svn commit: r280308 - head/sys/fs/devfs

[Konstantin Belousov]

On Sun, Mar 29, 2015 at 07:51:37PM +0200, Jilles Tjoelker wrote:
> On Sun, Mar 22, 2015 at 11:25:07AM -0700, Don Lewis wrote:
> > It's not totally worthless.  I think the mtime on tty devices is used to
> > calculate the idle time that is printed by the w command.  We just don't
> > need nanosecond accuracy for that.
> 
> Hmm. The idle time on tty devices breaking is a clear POLA violation,
> but I'm not sure what's the best way to fix it. By the way, w uses atime
> and who -u uses mtime.
> 
> Some options are:
> 
> * Bypass vfs_timestamp() and hard-code second accuracy in devfs;
>   futimens() will still set timestamps to the nanosecond, even with
>   UTIME_NOW. A timestamp update after UTIME_NOW setting may set back the
>   timestamp to an older value.
> 
> * Make vfs.timestamp_precision filesystem-specific. Since this does not
>   affect futimens() with explicit timestamps, it will not cause strange
>   situations with cp -p.
> 
> * Restrict file times on devfs to seconds.
> 
> * Somehow add a special case for TTY devices so they will always keep
>   timestamps.
> 
> * Somehow add a special case for "performance-critical" devices so they
>   will not keep timestamps.
> 
> * Change the default for vfs.timestamp_precision to 1, which still
>   provides non-zero nanoseconds (so much of the same bugs) but is not so
>   slow. The file server people won't like this though.
> 
> My proposal for delayed updates as in UFS clearly does not work for TTY
> idle times, so there is no point in that.
> 
Interesting complication with the devfs timestamp update is that
devfs_read_f() and devfs_write_f() do not lock the vnode. So whatever
update method is used, stat(2) on devfs might return inconsistent value,
since tv_src/tv_nsec cannot be updated or read by single op, without
locking.

That said, we could either:
- ignore the race.  Then the patch below might be good enough.
- Trim the precision of futimens() for devfs to always set tv_nsec to 0.
  Then, the race can be handled satisfactory with atomics.
- Add a lock (IMO this does not make sense).

diff --git a/sys/fs/devfs/devfs_vnops.c b/sys/fs/devfs/devfs_vnops.c
index 941f92c..e199d52 100644
--- a/sys/fs/devfs/devfs_vnops.c
+++ b/sys/fs/devfs/devfs_vnops.c
@@ -83,8 +83,24 @@ MTX_SYSINIT(cdevpriv_mtx, &cdevpriv_mtx, "cdevpriv lock", MTX_DEF);
 SYSCTL_DECL(_vfs_devfs);
 
 static int devfs_dotimes;
-SYSCTL_INT(_vfs_devfs, OID_AUTO, dotimes, CTLFLAG_RW,
-        &devfs_dotimes, 0, "Update timestamps on DEVFS");
+SYSCTL_INT(_vfs_devfs, OID_AUTO, dotimes, CTLFLAG_RW, &devfs_dotimes, 0,
+    "Update timestamps on DEVFS with default precision");
+
+static void
+devfs_timestamp(struct timespec *tsp)
+{
+	time_t ts;
+
+	if (devfs_dotimes) {
+		vfs_timestamp(tsp);
+	} else {
+		ts = time_second;
+		if (tsp->tv_sec < ts) {
+			tsp->tv_sec = ts;
+			tsp->tv_nsec = 0;
+		}
+	}
+}
 
 static int
 devfs_fp_check(struct file *fp, struct cdev **devp, struct cdevsw **dswp,
@@ -1228,9 +1244,8 @@ devfs_read_f(struct file *fp, struct uio *uio, struct ucred *cred,
 
 	foffset_lock_uio(fp, uio, flags | FOF_NOLOCK);
 	error = dsw->d_read(dev, uio, ioflag);
-	if (devfs_dotimes &&
-	    (uio->uio_resid != resid || (error == 0 && resid != 0)))
-		vfs_timestamp(&dev->si_atime);
+	if (uio->uio_resid != resid || (error == 0 && resid != 0))
+		devfs_timestamp(&dev->si_atime);
 	td->td_fpop = fpop;
 	dev_relthread(dev, ref);
 
@@ -1708,9 +1723,8 @@ devfs_write_f(struct file *fp, struct uio *uio, struct ucred *cred,
 	resid = uio->uio_resid;
 
 	error = dsw->d_write(dev, uio, ioflag);
-	if (devfs_dotimes &&
-	    (uio->uio_resid != resid || (error == 0 && resid != 0))) {
-		vfs_timestamp(&dev->si_ctime);
+	if (uio->uio_resid != resid || (error == 0 && resid != 0)) {
+		devfs_timestamp(&dev->si_ctime);
 		dev->si_mtime = dev->si_ctime;
 	}
 	td->td_fpop = fpop;