On 10/22/2024 11:33 PM, Michael Kelley wrote:
From: Naman Jain <namjain@xxxxxxxxxxxxxxxxxxx> Sent: Tuesday, October 22, 2024 2:50 AM
On 10/21/2024 10:03 AM, Michael Kelley wrote:
From: Naman Jain <namjain@xxxxxxxxxxxxxxxxxxx> Sent: Friday, October 18, 2024 4:58 AM
[snip]
@@ -1297,12 +1285,11 @@
EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
/*
* vmbus_onoffers_delivered -
* This is invoked when all offers have been delivered.
- *
- * Nothing to do here.
*/
I'm unclear on the meaning of the ALLOFFERS_DELIVERED message. In the
early days of Hyper-V, the set of virtual devices in a VM was fixed before a
VM started, so presumably ALLOFFERS_DELIVERED meant that offers for
that fixed set of devices had been delivered. That meaning makes sense
conceptually.
But more recent versions of Hyper-V support adding and removing devices
at any time during the life of the VM. If a device is added, a new offer is
generated. Existing devices (such as netvsc) can reconfigure their channels,
in which case new subchannel offers are generated. So the core concept of
"all offers delivered" isn't valid anymore.
To date Linux hasn't done anything with this message, so we didn't need
precision about what it means. There's no VMBus specification or
documentation, so we need some text here in the comments that
explains precisely what ALLOFFERS_DELIVERED means, and how that
meaning aligns with the fact that offers can be delivered anytime during
the life of the VM.
I'll note that in the past there's also been some behavior where during guest
boot, Hyper-V offers a virtual PCI device to the guest, immediately
rescinds the vPCI device (before Linux even finished processing the offer),
then offers it again. I wonder about how ALLOFFERS_DELIVERED interacts
with that situation.
I ran some tests in an Azure L16s_v3 VM, which has three vPCI devices:
2 NVMe devices and 1 Mellanox CX-5. The offers for the 2 NVMe devices
came before the ALLOFFERS_DELIVERED message, but the offer for
the Mellanox CX-5 came *after* the ALLOFFERS_DELIVERED message.
If that's the way the Mellanox CX-5 offers work, this patch breaks things
in the hibernation resume path because ALLOFFERS_DELIVERED isn't
sufficient to indicate that all primary channels have been re-offered
and the resume can proceed. All sub-channel offers came after the
ALLOFFERS_DELIVERED message, but that is what I expected and
shouldn't be a problem.
It's hard for me to review this patch set without some precision around
what ALLOFFERS_DELIVERED means.
Thanks for your valuable comments.
I checked this internally with John and here is the information you are
looking for.
CHANNELMSG_ALLOFFERS_DELIVERED message arrives after all the boot-time
offers are delivered. Other channels can be hot added later, even
immediately after the all-offers-delivered message.
A boot-time offer will be any of the virtual hardware the VM is
configured with at boot. The reason the Mellanox NIC is not included in
this list is because the Mellanoc NIC is an optional accelerator to the
synthetic vmbus NIC. It is hot added only after the vmbus NIC channel is
opened (once we know the guest can support it, via the sriov bit in the
netvsc protocol). So, it is not there at boot time, as we define it.
Good info about the meaning of ALLOFFERS_DELIVERED! Please
capture the info in comments in the code so it's available for future
readers.
Noted.
The reason the logs show the ordering below is that the vmbus driver
queues offer work to a work queue. The logs are out of order from the
actual messages that arrive. That’s why we have to flush the work queues
before we consider all the offers handled.
I'm not sure which logs are referenced in the above paragraph. When
I captured logs (which I didn't send), I did so in vmbus_on_msg_dpc(),
which is before the offer messages are put in the work queue. So my
logs reflected the order of the actual messages. But I agree that the
work queues need to be flushed.
My bad.
The logs, which I referred to here, was actually ftrace logs, for
different channels offer, request for offer and all offers delivered
event. My observation was then based on how these works are processed in
workqueues. I was observing that all offers delivered event was being
processed before channel offers. This was misleading, and sorry for the
confusion.
Regarding your comment for MLX CX-5, it is by design.The MLX device must
be logically hot removed and hot added back across a hibernate. There is
no guarantee the same accelerator device is present with the same
capabilities or device IDs or anything; the VHD could have been moved to
another machine with a different NIC, or to a VM without accelnet, or
whatever. This is all supported. So, it's not required to wait for the
MLX device to come back as part of hibernate resume.
Since Linux doesn't currently handle this correctly, the host works
around this by removing the MLX device before the guest hibernates. This
way, the guest does not expect an MLX device to still be present on
resume. This is inconvenient because it means the guest cannot initiate
hibernate on its own.
I wasn't aware of the VF handling. Where does the guest notify the
host that it is planning to hibernate? I went looking for such code, but
couldn't immediately find it. Is it in the netvsc driver? Is this the
sequence?
1) The guest notifies the host of the hibernate
2) The host sends a RESCIND_CHANNELOFFER message for each VF
in the VM
3) The guest waits for all VF rescind processing to complete, and
also must ensure that no new VFs get added in the meantime
4) Then the guest proceeds with the hibernation, knowing that there
are no open channels for VF devices
Dexuan has responded to this query.
The behavior we want is for the guest to hot remove the MLX device
driver on resume, even if the MLX device was still present at suspend,
so that the host does not need this special pre-hibernate behavior. This
patch series may not be sufficient to ensure this, though. It just moves
things in the right direction, by handling the all-offers-delivered
message.
This aspect of the patch might be worth calling out in the commit
message.
Will do.
static void vmbus_onoffers_delivered(
struct vmbus_channel_message_header *hdr)
{
+ complete(&vmbus_connection.all_offers_delivered_event);
}
/*
@@ -1578,7 +1565,8 @@ void vmbus_onmessage(struct vmbus_channel_message_header *hdr)
}
/*
- * vmbus_request_offers - Send a request to get all our pending offers.
+ * vmbus_request_offers - Send a request to get all our pending offers
+ * and wait for all offers to arrive.
*/
int vmbus_request_offers(void)
{
@@ -1596,6 +1584,10 @@ int vmbus_request_offers(void)
msg->msgtype = CHANNELMSG_REQUESTOFFERS;
+ /*
+ * This REQUESTOFFERS message will result in the host sending an all
+ * offers delivered message.
+ */
ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
true);
@@ -1607,6 +1599,22 @@ int vmbus_request_offers(void)
goto cleanup;
}
+ /* Wait for the host to send all offers. */
+ while (wait_for_completion_timeout(
+ &vmbus_connection.all_offers_delivered_event, msecs_to_jiffies(10 * 1000)) == 0) {
Maybe do !wait_for_completion_timeout( ...) instead of explicitly testing
for 0? I see that some existing code uses the explicit test for 0, but that's
not the usual kernel code idiom.
+ pr_warn("timed out waiting for all offers to be delivered...\n");
+ }
The while loop could continue forever. We don't want to trust the Hyper-V
host to be well-behaved, so there should be an uber timeout where it gives
up after 100 seconds (or pick some value). If the uber timeout is reached,
I'm not sure if the code should panic or just go on -- it's debatable. But doing
something explicit is probably better than repeatedly outputting the
warning message forever.
You're right its debatable. It seems to be a best effort mechanism, and
we can't say for sure what all offers are going to be provided. So
printing a warning and moving ahead seems right to me. I'll remove the
loop, and keep it simple.
+
+ /*
+ * Flush handling of offer messages (which may initiate work on
+ * other work queues).
+ */
+ flush_workqueue(vmbus_connection.work_queue);
+
+ /* Flush processing the incoming offers. */
+ flush_workqueue(vmbus_connection.handle_primary_chan_wq);
+ flush_workqueue(vmbus_connection.handle_sub_chan_wq);
Why does the sub-channel workqueue need to be flushed? Sub-channels
get created at the request of the Linux driver, in cooperation with the VSP
on the Hyper-V side. If the top-level goal is for user-space drivers to be
able to know that at least a core set of offers have been processed, it
seems like waiting for sub-channel offers isn't necessary. And new
subchannel offers could arrive immediately after this flush, so the flush
doesn't provide any guarantee about whether there are offers in that
workqueue. If there is a valid reason to wait, some explanatory
comments would be helpful.
I'll either remove it, or put a comment to manage the expectations that
this is just for best effort and there is no guarantee that all sub
channel offers will be entertained by now.
I lean toward removing it, as flushing the sub channel work queue
doesn't seem related to the stated goals of the patch. It just leaves a
dangling "why is this flush being done?" question.
Understood, will remove it. Thanks.
Michael
Regards,
Naman
