Loading...
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
43#include <asm/processor.h>
44#include <asm/pkru.h>
45#include <asm/fpu/sched.h>
46#include <asm/mmu_context.h>
47#include <asm/prctl.h>
48#include <asm/desc.h>
49#include <asm/proto.h>
50#include <asm/ia32.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/resctrl.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>
61#endif
62
63#include "process.h"
64
65/* Prints also some state that isn't saved in the pt_regs */
66void __show_regs(struct pt_regs *regs, enum show_regs_mode mode,
67 const char *log_lvl)
68{
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;
72 unsigned int ds, es;
73
74 show_iret_regs(regs, log_lvl);
75
76 if (regs->orig_ax != -1)
77 pr_cont(" ORIG_RAX: %016lx\n", regs->orig_ax);
78 else
79 pr_cont("\n");
80
81 printk("%sRAX: %016lx RBX: %016lx RCX: %016lx\n",
82 log_lvl, regs->ax, regs->bx, regs->cx);
83 printk("%sRDX: %016lx RSI: %016lx RDI: %016lx\n",
84 log_lvl, regs->dx, regs->si, regs->di);
85 printk("%sRBP: %016lx R08: %016lx R09: %016lx\n",
86 log_lvl, regs->bp, regs->r8, regs->r9);
87 printk("%sR10: %016lx R11: %016lx R12: %016lx\n",
88 log_lvl, regs->r10, regs->r11, regs->r12);
89 printk("%sR13: %016lx R14: %016lx R15: %016lx\n",
90 log_lvl, regs->r13, regs->r14, regs->r15);
91
92 if (mode == SHOW_REGS_SHORT)
93 return;
94
95 if (mode == SHOW_REGS_USER) {
96 rdmsrl(MSR_FS_BASE, fs);
97 rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
98 printk("%sFS: %016lx GS: %016lx\n",
99 log_lvl, fs, shadowgs);
100 return;
101 }
102
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));
107
108 rdmsrl(MSR_FS_BASE, fs);
109 rdmsrl(MSR_GS_BASE, gs);
110 rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
111
112 cr0 = read_cr0();
113 cr2 = read_cr2();
114 cr3 = __read_cr3();
115 cr4 = __read_cr4();
116
117 printk("%sFS: %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
118 log_lvl, fs, fsindex, gs, gsindex, shadowgs);
119 printk("%sCS: %04lx DS: %04x ES: %04x CR0: %016lx\n",
120 log_lvl, regs->cs, ds, es, cr0);
121 printk("%sCR2: %016lx CR3: %016lx CR4: %016lx\n",
122 log_lvl, cr2, cr3, cr4);
123
124 get_debugreg(d0, 0);
125 get_debugreg(d1, 1);
126 get_debugreg(d2, 2);
127 get_debugreg(d3, 3);
128 get_debugreg(d6, 6);
129 get_debugreg(d7, 7);
130
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("%sDR0: %016lx DR1: %016lx DR2: %016lx\n",
135 log_lvl, d0, d1, d2);
136 printk("%sDR3: %016lx DR6: %016lx DR7: %016lx\n",
137 log_lvl, d3, d6, d7);
138 }
139
140 if (cpu_feature_enabled(X86_FEATURE_OSPKE))
141 printk("%sPKRU: %08x\n", log_lvl, read_pkru());
142}
143
144void release_thread(struct task_struct *dead_task)
145{
146 WARN_ON(dead_task->mm);
147}
148
149enum which_selector {
150 FS,
151 GS
152};
153
154/*
155 * Out of line to be protected from kprobes and tracing. If this would be
156 * traced or probed than any access to a per CPU variable happens with
157 * the wrong GS.
158 *
159 * It is not used on Xen paravirt. When paravirt support is needed, it
160 * needs to be renamed with native_ prefix.
161 */
162static noinstr unsigned long __rdgsbase_inactive(void)
163{
164 unsigned long gsbase;
165
166 lockdep_assert_irqs_disabled();
167
168 if (!cpu_feature_enabled(X86_FEATURE_XENPV)) {
169 native_swapgs();
170 gsbase = rdgsbase();
171 native_swapgs();
172 } else {
173 instrumentation_begin();
174 rdmsrl(MSR_KERNEL_GS_BASE, gsbase);
175 instrumentation_end();
176 }
177
178 return gsbase;
179}
180
181/*
182 * Out of line to be protected from kprobes and tracing. If this would be
183 * traced or probed than any access to a per CPU variable happens with
184 * the wrong GS.
185 *
186 * It is not used on Xen paravirt. When paravirt support is needed, it
187 * needs to be renamed with native_ prefix.
188 */
189static noinstr void __wrgsbase_inactive(unsigned long gsbase)
190{
191 lockdep_assert_irqs_disabled();
192
193 if (!cpu_feature_enabled(X86_FEATURE_XENPV)) {
194 native_swapgs();
195 wrgsbase(gsbase);
196 native_swapgs();
197 } else {
198 instrumentation_begin();
199 wrmsrl(MSR_KERNEL_GS_BASE, gsbase);
200 instrumentation_end();
201 }
202}
203
204/*
205 * Saves the FS or GS base for an outgoing thread if FSGSBASE extensions are
206 * not available. The goal is to be reasonably fast on non-FSGSBASE systems.
207 * It's forcibly inlined because it'll generate better code and this function
208 * is hot.
209 */
210static __always_inline void save_base_legacy(struct task_struct *prev_p,
211 unsigned short selector,
212 enum which_selector which)
213{
214 if (likely(selector == 0)) {
215 /*
216 * On Intel (without X86_BUG_NULL_SEG), the segment base could
217 * be the pre-existing saved base or it could be zero. On AMD
218 * (with X86_BUG_NULL_SEG), the segment base could be almost
219 * anything.
220 *
221 * This branch is very hot (it's hit twice on almost every
222 * context switch between 64-bit programs), and avoiding
223 * the RDMSR helps a lot, so we just assume that whatever
224 * value is already saved is correct. This matches historical
225 * Linux behavior, so it won't break existing applications.
226 *
227 * To avoid leaking state, on non-X86_BUG_NULL_SEG CPUs, if we
228 * report that the base is zero, it needs to actually be zero:
229 * see the corresponding logic in load_seg_legacy.
230 */
231 } else {
232 /*
233 * If the selector is 1, 2, or 3, then the base is zero on
234 * !X86_BUG_NULL_SEG CPUs and could be anything on
235 * X86_BUG_NULL_SEG CPUs. In the latter case, Linux
236 * has never attempted to preserve the base across context
237 * switches.
238 *
239 * If selector > 3, then it refers to a real segment, and
240 * saving the base isn't necessary.
241 */
242 if (which == FS)
243 prev_p->thread.fsbase = 0;
244 else
245 prev_p->thread.gsbase = 0;
246 }
247}
248
249static __always_inline void save_fsgs(struct task_struct *task)
250{
251 savesegment(fs, task->thread.fsindex);
252 savesegment(gs, task->thread.gsindex);
253 if (static_cpu_has(X86_FEATURE_FSGSBASE)) {
254 /*
255 * If FSGSBASE is enabled, we can't make any useful guesses
256 * about the base, and user code expects us to save the current
257 * value. Fortunately, reading the base directly is efficient.
258 */
259 task->thread.fsbase = rdfsbase();
260 task->thread.gsbase = __rdgsbase_inactive();
261 } else {
262 save_base_legacy(task, task->thread.fsindex, FS);
263 save_base_legacy(task, task->thread.gsindex, GS);
264 }
265}
266
267/*
268 * While a process is running,current->thread.fsbase and current->thread.gsbase
269 * may not match the corresponding CPU registers (see save_base_legacy()).
270 */
271void current_save_fsgs(void)
272{
273 unsigned long flags;
274
275 /* Interrupts need to be off for FSGSBASE */
276 local_irq_save(flags);
277 save_fsgs(current);
278 local_irq_restore(flags);
279}
280#if IS_ENABLED(CONFIG_KVM)
281EXPORT_SYMBOL_GPL(current_save_fsgs);
282#endif
283
284static __always_inline void loadseg(enum which_selector which,
285 unsigned short sel)
286{
287 if (which == FS)
288 loadsegment(fs, sel);
289 else
290 load_gs_index(sel);
291}
292
293static __always_inline void load_seg_legacy(unsigned short prev_index,
294 unsigned long prev_base,
295 unsigned short next_index,
296 unsigned long next_base,
297 enum which_selector which)
298{
299 if (likely(next_index <= 3)) {
300 /*
301 * The next task is using 64-bit TLS, is not using this
302 * segment at all, or is having fun with arcane CPU features.
303 */
304 if (next_base == 0) {
305 /*
306 * Nasty case: on AMD CPUs, we need to forcibly zero
307 * the base.
308 */
309 if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
310 loadseg(which, __USER_DS);
311 loadseg(which, next_index);
312 } else {
313 /*
314 * We could try to exhaustively detect cases
315 * under which we can skip the segment load,
316 * but there's really only one case that matters
317 * for performance: if both the previous and
318 * next states are fully zeroed, we can skip
319 * the load.
320 *
321 * (This assumes that prev_base == 0 has no
322 * false positives. This is the case on
323 * Intel-style CPUs.)
324 */
325 if (likely(prev_index | next_index | prev_base))
326 loadseg(which, next_index);
327 }
328 } else {
329 if (prev_index != next_index)
330 loadseg(which, next_index);
331 wrmsrl(which == FS ? MSR_FS_BASE : MSR_KERNEL_GS_BASE,
332 next_base);
333 }
334 } else {
335 /*
336 * The next task is using a real segment. Loading the selector
337 * is sufficient.
338 */
339 loadseg(which, next_index);
340 }
341}
342
343/*
344 * Store prev's PKRU value and load next's PKRU value if they differ. PKRU
345 * is not XSTATE managed on context switch because that would require a
346 * lookup in the task's FPU xsave buffer and require to keep that updated
347 * in various places.
348 */
349static __always_inline void x86_pkru_load(struct thread_struct *prev,
350 struct thread_struct *next)
351{
352 if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
353 return;
354
355 /* Stash the prev task's value: */
356 prev->pkru = rdpkru();
357
358 /*
359 * PKRU writes are slightly expensive. Avoid them when not
360 * strictly necessary:
361 */
362 if (prev->pkru != next->pkru)
363 wrpkru(next->pkru);
364}
365
366static __always_inline void x86_fsgsbase_load(struct thread_struct *prev,
367 struct thread_struct *next)
368{
369 if (static_cpu_has(X86_FEATURE_FSGSBASE)) {
370 /* Update the FS and GS selectors if they could have changed. */
371 if (unlikely(prev->fsindex || next->fsindex))
372 loadseg(FS, next->fsindex);
373 if (unlikely(prev->gsindex || next->gsindex))
374 loadseg(GS, next->gsindex);
375
376 /* Update the bases. */
377 wrfsbase(next->fsbase);
378 __wrgsbase_inactive(next->gsbase);
379 } else {
380 load_seg_legacy(prev->fsindex, prev->fsbase,
381 next->fsindex, next->fsbase, FS);
382 load_seg_legacy(prev->gsindex, prev->gsbase,
383 next->gsindex, next->gsbase, GS);
384 }
385}
386
387unsigned long x86_fsgsbase_read_task(struct task_struct *task,
388 unsigned short selector)
389{
390 unsigned short idx = selector >> 3;
391 unsigned long base;
392
393 if (likely((selector & SEGMENT_TI_MASK) == 0)) {
394 if (unlikely(idx >= GDT_ENTRIES))
395 return 0;
396
397 /*
398 * There are no user segments in the GDT with nonzero bases
399 * other than the TLS segments.
400 */
401 if (idx < GDT_ENTRY_TLS_MIN || idx > GDT_ENTRY_TLS_MAX)
402 return 0;
403
404 idx -= GDT_ENTRY_TLS_MIN;
405 base = get_desc_base(&task->thread.tls_array[idx]);
406 } else {
407#ifdef CONFIG_MODIFY_LDT_SYSCALL
408 struct ldt_struct *ldt;
409
410 /*
411 * If performance here mattered, we could protect the LDT
412 * with RCU. This is a slow path, though, so we can just
413 * take the mutex.
414 */
415 mutex_lock(&task->mm->context.lock);
416 ldt = task->mm->context.ldt;
417 if (unlikely(!ldt || idx >= ldt->nr_entries))
418 base = 0;
419 else
420 base = get_desc_base(ldt->entries + idx);
421 mutex_unlock(&task->mm->context.lock);
422#else
423 base = 0;
424#endif
425 }
426
427 return base;
428}
429
430unsigned long x86_gsbase_read_cpu_inactive(void)
431{
432 unsigned long gsbase;
433
434 if (boot_cpu_has(X86_FEATURE_FSGSBASE)) {
435 unsigned long flags;
436
437 local_irq_save(flags);
438 gsbase = __rdgsbase_inactive();
439 local_irq_restore(flags);
440 } else {
441 rdmsrl(MSR_KERNEL_GS_BASE, gsbase);
442 }
443
444 return gsbase;
445}
446
447void x86_gsbase_write_cpu_inactive(unsigned long gsbase)
448{
449 if (boot_cpu_has(X86_FEATURE_FSGSBASE)) {
450 unsigned long flags;
451
452 local_irq_save(flags);
453 __wrgsbase_inactive(gsbase);
454 local_irq_restore(flags);
455 } else {
456 wrmsrl(MSR_KERNEL_GS_BASE, gsbase);
457 }
458}
459
460unsigned long x86_fsbase_read_task(struct task_struct *task)
461{
462 unsigned long fsbase;
463
464 if (task == current)
465 fsbase = x86_fsbase_read_cpu();
466 else if (boot_cpu_has(X86_FEATURE_FSGSBASE) ||
467 (task->thread.fsindex == 0))
468 fsbase = task->thread.fsbase;
469 else
470 fsbase = x86_fsgsbase_read_task(task, task->thread.fsindex);
471
472 return fsbase;
473}
474
475unsigned long x86_gsbase_read_task(struct task_struct *task)
476{
477 unsigned long gsbase;
478
479 if (task == current)
480 gsbase = x86_gsbase_read_cpu_inactive();
481 else if (boot_cpu_has(X86_FEATURE_FSGSBASE) ||
482 (task->thread.gsindex == 0))
483 gsbase = task->thread.gsbase;
484 else
485 gsbase = x86_fsgsbase_read_task(task, task->thread.gsindex);
486
487 return gsbase;
488}
489
490void x86_fsbase_write_task(struct task_struct *task, unsigned long fsbase)
491{
492 WARN_ON_ONCE(task == current);
493
494 task->thread.fsbase = fsbase;
495}
496
497void x86_gsbase_write_task(struct task_struct *task, unsigned long gsbase)
498{
499 WARN_ON_ONCE(task == current);
500
501 task->thread.gsbase = gsbase;
502}
503
504static void
505start_thread_common(struct pt_regs *regs, unsigned long new_ip,
506 unsigned long new_sp,
507 unsigned int _cs, unsigned int _ss, unsigned int _ds)
508{
509 WARN_ON_ONCE(regs != current_pt_regs());
510
511 if (static_cpu_has(X86_BUG_NULL_SEG)) {
512 /* Loading zero below won't clear the base. */
513 loadsegment(fs, __USER_DS);
514 load_gs_index(__USER_DS);
515 }
516
517 loadsegment(fs, 0);
518 loadsegment(es, _ds);
519 loadsegment(ds, _ds);
520 load_gs_index(0);
521
522 regs->ip = new_ip;
523 regs->sp = new_sp;
524 regs->cs = _cs;
525 regs->ss = _ss;
526 regs->flags = X86_EFLAGS_IF;
527}
528
529void
530start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
531{
532 start_thread_common(regs, new_ip, new_sp,
533 __USER_CS, __USER_DS, 0);
534}
535EXPORT_SYMBOL_GPL(start_thread);
536
537#ifdef CONFIG_COMPAT
538void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp, bool x32)
539{
540 start_thread_common(regs, new_ip, new_sp,
541 x32 ? __USER_CS : __USER32_CS,
542 __USER_DS, __USER_DS);
543}
544#endif
545
546/*
547 * switch_to(x,y) should switch tasks from x to y.
548 *
549 * This could still be optimized:
550 * - fold all the options into a flag word and test it with a single test.
551 * - could test fs/gs bitsliced
552 *
553 * Kprobes not supported here. Set the probe on schedule instead.
554 * Function graph tracer not supported too.
555 */
556__no_kmsan_checks
557__visible __notrace_funcgraph struct task_struct *
558__switch_to(struct task_struct *prev_p, struct task_struct *next_p)
559{
560 struct thread_struct *prev = &prev_p->thread;
561 struct thread_struct *next = &next_p->thread;
562 struct fpu *prev_fpu = &prev->fpu;
563 int cpu = smp_processor_id();
564
565 WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_ENTRY) &&
566 this_cpu_read(pcpu_hot.hardirq_stack_inuse));
567
568 if (!test_thread_flag(TIF_NEED_FPU_LOAD))
569 switch_fpu_prepare(prev_fpu, cpu);
570
571 /* We must save %fs and %gs before load_TLS() because
572 * %fs and %gs may be cleared by load_TLS().
573 *
574 * (e.g. xen_load_tls())
575 */
576 save_fsgs(prev_p);
577
578 /*
579 * Load TLS before restoring any segments so that segment loads
580 * reference the correct GDT entries.
581 */
582 load_TLS(next, cpu);
583
584 /*
585 * Leave lazy mode, flushing any hypercalls made here. This
586 * must be done after loading TLS entries in the GDT but before
587 * loading segments that might reference them.
588 */
589 arch_end_context_switch(next_p);
590
591 /* Switch DS and ES.
592 *
593 * Reading them only returns the selectors, but writing them (if
594 * nonzero) loads the full descriptor from the GDT or LDT. The
595 * LDT for next is loaded in switch_mm, and the GDT is loaded
596 * above.
597 *
598 * We therefore need to write new values to the segment
599 * registers on every context switch unless both the new and old
600 * values are zero.
601 *
602 * Note that we don't need to do anything for CS and SS, as
603 * those are saved and restored as part of pt_regs.
604 */
605 savesegment(es, prev->es);
606 if (unlikely(next->es | prev->es))
607 loadsegment(es, next->es);
608
609 savesegment(ds, prev->ds);
610 if (unlikely(next->ds | prev->ds))
611 loadsegment(ds, next->ds);
612
613 x86_fsgsbase_load(prev, next);
614
615 x86_pkru_load(prev, next);
616
617 /*
618 * Switch the PDA and FPU contexts.
619 */
620 raw_cpu_write(pcpu_hot.current_task, next_p);
621 raw_cpu_write(pcpu_hot.top_of_stack, task_top_of_stack(next_p));
622
623 switch_fpu_finish();
624
625 /* Reload sp0. */
626 update_task_stack(next_p);
627
628 switch_to_extra(prev_p, next_p);
629
630 if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
631 /*
632 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
633 * does not update the cached descriptor. As a result, if we
634 * do SYSRET while SS is NULL, we'll end up in user mode with
635 * SS apparently equal to __USER_DS but actually unusable.
636 *
637 * The straightforward workaround would be to fix it up just
638 * before SYSRET, but that would slow down the system call
639 * fast paths. Instead, we ensure that SS is never NULL in
640 * system call context. We do this by replacing NULL SS
641 * selectors at every context switch. SYSCALL sets up a valid
642 * SS, so the only way to get NULL is to re-enter the kernel
643 * from CPL 3 through an interrupt. Since that can't happen
644 * in the same task as a running syscall, we are guaranteed to
645 * context switch between every interrupt vector entry and a
646 * subsequent SYSRET.
647 *
648 * We read SS first because SS reads are much faster than
649 * writes. Out of caution, we force SS to __KERNEL_DS even if
650 * it previously had a different non-NULL value.
651 */
652 unsigned short ss_sel;
653 savesegment(ss, ss_sel);
654 if (ss_sel != __KERNEL_DS)
655 loadsegment(ss, __KERNEL_DS);
656 }
657
658 /* Load the Intel cache allocation PQR MSR. */
659 resctrl_sched_in();
660
661 return prev_p;
662}
663
664void set_personality_64bit(void)
665{
666 /* inherit personality from parent */
667
668 /* Make sure to be in 64bit mode */
669 clear_thread_flag(TIF_ADDR32);
670 /* Pretend that this comes from a 64bit execve */
671 task_pt_regs(current)->orig_ax = __NR_execve;
672 current_thread_info()->status &= ~TS_COMPAT;
673 if (current->mm)
674 current->mm->context.flags = MM_CONTEXT_HAS_VSYSCALL;
675
676 /* TBD: overwrites user setup. Should have two bits.
677 But 64bit processes have always behaved this way,
678 so it's not too bad. The main problem is just that
679 32bit children are affected again. */
680 current->personality &= ~READ_IMPLIES_EXEC;
681}
682
683static void __set_personality_x32(void)
684{
685#ifdef CONFIG_X86_X32_ABI
686 if (current->mm)
687 current->mm->context.flags = 0;
688
689 current->personality &= ~READ_IMPLIES_EXEC;
690 /*
691 * in_32bit_syscall() uses the presence of the x32 syscall bit
692 * flag to determine compat status. The x86 mmap() code relies on
693 * the syscall bitness so set x32 syscall bit right here to make
694 * in_32bit_syscall() work during exec().
695 *
696 * Pretend to come from a x32 execve.
697 */
698 task_pt_regs(current)->orig_ax = __NR_x32_execve | __X32_SYSCALL_BIT;
699 current_thread_info()->status &= ~TS_COMPAT;
700#endif
701}
702
703static void __set_personality_ia32(void)
704{
705#ifdef CONFIG_IA32_EMULATION
706 if (current->mm) {
707 /*
708 * uprobes applied to this MM need to know this and
709 * cannot use user_64bit_mode() at that time.
710 */
711 current->mm->context.flags = MM_CONTEXT_UPROBE_IA32;
712 }
713
714 current->personality |= force_personality32;
715 /* Prepare the first "return" to user space */
716 task_pt_regs(current)->orig_ax = __NR_ia32_execve;
717 current_thread_info()->status |= TS_COMPAT;
718#endif
719}
720
721void set_personality_ia32(bool x32)
722{
723 /* Make sure to be in 32bit mode */
724 set_thread_flag(TIF_ADDR32);
725
726 if (x32)
727 __set_personality_x32();
728 else
729 __set_personality_ia32();
730}
731EXPORT_SYMBOL_GPL(set_personality_ia32);
732
733#ifdef CONFIG_CHECKPOINT_RESTORE
734static long prctl_map_vdso(const struct vdso_image *image, unsigned long addr)
735{
736 int ret;
737
738 ret = map_vdso_once(image, addr);
739 if (ret)
740 return ret;
741
742 return (long)image->size;
743}
744#endif
745
746long do_arch_prctl_64(struct task_struct *task, int option, unsigned long arg2)
747{
748 int ret = 0;
749
750 switch (option) {
751 case ARCH_SET_GS: {
752 if (unlikely(arg2 >= TASK_SIZE_MAX))
753 return -EPERM;
754
755 preempt_disable();
756 /*
757 * ARCH_SET_GS has always overwritten the index
758 * and the base. Zero is the most sensible value
759 * to put in the index, and is the only value that
760 * makes any sense if FSGSBASE is unavailable.
761 */
762 if (task == current) {
763 loadseg(GS, 0);
764 x86_gsbase_write_cpu_inactive(arg2);
765
766 /*
767 * On non-FSGSBASE systems, save_base_legacy() expects
768 * that we also fill in thread.gsbase.
769 */
770 task->thread.gsbase = arg2;
771
772 } else {
773 task->thread.gsindex = 0;
774 x86_gsbase_write_task(task, arg2);
775 }
776 preempt_enable();
777 break;
778 }
779 case ARCH_SET_FS: {
780 /*
781 * Not strictly needed for %fs, but do it for symmetry
782 * with %gs
783 */
784 if (unlikely(arg2 >= TASK_SIZE_MAX))
785 return -EPERM;
786
787 preempt_disable();
788 /*
789 * Set the selector to 0 for the same reason
790 * as %gs above.
791 */
792 if (task == current) {
793 loadseg(FS, 0);
794 x86_fsbase_write_cpu(arg2);
795
796 /*
797 * On non-FSGSBASE systems, save_base_legacy() expects
798 * that we also fill in thread.fsbase.
799 */
800 task->thread.fsbase = arg2;
801 } else {
802 task->thread.fsindex = 0;
803 x86_fsbase_write_task(task, arg2);
804 }
805 preempt_enable();
806 break;
807 }
808 case ARCH_GET_FS: {
809 unsigned long base = x86_fsbase_read_task(task);
810
811 ret = put_user(base, (unsigned long __user *)arg2);
812 break;
813 }
814 case ARCH_GET_GS: {
815 unsigned long base = x86_gsbase_read_task(task);
816
817 ret = put_user(base, (unsigned long __user *)arg2);
818 break;
819 }
820
821#ifdef CONFIG_CHECKPOINT_RESTORE
822# ifdef CONFIG_X86_X32_ABI
823 case ARCH_MAP_VDSO_X32:
824 return prctl_map_vdso(&vdso_image_x32, arg2);
825# endif
826# if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
827 case ARCH_MAP_VDSO_32:
828 return prctl_map_vdso(&vdso_image_32, arg2);
829# endif
830 case ARCH_MAP_VDSO_64:
831 return prctl_map_vdso(&vdso_image_64, arg2);
832#endif
833
834 default:
835 ret = -EINVAL;
836 break;
837 }
838
839 return ret;
840}
841
842SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
843{
844 long ret;
845
846 ret = do_arch_prctl_64(current, option, arg2);
847 if (ret == -EINVAL)
848 ret = do_arch_prctl_common(option, arg2);
849
850 return ret;
851}
852
853#ifdef CONFIG_IA32_EMULATION
854COMPAT_SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
855{
856 return do_arch_prctl_common(option, arg2);
857}
858#endif
859
860unsigned long KSTK_ESP(struct task_struct *task)
861{
862 return task_pt_regs(task)->sp;
863}
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 */
67void __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
145void release_thread(struct task_struct *dead_task)
146{
147 WARN_ON(dead_task->mm);
148}
149
150enum 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 */
163static 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 */
190static 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 */
211static __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
250static __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 */
272void 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)
282EXPORT_SYMBOL_GPL(current_save_fsgs);
283#endif
284
285static __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
294static __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 */
350static __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
367static __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
388unsigned 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
431unsigned 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
448void 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
461unsigned 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
476unsigned 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
491void 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
498void 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
505static void
506start_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 reset_thread_features();
519
520 loadsegment(fs, 0);
521 loadsegment(es, _ds);
522 loadsegment(ds, _ds);
523 load_gs_index(0);
524
525 regs->ip = new_ip;
526 regs->sp = new_sp;
527 regs->cs = _cs;
528 regs->ss = _ss;
529 regs->flags = X86_EFLAGS_IF;
530}
531
532void
533start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
534{
535 start_thread_common(regs, new_ip, new_sp,
536 __USER_CS, __USER_DS, 0);
537}
538EXPORT_SYMBOL_GPL(start_thread);
539
540#ifdef CONFIG_COMPAT
541void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp, bool x32)
542{
543 start_thread_common(regs, new_ip, new_sp,
544 x32 ? __USER_CS : __USER32_CS,
545 __USER_DS, __USER_DS);
546}
547#endif
548
549/*
550 * switch_to(x,y) should switch tasks from x to y.
551 *
552 * This could still be optimized:
553 * - fold all the options into a flag word and test it with a single test.
554 * - could test fs/gs bitsliced
555 *
556 * Kprobes not supported here. Set the probe on schedule instead.
557 * Function graph tracer not supported too.
558 */
559__no_kmsan_checks
560__visible __notrace_funcgraph struct task_struct *
561__switch_to(struct task_struct *prev_p, struct task_struct *next_p)
562{
563 struct thread_struct *prev = &prev_p->thread;
564 struct thread_struct *next = &next_p->thread;
565 struct fpu *prev_fpu = &prev->fpu;
566 int cpu = smp_processor_id();
567
568 WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_ENTRY) &&
569 this_cpu_read(pcpu_hot.hardirq_stack_inuse));
570
571 if (!test_thread_flag(TIF_NEED_FPU_LOAD))
572 switch_fpu_prepare(prev_fpu, cpu);
573
574 /* We must save %fs and %gs before load_TLS() because
575 * %fs and %gs may be cleared by load_TLS().
576 *
577 * (e.g. xen_load_tls())
578 */
579 save_fsgs(prev_p);
580
581 /*
582 * Load TLS before restoring any segments so that segment loads
583 * reference the correct GDT entries.
584 */
585 load_TLS(next, cpu);
586
587 /*
588 * Leave lazy mode, flushing any hypercalls made here. This
589 * must be done after loading TLS entries in the GDT but before
590 * loading segments that might reference them.
591 */
592 arch_end_context_switch(next_p);
593
594 /* Switch DS and ES.
595 *
596 * Reading them only returns the selectors, but writing them (if
597 * nonzero) loads the full descriptor from the GDT or LDT. The
598 * LDT for next is loaded in switch_mm, and the GDT is loaded
599 * above.
600 *
601 * We therefore need to write new values to the segment
602 * registers on every context switch unless both the new and old
603 * values are zero.
604 *
605 * Note that we don't need to do anything for CS and SS, as
606 * those are saved and restored as part of pt_regs.
607 */
608 savesegment(es, prev->es);
609 if (unlikely(next->es | prev->es))
610 loadsegment(es, next->es);
611
612 savesegment(ds, prev->ds);
613 if (unlikely(next->ds | prev->ds))
614 loadsegment(ds, next->ds);
615
616 x86_fsgsbase_load(prev, next);
617
618 x86_pkru_load(prev, next);
619
620 /*
621 * Switch the PDA and FPU contexts.
622 */
623 raw_cpu_write(pcpu_hot.current_task, next_p);
624 raw_cpu_write(pcpu_hot.top_of_stack, task_top_of_stack(next_p));
625
626 switch_fpu_finish();
627
628 /* Reload sp0. */
629 update_task_stack(next_p);
630
631 switch_to_extra(prev_p, next_p);
632
633 if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
634 /*
635 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
636 * does not update the cached descriptor. As a result, if we
637 * do SYSRET while SS is NULL, we'll end up in user mode with
638 * SS apparently equal to __USER_DS but actually unusable.
639 *
640 * The straightforward workaround would be to fix it up just
641 * before SYSRET, but that would slow down the system call
642 * fast paths. Instead, we ensure that SS is never NULL in
643 * system call context. We do this by replacing NULL SS
644 * selectors at every context switch. SYSCALL sets up a valid
645 * SS, so the only way to get NULL is to re-enter the kernel
646 * from CPL 3 through an interrupt. Since that can't happen
647 * in the same task as a running syscall, we are guaranteed to
648 * context switch between every interrupt vector entry and a
649 * subsequent SYSRET.
650 *
651 * We read SS first because SS reads are much faster than
652 * writes. Out of caution, we force SS to __KERNEL_DS even if
653 * it previously had a different non-NULL value.
654 */
655 unsigned short ss_sel;
656 savesegment(ss, ss_sel);
657 if (ss_sel != __KERNEL_DS)
658 loadsegment(ss, __KERNEL_DS);
659 }
660
661 /* Load the Intel cache allocation PQR MSR. */
662 resctrl_sched_in(next_p);
663
664 return prev_p;
665}
666
667void set_personality_64bit(void)
668{
669 /* inherit personality from parent */
670
671 /* Make sure to be in 64bit mode */
672 clear_thread_flag(TIF_ADDR32);
673 /* Pretend that this comes from a 64bit execve */
674 task_pt_regs(current)->orig_ax = __NR_execve;
675 current_thread_info()->status &= ~TS_COMPAT;
676 if (current->mm)
677 __set_bit(MM_CONTEXT_HAS_VSYSCALL, ¤t->mm->context.flags);
678
679 /* TBD: overwrites user setup. Should have two bits.
680 But 64bit processes have always behaved this way,
681 so it's not too bad. The main problem is just that
682 32bit children are affected again. */
683 current->personality &= ~READ_IMPLIES_EXEC;
684}
685
686static void __set_personality_x32(void)
687{
688#ifdef CONFIG_X86_X32_ABI
689 if (current->mm)
690 current->mm->context.flags = 0;
691
692 current->personality &= ~READ_IMPLIES_EXEC;
693 /*
694 * in_32bit_syscall() uses the presence of the x32 syscall bit
695 * flag to determine compat status. The x86 mmap() code relies on
696 * the syscall bitness so set x32 syscall bit right here to make
697 * in_32bit_syscall() work during exec().
698 *
699 * Pretend to come from a x32 execve.
700 */
701 task_pt_regs(current)->orig_ax = __NR_x32_execve | __X32_SYSCALL_BIT;
702 current_thread_info()->status &= ~TS_COMPAT;
703#endif
704}
705
706static void __set_personality_ia32(void)
707{
708#ifdef CONFIG_IA32_EMULATION
709 if (current->mm) {
710 /*
711 * uprobes applied to this MM need to know this and
712 * cannot use user_64bit_mode() at that time.
713 */
714 __set_bit(MM_CONTEXT_UPROBE_IA32, ¤t->mm->context.flags);
715 }
716
717 current->personality |= force_personality32;
718 /* Prepare the first "return" to user space */
719 task_pt_regs(current)->orig_ax = __NR_ia32_execve;
720 current_thread_info()->status |= TS_COMPAT;
721#endif
722}
723
724void set_personality_ia32(bool x32)
725{
726 /* Make sure to be in 32bit mode */
727 set_thread_flag(TIF_ADDR32);
728
729 if (x32)
730 __set_personality_x32();
731 else
732 __set_personality_ia32();
733}
734EXPORT_SYMBOL_GPL(set_personality_ia32);
735
736#ifdef CONFIG_CHECKPOINT_RESTORE
737static long prctl_map_vdso(const struct vdso_image *image, unsigned long addr)
738{
739 int ret;
740
741 ret = map_vdso_once(image, addr);
742 if (ret)
743 return ret;
744
745 return (long)image->size;
746}
747#endif
748
749#ifdef CONFIG_ADDRESS_MASKING
750
751#define LAM_U57_BITS 6
752
753static int prctl_enable_tagged_addr(struct mm_struct *mm, unsigned long nr_bits)
754{
755 if (!cpu_feature_enabled(X86_FEATURE_LAM))
756 return -ENODEV;
757
758 /* PTRACE_ARCH_PRCTL */
759 if (current->mm != mm)
760 return -EINVAL;
761
762 if (mm_valid_pasid(mm) &&
763 !test_bit(MM_CONTEXT_FORCE_TAGGED_SVA, &mm->context.flags))
764 return -EINVAL;
765
766 if (mmap_write_lock_killable(mm))
767 return -EINTR;
768
769 if (test_bit(MM_CONTEXT_LOCK_LAM, &mm->context.flags)) {
770 mmap_write_unlock(mm);
771 return -EBUSY;
772 }
773
774 if (!nr_bits) {
775 mmap_write_unlock(mm);
776 return -EINVAL;
777 } else if (nr_bits <= LAM_U57_BITS) {
778 mm->context.lam_cr3_mask = X86_CR3_LAM_U57;
779 mm->context.untag_mask = ~GENMASK(62, 57);
780 } else {
781 mmap_write_unlock(mm);
782 return -EINVAL;
783 }
784
785 write_cr3(__read_cr3() | mm->context.lam_cr3_mask);
786 set_tlbstate_lam_mode(mm);
787 set_bit(MM_CONTEXT_LOCK_LAM, &mm->context.flags);
788
789 mmap_write_unlock(mm);
790
791 return 0;
792}
793#endif
794
795long do_arch_prctl_64(struct task_struct *task, int option, unsigned long arg2)
796{
797 int ret = 0;
798
799 switch (option) {
800 case ARCH_SET_GS: {
801 if (unlikely(arg2 >= TASK_SIZE_MAX))
802 return -EPERM;
803
804 preempt_disable();
805 /*
806 * ARCH_SET_GS has always overwritten the index
807 * and the base. Zero is the most sensible value
808 * to put in the index, and is the only value that
809 * makes any sense if FSGSBASE is unavailable.
810 */
811 if (task == current) {
812 loadseg(GS, 0);
813 x86_gsbase_write_cpu_inactive(arg2);
814
815 /*
816 * On non-FSGSBASE systems, save_base_legacy() expects
817 * that we also fill in thread.gsbase.
818 */
819 task->thread.gsbase = arg2;
820
821 } else {
822 task->thread.gsindex = 0;
823 x86_gsbase_write_task(task, arg2);
824 }
825 preempt_enable();
826 break;
827 }
828 case ARCH_SET_FS: {
829 /*
830 * Not strictly needed for %fs, but do it for symmetry
831 * with %gs
832 */
833 if (unlikely(arg2 >= TASK_SIZE_MAX))
834 return -EPERM;
835
836 preempt_disable();
837 /*
838 * Set the selector to 0 for the same reason
839 * as %gs above.
840 */
841 if (task == current) {
842 loadseg(FS, 0);
843 x86_fsbase_write_cpu(arg2);
844
845 /*
846 * On non-FSGSBASE systems, save_base_legacy() expects
847 * that we also fill in thread.fsbase.
848 */
849 task->thread.fsbase = arg2;
850 } else {
851 task->thread.fsindex = 0;
852 x86_fsbase_write_task(task, arg2);
853 }
854 preempt_enable();
855 break;
856 }
857 case ARCH_GET_FS: {
858 unsigned long base = x86_fsbase_read_task(task);
859
860 ret = put_user(base, (unsigned long __user *)arg2);
861 break;
862 }
863 case ARCH_GET_GS: {
864 unsigned long base = x86_gsbase_read_task(task);
865
866 ret = put_user(base, (unsigned long __user *)arg2);
867 break;
868 }
869
870#ifdef CONFIG_CHECKPOINT_RESTORE
871# ifdef CONFIG_X86_X32_ABI
872 case ARCH_MAP_VDSO_X32:
873 return prctl_map_vdso(&vdso_image_x32, arg2);
874# endif
875# if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
876 case ARCH_MAP_VDSO_32:
877 return prctl_map_vdso(&vdso_image_32, arg2);
878# endif
879 case ARCH_MAP_VDSO_64:
880 return prctl_map_vdso(&vdso_image_64, arg2);
881#endif
882#ifdef CONFIG_ADDRESS_MASKING
883 case ARCH_GET_UNTAG_MASK:
884 return put_user(task->mm->context.untag_mask,
885 (unsigned long __user *)arg2);
886 case ARCH_ENABLE_TAGGED_ADDR:
887 return prctl_enable_tagged_addr(task->mm, arg2);
888 case ARCH_FORCE_TAGGED_SVA:
889 if (current != task)
890 return -EINVAL;
891 set_bit(MM_CONTEXT_FORCE_TAGGED_SVA, &task->mm->context.flags);
892 return 0;
893 case ARCH_GET_MAX_TAG_BITS:
894 if (!cpu_feature_enabled(X86_FEATURE_LAM))
895 return put_user(0, (unsigned long __user *)arg2);
896 else
897 return put_user(LAM_U57_BITS, (unsigned long __user *)arg2);
898#endif
899 case ARCH_SHSTK_ENABLE:
900 case ARCH_SHSTK_DISABLE:
901 case ARCH_SHSTK_LOCK:
902 case ARCH_SHSTK_UNLOCK:
903 case ARCH_SHSTK_STATUS:
904 return shstk_prctl(task, option, arg2);
905 default:
906 ret = -EINVAL;
907 break;
908 }
909
910 return ret;
911}
912
913SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
914{
915 long ret;
916
917 ret = do_arch_prctl_64(current, option, arg2);
918 if (ret == -EINVAL)
919 ret = do_arch_prctl_common(option, arg2);
920
921 return ret;
922}
923
924#ifdef CONFIG_IA32_EMULATION
925COMPAT_SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
926{
927 return do_arch_prctl_common(option, arg2);
928}
929#endif
930
931unsigned long KSTK_ESP(struct task_struct *task)
932{
933 return task_pt_regs(task)->sp;
934}