hw/i386/microvm.c

   1 /*
   2  * Copyright (c) 2018 Intel Corporation
   3  * Copyright (c) 2019 Red Hat, Inc.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2 or later, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along with
  15  * this program.  If not, see <http://www.gnu.org/licenses/>.
  16  */
  17
  18 #include "qemu/osdep.h"
  19 #include "qemu/error-report.h"
  20 #include "qemu/cutils.h"
  21 #include "qemu/units.h"
  22 #include "qapi/error.h"
  23 #include "qapi/visitor.h"
  24 #include "qapi/qapi-visit-common.h"
  25 #include "sysemu/sysemu.h"
  26 #include "sysemu/cpus.h"
  27 #include "sysemu/numa.h"
  28 #include "sysemu/reset.h"
  29
  30 #include "hw/loader.h"
  31 #include "hw/irq.h"
  32 #include "hw/kvm/clock.h"
  33 #include "hw/i386/microvm.h"
  34 #include "hw/i386/x86.h"
  35 #include "target/i386/cpu.h"
  36 #include "hw/intc/i8259.h"
  37 #include "hw/timer/i8254.h"
  38 #include "hw/rtc/mc146818rtc.h"
  39 #include "hw/char/serial.h"
  40 #include "hw/i386/topology.h"
  41 #include "hw/i386/e820_memory_layout.h"
  42 #include "hw/i386/fw_cfg.h"
  43 #include "hw/virtio/virtio-mmio.h"
  44
  45 #include "cpu.h"
  46 #include "elf.h"
  47 #include "kvm_i386.h"
  48 #include "hw/xen/start_info.h"
  49
  50 #define MICROVM_BIOS_FILENAME "bios-microvm.bin"
  51
  52 static void microvm_set_rtc(MicrovmMachineState *mms, ISADevice *s)
  53 {
  54     X86MachineState *x86ms = X86_MACHINE(mms);
  55     int val;
  56
  57     val = MIN(x86ms->below_4g_mem_size / KiB, 640);
  58     rtc_set_memory(s, 0x15, val);
  59     rtc_set_memory(s, 0x16, val >> 8);
  60     /* extended memory (next 64MiB) */
  61     if (x86ms->below_4g_mem_size > 1 * MiB) {
  62         val = (x86ms->below_4g_mem_size - 1 * MiB) / KiB;
  63     } else {
  64         val = 0;
  65     }
  66     if (val > 65535) {
  67         val = 65535;
  68     }
  69     rtc_set_memory(s, 0x17, val);
  70     rtc_set_memory(s, 0x18, val >> 8);
  71     rtc_set_memory(s, 0x30, val);
  72     rtc_set_memory(s, 0x31, val >> 8);
  73     /* memory between 16MiB and 4GiB */
  74     if (x86ms->below_4g_mem_size > 16 * MiB) {
  75         val = (x86ms->below_4g_mem_size - 16 * MiB) / (64 * KiB);
  76     } else {
  77         val = 0;
  78     }
  79     if (val > 65535) {
  80         val = 65535;
  81     }
  82     rtc_set_memory(s, 0x34, val);
  83     rtc_set_memory(s, 0x35, val >> 8);
  84     /* memory above 4GiB */
  85     val = x86ms->above_4g_mem_size / 65536;
  86     rtc_set_memory(s, 0x5b, val);
  87     rtc_set_memory(s, 0x5c, val >> 8);
  88     rtc_set_memory(s, 0x5d, val >> 16);
  89 }
  90
  91 static void microvm_gsi_handler(void *opaque, int n, int level)
  92 {
  93     GSIState *s = opaque;
  94
  95     qemu_set_irq(s->ioapic_irq[n], level);
  96 }
  97
  98 static void microvm_devices_init(MicrovmMachineState *mms)
  99 {
 100     X86MachineState *x86ms = X86_MACHINE(mms);
 101     ISABus *isa_bus;
 102     ISADevice *rtc_state;
 103     GSIState *gsi_state;
 104     int i;
 105
 106     /* Core components */
 107
 108     gsi_state = g_malloc0(sizeof(*gsi_state));
 109     if (mms->pic == ON_OFF_AUTO_ON || mms->pic == ON_OFF_AUTO_AUTO) {
 110         x86ms->gsi = qemu_allocate_irqs(gsi_handler, gsi_state, GSI_NUM_PINS);
 111     } else {
 112         x86ms->gsi = qemu_allocate_irqs(microvm_gsi_handler,
 113                                         gsi_state, GSI_NUM_PINS);
 114     }
 115
 116     isa_bus = isa_bus_new(NULL, get_system_memory(), get_system_io(),
 117                           &error_abort);
 118     isa_bus_irqs(isa_bus, x86ms->gsi);
 119
 120     ioapic_init_gsi(gsi_state, "machine");
 121
 122     kvmclock_create();
 123
 124     for (i = 0; i < VIRTIO_NUM_TRANSPORTS; i++) {
 125         sysbus_create_simple("virtio-mmio",
 126                              VIRTIO_MMIO_BASE + i * 512,
 127                              x86ms->gsi[VIRTIO_IRQ_BASE + i]);
 128     }
 129
 130     /* Optional and legacy devices */
 131
 132     if (mms->pic == ON_OFF_AUTO_ON || mms->pic == ON_OFF_AUTO_AUTO) {
 133         qemu_irq *i8259;
 134
 135         i8259 = i8259_init(isa_bus, x86_allocate_cpu_irq());
 136         for (i = 0; i < ISA_NUM_IRQS; i++) {
 137             gsi_state->i8259_irq[i] = i8259[i];
 138         }
 139         g_free(i8259);
 140     }
 141
 142     if (mms->pit == ON_OFF_AUTO_ON || mms->pit == ON_OFF_AUTO_AUTO) {
 143         if (kvm_pit_in_kernel()) {
 144             kvm_pit_init(isa_bus, 0x40);
 145         } else {
 146             i8254_pit_init(isa_bus, 0x40, 0, NULL);
 147         }
 148     }
 149
 150     if (mms->rtc == ON_OFF_AUTO_ON ||
 151         (mms->rtc == ON_OFF_AUTO_AUTO && !kvm_enabled())) {
 152         rtc_state = mc146818_rtc_init(isa_bus, 2000, NULL);
 153         microvm_set_rtc(mms, rtc_state);
 154     }
 155
 156     if (mms->isa_serial) {
 157         serial_hds_isa_init(isa_bus, 0, 1);
 158     }
 159
 160     if (bios_name == NULL) {
 161         bios_name = MICROVM_BIOS_FILENAME;
 162     }
 163     x86_bios_rom_init(get_system_memory(), true);
 164 }
 165
 166 static void microvm_memory_init(MicrovmMachineState *mms)
 167 {
 168     MachineState *machine = MACHINE(mms);
 169     X86MachineState *x86ms = X86_MACHINE(mms);
 170     MemoryRegion *ram_below_4g, *ram_above_4g;
 171     MemoryRegion *system_memory = get_system_memory();
 172     FWCfgState *fw_cfg;
 173     ram_addr_t lowmem;
 174     int i;
 175
 176     /*
 177      * Check whether RAM fits below 4G (leaving 1/2 GByte for IO memory
 178      * and 256 Mbytes for PCI Express Enhanced Configuration Access Mapping
 179      * also known as MMCFG).
 180      * If it doesn't, we need to split it in chunks below and above 4G.
 181      * In any case, try to make sure that guest addresses aligned at
 182      * 1G boundaries get mapped to host addresses aligned at 1G boundaries.
 183      */
 184     if (machine->ram_size >= 0xb0000000) {
 185         lowmem = 0x80000000;
 186     } else {
 187         lowmem = 0xb0000000;
 188     }
 189
 190     /*
 191      * Handle the machine opt max-ram-below-4g.  It is basically doing
 192      * min(qemu limit, user limit).
 193      */
 194     if (!x86ms->max_ram_below_4g) {
 195         x86ms->max_ram_below_4g = 4 * GiB;
 196     }
 197     if (lowmem > x86ms->max_ram_below_4g) {
 198         lowmem = x86ms->max_ram_below_4g;
 199         if (machine->ram_size - lowmem > lowmem &&
 200             lowmem & (1 * GiB - 1)) {
 201             warn_report("There is possibly poor performance as the ram size "
 202                         " (0x%" PRIx64 ") is more then twice the size of"
 203                         " max-ram-below-4g (%"PRIu64") and"
 204                         " max-ram-below-4g is not a multiple of 1G.",
 205                         (uint64_t)machine->ram_size, x86ms->max_ram_below_4g);
 206         }
 207     }
 208
 209     if (machine->ram_size > lowmem) {
 210         x86ms->above_4g_mem_size = machine->ram_size - lowmem;
 211         x86ms->below_4g_mem_size = lowmem;
 212     } else {
 213         x86ms->above_4g_mem_size = 0;
 214         x86ms->below_4g_mem_size = machine->ram_size;
 215     }
 216
 217     ram_below_4g = g_malloc(sizeof(*ram_below_4g));
 218     memory_region_init_alias(ram_below_4g, NULL, "ram-below-4g", machine->ram,
 219                              0, x86ms->below_4g_mem_size);
 220     memory_region_add_subregion(system_memory, 0, ram_below_4g);
 221
 222     e820_add_entry(0, x86ms->below_4g_mem_size, E820_RAM);
 223
 224     if (x86ms->above_4g_mem_size > 0) {
 225         ram_above_4g = g_malloc(sizeof(*ram_above_4g));
 226         memory_region_init_alias(ram_above_4g, NULL, "ram-above-4g",
 227                                  machine->ram,
 228                                  x86ms->below_4g_mem_size,
 229                                  x86ms->above_4g_mem_size);
 230         memory_region_add_subregion(system_memory, 0x100000000ULL,
 231                                     ram_above_4g);
 232         e820_add_entry(0x100000000ULL, x86ms->above_4g_mem_size, E820_RAM);
 233     }
 234
 235     fw_cfg = fw_cfg_init_io_dma(FW_CFG_IO_BASE, FW_CFG_IO_BASE + 4,
 236                                 &address_space_memory);
 237
 238     fw_cfg_add_i16(fw_cfg, FW_CFG_NB_CPUS, machine->smp.cpus);
 239     fw_cfg_add_i16(fw_cfg, FW_CFG_MAX_CPUS, machine->smp.max_cpus);
 240     fw_cfg_add_i64(fw_cfg, FW_CFG_RAM_SIZE, (uint64_t)machine->ram_size);
 241     fw_cfg_add_i32(fw_cfg, FW_CFG_IRQ0_OVERRIDE, kvm_allows_irq0_override());
 242     fw_cfg_add_bytes(fw_cfg, FW_CFG_E820_TABLE,
 243                      &e820_reserve, sizeof(e820_reserve));
 244     fw_cfg_add_file(fw_cfg, "etc/e820", e820_table,
 245                     sizeof(struct e820_entry) * e820_get_num_entries());
 246
 247     rom_set_fw(fw_cfg);
 248
 249     if (machine->kernel_filename != NULL) {
 250         x86_load_linux(x86ms, fw_cfg, 0, true, true);
 251     }
 252
 253     if (mms->option_roms) {
 254         for (i = 0; i < nb_option_roms; i++) {
 255             rom_add_option(option_rom[i].name, option_rom[i].bootindex);
 256         }
 257     }
 258
 259     x86ms->fw_cfg = fw_cfg;
 260     x86ms->ioapic_as = &address_space_memory;
 261 }
 262
 263 static gchar *microvm_get_mmio_cmdline(gchar *name)
 264 {
 265     gchar *cmdline;
 266     gchar *separator;
 267     long int index;
 268     int ret;
 269
 270     separator = g_strrstr(name, ".");
 271     if (!separator) {
 272         return NULL;
 273     }
 274
 275     if (qemu_strtol(separator + 1, NULL, 10, &index) != 0) {
 276         return NULL;
 277     }
 278
 279     cmdline = g_malloc0(VIRTIO_CMDLINE_MAXLEN);
 280     ret = g_snprintf(cmdline, VIRTIO_CMDLINE_MAXLEN,
 281                      " virtio_mmio.device=512@0x%lx:%ld",
 282                      VIRTIO_MMIO_BASE + index * 512,
 283                      VIRTIO_IRQ_BASE + index);
 284     if (ret < 0 || ret >= VIRTIO_CMDLINE_MAXLEN) {
 285         g_free(cmdline);
 286         return NULL;
 287     }
 288
 289     return cmdline;
 290 }
 291
 292 static void microvm_fix_kernel_cmdline(MachineState *machine)
 293 {
 294     X86MachineState *x86ms = X86_MACHINE(machine);
 295     BusState *bus;
 296     BusChild *kid;
 297     char *cmdline;
 298
 299     /*
 300      * Find MMIO transports with attached devices, and add them to the kernel
 301      * command line.
 302      *
 303      * Yes, this is a hack, but one that heavily improves the UX without
 304      * introducing any significant issues.
 305      */
 306     cmdline = g_strdup(machine->kernel_cmdline);
 307     bus = sysbus_get_default();
 308     QTAILQ_FOREACH(kid, &bus->children, sibling) {
 309         DeviceState *dev = kid->child;
 310         ObjectClass *class = object_get_class(OBJECT(dev));
 311
 312         if (class == object_class_by_name(TYPE_VIRTIO_MMIO)) {
 313             VirtIOMMIOProxy *mmio = VIRTIO_MMIO(OBJECT(dev));
 314             VirtioBusState *mmio_virtio_bus = &mmio->bus;
 315             BusState *mmio_bus = &mmio_virtio_bus->parent_obj;
 316
 317             if (!QTAILQ_EMPTY(&mmio_bus->children)) {
 318                 gchar *mmio_cmdline = microvm_get_mmio_cmdline(mmio_bus->name);
 319                 if (mmio_cmdline) {
 320                     char *newcmd = g_strjoin(NULL, cmdline, mmio_cmdline, NULL);
 321                     g_free(mmio_cmdline);
 322                     g_free(cmdline);
 323                     cmdline = newcmd;
 324                 }
 325             }
 326         }
 327     }
 328
 329     fw_cfg_modify_i32(x86ms->fw_cfg, FW_CFG_CMDLINE_SIZE, strlen(cmdline) + 1);
 330     fw_cfg_modify_string(x86ms->fw_cfg, FW_CFG_CMDLINE_DATA, cmdline);
 331
 332     g_free(cmdline);
 333 }
 334
 335 static void microvm_machine_state_init(MachineState *machine)
 336 {
 337     MicrovmMachineState *mms = MICROVM_MACHINE(machine);
 338     X86MachineState *x86ms = X86_MACHINE(machine);
 339     Error *local_err = NULL;
 340
 341     microvm_memory_init(mms);
 342
 343     x86_cpus_init(x86ms, CPU_VERSION_LATEST);
 344     if (local_err) {
 345         error_report_err(local_err);
 346         exit(1);
 347     }
 348
 349     microvm_devices_init(mms);
 350 }
 351
 352 static void microvm_machine_reset(MachineState *machine)
 353 {
 354     MicrovmMachineState *mms = MICROVM_MACHINE(machine);
 355     CPUState *cs;
 356     X86CPU *cpu;
 357
 358     if (machine->kernel_filename != NULL &&
 359         mms->auto_kernel_cmdline && !mms->kernel_cmdline_fixed) {
 360         microvm_fix_kernel_cmdline(machine);
 361         mms->kernel_cmdline_fixed = true;
 362     }
 363
 364     qemu_devices_reset();
 365
 366     CPU_FOREACH(cs) {
 367         cpu = X86_CPU(cs);
 368
 369         if (cpu->apic_state) {
 370             device_legacy_reset(cpu->apic_state);
 371         }
 372     }
 373 }
 374
 375 static void microvm_machine_get_pic(Object *obj, Visitor *v, const char *name,
 376                                     void *opaque, Error **errp)
 377 {
 378     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 379     OnOffAuto pic = mms->pic;
 380
 381     visit_type_OnOffAuto(v, name, &pic, errp);
 382 }
 383
 384 static void microvm_machine_set_pic(Object *obj, Visitor *v, const char *name,
 385                                     void *opaque, Error **errp)
 386 {
 387     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 388
 389     visit_type_OnOffAuto(v, name, &mms->pic, errp);
 390 }
 391
 392 static void microvm_machine_get_pit(Object *obj, Visitor *v, const char *name,
 393                                     void *opaque, Error **errp)
 394 {
 395     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 396     OnOffAuto pit = mms->pit;
 397
 398     visit_type_OnOffAuto(v, name, &pit, errp);
 399 }
 400
 401 static void microvm_machine_set_pit(Object *obj, Visitor *v, const char *name,
 402                                     void *opaque, Error **errp)
 403 {
 404     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 405
 406     visit_type_OnOffAuto(v, name, &mms->pit, errp);
 407 }
 408
 409 static void microvm_machine_get_rtc(Object *obj, Visitor *v, const char *name,
 410                                     void *opaque, Error **errp)
 411 {
 412     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 413     OnOffAuto rtc = mms->rtc;
 414
 415     visit_type_OnOffAuto(v, name, &rtc, errp);
 416 }
 417
 418 static void microvm_machine_set_rtc(Object *obj, Visitor *v, const char *name,
 419                                     void *opaque, Error **errp)
 420 {
 421     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 422
 423     visit_type_OnOffAuto(v, name, &mms->rtc, errp);
 424 }
 425
 426 static bool microvm_machine_get_isa_serial(Object *obj, Error **errp)
 427 {
 428     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 429
 430     return mms->isa_serial;
 431 }
 432
 433 static void microvm_machine_set_isa_serial(Object *obj, bool value,
 434                                            Error **errp)
 435 {
 436     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 437
 438     mms->isa_serial = value;
 439 }
 440
 441 static bool microvm_machine_get_option_roms(Object *obj, Error **errp)
 442 {
 443     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 444
 445     return mms->option_roms;
 446 }
 447
 448 static void microvm_machine_set_option_roms(Object *obj, bool value,
 449                                             Error **errp)
 450 {
 451     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 452
 453     mms->option_roms = value;
 454 }
 455
 456 static bool microvm_machine_get_auto_kernel_cmdline(Object *obj, Error **errp)
 457 {
 458     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 459
 460     return mms->auto_kernel_cmdline;
 461 }
 462
 463 static void microvm_machine_set_auto_kernel_cmdline(Object *obj, bool value,
 464                                                     Error **errp)
 465 {
 466     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 467
 468     mms->auto_kernel_cmdline = value;
 469 }
 470
 471 static void microvm_machine_initfn(Object *obj)
 472 {
 473     MicrovmMachineState *mms = MICROVM_MACHINE(obj);
 474
 475     /* Configuration */
 476     mms->pic = ON_OFF_AUTO_AUTO;
 477     mms->pit = ON_OFF_AUTO_AUTO;
 478     mms->rtc = ON_OFF_AUTO_AUTO;
 479     mms->isa_serial = true;
 480     mms->option_roms = true;
 481     mms->auto_kernel_cmdline = true;
 482
 483     /* State */
 484     mms->kernel_cmdline_fixed = false;
 485 }
 486
 487 static void microvm_class_init(ObjectClass *oc, void *data)
 488 {
 489     MachineClass *mc = MACHINE_CLASS(oc);
 490
 491     mc->init = microvm_machine_state_init;
 492
 493     mc->family = "microvm_i386";
 494     mc->desc = "microvm (i386)";
 495     mc->units_per_default_bus = 1;
 496     mc->no_floppy = 1;
 497     mc->max_cpus = 288;
 498     mc->has_hotpluggable_cpus = false;
 499     mc->auto_enable_numa_with_memhp = false;
 500     mc->default_cpu_type = TARGET_DEFAULT_CPU_TYPE;
 501     mc->nvdimm_supported = false;
 502     mc->default_ram_id = "microvm.ram";
 503
 504     /* Avoid relying too much on kernel components */
 505     mc->default_kernel_irqchip_split = true;
 506
 507     /* Machine class handlers */
 508     mc->reset = microvm_machine_reset;
 509
 510     object_class_property_add(oc, MICROVM_MACHINE_PIC, "OnOffAuto",
 511                               microvm_machine_get_pic,
 512                               microvm_machine_set_pic,
 513                               NULL, NULL);
 514     object_class_property_set_description(oc, MICROVM_MACHINE_PIC,
 515         "Enable i8259 PIC");
 516
 517     object_class_property_add(oc, MICROVM_MACHINE_PIT, "OnOffAuto",
 518                               microvm_machine_get_pit,
 519                               microvm_machine_set_pit,
 520                               NULL, NULL);
 521     object_class_property_set_description(oc, MICROVM_MACHINE_PIT,
 522         "Enable i8254 PIT");
 523
 524     object_class_property_add(oc, MICROVM_MACHINE_RTC, "OnOffAuto",
 525                               microvm_machine_get_rtc,
 526                               microvm_machine_set_rtc,
 527                               NULL, NULL);
 528     object_class_property_set_description(oc, MICROVM_MACHINE_RTC,
 529         "Enable MC146818 RTC");
 530
 531     object_class_property_add_bool(oc, MICROVM_MACHINE_ISA_SERIAL,
 532                                    microvm_machine_get_isa_serial,
 533                                    microvm_machine_set_isa_serial);
 534     object_class_property_set_description(oc, MICROVM_MACHINE_ISA_SERIAL,
 535         "Set off to disable the instantiation an ISA serial port");
 536
 537     object_class_property_add_bool(oc, MICROVM_MACHINE_OPTION_ROMS,
 538                                    microvm_machine_get_option_roms,
 539                                    microvm_machine_set_option_roms);
 540     object_class_property_set_description(oc, MICROVM_MACHINE_OPTION_ROMS,
 541         "Set off to disable loading option ROMs");
 542
 543     object_class_property_add_bool(oc, MICROVM_MACHINE_AUTO_KERNEL_CMDLINE,
 544                                    microvm_machine_get_auto_kernel_cmdline,
 545                                    microvm_machine_set_auto_kernel_cmdline);
 546     object_class_property_set_description(oc,
 547         MICROVM_MACHINE_AUTO_KERNEL_CMDLINE,
 548         "Set off to disable adding virtio-mmio devices to the kernel cmdline");
 549 }
 550
 551 static const TypeInfo microvm_machine_info = {
 552     .name          = TYPE_MICROVM_MACHINE,
 553     .parent        = TYPE_X86_MACHINE,
 554     .instance_size = sizeof(MicrovmMachineState),
 555     .instance_init = microvm_machine_initfn,
 556     .class_size    = sizeof(MicrovmMachineClass),
 557     .class_init    = microvm_class_init,
 558     .interfaces = (InterfaceInfo[]) {
 559          { }
 560     },
 561 };
 562
 563 static void microvm_machine_init(void)
 564 {
 565     type_register_static(&microvm_machine_info);
 566 }
 567 type_init(microvm_machine_init);