic("could not allocate xen_cpu_initialized_map\n");
265
266 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
267
268 /* Restrict the possible_map according to max_cpus. */
269 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
270 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
271 continue;
272 set_cpu_possible(cpu, false);
273 }
274
275 for_each_possible_cpu(cpu)
276 set_cpu_present(cpu, true);
277 }
278
279 static int
cpu_initialize_context(unsigned int cpu,struct task_struct * idle)280 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
281 {
282 struct vcpu_guest_context *ctxt;
283 struct desc_struct *gdt;
284 unsigned long gdt_mfn;
285
286 /* used to tell cpu_init() that it can proceed with initialization */
287 cpumask_set_cpu(cpu, cpu_callout_mask);
288 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
289 return 0;
290
291 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
292 if (ctxt == NULL)
293 return -ENOMEM;
294
295 gdt = get_cpu_gdt_rw(cpu);
296
297 #ifdef CONFIG_X86_32
298 ctxt->user_regs.fs = __KERNEL_PERCPU;
299 ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
300 #endif
301 memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
302
303 /*
304 * Bring up the CPU in cpu_bringup_and_idle() with the stack
305 * pointing just below where pt_regs would be if it were a normal
306 * kernel entry.
307 */
308 ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
309 ctxt->flags = VGCF_IN_KERNEL;
310 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
311 ctxt->user_regs.ds = __USER_DS;
312 ctxt->user_regs.es = __USER_DS;
313 ctxt->user_regs.ss = __KERNEL_DS;
314 ctxt->user_regs.cs = __KERNEL_CS;
315 ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
316
317 xen_copy_trap_info(ctxt->trap_ctxt);
318
319 ctxt->ldt_ents = 0;
320
321 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
322
323 gdt_mfn = arbitrary_virt_to_mfn(gdt);
324 make_lowmem_page_readonly(gdt);
325 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
326
327 ctxt->gdt_frames[0] = gdt_mfn;
328 ctxt->gdt_ents = GDT_ENTRIES;
329
330 /*
331 * Set SS:SP that Xen will use when entering guest kernel mode
332 * from guest user mode. Subsequent calls to load_sp0() can
333 * change this value.
334 */
335 ctxt->kernel_ss = __KERNEL_DS;
336 ctxt->kernel_sp = task_top_of_stack(idle);
337
338 #ifdef CONFIG_X86_32
339 ctxt->event_callback_cs = __KERNEL_CS;
340 ctxt->failsafe_callback_cs = __KERNEL_CS;
341 #else
342 ctxt->gs_base_kernel = per_cpu_offset(cpu);
343 #endif
344 ctxt->event_callback_eip =
345 (unsigned long)xen_hypervisor_callback;
346 ctxt->failsafe_callback_eip =
347 (unsigned long)xen_failsafe_callback;
348 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
349
350 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
351 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
352 BUG();
353
354 kfree(ctxt);
355 return 0;
356 }
357
xen_pv_cpu_up(unsigned int cpu,struct task_struct * idle)358 static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle)
359 {
360 int rc;
361
362 common_cpu_up(cpu, idle);
363
364 xen_setup_runstate_info(cpu);
365
366 /*
367 * PV VCPUs are always successfully taken down (see 'while' loop
368 * in xen_cpu_die()), so -EBUSY is an error.
369 */
370 rc = cpu_check_up_prepare(cpu);
371 if (rc)
372 return rc;
373
374 /* make sure interrupts start blocked */
375 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
376
377 rc = cpu_initialize_context(cpu, idle);
378 if (rc)
379 return rc;
380
381 xen_pmu_init(cpu);
382
383 rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
384 BUG_ON(rc);
385
386 while (cpu_report_state(cpu) != CPU_ONLINE)
387 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
388
389 return 0;
390 }
391
392 #ifdef CONFIG_HOTPLUG_CPU
xen_pv_cpu_disable(void)393 static int xen_pv_cpu_disable(void)
394 {
395 unsigned int cpu = smp_processor_id();
396 if (cpu == 0)
397 return -EBUSY;
398
399 cpu_disable_common();
400
401 load_cr3(swapper_pg_dir);
402 return 0;
403 }
404
xen_pv_cpu_die(unsigned int cpu)405 static void xen_pv_cpu_die(unsigned int cpu)
406 {
407 while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
408 xen_vcpu_nr(cpu), NULL)) {
409 __set_current_state(TASK_UNINTERRUPTIBLE);
410 schedule_timeout(HZ/10);
411 }
412
413 if (common_cpu_die(cpu) == 0) {
414 xen_smp_intr_free(cpu);
415 xen_uninit_lock_cpu(cpu);
416 xen_teardown_timer(cpu);
417 xen_pmu_finish(cpu);
418 }
419 }
420
xen_pv_play_dead(void)421 static void xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
422 {
423 play_dead_common();
424 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
425 cpu_bringup();
426 /*
427 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
428 * clears certain data that the cpu_idle loop (which called us
429 * and that we return from) expects. The only way to get that
430 * data back is to call:
431 */
432 tick_nohz_idle_enter();
433 tick_nohz_idle_stop_tick_protected();
434
435 cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
436 }
437
438 #else /* !CONFIG_HOTPLUG_CPU */
xen_pv_cpu_disable(void)439 static int xen_pv_cpu_disable(void)
440 {
441 return -ENOSYS;
442 }
443
xen_pv_cpu_die(unsigned int cpu)444 static void xen_pv_cpu_die(unsigned int cpu)
445 {
446 BUG();
447 }
448
xen_pv_play_dead(void)449 static void xen_pv_play_dead(void)
450 {
451 BUG();
452 }
453
454 #endif
stop_self(void * v)455 static void stop_self(void *v)
456 {
457 int cpu = smp_processor_id();
458
459 /* make sure we're not pinning something down */
460 load_cr3(swapper_pg_dir);
461 /* should set up a minimal gdt */
462
463 set_cpu_online(cpu, false);
464
465 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
466 BUG();
467 }
468
xen_pv_stop_other_cpus(int wait)469 static void xen_pv_stop_other_cpus(int wait)
470 {
471 smp_call_function(stop_self, NULL, wait);
472 }
473
xen_irq_work_interrupt(int irq,void * dev_id)474 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
475 {
476 irq_enter();
477 irq_work_run();
478 inc_irq_stat(apic_irq_work_irqs);
479 irq_exit();
480
481 return IRQ_HANDLED;
482 }
483
484 static const struct smp_ops xen_smp_ops __initconst = {
485 .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
486 .smp_prepare_cpus = xen_pv_smp_prepare_cpus,
487 .smp_cpus_done = xen_smp_cpus_done,
488
489 .cpu_up = xen_pv_cpu_up,
490 .cpu_die = xen_pv_cpu_die,
491 .cpu_disable = xen_pv_cpu_disable,
492 .play_dead = xen_pv_play_dead,
493
494 .stop_other_cpus = xen_pv_stop_other_cpus,
495 .smp_send_reschedule = xen_smp_send_reschedule,
496
497 .send_call_func_ipi = xen_smp_send_call_function_ipi,
498 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
499 };
500
xen_smp_init(void)501 void __init xen_smp_init(void)
502 {
503 smp_ops = xen_smp_ops;
504 xen_fill_possible_map();
505 }
506
