arch/x86/xen/enlighten.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
   4 #include <linux/bootmem.h>
   5 #endif
   6 #include <linux/cpu.h>
   7 #include <linux/kexec.h>
   8 #include <linux/slab.h>
   9
  10 #include <xen/xen.h>
  11 #include <xen/features.h>
  12 #include <xen/page.h>
  13 #include <xen/interface/memory.h>
  14
  15 #include <asm/xen/hypercall.h>
  16 #include <asm/xen/hypervisor.h>
  17 #include <asm/cpu.h>
  18 #include <asm/e820/api.h>
  19
  20 #include "xen-ops.h"
  21 #include "smp.h"
  22 #include "pmu.h"
  23
  24 EXPORT_SYMBOL_GPL(hypercall_page);
  25
  26 /*
  27  * Pointer to the xen_vcpu_info structure or
  28  * &HYPERVISOR_shared_info->vcpu_info[cpu]. See xen_hvm_init_shared_info
  29  * and xen_vcpu_setup for details. By default it points to share_info->vcpu_info
  30  * but if the hypervisor supports VCPUOP_register_vcpu_info then it can point
  31  * to xen_vcpu_info. The pointer is used in __xen_evtchn_do_upcall to
  32  * acknowledge pending events.
  33  * Also more subtly it is used by the patched version of irq enable/disable
  34  * e.g. xen_irq_enable_direct and xen_iret in PV mode.
  35  *
  36  * The desire to be able to do those mask/unmask operations as a single
  37  * instruction by using the per-cpu offset held in %gs is the real reason
  38  * vcpu info is in a per-cpu pointer and the original reason for this
  39  * hypercall.
  40  *
  41  */
  42 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
  43
  44 /*
  45  * Per CPU pages used if hypervisor supports VCPUOP_register_vcpu_info
  46  * hypercall. This can be used both in PV and PVHVM mode. The structure
  47  * overrides the default per_cpu(xen_vcpu, cpu) value.
  48  */
  49 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
  50
  51 /* Linux <-> Xen vCPU id mapping */
  52 DEFINE_PER_CPU(uint32_t, xen_vcpu_id);
  53 EXPORT_PER_CPU_SYMBOL(xen_vcpu_id);
  54
  55 enum xen_domain_type xen_domain_type = XEN_NATIVE;
  56 EXPORT_SYMBOL_GPL(xen_domain_type);
  57
  58 unsigned long *machine_to_phys_mapping = (void *)MACH2PHYS_VIRT_START;
  59 EXPORT_SYMBOL(machine_to_phys_mapping);
  60 unsigned long  machine_to_phys_nr;
  61 EXPORT_SYMBOL(machine_to_phys_nr);
  62
  63 struct start_info *xen_start_info;
  64 EXPORT_SYMBOL_GPL(xen_start_info);
  65
  66 struct shared_info xen_dummy_shared_info;
  67
  68 __read_mostly int xen_have_vector_callback;
  69 EXPORT_SYMBOL_GPL(xen_have_vector_callback);
  70
  71 /*
  72  * NB: needs to live in .data because it's used by xen_prepare_pvh which runs
  73  * before clearing the bss.
  74  */
  75 uint32_t xen_start_flags __attribute__((section(".data"))) = 0;
  76 EXPORT_SYMBOL(xen_start_flags);
  77
  78 /*
  79  * Point at some empty memory to start with. We map the real shared_info
  80  * page as soon as fixmap is up and running.
  81  */
  82 struct shared_info *HYPERVISOR_shared_info = &xen_dummy_shared_info;
  83
  84 /*
  85  * Flag to determine whether vcpu info placement is available on all
  86  * VCPUs.  We assume it is to start with, and then set it to zero on
  87  * the first failure.  This is because it can succeed on some VCPUs
  88  * and not others, since it can involve hypervisor memory allocation,
  89  * or because the guest failed to guarantee all the appropriate
  90  * constraints on all VCPUs (ie buffer can't cross a page boundary).
  91  *
  92  * Note that any particular CPU may be using a placed vcpu structure,
  93  * but we can only optimise if the all are.
  94  *
  95  * 0: not available, 1: available
  96  */
  97 int xen_have_vcpu_info_placement = 1;
  98
  99 static int xen_cpu_up_online(unsigned int cpu)
 100 {
 101         xen_init_lock_cpu(cpu);
 102         return 0;
 103 }
 104
 105 int xen_cpuhp_setup(int (*cpu_up_prepare_cb)(unsigned int),
 106                     int (*cpu_dead_cb)(unsigned int))
 107 {
 108         int rc;
 109
 110         rc = cpuhp_setup_state_nocalls(CPUHP_XEN_PREPARE,
 111                                        "x86/xen/guest:prepare",
 112                                        cpu_up_prepare_cb, cpu_dead_cb);
 113         if (rc >= 0) {
 114                 rc = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 115                                                "x86/xen/guest:online",
 116                                                xen_cpu_up_online, NULL);
 117                 if (rc < 0)
 118                         cpuhp_remove_state_nocalls(CPUHP_XEN_PREPARE);
 119         }
 120
 121         return rc >= 0 ? 0 : rc;
 122 }
 123
 124 static int xen_vcpu_setup_restore(int cpu)
 125 {
 126         int rc = 0;
 127
 128         /* Any per_cpu(xen_vcpu) is stale, so reset it */
 129         xen_vcpu_info_reset(cpu);
 130
 131         /*
 132          * For PVH and PVHVM, setup online VCPUs only. The rest will
 133          * be handled by hotplug.
 134          */
 135         if (xen_pv_domain() ||
 136             (xen_hvm_domain() && cpu_online(cpu))) {
 137                 rc = xen_vcpu_setup(cpu);
 138         }
 139
 140         return rc;
 141 }
 142
 143 /*
 144  * On restore, set the vcpu placement up again.
 145  * If it fails, then we're in a bad state, since
 146  * we can't back out from using it...
 147  */
 148 void xen_vcpu_restore(void)
 149 {
 150         int cpu, rc;
 151
 152         for_each_possible_cpu(cpu) {
 153                 bool other_cpu = (cpu != smp_processor_id());
 154                 bool is_up;
 155
 156                 if (xen_vcpu_nr(cpu) == XEN_VCPU_ID_INVALID)
 157                         continue;
 158
 159                 /* Only Xen 4.5 and higher support this. */
 160                 is_up = HYPERVISOR_vcpu_op(VCPUOP_is_up,
 161                                            xen_vcpu_nr(cpu), NULL) > 0;
 162
 163                 if (other_cpu && is_up &&
 164                     HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL))
 165                         BUG();
 166
 167                 if (xen_pv_domain() || xen_feature(XENFEAT_hvm_safe_pvclock))
 168                         xen_setup_runstate_info(cpu);
 169
 170                 rc = xen_vcpu_setup_restore(cpu);
 171                 if (rc)
 172                         pr_emerg_once("vcpu restore failed for cpu=%d err=%d. "
 173                                         "System will hang.\n", cpu, rc);
 174                 /*
 175                  * In case xen_vcpu_setup_restore() fails, do not bring up the
 176                  * VCPU. This helps us avoid the resulting OOPS when the VCPU
 177                  * accesses pvclock_vcpu_time via xen_vcpu (which is NULL.)
 178                  * Note that this does not improve the situation much -- now the
 179                  * VM hangs instead of OOPSing -- with the VCPUs that did not
 180                  * fail, spinning in stop_machine(), waiting for the failed
 181                  * VCPUs to come up.
 182                  */
 183                 if (other_cpu && is_up && (rc == 0) &&
 184                     HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL))
 185                         BUG();
 186         }
 187 }
 188
 189 void xen_vcpu_info_reset(int cpu)
 190 {
 191         if (xen_vcpu_nr(cpu) < MAX_VIRT_CPUS) {
 192                 per_cpu(xen_vcpu, cpu) =
 193                         &HYPERVISOR_shared_info->vcpu_info[xen_vcpu_nr(cpu)];
 194         } else {
 195                 /* Set to NULL so that if somebody accesses it we get an OOPS */
 196                 per_cpu(xen_vcpu, cpu) = NULL;
 197         }
 198 }
 199
 200 int xen_vcpu_setup(int cpu)
 201 {
 202         struct vcpu_register_vcpu_info info;
 203         int err;
 204         struct vcpu_info *vcpup;
 205
 206         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
 207
 208         /*
 209          * This path is called on PVHVM at bootup (xen_hvm_smp_prepare_boot_cpu)
 210          * and at restore (xen_vcpu_restore). Also called for hotplugged
 211          * VCPUs (cpu_init -> xen_hvm_cpu_prepare_hvm).
 212          * However, the hypercall can only be done once (see below) so if a VCPU
 213          * is offlined and comes back online then let's not redo the hypercall.
 214          *
 215          * For PV it is called during restore (xen_vcpu_restore) and bootup
 216          * (xen_setup_vcpu_info_placement). The hotplug mechanism does not
 217          * use this function.
 218          */
 219         if (xen_hvm_domain()) {
 220                 if (per_cpu(xen_vcpu, cpu) == &per_cpu(xen_vcpu_info, cpu))
 221                         return 0;
 222         }
 223
 224         if (xen_have_vcpu_info_placement) {
 225                 vcpup = &per_cpu(xen_vcpu_info, cpu);
 226                 info.mfn = arbitrary_virt_to_mfn(vcpup);
 227                 info.offset = offset_in_page(vcpup);
 228
 229                 /*
 230                  * Check to see if the hypervisor will put the vcpu_info
 231                  * structure where we want it, which allows direct access via
 232                  * a percpu-variable.
 233                  * N.B. This hypercall can _only_ be called once per CPU.
 234                  * Subsequent calls will error out with -EINVAL. This is due to
 235                  * the fact that hypervisor has no unregister variant and this
 236                  * hypercall does not allow to over-write info.mfn and
 237                  * info.offset.
 238                  */
 239                 err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info,
 240                                          xen_vcpu_nr(cpu), &info);
 241
 242                 if (err) {
 243                         pr_warn_once("register_vcpu_info failed: cpu=%d err=%d\n",
 244                                      cpu, err);
 245                         xen_have_vcpu_info_placement = 0;
 246                 } else {
 247                         /*
 248                          * This cpu is using the registered vcpu info, even if
 249                          * later ones fail to.
 250                          */
 251                         per_cpu(xen_vcpu, cpu) = vcpup;
 252                 }
 253         }
 254
 255         if (!xen_have_vcpu_info_placement)
 256                 xen_vcpu_info_reset(cpu);
 257
 258         return ((per_cpu(xen_vcpu, cpu) == NULL) ? -ENODEV : 0);
 259 }
 260
 261 void xen_reboot(int reason)
 262 {
 263         struct sched_shutdown r = { .reason = reason };
 264         int cpu;
 265
 266         for_each_online_cpu(cpu)
 267                 xen_pmu_finish(cpu);
 268
 269         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
 270                 BUG();
 271 }
 272
 273 void xen_emergency_restart(void)
 274 {
 275         xen_reboot(SHUTDOWN_reboot);
 276 }
 277
 278 static int
 279 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
 280 {
 281         if (!kexec_crash_loaded())
 282                 xen_reboot(SHUTDOWN_crash);
 283         return NOTIFY_DONE;
 284 }
 285
 286 static struct notifier_block xen_panic_block = {
 287         .notifier_call = xen_panic_event,
 288         .priority = INT_MIN
 289 };
 290
 291 int xen_panic_handler_init(void)
 292 {
 293         atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block);
 294         return 0;
 295 }
 296
 297 void xen_pin_vcpu(int cpu)
 298 {
 299         static bool disable_pinning;
 300         struct sched_pin_override pin_override;
 301         int ret;
 302
 303         if (disable_pinning)
 304                 return;
 305
 306         pin_override.pcpu = cpu;
 307         ret = HYPERVISOR_sched_op(SCHEDOP_pin_override, &pin_override);
 308
 309         /* Ignore errors when removing override. */
 310         if (cpu < 0)
 311                 return;
 312
 313         switch (ret) {
 314         case -ENOSYS:
 315                 pr_warn("Unable to pin on physical cpu %d. In case of problems consider vcpu pinning.\n",
 316                         cpu);
 317                 disable_pinning = true;
 318                 break;
 319         case -EPERM:
 320                 WARN(1, "Trying to pin vcpu without having privilege to do so\n");
 321                 disable_pinning = true;
 322                 break;
 323         case -EINVAL:
 324         case -EBUSY:
 325                 pr_warn("Physical cpu %d not available for pinning. Check Xen cpu configuration.\n",
 326                         cpu);
 327                 break;
 328         case 0:
 329                 break;
 330         default:
 331                 WARN(1, "rc %d while trying to pin vcpu\n", ret);
 332                 disable_pinning = true;
 333         }
 334 }
 335
 336 #ifdef CONFIG_HOTPLUG_CPU
 337 void xen_arch_register_cpu(int num)
 338 {
 339         arch_register_cpu(num);
 340 }
 341 EXPORT_SYMBOL(xen_arch_register_cpu);
 342
 343 void xen_arch_unregister_cpu(int num)
 344 {
 345         arch_unregister_cpu(num);
 346 }
 347 EXPORT_SYMBOL(xen_arch_unregister_cpu);
 348 #endif
 349
 350 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 351 void __init arch_xen_balloon_init(struct resource *hostmem_resource)
 352 {
 353         struct xen_memory_map memmap;
 354         int rc;
 355         unsigned int i, last_guest_ram;
 356         phys_addr_t max_addr = PFN_PHYS(max_pfn);
 357         struct e820_table *xen_e820_table;
 358         const struct e820_entry *entry;
 359         struct resource *res;
 360
 361         if (!xen_initial_domain())
 362                 return;
 363
 364         xen_e820_table = kmalloc(sizeof(*xen_e820_table), GFP_KERNEL);
 365         if (!xen_e820_table)
 366                 return;
 367
 368         memmap.nr_entries = ARRAY_SIZE(xen_e820_table->entries);
 369         set_xen_guest_handle(memmap.buffer, xen_e820_table->entries);
 370         rc = HYPERVISOR_memory_op(XENMEM_machine_memory_map, &memmap);
 371         if (rc) {
 372                 pr_warn("%s: Can't read host e820 (%d)\n", __func__, rc);
 373                 goto out;
 374         }
 375
 376         last_guest_ram = 0;
 377         for (i = 0; i < memmap.nr_entries; i++) {
 378                 if (xen_e820_table->entries[i].addr >= max_addr)
 379                         break;
 380                 if (xen_e820_table->entries[i].type == E820_TYPE_RAM)
 381                         last_guest_ram = i;
 382         }
 383
 384         entry = &xen_e820_table->entries[last_guest_ram];
 385         if (max_addr >= entry->addr + entry->size)
 386                 goto out; /* No unallocated host RAM. */
 387
 388         hostmem_resource->start = max_addr;
 389         hostmem_resource->end = entry->addr + entry->size;
 390
 391         /*
 392          * Mark non-RAM regions between the end of dom0 RAM and end of host RAM
 393          * as unavailable. The rest of that region can be used for hotplug-based
 394          * ballooning.
 395          */
 396         for (; i < memmap.nr_entries; i++) {
 397                 entry = &xen_e820_table->entries[i];
 398
 399                 if (entry->type == E820_TYPE_RAM)
 400                         continue;
 401
 402                 if (entry->addr >= hostmem_resource->end)
 403                         break;
 404
 405                 res = kzalloc(sizeof(*res), GFP_KERNEL);
 406                 if (!res)
 407                         goto out;
 408
 409                 res->name = "Unavailable host RAM";
 410                 res->start = entry->addr;
 411                 res->end = (entry->addr + entry->size < hostmem_resource->end) ?
 412                             entry->addr + entry->size : hostmem_resource->end;
 413                 rc = insert_resource(hostmem_resource, res);
 414                 if (rc) {
 415                         pr_warn("%s: Can't insert [%llx - %llx) (%d)\n",
 416                                 __func__, res->start, res->end, rc);
 417                         kfree(res);
 418                         goto  out;
 419                 }
 420         }
 421
 422  out:
 423         kfree(xen_e820_table);
 424 }
 425 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */