Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Mon, 26 Sep 2022 12:32:55 +0200
From:   Petr Mladek <pmladek@suse.com>
To:     Conor Dooley <conor@kernel.org>
Cc:     senozhatsky@chromium.org, rostedt@goodmis.org,
        john.ogness@linutronix.de, linux-kernel@vger.kernel.org
Subject: Re: [resend][bug] low-probability console lockups since 5.19
Message-ID: <YzF/1wlROxiK4XvD@alley>
References: <Yy3dsSgpDnovBdEn@spud>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <Yy3dsSgpDnovBdEn@spud>
Precedence: bulk

On Fri 2022-09-23 17:24:17, Conor Dooley wrote:
> Hey all,
> 
> Been bisecting a bug that is causing a boot failure in my CI & have
> ended up here.. The bug in question is a low(ish) probability lock up
> of the serial console, I would estimate about 1-in-5 chance on the
> boards I could actually trigger it on which it has taken me so long
> to realise that this was an actual problem. Thinking back on it, there
> were other failures that I would retroactively attribute to this
> problem too, but I had earlycon disabled and I chalked the hangs at
> "Starting kernel" down to pre-release hardware since the other boots
> in a CI job would not fail (we boot the same config multiple times),
> Silly me I guess!
> 
> In a boot failure (with earlycon & keep_bootcon enabled) I see the
> likes of [0]:
> [    0.485063] NFS: Registering the id_resolver key type
> [    0.491051] Key type id_resolver registered
> [    0.495743] Key type id_legacy registered
> [    0.500528] nfs4filelayout_init: NFSv4 File Layout Driver Registering...
> [    0.508010] nfs4flexfilelayout_init: NFSv4 Flexfile Layout Driver Registering...
> [    0.516857] 9p: Installing v9fs 9p2000 file system support
> [    0.524032] NET: Registered PF_ALG protocol family
> [    0.529562] Block layer SCSI generic (bsg) driver version 0.4 loaded (major 247)
> [    0.537817] io scheduler mq-deadline registered
> [    0.542889] io scheduler kyber registered
> [    0.559682] microchip-pcie 2000000000.pcie: host bridge /soc/pcie@2000000000 ranges:
> [    0.568431] microchip-pcie 2000000000.pcie:      MEM 0x2008000000..0x2087ffffff -> 0x0008000000
> [    0.635633] microchip-pcie 2000000000.pcie: axi read request error
> [    0.642602] microchip-pcie 2000000000.pcie: axi read timeout
> [    0.776630] microchip-pcie 2000000000.pcie: sec error in axi2pcie buffer
> [    0.784133] microchip-pcie 2000000000.pcie: axi write request error
> [    0.791100] microchip-pcie 2000000000.pcie: sec error in axi2pcie buffer
> [    0.798543] microchip-pcie 2000000000.pcie: sec error in axi2pcie buffer
> [    0.805981] microchip-pcie 2000000000.pcie: sec error in axi2pcie buffer
> [    0.813421] microchip-pcie 2000000000.pcie: sec error in axi2pcie buffer
> [    0.820859] microchip-pcie 2000000000.pcie: sec error in axi2pcie buffer
> [    0.828295] microchip-pcie 2000000000.pcie: sec error in axi2pcie buffer
> [    1.579668] Freeing initrd memory: 7316K
> 
> The sec error stuff is unrelated & happens on good boots too - although
> there are far fewer lines of "sec error" there [1] (excuse the message
> duplication from the bootcon).
> 
> I was able to bisect the behaviour to this set of reverts, but I have
> low confidence as to which commit specifically caused it - it's hard
> to know that doing 20+ reboots on every commmit I built is enough to
> make sure I did not make a mistake. I had initially started the
> bisection between v6.0-rc6 & v6.0-rc1, and that lead me down a sub-tree
> of net which is why the bisect log starts at a "random" commit as that
> merge commit was the one blamed by the inital bisection.
> 
> git bisect start
> # status: waiting for both good and bad commits
> # bad: [5e8379351dbde61ea383e514f0f9ecb2c047cf4e] Merge tag 'net-5.19-rc5' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net
> git bisect bad 5e8379351dbde61ea383e514f0f9ecb2c047cf4e
> # good: [4b0986a3613c92f4ec1bdc7f60ec66fea135991f] Linux 5.18
> git bisect good 4b0986a3613c92f4ec1bdc7f60ec66fea135991f
> # good: [fbe86daca0ba878b04fa241b85e26e54d17d4229] Merge tag 'scsi-misc' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
> git bisect good fbe86daca0ba878b04fa241b85e26e54d17d4229
> # good: [2b8c612c6102f751e6e3e1bd425f64e9d3d3f638] kernel/reboot: Fix powering off using a non-syscall code paths
> git bisect good 2b8c612c6102f751e6e3e1bd425f64e9d3d3f638
> # good: [54c2cc79194c961a213c1d375fe3aa4165664cc4] Merge tag 'usb-5.19-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/usb
> git bisect good 54c2cc79194c961a213c1d375fe3aa4165664cc4
> # skip: [8f7ac50c97d30ae5ae48da3b516510d05a67b9e5] Merge tag 'sound-5.19-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai/sound
> git bisect skip 8f7ac50c97d30ae5ae48da3b516510d05a67b9e5
> # good: [cdd85786f4b3b9273e4376e69aa95a2d71722764] x86/tdx: Clarify RIP adjustments in #VE handler
> git bisect good cdd85786f4b3b9273e4376e69aa95a2d71722764
> # good: [8cc683833726912498130a0130fc3bd0d169ef59] Merge branch 'net-dp83822-fix-interrupt-floods'
> git bisect good 8cc683833726912498130a0130fc3bd0d169ef59
> # bad: [0840a7914caa14315a3191178a9f72c742477860] Merge tag 'char-misc-5.19-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/char-misc
> git bisect bad 0840a7914caa14315a3191178a9f72c742477860
> # bad: [38bc4ac431684498126f9baa3a530e5a132f0173] Merge tag 'drm-fixes-2022-06-24' of git://anongit.freedesktop.org/drm/drm
> git bisect bad 38bc4ac431684498126f9baa3a530e5a132f0173
> # good: [399bd66e219e331976fe6fa6ab81a023c0c97870] Merge tag 'net-5.19-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net
> git bisect good 399bd66e219e331976fe6fa6ab81a023c0c97870
> # bad: [a237cfd6b7469d6f5eeaa45f30128ab78b5281a5] Merge tag 'block-5.19-2022-06-24' of git://git.kernel.dk/linux-block
> git bisect bad a237cfd6b7469d6f5eeaa45f30128ab78b5281a5
> # good: [92f20ff72066d8d7e2ffb655c2236259ac9d1c5d] Merge tag 'pm-5.19-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm
> git bisect good 92f20ff72066d8d7e2ffb655c2236259ac9d1c5d
> # bad: [598f2404879336277a4320ac5000394b873e049a] Merge tag 'io_uring-5.19-2022-06-24' of git://git.kernel.dk/linux-block
> git bisect bad 598f2404879336277a4320ac5000394b873e049a
> # good: [386e4fb6962b9f248a80f8870aea0870ca603e89] io_uring: use original request task for inflight tracking
> git bisect good 386e4fb6962b9f248a80f8870aea0870ca603e89
> # good: [2d9ef940f89e0ab4fde7ba6f769d82f2a450c35a] Revert "printk: extend console_lock for per-console locking"
> git bisect good 2d9ef940f89e0ab4fde7ba6f769d82f2a450c35a
> # bad: [07a22b61946f0b80065b0ddcc703b715f84355f5] Revert "printk: add functions to prefer direct printing"
> git bisect bad 07a22b61946f0b80065b0ddcc703b715f84355f5
> # bad: [5831788afb17b89c5b531fb60cbd798613ccbb63] Revert "printk: add kthread console printers"
> git bisect bad 5831788afb17b89c5b531fb60cbd798613ccbb63
> # first bad commit: [5831788afb17b89c5b531fb60cbd798613ccbb63] Revert "printk: add kthread console printers"
> 
> On 5831788afb17 ("Revert "printk: add kthread console printers"") or on
> 07a22b61946f ("Revert "printk: add functions to prefer direct printing"")
> I was not able to reproduce the exact error conditions that I got on
> the current HEAD or any of the commits along my bisection, but instead I
> got something different [2]:

This kind of problems became visible with introducing printk console
kthreads in 5.19-rc1. And it was actually the main reason why they
were reverted in 5.19-rc4.

The mainline merge commits when they were introduced and reverted:

+ introduced: commit 537e62c865dcb9b91 ("Merge tag 'printk-for-5.19'
  of git://git.kernel.org/pub/scm/linux/kernel/git/printk/linux")

+ reverted: commit 9d882352bac8f2ff375 ("Merge tag 'printk-for-5.19-rc4'
  of git://git.kernel.org/pub/scm/linux/kernel/git/printk/linux


The problem might happen when console write() callback is not properly
serialized against other operations on the same port. It might usually
happen even without the kthreads. But the kthreads significantly
increased the chance to hit the problem.

The best solution is to fix the console driver. So far, we
noticed two typical problems:

1. Early console did not take port->lock. For example, see the commit
   589f892ac8ef244e47c5a0 ("serial: meson: acquire port->lock in startup()").

2. Console driver expected that the write() callback would be called
   with IRQs disabled. For example, see the commit aabdbb1b7a5819e18
   ("serial: msm_serial: disable interrupts in __msm_console_write().

Best Regards,
Petr