Method and apparatus for encoding special uncorrectable errors in an error correction code

[0001] This application is related to the commonly owned, concurrently filed application of the same inventor, Ser. No. 10/744,564, entitled “(18, 9) Error Correction Code for Double Error Correction and Triple Error Detection”.

[0002] 1. Field of the Invention

[0003] This invention relates, in general, to error correction codes for computers and data communications, and in particular to a special coding for encoding special uncorrectable errors for computer failure isolation.

[0004] 2. Description of the Related Art

[0005] Error correction codes (ECCs) have long been used in computers as well as data communications. Typically, such codes are constructed by appending r=n−k check symbols to k message symbols to form an n-symbol code word, using a linear matrix transformation of the form:
C=MG,
where C=(c₀, c₁, . . . , c_n-1) is a 1×n row vector representing the n-symbol code word, M=(m₀, m₁, . . . m_k-1) is a 1×k row vector representing the k-symbol message or data word, and G is a k×n matrix known as a generator matrix. (Alternatively, if C and M are assumed to be column vectors, the transformation becomes C=G^tM, where G^t is the transpose of G.) Although the symbols need not be bits, they are usually bits, and bits will be referred to in the discussion that follows. The code word C is either written to a storage medium or transmitted over a communication channel. Both the storage medium and the communication channel in the narrow sense may be regarded as “communication channels” in the broad sense.

[0006] On the decoding side, an n-bit word R is either retrieved from a storage medium or received over a communication channel. This word R is the sum of the originally generated code word C and an n-bit error word E (which may be zero) representing any errors that may have occurred. To determine whether the received word accurately represents the original code word C, the received word R is used to generate a k-bit syndrome vector using a matrix transformation of the form:
S=RH^t,
where S is the syndrome vector and H^t is the transpose of an r×n matrix H known as a parity check matrix. (Alternatively, if S is assumed to be a column vector, the transformation becomes S=HR^t, where R^t is the transpose of R.)

[0007] The parity check matrix H is selected so that its row vectors lie in the null space of those of the generator matrix G (i.e., GH^t=0), so that for an original code word C,
CH^t=0.

[0008] Since R=C+E,
S=(C+E)H^t=CH^t+EH^t=0+EH^t=EH^t.

[0009] In other words, the syndrome vector S is independent of the original code word C and a function only of the error word E. The decoder uses the syndrome vector S to reconstruct the error word (following maximum likelihood criteria), which is subtracted from the received word R to regenerate the code word C.

[0010] A code is the set of code words C generated from a given set of data words M. (Even if two generator matrices differ, their code spaces may be the same.) Codes are commonly classified by the number of symbols in their data word M and code word C. Thus, an (n, k) code has a code word of n symbols generated on a data word of k symbols.

[0011] The ability of a code to detect and correct errors depends on the so-called Hamming distance between different code words of the code. In general, the Hamming distance between two code words is the number of symbols in which the two code words differ. If the minimum Hamming distance of a code is t+1, then the code can detect up to t errors, since if the code word has t or fewer errors, it will not have changed into any other code word. Similarly, if the minimum Hamming distance of a code is 2t+1, the code can correct up to t errors, since a received word having t or fewer errors will be within a Hamming distance of t symbols on one and only one code word, and thus can be unambiguously decoded as that code word.

[0012] Furthermore, if the minimum Hamming distance of a code is 2t+2, the code can correct up to t errors and also can detect t+1 errors, since a received word having t+1 errors will not be within a Hamming distance of t symbols from any code word and thus will be detected as having uncorrectable errors (UEs). From the foregoing, it will be apparent that to correct 2 or fewer errors and simultaneously detect 3 errors, a code must have a minimum Hamming distance of 6 symbols. Such codes are commonly referred to as double error correcting and triple error detecting (DEC-TED) codes. To give another example, codes with a minimum Hamming distance of 4 symbols can correct a single error and detect up to 2 errors, and are known as single error correcting and double error detecting (SEC-DED) codes.

[0013] Single error correcting and double error detecting (SEC-DED) codes have been widely used to protect computer memory subsystems from failures. As certain critical data such as storage protection keys in computers requires a higher level of error protection, SEC-DED codes may not be adequate. In this case, a double error correcting, triple error detecting (DEC-TED) code may be desired.

[0014] Error correction codes capable of correcting double errors and detecting triple errors can be constructed based on the well-known BCH (Bose-Chaudhuri-Hocquenghem) theory (see W. Peterson and E. J. Weldon Jr., Error-Correcting Codes, 1972, MIT Press). A primitive BCH DEC-TED code of length n=2^m−1 with 2m+1 check bits is obtained with a parity check matrix, each column vector k of which consists of 1, α^k and α^3k, where α is a primitive element of the finite field of 2^m elements. Olderdissen describes a rapid decoding of the primitive BCH DEC-TED codes in U.S. Pat. No. 4,556,977. On the other hand, a non-primitive BCH DEC-TED code of length n=2^m+1 with 2m+1 check bits can be obtained with a parity check matrix, each column vector k of which consists of 1 and β^k, where β is a primitive root of xⁿ−1 in the finite field of 2^2m elements. A non-primitive BCH DEC-TED code provides two more data bits than a BCH DEC-TED code with the same number of check bits. One drawback of the Olderdissen decoding scheme is that it is not applicable to non-primitive BCH DEC-TED codes. In U.S. Pat. No. 4,117,458, Burghard and Coletti describe a decoding scheme based on a brute force table-look-up approach for a non-BCH code of length 17 (=2⁴+1) with 8 data bits and 9 check bits. In addition, their error detection scheme is limited to triple errors. The decoding table does not detect multiple errors beyond three that are theoretically detectable.

[0015] Recent ECC design for computer applications requires the ability to detect memory address errors as well as the ability to isolate component failures with invalid data indicators (see for examples U.S. Pat. No. 6,457,154 for memory address error detection and U.S. Pat. No. 6,519,736 for failure isolation of computer components with invalid data indicators). For memory address error detection, extra data bits are required for encoding the parity of a memory address. A special data invalid indicator, also known as a special UE (SPUE) indicator, is generated when the data sent out of a particular computer component to the memory is known to be bad. As the special UEs come from different computer components, it is desirable to be able to identify the source that generates a particular special UE when the data associated with the special UE is fetched from the memory. To meet this requirement, extra data bits are also required for the encoding of the special UEs. In the prior art, a plurality of data bits are reserved for multiple special UEs, which is inefficient in the usage of ECC data bits, especially when the number of available ECC data bits is limited.

[0016] The present invention contemplates a method and apparatus for encoding special uncorrectable errors in an information handling system. The system is one in which a data word that is susceptible to a plurality of types of special uncorrectable errors and contains a plurality of regular data symbols that are used to encode data in the absence of a special uncorrectable error is encoded using an error correction code to generate a code word. In accordance with the invention, in the encoder, the regular data symbols are modified in the presence of a special uncorrectable error to indicate the type of special uncorrectable error. A special uncorrectable error symbol indicating the presence of a special uncorrectable error is appended to the regular data symbols to form an extended data word, which is encoded to generate a code word. By modifying the regular data symbols in this manner to indicate the type of special uncorrectable error, the present invention makes it unnecessary to use additional special uncorrectable error symbols for this purpose.

[0017] The regular data symbols and the uncorrectable error symbol are typically bits, and the code word is typically generated by generating one or more check symbols as a function of the symbols of the extended data word and appending the check symbols to the extended data word to form the code word. The extended data word may also contain an address parity symbol indicating the parity of a memory address in which the data is stored.

[0018] The code word is decoded to regenerate the special uncorrectable error symbol and the modified regular data symbols. Preferably, the decoder decodes the code word using an assumed value for the special uncorrectable error symbol indicating the absence of a special uncorrectable error, and detects the presence of an error in the assumed value of the special uncorrectable error symbol to regenerate that symbol. More particularly, in the decoder, a syndrome generator would generate a syndrome vector from the code word using an assumed value for the special uncorrectable error symbol indicating the absence of a special uncorrectable error. A syndrome decoder would then detect the presence of an error in this assumed value from the syndrome vector to detect presence of a special uncorrectable error. By using its error detection logic in this manner, the decoder makes it unnecessary to actually store or transmit the special uncorrectable error symbol.

[0019] Additional features and advantages are realized through the techniques of the present invention. Other embodiments and aspects of the invention are described in detail herein and are considered a part of the claimed invention. For a better understanding of the invention with advantages and features, refer to the description and to the drawings.

[0031] The detailed description explains the preferred embodiments of the invention, together with advantages and features, by way of example with reference to the drawings.

[0032] FIG. 1 is a block diagram representation of one embodiment of a computer system 100 utilizing the error correction code of the present invention. In the embodiment shown, computer system 100 includes a memory 110 connected to a main storage controller (MSC) 120 via a data bus 115, a cache 130 connected to MSC 120 via an MSC-cache interface 116, and a plurality of central processing units (CPUs) 140 connected to cache 130 via data buses 117. MSC-cache interface 116 and data buses 117 are used to transfer data between MSC 120 and cache 130 and between cache 130 and CPUs 140, respectively. Transfer of data between memory 110 and MSC 120, on the other hand, occurs via data bus 115. Thus, data bus 115 facilitates the reading of data from memory 110 as well as the writing of data to memory 110 by MSC 120.

[0033] A subset of the data area in cache 130 is a storage protection key area (SP KEY) 118, which contains storage protection keys generated to assure data integrity in the cache. In accordance with the present invention, storage protection keys from key area 118 are constantly updated and stored in memory 110, as is all data from the cache 130. However, because storage protection keys are considered critical data that require a higher level of reliability, a more potent ECC is selected for error protection of these keys. In particular, a DEC-TED code is used to correct all single or double errors, detect all triple errors and also detect a plurality of multiple errors in an encoded ECC word, as described further below.

[0034] In the embodiment shown, a storage protection key consists of 7 bits. In addition, one data bit is used for encoding of the memory address parity, while another data bit is used for encoding of two special uncorrectable errors (UEs). Thus, a total of 9 data bits are required for the ECC. This leads to an (18, 9) DEC-TED code that consists of 18 bits in a code word with 9 data bits and 9 check bits. Note that a special UE (SPUE) is a data validity indicator generated when the data sent out of a particular computer component to the memory is known to be bad. As the special UEs come from different computer components, it is desirable to be able to identify the source that generates a particular special UE when the data associated with the special UE is fetched from the memory.

[0035] FIG. 2 shows the ECC word structure according to one embodiment of the invention. The first 7 bits (bits 0–6) contain the original storage protection key data. The next 9 bits (bits 7–15) are check bits generated from bits 0–6 and bits 16–17 according to the ECC equations to be described shortly. Bit 16 (SPUE) is assigned for special UEs. Finally, an address parity bit AP (bit 17) is assigned for the parity of the memory address. Only the first 16 bits (bits 0–15) are stored in memory 110 (FIG. 1). Though bits 16 and 17 are used in the generation of check bits, they are not stored in memory 110. In reading data from memory 110, the address parity bit AP is made available to the ECC decoder, while the value of the SPUE bit is assumed to be zero. If, however, the syndrome decoder to be described detects an error at bit location 16 (assuming a bit value of zero at that location), then it determines that the SPUE bit is one and that bits 0–6 encode a SPUE. Thus, even though the SPUE bit is not stored as such, it is effectively encoded in the 16 bits of the code word that are stored in memory 110 through its use in generating the check bits 7–15.

[0036] The value of the SPUE bit is 0 for a valid storage protection key. A key is marked invalid when the data received from other components of the computer system is known to be bad. In such case, the value of SPUE is set to 1. Conventionally, plural data bits are used in order to differentiate the sources of the bad data. In the present invention, by contrast, only one data bit (bit 16) is used. To identify the source of the bad data, the associated key data bits (bits 0–6) are modified so that different sources of bad data are represented by different pattern of bits 0–6. For example, to differentiate bad data from cache 130 from bad data from MSC 120 (FIG. 1), two 7-bit patterns (0000000) and (1111111) can be assigned to bits 0–6 (FIG. 2). As an illustration, (0000000) may be assigned to be the pattern of bits 0–6 and SPUE bit 16 set to 1 if the bad data originated from cache 130, while (1111111) may be assigned to be the pattern of bits 0–6 and SPUE bit 16 set to 1 if the bad data originated from MSC 120. Other possible pattern pairs are (1010101, 0101010) and (0001111, 1110000).

[0037] Now suppose that one of the SPUE patterns is stored in memory 110 and then retrieved. In accordance with the ECC decoding method to be described, if there is no error in the memory, a unique error syndrome 010001111 is generated and the ECC decoding flags bit 16 to be in error. This indicates that the data received is associated with a SPUE. The pattern in bits 0–6 is then used to identify the original source of bad data. If there is an error in the memory in which the SPUE data resided, the ECC decoding would detect the presence of two errors, one of which is a memory error and the other of which is bit 16. The error syndrome depends on the location of the memory error. In any case, the syndrome is a double error syndrome and will be decoded as such because the code is capable of correcting double errors. Again, upon the error detection of bit 16 the source of the bad data can be isolated.

[0038] An ECC can be specified by a set of equations that all encoded ECC words have to satisfy. Let C=(c₀, c₁, c₂, . . . , c₁₇) be a 1×18 row vector defining a code word. The (17, 8) DEC-TED code described in U.S. Pat. No. 4,117,458 is a non-primitive BCH code that can be lengthened by one bit to yield a (18, 9) DEC-TED code. Including the all-one 18-bit vector in the code space does this. The (18, 9) DEC-TED code of the present invention is defined by the following two equations:
c₀β³+c₁β⁶+c₂β¹²+c₃β⁷+c₄β¹⁴+c₅β¹¹+c₆β⁵+c₇β+c₈β²+c₉β⁴+c₁₀β⁸+c₁₁β¹⁶+c₁₂β¹⁵+c₁₃β¹³+c₁₄β⁹+c₁₅0+c₁₆β¹⁰+c₁₇β⁰=0
c₀+c₁+c₂+c₃+c₄+c₅+c₆+c₇+c₈+c₉+c₁₀+c₁₁+c₁₂+c₁₃+c₁₄+c₁₅+c₁₆+c₁₇=0

[0039] The additions in the above equations are performed according to the rules of the finite field of 256 elements. The first equation specifies the (17, 9) code listed on page 494 of the above-identified work of Peterson et al. and guarantees that the number of nonzero terms is at least 5 for a nonzero code word. The second equation says that the number of nonzero terms in a code word is even, since each c_i is binary. As explained on page 119 of the same work, combining both equations means that the number of nonzero terms of a nonzero code word is at least 6, i.e., the Hamming distance of the code is 6. Note that the first equation involves all 17 unique powers of β. In this first equation, the terms of the powers of β are not arranged in a sequential order. However, the ordering is not critical; any ordering works.

[0040] The symbol β in the above equations is a primitive root of x¹⁷−1 in the finite field of 256 elements. Specifically, β=α¹⁵, where α is a root of the binary primitive polynomial x⁸+x⁷+x⁶+x+1. The same code space can be defined using an irreducible polynomial. However, in practical application, the particular choice presented here has been empirically shown to lead to a simpler implementation. Notice that α is primitive element of the finite field of 256 elements, and β is also an element of the same finite field.

[0041] The above equations that define the ECC can be expressed in matrix form as CH^t=0, where H^t denotes the transpose of the matrix H:

[0042] $\begin{array}{cccccccccccccccccc}{\beta }^3& {\beta }^6& {\beta }^{12}& {\beta }^7& {\beta }^{14}& {\beta }^{11}& {\beta }^5& {\beta }^1& B^2& {\beta }^4& {\beta }^8& {\beta }^{16}& {\beta }^{15}& {\beta }^{13}& {\beta }^9& 0& {\beta }^{10}& 1\\ 1& 1& 1& 1& 1& 1& 1& 1& 1& 1& 1& 1& 1& 1& 1& 1& 1& 1.\end{array}$

[0043] It can be shown that the code defined by matrix H is a DEC-TED code. In addition, the columns of the matrix can be permuted in any order without reducing the capability of error correction and error detection.

[0044] The finite field elements in matrix H can be expressed in binary vectors to facilitate implementation using digital circuitry. Specifically, matrix H can be transformed into a 9×18 binary matrix H₁:

[0045] $\begin{array}{cc}\begin{array}{cccccccccccccccccc}0& 0& 1& 1& 1& 0& 1& 0& 0& 0& 0& 1& 1& 1& 0& 0& 0& 1\\ 0& 0& 0& 0& 1& 1& 1& 1& 0& 1& 0& 0& 0& 0& 0& 0& 1& 0\\ 1& 1& 1& 0& 0& 1& 0& 1& 1& 0& 1& 1& 1& 1& 0& 0& 0& 0\\ 0& 0& 1& 0& 1& 0& 0& 1& 0& 1& 0& 1& 0& 0& 1& 0& 0& 0\\ 1& 0& 0& 0& 1& 0& 0& 0& 0& 0& 0& 1& 1& 1& 1& 0& 0& 0\\ 0& 1& 1& 0& 0& 0& 1& 1& 0& 0& 0& 1& 1& 1& 0& 0& 1& 0\\ 0& 0& 1& 1& 0& 0& 1& 1& 1& 0& 1& 1& 1& 0& 1& 0& 1& 0\\ 1& 1& 1& 1& 0& 1& 0& 1& 1& 0& 0& 0& 0& 0& 0& 0& 1& 0\\ 0& 0& 1& 0& 1& 0& 1& 1& 0& 1& 1& 1& 0& 1& 0& 1& 1& 0.\end{array}& \left(1\right)\end{array}$

[0046] The derivation of H₁ from H maybe briefly explained. Assume, as before, that α is a root of x⁸+x⁷+x⁶+x+1 and β=α¹⁵. Then a power of β can be expressed as a polynomial in terms of the powers of α. The coefficients of the polynomial are a binary 8-bit vector listed in the first 8 bits of a column vector in the H₁ matrix above. Now, the field element 1 corresponds to the 8-bit vector 10000000. The second row of the H matrix above is an all ones vector. It is translated into a 8-row binary matrix with the first row being all ones and the rest of the rows being all zeros, which can be discarded. This explains how the original matrix is translated into a 9-row binary matrix. However, the last row of the H₁ matrix above is not all ones. The all ones row vector has been replaced by the sum (exclusive OR) of all 9 row vectors so that each column contains an odd number of ones. If you add all 9 row vectors together, you obtain an all ones vector. There is no difference in the spaces defined by H and H₁.

[0047] In reading data from the memory, matrix H₁ is used to check if an 18-bit received vector R is a legitimate code word by calculating the syndrome S by the formula S=RH₁^t, where H₁^t is the transpose of the vector H₁. Vector R is assumed to be a code word if the syndrome S is an all zeros vector. If S is not an all zeros vector, the ECC decoder to be described is used to determine if R contains one or two errors and also to determine the associated error positions. The decoder is also used to determine if R contains detectable uncorrectable errors (UEs), which include the set of all triple errors and some higher-order errors that are also detectable, though not correctible. Let S=(s₀, s₁, s₂, . . . , s₈). The received bits marked with the ones in row i of matrix H₁ are summed together using exclusive OR (XOR) operations to obtain the value of s_i. Specifically, the syndrome bits are obtained by the following formulas.
s₀=XOR of input bits 2, 3, 4, 6, 11, 12, 13, 17
s₁=XOR of bits 4, 5, 6, 7, 9, 16
s₂=XOR of bits 0, 1, 2, 5, 7, 8, 10, 11, 12, 13
s₃=XOR of bits 2, 4, 7, 9, 11, 14
s₄=XOR of bits 0, 4, 11, 12, 13, 14
s₅=XOR of bits 1, 2, 6, 7, 11, 12, 13, 16
s₆=XOR of bits 2, 3, 6, 7, 8, 10, 11, 12, 14, 16
s₇=XOR of bits 0, 1, 2, 3, 5, 7, 8, 16
s₈=XOR of bits 2, 4, 6, 7, 9, 10, 11, 13, 15, 16.  (2)

[0048] Let us label the columns of matrix H₁ as columns 0, 1, . . . , 17. For the generation of check bits, H₁ is multiplied by the inverse of the matrix formed by its columns 7–15 to obtain the matrix H₂.

[0049] $\begin{array}{cc}\begin{array}{cccccccccccccccccc}0& 1& 0& 1& 1& 0& 0& 1& 0& 0& 0& 0& 0& 0& 0& 0& 1& 1\\ 1& 0& 1& 0& 1& 1& 0& 0& 1& 0& 0& 0& 0& 0& 0& 0& 0& 1\\ 0& 1& 0& 1& 0& 1& 1& 0& 0& 1& 0& 0& 0& 0& 0& 0& 0& 1\\ 0& 0& 1& 0& 1& 0& 1& 0& 0& 0& 1& 0& 0& 0& 0& 0& 1& 1\\ 1& 0& 0& 1& 0& 1& 0& 0& 0& 0& 0& 1& 0& 0& 0& 0& 1& 1\\ 1& 1& 0& 0& 1& 0& 1& 0& 0& 0& 0& 0& 1& 0& 0& 0& 0& 1\\ 0& 1& 1& 0& 0& 1& 0& 0& 0& 0& 0& 0& 0& 1& 0& 0& 1& 1\\ 1& 0& 1& 1& 0& 0& 1& 0& 0& 0& 0& 0& 0& 0& 1& 0& 0& 1\\ 1& 1& 1& 1& 1& 1& 1& 0& 0& 0& 0& 0& 0& 0& 0& 1& 1& 1.\end{array}& \left(3\right)\end{array}$

[0050] Notice that columns 7–15 of H₂ form a 9×9 identity matrix. The value of each check bit is calculated from a row vector of H₂. Let C=(c₀, c₁, c₂, . . . , c₁₇) be a code word. Since for a properly formed code word C, CH₂^t=0, each row of H₂ is in effect a statement that the XOR sum of a given check bit and the data bits indicated by the ones in the row is zero or, equivalently, that the check bit is the XOR sum of those data bits. Given the values of bits 0–6 and bits 16–17 in the code word, the values of bits 7–15 are calculated from the following responding to the rows of H₂) in terms of exclusive OR operations:
c₇=XOR of bits 1, 3, 4, 16, 17
c₈=XOR of bits 0, 2, 4, 5, 17
c₉=XOR of bits 1, 3, 5, 6, 17
c₁₀=XOR of bits 2, 4, 6, 16, 17
c₁₁=XOR of bits 0, 3, 5, 16, 17
c₁₂=XOR of bits 0, 1, 4, 6, 17
c₁₃=XOR of bits 1, 2, 5, 16, 17
c₁₄=XOR of bits 0, 2, 3, 6, 17
c₁₅=XOR of bits 0, 1, 2, 3, 4, 5, 6, 16, 17.  (4)

[0051] To calculate the check bits above, it is not necessary to obtain an explicit value for a generator matrix G. However, it can readily be shown (see, e.g., chapter 3 of the Peterson et al. reference identified above) that matrix H₂ is a parity check matrix for a code having the following a generator matrix G:

[0052] $\begin{array}{cccccccccccccccccc}1& 0& 0& 0& 0& 0& 0& 0& 1& 0& 0& 1& 1& 0& 1& 1& 0& 0\\ 0& 1& 0& 0& 0& 0& 0& 1& 0& 1& 0& 0& 1& 1& 0& 1& 0& 0\\ 0& 0& 1& 0& 0& 0& 0& 0& 1& 0& 1& 0& 0& 1& 1& 1& 0& 0\\ 0& 0& 0& 1& 0& 0& 0& 1& 0& 1& 0& 1& 0& 0& 1& 1& 0& 0\\ 0& 0& 0& 0& 1& 0& 0& 1& 1& 0& 1& 0& 1& 0& 0& 1& 0& 0\\ 0& 0& 0& 0& 0& 1& 0& 0& 1& 1& 0& 1& 0& 1& 0& 1& 0& 0\\ 0& 0& 0& 0& 0& 0& 1& 0& 0& 1& 1& 0& 1& 0& 1& 1& 0& 0\\ 0& 0& 0& 0& 0& 0& 0& 1& 0& 0& 1& 1& 0& 1& 0& 1& 1& 0\\ 0& 0& 0& 0& 0& 0& 0& 1& 1& 1& 1& 1& 1& 1& 1& 1& 0& 1\end{array}$

[0053] Columns 0–6 and 16–17 of G form an identity matrix, while columns 7–15 form the transpose of the matrix formed by columns 0–6 and 7–15 of H₂. Since columns 0–6 and 16–17 of G form an identity matrix, bits 0–6 and 16–17 of a code word C are simply the corresponding bits of the original data word (hence their label as information bits). Check bits 7–15 may alternatively be calculated using columns 7–15 of generator matrix G (with the bits SPUE and AP being regarded as bits 7 and 8 of the original data word). Since columns 7–15 of G are simply the transpose of the matrix formed by columns 0–6 and 7–15 of H₂, the resulting XOR operations are identical to the operations (4) set forth above.

[0054] As described earlier, the syndrome S of a received vector R is used in the decoding to determine the nature of the errors if S is not zero. Let sp be the parity of the syndrome bits. That is, sp is the XOR of all 9 syndrome bits. Since each column of the parity check matrix H₁ contains an odd number of ones, an error in any one bit of the received vector R will invert (i.e., flip) an odd number of syndrome bits, thereby inverting their XOR sum sp. Accordingly, sp=0 if there is an even number of errors in R, and sp=1 if there is an odd number of errors in R. Thus, the decoder can easily determine whether the number of errors is even or odd. If the number of errors is odd, the decoder assumes that there is one error and it goes on to determine the single error position. On other hand, if the number of errors is even, the decoder assumes that there are two errors and it goes on to determine the locations of two errors.

[0055] Single error position is relatively easy to determine. If the first 8 bits of column i of H₁ are identical to the first 8 syndrome bits, then bit i is identified as the single error location.

[0056] Double error positions are not trivial to identify. An equation with the error locations as unknown variables has to be derived and solved. Let x₁ and x₂ be two unknown variables representing the error locations in terms of the finite field of 256 elements. Each variable is either a power of β or 0. Let S₁ be the first 8 bits of the syndrome and consider S₁ as an element of the finite field. From the first row of matrix H and the equation HC^t=0, the syndrome is related to the error locations x₁ and x₂ by the equation S₁=x₁+x₂. In addition, it can be shown that error locations x₁ and x₂ are roots of x¹⁸−x. That is, x₁¹⁸−x₁=0 and x₂¹⁸−x₂=0. Combining all these relations, it can be shown that x₁ and x₂ are solutions to the following equation with x as the unknown variable.
S₁¹⁷=x¹⁶ S₁+x S₁¹⁶  (5)

[0057] The values of x₁ and x₂ are obtained by solving equation (5) for x. Special attention is required in the case that one of the two errors is located at bit 15, which has a 0 as the field element in matrix H. In this case, equation (5) is not used. Instead, S₁ is treated as a single error syndrome, and the second error location is identified by matching S₁ with the column vectors of the first 8 rows of H₁.

[0058] Let E_i be the error indicator for bit i with the property that E_i=0 if bit i is not in error and E_i=1 if bit i is in error. The decoder is used to generate the values of E_i for all bit positions. One decoding algorithm (Algorithm A) is described below:

• • [0059] 1. If all 9-syndrome bits are zero, there is no error and the received data is not altered. Exit the algorithm.
  • [0060] 2. Set E₁₅=1 if (a) S₁=0 and sp=1; or (b) S₁¹⁷=1 and sp=0, where S₁ is the first 8 bits of the syndrome S and is considered a field element, and sp is the exclusive OR sum of all syndrome bits.
  • [0061] 3. For 0≦i≦17 and i≠15: set E_i=1 if (a) S₁=column i of the first 8 rows of H₁ and S₁¹⁷=1; or (b) the field element x_i of column i of the first 8 rows of H₁ satisfies x_i¹⁶ S+x_i S₁¹⁶=S₁¹⁷, S₁¹⁷≠1, S₁≠0, and sp=0.
  • [0062] 4. Set UE=1 if (a) S₁¹⁷≠1, S₁≠0 and sp=1; or (b) S₁¹⁷≠1, sp=0 and there is no solution to x_i¹⁶ S₁+x_i S₁¹⁶=S₁¹⁷.

[0063] If the value of E₁₆ obtained from Algorithm A is 1, the received data R is a SPUE. In this case, the data bits in bits 0–6 after error correction are used to determine the nature of the SPUE, which results in failure isolation.

[0064] The components of the 8-bit vector S₁¹⁷ are not independent. It can be shown that bits 0, 1, 2, and 4 are linearly independent and that the remaining 4 bits can be derived from these independent bits. (More generally, the exact positions of the independent bits depend on the polynomial defining the field, but the number of independent bits is always 4.) Thus, S₁¹⁷ can be replaced by S₁¹⁷(0, 1, 2, 4), which represents bits 0, 1, 2, 4 of S₁¹⁷ in Algorithm A. These 4 bits can be obtained from the following formulas:
S₁¹⁷(0)=XOR of S(0), S(2), S(3), S(6), S(0)S(1), S(0)S(5), S(0)S(7), S(1)S(2), S(1)S(6), S(1)S(7), S(2)S(4), S(2)S(6), S(3)S(6), S(3)S(7), S(4)S(5), S(4)S(7)
S₁¹⁷(1)=XOR of S(1), S(3), S(4), S(7), S(0)S(2), S(0)S(3), S(0)S(6), S(1)S(2), S(1)S(6), S(2)S(3), S(2)S(7), S(3)S(5), S(3)S(7), S(4)S(7), S(5)S(6)
S₁¹⁷(2)=XOR of S(1), S(2), S(5), S(0)S(1), S(0)S(5), S(0)S(6), S(0)S(7), S(1)S(3), S(1)S(5), S(2)S(5), S(2)S(6), S(3)S(4), S(3)S(6), S(3)S(7), S(4)S(7), S(5)S(7)
S₁¹⁷(4)=XOR of S(1), S(2), S(3), S(4), S(6), S(0)S(1), S(0)S(2), S(0)S(4), S(0)S(5), S(1)S(2), S(1)S(4), S(1)S(5), S(1)S(6), S(2)S(3), S(2)S(4), S(2)S(5), S(2), S(4)S(5), S(4)S(6), S(5)S(7), S(6)S(7)  (6)

[0065] In equation (6), S(i) denotes bit i of the syndrome S and S(i)S(j) is the product of S(i) and S(j).

[0066] One embodiment of the hardware implementation of the ECC encoding and decoding is described next.

[0067] FIG. 3a shows an encoder 200 for ECC encoding. Input data (DATA IN) 210 consists of the 7-bit storage protection key data to be encoded. Two other inputs to the encoder 200 are the special uncorrectable error bit SPUE and the memory address parity bit AP shown in FIG. 2. The key data 210 is sent to a modification circuit (DATA MODIFIED) 240 that has SPUE as the other input. If SPUE is 0, the key data 210 is not modified. On the other hand, if SPUE is 1, circuit 240 modifies the key data 210 according to the pre-defined SPUE data patterns as described above. The output of circuit 240 appears as output data (DATA OUT) 220. It is also sent to a check symbol generator 250 comprising an XOR gate array containing XOR circuits 260–268 (XOR0–XOR8). XOR circuits 260–268 generate check bits according to equation (4). Thus, FIG. 3b illustrates the generation of check bit 0 (bit 7 of the encoded word) using XOR circuit 260 according to equation (4).

[0068] The output of the XOR gate array 250 appears as check bits 230. Bits 0–15 of the encoded ECC word (bits 0–17) consist of the output data 220 (bits 0–6) and check bits 230 (bits 7–15). As noted above, bits 0–15 of the code word are stored in memory 110, while bit 16 (SPUE) is assumed to be zero and bit 17 (AP) of the code word is independently regenerated when bits 0–15 of the code word are later read out of the memory 110.

[0069] FIG. 4 is a block diagram of a decoder 300 for data read from memory 110. The received 16-bit word is stored in an input register (DATA IN) 400, of which the first 7 bits represent the data bits and the last 9 bits represent the check bits. The entire 16-bit word in input register 400 is sent to a syndrome generator (SYNDROME GEN) 500 that has the address parity bit AP as another input. (Since SPUE information is not available, it is assumed to be zero.) Syndrome generator 500 generates as an output all 9 syndrome bits, which are sent to a syndrome decoder (SYNDROME DECODE) 600 for the generation of error location indicators E_i and a one-bit uncorrectable error indicator UE. The error indicators E_i for 0≦i≦6 from syndrome decoder 600 and the 7 data bits from input register 400 are XORed bitwise by a data correction circuit 700 to produce corrected output data.

[0070] FIG. 5a shows the syndrome generator 500. The inputs are the 16 received bits—7 data bits (0–6) and 9 check bits (7–15)—stored in input register 400 and the AP bit (17). The 9-bit output is stored in a syndrome register (SYNDROME REG) 520. The 9 syndrome bits are generated by XOR blocks 530–538, each of which contains logic for performing an XOR operation specified in equation (2).

[0071] FIG. 5b shows, by way of illustration, the input bits (2–4, 6, 11–13, AP=17) for XOR block 530, which generates syndrome bit 0. Notice that bit 17 appears only once in equation (2); only XOR block 530 takes AP as an input.

[0072] FIG. 6 shows the syndrome decoder 600, which generates the error indicator E₁₅ as well as error indicators E_i for i≠15 in accordance with steps 2 and 3, respectively, of Algorithm A. Although not shown in FIG. 6, syndrome decoder 600 also contains uncorrectable error (UE) detection logic 660 (FIG. 8) for generating an uncorrectable error (UE) signal indicating the presence of an uncorrectable error. In this figure, for i≠15, E_{1, i}=1 indicates that S₁=column i of the first 8 rows of H₁, while E_{2, i}=1 indicates that the field element x_i of column i of the first 8 rows of H₁ satisfies the equation x_i¹⁶S+x_i S₁¹⁶=S₁¹⁷. Each AND block 632–636 outputs the logical AND of its inputs, while each OR block 641–643 outputs the logical OR of its inputs and each inverter block 651–653 (denoted by a triangle) outputs the logical inverse of its input. A syndrome parity bit sp is generated by an XOR circuit 602 that outputs the exclusive OR of all 9 syndrome bits, stored in a syndrome register (SYNDROME S) 601. Also, a vector S₁ (603) is extracted as the first 8 bits of the syndrome vector S.

[0073] Single error location logic 610 performs the function of matching input S₁ (603) with the column vectors of the first 8 rows of H₁ of equation (1). The output bits are single error indicators E_{1, i}. FIG. 7 illustrates by way of example the circuit for generating the single error indicator E_{1, 0}. (The triangles in the figure denote logical inverters.) The circuit matches an 8-bit input (0–7) with column 0 of the first 8 rows of H₁ to produce an output E_{1, 0} of 1 if and only if each input bit matches the corresponding bit of that column of H₁.

[0074] Logic 630 in FIG. 6 generates as the output S₁¹⁷(0, 1, 2, 4) for bits 0, 1, 2, and 4 of S₁¹⁷ according to equation (6). This generated output is used to represent S₁¹⁷.

[0075] Double error location logic 620 generates the double error indicators E_{2, i}. The inputs to logic 620 include S₁¹⁷(0, 1, 2, 4) from logic 630 and S₁ (603). The outputs E_{2, i} of logic 620 are generated as follows. Logic 620 first generates a set of comparison bits Fi from syndrome bits 0–7 according to the following formulas:
F1=syndrome bit 7,
F3=XOR of syndrome bits 6, 7,
F4=syndrome bit 5,
F14=XOR of syndrome bits 4, 5, 6,
F18=XOR of syndrome bits 3, 6,
F20=XOR of syndrome bits 3, 5,
F21=XOR of syndrome bits 3, 5, 7,
F23=XOR of syndrome bits 3, 5, 6, 7,
F26=XOR of syndrome bits 3, 4, 6,
F29=XOR of syndrome bits 3, 4, 5, 7,
F30=XOR of syndrome bits 3, 4, 5, 6,
F40=XOR of syndrome bits 2, 4,
F42=XOR of syndrome bits 2, 4, 6,
F44=XOR of syndrome bits 2, 4, 5,
F45=XOR of syndrome bits 2, 4, 5, 7,
F49=XOR of syndrome bits 2, 3, 7,
F50=XOR of syndrome bits 2, 3, 6,
F52=XOR of syndrome bits 2, 3, 5,
F58=XOR of syndrome bits 2, 3, 4, 6,
F69=XOR of syndrome bits 1, 5, 7,
F71=XOR of syndrome bits 1, 5, 6, 7,
F74=XOR of syndrome bits 1, 4, 6,
F85=XOR of syndrome bits 1, 3, 5, 7,
F86=XOR of syndrome bits 1, 3, 5, 6,
F89=XOR of syndrome bits 1, 3, 4, 7,
F96=XOR of syndrome bits 1, 2,
F98=XOR of syndrome bits 1, 2, 6,
F107=XOR of syndrome bits 1, 2, 4, 6, 7,
F108=XOR of syndrome bits 1, 2, 4, 5,
F123=XOR of syndrome bits 1, 2, 3, 4, 6, 7,
F130=XOR of syndrome bits 0, 6,
F131=XOR of syndrome bits 0, 6, 7,
F135=XOR of syndrome bits 0, 5, 6, 7,
F137=XOR of syndrome bits 0, 4, 7,
F139=XOR of syndrome bits 0, 4, 6, 7,
F142=XOR of syndrome bits 0, 4, 5, 6,
F143=XOR of syndrome bits 0, 4, 5, 6, 7,
F145=XOR of syndrome bits 0, 3, 7,
F148=XOR of syndrome bits 0, 3, 5,
F160=XOR of syndrome bits 0, 2,
F162=XOR of syndrome bits 0, 2, 6,
F168=XOR of syndrome bits 0, 2, 4,
F177=XOR of syndrome bits 0, 2, 3, 7,
F178=XOR of syndrome bits 0, 2, 3, 6,
F180=XOR of syndrome bits 0, 2, 3, 5,
F182=XOR of syndrome bits 0, 2, 3, 5, 6,
F183=XOR of syndrome bits 0, 2, 3, 5, 6, 7,
F184=XOR of syndrome bits 0, 2, 3, 4,
F190=XOR of syndrome bits 0, 2, 3, 4, 5, 6,
F192=XOR of syndrome bits 0, 1,
F198=XOR of syndrome bits 0, 1, 5, 6,
F199=XOR of syndrome bits 0, 1, 5, 6, 7,
F202=XOR of syndrome bits 0, 1, 4, 6,
F203=XOR of syndrome bits 0, 1, 4, 6, 7,
F204=XOR of syndrome bits 0, 1, 4, 5,
F206=XOR of syndrome bits 0, 1, 4, 5, 6,
F209=XOR of syndrome bits 0, 1, 3, 7,
F210=XOR of syndrome bits 0, 1, 3, 6,
F213=XOR of syndrome bits 0, 1, 3, 5, 7,
F215=XOR of syndrome bits 0, 1, 3, 5, 6, 7,
F223=XOR of syndrome bits 0, 1, 3, 4, 5, 6, 7,
F226=XOR of syndrome bits 0, 1, 2, 6,
F232=XOR of syndrome bits 0, 1, 2, 4,
F234=XOR of syndrome bits 0, 1, 2, 4, 6,
F237=XOR of syndrome bits 0, 1, 2, 4, 5, 7,
F240=XOR of syndrome bits 0, 1, 2, 3,
F247=XOR of syndrome bits 0, 1, 2, 3, 5, 6, 7,
F251=XOR of syndrome bits 0, 1, 2, 3, 4, 6, 7,

[0076] Note that the syndrome bits participating in the XOR operations for F_i correspond to the binary representation of the integer i. For example, the binary representation of 226 is 11100010, and F226 is the XOR of syndrome bits 0, 1, 2, 6.

[0077] Logic 620 then generates a set of error values E_{2, j} for 0≦j≦17 and j≠15 by comparing the bits of S₁¹⁷(0, 1, 2, 4) with selected bits Fi, as indicated below. Each generated error value E_{2, j} is one if the bits all match and is otherwise zero. More particularly:
E_{2, 0}=1 if S₁¹⁷(0, 1, 2, 4)=(F183, F232, F139, F29)
E_{2, 1}=1 if S₁¹⁷(0, 1, 2, 4)=(F26, F251, F123, F18)
E_{2, 2}=1 if S₁¹⁷(0, 1, 2, 4)=(F44, F168, F199, F23)
E_{2, 3}=1 if S₁¹⁷(0, 1, 2, 4)=(F237, F206, F107, F3)
E_{2, 4}=1 if S₁¹⁷(0, 1, 2, 4)=(F192, F180, F137, F4)
E_{2, 5}=1 if S₁¹⁷(0, 1, 2, 4)=(F49, F203, F14, F85)
E_{2, 6}=1 if S₁¹⁷(0, 1, 2, 4)=(F30, F198, F30, F98)
E_{2, 7}=1 if S₁¹⁷(0, 1, 2, 4)=(F202, F184, F20, F213)
E_{2, 8}=1 if S₁¹⁷(0, 1, 2, 4)=(F226, F45, F42, F178)
E_{2, 9}=1 if S₁¹⁷(0, 1, 2, 4)=(F160, F135, F223, F142)
E_{2, 10}=1 if S₁¹⁷(0, 1, 2, 4)=(F58, F21, F182, F148)
E_{2, 11}=1 if S₁¹⁷(0, 1, 2, 4)=(F209, F145, F74, F215)
E_{2, 12}=1 if S₁¹⁷(0, 1, 2, 4)=(F210, F52, F1, F247)
E_{2, 13}=1 if S₁¹⁷(0, 1, 2, 4)=(F162, F240, F177, F190)
E_{2, 14}=1 if S₁¹⁷(0, 1, 2, 4)=(F86, F96, F234, F143)
E_{2, 16}=1 if S₁¹⁷(0, 1, 2, 4)=(F131, F204, F89, F40)
E_{2, 17}=1 if S₁¹⁷(0, 1, 2, 4)=(F69, F50, F71, F108).

[0078] Syndrome decoder 600 combines the outputs of single error location logic 610 and double error location logic 620 to generate an error locator bit E_i for each bit i, where 0≦i≦17, in accordance with steps 2 and 3 of algorithm A. To accomplish this, a gate array 631 responsive to logic 630 produces a output of one whenever S₁¹⁷=1, that is, if S₁¹⁷(0, 1, 2, 4)=(1, 0, 0, 0). Also, an OR gate 641 produces a zero whenever S₁=0, that is, whenever the first eight bits of the syndrome vector S are all zero.

[0079] For i=15, in step 2 of the algorithm, if S₁=0 and sp=1, then both inputs to AND gate 633 are one, causing OR gate 642 to output a one to generate an E₁₅ locator bit. Similarly, if S₁¹⁷=1 sp=0, then both inputs to AND gate 635 are one, against causing OR gate 642 to output a one to generate an E₁₅ locator bit. In the first instance, the E₁₅ locator bit indicates a single error at bit location 15, while in the second, the E₁₅ locator bit indicates a double error involving bit 15 and one other bit location (as indicated by another E_i).

[0080] For 0≦i≦17 and i≠15, in step 3 of the algorithm, if S₁=column i of the first 8 rows of H₁ and S₁¹⁷=1, then logic 610 (E_{1, i}) and logic 631 input ones to the AND gate 632 for the particular i, causing the OR gate 643 for the particular i to generate an E_i locator bit, this time for a single error at bit location i. Similarly, if the field element x_i of column i of the first 8 rows of H₁ satisfies the equation x_i¹⁶ S+x_i S₁¹⁶=S₁¹⁷, then logic 620 (E_{2, i}) and AND gate 634 input ones to the AND gate 636 for the particular i, again causing the OR gate 643 for the particular i to generate an E_i locator bit, this time for a double error at bit location i and one other location (as indicated by another E_i).

[0081] FIG. 8 shows uncorrectable error (UE) detection logic 660 for generating the uncorrectable error (UE) signal (FIG. 4). UE detection logic 660 is a part of syndrome decoder 600. Each of the AND blocks 661, 662, and 663 outputs the logical AND of its inputs and each of the OR blocks 664 outputs the logical OR of its inputs. NOR 665 outputs the inverse of the logical OR of the inputs E_{2, i} from the output of logic 620.

[0082] UE detection logic 660 implements in hardware step 4 of Algorithm A. Thus, if (1) S₁¹⁷≠1, (2) S₁≠0 and (3) sp=1, then (1) circuit 631 outputs a zero to inverter 651, inverter to supply a first one to AND gate 661, (2) OR gate 641 supplies a second one to AND gate 661, causing that gate to supply a first one to AND gate 662, and (3) the sp line supplies a second one to AND gate 662; all of this causes AND gate 662 to input a one to OR gate 664, resulting in a one on the UE line. Alternatively, if (1) S₁¹⁷≠1, (2) sp=0 and (3) there is no solution to the equation x_i¹⁶ S₁+x_i S₁^16=S₁¹⁷, then (1) circuit 631 outputs a zero to inverter 651 causing that inverter to supply a first one to AND gate 663, (2) the sp line supplies a zero to inverter 652, causing that inverter to supply a second one to AND gate 663, and (3) logic 620 supplies all zeros to NOR gate 665, causing that gate to output a third one to AND gate 663; all of this causes AND gate 663 to input a one to OR gate 664, likewise resulting in a one on the UE line.

[0083] If (as shown in FIG. 4) errors in the check bits (7–15) are not corrected, the associated error locators need not be generated for this purpose. In such case, the circuits in single error location logic 610 for generating E_{1, i} for 7≦i≦14, as well as the circuits for generating E₁₅, may be omitted. On the other hand, since step 4 of Algorithm A involves all 17 E_{2, i} values (0≦i≦17, i≠15), it is still necessary for double error location logic 620 to generate all of these E_{2, i} values for UE detection logic 660 to fully implement this step of the algorithm. If the E_{2, i} values for 7≦i≦14 are not generated in logic 620, then the UE detection is reduced to determining whether S¹⁷≠1, S≠0 and sp=1. The decoder 300 still provides double error correcting and triple error detecting ability. However, it does not detect as many errors beyond triple errors as the full decoder. Note also that if logic 620 does generate E_{2, i} values for 7≦i≦14, the number of Fi XOR functions is reduced from 68 to 36. Thus, the overall syndrome decoding circuitry would be reduced nearly by half, but at the expense of reducing the probability of detecting four or more errors.

[0084] The capabilities of the present invention can be implemented in software, firmware, hardware or some combination thereof.

[0085] As one example, one or more aspects of the present invention can be included in an article of manufacture (e.g., one or more computer program products) having, for instance, computer usable media. The media has embodied therein, for instance, computer readable program code means for providing and facilitating the capabilities of the present invention. The article of manufacture can be included as a part of a computer system or sold separately.

[0086] Additionally, at least one program storage device readable by a machine, tangibly embodying at least one program of instructions executable by the machine to perform the capabilities of the present invention can be provided.

[0087] The flow diagrams depicted herein are just examples. There may be many variations to these diagrams or the steps (or operations) described therein without departing from the spirit of the invention. For instance, the steps may be performed in a differing order, or steps may be added, deleted or modified. All of these variations are considered a part of the claimed invention.

[0088] While the preferred embodiment to the invention has been described, it will be understood that those skilled in the art, both now and in the future, may make various improvements and enhancements which fall within the scope of the claims which follow. These claims should be construed to maintain the proper protection for the invention first described.