4 * Copyright (C) 1994 Linus Torvalds
6 * 29 dec 2001 - Fixed oopses caused by unchecked access to the vm86
7 * stack - Manfred Spraul <manfred@colorfullife.com>
9 * 22 mar 2002 - Manfred detected the stackfaults, but didn't handle
10 * them correctly. Now the emulation will be in a
11 * consistent state after stackfaults - Kasper Dupont
12 * <kasperd@daimi.au.dk>
14 * 22 mar 2002 - Added missing clear_IF in set_vflags_* Kasper Dupont
15 * <kasperd@daimi.au.dk>
17 * ?? ??? 2002 - Fixed premature returns from handle_vm86_fault
18 * caused by Kasper Dupont's changes - Stas Sergeev
20 * 4 apr 2002 - Fixed CHECK_IF_IN_TRAP broken by Stas' changes.
21 * Kasper Dupont <kasperd@daimi.au.dk>
23 * 9 apr 2002 - Changed syntax of macros in handle_vm86_fault.
24 * Kasper Dupont <kasperd@daimi.au.dk>
26 * 9 apr 2002 - Changed stack access macros to jump to a label
27 * instead of returning to userspace. This simplifies
28 * do_int, and is needed by handle_vm6_fault. Kasper
29 * Dupont <kasperd@daimi.au.dk>
33 #include <linux/capability.h>
34 #include <linux/config.h>
35 #include <linux/errno.h>
36 #include <linux/interrupt.h>
37 #include <linux/sched.h>
38 #include <linux/kernel.h>
39 #include <linux/signal.h>
40 #include <linux/string.h>
42 #include <linux/smp.h>
43 #include <linux/smp_lock.h>
44 #include <linux/highmem.h>
45 #include <linux/ptrace.h>
46 #include <linux/audit.h>
48 #include <asm/uaccess.h>
50 #include <asm/tlbflush.h>
56 * Interrupt handling is not guaranteed:
57 * - a real x86 will disable all interrupts for one instruction
58 * after a "mov ss,xx" to make stack handling atomic even without
59 * the 'lss' instruction. We can't guarantee this in v86 mode,
60 * as the next instruction might result in a page fault or similar.
61 * - a real x86 will have interrupts disabled for one instruction
62 * past the 'sti' that enables them. We don't bother with all the
65 * Let's hope these problems do not actually matter for anything.
69 #define KVM86 ((struct kernel_vm86_struct *)regs)
70 #define VMPI KVM86->vm86plus
74 * 8- and 16-bit register defines..
76 #define AL(regs) (((unsigned char *)&((regs)->eax))[0])
77 #define AH(regs) (((unsigned char *)&((regs)->eax))[1])
78 #define IP(regs) (*(unsigned short *)&((regs)->eip))
79 #define SP(regs) (*(unsigned short *)&((regs)->esp))
82 * virtual flags (16 and 32-bit versions)
84 #define VFLAGS (*(unsigned short *)&(current->thread.v86flags))
85 #define VEFLAGS (current->thread.v86flags)
87 #define set_flags(X,new,mask) \
88 ((X) = ((X) & ~(mask)) | ((new) & (mask)))
90 #define SAFE_MASK (0xDD5)
91 #define RETURN_MASK (0xDFF)
93 #define VM86_REGS_PART2 orig_eax
94 #define VM86_REGS_SIZE1 \
95 ( (unsigned)( & (((struct kernel_vm86_regs *)0)->VM86_REGS_PART2) ) )
96 #define VM86_REGS_SIZE2 (sizeof(struct kernel_vm86_regs) - VM86_REGS_SIZE1)
98 struct pt_regs * FASTCALL(save_v86_state(struct kernel_vm86_regs * regs));
99 struct pt_regs * fastcall save_v86_state(struct kernel_vm86_regs * regs)
101 struct tss_struct *tss;
106 * This gets called from entry.S with interrupts disabled, but
107 * from process context. Enable interrupts here, before trying
108 * to access user space.
112 if (!current->thread.vm86_info) {
113 printk("no vm86_info: BAD\n");
116 set_flags(regs->eflags, VEFLAGS, VIF_MASK | current->thread.v86mask);
117 tmp = copy_to_user(¤t->thread.vm86_info->regs,regs, VM86_REGS_SIZE1);
118 tmp += copy_to_user(¤t->thread.vm86_info->regs.VM86_REGS_PART2,
119 ®s->VM86_REGS_PART2, VM86_REGS_SIZE2);
120 tmp += put_user(current->thread.screen_bitmap,¤t->thread.vm86_info->screen_bitmap);
122 printk("vm86: could not access userspace vm86_info\n");
126 tss = &per_cpu(init_tss, get_cpu());
127 current->thread.esp0 = current->thread.saved_esp0;
128 current->thread.sysenter_cs = __KERNEL_CS;
129 load_esp0(tss, ¤t->thread);
130 current->thread.saved_esp0 = 0;
133 loadsegment(fs, current->thread.saved_fs);
134 loadsegment(gs, current->thread.saved_gs);
139 static void mark_screen_rdonly(struct mm_struct *mm)
148 pgd = pgd_offset(mm, 0xA0000);
149 if (pgd_none_or_clear_bad(pgd))
151 pud = pud_offset(pgd, 0xA0000);
152 if (pud_none_or_clear_bad(pud))
154 pmd = pmd_offset(pud, 0xA0000);
155 if (pmd_none_or_clear_bad(pmd))
157 pte = pte_offset_map_lock(mm, pmd, 0xA0000, &ptl);
158 for (i = 0; i < 32; i++) {
159 if (pte_present(*pte))
160 set_pte(pte, pte_wrprotect(*pte));
163 pte_unmap_unlock(pte, ptl);
170 static int do_vm86_irq_handling(int subfunction, int irqnumber);
171 static void do_sys_vm86(struct kernel_vm86_struct *info, struct task_struct *tsk);
173 asmlinkage int sys_vm86old(struct pt_regs regs)
175 struct vm86_struct __user *v86 = (struct vm86_struct __user *)regs.ebx;
176 struct kernel_vm86_struct info; /* declare this _on top_,
177 * this avoids wasting of stack space.
178 * This remains on the stack until we
179 * return to 32 bit user space.
181 struct task_struct *tsk;
182 int tmp, ret = -EPERM;
185 if (tsk->thread.saved_esp0)
187 tmp = copy_from_user(&info, v86, VM86_REGS_SIZE1);
188 tmp += copy_from_user(&info.regs.VM86_REGS_PART2, &v86->regs.VM86_REGS_PART2,
189 (long)&info.vm86plus - (long)&info.regs.VM86_REGS_PART2);
193 memset(&info.vm86plus, 0, (int)&info.regs32 - (int)&info.vm86plus);
195 tsk->thread.vm86_info = v86;
196 do_sys_vm86(&info, tsk);
197 ret = 0; /* we never return here */
203 asmlinkage int sys_vm86(struct pt_regs regs)
205 struct kernel_vm86_struct info; /* declare this _on top_,
206 * this avoids wasting of stack space.
207 * This remains on the stack until we
208 * return to 32 bit user space.
210 struct task_struct *tsk;
212 struct vm86plus_struct __user *v86;
216 case VM86_REQUEST_IRQ:
218 case VM86_GET_IRQ_BITS:
219 case VM86_GET_AND_RESET_IRQ:
220 ret = do_vm86_irq_handling(regs.ebx, (int)regs.ecx);
222 case VM86_PLUS_INSTALL_CHECK:
223 /* NOTE: on old vm86 stuff this will return the error
224 from access_ok(), because the subfunction is
225 interpreted as (invalid) address to vm86_struct.
226 So the installation check works.
232 /* we come here only for functions VM86_ENTER, VM86_ENTER_NO_BYPASS */
234 if (tsk->thread.saved_esp0)
236 v86 = (struct vm86plus_struct __user *)regs.ecx;
237 tmp = copy_from_user(&info, v86, VM86_REGS_SIZE1);
238 tmp += copy_from_user(&info.regs.VM86_REGS_PART2, &v86->regs.VM86_REGS_PART2,
239 (long)&info.regs32 - (long)&info.regs.VM86_REGS_PART2);
244 info.vm86plus.is_vm86pus = 1;
245 tsk->thread.vm86_info = (struct vm86_struct __user *)v86;
246 do_sys_vm86(&info, tsk);
247 ret = 0; /* we never return here */
253 static void do_sys_vm86(struct kernel_vm86_struct *info, struct task_struct *tsk)
255 struct tss_struct *tss;
258 * make sure the vm86() system call doesn't try to do anything silly
260 info->regs.__null_ds = 0;
261 info->regs.__null_es = 0;
263 /* we are clearing fs,gs later just before "jmp resume_userspace",
264 * because starting with Linux 2.1.x they aren't no longer saved/restored
268 * The eflags register is also special: we cannot trust that the user
269 * has set it up safely, so this makes sure interrupt etc flags are
270 * inherited from protected mode.
272 VEFLAGS = info->regs.eflags;
273 info->regs.eflags &= SAFE_MASK;
274 info->regs.eflags |= info->regs32->eflags & ~SAFE_MASK;
275 info->regs.eflags |= VM_MASK;
277 switch (info->cpu_type) {
279 tsk->thread.v86mask = 0;
282 tsk->thread.v86mask = NT_MASK | IOPL_MASK;
285 tsk->thread.v86mask = AC_MASK | NT_MASK | IOPL_MASK;
288 tsk->thread.v86mask = ID_MASK | AC_MASK | NT_MASK | IOPL_MASK;
293 * Save old state, set default return value (%eax) to 0
295 info->regs32->eax = 0;
296 tsk->thread.saved_esp0 = tsk->thread.esp0;
297 savesegment(fs, tsk->thread.saved_fs);
298 savesegment(gs, tsk->thread.saved_gs);
300 tss = &per_cpu(init_tss, get_cpu());
301 tsk->thread.esp0 = (unsigned long) &info->VM86_TSS_ESP0;
303 tsk->thread.sysenter_cs = 0;
304 load_esp0(tss, &tsk->thread);
307 tsk->thread.screen_bitmap = info->screen_bitmap;
308 if (info->flags & VM86_SCREEN_BITMAP)
309 mark_screen_rdonly(tsk->mm);
310 __asm__ __volatile__("xorl %eax,%eax; movl %eax,%fs; movl %eax,%gs\n\t");
311 __asm__ __volatile__("movl %%eax, %0\n" :"=r"(eax));
313 /*call audit_syscall_exit since we do not exit via the normal paths */
314 if (unlikely(current->audit_context))
315 audit_syscall_exit(current, AUDITSC_RESULT(eax), eax);
317 __asm__ __volatile__(
320 "jmp resume_userspace"
322 :"r" (&info->regs), "r" (task_thread_info(tsk)));
323 /* we never return here */
326 static inline void return_to_32bit(struct kernel_vm86_regs * regs16, int retval)
328 struct pt_regs * regs32;
330 regs32 = save_v86_state(regs16);
331 regs32->eax = retval;
332 __asm__ __volatile__("movl %0,%%esp\n\t"
334 "jmp resume_userspace"
335 : : "r" (regs32), "r" (current_thread_info()));
338 static inline void set_IF(struct kernel_vm86_regs * regs)
341 if (VEFLAGS & VIP_MASK)
342 return_to_32bit(regs, VM86_STI);
345 static inline void clear_IF(struct kernel_vm86_regs * regs)
347 VEFLAGS &= ~VIF_MASK;
350 static inline void clear_TF(struct kernel_vm86_regs * regs)
352 regs->eflags &= ~TF_MASK;
355 static inline void clear_AC(struct kernel_vm86_regs * regs)
357 regs->eflags &= ~AC_MASK;
360 /* It is correct to call set_IF(regs) from the set_vflags_*
361 * functions. However someone forgot to call clear_IF(regs)
362 * in the opposite case.
363 * After the command sequence CLI PUSHF STI POPF you should
364 * end up with interrups disabled, but you ended up with
365 * interrupts enabled.
366 * ( I was testing my own changes, but the only bug I
367 * could find was in a function I had not changed. )
371 static inline void set_vflags_long(unsigned long eflags, struct kernel_vm86_regs * regs)
373 set_flags(VEFLAGS, eflags, current->thread.v86mask);
374 set_flags(regs->eflags, eflags, SAFE_MASK);
375 if (eflags & IF_MASK)
381 static inline void set_vflags_short(unsigned short flags, struct kernel_vm86_regs * regs)
383 set_flags(VFLAGS, flags, current->thread.v86mask);
384 set_flags(regs->eflags, flags, SAFE_MASK);
391 static inline unsigned long get_vflags(struct kernel_vm86_regs * regs)
393 unsigned long flags = regs->eflags & RETURN_MASK;
395 if (VEFLAGS & VIF_MASK)
398 return flags | (VEFLAGS & current->thread.v86mask);
401 static inline int is_revectored(int nr, struct revectored_struct * bitmap)
403 __asm__ __volatile__("btl %2,%1\n\tsbbl %0,%0"
405 :"m" (*bitmap),"r" (nr));
409 #define val_byte(val, n) (((__u8 *)&val)[n])
411 #define pushb(base, ptr, val, err_label) \
415 if (put_user(__val, base + ptr) < 0) \
419 #define pushw(base, ptr, val, err_label) \
423 if (put_user(val_byte(__val, 1), base + ptr) < 0) \
426 if (put_user(val_byte(__val, 0), base + ptr) < 0) \
430 #define pushl(base, ptr, val, err_label) \
434 if (put_user(val_byte(__val, 3), base + ptr) < 0) \
437 if (put_user(val_byte(__val, 2), base + ptr) < 0) \
440 if (put_user(val_byte(__val, 1), base + ptr) < 0) \
443 if (put_user(val_byte(__val, 0), base + ptr) < 0) \
447 #define popb(base, ptr, err_label) \
450 if (get_user(__res, base + ptr) < 0) \
456 #define popw(base, ptr, err_label) \
459 if (get_user(val_byte(__res, 0), base + ptr) < 0) \
462 if (get_user(val_byte(__res, 1), base + ptr) < 0) \
468 #define popl(base, ptr, err_label) \
471 if (get_user(val_byte(__res, 0), base + ptr) < 0) \
474 if (get_user(val_byte(__res, 1), base + ptr) < 0) \
477 if (get_user(val_byte(__res, 2), base + ptr) < 0) \
480 if (get_user(val_byte(__res, 3), base + ptr) < 0) \
486 /* There are so many possible reasons for this function to return
487 * VM86_INTx, so adding another doesn't bother me. We can expect
488 * userspace programs to be able to handle it. (Getting a problem
489 * in userspace is always better than an Oops anyway.) [KD]
491 static void do_int(struct kernel_vm86_regs *regs, int i,
492 unsigned char __user * ssp, unsigned short sp)
494 unsigned long __user *intr_ptr;
495 unsigned long segoffs;
497 if (regs->cs == BIOSSEG)
499 if (is_revectored(i, &KVM86->int_revectored))
501 if (i==0x21 && is_revectored(AH(regs),&KVM86->int21_revectored))
503 intr_ptr = (unsigned long __user *) (i << 2);
504 if (get_user(segoffs, intr_ptr))
506 if ((segoffs >> 16) == BIOSSEG)
508 pushw(ssp, sp, get_vflags(regs), cannot_handle);
509 pushw(ssp, sp, regs->cs, cannot_handle);
510 pushw(ssp, sp, IP(regs), cannot_handle);
511 regs->cs = segoffs >> 16;
513 IP(regs) = segoffs & 0xffff;
520 return_to_32bit(regs, VM86_INTx + (i << 8));
523 int handle_vm86_trap(struct kernel_vm86_regs * regs, long error_code, int trapno)
525 if (VMPI.is_vm86pus) {
526 if ( (trapno==3) || (trapno==1) )
527 return_to_32bit(regs, VM86_TRAP + (trapno << 8));
528 do_int(regs, trapno, (unsigned char __user *) (regs->ss << 4), SP(regs));
532 return 1; /* we let this handle by the calling routine */
533 if (current->ptrace & PT_PTRACED) {
535 spin_lock_irqsave(¤t->sighand->siglock, flags);
536 sigdelset(¤t->blocked, SIGTRAP);
538 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
540 send_sig(SIGTRAP, current, 1);
541 current->thread.trap_no = trapno;
542 current->thread.error_code = error_code;
546 void handle_vm86_fault(struct kernel_vm86_regs * regs, long error_code)
548 unsigned char opcode;
549 unsigned char __user *csp;
550 unsigned char __user *ssp;
551 unsigned short ip, sp, orig_flags;
552 int data32, pref_done;
554 #define CHECK_IF_IN_TRAP \
555 if (VMPI.vm86dbg_active && VMPI.vm86dbg_TFpendig) \
557 #define VM86_FAULT_RETURN do { \
558 if (VMPI.force_return_for_pic && (VEFLAGS & (IF_MASK | VIF_MASK))) \
559 return_to_32bit(regs, VM86_PICRETURN); \
560 if (orig_flags & TF_MASK) \
561 handle_vm86_trap(regs, 0, 1); \
564 orig_flags = *(unsigned short *)®s->eflags;
566 csp = (unsigned char __user *) (regs->cs << 4);
567 ssp = (unsigned char __user *) (regs->ss << 4);
574 switch (opcode = popb(csp, ip, simulate_sigsegv)) {
575 case 0x66: /* 32-bit data */ data32=1; break;
576 case 0x67: /* 32-bit address */ break;
577 case 0x2e: /* CS */ break;
578 case 0x3e: /* DS */ break;
579 case 0x26: /* ES */ break;
580 case 0x36: /* SS */ break;
581 case 0x65: /* GS */ break;
582 case 0x64: /* FS */ break;
583 case 0xf2: /* repnz */ break;
584 case 0xf3: /* rep */ break;
585 default: pref_done = 1;
587 } while (!pref_done);
594 pushl(ssp, sp, get_vflags(regs), simulate_sigsegv);
597 pushw(ssp, sp, get_vflags(regs), simulate_sigsegv);
606 unsigned long newflags;
608 newflags=popl(ssp, sp, simulate_sigsegv);
611 newflags = popw(ssp, sp, simulate_sigsegv);
617 set_vflags_long(newflags, regs);
619 set_vflags_short(newflags, regs);
626 int intno=popb(csp, ip, simulate_sigsegv);
628 if (VMPI.vm86dbg_active) {
629 if ( (1 << (intno &7)) & VMPI.vm86dbg_intxxtab[intno >> 3] )
630 return_to_32bit(regs, VM86_INTx + (intno << 8));
632 do_int(regs, intno, ssp, sp);
641 unsigned long newflags;
643 newip=popl(ssp, sp, simulate_sigsegv);
644 newcs=popl(ssp, sp, simulate_sigsegv);
645 newflags=popl(ssp, sp, simulate_sigsegv);
648 newip = popw(ssp, sp, simulate_sigsegv);
649 newcs = popw(ssp, sp, simulate_sigsegv);
650 newflags = popw(ssp, sp, simulate_sigsegv);
657 set_vflags_long(newflags, regs);
659 set_vflags_short(newflags, regs);
672 * Damn. This is incorrect: the 'sti' instruction should actually
673 * enable interrupts after the /next/ instruction. Not good.
675 * Probably needs some horsing around with the TF flag. Aiee..
683 return_to_32bit(regs, VM86_UNKNOWN);
689 /* FIXME: After a long discussion with Stas we finally
690 * agreed, that this is wrong. Here we should
691 * really send a SIGSEGV to the user program.
692 * But how do we create the correct context? We
693 * are inside a general protection fault handler
694 * and has just returned from a page fault handler.
695 * The correct context for the signal handler
696 * should be a mixture of the two, but how do we
697 * get the information? [KD]
699 return_to_32bit(regs, VM86_UNKNOWN);
702 /* ---------------- vm86 special IRQ passing stuff ----------------- */
704 #define VM86_IRQNAME "vm86irq"
706 static struct vm86_irqs {
707 struct task_struct *tsk;
711 static DEFINE_SPINLOCK(irqbits_lock);
714 #define ALLOWED_SIGS ( 1 /* 0 = don't send a signal */ \
715 | (1 << SIGUSR1) | (1 << SIGUSR2) | (1 << SIGIO) | (1 << SIGURG) \
718 static irqreturn_t irq_handler(int intno, void *dev_id, struct pt_regs * regs)
723 spin_lock_irqsave(&irqbits_lock, flags);
724 irq_bit = 1 << intno;
725 if ((irqbits & irq_bit) || ! vm86_irqs[intno].tsk)
728 if (vm86_irqs[intno].sig)
729 send_sig(vm86_irqs[intno].sig, vm86_irqs[intno].tsk, 1);
731 * IRQ will be re-enabled when user asks for the irq (whether
732 * polling or as a result of the signal)
734 disable_irq_nosync(intno);
735 spin_unlock_irqrestore(&irqbits_lock, flags);
739 spin_unlock_irqrestore(&irqbits_lock, flags);
743 static inline void free_vm86_irq(int irqnumber)
747 free_irq(irqnumber, NULL);
748 vm86_irqs[irqnumber].tsk = NULL;
750 spin_lock_irqsave(&irqbits_lock, flags);
751 irqbits &= ~(1 << irqnumber);
752 spin_unlock_irqrestore(&irqbits_lock, flags);
755 void release_vm86_irqs(struct task_struct *task)
758 for (i = FIRST_VM86_IRQ ; i <= LAST_VM86_IRQ; i++)
759 if (vm86_irqs[i].tsk == task)
763 static inline int get_and_reset_irq(int irqnumber)
769 if (invalid_vm86_irq(irqnumber)) return 0;
770 if (vm86_irqs[irqnumber].tsk != current) return 0;
771 spin_lock_irqsave(&irqbits_lock, flags);
772 bit = irqbits & (1 << irqnumber);
775 enable_irq(irqnumber);
779 spin_unlock_irqrestore(&irqbits_lock, flags);
784 static int do_vm86_irq_handling(int subfunction, int irqnumber)
787 switch (subfunction) {
788 case VM86_GET_AND_RESET_IRQ: {
789 return get_and_reset_irq(irqnumber);
791 case VM86_GET_IRQ_BITS: {
794 case VM86_REQUEST_IRQ: {
795 int sig = irqnumber >> 8;
796 int irq = irqnumber & 255;
797 if (!capable(CAP_SYS_ADMIN)) return -EPERM;
798 if (!((1 << sig) & ALLOWED_SIGS)) return -EPERM;
799 if (invalid_vm86_irq(irq)) return -EPERM;
800 if (vm86_irqs[irq].tsk) return -EPERM;
801 ret = request_irq(irq, &irq_handler, 0, VM86_IRQNAME, NULL);
803 vm86_irqs[irq].sig = sig;
804 vm86_irqs[irq].tsk = current;
807 case VM86_FREE_IRQ: {
808 if (invalid_vm86_irq(irqnumber)) return -EPERM;
809 if (!vm86_irqs[irqnumber].tsk) return 0;
810 if (vm86_irqs[irqnumber].tsk != current) return -EPERM;
811 free_vm86_irq(irqnumber);