Re: Boot stall regression from "printk for 5.19" merge

[Marek Behún]

On Mon, 20 Jun 2022 19:13:37 +0900
Sergey Senozhatsky <senozhatsky@xxxxxxxxxxxx> wrote:

> On (22/06/20 12:02), Marek Behún wrote:
> > > On (22/06/19 20:49), Marek Behún wrote:
> > > [..]  
> > > > causes a regression on arm64 (Marvell CN9130-CRB board) where the
> > > > system boot freezes in most cases (and is unusable until restarted by
> > > > watchdog), or, in some cases boots, but the console output gets mangled
> > > > for a while (the serial console spits garbage characters).    
> > > 
> > > Can you please try disabling console kthreads and see how the boot
> > > process goes? Just `return 0` from printk_activate_kthreads() (I think
> > > this should do the trick).  
> > 
> > This indeed makes the problem go away...  
> 
> Oh... OK. Didn't expect that :)
> 
> > > > The garbage example:
> > > > 
> > > >   ...
> > > >   [    0.920951] raid6: using neon recovery algorithm
> > > >   [    0.921228] iommu: Default domain type: Translated
> > > >   %
> > > > 
> > > >           gb@k+cFL/[    4.954974] DSA: tree 0 setup
> > > >   [    4.955286] cfg80211: Loading compiled-in X.509 certificates for regulatory database    
> > > 
> > > This is pretty suspicious. I don't see how console kthreads would
> > > corrupt the output. I suspect that something else is going on, some
> > > memory corruption, etc.  
> > 
> > Maybe multiple threads are writing to serial registers, or something...  
> 
> That's possible. Console drivers usually should grab port->lock for
> write(), but maybe something is missing in the driver you use. What
> console driver are you using?

compatible = "snps,dw-apb-uart", so drivers/tty/serial/8250/8250_dw.c

seems that the function dw8250_serial_out() does not use the spinlock...

Marek