2 * Copyright (C) 1995 Linus Torvalds
4 * Pentium III FXSR, SSE support
5 * Gareth Hughes <gareth@valinux.com>, May 2000
10 * CPU hotplug support - ashok.raj@intel.com
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 #ifdef CONFIG_IA32_EMULATION
58 /* Not included via unistd.h */
59 #include <asm/unistd_32_ia32.h>
62 /* Prints also some state that isn't saved in the pt_regs */
63 void __show_regs(struct pt_regs
*regs
, int all
)
65 unsigned long cr0
= 0L, cr2
= 0L, cr3
= 0L, cr4
= 0L, fs
, gs
, shadowgs
;
66 unsigned long d0
, d1
, d2
, d3
, d6
, d7
;
67 unsigned int fsindex
, gsindex
;
68 unsigned int ds
, cs
, es
;
72 if (regs
->orig_ax
!= -1)
73 pr_cont(" ORIG_RAX: %016lx\n", regs
->orig_ax
);
77 printk(KERN_DEFAULT
"RAX: %016lx RBX: %016lx RCX: %016lx\n",
78 regs
->ax
, regs
->bx
, regs
->cx
);
79 printk(KERN_DEFAULT
"RDX: %016lx RSI: %016lx RDI: %016lx\n",
80 regs
->dx
, regs
->si
, regs
->di
);
81 printk(KERN_DEFAULT
"RBP: %016lx R08: %016lx R09: %016lx\n",
82 regs
->bp
, regs
->r8
, regs
->r9
);
83 printk(KERN_DEFAULT
"R10: %016lx R11: %016lx R12: %016lx\n",
84 regs
->r10
, regs
->r11
, regs
->r12
);
85 printk(KERN_DEFAULT
"R13: %016lx R14: %016lx R15: %016lx\n",
86 regs
->r13
, regs
->r14
, regs
->r15
);
91 asm("movl %%ds,%0" : "=r" (ds
));
92 asm("movl %%cs,%0" : "=r" (cs
));
93 asm("movl %%es,%0" : "=r" (es
));
94 asm("movl %%fs,%0" : "=r" (fsindex
));
95 asm("movl %%gs,%0" : "=r" (gsindex
));
97 rdmsrl(MSR_FS_BASE
, fs
);
98 rdmsrl(MSR_GS_BASE
, gs
);
99 rdmsrl(MSR_KERNEL_GS_BASE
, shadowgs
);
106 printk(KERN_DEFAULT
"FS: %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
107 fs
, fsindex
, gs
, gsindex
, shadowgs
);
108 printk(KERN_DEFAULT
"CS: %04x DS: %04x ES: %04x CR0: %016lx\n", cs
, ds
,
110 printk(KERN_DEFAULT
"CR2: %016lx CR3: %016lx CR4: %016lx\n", cr2
, cr3
,
120 /* Only print out debug registers if they are in their non-default state. */
121 if (!((d0
== 0) && (d1
== 0) && (d2
== 0) && (d3
== 0) &&
122 (d6
== DR6_RESERVED
) && (d7
== 0x400))) {
123 printk(KERN_DEFAULT
"DR0: %016lx DR1: %016lx DR2: %016lx\n",
125 printk(KERN_DEFAULT
"DR3: %016lx DR6: %016lx DR7: %016lx\n",
129 if (boot_cpu_has(X86_FEATURE_OSPKE
))
130 printk(KERN_DEFAULT
"PKRU: %08x\n", read_pkru());
133 void release_thread(struct task_struct
*dead_task
)
136 #ifdef CONFIG_MODIFY_LDT_SYSCALL
137 if (dead_task
->mm
->context
.ldt
) {
138 pr_warn("WARNING: dead process %s still has LDT? <%p/%d>\n",
140 dead_task
->mm
->context
.ldt
->entries
,
141 dead_task
->mm
->context
.ldt
->nr_entries
);
148 enum which_selector
{
154 * Saves the FS or GS base for an outgoing thread if FSGSBASE extensions are
155 * not available. The goal is to be reasonably fast on non-FSGSBASE systems.
156 * It's forcibly inlined because it'll generate better code and this function
159 static __always_inline
void save_base_legacy(struct task_struct
*prev_p
,
160 unsigned short selector
,
161 enum which_selector which
)
163 if (likely(selector
== 0)) {
165 * On Intel (without X86_BUG_NULL_SEG), the segment base could
166 * be the pre-existing saved base or it could be zero. On AMD
167 * (with X86_BUG_NULL_SEG), the segment base could be almost
170 * This branch is very hot (it's hit twice on almost every
171 * context switch between 64-bit programs), and avoiding
172 * the RDMSR helps a lot, so we just assume that whatever
173 * value is already saved is correct. This matches historical
174 * Linux behavior, so it won't break existing applications.
176 * To avoid leaking state, on non-X86_BUG_NULL_SEG CPUs, if we
177 * report that the base is zero, it needs to actually be zero:
178 * see the corresponding logic in load_seg_legacy.
182 * If the selector is 1, 2, or 3, then the base is zero on
183 * !X86_BUG_NULL_SEG CPUs and could be anything on
184 * X86_BUG_NULL_SEG CPUs. In the latter case, Linux
185 * has never attempted to preserve the base across context
188 * If selector > 3, then it refers to a real segment, and
189 * saving the base isn't necessary.
192 prev_p
->thread
.fsbase
= 0;
194 prev_p
->thread
.gsbase
= 0;
198 static __always_inline
void save_fsgs(struct task_struct
*task
)
200 savesegment(fs
, task
->thread
.fsindex
);
201 savesegment(gs
, task
->thread
.gsindex
);
202 save_base_legacy(task
, task
->thread
.fsindex
, FS
);
203 save_base_legacy(task
, task
->thread
.gsindex
, GS
);
206 #if IS_ENABLED(CONFIG_KVM)
208 * While a process is running,current->thread.fsbase and current->thread.gsbase
209 * may not match the corresponding CPU registers (see save_base_legacy()). KVM
210 * wants an efficient way to save and restore FSBASE and GSBASE.
211 * When FSGSBASE extensions are enabled, this will have to use RD{FS,GS}BASE.
213 void save_fsgs_for_kvm(void)
217 EXPORT_SYMBOL_GPL(save_fsgs_for_kvm
);
220 static __always_inline
void loadseg(enum which_selector which
,
224 loadsegment(fs
, sel
);
229 static __always_inline
void load_seg_legacy(unsigned short prev_index
,
230 unsigned long prev_base
,
231 unsigned short next_index
,
232 unsigned long next_base
,
233 enum which_selector which
)
235 if (likely(next_index
<= 3)) {
237 * The next task is using 64-bit TLS, is not using this
238 * segment at all, or is having fun with arcane CPU features.
240 if (next_base
== 0) {
242 * Nasty case: on AMD CPUs, we need to forcibly zero
245 if (static_cpu_has_bug(X86_BUG_NULL_SEG
)) {
246 loadseg(which
, __USER_DS
);
247 loadseg(which
, next_index
);
250 * We could try to exhaustively detect cases
251 * under which we can skip the segment load,
252 * but there's really only one case that matters
253 * for performance: if both the previous and
254 * next states are fully zeroed, we can skip
257 * (This assumes that prev_base == 0 has no
258 * false positives. This is the case on
261 if (likely(prev_index
| next_index
| prev_base
))
262 loadseg(which
, next_index
);
265 if (prev_index
!= next_index
)
266 loadseg(which
, next_index
);
267 wrmsrl(which
== FS
? MSR_FS_BASE
: MSR_KERNEL_GS_BASE
,
272 * The next task is using a real segment. Loading the selector
275 loadseg(which
, next_index
);
279 int copy_thread_tls(unsigned long clone_flags
, unsigned long sp
,
280 unsigned long arg
, struct task_struct
*p
, unsigned long tls
)
283 struct pt_regs
*childregs
;
284 struct fork_frame
*fork_frame
;
285 struct inactive_task_frame
*frame
;
286 struct task_struct
*me
= current
;
288 childregs
= task_pt_regs(p
);
289 fork_frame
= container_of(childregs
, struct fork_frame
, regs
);
290 frame
= &fork_frame
->frame
;
292 frame
->ret_addr
= (unsigned long) ret_from_fork
;
293 p
->thread
.sp
= (unsigned long) fork_frame
;
294 p
->thread
.io_bitmap_ptr
= NULL
;
296 savesegment(gs
, p
->thread
.gsindex
);
297 p
->thread
.gsbase
= p
->thread
.gsindex
? 0 : me
->thread
.gsbase
;
298 savesegment(fs
, p
->thread
.fsindex
);
299 p
->thread
.fsbase
= p
->thread
.fsindex
? 0 : me
->thread
.fsbase
;
300 savesegment(es
, p
->thread
.es
);
301 savesegment(ds
, p
->thread
.ds
);
302 memset(p
->thread
.ptrace_bps
, 0, sizeof(p
->thread
.ptrace_bps
));
304 if (unlikely(p
->flags
& PF_KTHREAD
)) {
306 memset(childregs
, 0, sizeof(struct pt_regs
));
307 frame
->bx
= sp
; /* function */
312 *childregs
= *current_pt_regs();
319 if (unlikely(test_tsk_thread_flag(me
, TIF_IO_BITMAP
))) {
320 p
->thread
.io_bitmap_ptr
= kmemdup(me
->thread
.io_bitmap_ptr
,
321 IO_BITMAP_BYTES
, GFP_KERNEL
);
322 if (!p
->thread
.io_bitmap_ptr
) {
323 p
->thread
.io_bitmap_max
= 0;
326 set_tsk_thread_flag(p
, TIF_IO_BITMAP
);
330 * Set a new TLS for the child thread?
332 if (clone_flags
& CLONE_SETTLS
) {
333 #ifdef CONFIG_IA32_EMULATION
334 if (in_ia32_syscall())
335 err
= do_set_thread_area(p
, -1,
336 (struct user_desc __user
*)tls
, 0);
339 err
= do_arch_prctl_64(p
, ARCH_SET_FS
, tls
);
345 if (err
&& p
->thread
.io_bitmap_ptr
) {
346 kfree(p
->thread
.io_bitmap_ptr
);
347 p
->thread
.io_bitmap_max
= 0;
354 start_thread_common(struct pt_regs
*regs
, unsigned long new_ip
,
355 unsigned long new_sp
,
356 unsigned int _cs
, unsigned int _ss
, unsigned int _ds
)
358 WARN_ON_ONCE(regs
!= current_pt_regs());
360 if (static_cpu_has(X86_BUG_NULL_SEG
)) {
361 /* Loading zero below won't clear the base. */
362 loadsegment(fs
, __USER_DS
);
363 load_gs_index(__USER_DS
);
367 loadsegment(es
, _ds
);
368 loadsegment(ds
, _ds
);
375 regs
->flags
= X86_EFLAGS_IF
;
380 start_thread(struct pt_regs
*regs
, unsigned long new_ip
, unsigned long new_sp
)
382 start_thread_common(regs
, new_ip
, new_sp
,
383 __USER_CS
, __USER_DS
, 0);
385 EXPORT_SYMBOL_GPL(start_thread
);
388 void compat_start_thread(struct pt_regs
*regs
, u32 new_ip
, u32 new_sp
)
390 start_thread_common(regs
, new_ip
, new_sp
,
391 test_thread_flag(TIF_X32
)
392 ? __USER_CS
: __USER32_CS
,
393 __USER_DS
, __USER_DS
);
398 * switch_to(x,y) should switch tasks from x to y.
400 * This could still be optimized:
401 * - fold all the options into a flag word and test it with a single test.
402 * - could test fs/gs bitsliced
404 * Kprobes not supported here. Set the probe on schedule instead.
405 * Function graph tracer not supported too.
407 __visible __notrace_funcgraph
struct task_struct
*
408 __switch_to(struct task_struct
*prev_p
, struct task_struct
*next_p
)
410 struct thread_struct
*prev
= &prev_p
->thread
;
411 struct thread_struct
*next
= &next_p
->thread
;
412 struct fpu
*prev_fpu
= &prev
->fpu
;
413 struct fpu
*next_fpu
= &next
->fpu
;
414 int cpu
= smp_processor_id();
415 struct tss_struct
*tss
= &per_cpu(cpu_tss_rw
, cpu
);
417 WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_ENTRY
) &&
418 this_cpu_read(irq_count
) != -1);
420 switch_fpu_prepare(prev_fpu
, cpu
);
422 /* We must save %fs and %gs before load_TLS() because
423 * %fs and %gs may be cleared by load_TLS().
425 * (e.g. xen_load_tls())
430 * Load TLS before restoring any segments so that segment loads
431 * reference the correct GDT entries.
436 * Leave lazy mode, flushing any hypercalls made here. This
437 * must be done after loading TLS entries in the GDT but before
438 * loading segments that might reference them, and and it must
439 * be done before fpu__restore(), so the TS bit is up to
442 arch_end_context_switch(next_p
);
446 * Reading them only returns the selectors, but writing them (if
447 * nonzero) loads the full descriptor from the GDT or LDT. The
448 * LDT for next is loaded in switch_mm, and the GDT is loaded
451 * We therefore need to write new values to the segment
452 * registers on every context switch unless both the new and old
455 * Note that we don't need to do anything for CS and SS, as
456 * those are saved and restored as part of pt_regs.
458 savesegment(es
, prev
->es
);
459 if (unlikely(next
->es
| prev
->es
))
460 loadsegment(es
, next
->es
);
462 savesegment(ds
, prev
->ds
);
463 if (unlikely(next
->ds
| prev
->ds
))
464 loadsegment(ds
, next
->ds
);
466 load_seg_legacy(prev
->fsindex
, prev
->fsbase
,
467 next
->fsindex
, next
->fsbase
, FS
);
468 load_seg_legacy(prev
->gsindex
, prev
->gsbase
,
469 next
->gsindex
, next
->gsbase
, GS
);
471 switch_fpu_finish(next_fpu
, cpu
);
474 * Switch the PDA and FPU contexts.
476 this_cpu_write(current_task
, next_p
);
477 this_cpu_write(cpu_current_top_of_stack
, task_top_of_stack(next_p
));
480 update_task_stack(next_p
);
483 * Now maybe reload the debug registers and handle I/O bitmaps
485 if (unlikely(task_thread_info(next_p
)->flags
& _TIF_WORK_CTXSW_NEXT
||
486 task_thread_info(prev_p
)->flags
& _TIF_WORK_CTXSW_PREV
))
487 __switch_to_xtra(prev_p
, next_p
, tss
);
491 * On Xen PV, IOPL bits in pt_regs->flags have no effect, and
492 * current_pt_regs()->flags may not match the current task's
493 * intended IOPL. We need to switch it manually.
495 if (unlikely(static_cpu_has(X86_FEATURE_XENPV
) &&
496 prev
->iopl
!= next
->iopl
))
497 xen_set_iopl_mask(next
->iopl
);
500 if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS
)) {
502 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
503 * does not update the cached descriptor. As a result, if we
504 * do SYSRET while SS is NULL, we'll end up in user mode with
505 * SS apparently equal to __USER_DS but actually unusable.
507 * The straightforward workaround would be to fix it up just
508 * before SYSRET, but that would slow down the system call
509 * fast paths. Instead, we ensure that SS is never NULL in
510 * system call context. We do this by replacing NULL SS
511 * selectors at every context switch. SYSCALL sets up a valid
512 * SS, so the only way to get NULL is to re-enter the kernel
513 * from CPL 3 through an interrupt. Since that can't happen
514 * in the same task as a running syscall, we are guaranteed to
515 * context switch between every interrupt vector entry and a
518 * We read SS first because SS reads are much faster than
519 * writes. Out of caution, we force SS to __KERNEL_DS even if
520 * it previously had a different non-NULL value.
522 unsigned short ss_sel
;
523 savesegment(ss
, ss_sel
);
524 if (ss_sel
!= __KERNEL_DS
)
525 loadsegment(ss
, __KERNEL_DS
);
528 /* Load the Intel cache allocation PQR MSR. */
529 intel_rdt_sched_in();
534 void set_personality_64bit(void)
536 /* inherit personality from parent */
538 /* Make sure to be in 64bit mode */
539 clear_thread_flag(TIF_IA32
);
540 clear_thread_flag(TIF_ADDR32
);
541 clear_thread_flag(TIF_X32
);
542 /* Pretend that this comes from a 64bit execve */
543 task_pt_regs(current
)->orig_ax
= __NR_execve
;
544 current_thread_info()->status
&= ~TS_COMPAT
;
546 /* Ensure the corresponding mm is not marked. */
548 current
->mm
->context
.ia32_compat
= 0;
550 /* TBD: overwrites user setup. Should have two bits.
551 But 64bit processes have always behaved this way,
552 so it's not too bad. The main problem is just that
553 32bit childs are affected again. */
554 current
->personality
&= ~READ_IMPLIES_EXEC
;
557 static void __set_personality_x32(void)
559 #ifdef CONFIG_X86_X32
560 clear_thread_flag(TIF_IA32
);
561 set_thread_flag(TIF_X32
);
563 current
->mm
->context
.ia32_compat
= TIF_X32
;
564 current
->personality
&= ~READ_IMPLIES_EXEC
;
566 * in_compat_syscall() uses the presence of the x32 syscall bit
567 * flag to determine compat status. The x86 mmap() code relies on
568 * the syscall bitness so set x32 syscall bit right here to make
569 * in_compat_syscall() work during exec().
571 * Pretend to come from a x32 execve.
573 task_pt_regs(current
)->orig_ax
= __NR_x32_execve
| __X32_SYSCALL_BIT
;
574 current_thread_info()->status
&= ~TS_COMPAT
;
578 static void __set_personality_ia32(void)
580 #ifdef CONFIG_IA32_EMULATION
581 set_thread_flag(TIF_IA32
);
582 clear_thread_flag(TIF_X32
);
584 current
->mm
->context
.ia32_compat
= TIF_IA32
;
585 current
->personality
|= force_personality32
;
586 /* Prepare the first "return" to user space */
587 task_pt_regs(current
)->orig_ax
= __NR_ia32_execve
;
588 current_thread_info()->status
|= TS_COMPAT
;
592 void set_personality_ia32(bool x32
)
594 /* Make sure to be in 32bit mode */
595 set_thread_flag(TIF_ADDR32
);
598 __set_personality_x32();
600 __set_personality_ia32();
602 EXPORT_SYMBOL_GPL(set_personality_ia32
);
604 #ifdef CONFIG_CHECKPOINT_RESTORE
605 static long prctl_map_vdso(const struct vdso_image
*image
, unsigned long addr
)
609 ret
= map_vdso_once(image
, addr
);
613 return (long)image
->size
;
617 long do_arch_prctl_64(struct task_struct
*task
, int option
, unsigned long arg2
)
620 int doit
= task
== current
;
625 if (arg2
>= TASK_SIZE_MAX
)
628 task
->thread
.gsindex
= 0;
629 task
->thread
.gsbase
= arg2
;
632 ret
= wrmsrl_safe(MSR_KERNEL_GS_BASE
, arg2
);
637 /* Not strictly needed for fs, but do it for symmetry
639 if (arg2
>= TASK_SIZE_MAX
)
642 task
->thread
.fsindex
= 0;
643 task
->thread
.fsbase
= arg2
;
645 /* set the selector to 0 to not confuse __switch_to */
647 ret
= wrmsrl_safe(MSR_FS_BASE
, arg2
);
655 rdmsrl(MSR_FS_BASE
, base
);
657 base
= task
->thread
.fsbase
;
658 ret
= put_user(base
, (unsigned long __user
*)arg2
);
665 rdmsrl(MSR_KERNEL_GS_BASE
, base
);
667 base
= task
->thread
.gsbase
;
668 ret
= put_user(base
, (unsigned long __user
*)arg2
);
672 #ifdef CONFIG_CHECKPOINT_RESTORE
673 # ifdef CONFIG_X86_X32_ABI
674 case ARCH_MAP_VDSO_X32
:
675 return prctl_map_vdso(&vdso_image_x32
, arg2
);
677 # if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
678 case ARCH_MAP_VDSO_32
:
679 return prctl_map_vdso(&vdso_image_32
, arg2
);
681 case ARCH_MAP_VDSO_64
:
682 return prctl_map_vdso(&vdso_image_64
, arg2
);
693 SYSCALL_DEFINE2(arch_prctl
, int, option
, unsigned long, arg2
)
697 ret
= do_arch_prctl_64(current
, option
, arg2
);
699 ret
= do_arch_prctl_common(current
, option
, arg2
);
704 #ifdef CONFIG_IA32_EMULATION
705 COMPAT_SYSCALL_DEFINE2(arch_prctl
, int, option
, unsigned long, arg2
)
707 return do_arch_prctl_common(current
, option
, arg2
);
711 unsigned long KSTK_ESP(struct task_struct
*task
)
713 return task_pt_regs(task
)->sp
;