Re: ath11k allocation failure on resume breaking wifi until power cycle

[Baochen Qiang]

On 2/23/2024 11:28 PM, Vlastimil Babka wrote:
> On 2/22/24 06:47, Manivannan Sadhasivam wrote:
> > On Wed, Feb 21, 2024 at 08:34:23AM -0800, Jeff Johnson wrote:
> > > On 2/21/2024 6:39 AM, Vlastimil Babka wrote:
> > > > Hi,
> > > >
> > > > starting with 6.8 rc series, I'm experiencing problems on resume from s2idle
> > > > on my laptop, which is Lenovo T14s Gen3:
> > > >
> > > > LENOVO 21CRS0K63K/21CRS0K63K, BIOS R22ET65W (1.35 )
> > > > ath11k_pci 0000:01:00.0: wcn6855 hw2.1
> > > > ath11k_pci 0000:01:00.0: chip_id 0x12 chip_family 0xb board_id 0xff soc_id 0x400c1211
> > > > ath11k_pci 0000:01:00.0: fw_version 0x1106196e fw_build_timestamp 2024-01-12 11:30 fw_build_id WLAN.HSP.1.1-03125-QCAHSPSWPL_V1_V2_SILICONZ_LITE-3.6510.37
> > > >
> > > > The problem is an allocation failure happening on resume from s2idle. After
> > > > that the wifi stops working and even a reboot won't fix it, only a
> > > > poweroff/poweron cycle of the laptop.
> >
> > Looks like WLAN is powered down during s2idle, which doesn't make sense. I hope
> > Jeff will figure out what's going on.
>
> You mean the firmware is supposed to power it down/up transparently without
> kernel involvement? Because it should be powered down to save the power, no?
Let me clarify: from backtrace info, seems you are using a kernel with 
the hibernation-support patches [1] applied, which are not accepted yet 
to mainline kernel or even 
git://git.kernel.org/pub/scm/linux/kernel/git/mani/mhi.git.

So this is why you see WLAN firmware is powered down during suspend.

[1] 
https://patchwork.kernel.org/project/linux-wireless/cover/20231127162022.518834-1-kvalo@xxxxxxxxxx/

> But I just found out that when I build my own kernel using the distro config
> as base but reduced by make localmodconfig, the "mhi mhi0: Requested to
> power ON" and related messages don't occur anymore, so there's something
> weird going on.
Here your own kernel doesn't include the hibernation-support patches, right?

> > But if you can share the dmesg after enabling the debug prints of both ath11k
> > and MHI, it will help a lot.
> >
> > - Mani
> >
> > > > This is order 4 (costly order), GFP_NOIO (maybe it's originally GFP_KERNEL
> > > > but we restrict to GFP_NOIO during resume) allocation, thus it's impossible
> > > > to do memory compaction and the page allocator gives up. Such high-order
> > > > allocations should have a fallback using smaller pages, or maybe it could at
> > > > least retry once the restricted GFP_NOIO context is gone.
> > > >
> > > > I don't know why it never happened before 6.8, didn't spot anything obvious
> > > > and it happens too unreliably to go bisect. Any idea?
> > >
> > > I've asked the development team to look at this, but in the interim can
> > > you apply the two hibernation patchsets to see if those cleanups also
> > > fix your problem:
> > >
> > > [PATCH 0/5] wifi: ath11k: prepare for hibernation support
> > > https://lore.kernel.org/linux-wireless/20240221024725.10057-1-quic_bqiang@xxxxxxxxxxx
> > >
> > > [PATCH 0/3] wifi: ath11k: hibernation support
> > > https://lore.kernel.org/linux-wireless/20240221030026.10553-1-quic_bqiang@xxxxxxxxxxx