2 * Copyright (C) 1995 Linus Torvalds
4 * Pentium III FXSR, SSE support
14 * This file handles the architecture-dependent parts of process handling..
17 #include <linux/cpu.h>
18 #include <linux/errno.h>
19 #include <linux/sched.h>
21 #include <linux/kernel.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/export.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>
37 #include <linux/ftrace.h>
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>
45 #include <asm/proto.h>
48 #include <asm/syscalls.h>
49 #include <asm/debugreg.h>
50 #include <asm/switch_to.h>
51 #include <asm/xen/hypervisor.h>
54 __visible DEFINE_PER_CPU(unsigned long, rsp_scratch);
56 /* Prints also some state that isn't saved in the pt_regs */
57 void __show_regs(struct pt_regs *regs, int all)
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;
64 printk(KERN_DEFAULT "RIP: %04lx:%pS\n", regs->cs & 0xffff,
66 printk(KERN_DEFAULT "RSP: %04lx:%016lx EFLAGS: %08lx", regs->ss,
67 regs->sp, regs->flags);
68 if (regs->orig_ax != -1)
69 pr_cont(" ORIG_RAX: %016lx\n", regs->orig_ax);
73 printk(KERN_DEFAULT "RAX: %016lx RBX: %016lx RCX: %016lx\n",
74 regs->ax, regs->bx, regs->cx);
75 printk(KERN_DEFAULT "RDX: %016lx RSI: %016lx RDI: %016lx\n",
76 regs->dx, regs->si, regs->di);
77 printk(KERN_DEFAULT "RBP: %016lx R08: %016lx R09: %016lx\n",
78 regs->bp, regs->r8, regs->r9);
79 printk(KERN_DEFAULT "R10: %016lx R11: %016lx R12: %016lx\n",
80 regs->r10, regs->r11, regs->r12);
81 printk(KERN_DEFAULT "R13: %016lx R14: %016lx R15: %016lx\n",
82 regs->r13, regs->r14, regs->r15);
84 asm("movl %%ds,%0" : "=r" (ds));
85 asm("movl %%cs,%0" : "=r" (cs));
86 asm("movl %%es,%0" : "=r" (es));
87 asm("movl %%fs,%0" : "=r" (fsindex));
88 asm("movl %%gs,%0" : "=r" (gsindex));
90 rdmsrl(MSR_FS_BASE, fs);
91 rdmsrl(MSR_GS_BASE, gs);
92 rdmsrl(MSR_KERNEL_GS_BASE, shadowgs);
102 printk(KERN_DEFAULT "FS: %016lx(%04x) GS:%016lx(%04x) knlGS:%016lx\n",
103 fs, fsindex, gs, gsindex, shadowgs);
104 printk(KERN_DEFAULT "CS: %04x DS: %04x ES: %04x CR0: %016lx\n", cs, ds,
106 printk(KERN_DEFAULT "CR2: %016lx CR3: %016lx CR4: %016lx\n", cr2, cr3,
116 /* Only print out debug registers if they are in their non-default state. */
117 if (!((d0 == 0) && (d1 == 0) && (d2 == 0) && (d3 == 0) &&
118 (d6 == DR6_RESERVED) && (d7 == 0x400))) {
119 printk(KERN_DEFAULT "DR0: %016lx DR1: %016lx DR2: %016lx\n",
121 printk(KERN_DEFAULT "DR3: %016lx DR6: %016lx DR7: %016lx\n",
125 if (boot_cpu_has(X86_FEATURE_OSPKE))
126 printk(KERN_DEFAULT "PKRU: %08x\n", read_pkru());
129 void release_thread(struct task_struct *dead_task)
132 #ifdef CONFIG_MODIFY_LDT_SYSCALL
133 if (dead_task->mm->context.ldt) {
134 pr_warn("WARNING: dead process %s still has LDT? <%p/%d>\n",
136 dead_task->mm->context.ldt->entries,
137 dead_task->mm->context.ldt->size);
144 int copy_thread_tls(unsigned long clone_flags, unsigned long sp,
145 unsigned long arg, struct task_struct *p, unsigned long tls)
148 struct pt_regs *childregs;
149 struct fork_frame *fork_frame;
150 struct inactive_task_frame *frame;
151 struct task_struct *me = current;
153 p->thread.sp0 = (unsigned long)task_stack_page(p) + THREAD_SIZE;
154 childregs = task_pt_regs(p);
155 fork_frame = container_of(childregs, struct fork_frame, regs);
156 frame = &fork_frame->frame;
158 frame->ret_addr = (unsigned long) ret_from_fork;
159 p->thread.sp = (unsigned long) fork_frame;
160 p->thread.io_bitmap_ptr = NULL;
162 savesegment(gs, p->thread.gsindex);
163 p->thread.gsbase = p->thread.gsindex ? 0 : me->thread.gsbase;
164 savesegment(fs, p->thread.fsindex);
165 p->thread.fsbase = p->thread.fsindex ? 0 : me->thread.fsbase;
166 savesegment(es, p->thread.es);
167 savesegment(ds, p->thread.ds);
168 memset(p->thread.ptrace_bps, 0, sizeof(p->thread.ptrace_bps));
170 if (unlikely(p->flags & PF_KTHREAD)) {
172 memset(childregs, 0, sizeof(struct pt_regs));
173 frame->bx = sp; /* function */
178 *childregs = *current_pt_regs();
185 if (unlikely(test_tsk_thread_flag(me, TIF_IO_BITMAP))) {
186 p->thread.io_bitmap_ptr = kmemdup(me->thread.io_bitmap_ptr,
187 IO_BITMAP_BYTES, GFP_KERNEL);
188 if (!p->thread.io_bitmap_ptr) {
189 p->thread.io_bitmap_max = 0;
192 set_tsk_thread_flag(p, TIF_IO_BITMAP);
196 * Set a new TLS for the child thread?
198 if (clone_flags & CLONE_SETTLS) {
199 #ifdef CONFIG_IA32_EMULATION
200 if (in_ia32_syscall())
201 err = do_set_thread_area(p, -1,
202 (struct user_desc __user *)tls, 0);
205 err = do_arch_prctl(p, ARCH_SET_FS, tls);
211 if (err && p->thread.io_bitmap_ptr) {
212 kfree(p->thread.io_bitmap_ptr);
213 p->thread.io_bitmap_max = 0;
220 start_thread_common(struct pt_regs *regs, unsigned long new_ip,
221 unsigned long new_sp,
222 unsigned int _cs, unsigned int _ss, unsigned int _ds)
225 loadsegment(es, _ds);
226 loadsegment(ds, _ds);
232 regs->flags = X86_EFLAGS_IF;
237 start_thread(struct pt_regs *regs, unsigned long new_ip, unsigned long new_sp)
239 start_thread_common(regs, new_ip, new_sp,
240 __USER_CS, __USER_DS, 0);
244 void compat_start_thread(struct pt_regs *regs, u32 new_ip, u32 new_sp)
246 start_thread_common(regs, new_ip, new_sp,
247 test_thread_flag(TIF_X32)
248 ? __USER_CS : __USER32_CS,
249 __USER_DS, __USER_DS);
254 * switch_to(x,y) should switch tasks from x to y.
256 * This could still be optimized:
257 * - fold all the options into a flag word and test it with a single test.
258 * - could test fs/gs bitsliced
260 * Kprobes not supported here. Set the probe on schedule instead.
261 * Function graph tracer not supported too.
263 __visible __notrace_funcgraph struct task_struct *
264 __switch_to(struct task_struct *prev_p, struct task_struct *next_p)
266 struct thread_struct *prev = &prev_p->thread;
267 struct thread_struct *next = &next_p->thread;
268 struct fpu *prev_fpu = &prev->fpu;
269 struct fpu *next_fpu = &next->fpu;
270 int cpu = smp_processor_id();
271 struct tss_struct *tss = &per_cpu(cpu_tss, cpu);
272 unsigned prev_fsindex, prev_gsindex;
273 fpu_switch_t fpu_switch;
275 fpu_switch = switch_fpu_prepare(prev_fpu, next_fpu, cpu);
277 /* We must save %fs and %gs before load_TLS() because
278 * %fs and %gs may be cleared by load_TLS().
280 * (e.g. xen_load_tls())
282 savesegment(fs, prev_fsindex);
283 savesegment(gs, prev_gsindex);
286 * Load TLS before restoring any segments so that segment loads
287 * reference the correct GDT entries.
292 * Leave lazy mode, flushing any hypercalls made here. This
293 * must be done after loading TLS entries in the GDT but before
294 * loading segments that might reference them, and and it must
295 * be done before fpu__restore(), so the TS bit is up to
298 arch_end_context_switch(next_p);
302 * Reading them only returns the selectors, but writing them (if
303 * nonzero) loads the full descriptor from the GDT or LDT. The
304 * LDT for next is loaded in switch_mm, and the GDT is loaded
307 * We therefore need to write new values to the segment
308 * registers on every context switch unless both the new and old
311 * Note that we don't need to do anything for CS and SS, as
312 * those are saved and restored as part of pt_regs.
314 savesegment(es, prev->es);
315 if (unlikely(next->es | prev->es))
316 loadsegment(es, next->es);
318 savesegment(ds, prev->ds);
319 if (unlikely(next->ds | prev->ds))
320 loadsegment(ds, next->ds);
325 * These are even more complicated than DS and ES: they have
326 * 64-bit bases are that controlled by arch_prctl. The bases
327 * don't necessarily match the selectors, as user code can do
328 * any number of things to cause them to be inconsistent.
330 * We don't promise to preserve the bases if the selectors are
331 * nonzero. We also don't promise to preserve the base if the
332 * selector is zero and the base doesn't match whatever was
333 * most recently passed to ARCH_SET_FS/GS. (If/when the
334 * FSGSBASE instructions are enabled, we'll need to offer
335 * stronger guarantees.)
338 * (fsbase != 0 && fsindex != 0) || (gsbase != 0 && gsindex != 0) is
342 /* Loading a nonzero value into FS sets the index and base. */
343 loadsegment(fs, next->fsindex);
346 /* Next index is zero but next base is nonzero. */
349 wrmsrl(MSR_FS_BASE, next->fsbase);
351 /* Next base and index are both zero. */
352 if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
354 * We don't know the previous base and can't
355 * find out without RDMSR. Forcibly clear it.
357 loadsegment(fs, __USER_DS);
361 * If the previous index is zero and ARCH_SET_FS
362 * didn't change the base, then the base is
363 * also zero and we don't need to do anything.
365 if (prev->fsbase || prev_fsindex)
371 * Save the old state and preserve the invariant.
372 * NB: if prev_fsindex == 0, then we can't reliably learn the base
373 * without RDMSR because Intel user code can zero it without telling
374 * us and AMD user code can program any 32-bit value without telling
379 prev->fsindex = prev_fsindex;
382 /* Loading a nonzero value into GS sets the index and base. */
383 load_gs_index(next->gsindex);
386 /* Next index is zero but next base is nonzero. */
389 wrmsrl(MSR_KERNEL_GS_BASE, next->gsbase);
391 /* Next base and index are both zero. */
392 if (static_cpu_has_bug(X86_BUG_NULL_SEG)) {
394 * We don't know the previous base and can't
395 * find out without RDMSR. Forcibly clear it.
397 * This contains a pointless SWAPGS pair.
398 * Fixing it would involve an explicit check
399 * for Xen or a new pvop.
401 load_gs_index(__USER_DS);
405 * If the previous index is zero and ARCH_SET_GS
406 * didn't change the base, then the base is
407 * also zero and we don't need to do anything.
409 if (prev->gsbase || prev_gsindex)
415 * Save the old state and preserve the invariant.
416 * NB: if prev_gsindex == 0, then we can't reliably learn the base
417 * without RDMSR because Intel user code can zero it without telling
418 * us and AMD user code can program any 32-bit value without telling
423 prev->gsindex = prev_gsindex;
425 switch_fpu_finish(next_fpu, fpu_switch);
428 * Switch the PDA and FPU contexts.
430 this_cpu_write(current_task, next_p);
432 /* Reload esp0 and ss1. This changes current_thread_info(). */
436 * Now maybe reload the debug registers and handle I/O bitmaps
438 if (unlikely(task_thread_info(next_p)->flags & _TIF_WORK_CTXSW_NEXT ||
439 task_thread_info(prev_p)->flags & _TIF_WORK_CTXSW_PREV))
440 __switch_to_xtra(prev_p, next_p, tss);
444 * On Xen PV, IOPL bits in pt_regs->flags have no effect, and
445 * current_pt_regs()->flags may not match the current task's
446 * intended IOPL. We need to switch it manually.
448 if (unlikely(static_cpu_has(X86_FEATURE_XENPV) &&
449 prev->iopl != next->iopl))
450 xen_set_iopl_mask(next->iopl);
453 if (static_cpu_has_bug(X86_BUG_SYSRET_SS_ATTRS)) {
455 * AMD CPUs have a misfeature: SYSRET sets the SS selector but
456 * does not update the cached descriptor. As a result, if we
457 * do SYSRET while SS is NULL, we'll end up in user mode with
458 * SS apparently equal to __USER_DS but actually unusable.
460 * The straightforward workaround would be to fix it up just
461 * before SYSRET, but that would slow down the system call
462 * fast paths. Instead, we ensure that SS is never NULL in
463 * system call context. We do this by replacing NULL SS
464 * selectors at every context switch. SYSCALL sets up a valid
465 * SS, so the only way to get NULL is to re-enter the kernel
466 * from CPL 3 through an interrupt. Since that can't happen
467 * in the same task as a running syscall, we are guaranteed to
468 * context switch between every interrupt vector entry and a
471 * We read SS first because SS reads are much faster than
472 * writes. Out of caution, we force SS to __KERNEL_DS even if
473 * it previously had a different non-NULL value.
475 unsigned short ss_sel;
476 savesegment(ss, ss_sel);
477 if (ss_sel != __KERNEL_DS)
478 loadsegment(ss, __KERNEL_DS);
484 void set_personality_64bit(void)
486 /* inherit personality from parent */
488 /* Make sure to be in 64bit mode */
489 clear_thread_flag(TIF_IA32);
490 clear_thread_flag(TIF_ADDR32);
491 clear_thread_flag(TIF_X32);
493 /* Ensure the corresponding mm is not marked. */
495 current->mm->context.ia32_compat = 0;
497 /* TBD: overwrites user setup. Should have two bits.
498 But 64bit processes have always behaved this way,
499 so it's not too bad. The main problem is just that
500 32bit childs are affected again. */
501 current->personality &= ~READ_IMPLIES_EXEC;
504 void set_personality_ia32(bool x32)
506 /* inherit personality from parent */
508 /* Make sure to be in 32bit mode */
509 set_thread_flag(TIF_ADDR32);
511 /* Mark the associated mm as containing 32-bit tasks. */
513 clear_thread_flag(TIF_IA32);
514 set_thread_flag(TIF_X32);
516 current->mm->context.ia32_compat = TIF_X32;
517 current->personality &= ~READ_IMPLIES_EXEC;
518 /* in_compat_syscall() uses the presence of the x32
519 syscall bit flag to determine compat status */
520 current->thread.status &= ~TS_COMPAT;
522 set_thread_flag(TIF_IA32);
523 clear_thread_flag(TIF_X32);
525 current->mm->context.ia32_compat = TIF_IA32;
526 current->personality |= force_personality32;
527 /* Prepare the first "return" to user space */
528 current->thread.status |= TS_COMPAT;
531 EXPORT_SYMBOL_GPL(set_personality_ia32);
533 #ifdef CONFIG_CHECKPOINT_RESTORE
534 static long prctl_map_vdso(const struct vdso_image *image, unsigned long addr)
538 ret = map_vdso_once(image, addr);
542 return (long)image->size;
546 long do_arch_prctl(struct task_struct *task, int code, unsigned long addr)
549 int doit = task == current;
554 if (addr >= TASK_SIZE_MAX)
557 task->thread.gsindex = 0;
558 task->thread.gsbase = addr;
561 ret = wrmsrl_safe(MSR_KERNEL_GS_BASE, addr);
566 /* Not strictly needed for fs, but do it for symmetry
568 if (addr >= TASK_SIZE_MAX)
571 task->thread.fsindex = 0;
572 task->thread.fsbase = addr;
574 /* set the selector to 0 to not confuse __switch_to */
576 ret = wrmsrl_safe(MSR_FS_BASE, addr);
583 rdmsrl(MSR_FS_BASE, base);
585 base = task->thread.fsbase;
586 ret = put_user(base, (unsigned long __user *)addr);
592 rdmsrl(MSR_KERNEL_GS_BASE, base);
594 base = task->thread.gsbase;
595 ret = put_user(base, (unsigned long __user *)addr);
599 #ifdef CONFIG_CHECKPOINT_RESTORE
600 # ifdef CONFIG_X86_X32_ABI
601 case ARCH_MAP_VDSO_X32:
602 return prctl_map_vdso(&vdso_image_x32, addr);
604 # if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
605 case ARCH_MAP_VDSO_32:
606 return prctl_map_vdso(&vdso_image_32, addr);
608 case ARCH_MAP_VDSO_64:
609 return prctl_map_vdso(&vdso_image_64, addr);
620 long sys_arch_prctl(int code, unsigned long addr)
622 return do_arch_prctl(current, code, addr);
625 unsigned long KSTK_ESP(struct task_struct *task)
627 return task_pt_regs(task)->sp;