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 extern void _paravirt_nop(void);
  41 asm (".pushsection .entry.text, \"ax\"\n"
  42      ".global _paravirt_nop\n"
  43      "_paravirt_nop:\n\t"
  44      "ret\n\t"
  45      ".size _paravirt_nop, . - _paravirt_nop\n\t"
  46      ".type _paravirt_nop, @function\n\t"
  47      ".popsection");
  48
  49 void __init default_banner(void)
  50 {
  51         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
  52                pv_info.name);
  53 }
  54
  55 /* Undefined instruction for dealing with missing ops pointers. */
  56 static void paravirt_BUG(void)
  57 {
  58         BUG();
  59 }
  60
  61 struct branch {
  62         unsigned char opcode;
  63         u32 delta;
  64 } __attribute__((packed));
  65
  66 static unsigned paravirt_patch_call(void *insn_buff, const void *target,
  67                                     unsigned long addr, unsigned len)
  68 {
  69         const int call_len = 5;
  70         struct branch *b = insn_buff;
  71         unsigned long delta = (unsigned long)target - (addr+call_len);
  72
  73         if (len < call_len) {
  74                 pr_warn("paravirt: Failed to patch indirect CALL at %ps\n", (void *)addr);
  75                 /* Kernel might not be viable if patching fails, bail out: */
  76                 BUG_ON(1);
  77         }
  78
  79         b->opcode = 0xe8; /* call */
  80         b->delta = delta;
  81         BUILD_BUG_ON(sizeof(*b) != call_len);
  82
  83         return call_len;
  84 }
  85
  86 #ifdef CONFIG_PARAVIRT_XXL
  87 /* identity function, which can be inlined */
  88 u64 notrace _paravirt_ident_64(u64 x)
  89 {
  90         return x;
  91 }
  92 #endif
  93
  94 DEFINE_STATIC_KEY_TRUE(virt_spin_lock_key);
  95
  96 void __init native_pv_lock_init(void)
  97 {
  98         if (!boot_cpu_has(X86_FEATURE_HYPERVISOR))
  99                 static_branch_disable(&virt_spin_lock_key);
 100 }
 101
 102 unsigned int paravirt_patch(u8 type, void *insn_buff, unsigned long addr,
 103                             unsigned int len)
 104 {
 105         /*
 106          * Neat trick to map patch type back to the call within the
 107          * corresponding structure.
 108          */
 109         void *opfunc = *((void **)&pv_ops + type);
 110         unsigned ret;
 111
 112         if (opfunc == NULL)
 113                 /* If there's no function, patch it with paravirt_BUG() */
 114                 ret = paravirt_patch_call(insn_buff, paravirt_BUG, addr, len);
 115         else if (opfunc == _paravirt_nop)
 116                 ret = 0;
 117         else
 118                 /* Otherwise call the function. */
 119                 ret = paravirt_patch_call(insn_buff, opfunc, addr, len);
 120
 121         return ret;
 122 }
 123
 124 struct static_key paravirt_steal_enabled;
 125 struct static_key paravirt_steal_rq_enabled;
 126
 127 static u64 native_steal_clock(int cpu)
 128 {
 129         return 0;
 130 }
 131
 132 DEFINE_STATIC_CALL(pv_steal_clock, native_steal_clock);
 133 DEFINE_STATIC_CALL(pv_sched_clock, native_sched_clock);
 134
 135 void paravirt_set_sched_clock(u64 (*func)(void))
 136 {
 137         static_call_update(pv_sched_clock, func);
 138 }
 139
 140 /* These are in entry.S */
 141 extern void native_iret(void);
 142
 143 static struct resource reserve_ioports = {
 144         .start = 0,
 145         .end = IO_SPACE_LIMIT,
 146         .name = "paravirt-ioport",
 147         .flags = IORESOURCE_IO | IORESOURCE_BUSY,
 148 };
 149
 150 /*
 151  * Reserve the whole legacy IO space to prevent any legacy drivers
 152  * from wasting time probing for their hardware.  This is a fairly
 153  * brute-force approach to disabling all non-virtual drivers.
 154  *
 155  * Note that this must be called very early to have any effect.
 156  */
 157 int paravirt_disable_iospace(void)
 158 {
 159         return request_resource(&ioport_resource, &reserve_ioports);
 160 }
 161
 162 static DEFINE_PER_CPU(enum paravirt_lazy_mode, paravirt_lazy_mode) = PARAVIRT_LAZY_NONE;
 163
 164 static inline void enter_lazy(enum paravirt_lazy_mode mode)
 165 {
 166         BUG_ON(this_cpu_read(paravirt_lazy_mode) != PARAVIRT_LAZY_NONE);
 167
 168         this_cpu_write(paravirt_lazy_mode, mode);
 169 }
 170
 171 static void leave_lazy(enum paravirt_lazy_mode mode)
 172 {
 173         BUG_ON(this_cpu_read(paravirt_lazy_mode) != mode);
 174
 175         this_cpu_write(paravirt_lazy_mode, PARAVIRT_LAZY_NONE);
 176 }
 177
 178 void paravirt_enter_lazy_mmu(void)
 179 {
 180         enter_lazy(PARAVIRT_LAZY_MMU);
 181 }
 182
 183 void paravirt_leave_lazy_mmu(void)
 184 {
 185         leave_lazy(PARAVIRT_LAZY_MMU);
 186 }
 187
 188 void paravirt_flush_lazy_mmu(void)
 189 {
 190         preempt_disable();
 191
 192         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU) {
 193                 arch_leave_lazy_mmu_mode();
 194                 arch_enter_lazy_mmu_mode();
 195         }
 196
 197         preempt_enable();
 198 }
 199
 200 #ifdef CONFIG_PARAVIRT_XXL
 201 void paravirt_start_context_switch(struct task_struct *prev)
 202 {
 203         BUG_ON(preemptible());
 204
 205         if (this_cpu_read(paravirt_lazy_mode) == PARAVIRT_LAZY_MMU) {
 206                 arch_leave_lazy_mmu_mode();
 207                 set_ti_thread_flag(task_thread_info(prev), TIF_LAZY_MMU_UPDATES);
 208         }
 209         enter_lazy(PARAVIRT_LAZY_CPU);
 210 }
 211
 212 void paravirt_end_context_switch(struct task_struct *next)
 213 {
 214         BUG_ON(preemptible());
 215
 216         leave_lazy(PARAVIRT_LAZY_CPU);
 217
 218         if (test_and_clear_ti_thread_flag(task_thread_info(next), TIF_LAZY_MMU_UPDATES))
 219                 arch_enter_lazy_mmu_mode();
 220 }
 221
 222 static noinstr unsigned long pv_native_read_cr2(void)
 223 {
 224         return native_read_cr2();
 225 }
 226
 227 static noinstr void pv_native_write_cr2(unsigned long val)
 228 {
 229         native_write_cr2(val);
 230 }
 231
 232 static noinstr unsigned long pv_native_get_debugreg(int regno)
 233 {
 234         return native_get_debugreg(regno);
 235 }
 236
 237 static noinstr void pv_native_set_debugreg(int regno, unsigned long val)
 238 {
 239         native_set_debugreg(regno, val);
 240 }
 241
 242 static noinstr void pv_native_irq_enable(void)
 243 {
 244         native_irq_enable();
 245 }
 246
 247 static noinstr void pv_native_irq_disable(void)
 248 {
 249         native_irq_disable();
 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         /* Cpu ops. */
 273         .cpu.io_delay           = native_io_delay,
 274
 275 #ifdef CONFIG_PARAVIRT_XXL
 276         .cpu.cpuid              = native_cpuid,
 277         .cpu.get_debugreg       = pv_native_get_debugreg,
 278         .cpu.set_debugreg       = pv_native_set_debugreg,
 279         .cpu.read_cr0           = native_read_cr0,
 280         .cpu.write_cr0          = native_write_cr0,
 281         .cpu.write_cr4          = native_write_cr4,
 282         .cpu.wbinvd             = native_wbinvd,
 283         .cpu.read_msr           = native_read_msr,
 284         .cpu.write_msr          = native_write_msr,
 285         .cpu.read_msr_safe      = native_read_msr_safe,
 286         .cpu.write_msr_safe     = native_write_msr_safe,
 287         .cpu.read_pmc           = native_read_pmc,
 288         .cpu.load_tr_desc       = native_load_tr_desc,
 289         .cpu.set_ldt            = native_set_ldt,
 290         .cpu.load_gdt           = native_load_gdt,
 291         .cpu.load_idt           = native_load_idt,
 292         .cpu.store_tr           = native_store_tr,
 293         .cpu.load_tls           = native_load_tls,
 294         .cpu.load_gs_index      = native_load_gs_index,
 295         .cpu.write_ldt_entry    = native_write_ldt_entry,
 296         .cpu.write_gdt_entry    = native_write_gdt_entry,
 297         .cpu.write_idt_entry    = native_write_idt_entry,
 298
 299         .cpu.alloc_ldt          = paravirt_nop,
 300         .cpu.free_ldt           = paravirt_nop,
 301
 302         .cpu.load_sp0           = native_load_sp0,
 303
 304 #ifdef CONFIG_X86_IOPL_IOPERM
 305         .cpu.invalidate_io_bitmap       = native_tss_invalidate_io_bitmap,
 306         .cpu.update_io_bitmap           = native_tss_update_io_bitmap,
 307 #endif
 308
 309         .cpu.start_context_switch       = paravirt_nop,
 310         .cpu.end_context_switch         = paravirt_nop,
 311
 312         /* Irq ops. */
 313         .irq.save_fl            = __PV_IS_CALLEE_SAVE(native_save_fl),
 314         .irq.irq_disable        = __PV_IS_CALLEE_SAVE(pv_native_irq_disable),
 315         .irq.irq_enable         = __PV_IS_CALLEE_SAVE(pv_native_irq_enable),
 316         .irq.safe_halt          = native_safe_halt,
 317         .irq.halt               = native_halt,
 318 #endif /* CONFIG_PARAVIRT_XXL */
 319
 320         /* Mmu ops. */
 321         .mmu.flush_tlb_user     = native_flush_tlb_local,
 322         .mmu.flush_tlb_kernel   = native_flush_tlb_global,
 323         .mmu.flush_tlb_one_user = native_flush_tlb_one_user,
 324         .mmu.flush_tlb_multi    = native_flush_tlb_multi,
 325         .mmu.tlb_remove_table   =
 326                         (void (*)(struct mmu_gather *, void *))tlb_remove_page,
 327
 328         .mmu.exit_mmap          = paravirt_nop,
 329
 330 #ifdef CONFIG_PARAVIRT_XXL
 331         .mmu.read_cr2           = __PV_IS_CALLEE_SAVE(pv_native_read_cr2),
 332         .mmu.write_cr2          = pv_native_write_cr2,
 333         .mmu.read_cr3           = __native_read_cr3,
 334         .mmu.write_cr3          = native_write_cr3,
 335
 336         .mmu.pgd_alloc          = __paravirt_pgd_alloc,
 337         .mmu.pgd_free           = paravirt_nop,
 338
 339         .mmu.alloc_pte          = paravirt_nop,
 340         .mmu.alloc_pmd          = paravirt_nop,
 341         .mmu.alloc_pud          = paravirt_nop,
 342         .mmu.alloc_p4d          = paravirt_nop,
 343         .mmu.release_pte        = paravirt_nop,
 344         .mmu.release_pmd        = paravirt_nop,
 345         .mmu.release_pud        = paravirt_nop,
 346         .mmu.release_p4d        = paravirt_nop,
 347
 348         .mmu.set_pte            = native_set_pte,
 349         .mmu.set_pmd            = native_set_pmd,
 350
 351         .mmu.ptep_modify_prot_start     = __ptep_modify_prot_start,
 352         .mmu.ptep_modify_prot_commit    = __ptep_modify_prot_commit,
 353
 354         .mmu.set_pud            = native_set_pud,
 355
 356         .mmu.pmd_val            = PTE_IDENT,
 357         .mmu.make_pmd           = PTE_IDENT,
 358
 359         .mmu.pud_val            = PTE_IDENT,
 360         .mmu.make_pud           = PTE_IDENT,
 361
 362         .mmu.set_p4d            = native_set_p4d,
 363
 364 #if CONFIG_PGTABLE_LEVELS >= 5
 365         .mmu.p4d_val            = PTE_IDENT,
 366         .mmu.make_p4d           = PTE_IDENT,
 367
 368         .mmu.set_pgd            = native_set_pgd,
 369 #endif /* CONFIG_PGTABLE_LEVELS >= 5 */
 370
 371         .mmu.pte_val            = PTE_IDENT,
 372         .mmu.pgd_val            = PTE_IDENT,
 373
 374         .mmu.make_pte           = PTE_IDENT,
 375         .mmu.make_pgd           = PTE_IDENT,
 376
 377         .mmu.dup_mmap           = paravirt_nop,
 378         .mmu.activate_mm        = paravirt_nop,
 379
 380         .mmu.lazy_mode = {
 381                 .enter          = paravirt_nop,
 382                 .leave          = paravirt_nop,
 383                 .flush          = paravirt_nop,
 384         },
 385
 386         .mmu.set_fixmap         = native_set_fixmap,
 387 #endif /* CONFIG_PARAVIRT_XXL */
 388
 389 #if defined(CONFIG_PARAVIRT_SPINLOCKS)
 390         /* Lock ops. */
 391 #ifdef CONFIG_SMP
 392         .lock.queued_spin_lock_slowpath = native_queued_spin_lock_slowpath,
 393         .lock.queued_spin_unlock        =
 394                                 PV_CALLEE_SAVE(__native_queued_spin_unlock),
 395         .lock.wait                      = paravirt_nop,
 396         .lock.kick                      = paravirt_nop,
 397         .lock.vcpu_is_preempted         =
 398                                 PV_CALLEE_SAVE(__native_vcpu_is_preempted),
 399 #endif /* SMP */
 400 #endif
 401 };
 402
 403 #ifdef CONFIG_PARAVIRT_XXL
 404 NOKPROBE_SYMBOL(native_load_idt);
 405
 406 void (*paravirt_iret)(void) = native_iret;
 407 #endif
 408
 409 EXPORT_SYMBOL(pv_ops);
 410 EXPORT_SYMBOL_GPL(pv_info);