Message ID | 20230328045122.25850-2-decui@microsoft.com (mailing list archive) |
---|---|
State | Not Applicable |
Headers | show |
Series | pci-hyper: fix race condition bugs for fast device hotplug | expand |
> -----Original Message----- > From: Dexuan Cui <decui@microsoft.com> > Sent: Tuesday, March 28, 2023 10:21 AM > To: bhelgaas@google.com; davem@davemloft.net; Dexuan Cui > <decui@microsoft.com>; edumazet@google.com; Haiyang Zhang > <haiyangz@microsoft.com>; Jake Oshins <jakeo@microsoft.com>; > kuba@kernel.org; kw@linux.com; KY Srinivasan <kys@microsoft.com>; > leon@kernel.org; linux-pci@vger.kernel.org; lpieralisi@kernel.org; Michael > Kelley (LINUX) <mikelley@microsoft.com>; pabeni@redhat.com; > robh@kernel.org; saeedm@nvidia.com; wei.liu@kernel.org; Long Li > <longli@microsoft.com>; boqun.feng@gmail.com > Cc: linux-hyperv@vger.kernel.org; linux-kernel@vger.kernel.org; linux- > rdma@vger.kernel.org; netdev@vger.kernel.org > Subject: [EXTERNAL] [PATCH 1/6] PCI: hv: fix a race condition bug in > hv_pci_query_relations() > > 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(). > > Fixes: 4daace0d8ce8 ("PCI: hv: Add paravirtual PCI front-end for Microsoft > Hyper-V VMs") > Signed-off-by: Dexuan Cui <decui@microsoft.com> > > --- > drivers/pci/controller/pci-hyperv.c | 13 +++++++++++++ > 1 file changed, 13 insertions(+) > > With the below debug code: > > @@ -2103,6 +2103,8 @@ static void survey_child_resources(struct > hv_pcibus_device *hbus) > } > > spin_unlock_irqrestore(&hbus->device_list_lock, flags); > + ssleep(15); > + printk("%s: completing %px\n", __func__, event); > complete(event); > } > > @@ -3305,8 +3307,12 @@ static int hv_pci_query_relations(struct hv_device > *hdev) > > ret = vmbus_sendpacket(hdev->channel, &message, sizeof(message), > 0, VM_PKT_DATA_INBAND, 0); > - if (!ret) > + if (!ret) { > + ssleep(10); // unassign the PCI device on the host during the > 10s > ret = wait_for_response(hdev, &comp); > + printk("%s: comp=%px is becoming invalid! ret=%d\n", > + __func__, &comp, ret); > + } > > return ret; > } > @@ -3635,6 +3641,8 @@ static int hv_pci_probe(struct hv_device *hdev, > > retry: > ret = hv_pci_query_relations(hdev); > + printk("hv_pci_query_relations() exited\n"); Can we use pr_* or the appropriate KERN_<LEVEL> in all the printk(s). > + > if (ret) > goto free_irq_domain; > > I'm able to repro the below hang issue: > > [ 74.544744] hv_pci b92a0085-468b-407a-a88a-d33fac8edc75: PCI VMBus > probing: Using version 0x10004 > [ 76.886944] hv_netvsc 818fe754-b912-4445-af51-1f584812e3c9 eth0: VF slot > 1 removed > [ 84.788266] hv_pci b92a0085-468b-407a-a88a-d33fac8edc75: The device is > gone. > [ 84.792586] hv_pci_query_relations: comp=ffffa7504012fb58 is becoming > invalid! ret=-19 > [ 84.797505] hv_pci_query_relations() exited > [ 89.652268] survey_child_resources: completing ffffa7504012fb58 > [ 150.392242] rcu: INFO: rcu_preempt detected stalls on CPUs/tasks: > [ 150.398447] rcu: 15-...0: (2 ticks this GP) > idle=867c/1/0x4000000000000000 softirq=947/947 fqs=5234 > [ 150.405851] rcu: (detected by 14, t=15004 jiffies, g=2553, q=4833 > ncpus=16) > [ 150.410870] Sending NMI from CPU 14 to CPUs 15: > [ 150.414836] NMI backtrace for cpu 15 > [ 150.414840] CPU: 15 PID: 10 Comm: kworker/u32:0 Tainted: G W E > 6.3.0-rc3-decui-dirty #34 > ... > [ 150.414849] Workqueue: hv_pci_468b pci_devices_present_work > [pci_hyperv] [ 150.414866] RIP: > 0010:__pv_queued_spin_lock_slowpath+0x10f/0x3c0 > ... > [ 150.414905] Call Trace: > [ 150.414907] <TASK> > [ 150.414911] _raw_spin_lock_irqsave+0x40/0x50 [ 150.414917] > complete+0x1d/0x60 [ 150.414924] pci_devices_present_work+0x5dd/0x680 > [pci_hyperv] [ 150.414946] process_one_work+0x21f/0x430 [ 150.414952] > worker_thread+0x4a/0x3c0 > > With this patch, the hang issue goes away: > > [ 186.143612] hv_pci b92a0085-468b-407a-a88a-d33fac8edc75: The device is > gone. > [ 186.148034] hv_pci_query_relations: comp=ffffa7cfd0aa3b50 is becoming > invalid! ret=-19 [ 191.263611] survey_child_resources: completing > ffffa7cfd0aa3b50 [ 191.267732] hv_pci_query_relations() exited > > diff --git a/drivers/pci/controller/pci-hyperv.c b/drivers/pci/controller/pci- > hyperv.c > index f33370b75628..b82c7cde19e6 100644 > --- a/drivers/pci/controller/pci-hyperv.c > +++ b/drivers/pci/controller/pci-hyperv.c > @@ -3308,6 +3308,19 @@ 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 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 > + * or sometimes a page fault. Flush hbus->wq before we exit from > + * hv_pci_query_relations() to avoid the issues. > + */ > + flush_workqueue(hbus->wq); > + > return ret; > } > > -- > 2.25.1
> From: Saurabh Singh Sengar <ssengar@microsoft.com> > Sent: Monday, March 27, 2023 10:29 PM > > ... > > --- Please note this special line "---". Anything after the special line and before the line "diff --git" is discarded automaticaly by 'git' and 'patch'. > > drivers/pci/controller/pci-hyperv.c | 13 +++++++++++++ > > 1 file changed, 13 insertions(+) > > > > @@ -3635,6 +3641,8 @@ static int hv_pci_probe(struct hv_device *hdev, > > > > retry: > > ret = hv_pci_query_relations(hdev); > > + printk("hv_pci_query_relations() exited\n"); > > Can we use pr_* or the appropriate KERN_<LEVEL> in all the printk(s). This is not part of the real patch :-) I just thought the debug code can help understand the issues resolved by the patches. I'll remove the debug code to avoid confusion if I need to post v2.
> diff --git a/drivers/pci/controller/pci-hyperv.c b/drivers/pci/controller/pci- > hyperv.c > index f33370b75628..b82c7cde19e6 100644 > --- a/drivers/pci/controller/pci-hyperv.c > +++ b/drivers/pci/controller/pci-hyperv.c > @@ -3308,6 +3308,19 @@ 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 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 > + * or sometimes a page fault. Flush hbus->wq before we exit from > + * hv_pci_query_relations() to avoid the issues. > + */ > + flush_workqueue(hbus->wq); Is it possible for PCI_BUS_RELATIONS to be scheduled arrive after calling flush_workqueue(hbus->wq)? > + > return ret; > } > > -- > 2.25.1
On Tue, Mar 28, 2023 at 05:38:59AM +0000, Dexuan Cui wrote: > > From: Saurabh Singh Sengar <ssengar@microsoft.com> > > Sent: Monday, March 27, 2023 10:29 PM > > > ... > > > --- > > Please note this special line "---". > Anything after the special line and before the line "diff --git" is discarded > automaticaly by 'git' and 'patch'. > > > > drivers/pci/controller/pci-hyperv.c | 13 +++++++++++++ > > > 1 file changed, 13 insertions(+) > > > > > > @@ -3635,6 +3641,8 @@ static int hv_pci_probe(struct hv_device *hdev, > > > > > > retry: > > > ret = hv_pci_query_relations(hdev); > > > + printk("hv_pci_query_relations() exited\n"); > > > > Can we use pr_* or the appropriate KERN_<LEVEL> in all the printk(s). > > This is not part of the real patch :-) > I just thought the debug code can help understand the issues > resolved by the patches. > I'll remove the debug code to avoid confusion if I need to post v2. I guess that means you *will* post a v2, right? Or do you expect somebody else to remove the debug code? If you do keep any debug or other logging, use pci_info() (or dev_info()) whenever possible. Also capitalize the subject line to match the others in the series. Bjorn
> From: Long Li <longli@microsoft.com> > Sent: Tuesday, March 28, 2023 9:49 AM > > --- a/drivers/pci/controller/pci-hyperv.c > > +++ b/drivers/pci/controller/pci-hyperv.c > > @@ -3308,6 +3308,19 @@ 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 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 > > + * or sometimes a page fault. Flush hbus->wq before we exit from > > + * hv_pci_query_relations() to avoid the issues. > > + */ > > + flush_workqueue(hbus->wq); > > Is it possible for PCI_BUS_RELATIONS to be scheduled arrive after calling > flush_workqueue(hbus->wq)? It's possible, but that doesn't matter: hv_pci_query_relations() is called only once, and it sets hbus->survey_event to point to the stack variable 'comp'. The first survey_child_resources() calls complete() for the 'comp' and sets hbus->survey_event to NULL. When the second survey_child_resources() is called, hbus->survey_event is NULL, so survey_child_resources() returns immediately. According to my test, after hv_pci_enter_d0() posts PCI_BUS_D0ENTRY, the guest receives a second PCI_BUS_RELATIONS2 message, which is the same as the first PCI_BUS_RELATIONS2 message, which is basically a no-op in pci_devices_present_work(), especially with the newly-introduced per-hbus state_lock mutex.
> From: Bjorn Helgaas <helgaas@kernel.org> > Sent: Tuesday, March 28, 2023 10:25 AM > To: Dexuan Cui <decui@microsoft.com> > ... > On Tue, Mar 28, 2023 at 05:38:59AM +0000, Dexuan Cui wrote: > > > From: Saurabh Singh Sengar <ssengar@microsoft.com> > > > Sent: Monday, March 27, 2023 10:29 PM > > > > ... > > > > --- > > > > Please note this special line "---". > > Anything after the special line and before the line "diff --git" is discarded > > automaticaly by 'git' and 'patch'. > > > > > > drivers/pci/controller/pci-hyperv.c | 13 +++++++++++++ > > > > 1 file changed, 13 insertions(+) > > > > > > > > @@ -3635,6 +3641,8 @@ static int hv_pci_probe(struct hv_device *hdev, > > > > > > > > retry: > > > > ret = hv_pci_query_relations(hdev); > > > > + printk("hv_pci_query_relations() exited\n"); > > > > > > Can we use pr_* or the appropriate KERN_<LEVEL> in all the printk(s). > > > > This is not part of the real patch :-) > > I just thought the debug code can help understand the issues > > resolved by the patches. > > I'll remove the debug code to avoid confusion if I need to post v2. > > I guess that means you *will* post a v2, right? I guess I didn't make myself clear, sorry. The "debug code" is not part of the real patch body -- if we run the "patch" program or "git am" to apply the patches, the "debug code" is automatically dropped because it's between the special "---" line and the real start of the patch body (i.e. the "diff --git" line). So far, IMO I don't have to post v2 because the patch body and the patch description (except for the part that's automatically removed by 'patch' and 'git') don't need any change. > Or do you expect > somebody else to remove the debug code? If you do keep any debug or > other logging, use pci_info() (or dev_info()) whenever possible. As I explained above, 'patch' and 'git' automatically remove the part that don't have to be in the git history. > > Also capitalize the subject line to match the others in the series. > > Bjorn Thanks for catching this! If this is the only thing to be fixed, I hope the PCI folks can help fix this when accepting the patch. If you think I should post v2, please let me know.
On Wed, Mar 29, 2023 at 08:37:20AM +0000, Dexuan Cui wrote: > > From: Bjorn Helgaas <helgaas@kernel.org> > > Sent: Tuesday, March 28, 2023 10:25 AM > > To: Dexuan Cui <decui@microsoft.com> > > ... > > On Tue, Mar 28, 2023 at 05:38:59AM +0000, Dexuan Cui wrote: > > > > From: Saurabh Singh Sengar <ssengar@microsoft.com> > > > > Sent: Monday, March 27, 2023 10:29 PM > > > > > ... > > > > > --- > > > > > > Please note this special line "---". > > > Anything after the special line and before the line "diff --git" is discarded > > > automaticaly by 'git' and 'patch'. > > > > > > > > drivers/pci/controller/pci-hyperv.c | 13 +++++++++++++ > > > > > 1 file changed, 13 insertions(+) > > > > > > > > > > @@ -3635,6 +3641,8 @@ static int hv_pci_probe(struct hv_device *hdev, > > > > > > > > > > retry: > > > > > ret = hv_pci_query_relations(hdev); > > > > > + printk("hv_pci_query_relations() exited\n"); > > > > > > > > Can we use pr_* or the appropriate KERN_<LEVEL> in all the printk(s). > > > > > > This is not part of the real patch :-) > > > I just thought the debug code can help understand the issues > > > resolved by the patches. > > > I'll remove the debug code to avoid confusion if I need to post v2. > > > > I guess that means you *will* post a v2, right? > > I guess I didn't make myself clear, sorry. The "debug code" is not > part of the real patch body -- if we run the "patch" program or "git am" > to apply the patches, the "debug code" is automatically dropped because > it's between the special "---" line and the real start of the patch body (i.e. > the "diff --git" line). > > So far, IMO I don't have to post v2 because the patch body and the patch > description (except for the part that's automatically removed by 'patch' > and 'git') don't need any change. Ah, sorry, I didn't notice that, even though you explicitly pointed it out earlier. No need to repost just to capitalize the subject, Lorenzo can do it or not as he chooses. Bjorn
diff --git a/drivers/pci/controller/pci-hyperv.c b/drivers/pci/controller/pci-hyperv.c index f33370b75628..b82c7cde19e6 100644 --- a/drivers/pci/controller/pci-hyperv.c +++ b/drivers/pci/controller/pci-hyperv.c @@ -3308,6 +3308,19 @@ 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 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 + * or sometimes a page fault. Flush hbus->wq before we exit from + * hv_pci_query_relations() to avoid the issues. + */ + flush_workqueue(hbus->wq); + return ret; }
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(). Fixes: 4daace0d8ce8 ("PCI: hv: Add paravirtual PCI front-end for Microsoft Hyper-V VMs") Signed-off-by: Dexuan Cui <decui@microsoft.com> --- drivers/pci/controller/pci-hyperv.c | 13 +++++++++++++ 1 file changed, 13 insertions(+) With the below debug code: @@ -2103,6 +2103,8 @@ static void survey_child_resources(struct hv_pcibus_device *hbus) } spin_unlock_irqrestore(&hbus->device_list_lock, flags); + ssleep(15); + printk("%s: completing %px\n", __func__, event); complete(event); } @@ -3305,8 +3307,12 @@ static int hv_pci_query_relations(struct hv_device *hdev) ret = vmbus_sendpacket(hdev->channel, &message, sizeof(message), 0, VM_PKT_DATA_INBAND, 0); - if (!ret) + if (!ret) { + ssleep(10); // unassign the PCI device on the host during the 10s ret = wait_for_response(hdev, &comp); + printk("%s: comp=%px is becoming invalid! ret=%d\n", + __func__, &comp, ret); + } return ret; } @@ -3635,6 +3641,8 @@ static int hv_pci_probe(struct hv_device *hdev, retry: ret = hv_pci_query_relations(hdev); + printk("hv_pci_query_relations() exited\n"); + if (ret) goto free_irq_domain; I'm able to repro the below hang issue: [ 74.544744] hv_pci b92a0085-468b-407a-a88a-d33fac8edc75: PCI VMBus probing: Using version 0x10004 [ 76.886944] hv_netvsc 818fe754-b912-4445-af51-1f584812e3c9 eth0: VF slot 1 removed [ 84.788266] hv_pci b92a0085-468b-407a-a88a-d33fac8edc75: The device is gone. [ 84.792586] hv_pci_query_relations: comp=ffffa7504012fb58 is becoming invalid! ret=-19 [ 84.797505] hv_pci_query_relations() exited [ 89.652268] survey_child_resources: completing ffffa7504012fb58 [ 150.392242] rcu: INFO: rcu_preempt detected stalls on CPUs/tasks: [ 150.398447] rcu: 15-...0: (2 ticks this GP) idle=867c/1/0x4000000000000000 softirq=947/947 fqs=5234 [ 150.405851] rcu: (detected by 14, t=15004 jiffies, g=2553, q=4833 ncpus=16) [ 150.410870] Sending NMI from CPU 14 to CPUs 15: [ 150.414836] NMI backtrace for cpu 15 [ 150.414840] CPU: 15 PID: 10 Comm: kworker/u32:0 Tainted: G W E 6.3.0-rc3-decui-dirty #34 ... [ 150.414849] Workqueue: hv_pci_468b pci_devices_present_work [pci_hyperv] [ 150.414866] RIP: 0010:__pv_queued_spin_lock_slowpath+0x10f/0x3c0 ... [ 150.414905] Call Trace: [ 150.414907] <TASK> [ 150.414911] _raw_spin_lock_irqsave+0x40/0x50 [ 150.414917] complete+0x1d/0x60 [ 150.414924] pci_devices_present_work+0x5dd/0x680 [pci_hyperv] [ 150.414946] process_one_work+0x21f/0x430 [ 150.414952] worker_thread+0x4a/0x3c0 With this patch, the hang issue goes away: [ 186.143612] hv_pci b92a0085-468b-407a-a88a-d33fac8edc75: The device is gone. [ 186.148034] hv_pci_query_relations: comp=ffffa7cfd0aa3b50 is becoming invalid! ret=-19 [ 191.263611] survey_child_resources: completing ffffa7cfd0aa3b50 [ 191.267732] hv_pci_query_relations() exited