arch/x86/xen/enlighten.c

   1 #include <linux/cpu.h>
   2 #include <linux/kexec.h>
   3
   4 #include <xen/features.h>
   5 #include <xen/page.h>
   6
   7 #include <asm/xen/hypercall.h>
   8 #include <asm/xen/hypervisor.h>
   9 #include <asm/cpu.h>
  10 #include <asm/e820/api.h>
  11
  12 #include "xen-ops.h"
  13 #include "smp.h"
  14 #include "pmu.h"
  15
  16 EXPORT_SYMBOL_GPL(hypercall_page);
  17
  18 /*
  19  * Pointer to the xen_vcpu_info structure or
  20  * &HYPERVISOR_shared_info->vcpu_info[cpu]. See xen_hvm_init_shared_info
  21  * and xen_vcpu_setup for details. By default it points to share_info->vcpu_info
  22  * but if the hypervisor supports VCPUOP_register_vcpu_info then it can point
  23  * to xen_vcpu_info. The pointer is used in __xen_evtchn_do_upcall to
  24  * acknowledge pending events.
  25  * Also more subtly it is used by the patched version of irq enable/disable
  26  * e.g. xen_irq_enable_direct and xen_iret in PV mode.
  27  *
  28  * The desire to be able to do those mask/unmask operations as a single
  29  * instruction by using the per-cpu offset held in %gs is the real reason
  30  * vcpu info is in a per-cpu pointer and the original reason for this
  31  * hypercall.
  32  *
  33  */
  34 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
  35
  36 /*
  37  * Per CPU pages used if hypervisor supports VCPUOP_register_vcpu_info
  38  * hypercall. This can be used both in PV and PVHVM mode. The structure
  39  * overrides the default per_cpu(xen_vcpu, cpu) value.
  40  */
  41 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
  42
  43 /* Linux <-> Xen vCPU id mapping */
  44 DEFINE_PER_CPU(uint32_t, xen_vcpu_id);
  45 EXPORT_PER_CPU_SYMBOL(xen_vcpu_id);
  46
  47 enum xen_domain_type xen_domain_type = XEN_NATIVE;
  48 EXPORT_SYMBOL_GPL(xen_domain_type);
  49
  50 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
  51 EXPORT_SYMBOL(machine_to_phys_mapping);
  52 unsigned long  machine_to_phys_nr;
  53 EXPORT_SYMBOL(machine_to_phys_nr);
  54
  55 struct start_info *xen_start_info;
  56 EXPORT_SYMBOL_GPL(xen_start_info);
  57
  58 struct shared_info xen_dummy_shared_info;
  59
  60 __read_mostly int xen_have_vector_callback;
  61 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
  62
  63 /*
  64  * Point at some empty memory to start with. We map the real shared_info
  65  * page as soon as fixmap is up and running.
  66  */
  67 struct shared_info *HYPERVISOR_shared_info = &xen_dummy_shared_info;
  68
  69 /*
  70  * Flag to determine whether vcpu info placement is available on all
  71  * VCPUs.  We assume it is to start with, and then set it to zero on
  72  * the first failure.  This is because it can succeed on some VCPUs
  73  * and not others, since it can involve hypervisor memory allocation,
  74  * or because the guest failed to guarantee all the appropriate
  75  * constraints on all VCPUs (ie buffer can't cross a page boundary).
  76  *
  77  * Note that any particular CPU may be using a placed vcpu structure,
  78  * but we can only optimise if the all are.
  79  *
  80  * 0: not available, 1: available
  81  */
  82 int xen_have_vcpu_info_placement = 1;
  83
  84 static int xen_cpu_up_online(unsigned int cpu)
  85 {
  86         xen_init_lock_cpu(cpu);
  87         return 0;
  88 }
  89
  90 int xen_cpuhp_setup(int (*cpu_up_prepare_cb)(unsigned int),
  91                     int (*cpu_dead_cb)(unsigned int))
  92 {
  93         int rc;
  94
  95         rc = cpuhp_setup_state_nocalls(CPUHP_XEN_PREPARE,
  96                                        "x86/xen/hvm_guest:prepare",
  97                                        cpu_up_prepare_cb, cpu_dead_cb);
  98         if (rc >= 0) {
  99                 rc = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 100                                                "x86/xen/hvm_guest:online",
 101                                                xen_cpu_up_online, NULL);
 102                 if (rc < 0)
 103                         cpuhp_remove_state_nocalls(CPUHP_XEN_PREPARE);
 104         }
 105
 106         return rc >= 0 ? 0 : rc;
 107 }
 108
 109 /*
 110  * On restore, set the vcpu placement up again.
 111  * If it fails, then we're in a bad state, since
 112  * we can't back out from using it...
 113  */
 114 void xen_vcpu_restore(void)
 115 {
 116         int cpu;
 117
 118         for_each_possible_cpu(cpu) {
 119                 bool other_cpu = (cpu != smp_processor_id());
 120                 bool is_up = HYPERVISOR_vcpu_op(VCPUOP_is_up, xen_vcpu_nr(cpu),
 121                                                 NULL);
 122
 123                 if (other_cpu && is_up &&
 124                     HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL))
 125                         BUG();
 126
 127                 xen_setup_runstate_info(cpu);
 128
 129                 if (xen_have_vcpu_info_placement)
 130                         xen_vcpu_setup(cpu);
 131
 132                 if (other_cpu && is_up &&
 133                     HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL))
 134                         BUG();
 135         }
 136 }
 137
 138 static void clamp_max_cpus(void)
 139 {
 140 #ifdef CONFIG_SMP
 141         if (setup_max_cpus > MAX_VIRT_CPUS)
 142                 setup_max_cpus = MAX_VIRT_CPUS;
 143 #endif
 144 }
 145
 146 void xen_vcpu_info_reset(int cpu)
 147 {
 148         if (xen_vcpu_nr(cpu) < MAX_VIRT_CPUS) {
 149                 per_cpu(xen_vcpu, cpu) =
 150                         &HYPERVISOR_shared_info->vcpu_info[xen_vcpu_nr(cpu)];
 151         } else {
 152                 /* Set to NULL so that if somebody accesses it we get an OOPS */
 153                 per_cpu(xen_vcpu, cpu) = NULL;
 154         }
 155 }
 156
 157 void xen_vcpu_setup(int cpu)
 158 {
 159         struct vcpu_register_vcpu_info info;
 160         int err;
 161         struct vcpu_info *vcpup;
 162
 163         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
 164
 165         /*
 166          * This path is called twice on PVHVM - first during bootup via
 167          * smp_init -> xen_hvm_cpu_notify, and then if the VCPU is being
 168          * hotplugged: cpu_up -> xen_hvm_cpu_notify.
 169          * As we can only do the VCPUOP_register_vcpu_info once lets
 170          * not over-write its result.
 171          *
 172          * For PV it is called during restore (xen_vcpu_restore) and bootup
 173          * (xen_setup_vcpu_info_placement). The hotplug mechanism does not
 174          * use this function.
 175          */
 176         if (xen_hvm_domain()) {
 177                 if (per_cpu(xen_vcpu, cpu) == &per_cpu(xen_vcpu_info, cpu))
 178                         return;
 179         }
 180
 181         xen_vcpu_info_reset(cpu);
 182
 183         if (xen_have_vcpu_info_placement) {
 184                 vcpup = &per_cpu(xen_vcpu_info, cpu);
 185                 info.mfn = arbitrary_virt_to_mfn(vcpup);
 186                 info.offset = offset_in_page(vcpup);
 187
 188                 /*
 189                  * Check to see if the hypervisor will put the vcpu_info
 190                  * structure where we want it, which allows direct access via
 191                  * a percpu-variable.
 192                  * N.B. This hypercall can _only_ be called once per CPU.
 193                  * Subsequent calls will error out with -EINVAL. This is due to
 194                  * the fact that hypervisor has no unregister variant and this
 195                  * hypercall does not allow to over-write info.mfn and
 196                  * info.offset.
 197                  */
 198                 err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info,
 199                                          xen_vcpu_nr(cpu), &info);
 200
 201                 if (err) {
 202                         pr_warn_once("register_vcpu_info failed: cpu=%d err=%d\n",
 203                                      cpu, err);
 204                         xen_have_vcpu_info_placement = 0;
 205                 } else {
 206                         /*
 207                          * This cpu is using the registered vcpu info, even if
 208                          * later ones fail to.
 209                          */
 210                         per_cpu(xen_vcpu, cpu) = vcpup;
 211                 }
 212         }
 213
 214         if (!xen_have_vcpu_info_placement) {
 215                 if (cpu >= MAX_VIRT_CPUS)
 216                         clamp_max_cpus();
 217                 return;
 218         }
 219 }
 220
 221 void xen_reboot(int reason)
 222 {
 223         struct sched_shutdown r = { .reason = reason };
 224         int cpu;
 225
 226         for_each_online_cpu(cpu)
 227                 xen_pmu_finish(cpu);
 228
 229         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
 230                 BUG();
 231 }
 232
 233 void xen_emergency_restart(void)
 234 {
 235         xen_reboot(SHUTDOWN_reboot);
 236 }
 237
 238 static int
 239 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
 240 {
 241         if (!kexec_crash_loaded())
 242                 xen_reboot(SHUTDOWN_crash);
 243         return NOTIFY_DONE;
 244 }
 245
 246 static struct notifier_block xen_panic_block = {
 247         .notifier_call = xen_panic_event,
 248         .priority = INT_MIN
 249 };
 250
 251 int xen_panic_handler_init(void)
 252 {
 253         atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
 254         return 0;
 255 }
 256
 257 void xen_pin_vcpu(int cpu)
 258 {
 259         static bool disable_pinning;
 260         struct sched_pin_override pin_override;
 261         int ret;
 262
 263         if (disable_pinning)
 264                 return;
 265
 266         pin_override.pcpu = cpu;
 267         ret = HYPERVISOR_sched_op(SCHEDOP_pin_override, &pin_override);
 268
 269         /* Ignore errors when removing override. */
 270         if (cpu < 0)
 271                 return;
 272
 273         switch (ret) {
 274         case -ENOSYS:
 275                 pr_warn("Unable to pin on physical cpu %d. In case of problems consider vcpu pinning.\n",
 276                         cpu);
 277                 disable_pinning = true;
 278                 break;
 279         case -EPERM:
 280                 WARN(1, "Trying to pin vcpu without having privilege to do so\n");
 281                 disable_pinning = true;
 282                 break;
 283         case -EINVAL:
 284         case -EBUSY:
 285                 pr_warn("Physical cpu %d not available for pinning. Check Xen cpu configuration.\n",
 286                         cpu);
 287                 break;
 288         case 0:
 289                 break;
 290         default:
 291                 WARN(1, "rc %d while trying to pin vcpu\n", ret);
 292                 disable_pinning = true;
 293         }
 294 }
 295
 296 #ifdef CONFIG_HOTPLUG_CPU
 297 void xen_arch_register_cpu(int num)
 298 {
 299         arch_register_cpu(num);
 300 }
 301 EXPORT_SYMBOL(xen_arch_register_cpu);
 302
 303 void xen_arch_unregister_cpu(int num)
 304 {
 305         arch_unregister_cpu(num);
 306 }
 307 EXPORT_SYMBOL(xen_arch_unregister_cpu);
 308 #endif