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 #include <linux/stackprotector.h>
26 #include <linux/pgtable.h>
27
28 #include <asm/paravirt.h>
29 #include <asm/idtentry.h>
30 #include <asm/desc.h>
31 #include <asm/cpu.h>
32 #include <asm/apic.h>
33 #include <asm/io_apic.h>
34
35 #include <xen/interface/xen.h>
36 #include <xen/interface/vcpu.h>
37 #include <xen/interface/xenpmu.h>
38
39 #include <asm/spec-ctrl.h>
40 #include <asm/xen/interface.h>
41 #include <asm/xen/hypercall.h>
42
43 #include <xen/xen.h>
44 #include <xen/page.h>
45 #include <xen/events.h>
46
47 #include <xen/hvc-console.h>
48 #include "xen-ops.h"
49
50 cpumask_var_t xen_cpu_initialized_map;
51
52 static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
53 static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
54
55 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
56
cpu_bringup(void)57 static void cpu_bringup(void)
58 {
59 int cpu;
60
61 cr4_init();
62 cpuhp_ap_sync_alive();
63 cpu_init();
64 fpu__init_cpu();
65 touch_softlockup_watchdog();
66
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_syscall();
70
71 cpu = smp_processor_id();
72 identify_secondary_cpu(cpu);
73 set_cpu_sibling_map(cpu);
74
75 speculative_store_bypass_ht_init();
76
77 xen_setup_cpu_clockevents();
78
79 notify_cpu_starting(cpu);
80
81 set_cpu_online(cpu, true);
82
83 smp_mb();
84
85 /* We can take interrupts now: we're officially "up". */
86 local_irq_enable();
87 }
88
cpu_bringup_and_idle(void)89 asmlinkage __visible void cpu_bringup_and_idle(void)
90 {
91 cpu_bringup();
92 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
93 }
94
xen_smp_intr_free_pv(unsigned int cpu)95 void xen_smp_intr_free_pv(unsigned int cpu)
96 {
97 kfree(per_cpu(xen_irq_work, cpu).name);
98 per_cpu(xen_irq_work, cpu).name = NULL;
99 if (per_cpu(xen_irq_work, cpu).irq >= 0) {
100 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
101 per_cpu(xen_irq_work, cpu).irq = -1;
102 }
103
104 kfree(per_cpu(xen_pmu_irq, cpu).name);
105 per_cpu(xen_pmu_irq, cpu).name = NULL;
106 if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
107 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
108 per_cpu(xen_pmu_irq, cpu).irq = -1;
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 per_cpu(xen_irq_work, cpu).name = callfunc_name;
119 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
120 cpu,
121 xen_irq_work_interrupt,
122 IRQF_PERCPU|IRQF_NOBALANCING,
123 callfunc_name,
124 NULL);
125 if (rc < 0)
126 goto fail;
127 per_cpu(xen_irq_work, cpu).irq = rc;
128
129 if (is_xen_pmu) {
130 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
131 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
132 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
133 xen_pmu_irq_handler,
134 IRQF_PERCPU|IRQF_NOBALANCING,
135 pmu_name, NULL);
136 if (rc < 0)
137 goto fail;
138 per_cpu(xen_pmu_irq, cpu).irq = rc;
139 }
140
141 return 0;
142
143 fail:
144 xen_smp_intr_free_pv(cpu);
145 return rc;
146 }
147
xen_pv_smp_config(void)148 static void __init xen_pv_smp_config(void)
149 {
150 u32 apicid = 0;
151 int i;
152
153 topology_register_boot_apic(apicid);
154
155 for (i = 0; i < nr_cpu_ids; i++)
156 topology_register_apic(apicid++, CPU_ACPIID_INVALID, true);
157
158 /* Pretend to be a proper enumerated system */
159 smp_found_config = 1;
160 }
161
xen_pv_smp_prepare_boot_cpu(void)162 static void __init xen_pv_smp_prepare_boot_cpu(void)
163 {
164 BUG_ON(smp_processor_id() != 0);
165 native_smp_prepare_boot_cpu();
166
167 if (!xen_feature(XENFEAT_writable_page_tables))
168 /* We've switched to the "real" per-cpu gdt, so make
169 * sure the old memory can be recycled. */
170 make_lowmem_page_readwrite(xen_initial_gdt);
171
172 xen_setup_vcpu_info_placement();
173
174 /*
175 * The alternative logic (which patches the unlock/lock) runs before
176 * the smp bootup up code is activated. Hence we need to set this up
177 * the core kernel is being patched. Otherwise we will have only
178 * modules patched but not core code.
179 */
180 xen_init_spinlocks();
181 }
182
xen_pv_smp_prepare_cpus(unsigned int max_cpus)183 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
184 {
185 unsigned cpu;
186
187 if (ioapic_is_disabled) {
188 char *m = (max_cpus == 0) ?
189 "The nosmp parameter is incompatible with Xen; " \
190 "use Xen dom0_max_vcpus=1 parameter" :
191 "The noapic parameter is incompatible with Xen";
192
193 xen_raw_printk(m);
194 panic(m);
195 }
196 xen_init_lock_cpu(0);
197
198 smp_prepare_cpus_common();
199
200 speculative_store_bypass_ht_init();
201
202 xen_pmu_init(0);
203
204 if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
205 BUG();
206
207 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
208 panic("could not allocate xen_cpu_initialized_map\n");
209
210 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
211
212 /* Restrict the possible_map according to max_cpus. */
213 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
214 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
215 continue;
216 set_cpu_possible(cpu, false);
217 }
218
219 for_each_possible_cpu(cpu)
220 set_cpu_present(cpu, true);
221 }
222
223 static int
cpu_initialize_context(unsigned int cpu,struct task_struct * idle)224 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
225 {
226 struct vcpu_guest_context *ctxt;
227 struct desc_struct *gdt;
228 unsigned long gdt_mfn;
229
230 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
231 return 0;
232
233 ctxt = kzalloc_obj(*ctxt);
234 if (ctxt == NULL) {
235 cpumask_clear_cpu(cpu, xen_cpu_initialized_map);
236 return -ENOMEM;
237 }
238
239 gdt = get_cpu_gdt_rw(cpu);
240
241 /*
242 * Bring up the CPU in cpu_bringup_and_idle() with the stack
243 * pointing just below where pt_regs would be if it were a normal
244 * kernel entry.
245 */
246 ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle;
247 ctxt->flags = VGCF_IN_KERNEL;
248 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
249 ctxt->user_regs.ds = __USER_DS;
250 ctxt->user_regs.es = __USER_DS;
251 ctxt->user_regs.ss = __KERNEL_DS;
252 ctxt->user_regs.cs = __KERNEL_CS;
253 ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
254
255 xen_copy_trap_info(ctxt->trap_ctxt);
256
257 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
258
259 gdt_mfn = arbitrary_virt_to_mfn(gdt);
260 make_lowmem_page_readonly(gdt);
261 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
262
263 ctxt->gdt_frames[0] = gdt_mfn;
264 ctxt->gdt_ents = GDT_ENTRIES;
265
266 /*
267 * Set SS:SP that Xen will use when entering guest kernel mode
268 * from guest user mode. Subsequent calls to load_sp0() can
269 * change this value.
270 */
271 ctxt->kernel_ss = __KERNEL_DS;
272 ctxt->kernel_sp = task_top_of_stack(idle);
273
274 ctxt->gs_base_kernel = per_cpu_offset(cpu);
275 ctxt->event_callback_eip =
276 (unsigned long)xen_asm_exc_xen_hypervisor_callback;
277 ctxt->failsafe_callback_eip =
278 (unsigned long)xen_failsafe_callback;
279 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
280
281 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
282 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
283 BUG();
284
285 kfree(ctxt);
286 return 0;
287 }
288
xen_pv_kick_ap(unsigned int cpu,struct task_struct * idle)289 static int xen_pv_kick_ap(unsigned int cpu, struct task_struct *idle)
290 {
291 int rc;
292
293 rc = common_cpu_up(cpu, idle);
294 if (rc)
295 return rc;
296
297 xen_setup_runstate_info(cpu);
298
299 /* make sure interrupts start blocked */
300 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
301
302 rc = cpu_initialize_context(cpu, idle);
303 if (rc)
304 return rc;
305
306 xen_pmu_init(cpu);
307
308 /*
309 * Why is this a BUG? If the hypercall fails then everything can be
310 * rolled back, no?
311 */
312 BUG_ON(HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL));
313
314 return 0;
315 }
316
xen_pv_poll_sync_state(void)317 static void xen_pv_poll_sync_state(void)
318 {
319 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
320 }
321
322 #ifdef CONFIG_HOTPLUG_CPU
xen_pv_cpu_disable(void)323 static int xen_pv_cpu_disable(void)
324 {
325 unsigned int cpu = smp_processor_id();
326 if (cpu == 0)
327 return -EBUSY;
328
329 cpu_disable_common();
330
331 load_cr3(swapper_pg_dir);
332 return 0;
333 }
334
xen_pv_cpu_die(unsigned int cpu)335 static void xen_pv_cpu_die(unsigned int cpu)
336 {
337 while (HYPERVISOR_vcpu_op(VCPUOP_is_up, xen_vcpu_nr(cpu), NULL)) {
338 __set_current_state(TASK_UNINTERRUPTIBLE);
339 schedule_timeout(HZ/10);
340 }
341 }
342
xen_pv_cleanup_dead_cpu(unsigned int cpu)343 static void xen_pv_cleanup_dead_cpu(unsigned int cpu)
344 {
345 xen_smp_intr_free(cpu);
346 xen_uninit_lock_cpu(cpu);
347 xen_teardown_timer(cpu);
348 xen_pmu_finish(cpu);
349 }
350
xen_pv_play_dead(void)351 static void __noreturn xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
352 {
353 play_dead_common();
354 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
355 xen_cpu_bringup_again((unsigned long)task_pt_regs(current));
356 BUG();
357 }
358
359 #else /* !CONFIG_HOTPLUG_CPU */
xen_pv_cpu_disable(void)360 static int xen_pv_cpu_disable(void)
361 {
362 return -ENOSYS;
363 }
364
xen_pv_cpu_die(unsigned int cpu)365 static void xen_pv_cpu_die(unsigned int cpu)
366 {
367 BUG();
368 }
369
xen_pv_cleanup_dead_cpu(unsigned int cpu)370 static void xen_pv_cleanup_dead_cpu(unsigned int cpu)
371 {
372 BUG();
373 }
374
xen_pv_play_dead(void)375 static void __noreturn xen_pv_play_dead(void)
376 {
377 BUG();
378 }
379
380 #endif
stop_self(void * v)381 static void stop_self(void *v)
382 {
383 int cpu = smp_processor_id();
384
385 /* make sure we're not pinning something down */
386 load_cr3(swapper_pg_dir);
387 /* should set up a minimal gdt */
388
389 set_cpu_online(cpu, false);
390
391 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
392 BUG();
393 }
394
xen_pv_stop_other_cpus(int wait)395 static void xen_pv_stop_other_cpus(int wait)
396 {
397 smp_call_function(stop_self, NULL, wait);
398 }
399
xen_irq_work_interrupt(int irq,void * dev_id)400 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
401 {
402 irq_work_run();
403 inc_irq_stat(apic_irq_work_irqs);
404
405 return IRQ_HANDLED;
406 }
407
xen_smp_count_cpus(void)408 void __init xen_smp_count_cpus(void)
409 {
410 unsigned int cpus;
411
412 for (cpus = 0; cpus < nr_cpu_ids; cpus++) {
413 if (HYPERVISOR_vcpu_op(VCPUOP_is_up, cpus, NULL) < 0)
414 break;
415 }
416
417 pr_info("Xen PV: Detected %u vCPUS\n", cpus);
418 if (cpus < nr_cpu_ids)
419 set_nr_cpu_ids(cpus);
420 }
421
422 static const struct smp_ops xen_smp_ops __initconst = {
423 .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
424 .smp_prepare_cpus = xen_pv_smp_prepare_cpus,
425 .smp_cpus_done = xen_smp_cpus_done,
426
427 .kick_ap_alive = xen_pv_kick_ap,
428 .cpu_die = xen_pv_cpu_die,
429 .cleanup_dead_cpu = xen_pv_cleanup_dead_cpu,
430 .poll_sync_state = xen_pv_poll_sync_state,
431 .cpu_disable = xen_pv_cpu_disable,
432 .play_dead = xen_pv_play_dead,
433
434 .stop_other_cpus = xen_pv_stop_other_cpus,
435 .smp_send_reschedule = xen_smp_send_reschedule,
436
437 .send_call_func_ipi = xen_smp_send_call_function_ipi,
438 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
439 };
440
xen_smp_init(void)441 void __init xen_smp_init(void)
442 {
443 smp_ops = xen_smp_ops;
444
445 /* Avoid searching for BIOS MP tables */
446 x86_init.mpparse.find_mptable = x86_init_noop;
447 x86_init.mpparse.early_parse_smp_cfg = x86_init_noop;
448
449 /* XEN/PV Dom0 has halfways sane topology information via CPUID/MADT */
450 if (xen_initial_domain())
451 x86_init.mpparse.parse_smp_cfg = x86_init_noop;
452 else
453 x86_init.mpparse.parse_smp_cfg = xen_pv_smp_config;
454 }
455