[PATCH] makedumpfile: Error on re-filtering the dump file with no free pages

[Michael Holzheu]

Am Thu, 18 May 2017 04:37:20 +0000
schrieb Atsushi Kumagai <ats-kumagai at wm.jp.nec.com>:

> Hello Zaslonko,
> 
> > When re-filtering the dump file after the free pages have already been
> > stripped out we get an error "Can't get necessary symbols for excluding
> > free pages" if newly specified dump level is below 16 (preserves free
> > pages).
> > According to the code, the check for the new dump level is done BEFORE
> > the new dump level is actually set (based on the dump level specified in
> > the command and the one from the input dump file).
> > Moving the new_dump_level calculation ahead would fix the error.
> 
> Well, I guess the real problem is different.
> The error you said is printed by exclude_free_page():
> 
>         if ((SYMBOL(node_data) == NOT_FOUND_SYMBOL)
>             && (SYMBOL(pgdat_list) == NOT_FOUND_SYMBOL)
>             && (SYMBOL(contig_page_data) == NOT_FOUND_SYMBOL)) {
>                 ERRMSG("Can't get necessary symbols for excluding free pages.\n");
>                 return FALSE;
>         }
> 
> I think the availability of these symbols are not related to free pages.
> exclude_free_page() is called if info->page_is_buddy is null, I estimate that
> this situation occurs only if the kernel is older(2.6.16 or before). 
> 
> However, I don't think you use such too old kernel, so I suspect that
> setup_page_is_buddy() should be updated for newer kernel.

Mikhail is on vacation now - so I try to explain:

The test case is as follows:

 1) We have a -d31 filtered dump "dump.d31"
 2) We want to compress the dump with "makedumpfile -c dump.31 dump31.compressed"

This fails with:

makedumpfile -c dump.31 dump.31.compressed
Excluding unnecessary pages        : [100.0 %]
exclude_free_page: Can't get necessary symbols for excluding free pages.
dump_level is changed to 31, because dump.31 was created by dump_level(31).
makedumpfile Failed.

The problem is that setup_page_is_buddy() is not called in this case because
info->dump_level is still 0 since it was not adjusted early enough:

        if (info->dump_level & DL_EXCLUDE_FREE)
                setup_page_is_buddy();

Because it is not set info->page_is_buddy is NULL and therefore the following
if condition gets true:

        if ((info->dump_level & DL_EXCLUDE_FREE) && !info->page_is_buddy)
                if (!exclude_free_page(cycle))
                        return FALSE;

Since we don't have the symbols in VMCOREINFO (and IMHO don't need it?) the
exclude_free_page() functions fails with the described error message.

So our fix is to adjust the info->level before setup_page_is_buddy() is called.

Best Regards
Michael

> Could you tell me your kernel version and how to reproduce this issue ?