Message-ID: <337ad5f0-bd1b-4dbf-a94b-681a21f0df45@gmail.com>
Date: Wed, 13 Mar 2024 16:29:48 -0700
From: Florian Fainelli <f.fainelli@...il.com>
To: "Russell King (Oracle)" <linux@...linux.org.uk>
Cc: Joel Fernandes <joel@...lfernandes.org>, Boqun Feng
 <boqun.feng@...il.com>, Anna-Maria Behnsen <anna-maria@...utronix.de>,
 Linus Torvalds <torvalds@...ux-foundation.org>,
 linux-kernel@...r.kernel.org, kernel-team@...a.com, paulmck@...nel.org,
 mingo@...nel.org, tglx@...utronix.de, rcu@...r.kernel.org,
 neeraj.upadhyay@....com, urezki@...il.com, qiang.zhang1211@...il.com,
 frederic@...nel.org, bigeasy@...utronix.de, chenzhongjin@...wei.com,
 yangjihong1@...wei.com, rostedt@...dmis.org,
 Justin Chen <justin.chen@...adcom.com>
Subject: Re: Unexplained long boot delays [Was Re: [GIT PULL] RCU changes for
 v6.9]

On 3/13/24 15:49, Russell King (Oracle) wrote:
> On Wed, Mar 13, 2024 at 03:04:26PM -0700, Florian Fainelli wrote:
>> On 3/13/24 14:59, Russell King (Oracle) wrote:
>>> On Wed, Mar 13, 2024 at 02:30:43PM -0700, Florian Fainelli wrote:
>>>> I will try to provide multiple answers for the sake of everyone having the
>>>> same context. Responding to Linus' specifically and his suggestion to use
>>>> "initcall_debug", this is what it gave me:
>>>>
>>>> [    6.970669] ata1: SATA link down (SStatus 0 SControl 300)
>>>> [  166.136366] probe of unimac-mdio-0:01 returned 0 after 159216218 usecs
>>>> [  166.142931] unimac-mdio unimac-mdio.0: Broadcom UniMAC MDIO bus
>>>> [  166.148900] probe of unimac-mdio.0 returned 0 after 159243553 usecs
>>>> [  166.155820] probe of f0480000.ethernet returned 0 after 159258794 usecs
>>>> [  166.166427] ehci-brcm f0b00300.ehci_v2: EHCI Host Controller
>>>>
>>>> Also got another occurrence happening resuming from suspend to DRAM with:
>>>>
>>>> [   22.570667] brcmstb-dpfe 9932000.dpfe-cpu: PM: calling
>>>> platform_pm_resume+0x0/0x54 @ 1574, parent: rdb
>>>> [  181.643809] brcmstb-dpfe 9932000.dpfe-cpu: PM:
>>>> platform_pm_resume+0x0/0x54 returned 0 after 159073134 usecs
>>>>
>>>> and also with the PCIe root complex driver:
>>>>
>>>> [   18.266279] brcm-pcie f0460000.pcie: PM: calling
>>>> brcm_pcie_resume_noirq+0x0/0x164 @ 1597, parent: platform
>>>> [  177.457219] brcm-pcie f0460000.pcie: clkreq-mode set to default
>>>> [  177.457225] brcm-pcie f0460000.pcie: link up, 2.5 GT/s PCIe x1 (!SSC)
>>>> [  177.457231] brcm-pcie f0460000.pcie: PM: brcm_pcie_resume_noirq+0x0/0x164
>>>> returned 0 after 159190939 usecs
>>>> [  177.457257] pcieport 0000:00:00.0: PM: calling
>>>> pci_pm_resume_noirq+0x0/0x160 @ 33, parent: pci0000:00
>>>>
>>>> Surprisingly those drivers are consistently reproducing the failures I am
>>>> seeing so at least this gave me a clue as to where the problem is.
>>>>
>>>> There were no changes to drivers/net/ethernet/broadcom/genet/, the two
>>>> changes done to drivers/net/mdio/mdio-bcm-unimac.c are correct, especially
>>>> the read_poll_timeout() conversion is correct, we properly break out of the
>>>> loop. The initial delay looked like a good culprit for a little while, but
>>>> it is not used on the affected platforms because instead we provide a
>>>> callback and we have an interrupt to signal the completion of a MDIO
>>>> operation, therefore unimac_mdio_poll() is not used at all. Finally
>>>> drivers/memory/brcmstb_dpfe.c also received a single change which is not
>>>> functional here (.remove function change do return void).
>>>>
>>>> I went back to a manual bisection and this time I believe that I have a more
>>>> plausible candidate with:
>>>>
>>>> 7ee988770326fca440472200c3eb58935fe712f6 ("timers: Implement the
>>>> hierarchical pull model")
>>>
>>> I haven't understood the code there yet, and how it would interact with
>>> arch code, but one thing that immediately jumps out to me is this:
>>>
>>> "    As long as a CPU is busy it expires both local and global timers. When a
>>>       CPU goes idle it arms for the first expiring local timer."
>>>
>>> So are local timers "armed" when they are enqueued while the cpu is
>>> "busy" during initialisation, and will they expire, and will that
>>> expiry be delivered in a timely manner?
>>>
>>> If not, this commit is basically broken, and would be the cause of the
>>> issue you are seeing. For the mdio case, we're talking about 2ms
>>> polling. For the dpfe case, it looks like we're talking about 1ms
>>> sleeps. I'm guessing that these end up being local timers.
>>>
>>> Looking at pcie-brcmstb, there's a 100ms msleep(), and then a polling
>>> for link up every 5ms - if the link was down and we msleep(5) I wonder
>>> if that's triggering the same issue.
>>>
>>> Why that would manifest itself on 32-bit but not 64-bit Arm, I can't
>>> say. I would imagine that the same hardware timer driver is being used
>>> (may be worth checking DT.) The same should be true for the interrupt
>>> driver as well. There's been no changes in that code.
>>
>> I just had it happen with ARM64 I was plagued by:
>>
>> https://lore.kernel.org/lkml/87wmqrjg8n.fsf@somnus/T/
>>
>> and my earlier bisections somehow did not have ARM64 fail, so I thought it
>> was immune but it fails with about the same failure rate as ARM 32-bit.
> 
> Okay, so if that's two architectures failing, it's probably not an
> architecture specific issue. It could still be a timer/interrupt driver
> issue though. It would be nice to have a reproducer on x86 which would
> confirm it.
> 
> Do you know the frequency that the failure happens?

On this specific commit 7ee988770326fca440472200c3eb58935fe712f6, there 
is a 100% failure for at least 3 devices out of the 16 that are running 
the test.
-- 
Florian

Powered by blists - more mailing lists