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 #include <asm/gsseg.h>
  36
  37 /*
  38  * nop stub, which must not clobber anything *including the stack* to
  39  * avoid confusing the entry prologues.
  40  */
  41 DEFINE_PARAVIRT_ASM(_paravirt_nop, "", .entry.text);
  42
  43 /* stub always returning 0. */
  44 DEFINE_PARAVIRT_ASM(paravirt_ret0, "xor %eax,%eax", .entry.text);
  45
  46 void __init default_banner(void)
  47 {
  48         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
  49                pv_info.name);
  50 }
  51
  52 /* Undefined instruction for dealing with missing ops pointers. */
  53 noinstr void paravirt_BUG(void)
  54 {
  55         BUG();
  56 }
  57
  58 static unsigned paravirt_patch_call(void *insn_buff, const void *target,
  59                                     unsigned long addr, unsigned len)
  60 {
  61         __text_gen_insn(insn_buff, CALL_INSN_OPCODE,
  62                         (void *)addr, target, CALL_INSN_SIZE);
  63         return CALL_INSN_SIZE;
  64 }
  65
  66 #ifdef CONFIG_PARAVIRT_XXL
  67 DEFINE_PARAVIRT_ASM(_paravirt_ident_64, "mov %rdi, %rax", .text);
  68 DEFINE_PARAVIRT_ASM(pv_native_save_fl, "pushf; pop %rax", .noinstr.text);
  69 DEFINE_PARAVIRT_ASM(pv_native_irq_disable, "cli", .noinstr.text);
  70 DEFINE_PARAVIRT_ASM(pv_native_irq_enable, "sti", .noinstr.text);
  71 DEFINE_PARAVIRT_ASM(pv_native_read_cr2, "mov %cr2, %rax", .noinstr.text);
  72 #endif
  73
  74 DEFINE_STATIC_KEY_TRUE(virt_spin_lock_key);
  75
  76 void __init native_pv_lock_init(void)
  77 {
  78         if (IS_ENABLED(CONFIG_PARAVIRT_SPINLOCKS) &&
  79             !boot_cpu_has(X86_FEATURE_HYPERVISOR))
  80                 static_branch_disable(&virt_spin_lock_key);
  81 }
  82
  83 static void native_tlb_remove_table(struct mmu_gather *tlb, void *table)
  84 {
  85         tlb_remove_page(tlb, table);
  86 }
  87
  88 unsigned int paravirt_patch(u8 type, void *insn_buff, unsigned long addr,
  89                             unsigned int len)
  90 {
  91         /*
  92          * Neat trick to map patch type back to the call within the
  93          * corresponding structure.
  94          */
  95         void *opfunc = *((void **)&pv_ops + type);
  96         unsigned ret;
  97
  98         if (opfunc == NULL)
  99                 /* If there's no function, patch it with paravirt_BUG() */
 100                 ret = paravirt_patch_call(insn_buff, paravirt_BUG, addr, len);
 101         else if (opfunc == _paravirt_nop)
 102                 ret = 0;
 103         else
 104                 /* Otherwise call the function. */
 105                 ret = paravirt_patch_call(insn_buff, opfunc, addr, len);
 106
 107         return ret;
 108 }
 109
 110 struct static_key paravirt_steal_enabled;
 111 struct static_key paravirt_steal_rq_enabled;
 112
 113 static u64 native_steal_clock(int cpu)
 114 {
 115         return 0;
 116 }
 117
 118 DEFINE_STATIC_CALL(pv_steal_clock, native_steal_clock);
 119 DEFINE_STATIC_CALL(pv_sched_clock, native_sched_clock);
 120
 121 void paravirt_set_sched_clock(u64 (*func)(void))
 122 {
 123         static_call_update(pv_sched_clock, func);
 124 }
 125
 126 /* These are in entry.S */
 127 static struct resource reserve_ioports = {
 128         .start = 0,
 129         .end = IO_SPACE_LIMIT,
 130         .name = "paravirt-ioport",
 131         .flags = IORESOURCE_IO | IORESOURCE_BUSY,
 132 };
 133
 134 /*
 135  * Reserve the whole legacy IO space to prevent any legacy drivers
 136  * from wasting time probing for their hardware.  This is a fairly
 137  * brute-force approach to disabling all non-virtual drivers.
 138  *
 139  * Note that this must be called very early to have any effect.
 140  */
 141 int paravirt_disable_iospace(void)
 142 {
 143         return request_resource(&ioport_resource, &reserve_ioports);
 144 }
 145
 146 static DEFINE_PER_CPU(enum paravirt_lazy_mode, paravirt_lazy_mode) = PARAVIRT_LAZY_NONE;
 147
 148 static inline void enter_lazy(enum paravirt_lazy_mode mode)
 149 {
 150         BUG_ON(this_cpu_read(paravirt_lazy_mode) != PARAVIRT_LAZY_NONE);
 151
 152         this_cpu_write(paravirt_lazy_mode, mode);
 153 }
 154
 155 static void leave_lazy(enum paravirt_lazy_mode mode)
 156 {
 157         BUG_ON(this_cpu_read(paravirt_lazy_mode) != mode);
 158
 159         this_cpu_write(paravirt_lazy_mode, PARAVIRT_LAZY_NONE);
 160 }
 161
 162 void paravirt_enter_lazy_mmu(void)
 163 {
 164         enter_lazy(PARAVIRT_LAZY_MMU);
 165 }
 166
 167 void paravirt_leave_lazy_mmu(void)
 168 {
 169         leave_lazy(PARAVIRT_LAZY_MMU);
 170 }
 171
 172 void paravirt_flush_lazy_mmu(void)
 173 {
 174         preempt_disable();
 175
 176         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU) {
 177                 arch_leave_lazy_mmu_mode();
 178                 arch_enter_lazy_mmu_mode();
 179         }
 180
 181         preempt_enable();
 182 }
 183
 184 #ifdef CONFIG_PARAVIRT_XXL
 185 void paravirt_start_context_switch(struct task_struct *prev)
 186 {
 187         BUG_ON(preemptible());
 188
 189         if (this_cpu_read(paravirt_lazy_mode) == PARAVIRT_LAZY_MMU) {
 190                 arch_leave_lazy_mmu_mode();
 191                 set_ti_thread_flag(task_thread_info(prev), TIF_LAZY_MMU_UPDATES);
 192         }
 193         enter_lazy(PARAVIRT_LAZY_CPU);
 194 }
 195
 196 void paravirt_end_context_switch(struct task_struct *next)
 197 {
 198         BUG_ON(preemptible());
 199
 200         leave_lazy(PARAVIRT_LAZY_CPU);
 201
 202         if (test_and_clear_ti_thread_flag(task_thread_info(next), TIF_LAZY_MMU_UPDATES))
 203                 arch_enter_lazy_mmu_mode();
 204 }
 205
 206 static noinstr void pv_native_write_cr2(unsigned long val)
 207 {
 208         native_write_cr2(val);
 209 }
 210
 211 static noinstr unsigned long pv_native_get_debugreg(int regno)
 212 {
 213         return native_get_debugreg(regno);
 214 }
 215
 216 static noinstr void pv_native_set_debugreg(int regno, unsigned long val)
 217 {
 218         native_set_debugreg(regno, val);
 219 }
 220
 221 noinstr void pv_native_wbinvd(void)
 222 {
 223         native_wbinvd();
 224 }
 225
 226 static noinstr void pv_native_safe_halt(void)
 227 {
 228         native_safe_halt();
 229 }
 230 #endif
 231
 232 enum paravirt_lazy_mode paravirt_get_lazy_mode(void)
 233 {
 234         if (in_interrupt())
 235                 return PARAVIRT_LAZY_NONE;
 236
 237         return this_cpu_read(paravirt_lazy_mode);
 238 }
 239
 240 struct pv_info pv_info = {
 241         .name = "bare hardware",
 242 #ifdef CONFIG_PARAVIRT_XXL
 243         .extra_user_64bit_cs = __USER_CS,
 244 #endif
 245 };
 246
 247 /* 64-bit pagetable entries */
 248 #define PTE_IDENT       __PV_IS_CALLEE_SAVE(_paravirt_ident_64)
 249
 250 struct paravirt_patch_template pv_ops = {
 251         /* Cpu ops. */
 252         .cpu.io_delay           = native_io_delay,
 253
 254 #ifdef CONFIG_PARAVIRT_XXL
 255         .cpu.cpuid              = native_cpuid,
 256         .cpu.get_debugreg       = pv_native_get_debugreg,
 257         .cpu.set_debugreg       = pv_native_set_debugreg,
 258         .cpu.read_cr0           = native_read_cr0,
 259         .cpu.write_cr0          = native_write_cr0,
 260         .cpu.write_cr4          = native_write_cr4,
 261         .cpu.wbinvd             = pv_native_wbinvd,
 262         .cpu.read_msr           = native_read_msr,
 263         .cpu.write_msr          = native_write_msr,
 264         .cpu.read_msr_safe      = native_read_msr_safe,
 265         .cpu.write_msr_safe     = native_write_msr_safe,
 266         .cpu.read_pmc           = native_read_pmc,
 267         .cpu.load_tr_desc       = native_load_tr_desc,
 268         .cpu.set_ldt            = native_set_ldt,
 269         .cpu.load_gdt           = native_load_gdt,
 270         .cpu.load_idt           = native_load_idt,
 271         .cpu.store_tr           = native_store_tr,
 272         .cpu.load_tls           = native_load_tls,
 273         .cpu.load_gs_index      = native_load_gs_index,
 274         .cpu.write_ldt_entry    = native_write_ldt_entry,
 275         .cpu.write_gdt_entry    = native_write_gdt_entry,
 276         .cpu.write_idt_entry    = native_write_idt_entry,
 277
 278         .cpu.alloc_ldt          = paravirt_nop,
 279         .cpu.free_ldt           = paravirt_nop,
 280
 281         .cpu.load_sp0           = native_load_sp0,
 282
 283 #ifdef CONFIG_X86_IOPL_IOPERM
 284         .cpu.invalidate_io_bitmap       = native_tss_invalidate_io_bitmap,
 285         .cpu.update_io_bitmap           = native_tss_update_io_bitmap,
 286 #endif
 287
 288         .cpu.start_context_switch       = paravirt_nop,
 289         .cpu.end_context_switch         = paravirt_nop,
 290
 291         /* Irq ops. */
 292         .irq.save_fl            = __PV_IS_CALLEE_SAVE(pv_native_save_fl),
 293         .irq.irq_disable        = __PV_IS_CALLEE_SAVE(pv_native_irq_disable),
 294         .irq.irq_enable         = __PV_IS_CALLEE_SAVE(pv_native_irq_enable),
 295         .irq.safe_halt          = pv_native_safe_halt,
 296         .irq.halt               = native_halt,
 297 #endif /* CONFIG_PARAVIRT_XXL */
 298
 299         /* Mmu ops. */
 300         .mmu.flush_tlb_user     = native_flush_tlb_local,
 301         .mmu.flush_tlb_kernel   = native_flush_tlb_global,
 302         .mmu.flush_tlb_one_user = native_flush_tlb_one_user,
 303         .mmu.flush_tlb_multi    = native_flush_tlb_multi,
 304         .mmu.tlb_remove_table   = native_tlb_remove_table,
 305
 306         .mmu.exit_mmap          = paravirt_nop,
 307         .mmu.notify_page_enc_status_changed     = paravirt_nop,
 308
 309 #ifdef CONFIG_PARAVIRT_XXL
 310         .mmu.read_cr2           = __PV_IS_CALLEE_SAVE(pv_native_read_cr2),
 311         .mmu.write_cr2          = pv_native_write_cr2,
 312         .mmu.read_cr3           = __native_read_cr3,
 313         .mmu.write_cr3          = native_write_cr3,
 314
 315         .mmu.pgd_alloc          = __paravirt_pgd_alloc,
 316         .mmu.pgd_free           = paravirt_nop,
 317
 318         .mmu.alloc_pte          = paravirt_nop,
 319         .mmu.alloc_pmd          = paravirt_nop,
 320         .mmu.alloc_pud          = paravirt_nop,
 321         .mmu.alloc_p4d          = paravirt_nop,
 322         .mmu.release_pte        = paravirt_nop,
 323         .mmu.release_pmd        = paravirt_nop,
 324         .mmu.release_pud        = paravirt_nop,
 325         .mmu.release_p4d        = paravirt_nop,
 326
 327         .mmu.set_pte            = native_set_pte,
 328         .mmu.set_pmd            = native_set_pmd,
 329
 330         .mmu.ptep_modify_prot_start     = __ptep_modify_prot_start,
 331         .mmu.ptep_modify_prot_commit    = __ptep_modify_prot_commit,
 332
 333         .mmu.set_pud            = native_set_pud,
 334
 335         .mmu.pmd_val            = PTE_IDENT,
 336         .mmu.make_pmd           = PTE_IDENT,
 337
 338         .mmu.pud_val            = PTE_IDENT,
 339         .mmu.make_pud           = PTE_IDENT,
 340
 341         .mmu.set_p4d            = native_set_p4d,
 342
 343 #if CONFIG_PGTABLE_LEVELS >= 5
 344         .mmu.p4d_val            = PTE_IDENT,
 345         .mmu.make_p4d           = PTE_IDENT,
 346
 347         .mmu.set_pgd            = native_set_pgd,
 348 #endif /* CONFIG_PGTABLE_LEVELS >= 5 */
 349
 350         .mmu.pte_val            = PTE_IDENT,
 351         .mmu.pgd_val            = PTE_IDENT,
 352
 353         .mmu.make_pte           = PTE_IDENT,
 354         .mmu.make_pgd           = PTE_IDENT,
 355
 356         .mmu.enter_mmap         = 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);