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#include <asm/fred.h>
60#ifdef CONFIG_IA32_EMULATION
61/* Not included via unistd.h */
62#include <asm/unistd_32_ia32.h>
63#endif
64
65#include "process.h"
66
67/* Prints also some state that isn't saved in the pt_regs */
68void __show_regs(struct pt_regs *regs, enum show_regs_mode mode,
69 const char *log_lvl)
70{
71 unsigned long cr0 = 0L, cr2 = 0L, cr3 = 0L, cr4 = 0L, fs, gs, shadowgs;
72 unsigned long d0, d1, d2, d3, d6, d7;
73 unsigned int fsindex, gsindex;
74 unsigned int ds, es;
75
76 show_iret_regs(regs, log_lvl);
77
78 if (regs->orig_ax != -1)
79 pr_cont(" ORIG_RAX: %016lx\n", regs->orig_ax);
80 else
81 pr_cont("\n");
82
83 printk("%sRAX: %016lx RBX: %016lx RCX: %016lx\n",
84 log_lvl, regs->ax, regs->bx, regs->cx);
85 printk("%sRDX: %016lx RSI: %016lx RDI: %016lx\n",
86 log_lvl, regs->dx, regs->si, regs->di);
87 printk("%sRBP: %016lx R08: %016lx R09: %016lx\n",
88 log_lvl, regs->bp, regs->r8, regs->r9);
89 printk("%sR10: %016lx R11: %016lx R12: %016lx\n",
90 log_lvl, regs->r10, regs->r11, regs->r12);
91 printk("%sR13: %016lx R14: %016lx R15: %016lx\n",
92 log_lvl, regs->r13, regs->r14, regs->r15);
93
94 if (mode == SHOW_REGS_SHORT)
95 return;
96
97 if (mode == SHOW_REGS_USER) {
98 rdmsrl(MSR_FS_BASE, fs);
99 rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
100 printk("%sFS: %016lx GS: %016lx\n",
101 log_lvl, fs, shadowgs);
102 return;
103 }
104
105 asm("movl %%ds,%0" : "=r" (ds));
106 asm("movl %%es,%0" : "=r" (es));
107 asm("movl %%fs,%0" : "=r" (fsindex));
108 asm("movl %%gs,%0" : "=r" (gsindex));
109
110 rdmsrl(MSR_FS_BASE, fs);
111 rdmsrl(MSR_GS_BASE, gs);
112 rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
113
114 cr0 = read_cr0();
115 cr2 = read_cr2();
116 cr3 = __read_cr3();
117 cr4 = __read_cr4();
118
119 printk("%sFS: %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
120 log_lvl, fs, fsindex, gs, gsindex, shadowgs);
121 printk("%sCS: %04x DS: %04x ES: %04x CR0: %016lx\n",
122 log_lvl, regs->cs, ds, es, cr0);
123 printk("%sCR2: %016lx CR3: %016lx CR4: %016lx\n",
124 log_lvl, cr2, cr3, cr4);
125
126 get_debugreg(d0, 0);
127 get_debugreg(d1, 1);
128 get_debugreg(d2, 2);
129 get_debugreg(d3, 3);
130 get_debugreg(d6, 6);
131 get_debugreg(d7, 7);
132
133 /* Only print out debug registers if they are in their non-default state. */
134 if (!((d0 == 0) && (d1 == 0) && (d2 == 0) && (d3 == 0) &&
135 (d6 == DR6_RESERVED) && (d7 == 0x400))) {
136 printk("%sDR0: %016lx DR1: %016lx DR2: %016lx\n",
137 log_lvl, d0, d1, d2);
138 printk("%sDR3: %016lx DR6: %016lx DR7: %016lx\n",
139 log_lvl, d3, d6, d7);
140 }
141
142 if (cpu_feature_enabled(X86_FEATURE_OSPKE))
143 printk("%sPKRU: %08x\n", log_lvl, read_pkru());
144}
145
146void release_thread(struct task_struct *dead_task)
147{
148 WARN_ON(dead_task->mm);
149}
150
151enum which_selector {
152 FS,
153 GS
154};
155
156/*
157 * Out of line to be protected from kprobes and tracing. If this would be
158 * traced or probed than any access to a per CPU variable happens with
159 * the wrong GS.
160 *
161 * It is not used on Xen paravirt. When paravirt support is needed, it
162 * needs to be renamed with native_ prefix.
163 */
164static noinstr unsigned long __rdgsbase_inactive(void)
165{
166 unsigned long gsbase;
167
168 lockdep_assert_irqs_disabled();
169
170 /*
171 * SWAPGS is no longer needed thus NOT allowed with FRED because
172 * FRED transitions ensure that an operating system can _always_
173 * operate with its own GS base address:
174 * - For events that occur in ring 3, FRED event delivery swaps
175 * the GS base address with the IA32_KERNEL_GS_BASE MSR.
176 * - ERETU (the FRED transition that returns to ring 3) also swaps
177 * the GS base address with the IA32_KERNEL_GS_BASE MSR.
178 *
179 * And the operating system can still setup the GS segment for a
180 * user thread without the need of loading a user thread GS with:
181 * - Using LKGS, available with FRED, to modify other attributes
182 * of the GS segment without compromising its ability always to
183 * operate with its own GS base address.
184 * - Accessing the GS segment base address for a user thread as
185 * before using RDMSR or WRMSR on the IA32_KERNEL_GS_BASE MSR.
186 *
187 * Note, LKGS loads the GS base address into the IA32_KERNEL_GS_BASE
188 * MSR instead of the GS segment’s descriptor cache. As such, the
189 * operating system never changes its runtime GS base address.
190 */
191 if (!cpu_feature_enabled(X86_FEATURE_FRED) &&
192 !cpu_feature_enabled(X86_FEATURE_XENPV)) {
193 native_swapgs();
194 gsbase = rdgsbase();
195 native_swapgs();
196 } else {
197 instrumentation_begin();
198 rdmsrl(MSR_KERNEL_GS_BASE, gsbase);
199 instrumentation_end();
200 }
201
202 return gsbase;
203}
204
205/*
206 * Out of line to be protected from kprobes and tracing. If this would be
207 * traced or probed than any access to a per CPU variable happens with
208 * the wrong GS.
209 *
210 * It is not used on Xen paravirt. When paravirt support is needed, it
211 * needs to be renamed with native_ prefix.
212 */
213static noinstr void __wrgsbase_inactive(unsigned long gsbase)
214{
215 lockdep_assert_irqs_disabled();
216
217 if (!cpu_feature_enabled(X86_FEATURE_FRED) &&
218 !cpu_feature_enabled(X86_FEATURE_XENPV)) {
219 native_swapgs();
220 wrgsbase(gsbase);
221 native_swapgs();
222 } else {
223 instrumentation_begin();
224 wrmsrl(MSR_KERNEL_GS_BASE, val: gsbase);
225 instrumentation_end();
226 }
227}
228
229/*
230 * Saves the FS or GS base for an outgoing thread if FSGSBASE extensions are
231 * not available. The goal is to be reasonably fast on non-FSGSBASE systems.
232 * It's forcibly inlined because it'll generate better code and this function
233 * is hot.
234 */
235static __always_inline void save_base_legacy(struct task_struct *prev_p,
236 unsigned short selector,
237 enum which_selector which)
238{
239 if (likely(selector == 0)) {
240 /*
241 * On Intel (without X86_BUG_NULL_SEG), the segment base could
242 * be the pre-existing saved base or it could be zero. On AMD
243 * (with X86_BUG_NULL_SEG), the segment base could be almost
244 * anything.
245 *
246 * This branch is very hot (it's hit twice on almost every
247 * context switch between 64-bit programs), and avoiding
248 * the RDMSR helps a lot, so we just assume that whatever
249 * value is already saved is correct. This matches historical
250 * Linux behavior, so it won't break existing applications.
251 *
252 * To avoid leaking state, on non-X86_BUG_NULL_SEG CPUs, if we
253 * report that the base is zero, it needs to actually be zero:
254 * see the corresponding logic in load_seg_legacy.
255 */
256 } else {
257 /*
258 * If the selector is 1, 2, or 3, then the base is zero on
259 * !X86_BUG_NULL_SEG CPUs and could be anything on
260 * X86_BUG_NULL_SEG CPUs. In the latter case, Linux
261 * has never attempted to preserve the base across context
262 * switches.
263 *
264 * If selector > 3, then it refers to a real segment, and
265 * saving the base isn't necessary.
266 */
267 if (which == FS)
268 prev_p->thread.fsbase = 0;
269 else
270 prev_p->thread.gsbase = 0;
271 }
272}
273
274static __always_inline void save_fsgs(struct task_struct *task)
275{
276 savesegment(fs, task->thread.fsindex);
277 savesegment(gs, task->thread.gsindex);
278 if (static_cpu_has(X86_FEATURE_FSGSBASE)) {
279 /*
280 * If FSGSBASE is enabled, we can't make any useful guesses
281 * about the base, and user code expects us to save the current
282 * value. Fortunately, reading the base directly is efficient.
283 */
284 task->thread.fsbase = rdfsbase();
285 task->thread.gsbase = __rdgsbase_inactive();
286 } else {
287 save_base_legacy(prev_p: task, selector: task->thread.fsindex, which: FS);
288 save_base_legacy(prev_p: task, selector: task->thread.gsindex, which: GS);
289 }
290}
291
292/*
293 * While a process is running,current->thread.fsbase and current->thread.gsbase
294 * may not match the corresponding CPU registers (see save_base_legacy()).
295 */
296void current_save_fsgs(void)
297{
298 unsigned long flags;
299
300 /* Interrupts need to be off for FSGSBASE */
301 local_irq_save(flags);
302 save_fsgs(current);
303 local_irq_restore(flags);
304}
305#if IS_ENABLED(CONFIG_KVM)
306EXPORT_SYMBOL_GPL(current_save_fsgs);
307#endif
308
309static __always_inline void loadseg(enum which_selector which,
310 unsigned short sel)
311{
312 if (which == FS)
313 loadsegment(fs, sel);
314 else
315 load_gs_index(gs: sel);
316}
317
318static __always_inline void load_seg_legacy(unsigned short prev_index,
319 unsigned long prev_base,
320 unsigned short next_index,
321 unsigned long next_base,
322 enum which_selector which)
323{
324 if (likely(next_index <= 3)) {
325 /*
326 * The next task is using 64-bit TLS, is not using this
327 * segment at all, or is having fun with arcane CPU features.
328 */
329 if (next_base == 0) {
330 /*
331 * Nasty case: on AMD CPUs, we need to forcibly zero
332 * the base.
333 */
334 if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
335 loadseg(which, __USER_DS);
336 loadseg(which, sel: next_index);
337 } else {
338 /*
339 * We could try to exhaustively detect cases
340 * under which we can skip the segment load,
341 * but there's really only one case that matters
342 * for performance: if both the previous and
343 * next states are fully zeroed, we can skip
344 * the load.
345 *
346 * (This assumes that prev_base == 0 has no
347 * false positives. This is the case on
348 * Intel-style CPUs.)
349 */
350 if (likely(prev_index | next_index | prev_base))
351 loadseg(which, sel: next_index);
352 }
353 } else {
354 if (prev_index != next_index)
355 loadseg(which, sel: next_index);
356 wrmsrl(msr: which == FS ? MSR_FS_BASE : MSR_KERNEL_GS_BASE,
357 val: next_base);
358 }
359 } else {
360 /*
361 * The next task is using a real segment. Loading the selector
362 * is sufficient.
363 */
364 loadseg(which, sel: next_index);
365 }
366}
367
368/*
369 * Store prev's PKRU value and load next's PKRU value if they differ. PKRU
370 * is not XSTATE managed on context switch because that would require a
371 * lookup in the task's FPU xsave buffer and require to keep that updated
372 * in various places.
373 */
374static __always_inline void x86_pkru_load(struct thread_struct *prev,
375 struct thread_struct *next)
376{
377 if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
378 return;
379
380 /* Stash the prev task's value: */
381 prev->pkru = rdpkru();
382
383 /*
384 * PKRU writes are slightly expensive. Avoid them when not
385 * strictly necessary:
386 */
387 if (prev->pkru != next->pkru)
388 wrpkru(pkru: next->pkru);
389}
390
391static __always_inline void x86_fsgsbase_load(struct thread_struct *prev,
392 struct thread_struct *next)
393{
394 if (static_cpu_has(X86_FEATURE_FSGSBASE)) {
395 /* Update the FS and GS selectors if they could have changed. */
396 if (unlikely(prev->fsindex || next->fsindex))
397 loadseg(which: FS, sel: next->fsindex);
398 if (unlikely(prev->gsindex || next->gsindex))
399 loadseg(which: GS, sel: next->gsindex);
400
401 /* Update the bases. */
402 wrfsbase(fsbase: next->fsbase);
403 __wrgsbase_inactive(gsbase: next->gsbase);
404 } else {
405 load_seg_legacy(prev_index: prev->fsindex, prev_base: prev->fsbase,
406 next_index: next->fsindex, next_base: next->fsbase, which: FS);
407 load_seg_legacy(prev_index: prev->gsindex, prev_base: prev->gsbase,
408 next_index: next->gsindex, next_base: next->gsbase, which: GS);
409 }
410}
411
412unsigned long x86_fsgsbase_read_task(struct task_struct *task,
413 unsigned short selector)
414{
415 unsigned short idx = selector >> 3;
416 unsigned long base;
417
418 if (likely((selector & SEGMENT_TI_MASK) == 0)) {
419 if (unlikely(idx >= GDT_ENTRIES))
420 return 0;
421
422 /*
423 * There are no user segments in the GDT with nonzero bases
424 * other than the TLS segments.
425 */
426 if (idx < GDT_ENTRY_TLS_MIN || idx > GDT_ENTRY_TLS_MAX)
427 return 0;
428
429 idx -= GDT_ENTRY_TLS_MIN;
430 base = get_desc_base(desc: &task->thread.tls_array[idx]);
431 } else {
432#ifdef CONFIG_MODIFY_LDT_SYSCALL
433 struct ldt_struct *ldt;
434
435 /*
436 * If performance here mattered, we could protect the LDT
437 * with RCU. This is a slow path, though, so we can just
438 * take the mutex.
439 */
440 mutex_lock(&task->mm->context.lock);
441 ldt = task->mm->context.ldt;
442 if (unlikely(!ldt || idx >= ldt->nr_entries))
443 base = 0;
444 else
445 base = get_desc_base(desc: ldt->entries + idx);
446 mutex_unlock(lock: &task->mm->context.lock);
447#else
448 base = 0;
449#endif
450 }
451
452 return base;
453}
454
455unsigned long x86_gsbase_read_cpu_inactive(void)
456{
457 unsigned long gsbase;
458
459 if (boot_cpu_has(X86_FEATURE_FSGSBASE)) {
460 unsigned long flags;
461
462 local_irq_save(flags);
463 gsbase = __rdgsbase_inactive();
464 local_irq_restore(flags);
465 } else {
466 rdmsrl(MSR_KERNEL_GS_BASE, gsbase);
467 }
468
469 return gsbase;
470}
471
472void x86_gsbase_write_cpu_inactive(unsigned long gsbase)
473{
474 if (boot_cpu_has(X86_FEATURE_FSGSBASE)) {
475 unsigned long flags;
476
477 local_irq_save(flags);
478 __wrgsbase_inactive(gsbase);
479 local_irq_restore(flags);
480 } else {
481 wrmsrl(MSR_KERNEL_GS_BASE, val: gsbase);
482 }
483}
484
485unsigned long x86_fsbase_read_task(struct task_struct *task)
486{
487 unsigned long fsbase;
488
489 if (task == current)
490 fsbase = x86_fsbase_read_cpu();
491 else if (boot_cpu_has(X86_FEATURE_FSGSBASE) ||
492 (task->thread.fsindex == 0))
493 fsbase = task->thread.fsbase;
494 else
495 fsbase = x86_fsgsbase_read_task(task, selector: task->thread.fsindex);
496
497 return fsbase;
498}
499
500unsigned long x86_gsbase_read_task(struct task_struct *task)
501{
502 unsigned long gsbase;
503
504 if (task == current)
505 gsbase = x86_gsbase_read_cpu_inactive();
506 else if (boot_cpu_has(X86_FEATURE_FSGSBASE) ||
507 (task->thread.gsindex == 0))
508 gsbase = task->thread.gsbase;
509 else
510 gsbase = x86_fsgsbase_read_task(task, selector: task->thread.gsindex);
511
512 return gsbase;
513}
514
515void x86_fsbase_write_task(struct task_struct *task, unsigned long fsbase)
516{
517 WARN_ON_ONCE(task == current);
518
519 task->thread.fsbase = fsbase;
520}
521
522void x86_gsbase_write_task(struct task_struct *task, unsigned long gsbase)
523{
524 WARN_ON_ONCE(task == current);
525
526 task->thread.gsbase = gsbase;
527}
528
529static void
530start_thread_common(struct pt_regs *regs, unsigned long new_ip,
531 unsigned long new_sp,
532 u16 _cs, u16 _ss, u16 _ds)
533{
534 WARN_ON_ONCE(regs != current_pt_regs());
535
536 if (static_cpu_has(X86_BUG_NULL_SEG)) {
537 /* Loading zero below won't clear the base. */
538 loadsegment(fs, __USER_DS);
539 load_gs_index(__USER_DS);
540 }
541
542 reset_thread_features();
543
544 loadsegment(fs, 0);
545 loadsegment(es, _ds);
546 loadsegment(ds, _ds);
547 load_gs_index(gs: 0);
548
549 regs->ip = new_ip;
550 regs->sp = new_sp;
551 regs->csx = _cs;
552 regs->ssx = _ss;
553 /*
554 * Allow single-step trap and NMI when starting a new task, thus
555 * once the new task enters user space, single-step trap and NMI
556 * are both enabled immediately.
557 *
558 * Entering a new task is logically speaking a return from a
559 * system call (exec, fork, clone, etc.). As such, if ptrace
560 * enables single stepping a single step exception should be
561 * allowed to trigger immediately upon entering user space.
562 * This is not optional.
563 *
564 * NMI should *never* be disabled in user space. As such, this
565 * is an optional, opportunistic way to catch errors.
566 *
567 * Paranoia: High-order 48 bits above the lowest 16 bit SS are
568 * discarded by the legacy IRET instruction on all Intel, AMD,
569 * and Cyrix/Centaur/VIA CPUs, thus can be set unconditionally,
570 * even when FRED is not enabled. But we choose the safer side
571 * to use these bits only when FRED is enabled.
572 */
573 if (cpu_feature_enabled(X86_FEATURE_FRED)) {
574 regs->fred_ss.swevent = true;
575 regs->fred_ss.nmi = true;
576 }
577
578 regs->flags = X86_EFLAGS_IF | X86_EFLAGS_FIXED;
579}
580
581void
582start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
583{
584 start_thread_common(regs, new_ip, new_sp,
585 __USER_CS, __USER_DS, ds: 0);
586}
587EXPORT_SYMBOL_GPL(start_thread);
588
589#ifdef CONFIG_COMPAT
590void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp, bool x32)
591{
592 start_thread_common(regs, new_ip, new_sp,
593 cs: x32 ? __USER_CS : __USER32_CS,
594 __USER_DS, __USER_DS);
595}
596#endif
597
598/*
599 * switch_to(x,y) should switch tasks from x to y.
600 *
601 * This could still be optimized:
602 * - fold all the options into a flag word and test it with a single test.
603 * - could test fs/gs bitsliced
604 *
605 * Kprobes not supported here. Set the probe on schedule instead.
606 * Function graph tracer not supported too.
607 */
608__no_kmsan_checks
609__visible __notrace_funcgraph struct task_struct *
610__switch_to(struct task_struct *prev_p, struct task_struct *next_p)
611{
612 struct thread_struct *prev = &prev_p->thread;
613 struct thread_struct *next = &next_p->thread;
614 int cpu = smp_processor_id();
615
616 WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_ENTRY) &&
617 this_cpu_read(pcpu_hot.hardirq_stack_inuse));
618
619 if (!test_tsk_thread_flag(tsk: prev_p, TIF_NEED_FPU_LOAD))
620 switch_fpu_prepare(old: prev_p, cpu);
621
622 /* We must save %fs and %gs before load_TLS() because
623 * %fs and %gs may be cleared by load_TLS().
624 *
625 * (e.g. xen_load_tls())
626 */
627 save_fsgs(task: prev_p);
628
629 /*
630 * Load TLS before restoring any segments so that segment loads
631 * reference the correct GDT entries.
632 */
633 load_TLS(t: next, cpu);
634
635 /*
636 * Leave lazy mode, flushing any hypercalls made here. This
637 * must be done after loading TLS entries in the GDT but before
638 * loading segments that might reference them.
639 */
640 arch_end_context_switch(next: next_p);
641
642 /* Switch DS and ES.
643 *
644 * Reading them only returns the selectors, but writing them (if
645 * nonzero) loads the full descriptor from the GDT or LDT. The
646 * LDT for next is loaded in switch_mm, and the GDT is loaded
647 * above.
648 *
649 * We therefore need to write new values to the segment
650 * registers on every context switch unless both the new and old
651 * values are zero.
652 *
653 * Note that we don't need to do anything for CS and SS, as
654 * those are saved and restored as part of pt_regs.
655 */
656 savesegment(es, prev->es);
657 if (unlikely(next->es | prev->es))
658 loadsegment(es, next->es);
659
660 savesegment(ds, prev->ds);
661 if (unlikely(next->ds | prev->ds))
662 loadsegment(ds, next->ds);
663
664 x86_fsgsbase_load(prev, next);
665
666 x86_pkru_load(prev, next);
667
668 /*
669 * Switch the PDA and FPU contexts.
670 */
671 raw_cpu_write(pcpu_hot.current_task, next_p);
672 raw_cpu_write(pcpu_hot.top_of_stack, task_top_of_stack(next_p));
673
674 switch_fpu_finish(new: next_p);
675
676 /* Reload sp0. */
677 update_task_stack(task: next_p);
678
679 switch_to_extra(prev: prev_p, next: next_p);
680
681 if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
682 /*
683 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
684 * does not update the cached descriptor. As a result, if we
685 * do SYSRET while SS is NULL, we'll end up in user mode with
686 * SS apparently equal to __USER_DS but actually unusable.
687 *
688 * The straightforward workaround would be to fix it up just
689 * before SYSRET, but that would slow down the system call
690 * fast paths. Instead, we ensure that SS is never NULL in
691 * system call context. We do this by replacing NULL SS
692 * selectors at every context switch. SYSCALL sets up a valid
693 * SS, so the only way to get NULL is to re-enter the kernel
694 * from CPL 3 through an interrupt. Since that can't happen
695 * in the same task as a running syscall, we are guaranteed to
696 * context switch between every interrupt vector entry and a
697 * subsequent SYSRET.
698 *
699 * We read SS first because SS reads are much faster than
700 * writes. Out of caution, we force SS to __KERNEL_DS even if
701 * it previously had a different non-NULL value.
702 */
703 unsigned short ss_sel;
704 savesegment(ss, ss_sel);
705 if (ss_sel != __KERNEL_DS)
706 loadsegment(ss, __KERNEL_DS);
707 }
708
709 /* Load the Intel cache allocation PQR MSR. */
710 resctrl_sched_in(tsk: next_p);
711
712 return prev_p;
713}
714
715void set_personality_64bit(void)
716{
717 /* inherit personality from parent */
718
719 /* Make sure to be in 64bit mode */
720 clear_thread_flag(TIF_ADDR32);
721 /* Pretend that this comes from a 64bit execve */
722 task_pt_regs(current)->orig_ax = __NR_execve;
723 current_thread_info()->status &= ~TS_COMPAT;
724 if (current->mm)
725 __set_bit(MM_CONTEXT_HAS_VSYSCALL, &current->mm->context.flags);
726
727 /* TBD: overwrites user setup. Should have two bits.
728 But 64bit processes have always behaved this way,
729 so it's not too bad. The main problem is just that
730 32bit children are affected again. */
731 current->personality &= ~READ_IMPLIES_EXEC;
732}
733
734static void __set_personality_x32(void)
735{
736#ifdef CONFIG_X86_X32_ABI
737 if (current->mm)
738 current->mm->context.flags = 0;
739
740 current->personality &= ~READ_IMPLIES_EXEC;
741 /*
742 * in_32bit_syscall() uses the presence of the x32 syscall bit
743 * flag to determine compat status. The x86 mmap() code relies on
744 * the syscall bitness so set x32 syscall bit right here to make
745 * in_32bit_syscall() work during exec().
746 *
747 * Pretend to come from a x32 execve.
748 */
749 task_pt_regs(current)->orig_ax = __NR_x32_execve | __X32_SYSCALL_BIT;
750 current_thread_info()->status &= ~TS_COMPAT;
751#endif
752}
753
754static void __set_personality_ia32(void)
755{
756#ifdef CONFIG_IA32_EMULATION
757 if (current->mm) {
758 /*
759 * uprobes applied to this MM need to know this and
760 * cannot use user_64bit_mode() at that time.
761 */
762 __set_bit(MM_CONTEXT_UPROBE_IA32, &current->mm->context.flags);
763 }
764
765 current->personality |= force_personality32;
766 /* Prepare the first "return" to user space */
767 task_pt_regs(current)->orig_ax = __NR_ia32_execve;
768 current_thread_info()->status |= TS_COMPAT;
769#endif
770}
771
772void set_personality_ia32(bool x32)
773{
774 /* Make sure to be in 32bit mode */
775 set_thread_flag(TIF_ADDR32);
776
777 if (x32)
778 __set_personality_x32();
779 else
780 __set_personality_ia32();
781}
782EXPORT_SYMBOL_GPL(set_personality_ia32);
783
784#ifdef CONFIG_CHECKPOINT_RESTORE
785static long prctl_map_vdso(const struct vdso_image *image, unsigned long addr)
786{
787 int ret;
788
789 ret = map_vdso_once(image, addr);
790 if (ret)
791 return ret;
792
793 return (long)image->size;
794}
795#endif
796
797#ifdef CONFIG_ADDRESS_MASKING
798
799#define LAM_U57_BITS 6
800
801static int prctl_enable_tagged_addr(struct mm_struct *mm, unsigned long nr_bits)
802{
803 if (!cpu_feature_enabled(X86_FEATURE_LAM))
804 return -ENODEV;
805
806 /* PTRACE_ARCH_PRCTL */
807 if (current->mm != mm)
808 return -EINVAL;
809
810 if (mm_valid_pasid(mm) &&
811 !test_bit(MM_CONTEXT_FORCE_TAGGED_SVA, &mm->context.flags))
812 return -EINVAL;
813
814 if (mmap_write_lock_killable(mm))
815 return -EINTR;
816
817 if (test_bit(MM_CONTEXT_LOCK_LAM, &mm->context.flags)) {
818 mmap_write_unlock(mm);
819 return -EBUSY;
820 }
821
822 if (!nr_bits) {
823 mmap_write_unlock(mm);
824 return -EINVAL;
825 } else if (nr_bits <= LAM_U57_BITS) {
826 mm->context.lam_cr3_mask = X86_CR3_LAM_U57;
827 mm->context.untag_mask = ~GENMASK(62, 57);
828 } else {
829 mmap_write_unlock(mm);
830 return -EINVAL;
831 }
832
833 write_cr3(x: __read_cr3() | mm->context.lam_cr3_mask);
834 set_tlbstate_lam_mode(mm);
835 set_bit(MM_CONTEXT_LOCK_LAM, addr: &mm->context.flags);
836
837 mmap_write_unlock(mm);
838
839 return 0;
840}
841#endif
842
843long do_arch_prctl_64(struct task_struct *task, int option, unsigned long arg2)
844{
845 int ret = 0;
846
847 switch (option) {
848 case ARCH_SET_GS: {
849 if (unlikely(arg2 >= TASK_SIZE_MAX))
850 return -EPERM;
851
852 preempt_disable();
853 /*
854 * ARCH_SET_GS has always overwritten the index
855 * and the base. Zero is the most sensible value
856 * to put in the index, and is the only value that
857 * makes any sense if FSGSBASE is unavailable.
858 */
859 if (task == current) {
860 loadseg(which: GS, sel: 0);
861 x86_gsbase_write_cpu_inactive(gsbase: arg2);
862
863 /*
864 * On non-FSGSBASE systems, save_base_legacy() expects
865 * that we also fill in thread.gsbase.
866 */
867 task->thread.gsbase = arg2;
868
869 } else {
870 task->thread.gsindex = 0;
871 x86_gsbase_write_task(task, gsbase: arg2);
872 }
873 preempt_enable();
874 break;
875 }
876 case ARCH_SET_FS: {
877 /*
878 * Not strictly needed for %fs, but do it for symmetry
879 * with %gs
880 */
881 if (unlikely(arg2 >= TASK_SIZE_MAX))
882 return -EPERM;
883
884 preempt_disable();
885 /*
886 * Set the selector to 0 for the same reason
887 * as %gs above.
888 */
889 if (task == current) {
890 loadseg(which: FS, sel: 0);
891 x86_fsbase_write_cpu(fsbase: arg2);
892
893 /*
894 * On non-FSGSBASE systems, save_base_legacy() expects
895 * that we also fill in thread.fsbase.
896 */
897 task->thread.fsbase = arg2;
898 } else {
899 task->thread.fsindex = 0;
900 x86_fsbase_write_task(task, fsbase: arg2);
901 }
902 preempt_enable();
903 break;
904 }
905 case ARCH_GET_FS: {
906 unsigned long base = x86_fsbase_read_task(task);
907
908 ret = put_user(base, (unsigned long __user *)arg2);
909 break;
910 }
911 case ARCH_GET_GS: {
912 unsigned long base = x86_gsbase_read_task(task);
913
914 ret = put_user(base, (unsigned long __user *)arg2);
915 break;
916 }
917
918#ifdef CONFIG_CHECKPOINT_RESTORE
919# ifdef CONFIG_X86_X32_ABI
920 case ARCH_MAP_VDSO_X32:
921 return prctl_map_vdso(image: &vdso_image_x32, addr: arg2);
922# endif
923# if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
924 case ARCH_MAP_VDSO_32:
925 return prctl_map_vdso(image: &vdso_image_32, addr: arg2);
926# endif
927 case ARCH_MAP_VDSO_64:
928 return prctl_map_vdso(image: &vdso_image_64, addr: arg2);
929#endif
930#ifdef CONFIG_ADDRESS_MASKING
931 case ARCH_GET_UNTAG_MASK:
932 return put_user(task->mm->context.untag_mask,
933 (unsigned long __user *)arg2);
934 case ARCH_ENABLE_TAGGED_ADDR:
935 return prctl_enable_tagged_addr(mm: task->mm, nr_bits: arg2);
936 case ARCH_FORCE_TAGGED_SVA:
937 if (current != task)
938 return -EINVAL;
939 set_bit(MM_CONTEXT_FORCE_TAGGED_SVA, addr: &task->mm->context.flags);
940 return 0;
941 case ARCH_GET_MAX_TAG_BITS:
942 if (!cpu_feature_enabled(X86_FEATURE_LAM))
943 return put_user(0, (unsigned long __user *)arg2);
944 else
945 return put_user(LAM_U57_BITS, (unsigned long __user *)arg2);
946#endif
947 case ARCH_SHSTK_ENABLE:
948 case ARCH_SHSTK_DISABLE:
949 case ARCH_SHSTK_LOCK:
950 case ARCH_SHSTK_UNLOCK:
951 case ARCH_SHSTK_STATUS:
952 return shstk_prctl(task, option, arg2);
953 default:
954 ret = -EINVAL;
955 break;
956 }
957
958 return ret;
959}
960
961SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
962{
963 long ret;
964
965 ret = do_arch_prctl_64(current, option, arg2);
966 if (ret == -EINVAL)
967 ret = do_arch_prctl_common(option, arg2);
968
969 return ret;
970}
971
972#ifdef CONFIG_IA32_EMULATION
973COMPAT_SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
974{
975 return do_arch_prctl_common(option, arg2);
976}
977#endif
978
979unsigned long KSTK_ESP(struct task_struct *task)
980{
981 return task_pt_regs(task)->sp;
982}
983

source code of linux/arch/x86/kernel/process_64.c