xref: /freebsd/sys/netinet/tcp_log_buf.c (revision dd41de95a84d979615a2ef11df6850622bf6184e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2016-2018 Netflix, Inc.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  *
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/arb.h>
34 #include <sys/kernel.h>
35 #include <sys/lock.h>
36 #include <sys/malloc.h>
37 #include <sys/mutex.h>
38 #include <sys/qmath.h>
39 #include <sys/queue.h>
40 #include <sys/refcount.h>
41 #include <sys/rwlock.h>
42 #include <sys/socket.h>
43 #include <sys/socketvar.h>
44 #include <sys/sysctl.h>
45 #include <sys/tree.h>
46 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
47 #include <sys/counter.h>
48 
49 #include <dev/tcp_log/tcp_log_dev.h>
50 
51 #include <net/if.h>
52 #include <net/if_var.h>
53 #include <net/vnet.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/in_pcb.h>
57 #include <netinet/in_var.h>
58 #include <netinet/tcp_var.h>
59 #include <netinet/tcp_log_buf.h>
60 
61 /* Default expiry time */
62 #define	TCP_LOG_EXPIRE_TIME	((sbintime_t)60 * SBT_1S)
63 
64 /* Max interval at which to run the expiry timer */
65 #define	TCP_LOG_EXPIRE_INTVL	((sbintime_t)5 * SBT_1S)
66 
67 bool	tcp_log_verbose;
68 static uma_zone_t tcp_log_bucket_zone, tcp_log_node_zone, tcp_log_zone;
69 static int	tcp_log_session_limit = TCP_LOG_BUF_DEFAULT_SESSION_LIMIT;
70 static uint32_t	tcp_log_version = TCP_LOG_BUF_VER;
71 RB_HEAD(tcp_log_id_tree, tcp_log_id_bucket);
72 static struct tcp_log_id_tree tcp_log_id_head;
73 static STAILQ_HEAD(, tcp_log_id_node) tcp_log_expireq_head =
74     STAILQ_HEAD_INITIALIZER(tcp_log_expireq_head);
75 static struct mtx tcp_log_expireq_mtx;
76 static struct callout tcp_log_expireq_callout;
77 static u_long tcp_log_auto_ratio = 0;
78 static volatile u_long tcp_log_auto_ratio_cur = 0;
79 static uint32_t tcp_log_auto_mode = TCP_LOG_STATE_TAIL;
80 static bool tcp_log_auto_all = false;
81 static uint32_t tcp_disable_all_bb_logs = 0;
82 
83 RB_PROTOTYPE_STATIC(tcp_log_id_tree, tcp_log_id_bucket, tlb_rb, tcp_log_id_cmp)
84 
85 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, bb, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
86     "TCP Black Box controls");
87 
88 SYSCTL_BOOL(_net_inet_tcp_bb, OID_AUTO, log_verbose, CTLFLAG_RW, &tcp_log_verbose,
89     0, "Force verbose logging for TCP traces");
90 
91 SYSCTL_INT(_net_inet_tcp_bb, OID_AUTO, log_session_limit,
92     CTLFLAG_RW, &tcp_log_session_limit, 0,
93     "Maximum number of events maintained for each TCP session");
94 
95 SYSCTL_UMA_MAX(_net_inet_tcp_bb, OID_AUTO, log_global_limit, CTLFLAG_RW,
96     &tcp_log_zone, "Maximum number of events maintained for all TCP sessions");
97 
98 SYSCTL_UMA_CUR(_net_inet_tcp_bb, OID_AUTO, log_global_entries, CTLFLAG_RD,
99     &tcp_log_zone, "Current number of events maintained for all TCP sessions");
100 
101 SYSCTL_UMA_MAX(_net_inet_tcp_bb, OID_AUTO, log_id_limit, CTLFLAG_RW,
102     &tcp_log_bucket_zone, "Maximum number of log IDs");
103 
104 SYSCTL_UMA_CUR(_net_inet_tcp_bb, OID_AUTO, log_id_entries, CTLFLAG_RD,
105     &tcp_log_bucket_zone, "Current number of log IDs");
106 
107 SYSCTL_UMA_MAX(_net_inet_tcp_bb, OID_AUTO, log_id_tcpcb_limit, CTLFLAG_RW,
108     &tcp_log_node_zone, "Maximum number of tcpcbs with log IDs");
109 
110 SYSCTL_UMA_CUR(_net_inet_tcp_bb, OID_AUTO, log_id_tcpcb_entries, CTLFLAG_RD,
111     &tcp_log_node_zone, "Current number of tcpcbs with log IDs");
112 
113 SYSCTL_U32(_net_inet_tcp_bb, OID_AUTO, log_version, CTLFLAG_RD, &tcp_log_version,
114     0, "Version of log formats exported");
115 
116 SYSCTL_U32(_net_inet_tcp_bb, OID_AUTO, disable_all, CTLFLAG_RW,
117     &tcp_disable_all_bb_logs, TCP_LOG_STATE_HEAD_AUTO,
118     "Disable all BB logging for all connections");
119 
120 SYSCTL_ULONG(_net_inet_tcp_bb, OID_AUTO, log_auto_ratio, CTLFLAG_RW,
121     &tcp_log_auto_ratio, 0, "Do auto capturing for 1 out of N sessions");
122 
123 SYSCTL_U32(_net_inet_tcp_bb, OID_AUTO, log_auto_mode, CTLFLAG_RW,
124     &tcp_log_auto_mode, TCP_LOG_STATE_HEAD_AUTO,
125     "Logging mode for auto-selected sessions (default is TCP_LOG_STATE_HEAD_AUTO)");
126 
127 SYSCTL_BOOL(_net_inet_tcp_bb, OID_AUTO, log_auto_all, CTLFLAG_RW,
128     &tcp_log_auto_all, false,
129     "Auto-select from all sessions (rather than just those with IDs)");
130 
131 #ifdef TCPLOG_DEBUG_COUNTERS
132 counter_u64_t tcp_log_queued;
133 counter_u64_t tcp_log_que_fail1;
134 counter_u64_t tcp_log_que_fail2;
135 counter_u64_t tcp_log_que_fail3;
136 counter_u64_t tcp_log_que_fail4;
137 counter_u64_t tcp_log_que_fail5;
138 counter_u64_t tcp_log_que_copyout;
139 counter_u64_t tcp_log_que_read;
140 counter_u64_t tcp_log_que_freed;
141 
142 SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, queued, CTLFLAG_RD,
143     &tcp_log_queued, "Number of entries queued");
144 SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, fail1, CTLFLAG_RD,
145     &tcp_log_que_fail1, "Number of entries queued but fail 1");
146 SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, fail2, CTLFLAG_RD,
147     &tcp_log_que_fail2, "Number of entries queued but fail 2");
148 SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, fail3, CTLFLAG_RD,
149     &tcp_log_que_fail3, "Number of entries queued but fail 3");
150 SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, fail4, CTLFLAG_RD,
151     &tcp_log_que_fail4, "Number of entries queued but fail 4");
152 SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, fail5, CTLFLAG_RD,
153     &tcp_log_que_fail5, "Number of entries queued but fail 4");
154 SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, copyout, CTLFLAG_RD,
155     &tcp_log_que_copyout, "Number of entries copied out");
156 SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, read, CTLFLAG_RD,
157     &tcp_log_que_read, "Number of entries read from the queue");
158 SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, freed, CTLFLAG_RD,
159     &tcp_log_que_freed, "Number of entries freed after reading");
160 #endif
161 
162 #ifdef INVARIANTS
163 #define	TCPLOG_DEBUG_RINGBUF
164 #endif
165 /* Number of requests to consider a PBCID "active". */
166 #define	ACTIVE_REQUEST_COUNT	10
167 
168 /* Statistic tracking for "active" PBCIDs. */
169 static counter_u64_t tcp_log_pcb_ids_cur;
170 static counter_u64_t tcp_log_pcb_ids_tot;
171 
172 SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, pcb_ids_cur, CTLFLAG_RD,
173     &tcp_log_pcb_ids_cur, "Number of pcb IDs allocated in the system");
174 SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, pcb_ids_tot, CTLFLAG_RD,
175     &tcp_log_pcb_ids_tot, "Total number of pcb IDs that have been allocated");
176 
177 struct tcp_log_mem
178 {
179 	STAILQ_ENTRY(tcp_log_mem) tlm_queue;
180 	struct tcp_log_buffer	tlm_buf;
181 	struct tcp_log_verbose	tlm_v;
182 #ifdef TCPLOG_DEBUG_RINGBUF
183 	volatile int		tlm_refcnt;
184 #endif
185 };
186 
187 /* 60 bytes for the header, + 16 bytes for padding */
188 static uint8_t	zerobuf[76];
189 
190 /*
191  * Lock order:
192  * 1. TCPID_TREE
193  * 2. TCPID_BUCKET
194  * 3. INP
195  *
196  * Rules:
197  * A. You need a lock on the Tree to add/remove buckets.
198  * B. You need a lock on the bucket to add/remove nodes from the bucket.
199  * C. To change information in a node, you need the INP lock if the tln_closed
200  *    field is false. Otherwise, you need the bucket lock. (Note that the
201  *    tln_closed field can change at any point, so you need to recheck the
202  *    entry after acquiring the INP lock.)
203  * D. To remove a node from the bucket, you must have that entry locked,
204  *    according to the criteria of Rule C. Also, the node must not be on
205  *    the expiry queue.
206  * E. The exception to C is the expiry queue fields, which are locked by
207  *    the TCPLOG_EXPIREQ lock.
208  *
209  * Buckets have a reference count. Each node is a reference. Further,
210  * other callers may add reference counts to keep a bucket from disappearing.
211  * You can add a reference as long as you own a lock sufficient to keep the
212  * bucket from disappearing. For example, a common use is:
213  *   a. Have a locked INP, but need to lock the TCPID_BUCKET.
214  *   b. Add a refcount on the bucket. (Safe because the INP lock prevents
215  *      the TCPID_BUCKET from going away.)
216  *   c. Drop the INP lock.
217  *   d. Acquire a lock on the TCPID_BUCKET.
218  *   e. Acquire a lock on the INP.
219  *   f. Drop the refcount on the bucket.
220  *      (At this point, the bucket may disappear.)
221  *
222  * Expire queue lock:
223  * You can acquire this with either the bucket or INP lock. Don't reverse it.
224  * When the expire code has committed to freeing a node, it resets the expiry
225  * time to SBT_MAX. That is the signal to everyone else that they should
226  * leave that node alone.
227  */
228 static struct rwlock tcp_id_tree_lock;
229 #define	TCPID_TREE_WLOCK()		rw_wlock(&tcp_id_tree_lock)
230 #define	TCPID_TREE_RLOCK()		rw_rlock(&tcp_id_tree_lock)
231 #define	TCPID_TREE_UPGRADE()		rw_try_upgrade(&tcp_id_tree_lock)
232 #define	TCPID_TREE_WUNLOCK()		rw_wunlock(&tcp_id_tree_lock)
233 #define	TCPID_TREE_RUNLOCK()		rw_runlock(&tcp_id_tree_lock)
234 #define	TCPID_TREE_WLOCK_ASSERT()	rw_assert(&tcp_id_tree_lock, RA_WLOCKED)
235 #define	TCPID_TREE_RLOCK_ASSERT()	rw_assert(&tcp_id_tree_lock, RA_RLOCKED)
236 #define	TCPID_TREE_UNLOCK_ASSERT()	rw_assert(&tcp_id_tree_lock, RA_UNLOCKED)
237 
238 #define	TCPID_BUCKET_LOCK_INIT(tlb)	mtx_init(&((tlb)->tlb_mtx), "tcp log id bucket", NULL, MTX_DEF)
239 #define	TCPID_BUCKET_LOCK_DESTROY(tlb)	mtx_destroy(&((tlb)->tlb_mtx))
240 #define	TCPID_BUCKET_LOCK(tlb)		mtx_lock(&((tlb)->tlb_mtx))
241 #define	TCPID_BUCKET_UNLOCK(tlb)	mtx_unlock(&((tlb)->tlb_mtx))
242 #define	TCPID_BUCKET_LOCK_ASSERT(tlb)	mtx_assert(&((tlb)->tlb_mtx), MA_OWNED)
243 #define	TCPID_BUCKET_UNLOCK_ASSERT(tlb) mtx_assert(&((tlb)->tlb_mtx), MA_NOTOWNED)
244 
245 #define	TCPID_BUCKET_REF(tlb)		refcount_acquire(&((tlb)->tlb_refcnt))
246 #define	TCPID_BUCKET_UNREF(tlb)		refcount_release(&((tlb)->tlb_refcnt))
247 
248 #define	TCPLOG_EXPIREQ_LOCK()		mtx_lock(&tcp_log_expireq_mtx)
249 #define	TCPLOG_EXPIREQ_UNLOCK()		mtx_unlock(&tcp_log_expireq_mtx)
250 
251 SLIST_HEAD(tcp_log_id_head, tcp_log_id_node);
252 
253 struct tcp_log_id_bucket
254 {
255 	/*
256 	 * tlb_id must be first. This lets us use strcmp on
257 	 * (struct tcp_log_id_bucket *) and (char *) interchangeably.
258 	 */
259 	char				tlb_id[TCP_LOG_ID_LEN];
260 	char				tlb_tag[TCP_LOG_TAG_LEN];
261 	RB_ENTRY(tcp_log_id_bucket)	tlb_rb;
262 	struct tcp_log_id_head		tlb_head;
263 	struct mtx			tlb_mtx;
264 	volatile u_int			tlb_refcnt;
265 	volatile u_int			tlb_reqcnt;
266 	uint32_t			tlb_loglimit;
267 	uint8_t				tlb_logstate;
268 };
269 
270 struct tcp_log_id_node
271 {
272 	SLIST_ENTRY(tcp_log_id_node) tln_list;
273 	STAILQ_ENTRY(tcp_log_id_node) tln_expireq; /* Locked by the expireq lock */
274 	sbintime_t		tln_expiretime;	/* Locked by the expireq lock */
275 
276 	/*
277 	 * If INP is NULL, that means the connection has closed. We've
278 	 * saved the connection endpoint information and the log entries
279 	 * in the tln_ie and tln_entries members. We've also saved a pointer
280 	 * to the enclosing bucket here. If INP is not NULL, the information is
281 	 * in the PCB and not here.
282 	 */
283 	struct inpcb		*tln_inp;
284 	struct tcpcb		*tln_tp;
285 	struct tcp_log_id_bucket *tln_bucket;
286 	struct in_endpoints	tln_ie;
287 	struct tcp_log_stailq	tln_entries;
288 	int			tln_count;
289 	volatile int		tln_closed;
290 	uint8_t			tln_af;
291 };
292 
293 enum tree_lock_state {
294 	TREE_UNLOCKED = 0,
295 	TREE_RLOCKED,
296 	TREE_WLOCKED,
297 };
298 
299 /* Do we want to select this session for auto-logging? */
300 static __inline bool
301 tcp_log_selectauto(void)
302 {
303 
304 	/*
305 	 * If we are doing auto-capturing, figure out whether we will capture
306 	 * this session.
307 	 */
308 	if (tcp_log_auto_ratio &&
309 	    (tcp_disable_all_bb_logs == 0) &&
310 	    (atomic_fetchadd_long(&tcp_log_auto_ratio_cur, 1) %
311 	    tcp_log_auto_ratio) == 0)
312 		return (true);
313 	return (false);
314 }
315 
316 static __inline int
317 tcp_log_id_cmp(struct tcp_log_id_bucket *a, struct tcp_log_id_bucket *b)
318 {
319 	KASSERT(a != NULL, ("tcp_log_id_cmp: argument a is unexpectedly NULL"));
320 	KASSERT(b != NULL, ("tcp_log_id_cmp: argument b is unexpectedly NULL"));
321 	return strncmp(a->tlb_id, b->tlb_id, TCP_LOG_ID_LEN);
322 }
323 
324 RB_GENERATE_STATIC(tcp_log_id_tree, tcp_log_id_bucket, tlb_rb, tcp_log_id_cmp)
325 
326 static __inline void
327 tcp_log_id_validate_tree_lock(int tree_locked)
328 {
329 
330 #ifdef INVARIANTS
331 	switch (tree_locked) {
332 	case TREE_WLOCKED:
333 		TCPID_TREE_WLOCK_ASSERT();
334 		break;
335 	case TREE_RLOCKED:
336 		TCPID_TREE_RLOCK_ASSERT();
337 		break;
338 	case TREE_UNLOCKED:
339 		TCPID_TREE_UNLOCK_ASSERT();
340 		break;
341 	default:
342 		kassert_panic("%s:%d: unknown tree lock state", __func__,
343 		    __LINE__);
344 	}
345 #endif
346 }
347 
348 static __inline void
349 tcp_log_remove_bucket(struct tcp_log_id_bucket *tlb)
350 {
351 
352 	TCPID_TREE_WLOCK_ASSERT();
353 	KASSERT(SLIST_EMPTY(&tlb->tlb_head),
354 	    ("%s: Attempt to remove non-empty bucket", __func__));
355 	if (RB_REMOVE(tcp_log_id_tree, &tcp_log_id_head, tlb) == NULL) {
356 #ifdef INVARIANTS
357 		kassert_panic("%s:%d: error removing element from tree",
358 			    __func__, __LINE__);
359 #endif
360 	}
361 	TCPID_BUCKET_LOCK_DESTROY(tlb);
362 	counter_u64_add(tcp_log_pcb_ids_cur, (int64_t)-1);
363 	uma_zfree(tcp_log_bucket_zone, tlb);
364 }
365 
366 /*
367  * Call with a referenced and locked bucket.
368  * Will return true if the bucket was freed; otherwise, false.
369  * tlb: The bucket to unreference.
370  * tree_locked: A pointer to the state of the tree lock. If the tree lock
371  *    state changes, the function will update it.
372  * inp: If not NULL and the function needs to drop the inp lock to relock the
373  *    tree, it will do so. (The caller must ensure inp will not become invalid,
374  *    probably by holding a reference to it.)
375  */
376 static bool
377 tcp_log_unref_bucket(struct tcp_log_id_bucket *tlb, int *tree_locked,
378     struct inpcb *inp)
379 {
380 
381 	KASSERT(tlb != NULL, ("%s: called with NULL tlb", __func__));
382 	KASSERT(tree_locked != NULL, ("%s: called with NULL tree_locked",
383 	    __func__));
384 
385 	tcp_log_id_validate_tree_lock(*tree_locked);
386 
387 	/*
388 	 * Did we hold the last reference on the tlb? If so, we may need
389 	 * to free it. (Note that we can realistically only execute the
390 	 * loop twice: once without a write lock and once with a write
391 	 * lock.)
392 	 */
393 	while (TCPID_BUCKET_UNREF(tlb)) {
394 		/*
395 		 * We need a write lock on the tree to free this.
396 		 * If we can upgrade the tree lock, this is "easy". If we
397 		 * can't upgrade the tree lock, we need to do this the
398 		 * "hard" way: unwind all our locks and relock everything.
399 		 * In the meantime, anything could have changed. We even
400 		 * need to validate that we still need to free the bucket.
401 		 */
402 		if (*tree_locked == TREE_RLOCKED && TCPID_TREE_UPGRADE())
403 			*tree_locked = TREE_WLOCKED;
404 		else if (*tree_locked != TREE_WLOCKED) {
405 			TCPID_BUCKET_REF(tlb);
406 			if (inp != NULL)
407 				INP_WUNLOCK(inp);
408 			TCPID_BUCKET_UNLOCK(tlb);
409 			if (*tree_locked == TREE_RLOCKED)
410 				TCPID_TREE_RUNLOCK();
411 			TCPID_TREE_WLOCK();
412 			*tree_locked = TREE_WLOCKED;
413 			TCPID_BUCKET_LOCK(tlb);
414 			if (inp != NULL)
415 				INP_WLOCK(inp);
416 			continue;
417 		}
418 
419 		/*
420 		 * We have an empty bucket and a write lock on the tree.
421 		 * Remove the empty bucket.
422 		 */
423 		tcp_log_remove_bucket(tlb);
424 		return (true);
425 	}
426 	return (false);
427 }
428 
429 /*
430  * Call with a locked bucket. This function will release the lock on the
431  * bucket before returning.
432  *
433  * The caller is responsible for freeing the tp->t_lin/tln node!
434  *
435  * Note: one of tp or both tlb and tln must be supplied.
436  *
437  * inp: A pointer to the inp. If the function needs to drop the inp lock to
438  *    acquire the tree write lock, it will do so. (The caller must ensure inp
439  *    will not become invalid, probably by holding a reference to it.)
440  * tp: A pointer to the tcpcb. (optional; if specified, tlb and tln are ignored)
441  * tlb: A pointer to the bucket. (optional; ignored if tp is specified)
442  * tln: A pointer to the node. (optional; ignored if tp is specified)
443  * tree_locked: A pointer to the state of the tree lock. If the tree lock
444  *    state changes, the function will update it.
445  *
446  * Will return true if the INP lock was reacquired; otherwise, false.
447  */
448 static bool
449 tcp_log_remove_id_node(struct inpcb *inp, struct tcpcb *tp,
450     struct tcp_log_id_bucket *tlb, struct tcp_log_id_node *tln,
451     int *tree_locked)
452 {
453 	int orig_tree_locked;
454 
455 	KASSERT(tp != NULL || (tlb != NULL && tln != NULL),
456 	    ("%s: called with tp=%p, tlb=%p, tln=%p", __func__,
457 	    tp, tlb, tln));
458 	KASSERT(tree_locked != NULL, ("%s: called with NULL tree_locked",
459 	    __func__));
460 
461 	if (tp != NULL) {
462 		tlb = tp->t_lib;
463 		tln = tp->t_lin;
464 		KASSERT(tlb != NULL, ("%s: unexpectedly NULL tlb", __func__));
465 		KASSERT(tln != NULL, ("%s: unexpectedly NULL tln", __func__));
466 	}
467 
468 	tcp_log_id_validate_tree_lock(*tree_locked);
469 	TCPID_BUCKET_LOCK_ASSERT(tlb);
470 
471 	/*
472 	 * Remove the node, clear the log bucket and node from the TCPCB, and
473 	 * decrement the bucket refcount. In the process, if this is the
474 	 * last reference, the bucket will be freed.
475 	 */
476 	SLIST_REMOVE(&tlb->tlb_head, tln, tcp_log_id_node, tln_list);
477 	if (tp != NULL) {
478 		tp->t_lib = NULL;
479 		tp->t_lin = NULL;
480 	}
481 	orig_tree_locked = *tree_locked;
482 	if (!tcp_log_unref_bucket(tlb, tree_locked, inp))
483 		TCPID_BUCKET_UNLOCK(tlb);
484 	return (*tree_locked != orig_tree_locked);
485 }
486 
487 #define	RECHECK_INP_CLEAN(cleanup)	do {			\
488 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {	\
489 		rv = ECONNRESET;				\
490 		cleanup;					\
491 		goto done;					\
492 	}							\
493 	tp = intotcpcb(inp);					\
494 } while (0)
495 
496 #define	RECHECK_INP()	RECHECK_INP_CLEAN(/* noop */)
497 
498 static void
499 tcp_log_grow_tlb(char *tlb_id, struct tcpcb *tp)
500 {
501 
502 	INP_WLOCK_ASSERT(tp->t_inpcb);
503 
504 #ifdef STATS
505 	if (V_tcp_perconn_stats_enable == 2 && tp->t_stats == NULL)
506 		(void)tcp_stats_sample_rollthedice(tp, tlb_id, strlen(tlb_id));
507 #endif
508 }
509 
510 static void
511 tcp_log_increment_reqcnt(struct tcp_log_id_bucket *tlb)
512 {
513 
514 	atomic_fetchadd_int(&tlb->tlb_reqcnt, 1);
515 }
516 
517 /*
518  * Associate the specified tag with a particular TCP log ID.
519  * Called with INPCB locked. Returns with it unlocked.
520  * Returns 0 on success or EOPNOTSUPP if the connection has no TCP log ID.
521  */
522 int
523 tcp_log_set_tag(struct tcpcb *tp, char *tag)
524 {
525 	struct tcp_log_id_bucket *tlb;
526 	int tree_locked;
527 
528 	INP_WLOCK_ASSERT(tp->t_inpcb);
529 
530 	tree_locked = TREE_UNLOCKED;
531 	tlb = tp->t_lib;
532 	if (tlb == NULL) {
533 		INP_WUNLOCK(tp->t_inpcb);
534 		return (EOPNOTSUPP);
535 	}
536 
537 	TCPID_BUCKET_REF(tlb);
538 	INP_WUNLOCK(tp->t_inpcb);
539 	TCPID_BUCKET_LOCK(tlb);
540 	strlcpy(tlb->tlb_tag, tag, TCP_LOG_TAG_LEN);
541 	if (!tcp_log_unref_bucket(tlb, &tree_locked, NULL))
542 		TCPID_BUCKET_UNLOCK(tlb);
543 
544 	if (tree_locked == TREE_WLOCKED) {
545 		TCPID_TREE_WLOCK_ASSERT();
546 		TCPID_TREE_WUNLOCK();
547 	} else if (tree_locked == TREE_RLOCKED) {
548 		TCPID_TREE_RLOCK_ASSERT();
549 		TCPID_TREE_RUNLOCK();
550 	} else
551 		TCPID_TREE_UNLOCK_ASSERT();
552 
553 	return (0);
554 }
555 
556 /*
557  * Set the TCP log ID for a TCPCB.
558  * Called with INPCB locked. Returns with it unlocked.
559  */
560 int
561 tcp_log_set_id(struct tcpcb *tp, char *id)
562 {
563 	struct tcp_log_id_bucket *tlb, *tmp_tlb;
564 	struct tcp_log_id_node *tln;
565 	struct inpcb *inp;
566 	int tree_locked, rv;
567 	bool bucket_locked;
568 
569 	tlb = NULL;
570 	tln = NULL;
571 	inp = tp->t_inpcb;
572 	tree_locked = TREE_UNLOCKED;
573 	bucket_locked = false;
574 
575 restart:
576 	INP_WLOCK_ASSERT(inp);
577 
578 	/* See if the ID is unchanged. */
579 	if ((tp->t_lib != NULL && !strcmp(tp->t_lib->tlb_id, id)) ||
580 	    (tp->t_lib == NULL && *id == 0)) {
581 		if (tp->t_lib != NULL) {
582 			tcp_log_increment_reqcnt(tp->t_lib);
583 			if ((tp->t_lib->tlb_logstate) &&
584 			    (tp->t_log_state_set == 0)) {
585 				/* Clone in any logging */
586 
587 				tp->t_logstate = tp->t_lib->tlb_logstate;
588 			}
589 			if ((tp->t_lib->tlb_loglimit) &&
590 			    (tp->t_log_state_set == 0)) {
591 				/* We also have a limit set */
592 
593 				tp->t_loglimit = tp->t_lib->tlb_loglimit;
594 			}
595 		}
596 		rv = 0;
597 		goto done;
598 	}
599 
600 	/*
601 	 * If the TCPCB had a previous ID, we need to extricate it from
602 	 * the previous list.
603 	 *
604 	 * Drop the TCPCB lock and lock the tree and the bucket.
605 	 * Because this is called in the socket context, we (theoretically)
606 	 * don't need to worry about the INPCB completely going away
607 	 * while we are gone.
608 	 */
609 	if (tp->t_lib != NULL) {
610 		tlb = tp->t_lib;
611 		TCPID_BUCKET_REF(tlb);
612 		INP_WUNLOCK(inp);
613 
614 		if (tree_locked == TREE_UNLOCKED) {
615 			TCPID_TREE_RLOCK();
616 			tree_locked = TREE_RLOCKED;
617 		}
618 		TCPID_BUCKET_LOCK(tlb);
619 		bucket_locked = true;
620 		INP_WLOCK(inp);
621 
622 		/*
623 		 * Unreference the bucket. If our bucket went away, it is no
624 		 * longer locked or valid.
625 		 */
626 		if (tcp_log_unref_bucket(tlb, &tree_locked, inp)) {
627 			bucket_locked = false;
628 			tlb = NULL;
629 		}
630 
631 		/* Validate the INP. */
632 		RECHECK_INP();
633 
634 		/*
635 		 * Evaluate whether the bucket changed while we were unlocked.
636 		 *
637 		 * Possible scenarios here:
638 		 * 1. Bucket is unchanged and the same one we started with.
639 		 * 2. The TCPCB no longer has a bucket and our bucket was
640 		 *    freed.
641 		 * 3. The TCPCB has a new bucket, whether ours was freed.
642 		 * 4. The TCPCB no longer has a bucket and our bucket was
643 		 *    not freed.
644 		 *
645 		 * In cases 2-4, we will start over. In case 1, we will
646 		 * proceed here to remove the bucket.
647 		 */
648 		if (tlb == NULL || tp->t_lib != tlb) {
649 			KASSERT(bucket_locked || tlb == NULL,
650 			    ("%s: bucket_locked (%d) and tlb (%p) are "
651 			    "inconsistent", __func__, bucket_locked, tlb));
652 
653 			if (bucket_locked) {
654 				TCPID_BUCKET_UNLOCK(tlb);
655 				bucket_locked = false;
656 				tlb = NULL;
657 			}
658 			goto restart;
659 		}
660 
661 		/*
662 		 * Store the (struct tcp_log_id_node) for reuse. Then, remove
663 		 * it from the bucket. In the process, we may end up relocking.
664 		 * If so, we need to validate that the INP is still valid, and
665 		 * the TCPCB entries match we expect.
666 		 *
667 		 * We will clear tlb and change the bucket_locked state just
668 		 * before calling tcp_log_remove_id_node(), since that function
669 		 * will unlock the bucket.
670 		 */
671 		if (tln != NULL)
672 			uma_zfree(tcp_log_node_zone, tln);
673 		tln = tp->t_lin;
674 		tlb = NULL;
675 		bucket_locked = false;
676 		if (tcp_log_remove_id_node(inp, tp, NULL, NULL, &tree_locked)) {
677 			RECHECK_INP();
678 
679 			/*
680 			 * If the TCPCB moved to a new bucket while we had
681 			 * dropped the lock, restart.
682 			 */
683 			if (tp->t_lib != NULL || tp->t_lin != NULL)
684 				goto restart;
685 		}
686 
687 		/*
688 		 * Yay! We successfully removed the TCPCB from its old
689 		 * bucket. Phew!
690 		 *
691 		 * On to bigger and better things...
692 		 */
693 	}
694 
695 	/* At this point, the TCPCB should not be in any bucket. */
696 	KASSERT(tp->t_lib == NULL, ("%s: tp->t_lib is not NULL", __func__));
697 
698 	/*
699 	 * If the new ID is not empty, we need to now assign this TCPCB to a
700 	 * new bucket.
701 	 */
702 	if (*id) {
703 		/* Get a new tln, if we don't already have one to reuse. */
704 		if (tln == NULL) {
705 			tln = uma_zalloc(tcp_log_node_zone, M_NOWAIT | M_ZERO);
706 			if (tln == NULL) {
707 				rv = ENOBUFS;
708 				goto done;
709 			}
710 			tln->tln_inp = inp;
711 			tln->tln_tp = tp;
712 		}
713 
714 		/*
715 		 * Drop the INP lock for a bit. We don't need it, and dropping
716 		 * it prevents lock order reversals.
717 		 */
718 		INP_WUNLOCK(inp);
719 
720 		/* Make sure we have at least a read lock on the tree. */
721 		tcp_log_id_validate_tree_lock(tree_locked);
722 		if (tree_locked == TREE_UNLOCKED) {
723 			TCPID_TREE_RLOCK();
724 			tree_locked = TREE_RLOCKED;
725 		}
726 
727 refind:
728 		/*
729 		 * Remember that we constructed (struct tcp_log_id_node) so
730 		 * we can safely cast the id to it for the purposes of finding.
731 		 */
732 		KASSERT(tlb == NULL, ("%s:%d tlb unexpectedly non-NULL",
733 		    __func__, __LINE__));
734 		tmp_tlb = RB_FIND(tcp_log_id_tree, &tcp_log_id_head,
735 		    (struct tcp_log_id_bucket *) id);
736 
737 		/*
738 		 * If we didn't find a matching bucket, we need to add a new
739 		 * one. This requires a write lock. But, of course, we will
740 		 * need to recheck some things when we re-acquire the lock.
741 		 */
742 		if (tmp_tlb == NULL && tree_locked != TREE_WLOCKED) {
743 			tree_locked = TREE_WLOCKED;
744 			if (!TCPID_TREE_UPGRADE()) {
745 				TCPID_TREE_RUNLOCK();
746 				TCPID_TREE_WLOCK();
747 
748 				/*
749 				 * The tree may have changed while we were
750 				 * unlocked.
751 				 */
752 				goto refind;
753 			}
754 		}
755 
756 		/* If we need to add a new bucket, do it now. */
757 		if (tmp_tlb == NULL) {
758 			/* Allocate new bucket. */
759 			tlb = uma_zalloc(tcp_log_bucket_zone, M_NOWAIT);
760 			if (tlb == NULL) {
761 				rv = ENOBUFS;
762 				goto done_noinp;
763 			}
764 			counter_u64_add(tcp_log_pcb_ids_cur, 1);
765 			counter_u64_add(tcp_log_pcb_ids_tot, 1);
766 
767 			if ((tcp_log_auto_all == false) &&
768 			    tcp_log_auto_mode &&
769 			    tcp_log_selectauto()) {
770 				/* Save off the log state */
771 				tlb->tlb_logstate = tcp_log_auto_mode;
772 			} else
773 				tlb->tlb_logstate = TCP_LOG_STATE_OFF;
774 			tlb->tlb_loglimit = 0;
775 			tlb->tlb_tag[0] = '\0'; /* Default to an empty tag. */
776 
777 			/*
778 			 * Copy the ID to the bucket.
779 			 * NB: Don't use strlcpy() unless you are sure
780 			 * we've always validated NULL termination.
781 			 *
782 			 * TODO: When I'm done writing this, see if we
783 			 * we have correctly validated NULL termination and
784 			 * can use strlcpy(). :-)
785 			 */
786 			strncpy(tlb->tlb_id, id, TCP_LOG_ID_LEN - 1);
787 			tlb->tlb_id[TCP_LOG_ID_LEN - 1] = '\0';
788 
789 			/*
790 			 * Take the refcount for the first node and go ahead
791 			 * and lock this. Note that we zero the tlb_mtx
792 			 * structure, since 0xdeadc0de flips the right bits
793 			 * for the code to think that this mutex has already
794 			 * been initialized. :-(
795 			 */
796 			SLIST_INIT(&tlb->tlb_head);
797 			refcount_init(&tlb->tlb_refcnt, 1);
798 			tlb->tlb_reqcnt = 1;
799 			memset(&tlb->tlb_mtx, 0, sizeof(struct mtx));
800 			TCPID_BUCKET_LOCK_INIT(tlb);
801 			TCPID_BUCKET_LOCK(tlb);
802 			bucket_locked = true;
803 
804 #define	FREE_NEW_TLB()	do {				\
805 	TCPID_BUCKET_LOCK_DESTROY(tlb);			\
806 	uma_zfree(tcp_log_bucket_zone, tlb);		\
807 	counter_u64_add(tcp_log_pcb_ids_cur, (int64_t)-1);	\
808 	counter_u64_add(tcp_log_pcb_ids_tot, (int64_t)-1);	\
809 	bucket_locked = false;				\
810 	tlb = NULL;					\
811 } while (0)
812 			/*
813 			 * Relock the INP and make sure we are still
814 			 * unassigned.
815 			 */
816 			INP_WLOCK(inp);
817 			RECHECK_INP_CLEAN(FREE_NEW_TLB());
818 			if (tp->t_lib != NULL) {
819 				FREE_NEW_TLB();
820 				goto restart;
821 			}
822 
823 			/* Add the new bucket to the tree. */
824 			tmp_tlb = RB_INSERT(tcp_log_id_tree, &tcp_log_id_head,
825 			    tlb);
826 			KASSERT(tmp_tlb == NULL,
827 			    ("%s: Unexpected conflicting bucket (%p) while "
828 			    "adding new bucket (%p)", __func__, tmp_tlb, tlb));
829 
830 			/*
831 			 * If we found a conflicting bucket, free the new
832 			 * one we made and fall through to use the existing
833 			 * bucket.
834 			 */
835 			if (tmp_tlb != NULL) {
836 				FREE_NEW_TLB();
837 				INP_WUNLOCK(inp);
838 			}
839 #undef	FREE_NEW_TLB
840 		}
841 
842 		/* If we found an existing bucket, use it. */
843 		if (tmp_tlb != NULL) {
844 			tlb = tmp_tlb;
845 			TCPID_BUCKET_LOCK(tlb);
846 			bucket_locked = true;
847 
848 			/*
849 			 * Relock the INP and make sure we are still
850 			 * unassigned.
851 			 */
852 			INP_UNLOCK_ASSERT(inp);
853 			INP_WLOCK(inp);
854 			RECHECK_INP();
855 			if (tp->t_lib != NULL) {
856 				TCPID_BUCKET_UNLOCK(tlb);
857 				bucket_locked = false;
858 				tlb = NULL;
859 				goto restart;
860 			}
861 
862 			/* Take a reference on the bucket. */
863 			TCPID_BUCKET_REF(tlb);
864 
865 			/* Record the request. */
866 			tcp_log_increment_reqcnt(tlb);
867 		}
868 
869 		tcp_log_grow_tlb(tlb->tlb_id, tp);
870 
871 		/* Add the new node to the list. */
872 		SLIST_INSERT_HEAD(&tlb->tlb_head, tln, tln_list);
873 		tp->t_lib = tlb;
874 		tp->t_lin = tln;
875 		if (tp->t_lib->tlb_logstate) {
876 			/* Clone in any logging */
877 
878 			tp->t_logstate = tp->t_lib->tlb_logstate;
879 		}
880 		if (tp->t_lib->tlb_loglimit) {
881 			/* The loglimit too */
882 
883 			tp->t_loglimit = tp->t_lib->tlb_loglimit;
884 		}
885 		tln = NULL;
886 	}
887 
888 	rv = 0;
889 
890 done:
891 	/* Unlock things, as needed, and return. */
892 	INP_WUNLOCK(inp);
893 done_noinp:
894 	INP_UNLOCK_ASSERT(inp);
895 	if (bucket_locked) {
896 		TCPID_BUCKET_LOCK_ASSERT(tlb);
897 		TCPID_BUCKET_UNLOCK(tlb);
898 	} else if (tlb != NULL)
899 		TCPID_BUCKET_UNLOCK_ASSERT(tlb);
900 	if (tree_locked == TREE_WLOCKED) {
901 		TCPID_TREE_WLOCK_ASSERT();
902 		TCPID_TREE_WUNLOCK();
903 	} else if (tree_locked == TREE_RLOCKED) {
904 		TCPID_TREE_RLOCK_ASSERT();
905 		TCPID_TREE_RUNLOCK();
906 	} else
907 		TCPID_TREE_UNLOCK_ASSERT();
908 	if (tln != NULL)
909 		uma_zfree(tcp_log_node_zone, tln);
910 	return (rv);
911 }
912 
913 /*
914  * Get the TCP log ID for a TCPCB.
915  * Called with INPCB locked.
916  * 'buf' must point to a buffer that is at least TCP_LOG_ID_LEN bytes long.
917  * Returns number of bytes copied.
918  */
919 size_t
920 tcp_log_get_id(struct tcpcb *tp, char *buf)
921 {
922 	size_t len;
923 
924 	INP_LOCK_ASSERT(tp->t_inpcb);
925 	if (tp->t_lib != NULL) {
926 		len = strlcpy(buf, tp->t_lib->tlb_id, TCP_LOG_ID_LEN);
927 		KASSERT(len < TCP_LOG_ID_LEN,
928 		    ("%s:%d: tp->t_lib->tlb_id too long (%zu)",
929 		    __func__, __LINE__, len));
930 	} else {
931 		*buf = '\0';
932 		len = 0;
933 	}
934 	return (len);
935 }
936 
937 /*
938  * Get the tag associated with the TCPCB's log ID.
939  * Called with INPCB locked. Returns with it unlocked.
940  * 'buf' must point to a buffer that is at least TCP_LOG_TAG_LEN bytes long.
941  * Returns number of bytes copied.
942  */
943 size_t
944 tcp_log_get_tag(struct tcpcb *tp, char *buf)
945 {
946 	struct tcp_log_id_bucket *tlb;
947 	size_t len;
948 	int tree_locked;
949 
950 	INP_WLOCK_ASSERT(tp->t_inpcb);
951 
952 	tree_locked = TREE_UNLOCKED;
953 	tlb = tp->t_lib;
954 
955 	if (tlb != NULL) {
956 		TCPID_BUCKET_REF(tlb);
957 		INP_WUNLOCK(tp->t_inpcb);
958 		TCPID_BUCKET_LOCK(tlb);
959 		len = strlcpy(buf, tlb->tlb_tag, TCP_LOG_TAG_LEN);
960 		KASSERT(len < TCP_LOG_TAG_LEN,
961 		    ("%s:%d: tp->t_lib->tlb_tag too long (%zu)",
962 		    __func__, __LINE__, len));
963 		if (!tcp_log_unref_bucket(tlb, &tree_locked, NULL))
964 			TCPID_BUCKET_UNLOCK(tlb);
965 
966 		if (tree_locked == TREE_WLOCKED) {
967 			TCPID_TREE_WLOCK_ASSERT();
968 			TCPID_TREE_WUNLOCK();
969 		} else if (tree_locked == TREE_RLOCKED) {
970 			TCPID_TREE_RLOCK_ASSERT();
971 			TCPID_TREE_RUNLOCK();
972 		} else
973 			TCPID_TREE_UNLOCK_ASSERT();
974 	} else {
975 		INP_WUNLOCK(tp->t_inpcb);
976 		*buf = '\0';
977 		len = 0;
978 	}
979 
980 	return (len);
981 }
982 
983 /*
984  * Get number of connections with the same log ID.
985  * Log ID is taken from given TCPCB.
986  * Called with INPCB locked.
987  */
988 u_int
989 tcp_log_get_id_cnt(struct tcpcb *tp)
990 {
991 
992 	INP_WLOCK_ASSERT(tp->t_inpcb);
993 	return ((tp->t_lib == NULL) ? 0 : tp->t_lib->tlb_refcnt);
994 }
995 
996 #ifdef TCPLOG_DEBUG_RINGBUF
997 /*
998  * Functions/macros to increment/decrement reference count for a log
999  * entry. This should catch when we do a double-free/double-remove or
1000  * a double-add.
1001  */
1002 static inline void
1003 _tcp_log_entry_refcnt_add(struct tcp_log_mem *log_entry, const char *func,
1004     int line)
1005 {
1006 	int refcnt;
1007 
1008 	refcnt = atomic_fetchadd_int(&log_entry->tlm_refcnt, 1);
1009 	if (refcnt != 0)
1010 		panic("%s:%d: log_entry(%p)->tlm_refcnt is %d (expected 0)",
1011 		    func, line, log_entry, refcnt);
1012 }
1013 #define	tcp_log_entry_refcnt_add(l)	\
1014     _tcp_log_entry_refcnt_add((l), __func__, __LINE__)
1015 
1016 static inline void
1017 _tcp_log_entry_refcnt_rem(struct tcp_log_mem *log_entry, const char *func,
1018     int line)
1019 {
1020 	int refcnt;
1021 
1022 	refcnt = atomic_fetchadd_int(&log_entry->tlm_refcnt, -1);
1023 	if (refcnt != 1)
1024 		panic("%s:%d: log_entry(%p)->tlm_refcnt is %d (expected 1)",
1025 		    func, line, log_entry, refcnt);
1026 }
1027 #define	tcp_log_entry_refcnt_rem(l)	\
1028     _tcp_log_entry_refcnt_rem((l), __func__, __LINE__)
1029 
1030 #else /* !TCPLOG_DEBUG_RINGBUF */
1031 
1032 #define	tcp_log_entry_refcnt_add(l)
1033 #define	tcp_log_entry_refcnt_rem(l)
1034 
1035 #endif
1036 
1037 /*
1038  * Cleanup after removing a log entry, but only decrement the count if we
1039  * are running INVARIANTS.
1040  */
1041 static inline void
1042 tcp_log_free_log_common(struct tcp_log_mem *log_entry, int *count __unused)
1043 {
1044 
1045 	uma_zfree(tcp_log_zone, log_entry);
1046 #ifdef INVARIANTS
1047 	(*count)--;
1048 	KASSERT(*count >= 0,
1049 	    ("%s: count unexpectedly negative", __func__));
1050 #endif
1051 }
1052 
1053 static void
1054 tcp_log_free_entries(struct tcp_log_stailq *head, int *count)
1055 {
1056 	struct tcp_log_mem *log_entry;
1057 
1058 	/* Free the entries. */
1059 	while ((log_entry = STAILQ_FIRST(head)) != NULL) {
1060 		STAILQ_REMOVE_HEAD(head, tlm_queue);
1061 		tcp_log_entry_refcnt_rem(log_entry);
1062 		tcp_log_free_log_common(log_entry, count);
1063 	}
1064 }
1065 
1066 /* Cleanup after removing a log entry. */
1067 static inline void
1068 tcp_log_remove_log_cleanup(struct tcpcb *tp, struct tcp_log_mem *log_entry)
1069 {
1070 	uma_zfree(tcp_log_zone, log_entry);
1071 	tp->t_lognum--;
1072 	KASSERT(tp->t_lognum >= 0,
1073 	    ("%s: tp->t_lognum unexpectedly negative", __func__));
1074 }
1075 
1076 /* Remove a log entry from the head of a list. */
1077 static inline void
1078 tcp_log_remove_log_head(struct tcpcb *tp, struct tcp_log_mem *log_entry)
1079 {
1080 
1081 	KASSERT(log_entry == STAILQ_FIRST(&tp->t_logs),
1082 	    ("%s: attempt to remove non-HEAD log entry", __func__));
1083 	STAILQ_REMOVE_HEAD(&tp->t_logs, tlm_queue);
1084 	tcp_log_entry_refcnt_rem(log_entry);
1085 	tcp_log_remove_log_cleanup(tp, log_entry);
1086 }
1087 
1088 #ifdef TCPLOG_DEBUG_RINGBUF
1089 /*
1090  * Initialize the log entry's reference count, which we want to
1091  * survive allocations.
1092  */
1093 static int
1094 tcp_log_zone_init(void *mem, int size, int flags __unused)
1095 {
1096 	struct tcp_log_mem *tlm;
1097 
1098 	KASSERT(size >= sizeof(struct tcp_log_mem),
1099 	    ("%s: unexpectedly short (%d) allocation", __func__, size));
1100 	tlm = (struct tcp_log_mem *)mem;
1101 	tlm->tlm_refcnt = 0;
1102 	return (0);
1103 }
1104 
1105 /*
1106  * Double check that the refcnt is zero on allocation and return.
1107  */
1108 static int
1109 tcp_log_zone_ctor(void *mem, int size, void *args __unused, int flags __unused)
1110 {
1111 	struct tcp_log_mem *tlm;
1112 
1113 	KASSERT(size >= sizeof(struct tcp_log_mem),
1114 	    ("%s: unexpectedly short (%d) allocation", __func__, size));
1115 	tlm = (struct tcp_log_mem *)mem;
1116 	if (tlm->tlm_refcnt != 0)
1117 		panic("%s:%d: tlm(%p)->tlm_refcnt is %d (expected 0)",
1118 		    __func__, __LINE__, tlm, tlm->tlm_refcnt);
1119 	return (0);
1120 }
1121 
1122 static void
1123 tcp_log_zone_dtor(void *mem, int size, void *args __unused)
1124 {
1125 	struct tcp_log_mem *tlm;
1126 
1127 	KASSERT(size >= sizeof(struct tcp_log_mem),
1128 	    ("%s: unexpectedly short (%d) allocation", __func__, size));
1129 	tlm = (struct tcp_log_mem *)mem;
1130 	if (tlm->tlm_refcnt != 0)
1131 		panic("%s:%d: tlm(%p)->tlm_refcnt is %d (expected 0)",
1132 		    __func__, __LINE__, tlm, tlm->tlm_refcnt);
1133 }
1134 #endif /* TCPLOG_DEBUG_RINGBUF */
1135 
1136 /* Do global initialization. */
1137 void
1138 tcp_log_init(void)
1139 {
1140 
1141 	tcp_log_zone = uma_zcreate("tcp_log", sizeof(struct tcp_log_mem),
1142 #ifdef TCPLOG_DEBUG_RINGBUF
1143 	    tcp_log_zone_ctor, tcp_log_zone_dtor, tcp_log_zone_init,
1144 #else
1145 	    NULL, NULL, NULL,
1146 #endif
1147 	    NULL, UMA_ALIGN_PTR, 0);
1148 	(void)uma_zone_set_max(tcp_log_zone, TCP_LOG_BUF_DEFAULT_GLOBAL_LIMIT);
1149 	tcp_log_bucket_zone = uma_zcreate("tcp_log_bucket",
1150 	    sizeof(struct tcp_log_id_bucket), NULL, NULL, NULL, NULL,
1151 	    UMA_ALIGN_PTR, 0);
1152 	tcp_log_node_zone = uma_zcreate("tcp_log_node",
1153 	    sizeof(struct tcp_log_id_node), NULL, NULL, NULL, NULL,
1154 	    UMA_ALIGN_PTR, 0);
1155 #ifdef TCPLOG_DEBUG_COUNTERS
1156 	tcp_log_queued = counter_u64_alloc(M_WAITOK);
1157 	tcp_log_que_fail1 = counter_u64_alloc(M_WAITOK);
1158 	tcp_log_que_fail2 = counter_u64_alloc(M_WAITOK);
1159 	tcp_log_que_fail3 = counter_u64_alloc(M_WAITOK);
1160 	tcp_log_que_fail4 = counter_u64_alloc(M_WAITOK);
1161 	tcp_log_que_fail5 = counter_u64_alloc(M_WAITOK);
1162 	tcp_log_que_copyout = counter_u64_alloc(M_WAITOK);
1163 	tcp_log_que_read = counter_u64_alloc(M_WAITOK);
1164 	tcp_log_que_freed = counter_u64_alloc(M_WAITOK);
1165 #endif
1166 	tcp_log_pcb_ids_cur = counter_u64_alloc(M_WAITOK);
1167 	tcp_log_pcb_ids_tot = counter_u64_alloc(M_WAITOK);
1168 
1169 	rw_init_flags(&tcp_id_tree_lock, "TCP ID tree", RW_NEW);
1170 	mtx_init(&tcp_log_expireq_mtx, "TCP log expireq", NULL, MTX_DEF);
1171 	callout_init(&tcp_log_expireq_callout, 1);
1172 }
1173 
1174 /* Do per-TCPCB initialization. */
1175 void
1176 tcp_log_tcpcbinit(struct tcpcb *tp)
1177 {
1178 
1179 	/* A new TCPCB should start out zero-initialized. */
1180 	STAILQ_INIT(&tp->t_logs);
1181 
1182 	/*
1183 	 * If we are doing auto-capturing, figure out whether we will capture
1184 	 * this session.
1185 	 */
1186 	tp->t_loglimit = tcp_log_session_limit;
1187 	if ((tcp_log_auto_all == true) &&
1188 	    tcp_log_auto_mode &&
1189 	    tcp_log_selectauto()) {
1190 		tp->t_logstate = tcp_log_auto_mode;
1191 		tp->t_flags2 |= TF2_LOG_AUTO;
1192 	}
1193 }
1194 
1195 /* Remove entries */
1196 static void
1197 tcp_log_expire(void *unused __unused)
1198 {
1199 	struct tcp_log_id_bucket *tlb;
1200 	struct tcp_log_id_node *tln;
1201 	sbintime_t expiry_limit;
1202 	int tree_locked;
1203 
1204 	TCPLOG_EXPIREQ_LOCK();
1205 	if (callout_pending(&tcp_log_expireq_callout)) {
1206 		/* Callout was reset. */
1207 		TCPLOG_EXPIREQ_UNLOCK();
1208 		return;
1209 	}
1210 
1211 	/*
1212 	 * Process entries until we reach one that expires too far in the
1213 	 * future. Look one second in the future.
1214 	 */
1215 	expiry_limit = getsbinuptime() + SBT_1S;
1216 	tree_locked = TREE_UNLOCKED;
1217 
1218 	while ((tln = STAILQ_FIRST(&tcp_log_expireq_head)) != NULL &&
1219 	    tln->tln_expiretime <= expiry_limit) {
1220 		if (!callout_active(&tcp_log_expireq_callout)) {
1221 			/*
1222 			 * Callout was stopped. I guess we should
1223 			 * just quit at this point.
1224 			 */
1225 			TCPLOG_EXPIREQ_UNLOCK();
1226 			return;
1227 		}
1228 
1229 		/*
1230 		 * Remove the node from the head of the list and unlock
1231 		 * the list. Change the expiry time to SBT_MAX as a signal
1232 		 * to other threads that we now own this.
1233 		 */
1234 		STAILQ_REMOVE_HEAD(&tcp_log_expireq_head, tln_expireq);
1235 		tln->tln_expiretime = SBT_MAX;
1236 		TCPLOG_EXPIREQ_UNLOCK();
1237 
1238 		/*
1239 		 * Remove the node from the bucket.
1240 		 */
1241 		tlb = tln->tln_bucket;
1242 		TCPID_BUCKET_LOCK(tlb);
1243 		if (tcp_log_remove_id_node(NULL, NULL, tlb, tln, &tree_locked)) {
1244 			tcp_log_id_validate_tree_lock(tree_locked);
1245 			if (tree_locked == TREE_WLOCKED)
1246 				TCPID_TREE_WUNLOCK();
1247 			else
1248 				TCPID_TREE_RUNLOCK();
1249 			tree_locked = TREE_UNLOCKED;
1250 		}
1251 
1252 		/* Drop the INP reference. */
1253 		INP_WLOCK(tln->tln_inp);
1254 		if (!in_pcbrele_wlocked(tln->tln_inp))
1255 			INP_WUNLOCK(tln->tln_inp);
1256 
1257 		/* Free the log records. */
1258 		tcp_log_free_entries(&tln->tln_entries, &tln->tln_count);
1259 
1260 		/* Free the node. */
1261 		uma_zfree(tcp_log_node_zone, tln);
1262 
1263 		/* Relock the expiry queue. */
1264 		TCPLOG_EXPIREQ_LOCK();
1265 	}
1266 
1267 	/*
1268 	 * We've expired all the entries we can. Do we need to reschedule
1269 	 * ourselves?
1270 	 */
1271 	callout_deactivate(&tcp_log_expireq_callout);
1272 	if (tln != NULL) {
1273 		/*
1274 		 * Get max(now + TCP_LOG_EXPIRE_INTVL, tln->tln_expiretime) and
1275 		 * set the next callout to that. (This helps ensure we generally
1276 		 * run the callout no more often than desired.)
1277 		 */
1278 		expiry_limit = getsbinuptime() + TCP_LOG_EXPIRE_INTVL;
1279 		if (expiry_limit < tln->tln_expiretime)
1280 			expiry_limit = tln->tln_expiretime;
1281 		callout_reset_sbt(&tcp_log_expireq_callout, expiry_limit,
1282 		    SBT_1S, tcp_log_expire, NULL, C_ABSOLUTE);
1283 	}
1284 
1285 	/* We're done. */
1286 	TCPLOG_EXPIREQ_UNLOCK();
1287 	return;
1288 }
1289 
1290 /*
1291  * Move log data from the TCPCB to a new node. This will reset the TCPCB log
1292  * entries and log count; however, it will not touch other things from the
1293  * TCPCB (e.g. t_lin, t_lib).
1294  *
1295  * NOTE: Must hold a lock on the INP.
1296  */
1297 static void
1298 tcp_log_move_tp_to_node(struct tcpcb *tp, struct tcp_log_id_node *tln)
1299 {
1300 
1301 	INP_WLOCK_ASSERT(tp->t_inpcb);
1302 
1303 	tln->tln_ie = tp->t_inpcb->inp_inc.inc_ie;
1304 	if (tp->t_inpcb->inp_inc.inc_flags & INC_ISIPV6)
1305 		tln->tln_af = AF_INET6;
1306 	else
1307 		tln->tln_af = AF_INET;
1308 	tln->tln_entries = tp->t_logs;
1309 	tln->tln_count = tp->t_lognum;
1310 	tln->tln_bucket = tp->t_lib;
1311 
1312 	/* Clear information from the PCB. */
1313 	STAILQ_INIT(&tp->t_logs);
1314 	tp->t_lognum = 0;
1315 }
1316 
1317 /* Do per-TCPCB cleanup */
1318 void
1319 tcp_log_tcpcbfini(struct tcpcb *tp)
1320 {
1321 	struct tcp_log_id_node *tln, *tln_first;
1322 	struct tcp_log_mem *log_entry;
1323 	sbintime_t callouttime;
1324 
1325 	INP_WLOCK_ASSERT(tp->t_inpcb);
1326 
1327 	TCP_LOG_EVENT(tp, NULL, NULL, NULL, TCP_LOG_CONNEND, 0, 0, NULL, false);
1328 
1329 	/*
1330 	 * If we were gathering packets to be automatically dumped, try to do
1331 	 * it now. If this succeeds, the log information in the TCPCB will be
1332 	 * cleared. Otherwise, we'll handle the log information as we do
1333 	 * for other states.
1334 	 */
1335 	switch(tp->t_logstate) {
1336 	case TCP_LOG_STATE_HEAD_AUTO:
1337 		(void)tcp_log_dump_tp_logbuf(tp, "auto-dumped from head",
1338 		    M_NOWAIT, false);
1339 		break;
1340 	case TCP_LOG_STATE_TAIL_AUTO:
1341 		(void)tcp_log_dump_tp_logbuf(tp, "auto-dumped from tail",
1342 		    M_NOWAIT, false);
1343 		break;
1344 	case TCP_LOG_STATE_CONTINUAL:
1345 		(void)tcp_log_dump_tp_logbuf(tp, "auto-dumped from continual",
1346 		    M_NOWAIT, false);
1347 		break;
1348 	}
1349 
1350 	/*
1351 	 * There are two ways we could keep logs: per-socket or per-ID. If
1352 	 * we are tracking logs with an ID, then the logs survive the
1353 	 * destruction of the TCPCB.
1354 	 *
1355 	 * If the TCPCB is associated with an ID node, move the logs from the
1356 	 * TCPCB to the ID node. In theory, this is safe, for reasons which I
1357 	 * will now explain for my own benefit when I next need to figure out
1358 	 * this code. :-)
1359 	 *
1360 	 * We own the INP lock. Therefore, no one else can change the contents
1361 	 * of this node (Rule C). Further, no one can remove this node from
1362 	 * the bucket while we hold the lock (Rule D). Basically, no one can
1363 	 * mess with this node. That leaves two states in which we could be:
1364 	 *
1365 	 * 1. Another thread is currently waiting to acquire the INP lock, with
1366 	 *    plans to do something with this node. When we drop the INP lock,
1367 	 *    they will have a chance to do that. They will recheck the
1368 	 *    tln_closed field (see note to Rule C) and then acquire the
1369 	 *    bucket lock before proceeding further.
1370 	 *
1371 	 * 2. Another thread will try to acquire a lock at some point in the
1372 	 *    future. If they try to acquire a lock before we set the
1373 	 *    tln_closed field, they will follow state #1. If they try to
1374 	 *    acquire a lock after we set the tln_closed field, they will be
1375 	 *    able to make changes to the node, at will, following Rule C.
1376 	 *
1377 	 * Therefore, we currently own this node and can make any changes
1378 	 * we want. But, as soon as we set the tln_closed field to true, we
1379 	 * have effectively dropped our lock on the node. (For this reason, we
1380 	 * also need to make sure our writes are ordered correctly. An atomic
1381 	 * operation with "release" semantics should be sufficient.)
1382 	 */
1383 
1384 	if (tp->t_lin != NULL) {
1385 		/* Copy the relevant information to the log entry. */
1386 		tln = tp->t_lin;
1387 		KASSERT(tln->tln_inp == tp->t_inpcb,
1388 		    ("%s: Mismatched inp (tln->tln_inp=%p, tp->t_inpcb=%p)",
1389 		    __func__, tln->tln_inp, tp->t_inpcb));
1390 		tcp_log_move_tp_to_node(tp, tln);
1391 
1392 		/* Clear information from the PCB. */
1393 		tp->t_lin = NULL;
1394 		tp->t_lib = NULL;
1395 
1396 		/*
1397 		 * Take a reference on the INP. This ensures that the INP
1398 		 * remains valid while the node is on the expiry queue. This
1399 		 * ensures the INP is valid for other threads that may be
1400 		 * racing to lock this node when we move it to the expire
1401 		 * queue.
1402 		 */
1403 		in_pcbref(tp->t_inpcb);
1404 
1405 		/*
1406 		 * Store the entry on the expiry list. The exact behavior
1407 		 * depends on whether we have entries to keep. If so, we
1408 		 * put the entry at the tail of the list and expire in
1409 		 * TCP_LOG_EXPIRE_TIME. Otherwise, we expire "now" and put
1410 		 * the entry at the head of the list. (Handling the cleanup
1411 		 * via the expiry timer lets us avoid locking messy-ness here.)
1412 		 */
1413 		tln->tln_expiretime = getsbinuptime();
1414 		TCPLOG_EXPIREQ_LOCK();
1415 		if (tln->tln_count) {
1416 			tln->tln_expiretime += TCP_LOG_EXPIRE_TIME;
1417 			if (STAILQ_EMPTY(&tcp_log_expireq_head) &&
1418 			    !callout_active(&tcp_log_expireq_callout)) {
1419 				/*
1420 				 * We are adding the first entry and a callout
1421 				 * is not currently scheduled; therefore, we
1422 				 * need to schedule one.
1423 				 */
1424 				callout_reset_sbt(&tcp_log_expireq_callout,
1425 				    tln->tln_expiretime, SBT_1S, tcp_log_expire,
1426 				    NULL, C_ABSOLUTE);
1427 			}
1428 			STAILQ_INSERT_TAIL(&tcp_log_expireq_head, tln,
1429 			    tln_expireq);
1430 		} else {
1431 			callouttime = tln->tln_expiretime +
1432 			    TCP_LOG_EXPIRE_INTVL;
1433 			tln_first = STAILQ_FIRST(&tcp_log_expireq_head);
1434 
1435 			if ((tln_first == NULL ||
1436 			    callouttime < tln_first->tln_expiretime) &&
1437 			    (callout_pending(&tcp_log_expireq_callout) ||
1438 			    !callout_active(&tcp_log_expireq_callout))) {
1439 				/*
1440 				 * The list is empty, or we want to run the
1441 				 * expire code before the first entry's timer
1442 				 * fires. Also, we are in a case where a callout
1443 				 * is not actively running. We want to reset
1444 				 * the callout to occur sooner.
1445 				 */
1446 				callout_reset_sbt(&tcp_log_expireq_callout,
1447 				    callouttime, SBT_1S, tcp_log_expire, NULL,
1448 				    C_ABSOLUTE);
1449 			}
1450 
1451 			/*
1452 			 * Insert to the head, or just after the head, as
1453 			 * appropriate. (This might result in small
1454 			 * mis-orderings as a bunch of "expire now" entries
1455 			 * gather at the start of the list, but that should
1456 			 * not produce big problems, since the expire timer
1457 			 * will walk through all of them.)
1458 			 */
1459 			if (tln_first == NULL ||
1460 			    tln->tln_expiretime < tln_first->tln_expiretime)
1461 				STAILQ_INSERT_HEAD(&tcp_log_expireq_head, tln,
1462 				    tln_expireq);
1463 			else
1464 				STAILQ_INSERT_AFTER(&tcp_log_expireq_head,
1465 				    tln_first, tln, tln_expireq);
1466 		}
1467 		TCPLOG_EXPIREQ_UNLOCK();
1468 
1469 		/*
1470 		 * We are done messing with the tln. After this point, we
1471 		 * can't touch it. (Note that the "release" semantics should
1472 		 * be included with the TCPLOG_EXPIREQ_UNLOCK() call above.
1473 		 * Therefore, they should be unnecessary here. However, it
1474 		 * seems like a good idea to include them anyway, since we
1475 		 * really are releasing a lock here.)
1476 		 */
1477 		atomic_store_rel_int(&tln->tln_closed, 1);
1478 	} else {
1479 		/* Remove log entries. */
1480 		while ((log_entry = STAILQ_FIRST(&tp->t_logs)) != NULL)
1481 			tcp_log_remove_log_head(tp, log_entry);
1482 		KASSERT(tp->t_lognum == 0,
1483 		    ("%s: After freeing entries, tp->t_lognum=%d (expected 0)",
1484 			__func__, tp->t_lognum));
1485 	}
1486 
1487 	/*
1488 	 * Change the log state to off (just in case anything tries to sneak
1489 	 * in a last-minute log).
1490 	 */
1491 	tp->t_logstate = TCP_LOG_STATE_OFF;
1492 }
1493 
1494 static void
1495 tcp_log_purge_tp_logbuf(struct tcpcb *tp)
1496 {
1497 	struct tcp_log_mem *log_entry;
1498 	struct inpcb *inp;
1499 
1500 	inp = tp->t_inpcb;
1501 	INP_WLOCK_ASSERT(inp);
1502 	if (tp->t_lognum == 0)
1503 		return;
1504 
1505 	while ((log_entry = STAILQ_FIRST(&tp->t_logs)) != NULL)
1506 		tcp_log_remove_log_head(tp, log_entry);
1507 	KASSERT(tp->t_lognum == 0,
1508 		("%s: After freeing entries, tp->t_lognum=%d (expected 0)",
1509 		 __func__, tp->t_lognum));
1510 	tp->t_logstate = TCP_LOG_STATE_OFF;
1511 }
1512 
1513 /*
1514  * This logs an event for a TCP socket. Normally, this is called via
1515  * TCP_LOG_EVENT or TCP_LOG_EVENT_VERBOSE. See the documentation for
1516  * TCP_LOG_EVENT().
1517  */
1518 
1519 struct tcp_log_buffer *
1520 tcp_log_event_(struct tcpcb *tp, struct tcphdr *th, struct sockbuf *rxbuf,
1521     struct sockbuf *txbuf, uint8_t eventid, int errornum, uint32_t len,
1522     union tcp_log_stackspecific *stackinfo, int th_hostorder,
1523     const char *output_caller, const char *func, int line, const struct timeval *itv)
1524 {
1525 	struct tcp_log_mem *log_entry;
1526 	struct tcp_log_buffer *log_buf;
1527 	int attempt_count = 0;
1528 	struct tcp_log_verbose *log_verbose;
1529 	uint32_t logsn;
1530 
1531 	KASSERT((func == NULL && line == 0) || (func != NULL && line > 0),
1532 	    ("%s called with inconsistent func (%p) and line (%d) arguments",
1533 		__func__, func, line));
1534 
1535 	INP_WLOCK_ASSERT(tp->t_inpcb);
1536 	if (tcp_disable_all_bb_logs) {
1537 		/*
1538 		 * The global shutdown logging
1539 		 * switch has been thrown. Call
1540 		 * the purge function that frees
1541 		 * purges out the logs and
1542 		 * turns off logging.
1543 		 */
1544 		tcp_log_purge_tp_logbuf(tp);
1545 		return (NULL);
1546 	}
1547 	KASSERT(tp->t_logstate == TCP_LOG_STATE_HEAD ||
1548 	    tp->t_logstate == TCP_LOG_STATE_TAIL ||
1549 	    tp->t_logstate == TCP_LOG_STATE_CONTINUAL ||
1550 	    tp->t_logstate == TCP_LOG_STATE_HEAD_AUTO ||
1551 	    tp->t_logstate == TCP_LOG_STATE_TAIL_AUTO,
1552 	    ("%s called with unexpected tp->t_logstate (%d)", __func__,
1553 		tp->t_logstate));
1554 
1555 	/*
1556 	 * Get the serial number. We do this early so it will
1557 	 * increment even if we end up skipping the log entry for some
1558 	 * reason.
1559 	 */
1560 	logsn = tp->t_logsn++;
1561 
1562 	/*
1563 	 * Can we get a new log entry? If so, increment the lognum counter
1564 	 * here.
1565 	 */
1566 retry:
1567 	if (tp->t_lognum < tp->t_loglimit) {
1568 		if ((log_entry = uma_zalloc(tcp_log_zone, M_NOWAIT)) != NULL)
1569 			tp->t_lognum++;
1570 	} else
1571 		log_entry = NULL;
1572 
1573 	/* Do we need to try to reuse? */
1574 	if (log_entry == NULL) {
1575 		/*
1576 		 * Sacrifice auto-logged sessions without a log ID if
1577 		 * tcp_log_auto_all is false. (If they don't have a log
1578 		 * ID by now, it is probable that either they won't get one
1579 		 * or we are resource-constrained.)
1580 		 */
1581 		if (tp->t_lib == NULL && (tp->t_flags2 & TF2_LOG_AUTO) &&
1582 		    !tcp_log_auto_all) {
1583 			if (tcp_log_state_change(tp, TCP_LOG_STATE_CLEAR)) {
1584 #ifdef INVARIANTS
1585 				panic("%s:%d: tcp_log_state_change() failed "
1586 				    "to set tp %p to TCP_LOG_STATE_CLEAR",
1587 				    __func__, __LINE__, tp);
1588 #endif
1589 				tp->t_logstate = TCP_LOG_STATE_OFF;
1590 			}
1591 			return (NULL);
1592 		}
1593 		/*
1594 		 * If we are in TCP_LOG_STATE_HEAD_AUTO state, try to dump
1595 		 * the buffers. If successful, deactivate tracing. Otherwise,
1596 		 * leave it active so we will retry.
1597 		 */
1598 		if (tp->t_logstate == TCP_LOG_STATE_HEAD_AUTO &&
1599 		    !tcp_log_dump_tp_logbuf(tp, "auto-dumped from head",
1600 		    M_NOWAIT, false)) {
1601 			tp->t_logstate = TCP_LOG_STATE_OFF;
1602 			return(NULL);
1603 		} else if ((tp->t_logstate == TCP_LOG_STATE_CONTINUAL) &&
1604 		    !tcp_log_dump_tp_logbuf(tp, "auto-dumped from continual",
1605 		    M_NOWAIT, false)) {
1606 			if (attempt_count == 0) {
1607 				attempt_count++;
1608 				goto retry;
1609 			}
1610 #ifdef TCPLOG_DEBUG_COUNTERS
1611 			counter_u64_add(tcp_log_que_fail4, 1);
1612 #endif
1613 			return(NULL);
1614 		} else if (tp->t_logstate == TCP_LOG_STATE_HEAD_AUTO)
1615 			return(NULL);
1616 
1617 		/* If in HEAD state, just deactivate the tracing and return. */
1618 		if (tp->t_logstate == TCP_LOG_STATE_HEAD) {
1619 			tp->t_logstate = TCP_LOG_STATE_OFF;
1620 			return(NULL);
1621 		}
1622 
1623 		/*
1624 		 * Get a buffer to reuse. If that fails, just give up.
1625 		 * (We can't log anything without a buffer in which to
1626 		 * put it.)
1627 		 *
1628 		 * Note that we don't change the t_lognum counter
1629 		 * here. Because we are re-using the buffer, the total
1630 		 * number won't change.
1631 		 */
1632 		if ((log_entry = STAILQ_FIRST(&tp->t_logs)) == NULL)
1633 			return(NULL);
1634 		STAILQ_REMOVE_HEAD(&tp->t_logs, tlm_queue);
1635 		tcp_log_entry_refcnt_rem(log_entry);
1636 	}
1637 
1638 	KASSERT(log_entry != NULL,
1639 	    ("%s: log_entry unexpectedly NULL", __func__));
1640 
1641 	/* Extract the log buffer and verbose buffer pointers. */
1642 	log_buf = &log_entry->tlm_buf;
1643 	log_verbose = &log_entry->tlm_v;
1644 
1645 	/* Basic entries. */
1646 	if (itv == NULL)
1647 		getmicrouptime(&log_buf->tlb_tv);
1648 	else
1649 		memcpy(&log_buf->tlb_tv, itv, sizeof(struct timeval));
1650 	log_buf->tlb_ticks = ticks;
1651 	log_buf->tlb_sn = logsn;
1652 	log_buf->tlb_stackid = tp->t_fb->tfb_id;
1653 	log_buf->tlb_eventid = eventid;
1654 	log_buf->tlb_eventflags = 0;
1655 	log_buf->tlb_errno = errornum;
1656 
1657 	/* Socket buffers */
1658 	if (rxbuf != NULL) {
1659 		log_buf->tlb_eventflags |= TLB_FLAG_RXBUF;
1660 		log_buf->tlb_rxbuf.tls_sb_acc = rxbuf->sb_acc;
1661 		log_buf->tlb_rxbuf.tls_sb_ccc = rxbuf->sb_ccc;
1662 		log_buf->tlb_rxbuf.tls_sb_spare = 0;
1663 	}
1664 	if (txbuf != NULL) {
1665 		log_buf->tlb_eventflags |= TLB_FLAG_TXBUF;
1666 		log_buf->tlb_txbuf.tls_sb_acc = txbuf->sb_acc;
1667 		log_buf->tlb_txbuf.tls_sb_ccc = txbuf->sb_ccc;
1668 		log_buf->tlb_txbuf.tls_sb_spare = 0;
1669 	}
1670 	/* Copy values from tp to the log entry. */
1671 #define	COPY_STAT(f)	log_buf->tlb_ ## f = tp->f
1672 #define	COPY_STAT_T(f)	log_buf->tlb_ ## f = tp->t_ ## f
1673 	COPY_STAT_T(state);
1674 	COPY_STAT_T(starttime);
1675 	COPY_STAT(iss);
1676 	COPY_STAT_T(flags);
1677 	COPY_STAT(snd_una);
1678 	COPY_STAT(snd_max);
1679 	COPY_STAT(snd_cwnd);
1680 	COPY_STAT(snd_nxt);
1681 	COPY_STAT(snd_recover);
1682 	COPY_STAT(snd_wnd);
1683 	COPY_STAT(snd_ssthresh);
1684 	COPY_STAT_T(srtt);
1685 	COPY_STAT_T(rttvar);
1686 	COPY_STAT(rcv_up);
1687 	COPY_STAT(rcv_adv);
1688 	COPY_STAT(rcv_nxt);
1689 	COPY_STAT(rcv_wnd);
1690 	COPY_STAT_T(dupacks);
1691 	COPY_STAT_T(segqlen);
1692 	COPY_STAT(snd_numholes);
1693 	COPY_STAT(snd_scale);
1694 	COPY_STAT(rcv_scale);
1695 	COPY_STAT_T(flags2);
1696 	COPY_STAT_T(fbyte_in);
1697 	COPY_STAT_T(fbyte_out);
1698 #undef COPY_STAT
1699 #undef COPY_STAT_T
1700 	log_buf->tlb_flex1 = 0;
1701 	log_buf->tlb_flex2 = 0;
1702 	/* Copy stack-specific info. */
1703 	if (stackinfo != NULL) {
1704 		memcpy(&log_buf->tlb_stackinfo, stackinfo,
1705 		    sizeof(log_buf->tlb_stackinfo));
1706 		log_buf->tlb_eventflags |= TLB_FLAG_STACKINFO;
1707 	}
1708 
1709 	/* The packet */
1710 	log_buf->tlb_len = len;
1711 	if (th) {
1712 		int optlen;
1713 
1714 		log_buf->tlb_eventflags |= TLB_FLAG_HDR;
1715 		log_buf->tlb_th = *th;
1716 		if (th_hostorder)
1717 			tcp_fields_to_net(&log_buf->tlb_th);
1718 		optlen = (th->th_off << 2) - sizeof (struct tcphdr);
1719 		if (optlen > 0)
1720 			memcpy(log_buf->tlb_opts, th + 1, optlen);
1721 	}
1722 
1723 	/* Verbose information */
1724 	if (func != NULL) {
1725 		log_buf->tlb_eventflags |= TLB_FLAG_VERBOSE;
1726 		if (output_caller != NULL)
1727 			strlcpy(log_verbose->tlv_snd_frm, output_caller,
1728 			    TCP_FUNC_LEN);
1729 		else
1730 			*log_verbose->tlv_snd_frm = 0;
1731 		strlcpy(log_verbose->tlv_trace_func, func, TCP_FUNC_LEN);
1732 		log_verbose->tlv_trace_line = line;
1733 	}
1734 
1735 	/* Insert the new log at the tail. */
1736 	STAILQ_INSERT_TAIL(&tp->t_logs, log_entry, tlm_queue);
1737 	tcp_log_entry_refcnt_add(log_entry);
1738 	return (log_buf);
1739 }
1740 
1741 /*
1742  * Change the logging state for a TCPCB. Returns 0 on success or an
1743  * error code on failure.
1744  */
1745 int
1746 tcp_log_state_change(struct tcpcb *tp, int state)
1747 {
1748 	struct tcp_log_mem *log_entry;
1749 
1750 	INP_WLOCK_ASSERT(tp->t_inpcb);
1751 	switch(state) {
1752 	case TCP_LOG_STATE_CLEAR:
1753 		while ((log_entry = STAILQ_FIRST(&tp->t_logs)) != NULL)
1754 			tcp_log_remove_log_head(tp, log_entry);
1755 		/* Fall through */
1756 
1757 	case TCP_LOG_STATE_OFF:
1758 		tp->t_logstate = TCP_LOG_STATE_OFF;
1759 		break;
1760 
1761 	case TCP_LOG_STATE_TAIL:
1762 	case TCP_LOG_STATE_HEAD:
1763 	case TCP_LOG_STATE_CONTINUAL:
1764 	case TCP_LOG_STATE_HEAD_AUTO:
1765 	case TCP_LOG_STATE_TAIL_AUTO:
1766 		tp->t_logstate = state;
1767 		break;
1768 
1769 	default:
1770 		return (EINVAL);
1771 	}
1772 	if (tcp_disable_all_bb_logs) {
1773 		/* We are prohibited from doing any logs */
1774 		tp->t_logstate = TCP_LOG_STATE_OFF;
1775 	}
1776 	tp->t_flags2 &= ~(TF2_LOG_AUTO);
1777 
1778 	return (0);
1779 }
1780 
1781 /* If tcp_drain() is called, flush half the log entries. */
1782 void
1783 tcp_log_drain(struct tcpcb *tp)
1784 {
1785 	struct tcp_log_mem *log_entry, *next;
1786 	int target, skip;
1787 
1788 	INP_WLOCK_ASSERT(tp->t_inpcb);
1789 	if ((target = tp->t_lognum / 2) == 0)
1790 		return;
1791 
1792 	/*
1793 	 * If we are logging the "head" packets, we want to discard
1794 	 * from the tail of the queue. Otherwise, we want to discard
1795 	 * from the head.
1796 	 */
1797 	if (tp->t_logstate == TCP_LOG_STATE_HEAD ||
1798 	    tp->t_logstate == TCP_LOG_STATE_HEAD_AUTO) {
1799 		skip = tp->t_lognum - target;
1800 		STAILQ_FOREACH(log_entry, &tp->t_logs, tlm_queue)
1801 			if (!--skip)
1802 				break;
1803 		KASSERT(log_entry != NULL,
1804 		    ("%s: skipped through all entries!", __func__));
1805 		if (log_entry == NULL)
1806 			return;
1807 		while ((next = STAILQ_NEXT(log_entry, tlm_queue)) != NULL) {
1808 			STAILQ_REMOVE_AFTER(&tp->t_logs, log_entry, tlm_queue);
1809 			tcp_log_entry_refcnt_rem(next);
1810 			tcp_log_remove_log_cleanup(tp, next);
1811 #ifdef INVARIANTS
1812 			target--;
1813 #endif
1814 		}
1815 		KASSERT(target == 0,
1816 		    ("%s: After removing from tail, target was %d", __func__,
1817 			target));
1818 	} else if (tp->t_logstate == TCP_LOG_STATE_CONTINUAL) {
1819 		(void)tcp_log_dump_tp_logbuf(tp, "auto-dumped from continual",
1820 		    M_NOWAIT, false);
1821 	} else {
1822 		while ((log_entry = STAILQ_FIRST(&tp->t_logs)) != NULL &&
1823 		    target--)
1824 			tcp_log_remove_log_head(tp, log_entry);
1825 		KASSERT(target <= 0,
1826 		    ("%s: After removing from head, target was %d", __func__,
1827 			target));
1828 		KASSERT(tp->t_lognum > 0,
1829 		    ("%s: After removing from head, tp->t_lognum was %d",
1830 			__func__, target));
1831 		KASSERT(log_entry != NULL,
1832 		    ("%s: After removing from head, the tailq was empty",
1833 			__func__));
1834 	}
1835 }
1836 
1837 static inline int
1838 tcp_log_copyout(struct sockopt *sopt, void *src, void *dst, size_t len)
1839 {
1840 
1841 	if (sopt->sopt_td != NULL)
1842 		return (copyout(src, dst, len));
1843 	bcopy(src, dst, len);
1844 	return (0);
1845 }
1846 
1847 static int
1848 tcp_log_logs_to_buf(struct sockopt *sopt, struct tcp_log_stailq *log_tailqp,
1849     struct tcp_log_buffer **end, int count)
1850 {
1851 	struct tcp_log_buffer *out_entry;
1852 	struct tcp_log_mem *log_entry;
1853 	size_t entrysize;
1854 	int error;
1855 #ifdef INVARIANTS
1856 	int orig_count = count;
1857 #endif
1858 
1859 	/* Copy the data out. */
1860 	error = 0;
1861 	out_entry = (struct tcp_log_buffer *) sopt->sopt_val;
1862 	STAILQ_FOREACH(log_entry, log_tailqp, tlm_queue) {
1863 		count--;
1864 		KASSERT(count >= 0,
1865 		    ("%s:%d: Exceeded expected count (%d) processing list %p",
1866 		    __func__, __LINE__, orig_count, log_tailqp));
1867 
1868 #ifdef TCPLOG_DEBUG_COUNTERS
1869 		counter_u64_add(tcp_log_que_copyout, 1);
1870 #endif
1871 
1872 		/*
1873 		 * Skip copying out the header if it isn't present.
1874 		 * Instead, copy out zeros (to ensure we don't leak info).
1875 		 * TODO: Make sure we truly do zero everything we don't
1876 		 * explicitly set.
1877 		 */
1878 		if (log_entry->tlm_buf.tlb_eventflags & TLB_FLAG_HDR)
1879 			entrysize = sizeof(struct tcp_log_buffer);
1880 		else
1881 			entrysize = offsetof(struct tcp_log_buffer, tlb_th);
1882 		error = tcp_log_copyout(sopt, &log_entry->tlm_buf, out_entry,
1883 		    entrysize);
1884 		if (error)
1885 			break;
1886 		if (!(log_entry->tlm_buf.tlb_eventflags & TLB_FLAG_HDR)) {
1887 			error = tcp_log_copyout(sopt, zerobuf,
1888 			    ((uint8_t *)out_entry) + entrysize,
1889 			    sizeof(struct tcp_log_buffer) - entrysize);
1890 		}
1891 
1892 		/*
1893 		 * Copy out the verbose bit, if needed. Either way,
1894 		 * increment the output pointer the correct amount.
1895 		 */
1896 		if (log_entry->tlm_buf.tlb_eventflags & TLB_FLAG_VERBOSE) {
1897 			error = tcp_log_copyout(sopt, &log_entry->tlm_v,
1898 			    out_entry->tlb_verbose,
1899 			    sizeof(struct tcp_log_verbose));
1900 			if (error)
1901 				break;
1902 			out_entry = (struct tcp_log_buffer *)
1903 			    (((uint8_t *) (out_entry + 1)) +
1904 			    sizeof(struct tcp_log_verbose));
1905 		} else
1906 			out_entry++;
1907 	}
1908 	*end = out_entry;
1909 	KASSERT(error || count == 0,
1910 	    ("%s:%d: Less than expected count (%d) processing list %p"
1911 	    " (%d remain)", __func__, __LINE__, orig_count,
1912 	    log_tailqp, count));
1913 
1914 	return (error);
1915 }
1916 
1917 /*
1918  * Copy out the buffer. Note that we do incremental copying, so
1919  * sooptcopyout() won't work. However, the goal is to produce the same
1920  * end result as if we copied in the entire user buffer, updated it,
1921  * and then used sooptcopyout() to copy it out.
1922  *
1923  * NOTE: This should be called with a write lock on the PCB; however,
1924  * the function will drop it after it extracts the data from the TCPCB.
1925  */
1926 int
1927 tcp_log_getlogbuf(struct sockopt *sopt, struct tcpcb *tp)
1928 {
1929 	struct tcp_log_stailq log_tailq;
1930 	struct tcp_log_mem *log_entry, *log_next;
1931 	struct tcp_log_buffer *out_entry;
1932 	struct inpcb *inp;
1933 	size_t outsize, entrysize;
1934 	int error, outnum;
1935 
1936 	INP_WLOCK_ASSERT(tp->t_inpcb);
1937 	inp = tp->t_inpcb;
1938 
1939 	/*
1940 	 * Determine which log entries will fit in the buffer. As an
1941 	 * optimization, skip this if all the entries will clearly fit
1942 	 * in the buffer. (However, get an exact size if we are using
1943 	 * INVARIANTS.)
1944 	 */
1945 #ifndef INVARIANTS
1946 	if (sopt->sopt_valsize / (sizeof(struct tcp_log_buffer) +
1947 	    sizeof(struct tcp_log_verbose)) >= tp->t_lognum) {
1948 		log_entry = STAILQ_LAST(&tp->t_logs, tcp_log_mem, tlm_queue);
1949 		log_next = NULL;
1950 		outsize = 0;
1951 		outnum = tp->t_lognum;
1952 	} else {
1953 #endif
1954 		outsize = outnum = 0;
1955 		log_entry = NULL;
1956 		STAILQ_FOREACH(log_next, &tp->t_logs, tlm_queue) {
1957 			entrysize = sizeof(struct tcp_log_buffer);
1958 			if (log_next->tlm_buf.tlb_eventflags &
1959 			    TLB_FLAG_VERBOSE)
1960 				entrysize += sizeof(struct tcp_log_verbose);
1961 			if ((sopt->sopt_valsize - outsize) < entrysize)
1962 				break;
1963 			outsize += entrysize;
1964 			outnum++;
1965 			log_entry = log_next;
1966 		}
1967 		KASSERT(outsize <= sopt->sopt_valsize,
1968 		    ("%s: calculated output size (%zu) greater than available"
1969 			"space (%zu)", __func__, outsize, sopt->sopt_valsize));
1970 #ifndef INVARIANTS
1971 	}
1972 #endif
1973 
1974 	/*
1975 	 * Copy traditional sooptcopyout() behavior: if sopt->sopt_val
1976 	 * is NULL, silently skip the copy. However, in this case, we
1977 	 * will leave the list alone and return. Functionally, this
1978 	 * gives userspace a way to poll for an approximate buffer
1979 	 * size they will need to get the log entries.
1980 	 */
1981 	if (sopt->sopt_val == NULL) {
1982 		INP_WUNLOCK(inp);
1983 		if (outsize == 0) {
1984 			outsize = outnum * (sizeof(struct tcp_log_buffer) +
1985 			    sizeof(struct tcp_log_verbose));
1986 		}
1987 		if (sopt->sopt_valsize > outsize)
1988 			sopt->sopt_valsize = outsize;
1989 		return (0);
1990 	}
1991 
1992 	/*
1993 	 * Break apart the list. We'll save the ones we want to copy
1994 	 * out locally and remove them from the TCPCB list. We can
1995 	 * then drop the INPCB lock while we do the copyout.
1996 	 *
1997 	 * There are roughly three cases:
1998 	 * 1. There was nothing to copy out. That's easy: drop the
1999 	 * lock and return.
2000 	 * 2. We are copying out the entire list. Again, that's easy:
2001 	 * move the whole list.
2002 	 * 3. We are copying out a partial list. That's harder. We
2003 	 * need to update the list book-keeping entries.
2004 	 */
2005 	if (log_entry != NULL && log_next == NULL) {
2006 		/* Move entire list. */
2007 		KASSERT(outnum == tp->t_lognum,
2008 		    ("%s:%d: outnum (%d) should match tp->t_lognum (%d)",
2009 			__func__, __LINE__, outnum, tp->t_lognum));
2010 		log_tailq = tp->t_logs;
2011 		tp->t_lognum = 0;
2012 		STAILQ_INIT(&tp->t_logs);
2013 	} else if (log_entry != NULL) {
2014 		/* Move partial list. */
2015 		KASSERT(outnum < tp->t_lognum,
2016 		    ("%s:%d: outnum (%d) not less than tp->t_lognum (%d)",
2017 			__func__, __LINE__, outnum, tp->t_lognum));
2018 		STAILQ_FIRST(&log_tailq) = STAILQ_FIRST(&tp->t_logs);
2019 		STAILQ_FIRST(&tp->t_logs) = STAILQ_NEXT(log_entry, tlm_queue);
2020 		KASSERT(STAILQ_NEXT(log_entry, tlm_queue) != NULL,
2021 		    ("%s:%d: tp->t_logs is unexpectedly shorter than expected"
2022 		    "(tp: %p, log_tailq: %p, outnum: %d, tp->t_lognum: %d)",
2023 		    __func__, __LINE__, tp, &log_tailq, outnum, tp->t_lognum));
2024 		STAILQ_NEXT(log_entry, tlm_queue) = NULL;
2025 		log_tailq.stqh_last = &STAILQ_NEXT(log_entry, tlm_queue);
2026 		tp->t_lognum -= outnum;
2027 	} else
2028 		STAILQ_INIT(&log_tailq);
2029 
2030 	/* Drop the PCB lock. */
2031 	INP_WUNLOCK(inp);
2032 
2033 	/* Copy the data out. */
2034 	error = tcp_log_logs_to_buf(sopt, &log_tailq, &out_entry, outnum);
2035 
2036 	if (error) {
2037 		/* Restore list */
2038 		INP_WLOCK(inp);
2039 		if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0) {
2040 			tp = intotcpcb(inp);
2041 
2042 			/* Merge the two lists. */
2043 			STAILQ_CONCAT(&log_tailq, &tp->t_logs);
2044 			tp->t_logs = log_tailq;
2045 			tp->t_lognum += outnum;
2046 		}
2047 		INP_WUNLOCK(inp);
2048 	} else {
2049 		/* Sanity check entries */
2050 		KASSERT(((caddr_t)out_entry - (caddr_t)sopt->sopt_val)  ==
2051 		    outsize, ("%s: Actual output size (%zu) != "
2052 			"calculated output size (%zu)", __func__,
2053 			(size_t)((caddr_t)out_entry - (caddr_t)sopt->sopt_val),
2054 			outsize));
2055 
2056 		/* Free the entries we just copied out. */
2057 		STAILQ_FOREACH_SAFE(log_entry, &log_tailq, tlm_queue, log_next) {
2058 			tcp_log_entry_refcnt_rem(log_entry);
2059 			uma_zfree(tcp_log_zone, log_entry);
2060 		}
2061 	}
2062 
2063 	sopt->sopt_valsize = (size_t)((caddr_t)out_entry -
2064 	    (caddr_t)sopt->sopt_val);
2065 	return (error);
2066 }
2067 
2068 static void
2069 tcp_log_free_queue(struct tcp_log_dev_queue *param)
2070 {
2071 	struct tcp_log_dev_log_queue *entry;
2072 
2073 	KASSERT(param != NULL, ("%s: called with NULL param", __func__));
2074 	if (param == NULL)
2075 		return;
2076 
2077 	entry = (struct tcp_log_dev_log_queue *)param;
2078 
2079 	/* Free the entries. */
2080 	tcp_log_free_entries(&entry->tldl_entries, &entry->tldl_count);
2081 
2082 	/* Free the buffer, if it is allocated. */
2083 	if (entry->tldl_common.tldq_buf != NULL)
2084 		free(entry->tldl_common.tldq_buf, M_TCPLOGDEV);
2085 
2086 	/* Free the queue entry. */
2087 	free(entry, M_TCPLOGDEV);
2088 }
2089 
2090 static struct tcp_log_common_header *
2091 tcp_log_expandlogbuf(struct tcp_log_dev_queue *param)
2092 {
2093 	struct tcp_log_dev_log_queue *entry;
2094 	struct tcp_log_header *hdr;
2095 	uint8_t *end;
2096 	struct sockopt sopt;
2097 	int error;
2098 
2099 	entry = (struct tcp_log_dev_log_queue *)param;
2100 
2101 	/* Take a worst-case guess at space needs. */
2102 	sopt.sopt_valsize = sizeof(struct tcp_log_header) +
2103 	    entry->tldl_count * (sizeof(struct tcp_log_buffer) +
2104 	    sizeof(struct tcp_log_verbose));
2105 	hdr = malloc(sopt.sopt_valsize, M_TCPLOGDEV, M_NOWAIT);
2106 	if (hdr == NULL) {
2107 #ifdef TCPLOG_DEBUG_COUNTERS
2108 		counter_u64_add(tcp_log_que_fail5, entry->tldl_count);
2109 #endif
2110 		return (NULL);
2111 	}
2112 	sopt.sopt_val = hdr + 1;
2113 	sopt.sopt_valsize -= sizeof(struct tcp_log_header);
2114 	sopt.sopt_td = NULL;
2115 
2116 	error = tcp_log_logs_to_buf(&sopt, &entry->tldl_entries,
2117 	    (struct tcp_log_buffer **)&end, entry->tldl_count);
2118 	if (error) {
2119 		free(hdr, M_TCPLOGDEV);
2120 		return (NULL);
2121 	}
2122 
2123 	/* Free the entries. */
2124 	tcp_log_free_entries(&entry->tldl_entries, &entry->tldl_count);
2125 	entry->tldl_count = 0;
2126 
2127 	memset(hdr, 0, sizeof(struct tcp_log_header));
2128 	hdr->tlh_version = TCP_LOG_BUF_VER;
2129 	hdr->tlh_type = TCP_LOG_DEV_TYPE_BBR;
2130 	hdr->tlh_length = end - (uint8_t *)hdr;
2131 	hdr->tlh_ie = entry->tldl_ie;
2132 	hdr->tlh_af = entry->tldl_af;
2133 	getboottime(&hdr->tlh_offset);
2134 	strlcpy(hdr->tlh_id, entry->tldl_id, TCP_LOG_ID_LEN);
2135 	strlcpy(hdr->tlh_tag, entry->tldl_tag, TCP_LOG_TAG_LEN);
2136 	strlcpy(hdr->tlh_reason, entry->tldl_reason, TCP_LOG_REASON_LEN);
2137 	return ((struct tcp_log_common_header *)hdr);
2138 }
2139 
2140 /*
2141  * Queue the tcpcb's log buffer for transmission via the log buffer facility.
2142  *
2143  * NOTE: This should be called with a write lock on the PCB.
2144  *
2145  * how should be M_WAITOK or M_NOWAIT. If M_WAITOK, the function will drop
2146  * and reacquire the INP lock if it needs to do so.
2147  *
2148  * If force is false, this will only dump auto-logged sessions if
2149  * tcp_log_auto_all is true or if there is a log ID defined for the session.
2150  */
2151 int
2152 tcp_log_dump_tp_logbuf(struct tcpcb *tp, char *reason, int how, bool force)
2153 {
2154 	struct tcp_log_dev_log_queue *entry;
2155 	struct inpcb *inp;
2156 #ifdef TCPLOG_DEBUG_COUNTERS
2157 	int num_entries;
2158 #endif
2159 
2160 	inp = tp->t_inpcb;
2161 	INP_WLOCK_ASSERT(inp);
2162 
2163 	/* If there are no log entries, there is nothing to do. */
2164 	if (tp->t_lognum == 0)
2165 		return (0);
2166 
2167 	/* Check for a log ID. */
2168 	if (tp->t_lib == NULL && (tp->t_flags2 & TF2_LOG_AUTO) &&
2169 	    !tcp_log_auto_all && !force) {
2170 		struct tcp_log_mem *log_entry;
2171 
2172 		/*
2173 		 * We needed a log ID and none was found. Free the log entries
2174 		 * and return success. Also, cancel further logging. If the
2175 		 * session doesn't have a log ID by now, we'll assume it isn't
2176 		 * going to get one.
2177 		 */
2178 		while ((log_entry = STAILQ_FIRST(&tp->t_logs)) != NULL)
2179 			tcp_log_remove_log_head(tp, log_entry);
2180 		KASSERT(tp->t_lognum == 0,
2181 		    ("%s: After freeing entries, tp->t_lognum=%d (expected 0)",
2182 			__func__, tp->t_lognum));
2183 		tp->t_logstate = TCP_LOG_STATE_OFF;
2184 		return (0);
2185 	}
2186 
2187 	/*
2188 	 * Allocate memory. If we must wait, we'll need to drop the locks
2189 	 * and reacquire them (and do all the related business that goes
2190 	 * along with that).
2191 	 */
2192 	entry = malloc(sizeof(struct tcp_log_dev_log_queue), M_TCPLOGDEV,
2193 	    M_NOWAIT);
2194 	if (entry == NULL && (how & M_NOWAIT)) {
2195 #ifdef TCPLOG_DEBUG_COUNTERS
2196 		counter_u64_add(tcp_log_que_fail3, 1);
2197 #endif
2198 		return (ENOBUFS);
2199 	}
2200 	if (entry == NULL) {
2201 		INP_WUNLOCK(inp);
2202 		entry = malloc(sizeof(struct tcp_log_dev_log_queue),
2203 		    M_TCPLOGDEV, M_WAITOK);
2204 		INP_WLOCK(inp);
2205 		/*
2206 		 * Note that this check is slightly overly-restrictive in
2207 		 * that the TCB can survive either of these events.
2208 		 * However, there is currently not a good way to ensure
2209 		 * that is the case. So, if we hit this M_WAIT path, we
2210 		 * may end up dropping some entries. That seems like a
2211 		 * small price to pay for safety.
2212 		 */
2213 		if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
2214 			free(entry, M_TCPLOGDEV);
2215 #ifdef TCPLOG_DEBUG_COUNTERS
2216 			counter_u64_add(tcp_log_que_fail2, 1);
2217 #endif
2218 			return (ECONNRESET);
2219 		}
2220 		tp = intotcpcb(inp);
2221 		if (tp->t_lognum == 0) {
2222 			free(entry, M_TCPLOGDEV);
2223 			return (0);
2224 		}
2225 	}
2226 
2227 	/* Fill in the unique parts of the queue entry. */
2228 	if (tp->t_lib != NULL) {
2229 		strlcpy(entry->tldl_id, tp->t_lib->tlb_id, TCP_LOG_ID_LEN);
2230 		strlcpy(entry->tldl_tag, tp->t_lib->tlb_tag, TCP_LOG_TAG_LEN);
2231 	} else {
2232 		strlcpy(entry->tldl_id, "UNKNOWN", TCP_LOG_ID_LEN);
2233 		strlcpy(entry->tldl_tag, "UNKNOWN", TCP_LOG_TAG_LEN);
2234 	}
2235 	if (reason != NULL)
2236 		strlcpy(entry->tldl_reason, reason, TCP_LOG_REASON_LEN);
2237 	else
2238 		strlcpy(entry->tldl_reason, "UNKNOWN", TCP_LOG_ID_LEN);
2239 	entry->tldl_ie = inp->inp_inc.inc_ie;
2240 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
2241 		entry->tldl_af = AF_INET6;
2242 	else
2243 		entry->tldl_af = AF_INET;
2244 	entry->tldl_entries = tp->t_logs;
2245 	entry->tldl_count = tp->t_lognum;
2246 
2247 	/* Fill in the common parts of the queue entry. */
2248 	entry->tldl_common.tldq_buf = NULL;
2249 	entry->tldl_common.tldq_xform = tcp_log_expandlogbuf;
2250 	entry->tldl_common.tldq_dtor = tcp_log_free_queue;
2251 
2252 	/* Clear the log data from the TCPCB. */
2253 #ifdef TCPLOG_DEBUG_COUNTERS
2254 	num_entries = tp->t_lognum;
2255 #endif
2256 	tp->t_lognum = 0;
2257 	STAILQ_INIT(&tp->t_logs);
2258 
2259 	/* Add the entry. If no one is listening, free the entry. */
2260 	if (tcp_log_dev_add_log((struct tcp_log_dev_queue *)entry)) {
2261 		tcp_log_free_queue((struct tcp_log_dev_queue *)entry);
2262 #ifdef TCPLOG_DEBUG_COUNTERS
2263 		counter_u64_add(tcp_log_que_fail1, num_entries);
2264 	} else {
2265 		counter_u64_add(tcp_log_queued, num_entries);
2266 #endif
2267 	}
2268 	return (0);
2269 }
2270 
2271 /*
2272  * Queue the log_id_node's log buffers for transmission via the log buffer
2273  * facility.
2274  *
2275  * NOTE: This should be called with the bucket locked and referenced.
2276  *
2277  * how should be M_WAITOK or M_NOWAIT. If M_WAITOK, the function will drop
2278  * and reacquire the bucket lock if it needs to do so. (The caller must
2279  * ensure that the tln is no longer on any lists so no one else will mess
2280  * with this while the lock is dropped!)
2281  */
2282 static int
2283 tcp_log_dump_node_logbuf(struct tcp_log_id_node *tln, char *reason, int how)
2284 {
2285 	struct tcp_log_dev_log_queue *entry;
2286 	struct tcp_log_id_bucket *tlb;
2287 
2288 	tlb = tln->tln_bucket;
2289 	TCPID_BUCKET_LOCK_ASSERT(tlb);
2290 	KASSERT(tlb->tlb_refcnt > 0,
2291 	    ("%s:%d: Called with unreferenced bucket (tln=%p, tlb=%p)",
2292 	    __func__, __LINE__, tln, tlb));
2293 	KASSERT(tln->tln_closed,
2294 	    ("%s:%d: Called for node with tln_closed==false (tln=%p)",
2295 	    __func__, __LINE__, tln));
2296 
2297 	/* If there are no log entries, there is nothing to do. */
2298 	if (tln->tln_count == 0)
2299 		return (0);
2300 
2301 	/*
2302 	 * Allocate memory. If we must wait, we'll need to drop the locks
2303 	 * and reacquire them (and do all the related business that goes
2304 	 * along with that).
2305 	 */
2306 	entry = malloc(sizeof(struct tcp_log_dev_log_queue), M_TCPLOGDEV,
2307 	    M_NOWAIT);
2308 	if (entry == NULL && (how & M_NOWAIT))
2309 		return (ENOBUFS);
2310 	if (entry == NULL) {
2311 		TCPID_BUCKET_UNLOCK(tlb);
2312 		entry = malloc(sizeof(struct tcp_log_dev_log_queue),
2313 		    M_TCPLOGDEV, M_WAITOK);
2314 		TCPID_BUCKET_LOCK(tlb);
2315 	}
2316 
2317 	/* Fill in the common parts of the queue entry.. */
2318 	entry->tldl_common.tldq_buf = NULL;
2319 	entry->tldl_common.tldq_xform = tcp_log_expandlogbuf;
2320 	entry->tldl_common.tldq_dtor = tcp_log_free_queue;
2321 
2322 	/* Fill in the unique parts of the queue entry. */
2323 	strlcpy(entry->tldl_id, tlb->tlb_id, TCP_LOG_ID_LEN);
2324 	strlcpy(entry->tldl_tag, tlb->tlb_tag, TCP_LOG_TAG_LEN);
2325 	if (reason != NULL)
2326 		strlcpy(entry->tldl_reason, reason, TCP_LOG_REASON_LEN);
2327 	else
2328 		strlcpy(entry->tldl_reason, "UNKNOWN", TCP_LOG_ID_LEN);
2329 	entry->tldl_ie = tln->tln_ie;
2330 	entry->tldl_entries = tln->tln_entries;
2331 	entry->tldl_count = tln->tln_count;
2332 	entry->tldl_af = tln->tln_af;
2333 
2334 	/* Add the entry. If no one is listening, free the entry. */
2335 	if (tcp_log_dev_add_log((struct tcp_log_dev_queue *)entry))
2336 		tcp_log_free_queue((struct tcp_log_dev_queue *)entry);
2337 
2338 	return (0);
2339 }
2340 
2341 /*
2342  * Queue the log buffers for all sessions in a bucket for transmissions via
2343  * the log buffer facility.
2344  *
2345  * NOTE: This should be called with a locked bucket; however, the function
2346  * will drop the lock.
2347  */
2348 #define	LOCAL_SAVE	10
2349 static void
2350 tcp_log_dumpbucketlogs(struct tcp_log_id_bucket *tlb, char *reason)
2351 {
2352 	struct tcp_log_id_node local_entries[LOCAL_SAVE];
2353 	struct inpcb *inp;
2354 	struct tcpcb *tp;
2355 	struct tcp_log_id_node *cur_tln, *prev_tln, *tmp_tln;
2356 	int i, num_local_entries, tree_locked;
2357 	bool expireq_locked;
2358 
2359 	TCPID_BUCKET_LOCK_ASSERT(tlb);
2360 
2361 	/*
2362 	 * Take a reference on the bucket to keep it from disappearing until
2363 	 * we are done.
2364 	 */
2365 	TCPID_BUCKET_REF(tlb);
2366 
2367 	/*
2368 	 * We'll try to create these without dropping locks. However, we
2369 	 * might very well need to drop locks to get memory. If that's the
2370 	 * case, we'll save up to 10 on the stack, and sacrifice the rest.
2371 	 * (Otherwise, we need to worry about finding our place again in a
2372 	 * potentially changed list. It just doesn't seem worth the trouble
2373 	 * to do that.
2374 	 */
2375 	expireq_locked = false;
2376 	num_local_entries = 0;
2377 	prev_tln = NULL;
2378 	tree_locked = TREE_UNLOCKED;
2379 	SLIST_FOREACH_SAFE(cur_tln, &tlb->tlb_head, tln_list, tmp_tln) {
2380 		/*
2381 		 * If this isn't associated with a TCPCB, we can pull it off
2382 		 * the list now. We need to be careful that the expire timer
2383 		 * hasn't already taken ownership (tln_expiretime == SBT_MAX).
2384 		 * If so, we let the expire timer code free the data.
2385 		 */
2386 		if (cur_tln->tln_closed) {
2387 no_inp:
2388 			/*
2389 			 * Get the expireq lock so we can get a consistent
2390 			 * read of tln_expiretime and so we can remove this
2391 			 * from the expireq.
2392 			 */
2393 			if (!expireq_locked) {
2394 				TCPLOG_EXPIREQ_LOCK();
2395 				expireq_locked = true;
2396 			}
2397 
2398 			/*
2399 			 * We ignore entries with tln_expiretime == SBT_MAX.
2400 			 * The expire timer code already owns those.
2401 			 */
2402 			KASSERT(cur_tln->tln_expiretime > (sbintime_t) 0,
2403 			    ("%s:%d: node on the expire queue without positive "
2404 			    "expire time", __func__, __LINE__));
2405 			if (cur_tln->tln_expiretime == SBT_MAX) {
2406 				prev_tln = cur_tln;
2407 				continue;
2408 			}
2409 
2410 			/* Remove the entry from the expireq. */
2411 			STAILQ_REMOVE(&tcp_log_expireq_head, cur_tln,
2412 			    tcp_log_id_node, tln_expireq);
2413 
2414 			/* Remove the entry from the bucket. */
2415 			if (prev_tln != NULL)
2416 				SLIST_REMOVE_AFTER(prev_tln, tln_list);
2417 			else
2418 				SLIST_REMOVE_HEAD(&tlb->tlb_head, tln_list);
2419 
2420 			/*
2421 			 * Drop the INP and bucket reference counts. Due to
2422 			 * lock-ordering rules, we need to drop the expire
2423 			 * queue lock.
2424 			 */
2425 			TCPLOG_EXPIREQ_UNLOCK();
2426 			expireq_locked = false;
2427 
2428 			/* Drop the INP reference. */
2429 			INP_WLOCK(cur_tln->tln_inp);
2430 			if (!in_pcbrele_wlocked(cur_tln->tln_inp))
2431 				INP_WUNLOCK(cur_tln->tln_inp);
2432 
2433 			if (tcp_log_unref_bucket(tlb, &tree_locked, NULL)) {
2434 #ifdef INVARIANTS
2435 				panic("%s: Bucket refcount unexpectedly 0.",
2436 				    __func__);
2437 #endif
2438 				/*
2439 				 * Recover as best we can: free the entry we
2440 				 * own.
2441 				 */
2442 				tcp_log_free_entries(&cur_tln->tln_entries,
2443 				    &cur_tln->tln_count);
2444 				uma_zfree(tcp_log_node_zone, cur_tln);
2445 				goto done;
2446 			}
2447 
2448 			if (tcp_log_dump_node_logbuf(cur_tln, reason,
2449 			    M_NOWAIT)) {
2450 				/*
2451 				 * If we have sapce, save the entries locally.
2452 				 * Otherwise, free them.
2453 				 */
2454 				if (num_local_entries < LOCAL_SAVE) {
2455 					local_entries[num_local_entries] =
2456 					    *cur_tln;
2457 					num_local_entries++;
2458 				} else {
2459 					tcp_log_free_entries(
2460 					    &cur_tln->tln_entries,
2461 					    &cur_tln->tln_count);
2462 				}
2463 			}
2464 
2465 			/* No matter what, we are done with the node now. */
2466 			uma_zfree(tcp_log_node_zone, cur_tln);
2467 
2468 			/*
2469 			 * Because we removed this entry from the list, prev_tln
2470 			 * (which tracks the previous entry still on the tlb
2471 			 * list) remains unchanged.
2472 			 */
2473 			continue;
2474 		}
2475 
2476 		/*
2477 		 * If we get to this point, the session data is still held in
2478 		 * the TCPCB. So, we need to pull the data out of that.
2479 		 *
2480 		 * We will need to drop the expireq lock so we can lock the INP.
2481 		 * We can then try to extract the data the "easy" way. If that
2482 		 * fails, we'll save the log entries for later.
2483 		 */
2484 		if (expireq_locked) {
2485 			TCPLOG_EXPIREQ_UNLOCK();
2486 			expireq_locked = false;
2487 		}
2488 
2489 		/* Lock the INP and then re-check the state. */
2490 		inp = cur_tln->tln_inp;
2491 		INP_WLOCK(inp);
2492 		/*
2493 		 * If we caught this while it was transitioning, the data
2494 		 * might have moved from the TCPCB to the tln (signified by
2495 		 * setting tln_closed to true. If so, treat this like an
2496 		 * inactive connection.
2497 		 */
2498 		if (cur_tln->tln_closed) {
2499 			/*
2500 			 * It looks like we may have caught this connection
2501 			 * while it was transitioning from active to inactive.
2502 			 * Treat this like an inactive connection.
2503 			 */
2504 			INP_WUNLOCK(inp);
2505 			goto no_inp;
2506 		}
2507 
2508 		/*
2509 		 * Try to dump the data from the tp without dropping the lock.
2510 		 * If this fails, try to save off the data locally.
2511 		 */
2512 		tp = cur_tln->tln_tp;
2513 		if (tcp_log_dump_tp_logbuf(tp, reason, M_NOWAIT, true) &&
2514 		    num_local_entries < LOCAL_SAVE) {
2515 			tcp_log_move_tp_to_node(tp,
2516 			    &local_entries[num_local_entries]);
2517 			local_entries[num_local_entries].tln_closed = 1;
2518 			KASSERT(local_entries[num_local_entries].tln_bucket ==
2519 			    tlb, ("%s: %d: bucket mismatch for node %p",
2520 			    __func__, __LINE__, cur_tln));
2521 			num_local_entries++;
2522 		}
2523 
2524 		INP_WUNLOCK(inp);
2525 
2526 		/*
2527 		 * We are goint to leave the current tln on the list. It will
2528 		 * become the previous tln.
2529 		 */
2530 		prev_tln = cur_tln;
2531 	}
2532 
2533 	/* Drop our locks, if any. */
2534 	KASSERT(tree_locked == TREE_UNLOCKED,
2535 	    ("%s: %d: tree unexpectedly locked", __func__, __LINE__));
2536 	switch (tree_locked) {
2537 	case TREE_WLOCKED:
2538 		TCPID_TREE_WUNLOCK();
2539 		tree_locked = TREE_UNLOCKED;
2540 		break;
2541 	case TREE_RLOCKED:
2542 		TCPID_TREE_RUNLOCK();
2543 		tree_locked = TREE_UNLOCKED;
2544 		break;
2545 	}
2546 	if (expireq_locked) {
2547 		TCPLOG_EXPIREQ_UNLOCK();
2548 		expireq_locked = false;
2549 	}
2550 
2551 	/*
2552 	 * Try again for any saved entries. tcp_log_dump_node_logbuf() is
2553 	 * guaranteed to free the log entries within the node. And, since
2554 	 * the node itself is on our stack, we don't need to free it.
2555 	 */
2556 	for (i = 0; i < num_local_entries; i++)
2557 		tcp_log_dump_node_logbuf(&local_entries[i], reason, M_WAITOK);
2558 
2559 	/* Drop our reference. */
2560 	if (!tcp_log_unref_bucket(tlb, &tree_locked, NULL))
2561 		TCPID_BUCKET_UNLOCK(tlb);
2562 
2563 done:
2564 	/* Drop our locks, if any. */
2565 	switch (tree_locked) {
2566 	case TREE_WLOCKED:
2567 		TCPID_TREE_WUNLOCK();
2568 		break;
2569 	case TREE_RLOCKED:
2570 		TCPID_TREE_RUNLOCK();
2571 		break;
2572 	}
2573 	if (expireq_locked)
2574 		TCPLOG_EXPIREQ_UNLOCK();
2575 }
2576 #undef	LOCAL_SAVE
2577 
2578 /*
2579  * Queue the log buffers for all sessions in a bucket for transmissions via
2580  * the log buffer facility.
2581  *
2582  * NOTE: This should be called with a locked INP; however, the function
2583  * will drop the lock.
2584  */
2585 void
2586 tcp_log_dump_tp_bucket_logbufs(struct tcpcb *tp, char *reason)
2587 {
2588 	struct tcp_log_id_bucket *tlb;
2589 	int tree_locked;
2590 
2591 	/* Figure out our bucket and lock it. */
2592 	INP_WLOCK_ASSERT(tp->t_inpcb);
2593 	tlb = tp->t_lib;
2594 	if (tlb == NULL) {
2595 		/*
2596 		 * No bucket; treat this like a request to dump a single
2597 		 * session's traces.
2598 		 */
2599 		(void)tcp_log_dump_tp_logbuf(tp, reason, M_WAITOK, true);
2600 		INP_WUNLOCK(tp->t_inpcb);
2601 		return;
2602 	}
2603 	TCPID_BUCKET_REF(tlb);
2604 	INP_WUNLOCK(tp->t_inpcb);
2605 	TCPID_BUCKET_LOCK(tlb);
2606 
2607 	/* If we are the last reference, we have nothing more to do here. */
2608 	tree_locked = TREE_UNLOCKED;
2609 	if (tcp_log_unref_bucket(tlb, &tree_locked, NULL)) {
2610 		switch (tree_locked) {
2611 		case TREE_WLOCKED:
2612 			TCPID_TREE_WUNLOCK();
2613 			break;
2614 		case TREE_RLOCKED:
2615 			TCPID_TREE_RUNLOCK();
2616 			break;
2617 		}
2618 		return;
2619 	}
2620 
2621 	/* Turn this over to tcp_log_dumpbucketlogs() to finish the work. */
2622 	tcp_log_dumpbucketlogs(tlb, reason);
2623 }
2624 
2625 /*
2626  * Mark the end of a flow with the current stack. A stack can add
2627  * stack-specific info to this trace event by overriding this
2628  * function (see bbr_log_flowend() for example).
2629  */
2630 void
2631 tcp_log_flowend(struct tcpcb *tp)
2632 {
2633 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2634 		struct socket *so = tp->t_inpcb->inp_socket;
2635 		TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd,
2636 				TCP_LOG_FLOWEND, 0, 0, NULL, false);
2637 	}
2638 }
2639