On Fri, Jun 5, 2020 at 12:57 PM Eric Biggers <[email protected]> wrote:
>
> On Fri, Jun 05, 2020 at 12:42:59PM -0700, Nick Desaulniers wrote:
> > On Fri, Jun 5, 2020 at 12:34 PM Eric Biggers <[email protected]> wrote:
> > >
> > > On Fri, Jun 05, 2020 at 09:45:43AM -0700, Nick Desaulniers wrote:
> > > > On Fri, Jun 5, 2020 at 9:08 AM Eric Biggers <[email protected]> wrote:
> > > > >
> > > > > On Fri, Jun 05, 2020 at 07:55:46AM -0700, Jaegeuk Kim wrote:
> > > > > > Eric,
> > > > > >
> > > > > > Could you PTAL?
> > > > > >
> > > > > > On 06/05, kernel test robot wrote:
> > > > > > > tree:   
> > > > > > > https://git.kernel.org/pub/scm/linux/kernel/git/jaegeuk/f2fs.git 
> > > > > > > dev-test
> > > > > > > head:   adf3d3a53cf13d0998c699ba43d8582c875179e3
> > > > > > > commit: adf3d3a53cf13d0998c699ba43d8582c875179e3 [48/48] f2fs: 
> > > > > > > don't return vmalloc() memory from f2fs_kmalloc()
> > > > > > > config: powerpc64-randconfig-r011-20200605 (attached as .config)
> > > > > > > compiler: clang version 11.0.0 
> > > > > > > (https://github.com/llvm/llvm-project 
> > > > > > > ac47588bc4ff5927a01ed6fcd269ce86aba52a7c)
> > > > > > > reproduce (this is a W=1 build):
> > > > > > >         wget 
> > > > > > > https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross
> > > > > > >  -O ~/bin/make.cross
> > > > > > >         chmod +x ~/bin/make.cross
> > > > > > >         # install powerpc64 cross compiling tool for clang build
> > > > > > >         # apt-get install binutils-powerpc64-linux-gnu
> > > > > > >         git checkout adf3d3a53cf13d0998c699ba43d8582c875179e3
> > > > > > >         # save the attached .config to linux build tree
> > > > > > >         COMPILER_INSTALL_PATH=$HOME/0day COMPILER=clang 
> > > > > > > make.cross ARCH=powerpc64
> > > > > > >
> > > > > > > If you fix the issue, kindly add following tag as appropriate
> > > > > > > Reported-by: kernel test robot <[email protected]>
> > > > >
> > > > > I don't know what's causing this.  It doesn't look at all related to 
> > > > > my commit,
> > > > > and I don't see what's using so much stack in f2fs_fill_super().
> > > > >
> > > > > @kernel test robot: the directions given above don't actually work.
> > > > > First I got an error due to powerpc64-linux-gnu-elfedit not existing.
> > > > > So I had to build binutils for powerpc64 myself.
> > > > >
> > > > > Then I still got an error:
> > > > >
> > > > >         make: *** No rule to make target 'arch/powerpc64/Makefile'.  
> > > > > Stop.
> > > >
> > > > If you have the config, then
> > > > $ ARCH=powerpc CROSS_COMPILE=powerpc64-linux-gnu- make CC=clang -j71
> > > > If you recompile with CONFIG_DEBUG_INFO, you can get the stack frame
> > > > information. I wrote a tool to parse this for these cryptic warnings.
> > > > https://github.com/ClangBuiltLinux/frame-larger-than
> > >
> > > I can build the file and get the warning now, but the frame_larger_than.py
> > > script doesn't work:
> > >
> > > $ ARCH=powerpc CROSS_COMPILE=powerpc64-linux-gnu- make CC=clang 
> > > fs/f2fs/super.o
> > >   CALL    scripts/checksyscalls.sh
> > >   CALL    scripts/atomic/check-atomics.sh
> > >   CC [M]  fs/f2fs/super.o
> > > fs/f2fs/super.c:3303:12: warning: stack frame size of 2064 bytes in 
> > > function 'f2fs_fill_super' [-Wframe-larger-than=]
> > > static int f2fs_fill_super(struct super_block *sb, void *data, int silent)
> > >            ^
> > > 1 warning generated.
> > >
> > > $ python3 ~/src/frame-larger-than/frame_larger_than.py fs/f2fs/super.o 
> > > f2fs_fill_super
> > > failed to parse elf: Unsupported relocation type: 1
> >
> > Looks like the python library I'm using to parse the DWARF debug info
> > doesn't understand some ppc64 specific relocation kind, and is
> > throwing an ELFError.  Let me report it upstream.
> >
> > The hard part for these is inlining; it can be hard to tell which
> > child has a large local allocation when looking at the warning about
> > the parent.
> >
> > Then again, my script is just nicer output than `llvm-dwarfdump foo.o`
> > or `readelf --debug-dump=info foo.o` for this specific issue.
> >
>
> I did confirm that my commit "f2fs: don't return vmalloc() memory from
> f2fs_kmalloc()" actually caused this, in particular the following part:
>
> diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
> index a71da699cb2d55..f3c15116954218 100644
> --- a/fs/f2fs/super.c
> +++ b/fs/f2fs/super.c
> @@ -3033,7 +3033,7 @@ static int init_blkz_info(struct f2fs_sb_info *sbi, int 
> devi)
>         if (nr_sectors & (bdev_zone_sectors(bdev) - 1))
>                 FDEV(devi).nr_blkz++;
>
> -       FDEV(devi).blkz_seq = f2fs_kzalloc(sbi,
> +       FDEV(devi).blkz_seq = f2fs_kvzalloc(sbi,
>                                         BITS_TO_LONGS(FDEV(devi).nr_blkz)
>                                         * sizeof(unsigned long),
>                                         GFP_KERNEL);
>
> This is inlined as:
>
>     f2fs_fill_super()
>      => f2fs_scan_devices()
>         => init_blkz_info()
>            => f2fs_kvzalloc()
>
> Somehow this change increased the frame size of f2fs_fill_super() from 1984 to
> 2064 bytes.  That exceeds the 2048-byte limit, triggering the warning.
>
> So, f2fs_fill_super() was very close to the limit already.  I'm not sure why; 
> I
> don't see any huge stack allocations in it or in static functions it calls.
> Maybe there is a kconfig option that is increasing the stack usage?
>
> If I add noinline_for_stack to f2fs_scan_devices(), the frame drops to 1744
> bytes.  Or if I add noinline_for_stack to read_raw_super_block(), the frame
> drops to 1776 bytes.  We could do either (or both) of those to avoid the
> warning.  But we'd still wouldn't be that far from this 2048-byte "limit".

Right, so my script would have printed out the list of all local
variables in f2fs_fill_super() and their sizes.  With that information
handy, we could assess if there were any smoking guns of clearly
incorrect large stack allocations vs death by a thousand cuts.  Your
change may not have added a new large local allocation, simply tipped
the scale or changed inlining decisions.  They may be other local
variables in this call chain that we should reassess allocation
strategy; ie. dynamic or static rather than local, to avoid the
potential for exhausting kernel stack.
-- 
Thanks,
~Nick Desaulniers


_______________________________________________
Linux-f2fs-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel

Reply via email to