Hi Glenn.
I analysed BSS size gzip applet and found may be mistake:
updcrc() checking if (crc_table_empty) but not resetted this var.
This do make slow gzip applet ;-)
--w
vodz
Use the old fork() method of tar compression support, rather than
read_bz2....
- (*uncompress)(int in, int out) seems like a more natural interface
for compression code.
- it might improve performance by seperating the work into one cpu
bound and one io bound process.
- There is extra code required to do read_[gz|bunzip] since (*uncompress)(int in,
int out) will normally be used by the standalone compression applet.
There have been problems with this method so if you see a "Short read"
error let me know.
This hides a bug related to the new bunzip code in the tar and dpkg[-deb]
applets.
It will also reduce compile time a little as some unused files wont be
compiled.
the busybox menuconfig triggered my "inacceptable number of spelling mistakes"
upper level, so I decided to make a patch ;-)
I also improved some wording to describe some things in a better way.
Many thanks for an incredible piece of software!
Andreas Mohr, random OSS developer
Hello Rob,
Here's a patch to your bunzip-3.c file. Nice work btw.
One minor bug fix... checking for error return when read()ing.
Some size/performance optimizations as well. One instance of
memset() seems unnecssary. You might want to take a look.
Anyway, on my machine, decompressing linux-2.6.0-test7.tar.bz2
to /dev/null gave the following times:
bunzip-3.c bzcat (system) bunzip-3.c (patched)
real 0m24.420s 0m22.725s 0m20.701s
user 0m23.930s 0m22.170s 0m20.180s
sys 0m0.070s 0m0.080s 0m0.140s
Size of the patched version is comparable (slightly larger or
smaller depending on compiler flags).
Manuel
The API for using partial writes, as described in my last message, sucked.
So here's a patch against my last patch that changes things so that
write_bunzip_data calls read_bunzip_data itself behind the scenes whenever
necessary. So usage is now just start_bunzip(), write_bunzip_data() until it
returns a negative number, and then the cleanup at the end of
uncompressStream.
It adds 32 bytes to the executable, but it should allow the caller (tar) to be
simplified enough to compensate. Total -Os stripped exe size now 6856 bytes.
Rob
P.S. I attached the whole C file so you don't have to keep incremental
patches straight if you don't want to. :)
P.S. In the version I'm banging on now, I've simplified the license to just
LGPL. I read the OSL a bit more closely and the patent termination clause
would have bit IBM in their counter-suit of SCO if the code in question had
been OSL instead of GPL, and I've decided I just don't want to beta-test
legal code right now.
Need to chdir after the tar file is opened, so make common tar filename
parsing and send the file descriptor rather than filename to
writeTarFile.
Modify the verboseFlag operation to determine wether to display on
stderr or stdout at display time, simpler than doing it in tar_main.
The tar -x command in busybox does not restore the file mode correctly.
The reason is most probably this code in
archival/libunarachive/data_extract_all.c:
chmod(file_header->name, file_header->mode);
chown(file_header->name, file_header->uid, file_header->gid);
chown clears the set*id bits (on current versions of linux :). Flipping
the order around fixes the problem.
(tested with 1.00pre3 from cvs).
directory.
From http://www.gnu.org/manual/tar/html_node/tar_123.html
REGTYPE
AREGTYPE
These flags represent a regular file. In order to be compatible with
older versions of tar, a typeflag value of AREGTYPE should be silently
recognized as a regular file. New archives should be created using
REGTYPE. Also, for backward compatibility, tar treats a regular file
whose name ends with a slash as a directory.
Always preserve creation date
Disable the -p option its for modification date
Remove some cpio header debugging noise
Syncronise file listing behaviour with upstream.
If we read the contents of compressed files within the ar archive,
e.g. control.tar.gz, then file position gets all out of whack, so
it has to be reset before reading thenext header.
Hello all,
This patch adds more "Help" text to the config system. Almost
all applets now have a help entry. Also, I cleaned up the spacing of
the existing text so that things are consistent. This patch is against
this morning's CVS.
Thomas Cameron
CEI Systems, Inc.
bb-tar "cjf" does not create a valid tbz2-archive -- if fact the result is a
plain tar-file (no compression) -- but does not warn about the unrecognized
parameter combination "cj" (bb does not have bzip2-compression yet, right?).
to fix this I have added an error message stating this does not work.
He also reported
cosmetic: versose "-v" does not show any output when used with "create"
which I have now fixed as well.
-Erik
He took a look into the recent reports of tar problems, and found an obvious
typo in last_patch91 from vodz which converted tar to use bb_getopt_ulflags.
complain "unknown file type" if it tries to extract an oldgnu tar file
and TAR_FEATURE_OLDGNU_COMPATABILITY sint defined.
Print a warning if unisupported gnu extensions are encountered.
modified Kbuild system I put into uClibc. With this, there should be no more
need to modify Rules.mak since I've moved all the interesting options into the
config system. I think I've got everything updated, but you never know, I may
have made some mistakes, so watch closely.
-Erik
It seems that under busybox unstable, "tar -c -f - blabla" create
a tar file named "-" instead of writing to stdout.
The included patch should fix this.
New complex patch for decrease size devel version. Requires previous patch.
Also removed small problems from dutmp and tar applets.
Also includes vodz' last_patch61_2:
Last patch correcting comment for #endif and more integrated
with libbb (very reduce size if used "cat" applet also).
Requires last_patch61 for modutils/config.in.
Hi, Erik.
my_getpw(uid/gid) and applets used it have problem:
if username for uid not found, applets can`t detect it
(but code pessent). Also "%8ld " format is bad:
spaces not required (applets have self format
or spec format (tar applet) and overflow for "id" applet...)
This problem also pressent in stable version.
Patch for unstable in attach.
--w
vodz
In most cases, dirname returns the same argument it was given, so this code
works nice, but there's one special case: when the name contains no
dirname, it returns "." (stored statically in the body of itself), and we
get a segfault in attempt to free() it.
This patch fixes this problem.
Also tried to clean up the logic a little, and ensure that read errors
or invalid archives resulted in error returns. This could use a lot more
work... Volunteers?
#49: I found one memory overflow and memory leak in "ln" applet.
Last patch reduced also 54 bytes. ;)
#50: I found bug in loginutils/Makefile.in.
New patch have also new function to libbb and
aplied this to applets and other cosmetic changes.
still needs major attention (i.e. removal).
Removed references to uninitialized variables like progName,
smallMode, noisy, etc. Remove functions and code for handling "small"
decompression mode, since it is all unreachable. Remove
total_{in,out}* counters - they are never used. Remove panic()
function and assert_h and their uses because they are all for "should
never happen" circumstances. Replace internal malloc/free wrappers
with xmalloc and free. Remove conditional in if(foo)free(foo);
situations. Remove bogus
if (sizeof(int) != 4) {
return BZ_CONFIG_ERROR;
}
(...etc...) code and code for handling BZ_CONFIG_ERROR. Someone should
go through and change the applet to use well-defined types when
appropriate - it expects sizeof(short)==2 && sizeof(int)==4. Until
this commit these were explicitly checked for, and the applet would
exit if these types were not the right size. (I think this is wrong
even as an interim solution.)
With gcc 3.1 reduces size of binary on i386-linux by about 700 bytes.
Lightly tested.
