arch/x86/kernel/paravirt.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*  Paravirtualization interfaces
   3     Copyright (C) 2006 Rusty Russell IBM Corporation
   4
   5
   6     2007 - x86_64 support added by Glauber de Oliveira Costa, Red Hat Inc
   7 */
   8
   9 #include <linux/errno.h>
  10 #include <linux/init.h>
  11 #include <linux/export.h>
  12 #include <linux/efi.h>
  13 #include <linux/bcd.h>
  14 #include <linux/highmem.h>
  15 #include <linux/kprobes.h>
  16 #include <linux/pgtable.h>
  17
  18 #include <asm/bug.h>
  19 #include <asm/paravirt.h>
  20 #include <asm/debugreg.h>
  21 #include <asm/desc.h>
  22 #include <asm/setup.h>
  23 #include <asm/time.h>
  24 #include <asm/pgalloc.h>
  25 #include <asm/irq.h>
  26 #include <asm/delay.h>
  27 #include <asm/fixmap.h>
  28 #include <asm/apic.h>
  29 #include <asm/tlbflush.h>
  30 #include <asm/timer.h>
  31 #include <asm/special_insns.h>
  32 #include <asm/tlb.h>
  33 #include <asm/io_bitmap.h>
  34
  35 /*
  36  * nop stub, which must not clobber anything *including the stack* to
  37  * avoid confusing the entry prologues.
  38  */
  39 extern void _paravirt_nop(void);
  40 asm (".pushsection .entry.text, \"ax\"\n"
  41      ".global _paravirt_nop\n"
  42      "_paravirt_nop:\n\t"
  43      "ret\n\t"
  44      ".size _paravirt_nop, . - _paravirt_nop\n\t"
  45      ".type _paravirt_nop, @function\n\t"
  46      ".popsection");
  47
  48 void __init default_banner(void)
  49 {
  50         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
  51                pv_info.name);
  52 }
  53
  54 /* Undefined instruction for dealing with missing ops pointers. */
  55 static const unsigned char ud2a[] = { 0x0f, 0x0b };
  56
  57 struct branch {
  58         unsigned char opcode;
  59         u32 delta;
  60 } __attribute__((packed));
  61
  62 static unsigned paravirt_patch_call(void *insn_buff, const void *target,
  63                                     unsigned long addr, unsigned len)
  64 {
  65         const int call_len = 5;
  66         struct branch *b = insn_buff;
  67         unsigned long delta = (unsigned long)target - (addr+call_len);
  68
  69         if (len < call_len) {
  70                 pr_warn("paravirt: Failed to patch indirect CALL at %ps\n", (void *)addr);
  71                 /* Kernel might not be viable if patching fails, bail out: */
  72                 BUG_ON(1);
  73         }
  74
  75         b->opcode = 0xe8; /* call */
  76         b->delta = delta;
  77         BUILD_BUG_ON(sizeof(*b) != call_len);
  78
  79         return call_len;
  80 }
  81
  82 #ifdef CONFIG_PARAVIRT_XXL
  83 /* identity function, which can be inlined */
  84 u64 notrace _paravirt_ident_64(u64 x)
  85 {
  86         return x;
  87 }
  88
  89 static unsigned paravirt_patch_jmp(void *insn_buff, const void *target,
  90                                    unsigned long addr, unsigned len)
  91 {
  92         struct branch *b = insn_buff;
  93         unsigned long delta = (unsigned long)target - (addr+5);
  94
  95         if (len < 5) {
  96 #ifdef CONFIG_RETPOLINE
  97                 WARN_ONCE(1, "Failing to patch indirect JMP in %ps\n", (void *)addr);
  98 #endif
  99                 return len;     /* call too long for patch site */
 100         }
 101
 102         b->opcode = 0xe9;       /* jmp */
 103         b->delta = delta;
 104
 105         return 5;
 106 }
 107 #endif
 108
 109 DEFINE_STATIC_KEY_TRUE(virt_spin_lock_key);
 110
 111 void __init native_pv_lock_init(void)
 112 {
 113         if (!boot_cpu_has(X86_FEATURE_HYPERVISOR))
 114                 static_branch_disable(&virt_spin_lock_key);
 115 }
 116
 117 unsigned paravirt_patch_default(u8 type, void *insn_buff,
 118                                 unsigned long addr, unsigned len)
 119 {
 120         /*
 121          * Neat trick to map patch type back to the call within the
 122          * corresponding structure.
 123          */
 124         void *opfunc = *((void **)&pv_ops + type);
 125         unsigned ret;
 126
 127         if (opfunc == NULL)
 128                 /* If there's no function, patch it with a ud2a (BUG) */
 129                 ret = paravirt_patch_insns(insn_buff, len, ud2a, ud2a+sizeof(ud2a));
 130         else if (opfunc == _paravirt_nop)
 131                 ret = 0;
 132
 133 #ifdef CONFIG_PARAVIRT_XXL
 134         /* identity functions just return their single argument */
 135         else if (opfunc == _paravirt_ident_64)
 136                 ret = paravirt_patch_ident_64(insn_buff, len);
 137
 138         else if (type == PARAVIRT_PATCH(cpu.iret))
 139                 /* If operation requires a jmp, then jmp */
 140                 ret = paravirt_patch_jmp(insn_buff, opfunc, addr, len);
 141 #endif
 142         else
 143                 /* Otherwise call the function. */
 144                 ret = paravirt_patch_call(insn_buff, opfunc, addr, len);
 145
 146         return ret;
 147 }
 148
 149 unsigned paravirt_patch_insns(void *insn_buff, unsigned len,
 150                               const char *start, const char *end)
 151 {
 152         unsigned insn_len = end - start;
 153
 154         /* Alternative instruction is too large for the patch site and we cannot continue: */
 155         BUG_ON(insn_len > len || start == NULL);
 156
 157         memcpy(insn_buff, start, insn_len);
 158
 159         return insn_len;
 160 }
 161
 162 struct static_key paravirt_steal_enabled;
 163 struct static_key paravirt_steal_rq_enabled;
 164
 165 static u64 native_steal_clock(int cpu)
 166 {
 167         return 0;
 168 }
 169
 170 /* These are in entry.S */
 171 extern void native_iret(void);
 172
 173 static struct resource reserve_ioports = {
 174         .start = 0,
 175         .end = IO_SPACE_LIMIT,
 176         .name = "paravirt-ioport",
 177         .flags = IORESOURCE_IO | IORESOURCE_BUSY,
 178 };
 179
 180 /*
 181  * Reserve the whole legacy IO space to prevent any legacy drivers
 182  * from wasting time probing for their hardware.  This is a fairly
 183  * brute-force approach to disabling all non-virtual drivers.
 184  *
 185  * Note that this must be called very early to have any effect.
 186  */
 187 int paravirt_disable_iospace(void)
 188 {
 189         return request_resource(&ioport_resource, &reserve_ioports);
 190 }
 191
 192 static DEFINE_PER_CPU(enum paravirt_lazy_mode, paravirt_lazy_mode) = PARAVIRT_LAZY_NONE;
 193
 194 static inline void enter_lazy(enum paravirt_lazy_mode mode)
 195 {
 196         BUG_ON(this_cpu_read(paravirt_lazy_mode) != PARAVIRT_LAZY_NONE);
 197
 198         this_cpu_write(paravirt_lazy_mode, mode);
 199 }
 200
 201 static void leave_lazy(enum paravirt_lazy_mode mode)
 202 {
 203         BUG_ON(this_cpu_read(paravirt_lazy_mode) != mode);
 204
 205         this_cpu_write(paravirt_lazy_mode, PARAVIRT_LAZY_NONE);
 206 }
 207
 208 void paravirt_enter_lazy_mmu(void)
 209 {
 210         enter_lazy(PARAVIRT_LAZY_MMU);
 211 }
 212
 213 void paravirt_leave_lazy_mmu(void)
 214 {
 215         leave_lazy(PARAVIRT_LAZY_MMU);
 216 }
 217
 218 void paravirt_flush_lazy_mmu(void)
 219 {
 220         preempt_disable();
 221
 222         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU) {
 223                 arch_leave_lazy_mmu_mode();
 224                 arch_enter_lazy_mmu_mode();
 225         }
 226
 227         preempt_enable();
 228 }
 229
 230 #ifdef CONFIG_PARAVIRT_XXL
 231 void paravirt_start_context_switch(struct task_struct *prev)
 232 {
 233         BUG_ON(preemptible());
 234
 235         if (this_cpu_read(paravirt_lazy_mode) == PARAVIRT_LAZY_MMU) {
 236                 arch_leave_lazy_mmu_mode();
 237                 set_ti_thread_flag(task_thread_info(prev), TIF_LAZY_MMU_UPDATES);
 238         }
 239         enter_lazy(PARAVIRT_LAZY_CPU);
 240 }
 241
 242 void paravirt_end_context_switch(struct task_struct *next)
 243 {
 244         BUG_ON(preemptible());
 245
 246         leave_lazy(PARAVIRT_LAZY_CPU);
 247
 248         if (test_and_clear_ti_thread_flag(task_thread_info(next), TIF_LAZY_MMU_UPDATES))
 249                 arch_enter_lazy_mmu_mode();
 250 }
 251 #endif
 252
 253 enum paravirt_lazy_mode paravirt_get_lazy_mode(void)
 254 {
 255         if (in_interrupt())
 256                 return PARAVIRT_LAZY_NONE;
 257
 258         return this_cpu_read(paravirt_lazy_mode);
 259 }
 260
 261 struct pv_info pv_info = {
 262         .name = "bare hardware",
 263 #ifdef CONFIG_PARAVIRT_XXL
 264         .extra_user_64bit_cs = __USER_CS,
 265 #endif
 266 };
 267
 268 /* 64-bit pagetable entries */
 269 #define PTE_IDENT       __PV_IS_CALLEE_SAVE(_paravirt_ident_64)
 270
 271 struct paravirt_patch_template pv_ops = {
 272         /* Init ops. */
 273         .init.patch             = native_patch,
 274
 275         /* Time ops. */
 276         .time.sched_clock       = native_sched_clock,
 277         .time.steal_clock       = native_steal_clock,
 278
 279         /* Cpu ops. */
 280         .cpu.io_delay           = native_io_delay,
 281
 282 #ifdef CONFIG_PARAVIRT_XXL
 283         .cpu.cpuid              = native_cpuid,
 284         .cpu.get_debugreg       = native_get_debugreg,
 285         .cpu.set_debugreg       = native_set_debugreg,
 286         .cpu.read_cr0           = native_read_cr0,
 287         .cpu.write_cr0          = native_write_cr0,
 288         .cpu.write_cr4          = native_write_cr4,
 289         .cpu.wbinvd             = native_wbinvd,
 290         .cpu.read_msr           = native_read_msr,
 291         .cpu.write_msr          = native_write_msr,
 292         .cpu.read_msr_safe      = native_read_msr_safe,
 293         .cpu.write_msr_safe     = native_write_msr_safe,
 294         .cpu.read_pmc           = native_read_pmc,
 295         .cpu.load_tr_desc       = native_load_tr_desc,
 296         .cpu.set_ldt            = native_set_ldt,
 297         .cpu.load_gdt           = native_load_gdt,
 298         .cpu.load_idt           = native_load_idt,
 299         .cpu.store_tr           = native_store_tr,
 300         .cpu.load_tls           = native_load_tls,
 301         .cpu.load_gs_index      = native_load_gs_index,
 302         .cpu.write_ldt_entry    = native_write_ldt_entry,
 303         .cpu.write_gdt_entry    = native_write_gdt_entry,
 304         .cpu.write_idt_entry    = native_write_idt_entry,
 305
 306         .cpu.alloc_ldt          = paravirt_nop,
 307         .cpu.free_ldt           = paravirt_nop,
 308
 309         .cpu.load_sp0           = native_load_sp0,
 310
 311         .cpu.iret               = native_iret,
 312
 313 #ifdef CONFIG_X86_IOPL_IOPERM
 314         .cpu.invalidate_io_bitmap       = native_tss_invalidate_io_bitmap,
 315         .cpu.update_io_bitmap           = native_tss_update_io_bitmap,
 316 #endif
 317
 318         .cpu.start_context_switch       = paravirt_nop,
 319         .cpu.end_context_switch         = paravirt_nop,
 320
 321         /* Irq ops. */
 322         .irq.save_fl            = __PV_IS_CALLEE_SAVE(native_save_fl),
 323         .irq.irq_disable        = __PV_IS_CALLEE_SAVE(native_irq_disable),
 324         .irq.irq_enable         = __PV_IS_CALLEE_SAVE(native_irq_enable),
 325         .irq.safe_halt          = native_safe_halt,
 326         .irq.halt               = native_halt,
 327 #endif /* CONFIG_PARAVIRT_XXL */
 328
 329         /* Mmu ops. */
 330         .mmu.flush_tlb_user     = native_flush_tlb_local,
 331         .mmu.flush_tlb_kernel   = native_flush_tlb_global,
 332         .mmu.flush_tlb_one_user = native_flush_tlb_one_user,
 333         .mmu.flush_tlb_multi    = native_flush_tlb_multi,
 334         .mmu.tlb_remove_table   =
 335                         (void (*)(struct mmu_gather *, void *))tlb_remove_page,
 336
 337         .mmu.exit_mmap          = paravirt_nop,
 338
 339 #ifdef CONFIG_PARAVIRT_XXL
 340         .mmu.read_cr2           = __PV_IS_CALLEE_SAVE(native_read_cr2),
 341         .mmu.write_cr2          = native_write_cr2,
 342         .mmu.read_cr3           = __native_read_cr3,
 343         .mmu.write_cr3          = native_write_cr3,
 344
 345         .mmu.pgd_alloc          = __paravirt_pgd_alloc,
 346         .mmu.pgd_free           = paravirt_nop,
 347
 348         .mmu.alloc_pte          = paravirt_nop,
 349         .mmu.alloc_pmd          = paravirt_nop,
 350         .mmu.alloc_pud          = paravirt_nop,
 351         .mmu.alloc_p4d          = paravirt_nop,
 352         .mmu.release_pte        = paravirt_nop,
 353         .mmu.release_pmd        = paravirt_nop,
 354         .mmu.release_pud        = paravirt_nop,
 355         .mmu.release_p4d        = paravirt_nop,
 356
 357         .mmu.set_pte            = native_set_pte,
 358         .mmu.set_pmd            = native_set_pmd,
 359
 360         .mmu.ptep_modify_prot_start     = __ptep_modify_prot_start,
 361         .mmu.ptep_modify_prot_commit    = __ptep_modify_prot_commit,
 362
 363         .mmu.set_pud            = native_set_pud,
 364
 365         .mmu.pmd_val            = PTE_IDENT,
 366         .mmu.make_pmd           = PTE_IDENT,
 367
 368         .mmu.pud_val            = PTE_IDENT,
 369         .mmu.make_pud           = PTE_IDENT,
 370
 371         .mmu.set_p4d            = native_set_p4d,
 372
 373 #if CONFIG_PGTABLE_LEVELS >= 5
 374         .mmu.p4d_val            = PTE_IDENT,
 375         .mmu.make_p4d           = PTE_IDENT,
 376
 377         .mmu.set_pgd            = native_set_pgd,
 378 #endif /* CONFIG_PGTABLE_LEVELS >= 5 */
 379
 380         .mmu.pte_val            = PTE_IDENT,
 381         .mmu.pgd_val            = PTE_IDENT,
 382
 383         .mmu.make_pte           = PTE_IDENT,
 384         .mmu.make_pgd           = PTE_IDENT,
 385
 386         .mmu.dup_mmap           = paravirt_nop,
 387         .mmu.activate_mm        = paravirt_nop,
 388
 389         .mmu.lazy_mode = {
 390                 .enter          = paravirt_nop,
 391                 .leave          = paravirt_nop,
 392                 .flush          = paravirt_nop,
 393         },
 394
 395         .mmu.set_fixmap         = native_set_fixmap,
 396 #endif /* CONFIG_PARAVIRT_XXL */
 397
 398 #if defined(CONFIG_PARAVIRT_SPINLOCKS)
 399         /* Lock ops. */
 400 #ifdef CONFIG_SMP
 401         .lock.queued_spin_lock_slowpath = native_queued_spin_lock_slowpath,
 402         .lock.queued_spin_unlock        =
 403                                 PV_CALLEE_SAVE(__native_queued_spin_unlock),
 404         .lock.wait                      = paravirt_nop,
 405         .lock.kick                      = paravirt_nop,
 406         .lock.vcpu_is_preempted         =
 407                                 PV_CALLEE_SAVE(__native_vcpu_is_preempted),
 408 #endif /* SMP */
 409 #endif
 410 };
 411
 412 #ifdef CONFIG_PARAVIRT_XXL
 413 /* At this point, native_get/set_debugreg has real function entries */
 414 NOKPROBE_SYMBOL(native_get_debugreg);
 415 NOKPROBE_SYMBOL(native_set_debugreg);
 416 NOKPROBE_SYMBOL(native_load_idt);
 417 #endif
 418
 419 EXPORT_SYMBOL(pv_ops);
 420 EXPORT_SYMBOL_GPL(pv_info);