[makedumpfile PATCH RFC v0.1] Implemented the --fill-excluded-pages=<value> feature

[Dave Anderson]

----- Original Message -----

> When a page is excluded by any of the existing dump levels,
> that page may still be written to the ELF dump file, depending
> upon the PFN_EXCLUDED mechanism.
> 
> The PFN_EXCLUDED mechanism looks for N consecutive "not
> dumpable" pages, and if found, the current ELF segment is
> closed out and a new ELF segment started, at the next dumpable
> page. Otherwise, if the PFN_EXCLUDED criteria is not meet (that
> is, there is a mix of dumpable and not dumpable pages, but not
> N consecutive not dumpable pages) all pages are written to the
> dump file.
> 
> This patch implements a mechanism for those "not dumpable" pages
> that are written to the ELF dump file to fill those pages with
> constant data, rather than the original data. In other words,
> the dump file still contains the page, but its data is wiped.
> 
> The motivation for doing this is to protect real user data from
> "leaking" through to a dump file when that data was asked to be
> omitted. This is especially important for effort I currently am
> working on to allow further refinement of what is allowed to be
> dumped, all in an effort to protect user (customer) data.
> 
> The patch is simple enough, however, it causes problems with
> crash; crash is unable to load the resulting ELF dump file.
> For example, I do the following as a test scenario for this
> change:
> 
> - Obtain a non-filtered dump file (eg. dump level 0, no -d option,
>   or straight copy of /proc/vmcore)
> - Run vmcore through 'crash' to ensure loads ok, test with
>   commands like: ps, files, etc.
>   % crash vmlinux vmcore
> - Apply this patch and rebuild makedumpfile
> - Run vmcore through makedumpfile *without* --fill-excluded-pages
>   and with filtering to ensure no uintended side effects of patch:
>   % ./makedumpfile -E -d31 -x vmlinux vmcore newvmcore
> - Run new vmcore through crash to ensure still loads ok, test
>   with commands like: ps, files, etc.
>   % crash vmlinux newvmcore
> - Run vmcore through makedumpfile *with* --fill-excluded-pages
>   and with filtering to check side effects of patch:
>   % ./makedumpfile -E -d31 --fill-excluded-pages=0 -x vmlinux vmcore
>   newvmcore2
> - Run new vmcore through crash to ensure still loads ok, test
>   with commands like: ps, files, etc.
>   % crash vmlinux newvmcore2
> 
> But crash yields errors like:
>   [...]
>   This GDB was configured as "x86_64-unknown-linux-gnu"...
> 
>   crash: cannot determine thread return address
>   please wait... (gathering kmem slab cache data)
>   crash: invalid kernel virtual address: 1c  type: "kmem_cache
>   objsize/object_size"
> 
> If the patch is correct/accurate, then that may mean that crash
> is using data which it should not be.

Why would the crash utility be "using data which it should not be"
if your patch is applied?

The two error messages above come from attempting to read memory
from the kernel text region (the "thread return" message), and then
the kmem_cache.object_size field of the kernel's kmem_cache data structure
pointed to by its "kmem_cache" pointer.  It looks like the patch is
causing bogus data to be returned for a given physical address.

Dave


> 
> The more likely scenario is that the patch is not correct/accurate,
> and I'm corrupting the dump file.
> 
> Please provide feedback!!