Re: [PATCH v2 03/20] vmcore, sysfs: export ELF note segment size instead of vmcoreinfo data size

[HATAYAMA Daisuke <d.hatayama@jp.fujitsu.com>]

From: HATAYAMA Daisuke <d.hatayama@jp.fujitsu.com>
Subject: [PATCH v2 03/20] vmcore, sysfs: export ELF note segment size instead of vmcoreinfo data size
Date: Sat, 2 Mar 2013 17:36:05 +0900

> p_memsz member of program header entry with PT_NOTE type needs to have
> size of the corresponding ELF note segment. Currently, vmcoreinfo
> exports data part only. If vmcoreinfo reachs vmcoreinfo_max_size, then
> in merge_note_headers_elf{32,64}, empty ELF note header cannot be
> found or buffer overrun can happen.

Sorry, I noticed this "buffer overrun can happen" was completely
wrong. In merge_note_headers_elf{32,64}, the size is being checked to
avoid buffer overrun.

                int j;
                void *notes_section;
                struct vmcore *new;
                u64 offset, max_sz, sz, real_sz = 0;
...
                for (j = 0; j < max_sz; j += sz) {
                        if (nhdr_ptr->n_namesz == 0)
                                break;
                        sz = sizeof(Elf32_Nhdr) +
                                ((nhdr_ptr->n_namesz + 3) & ~3) +
                                ((nhdr_ptr->n_descsz + 3) & ~3);
                        real_sz += sz;
                        nhdr_ptr = (Elf32_Nhdr*)((char*)nhdr_ptr + sz);
                }

But later patch changes teminator of ELF note segments from the null
not header to NT_VMCORE_PAD note type. It's important to export a
whole buffer for ELF note segments, not data part only. This patch
description doesn't explain this, and I'll add this explanation in the
next version.

Also, here j has int type but the other variables compared with the j
have u64 type. This is strange, and in fact verbose because for the
purpose of the j, real_sz seems exact. I'll replace the for statement
by while statement in additional clean-up patch as:

  while (real_sz < max_sz) {
    ..
  }

Thanks.
HATAYAMA, Daisuke