Resending patch which also applies to 2.6.11-rc4. On Thu, 2004-12-23 at 19:26, Matthew Wilcox wrote: > On Thu, Dec 23, 2004 at 03:01:20PM -0800, Bruce Allan wrote: > > A number of Connectathon (http://www.connectathon.org/nfstests.html) > > POSIX/fcntl() locking tests fail (even on local filesystems) at various > > edge cases (i.e. around maximum allowable offsets) on 64-bit > > architectures. > > OK, that's bad and needs to be fixed. > > > The overflow tests in fs/compat.c were superfluous where they were > > located because if there was a conflicting lock, l_start and l_len would > > have been overwritten with the values owned by the conflicting lock; if > > no conflicting lock, sys_fcntl() would have returned any applicable > > error. The tests are moved above the call to sys_fcntl() to capture > > overflow errors which would not have been caught by sys_fcntl(), eg. > > obvious overflow when _FILE_OFFSET_BITS == 32. > > I don't buy this explanation though. With your patch, we're testing > the lock the user passed in to see if it'd overflow. Clearly, that > can never happen. The checks are supposed to be testing whether the > conflicting lock is outside the limits that a program using a 32-bit > off_t can cope with.
Hi Matthew, I now agree with you not buying my initial explanation, but have some follow-up comments/questions. How is it clear that a user can never pass in a lock request that would overflow? AFAICT, there is no reason a 32-bit application could not pass down a request for a lock (eg. l_whence=SEEK_SET, l_start=0x7fffffff, l_len=2) which should overflow. For 64-bit applications on a 64-bit architecture (and 32-bit apps/32-bit archs) any overflow error in a lock request would be caught in flock_to_posix_lock() whether or not there is a conflicting lock. For 32-bit applications with 32-bit userland off_t on a 64-bit architecture, flock_to_posix_lock() does not capture overflow errors on requested locks because that function is checking for overflow assuming it is a 64-bit value. If there is no conflicting lock these values are passed back to compat_sys_fcntl64() and the overflow check in that function should catch it (provided l_whence==SEEK_SET, otherwise that check may not catch it), but if there is a conflicting lock the overflow check in compat_sys_fcntl64() is checking the values on the conflicting lock instead of the requested lock. I do see your point now about compat_sys_fcntl64() "checks are supposed to be testing whether the conflicting lock is outside the limits that a program using a 32-bit off_t can cope with", but it seems to me there still needs to be an overflow check of the requested lock in compat_sys_fcntl64() for the case of F_GETLK/F_SETLK/F_SETLKW; something akin to the check in flock_to_posix_lock(). It's duplicate code, I know, but at the moment I'm at a loss for a better alternative (I suppose it could be setup instead to pass the maximum filesize all the way to flock_to_posix_lock()). This check would be unnecessary for F_GETLK64/F_SETLK64/F_SETLKW64 as it would be properly handled in flock_to_posix_lock(). Comments? > > These tests also had a couple 'off by one' errors when comparing with > > the maximum allowable offset. > > Perhaps just fix that, and don't move the tests around? [snip] > I hate assignments inside if statements. Make this: > > start += l->l_start; > if (start < 0) > return -EINVAL; Done. Signed-off-by: Bruce Allan <[EMAIL PROTECTED]> diff -Nurp -Xdontdiff linux-2.6.10-rc3-vanilla/fs/compat.c linux-2.6.10-rc3/fs/compat.c --- linux-2.6.10-rc3-vanilla/fs/compat.c 2004-12-23 11:52:50.000000000 -0800 +++ linux-2.6.10-rc3/fs/compat.c 2004-12-30 13:33:52.525317789 -0800 @@ -523,6 +523,40 @@ static int put_compat_flock64(struct flo } #endif +static int check_compat_flock(unsigned int fd, struct flock *l) +{ + compat_off_t start, end; + struct file *filp = fget(fd); + + switch (l->l_whence) { + case 0: /*SEEK_SET*/ + start = 0; + break; + case 1: /*SEEK_CUR*/ + start = filp->f_pos; + break; + case 2: /*SEEK_END*/ + start = i_size_read(filp->f_dentry->d_inode); + break; + default: + return -EINVAL; + } + + /* POSIX-1996 leaves the case l->l_len < 0 undefined; + POSIX-2001 defines it. */ + start += l->l_start; + end = start + l->l_len - 1; + if (l->l_len < 0) { + end = start - 1; + start += l->l_len; + } + if (start < 0) + return -EINVAL; + if (l->l_len > 0 && end < 0) + return -EOVERFLOW; + return 0; +} + asmlinkage long compat_sys_fcntl64(unsigned int fd, unsigned int cmd, unsigned long arg) { @@ -537,13 +571,18 @@ asmlinkage long compat_sys_fcntl64(unsig ret = get_compat_flock(&f, compat_ptr(arg)); if (ret != 0) break; + ret = check_compat_flock(fd, &f); + if (ret != 0) + break; old_fs = get_fs(); set_fs(KERNEL_DS); ret = sys_fcntl(fd, cmd, (unsigned long)&f); set_fs(old_fs); if (cmd == F_GETLK && ret == 0) { - if ((f.l_start >= COMPAT_OFF_T_MAX) || - ((f.l_start + f.l_len) > COMPAT_OFF_T_MAX)) + /* check for overflow of conflicting lock if any */ + if ((f.l_whence != F_UNLCK) && + ((f.l_start > COMPAT_OFF_T_MAX) || + ((f.l_start + f.l_len - 1) > COMPAT_OFF_T_MAX))) ret = -EOVERFLOW; if (ret == 0) ret = put_compat_flock(&f, compat_ptr(arg)); @@ -563,8 +602,10 @@ asmlinkage long compat_sys_fcntl64(unsig (unsigned long)&f); set_fs(old_fs); if (cmd == F_GETLK64 && ret == 0) { - if ((f.l_start >= COMPAT_LOFF_T_MAX) || - ((f.l_start + f.l_len) > COMPAT_LOFF_T_MAX)) + /* check for overflow of conflicting lock if any */ + if ((f.l_whence != F_UNLCK) && + ((f.l_start > COMPAT_LOFF_T_MAX) || + ((f.l_start + f.l_len - 1) > COMPAT_LOFF_T_MAX))) ret = -EOVERFLOW; if (ret == 0) ret = put_compat_flock64(&f, compat_ptr(arg)); diff -Nurp -Xdontdiff linux-2.6.10-rc3-vanilla/fs/locks.c linux-2.6.10-rc3/fs/locks.c --- linux-2.6.10-rc3-vanilla/fs/locks.c 2004-12-23 11:52:50.000000000 -0800 +++ linux-2.6.10-rc3/fs/locks.c 2004-12-30 13:33:52.529317402 -0800 @@ -315,6 +315,8 @@ static int flock_to_posix_lock(struct fi /* POSIX-1996 leaves the case l->l_len < 0 undefined; POSIX-2001 defines it. */ start += l->l_start; + if (start < 0) + return -EINVAL; end = start + l->l_len - 1; if (l->l_len < 0) { end = start - 1; --- Bruce Allan <[EMAIL PROTECTED]> Software Engineer, Linux Technology Center IBM Corporation, Beaverton OR USA - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/