2 * Copyright (C) 1995 Linus Torvalds
4 * Pentium III FXSR, SSE support
14 * This file handles the architecture-dependent parts of process handling..
17 #include <linux/cpu.h>
18 #include <linux/errno.h>
19 #include <linux/sched.h>
20 #include <linux/sched/task.h>
21 #include <linux/sched/task_stack.h>
23 #include <linux/kernel.h>
25 #include <linux/elfcore.h>
26 #include <linux/smp.h>
27 #include <linux/slab.h>
28 #include <linux/user.h>
29 #include <linux/interrupt.h>
30 #include <linux/delay.h>
31 #include <linux/export.h>
32 #include <linux/ptrace.h>
33 #include <linux/notifier.h>
34 #include <linux/kprobes.h>
35 #include <linux/kdebug.h>
36 #include <linux/prctl.h>
37 #include <linux/uaccess.h>
39 #include <linux/ftrace.h>
40 #include <linux/syscalls.h>
42 #include <asm/pgtable.h>
43 #include <asm/processor.h>
44 #include <asm/fpu/internal.h>
45 #include <asm/mmu_context.h>
46 #include <asm/prctl.h>
48 #include <asm/proto.h>
50 #include <asm/syscalls.h>
51 #include <asm/debugreg.h>
52 #include <asm/switch_to.h>
53 #include <asm/xen/hypervisor.h>
55 #include <asm/intel_rdt_sched.h>
56 #include <asm/unistd.h>
57 #include <asm/fsgsbase.h>
58 #ifdef CONFIG_IA32_EMULATION
59 /* Not included via unistd.h */
60 #include <asm/unistd_32_ia32.h>
65 /* Prints also some state that isn't saved in the pt_regs */
66 void __show_regs(struct pt_regs *regs, enum show_regs_mode mode)
68 unsigned long cr0 = 0L, cr2 = 0L, cr3 = 0L, cr4 = 0L, fs, gs, shadowgs;
69 unsigned long d0, d1, d2, d3, d6, d7;
70 unsigned int fsindex, gsindex;
71 unsigned int ds, cs, es;
75 if (regs->orig_ax != -1)
76 pr_cont(" ORIG_RAX: %016lx\n", regs->orig_ax);
80 printk(KERN_DEFAULT "RAX: %016lx RBX: %016lx RCX: %016lx\n",
81 regs->ax, regs->bx, regs->cx);
82 printk(KERN_DEFAULT "RDX: %016lx RSI: %016lx RDI: %016lx\n",
83 regs->dx, regs->si, regs->di);
84 printk(KERN_DEFAULT "RBP: %016lx R08: %016lx R09: %016lx\n",
85 regs->bp, regs->r8, regs->r9);
86 printk(KERN_DEFAULT "R10: %016lx R11: %016lx R12: %016lx\n",
87 regs->r10, regs->r11, regs->r12);
88 printk(KERN_DEFAULT "R13: %016lx R14: %016lx R15: %016lx\n",
89 regs->r13, regs->r14, regs->r15);
91 if (mode == SHOW_REGS_SHORT)
94 if (mode == SHOW_REGS_USER) {
95 rdmsrl(MSR_FS_BASE, fs);
96 rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
97 printk(KERN_DEFAULT "FS: %016lx GS: %016lx\n",
102 asm("movl %%ds,%0" : "=r" (ds));
103 asm("movl %%cs,%0" : "=r" (cs));
104 asm("movl %%es,%0" : "=r" (es));
105 asm("movl %%fs,%0" : "=r" (fsindex));
106 asm("movl %%gs,%0" : "=r" (gsindex));
108 rdmsrl(MSR_FS_BASE, fs);
109 rdmsrl(MSR_GS_BASE, gs);
110 rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
117 printk(KERN_DEFAULT "FS: %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
118 fs, fsindex, gs, gsindex, shadowgs);
119 printk(KERN_DEFAULT "CS: %04x DS: %04x ES: %04x CR0: %016lx\n", cs, ds,
121 printk(KERN_DEFAULT "CR2: %016lx CR3: %016lx CR4: %016lx\n", cr2, cr3,
131 /* Only print out debug registers if they are in their non-default state. */
132 if (!((d0 == 0) && (d1 == 0) && (d2 == 0) && (d3 == 0) &&
133 (d6 == DR6_RESERVED) && (d7 == 0x400))) {
134 printk(KERN_DEFAULT "DR0: %016lx DR1: %016lx DR2: %016lx\n",
136 printk(KERN_DEFAULT "DR3: %016lx DR6: %016lx DR7: %016lx\n",
140 if (boot_cpu_has(X86_FEATURE_OSPKE))
141 printk(KERN_DEFAULT "PKRU: %08x\n", read_pkru());
144 void release_thread(struct task_struct *dead_task)
147 #ifdef CONFIG_MODIFY_LDT_SYSCALL
148 if (dead_task->mm->context.ldt) {
149 pr_warn("WARNING: dead process %s still has LDT? <%p/%d>\n",
151 dead_task->mm->context.ldt->entries,
152 dead_task->mm->context.ldt->nr_entries);
159 enum which_selector {
165 * Saves the FS or GS base for an outgoing thread if FSGSBASE extensions are
166 * not available. The goal is to be reasonably fast on non-FSGSBASE systems.
167 * It's forcibly inlined because it'll generate better code and this function
170 static __always_inline void save_base_legacy(struct task_struct *prev_p,
171 unsigned short selector,
172 enum which_selector which)
174 if (likely(selector == 0)) {
176 * On Intel (without X86_BUG_NULL_SEG), the segment base could
177 * be the pre-existing saved base or it could be zero. On AMD
178 * (with X86_BUG_NULL_SEG), the segment base could be almost
181 * This branch is very hot (it's hit twice on almost every
182 * context switch between 64-bit programs), and avoiding
183 * the RDMSR helps a lot, so we just assume that whatever
184 * value is already saved is correct. This matches historical
185 * Linux behavior, so it won't break existing applications.
187 * To avoid leaking state, on non-X86_BUG_NULL_SEG CPUs, if we
188 * report that the base is zero, it needs to actually be zero:
189 * see the corresponding logic in load_seg_legacy.
193 * If the selector is 1, 2, or 3, then the base is zero on
194 * !X86_BUG_NULL_SEG CPUs and could be anything on
195 * X86_BUG_NULL_SEG CPUs. In the latter case, Linux
196 * has never attempted to preserve the base across context
199 * If selector > 3, then it refers to a real segment, and
200 * saving the base isn't necessary.
203 prev_p->thread.fsbase = 0;
205 prev_p->thread.gsbase = 0;
209 static __always_inline void save_fsgs(struct task_struct *task)
211 savesegment(fs, task->thread.fsindex);
212 savesegment(gs, task->thread.gsindex);
213 save_base_legacy(task, task->thread.fsindex, FS);
214 save_base_legacy(task, task->thread.gsindex, GS);
217 #if IS_ENABLED(CONFIG_KVM)
219 * While a process is running,current->thread.fsbase and current->thread.gsbase
220 * may not match the corresponding CPU registers (see save_base_legacy()). KVM
221 * wants an efficient way to save and restore FSBASE and GSBASE.
222 * When FSGSBASE extensions are enabled, this will have to use RD{FS,GS}BASE.
224 void save_fsgs_for_kvm(void)
228 EXPORT_SYMBOL_GPL(save_fsgs_for_kvm);
231 static __always_inline void loadseg(enum which_selector which,
235 loadsegment(fs, sel);
240 static __always_inline void load_seg_legacy(unsigned short prev_index,
241 unsigned long prev_base,
242 unsigned short next_index,
243 unsigned long next_base,
244 enum which_selector which)
246 if (likely(next_index <= 3)) {
248 * The next task is using 64-bit TLS, is not using this
249 * segment at all, or is having fun with arcane CPU features.
251 if (next_base == 0) {
253 * Nasty case: on AMD CPUs, we need to forcibly zero
256 if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
257 loadseg(which, __USER_DS);
258 loadseg(which, next_index);
261 * We could try to exhaustively detect cases
262 * under which we can skip the segment load,
263 * but there's really only one case that matters
264 * for performance: if both the previous and
265 * next states are fully zeroed, we can skip
268 * (This assumes that prev_base == 0 has no
269 * false positives. This is the case on
272 if (likely(prev_index | next_index | prev_base))
273 loadseg(which, next_index);
276 if (prev_index != next_index)
277 loadseg(which, next_index);
278 wrmsrl(which == FS ? MSR_FS_BASE : MSR_KERNEL_GS_BASE,
283 * The next task is using a real segment. Loading the selector
286 loadseg(which, next_index);
290 static __always_inline void x86_fsgsbase_load(struct thread_struct *prev,
291 struct thread_struct *next)
293 load_seg_legacy(prev->fsindex, prev->fsbase,
294 next->fsindex, next->fsbase, FS);
295 load_seg_legacy(prev->gsindex, prev->gsbase,
296 next->gsindex, next->gsbase, GS);
299 static unsigned long x86_fsgsbase_read_task(struct task_struct *task,
300 unsigned short selector)
302 unsigned short idx = selector >> 3;
305 if (likely((selector & SEGMENT_TI_MASK) == 0)) {
306 if (unlikely(idx >= GDT_ENTRIES))
310 * There are no user segments in the GDT with nonzero bases
311 * other than the TLS segments.
313 if (idx < GDT_ENTRY_TLS_MIN || idx > GDT_ENTRY_TLS_MAX)
316 idx -= GDT_ENTRY_TLS_MIN;
317 base = get_desc_base(&task->thread.tls_array[idx]);
319 #ifdef CONFIG_MODIFY_LDT_SYSCALL
320 struct ldt_struct *ldt;
323 * If performance here mattered, we could protect the LDT
324 * with RCU. This is a slow path, though, so we can just
327 mutex_lock(&task->mm->context.lock);
328 ldt = task->mm->context.ldt;
329 if (unlikely(idx >= ldt->nr_entries))
332 base = get_desc_base(ldt->entries + idx);
333 mutex_unlock(&task->mm->context.lock);
342 void x86_fsbase_write_cpu(unsigned long fsbase)
345 * Set the selector to 0 as a notion, that the segment base is
346 * overwritten, which will be checked for skipping the segment load
347 * during context switch.
350 wrmsrl(MSR_FS_BASE, fsbase);
353 void x86_gsbase_write_cpu_inactive(unsigned long gsbase)
355 /* Set the selector to 0 for the same reason as %fs above. */
357 wrmsrl(MSR_KERNEL_GS_BASE, gsbase);
360 unsigned long x86_fsbase_read_task(struct task_struct *task)
362 unsigned long fsbase;
365 fsbase = x86_fsbase_read_cpu();
366 else if (task->thread.fsindex == 0)
367 fsbase = task->thread.fsbase;
369 fsbase = x86_fsgsbase_read_task(task, task->thread.fsindex);
374 unsigned long x86_gsbase_read_task(struct task_struct *task)
376 unsigned long gsbase;
379 gsbase = x86_gsbase_read_cpu_inactive();
380 else if (task->thread.gsindex == 0)
381 gsbase = task->thread.gsbase;
383 gsbase = x86_fsgsbase_read_task(task, task->thread.gsindex);
388 int x86_fsbase_write_task(struct task_struct *task, unsigned long fsbase)
391 * Not strictly needed for %fs, but do it for symmetry
394 if (unlikely(fsbase >= TASK_SIZE_MAX))
398 task->thread.fsbase = fsbase;
400 x86_fsbase_write_cpu(fsbase);
401 task->thread.fsindex = 0;
407 int x86_gsbase_write_task(struct task_struct *task, unsigned long gsbase)
409 if (unlikely(gsbase >= TASK_SIZE_MAX))
413 task->thread.gsbase = gsbase;
415 x86_gsbase_write_cpu_inactive(gsbase);
416 task->thread.gsindex = 0;
422 int copy_thread_tls(unsigned long clone_flags, unsigned long sp,
423 unsigned long arg, struct task_struct *p, unsigned long tls)
426 struct pt_regs *childregs;
427 struct fork_frame *fork_frame;
428 struct inactive_task_frame *frame;
429 struct task_struct *me = current;
431 childregs = task_pt_regs(p);
432 fork_frame = container_of(childregs, struct fork_frame, regs);
433 frame = &fork_frame->frame;
435 frame->ret_addr = (unsigned long) ret_from_fork;
436 p->thread.sp = (unsigned long) fork_frame;
437 p->thread.io_bitmap_ptr = NULL;
439 savesegment(gs, p->thread.gsindex);
440 p->thread.gsbase = p->thread.gsindex ? 0 : me->thread.gsbase;
441 savesegment(fs, p->thread.fsindex);
442 p->thread.fsbase = p->thread.fsindex ? 0 : me->thread.fsbase;
443 savesegment(es, p->thread.es);
444 savesegment(ds, p->thread.ds);
445 memset(p->thread.ptrace_bps, 0, sizeof(p->thread.ptrace_bps));
447 if (unlikely(p->flags & PF_KTHREAD)) {
449 memset(childregs, 0, sizeof(struct pt_regs));
450 frame->bx = sp; /* function */
455 *childregs = *current_pt_regs();
462 if (unlikely(test_tsk_thread_flag(me, TIF_IO_BITMAP))) {
463 p->thread.io_bitmap_ptr = kmemdup(me->thread.io_bitmap_ptr,
464 IO_BITMAP_BYTES, GFP_KERNEL);
465 if (!p->thread.io_bitmap_ptr) {
466 p->thread.io_bitmap_max = 0;
469 set_tsk_thread_flag(p, TIF_IO_BITMAP);
473 * Set a new TLS for the child thread?
475 if (clone_flags & CLONE_SETTLS) {
476 #ifdef CONFIG_IA32_EMULATION
477 if (in_ia32_syscall())
478 err = do_set_thread_area(p, -1,
479 (struct user_desc __user *)tls, 0);
482 err = do_arch_prctl_64(p, ARCH_SET_FS, tls);
488 if (err && p->thread.io_bitmap_ptr) {
489 kfree(p->thread.io_bitmap_ptr);
490 p->thread.io_bitmap_max = 0;
497 start_thread_common(struct pt_regs *regs, unsigned long new_ip,
498 unsigned long new_sp,
499 unsigned int _cs, unsigned int _ss, unsigned int _ds)
501 WARN_ON_ONCE(regs != current_pt_regs());
503 if (static_cpu_has(X86_BUG_NULL_SEG)) {
504 /* Loading zero below won't clear the base. */
505 loadsegment(fs, __USER_DS);
506 load_gs_index(__USER_DS);
510 loadsegment(es, _ds);
511 loadsegment(ds, _ds);
518 regs->flags = X86_EFLAGS_IF;
523 start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
525 start_thread_common(regs, new_ip, new_sp,
526 __USER_CS, __USER_DS, 0);
528 EXPORT_SYMBOL_GPL(start_thread);
531 void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp)
533 start_thread_common(regs, new_ip, new_sp,
534 test_thread_flag(TIF_X32)
535 ? __USER_CS : __USER32_CS,
536 __USER_DS, __USER_DS);
541 * switch_to(x,y) should switch tasks from x to y.
543 * This could still be optimized:
544 * - fold all the options into a flag word and test it with a single test.
545 * - could test fs/gs bitsliced
547 * Kprobes not supported here. Set the probe on schedule instead.
548 * Function graph tracer not supported too.
550 __visible __notrace_funcgraph struct task_struct *
551 __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
553 struct thread_struct *prev = &prev_p->thread;
554 struct thread_struct *next = &next_p->thread;
555 struct fpu *prev_fpu = &prev->fpu;
556 struct fpu *next_fpu = &next->fpu;
557 int cpu = smp_processor_id();
559 WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_ENTRY) &&
560 this_cpu_read(irq_count) != -1);
562 switch_fpu_prepare(prev_fpu, cpu);
564 /* We must save %fs and %gs before load_TLS() because
565 * %fs and %gs may be cleared by load_TLS().
567 * (e.g. xen_load_tls())
572 * Load TLS before restoring any segments so that segment loads
573 * reference the correct GDT entries.
578 * Leave lazy mode, flushing any hypercalls made here. This
579 * must be done after loading TLS entries in the GDT but before
580 * loading segments that might reference them, and and it must
581 * be done before fpu__restore(), so the TS bit is up to
584 arch_end_context_switch(next_p);
588 * Reading them only returns the selectors, but writing them (if
589 * nonzero) loads the full descriptor from the GDT or LDT. The
590 * LDT for next is loaded in switch_mm, and the GDT is loaded
593 * We therefore need to write new values to the segment
594 * registers on every context switch unless both the new and old
597 * Note that we don't need to do anything for CS and SS, as
598 * those are saved and restored as part of pt_regs.
600 savesegment(es, prev->es);
601 if (unlikely(next->es | prev->es))
602 loadsegment(es, next->es);
604 savesegment(ds, prev->ds);
605 if (unlikely(next->ds | prev->ds))
606 loadsegment(ds, next->ds);
608 x86_fsgsbase_load(prev, next);
610 switch_fpu_finish(next_fpu, cpu);
613 * Switch the PDA and FPU contexts.
615 this_cpu_write(current_task, next_p);
616 this_cpu_write(cpu_current_top_of_stack, task_top_of_stack(next_p));
619 update_task_stack(next_p);
621 switch_to_extra(prev_p, next_p);
625 * On Xen PV, IOPL bits in pt_regs->flags have no effect, and
626 * current_pt_regs()->flags may not match the current task's
627 * intended IOPL. We need to switch it manually.
629 if (unlikely(static_cpu_has(X86_FEATURE_XENPV) &&
630 prev->iopl != next->iopl))
631 xen_set_iopl_mask(next->iopl);
634 if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
636 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
637 * does not update the cached descriptor. As a result, if we
638 * do SYSRET while SS is NULL, we'll end up in user mode with
639 * SS apparently equal to __USER_DS but actually unusable.
641 * The straightforward workaround would be to fix it up just
642 * before SYSRET, but that would slow down the system call
643 * fast paths. Instead, we ensure that SS is never NULL in
644 * system call context. We do this by replacing NULL SS
645 * selectors at every context switch. SYSCALL sets up a valid
646 * SS, so the only way to get NULL is to re-enter the kernel
647 * from CPL 3 through an interrupt. Since that can't happen
648 * in the same task as a running syscall, we are guaranteed to
649 * context switch between every interrupt vector entry and a
652 * We read SS first because SS reads are much faster than
653 * writes. Out of caution, we force SS to __KERNEL_DS even if
654 * it previously had a different non-NULL value.
656 unsigned short ss_sel;
657 savesegment(ss, ss_sel);
658 if (ss_sel != __KERNEL_DS)
659 loadsegment(ss, __KERNEL_DS);
662 /* Load the Intel cache allocation PQR MSR. */
663 intel_rdt_sched_in();
668 void set_personality_64bit(void)
670 /* inherit personality from parent */
672 /* Make sure to be in 64bit mode */
673 clear_thread_flag(TIF_IA32);
674 clear_thread_flag(TIF_ADDR32);
675 clear_thread_flag(TIF_X32);
676 /* Pretend that this comes from a 64bit execve */
677 task_pt_regs(current)->orig_ax = __NR_execve;
678 current_thread_info()->status &= ~TS_COMPAT;
680 /* Ensure the corresponding mm is not marked. */
682 current->mm->context.ia32_compat = 0;
684 /* TBD: overwrites user setup. Should have two bits.
685 But 64bit processes have always behaved this way,
686 so it's not too bad. The main problem is just that
687 32bit childs are affected again. */
688 current->personality &= ~READ_IMPLIES_EXEC;
691 static void __set_personality_x32(void)
693 #ifdef CONFIG_X86_X32
694 clear_thread_flag(TIF_IA32);
695 set_thread_flag(TIF_X32);
697 current->mm->context.ia32_compat = TIF_X32;
698 current->personality &= ~READ_IMPLIES_EXEC;
700 * in_32bit_syscall() uses the presence of the x32 syscall bit
701 * flag to determine compat status. The x86 mmap() code relies on
702 * the syscall bitness so set x32 syscall bit right here to make
703 * in_32bit_syscall() work during exec().
705 * Pretend to come from a x32 execve.
707 task_pt_regs(current)->orig_ax = __NR_x32_execve | __X32_SYSCALL_BIT;
708 current_thread_info()->status &= ~TS_COMPAT;
712 static void __set_personality_ia32(void)
714 #ifdef CONFIG_IA32_EMULATION
715 set_thread_flag(TIF_IA32);
716 clear_thread_flag(TIF_X32);
718 current->mm->context.ia32_compat = TIF_IA32;
719 current->personality |= force_personality32;
720 /* Prepare the first "return" to user space */
721 task_pt_regs(current)->orig_ax = __NR_ia32_execve;
722 current_thread_info()->status |= TS_COMPAT;
726 void set_personality_ia32(bool x32)
728 /* Make sure to be in 32bit mode */
729 set_thread_flag(TIF_ADDR32);
732 __set_personality_x32();
734 __set_personality_ia32();
736 EXPORT_SYMBOL_GPL(set_personality_ia32);
738 #ifdef CONFIG_CHECKPOINT_RESTORE
739 static long prctl_map_vdso(const struct vdso_image *image, unsigned long addr)
743 ret = map_vdso_once(image, addr);
747 return (long)image->size;
751 long do_arch_prctl_64(struct task_struct *task, int option, unsigned long arg2)
757 ret = x86_gsbase_write_task(task, arg2);
761 ret = x86_fsbase_write_task(task, arg2);
765 unsigned long base = x86_fsbase_read_task(task);
767 ret = put_user(base, (unsigned long __user *)arg2);
771 unsigned long base = x86_gsbase_read_task(task);
773 ret = put_user(base, (unsigned long __user *)arg2);
777 #ifdef CONFIG_CHECKPOINT_RESTORE
778 # ifdef CONFIG_X86_X32_ABI
779 case ARCH_MAP_VDSO_X32:
780 return prctl_map_vdso(&vdso_image_x32, arg2);
782 # if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
783 case ARCH_MAP_VDSO_32:
784 return prctl_map_vdso(&vdso_image_32, arg2);
786 case ARCH_MAP_VDSO_64:
787 return prctl_map_vdso(&vdso_image_64, arg2);
798 SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
802 ret = do_arch_prctl_64(current, option, arg2);
804 ret = do_arch_prctl_common(current, option, arg2);
809 #ifdef CONFIG_IA32_EMULATION
810 COMPAT_SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
812 return do_arch_prctl_common(current, option, arg2);
816 unsigned long KSTK_ESP(struct task_struct *task)
818 return task_pt_regs(task)->sp;