Re: [RFC patch 5/7] crash: Add VMCOREINFO_FIELD_AND_OFFSET()

[Petr Mladek]

On Wed 2017-11-15 19:15:36, Thomas Gleixner wrote:
> Changing the time stamp storage in the printk ringbuffer entries would
> break existing tools which rely on the availability of the 'ts_nsec' field
> in vmcore info.
> 
> Provide a new macro which allows to store the offset of a member of struct
> printk_log with a different name. This allows to change the underlying
> storage and without breaking tools.

It seems that this cheating is not enough. I tried to use it in printk
(6th patch in this patch set) and the "crash" tool complained:

===== cut =====
WARNING: log buf data structure(s) have changed

log: invalid structure member offset: log_ts_nsec
     FILE: kernel.c  LINE: 5055  FUNCTION: dump_log_entry()
===== cut =====


The crash tool get this value using the following code:

  MEMBER_OFFSET_INIT(log_ts_nsec, "printk_log", "ts_nsec");
  + MEMBER_OFFSET()
    + datatype_info((X), (Y), NULL)

, where the last function does something like:

long
datatype_info(char *name, char *member, struct datatype_member *dm)
{
	struct gnu_request *req;

	req->command = GNU_GET_DATATYPE;
	req->name = name;
	req->member = member;

	gdb_interface(req);

I am not completely sure what it does. I guess that gdb
is able to get information about the real types either
from vmlinux binary or vmcore. It is not able to
find "ts_nsec" in the struct "printk_log" and fails.


I am afraid that we would need to fix crash for
the modified struct printk_log. Then this hack
will not be needed.

Best Regards,
Petr