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").
140 Reads an Internet message (RFC 5322), or one or more messages stored in a
141 maildrop in mbox (RFC 4155) or MMDF format, from a file stream. Each call
142 to m_getfld() reads one header field, or a portion of the body, in sequence.
146 state: message parse state
147 bufsz: maximum number of characters to load into buf
148 iob: input file stream
152 name: header field name (array of size NAMESZ=999)
153 buf: either a header field body or message body
154 (return value): message parse state on return from function
155 (global) int msg_count: number of characters loaded into buf
157 Functions (part of Inputs, really)
159 void m_unknown(FILE *iob): Determines the message delimiter string for the
160 maildrop. Called by inc, scan, and msh when reading from a maildrop file.
162 void m_eomsbr (int (*action)(int)): Sets the hook to check for end of
163 message in a maildrop. Called only by msh.
165 Those functions save state in the State variables listed below.
170 FLD // Field returned
171 FLDPLUS // Field returned with more to come
172 FLDEOF // Field returned ending at eom
173 BODY // Body returned with more to come
174 BODYEOF // Body returned ending at eom
175 FILEEOF // Reached end of input file
176 FMTERR // Message Format error
177 LENERR // Name too long error from getfld
179 msg_style is maildrop style, one of:
180 MS_UNKNOWN // type not known yet
181 MS_DEFAULT // default (one msg per file)
182 MS_MBOX // Unix-style "from" lines
183 MS_MMDF // string mmdlm2
186 State variables (part of Outputs)
188 m_getfld() retains state internally between calls in some state variables.
190 These two variables are global, but only used internally by m_getfld.c:
194 These are used for the end-of-message matcher when reading maildrops:
195 static unsigned char **pat_map
196 static unsigned char *fdelim
197 static unsigned char *delimend
199 static unsigned char *edelim
204 m_getfld() is restricted to operate on one file stream at a time because of
205 the retained state (see "State variables" above).
209 The first call to m_getfld() on a file stream is with a state of FLD.
210 Subsequent calls provide the state returned by the previous call.
212 Along the way, I thought of these possible interface changes that we
213 might want to consider before rototilling the internals:
215 1) To remove a global:
216 Change bufsz to be in-out instead of in, and therefore int * instead of
217 int, and use that instead of global msg_count. There are only 3 call
218 sites that use msg_count so it wouldn't take much effort to remove use of
219 it. Of course, all call sites would have to change to provide an int *
220 instead of an int. Some now pass constants.
222 2) To remove the state argument from the signature:
223 Given the Current usage and Restriction above, the state variable could
224 be removed from the signature and just retained internally.
226 3) To remove the Restriction above:
227 One approach would be for m_getfld() to retain multiple copies of that
228 state, one per iob that it sees. Another approach would be for the
229 caller to store it in an opaque struct, the address of which is passed
230 through the interface.
236 static int m_Eom (int, FILE *);
237 static unsigned char *matchc(int, char *, int, char *);
238 static unsigned char *locc(int, unsigned char *, unsigned char);
240 #define Getc(iob) getc(iob)
241 #define eom(c,iob) (msg_style != MS_DEFAULT && \
242 (((c) == *msg_delim && m_Eom(c,iob)) ||\
243 (eom_action && (*eom_action)(c))))
245 static unsigned char **pat_map;
248 * defined in sbr/m_msgdef.c = 0
249 * This is a disgusting hack for "inc" so it can know how many
250 * characters were stuffed in the buffer on the last call
251 * (see comments in uip/scansbr.c).
253 extern int msg_count;
255 static int msg_style = MS_DEFAULT;
258 * The "full" delimiter string for a packed maildrop consists
259 * of a newline followed by the actual delimiter. E.g., the
260 * full string for a Unix maildrop would be: "\n\nFrom ".
261 * "Fdelim" points to the start of the full string and is used
262 * in the BODY case of the main routine to search the buffer for
263 * a possible eom. Msg_delim points to the first character of
264 * the actual delim. string (i.e., fdelim+1). Edelim
265 * points to the 2nd character of actual delimiter string. It
266 * is used in m_Eom because the first character of the string
267 * has been read and matched before m_Eom is called.
269 static char *msg_delim = "";
270 static unsigned char *fdelim;
271 static unsigned char *delimend;
272 static int fdelimlen;
273 static unsigned char *edelim;
274 static int edelimlen;
276 static int (*eom_action)(int) = NULL;
279 # define _ptr _p /* Gag */
280 # define _cnt _r /* Retch */
281 # define _filbuf __srget /* Puke */
282 # define DEFINED__FILBUF_TO_SOMETHING_SPECIFIC
284 # if defined __CYGWIN__
285 /* Cygwin's stdio.h does not declare __srget(). */
287 # endif /* __CYGWIN__ */
290 #ifndef DEFINED__FILBUF_TO_SOMETHING_SPECIFIC
291 extern int _filbuf(FILE*);
296 m_getfld (int state, unsigned char name[NAMESZ], unsigned char *buf,
297 int bufsz, FILE *iob)
299 register unsigned char *bp, *cp, *ep, *sp;
300 register int cnt, c, i, j;
302 if ((c = Getc(iob)) < 0) {
309 /* flush null messages */
310 while ((c = Getc(iob)) >= 0 && eom (c, iob))
324 if (c == '\n' || c == '-') {
325 /* we hit the header/body separator */
326 while (c != '\n' && (c = Getc(iob)) >= 0)
329 if (c < 0 || (c = Getc(iob)) < 0 || eom (c, iob)) {
331 /* flush null messages */
332 while ((c = Getc(iob)) >= 0 && eom (c, iob))
345 * get the name of this component. take characters up
346 * to a ':', a newline or NAMESZ-1 characters, whichever
353 bp = sp = (unsigned char *) iob->_IO_read_ptr - 1;
354 j = (cnt = ((long) iob->_IO_read_end -
355 (long) iob->_IO_read_ptr) + 1) < i ? cnt : i;
356 #elif defined(__DragonFly__)
357 bp = sp = (unsigned char *) ((struct __FILE_public *)iob)->_p - 1;
358 j = (cnt = ((struct __FILE_public *)iob)->_r+1) < i ? cnt : i;
360 bp = sp = (unsigned char *) iob->_ptr - 1;
361 j = (cnt = iob->_cnt+1) < i ? cnt : i;
363 while (--j >= 0 && (c = *bp++) != ':' && c != '\n')
367 if ((cnt -= j) <= 0) {
369 iob->_IO_read_ptr = iob->_IO_read_end;
370 if (__underflow(iob) == EOF) {
371 #elif defined(__DragonFly__)
372 if (__srget(iob) == EOF) {
374 if (_filbuf(iob) == EOF) {
377 advise (NULL, "eof encountered in field \"%s\"", name);
381 iob->_IO_read_ptr++; /* NOT automatic in __underflow()! */
385 iob->_IO_read_ptr = bp + 1;
386 #elif defined(__DragonFly__)
387 ((struct __FILE_public *)iob)->_p = bp + 1;
388 ((struct __FILE_public *)iob)->_r = cnt - 1;
398 * something went wrong. possibilities are:
399 * . hit a newline (error)
400 * . got more than namesz chars. (error)
401 * . hit the end of the buffer. (loop)
404 /* We hit the end of the line without seeing ':' to
405 * terminate the field name. This is usually (always?)
406 * spam. But, blowing up is lame, especially when
407 * scan(1)ing a folder with such messages. Pretend such
408 * lines are the first of the body (at least mutt also
409 * handles it this way). */
411 /* See if buf can hold this line, since we were assuming
412 * we had a buffer of NAMESZ, not bufsz. */
413 /* + 1 for the newline */
415 /* No, it can't. Oh well, guess we'll blow up. */
417 advise (NULL, "eol encountered in field \"%s\"", name);
421 memcpy (buf, name, j - 1);
424 /* mhparse.c:get_content wants to find the position of the
425 * body start, but it thinks there's a blank line between
426 * the header and the body (naturally!), so seek back so
427 * that things line up even though we don't have that
428 * blank line in this case. Simpler parsers (e.g. mhl)
429 * get extra newlines, but that should be harmless enough,
430 * right? This is a corrupt message anyway. */
431 fseek (iob, ftell (iob) - 2, SEEK_SET);
436 advise (NULL, "field name \"%s\" exceeds %d bytes", name, NAMESZ - 2);
442 while (isspace (*--cp) && cp >= name)
449 * get (more of) the text of a field. take
450 * characters up to the end of this field (newline
451 * followed by non-blank) or bufsz-1 characters.
453 cp = buf; i = bufsz-1;
456 cnt = (long) iob->_IO_read_end - (long) iob->_IO_read_ptr;
457 bp = (unsigned char *) --iob->_IO_read_ptr;
458 #elif defined(__DragonFly__)
459 cnt = ((struct __FILE_public *)iob)->_r++;
460 bp = (unsigned char *) --((struct __FILE_public *)iob)->_p;
463 bp = (unsigned char *) --iob->_ptr;
465 c = cnt < i ? cnt : i;
466 while ((ep = locc( c, bp, '\n' ))) {
468 * if we hit the end of this field, return.
470 if ((j = *++ep) != ' ' && j != '\t') {
472 j = ep - (unsigned char *) iob->_IO_read_ptr;
473 memcpy (cp, iob->_IO_read_ptr, j);
474 iob->_IO_read_ptr = ep;
475 #elif defined(__DragonFly__)
476 j = ep - (unsigned char *) ((struct __FILE_public *)iob)->_p;
477 memcpy (cp, ((struct __FILE_public *)iob)->_p, j);
478 ((struct __FILE_public *)iob)->_p = ep;
479 ((struct __FILE_public *)iob)->_r -= j;
481 j = ep - (unsigned char *) iob->_ptr;
482 memcpy (cp, iob->_ptr, j);
494 * end of input or dest buffer - copy what we've found.
497 c += bp - (unsigned char *) iob->_IO_read_ptr;
498 memcpy( cp, iob->_IO_read_ptr, c);
499 #elif defined(__DragonFly__)
500 c += bp - (unsigned char *) ((struct __FILE_public *)iob)->_p;
501 memcpy( cp, ((struct __FILE_public *)iob)->_p, c);
503 c += bp - (unsigned char *) iob->_ptr;
504 memcpy( cp, iob->_ptr, c);
509 /* the dest buffer is full */
511 iob->_IO_read_ptr += c;
512 #elif defined(__DragonFly__)
513 ((struct __FILE_public *)iob)->_r -= c;
514 ((struct __FILE_public *)iob)->_p += c;
523 * There's one character left in the input buffer.
524 * Copy it & fill the buffer. If the last char
525 * was a newline and the next char is not whitespace,
526 * this is the end of the field. Otherwise loop.
530 *cp++ = j = *(iob->_IO_read_ptr + c);
531 iob->_IO_read_ptr = iob->_IO_read_end;
532 c = __underflow(iob);
533 iob->_IO_read_ptr++; /* NOT automatic! */
534 #elif defined(__DragonFly__)
535 *cp++ =j = *(((struct __FILE_public *)iob)->_p + c);
538 *cp++ = j = *(iob->_ptr + c);
542 ((j == '\0' || j == '\n') && c != ' ' && c != '\t')) {
546 #elif defined(__DragonFly__)
547 --((struct __FILE_public *)iob)->_p;
548 ++((struct __FILE_public *)iob)->_r;
563 * get the message body up to bufsz characters or the
564 * end of the message. Sleazy hack: if bufsz is negative
565 * we assume that we were called to copy directly into
566 * the output buffer and we don't add an eos.
568 i = (bufsz < 0) ? -bufsz : bufsz-1;
570 bp = (unsigned char *) --iob->_IO_read_ptr;
571 cnt = (long) iob->_IO_read_end - (long) iob->_IO_read_ptr;
572 #elif defined(__DragonFly__)
573 bp = (unsigned char *) --((struct __FILE_public *)iob)->_p;
574 cnt = ++((struct __FILE_public *)iob)->_r;
576 bp = (unsigned char *) --iob->_ptr;
579 c = (cnt < i ? cnt : i);
580 if (msg_style != MS_DEFAULT && c > 1) {
582 * packed maildrop - only take up to the (possible)
583 * start of the next message. This "matchc" should
584 * probably be a Boyer-Moore matcher for non-vaxen,
585 * particularly since we have the alignment table
586 * all built for the end-of-buffer test (next).
587 * But our vax timings indicate that the "matchc"
588 * instruction is 50% faster than a carefully coded
589 * B.M. matcher for most strings. (So much for elegant
590 * algorithms vs. brute force.) Since I (currently)
591 * run MH on a vax, we use the matchc instruction. --vj
593 if ((ep = matchc( fdelimlen, fdelim, c, bp )))
597 * There's no delim in the buffer but there may be
598 * a partial one at the end. If so, we want to leave
599 * it so the "eom" check on the next call picks it up.
600 * Use a modified Boyer-Moore matcher to make this
601 * check relatively cheap. The first "if" figures
602 * out what position in the pattern matches the last
603 * character in the buffer. The inner "while" matches
604 * the pattern against the buffer, backwards starting
605 * at that position. Note that unless the buffer
606 * ends with one of the characters in the pattern
607 * (excluding the first and last), we do only one test.
610 if ((sp = pat_map[*ep])) {
612 /* This if() is true unless (a) the buffer is too
613 * small to contain this delimiter prefix, or
614 * (b) it contains exactly enough chars for the
616 * For case (a) obviously we aren't going to match.
617 * For case (b), if the buffer really contained exactly
618 * a delim prefix, then the m_eom call at entry
619 * should have found it. Thus it's not a delim
620 * and we know we won't get a match.
622 if (((sp - fdelim) + 2) <= c) {
624 /* Unfortunately although fdelim has a preceding NUL
625 * we can't use this as a sentinel in case the buffer
626 * contains a NUL in exactly the wrong place (this
627 * would cause us to run off the front of fdelim).
629 while (*--ep == *--cp)
633 /* we matched the entire delim prefix,
634 * so only take the buffer up to there.
635 * we know ep >= bp -- check above prevents underrun
641 /* try matching one less char of delim string */
643 } while (--sp > fdelim);
647 memcpy( buf, bp, c );
649 iob->_IO_read_ptr += c;
650 #elif defined(__DragonFly__)
651 ((struct __FILE_public *)iob)->_r -= c;
652 ((struct __FILE_public *)iob)->_p += c;
665 adios (NULL, "m_getfld() called with bogus state of %d", state);
669 msg_count = cp - buf;
681 register char *delimstr;
684 * Figure out what the message delimitter string is for this
685 * maildrop. (This used to be part of m_Eom but I didn't like
686 * the idea of an "if" statement that could only succeed on the
687 * first call to m_Eom getting executed on each call, i.e., at
688 * every newline in the message).
690 * If the first line of the maildrop is a Unix "From " line, we
691 * say the style is MBOX and eat the rest of the line. Otherwise
692 * we say the style is MMDF and look for the delimiter string
693 * specified when nmh was built (or from the mts.conf file).
696 msg_style = MS_UNKNOWN;
699 if (fread (text, sizeof(*text), 5, iob) == 5
700 && strncmp (text, "From ", 5) == 0) {
702 delimstr = "\nFrom ";
703 while ((c = getc (iob)) != '\n' && c >= 0)
706 /* not a Unix style maildrop */
707 fseek (iob, pos, SEEK_SET);
708 if (mmdlm2 == NULL || *mmdlm2 == 0)
709 mmdlm2 = "\001\001\001\001\n";
713 c = strlen (delimstr);
714 fdelim = (unsigned char *) mh_xmalloc((size_t) (c + 3));
717 msg_delim = (char *)fdelim+1;
718 edelim = (unsigned char *)msg_delim+1;
721 strcpy (msg_delim, delimstr);
722 delimend = (unsigned char *)msg_delim + edelimlen;
724 adios (NULL, "maildrop delimiter must be at least 2 bytes");
726 * build a Boyer-Moore end-position map for the matcher in m_getfld.
727 * N.B. - we don't match just the first char (since it's the newline
728 * separator) or the last char (since the matchc would have found it
729 * if it was a real delim).
731 pat_map = (unsigned char **) calloc (256, sizeof(unsigned char *));
733 for (cp = (char *) fdelim + 1; cp < (char *) delimend; cp++ )
734 pat_map[(unsigned char)*cp] = (unsigned char *) cp;
736 if (msg_style == MS_MMDF) {
737 /* flush extra msg hdrs */
738 while ((c = Getc(iob)) >= 0 && eom (c, iob))
747 m_eomsbr (int (*action)(int))
749 if ((eom_action = action)) {
756 msg_delim = (char *)fdelim + 1;
757 fdelimlen = strlen((char *)fdelim);
758 delimend = (unsigned char *)(msg_delim + edelimlen);
764 * test for msg delimiter string
768 m_Eom (int c, FILE *iob)
770 register long pos = 0L;
775 if ((i = fread (text, sizeof *text, edelimlen, iob)) != edelimlen
776 || strncmp (text, (char *)edelim, edelimlen)) {
777 if (i == 0 && msg_style == MS_MBOX)
778 /* the final newline in the (brain damaged) unix-format
779 * maildrop is part of the delimitter - delete it.
784 fseek (iob, pos, SEEK_SET);
787 fseek (iob, (long)(pos-1), SEEK_SET);
788 getc (iob); /* should be OK */
792 if (msg_style == MS_MBOX) {
793 while ((c = getc (iob)) != '\n')
802 static unsigned char *
803 matchc(int patln, char *pat, int strln, char *str)
805 register char *es = str + strln - patln;
808 register char *ep = pat + patln;
809 register char pc = *pat++;
818 while (pp < ep && *sp++ == *pp)
821 return ((unsigned char *)--str);
827 * Locate character "term" in the next "cnt" characters of "src".
828 * If found, return its address, otherwise return 0.
831 static unsigned char *
832 locc(int cnt, unsigned char *src, unsigned char term)
834 while (*src++ != term && --cnt > 0);
836 return (cnt > 0 ? --src : (unsigned char *)0);