timer: Fix jiffies wrap behavior of round_jiffies*()

Make sure that the round_jiffies*() functions return a time that is in the future when the jiffies counter is about to wrap.

---
 kernel/timer.c |    4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

index 15ffdb3..aa8b964 100644
@@ -149,9 +149,7 @@ static unsigned long round_jiffies_common(unsigned long j, int cpu,
 	/* now that we have rounded, subtract the extra skew again */
 	j -= cpu * 3;
-	if (j <= jiffies) /* rounding ate our timeout entirely; */
-		return original;
-	return j;
+	return time_is_after_jiffies(j) ? j : original;
 }
  /**

1.7.10.4

Actually "when the jiffies counter has recently wrapped".

I assume this was found by inspection?

This list probably isn't comprehensive.

$ git grep -E "if\s*\((\s*jiffies\s*[\<\>]|.*[\<\>]=?\s*jiffies\b)"
drivers/acpi/acpi_pad.c:                if (last_jiffies + round_robin_time * HZ < jiffies) {
drivers/acpi/acpi_pad.c:                        if (jiffies > expire_time) {
drivers/dma/ioat/dma_v2.c:      if (jiffies > chan->timer.expires && timer_pending(&chan->timer)) {
drivers/infiniband/hw/ipath/ipath_sdma.c:               if (jiffies < dd->ipath_sdma_abort_intr_timeout)
drivers/infiniband/hw/ipath/ipath_sdma.c:       if (jiffies > dd->ipath_sdma_abort_jiffies) {
drivers/infiniband/ulp/ipoib/ipoib_cm.c:                if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE
drivers/infiniband/ulp/ipoib/ipoib_cm.c:                if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
drivers/input/misc/ati_remote2.c:               if (!time_after_eq(jiffies, ar2->jiffies))
drivers/md/dm-log-userspace-base.c:     else if (jiffies < limit)
drivers/misc/sgi-gru/grumain.c:                 if (gts->ts_steal_jiffies + GRU_STEAL_DELAY < jiffies)
drivers/net/ethernet/intel/e1000/e1000_ethtool.c:               if (jiffies >= (time + 2)) {
drivers/net/ethernet/intel/e1000e/ethtool.c:            if (jiffies >= (time + 20)) {
drivers/net/ethernet/micrel/ksz884x.c:                  if (next_jiffies < jiffies)
drivers/net/ethernet/micrel/ksz884x.c:          } else if (jiffies >= hw_priv->counter[i].time) {
drivers/net/ethernet/micrel/ksz884x.c:          if (hw_priv->pme_wait <= jiffies) {
drivers/net/ethernet/neterion/vxge/vxge-main.c: if (jiffies > fifo->jiffies + HZ / 100) {
drivers/net/ethernet/neterion/vxge/vxge-main.c: if (jiffies > ring->jiffies + HZ / 100) {
drivers/net/ethernet/qlogic/qlcnic/qlcnic_hw.c:                 if (jiffies > (QLCNIC_FILTER_AGE * HZ + time)) {
drivers/net/ethernet/qlogic/qlcnic/qlcnic_hw.c:                 if (jiffies > (QLCNIC_FILTER_AGE * HZ + time)) {
drivers/net/ethernet/qlogic/qlcnic/qlcnic_io.c:                         if (jiffies > (QLCNIC_READD_AGE * HZ + time))
drivers/net/ethernet/qlogic/qlcnic/qlcnic_io.c:                 if (jiffies > (QLCNIC_READD_AGE * HZ + tmp_fil->ftime))
drivers/scsi/lpfc/lpfc_scsi.c:  if ((phba->last_ramp_down_time + QUEUE_RAMP_DOWN_INTERVAL) > jiffies) {
drivers/staging/bcm/InterfaceIdleMode.c:                if (timeout < jiffies )
drivers/staging/lustre/include/linux/libcfs/linux/linux-prim.h:                 if (jiffies > expire) {           \
drivers/staging/speakup/speakup_acntpc.c:               if (jiffies >= jiff_max && ch == SPACE) {
drivers/staging/speakup/speakup_decext.c:                       if (jiffies >= jiff_max) {
drivers/staging/speakup/speakup_decpc.c:                        if (jiffies >= jiff_max) {
drivers/staging/speakup/speakup_dectlk.c:                       if (jiffies >= jiff_max) {
kernel/timer.c: if (j <= jiffies) /* rounding ate our timeout entirely; */
net/rds/ib_send.c:                      if (ic->i_ack_queued + HZ/2 < jiffies)
net/rds/ib_send.c:                      if (send->s_queued + HZ/2 < jiffies)
net/rds/iw_send.c:                      if (ic->i_ack_queued + HZ/2 < jiffies)
net/rds/iw_send.c:                      if (send->s_queued + HZ/2 < jiffies)

other variables like?

Grepping for time_after finds a bunch. There's no real pattern to it though.

Yup, that's the problem.
Thought I'd ask what I was missing though.

No variable really stands out as testable other
than jiffies.

I added this to my local queue and I'll send it later.

# check for comparisons of jiffies
		if ($line =~ /\bif\s*\((\s*jiffies\s*[\<\>]|.*[\<\>]=?\s*jiffies\b)/) {
			WARN("JIFFIES_COMPARISON",
			     "Comparing jiffies is almost always wrong; prefer time_after, time_before and friends\n" . $herecurr);
		}

get_jiffies_64() should probably be added as
another test too.

Also, these might be wrong:

arch/arm/kernel/smp_twd.c:              while (get_jiffies_64() < waitjiffies)
arch/arm/kernel/smp_twd.c:              while (get_jiffies_64() < waitjiffies)
fs/fuse/dir.c:  else if (fuse_dentry_time(entry) < get_jiffies_64()) {
fs/fuse/dir.c:  if (fi->i_time < get_jiffies_64()) {
fs/fuse/dir.c:          if (fi->i_time < get_jiffies_64()) {

Yup. Normally a 64-bit jiffy will wrap around shortly after the heat death of the universe, but

a) it's derived from jiffies, which we evilly cause to wrap after 5 minutes uptime and

b) it's derived from jiffies, which is 32-bit on 32-bit and hence wraps every 49 days (HZ=1000).

Hello Andrew,

You are correct, this was found via source code inspection. I started reviewing the round_jiffies*() implementation because I was chasing a bug in a kernel driver using one of these functions.

Sorry. Will take more care in the future.

Bart.