On Thu, 19 Sep 2013 05:00:02 +0000 Atsushi Kumagai <kumagai-atsushi at mxc.nes.nec.co.jp> wrote: > (2013/09/19 11:13), HATAYAMA Daisuke wrote: > > (2013/09/18 12:00), Atsushi Kumagai wrote: > >> (2013/08/30 10:33), HATAYAMA Daisuke wrote: > >>> (2013/08/29 7:08), Cliff Wickman wrote: > >>>> From: Cliff Wickman <cpw at sgi.com> > >>>> > >>>> I am submitting 6 patches that I have found helpful in speeding the dump > >>>> process or clarifying the progress report. > >>>> They are not a series, and should not be interdependent. But if you > >>>> find any dependencies I apply them in this order: > >>>> [PATCH] makedumpfile: reverse -c and -p if using snappy compression > >>>> [PATCH] makedumpfile: use non-cyclic when possible > >>>> [PATCH] makedumpfile: shorten cyclic exclude-unnecessary passes > >>>> [PATCH] makedumpfile: shorten cyclic unnecessary-page scans > >>>> [PATCH] makedumpfile: show needed memory > >>>> [PATCH] makedumpfile: search for a debug vmlinux > >>>> > >>>> The last one (search for a debug vmlinux) is useful in identifying huge > >>>> pages with the PG_head/PG-tail flags. There was a patch from Petr Tesarik > >>>> that enables that huge page filtering. I don't think you are taking that one > >>>> as-is, but are reworking it. Seems like Hatayama-san was doing that work. > >>> > >>> No. If I have good memory, Kumagai-san was investigating how to integrate > >>> huge page filtering into current memory types currently supported by > >>> makedumpfile. > >> > >> Yes, I was investigating it but I'm not working for it now. > >> > >> I think main features should work without vmlinux, > >> but it was impossible about his patch as said by himself: > >> > >>> This patch depends on exporting the relevant PG_* flags from the > >>> kernel (in VMCOREINFO), and that's where I got stuck, because depending > >>> on the number of available bits for the page flags, the kernel either > >>> has PG_head and PG_tail, or only PG_compound, so I needed a #ifdef, and > >>> the kernel maintainers didn't like the conditional. > >>> I can restart the discussion with kernel maintainers and see what I can do > >> > >> Therefore, I waited that his work is finished and I was going to continue > >> my work, but I didn't say my thinking definitely, sorry. > >> > >> Anyway, I should cooperate with Petr to develop huge page filtering, > >> so could you let me know the status of your work ? > >> > > > > To whom do you make this question? It seems Cliff according to content of this question. > > If so you should resend. If to me, I have not done huge page filtering work at all so far > > as I already said in the previous mail in this thread. > > Sorry for confusing you, I asked Petr. > According to Cliff's 1st mail: > > > From: Cliff Wickman <cpw at sgi.com> > > > > This fix comes from Petr Tesarik <ptesarik at suse.cz>, and was applied > > to the SuSE version of makedumpfile. > > I don't see it in the mmap version. > > It is important to filtering out a great deal of user memory. > > > > But this from Petr on 5/15: > >> This patch depends on exporting the relevant PG_* flags from the > >> kernel (in VMCOREINFO), and that's where I got stuck, because depending > >> on the number of available bits for the page flags, the kernel either > >> has PG_head and PG_tail, or only PG_compound, so I needed a #ifdef, and > >> the kernel maintainers didn't like the conditional. > >> I can restart the discussion with kernel maintainers and see what I can do > > So I'm not sending to the kexec list until he's ready. > > I understood Petr tried to add PG_* flags into VMCOREINFO, > but it seems not completed in linux-next at this time. > To prepare the flags is important to huge page filtering, > I want to know the status of Petr's work. This got blocked when the maintainer of that subsystem didn't answer me what he considered a clean solution. And then it got sidetracked by other issues. I can resume from there, because I also need these flags properly exported in VMCOREINFO. OK, time to ask again. Petr T
