[Crash-utility] [RFC 0/4] Add support to extract hardware device dumps from vmcore

[Dave Anderson]

----- Original Message -----
> Hi Dave,
> 
> On Saturday, April 04/13/19, 2019 at 00:39:09 +0530, Dave Anderson wrote:
> > 
> > Hi Surenda,
> > 
> > Great -- I've been looking forward for this patch set to arrive.
> > 
> > A couple things...
> > 
> > First, I'm going to need two sample vmcores (one ELF, one compressed kdump)
> > along with the associated vmlinux.  You can contact me off-list with
> > details
> > on how we can arrange a transfer.
> > 
> > Second, I do not want to add a new command.  I rarely do so -- in fact,
> > the only times since the original crash utility was released was in 2012,
> > when the "ipcs" and "tree" commands were added in crash-6.0.7 and
> > crash-6.0.8.
> > New functionality is typically added as an option to an existing command.
> > And in this case, the natural location to put it is in the existing "dev"
> > command, and the devdump_extract() function can be moved into dev.c.
> > (FWIW, you can add your Chelsio copyright at the top of that file)
> > 
> 
> Ok, thanks for the suggestion. We will move the logic to dev command,
> instead.
> 
> We're thinking along the lines of following sample commands:
> 
> Display the available device dumps
> crash> dev -v
> INDEX                         NAME                      OFFSET      SIZE
> 0                      cxgb4_0000:02:00.4               0x278       33558464
> 1                      cxgb4_0000:03:00.4               0x2001278   33558464
> 
> Extract device dump at specified index
> crash> devdump -v 0 -f device_dump_0.bin
> 33558464 bytes copied from 0x278 to device_dump_0.bin
> 
> Let us know your thoughts.

Hi Surenda,

I've got your sample files -- thanks for them, I really appreciate it.

Here are my thoughts...

You probably want to make "-V" display the list of available device dumps in the
vmcore, and make "-v <index>" select a singular device for dumping.

I also have a question re: the note contents.  Is it up to the individual device
as to what format the dump contents are made up of?  Are they always binary
dumps, or could a device dump ASCII log data or something to that effect?

I ask because I see that you are calling display_memory() using these arguments:

  void
  display_memory_from_file_offset(ulonglong addr, long count, void *opt)
  {
          display_memory(addr, count, DISPLAY_RAW | ASCII_ENDLINE | HEXADECIMAL,
                         FILEADDR, opt);
  }

Since you are using DISPLAY_RAW, display_memory() will simply copy the
note data unmodified directly to the file, and the ASCII_ENDLINE and 
HEXADECIMAL arguments are ignored.  So it's not clear why you added them?  
But since you did, I'm now wondering whether it would be useful for
a user to optionally dump a human-readable HEXADECIMAL/ASCII_ENDLINE 
formatted display of the data to the screen?  If so, then perhaps if 
the "-v index" option is used alone *without* a file specified,  
why not just do a translated device dump to the screen?  

So to summarize, my suggestions would be:

  dev -V                        dump the list of device dumps
  dev -v <index> [ -r file ]    select and display one device dump, either in
                                human-readable format to the screen by default,
                                or optionally copy it raw to a file

Thanks,
  Dave




> 
> > Third, there are some aesthetic changes that should be made in order
> > to have the display use a more traditional output format like those
> > used by other commands, e.g., with a single header with NAME, OFFSET and
> > SIZE columns.
> > 
> 
> Ok, We will fix this in next version.
> 
> > Other than that, this looks good on paper!
> > 
> > Thanks,
> >   Dave
> > 
> 
> Thanks
> Surendra
> 
> 
> > 
> > ----- Original Message -----
> > > When kernel panic happens and kdump crash kernel is loaded, device
> > > drivers enabled in the kdump crash kernel collect device specific
> > > snapshot of the hardware/firmware state of their underlying devices.
> > > These snapshots are exported as ELF notes with note type NT_VMCOREDD
> > > (i.e., 0x700) in vmcore [1].
> > > 
> > > This series of patches enhance crash utility to analyze and
> > > extract these hardware specific device dumps from vmcore using
> > > a new "devdump" command.
> > > 
> > > Patches 1 and 2 enhance help -D to parse NT_VMCOREDD ELF notes
> > > in ELF vmcore and KDUMP vmcore, respectively.
> > > 
> > > Patches 3 and 4 implement devdump command to analyze and extract
> > > hardware specific device dumps from ELF vmcore and KDUMP vmcore,
> > > respectively.
> > > 
> > > Suggestions and feedback will be much appreciated.
> > > 
> > > Thanks,
> > > Surendra
> > > 
> > > [1] https://lkml.org/lkml/2018/5/2/190
> > > 
> > > Surendra Mobiya (4):
> > >   parse NT_VMCOREDD ELF notes in ELF vmcore
> > >   parse NT_VMCOREDD ELF notes in KDUMP vmcore
> > >   add devdump command to extract NT_VMCOREDD from ELF vmcore
> > >   enhance devdump command to extract NT_VMCOREDD from KDUMP vmcore
> > > 
> > >  Makefile      |   4 +--
> > >  defs.h        |  16 +++++++++
> > >  devdump.c     | 114
> > >  ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
> > >  diskdump.c    |  71 +++++++++++++++++++++++++++++++++++-
> > >  global_data.c |   1 +
> > >  help.c        |  23 ++++++++++++
> > >  memory.c      |   7 ++++
> > >  netdump.c     |  89 +++++++++++++++++++++++++++++++++++++++++++--
> > >  netdump.h     |   3 ++
> > >  vmcore.h      |  36 +++++++++++++++++++
> > >  10 files changed, 359 insertions(+), 5 deletions(-)
> > >  create mode 100644 devdump.c
> > >  create mode 100644 vmcore.h
> > > 
> > > --
> > > 1.8.3.1
> > > 
> > > 
>