xref: /linux/arch/x86/kernel/process_64.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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/fpu/internal.h>
42 #include <asm/mmu_context.h>
43 #include <asm/prctl.h>
44 #include <asm/desc.h>
45 #include <asm/proto.h>
46 #include <asm/ia32.h>
47 #include <asm/idle.h>
48 #include <asm/syscalls.h>
49 #include <asm/debugreg.h>
50 #include <asm/switch_to.h>
51 
52 asmlinkage extern void ret_from_fork(void);
53 
54 __visible DEFINE_PER_CPU(unsigned long, rsp_scratch);
55 
56 /* Prints also some state that isn't saved in the pt_regs */
57 void __show_regs(struct pt_regs *regs, int all)
58 {
59 	unsigned long cr0 = 0L, cr2 = 0L, cr3 = 0L, cr4 = 0L, fs, gs, shadowgs;
60 	unsigned long d0, d1, d2, d3, d6, d7;
61 	unsigned int fsindex, gsindex;
62 	unsigned int ds, cs, es;
63 
64 	printk(KERN_DEFAULT "RIP: %04lx:[<%016lx>] ", regs->cs & 0xffff, regs->ip);
65 	printk_address(regs->ip);
66 	printk(KERN_DEFAULT "RSP: %04lx:%016lx  EFLAGS: %08lx\n", regs->ss,
67 			regs->sp, regs->flags);
68 	printk(KERN_DEFAULT "RAX: %016lx RBX: %016lx RCX: %016lx\n",
69 	       regs->ax, regs->bx, regs->cx);
70 	printk(KERN_DEFAULT "RDX: %016lx RSI: %016lx RDI: %016lx\n",
71 	       regs->dx, regs->si, regs->di);
72 	printk(KERN_DEFAULT "RBP: %016lx R08: %016lx R09: %016lx\n",
73 	       regs->bp, regs->r8, regs->r9);
74 	printk(KERN_DEFAULT "R10: %016lx R11: %016lx R12: %016lx\n",
75 	       regs->r10, regs->r11, regs->r12);
76 	printk(KERN_DEFAULT "R13: %016lx R14: %016lx R15: %016lx\n",
77 	       regs->r13, regs->r14, regs->r15);
78 
79 	asm("movl %%ds,%0" : "=r" (ds));
80 	asm("movl %%cs,%0" : "=r" (cs));
81 	asm("movl %%es,%0" : "=r" (es));
82 	asm("movl %%fs,%0" : "=r" (fsindex));
83 	asm("movl %%gs,%0" : "=r" (gsindex));
84 
85 	rdmsrl(MSR_FS_BASE, fs);
86 	rdmsrl(MSR_GS_BASE, gs);
87 	rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
88 
89 	if (!all)
90 		return;
91 
92 	cr0 = read_cr0();
93 	cr2 = read_cr2();
94 	cr3 = read_cr3();
95 	cr4 = __read_cr4();
96 
97 	printk(KERN_DEFAULT "FS:  %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
98 	       fs, fsindex, gs, gsindex, shadowgs);
99 	printk(KERN_DEFAULT "CS:  %04x DS: %04x ES: %04x CR0: %016lx\n", cs, ds,
100 			es, cr0);
101 	printk(KERN_DEFAULT "CR2: %016lx CR3: %016lx CR4: %016lx\n", cr2, cr3,
102 			cr4);
103 
104 	get_debugreg(d0, 0);
105 	get_debugreg(d1, 1);
106 	get_debugreg(d2, 2);
107 	get_debugreg(d3, 3);
108 	get_debugreg(d6, 6);
109 	get_debugreg(d7, 7);
110 
111 	/* Only print out debug registers if they are in their non-default state. */
112 	if ((d0 == 0) && (d1 == 0) && (d2 == 0) && (d3 == 0) &&
113 	    (d6 == DR6_RESERVED) && (d7 == 0x400))
114 		return;
115 
116 	printk(KERN_DEFAULT "DR0: %016lx DR1: %016lx DR2: %016lx\n", d0, d1, d2);
117 	printk(KERN_DEFAULT "DR3: %016lx DR6: %016lx DR7: %016lx\n", d3, d6, d7);
118 
119 }
120 
121 void release_thread(struct task_struct *dead_task)
122 {
123 	if (dead_task->mm) {
124 #ifdef CONFIG_MODIFY_LDT_SYSCALL
125 		if (dead_task->mm->context.ldt) {
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.ldt->size);
130 			BUG();
131 		}
132 #endif
133 	}
134 }
135 
136 static inline void set_32bit_tls(struct task_struct *t, int tls, u32 addr)
137 {
138 	struct user_desc ud = {
139 		.base_addr = addr,
140 		.limit = 0xfffff,
141 		.seg_32bit = 1,
142 		.limit_in_pages = 1,
143 		.useable = 1,
144 	};
145 	struct desc_struct *desc = t->thread.tls_array;
146 	desc += tls;
147 	fill_ldt(desc, &ud);
148 }
149 
150 static inline u32 read_32bit_tls(struct task_struct *t, int tls)
151 {
152 	return get_desc_base(&t->thread.tls_array[tls]);
153 }
154 
155 int copy_thread_tls(unsigned long clone_flags, unsigned long sp,
156 		unsigned long arg, struct task_struct *p, unsigned long tls)
157 {
158 	int err;
159 	struct pt_regs *childregs;
160 	struct task_struct *me = current;
161 
162 	p->thread.sp0 = (unsigned long)task_stack_page(p) + THREAD_SIZE;
163 	childregs = task_pt_regs(p);
164 	p->thread.sp = (unsigned long) childregs;
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 (is_ia32_task())
211 			err = do_set_thread_area(p, -1,
212 				(struct user_desc __user *)tls, 0);
213 		else
214 #endif
215 			err = do_arch_prctl(p, ARCH_SET_FS, tls);
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 	regs->ip		= new_ip;
239 	regs->sp		= new_sp;
240 	regs->cs		= _cs;
241 	regs->ss		= _ss;
242 	regs->flags		= X86_EFLAGS_IF;
243 	force_iret();
244 }
245 
246 void
247 start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
248 {
249 	start_thread_common(regs, new_ip, new_sp,
250 			    __USER_CS, __USER_DS, 0);
251 }
252 
253 #ifdef CONFIG_COMPAT
254 void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp)
255 {
256 	start_thread_common(regs, new_ip, new_sp,
257 			    test_thread_flag(TIF_X32)
258 			    ? __USER_CS : __USER32_CS,
259 			    __USER_DS, __USER_DS);
260 }
261 #endif
262 
263 /*
264  *	switch_to(x,y) should switch tasks from x to y.
265  *
266  * This could still be optimized:
267  * - fold all the options into a flag word and test it with a single test.
268  * - could test fs/gs bitsliced
269  *
270  * Kprobes not supported here. Set the probe on schedule instead.
271  * Function graph tracer not supported too.
272  */
273 __visible __notrace_funcgraph struct task_struct *
274 __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
275 {
276 	struct thread_struct *prev = &prev_p->thread;
277 	struct thread_struct *next = &next_p->thread;
278 	struct fpu *prev_fpu = &prev->fpu;
279 	struct fpu *next_fpu = &next->fpu;
280 	int cpu = smp_processor_id();
281 	struct tss_struct *tss = &per_cpu(cpu_tss, cpu);
282 	unsigned fsindex, gsindex;
283 	fpu_switch_t fpu_switch;
284 
285 	fpu_switch = switch_fpu_prepare(prev_fpu, next_fpu, cpu);
286 
287 	/* We must save %fs and %gs before load_TLS() because
288 	 * %fs and %gs may be cleared by load_TLS().
289 	 *
290 	 * (e.g. xen_load_tls())
291 	 */
292 	savesegment(fs, fsindex);
293 	savesegment(gs, gsindex);
294 
295 	/*
296 	 * Load TLS before restoring any segments so that segment loads
297 	 * reference the correct GDT entries.
298 	 */
299 	load_TLS(next, cpu);
300 
301 	/*
302 	 * Leave lazy mode, flushing any hypercalls made here.  This
303 	 * must be done after loading TLS entries in the GDT but before
304 	 * loading segments that might reference them, and and it must
305 	 * be done before fpu__restore(), so the TS bit is up to
306 	 * date.
307 	 */
308 	arch_end_context_switch(next_p);
309 
310 	/* Switch DS and ES.
311 	 *
312 	 * Reading them only returns the selectors, but writing them (if
313 	 * nonzero) loads the full descriptor from the GDT or LDT.  The
314 	 * LDT for next is loaded in switch_mm, and the GDT is loaded
315 	 * above.
316 	 *
317 	 * We therefore need to write new values to the segment
318 	 * registers on every context switch unless both the new and old
319 	 * values are zero.
320 	 *
321 	 * Note that we don't need to do anything for CS and SS, as
322 	 * those are saved and restored as part of pt_regs.
323 	 */
324 	savesegment(es, prev->es);
325 	if (unlikely(next->es | prev->es))
326 		loadsegment(es, next->es);
327 
328 	savesegment(ds, prev->ds);
329 	if (unlikely(next->ds | prev->ds))
330 		loadsegment(ds, next->ds);
331 
332 	/*
333 	 * Switch FS and GS.
334 	 *
335 	 * These are even more complicated than FS and GS: they have
336 	 * 64-bit bases are that controlled by arch_prctl.  Those bases
337 	 * only differ from the values in the GDT or LDT if the selector
338 	 * is 0.
339 	 *
340 	 * Loading the segment register resets the hidden base part of
341 	 * the register to 0 or the value from the GDT / LDT.  If the
342 	 * next base address zero, writing 0 to the segment register is
343 	 * much faster than using wrmsr to explicitly zero the base.
344 	 *
345 	 * The thread_struct.fs and thread_struct.gs values are 0
346 	 * if the fs and gs bases respectively are not overridden
347 	 * from the values implied by fsindex and gsindex.  They
348 	 * are nonzero, and store the nonzero base addresses, if
349 	 * the bases are overridden.
350 	 *
351 	 * (fs != 0 && fsindex != 0) || (gs != 0 && gsindex != 0) should
352 	 * be impossible.
353 	 *
354 	 * Therefore we need to reload the segment registers if either
355 	 * the old or new selector is nonzero, and we need to override
356 	 * the base address if next thread expects it to be overridden.
357 	 *
358 	 * This code is unnecessarily slow in the case where the old and
359 	 * new indexes are zero and the new base is nonzero -- it will
360 	 * unnecessarily write 0 to the selector before writing the new
361 	 * base address.
362 	 *
363 	 * Note: This all depends on arch_prctl being the only way that
364 	 * user code can override the segment base.  Once wrfsbase and
365 	 * wrgsbase are enabled, most of this code will need to change.
366 	 */
367 	if (unlikely(fsindex | next->fsindex | prev->fs)) {
368 		loadsegment(fs, next->fsindex);
369 
370 		/*
371 		 * If user code wrote a nonzero value to FS, then it also
372 		 * cleared the overridden base address.
373 		 *
374 		 * XXX: if user code wrote 0 to FS and cleared the base
375 		 * address itself, we won't notice and we'll incorrectly
376 		 * restore the prior base address next time we reschdule
377 		 * the process.
378 		 */
379 		if (fsindex)
380 			prev->fs = 0;
381 	}
382 	if (next->fs)
383 		wrmsrl(MSR_FS_BASE, next->fs);
384 	prev->fsindex = fsindex;
385 
386 	if (unlikely(gsindex | next->gsindex | prev->gs)) {
387 		load_gs_index(next->gsindex);
388 
389 		/* This works (and fails) the same way as fsindex above. */
390 		if (gsindex)
391 			prev->gs = 0;
392 	}
393 	if (next->gs)
394 		wrmsrl(MSR_KERNEL_GS_BASE, next->gs);
395 	prev->gsindex = gsindex;
396 
397 	switch_fpu_finish(next_fpu, fpu_switch);
398 
399 	/*
400 	 * Switch the PDA and FPU contexts.
401 	 */
402 	this_cpu_write(current_task, next_p);
403 
404 	/*
405 	 * If it were not for PREEMPT_ACTIVE we could guarantee that the
406 	 * preempt_count of all tasks was equal here and this would not be
407 	 * needed.
408 	 */
409 	task_thread_info(prev_p)->saved_preempt_count = this_cpu_read(__preempt_count);
410 	this_cpu_write(__preempt_count, task_thread_info(next_p)->saved_preempt_count);
411 
412 	/* Reload esp0 and ss1.  This changes current_thread_info(). */
413 	load_sp0(tss, next);
414 
415 	/*
416 	 * Now maybe reload the debug registers and handle I/O bitmaps
417 	 */
418 	if (unlikely(task_thread_info(next_p)->flags & _TIF_WORK_CTXSW_NEXT ||
419 		     task_thread_info(prev_p)->flags & _TIF_WORK_CTXSW_PREV))
420 		__switch_to_xtra(prev_p, next_p, tss);
421 
422 	if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
423 		/*
424 		 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
425 		 * does not update the cached descriptor.  As a result, if we
426 		 * do SYSRET while SS is NULL, we'll end up in user mode with
427 		 * SS apparently equal to __USER_DS but actually unusable.
428 		 *
429 		 * The straightforward workaround would be to fix it up just
430 		 * before SYSRET, but that would slow down the system call
431 		 * fast paths.  Instead, we ensure that SS is never NULL in
432 		 * system call context.  We do this by replacing NULL SS
433 		 * selectors at every context switch.  SYSCALL sets up a valid
434 		 * SS, so the only way to get NULL is to re-enter the kernel
435 		 * from CPL 3 through an interrupt.  Since that can't happen
436 		 * in the same task as a running syscall, we are guaranteed to
437 		 * context switch between every interrupt vector entry and a
438 		 * subsequent SYSRET.
439 		 *
440 		 * We read SS first because SS reads are much faster than
441 		 * writes.  Out of caution, we force SS to __KERNEL_DS even if
442 		 * it previously had a different non-NULL value.
443 		 */
444 		unsigned short ss_sel;
445 		savesegment(ss, ss_sel);
446 		if (ss_sel != __KERNEL_DS)
447 			loadsegment(ss, __KERNEL_DS);
448 	}
449 
450 	return prev_p;
451 }
452 
453 void set_personality_64bit(void)
454 {
455 	/* inherit personality from parent */
456 
457 	/* Make sure to be in 64bit mode */
458 	clear_thread_flag(TIF_IA32);
459 	clear_thread_flag(TIF_ADDR32);
460 	clear_thread_flag(TIF_X32);
461 
462 	/* Ensure the corresponding mm is not marked. */
463 	if (current->mm)
464 		current->mm->context.ia32_compat = 0;
465 
466 	/* TBD: overwrites user setup. Should have two bits.
467 	   But 64bit processes have always behaved this way,
468 	   so it's not too bad. The main problem is just that
469 	   32bit childs are affected again. */
470 	current->personality &= ~READ_IMPLIES_EXEC;
471 }
472 
473 void set_personality_ia32(bool x32)
474 {
475 	/* inherit personality from parent */
476 
477 	/* Make sure to be in 32bit mode */
478 	set_thread_flag(TIF_ADDR32);
479 
480 	/* Mark the associated mm as containing 32-bit tasks. */
481 	if (x32) {
482 		clear_thread_flag(TIF_IA32);
483 		set_thread_flag(TIF_X32);
484 		if (current->mm)
485 			current->mm->context.ia32_compat = TIF_X32;
486 		current->personality &= ~READ_IMPLIES_EXEC;
487 		/* is_compat_task() uses the presence of the x32
488 		   syscall bit flag to determine compat status */
489 		current_thread_info()->status &= ~TS_COMPAT;
490 	} else {
491 		set_thread_flag(TIF_IA32);
492 		clear_thread_flag(TIF_X32);
493 		if (current->mm)
494 			current->mm->context.ia32_compat = TIF_IA32;
495 		current->personality |= force_personality32;
496 		/* Prepare the first "return" to user space */
497 		current_thread_info()->status |= TS_COMPAT;
498 	}
499 }
500 EXPORT_SYMBOL_GPL(set_personality_ia32);
501 
502 long do_arch_prctl(struct task_struct *task, int code, unsigned long addr)
503 {
504 	int ret = 0;
505 	int doit = task == current;
506 	int cpu;
507 
508 	switch (code) {
509 	case ARCH_SET_GS:
510 		if (addr >= TASK_SIZE_OF(task))
511 			return -EPERM;
512 		cpu = get_cpu();
513 		/* handle small bases via the GDT because that's faster to
514 		   switch. */
515 		if (addr <= 0xffffffff) {
516 			set_32bit_tls(task, GS_TLS, addr);
517 			if (doit) {
518 				load_TLS(&task->thread, cpu);
519 				load_gs_index(GS_TLS_SEL);
520 			}
521 			task->thread.gsindex = GS_TLS_SEL;
522 			task->thread.gs = 0;
523 		} else {
524 			task->thread.gsindex = 0;
525 			task->thread.gs = addr;
526 			if (doit) {
527 				load_gs_index(0);
528 				ret = wrmsrl_safe(MSR_KERNEL_GS_BASE, addr);
529 			}
530 		}
531 		put_cpu();
532 		break;
533 	case ARCH_SET_FS:
534 		/* Not strictly needed for fs, but do it for symmetry
535 		   with gs */
536 		if (addr >= TASK_SIZE_OF(task))
537 			return -EPERM;
538 		cpu = get_cpu();
539 		/* handle small bases via the GDT because that's faster to
540 		   switch. */
541 		if (addr <= 0xffffffff) {
542 			set_32bit_tls(task, FS_TLS, addr);
543 			if (doit) {
544 				load_TLS(&task->thread, cpu);
545 				loadsegment(fs, FS_TLS_SEL);
546 			}
547 			task->thread.fsindex = FS_TLS_SEL;
548 			task->thread.fs = 0;
549 		} else {
550 			task->thread.fsindex = 0;
551 			task->thread.fs = addr;
552 			if (doit) {
553 				/* set the selector to 0 to not confuse
554 				   __switch_to */
555 				loadsegment(fs, 0);
556 				ret = wrmsrl_safe(MSR_FS_BASE, addr);
557 			}
558 		}
559 		put_cpu();
560 		break;
561 	case ARCH_GET_FS: {
562 		unsigned long base;
563 		if (task->thread.fsindex == FS_TLS_SEL)
564 			base = read_32bit_tls(task, FS_TLS);
565 		else if (doit)
566 			rdmsrl(MSR_FS_BASE, base);
567 		else
568 			base = task->thread.fs;
569 		ret = put_user(base, (unsigned long __user *)addr);
570 		break;
571 	}
572 	case ARCH_GET_GS: {
573 		unsigned long base;
574 		unsigned gsindex;
575 		if (task->thread.gsindex == GS_TLS_SEL)
576 			base = read_32bit_tls(task, GS_TLS);
577 		else if (doit) {
578 			savesegment(gs, gsindex);
579 			if (gsindex)
580 				rdmsrl(MSR_KERNEL_GS_BASE, base);
581 			else
582 				base = task->thread.gs;
583 		} else
584 			base = task->thread.gs;
585 		ret = put_user(base, (unsigned long __user *)addr);
586 		break;
587 	}
588 
589 	default:
590 		ret = -EINVAL;
591 		break;
592 	}
593 
594 	return ret;
595 }
596 
597 long sys_arch_prctl(int code, unsigned long addr)
598 {
599 	return do_arch_prctl(current, code, addr);
600 }
601 
602 unsigned long KSTK_ESP(struct task_struct *task)
603 {
604 	return task_pt_regs(task)->sp;
605 }
606