cannam@89: /* cannam@89: * gzlog.c cannam@89: * Copyright (C) 2004, 2008 Mark Adler, all rights reserved cannam@89: * For conditions of distribution and use, see copyright notice in gzlog.h cannam@89: * version 2.0, 25 Apr 2008 cannam@89: */ cannam@89: cannam@89: /* cannam@89: gzlog provides a mechanism for frequently appending short strings to a gzip cannam@89: file that is efficient both in execution time and compression ratio. The cannam@89: strategy is to write the short strings in an uncompressed form to the end of cannam@89: the gzip file, only compressing when the amount of uncompressed data has cannam@89: reached a given threshold. cannam@89: cannam@89: gzlog also provides protection against interruptions in the process due to cannam@89: system crashes. The status of the operation is recorded in an extra field cannam@89: in the gzip file, and is only updated once the gzip file is brought to a cannam@89: valid state. The last data to be appended or compressed is saved in an cannam@89: auxiliary file, so that if the operation is interrupted, it can be completed cannam@89: the next time an append operation is attempted. cannam@89: cannam@89: gzlog maintains another auxiliary file with the last 32K of data from the cannam@89: compressed portion, which is preloaded for the compression of the subsequent cannam@89: data. This minimizes the impact to the compression ratio of appending. cannam@89: */ cannam@89: cannam@89: /* cannam@89: Operations Concept: cannam@89: cannam@89: Files (log name "foo"): cannam@89: foo.gz -- gzip file with the complete log cannam@89: foo.add -- last message to append or last data to compress cannam@89: foo.dict -- dictionary of the last 32K of data for next compression cannam@89: foo.temp -- temporary dictionary file for compression after this one cannam@89: foo.lock -- lock file for reading and writing the other files cannam@89: foo.repairs -- log file for log file recovery operations (not compressed) cannam@89: cannam@89: gzip file structure: cannam@89: - fixed-length (no file name) header with extra field (see below) cannam@89: - compressed data ending initially with empty stored block cannam@89: - uncompressed data filling out originally empty stored block and cannam@89: subsequent stored blocks as needed (16K max each) cannam@89: - gzip trailer cannam@89: - no junk at end (no other gzip streams) cannam@89: cannam@89: When appending data, the information in the first three items above plus the cannam@89: foo.add file are sufficient to recover an interrupted append operation. The cannam@89: extra field has the necessary information to restore the start of the last cannam@89: stored block and determine where to append the data in the foo.add file, as cannam@89: well as the crc and length of the gzip data before the append operation. cannam@89: cannam@89: The foo.add file is created before the gzip file is marked for append, and cannam@89: deleted after the gzip file is marked as complete. So if the append cannam@89: operation is interrupted, the data to add will still be there. If due to cannam@89: some external force, the foo.add file gets deleted between when the append cannam@89: operation was interrupted and when recovery is attempted, the gzip file will cannam@89: still be restored, but without the appended data. cannam@89: cannam@89: When compressing data, the information in the first two items above plus the cannam@89: foo.add file are sufficient to recover an interrupted compress operation. cannam@89: The extra field has the necessary information to find the end of the cannam@89: compressed data, and contains both the crc and length of just the compressed cannam@89: data and of the complete set of data including the contents of the foo.add cannam@89: file. cannam@89: cannam@89: Again, the foo.add file is maintained during the compress operation in case cannam@89: of an interruption. If in the unlikely event the foo.add file with the data cannam@89: to be compressed is missing due to some external force, a gzip file with cannam@89: just the previous compressed data will be reconstructed. In this case, all cannam@89: of the data that was to be compressed is lost (approximately one megabyte). cannam@89: This will not occur if all that happened was an interruption of the compress cannam@89: operation. cannam@89: cannam@89: The third state that is marked is the replacement of the old dictionary with cannam@89: the new dictionary after a compress operation. Once compression is cannam@89: complete, the gzip file is marked as being in the replace state. This cannam@89: completes the gzip file, so an interrupt after being so marked does not cannam@89: result in recompression. Then the dictionary file is replaced, and the gzip cannam@89: file is marked as completed. This state prevents the possibility of cannam@89: restarting compression with the wrong dictionary file. cannam@89: cannam@89: All three operations are wrapped by a lock/unlock procedure. In order to cannam@89: gain exclusive access to the log files, first a foo.lock file must be cannam@89: exclusively created. When all operations are complete, the lock is cannam@89: released by deleting the foo.lock file. If when attempting to create the cannam@89: lock file, it already exists and the modify time of the lock file is more cannam@89: than five minutes old (set by the PATIENCE define below), then the old cannam@89: lock file is considered stale and deleted, and the exclusive creation of cannam@89: the lock file is retried. To assure that there are no false assessments cannam@89: of the staleness of the lock file, the operations periodically touch the cannam@89: lock file to update the modified date. cannam@89: cannam@89: Following is the definition of the extra field with all of the information cannam@89: required to enable the above append and compress operations and their cannam@89: recovery if interrupted. Multi-byte values are stored little endian cannam@89: (consistent with the gzip format). File pointers are eight bytes long. cannam@89: The crc's and lengths for the gzip trailer are four bytes long. (Note that cannam@89: the length at the end of a gzip file is used for error checking only, and cannam@89: for large files is actually the length modulo 2^32.) The stored block cannam@89: length is two bytes long. The gzip extra field two-byte identification is cannam@89: "ap" for append. It is assumed that writing the extra field to the file is cannam@89: an "atomic" operation. That is, either all of the extra field is written cannam@89: to the file, or none of it is, if the operation is interrupted right at the cannam@89: point of updating the extra field. This is a reasonable assumption, since cannam@89: the extra field is within the first 52 bytes of the file, which is smaller cannam@89: than any expected block size for a mass storage device (usually 512 bytes or cannam@89: larger). cannam@89: cannam@89: Extra field (35 bytes): cannam@89: - Pointer to first stored block length -- this points to the two-byte length cannam@89: of the first stored block, which is followed by the two-byte, one's cannam@89: complement of that length. The stored block length is preceded by the cannam@89: three-bit header of the stored block, which is the actual start of the cannam@89: stored block in the deflate format. See the bit offset field below. cannam@89: - Pointer to the last stored block length. This is the same as above, but cannam@89: for the last stored block of the uncompressed data in the gzip file. cannam@89: Initially this is the same as the first stored block length pointer. cannam@89: When the stored block gets to 16K (see the MAX_STORE define), then a new cannam@89: stored block as added, at which point the last stored block length pointer cannam@89: is different from the first stored block length pointer. When they are cannam@89: different, the first bit of the last stored block header is eight bits, or cannam@89: one byte back from the block length. cannam@89: - Compressed data crc and length. This is the crc and length of the data cannam@89: that is in the compressed portion of the deflate stream. These are used cannam@89: only in the event that the foo.add file containing the data to compress is cannam@89: lost after a compress operation is interrupted. cannam@89: - Total data crc and length. This is the crc and length of all of the data cannam@89: stored in the gzip file, compressed and uncompressed. It is used to cannam@89: reconstruct the gzip trailer when compressing, as well as when recovering cannam@89: interrupted operations. cannam@89: - Final stored block length. This is used to quickly find where to append, cannam@89: and allows the restoration of the original final stored block state when cannam@89: an append operation is interrupted. cannam@89: - First stored block start as the number of bits back from the final stored cannam@89: block first length byte. This value is in the range of 3..10, and is cannam@89: stored as the low three bits of the final byte of the extra field after cannam@89: subtracting three (0..7). This allows the last-block bit of the stored cannam@89: block header to be updated when a new stored block is added, for the case cannam@89: when the first stored block and the last stored block are the same. (When cannam@89: they are different, the numbers of bits back is known to be eight.) This cannam@89: also allows for new compressed data to be appended to the old compressed cannam@89: data in the compress operation, overwriting the previous first stored cannam@89: block, or for the compressed data to be terminated and a valid gzip file cannam@89: reconstructed on the off chance that a compression operation was cannam@89: interrupted and the data to compress in the foo.add file was deleted. cannam@89: - The operation in process. This is the next two bits in the last byte (the cannam@89: bits under the mask 0x18). The are interpreted as 0: nothing in process, cannam@89: 1: append in process, 2: compress in process, 3: replace in process. cannam@89: - The top three bits of the last byte in the extra field are reserved and cannam@89: are currently set to zero. cannam@89: cannam@89: Main procedure: cannam@89: - Exclusively create the foo.lock file using the O_CREAT and O_EXCL modes of cannam@89: the system open() call. If the modify time of an existing lock file is cannam@89: more than PATIENCE seconds old, then the lock file is deleted and the cannam@89: exclusive create is retried. cannam@89: - Load the extra field from the foo.gz file, and see if an operation was in cannam@89: progress but not completed. If so, apply the recovery procedure below. cannam@89: - Perform the append procedure with the provided data. cannam@89: - If the uncompressed data in the foo.gz file is 1MB or more, apply the cannam@89: compress procedure. cannam@89: - Delete the foo.lock file. cannam@89: cannam@89: Append procedure: cannam@89: - Put what to append in the foo.add file so that the operation can be cannam@89: restarted if this procedure is interrupted. cannam@89: - Mark the foo.gz extra field with the append operation in progress. cannam@89: + Restore the original last-block bit and stored block length of the last cannam@89: stored block from the information in the extra field, in case a previous cannam@89: append operation was interrupted. cannam@89: - Append the provided data to the last stored block, creating new stored cannam@89: blocks as needed and updating the stored blocks last-block bits and cannam@89: lengths. cannam@89: - Update the crc and length with the new data, and write the gzip trailer. cannam@89: - Write over the extra field (with a single write operation) with the new cannam@89: pointers, lengths, and crc's, and mark the gzip file as not in process. cannam@89: Though there is still a foo.add file, it will be ignored since nothing cannam@89: is in process. If a foo.add file is leftover from a previously cannam@89: completed operation, it is truncated when writing new data to it. cannam@89: - Delete the foo.add file. cannam@89: cannam@89: Compress and replace procedures: cannam@89: - Read all of the uncompressed data in the stored blocks in foo.gz and write cannam@89: it to foo.add. Also write foo.temp with the last 32K of that data to cannam@89: provide a dictionary for the next invocation of this procedure. cannam@89: - Rewrite the extra field marking foo.gz with a compression in process. cannam@89: * If there is no data provided to compress (due to a missing foo.add file cannam@89: when recovering), reconstruct and truncate the foo.gz file to contain cannam@89: only the previous compressed data and proceed to the step after the next cannam@89: one. Otherwise ... cannam@89: - Compress the data with the dictionary in foo.dict, and write to the cannam@89: foo.gz file starting at the bit immediately following the last previously cannam@89: compressed block. If there is no foo.dict, proceed anyway with the cannam@89: compression at slightly reduced efficiency. (For the foo.dict file to be cannam@89: missing requires some external failure beyond simply the interruption of cannam@89: a compress operation.) During this process, the foo.lock file is cannam@89: periodically touched to assure that that file is not considered stale by cannam@89: another process before we're done. The deflation is terminated with a cannam@89: non-last empty static block (10 bits long), that is then located and cannam@89: written over by a last-bit-set empty stored block. cannam@89: - Append the crc and length of the data in the gzip file (previously cannam@89: calculated during the append operations). cannam@89: - Write over the extra field with the updated stored block offsets, bits cannam@89: back, crc's, and lengths, and mark foo.gz as in process for a replacement cannam@89: of the dictionary. cannam@89: @ Delete the foo.add file. cannam@89: - Replace foo.dict with foo.temp. cannam@89: - Write over the extra field, marking foo.gz as complete. cannam@89: cannam@89: Recovery procedure: cannam@89: - If not a replace recovery, read in the foo.add file, and provide that data cannam@89: to the appropriate recovery below. If there is no foo.add file, provide cannam@89: a zero data length to the recovery. In that case, the append recovery cannam@89: restores the foo.gz to the previous compressed + uncompressed data state. cannam@89: For the the compress recovery, a missing foo.add file results in foo.gz cannam@89: being restored to the previous compressed-only data state. cannam@89: - Append recovery: cannam@89: - Pick up append at + step above cannam@89: - Compress recovery: cannam@89: - Pick up compress at * step above cannam@89: - Replace recovery: cannam@89: - Pick up compress at @ step above cannam@89: - Log the repair with a date stamp in foo.repairs cannam@89: */ cannam@89: cannam@89: #include cannam@89: #include /* rename, fopen, fprintf, fclose */ cannam@89: #include /* malloc, free */ cannam@89: #include /* strlen, strrchr, strcpy, strncpy, strcmp */ cannam@89: #include /* open */ cannam@89: #include /* lseek, read, write, close, unlink, sleep, */ cannam@89: /* ftruncate, fsync */ cannam@89: #include /* errno */ cannam@89: #include /* time, ctime */ cannam@89: #include /* stat */ cannam@89: #include /* utimes */ cannam@89: #include "zlib.h" /* crc32 */ cannam@89: cannam@89: #include "gzlog.h" /* header for external access */ cannam@89: cannam@89: #define local static cannam@89: typedef unsigned int uint; cannam@89: typedef unsigned long ulong; cannam@89: cannam@89: /* Macro for debugging to deterministically force recovery operations */ cannam@89: #ifdef DEBUG cannam@89: #include /* longjmp */ cannam@89: jmp_buf gzlog_jump; /* where to go back to */ cannam@89: int gzlog_bail = 0; /* which point to bail at (1..8) */ cannam@89: int gzlog_count = -1; /* number of times through to wait */ cannam@89: # define BAIL(n) do { if (n == gzlog_bail && gzlog_count-- == 0) \ cannam@89: longjmp(gzlog_jump, gzlog_bail); } while (0) cannam@89: #else cannam@89: # define BAIL(n) cannam@89: #endif cannam@89: cannam@89: /* how old the lock file can be in seconds before considering it stale */ cannam@89: #define PATIENCE 300 cannam@89: cannam@89: /* maximum stored block size in Kbytes -- must be in 1..63 */ cannam@89: #define MAX_STORE 16 cannam@89: cannam@89: /* number of stored Kbytes to trigger compression (must be >= 32 to allow cannam@89: dictionary construction, and <= 204 * MAX_STORE, in order for >> 10 to cannam@89: discard the stored block headers contribution of five bytes each) */ cannam@89: #define TRIGGER 1024 cannam@89: cannam@89: /* size of a deflate dictionary (this cannot be changed) */ cannam@89: #define DICT 32768U cannam@89: cannam@89: /* values for the operation (2 bits) */ cannam@89: #define NO_OP 0 cannam@89: #define APPEND_OP 1 cannam@89: #define COMPRESS_OP 2 cannam@89: #define REPLACE_OP 3 cannam@89: cannam@89: /* macros to extract little-endian integers from an unsigned byte buffer */ cannam@89: #define PULL2(p) ((p)[0]+((uint)((p)[1])<<8)) cannam@89: #define PULL4(p) (PULL2(p)+((ulong)PULL2(p+2)<<16)) cannam@89: #define PULL8(p) (PULL4(p)+((off_t)PULL4(p+4)<<32)) cannam@89: cannam@89: /* macros to store integers into a byte buffer in little-endian order */ cannam@89: #define PUT2(p,a) do {(p)[0]=a;(p)[1]=(a)>>8;} while(0) cannam@89: #define PUT4(p,a) do {PUT2(p,a);PUT2(p+2,a>>16);} while(0) cannam@89: #define PUT8(p,a) do {PUT4(p,a);PUT4(p+4,a>>32);} while(0) cannam@89: cannam@89: /* internal structure for log information */ cannam@89: #define LOGID "\106\035\172" /* should be three non-zero characters */ cannam@89: struct log { cannam@89: char id[4]; /* contains LOGID to detect inadvertent overwrites */ cannam@89: int fd; /* file descriptor for .gz file, opened read/write */ cannam@89: char *path; /* allocated path, e.g. "/var/log/foo" or "foo" */ cannam@89: char *end; /* end of path, for appending suffices such as ".gz" */ cannam@89: off_t first; /* offset of first stored block first length byte */ cannam@89: int back; /* location of first block id in bits back from first */ cannam@89: uint stored; /* bytes currently in last stored block */ cannam@89: off_t last; /* offset of last stored block first length byte */ cannam@89: ulong ccrc; /* crc of compressed data */ cannam@89: ulong clen; /* length (modulo 2^32) of compressed data */ cannam@89: ulong tcrc; /* crc of total data */ cannam@89: ulong tlen; /* length (modulo 2^32) of total data */ cannam@89: time_t lock; /* last modify time of our lock file */ cannam@89: }; cannam@89: cannam@89: /* gzip header for gzlog */ cannam@89: local unsigned char log_gzhead[] = { cannam@89: 0x1f, 0x8b, /* magic gzip id */ cannam@89: 8, /* compression method is deflate */ cannam@89: 4, /* there is an extra field (no file name) */ cannam@89: 0, 0, 0, 0, /* no modification time provided */ cannam@89: 0, 0xff, /* no extra flags, no OS specified */ cannam@89: 39, 0, 'a', 'p', 35, 0 /* extra field with "ap" subfield */ cannam@89: /* 35 is EXTRA, 39 is EXTRA + 4 */ cannam@89: }; cannam@89: cannam@89: #define HEAD sizeof(log_gzhead) /* should be 16 */ cannam@89: cannam@89: /* initial gzip extra field content (52 == HEAD + EXTRA + 1) */ cannam@89: local unsigned char log_gzext[] = { cannam@89: 52, 0, 0, 0, 0, 0, 0, 0, /* offset of first stored block length */ cannam@89: 52, 0, 0, 0, 0, 0, 0, 0, /* offset of last stored block length */ cannam@89: 0, 0, 0, 0, 0, 0, 0, 0, /* compressed data crc and length */ cannam@89: 0, 0, 0, 0, 0, 0, 0, 0, /* total data crc and length */ cannam@89: 0, 0, /* final stored block data length */ cannam@89: 5 /* op is NO_OP, last bit 8 bits back */ cannam@89: }; cannam@89: cannam@89: #define EXTRA sizeof(log_gzext) /* should be 35 */ cannam@89: cannam@89: /* initial gzip data and trailer */ cannam@89: local unsigned char log_gzbody[] = { cannam@89: 1, 0, 0, 0xff, 0xff, /* empty stored block (last) */ cannam@89: 0, 0, 0, 0, /* crc */ cannam@89: 0, 0, 0, 0 /* uncompressed length */ cannam@89: }; cannam@89: cannam@89: #define BODY sizeof(log_gzbody) cannam@89: cannam@89: /* Exclusively create foo.lock in order to negotiate exclusive access to the cannam@89: foo.* files. If the modify time of an existing lock file is greater than cannam@89: PATIENCE seconds in the past, then consider the lock file to have been cannam@89: abandoned, delete it, and try the exclusive create again. Save the lock cannam@89: file modify time for verification of ownership. Return 0 on success, or -1 cannam@89: on failure, usually due to an access restriction or invalid path. Note that cannam@89: if stat() or unlink() fails, it may be due to another process noticing the cannam@89: abandoned lock file a smidge sooner and deleting it, so those are not cannam@89: flagged as an error. */ cannam@89: local int log_lock(struct log *log) cannam@89: { cannam@89: int fd; cannam@89: struct stat st; cannam@89: cannam@89: strcpy(log->end, ".lock"); cannam@89: while ((fd = open(log->path, O_CREAT | O_EXCL, 0644)) < 0) { cannam@89: if (errno != EEXIST) cannam@89: return -1; cannam@89: if (stat(log->path, &st) == 0 && time(NULL) - st.st_mtime > PATIENCE) { cannam@89: unlink(log->path); cannam@89: continue; cannam@89: } cannam@89: sleep(2); /* relinquish the CPU for two seconds while waiting */ cannam@89: } cannam@89: close(fd); cannam@89: if (stat(log->path, &st) == 0) cannam@89: log->lock = st.st_mtime; cannam@89: return 0; cannam@89: } cannam@89: cannam@89: /* Update the modify time of the lock file to now, in order to prevent another cannam@89: task from thinking that the lock is stale. Save the lock file modify time cannam@89: for verification of ownership. */ cannam@89: local void log_touch(struct log *log) cannam@89: { cannam@89: struct stat st; cannam@89: cannam@89: strcpy(log->end, ".lock"); cannam@89: utimes(log->path, NULL); cannam@89: if (stat(log->path, &st) == 0) cannam@89: log->lock = st.st_mtime; cannam@89: } cannam@89: cannam@89: /* Check the log file modify time against what is expected. Return true if cannam@89: this is not our lock. If it is our lock, touch it to keep it. */ cannam@89: local int log_check(struct log *log) cannam@89: { cannam@89: struct stat st; cannam@89: cannam@89: strcpy(log->end, ".lock"); cannam@89: if (stat(log->path, &st) || st.st_mtime != log->lock) cannam@89: return 1; cannam@89: log_touch(log); cannam@89: return 0; cannam@89: } cannam@89: cannam@89: /* Unlock a previously acquired lock, but only if it's ours. */ cannam@89: local void log_unlock(struct log *log) cannam@89: { cannam@89: if (log_check(log)) cannam@89: return; cannam@89: strcpy(log->end, ".lock"); cannam@89: unlink(log->path); cannam@89: log->lock = 0; cannam@89: } cannam@89: cannam@89: /* Check the gzip header and read in the extra field, filling in the values in cannam@89: the log structure. Return op on success or -1 if the gzip header was not as cannam@89: expected. op is the current operation in progress last written to the extra cannam@89: field. This assumes that the gzip file has already been opened, with the cannam@89: file descriptor log->fd. */ cannam@89: local int log_head(struct log *log) cannam@89: { cannam@89: int op; cannam@89: unsigned char buf[HEAD + EXTRA]; cannam@89: cannam@89: if (lseek(log->fd, 0, SEEK_SET) < 0 || cannam@89: read(log->fd, buf, HEAD + EXTRA) != HEAD + EXTRA || cannam@89: memcmp(buf, log_gzhead, HEAD)) { cannam@89: return -1; cannam@89: } cannam@89: log->first = PULL8(buf + HEAD); cannam@89: log->last = PULL8(buf + HEAD + 8); cannam@89: log->ccrc = PULL4(buf + HEAD + 16); cannam@89: log->clen = PULL4(buf + HEAD + 20); cannam@89: log->tcrc = PULL4(buf + HEAD + 24); cannam@89: log->tlen = PULL4(buf + HEAD + 28); cannam@89: log->stored = PULL2(buf + HEAD + 32); cannam@89: log->back = 3 + (buf[HEAD + 34] & 7); cannam@89: op = (buf[HEAD + 34] >> 3) & 3; cannam@89: return op; cannam@89: } cannam@89: cannam@89: /* Write over the extra field contents, marking the operation as op. Use fsync cannam@89: to assure that the device is written to, and in the requested order. This cannam@89: operation, and only this operation, is assumed to be atomic in order to cannam@89: assure that the log is recoverable in the event of an interruption at any cannam@89: point in the process. Return -1 if the write to foo.gz failed. */ cannam@89: local int log_mark(struct log *log, int op) cannam@89: { cannam@89: int ret; cannam@89: unsigned char ext[EXTRA]; cannam@89: cannam@89: PUT8(ext, log->first); cannam@89: PUT8(ext + 8, log->last); cannam@89: PUT4(ext + 16, log->ccrc); cannam@89: PUT4(ext + 20, log->clen); cannam@89: PUT4(ext + 24, log->tcrc); cannam@89: PUT4(ext + 28, log->tlen); cannam@89: PUT2(ext + 32, log->stored); cannam@89: ext[34] = log->back - 3 + (op << 3); cannam@89: fsync(log->fd); cannam@89: ret = lseek(log->fd, HEAD, SEEK_SET) < 0 || cannam@89: write(log->fd, ext, EXTRA) != EXTRA ? -1 : 0; cannam@89: fsync(log->fd); cannam@89: return ret; cannam@89: } cannam@89: cannam@89: /* Rewrite the last block header bits and subsequent zero bits to get to a byte cannam@89: boundary, setting the last block bit if last is true, and then write the cannam@89: remainder of the stored block header (length and one's complement). Leave cannam@89: the file pointer after the end of the last stored block data. Return -1 if cannam@89: there is a read or write failure on the foo.gz file */ cannam@89: local int log_last(struct log *log, int last) cannam@89: { cannam@89: int back, len, mask; cannam@89: unsigned char buf[6]; cannam@89: cannam@89: /* determine the locations of the bytes and bits to modify */ cannam@89: back = log->last == log->first ? log->back : 8; cannam@89: len = back > 8 ? 2 : 1; /* bytes back from log->last */ cannam@89: mask = 0x80 >> ((back - 1) & 7); /* mask for block last-bit */ cannam@89: cannam@89: /* get the byte to modify (one or two back) into buf[0] -- don't need to cannam@89: read the byte if the last-bit is eight bits back, since in that case cannam@89: the entire byte will be modified */ cannam@89: buf[0] = 0; cannam@89: if (back != 8 && (lseek(log->fd, log->last - len, SEEK_SET) < 0 || cannam@89: read(log->fd, buf, 1) != 1)) cannam@89: return -1; cannam@89: cannam@89: /* change the last-bit of the last stored block as requested -- note cannam@89: that all bits above the last-bit are set to zero, per the type bits cannam@89: of a stored block being 00 and per the convention that the bits to cannam@89: bring the stream to a byte boundary are also zeros */ cannam@89: buf[1] = 0; cannam@89: buf[2 - len] = (*buf & (mask - 1)) + (last ? mask : 0); cannam@89: cannam@89: /* write the modified stored block header and lengths, move the file cannam@89: pointer to after the last stored block data */ cannam@89: PUT2(buf + 2, log->stored); cannam@89: PUT2(buf + 4, log->stored ^ 0xffff); cannam@89: return lseek(log->fd, log->last - len, SEEK_SET) < 0 || cannam@89: write(log->fd, buf + 2 - len, len + 4) != len + 4 || cannam@89: lseek(log->fd, log->stored, SEEK_CUR) < 0 ? -1 : 0; cannam@89: } cannam@89: cannam@89: /* Append len bytes from data to the locked and open log file. len may be zero cannam@89: if recovering and no .add file was found. In that case, the previous state cannam@89: of the foo.gz file is restored. The data is appended uncompressed in cannam@89: deflate stored blocks. Return -1 if there was an error reading or writing cannam@89: the foo.gz file. */ cannam@89: local int log_append(struct log *log, unsigned char *data, size_t len) cannam@89: { cannam@89: uint put; cannam@89: off_t end; cannam@89: unsigned char buf[8]; cannam@89: cannam@89: /* set the last block last-bit and length, in case recovering an cannam@89: interrupted append, then position the file pointer to append to the cannam@89: block */ cannam@89: if (log_last(log, 1)) cannam@89: return -1; cannam@89: cannam@89: /* append, adding stored blocks and updating the offset of the last stored cannam@89: block as needed, and update the total crc and length */ cannam@89: while (len) { cannam@89: /* append as much as we can to the last block */ cannam@89: put = (MAX_STORE << 10) - log->stored; cannam@89: if (put > len) cannam@89: put = (uint)len; cannam@89: if (put) { cannam@89: if (write(log->fd, data, put) != put) cannam@89: return -1; cannam@89: BAIL(1); cannam@89: log->tcrc = crc32(log->tcrc, data, put); cannam@89: log->tlen += put; cannam@89: log->stored += put; cannam@89: data += put; cannam@89: len -= put; cannam@89: } cannam@89: cannam@89: /* if we need to, add a new empty stored block */ cannam@89: if (len) { cannam@89: /* mark current block as not last */ cannam@89: if (log_last(log, 0)) cannam@89: return -1; cannam@89: cannam@89: /* point to new, empty stored block */ cannam@89: log->last += 4 + log->stored + 1; cannam@89: log->stored = 0; cannam@89: } cannam@89: cannam@89: /* mark last block as last, update its length */ cannam@89: if (log_last(log, 1)) cannam@89: return -1; cannam@89: BAIL(2); cannam@89: } cannam@89: cannam@89: /* write the new crc and length trailer, and truncate just in case (could cannam@89: be recovering from partial append with a missing foo.add file) */ cannam@89: PUT4(buf, log->tcrc); cannam@89: PUT4(buf + 4, log->tlen); cannam@89: if (write(log->fd, buf, 8) != 8 || cannam@89: (end = lseek(log->fd, 0, SEEK_CUR)) < 0 || ftruncate(log->fd, end)) cannam@89: return -1; cannam@89: cannam@89: /* write the extra field, marking the log file as done, delete .add file */ cannam@89: if (log_mark(log, NO_OP)) cannam@89: return -1; cannam@89: strcpy(log->end, ".add"); cannam@89: unlink(log->path); /* ignore error, since may not exist */ cannam@89: return 0; cannam@89: } cannam@89: cannam@89: /* Replace the foo.dict file with the foo.temp file. Also delete the foo.add cannam@89: file, since the compress operation may have been interrupted before that was cannam@89: done. Returns 1 if memory could not be allocated, or -1 if reading or cannam@89: writing foo.gz fails, or if the rename fails for some reason other than cannam@89: foo.temp not existing. foo.temp not existing is a permitted error, since cannam@89: the replace operation may have been interrupted after the rename is done, cannam@89: but before foo.gz is marked as complete. */ cannam@89: local int log_replace(struct log *log) cannam@89: { cannam@89: int ret; cannam@89: char *dest; cannam@89: cannam@89: /* delete foo.add file */ cannam@89: strcpy(log->end, ".add"); cannam@89: unlink(log->path); /* ignore error, since may not exist */ cannam@89: BAIL(3); cannam@89: cannam@89: /* rename foo.name to foo.dict, replacing foo.dict if it exists */ cannam@89: strcpy(log->end, ".dict"); cannam@89: dest = malloc(strlen(log->path) + 1); cannam@89: if (dest == NULL) cannam@89: return -2; cannam@89: strcpy(dest, log->path); cannam@89: strcpy(log->end, ".temp"); cannam@89: ret = rename(log->path, dest); cannam@89: free(dest); cannam@89: if (ret && errno != ENOENT) cannam@89: return -1; cannam@89: BAIL(4); cannam@89: cannam@89: /* mark the foo.gz file as done */ cannam@89: return log_mark(log, NO_OP); cannam@89: } cannam@89: cannam@89: /* Compress the len bytes at data and append the compressed data to the cannam@89: foo.gz deflate data immediately after the previous compressed data. This cannam@89: overwrites the previous uncompressed data, which was stored in foo.add cannam@89: and is the data provided in data[0..len-1]. If this operation is cannam@89: interrupted, it picks up at the start of this routine, with the foo.add cannam@89: file read in again. If there is no data to compress (len == 0), then we cannam@89: simply terminate the foo.gz file after the previously compressed data, cannam@89: appending a final empty stored block and the gzip trailer. Return -1 if cannam@89: reading or writing the log.gz file failed, or -2 if there was a memory cannam@89: allocation failure. */ cannam@89: local int log_compress(struct log *log, unsigned char *data, size_t len) cannam@89: { cannam@89: int fd; cannam@89: uint got, max; cannam@89: ssize_t dict; cannam@89: off_t end; cannam@89: z_stream strm; cannam@89: unsigned char buf[DICT]; cannam@89: cannam@89: /* compress and append compressed data */ cannam@89: if (len) { cannam@89: /* set up for deflate, allocating memory */ cannam@89: strm.zalloc = Z_NULL; cannam@89: strm.zfree = Z_NULL; cannam@89: strm.opaque = Z_NULL; cannam@89: if (deflateInit2(&strm, Z_DEFAULT_COMPRESSION, Z_DEFLATED, -15, 8, cannam@89: Z_DEFAULT_STRATEGY) != Z_OK) cannam@89: return -2; cannam@89: cannam@89: /* read in dictionary (last 32K of data that was compressed) */ cannam@89: strcpy(log->end, ".dict"); cannam@89: fd = open(log->path, O_RDONLY, 0); cannam@89: if (fd >= 0) { cannam@89: dict = read(fd, buf, DICT); cannam@89: close(fd); cannam@89: if (dict < 0) { cannam@89: deflateEnd(&strm); cannam@89: return -1; cannam@89: } cannam@89: if (dict) cannam@89: deflateSetDictionary(&strm, buf, (uint)dict); cannam@89: } cannam@89: log_touch(log); cannam@89: cannam@89: /* prime deflate with last bits of previous block, position write cannam@89: pointer to write those bits and overwrite what follows */ cannam@89: if (lseek(log->fd, log->first - (log->back > 8 ? 2 : 1), cannam@89: SEEK_SET) < 0 || cannam@89: read(log->fd, buf, 1) != 1 || lseek(log->fd, -1, SEEK_CUR) < 0) { cannam@89: deflateEnd(&strm); cannam@89: return -1; cannam@89: } cannam@89: deflatePrime(&strm, (8 - log->back) & 7, *buf); cannam@89: cannam@89: /* compress, finishing with a partial non-last empty static block */ cannam@89: strm.next_in = data; cannam@89: max = (((uint)0 - 1) >> 1) + 1; /* in case int smaller than size_t */ cannam@89: do { cannam@89: strm.avail_in = len > max ? max : (uint)len; cannam@89: len -= strm.avail_in; cannam@89: do { cannam@89: strm.avail_out = DICT; cannam@89: strm.next_out = buf; cannam@89: deflate(&strm, len ? Z_NO_FLUSH : Z_PARTIAL_FLUSH); cannam@89: got = DICT - strm.avail_out; cannam@89: if (got && write(log->fd, buf, got) != got) { cannam@89: deflateEnd(&strm); cannam@89: return -1; cannam@89: } cannam@89: log_touch(log); cannam@89: } while (strm.avail_out == 0); cannam@89: } while (len); cannam@89: deflateEnd(&strm); cannam@89: BAIL(5); cannam@89: cannam@89: /* find start of empty static block -- scanning backwards the first one cannam@89: bit is the second bit of the block, if the last byte is zero, then cannam@89: we know the byte before that has a one in the top bit, since an cannam@89: empty static block is ten bits long */ cannam@89: if ((log->first = lseek(log->fd, -1, SEEK_CUR)) < 0 || cannam@89: read(log->fd, buf, 1) != 1) cannam@89: return -1; cannam@89: log->first++; cannam@89: if (*buf) { cannam@89: log->back = 1; cannam@89: while ((*buf & ((uint)1 << (8 - log->back++))) == 0) cannam@89: ; /* guaranteed to terminate, since *buf != 0 */ cannam@89: } cannam@89: else cannam@89: log->back = 10; cannam@89: cannam@89: /* update compressed crc and length */ cannam@89: log->ccrc = log->tcrc; cannam@89: log->clen = log->tlen; cannam@89: } cannam@89: else { cannam@89: /* no data to compress -- fix up existing gzip stream */ cannam@89: log->tcrc = log->ccrc; cannam@89: log->tlen = log->clen; cannam@89: } cannam@89: cannam@89: /* complete and truncate gzip stream */ cannam@89: log->last = log->first; cannam@89: log->stored = 0; cannam@89: PUT4(buf, log->tcrc); cannam@89: PUT4(buf + 4, log->tlen); cannam@89: if (log_last(log, 1) || write(log->fd, buf, 8) != 8 || cannam@89: (end = lseek(log->fd, 0, SEEK_CUR)) < 0 || ftruncate(log->fd, end)) cannam@89: return -1; cannam@89: BAIL(6); cannam@89: cannam@89: /* mark as being in the replace operation */ cannam@89: if (log_mark(log, REPLACE_OP)) cannam@89: return -1; cannam@89: cannam@89: /* execute the replace operation and mark the file as done */ cannam@89: return log_replace(log); cannam@89: } cannam@89: cannam@89: /* log a repair record to the .repairs file */ cannam@89: local void log_log(struct log *log, int op, char *record) cannam@89: { cannam@89: time_t now; cannam@89: FILE *rec; cannam@89: cannam@89: now = time(NULL); cannam@89: strcpy(log->end, ".repairs"); cannam@89: rec = fopen(log->path, "a"); cannam@89: if (rec == NULL) cannam@89: return; cannam@89: fprintf(rec, "%.24s %s recovery: %s\n", ctime(&now), op == APPEND_OP ? cannam@89: "append" : (op == COMPRESS_OP ? "compress" : "replace"), record); cannam@89: fclose(rec); cannam@89: return; cannam@89: } cannam@89: cannam@89: /* Recover the interrupted operation op. First read foo.add for recovering an cannam@89: append or compress operation. Return -1 if there was an error reading or cannam@89: writing foo.gz or reading an existing foo.add, or -2 if there was a memory cannam@89: allocation failure. */ cannam@89: local int log_recover(struct log *log, int op) cannam@89: { cannam@89: int fd, ret = 0; cannam@89: unsigned char *data = NULL; cannam@89: size_t len = 0; cannam@89: struct stat st; cannam@89: cannam@89: /* log recovery */ cannam@89: log_log(log, op, "start"); cannam@89: cannam@89: /* load foo.add file if expected and present */ cannam@89: if (op == APPEND_OP || op == COMPRESS_OP) { cannam@89: strcpy(log->end, ".add"); cannam@89: if (stat(log->path, &st) == 0 && st.st_size) { cannam@89: len = (size_t)(st.st_size); cannam@89: if (len != st.st_size || (data = malloc(st.st_size)) == NULL) { cannam@89: log_log(log, op, "allocation failure"); cannam@89: return -2; cannam@89: } cannam@89: if ((fd = open(log->path, O_RDONLY, 0)) < 0) { cannam@89: log_log(log, op, ".add file read failure"); cannam@89: return -1; cannam@89: } cannam@89: ret = read(fd, data, len) != len; cannam@89: close(fd); cannam@89: if (ret) { cannam@89: log_log(log, op, ".add file read failure"); cannam@89: return -1; cannam@89: } cannam@89: log_log(log, op, "loaded .add file"); cannam@89: } cannam@89: else cannam@89: log_log(log, op, "missing .add file!"); cannam@89: } cannam@89: cannam@89: /* recover the interrupted operation */ cannam@89: switch (op) { cannam@89: case APPEND_OP: cannam@89: ret = log_append(log, data, len); cannam@89: break; cannam@89: case COMPRESS_OP: cannam@89: ret = log_compress(log, data, len); cannam@89: break; cannam@89: case REPLACE_OP: cannam@89: ret = log_replace(log); cannam@89: } cannam@89: cannam@89: /* log status */ cannam@89: log_log(log, op, ret ? "failure" : "complete"); cannam@89: cannam@89: /* clean up */ cannam@89: if (data != NULL) cannam@89: free(data); cannam@89: return ret; cannam@89: } cannam@89: cannam@89: /* Close the foo.gz file (if open) and release the lock. */ cannam@89: local void log_close(struct log *log) cannam@89: { cannam@89: if (log->fd >= 0) cannam@89: close(log->fd); cannam@89: log->fd = -1; cannam@89: log_unlock(log); cannam@89: } cannam@89: cannam@89: /* Open foo.gz, verify the header, and load the extra field contents, after cannam@89: first creating the foo.lock file to gain exclusive access to the foo.* cannam@89: files. If foo.gz does not exist or is empty, then write the initial header, cannam@89: extra, and body content of an empty foo.gz log file. If there is an error cannam@89: creating the lock file due to access restrictions, or an error reading or cannam@89: writing the foo.gz file, or if the foo.gz file is not a proper log file for cannam@89: this object (e.g. not a gzip file or does not contain the expected extra cannam@89: field), then return true. If there is an error, the lock is released. cannam@89: Otherwise, the lock is left in place. */ cannam@89: local int log_open(struct log *log) cannam@89: { cannam@89: int op; cannam@89: cannam@89: /* release open file resource if left over -- can occur if lock lost cannam@89: between gzlog_open() and gzlog_write() */ cannam@89: if (log->fd >= 0) cannam@89: close(log->fd); cannam@89: log->fd = -1; cannam@89: cannam@89: /* negotiate exclusive access */ cannam@89: if (log_lock(log) < 0) cannam@89: return -1; cannam@89: cannam@89: /* open the log file, foo.gz */ cannam@89: strcpy(log->end, ".gz"); cannam@89: log->fd = open(log->path, O_RDWR | O_CREAT, 0644); cannam@89: if (log->fd < 0) { cannam@89: log_close(log); cannam@89: return -1; cannam@89: } cannam@89: cannam@89: /* if new, initialize foo.gz with an empty log, delete old dictionary */ cannam@89: if (lseek(log->fd, 0, SEEK_END) == 0) { cannam@89: if (write(log->fd, log_gzhead, HEAD) != HEAD || cannam@89: write(log->fd, log_gzext, EXTRA) != EXTRA || cannam@89: write(log->fd, log_gzbody, BODY) != BODY) { cannam@89: log_close(log); cannam@89: return -1; cannam@89: } cannam@89: strcpy(log->end, ".dict"); cannam@89: unlink(log->path); cannam@89: } cannam@89: cannam@89: /* verify log file and load extra field information */ cannam@89: if ((op = log_head(log)) < 0) { cannam@89: log_close(log); cannam@89: return -1; cannam@89: } cannam@89: cannam@89: /* check for interrupted process and if so, recover */ cannam@89: if (op != NO_OP && log_recover(log, op)) { cannam@89: log_close(log); cannam@89: return -1; cannam@89: } cannam@89: cannam@89: /* touch the lock file to prevent another process from grabbing it */ cannam@89: log_touch(log); cannam@89: return 0; cannam@89: } cannam@89: cannam@89: /* See gzlog.h for the description of the external methods below */ cannam@89: gzlog *gzlog_open(char *path) cannam@89: { cannam@89: size_t n; cannam@89: struct log *log; cannam@89: cannam@89: /* check arguments */ cannam@89: if (path == NULL || *path == 0) cannam@89: return NULL; cannam@89: cannam@89: /* allocate and initialize log structure */ cannam@89: log = malloc(sizeof(struct log)); cannam@89: if (log == NULL) cannam@89: return NULL; cannam@89: strcpy(log->id, LOGID); cannam@89: log->fd = -1; cannam@89: cannam@89: /* save path and end of path for name construction */ cannam@89: n = strlen(path); cannam@89: log->path = malloc(n + 9); /* allow for ".repairs" */ cannam@89: if (log->path == NULL) { cannam@89: free(log); cannam@89: return NULL; cannam@89: } cannam@89: strcpy(log->path, path); cannam@89: log->end = log->path + n; cannam@89: cannam@89: /* gain exclusive access and verify log file -- may perform a cannam@89: recovery operation if needed */ cannam@89: if (log_open(log)) { cannam@89: free(log->path); cannam@89: free(log); cannam@89: return NULL; cannam@89: } cannam@89: cannam@89: /* return pointer to log structure */ cannam@89: return log; cannam@89: } cannam@89: cannam@89: /* gzlog_compress() return values: cannam@89: 0: all good cannam@89: -1: file i/o error (usually access issue) cannam@89: -2: memory allocation failure cannam@89: -3: invalid log pointer argument */ cannam@89: int gzlog_compress(gzlog *logd) cannam@89: { cannam@89: int fd, ret; cannam@89: uint block; cannam@89: size_t len, next; cannam@89: unsigned char *data, buf[5]; cannam@89: struct log *log = logd; cannam@89: cannam@89: /* check arguments */ cannam@89: if (log == NULL || strcmp(log->id, LOGID) || len < 0) cannam@89: return -3; cannam@89: cannam@89: /* see if we lost the lock -- if so get it again and reload the extra cannam@89: field information (it probably changed), recover last operation if cannam@89: necessary */ cannam@89: if (log_check(log) && log_open(log)) cannam@89: return -1; cannam@89: cannam@89: /* create space for uncompressed data */ cannam@89: len = ((size_t)(log->last - log->first) & ~(((size_t)1 << 10) - 1)) + cannam@89: log->stored; cannam@89: if ((data = malloc(len)) == NULL) cannam@89: return -2; cannam@89: cannam@89: /* do statement here is just a cheap trick for error handling */ cannam@89: do { cannam@89: /* read in the uncompressed data */ cannam@89: if (lseek(log->fd, log->first - 1, SEEK_SET) < 0) cannam@89: break; cannam@89: next = 0; cannam@89: while (next < len) { cannam@89: if (read(log->fd, buf, 5) != 5) cannam@89: break; cannam@89: block = PULL2(buf + 1); cannam@89: if (next + block > len || cannam@89: read(log->fd, (char *)data + next, block) != block) cannam@89: break; cannam@89: next += block; cannam@89: } cannam@89: if (lseek(log->fd, 0, SEEK_CUR) != log->last + 4 + log->stored) cannam@89: break; cannam@89: log_touch(log); cannam@89: cannam@89: /* write the uncompressed data to the .add file */ cannam@89: strcpy(log->end, ".add"); cannam@89: fd = open(log->path, O_WRONLY | O_CREAT | O_TRUNC, 0644); cannam@89: if (fd < 0) cannam@89: break; cannam@89: ret = write(fd, data, len) != len; cannam@89: if (ret | close(fd)) cannam@89: break; cannam@89: log_touch(log); cannam@89: cannam@89: /* write the dictionary for the next compress to the .temp file */ cannam@89: strcpy(log->end, ".temp"); cannam@89: fd = open(log->path, O_WRONLY | O_CREAT | O_TRUNC, 0644); cannam@89: if (fd < 0) cannam@89: break; cannam@89: next = DICT > len ? len : DICT; cannam@89: ret = write(fd, (char *)data + len - next, next) != next; cannam@89: if (ret | close(fd)) cannam@89: break; cannam@89: log_touch(log); cannam@89: cannam@89: /* roll back to compressed data, mark the compress in progress */ cannam@89: log->last = log->first; cannam@89: log->stored = 0; cannam@89: if (log_mark(log, COMPRESS_OP)) cannam@89: break; cannam@89: BAIL(7); cannam@89: cannam@89: /* compress and append the data (clears mark) */ cannam@89: ret = log_compress(log, data, len); cannam@89: free(data); cannam@89: return ret; cannam@89: } while (0); cannam@89: cannam@89: /* broke out of do above on i/o error */ cannam@89: free(data); cannam@89: return -1; cannam@89: } cannam@89: cannam@89: /* gzlog_write() return values: cannam@89: 0: all good cannam@89: -1: file i/o error (usually access issue) cannam@89: -2: memory allocation failure cannam@89: -3: invalid log pointer argument */ cannam@89: int gzlog_write(gzlog *logd, void *data, size_t len) cannam@89: { cannam@89: int fd, ret; cannam@89: struct log *log = logd; cannam@89: cannam@89: /* check arguments */ cannam@89: if (log == NULL || strcmp(log->id, LOGID) || len < 0) cannam@89: return -3; cannam@89: if (data == NULL || len == 0) cannam@89: return 0; cannam@89: cannam@89: /* see if we lost the lock -- if so get it again and reload the extra cannam@89: field information (it probably changed), recover last operation if cannam@89: necessary */ cannam@89: if (log_check(log) && log_open(log)) cannam@89: return -1; cannam@89: cannam@89: /* create and write .add file */ cannam@89: strcpy(log->end, ".add"); cannam@89: fd = open(log->path, O_WRONLY | O_CREAT | O_TRUNC, 0644); cannam@89: if (fd < 0) cannam@89: return -1; cannam@89: ret = write(fd, data, len) != len; cannam@89: if (ret | close(fd)) cannam@89: return -1; cannam@89: log_touch(log); cannam@89: cannam@89: /* mark log file with append in progress */ cannam@89: if (log_mark(log, APPEND_OP)) cannam@89: return -1; cannam@89: BAIL(8); cannam@89: cannam@89: /* append data (clears mark) */ cannam@89: if (log_append(log, data, len)) cannam@89: return -1; cannam@89: cannam@89: /* check to see if it's time to compress -- if not, then done */ cannam@89: if (((log->last - log->first) >> 10) + (log->stored >> 10) < TRIGGER) cannam@89: return 0; cannam@89: cannam@89: /* time to compress */ cannam@89: return gzlog_compress(log); cannam@89: } cannam@89: cannam@89: /* gzlog_close() return values: cannam@89: 0: ok cannam@89: -3: invalid log pointer argument */ cannam@89: int gzlog_close(gzlog *logd) cannam@89: { cannam@89: struct log *log = logd; cannam@89: cannam@89: /* check arguments */ cannam@89: if (log == NULL || strcmp(log->id, LOGID)) cannam@89: return -3; cannam@89: cannam@89: /* close the log file and release the lock */ cannam@89: log_close(log); cannam@89: cannam@89: /* free structure and return */ cannam@89: if (log->path != NULL) cannam@89: free(log->path); cannam@89: strcpy(log->id, "bad"); cannam@89: free(log); cannam@89: return 0; cannam@89: }