arch/x86/kernel/process_64.c

   1 /*
   2  *  Copyright (C) 1995  Linus Torvalds
   3  *
   4  *  Pentium III FXSR, SSE support
   5  *      Gareth Hughes <gareth@valinux.com>, May 2000
   6  *
   7  *  X86-64 port
   8  *      Andi Kleen.
   9  *
  10  *      CPU hotplug support - ashok.raj@intel.com
  11  */
  12
  13 /*
  14  * This file handles the architecture-dependent parts of process handling..
  15  */
  16
  17 #include <linux/cpu.h>
  18 #include <linux/errno.h>
  19 #include <linux/sched.h>
  20 #include <linux/fs.h>
  21 #include <linux/kernel.h>
  22 #include <linux/mm.h>
  23 #include <linux/elfcore.h>
  24 #include <linux/smp.h>
  25 #include <linux/slab.h>
  26 #include <linux/user.h>
  27 #include <linux/interrupt.h>
  28 #include <linux/delay.h>
  29 #include <linux/module.h>
  30 #include <linux/ptrace.h>
  31 #include <linux/notifier.h>
  32 #include <linux/kprobes.h>
  33 #include <linux/kdebug.h>
  34 #include <linux/prctl.h>
  35 #include <linux/uaccess.h>
  36 #include <linux/io.h>
  37 #include <linux/ftrace.h>
  38
  39 #include <asm/pgtable.h>
  40 #include <asm/processor.h>
  41 #include <asm/fpu/internal.h>
  42 #include <asm/mmu_context.h>
  43 #include <asm/prctl.h>
  44 #include <asm/desc.h>
  45 #include <asm/proto.h>
  46 #include <asm/ia32.h>
  47 #include <asm/idle.h>
  48 #include <asm/syscalls.h>
  49 #include <asm/debugreg.h>
  50 #include <asm/switch_to.h>
  51
  52 asmlinkage extern void ret_from_fork(void);
  53
  54 __visible DEFINE_PER_CPU(unsigned long, rsp_scratch);
  55
  56 /* Prints also some state that isn't saved in the pt_regs */
  57 void __show_regs(struct pt_regs *regs, int all)
  58 {
  59         unsigned long cr0 = 0L, cr2 = 0L, cr3 = 0L, cr4 = 0L, fs, gs, shadowgs;
  60         unsigned long d0, d1, d2, d3, d6, d7;
  61         unsigned int fsindex, gsindex;
  62         unsigned int ds, cs, es;
  63
  64         printk(KERN_DEFAULT "RIP: %04lx:[<%016lx>] ", regs->cs & 0xffff, regs->ip);
  65         printk_address(regs->ip);
  66         printk(KERN_DEFAULT "RSP: %04lx:%016lx  EFLAGS: %08lx\n", regs->ss,
  67                         regs->sp, regs->flags);
  68         printk(KERN_DEFAULT "RAX: %016lx RBX: %016lx RCX: %016lx\n",
  69                regs->ax, regs->bx, regs->cx);
  70         printk(KERN_DEFAULT "RDX: %016lx RSI: %016lx RDI: %016lx\n",
  71                regs->dx, regs->si, regs->di);
  72         printk(KERN_DEFAULT "RBP: %016lx R08: %016lx R09: %016lx\n",
  73                regs->bp, regs->r8, regs->r9);
  74         printk(KERN_DEFAULT "R10: %016lx R11: %016lx R12: %016lx\n",
  75                regs->r10, regs->r11, regs->r12);
  76         printk(KERN_DEFAULT "R13: %016lx R14: %016lx R15: %016lx\n",
  77                regs->r13, regs->r14, regs->r15);
  78
  79         asm("movl %%ds,%0" : "=r" (ds));
  80         asm("movl %%cs,%0" : "=r" (cs));
  81         asm("movl %%es,%0" : "=r" (es));
  82         asm("movl %%fs,%0" : "=r" (fsindex));
  83         asm("movl %%gs,%0" : "=r" (gsindex));
  84
  85         rdmsrl(MSR_FS_BASE, fs);
  86         rdmsrl(MSR_GS_BASE, gs);
  87         rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
  88
  89         if (!all)
  90                 return;
  91
  92         cr0 = read_cr0();
  93         cr2 = read_cr2();
  94         cr3 = read_cr3();
  95         cr4 = __read_cr4();
  96
  97         printk(KERN_DEFAULT "FS:  %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
  98                fs, fsindex, gs, gsindex, shadowgs);
  99         printk(KERN_DEFAULT "CS:  %04x DS: %04x ES: %04x CR0: %016lx\n", cs, ds,
 100                         es, cr0);
 101         printk(KERN_DEFAULT "CR2: %016lx CR3: %016lx CR4: %016lx\n", cr2, cr3,
 102                         cr4);
 103
 104         get_debugreg(d0, 0);
 105         get_debugreg(d1, 1);
 106         get_debugreg(d2, 2);
 107         get_debugreg(d3, 3);
 108         get_debugreg(d6, 6);
 109         get_debugreg(d7, 7);
 110
 111         /* Only print out debug registers if they are in their non-default state. */
 112         if ((d0 == 0) && (d1 == 0) && (d2 == 0) && (d3 == 0) &&
 113             (d6 == DR6_RESERVED) && (d7 == 0x400))
 114                 return;
 115
 116         printk(KERN_DEFAULT "DR0: %016lx DR1: %016lx DR2: %016lx\n", d0, d1, d2);
 117         printk(KERN_DEFAULT "DR3: %016lx DR6: %016lx DR7: %016lx\n", d3, d6, d7);
 118
 119         if (boot_cpu_has(X86_FEATURE_OSPKE))
 120                 printk(KERN_DEFAULT "PKRU: %08x\n", read_pkru());
 121 }
 122
 123 void release_thread(struct task_struct *dead_task)
 124 {
 125         if (dead_task->mm) {
 126 #ifdef CONFIG_MODIFY_LDT_SYSCALL
 127                 if (dead_task->mm->context.ldt) {
 128                         pr_warn("WARNING: dead process %s still has LDT? <%p/%d>\n",
 129                                 dead_task->comm,
 130                                 dead_task->mm->context.ldt->entries,
 131                                 dead_task->mm->context.ldt->size);
 132                         BUG();
 133                 }
 134 #endif
 135         }
 136 }
 137
 138 static inline void set_32bit_tls(struct task_struct *t, int tls, u32 addr)
 139 {
 140         struct user_desc ud = {
 141                 .base_addr = addr,
 142                 .limit = 0xfffff,
 143                 .seg_32bit = 1,
 144                 .limit_in_pages = 1,
 145                 .useable = 1,
 146         };
 147         struct desc_struct *desc = t->thread.tls_array;
 148         desc += tls;
 149         fill_ldt(desc, &ud);
 150 }
 151
 152 static inline u32 read_32bit_tls(struct task_struct *t, int tls)
 153 {
 154         return get_desc_base(&t->thread.tls_array[tls]);
 155 }
 156
 157 int copy_thread_tls(unsigned long clone_flags, unsigned long sp,
 158                 unsigned long arg, struct task_struct *p, unsigned long tls)
 159 {
 160         int err;
 161         struct pt_regs *childregs;
 162         struct task_struct *me = current;
 163
 164         p->thread.sp0 = (unsigned long)task_stack_page(p) + THREAD_SIZE;
 165         childregs = task_pt_regs(p);
 166         p->thread.sp = (unsigned long) childregs;
 167         set_tsk_thread_flag(p, TIF_FORK);
 168         p->thread.io_bitmap_ptr = NULL;
 169
 170         savesegment(gs, p->thread.gsindex);
 171         p->thread.gs = p->thread.gsindex ? 0 : me->thread.gs;
 172         savesegment(fs, p->thread.fsindex);
 173         p->thread.fs = p->thread.fsindex ? 0 : me->thread.fs;
 174         savesegment(es, p->thread.es);
 175         savesegment(ds, p->thread.ds);
 176         memset(p->thread.ptrace_bps, 0, sizeof(p->thread.ptrace_bps));
 177
 178         if (unlikely(p->flags & PF_KTHREAD)) {
 179                 /* kernel thread */
 180                 memset(childregs, 0, sizeof(struct pt_regs));
 181                 childregs->sp = (unsigned long)childregs;
 182                 childregs->ss = __KERNEL_DS;
 183                 childregs->bx = sp; /* function */
 184                 childregs->bp = arg;
 185                 childregs->orig_ax = -1;
 186                 childregs->cs = __KERNEL_CS | get_kernel_rpl();
 187                 childregs->flags = X86_EFLAGS_IF | X86_EFLAGS_FIXED;
 188                 return 0;
 189         }
 190         *childregs = *current_pt_regs();
 191
 192         childregs->ax = 0;
 193         if (sp)
 194                 childregs->sp = sp;
 195
 196         err = -ENOMEM;
 197         if (unlikely(test_tsk_thread_flag(me, TIF_IO_BITMAP))) {
 198                 p->thread.io_bitmap_ptr = kmemdup(me->thread.io_bitmap_ptr,
 199                                                   IO_BITMAP_BYTES, GFP_KERNEL);
 200                 if (!p->thread.io_bitmap_ptr) {
 201                         p->thread.io_bitmap_max = 0;
 202                         return -ENOMEM;
 203                 }
 204                 set_tsk_thread_flag(p, TIF_IO_BITMAP);
 205         }
 206
 207         /*
 208          * Set a new TLS for the child thread?
 209          */
 210         if (clone_flags & CLONE_SETTLS) {
 211 #ifdef CONFIG_IA32_EMULATION
 212                 if (is_ia32_task())
 213                         err = do_set_thread_area(p, -1,
 214                                 (struct user_desc __user *)tls, 0);
 215                 else
 216 #endif
 217                         err = do_arch_prctl(p, ARCH_SET_FS, tls);
 218                 if (err)
 219                         goto out;
 220         }
 221         err = 0;
 222 out:
 223         if (err && p->thread.io_bitmap_ptr) {
 224                 kfree(p->thread.io_bitmap_ptr);
 225                 p->thread.io_bitmap_max = 0;
 226         }
 227
 228         return err;
 229 }
 230
 231 static void
 232 start_thread_common(struct pt_regs *regs, unsigned long new_ip,
 233                     unsigned long new_sp,
 234                     unsigned int _cs, unsigned int _ss, unsigned int _ds)
 235 {
 236         loadsegment(fs, 0);
 237         loadsegment(es, _ds);
 238         loadsegment(ds, _ds);
 239         load_gs_index(0);
 240         regs->ip                = new_ip;
 241         regs->sp                = new_sp;
 242         regs->cs                = _cs;
 243         regs->ss                = _ss;
 244         regs->flags             = X86_EFLAGS_IF;
 245         force_iret();
 246 }
 247
 248 void
 249 start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
 250 {
 251         start_thread_common(regs, new_ip, new_sp,
 252                             __USER_CS, __USER_DS, 0);
 253 }
 254
 255 #ifdef CONFIG_COMPAT
 256 void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp)
 257 {
 258         start_thread_common(regs, new_ip, new_sp,
 259                             test_thread_flag(TIF_X32)
 260                             ? __USER_CS : __USER32_CS,
 261                             __USER_DS, __USER_DS);
 262 }
 263 #endif
 264
 265 /*
 266  *      switch_to(x,y) should switch tasks from x to y.
 267  *
 268  * This could still be optimized:
 269  * - fold all the options into a flag word and test it with a single test.
 270  * - could test fs/gs bitsliced
 271  *
 272  * Kprobes not supported here. Set the probe on schedule instead.
 273  * Function graph tracer not supported too.
 274  */
 275 __visible __notrace_funcgraph struct task_struct *
 276 __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
 277 {
 278         struct thread_struct *prev = &prev_p->thread;
 279         struct thread_struct *next = &next_p->thread;
 280         struct fpu *prev_fpu = &prev->fpu;
 281         struct fpu *next_fpu = &next->fpu;
 282         int cpu = smp_processor_id();
 283         struct tss_struct *tss = &per_cpu(cpu_tss, cpu);
 284         unsigned fsindex, gsindex;
 285         fpu_switch_t fpu_switch;
 286
 287         fpu_switch = switch_fpu_prepare(prev_fpu, next_fpu, cpu);
 288
 289         /* We must save %fs and %gs before load_TLS() because
 290          * %fs and %gs may be cleared by load_TLS().
 291          *
 292          * (e.g. xen_load_tls())
 293          */
 294         savesegment(fs, fsindex);
 295         savesegment(gs, gsindex);
 296
 297         /*
 298          * Load TLS before restoring any segments so that segment loads
 299          * reference the correct GDT entries.
 300          */
 301         load_TLS(next, cpu);
 302
 303         /*
 304          * Leave lazy mode, flushing any hypercalls made here.  This
 305          * must be done after loading TLS entries in the GDT but before
 306          * loading segments that might reference them, and and it must
 307          * be done before fpu__restore(), so the TS bit is up to
 308          * date.
 309          */
 310         arch_end_context_switch(next_p);
 311
 312         /* Switch DS and ES.
 313          *
 314          * Reading them only returns the selectors, but writing them (if
 315          * nonzero) loads the full descriptor from the GDT or LDT.  The
 316          * LDT for next is loaded in switch_mm, and the GDT is loaded
 317          * above.
 318          *
 319          * We therefore need to write new values to the segment
 320          * registers on every context switch unless both the new and old
 321          * values are zero.
 322          *
 323          * Note that we don't need to do anything for CS and SS, as
 324          * those are saved and restored as part of pt_regs.
 325          */
 326         savesegment(es, prev->es);
 327         if (unlikely(next->es | prev->es))
 328                 loadsegment(es, next->es);
 329
 330         savesegment(ds, prev->ds);
 331         if (unlikely(next->ds | prev->ds))
 332                 loadsegment(ds, next->ds);
 333
 334         /*
 335          * Switch FS and GS.
 336          *
 337          * These are even more complicated than DS and ES: they have
 338          * 64-bit bases are that controlled by arch_prctl.  Those bases
 339          * only differ from the values in the GDT or LDT if the selector
 340          * is 0.
 341          *
 342          * Loading the segment register resets the hidden base part of
 343          * the register to 0 or the value from the GDT / LDT.  If the
 344          * next base address zero, writing 0 to the segment register is
 345          * much faster than using wrmsr to explicitly zero the base.
 346          *
 347          * The thread_struct.fs and thread_struct.gs values are 0
 348          * if the fs and gs bases respectively are not overridden
 349          * from the values implied by fsindex and gsindex.  They
 350          * are nonzero, and store the nonzero base addresses, if
 351          * the bases are overridden.
 352          *
 353          * (fs != 0 && fsindex != 0) || (gs != 0 && gsindex != 0) should
 354          * be impossible.
 355          *
 356          * Therefore we need to reload the segment registers if either
 357          * the old or new selector is nonzero, and we need to override
 358          * the base address if next thread expects it to be overridden.
 359          *
 360          * This code is unnecessarily slow in the case where the old and
 361          * new indexes are zero and the new base is nonzero -- it will
 362          * unnecessarily write 0 to the selector before writing the new
 363          * base address.
 364          *
 365          * Note: This all depends on arch_prctl being the only way that
 366          * user code can override the segment base.  Once wrfsbase and
 367          * wrgsbase are enabled, most of this code will need to change.
 368          */
 369         if (unlikely(fsindex | next->fsindex | prev->fs)) {
 370                 loadsegment(fs, next->fsindex);
 371
 372                 /*
 373                  * If user code wrote a nonzero value to FS, then it also
 374                  * cleared the overridden base address.
 375                  *
 376                  * XXX: if user code wrote 0 to FS and cleared the base
 377                  * address itself, we won't notice and we'll incorrectly
 378                  * restore the prior base address next time we reschdule
 379                  * the process.
 380                  */
 381                 if (fsindex)
 382                         prev->fs = 0;
 383         }
 384         if (next->fs)
 385                 wrmsrl(MSR_FS_BASE, next->fs);
 386         prev->fsindex = fsindex;
 387
 388         if (unlikely(gsindex | next->gsindex | prev->gs)) {
 389                 load_gs_index(next->gsindex);
 390
 391                 /* This works (and fails) the same way as fsindex above. */
 392                 if (gsindex)
 393                         prev->gs = 0;
 394         }
 395         if (next->gs)
 396                 wrmsrl(MSR_KERNEL_GS_BASE, next->gs);
 397         prev->gsindex = gsindex;
 398
 399         switch_fpu_finish(next_fpu, fpu_switch);
 400
 401         /*
 402          * Switch the PDA and FPU contexts.
 403          */
 404         this_cpu_write(current_task, next_p);
 405
 406         /* Reload esp0 and ss1.  This changes current_thread_info(). */
 407         load_sp0(tss, next);
 408
 409         /*
 410          * Now maybe reload the debug registers and handle I/O bitmaps
 411          */
 412         if (unlikely(task_thread_info(next_p)->flags & _TIF_WORK_CTXSW_NEXT ||
 413                      task_thread_info(prev_p)->flags & _TIF_WORK_CTXSW_PREV))
 414                 __switch_to_xtra(prev_p, next_p, tss);
 415
 416         if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
 417                 /*
 418                  * AMD CPUs have a misfeature: SYSRET sets the SS selector but
 419                  * does not update the cached descriptor.  As a result, if we
 420                  * do SYSRET while SS is NULL, we'll end up in user mode with
 421                  * SS apparently equal to __USER_DS but actually unusable.
 422                  *
 423                  * The straightforward workaround would be to fix it up just
 424                  * before SYSRET, but that would slow down the system call
 425                  * fast paths.  Instead, we ensure that SS is never NULL in
 426                  * system call context.  We do this by replacing NULL SS
 427                  * selectors at every context switch.  SYSCALL sets up a valid
 428                  * SS, so the only way to get NULL is to re-enter the kernel
 429                  * from CPL 3 through an interrupt.  Since that can't happen
 430                  * in the same task as a running syscall, we are guaranteed to
 431                  * context switch between every interrupt vector entry and a
 432                  * subsequent SYSRET.
 433                  *
 434                  * We read SS first because SS reads are much faster than
 435                  * writes.  Out of caution, we force SS to __KERNEL_DS even if
 436                  * it previously had a different non-NULL value.
 437                  */
 438                 unsigned short ss_sel;
 439                 savesegment(ss, ss_sel);
 440                 if (ss_sel != __KERNEL_DS)
 441                         loadsegment(ss, __KERNEL_DS);
 442         }
 443
 444         return prev_p;
 445 }
 446
 447 void set_personality_64bit(void)
 448 {
 449         /* inherit personality from parent */
 450
 451         /* Make sure to be in 64bit mode */
 452         clear_thread_flag(TIF_IA32);
 453         clear_thread_flag(TIF_ADDR32);
 454         clear_thread_flag(TIF_X32);
 455
 456         /* Ensure the corresponding mm is not marked. */
 457         if (current->mm)
 458                 current->mm->context.ia32_compat = 0;
 459
 460         /* TBD: overwrites user setup. Should have two bits.
 461            But 64bit processes have always behaved this way,
 462            so it's not too bad. The main problem is just that
 463            32bit childs are affected again. */
 464         current->personality &= ~READ_IMPLIES_EXEC;
 465 }
 466
 467 void set_personality_ia32(bool x32)
 468 {
 469         /* inherit personality from parent */
 470
 471         /* Make sure to be in 32bit mode */
 472         set_thread_flag(TIF_ADDR32);
 473
 474         /* Mark the associated mm as containing 32-bit tasks. */
 475         if (x32) {
 476                 clear_thread_flag(TIF_IA32);
 477                 set_thread_flag(TIF_X32);
 478                 if (current->mm)
 479                         current->mm->context.ia32_compat = TIF_X32;
 480                 current->personality &= ~READ_IMPLIES_EXEC;
 481                 /* is_compat_task() uses the presence of the x32
 482                    syscall bit flag to determine compat status */
 483                 current_thread_info()->status &= ~TS_COMPAT;
 484         } else {
 485                 set_thread_flag(TIF_IA32);
 486                 clear_thread_flag(TIF_X32);
 487                 if (current->mm)
 488                         current->mm->context.ia32_compat = TIF_IA32;
 489                 current->personality |= force_personality32;
 490                 /* Prepare the first "return" to user space */
 491                 current_thread_info()->status |= TS_COMPAT;
 492         }
 493 }
 494 EXPORT_SYMBOL_GPL(set_personality_ia32);
 495
 496 long do_arch_prctl(struct task_struct *task, int code, unsigned long addr)
 497 {
 498         int ret = 0;
 499         int doit = task == current;
 500         int cpu;
 501
 502         switch (code) {
 503         case ARCH_SET_GS:
 504                 if (addr >= TASK_SIZE_OF(task))
 505                         return -EPERM;
 506                 cpu = get_cpu();
 507                 /* handle small bases via the GDT because that's faster to
 508                    switch. */
 509                 if (addr <= 0xffffffff) {
 510                         set_32bit_tls(task, GS_TLS, addr);
 511                         if (doit) {
 512                                 load_TLS(&task->thread, cpu);
 513                                 load_gs_index(GS_TLS_SEL);
 514                         }
 515                         task->thread.gsindex = GS_TLS_SEL;
 516                         task->thread.gs = 0;
 517                 } else {
 518                         task->thread.gsindex = 0;
 519                         task->thread.gs = addr;
 520                         if (doit) {
 521                                 load_gs_index(0);
 522                                 ret = wrmsrl_safe(MSR_KERNEL_GS_BASE, addr);
 523                         }
 524                 }
 525                 put_cpu();
 526                 break;
 527         case ARCH_SET_FS:
 528                 /* Not strictly needed for fs, but do it for symmetry
 529                    with gs */
 530                 if (addr >= TASK_SIZE_OF(task))
 531                         return -EPERM;
 532                 cpu = get_cpu();
 533                 /* handle small bases via the GDT because that's faster to
 534                    switch. */
 535                 if (addr <= 0xffffffff) {
 536                         set_32bit_tls(task, FS_TLS, addr);
 537                         if (doit) {
 538                                 load_TLS(&task->thread, cpu);
 539                                 loadsegment(fs, FS_TLS_SEL);
 540                         }
 541                         task->thread.fsindex = FS_TLS_SEL;
 542                         task->thread.fs = 0;
 543                 } else {
 544                         task->thread.fsindex = 0;
 545                         task->thread.fs = addr;
 546                         if (doit) {
 547                                 /* set the selector to 0 to not confuse
 548                                    __switch_to */
 549                                 loadsegment(fs, 0);
 550                                 ret = wrmsrl_safe(MSR_FS_BASE, addr);
 551                         }
 552                 }
 553                 put_cpu();
 554                 break;
 555         case ARCH_GET_FS: {
 556                 unsigned long base;
 557                 if (task->thread.fsindex == FS_TLS_SEL)
 558                         base = read_32bit_tls(task, FS_TLS);
 559                 else if (doit)
 560                         rdmsrl(MSR_FS_BASE, base);
 561                 else
 562                         base = task->thread.fs;
 563                 ret = put_user(base, (unsigned long __user *)addr);
 564                 break;
 565         }
 566         case ARCH_GET_GS: {
 567                 unsigned long base;
 568                 unsigned gsindex;
 569                 if (task->thread.gsindex == GS_TLS_SEL)
 570                         base = read_32bit_tls(task, GS_TLS);
 571                 else if (doit) {
 572                         savesegment(gs, gsindex);
 573                         if (gsindex)
 574                                 rdmsrl(MSR_KERNEL_GS_BASE, base);
 575                         else
 576                                 base = task->thread.gs;
 577                 } else
 578                         base = task->thread.gs;
 579                 ret = put_user(base, (unsigned long __user *)addr);
 580                 break;
 581         }
 582
 583         default:
 584                 ret = -EINVAL;
 585                 break;
 586         }
 587
 588         return ret;
 589 }
 590
 591 long sys_arch_prctl(int code, unsigned long addr)
 592 {
 593         return do_arch_prctl(current, code, addr);
 594 }
 595
 596 unsigned long KSTK_ESP(struct task_struct *task)
 597 {
 598         return task_pt_regs(task)->sp;
 599 }