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/pgtable.h>
44#include <asm/processor.h>
45#include <asm/fpu/internal.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/syscalls.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_sched.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{
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);
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(KERN_DEFAULT "RAX: %016lx RBX: %016lx RCX: %016lx\n",
82 regs->ax, regs->bx, regs->cx);
83 printk(KERN_DEFAULT "RDX: %016lx RSI: %016lx RDI: %016lx\n",
84 regs->dx, regs->si, regs->di);
85 printk(KERN_DEFAULT "RBP: %016lx R08: %016lx R09: %016lx\n",
86 regs->bp, regs->r8, regs->r9);
87 printk(KERN_DEFAULT "R10: %016lx R11: %016lx R12: %016lx\n",
88 regs->r10, regs->r11, regs->r12);
89 printk(KERN_DEFAULT "R13: %016lx R14: %016lx R15: %016lx\n",
90 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(KERN_DEFAULT "FS: %016lx GS: %016lx\n",
99 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(KERN_DEFAULT "FS: %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
118 fs, fsindex, gs, gsindex, shadowgs);
119 printk(KERN_DEFAULT "CS: %04lx DS: %04x ES: %04x CR0: %016lx\n", regs->cs, ds,
120 es, cr0);
121 printk(KERN_DEFAULT "CR2: %016lx CR3: %016lx CR4: %016lx\n", cr2, cr3,
122 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(KERN_DEFAULT "DR0: %016lx DR1: %016lx DR2: %016lx\n",
135 d0, d1, d2);
136 printk(KERN_DEFAULT "DR3: %016lx DR6: %016lx DR7: %016lx\n",
137 d3, d6, d7);
138 }
139
140 if (boot_cpu_has(X86_FEATURE_OSPKE))
141 printk(KERN_DEFAULT "PKRU: %08x\n", 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 * Saves the FS or GS base for an outgoing thread if FSGSBASE extensions are
156 * not available. The goal is to be reasonably fast on non-FSGSBASE systems.
157 * It's forcibly inlined because it'll generate better code and this function
158 * is hot.
159 */
160static __always_inline void save_base_legacy(struct task_struct *prev_p,
161 unsigned short selector,
162 enum which_selector which)
163{
164 if (likely(selector == 0)) {
165 /*
166 * On Intel (without X86_BUG_NULL_SEG), the segment base could
167 * be the pre-existing saved base or it could be zero. On AMD
168 * (with X86_BUG_NULL_SEG), the segment base could be almost
169 * anything.
170 *
171 * This branch is very hot (it's hit twice on almost every
172 * context switch between 64-bit programs), and avoiding
173 * the RDMSR helps a lot, so we just assume that whatever
174 * value is already saved is correct. This matches historical
175 * Linux behavior, so it won't break existing applications.
176 *
177 * To avoid leaking state, on non-X86_BUG_NULL_SEG CPUs, if we
178 * report that the base is zero, it needs to actually be zero:
179 * see the corresponding logic in load_seg_legacy.
180 */
181 } else {
182 /*
183 * If the selector is 1, 2, or 3, then the base is zero on
184 * !X86_BUG_NULL_SEG CPUs and could be anything on
185 * X86_BUG_NULL_SEG CPUs. In the latter case, Linux
186 * has never attempted to preserve the base across context
187 * switches.
188 *
189 * If selector > 3, then it refers to a real segment, and
190 * saving the base isn't necessary.
191 */
192 if (which == FS)
193 prev_p->thread.fsbase = 0;
194 else
195 prev_p->thread.gsbase = 0;
196 }
197}
198
199static __always_inline void save_fsgs(struct task_struct *task)
200{
201 savesegment(fs, task->thread.fsindex);
202 savesegment(gs, task->thread.gsindex);
203 save_base_legacy(task, task->thread.fsindex, FS);
204 save_base_legacy(task, task->thread.gsindex, GS);
205}
206
207#if IS_ENABLED(CONFIG_KVM)
208/*
209 * While a process is running,current->thread.fsbase and current->thread.gsbase
210 * may not match the corresponding CPU registers (see save_base_legacy()). KVM
211 * wants an efficient way to save and restore FSBASE and GSBASE.
212 * When FSGSBASE extensions are enabled, this will have to use RD{FS,GS}BASE.
213 */
214void save_fsgs_for_kvm(void)
215{
216 save_fsgs(current);
217}
218EXPORT_SYMBOL_GPL(save_fsgs_for_kvm);
219#endif
220
221static __always_inline void loadseg(enum which_selector which,
222 unsigned short sel)
223{
224 if (which == FS)
225 loadsegment(fs, sel);
226 else
227 load_gs_index(sel);
228}
229
230static __always_inline void load_seg_legacy(unsigned short prev_index,
231 unsigned long prev_base,
232 unsigned short next_index,
233 unsigned long next_base,
234 enum which_selector which)
235{
236 if (likely(next_index <= 3)) {
237 /*
238 * The next task is using 64-bit TLS, is not using this
239 * segment at all, or is having fun with arcane CPU features.
240 */
241 if (next_base == 0) {
242 /*
243 * Nasty case: on AMD CPUs, we need to forcibly zero
244 * the base.
245 */
246 if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
247 loadseg(which, __USER_DS);
248 loadseg(which, next_index);
249 } else {
250 /*
251 * We could try to exhaustively detect cases
252 * under which we can skip the segment load,
253 * but there's really only one case that matters
254 * for performance: if both the previous and
255 * next states are fully zeroed, we can skip
256 * the load.
257 *
258 * (This assumes that prev_base == 0 has no
259 * false positives. This is the case on
260 * Intel-style CPUs.)
261 */
262 if (likely(prev_index | next_index | prev_base))
263 loadseg(which, next_index);
264 }
265 } else {
266 if (prev_index != next_index)
267 loadseg(which, next_index);
268 wrmsrl(which == FS ? MSR_FS_BASE : MSR_KERNEL_GS_BASE,
269 next_base);
270 }
271 } else {
272 /*
273 * The next task is using a real segment. Loading the selector
274 * is sufficient.
275 */
276 loadseg(which, next_index);
277 }
278}
279
280static __always_inline void x86_fsgsbase_load(struct thread_struct *prev,
281 struct thread_struct *next)
282{
283 load_seg_legacy(prev->fsindex, prev->fsbase,
284 next->fsindex, next->fsbase, FS);
285 load_seg_legacy(prev->gsindex, prev->gsbase,
286 next->gsindex, next->gsbase, GS);
287}
288
289static unsigned long x86_fsgsbase_read_task(struct task_struct *task,
290 unsigned short selector)
291{
292 unsigned short idx = selector >> 3;
293 unsigned long base;
294
295 if (likely((selector & SEGMENT_TI_MASK) == 0)) {
296 if (unlikely(idx >= GDT_ENTRIES))
297 return 0;
298
299 /*
300 * There are no user segments in the GDT with nonzero bases
301 * other than the TLS segments.
302 */
303 if (idx < GDT_ENTRY_TLS_MIN || idx > GDT_ENTRY_TLS_MAX)
304 return 0;
305
306 idx -= GDT_ENTRY_TLS_MIN;
307 base = get_desc_base(&task->thread.tls_array[idx]);
308 } else {
309#ifdef CONFIG_MODIFY_LDT_SYSCALL
310 struct ldt_struct *ldt;
311
312 /*
313 * If performance here mattered, we could protect the LDT
314 * with RCU. This is a slow path, though, so we can just
315 * take the mutex.
316 */
317 mutex_lock(&task->mm->context.lock);
318 ldt = task->mm->context.ldt;
319 if (unlikely(idx >= ldt->nr_entries))
320 base = 0;
321 else
322 base = get_desc_base(ldt->entries + idx);
323 mutex_unlock(&task->mm->context.lock);
324#else
325 base = 0;
326#endif
327 }
328
329 return base;
330}
331
332unsigned long x86_fsbase_read_task(struct task_struct *task)
333{
334 unsigned long fsbase;
335
336 if (task == current)
337 fsbase = x86_fsbase_read_cpu();
338 else if (task->thread.fsindex == 0)
339 fsbase = task->thread.fsbase;
340 else
341 fsbase = x86_fsgsbase_read_task(task, task->thread.fsindex);
342
343 return fsbase;
344}
345
346unsigned long x86_gsbase_read_task(struct task_struct *task)
347{
348 unsigned long gsbase;
349
350 if (task == current)
351 gsbase = x86_gsbase_read_cpu_inactive();
352 else if (task->thread.gsindex == 0)
353 gsbase = task->thread.gsbase;
354 else
355 gsbase = x86_fsgsbase_read_task(task, task->thread.gsindex);
356
357 return gsbase;
358}
359
360void x86_fsbase_write_task(struct task_struct *task, unsigned long fsbase)
361{
362 WARN_ON_ONCE(task == current);
363
364 task->thread.fsbase = fsbase;
365}
366
367void x86_gsbase_write_task(struct task_struct *task, unsigned long gsbase)
368{
369 WARN_ON_ONCE(task == current);
370
371 task->thread.gsbase = gsbase;
372}
373
374int copy_thread_tls(unsigned long clone_flags, unsigned long sp,
375 unsigned long arg, struct task_struct *p, unsigned long tls)
376{
377 int err;
378 struct pt_regs *childregs;
379 struct fork_frame *fork_frame;
380 struct inactive_task_frame *frame;
381 struct task_struct *me = current;
382
383 childregs = task_pt_regs(p);
384 fork_frame = container_of(childregs, struct fork_frame, regs);
385 frame = &fork_frame->frame;
386
387 frame->bp = 0;
388 frame->ret_addr = (unsigned long) ret_from_fork;
389 p->thread.sp = (unsigned long) fork_frame;
390 p->thread.io_bitmap_ptr = NULL;
391
392 savesegment(gs, p->thread.gsindex);
393 p->thread.gsbase = p->thread.gsindex ? 0 : me->thread.gsbase;
394 savesegment(fs, p->thread.fsindex);
395 p->thread.fsbase = p->thread.fsindex ? 0 : me->thread.fsbase;
396 savesegment(es, p->thread.es);
397 savesegment(ds, p->thread.ds);
398 memset(p->thread.ptrace_bps, 0, sizeof(p->thread.ptrace_bps));
399
400 if (unlikely(p->flags & PF_KTHREAD)) {
401 /* kernel thread */
402 memset(childregs, 0, sizeof(struct pt_regs));
403 frame->bx = sp; /* function */
404 frame->r12 = arg;
405 return 0;
406 }
407 frame->bx = 0;
408 *childregs = *current_pt_regs();
409
410 childregs->ax = 0;
411 if (sp)
412 childregs->sp = sp;
413
414 err = -ENOMEM;
415 if (unlikely(test_tsk_thread_flag(me, TIF_IO_BITMAP))) {
416 p->thread.io_bitmap_ptr = kmemdup(me->thread.io_bitmap_ptr,
417 IO_BITMAP_BYTES, GFP_KERNEL);
418 if (!p->thread.io_bitmap_ptr) {
419 p->thread.io_bitmap_max = 0;
420 return -ENOMEM;
421 }
422 set_tsk_thread_flag(p, TIF_IO_BITMAP);
423 }
424
425 /*
426 * Set a new TLS for the child thread?
427 */
428 if (clone_flags & CLONE_SETTLS) {
429#ifdef CONFIG_IA32_EMULATION
430 if (in_ia32_syscall())
431 err = do_set_thread_area(p, -1,
432 (struct user_desc __user *)tls, 0);
433 else
434#endif
435 err = do_arch_prctl_64(p, ARCH_SET_FS, tls);
436 if (err)
437 goto out;
438 }
439 err = 0;
440out:
441 if (err && p->thread.io_bitmap_ptr) {
442 kfree(p->thread.io_bitmap_ptr);
443 p->thread.io_bitmap_max = 0;
444 }
445
446 return err;
447}
448
449static void
450start_thread_common(struct pt_regs *regs, unsigned long new_ip,
451 unsigned long new_sp,
452 unsigned int _cs, unsigned int _ss, unsigned int _ds)
453{
454 WARN_ON_ONCE(regs != current_pt_regs());
455
456 if (static_cpu_has(X86_BUG_NULL_SEG)) {
457 /* Loading zero below won't clear the base. */
458 loadsegment(fs, __USER_DS);
459 load_gs_index(__USER_DS);
460 }
461
462 loadsegment(fs, 0);
463 loadsegment(es, _ds);
464 loadsegment(ds, _ds);
465 load_gs_index(0);
466
467 regs->ip = new_ip;
468 regs->sp = new_sp;
469 regs->cs = _cs;
470 regs->ss = _ss;
471 regs->flags = X86_EFLAGS_IF;
472 force_iret();
473}
474
475void
476start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
477{
478 start_thread_common(regs, new_ip, new_sp,
479 __USER_CS, __USER_DS, 0);
480}
481EXPORT_SYMBOL_GPL(start_thread);
482
483#ifdef CONFIG_COMPAT
484void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp)
485{
486 start_thread_common(regs, new_ip, new_sp,
487 test_thread_flag(TIF_X32)
488 ? __USER_CS : __USER32_CS,
489 __USER_DS, __USER_DS);
490}
491#endif
492
493/*
494 * switch_to(x,y) should switch tasks from x to y.
495 *
496 * This could still be optimized:
497 * - fold all the options into a flag word and test it with a single test.
498 * - could test fs/gs bitsliced
499 *
500 * Kprobes not supported here. Set the probe on schedule instead.
501 * Function graph tracer not supported too.
502 */
503__visible __notrace_funcgraph struct task_struct *
504__switch_to(struct task_struct *prev_p, struct task_struct *next_p)
505{
506 struct thread_struct *prev = &prev_p->thread;
507 struct thread_struct *next = &next_p->thread;
508 struct fpu *prev_fpu = &prev->fpu;
509 struct fpu *next_fpu = &next->fpu;
510 int cpu = smp_processor_id();
511
512 WARN_ON_ONCE(IS_ENABLED(CONFIG_DEBUG_ENTRY) &&
513 this_cpu_read(irq_count) != -1);
514
515 if (!test_thread_flag(TIF_NEED_FPU_LOAD))
516 switch_fpu_prepare(prev_fpu, cpu);
517
518 /* We must save %fs and %gs before load_TLS() because
519 * %fs and %gs may be cleared by load_TLS().
520 *
521 * (e.g. xen_load_tls())
522 */
523 save_fsgs(prev_p);
524
525 /*
526 * Load TLS before restoring any segments so that segment loads
527 * reference the correct GDT entries.
528 */
529 load_TLS(next, cpu);
530
531 /*
532 * Leave lazy mode, flushing any hypercalls made here. This
533 * must be done after loading TLS entries in the GDT but before
534 * loading segments that might reference them.
535 */
536 arch_end_context_switch(next_p);
537
538 /* Switch DS and ES.
539 *
540 * Reading them only returns the selectors, but writing them (if
541 * nonzero) loads the full descriptor from the GDT or LDT. The
542 * LDT for next is loaded in switch_mm, and the GDT is loaded
543 * above.
544 *
545 * We therefore need to write new values to the segment
546 * registers on every context switch unless both the new and old
547 * values are zero.
548 *
549 * Note that we don't need to do anything for CS and SS, as
550 * those are saved and restored as part of pt_regs.
551 */
552 savesegment(es, prev->es);
553 if (unlikely(next->es | prev->es))
554 loadsegment(es, next->es);
555
556 savesegment(ds, prev->ds);
557 if (unlikely(next->ds | prev->ds))
558 loadsegment(ds, next->ds);
559
560 x86_fsgsbase_load(prev, next);
561
562 /*
563 * Switch the PDA and FPU contexts.
564 */
565 this_cpu_write(current_task, next_p);
566 this_cpu_write(cpu_current_top_of_stack, task_top_of_stack(next_p));
567
568 switch_fpu_finish(next_fpu);
569
570 /* Reload sp0. */
571 update_task_stack(next_p);
572
573 switch_to_extra(prev_p, next_p);
574
575#ifdef CONFIG_XEN_PV
576 /*
577 * On Xen PV, IOPL bits in pt_regs->flags have no effect, and
578 * current_pt_regs()->flags may not match the current task's
579 * intended IOPL. We need to switch it manually.
580 */
581 if (unlikely(static_cpu_has(X86_FEATURE_XENPV) &&
582 prev->iopl != next->iopl))
583 xen_set_iopl_mask(next->iopl);
584#endif
585
586 if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
587 /*
588 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
589 * does not update the cached descriptor. As a result, if we
590 * do SYSRET while SS is NULL, we'll end up in user mode with
591 * SS apparently equal to __USER_DS but actually unusable.
592 *
593 * The straightforward workaround would be to fix it up just
594 * before SYSRET, but that would slow down the system call
595 * fast paths. Instead, we ensure that SS is never NULL in
596 * system call context. We do this by replacing NULL SS
597 * selectors at every context switch. SYSCALL sets up a valid
598 * SS, so the only way to get NULL is to re-enter the kernel
599 * from CPL 3 through an interrupt. Since that can't happen
600 * in the same task as a running syscall, we are guaranteed to
601 * context switch between every interrupt vector entry and a
602 * subsequent SYSRET.
603 *
604 * We read SS first because SS reads are much faster than
605 * writes. Out of caution, we force SS to __KERNEL_DS even if
606 * it previously had a different non-NULL value.
607 */
608 unsigned short ss_sel;
609 savesegment(ss, ss_sel);
610 if (ss_sel != __KERNEL_DS)
611 loadsegment(ss, __KERNEL_DS);
612 }
613
614 /* Load the Intel cache allocation PQR MSR. */
615 resctrl_sched_in();
616
617 return prev_p;
618}
619
620void set_personality_64bit(void)
621{
622 /* inherit personality from parent */
623
624 /* Make sure to be in 64bit mode */
625 clear_thread_flag(TIF_IA32);
626 clear_thread_flag(TIF_ADDR32);
627 clear_thread_flag(TIF_X32);
628 /* Pretend that this comes from a 64bit execve */
629 task_pt_regs(current)->orig_ax = __NR_execve;
630 current_thread_info()->status &= ~TS_COMPAT;
631
632 /* Ensure the corresponding mm is not marked. */
633 if (current->mm)
634 current->mm->context.ia32_compat = 0;
635
636 /* TBD: overwrites user setup. Should have two bits.
637 But 64bit processes have always behaved this way,
638 so it's not too bad. The main problem is just that
639 32bit children are affected again. */
640 current->personality &= ~READ_IMPLIES_EXEC;
641}
642
643static void __set_personality_x32(void)
644{
645#ifdef CONFIG_X86_X32
646 clear_thread_flag(TIF_IA32);
647 set_thread_flag(TIF_X32);
648 if (current->mm)
649 current->mm->context.ia32_compat = TIF_X32;
650 current->personality &= ~READ_IMPLIES_EXEC;
651 /*
652 * in_32bit_syscall() uses the presence of the x32 syscall bit
653 * flag to determine compat status. The x86 mmap() code relies on
654 * the syscall bitness so set x32 syscall bit right here to make
655 * in_32bit_syscall() work during exec().
656 *
657 * Pretend to come from a x32 execve.
658 */
659 task_pt_regs(current)->orig_ax = __NR_x32_execve | __X32_SYSCALL_BIT;
660 current_thread_info()->status &= ~TS_COMPAT;
661#endif
662}
663
664static void __set_personality_ia32(void)
665{
666#ifdef CONFIG_IA32_EMULATION
667 set_thread_flag(TIF_IA32);
668 clear_thread_flag(TIF_X32);
669 if (current->mm)
670 current->mm->context.ia32_compat = TIF_IA32;
671 current->personality |= force_personality32;
672 /* Prepare the first "return" to user space */
673 task_pt_regs(current)->orig_ax = __NR_ia32_execve;
674 current_thread_info()->status |= TS_COMPAT;
675#endif
676}
677
678void set_personality_ia32(bool x32)
679{
680 /* Make sure to be in 32bit mode */
681 set_thread_flag(TIF_ADDR32);
682
683 if (x32)
684 __set_personality_x32();
685 else
686 __set_personality_ia32();
687}
688EXPORT_SYMBOL_GPL(set_personality_ia32);
689
690#ifdef CONFIG_CHECKPOINT_RESTORE
691static long prctl_map_vdso(const struct vdso_image *image, unsigned long addr)
692{
693 int ret;
694
695 ret = map_vdso_once(image, addr);
696 if (ret)
697 return ret;
698
699 return (long)image->size;
700}
701#endif
702
703long do_arch_prctl_64(struct task_struct *task, int option, unsigned long arg2)
704{
705 int ret = 0;
706
707 switch (option) {
708 case ARCH_SET_GS: {
709 if (unlikely(arg2 >= TASK_SIZE_MAX))
710 return -EPERM;
711
712 preempt_disable();
713 /*
714 * ARCH_SET_GS has always overwritten the index
715 * and the base. Zero is the most sensible value
716 * to put in the index, and is the only value that
717 * makes any sense if FSGSBASE is unavailable.
718 */
719 if (task == current) {
720 loadseg(GS, 0);
721 x86_gsbase_write_cpu_inactive(arg2);
722
723 /*
724 * On non-FSGSBASE systems, save_base_legacy() expects
725 * that we also fill in thread.gsbase.
726 */
727 task->thread.gsbase = arg2;
728
729 } else {
730 task->thread.gsindex = 0;
731 x86_gsbase_write_task(task, arg2);
732 }
733 preempt_enable();
734 break;
735 }
736 case ARCH_SET_FS: {
737 /*
738 * Not strictly needed for %fs, but do it for symmetry
739 * with %gs
740 */
741 if (unlikely(arg2 >= TASK_SIZE_MAX))
742 return -EPERM;
743
744 preempt_disable();
745 /*
746 * Set the selector to 0 for the same reason
747 * as %gs above.
748 */
749 if (task == current) {
750 loadseg(FS, 0);
751 x86_fsbase_write_cpu(arg2);
752
753 /*
754 * On non-FSGSBASE systems, save_base_legacy() expects
755 * that we also fill in thread.fsbase.
756 */
757 task->thread.fsbase = arg2;
758 } else {
759 task->thread.fsindex = 0;
760 x86_fsbase_write_task(task, arg2);
761 }
762 preempt_enable();
763 break;
764 }
765 case ARCH_GET_FS: {
766 unsigned long base = x86_fsbase_read_task(task);
767
768 ret = put_user(base, (unsigned long __user *)arg2);
769 break;
770 }
771 case ARCH_GET_GS: {
772 unsigned long base = x86_gsbase_read_task(task);
773
774 ret = put_user(base, (unsigned long __user *)arg2);
775 break;
776 }
777
778#ifdef CONFIG_CHECKPOINT_RESTORE
779# ifdef CONFIG_X86_X32_ABI
780 case ARCH_MAP_VDSO_X32:
781 return prctl_map_vdso(&vdso_image_x32, arg2);
782# endif
783# if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
784 case ARCH_MAP_VDSO_32:
785 return prctl_map_vdso(&vdso_image_32, arg2);
786# endif
787 case ARCH_MAP_VDSO_64:
788 return prctl_map_vdso(&vdso_image_64, arg2);
789#endif
790
791 default:
792 ret = -EINVAL;
793 break;
794 }
795
796 return ret;
797}
798
799SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
800{
801 long ret;
802
803 ret = do_arch_prctl_64(current, option, arg2);
804 if (ret == -EINVAL)
805 ret = do_arch_prctl_common(current, option, arg2);
806
807 return ret;
808}
809
810#ifdef CONFIG_IA32_EMULATION
811COMPAT_SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
812{
813 return do_arch_prctl_common(current, option, arg2);
814}
815#endif
816
817unsigned long KSTK_ESP(struct task_struct *task)
818{
819 return task_pt_regs(task)->sp;
820}
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}