Message-ID: <Z4GOWBkqX-vx_0sO@boqun-archlinux>
Date: Fri, 10 Jan 2025 13:17:12 -0800
From: Boqun Feng <boqun.feng@...il.com>
To: Hamza Mahfooz <hamzamahfooz@...ux.microsoft.com>
Cc: linux-hyperv@...r.kernel.org, Wei Liu <wei.liu@...nel.org>,
	"K. Y. Srinivasan" <kys@...rosoft.com>,
	Haiyang Zhang <haiyangz@...rosoft.com>,
	Dexuan Cui <decui@...rosoft.com>, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] drivers/hv: add CPU offlining support

Hi Hamza,

Thanks for the patch, a few comments below:

On Fri, Jan 10, 2025 at 03:05:06PM -0500, Hamza Mahfooz wrote:
> Currently, it is effectively impossible to offline CPUs. Since, most
> CPUs will have vmbus channels attached to them. So, as made mention of
> in commit d570aec0f2154 ("Drivers: hv: vmbus: Synchronize
> init_vp_index() vs. CPU hotplug"), rebind channels associated with CPUs
> that a user is trying to offline to a new "randomly" selected CPU.
> 
> Cc: Boqun Feng <boqun.feng@...il.com>
> Cc: Wei Liu <wei.liu@...nel.org>
> Signed-off-by: Hamza Mahfooz <hamzamahfooz@...ux.microsoft.com>
> ---
>  drivers/hv/hv.c        | 57 +++++++++++++++++++++++++++++++-----------
>  drivers/hv/vmbus_drv.c | 51 +++++++++++++++++++++----------------
>  include/linux/hyperv.h |  1 +
>  3 files changed, 73 insertions(+), 36 deletions(-)
> 
> diff --git a/drivers/hv/hv.c b/drivers/hv/hv.c
> index 36d9ba097ff5..42270a7a7a19 100644
> --- a/drivers/hv/hv.c
> +++ b/drivers/hv/hv.c
> @@ -433,13 +433,40 @@ static bool hv_synic_event_pending(void)
>  	return pending;
>  }
>  
> +static int hv_pick_new_cpu(struct vmbus_channel *channel,
> +			   unsigned int current_cpu)
> +{
> +	int ret = 0;
> +	int cpu;
> +
> +	lockdep_assert_held(&vmbus_connection.channel_mutex);
> +
> +	/*
> +	 * We can't assume that the relevant interrupts will be sent before
> +	 * the cpu is offlined on older versions of hyperv.
> +	 */
> +	if (vmbus_proto_version < VERSION_WIN10_V5_3)
> +		return -EBUSY;
> +
> +	cpus_read_lock();

hv_pick_new_cpu() is only called inside hv_synic_cleanup(), which is
only called with cpus_read_lock() held (because it's registered via
cpuhp_setup_state_nocalls_cpuslocked()). So the cpus_read_lock() is not
necessary here if I'm not missing anything. Moreover, given
cpus_read_lock() is a non-recursive read lock, it's actually incorrect
to re-acquire it here, see:

	https://docs.kernel.org/locking/lockdep-design.html#recursive-read-locks

for more information.

> +	cpu = cpumask_next(get_random_u32_below(nr_cpu_ids), cpu_online_mask);
> +
> +	if (cpu >= nr_cpu_ids || cpu == current_cpu)
> +		cpu = VMBUS_CONNECT_CPU;
> +
> +	ret = vmbus_channel_set_cpu(channel, cpu);
> +	cpus_read_unlock();
> +
> +	return ret;
> +}
> +
>  /*
>   * hv_synic_cleanup - Cleanup routine for hv_synic_init().
>   */
>  int hv_synic_cleanup(unsigned int cpu)
>  {
>  	struct vmbus_channel *channel, *sc;
> -	bool channel_found = false;
> +	int ret = 0;
>  
>  	if (vmbus_connection.conn_state != CONNECTED)
>  		goto always_cleanup;
> @@ -456,31 +483,31 @@ int hv_synic_cleanup(unsigned int cpu)
>  
>  	/*
>  	 * Search for channels which are bound to the CPU we're about to
> -	 * cleanup.  In case we find one and vmbus is still connected, we
> -	 * fail; this will effectively prevent CPU offlining.
> -	 *
> -	 * TODO: Re-bind the channels to different CPUs.
> +	 * cleanup.
>  	 */
>  	mutex_lock(&vmbus_connection.channel_mutex);
>  	list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
>  		if (channel->target_cpu == cpu) {
> -			channel_found = true;
> -			break;
> +			ret = hv_pick_new_cpu(channel, cpu);
> +
> +			if (ret) {
> +				mutex_unlock(&vmbus_connection.channel_mutex);
> +				return ret;
> +			}
>  		}
>  		list_for_each_entry(sc, &channel->sc_list, sc_list) {
>  			if (sc->target_cpu == cpu) {
> -				channel_found = true;
> -				break;
> +				ret = hv_pick_new_cpu(channel, cpu);
> +
> +				if (ret) {
> +					mutex_unlock(&vmbus_connection.channel_mutex);
> +					return ret;
> +				}
>  			}
>  		}
> -		if (channel_found)
> -			break;
>  	}
>  	mutex_unlock(&vmbus_connection.channel_mutex);
>  
> -	if (channel_found)
> -		return -EBUSY;
> -
>  	/*
>  	 * channel_found == false means that any channels that were previously
>  	 * assigned to the CPU have been reassigned elsewhere with a call of
> @@ -497,5 +524,5 @@ int hv_synic_cleanup(unsigned int cpu)
>  
>  	hv_synic_disable_regs(cpu);
>  
> -	return 0;
> +	return ret;
>  }
> diff --git a/drivers/hv/vmbus_drv.c b/drivers/hv/vmbus_drv.c
> index 2892b8da20a5..c256e02fa66b 100644
> --- a/drivers/hv/vmbus_drv.c
> +++ b/drivers/hv/vmbus_drv.c
> @@ -1611,16 +1611,15 @@ static ssize_t target_cpu_show(struct vmbus_channel *channel, char *buf)
>  {
>  	return sprintf(buf, "%u\n", channel->target_cpu);
>  }
> -static ssize_t target_cpu_store(struct vmbus_channel *channel,

It'll be nice if you could separate this change into a different patch,
that is one patch for refactoring target_cpu_store() with
vmbus_channel_set_cpu() and the other patch to introduce the support of
cpu offline. Make it easier for reviewing.

> -				const char *buf, size_t count)
> +
> +int vmbus_channel_set_cpu(struct vmbus_channel *channel, u32 target_cpu)
>  {
> -	u32 target_cpu, origin_cpu;
> -	ssize_t ret = count;
> +	u32 origin_cpu;
> +	int ret = 0;
>  
> -	if (vmbus_proto_version < VERSION_WIN10_V4_1)
> -		return -EIO;
> +	lockdep_assert_held(&vmbus_connection.channel_mutex);

Add

	lockdep_assert_cpus_held();

as well?

Regards,
Boqun

>  
> -	if (sscanf(buf, "%uu", &target_cpu) != 1)
> +	if (vmbus_proto_version < VERSION_WIN10_V4_1)
>  		return -EIO;
>  
>  	/* Validate target_cpu for the cpumask_test_cpu() operation below. */
> @@ -1630,22 +1629,17 @@ static ssize_t target_cpu_store(struct vmbus_channel *channel,
>  	if (!cpumask_test_cpu(target_cpu, housekeeping_cpumask(HK_TYPE_MANAGED_IRQ)))
>  		return -EINVAL;
>  
> -	/* No CPUs should come up or down during this. */
> -	cpus_read_lock();
> -
> -	if (!cpu_online(target_cpu)) {
> -		cpus_read_unlock();
> +	if (!cpu_online(target_cpu))
>  		return -EINVAL;
> -	}
>  
>  	/*
> -	 * Synchronizes target_cpu_store() and channel closure:
> +	 * Synchronizes vmbus_channel_set_cpu() and channel closure:
>  	 *
>  	 * { Initially: state = CHANNEL_OPENED }
>  	 *
>  	 * CPU1				CPU2
>  	 *
> -	 * [target_cpu_store()]		[vmbus_disconnect_ring()]
> +	 * [vmbus_channel_set_cpu()]	[vmbus_disconnect_ring()]
>  	 *
>  	 * LOCK channel_mutex		LOCK channel_mutex
>  	 * LOAD r1 = state		LOAD r2 = state
> @@ -1660,7 +1654,6 @@ static ssize_t target_cpu_store(struct vmbus_channel *channel,
>  	 * Note.  The host processes the channel messages "sequentially", in
>  	 * the order in which they are received on a per-partition basis.
>  	 */
> -	mutex_lock(&vmbus_connection.channel_mutex);
>  
>  	/*
>  	 * Hyper-V will ignore MODIFYCHANNEL messages for "non-open" channels;
> @@ -1668,17 +1661,17 @@ static ssize_t target_cpu_store(struct vmbus_channel *channel,
>  	 */
>  	if (channel->state != CHANNEL_OPENED_STATE) {
>  		ret = -EIO;
> -		goto cpu_store_unlock;
> +		goto end;
>  	}
>  
>  	origin_cpu = channel->target_cpu;
>  	if (target_cpu == origin_cpu)
> -		goto cpu_store_unlock;
> +		goto end;
>  
>  	if (vmbus_send_modifychannel(channel,
>  				     hv_cpu_number_to_vp_number(target_cpu))) {
>  		ret = -EIO;
> -		goto cpu_store_unlock;
> +		goto end;
>  	}
>  
>  	/*
> @@ -1708,9 +1701,25 @@ static ssize_t target_cpu_store(struct vmbus_channel *channel,
>  				origin_cpu, target_cpu);
>  	}
>  
> -cpu_store_unlock:
> -	mutex_unlock(&vmbus_connection.channel_mutex);
> +end:
> +	return ret;
> +}
> +
> +static ssize_t target_cpu_store(struct vmbus_channel *channel,
> +				const char *buf, size_t count)
> +{
> +	ssize_t ret = count;
> +	u32 target_cpu;
> +
> +	if (sscanf(buf, "%uu", &target_cpu) != 1)
> +		return -EIO;
> +
> +	mutex_lock(&vmbus_connection.channel_mutex);
> +	cpus_read_lock();
> +	ret = vmbus_channel_set_cpu(channel, target_cpu);
>  	cpus_read_unlock();
> +	mutex_unlock(&vmbus_connection.channel_mutex);
> +
>  	return ret;
>  }
>  static VMBUS_CHAN_ATTR(cpu, 0644, target_cpu_show, target_cpu_store);
> diff --git a/include/linux/hyperv.h b/include/linux/hyperv.h
> index 02a226bcf0ed..25e9e982f1b0 100644
> --- a/include/linux/hyperv.h
> +++ b/include/linux/hyperv.h
> @@ -1670,6 +1670,7 @@ int vmbus_send_tl_connect_request(const guid_t *shv_guest_servie_id,
>  				  const guid_t *shv_host_servie_id);
>  int vmbus_send_modifychannel(struct vmbus_channel *channel, u32 target_vp);
>  void vmbus_set_event(struct vmbus_channel *channel);
> +int vmbus_channel_set_cpu(struct vmbus_channel *channel, u32 target_cpu);
>  
>  /* Get the start of the ring buffer. */
>  static inline void *
> -- 
> 2.47.1
> 

Powered by blists - more mailing lists