UFC-crypt: ultra fast 'crypt' implementation ============================================ @(#)README 2.27 11 Sep 1996 Design goals/non goals: ---------------------- - Crypt implementation plugin compatible with crypt(3)/fcrypt. - High performance when used for password cracking. - Portable to most 32/64 bit machines. - Startup time/mixed salt performance not critical. Features of the implementation: ------------------------------ - On most machines, UFC-crypt runs 30-60 times faster than crypt(3) when invoked repeated times with the same salt and varying passwords. - With mostly constant salts, performance is about two to three times that of the default fcrypt implementation shipped with Alec Muffets 'Crack' password cracker. For instructions on how to plug UFC-crypt into 'Crack', see below. - With alternating salts, performance is only about twice that of crypt(3). - Tested on 680x0, 386, SPARC, MIPS, HP-PA, Convex, Cray, Pyramid and IBM RS/6000 systems as well as with gcc on IBM PS/2(DOS) and Linux on a 386 PC. - Requires 165 kb for tables. - UFC-crypt is known to have compilation problems with some micro computer C compilers (e.g. Turbo C++) due to its table sizes. Flame the vendors for placing arbitrary limitations on their products. Use & support the GNU C compiler, gcc. - Also includes 'fcrypt16' - the function used with DEC's enhanced security setup. Author & licensing etc ---------------------- UFC-crypt is created by Michael Glad, email: glad@daimi.aau.dk, and has been donated to the Free Software Foundation, Inc. It is covered by the GNU library license version 2, see the file 'COPYING'. NOTES FOR USERS OUTSIDE THE US: ------------------------------ The US government limits the export of DES based software/hardware. This software is written in Aarhus, Denmark. It can therefore be retrieved from ftp sites outside the US without breaking US law. Please do not ftp it from american sites. Installing UFC-crypt -------------------- Installing as GNU crypt Library: **********************(******** Unpack the library in some place. The simplest place is the toplevel directory of the GNU C library sources. Now mention `crypt' as on of the add-ons used for GNU libc. I.e., when the sources are in the toplevel source directory simply run the `configure' script of the *libc* as: configure --enable-add-ons=crypt,XXX where XXX are the names of possibly existing other add-ons. That's all. Ordinary UNIX install ********************* Copy the file Makefile.non-gnu to Makefile Edit Makefile setting the variables CRYPT: The encryption module to use; crypt.o should always work. If compiling for one of the machines for which special support is available, select the appropriate module. CC: The compiler to use. OFLAGS: The highest level of optimization available. Now run 'make'. UFC-crypt is compiled into 'libufc.a'. A test program: ufc is also linked. Try it out: './ufc 1' to test proper operation. For a more thorough test, run 'make tests'. This compiles and invokes a DES validation suite as well as two benchmark programs comparing UFC-crypt with the native crypt(3) implementation. If your friendly vendor has omitted crypt(3) from libc, compilation of the native benchmark program 'speedc' will fail. Compilation of the 'speed*' programs may also fail as they use timer facilities not present in the same form in all UNIX implementations. If so is, you may choose to run '/bin/time ./ufc 10000' and divide the CPU time used by 10000. 'libufc.a' can be linked into your applications. It is compatible with both crypt(3) and the fcrypt shipped with Alec Muffett's Crack program. DOS install *********** UFC-crypt compiles under DOS at least with MSC 6.00A. To compile, move 'patchlevel.h' to 'pl.h', 'ufc-crypt.h' to 'ufc.h' and 'crypt_util.c' to 'cryptutl.c'. When compiling, define the macro 'DOS' ( -DDOS ). You should compile & link the programs by hand. Refer to the Makefile to see which modules is needed for a program. You should be able to compile & run the verification programs 'cert' and 'ufc' but not the benchmark program 'speedf'. For benchmarking, measure the running time of 'ufc 10000'. Having compiled the modules, you can link 'cryptutl.o' and 'crypt.o' into your application programs or put them into a library. Installing UFC-crypt into Crack: ******************************* Crack Release 4.0a: in 'Sources/Makefile', change the CRACKCRYPT macro to a path leading to 'libufc.a' and invoke the Crack script as usual. 4.1 and later: Crack knows about UFC-crypt. Refer to the Crack docs for instructions. Benchmark table: --------------- The table shows how many operations per second UFC-crypt can do on various machines. |--------------|-------------------------------------------| |Machine | SUN* SUN* HP* DecStation HP | | | 3/50 ELC 9000/425e 3100 9000/720 | |--------------|-------------------------------------------| | Crypt(3)/sec | 4.6 30 15 25 57 | | Ufc/sec | 220 990 780 1015 3500 | |--------------|-------------------------------------------| | Speedup | 48 30 52 40 60 | |--------------|-------------------------------------------| *) Compiled using special assembly language support module. It seems as if performance is limited by CPU bus and data cache capacity. This also makes the benchmarks debatable compared to a real test with UFC-crypt wired into Crack. However, the table gives an outline of what can be expected. Optimizations: ------------- Here are the optimizations used relative to an ordinary implementation such as the one said to be used in crypt(3). Major optimizations ******************* - Keep data packed as bits in integer variables -- allows for fast permutations & parallel xor's in CPU hardware. - Let adjacent final & initial permutations collapse. - Keep working data in 'E expanded' format all the time. - Implement DES 'f' function mostly by table lookup - Calculate the above function on 12 bit basis rather than 6 as would be the most natural. - Implement setup routines so that performance is limited by the DES inner loops only. - Instead of doing salting in the DES inner loops, modify the above tables each time a new salt is seen. According to the BSD crypt code this is ugly :-) Minor (dirty) optimizations *************************** - combine iterations of DES inner loop so that DES only loops 8 times. This saves a lot of variable swapping. - Implement key access by a walking pointer rather than coding as array indexing. - As described, the table based f function uses a 3 dimensional array: sb ['number of 12 bit segment']['12 bit index']['48 bit half index'] Code the routine with 4 (one dimensional) vectors. - Design the internal data format & uglify the DES loops so that the compiler does not need to do bit shifts when indexing vectors. Portability issues ****************** UFC-crypt does not need to know the byte endianness of the machine is runs on. To speed up the DES inner loop, it does a dirty trick requiring the availability of a integer flavoured data type occupying exactly 32 (or 64) bits. This is normally the case of 'long'. The header file 'ufc-crypt.h' contains typedefs for this type. If you have to change it (or any other part) to get things working, please drop me a note. UFC-crypt can take advantage of 64 bit integers. At the moment, it is configured to do so automatically for Convex & Cray machines. Revision history **************** UFC patchlevel 0: base version; released to alt.sources on Sep 24 1991 UFC patchlevel 1: patch released to alt.sources on Sep 27 1991. No longer rebuilds sb tables when seeing a new salt. UFC-crypt pl0: Essentially UFC pl 1. Released to comp.sources.misc on Oct 22 1991. UFC-crypt pl1: Released to comp.sources.misc in march 1992 * setkey/encrypt routines added * added validation/benchmarking programs * reworked keyschedule setup code * memory demands reduced * 64 bit support added