Skip to contentSkip to conversation

PCI/hotplug: Replaced down_write_nested with hotplug_slot_rwsem if ctrl->depth > 0 when taking the ctrl->reset_lock.

PATCH In review Anatoli Antonovitch 5 participants 10 messages Jan 13, 2023 – Apr 10, 2023 PCI

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 PCI/hotplug: Replaced down_write_nested with hotplug_slot_rwsem if ctrl->depth > 0 when taking the ctrl->reset_lock.

From: Anatoli Antonovitch <[email protected]>

It is to avoid any potential issues when S3 resume but at the same time we want to hot-unplug.

To fix the race between pciehp and AER reported in https://bugzilla.kernel.org/show_bug.cgi?id=215590

INFO: task irq/26-aerdrv:95 blocked for more than 120 seconds.
Tainted: G        W          6.2.0-rc3-custom-norework-jan11+ #29
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
task:irq/26-aerdrv   state:D stack:0     pid:95    ppid:2      flags:0x00004000
Call Trace:
<TASK>
__schedule+0x3df/0xec0
? rcu_read_lock_held_common+0x12/0x60
schedule+0x6b/0x100
rwsem_down_write_slowpath+0x3b2/0x9c0
down_write_nested+0x16b/0x220
pciehp_reset_slot+0x63/0x160
pci_reset_hotplug_slot+0x44/0x80
pci_slot_reset+0x10d/0x1a0
pci_bus_error_reset+0xb2/0xe0
aer_root_reset+0x144/0x1a0
pcie_do_recovery+0x15a/0x280
? __pfx_aer_root_reset+0x20/0x20
aer_process_err_devices+0xfa/0x115
aer_isr.cold+0x52/0xa1
? __kmem_cache_free+0x36a/0x3c0
? irq_thread+0xb0/0x1e0
? irq_thread+0xb0/0x1e0
irq_thread_fn+0x28/0x80
irq_thread+0x106/0x1e0
? __pfx_irq_thread_fn+0x20/0x20
? __pfx_irq_thread_dtor+0x20/0x20
? __pfx_irq_thread+0x20/0x20
kthread+0x10a/0x140
? __pfx_kthread+0x20/0x20
ret_from_fork+0x35/0x60
</TASK>

INFO: task irq/26-pciehp:96 blocked for more than 120 seconds.
Tainted: G        W          6.2.0-rc3-custom-norework-jan11+ #29
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
task:irq/26-pciehp   state:D stack:0     pid:96    ppid:2      flags:0x00004000
Call Trace:
<TASK>
__schedule+0x3df/0xec0
? rcu_read_lock_sched_held+0x25/0x80
schedule+0x6b/0x100
schedule_preempt_disabled+0x18/0x40
__mutex_lock+0x685/0xf60
? rcu_read_lock_sched_held+0x25/0x80
? rcu_read_lock_held_common+0x12/0x60
? pci_dev_set_disconnected+0x1b/0x80
mutex_lock_nested+0x1b/0x40
? mutex_lock_nested+0x1b/0x40
pci_dev_set_disconnected+0x1b/0x80
? __pfx_pci_dev_set_disconnected+0x20/0x20
pci_walk_bus+0x48/0xa0
pciehp_unconfigure_device+0x129/0x140
pciehp_disable_slot+0x6e/0x120
pciehp_handle_presence_or_link_change+0xf1/0x320
pciehp_ist+0x1a0/0x1c0
? irq_thread+0xb0/0x1e0
irq_thread_fn+0x28/0x80
irq_thread+0x106/0x1e0
? __pfx_irq_thread_fn+0x20/0x20
? __pfx_irq_thread_dtor+0x20/0x20
? __pfx_irq_thread+0x20/0x20
kthread+0x10a/0x140
? __pfx_kthread+0x20/0x20
ret_from_fork+0x35/0x60
</TASK>
---
 drivers/pci/hotplug/pciehp_hpc.c | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)
drivers/pci/hotplug/pciehp_hpc.c +11 -2
index 10e9670eea0b..b1084e67f798 100644
@@ -27,6 +27,8 @@
 #include "../pci.h"
 #include "pciehp.h"
 
+static DECLARE_RWSEM(hotplug_slot_rwsem);
+
 static const struct dmi_system_id inband_presence_disabled_dmi_table[] = {
 	/*
 	 * Match all Dell systems, as some Dell systems have inband
@@ -911,7 +913,10 @@ int pciehp_reset_slot(struct hotplug_slot *hotplug_slot, bool probe)
 	if (probe)
 		return 0;
 
-	down_write_nested(&ctrl->reset_lock, ctrl->depth);
+	if (ctrl->depth > 0)
+		down_write_nested(&ctrl->reset_lock, ctrl->depth);
+	else
+		down_write(&hotplug_slot_rwsem);
 
 	if (!ATTN_BUTTN(ctrl)) {
 		ctrl_mask |= PCI_EXP_SLTCTL_PDCE;
@@ -931,7 +936,11 @@ int pciehp_reset_slot(struct hotplug_slot *hotplug_slot, bool probe)
 	ctrl_dbg(ctrl, "%s: SLOTCTRL %x write cmd %x\n", __func__,
 		 pci_pcie_cap(ctrl->pcie->port) + PCI_EXP_SLTCTL, ctrl_mask);
 
-	up_write(&ctrl->reset_lock);
+	if (ctrl->depth > 0)
+		up_write(&ctrl->reset_lock);
+	else
+		up_write(&hotplug_slot_rwsem);
+
 	return rc;
 }
 
Signature · 1 line
2.25.1
#1 lore raw

Hi Anatoli,

Quoted 5 lines · Anatoli Antonovitch

It is to avoid any potential issues when S3 resume but at the same time we want to hot-unplug.

To fix the race between pciehp and AER reported in
https://bugzilla.kernel.org/show_bug.cgi?id=215590

I've just submitted an alternative patch to fix this, could you give it a spin and see if the issue goes away?

https://patchwork.kernel.org/project/linux-pci/patch/3dc88ea82bdc0e37d9000e413d5ebce481cbd629.1674205689.git.lukas@wunner.de/

That alternative approach is preferable IMO because it also solves the problem that marking devices as permanently offline isn't possible concurrently to driver bind/unbind at the moment. Additionally, the alternative patch simplifies locking and reduces code size.

Thanks and sorry for my belated response.

Lukas

#2 lore raw

Hi Lukas,

The patch has been tested on the same setup.
Unfortunately, this alternative approach with optimization and simplified locking is not resolve the issue in
https://bugzilla.kernel.org/show_bug.cgi?id=215590

I have uploaded the log dmesg_6_2_rc4_hotadd_aer_fix_a6bd101b8f84.txt into the bugzilla for your patch.

Thanks,

Anatoli

Quoted 4 lines · Lukas Wunner

I've just submitted an alternative patch to fix this, could you give it a spin and see if the issue goes away?

https://patchwork.kernel.org/project/linux-pci/patch/3dc88ea82bdc0e37d9000e413d5ebce481cbd629.1674205689.git.lukas@wunner.de/

#3 lore raw
Quoted 6 lines · Anatoli Antonovitch
Quoted 4 lines · Lukas Wunner

I've just submitted an alternative patch to fix this, could you give it a spin and see if the issue goes away?

https://patchwork.kernel.org/project/linux-pci/patch/3dc88ea82bdc0e37d9000e413d5ebce481cbd629.1674205689.git.lukas@wunner.de/

Quoted 7 lines

The patch has been tested on the same setup.
Unfortunately, this alternative approach with optimization and simplified locking is not resolve the issue in
https://bugzilla.kernel.org/show_bug.cgi?id=215590

I have uploaded the log dmesg_6_2_rc4_hotadd_aer_fix_a6bd101b8f84.txt into the bugzilla for your patch.

You're now getting a different deadlock. That one is addressed by this old patch (it's already linked from the bugzilla):

https://lore.kernel.org/linux-pci/908047f7699d9de9ec2efd6b79aa752d73dab4b6.1595329748.git.lukas@wunner.de/

If you apply that patch plus the new one, do you still see a deadlock?

Thanks,

Lukas

#4 lore raw

I do not see a deadlock, when applying the following old patch:
https://lore.kernel.org/linux-pci/908047f7699d9de9ec2efd6b79aa752d73dab4b6.1595329748.git.lukas@wunner.de/
after merge for the kernel 6.2.0-rc5, and applied the alternative patch:
https://patchwork.kernel.org/project/linux-pci/patch/3dc88ea82bdc0e37d9000e413d5ebce481cbd629.1674205689.git.lukas@wunner.de/

I have uploaded the merged patch and the system log for the upstream kernel.

Anatoli

Quoted 6 lines · Lukas Wunner

You're now getting a different deadlock. That one is addressed by this old patch (it's already linked from the bugzilla):

https://lore.kernel.org/linux-pci/908047f7699d9de9ec2efd6b79aa752d73dab4b6.1595329748.git.lukas@wunner.de/

If you apply that patch plus the new one, do you still see a deadlock?

#5 lore raw

Hi Lukas,

Can we revisit the patches again to get a fix?
The issue still reproduce and visible in the kernel 6.2.0-rc8.

Thanks,
Anatoli

Quoted 20 lines · Anatoli Antonovitch

I do not see a deadlock, when applying the following old patch:
https://lore.kernel.org/linux-pci/908047f7699d9de9ec2efd6b79aa752d73dab4b6.1595329748.git.lukas@wunner.de/

after merge for the kernel 6.2.0-rc5, and applied the alternative patch:
https://patchwork.kernel.org/project/linux-pci/patch/3dc88ea82bdc0e37d9000e413d5ebce481cbd629.1674205689.git.lukas@wunner.de/

I have uploaded the merged patch and the system log for the upstream kernel.

Anatoli

Quoted 7 lines · Lukas Wunner

You're now getting a different deadlock. That one is addressed by this old patch (it's already linked from the bugzilla):

https://lore.kernel.org/linux-pci/908047f7699d9de9ec2efd6b79aa752d73dab4b6.1595329748.git.lukas@wunner.de/

If you apply that patch plus the new one, do you still see a deadlock?

#6 lore raw
Quoted 8 lines · Anatoli Antonovitch

Hi Lukas,

Can we revisit the patches again to get a fix?
The issue still reproduce and visible in the kernel 6.2.0-rc8.

Quoted 2 lines · Anatoli Antonovitch

I do not see a deadlock, when applying the following old patch:
https://lore.kernel.org/linux-pci/908047f7699d9de9ec2efd6b79aa752d73dab4b6.1595329748.git.lukas@wunner.de/

This old patch would need to be updated and reposted. There was a 0-day bot issue and a question to be resolved. Maybe this is all already resolved, but it needs to be posted and tested with a current kernel.

Quoted 2 lines
Quoted 2 lines

after merge for the kernel 6.2.0-rc5, and applied the alternative patch:
https://patchwork.kernel.org/project/linux-pci/patch/3dc88ea82bdc0e37d9000e413d5ebce481cbd629.1674205689.git.lukas@wunner.de/

This one is on track to appear in v6.3-rc1:
https://git.kernel.org/cgit/linux/kernel/git/pci/pci.git/commit/?id=74ff8864cc84

Quoted 14 lines
Quoted 14 lines

I have uploaded the merged patch and the system log for the upstream kernel.

Anatoli

Quoted 7 lines · Lukas Wunner

You're now getting a different deadlock. That one is addressed by this old patch (it's already linked from the bugzilla):

https://lore.kernel.org/linux-pci/908047f7699d9de9ec2efd6b79aa752d73dab4b6.1595329748.git.lukas@wunner.de/

If you apply that patch plus the new one, do you still see a deadlock?

#7 lore raw

[Public]

Quoted 18 lines

-----Original Message-----
From: Bjorn Helgaas <[email protected]>
Sent: Friday, February 17, 2023 11:04 AM
To: Antonovitch, Anatoli <[email protected]>
Cc: Lukas Wunner <[email protected]>; Anatoli Antonovitch <[email protected]>; [email protected]; [email protected]; Deucher, Alexander <[email protected]>; Koenig, Christian <[email protected]>
Subject: Re: [PATCH] PCI/hotplug: Replaced down_write_nested with hotplug_slot_rwsem if ctrl->depth > 0 when taking the ctrl->reset_lock.

Quoted 4 lines · Anatoli Antonovitch

Hi Lukas,

Can we revisit the patches again to get a fix?
The issue still reproduce and visible in the kernel 6.2.0-rc8.

Quoted 8 lines
Quoted 4 lines · Anatoli Antonovitch
Quoted 3 lines

I do not see a deadlock, when applying the following old patch:
https://lore.kernel.org/linux-pci/908047f7699d9de9ec2efd6b79aa752d73
[email protected]/

This old patch would need to be updated and reposted. There was a 0-day bot issue and a question to be resolved. Maybe this is all already resolved, but it needs to be posted and tested with a current kernel.

Lukas, can you resend that patch? We can test it.

Alex

Quoted 24 lines
Quoted 4 lines
Quoted 4 lines

after merge for the kernel 6.2.0-rc5, and applied the alternative patch:
https://patchwork.kernel.org/project/linux-pci/patch/3dc88ea82bdc0e3
[email protected]/

This one is on track to appear in v6.3-rc1:
https://git.kernel.org/cgit/linux/kernel/git/pci/pci.git/commit/?id=74ff8864c
c84

Quoted 15 lines
Quoted 15 lines

I have uploaded the merged patch and the system log for the upstream kernel.

Anatoli

Quoted 8 lines · Lukas Wunner

You're now getting a different deadlock. That one is addressed by this old patch (it's already linked from the bugzilla):

https://lore.kernel.org/linux-pci/908047f7699d9de9ec2efd6b79aa752d
[email protected]/

If you apply that patch plus the new one, do you still see a deadlock?

#8 lore raw
Quoted 6 lines · Alexander
Quoted 6 lines

From: Bjorn Helgaas <[email protected]>

Quoted 4 lines · Anatoli Antonovitch
Quoted 2 lines · Anatoli Antonovitch

I do not see a deadlock, when applying the following old patch:
https://lore.kernel.org/linux-pci/908047f7699d9de9ec2efd6b79aa752d73dab4b6.1595329748.git.lukas@wunner.de/

Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

Can we revisit the patches again to get a fix?
The issue still reproduce and visible in the kernel 6.2.0-rc8.

This old patch would need to be updated and reposted. There was a 0-day bot issue and a question to be resolved. Maybe this is all already resolved, but it needs to be posted and tested with a current kernel.

Lukas, can you resend that patch? We can test it.

I'm working on a patch which aims to solve these deadlocks differently, by reducing the critical sections for which the reset_lock is held. Please stand by.

Thanks,

Lukas

#9 lore raw

Thanks Lukas.

The patch has been tested with current kernel 6.3.0-rc5 on the same setup. The deadlock between reset_lock and device_lock has been fixed.

See details in the dmesg log: dmesg_6.3.0-rc5_fix.txt in bugzilla:

https://bugzilla.kernel.org/show_bug.cgi?id=215590

Thanks,
Anatoli

Quoted 19 lines · Lukas Wunner
Quoted 11 lines · Alexander
Quoted 10 lines

From: Bjorn Helgaas <[email protected]>

Quoted 5 lines · Anatoli Antonovitch
Quoted 2 lines · Anatoli Antonovitch

I do not see a deadlock, when applying the following old patch:
https://lore.kernel.org/linux-pci/908047f7699d9de9ec2efd6b79aa752d73dab4b6.1595329748.git.lukas@wunner.de/

Can we revisit the patches again to get a fix?
The issue still reproduce and visible in the kernel 6.2.0-rc8.

This old patch would need to be updated and reposted. There was a 0-day bot issue and a question to be resolved. Maybe this is all already resolved, but it needs to be posted and tested with a current kernel.

Lukas, can you resend that patch? We can test it.

I'm working on a patch which aims to solve these deadlocks differently, by reducing the critical sections for which the reset_lock is held. Please stand by.

Thanks,

Lukas