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 static int xen_vcpu_setup_restore(int cpu)
 110 {
 111         int rc = 0;
 112
 113         /* Any per_cpu(xen_vcpu) is stale, so reset it */
 114         xen_vcpu_info_reset(cpu);
 115
 116         /*
 117          * For PVH and PVHVM, setup online VCPUs only. The rest will
 118          * be handled by hotplug.
 119          */
 120         if (xen_pv_domain() ||
 121             (xen_hvm_domain() && cpu_online(cpu))) {
 122                 rc = xen_vcpu_setup(cpu);
 123         }
 124
 125         return rc;
 126 }
 127
 128 /*
 129  * On restore, set the vcpu placement up again.
 130  * If it fails, then we're in a bad state, since
 131  * we can't back out from using it...
 132  */
 133 void xen_vcpu_restore(void)
 134 {
 135         int cpu, rc;
 136
 137         for_each_possible_cpu(cpu) {
 138                 bool other_cpu = (cpu != smp_processor_id());
 139                 bool is_up;
 140
 141                 if (xen_vcpu_nr(cpu) == XEN_VCPU_ID_INVALID)
 142                         continue;
 143
 144                 /* Only Xen 4.5 and higher support this. */
 145                 is_up = HYPERVISOR_vcpu_op(VCPUOP_is_up,
 146                                            xen_vcpu_nr(cpu), NULL) > 0;
 147
 148                 if (other_cpu && is_up &&
 149                     HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL))
 150                         BUG();
 151
 152                 if (xen_pv_domain() || xen_feature(XENFEAT_hvm_safe_pvclock))
 153                         xen_setup_runstate_info(cpu);
 154
 155                 rc = xen_vcpu_setup_restore(cpu);
 156                 if (rc)
 157                         pr_emerg_once("vcpu restore failed for cpu=%d err=%d. "
 158                                         "System will hang.\n", cpu, rc);
 159                 /*
 160                  * In case xen_vcpu_setup_restore() fails, do not bring up the
 161                  * VCPU. This helps us avoid the resulting OOPS when the VCPU
 162                  * accesses pvclock_vcpu_time via xen_vcpu (which is NULL.)
 163                  * Note that this does not improve the situation much -- now the
 164                  * VM hangs instead of OOPSing -- with the VCPUs that did not
 165                  * fail, spinning in stop_machine(), waiting for the failed
 166                  * VCPUs to come up.
 167                  */
 168                 if (other_cpu && is_up && (rc == 0) &&
 169                     HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL))
 170                         BUG();
 171         }
 172 }
 173
 174 void xen_vcpu_info_reset(int cpu)
 175 {
 176         if (xen_vcpu_nr(cpu) < MAX_VIRT_CPUS) {
 177                 per_cpu(xen_vcpu, cpu) =
 178                         &HYPERVISOR_shared_info->vcpu_info[xen_vcpu_nr(cpu)];
 179         } else {
 180                 /* Set to NULL so that if somebody accesses it we get an OOPS */
 181                 per_cpu(xen_vcpu, cpu) = NULL;
 182         }
 183 }
 184
 185 int xen_vcpu_setup(int cpu)
 186 {
 187         struct vcpu_register_vcpu_info info;
 188         int err;
 189         struct vcpu_info *vcpup;
 190
 191         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
 192
 193         /*
 194          * This path is called on PVHVM at bootup (xen_hvm_smp_prepare_boot_cpu)
 195          * and at restore (xen_vcpu_restore). Also called for hotplugged
 196          * VCPUs (cpu_init -> xen_hvm_cpu_prepare_hvm).
 197          * However, the hypercall can only be done once (see below) so if a VCPU
 198          * is offlined and comes back online then let's not redo the hypercall.
 199          *
 200          * For PV it is called during restore (xen_vcpu_restore) and bootup
 201          * (xen_setup_vcpu_info_placement). The hotplug mechanism does not
 202          * use this function.
 203          */
 204         if (xen_hvm_domain()) {
 205                 if (per_cpu(xen_vcpu, cpu) == &per_cpu(xen_vcpu_info, cpu))
 206                         return 0;
 207         }
 208
 209         if (xen_have_vcpu_info_placement) {
 210                 vcpup = &per_cpu(xen_vcpu_info, cpu);
 211                 info.mfn = arbitrary_virt_to_mfn(vcpup);
 212                 info.offset = offset_in_page(vcpup);
 213
 214                 /*
 215                  * Check to see if the hypervisor will put the vcpu_info
 216                  * structure where we want it, which allows direct access via
 217                  * a percpu-variable.
 218                  * N.B. This hypercall can _only_ be called once per CPU.
 219                  * Subsequent calls will error out with -EINVAL. This is due to
 220                  * the fact that hypervisor has no unregister variant and this
 221                  * hypercall does not allow to over-write info.mfn and
 222                  * info.offset.
 223                  */
 224                 err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info,
 225                                          xen_vcpu_nr(cpu), &info);
 226
 227                 if (err) {
 228                         pr_warn_once("register_vcpu_info failed: cpu=%d err=%d\n",
 229                                      cpu, err);
 230                         xen_have_vcpu_info_placement = 0;
 231                 } else {
 232                         /*
 233                          * This cpu is using the registered vcpu info, even if
 234                          * later ones fail to.
 235                          */
 236                         per_cpu(xen_vcpu, cpu) = vcpup;
 237                 }
 238         }
 239
 240         if (!xen_have_vcpu_info_placement)
 241                 xen_vcpu_info_reset(cpu);
 242
 243         return ((per_cpu(xen_vcpu, cpu) == NULL) ? -ENODEV : 0);
 244 }
 245
 246 void xen_reboot(int reason)
 247 {
 248         struct sched_shutdown r = { .reason = reason };
 249         int cpu;
 250
 251         for_each_online_cpu(cpu)
 252                 xen_pmu_finish(cpu);
 253
 254         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
 255                 BUG();
 256 }
 257
 258 void xen_emergency_restart(void)
 259 {
 260         xen_reboot(SHUTDOWN_reboot);
 261 }
 262
 263 static int
 264 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
 265 {
 266         if (!kexec_crash_loaded())
 267                 xen_reboot(SHUTDOWN_crash);
 268         return NOTIFY_DONE;
 269 }
 270
 271 static struct notifier_block xen_panic_block = {
 272         .notifier_call = xen_panic_event,
 273         .priority = INT_MIN
 274 };
 275
 276 int xen_panic_handler_init(void)
 277 {
 278         atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
 279         return 0;
 280 }
 281
 282 void xen_pin_vcpu(int cpu)
 283 {
 284         static bool disable_pinning;
 285         struct sched_pin_override pin_override;
 286         int ret;
 287
 288         if (disable_pinning)
 289                 return;
 290
 291         pin_override.pcpu = cpu;
 292         ret = HYPERVISOR_sched_op(SCHEDOP_pin_override, &pin_override);
 293
 294         /* Ignore errors when removing override. */
 295         if (cpu < 0)
 296                 return;
 297
 298         switch (ret) {
 299         case -ENOSYS:
 300                 pr_warn("Unable to pin on physical cpu %d. In case of problems consider vcpu pinning.\n",
 301                         cpu);
 302                 disable_pinning = true;
 303                 break;
 304         case -EPERM:
 305                 WARN(1, "Trying to pin vcpu without having privilege to do so\n");
 306                 disable_pinning = true;
 307                 break;
 308         case -EINVAL:
 309         case -EBUSY:
 310                 pr_warn("Physical cpu %d not available for pinning. Check Xen cpu configuration.\n",
 311                         cpu);
 312                 break;
 313         case 0:
 314                 break;
 315         default:
 316                 WARN(1, "rc %d while trying to pin vcpu\n", ret);
 317                 disable_pinning = true;
 318         }
 319 }
 320
 321 #ifdef CONFIG_HOTPLUG_CPU
 322 void xen_arch_register_cpu(int num)
 323 {
 324         arch_register_cpu(num);
 325 }
 326 EXPORT_SYMBOL(xen_arch_register_cpu);
 327
 328 void xen_arch_unregister_cpu(int num)
 329 {
 330         arch_unregister_cpu(num);
 331 }
 332 EXPORT_SYMBOL(xen_arch_unregister_cpu);
 333 #endif