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/io_apic.h>
33
34 #include <xen/interface/xen.h>
35 #include <xen/interface/vcpu.h>
36 #include <xen/interface/xenpmu.h>
37
38 #include <asm/spec-ctrl.h>
39 #include <asm/xen/interface.h>
40 #include <asm/xen/hypercall.h>
41
42 #include <xen/xen.h>
43 #include <xen/page.h>
44 #include <xen/events.h>
45
46 #include <xen/hvc-console.h>
47 #include "xen-ops.h"
48 #include "mmu.h"
49 #include "smp.h"
50 #include "pmu.h"
51
52 cpumask_var_t xen_cpu_initialized_map;
53
54 static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
55 static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
56
57 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
58 void asm_cpu_bringup_and_idle(void);
59
cpu_bringup(void)60 static void cpu_bringup(void)
61 {
62 int cpu;
63
64 cr4_init();
65 cpu_init();
66 touch_softlockup_watchdog();
67
68 /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
69 if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
70 xen_enable_sysenter();
71 xen_enable_syscall();
72 }
73 cpu = smp_processor_id();
74 smp_store_cpu_info(cpu);
75 cpu_data(cpu).x86_max_cores = 1;
76 set_cpu_sibling_map(cpu);
77
78 speculative_store_bypass_ht_init();
79
80 xen_setup_cpu_clockevents();
81
82 notify_cpu_starting(cpu);
83
84 set_cpu_online(cpu, true);
85
86 cpu_set_state_online(cpu); /* Implies full memory barrier. */
87
88 /* We can take interrupts now: we're officially "up". */
89 local_irq_enable();
90 }
91
cpu_bringup_and_idle(void)92 asmlinkage __visible void cpu_bringup_and_idle(void)
93 {
94 cpu_bringup();
95 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
96 }
97
xen_smp_intr_free_pv(unsigned int cpu)98 void xen_smp_intr_free_pv(unsigned int cpu)
99 {
100 kfree(per_cpu(xen_irq_work, cpu).name);
101 per_cpu(xen_irq_work, cpu).name = NULL;
102 if (per_cpu(xen_irq_work, cpu).irq >= 0) {
103 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
104 per_cpu(xen_irq_work, cpu).irq = -1;
105 }
106
107 kfree(per_cpu(xen_pmu_irq, cpu).name);
108 per_cpu(xen_pmu_irq, cpu).name = NULL;
109 if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
110 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
111 per_cpu(xen_pmu_irq, cpu).irq = -1;
112 }
113 }
114
xen_smp_intr_init_pv(unsigned int cpu)115 int xen_smp_intr_init_pv(unsigned int cpu)
116 {
117 int rc;
118 char *callfunc_name, *pmu_name;
119
120 callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
121 per_cpu(xen_irq_work, cpu).name = callfunc_name;
122 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
123 cpu,
124 xen_irq_work_interrupt,
125 IRQF_PERCPU|IRQF_NOBALANCING,
126 callfunc_name,
127 NULL);
128 if (rc < 0)
129 goto fail;
130 per_cpu(xen_irq_work, cpu).irq = rc;
131
132 if (is_xen_pmu) {
133 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
134 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
135 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
136 xen_pmu_irq_handler,
137 IRQF_PERCPU|IRQF_NOBALANCING,
138 pmu_name, NULL);
139 if (rc < 0)
140 goto fail;
141 per_cpu(xen_pmu_irq, cpu).irq = rc;
142 }
143
144 return 0;
145
146 fail:
147 xen_smp_intr_free_pv(cpu);
148 return rc;
149 }
150
_get_smp_config(unsigned int early)151 static void __init _get_smp_config(unsigned int early)
152 {
153 int i, rc;
154 unsigned int subtract = 0;
155
156 if (early)
157 return;
158
159 num_processors = 0;
160 disabled_cpus = 0;
161 for (i = 0; i < nr_cpu_ids; i++) {
162 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
163 if (rc >= 0) {
164 num_processors++;
165 set_cpu_possible(i, true);
166 } else {
167 set_cpu_possible(i, false);
168 set_cpu_present(i, false);
169 subtract++;
170 }
171 }
172 #ifdef CONFIG_HOTPLUG_CPU
173 /* This is akin to using 'nr_cpus' on the Linux command line.
174 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
175 * have up to X, while nr_cpu_ids is greater than X. This
176 * normally is not a problem, except when CPU hotplugging
177 * is involved and then there might be more than X CPUs
178 * in the guest - which will not work as there is no
179 * hypercall to expand the max number of VCPUs an already
180 * running guest has. So cap it up to X. */
181 if (subtract)
182 set_nr_cpu_ids(nr_cpu_ids - subtract);
183 #endif
184
185 }
186
xen_pv_smp_prepare_boot_cpu(void)187 static void __init xen_pv_smp_prepare_boot_cpu(void)
188 {
189 BUG_ON(smp_processor_id() != 0);
190 native_smp_prepare_boot_cpu();
191
192 if (!xen_feature(XENFEAT_writable_page_tables))
193 /* We've switched to the "real" per-cpu gdt, so make
194 * sure the old memory can be recycled. */
195 make_lowmem_page_readwrite(xen_initial_gdt);
196
197 xen_setup_vcpu_info_placement();
198
199 /*
200 * The alternative logic (which patches the unlock/lock) runs before
201 * the smp bootup up code is activated. Hence we need to set this up
202 * the core kernel is being patched. Otherwise we will have only
203 * modules patched but not core code.
204 */
205 xen_init_spinlocks();
206 }
207
xen_pv_smp_prepare_cpus(unsigned int max_cpus)208 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
209 {
210 unsigned cpu;
211
212 if (skip_ioapic_setup) {
213 char *m = (max_cpus == 0) ?
214 "The nosmp parameter is incompatible with Xen; " \
215 "use Xen dom0_max_vcpus=1 parameter" :
216 "The noapic parameter is incompatible with Xen";
217
218 xen_raw_printk(m);
219 panic(m);
220 }
221 xen_init_lock_cpu(0);
222
223 smp_prepare_cpus_common();
224
225 cpu_data(0).x86_max_cores = 1;
226
227 speculative_store_bypass_ht_init();
228
229 xen_pmu_init(0);
230
231 if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
232 BUG();
233
234 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
235 panic("could not allocate xen_cpu_initialized_map\n");
236
237 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
238
239 /* Restrict the possible_map according to max_cpus. */
240 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
241 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
242 continue;
243 set_cpu_possible(cpu, false);
244 }
245
246 for_each_possible_cpu(cpu)
247 set_cpu_present(cpu, true);
248 }
249
250 static int
cpu_initialize_context(unsigned int cpu,struct task_struct * idle)251 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
252 {
253 struct vcpu_guest_context *ctxt;
254 struct desc_struct *gdt;
255 unsigned long gdt_mfn;
256
257 /* used to tell cpu_init() that it can proceed with initialization */
258 cpumask_set_cpu(cpu, cpu_callout_mask);
259 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
260 return 0;
261
262 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
263 if (ctxt == NULL) {
264 cpumask_clear_cpu(cpu, xen_cpu_initialized_map);
265 cpumask_clear_cpu(cpu, cpu_callout_mask);
266 return -ENOMEM;
267 }
268
269 gdt = get_cpu_gdt_rw(cpu);
270
271 /*
272 * Bring up the CPU in cpu_bringup_and_idle() with the stack
273 * pointing just below where pt_regs would be if it were a normal
274 * kernel entry.
275 */
276 ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle;
277 ctxt->flags = VGCF_IN_KERNEL;
278 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
279 ctxt->user_regs.ds = __USER_DS;
280 ctxt->user_regs.es = __USER_DS;
281 ctxt->user_regs.ss = __KERNEL_DS;
282 ctxt->user_regs.cs = __KERNEL_CS;
283 ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
284
285 xen_copy_trap_info(ctxt->trap_ctxt);
286
287 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
288
289 gdt_mfn = arbitrary_virt_to_mfn(gdt);
290 make_lowmem_page_readonly(gdt);
291 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
292
293 ctxt->gdt_frames[0] = gdt_mfn;
294 ctxt->gdt_ents = GDT_ENTRIES;
295
296 /*
297 * Set SS:SP that Xen will use when entering guest kernel mode
298 * from guest user mode. Subsequent calls to load_sp0() can
299 * change this value.
300 */
301 ctxt->kernel_ss = __KERNEL_DS;
302 ctxt->kernel_sp = task_top_of_stack(idle);
303
304 ctxt->gs_base_kernel = per_cpu_offset(cpu);
305 ctxt->event_callback_eip =
306 (unsigned long)xen_asm_exc_xen_hypervisor_callback;
307 ctxt->failsafe_callback_eip =
308 (unsigned long)xen_failsafe_callback;
309 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
310
311 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
312 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
313 BUG();
314
315 kfree(ctxt);
316 return 0;
317 }
318
xen_pv_cpu_up(unsigned int cpu,struct task_struct * idle)319 static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle)
320 {
321 int rc;
322
323 rc = common_cpu_up(cpu, idle);
324 if (rc)
325 return rc;
326
327 xen_setup_runstate_info(cpu);
328
329 /*
330 * PV VCPUs are always successfully taken down (see 'while' loop
331 * in xen_cpu_die()), so -EBUSY is an error.
332 */
333 rc = cpu_check_up_prepare(cpu);
334 if (rc)
335 return rc;
336
337 /* make sure interrupts start blocked */
338 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
339
340 rc = cpu_initialize_context(cpu, idle);
341 if (rc)
342 return rc;
343
344 xen_pmu_init(cpu);
345
346 rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
347 BUG_ON(rc);
348
349 while (cpu_report_state(cpu) != CPU_ONLINE)
350 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
351
352 return 0;
353 }
354
355 #ifdef CONFIG_HOTPLUG_CPU
xen_pv_cpu_disable(void)356 static int xen_pv_cpu_disable(void)
357 {
358 unsigned int cpu = smp_processor_id();
359 if (cpu == 0)
360 return -EBUSY;
361
362 cpu_disable_common();
363
364 load_cr3(swapper_pg_dir);
365 return 0;
366 }
367
xen_pv_cpu_die(unsigned int cpu)368 static void xen_pv_cpu_die(unsigned int cpu)
369 {
370 while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
371 xen_vcpu_nr(cpu), NULL)) {
372 __set_current_state(TASK_UNINTERRUPTIBLE);
373 schedule_timeout(HZ/10);
374 }
375
376 if (common_cpu_die(cpu) == 0) {
377 xen_smp_intr_free(cpu);
378 xen_uninit_lock_cpu(cpu);
379 xen_teardown_timer(cpu);
380 xen_pmu_finish(cpu);
381 }
382 }
383
xen_pv_play_dead(void)384 static void xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
385 {
386 play_dead_common();
387 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
388 cpu_bringup();
389 /*
390 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
391 * clears certain data that the cpu_idle loop (which called us
392 * and that we return from) expects. The only way to get that
393 * data back is to call:
394 */
395 tick_nohz_idle_enter();
396 tick_nohz_idle_stop_tick_protected();
397
398 cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
399 }
400
401 #else /* !CONFIG_HOTPLUG_CPU */
xen_pv_cpu_disable(void)402 static int xen_pv_cpu_disable(void)
403 {
404 return -ENOSYS;
405 }
406
xen_pv_cpu_die(unsigned int cpu)407 static void xen_pv_cpu_die(unsigned int cpu)
408 {
409 BUG();
410 }
411
xen_pv_play_dead(void)412 static void xen_pv_play_dead(void)
413 {
414 BUG();
415 }
416
417 #endif
stop_self(void * v)418 static void stop_self(void *v)
419 {
420 int cpu = smp_processor_id();
421
422 /* make sure we're not pinning something down */
423 load_cr3(swapper_pg_dir);
424 /* should set up a minimal gdt */
425
426 set_cpu_online(cpu, false);
427
428 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
429 BUG();
430 }
431
xen_pv_stop_other_cpus(int wait)432 static void xen_pv_stop_other_cpus(int wait)
433 {
434 smp_call_function(stop_self, NULL, wait);
435 }
436
xen_irq_work_interrupt(int irq,void * dev_id)437 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
438 {
439 irq_work_run();
440 inc_irq_stat(apic_irq_work_irqs);
441
442 return IRQ_HANDLED;
443 }
444
445 static const struct smp_ops xen_smp_ops __initconst = {
446 .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
447 .smp_prepare_cpus = xen_pv_smp_prepare_cpus,
448 .smp_cpus_done = xen_smp_cpus_done,
449
450 .cpu_up = xen_pv_cpu_up,
451 .cpu_die = xen_pv_cpu_die,
452 .cpu_disable = xen_pv_cpu_disable,
453 .play_dead = xen_pv_play_dead,
454
455 .stop_other_cpus = xen_pv_stop_other_cpus,
456 .smp_send_reschedule = xen_smp_send_reschedule,
457
458 .send_call_func_ipi = xen_smp_send_call_function_ipi,
459 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
460 };
461
xen_smp_init(void)462 void __init xen_smp_init(void)
463 {
464 smp_ops = xen_smp_ops;
465
466 /* Avoid searching for BIOS MP tables */
467 x86_init.mpparse.find_smp_config = x86_init_noop;
468 x86_init.mpparse.get_smp_config = _get_smp_config;
469 }
470