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 #include <linux/static_call.h>
  18
  19 #include <asm/bug.h>
  20 #include <asm/paravirt.h>
  21 #include <asm/debugreg.h>
  22 #include <asm/desc.h>
  23 #include <asm/setup.h>
  24 #include <asm/time.h>
  25 #include <asm/pgalloc.h>
  26 #include <asm/irq.h>
  27 #include <asm/delay.h>
  28 #include <asm/fixmap.h>
  29 #include <asm/apic.h>
  30 #include <asm/tlbflush.h>
  31 #include <asm/timer.h>
  32 #include <asm/special_insns.h>
  33 #include <asm/tlb.h>
  34 #include <asm/io_bitmap.h>
  35
  36 /*
  37  * nop stub, which must not clobber anything *including the stack* to
  38  * avoid confusing the entry prologues.
  39  */
  40 DEFINE_PARAVIRT_ASM(_paravirt_nop, "", .entry.text);
  41
  42 /* stub always returning 0. */
  43 DEFINE_PARAVIRT_ASM(paravirt_ret0, "xor %eax,%eax", .entry.text);
  44
  45 void __init default_banner(void)
  46 {
  47         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
  48                pv_info.name);
  49 }
  50
  51 /* Undefined instruction for dealing with missing ops pointers. */
  52 noinstr void paravirt_BUG(void)
  53 {
  54         BUG();
  55 }
  56
  57 static unsigned paravirt_patch_call(void *insn_buff, const void *target,
  58                                     unsigned long addr, unsigned len)
  59 {
  60         __text_gen_insn(insn_buff, CALL_INSN_OPCODE,
  61                         (void *)addr, target, CALL_INSN_SIZE);
  62         return CALL_INSN_SIZE;
  63 }
  64
  65 #ifdef CONFIG_PARAVIRT_XXL
  66 /* identity function, which can be inlined */
  67 u64 notrace _paravirt_ident_64(u64 x)
  68 {
  69         return x;
  70 }
  71 #endif
  72
  73 DEFINE_STATIC_KEY_TRUE(virt_spin_lock_key);
  74
  75 void __init native_pv_lock_init(void)
  76 {
  77         if (!boot_cpu_has(X86_FEATURE_HYPERVISOR))
  78                 static_branch_disable(&virt_spin_lock_key);
  79 }
  80
  81 unsigned int paravirt_patch(u8 type, void *insn_buff, unsigned long addr,
  82                             unsigned int len)
  83 {
  84         /*
  85          * Neat trick to map patch type back to the call within the
  86          * corresponding structure.
  87          */
  88         void *opfunc = *((void **)&pv_ops + type);
  89         unsigned ret;
  90
  91         if (opfunc == NULL)
  92                 /* If there's no function, patch it with paravirt_BUG() */
  93                 ret = paravirt_patch_call(insn_buff, paravirt_BUG, addr, len);
  94         else if (opfunc == _paravirt_nop)
  95                 ret = 0;
  96         else
  97                 /* Otherwise call the function. */
  98                 ret = paravirt_patch_call(insn_buff, opfunc, addr, len);
  99
 100         return ret;
 101 }
 102
 103 struct static_key paravirt_steal_enabled;
 104 struct static_key paravirt_steal_rq_enabled;
 105
 106 static u64 native_steal_clock(int cpu)
 107 {
 108         return 0;
 109 }
 110
 111 DEFINE_STATIC_CALL(pv_steal_clock, native_steal_clock);
 112 DEFINE_STATIC_CALL(pv_sched_clock, native_sched_clock);
 113
 114 void paravirt_set_sched_clock(u64 (*func)(void))
 115 {
 116         static_call_update(pv_sched_clock, func);
 117 }
 118
 119 /* These are in entry.S */
 120 static struct resource reserve_ioports = {
 121         .start = 0,
 122         .end = IO_SPACE_LIMIT,
 123         .name = "paravirt-ioport",
 124         .flags = IORESOURCE_IO | IORESOURCE_BUSY,
 125 };
 126
 127 /*
 128  * Reserve the whole legacy IO space to prevent any legacy drivers
 129  * from wasting time probing for their hardware.  This is a fairly
 130  * brute-force approach to disabling all non-virtual drivers.
 131  *
 132  * Note that this must be called very early to have any effect.
 133  */
 134 int paravirt_disable_iospace(void)
 135 {
 136         return request_resource(&ioport_resource, &reserve_ioports);
 137 }
 138
 139 static DEFINE_PER_CPU(enum paravirt_lazy_mode, paravirt_lazy_mode) = PARAVIRT_LAZY_NONE;
 140
 141 static inline void enter_lazy(enum paravirt_lazy_mode mode)
 142 {
 143         BUG_ON(this_cpu_read(paravirt_lazy_mode) != PARAVIRT_LAZY_NONE);
 144
 145         this_cpu_write(paravirt_lazy_mode, mode);
 146 }
 147
 148 static void leave_lazy(enum paravirt_lazy_mode mode)
 149 {
 150         BUG_ON(this_cpu_read(paravirt_lazy_mode) != mode);
 151
 152         this_cpu_write(paravirt_lazy_mode, PARAVIRT_LAZY_NONE);
 153 }
 154
 155 void paravirt_enter_lazy_mmu(void)
 156 {
 157         enter_lazy(PARAVIRT_LAZY_MMU);
 158 }
 159
 160 void paravirt_leave_lazy_mmu(void)
 161 {
 162         leave_lazy(PARAVIRT_LAZY_MMU);
 163 }
 164
 165 void paravirt_flush_lazy_mmu(void)
 166 {
 167         preempt_disable();
 168
 169         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU) {
 170                 arch_leave_lazy_mmu_mode();
 171                 arch_enter_lazy_mmu_mode();
 172         }
 173
 174         preempt_enable();
 175 }
 176
 177 #ifdef CONFIG_PARAVIRT_XXL
 178 void paravirt_start_context_switch(struct task_struct *prev)
 179 {
 180         BUG_ON(preemptible());
 181
 182         if (this_cpu_read(paravirt_lazy_mode) == PARAVIRT_LAZY_MMU) {
 183                 arch_leave_lazy_mmu_mode();
 184                 set_ti_thread_flag(task_thread_info(prev), TIF_LAZY_MMU_UPDATES);
 185         }
 186         enter_lazy(PARAVIRT_LAZY_CPU);
 187 }
 188
 189 void paravirt_end_context_switch(struct task_struct *next)
 190 {
 191         BUG_ON(preemptible());
 192
 193         leave_lazy(PARAVIRT_LAZY_CPU);
 194
 195         if (test_and_clear_ti_thread_flag(task_thread_info(next), TIF_LAZY_MMU_UPDATES))
 196                 arch_enter_lazy_mmu_mode();
 197 }
 198
 199 static noinstr unsigned long pv_native_read_cr2(void)
 200 {
 201         return native_read_cr2();
 202 }
 203
 204 static noinstr void pv_native_write_cr2(unsigned long val)
 205 {
 206         native_write_cr2(val);
 207 }
 208
 209 static noinstr unsigned long pv_native_get_debugreg(int regno)
 210 {
 211         return native_get_debugreg(regno);
 212 }
 213
 214 static noinstr void pv_native_set_debugreg(int regno, unsigned long val)
 215 {
 216         native_set_debugreg(regno, val);
 217 }
 218
 219 static noinstr void pv_native_irq_enable(void)
 220 {
 221         native_irq_enable();
 222 }
 223
 224 static noinstr void pv_native_irq_disable(void)
 225 {
 226         native_irq_disable();
 227 }
 228 #endif
 229
 230 enum paravirt_lazy_mode paravirt_get_lazy_mode(void)
 231 {
 232         if (in_interrupt())
 233                 return PARAVIRT_LAZY_NONE;
 234
 235         return this_cpu_read(paravirt_lazy_mode);
 236 }
 237
 238 struct pv_info pv_info = {
 239         .name = "bare hardware",
 240 #ifdef CONFIG_PARAVIRT_XXL
 241         .extra_user_64bit_cs = __USER_CS,
 242 #endif
 243 };
 244
 245 /* 64-bit pagetable entries */
 246 #define PTE_IDENT       __PV_IS_CALLEE_SAVE(_paravirt_ident_64)
 247
 248 struct paravirt_patch_template pv_ops = {
 249         /* Cpu ops. */
 250         .cpu.io_delay           = native_io_delay,
 251
 252 #ifdef CONFIG_PARAVIRT_XXL
 253         .cpu.cpuid              = native_cpuid,
 254         .cpu.get_debugreg       = pv_native_get_debugreg,
 255         .cpu.set_debugreg       = pv_native_set_debugreg,
 256         .cpu.read_cr0           = native_read_cr0,
 257         .cpu.write_cr0          = native_write_cr0,
 258         .cpu.write_cr4          = native_write_cr4,
 259         .cpu.wbinvd             = native_wbinvd,
 260         .cpu.read_msr           = native_read_msr,
 261         .cpu.write_msr          = native_write_msr,
 262         .cpu.read_msr_safe      = native_read_msr_safe,
 263         .cpu.write_msr_safe     = native_write_msr_safe,
 264         .cpu.read_pmc           = native_read_pmc,
 265         .cpu.load_tr_desc       = native_load_tr_desc,
 266         .cpu.set_ldt            = native_set_ldt,
 267         .cpu.load_gdt           = native_load_gdt,
 268         .cpu.load_idt           = native_load_idt,
 269         .cpu.store_tr           = native_store_tr,
 270         .cpu.load_tls           = native_load_tls,
 271         .cpu.load_gs_index      = native_load_gs_index,
 272         .cpu.write_ldt_entry    = native_write_ldt_entry,
 273         .cpu.write_gdt_entry    = native_write_gdt_entry,
 274         .cpu.write_idt_entry    = native_write_idt_entry,
 275
 276         .cpu.alloc_ldt          = paravirt_nop,
 277         .cpu.free_ldt           = paravirt_nop,
 278
 279         .cpu.load_sp0           = native_load_sp0,
 280
 281 #ifdef CONFIG_X86_IOPL_IOPERM
 282         .cpu.invalidate_io_bitmap       = native_tss_invalidate_io_bitmap,
 283         .cpu.update_io_bitmap           = native_tss_update_io_bitmap,
 284 #endif
 285
 286         .cpu.start_context_switch       = paravirt_nop,
 287         .cpu.end_context_switch         = paravirt_nop,
 288
 289         /* Irq ops. */
 290         .irq.save_fl            = __PV_IS_CALLEE_SAVE(native_save_fl),
 291         .irq.irq_disable        = __PV_IS_CALLEE_SAVE(pv_native_irq_disable),
 292         .irq.irq_enable         = __PV_IS_CALLEE_SAVE(pv_native_irq_enable),
 293         .irq.safe_halt          = native_safe_halt,
 294         .irq.halt               = native_halt,
 295 #endif /* CONFIG_PARAVIRT_XXL */
 296
 297         /* Mmu ops. */
 298         .mmu.flush_tlb_user     = native_flush_tlb_local,
 299         .mmu.flush_tlb_kernel   = native_flush_tlb_global,
 300         .mmu.flush_tlb_one_user = native_flush_tlb_one_user,
 301         .mmu.flush_tlb_multi    = native_flush_tlb_multi,
 302         .mmu.tlb_remove_table   =
 303                         (void (*)(struct mmu_gather *, void *))tlb_remove_page,
 304
 305         .mmu.exit_mmap          = paravirt_nop,
 306         .mmu.notify_page_enc_status_changed     = paravirt_nop,
 307
 308 #ifdef CONFIG_PARAVIRT_XXL
 309         .mmu.read_cr2           = __PV_IS_CALLEE_SAVE(pv_native_read_cr2),
 310         .mmu.write_cr2          = pv_native_write_cr2,
 311         .mmu.read_cr3           = __native_read_cr3,
 312         .mmu.write_cr3          = native_write_cr3,
 313
 314         .mmu.pgd_alloc          = __paravirt_pgd_alloc,
 315         .mmu.pgd_free           = paravirt_nop,
 316
 317         .mmu.alloc_pte          = paravirt_nop,
 318         .mmu.alloc_pmd          = paravirt_nop,
 319         .mmu.alloc_pud          = paravirt_nop,
 320         .mmu.alloc_p4d          = paravirt_nop,
 321         .mmu.release_pte        = paravirt_nop,
 322         .mmu.release_pmd        = paravirt_nop,
 323         .mmu.release_pud        = paravirt_nop,
 324         .mmu.release_p4d        = paravirt_nop,
 325
 326         .mmu.set_pte            = native_set_pte,
 327         .mmu.set_pmd            = native_set_pmd,
 328
 329         .mmu.ptep_modify_prot_start     = __ptep_modify_prot_start,
 330         .mmu.ptep_modify_prot_commit    = __ptep_modify_prot_commit,
 331
 332         .mmu.set_pud            = native_set_pud,
 333
 334         .mmu.pmd_val            = PTE_IDENT,
 335         .mmu.make_pmd           = PTE_IDENT,
 336
 337         .mmu.pud_val            = PTE_IDENT,
 338         .mmu.make_pud           = PTE_IDENT,
 339
 340         .mmu.set_p4d            = native_set_p4d,
 341
 342 #if CONFIG_PGTABLE_LEVELS >= 5
 343         .mmu.p4d_val            = PTE_IDENT,
 344         .mmu.make_p4d           = PTE_IDENT,
 345
 346         .mmu.set_pgd            = native_set_pgd,
 347 #endif /* CONFIG_PGTABLE_LEVELS >= 5 */
 348
 349         .mmu.pte_val            = PTE_IDENT,
 350         .mmu.pgd_val            = PTE_IDENT,
 351
 352         .mmu.make_pte           = PTE_IDENT,
 353         .mmu.make_pgd           = PTE_IDENT,
 354
 355         .mmu.dup_mmap           = paravirt_nop,
 356         .mmu.activate_mm        = paravirt_nop,
 357
 358         .mmu.lazy_mode = {
 359                 .enter          = paravirt_nop,
 360                 .leave          = paravirt_nop,
 361                 .flush          = paravirt_nop,
 362         },
 363
 364         .mmu.set_fixmap         = native_set_fixmap,
 365 #endif /* CONFIG_PARAVIRT_XXL */
 366
 367 #if defined(CONFIG_PARAVIRT_SPINLOCKS)
 368         /* Lock ops. */
 369 #ifdef CONFIG_SMP
 370         .lock.queued_spin_lock_slowpath = native_queued_spin_lock_slowpath,
 371         .lock.queued_spin_unlock        =
 372                                 PV_CALLEE_SAVE(__native_queued_spin_unlock),
 373         .lock.wait                      = paravirt_nop,
 374         .lock.kick                      = paravirt_nop,
 375         .lock.vcpu_is_preempted         =
 376                                 PV_CALLEE_SAVE(__native_vcpu_is_preempted),
 377 #endif /* SMP */
 378 #endif
 379 };
 380
 381 #ifdef CONFIG_PARAVIRT_XXL
 382 NOKPROBE_SYMBOL(native_load_idt);
 383 #endif
 384
 385 EXPORT_SYMBOL(pv_ops);
 386 EXPORT_SYMBOL_GPL(pv_info);