4 * This file implements the Xen versions of smp_ops. SMP under Xen is
5 * very straightforward. Bringing a CPU up is simply a matter of
6 * loading its initial context and setting it running.
8 * IPIs are handled through the Xen event mechanism.
10 * Because virtual CPUs can be scheduled onto any real CPU, there's no
11 * useful topology information for the kernel to make use of. As a
12 * result, all CPUs are treated as if they're single-core and
15 * This does not handle HOTPLUG_CPU yet.
17 #include <linux/sched.h>
18 #include <linux/err.h>
19 #include <linux/smp.h>
21 #include <asm/paravirt.h>
23 #include <asm/pgtable.h>
26 #include <xen/interface/xen.h>
27 #include <xen/interface/vcpu.h>
29 #include <asm/xen/interface.h>
30 #include <asm/xen/hypercall.h>
33 #include <xen/events.h>
38 static cpumask_t xen_cpu_initialized_map;
39 static DEFINE_PER_CPU(int, resched_irq);
40 static DEFINE_PER_CPU(int, callfunc_irq);
43 * Structure and data for smp_call_function(). This is designed to minimise
44 * static memory requirements. It also looks cleaner.
46 static DEFINE_SPINLOCK(call_lock);
48 struct call_data_struct {
49 void (*func) (void *info);
56 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
58 static struct call_data_struct *call_data;
61 * Reschedule call back. Nothing to do,
62 * all the work is done automatically when
63 * we return from the interrupt.
65 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
70 static __cpuinit void cpu_bringup_and_idle(void)
72 int cpu = smp_processor_id();
75 xen_enable_sysenter();
78 per_cpu(cpu_state, cpu) = CPU_ONLINE;
80 xen_setup_cpu_clockevents();
82 /* We can take interrupts now: we're officially "up". */
85 wmb(); /* make sure everything is out */
89 static int xen_smp_intr_init(unsigned int cpu)
92 const char *resched_name, *callfunc_name;
94 per_cpu(resched_irq, cpu) = per_cpu(callfunc_irq, cpu) = -1;
96 resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
97 rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
99 xen_reschedule_interrupt,
100 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
105 per_cpu(resched_irq, cpu) = rc;
107 callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
108 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
110 xen_call_function_interrupt,
111 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
116 per_cpu(callfunc_irq, cpu) = rc;
121 if (per_cpu(resched_irq, cpu) >= 0)
122 unbind_from_irqhandler(per_cpu(resched_irq, cpu), NULL);
123 if (per_cpu(callfunc_irq, cpu) >= 0)
124 unbind_from_irqhandler(per_cpu(callfunc_irq, cpu), NULL);
128 void __init xen_fill_possible_map(void)
132 for (i = 0; i < NR_CPUS; i++) {
133 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
135 cpu_set(i, cpu_possible_map);
139 void __init xen_smp_prepare_boot_cpu(void)
143 BUG_ON(smp_processor_id() != 0);
144 native_smp_prepare_boot_cpu();
146 /* We've switched to the "real" per-cpu gdt, so make sure the
147 old memory can be recycled */
148 make_lowmem_page_readwrite(&per_cpu__gdt_page);
150 for_each_possible_cpu(cpu) {
151 cpus_clear(per_cpu(cpu_sibling_map, cpu));
153 * cpu_core_map lives in a per cpu area that is cleared
154 * when the per cpu array is allocated.
156 * cpus_clear(per_cpu(cpu_core_map, cpu));
160 xen_setup_vcpu_info_placement();
163 void __init xen_smp_prepare_cpus(unsigned int max_cpus)
167 for_each_possible_cpu(cpu) {
168 cpus_clear(per_cpu(cpu_sibling_map, cpu));
170 * cpu_core_ map will be zeroed when the per
171 * cpu area is allocated.
173 * cpus_clear(per_cpu(cpu_core_map, cpu));
177 smp_store_cpu_info(0);
178 set_cpu_sibling_map(0);
180 if (xen_smp_intr_init(0))
183 xen_cpu_initialized_map = cpumask_of_cpu(0);
185 /* Restrict the possible_map according to max_cpus. */
186 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
187 for (cpu = NR_CPUS-1; !cpu_isset(cpu, cpu_possible_map); cpu--)
189 cpu_clear(cpu, cpu_possible_map);
192 for_each_possible_cpu (cpu) {
193 struct task_struct *idle;
198 idle = fork_idle(cpu);
200 panic("failed fork for CPU %d", cpu);
202 cpu_set(cpu, cpu_present_map);
205 //init_xenbus_allowed_cpumask();
209 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
211 struct vcpu_guest_context *ctxt;
212 struct gdt_page *gdt = &per_cpu(gdt_page, cpu);
214 if (cpu_test_and_set(cpu, xen_cpu_initialized_map))
217 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
221 ctxt->flags = VGCF_IN_KERNEL;
222 ctxt->user_regs.ds = __USER_DS;
223 ctxt->user_regs.es = __USER_DS;
224 ctxt->user_regs.fs = __KERNEL_PERCPU;
225 ctxt->user_regs.gs = 0;
226 ctxt->user_regs.ss = __KERNEL_DS;
227 ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
228 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
230 memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
232 xen_copy_trap_info(ctxt->trap_ctxt);
236 BUG_ON((unsigned long)gdt->gdt & ~PAGE_MASK);
237 make_lowmem_page_readonly(gdt->gdt);
239 ctxt->gdt_frames[0] = virt_to_mfn(gdt->gdt);
240 ctxt->gdt_ents = ARRAY_SIZE(gdt->gdt);
242 ctxt->user_regs.cs = __KERNEL_CS;
243 ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
245 ctxt->kernel_ss = __KERNEL_DS;
246 ctxt->kernel_sp = idle->thread.sp0;
248 ctxt->event_callback_cs = __KERNEL_CS;
249 ctxt->event_callback_eip = (unsigned long)xen_hypervisor_callback;
250 ctxt->failsafe_callback_cs = __KERNEL_CS;
251 ctxt->failsafe_callback_eip = (unsigned long)xen_failsafe_callback;
253 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
254 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
256 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
263 int __cpuinit xen_cpu_up(unsigned int cpu)
265 struct task_struct *idle = idle_task(cpu);
269 rc = cpu_up_check(cpu);
275 per_cpu(current_task, cpu) = idle;
277 xen_setup_timer(cpu);
279 /* make sure interrupts start blocked */
280 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
282 rc = cpu_initialize_context(cpu, idle);
286 if (num_online_cpus() == 1)
287 alternatives_smp_switch(1);
289 rc = xen_smp_intr_init(cpu);
293 smp_store_cpu_info(cpu);
294 set_cpu_sibling_map(cpu);
295 /* This must be done before setting cpu_online_map */
298 cpu_set(cpu, cpu_online_map);
300 rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
306 void xen_smp_cpus_done(unsigned int max_cpus)
310 static void stop_self(void *v)
312 int cpu = smp_processor_id();
314 /* make sure we're not pinning something down */
315 load_cr3(swapper_pg_dir);
316 /* should set up a minimal gdt */
318 HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
322 void xen_smp_send_stop(void)
324 smp_call_function(stop_self, NULL, 0, 0);
327 void xen_smp_send_reschedule(int cpu)
329 xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
333 static void xen_send_IPI_mask(cpumask_t mask, enum ipi_vector vector)
337 cpus_and(mask, mask, cpu_online_map);
339 for_each_cpu_mask(cpu, mask)
340 xen_send_IPI_one(cpu, vector);
343 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
345 void (*func) (void *info) = call_data->func;
346 void *info = call_data->info;
347 int wait = call_data->wait;
350 * Notify initiating CPU that I've grabbed the data and am
351 * about to execute the function
354 atomic_inc(&call_data->started);
356 * At this point the info structure may be out of scope unless wait==1
360 __get_cpu_var(irq_stat).irq_call_count++;
364 mb(); /* commit everything before setting finished */
365 atomic_inc(&call_data->finished);
371 int xen_smp_call_function_mask(cpumask_t mask, void (*func)(void *),
372 void *info, int wait)
374 struct call_data_struct data;
378 /* Holding any lock stops cpus from going down. */
379 spin_lock(&call_lock);
381 cpu_clear(smp_processor_id(), mask);
383 cpus = cpus_weight(mask);
385 spin_unlock(&call_lock);
389 /* Can deadlock when called with interrupts disabled */
390 WARN_ON(irqs_disabled());
394 atomic_set(&data.started, 0);
397 atomic_set(&data.finished, 0);
400 mb(); /* write everything before IPI */
402 /* Send a message to other CPUs and wait for them to respond */
403 xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
405 /* Make sure other vcpus get a chance to run if they need to. */
407 for_each_cpu_mask(cpu, mask)
408 if (xen_vcpu_stolen(cpu))
412 HYPERVISOR_sched_op(SCHEDOP_yield, 0);
414 /* Wait for response */
415 while (atomic_read(&data.started) != cpus ||
416 (wait && atomic_read(&data.finished) != cpus))
419 spin_unlock(&call_lock);