Re: [PATCH] Drivers: hv: vmbus: fix the race when querying & updating the percpu list

[Vitaly Kuznetsov]

Dexuan Cui <decui@xxxxxxxxxxxxx> writes:

> There is a rare race when we remove an entry from the global list
> hv_context.percpu_list[cpu] in hv_process_channel_removal() ->
> percpu_channel_deq() -> list_del(): at this time, if vmbus_on_event() ->
> process_chn_event() -> pcpu_relid2channel() is trying to query the list,
> we can get the general protection fault:
>
> general protection fault: 0000 [#1] SMP
> ...
> RIP: 0010:[<ffffffff81461b6b>]  [<ffffffff81461b6b>] vmbus_on_event+0xc4/0x149
>
> Similarly, we also have the issue in the code path: vmbus_process_offer() ->
> percpu_channel_enq().
>
> We can resolve the issue by disabling the tasklet when updating the list.
>
> Reported-by: Rolf Neugebauer <rolf.neugebauer@xxxxxxxxxx>
> Signed-off-by: Dexuan Cui <decui@xxxxxxxxxxxxx>
> ---
>  drivers/hv/channel.c      |  3 +++
>  drivers/hv/channel_mgmt.c | 20 +++++++++-----------
>  include/linux/hyperv.h    |  3 +++
>  3 files changed, 15 insertions(+), 11 deletions(-)
>
> diff --git a/drivers/hv/channel.c b/drivers/hv/channel.c
> index 56dd261..7811cf9 100644
> --- a/drivers/hv/channel.c
> +++ b/drivers/hv/channel.c
> @@ -546,8 +546,11 @@ static int vmbus_close_internal(struct vmbus_channel *channel)
>  		put_cpu();
>  		smp_call_function_single(channel->target_cpu, reset_channel_cb,
>  					 channel, true);
> +		smp_call_function_single(channel->target_cpu,
> +					 percpu_channel_deq, channel, true);
>  	} else {
>  		reset_channel_cb(channel);
> +		percpu_channel_deq(channel);
>  		put_cpu();
>  	}
>
> diff --git a/drivers/hv/channel_mgmt.c b/drivers/hv/channel_mgmt.c
> index 38b682ba..c4b5c07 100644
> --- a/drivers/hv/channel_mgmt.c
> +++ b/drivers/hv/channel_mgmt.c
> @@ -21,6 +21,7 @@
>  #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
>
>  #include <linux/kernel.h>
> +#include <linux/interrupt.h>
>  #include <linux/sched.h>
>  #include <linux/wait.h>
>  #include <linux/mm.h>
> @@ -277,7 +278,7 @@ static void free_channel(struct vmbus_channel *channel)
>  	kfree(channel);
>  }
>
> -static void percpu_channel_enq(void *arg)
> +void percpu_channel_enq(void *arg)
>  {
>  	struct vmbus_channel *channel = arg;
>  	int cpu = smp_processor_id();
> @@ -285,7 +286,7 @@ static void percpu_channel_enq(void *arg)
>  	list_add_tail(&channel->percpu_list, &hv_context.percpu_list[cpu]);
>  }
>
> -static void percpu_channel_deq(void *arg)
> +void percpu_channel_deq(void *arg)
>  {
>  	struct vmbus_channel *channel = arg;
>
> @@ -313,15 +314,6 @@ void hv_process_channel_removal(struct vmbus_channel *channel, u32 relid)
>  	BUG_ON(!channel->rescind);
>  	BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
>
> -	if (channel->target_cpu != get_cpu()) {
> -		put_cpu();
> -		smp_call_function_single(channel->target_cpu,
> -					 percpu_channel_deq, channel, true);
> -	} else {
> -		percpu_channel_deq(channel);
> -		put_cpu();
> -	}
> -
>  	if (channel->primary_channel == NULL) {
>  		list_del(&channel->listentry);
>
> @@ -363,6 +355,7 @@ void vmbus_free_channels(void)
>   */
>  static void vmbus_process_offer(struct vmbus_channel *newchannel)
>  {
> +	struct tasklet_struct *tasklet;
>  	struct vmbus_channel *channel;
>  	bool fnew = true;
>  	unsigned long flags;
> @@ -409,6 +402,8 @@ static void vmbus_process_offer(struct vmbus_channel *newchannel)
>
>  	init_vp_index(newchannel, dev_type);
>
> +	tasklet = hv_context.event_dpc[newchannel->target_cpu];
> +	tasklet_disable(tasklet);
>  	if (newchannel->target_cpu != get_cpu()) {
>  		put_cpu();
>  		smp_call_function_single(newchannel->target_cpu,
> @@ -418,6 +413,7 @@ static void vmbus_process_offer(struct vmbus_channel *newchannel)
>  		percpu_channel_enq(newchannel);
>  		put_cpu();
>  	}
> +	tasklet_enable(tasklet);

Do we need to do tasklet_schedule() to make sure there are no events
pending? This is probably not a big issue as some other event will
trigger scheduling but in some corner cases it may bite. Same question
applies to the code below and to vmbus_close_internal().

>
>  	/*
>  	 * This state is used to indicate a successful open
> @@ -469,6 +465,7 @@ static void vmbus_process_offer(struct vmbus_channel *newchannel)
>  	list_del(&newchannel->listentry);
>  	mutex_unlock(&vmbus_connection.channel_mutex);
>
> +	tasklet_disable(tasklet);
>  	if (newchannel->target_cpu != get_cpu()) {
>  		put_cpu();
>  		smp_call_function_single(newchannel->target_cpu,
> @@ -477,6 +474,7 @@ static void vmbus_process_offer(struct vmbus_channel *newchannel)
>  		percpu_channel_deq(newchannel);
>  		put_cpu();
>  	}
> +	tasklet_enable(tasklet);
>
>  err_free_chan:
>  	free_channel(newchannel);
> diff --git a/include/linux/hyperv.h b/include/linux/hyperv.h
> index 7be7237..95aea09 100644
> --- a/include/linux/hyperv.h
> +++ b/include/linux/hyperv.h
> @@ -1328,6 +1328,9 @@ extern bool vmbus_prep_negotiate_resp(struct icmsg_hdr *,
>  					struct icmsg_negotiate *, u8 *, int,
>  					int);
>
> +void percpu_channel_enq(void *arg);
> +void percpu_channel_deq(void *arg);
> +
>  void hv_process_channel_removal(struct vmbus_channel *channel, u32 relid);
>
>  /*

-- 
  Vitaly