During initialization, the channel initialization code schedules the
tasklet to scan the VMBUS receive event page (i.e. simulates an
interrupt). The problem was that it invokes the tasklet on a different
CPU from where it normally runs and therefore if an event is present,
it will clear the bit but not find the associated channel.
This can lead to missed events, typically stuck tasks, during bootup
when sub channels are being initialized. Typically seen as stuck
boot with 8 or more CPU's.
This patch is not necessary for upstream (4.11 and later) since
commit 631e63a9f346 ("vmbus: change to per channel tasklet").
This changed vmbus code to get rid of common tasklet which
caused the problem.
Cc: stable@xxxxxxxxxxxxxxx
Fixes: 638fea33aee8 ("Drivers: hv: vmbus: fix the race when querying & updating the percpu list")
Signed-off-by: Stephen Hemminger <sthemmin@xxxxxxxxxxxxx>
---
v2 - fix generic in hv_event_tasklet_enable
drivers/hv/channel_mgmt.c | 13 +++++++++++--
1 file changed, 11 insertions(+), 2 deletions(-)
diff --git a/drivers/hv/channel_mgmt.c b/drivers/hv/channel_mgmt.c
index a58cd102af1b..51be9ec8a51b 100644
--- a/drivers/hv/channel_mgmt.c
+++ b/drivers/hv/channel_mgmt.c
@@ -361,8 +361,17 @@ void hv_event_tasklet_enable(struct vmbus_channel *channel)
tasklet = hv_context.event_dpc[channel->target_cpu];
tasklet_enable(tasklet);
- /* In case there is any pending event */
- tasklet_schedule(tasklet);
+ /*
+ * In case there is any pending event schedule a rescan
+ * but must be on the correct CPU for the channel.
+ */
+ if (channel->target_cpu == get_cpu())
+ tasklet_schedule(tasklet);
+ else
+ smp_call_function_single(channel->target_cpu,
+ (smp_call_func_t)tasklet_schedule,
+ tasklet, false);
+ put_cpu();
}
void hv_process_channel_removal(struct vmbus_channel *channel, u32 relid)
--
2.11.0