Re: [PATCH v13 0/8] Coresight for Kernel panic and watchdog reset

[James Clark]

On 16/12/2024 5:30 am, Linu Cherian wrote:
> This patch series is rebased on coresight-next-v6.12.rc4
>
> * Patches 1 & 2 adds support for allocation of trace buffer pages from
>    reserved RAM
> * Patches 3 & 4 adds support for saving metadata at the time of kernel panic
> * Patch 5 adds support for reading trace data captured at the time of panic
> * Patches 6 & 7 adds support for disabling coresight blocks at the time of panic
> * Patch 8: Gives the full description about this feature as part of documentation
>
> v12 is posted here,
> https://lore.kernel.org/linux-arm-kernel/20241129084714.3057080-1-lcherian@xxxxxxxxxxx/
>
> Changelog from v12:
> * Fixed wrong buffer pointer passed to coresigh_insert_barrier_packet
> * tmc_read_prepare/unprepare_crashdata need to be called only once and
>    hence removed from read path and added to tmc_probe
> * tmc_read_prepare_crashdata renamed to tmc_prepare_crashdata and
>    avoid taking locks  as its moved to probe function.
> * Introduced read status flag, "reading" specific to reserved buffer to keep the
>    reserved buffer reading independent of the regular buffer.
> * open/release ops for reserved buffer has to take care only about the
>    set/unset the "reading" status flag as the reserved buffer is prepared
>    during the probe time itself.
> * Few other trivial changes

Hi Linu,

I tested that decoding a crash dump of ETM1 (trace ID 17) from panic 
kernel works:

  $ ./ptm2human -i cstrace.bin

  ...
  There is no valid data in the stream of ID 16
  Decode trace stream of ID 17
  Syncing the trace stream...
  Decoding the trace stream...
  instruction addr at 0x140c9afc, ARM state, secure state,
  ...

I noticed that once in the panic kernel Coresight becomes unusable, and 
the Perf Coresight tests fail, with no obvious way to reset it other 
than a cold boot:

 $ perf record -e cs_etm//u -- true
 $ perf report -D | grep AUX
 ...
 AUX data lost 27 times out of 27!
 ...

I didn't debug it yet. I thought it might be something to do with the 
RESRV buffer mode, but it doesn't look like that should be the case from 
the code. Perhaps its the claim tags and coresight_is_claimed_any() 
lingering, so it's not really an issue that's introduced by this change?

Thanks
James