arch/x86/xen/enlighten.c

   1 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
   2 #include <linux/bootmem.h>
   3 #endif
   4 #include <linux/cpu.h>
   5 #include <linux/kexec.h>
   6 #include <linux/slab.h>
   7
   8 #include <xen/features.h>
   9 #include <xen/page.h>
  10
  11 #include <asm/xen/hypercall.h>
  12 #include <asm/xen/hypervisor.h>
  13 #include <asm/cpu.h>
  14 #include <asm/e820/api.h>
  15
  16 #include "xen-ops.h"
  17 #include "smp.h"
  18 #include "pmu.h"
  19
  20 EXPORT_SYMBOL_GPL(hypercall_page);
  21
  22 /*
  23  * Pointer to the xen_vcpu_info structure or
  24  * &HYPERVISOR_shared_info->vcpu_info[cpu]. See xen_hvm_init_shared_info
  25  * and xen_vcpu_setup for details. By default it points to share_info->vcpu_info
  26  * but if the hypervisor supports VCPUOP_register_vcpu_info then it can point
  27  * to xen_vcpu_info. The pointer is used in __xen_evtchn_do_upcall to
  28  * acknowledge pending events.
  29  * Also more subtly it is used by the patched version of irq enable/disable
  30  * e.g. xen_irq_enable_direct and xen_iret in PV mode.
  31  *
  32  * The desire to be able to do those mask/unmask operations as a single
  33  * instruction by using the per-cpu offset held in %gs is the real reason
  34  * vcpu info is in a per-cpu pointer and the original reason for this
  35  * hypercall.
  36  *
  37  */
  38 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
  39
  40 /*
  41  * Per CPU pages used if hypervisor supports VCPUOP_register_vcpu_info
  42  * hypercall. This can be used both in PV and PVHVM mode. The structure
  43  * overrides the default per_cpu(xen_vcpu, cpu) value.
  44  */
  45 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
  46
  47 /* Linux <-> Xen vCPU id mapping */
  48 DEFINE_PER_CPU(uint32_t, xen_vcpu_id);
  49 EXPORT_PER_CPU_SYMBOL(xen_vcpu_id);
  50
  51 enum xen_domain_type xen_domain_type = XEN_NATIVE;
  52 EXPORT_SYMBOL_GPL(xen_domain_type);
  53
  54 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
  55 EXPORT_SYMBOL(machine_to_phys_mapping);
  56 unsigned long  machine_to_phys_nr;
  57 EXPORT_SYMBOL(machine_to_phys_nr);
  58
  59 struct start_info *xen_start_info;
  60 EXPORT_SYMBOL_GPL(xen_start_info);
  61
  62 struct shared_info xen_dummy_shared_info;
  63
  64 __read_mostly int xen_have_vector_callback;
  65 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
  66
  67 /*
  68  * NB: needs to live in .data because it's used by xen_prepare_pvh which runs
  69  * before clearing the bss.
  70  */
  71 uint32_t xen_start_flags __attribute__((section(".data"))) = 0;
  72 EXPORT_SYMBOL(xen_start_flags);
  73
  74 /*
  75  * Point at some empty memory to start with. We map the real shared_info
  76  * page as soon as fixmap is up and running.
  77  */
  78 struct shared_info *HYPERVISOR_shared_info = &xen_dummy_shared_info;
  79
  80 /*
  81  * Flag to determine whether vcpu info placement is available on all
  82  * VCPUs.  We assume it is to start with, and then set it to zero on
  83  * the first failure.  This is because it can succeed on some VCPUs
  84  * and not others, since it can involve hypervisor memory allocation,
  85  * or because the guest failed to guarantee all the appropriate
  86  * constraints on all VCPUs (ie buffer can't cross a page boundary).
  87  *
  88  * Note that any particular CPU may be using a placed vcpu structure,
  89  * but we can only optimise if the all are.
  90  *
  91  * 0: not available, 1: available
  92  */
  93 int xen_have_vcpu_info_placement = 1;
  94
  95 static int xen_cpu_up_online(unsigned int cpu)
  96 {
  97         xen_init_lock_cpu(cpu);
  98         return 0;
  99 }
 100
 101 int xen_cpuhp_setup(int (*cpu_up_prepare_cb)(unsigned int),
 102                     int (*cpu_dead_cb)(unsigned int))
 103 {
 104         int rc;
 105
 106         rc = cpuhp_setup_state_nocalls(CPUHP_XEN_PREPARE,
 107                                        "x86/xen/guest:prepare",
 108                                        cpu_up_prepare_cb, cpu_dead_cb);
 109         if (rc >= 0) {
 110                 rc = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 111                                                "x86/xen/guest:online",
 112                                                xen_cpu_up_online, NULL);
 113                 if (rc < 0)
 114                         cpuhp_remove_state_nocalls(CPUHP_XEN_PREPARE);
 115         }
 116
 117         return rc >= 0 ? 0 : rc;
 118 }
 119
 120 static int xen_vcpu_setup_restore(int cpu)
 121 {
 122         int rc = 0;
 123
 124         /* Any per_cpu(xen_vcpu) is stale, so reset it */
 125         xen_vcpu_info_reset(cpu);
 126
 127         /*
 128          * For PVH and PVHVM, setup online VCPUs only. The rest will
 129          * be handled by hotplug.
 130          */
 131         if (xen_pv_domain() ||
 132             (xen_hvm_domain() && cpu_online(cpu))) {
 133                 rc = xen_vcpu_setup(cpu);
 134         }
 135
 136         return rc;
 137 }
 138
 139 /*
 140  * On restore, set the vcpu placement up again.
 141  * If it fails, then we're in a bad state, since
 142  * we can't back out from using it...
 143  */
 144 void xen_vcpu_restore(void)
 145 {
 146         int cpu, rc;
 147
 148         for_each_possible_cpu(cpu) {
 149                 bool other_cpu = (cpu != smp_processor_id());
 150                 bool is_up;
 151
 152                 if (xen_vcpu_nr(cpu) == XEN_VCPU_ID_INVALID)
 153                         continue;
 154
 155                 /* Only Xen 4.5 and higher support this. */
 156                 is_up = HYPERVISOR_vcpu_op(VCPUOP_is_up,
 157                                            xen_vcpu_nr(cpu), NULL) > 0;
 158
 159                 if (other_cpu && is_up &&
 160                     HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL))
 161                         BUG();
 162
 163                 if (xen_pv_domain() || xen_feature(XENFEAT_hvm_safe_pvclock))
 164                         xen_setup_runstate_info(cpu);
 165
 166                 rc = xen_vcpu_setup_restore(cpu);
 167                 if (rc)
 168                         pr_emerg_once("vcpu restore failed for cpu=%d err=%d. "
 169                                         "System will hang.\n", cpu, rc);
 170                 /*
 171                  * In case xen_vcpu_setup_restore() fails, do not bring up the
 172                  * VCPU. This helps us avoid the resulting OOPS when the VCPU
 173                  * accesses pvclock_vcpu_time via xen_vcpu (which is NULL.)
 174                  * Note that this does not improve the situation much -- now the
 175                  * VM hangs instead of OOPSing -- with the VCPUs that did not
 176                  * fail, spinning in stop_machine(), waiting for the failed
 177                  * VCPUs to come up.
 178                  */
 179                 if (other_cpu && is_up && (rc == 0) &&
 180                     HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL))
 181                         BUG();
 182         }
 183 }
 184
 185 void xen_vcpu_info_reset(int cpu)
 186 {
 187         if (xen_vcpu_nr(cpu) < MAX_VIRT_CPUS) {
 188                 per_cpu(xen_vcpu, cpu) =
 189                         &HYPERVISOR_shared_info->vcpu_info[xen_vcpu_nr(cpu)];
 190         } else {
 191                 /* Set to NULL so that if somebody accesses it we get an OOPS */
 192                 per_cpu(xen_vcpu, cpu) = NULL;
 193         }
 194 }
 195
 196 int xen_vcpu_setup(int cpu)
 197 {
 198         struct vcpu_register_vcpu_info info;
 199         int err;
 200         struct vcpu_info *vcpup;
 201
 202         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
 203
 204         /*
 205          * This path is called on PVHVM at bootup (xen_hvm_smp_prepare_boot_cpu)
 206          * and at restore (xen_vcpu_restore). Also called for hotplugged
 207          * VCPUs (cpu_init -> xen_hvm_cpu_prepare_hvm).
 208          * However, the hypercall can only be done once (see below) so if a VCPU
 209          * is offlined and comes back online then let's not redo the hypercall.
 210          *
 211          * For PV it is called during restore (xen_vcpu_restore) and bootup
 212          * (xen_setup_vcpu_info_placement). The hotplug mechanism does not
 213          * use this function.
 214          */
 215         if (xen_hvm_domain()) {
 216                 if (per_cpu(xen_vcpu, cpu) == &per_cpu(xen_vcpu_info, cpu))
 217                         return 0;
 218         }
 219
 220         if (xen_have_vcpu_info_placement) {
 221                 vcpup = &per_cpu(xen_vcpu_info, cpu);
 222                 info.mfn = arbitrary_virt_to_mfn(vcpup);
 223                 info.offset = offset_in_page(vcpup);
 224
 225                 /*
 226                  * Check to see if the hypervisor will put the vcpu_info
 227                  * structure where we want it, which allows direct access via
 228                  * a percpu-variable.
 229                  * N.B. This hypercall can _only_ be called once per CPU.
 230                  * Subsequent calls will error out with -EINVAL. This is due to
 231                  * the fact that hypervisor has no unregister variant and this
 232                  * hypercall does not allow to over-write info.mfn and
 233                  * info.offset.
 234                  */
 235                 err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info,
 236                                          xen_vcpu_nr(cpu), &info);
 237
 238                 if (err) {
 239                         pr_warn_once("register_vcpu_info failed: cpu=%d err=%d\n",
 240                                      cpu, err);
 241                         xen_have_vcpu_info_placement = 0;
 242                 } else {
 243                         /*
 244                          * This cpu is using the registered vcpu info, even if
 245                          * later ones fail to.
 246                          */
 247                         per_cpu(xen_vcpu, cpu) = vcpup;
 248                 }
 249         }
 250
 251         if (!xen_have_vcpu_info_placement)
 252                 xen_vcpu_info_reset(cpu);
 253
 254         return ((per_cpu(xen_vcpu, cpu) == NULL) ? -ENODEV : 0);
 255 }
 256
 257 void xen_reboot(int reason)
 258 {
 259         struct sched_shutdown r = { .reason = reason };
 260         int cpu;
 261
 262         for_each_online_cpu(cpu)
 263                 xen_pmu_finish(cpu);
 264
 265         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
 266                 BUG();
 267 }
 268
 269 static int reboot_reason = SHUTDOWN_reboot;
 270 static bool xen_legacy_crash;
 271 void xen_emergency_restart(void)
 272 {
 273         xen_reboot(reboot_reason);
 274 }
 275
 276 static int
 277 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
 278 {
 279         if (!kexec_crash_loaded()) {
 280                 if (xen_legacy_crash)
 281                         xen_reboot(SHUTDOWN_crash);
 282
 283                 reboot_reason = SHUTDOWN_crash;
 284
 285                 /*
 286                  * If panic_timeout==0 then we are supposed to wait forever.
 287                  * However, to preserve original dom0 behavior we have to drop
 288                  * into hypervisor. (domU behavior is controlled by its
 289                  * config file)
 290                  */
 291                 if (panic_timeout == 0)
 292                         panic_timeout = -1;
 293         }
 294         return NOTIFY_DONE;
 295 }
 296
 297 static int __init parse_xen_legacy_crash(char *arg)
 298 {
 299         xen_legacy_crash = true;
 300         return 0;
 301 }
 302 early_param("xen_legacy_crash", parse_xen_legacy_crash);
 303
 304 static struct notifier_block xen_panic_block = {
 305         .notifier_call = xen_panic_event,
 306         .priority = INT_MIN
 307 };
 308
 309 int xen_panic_handler_init(void)
 310 {
 311         atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
 312         return 0;
 313 }
 314
 315 void xen_pin_vcpu(int cpu)
 316 {
 317         static bool disable_pinning;
 318         struct sched_pin_override pin_override;
 319         int ret;
 320
 321         if (disable_pinning)
 322                 return;
 323
 324         pin_override.pcpu = cpu;
 325         ret = HYPERVISOR_sched_op(SCHEDOP_pin_override, &pin_override);
 326
 327         /* Ignore errors when removing override. */
 328         if (cpu < 0)
 329                 return;
 330
 331         switch (ret) {
 332         case -ENOSYS:
 333                 pr_warn("Unable to pin on physical cpu %d. In case of problems consider vcpu pinning.\n",
 334                         cpu);
 335                 disable_pinning = true;
 336                 break;
 337         case -EPERM:
 338                 WARN(1, "Trying to pin vcpu without having privilege to do so\n");
 339                 disable_pinning = true;
 340                 break;
 341         case -EINVAL:
 342         case -EBUSY:
 343                 pr_warn("Physical cpu %d not available for pinning. Check Xen cpu configuration.\n",
 344                         cpu);
 345                 break;
 346         case 0:
 347                 break;
 348         default:
 349                 WARN(1, "rc %d while trying to pin vcpu\n", ret);
 350                 disable_pinning = true;
 351         }
 352 }
 353
 354 #ifdef CONFIG_HOTPLUG_CPU
 355 void xen_arch_register_cpu(int num)
 356 {
 357         arch_register_cpu(num);
 358 }
 359 EXPORT_SYMBOL(xen_arch_register_cpu);
 360
 361 void xen_arch_unregister_cpu(int num)
 362 {
 363         arch_unregister_cpu(num);
 364 }
 365 EXPORT_SYMBOL(xen_arch_unregister_cpu);
 366 #endif