2 * Copyright (C) 1994 Linus Torvalds
4 * 29 dec 2001 - Fixed oopses caused by unchecked access to the vm86
5 * stack - Manfred Spraul <manfred@colorfullife.com>
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>
12 * 22 mar 2002 - Added missing clear_IF in set_vflags_* Kasper Dupont
13 * <kasperd@daimi.au.dk>
15 * ?? ??? 2002 - Fixed premature returns from handle_vm86_fault
16 * caused by Kasper Dupont's changes - Stas Sergeev
18 * 4 apr 2002 - Fixed CHECK_IF_IN_TRAP broken by Stas' changes.
19 * Kasper Dupont <kasperd@daimi.au.dk>
21 * 9 apr 2002 - Changed syntax of macros in handle_vm86_fault.
22 * Kasper Dupont <kasperd@daimi.au.dk>
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>
31 #include <linux/capability.h>
32 #include <linux/errno.h>
33 #include <linux/interrupt.h>
34 #include <linux/sched.h>
35 #include <linux/kernel.h>
36 #include <linux/signal.h>
37 #include <linux/string.h>
39 #include <linux/smp.h>
40 #include <linux/highmem.h>
41 #include <linux/ptrace.h>
42 #include <linux/audit.h>
43 #include <linux/stddef.h>
45 #include <asm/uaccess.h>
47 #include <asm/tlbflush.h>
53 * Interrupt handling is not guaranteed:
54 * - a real x86 will disable all interrupts for one instruction
55 * after a "mov ss,xx" to make stack handling atomic even without
56 * the 'lss' instruction. We can't guarantee this in v86 mode,
57 * as the next instruction might result in a page fault or similar.
58 * - a real x86 will have interrupts disabled for one instruction
59 * past the 'sti' that enables them. We don't bother with all the
62 * Let's hope these problems do not actually matter for anything.
66 #define KVM86 ((struct kernel_vm86_struct *)regs)
67 #define VMPI KVM86->vm86plus
71 * 8- and 16-bit register defines..
73 #define AL(regs) (((unsigned char *)&((regs)->pt.eax))[0])
74 #define AH(regs) (((unsigned char *)&((regs)->pt.eax))[1])
75 #define IP(regs) (*(unsigned short *)&((regs)->pt.eip))
76 #define SP(regs) (*(unsigned short *)&((regs)->pt.esp))
79 * virtual flags (16 and 32-bit versions)
81 #define VFLAGS (*(unsigned short *)&(current->thread.v86flags))
82 #define VEFLAGS (current->thread.v86flags)
84 #define set_flags(X,new,mask) \
85 ((X) = ((X) & ~(mask)) | ((new) & (mask)))
87 #define SAFE_MASK (0xDD5)
88 #define RETURN_MASK (0xDFF)
90 /* convert kernel_vm86_regs to vm86_regs */
91 static int copy_vm86_regs_to_user(struct vm86_regs __user *user,
92 const struct kernel_vm86_regs *regs)
96 /* kernel_vm86_regs is missing xgs, so copy everything up to
97 (but not including) orig_eax, and then rest including orig_eax. */
98 ret += copy_to_user(user, regs, offsetof(struct kernel_vm86_regs, pt.orig_eax));
99 ret += copy_to_user(&user->orig_eax, ®s->pt.orig_eax,
100 sizeof(struct kernel_vm86_regs) -
101 offsetof(struct kernel_vm86_regs, pt.orig_eax));
106 /* convert vm86_regs to kernel_vm86_regs */
107 static int copy_vm86_regs_from_user(struct kernel_vm86_regs *regs,
108 const struct vm86_regs __user *user,
113 /* copy eax-xfs inclusive */
114 ret += copy_from_user(regs, user, offsetof(struct kernel_vm86_regs, pt.orig_eax));
115 /* copy orig_eax-__gsh+extra */
116 ret += copy_from_user(®s->pt.orig_eax, &user->orig_eax,
117 sizeof(struct kernel_vm86_regs) -
118 offsetof(struct kernel_vm86_regs, pt.orig_eax) +
123 struct pt_regs * FASTCALL(save_v86_state(struct kernel_vm86_regs * regs));
124 struct pt_regs * fastcall save_v86_state(struct kernel_vm86_regs * regs)
126 struct tss_struct *tss;
131 * This gets called from entry.S with interrupts disabled, but
132 * from process context. Enable interrupts here, before trying
133 * to access user space.
137 if (!current->thread.vm86_info) {
138 printk("no vm86_info: BAD\n");
141 set_flags(regs->pt.eflags, VEFLAGS, VIF_MASK | current->thread.v86mask);
142 tmp = copy_vm86_regs_to_user(¤t->thread.vm86_info->regs,regs);
143 tmp += put_user(current->thread.screen_bitmap,¤t->thread.vm86_info->screen_bitmap);
145 printk("vm86: could not access userspace vm86_info\n");
149 tss = &per_cpu(init_tss, get_cpu());
150 current->thread.esp0 = current->thread.saved_esp0;
151 current->thread.sysenter_cs = __KERNEL_CS;
152 load_esp0(tss, ¤t->thread);
153 current->thread.saved_esp0 = 0;
158 ret->xfs = current->thread.saved_fs;
159 loadsegment(gs, current->thread.saved_gs);
164 static void mark_screen_rdonly(struct mm_struct *mm)
173 pgd = pgd_offset(mm, 0xA0000);
174 if (pgd_none_or_clear_bad(pgd))
176 pud = pud_offset(pgd, 0xA0000);
177 if (pud_none_or_clear_bad(pud))
179 pmd = pmd_offset(pud, 0xA0000);
180 if (pmd_none_or_clear_bad(pmd))
182 pte = pte_offset_map_lock(mm, pmd, 0xA0000, &ptl);
183 for (i = 0; i < 32; i++) {
184 if (pte_present(*pte))
185 set_pte(pte, pte_wrprotect(*pte));
188 pte_unmap_unlock(pte, ptl);
195 static int do_vm86_irq_handling(int subfunction, int irqnumber);
196 static void do_sys_vm86(struct kernel_vm86_struct *info, struct task_struct *tsk);
198 asmlinkage int sys_vm86old(struct pt_regs regs)
200 struct vm86_struct __user *v86 = (struct vm86_struct __user *)regs.ebx;
201 struct kernel_vm86_struct info; /* declare this _on top_,
202 * this avoids wasting of stack space.
203 * This remains on the stack until we
204 * return to 32 bit user space.
206 struct task_struct *tsk;
207 int tmp, ret = -EPERM;
210 if (tsk->thread.saved_esp0)
212 tmp = copy_vm86_regs_from_user(&info.regs, &v86->regs,
213 offsetof(struct kernel_vm86_struct, vm86plus) -
218 memset(&info.vm86plus, 0, (int)&info.regs32 - (int)&info.vm86plus);
220 tsk->thread.vm86_info = v86;
221 do_sys_vm86(&info, tsk);
222 ret = 0; /* we never return here */
228 asmlinkage int sys_vm86(struct pt_regs regs)
230 struct kernel_vm86_struct info; /* declare this _on top_,
231 * this avoids wasting of stack space.
232 * This remains on the stack until we
233 * return to 32 bit user space.
235 struct task_struct *tsk;
237 struct vm86plus_struct __user *v86;
241 case VM86_REQUEST_IRQ:
243 case VM86_GET_IRQ_BITS:
244 case VM86_GET_AND_RESET_IRQ:
245 ret = do_vm86_irq_handling(regs.ebx, (int)regs.ecx);
247 case VM86_PLUS_INSTALL_CHECK:
248 /* NOTE: on old vm86 stuff this will return the error
249 from access_ok(), because the subfunction is
250 interpreted as (invalid) address to vm86_struct.
251 So the installation check works.
257 /* we come here only for functions VM86_ENTER, VM86_ENTER_NO_BYPASS */
259 if (tsk->thread.saved_esp0)
261 v86 = (struct vm86plus_struct __user *)regs.ecx;
262 tmp = copy_vm86_regs_from_user(&info.regs, &v86->regs,
263 offsetof(struct kernel_vm86_struct, regs32) -
269 info.vm86plus.is_vm86pus = 1;
270 tsk->thread.vm86_info = (struct vm86_struct __user *)v86;
271 do_sys_vm86(&info, tsk);
272 ret = 0; /* we never return here */
278 static void do_sys_vm86(struct kernel_vm86_struct *info, struct task_struct *tsk)
280 struct tss_struct *tss;
282 * make sure the vm86() system call doesn't try to do anything silly
284 info->regs.pt.xds = 0;
285 info->regs.pt.xes = 0;
286 info->regs.pt.xfs = 0;
288 /* we are clearing gs later just before "jmp resume_userspace",
289 * because it is not saved/restored.
293 * The eflags register is also special: we cannot trust that the user
294 * has set it up safely, so this makes sure interrupt etc flags are
295 * inherited from protected mode.
297 VEFLAGS = info->regs.pt.eflags;
298 info->regs.pt.eflags &= SAFE_MASK;
299 info->regs.pt.eflags |= info->regs32->eflags & ~SAFE_MASK;
300 info->regs.pt.eflags |= VM_MASK;
302 switch (info->cpu_type) {
304 tsk->thread.v86mask = 0;
307 tsk->thread.v86mask = NT_MASK | IOPL_MASK;
310 tsk->thread.v86mask = AC_MASK | NT_MASK | IOPL_MASK;
313 tsk->thread.v86mask = ID_MASK | AC_MASK | NT_MASK | IOPL_MASK;
318 * Save old state, set default return value (%eax) to 0
320 info->regs32->eax = 0;
321 tsk->thread.saved_esp0 = tsk->thread.esp0;
322 tsk->thread.saved_fs = info->regs32->xfs;
323 savesegment(gs, tsk->thread.saved_gs);
325 tss = &per_cpu(init_tss, get_cpu());
326 tsk->thread.esp0 = (unsigned long) &info->VM86_TSS_ESP0;
328 tsk->thread.sysenter_cs = 0;
329 load_esp0(tss, &tsk->thread);
332 tsk->thread.screen_bitmap = info->screen_bitmap;
333 if (info->flags & VM86_SCREEN_BITMAP)
334 mark_screen_rdonly(tsk->mm);
336 /*call audit_syscall_exit since we do not exit via the normal paths */
337 if (unlikely(current->audit_context))
338 audit_syscall_exit(AUDITSC_RESULT(0), 0);
340 __asm__ __volatile__(
344 "jmp resume_userspace"
346 :"r" (&info->regs), "r" (task_thread_info(tsk)), "r" (0));
347 /* we never return here */
350 static inline void return_to_32bit(struct kernel_vm86_regs * regs16, int retval)
352 struct pt_regs * regs32;
354 regs32 = save_v86_state(regs16);
355 regs32->eax = retval;
356 __asm__ __volatile__("movl %0,%%esp\n\t"
358 "jmp resume_userspace"
359 : : "r" (regs32), "r" (current_thread_info()));
362 static inline void set_IF(struct kernel_vm86_regs * regs)
365 if (VEFLAGS & VIP_MASK)
366 return_to_32bit(regs, VM86_STI);
369 static inline void clear_IF(struct kernel_vm86_regs * regs)
371 VEFLAGS &= ~VIF_MASK;
374 static inline void clear_TF(struct kernel_vm86_regs * regs)
376 regs->pt.eflags &= ~TF_MASK;
379 static inline void clear_AC(struct kernel_vm86_regs * regs)
381 regs->pt.eflags &= ~AC_MASK;
384 /* It is correct to call set_IF(regs) from the set_vflags_*
385 * functions. However someone forgot to call clear_IF(regs)
386 * in the opposite case.
387 * After the command sequence CLI PUSHF STI POPF you should
388 * end up with interrups disabled, but you ended up with
389 * interrupts enabled.
390 * ( I was testing my own changes, but the only bug I
391 * could find was in a function I had not changed. )
395 static inline void set_vflags_long(unsigned long eflags, struct kernel_vm86_regs * regs)
397 set_flags(VEFLAGS, eflags, current->thread.v86mask);
398 set_flags(regs->pt.eflags, eflags, SAFE_MASK);
399 if (eflags & IF_MASK)
405 static inline void set_vflags_short(unsigned short flags, struct kernel_vm86_regs * regs)
407 set_flags(VFLAGS, flags, current->thread.v86mask);
408 set_flags(regs->pt.eflags, flags, SAFE_MASK);
415 static inline unsigned long get_vflags(struct kernel_vm86_regs * regs)
417 unsigned long flags = regs->pt.eflags & RETURN_MASK;
419 if (VEFLAGS & VIF_MASK)
422 return flags | (VEFLAGS & current->thread.v86mask);
425 static inline int is_revectored(int nr, struct revectored_struct * bitmap)
427 __asm__ __volatile__("btl %2,%1\n\tsbbl %0,%0"
429 :"m" (*bitmap),"r" (nr));
433 #define val_byte(val, n) (((__u8 *)&val)[n])
435 #define pushb(base, ptr, val, err_label) \
439 if (put_user(__val, base + ptr) < 0) \
443 #define pushw(base, ptr, val, err_label) \
447 if (put_user(val_byte(__val, 1), base + ptr) < 0) \
450 if (put_user(val_byte(__val, 0), base + ptr) < 0) \
454 #define pushl(base, ptr, val, err_label) \
458 if (put_user(val_byte(__val, 3), base + ptr) < 0) \
461 if (put_user(val_byte(__val, 2), base + ptr) < 0) \
464 if (put_user(val_byte(__val, 1), base + ptr) < 0) \
467 if (put_user(val_byte(__val, 0), base + ptr) < 0) \
471 #define popb(base, ptr, err_label) \
474 if (get_user(__res, base + ptr) < 0) \
480 #define popw(base, ptr, err_label) \
483 if (get_user(val_byte(__res, 0), base + ptr) < 0) \
486 if (get_user(val_byte(__res, 1), base + ptr) < 0) \
492 #define popl(base, ptr, err_label) \
495 if (get_user(val_byte(__res, 0), base + ptr) < 0) \
498 if (get_user(val_byte(__res, 1), base + ptr) < 0) \
501 if (get_user(val_byte(__res, 2), base + ptr) < 0) \
504 if (get_user(val_byte(__res, 3), base + ptr) < 0) \
510 /* There are so many possible reasons for this function to return
511 * VM86_INTx, so adding another doesn't bother me. We can expect
512 * userspace programs to be able to handle it. (Getting a problem
513 * in userspace is always better than an Oops anyway.) [KD]
515 static void do_int(struct kernel_vm86_regs *regs, int i,
516 unsigned char __user * ssp, unsigned short sp)
518 unsigned long __user *intr_ptr;
519 unsigned long segoffs;
521 if (regs->pt.xcs == BIOSSEG)
523 if (is_revectored(i, &KVM86->int_revectored))
525 if (i==0x21 && is_revectored(AH(regs),&KVM86->int21_revectored))
527 intr_ptr = (unsigned long __user *) (i << 2);
528 if (get_user(segoffs, intr_ptr))
530 if ((segoffs >> 16) == BIOSSEG)
532 pushw(ssp, sp, get_vflags(regs), cannot_handle);
533 pushw(ssp, sp, regs->pt.xcs, cannot_handle);
534 pushw(ssp, sp, IP(regs), cannot_handle);
535 regs->pt.xcs = segoffs >> 16;
537 IP(regs) = segoffs & 0xffff;
544 return_to_32bit(regs, VM86_INTx + (i << 8));
547 int handle_vm86_trap(struct kernel_vm86_regs * regs, long error_code, int trapno)
549 if (VMPI.is_vm86pus) {
550 if ( (trapno==3) || (trapno==1) )
551 return_to_32bit(regs, VM86_TRAP + (trapno << 8));
552 do_int(regs, trapno, (unsigned char __user *) (regs->pt.xss << 4), SP(regs));
556 return 1; /* we let this handle by the calling routine */
557 if (current->ptrace & PT_PTRACED) {
559 spin_lock_irqsave(¤t->sighand->siglock, flags);
560 sigdelset(¤t->blocked, SIGTRAP);
562 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
564 send_sig(SIGTRAP, current, 1);
565 current->thread.trap_no = trapno;
566 current->thread.error_code = error_code;
570 void handle_vm86_fault(struct kernel_vm86_regs * regs, long error_code)
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;
578 #define CHECK_IF_IN_TRAP \
579 if (VMPI.vm86dbg_active && VMPI.vm86dbg_TFpendig) \
581 #define VM86_FAULT_RETURN do { \
582 if (VMPI.force_return_for_pic && (VEFLAGS & (IF_MASK | VIF_MASK))) \
583 return_to_32bit(regs, VM86_PICRETURN); \
584 if (orig_flags & TF_MASK) \
585 handle_vm86_trap(regs, 0, 1); \
588 orig_flags = *(unsigned short *)®s->pt.eflags;
590 csp = (unsigned char __user *) (regs->pt.xcs << 4);
591 ssp = (unsigned char __user *) (regs->pt.xss << 4);
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;
611 } while (!pref_done);
618 pushl(ssp, sp, get_vflags(regs), simulate_sigsegv);
621 pushw(ssp, sp, get_vflags(regs), simulate_sigsegv);
630 unsigned long newflags;
632 newflags=popl(ssp, sp, simulate_sigsegv);
635 newflags = popw(ssp, sp, simulate_sigsegv);
641 set_vflags_long(newflags, regs);
643 set_vflags_short(newflags, regs);
650 int intno=popb(csp, ip, simulate_sigsegv);
652 if (VMPI.vm86dbg_active) {
653 if ( (1 << (intno &7)) & VMPI.vm86dbg_intxxtab[intno >> 3] )
654 return_to_32bit(regs, VM86_INTx + (intno << 8));
656 do_int(regs, intno, ssp, sp);
665 unsigned long newflags;
667 newip=popl(ssp, sp, simulate_sigsegv);
668 newcs=popl(ssp, sp, simulate_sigsegv);
669 newflags=popl(ssp, sp, simulate_sigsegv);
672 newip = popw(ssp, sp, simulate_sigsegv);
673 newcs = popw(ssp, sp, simulate_sigsegv);
674 newflags = popw(ssp, sp, simulate_sigsegv);
678 regs->pt.xcs = newcs;
681 set_vflags_long(newflags, regs);
683 set_vflags_short(newflags, regs);
696 * Damn. This is incorrect: the 'sti' instruction should actually
697 * enable interrupts after the /next/ instruction. Not good.
699 * Probably needs some horsing around with the TF flag. Aiee..
707 return_to_32bit(regs, VM86_UNKNOWN);
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]
723 return_to_32bit(regs, VM86_UNKNOWN);
726 /* ---------------- vm86 special IRQ passing stuff ----------------- */
728 #define VM86_IRQNAME "vm86irq"
730 static struct vm86_irqs {
731 struct task_struct *tsk;
735 static DEFINE_SPINLOCK(irqbits_lock);
738 #define ALLOWED_SIGS ( 1 /* 0 = don't send a signal */ \
739 | (1 << SIGUSR1) | (1 << SIGUSR2) | (1 << SIGIO) | (1 << SIGURG) \
742 static irqreturn_t irq_handler(int intno, void *dev_id)
747 spin_lock_irqsave(&irqbits_lock, flags);
748 irq_bit = 1 << intno;
749 if ((irqbits & irq_bit) || ! vm86_irqs[intno].tsk)
752 if (vm86_irqs[intno].sig)
753 send_sig(vm86_irqs[intno].sig, vm86_irqs[intno].tsk, 1);
755 * IRQ will be re-enabled when user asks for the irq (whether
756 * polling or as a result of the signal)
758 disable_irq_nosync(intno);
759 spin_unlock_irqrestore(&irqbits_lock, flags);
763 spin_unlock_irqrestore(&irqbits_lock, flags);
767 static inline void free_vm86_irq(int irqnumber)
771 free_irq(irqnumber, NULL);
772 vm86_irqs[irqnumber].tsk = NULL;
774 spin_lock_irqsave(&irqbits_lock, flags);
775 irqbits &= ~(1 << irqnumber);
776 spin_unlock_irqrestore(&irqbits_lock, flags);
779 void release_vm86_irqs(struct task_struct *task)
782 for (i = FIRST_VM86_IRQ ; i <= LAST_VM86_IRQ; i++)
783 if (vm86_irqs[i].tsk == task)
787 static inline int get_and_reset_irq(int irqnumber)
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);
799 enable_irq(irqnumber);
803 spin_unlock_irqrestore(&irqbits_lock, flags);
808 static int do_vm86_irq_handling(int subfunction, int irqnumber)
811 switch (subfunction) {
812 case VM86_GET_AND_RESET_IRQ: {
813 return get_and_reset_irq(irqnumber);
815 case VM86_GET_IRQ_BITS: {
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);
827 vm86_irqs[irq].sig = sig;
828 vm86_irqs[irq].tsk = current;
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);