From 849369d6c66d3054688672f97d31fceb8e8230fb Mon Sep 17 00:00:00 2001 From: root Date: Fri, 25 Dec 2015 04:40:36 +0000 Subject: initial_commit --- Documentation/crypto/api-intro.txt | 248 ++++++++++++++++++++++ Documentation/crypto/async-tx-api.txt | 226 ++++++++++++++++++++ Documentation/crypto/descore-readme.txt | 352 ++++++++++++++++++++++++++++++++ 3 files changed, 826 insertions(+) create mode 100644 Documentation/crypto/api-intro.txt create mode 100644 Documentation/crypto/async-tx-api.txt create mode 100644 Documentation/crypto/descore-readme.txt (limited to 'Documentation/crypto') diff --git a/Documentation/crypto/api-intro.txt b/Documentation/crypto/api-intro.txt new file mode 100644 index 00000000..8b493027 --- /dev/null +++ b/Documentation/crypto/api-intro.txt @@ -0,0 +1,248 @@ + + Scatterlist Cryptographic API + +INTRODUCTION + +The Scatterlist Crypto API takes page vectors (scatterlists) as +arguments, and works directly on pages. In some cases (e.g. ECB +mode ciphers), this will allow for pages to be encrypted in-place +with no copying. + +One of the initial goals of this design was to readily support IPsec, +so that processing can be applied to paged skb's without the need +for linearization. + + +DETAILS + +At the lowest level are algorithms, which register dynamically with the +API. + +'Transforms' are user-instantiated objects, which maintain state, handle all +of the implementation logic (e.g. manipulating page vectors) and provide an +abstraction to the underlying algorithms. However, at the user +level they are very simple. + +Conceptually, the API layering looks like this: + + [transform api] (user interface) + [transform ops] (per-type logic glue e.g. cipher.c, compress.c) + [algorithm api] (for registering algorithms) + +The idea is to make the user interface and algorithm registration API +very simple, while hiding the core logic from both. Many good ideas +from existing APIs such as Cryptoapi and Nettle have been adapted for this. + +The API currently supports five main types of transforms: AEAD (Authenticated +Encryption with Associated Data), Block Ciphers, Ciphers, Compressors and +Hashes. + +Please note that Block Ciphers is somewhat of a misnomer. It is in fact +meant to support all ciphers including stream ciphers. The difference +between Block Ciphers and Ciphers is that the latter operates on exactly +one block while the former can operate on an arbitrary amount of data, +subject to block size requirements (i.e., non-stream ciphers can only +process multiples of blocks). + +Support for hardware crypto devices via an asynchronous interface is +under development. + +Here's an example of how to use the API: + + #include + #include + #include + + struct scatterlist sg[2]; + char result[128]; + struct crypto_hash *tfm; + struct hash_desc desc; + + tfm = crypto_alloc_hash("md5", 0, CRYPTO_ALG_ASYNC); + if (IS_ERR(tfm)) + fail(); + + /* ... set up the scatterlists ... */ + + desc.tfm = tfm; + desc.flags = 0; + + if (crypto_hash_digest(&desc, sg, 2, result)) + fail(); + + crypto_free_hash(tfm); + + +Many real examples are available in the regression test module (tcrypt.c). + + +DEVELOPER NOTES + +Transforms may only be allocated in user context, and cryptographic +methods may only be called from softirq and user contexts. For +transforms with a setkey method it too should only be called from +user context. + +When using the API for ciphers, performance will be optimal if each +scatterlist contains data which is a multiple of the cipher's block +size (typically 8 bytes). This prevents having to do any copying +across non-aligned page fragment boundaries. + + +ADDING NEW ALGORITHMS + +When submitting a new algorithm for inclusion, a mandatory requirement +is that at least a few test vectors from known sources (preferably +standards) be included. + +Converting existing well known code is preferred, as it is more likely +to have been reviewed and widely tested. If submitting code from LGPL +sources, please consider changing the license to GPL (see section 3 of +the LGPL). + +Algorithms submitted must also be generally patent-free (e.g. IDEA +will not be included in the mainline until around 2011), and be based +on a recognized standard and/or have been subjected to appropriate +peer review. + +Also check for any RFCs which may relate to the use of specific algorithms, +as well as general application notes such as RFC2451 ("The ESP CBC-Mode +Cipher Algorithms"). + +It's a good idea to avoid using lots of macros and use inlined functions +instead, as gcc does a good job with inlining, while excessive use of +macros can cause compilation problems on some platforms. + +Also check the TODO list at the web site listed below to see what people +might already be working on. + + +BUGS + +Send bug reports to: +linux-crypto@vger.kernel.org +Cc: Herbert Xu , + David S. Miller + + +FURTHER INFORMATION + +For further patches and various updates, including the current TODO +list, see: +http://gondor.apana.org.au/~herbert/crypto/ + + +AUTHORS + +James Morris +David S. Miller +Herbert Xu + + +CREDITS + +The following people provided invaluable feedback during the development +of the API: + + Alexey Kuznetzov + Rusty Russell + Herbert Valerio Riedel + Jeff Garzik + Michael Richardson + Andrew Morton + Ingo Oeser + Christoph Hellwig + +Portions of this API were derived from the following projects: + + Kerneli Cryptoapi (http://www.kerneli.org/) + Alexander Kjeldaas + Herbert Valerio Riedel + Kyle McMartin + Jean-Luc Cooke + David Bryson + Clemens Fruhwirth + Tobias Ringstrom + Harald Welte + +and; + + Nettle (http://www.lysator.liu.se/~nisse/nettle/) + Niels Möller + +Original developers of the crypto algorithms: + + Dana L. How (DES) + Andrew Tridgell and Steve French (MD4) + Colin Plumb (MD5) + Steve Reid (SHA1) + Jean-Luc Cooke (SHA256, SHA384, SHA512) + Kazunori Miyazawa / USAGI (HMAC) + Matthew Skala (Twofish) + Dag Arne Osvik (Serpent) + Brian Gladman (AES) + Kartikey Mahendra Bhatt (CAST6) + Jon Oberheide (ARC4) + Jouni Malinen (Michael MIC) + NTT(Nippon Telegraph and Telephone Corporation) (Camellia) + +SHA1 algorithm contributors: + Jean-Francois Dive + +DES algorithm contributors: + Raimar Falke + Gisle Sælensminde + Niels Möller + +Blowfish algorithm contributors: + Herbert Valerio Riedel + Kyle McMartin + +Twofish algorithm contributors: + Werner Koch + Marc Mutz + +SHA256/384/512 algorithm contributors: + Andrew McDonald + Kyle McMartin + Herbert Valerio Riedel + +AES algorithm contributors: + Alexander Kjeldaas + Herbert Valerio Riedel + Kyle McMartin + Adam J. Richter + Fruhwirth Clemens (i586) + Linus Torvalds (i586) + +CAST5 algorithm contributors: + Kartikey Mahendra Bhatt (original developers unknown, FSF copyright). + +TEA/XTEA algorithm contributors: + Aaron Grothe + Michael Ringe + +Khazad algorithm contributors: + Aaron Grothe + +Whirlpool algorithm contributors: + Aaron Grothe + Jean-Luc Cooke + +Anubis algorithm contributors: + Aaron Grothe + +Tiger algorithm contributors: + Aaron Grothe + +VIA PadLock contributors: + Michal Ludvig + +Camellia algorithm contributors: + NTT(Nippon Telegraph and Telephone Corporation) (Camellia) + +Generic scatterwalk code by Adam J. Richter + +Please send any credits updates or corrections to: +Herbert Xu + diff --git a/Documentation/crypto/async-tx-api.txt b/Documentation/crypto/async-tx-api.txt new file mode 100644 index 00000000..ba046b8f --- /dev/null +++ b/Documentation/crypto/async-tx-api.txt @@ -0,0 +1,226 @@ + Asynchronous Transfers/Transforms API + +1 INTRODUCTION + +2 GENEALOGY + +3 USAGE +3.1 General format of the API +3.2 Supported operations +3.3 Descriptor management +3.4 When does the operation execute? +3.5 When does the operation complete? +3.6 Constraints +3.7 Example + +4 DMAENGINE DRIVER DEVELOPER NOTES +4.1 Conformance points +4.2 "My application needs exclusive control of hardware channels" + +5 SOURCE + +--- + +1 INTRODUCTION + +The async_tx API provides methods for describing a chain of asynchronous +bulk memory transfers/transforms with support for inter-transactional +dependencies. It is implemented as a dmaengine client that smooths over +the details of different hardware offload engine implementations. Code +that is written to the API can optimize for asynchronous operation and +the API will fit the chain of operations to the available offload +resources. + +2 GENEALOGY + +The API was initially designed to offload the memory copy and +xor-parity-calculations of the md-raid5 driver using the offload engines +present in the Intel(R) Xscale series of I/O processors. It also built +on the 'dmaengine' layer developed for offloading memory copies in the +network stack using Intel(R) I/OAT engines. The following design +features surfaced as a result: +1/ implicit synchronous path: users of the API do not need to know if + the platform they are running on has offload capabilities. The + operation will be offloaded when an engine is available and carried out + in software otherwise. +2/ cross channel dependency chains: the API allows a chain of dependent + operations to be submitted, like xor->copy->xor in the raid5 case. The + API automatically handles cases where the transition from one operation + to another implies a hardware channel switch. +3/ dmaengine extensions to support multiple clients and operation types + beyond 'memcpy' + +3 USAGE + +3.1 General format of the API: +struct dma_async_tx_descriptor * +async_(, struct async_submit ctl *submit) + +3.2 Supported operations: +memcpy - memory copy between a source and a destination buffer +memset - fill a destination buffer with a byte value +xor - xor a series of source buffers and write the result to a + destination buffer +xor_val - xor a series of source buffers and set a flag if the + result is zero. The implementation attempts to prevent + writes to memory +pq - generate the p+q (raid6 syndrome) from a series of source buffers +pq_val - validate that a p and or q buffer are in sync with a given series of + sources +datap - (raid6_datap_recov) recover a raid6 data block and the p block + from the given sources +2data - (raid6_2data_recov) recover 2 raid6 data blocks from the given + sources + +3.3 Descriptor management: +The return value is non-NULL and points to a 'descriptor' when the operation +has been queued to execute asynchronously. Descriptors are recycled +resources, under control of the offload engine driver, to be reused as +operations complete. When an application needs to submit a chain of +operations it must guarantee that the descriptor is not automatically recycled +before the dependency is submitted. This requires that all descriptors be +acknowledged by the application before the offload engine driver is allowed to +recycle (or free) the descriptor. A descriptor can be acked by one of the +following methods: +1/ setting the ASYNC_TX_ACK flag if no child operations are to be submitted +2/ submitting an unacknowledged descriptor as a dependency to another + async_tx call will implicitly set the acknowledged state. +3/ calling async_tx_ack() on the descriptor. + +3.4 When does the operation execute? +Operations do not immediately issue after return from the +async_ call. Offload engine drivers batch operations to +improve performance by reducing the number of mmio cycles needed to +manage the channel. Once a driver-specific threshold is met the driver +automatically issues pending operations. An application can force this +event by calling async_tx_issue_pending_all(). This operates on all +channels since the application has no knowledge of channel to operation +mapping. + +3.5 When does the operation complete? +There are two methods for an application to learn about the completion +of an operation. +1/ Call dma_wait_for_async_tx(). This call causes the CPU to spin while + it polls for the completion of the operation. It handles dependency + chains and issuing pending operations. +2/ Specify a completion callback. The callback routine runs in tasklet + context if the offload engine driver supports interrupts, or it is + called in application context if the operation is carried out + synchronously in software. The callback can be set in the call to + async_, or when the application needs to submit a chain of + unknown length it can use the async_trigger_callback() routine to set a + completion interrupt/callback at the end of the chain. + +3.6 Constraints: +1/ Calls to async_ are not permitted in IRQ context. Other + contexts are permitted provided constraint #2 is not violated. +2/ Completion callback routines cannot submit new operations. This + results in recursion in the synchronous case and spin_locks being + acquired twice in the asynchronous case. + +3.7 Example: +Perform a xor->copy->xor operation where each operation depends on the +result from the previous operation: + +void callback(void *param) +{ + struct completion *cmp = param; + + complete(cmp); +} + +void run_xor_copy_xor(struct page **xor_srcs, + int xor_src_cnt, + struct page *xor_dest, + size_t xor_len, + struct page *copy_src, + struct page *copy_dest, + size_t copy_len) +{ + struct dma_async_tx_descriptor *tx; + addr_conv_t addr_conv[xor_src_cnt]; + struct async_submit_ctl submit; + addr_conv_t addr_conv[NDISKS]; + struct completion cmp; + + init_async_submit(&submit, ASYNC_TX_XOR_DROP_DST, NULL, NULL, NULL, + addr_conv); + tx = async_xor(xor_dest, xor_srcs, 0, xor_src_cnt, xor_len, &submit) + + submit->depend_tx = tx; + tx = async_memcpy(copy_dest, copy_src, 0, 0, copy_len, &submit); + + init_completion(&cmp); + init_async_submit(&submit, ASYNC_TX_XOR_DROP_DST | ASYNC_TX_ACK, tx, + callback, &cmp, addr_conv); + tx = async_xor(xor_dest, xor_srcs, 0, xor_src_cnt, xor_len, &submit); + + async_tx_issue_pending_all(); + + wait_for_completion(&cmp); +} + +See include/linux/async_tx.h for more information on the flags. See the +ops_run_* and ops_complete_* routines in drivers/md/raid5.c for more +implementation examples. + +4 DRIVER DEVELOPMENT NOTES + +4.1 Conformance points: +There are a few conformance points required in dmaengine drivers to +accommodate assumptions made by applications using the async_tx API: +1/ Completion callbacks are expected to happen in tasklet context +2/ dma_async_tx_descriptor fields are never manipulated in IRQ context +3/ Use async_tx_run_dependencies() in the descriptor clean up path to + handle submission of dependent operations + +4.2 "My application needs exclusive control of hardware channels" +Primarily this requirement arises from cases where a DMA engine driver +is being used to support device-to-memory operations. A channel that is +performing these operations cannot, for many platform specific reasons, +be shared. For these cases the dma_request_channel() interface is +provided. + +The interface is: +struct dma_chan *dma_request_channel(dma_cap_mask_t mask, + dma_filter_fn filter_fn, + void *filter_param); + +Where dma_filter_fn is defined as: +typedef bool (*dma_filter_fn)(struct dma_chan *chan, void *filter_param); + +When the optional 'filter_fn' parameter is set to NULL +dma_request_channel simply returns the first channel that satisfies the +capability mask. Otherwise, when the mask parameter is insufficient for +specifying the necessary channel, the filter_fn routine can be used to +disposition the available channels in the system. The filter_fn routine +is called once for each free channel in the system. Upon seeing a +suitable channel filter_fn returns DMA_ACK which flags that channel to +be the return value from dma_request_channel. A channel allocated via +this interface is exclusive to the caller, until dma_release_channel() +is called. + +The DMA_PRIVATE capability flag is used to tag dma devices that should +not be used by the general-purpose allocator. It can be set at +initialization time if it is known that a channel will always be +private. Alternatively, it is set when dma_request_channel() finds an +unused "public" channel. + +A couple caveats to note when implementing a driver and consumer: +1/ Once a channel has been privately allocated it will no longer be + considered by the general-purpose allocator even after a call to + dma_release_channel(). +2/ Since capabilities are specified at the device level a dma_device + with multiple channels will either have all channels public, or all + channels private. + +5 SOURCE + +include/linux/dmaengine.h: core header file for DMA drivers and api users +drivers/dma/dmaengine.c: offload engine channel management routines +drivers/dma/: location for offload engine drivers +include/linux/async_tx.h: core header file for the async_tx api +crypto/async_tx/async_tx.c: async_tx interface to dmaengine and common code +crypto/async_tx/async_memcpy.c: copy offload +crypto/async_tx/async_memset.c: memory fill offload +crypto/async_tx/async_xor.c: xor and xor zero sum offload diff --git a/Documentation/crypto/descore-readme.txt b/Documentation/crypto/descore-readme.txt new file mode 100644 index 00000000..16e9e635 --- /dev/null +++ b/Documentation/crypto/descore-readme.txt @@ -0,0 +1,352 @@ +Below is the original README file from the descore.shar package. +------------------------------------------------------------------------------ + +des - fast & portable DES encryption & decryption. +Copyright (C) 1992 Dana L. How + +This program is free software; you can redistribute it and/or modify +it under the terms of the GNU Library General Public License as published by +the Free Software Foundation; either version 2 of the License, or +(at your option) any later version. + +This program is distributed in the hope that it will be useful, +but WITHOUT ANY WARRANTY; without even the implied warranty of +MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +GNU Library General Public License for more details. + +You should have received a copy of the GNU Library General Public License +along with this program; if not, write to the Free Software +Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + +Author's address: how@isl.stanford.edu + +$Id: README,v 1.15 1992/05/20 00:25:32 how E $ + + +==>> To compile after untarring/unsharring, just `make' <<== + + +This package was designed with the following goals: +1. Highest possible encryption/decryption PERFORMANCE. +2. PORTABILITY to any byte-addressable host with a 32bit unsigned C type +3. Plug-compatible replacement for KERBEROS's low-level routines. + +This second release includes a number of performance enhancements for +register-starved machines. My discussions with Richard Outerbridge, +71755.204@compuserve.com, sparked a number of these enhancements. + +To more rapidly understand the code in this package, inspect desSmallFips.i +(created by typing `make') BEFORE you tackle desCode.h. The latter is set +up in a parameterized fashion so it can easily be modified by speed-daemon +hackers in pursuit of that last microsecond. You will find it more +illuminating to inspect one specific implementation, +and then move on to the common abstract skeleton with this one in mind. + + +performance comparison to other available des code which i could +compile on a SPARCStation 1 (cc -O4, gcc -O2): + +this code (byte-order independent): + 30us per encryption (options: 64k tables, no IP/FP) + 33us per encryption (options: 64k tables, FIPS standard bit ordering) + 45us per encryption (options: 2k tables, no IP/FP) + 48us per encryption (options: 2k tables, FIPS standard bit ordering) + 275us to set a new key (uses 1k of key tables) + this has the quickest encryption/decryption routines i've seen. + since i was interested in fast des filters rather than crypt(3) + and password cracking, i haven't really bothered yet to speed up + the key setting routine. also, i have no interest in re-implementing + all the other junk in the mit kerberos des library, so i've just + provided my routines with little stub interfaces so they can be + used as drop-in replacements with mit's code or any of the mit- + compatible packages below. (note that the first two timings above + are highly variable because of cache effects). + +kerberos des replacement from australia (version 1.95): + 53us per encryption (uses 2k of tables) + 96us to set a new key (uses 2.25k of key tables) + so despite the author's inclusion of some of the performance + improvements i had suggested to him, this package's + encryption/decryption is still slower on the sparc and 68000. + more specifically, 19-40% slower on the 68020 and 11-35% slower + on the sparc, depending on the compiler; + in full gory detail (ALT_ECB is a libdes variant): + compiler machine desCore libdes ALT_ECB slower by + gcc 2.1 -O2 Sun 3/110 304 uS 369.5uS 461.8uS 22% + cc -O1 Sun 3/110 336 uS 436.6uS 399.3uS 19% + cc -O2 Sun 3/110 360 uS 532.4uS 505.1uS 40% + cc -O4 Sun 3/110 365 uS 532.3uS 505.3uS 38% + gcc 2.1 -O2 Sun 4/50 48 uS 53.4uS 57.5uS 11% + cc -O2 Sun 4/50 48 uS 64.6uS 64.7uS 35% + cc -O4 Sun 4/50 48 uS 64.7uS 64.9uS 35% + (my time measurements are not as accurate as his). + the comments in my first release of desCore on version 1.92: + 68us per encryption (uses 2k of tables) + 96us to set a new key (uses 2.25k of key tables) + this is a very nice package which implements the most important + of the optimizations which i did in my encryption routines. + it's a bit weak on common low-level optimizations which is why + it's 39%-106% slower. because he was interested in fast crypt(3) and + password-cracking applications, he also used the same ideas to + speed up the key-setting routines with impressive results. + (at some point i may do the same in my package). he also implements + the rest of the mit des library. + (code from eay@psych.psy.uq.oz.au via comp.sources.misc) + +fast crypt(3) package from denmark: + the des routine here is buried inside a loop to do the + crypt function and i didn't feel like ripping it out and measuring + performance. his code takes 26 sparc instructions to compute one + des iteration; above, Quick (64k) takes 21 and Small (2k) takes 37. + he claims to use 280k of tables but the iteration calculation seems + to use only 128k. his tables and code are machine independent. + (code from glad@daimi.aau.dk via alt.sources or comp.sources.misc) + +swedish reimplementation of Kerberos des library + 108us per encryption (uses 34k worth of tables) + 134us to set a new key (uses 32k of key tables to get this speed!) + the tables used seem to be machine-independent; + he seems to have included a lot of special case code + so that, e.g., `long' loads can be used instead of 4 `char' loads + when the machine's architecture allows it. + (code obtained from chalmers.se:pub/des) + +crack 3.3c package from england: + as in crypt above, the des routine is buried in a loop. it's + also very modified for crypt. his iteration code uses 16k + of tables and appears to be slow. + (code obtained from aem@aber.ac.uk via alt.sources or comp.sources.misc) + +``highly optimized'' and tweaked Kerberos/Athena code (byte-order dependent): + 165us per encryption (uses 6k worth of tables) + 478us to set a new key (uses <1k of key tables) + so despite the comments in this code, it was possible to get + faster code AND smaller tables, as well as making the tables + machine-independent. + (code obtained from prep.ai.mit.edu) + +UC Berkeley code (depends on machine-endedness): + 226us per encryption +10848us to set a new key + table sizes are unclear, but they don't look very small + (code obtained from wuarchive.wustl.edu) + + +motivation and history + +a while ago i wanted some des routines and the routines documented on sun's +man pages either didn't exist or dumped core. i had heard of kerberos, +and knew that it used des, so i figured i'd use its routines. but once +i got it and looked at the code, it really set off a lot of pet peeves - +it was too convoluted, the code had been written without taking +advantage of the regular structure of operations such as IP, E, and FP +(i.e. the author didn't sit down and think before coding), +it was excessively slow, the author had attempted to clarify the code +by adding MORE statements to make the data movement more `consistent' +instead of simplifying his implementation and cutting down on all data +movement (in particular, his use of L1, R1, L2, R2), and it was full of +idiotic `tweaks' for particular machines which failed to deliver significant +speedups but which did obfuscate everything. so i took the test data +from his verification program and rewrote everything else. + +a while later i ran across the great crypt(3) package mentioned above. +the fact that this guy was computing 2 sboxes per table lookup rather +than one (and using a MUCH larger table in the process) emboldened me to +do the same - it was a trivial change from which i had been scared away +by the larger table size. in his case he didn't realize you don't need to keep +the working data in TWO forms, one for easy use of half the sboxes in +indexing, the other for easy use of the other half; instead you can keep +it in the form for the first half and use a simple rotate to get the other +half. this means i have (almost) half the data manipulation and half +the table size. in fairness though he might be encoding something particular +to crypt(3) in his tables - i didn't check. + +i'm glad that i implemented it the way i did, because this C version is +portable (the ifdef's are performance enhancements) and it is faster +than versions hand-written in assembly for the sparc! + + +porting notes + +one thing i did not want to do was write an enormous mess +which depended on endedness and other machine quirks, +and which necessarily produced different code and different lookup tables +for different machines. see the kerberos code for an example +of what i didn't want to do; all their endedness-specific `optimizations' +obfuscate the code and in the end were slower than a simpler machine +independent approach. however, there are always some portability +considerations of some kind, and i have included some options +for varying numbers of register variables. +perhaps some will still regard the result as a mess! + +1) i assume everything is byte addressable, although i don't actually + depend on the byte order, and that bytes are 8 bits. + i assume word pointers can be freely cast to and from char pointers. + note that 99% of C programs make these assumptions. + i always use unsigned char's if the high bit could be set. +2) the typedef `word' means a 32 bit unsigned integral type. + if `unsigned long' is not 32 bits, change the typedef in desCore.h. + i assume sizeof(word) == 4 EVERYWHERE. + +the (worst-case) cost of my NOT doing endedness-specific optimizations +in the data loading and storing code surrounding the key iterations +is less than 12%. also, there is the added benefit that +the input and output work areas do not need to be word-aligned. + + +OPTIONAL performance optimizations + +1) you should define one of `i386,' `vax,' `mc68000,' or `sparc,' + whichever one is closest to the capabilities of your machine. + see the start of desCode.h to see exactly what this selection implies. + note that if you select the wrong one, the des code will still work; + these are just performance tweaks. +2) for those with functional `asm' keywords: you should change the + ROR and ROL macros to use machine rotate instructions if you have them. + this will save 2 instructions and a temporary per use, + or about 32 to 40 instructions per en/decryption. + note that gcc is smart enough to translate the ROL/R macros into + machine rotates! + +these optimizations are all rather persnickety, yet with them you should +be able to get performance equal to assembly-coding, except that: +1) with the lack of a bit rotate operator in C, rotates have to be synthesized + from shifts. so access to `asm' will speed things up if your machine + has rotates, as explained above in (3) (not necessary if you use gcc). +2) if your machine has less than 12 32-bit registers i doubt your compiler will + generate good code. + `i386' tries to configure the code for a 386 by only declaring 3 registers + (it appears that gcc can use ebx, esi and edi to hold register variables). + however, if you like assembly coding, the 386 does have 7 32-bit registers, + and if you use ALL of them, use `scaled by 8' address modes with displacement + and other tricks, you can get reasonable routines for DesQuickCore... with + about 250 instructions apiece. For DesSmall... it will help to rearrange + des_keymap, i.e., now the sbox # is the high part of the index and + the 6 bits of data is the low part; it helps to exchange these. + since i have no way to conveniently test it i have not provided my + shoehorned 386 version. note that with this release of desCore, gcc is able + to put everything in registers(!), and generate about 370 instructions apiece + for the DesQuickCore... routines! + +coding notes + +the en/decryption routines each use 6 necessary register variables, +with 4 being actively used at once during the inner iterations. +if you don't have 4 register variables get a new machine. +up to 8 more registers are used to hold constants in some configurations. + +i assume that the use of a constant is more expensive than using a register: +a) additionally, i have tried to put the larger constants in registers. + registering priority was by the following: + anything more than 12 bits (bad for RISC and CISC) + greater than 127 in value (can't use movq or byte immediate on CISC) + 9-127 (may not be able to use CISC shift immediate or add/sub quick), + 1-8 were never registered, being the cheapest constants. +b) the compiler may be too stupid to realize table and table+256 should + be assigned to different constant registers and instead repetitively + do the arithmetic, so i assign these to explicit `m' register variables + when possible and helpful. + +i assume that indexing is cheaper or equivalent to auto increment/decrement, +where the index is 7 bits unsigned or smaller. +this assumption is reversed for 68k and vax. + +i assume that addresses can be cheaply formed from two registers, +or from a register and a small constant. +for the 68000, the `two registers and small offset' form is used sparingly. +all index scaling is done explicitly - no hidden shifts by log2(sizeof). + +the code is written so that even a dumb compiler +should never need more than one hidden temporary, +increasing the chance that everything will fit in the registers. +KEEP THIS MORE SUBTLE POINT IN MIND IF YOU REWRITE ANYTHING. +(actually, there are some code fragments now which do require two temps, +but fixing it would either break the structure of the macros or +require declaring another temporary). + + +special efficient data format + +bits are manipulated in this arrangement most of the time (S7 S5 S3 S1): + 003130292827xxxx242322212019xxxx161514131211xxxx080706050403xxxx +(the x bits are still there, i'm just emphasizing where the S boxes are). +bits are rotated left 4 when computing S6 S4 S2 S0: + 282726252423xxxx201918171615xxxx121110090807xxxx040302010031xxxx +the rightmost two bits are usually cleared so the lower byte can be used +as an index into an sbox mapping table. the next two x'd bits are set +to various values to access different parts of the tables. + + +how to use the routines + +datatypes: + pointer to 8 byte area of type DesData + used to hold keys and input/output blocks to des. + + pointer to 128 byte area of type DesKeys + used to hold full 768-bit key. + must be long-aligned. + +DesQuickInit() + call this before using any other routine with `Quick' in its name. + it generates the special 64k table these routines need. +DesQuickDone() + frees this table + +DesMethod(m, k) + m points to a 128byte block, k points to an 8 byte des key + which must have odd parity (or -1 is returned) and which must + not be a (semi-)weak key (or -2 is returned). + normally DesMethod() returns 0. + m is filled in from k so that when one of the routines below + is called with m, the routine will act like standard des + en/decryption with the key k. if you use DesMethod, + you supply a standard 56bit key; however, if you fill in + m yourself, you will get a 768bit key - but then it won't + be standard. it's 768bits not 1024 because the least significant + two bits of each byte are not used. note that these two bits + will be set to magic constants which speed up the encryption/decryption + on some machines. and yes, each byte controls + a specific sbox during a specific iteration. + you really shouldn't use the 768bit format directly; i should + provide a routine that converts 128 6-bit bytes (specified in + S-box mapping order or something) into the right format for you. + this would entail some byte concatenation and rotation. + +Des{Small|Quick}{Fips|Core}{Encrypt|Decrypt}(d, m, s) + performs des on the 8 bytes at s into the 8 bytes at d. (d,s: char *). + uses m as a 768bit key as explained above. + the Encrypt|Decrypt choice is obvious. + Fips|Core determines whether a completely standard FIPS initial + and final permutation is done; if not, then the data is loaded + and stored in a nonstandard bit order (FIPS w/o IP/FP). + Fips slows down Quick by 10%, Small by 9%. + Small|Quick determines whether you use the normal routine + or the crazy quick one which gobbles up 64k more of memory. + Small is 50% slower then Quick, but Quick needs 32 times as much + memory. Quick is included for programs that do nothing but DES, + e.g., encryption filters, etc. + + +Getting it to compile on your machine + +there are no machine-dependencies in the code (see porting), +except perhaps the `now()' macro in desTest.c. +ALL generated tables are machine independent. +you should edit the Makefile with the appropriate optimization flags +for your compiler (MAX optimization). + + +Speeding up kerberos (and/or its des library) + +note that i have included a kerberos-compatible interface in desUtil.c +through the functions des_key_sched() and des_ecb_encrypt(). +to use these with kerberos or kerberos-compatible code put desCore.a +ahead of the kerberos-compatible library on your linker's command line. +you should not need to #include desCore.h; just include the header +file provided with the kerberos library. + +Other uses + +the macros in desCode.h would be very useful for putting inline des +functions in more complicated encryption routines. -- cgit v1.2.3