[LinuxBIOS] selection of compression algorithm

[Carl-Daniel Hailfinger]

Hi,

what was the motivation to select nrv2b for compression?
Speed? Size of decompressor? Coding style?

Size   File
843910 vmlinux.bin
437684 vmlinux.bin.nrv2b
430890 vmlinux.bin.gz
414679 vmlinux.bin.bz2
357609 vmlinux.bin.7z

If the 7z decompressor is less than 80 kb bigger than
the nrv2b decompressor, would it make sense to switch?

Regards,
Carl-Daniel

[Ronald G Minnich]

Carl-Daniel Hailfinger wrote:
> Hi,
>
> what was the motivation to select nrv2b for compression?
> Speed? Size of decompressor? Coding style?
>
> Size   File
> 843910 vmlinux.bin
> 437684 vmlinux.bin.nrv2b
> 430890 vmlinux.bin.gz
> 414679 vmlinux.bin.bz2
> 357609 vmlinux.bin.7z
>
> If the 7z decompressor is less than 80 kb bigger than
> the nrv2b decompressor, would it make sense to switch?

It's fine with me. As i said, I had a gzip decompressor in early 
linuxbios, and we can look at getting it back. I just stole the one from 
the kernel. nrv2b is nice, I understand, for the romcc section where you 
don't have many variables to play with, but it's been so long since I 
looked at the gunzip code that I have no idea if this is true.

ron

[Stefan Reinauer]

* Ronald G Minnich  [060522 07:22]:
> It's fine with me. As i said, I had a gzip decompressor in early 
> linuxbios, and we can look at getting it back. I just stole the one from 
> the kernel. nrv2b is nice, I understand, for the romcc section where you 
> don't have many variables to play with, but it's been so long since I 
> looked at the gunzip code that I have no idea if this is true.

We never actually compile nrv2b with romcc, do we? At the point we run
nrv2b ram has to be enabled anyways - we uncompress the code somewhere.
So our limits are far bigger than that.

[yhlu]

in crt0.s, and it is used to uncompress linuxbios_ram when using romcc.

YH

On 5/22/06, Stefan Reinauer  wrote:
> Ronald G Minnich  [060522 07:22]:
> > It's fine with me. As i said, I had a gzip decompressor in early
> > linuxbios, and we can look at getting it back. I just stole the one from
> > the kernel. nrv2b is nice, I understand, for the romcc section where you
> > don't have many variables to play with, but it's been so long since I
> > looked at the gunzip code that I have no idea if this is true.
>
> We never actually compile nrv2b with romcc, do we? At the point we run
> nrv2b ram has to be enabled anyways - we uncompress the code somewhere.
> So our limits are far bigger than that.

[Carl-Daniel Hailfinger]

> On 5/22/06, Stefan Reinauer  wrote:
> > We never actually compile nrv2b with romcc, do we? At the point we run
> > nrv2b ram has to be enabled anyways - we uncompress the code somewhere.
> > So our limits are far bigger than that.

yhlu wrote:
> in crt0.s, and it is used to uncompress linuxbios_ram when using romcc.

You mean copy_and_run() is compiled with romcc? Can we change that?

What's the difference between CONFIG_COMPRESSED_ROM_STREAM and
CONFIG_COMPRESS? Can we use lzma compression for both?

Regards,
Carl-Daniel

[Stefan Reinauer]

* Carl-Daniel Hailfinger  [060522 21:29]:
> > On 5/22/06, Stefan Reinauer  wrote:
> > > We never actually compile nrv2b with romcc, do we? At the point we run
> > > nrv2b ram has to be enabled anyways - we uncompress the code somewhere.
> > > So our limits are far bigger than that.
>
> yhlu wrote:
> > in crt0.s, and it is used to uncompress linuxbios_ram when using romcc.
>
> You mean copy_and_run() is compiled with romcc? Can we change that?

No, it is not. But src/arch/i386/init/crt0.S.lb contains another
assembly written version of an nrv2b decompressor.

crt0.s should be changed to call the C version of unrv2b.

Then it can easily be changed later to use any other algorithm.

Using assembly at this point even though ram is enabled already must be
some code fragment from old times when we did a lot more assembly.

This might require though that we have

1. auto.c stage           (romcc)
2. uncompression stage    (gcc)
3. compressed linuxbios stage (gcc, initially compressed)

This reads ugly in a mail but is by no means uglier than keeping
assembler versions of decompressors in the code.
> What's the difference between CONFIG_COMPRESSED_ROM_STREAM and
> CONFIG_COMPRESS? Can we use lzma compression for both?

CONFIG_COMPRESSED_ROM_STREAM says: the payload is compressed.
CONFIG_COMPRESS says: LinuxBIOS is compressed.

[Richard Smith]

> > What's the difference between CONFIG_COMPRESSED_ROM_STREAM and
> > CONFIG_COMPRESS? Can we use lzma compression for both?
>
> CONFIG_COMPRESSED_ROM_STREAM says: the payload is compressed.
> CONFIG_COMPRESS says: LinuxBIOS is compressed.

Perhaps we should rename those to something more indicative of what they do?
Say:

CONFIG_COMPRESSED_ROM_STREAM -> CONFIG_COMPRESSED_PAYLOAD
CONFIG_COMPRESS -> CONFIG_COMPRESSED_LINUXBIOS

[Ronald G Minnich]

Stefan Reinauer wrote:
> auto.c stage           (romcc)
> uncompression stage    (gcc)
> compressed linuxbios stage (gcc, initially compressed)

What I think of as the long term goal for OLPC:
1. auto.c stage: CAR, gunzip gcc code to ram
2. gcc stage: turn on the hardware, gunzip vga bios, other pieces, and 
payload to ram
3. payload runs, may do further gunzip work

the gunzip was removed from V1 because linux had its own decompressor in 
the bzimage. Given that there was, at the time, only one gunzip step, 
using the kernel's gunzip was deemed to save space. On OLPC, which is 
representative of the new systems, there are several distinct compressed 
bits:
1. nrv2b for gcc-compiled linuxbios
2. nrv2b for OLPC VSA binary
3. nrv2b for linux+initrd
4. nrv2b for VGA BIOS

it's clear that the balance of compression is now in linuxbios. So, in 
my view, it's time to move the decompressor back into linuxbios and dump 
the one we use in the kernel. This will help Plan 9 as well. We should 
also move to a better compression algorithm, I think.

I've already modified mkelfimage to just create 3 segments
1. kernel
2. cmdline+params (at 0x90000)
3. initrd (at 0x8000000)

So this is no real problem.

ron

[Carl-Daniel Hailfinger]

Ronald G Minnich wrote:
> Stefan Reinauer wrote:
> > auto.c stage           (romcc)
> > uncompression stage    (gcc)
> > compressed linuxbios stage (gcc, initially compressed)
>
> What I think of as the long term goal for OLPC:
> auto.c stage: CAR, gunzip gcc code to ram
> gcc stage: turn on the hardware, gunzip vga bios, other pieces, and
>
> payload to ram
> 3. payload runs, may do further gunzip work

Do we really want to gunzip anything in CAR stage? How big is the code
to turn on RAM? Will it really benefit from being compressed? I would
turn on RAM first and only then start uncompression.

At what stage am I allowed to use malloc? gcc stage? payload stage?
Can I abuse the function stack before?

Another thing that bothers me: It seems we include the uncompression code
multiple times. Can we include it only once and make all stages use it?
> the gunzip was removed from V1 because linux had its own decompressor in 
> the bzimage. Given that there was, at the time, only one gunzip step, 
> using the kernel's gunzip was deemed to save space. On OLPC, which is 
> representative of the new systems, there are several distinct compressed 
> bits:
> nrv2b for gcc-compiled linuxbios
> nrv2b for OLPC VSA binary
> nrv2b for linux+initrd
> nrv2b for VGA BIOS

Can you upload all these files somewhere so I can make a size comparison
and work out the best compression parameters for lzma?

I have a (not completely finished) script called squeeze_lzma which will
run lzma with different parameters and figure out which yield the best
compression for the unlimited RAM case and for the small RAM case. So
we can use compression which needs only 5k RAM for decompression in
stage 1 and 60k RAM in stage 2+3.

Regards,
Carl-Daniel

[yhlu]

We at least need two times
1. one is in car stage, and the code is in rom, and it is used to
uncompress linuxbios_ram
2. one is in linuxbios_ram, and it is in ram, it is used to uncompress payload.

car stage, before call copy_and_run, the stack is already in ram. but
code is still in rom. So at this time we need to balance the lzma
benifit on compress linuxbios_ram and it's own code size increase in
rom. also need to set one extra buffer in ram for every ap code.

Hope we can squash car stage code + linuxbios_ram code into 64k again.

YH

[Carl-Daniel Hailfinger]

yhlu wrote:
> We at least need two times
> one is in car stage, and the code is in rom, and it is used to
>
> uncompress linuxbios_ram
> 2. one is in linuxbios_ram, and it is in ram, it is used to uncompress payload.
>
> car stage, before call copy_and_run, the stack is already in ram. but
> code is still in rom. So at this time we need to balance the lzma
> benifit on compress linuxbios_ram and it's own code size increase in
> rom. also need to set one extra buffer in ram for every ap code.

OK, so we have to check which of the two solutions is smaller.
1) lzma decompressor + lzma compressed linuxbios_ram
2) nrv2b decompressor + nrv2b compressed (linuxbios_ram + lzma decompressor)

I just checked this for all abuild targets. The second variant wins
for half of the targets. That brings me to a third option:
3) nrv2b decompressor + nrv2b compressed lzma decompressor + lzma
   compressed linuxbios_ram
nrv2b decompressor needs IIRC 512 bytes, lzma decompressor needs 4096
bytes, so if nrv2b only compresses the lzma decompressor and we assume
~50% savings by compression, option 3 is always 1.5k smaller than option 1
and option 3 is 1.5k-7.3k smaller (2k for OLPC) than option 2.

How difficult would it be for us to build option 3?
> Hope we can squash car stage code + linuxbios_ram code into 64k again.

Depending on the board, that might be difficult. For OLPC, everything fits
nicely into 32k. For most Tyan boards, we suffer from the fact that auto.o
is uncompressed and ~28k alone. That is really horrible. Any reasons for
that?

Regards,
Carl-Daniel

[Stefan Reinauer]

* yhlu  [060523 21:49]:
> one is in linuxbios_ram, and it is in ram, it is used to uncompress
>
> payload.
>
> car stage, before call copy_and_run, the stack is already in ram. but
> code is still in rom. So at this time we need to balance the lzma
> benifit on compress linuxbios_ram and it's own code size increase in

We have to be careful here, indeed.
> rom. also need to set one extra buffer in ram for every ap code.

Can't the APs share the code?

[yhlu]

On 5/23/06, Carl-Daniel Hailfinger  wrote:
> Depending on the board, that might be difficult. For OLPC, everything fits
> nicely into 32k. For most Tyan boards, we suffer from the fact that auto.o
> is uncompressed and ~28k alone. That is really horrible. Any reasons for

That is because of romcc can not have function call. with car
cache_as_ram_auto.c you can have that smaller.

YH

[yhlu]

uncompress Code is shared because it is rom. and every CPU has its own
stack on its cache.

YH

On 5/24/06, Stefan Reinauer  wrote:
> yhlu  [060523 21:49]:
> > one is in linuxbios_ram, and it is in ram, it is used to uncompress
> >
> > payload.
> >
> > car stage, before call copy_and_run, the stack is already in ram. but
> > code is still in rom. So at this time we need to balance the lzma
> > benifit on compress linuxbios_ram and it's own code size increase in
>
> We have to be careful here, indeed.
> > rom. also need to set one extra buffer in ram for every ap code.
>
> Can't the APs share the code?

[yhlu]

OK, then We may let linuxbios_ram.rom and linuxbios_apc.rom and vsm code
stay with nrv2b if there is more overhead with lzma in CAR stage.

YH

On 5/23/06, Carl-Daniel Hailfinger  wrote:
> yhlu wrote:
> > We at least need two times
> > one is in car stage, and the code is in rom, and it is used to
> >
> > uncompress linuxbios_ram
> > 2. one is in linuxbios_ram, and it is in ram, it is used to uncompress payload.
> >
> > car stage, before call copy_and_run, the stack is already in ram. but
> > code is still in rom. So at this time we need to balance the lzma
> > benifit on compress linuxbios_ram and it's own code size increase in
> > rom. also need to set one extra buffer in ram for every ap code.
>
> OK, so we have to check which of the two solutions is smaller.
> lzma decompressor + lzma compressed linuxbios_ram
> nrv2b decompressor + nrv2b compressed (linuxbios_ram + lzma decompressor)
>
> I just checked this for all abuild targets. The second variant wins
> for half of the targets. That brings me to a third option:
> 3) nrv2b decompressor + nrv2b compressed lzma decompressor + lzma
>   compressed linuxbios_ram
> nrv2b decompressor needs IIRC 512 bytes, lzma decompressor needs 4096
> bytes, so if nrv2b only compresses the lzma decompressor and we assume
> ~50% savings by compression, option 3 is always 1.5k smaller than option 1
> and option 3 is 1.5k-7.3k smaller (2k for OLPC) than option 2.
>
> How difficult would it be for us to build option 3?
> > Hope we can squash car stage code + linuxbios_ram code into 64k again.
>
> Depending on the board, that might be difficult. For OLPC, everything fits
> nicely into 32k. For most Tyan boards, we suffer from the fact that auto.o
> is uncompressed and ~28k alone. That is really horrible. Any reasons for
> that?
>
> Regards,
> Carl-Daniel
> --
> http://www.hailfinger.org/

[Carl-Daniel Hailfinger]

yhlu wrote:
> uncompress Code is shared because it is rom. and every CPU has its own
> stack on its cache.

So a decompressor is in ROM, reads from ROM, writes to RAM and has its
stack in cache? Or do we uncompress to cache?

How big is that cache? I can modify lzma to only need 6000 bytes stack
and no additional memory allocations.

Are reads from ROM special (alignment problems etc)? Can I treat stack
like normal RAM?

Regards,
Carl-Daniel

[yhlu]

just let the the uncompress take more parameter for the properties
buf, and every cpu find the correct buffer in RAM in BSP should be
OK...because the RAM on first node is ready at that time.

YH

On 5/24/06, Carl-Daniel Hailfinger  wrote:
> yhlu wrote:
> > uncompress Code is shared because it is rom. and every CPU has its own
> > stack on its cache.
>
> So a decompressor is in ROM, reads from ROM, writes to RAM and has its
> stack in cache? Or do we uncompress to cache?
>
> How big is that cache? I can modify lzma to only need 6000 bytes stack
> and no additional memory allocations.
>
> Are reads from ROM special (alignment problems etc)? Can I treat stack
> like normal RAM?
>
> Regards,
> Carl-Daniel
> --
> http://www.hailfinger.org/

[Carl-Daniel Hailfinger]

yhlu wrote:
> just let the the uncompress take more parameter for the properties
> buf, and every cpu find the correct buffer in RAM in BSP should be
> OK...because the RAM on first node is ready at that time.

How about

int uncompress_generic(void *src, void *dst, void *buf, int bufsize);

and uncompress_generic can assume that buf can be treated like RAM.
If bufsize is too small, it can return an error. An additional function

int needed_bufsize(void *src);

could be written to compute the minimum buffer size for a given
compressed image.

What do you think?

Regards,
Carl-Daniel

[Ronald G Minnich]

Carl-Daniel Hailfinger wrote:
> yhlu wrote:
> > just let the the uncompress take more parameter for the properties
> > buf, and every cpu find the correct buffer in RAM in BSP should be
> > OK...because the RAM on first node is ready at that time.
>
> How about
>
> int uncompress_generic(void *src, void *dst, void *buf, int bufsize);
>
> and uncompress_generic can assume that buf can be treated like RAM.
> If bufsize is too small, it can return an error. An additional function
>
> int needed_bufsize(void *src);
>
> could be written to compute the minimum buffer size for a given
> compressed image.
>
> What do you think?
>
> Regards,
> Carl-Danie

int needed_bufsize(void *src);

good. And in most cases it is:
size = *(unsigned long *)src;

int uncompress_generic(void *src, void *dst, void *buf)

if the programmer is too stupid to call needed_bufsize, they're probably 
too stupid to pass in a correct bufsize to this function, I think it is 
not needed.

ron

[yhlu]

int uncompress_generic(void *src, void *dst, void *buf)

==>

int uncompress_generic(void *src, void *dst, uint32_t *ilen, void *op_buf)

YH

On 5/25/06, Ronald G Minnich  wrote:
> Carl-Daniel Hailfinger wrote:
> > yhlu wrote:
> > > just let the the uncompress take more parameter for the properties
> > > buf, and every cpu find the correct buffer in RAM in BSP should be
> > > OK...because the RAM on first node is ready at that time.
> >
> > How about
> >
> > int uncompress_generic(void *src, void *dst, void *buf, int bufsize);
> >
> > and uncompress_generic can assume that buf can be treated like RAM.
> > If bufsize is too small, it can return an error. An additional function
> >
> > int needed_bufsize(void *src);
> >
> > could be written to compute the minimum buffer size for a given
> > compressed image.
> >
> > What do you think?
> >
> > Regards,
> > Carl-Danie
>
> int needed_bufsize(void *src);
>
> good. And in most cases it is:
> size = *(unsigned long *)src;
>
> int uncompress_generic(void *src, void *dst, void *buf)
>
> if the programmer is too stupid to call needed_bufsize, they're probably
> too stupid to pass in a correct bufsize to this function, I think it is
> not needed.
>
> ron