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 | #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 | |
60 | static void cpu_bringup(void) |
61 | { |
62 | int cpu; |
63 | |
64 | cr4_init(); |
65 | cpuhp_ap_sync_alive(); |
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(id: cpu); |
77 | set_cpu_sibling_map(cpu); |
78 | |
79 | speculative_store_bypass_ht_init(); |
80 | |
81 | xen_setup_cpu_clockevents(); |
82 | |
83 | notify_cpu_starting(cpu); |
84 | |
85 | set_cpu_online(cpu, online: true); |
86 | |
87 | smp_mb(); |
88 | |
89 | /* We can take interrupts now: we're officially "up". */ |
90 | local_irq_enable(); |
91 | } |
92 | |
93 | asmlinkage __visible void cpu_bringup_and_idle(void) |
94 | { |
95 | cpu_bringup(); |
96 | cpu_startup_entry(state: CPUHP_AP_ONLINE_IDLE); |
97 | } |
98 | |
99 | void xen_smp_intr_free_pv(unsigned int cpu) |
100 | { |
101 | kfree(per_cpu(xen_irq_work, cpu).name); |
102 | per_cpu(xen_irq_work, cpu).name = NULL; |
103 | if (per_cpu(xen_irq_work, cpu).irq >= 0) { |
104 | unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL); |
105 | per_cpu(xen_irq_work, cpu).irq = -1; |
106 | } |
107 | |
108 | kfree(per_cpu(xen_pmu_irq, cpu).name); |
109 | per_cpu(xen_pmu_irq, cpu).name = NULL; |
110 | if (per_cpu(xen_pmu_irq, cpu).irq >= 0) { |
111 | unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL); |
112 | per_cpu(xen_pmu_irq, cpu).irq = -1; |
113 | } |
114 | } |
115 | |
116 | int xen_smp_intr_init_pv(unsigned int cpu) |
117 | { |
118 | int rc; |
119 | char *callfunc_name, *pmu_name; |
120 | |
121 | callfunc_name = kasprintf(GFP_KERNEL, fmt: "irqwork%d" , cpu); |
122 | per_cpu(xen_irq_work, cpu).name = callfunc_name; |
123 | rc = bind_ipi_to_irqhandler(ipi: XEN_IRQ_WORK_VECTOR, |
124 | cpu, |
125 | handler: xen_irq_work_interrupt, |
126 | IRQF_PERCPU|IRQF_NOBALANCING, |
127 | devname: callfunc_name, |
128 | NULL); |
129 | if (rc < 0) |
130 | goto fail; |
131 | per_cpu(xen_irq_work, cpu).irq = rc; |
132 | |
133 | if (is_xen_pmu) { |
134 | pmu_name = kasprintf(GFP_KERNEL, fmt: "pmu%d" , cpu); |
135 | per_cpu(xen_pmu_irq, cpu).name = pmu_name; |
136 | rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu, |
137 | handler: xen_pmu_irq_handler, |
138 | IRQF_PERCPU|IRQF_NOBALANCING, |
139 | devname: pmu_name, NULL); |
140 | if (rc < 0) |
141 | goto fail; |
142 | per_cpu(xen_pmu_irq, cpu).irq = rc; |
143 | } |
144 | |
145 | return 0; |
146 | |
147 | fail: |
148 | xen_smp_intr_free_pv(cpu); |
149 | return rc; |
150 | } |
151 | |
152 | static void __init xen_pv_smp_config(void) |
153 | { |
154 | u32 apicid = 0; |
155 | int i; |
156 | |
157 | topology_register_boot_apic(apic_id: apicid++); |
158 | |
159 | for (i = 1; i < nr_cpu_ids; i++) |
160 | topology_register_apic(apic_id: apicid++, CPU_ACPIID_INVALID, present: true); |
161 | |
162 | /* Pretend to be a proper enumerated system */ |
163 | smp_found_config = 1; |
164 | } |
165 | |
166 | static void __init xen_pv_smp_prepare_boot_cpu(void) |
167 | { |
168 | BUG_ON(smp_processor_id() != 0); |
169 | native_smp_prepare_boot_cpu(); |
170 | |
171 | if (!xen_feature(XENFEAT_writable_page_tables)) |
172 | /* We've switched to the "real" per-cpu gdt, so make |
173 | * sure the old memory can be recycled. */ |
174 | make_lowmem_page_readwrite(vaddr: xen_initial_gdt); |
175 | |
176 | xen_setup_vcpu_info_placement(); |
177 | |
178 | /* |
179 | * The alternative logic (which patches the unlock/lock) runs before |
180 | * the smp bootup up code is activated. Hence we need to set this up |
181 | * the core kernel is being patched. Otherwise we will have only |
182 | * modules patched but not core code. |
183 | */ |
184 | xen_init_spinlocks(); |
185 | } |
186 | |
187 | static void __init xen_pv_smp_prepare_cpus(unsigned int max_cpus) |
188 | { |
189 | unsigned cpu; |
190 | |
191 | if (ioapic_is_disabled) { |
192 | char *m = (max_cpus == 0) ? |
193 | "The nosmp parameter is incompatible with Xen; " \ |
194 | "use Xen dom0_max_vcpus=1 parameter" : |
195 | "The noapic parameter is incompatible with Xen" ; |
196 | |
197 | xen_raw_printk(fmt: m); |
198 | panic(fmt: m); |
199 | } |
200 | xen_init_lock_cpu(cpu: 0); |
201 | |
202 | smp_prepare_cpus_common(); |
203 | |
204 | speculative_store_bypass_ht_init(); |
205 | |
206 | xen_pmu_init(cpu: 0); |
207 | |
208 | if (xen_smp_intr_init(cpu: 0) || xen_smp_intr_init_pv(cpu: 0)) |
209 | BUG(); |
210 | |
211 | if (!alloc_cpumask_var(mask: &xen_cpu_initialized_map, GFP_KERNEL)) |
212 | panic(fmt: "could not allocate xen_cpu_initialized_map\n" ); |
213 | |
214 | cpumask_copy(dstp: xen_cpu_initialized_map, cpumask_of(0)); |
215 | |
216 | /* Restrict the possible_map according to max_cpus. */ |
217 | while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) { |
218 | for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--) |
219 | continue; |
220 | set_cpu_possible(cpu, possible: false); |
221 | } |
222 | |
223 | for_each_possible_cpu(cpu) |
224 | set_cpu_present(cpu, present: true); |
225 | } |
226 | |
227 | static int |
228 | cpu_initialize_context(unsigned int cpu, struct task_struct *idle) |
229 | { |
230 | struct vcpu_guest_context *ctxt; |
231 | struct desc_struct *gdt; |
232 | unsigned long gdt_mfn; |
233 | |
234 | if (cpumask_test_and_set_cpu(cpu, cpumask: xen_cpu_initialized_map)) |
235 | return 0; |
236 | |
237 | ctxt = kzalloc(size: sizeof(*ctxt), GFP_KERNEL); |
238 | if (ctxt == NULL) { |
239 | cpumask_clear_cpu(cpu, dstp: xen_cpu_initialized_map); |
240 | return -ENOMEM; |
241 | } |
242 | |
243 | gdt = get_cpu_gdt_rw(cpu); |
244 | |
245 | /* |
246 | * Bring up the CPU in cpu_bringup_and_idle() with the stack |
247 | * pointing just below where pt_regs would be if it were a normal |
248 | * kernel entry. |
249 | */ |
250 | ctxt->user_regs.eip = (unsigned long)asm_cpu_bringup_and_idle; |
251 | ctxt->flags = VGCF_IN_KERNEL; |
252 | ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */ |
253 | ctxt->user_regs.ds = __USER_DS; |
254 | ctxt->user_regs.es = __USER_DS; |
255 | ctxt->user_regs.ss = __KERNEL_DS; |
256 | ctxt->user_regs.cs = __KERNEL_CS; |
257 | ctxt->user_regs.esp = (unsigned long)task_pt_regs(idle); |
258 | |
259 | xen_copy_trap_info(traps: ctxt->trap_ctxt); |
260 | |
261 | BUG_ON((unsigned long)gdt & ~PAGE_MASK); |
262 | |
263 | gdt_mfn = arbitrary_virt_to_mfn(vaddr: gdt); |
264 | make_lowmem_page_readonly(vaddr: gdt); |
265 | make_lowmem_page_readonly(mfn_to_virt(gdt_mfn)); |
266 | |
267 | ctxt->gdt_frames[0] = gdt_mfn; |
268 | ctxt->gdt_ents = GDT_ENTRIES; |
269 | |
270 | /* |
271 | * Set SS:SP that Xen will use when entering guest kernel mode |
272 | * from guest user mode. Subsequent calls to load_sp0() can |
273 | * change this value. |
274 | */ |
275 | ctxt->kernel_ss = __KERNEL_DS; |
276 | ctxt->kernel_sp = task_top_of_stack(idle); |
277 | |
278 | ctxt->gs_base_kernel = per_cpu_offset(cpu); |
279 | ctxt->event_callback_eip = |
280 | (unsigned long)xen_asm_exc_xen_hypervisor_callback; |
281 | ctxt->failsafe_callback_eip = |
282 | (unsigned long)xen_failsafe_callback; |
283 | per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir); |
284 | |
285 | ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir)); |
286 | if (HYPERVISOR_vcpu_op(VCPUOP_initialise, vcpuid: xen_vcpu_nr(cpu), extra_args: ctxt)) |
287 | BUG(); |
288 | |
289 | kfree(objp: ctxt); |
290 | return 0; |
291 | } |
292 | |
293 | static int xen_pv_kick_ap(unsigned int cpu, struct task_struct *idle) |
294 | { |
295 | int rc; |
296 | |
297 | rc = common_cpu_up(cpunum: cpu, tidle: idle); |
298 | if (rc) |
299 | return rc; |
300 | |
301 | xen_setup_runstate_info(cpu); |
302 | |
303 | /* make sure interrupts start blocked */ |
304 | per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1; |
305 | |
306 | rc = cpu_initialize_context(cpu, idle); |
307 | if (rc) |
308 | return rc; |
309 | |
310 | xen_pmu_init(cpu); |
311 | |
312 | /* |
313 | * Why is this a BUG? If the hypercall fails then everything can be |
314 | * rolled back, no? |
315 | */ |
316 | BUG_ON(HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL)); |
317 | |
318 | return 0; |
319 | } |
320 | |
321 | static void xen_pv_poll_sync_state(void) |
322 | { |
323 | HYPERVISOR_sched_op(SCHEDOP_yield, NULL); |
324 | } |
325 | |
326 | #ifdef CONFIG_HOTPLUG_CPU |
327 | static int xen_pv_cpu_disable(void) |
328 | { |
329 | unsigned int cpu = smp_processor_id(); |
330 | if (cpu == 0) |
331 | return -EBUSY; |
332 | |
333 | cpu_disable_common(); |
334 | |
335 | load_cr3(swapper_pg_dir); |
336 | return 0; |
337 | } |
338 | |
339 | static void xen_pv_cpu_die(unsigned int cpu) |
340 | { |
341 | while (HYPERVISOR_vcpu_op(VCPUOP_is_up, vcpuid: xen_vcpu_nr(cpu), NULL)) { |
342 | __set_current_state(TASK_UNINTERRUPTIBLE); |
343 | schedule_timeout(HZ/10); |
344 | } |
345 | } |
346 | |
347 | static void xen_pv_cleanup_dead_cpu(unsigned int cpu) |
348 | { |
349 | xen_smp_intr_free(cpu); |
350 | xen_uninit_lock_cpu(cpu); |
351 | xen_teardown_timer(cpu); |
352 | xen_pmu_finish(cpu); |
353 | } |
354 | |
355 | static void __noreturn xen_pv_play_dead(void) /* used only with HOTPLUG_CPU */ |
356 | { |
357 | play_dead_common(); |
358 | HYPERVISOR_vcpu_op(VCPUOP_down, vcpuid: xen_vcpu_nr(smp_processor_id()), NULL); |
359 | xen_cpu_bringup_again(stack: (unsigned long)task_pt_regs(current)); |
360 | BUG(); |
361 | } |
362 | |
363 | #else /* !CONFIG_HOTPLUG_CPU */ |
364 | static int xen_pv_cpu_disable(void) |
365 | { |
366 | return -ENOSYS; |
367 | } |
368 | |
369 | static void xen_pv_cpu_die(unsigned int cpu) |
370 | { |
371 | BUG(); |
372 | } |
373 | |
374 | static void xen_pv_cleanup_dead_cpu(unsigned int cpu) |
375 | { |
376 | BUG(); |
377 | } |
378 | |
379 | static void __noreturn xen_pv_play_dead(void) |
380 | { |
381 | BUG(); |
382 | } |
383 | |
384 | #endif |
385 | static void stop_self(void *v) |
386 | { |
387 | int cpu = smp_processor_id(); |
388 | |
389 | /* make sure we're not pinning something down */ |
390 | load_cr3(swapper_pg_dir); |
391 | /* should set up a minimal gdt */ |
392 | |
393 | set_cpu_online(cpu, online: false); |
394 | |
395 | HYPERVISOR_vcpu_op(VCPUOP_down, vcpuid: xen_vcpu_nr(cpu), NULL); |
396 | BUG(); |
397 | } |
398 | |
399 | static void xen_pv_stop_other_cpus(int wait) |
400 | { |
401 | smp_call_function(func: stop_self, NULL, wait); |
402 | } |
403 | |
404 | static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id) |
405 | { |
406 | irq_work_run(); |
407 | inc_irq_stat(apic_irq_work_irqs); |
408 | |
409 | return IRQ_HANDLED; |
410 | } |
411 | |
412 | void __init xen_smp_count_cpus(void) |
413 | { |
414 | unsigned int cpus; |
415 | |
416 | for (cpus = 0; cpus < nr_cpu_ids; cpus++) { |
417 | if (HYPERVISOR_vcpu_op(VCPUOP_is_up, vcpuid: cpus, NULL) < 0) |
418 | break; |
419 | } |
420 | |
421 | pr_info("Xen PV: Detected %u vCPUS\n" , cpus); |
422 | if (cpus < nr_cpu_ids) |
423 | set_nr_cpu_ids(cpus); |
424 | } |
425 | |
426 | static const struct smp_ops xen_smp_ops __initconst = { |
427 | .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu, |
428 | .smp_prepare_cpus = xen_pv_smp_prepare_cpus, |
429 | .smp_cpus_done = xen_smp_cpus_done, |
430 | |
431 | .kick_ap_alive = xen_pv_kick_ap, |
432 | .cpu_die = xen_pv_cpu_die, |
433 | .cleanup_dead_cpu = xen_pv_cleanup_dead_cpu, |
434 | .poll_sync_state = xen_pv_poll_sync_state, |
435 | .cpu_disable = xen_pv_cpu_disable, |
436 | .play_dead = xen_pv_play_dead, |
437 | |
438 | .stop_other_cpus = xen_pv_stop_other_cpus, |
439 | .smp_send_reschedule = xen_smp_send_reschedule, |
440 | |
441 | .send_call_func_ipi = xen_smp_send_call_function_ipi, |
442 | .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi, |
443 | }; |
444 | |
445 | void __init xen_smp_init(void) |
446 | { |
447 | smp_ops = xen_smp_ops; |
448 | |
449 | /* Avoid searching for BIOS MP tables */ |
450 | x86_init.mpparse.find_mptable = x86_init_noop; |
451 | x86_init.mpparse.early_parse_smp_cfg = x86_init_noop; |
452 | |
453 | /* XEN/PV Dom0 has halfways sane topology information via CPUID/MADT */ |
454 | if (xen_initial_domain()) |
455 | x86_init.mpparse.parse_smp_cfg = x86_init_noop; |
456 | else |
457 | x86_init.mpparse.parse_smp_cfg = xen_pv_smp_config; |
458 | } |
459 | |