34
35#include "opt_param.h"
36
37#include <sys/param.h>
38#include <sys/aio.h> /* for aio_swake proto */
39#include <sys/kernel.h>
40#include <sys/lock.h>
41#include <sys/mbuf.h>
42#include <sys/mutex.h>
43#include <sys/proc.h>
44#include <sys/protosw.h>
45#include <sys/resourcevar.h>
46#include <sys/signalvar.h>
47#include <sys/socket.h>
48#include <sys/socketvar.h>
49#include <sys/sx.h>
50#include <sys/sysctl.h>
51
52/*
53 * Function pointer set by the AIO routines so that the socket buffer code
54 * can call back into the AIO module if it is loaded.
55 */
56void (*aio_swake)(struct socket *, struct sockbuf *);
57
58/*
59 * Primitive routines for operating on socket buffers
60 */
61
62u_long sb_max = SB_MAX;
63u_long sb_max_adj =
64 (quad_t)SB_MAX * MCLBYTES / (MSIZE + MCLBYTES); /* adjusted sb_max */
65
66static u_long sb_efficiency = 8; /* parameter for sbreserve() */
67
68static void sbdrop_internal(struct sockbuf *sb, int len);
69static void sbflush_internal(struct sockbuf *sb);
70
71/*
72 * Socantsendmore indicates that no more data will be sent on the socket; it
73 * would normally be applied to a socket when the user informs the system
74 * that no more data is to be sent, by the protocol code (in case
75 * PRU_SHUTDOWN). Socantrcvmore indicates that no more data will be
76 * received, and will normally be applied to the socket by a protocol when it
77 * detects that the peer will send no more data. Data queued for reading in
78 * the socket may yet be read.
79 */
80void
81socantsendmore_locked(struct socket *so)
82{
83
84 SOCKBUF_LOCK_ASSERT(&so->so_snd);
85
86 so->so_snd.sb_state |= SBS_CANTSENDMORE;
87 sowwakeup_locked(so);
88 mtx_assert(SOCKBUF_MTX(&so->so_snd), MA_NOTOWNED);
89}
90
91void
92socantsendmore(struct socket *so)
93{
94
95 SOCKBUF_LOCK(&so->so_snd);
96 socantsendmore_locked(so);
97 mtx_assert(SOCKBUF_MTX(&so->so_snd), MA_NOTOWNED);
98}
99
100void
101socantrcvmore_locked(struct socket *so)
102{
103
104 SOCKBUF_LOCK_ASSERT(&so->so_rcv);
105
106 so->so_rcv.sb_state |= SBS_CANTRCVMORE;
107 sorwakeup_locked(so);
108 mtx_assert(SOCKBUF_MTX(&so->so_rcv), MA_NOTOWNED);
109}
110
111void
112socantrcvmore(struct socket *so)
113{
114
115 SOCKBUF_LOCK(&so->so_rcv);
116 socantrcvmore_locked(so);
117 mtx_assert(SOCKBUF_MTX(&so->so_rcv), MA_NOTOWNED);
118}
119
120/*
121 * Wait for data to arrive at/drain from a socket buffer.
122 */
123int
124sbwait(struct sockbuf *sb)
125{
126
127 SOCKBUF_LOCK_ASSERT(sb);
128
129 sb->sb_flags |= SB_WAIT;
130 return (msleep(&sb->sb_cc, &sb->sb_mtx,
131 (sb->sb_flags & SB_NOINTR) ? PSOCK : PSOCK | PCATCH, "sbwait",
132 sb->sb_timeo));
133}
134
135int
136sblock(struct sockbuf *sb, int flags)
137{
138
139 KASSERT((flags & SBL_VALID) == flags,
140 ("sblock: flags invalid (0x%x)", flags));
141
142 if (flags & SBL_WAIT) {
143 if ((sb->sb_flags & SB_NOINTR) ||
144 (flags & SBL_NOINTR)) {
145 sx_xlock(&sb->sb_sx);
146 return (0);
147 }
148 return (sx_xlock_sig(&sb->sb_sx));
149 } else {
150 if (sx_try_xlock(&sb->sb_sx) == 0)
151 return (EWOULDBLOCK);
152 return (0);
153 }
154}
155
156void
157sbunlock(struct sockbuf *sb)
158{
159
160 sx_xunlock(&sb->sb_sx);
161}
162
163/*
164 * Wakeup processes waiting on a socket buffer. Do asynchronous notification
165 * via SIGIO if the socket has the SS_ASYNC flag set.
166 *
167 * Called with the socket buffer lock held; will release the lock by the end
168 * of the function. This allows the caller to acquire the socket buffer lock
169 * while testing for the need for various sorts of wakeup and hold it through
170 * to the point where it's no longer required. We currently hold the lock
171 * through calls out to other subsystems (with the exception of kqueue), and
172 * then release it to avoid lock order issues. It's not clear that's
173 * correct.
174 */
175void
176sowakeup(struct socket *so, struct sockbuf *sb)
177{
178 int ret;
179
180 SOCKBUF_LOCK_ASSERT(sb);
181
182 selwakeuppri(&sb->sb_sel, PSOCK);
183 if (!SEL_WAITING(&sb->sb_sel))
184 sb->sb_flags &= ~SB_SEL;
185 if (sb->sb_flags & SB_WAIT) {
186 sb->sb_flags &= ~SB_WAIT;
187 wakeup(&sb->sb_cc);
188 }
189 KNOTE_LOCKED(&sb->sb_sel.si_note, 0);
190 if (sb->sb_upcall != NULL) {
191 ret = sb->sb_upcall(so, sb->sb_upcallarg, M_DONTWAIT);
192 if (ret == SU_ISCONNECTED) {
193 KASSERT(sb == &so->so_rcv,
194 ("SO_SND upcall returned SU_ISCONNECTED"));
195 soupcall_clear(so, SO_RCV);
196 }
197 } else
198 ret = SU_OK;
199 if (sb->sb_flags & SB_AIO)
200 aio_swake(so, sb);
201 SOCKBUF_UNLOCK(sb);
202 if (ret == SU_ISCONNECTED)
203 soisconnected(so);
204 if ((so->so_state & SS_ASYNC) && so->so_sigio != NULL)
205 pgsigio(&so->so_sigio, SIGIO, 0);
206 mtx_assert(SOCKBUF_MTX(sb), MA_NOTOWNED);
207}
208
209/*
210 * Socket buffer (struct sockbuf) utility routines.
211 *
212 * Each socket contains two socket buffers: one for sending data and one for
213 * receiving data. Each buffer contains a queue of mbufs, information about
214 * the number of mbufs and amount of data in the queue, and other fields
215 * allowing select() statements and notification on data availability to be
216 * implemented.
217 *
218 * Data stored in a socket buffer is maintained as a list of records. Each
219 * record is a list of mbufs chained together with the m_next field. Records
220 * are chained together with the m_nextpkt field. The upper level routine
221 * soreceive() expects the following conventions to be observed when placing
222 * information in the receive buffer:
223 *
224 * 1. If the protocol requires each message be preceded by the sender's name,
225 * then a record containing that name must be present before any
226 * associated data (mbuf's must be of type MT_SONAME).
227 * 2. If the protocol supports the exchange of ``access rights'' (really just
228 * additional data associated with the message), and there are ``rights''
229 * to be received, then a record containing this data should be present
230 * (mbuf's must be of type MT_RIGHTS).
231 * 3. If a name or rights record exists, then it must be followed by a data
232 * record, perhaps of zero length.
233 *
234 * Before using a new socket structure it is first necessary to reserve
235 * buffer space to the socket, by calling sbreserve(). This should commit
236 * some of the available buffer space in the system buffer pool for the
237 * socket (currently, it does nothing but enforce limits). The space should
238 * be released by calling sbrelease() when the socket is destroyed.
239 */
240int
241soreserve(struct socket *so, u_long sndcc, u_long rcvcc)
242{
243 struct thread *td = curthread;
244
245 SOCKBUF_LOCK(&so->so_snd);
246 SOCKBUF_LOCK(&so->so_rcv);
247 if (sbreserve_locked(&so->so_snd, sndcc, so, td) == 0)
248 goto bad;
249 if (sbreserve_locked(&so->so_rcv, rcvcc, so, td) == 0)
250 goto bad2;
251 if (so->so_rcv.sb_lowat == 0)
252 so->so_rcv.sb_lowat = 1;
253 if (so->so_snd.sb_lowat == 0)
254 so->so_snd.sb_lowat = MCLBYTES;
255 if (so->so_snd.sb_lowat > so->so_snd.sb_hiwat)
256 so->so_snd.sb_lowat = so->so_snd.sb_hiwat;
257 SOCKBUF_UNLOCK(&so->so_rcv);
258 SOCKBUF_UNLOCK(&so->so_snd);
259 return (0);
260bad2:
261 sbrelease_locked(&so->so_snd, so);
262bad:
263 SOCKBUF_UNLOCK(&so->so_rcv);
264 SOCKBUF_UNLOCK(&so->so_snd);
265 return (ENOBUFS);
266}
267
268static int
269sysctl_handle_sb_max(SYSCTL_HANDLER_ARGS)
270{
271 int error = 0;
272 u_long tmp_sb_max = sb_max;
273
274 error = sysctl_handle_long(oidp, &tmp_sb_max, arg2, req);
275 if (error || !req->newptr)
276 return (error);
277 if (tmp_sb_max < MSIZE + MCLBYTES)
278 return (EINVAL);
279 sb_max = tmp_sb_max;
280 sb_max_adj = (u_quad_t)sb_max * MCLBYTES / (MSIZE + MCLBYTES);
281 return (0);
282}
283
284/*
285 * Allot mbufs to a sockbuf. Attempt to scale mbmax so that mbcnt doesn't
286 * become limiting if buffering efficiency is near the normal case.
287 */
288int
289sbreserve_locked(struct sockbuf *sb, u_long cc, struct socket *so,
290 struct thread *td)
291{
292 rlim_t sbsize_limit;
293
294 SOCKBUF_LOCK_ASSERT(sb);
295
296 /*
297 * When a thread is passed, we take into account the thread's socket
298 * buffer size limit. The caller will generally pass curthread, but
299 * in the TCP input path, NULL will be passed to indicate that no
300 * appropriate thread resource limits are available. In that case,
301 * we don't apply a process limit.
302 */
303 if (cc > sb_max_adj)
304 return (0);
305 if (td != NULL) {
306 PROC_LOCK(td->td_proc);
307 sbsize_limit = lim_cur(td->td_proc, RLIMIT_SBSIZE);
308 PROC_UNLOCK(td->td_proc);
309 } else
310 sbsize_limit = RLIM_INFINITY;
311 if (!chgsbsize(so->so_cred->cr_uidinfo, &sb->sb_hiwat, cc,
312 sbsize_limit))
313 return (0);
314 sb->sb_mbmax = min(cc * sb_efficiency, sb_max);
315 if (sb->sb_lowat > sb->sb_hiwat)
316 sb->sb_lowat = sb->sb_hiwat;
317 return (1);
318}
319
320int
321sbreserve(struct sockbuf *sb, u_long cc, struct socket *so,
322 struct thread *td)
323{
324 int error;
325
326 SOCKBUF_LOCK(sb);
327 error = sbreserve_locked(sb, cc, so, td);
328 SOCKBUF_UNLOCK(sb);
329 return (error);
330}
331
332/*
333 * Free mbufs held by a socket, and reserved mbuf space.
334 */
335void
336sbrelease_internal(struct sockbuf *sb, struct socket *so)
337{
338
339 sbflush_internal(sb);
340 (void)chgsbsize(so->so_cred->cr_uidinfo, &sb->sb_hiwat, 0,
341 RLIM_INFINITY);
342 sb->sb_mbmax = 0;
343}
344
345void
346sbrelease_locked(struct sockbuf *sb, struct socket *so)
347{
348
349 SOCKBUF_LOCK_ASSERT(sb);
350
351 sbrelease_internal(sb, so);
352}
353
354void
355sbrelease(struct sockbuf *sb, struct socket *so)
356{
357
358 SOCKBUF_LOCK(sb);
359 sbrelease_locked(sb, so);
360 SOCKBUF_UNLOCK(sb);
361}
362
363void
364sbdestroy(struct sockbuf *sb, struct socket *so)
365{
366
367 sbrelease_internal(sb, so);
368}
369
370/*
371 * Routines to add and remove data from an mbuf queue.
372 *
373 * The routines sbappend() or sbappendrecord() are normally called to append
374 * new mbufs to a socket buffer, after checking that adequate space is
375 * available, comparing the function sbspace() with the amount of data to be
376 * added. sbappendrecord() differs from sbappend() in that data supplied is
377 * treated as the beginning of a new record. To place a sender's address,
378 * optional access rights, and data in a socket receive buffer,
379 * sbappendaddr() should be used. To place access rights and data in a
380 * socket receive buffer, sbappendrights() should be used. In either case,
381 * the new data begins a new record. Note that unlike sbappend() and
382 * sbappendrecord(), these routines check for the caller that there will be
383 * enough space to store the data. Each fails if there is not enough space,
384 * or if it cannot find mbufs to store additional information in.
385 *
386 * Reliable protocols may use the socket send buffer to hold data awaiting
387 * acknowledgement. Data is normally copied from a socket send buffer in a
388 * protocol with m_copy for output to a peer, and then removing the data from
389 * the socket buffer with sbdrop() or sbdroprecord() when the data is
390 * acknowledged by the peer.
391 */
392#ifdef SOCKBUF_DEBUG
393void
394sblastrecordchk(struct sockbuf *sb, const char *file, int line)
395{
396 struct mbuf *m = sb->sb_mb;
397
398 SOCKBUF_LOCK_ASSERT(sb);
399
400 while (m && m->m_nextpkt)
401 m = m->m_nextpkt;
402
403 if (m != sb->sb_lastrecord) {
404 printf("%s: sb_mb %p sb_lastrecord %p last %p\n",
405 __func__, sb->sb_mb, sb->sb_lastrecord, m);
406 printf("packet chain:\n");
407 for (m = sb->sb_mb; m != NULL; m = m->m_nextpkt)
408 printf("\t%p\n", m);
409 panic("%s from %s:%u", __func__, file, line);
410 }
411}
412
413void
414sblastmbufchk(struct sockbuf *sb, const char *file, int line)
415{
416 struct mbuf *m = sb->sb_mb;
417 struct mbuf *n;
418
419 SOCKBUF_LOCK_ASSERT(sb);
420
421 while (m && m->m_nextpkt)
422 m = m->m_nextpkt;
423
424 while (m && m->m_next)
425 m = m->m_next;
426
427 if (m != sb->sb_mbtail) {
428 printf("%s: sb_mb %p sb_mbtail %p last %p\n",
429 __func__, sb->sb_mb, sb->sb_mbtail, m);
430 printf("packet tree:\n");
431 for (m = sb->sb_mb; m != NULL; m = m->m_nextpkt) {
432 printf("\t");
433 for (n = m; n != NULL; n = n->m_next)
434 printf("%p ", n);
435 printf("\n");
436 }
437 panic("%s from %s:%u", __func__, file, line);
438 }
439}
440#endif /* SOCKBUF_DEBUG */
441
442#define SBLINKRECORD(sb, m0) do { \
443 SOCKBUF_LOCK_ASSERT(sb); \
444 if ((sb)->sb_lastrecord != NULL) \
445 (sb)->sb_lastrecord->m_nextpkt = (m0); \
446 else \
447 (sb)->sb_mb = (m0); \
448 (sb)->sb_lastrecord = (m0); \
449} while (/*CONSTCOND*/0)
450
451/*
452 * Append mbuf chain m to the last record in the socket buffer sb. The
453 * additional space associated the mbuf chain is recorded in sb. Empty mbufs
454 * are discarded and mbufs are compacted where possible.
455 */
456void
457sbappend_locked(struct sockbuf *sb, struct mbuf *m)
458{
459 struct mbuf *n;
460
461 SOCKBUF_LOCK_ASSERT(sb);
462
463 if (m == 0)
464 return;
465
466 SBLASTRECORDCHK(sb);
467 n = sb->sb_mb;
468 if (n) {
469 while (n->m_nextpkt)
470 n = n->m_nextpkt;
471 do {
472 if (n->m_flags & M_EOR) {
473 sbappendrecord_locked(sb, m); /* XXXXXX!!!! */
474 return;
475 }
476 } while (n->m_next && (n = n->m_next));
477 } else {
478 /*
479 * XXX Would like to simply use sb_mbtail here, but
480 * XXX I need to verify that I won't miss an EOR that
481 * XXX way.
482 */
483 if ((n = sb->sb_lastrecord) != NULL) {
484 do {
485 if (n->m_flags & M_EOR) {
486 sbappendrecord_locked(sb, m); /* XXXXXX!!!! */
487 return;
488 }
489 } while (n->m_next && (n = n->m_next));
490 } else {
491 /*
492 * If this is the first record in the socket buffer,
493 * it's also the last record.
494 */
495 sb->sb_lastrecord = m;
496 }
497 }
498 sbcompress(sb, m, n);
499 SBLASTRECORDCHK(sb);
500}
501
502/*
503 * Append mbuf chain m to the last record in the socket buffer sb. The
504 * additional space associated the mbuf chain is recorded in sb. Empty mbufs
505 * are discarded and mbufs are compacted where possible.
506 */
507void
508sbappend(struct sockbuf *sb, struct mbuf *m)
509{
510
511 SOCKBUF_LOCK(sb);
512 sbappend_locked(sb, m);
513 SOCKBUF_UNLOCK(sb);
514}
515
516/*
517 * This version of sbappend() should only be used when the caller absolutely
518 * knows that there will never be more than one record in the socket buffer,
519 * that is, a stream protocol (such as TCP).
520 */
521void
522sbappendstream_locked(struct sockbuf *sb, struct mbuf *m)
523{
524 SOCKBUF_LOCK_ASSERT(sb);
525
526 KASSERT(m->m_nextpkt == NULL,("sbappendstream 0"));
527 KASSERT(sb->sb_mb == sb->sb_lastrecord,("sbappendstream 1"));
528
529 SBLASTMBUFCHK(sb);
530
531 sbcompress(sb, m, sb->sb_mbtail);
532
533 sb->sb_lastrecord = sb->sb_mb;
534 SBLASTRECORDCHK(sb);
535}
536
537/*
538 * This version of sbappend() should only be used when the caller absolutely
539 * knows that there will never be more than one record in the socket buffer,
540 * that is, a stream protocol (such as TCP).
541 */
542void
543sbappendstream(struct sockbuf *sb, struct mbuf *m)
544{
545
546 SOCKBUF_LOCK(sb);
547 sbappendstream_locked(sb, m);
548 SOCKBUF_UNLOCK(sb);
549}
550
551#ifdef SOCKBUF_DEBUG
552void
553sbcheck(struct sockbuf *sb)
554{
555 struct mbuf *m;
556 struct mbuf *n = 0;
557 u_long len = 0, mbcnt = 0;
558
559 SOCKBUF_LOCK_ASSERT(sb);
560
561 for (m = sb->sb_mb; m; m = n) {
562 n = m->m_nextpkt;
563 for (; m; m = m->m_next) {
564 len += m->m_len;
565 mbcnt += MSIZE;
566 if (m->m_flags & M_EXT) /*XXX*/ /* pretty sure this is bogus */
567 mbcnt += m->m_ext.ext_size;
568 }
569 }
570 if (len != sb->sb_cc || mbcnt != sb->sb_mbcnt) {
571 printf("cc %ld != %u || mbcnt %ld != %u\n", len, sb->sb_cc,
572 mbcnt, sb->sb_mbcnt);
573 panic("sbcheck");
574 }
575}
576#endif
577
578/*
579 * As above, except the mbuf chain begins a new record.
580 */
581void
582sbappendrecord_locked(struct sockbuf *sb, struct mbuf *m0)
583{
584 struct mbuf *m;
585
586 SOCKBUF_LOCK_ASSERT(sb);
587
588 if (m0 == 0)
589 return;
590 /*
591 * Put the first mbuf on the queue. Note this permits zero length
592 * records.
593 */
594 sballoc(sb, m0);
595 SBLASTRECORDCHK(sb);
596 SBLINKRECORD(sb, m0);
597 sb->sb_mbtail = m0;
598 m = m0->m_next;
599 m0->m_next = 0;
600 if (m && (m0->m_flags & M_EOR)) {
601 m0->m_flags &= ~M_EOR;
602 m->m_flags |= M_EOR;
603 }
604 /* always call sbcompress() so it can do SBLASTMBUFCHK() */
605 sbcompress(sb, m, m0);
606}
607
608/*
609 * As above, except the mbuf chain begins a new record.
610 */
611void
612sbappendrecord(struct sockbuf *sb, struct mbuf *m0)
613{
614
615 SOCKBUF_LOCK(sb);
616 sbappendrecord_locked(sb, m0);
617 SOCKBUF_UNLOCK(sb);
618}
619
620/*
621 * Append address and data, and optionally, control (ancillary) data to the
622 * receive queue of a socket. If present, m0 must include a packet header
623 * with total length. Returns 0 if no space in sockbuf or insufficient
624 * mbufs.
625 */
626int
627sbappendaddr_locked(struct sockbuf *sb, const struct sockaddr *asa,
628 struct mbuf *m0, struct mbuf *control)
629{
630 struct mbuf *m, *n, *nlast;
631 int space = asa->sa_len;
632
633 SOCKBUF_LOCK_ASSERT(sb);
634
635 if (m0 && (m0->m_flags & M_PKTHDR) == 0)
636 panic("sbappendaddr_locked");
637 if (m0)
638 space += m0->m_pkthdr.len;
639 space += m_length(control, &n);
640
641 if (space > sbspace(sb))
642 return (0);
643#if MSIZE <= 256
644 if (asa->sa_len > MLEN)
645 return (0);
646#endif
647 MGET(m, M_DONTWAIT, MT_SONAME);
648 if (m == 0)
649 return (0);
650 m->m_len = asa->sa_len;
651 bcopy(asa, mtod(m, caddr_t), asa->sa_len);
652 if (n)
653 n->m_next = m0; /* concatenate data to control */
654 else
655 control = m0;
656 m->m_next = control;
657 for (n = m; n->m_next != NULL; n = n->m_next)
658 sballoc(sb, n);
659 sballoc(sb, n);
660 nlast = n;
661 SBLINKRECORD(sb, m);
662
663 sb->sb_mbtail = nlast;
664 SBLASTMBUFCHK(sb);
665
666 SBLASTRECORDCHK(sb);
667 return (1);
668}
669
670/*
671 * Append address and data, and optionally, control (ancillary) data to the
672 * receive queue of a socket. If present, m0 must include a packet header
673 * with total length. Returns 0 if no space in sockbuf or insufficient
674 * mbufs.
675 */
676int
677sbappendaddr(struct sockbuf *sb, const struct sockaddr *asa,
678 struct mbuf *m0, struct mbuf *control)
679{
680 int retval;
681
682 SOCKBUF_LOCK(sb);
683 retval = sbappendaddr_locked(sb, asa, m0, control);
684 SOCKBUF_UNLOCK(sb);
685 return (retval);
686}
687
688int
689sbappendcontrol_locked(struct sockbuf *sb, struct mbuf *m0,
690 struct mbuf *control)
691{
692 struct mbuf *m, *n, *mlast;
693 int space;
694
695 SOCKBUF_LOCK_ASSERT(sb);
696
697 if (control == 0)
698 panic("sbappendcontrol_locked");
699 space = m_length(control, &n) + m_length(m0, NULL);
700
701 if (space > sbspace(sb))
702 return (0);
703 n->m_next = m0; /* concatenate data to control */
704
705 SBLASTRECORDCHK(sb);
706
707 for (m = control; m->m_next; m = m->m_next)
708 sballoc(sb, m);
709 sballoc(sb, m);
710 mlast = m;
711 SBLINKRECORD(sb, control);
712
713 sb->sb_mbtail = mlast;
714 SBLASTMBUFCHK(sb);
715
716 SBLASTRECORDCHK(sb);
717 return (1);
718}
719
720int
721sbappendcontrol(struct sockbuf *sb, struct mbuf *m0, struct mbuf *control)
722{
723 int retval;
724
725 SOCKBUF_LOCK(sb);
726 retval = sbappendcontrol_locked(sb, m0, control);
727 SOCKBUF_UNLOCK(sb);
728 return (retval);
729}
730
731/*
732 * Append the data in mbuf chain (m) into the socket buffer sb following mbuf
733 * (n). If (n) is NULL, the buffer is presumed empty.
734 *
735 * When the data is compressed, mbufs in the chain may be handled in one of
736 * three ways:
737 *
738 * (1) The mbuf may simply be dropped, if it contributes nothing (no data, no
739 * record boundary, and no change in data type).
740 *
741 * (2) The mbuf may be coalesced -- i.e., data in the mbuf may be copied into
742 * an mbuf already in the socket buffer. This can occur if an
743 * appropriate mbuf exists, there is room, and no merging of data types
744 * will occur.
745 *
746 * (3) The mbuf may be appended to the end of the existing mbuf chain.
747 *
748 * If any of the new mbufs is marked as M_EOR, mark the last mbuf appended as
749 * end-of-record.
750 */
751void
752sbcompress(struct sockbuf *sb, struct mbuf *m, struct mbuf *n)
753{
754 int eor = 0;
755 struct mbuf *o;
756
757 SOCKBUF_LOCK_ASSERT(sb);
758
759 while (m) {
760 eor |= m->m_flags & M_EOR;
761 if (m->m_len == 0 &&
762 (eor == 0 ||
763 (((o = m->m_next) || (o = n)) &&
764 o->m_type == m->m_type))) {
765 if (sb->sb_lastrecord == m)
766 sb->sb_lastrecord = m->m_next;
767 m = m_free(m);
768 continue;
769 }
770 if (n && (n->m_flags & M_EOR) == 0 &&
771 M_WRITABLE(n) &&
772 ((sb->sb_flags & SB_NOCOALESCE) == 0) &&
773 m->m_len <= MCLBYTES / 4 && /* XXX: Don't copy too much */
774 m->m_len <= M_TRAILINGSPACE(n) &&
775 n->m_type == m->m_type) {
776 bcopy(mtod(m, caddr_t), mtod(n, caddr_t) + n->m_len,
777 (unsigned)m->m_len);
778 n->m_len += m->m_len;
779 sb->sb_cc += m->m_len;
780 if (m->m_type != MT_DATA && m->m_type != MT_OOBDATA)
781 /* XXX: Probably don't need.*/
782 sb->sb_ctl += m->m_len;
783 m = m_free(m);
784 continue;
785 }
786 if (n)
787 n->m_next = m;
788 else
789 sb->sb_mb = m;
790 sb->sb_mbtail = m;
791 sballoc(sb, m);
792 n = m;
793 m->m_flags &= ~M_EOR;
794 m = m->m_next;
795 n->m_next = 0;
796 }
797 if (eor) {
798 KASSERT(n != NULL, ("sbcompress: eor && n == NULL"));
799 n->m_flags |= eor;
800 }
801 SBLASTMBUFCHK(sb);
802}
803
804/*
805 * Free all mbufs in a sockbuf. Check that all resources are reclaimed.
806 */
807static void
808sbflush_internal(struct sockbuf *sb)
809{
810
811 while (sb->sb_mbcnt) {
812 /*
813 * Don't call sbdrop(sb, 0) if the leading mbuf is non-empty:
814 * we would loop forever. Panic instead.
815 */
816 if (!sb->sb_cc && (sb->sb_mb == NULL || sb->sb_mb->m_len))
817 break;
818 sbdrop_internal(sb, (int)sb->sb_cc);
819 }
820 if (sb->sb_cc || sb->sb_mb || sb->sb_mbcnt)
821 panic("sbflush_internal: cc %u || mb %p || mbcnt %u",
822 sb->sb_cc, (void *)sb->sb_mb, sb->sb_mbcnt);
823}
824
825void
826sbflush_locked(struct sockbuf *sb)
827{
828
829 SOCKBUF_LOCK_ASSERT(sb);
830 sbflush_internal(sb);
831}
832
833void
834sbflush(struct sockbuf *sb)
835{
836
837 SOCKBUF_LOCK(sb);
838 sbflush_locked(sb);
839 SOCKBUF_UNLOCK(sb);
840}
841
842/*
843 * Drop data from (the front of) a sockbuf.
844 */
845static void
846sbdrop_internal(struct sockbuf *sb, int len)
847{
848 struct mbuf *m;
849 struct mbuf *next;
850
851 next = (m = sb->sb_mb) ? m->m_nextpkt : 0;
852 while (len > 0) {
853 if (m == 0) {
854 if (next == 0)
855 panic("sbdrop");
856 m = next;
857 next = m->m_nextpkt;
858 continue;
859 }
860 if (m->m_len > len) {
861 m->m_len -= len;
862 m->m_data += len;
863 sb->sb_cc -= len;
864 if (sb->sb_sndptroff != 0)
865 sb->sb_sndptroff -= len;
866 if (m->m_type != MT_DATA && m->m_type != MT_OOBDATA)
867 sb->sb_ctl -= len;
868 break;
869 }
870 len -= m->m_len;
871 sbfree(sb, m);
872 m = m_free(m);
873 }
874 while (m && m->m_len == 0) {
875 sbfree(sb, m);
876 m = m_free(m);
877 }
878 if (m) {
879 sb->sb_mb = m;
880 m->m_nextpkt = next;
881 } else
882 sb->sb_mb = next;
883 /*
884 * First part is an inline SB_EMPTY_FIXUP(). Second part makes sure
885 * sb_lastrecord is up-to-date if we dropped part of the last record.
886 */
887 m = sb->sb_mb;
888 if (m == NULL) {
889 sb->sb_mbtail = NULL;
890 sb->sb_lastrecord = NULL;
891 } else if (m->m_nextpkt == NULL) {
892 sb->sb_lastrecord = m;
893 }
894}
895
896/*
897 * Drop data from (the front of) a sockbuf.
898 */
899void
900sbdrop_locked(struct sockbuf *sb, int len)
901{
902
903 SOCKBUF_LOCK_ASSERT(sb);
904
905 sbdrop_internal(sb, len);
906}
907
908void
909sbdrop(struct sockbuf *sb, int len)
910{
911
912 SOCKBUF_LOCK(sb);
913 sbdrop_locked(sb, len);
914 SOCKBUF_UNLOCK(sb);
915}
916
917/*
918 * Maintain a pointer and offset pair into the socket buffer mbuf chain to
919 * avoid traversal of the entire socket buffer for larger offsets.
920 */
921struct mbuf *
922sbsndptr(struct sockbuf *sb, u_int off, u_int len, u_int *moff)
923{
924 struct mbuf *m, *ret;
925
926 KASSERT(sb->sb_mb != NULL, ("%s: sb_mb is NULL", __func__));
927 KASSERT(off + len <= sb->sb_cc, ("%s: beyond sb", __func__));
928 KASSERT(sb->sb_sndptroff <= sb->sb_cc, ("%s: sndptroff broken", __func__));
929
930 /*
931 * Is off below stored offset? Happens on retransmits.
932 * Just return, we can't help here.
933 */
934 if (sb->sb_sndptroff > off) {
935 *moff = off;
936 return (sb->sb_mb);
937 }
938
939 /* Return closest mbuf in chain for current offset. */
940 *moff = off - sb->sb_sndptroff;
941 m = ret = sb->sb_sndptr ? sb->sb_sndptr : sb->sb_mb;
942
943 /* Advance by len to be as close as possible for the next transmit. */
944 for (off = off - sb->sb_sndptroff + len - 1;
945 off > 0 && m != NULL && off >= m->m_len;
946 m = m->m_next) {
947 sb->sb_sndptroff += m->m_len;
948 off -= m->m_len;
949 }
950 if (off > 0 && m == NULL)
951 panic("%s: sockbuf %p and mbuf %p clashing", __func__, sb, ret);
952 sb->sb_sndptr = m;
953
954 return (ret);
955}
956
957/*
958 * Drop a record off the front of a sockbuf and move the next record to the
959 * front.
960 */
961void
962sbdroprecord_locked(struct sockbuf *sb)
963{
964 struct mbuf *m;
965
966 SOCKBUF_LOCK_ASSERT(sb);
967
968 m = sb->sb_mb;
969 if (m) {
970 sb->sb_mb = m->m_nextpkt;
971 do {
972 sbfree(sb, m);
973 m = m_free(m);
974 } while (m);
975 }
976 SB_EMPTY_FIXUP(sb);
977}
978
979/*
980 * Drop a record off the front of a sockbuf and move the next record to the
981 * front.
982 */
983void
984sbdroprecord(struct sockbuf *sb)
985{
986
987 SOCKBUF_LOCK(sb);
988 sbdroprecord_locked(sb);
989 SOCKBUF_UNLOCK(sb);
990}
991
992/*
993 * Create a "control" mbuf containing the specified data with the specified
994 * type for presentation on a socket buffer.
995 */
996struct mbuf *
997sbcreatecontrol(caddr_t p, int size, int type, int level)
998{
999 struct cmsghdr *cp;
1000 struct mbuf *m;
1001
1002 if (CMSG_SPACE((u_int)size) > MCLBYTES)
1003 return ((struct mbuf *) NULL);
1004 if (CMSG_SPACE((u_int)size) > MLEN)
1005 m = m_getcl(M_DONTWAIT, MT_CONTROL, 0);
1006 else
1007 m = m_get(M_DONTWAIT, MT_CONTROL);
1008 if (m == NULL)
1009 return ((struct mbuf *) NULL);
1010 cp = mtod(m, struct cmsghdr *);
1011 m->m_len = 0;
1012 KASSERT(CMSG_SPACE((u_int)size) <= M_TRAILINGSPACE(m),
1013 ("sbcreatecontrol: short mbuf"));
1014 if (p != NULL)
1015 (void)memcpy(CMSG_DATA(cp), p, size);
1016 m->m_len = CMSG_SPACE(size);
1017 cp->cmsg_len = CMSG_LEN(size);
1018 cp->cmsg_level = level;
1019 cp->cmsg_type = type;
1020 return (m);
1021}
1022
1023/*
1024 * This does the same for socket buffers that sotoxsocket does for sockets:
1025 * generate an user-format data structure describing the socket buffer. Note
1026 * that the xsockbuf structure, since it is always embedded in a socket, does
1027 * not include a self pointer nor a length. We make this entry point public
1028 * in case some other mechanism needs it.
1029 */
1030void
1031sbtoxsockbuf(struct sockbuf *sb, struct xsockbuf *xsb)
1032{
1033
1034 xsb->sb_cc = sb->sb_cc;
1035 xsb->sb_hiwat = sb->sb_hiwat;
1036 xsb->sb_mbcnt = sb->sb_mbcnt;
1037 xsb->sb_mcnt = sb->sb_mcnt;
1038 xsb->sb_ccnt = sb->sb_ccnt;
1039 xsb->sb_mbmax = sb->sb_mbmax;
1040 xsb->sb_lowat = sb->sb_lowat;
1041 xsb->sb_flags = sb->sb_flags;
1042 xsb->sb_timeo = sb->sb_timeo;
1043}
1044
1045/* This takes the place of kern.maxsockbuf, which moved to kern.ipc. */
1046static int dummy;
1047SYSCTL_INT(_kern, KERN_DUMMY, dummy, CTLFLAG_RW, &dummy, 0, "");
1048SYSCTL_OID(_kern_ipc, KIPC_MAXSOCKBUF, maxsockbuf, CTLTYPE_ULONG|CTLFLAG_RW,
1049 &sb_max, 0, sysctl_handle_sb_max, "LU", "Maximum socket buffer size");
1050SYSCTL_ULONG(_kern_ipc, KIPC_SOCKBUF_WASTE, sockbuf_waste_factor, CTLFLAG_RW,
| 34
35#include "opt_param.h"
36
37#include <sys/param.h>
38#include <sys/aio.h> /* for aio_swake proto */
39#include <sys/kernel.h>
40#include <sys/lock.h>
41#include <sys/mbuf.h>
42#include <sys/mutex.h>
43#include <sys/proc.h>
44#include <sys/protosw.h>
45#include <sys/resourcevar.h>
46#include <sys/signalvar.h>
47#include <sys/socket.h>
48#include <sys/socketvar.h>
49#include <sys/sx.h>
50#include <sys/sysctl.h>
51
52/*
53 * Function pointer set by the AIO routines so that the socket buffer code
54 * can call back into the AIO module if it is loaded.
55 */
56void (*aio_swake)(struct socket *, struct sockbuf *);
57
58/*
59 * Primitive routines for operating on socket buffers
60 */
61
62u_long sb_max = SB_MAX;
63u_long sb_max_adj =
64 (quad_t)SB_MAX * MCLBYTES / (MSIZE + MCLBYTES); /* adjusted sb_max */
65
66static u_long sb_efficiency = 8; /* parameter for sbreserve() */
67
68static void sbdrop_internal(struct sockbuf *sb, int len);
69static void sbflush_internal(struct sockbuf *sb);
70
71/*
72 * Socantsendmore indicates that no more data will be sent on the socket; it
73 * would normally be applied to a socket when the user informs the system
74 * that no more data is to be sent, by the protocol code (in case
75 * PRU_SHUTDOWN). Socantrcvmore indicates that no more data will be
76 * received, and will normally be applied to the socket by a protocol when it
77 * detects that the peer will send no more data. Data queued for reading in
78 * the socket may yet be read.
79 */
80void
81socantsendmore_locked(struct socket *so)
82{
83
84 SOCKBUF_LOCK_ASSERT(&so->so_snd);
85
86 so->so_snd.sb_state |= SBS_CANTSENDMORE;
87 sowwakeup_locked(so);
88 mtx_assert(SOCKBUF_MTX(&so->so_snd), MA_NOTOWNED);
89}
90
91void
92socantsendmore(struct socket *so)
93{
94
95 SOCKBUF_LOCK(&so->so_snd);
96 socantsendmore_locked(so);
97 mtx_assert(SOCKBUF_MTX(&so->so_snd), MA_NOTOWNED);
98}
99
100void
101socantrcvmore_locked(struct socket *so)
102{
103
104 SOCKBUF_LOCK_ASSERT(&so->so_rcv);
105
106 so->so_rcv.sb_state |= SBS_CANTRCVMORE;
107 sorwakeup_locked(so);
108 mtx_assert(SOCKBUF_MTX(&so->so_rcv), MA_NOTOWNED);
109}
110
111void
112socantrcvmore(struct socket *so)
113{
114
115 SOCKBUF_LOCK(&so->so_rcv);
116 socantrcvmore_locked(so);
117 mtx_assert(SOCKBUF_MTX(&so->so_rcv), MA_NOTOWNED);
118}
119
120/*
121 * Wait for data to arrive at/drain from a socket buffer.
122 */
123int
124sbwait(struct sockbuf *sb)
125{
126
127 SOCKBUF_LOCK_ASSERT(sb);
128
129 sb->sb_flags |= SB_WAIT;
130 return (msleep(&sb->sb_cc, &sb->sb_mtx,
131 (sb->sb_flags & SB_NOINTR) ? PSOCK : PSOCK | PCATCH, "sbwait",
132 sb->sb_timeo));
133}
134
135int
136sblock(struct sockbuf *sb, int flags)
137{
138
139 KASSERT((flags & SBL_VALID) == flags,
140 ("sblock: flags invalid (0x%x)", flags));
141
142 if (flags & SBL_WAIT) {
143 if ((sb->sb_flags & SB_NOINTR) ||
144 (flags & SBL_NOINTR)) {
145 sx_xlock(&sb->sb_sx);
146 return (0);
147 }
148 return (sx_xlock_sig(&sb->sb_sx));
149 } else {
150 if (sx_try_xlock(&sb->sb_sx) == 0)
151 return (EWOULDBLOCK);
152 return (0);
153 }
154}
155
156void
157sbunlock(struct sockbuf *sb)
158{
159
160 sx_xunlock(&sb->sb_sx);
161}
162
163/*
164 * Wakeup processes waiting on a socket buffer. Do asynchronous notification
165 * via SIGIO if the socket has the SS_ASYNC flag set.
166 *
167 * Called with the socket buffer lock held; will release the lock by the end
168 * of the function. This allows the caller to acquire the socket buffer lock
169 * while testing for the need for various sorts of wakeup and hold it through
170 * to the point where it's no longer required. We currently hold the lock
171 * through calls out to other subsystems (with the exception of kqueue), and
172 * then release it to avoid lock order issues. It's not clear that's
173 * correct.
174 */
175void
176sowakeup(struct socket *so, struct sockbuf *sb)
177{
178 int ret;
179
180 SOCKBUF_LOCK_ASSERT(sb);
181
182 selwakeuppri(&sb->sb_sel, PSOCK);
183 if (!SEL_WAITING(&sb->sb_sel))
184 sb->sb_flags &= ~SB_SEL;
185 if (sb->sb_flags & SB_WAIT) {
186 sb->sb_flags &= ~SB_WAIT;
187 wakeup(&sb->sb_cc);
188 }
189 KNOTE_LOCKED(&sb->sb_sel.si_note, 0);
190 if (sb->sb_upcall != NULL) {
191 ret = sb->sb_upcall(so, sb->sb_upcallarg, M_DONTWAIT);
192 if (ret == SU_ISCONNECTED) {
193 KASSERT(sb == &so->so_rcv,
194 ("SO_SND upcall returned SU_ISCONNECTED"));
195 soupcall_clear(so, SO_RCV);
196 }
197 } else
198 ret = SU_OK;
199 if (sb->sb_flags & SB_AIO)
200 aio_swake(so, sb);
201 SOCKBUF_UNLOCK(sb);
202 if (ret == SU_ISCONNECTED)
203 soisconnected(so);
204 if ((so->so_state & SS_ASYNC) && so->so_sigio != NULL)
205 pgsigio(&so->so_sigio, SIGIO, 0);
206 mtx_assert(SOCKBUF_MTX(sb), MA_NOTOWNED);
207}
208
209/*
210 * Socket buffer (struct sockbuf) utility routines.
211 *
212 * Each socket contains two socket buffers: one for sending data and one for
213 * receiving data. Each buffer contains a queue of mbufs, information about
214 * the number of mbufs and amount of data in the queue, and other fields
215 * allowing select() statements and notification on data availability to be
216 * implemented.
217 *
218 * Data stored in a socket buffer is maintained as a list of records. Each
219 * record is a list of mbufs chained together with the m_next field. Records
220 * are chained together with the m_nextpkt field. The upper level routine
221 * soreceive() expects the following conventions to be observed when placing
222 * information in the receive buffer:
223 *
224 * 1. If the protocol requires each message be preceded by the sender's name,
225 * then a record containing that name must be present before any
226 * associated data (mbuf's must be of type MT_SONAME).
227 * 2. If the protocol supports the exchange of ``access rights'' (really just
228 * additional data associated with the message), and there are ``rights''
229 * to be received, then a record containing this data should be present
230 * (mbuf's must be of type MT_RIGHTS).
231 * 3. If a name or rights record exists, then it must be followed by a data
232 * record, perhaps of zero length.
233 *
234 * Before using a new socket structure it is first necessary to reserve
235 * buffer space to the socket, by calling sbreserve(). This should commit
236 * some of the available buffer space in the system buffer pool for the
237 * socket (currently, it does nothing but enforce limits). The space should
238 * be released by calling sbrelease() when the socket is destroyed.
239 */
240int
241soreserve(struct socket *so, u_long sndcc, u_long rcvcc)
242{
243 struct thread *td = curthread;
244
245 SOCKBUF_LOCK(&so->so_snd);
246 SOCKBUF_LOCK(&so->so_rcv);
247 if (sbreserve_locked(&so->so_snd, sndcc, so, td) == 0)
248 goto bad;
249 if (sbreserve_locked(&so->so_rcv, rcvcc, so, td) == 0)
250 goto bad2;
251 if (so->so_rcv.sb_lowat == 0)
252 so->so_rcv.sb_lowat = 1;
253 if (so->so_snd.sb_lowat == 0)
254 so->so_snd.sb_lowat = MCLBYTES;
255 if (so->so_snd.sb_lowat > so->so_snd.sb_hiwat)
256 so->so_snd.sb_lowat = so->so_snd.sb_hiwat;
257 SOCKBUF_UNLOCK(&so->so_rcv);
258 SOCKBUF_UNLOCK(&so->so_snd);
259 return (0);
260bad2:
261 sbrelease_locked(&so->so_snd, so);
262bad:
263 SOCKBUF_UNLOCK(&so->so_rcv);
264 SOCKBUF_UNLOCK(&so->so_snd);
265 return (ENOBUFS);
266}
267
268static int
269sysctl_handle_sb_max(SYSCTL_HANDLER_ARGS)
270{
271 int error = 0;
272 u_long tmp_sb_max = sb_max;
273
274 error = sysctl_handle_long(oidp, &tmp_sb_max, arg2, req);
275 if (error || !req->newptr)
276 return (error);
277 if (tmp_sb_max < MSIZE + MCLBYTES)
278 return (EINVAL);
279 sb_max = tmp_sb_max;
280 sb_max_adj = (u_quad_t)sb_max * MCLBYTES / (MSIZE + MCLBYTES);
281 return (0);
282}
283
284/*
285 * Allot mbufs to a sockbuf. Attempt to scale mbmax so that mbcnt doesn't
286 * become limiting if buffering efficiency is near the normal case.
287 */
288int
289sbreserve_locked(struct sockbuf *sb, u_long cc, struct socket *so,
290 struct thread *td)
291{
292 rlim_t sbsize_limit;
293
294 SOCKBUF_LOCK_ASSERT(sb);
295
296 /*
297 * When a thread is passed, we take into account the thread's socket
298 * buffer size limit. The caller will generally pass curthread, but
299 * in the TCP input path, NULL will be passed to indicate that no
300 * appropriate thread resource limits are available. In that case,
301 * we don't apply a process limit.
302 */
303 if (cc > sb_max_adj)
304 return (0);
305 if (td != NULL) {
306 PROC_LOCK(td->td_proc);
307 sbsize_limit = lim_cur(td->td_proc, RLIMIT_SBSIZE);
308 PROC_UNLOCK(td->td_proc);
309 } else
310 sbsize_limit = RLIM_INFINITY;
311 if (!chgsbsize(so->so_cred->cr_uidinfo, &sb->sb_hiwat, cc,
312 sbsize_limit))
313 return (0);
314 sb->sb_mbmax = min(cc * sb_efficiency, sb_max);
315 if (sb->sb_lowat > sb->sb_hiwat)
316 sb->sb_lowat = sb->sb_hiwat;
317 return (1);
318}
319
320int
321sbreserve(struct sockbuf *sb, u_long cc, struct socket *so,
322 struct thread *td)
323{
324 int error;
325
326 SOCKBUF_LOCK(sb);
327 error = sbreserve_locked(sb, cc, so, td);
328 SOCKBUF_UNLOCK(sb);
329 return (error);
330}
331
332/*
333 * Free mbufs held by a socket, and reserved mbuf space.
334 */
335void
336sbrelease_internal(struct sockbuf *sb, struct socket *so)
337{
338
339 sbflush_internal(sb);
340 (void)chgsbsize(so->so_cred->cr_uidinfo, &sb->sb_hiwat, 0,
341 RLIM_INFINITY);
342 sb->sb_mbmax = 0;
343}
344
345void
346sbrelease_locked(struct sockbuf *sb, struct socket *so)
347{
348
349 SOCKBUF_LOCK_ASSERT(sb);
350
351 sbrelease_internal(sb, so);
352}
353
354void
355sbrelease(struct sockbuf *sb, struct socket *so)
356{
357
358 SOCKBUF_LOCK(sb);
359 sbrelease_locked(sb, so);
360 SOCKBUF_UNLOCK(sb);
361}
362
363void
364sbdestroy(struct sockbuf *sb, struct socket *so)
365{
366
367 sbrelease_internal(sb, so);
368}
369
370/*
371 * Routines to add and remove data from an mbuf queue.
372 *
373 * The routines sbappend() or sbappendrecord() are normally called to append
374 * new mbufs to a socket buffer, after checking that adequate space is
375 * available, comparing the function sbspace() with the amount of data to be
376 * added. sbappendrecord() differs from sbappend() in that data supplied is
377 * treated as the beginning of a new record. To place a sender's address,
378 * optional access rights, and data in a socket receive buffer,
379 * sbappendaddr() should be used. To place access rights and data in a
380 * socket receive buffer, sbappendrights() should be used. In either case,
381 * the new data begins a new record. Note that unlike sbappend() and
382 * sbappendrecord(), these routines check for the caller that there will be
383 * enough space to store the data. Each fails if there is not enough space,
384 * or if it cannot find mbufs to store additional information in.
385 *
386 * Reliable protocols may use the socket send buffer to hold data awaiting
387 * acknowledgement. Data is normally copied from a socket send buffer in a
388 * protocol with m_copy for output to a peer, and then removing the data from
389 * the socket buffer with sbdrop() or sbdroprecord() when the data is
390 * acknowledged by the peer.
391 */
392#ifdef SOCKBUF_DEBUG
393void
394sblastrecordchk(struct sockbuf *sb, const char *file, int line)
395{
396 struct mbuf *m = sb->sb_mb;
397
398 SOCKBUF_LOCK_ASSERT(sb);
399
400 while (m && m->m_nextpkt)
401 m = m->m_nextpkt;
402
403 if (m != sb->sb_lastrecord) {
404 printf("%s: sb_mb %p sb_lastrecord %p last %p\n",
405 __func__, sb->sb_mb, sb->sb_lastrecord, m);
406 printf("packet chain:\n");
407 for (m = sb->sb_mb; m != NULL; m = m->m_nextpkt)
408 printf("\t%p\n", m);
409 panic("%s from %s:%u", __func__, file, line);
410 }
411}
412
413void
414sblastmbufchk(struct sockbuf *sb, const char *file, int line)
415{
416 struct mbuf *m = sb->sb_mb;
417 struct mbuf *n;
418
419 SOCKBUF_LOCK_ASSERT(sb);
420
421 while (m && m->m_nextpkt)
422 m = m->m_nextpkt;
423
424 while (m && m->m_next)
425 m = m->m_next;
426
427 if (m != sb->sb_mbtail) {
428 printf("%s: sb_mb %p sb_mbtail %p last %p\n",
429 __func__, sb->sb_mb, sb->sb_mbtail, m);
430 printf("packet tree:\n");
431 for (m = sb->sb_mb; m != NULL; m = m->m_nextpkt) {
432 printf("\t");
433 for (n = m; n != NULL; n = n->m_next)
434 printf("%p ", n);
435 printf("\n");
436 }
437 panic("%s from %s:%u", __func__, file, line);
438 }
439}
440#endif /* SOCKBUF_DEBUG */
441
442#define SBLINKRECORD(sb, m0) do { \
443 SOCKBUF_LOCK_ASSERT(sb); \
444 if ((sb)->sb_lastrecord != NULL) \
445 (sb)->sb_lastrecord->m_nextpkt = (m0); \
446 else \
447 (sb)->sb_mb = (m0); \
448 (sb)->sb_lastrecord = (m0); \
449} while (/*CONSTCOND*/0)
450
451/*
452 * Append mbuf chain m to the last record in the socket buffer sb. The
453 * additional space associated the mbuf chain is recorded in sb. Empty mbufs
454 * are discarded and mbufs are compacted where possible.
455 */
456void
457sbappend_locked(struct sockbuf *sb, struct mbuf *m)
458{
459 struct mbuf *n;
460
461 SOCKBUF_LOCK_ASSERT(sb);
462
463 if (m == 0)
464 return;
465
466 SBLASTRECORDCHK(sb);
467 n = sb->sb_mb;
468 if (n) {
469 while (n->m_nextpkt)
470 n = n->m_nextpkt;
471 do {
472 if (n->m_flags & M_EOR) {
473 sbappendrecord_locked(sb, m); /* XXXXXX!!!! */
474 return;
475 }
476 } while (n->m_next && (n = n->m_next));
477 } else {
478 /*
479 * XXX Would like to simply use sb_mbtail here, but
480 * XXX I need to verify that I won't miss an EOR that
481 * XXX way.
482 */
483 if ((n = sb->sb_lastrecord) != NULL) {
484 do {
485 if (n->m_flags & M_EOR) {
486 sbappendrecord_locked(sb, m); /* XXXXXX!!!! */
487 return;
488 }
489 } while (n->m_next && (n = n->m_next));
490 } else {
491 /*
492 * If this is the first record in the socket buffer,
493 * it's also the last record.
494 */
495 sb->sb_lastrecord = m;
496 }
497 }
498 sbcompress(sb, m, n);
499 SBLASTRECORDCHK(sb);
500}
501
502/*
503 * Append mbuf chain m to the last record in the socket buffer sb. The
504 * additional space associated the mbuf chain is recorded in sb. Empty mbufs
505 * are discarded and mbufs are compacted where possible.
506 */
507void
508sbappend(struct sockbuf *sb, struct mbuf *m)
509{
510
511 SOCKBUF_LOCK(sb);
512 sbappend_locked(sb, m);
513 SOCKBUF_UNLOCK(sb);
514}
515
516/*
517 * This version of sbappend() should only be used when the caller absolutely
518 * knows that there will never be more than one record in the socket buffer,
519 * that is, a stream protocol (such as TCP).
520 */
521void
522sbappendstream_locked(struct sockbuf *sb, struct mbuf *m)
523{
524 SOCKBUF_LOCK_ASSERT(sb);
525
526 KASSERT(m->m_nextpkt == NULL,("sbappendstream 0"));
527 KASSERT(sb->sb_mb == sb->sb_lastrecord,("sbappendstream 1"));
528
529 SBLASTMBUFCHK(sb);
530
531 sbcompress(sb, m, sb->sb_mbtail);
532
533 sb->sb_lastrecord = sb->sb_mb;
534 SBLASTRECORDCHK(sb);
535}
536
537/*
538 * This version of sbappend() should only be used when the caller absolutely
539 * knows that there will never be more than one record in the socket buffer,
540 * that is, a stream protocol (such as TCP).
541 */
542void
543sbappendstream(struct sockbuf *sb, struct mbuf *m)
544{
545
546 SOCKBUF_LOCK(sb);
547 sbappendstream_locked(sb, m);
548 SOCKBUF_UNLOCK(sb);
549}
550
551#ifdef SOCKBUF_DEBUG
552void
553sbcheck(struct sockbuf *sb)
554{
555 struct mbuf *m;
556 struct mbuf *n = 0;
557 u_long len = 0, mbcnt = 0;
558
559 SOCKBUF_LOCK_ASSERT(sb);
560
561 for (m = sb->sb_mb; m; m = n) {
562 n = m->m_nextpkt;
563 for (; m; m = m->m_next) {
564 len += m->m_len;
565 mbcnt += MSIZE;
566 if (m->m_flags & M_EXT) /*XXX*/ /* pretty sure this is bogus */
567 mbcnt += m->m_ext.ext_size;
568 }
569 }
570 if (len != sb->sb_cc || mbcnt != sb->sb_mbcnt) {
571 printf("cc %ld != %u || mbcnt %ld != %u\n", len, sb->sb_cc,
572 mbcnt, sb->sb_mbcnt);
573 panic("sbcheck");
574 }
575}
576#endif
577
578/*
579 * As above, except the mbuf chain begins a new record.
580 */
581void
582sbappendrecord_locked(struct sockbuf *sb, struct mbuf *m0)
583{
584 struct mbuf *m;
585
586 SOCKBUF_LOCK_ASSERT(sb);
587
588 if (m0 == 0)
589 return;
590 /*
591 * Put the first mbuf on the queue. Note this permits zero length
592 * records.
593 */
594 sballoc(sb, m0);
595 SBLASTRECORDCHK(sb);
596 SBLINKRECORD(sb, m0);
597 sb->sb_mbtail = m0;
598 m = m0->m_next;
599 m0->m_next = 0;
600 if (m && (m0->m_flags & M_EOR)) {
601 m0->m_flags &= ~M_EOR;
602 m->m_flags |= M_EOR;
603 }
604 /* always call sbcompress() so it can do SBLASTMBUFCHK() */
605 sbcompress(sb, m, m0);
606}
607
608/*
609 * As above, except the mbuf chain begins a new record.
610 */
611void
612sbappendrecord(struct sockbuf *sb, struct mbuf *m0)
613{
614
615 SOCKBUF_LOCK(sb);
616 sbappendrecord_locked(sb, m0);
617 SOCKBUF_UNLOCK(sb);
618}
619
620/*
621 * Append address and data, and optionally, control (ancillary) data to the
622 * receive queue of a socket. If present, m0 must include a packet header
623 * with total length. Returns 0 if no space in sockbuf or insufficient
624 * mbufs.
625 */
626int
627sbappendaddr_locked(struct sockbuf *sb, const struct sockaddr *asa,
628 struct mbuf *m0, struct mbuf *control)
629{
630 struct mbuf *m, *n, *nlast;
631 int space = asa->sa_len;
632
633 SOCKBUF_LOCK_ASSERT(sb);
634
635 if (m0 && (m0->m_flags & M_PKTHDR) == 0)
636 panic("sbappendaddr_locked");
637 if (m0)
638 space += m0->m_pkthdr.len;
639 space += m_length(control, &n);
640
641 if (space > sbspace(sb))
642 return (0);
643#if MSIZE <= 256
644 if (asa->sa_len > MLEN)
645 return (0);
646#endif
647 MGET(m, M_DONTWAIT, MT_SONAME);
648 if (m == 0)
649 return (0);
650 m->m_len = asa->sa_len;
651 bcopy(asa, mtod(m, caddr_t), asa->sa_len);
652 if (n)
653 n->m_next = m0; /* concatenate data to control */
654 else
655 control = m0;
656 m->m_next = control;
657 for (n = m; n->m_next != NULL; n = n->m_next)
658 sballoc(sb, n);
659 sballoc(sb, n);
660 nlast = n;
661 SBLINKRECORD(sb, m);
662
663 sb->sb_mbtail = nlast;
664 SBLASTMBUFCHK(sb);
665
666 SBLASTRECORDCHK(sb);
667 return (1);
668}
669
670/*
671 * Append address and data, and optionally, control (ancillary) data to the
672 * receive queue of a socket. If present, m0 must include a packet header
673 * with total length. Returns 0 if no space in sockbuf or insufficient
674 * mbufs.
675 */
676int
677sbappendaddr(struct sockbuf *sb, const struct sockaddr *asa,
678 struct mbuf *m0, struct mbuf *control)
679{
680 int retval;
681
682 SOCKBUF_LOCK(sb);
683 retval = sbappendaddr_locked(sb, asa, m0, control);
684 SOCKBUF_UNLOCK(sb);
685 return (retval);
686}
687
688int
689sbappendcontrol_locked(struct sockbuf *sb, struct mbuf *m0,
690 struct mbuf *control)
691{
692 struct mbuf *m, *n, *mlast;
693 int space;
694
695 SOCKBUF_LOCK_ASSERT(sb);
696
697 if (control == 0)
698 panic("sbappendcontrol_locked");
699 space = m_length(control, &n) + m_length(m0, NULL);
700
701 if (space > sbspace(sb))
702 return (0);
703 n->m_next = m0; /* concatenate data to control */
704
705 SBLASTRECORDCHK(sb);
706
707 for (m = control; m->m_next; m = m->m_next)
708 sballoc(sb, m);
709 sballoc(sb, m);
710 mlast = m;
711 SBLINKRECORD(sb, control);
712
713 sb->sb_mbtail = mlast;
714 SBLASTMBUFCHK(sb);
715
716 SBLASTRECORDCHK(sb);
717 return (1);
718}
719
720int
721sbappendcontrol(struct sockbuf *sb, struct mbuf *m0, struct mbuf *control)
722{
723 int retval;
724
725 SOCKBUF_LOCK(sb);
726 retval = sbappendcontrol_locked(sb, m0, control);
727 SOCKBUF_UNLOCK(sb);
728 return (retval);
729}
730
731/*
732 * Append the data in mbuf chain (m) into the socket buffer sb following mbuf
733 * (n). If (n) is NULL, the buffer is presumed empty.
734 *
735 * When the data is compressed, mbufs in the chain may be handled in one of
736 * three ways:
737 *
738 * (1) The mbuf may simply be dropped, if it contributes nothing (no data, no
739 * record boundary, and no change in data type).
740 *
741 * (2) The mbuf may be coalesced -- i.e., data in the mbuf may be copied into
742 * an mbuf already in the socket buffer. This can occur if an
743 * appropriate mbuf exists, there is room, and no merging of data types
744 * will occur.
745 *
746 * (3) The mbuf may be appended to the end of the existing mbuf chain.
747 *
748 * If any of the new mbufs is marked as M_EOR, mark the last mbuf appended as
749 * end-of-record.
750 */
751void
752sbcompress(struct sockbuf *sb, struct mbuf *m, struct mbuf *n)
753{
754 int eor = 0;
755 struct mbuf *o;
756
757 SOCKBUF_LOCK_ASSERT(sb);
758
759 while (m) {
760 eor |= m->m_flags & M_EOR;
761 if (m->m_len == 0 &&
762 (eor == 0 ||
763 (((o = m->m_next) || (o = n)) &&
764 o->m_type == m->m_type))) {
765 if (sb->sb_lastrecord == m)
766 sb->sb_lastrecord = m->m_next;
767 m = m_free(m);
768 continue;
769 }
770 if (n && (n->m_flags & M_EOR) == 0 &&
771 M_WRITABLE(n) &&
772 ((sb->sb_flags & SB_NOCOALESCE) == 0) &&
773 m->m_len <= MCLBYTES / 4 && /* XXX: Don't copy too much */
774 m->m_len <= M_TRAILINGSPACE(n) &&
775 n->m_type == m->m_type) {
776 bcopy(mtod(m, caddr_t), mtod(n, caddr_t) + n->m_len,
777 (unsigned)m->m_len);
778 n->m_len += m->m_len;
779 sb->sb_cc += m->m_len;
780 if (m->m_type != MT_DATA && m->m_type != MT_OOBDATA)
781 /* XXX: Probably don't need.*/
782 sb->sb_ctl += m->m_len;
783 m = m_free(m);
784 continue;
785 }
786 if (n)
787 n->m_next = m;
788 else
789 sb->sb_mb = m;
790 sb->sb_mbtail = m;
791 sballoc(sb, m);
792 n = m;
793 m->m_flags &= ~M_EOR;
794 m = m->m_next;
795 n->m_next = 0;
796 }
797 if (eor) {
798 KASSERT(n != NULL, ("sbcompress: eor && n == NULL"));
799 n->m_flags |= eor;
800 }
801 SBLASTMBUFCHK(sb);
802}
803
804/*
805 * Free all mbufs in a sockbuf. Check that all resources are reclaimed.
806 */
807static void
808sbflush_internal(struct sockbuf *sb)
809{
810
811 while (sb->sb_mbcnt) {
812 /*
813 * Don't call sbdrop(sb, 0) if the leading mbuf is non-empty:
814 * we would loop forever. Panic instead.
815 */
816 if (!sb->sb_cc && (sb->sb_mb == NULL || sb->sb_mb->m_len))
817 break;
818 sbdrop_internal(sb, (int)sb->sb_cc);
819 }
820 if (sb->sb_cc || sb->sb_mb || sb->sb_mbcnt)
821 panic("sbflush_internal: cc %u || mb %p || mbcnt %u",
822 sb->sb_cc, (void *)sb->sb_mb, sb->sb_mbcnt);
823}
824
825void
826sbflush_locked(struct sockbuf *sb)
827{
828
829 SOCKBUF_LOCK_ASSERT(sb);
830 sbflush_internal(sb);
831}
832
833void
834sbflush(struct sockbuf *sb)
835{
836
837 SOCKBUF_LOCK(sb);
838 sbflush_locked(sb);
839 SOCKBUF_UNLOCK(sb);
840}
841
842/*
843 * Drop data from (the front of) a sockbuf.
844 */
845static void
846sbdrop_internal(struct sockbuf *sb, int len)
847{
848 struct mbuf *m;
849 struct mbuf *next;
850
851 next = (m = sb->sb_mb) ? m->m_nextpkt : 0;
852 while (len > 0) {
853 if (m == 0) {
854 if (next == 0)
855 panic("sbdrop");
856 m = next;
857 next = m->m_nextpkt;
858 continue;
859 }
860 if (m->m_len > len) {
861 m->m_len -= len;
862 m->m_data += len;
863 sb->sb_cc -= len;
864 if (sb->sb_sndptroff != 0)
865 sb->sb_sndptroff -= len;
866 if (m->m_type != MT_DATA && m->m_type != MT_OOBDATA)
867 sb->sb_ctl -= len;
868 break;
869 }
870 len -= m->m_len;
871 sbfree(sb, m);
872 m = m_free(m);
873 }
874 while (m && m->m_len == 0) {
875 sbfree(sb, m);
876 m = m_free(m);
877 }
878 if (m) {
879 sb->sb_mb = m;
880 m->m_nextpkt = next;
881 } else
882 sb->sb_mb = next;
883 /*
884 * First part is an inline SB_EMPTY_FIXUP(). Second part makes sure
885 * sb_lastrecord is up-to-date if we dropped part of the last record.
886 */
887 m = sb->sb_mb;
888 if (m == NULL) {
889 sb->sb_mbtail = NULL;
890 sb->sb_lastrecord = NULL;
891 } else if (m->m_nextpkt == NULL) {
892 sb->sb_lastrecord = m;
893 }
894}
895
896/*
897 * Drop data from (the front of) a sockbuf.
898 */
899void
900sbdrop_locked(struct sockbuf *sb, int len)
901{
902
903 SOCKBUF_LOCK_ASSERT(sb);
904
905 sbdrop_internal(sb, len);
906}
907
908void
909sbdrop(struct sockbuf *sb, int len)
910{
911
912 SOCKBUF_LOCK(sb);
913 sbdrop_locked(sb, len);
914 SOCKBUF_UNLOCK(sb);
915}
916
917/*
918 * Maintain a pointer and offset pair into the socket buffer mbuf chain to
919 * avoid traversal of the entire socket buffer for larger offsets.
920 */
921struct mbuf *
922sbsndptr(struct sockbuf *sb, u_int off, u_int len, u_int *moff)
923{
924 struct mbuf *m, *ret;
925
926 KASSERT(sb->sb_mb != NULL, ("%s: sb_mb is NULL", __func__));
927 KASSERT(off + len <= sb->sb_cc, ("%s: beyond sb", __func__));
928 KASSERT(sb->sb_sndptroff <= sb->sb_cc, ("%s: sndptroff broken", __func__));
929
930 /*
931 * Is off below stored offset? Happens on retransmits.
932 * Just return, we can't help here.
933 */
934 if (sb->sb_sndptroff > off) {
935 *moff = off;
936 return (sb->sb_mb);
937 }
938
939 /* Return closest mbuf in chain for current offset. */
940 *moff = off - sb->sb_sndptroff;
941 m = ret = sb->sb_sndptr ? sb->sb_sndptr : sb->sb_mb;
942
943 /* Advance by len to be as close as possible for the next transmit. */
944 for (off = off - sb->sb_sndptroff + len - 1;
945 off > 0 && m != NULL && off >= m->m_len;
946 m = m->m_next) {
947 sb->sb_sndptroff += m->m_len;
948 off -= m->m_len;
949 }
950 if (off > 0 && m == NULL)
951 panic("%s: sockbuf %p and mbuf %p clashing", __func__, sb, ret);
952 sb->sb_sndptr = m;
953
954 return (ret);
955}
956
957/*
958 * Drop a record off the front of a sockbuf and move the next record to the
959 * front.
960 */
961void
962sbdroprecord_locked(struct sockbuf *sb)
963{
964 struct mbuf *m;
965
966 SOCKBUF_LOCK_ASSERT(sb);
967
968 m = sb->sb_mb;
969 if (m) {
970 sb->sb_mb = m->m_nextpkt;
971 do {
972 sbfree(sb, m);
973 m = m_free(m);
974 } while (m);
975 }
976 SB_EMPTY_FIXUP(sb);
977}
978
979/*
980 * Drop a record off the front of a sockbuf and move the next record to the
981 * front.
982 */
983void
984sbdroprecord(struct sockbuf *sb)
985{
986
987 SOCKBUF_LOCK(sb);
988 sbdroprecord_locked(sb);
989 SOCKBUF_UNLOCK(sb);
990}
991
992/*
993 * Create a "control" mbuf containing the specified data with the specified
994 * type for presentation on a socket buffer.
995 */
996struct mbuf *
997sbcreatecontrol(caddr_t p, int size, int type, int level)
998{
999 struct cmsghdr *cp;
1000 struct mbuf *m;
1001
1002 if (CMSG_SPACE((u_int)size) > MCLBYTES)
1003 return ((struct mbuf *) NULL);
1004 if (CMSG_SPACE((u_int)size) > MLEN)
1005 m = m_getcl(M_DONTWAIT, MT_CONTROL, 0);
1006 else
1007 m = m_get(M_DONTWAIT, MT_CONTROL);
1008 if (m == NULL)
1009 return ((struct mbuf *) NULL);
1010 cp = mtod(m, struct cmsghdr *);
1011 m->m_len = 0;
1012 KASSERT(CMSG_SPACE((u_int)size) <= M_TRAILINGSPACE(m),
1013 ("sbcreatecontrol: short mbuf"));
1014 if (p != NULL)
1015 (void)memcpy(CMSG_DATA(cp), p, size);
1016 m->m_len = CMSG_SPACE(size);
1017 cp->cmsg_len = CMSG_LEN(size);
1018 cp->cmsg_level = level;
1019 cp->cmsg_type = type;
1020 return (m);
1021}
1022
1023/*
1024 * This does the same for socket buffers that sotoxsocket does for sockets:
1025 * generate an user-format data structure describing the socket buffer. Note
1026 * that the xsockbuf structure, since it is always embedded in a socket, does
1027 * not include a self pointer nor a length. We make this entry point public
1028 * in case some other mechanism needs it.
1029 */
1030void
1031sbtoxsockbuf(struct sockbuf *sb, struct xsockbuf *xsb)
1032{
1033
1034 xsb->sb_cc = sb->sb_cc;
1035 xsb->sb_hiwat = sb->sb_hiwat;
1036 xsb->sb_mbcnt = sb->sb_mbcnt;
1037 xsb->sb_mcnt = sb->sb_mcnt;
1038 xsb->sb_ccnt = sb->sb_ccnt;
1039 xsb->sb_mbmax = sb->sb_mbmax;
1040 xsb->sb_lowat = sb->sb_lowat;
1041 xsb->sb_flags = sb->sb_flags;
1042 xsb->sb_timeo = sb->sb_timeo;
1043}
1044
1045/* This takes the place of kern.maxsockbuf, which moved to kern.ipc. */
1046static int dummy;
1047SYSCTL_INT(_kern, KERN_DUMMY, dummy, CTLFLAG_RW, &dummy, 0, "");
1048SYSCTL_OID(_kern_ipc, KIPC_MAXSOCKBUF, maxsockbuf, CTLTYPE_ULONG|CTLFLAG_RW,
1049 &sb_max, 0, sysctl_handle_sb_max, "LU", "Maximum socket buffer size");
1050SYSCTL_ULONG(_kern_ipc, KIPC_SOCKBUF_WASTE, sockbuf_waste_factor, CTLFLAG_RW,
|