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.
13 ** This module has a long and checkered history.
15 ** [ Here had been some history of delimiter problems in MMDF maildrops ... ]
17 ** Unfortunately the speed issue finally caught up with us since this
18 ** routine is at the very heart of MH. To speed things up considerably, the
19 ** routine Eom() was made an auxilary function called by the macro eom().
20 ** Unless we are bursting a maildrop, the eom() macro returns FALSE saying
21 ** we aren't at the end of the message.
23 ** [ ... and here had been some more of it. ]
26 ** ------------------------
27 ** (Written by Van Jacobson for the mh6 m_getfld, January, 1986):
29 ** This routine was accounting for 60% of the cpu time used by most mh
30 ** programs. I spent a bit of time tuning and it now accounts for <10%
31 ** of the time used. Like any heavily tuned routine, it's a bit
32 ** complex and you want to be sure you understand everything that it's
33 ** doing before you start hacking on it. Let me try to emphasize
34 ** that: every line in this atrocity depends on every other line,
35 ** sometimes in subtle ways. You should understand it all, in detail,
36 ** before trying to change any part. If you do change it, test the
37 ** result thoroughly (I use a hand-constructed test file that exercises
38 ** all the ways a header name, header body, header continuation,
39 ** header-body separator, body line and body eom can align themselves
40 ** with respect to a buffer boundary). "Minor" bugs in this routine
41 ** result in garbaged or lost mail.
43 ** If you hack on this and slow it down, I, my children and my
44 ** children's children will curse you.
46 ** This routine gets used on two different types of files: normal,
47 ** single msg files and "packed" unix mailboxs (when used by inc).
48 ** The biggest impact of different file types is in "eom" testing. The
49 ** code has been carefully organized to test for eom at appropriate
50 ** times and at no other times (since the check is quite expensive).
51 ** I have tried to arrange things so that the eom check need only be
52 ** done on entry to this routine. Since an eom can only occur after a
53 ** newline, this is easy to manage for header fields. For the msg
54 ** body, we try to efficiently search the input buffer to see if
55 ** contains the eom delimiter. If it does, we take up to the
56 ** delimiter, otherwise we take everything in the buffer. (The change
57 ** to the body eom/copy processing produced the most noticeable
58 ** performance difference, particularly for "inc" and "show".)
60 ** There are three qualitatively different things this routine busts
61 ** out of a message: field names, field text and msg bodies. Field
62 ** names are typically short (~8 char) and the loop that extracts them
63 ** might terminate on a colon, newline or max width. I considered
64 ** using a Vax "scanc" to locate the end of the field followed by a
65 ** "bcopy" but the routine call overhead on a Vax is too large for this
66 ** to work on short names. If Berkeley ever makes "inline" part of the
67 ** C optimiser (so things like "scanc" turn into inline instructions) a
68 ** change here would be worthwhile.
70 ** Field text is typically 60 - 100 characters so there's (barely)
71 ** a win in doing a routine call to something that does a "locc"
72 ** followed by a "bmove". About 30% of the fields have continuations
73 ** (usually the 822 "received:" lines) and each continuation generates
74 ** another routine call. "Inline" would be a big win here, as well.
76 ** Messages, as of this writing, seem to come in two flavors: small
77 ** (~1K) and long (>2K). Most messages have 400 - 600 bytes of headers
78 ** so message bodies average at least a few hundred characters.
79 ** Assuming your system uses reasonably sized stdio buffers (1K or
80 ** more), this routine should be able to remove the body in large
81 ** (>500 byte) chunks. The makes the cost of a call to "bcopy"
82 ** small but there is a premium on checking for the eom in packed
83 ** maildrops. The eom pattern is always a simple string so we can
84 ** construct an efficient pattern matcher for it (e.g., a Vax "matchc"
85 ** instruction). Some thought went into recognizing the start of
86 ** an eom that has been split across two buffers.
88 ** This routine wants to deal with large chunks of data so, rather
89 ** than "getc" into a local buffer, it uses stdio's buffer. If
90 ** you try to use it on a non-buffered file, you'll get what you
91 ** deserve. This routine "knows" that struct FILEs have a _ptr
92 ** and a _cnt to describe the current state of the buffer and
93 ** it knows that _filbuf ignores the _ptr & _cnt and simply fills
94 ** the buffer. If stdio on your system doesn't work this way, you
95 ** may have to make small changes in this routine.
97 ** This routine also "knows" that an EOF indication on a stream is
98 ** "sticky" (i.e., you will keep getting EOF until you reposition the
99 ** stream). If your system doesn't work this way it is broken and you
100 ** should complain to the vendor. As a consequence of the sticky
101 ** EOF, this routine will never return any kind of EOF status when
102 ** there is data in "name" or "buf").
109 static int m_Eom(int, FILE *);
110 static unsigned char *matchc(int, char *, int, char *);
111 static unsigned char *locc(int, unsigned char *, unsigned char);
113 #define eom(c,iob) (ismbox && \
114 (((c) == *msg_delim && m_Eom(c,iob)) ||\
115 (eom_action && (*eom_action)(c))))
117 static unsigned char **pat_map;
120 ** This is a disgusting hack for "inc" so it can know how many
121 ** characters were stuffed in the buffer on the last call
122 ** (see comments in uip/scansbr.c).
129 ** The "full" delimiter string for a packed maildrop consists
130 ** of a newline followed by the actual delimiter. E.g., the
131 ** full string for a Unix maildrop would be: "\n\nFrom ".
132 ** "Fdelim" points to the start of the full string and is used
133 ** in the BODY case of the main routine to search the buffer for
134 ** a possible eom. Msg_delim points to the first character of
135 ** the actual delim. string (i.e., fdelim+1). Edelim
136 ** points to the 2nd character of actual delimiter string. It
137 ** is used in m_Eom because the first character of the string
138 ** has been read and matched before m_Eom is called.
140 char *msg_delim = "";
142 static unsigned char *fdelim;
143 static unsigned char *delimend;
144 static int fdelimlen;
145 static unsigned char *edelim;
146 static int edelimlen;
148 static int (*eom_action)(int) = NULL;
151 # define _ptr _p /* Gag */
152 # define _cnt _r /* Retch */
153 # define _filbuf __srget /* Puke */
154 # define DEFINED__FILBUF_TO_SOMETHING_SPECIFIC
157 #ifndef DEFINED__FILBUF_TO_SOMETHING_SPECIFIC
158 extern int _filbuf(FILE*);
163 m_getfld(int state, unsigned char *name, unsigned char *buf,
164 int bufsz, FILE *iob)
166 register unsigned char *bp, *cp, *ep, *sp;
167 register int cnt, c, i, j;
169 if ((c = getc(iob)) < 0) {
176 /* flush null messages */
177 while ((c = getc(iob)) >= 0 && eom(c, iob))
191 if (c == '\n' || c == '-') {
192 /* we hit the header/body separator */
193 while (c != '\n' && (c = getc(iob)) >= 0)
196 if (c < 0 || (c = getc(iob)) < 0 || eom(c, iob)) {
198 /* flush null messages */
199 while ((c = getc(iob)) >= 0 && eom(c, iob))
212 ** get the name of this component. take characters up
213 ** to a ':', a newline or NAMESZ-1 characters,
214 ** whichever comes first.
220 bp = sp = (unsigned char *) iob->_IO_read_ptr - 1;
221 j = (cnt = ((long) iob->_IO_read_end -
222 (long) iob->_IO_read_ptr) + 1) < i ? cnt : i;
223 #elif defined(__DragonFly__)
224 bp = sp = (unsigned char *) ((struct __FILE_public *)iob)->_p - 1;
225 j = (cnt = ((struct __FILE_public *)iob)->_r+1) < i ? cnt : i;
227 bp = sp = (unsigned char *) iob->_ptr - 1;
228 j = (cnt = iob->_cnt+1) < i ? cnt : i;
230 while (--j >= 0 && (c = *bp++) != ':' && c != '\n')
234 if ((cnt -= j) <= 0) {
236 iob->_IO_read_ptr = iob->_IO_read_end;
237 if (__underflow(iob) == EOF) {
238 #elif defined(__DragonFly__)
239 if (__srget(iob) == EOF) {
241 if (_filbuf(iob) == EOF) {
244 advise(NULL, "eof encountered in field \"%s\"", name);
248 iob->_IO_read_ptr++; /* NOT automatic in __underflow()! */
252 iob->_IO_read_ptr = bp + 1;
253 #elif defined(__DragonFly__)
254 ((struct __FILE_public *)iob)->_p = bp + 1;
255 ((struct __FILE_public *)iob)->_r = cnt - 1;
265 ** something went wrong. possibilities are:
266 ** . hit a newline (error)
267 ** . got more than namesz chars. (error)
268 ** . hit the end of the buffer. (loop)
272 ** We hit the end of the line without
273 ** seeing ':' to terminate the field name.
274 ** This is usually (always?) spam. But,
275 ** blowing up is lame, especially when
276 ** scan(1)ing a folder with such messages.
277 ** Pretend such lines are the first of
278 ** the body (at least mutt also handles
283 ** See if buf can hold this line, since we
284 ** were assuming we had a buffer of NAMESZ,
287 /* + 1 for the newline */
290 ** No, it can't. Oh well,
291 ** guess we'll blow up.
294 advise(NULL, "eol encountered in field \"%s\"", name);
298 memcpy(buf, name, j - 1);
302 ** mhparse.c:get_content wants to find
303 ** the position of the body start, but
304 ** it thinks there's a blank line between
305 ** the header and the body (naturally!),
306 ** so seek back so that things line up
307 ** even though we don't have that blank
308 ** line in this case. Simpler parsers
309 ** (e.g. mhl) get extra newlines, but
310 ** that should be harmless enough, right?
311 ** This is a corrupt message anyway.
313 fseek(iob, ftell(iob) - 2, SEEK_SET);
318 advise(NULL, "field name \"%s\" exceeds %d bytes", name, NAMESZ - 2);
324 while (isspace(*--cp) && cp >= name)
331 ** get (more of) the text of a field. take
332 ** characters up to the end of this field (newline
333 ** followed by non-blank) or bufsz-1 characters.
335 cp = buf; i = bufsz-1;
338 cnt = (long) iob->_IO_read_end - (long) iob->_IO_read_ptr;
339 bp = (unsigned char *) --iob->_IO_read_ptr;
340 #elif defined(__DragonFly__)
341 cnt = ((struct __FILE_public *)iob)->_r++;
342 bp = (unsigned char *) --((struct __FILE_public *)iob)->_p;
345 bp = (unsigned char *) --iob->_ptr;
347 c = cnt < i ? cnt : i;
348 while ((ep = locc( c, bp, '\n' ))) {
350 ** if we hit the end of this field,
353 if ((j = *++ep) != ' ' && j != '\t') {
355 j = ep - (unsigned char *) iob->_IO_read_ptr;
356 memcpy(cp, iob->_IO_read_ptr, j);
357 iob->_IO_read_ptr = ep;
358 #elif defined(__DragonFly__)
359 j = ep - (unsigned char *) ((struct __FILE_public *)iob)->_p;
360 memcpy(cp, ((struct __FILE_public *)iob)->_p, j);
361 ((struct __FILE_public *)iob)->_p = ep;
362 ((struct __FILE_public *)iob)->_r -= j;
364 j = ep - (unsigned char *) iob->_ptr;
365 memcpy(cp, iob->_ptr, j);
377 ** end of input or dest buffer - copy what
381 c += bp - (unsigned char *) iob->_IO_read_ptr;
382 memcpy(cp, iob->_IO_read_ptr, c);
383 #elif defined(__DragonFly__)
384 c += bp - (unsigned char *) ((struct __FILE_public *)iob)->_p;
385 memcpy(cp, ((struct __FILE_public *)iob)->_p, c);
387 c += bp - (unsigned char *) iob->_ptr;
388 memcpy(cp, iob->_ptr, c);
393 /* the dest buffer is full */
395 iob->_IO_read_ptr += c;
396 #elif defined(__DragonFly__)
397 ((struct __FILE_public *)iob)->_r -= c;
398 ((struct __FILE_public *)iob)->_p += c;
407 ** There's one character left in the input
408 ** buffer. Copy it & fill the buffer.
409 ** If the last char was a newline and the
410 ** next char is not whitespace, this is
411 ** the end of the field. Otherwise loop.
415 *cp++ = j = *(iob->_IO_read_ptr + c);
416 iob->_IO_read_ptr = iob->_IO_read_end;
417 c = __underflow(iob);
418 iob->_IO_read_ptr++; /* NOT automatic! */
419 #elif defined(__DragonFly__)
420 *cp++ =j = *(((struct __FILE_public *)iob)->_p + c);
423 *cp++ = j = *(iob->_ptr + c);
427 ((j == '\0' || j == '\n') && c != ' ' && c != '\t')) {
431 #elif defined(__DragonFly__)
432 --((struct __FILE_public *)iob)->_p;
433 ++((struct __FILE_public *)iob)->_r;
448 ** get the message body up to bufsz characters or
449 ** the end of the message. Sleazy hack: if bufsz
450 ** is negative we assume that we were called to
451 ** copy directly into the output buffer and we
454 i = (bufsz < 0) ? -bufsz : bufsz-1;
456 bp = (unsigned char *) --iob->_IO_read_ptr;
457 cnt = (long) iob->_IO_read_end - (long) iob->_IO_read_ptr;
458 #elif defined(__DragonFly__)
459 bp = (unsigned char *) --((struct __FILE_public *)iob)->_p;
460 cnt = ++((struct __FILE_public *)iob)->_r;
462 bp = (unsigned char *) --iob->_ptr;
465 c = (cnt < i ? cnt : i);
466 if (ismbox && c > 1) {
468 ** packed maildrop - only take up to the (possible)
469 ** start of the next message. This "matchc" should
470 ** probably be a Boyer-Moore matcher for non-vaxen,
471 ** particularly since we have the alignment table
472 ** all built for the end-of-buffer test (next).
473 ** But our vax timings indicate that the "matchc"
474 ** instruction is 50% faster than a carefully coded
475 ** B.M. matcher for most strings. (So much for
476 ** elegant algorithms vs. brute force.) Since I
477 ** (currently) run MH on a vax, we use the matchc
480 if ((ep = matchc( fdelimlen, fdelim, c, bp )))
484 ** There's no delim in the buffer but
485 ** there may be a partial one at the end.
486 ** If so, we want to leave it so the "eom"
487 ** check on the next call picks it up. Use a
488 ** modified Boyer-Moore matcher to make this
489 ** check relatively cheap. The first "if"
490 ** figures out what position in the pattern
491 ** matches the last character in the buffer.
492 ** The inner "while" matches the pattern
493 ** against the buffer, backwards starting
494 ** at that position. Note that unless the
495 ** buffer ends with one of the characters
496 ** in the pattern (excluding the first
497 ** and last), we do only one test.
500 if ((sp = pat_map[*ep])) {
503 ** This if() is true unless
504 ** (a) the buffer is too
505 ** small to contain this
507 ** or (b) it contains
508 ** exactly enough chars for
509 ** the delimiter prefix.
510 ** For case (a) obviously we
511 ** aren't going to match.
512 ** For case (b), if the
513 ** buffer really contained
514 ** exactly a delim prefix,
515 ** then the m_eom call
516 ** at entry should have
517 ** found it. Thus it's
518 ** not a delim and we know
519 ** we won't get a match.
521 if (((sp - fdelim) + 2) <= c) {
524 ** Unfortunately although fdelim has a preceding NUL
525 ** we can't use this as a sentinel in case the buffer
526 ** contains a NUL in exactly the wrong place (this
527 ** would cause us to run off the front of fdelim).
529 while (*--ep == *--cp)
533 /* we matched the entire delim prefix,
534 ** so only take the buffer up to there.
535 ** we know ep >= bp -- check above prevents underrun
541 /* try matching one less char of delim string */
543 } while (--sp > fdelim);
547 memcpy( buf, bp, c );
549 iob->_IO_read_ptr += c;
550 #elif defined(__DragonFly__)
551 ((struct __FILE_public *)iob)->_r -= c;
552 ((struct __FILE_public *)iob)->_p += c;
565 adios(NULL, "m_getfld() called with bogus state of %d", state);
569 msg_count = cp - buf;
575 thisisanmbox(FILE *iob)
581 register char *delimstr;
584 ** Figure out what the message delimitter string is for this
585 ** maildrop. (This used to be part of m_Eom but I didn't like
586 ** the idea of an "if" statement that could only succeed on the
587 ** first call to m_Eom getting executed on each call, i.e., at
588 ** every newline in the message).
590 ** If the first line of the maildrop is a Unix "From " line, we
591 ** say the style is MBOX and eat the rest of the line. Otherwise
596 if (fread(text, sizeof(*text), 5, iob) != 5) {
597 adios(NULL, "Read error");
599 if (strncmp(text, "From ", 5)!=0) {
600 adios(NULL, "No Unix style (mbox) maildrop.");
603 delimstr = "\nFrom ";
604 while ((c = getc(iob)) != '\n' && c >= 0) {
607 c = strlen(delimstr);
608 fdelim = (unsigned char *) mh_xmalloc((size_t) (c + 3));
611 msg_delim = (char *)fdelim+1;
612 edelim = (unsigned char *)msg_delim+1;
615 strcpy(msg_delim, delimstr);
616 delimend = (unsigned char *)msg_delim + edelimlen;
618 adios(NULL, "maildrop delimiter must be at least 2 bytes");
620 ** build a Boyer-Moore end-position map for the matcher in m_getfld.
621 ** N.B. - we don't match just the first char (since it's the newline
622 ** separator) or the last char (since the matchc would have found it
623 ** if it was a real delim).
625 pat_map = (unsigned char **) calloc(256, sizeof(unsigned char *));
627 for (cp = (char *) fdelim + 1; cp < (char *) delimend; cp++ )
628 pat_map[(unsigned char)*cp] = (unsigned char *) cp;
633 ** test for msg delimiter string
637 m_Eom(int c, FILE *iob)
639 register long pos = 0L;
644 if ((i = fread(text, sizeof *text, edelimlen, iob)) != edelimlen ||
645 (strncmp(text, (char *)edelim, edelimlen)!=0)) {
646 if (i == 0 && ismbox)
648 ** the final newline in the (brain damaged) unix-format
649 ** maildrop is part of the delimitter - delete it.
653 fseek(iob, (long)(pos-1), SEEK_SET);
654 getc(iob); /* should be OK */
659 while ((c = getc(iob)) != '\n' && c >= 0) {
668 static unsigned char *
669 matchc(int patln, char *pat, int strln, char *str)
671 register char *es = str + strln - patln;
674 register char *ep = pat + patln;
675 register char pc = *pat++;
684 while (pp < ep && *sp++ == *pp)
687 return ((unsigned char *)--str);
693 ** Locate character "term" in the next "cnt" characters of "src".
694 ** If found, return its address, otherwise return 0.
697 static unsigned char *
698 locc(int cnt, unsigned char *src, unsigned char term)
700 while (*src++ != term && --cnt > 0)
703 return (cnt > 0 ? --src : (unsigned char *)0);