mm/mlock.c

32761 /*
32762  *      linux/mm/mlock.c
32763  *
32764  *  (C) Copyright 1995 Linus Torvalds
32765  */
32766 #include <linux/slab.h>
32767 #include <linux/shm.h>
32768 #include <linux/mman.h>
32769 #include <linux/smp_lock.h>
32770 
32771 #include <asm/uaccess.h>
32772 #include <asm/pgtable.h>
32773 
32774 static inline int mlock_fixup_all(
32775   struct vm_area_struct * vma, int newflags)
32776 {
32777   vma->vm_flags = newflags;
32778   return 0;
32779 }
32780 
32781 static inline int mlock_fixup_start(
32782   struct vm_area_struct * vma, unsigned long end,
32783   int newflags)
32784 {
32785   struct vm_area_struct * n;
32786 
32787   n = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
32788   if (!n)
32789     return -EAGAIN;
32790   *n = *vma;
32791   vma->vm_start = end;
32792   n->vm_end = end;
32793   vma->vm_offset += vma->vm_start - n->vm_start;
32794   n->vm_flags = newflags;
32795   if (n->vm_file)
32796     n->vm_file->f_count++;
32797   if (n->vm_ops && n->vm_ops->open)
32798     n->vm_ops->open(n);
32799   insert_vm_struct(current->mm, n);
32800   return 0;
32801 }
32802 
32803 static inline int mlock_fixup_end(
32804   struct vm_area_struct * vma,
32805   unsigned long start, int newflags)
32806 {
32807   struct vm_area_struct * n;
32808 
32809   n = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
32810   if (!n)
32811     return -EAGAIN;
32812   *n = *vma;
32813   vma->vm_end = start;
32814   n->vm_start = start;
32815   n->vm_offset += n->vm_start - vma->vm_start;
32816   n->vm_flags = newflags;
32817   if (n->vm_file)
32818     n->vm_file->f_count++;
32819   if (n->vm_ops && n->vm_ops->open)
32820     n->vm_ops->open(n);
32821   insert_vm_struct(current->mm, n);
32822   return 0;
32823 }
32824 
32825 static inline int mlock_fixup_middle(
32826   struct vm_area_struct * vma,
32827   unsigned long start, unsigned long end, int newflags)
32828 {
32829   struct vm_area_struct * left, * right;
32830 
32831   left = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
32832   if (!left)
32833     return -EAGAIN;
32834   right = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
32835   if (!right) {
32836     kmem_cache_free(vm_area_cachep, left);
32837     return -EAGAIN;
32838   }
32839   *left = *vma;
32840   *right = *vma;
32841   left->vm_end = start;
32842   vma->vm_start = start;
32843   vma->vm_end = end;
32844   right->vm_start = end;
32845   vma->vm_offset += vma->vm_start - left->vm_start;
32846   right->vm_offset += right->vm_start - left->vm_start;
32847   vma->vm_flags = newflags;
32848   if (vma->vm_file)
32849     vma->vm_file->f_count += 2;
32850 
32851   if (vma->vm_ops && vma->vm_ops->open) {
32852     vma->vm_ops->open(left);
32853     vma->vm_ops->open(right);
32854   }
32855   insert_vm_struct(current->mm, left);
32856   insert_vm_struct(current->mm, right);
32857   return 0;
32858 }
32859 
32860 static int mlock_fixup(struct vm_area_struct * vma,
32861   unsigned long start, unsigned long end,
32862   unsigned int newflags)
32863 {
32864   int pages, retval;
32865 
32866   if (newflags == vma->vm_flags)
32867     return 0;
32868 
32869   if (start == vma->vm_start) {
32870     if (end == vma->vm_end)
32871       retval = mlock_fixup_all(vma, newflags);
32872     else
32873       retval = mlock_fixup_start(vma, end, newflags);
32874   } else {
32875     if (end == vma->vm_end)
32876       retval = mlock_fixup_end(vma, start, newflags);
32877     else
32878       retval = mlock_fixup_middle(vma, start, end,
32879                                   newflags);
32880   }
32881   if (!retval) {
32882     /* keep track of amount of locked VM */
32883     pages = (end - start) >> PAGE_SHIFT;
32884     if (!(newflags & VM_LOCKED))
32885       pages = -pages;
32886     vma->vm_mm->locked_vm += pages;
32887     make_pages_present(start, end);
32888   }
32889   return retval;
32890 }
32891 
32892 static int do_mlock(unsigned long start, size_t len,
32893                     int on)
32894 {
32895   unsigned long nstart, end, tmp;
32896   struct vm_area_struct * vma, * next;
32897   int error;
32898 
32899   if (!capable(CAP_IPC_LOCK))
32900     return -EPERM;
32901   len = (len + ~PAGE_MASK) & PAGE_MASK;
32902   end = start + len;
32903   if (end < start)
32904     return -EINVAL;
32905   if (end == start)
32906     return 0;
32907   vma = find_vma(current->mm, start);
32908   if (!vma || vma->vm_start > start)
32909     return -ENOMEM;
32910 
32911   for (nstart = start ; ; ) {
32912     unsigned int newflags;
32913 
32914     /* Here we know that
32915      * vma->vm_start <= nstart < vma->vm_end. */
32916     newflags = vma->vm_flags | VM_LOCKED;
32917     if (!on)
32918       newflags &= ~VM_LOCKED;
32919 
32920     if (vma->vm_end >= end) {
32921       error = mlock_fixup(vma, nstart, end, newflags);
32922       break;
32923     }
32924 
32925     tmp = vma->vm_end;
32926     next = vma->vm_next;
32927     error = mlock_fixup(vma, nstart, tmp, newflags);
32928     if (error)
32929       break;
32930     nstart = tmp;
32931     vma = next;
32932     if (!vma || vma->vm_start != nstart) {
32933       error = -ENOMEM;
32934       break;
32935     }
32936   }
32937   merge_segments(current->mm, start, end);
32938   return error;
32939 }
32940 
32941 asmlinkage int sys_mlock(unsigned long start, size_t len)
32942 {
32943   unsigned long locked;
32944   unsigned long lock_limit;
32945   int error = -ENOMEM;
32946 
32947   down(&current->mm->mmap_sem);
32948   lock_kernel();
32949   len = (len + (start & ~PAGE_MASK) + ~PAGE_MASK) &
32950         PAGE_MASK;
32951   start &= PAGE_MASK;
32952 
32953   locked = len >> PAGE_SHIFT;
32954   locked += current->mm->locked_vm;
32955 
32956   lock_limit = current->rlim[RLIMIT_MEMLOCK].rlim_cur;
32957   lock_limit >>= PAGE_SHIFT;
32958 
32959   /* check against resource limits */
32960   if (locked > lock_limit)
32961     goto out;
32962 
32963   /* we may lock at most half of physical memory... */
32964   /* (this check is pretty bogus, but doesn't hurt) */
32965   if (locked > num_physpages/2)
32966     goto out;
32967 
32968   error = do_mlock(start, len, 1);
32969 out:
32970   unlock_kernel();
32971   up(&current->mm->mmap_sem);
32972   return error;
32973 }
32974 
32975 asmlinkage int sys_munlock(unsigned long start,
32976                            size_t len)
32977 {
32978   int ret;
32979 
32980   down(&current->mm->mmap_sem);
32981   lock_kernel();
32982   len = (len + (start & ~PAGE_MASK) + ~PAGE_MASK) &
32983         PAGE_MASK;
32984   start &= PAGE_MASK;
32985   ret = do_mlock(start, len, 0);
32986   unlock_kernel();
32987   up(&current->mm->mmap_sem);
32988   return ret;
32989 }
32990 
32991 static int do_mlockall(int flags)
32992 {
32993   int error;
32994   unsigned int def_flags;
32995   struct vm_area_struct * vma;
32996 
32997   if (!capable(CAP_IPC_LOCK))
32998     return -EPERM;
32999 
33000   def_flags = 0;
33001   if (flags & MCL_FUTURE)
33002     def_flags = VM_LOCKED;
33003   current->mm->def_flags = def_flags;
33004 
33005   error = 0;
33006   for (vma = current->mm->mmap; vma;
33007        vma = vma->vm_next) {
33008     unsigned int newflags;
33009 
33010     newflags = vma->vm_flags | VM_LOCKED;
33011     if (!(flags & MCL_CURRENT))
33012       newflags &= ~VM_LOCKED;
33013     error = mlock_fixup(vma, vma->vm_start, vma->vm_end,
33014                         newflags);
33015     if (error)
33016       break;
33017   }
33018   merge_segments(current->mm, 0, TASK_SIZE);
33019   return error;
33020 }
33021 
33022 asmlinkage int sys_mlockall(int flags)
33023 {
33024   unsigned long lock_limit;
33025   int ret = -EINVAL;
33026 
33027   down(&current->mm->mmap_sem);
33028   lock_kernel();
33029   if (!flags || (flags & ~(MCL_CURRENT | MCL_FUTURE)))
33030     goto out;
33031 
33032   lock_limit = current->rlim[RLIMIT_MEMLOCK].rlim_cur;
33033   lock_limit >>= PAGE_SHIFT;
33034 
33035   ret = -ENOMEM;
33036   if (current->mm->total_vm > lock_limit)
33037     goto out;
33038 
33039   /* we may lock at most half of physical memory... */
33040   /* (this check is pretty bogus, but doesn't hurt) */
33041   if (current->mm->total_vm > num_physpages/2)
33042     goto out;
33043 
33044   ret = do_mlockall(flags);
33045 out:
33046   unlock_kernel();
33047   up(&current->mm->mmap_sem);
33048   return ret;
33049 }
33050 
33051 asmlinkage int sys_munlockall(void)
33052 {
33053   int ret;
33054 
33055   down(&current->mm->mmap_sem);
33056   lock_kernel();
33057   ret = do_mlockall(0);
33058   unlock_kernel();
33059   up(&current->mm->mmap_sem);
33060   return ret;
33061 }