Re: [PATCH 1/4] zstd: pass pointer rathen than structure to functions

From: David Sterba
Date: Wed Jun 05 2019 - 08:36:06 EST


On Wed, Jun 05, 2019 at 01:57:03PM +0200, David Sterba wrote:
> On Tue, Jun 04, 2019 at 03:43:26PM -0700, Andrew Morton wrote:
> > On Mon, 3 Jun 2019 14:32:03 +0530 Maninder Singh <maninder1.s@xxxxxxxxxxx> wrote:
> >
> > > currently params structure is passed in all functions, which increases
> > > stack usage in all the function and lead to stack overflow on target like
> > > ARM with kernel stack size of 8 KB so better to pass pointer.
> > >
> > > Checked for ARM:
> > >
> > > Original Patched
> > > Call FLow Size: 1264 1040
> > > ....
> > > (HUF_sort) -> 296
> > > (HUF_buildCTable_wksp) -> 144
> > > (HUF_compress4X_repeat) -> 88
> > > (ZSTD_compressBlock_internal) -> 200
> > > (ZSTD_compressContinue_internal)-> 136 -> 88
> > > (ZSTD_compressCCtx) -> 192 -> 64
> > > (zstd_compress) -> 144 -> 96
> > > (crypto_compress) -> 32
> > > (zcomp_compress) -> 32
> > > ....
> > >
> > > Signed-off-by: Maninder Singh <maninder1.s@xxxxxxxxxxx>
> > > Signed-off-by: Vaneet Narang <v.narang@xxxxxxxxxxx>
> > >
> >
> > You missed btrfs. This needs review, please - particularly the
> > kernel-wide static ZSTD_parameters in zstd_get_btrfs_parameters().
>
> >
> > The base patch is here:
> >
> > http://lkml.kernel.org/r/1559552526-4317-2-git-send-email-maninder1.s@xxxxxxxxxxx
> >
> > --- a/fs/btrfs/zstd.c~zstd-pass-pointer-rathen-than-structure-to-functions-fix
> > +++ a/fs/btrfs/zstd.c
> > @@ -27,15 +27,17 @@
> > /* 307s to avoid pathologically clashing with transaction commit */
> > #define ZSTD_BTRFS_RECLAIM_JIFFIES (307 * HZ)
> >
> > -static ZSTD_parameters zstd_get_btrfs_parameters(unsigned int level,
> > +static ZSTD_parameters *zstd_get_btrfs_parameters(unsigned int level,
> > size_t src_len)
> > {
> > - ZSTD_parameters params = ZSTD_getParams(level, src_len, 0);
> > + static ZSTD_parameters params;
>
> > +
> > + params = ZSTD_getParams(level, src_len, 0);
>
> No thats' broken, the params can't be static as it depends on level and
> src_len. What happens if there are several requests in parallel with
> eg. different levels?
>
> Would be really great if the mailinglist is CCed when the code is
> changed in a non-trivial way.

So this does not compile fs/btrfs/zstd.o which Andrew probably found
too, otherwise btrfs is the only in-tree user of the function outside of
lib/ and crypto/.

I think that Nick Terrell should have been CCed too, as he ported zstd
to linux.