xref: /linux/arch/x86/kernel/process_64.c (revision 5e3992fe72748ed3892be876f09d4d990548b7af)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  Copyright (C) 1995  Linus Torvalds
4  *
5  *  Pentium III FXSR, SSE support
6  *	Gareth Hughes <gareth@valinux.com>, May 2000
7  *
8  *  X86-64 port
9  *	Andi Kleen.
10  *
11  *	CPU hotplug support - ashok.raj@intel.com
12  */
13 
14 /*
15  * This file handles the architecture-dependent parts of process handling..
16  */
17 
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>
23 #include <linux/fs.h>
24 #include <linux/kernel.h>
25 #include <linux/mm.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>
39 #include <linux/io.h>
40 #include <linux/ftrace.h>
41 #include <linux/syscalls.h>
42 #include <linux/iommu.h>
43 
44 #include <asm/processor.h>
45 #include <asm/pkru.h>
46 #include <asm/fpu/sched.h>
47 #include <asm/mmu_context.h>
48 #include <asm/prctl.h>
49 #include <asm/desc.h>
50 #include <asm/proto.h>
51 #include <asm/ia32.h>
52 #include <asm/debugreg.h>
53 #include <asm/switch_to.h>
54 #include <asm/xen/hypervisor.h>
55 #include <asm/vdso.h>
56 #include <asm/resctrl.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>
62 #endif
63 
64 #include "process.h"
65 
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,
68 		 const char *log_lvl)
69 {
70 	unsigned long cr0 = 0L, cr2 = 0L, cr3 = 0L, cr4 = 0L, fs, gs, shadowgs;
71 	unsigned long d0, d1, d2, d3, d6, d7;
72 	unsigned int fsindex, gsindex;
73 	unsigned int ds, es;
74 
75 	show_iret_regs(regs, log_lvl);
76 
77 	if (regs->orig_ax != -1)
78 		pr_cont(" ORIG_RAX: %016lx\n", regs->orig_ax);
79 	else
80 		pr_cont("\n");
81 
82 	printk("%sRAX: %016lx RBX: %016lx RCX: %016lx\n",
83 	       log_lvl, regs->ax, regs->bx, regs->cx);
84 	printk("%sRDX: %016lx RSI: %016lx RDI: %016lx\n",
85 	       log_lvl, regs->dx, regs->si, regs->di);
86 	printk("%sRBP: %016lx R08: %016lx R09: %016lx\n",
87 	       log_lvl, regs->bp, regs->r8, regs->r9);
88 	printk("%sR10: %016lx R11: %016lx R12: %016lx\n",
89 	       log_lvl, regs->r10, regs->r11, regs->r12);
90 	printk("%sR13: %016lx R14: %016lx R15: %016lx\n",
91 	       log_lvl, regs->r13, regs->r14, regs->r15);
92 
93 	if (mode == SHOW_REGS_SHORT)
94 		return;
95 
96 	if (mode == SHOW_REGS_USER) {
97 		rdmsrl(MSR_FS_BASE, fs);
98 		rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
99 		printk("%sFS:  %016lx GS:  %016lx\n",
100 		       log_lvl, fs, shadowgs);
101 		return;
102 	}
103 
104 	asm("movl %%ds,%0" : "=r" (ds));
105 	asm("movl %%es,%0" : "=r" (es));
106 	asm("movl %%fs,%0" : "=r" (fsindex));
107 	asm("movl %%gs,%0" : "=r" (gsindex));
108 
109 	rdmsrl(MSR_FS_BASE, fs);
110 	rdmsrl(MSR_GS_BASE, gs);
111 	rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
112 
113 	cr0 = read_cr0();
114 	cr2 = read_cr2();
115 	cr3 = __read_cr3();
116 	cr4 = __read_cr4();
117 
118 	printk("%sFS:  %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
119 	       log_lvl, fs, fsindex, gs, gsindex, shadowgs);
120 	printk("%sCS:  %04lx DS: %04x ES: %04x CR0: %016lx\n",
121 		log_lvl, regs->cs, ds, es, cr0);
122 	printk("%sCR2: %016lx CR3: %016lx CR4: %016lx\n",
123 		log_lvl, cr2, cr3, cr4);
124 
125 	get_debugreg(d0, 0);
126 	get_debugreg(d1, 1);
127 	get_debugreg(d2, 2);
128 	get_debugreg(d3, 3);
129 	get_debugreg(d6, 6);
130 	get_debugreg(d7, 7);
131 
132 	/* Only print out debug registers if they are in their non-default state. */
133 	if (!((d0 == 0) && (d1 == 0) && (d2 == 0) && (d3 == 0) &&
134 	    (d6 == DR6_RESERVED) && (d7 == 0x400))) {
135 		printk("%sDR0: %016lx DR1: %016lx DR2: %016lx\n",
136 		       log_lvl, d0, d1, d2);
137 		printk("%sDR3: %016lx DR6: %016lx DR7: %016lx\n",
138 		       log_lvl, d3, d6, d7);
139 	}
140 
141 	if (cpu_feature_enabled(X86_FEATURE_OSPKE))
142 		printk("%sPKRU: %08x\n", log_lvl, read_pkru());
143 }
144 
145 void release_thread(struct task_struct *dead_task)
146 {
147 	WARN_ON(dead_task->mm);
148 }
149 
150 enum which_selector {
151 	FS,
152 	GS
153 };
154 
155 /*
156  * Out of line to be protected from kprobes and tracing. If this would be
157  * traced or probed than any access to a per CPU variable happens with
158  * the wrong GS.
159  *
160  * It is not used on Xen paravirt. When paravirt support is needed, it
161  * needs to be renamed with native_ prefix.
162  */
163 static noinstr unsigned long __rdgsbase_inactive(void)
164 {
165 	unsigned long gsbase;
166 
167 	lockdep_assert_irqs_disabled();
168 
169 	if (!cpu_feature_enabled(X86_FEATURE_XENPV)) {
170 		native_swapgs();
171 		gsbase = rdgsbase();
172 		native_swapgs();
173 	} else {
174 		instrumentation_begin();
175 		rdmsrl(MSR_KERNEL_GS_BASE, gsbase);
176 		instrumentation_end();
177 	}
178 
179 	return gsbase;
180 }
181 
182 /*
183  * Out of line to be protected from kprobes and tracing. If this would be
184  * traced or probed than any access to a per CPU variable happens with
185  * the wrong GS.
186  *
187  * It is not used on Xen paravirt. When paravirt support is needed, it
188  * needs to be renamed with native_ prefix.
189  */
190 static noinstr void __wrgsbase_inactive(unsigned long gsbase)
191 {
192 	lockdep_assert_irqs_disabled();
193 
194 	if (!cpu_feature_enabled(X86_FEATURE_XENPV)) {
195 		native_swapgs();
196 		wrgsbase(gsbase);
197 		native_swapgs();
198 	} else {
199 		instrumentation_begin();
200 		wrmsrl(MSR_KERNEL_GS_BASE, gsbase);
201 		instrumentation_end();
202 	}
203 }
204 
205 /*
206  * Saves the FS or GS base for an outgoing thread if FSGSBASE extensions are
207  * not available.  The goal is to be reasonably fast on non-FSGSBASE systems.
208  * It's forcibly inlined because it'll generate better code and this function
209  * is hot.
210  */
211 static __always_inline void save_base_legacy(struct task_struct *prev_p,
212 					     unsigned short selector,
213 					     enum which_selector which)
214 {
215 	if (likely(selector == 0)) {
216 		/*
217 		 * On Intel (without X86_BUG_NULL_SEG), the segment base could
218 		 * be the pre-existing saved base or it could be zero.  On AMD
219 		 * (with X86_BUG_NULL_SEG), the segment base could be almost
220 		 * anything.
221 		 *
222 		 * This branch is very hot (it's hit twice on almost every
223 		 * context switch between 64-bit programs), and avoiding
224 		 * the RDMSR helps a lot, so we just assume that whatever
225 		 * value is already saved is correct.  This matches historical
226 		 * Linux behavior, so it won't break existing applications.
227 		 *
228 		 * To avoid leaking state, on non-X86_BUG_NULL_SEG CPUs, if we
229 		 * report that the base is zero, it needs to actually be zero:
230 		 * see the corresponding logic in load_seg_legacy.
231 		 */
232 	} else {
233 		/*
234 		 * If the selector is 1, 2, or 3, then the base is zero on
235 		 * !X86_BUG_NULL_SEG CPUs and could be anything on
236 		 * X86_BUG_NULL_SEG CPUs.  In the latter case, Linux
237 		 * has never attempted to preserve the base across context
238 		 * switches.
239 		 *
240 		 * If selector > 3, then it refers to a real segment, and
241 		 * saving the base isn't necessary.
242 		 */
243 		if (which == FS)
244 			prev_p->thread.fsbase = 0;
245 		else
246 			prev_p->thread.gsbase = 0;
247 	}
248 }
249 
250 static __always_inline void save_fsgs(struct task_struct *task)
251 {
252 	savesegment(fs, task->thread.fsindex);
253 	savesegment(gs, task->thread.gsindex);
254 	if (static_cpu_has(X86_FEATURE_FSGSBASE)) {
255 		/*
256 		 * If FSGSBASE is enabled, we can't make any useful guesses
257 		 * about the base, and user code expects us to save the current
258 		 * value.  Fortunately, reading the base directly is efficient.
259 		 */
260 		task->thread.fsbase = rdfsbase();
261 		task->thread.gsbase = __rdgsbase_inactive();
262 	} else {
263 		save_base_legacy(task, task->thread.fsindex, FS);
264 		save_base_legacy(task, task->thread.gsindex, GS);
265 	}
266 }
267 
268 /*
269  * While a process is running,current->thread.fsbase and current->thread.gsbase
270  * may not match the corresponding CPU registers (see save_base_legacy()).
271  */
272 void current_save_fsgs(void)
273 {
274 	unsigned long flags;
275 
276 	/* Interrupts need to be off for FSGSBASE */
277 	local_irq_save(flags);
278 	save_fsgs(current);
279 	local_irq_restore(flags);
280 }
281 #if IS_ENABLED(CONFIG_KVM)
282 EXPORT_SYMBOL_GPL(current_save_fsgs);
283 #endif
284 
285 static __always_inline void loadseg(enum which_selector which,
286 				    unsigned short sel)
287 {
288 	if (which == FS)
289 		loadsegment(fs, sel);
290 	else
291 		load_gs_index(sel);
292 }
293 
294 static __always_inline void load_seg_legacy(unsigned short prev_index,
295 					    unsigned long prev_base,
296 					    unsigned short next_index,
297 					    unsigned long next_base,
298 					    enum which_selector which)
299 {
300 	if (likely(next_index <= 3)) {
301 		/*
302 		 * The next task is using 64-bit TLS, is not using this
303 		 * segment at all, or is having fun with arcane CPU features.
304 		 */
305 		if (next_base == 0) {
306 			/*
307 			 * Nasty case: on AMD CPUs, we need to forcibly zero
308 			 * the base.
309 			 */
310 			if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
311 				loadseg(which, __USER_DS);
312 				loadseg(which, next_index);
313 			} else {
314 				/*
315 				 * We could try to exhaustively detect cases
316 				 * under which we can skip the segment load,
317 				 * but there's really only one case that matters
318 				 * for performance: if both the previous and
319 				 * next states are fully zeroed, we can skip
320 				 * the load.
321 				 *
322 				 * (This assumes that prev_base == 0 has no
323 				 * false positives.  This is the case on
324 				 * Intel-style CPUs.)
325 				 */
326 				if (likely(prev_index | next_index | prev_base))
327 					loadseg(which, next_index);
328 			}
329 		} else {
330 			if (prev_index != next_index)
331 				loadseg(which, next_index);
332 			wrmsrl(which == FS ? MSR_FS_BASE : MSR_KERNEL_GS_BASE,
333 			       next_base);
334 		}
335 	} else {
336 		/*
337 		 * The next task is using a real segment.  Loading the selector
338 		 * is sufficient.
339 		 */
340 		loadseg(which, next_index);
341 	}
342 }
343 
344 /*
345  * Store prev's PKRU value and load next's PKRU value if they differ. PKRU
346  * is not XSTATE managed on context switch because that would require a
347  * lookup in the task's FPU xsave buffer and require to keep that updated
348  * in various places.
349  */
350 static __always_inline void x86_pkru_load(struct thread_struct *prev,
351 					  struct thread_struct *next)
352 {
353 	if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
354 		return;
355 
356 	/* Stash the prev task's value: */
357 	prev->pkru = rdpkru();
358 
359 	/*
360 	 * PKRU writes are slightly expensive.  Avoid them when not
361 	 * strictly necessary:
362 	 */
363 	if (prev->pkru != next->pkru)
364 		wrpkru(next->pkru);
365 }
366 
367 static __always_inline void x86_fsgsbase_load(struct thread_struct *prev,
368 					      struct thread_struct *next)
369 {
370 	if (static_cpu_has(X86_FEATURE_FSGSBASE)) {
371 		/* Update the FS and GS selectors if they could have changed. */
372 		if (unlikely(prev->fsindex || next->fsindex))
373 			loadseg(FS, next->fsindex);
374 		if (unlikely(prev->gsindex || next->gsindex))
375 			loadseg(GS, next->gsindex);
376 
377 		/* Update the bases. */
378 		wrfsbase(next->fsbase);
379 		__wrgsbase_inactive(next->gsbase);
380 	} else {
381 		load_seg_legacy(prev->fsindex, prev->fsbase,
382 				next->fsindex, next->fsbase, FS);
383 		load_seg_legacy(prev->gsindex, prev->gsbase,
384 				next->gsindex, next->gsbase, GS);
385 	}
386 }
387 
388 unsigned long x86_fsgsbase_read_task(struct task_struct *task,
389 				     unsigned short selector)
390 {
391 	unsigned short idx = selector >> 3;
392 	unsigned long base;
393 
394 	if (likely((selector & SEGMENT_TI_MASK) == 0)) {
395 		if (unlikely(idx >= GDT_ENTRIES))
396 			return 0;
397 
398 		/*
399 		 * There are no user segments in the GDT with nonzero bases
400 		 * other than the TLS segments.
401 		 */
402 		if (idx < GDT_ENTRY_TLS_MIN || idx > GDT_ENTRY_TLS_MAX)
403 			return 0;
404 
405 		idx -= GDT_ENTRY_TLS_MIN;
406 		base = get_desc_base(&task->thread.tls_array[idx]);
407 	} else {
408 #ifdef CONFIG_MODIFY_LDT_SYSCALL
409 		struct ldt_struct *ldt;
410 
411 		/*
412 		 * If performance here mattered, we could protect the LDT
413 		 * with RCU.  This is a slow path, though, so we can just
414 		 * take the mutex.
415 		 */
416 		mutex_lock(&task->mm->context.lock);
417 		ldt = task->mm->context.ldt;
418 		if (unlikely(!ldt || idx >= ldt->nr_entries))
419 			base = 0;
420 		else
421 			base = get_desc_base(ldt->entries + idx);
422 		mutex_unlock(&task->mm->context.lock);
423 #else
424 		base = 0;
425 #endif
426 	}
427 
428 	return base;
429 }
430 
431 unsigned long x86_gsbase_read_cpu_inactive(void)
432 {
433 	unsigned long gsbase;
434 
435 	if (boot_cpu_has(X86_FEATURE_FSGSBASE)) {
436 		unsigned long flags;
437 
438 		local_irq_save(flags);
439 		gsbase = __rdgsbase_inactive();
440 		local_irq_restore(flags);
441 	} else {
442 		rdmsrl(MSR_KERNEL_GS_BASE, gsbase);
443 	}
444 
445 	return gsbase;
446 }
447 
448 void x86_gsbase_write_cpu_inactive(unsigned long gsbase)
449 {
450 	if (boot_cpu_has(X86_FEATURE_FSGSBASE)) {
451 		unsigned long flags;
452 
453 		local_irq_save(flags);
454 		__wrgsbase_inactive(gsbase);
455 		local_irq_restore(flags);
456 	} else {
457 		wrmsrl(MSR_KERNEL_GS_BASE, gsbase);
458 	}
459 }
460 
461 unsigned long x86_fsbase_read_task(struct task_struct *task)
462 {
463 	unsigned long fsbase;
464 
465 	if (task == current)
466 		fsbase = x86_fsbase_read_cpu();
467 	else if (boot_cpu_has(X86_FEATURE_FSGSBASE) ||
468 		 (task->thread.fsindex == 0))
469 		fsbase = task->thread.fsbase;
470 	else
471 		fsbase = x86_fsgsbase_read_task(task, task->thread.fsindex);
472 
473 	return fsbase;
474 }
475 
476 unsigned long x86_gsbase_read_task(struct task_struct *task)
477 {
478 	unsigned long gsbase;
479 
480 	if (task == current)
481 		gsbase = x86_gsbase_read_cpu_inactive();
482 	else if (boot_cpu_has(X86_FEATURE_FSGSBASE) ||
483 		 (task->thread.gsindex == 0))
484 		gsbase = task->thread.gsbase;
485 	else
486 		gsbase = x86_fsgsbase_read_task(task, task->thread.gsindex);
487 
488 	return gsbase;
489 }
490 
491 void x86_fsbase_write_task(struct task_struct *task, unsigned long fsbase)
492 {
493 	WARN_ON_ONCE(task == current);
494 
495 	task->thread.fsbase = fsbase;
496 }
497 
498 void x86_gsbase_write_task(struct task_struct *task, unsigned long gsbase)
499 {
500 	WARN_ON_ONCE(task == current);
501 
502 	task->thread.gsbase = gsbase;
503 }
504 
505 static void
506 start_thread_common(struct pt_regs *regs, unsigned long new_ip,
507 		    unsigned long new_sp,
508 		    unsigned int _cs, unsigned int _ss, unsigned int _ds)
509 {
510 	WARN_ON_ONCE(regs != current_pt_regs());
511 
512 	if (static_cpu_has(X86_BUG_NULL_SEG)) {
513 		/* Loading zero below won't clear the base. */
514 		loadsegment(fs, __USER_DS);
515 		load_gs_index(__USER_DS);
516 	}
517 
518 	loadsegment(fs, 0);
519 	loadsegment(es, _ds);
520 	loadsegment(ds, _ds);
521 	load_gs_index(0);
522 
523 	regs->ip		= new_ip;
524 	regs->sp		= new_sp;
525 	regs->cs		= _cs;
526 	regs->ss		= _ss;
527 	regs->flags		= X86_EFLAGS_IF;
528 }
529 
530 void
531 start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
532 {
533 	start_thread_common(regs, new_ip, new_sp,
534 			    __USER_CS, __USER_DS, 0);
535 }
536 EXPORT_SYMBOL_GPL(start_thread);
537 
538 #ifdef CONFIG_COMPAT
539 void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp, bool x32)
540 {
541 	start_thread_common(regs, new_ip, new_sp,
542 			    x32 ? __USER_CS : __USER32_CS,
543 			    __USER_DS, __USER_DS);
544 }
545 #endif
546 
547 /*
548  *	switch_to(x,y) should switch tasks from x to y.
549  *
550  * This could still be optimized:
551  * - fold all the options into a flag word and test it with a single test.
552  * - could test fs/gs bitsliced
553  *
554  * Kprobes not supported here. Set the probe on schedule instead.
555  * Function graph tracer not supported too.
556  */
557 __no_kmsan_checks
558 __visible __notrace_funcgraph struct task_struct *
559 __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
560 {
561 	struct thread_struct *prev = &prev_p->thread;
562 	struct thread_struct *next = &next_p->thread;
563 	struct fpu *prev_fpu = &prev->fpu;
564 	int cpu = smp_processor_id();
565 
566 	WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_ENTRY) &&
567 		     this_cpu_read(pcpu_hot.hardirq_stack_inuse));
568 
569 	if (!test_thread_flag(TIF_NEED_FPU_LOAD))
570 		switch_fpu_prepare(prev_fpu, cpu);
571 
572 	/* We must save %fs and %gs before load_TLS() because
573 	 * %fs and %gs may be cleared by load_TLS().
574 	 *
575 	 * (e.g. xen_load_tls())
576 	 */
577 	save_fsgs(prev_p);
578 
579 	/*
580 	 * Load TLS before restoring any segments so that segment loads
581 	 * reference the correct GDT entries.
582 	 */
583 	load_TLS(next, cpu);
584 
585 	/*
586 	 * Leave lazy mode, flushing any hypercalls made here.  This
587 	 * must be done after loading TLS entries in the GDT but before
588 	 * loading segments that might reference them.
589 	 */
590 	arch_end_context_switch(next_p);
591 
592 	/* Switch DS and ES.
593 	 *
594 	 * Reading them only returns the selectors, but writing them (if
595 	 * nonzero) loads the full descriptor from the GDT or LDT.  The
596 	 * LDT for next is loaded in switch_mm, and the GDT is loaded
597 	 * above.
598 	 *
599 	 * We therefore need to write new values to the segment
600 	 * registers on every context switch unless both the new and old
601 	 * values are zero.
602 	 *
603 	 * Note that we don't need to do anything for CS and SS, as
604 	 * those are saved and restored as part of pt_regs.
605 	 */
606 	savesegment(es, prev->es);
607 	if (unlikely(next->es | prev->es))
608 		loadsegment(es, next->es);
609 
610 	savesegment(ds, prev->ds);
611 	if (unlikely(next->ds | prev->ds))
612 		loadsegment(ds, next->ds);
613 
614 	x86_fsgsbase_load(prev, next);
615 
616 	x86_pkru_load(prev, next);
617 
618 	/*
619 	 * Switch the PDA and FPU contexts.
620 	 */
621 	raw_cpu_write(pcpu_hot.current_task, next_p);
622 	raw_cpu_write(pcpu_hot.top_of_stack, task_top_of_stack(next_p));
623 
624 	switch_fpu_finish();
625 
626 	/* Reload sp0. */
627 	update_task_stack(next_p);
628 
629 	switch_to_extra(prev_p, next_p);
630 
631 	if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
632 		/*
633 		 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
634 		 * does not update the cached descriptor.  As a result, if we
635 		 * do SYSRET while SS is NULL, we'll end up in user mode with
636 		 * SS apparently equal to __USER_DS but actually unusable.
637 		 *
638 		 * The straightforward workaround would be to fix it up just
639 		 * before SYSRET, but that would slow down the system call
640 		 * fast paths.  Instead, we ensure that SS is never NULL in
641 		 * system call context.  We do this by replacing NULL SS
642 		 * selectors at every context switch.  SYSCALL sets up a valid
643 		 * SS, so the only way to get NULL is to re-enter the kernel
644 		 * from CPL 3 through an interrupt.  Since that can't happen
645 		 * in the same task as a running syscall, we are guaranteed to
646 		 * context switch between every interrupt vector entry and a
647 		 * subsequent SYSRET.
648 		 *
649 		 * We read SS first because SS reads are much faster than
650 		 * writes.  Out of caution, we force SS to __KERNEL_DS even if
651 		 * it previously had a different non-NULL value.
652 		 */
653 		unsigned short ss_sel;
654 		savesegment(ss, ss_sel);
655 		if (ss_sel != __KERNEL_DS)
656 			loadsegment(ss, __KERNEL_DS);
657 	}
658 
659 	/* Load the Intel cache allocation PQR MSR. */
660 	resctrl_sched_in(next_p);
661 
662 	return prev_p;
663 }
664 
665 void set_personality_64bit(void)
666 {
667 	/* inherit personality from parent */
668 
669 	/* Make sure to be in 64bit mode */
670 	clear_thread_flag(TIF_ADDR32);
671 	/* Pretend that this comes from a 64bit execve */
672 	task_pt_regs(current)->orig_ax = __NR_execve;
673 	current_thread_info()->status &= ~TS_COMPAT;
674 	if (current->mm)
675 		__set_bit(MM_CONTEXT_HAS_VSYSCALL, &current->mm->context.flags);
676 
677 	/* TBD: overwrites user setup. Should have two bits.
678 	   But 64bit processes have always behaved this way,
679 	   so it's not too bad. The main problem is just that
680 	   32bit children are affected again. */
681 	current->personality &= ~READ_IMPLIES_EXEC;
682 }
683 
684 static void __set_personality_x32(void)
685 {
686 #ifdef CONFIG_X86_X32_ABI
687 	if (current->mm)
688 		current->mm->context.flags = 0;
689 
690 	current->personality &= ~READ_IMPLIES_EXEC;
691 	/*
692 	 * in_32bit_syscall() uses the presence of the x32 syscall bit
693 	 * flag to determine compat status.  The x86 mmap() code relies on
694 	 * the syscall bitness so set x32 syscall bit right here to make
695 	 * in_32bit_syscall() work during exec().
696 	 *
697 	 * Pretend to come from a x32 execve.
698 	 */
699 	task_pt_regs(current)->orig_ax = __NR_x32_execve | __X32_SYSCALL_BIT;
700 	current_thread_info()->status &= ~TS_COMPAT;
701 #endif
702 }
703 
704 static void __set_personality_ia32(void)
705 {
706 #ifdef CONFIG_IA32_EMULATION
707 	if (current->mm) {
708 		/*
709 		 * uprobes applied to this MM need to know this and
710 		 * cannot use user_64bit_mode() at that time.
711 		 */
712 		__set_bit(MM_CONTEXT_UPROBE_IA32, &current->mm->context.flags);
713 	}
714 
715 	current->personality |= force_personality32;
716 	/* Prepare the first "return" to user space */
717 	task_pt_regs(current)->orig_ax = __NR_ia32_execve;
718 	current_thread_info()->status |= TS_COMPAT;
719 #endif
720 }
721 
722 void set_personality_ia32(bool x32)
723 {
724 	/* Make sure to be in 32bit mode */
725 	set_thread_flag(TIF_ADDR32);
726 
727 	if (x32)
728 		__set_personality_x32();
729 	else
730 		__set_personality_ia32();
731 }
732 EXPORT_SYMBOL_GPL(set_personality_ia32);
733 
734 #ifdef CONFIG_CHECKPOINT_RESTORE
735 static long prctl_map_vdso(const struct vdso_image *image, unsigned long addr)
736 {
737 	int ret;
738 
739 	ret = map_vdso_once(image, addr);
740 	if (ret)
741 		return ret;
742 
743 	return (long)image->size;
744 }
745 #endif
746 
747 #ifdef CONFIG_ADDRESS_MASKING
748 
749 #define LAM_U57_BITS 6
750 
751 static int prctl_enable_tagged_addr(struct mm_struct *mm, unsigned long nr_bits)
752 {
753 	if (!cpu_feature_enabled(X86_FEATURE_LAM))
754 		return -ENODEV;
755 
756 	/* PTRACE_ARCH_PRCTL */
757 	if (current->mm != mm)
758 		return -EINVAL;
759 
760 	if (mm_valid_pasid(mm) &&
761 	    !test_bit(MM_CONTEXT_FORCE_TAGGED_SVA, &mm->context.flags))
762 		return -EINVAL;
763 
764 	if (mmap_write_lock_killable(mm))
765 		return -EINTR;
766 
767 	if (test_bit(MM_CONTEXT_LOCK_LAM, &mm->context.flags)) {
768 		mmap_write_unlock(mm);
769 		return -EBUSY;
770 	}
771 
772 	if (!nr_bits) {
773 		mmap_write_unlock(mm);
774 		return -EINVAL;
775 	} else if (nr_bits <= LAM_U57_BITS) {
776 		mm->context.lam_cr3_mask = X86_CR3_LAM_U57;
777 		mm->context.untag_mask =  ~GENMASK(62, 57);
778 	} else {
779 		mmap_write_unlock(mm);
780 		return -EINVAL;
781 	}
782 
783 	write_cr3(__read_cr3() | mm->context.lam_cr3_mask);
784 	set_tlbstate_lam_mode(mm);
785 	set_bit(MM_CONTEXT_LOCK_LAM, &mm->context.flags);
786 
787 	mmap_write_unlock(mm);
788 
789 	return 0;
790 }
791 #endif
792 
793 long do_arch_prctl_64(struct task_struct *task, int option, unsigned long arg2)
794 {
795 	int ret = 0;
796 
797 	switch (option) {
798 	case ARCH_SET_GS: {
799 		if (unlikely(arg2 >= TASK_SIZE_MAX))
800 			return -EPERM;
801 
802 		preempt_disable();
803 		/*
804 		 * ARCH_SET_GS has always overwritten the index
805 		 * and the base. Zero is the most sensible value
806 		 * to put in the index, and is the only value that
807 		 * makes any sense if FSGSBASE is unavailable.
808 		 */
809 		if (task == current) {
810 			loadseg(GS, 0);
811 			x86_gsbase_write_cpu_inactive(arg2);
812 
813 			/*
814 			 * On non-FSGSBASE systems, save_base_legacy() expects
815 			 * that we also fill in thread.gsbase.
816 			 */
817 			task->thread.gsbase = arg2;
818 
819 		} else {
820 			task->thread.gsindex = 0;
821 			x86_gsbase_write_task(task, arg2);
822 		}
823 		preempt_enable();
824 		break;
825 	}
826 	case ARCH_SET_FS: {
827 		/*
828 		 * Not strictly needed for %fs, but do it for symmetry
829 		 * with %gs
830 		 */
831 		if (unlikely(arg2 >= TASK_SIZE_MAX))
832 			return -EPERM;
833 
834 		preempt_disable();
835 		/*
836 		 * Set the selector to 0 for the same reason
837 		 * as %gs above.
838 		 */
839 		if (task == current) {
840 			loadseg(FS, 0);
841 			x86_fsbase_write_cpu(arg2);
842 
843 			/*
844 			 * On non-FSGSBASE systems, save_base_legacy() expects
845 			 * that we also fill in thread.fsbase.
846 			 */
847 			task->thread.fsbase = arg2;
848 		} else {
849 			task->thread.fsindex = 0;
850 			x86_fsbase_write_task(task, arg2);
851 		}
852 		preempt_enable();
853 		break;
854 	}
855 	case ARCH_GET_FS: {
856 		unsigned long base = x86_fsbase_read_task(task);
857 
858 		ret = put_user(base, (unsigned long __user *)arg2);
859 		break;
860 	}
861 	case ARCH_GET_GS: {
862 		unsigned long base = x86_gsbase_read_task(task);
863 
864 		ret = put_user(base, (unsigned long __user *)arg2);
865 		break;
866 	}
867 
868 #ifdef CONFIG_CHECKPOINT_RESTORE
869 # ifdef CONFIG_X86_X32_ABI
870 	case ARCH_MAP_VDSO_X32:
871 		return prctl_map_vdso(&vdso_image_x32, arg2);
872 # endif
873 # if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
874 	case ARCH_MAP_VDSO_32:
875 		return prctl_map_vdso(&vdso_image_32, arg2);
876 # endif
877 	case ARCH_MAP_VDSO_64:
878 		return prctl_map_vdso(&vdso_image_64, arg2);
879 #endif
880 #ifdef CONFIG_ADDRESS_MASKING
881 	case ARCH_GET_UNTAG_MASK:
882 		return put_user(task->mm->context.untag_mask,
883 				(unsigned long __user *)arg2);
884 	case ARCH_ENABLE_TAGGED_ADDR:
885 		return prctl_enable_tagged_addr(task->mm, arg2);
886 	case ARCH_FORCE_TAGGED_SVA:
887 		if (current != task)
888 			return -EINVAL;
889 		set_bit(MM_CONTEXT_FORCE_TAGGED_SVA, &task->mm->context.flags);
890 		return 0;
891 	case ARCH_GET_MAX_TAG_BITS:
892 		if (!cpu_feature_enabled(X86_FEATURE_LAM))
893 			return put_user(0, (unsigned long __user *)arg2);
894 		else
895 			return put_user(LAM_U57_BITS, (unsigned long __user *)arg2);
896 #endif
897 	default:
898 		ret = -EINVAL;
899 		break;
900 	}
901 
902 	return ret;
903 }
904 
905 SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
906 {
907 	long ret;
908 
909 	ret = do_arch_prctl_64(current, option, arg2);
910 	if (ret == -EINVAL)
911 		ret = do_arch_prctl_common(option, arg2);
912 
913 	return ret;
914 }
915 
916 #ifdef CONFIG_IA32_EMULATION
917 COMPAT_SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
918 {
919 	return do_arch_prctl_common(option, arg2);
920 }
921 #endif
922 
923 unsigned long KSTK_ESP(struct task_struct *task)
924 {
925 	return task_pt_regs(task)->sp;
926 }
927