On Thu, Nov 25, 2021 at 12:30 AM Alex Xu (Hello71) <alex_y_xu@xxxxxxxx> wrote: > > Otherwise, it allocates 2 GB of memory at once. Even though the majority > of this memory is never touched, the default heuristic overcommit > refuses this request if less than 2 GB of RAM+swap is currently > available. This results in "zstd: error 11 : Allocation error : not > enough memory" and the kernel failing to build. > > When the size is specified, zstd will reduce the memory request > appropriately. For typical kernel sizes of ~32 MB, the largest mmap > request will be reduced to 512 MB, which will succeed on all but the > smallest devices. > > For inputs around this size, --stream-size --no-content-size may > slightly decrease the compressed size, or slightly increase it: > https://github.com/facebook/zstd/issues/2848. > > Signed-off-by: Alex Xu (Hello71) <alex_y_xu@xxxxxxxx> The reason why we need this workaround is just because we do "cat and compress". zstd must allocate a huge memory beforehand since it cannot predict how long the stream it will receive. If zstd is given with a file name, it can fstat it to know its file size and allocate the minimal amount of memory. This is my test. I used 'ulimit' to set the upper limit of the memory the zstd can use. [test steps] # Create a 1kB file $ truncate --size=1k dummy # Set the memory size limit to 10MB $ ulimit -S -v 10240 # Pass the file as a argument; success $ zstd -19 -o dummy.zst dummy dummy : 2.15% ( 1024 => 22 bytes, dummy.zst) # cat and zstd; fail $ cat dummy | zstd -19 > dummy.zst zstd: error 11 : Allocation error : not enough memory # cat and zstd --stream-size; success $ cat dummy | zstd -19 --stream-size=1024 > dummy.zst scripts/Makefile.modinst was written in such a way that zstd can know the file size by itself. cmd_zstd = $(ZSTD) -T0 --rm -f -q $< We cannot rewrite scripts/Makefile.lib in that way because arch/x86/boot/compress/Makefile concatenates two files before compression. And this is the only use-case of this feature. So, I am seriously considering to revert this commit: commit d3dd3b5a29bb9582957451531fed461628dfc834 Author: H. Peter Anvin <hpa@xxxxxxxxx> Date: Tue May 5 21:17:15 2009 -0700 kbuild: allow compressors (gzip, bzip2, lzma) to take multiple inputs With that commit reverted, zstd will take a single input file, and we can do "zstd -o <output> <input>". So, I will take some time to investigate that approach. > --- > scripts/Makefile.lib | 12 ++++++++++-- > 1 file changed, 10 insertions(+), 2 deletions(-) > > diff --git a/scripts/Makefile.lib b/scripts/Makefile.lib > index ca901814986a..c98a82ca38e6 100644 > --- a/scripts/Makefile.lib > +++ b/scripts/Makefile.lib > @@ -466,12 +466,20 @@ quiet_cmd_xzmisc = XZMISC $@ > # single pass, so zstd doesn't need to allocate a window buffer. When streaming > # decompression is used, like initramfs decompression, zstd22 should likely not > # be used because it would require zstd to allocate a 128 MB buffer. > +# > +# --stream-size to reduce zstd memory usage (otherwise zstd -22 --ultra > +# allocates, but does not use, 2 GB) and potentially improve compression. > +# > +# --no-content-size to save three bytes which we do not use (we use size_append). > + > +# zstd --stream-size is only supported since 1.4.4 > +zstd_stream_size = $(shell $(ZSTD) -1c --stream-size=0 --no-content-size </dev/null >/dev/null 2>&1 && printf '%s' '--stream-size=$(total_size) --no-content-size') > > quiet_cmd_zstd = ZSTD $@ > - cmd_zstd = { cat $(real-prereqs) | $(ZSTD) -19; $(size_append); } > $@ > + cmd_zstd = { cat $(real-prereqs) | $(ZSTD) $(zstd_stream_size) -19; $(size_append); } > $@ > > quiet_cmd_zstd22 = ZSTD22 $@ > - cmd_zstd22 = { cat $(real-prereqs) | $(ZSTD) -22 --ultra; $(size_append); } > $@ > + cmd_zstd22 = { cat $(real-prereqs) | $(ZSTD) $(zstd_stream_size) -22 --ultra; $(size_append); } > $@ > > # ASM offsets > # --------------------------------------------------------------------------- > -- > 2.34.0 > -- Best Regards Masahiro Yamada
