The patch titled Subject: crc32: move long comment about crc32 fundamentals to Documentation/ has been added to the -mm tree. Its filename is crc32-move-long-comment-about-crc32-fundamentals-to-documentation.patch Before you just go and hit "reply", please: a) Consider who else should be cc'ed b) Prefer to cc a suitable mailing list as well c) Ideally: find the original patch on the mailing list and do a reply-to-all to that, adding suitable additional cc's *** Remember to use Documentation/SubmitChecklist when testing your code *** The -mm tree is included into linux-next and is updated there every 3-4 working days ------------------------------------------------------ From: Bob Pearson <rpearson@xxxxxxxxxxxxxxxxxxxxx> Subject: crc32: move long comment about crc32 fundamentals to Documentation/ Move a long comment from lib/crc32.c to Documentation/crc32.txt where it will more likely get read. Edited the resulting document to add an explanation of the slicing-by-n algorithm. Signed-off-by: George Spelvin <linux@xxxxxxxxxxx> Signed-off-by: Bob Pearson <rpearson@xxxxxxxxxxxxxxxxxxxxx> [djwong@xxxxxxxxxx: Minor changelog tweaks] Signed-off-by: Darrick J. Wong <djwong@xxxxxxxxxx> Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> --- Documentation/00-INDEX | 2 Documentation/crc32.txt | 183 ++++++++++++++++++++++++++++++++++++++ lib/crc32.c | 129 -------------------------- 3 files changed, 187 insertions(+), 127 deletions(-) diff -puN Documentation/00-INDEX~crc32-move-long-comment-about-crc32-fundamentals-to-documentation Documentation/00-INDEX --- a/Documentation/00-INDEX~crc32-move-long-comment-about-crc32-fundamentals-to-documentation +++ a/Documentation/00-INDEX @@ -104,6 +104,8 @@ cpuidle/ - info on CPU_IDLE, CPU idle state management subsystem. cputopology.txt - documentation on how CPU topology info is exported via sysfs. +crc32.txt + - brief tutorial on CRC computation cris/ - directory with info about Linux on CRIS architecture. crypto/ diff -puN /dev/null Documentation/crc32.txt --- /dev/null +++ a/Documentation/crc32.txt @@ -0,0 +1,183 @@ +A brief CRC tutorial. + +A CRC is a long-division remainder. You add the CRC to the message, +and the whole thing (message+CRC) is a multiple of the given +CRC polynomial. To check the CRC, you can either check that the +CRC matches the recomputed value, *or* you can check that the +remainder computed on the message+CRC is 0. This latter approach +is used by a lot of hardware implementations, and is why so many +protocols put the end-of-frame flag after the CRC. + +It's actually the same long division you learned in school, except that +- We're working in binary, so the digits are only 0 and 1, and +- When dividing polynomials, there are no carries. Rather than add and + subtract, we just xor. Thus, we tend to get a bit sloppy about + the difference between adding and subtracting. + +Like all division, the remainder is always smaller than the divisor. +To produce a 32-bit CRC, the divisor is actually a 33-bit CRC polynomial. +Since it's 33 bits long, bit 32 is always going to be set, so usually the +CRC is written in hex with the most significant bit omitted. (If you're +familiar with the IEEE 754 floating-point format, it's the same idea.) + +Note that a CRC is computed over a string of *bits*, so you have +to decide on the endianness of the bits within each byte. To get +the best error-detecting properties, this should correspond to the +order they're actually sent. For example, standard RS-232 serial is +little-endian; the most significant bit (sometimes used for parity) +is sent last. And when appending a CRC word to a message, you should +do it in the right order, matching the endianness. + +Just like with ordinary division, you proceed one digit (bit) at a time. +Each step of the division, division, you take one more digit (bit) of the +dividend and append it to the current remainder. Then you figure out the +appropriate multiple of the divisor to subtract to being the remainder +back into range. In binary, this is easy - it has to be either 0 or 1, +and to make the XOR cancel, it's just a copy of bit 32 of the remainder. + +When computing a CRC, we don't care about the quotient, so we can +throw the quotient bit away, but subtract the appropriate multiple of +the polynomial from the remainder and we're back to where we started, +ready to process the next bit. + +A big-endian CRC written this way would be coded like: +for (i = 0; i < input_bits; i++) { + multiple = remainder & 0x80000000 ? CRCPOLY : 0; + remainder = (remainder << 1 | next_input_bit()) ^ multiple; +} + +Notice how, to get at bit 32 of the shifted remainder, we look +at bit 31 of the remainder *before* shifting it. + +But also notice how the next_input_bit() bits we're shifting into +the remainder don't actually affect any decision-making until +32 bits later. Thus, the first 32 cycles of this are pretty boring. +Also, to add the CRC to a message, we need a 32-bit-long hole for it at +the end, so we have to add 32 extra cycles shifting in zeros at the +end of every message, + +These details lead to a standard trick: rearrange merging in the +next_input_bit() until the moment it's needed. Then the first 32 cycles +can be precomputed, and merging in the final 32 zero bits to make room +for the CRC can be skipped entirely. This changes the code to: + +for (i = 0; i < input_bits; i++) { + remainder ^= next_input_bit() << 31; + multiple = (remainder & 0x80000000) ? CRCPOLY : 0; + remainder = (remainder << 1) ^ multiple; +} + +With this optimization, the little-endian code is particularly simple: +for (i = 0; i < input_bits; i++) { + remainder ^= next_input_bit(); + multiple = (remainder & 1) ? CRCPOLY : 0; + remainder = (remainder >> 1) ^ multiple; +} + +The most significant coefficient of the remainder polynomial is stored +in the least significant bit of the binary "remainder" variable. +The other details of endianness have been hidden in CRCPOLY (which must +be bit-reversed) and next_input_bit(). + +As long as next_input_bit is returning the bits in a sensible order, we don't +*have* to wait until the last possible moment to merge in additional bits. +We can do it 8 bits at a time rather than 1 bit at a time: +for (i = 0; i < input_bytes; i++) { + remainder ^= next_input_byte() << 24; + for (j = 0; j < 8; j++) { + multiple = (remainder & 0x80000000) ? CRCPOLY : 0; + remainder = (remainder << 1) ^ multiple; + } +} + +Or in little-endian: +for (i = 0; i < input_bytes; i++) { + remainder ^= next_input_byte(); + for (j = 0; j < 8; j++) { + multiple = (remainder & 1) ? CRCPOLY : 0; + remainder = (remainder >> 1) ^ multiple; + } +} + +If the input is a multiple of 32 bits, you can even XOR in a 32-bit +word at a time and increase the inner loop count to 32. + +You can also mix and match the two loop styles, for example doing the +bulk of a message byte-at-a-time and adding bit-at-a-time processing +for any fractional bytes at the end. + +To reduce the number of conditional branches, software commonly uses +the byte-at-a-time table method, popularized by Dilip V. Sarwate, +"Computation of Cyclic Redundancy Checks via Table Look-Up", Comm. ACM +v.31 no.8 (August 1998) p. 1008-1013. + +Here, rather than just shifting one bit of the remainder to decide +in the correct multiple to subtract, we can shift a byte at a time. +This produces a 40-bit (rather than a 33-bit) intermediate remainder, +and the correct multiple of the polynomial to subtract is found using +a 256-entry lookup table indexed by the high 8 bits. + +(The table entries are simply the CRC-32 of the given one-byte messages.) + +When space is more constrained, smaller tables can be used, e.g. two +4-bit shifts followed by a lookup in a 16-entry table. + +It is not practical to process much more than 8 bits at a time using this +technique, because tables larger than 256 entries use too much memory and, +more importantly, too much of the L1 cache. + +To get higher software performance, a "slicing" technique can be used. +See "High Octane CRC Generation with the Intel Slicing-by-8 Algorithm", +ftp://download.intel.com/technology/comms/perfnet/download/slicing-by-8.pdf + +This does not change the number of table lookups, but does increase +the parallelism. With the classic Sarwate algorithm, each table lookup +must be completed before the index of the next can be computed. + +A "slicing by 2" technique would shift the remainder 16 bits at a time, +producing a 48-bit intermediate remainder. Rather than doing a single +lookup in a 65536-entry table, the two high bytes are looked up in +two different 256-entry tables. Each contains the remainder required +to cancel out the corresponding byte. The tables are different because the +polynomials to cancel are different. One has non-zero coefficients from +x^32 to x^39, while the other goes from x^40 to x^47. + +Since modern processors can handle many parallel memory operations, this +takes barely longer than a single table look-up and thus performs almost +twice as fast as the basic Sarwate algorithm. + +This can be extended to "slicing by 4" using 4 256-entry tables. +Each step, 32 bits of data is fetched, XORed with the CRC, and the result +broken into bytes and looked up in the tables. Because the 32-bit shift +leaves the low-order bits of the intermediate remainder zero, the +final CRC is simply the XOR of the 4 table look-ups. + +But this still enforces sequential execution: a second group of table +look-ups cannot begin until the previous groups 4 table look-ups have all +been completed. Thus, the processor's load/store unit is sometimes idle. + +To make maximum use of the processor, "slicing by 8" performs 8 look-ups +in parallel. Each step, the 32-bit CRC is shifted 64 bits and XORed +with 64 bits of input data. What is important to note is that 4 of +those 8 bytes are simply copies of the input data; they do not depend +on the previous CRC at all. Thus, those 4 table look-ups may commence +immediately, without waiting for the previous loop iteration. + +By always having 4 loads in flight, a modern superscalar processor can +be kept busy and make full use of its L1 cache. + +Two more details about CRC implementation in the real world: + +Normally, appending zero bits to a message which is already a multiple +of a polynomial produces a larger multiple of that polynomial. Thus, +a basic CRC will not detect appended zero bits (or bytes). To enable +a CRC to detect this condition, it's common to invert the CRC before +appending it. This makes the remainder of the message+crc come out not +as zero, but some fixed non-zero value. (The CRC of the inversion +pattern, 0xffffffff.) + +The same problem applies to zero bits prepended to the message, and a +similar solution is used. Instead of starting the CRC computation with +a remainder of 0, an initial remainder of all ones is used. As long as +you start the same way on decoding, it doesn't make a difference. + diff -puN lib/crc32.c~crc32-move-long-comment-about-crc32-fundamentals-to-documentation lib/crc32.c --- a/lib/crc32.c~crc32-move-long-comment-about-crc32-fundamentals-to-documentation +++ a/lib/crc32.c @@ -20,6 +20,8 @@ * Version 2. See the file COPYING for more details. */ +/* see: Documentation/crc32.txt for a description of algorithms */ + #include <linux/crc32.h> #include <linux/kernel.h> #include <linux/module.h> @@ -209,133 +211,6 @@ u32 __pure crc32_be(u32 crc, unsigned ch EXPORT_SYMBOL(crc32_le); EXPORT_SYMBOL(crc32_be); -/* - * A brief CRC tutorial. - * - * A CRC is a long-division remainder. You add the CRC to the message, - * and the whole thing (message+CRC) is a multiple of the given - * CRC polynomial. To check the CRC, you can either check that the - * CRC matches the recomputed value, *or* you can check that the - * remainder computed on the message+CRC is 0. This latter approach - * is used by a lot of hardware implementations, and is why so many - * protocols put the end-of-frame flag after the CRC. - * - * It's actually the same long division you learned in school, except that - * - We're working in binary, so the digits are only 0 and 1, and - * - When dividing polynomials, there are no carries. Rather than add and - * subtract, we just xor. Thus, we tend to get a bit sloppy about - * the difference between adding and subtracting. - * - * A 32-bit CRC polynomial is actually 33 bits long. But since it's - * 33 bits long, bit 32 is always going to be set, so usually the CRC - * is written in hex with the most significant bit omitted. (If you're - * familiar with the IEEE 754 floating-point format, it's the same idea.) - * - * Note that a CRC is computed over a string of *bits*, so you have - * to decide on the endianness of the bits within each byte. To get - * the best error-detecting properties, this should correspond to the - * order they're actually sent. For example, standard RS-232 serial is - * little-endian; the most significant bit (sometimes used for parity) - * is sent last. And when appending a CRC word to a message, you should - * do it in the right order, matching the endianness. - * - * Just like with ordinary division, the remainder is always smaller than - * the divisor (the CRC polynomial) you're dividing by. Each step of the - * division, you take one more digit (bit) of the dividend and append it - * to the current remainder. Then you figure out the appropriate multiple - * of the divisor to subtract to being the remainder back into range. - * In binary, it's easy - it has to be either 0 or 1, and to make the - * XOR cancel, it's just a copy of bit 32 of the remainder. - * - * When computing a CRC, we don't care about the quotient, so we can - * throw the quotient bit away, but subtract the appropriate multiple of - * the polynomial from the remainder and we're back to where we started, - * ready to process the next bit. - * - * A big-endian CRC written this way would be coded like: - * for (i = 0; i < input_bits; i++) { - * multiple = remainder & 0x80000000 ? CRCPOLY : 0; - * remainder = (remainder << 1 | next_input_bit()) ^ multiple; - * } - * Notice how, to get at bit 32 of the shifted remainder, we look - * at bit 31 of the remainder *before* shifting it. - * - * But also notice how the next_input_bit() bits we're shifting into - * the remainder don't actually affect any decision-making until - * 32 bits later. Thus, the first 32 cycles of this are pretty boring. - * Also, to add the CRC to a message, we need a 32-bit-long hole for it at - * the end, so we have to add 32 extra cycles shifting in zeros at the - * end of every message, - * - * So the standard trick is to rearrage merging in the next_input_bit() - * until the moment it's needed. Then the first 32 cycles can be precomputed, - * and merging in the final 32 zero bits to make room for the CRC can be - * skipped entirely. - * This changes the code to: - * for (i = 0; i < input_bits; i++) { - * remainder ^= next_input_bit() << 31; - * multiple = (remainder & 0x80000000) ? CRCPOLY : 0; - * remainder = (remainder << 1) ^ multiple; - * } - * With this optimization, the little-endian code is simpler: - * for (i = 0; i < input_bits; i++) { - * remainder ^= next_input_bit(); - * multiple = (remainder & 1) ? CRCPOLY : 0; - * remainder = (remainder >> 1) ^ multiple; - * } - * - * Note that the other details of endianness have been hidden in CRCPOLY - * (which must be bit-reversed) and next_input_bit(). - * - * However, as long as next_input_bit is returning the bits in a sensible - * order, we can actually do the merging 8 or more bits at a time rather - * than one bit at a time: - * for (i = 0; i < input_bytes; i++) { - * remainder ^= next_input_byte() << 24; - * for (j = 0; j < 8; j++) { - * multiple = (remainder & 0x80000000) ? CRCPOLY : 0; - * remainder = (remainder << 1) ^ multiple; - * } - * } - * Or in little-endian: - * for (i = 0; i < input_bytes; i++) { - * remainder ^= next_input_byte(); - * for (j = 0; j < 8; j++) { - * multiple = (remainder & 1) ? CRCPOLY : 0; - * remainder = (remainder << 1) ^ multiple; - * } - * } - * If the input is a multiple of 32 bits, you can even XOR in a 32-bit - * word at a time and increase the inner loop count to 32. - * - * You can also mix and match the two loop styles, for example doing the - * bulk of a message byte-at-a-time and adding bit-at-a-time processing - * for any fractional bytes at the end. - * - * The only remaining optimization is to the byte-at-a-time table method. - * Here, rather than just shifting one bit of the remainder to decide - * in the correct multiple to subtract, we can shift a byte at a time. - * This produces a 40-bit (rather than a 33-bit) intermediate remainder, - * but again the multiple of the polynomial to subtract depends only on - * the high bits, the high 8 bits in this case. - * - * The multiple we need in that case is the low 32 bits of a 40-bit - * value whose high 8 bits are given, and which is a multiple of the - * generator polynomial. This is simply the CRC-32 of the given - * one-byte message. - * - * Two more details: normally, appending zero bits to a message which - * is already a multiple of a polynomial produces a larger multiple of that - * polynomial. To enable a CRC to detect this condition, it's common to - * invert the CRC before appending it. This makes the remainder of the - * message+crc come out not as zero, but some fixed non-zero value. - * - * The same problem applies to zero bits prepended to the message, and - * a similar solution is used. Instead of starting with a remainder of - * 0, an initial remainder of all ones is used. As long as you start - * the same way on decoding, it doesn't make a difference. - */ - #ifdef UNITTEST #include <stdlib.h> _ Subject: Subject: crc32: move long comment about crc32 fundamentals to Documentation/ Patches currently in -mm which might be from rpearson@xxxxxxxxxxxxxxxxxxxxx are origin.patch crc32-removed-two-instances-of-trailing-whitespaces.patch crc32-move-long-comment-about-crc32-fundamentals-to-documentation.patch crc32-simplify-unit-test-code.patch crc32-miscellaneous-cleanups.patch crc32-fix-mixing-of-endian-specific-types.patch crc32-make-crc__bits-definition-correspond-to-actual-bit-counts.patch crc32-add-slice-by-8-algorithm-to-existing-code.patch crc32-optimize-loop-counter-for-x86.patch crc32-add-note-about-this-patchset-to-crc32c.patch crc32-bolt-on-crc32c.patch crypto-crc32c-should-use-library-implementation.patch crc32-add-self-test-code-for-crc32c.patch crc32-select-an-algorithm-via-kconfig.patch -- To unsubscribe from this list: send the line "unsubscribe mm-commits" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html