1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 1995 Linus Torvalds
5 * Pentium III FXSR, SSE support
6 * Gareth Hughes <gareth@valinux.com>, May 2000
11 * CPU hotplug support - ashok.raj@intel.com
15 * This file handles the architecture-dependent parts of process handling..
18 #include <linux/cpu.h>
19 #include <linux/errno.h>
20 #include <linux/sched.h>
21 #include <linux/sched/task.h>
22 #include <linux/sched/task_stack.h>
24 #include <linux/kernel.h>
26 #include <linux/elfcore.h>
27 #include <linux/smp.h>
28 #include <linux/slab.h>
29 #include <linux/user.h>
30 #include <linux/interrupt.h>
31 #include <linux/delay.h>
32 #include <linux/export.h>
33 #include <linux/ptrace.h>
34 #include <linux/notifier.h>
35 #include <linux/kprobes.h>
36 #include <linux/kdebug.h>
37 #include <linux/prctl.h>
38 #include <linux/uaccess.h>
40 #include <linux/ftrace.h>
41 #include <linux/syscalls.h>
43 #include <asm/pgtable.h>
44 #include <asm/processor.h>
45 #include <asm/fpu/internal.h>
46 #include <asm/mmu_context.h>
47 #include <asm/prctl.h>
49 #include <asm/proto.h>
51 #include <asm/syscalls.h>
52 #include <asm/debugreg.h>
53 #include <asm/switch_to.h>
54 #include <asm/xen/hypervisor.h>
56 #include <asm/resctrl_sched.h>
57 #include <asm/unistd.h>
58 #include <asm/fsgsbase.h>
59 #ifdef CONFIG_IA32_EMULATION
60 /* Not included via unistd.h */
61 #include <asm/unistd_32_ia32.h>
66 /* Prints also some state that isn't saved in the pt_regs */
67 void __show_regs(struct pt_regs *regs, enum show_regs_mode mode)
69 unsigned long cr0 = 0L, cr2 = 0L, cr3 = 0L, cr4 = 0L, fs, gs, shadowgs;
70 unsigned long d0, d1, d2, d3, d6, d7;
71 unsigned int fsindex, gsindex;
76 if (regs->orig_ax != -1)
77 pr_cont(" ORIG_RAX: %016lx\n", regs->orig_ax);
81 printk(KERN_DEFAULT "RAX: %016lx RBX: %016lx RCX: %016lx\n",
82 regs->ax, regs->bx, regs->cx);
83 printk(KERN_DEFAULT "RDX: %016lx RSI: %016lx RDI: %016lx\n",
84 regs->dx, regs->si, regs->di);
85 printk(KERN_DEFAULT "RBP: %016lx R08: %016lx R09: %016lx\n",
86 regs->bp, regs->r8, regs->r9);
87 printk(KERN_DEFAULT "R10: %016lx R11: %016lx R12: %016lx\n",
88 regs->r10, regs->r11, regs->r12);
89 printk(KERN_DEFAULT "R13: %016lx R14: %016lx R15: %016lx\n",
90 regs->r13, regs->r14, regs->r15);
92 if (mode == SHOW_REGS_SHORT)
95 if (mode == SHOW_REGS_USER) {
96 rdmsrl(MSR_FS_BASE, fs);
97 rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
98 printk(KERN_DEFAULT "FS: %016lx GS: %016lx\n",
103 asm("movl %%ds,%0" : "=r" (ds));
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: %04lx DS: %04x ES: %04x CR0: %016lx\n", regs->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)
146 WARN_ON(dead_task->mm);
149 enum which_selector {
155 * Saves the FS or GS base for an outgoing thread if FSGSBASE extensions are
156 * not available. The goal is to be reasonably fast on non-FSGSBASE systems.
157 * It's forcibly inlined because it'll generate better code and this function
160 static __always_inline void save_base_legacy(struct task_struct *prev_p,
161 unsigned short selector,
162 enum which_selector which)
164 if (likely(selector == 0)) {
166 * On Intel (without X86_BUG_NULL_SEG), the segment base could
167 * be the pre-existing saved base or it could be zero. On AMD
168 * (with X86_BUG_NULL_SEG), the segment base could be almost
171 * This branch is very hot (it's hit twice on almost every
172 * context switch between 64-bit programs), and avoiding
173 * the RDMSR helps a lot, so we just assume that whatever
174 * value is already saved is correct. This matches historical
175 * Linux behavior, so it won't break existing applications.
177 * To avoid leaking state, on non-X86_BUG_NULL_SEG CPUs, if we
178 * report that the base is zero, it needs to actually be zero:
179 * see the corresponding logic in load_seg_legacy.
183 * If the selector is 1, 2, or 3, then the base is zero on
184 * !X86_BUG_NULL_SEG CPUs and could be anything on
185 * X86_BUG_NULL_SEG CPUs. In the latter case, Linux
186 * has never attempted to preserve the base across context
189 * If selector > 3, then it refers to a real segment, and
190 * saving the base isn't necessary.
193 prev_p->thread.fsbase = 0;
195 prev_p->thread.gsbase = 0;
199 static __always_inline void save_fsgs(struct task_struct *task)
201 savesegment(fs, task->thread.fsindex);
202 savesegment(gs, task->thread.gsindex);
203 save_base_legacy(task, task->thread.fsindex, FS);
204 save_base_legacy(task, task->thread.gsindex, GS);
207 #if IS_ENABLED(CONFIG_KVM)
209 * While a process is running,current->thread.fsbase and current->thread.gsbase
210 * may not match the corresponding CPU registers (see save_base_legacy()). KVM
211 * wants an efficient way to save and restore FSBASE and GSBASE.
212 * When FSGSBASE extensions are enabled, this will have to use RD{FS,GS}BASE.
214 void save_fsgs_for_kvm(void)
218 EXPORT_SYMBOL_GPL(save_fsgs_for_kvm);
221 static __always_inline void loadseg(enum which_selector which,
225 loadsegment(fs, sel);
230 static __always_inline void load_seg_legacy(unsigned short prev_index,
231 unsigned long prev_base,
232 unsigned short next_index,
233 unsigned long next_base,
234 enum which_selector which)
236 if (likely(next_index <= 3)) {
238 * The next task is using 64-bit TLS, is not using this
239 * segment at all, or is having fun with arcane CPU features.
241 if (next_base == 0) {
243 * Nasty case: on AMD CPUs, we need to forcibly zero
246 if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
247 loadseg(which, __USER_DS);
248 loadseg(which, next_index);
251 * We could try to exhaustively detect cases
252 * under which we can skip the segment load,
253 * but there's really only one case that matters
254 * for performance: if both the previous and
255 * next states are fully zeroed, we can skip
258 * (This assumes that prev_base == 0 has no
259 * false positives. This is the case on
262 if (likely(prev_index | next_index | prev_base))
263 loadseg(which, next_index);
266 if (prev_index != next_index)
267 loadseg(which, next_index);
268 wrmsrl(which == FS ? MSR_FS_BASE : MSR_KERNEL_GS_BASE,
273 * The next task is using a real segment. Loading the selector
276 loadseg(which, next_index);
280 static __always_inline void x86_fsgsbase_load(struct thread_struct *prev,
281 struct thread_struct *next)
283 load_seg_legacy(prev->fsindex, prev->fsbase,
284 next->fsindex, next->fsbase, FS);
285 load_seg_legacy(prev->gsindex, prev->gsbase,
286 next->gsindex, next->gsbase, GS);
289 static unsigned long x86_fsgsbase_read_task(struct task_struct *task,
290 unsigned short selector)
292 unsigned short idx = selector >> 3;
295 if (likely((selector & SEGMENT_TI_MASK) == 0)) {
296 if (unlikely(idx >= GDT_ENTRIES))
300 * There are no user segments in the GDT with nonzero bases
301 * other than the TLS segments.
303 if (idx < GDT_ENTRY_TLS_MIN || idx > GDT_ENTRY_TLS_MAX)
306 idx -= GDT_ENTRY_TLS_MIN;
307 base = get_desc_base(&task->thread.tls_array[idx]);
309 #ifdef CONFIG_MODIFY_LDT_SYSCALL
310 struct ldt_struct *ldt;
313 * If performance here mattered, we could protect the LDT
314 * with RCU. This is a slow path, though, so we can just
317 mutex_lock(&task->mm->context.lock);
318 ldt = task->mm->context.ldt;
319 if (unlikely(!ldt || idx >= ldt->nr_entries))
322 base = get_desc_base(ldt->entries + idx);
323 mutex_unlock(&task->mm->context.lock);
332 unsigned long x86_fsbase_read_task(struct task_struct *task)
334 unsigned long fsbase;
337 fsbase = x86_fsbase_read_cpu();
338 else if (task->thread.fsindex == 0)
339 fsbase = task->thread.fsbase;
341 fsbase = x86_fsgsbase_read_task(task, task->thread.fsindex);
346 unsigned long x86_gsbase_read_task(struct task_struct *task)
348 unsigned long gsbase;
351 gsbase = x86_gsbase_read_cpu_inactive();
352 else if (task->thread.gsindex == 0)
353 gsbase = task->thread.gsbase;
355 gsbase = x86_fsgsbase_read_task(task, task->thread.gsindex);
360 void x86_fsbase_write_task(struct task_struct *task, unsigned long fsbase)
362 WARN_ON_ONCE(task == current);
364 task->thread.fsbase = fsbase;
367 void x86_gsbase_write_task(struct task_struct *task, unsigned long gsbase)
369 WARN_ON_ONCE(task == current);
371 task->thread.gsbase = gsbase;
374 int copy_thread_tls(unsigned long clone_flags, unsigned long sp,
375 unsigned long arg, struct task_struct *p, unsigned long tls)
378 struct pt_regs *childregs;
379 struct fork_frame *fork_frame;
380 struct inactive_task_frame *frame;
381 struct task_struct *me = current;
383 childregs = task_pt_regs(p);
384 fork_frame = container_of(childregs, struct fork_frame, regs);
385 frame = &fork_frame->frame;
388 frame->ret_addr = (unsigned long) ret_from_fork;
389 p->thread.sp = (unsigned long) fork_frame;
390 p->thread.io_bitmap_ptr = NULL;
392 savesegment(gs, p->thread.gsindex);
393 p->thread.gsbase = p->thread.gsindex ? 0 : me->thread.gsbase;
394 savesegment(fs, p->thread.fsindex);
395 p->thread.fsbase = p->thread.fsindex ? 0 : me->thread.fsbase;
396 savesegment(es, p->thread.es);
397 savesegment(ds, p->thread.ds);
398 memset(p->thread.ptrace_bps, 0, sizeof(p->thread.ptrace_bps));
400 if (unlikely(p->flags & PF_KTHREAD)) {
402 memset(childregs, 0, sizeof(struct pt_regs));
403 frame->bx = sp; /* function */
408 *childregs = *current_pt_regs();
415 if (unlikely(test_tsk_thread_flag(me, TIF_IO_BITMAP))) {
416 p->thread.io_bitmap_ptr = kmemdup(me->thread.io_bitmap_ptr,
417 IO_BITMAP_BYTES, GFP_KERNEL);
418 if (!p->thread.io_bitmap_ptr) {
419 p->thread.io_bitmap_max = 0;
422 set_tsk_thread_flag(p, TIF_IO_BITMAP);
426 * Set a new TLS for the child thread?
428 if (clone_flags & CLONE_SETTLS) {
429 #ifdef CONFIG_IA32_EMULATION
430 if (in_ia32_syscall())
431 err = do_set_thread_area(p, -1,
432 (struct user_desc __user *)tls, 0);
435 err = do_arch_prctl_64(p, ARCH_SET_FS, tls);
441 if (err && p->thread.io_bitmap_ptr) {
442 kfree(p->thread.io_bitmap_ptr);
443 p->thread.io_bitmap_max = 0;
450 start_thread_common(struct pt_regs *regs, unsigned long new_ip,
451 unsigned long new_sp,
452 unsigned int _cs, unsigned int _ss, unsigned int _ds)
454 WARN_ON_ONCE(regs != current_pt_regs());
456 if (static_cpu_has(X86_BUG_NULL_SEG)) {
457 /* Loading zero below won't clear the base. */
458 loadsegment(fs, __USER_DS);
459 load_gs_index(__USER_DS);
463 loadsegment(es, _ds);
464 loadsegment(ds, _ds);
471 regs->flags = X86_EFLAGS_IF;
476 start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
478 start_thread_common(regs, new_ip, new_sp,
479 __USER_CS, __USER_DS, 0);
481 EXPORT_SYMBOL_GPL(start_thread);
484 void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp)
486 start_thread_common(regs, new_ip, new_sp,
487 test_thread_flag(TIF_X32)
488 ? __USER_CS : __USER32_CS,
489 __USER_DS, __USER_DS);
494 * switch_to(x,y) should switch tasks from x to y.
496 * This could still be optimized:
497 * - fold all the options into a flag word and test it with a single test.
498 * - could test fs/gs bitsliced
500 * Kprobes not supported here. Set the probe on schedule instead.
501 * Function graph tracer not supported too.
503 __visible __notrace_funcgraph struct task_struct *
504 __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
506 struct thread_struct *prev = &prev_p->thread;
507 struct thread_struct *next = &next_p->thread;
508 int cpu = smp_processor_id();
510 WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_ENTRY) &&
511 this_cpu_read(irq_count) != -1);
513 if (!test_thread_flag(TIF_NEED_FPU_LOAD))
514 switch_fpu_prepare(prev_p, cpu);
516 /* We must save %fs and %gs before load_TLS() because
517 * %fs and %gs may be cleared by load_TLS().
519 * (e.g. xen_load_tls())
524 * Load TLS before restoring any segments so that segment loads
525 * reference the correct GDT entries.
530 * Leave lazy mode, flushing any hypercalls made here. This
531 * must be done after loading TLS entries in the GDT but before
532 * loading segments that might reference them.
534 arch_end_context_switch(next_p);
538 * Reading them only returns the selectors, but writing them (if
539 * nonzero) loads the full descriptor from the GDT or LDT. The
540 * LDT for next is loaded in switch_mm, and the GDT is loaded
543 * We therefore need to write new values to the segment
544 * registers on every context switch unless both the new and old
547 * Note that we don't need to do anything for CS and SS, as
548 * those are saved and restored as part of pt_regs.
550 savesegment(es, prev->es);
551 if (unlikely(next->es | prev->es))
552 loadsegment(es, next->es);
554 savesegment(ds, prev->ds);
555 if (unlikely(next->ds | prev->ds))
556 loadsegment(ds, next->ds);
558 x86_fsgsbase_load(prev, next);
561 * Switch the PDA and FPU contexts.
563 this_cpu_write(current_task, next_p);
564 this_cpu_write(cpu_current_top_of_stack, task_top_of_stack(next_p));
566 switch_fpu_finish(next_p);
569 update_task_stack(next_p);
571 switch_to_extra(prev_p, next_p);
575 * On Xen PV, IOPL bits in pt_regs->flags have no effect, and
576 * current_pt_regs()->flags may not match the current task's
577 * intended IOPL. We need to switch it manually.
579 if (unlikely(static_cpu_has(X86_FEATURE_XENPV) &&
580 prev->iopl != next->iopl))
581 xen_set_iopl_mask(next->iopl);
584 if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
586 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
587 * does not update the cached descriptor. As a result, if we
588 * do SYSRET while SS is NULL, we'll end up in user mode with
589 * SS apparently equal to __USER_DS but actually unusable.
591 * The straightforward workaround would be to fix it up just
592 * before SYSRET, but that would slow down the system call
593 * fast paths. Instead, we ensure that SS is never NULL in
594 * system call context. We do this by replacing NULL SS
595 * selectors at every context switch. SYSCALL sets up a valid
596 * SS, so the only way to get NULL is to re-enter the kernel
597 * from CPL 3 through an interrupt. Since that can't happen
598 * in the same task as a running syscall, we are guaranteed to
599 * context switch between every interrupt vector entry and a
602 * We read SS first because SS reads are much faster than
603 * writes. Out of caution, we force SS to __KERNEL_DS even if
604 * it previously had a different non-NULL value.
606 unsigned short ss_sel;
607 savesegment(ss, ss_sel);
608 if (ss_sel != __KERNEL_DS)
609 loadsegment(ss, __KERNEL_DS);
612 /* Load the Intel cache allocation PQR MSR. */
613 resctrl_sched_in(next_p);
618 void set_personality_64bit(void)
620 /* inherit personality from parent */
622 /* Make sure to be in 64bit mode */
623 clear_thread_flag(TIF_IA32);
624 clear_thread_flag(TIF_ADDR32);
625 clear_thread_flag(TIF_X32);
626 /* Pretend that this comes from a 64bit execve */
627 task_pt_regs(current)->orig_ax = __NR_execve;
628 current_thread_info()->status &= ~TS_COMPAT;
630 /* Ensure the corresponding mm is not marked. */
632 current->mm->context.ia32_compat = 0;
634 /* TBD: overwrites user setup. Should have two bits.
635 But 64bit processes have always behaved this way,
636 so it's not too bad. The main problem is just that
637 32bit children are affected again. */
638 current->personality &= ~READ_IMPLIES_EXEC;
641 static void __set_personality_x32(void)
643 #ifdef CONFIG_X86_X32
644 clear_thread_flag(TIF_IA32);
645 set_thread_flag(TIF_X32);
647 current->mm->context.ia32_compat = TIF_X32;
648 current->personality &= ~READ_IMPLIES_EXEC;
650 * in_32bit_syscall() uses the presence of the x32 syscall bit
651 * flag to determine compat status. The x86 mmap() code relies on
652 * the syscall bitness so set x32 syscall bit right here to make
653 * in_32bit_syscall() work during exec().
655 * Pretend to come from a x32 execve.
657 task_pt_regs(current)->orig_ax = __NR_x32_execve | __X32_SYSCALL_BIT;
658 current_thread_info()->status &= ~TS_COMPAT;
662 static void __set_personality_ia32(void)
664 #ifdef CONFIG_IA32_EMULATION
665 set_thread_flag(TIF_IA32);
666 clear_thread_flag(TIF_X32);
668 current->mm->context.ia32_compat = TIF_IA32;
669 current->personality |= force_personality32;
670 /* Prepare the first "return" to user space */
671 task_pt_regs(current)->orig_ax = __NR_ia32_execve;
672 current_thread_info()->status |= TS_COMPAT;
676 void set_personality_ia32(bool x32)
678 /* Make sure to be in 32bit mode */
679 set_thread_flag(TIF_ADDR32);
682 __set_personality_x32();
684 __set_personality_ia32();
686 EXPORT_SYMBOL_GPL(set_personality_ia32);
688 #ifdef CONFIG_CHECKPOINT_RESTORE
689 static long prctl_map_vdso(const struct vdso_image *image, unsigned long addr)
693 ret = map_vdso_once(image, addr);
697 return (long)image->size;
701 long do_arch_prctl_64(struct task_struct *task, int option, unsigned long arg2)
707 if (unlikely(arg2 >= TASK_SIZE_MAX))
712 * ARCH_SET_GS has always overwritten the index
713 * and the base. Zero is the most sensible value
714 * to put in the index, and is the only value that
715 * makes any sense if FSGSBASE is unavailable.
717 if (task == current) {
719 x86_gsbase_write_cpu_inactive(arg2);
722 * On non-FSGSBASE systems, save_base_legacy() expects
723 * that we also fill in thread.gsbase.
725 task->thread.gsbase = arg2;
728 task->thread.gsindex = 0;
729 x86_gsbase_write_task(task, arg2);
736 * Not strictly needed for %fs, but do it for symmetry
739 if (unlikely(arg2 >= TASK_SIZE_MAX))
744 * Set the selector to 0 for the same reason
747 if (task == current) {
749 x86_fsbase_write_cpu(arg2);
752 * On non-FSGSBASE systems, save_base_legacy() expects
753 * that we also fill in thread.fsbase.
755 task->thread.fsbase = arg2;
757 task->thread.fsindex = 0;
758 x86_fsbase_write_task(task, arg2);
764 unsigned long base = x86_fsbase_read_task(task);
766 ret = put_user(base, (unsigned long __user *)arg2);
770 unsigned long base = x86_gsbase_read_task(task);
772 ret = put_user(base, (unsigned long __user *)arg2);
776 #ifdef CONFIG_CHECKPOINT_RESTORE
777 # ifdef CONFIG_X86_X32_ABI
778 case ARCH_MAP_VDSO_X32:
779 return prctl_map_vdso(&vdso_image_x32, arg2);
781 # if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
782 case ARCH_MAP_VDSO_32:
783 return prctl_map_vdso(&vdso_image_32, arg2);
785 case ARCH_MAP_VDSO_64:
786 return prctl_map_vdso(&vdso_image_64, arg2);
797 SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
801 ret = do_arch_prctl_64(current, option, arg2);
803 ret = do_arch_prctl_common(current, option, arg2);
808 #ifdef CONFIG_IA32_EMULATION
809 COMPAT_SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
811 return do_arch_prctl_common(current, option, arg2);
815 unsigned long KSTK_ESP(struct task_struct *task)
817 return task_pt_regs(task)->sp;