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 prevent_tail_call_optimization();
93 }
94
xen_smp_intr_free_pv(unsigned int cpu)95 void xen_smp_intr_free_pv(unsigned int cpu)
96 {
97 if (per_cpu(xen_irq_work, cpu).irq >= 0) {
98 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
99 per_cpu(xen_irq_work, cpu).irq = -1;
100 kfree(per_cpu(xen_irq_work, cpu).name);
101 per_cpu(xen_irq_work, cpu).name = NULL;
102 }
103
104 if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
105 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
106 per_cpu(xen_pmu_irq, cpu).irq = -1;
107 kfree(per_cpu(xen_pmu_irq, cpu).name);
108 per_cpu(xen_pmu_irq, cpu).name = NULL;
109 }
110 }
111
xen_smp_intr_init_pv(unsigned int cpu)112 int xen_smp_intr_init_pv(unsigned int cpu)
113 {
114 int rc;
115 char *callfunc_name, *pmu_name;
116
117 callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
118 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
119 cpu,
120 xen_irq_work_interrupt,
121 IRQF_PERCPU|IRQF_NOBALANCING,
122 callfunc_name,
123 NULL);
124 if (rc < 0)
125 goto fail;
126 per_cpu(xen_irq_work, cpu).irq = rc;
127 per_cpu(xen_irq_work, cpu).name = callfunc_name;
128
129 if (is_xen_pmu(cpu)) {
130 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
131 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
132 xen_pmu_irq_handler,
133 IRQF_PERCPU|IRQF_NOBALANCING,
134 pmu_name, NULL);
135 if (rc < 0)
136 goto fail;
137 per_cpu(xen_pmu_irq, cpu).irq = rc;
138 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
139 }
140
141 return 0;
142
143 fail:
144 xen_smp_intr_free_pv(cpu);
145 return rc;
146 }
147
xen_fill_possible_map(void)148 static void __init xen_fill_possible_map(void)
149 {
150 int i, rc;
151
152 if (xen_initial_domain())
153 return;
154
155 for (i = 0; i < nr_cpu_ids; i++) {
156 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
157 if (rc >= 0) {
158 num_processors++;
159 set_cpu_possible(i, true);
160 }
161 }
162 }
163
xen_filter_cpu_maps(void)164 static void __init xen_filter_cpu_maps(void)
165 {
166 int i, rc;
167 unsigned int subtract = 0;
168
169 if (!xen_initial_domain())
170 return;
171
172 num_processors = 0;
173 disabled_cpus = 0;
174 for (i = 0; i < nr_cpu_ids; i++) {
175 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
176 if (rc >= 0) {
177 num_processors++;
178 set_cpu_possible(i, true);
179 } else {
180 set_cpu_possible(i, false);
181 set_cpu_present(i, false);
182 subtract++;
183 }
184 }
185 #ifdef CONFIG_HOTPLUG_CPU
186 /* This is akin to using 'nr_cpus' on the Linux command line.
187 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
188 * have up to X, while nr_cpu_ids is greater than X. This
189 * normally is not a problem, except when CPU hotplugging
190 * is involved and then there might be more than X CPUs
191 * in the guest - which will not work as there is no
192 * hypercall to expand the max number of VCPUs an already
193 * running guest has. So cap it up to X. */
194 if (subtract)
195 nr_cpu_ids = nr_cpu_ids - subtract;
196 #endif
197
198 }
199
xen_pv_smp_prepare_boot_cpu(void)200 static void __init xen_pv_smp_prepare_boot_cpu(void)
201 {
202 BUG_ON(smp_processor_id() != 0);
203 native_smp_prepare_boot_cpu();
204
205 if (!xen_feature(XENFEAT_writable_page_tables))
206 /* We've switched to the "real" per-cpu gdt, so make
207 * sure the old memory can be recycled. */
208 make_lowmem_page_readwrite(xen_initial_gdt);
209
210 #ifdef CONFIG_X86_32
211 /*
212 * Xen starts us with XEN_FLAT_RING1_DS, but linux code
213 * expects __USER_DS
214 */
215 loadsegment(ds, __USER_DS);
216 loadsegment(es, __USER_DS);
217 #endif
218
219 xen_filter_cpu_maps();
220 xen_setup_vcpu_info_placement();
221
222 /*
223 * The alternative logic (which patches the unlock/lock) runs before
224 * the smp bootup up code is activated. Hence we need to set this up
225 * the core kernel is being patched. Otherwise we will have only
226 * modules patched but not core code.
227 */
228 xen_init_spinlocks();
229 }
230
xen_pv_smp_prepare_cpus(unsigned int max_cpus)231 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
232 {
233 unsigned cpu;
234 unsigned int i;
235
236 if (skip_ioapic_setup) {
237 char *m = (max_cpus == 0) ?
238 "The nosmp parameter is incompatible with Xen; " \
239 "use Xen dom0_max_vcpus=1 parameter" :
240 "The noapic parameter is incompatible with Xen";
241
242 xen_raw_printk(m);
243 panic(m);
244 }
245 xen_init_lock_cpu(0);
246
247 smp_store_boot_cpu_info();
248 cpu_data(0).x86_max_cores = 1;
249
250 for_each_possible_cpu(i) {
251 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
252 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
253 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
254 }
255 set_cpu_sibling_map(0);
256
257 speculative_store_bypass_ht_init();
258
259 xen_pmu_init(0);
260
261 if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
262 BUG();
263
264 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
265 panic("could not allocate xen_cpu_initialized_map\n");
266
267 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
268
269 /* Restrict the possible_map according to max_cpus. */
270 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
271 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
272 continue;
273 set_cpu_possible(cpu, false);
274 }
275
276 for_each_possible_cpu(cpu)
277 set_cpu_present(cpu, true);
278 }
279
280 static int
cpu_initialize_context(unsigned int cpu,struct task_struct * idle)281 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
282 {
283 struct vcpu_guest_context *ctxt;
284 struct desc_struct *gdt;
285 unsigned long gdt_mfn;
286
287 /* used to tell cpu_init() that it can proceed with initialization */
288 cpumask_set_cpu(cpu, cpu_callout_mask);
289 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
290 return 0;
291
292 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
293 if (ctxt == NULL)
294 return -ENOMEM;
295
296 gdt = get_cpu_gdt_rw(cpu);
297
298 #ifdef CONFIG_X86_32
299 ctxt->user_regs.fs = __KERNEL_PERCPU;
300 ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
301 #endif
302 memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
303
304 /*
305 * Bring up the CPU in cpu_bringup_and_idle() with the stack
306 * pointing just below where pt_regs would be if it were a normal
307 * kernel entry.
308 */
309 ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
310 ctxt->flags = VGCF_IN_KERNEL;
311 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
312 ctxt->user_regs.ds = __USER_DS;
313 ctxt->user_regs.es = __USER_DS;
314 ctxt->user_regs.ss = __KERNEL_DS;
315 ctxt->user_regs.cs = __KERNEL_CS;
316 ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
317
318 xen_copy_trap_info(ctxt->trap_ctxt);
319
320 ctxt->ldt_ents = 0;
321
322 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
323
324 gdt_mfn = arbitrary_virt_to_mfn(gdt);
325 make_lowmem_page_readonly(gdt);
326 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
327
328 ctxt->gdt_frames[0] = gdt_mfn;
329 ctxt->gdt_ents = GDT_ENTRIES;
330
331 /*
332 * Set SS:SP that Xen will use when entering guest kernel mode
333 * from guest user mode. Subsequent calls to load_sp0() can
334 * change this value.
335 */
336 ctxt->kernel_ss = __KERNEL_DS;
337 ctxt->kernel_sp = task_top_of_stack(idle);
338
339 #ifdef CONFIG_X86_32
340 ctxt->event_callback_cs = __KERNEL_CS;
341 ctxt->failsafe_callback_cs = __KERNEL_CS;
342 #else
343 ctxt->gs_base_kernel = per_cpu_offset(cpu);
344 #endif
345 ctxt->event_callback_eip =
346 (unsigned long)xen_hypervisor_callback;
347 ctxt->failsafe_callback_eip =
348 (unsigned long)xen_failsafe_callback;
349 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
350
351 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
352 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
353 BUG();
354
355 kfree(ctxt);
356 return 0;
357 }
358
xen_pv_cpu_up(unsigned int cpu,struct task_struct * idle)359 static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle)
360 {
361 int rc;
362
363 common_cpu_up(cpu, idle);
364
365 xen_setup_runstate_info(cpu);
366
367 /*
368 * PV VCPUs are always successfully taken down (see 'while' loop
369 * in xen_cpu_die()), so -EBUSY is an error.
370 */
371 rc = cpu_check_up_prepare(cpu);
372 if (rc)
373 return rc;
374
375 /* make sure interrupts start blocked */
376 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
377
378 rc = cpu_initialize_context(cpu, idle);
379 if (rc)
380 return rc;
381
382 xen_pmu_init(cpu);
383
384 rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
385 BUG_ON(rc);
386
387 while (cpu_report_state(cpu) != CPU_ONLINE)
388 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
389
390 return 0;
391 }
392
393 #ifdef CONFIG_HOTPLUG_CPU
xen_pv_cpu_disable(void)394 static int xen_pv_cpu_disable(void)
395 {
396 unsigned int cpu = smp_processor_id();
397 if (cpu == 0)
398 return -EBUSY;
399
400 cpu_disable_common();
401
402 load_cr3(swapper_pg_dir);
403 return 0;
404 }
405
xen_pv_cpu_die(unsigned int cpu)406 static void xen_pv_cpu_die(unsigned int cpu)
407 {
408 while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
409 xen_vcpu_nr(cpu), NULL)) {
410 __set_current_state(TASK_UNINTERRUPTIBLE);
411 schedule_timeout(HZ/10);
412 }
413
414 if (common_cpu_die(cpu) == 0) {
415 xen_smp_intr_free(cpu);
416 xen_uninit_lock_cpu(cpu);
417 xen_teardown_timer(cpu);
418 xen_pmu_finish(cpu);
419 }
420 }
421
xen_pv_play_dead(void)422 static void xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
423 {
424 play_dead_common();
425 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
426 cpu_bringup();
427 /*
428 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
429 * clears certain data that the cpu_idle loop (which called us
430 * and that we return from) expects. The only way to get that
431 * data back is to call:
432 */
433 tick_nohz_idle_enter();
434 tick_nohz_idle_stop_tick_protected();
435
436 cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
437 }
438
439 #else /* !CONFIG_HOTPLUG_CPU */
xen_pv_cpu_disable(void)440 static int xen_pv_cpu_disable(void)
441 {
442 return -ENOSYS;
443 }
444
xen_pv_cpu_die(unsigned int cpu)445 static void xen_pv_cpu_die(unsigned int cpu)
446 {
447 BUG();
448 }
449
xen_pv_play_dead(void)450 static void xen_pv_play_dead(void)
451 {
452 BUG();
453 }
454
455 #endif
stop_self(void * v)456 static void stop_self(void *v)
457 {
458 int cpu = smp_processor_id();
459
460 /* make sure we're not pinning something down */
461 load_cr3(swapper_pg_dir);
462 /* should set up a minimal gdt */
463
464 set_cpu_online(cpu, false);
465
466 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
467 BUG();
468 }
469
xen_pv_stop_other_cpus(int wait)470 static void xen_pv_stop_other_cpus(int wait)
471 {
472 smp_call_function(stop_self, NULL, wait);
473 }
474
xen_irq_work_interrupt(int irq,void * dev_id)475 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
476 {
477 irq_enter();
478 irq_work_run();
479 inc_irq_stat(apic_irq_work_irqs);
480 irq_exit();
481
482 return IRQ_HANDLED;
483 }
484
485 static const struct smp_ops xen_smp_ops __initconst = {
486 .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
487 .smp_prepare_cpus = xen_pv_smp_prepare_cpus,
488 .smp_cpus_done = xen_smp_cpus_done,
489
490 .cpu_up = xen_pv_cpu_up,
491 .cpu_die = xen_pv_cpu_die,
492 .cpu_disable = xen_pv_cpu_disable,
493 .play_dead = xen_pv_play_dead,
494
495 .stop_other_cpus = xen_pv_stop_other_cpus,
496 .smp_send_reschedule = xen_smp_send_reschedule,
497
498 .send_call_func_ipi = xen_smp_send_call_function_ipi,
499 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
500 };
501
xen_smp_init(void)502 void __init xen_smp_init(void)
503 {
504 smp_ops = xen_smp_ops;
505 xen_fill_possible_map();
506 }
507