> -----Original Message-----
> From: Michael Kelley <mikelley@xxxxxxxxxxxxx>
> Sent: Thursday, July 15, 2021 2:18 PM
> To: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>; linux-
> hyperv@xxxxxxxxxxxxxxx
> Cc: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>; KY Srinivasan
> <kys@xxxxxxxxxxxxx>; linux-kernel@xxxxxxxxxxxxxxx
> Subject: RE: [PATCH hyperv-fixes] Drivers: hv: vmbus: Fix duplicate CPU
> assignments within a device
>
> From: LKML haiyangz <lkmlhyz@xxxxxxxxxxxxx> On Behalf Of Haiyang Zhang
> Sent: Thursday, July 15, 2021 9:11 AM
> >
> > The vmbus module uses a rotational algorithm to assign target CPUs to
> > device's channels. Depends on the timing of different device's channel
> > offers, different channels of a device may be assigned to the same CPU.
> >
> > For example on a VM with 2 CPUs, if the NIC A and B's channels offered
>
> s/the NIC A/NIC A/
> s/offered/are offered/
Thanks, I will update this and other descriptions.
>
> > in the following order, the NIC A will have both channels on CPU0, and
>
> s/the NIC A/NIC A/
>
> > NIC B will have both channels on CPU1 -- see below. This kind of
> > assignments cause RSS spreading loads among different channels ends up
>
> "assignment causes RSS load that is spread across different channels to end
> up"
>
> > on the same CPU.
> >
> > Timing of channel offers:
> > NIC A channel 0
> > NIC B channel 0
> > NIC A channel 1
> > NIC B channel 1
> >
> > VMBUS ID 14: Class_ID = {f8615163-df3e-46c5-913f-f2d2f965ed0e} -
> Synthetic network adapter
> > Device_ID = {cab064cd-1f31-47d5-a8b4-9d57e320cccd}
> > Sysfs path: /sys/bus/vmbus/devices/cab064cd-1f31-47d5-a8b4-
> 9d57e320cccd
> > Rel_ID=14, target_cpu=0
> > Rel_ID=17, target_cpu=0
> >
> > VMBUS ID 16: Class_ID = {f8615163-df3e-46c5-913f-f2d2f965ed0e} -
> Synthetic network adapter
> > Device_ID = {244225ca-743e-4020-a17d-d7baa13d6cea}
> > Sysfs path: /sys/bus/vmbus/devices/244225ca-743e-4020-a17d-
> d7baa13d6cea
> > Rel_ID=16, target_cpu=1
> > Rel_ID=18, target_cpu=1
> >
> >
> > Update the vmbus' CPU assignment algorithm to avoid duplicate CPU
>
> s/vmbus'/vmbus/
>
> > assignments within a device.
> >
> > The new algorithm iterates 2 * #NUMA_Node + 1 times. In the first
> > round of checking all NUMA nodes, it tries to find previously unassigned
> > CPUs by this and other devices. If not available, it clears the
> > allocated CPU mask.
> > In the second round, it tries to find unassigned CPUs by the same
> > device.
> > In the last iteration, it assigns the channel to the first available CPU.
> > This is not normally expected, because during device probe, we limit the
> > number of channels of a device to be <= number of online CPUs.
> >
> > Signed-off-by: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>
> >
> > ---
> > drivers/hv/channel_mgmt.c | 95 ++++++++++++++++++++++++++---------
> ----
> > 1 file changed, 65 insertions(+), 30 deletions(-)
> >
> > diff --git a/drivers/hv/channel_mgmt.c b/drivers/hv/channel_mgmt.c
> > index caf6d0c4bc1b..fbddc4954f57 100644
> > --- a/drivers/hv/channel_mgmt.c
> > +++ b/drivers/hv/channel_mgmt.c
> > @@ -605,6 +605,17 @@ static void vmbus_process_offer(struct
> vmbus_channel *newchannel)
> > */
> > mutex_lock(&vmbus_connection.channel_mutex);
> >
> > + list_for_each_entry(channel, &vmbus_connection.chn_list, listentry)
> {
> > + if (guid_equal(&channel->offermsg.offer.if_type,
> > + &newchannel->offermsg.offer.if_type) &&
> > + guid_equal(&channel->offermsg.offer.if_instance,
> > + &newchannel->offermsg.offer.if_instance)) {
> > + fnew = false;
> > + newchannel->primary_channel = channel;
> > + break;
> > + }
> > + }
> > +
> > init_vp_index(newchannel);
> >
> > /* Remember the channels that should be cleaned up upon suspend.
> */
> > @@ -617,16 +628,6 @@ static void vmbus_process_offer(struct
> vmbus_channel *newchannel)
> > */
> > atomic_dec(&vmbus_connection.offer_in_progress);
> >
> > - list_for_each_entry(channel, &vmbus_connection.chn_list, listentry)
> {
> > - if (guid_equal(&channel->offermsg.offer.if_type,
> > - &newchannel->offermsg.offer.if_type) &&
> > - guid_equal(&channel->offermsg.offer.if_instance,
> > - &newchannel->offermsg.offer.if_instance)) {
> > - fnew = false;
> > - break;
> > - }
> > - }
> > -
> > if (fnew) {
> > list_add_tail(&newchannel->listentry,
> > &vmbus_connection.chn_list);
> > @@ -647,7 +648,6 @@ static void vmbus_process_offer(struct
> vmbus_channel *newchannel)
> > /*
> > * Process the sub-channel.
> > */
> > - newchannel->primary_channel = channel;
> > list_add_tail(&newchannel->sc_list, &channel->sc_list);
> > }
> >
> > @@ -683,6 +683,29 @@ static void vmbus_process_offer(struct
> vmbus_channel *newchannel)
> > queue_work(wq, &newchannel->add_channel_work);
> > }
> >
> > +/*
> > + * Clear CPUs used by other channels of the same device.
> > + * It's should only be called by init_vp_index().
> > + */
> > +static bool hv_clear_usedcpu(struct cpumask *cmask, struct
> vmbus_channel *chn)
> > +{
> > + struct vmbus_channel *primary = chn->primary_channel;
> > + struct vmbus_channel *sc;
> > +
> > + lockdep_assert_held(&vmbus_connection.channel_mutex);
> > +
> > + if (!primary)
> > + return !cpumask_empty(cmask);
>
> Minor note: As this function is currently used, the cmask parameter should
> never be empty. Calling cpumask_empty() doesn't hurt anything but
> this line should always return "true".
>
> > +
> > + cpumask_clear_cpu(primary->target_cpu, cmask);
> > +
> > + list_for_each_entry(sc, &primary->sc_list, sc_list)
> > + if (sc != chn)
> > + cpumask_clear_cpu(sc->target_cpu, cmask);
> > +
> > + return !cpumask_empty(cmask);
> > +}
> > +
> > /*
> > * We use this state to statically distribute the channel interrupt load.
> > */
> > @@ -705,7 +728,7 @@ static void init_vp_index(struct vmbus_channel
> *channel)
> > cpumask_var_t available_mask;
> > struct cpumask *alloced_mask;
> > u32 target_cpu;
> > - int numa_node;
> > + int numa_node, i;
> >
> > if ((vmbus_proto_version == VERSION_WS2008) ||
> > (vmbus_proto_version == VERSION_WIN7) || (!perf_chn) ||
> > @@ -724,29 +747,41 @@ static void init_vp_index(struct vmbus_channel
> *channel)
> > return;
> > }
> >
> > - while (true) {
> > - numa_node = next_numa_node_id++;
> > - if (numa_node == nr_node_ids) {
> > - next_numa_node_id = 0;
> > - continue;
> > + for (i = 1; i <= nr_node_ids * 2 + 1; i++) {
> > + while (true) {
> > + numa_node = next_numa_node_id++;
> > + if (numa_node == nr_node_ids) {
> > + next_numa_node_id = 0;
> > + continue;
> > + }
> > + if
> (cpumask_empty(cpumask_of_node(numa_node)))
> > + continue;
>
> This test has the potential to get the next_numa_node_id value out of sync
> with the index "i" in the containing "for" loop. The intent is to go through all
> NUMA nodes up to 2 times, plus 1 more time. But if a NUMA node is
> encountered
> with no online CPUs, next_numa_node_id may get incremented multiple
> times
> in a single iteration of the "for" loop. So you could end up cycling through
> some
> of the NUMA nodes more than twice before doing the final iteration.
I will switch to an update algorithm: see blow.
>
> > + break;
> > }
> > - if (cpumask_empty(cpumask_of_node(numa_node)))
> > - continue;
> > - break;
> > - }
> > - alloced_mask = &hv_context.hv_numa_map[numa_node];
> > + alloced_mask = &hv_context.hv_numa_map[numa_node];
> > +
> > + if (cpumask_weight(alloced_mask) ==
> > + cpumask_weight(cpumask_of_node(numa_node))) {
> > + /*
> > + * We have cycled through all the CPUs in the node;
> > + * reset the alloced map.
> > + */
> > + cpumask_clear(alloced_mask);
> > + }
> > +
> > + cpumask_xor(available_mask, alloced_mask,
> > + cpumask_of_node(numa_node));
> > +
> > + /* Try to avoid duplicate cpus within a device */
> > + if (channel->offermsg.offer.sub_channel_index >=
> > + num_online_cpus() ||
> > + i > nr_node_ids * 2 ||
> > + hv_clear_usedcpu(available_mask, channel))
> > + break;
> >
> > - if (cpumask_weight(alloced_mask) ==
> > - cpumask_weight(cpumask_of_node(numa_node))) {
> > - /*
> > - * We have cycled through all the CPUs in the node;
> > - * reset the alloced map.
> > - */
> > cpumask_clear(alloced_mask);
>
> This clearing of the "alloced_mask" concerns me. It is done incrementally
> as the NUMA nodes are cycled through the first time. So if an available CPU
> is found in the 3rd NUMA node on the first cycle, the alloced_mask will
> have been cleared for NUMA nodes 0, 1, and 2. A subsequent call to
> this function for a different device will find the CPUs in those NUMA nodes
> unalloc'ed, and therefore useable. It seems like this will tend to bunch the
> CPU
> assignments across multiple devices much more toward NUMA node 0 than
> the existing algorithm does. Or am I missing something about how this works?
The assignment is a rotational algorithm, and next_numa_node_id is a global
variable, which keeps the last assignment's numa id. So if the last device used
numa#3, the next device will try numa#4 (or numa 0, if there are only 4 nodes).
Then the next channel of this device will try the next numa node #5. And the
next one will try #6, etc... Because the rotational algorithm keeps rotating,
there is no preference towards any NUMA node.
But, I believe since we use cpumask_first(available_mask) instead of a rotation
within a node, the new algorithm does have preference on the "small" cpu
numbers WITHIN a numa node due to possible "extra" cpumask_clear() than
the old algorithm.
So I updated the algorithm like this below. Basically, it keeps running the rotational
algorithm until it finds a cpu not used by self. And any new device will start rotation
after the point of last cpu assignment. So, it doesn't have any preference on any
node or cpu.
Existing_find_next() {
Get the next numa node
Clear this node if all occupied
Cpu = the next available cpu of this node;
cpumask_set_cpu(cpu);
return cpu;
}
Updated_find_next() {
N = number_online_cpu()
for (i=1, i<= N+1, i++) {
cpu= Existing_find_next()
If (chn_index >= N ||
i > N ||
NotUsedBySelf(cpu))
Break;
}
channel->target_cpu = cpu;
}
Thanks,
- Haiyang
