xref: /linux/arch/x86/kernel/vm86_32.c (revision c0e297dc61f8d4453e07afbea1fa8d0e67cd4a34)
1 /*
2  *  Copyright (C) 1994  Linus Torvalds
3  *
4  *  29 dec 2001 - Fixed oopses caused by unchecked access to the vm86
5  *                stack - Manfred Spraul <manfred@colorfullife.com>
6  *
7  *  22 mar 2002 - Manfred detected the stackfaults, but didn't handle
8  *                them correctly. Now the emulation will be in a
9  *                consistent state after stackfaults - Kasper Dupont
10  *                <kasperd@daimi.au.dk>
11  *
12  *  22 mar 2002 - Added missing clear_IF in set_vflags_* Kasper Dupont
13  *                <kasperd@daimi.au.dk>
14  *
15  *  ?? ??? 2002 - Fixed premature returns from handle_vm86_fault
16  *                caused by Kasper Dupont's changes - Stas Sergeev
17  *
18  *   4 apr 2002 - Fixed CHECK_IF_IN_TRAP broken by Stas' changes.
19  *                Kasper Dupont <kasperd@daimi.au.dk>
20  *
21  *   9 apr 2002 - Changed syntax of macros in handle_vm86_fault.
22  *                Kasper Dupont <kasperd@daimi.au.dk>
23  *
24  *   9 apr 2002 - Changed stack access macros to jump to a label
25  *                instead of returning to userspace. This simplifies
26  *                do_int, and is needed by handle_vm6_fault. Kasper
27  *                Dupont <kasperd@daimi.au.dk>
28  *
29  */
30 
31 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
32 
33 #include <linux/capability.h>
34 #include <linux/errno.h>
35 #include <linux/interrupt.h>
36 #include <linux/syscalls.h>
37 #include <linux/sched.h>
38 #include <linux/kernel.h>
39 #include <linux/signal.h>
40 #include <linux/string.h>
41 #include <linux/mm.h>
42 #include <linux/smp.h>
43 #include <linux/highmem.h>
44 #include <linux/ptrace.h>
45 #include <linux/audit.h>
46 #include <linux/stddef.h>
47 
48 #include <asm/uaccess.h>
49 #include <asm/io.h>
50 #include <asm/tlbflush.h>
51 #include <asm/irq.h>
52 
53 /*
54  * Known problems:
55  *
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
63  *   details yet.
64  *
65  * Let's hope these problems do not actually matter for anything.
66  */
67 
68 
69 #define KVM86	((struct kernel_vm86_struct *)regs)
70 #define VMPI	KVM86->vm86plus
71 
72 
73 /*
74  * 8- and 16-bit register defines..
75  */
76 #define AL(regs)	(((unsigned char *)&((regs)->pt.ax))[0])
77 #define AH(regs)	(((unsigned char *)&((regs)->pt.ax))[1])
78 #define IP(regs)	(*(unsigned short *)&((regs)->pt.ip))
79 #define SP(regs)	(*(unsigned short *)&((regs)->pt.sp))
80 
81 /*
82  * virtual flags (16 and 32-bit versions)
83  */
84 #define VFLAGS	(*(unsigned short *)&(current->thread.v86flags))
85 #define VEFLAGS	(current->thread.v86flags)
86 
87 #define set_flags(X, new, mask) \
88 ((X) = ((X) & ~(mask)) | ((new) & (mask)))
89 
90 #define SAFE_MASK	(0xDD5)
91 #define RETURN_MASK	(0xDFF)
92 
93 /* convert kernel_vm86_regs to vm86_regs */
94 static int copy_vm86_regs_to_user(struct vm86_regs __user *user,
95 				  const struct kernel_vm86_regs *regs)
96 {
97 	int ret = 0;
98 
99 	/*
100 	 * kernel_vm86_regs is missing gs, so copy everything up to
101 	 * (but not including) orig_eax, and then rest including orig_eax.
102 	 */
103 	ret += copy_to_user(user, regs, offsetof(struct kernel_vm86_regs, pt.orig_ax));
104 	ret += copy_to_user(&user->orig_eax, &regs->pt.orig_ax,
105 			    sizeof(struct kernel_vm86_regs) -
106 			    offsetof(struct kernel_vm86_regs, pt.orig_ax));
107 
108 	return ret;
109 }
110 
111 /* convert vm86_regs to kernel_vm86_regs */
112 static int copy_vm86_regs_from_user(struct kernel_vm86_regs *regs,
113 				    const struct vm86_regs __user *user,
114 				    unsigned extra)
115 {
116 	int ret = 0;
117 
118 	/* copy ax-fs inclusive */
119 	ret += copy_from_user(regs, user, offsetof(struct kernel_vm86_regs, pt.orig_ax));
120 	/* copy orig_ax-__gsh+extra */
121 	ret += copy_from_user(&regs->pt.orig_ax, &user->orig_eax,
122 			      sizeof(struct kernel_vm86_regs) -
123 			      offsetof(struct kernel_vm86_regs, pt.orig_ax) +
124 			      extra);
125 	return ret;
126 }
127 
128 struct pt_regs *save_v86_state(struct kernel_vm86_regs *regs)
129 {
130 	struct tss_struct *tss;
131 	struct pt_regs *ret;
132 	unsigned long tmp;
133 
134 	/*
135 	 * This gets called from entry.S with interrupts disabled, but
136 	 * from process context. Enable interrupts here, before trying
137 	 * to access user space.
138 	 */
139 	local_irq_enable();
140 
141 	if (!current->thread.vm86_info) {
142 		pr_alert("no vm86_info: BAD\n");
143 		do_exit(SIGSEGV);
144 	}
145 	set_flags(regs->pt.flags, VEFLAGS, X86_EFLAGS_VIF | current->thread.v86mask);
146 	tmp = copy_vm86_regs_to_user(&current->thread.vm86_info->regs, regs);
147 	tmp += put_user(current->thread.screen_bitmap, &current->thread.vm86_info->screen_bitmap);
148 	if (tmp) {
149 		pr_alert("could not access userspace vm86_info\n");
150 		do_exit(SIGSEGV);
151 	}
152 
153 	tss = &per_cpu(cpu_tss, get_cpu());
154 	current->thread.sp0 = current->thread.saved_sp0;
155 	current->thread.sysenter_cs = __KERNEL_CS;
156 	load_sp0(tss, &current->thread);
157 	current->thread.saved_sp0 = 0;
158 	put_cpu();
159 
160 	ret = KVM86->regs32;
161 
162 	ret->fs = current->thread.saved_fs;
163 	set_user_gs(ret, current->thread.saved_gs);
164 
165 	return ret;
166 }
167 
168 static void mark_screen_rdonly(struct mm_struct *mm)
169 {
170 	pgd_t *pgd;
171 	pud_t *pud;
172 	pmd_t *pmd;
173 	pte_t *pte;
174 	spinlock_t *ptl;
175 	int i;
176 
177 	down_write(&mm->mmap_sem);
178 	pgd = pgd_offset(mm, 0xA0000);
179 	if (pgd_none_or_clear_bad(pgd))
180 		goto out;
181 	pud = pud_offset(pgd, 0xA0000);
182 	if (pud_none_or_clear_bad(pud))
183 		goto out;
184 	pmd = pmd_offset(pud, 0xA0000);
185 	split_huge_page_pmd_mm(mm, 0xA0000, pmd);
186 	if (pmd_none_or_clear_bad(pmd))
187 		goto out;
188 	pte = pte_offset_map_lock(mm, pmd, 0xA0000, &ptl);
189 	for (i = 0; i < 32; i++) {
190 		if (pte_present(*pte))
191 			set_pte(pte, pte_wrprotect(*pte));
192 		pte++;
193 	}
194 	pte_unmap_unlock(pte, ptl);
195 out:
196 	up_write(&mm->mmap_sem);
197 	flush_tlb();
198 }
199 
200 
201 
202 static int do_vm86_irq_handling(int subfunction, int irqnumber);
203 static void do_sys_vm86(struct kernel_vm86_struct *info, struct task_struct *tsk);
204 
205 SYSCALL_DEFINE1(vm86old, struct vm86_struct __user *, v86)
206 {
207 	struct kernel_vm86_struct info; /* declare this _on top_,
208 					 * this avoids wasting of stack space.
209 					 * This remains on the stack until we
210 					 * return to 32 bit user space.
211 					 */
212 	struct task_struct *tsk = current;
213 	int tmp;
214 
215 	if (tsk->thread.saved_sp0)
216 		return -EPERM;
217 	tmp = copy_vm86_regs_from_user(&info.regs, &v86->regs,
218 				       offsetof(struct kernel_vm86_struct, vm86plus) -
219 				       sizeof(info.regs));
220 	if (tmp)
221 		return -EFAULT;
222 	memset(&info.vm86plus, 0, (int)&info.regs32 - (int)&info.vm86plus);
223 	info.regs32 = current_pt_regs();
224 	tsk->thread.vm86_info = v86;
225 	do_sys_vm86(&info, tsk);
226 	return 0;	/* we never return here */
227 }
228 
229 
230 SYSCALL_DEFINE2(vm86, unsigned long, cmd, unsigned long, arg)
231 {
232 	struct kernel_vm86_struct info; /* declare this _on top_,
233 					 * this avoids wasting of stack space.
234 					 * This remains on the stack until we
235 					 * return to 32 bit user space.
236 					 */
237 	struct task_struct *tsk;
238 	int tmp;
239 	struct vm86plus_struct __user *v86;
240 
241 	tsk = current;
242 	switch (cmd) {
243 	case VM86_REQUEST_IRQ:
244 	case VM86_FREE_IRQ:
245 	case VM86_GET_IRQ_BITS:
246 	case VM86_GET_AND_RESET_IRQ:
247 		return do_vm86_irq_handling(cmd, (int)arg);
248 	case VM86_PLUS_INSTALL_CHECK:
249 		/*
250 		 * NOTE: on old vm86 stuff this will return the error
251 		 *  from access_ok(), because the subfunction is
252 		 *  interpreted as (invalid) address to vm86_struct.
253 		 *  So the installation check works.
254 		 */
255 		return 0;
256 	}
257 
258 	/* we come here only for functions VM86_ENTER, VM86_ENTER_NO_BYPASS */
259 	if (tsk->thread.saved_sp0)
260 		return -EPERM;
261 	v86 = (struct vm86plus_struct __user *)arg;
262 	tmp = copy_vm86_regs_from_user(&info.regs, &v86->regs,
263 				       offsetof(struct kernel_vm86_struct, regs32) -
264 				       sizeof(info.regs));
265 	if (tmp)
266 		return -EFAULT;
267 	info.regs32 = current_pt_regs();
268 	info.vm86plus.is_vm86pus = 1;
269 	tsk->thread.vm86_info = (struct vm86_struct __user *)v86;
270 	do_sys_vm86(&info, tsk);
271 	return 0;	/* we never return here */
272 }
273 
274 
275 static void do_sys_vm86(struct kernel_vm86_struct *info, struct task_struct *tsk)
276 {
277 	struct tss_struct *tss;
278 /*
279  * make sure the vm86() system call doesn't try to do anything silly
280  */
281 	info->regs.pt.ds = 0;
282 	info->regs.pt.es = 0;
283 	info->regs.pt.fs = 0;
284 #ifndef CONFIG_X86_32_LAZY_GS
285 	info->regs.pt.gs = 0;
286 #endif
287 
288 /*
289  * The flags register is also special: we cannot trust that the user
290  * has set it up safely, so this makes sure interrupt etc flags are
291  * inherited from protected mode.
292  */
293 	VEFLAGS = info->regs.pt.flags;
294 	info->regs.pt.flags &= SAFE_MASK;
295 	info->regs.pt.flags |= info->regs32->flags & ~SAFE_MASK;
296 	info->regs.pt.flags |= X86_VM_MASK;
297 
298 	switch (info->cpu_type) {
299 	case CPU_286:
300 		tsk->thread.v86mask = 0;
301 		break;
302 	case CPU_386:
303 		tsk->thread.v86mask = X86_EFLAGS_NT | X86_EFLAGS_IOPL;
304 		break;
305 	case CPU_486:
306 		tsk->thread.v86mask = X86_EFLAGS_AC | X86_EFLAGS_NT | X86_EFLAGS_IOPL;
307 		break;
308 	default:
309 		tsk->thread.v86mask = X86_EFLAGS_ID | X86_EFLAGS_AC | X86_EFLAGS_NT | X86_EFLAGS_IOPL;
310 		break;
311 	}
312 
313 /*
314  * Save old state, set default return value (%ax) to 0 (VM86_SIGNAL)
315  */
316 	info->regs32->ax = VM86_SIGNAL;
317 	tsk->thread.saved_sp0 = tsk->thread.sp0;
318 	tsk->thread.saved_fs = info->regs32->fs;
319 	tsk->thread.saved_gs = get_user_gs(info->regs32);
320 
321 	tss = &per_cpu(cpu_tss, get_cpu());
322 	tsk->thread.sp0 = (unsigned long) &info->VM86_TSS_ESP0;
323 	if (cpu_has_sep)
324 		tsk->thread.sysenter_cs = 0;
325 	load_sp0(tss, &tsk->thread);
326 	put_cpu();
327 
328 	tsk->thread.screen_bitmap = info->screen_bitmap;
329 	if (info->flags & VM86_SCREEN_BITMAP)
330 		mark_screen_rdonly(tsk->mm);
331 
332 	/*call __audit_syscall_exit since we do not exit via the normal paths */
333 #ifdef CONFIG_AUDITSYSCALL
334 	if (unlikely(current->audit_context))
335 		__audit_syscall_exit(1, 0);
336 #endif
337 
338 	__asm__ __volatile__(
339 		"movl %0,%%esp\n\t"
340 		"movl %1,%%ebp\n\t"
341 #ifdef CONFIG_X86_32_LAZY_GS
342 		"mov  %2, %%gs\n\t"
343 #endif
344 		"jmp resume_userspace"
345 		: /* no outputs */
346 		:"r" (&info->regs), "r" (task_thread_info(tsk)), "r" (0));
347 	/* we never return here */
348 }
349 
350 static inline void return_to_32bit(struct kernel_vm86_regs *regs16, int retval)
351 {
352 	struct pt_regs *regs32;
353 
354 	regs32 = save_v86_state(regs16);
355 	regs32->ax = retval;
356 	__asm__ __volatile__("movl %0,%%esp\n\t"
357 		"movl %1,%%ebp\n\t"
358 		"jmp resume_userspace"
359 		: : "r" (regs32), "r" (current_thread_info()));
360 }
361 
362 static inline void set_IF(struct kernel_vm86_regs *regs)
363 {
364 	VEFLAGS |= X86_EFLAGS_VIF;
365 	if (VEFLAGS & X86_EFLAGS_VIP)
366 		return_to_32bit(regs, VM86_STI);
367 }
368 
369 static inline void clear_IF(struct kernel_vm86_regs *regs)
370 {
371 	VEFLAGS &= ~X86_EFLAGS_VIF;
372 }
373 
374 static inline void clear_TF(struct kernel_vm86_regs *regs)
375 {
376 	regs->pt.flags &= ~X86_EFLAGS_TF;
377 }
378 
379 static inline void clear_AC(struct kernel_vm86_regs *regs)
380 {
381 	regs->pt.flags &= ~X86_EFLAGS_AC;
382 }
383 
384 /*
385  * It is correct to call set_IF(regs) from the set_vflags_*
386  * functions. However someone forgot to call clear_IF(regs)
387  * in the opposite case.
388  * After the command sequence CLI PUSHF STI POPF you should
389  * end up with interrupts disabled, but you ended up with
390  * interrupts enabled.
391  *  ( I was testing my own changes, but the only bug I
392  *    could find was in a function I had not changed. )
393  * [KD]
394  */
395 
396 static inline void set_vflags_long(unsigned long flags, struct kernel_vm86_regs *regs)
397 {
398 	set_flags(VEFLAGS, flags, current->thread.v86mask);
399 	set_flags(regs->pt.flags, flags, SAFE_MASK);
400 	if (flags & X86_EFLAGS_IF)
401 		set_IF(regs);
402 	else
403 		clear_IF(regs);
404 }
405 
406 static inline void set_vflags_short(unsigned short flags, struct kernel_vm86_regs *regs)
407 {
408 	set_flags(VFLAGS, flags, current->thread.v86mask);
409 	set_flags(regs->pt.flags, flags, SAFE_MASK);
410 	if (flags & X86_EFLAGS_IF)
411 		set_IF(regs);
412 	else
413 		clear_IF(regs);
414 }
415 
416 static inline unsigned long get_vflags(struct kernel_vm86_regs *regs)
417 {
418 	unsigned long flags = regs->pt.flags & RETURN_MASK;
419 
420 	if (VEFLAGS & X86_EFLAGS_VIF)
421 		flags |= X86_EFLAGS_IF;
422 	flags |= X86_EFLAGS_IOPL;
423 	return flags | (VEFLAGS & current->thread.v86mask);
424 }
425 
426 static inline int is_revectored(int nr, struct revectored_struct *bitmap)
427 {
428 	__asm__ __volatile__("btl %2,%1\n\tsbbl %0,%0"
429 		:"=r" (nr)
430 		:"m" (*bitmap), "r" (nr));
431 	return nr;
432 }
433 
434 #define val_byte(val, n) (((__u8 *)&val)[n])
435 
436 #define pushb(base, ptr, val, err_label) \
437 	do { \
438 		__u8 __val = val; \
439 		ptr--; \
440 		if (put_user(__val, base + ptr) < 0) \
441 			goto err_label; \
442 	} while (0)
443 
444 #define pushw(base, ptr, val, err_label) \
445 	do { \
446 		__u16 __val = val; \
447 		ptr--; \
448 		if (put_user(val_byte(__val, 1), base + ptr) < 0) \
449 			goto err_label; \
450 		ptr--; \
451 		if (put_user(val_byte(__val, 0), base + ptr) < 0) \
452 			goto err_label; \
453 	} while (0)
454 
455 #define pushl(base, ptr, val, err_label) \
456 	do { \
457 		__u32 __val = val; \
458 		ptr--; \
459 		if (put_user(val_byte(__val, 3), base + ptr) < 0) \
460 			goto err_label; \
461 		ptr--; \
462 		if (put_user(val_byte(__val, 2), base + ptr) < 0) \
463 			goto err_label; \
464 		ptr--; \
465 		if (put_user(val_byte(__val, 1), base + ptr) < 0) \
466 			goto err_label; \
467 		ptr--; \
468 		if (put_user(val_byte(__val, 0), base + ptr) < 0) \
469 			goto err_label; \
470 	} while (0)
471 
472 #define popb(base, ptr, err_label) \
473 	({ \
474 		__u8 __res; \
475 		if (get_user(__res, base + ptr) < 0) \
476 			goto err_label; \
477 		ptr++; \
478 		__res; \
479 	})
480 
481 #define popw(base, ptr, err_label) \
482 	({ \
483 		__u16 __res; \
484 		if (get_user(val_byte(__res, 0), base + ptr) < 0) \
485 			goto err_label; \
486 		ptr++; \
487 		if (get_user(val_byte(__res, 1), base + ptr) < 0) \
488 			goto err_label; \
489 		ptr++; \
490 		__res; \
491 	})
492 
493 #define popl(base, ptr, err_label) \
494 	({ \
495 		__u32 __res; \
496 		if (get_user(val_byte(__res, 0), base + ptr) < 0) \
497 			goto err_label; \
498 		ptr++; \
499 		if (get_user(val_byte(__res, 1), base + ptr) < 0) \
500 			goto err_label; \
501 		ptr++; \
502 		if (get_user(val_byte(__res, 2), base + ptr) < 0) \
503 			goto err_label; \
504 		ptr++; \
505 		if (get_user(val_byte(__res, 3), base + ptr) < 0) \
506 			goto err_label; \
507 		ptr++; \
508 		__res; \
509 	})
510 
511 /* There are so many possible reasons for this function to return
512  * VM86_INTx, so adding another doesn't bother me. We can expect
513  * userspace programs to be able to handle it. (Getting a problem
514  * in userspace is always better than an Oops anyway.) [KD]
515  */
516 static void do_int(struct kernel_vm86_regs *regs, int i,
517     unsigned char __user *ssp, unsigned short sp)
518 {
519 	unsigned long __user *intr_ptr;
520 	unsigned long segoffs;
521 
522 	if (regs->pt.cs == BIOSSEG)
523 		goto cannot_handle;
524 	if (is_revectored(i, &KVM86->int_revectored))
525 		goto cannot_handle;
526 	if (i == 0x21 && is_revectored(AH(regs), &KVM86->int21_revectored))
527 		goto cannot_handle;
528 	intr_ptr = (unsigned long __user *) (i << 2);
529 	if (get_user(segoffs, intr_ptr))
530 		goto cannot_handle;
531 	if ((segoffs >> 16) == BIOSSEG)
532 		goto cannot_handle;
533 	pushw(ssp, sp, get_vflags(regs), cannot_handle);
534 	pushw(ssp, sp, regs->pt.cs, cannot_handle);
535 	pushw(ssp, sp, IP(regs), cannot_handle);
536 	regs->pt.cs = segoffs >> 16;
537 	SP(regs) -= 6;
538 	IP(regs) = segoffs & 0xffff;
539 	clear_TF(regs);
540 	clear_IF(regs);
541 	clear_AC(regs);
542 	return;
543 
544 cannot_handle:
545 	return_to_32bit(regs, VM86_INTx + (i << 8));
546 }
547 
548 int handle_vm86_trap(struct kernel_vm86_regs *regs, long error_code, int trapno)
549 {
550 	if (VMPI.is_vm86pus) {
551 		if ((trapno == 3) || (trapno == 1)) {
552 			KVM86->regs32->ax = VM86_TRAP + (trapno << 8);
553 			/* setting this flag forces the code in entry_32.S to
554 			   the path where we call save_v86_state() and change
555 			   the stack pointer to KVM86->regs32 */
556 			set_thread_flag(TIF_NOTIFY_RESUME);
557 			return 0;
558 		}
559 		do_int(regs, trapno, (unsigned char __user *) (regs->pt.ss << 4), SP(regs));
560 		return 0;
561 	}
562 	if (trapno != 1)
563 		return 1; /* we let this handle by the calling routine */
564 	current->thread.trap_nr = trapno;
565 	current->thread.error_code = error_code;
566 	force_sig(SIGTRAP, current);
567 	return 0;
568 }
569 
570 void handle_vm86_fault(struct kernel_vm86_regs *regs, long error_code)
571 {
572 	unsigned char opcode;
573 	unsigned char __user *csp;
574 	unsigned char __user *ssp;
575 	unsigned short ip, sp, orig_flags;
576 	int data32, pref_done;
577 
578 #define CHECK_IF_IN_TRAP \
579 	if (VMPI.vm86dbg_active && VMPI.vm86dbg_TFpendig) \
580 		newflags |= X86_EFLAGS_TF
581 #define VM86_FAULT_RETURN do { \
582 	if (VMPI.force_return_for_pic  && (VEFLAGS & (X86_EFLAGS_IF | X86_EFLAGS_VIF))) \
583 		return_to_32bit(regs, VM86_PICRETURN); \
584 	if (orig_flags & X86_EFLAGS_TF) \
585 		handle_vm86_trap(regs, 0, 1); \
586 	return; } while (0)
587 
588 	orig_flags = *(unsigned short *)&regs->pt.flags;
589 
590 	csp = (unsigned char __user *) (regs->pt.cs << 4);
591 	ssp = (unsigned char __user *) (regs->pt.ss << 4);
592 	sp = SP(regs);
593 	ip = IP(regs);
594 
595 	data32 = 0;
596 	pref_done = 0;
597 	do {
598 		switch (opcode = popb(csp, ip, simulate_sigsegv)) {
599 		case 0x66:      /* 32-bit data */     data32 = 1; break;
600 		case 0x67:      /* 32-bit address */  break;
601 		case 0x2e:      /* CS */              break;
602 		case 0x3e:      /* DS */              break;
603 		case 0x26:      /* ES */              break;
604 		case 0x36:      /* SS */              break;
605 		case 0x65:      /* GS */              break;
606 		case 0x64:      /* FS */              break;
607 		case 0xf2:      /* repnz */       break;
608 		case 0xf3:      /* rep */             break;
609 		default: pref_done = 1;
610 		}
611 	} while (!pref_done);
612 
613 	switch (opcode) {
614 
615 	/* pushf */
616 	case 0x9c:
617 		if (data32) {
618 			pushl(ssp, sp, get_vflags(regs), simulate_sigsegv);
619 			SP(regs) -= 4;
620 		} else {
621 			pushw(ssp, sp, get_vflags(regs), simulate_sigsegv);
622 			SP(regs) -= 2;
623 		}
624 		IP(regs) = ip;
625 		VM86_FAULT_RETURN;
626 
627 	/* popf */
628 	case 0x9d:
629 		{
630 		unsigned long newflags;
631 		if (data32) {
632 			newflags = popl(ssp, sp, simulate_sigsegv);
633 			SP(regs) += 4;
634 		} else {
635 			newflags = popw(ssp, sp, simulate_sigsegv);
636 			SP(regs) += 2;
637 		}
638 		IP(regs) = ip;
639 		CHECK_IF_IN_TRAP;
640 		if (data32)
641 			set_vflags_long(newflags, regs);
642 		else
643 			set_vflags_short(newflags, regs);
644 
645 		VM86_FAULT_RETURN;
646 		}
647 
648 	/* int xx */
649 	case 0xcd: {
650 		int intno = popb(csp, ip, simulate_sigsegv);
651 		IP(regs) = ip;
652 		if (VMPI.vm86dbg_active) {
653 			if ((1 << (intno & 7)) & VMPI.vm86dbg_intxxtab[intno >> 3])
654 				return_to_32bit(regs, VM86_INTx + (intno << 8));
655 		}
656 		do_int(regs, intno, ssp, sp);
657 		return;
658 	}
659 
660 	/* iret */
661 	case 0xcf:
662 		{
663 		unsigned long newip;
664 		unsigned long newcs;
665 		unsigned long newflags;
666 		if (data32) {
667 			newip = popl(ssp, sp, simulate_sigsegv);
668 			newcs = popl(ssp, sp, simulate_sigsegv);
669 			newflags = popl(ssp, sp, simulate_sigsegv);
670 			SP(regs) += 12;
671 		} else {
672 			newip = popw(ssp, sp, simulate_sigsegv);
673 			newcs = popw(ssp, sp, simulate_sigsegv);
674 			newflags = popw(ssp, sp, simulate_sigsegv);
675 			SP(regs) += 6;
676 		}
677 		IP(regs) = newip;
678 		regs->pt.cs = newcs;
679 		CHECK_IF_IN_TRAP;
680 		if (data32) {
681 			set_vflags_long(newflags, regs);
682 		} else {
683 			set_vflags_short(newflags, regs);
684 		}
685 		VM86_FAULT_RETURN;
686 		}
687 
688 	/* cli */
689 	case 0xfa:
690 		IP(regs) = ip;
691 		clear_IF(regs);
692 		VM86_FAULT_RETURN;
693 
694 	/* sti */
695 	/*
696 	 * Damn. This is incorrect: the 'sti' instruction should actually
697 	 * enable interrupts after the /next/ instruction. Not good.
698 	 *
699 	 * Probably needs some horsing around with the TF flag. Aiee..
700 	 */
701 	case 0xfb:
702 		IP(regs) = ip;
703 		set_IF(regs);
704 		VM86_FAULT_RETURN;
705 
706 	default:
707 		return_to_32bit(regs, VM86_UNKNOWN);
708 	}
709 
710 	return;
711 
712 simulate_sigsegv:
713 	/* FIXME: After a long discussion with Stas we finally
714 	 *        agreed, that this is wrong. Here we should
715 	 *        really send a SIGSEGV to the user program.
716 	 *        But how do we create the correct context? We
717 	 *        are inside a general protection fault handler
718 	 *        and has just returned from a page fault handler.
719 	 *        The correct context for the signal handler
720 	 *        should be a mixture of the two, but how do we
721 	 *        get the information? [KD]
722 	 */
723 	return_to_32bit(regs, VM86_UNKNOWN);
724 }
725 
726 /* ---------------- vm86 special IRQ passing stuff ----------------- */
727 
728 #define VM86_IRQNAME		"vm86irq"
729 
730 static struct vm86_irqs {
731 	struct task_struct *tsk;
732 	int sig;
733 } vm86_irqs[16];
734 
735 static DEFINE_SPINLOCK(irqbits_lock);
736 static int irqbits;
737 
738 #define ALLOWED_SIGS (1 /* 0 = don't send a signal */ \
739 	| (1 << SIGUSR1) | (1 << SIGUSR2) | (1 << SIGIO)  | (1 << SIGURG) \
740 	| (1 << SIGUNUSED))
741 
742 static irqreturn_t irq_handler(int intno, void *dev_id)
743 {
744 	int irq_bit;
745 	unsigned long flags;
746 
747 	spin_lock_irqsave(&irqbits_lock, flags);
748 	irq_bit = 1 << intno;
749 	if ((irqbits & irq_bit) || !vm86_irqs[intno].tsk)
750 		goto out;
751 	irqbits |= irq_bit;
752 	if (vm86_irqs[intno].sig)
753 		send_sig(vm86_irqs[intno].sig, vm86_irqs[intno].tsk, 1);
754 	/*
755 	 * IRQ will be re-enabled when user asks for the irq (whether
756 	 * polling or as a result of the signal)
757 	 */
758 	disable_irq_nosync(intno);
759 	spin_unlock_irqrestore(&irqbits_lock, flags);
760 	return IRQ_HANDLED;
761 
762 out:
763 	spin_unlock_irqrestore(&irqbits_lock, flags);
764 	return IRQ_NONE;
765 }
766 
767 static inline void free_vm86_irq(int irqnumber)
768 {
769 	unsigned long flags;
770 
771 	free_irq(irqnumber, NULL);
772 	vm86_irqs[irqnumber].tsk = NULL;
773 
774 	spin_lock_irqsave(&irqbits_lock, flags);
775 	irqbits &= ~(1 << irqnumber);
776 	spin_unlock_irqrestore(&irqbits_lock, flags);
777 }
778 
779 void release_vm86_irqs(struct task_struct *task)
780 {
781 	int i;
782 	for (i = FIRST_VM86_IRQ ; i <= LAST_VM86_IRQ; i++)
783 	    if (vm86_irqs[i].tsk == task)
784 		free_vm86_irq(i);
785 }
786 
787 static inline int get_and_reset_irq(int irqnumber)
788 {
789 	int bit;
790 	unsigned long flags;
791 	int ret = 0;
792 
793 	if (invalid_vm86_irq(irqnumber)) return 0;
794 	if (vm86_irqs[irqnumber].tsk != current) return 0;
795 	spin_lock_irqsave(&irqbits_lock, flags);
796 	bit = irqbits & (1 << irqnumber);
797 	irqbits &= ~bit;
798 	if (bit) {
799 		enable_irq(irqnumber);
800 		ret = 1;
801 	}
802 
803 	spin_unlock_irqrestore(&irqbits_lock, flags);
804 	return ret;
805 }
806 
807 
808 static int do_vm86_irq_handling(int subfunction, int irqnumber)
809 {
810 	int ret;
811 	switch (subfunction) {
812 		case VM86_GET_AND_RESET_IRQ: {
813 			return get_and_reset_irq(irqnumber);
814 		}
815 		case VM86_GET_IRQ_BITS: {
816 			return irqbits;
817 		}
818 		case VM86_REQUEST_IRQ: {
819 			int sig = irqnumber >> 8;
820 			int irq = irqnumber & 255;
821 			if (!capable(CAP_SYS_ADMIN)) return -EPERM;
822 			if (!((1 << sig) & ALLOWED_SIGS)) return -EPERM;
823 			if (invalid_vm86_irq(irq)) return -EPERM;
824 			if (vm86_irqs[irq].tsk) return -EPERM;
825 			ret = request_irq(irq, &irq_handler, 0, VM86_IRQNAME, NULL);
826 			if (ret) return ret;
827 			vm86_irqs[irq].sig = sig;
828 			vm86_irqs[irq].tsk = current;
829 			return irq;
830 		}
831 		case  VM86_FREE_IRQ: {
832 			if (invalid_vm86_irq(irqnumber)) return -EPERM;
833 			if (!vm86_irqs[irqnumber].tsk) return 0;
834 			if (vm86_irqs[irqnumber].tsk != current) return -EPERM;
835 			free_vm86_irq(irqnumber);
836 			return 0;
837 		}
838 	}
839 	return -EINVAL;
840 }
841 
842