4 * Copyright (C) 1994 Linus Torvalds
6 * 29 dec 2001 - Fixed oopses caused by unchecked access to the vm86
7 * stack - Manfred Spraul <manfreds@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/config.h>
34 #include <linux/errno.h>
35 #include <linux/interrupt.h>
36 #include <linux/sched.h>
37 #include <linux/kernel.h>
38 #include <linux/signal.h>
39 #include <linux/string.h>
41 #include <linux/smp.h>
42 #include <linux/smp_lock.h>
43 #include <linux/highmem.h>
44 #include <linux/ptrace.h>
46 #include <asm/uaccess.h>
48 #include <asm/tlbflush.h>
54 * Interrupt handling is not guaranteed:
55 * - a real x86 will disable all interrupts for one instruction
56 * after a "mov ss,xx" to make stack handling atomic even without
57 * the 'lss' instruction. We can't guarantee this in v86 mode,
58 * as the next instruction might result in a page fault or similar.
59 * - a real x86 will have interrupts disabled for one instruction
60 * past the 'sti' that enables them. We don't bother with all the
63 * Let's hope these problems do not actually matter for anything.
67 #define KVM86 ((struct kernel_vm86_struct *)regs)
68 #define VMPI KVM86->vm86plus
72 * 8- and 16-bit register defines..
74 #define AL(regs) (((unsigned char *)&((regs)->eax))[0])
75 #define AH(regs) (((unsigned char *)&((regs)->eax))[1])
76 #define IP(regs) (*(unsigned short *)&((regs)->eip))
77 #define SP(regs) (*(unsigned short *)&((regs)->esp))
80 * virtual flags (16 and 32-bit versions)
82 #define VFLAGS (*(unsigned short *)&(current->thread.v86flags))
83 #define VEFLAGS (current->thread.v86flags)
85 #define set_flags(X,new,mask) \
86 ((X) = ((X) & ~(mask)) | ((new) & (mask)))
88 #define SAFE_MASK (0xDD5)
89 #define RETURN_MASK (0xDFF)
91 #define VM86_REGS_PART2 orig_eax
92 #define VM86_REGS_SIZE1 \
93 ( (unsigned)( & (((struct kernel_vm86_regs *)0)->VM86_REGS_PART2) ) )
94 #define VM86_REGS_SIZE2 (sizeof(struct kernel_vm86_regs) - VM86_REGS_SIZE1)
96 struct pt_regs * FASTCALL(save_v86_state(struct kernel_vm86_regs * regs));
97 struct pt_regs * fastcall save_v86_state(struct kernel_vm86_regs * regs)
99 struct tss_struct *tss;
104 * This gets called from entry.S with interrupts disabled, but
105 * from process context. Enable interrupts here, before trying
106 * to access user space.
110 if (!current->thread.vm86_info) {
111 printk("no vm86_info: BAD\n");
114 set_flags(regs->eflags, VEFLAGS, VIF_MASK | current->thread.v86mask);
115 tmp = copy_to_user(¤t->thread.vm86_info->regs,regs, VM86_REGS_SIZE1);
116 tmp += copy_to_user(¤t->thread.vm86_info->regs.VM86_REGS_PART2,
117 ®s->VM86_REGS_PART2, VM86_REGS_SIZE2);
118 tmp += put_user(current->thread.screen_bitmap,¤t->thread.vm86_info->screen_bitmap);
120 printk("vm86: could not access userspace vm86_info\n");
124 tss = &per_cpu(init_tss, get_cpu());
125 current->thread.esp0 = current->thread.saved_esp0;
126 current->thread.sysenter_cs = __KERNEL_CS;
127 load_esp0(tss, ¤t->thread);
128 current->thread.saved_esp0 = 0;
131 loadsegment(fs, current->thread.saved_fs);
132 loadsegment(gs, current->thread.saved_gs);
137 static void mark_screen_rdonly(struct mm_struct *mm)
146 pgd = pgd_offset(mm, 0xA0000);
147 if (pgd_none_or_clear_bad(pgd))
149 pud = pud_offset(pgd, 0xA0000);
150 if (pud_none_or_clear_bad(pud))
152 pmd = pmd_offset(pud, 0xA0000);
153 if (pmd_none_or_clear_bad(pmd))
155 pte = pte_offset_map_lock(mm, pmd, 0xA0000, &ptl);
156 for (i = 0; i < 32; i++) {
157 if (pte_present(*pte))
158 set_pte(pte, pte_wrprotect(*pte));
161 pte_unmap_unlock(pte, ptl);
168 static int do_vm86_irq_handling(int subfunction, int irqnumber);
169 static void do_sys_vm86(struct kernel_vm86_struct *info, struct task_struct *tsk);
171 asmlinkage int sys_vm86old(struct pt_regs regs)
173 struct vm86_struct __user *v86 = (struct vm86_struct __user *)regs.ebx;
174 struct kernel_vm86_struct info; /* declare this _on top_,
175 * this avoids wasting of stack space.
176 * This remains on the stack until we
177 * return to 32 bit user space.
179 struct task_struct *tsk;
180 int tmp, ret = -EPERM;
183 if (tsk->thread.saved_esp0)
185 tmp = copy_from_user(&info, v86, VM86_REGS_SIZE1);
186 tmp += copy_from_user(&info.regs.VM86_REGS_PART2, &v86->regs.VM86_REGS_PART2,
187 (long)&info.vm86plus - (long)&info.regs.VM86_REGS_PART2);
191 memset(&info.vm86plus, 0, (int)&info.regs32 - (int)&info.vm86plus);
193 tsk->thread.vm86_info = v86;
194 do_sys_vm86(&info, tsk);
195 ret = 0; /* we never return here */
201 asmlinkage int sys_vm86(struct pt_regs regs)
203 struct kernel_vm86_struct info; /* declare this _on top_,
204 * this avoids wasting of stack space.
205 * This remains on the stack until we
206 * return to 32 bit user space.
208 struct task_struct *tsk;
210 struct vm86plus_struct __user *v86;
214 case VM86_REQUEST_IRQ:
216 case VM86_GET_IRQ_BITS:
217 case VM86_GET_AND_RESET_IRQ:
218 ret = do_vm86_irq_handling(regs.ebx, (int)regs.ecx);
220 case VM86_PLUS_INSTALL_CHECK:
221 /* NOTE: on old vm86 stuff this will return the error
222 from access_ok(), because the subfunction is
223 interpreted as (invalid) address to vm86_struct.
224 So the installation check works.
230 /* we come here only for functions VM86_ENTER, VM86_ENTER_NO_BYPASS */
232 if (tsk->thread.saved_esp0)
234 v86 = (struct vm86plus_struct __user *)regs.ecx;
235 tmp = copy_from_user(&info, v86, VM86_REGS_SIZE1);
236 tmp += copy_from_user(&info.regs.VM86_REGS_PART2, &v86->regs.VM86_REGS_PART2,
237 (long)&info.regs32 - (long)&info.regs.VM86_REGS_PART2);
242 info.vm86plus.is_vm86pus = 1;
243 tsk->thread.vm86_info = (struct vm86_struct __user *)v86;
244 do_sys_vm86(&info, tsk);
245 ret = 0; /* we never return here */
251 static void do_sys_vm86(struct kernel_vm86_struct *info, struct task_struct *tsk)
253 struct tss_struct *tss;
255 * make sure the vm86() system call doesn't try to do anything silly
257 info->regs.__null_ds = 0;
258 info->regs.__null_es = 0;
260 /* we are clearing fs,gs later just before "jmp resume_userspace",
261 * because starting with Linux 2.1.x they aren't no longer saved/restored
265 * The eflags register is also special: we cannot trust that the user
266 * has set it up safely, so this makes sure interrupt etc flags are
267 * inherited from protected mode.
269 VEFLAGS = info->regs.eflags;
270 info->regs.eflags &= SAFE_MASK;
271 info->regs.eflags |= info->regs32->eflags & ~SAFE_MASK;
272 info->regs.eflags |= VM_MASK;
274 switch (info->cpu_type) {
276 tsk->thread.v86mask = 0;
279 tsk->thread.v86mask = NT_MASK | IOPL_MASK;
282 tsk->thread.v86mask = AC_MASK | NT_MASK | IOPL_MASK;
285 tsk->thread.v86mask = ID_MASK | AC_MASK | NT_MASK | IOPL_MASK;
290 * Save old state, set default return value (%eax) to 0
292 info->regs32->eax = 0;
293 tsk->thread.saved_esp0 = tsk->thread.esp0;
294 savesegment(fs, tsk->thread.saved_fs);
295 savesegment(gs, tsk->thread.saved_gs);
297 tss = &per_cpu(init_tss, get_cpu());
298 tsk->thread.esp0 = (unsigned long) &info->VM86_TSS_ESP0;
300 tsk->thread.sysenter_cs = 0;
301 load_esp0(tss, &tsk->thread);
304 tsk->thread.screen_bitmap = info->screen_bitmap;
305 if (info->flags & VM86_SCREEN_BITMAP)
306 mark_screen_rdonly(tsk->mm);
307 __asm__ __volatile__(
308 "xorl %%eax,%%eax; movl %%eax,%%fs; movl %%eax,%%gs\n\t"
311 "jmp resume_userspace"
313 :"r" (&info->regs), "r" (tsk->thread_info) : "ax");
314 /* we never return here */
317 static inline void return_to_32bit(struct kernel_vm86_regs * regs16, int retval)
319 struct pt_regs * regs32;
321 regs32 = save_v86_state(regs16);
322 regs32->eax = retval;
323 __asm__ __volatile__("movl %0,%%esp\n\t"
325 "jmp resume_userspace"
326 : : "r" (regs32), "r" (current_thread_info()));
329 static inline void set_IF(struct kernel_vm86_regs * regs)
332 if (VEFLAGS & VIP_MASK)
333 return_to_32bit(regs, VM86_STI);
336 static inline void clear_IF(struct kernel_vm86_regs * regs)
338 VEFLAGS &= ~VIF_MASK;
341 static inline void clear_TF(struct kernel_vm86_regs * regs)
343 regs->eflags &= ~TF_MASK;
346 static inline void clear_AC(struct kernel_vm86_regs * regs)
348 regs->eflags &= ~AC_MASK;
351 /* It is correct to call set_IF(regs) from the set_vflags_*
352 * functions. However someone forgot to call clear_IF(regs)
353 * in the opposite case.
354 * After the command sequence CLI PUSHF STI POPF you should
355 * end up with interrups disabled, but you ended up with
356 * interrupts enabled.
357 * ( I was testing my own changes, but the only bug I
358 * could find was in a function I had not changed. )
362 static inline void set_vflags_long(unsigned long eflags, struct kernel_vm86_regs * regs)
364 set_flags(VEFLAGS, eflags, current->thread.v86mask);
365 set_flags(regs->eflags, eflags, SAFE_MASK);
366 if (eflags & IF_MASK)
372 static inline void set_vflags_short(unsigned short flags, struct kernel_vm86_regs * regs)
374 set_flags(VFLAGS, flags, current->thread.v86mask);
375 set_flags(regs->eflags, flags, SAFE_MASK);
382 static inline unsigned long get_vflags(struct kernel_vm86_regs * regs)
384 unsigned long flags = regs->eflags & RETURN_MASK;
386 if (VEFLAGS & VIF_MASK)
389 return flags | (VEFLAGS & current->thread.v86mask);
392 static inline int is_revectored(int nr, struct revectored_struct * bitmap)
394 __asm__ __volatile__("btl %2,%1\n\tsbbl %0,%0"
396 :"m" (*bitmap),"r" (nr));
400 #define val_byte(val, n) (((__u8 *)&val)[n])
402 #define pushb(base, ptr, val, err_label) \
406 if (put_user(__val, base + ptr) < 0) \
410 #define pushw(base, ptr, val, err_label) \
414 if (put_user(val_byte(__val, 1), base + ptr) < 0) \
417 if (put_user(val_byte(__val, 0), base + ptr) < 0) \
421 #define pushl(base, ptr, val, err_label) \
425 if (put_user(val_byte(__val, 3), base + ptr) < 0) \
428 if (put_user(val_byte(__val, 2), base + ptr) < 0) \
431 if (put_user(val_byte(__val, 1), base + ptr) < 0) \
434 if (put_user(val_byte(__val, 0), base + ptr) < 0) \
438 #define popb(base, ptr, err_label) \
441 if (get_user(__res, base + ptr) < 0) \
447 #define popw(base, ptr, err_label) \
450 if (get_user(val_byte(__res, 0), base + ptr) < 0) \
453 if (get_user(val_byte(__res, 1), base + ptr) < 0) \
459 #define popl(base, ptr, err_label) \
462 if (get_user(val_byte(__res, 0), base + ptr) < 0) \
465 if (get_user(val_byte(__res, 1), base + ptr) < 0) \
468 if (get_user(val_byte(__res, 2), base + ptr) < 0) \
471 if (get_user(val_byte(__res, 3), base + ptr) < 0) \
477 /* There are so many possible reasons for this function to return
478 * VM86_INTx, so adding another doesn't bother me. We can expect
479 * userspace programs to be able to handle it. (Getting a problem
480 * in userspace is always better than an Oops anyway.) [KD]
482 static void do_int(struct kernel_vm86_regs *regs, int i,
483 unsigned char __user * ssp, unsigned short sp)
485 unsigned long __user *intr_ptr;
486 unsigned long segoffs;
488 if (regs->cs == BIOSSEG)
490 if (is_revectored(i, &KVM86->int_revectored))
492 if (i==0x21 && is_revectored(AH(regs),&KVM86->int21_revectored))
494 intr_ptr = (unsigned long __user *) (i << 2);
495 if (get_user(segoffs, intr_ptr))
497 if ((segoffs >> 16) == BIOSSEG)
499 pushw(ssp, sp, get_vflags(regs), cannot_handle);
500 pushw(ssp, sp, regs->cs, cannot_handle);
501 pushw(ssp, sp, IP(regs), cannot_handle);
502 regs->cs = segoffs >> 16;
504 IP(regs) = segoffs & 0xffff;
511 return_to_32bit(regs, VM86_INTx + (i << 8));
514 int handle_vm86_trap(struct kernel_vm86_regs * regs, long error_code, int trapno)
516 if (VMPI.is_vm86pus) {
517 if ( (trapno==3) || (trapno==1) )
518 return_to_32bit(regs, VM86_TRAP + (trapno << 8));
519 do_int(regs, trapno, (unsigned char __user *) (regs->ss << 4), SP(regs));
523 return 1; /* we let this handle by the calling routine */
524 if (current->ptrace & PT_PTRACED) {
526 spin_lock_irqsave(¤t->sighand->siglock, flags);
527 sigdelset(¤t->blocked, SIGTRAP);
529 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
531 send_sig(SIGTRAP, current, 1);
532 current->thread.trap_no = trapno;
533 current->thread.error_code = error_code;
537 void handle_vm86_fault(struct kernel_vm86_regs * regs, long error_code)
539 unsigned char opcode;
540 unsigned char __user *csp;
541 unsigned char __user *ssp;
542 unsigned short ip, sp, orig_flags;
543 int data32, pref_done;
545 #define CHECK_IF_IN_TRAP \
546 if (VMPI.vm86dbg_active && VMPI.vm86dbg_TFpendig) \
548 #define VM86_FAULT_RETURN do { \
549 if (VMPI.force_return_for_pic && (VEFLAGS & (IF_MASK | VIF_MASK))) \
550 return_to_32bit(regs, VM86_PICRETURN); \
551 if (orig_flags & TF_MASK) \
552 handle_vm86_trap(regs, 0, 1); \
555 orig_flags = *(unsigned short *)®s->eflags;
557 csp = (unsigned char __user *) (regs->cs << 4);
558 ssp = (unsigned char __user *) (regs->ss << 4);
565 switch (opcode = popb(csp, ip, simulate_sigsegv)) {
566 case 0x66: /* 32-bit data */ data32=1; break;
567 case 0x67: /* 32-bit address */ break;
568 case 0x2e: /* CS */ break;
569 case 0x3e: /* DS */ break;
570 case 0x26: /* ES */ break;
571 case 0x36: /* SS */ break;
572 case 0x65: /* GS */ break;
573 case 0x64: /* FS */ break;
574 case 0xf2: /* repnz */ break;
575 case 0xf3: /* rep */ break;
576 default: pref_done = 1;
578 } while (!pref_done);
585 pushl(ssp, sp, get_vflags(regs), simulate_sigsegv);
588 pushw(ssp, sp, get_vflags(regs), simulate_sigsegv);
597 unsigned long newflags;
599 newflags=popl(ssp, sp, simulate_sigsegv);
602 newflags = popw(ssp, sp, simulate_sigsegv);
608 set_vflags_long(newflags, regs);
610 set_vflags_short(newflags, regs);
617 int intno=popb(csp, ip, simulate_sigsegv);
619 if (VMPI.vm86dbg_active) {
620 if ( (1 << (intno &7)) & VMPI.vm86dbg_intxxtab[intno >> 3] )
621 return_to_32bit(regs, VM86_INTx + (intno << 8));
623 do_int(regs, intno, ssp, sp);
632 unsigned long newflags;
634 newip=popl(ssp, sp, simulate_sigsegv);
635 newcs=popl(ssp, sp, simulate_sigsegv);
636 newflags=popl(ssp, sp, simulate_sigsegv);
639 newip = popw(ssp, sp, simulate_sigsegv);
640 newcs = popw(ssp, sp, simulate_sigsegv);
641 newflags = popw(ssp, sp, simulate_sigsegv);
648 set_vflags_long(newflags, regs);
650 set_vflags_short(newflags, regs);
663 * Damn. This is incorrect: the 'sti' instruction should actually
664 * enable interrupts after the /next/ instruction. Not good.
666 * Probably needs some horsing around with the TF flag. Aiee..
674 return_to_32bit(regs, VM86_UNKNOWN);
680 /* FIXME: After a long discussion with Stas we finally
681 * agreed, that this is wrong. Here we should
682 * really send a SIGSEGV to the user program.
683 * But how do we create the correct context? We
684 * are inside a general protection fault handler
685 * and has just returned from a page fault handler.
686 * The correct context for the signal handler
687 * should be a mixture of the two, but how do we
688 * get the information? [KD]
690 return_to_32bit(regs, VM86_UNKNOWN);
693 /* ---------------- vm86 special IRQ passing stuff ----------------- */
695 #define VM86_IRQNAME "vm86irq"
697 static struct vm86_irqs {
698 struct task_struct *tsk;
702 static DEFINE_SPINLOCK(irqbits_lock);
705 #define ALLOWED_SIGS ( 1 /* 0 = don't send a signal */ \
706 | (1 << SIGUSR1) | (1 << SIGUSR2) | (1 << SIGIO) | (1 << SIGURG) \
709 static irqreturn_t irq_handler(int intno, void *dev_id, struct pt_regs * regs)
714 spin_lock_irqsave(&irqbits_lock, flags);
715 irq_bit = 1 << intno;
716 if ((irqbits & irq_bit) || ! vm86_irqs[intno].tsk)
719 if (vm86_irqs[intno].sig)
720 send_sig(vm86_irqs[intno].sig, vm86_irqs[intno].tsk, 1);
722 * IRQ will be re-enabled when user asks for the irq (whether
723 * polling or as a result of the signal)
725 disable_irq_nosync(intno);
726 spin_unlock_irqrestore(&irqbits_lock, flags);
730 spin_unlock_irqrestore(&irqbits_lock, flags);
734 static inline void free_vm86_irq(int irqnumber)
738 free_irq(irqnumber, NULL);
739 vm86_irqs[irqnumber].tsk = NULL;
741 spin_lock_irqsave(&irqbits_lock, flags);
742 irqbits &= ~(1 << irqnumber);
743 spin_unlock_irqrestore(&irqbits_lock, flags);
746 void release_vm86_irqs(struct task_struct *task)
749 for (i = FIRST_VM86_IRQ ; i <= LAST_VM86_IRQ; i++)
750 if (vm86_irqs[i].tsk == task)
754 static inline int get_and_reset_irq(int irqnumber)
760 if (invalid_vm86_irq(irqnumber)) return 0;
761 if (vm86_irqs[irqnumber].tsk != current) return 0;
762 spin_lock_irqsave(&irqbits_lock, flags);
763 bit = irqbits & (1 << irqnumber);
766 enable_irq(irqnumber);
770 spin_unlock_irqrestore(&irqbits_lock, flags);
775 static int do_vm86_irq_handling(int subfunction, int irqnumber)
778 switch (subfunction) {
779 case VM86_GET_AND_RESET_IRQ: {
780 return get_and_reset_irq(irqnumber);
782 case VM86_GET_IRQ_BITS: {
785 case VM86_REQUEST_IRQ: {
786 int sig = irqnumber >> 8;
787 int irq = irqnumber & 255;
788 if (!capable(CAP_SYS_ADMIN)) return -EPERM;
789 if (!((1 << sig) & ALLOWED_SIGS)) return -EPERM;
790 if (invalid_vm86_irq(irq)) return -EPERM;
791 if (vm86_irqs[irq].tsk) return -EPERM;
792 ret = request_irq(irq, &irq_handler, 0, VM86_IRQNAME, NULL);
794 vm86_irqs[irq].sig = sig;
795 vm86_irqs[irq].tsk = current;
798 case VM86_FREE_IRQ: {
799 if (invalid_vm86_irq(irqnumber)) return -EPERM;
800 if (!vm86_irqs[irqnumber].tsk) return 0;
801 if (vm86_irqs[irqnumber].tsk != current) return -EPERM;
802 free_vm86_irq(irqnumber);