From: Lorenzo Pieralisi <lpieralisi@kernel.org>
Sent: Thursday, May 25, 2023 1:14 AM
To: Dexuan Cui <decui@microsoft.com>

On Wed, Apr 19, 2023 at 07:40:32PM -0700, Dexuan Cui wrote:

quoted

Fix the longstanding race between hv_pci_query_relations() and
survey_child_resources() by flushing the workqueue before we exit from
hv_pci_query_relations().

"Fix the longstanding race" is vague. Please describe the race
succinctly at least to give an idea of what the problem is.

Hi Lozenzo, I'm sorry for the late response -- finally I'm back on the patchset...

I'll use the below commit message in v4:
 
 Since day 1 of the driver, there has been a race between
 hv_pci_query_relations() and survey_child_resources(): during fast
 device hotplug, hv_pci_query_relations() may error out due to
 device-remove and the stack variable 'comp' is no longer valid;
 however, pci_devices_present_work() -> survey_child_resources() ->
 complete() may be running on another CPU and accessing the no-longer-valid
 'comp'. Fix the race by flushing the workqueue before we exit from
 hv_pci_query_relations().

I'll also update the comment to share more details of the race:

--- a/drivers/pci/controller/pci-hyperv.c
+++ b/drivers/pci/controller/pci-hyperv.c

@@ -3401,6 +3401,24 @@ static int hv_pci_query_relations(struct hv_device *hdev)
        if (!ret)
                ret = wait_for_response(hdev, &comp);

+       /*
+        * In the case of fast device addition/removal, it's possible that
+        * vmbus_sendpacket() or wait_for_response() returns -ENODEV but we
+        * already got a PCI_BUS_RELATIONS* message from the host and the
+        * channel callback already scheduled a work to hbus->wq, which can be
+        * running pci_devices_present_work() -> survey_child_resources() ->
+        * complete(&hbus->survey_event), even after hv_pci_query_relations()
+        * exits and the stack variable 'comp' is no longer valid; as a result,
+        * a hang or a page fault may happen when the complete() calls
+        * raw_spin_lock_irqsave(). Flush hbus->wq before we exit from
+        * hv_pci_query_relations() to avoid the issues. Note: if 'ret' is
+        * -ENODEV, there can't be any more work item scheduled to hbus->wq
+        * after the flush_workqueue(): see vmbus_onoffer_rescind() ->
+        * vmbus_reset_channel_cb(), vmbus_rescind_cleanup() ->
+        * channel->rescind = true.
+        */
+       flush_workqueue(hbus->wq);
+
        return ret;
 }

quoted

+	 * In the case of fast device addition/removal, it's possible that
+	 * vmbus_sendpacket() or wait_for_response() returns -ENODEV but we
+	 * already got a PCI_BUS_RELATIONS* message from the host and the
+	 * channel callback already scheduled a work to hbus->wq, which can
 be
+	 * running survey_child_resources() -> complete(&hbus->survey_event),
+	 * even after hv_pci_query_relations() exits and the stack variable
+	 * 'comp' is no longer valid. This can cause a strange hang issue

"A strange hang" sounds like we don't understand what's happening, it
does not feel like it is a solid understanding of the issue.

I would remove it - given that you already explain that comp is no
longer valid - that is already a bug that needs fixing.

I share more details in the comment (see the above).
 

Acked-by: Lorenzo Pieralisi <lpieralisi@kernel.org>

Thanks!