Re: [PATCH] igb: Free IRQs when device is hotplugged

[Lyude Paul]

On Mon, 2017-12-11 at 16:34 -0800, Stephen Hemminger wrote:
> On Mon, 11 Dec 2017 18:45:02 -0500
> Lyude Paul <lyude@xxxxxxxxxx> wrote:
> 
> > Recently I got a Caldigit TS3 Thunderbolt 3 dock, and noticed that upon
> > hotplugging my kernel would immediately crash due to igb:
> > 
> > [  680.825801] kernel BUG at drivers/pci/msi.c:352!
> > [  680.828388] invalid opcode: 0000 [#1] SMP
> > [  680.829194] Modules linked in: igb(O) thunderbolt i2c_algo_bit joydev
> > vfat fat btusb btrtl btbcm btintel bluetooth ecdh_generic hp_wmi
> > sparse_keymap rfkill wmi_bmof iTCO_wdt intel_rapl x86_pkg_temp_thermal
> > coretemp crc32_pclmul snd_pcm rtsx_pci_ms mei_me snd_timer memstick snd
> > pcspkr mei soundcore i2c_i801 tpm_tis psmouse shpchp wmi tpm_tis_core tpm
> > video hp_wireless acpi_pad rtsx_pci_sdmmc mmc_core crc32c_intel serio_raw
> > rtsx_pci mfd_core xhci_pci xhci_hcd i2c_hid i2c_core [last unloaded: igb]
> > [  680.831085] CPU: 1 PID: 78 Comm: kworker/u16:1 Tainted:
> > G           O     4.15.0-rc3Lyude-Test+ #6
> > [  680.831596] Hardware name: HP HP ZBook Studio G4/826B, BIOS P71 Ver.
> > 01.03 06/09/2017
> > [  680.832168] Workqueue: kacpi_hotplug acpi_hotplug_work_fn
> > [  680.832687] RIP: 0010:free_msi_irqs+0x180/0x1b0
> > [  680.833271] RSP: 0018:ffffc9000030fbf0 EFLAGS: 00010286
> > [  680.833761] RAX: ffff8803405f9c00 RBX: ffff88033e3d2e40 RCX:
> > 000000000000002c
> > [  680.834278] RDX: 0000000000000000 RSI: 00000000000000ac RDI:
> > ffff880340be2178
> > [  680.834832] RBP: 0000000000000000 R08: ffff880340be1ff0 R09:
> > ffff8803405f9c00
> > [  680.835342] R10: 0000000000000000 R11: 0000000000000040 R12:
> > ffff88033d63a298
> > [  680.835822] R13: ffff88033d63a000 R14: 0000000000000060 R15:
> > ffff880341959000
> > [  680.836332] FS:  0000000000000000(0000) GS:ffff88034f440000(0000)
> > knlGS:0000000000000000
> > [  680.836817] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > [  680.837360] CR2: 000055e64044afdf CR3: 0000000001c09002 CR4:
> > 00000000003606e0
> > [  680.837954] Call Trace:
> > [  680.838853]  pci_disable_msix+0xce/0xf0
> > [  680.839616]  igb_reset_interrupt_capability+0x5d/0x60 [igb]
> > [  680.840278]  igb_remove+0x9d/0x110 [igb]
> > [  680.840764]  pci_device_remove+0x36/0xb0
> > [  680.841279]  device_release_driver_internal+0x157/0x220
> > [  680.841739]  pci_stop_bus_device+0x7d/0xa0
> > [  680.842255]  pci_stop_bus_device+0x2b/0xa0
> > [  680.842722]  pci_stop_bus_device+0x3d/0xa0
> > [  680.843189]  pci_stop_and_remove_bus_device+0xe/0x20
> > [  680.843627]  trim_stale_devices+0xf3/0x140
> > [  680.844086]  trim_stale_devices+0x94/0x140
> > [  680.844532]  trim_stale_devices+0xa6/0x140
> > [  680.845031]  ? get_slot_status+0x90/0xc0
> > [  680.845536]  acpiphp_check_bridge.part.5+0xfe/0x140
> > [  680.846021]  acpiphp_hotplug_notify+0x175/0x200
> > [  680.846581]  ? free_bridge+0x100/0x100
> > [  680.847113]  acpi_device_hotplug+0x8a/0x490
> > [  680.847535]  acpi_hotplug_work_fn+0x1a/0x30
> > [  680.848076]  process_one_work+0x182/0x3a0
> > [  680.848543]  worker_thread+0x2e/0x380
> > [  680.848963]  ? process_one_work+0x3a0/0x3a0
> > [  680.849373]  kthread+0x111/0x130
> > [  680.849776]  ? kthread_create_worker_on_cpu+0x50/0x50
> > [  680.850188]  ret_from_fork+0x1f/0x30
> > [  680.850601] Code: 43 14 85 c0 0f 84 d5 fe ff ff 31 ed eb 0f 83 c5 01 39
> > 6b 14 0f 86 c5 fe ff ff 8b 7b 10 01 ef e8 b7 e4 d2 ff 48 83 78 70 00 74 e3
> > <0f> 0b 49 8d b5 a0 00 00 00 e8 62 6f d3 ff e9 c7 fe ff ff 48 8b
> > [  680.851497] RIP: free_msi_irqs+0x180/0x1b0 RSP: ffffc9000030fbf0
> > 
> > As it turns out, normally the freeing of IRQs that would fix this is called
> > inside of the scope of __igb_close(). However, since the device is
> > already gone by the point we try to unregister the netdevice from the
> > driver due to a hotplug we end up seeing that the netif isn't present
> > and thus, forget to free any of the device IRQs.
> > 
> > So: after unregistering the netdev in igb_remove() check whether the PCI
> > device is stale and if so, free it's IRQs and tx/rx resources.
> > 
> > Signed-off-by: Lyude Paul <lyude@xxxxxxxxxx>
> > Fixes: 9474933caf21 ("igb: close/suspend race in netif_device_detach")
> > Cc: Todd Fujinaka <todd.fujinaka@xxxxxxxxx>
> > Cc: stable@xxxxxxxxxxxxxxx
> > ---
> >  drivers/net/ethernet/intel/igb/igb_main.c | 10 ++++++++++
> >  1 file changed, 10 insertions(+)
> > 
> > diff --git a/drivers/net/ethernet/intel/igb/igb_main.c
> > b/drivers/net/ethernet/intel/igb/igb_main.c
> > index c208753ff5b7..e650348b4bd7 100644
> > --- a/drivers/net/ethernet/intel/igb/igb_main.c
> > +++ b/drivers/net/ethernet/intel/igb/igb_main.c
> > @@ -3325,6 +3325,16 @@ static void igb_remove(struct pci_dev *pdev)
> >  
> >  	unregister_netdev(netdev);
> >  
> > +	/* If the PCI device has already been physically removed (e.g. user
> > +	 * unplugged a thunderbolt dock containing our hw) then the netif
> > will
> > +	 * already be down, so unregistering the netdev won't free the IRQs
> > +	 */
> > +	if (!pci_device_is_present(pdev)) {
> > +		igb_free_irq(adapter);
> > +		igb_free_all_tx_resources(adapter);
> > +		igb_free_all_rx_resources(adapter);
> > +	}
> > +
> >  	igb_clear_interrupt_scheme(adapter);
> >  
> >  	pci_iounmap(pdev, adapter->io_addr);
> 
> This looks like you are making a special case out of something that should
> be fixed in igb_close instead.
> 
> Most likely something is wrong with earlier commit.
> 
> commit 9474933caf21a4cb5147223dca1551f527aaac36
> Author: Todd Fujinaka <todd.fujinaka@xxxxxxxxx>
> Date:   Tue Nov 15 08:54:26 2016 -0800
> 
>     igb: close/suspend race in netif_device_detach
>     
>     Similar to ixgbe, when an interface is part of a namespace it is
>     possible that igb_close() may be called while __igb_shutdown() is
>     running which ends up in a double free WARN and/or a BUG in
>     free_msi_irqs().
>     
>     Extend the rtnl_lock() to protect the call to netif_device_detach() and
>     igb_clear_interrupt_scheme() in __igb_shutdown() and check for
>     netif_device_present() to avoid calling igb_clear_interrupt_scheme() a
>     second time in igb_close().
>     
>     Also extend the rtnl lock in igb_resume() to netif_device_attach().
>     
>     Signed-off-by: Todd Fujinaka <todd.fujinaka@xxxxxxxxx>
>     Acked-by: Alexander Duyck <alexander.h.duyck@xxxxxxxxx>
>     Tested-by: Aaron Brown <aaron.f.brown@xxxxxxxxx>
>     Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@xxxxxxxxx>
> 
> 
> Layers on layers of workarounds does not lead to stability.'

Agreed, although it should be noted though I mentioned that exact commit in the
patch I sent :P. Will rework patch and send out a new version.

>