2 ** m_getfld.c -- read/parse a message
4 ** This code is Copyright (c) 2002, by the authors of nmh. See the
5 ** COPYRIGHT file in the root directory of the nmh distribution for
6 ** complete copyright information.
15 ** This module has a long and checkered history.
17 ** [ Here had been some history of delimiter problems in MMDF maildrops ... ]
19 ** Unfortunately the speed issue finally caught up with us since this
20 ** routine is at the very heart of MH. To speed things up considerably, the
21 ** routine Eom() was made an auxilary function called by the macro eom().
22 ** Unless we are bursting a maildrop, the eom() macro returns FALSE saying
23 ** we aren't at the end of the message.
25 ** [ ... and here had been some more of it. ]
28 ** ------------------------
29 ** (Written by Van Jacobson for the mh6 m_getfld, January, 1986):
31 ** This routine was accounting for 60% of the cpu time used by most mh
32 ** programs. I spent a bit of time tuning and it now accounts for <10%
33 ** of the time used. Like any heavily tuned routine, it's a bit
34 ** complex and you want to be sure you understand everything that it's
35 ** doing before you start hacking on it. Let me try to emphasize
36 ** that: every line in this atrocity depends on every other line,
37 ** sometimes in subtle ways. You should understand it all, in detail,
38 ** before trying to change any part. If you do change it, test the
39 ** result thoroughly (I use a hand-constructed test file that exercises
40 ** all the ways a header name, header body, header continuation,
41 ** header-body separator, body line and body eom can align themselves
42 ** with respect to a buffer boundary). "Minor" bugs in this routine
43 ** result in garbaged or lost mail.
45 ** If you hack on this and slow it down, I, my children and my
46 ** children's children will curse you.
48 ** This routine gets used on two different types of files: normal,
49 ** single msg files and "packed" unix mailboxs (when used by inc).
50 ** The biggest impact of different file types is in "eom" testing. The
51 ** code has been carefully organized to test for eom at appropriate
52 ** times and at no other times (since the check is quite expensive).
53 ** I have tried to arrange things so that the eom check need only be
54 ** done on entry to this routine. Since an eom can only occur after a
55 ** newline, this is easy to manage for header fields. For the msg
56 ** body, we try to efficiently search the input buffer to see if
57 ** contains the eom delimiter. If it does, we take up to the
58 ** delimiter, otherwise we take everything in the buffer. (The change
59 ** to the body eom/copy processing produced the most noticeable
60 ** performance difference, particularly for "inc" and "show".)
62 ** There are three qualitatively different things this routine busts
63 ** out of a message: field names, field text and msg bodies. Field
64 ** names are typically short (~8 char) and the loop that extracts them
65 ** might terminate on a colon, newline or max width. I considered
66 ** using a Vax "scanc" to locate the end of the field followed by a
67 ** "bcopy" but the routine call overhead on a Vax is too large for this
68 ** to work on short names. If Berkeley ever makes "inline" part of the
69 ** C optimiser (so things like "scanc" turn into inline instructions) a
70 ** change here would be worthwhile.
72 ** Field text is typically 60 - 100 characters so there's (barely)
73 ** a win in doing a routine call to something that does a "locc"
74 ** followed by a "bmove". About 30% of the fields have continuations
75 ** (usually the 822 "received:" lines) and each continuation generates
76 ** another routine call. "Inline" would be a big win here, as well.
78 ** Messages, as of this writing, seem to come in two flavors: small
79 ** (~1K) and long (>2K). Most messages have 400 - 600 bytes of headers
80 ** so message bodies average at least a few hundred characters.
81 ** Assuming your system uses reasonably sized stdio buffers (1K or
82 ** more), this routine should be able to remove the body in large
83 ** (>500 byte) chunks. The makes the cost of a call to "bcopy"
84 ** small but there is a premium on checking for the eom in packed
85 ** maildrops. The eom pattern is always a simple string so we can
86 ** construct an efficient pattern matcher for it (e.g., a Vax "matchc"
87 ** instruction). Some thought went into recognizing the start of
88 ** an eom that has been split across two buffers.
90 ** This routine wants to deal with large chunks of data so, rather
91 ** than "getc" into a local buffer, it uses stdio's buffer. If
92 ** you try to use it on a non-buffered file, you'll get what you
93 ** deserve. This routine "knows" that struct FILEs have a _ptr
94 ** and a _cnt to describe the current state of the buffer and
95 ** it knows that _filbuf ignores the _ptr & _cnt and simply fills
96 ** the buffer. If stdio on your system doesn't work this way, you
97 ** may have to make small changes in this routine.
99 ** This routine also "knows" that an EOF indication on a stream is
100 ** "sticky" (i.e., you will keep getting EOF until you reposition the
101 ** stream). If your system doesn't work this way it is broken and you
102 ** should complain to the vendor. As a consequence of the sticky
103 ** EOF, this routine will never return any kind of EOF status when
104 ** there is data in "name" or "buf").
111 static int m_Eom(int, FILE *);
112 static unsigned char *matchc(int, char *, int, char *);
113 static unsigned char *locc(int, unsigned char *, unsigned char);
115 #define eom(c,iob) (ismbox && \
116 (((c) == *msg_delim && m_Eom(c,iob)) ||\
117 (eom_action && (*eom_action)(c))))
119 static unsigned char **pat_map;
122 ** This is a disgusting hack for "inc" so it can know how many
123 ** characters were stuffed in the buffer on the last call
124 ** (see comments in uip/scansbr.c).
131 ** The "full" delimiter string for a packed maildrop consists
132 ** of a newline followed by the actual delimiter. E.g., the
133 ** full string for a Unix maildrop would be: "\n\nFrom ".
134 ** "Fdelim" points to the start of the full string and is used
135 ** in the BODY case of the main routine to search the buffer for
136 ** a possible eom. Msg_delim points to the first character of
137 ** the actual delim. string (i.e., fdelim+1). Edelim
138 ** points to the 2nd character of actual delimiter string. It
139 ** is used in m_Eom because the first character of the string
140 ** has been read and matched before m_Eom is called.
142 static char *msg_delim = "";
144 static unsigned char *fdelim;
145 static unsigned char *delimend;
146 static int fdelimlen;
147 static unsigned char *edelim;
148 static int edelimlen;
150 static int (*eom_action)(int) = NULL;
153 # define _ptr _p /* Gag */
154 # define _cnt _r /* Retch */
155 # define _filbuf __srget /* Puke */
156 # define DEFINED__FILBUF_TO_SOMETHING_SPECIFIC
159 #ifndef DEFINED__FILBUF_TO_SOMETHING_SPECIFIC
160 extern int _filbuf(FILE*);
165 m_getfld(int state, unsigned char *name, unsigned char *buf,
166 int bufsz, FILE *iob)
168 unsigned char *bp, *cp, *ep, *sp;
171 if ((c = getc(iob)) < 0) {
178 /* flush null messages */
179 while ((c = getc(iob)) >= 0 && eom(c, iob))
193 if (c == '\n' || c == '-') {
194 /* we hit the header/body separator */
195 while (c != '\n' && (c = getc(iob)) >= 0)
198 if (c < 0 || (c = getc(iob)) < 0 || eom(c, iob)) {
200 /* flush null messages */
201 while ((c = getc(iob)) >= 0 && eom(c, iob))
214 ** get the name of this component. take characters up
215 ** to a ':', a newline or NAMESZ-1 characters,
216 ** whichever comes first.
222 bp = sp = (unsigned char *) iob->_IO_read_ptr - 1;
223 j = (cnt = ((long) iob->_IO_read_end -
224 (long) iob->_IO_read_ptr) + 1) < i ? cnt : i;
225 #elif defined(__DragonFly__)
226 bp = sp = (unsigned char *) ((struct __FILE_public *)iob)->_p - 1;
227 j = (cnt = ((struct __FILE_public *)iob)->_r+1) < i ? cnt : i;
229 bp = sp = (unsigned char *) iob->_ptr - 1;
230 j = (cnt = iob->_cnt+1) < i ? cnt : i;
232 while (--j >= 0 && (c = *bp++) != ':' && c != '\n')
236 if ((cnt -= j) <= 0) {
238 iob->_IO_read_ptr = iob->_IO_read_end;
239 if (__underflow(iob) == EOF) {
240 #elif defined(__DragonFly__)
241 if (__srget(iob) == EOF) {
243 if (_filbuf(iob) == EOF) {
246 advise(NULL, "eof encountered in field \"%s\"", name);
250 iob->_IO_read_ptr++; /* NOT automatic in __underflow()! */
254 iob->_IO_read_ptr = bp + 1;
255 #elif defined(__DragonFly__)
256 ((struct __FILE_public *)iob)->_p = bp + 1;
257 ((struct __FILE_public *)iob)->_r = cnt - 1;
267 ** something went wrong. possibilities are:
268 ** . hit a newline (error)
269 ** . got more than namesz chars. (error)
270 ** . hit the end of the buffer. (loop)
274 ** We hit the end of the line without
275 ** seeing ':' to terminate the field name.
276 ** This is usually (always?) spam. But,
277 ** blowing up is lame, especially when
278 ** scan(1)ing a folder with such messages.
279 ** Pretend such lines are the first of
280 ** the body (at least mutt also handles
285 ** See if buf can hold this line, since we
286 ** were assuming we had a buffer of NAMESZ,
289 /* + 1 for the newline */
292 ** No, it can't. Oh well,
293 ** guess we'll blow up.
296 advise(NULL, "eol encountered in field \"%s\"", name);
300 memcpy(buf, name, j - 1);
304 ** mhparse.c:get_content wants to find
305 ** the position of the body start, but
306 ** it thinks there's a blank line between
307 ** the header and the body (naturally!),
308 ** so seek back so that things line up
309 ** even though we don't have that blank
310 ** line in this case. Simpler parsers
311 ** (e.g. mhl) get extra newlines, but
312 ** that should be harmless enough, right?
313 ** This is a corrupt message anyway.
315 fseek(iob, ftell(iob) - 2, SEEK_SET);
320 advise(NULL, "field name \"%s\" exceeds %d bytes", name, NAMESZ - 2);
326 while (isspace(*--cp) && cp >= name)
333 ** get (more of) the text of a field. take
334 ** characters up to the end of this field (newline
335 ** followed by non-blank) or bufsz-1 characters.
337 cp = buf; i = bufsz-1;
340 cnt = (long) iob->_IO_read_end - (long) iob->_IO_read_ptr;
341 bp = (unsigned char *) --iob->_IO_read_ptr;
342 #elif defined(__DragonFly__)
343 cnt = ((struct __FILE_public *)iob)->_r++;
344 bp = (unsigned char *) --((struct __FILE_public *)iob)->_p;
347 bp = (unsigned char *) --iob->_ptr;
349 c = cnt < i ? cnt : i;
350 while ((ep = locc( c, bp, '\n' ))) {
352 ** if we hit the end of this field,
355 if ((j = *++ep) != ' ' && j != '\t') {
357 j = ep - (unsigned char *) iob->_IO_read_ptr;
358 memcpy(cp, iob->_IO_read_ptr, j);
359 iob->_IO_read_ptr = ep;
360 #elif defined(__DragonFly__)
361 j = ep - (unsigned char *) ((struct __FILE_public *)iob)->_p;
362 memcpy(cp, ((struct __FILE_public *)iob)->_p, j);
363 ((struct __FILE_public *)iob)->_p = ep;
364 ((struct __FILE_public *)iob)->_r -= j;
366 j = ep - (unsigned char *) iob->_ptr;
367 memcpy(cp, iob->_ptr, j);
379 ** end of input or dest buffer - copy what
383 c += bp - (unsigned char *) iob->_IO_read_ptr;
384 memcpy(cp, iob->_IO_read_ptr, c);
385 #elif defined(__DragonFly__)
386 c += bp - (unsigned char *) ((struct __FILE_public *)iob)->_p;
387 memcpy(cp, ((struct __FILE_public *)iob)->_p, c);
389 c += bp - (unsigned char *) iob->_ptr;
390 memcpy(cp, iob->_ptr, c);
395 /* the dest buffer is full */
397 iob->_IO_read_ptr += c;
398 #elif defined(__DragonFly__)
399 ((struct __FILE_public *)iob)->_r -= c;
400 ((struct __FILE_public *)iob)->_p += c;
409 ** There's one character left in the input
410 ** buffer. Copy it & fill the buffer.
411 ** If the last char was a newline and the
412 ** next char is not whitespace, this is
413 ** the end of the field. Otherwise loop.
417 *cp++ = j = *(iob->_IO_read_ptr + c);
418 iob->_IO_read_ptr = iob->_IO_read_end;
419 c = __underflow(iob);
420 iob->_IO_read_ptr++; /* NOT automatic! */
421 #elif defined(__DragonFly__)
422 *cp++ =j = *(((struct __FILE_public *)iob)->_p + c);
425 *cp++ = j = *(iob->_ptr + c);
428 if (c == EOF || ((j == '\0' || j == '\n')
429 && c != ' ' && c != '\t')) {
433 #elif defined(__DragonFly__)
434 --((struct __FILE_public *)iob)->_p;
435 ++((struct __FILE_public *)iob)->_r;
450 ** get the message body up to bufsz characters or
451 ** the end of the message. Sleazy hack: if bufsz
452 ** is negative we assume that we were called to
453 ** copy directly into the output buffer and we
456 i = (bufsz < 0) ? -bufsz : bufsz-1;
458 bp = (unsigned char *) --iob->_IO_read_ptr;
459 cnt = (long) iob->_IO_read_end - (long) iob->_IO_read_ptr;
460 #elif defined(__DragonFly__)
461 bp = (unsigned char *) --((struct __FILE_public *)iob)->_p;
462 cnt = ++((struct __FILE_public *)iob)->_r;
464 bp = (unsigned char *) --iob->_ptr;
467 c = (cnt < i ? cnt : i);
468 if (ismbox && c > 1) {
470 ** packed maildrop - only take up to the (possible)
471 ** start of the next message. This "matchc" should
472 ** probably be a Boyer-Moore matcher for non-vaxen,
473 ** particularly since we have the alignment table
474 ** all built for the end-of-buffer test (next).
475 ** But our vax timings indicate that the "matchc"
476 ** instruction is 50% faster than a carefully coded
477 ** B.M. matcher for most strings. (So much for
478 ** elegant algorithms vs. brute force.) Since I
479 ** (currently) run MH on a vax, we use the matchc
482 if ((ep = matchc( fdelimlen, fdelim, c, bp )))
486 ** There's no delim in the buffer but
487 ** there may be a partial one at the end.
488 ** If so, we want to leave it so the "eom"
489 ** check on the next call picks it up. Use a
490 ** modified Boyer-Moore matcher to make this
491 ** check relatively cheap. The first "if"
492 ** figures out what position in the pattern
493 ** matches the last character in the buffer.
494 ** The inner "while" matches the pattern
495 ** against the buffer, backwards starting
496 ** at that position. Note that unless the
497 ** buffer ends with one of the characters
498 ** in the pattern (excluding the first
499 ** and last), we do only one test.
502 if ((sp = pat_map[*ep])) {
505 ** This if() is true unless
506 ** (a) the buffer is too
507 ** small to contain this
509 ** or (b) it contains
510 ** exactly enough chars for
511 ** the delimiter prefix.
512 ** For case (a) obviously we
513 ** aren't going to match.
514 ** For case (b), if the
515 ** buffer really contained
516 ** exactly a delim prefix,
517 ** then the m_eom call
518 ** at entry should have
519 ** found it. Thus it's
520 ** not a delim and we know
521 ** we won't get a match.
523 if (((sp - fdelim) + 2) <= c) {
526 ** Unfortunately although fdelim has a preceding NUL
527 ** we can't use this as a sentinel in case the buffer
528 ** contains a NUL in exactly the wrong place (this
529 ** would cause us to run off the front of fdelim).
531 while (*--ep == *--cp)
535 /* we matched the entire delim prefix,
536 ** so only take the buffer up to there.
537 ** we know ep >= bp -- check above prevents underrun
543 /* try matching one less char of delim string */
545 } while (--sp > fdelim);
549 memcpy( buf, bp, c );
551 iob->_IO_read_ptr += c;
552 #elif defined(__DragonFly__)
553 ((struct __FILE_public *)iob)->_r -= c;
554 ((struct __FILE_public *)iob)->_p += c;
567 adios(EX_SOFTWARE, NULL, "m_getfld() called with bogus state of %d", state);
571 msg_count = cp - buf;
577 thisisanmbox(FILE *iob)
591 ** Figure out what the message delimitter string is for this
592 ** maildrop. (This used to be part of m_Eom but I didn't like
593 ** the idea of an "if" statement that could only succeed on the
594 ** first call to m_Eom getting executed on each call, i.e., at
595 ** every newline in the message).
597 ** If the first line of the maildrop is a Unix "From " line, we
598 ** say the style is MBOX and eat the rest of the line. Otherwise
602 if (fread(text, sizeof(*text), 5, iob) != 5) {
603 adios(EX_IOERR, NULL, "Read error");
605 if (strncmp(text, "From ", 5)!=0) {
606 adios(EX_USAGE, NULL, "No Unix style (mbox) maildrop.");
609 delimstr = "\nFrom ";
610 while ((c = getc(iob)) != '\n' && c >= 0) {
613 c = strlen(delimstr);
614 fdelim = (unsigned char *) mh_xmalloc((size_t) (c + 3));
617 msg_delim = (char *)fdelim+1;
618 edelim = (unsigned char *)msg_delim+1;
621 strcpy(msg_delim, delimstr);
622 delimend = (unsigned char *)msg_delim + edelimlen;
624 adios(EX_DATAERR, NULL, "maildrop delimiter must be at least 2 bytes");
626 ** build a Boyer-Moore end-position map for the matcher in m_getfld.
627 ** N.B. - we don't match just the first char (since it's the newline
628 ** separator) or the last char (since the matchc would have found it
629 ** if it was a real delim).
631 pat_map = (unsigned char **) calloc(256, sizeof(unsigned char *));
633 for (cp = (char *) fdelim + 1; cp < (char *) delimend; cp++ )
634 pat_map[(unsigned char)*cp] = (unsigned char *) cp;
639 ** test for msg delimiter string
643 m_Eom(int c, FILE *iob)
650 if ((i = fread(text, sizeof *text, edelimlen, iob)) != edelimlen ||
651 (strncmp(text, (char *)edelim, edelimlen)!=0)) {
652 if (i == 0 && ismbox)
654 ** the final newline in the (brain damaged) unix-format
655 ** maildrop is part of the delimitter - delete it.
659 fseek(iob, (long)(pos-1), SEEK_SET);
660 getc(iob); /* should be OK */
665 while ((c = getc(iob)) != '\n' && c >= 0) {
674 static unsigned char *
675 matchc(int patln, char *pat, int strln, char *str)
677 char *es = str + strln - patln;
680 char *ep = pat + patln;
690 while (pp < ep && *sp++ == *pp)
693 return ((unsigned char *)--str);
699 ** Locate character "term" in the next "cnt" characters of "src".
700 ** If found, return its address, otherwise return 0.
703 static unsigned char *
704 locc(int cnt, unsigned char *src, unsigned char term)
706 while (*src++ != term && --cnt > 0)
709 return (cnt > 0 ? --src : (unsigned char *)0);