1 /* 2 * Core of Xen paravirt_ops implementation. 3 * 4 * This file contains the xen_paravirt_ops structure itself, and the 5 * implementations for: 6 * - privileged instructions 7 * - interrupt flags 8 * - segment operations 9 * - booting and setup 10 * 11 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007 12 */ 13 14 #include <linux/cpu.h> 15 #include <linux/kernel.h> 16 #include <linux/init.h> 17 #include <linux/smp.h> 18 #include <linux/preempt.h> 19 #include <linux/hardirq.h> 20 #include <linux/percpu.h> 21 #include <linux/delay.h> 22 #include <linux/start_kernel.h> 23 #include <linux/sched.h> 24 #include <linux/kprobes.h> 25 #include <linux/bootmem.h> 26 #include <linux/module.h> 27 #include <linux/mm.h> 28 #include <linux/page-flags.h> 29 #include <linux/highmem.h> 30 #include <linux/console.h> 31 #include <linux/pci.h> 32 #include <linux/gfp.h> 33 34 #include <xen/xen.h> 35 #include <xen/interface/xen.h> 36 #include <xen/interface/version.h> 37 #include <xen/interface/physdev.h> 38 #include <xen/interface/vcpu.h> 39 #include <xen/interface/memory.h> 40 #include <xen/features.h> 41 #include <xen/page.h> 42 #include <xen/hvm.h> 43 #include <xen/hvc-console.h> 44 45 #include <asm/paravirt.h> 46 #include <asm/apic.h> 47 #include <asm/page.h> 48 #include <asm/xen/hypercall.h> 49 #include <asm/xen/hypervisor.h> 50 #include <asm/fixmap.h> 51 #include <asm/processor.h> 52 #include <asm/proto.h> 53 #include <asm/msr-index.h> 54 #include <asm/traps.h> 55 #include <asm/setup.h> 56 #include <asm/desc.h> 57 #include <asm/pgalloc.h> 58 #include <asm/pgtable.h> 59 #include <asm/tlbflush.h> 60 #include <asm/reboot.h> 61 #include <asm/setup.h> 62 #include <asm/stackprotector.h> 63 #include <asm/hypervisor.h> 64 65 #include "xen-ops.h" 66 #include "mmu.h" 67 #include "multicalls.h" 68 69 EXPORT_SYMBOL_GPL(hypercall_page); 70 71 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu); 72 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info); 73 74 enum xen_domain_type xen_domain_type = XEN_NATIVE; 75 EXPORT_SYMBOL_GPL(xen_domain_type); 76 77 struct start_info *xen_start_info; 78 EXPORT_SYMBOL_GPL(xen_start_info); 79 80 struct shared_info xen_dummy_shared_info; 81 82 void *xen_initial_gdt; 83 84 RESERVE_BRK(shared_info_page_brk, PAGE_SIZE); 85 __read_mostly int xen_have_vector_callback; 86 EXPORT_SYMBOL_GPL(xen_have_vector_callback); 87 88 /* 89 * Point at some empty memory to start with. We map the real shared_info 90 * page as soon as fixmap is up and running. 91 */ 92 struct shared_info *HYPERVISOR_shared_info = (void *)&xen_dummy_shared_info; 93 94 /* 95 * Flag to determine whether vcpu info placement is available on all 96 * VCPUs. We assume it is to start with, and then set it to zero on 97 * the first failure. This is because it can succeed on some VCPUs 98 * and not others, since it can involve hypervisor memory allocation, 99 * or because the guest failed to guarantee all the appropriate 100 * constraints on all VCPUs (ie buffer can't cross a page boundary). 101 * 102 * Note that any particular CPU may be using a placed vcpu structure, 103 * but we can only optimise if the all are. 104 * 105 * 0: not available, 1: available 106 */ 107 static int have_vcpu_info_placement = 1; 108 109 static void clamp_max_cpus(void) 110 { 111 #ifdef CONFIG_SMP 112 if (setup_max_cpus > MAX_VIRT_CPUS) 113 setup_max_cpus = MAX_VIRT_CPUS; 114 #endif 115 } 116 117 static void xen_vcpu_setup(int cpu) 118 { 119 struct vcpu_register_vcpu_info info; 120 int err; 121 struct vcpu_info *vcpup; 122 123 BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info); 124 125 if (cpu < MAX_VIRT_CPUS) 126 per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu]; 127 128 if (!have_vcpu_info_placement) { 129 if (cpu >= MAX_VIRT_CPUS) 130 clamp_max_cpus(); 131 return; 132 } 133 134 vcpup = &per_cpu(xen_vcpu_info, cpu); 135 info.mfn = arbitrary_virt_to_mfn(vcpup); 136 info.offset = offset_in_page(vcpup); 137 138 printk(KERN_DEBUG "trying to map vcpu_info %d at %p, mfn %llx, offset %d\n", 139 cpu, vcpup, info.mfn, info.offset); 140 141 /* Check to see if the hypervisor will put the vcpu_info 142 structure where we want it, which allows direct access via 143 a percpu-variable. */ 144 err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info); 145 146 if (err) { 147 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err); 148 have_vcpu_info_placement = 0; 149 clamp_max_cpus(); 150 } else { 151 /* This cpu is using the registered vcpu info, even if 152 later ones fail to. */ 153 per_cpu(xen_vcpu, cpu) = vcpup; 154 155 printk(KERN_DEBUG "cpu %d using vcpu_info at %p\n", 156 cpu, vcpup); 157 } 158 } 159 160 /* 161 * On restore, set the vcpu placement up again. 162 * If it fails, then we're in a bad state, since 163 * we can't back out from using it... 164 */ 165 void xen_vcpu_restore(void) 166 { 167 int cpu; 168 169 for_each_online_cpu(cpu) { 170 bool other_cpu = (cpu != smp_processor_id()); 171 172 if (other_cpu && 173 HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL)) 174 BUG(); 175 176 xen_setup_runstate_info(cpu); 177 178 if (have_vcpu_info_placement) 179 xen_vcpu_setup(cpu); 180 181 if (other_cpu && 182 HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL)) 183 BUG(); 184 } 185 } 186 187 static void __init xen_banner(void) 188 { 189 unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL); 190 struct xen_extraversion extra; 191 HYPERVISOR_xen_version(XENVER_extraversion, &extra); 192 193 printk(KERN_INFO "Booting paravirtualized kernel on %s\n", 194 pv_info.name); 195 printk(KERN_INFO "Xen version: %d.%d%s%s\n", 196 version >> 16, version & 0xffff, extra.extraversion, 197 xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : ""); 198 } 199 200 static __read_mostly unsigned int cpuid_leaf1_edx_mask = ~0; 201 static __read_mostly unsigned int cpuid_leaf1_ecx_mask = ~0; 202 203 static void xen_cpuid(unsigned int *ax, unsigned int *bx, 204 unsigned int *cx, unsigned int *dx) 205 { 206 unsigned maskebx = ~0; 207 unsigned maskecx = ~0; 208 unsigned maskedx = ~0; 209 210 /* 211 * Mask out inconvenient features, to try and disable as many 212 * unsupported kernel subsystems as possible. 213 */ 214 switch (*ax) { 215 case 1: 216 maskecx = cpuid_leaf1_ecx_mask; 217 maskedx = cpuid_leaf1_edx_mask; 218 break; 219 220 case 0xb: 221 /* Suppress extended topology stuff */ 222 maskebx = 0; 223 break; 224 } 225 226 asm(XEN_EMULATE_PREFIX "cpuid" 227 : "=a" (*ax), 228 "=b" (*bx), 229 "=c" (*cx), 230 "=d" (*dx) 231 : "0" (*ax), "2" (*cx)); 232 233 *bx &= maskebx; 234 *cx &= maskecx; 235 *dx &= maskedx; 236 } 237 238 static __init void xen_init_cpuid_mask(void) 239 { 240 unsigned int ax, bx, cx, dx; 241 242 cpuid_leaf1_edx_mask = 243 ~((1 << X86_FEATURE_MCE) | /* disable MCE */ 244 (1 << X86_FEATURE_MCA) | /* disable MCA */ 245 (1 << X86_FEATURE_ACC)); /* thermal monitoring */ 246 247 if (!xen_initial_domain()) 248 cpuid_leaf1_edx_mask &= 249 ~((1 << X86_FEATURE_APIC) | /* disable local APIC */ 250 (1 << X86_FEATURE_ACPI)); /* disable ACPI */ 251 252 ax = 1; 253 cx = 0; 254 xen_cpuid(&ax, &bx, &cx, &dx); 255 256 /* cpuid claims we support xsave; try enabling it to see what happens */ 257 if (cx & (1 << (X86_FEATURE_XSAVE % 32))) { 258 unsigned long cr4; 259 260 set_in_cr4(X86_CR4_OSXSAVE); 261 262 cr4 = read_cr4(); 263 264 if ((cr4 & X86_CR4_OSXSAVE) == 0) 265 cpuid_leaf1_ecx_mask &= ~(1 << (X86_FEATURE_XSAVE % 32)); 266 267 clear_in_cr4(X86_CR4_OSXSAVE); 268 } 269 } 270 271 static void xen_set_debugreg(int reg, unsigned long val) 272 { 273 HYPERVISOR_set_debugreg(reg, val); 274 } 275 276 static unsigned long xen_get_debugreg(int reg) 277 { 278 return HYPERVISOR_get_debugreg(reg); 279 } 280 281 static void xen_end_context_switch(struct task_struct *next) 282 { 283 xen_mc_flush(); 284 paravirt_end_context_switch(next); 285 } 286 287 static unsigned long xen_store_tr(void) 288 { 289 return 0; 290 } 291 292 /* 293 * Set the page permissions for a particular virtual address. If the 294 * address is a vmalloc mapping (or other non-linear mapping), then 295 * find the linear mapping of the page and also set its protections to 296 * match. 297 */ 298 static void set_aliased_prot(void *v, pgprot_t prot) 299 { 300 int level; 301 pte_t *ptep; 302 pte_t pte; 303 unsigned long pfn; 304 struct page *page; 305 306 ptep = lookup_address((unsigned long)v, &level); 307 BUG_ON(ptep == NULL); 308 309 pfn = pte_pfn(*ptep); 310 page = pfn_to_page(pfn); 311 312 pte = pfn_pte(pfn, prot); 313 314 if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0)) 315 BUG(); 316 317 if (!PageHighMem(page)) { 318 void *av = __va(PFN_PHYS(pfn)); 319 320 if (av != v) 321 if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0)) 322 BUG(); 323 } else 324 kmap_flush_unused(); 325 } 326 327 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries) 328 { 329 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE; 330 int i; 331 332 for(i = 0; i < entries; i += entries_per_page) 333 set_aliased_prot(ldt + i, PAGE_KERNEL_RO); 334 } 335 336 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries) 337 { 338 const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE; 339 int i; 340 341 for(i = 0; i < entries; i += entries_per_page) 342 set_aliased_prot(ldt + i, PAGE_KERNEL); 343 } 344 345 static void xen_set_ldt(const void *addr, unsigned entries) 346 { 347 struct mmuext_op *op; 348 struct multicall_space mcs = xen_mc_entry(sizeof(*op)); 349 350 op = mcs.args; 351 op->cmd = MMUEXT_SET_LDT; 352 op->arg1.linear_addr = (unsigned long)addr; 353 op->arg2.nr_ents = entries; 354 355 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF); 356 357 xen_mc_issue(PARAVIRT_LAZY_CPU); 358 } 359 360 static void xen_load_gdt(const struct desc_ptr *dtr) 361 { 362 unsigned long va = dtr->address; 363 unsigned int size = dtr->size + 1; 364 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE; 365 unsigned long frames[pages]; 366 int f; 367 368 /* 369 * A GDT can be up to 64k in size, which corresponds to 8192 370 * 8-byte entries, or 16 4k pages.. 371 */ 372 373 BUG_ON(size > 65536); 374 BUG_ON(va & ~PAGE_MASK); 375 376 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) { 377 int level; 378 pte_t *ptep; 379 unsigned long pfn, mfn; 380 void *virt; 381 382 /* 383 * The GDT is per-cpu and is in the percpu data area. 384 * That can be virtually mapped, so we need to do a 385 * page-walk to get the underlying MFN for the 386 * hypercall. The page can also be in the kernel's 387 * linear range, so we need to RO that mapping too. 388 */ 389 ptep = lookup_address(va, &level); 390 BUG_ON(ptep == NULL); 391 392 pfn = pte_pfn(*ptep); 393 mfn = pfn_to_mfn(pfn); 394 virt = __va(PFN_PHYS(pfn)); 395 396 frames[f] = mfn; 397 398 make_lowmem_page_readonly((void *)va); 399 make_lowmem_page_readonly(virt); 400 } 401 402 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct))) 403 BUG(); 404 } 405 406 /* 407 * load_gdt for early boot, when the gdt is only mapped once 408 */ 409 static __init void xen_load_gdt_boot(const struct desc_ptr *dtr) 410 { 411 unsigned long va = dtr->address; 412 unsigned int size = dtr->size + 1; 413 unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE; 414 unsigned long frames[pages]; 415 int f; 416 417 /* 418 * A GDT can be up to 64k in size, which corresponds to 8192 419 * 8-byte entries, or 16 4k pages.. 420 */ 421 422 BUG_ON(size > 65536); 423 BUG_ON(va & ~PAGE_MASK); 424 425 for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) { 426 pte_t pte; 427 unsigned long pfn, mfn; 428 429 pfn = virt_to_pfn(va); 430 mfn = pfn_to_mfn(pfn); 431 432 pte = pfn_pte(pfn, PAGE_KERNEL_RO); 433 434 if (HYPERVISOR_update_va_mapping((unsigned long)va, pte, 0)) 435 BUG(); 436 437 frames[f] = mfn; 438 } 439 440 if (HYPERVISOR_set_gdt(frames, size / sizeof(struct desc_struct))) 441 BUG(); 442 } 443 444 static void load_TLS_descriptor(struct thread_struct *t, 445 unsigned int cpu, unsigned int i) 446 { 447 struct desc_struct *gdt = get_cpu_gdt_table(cpu); 448 xmaddr_t maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]); 449 struct multicall_space mc = __xen_mc_entry(0); 450 451 MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]); 452 } 453 454 static void xen_load_tls(struct thread_struct *t, unsigned int cpu) 455 { 456 /* 457 * XXX sleazy hack: If we're being called in a lazy-cpu zone 458 * and lazy gs handling is enabled, it means we're in a 459 * context switch, and %gs has just been saved. This means we 460 * can zero it out to prevent faults on exit from the 461 * hypervisor if the next process has no %gs. Either way, it 462 * has been saved, and the new value will get loaded properly. 463 * This will go away as soon as Xen has been modified to not 464 * save/restore %gs for normal hypercalls. 465 * 466 * On x86_64, this hack is not used for %gs, because gs points 467 * to KERNEL_GS_BASE (and uses it for PDA references), so we 468 * must not zero %gs on x86_64 469 * 470 * For x86_64, we need to zero %fs, otherwise we may get an 471 * exception between the new %fs descriptor being loaded and 472 * %fs being effectively cleared at __switch_to(). 473 */ 474 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) { 475 #ifdef CONFIG_X86_32 476 lazy_load_gs(0); 477 #else 478 loadsegment(fs, 0); 479 #endif 480 } 481 482 xen_mc_batch(); 483 484 load_TLS_descriptor(t, cpu, 0); 485 load_TLS_descriptor(t, cpu, 1); 486 load_TLS_descriptor(t, cpu, 2); 487 488 xen_mc_issue(PARAVIRT_LAZY_CPU); 489 } 490 491 #ifdef CONFIG_X86_64 492 static void xen_load_gs_index(unsigned int idx) 493 { 494 if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx)) 495 BUG(); 496 } 497 #endif 498 499 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum, 500 const void *ptr) 501 { 502 xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]); 503 u64 entry = *(u64 *)ptr; 504 505 preempt_disable(); 506 507 xen_mc_flush(); 508 if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry)) 509 BUG(); 510 511 preempt_enable(); 512 } 513 514 static int cvt_gate_to_trap(int vector, const gate_desc *val, 515 struct trap_info *info) 516 { 517 unsigned long addr; 518 519 if (val->type != GATE_TRAP && val->type != GATE_INTERRUPT) 520 return 0; 521 522 info->vector = vector; 523 524 addr = gate_offset(*val); 525 #ifdef CONFIG_X86_64 526 /* 527 * Look for known traps using IST, and substitute them 528 * appropriately. The debugger ones are the only ones we care 529 * about. Xen will handle faults like double_fault and 530 * machine_check, so we should never see them. Warn if 531 * there's an unexpected IST-using fault handler. 532 */ 533 if (addr == (unsigned long)debug) 534 addr = (unsigned long)xen_debug; 535 else if (addr == (unsigned long)int3) 536 addr = (unsigned long)xen_int3; 537 else if (addr == (unsigned long)stack_segment) 538 addr = (unsigned long)xen_stack_segment; 539 else if (addr == (unsigned long)double_fault || 540 addr == (unsigned long)nmi) { 541 /* Don't need to handle these */ 542 return 0; 543 #ifdef CONFIG_X86_MCE 544 } else if (addr == (unsigned long)machine_check) { 545 return 0; 546 #endif 547 } else { 548 /* Some other trap using IST? */ 549 if (WARN_ON(val->ist != 0)) 550 return 0; 551 } 552 #endif /* CONFIG_X86_64 */ 553 info->address = addr; 554 555 info->cs = gate_segment(*val); 556 info->flags = val->dpl; 557 /* interrupt gates clear IF */ 558 if (val->type == GATE_INTERRUPT) 559 info->flags |= 1 << 2; 560 561 return 1; 562 } 563 564 /* Locations of each CPU's IDT */ 565 static DEFINE_PER_CPU(struct desc_ptr, idt_desc); 566 567 /* Set an IDT entry. If the entry is part of the current IDT, then 568 also update Xen. */ 569 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g) 570 { 571 unsigned long p = (unsigned long)&dt[entrynum]; 572 unsigned long start, end; 573 574 preempt_disable(); 575 576 start = __get_cpu_var(idt_desc).address; 577 end = start + __get_cpu_var(idt_desc).size + 1; 578 579 xen_mc_flush(); 580 581 native_write_idt_entry(dt, entrynum, g); 582 583 if (p >= start && (p + 8) <= end) { 584 struct trap_info info[2]; 585 586 info[1].address = 0; 587 588 if (cvt_gate_to_trap(entrynum, g, &info[0])) 589 if (HYPERVISOR_set_trap_table(info)) 590 BUG(); 591 } 592 593 preempt_enable(); 594 } 595 596 static void xen_convert_trap_info(const struct desc_ptr *desc, 597 struct trap_info *traps) 598 { 599 unsigned in, out, count; 600 601 count = (desc->size+1) / sizeof(gate_desc); 602 BUG_ON(count > 256); 603 604 for (in = out = 0; in < count; in++) { 605 gate_desc *entry = (gate_desc*)(desc->address) + in; 606 607 if (cvt_gate_to_trap(in, entry, &traps[out])) 608 out++; 609 } 610 traps[out].address = 0; 611 } 612 613 void xen_copy_trap_info(struct trap_info *traps) 614 { 615 const struct desc_ptr *desc = &__get_cpu_var(idt_desc); 616 617 xen_convert_trap_info(desc, traps); 618 } 619 620 /* Load a new IDT into Xen. In principle this can be per-CPU, so we 621 hold a spinlock to protect the static traps[] array (static because 622 it avoids allocation, and saves stack space). */ 623 static void xen_load_idt(const struct desc_ptr *desc) 624 { 625 static DEFINE_SPINLOCK(lock); 626 static struct trap_info traps[257]; 627 628 spin_lock(&lock); 629 630 __get_cpu_var(idt_desc) = *desc; 631 632 xen_convert_trap_info(desc, traps); 633 634 xen_mc_flush(); 635 if (HYPERVISOR_set_trap_table(traps)) 636 BUG(); 637 638 spin_unlock(&lock); 639 } 640 641 /* Write a GDT descriptor entry. Ignore LDT descriptors, since 642 they're handled differently. */ 643 static void xen_write_gdt_entry(struct desc_struct *dt, int entry, 644 const void *desc, int type) 645 { 646 preempt_disable(); 647 648 switch (type) { 649 case DESC_LDT: 650 case DESC_TSS: 651 /* ignore */ 652 break; 653 654 default: { 655 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]); 656 657 xen_mc_flush(); 658 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc)) 659 BUG(); 660 } 661 662 } 663 664 preempt_enable(); 665 } 666 667 /* 668 * Version of write_gdt_entry for use at early boot-time needed to 669 * update an entry as simply as possible. 670 */ 671 static __init void xen_write_gdt_entry_boot(struct desc_struct *dt, int entry, 672 const void *desc, int type) 673 { 674 switch (type) { 675 case DESC_LDT: 676 case DESC_TSS: 677 /* ignore */ 678 break; 679 680 default: { 681 xmaddr_t maddr = virt_to_machine(&dt[entry]); 682 683 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc)) 684 dt[entry] = *(struct desc_struct *)desc; 685 } 686 687 } 688 } 689 690 static void xen_load_sp0(struct tss_struct *tss, 691 struct thread_struct *thread) 692 { 693 struct multicall_space mcs = xen_mc_entry(0); 694 MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0); 695 xen_mc_issue(PARAVIRT_LAZY_CPU); 696 } 697 698 static void xen_set_iopl_mask(unsigned mask) 699 { 700 struct physdev_set_iopl set_iopl; 701 702 /* Force the change at ring 0. */ 703 set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3; 704 HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl); 705 } 706 707 static void xen_io_delay(void) 708 { 709 } 710 711 #ifdef CONFIG_X86_LOCAL_APIC 712 static u32 xen_apic_read(u32 reg) 713 { 714 return 0; 715 } 716 717 static void xen_apic_write(u32 reg, u32 val) 718 { 719 /* Warn to see if there's any stray references */ 720 WARN_ON(1); 721 } 722 723 static u64 xen_apic_icr_read(void) 724 { 725 return 0; 726 } 727 728 static void xen_apic_icr_write(u32 low, u32 id) 729 { 730 /* Warn to see if there's any stray references */ 731 WARN_ON(1); 732 } 733 734 static void xen_apic_wait_icr_idle(void) 735 { 736 return; 737 } 738 739 static u32 xen_safe_apic_wait_icr_idle(void) 740 { 741 return 0; 742 } 743 744 static void set_xen_basic_apic_ops(void) 745 { 746 apic->read = xen_apic_read; 747 apic->write = xen_apic_write; 748 apic->icr_read = xen_apic_icr_read; 749 apic->icr_write = xen_apic_icr_write; 750 apic->wait_icr_idle = xen_apic_wait_icr_idle; 751 apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle; 752 } 753 754 #endif 755 756 static void xen_clts(void) 757 { 758 struct multicall_space mcs; 759 760 mcs = xen_mc_entry(0); 761 762 MULTI_fpu_taskswitch(mcs.mc, 0); 763 764 xen_mc_issue(PARAVIRT_LAZY_CPU); 765 } 766 767 static DEFINE_PER_CPU(unsigned long, xen_cr0_value); 768 769 static unsigned long xen_read_cr0(void) 770 { 771 unsigned long cr0 = percpu_read(xen_cr0_value); 772 773 if (unlikely(cr0 == 0)) { 774 cr0 = native_read_cr0(); 775 percpu_write(xen_cr0_value, cr0); 776 } 777 778 return cr0; 779 } 780 781 static void xen_write_cr0(unsigned long cr0) 782 { 783 struct multicall_space mcs; 784 785 percpu_write(xen_cr0_value, cr0); 786 787 /* Only pay attention to cr0.TS; everything else is 788 ignored. */ 789 mcs = xen_mc_entry(0); 790 791 MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0); 792 793 xen_mc_issue(PARAVIRT_LAZY_CPU); 794 } 795 796 static void xen_write_cr4(unsigned long cr4) 797 { 798 cr4 &= ~X86_CR4_PGE; 799 cr4 &= ~X86_CR4_PSE; 800 801 native_write_cr4(cr4); 802 } 803 804 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high) 805 { 806 int ret; 807 808 ret = 0; 809 810 switch (msr) { 811 #ifdef CONFIG_X86_64 812 unsigned which; 813 u64 base; 814 815 case MSR_FS_BASE: which = SEGBASE_FS; goto set; 816 case MSR_KERNEL_GS_BASE: which = SEGBASE_GS_USER; goto set; 817 case MSR_GS_BASE: which = SEGBASE_GS_KERNEL; goto set; 818 819 set: 820 base = ((u64)high << 32) | low; 821 if (HYPERVISOR_set_segment_base(which, base) != 0) 822 ret = -EIO; 823 break; 824 #endif 825 826 case MSR_STAR: 827 case MSR_CSTAR: 828 case MSR_LSTAR: 829 case MSR_SYSCALL_MASK: 830 case MSR_IA32_SYSENTER_CS: 831 case MSR_IA32_SYSENTER_ESP: 832 case MSR_IA32_SYSENTER_EIP: 833 /* Fast syscall setup is all done in hypercalls, so 834 these are all ignored. Stub them out here to stop 835 Xen console noise. */ 836 break; 837 838 default: 839 ret = native_write_msr_safe(msr, low, high); 840 } 841 842 return ret; 843 } 844 845 void xen_setup_shared_info(void) 846 { 847 if (!xen_feature(XENFEAT_auto_translated_physmap)) { 848 set_fixmap(FIX_PARAVIRT_BOOTMAP, 849 xen_start_info->shared_info); 850 851 HYPERVISOR_shared_info = 852 (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP); 853 } else 854 HYPERVISOR_shared_info = 855 (struct shared_info *)__va(xen_start_info->shared_info); 856 857 #ifndef CONFIG_SMP 858 /* In UP this is as good a place as any to set up shared info */ 859 xen_setup_vcpu_info_placement(); 860 #endif 861 862 xen_setup_mfn_list_list(); 863 } 864 865 /* This is called once we have the cpu_possible_map */ 866 void xen_setup_vcpu_info_placement(void) 867 { 868 int cpu; 869 870 for_each_possible_cpu(cpu) 871 xen_vcpu_setup(cpu); 872 873 /* xen_vcpu_setup managed to place the vcpu_info within the 874 percpu area for all cpus, so make use of it */ 875 if (have_vcpu_info_placement) { 876 printk(KERN_INFO "Xen: using vcpu_info placement\n"); 877 878 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct); 879 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct); 880 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct); 881 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct); 882 pv_mmu_ops.read_cr2 = xen_read_cr2_direct; 883 } 884 } 885 886 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf, 887 unsigned long addr, unsigned len) 888 { 889 char *start, *end, *reloc; 890 unsigned ret; 891 892 start = end = reloc = NULL; 893 894 #define SITE(op, x) \ 895 case PARAVIRT_PATCH(op.x): \ 896 if (have_vcpu_info_placement) { \ 897 start = (char *)xen_##x##_direct; \ 898 end = xen_##x##_direct_end; \ 899 reloc = xen_##x##_direct_reloc; \ 900 } \ 901 goto patch_site 902 903 switch (type) { 904 SITE(pv_irq_ops, irq_enable); 905 SITE(pv_irq_ops, irq_disable); 906 SITE(pv_irq_ops, save_fl); 907 SITE(pv_irq_ops, restore_fl); 908 #undef SITE 909 910 patch_site: 911 if (start == NULL || (end-start) > len) 912 goto default_patch; 913 914 ret = paravirt_patch_insns(insnbuf, len, start, end); 915 916 /* Note: because reloc is assigned from something that 917 appears to be an array, gcc assumes it's non-null, 918 but doesn't know its relationship with start and 919 end. */ 920 if (reloc > start && reloc < end) { 921 int reloc_off = reloc - start; 922 long *relocp = (long *)(insnbuf + reloc_off); 923 long delta = start - (char *)addr; 924 925 *relocp += delta; 926 } 927 break; 928 929 default_patch: 930 default: 931 ret = paravirt_patch_default(type, clobbers, insnbuf, 932 addr, len); 933 break; 934 } 935 936 return ret; 937 } 938 939 static const struct pv_info xen_info __initdata = { 940 .paravirt_enabled = 1, 941 .shared_kernel_pmd = 0, 942 943 .name = "Xen", 944 }; 945 946 static const struct pv_init_ops xen_init_ops __initdata = { 947 .patch = xen_patch, 948 }; 949 950 static const struct pv_cpu_ops xen_cpu_ops __initdata = { 951 .cpuid = xen_cpuid, 952 953 .set_debugreg = xen_set_debugreg, 954 .get_debugreg = xen_get_debugreg, 955 956 .clts = xen_clts, 957 958 .read_cr0 = xen_read_cr0, 959 .write_cr0 = xen_write_cr0, 960 961 .read_cr4 = native_read_cr4, 962 .read_cr4_safe = native_read_cr4_safe, 963 .write_cr4 = xen_write_cr4, 964 965 .wbinvd = native_wbinvd, 966 967 .read_msr = native_read_msr_safe, 968 .write_msr = xen_write_msr_safe, 969 .read_tsc = native_read_tsc, 970 .read_pmc = native_read_pmc, 971 972 .iret = xen_iret, 973 .irq_enable_sysexit = xen_sysexit, 974 #ifdef CONFIG_X86_64 975 .usergs_sysret32 = xen_sysret32, 976 .usergs_sysret64 = xen_sysret64, 977 #endif 978 979 .load_tr_desc = paravirt_nop, 980 .set_ldt = xen_set_ldt, 981 .load_gdt = xen_load_gdt, 982 .load_idt = xen_load_idt, 983 .load_tls = xen_load_tls, 984 #ifdef CONFIG_X86_64 985 .load_gs_index = xen_load_gs_index, 986 #endif 987 988 .alloc_ldt = xen_alloc_ldt, 989 .free_ldt = xen_free_ldt, 990 991 .store_gdt = native_store_gdt, 992 .store_idt = native_store_idt, 993 .store_tr = xen_store_tr, 994 995 .write_ldt_entry = xen_write_ldt_entry, 996 .write_gdt_entry = xen_write_gdt_entry, 997 .write_idt_entry = xen_write_idt_entry, 998 .load_sp0 = xen_load_sp0, 999 1000 .set_iopl_mask = xen_set_iopl_mask, 1001 .io_delay = xen_io_delay, 1002 1003 /* Xen takes care of %gs when switching to usermode for us */ 1004 .swapgs = paravirt_nop, 1005 1006 .start_context_switch = paravirt_start_context_switch, 1007 .end_context_switch = xen_end_context_switch, 1008 }; 1009 1010 static const struct pv_apic_ops xen_apic_ops __initdata = { 1011 #ifdef CONFIG_X86_LOCAL_APIC 1012 .startup_ipi_hook = paravirt_nop, 1013 #endif 1014 }; 1015 1016 static void xen_reboot(int reason) 1017 { 1018 struct sched_shutdown r = { .reason = reason }; 1019 1020 #ifdef CONFIG_SMP 1021 smp_send_stop(); 1022 #endif 1023 1024 if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r)) 1025 BUG(); 1026 } 1027 1028 static void xen_restart(char *msg) 1029 { 1030 xen_reboot(SHUTDOWN_reboot); 1031 } 1032 1033 static void xen_emergency_restart(void) 1034 { 1035 xen_reboot(SHUTDOWN_reboot); 1036 } 1037 1038 static void xen_machine_halt(void) 1039 { 1040 xen_reboot(SHUTDOWN_poweroff); 1041 } 1042 1043 static void xen_crash_shutdown(struct pt_regs *regs) 1044 { 1045 xen_reboot(SHUTDOWN_crash); 1046 } 1047 1048 static int 1049 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr) 1050 { 1051 xen_reboot(SHUTDOWN_crash); 1052 return NOTIFY_DONE; 1053 } 1054 1055 static struct notifier_block xen_panic_block = { 1056 .notifier_call= xen_panic_event, 1057 }; 1058 1059 int xen_panic_handler_init(void) 1060 { 1061 atomic_notifier_chain_register(&panic_notifier_list, &xen_panic_block); 1062 return 0; 1063 } 1064 1065 static const struct machine_ops __initdata xen_machine_ops = { 1066 .restart = xen_restart, 1067 .halt = xen_machine_halt, 1068 .power_off = xen_machine_halt, 1069 .shutdown = xen_machine_halt, 1070 .crash_shutdown = xen_crash_shutdown, 1071 .emergency_restart = xen_emergency_restart, 1072 }; 1073 1074 /* 1075 * Set up the GDT and segment registers for -fstack-protector. Until 1076 * we do this, we have to be careful not to call any stack-protected 1077 * function, which is most of the kernel. 1078 */ 1079 static void __init xen_setup_stackprotector(void) 1080 { 1081 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry_boot; 1082 pv_cpu_ops.load_gdt = xen_load_gdt_boot; 1083 1084 setup_stack_canary_segment(0); 1085 switch_to_new_gdt(0); 1086 1087 pv_cpu_ops.write_gdt_entry = xen_write_gdt_entry; 1088 pv_cpu_ops.load_gdt = xen_load_gdt; 1089 } 1090 1091 /* First C function to be called on Xen boot */ 1092 asmlinkage void __init xen_start_kernel(void) 1093 { 1094 pgd_t *pgd; 1095 1096 if (!xen_start_info) 1097 return; 1098 1099 xen_domain_type = XEN_PV_DOMAIN; 1100 1101 /* Install Xen paravirt ops */ 1102 pv_info = xen_info; 1103 pv_init_ops = xen_init_ops; 1104 pv_cpu_ops = xen_cpu_ops; 1105 pv_apic_ops = xen_apic_ops; 1106 1107 x86_init.resources.memory_setup = xen_memory_setup; 1108 x86_init.oem.arch_setup = xen_arch_setup; 1109 x86_init.oem.banner = xen_banner; 1110 1111 xen_init_time_ops(); 1112 1113 /* 1114 * Set up some pagetable state before starting to set any ptes. 1115 */ 1116 1117 xen_init_mmu_ops(); 1118 1119 /* Prevent unwanted bits from being set in PTEs. */ 1120 __supported_pte_mask &= ~_PAGE_GLOBAL; 1121 if (!xen_initial_domain()) 1122 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD); 1123 1124 __supported_pte_mask |= _PAGE_IOMAP; 1125 1126 /* 1127 * Prevent page tables from being allocated in highmem, even 1128 * if CONFIG_HIGHPTE is enabled. 1129 */ 1130 __userpte_alloc_gfp &= ~__GFP_HIGHMEM; 1131 1132 /* Work out if we support NX */ 1133 x86_configure_nx(); 1134 1135 xen_setup_features(); 1136 1137 /* Get mfn list */ 1138 if (!xen_feature(XENFEAT_auto_translated_physmap)) 1139 xen_build_dynamic_phys_to_machine(); 1140 1141 /* 1142 * Set up kernel GDT and segment registers, mainly so that 1143 * -fstack-protector code can be executed. 1144 */ 1145 xen_setup_stackprotector(); 1146 1147 xen_init_irq_ops(); 1148 xen_init_cpuid_mask(); 1149 1150 #ifdef CONFIG_X86_LOCAL_APIC 1151 /* 1152 * set up the basic apic ops. 1153 */ 1154 set_xen_basic_apic_ops(); 1155 #endif 1156 1157 if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) { 1158 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start; 1159 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit; 1160 } 1161 1162 machine_ops = xen_machine_ops; 1163 1164 /* 1165 * The only reliable way to retain the initial address of the 1166 * percpu gdt_page is to remember it here, so we can go and 1167 * mark it RW later, when the initial percpu area is freed. 1168 */ 1169 xen_initial_gdt = &per_cpu(gdt_page, 0); 1170 1171 xen_smp_init(); 1172 1173 pgd = (pgd_t *)xen_start_info->pt_base; 1174 1175 if (!xen_initial_domain()) 1176 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD); 1177 1178 __supported_pte_mask |= _PAGE_IOMAP; 1179 /* Don't do the full vcpu_info placement stuff until we have a 1180 possible map and a non-dummy shared_info. */ 1181 per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0]; 1182 1183 local_irq_disable(); 1184 early_boot_irqs_off(); 1185 1186 xen_raw_console_write("mapping kernel into physical memory\n"); 1187 pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages); 1188 1189 init_mm.pgd = pgd; 1190 1191 /* keep using Xen gdt for now; no urgent need to change it */ 1192 1193 #ifdef CONFIG_X86_32 1194 pv_info.kernel_rpl = 1; 1195 if (xen_feature(XENFEAT_supervisor_mode_kernel)) 1196 pv_info.kernel_rpl = 0; 1197 #else 1198 pv_info.kernel_rpl = 0; 1199 #endif 1200 1201 /* set the limit of our address space */ 1202 xen_reserve_top(); 1203 1204 #ifdef CONFIG_X86_32 1205 /* set up basic CPUID stuff */ 1206 cpu_detect(&new_cpu_data); 1207 new_cpu_data.hard_math = 1; 1208 new_cpu_data.wp_works_ok = 1; 1209 new_cpu_data.x86_capability[0] = cpuid_edx(1); 1210 #endif 1211 1212 /* Poke various useful things into boot_params */ 1213 boot_params.hdr.type_of_loader = (9 << 4) | 0; 1214 boot_params.hdr.ramdisk_image = xen_start_info->mod_start 1215 ? __pa(xen_start_info->mod_start) : 0; 1216 boot_params.hdr.ramdisk_size = xen_start_info->mod_len; 1217 boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line); 1218 1219 if (!xen_initial_domain()) { 1220 add_preferred_console("xenboot", 0, NULL); 1221 add_preferred_console("tty", 0, NULL); 1222 add_preferred_console("hvc", 0, NULL); 1223 } else { 1224 /* Make sure ACS will be enabled */ 1225 pci_request_acs(); 1226 } 1227 1228 1229 xen_raw_console_write("about to get started...\n"); 1230 1231 xen_setup_runstate_info(0); 1232 1233 /* Start the world */ 1234 #ifdef CONFIG_X86_32 1235 i386_start_kernel(); 1236 #else 1237 x86_64_start_reservations((char *)__pa_symbol(&boot_params)); 1238 #endif 1239 } 1240 1241 static uint32_t xen_cpuid_base(void) 1242 { 1243 uint32_t base, eax, ebx, ecx, edx; 1244 char signature[13]; 1245 1246 for (base = 0x40000000; base < 0x40010000; base += 0x100) { 1247 cpuid(base, &eax, &ebx, &ecx, &edx); 1248 *(uint32_t *)(signature + 0) = ebx; 1249 *(uint32_t *)(signature + 4) = ecx; 1250 *(uint32_t *)(signature + 8) = edx; 1251 signature[12] = 0; 1252 1253 if (!strcmp("XenVMMXenVMM", signature) && ((eax - base) >= 2)) 1254 return base; 1255 } 1256 1257 return 0; 1258 } 1259 1260 static int init_hvm_pv_info(int *major, int *minor) 1261 { 1262 uint32_t eax, ebx, ecx, edx, pages, msr, base; 1263 u64 pfn; 1264 1265 base = xen_cpuid_base(); 1266 cpuid(base + 1, &eax, &ebx, &ecx, &edx); 1267 1268 *major = eax >> 16; 1269 *minor = eax & 0xffff; 1270 printk(KERN_INFO "Xen version %d.%d.\n", *major, *minor); 1271 1272 cpuid(base + 2, &pages, &msr, &ecx, &edx); 1273 1274 pfn = __pa(hypercall_page); 1275 wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32)); 1276 1277 xen_setup_features(); 1278 1279 pv_info = xen_info; 1280 pv_info.kernel_rpl = 0; 1281 1282 xen_domain_type = XEN_HVM_DOMAIN; 1283 1284 return 0; 1285 } 1286 1287 void xen_hvm_init_shared_info(void) 1288 { 1289 int cpu; 1290 struct xen_add_to_physmap xatp; 1291 static struct shared_info *shared_info_page = 0; 1292 1293 if (!shared_info_page) 1294 shared_info_page = (struct shared_info *) 1295 extend_brk(PAGE_SIZE, PAGE_SIZE); 1296 xatp.domid = DOMID_SELF; 1297 xatp.idx = 0; 1298 xatp.space = XENMAPSPACE_shared_info; 1299 xatp.gpfn = __pa(shared_info_page) >> PAGE_SHIFT; 1300 if (HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp)) 1301 BUG(); 1302 1303 HYPERVISOR_shared_info = (struct shared_info *)shared_info_page; 1304 1305 /* xen_vcpu is a pointer to the vcpu_info struct in the shared_info 1306 * page, we use it in the event channel upcall and in some pvclock 1307 * related functions. We don't need the vcpu_info placement 1308 * optimizations because we don't use any pv_mmu or pv_irq op on 1309 * HVM. 1310 * When xen_hvm_init_shared_info is run at boot time only vcpu 0 is 1311 * online but xen_hvm_init_shared_info is run at resume time too and 1312 * in that case multiple vcpus might be online. */ 1313 for_each_online_cpu(cpu) { 1314 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu]; 1315 } 1316 } 1317 1318 #ifdef CONFIG_XEN_PVHVM 1319 static int __cpuinit xen_hvm_cpu_notify(struct notifier_block *self, 1320 unsigned long action, void *hcpu) 1321 { 1322 int cpu = (long)hcpu; 1323 switch (action) { 1324 case CPU_UP_PREPARE: 1325 per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu]; 1326 break; 1327 default: 1328 break; 1329 } 1330 return NOTIFY_OK; 1331 } 1332 1333 static struct notifier_block __cpuinitdata xen_hvm_cpu_notifier = { 1334 .notifier_call = xen_hvm_cpu_notify, 1335 }; 1336 1337 static void __init xen_hvm_guest_init(void) 1338 { 1339 int r; 1340 int major, minor; 1341 1342 r = init_hvm_pv_info(&major, &minor); 1343 if (r < 0) 1344 return; 1345 1346 xen_hvm_init_shared_info(); 1347 1348 if (xen_feature(XENFEAT_hvm_callback_vector)) 1349 xen_have_vector_callback = 1; 1350 register_cpu_notifier(&xen_hvm_cpu_notifier); 1351 xen_unplug_emulated_devices(); 1352 have_vcpu_info_placement = 0; 1353 x86_init.irqs.intr_init = xen_init_IRQ; 1354 xen_hvm_init_time_ops(); 1355 xen_hvm_init_mmu_ops(); 1356 } 1357 1358 static bool __init xen_hvm_platform(void) 1359 { 1360 if (xen_pv_domain()) 1361 return false; 1362 1363 if (!xen_cpuid_base()) 1364 return false; 1365 1366 return true; 1367 } 1368 1369 const __refconst struct hypervisor_x86 x86_hyper_xen_hvm = { 1370 .name = "Xen HVM", 1371 .detect = xen_hvm_platform, 1372 .init_platform = xen_hvm_guest_init, 1373 }; 1374 EXPORT_SYMBOL(x86_hyper_xen_hvm); 1375 #endif 1376