sys/x86/xen/hvm.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause
   3  *
   4  * Copyright (c) 2008, 2013 Citrix Systems, Inc.
   5  * Copyright (c) 2012 Spectra Logic Corporation
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  */
  29
  30 #include <sys/param.h>
  31 #include <sys/bus.h>
  32 #include <sys/kernel.h>
  33 #include <sys/linker.h>
  34 #include <sys/malloc.h>
  35 #include <sys/proc.h>
  36 #include <sys/smp.h>
  37 #include <sys/systm.h>
  38
  39 #include <vm/vm.h>
  40 #include <vm/pmap.h>
  41 #include <vm/vm_param.h>
  42
  43 #include <dev/pci/pcivar.h>
  44
  45 #include <machine/_inttypes.h>
  46 #include <machine/cpufunc.h>
  47 #include <machine/cpu.h>
  48 #include <machine/md_var.h>
  49 #include <machine/metadata.h>
  50 #include <machine/smp.h>
  51
  52 #include <x86/apicreg.h>
  53
  54 #include <xen/xen-os.h>
  55 #include <xen/error.h>
  56 #include <xen/features.h>
  57 #include <xen/gnttab.h>
  58 #include <xen/hypervisor.h>
  59 #include <xen/hvm.h>
  60 #include <xen/xen_intr.h>
  61
  62 #include <contrib/xen/arch-x86/cpuid.h>
  63 #include <contrib/xen/hvm/params.h>
  64 #include <contrib/xen/vcpu.h>
  65
  66 /*--------------------------- Forward Declarations ---------------------------*/
  67 static void xen_hvm_cpu_init(void);
  68
  69 /*-------------------------------- Global Data -------------------------------*/
  70 #ifdef SMP
  71 struct cpu_ops xen_hvm_cpu_ops = {
  72         .cpu_init       = xen_hvm_cpu_init,
  73         .cpu_resume     = xen_hvm_cpu_init
  74 };
  75 #endif
  76
  77 static MALLOC_DEFINE(M_XENHVM, "xen_hvm", "Xen HVM PV Support");
  78
  79 /**
  80  * If non-zero, the hypervisor has been configured to use a direct
  81  * IDT event callback for interrupt injection.
  82  */
  83 int xen_vector_callback_enabled;
  84
  85 /**
  86  * Signal whether the vector injected for the event channel upcall requires to
  87  * be EOI'ed on the local APIC.
  88  */
  89 bool xen_evtchn_needs_ack;
  90
  91 /*------------------------------- Per-CPU Data -------------------------------*/
  92 DPCPU_DECLARE(struct vcpu_info *, vcpu_info);
  93
  94 /*------------------------------ Sysctl tunables -----------------------------*/
  95 int xen_disable_pv_disks = 0;
  96 int xen_disable_pv_nics = 0;
  97 TUNABLE_INT("hw.xen.disable_pv_disks", &xen_disable_pv_disks);
  98 TUNABLE_INT("hw.xen.disable_pv_nics", &xen_disable_pv_nics);
  99
 100 /*---------------------- XEN Hypervisor Probe and Setup ----------------------*/
 101
 102 void xen_emergency_print(const char *str, size_t size)
 103 {
 104         outsb(XEN_HVM_DEBUGCONS_IOPORT, str, size);
 105 }
 106
 107 uint32_t xen_cpuid_base;
 108
 109 static uint32_t
 110 xen_hvm_cpuid_base(void)
 111 {
 112         uint32_t base, regs[4];
 113
 114         for (base = 0x40000000; base < 0x40010000; base += 0x100) {
 115                 do_cpuid(base, regs);
 116                 if (!memcmp("XenVMMXenVMM", &regs[1], 12)
 117                     && (regs[0] - base) >= 2)
 118                         return (base);
 119         }
 120         return (0);
 121 }
 122
 123 static void
 124 hypervisor_quirks(unsigned int major, unsigned int minor)
 125 {
 126 #ifdef SMP
 127         if (((major < 4) || (major == 4 && minor <= 5)) &&
 128             msix_disable_migration == -1) {
 129                 /*
 130                  * Xen hypervisors prior to 4.6.0 do not properly
 131                  * handle updates to enabled MSI-X table entries,
 132                  * so disable MSI-X interrupt migration in that
 133                  * case.
 134                  */
 135                 if (bootverbose)
 136                         printf(
 137 "Disabling MSI-X interrupt migration due to Xen hypervisor bug.\n"
 138 "Set machdep.msix_disable_migration=0 to forcefully enable it.\n");
 139                 msix_disable_migration = 1;
 140         }
 141 #endif
 142 }
 143
 144 static void
 145 hypervisor_version(void)
 146 {
 147         uint32_t regs[4];
 148         int major, minor;
 149
 150         do_cpuid(xen_cpuid_base + 1, regs);
 151
 152         major = regs[0] >> 16;
 153         minor = regs[0] & 0xffff;
 154         printf("XEN: Hypervisor version %d.%d detected.\n", major, minor);
 155
 156         hypervisor_quirks(major, minor);
 157 }
 158
 159 /*
 160  * Allocate and fill in the hypcall page.
 161  */
 162 int
 163 xen_hvm_init_hypercall_stubs(enum xen_hvm_init_type init_type)
 164 {
 165         uint32_t regs[4];
 166
 167         if (xen_cpuid_base != 0)
 168                 /* Already setup. */
 169                 goto out;
 170
 171         xen_cpuid_base = xen_hvm_cpuid_base();
 172         if (xen_cpuid_base == 0)
 173                 return (ENXIO);
 174
 175         /*
 176          * Find the hypercall pages.
 177          */
 178         do_cpuid(xen_cpuid_base + 2, regs);
 179         if (regs[0] != 1)
 180                 return (EINVAL);
 181
 182         wrmsr(regs[1], (init_type == XEN_HVM_INIT_EARLY)
 183             ? (vm_paddr_t)((uintptr_t)&hypercall_page - KERNBASE)
 184             : vtophys(&hypercall_page));
 185
 186 out:
 187         hypervisor_version();
 188         return (0);
 189 }
 190
 191 /*
 192  * Translate linear to physical address when still running on the bootloader
 193  * created page-tables.
 194  */
 195 static vm_paddr_t
 196 early_init_vtop(void *addr)
 197 {
 198
 199         /*
 200          * Using a KASSERT won't print anything, as this is before console
 201          * initialization.
 202          */
 203         if (__predict_false((uintptr_t)addr < KERNBASE)) {
 204                 xc_printf("invalid linear address: %#lx\n", (uintptr_t)addr);
 205                 halt();
 206         }
 207
 208         return ((uintptr_t)addr - KERNBASE
 209 #ifdef __amd64__
 210             + kernphys - KERNLOAD
 211 #endif
 212             );
 213 }
 214
 215 static int
 216 map_shared_info(void)
 217 {
 218         /*
 219          * TODO shared info page should be mapped in an unpopulated (IOW:
 220          * non-RAM) address.  But finding one at this point in boot is
 221          * complicated, hence re-use a RAM address for the time being.  This
 222          * sadly causes super-page shattering in the second stage translation
 223          * page tables.
 224          */
 225         static union {
 226                 shared_info_t shared_info;
 227                 uint8_t raw[PAGE_SIZE];
 228         } shared_page __attribute__((aligned(PAGE_SIZE)));
 229         static struct xen_add_to_physmap xatp = {
 230             .domid = DOMID_SELF,
 231             .space = XENMAPSPACE_shared_info,
 232         };
 233         int rc;
 234
 235         _Static_assert(sizeof(shared_page) == PAGE_SIZE,
 236             "invalid Xen shared_info struct size");
 237
 238         if (xatp.gpfn == 0)
 239                 xatp.gpfn = atop(early_init_vtop(&shared_page.shared_info));
 240
 241         rc = HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp);
 242         if (rc != 0) {
 243                 xc_printf("cannot map shared info page: %d\n", rc);
 244                 HYPERVISOR_shared_info = NULL;
 245         } else if (HYPERVISOR_shared_info == NULL)
 246                 HYPERVISOR_shared_info = &shared_page.shared_info;
 247
 248         return (rc);
 249 }
 250
 251 static void
 252 fixup_console(void)
 253 {
 254         struct xen_platform_op op = {
 255                 .cmd = XENPF_get_dom0_console,
 256         };
 257         xenpf_dom0_console_t *console = &op.u.dom0_console;
 258         union {
 259                 struct efi_fb efi;
 260                 struct vbe_fb vbe;
 261         } *fb = NULL;
 262         int size;
 263         caddr_t kmdp;
 264
 265         kmdp = preload_search_by_type("elf kernel");
 266         if (kmdp == NULL)
 267                 kmdp = preload_search_by_type("elf64 kernel");
 268         if (kmdp == NULL) {
 269                 xc_printf("Unable to find kernel metadata\n");
 270                 return;
 271         }
 272
 273         size = HYPERVISOR_platform_op(&op);
 274         if (size < 0) {
 275                 xc_printf("Failed to get video console info: %d\n", size);
 276                 return;
 277         }
 278
 279         switch (console->video_type) {
 280         case XEN_VGATYPE_VESA_LFB:
 281                 fb = (__typeof__ (fb))preload_search_info(kmdp,
 282                     MODINFO_METADATA | MODINFOMD_VBE_FB);
 283
 284                 if (fb == NULL) {
 285                         xc_printf("No VBE FB in kernel metadata\n");
 286                         return;
 287                 }
 288
 289                 _Static_assert(offsetof(struct vbe_fb, fb_bpp) ==
 290                     offsetof(struct efi_fb, fb_mask_reserved) +
 291                     sizeof(fb->efi.fb_mask_reserved),
 292                     "Bad structure overlay\n");
 293                 fb->vbe.fb_bpp = console->u.vesa_lfb.bits_per_pixel;
 294                 /* FALLTHROUGH */
 295         case XEN_VGATYPE_EFI_LFB:
 296                 if (fb == NULL) {
 297                         fb = (__typeof__ (fb))preload_search_info(kmdp,
 298                             MODINFO_METADATA | MODINFOMD_EFI_FB);
 299                         if (fb == NULL) {
 300                                 xc_printf("No EFI FB in kernel metadata\n");
 301                                 return;
 302                         }
 303                 }
 304
 305                 fb->efi.fb_addr = console->u.vesa_lfb.lfb_base;
 306                 if (size >
 307                     offsetof(xenpf_dom0_console_t, u.vesa_lfb.ext_lfb_base))
 308                         fb->efi.fb_addr |=
 309                             (uint64_t)console->u.vesa_lfb.ext_lfb_base << 32;
 310                 fb->efi.fb_size = console->u.vesa_lfb.lfb_size << 16;
 311                 fb->efi.fb_height = console->u.vesa_lfb.height;
 312                 fb->efi.fb_width = console->u.vesa_lfb.width;
 313                 fb->efi.fb_stride = (console->u.vesa_lfb.bytes_per_line << 3) /
 314                     console->u.vesa_lfb.bits_per_pixel;
 315 #define FBMASK(c) \
 316     ((~0u << console->u.vesa_lfb.c ## _pos) & \
 317     (~0u >> (32 - console->u.vesa_lfb.c ## _pos - \
 318     console->u.vesa_lfb.c ## _size)))
 319                 fb->efi.fb_mask_red = FBMASK(red);
 320                 fb->efi.fb_mask_green = FBMASK(green);
 321                 fb->efi.fb_mask_blue = FBMASK(blue);
 322                 fb->efi.fb_mask_reserved = FBMASK(rsvd);
 323 #undef FBMASK
 324                 break;
 325
 326         default:
 327                 xc_printf("Video console type unsupported\n");
 328                 return;
 329         }
 330 }
 331
 332 /* Early initialization when running as a Xen guest. */
 333 void
 334 xen_early_init(void)
 335 {
 336         uint32_t regs[4];
 337         int rc;
 338
 339         xen_cpuid_base = xen_hvm_cpuid_base();
 340         if (xen_cpuid_base == 0)
 341                 return;
 342
 343         /* Find the hypercall pages. */
 344         do_cpuid(xen_cpuid_base + 2, regs);
 345         if (regs[0] != 1) {
 346                 xc_printf("Invalid number of hypercall pages %u\n",
 347                     regs[0]);
 348                 vm_guest = VM_GUEST_VM;
 349                 return;
 350         }
 351
 352         wrmsr(regs[1], early_init_vtop(&hypercall_page));
 353
 354         rc = map_shared_info();
 355         if (rc != 0) {
 356                 vm_guest = VM_GUEST_VM;
 357                 return;
 358         }
 359
 360         if (xen_initial_domain())
 361             /* Fixup video console information in case Xen changed the mode. */
 362             fixup_console();
 363 }
 364
 365 static void
 366 xen_hvm_init_shared_info_page(void)
 367 {
 368         struct xen_add_to_physmap xatp;
 369
 370         if (xen_pv_domain()) {
 371                 /*
 372                  * Already setup in the PV case, shared_info is passed inside
 373                  * of the start_info struct at start of day.
 374                  */
 375                 return;
 376         }
 377
 378         if (HYPERVISOR_shared_info == NULL) {
 379                 HYPERVISOR_shared_info = malloc(PAGE_SIZE, M_XENHVM, M_NOWAIT);
 380                 if (HYPERVISOR_shared_info == NULL)
 381                         panic("Unable to allocate Xen shared info page");
 382         }
 383
 384         xatp.domid = DOMID_SELF;
 385         xatp.idx = 0;
 386         xatp.space = XENMAPSPACE_shared_info;
 387         xatp.gpfn = vtophys(HYPERVISOR_shared_info) >> PAGE_SHIFT;
 388         if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp))
 389                 panic("HYPERVISOR_memory_op failed");
 390 }
 391
 392 static int
 393 set_percpu_callback(unsigned int vcpu)
 394 {
 395         struct xen_hvm_evtchn_upcall_vector vec;
 396         int error;
 397
 398         vec.vcpu = vcpu;
 399         vec.vector = IDT_EVTCHN;
 400         error = HYPERVISOR_hvm_op(HVMOP_set_evtchn_upcall_vector, &vec);
 401
 402         return (error != 0 ? xen_translate_error(error) : 0);
 403 }
 404
 405 /*
 406  * Tell the hypervisor how to contact us for event channel callbacks.
 407  */
 408 void
 409 xen_hvm_set_callback(device_t dev)
 410 {
 411         struct xen_hvm_param xhp;
 412         int irq;
 413
 414         if (xen_vector_callback_enabled)
 415                 return;
 416
 417         xhp.domid = DOMID_SELF;
 418         xhp.index = HVM_PARAM_CALLBACK_IRQ;
 419         if (xen_feature(XENFEAT_hvm_callback_vector) != 0) {
 420                 int error;
 421
 422                 error = set_percpu_callback(0);
 423                 if (error == 0) {
 424                         xen_evtchn_needs_ack = true;
 425                         /* Trick toolstack to think we are enlightened */
 426                         xhp.value = 1;
 427                 } else
 428                         xhp.value = HVM_CALLBACK_VECTOR(IDT_EVTCHN);
 429                 error = HYPERVISOR_hvm_op(HVMOP_set_param, &xhp);
 430                 if (error == 0) {
 431                         xen_vector_callback_enabled = 1;
 432                         return;
 433                 } else if (xen_evtchn_needs_ack)
 434                         panic("Unable to setup fake HVM param: %d", error);
 435
 436                 printf("Xen HVM callback vector registration failed (%d). "
 437                     "Falling back to emulated device interrupt\n", error);
 438         }
 439         xen_vector_callback_enabled = 0;
 440         if (dev == NULL) {
 441                 /*
 442                  * Called from early boot or resume.
 443                  * xenpci will invoke us again later.
 444                  */
 445                 return;
 446         }
 447
 448         irq = pci_get_irq(dev);
 449         if (irq < 16) {
 450                 xhp.value = HVM_CALLBACK_GSI(irq);
 451         } else {
 452                 u_int slot;
 453                 u_int pin;
 454
 455                 slot = pci_get_slot(dev);
 456                 pin = pci_get_intpin(dev) - 1;
 457                 xhp.value = HVM_CALLBACK_PCI_INTX(slot, pin);
 458         }
 459
 460         if (HYPERVISOR_hvm_op(HVMOP_set_param, &xhp) != 0)
 461                 panic("Can't set evtchn callback");
 462 }
 463
 464 #define XEN_MAGIC_IOPORT 0x10
 465 enum {
 466         XMI_MAGIC                        = 0x49d2,
 467         XMI_UNPLUG_IDE_DISKS             = 0x01,
 468         XMI_UNPLUG_NICS                  = 0x02,
 469         XMI_UNPLUG_IDE_EXCEPT_PRI_MASTER = 0x04
 470 };
 471
 472 static void
 473 xen_hvm_disable_emulated_devices(void)
 474 {
 475         u_short disable_devs = 0;
 476
 477         if (xen_pv_domain()) {
 478                 /*
 479                  * No emulated devices in the PV case, so no need to unplug
 480                  * anything.
 481                  */
 482                 if (xen_disable_pv_disks != 0 || xen_disable_pv_nics != 0)
 483                         printf("PV devices cannot be disabled in PV guests\n");
 484                 return;
 485         }
 486
 487         if (inw(XEN_MAGIC_IOPORT) != XMI_MAGIC)
 488                 return;
 489
 490         if (xen_disable_pv_disks == 0) {
 491                 if (bootverbose)
 492                         printf("XEN: disabling emulated disks\n");
 493                 disable_devs |= XMI_UNPLUG_IDE_DISKS;
 494         }
 495         if (xen_disable_pv_nics == 0) {
 496                 if (bootverbose)
 497                         printf("XEN: disabling emulated nics\n");
 498                 disable_devs |= XMI_UNPLUG_NICS;
 499         }
 500
 501         if (disable_devs != 0)
 502                 outw(XEN_MAGIC_IOPORT, disable_devs);
 503 }
 504
 505 static void
 506 xen_hvm_init(enum xen_hvm_init_type init_type)
 507 {
 508         int error;
 509         int i;
 510
 511         if (!xen_domain() ||
 512             init_type == XEN_HVM_INIT_CANCELLED_SUSPEND)
 513                 return;
 514
 515         error = xen_hvm_init_hypercall_stubs(init_type);
 516
 517         switch (init_type) {
 518         case XEN_HVM_INIT_LATE:
 519                 if (error != 0)
 520                         return;
 521
 522                 setup_xen_features();
 523 #ifdef SMP
 524                 cpu_ops = xen_hvm_cpu_ops;
 525 #endif
 526                 break;
 527         case XEN_HVM_INIT_RESUME:
 528                 if (error != 0)
 529                         panic("Unable to init Xen hypercall stubs on resume");
 530
 531                 /* Clear stale vcpu_info. */
 532                 CPU_FOREACH(i)
 533                         DPCPU_ID_SET(i, vcpu_info, NULL);
 534                 break;
 535         default:
 536                 panic("Unsupported HVM initialization type");
 537         }
 538
 539         xen_vector_callback_enabled = 0;
 540         xen_evtchn_needs_ack = false;
 541         xen_hvm_set_callback(NULL);
 542
 543         /*
 544          * On (PV)HVM domains we need to request the hypervisor to
 545          * fill the shared info page, for PVH guest the shared_info page
 546          * is passed inside the start_info struct and is already set, so this
 547          * functions are no-ops.
 548          */
 549         xen_hvm_init_shared_info_page();
 550         xen_hvm_disable_emulated_devices();
 551 }
 552
 553 void
 554 xen_hvm_suspend(void)
 555 {
 556 }
 557
 558 void
 559 xen_hvm_resume(bool suspend_cancelled)
 560 {
 561
 562         xen_hvm_init(suspend_cancelled ?
 563             XEN_HVM_INIT_CANCELLED_SUSPEND : XEN_HVM_INIT_RESUME);
 564
 565         /* Register vcpu_info area for CPU#0. */
 566         xen_hvm_cpu_init();
 567 }
 568
 569 static void
 570 xen_hvm_sysinit(void *arg __unused)
 571 {
 572         xen_hvm_init(XEN_HVM_INIT_LATE);
 573 }
 574 SYSINIT(xen_hvm_init, SI_SUB_HYPERVISOR, SI_ORDER_FIRST, xen_hvm_sysinit, NULL);
 575
 576 static void
 577 xen_hvm_cpu_init(void)
 578 {
 579         uint32_t regs[4];
 580         int rc;
 581
 582         if (!xen_domain())
 583                 return;
 584
 585         if (DPCPU_GET(vcpu_info) != NULL) {
 586                 /*
 587                  * vcpu_info is already set.  We're resuming
 588                  * from a failed migration and our pre-suspend
 589                  * configuration is still valid.
 590                  */
 591                 return;
 592         }
 593
 594         /*
 595          * Set vCPU ID. If available fetch the ID from CPUID, if not just use
 596          * the ACPI ID.
 597          */
 598         KASSERT(xen_cpuid_base != 0, ("Invalid base Xen CPUID leaf"));
 599         cpuid_count(xen_cpuid_base + 4, 0, regs);
 600         KASSERT((regs[0] & XEN_HVM_CPUID_VCPU_ID_PRESENT) ||
 601             !xen_pv_domain(),
 602             ("Xen PV domain without vcpu_id in cpuid"));
 603         PCPU_SET(vcpu_id, (regs[0] & XEN_HVM_CPUID_VCPU_ID_PRESENT) ?
 604             regs[1] : PCPU_GET(acpi_id));
 605
 606         if (xen_evtchn_needs_ack && !IS_BSP()) {
 607                 /*
 608                  * Setup the per-vpcu event channel upcall vector. This is only
 609                  * required when using the new HVMOP_set_evtchn_upcall_vector
 610                  * hypercall, which allows using a different vector for each
 611                  * vCPU. Note that FreeBSD uses the same vector for all vCPUs
 612                  * because it's not dynamically allocated.
 613                  */
 614                 rc = set_percpu_callback(PCPU_GET(vcpu_id));
 615                 if (rc != 0)
 616                         panic("Event channel upcall vector setup failed: %d",
 617                             rc);
 618         }
 619
 620         xen_setup_vcpu_info();
 621 }
 622 SYSINIT(xen_hvm_cpu_init, SI_SUB_INTR, SI_ORDER_FIRST, xen_hvm_cpu_init, NULL);
 623
 624 bool
 625 xen_has_iommu_maps(void)
 626 {
 627         uint32_t regs[4];
 628
 629         KASSERT(xen_cpuid_base != 0, ("Invalid base Xen CPUID leaf"));
 630         cpuid_count(xen_cpuid_base + 4, 0, regs);
 631
 632         return (regs[0] & XEN_HVM_CPUID_IOMMU_MAPPINGS);
 633 }