With the recent introduction of per-channel tasklet, we need to update
the way we handle the 3 concurrency issues:
1. hv_process_channel_removal -> percpu_channel_deq vs.
vmbus_chan_sched -> list_for_each_entry(..., percpu_list);
2. vmbus_process_offer -> percpu_channel_enq/deq vs. vmbus_chan_sched.
3. vmbus_close_internal vs. the per-channel tasklet vmbus_on_event;
The first 2 issues can be handled by Stephen's recent patch
"vmbus: use rcu for per-cpu channel list", and the third issue
can be handled by calling tasklet_disable in vmbus_close_internal here.
We don't need the original hv_event_tasklet_disable/enable since we
now use per-channel tasklet instead of the previous per-CPU tasklet,
and actually we must remove them due to the side effect now:
vmbus_process_offer -> hv_event_tasklet_enable -> tasklet_schedule will
start the per-channel callback prematurely, cauing NULL dereferencing
(the channel may haven't been properly configured to run the callback yet).
Fixes: 631e63a9f346 ("vmbus: change to per channel tasklet")
Signed-off-by: Dexuan Cui <[email protected]>
Cc: "K. Y. Srinivasan" <[email protected]>
Cc: Haiyang Zhang <[email protected]>
Cc: Stephen Hemminger <[email protected]>
---
Stephen Hemminger's patch ("vmbus: use rcu for per-cpu channel list")
was posted to the community last month, but it hasn't been in char-misc
yet, as far as I know.
drivers/hv/channel.c | 12 ++++--------
drivers/hv/channel_mgmt.c | 19 -------------------
include/linux/hyperv.h | 3 ---
3 files changed, 4 insertions(+), 30 deletions(-)
diff --git a/drivers/hv/channel.c b/drivers/hv/channel.c
index bd0d198..57b2958 100644
--- a/drivers/hv/channel.c
+++ b/drivers/hv/channel.c
@@ -530,15 +530,13 @@ static int vmbus_close_internal(struct vmbus_channel *channel)
int ret;
/*
- * vmbus_on_event(), running in the tasklet, can race
+ * vmbus_on_event(), running in the per-channel tasklet, can race
* with vmbus_close_internal() in the case of SMP guest, e.g., when
* the former is accessing channel->inbound.ring_buffer, the latter
- * could be freeing the ring_buffer pages.
- *
- * To resolve the race, we can serialize them by disabling the
- * tasklet when the latter is running here.
+ * could be freeing the ring_buffer pages, so here we must stop it
+ * first.
*/
- hv_event_tasklet_disable(channel);
+ tasklet_disable(&channel->callback_event);
/*
* In case a device driver's probe() fails (e.g.,
@@ -605,8 +603,6 @@ static int vmbus_close_internal(struct vmbus_channel *channel)
get_order(channel->ringbuffer_pagecount * PAGE_SIZE));
out:
- hv_event_tasklet_enable(channel);
-
return ret;
}
diff --git a/drivers/hv/channel_mgmt.c b/drivers/hv/channel_mgmt.c
index d2cfa3e..bf846d0 100644
--- a/drivers/hv/channel_mgmt.c
+++ b/drivers/hv/channel_mgmt.c
@@ -382,19 +382,6 @@ static void vmbus_release_relid(u32 relid)
true);
}
-void hv_event_tasklet_disable(struct vmbus_channel *channel)
-{
- tasklet_disable(&channel->callback_event);
-}
-
-void hv_event_tasklet_enable(struct vmbus_channel *channel)
-{
- tasklet_enable(&channel->callback_event);
-
- /* In case there is any pending event */
- tasklet_schedule(&channel->callback_event);
-}
-
void hv_process_channel_removal(struct vmbus_channel *channel, u32 relid)
{
unsigned long flags;
@@ -403,7 +390,6 @@ void hv_process_channel_removal(struct vmbus_channel *channel, u32 relid)
BUG_ON(!channel->rescind);
BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
- hv_event_tasklet_disable(channel);
if (channel->target_cpu != get_cpu()) {
put_cpu();
smp_call_function_single(channel->target_cpu,
@@ -412,7 +398,6 @@ void hv_process_channel_removal(struct vmbus_channel *channel, u32 relid)
percpu_channel_deq(channel);
put_cpu();
}
- hv_event_tasklet_enable(channel);
if (channel->primary_channel == NULL) {
list_del(&channel->listentry);
@@ -506,7 +491,6 @@ static void vmbus_process_offer(struct vmbus_channel *newchannel)
init_vp_index(newchannel, dev_type);
- hv_event_tasklet_disable(newchannel);
if (newchannel->target_cpu != get_cpu()) {
put_cpu();
smp_call_function_single(newchannel->target_cpu,
@@ -516,7 +500,6 @@ static void vmbus_process_offer(struct vmbus_channel *newchannel)
percpu_channel_enq(newchannel);
put_cpu();
}
- hv_event_tasklet_enable(newchannel);
/*
* This state is used to indicate a successful open
@@ -566,7 +549,6 @@ static void vmbus_process_offer(struct vmbus_channel *newchannel)
list_del(&newchannel->listentry);
mutex_unlock(&vmbus_connection.channel_mutex);
- hv_event_tasklet_disable(newchannel);
if (newchannel->target_cpu != get_cpu()) {
put_cpu();
smp_call_function_single(newchannel->target_cpu,
@@ -575,7 +557,6 @@ static void vmbus_process_offer(struct vmbus_channel *newchannel)
percpu_channel_deq(newchannel);
put_cpu();
}
- hv_event_tasklet_enable(newchannel);
vmbus_release_relid(newchannel->offermsg.child_relid);
diff --git a/include/linux/hyperv.h b/include/linux/hyperv.h
index c4c7ae9..970771a 100644
--- a/include/linux/hyperv.h
+++ b/include/linux/hyperv.h
@@ -1437,9 +1437,6 @@ extern bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp, u8 *buf,
const int *srv_version, int srv_vercnt,
int *nego_fw_version, int *nego_srv_version);
-void hv_event_tasklet_disable(struct vmbus_channel *channel);
-void hv_event_tasklet_enable(struct vmbus_channel *channel);
-
void hv_process_channel_removal(struct vmbus_channel *channel, u32 relid);
void vmbus_setevent(struct vmbus_channel *channel);
--
2.7.4
Dexuan Cui <[email protected]> writes:
> With the recent introduction of per-channel tasklet, we need to update
> the way we handle the 3 concurrency issues:
>
> 1. hv_process_channel_removal -> percpu_channel_deq vs.
> vmbus_chan_sched -> list_for_each_entry(..., percpu_list);
>
> 2. vmbus_process_offer -> percpu_channel_enq/deq vs. vmbus_chan_sched.
>
> 3. vmbus_close_internal vs. the per-channel tasklet vmbus_on_event;
>
> The first 2 issues can be handled by Stephen's recent patch
> "vmbus: use rcu for per-cpu channel list", and the third issue
> can be handled by calling tasklet_disable in vmbus_close_internal here.
>
> We don't need the original hv_event_tasklet_disable/enable since we
> now use per-channel tasklet instead of the previous per-CPU tasklet,
> and actually we must remove them due to the side effect now:
> vmbus_process_offer -> hv_event_tasklet_enable -> tasklet_schedule will
> start the per-channel callback prematurely, cauing NULL dereferencing
> (the channel may haven't been properly configured to run the callback yet).
>
> Fixes: 631e63a9f346 ("vmbus: change to per channel tasklet")
>
> Signed-off-by: Dexuan Cui <[email protected]>
> Cc: "K. Y. Srinivasan" <[email protected]>
> Cc: Haiyang Zhang <[email protected]>
> Cc: Stephen Hemminger <[email protected]>
Tested-by: Vitaly Kuznetsov <[email protected]>
This patch fixes the following crash on boot:
[ 1.451648] BUG: unable to handle kernel NULL pointer dereference at 0000000000000004
[ 1.452255] IP: netvsc_channel_cb+0x90/0x7b0 [hv_netvsc]
[ 1.452255] PGD 0
[ 1.452255]
[ 1.452255] Oops: 0000 [#1] SMP
[ 1.452255] Modules linked in: hv_storvsc hv_netvsc(+) scsi_transport_fc hyperv_fb hyperv_keyboard hid_hyperv hv_vmbus
[ 1.452255] CPU: 1 PID: 19 Comm: ksoftirqd/1 Not tainted 4.10.0_test+ #911
[ 1.452255] Hardware name: Microsoft Corporation Virtual Machine/Virtual Machine, BIOS Hyper-V UEFI Release v1.0 11/26/2012
[ 1.452255] task: ffff880007fd2b00 task.stack: ffffc90000e34000
[ 1.452255] RIP: 0010:netvsc_channel_cb+0x90/0x7b0 [hv_netvsc]
...
[ 1.452255] Call Trace:
[ 1.452255] vmbus_on_event+0x22/0x90 [hv_vmbus]
[ 1.452255] tasklet_action+0x5e/0x110
[ 1.452255] __do_softirq+0x104/0x2af
[ 1.452255] run_ksoftirqd+0x29/0x40
...
[ 1.548068] RIP: netvsc_channel_cb+0x90/0x7b0 [hv_netvsc] RSP: ffffc90000e37d88
[ 1.548068] CR2: 0000000000000004
[ 1.548068] ---[ end trace 601fd9d6588b21e5 ]---
[ 1.548068] Kernel panic - not syncing: Fatal exception in interrupt
[ 1.548068] Kernel Offset: disabled
[ 1.548068] ---[ end Kernel panic - not syncing: Fatal exception in interrupt
[ 1.572155] ------------[ cut here ]------------
The crash is not imminent but it happens pretty often on boot, I think
we need to push it to 4.11.
--
Vitaly
On Fri, 03 Mar 2017 17:40:47 +0100
Vitaly Kuznetsov <[email protected]> wrote:
> Dexuan Cui <[email protected]> writes:
>
> > With the recent introduction of per-channel tasklet, we need to update
> > the way we handle the 3 concurrency issues:
> >
> > 1. hv_process_channel_removal -> percpu_channel_deq vs.
> > vmbus_chan_sched -> list_for_each_entry(..., percpu_list);
> >
> > 2. vmbus_process_offer -> percpu_channel_enq/deq vs. vmbus_chan_sched.
> >
> > 3. vmbus_close_internal vs. the per-channel tasklet vmbus_on_event;
> >
> > The first 2 issues can be handled by Stephen's recent patch
> > "vmbus: use rcu for per-cpu channel list", and the third issue
> > can be handled by calling tasklet_disable in vmbus_close_internal here.
> >
> > We don't need the original hv_event_tasklet_disable/enable since we
> > now use per-channel tasklet instead of the previous per-CPU tasklet,
> > and actually we must remove them due to the side effect now:
> > vmbus_process_offer -> hv_event_tasklet_enable -> tasklet_schedule will
> > start the per-channel callback prematurely, cauing NULL dereferencing
> > (the channel may haven't been properly configured to run the callback yet).
> >
> > Fixes: 631e63a9f346 ("vmbus: change to per channel tasklet")
> >
> > Signed-off-by: Dexuan Cui <[email protected]>
> > Cc: "K. Y. Srinivasan" <[email protected]>
> > Cc: Haiyang Zhang <[email protected]>
> > Cc: Stephen Hemminger <[email protected]>
>
> Tested-by: Vitaly Kuznetsov <[email protected]>
>
> This patch fixes the following crash on boot:
>
> [ 1.451648] BUG: unable to handle kernel NULL pointer dereference at 0000000000000004
> [ 1.452255] IP: netvsc_channel_cb+0x90/0x7b0 [hv_netvsc]
> [ 1.452255] PGD 0
> [ 1.452255]
> [ 1.452255] Oops: 0000 [#1] SMP
> [ 1.452255] Modules linked in: hv_storvsc hv_netvsc(+) scsi_transport_fc hyperv_fb hyperv_keyboard hid_hyperv hv_vmbus
> [ 1.452255] CPU: 1 PID: 19 Comm: ksoftirqd/1 Not tainted 4.10.0_test+ #911
> [ 1.452255] Hardware name: Microsoft Corporation Virtual Machine/Virtual Machine, BIOS Hyper-V UEFI Release v1.0 11/26/2012
> [ 1.452255] task: ffff880007fd2b00 task.stack: ffffc90000e34000
> [ 1.452255] RIP: 0010:netvsc_channel_cb+0x90/0x7b0 [hv_netvsc]
> ...
> [ 1.452255] Call Trace:
> [ 1.452255] vmbus_on_event+0x22/0x90 [hv_vmbus]
> [ 1.452255] tasklet_action+0x5e/0x110
> [ 1.452255] __do_softirq+0x104/0x2af
> [ 1.452255] run_ksoftirqd+0x29/0x40
> ...
> [ 1.548068] RIP: netvsc_channel_cb+0x90/0x7b0 [hv_netvsc] RSP: ffffc90000e37d88
> [ 1.548068] CR2: 0000000000000004
> [ 1.548068] ---[ end trace 601fd9d6588b21e5 ]---
> [ 1.548068] Kernel panic - not syncing: Fatal exception in interrupt
> [ 1.548068] Kernel Offset: disabled
> [ 1.548068] ---[ end Kernel panic - not syncing: Fatal exception in interrupt
> [ 1.572155] ------------[ cut here ]------------
>
> The crash is not imminent but it happens pretty often on boot, I think
> we need to push it to 4.11.
>
Agree that this needs to be in 4.11, but when NAPI is merged it also will not happen.
On Thu, 2 Mar 2017 12:32:55 +0000
Dexuan Cui <[email protected]> wrote:
> With the recent introduction of per-channel tasklet, we need to update
> the way we handle the 3 concurrency issues:
>
> 1. hv_process_channel_removal -> percpu_channel_deq vs.
> vmbus_chan_sched -> list_for_each_entry(..., percpu_list);
>
> 2. vmbus_process_offer -> percpu_channel_enq/deq vs. vmbus_chan_sched.
>
> 3. vmbus_close_internal vs. the per-channel tasklet vmbus_on_event;
>
> The first 2 issues can be handled by Stephen's recent patch
> "vmbus: use rcu for per-cpu channel list", and the third issue
> can be handled by calling tasklet_disable in vmbus_close_internal here.
>
> We don't need the original hv_event_tasklet_disable/enable since we
> now use per-channel tasklet instead of the previous per-CPU tasklet,
> and actually we must remove them due to the side effect now:
> vmbus_process_offer -> hv_event_tasklet_enable -> tasklet_schedule will
> start the per-channel callback prematurely, cauing NULL dereferencing
> (the channel may haven't been properly configured to run the callback yet).
>
> Fixes: 631e63a9f346 ("vmbus: change to per channel tasklet")
>
> Signed-off-by: Dexuan Cui <[email protected]>
> Cc: "K. Y. Srinivasan" <[email protected]>
> Cc: Haiyang Zhang <[email protected]>
> Cc: Stephen Hemminger <[email protected]>
It also fixes a number of missed interrupts causing stuck tasks on boot.
Please put it in 4.11
Reviewed-by: Stephen Hemminger <[email protected]>