xref: /openbmc/linux/arch/x86/kernel/process_64.c (revision 82e6fdd6)
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/sched/task.h>
21 #include <linux/sched/task_stack.h>
22 #include <linux/fs.h>
23 #include <linux/kernel.h>
24 #include <linux/mm.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>
38 #include <linux/io.h>
39 #include <linux/ftrace.h>
40 #include <linux/syscalls.h>
41 
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>
47 #include <asm/desc.h>
48 #include <asm/proto.h>
49 #include <asm/ia32.h>
50 #include <asm/syscalls.h>
51 #include <asm/debugreg.h>
52 #include <asm/switch_to.h>
53 #include <asm/xen/hypervisor.h>
54 #include <asm/vdso.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>
60 #endif
61 
62 __visible DEFINE_PER_CPU(unsigned long, rsp_scratch);
63 
64 /* Prints also some state that isn't saved in the pt_regs */
65 void __show_regs(struct pt_regs *regs, int all)
66 {
67 	unsigned long cr0 = 0L, cr2 = 0L, cr3 = 0L, cr4 = 0L, fs, gs, shadowgs;
68 	unsigned long d0, d1, d2, d3, d6, d7;
69 	unsigned int fsindex, gsindex;
70 	unsigned int ds, cs, es;
71 
72 	show_iret_regs(regs);
73 
74 	if (regs->orig_ax != -1)
75 		pr_cont(" ORIG_RAX: %016lx\n", regs->orig_ax);
76 	else
77 		pr_cont("\n");
78 
79 	printk(KERN_DEFAULT "RAX: %016lx RBX: %016lx RCX: %016lx\n",
80 	       regs->ax, regs->bx, regs->cx);
81 	printk(KERN_DEFAULT "RDX: %016lx RSI: %016lx RDI: %016lx\n",
82 	       regs->dx, regs->si, regs->di);
83 	printk(KERN_DEFAULT "RBP: %016lx R08: %016lx R09: %016lx\n",
84 	       regs->bp, regs->r8, regs->r9);
85 	printk(KERN_DEFAULT "R10: %016lx R11: %016lx R12: %016lx\n",
86 	       regs->r10, regs->r11, regs->r12);
87 	printk(KERN_DEFAULT "R13: %016lx R14: %016lx R15: %016lx\n",
88 	       regs->r13, regs->r14, regs->r15);
89 
90 	if (!all)
91 		return;
92 
93 	asm("movl %%ds,%0" : "=r" (ds));
94 	asm("movl %%cs,%0" : "=r" (cs));
95 	asm("movl %%es,%0" : "=r" (es));
96 	asm("movl %%fs,%0" : "=r" (fsindex));
97 	asm("movl %%gs,%0" : "=r" (gsindex));
98 
99 	rdmsrl(MSR_FS_BASE, fs);
100 	rdmsrl(MSR_GS_BASE, gs);
101 	rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
102 
103 	cr0 = read_cr0();
104 	cr2 = read_cr2();
105 	cr3 = __read_cr3();
106 	cr4 = __read_cr4();
107 
108 	printk(KERN_DEFAULT "FS:  %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
109 	       fs, fsindex, gs, gsindex, shadowgs);
110 	printk(KERN_DEFAULT "CS:  %04x DS: %04x ES: %04x CR0: %016lx\n", cs, ds,
111 			es, cr0);
112 	printk(KERN_DEFAULT "CR2: %016lx CR3: %016lx CR4: %016lx\n", cr2, cr3,
113 			cr4);
114 
115 	get_debugreg(d0, 0);
116 	get_debugreg(d1, 1);
117 	get_debugreg(d2, 2);
118 	get_debugreg(d3, 3);
119 	get_debugreg(d6, 6);
120 	get_debugreg(d7, 7);
121 
122 	/* Only print out debug registers if they are in their non-default state. */
123 	if (!((d0 == 0) && (d1 == 0) && (d2 == 0) && (d3 == 0) &&
124 	    (d6 == DR6_RESERVED) && (d7 == 0x400))) {
125 		printk(KERN_DEFAULT "DR0: %016lx DR1: %016lx DR2: %016lx\n",
126 		       d0, d1, d2);
127 		printk(KERN_DEFAULT "DR3: %016lx DR6: %016lx DR7: %016lx\n",
128 		       d3, d6, d7);
129 	}
130 
131 	if (boot_cpu_has(X86_FEATURE_OSPKE))
132 		printk(KERN_DEFAULT "PKRU: %08x\n", read_pkru());
133 }
134 
135 void release_thread(struct task_struct *dead_task)
136 {
137 	if (dead_task->mm) {
138 #ifdef CONFIG_MODIFY_LDT_SYSCALL
139 		if (dead_task->mm->context.ldt) {
140 			pr_warn("WARNING: dead process %s still has LDT? <%p/%d>\n",
141 				dead_task->comm,
142 				dead_task->mm->context.ldt->entries,
143 				dead_task->mm->context.ldt->nr_entries);
144 			BUG();
145 		}
146 #endif
147 	}
148 }
149 
150 enum which_selector {
151 	FS,
152 	GS
153 };
154 
155 /*
156  * Saves the FS or GS base for an outgoing thread if FSGSBASE extensions are
157  * not available.  The goal is to be reasonably fast on non-FSGSBASE systems.
158  * It's forcibly inlined because it'll generate better code and this function
159  * is hot.
160  */
161 static __always_inline void save_base_legacy(struct task_struct *prev_p,
162 					     unsigned short selector,
163 					     enum which_selector which)
164 {
165 	if (likely(selector == 0)) {
166 		/*
167 		 * On Intel (without X86_BUG_NULL_SEG), the segment base could
168 		 * be the pre-existing saved base or it could be zero.  On AMD
169 		 * (with X86_BUG_NULL_SEG), the segment base could be almost
170 		 * anything.
171 		 *
172 		 * This branch is very hot (it's hit twice on almost every
173 		 * context switch between 64-bit programs), and avoiding
174 		 * the RDMSR helps a lot, so we just assume that whatever
175 		 * value is already saved is correct.  This matches historical
176 		 * Linux behavior, so it won't break existing applications.
177 		 *
178 		 * To avoid leaking state, on non-X86_BUG_NULL_SEG CPUs, if we
179 		 * report that the base is zero, it needs to actually be zero:
180 		 * see the corresponding logic in load_seg_legacy.
181 		 */
182 	} else {
183 		/*
184 		 * If the selector is 1, 2, or 3, then the base is zero on
185 		 * !X86_BUG_NULL_SEG CPUs and could be anything on
186 		 * X86_BUG_NULL_SEG CPUs.  In the latter case, Linux
187 		 * has never attempted to preserve the base across context
188 		 * switches.
189 		 *
190 		 * If selector > 3, then it refers to a real segment, and
191 		 * saving the base isn't necessary.
192 		 */
193 		if (which == FS)
194 			prev_p->thread.fsbase = 0;
195 		else
196 			prev_p->thread.gsbase = 0;
197 	}
198 }
199 
200 static __always_inline void save_fsgs(struct task_struct *task)
201 {
202 	savesegment(fs, task->thread.fsindex);
203 	savesegment(gs, task->thread.gsindex);
204 	save_base_legacy(task, task->thread.fsindex, FS);
205 	save_base_legacy(task, task->thread.gsindex, GS);
206 }
207 
208 static __always_inline void loadseg(enum which_selector which,
209 				    unsigned short sel)
210 {
211 	if (which == FS)
212 		loadsegment(fs, sel);
213 	else
214 		load_gs_index(sel);
215 }
216 
217 static __always_inline void load_seg_legacy(unsigned short prev_index,
218 					    unsigned long prev_base,
219 					    unsigned short next_index,
220 					    unsigned long next_base,
221 					    enum which_selector which)
222 {
223 	if (likely(next_index <= 3)) {
224 		/*
225 		 * The next task is using 64-bit TLS, is not using this
226 		 * segment at all, or is having fun with arcane CPU features.
227 		 */
228 		if (next_base == 0) {
229 			/*
230 			 * Nasty case: on AMD CPUs, we need to forcibly zero
231 			 * the base.
232 			 */
233 			if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
234 				loadseg(which, __USER_DS);
235 				loadseg(which, next_index);
236 			} else {
237 				/*
238 				 * We could try to exhaustively detect cases
239 				 * under which we can skip the segment load,
240 				 * but there's really only one case that matters
241 				 * for performance: if both the previous and
242 				 * next states are fully zeroed, we can skip
243 				 * the load.
244 				 *
245 				 * (This assumes that prev_base == 0 has no
246 				 * false positives.  This is the case on
247 				 * Intel-style CPUs.)
248 				 */
249 				if (likely(prev_index | next_index | prev_base))
250 					loadseg(which, next_index);
251 			}
252 		} else {
253 			if (prev_index != next_index)
254 				loadseg(which, next_index);
255 			wrmsrl(which == FS ? MSR_FS_BASE : MSR_KERNEL_GS_BASE,
256 			       next_base);
257 		}
258 	} else {
259 		/*
260 		 * The next task is using a real segment.  Loading the selector
261 		 * is sufficient.
262 		 */
263 		loadseg(which, next_index);
264 	}
265 }
266 
267 int copy_thread_tls(unsigned long clone_flags, unsigned long sp,
268 		unsigned long arg, struct task_struct *p, unsigned long tls)
269 {
270 	int err;
271 	struct pt_regs *childregs;
272 	struct fork_frame *fork_frame;
273 	struct inactive_task_frame *frame;
274 	struct task_struct *me = current;
275 
276 	childregs = task_pt_regs(p);
277 	fork_frame = container_of(childregs, struct fork_frame, regs);
278 	frame = &fork_frame->frame;
279 	frame->bp = 0;
280 	frame->ret_addr = (unsigned long) ret_from_fork;
281 	p->thread.sp = (unsigned long) fork_frame;
282 	p->thread.io_bitmap_ptr = NULL;
283 
284 	savesegment(gs, p->thread.gsindex);
285 	p->thread.gsbase = p->thread.gsindex ? 0 : me->thread.gsbase;
286 	savesegment(fs, p->thread.fsindex);
287 	p->thread.fsbase = p->thread.fsindex ? 0 : me->thread.fsbase;
288 	savesegment(es, p->thread.es);
289 	savesegment(ds, p->thread.ds);
290 	memset(p->thread.ptrace_bps, 0, sizeof(p->thread.ptrace_bps));
291 
292 	if (unlikely(p->flags & PF_KTHREAD)) {
293 		/* kernel thread */
294 		memset(childregs, 0, sizeof(struct pt_regs));
295 		frame->bx = sp;		/* function */
296 		frame->r12 = arg;
297 		return 0;
298 	}
299 	frame->bx = 0;
300 	*childregs = *current_pt_regs();
301 
302 	childregs->ax = 0;
303 	if (sp)
304 		childregs->sp = sp;
305 
306 	err = -ENOMEM;
307 	if (unlikely(test_tsk_thread_flag(me, TIF_IO_BITMAP))) {
308 		p->thread.io_bitmap_ptr = kmemdup(me->thread.io_bitmap_ptr,
309 						  IO_BITMAP_BYTES, GFP_KERNEL);
310 		if (!p->thread.io_bitmap_ptr) {
311 			p->thread.io_bitmap_max = 0;
312 			return -ENOMEM;
313 		}
314 		set_tsk_thread_flag(p, TIF_IO_BITMAP);
315 	}
316 
317 	/*
318 	 * Set a new TLS for the child thread?
319 	 */
320 	if (clone_flags & CLONE_SETTLS) {
321 #ifdef CONFIG_IA32_EMULATION
322 		if (in_ia32_syscall())
323 			err = do_set_thread_area(p, -1,
324 				(struct user_desc __user *)tls, 0);
325 		else
326 #endif
327 			err = do_arch_prctl_64(p, ARCH_SET_FS, tls);
328 		if (err)
329 			goto out;
330 	}
331 	err = 0;
332 out:
333 	if (err && p->thread.io_bitmap_ptr) {
334 		kfree(p->thread.io_bitmap_ptr);
335 		p->thread.io_bitmap_max = 0;
336 	}
337 
338 	return err;
339 }
340 
341 static void
342 start_thread_common(struct pt_regs *regs, unsigned long new_ip,
343 		    unsigned long new_sp,
344 		    unsigned int _cs, unsigned int _ss, unsigned int _ds)
345 {
346 	WARN_ON_ONCE(regs != current_pt_regs());
347 
348 	if (static_cpu_has(X86_BUG_NULL_SEG)) {
349 		/* Loading zero below won't clear the base. */
350 		loadsegment(fs, __USER_DS);
351 		load_gs_index(__USER_DS);
352 	}
353 
354 	loadsegment(fs, 0);
355 	loadsegment(es, _ds);
356 	loadsegment(ds, _ds);
357 	load_gs_index(0);
358 
359 	regs->ip		= new_ip;
360 	regs->sp		= new_sp;
361 	regs->cs		= _cs;
362 	regs->ss		= _ss;
363 	regs->flags		= X86_EFLAGS_IF;
364 	force_iret();
365 }
366 
367 void
368 start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
369 {
370 	start_thread_common(regs, new_ip, new_sp,
371 			    __USER_CS, __USER_DS, 0);
372 }
373 
374 #ifdef CONFIG_COMPAT
375 void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp)
376 {
377 	start_thread_common(regs, new_ip, new_sp,
378 			    test_thread_flag(TIF_X32)
379 			    ? __USER_CS : __USER32_CS,
380 			    __USER_DS, __USER_DS);
381 }
382 #endif
383 
384 /*
385  *	switch_to(x,y) should switch tasks from x to y.
386  *
387  * This could still be optimized:
388  * - fold all the options into a flag word and test it with a single test.
389  * - could test fs/gs bitsliced
390  *
391  * Kprobes not supported here. Set the probe on schedule instead.
392  * Function graph tracer not supported too.
393  */
394 __visible __notrace_funcgraph struct task_struct *
395 __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
396 {
397 	struct thread_struct *prev = &prev_p->thread;
398 	struct thread_struct *next = &next_p->thread;
399 	struct fpu *prev_fpu = &prev->fpu;
400 	struct fpu *next_fpu = &next->fpu;
401 	int cpu = smp_processor_id();
402 	struct tss_struct *tss = &per_cpu(cpu_tss_rw, cpu);
403 
404 	WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_ENTRY) &&
405 		     this_cpu_read(irq_count) != -1);
406 
407 	switch_fpu_prepare(prev_fpu, cpu);
408 
409 	/* We must save %fs and %gs before load_TLS() because
410 	 * %fs and %gs may be cleared by load_TLS().
411 	 *
412 	 * (e.g. xen_load_tls())
413 	 */
414 	save_fsgs(prev_p);
415 
416 	/*
417 	 * Load TLS before restoring any segments so that segment loads
418 	 * reference the correct GDT entries.
419 	 */
420 	load_TLS(next, cpu);
421 
422 	/*
423 	 * Leave lazy mode, flushing any hypercalls made here.  This
424 	 * must be done after loading TLS entries in the GDT but before
425 	 * loading segments that might reference them, and and it must
426 	 * be done before fpu__restore(), so the TS bit is up to
427 	 * date.
428 	 */
429 	arch_end_context_switch(next_p);
430 
431 	/* Switch DS and ES.
432 	 *
433 	 * Reading them only returns the selectors, but writing them (if
434 	 * nonzero) loads the full descriptor from the GDT or LDT.  The
435 	 * LDT for next is loaded in switch_mm, and the GDT is loaded
436 	 * above.
437 	 *
438 	 * We therefore need to write new values to the segment
439 	 * registers on every context switch unless both the new and old
440 	 * values are zero.
441 	 *
442 	 * Note that we don't need to do anything for CS and SS, as
443 	 * those are saved and restored as part of pt_regs.
444 	 */
445 	savesegment(es, prev->es);
446 	if (unlikely(next->es | prev->es))
447 		loadsegment(es, next->es);
448 
449 	savesegment(ds, prev->ds);
450 	if (unlikely(next->ds | prev->ds))
451 		loadsegment(ds, next->ds);
452 
453 	load_seg_legacy(prev->fsindex, prev->fsbase,
454 			next->fsindex, next->fsbase, FS);
455 	load_seg_legacy(prev->gsindex, prev->gsbase,
456 			next->gsindex, next->gsbase, GS);
457 
458 	switch_fpu_finish(next_fpu, cpu);
459 
460 	/*
461 	 * Switch the PDA and FPU contexts.
462 	 */
463 	this_cpu_write(current_task, next_p);
464 	this_cpu_write(cpu_current_top_of_stack, task_top_of_stack(next_p));
465 
466 	/* Reload sp0. */
467 	update_sp0(next_p);
468 
469 	/*
470 	 * Now maybe reload the debug registers and handle I/O bitmaps
471 	 */
472 	if (unlikely(task_thread_info(next_p)->flags & _TIF_WORK_CTXSW_NEXT ||
473 		     task_thread_info(prev_p)->flags & _TIF_WORK_CTXSW_PREV))
474 		__switch_to_xtra(prev_p, next_p, tss);
475 
476 #ifdef CONFIG_XEN_PV
477 	/*
478 	 * On Xen PV, IOPL bits in pt_regs->flags have no effect, and
479 	 * current_pt_regs()->flags may not match the current task's
480 	 * intended IOPL.  We need to switch it manually.
481 	 */
482 	if (unlikely(static_cpu_has(X86_FEATURE_XENPV) &&
483 		     prev->iopl != next->iopl))
484 		xen_set_iopl_mask(next->iopl);
485 #endif
486 
487 	if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
488 		/*
489 		 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
490 		 * does not update the cached descriptor.  As a result, if we
491 		 * do SYSRET while SS is NULL, we'll end up in user mode with
492 		 * SS apparently equal to __USER_DS but actually unusable.
493 		 *
494 		 * The straightforward workaround would be to fix it up just
495 		 * before SYSRET, but that would slow down the system call
496 		 * fast paths.  Instead, we ensure that SS is never NULL in
497 		 * system call context.  We do this by replacing NULL SS
498 		 * selectors at every context switch.  SYSCALL sets up a valid
499 		 * SS, so the only way to get NULL is to re-enter the kernel
500 		 * from CPL 3 through an interrupt.  Since that can't happen
501 		 * in the same task as a running syscall, we are guaranteed to
502 		 * context switch between every interrupt vector entry and a
503 		 * subsequent SYSRET.
504 		 *
505 		 * We read SS first because SS reads are much faster than
506 		 * writes.  Out of caution, we force SS to __KERNEL_DS even if
507 		 * it previously had a different non-NULL value.
508 		 */
509 		unsigned short ss_sel;
510 		savesegment(ss, ss_sel);
511 		if (ss_sel != __KERNEL_DS)
512 			loadsegment(ss, __KERNEL_DS);
513 	}
514 
515 	/* Load the Intel cache allocation PQR MSR. */
516 	intel_rdt_sched_in();
517 
518 	return prev_p;
519 }
520 
521 void set_personality_64bit(void)
522 {
523 	/* inherit personality from parent */
524 
525 	/* Make sure to be in 64bit mode */
526 	clear_thread_flag(TIF_IA32);
527 	clear_thread_flag(TIF_ADDR32);
528 	clear_thread_flag(TIF_X32);
529 	/* Pretend that this comes from a 64bit execve */
530 	task_pt_regs(current)->orig_ax = __NR_execve;
531 
532 	/* Ensure the corresponding mm is not marked. */
533 	if (current->mm)
534 		current->mm->context.ia32_compat = 0;
535 
536 	/* TBD: overwrites user setup. Should have two bits.
537 	   But 64bit processes have always behaved this way,
538 	   so it's not too bad. The main problem is just that
539 	   32bit childs are affected again. */
540 	current->personality &= ~READ_IMPLIES_EXEC;
541 }
542 
543 static void __set_personality_x32(void)
544 {
545 #ifdef CONFIG_X86_X32
546 	clear_thread_flag(TIF_IA32);
547 	set_thread_flag(TIF_X32);
548 	if (current->mm)
549 		current->mm->context.ia32_compat = TIF_X32;
550 	current->personality &= ~READ_IMPLIES_EXEC;
551 	/*
552 	 * in_compat_syscall() uses the presence of the x32 syscall bit
553 	 * flag to determine compat status.  The x86 mmap() code relies on
554 	 * the syscall bitness so set x32 syscall bit right here to make
555 	 * in_compat_syscall() work during exec().
556 	 *
557 	 * Pretend to come from a x32 execve.
558 	 */
559 	task_pt_regs(current)->orig_ax = __NR_x32_execve | __X32_SYSCALL_BIT;
560 	current_thread_info()->status &= ~TS_COMPAT;
561 #endif
562 }
563 
564 static void __set_personality_ia32(void)
565 {
566 #ifdef CONFIG_IA32_EMULATION
567 	set_thread_flag(TIF_IA32);
568 	clear_thread_flag(TIF_X32);
569 	if (current->mm)
570 		current->mm->context.ia32_compat = TIF_IA32;
571 	current->personality |= force_personality32;
572 	/* Prepare the first "return" to user space */
573 	task_pt_regs(current)->orig_ax = __NR_ia32_execve;
574 	current_thread_info()->status |= TS_COMPAT;
575 #endif
576 }
577 
578 void set_personality_ia32(bool x32)
579 {
580 	/* Make sure to be in 32bit mode */
581 	set_thread_flag(TIF_ADDR32);
582 
583 	if (x32)
584 		__set_personality_x32();
585 	else
586 		__set_personality_ia32();
587 }
588 EXPORT_SYMBOL_GPL(set_personality_ia32);
589 
590 #ifdef CONFIG_CHECKPOINT_RESTORE
591 static long prctl_map_vdso(const struct vdso_image *image, unsigned long addr)
592 {
593 	int ret;
594 
595 	ret = map_vdso_once(image, addr);
596 	if (ret)
597 		return ret;
598 
599 	return (long)image->size;
600 }
601 #endif
602 
603 long do_arch_prctl_64(struct task_struct *task, int option, unsigned long arg2)
604 {
605 	int ret = 0;
606 	int doit = task == current;
607 	int cpu;
608 
609 	switch (option) {
610 	case ARCH_SET_GS:
611 		if (arg2 >= TASK_SIZE_MAX)
612 			return -EPERM;
613 		cpu = get_cpu();
614 		task->thread.gsindex = 0;
615 		task->thread.gsbase = arg2;
616 		if (doit) {
617 			load_gs_index(0);
618 			ret = wrmsrl_safe(MSR_KERNEL_GS_BASE, arg2);
619 		}
620 		put_cpu();
621 		break;
622 	case ARCH_SET_FS:
623 		/* Not strictly needed for fs, but do it for symmetry
624 		   with gs */
625 		if (arg2 >= TASK_SIZE_MAX)
626 			return -EPERM;
627 		cpu = get_cpu();
628 		task->thread.fsindex = 0;
629 		task->thread.fsbase = arg2;
630 		if (doit) {
631 			/* set the selector to 0 to not confuse __switch_to */
632 			loadsegment(fs, 0);
633 			ret = wrmsrl_safe(MSR_FS_BASE, arg2);
634 		}
635 		put_cpu();
636 		break;
637 	case ARCH_GET_FS: {
638 		unsigned long base;
639 
640 		if (doit)
641 			rdmsrl(MSR_FS_BASE, base);
642 		else
643 			base = task->thread.fsbase;
644 		ret = put_user(base, (unsigned long __user *)arg2);
645 		break;
646 	}
647 	case ARCH_GET_GS: {
648 		unsigned long base;
649 
650 		if (doit)
651 			rdmsrl(MSR_KERNEL_GS_BASE, base);
652 		else
653 			base = task->thread.gsbase;
654 		ret = put_user(base, (unsigned long __user *)arg2);
655 		break;
656 	}
657 
658 #ifdef CONFIG_CHECKPOINT_RESTORE
659 # ifdef CONFIG_X86_X32_ABI
660 	case ARCH_MAP_VDSO_X32:
661 		return prctl_map_vdso(&vdso_image_x32, arg2);
662 # endif
663 # if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
664 	case ARCH_MAP_VDSO_32:
665 		return prctl_map_vdso(&vdso_image_32, arg2);
666 # endif
667 	case ARCH_MAP_VDSO_64:
668 		return prctl_map_vdso(&vdso_image_64, arg2);
669 #endif
670 
671 	default:
672 		ret = -EINVAL;
673 		break;
674 	}
675 
676 	return ret;
677 }
678 
679 SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
680 {
681 	long ret;
682 
683 	ret = do_arch_prctl_64(current, option, arg2);
684 	if (ret == -EINVAL)
685 		ret = do_arch_prctl_common(current, option, arg2);
686 
687 	return ret;
688 }
689 
690 #ifdef CONFIG_IA32_EMULATION
691 COMPAT_SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
692 {
693 	return do_arch_prctl_common(current, option, arg2);
694 }
695 #endif
696 
697 unsigned long KSTK_ESP(struct task_struct *task)
698 {
699 	return task_pt_regs(task)->sp;
700 }
701