Jeremy Fitzhardinge | f87e4ca | 2007-07-17 18:37:06 -0700 | [diff] [blame] | 1 | /* |
| 2 | * Xen SMP support |
| 3 | * |
| 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. |
| 7 | * |
| 8 | * IPIs are handled through the Xen event mechanism. |
| 9 | * |
| 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 |
| 13 | * single-threaded. |
| 14 | * |
| 15 | * This does not handle HOTPLUG_CPU yet. |
| 16 | */ |
| 17 | #include <linux/sched.h> |
| 18 | #include <linux/err.h> |
| 19 | #include <linux/smp.h> |
| 20 | |
| 21 | #include <asm/paravirt.h> |
| 22 | #include <asm/desc.h> |
| 23 | #include <asm/pgtable.h> |
| 24 | #include <asm/cpu.h> |
| 25 | |
| 26 | #include <xen/interface/xen.h> |
| 27 | #include <xen/interface/vcpu.h> |
| 28 | |
| 29 | #include <asm/xen/interface.h> |
| 30 | #include <asm/xen/hypercall.h> |
| 31 | |
| 32 | #include <xen/page.h> |
| 33 | #include <xen/events.h> |
| 34 | |
| 35 | #include "xen-ops.h" |
| 36 | #include "mmu.h" |
| 37 | |
| 38 | static cpumask_t cpu_initialized_map; |
| 39 | static DEFINE_PER_CPU(int, resched_irq); |
| 40 | static DEFINE_PER_CPU(int, callfunc_irq); |
| 41 | |
| 42 | /* |
| 43 | * Structure and data for smp_call_function(). This is designed to minimise |
| 44 | * static memory requirements. It also looks cleaner. |
| 45 | */ |
| 46 | static DEFINE_SPINLOCK(call_lock); |
| 47 | |
| 48 | struct call_data_struct { |
| 49 | void (*func) (void *info); |
| 50 | void *info; |
| 51 | atomic_t started; |
| 52 | atomic_t finished; |
| 53 | int wait; |
| 54 | }; |
| 55 | |
| 56 | static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id); |
| 57 | |
| 58 | static struct call_data_struct *call_data; |
| 59 | |
| 60 | /* |
| 61 | * Reschedule call back. Nothing to do, |
| 62 | * all the work is done automatically when |
| 63 | * we return from the interrupt. |
| 64 | */ |
| 65 | static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id) |
| 66 | { |
| 67 | return IRQ_HANDLED; |
| 68 | } |
| 69 | |
| 70 | static __cpuinit void cpu_bringup_and_idle(void) |
| 71 | { |
| 72 | int cpu = smp_processor_id(); |
| 73 | |
| 74 | cpu_init(); |
| 75 | |
| 76 | preempt_disable(); |
| 77 | per_cpu(cpu_state, cpu) = CPU_ONLINE; |
| 78 | |
| 79 | xen_setup_cpu_clockevents(); |
| 80 | |
| 81 | /* We can take interrupts now: we're officially "up". */ |
| 82 | local_irq_enable(); |
| 83 | |
| 84 | wmb(); /* make sure everything is out */ |
| 85 | cpu_idle(); |
| 86 | } |
| 87 | |
| 88 | static int xen_smp_intr_init(unsigned int cpu) |
| 89 | { |
| 90 | int rc; |
| 91 | const char *resched_name, *callfunc_name; |
| 92 | |
| 93 | per_cpu(resched_irq, cpu) = per_cpu(callfunc_irq, cpu) = -1; |
| 94 | |
| 95 | resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu); |
| 96 | rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR, |
| 97 | cpu, |
| 98 | xen_reschedule_interrupt, |
| 99 | IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING, |
| 100 | resched_name, |
| 101 | NULL); |
| 102 | if (rc < 0) |
| 103 | goto fail; |
| 104 | per_cpu(resched_irq, cpu) = rc; |
| 105 | |
| 106 | callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu); |
| 107 | rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR, |
| 108 | cpu, |
| 109 | xen_call_function_interrupt, |
| 110 | IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING, |
| 111 | callfunc_name, |
| 112 | NULL); |
| 113 | if (rc < 0) |
| 114 | goto fail; |
| 115 | per_cpu(callfunc_irq, cpu) = rc; |
| 116 | |
| 117 | return 0; |
| 118 | |
| 119 | fail: |
| 120 | if (per_cpu(resched_irq, cpu) >= 0) |
| 121 | unbind_from_irqhandler(per_cpu(resched_irq, cpu), NULL); |
| 122 | if (per_cpu(callfunc_irq, cpu) >= 0) |
| 123 | unbind_from_irqhandler(per_cpu(callfunc_irq, cpu), NULL); |
| 124 | return rc; |
| 125 | } |
| 126 | |
| 127 | void __init xen_fill_possible_map(void) |
| 128 | { |
| 129 | int i, rc; |
| 130 | |
| 131 | for (i = 0; i < NR_CPUS; i++) { |
| 132 | rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL); |
| 133 | if (rc >= 0) |
| 134 | cpu_set(i, cpu_possible_map); |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | void __init xen_smp_prepare_boot_cpu(void) |
| 139 | { |
| 140 | int cpu; |
| 141 | |
| 142 | BUG_ON(smp_processor_id() != 0); |
| 143 | native_smp_prepare_boot_cpu(); |
| 144 | |
Jeremy Fitzhardinge | f87e4ca | 2007-07-17 18:37:06 -0700 | [diff] [blame] | 145 | /* We've switched to the "real" per-cpu gdt, so make sure the |
| 146 | old memory can be recycled */ |
| 147 | make_lowmem_page_readwrite(&per_cpu__gdt_page); |
| 148 | |
| 149 | for (cpu = 0; cpu < NR_CPUS; cpu++) { |
| 150 | cpus_clear(cpu_sibling_map[cpu]); |
Mike Travis | 0835761 | 2007-10-16 01:24:04 -0700 | [diff] [blame^] | 151 | /* |
| 152 | * cpu_core_map lives in a per cpu area that is cleared |
| 153 | * when the per cpu array is allocated. |
| 154 | * |
| 155 | * cpus_clear(per_cpu(cpu_core_map, cpu)); |
| 156 | */ |
Jeremy Fitzhardinge | f87e4ca | 2007-07-17 18:37:06 -0700 | [diff] [blame] | 157 | } |
Jeremy Fitzhardinge | 60223a3 | 2007-07-17 18:37:07 -0700 | [diff] [blame] | 158 | |
| 159 | xen_setup_vcpu_info_placement(); |
Jeremy Fitzhardinge | f87e4ca | 2007-07-17 18:37:06 -0700 | [diff] [blame] | 160 | } |
| 161 | |
| 162 | void __init xen_smp_prepare_cpus(unsigned int max_cpus) |
| 163 | { |
| 164 | unsigned cpu; |
| 165 | |
| 166 | for (cpu = 0; cpu < NR_CPUS; cpu++) { |
| 167 | cpus_clear(cpu_sibling_map[cpu]); |
Mike Travis | 0835761 | 2007-10-16 01:24:04 -0700 | [diff] [blame^] | 168 | /* |
| 169 | * cpu_core_ map will be zeroed when the per |
| 170 | * cpu area is allocated. |
| 171 | * |
| 172 | * cpus_clear(per_cpu(cpu_core_map, cpu)); |
| 173 | */ |
Jeremy Fitzhardinge | f87e4ca | 2007-07-17 18:37:06 -0700 | [diff] [blame] | 174 | } |
| 175 | |
| 176 | smp_store_cpu_info(0); |
| 177 | set_cpu_sibling_map(0); |
| 178 | |
| 179 | if (xen_smp_intr_init(0)) |
| 180 | BUG(); |
| 181 | |
| 182 | cpu_initialized_map = cpumask_of_cpu(0); |
| 183 | |
| 184 | /* Restrict the possible_map according to max_cpus. */ |
| 185 | while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) { |
| 186 | for (cpu = NR_CPUS-1; !cpu_isset(cpu, cpu_possible_map); cpu--) |
| 187 | continue; |
| 188 | cpu_clear(cpu, cpu_possible_map); |
| 189 | } |
| 190 | |
| 191 | for_each_possible_cpu (cpu) { |
| 192 | struct task_struct *idle; |
| 193 | |
| 194 | if (cpu == 0) |
| 195 | continue; |
| 196 | |
| 197 | idle = fork_idle(cpu); |
| 198 | if (IS_ERR(idle)) |
| 199 | panic("failed fork for CPU %d", cpu); |
| 200 | |
| 201 | cpu_set(cpu, cpu_present_map); |
| 202 | } |
| 203 | |
| 204 | //init_xenbus_allowed_cpumask(); |
| 205 | } |
| 206 | |
| 207 | static __cpuinit int |
| 208 | cpu_initialize_context(unsigned int cpu, struct task_struct *idle) |
| 209 | { |
| 210 | struct vcpu_guest_context *ctxt; |
| 211 | struct gdt_page *gdt = &per_cpu(gdt_page, cpu); |
| 212 | |
| 213 | if (cpu_test_and_set(cpu, cpu_initialized_map)) |
| 214 | return 0; |
| 215 | |
| 216 | ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL); |
| 217 | if (ctxt == NULL) |
| 218 | return -ENOMEM; |
| 219 | |
| 220 | ctxt->flags = VGCF_IN_KERNEL; |
| 221 | ctxt->user_regs.ds = __USER_DS; |
| 222 | ctxt->user_regs.es = __USER_DS; |
| 223 | ctxt->user_regs.fs = __KERNEL_PERCPU; |
| 224 | ctxt->user_regs.gs = 0; |
| 225 | ctxt->user_regs.ss = __KERNEL_DS; |
| 226 | ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle; |
| 227 | ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */ |
| 228 | |
| 229 | memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt)); |
| 230 | |
| 231 | xen_copy_trap_info(ctxt->trap_ctxt); |
| 232 | |
| 233 | ctxt->ldt_ents = 0; |
| 234 | |
| 235 | BUG_ON((unsigned long)gdt->gdt & ~PAGE_MASK); |
| 236 | make_lowmem_page_readonly(gdt->gdt); |
| 237 | |
| 238 | ctxt->gdt_frames[0] = virt_to_mfn(gdt->gdt); |
| 239 | ctxt->gdt_ents = ARRAY_SIZE(gdt->gdt); |
| 240 | |
| 241 | ctxt->user_regs.cs = __KERNEL_CS; |
| 242 | ctxt->user_regs.esp = idle->thread.esp0 - sizeof(struct pt_regs); |
| 243 | |
| 244 | ctxt->kernel_ss = __KERNEL_DS; |
| 245 | ctxt->kernel_sp = idle->thread.esp0; |
| 246 | |
| 247 | ctxt->event_callback_cs = __KERNEL_CS; |
| 248 | ctxt->event_callback_eip = (unsigned long)xen_hypervisor_callback; |
| 249 | ctxt->failsafe_callback_cs = __KERNEL_CS; |
| 250 | ctxt->failsafe_callback_eip = (unsigned long)xen_failsafe_callback; |
| 251 | |
| 252 | per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir); |
| 253 | ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir)); |
| 254 | |
| 255 | if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt)) |
| 256 | BUG(); |
| 257 | |
| 258 | kfree(ctxt); |
| 259 | return 0; |
| 260 | } |
| 261 | |
| 262 | int __cpuinit xen_cpu_up(unsigned int cpu) |
| 263 | { |
| 264 | struct task_struct *idle = idle_task(cpu); |
| 265 | int rc; |
| 266 | |
| 267 | #if 0 |
| 268 | rc = cpu_up_check(cpu); |
| 269 | if (rc) |
| 270 | return rc; |
| 271 | #endif |
| 272 | |
| 273 | init_gdt(cpu); |
| 274 | per_cpu(current_task, cpu) = idle; |
Jeremy Fitzhardinge | f87e4ca | 2007-07-17 18:37:06 -0700 | [diff] [blame] | 275 | irq_ctx_init(cpu); |
| 276 | xen_setup_timer(cpu); |
| 277 | |
| 278 | /* make sure interrupts start blocked */ |
| 279 | per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1; |
| 280 | |
| 281 | rc = cpu_initialize_context(cpu, idle); |
| 282 | if (rc) |
| 283 | return rc; |
| 284 | |
| 285 | if (num_online_cpus() == 1) |
| 286 | alternatives_smp_switch(1); |
| 287 | |
| 288 | rc = xen_smp_intr_init(cpu); |
| 289 | if (rc) |
| 290 | return rc; |
| 291 | |
| 292 | smp_store_cpu_info(cpu); |
| 293 | set_cpu_sibling_map(cpu); |
| 294 | /* This must be done before setting cpu_online_map */ |
| 295 | wmb(); |
| 296 | |
| 297 | cpu_set(cpu, cpu_online_map); |
| 298 | |
| 299 | rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL); |
| 300 | BUG_ON(rc); |
| 301 | |
| 302 | return 0; |
| 303 | } |
| 304 | |
| 305 | void xen_smp_cpus_done(unsigned int max_cpus) |
| 306 | { |
| 307 | } |
| 308 | |
| 309 | static void stop_self(void *v) |
| 310 | { |
| 311 | int cpu = smp_processor_id(); |
| 312 | |
| 313 | /* make sure we're not pinning something down */ |
| 314 | load_cr3(swapper_pg_dir); |
| 315 | /* should set up a minimal gdt */ |
| 316 | |
| 317 | HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL); |
| 318 | BUG(); |
| 319 | } |
| 320 | |
| 321 | void xen_smp_send_stop(void) |
| 322 | { |
Jeremy Fitzhardinge | fefa629 | 2007-07-17 18:37:07 -0700 | [diff] [blame] | 323 | smp_call_function(stop_self, NULL, 0, 0); |
Jeremy Fitzhardinge | f87e4ca | 2007-07-17 18:37:06 -0700 | [diff] [blame] | 324 | } |
| 325 | |
| 326 | void xen_smp_send_reschedule(int cpu) |
| 327 | { |
| 328 | xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR); |
| 329 | } |
| 330 | |
| 331 | |
| 332 | static void xen_send_IPI_mask(cpumask_t mask, enum ipi_vector vector) |
| 333 | { |
| 334 | unsigned cpu; |
| 335 | |
| 336 | cpus_and(mask, mask, cpu_online_map); |
| 337 | |
| 338 | for_each_cpu_mask(cpu, mask) |
| 339 | xen_send_IPI_one(cpu, vector); |
| 340 | } |
| 341 | |
| 342 | static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id) |
| 343 | { |
| 344 | void (*func) (void *info) = call_data->func; |
| 345 | void *info = call_data->info; |
| 346 | int wait = call_data->wait; |
| 347 | |
| 348 | /* |
| 349 | * Notify initiating CPU that I've grabbed the data and am |
| 350 | * about to execute the function |
| 351 | */ |
| 352 | mb(); |
| 353 | atomic_inc(&call_data->started); |
| 354 | /* |
| 355 | * At this point the info structure may be out of scope unless wait==1 |
| 356 | */ |
| 357 | irq_enter(); |
| 358 | (*func)(info); |
| 359 | irq_exit(); |
| 360 | |
| 361 | if (wait) { |
| 362 | mb(); /* commit everything before setting finished */ |
| 363 | atomic_inc(&call_data->finished); |
| 364 | } |
| 365 | |
| 366 | return IRQ_HANDLED; |
| 367 | } |
| 368 | |
| 369 | int xen_smp_call_function_mask(cpumask_t mask, void (*func)(void *), |
| 370 | void *info, int wait) |
| 371 | { |
| 372 | struct call_data_struct data; |
| 373 | int cpus; |
| 374 | |
| 375 | /* Holding any lock stops cpus from going down. */ |
| 376 | spin_lock(&call_lock); |
| 377 | |
| 378 | cpu_clear(smp_processor_id(), mask); |
| 379 | |
| 380 | cpus = cpus_weight(mask); |
| 381 | if (!cpus) { |
| 382 | spin_unlock(&call_lock); |
| 383 | return 0; |
| 384 | } |
| 385 | |
| 386 | /* Can deadlock when called with interrupts disabled */ |
| 387 | WARN_ON(irqs_disabled()); |
| 388 | |
| 389 | data.func = func; |
| 390 | data.info = info; |
| 391 | atomic_set(&data.started, 0); |
| 392 | data.wait = wait; |
| 393 | if (wait) |
| 394 | atomic_set(&data.finished, 0); |
| 395 | |
| 396 | call_data = &data; |
| 397 | mb(); /* write everything before IPI */ |
| 398 | |
| 399 | /* Send a message to other CPUs and wait for them to respond */ |
| 400 | xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR); |
| 401 | |
| 402 | /* Make sure other vcpus get a chance to run. |
| 403 | XXX too severe? Maybe we should check the other CPU's states? */ |
| 404 | HYPERVISOR_sched_op(SCHEDOP_yield, 0); |
| 405 | |
| 406 | /* Wait for response */ |
| 407 | while (atomic_read(&data.started) != cpus || |
| 408 | (wait && atomic_read(&data.finished) != cpus)) |
| 409 | cpu_relax(); |
| 410 | |
| 411 | spin_unlock(&call_lock); |
| 412 | |
| 413 | return 0; |
| 414 | } |