Re: Invalid pstore_blk use?

From: Kamal Dasu
Date: Wed Dec 07 2022 - 13:31:52 EST


+ Adrain Hunter
+ Ulf Hansoon



On Fri, Nov 18, 2022 at 3:41 PM Kamal Dasu <kdasu.kdev@xxxxxxxxx> wrote:
>
> Kees,
>
> I am in the process of implementing mmcpstore backend for mmc based
> on the mtdpstore driver
>
> This is what is registered with register_pstore_device(&cxt->dev);
> cxt->dev.flags = PSTORE_FLAGS_DMESG;
> cxt->dev.zone.read = mmcpstore_read;
> cxt->dev.zone.write = mmcpstore_write;
> cxt->dev.zone.erase = mmcpstore_erase;
> cxt->dev.zone.panic_write = mmcpstore_panic_write;
>
> # dmesg | grep pstor
> [ 0.000000] Kernel command line: pstore_blk.blkdev=/dev/mmcblk1p8
> crash_kexec_post_notifiers printk.always_kmsg_dump
> [ 1.993986] pstore_zone: registered pstore_blk as backend for
> kmsg(Oops,panic_write) pmsg
> [ 2.002582] pstore: Using crash dump compression: deflate
> [ 2.008133] pstore: Registered pstore_blk as persistent store backend
> [ 2.020907] mmcpstore: /dev/mmcblk1p8 size 131072 start sector
> 34468 registered as psblk backend
> [ 17.868753] psz_kmsg_recover_meta: pstore_zone: no valid data in
> kmsg dump zone 0
> [ 18.298933] psz_recover_zone: pstore_zone: no valid data in zone pmsg
> [ 18.305398] psz_recovery: pstore_zone: recover end!
>
> The driver is successfully registered and the read path works when
> /sys/fs/pstor is mounted , however mmc_pstore_panic_write is not
> called.
> Need help in understanding what could be missing. I am using the
> latest upstream kernel for testing.
>
> Kamal
>
> On Tue, Aug 9, 2022 at 7:06 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> >
> > On Tue, Aug 09, 2022 at 11:35:08AM -0700, Florian Fainelli wrote:
> > > Hi Kees, WeiXiong,
> > >
> > > On 7/14/22 20:49, Florian Fainelli wrote:
> > > > Hi Kees, WeiXiong,
> > > >
> > > > I am trying to make use of pstore_blk which is BTW exactly what I had
> > > > been looking for to store panic/console logs onto an eMMC partition.
> > > >
> > > > Using the 5.10 kernel plus:
> > > >
> > > > 7e2e92e9861b Revert "mark pstore-blk as broken"
> > > > 01c28bc8f389 pstore/blk: Use the normal block device I/O path
> > > > 2a7507999638 pstore/blk: remove {un,}register_pstore_blk
> > > > fef0b337cd25 pstore/zone: cap the maximum device size
> > > >
> > > > or the android13-5.15 (at Merge 5.15.40 into android13-5.15) kernel with
> > > > no changes and using:
> > > >
> > > > mount -t pstore pstore /sys/fs/pstore
> > > > modprobe pstore_blk blkdev=/dev/mmcblk1p9 best_effort=yes
> > > >
> > > > upon triggering a crash with:
> > > >
> > > > echo c > /proc/sysrq-trigger
> > > >
> > > > and rebooting and remounting the pstore filesystem and loading
> > > > pstore_blk, I only have:
> > > >
> > > > # ls /sys/fs/pstore/
> > > > console-pstore_blk-0
> > > >
> > > > which contains the entire console log up to, but excluding the crash.
> > > > The kernel does show that pstore_blk was used for all 3 types of kmsg,
> > > > pmsg and console:
> > > >
> > > > [ 28.649514] pstore_zone: capping size to 128MiB
> > > > [ 28.712894] pstore_zone: registered pstore_blk as backend for
> > > > kmsg(Oops) pmsg console
> > > > [ 28.721145] pstore: Using crash dump compression: deflate
> > > > [ 28.906253] printk: console [pstore_blk-1] enabled
> > > > [ 28.911229] pstore: Registered pstore_blk as persistent store backend
> > > > [ 28.917735] pstore_blk: attached pstore_blk:/dev/mmcblk1p9
> > > > (134217728) (no dedicated panic_write!)
> > > >
> > > > there is no automatic reboot upon panic, so I just tend to reboot after
> > > > 2-3 seconds manually. The kernel is configured with the default
> > > > CONFIG_PSTORE_* options.
> > > >
> > > > Is the observed behavior a limitation of the best_effort mode? If so, do
> > > > we have any plans to implementing a non-best effort mode for eMMC
> > > > devices?
> > >
> > > Any feedback on my email? I did try to get kernel panics to be dumped out to
> >
> > Hi! Sorry I lost this email originally. :)
> >
> > > a dedicated /dev/mtdblock* partition for which there ought to be support for
> > > mtd->panic_write, but it still did not work any better. Is there something
> >
> > With the mtdblock driver, do you still see:
> >
> > pstore_blk: attached pstore_blk:/dev/... (no dedicated panic_write!)
> > ^^^^^^^^^^^^^^^^^^^^^^^^^
> >
> > > obvious that I am missing which prevents kernel panics from being logged?
> >
> > Unfortunately it really depends on how the drivers are built. If the
> > block layer is shut down during a panic, pstore_blk won't catch the
> > panic. :(
> >
> > --
> > Kees Cook