Re: [PATCH] kexec: fix out of the ELF headers buffer issue in syscall kexec_file_load()

[joeyli <jlee@suse.com>]

Hi,

On Mon, Sep 28, 2015 at 03:16:41PM +0800, Baoquan He wrote:
> Hi Chun-Yi,
> 
> On 09/28/15 at 02:41pm, Lee, Chun-Yi wrote:
> > On big machines have CPU number that's very nearly to consume whole ELF
> > headers buffer that's page aligned, 4096, 8192... Then the page fault error
> > randomly happened.
> > 
> > This patch modified the code in fill_up_crash_elf_data() by using
> > walk_system_ram_res() instead of walk_system_ram_range() to count the max
> > number of crash memory ranges. That's because the walk_system_ram_range()
> > filters out small memory regions that reside the same page, but
> > walk_system_ram_res() does not.
> > 
> > The oringial page fault issue sometimes happened on big machines when
> > preparing ELF headers:
> > 
> > [  305.291522] BUG: unable to handle kernel paging request at ffffc90613fc9000
> > [  305.299621] IP: [<ffffffff8103d645>] prepare_elf64_ram_headers_callback+0x165/0x260
> > [  305.308300] PGD e000032067 PUD 6dcbec54067 PMD 9dc9bdeb067 PTE 0
> > [  305.315393] Oops: 0002 [#1] SMP
> > [...snip]
> > [  305.420953] task: ffff8e1c01ced600 ti: ffff8e1c03ec2000 task.ti: ffff8e1c03ec2000
> > [  305.429292] RIP: 0010:[<ffffffff8103d645>]  [<ffffffff8103d645>] prepare_elf64_ra
> > m_headers_callback+0x165/0x260
> > [...snip]
> > 
> > After tracing prepare_elf64_headers() and prepare_elf64_ram_headers_callback(),
> > the code uses walk_system_ram_res() to fill-in crash memory regions information
> > to program header, so it counts those small memory regions that reside in a
> > page area. But, when kernel was using walk_system_ram_range() in
> > fill_up_crash_elf_data() to count the number of crash memory regions, it
> > filters out small regions.
> > 
> > I printed those small memory regions, for example:
> > 
> > kexec: Get nr_ram ranges. vaddr=0xffff880077592258 paddr=0x77592258, sz=0xdc0
> > 
> > Base on the logic of walk_system_ram_range(), this memory region will be
> > filter out:
> > 
> > pfn = (0x77592258 + 0x1000 - 1) >> 12 = 0x77593
> > end_pfn = (0x77592258 + 0xfc0 -1 + 1) >> 12 = 0x77593
> > end_pfn - pfn = 0x77593 - 0x77593 = 0  <=== if (end_pfn > pfn)      [FAIL]
> > 
> > So, the max_nr_ranges that counted by kernel doesn't include small memory
> > regions. That causes the page fault issue happened in later code path for
> > preparing EFL headers,
> > 
> > This issue was hided on small machine that doesn't have too many CPU because
> > the free space of ELF headers buffer can cover the number of small memory
> > regions. But, when the machine has more CPUs or the number of memory regions
> > very nearly to consume whole page aligned buffer, e.g. 4096, 8192... Then
> > issue will happen randomly.
> 
> It's a good finding and fix sounds reasonable. I didn't get why too many
> CPUs will cause this bug. From your big machine can you check which
> regions they are and what they are used for? I guess you mean the
> crash_notes region, but not very sure.
> 

In prepare_elf64_headers, the logic to allocate ELF header buffer is:

        /* extra phdr for vmcoreinfo elf note */
        nr_phdr = nr_cpus + 1;
        nr_phdr += ced->max_nr_ranges;

        /*
         * kexec-tools creates an extra PT_LOAD phdr for kernel text mapping
         * area on x86_64 (ffffffff80000000 - ffffffffa0000000).
         * I think this is required by tools like gdb. So same physical
         * memory will be mapped in two elf headers. One will contain kernel
         * text virtual addresses and other will have __va(physical) addresses.
         */
        nr_phdr++;
        elf_sz = sizeof(Elf64_Ehdr) + nr_phdr * sizeof(Elf64_Phdr);
        elf_sz = ALIGN(elf_sz, ELF_CORE_HEADER_ALIGN);

So whole buffer will be consumed as following:
0                                                                                                                       4096
+------------+--------------------+--------------------+---------------------------+---------------------------+---------+
| ELF header | each cpu PT_NOTE...| vmcoreinfo PT_NOTE | kernel text region PT_NOTE| PT_NOTE for memory regions|   free  |
| (64 bytes) | (n * 56 bytes)     | (56 bytes)         | (56 bytes)                | (n * 56 bytes)            |         |
+------------+--------------------+--------------------+---------------------------+---------------------------+---------+

When the free space can cover the number of small memory regions, means the
difference between walk_system_ram_range() and walk_system_ram_res(), then
this issue will not trigger.

But, when the CPU number grows to very nearly to consume whole 4096 buffer
then the issue will be happen.


Thanks a lot!
Joey Lee


> > 
> > Signed-off-by: Lee, Chun-Yi <jlee@suse.com>
> > ---
> >  arch/x86/kernel/crash.c | 5 ++---
> >  1 file changed, 2 insertions(+), 3 deletions(-)
> > 
> > diff --git a/arch/x86/kernel/crash.c b/arch/x86/kernel/crash.c
> > index e068d66..ad273b3d 100644
> > --- a/arch/x86/kernel/crash.c
> > +++ b/arch/x86/kernel/crash.c
> > @@ -185,8 +185,7 @@ void native_machine_crash_shutdown(struct pt_regs *regs)
> >  }
> >  
> >  #ifdef CONFIG_KEXEC_FILE
> > -static int get_nr_ram_ranges_callback(unsigned long start_pfn,
> > -				unsigned long nr_pfn, void *arg)
> > +static int get_nr_ram_ranges_callback(u64 start, u64 end, void *arg)
> >  {
> >  	int *nr_ranges = arg;
> >  
> > @@ -214,7 +213,7 @@ static void fill_up_crash_elf_data(struct crash_elf_data *ced,
> >  
> >  	ced->image = image;
> >  
> > -	walk_system_ram_range(0, -1, &nr_ranges,
> > +	walk_system_ram_res(0, -1, &nr_ranges,
> >  				get_nr_ram_ranges_callback);
> >  
> >  	ced->max_nr_ranges = nr_ranges;
> > -- 
> > 2.1.4
> >