xref: /openbmc/linux/arch/x86/kernel/vm86_32.c (revision bc5aa3a0)
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 #include <linux/slab.h>
48 #include <linux/security.h>
49 
50 #include <asm/uaccess.h>
51 #include <asm/io.h>
52 #include <asm/tlbflush.h>
53 #include <asm/irq.h>
54 #include <asm/traps.h>
55 #include <asm/vm86.h>
56 
57 /*
58  * Known problems:
59  *
60  * Interrupt handling is not guaranteed:
61  * - a real x86 will disable all interrupts for one instruction
62  *   after a "mov ss,xx" to make stack handling atomic even without
63  *   the 'lss' instruction. We can't guarantee this in v86 mode,
64  *   as the next instruction might result in a page fault or similar.
65  * - a real x86 will have interrupts disabled for one instruction
66  *   past the 'sti' that enables them. We don't bother with all the
67  *   details yet.
68  *
69  * Let's hope these problems do not actually matter for anything.
70  */
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.vm86->veflags))
85 #define VEFLAGS	(current->thread.vm86->veflags)
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 void save_v86_state(struct kernel_vm86_regs *regs, int retval)
94 {
95 	struct tss_struct *tss;
96 	struct task_struct *tsk = current;
97 	struct vm86plus_struct __user *user;
98 	struct vm86 *vm86 = current->thread.vm86;
99 	long err = 0;
100 
101 	/*
102 	 * This gets called from entry.S with interrupts disabled, but
103 	 * from process context. Enable interrupts here, before trying
104 	 * to access user space.
105 	 */
106 	local_irq_enable();
107 
108 	if (!vm86 || !vm86->user_vm86) {
109 		pr_alert("no user_vm86: BAD\n");
110 		do_exit(SIGSEGV);
111 	}
112 	set_flags(regs->pt.flags, VEFLAGS, X86_EFLAGS_VIF | vm86->veflags_mask);
113 	user = vm86->user_vm86;
114 
115 	if (!access_ok(VERIFY_WRITE, user, vm86->vm86plus.is_vm86pus ?
116 		       sizeof(struct vm86plus_struct) :
117 		       sizeof(struct vm86_struct))) {
118 		pr_alert("could not access userspace vm86 info\n");
119 		do_exit(SIGSEGV);
120 	}
121 
122 	put_user_try {
123 		put_user_ex(regs->pt.bx, &user->regs.ebx);
124 		put_user_ex(regs->pt.cx, &user->regs.ecx);
125 		put_user_ex(regs->pt.dx, &user->regs.edx);
126 		put_user_ex(regs->pt.si, &user->regs.esi);
127 		put_user_ex(regs->pt.di, &user->regs.edi);
128 		put_user_ex(regs->pt.bp, &user->regs.ebp);
129 		put_user_ex(regs->pt.ax, &user->regs.eax);
130 		put_user_ex(regs->pt.ip, &user->regs.eip);
131 		put_user_ex(regs->pt.cs, &user->regs.cs);
132 		put_user_ex(regs->pt.flags, &user->regs.eflags);
133 		put_user_ex(regs->pt.sp, &user->regs.esp);
134 		put_user_ex(regs->pt.ss, &user->regs.ss);
135 		put_user_ex(regs->es, &user->regs.es);
136 		put_user_ex(regs->ds, &user->regs.ds);
137 		put_user_ex(regs->fs, &user->regs.fs);
138 		put_user_ex(regs->gs, &user->regs.gs);
139 
140 		put_user_ex(vm86->screen_bitmap, &user->screen_bitmap);
141 	} put_user_catch(err);
142 	if (err) {
143 		pr_alert("could not access userspace vm86 info\n");
144 		do_exit(SIGSEGV);
145 	}
146 
147 	tss = &per_cpu(cpu_tss, get_cpu());
148 	tsk->thread.sp0 = vm86->saved_sp0;
149 	tsk->thread.sysenter_cs = __KERNEL_CS;
150 	load_sp0(tss, &tsk->thread);
151 	vm86->saved_sp0 = 0;
152 	put_cpu();
153 
154 	memcpy(&regs->pt, &vm86->regs32, sizeof(struct pt_regs));
155 
156 	lazy_load_gs(vm86->regs32.gs);
157 
158 	regs->pt.ax = retval;
159 }
160 
161 static void mark_screen_rdonly(struct mm_struct *mm)
162 {
163 	pgd_t *pgd;
164 	pud_t *pud;
165 	pmd_t *pmd;
166 	pte_t *pte;
167 	spinlock_t *ptl;
168 	int i;
169 
170 	down_write(&mm->mmap_sem);
171 	pgd = pgd_offset(mm, 0xA0000);
172 	if (pgd_none_or_clear_bad(pgd))
173 		goto out;
174 	pud = pud_offset(pgd, 0xA0000);
175 	if (pud_none_or_clear_bad(pud))
176 		goto out;
177 	pmd = pmd_offset(pud, 0xA0000);
178 
179 	if (pmd_trans_huge(*pmd)) {
180 		struct vm_area_struct *vma = find_vma(mm, 0xA0000);
181 		split_huge_pmd(vma, pmd, 0xA0000);
182 	}
183 	if (pmd_none_or_clear_bad(pmd))
184 		goto out;
185 	pte = pte_offset_map_lock(mm, pmd, 0xA0000, &ptl);
186 	for (i = 0; i < 32; i++) {
187 		if (pte_present(*pte))
188 			set_pte(pte, pte_wrprotect(*pte));
189 		pte++;
190 	}
191 	pte_unmap_unlock(pte, ptl);
192 out:
193 	up_write(&mm->mmap_sem);
194 	flush_tlb();
195 }
196 
197 
198 
199 static int do_vm86_irq_handling(int subfunction, int irqnumber);
200 static long do_sys_vm86(struct vm86plus_struct __user *user_vm86, bool plus);
201 
202 SYSCALL_DEFINE1(vm86old, struct vm86_struct __user *, user_vm86)
203 {
204 	return do_sys_vm86((struct vm86plus_struct __user *) user_vm86, false);
205 }
206 
207 
208 SYSCALL_DEFINE2(vm86, unsigned long, cmd, unsigned long, arg)
209 {
210 	switch (cmd) {
211 	case VM86_REQUEST_IRQ:
212 	case VM86_FREE_IRQ:
213 	case VM86_GET_IRQ_BITS:
214 	case VM86_GET_AND_RESET_IRQ:
215 		return do_vm86_irq_handling(cmd, (int)arg);
216 	case VM86_PLUS_INSTALL_CHECK:
217 		/*
218 		 * NOTE: on old vm86 stuff this will return the error
219 		 *  from access_ok(), because the subfunction is
220 		 *  interpreted as (invalid) address to vm86_struct.
221 		 *  So the installation check works.
222 		 */
223 		return 0;
224 	}
225 
226 	/* we come here only for functions VM86_ENTER, VM86_ENTER_NO_BYPASS */
227 	return do_sys_vm86((struct vm86plus_struct __user *) arg, true);
228 }
229 
230 
231 static long do_sys_vm86(struct vm86plus_struct __user *user_vm86, bool plus)
232 {
233 	struct tss_struct *tss;
234 	struct task_struct *tsk = current;
235 	struct vm86 *vm86 = tsk->thread.vm86;
236 	struct kernel_vm86_regs vm86regs;
237 	struct pt_regs *regs = current_pt_regs();
238 	unsigned long err = 0;
239 
240 	err = security_mmap_addr(0);
241 	if (err) {
242 		/*
243 		 * vm86 cannot virtualize the address space, so vm86 users
244 		 * need to manage the low 1MB themselves using mmap.  Given
245 		 * that BIOS places important data in the first page, vm86
246 		 * is essentially useless if mmap_min_addr != 0.  DOSEMU,
247 		 * for example, won't even bother trying to use vm86 if it
248 		 * can't map a page at virtual address 0.
249 		 *
250 		 * To reduce the available kernel attack surface, simply
251 		 * disallow vm86(old) for users who cannot mmap at va 0.
252 		 *
253 		 * The implementation of security_mmap_addr will allow
254 		 * suitably privileged users to map va 0 even if
255 		 * vm.mmap_min_addr is set above 0, and we want this
256 		 * behavior for vm86 as well, as it ensures that legacy
257 		 * tools like vbetool will not fail just because of
258 		 * vm.mmap_min_addr.
259 		 */
260 		pr_info_once("Denied a call to vm86(old) from %s[%d] (uid: %d).  Set the vm.mmap_min_addr sysctl to 0 and/or adjust LSM mmap_min_addr policy to enable vm86 if you are using a vm86-based DOS emulator.\n",
261 			     current->comm, task_pid_nr(current),
262 			     from_kuid_munged(&init_user_ns, current_uid()));
263 		return -EPERM;
264 	}
265 
266 	if (!vm86) {
267 		if (!(vm86 = kzalloc(sizeof(*vm86), GFP_KERNEL)))
268 			return -ENOMEM;
269 		tsk->thread.vm86 = vm86;
270 	}
271 	if (vm86->saved_sp0)
272 		return -EPERM;
273 
274 	if (!access_ok(VERIFY_READ, user_vm86, plus ?
275 		       sizeof(struct vm86_struct) :
276 		       sizeof(struct vm86plus_struct)))
277 		return -EFAULT;
278 
279 	memset(&vm86regs, 0, sizeof(vm86regs));
280 	get_user_try {
281 		unsigned short seg;
282 		get_user_ex(vm86regs.pt.bx, &user_vm86->regs.ebx);
283 		get_user_ex(vm86regs.pt.cx, &user_vm86->regs.ecx);
284 		get_user_ex(vm86regs.pt.dx, &user_vm86->regs.edx);
285 		get_user_ex(vm86regs.pt.si, &user_vm86->regs.esi);
286 		get_user_ex(vm86regs.pt.di, &user_vm86->regs.edi);
287 		get_user_ex(vm86regs.pt.bp, &user_vm86->regs.ebp);
288 		get_user_ex(vm86regs.pt.ax, &user_vm86->regs.eax);
289 		get_user_ex(vm86regs.pt.ip, &user_vm86->regs.eip);
290 		get_user_ex(seg, &user_vm86->regs.cs);
291 		vm86regs.pt.cs = seg;
292 		get_user_ex(vm86regs.pt.flags, &user_vm86->regs.eflags);
293 		get_user_ex(vm86regs.pt.sp, &user_vm86->regs.esp);
294 		get_user_ex(seg, &user_vm86->regs.ss);
295 		vm86regs.pt.ss = seg;
296 		get_user_ex(vm86regs.es, &user_vm86->regs.es);
297 		get_user_ex(vm86regs.ds, &user_vm86->regs.ds);
298 		get_user_ex(vm86regs.fs, &user_vm86->regs.fs);
299 		get_user_ex(vm86regs.gs, &user_vm86->regs.gs);
300 
301 		get_user_ex(vm86->flags, &user_vm86->flags);
302 		get_user_ex(vm86->screen_bitmap, &user_vm86->screen_bitmap);
303 		get_user_ex(vm86->cpu_type, &user_vm86->cpu_type);
304 	} get_user_catch(err);
305 	if (err)
306 		return err;
307 
308 	if (copy_from_user(&vm86->int_revectored,
309 			   &user_vm86->int_revectored,
310 			   sizeof(struct revectored_struct)))
311 		return -EFAULT;
312 	if (copy_from_user(&vm86->int21_revectored,
313 			   &user_vm86->int21_revectored,
314 			   sizeof(struct revectored_struct)))
315 		return -EFAULT;
316 	if (plus) {
317 		if (copy_from_user(&vm86->vm86plus, &user_vm86->vm86plus,
318 				   sizeof(struct vm86plus_info_struct)))
319 			return -EFAULT;
320 		vm86->vm86plus.is_vm86pus = 1;
321 	} else
322 		memset(&vm86->vm86plus, 0,
323 		       sizeof(struct vm86plus_info_struct));
324 
325 	memcpy(&vm86->regs32, regs, sizeof(struct pt_regs));
326 	vm86->user_vm86 = user_vm86;
327 
328 /*
329  * The flags register is also special: we cannot trust that the user
330  * has set it up safely, so this makes sure interrupt etc flags are
331  * inherited from protected mode.
332  */
333 	VEFLAGS = vm86regs.pt.flags;
334 	vm86regs.pt.flags &= SAFE_MASK;
335 	vm86regs.pt.flags |= regs->flags & ~SAFE_MASK;
336 	vm86regs.pt.flags |= X86_VM_MASK;
337 
338 	vm86regs.pt.orig_ax = regs->orig_ax;
339 
340 	switch (vm86->cpu_type) {
341 	case CPU_286:
342 		vm86->veflags_mask = 0;
343 		break;
344 	case CPU_386:
345 		vm86->veflags_mask = X86_EFLAGS_NT | X86_EFLAGS_IOPL;
346 		break;
347 	case CPU_486:
348 		vm86->veflags_mask = X86_EFLAGS_AC | X86_EFLAGS_NT | X86_EFLAGS_IOPL;
349 		break;
350 	default:
351 		vm86->veflags_mask = X86_EFLAGS_ID | X86_EFLAGS_AC | X86_EFLAGS_NT | X86_EFLAGS_IOPL;
352 		break;
353 	}
354 
355 /*
356  * Save old state
357  */
358 	vm86->saved_sp0 = tsk->thread.sp0;
359 	lazy_save_gs(vm86->regs32.gs);
360 
361 	tss = &per_cpu(cpu_tss, get_cpu());
362 	/* make room for real-mode segments */
363 	tsk->thread.sp0 += 16;
364 
365 	if (static_cpu_has(X86_FEATURE_SEP))
366 		tsk->thread.sysenter_cs = 0;
367 
368 	load_sp0(tss, &tsk->thread);
369 	put_cpu();
370 
371 	if (vm86->flags & VM86_SCREEN_BITMAP)
372 		mark_screen_rdonly(tsk->mm);
373 
374 	memcpy((struct kernel_vm86_regs *)regs, &vm86regs, sizeof(vm86regs));
375 	force_iret();
376 	return regs->ax;
377 }
378 
379 static inline void set_IF(struct kernel_vm86_regs *regs)
380 {
381 	VEFLAGS |= X86_EFLAGS_VIF;
382 }
383 
384 static inline void clear_IF(struct kernel_vm86_regs *regs)
385 {
386 	VEFLAGS &= ~X86_EFLAGS_VIF;
387 }
388 
389 static inline void clear_TF(struct kernel_vm86_regs *regs)
390 {
391 	regs->pt.flags &= ~X86_EFLAGS_TF;
392 }
393 
394 static inline void clear_AC(struct kernel_vm86_regs *regs)
395 {
396 	regs->pt.flags &= ~X86_EFLAGS_AC;
397 }
398 
399 /*
400  * It is correct to call set_IF(regs) from the set_vflags_*
401  * functions. However someone forgot to call clear_IF(regs)
402  * in the opposite case.
403  * After the command sequence CLI PUSHF STI POPF you should
404  * end up with interrupts disabled, but you ended up with
405  * interrupts enabled.
406  *  ( I was testing my own changes, but the only bug I
407  *    could find was in a function I had not changed. )
408  * [KD]
409  */
410 
411 static inline void set_vflags_long(unsigned long flags, struct kernel_vm86_regs *regs)
412 {
413 	set_flags(VEFLAGS, flags, current->thread.vm86->veflags_mask);
414 	set_flags(regs->pt.flags, flags, SAFE_MASK);
415 	if (flags & X86_EFLAGS_IF)
416 		set_IF(regs);
417 	else
418 		clear_IF(regs);
419 }
420 
421 static inline void set_vflags_short(unsigned short flags, struct kernel_vm86_regs *regs)
422 {
423 	set_flags(VFLAGS, flags, current->thread.vm86->veflags_mask);
424 	set_flags(regs->pt.flags, flags, SAFE_MASK);
425 	if (flags & X86_EFLAGS_IF)
426 		set_IF(regs);
427 	else
428 		clear_IF(regs);
429 }
430 
431 static inline unsigned long get_vflags(struct kernel_vm86_regs *regs)
432 {
433 	unsigned long flags = regs->pt.flags & RETURN_MASK;
434 
435 	if (VEFLAGS & X86_EFLAGS_VIF)
436 		flags |= X86_EFLAGS_IF;
437 	flags |= X86_EFLAGS_IOPL;
438 	return flags | (VEFLAGS & current->thread.vm86->veflags_mask);
439 }
440 
441 static inline int is_revectored(int nr, struct revectored_struct *bitmap)
442 {
443 	return test_bit(nr, bitmap->__map);
444 }
445 
446 #define val_byte(val, n) (((__u8 *)&val)[n])
447 
448 #define pushb(base, ptr, val, err_label) \
449 	do { \
450 		__u8 __val = val; \
451 		ptr--; \
452 		if (put_user(__val, base + ptr) < 0) \
453 			goto err_label; \
454 	} while (0)
455 
456 #define pushw(base, ptr, val, err_label) \
457 	do { \
458 		__u16 __val = val; \
459 		ptr--; \
460 		if (put_user(val_byte(__val, 1), base + ptr) < 0) \
461 			goto err_label; \
462 		ptr--; \
463 		if (put_user(val_byte(__val, 0), base + ptr) < 0) \
464 			goto err_label; \
465 	} while (0)
466 
467 #define pushl(base, ptr, val, err_label) \
468 	do { \
469 		__u32 __val = val; \
470 		ptr--; \
471 		if (put_user(val_byte(__val, 3), base + ptr) < 0) \
472 			goto err_label; \
473 		ptr--; \
474 		if (put_user(val_byte(__val, 2), base + ptr) < 0) \
475 			goto err_label; \
476 		ptr--; \
477 		if (put_user(val_byte(__val, 1), base + ptr) < 0) \
478 			goto err_label; \
479 		ptr--; \
480 		if (put_user(val_byte(__val, 0), base + ptr) < 0) \
481 			goto err_label; \
482 	} while (0)
483 
484 #define popb(base, ptr, err_label) \
485 	({ \
486 		__u8 __res; \
487 		if (get_user(__res, base + ptr) < 0) \
488 			goto err_label; \
489 		ptr++; \
490 		__res; \
491 	})
492 
493 #define popw(base, ptr, err_label) \
494 	({ \
495 		__u16 __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 		__res; \
503 	})
504 
505 #define popl(base, ptr, err_label) \
506 	({ \
507 		__u32 __res; \
508 		if (get_user(val_byte(__res, 0), base + ptr) < 0) \
509 			goto err_label; \
510 		ptr++; \
511 		if (get_user(val_byte(__res, 1), base + ptr) < 0) \
512 			goto err_label; \
513 		ptr++; \
514 		if (get_user(val_byte(__res, 2), base + ptr) < 0) \
515 			goto err_label; \
516 		ptr++; \
517 		if (get_user(val_byte(__res, 3), base + ptr) < 0) \
518 			goto err_label; \
519 		ptr++; \
520 		__res; \
521 	})
522 
523 /* There are so many possible reasons for this function to return
524  * VM86_INTx, so adding another doesn't bother me. We can expect
525  * userspace programs to be able to handle it. (Getting a problem
526  * in userspace is always better than an Oops anyway.) [KD]
527  */
528 static void do_int(struct kernel_vm86_regs *regs, int i,
529     unsigned char __user *ssp, unsigned short sp)
530 {
531 	unsigned long __user *intr_ptr;
532 	unsigned long segoffs;
533 	struct vm86 *vm86 = current->thread.vm86;
534 
535 	if (regs->pt.cs == BIOSSEG)
536 		goto cannot_handle;
537 	if (is_revectored(i, &vm86->int_revectored))
538 		goto cannot_handle;
539 	if (i == 0x21 && is_revectored(AH(regs), &vm86->int21_revectored))
540 		goto cannot_handle;
541 	intr_ptr = (unsigned long __user *) (i << 2);
542 	if (get_user(segoffs, intr_ptr))
543 		goto cannot_handle;
544 	if ((segoffs >> 16) == BIOSSEG)
545 		goto cannot_handle;
546 	pushw(ssp, sp, get_vflags(regs), cannot_handle);
547 	pushw(ssp, sp, regs->pt.cs, cannot_handle);
548 	pushw(ssp, sp, IP(regs), cannot_handle);
549 	regs->pt.cs = segoffs >> 16;
550 	SP(regs) -= 6;
551 	IP(regs) = segoffs & 0xffff;
552 	clear_TF(regs);
553 	clear_IF(regs);
554 	clear_AC(regs);
555 	return;
556 
557 cannot_handle:
558 	save_v86_state(regs, VM86_INTx + (i << 8));
559 }
560 
561 int handle_vm86_trap(struct kernel_vm86_regs *regs, long error_code, int trapno)
562 {
563 	struct vm86 *vm86 = current->thread.vm86;
564 
565 	if (vm86->vm86plus.is_vm86pus) {
566 		if ((trapno == 3) || (trapno == 1)) {
567 			save_v86_state(regs, VM86_TRAP + (trapno << 8));
568 			return 0;
569 		}
570 		do_int(regs, trapno, (unsigned char __user *) (regs->pt.ss << 4), SP(regs));
571 		return 0;
572 	}
573 	if (trapno != 1)
574 		return 1; /* we let this handle by the calling routine */
575 	current->thread.trap_nr = trapno;
576 	current->thread.error_code = error_code;
577 	force_sig(SIGTRAP, current);
578 	return 0;
579 }
580 
581 void handle_vm86_fault(struct kernel_vm86_regs *regs, long error_code)
582 {
583 	unsigned char opcode;
584 	unsigned char __user *csp;
585 	unsigned char __user *ssp;
586 	unsigned short ip, sp, orig_flags;
587 	int data32, pref_done;
588 	struct vm86plus_info_struct *vmpi = &current->thread.vm86->vm86plus;
589 
590 #define CHECK_IF_IN_TRAP \
591 	if (vmpi->vm86dbg_active && vmpi->vm86dbg_TFpendig) \
592 		newflags |= X86_EFLAGS_TF
593 
594 	orig_flags = *(unsigned short *)&regs->pt.flags;
595 
596 	csp = (unsigned char __user *) (regs->pt.cs << 4);
597 	ssp = (unsigned char __user *) (regs->pt.ss << 4);
598 	sp = SP(regs);
599 	ip = IP(regs);
600 
601 	data32 = 0;
602 	pref_done = 0;
603 	do {
604 		switch (opcode = popb(csp, ip, simulate_sigsegv)) {
605 		case 0x66:      /* 32-bit data */     data32 = 1; break;
606 		case 0x67:      /* 32-bit address */  break;
607 		case 0x2e:      /* CS */              break;
608 		case 0x3e:      /* DS */              break;
609 		case 0x26:      /* ES */              break;
610 		case 0x36:      /* SS */              break;
611 		case 0x65:      /* GS */              break;
612 		case 0x64:      /* FS */              break;
613 		case 0xf2:      /* repnz */       break;
614 		case 0xf3:      /* rep */             break;
615 		default: pref_done = 1;
616 		}
617 	} while (!pref_done);
618 
619 	switch (opcode) {
620 
621 	/* pushf */
622 	case 0x9c:
623 		if (data32) {
624 			pushl(ssp, sp, get_vflags(regs), simulate_sigsegv);
625 			SP(regs) -= 4;
626 		} else {
627 			pushw(ssp, sp, get_vflags(regs), simulate_sigsegv);
628 			SP(regs) -= 2;
629 		}
630 		IP(regs) = ip;
631 		goto vm86_fault_return;
632 
633 	/* popf */
634 	case 0x9d:
635 		{
636 		unsigned long newflags;
637 		if (data32) {
638 			newflags = popl(ssp, sp, simulate_sigsegv);
639 			SP(regs) += 4;
640 		} else {
641 			newflags = popw(ssp, sp, simulate_sigsegv);
642 			SP(regs) += 2;
643 		}
644 		IP(regs) = ip;
645 		CHECK_IF_IN_TRAP;
646 		if (data32)
647 			set_vflags_long(newflags, regs);
648 		else
649 			set_vflags_short(newflags, regs);
650 
651 		goto check_vip;
652 		}
653 
654 	/* int xx */
655 	case 0xcd: {
656 		int intno = popb(csp, ip, simulate_sigsegv);
657 		IP(regs) = ip;
658 		if (vmpi->vm86dbg_active) {
659 			if ((1 << (intno & 7)) & vmpi->vm86dbg_intxxtab[intno >> 3]) {
660 				save_v86_state(regs, VM86_INTx + (intno << 8));
661 				return;
662 			}
663 		}
664 		do_int(regs, intno, ssp, sp);
665 		return;
666 	}
667 
668 	/* iret */
669 	case 0xcf:
670 		{
671 		unsigned long newip;
672 		unsigned long newcs;
673 		unsigned long newflags;
674 		if (data32) {
675 			newip = popl(ssp, sp, simulate_sigsegv);
676 			newcs = popl(ssp, sp, simulate_sigsegv);
677 			newflags = popl(ssp, sp, simulate_sigsegv);
678 			SP(regs) += 12;
679 		} else {
680 			newip = popw(ssp, sp, simulate_sigsegv);
681 			newcs = popw(ssp, sp, simulate_sigsegv);
682 			newflags = popw(ssp, sp, simulate_sigsegv);
683 			SP(regs) += 6;
684 		}
685 		IP(regs) = newip;
686 		regs->pt.cs = newcs;
687 		CHECK_IF_IN_TRAP;
688 		if (data32) {
689 			set_vflags_long(newflags, regs);
690 		} else {
691 			set_vflags_short(newflags, regs);
692 		}
693 		goto check_vip;
694 		}
695 
696 	/* cli */
697 	case 0xfa:
698 		IP(regs) = ip;
699 		clear_IF(regs);
700 		goto vm86_fault_return;
701 
702 	/* sti */
703 	/*
704 	 * Damn. This is incorrect: the 'sti' instruction should actually
705 	 * enable interrupts after the /next/ instruction. Not good.
706 	 *
707 	 * Probably needs some horsing around with the TF flag. Aiee..
708 	 */
709 	case 0xfb:
710 		IP(regs) = ip;
711 		set_IF(regs);
712 		goto check_vip;
713 
714 	default:
715 		save_v86_state(regs, VM86_UNKNOWN);
716 	}
717 
718 	return;
719 
720 check_vip:
721 	if (VEFLAGS & X86_EFLAGS_VIP) {
722 		save_v86_state(regs, VM86_STI);
723 		return;
724 	}
725 
726 vm86_fault_return:
727 	if (vmpi->force_return_for_pic  && (VEFLAGS & (X86_EFLAGS_IF | X86_EFLAGS_VIF))) {
728 		save_v86_state(regs, VM86_PICRETURN);
729 		return;
730 	}
731 	if (orig_flags & X86_EFLAGS_TF)
732 		handle_vm86_trap(regs, 0, X86_TRAP_DB);
733 	return;
734 
735 simulate_sigsegv:
736 	/* FIXME: After a long discussion with Stas we finally
737 	 *        agreed, that this is wrong. Here we should
738 	 *        really send a SIGSEGV to the user program.
739 	 *        But how do we create the correct context? We
740 	 *        are inside a general protection fault handler
741 	 *        and has just returned from a page fault handler.
742 	 *        The correct context for the signal handler
743 	 *        should be a mixture of the two, but how do we
744 	 *        get the information? [KD]
745 	 */
746 	save_v86_state(regs, VM86_UNKNOWN);
747 }
748 
749 /* ---------------- vm86 special IRQ passing stuff ----------------- */
750 
751 #define VM86_IRQNAME		"vm86irq"
752 
753 static struct vm86_irqs {
754 	struct task_struct *tsk;
755 	int sig;
756 } vm86_irqs[16];
757 
758 static DEFINE_SPINLOCK(irqbits_lock);
759 static int irqbits;
760 
761 #define ALLOWED_SIGS (1 /* 0 = don't send a signal */ \
762 	| (1 << SIGUSR1) | (1 << SIGUSR2) | (1 << SIGIO)  | (1 << SIGURG) \
763 	| (1 << SIGUNUSED))
764 
765 static irqreturn_t irq_handler(int intno, void *dev_id)
766 {
767 	int irq_bit;
768 	unsigned long flags;
769 
770 	spin_lock_irqsave(&irqbits_lock, flags);
771 	irq_bit = 1 << intno;
772 	if ((irqbits & irq_bit) || !vm86_irqs[intno].tsk)
773 		goto out;
774 	irqbits |= irq_bit;
775 	if (vm86_irqs[intno].sig)
776 		send_sig(vm86_irqs[intno].sig, vm86_irqs[intno].tsk, 1);
777 	/*
778 	 * IRQ will be re-enabled when user asks for the irq (whether
779 	 * polling or as a result of the signal)
780 	 */
781 	disable_irq_nosync(intno);
782 	spin_unlock_irqrestore(&irqbits_lock, flags);
783 	return IRQ_HANDLED;
784 
785 out:
786 	spin_unlock_irqrestore(&irqbits_lock, flags);
787 	return IRQ_NONE;
788 }
789 
790 static inline void free_vm86_irq(int irqnumber)
791 {
792 	unsigned long flags;
793 
794 	free_irq(irqnumber, NULL);
795 	vm86_irqs[irqnumber].tsk = NULL;
796 
797 	spin_lock_irqsave(&irqbits_lock, flags);
798 	irqbits &= ~(1 << irqnumber);
799 	spin_unlock_irqrestore(&irqbits_lock, flags);
800 }
801 
802 void release_vm86_irqs(struct task_struct *task)
803 {
804 	int i;
805 	for (i = FIRST_VM86_IRQ ; i <= LAST_VM86_IRQ; i++)
806 	    if (vm86_irqs[i].tsk == task)
807 		free_vm86_irq(i);
808 }
809 
810 static inline int get_and_reset_irq(int irqnumber)
811 {
812 	int bit;
813 	unsigned long flags;
814 	int ret = 0;
815 
816 	if (invalid_vm86_irq(irqnumber)) return 0;
817 	if (vm86_irqs[irqnumber].tsk != current) return 0;
818 	spin_lock_irqsave(&irqbits_lock, flags);
819 	bit = irqbits & (1 << irqnumber);
820 	irqbits &= ~bit;
821 	if (bit) {
822 		enable_irq(irqnumber);
823 		ret = 1;
824 	}
825 
826 	spin_unlock_irqrestore(&irqbits_lock, flags);
827 	return ret;
828 }
829 
830 
831 static int do_vm86_irq_handling(int subfunction, int irqnumber)
832 {
833 	int ret;
834 	switch (subfunction) {
835 		case VM86_GET_AND_RESET_IRQ: {
836 			return get_and_reset_irq(irqnumber);
837 		}
838 		case VM86_GET_IRQ_BITS: {
839 			return irqbits;
840 		}
841 		case VM86_REQUEST_IRQ: {
842 			int sig = irqnumber >> 8;
843 			int irq = irqnumber & 255;
844 			if (!capable(CAP_SYS_ADMIN)) return -EPERM;
845 			if (!((1 << sig) & ALLOWED_SIGS)) return -EPERM;
846 			if (invalid_vm86_irq(irq)) return -EPERM;
847 			if (vm86_irqs[irq].tsk) return -EPERM;
848 			ret = request_irq(irq, &irq_handler, 0, VM86_IRQNAME, NULL);
849 			if (ret) return ret;
850 			vm86_irqs[irq].sig = sig;
851 			vm86_irqs[irq].tsk = current;
852 			return irq;
853 		}
854 		case  VM86_FREE_IRQ: {
855 			if (invalid_vm86_irq(irqnumber)) return -EPERM;
856 			if (!vm86_irqs[irqnumber].tsk) return 0;
857 			if (vm86_irqs[irqnumber].tsk != current) return -EPERM;
858 			free_vm86_irq(irqnumber);
859 			return 0;
860 		}
861 	}
862 	return -EINVAL;
863 }
864 
865