[PATCH v2][makedumpfile] Fix a data race in multi-threading mode (--num-threads=N)

[Tao Liu]

Hi Kazu,

On Wed, Jul 2, 2025 at 12:13 PM HAGIO KAZUHITO(萩尾　一仁)
<k-hagio-ab at nec.com> wrote:
>
> On 2025/07/01 16:59, Tao Liu wrote:
> > Hi Kazu,
> >
> > Thanks for your comments!
> >
> > On Tue, Jul 1, 2025 at 7:38 PM HAGIO KAZUHITO(萩尾　一仁) <k-hagio-ab at nec.com> wrote:
> >>
> >> Hi Tao,
> >>
> >> thank you for the patch.
> >>
> >> On 2025/06/25 11:23, Tao Liu wrote:
> >>> A vmcore corrupt issue has been noticed in powerpc arch [1]. It can be
> >>> reproduced with upstream makedumpfile.
> >>>
> >>> When analyzing the corrupt vmcore using crash, the following error
> >>> message will output:
> >>>
> >>>       crash: compressed kdump: uncompress failed: 0
> >>>       crash: read error: kernel virtual address: c0001e2d2fe48000  type:
> >>>       "hardirq thread_union"
> >>>       crash: cannot read hardirq_ctx[930] at c0001e2d2fe48000
> >>>       crash: compressed kdump: uncompress failed: 0
> >>>
> >>> If the vmcore is generated without num-threads option, then no such
> >>> errors are noticed.
> >>>
> >>> With --num-threads=N enabled, there will be N sub-threads created. All
> >>> sub-threads are producers which responsible for mm page processing, e.g.
> >>> compression. The main thread is the consumer which responsible for
> >>> writing the compressed data into file. page_flag_buf->ready is used to
> >>> sync main and sub-threads. When a sub-thread finishes page processing,
> >>> it will set ready flag to be FLAG_READY. In the meantime, main thread
> >>> looply check all threads of the ready flags, and break the loop when
> >>> find FLAG_READY.
> >>
> >> I've tried to reproduce the issue, but I couldn't on x86_64.
> >
> > Yes, I cannot reproduce it on x86_64 either, but the issue is very
> > easily reproduced on ppc64 arch, which is where our QE reported.
> > Recently we have enabled --num-threads=N in rhel by default. N ==
> > nr_cpus in 2nd kernel, so QE noticed the issue.
>
> I see, thank you for the information.
>
> >
> >>
> >> Do you have any possible scenario that breaks a vmcore?  I could not
> >> think of it only by looking at the code.
> >
> > I guess the issue only been observed on ppc might be due to ppc's
> > memory model, multi-thread scheduling algorithm etc. I'm not an expert
> > on those. So I cannot give a clear explanation, sorry...
>
> ok, I also don't think of how to debug this well..
>
> >
> > The page_flag_buf->ready is an integer that r/w by main and sub
> > threads simultaneously. And the assignment operation, like
> > page_flag_buf->ready = 1, might be composed of several assembly
> > instructions. Without atomic r/w (memory) protection, there might be
> > racing r/w just within the few instructions, which caused the data
> > inconsistency. Frankly the ppc assembly consists of more instructions
> > than x86_64 for the same c code, which enlarged the possibility of
> > data racing.
> >
> > We can observe the issue without the help of crash, just compare the
> > binary output of vmcore generated from the same core file, and
> > compress it with or without --num-threads option. Then compare it with
> > "cmp vmcore1 vmcore2" cmdline, and cmp will output bytes differ for
> > the 2 vmcores, and this is unexpected.
> >
> >>
> >> and this is just out of curiosity, is the issue reproduced with
> >> makedumpfile compiled with -O0 too?
> >
> > Sorry, I haven't done the -O0 experiment, I can do it tomorrow and
> > share my findings...
>
> Thanks, we have to fix this anyway, I want a clue to think about a
> possible scenario..

1) Compiled with -O2 flag:

[root at ibm-p10-01-lp45 makedumpfile]# ./makedumpfile -d 31 -l ~/vmcore /tmp/out1
Copying data                                      : [100.0 %] /
   eta: 0s

The dumpfile is saved to /tmp/out1.

makedumpfile Completed.
[root at ibm-p10-01-lp45 makedumpfile]# ./makedumpfile --num-threads=2 -d
31 -l ~/vmcore /tmp/out2
Copying data                                      : [100.0 %] |
   eta: 0s
Copying data                                      : [100.0 %] \
   eta: 0s

The dumpfile is saved to /tmp/out2.

makedumpfile Completed.
[root at ibm-p10-01-lp45 makedumpfile]# cd /tmp
[root at ibm-p10-01-lp45 tmp]# cmp out1 out2
out1 out2 differ: byte 20786414, line 108064

2) Compiled with -O0 flag:

[root at ibm-p10-01-lp45 makedumpfile]# ./makedumpfile -d 31 -l ~/vmcore /tmp/out3
Copying data                                      : [100.0 %] /
   eta: 0s

The dumpfile is saved to /tmp/out3.

makedumpfile Completed.
[root at ibm-p10-01-lp45 makedumpfile]# ./makedumpfile --num-threads=2 -d
31 -l ~/vmcore /tmp/out4
Copying data                                      : [100.0 %] |
   eta: 0s
Copying data                                      : [100.0 %] \
   eta: 0s

The dumpfile is saved to /tmp/out4.

makedumpfile Completed.
[root at ibm-p10-01-lp45 makedumpfile]# cd /tmp
[root at ibm-p10-01-lp45 tmp]# cmp out3 out4
out3 out4 differ: byte 23948282, line 151739

Looks to me the O0/O2 have no difference for this case. If no problem,
the /tmp/outX generated from both single/multi thread should be
exactly the same, however the cmp reports there are differences. With
the v2 patch applied, there is no such difference:

[root at ibm-p10-01-lp45 makedumpfile]# ./makedumpfile -d 31 -l ~/vmcore /tmp/out5
Copying data                                      : [100.0 %] /
   eta: 0s

The dumpfile is saved to /tmp/out5.

makedumpfile Completed.
[root at ibm-p10-01-lp45 makedumpfile]# ./makedumpfile --num-threads=2 -d
31 -l ~/vmcore /tmp/out6
Copying data                                      : [100.0 %] |
   eta: 0s
Copying data                                      : [100.0 %] \
   eta: 0s

The dumpfile is saved to /tmp/out6.

makedumpfile Completed.
[root at ibm-p10-01-lp45 makedumpfile]# cmp /tmp/out5 /tmp/out6
[root at ibm-p10-01-lp45 makedumpfile]#

Thanks,
Tao Liu
>
> Thanks,
> Kazu