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 #include <asm/fpu/internal.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 #include "mmu.h"
50 #include "smp.h"
51 #include "pmu.h"
52
53 cpumask_var_t xen_cpu_initialized_map;
54
55 static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
56 static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
57
58 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
59 void asm_cpu_bringup_and_idle(void);
60
cpu_bringup(void)61 static void cpu_bringup(void)
62 {
63 int cpu;
64
65 cr4_init();
66 cpu_init();
67 fpu__init_cpu();
68 touch_softlockup_watchdog();
69
70 /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
71 if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
72 xen_enable_sysenter();
73 xen_enable_syscall();
74 }
75 cpu = smp_processor_id();
76 smp_store_cpu_info(cpu);
77 cpu_data(cpu).x86_max_cores = 1;
78 set_cpu_sibling_map(cpu);
79
80 speculative_store_bypass_ht_init();
81
82 xen_setup_cpu_clockevents();
83
84 notify_cpu_starting(cpu);
85
86 set_cpu_online(cpu, true);
87
88 cpu_set_state_online(cpu); /* Implies full memory barrier. */
89
90 /* We can take interrupts now: we're officially "up". */
91 local_irq_enable();
92 }
93
cpu_bringup_and_idle(void)94 asmlinkage __visible void cpu_bringup_and_idle(void)
95 {
96 cpu_bringup();
97 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
98 }
99
xen_smp_intr_free_pv(unsigned int cpu)100 void xen_smp_intr_free_pv(unsigned int cpu)
101 {
102 kfree(per_cpu(xen_irq_work, cpu).name);
103 per_cpu(xen_irq_work, cpu).name = NULL;
104 if (per_cpu(xen_irq_work, cpu).irq >= 0) {
105 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
106 per_cpu(xen_irq_work, cpu).irq = -1;
107 }
108
109 kfree(per_cpu(xen_pmu_irq, cpu).name);
110 per_cpu(xen_pmu_irq, cpu).name = NULL;
111 if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
112 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
113 per_cpu(xen_pmu_irq, cpu).irq = -1;
114 }
115 }
116
xen_smp_intr_init_pv(unsigned int cpu)117 int xen_smp_intr_init_pv(unsigned int cpu)
118 {
119 int rc;
120 char *callfunc_name, *pmu_name;
121
122 callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
123 per_cpu(xen_irq_work, cpu).name = callfunc_name;
124 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
125 cpu,
126 xen_irq_work_interrupt,
127 IRQF_PERCPU|IRQF_NOBALANCING,
128 callfunc_name,
129 NULL);
130 if (rc < 0)
131 goto fail;
132 per_cpu(xen_irq_work, cpu).irq = rc;
133
134 if (is_xen_pmu) {
135 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
136 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
137 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
138 xen_pmu_irq_handler,
139 IRQF_PERCPU|IRQF_NOBALANCING,
140 pmu_name, NULL);
141 if (rc < 0)
142 goto fail;
143 per_cpu(xen_pmu_irq, cpu).irq = rc;
144 }
145
146 return 0;
147
148 fail:
149 xen_smp_intr_free_pv(cpu);
150 return rc;
151 }
152
_get_smp_config(unsigned int early)153 static void __init _get_smp_config(unsigned int early)
154 {
155 int i, rc;
156 unsigned int subtract = 0;
157
158 if (early)
159 return;
160
161 num_processors = 0;
162 disabled_cpus = 0;
163 for (i = 0; i < nr_cpu_ids; i++) {
164 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
165 if (rc >= 0) {
166 num_processors++;
167 set_cpu_possible(i, true);
168 } else {
169 set_cpu_possible(i, false);
170 set_cpu_present(i, false);
171 subtract++;
172 }
173 }
174 #ifdef CONFIG_HOTPLUG_CPU
175 /* This is akin to using 'nr_cpus' on the Linux command line.
176 * Which is OK as when we use 'dom0_max_vcpus=X' we can only
177 * have up to X, while nr_cpu_ids is greater than X. This
178 * normally is not a problem, except when CPU hotplugging
179 * is involved and then there might be more than X CPUs
180 * in the guest - which will not work as there is no
181 * hypercall to expand the max number of VCPUs an already
182 * running guest has. So cap it up to X. */
183 if (subtract)
184 nr_cpu_ids = nr_cpu_ids - subtract;
185 #endif
186
187 }
188
xen_pv_smp_prepare_boot_cpu(void)189 static void __init xen_pv_smp_prepare_boot_cpu(void)
190 {
191 BUG_ON(smp_processor_id() != 0);
192 native_smp_prepare_boot_cpu();
193
194 if (!xen_feature(XENFEAT_writable_page_tables))
195 /* We've switched to the "real" per-cpu gdt, so make
196 * sure the old memory can be recycled. */
197 make_lowmem_page_readwrite(xen_initial_gdt);
198
199 xen_setup_vcpu_info_placement();
200
201 /*
202 * The alternative logic (which patches the unlock/lock) runs before
203 * the smp bootup up code is activated. Hence we need to set this up
204 * the core kernel is being patched. Otherwise we will have only
205 * modules patched but not core code.
206 */
207 xen_init_spinlocks();
208 }
209
xen_pv_smp_prepare_cpus(unsigned int max_cpus)210 static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus)
211 {
212 unsigned cpu;
213 unsigned int i;
214
215 if (skip_ioapic_setup) {
216 char *m = (max_cpus == 0) ?
217 "The nosmp parameter is incompatible with Xen; " \
218 "use Xen dom0_max_vcpus=1 parameter" :
219 "The noapic parameter is incompatible with Xen";
220
221 xen_raw_printk(m);
222 panic(m);
223 }
224 xen_init_lock_cpu(0);
225
226 smp_store_boot_cpu_info();
227 cpu_data(0).x86_max_cores = 1;
228
229 for_each_possible_cpu(i) {
230 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
231 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
232 zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
233 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
234 }
235 set_cpu_sibling_map(0);
236
237 speculative_store_bypass_ht_init();
238
239 xen_pmu_init(0);
240
241 if (xen_smp_intr_init(0) || xen_smp_intr_init_pv(0))
242 BUG();
243
244 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
245 panic("could not allocate xen_cpu_initialized_map\n");
246
247 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
248
249 /* Restrict the possible_map according to max_cpus. */
250 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
251 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
252 continue;
253 set_cpu_possible(cpu, false);
254 }
255
256 for_each_possible_cpu(cpu)
257 set_cpu_present(cpu, true);
258 }
259
260 static int
cpu_initialize_context(unsigned int cpu,struct task_struct * idle)261 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
262 {
263 struct vcpu_guest_context *ctxt;
264 struct desc_struct *gdt;
265 unsigned long gdt_mfn;
266
267 /* used to tell cpu_init() that it can proceed with initialization */
268 cpumask_set_cpu(cpu, cpu_callout_mask);
269 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
270 return 0;
271
272 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
273 if (ctxt == NULL)
274 return -ENOMEM;
275
276 gdt = get_cpu_gdt_rw(cpu);
277
278 /*
279 * Bring up the CPU in cpu_bringup_and_idle() with the stack
280 * pointing just below where pt_regs would be if it were a normal
281 * kernel entry.
282 */
283 ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle;
284 ctxt->flags = VGCF_IN_KERNEL;
285 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
286 ctxt->user_regs.ds = __USER_DS;
287 ctxt->user_regs.es = __USER_DS;
288 ctxt->user_regs.ss = __KERNEL_DS;
289 ctxt->user_regs.cs = __KERNEL_CS;
290 ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle);
291
292 xen_copy_trap_info(ctxt->trap_ctxt);
293
294 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
295
296 gdt_mfn = arbitrary_virt_to_mfn(gdt);
297 make_lowmem_page_readonly(gdt);
298 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
299
300 ctxt->gdt_frames[0] = gdt_mfn;
301 ctxt->gdt_ents = GDT_ENTRIES;
302
303 /*
304 * Set SS:SP that Xen will use when entering guest kernel mode
305 * from guest user mode. Subsequent calls to load_sp0() can
306 * change this value.
307 */
308 ctxt->kernel_ss = __KERNEL_DS;
309 ctxt->kernel_sp = task_top_of_stack(idle);
310
311 ctxt->gs_base_kernel = per_cpu_offset(cpu);
312 ctxt->event_callback_eip =
313 (unsigned long)xen_asm_exc_xen_hypervisor_callback;
314 ctxt->failsafe_callback_eip =
315 (unsigned long)xen_failsafe_callback;
316 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
317
318 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
319 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
320 BUG();
321
322 kfree(ctxt);
323 return 0;
324 }
325
xen_pv_cpu_up(unsigned int cpu,struct task_struct * idle)326 static int xen_pv_cpu_up(unsigned int cpu, struct task_struct *idle)
327 {
328 int rc;
329
330 rc = common_cpu_up(cpu, idle);
331 if (rc)
332 return rc;
333
334 xen_setup_runstate_info(cpu);
335
336 /*
337 * PV VCPUs are always successfully taken down (see 'while' loop
338 * in xen_cpu_die()), so -EBUSY is an error.
339 */
340 rc = cpu_check_up_prepare(cpu);
341 if (rc)
342 return rc;
343
344 /* make sure interrupts start blocked */
345 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
346
347 rc = cpu_initialize_context(cpu, idle);
348 if (rc)
349 return rc;
350
351 xen_pmu_init(cpu);
352
353 rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
354 BUG_ON(rc);
355
356 while (cpu_report_state(cpu) != CPU_ONLINE)
357 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
358
359 return 0;
360 }
361
362 #ifdef CONFIG_HOTPLUG_CPU
xen_pv_cpu_disable(void)363 static int xen_pv_cpu_disable(void)
364 {
365 unsigned int cpu = smp_processor_id();
366 if (cpu == 0)
367 return -EBUSY;
368
369 cpu_disable_common();
370
371 load_cr3(swapper_pg_dir);
372 return 0;
373 }
374
xen_pv_cpu_die(unsigned int cpu)375 static void xen_pv_cpu_die(unsigned int cpu)
376 {
377 while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
378 xen_vcpu_nr(cpu), NULL)) {
379 __set_current_state(TASK_UNINTERRUPTIBLE);
380 schedule_timeout(HZ/10);
381 }
382
383 if (common_cpu_die(cpu) == 0) {
384 xen_smp_intr_free(cpu);
385 xen_uninit_lock_cpu(cpu);
386 xen_teardown_timer(cpu);
387 xen_pmu_finish(cpu);
388 }
389 }
390
xen_pv_play_dead(void)391 static void xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */
392 {
393 play_dead_common();
394 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
395 cpu_bringup();
396 /*
397 * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
398 * clears certain data that the cpu_idle loop (which called us
399 * and that we return from) expects. The only way to get that
400 * data back is to call:
401 */
402 tick_nohz_idle_enter();
403 tick_nohz_idle_stop_tick_protected();
404
405 cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
406 }
407
408 #else /* !CONFIG_HOTPLUG_CPU */
xen_pv_cpu_disable(void)409 static int xen_pv_cpu_disable(void)
410 {
411 return -ENOSYS;
412 }
413
xen_pv_cpu_die(unsigned int cpu)414 static void xen_pv_cpu_die(unsigned int cpu)
415 {
416 BUG();
417 }
418
xen_pv_play_dead(void)419 static void xen_pv_play_dead(void)
420 {
421 BUG();
422 }
423
424 #endif
stop_self(void * v)425 static void stop_self(void *v)
426 {
427 int cpu = smp_processor_id();
428
429 /* make sure we're not pinning something down */
430 load_cr3(swapper_pg_dir);
431 /* should set up a minimal gdt */
432
433 set_cpu_online(cpu, false);
434
435 HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
436 BUG();
437 }
438
xen_pv_stop_other_cpus(int wait)439 static void xen_pv_stop_other_cpus(int wait)
440 {
441 smp_call_function(stop_self, NULL, wait);
442 }
443
xen_irq_work_interrupt(int irq,void * dev_id)444 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
445 {
446 irq_enter();
447 irq_work_run();
448 inc_irq_stat(apic_irq_work_irqs);
449 irq_exit();
450
451 return IRQ_HANDLED;
452 }
453
454 static const struct smp_ops xen_smp_ops __initconst = {
455 .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
456 .smp_prepare_cpus = xen_pv_smp_prepare_cpus,
457 .smp_cpus_done = xen_smp_cpus_done,
458
459 .cpu_up = xen_pv_cpu_up,
460 .cpu_die = xen_pv_cpu_die,
461 .cpu_disable = xen_pv_cpu_disable,
462 .play_dead = xen_pv_play_dead,
463
464 .stop_other_cpus = xen_pv_stop_other_cpus,
465 .smp_send_reschedule = xen_smp_send_reschedule,
466
467 .send_call_func_ipi = xen_smp_send_call_function_ipi,
468 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
469 };
470
xen_smp_init(void)471 void __init xen_smp_init(void)
472 {
473 smp_ops = xen_smp_ops;
474
475 /* Avoid searching for BIOS MP tables */
476 x86_init.mpparse.find_smp_config = x86_init_noop;
477 x86_init.mpparse.get_smp_config = _get_smp_config;
478 }
479