Loading...
1/*
2 * Copyright (C) 1995 Linus Torvalds
3 *
4 * Pentium III FXSR, SSE support
5 * Gareth Hughes <gareth@valinux.com>, May 2000
6 *
7 * X86-64 port
8 * Andi Kleen.
9 *
10 * CPU hotplug support - ashok.raj@intel.com
11 */
12
13/*
14 * This file handles the architecture-dependent parts of process handling..
15 */
16
17#include <linux/cpu.h>
18#include <linux/errno.h>
19#include <linux/sched.h>
20#include <linux/fs.h>
21#include <linux/kernel.h>
22#include <linux/mm.h>
23#include <linux/elfcore.h>
24#include <linux/smp.h>
25#include <linux/slab.h>
26#include <linux/user.h>
27#include <linux/interrupt.h>
28#include <linux/delay.h>
29#include <linux/module.h>
30#include <linux/ptrace.h>
31#include <linux/notifier.h>
32#include <linux/kprobes.h>
33#include <linux/kdebug.h>
34#include <linux/prctl.h>
35#include <linux/uaccess.h>
36#include <linux/io.h>
37#include <linux/ftrace.h>
38
39#include <asm/pgtable.h>
40#include <asm/processor.h>
41#include <asm/i387.h>
42#include <asm/fpu-internal.h>
43#include <asm/mmu_context.h>
44#include <asm/prctl.h>
45#include <asm/desc.h>
46#include <asm/proto.h>
47#include <asm/ia32.h>
48#include <asm/idle.h>
49#include <asm/syscalls.h>
50#include <asm/debugreg.h>
51#include <asm/switch_to.h>
52
53asmlinkage extern void ret_from_fork(void);
54
55__visible DEFINE_PER_CPU(unsigned long, old_rsp);
56
57/* Prints also some state that isn't saved in the pt_regs */
58void __show_regs(struct pt_regs *regs, int all)
59{
60 unsigned long cr0 = 0L, cr2 = 0L, cr3 = 0L, cr4 = 0L, fs, gs, shadowgs;
61 unsigned long d0, d1, d2, d3, d6, d7;
62 unsigned int fsindex, gsindex;
63 unsigned int ds, cs, es;
64
65 printk(KERN_DEFAULT "RIP: %04lx:[<%016lx>] ", regs->cs & 0xffff, regs->ip);
66 printk_address(regs->ip);
67 printk(KERN_DEFAULT "RSP: %04lx:%016lx EFLAGS: %08lx\n", regs->ss,
68 regs->sp, regs->flags);
69 printk(KERN_DEFAULT "RAX: %016lx RBX: %016lx RCX: %016lx\n",
70 regs->ax, regs->bx, regs->cx);
71 printk(KERN_DEFAULT "RDX: %016lx RSI: %016lx RDI: %016lx\n",
72 regs->dx, regs->si, regs->di);
73 printk(KERN_DEFAULT "RBP: %016lx R08: %016lx R09: %016lx\n",
74 regs->bp, regs->r8, regs->r9);
75 printk(KERN_DEFAULT "R10: %016lx R11: %016lx R12: %016lx\n",
76 regs->r10, regs->r11, regs->r12);
77 printk(KERN_DEFAULT "R13: %016lx R14: %016lx R15: %016lx\n",
78 regs->r13, regs->r14, regs->r15);
79
80 asm("movl %%ds,%0" : "=r" (ds));
81 asm("movl %%cs,%0" : "=r" (cs));
82 asm("movl %%es,%0" : "=r" (es));
83 asm("movl %%fs,%0" : "=r" (fsindex));
84 asm("movl %%gs,%0" : "=r" (gsindex));
85
86 rdmsrl(MSR_FS_BASE, fs);
87 rdmsrl(MSR_GS_BASE, gs);
88 rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
89
90 if (!all)
91 return;
92
93 cr0 = read_cr0();
94 cr2 = read_cr2();
95 cr3 = read_cr3();
96 cr4 = read_cr4();
97
98 printk(KERN_DEFAULT "FS: %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
99 fs, fsindex, gs, gsindex, shadowgs);
100 printk(KERN_DEFAULT "CS: %04x DS: %04x ES: %04x CR0: %016lx\n", cs, ds,
101 es, cr0);
102 printk(KERN_DEFAULT "CR2: %016lx CR3: %016lx CR4: %016lx\n", cr2, cr3,
103 cr4);
104
105 get_debugreg(d0, 0);
106 get_debugreg(d1, 1);
107 get_debugreg(d2, 2);
108 get_debugreg(d3, 3);
109 get_debugreg(d6, 6);
110 get_debugreg(d7, 7);
111
112 /* Only print out debug registers if they are in their non-default state. */
113 if ((d0 == 0) && (d1 == 0) && (d2 == 0) && (d3 == 0) &&
114 (d6 == DR6_RESERVED) && (d7 == 0x400))
115 return;
116
117 printk(KERN_DEFAULT "DR0: %016lx DR1: %016lx DR2: %016lx\n", d0, d1, d2);
118 printk(KERN_DEFAULT "DR3: %016lx DR6: %016lx DR7: %016lx\n", d3, d6, d7);
119
120}
121
122void release_thread(struct task_struct *dead_task)
123{
124 if (dead_task->mm) {
125 if (dead_task->mm->context.size) {
126 pr_warn("WARNING: dead process %s still has LDT? <%p/%d>\n",
127 dead_task->comm,
128 dead_task->mm->context.ldt,
129 dead_task->mm->context.size);
130 BUG();
131 }
132 }
133}
134
135static inline void set_32bit_tls(struct task_struct *t, int tls, u32 addr)
136{
137 struct user_desc ud = {
138 .base_addr = addr,
139 .limit = 0xfffff,
140 .seg_32bit = 1,
141 .limit_in_pages = 1,
142 .useable = 1,
143 };
144 struct desc_struct *desc = t->thread.tls_array;
145 desc += tls;
146 fill_ldt(desc, &ud);
147}
148
149static inline u32 read_32bit_tls(struct task_struct *t, int tls)
150{
151 return get_desc_base(&t->thread.tls_array[tls]);
152}
153
154int copy_thread(unsigned long clone_flags, unsigned long sp,
155 unsigned long arg, struct task_struct *p)
156{
157 int err;
158 struct pt_regs *childregs;
159 struct task_struct *me = current;
160
161 p->thread.sp0 = (unsigned long)task_stack_page(p) + THREAD_SIZE;
162 childregs = task_pt_regs(p);
163 p->thread.sp = (unsigned long) childregs;
164 p->thread.usersp = me->thread.usersp;
165 set_tsk_thread_flag(p, TIF_FORK);
166 p->thread.fpu_counter = 0;
167 p->thread.io_bitmap_ptr = NULL;
168
169 savesegment(gs, p->thread.gsindex);
170 p->thread.gs = p->thread.gsindex ? 0 : me->thread.gs;
171 savesegment(fs, p->thread.fsindex);
172 p->thread.fs = p->thread.fsindex ? 0 : me->thread.fs;
173 savesegment(es, p->thread.es);
174 savesegment(ds, p->thread.ds);
175 memset(p->thread.ptrace_bps, 0, sizeof(p->thread.ptrace_bps));
176
177 if (unlikely(p->flags & PF_KTHREAD)) {
178 /* kernel thread */
179 memset(childregs, 0, sizeof(struct pt_regs));
180 childregs->sp = (unsigned long)childregs;
181 childregs->ss = __KERNEL_DS;
182 childregs->bx = sp; /* function */
183 childregs->bp = arg;
184 childregs->orig_ax = -1;
185 childregs->cs = __KERNEL_CS | get_kernel_rpl();
186 childregs->flags = X86_EFLAGS_IF | X86_EFLAGS_FIXED;
187 return 0;
188 }
189 *childregs = *current_pt_regs();
190
191 childregs->ax = 0;
192 if (sp)
193 childregs->sp = sp;
194
195 err = -ENOMEM;
196 memset(p->thread.ptrace_bps, 0, sizeof(p->thread.ptrace_bps));
197
198 if (unlikely(test_tsk_thread_flag(me, TIF_IO_BITMAP))) {
199 p->thread.io_bitmap_ptr = kmemdup(me->thread.io_bitmap_ptr,
200 IO_BITMAP_BYTES, GFP_KERNEL);
201 if (!p->thread.io_bitmap_ptr) {
202 p->thread.io_bitmap_max = 0;
203 return -ENOMEM;
204 }
205 set_tsk_thread_flag(p, TIF_IO_BITMAP);
206 }
207
208 /*
209 * Set a new TLS for the child thread?
210 */
211 if (clone_flags & CLONE_SETTLS) {
212#ifdef CONFIG_IA32_EMULATION
213 if (test_thread_flag(TIF_IA32))
214 err = do_set_thread_area(p, -1,
215 (struct user_desc __user *)childregs->si, 0);
216 else
217#endif
218 err = do_arch_prctl(p, ARCH_SET_FS, childregs->r8);
219 if (err)
220 goto out;
221 }
222 err = 0;
223out:
224 if (err && p->thread.io_bitmap_ptr) {
225 kfree(p->thread.io_bitmap_ptr);
226 p->thread.io_bitmap_max = 0;
227 }
228
229 return err;
230}
231
232static void
233start_thread_common(struct pt_regs *regs, unsigned long new_ip,
234 unsigned long new_sp,
235 unsigned int _cs, unsigned int _ss, unsigned int _ds)
236{
237 loadsegment(fs, 0);
238 loadsegment(es, _ds);
239 loadsegment(ds, _ds);
240 load_gs_index(0);
241 current->thread.usersp = new_sp;
242 regs->ip = new_ip;
243 regs->sp = new_sp;
244 this_cpu_write(old_rsp, new_sp);
245 regs->cs = _cs;
246 regs->ss = _ss;
247 regs->flags = X86_EFLAGS_IF;
248}
249
250void
251start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
252{
253 start_thread_common(regs, new_ip, new_sp,
254 __USER_CS, __USER_DS, 0);
255}
256
257#ifdef CONFIG_IA32_EMULATION
258void start_thread_ia32(struct pt_regs *regs, u32 new_ip, u32 new_sp)
259{
260 start_thread_common(regs, new_ip, new_sp,
261 test_thread_flag(TIF_X32)
262 ? __USER_CS : __USER32_CS,
263 __USER_DS, __USER_DS);
264}
265#endif
266
267/*
268 * switch_to(x,y) should switch tasks from x to y.
269 *
270 * This could still be optimized:
271 * - fold all the options into a flag word and test it with a single test.
272 * - could test fs/gs bitsliced
273 *
274 * Kprobes not supported here. Set the probe on schedule instead.
275 * Function graph tracer not supported too.
276 */
277__visible __notrace_funcgraph struct task_struct *
278__switch_to(struct task_struct *prev_p, struct task_struct *next_p)
279{
280 struct thread_struct *prev = &prev_p->thread;
281 struct thread_struct *next = &next_p->thread;
282 int cpu = smp_processor_id();
283 struct tss_struct *tss = &per_cpu(init_tss, cpu);
284 unsigned fsindex, gsindex;
285 fpu_switch_t fpu;
286
287 fpu = switch_fpu_prepare(prev_p, next_p, cpu);
288
289 /*
290 * Reload esp0, LDT and the page table pointer:
291 */
292 load_sp0(tss, next);
293
294 /*
295 * Switch DS and ES.
296 * This won't pick up thread selector changes, but I guess that is ok.
297 */
298 savesegment(es, prev->es);
299 if (unlikely(next->es | prev->es))
300 loadsegment(es, next->es);
301
302 savesegment(ds, prev->ds);
303 if (unlikely(next->ds | prev->ds))
304 loadsegment(ds, next->ds);
305
306
307 /* We must save %fs and %gs before load_TLS() because
308 * %fs and %gs may be cleared by load_TLS().
309 *
310 * (e.g. xen_load_tls())
311 */
312 savesegment(fs, fsindex);
313 savesegment(gs, gsindex);
314
315 load_TLS(next, cpu);
316
317 /*
318 * Leave lazy mode, flushing any hypercalls made here.
319 * This must be done before restoring TLS segments so
320 * the GDT and LDT are properly updated, and must be
321 * done before math_state_restore, so the TS bit is up
322 * to date.
323 */
324 arch_end_context_switch(next_p);
325
326 /*
327 * Switch FS and GS.
328 *
329 * Segment register != 0 always requires a reload. Also
330 * reload when it has changed. When prev process used 64bit
331 * base always reload to avoid an information leak.
332 */
333 if (unlikely(fsindex | next->fsindex | prev->fs)) {
334 loadsegment(fs, next->fsindex);
335 /*
336 * Check if the user used a selector != 0; if yes
337 * clear 64bit base, since overloaded base is always
338 * mapped to the Null selector
339 */
340 if (fsindex)
341 prev->fs = 0;
342 }
343 /* when next process has a 64bit base use it */
344 if (next->fs)
345 wrmsrl(MSR_FS_BASE, next->fs);
346 prev->fsindex = fsindex;
347
348 if (unlikely(gsindex | next->gsindex | prev->gs)) {
349 load_gs_index(next->gsindex);
350 if (gsindex)
351 prev->gs = 0;
352 }
353 if (next->gs)
354 wrmsrl(MSR_KERNEL_GS_BASE, next->gs);
355 prev->gsindex = gsindex;
356
357 switch_fpu_finish(next_p, fpu);
358
359 /*
360 * Switch the PDA and FPU contexts.
361 */
362 prev->usersp = this_cpu_read(old_rsp);
363 this_cpu_write(old_rsp, next->usersp);
364 this_cpu_write(current_task, next_p);
365
366 /*
367 * If it were not for PREEMPT_ACTIVE we could guarantee that the
368 * preempt_count of all tasks was equal here and this would not be
369 * needed.
370 */
371 task_thread_info(prev_p)->saved_preempt_count = this_cpu_read(__preempt_count);
372 this_cpu_write(__preempt_count, task_thread_info(next_p)->saved_preempt_count);
373
374 this_cpu_write(kernel_stack,
375 (unsigned long)task_stack_page(next_p) +
376 THREAD_SIZE - KERNEL_STACK_OFFSET);
377
378 /*
379 * Now maybe reload the debug registers and handle I/O bitmaps
380 */
381 if (unlikely(task_thread_info(next_p)->flags & _TIF_WORK_CTXSW_NEXT ||
382 task_thread_info(prev_p)->flags & _TIF_WORK_CTXSW_PREV))
383 __switch_to_xtra(prev_p, next_p, tss);
384
385 return prev_p;
386}
387
388void set_personality_64bit(void)
389{
390 /* inherit personality from parent */
391
392 /* Make sure to be in 64bit mode */
393 clear_thread_flag(TIF_IA32);
394 clear_thread_flag(TIF_ADDR32);
395 clear_thread_flag(TIF_X32);
396
397 /* Ensure the corresponding mm is not marked. */
398 if (current->mm)
399 current->mm->context.ia32_compat = 0;
400
401 /* TBD: overwrites user setup. Should have two bits.
402 But 64bit processes have always behaved this way,
403 so it's not too bad. The main problem is just that
404 32bit childs are affected again. */
405 current->personality &= ~READ_IMPLIES_EXEC;
406}
407
408void set_personality_ia32(bool x32)
409{
410 /* inherit personality from parent */
411
412 /* Make sure to be in 32bit mode */
413 set_thread_flag(TIF_ADDR32);
414
415 /* Mark the associated mm as containing 32-bit tasks. */
416 if (current->mm)
417 current->mm->context.ia32_compat = 1;
418
419 if (x32) {
420 clear_thread_flag(TIF_IA32);
421 set_thread_flag(TIF_X32);
422 current->personality &= ~READ_IMPLIES_EXEC;
423 /* is_compat_task() uses the presence of the x32
424 syscall bit flag to determine compat status */
425 current_thread_info()->status &= ~TS_COMPAT;
426 } else {
427 set_thread_flag(TIF_IA32);
428 clear_thread_flag(TIF_X32);
429 current->personality |= force_personality32;
430 /* Prepare the first "return" to user space */
431 current_thread_info()->status |= TS_COMPAT;
432 }
433}
434EXPORT_SYMBOL_GPL(set_personality_ia32);
435
436unsigned long get_wchan(struct task_struct *p)
437{
438 unsigned long stack;
439 u64 fp, ip;
440 int count = 0;
441
442 if (!p || p == current || p->state == TASK_RUNNING)
443 return 0;
444 stack = (unsigned long)task_stack_page(p);
445 if (p->thread.sp < stack || p->thread.sp >= stack+THREAD_SIZE)
446 return 0;
447 fp = *(u64 *)(p->thread.sp);
448 do {
449 if (fp < (unsigned long)stack ||
450 fp >= (unsigned long)stack+THREAD_SIZE)
451 return 0;
452 ip = *(u64 *)(fp+8);
453 if (!in_sched_functions(ip))
454 return ip;
455 fp = *(u64 *)fp;
456 } while (count++ < 16);
457 return 0;
458}
459
460long do_arch_prctl(struct task_struct *task, int code, unsigned long addr)
461{
462 int ret = 0;
463 int doit = task == current;
464 int cpu;
465
466 switch (code) {
467 case ARCH_SET_GS:
468 if (addr >= TASK_SIZE_OF(task))
469 return -EPERM;
470 cpu = get_cpu();
471 /* handle small bases via the GDT because that's faster to
472 switch. */
473 if (addr <= 0xffffffff) {
474 set_32bit_tls(task, GS_TLS, addr);
475 if (doit) {
476 load_TLS(&task->thread, cpu);
477 load_gs_index(GS_TLS_SEL);
478 }
479 task->thread.gsindex = GS_TLS_SEL;
480 task->thread.gs = 0;
481 } else {
482 task->thread.gsindex = 0;
483 task->thread.gs = addr;
484 if (doit) {
485 load_gs_index(0);
486 ret = wrmsrl_safe(MSR_KERNEL_GS_BASE, addr);
487 }
488 }
489 put_cpu();
490 break;
491 case ARCH_SET_FS:
492 /* Not strictly needed for fs, but do it for symmetry
493 with gs */
494 if (addr >= TASK_SIZE_OF(task))
495 return -EPERM;
496 cpu = get_cpu();
497 /* handle small bases via the GDT because that's faster to
498 switch. */
499 if (addr <= 0xffffffff) {
500 set_32bit_tls(task, FS_TLS, addr);
501 if (doit) {
502 load_TLS(&task->thread, cpu);
503 loadsegment(fs, FS_TLS_SEL);
504 }
505 task->thread.fsindex = FS_TLS_SEL;
506 task->thread.fs = 0;
507 } else {
508 task->thread.fsindex = 0;
509 task->thread.fs = addr;
510 if (doit) {
511 /* set the selector to 0 to not confuse
512 __switch_to */
513 loadsegment(fs, 0);
514 ret = wrmsrl_safe(MSR_FS_BASE, addr);
515 }
516 }
517 put_cpu();
518 break;
519 case ARCH_GET_FS: {
520 unsigned long base;
521 if (task->thread.fsindex == FS_TLS_SEL)
522 base = read_32bit_tls(task, FS_TLS);
523 else if (doit)
524 rdmsrl(MSR_FS_BASE, base);
525 else
526 base = task->thread.fs;
527 ret = put_user(base, (unsigned long __user *)addr);
528 break;
529 }
530 case ARCH_GET_GS: {
531 unsigned long base;
532 unsigned gsindex;
533 if (task->thread.gsindex == GS_TLS_SEL)
534 base = read_32bit_tls(task, GS_TLS);
535 else if (doit) {
536 savesegment(gs, gsindex);
537 if (gsindex)
538 rdmsrl(MSR_KERNEL_GS_BASE, base);
539 else
540 base = task->thread.gs;
541 } else
542 base = task->thread.gs;
543 ret = put_user(base, (unsigned long __user *)addr);
544 break;
545 }
546
547 default:
548 ret = -EINVAL;
549 break;
550 }
551
552 return ret;
553}
554
555long sys_arch_prctl(int code, unsigned long addr)
556{
557 return do_arch_prctl(current, code, addr);
558}
559
560unsigned long KSTK_ESP(struct task_struct *task)
561{
562 return (test_tsk_thread_flag(task, TIF_IA32)) ?
563 (task_pt_regs(task)->sp) : ((task)->thread.usersp);
564}
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}