xref: /linux/arch/x86/kernel/process_64.c (revision 2ba9268dd603d23e17643437b2246acb6844953b)
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 
53 asmlinkage 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 */
58 void __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 
122 void 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 
135 static 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 
149 static inline u32 read_32bit_tls(struct task_struct *t, int tls)
150 {
151 	return get_desc_base(&t->thread.tls_array[tls]);
152 }
153 
154 int 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.io_bitmap_ptr = NULL;
167 
168 	savesegment(gs, p->thread.gsindex);
169 	p->thread.gs = p->thread.gsindex ? 0 : me->thread.gs;
170 	savesegment(fs, p->thread.fsindex);
171 	p->thread.fs = p->thread.fsindex ? 0 : me->thread.fs;
172 	savesegment(es, p->thread.es);
173 	savesegment(ds, p->thread.ds);
174 	memset(p->thread.ptrace_bps, 0, sizeof(p->thread.ptrace_bps));
175 
176 	if (unlikely(p->flags & PF_KTHREAD)) {
177 		/* kernel thread */
178 		memset(childregs, 0, sizeof(struct pt_regs));
179 		childregs->sp = (unsigned long)childregs;
180 		childregs->ss = __KERNEL_DS;
181 		childregs->bx = sp; /* function */
182 		childregs->bp = arg;
183 		childregs->orig_ax = -1;
184 		childregs->cs = __KERNEL_CS | get_kernel_rpl();
185 		childregs->flags = X86_EFLAGS_IF | X86_EFLAGS_FIXED;
186 		return 0;
187 	}
188 	*childregs = *current_pt_regs();
189 
190 	childregs->ax = 0;
191 	if (sp)
192 		childregs->sp = sp;
193 
194 	err = -ENOMEM;
195 	if (unlikely(test_tsk_thread_flag(me, TIF_IO_BITMAP))) {
196 		p->thread.io_bitmap_ptr = kmemdup(me->thread.io_bitmap_ptr,
197 						  IO_BITMAP_BYTES, GFP_KERNEL);
198 		if (!p->thread.io_bitmap_ptr) {
199 			p->thread.io_bitmap_max = 0;
200 			return -ENOMEM;
201 		}
202 		set_tsk_thread_flag(p, TIF_IO_BITMAP);
203 	}
204 
205 	/*
206 	 * Set a new TLS for the child thread?
207 	 */
208 	if (clone_flags & CLONE_SETTLS) {
209 #ifdef CONFIG_IA32_EMULATION
210 		if (test_thread_flag(TIF_IA32))
211 			err = do_set_thread_area(p, -1,
212 				(struct user_desc __user *)childregs->si, 0);
213 		else
214 #endif
215 			err = do_arch_prctl(p, ARCH_SET_FS, childregs->r8);
216 		if (err)
217 			goto out;
218 	}
219 	err = 0;
220 out:
221 	if (err && p->thread.io_bitmap_ptr) {
222 		kfree(p->thread.io_bitmap_ptr);
223 		p->thread.io_bitmap_max = 0;
224 	}
225 
226 	return err;
227 }
228 
229 static void
230 start_thread_common(struct pt_regs *regs, unsigned long new_ip,
231 		    unsigned long new_sp,
232 		    unsigned int _cs, unsigned int _ss, unsigned int _ds)
233 {
234 	loadsegment(fs, 0);
235 	loadsegment(es, _ds);
236 	loadsegment(ds, _ds);
237 	load_gs_index(0);
238 	current->thread.usersp	= new_sp;
239 	regs->ip		= new_ip;
240 	regs->sp		= new_sp;
241 	this_cpu_write(old_rsp, new_sp);
242 	regs->cs		= _cs;
243 	regs->ss		= _ss;
244 	regs->flags		= X86_EFLAGS_IF;
245 }
246 
247 void
248 start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
249 {
250 	start_thread_common(regs, new_ip, new_sp,
251 			    __USER_CS, __USER_DS, 0);
252 }
253 
254 #ifdef CONFIG_IA32_EMULATION
255 void start_thread_ia32(struct pt_regs *regs, u32 new_ip, u32 new_sp)
256 {
257 	start_thread_common(regs, new_ip, new_sp,
258 			    test_thread_flag(TIF_X32)
259 			    ? __USER_CS : __USER32_CS,
260 			    __USER_DS, __USER_DS);
261 }
262 #endif
263 
264 /*
265  *	switch_to(x,y) should switch tasks from x to y.
266  *
267  * This could still be optimized:
268  * - fold all the options into a flag word and test it with a single test.
269  * - could test fs/gs bitsliced
270  *
271  * Kprobes not supported here. Set the probe on schedule instead.
272  * Function graph tracer not supported too.
273  */
274 __visible __notrace_funcgraph struct task_struct *
275 __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
276 {
277 	struct thread_struct *prev = &prev_p->thread;
278 	struct thread_struct *next = &next_p->thread;
279 	int cpu = smp_processor_id();
280 	struct tss_struct *tss = &per_cpu(init_tss, cpu);
281 	unsigned fsindex, gsindex;
282 	fpu_switch_t fpu;
283 
284 	fpu = switch_fpu_prepare(prev_p, next_p, cpu);
285 
286 	/* Reload esp0 and ss1. */
287 	load_sp0(tss, next);
288 
289 	/* We must save %fs and %gs before load_TLS() because
290 	 * %fs and %gs may be cleared by load_TLS().
291 	 *
292 	 * (e.g. xen_load_tls())
293 	 */
294 	savesegment(fs, fsindex);
295 	savesegment(gs, gsindex);
296 
297 	/*
298 	 * Load TLS before restoring any segments so that segment loads
299 	 * reference the correct GDT entries.
300 	 */
301 	load_TLS(next, cpu);
302 
303 	/*
304 	 * Leave lazy mode, flushing any hypercalls made here.  This
305 	 * must be done after loading TLS entries in the GDT but before
306 	 * loading segments that might reference them, and and it must
307 	 * be done before math_state_restore, so the TS bit is up to
308 	 * date.
309 	 */
310 	arch_end_context_switch(next_p);
311 
312 	/* Switch DS and ES.
313 	 *
314 	 * Reading them only returns the selectors, but writing them (if
315 	 * nonzero) loads the full descriptor from the GDT or LDT.  The
316 	 * LDT for next is loaded in switch_mm, and the GDT is loaded
317 	 * above.
318 	 *
319 	 * We therefore need to write new values to the segment
320 	 * registers on every context switch unless both the new and old
321 	 * values are zero.
322 	 *
323 	 * Note that we don't need to do anything for CS and SS, as
324 	 * those are saved and restored as part of pt_regs.
325 	 */
326 	savesegment(es, prev->es);
327 	if (unlikely(next->es | prev->es))
328 		loadsegment(es, next->es);
329 
330 	savesegment(ds, prev->ds);
331 	if (unlikely(next->ds | prev->ds))
332 		loadsegment(ds, next->ds);
333 
334 	/*
335 	 * Switch FS and GS.
336 	 *
337 	 * These are even more complicated than FS and GS: they have
338 	 * 64-bit bases are that controlled by arch_prctl.  Those bases
339 	 * only differ from the values in the GDT or LDT if the selector
340 	 * is 0.
341 	 *
342 	 * Loading the segment register resets the hidden base part of
343 	 * the register to 0 or the value from the GDT / LDT.  If the
344 	 * next base address zero, writing 0 to the segment register is
345 	 * much faster than using wrmsr to explicitly zero the base.
346 	 *
347 	 * The thread_struct.fs and thread_struct.gs values are 0
348 	 * if the fs and gs bases respectively are not overridden
349 	 * from the values implied by fsindex and gsindex.  They
350 	 * are nonzero, and store the nonzero base addresses, if
351 	 * the bases are overridden.
352 	 *
353 	 * (fs != 0 && fsindex != 0) || (gs != 0 && gsindex != 0) should
354 	 * be impossible.
355 	 *
356 	 * Therefore we need to reload the segment registers if either
357 	 * the old or new selector is nonzero, and we need to override
358 	 * the base address if next thread expects it to be overridden.
359 	 *
360 	 * This code is unnecessarily slow in the case where the old and
361 	 * new indexes are zero and the new base is nonzero -- it will
362 	 * unnecessarily write 0 to the selector before writing the new
363 	 * base address.
364 	 *
365 	 * Note: This all depends on arch_prctl being the only way that
366 	 * user code can override the segment base.  Once wrfsbase and
367 	 * wrgsbase are enabled, most of this code will need to change.
368 	 */
369 	if (unlikely(fsindex | next->fsindex | prev->fs)) {
370 		loadsegment(fs, next->fsindex);
371 
372 		/*
373 		 * If user code wrote a nonzero value to FS, then it also
374 		 * cleared the overridden base address.
375 		 *
376 		 * XXX: if user code wrote 0 to FS and cleared the base
377 		 * address itself, we won't notice and we'll incorrectly
378 		 * restore the prior base address next time we reschdule
379 		 * the process.
380 		 */
381 		if (fsindex)
382 			prev->fs = 0;
383 	}
384 	if (next->fs)
385 		wrmsrl(MSR_FS_BASE, next->fs);
386 	prev->fsindex = fsindex;
387 
388 	if (unlikely(gsindex | next->gsindex | prev->gs)) {
389 		load_gs_index(next->gsindex);
390 
391 		/* This works (and fails) the same way as fsindex above. */
392 		if (gsindex)
393 			prev->gs = 0;
394 	}
395 	if (next->gs)
396 		wrmsrl(MSR_KERNEL_GS_BASE, next->gs);
397 	prev->gsindex = gsindex;
398 
399 	switch_fpu_finish(next_p, fpu);
400 
401 	/*
402 	 * Switch the PDA and FPU contexts.
403 	 */
404 	prev->usersp = this_cpu_read(old_rsp);
405 	this_cpu_write(old_rsp, next->usersp);
406 	this_cpu_write(current_task, next_p);
407 
408 	/*
409 	 * If it were not for PREEMPT_ACTIVE we could guarantee that the
410 	 * preempt_count of all tasks was equal here and this would not be
411 	 * needed.
412 	 */
413 	task_thread_info(prev_p)->saved_preempt_count = this_cpu_read(__preempt_count);
414 	this_cpu_write(__preempt_count, task_thread_info(next_p)->saved_preempt_count);
415 
416 	this_cpu_write(kernel_stack,
417 		  (unsigned long)task_stack_page(next_p) +
418 		  THREAD_SIZE - KERNEL_STACK_OFFSET);
419 
420 	/*
421 	 * Now maybe reload the debug registers and handle I/O bitmaps
422 	 */
423 	if (unlikely(task_thread_info(next_p)->flags & _TIF_WORK_CTXSW_NEXT ||
424 		     task_thread_info(prev_p)->flags & _TIF_WORK_CTXSW_PREV))
425 		__switch_to_xtra(prev_p, next_p, tss);
426 
427 	return prev_p;
428 }
429 
430 void set_personality_64bit(void)
431 {
432 	/* inherit personality from parent */
433 
434 	/* Make sure to be in 64bit mode */
435 	clear_thread_flag(TIF_IA32);
436 	clear_thread_flag(TIF_ADDR32);
437 	clear_thread_flag(TIF_X32);
438 
439 	/* Ensure the corresponding mm is not marked. */
440 	if (current->mm)
441 		current->mm->context.ia32_compat = 0;
442 
443 	/* TBD: overwrites user setup. Should have two bits.
444 	   But 64bit processes have always behaved this way,
445 	   so it's not too bad. The main problem is just that
446 	   32bit childs are affected again. */
447 	current->personality &= ~READ_IMPLIES_EXEC;
448 }
449 
450 void set_personality_ia32(bool x32)
451 {
452 	/* inherit personality from parent */
453 
454 	/* Make sure to be in 32bit mode */
455 	set_thread_flag(TIF_ADDR32);
456 
457 	/* Mark the associated mm as containing 32-bit tasks. */
458 	if (x32) {
459 		clear_thread_flag(TIF_IA32);
460 		set_thread_flag(TIF_X32);
461 		if (current->mm)
462 			current->mm->context.ia32_compat = TIF_X32;
463 		current->personality &= ~READ_IMPLIES_EXEC;
464 		/* is_compat_task() uses the presence of the x32
465 		   syscall bit flag to determine compat status */
466 		current_thread_info()->status &= ~TS_COMPAT;
467 	} else {
468 		set_thread_flag(TIF_IA32);
469 		clear_thread_flag(TIF_X32);
470 		if (current->mm)
471 			current->mm->context.ia32_compat = TIF_IA32;
472 		current->personality |= force_personality32;
473 		/* Prepare the first "return" to user space */
474 		current_thread_info()->status |= TS_COMPAT;
475 	}
476 }
477 EXPORT_SYMBOL_GPL(set_personality_ia32);
478 
479 unsigned long get_wchan(struct task_struct *p)
480 {
481 	unsigned long stack;
482 	u64 fp, ip;
483 	int count = 0;
484 
485 	if (!p || p == current || p->state == TASK_RUNNING)
486 		return 0;
487 	stack = (unsigned long)task_stack_page(p);
488 	if (p->thread.sp < stack || p->thread.sp >= stack+THREAD_SIZE)
489 		return 0;
490 	fp = *(u64 *)(p->thread.sp);
491 	do {
492 		if (fp < (unsigned long)stack ||
493 		    fp >= (unsigned long)stack+THREAD_SIZE)
494 			return 0;
495 		ip = *(u64 *)(fp+8);
496 		if (!in_sched_functions(ip))
497 			return ip;
498 		fp = *(u64 *)fp;
499 	} while (count++ < 16);
500 	return 0;
501 }
502 
503 long do_arch_prctl(struct task_struct *task, int code, unsigned long addr)
504 {
505 	int ret = 0;
506 	int doit = task == current;
507 	int cpu;
508 
509 	switch (code) {
510 	case ARCH_SET_GS:
511 		if (addr >= TASK_SIZE_OF(task))
512 			return -EPERM;
513 		cpu = get_cpu();
514 		/* handle small bases via the GDT because that's faster to
515 		   switch. */
516 		if (addr <= 0xffffffff) {
517 			set_32bit_tls(task, GS_TLS, addr);
518 			if (doit) {
519 				load_TLS(&task->thread, cpu);
520 				load_gs_index(GS_TLS_SEL);
521 			}
522 			task->thread.gsindex = GS_TLS_SEL;
523 			task->thread.gs = 0;
524 		} else {
525 			task->thread.gsindex = 0;
526 			task->thread.gs = addr;
527 			if (doit) {
528 				load_gs_index(0);
529 				ret = wrmsrl_safe(MSR_KERNEL_GS_BASE, addr);
530 			}
531 		}
532 		put_cpu();
533 		break;
534 	case ARCH_SET_FS:
535 		/* Not strictly needed for fs, but do it for symmetry
536 		   with gs */
537 		if (addr >= TASK_SIZE_OF(task))
538 			return -EPERM;
539 		cpu = get_cpu();
540 		/* handle small bases via the GDT because that's faster to
541 		   switch. */
542 		if (addr <= 0xffffffff) {
543 			set_32bit_tls(task, FS_TLS, addr);
544 			if (doit) {
545 				load_TLS(&task->thread, cpu);
546 				loadsegment(fs, FS_TLS_SEL);
547 			}
548 			task->thread.fsindex = FS_TLS_SEL;
549 			task->thread.fs = 0;
550 		} else {
551 			task->thread.fsindex = 0;
552 			task->thread.fs = addr;
553 			if (doit) {
554 				/* set the selector to 0 to not confuse
555 				   __switch_to */
556 				loadsegment(fs, 0);
557 				ret = wrmsrl_safe(MSR_FS_BASE, addr);
558 			}
559 		}
560 		put_cpu();
561 		break;
562 	case ARCH_GET_FS: {
563 		unsigned long base;
564 		if (task->thread.fsindex == FS_TLS_SEL)
565 			base = read_32bit_tls(task, FS_TLS);
566 		else if (doit)
567 			rdmsrl(MSR_FS_BASE, base);
568 		else
569 			base = task->thread.fs;
570 		ret = put_user(base, (unsigned long __user *)addr);
571 		break;
572 	}
573 	case ARCH_GET_GS: {
574 		unsigned long base;
575 		unsigned gsindex;
576 		if (task->thread.gsindex == GS_TLS_SEL)
577 			base = read_32bit_tls(task, GS_TLS);
578 		else if (doit) {
579 			savesegment(gs, gsindex);
580 			if (gsindex)
581 				rdmsrl(MSR_KERNEL_GS_BASE, base);
582 			else
583 				base = task->thread.gs;
584 		} else
585 			base = task->thread.gs;
586 		ret = put_user(base, (unsigned long __user *)addr);
587 		break;
588 	}
589 
590 	default:
591 		ret = -EINVAL;
592 		break;
593 	}
594 
595 	return ret;
596 }
597 
598 long sys_arch_prctl(int code, unsigned long addr)
599 {
600 	return do_arch_prctl(current, code, addr);
601 }
602 
603 unsigned long KSTK_ESP(struct task_struct *task)
604 {
605 	return (test_tsk_thread_flag(task, TIF_IA32)) ?
606 			(task_pt_regs(task)->sp) : ((task)->thread.usersp);
607 }
608