3 * m_getfld.c -- read/parse a message
5 * This code is Copyright (c) 2002, by the authors of nmh. See the
6 * COPYRIGHT file in the root directory of the nmh distribution for
7 * complete copyright information.
14 /* This module has a long and checkered history. First, it didn't burst
15 maildrops correctly because it considered two CTRL-A:s in a row to be
16 an inter-message delimiter. It really is four CTRL-A:s followed by a
17 newline. Unfortunately, MMDF will convert this delimiter *inside* a
18 message to a CTRL-B followed by three CTRL-A:s and a newline. This
19 caused the old version of m_getfld() to declare eom prematurely. The
20 fix was a lot slower than
22 c == '\001' && peekc (iob) == '\001'
24 but it worked, and to increase generality, MBOX style maildrops could
25 be parsed as well. Unfortunately the speed issue finally caught up with
26 us since this routine is at the very heart of MH.
28 To speed things up considerably, the routine Eom() was made an auxilary
29 function called by the macro eom(). Unless we are bursting a maildrop,
30 the eom() macro returns FALSE saying we aren't at the end of the
33 The next thing to do is to read the mts.conf file and initialize
34 delimiter[] and delimlen accordingly...
36 After mhl was made a built-in in msh, m_getfld() worked just fine
37 (using m_unknown() at startup). Until one day: a message which was
38 the result of a bursting was shown. Then, since the burst boundaries
39 aren't CTRL-A:s, m_getfld() would blinding plunge on past the boundary.
40 Very sad. The solution: introduce m_eomsbr(). This hook gets called
41 after the end of each line (since testing for eom involves an fseek()).
42 This worked fine, until one day: a message with no body portion arrived.
45 while (eom (c = Getc (iob), iob))
48 loop caused m_getfld() to return FMTERR. So, that logic was changed to
49 check for (*eom_action) and act accordingly.
51 This worked fine, until one day: someone didn't use four CTRL:A's as
52 their delimiters. So, the bullet got bit and we read mts.h and
53 continue to struggle on. It's not that bad though, since the only time
54 the code gets executed is when inc (or msh) calls it, and both of these
55 have already called mts_init().
57 ------------------------
58 (Written by Van Jacobson for the mh6 m_getfld, January, 1986):
60 This routine was accounting for 60% of the cpu time used by most mh
61 programs. I spent a bit of time tuning and it now accounts for <10%
62 of the time used. Like any heavily tuned routine, it's a bit
63 complex and you want to be sure you understand everything that it's
64 doing before you start hacking on it. Let me try to emphasize
65 that: every line in this atrocity depends on every other line,
66 sometimes in subtle ways. You should understand it all, in detail,
67 before trying to change any part. If you do change it, test the
68 result thoroughly (I use a hand-constructed test file that exercises
69 all the ways a header name, header body, header continuation,
70 header-body separator, body line and body eom can align themselves
71 with respect to a buffer boundary). "Minor" bugs in this routine
72 result in garbaged or lost mail.
74 If you hack on this and slow it down, I, my children and my
75 children's children will curse you.
77 This routine gets used on three different types of files: normal,
78 single msg files, "packed" unix or mmdf mailboxs (when used by inc)
79 and packed, directoried bulletin board files (when used by msh).
80 The biggest impact of different file types is in "eom" testing. The
81 code has been carefully organized to test for eom at appropriate
82 times and at no other times (since the check is quite expensive).
83 I have tried to arrange things so that the eom check need only be
84 done on entry to this routine. Since an eom can only occur after a
85 newline, this is easy to manage for header fields. For the msg
86 body, we try to efficiently search the input buffer to see if
87 contains the eom delimiter. If it does, we take up to the
88 delimiter, otherwise we take everything in the buffer. (The change
89 to the body eom/copy processing produced the most noticeable
90 performance difference, particularly for "inc" and "show".)
92 There are three qualitatively different things this routine busts
93 out of a message: field names, field text and msg bodies. Field
94 names are typically short (~8 char) and the loop that extracts them
95 might terminate on a colon, newline or max width. I considered
96 using a Vax "scanc" to locate the end of the field followed by a
97 "bcopy" but the routine call overhead on a Vax is too large for this
98 to work on short names. If Berkeley ever makes "inline" part of the
99 C optimiser (so things like "scanc" turn into inline instructions) a
100 change here would be worthwhile.
102 Field text is typically 60 - 100 characters so there's (barely)
103 a win in doing a routine call to something that does a "locc"
104 followed by a "bmove". About 30% of the fields have continuations
105 (usually the 822 "received:" lines) and each continuation generates
106 another routine call. "Inline" would be a big win here, as well.
108 Messages, as of this writing, seem to come in two flavors: small
109 (~1K) and long (>2K). Most messages have 400 - 600 bytes of headers
110 so message bodies average at least a few hundred characters.
111 Assuming your system uses reasonably sized stdio buffers (1K or
112 more), this routine should be able to remove the body in large
113 (>500 byte) chunks. The makes the cost of a call to "bcopy"
114 small but there is a premium on checking for the eom in packed
115 maildrops. The eom pattern is always a simple string so we can
116 construct an efficient pattern matcher for it (e.g., a Vax "matchc"
117 instruction). Some thought went into recognizing the start of
118 an eom that has been split across two buffers.
120 This routine wants to deal with large chunks of data so, rather
121 than "getc" into a local buffer, it uses stdio's buffer. If
122 you try to use it on a non-buffered file, you'll get what you
123 deserve. This routine "knows" that struct FILEs have a _ptr
124 and a _cnt to describe the current state of the buffer and
125 it knows that _filbuf ignores the _ptr & _cnt and simply fills
126 the buffer. If stdio on your system doesn't work this way, you
127 may have to make small changes in this routine.
129 This routine also "knows" that an EOF indication on a stream is
130 "sticky" (i.e., you will keep getting EOF until you reposition the
131 stream). If your system doesn't work this way it is broken and you
132 should complain to the vendor. As a consequence of the sticky
133 EOF, this routine will never return any kind of EOF status when
134 there is data in "name" or "buf").
141 static int m_Eom (int, FILE *);
142 static unsigned char *matchc(int, char *, int, char *);
143 static unsigned char *locc(int, unsigned char *, unsigned char);
145 #define Getc(iob) getc(iob)
146 #define eom(c,iob) (msg_style != MS_DEFAULT && \
147 (((c) == *msg_delim && m_Eom(c,iob)) ||\
148 (eom_action && (*eom_action)(c))))
150 static unsigned char **pat_map;
153 * defined in sbr/m_msgdef.c = 0
154 * This is a disgusting hack for "inc" so it can know how many
155 * characters were stuffed in the buffer on the last call
156 * (see comments in uip/scansbr.c).
158 extern int msg_count;
161 * defined in sbr/m_msgdef.c = MS_DEFAULT
163 extern int msg_style;
166 * The "full" delimiter string for a packed maildrop consists
167 * of a newline followed by the actual delimiter. E.g., the
168 * full string for a Unix maildrop would be: "\n\nFrom ".
169 * "Fdelim" points to the start of the full string and is used
170 * in the BODY case of the main routine to search the buffer for
171 * a possible eom. Msg_delim points to the first character of
172 * the actual delim. string (i.e., fdelim+1). Edelim
173 * points to the 2nd character of actual delimiter string. It
174 * is used in m_Eom because the first character of the string
175 * has been read and matched before m_Eom is called.
177 extern char *msg_delim; /* defined in sbr/m_msgdef.c = "" */
178 static unsigned char *fdelim;
179 static unsigned char *delimend;
180 static int fdelimlen;
181 static unsigned char *edelim;
182 static int edelimlen;
184 static int (*eom_action)(int) = NULL;
187 # define _ptr _p /* Gag */
188 # define _cnt _r /* Retch */
189 # define _filbuf __srget /* Puke */
190 # define DEFINED__FILBUF_TO_SOMETHING_SPECIFIC
192 # if defined __CYGWIN__
193 /* Cygwin's stdio.h does not declare __srget(). */
195 # endif /* __CYGWIN__ */
198 #ifndef DEFINED__FILBUF_TO_SOMETHING_SPECIFIC
199 extern int _filbuf(FILE*);
204 m_getfld (int state, unsigned char *name, unsigned char *buf,
205 int bufsz, FILE *iob)
207 register unsigned char *bp, *cp, *ep, *sp;
208 register int cnt, c, i, j;
210 if ((c = Getc(iob)) < 0) {
217 /* flush null messages */
218 while ((c = Getc(iob)) >= 0 && eom (c, iob))
232 if (c == '\n' || c == '-') {
233 /* we hit the header/body separator */
234 while (c != '\n' && (c = Getc(iob)) >= 0)
237 if (c < 0 || (c = Getc(iob)) < 0 || eom (c, iob)) {
239 /* flush null messages */
240 while ((c = Getc(iob)) >= 0 && eom (c, iob))
253 * get the name of this component. take characters up
254 * to a ':', a newline or NAMESZ-1 characters, whichever
261 bp = sp = (unsigned char *) iob->_IO_read_ptr - 1;
262 j = (cnt = ((long) iob->_IO_read_end -
263 (long) iob->_IO_read_ptr) + 1) < i ? cnt : i;
264 #elif defined(__DragonFly__)
265 bp = sp = (unsigned char *) ((struct __FILE_public *)iob)->_p - 1;
266 j = (cnt = ((struct __FILE_public *)iob)->_r+1) < i ? cnt : i;
268 bp = sp = (unsigned char *) iob->_ptr - 1;
269 j = (cnt = iob->_cnt+1) < i ? cnt : i;
271 while (--j >= 0 && (c = *bp++) != ':' && c != '\n')
275 if ((cnt -= j) <= 0) {
277 iob->_IO_read_ptr = iob->_IO_read_end;
278 if (__underflow(iob) == EOF) {
279 #elif defined(__DragonFly__)
280 if (__srget(iob) == EOF) {
282 if (_filbuf(iob) == EOF) {
285 advise (NULL, "eof encountered in field \"%s\"", name);
289 iob->_IO_read_ptr++; /* NOT automatic in __underflow()! */
293 iob->_IO_read_ptr = bp + 1;
294 #elif defined(__DragonFly__)
295 ((struct __FILE_public *)iob)->_p = bp + 1;
296 ((struct __FILE_public *)iob)->_r = cnt - 1;
306 * something went wrong. possibilities are:
307 * . hit a newline (error)
308 * . got more than namesz chars. (error)
309 * . hit the end of the buffer. (loop)
312 /* We hit the end of the line without seeing ':' to
313 * terminate the field name. This is usually (always?)
314 * spam. But, blowing up is lame, especially when
315 * scan(1)ing a folder with such messages. Pretend such
316 * lines are the first of the body (at least mutt also
317 * handles it this way). */
319 /* See if buf can hold this line, since we were assuming
320 * we had a buffer of NAMESZ, not bufsz. */
321 /* + 1 for the newline */
323 /* No, it can't. Oh well, guess we'll blow up. */
325 advise (NULL, "eol encountered in field \"%s\"", name);
329 memcpy (buf, name, j - 1);
332 /* mhparse.c:get_content wants to find the position of the
333 * body start, but it thinks there's a blank line between
334 * the header and the body (naturally!), so seek back so
335 * that things line up even though we don't have that
336 * blank line in this case. Simpler parsers (e.g. mhl)
337 * get extra newlines, but that should be harmless enough,
338 * right? This is a corrupt message anyway. */
339 fseek (iob, ftell (iob) - 2, SEEK_SET);
344 advise (NULL, "field name \"%s\" exceeds %d bytes", name, NAMESZ - 2);
350 while (isspace (*--cp) && cp >= name)
357 * get (more of) the text of a field. take
358 * characters up to the end of this field (newline
359 * followed by non-blank) or bufsz-1 characters.
361 cp = buf; i = bufsz-1;
364 cnt = (long) iob->_IO_read_end - (long) iob->_IO_read_ptr;
365 bp = (unsigned char *) --iob->_IO_read_ptr;
366 #elif defined(__DragonFly__)
367 cnt = ((struct __FILE_public *)iob)->_r++;
368 bp = (unsigned char *) --((struct __FILE_public *)iob)->_p;
371 bp = (unsigned char *) --iob->_ptr;
373 c = cnt < i ? cnt : i;
374 while ((ep = locc( c, bp, '\n' ))) {
376 * if we hit the end of this field, return.
378 if ((j = *++ep) != ' ' && j != '\t') {
380 j = ep - (unsigned char *) iob->_IO_read_ptr;
381 memcpy (cp, iob->_IO_read_ptr, j);
382 iob->_IO_read_ptr = ep;
383 #elif defined(__DragonFly__)
384 j = ep - (unsigned char *) ((struct __FILE_public *)iob)->_p;
385 memcpy (cp, ((struct __FILE_public *)iob)->_p, j);
386 ((struct __FILE_public *)iob)->_p = ep;
387 ((struct __FILE_public *)iob)->_r -= j;
389 j = ep - (unsigned char *) iob->_ptr;
390 memcpy (cp, iob->_ptr, j);
402 * end of input or dest buffer - copy what we've found.
405 c += bp - (unsigned char *) iob->_IO_read_ptr;
406 memcpy( cp, iob->_IO_read_ptr, c);
407 #elif defined(__DragonFly__)
408 c += bp - (unsigned char *) ((struct __FILE_public *)iob)->_p;
409 memcpy( cp, ((struct __FILE_public *)iob)->_p, c);
411 c += bp - (unsigned char *) iob->_ptr;
412 memcpy( cp, iob->_ptr, c);
417 /* the dest buffer is full */
419 iob->_IO_read_ptr += c;
420 #elif defined(__DragonFly__)
421 ((struct __FILE_public *)iob)->_r -= c;
422 ((struct __FILE_public *)iob)->_p += c;
431 * There's one character left in the input buffer.
432 * Copy it & fill the buffer. If the last char
433 * was a newline and the next char is not whitespace,
434 * this is the end of the field. Otherwise loop.
438 *cp++ = j = *(iob->_IO_read_ptr + c);
439 iob->_IO_read_ptr = iob->_IO_read_end;
440 c = __underflow(iob);
441 iob->_IO_read_ptr++; /* NOT automatic! */
442 #elif defined(__DragonFly__)
443 *cp++ =j = *(((struct __FILE_public *)iob)->_p + c);
446 *cp++ = j = *(iob->_ptr + c);
450 ((j == '\0' || j == '\n') && c != ' ' && c != '\t')) {
454 #elif defined(__DragonFly__)
455 --((struct __FILE_public *)iob)->_p;
456 ++((struct __FILE_public *)iob)->_r;
471 * get the message body up to bufsz characters or the
472 * end of the message. Sleazy hack: if bufsz is negative
473 * we assume that we were called to copy directly into
474 * the output buffer and we don't add an eos.
476 i = (bufsz < 0) ? -bufsz : bufsz-1;
478 bp = (unsigned char *) --iob->_IO_read_ptr;
479 cnt = (long) iob->_IO_read_end - (long) iob->_IO_read_ptr;
480 #elif defined(__DragonFly__)
481 bp = (unsigned char *) --((struct __FILE_public *)iob)->_p;
482 cnt = ++((struct __FILE_public *)iob)->_r;
484 bp = (unsigned char *) --iob->_ptr;
487 c = (cnt < i ? cnt : i);
488 if (msg_style != MS_DEFAULT && c > 1) {
490 * packed maildrop - only take up to the (possible)
491 * start of the next message. This "matchc" should
492 * probably be a Boyer-Moore matcher for non-vaxen,
493 * particularly since we have the alignment table
494 * all built for the end-of-buffer test (next).
495 * But our vax timings indicate that the "matchc"
496 * instruction is 50% faster than a carefully coded
497 * B.M. matcher for most strings. (So much for elegant
498 * algorithms vs. brute force.) Since I (currently)
499 * run MH on a vax, we use the matchc instruction. --vj
501 if ((ep = matchc( fdelimlen, fdelim, c, bp )))
505 * There's no delim in the buffer but there may be
506 * a partial one at the end. If so, we want to leave
507 * it so the "eom" check on the next call picks it up.
508 * Use a modified Boyer-Moore matcher to make this
509 * check relatively cheap. The first "if" figures
510 * out what position in the pattern matches the last
511 * character in the buffer. The inner "while" matches
512 * the pattern against the buffer, backwards starting
513 * at that position. Note that unless the buffer
514 * ends with one of the characters in the pattern
515 * (excluding the first and last), we do only one test.
518 if ((sp = pat_map[*ep])) {
520 /* This if() is true unless (a) the buffer is too
521 * small to contain this delimiter prefix, or
522 * (b) it contains exactly enough chars for the
524 * For case (a) obviously we aren't going to match.
525 * For case (b), if the buffer really contained exactly
526 * a delim prefix, then the m_eom call at entry
527 * should have found it. Thus it's not a delim
528 * and we know we won't get a match.
530 if (((sp - fdelim) + 2) <= c) {
532 /* Unfortunately although fdelim has a preceding NUL
533 * we can't use this as a sentinel in case the buffer
534 * contains a NUL in exactly the wrong place (this
535 * would cause us to run off the front of fdelim).
537 while (*--ep == *--cp)
541 /* we matched the entire delim prefix,
542 * so only take the buffer up to there.
543 * we know ep >= bp -- check above prevents underrun
549 /* try matching one less char of delim string */
551 } while (--sp > fdelim);
555 memcpy( buf, bp, c );
557 iob->_IO_read_ptr += c;
558 #elif defined(__DragonFly__)
559 ((struct __FILE_public *)iob)->_r -= c;
560 ((struct __FILE_public *)iob)->_p += c;
573 adios (NULL, "m_getfld() called with bogus state of %d", state);
577 msg_count = cp - buf;
589 register char *delimstr;
592 * Figure out what the message delimitter string is for this
593 * maildrop. (This used to be part of m_Eom but I didn't like
594 * the idea of an "if" statement that could only succeed on the
595 * first call to m_Eom getting executed on each call, i.e., at
596 * every newline in the message).
598 * If the first line of the maildrop is a Unix "From " line, we
599 * say the style is MBOX and eat the rest of the line. Otherwise
600 * we say the style is MMDF and look for the delimiter string
601 * specified when nmh was built (or from the mts.conf file).
604 msg_style = MS_UNKNOWN;
607 if (fread (text, sizeof(*text), 5, iob) == 5
608 && strncmp (text, "From ", 5) == 0) {
610 delimstr = "\nFrom ";
611 while ((c = getc (iob)) != '\n' && c >= 0)
614 /* not a Unix style maildrop */
615 fseek (iob, pos, SEEK_SET);
616 if (mmdlm2 == NULL || *mmdlm2 == 0)
617 mmdlm2 = "\001\001\001\001\n";
621 c = strlen (delimstr);
622 fdelim = (unsigned char *) mh_xmalloc((size_t) (c + 3));
625 msg_delim = (char *)fdelim+1;
626 edelim = (unsigned char *)msg_delim+1;
629 strcpy (msg_delim, delimstr);
630 delimend = (unsigned char *)msg_delim + edelimlen;
632 adios (NULL, "maildrop delimiter must be at least 2 bytes");
634 * build a Boyer-Moore end-position map for the matcher in m_getfld.
635 * N.B. - we don't match just the first char (since it's the newline
636 * separator) or the last char (since the matchc would have found it
637 * if it was a real delim).
639 pat_map = (unsigned char **) calloc (256, sizeof(unsigned char *));
641 for (cp = (char *) fdelim + 1; cp < (char *) delimend; cp++ )
642 pat_map[(unsigned char)*cp] = (unsigned char *) cp;
644 if (msg_style == MS_MMDF) {
645 /* flush extra msg hdrs */
646 while ((c = Getc(iob)) >= 0 && eom (c, iob))
655 m_eomsbr (int (*action)(int))
657 if ((eom_action = action)) {
664 msg_delim = (char *)fdelim + 1;
665 fdelimlen = strlen((char *)fdelim);
666 delimend = (unsigned char *)(msg_delim + edelimlen);
672 * test for msg delimiter string
676 m_Eom (int c, FILE *iob)
678 register long pos = 0L;
683 if ((i = fread (text, sizeof *text, edelimlen, iob)) != edelimlen
684 || strncmp (text, (char *)edelim, edelimlen)) {
685 if (i == 0 && msg_style == MS_MBOX)
686 /* the final newline in the (brain damaged) unix-format
687 * maildrop is part of the delimitter - delete it.
692 fseek (iob, pos, SEEK_SET);
695 fseek (iob, (long)(pos-1), SEEK_SET);
696 getc (iob); /* should be OK */
700 if (msg_style == MS_MBOX) {
701 while ((c = getc (iob)) != '\n')
710 static unsigned char *
711 matchc(int patln, char *pat, int strln, char *str)
713 register char *es = str + strln - patln;
716 register char *ep = pat + patln;
717 register char pc = *pat++;
726 while (pp < ep && *sp++ == *pp)
729 return ((unsigned char *)--str);
735 * Locate character "term" in the next "cnt" characters of "src".
736 * If found, return its address, otherwise return 0.
739 static unsigned char *
740 locc(int cnt, unsigned char *src, unsigned char term)
742 while (*src++ != term && --cnt > 0);
744 return (cnt > 0 ? --src : (unsigned char *)0);