Linux-HyperV Archive on lore.kernel.org
 help / color / Atom feed
From: Michael Kelley <mikelley@microsoft.com>
To: "Andrea Parri (Microsoft)" <parri.andrea@gmail.com>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Cc: KY Srinivasan <kys@microsoft.com>,
	Haiyang Zhang <haiyangz@microsoft.com>,
	Stephen Hemminger <sthemmin@microsoft.com>,
	Wei Liu <wei.liu@kernel.org>,
	"linux-hyperv@vger.kernel.org" <linux-hyperv@vger.kernel.org>,
	Dexuan Cui <decui@microsoft.com>,
	Boqun Feng <boqun.feng@gmail.com>, vkuznets <vkuznets@redhat.com>,
	"James E.J. Bottomley" <jejb@linux.ibm.com>,
	"Martin K. Petersen" <martin.petersen@oracle.com>,
	"linux-scsi@vger.kernel.org" <linux-scsi@vger.kernel.org>
Subject: RE: [RFC PATCH 11/11] scsi: storvsc: Re-init stor_chns when a channel interrupt is re-assigned
Date: Mon, 30 Mar 2020 16:42:19 +0000
Message-ID: <MW2PR2101MB105208138683A6DE0564745AD7CB0@MW2PR2101MB1052.namprd21.prod.outlook.com> (raw)
In-Reply-To: <20200325225505.23998-12-parri.andrea@gmail.com>

From: Andrea Parri (Microsoft) <parri.andrea@gmail.com>  Sent: Wednesday, March 25, 2020 3:55 PM
> 
> For each storvsc_device, storvsc keeps track of the channel target CPUs
> associated to the device (alloced_cpus) and it uses this information to
> fill a "cache" (stor_chns) mapping CPU->channel according to a certain
> heuristic.  Update the alloced_cpus mask and the stor_chns array when a
> channel of the storvsc device is re-assigned to a different CPU.
> 
> Signed-off-by: Andrea Parri (Microsoft) <parri.andrea@gmail.com>
> Cc: "James E.J. Bottomley" <jejb@linux.ibm.com>
> Cc: "Martin K. Petersen" <martin.petersen@oracle.com>
> Cc: <linux-scsi@vger.kernel.org>
> ---
>  drivers/hv/vmbus_drv.c     |  4 ++
>  drivers/scsi/storvsc_drv.c | 95 ++++++++++++++++++++++++++++++++++----
>  include/linux/hyperv.h     |  3 ++
>  3 files changed, 94 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/hv/vmbus_drv.c b/drivers/hv/vmbus_drv.c
> index 84d2f22c569aa..7199fee2b5869 100644
> --- a/drivers/hv/vmbus_drv.c
> +++ b/drivers/hv/vmbus_drv.c
> @@ -1721,6 +1721,10 @@ static ssize_t target_cpu_store(struct vmbus_channel *channel,
>  	 * in on a CPU that is different from the channel target_cpu value.
>  	 */
> 
> +	if (channel->change_target_cpu_callback)
> +		(*channel->change_target_cpu_callback)(channel,
> +				channel->target_cpu, target_cpu);
> +
>  	channel->target_cpu = target_cpu;
>  	channel->target_vp = hv_cpu_number_to_vp_number(target_cpu);
>  	channel->numa_node = cpu_to_node(target_cpu);

I think there's an ordering problem here.  The change_target_cpu_callback
will allow storvsc to flush the cache that it is keeping, but there's a window
after the storvsc callback releases the spin lock and before this function
changes channel->target_cpu to the new value.  In that window, the cache
could get refilled based on the old value of channel->target_cpu, which is
exactly what we don't want.  Generally with caches, you have to set the new
value first, then flush the cache, and I think that works in this case.  The
callback function doesn't depend on the value of channel->target_cpu,
and any cache filling that might happen after channel->target_cpu is set
to the new value but before the callback function runs is OK.   But please
double-check my thinking. :-)


> diff --git a/drivers/scsi/storvsc_drv.c b/drivers/scsi/storvsc_drv.c
> index fb41636519ee8..a680592b9d32a 100644
> --- a/drivers/scsi/storvsc_drv.c
> +++ b/drivers/scsi/storvsc_drv.c
> @@ -621,6 +621,63 @@ static inline struct storvsc_device *get_in_stor_device(
> 
>  }
> 
> +void storvsc_change_target_cpu(struct vmbus_channel *channel, u32 old, u32 new)
> +{
> +	struct storvsc_device *stor_device;
> +	struct vmbus_channel *cur_chn;
> +	bool old_is_alloced = false;
> +	struct hv_device *device;
> +	unsigned long flags;
> +	int cpu;
> +
> +	device = channel->primary_channel ?
> +			channel->primary_channel->device_obj
> +				: channel->device_obj;
> +	stor_device = get_out_stor_device(device);
> +	if (!stor_device)
> +		return;
> +
> +	/* See storvsc_do_io() -> get_og_chn(). */
> +	spin_lock_irqsave(&device->channel->lock, flags);
> +
> +	/*
> +	 * Determines if the storvsc device has other channels assigned to
> +	 * the "old" CPU to update the alloced_cpus mask and the stor_chns
> +	 * array.
> +	 */
> +	if (device->channel != channel && device->channel->target_cpu == old) {
> +		cur_chn = device->channel;
> +		old_is_alloced = true;
> +		goto old_is_alloced;
> +	}
> +	list_for_each_entry(cur_chn, &device->channel->sc_list, sc_list) {
> +		if (cur_chn == channel)
> +			continue;
> +		if (cur_chn->target_cpu == old) {
> +			old_is_alloced = true;
> +			goto old_is_alloced;
> +		}
> +	}
> +
> +old_is_alloced:
> +	if (old_is_alloced)
> +		WRITE_ONCE(stor_device->stor_chns[old], cur_chn);
> +	else
> +		cpumask_clear_cpu(old, &stor_device->alloced_cpus);

I think target_cpu_store() can get called in parallel on multiple CPUs for different
channels on the same storvsc device, but multiple changes to a single channel are
serialized by higher levels of sysfs.  So this function could run after multiple
channels have been changed, in which case there's not just a single "old" value,
and the above algorithm might not work, especially if channel->target_cpu is
updated before calling this function per my earlier comment.   I can see a
couple of possible ways to deal with this.  One is to put the update of
channel->target_cpu in this function, within the spin lock boundaries so
that the cache flush and target_cpu update are atomic.  Another idea is to
process multiple changes in this function, by building a temp copy of
alloced_cpus by walking the channel list, use XOR to create a cpumask
with changes, and then process all the changes in a loop instead of
just handling a single change as with the current code at the old_is_alloced
label.  But I haven't completely thought through this idea.

> +
> +	/* "Flush" the stor_chns array. */
> +	for_each_possible_cpu(cpu) {
> +		if (stor_device->stor_chns[cpu] && !cpumask_test_cpu(
> +					cpu, &stor_device->alloced_cpus))
> +			WRITE_ONCE(stor_device->stor_chns[cpu], NULL);
> +	}
> +
> +	WRITE_ONCE(stor_device->stor_chns[new], channel);
> +	cpumask_set_cpu(new, &stor_device->alloced_cpus);
> +
> +	spin_unlock_irqrestore(&device->channel->lock, flags);
> +}
> +
>  static void handle_sc_creation(struct vmbus_channel *new_sc)
>  {
>  	struct hv_device *device = new_sc->primary_channel->device_obj;
> @@ -648,6 +705,8 @@ static void handle_sc_creation(struct vmbus_channel *new_sc)
>  		return;
>  	}
> 
> +	new_sc->change_target_cpu_callback = storvsc_change_target_cpu;
> +
>  	/* Add the sub-channel to the array of available channels. */
>  	stor_device->stor_chns[new_sc->target_cpu] = new_sc;
>  	cpumask_set_cpu(new_sc->target_cpu, &stor_device->alloced_cpus);
> @@ -876,6 +935,8 @@ static int storvsc_channel_init(struct hv_device *device, bool is_fc)
>  	if (stor_device->stor_chns == NULL)
>  		return -ENOMEM;
> 
> +	device->channel->change_target_cpu_callback = storvsc_change_target_cpu;
> +
>  	stor_device->stor_chns[device->channel->target_cpu] = device->channel;
>  	cpumask_set_cpu(device->channel->target_cpu,
>  			&stor_device->alloced_cpus);
> @@ -1248,8 +1309,10 @@ static struct vmbus_channel *get_og_chn(struct storvsc_device
> *stor_device,
>  	const struct cpumask *node_mask;
>  	int num_channels, tgt_cpu;
> 
> -	if (stor_device->num_sc == 0)
> +	if (stor_device->num_sc == 0) {
> +		stor_device->stor_chns[q_num] = stor_device->device->channel;
>  		return stor_device->device->channel;
> +	}
> 
>  	/*
>  	 * Our channel array is sparsley populated and we
> @@ -1258,7 +1321,6 @@ static struct vmbus_channel *get_og_chn(struct storvsc_device
> *stor_device,
>  	 * The strategy is simple:
>  	 * I. Ensure NUMA locality
>  	 * II. Distribute evenly (best effort)
> -	 * III. Mapping is persistent.
>  	 */
> 
>  	node_mask = cpumask_of_node(cpu_to_node(q_num));
> @@ -1268,8 +1330,10 @@ static struct vmbus_channel *get_og_chn(struct storvsc_device
> *stor_device,
>  		if (cpumask_test_cpu(tgt_cpu, node_mask))
>  			num_channels++;
>  	}
> -	if (num_channels == 0)
> +	if (num_channels == 0) {
> +		stor_device->stor_chns[q_num] = stor_device->device->channel;

Is the above added line just fixing a bug in the existing code?  I'm not seeing how
it would derive from the other changes in this patch.

>  		return stor_device->device->channel;
> +	}
> 
>  	hash_qnum = q_num;
>  	while (hash_qnum >= num_channels)
> @@ -1295,6 +1359,7 @@ static int storvsc_do_io(struct hv_device *device,
>  	struct storvsc_device *stor_device;
>  	struct vstor_packet *vstor_packet;
>  	struct vmbus_channel *outgoing_channel, *channel;
> +	unsigned long flags;
>  	int ret = 0;
>  	const struct cpumask *node_mask;
>  	int tgt_cpu;
> @@ -1308,10 +1373,11 @@ static int storvsc_do_io(struct hv_device *device,
> 
>  	request->device  = device;
>  	/*
> -	 * Select an an appropriate channel to send the request out.
> +	 * Select an appropriate channel to send the request out.
>  	 */
> -	if (stor_device->stor_chns[q_num] != NULL) {
> -		outgoing_channel = stor_device->stor_chns[q_num];
> +	/* See storvsc_change_target_cpu(). */
> +	outgoing_channel = READ_ONCE(stor_device->stor_chns[q_num]);
> +	if (outgoing_channel != NULL) {
>  		if (outgoing_channel->target_cpu == q_num) {
>  			/*
>  			 * Ideally, we want to pick a different channel if
> @@ -1324,7 +1390,10 @@ static int storvsc_do_io(struct hv_device *device,
>  					continue;
>  				if (tgt_cpu == q_num)
>  					continue;
> -				channel = stor_device->stor_chns[tgt_cpu];
> +				channel = READ_ONCE(
> +					stor_device->stor_chns[tgt_cpu]);
> +				if (channel == NULL)
> +					continue;

The channel == NULL case is new because a cache flush could be happening
in parallel on another CPU.  I'm wondering about the tradeoffs of
continuing in the loop (as you have coded in this patch) vs. a "goto" back to
the top level "if" statement.   With the "continue" you might finish the
loop without finding any matches, and fall through to the next approach.
But it's only a single I/O operation, and if it comes up with a less than
optimal channel choice, it's no big deal.  So I guess it's really a wash.

>  				if (hv_get_avail_to_write_percent(
>  							&channel->outbound)
>  						> ring_avail_percent_lowater) {
> @@ -1350,7 +1419,10 @@ static int storvsc_do_io(struct hv_device *device,
>  			for_each_cpu(tgt_cpu, &stor_device->alloced_cpus) {
>  				if (cpumask_test_cpu(tgt_cpu, node_mask))
>  					continue;
> -				channel = stor_device->stor_chns[tgt_cpu];
> +				channel = READ_ONCE(
> +					stor_device->stor_chns[tgt_cpu]);
> +				if (channel == NULL)
> +					continue;

Same comment here.

>  				if (hv_get_avail_to_write_percent(
>  							&channel->outbound)
>  						> ring_avail_percent_lowater) {
> @@ -1360,7 +1432,14 @@ static int storvsc_do_io(struct hv_device *device,
>  			}
>  		}
>  	} else {
> +		spin_lock_irqsave(&device->channel->lock, flags);
> +		outgoing_channel = stor_device->stor_chns[q_num];
> +		if (outgoing_channel != NULL) {
> +			spin_unlock_irqrestore(&device->channel->lock, flags);
> +			goto found_channel;
> +		}
>  		outgoing_channel = get_og_chn(stor_device, q_num);
> +		spin_unlock_irqrestore(&device->channel->lock, flags);
>  	}
> 
>  found_channel:
> diff --git a/include/linux/hyperv.h b/include/linux/hyperv.h
> index edfcd42319ef3..9018b89614b78 100644
> --- a/include/linux/hyperv.h
> +++ b/include/linux/hyperv.h
> @@ -773,6 +773,9 @@ struct vmbus_channel {
>  	void (*onchannel_callback)(void *context);
>  	void *channel_callback_context;
> 
> +	void (*change_target_cpu_callback)(struct vmbus_channel *channel,
> +			u32 old, u32 new);
> +
>  	/*
>  	 * Synchronize channel scheduling and channel removal; see the inline
>  	 * comments in vmbus_chan_sched() and vmbus_reset_channel_cb().
> --
> 2.24.0


  reply index

Thread overview: 38+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-03-25 22:54 [RFC PATCH 00/11] VMBus channel interrupt reassignment Andrea Parri (Microsoft)
2020-03-25 22:54 ` [RFC PATCH 01/11] Drivers: hv: vmbus: Always handle the VMBus messages on CPU0 Andrea Parri (Microsoft)
2020-03-26 14:05   ` Vitaly Kuznetsov
2020-03-28 18:50     ` Andrea Parri
2020-03-25 22:54 ` [RFC PATCH 02/11] Drivers: hv: vmbus: Don't bind the offer&rescind works to a specific CPU Andrea Parri (Microsoft)
2020-03-26 14:16   ` Vitaly Kuznetsov
2020-03-26 15:47     ` Andrea Parri
2020-03-26 17:26       ` Vitaly Kuznetsov
2020-03-28 17:08         ` Andrea Parri
2020-03-29  3:43           ` Michael Kelley
2020-03-30 12:24             ` Vitaly Kuznetsov
2020-04-03 12:04               ` Andrea Parri
2020-03-25 22:54 ` [RFC PATCH 03/11] Drivers: hv: vmbus: Replace the per-CPU channel lists with a global array of channels Andrea Parri (Microsoft)
2020-03-26 14:31   ` Vitaly Kuznetsov
2020-03-26 17:05     ` Andrea Parri
2020-03-26 17:43       ` Vitaly Kuznetsov
2020-03-28 18:21         ` Andrea Parri
2020-03-29  3:49           ` Michael Kelley
2020-03-30 12:45           ` Vitaly Kuznetsov
2020-04-03 13:38             ` Andrea Parri
2020-04-03 14:56               ` Vitaly Kuznetsov
2020-03-25 22:54 ` [RFC PATCH 04/11] hv_netvsc: Disable NAPI before closing the VMBus channel Andrea Parri (Microsoft)
2020-03-26 15:26   ` Stephen Hemminger
2020-03-26 17:55     ` Andrea Parri
2020-03-25 22:54 ` [RFC PATCH 05/11] hv_utils: Always execute the fcopy and vss callbacks in a tasklet Andrea Parri (Microsoft)
2020-03-25 22:55 ` [RFC PATCH 06/11] Drivers: hv: vmbus: Use a spin lock for synchronizing channel scheduling vs. channel removal Andrea Parri (Microsoft)
2020-03-25 22:55 ` [RFC PATCH 07/11] PCI: hv: Prepare hv_compose_msi_msg() for the VMBus-channel-interrupt-to-vCPU reassignment functionality Andrea Parri (Microsoft)
2020-03-25 22:55 ` [RFC PATCH 08/11] Drivers: hv: vmbus: Remove the unused HV_LOCALIZED channel affinity logic Andrea Parri (Microsoft)
2020-03-25 22:55 ` [RFC PATCH 09/11] Drivers: hv: vmbus: Synchronize init_vp_index() vs. CPU hotplug Andrea Parri (Microsoft)
2020-03-25 22:55 ` [RFC PATCH 10/11] Drivers: hv: vmbus: Introduce the CHANNELMSG_MODIFYCHANNEL message type Andrea Parri (Microsoft)
2020-03-26 14:46   ` Vitaly Kuznetsov
2020-03-28 18:48     ` Andrea Parri
2020-04-03 14:55       ` Andrea Parri
2020-03-25 22:55 ` [RFC PATCH 11/11] scsi: storvsc: Re-init stor_chns when a channel interrupt is re-assigned Andrea Parri (Microsoft)
2020-03-30 16:42   ` Michael Kelley [this message]
2020-03-30 18:55     ` Andrea Parri
2020-03-30 19:49       ` Michael Kelley
2020-04-03 13:41         ` Andrea Parri

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=MW2PR2101MB105208138683A6DE0564745AD7CB0@MW2PR2101MB1052.namprd21.prod.outlook.com \
    --to=mikelley@microsoft.com \
    --cc=boqun.feng@gmail.com \
    --cc=decui@microsoft.com \
    --cc=haiyangz@microsoft.com \
    --cc=jejb@linux.ibm.com \
    --cc=kys@microsoft.com \
    --cc=linux-hyperv@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-scsi@vger.kernel.org \
    --cc=martin.petersen@oracle.com \
    --cc=parri.andrea@gmail.com \
    --cc=sthemmin@microsoft.com \
    --cc=vkuznets@redhat.com \
    --cc=wei.liu@kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

Linux-HyperV Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/linux-hyperv/0 linux-hyperv/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 linux-hyperv linux-hyperv/ https://lore.kernel.org/linux-hyperv \
		linux-hyperv@vger.kernel.org
	public-inbox-index linux-hyperv

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.kernel.vger.linux-hyperv


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git