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/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>
45 #include <linux/audit.h>
47 #include <asm/uaccess.h>
49 #include <asm/tlbflush.h>
55 * Interrupt handling is not guaranteed:
56 * - a real x86 will disable all interrupts for one instruction
57 * after a "mov ss,xx" to make stack handling atomic even without
58 * the 'lss' instruction. We can't guarantee this in v86 mode,
59 * as the next instruction might result in a page fault or similar.
60 * - a real x86 will have interrupts disabled for one instruction
61 * past the 'sti' that enables them. We don't bother with all the
64 * Let's hope these problems do not actually matter for anything.
68 #define KVM86 ((struct kernel_vm86_struct *)regs)
69 #define VMPI KVM86->vm86plus
73 * 8- and 16-bit register defines..
75 #define AL(regs) (((unsigned char *)&((regs)->eax))[0])
76 #define AH(regs) (((unsigned char *)&((regs)->eax))[1])
77 #define IP(regs) (*(unsigned short *)&((regs)->eip))
78 #define SP(regs) (*(unsigned short *)&((regs)->esp))
81 * virtual flags (16 and 32-bit versions)
83 #define VFLAGS (*(unsigned short *)&(current->thread.v86flags))
84 #define VEFLAGS (current->thread.v86flags)
86 #define set_flags(X,new,mask) \
87 ((X) = ((X) & ~(mask)) | ((new) & (mask)))
89 #define SAFE_MASK (0xDD5)
90 #define RETURN_MASK (0xDFF)
92 #define VM86_REGS_PART2 orig_eax
93 #define VM86_REGS_SIZE1 \
94 ( (unsigned)( & (((struct kernel_vm86_regs *)0)->VM86_REGS_PART2) ) )
95 #define VM86_REGS_SIZE2 (sizeof(struct kernel_vm86_regs) - VM86_REGS_SIZE1)
97 struct pt_regs * FASTCALL(save_v86_state(struct kernel_vm86_regs * regs));
98 struct pt_regs * fastcall save_v86_state(struct kernel_vm86_regs * regs)
100 struct tss_struct *tss;
105 * This gets called from entry.S with interrupts disabled, but
106 * from process context. Enable interrupts here, before trying
107 * to access user space.
111 if (!current->thread.vm86_info) {
112 printk("no vm86_info: BAD\n");
115 set_flags(regs->eflags, VEFLAGS, VIF_MASK | current->thread.v86mask);
116 tmp = copy_to_user(¤t->thread.vm86_info->regs,regs, VM86_REGS_SIZE1);
117 tmp += copy_to_user(¤t->thread.vm86_info->regs.VM86_REGS_PART2,
118 ®s->VM86_REGS_PART2, VM86_REGS_SIZE2);
119 tmp += put_user(current->thread.screen_bitmap,¤t->thread.vm86_info->screen_bitmap);
121 printk("vm86: could not access userspace vm86_info\n");
125 tss = &per_cpu(init_tss, get_cpu());
126 current->thread.esp0 = current->thread.saved_esp0;
127 current->thread.sysenter_cs = __KERNEL_CS;
128 load_esp0(tss, ¤t->thread);
129 current->thread.saved_esp0 = 0;
132 loadsegment(fs, current->thread.saved_fs);
133 loadsegment(gs, current->thread.saved_gs);
138 static void mark_screen_rdonly(struct mm_struct *mm)
147 pgd = pgd_offset(mm, 0xA0000);
148 if (pgd_none_or_clear_bad(pgd))
150 pud = pud_offset(pgd, 0xA0000);
151 if (pud_none_or_clear_bad(pud))
153 pmd = pmd_offset(pud, 0xA0000);
154 if (pmd_none_or_clear_bad(pmd))
156 pte = pte_offset_map_lock(mm, pmd, 0xA0000, &ptl);
157 for (i = 0; i < 32; i++) {
158 if (pte_present(*pte))
159 set_pte(pte, pte_wrprotect(*pte));
162 pte_unmap_unlock(pte, ptl);
169 static int do_vm86_irq_handling(int subfunction, int irqnumber);
170 static void do_sys_vm86(struct kernel_vm86_struct *info, struct task_struct *tsk);
172 asmlinkage int sys_vm86old(struct pt_regs regs)
174 struct vm86_struct __user *v86 = (struct vm86_struct __user *)regs.ebx;
175 struct kernel_vm86_struct info; /* declare this _on top_,
176 * this avoids wasting of stack space.
177 * This remains on the stack until we
178 * return to 32 bit user space.
180 struct task_struct *tsk;
181 int tmp, ret = -EPERM;
184 if (tsk->thread.saved_esp0)
186 tmp = copy_from_user(&info, v86, VM86_REGS_SIZE1);
187 tmp += copy_from_user(&info.regs.VM86_REGS_PART2, &v86->regs.VM86_REGS_PART2,
188 (long)&info.vm86plus - (long)&info.regs.VM86_REGS_PART2);
192 memset(&info.vm86plus, 0, (int)&info.regs32 - (int)&info.vm86plus);
194 tsk->thread.vm86_info = v86;
195 do_sys_vm86(&info, tsk);
196 ret = 0; /* we never return here */
202 asmlinkage int sys_vm86(struct pt_regs regs)
204 struct kernel_vm86_struct info; /* declare this _on top_,
205 * this avoids wasting of stack space.
206 * This remains on the stack until we
207 * return to 32 bit user space.
209 struct task_struct *tsk;
211 struct vm86plus_struct __user *v86;
215 case VM86_REQUEST_IRQ:
217 case VM86_GET_IRQ_BITS:
218 case VM86_GET_AND_RESET_IRQ:
219 ret = do_vm86_irq_handling(regs.ebx, (int)regs.ecx);
221 case VM86_PLUS_INSTALL_CHECK:
222 /* NOTE: on old vm86 stuff this will return the error
223 from access_ok(), because the subfunction is
224 interpreted as (invalid) address to vm86_struct.
225 So the installation check works.
231 /* we come here only for functions VM86_ENTER, VM86_ENTER_NO_BYPASS */
233 if (tsk->thread.saved_esp0)
235 v86 = (struct vm86plus_struct __user *)regs.ecx;
236 tmp = copy_from_user(&info, v86, VM86_REGS_SIZE1);
237 tmp += copy_from_user(&info.regs.VM86_REGS_PART2, &v86->regs.VM86_REGS_PART2,
238 (long)&info.regs32 - (long)&info.regs.VM86_REGS_PART2);
243 info.vm86plus.is_vm86pus = 1;
244 tsk->thread.vm86_info = (struct vm86_struct __user *)v86;
245 do_sys_vm86(&info, tsk);
246 ret = 0; /* we never return here */
252 static void do_sys_vm86(struct kernel_vm86_struct *info, struct task_struct *tsk)
254 struct tss_struct *tss;
257 * make sure the vm86() system call doesn't try to do anything silly
259 info->regs.__null_ds = 0;
260 info->regs.__null_es = 0;
262 /* we are clearing fs,gs later just before "jmp resume_userspace",
263 * because starting with Linux 2.1.x they aren't no longer saved/restored
267 * The eflags register is also special: we cannot trust that the user
268 * has set it up safely, so this makes sure interrupt etc flags are
269 * inherited from protected mode.
271 VEFLAGS = info->regs.eflags;
272 info->regs.eflags &= SAFE_MASK;
273 info->regs.eflags |= info->regs32->eflags & ~SAFE_MASK;
274 info->regs.eflags |= VM_MASK;
276 switch (info->cpu_type) {
278 tsk->thread.v86mask = 0;
281 tsk->thread.v86mask = NT_MASK | IOPL_MASK;
284 tsk->thread.v86mask = AC_MASK | NT_MASK | IOPL_MASK;
287 tsk->thread.v86mask = ID_MASK | AC_MASK | NT_MASK | IOPL_MASK;
292 * Save old state, set default return value (%eax) to 0
294 info->regs32->eax = 0;
295 tsk->thread.saved_esp0 = tsk->thread.esp0;
296 savesegment(fs, tsk->thread.saved_fs);
297 savesegment(gs, tsk->thread.saved_gs);
299 tss = &per_cpu(init_tss, get_cpu());
300 tsk->thread.esp0 = (unsigned long) &info->VM86_TSS_ESP0;
302 tsk->thread.sysenter_cs = 0;
303 load_esp0(tss, &tsk->thread);
306 tsk->thread.screen_bitmap = info->screen_bitmap;
307 if (info->flags & VM86_SCREEN_BITMAP)
308 mark_screen_rdonly(tsk->mm);
309 __asm__ __volatile__("xorl %eax,%eax; movl %eax,%fs; movl %eax,%gs\n\t");
310 __asm__ __volatile__("movl %%eax, %0\n" :"=r"(eax));
312 /*call audit_syscall_exit since we do not exit via the normal paths */
313 if (unlikely(current->audit_context))
314 audit_syscall_exit(AUDITSC_RESULT(eax), eax);
316 __asm__ __volatile__(
319 "jmp resume_userspace"
321 :"r" (&info->regs), "r" (task_thread_info(tsk)));
322 /* we never return here */
325 static inline void return_to_32bit(struct kernel_vm86_regs * regs16, int retval)
327 struct pt_regs * regs32;
329 regs32 = save_v86_state(regs16);
330 regs32->eax = retval;
331 __asm__ __volatile__("movl %0,%%esp\n\t"
333 "jmp resume_userspace"
334 : : "r" (regs32), "r" (current_thread_info()));
337 static inline void set_IF(struct kernel_vm86_regs * regs)
340 if (VEFLAGS & VIP_MASK)
341 return_to_32bit(regs, VM86_STI);
344 static inline void clear_IF(struct kernel_vm86_regs * regs)
346 VEFLAGS &= ~VIF_MASK;
349 static inline void clear_TF(struct kernel_vm86_regs * regs)
351 regs->eflags &= ~TF_MASK;
354 static inline void clear_AC(struct kernel_vm86_regs * regs)
356 regs->eflags &= ~AC_MASK;
359 /* It is correct to call set_IF(regs) from the set_vflags_*
360 * functions. However someone forgot to call clear_IF(regs)
361 * in the opposite case.
362 * After the command sequence CLI PUSHF STI POPF you should
363 * end up with interrups disabled, but you ended up with
364 * interrupts enabled.
365 * ( I was testing my own changes, but the only bug I
366 * could find was in a function I had not changed. )
370 static inline void set_vflags_long(unsigned long eflags, struct kernel_vm86_regs * regs)
372 set_flags(VEFLAGS, eflags, current->thread.v86mask);
373 set_flags(regs->eflags, eflags, SAFE_MASK);
374 if (eflags & IF_MASK)
380 static inline void set_vflags_short(unsigned short flags, struct kernel_vm86_regs * regs)
382 set_flags(VFLAGS, flags, current->thread.v86mask);
383 set_flags(regs->eflags, flags, SAFE_MASK);
390 static inline unsigned long get_vflags(struct kernel_vm86_regs * regs)
392 unsigned long flags = regs->eflags & RETURN_MASK;
394 if (VEFLAGS & VIF_MASK)
397 return flags | (VEFLAGS & current->thread.v86mask);
400 static inline int is_revectored(int nr, struct revectored_struct * bitmap)
402 __asm__ __volatile__("btl %2,%1\n\tsbbl %0,%0"
404 :"m" (*bitmap),"r" (nr));
408 #define val_byte(val, n) (((__u8 *)&val)[n])
410 #define pushb(base, ptr, val, err_label) \
414 if (put_user(__val, base + ptr) < 0) \
418 #define pushw(base, ptr, val, err_label) \
422 if (put_user(val_byte(__val, 1), base + ptr) < 0) \
425 if (put_user(val_byte(__val, 0), base + ptr) < 0) \
429 #define pushl(base, ptr, val, err_label) \
433 if (put_user(val_byte(__val, 3), base + ptr) < 0) \
436 if (put_user(val_byte(__val, 2), base + ptr) < 0) \
439 if (put_user(val_byte(__val, 1), base + ptr) < 0) \
442 if (put_user(val_byte(__val, 0), base + ptr) < 0) \
446 #define popb(base, ptr, err_label) \
449 if (get_user(__res, base + ptr) < 0) \
455 #define popw(base, ptr, err_label) \
458 if (get_user(val_byte(__res, 0), base + ptr) < 0) \
461 if (get_user(val_byte(__res, 1), base + ptr) < 0) \
467 #define popl(base, ptr, err_label) \
470 if (get_user(val_byte(__res, 0), base + ptr) < 0) \
473 if (get_user(val_byte(__res, 1), base + ptr) < 0) \
476 if (get_user(val_byte(__res, 2), base + ptr) < 0) \
479 if (get_user(val_byte(__res, 3), base + ptr) < 0) \
485 /* There are so many possible reasons for this function to return
486 * VM86_INTx, so adding another doesn't bother me. We can expect
487 * userspace programs to be able to handle it. (Getting a problem
488 * in userspace is always better than an Oops anyway.) [KD]
490 static void do_int(struct kernel_vm86_regs *regs, int i,
491 unsigned char __user * ssp, unsigned short sp)
493 unsigned long __user *intr_ptr;
494 unsigned long segoffs;
496 if (regs->cs == BIOSSEG)
498 if (is_revectored(i, &KVM86->int_revectored))
500 if (i==0x21 && is_revectored(AH(regs),&KVM86->int21_revectored))
502 intr_ptr = (unsigned long __user *) (i << 2);
503 if (get_user(segoffs, intr_ptr))
505 if ((segoffs >> 16) == BIOSSEG)
507 pushw(ssp, sp, get_vflags(regs), cannot_handle);
508 pushw(ssp, sp, regs->cs, cannot_handle);
509 pushw(ssp, sp, IP(regs), cannot_handle);
510 regs->cs = segoffs >> 16;
512 IP(regs) = segoffs & 0xffff;
519 return_to_32bit(regs, VM86_INTx + (i << 8));
522 int handle_vm86_trap(struct kernel_vm86_regs * regs, long error_code, int trapno)
524 if (VMPI.is_vm86pus) {
525 if ( (trapno==3) || (trapno==1) )
526 return_to_32bit(regs, VM86_TRAP + (trapno << 8));
527 do_int(regs, trapno, (unsigned char __user *) (regs->ss << 4), SP(regs));
531 return 1; /* we let this handle by the calling routine */
532 if (current->ptrace & PT_PTRACED) {
534 spin_lock_irqsave(¤t->sighand->siglock, flags);
535 sigdelset(¤t->blocked, SIGTRAP);
537 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
539 send_sig(SIGTRAP, current, 1);
540 current->thread.trap_no = trapno;
541 current->thread.error_code = error_code;
545 void handle_vm86_fault(struct kernel_vm86_regs * regs, long error_code)
547 unsigned char opcode;
548 unsigned char __user *csp;
549 unsigned char __user *ssp;
550 unsigned short ip, sp, orig_flags;
551 int data32, pref_done;
553 #define CHECK_IF_IN_TRAP \
554 if (VMPI.vm86dbg_active && VMPI.vm86dbg_TFpendig) \
556 #define VM86_FAULT_RETURN do { \
557 if (VMPI.force_return_for_pic && (VEFLAGS & (IF_MASK | VIF_MASK))) \
558 return_to_32bit(regs, VM86_PICRETURN); \
559 if (orig_flags & TF_MASK) \
560 handle_vm86_trap(regs, 0, 1); \
563 orig_flags = *(unsigned short *)®s->eflags;
565 csp = (unsigned char __user *) (regs->cs << 4);
566 ssp = (unsigned char __user *) (regs->ss << 4);
573 switch (opcode = popb(csp, ip, simulate_sigsegv)) {
574 case 0x66: /* 32-bit data */ data32=1; break;
575 case 0x67: /* 32-bit address */ break;
576 case 0x2e: /* CS */ break;
577 case 0x3e: /* DS */ break;
578 case 0x26: /* ES */ break;
579 case 0x36: /* SS */ break;
580 case 0x65: /* GS */ break;
581 case 0x64: /* FS */ break;
582 case 0xf2: /* repnz */ break;
583 case 0xf3: /* rep */ break;
584 default: pref_done = 1;
586 } while (!pref_done);
593 pushl(ssp, sp, get_vflags(regs), simulate_sigsegv);
596 pushw(ssp, sp, get_vflags(regs), simulate_sigsegv);
605 unsigned long newflags;
607 newflags=popl(ssp, sp, simulate_sigsegv);
610 newflags = popw(ssp, sp, simulate_sigsegv);
616 set_vflags_long(newflags, regs);
618 set_vflags_short(newflags, regs);
625 int intno=popb(csp, ip, simulate_sigsegv);
627 if (VMPI.vm86dbg_active) {
628 if ( (1 << (intno &7)) & VMPI.vm86dbg_intxxtab[intno >> 3] )
629 return_to_32bit(regs, VM86_INTx + (intno << 8));
631 do_int(regs, intno, ssp, sp);
640 unsigned long newflags;
642 newip=popl(ssp, sp, simulate_sigsegv);
643 newcs=popl(ssp, sp, simulate_sigsegv);
644 newflags=popl(ssp, sp, simulate_sigsegv);
647 newip = popw(ssp, sp, simulate_sigsegv);
648 newcs = popw(ssp, sp, simulate_sigsegv);
649 newflags = popw(ssp, sp, simulate_sigsegv);
656 set_vflags_long(newflags, regs);
658 set_vflags_short(newflags, regs);
671 * Damn. This is incorrect: the 'sti' instruction should actually
672 * enable interrupts after the /next/ instruction. Not good.
674 * Probably needs some horsing around with the TF flag. Aiee..
682 return_to_32bit(regs, VM86_UNKNOWN);
688 /* FIXME: After a long discussion with Stas we finally
689 * agreed, that this is wrong. Here we should
690 * really send a SIGSEGV to the user program.
691 * But how do we create the correct context? We
692 * are inside a general protection fault handler
693 * and has just returned from a page fault handler.
694 * The correct context for the signal handler
695 * should be a mixture of the two, but how do we
696 * get the information? [KD]
698 return_to_32bit(regs, VM86_UNKNOWN);
701 /* ---------------- vm86 special IRQ passing stuff ----------------- */
703 #define VM86_IRQNAME "vm86irq"
705 static struct vm86_irqs {
706 struct task_struct *tsk;
710 static DEFINE_SPINLOCK(irqbits_lock);
713 #define ALLOWED_SIGS ( 1 /* 0 = don't send a signal */ \
714 | (1 << SIGUSR1) | (1 << SIGUSR2) | (1 << SIGIO) | (1 << SIGURG) \
717 static irqreturn_t irq_handler(int intno, void *dev_id)
722 spin_lock_irqsave(&irqbits_lock, flags);
723 irq_bit = 1 << intno;
724 if ((irqbits & irq_bit) || ! vm86_irqs[intno].tsk)
727 if (vm86_irqs[intno].sig)
728 send_sig(vm86_irqs[intno].sig, vm86_irqs[intno].tsk, 1);
730 * IRQ will be re-enabled when user asks for the irq (whether
731 * polling or as a result of the signal)
733 disable_irq_nosync(intno);
734 spin_unlock_irqrestore(&irqbits_lock, flags);
738 spin_unlock_irqrestore(&irqbits_lock, flags);
742 static inline void free_vm86_irq(int irqnumber)
746 free_irq(irqnumber, NULL);
747 vm86_irqs[irqnumber].tsk = NULL;
749 spin_lock_irqsave(&irqbits_lock, flags);
750 irqbits &= ~(1 << irqnumber);
751 spin_unlock_irqrestore(&irqbits_lock, flags);
754 void release_vm86_irqs(struct task_struct *task)
757 for (i = FIRST_VM86_IRQ ; i <= LAST_VM86_IRQ; i++)
758 if (vm86_irqs[i].tsk == task)
762 static inline int get_and_reset_irq(int irqnumber)
768 if (invalid_vm86_irq(irqnumber)) return 0;
769 if (vm86_irqs[irqnumber].tsk != current) return 0;
770 spin_lock_irqsave(&irqbits_lock, flags);
771 bit = irqbits & (1 << irqnumber);
774 enable_irq(irqnumber);
778 spin_unlock_irqrestore(&irqbits_lock, flags);
783 static int do_vm86_irq_handling(int subfunction, int irqnumber)
786 switch (subfunction) {
787 case VM86_GET_AND_RESET_IRQ: {
788 return get_and_reset_irq(irqnumber);
790 case VM86_GET_IRQ_BITS: {
793 case VM86_REQUEST_IRQ: {
794 int sig = irqnumber >> 8;
795 int irq = irqnumber & 255;
796 if (!capable(CAP_SYS_ADMIN)) return -EPERM;
797 if (!((1 << sig) & ALLOWED_SIGS)) return -EPERM;
798 if (invalid_vm86_irq(irq)) return -EPERM;
799 if (vm86_irqs[irq].tsk) return -EPERM;
800 ret = request_irq(irq, &irq_handler, 0, VM86_IRQNAME, NULL);
802 vm86_irqs[irq].sig = sig;
803 vm86_irqs[irq].tsk = current;
806 case VM86_FREE_IRQ: {
807 if (invalid_vm86_irq(irqnumber)) return -EPERM;
808 if (!vm86_irqs[irqnumber].tsk) return 0;
809 if (vm86_irqs[irqnumber].tsk != current) return -EPERM;
810 free_vm86_irq(irqnumber);