Re: [PATCH] pci/quirks: Add quirk to reset nvgpu at boot for the Lenovo ThinkPad P50

From: Lyude Paul
Date: Wed Mar 13 2019 - 18:25:10 EST


[note to David Ober: you -should- be able to reply to this, hopefully, but I
haven't actually tested that so results may vary]

Hi again! Sorry I didn't fully answer all of the questions you originally
asked in this email, I had to get in contact with Lenovo to make sure that it
was OK for me to disclose more details on this bug (and I had PTO scheduled
immediately after I asked). I've added David Ober from Lenovo to this thread
as well. So now that I've got Lenovo's approval I can answer those questions,
and give some better answers for the others! (see below)


On Fri, 2019-02-15 at 16:17 -0500, Lyude Paul wrote:
> On Thu, 2019-02-14 at 18:43 -0600, Bjorn Helgaas wrote:
> > Hi Lyude,
> >
> > On Tue, Feb 12, 2019 at 05:02:30PM -0500, Lyude Paul wrote:
> > > On a very specific subset of ThinkPad P50 SKUs, particularly ones that
> > > come with a Quadro M1000M chip instead of the M2000M variant, the BIOS
> > > seems to have a very nasty habit of not always resetting the secondary
> > > Nvidia GPU between full reboots if the laptop is configured in Hybrid
> > > Graphics mode. The reason for this happening is unknown, but the
> > > following steps and possibly a good bit of patience will reproduce the
> > > issue:
> > >
> > > 1. Boot up the laptop normally in Hybrid graphics mode
> > > 2. Make sure nouveau is loaded and that the GPU is awake
> > > 2. Allow the nvidia GPU to runtime suspend itself after being idle
> > > 3. Reboot the machine, the more sudden the better (e.g sysrq-b may help)
> > > 4. If nouveau loads up properly, reboot the machine again and go back to
> > > step 2 until you reproduce the issue
> > >
> > > This results in some very strange behavior: the GPU will
> > > quite literally be left in exactly the same state it was in when the
> > > previously booted kernel started the reboot. This has all sorts of bad
> > > sideaffects: for starters, this completely breaks nouveau starting with
> > > a
> > > mysterious EVO channel failure that happens well before we've actually
> > > used the EVO channel for anything:
> >
> > There are a lot of moving parts here that are probably obvious to you
> > but not to me. I need help untangling this a bit so I'm comfortable
> > that we got to the root cause and that we're doing something logical
> > as opposed to something that just happens to make things work. I
> > really don't know enough to even ask the right questions...
>
> I completely understand! I'm pretty sure I'd be just as skeptical if I was
> in
> your position reviewing a patch like this :P
>
> > Is there a bug report for this? Bugzilla.kernel.org would be ideal,
> > including "lspci -vvxxx" and dmidecode for the system.
> >
> Not yet, but there has been discussion about this between nouveau developers
> on our IRC channel.
I lied: yes there actually is a bug report for this, but it's currently on the
Red Hat bugzilla. I can get more information from it if you need (with
lenovo's approval of course).

>
> > Is this running a current BIOS? The date in your log below looks
> > pretty recent, so I assume it is current.
>
> Yes, this is the most up to date BIOS available for this system.

And additionally: I've been working with Lenovo on this issue for a couple of
months now, and we've gone through dozens of different trial BIOSes with no
success thus far. However, Lenovo is currently working on trying to add this
workaround into their BIOS but I've been told that this change is going to
take a decent amount of time since they need to test it across multiple
operating systems. I'd be happy to come back and add a conditional later to
turn this workaround off for later BIOS versions once Lenovo has released a
proper fix.

With all of that being said, [how] do you think we should proceed?

>
> > I assume "hybrid graphics" means you have two GPUs. Do you select
> > hybrid graphics mode in the BIOS?
>
> Yes, the P50 has two available modes in the BIOS: Dedicated (e.g. only
> the nvidia GPU is used for everything), and Hybrid (i915 drives the
> built-in display panel, nouveau drives everything else). This bug only
> seems to occur in Hybrid mode.
>
> > I assume when you say the Nvidia GPU doesn't get reset on a full
> > reboot, you're talking about a "warm reboot", and that if you actually
> > remove the power and do a cold reboot, there's no problem?
>
> If you meant "unplugging the power adapter" when you said cutting the
> power we don't need to go that far, but shutting down the machine and
> restarting it by hand does avoid the problem yes.
>
> > I assume Nvidia GPU being active means you are using the performance
> > GPU. Does that mean the integrated GPU is completely unused and Linux
> > does nothing at all with it? Is Linux doing any switching between
> > them? If so, how? I am not 100% confident in the code I've seen that
> > does the switching.
>
> "Switching" isn't really the right word to describe it these days as the
> process for how this is handled has changed quite a bit in the last few
> years. As I mentioned above, the main intel GPU is used for driving the
> built-in display. As for the nvidia GPU, it's used to drive any kind of
> external displays connected to the P50 and can also be used to do
> rendering through DRI PRIME. The connector hookups are mostly hardcoded
> as there isn't really much of a mux, unlike the old days where we could
> actually do things like switch which GPU was driving the internal
> display on the fly with vga-switcheroo using the _DSM methods provided
> by ACPI.
>
> Nowadays muxless laptops like the ThinkPad P50 don't make much use of
> _DSM for power management and instead just runtime suspend the dedicated
> nvidia GPU when it's not in use. We rely on the PCI subsystem to invoke
> the _PR3 ACPI methods for suspending and resuming the GPU and don't
> handle much of it ourselves anymore:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpu/drm/nouveau/nouveau_drm.c?h=v5.0-rc6#n860
>
> (You can ignore the call to nouveau_switcheroo_optimus_dsm() there; it
> doesn't do anything on systems like the P50 where we detect _PR3)
>
> Additionally-if you see any problems with how we're handling things
> there we're all ears! We've been trying to fix as many issues with
> runtime PM as we can find.
>
> Something else to consider here too: I've only ever managed to reproduce
> this issue on this specific P50 SKU. Since I work on Desktop engineering
> at Red Hat we have access to a very large number of laptops with hybrid
> GPU setups. Interestingly I haven't ever reproduced anything like this
> on any of them, even P50 SKUs with the M2000M GPU instead of the M1000M
> (which are almost identical, they're both GM107 chips) I can follow the
> same steps that I followed to reproduce this bug and it never appears.
>
> Another thing to note: this bug also doesn't appear to happen 100% of
> the time. Some warm reboots the GPU will be reset by the BIOS perfectly
> fine.
>
> > > nouveau 0000:01:00.0: disp: chid 0 mthd 0000 data 00000400 00001000
> > > 00000002
> > >
> > > Later on, this causes us to timeout trying to bring up the GR ctx:
> > >
> > > ------------[ cut here ]------------
> > > nouveau 0000:01:00.0: timeout
> > > WARNING: CPU: 0 PID: 12 at
> > > drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf100.c:1547
> > > gf100_grctx_generate+0x7b2/0x850 [nouveau]
> > > Modules linked in: nouveau mxm_wmi i915 crc32c_intel ttm i2c_algo_bit
> > > serio_raw drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops
> > > xhci_pci drm xhci_hcd i2c_core wmi video
> > > CPU: 0 PID: 12 Comm: kworker/0:1 Not tainted 5.0.0-rc5Lyude-Test+ #29
> > > Hardware name: LENOVO 20EQS64N0B/20EQS64N0B, BIOS N1EET82W (1.55 )
> > > 12/18/2018
> > > Workqueue: events_long drm_dp_mst_link_probe_work [drm_kms_helper]
> > > RIP: 0010:gf100_grctx_generate+0x7b2/0x850 [nouveau]
> > > Code: 85 d2 75 04 48 8b 57 10 48 89 95 28 ff ff ff e8 b4 37 0e e1 48 8b
> > > 95 28 ff ff ff 48 c7 c7 b1 97 57 a0 48 89 c6 e8 5a 38 c0 e0 <0f> 0b e9
> > > b9 fd ff ff 48 8b 85 60 ff ff ff 48 8b 40 10 48 8b 78 10
> > > RSP: 0018:ffffc900000b77f0 EFLAGS: 00010286
> > > RAX: 0000000000000000 RBX: ffff888871af8000 RCX: 0000000000000000
> > > RDX: ffff88887f41dfe0 RSI: ffff88887f415698 RDI: ffff88887f415698
> > > RBP: ffffc900000b78c8 R08: 0000000000000000 R09: 0000000000000000
> > > R10: 0000000000000000 R11: 0000000000000000 R12: ffff888872118000
> > > R13: 0000000000000000 R14: ffffffffa0551420 R15: ffffc900000b7818
> > > FS: 0000000000000000(0000) GS:ffff88887f400000(0000)
> > > knlGS:0000000000000000
> > > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > CR2: 00005644d0556ca8 CR3: 0000000002214006 CR4: 00000000003606f0
> > > DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > > DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> > > Call Trace:
> > > gf100_gr_init_ctxctl+0x27b/0x2d0 [nouveau]
> > > gf100_gr_init+0x5bd/0x5e0 [nouveau]
> > > gf100_gr_init_+0x61/0x70 [nouveau]
> > > nvkm_gr_init+0x1d/0x20 [nouveau]
> > > nvkm_engine_init+0xcb/0x210 [nouveau]
> > > nvkm_subdev_init+0xd6/0x230 [nouveau]
> > > nvkm_engine_ref.part.0+0x52/0x70 [nouveau]
> > > nvkm_engine_ref+0x13/0x20 [nouveau]
> > > nvkm_ioctl_new+0x12c/0x260 [nouveau]
> > > ? nvkm_fifo_chan_child_del+0xa0/0xa0 [nouveau]
> > > ? gf100_gr_dtor+0xe0/0xe0 [nouveau]
> > > nvkm_ioctl+0xe2/0x180 [nouveau]
> > > nvkm_client_ioctl+0x12/0x20 [nouveau]
> > > nvif_object_ioctl+0x47/0x50 [nouveau]
> > > nvif_object_init+0xc8/0x120 [nouveau]
> > > nvc0_fbcon_accel_init+0x5c/0x960 [nouveau]
> > > nouveau_fbcon_create+0x5a5/0x5d0 [nouveau]
> > > ? drm_setup_crtcs+0x27b/0xcb0 [drm_kms_helper]
> > > ? __lock_is_held+0x5e/0xa0
> > > __drm_fb_helper_initial_config_and_unlock+0x27c/0x520 [drm_kms_helper]
> > > drm_fb_helper_hotplug_event.part.29+0xae/0xc0 [drm_kms_helper]
> > > drm_fb_helper_hotplug_event+0x1c/0x30 [drm_kms_helper]
> > > nouveau_fbcon_output_poll_changed+0xb8/0x110 [nouveau]
> > > drm_kms_helper_hotplug_event+0x2a/0x40 [drm_kms_helper]
> > > drm_dp_send_link_address+0x176/0x1c0 [drm_kms_helper]
> > > drm_dp_check_and_send_link_address+0xa0/0xb0 [drm_kms_helper]
> > > drm_dp_mst_link_probe_work+0xa4/0xc0 [drm_kms_helper]
> > > process_one_work+0x22f/0x5c0
> > > worker_thread+0x44/0x3a0
> > > kthread+0x12b/0x150
> > > ? wq_pool_ids_show+0x140/0x140
> > > ? kthread_create_on_node+0x60/0x60
> > > ret_from_fork+0x3a/0x50
> > > irq event stamp: 22490
> > > hardirqs last enabled at (22489): [<ffffffff8113281d>]
> > > console_unlock+0x44d/0x5f0
> > > hardirqs last disabled at (22490): [<ffffffff81001c03>]
> > > trace_hardirqs_off_thunk+0x1a/0x1c
> > > softirqs last enabled at (22486): [<ffffffff81c00330>]
> > > __do_softirq+0x330/0x44d
> > > softirqs last disabled at (22479): [<ffffffff810c3105>]
> > > irq_exit+0xe5/0xf0
> > > WARNING: CPU: 0 PID: 12 at
> > > drivers/gpu/drm/nouveau/nvkm/engine/gr/ctxgf100.c:1547
> > > gf100_grctx_generate+0x7b2/0x850 [nouveau]
> > > ---[ end trace bf0976ed88b122a8 ]---
> > > nouveau 0000:01:00.0: gr: wait for idle timeout (en: 1, ctxsw: 0, busy:
> > > 1)
> > > nouveau 0000:01:00.0: gr: wait for idle timeout (en: 1, ctxsw: 0, busy:
> > > 1)
> > > nouveau 0000:01:00.0: fifo: fault 01 [WRITE] at 0000000000008000 engine
> > > 00 [GR] client 15 [HUB/SCC_NB] reason c4 [] on channel -1 [0000000000
> > > unknown]
> > >
> > > From which the GPU never manages to recover. Booting without nouveau
> > > loading causes issues as well, since the GPU starts sending spurious
> > > interrupts that cause other device's IRQs to get disabled by the kernel:
> > >
> > > irq 16: nobody cared (try booting with the "irqpoll" option)
> > > â
> > > handlers:
> > > [<000000007faa9e99>] i801_isr [i2c_i801]
> > > Disabling IRQ #16
> > > â
> > > serio: RMI4 PS/2 pass-through port at rmi4-00.fn03
> > > i801_smbus 0000:00:1f.4: Timeout waiting for interrupt!
> > > i801_smbus 0000:00:1f.4: Transaction timeout
> > > rmi4_f03 rmi4-00.fn03: rmi_f03_pt_write: Failed to write to F03 TX
> > > register (-110).
> > > i801_smbus 0000:00:1f.4: Timeout waiting for interrupt!
> > > i801_smbus 0000:00:1f.4: Transaction timeout
> > > rmi4_physical rmi4-00: rmi_driver_set_irq_bits: Failed to change enabled
> > > interrupts!
> > >
> > > Which in turn causes the touchpad and sometimes even other things to get
> > > disabled.
> > >
> > > Since the GPU staying on causes problems even without nouveau's
> > > intervention, we can't fix this problem from nouveau itself. We have to
> > > fix it as early as possible in the boot sequence in order to make sure
> > > that the GPU is in a clean state before it has a chance to spam us with
> > > interrupts and break things.
> >
> > Was nouveau loaded *before* the reboot? Or can you reproduce the
> > spurious interrupts even if you do a cold reboot without nouveau, then
> > a warm reboot again without nouveau?
>
> Nouveau was loaded before the reboot, yes, and as far as I'm aware this
> bug won't show up following the steps you described here. We wouldn't
> really notice the interrupts if this bug happened without nouveau loaded
> (since the BIOS doesn't arm them). However, I can confirm it's being
> reset each on each warm reboot when nouveau's completely disabled by
> sticking values into some MMIO registers we know the VBIOS doesn't touch
> followed by rebooting in order to see if they retain their value (this
> is how I discovered that the GPU wasn't being power cycled in the first
> place, if the values I stick into said registers gets cleared after
> reboot then the GPU was reset, otherwise it wasn't).
>
>
> > > So to do this, we add a new pci quirk using
> > > DECLARE_PCI_FIXUP_CLASS_FINAL that will be invoked before the PCI probe
> > > at boot finishes. From there, we check to make sure that this is indeed
> > > the specific P50 variant of this GPU. We also make sure that the GPU PCI
> > > device is advertising NoReset- in order to prevent us from trying to
> > > reset the GPU when the machine is in Dedicated graphics mode (where the
> > > GPU being initialized by the BIOS is normal and expected). Finally, we
> > > try mapping the MMIO space for the GPU which should only work if the GPU
> > > is actually active in D0 mode. We can then read the magic 0x2240c
> > > register on the GPU, which will have bit 1 set if the GPU's firmware has
> > > already been posted during a previous boot. Once we've confirmed all of
> > > this, we reset the PCI device and re-disable it - bringing the GPU back
> > > into a healthy state.
> > >
> > > Signed-off-by: Lyude Paul <lyude@xxxxxxxxxx>
> > > Cc: nouveau@xxxxxxxxxxxxxxxxxxxxx
> > > Cc: dri-devel@xxxxxxxxxxxxxxxxxxxxx
> > > Cc: Karol Herbst <kherbst@xxxxxxxxxx>
> > > Cc: Ben Skeggs <skeggsb@xxxxxxxxx>
> > > Cc: stable@xxxxxxxxxxxxxxx
> > > ---
> > > drivers/pci/quirks.c | 65 ++++++++++++++++++++++++++++++++++++++++++++
> > > 1 file changed, 65 insertions(+)
> > >
> > > diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
> > > index b0a413f3f7ca..948492fda8bf 100644
> > > --- a/drivers/pci/quirks.c
> > > +++ b/drivers/pci/quirks.c
> > > @@ -5117,3 +5117,68 @@ SWITCHTEC_QUIRK(0x8573); /* PFXI 48XG3 */
> > > SWITCHTEC_QUIRK(0x8574); /* PFXI 64XG3 */
> > > SWITCHTEC_QUIRK(0x8575); /* PFXI 80XG3 */
> > > SWITCHTEC_QUIRK(0x8576); /* PFXI 96XG3 */
> > > +
> > > +/*
> > > + * On certain Lenovo Thinkpad P50 SKUs, specifically those with a
> > > Nvidia
> > > + * Quadro M1000M, the BIOS will occasionally make the mistake of not
> > > resetting
> > > + * the nvidia GPU between reboots if the system is configured to use
> > > hybrid
> > > + * graphics mode. This results in the GPU being left in whatever state
> > > it
> > > was
> > > + * in during the previous boot which causes spurious interrupts from
> > > the
> > > GPU,
> > > + * which in turn cause us to disable the wrong IRQs and end up breaking
> > > the
> > > + * touchpad. Unsurprisingly, this also completely breaks nouveau.
> > > + *
> > > + * Luckily, it seems a simple reset of the PCI device for the nvidia
> > > GPU
> > > + * manages to bring the GPU back into a clean state and fix all of
> > > these
> > > + * issues. Additionally since the GPU will report NoReset+ when the
> > > machine is
> > > + * configured in Dedicated display mode, we don't need to worry about
> > > + * accidentally resetting the GPU when it's supposed to already be
> > > + * initialized.
> > > + */
> > > +static void
> > > +quirk_lenovo_thinkpad_p50_nvgpu_survives_reboot(struct pci_dev *pdev)
> > > +{
> > > + void __iomem *map;
> > > + int ret;
> > > +
> > > + if (pdev->subsystem_vendor != PCI_VENDOR_ID_LENOVO ||
> > > + pdev->subsystem_device != 0x222e ||
> > > + !pdev->reset_fn)
> > > + return;
> > > +
> > > + /*
> > > + * If we can't enable the device's mmio space, it's probably not even
> > > + * initialized. This is fine, and means we can just skip the quirk
> > > + * entirely.
> > > + */
> > > + if (pci_enable_device_mem(pdev)) {
> > > + pci_dbg(pdev, "Can't enable device mem, no reset needed\n");
> > > + return;
> > > + }
> > > +
> > > + /* Taken from drivers/gpu/drm/nouveau/engine/device/base.c */
> > > + map = ioremap(pci_resource_start(pdev, 0), 0x102000);
> > > + if (!map) {
> > > + pci_err(pdev, "Can't map MMIO space, this is probably very
> > > bad\n");
> > > + goto out_disable;
> > > + }
> > > +
> > > + /*
> > > + * Be extra careful, and make sure that the GPU firmware is posted
> > > + * before trying a reset
> > > + */
> > > + if (ioread32(map + 0x2240c) & 0x2) {
> > > + pci_info(pdev,
> > > + FW_BUG "GPU left initialized by EFI, resetting\n");
> > > + ret = pci_reset_function(pdev);
> > > + if (ret < 0)
> > > + pci_err(pdev, "Failed to reset GPU: %d\n", ret);
> > > + }
> > > +
> > > + iounmap(map);
> > > +out_disable:
> > > + pci_disable_device(pdev);
> > > +}
> > > +
> > > +DECLARE_PCI_FIXUP_CLASS_FINAL(PCI_VENDOR_ID_NVIDIA, 0x13b1,
> > > + PCI_CLASS_DISPLAY_VGA, 8,
> > > + quirk_lenovo_thinkpad_p50_nvgpu_survives_reboot)
> > > ;
> > > --
> > > 2.20.1
> > >
--
Cheers,
Lyude Paul