xref: /xnu-12377.1.9/bsd/net/classq/classq_fq_codel.c (revision f6217f891ac0bb64f3d375211650a4c1ff8ca1ea)
1 /*
2  * Copyright (c) 2016-2021 Apple Inc. All rights reserved.
3  *
4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5  *
6  * This file contains Original Code and/or Modifications of Original Code
7  * as defined in and that are subject to the Apple Public Source License
8  * Version 2.0 (the 'License'). You may not use this file except in
9  * compliance with the License. The rights granted to you under the License
10  * may not be used to create, or enable the creation or redistribution of,
11  * unlawful or unlicensed copies of an Apple operating system, or to
12  * circumvent, violate, or enable the circumvention or violation of, any
13  * terms of an Apple operating system software license agreement.
14  *
15  * Please obtain a copy of the License at
16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
17  *
18  * The Original Code and all software distributed under the License are
19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23  * Please see the License for the specific language governing rights and
24  * limitations under the License.
25  *
26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27  */
28 
29 /*
30  * The migration of flow queue between the different states is summarised in
31  * the below state diagram. (RFC 8290)
32  *
33  * +-----------------+                +------------------+
34  * |                 |     Empty      |                  |
35  * |     Empty       |<---------------+       Old        +----+
36  * |                 |                |                  |    |
37  * +-------+---------+                +------------------+    |
38  *         |                             ^            ^       |Credits
39  *         |Arrival                      |            |       |Exhausted
40  *         v                             |            |       |
41  * +-----------------+                   |            |       |
42  * |                 |      Empty or     |            |       |
43  * |      New        +-------------------+            +-------+
44  * |                 | Credits Exhausted
45  * +-----------------+
46  *
47  * In this implementation of FQ-CODEL, flow queue is a dynamically allocated
48  * object. An active flow queue goes through the above cycle of state
49  * transitions very often. To avoid the cost of frequent flow queue object
50  * allocation/free, this implementation retains the flow queue object in
51  * [Empty] state on an Empty flow queue list with an active reference in flow
52  * queue hash table. The flow queue objects on the Empty flow queue list have
53  * an associated age and are purged accordingly.
54  */
55 
56 #include <sys/cdefs.h>
57 #include <sys/param.h>
58 #include <sys/mbuf.h>
59 #include <sys/socket.h>
60 #include <sys/sockio.h>
61 #include <sys/systm.h>
62 #include <sys/syslog.h>
63 #include <sys/proc.h>
64 #include <sys/errno.h>
65 #include <sys/kernel.h>
66 #include <sys/kauth.h>
67 #include <sys/sdt.h>
68 #include <kern/zalloc.h>
69 #include <netinet/in.h>
70 
71 #include <net/classq/classq.h>
72 #include <net/classq/if_classq.h>
73 #include <net/pktsched/pktsched.h>
74 #include <net/pktsched/pktsched_fq_codel.h>
75 #include <net/classq/classq_fq_codel.h>
76 #include <net/droptap.h>
77 
78 #include <netinet/tcp_var.h>
79 
80 #define FQ_ZONE_MAX     (32 * 1024)     /* across all interfaces */
81 
82 #define DTYPE_NODROP    0       /* no drop */
83 #define DTYPE_FORCED    1       /* a "forced" drop */
84 #define DTYPE_EARLY     2       /* an "unforced" (early) drop */
85 
86 static uint32_t pkt_compressor = 1;
87 static uint64_t l4s_ce_threshold = 0; /* in usec */
88 static uint32_t l4s_local_ce_report = 0;
89 static uint64_t pkt_pacing_leeway = 0; /* in usec */
90 static uint64_t max_pkt_pacing_interval = 3 * NSEC_PER_SEC;
91 static uint64_t l4s_min_delay_threshold = 20 * NSEC_PER_MSEC; /* 20 ms */
92 #if (DEBUG || DEVELOPMENT)
93 SYSCTL_NODE(_net_classq, OID_AUTO, flow_q, CTLFLAG_RW | CTLFLAG_LOCKED,
94     0, "FQ-CODEL parameters");
95 
96 SYSCTL_UINT(_net_classq_flow_q, OID_AUTO, pkt_compressor,
97     CTLFLAG_RW | CTLFLAG_LOCKED, &pkt_compressor, 0, "enable pkt compression");
98 
99 SYSCTL_QUAD(_net_classq, OID_AUTO, l4s_ce_threshold,
100     CTLFLAG_RW | CTLFLAG_LOCKED, &l4s_ce_threshold,
101     "L4S CE threshold");
102 
103 SYSCTL_UINT(_net_classq_flow_q, OID_AUTO, l4s_local_ce_report,
104     CTLFLAG_RW | CTLFLAG_LOCKED, &l4s_local_ce_report, 0,
105     "enable L4S local CE report");
106 
107 SYSCTL_QUAD(_net_classq_flow_q, OID_AUTO, pkt_pacing_leeway,
108     CTLFLAG_RW | CTLFLAG_LOCKED, &pkt_pacing_leeway, "packet pacing leeway");
109 
110 SYSCTL_QUAD(_net_classq_flow_q, OID_AUTO, max_pkt_pacing_interval,
111     CTLFLAG_RW | CTLFLAG_LOCKED, &max_pkt_pacing_interval, "max packet pacing interval");
112 
113 SYSCTL_QUAD(_net_classq_flow_q, OID_AUTO, l4s_min_delay_threshold,
114     CTLFLAG_RW | CTLFLAG_LOCKED, &l4s_min_delay_threshold, "l4s min delay threshold");
115 #endif /* (DEBUG || DEVELOPMENT) */
116 
117 void
fq_codel_init(void)118 fq_codel_init(void)
119 {
120 	static_assert(AQM_KTRACE_AON_FLOW_HIGH_DELAY == 0x8300004);
121 	static_assert(AQM_KTRACE_AON_THROTTLE == 0x8300008);
122 	static_assert(AQM_KTRACE_AON_FLOW_OVERWHELMING == 0x830000c);
123 	static_assert(AQM_KTRACE_AON_FLOW_DQ_STALL == 0x8300010);
124 
125 	static_assert(AQM_KTRACE_STATS_FLOW_ENQUEUE == 0x8310004);
126 	static_assert(AQM_KTRACE_STATS_FLOW_DEQUEUE == 0x8310008);
127 	static_assert(AQM_KTRACE_STATS_FLOW_CTL == 0x831000c);
128 	static_assert(AQM_KTRACE_STATS_FLOW_ALLOC == 0x8310010);
129 	static_assert(AQM_KTRACE_STATS_FLOW_DESTROY == 0x8310014);
130 	static_assert(AQM_KTRACE_STATS_FLOW_REPORT_CE == 0x8310018);
131 	static_assert(AQM_KTRACE_STATS_GET_QLEN == 0x831001c);
132 	static_assert(AQM_KTRACE_TX_NOT_READY == 0x8310020);
133 	static_assert(AQM_KTRACE_TX_PACEMAKER == 0x8310024);
134 	static_assert(AQM_KTRACE_PKT_DROP == 0x8310028);
135 	static_assert(AQM_KTRACE_OK_TO_DROP == 0x831002c);
136 	static_assert(AQM_KTRACE_CONGESTION_INC == 0x8310030);
137 	static_assert(AQM_KTRACE_CONGESTION_NOTIFIED == 0x8310034);
138 }
139 
140 fq_t *
fq_alloc(classq_pkt_type_t ptype)141 fq_alloc(classq_pkt_type_t ptype)
142 {
143 	fq_t *fq = NULL;
144 
145 	fq = kalloc_type(fq_t, Z_WAITOK_ZERO);
146 	if (ptype == QP_MBUF) {
147 		MBUFQ_INIT(&fq->fq_mbufq);
148 	}
149 #if SKYWALK
150 	else {
151 		VERIFY(ptype == QP_PACKET);
152 		KPKTQ_INIT(&fq->fq_kpktq);
153 	}
154 #endif /* SKYWALK */
155 	CLASSQ_PKT_INIT(&fq->fq_dq_head);
156 	CLASSQ_PKT_INIT(&fq->fq_dq_tail);
157 	fq->fq_in_dqlist = false;
158 
159 	return fq;
160 }
161 
162 void
fq_destroy(fq_t * fq,classq_pkt_type_t ptype)163 fq_destroy(fq_t *fq, classq_pkt_type_t ptype)
164 {
165 	VERIFY(!fq->fq_in_dqlist);
166 	VERIFY(fq_empty(fq, ptype));
167 	VERIFY(!(fq->fq_flags & (FQF_NEW_FLOW | FQF_OLD_FLOW |
168 	    FQF_EMPTY_FLOW)));
169 	VERIFY(fq->fq_bytes == 0);
170 	kfree_type(fq_t, fq);
171 }
172 
173 static inline void
fq_detect_dequeue_stall(fq_if_t * fqs,fq_t * flowq,fq_if_classq_t * fq_cl,u_int64_t * now)174 fq_detect_dequeue_stall(fq_if_t *fqs, fq_t *flowq, fq_if_classq_t *fq_cl,
175     u_int64_t *now)
176 {
177 	u_int64_t maxgetqtime, update_interval;
178 	if (FQ_IS_DELAY_HIGH(flowq) || flowq->fq_getqtime == 0 ||
179 	    fq_empty(flowq, fqs->fqs_ptype) ||
180 	    flowq->fq_bytes < FQ_MIN_FC_THRESHOLD_BYTES) {
181 		return;
182 	}
183 
184 	update_interval = FQ_UPDATE_INTERVAL(flowq);
185 	maxgetqtime = flowq->fq_getqtime + update_interval;
186 	if ((*now) > maxgetqtime) {
187 		/*
188 		 * there was no dequeue in an update interval worth of
189 		 * time. It means that the queue is stalled.
190 		 */
191 		FQ_SET_DELAY_HIGH(flowq);
192 		fq_cl->fcl_stat.fcl_dequeue_stall++;
193 		os_log_error(OS_LOG_DEFAULT, "%s:num: %d, "
194 		    "scidx: %d, flow: 0x%x, iface: %s grp: %hhu", __func__,
195 		    fq_cl->fcl_stat.fcl_dequeue_stall, flowq->fq_sc_index,
196 		    flowq->fq_flowhash, if_name(fqs->fqs_ifq->ifcq_ifp),
197 		    FQ_GROUP(flowq)->fqg_index);
198 		KDBG(AQM_KTRACE_AON_FLOW_DQ_STALL, flowq->fq_flowhash,
199 		    AQM_KTRACE_FQ_GRP_SC_IDX(flowq), flowq->fq_bytes,
200 		    (*now) - flowq->fq_getqtime);
201 	}
202 }
203 
204 void
fq_head_drop(fq_if_t * fqs,fq_t * fq)205 fq_head_drop(fq_if_t *fqs, fq_t *fq)
206 {
207 	pktsched_pkt_t pkt;
208 	volatile uint32_t *__single pkt_flags;
209 	uint64_t *__single pkt_timestamp;
210 	struct ifclassq *ifq = fqs->fqs_ifq;
211 
212 	_PKTSCHED_PKT_INIT(&pkt);
213 	fq_getq_flow_internal(fqs, fq, &pkt);
214 	if (pkt.pktsched_pkt_mbuf == NULL) {
215 		return;
216 	}
217 
218 	pktsched_get_pkt_vars(&pkt, &pkt_flags, &pkt_timestamp, NULL, NULL,
219 	    NULL, NULL, NULL);
220 
221 	*pkt_timestamp = 0;
222 	switch (pkt.pktsched_ptype) {
223 	case QP_MBUF:
224 		*pkt_flags &= ~PKTF_PRIV_GUARDED;
225 		break;
226 #if SKYWALK
227 	case QP_PACKET:
228 		/* sanity check */
229 		ASSERT((*pkt_flags & ~PKT_F_COMMON_MASK) == 0);
230 		break;
231 #endif /* SKYWALK */
232 	default:
233 		VERIFY(0);
234 		/* NOTREACHED */
235 		__builtin_unreachable();
236 	}
237 
238 	IFCQ_DROP_ADD(ifq, 1, pktsched_get_pkt_len(&pkt));
239 	IFCQ_CONVERT_LOCK(ifq);
240 	pktsched_free_pkt(&pkt);
241 }
242 
243 
244 static int
fq_compressor(fq_if_t * fqs,fq_t * fq,fq_if_classq_t * fq_cl,pktsched_pkt_t * pkt)245 fq_compressor(fq_if_t *fqs, fq_t *fq, fq_if_classq_t *fq_cl,
246     pktsched_pkt_t *pkt)
247 {
248 	classq_pkt_type_t ptype = fqs->fqs_ptype;
249 	uint32_t comp_gencnt = 0;
250 	uint64_t *__single pkt_timestamp;
251 	uint64_t old_timestamp = 0;
252 	uint32_t old_pktlen = 0;
253 	struct ifclassq *ifq = fqs->fqs_ifq;
254 
255 	if (__improbable(pkt_compressor == 0)) {
256 		return 0;
257 	}
258 
259 	pktsched_get_pkt_vars(pkt, NULL, &pkt_timestamp, NULL, NULL, NULL,
260 	    &comp_gencnt, NULL);
261 
262 	if (comp_gencnt == 0) {
263 		return 0;
264 	}
265 
266 	fq_cl->fcl_stat.fcl_pkts_compressible++;
267 
268 	if (fq_empty(fq, fqs->fqs_ptype)) {
269 		return 0;
270 	}
271 
272 	if (ptype == QP_MBUF) {
273 		struct mbuf *m = MBUFQ_LAST(&fq->fq_mbufq);
274 
275 		if (comp_gencnt != m->m_pkthdr.comp_gencnt) {
276 			return 0;
277 		}
278 
279 		/* If we got until here, we should merge/replace the segment */
280 		MBUFQ_REMOVE(&fq->fq_mbufq, m);
281 		old_pktlen = m_pktlen(m);
282 		old_timestamp = m->m_pkthdr.pkt_timestamp;
283 
284 		IFCQ_CONVERT_LOCK(fqs->fqs_ifq);
285 
286 		if (__improbable(droptap_verbose > 0)) {
287 			droptap_output_mbuf(m, DROP_REASON_AQM_COMPRESSED, NULL, 0, 0,
288 			    fqs->fqs_ifq->ifcq_ifp);
289 		}
290 
291 		m_freem(m);
292 	}
293 #if SKYWALK
294 	else {
295 		struct __kern_packet *kpkt = KPKTQ_LAST(&fq->fq_kpktq);
296 
297 		if (comp_gencnt != kpkt->pkt_comp_gencnt) {
298 			return 0;
299 		}
300 
301 		/* If we got until here, we should merge/replace the segment */
302 		KPKTQ_REMOVE(&fq->fq_kpktq, kpkt);
303 		old_pktlen = kpkt->pkt_length;
304 		old_timestamp = kpkt->pkt_timestamp;
305 
306 		IFCQ_CONVERT_LOCK(fqs->fqs_ifq);
307 
308 		if (__improbable(droptap_verbose > 0)) {
309 			droptap_output_packet(SK_PKT2PH(kpkt), DROP_REASON_AQM_COMPRESSED, NULL, 0, 0,
310 			    fqs->fqs_ifq->ifcq_ifp, 0, NULL, -1, NULL, 0, 0);
311 		}
312 
313 		struct kern_pbufpool *pp =
314 		    __DECONST(struct kern_pbufpool *, ((struct __kern_quantum *)kpkt)->qum_pp);
315 		pp_free_packet(pp, (uint64_t)kpkt);
316 	}
317 #endif /* SKYWALK */
318 
319 	fq->fq_bytes -= old_pktlen;
320 	fq_cl->fcl_stat.fcl_byte_cnt -= old_pktlen;
321 	fq_cl->fcl_stat.fcl_pkt_cnt--;
322 	IFCQ_DEC_LEN(ifq);
323 	IFCQ_DEC_BYTES(ifq, old_pktlen);
324 
325 	FQ_GRP_DEC_LEN(fq);
326 	FQ_GRP_DEC_BYTES(fq, old_pktlen);
327 
328 	*pkt_timestamp = old_timestamp;
329 
330 	return CLASSQEQ_COMPRESSED;
331 }
332 
333 int
fq_codel_enq_legacy(void * fqs_p,fq_if_group_t * fq_grp,pktsched_pkt_t * pkt,fq_if_classq_t * fq_cl)334 fq_codel_enq_legacy(void *fqs_p, fq_if_group_t *fq_grp, pktsched_pkt_t *pkt,
335     fq_if_classq_t *fq_cl)
336 {
337 	fq_if_t *fqs = (fq_if_t *)fqs_p;
338 	int droptype = DTYPE_NODROP, fc_adv = 0, ret = CLASSQEQ_SUCCESS;
339 	u_int64_t now;
340 	fq_t *fq = NULL;
341 	uint64_t *__single pkt_timestamp;
342 	volatile uint32_t *__single pkt_flags;
343 	uint32_t pkt_flowid, cnt;
344 	uint8_t pkt_proto, pkt_flowsrc;
345 	fq_tfc_type_t tfc_type = FQ_TFC_C;
346 
347 	cnt = pkt->pktsched_pcnt;
348 	pktsched_get_pkt_vars(pkt, &pkt_flags, &pkt_timestamp, &pkt_flowid,
349 	    &pkt_flowsrc, &pkt_proto, NULL, NULL);
350 
351 	/*
352 	 * XXX Not walking the chain to set this flag on every packet.
353 	 * This flag is only used for debugging. Nothing is affected if it's
354 	 * not set.
355 	 */
356 	switch (pkt->pktsched_ptype) {
357 	case QP_MBUF:
358 		/* See comments in <rdar://problem/14040693> */
359 		VERIFY(!(*pkt_flags & PKTF_PRIV_GUARDED));
360 		*pkt_flags |= PKTF_PRIV_GUARDED;
361 		break;
362 #if SKYWALK
363 	case QP_PACKET:
364 		/* sanity check */
365 		ASSERT((*pkt_flags & ~PKT_F_COMMON_MASK) == 0);
366 		break;
367 #endif /* SKYWALK */
368 	default:
369 		VERIFY(0);
370 		/* NOTREACHED */
371 		__builtin_unreachable();
372 	}
373 
374 	if (fq_codel_enable_l4s) {
375 		tfc_type = pktsched_is_pkt_l4s(pkt) ? FQ_TFC_L4S : FQ_TFC_C;
376 	}
377 
378 	/*
379 	 * Timestamps for every packet must be set prior to entering this path.
380 	 */
381 	now = *pkt_timestamp;
382 	ASSERT(now > 0);
383 
384 	/* find the flowq for this packet */
385 	fq = fq_if_hash_pkt(fqs, fq_grp, pkt_flowid, pktsched_get_pkt_svc(pkt),
386 	    now, pkt_proto, pkt_flowsrc, tfc_type, true);
387 	if (__improbable(fq == NULL)) {
388 		DTRACE_IP1(memfail__drop, fq_if_t *, fqs);
389 		/* drop the packet if we could not allocate a flow queue */
390 		fq_cl->fcl_stat.fcl_drop_memfailure += cnt;
391 		return CLASSQEQ_DROP;
392 	}
393 	VERIFY(fq->fq_group == fq_grp);
394 	VERIFY(fqs->fqs_ptype == pkt->pktsched_ptype);
395 
396 	KDBG(AQM_KTRACE_STATS_FLOW_ENQUEUE, fq->fq_flowhash,
397 	    AQM_KTRACE_FQ_GRP_SC_IDX(fq),
398 	    fq->fq_bytes, pktsched_get_pkt_len(pkt));
399 
400 	fq_detect_dequeue_stall(fqs, fq, fq_cl, &now);
401 
402 	/*
403 	 * Skip the dropping part if it's L4S. Flow control or ECN marking decision
404 	 * will be made at dequeue time.
405 	 */
406 	if (fq_codel_enable_l4s && tfc_type == FQ_TFC_L4S) {
407 		fq_cl->fcl_stat.fcl_l4s_pkts += cnt;
408 		droptype = DTYPE_NODROP;
409 	}
410 
411 	if (__improbable(FQ_IS_DELAY_HIGH(fq) || FQ_IS_OVERWHELMING(fq))) {
412 		if ((fq->fq_flags & FQF_FLOWCTL_CAPABLE) &&
413 		    (*pkt_flags & PKTF_FLOW_ADV)) {
414 			fc_adv = 1;
415 			/*
416 			 * If the flow is suspended or it is not
417 			 * TCP/QUIC, drop the chain.
418 			 */
419 			if ((pkt_proto != IPPROTO_TCP) &&
420 			    (pkt_proto != IPPROTO_QUIC)) {
421 				droptype = DTYPE_EARLY;
422 				fq_cl->fcl_stat.fcl_drop_early += cnt;
423 				IFCQ_DROP_ADD(fqs->fqs_ifq, cnt, pktsched_get_pkt_len(pkt));
424 			}
425 			DTRACE_IP6(flow__adv, fq_if_t *, fqs,
426 			    fq_if_classq_t *, fq_cl, fq_t *, fq,
427 			    int, droptype, pktsched_pkt_t *, pkt,
428 			    uint32_t, cnt);
429 		} else {
430 			/*
431 			 * Need to drop packets to make room for the new
432 			 * ones. Try to drop from the head of the queue
433 			 * instead of the latest packets.
434 			 */
435 			if (!fq_empty(fq, fqs->fqs_ptype)) {
436 				uint32_t i;
437 
438 				for (i = 0; i < cnt; i++) {
439 					fq_head_drop(fqs, fq);
440 				}
441 				droptype = DTYPE_NODROP;
442 			} else {
443 				droptype = DTYPE_EARLY;
444 			}
445 			fq_cl->fcl_stat.fcl_drop_early += cnt;
446 
447 			DTRACE_IP6(no__flow__adv, fq_if_t *, fqs,
448 			    fq_if_classq_t *, fq_cl, fq_t *, fq,
449 			    int, droptype, pktsched_pkt_t *, pkt,
450 			    uint32_t, cnt);
451 		}
452 	}
453 
454 	/* Set the return code correctly */
455 	if (__improbable(fc_adv == 1 && droptype != DTYPE_FORCED)) {
456 		if (fq_if_add_fcentry(fqs, pkt, pkt_flowsrc, fq, fq_cl)) {
457 			fq->fq_flags |= FQF_FLOWCTL_ON;
458 			/* deliver flow control advisory error */
459 			if (droptype == DTYPE_NODROP) {
460 				ret = CLASSQEQ_SUCCESS_FC;
461 			} else {
462 				/* dropped due to flow control */
463 				ret = CLASSQEQ_DROP_FC;
464 			}
465 		} else {
466 			/*
467 			 * if we could not flow control the flow, it is
468 			 * better to drop
469 			 */
470 			droptype = DTYPE_FORCED;
471 			ret = CLASSQEQ_DROP_FC;
472 			fq_cl->fcl_stat.fcl_flow_control_fail++;
473 		}
474 		DTRACE_IP3(fc__ret, fq_if_t *, fqs, int, droptype, int, ret);
475 	}
476 
477 	/*
478 	 * If the queue length hits the queue limit, drop a chain with the
479 	 * same number of packets from the front of the queue for a flow with
480 	 * maximum number of bytes. This will penalize heavy and unresponsive
481 	 * flows. It will also avoid a tail drop.
482 	 */
483 	if (__improbable(droptype == DTYPE_NODROP &&
484 	    fq_if_at_drop_limit(fqs))) {
485 		uint32_t i;
486 
487 		if (fqs->fqs_large_flow == fq) {
488 			/*
489 			 * Drop from the head of the current fq. Since a
490 			 * new packet will be added to the tail, it is ok
491 			 * to leave fq in place.
492 			 */
493 			DTRACE_IP5(large__flow, fq_if_t *, fqs,
494 			    fq_if_classq_t *, fq_cl, fq_t *, fq,
495 			    pktsched_pkt_t *, pkt, uint32_t, cnt);
496 
497 			for (i = 0; i < cnt; i++) {
498 				fq_head_drop(fqs, fq);
499 			}
500 			fq_cl->fcl_stat.fcl_drop_overflow += cnt;
501 
502 			/*
503 			 * TCP and QUIC will react to the loss of those head dropped pkts
504 			 * and adjust send rate.
505 			 */
506 			if ((fq->fq_flags & FQF_FLOWCTL_CAPABLE) &&
507 			    (*pkt_flags & PKTF_FLOW_ADV) &&
508 			    (pkt_proto != IPPROTO_TCP) &&
509 			    (pkt_proto != IPPROTO_QUIC)) {
510 				if (fq_if_add_fcentry(fqs, pkt, pkt_flowsrc, fq, fq_cl)) {
511 					fq->fq_flags |= FQF_FLOWCTL_ON;
512 					FQ_SET_OVERWHELMING(fq);
513 					fq_cl->fcl_stat.fcl_overwhelming++;
514 					/* deliver flow control advisory error */
515 					ret = CLASSQEQ_SUCCESS_FC;
516 				}
517 			}
518 		} else {
519 			if (fqs->fqs_large_flow == NULL) {
520 				droptype = DTYPE_FORCED;
521 				fq_cl->fcl_stat.fcl_drop_overflow += cnt;
522 				ret = CLASSQEQ_DROP;
523 
524 				DTRACE_IP5(no__large__flow, fq_if_t *, fqs,
525 				    fq_if_classq_t *, fq_cl, fq_t *, fq,
526 				    pktsched_pkt_t *, pkt, uint32_t, cnt);
527 
528 				/*
529 				 * if this fq was freshly created and there
530 				 * is nothing to enqueue, move it to empty list
531 				 */
532 				if (fq_empty(fq, fqs->fqs_ptype) &&
533 				    !(fq->fq_flags & (FQF_NEW_FLOW |
534 				    FQF_OLD_FLOW))) {
535 					fq_if_move_to_empty_flow(fqs, fq_cl,
536 					    fq, now);
537 					fq = NULL;
538 				}
539 			} else {
540 				DTRACE_IP5(different__large__flow,
541 				    fq_if_t *, fqs, fq_if_classq_t *, fq_cl,
542 				    fq_t *, fq, pktsched_pkt_t *, pkt,
543 				    uint32_t, cnt);
544 
545 				for (i = 0; i < cnt; i++) {
546 					fq_if_drop_packet(fqs, now);
547 				}
548 			}
549 		}
550 	}
551 
552 	fq_cl->fcl_flags &= ~FCL_PACED;
553 
554 	if (__probable(droptype == DTYPE_NODROP)) {
555 		uint32_t chain_len = pktsched_get_pkt_len(pkt);
556 		int ret_compress = 0;
557 
558 		/*
559 		 * We do not compress if we are enqueuing a chain.
560 		 * Traversing the chain to look for acks would defeat the
561 		 * purpose of batch enqueueing.
562 		 */
563 		if (cnt == 1) {
564 			ret_compress = fq_compressor(fqs, fq, fq_cl, pkt);
565 			if (ret_compress == CLASSQEQ_COMPRESSED) {
566 				fq_cl->fcl_stat.fcl_pkts_compressed++;
567 			}
568 		}
569 		DTRACE_IP5(fq_enqueue, fq_if_t *, fqs, fq_if_classq_t *, fq_cl,
570 		    fq_t *, fq, pktsched_pkt_t *, pkt, uint32_t, cnt);
571 		fq_enqueue(fq, pkt->pktsched_pkt, pkt->pktsched_tail, cnt,
572 		    pkt->pktsched_ptype);
573 
574 		fq->fq_bytes += chain_len;
575 		fq_cl->fcl_stat.fcl_byte_cnt += chain_len;
576 		fq_cl->fcl_stat.fcl_pkt_cnt += cnt;
577 
578 		/*
579 		 * check if this queue will qualify to be the next
580 		 * victim queue
581 		 */
582 		fq_if_is_flow_heavy(fqs, fq);
583 	} else {
584 		DTRACE_IP3(fq_drop, fq_if_t *, fqs, int, droptype, int, ret);
585 		return (ret != CLASSQEQ_SUCCESS) ? ret : CLASSQEQ_DROP;
586 	}
587 
588 	/*
589 	 * If the queue is not currently active, add it to the end of new
590 	 * flows list for that service class.
591 	 */
592 	if ((fq->fq_flags & (FQF_NEW_FLOW | FQF_OLD_FLOW)) == 0) {
593 		VERIFY(STAILQ_NEXT(fq, fq_actlink) == NULL);
594 		STAILQ_INSERT_TAIL(&fq_cl->fcl_new_flows, fq, fq_actlink);
595 		fq->fq_flags |= FQF_NEW_FLOW;
596 
597 		fq_cl->fcl_stat.fcl_newflows_cnt++;
598 
599 		fq->fq_deficit = fq_cl->fcl_quantum;
600 	}
601 	return ret;
602 }
603 
604 void
fq_getq_flow_internal(fq_if_t * fqs,fq_t * fq,pktsched_pkt_t * pkt)605 fq_getq_flow_internal(fq_if_t *fqs, fq_t *fq, pktsched_pkt_t *pkt)
606 {
607 	classq_pkt_t p = CLASSQ_PKT_INITIALIZER(p);
608 	uint32_t plen;
609 	fq_if_classq_t *fq_cl;
610 	struct ifclassq *ifq = fqs->fqs_ifq;
611 
612 	fq_dequeue(fq, &p, fqs->fqs_ptype);
613 	if (p.cp_ptype == QP_INVALID) {
614 		VERIFY(p.cp_mbuf == NULL);
615 		return;
616 	}
617 
618 	fq->fq_next_tx_time = FQ_INVALID_TX_TS;
619 
620 	pktsched_pkt_encap(pkt, &p);
621 	plen = pktsched_get_pkt_len(pkt);
622 
623 	VERIFY(fq->fq_bytes >= plen);
624 	fq->fq_bytes -= plen;
625 
626 	fq_cl = &FQ_CLASSQ(fq);
627 	fq_cl->fcl_stat.fcl_byte_cnt -= plen;
628 	fq_cl->fcl_stat.fcl_pkt_cnt--;
629 	fq_cl->fcl_flags &= ~FCL_PACED;
630 
631 	IFCQ_DEC_LEN(ifq);
632 	IFCQ_DEC_BYTES(ifq, plen);
633 
634 	FQ_GRP_DEC_LEN(fq);
635 	FQ_GRP_DEC_BYTES(fq, plen);
636 
637 	/* Reset getqtime so that we don't count idle times */
638 	if (fq_empty(fq, fqs->fqs_ptype)) {
639 		fq->fq_getqtime = 0;
640 	}
641 }
642 
643 /*
644  * fq_get_next_tx_time returns FQ_INVALID_TX_TS when there is no tx time in fq
645  */
646 static uint64_t
fq_get_next_tx_time(fq_if_t * fqs,fq_t * fq)647 fq_get_next_tx_time(fq_if_t *fqs, fq_t *fq)
648 {
649 	uint64_t tx_time = FQ_INVALID_TX_TS;
650 
651 	/*
652 	 * Check the cached value in fq
653 	 */
654 	if (fq->fq_next_tx_time != FQ_INVALID_TX_TS) {
655 		return fq->fq_next_tx_time;
656 	}
657 
658 	switch (fqs->fqs_ptype) {
659 	case QP_MBUF: {
660 		struct mbuf *m;
661 		if ((m = MBUFQ_FIRST(&fq->fq_mbufq)) != NULL) {
662 			struct m_tag *tag;
663 			tag = m_tag_locate(m, KERNEL_MODULE_TAG_ID,
664 			    KERNEL_TAG_TYPE_AQM);
665 			if (tag != NULL) {
666 				tx_time = *(uint64_t *)tag->m_tag_data;
667 			}
668 		}
669 		break;
670 	}
671 	case QP_PACKET: {
672 		struct __kern_packet *p = KPKTQ_FIRST(&fq->fq_kpktq);
673 		if (__probable(p != NULL)) {
674 			tx_time = __packet_get_tx_timestamp(SK_PKT2PH(p));
675 		}
676 		break;
677 	}
678 	default:
679 		VERIFY(0);
680 		/* NOTREACHED */
681 		__builtin_unreachable();
682 	}
683 
684 	/*
685 	 * Cache the tx time in fq. The cache will be clear after dequeue or drop
686 	 * from the fq.
687 	 */
688 	fq->fq_next_tx_time = tx_time;
689 
690 	return tx_time;
691 }
692 
693 /*
694  * fq_tx_time_ready returns true if the fq is empty so that it doesn't
695  * affect caller logics that handles empty flow.
696  */
697 boolean_t
fq_tx_time_ready(fq_if_t * fqs,fq_t * fq,uint64_t now,uint64_t * ready_time)698 fq_tx_time_ready(fq_if_t *fqs, fq_t *fq, uint64_t now, uint64_t *ready_time)
699 {
700 	uint64_t pkt_tx_time;
701 	fq_if_classq_t *fq_cl = &FQ_CLASSQ(fq);
702 
703 	if (!fq_codel_enable_pacing) {
704 		return TRUE;
705 	}
706 
707 	pkt_tx_time = fq_get_next_tx_time(fqs, fq);
708 	if (ready_time != NULL) {
709 		*ready_time = pkt_tx_time;
710 	}
711 
712 	if (pkt_tx_time <= now + pkt_pacing_leeway ||
713 	    pkt_tx_time == FQ_INVALID_TX_TS) {
714 		return TRUE;
715 	}
716 
717 	/*
718 	 * Ignore the tx time if it's scheduled too far in the future
719 	 */
720 	if (pkt_tx_time > max_pkt_pacing_interval + now) {
721 		fq_cl->fcl_stat.fcl_ignore_tx_time++;
722 		return TRUE;
723 	}
724 
725 	ASSERT(pkt_tx_time != FQ_INVALID_TX_TS);
726 	KDBG(AQM_KTRACE_TX_NOT_READY, fq->fq_flowhash,
727 	    AQM_KTRACE_FQ_GRP_SC_IDX(fq), now, pkt_tx_time);
728 	return FALSE;
729 }
730 
731 void
fq_codel_dq_legacy(void * fqs_p,void * fq_p,pktsched_pkt_t * pkt,uint64_t now)732 fq_codel_dq_legacy(void *fqs_p, void *fq_p, pktsched_pkt_t *pkt, uint64_t now)
733 {
734 	fq_if_t *fqs = (fq_if_t *)fqs_p;
735 	fq_t *fq = (fq_t *)fq_p;
736 	fq_if_classq_t *fq_cl = &FQ_CLASSQ(fq);
737 	int64_t qdelay = 0;
738 	volatile uint32_t *__single pkt_flags;
739 	uint64_t *__single pkt_timestamp, pkt_tx_time = 0, pacing_delay = 0;
740 	uint64_t fq_min_delay_threshold = FQ_TARGET_DELAY(fq);
741 	uint8_t pkt_flowsrc;
742 	boolean_t l4s_pkt;
743 
744 	fq_getq_flow_internal(fqs, fq, pkt);
745 	if (pkt->pktsched_ptype == QP_INVALID) {
746 		VERIFY(pkt->pktsched_pkt_mbuf == NULL);
747 		return;
748 	}
749 
750 	pktsched_get_pkt_vars(pkt, &pkt_flags, &pkt_timestamp, NULL, &pkt_flowsrc,
751 	    NULL, NULL, &pkt_tx_time);
752 	l4s_pkt = pktsched_is_pkt_l4s(pkt);
753 	if (fq_codel_enable_pacing && fq_codel_enable_l4s) {
754 		if (pkt_tx_time > *pkt_timestamp) {
755 			pacing_delay = pkt_tx_time - *pkt_timestamp;
756 			fq_cl->fcl_stat.fcl_paced_pkts++;
757 			DTRACE_SKYWALK3(aqm__pacing__delta, uint64_t, now - pkt_tx_time,
758 			    fq_if_t *, fqs, fq_t *, fq);
759 		}
760 #if (DEVELOPMENT || DEBUG)
761 		else if (pkt_tx_time != 0) {
762 			DTRACE_SKYWALK5(aqm__miss__pacing__delay, uint64_t, *pkt_timestamp,
763 			    uint64_t, pkt_tx_time, uint64_t, now, fq_if_t *,
764 			    fqs, fq_t *, fq);
765 		}
766 #endif // (DEVELOPMENT || DEBUG)
767 	}
768 
769 	/* this will compute qdelay in nanoseconds */
770 	if (now > *pkt_timestamp) {
771 		qdelay = now - *pkt_timestamp;
772 	}
773 
774 	/* Update min/max/avg qdelay for the respective class */
775 	if (fq_cl->fcl_stat.fcl_min_qdelay == 0 ||
776 	    (qdelay > 0 && (u_int64_t)qdelay < fq_cl->fcl_stat.fcl_min_qdelay)) {
777 		fq_cl->fcl_stat.fcl_min_qdelay = qdelay;
778 	}
779 
780 	if (fq_cl->fcl_stat.fcl_max_qdelay == 0 ||
781 	    (qdelay > 0 && (u_int64_t)qdelay > fq_cl->fcl_stat.fcl_max_qdelay)) {
782 		fq_cl->fcl_stat.fcl_max_qdelay = qdelay;
783 	}
784 
785 	uint64_t num_dequeues = fq_cl->fcl_stat.fcl_dequeue;
786 
787 	if (num_dequeues == 0) {
788 		fq_cl->fcl_stat.fcl_avg_qdelay = qdelay;
789 	} else if (qdelay > 0) {
790 		uint64_t res = 0;
791 		if (os_add_overflow(num_dequeues, 1, &res)) {
792 			/* Reset the dequeue num and dequeue bytes */
793 			fq_cl->fcl_stat.fcl_dequeue = num_dequeues = 0;
794 			fq_cl->fcl_stat.fcl_dequeue_bytes = 0;
795 			fq_cl->fcl_stat.fcl_avg_qdelay = qdelay;
796 			os_log_info(OS_LOG_DEFAULT, "%s: dequeue num overflow, "
797 			    "flow: 0x%x, iface: %s", __func__, fq->fq_flowhash,
798 			    if_name(fqs->fqs_ifq->ifcq_ifp));
799 		} else {
800 			uint64_t product = 0;
801 			if (os_mul_overflow(fq_cl->fcl_stat.fcl_avg_qdelay,
802 			    num_dequeues, &product) || os_add_overflow(product, qdelay, &res)) {
803 				fq_cl->fcl_stat.fcl_avg_qdelay = qdelay;
804 			} else {
805 				fq_cl->fcl_stat.fcl_avg_qdelay = res /
806 				    (num_dequeues + 1);
807 			}
808 		}
809 	}
810 
811 	fq->fq_pkts_since_last_report++;
812 	if (fq_codel_enable_l4s && l4s_pkt) {
813 		/*
814 		 * A safe guard to make sure that L4S is not going to build a huge
815 		 * queue if we encounter unexpected problems (for eg., if ACKs don't
816 		 * arrive in timely manner due to congestion in reverse path).
817 		 */
818 		fq_min_delay_threshold = l4s_min_delay_threshold;
819 
820 		if ((l4s_ce_threshold != 0 && qdelay > l4s_ce_threshold + pacing_delay) ||
821 		    (l4s_ce_threshold == 0 && qdelay > FQ_TARGET_DELAY(fq) + pacing_delay)) {
822 			DTRACE_SKYWALK4(aqm__mark__ce, uint64_t, qdelay, uint64_t, pacing_delay,
823 			    fq_if_t *, fqs, fq_t *, fq);
824 			KDBG(AQM_KTRACE_STATS_FLOW_REPORT_CE, fq->fq_flowhash,
825 			    AQM_KTRACE_FQ_GRP_SC_IDX(fq), qdelay, pacing_delay);
826 			/*
827 			 * The packet buffer that pktsched_mark_ecn writes to can be pageable.
828 			 * Since it is not safe to write to pageable memory while preemption
829 			 * is disabled, convert the spin lock into mutex.
830 			 */
831 			IFCQ_CONVERT_LOCK(fqs->fqs_ifq);
832 			if (__improbable(l4s_local_ce_report != 0) &&
833 			    (*pkt_flags & PKTF_FLOW_ADV) != 0 &&
834 			    fq_if_report_congestion(fqs, pkt, 0, 1, fq->fq_pkts_since_last_report)) {
835 				fq->fq_pkts_since_last_report = 0;
836 				fq_cl->fcl_stat.fcl_ce_reported++;
837 			} else if (pktsched_mark_ecn(pkt) == 0) {
838 				fq_cl->fcl_stat.fcl_ce_marked++;
839 			} else {
840 				fq_cl->fcl_stat.fcl_ce_mark_failures++;
841 			}
842 		}
843 	}
844 
845 	ASSERT(pacing_delay <= INT64_MAX);
846 	qdelay = MAX(0, qdelay - (int64_t)pacing_delay);
847 	if (fq->fq_min_qdelay == 0 ||
848 	    (u_int64_t)qdelay < fq->fq_min_qdelay) {
849 		fq->fq_min_qdelay = qdelay;
850 	}
851 
852 	if (now >= fq->fq_updatetime) {
853 		if (fq->fq_min_qdelay > fq_min_delay_threshold) {
854 			if (!FQ_IS_DELAY_HIGH(fq)) {
855 				FQ_SET_DELAY_HIGH(fq);
856 			}
857 		} else {
858 			FQ_CLEAR_DELAY_HIGH(fq);
859 		}
860 		/* Reset measured queue delay and update time */
861 		fq->fq_updatetime = now + FQ_UPDATE_INTERVAL(fq);
862 		fq->fq_min_qdelay = 0;
863 	}
864 
865 	if (fqs->fqs_large_flow != fq || !fq_if_almost_at_drop_limit(fqs)) {
866 		FQ_CLEAR_OVERWHELMING(fq);
867 	}
868 	if (!FQ_IS_DELAY_HIGH(fq) || fq_empty(fq, fqs->fqs_ptype)) {
869 		FQ_CLEAR_DELAY_HIGH(fq);
870 	}
871 
872 	if ((fq->fq_flags & FQF_FLOWCTL_ON) &&
873 	    !FQ_IS_DELAY_HIGH(fq) && !FQ_IS_OVERWHELMING(fq)) {
874 		fq_if_flow_feedback(fqs, fq, fq_cl);
875 	}
876 
877 	if (fq_empty(fq, fqs->fqs_ptype)) {
878 		/* Reset getqtime so that we don't count idle times */
879 		fq->fq_getqtime = 0;
880 	} else {
881 		fq->fq_getqtime = now;
882 	}
883 	fq_if_is_flow_heavy(fqs, fq);
884 
885 	*pkt_timestamp = 0;
886 	switch (pkt->pktsched_ptype) {
887 	case QP_MBUF:
888 		*pkt_flags &= ~PKTF_PRIV_GUARDED;
889 		break;
890 #if SKYWALK
891 	case QP_PACKET:
892 		/* sanity check */
893 		ASSERT((*pkt_flags & ~PKT_F_COMMON_MASK) == 0);
894 		break;
895 #endif /* SKYWALK */
896 	default:
897 		VERIFY(0);
898 		/* NOTREACHED */
899 		__builtin_unreachable();
900 	}
901 }
902 
903 int
fq_codel_enq(void * fqs_p,fq_if_group_t * fq_grp,pktsched_pkt_t * pkt,fq_if_classq_t * fq_cl)904 fq_codel_enq(void *fqs_p, fq_if_group_t *fq_grp, pktsched_pkt_t *pkt,
905     fq_if_classq_t *fq_cl)
906 {
907 	fq_if_t *fqs = (fq_if_t *)fqs_p;
908 	int droptype = DTYPE_NODROP, ret = CLASSQEQ_SUCCESS;
909 	u_int64_t now;
910 	fq_t *fq = NULL;
911 	uint64_t *__single pkt_timestamp;
912 	volatile uint32_t *__single pkt_flags;
913 	uint32_t pkt_flowid, cnt;
914 	uint8_t pkt_proto, pkt_flowsrc;
915 	fq_tfc_type_t tfc_type = FQ_TFC_C;
916 
917 	cnt = pkt->pktsched_pcnt;
918 	pktsched_get_pkt_vars(pkt, &pkt_flags, &pkt_timestamp, &pkt_flowid,
919 	    &pkt_flowsrc, &pkt_proto, NULL, NULL);
920 
921 	/*
922 	 * XXX Not walking the chain to set this flag on every packet.
923 	 * This flag is only used for debugging. Nothing is affected if it's
924 	 * not set.
925 	 */
926 	switch (pkt->pktsched_ptype) {
927 	case QP_MBUF:
928 		/* See comments in <rdar://problem/14040693> */
929 		VERIFY(!(*pkt_flags & PKTF_PRIV_GUARDED));
930 		break;
931 #if SKYWALK
932 	case QP_PACKET:
933 		/* sanity check */
934 		ASSERT((*pkt_flags & ~PKT_F_COMMON_MASK) == 0);
935 		break;
936 #endif /* SKYWALK */
937 	default:
938 		VERIFY(0);
939 		/* NOTREACHED */
940 		__builtin_unreachable();
941 	}
942 
943 	if (fq_codel_enable_l4s) {
944 		tfc_type = pktsched_is_pkt_l4s(pkt) ? FQ_TFC_L4S : FQ_TFC_C;
945 	}
946 
947 	/*
948 	 * Timestamps for every packet must be set prior to entering this path.
949 	 */
950 	now = *pkt_timestamp;
951 	ASSERT(now > 0);
952 
953 	/* find the flowq for this packet */
954 	fq = fq_if_hash_pkt(fqs, fq_grp, pkt_flowid, pktsched_get_pkt_svc(pkt),
955 	    now, pkt_proto, pkt_flowsrc, tfc_type, true);
956 	if (__improbable(fq == NULL)) {
957 		DTRACE_IP1(memfail__drop, fq_if_t *, fqs);
958 		/* drop the packet if we could not allocate a flow queue */
959 		fq_cl->fcl_stat.fcl_drop_memfailure += cnt;
960 		return CLASSQEQ_DROP;
961 	}
962 	VERIFY(fq->fq_group == fq_grp);
963 	VERIFY(fqs->fqs_ptype == pkt->pktsched_ptype);
964 
965 	KDBG(AQM_KTRACE_STATS_FLOW_ENQUEUE, fq->fq_flowhash,
966 	    AQM_KTRACE_FQ_GRP_SC_IDX(fq),
967 	    fq->fq_bytes, pktsched_get_pkt_len(pkt));
968 
969 	/*
970 	 * Skip the dropping part if it's L4S. Flow control or ECN marking decision
971 	 * will be made at dequeue time.
972 	 */
973 	if (fq_codel_enable_l4s && tfc_type == FQ_TFC_L4S) {
974 		fq_cl->fcl_stat.fcl_l4s_pkts += cnt;
975 		droptype = DTYPE_NODROP;
976 	}
977 
978 	/*
979 	 * If the queue length hits the queue limit, drop a chain with the
980 	 * same number of packets from the front of the queue for a flow with
981 	 * maximum number of bytes. This will penalize heavy and unresponsive
982 	 * flows. It will also avoid a tail drop.
983 	 */
984 	if (__improbable(droptype == DTYPE_NODROP &&
985 	    fq_if_at_drop_limit(fqs))) {
986 		uint32_t i;
987 
988 		if (fqs->fqs_large_flow == fq) {
989 			/*
990 			 * Drop from the head of the current fq. Since a
991 			 * new packet will be added to the tail, it is ok
992 			 * to leave fq in place.
993 			 */
994 			DTRACE_IP5(large__flow, fq_if_t *, fqs,
995 			    fq_if_classq_t *, fq_cl, fq_t *, fq,
996 			    pktsched_pkt_t *, pkt, uint32_t, cnt);
997 
998 			for (i = 0; i < cnt; i++) {
999 				fq_head_drop(fqs, fq);
1000 			}
1001 			fq_cl->fcl_stat.fcl_drop_overflow += cnt;
1002 			/*
1003 			 * For UDP, flow control it here so that we won't waste too much
1004 			 * CPU dropping packets.
1005 			 */
1006 			if ((fq->fq_flags & FQF_FLOWCTL_CAPABLE) &&
1007 			    (*pkt_flags & PKTF_FLOW_ADV) &&
1008 			    (pkt_proto != IPPROTO_TCP) &&
1009 			    (pkt_proto != IPPROTO_QUIC)) {
1010 				if (fq_if_add_fcentry(fqs, pkt, pkt_flowsrc, fq, fq_cl)) {
1011 					fq->fq_flags |= FQF_FLOWCTL_ON;
1012 					FQ_SET_OVERWHELMING(fq);
1013 					fq_cl->fcl_stat.fcl_overwhelming++;
1014 					/* deliver flow control advisory error */
1015 					ret = CLASSQEQ_SUCCESS_FC;
1016 				}
1017 			}
1018 		} else {
1019 			if (fqs->fqs_large_flow == NULL) {
1020 				droptype = DTYPE_FORCED;
1021 				fq_cl->fcl_stat.fcl_drop_overflow += cnt;
1022 				ret = CLASSQEQ_DROP;
1023 
1024 				DTRACE_IP5(no__large__flow, fq_if_t *, fqs,
1025 				    fq_if_classq_t *, fq_cl, fq_t *, fq,
1026 				    pktsched_pkt_t *, pkt, uint32_t, cnt);
1027 
1028 				/*
1029 				 * if this fq was freshly created and there
1030 				 * is nothing to enqueue, move it to empty list
1031 				 */
1032 				if (fq_empty(fq, fqs->fqs_ptype) &&
1033 				    !(fq->fq_flags & (FQF_NEW_FLOW |
1034 				    FQF_OLD_FLOW))) {
1035 					fq_if_move_to_empty_flow(fqs, fq_cl,
1036 					    fq, now);
1037 					fq = NULL;
1038 				}
1039 			} else {
1040 				DTRACE_IP5(different__large__flow,
1041 				    fq_if_t *, fqs, fq_if_classq_t *, fq_cl,
1042 				    fq_t *, fq, pktsched_pkt_t *, pkt,
1043 				    uint32_t, cnt);
1044 
1045 				for (i = 0; i < cnt; i++) {
1046 					fq_if_drop_packet(fqs, now);
1047 				}
1048 			}
1049 		}
1050 	}
1051 
1052 	fq_cl->fcl_flags &= ~FCL_PACED;
1053 
1054 	if (__probable(droptype == DTYPE_NODROP)) {
1055 		uint32_t chain_len = pktsched_get_pkt_len(pkt);
1056 		int ret_compress = 0;
1057 
1058 		/*
1059 		 * We do not compress if we are enqueuing a chain.
1060 		 * Traversing the chain to look for acks would defeat the
1061 		 * purpose of batch enqueueing.
1062 		 */
1063 		if (cnt == 1) {
1064 			ret_compress = fq_compressor(fqs, fq, fq_cl, pkt);
1065 			if (ret_compress == CLASSQEQ_COMPRESSED) {
1066 				fq_cl->fcl_stat.fcl_pkts_compressed++;
1067 			}
1068 		}
1069 		DTRACE_IP5(fq_enqueue, fq_if_t *, fqs, fq_if_classq_t *, fq_cl,
1070 		    fq_t *, fq, pktsched_pkt_t *, pkt, uint32_t, cnt);
1071 		fq_enqueue(fq, pkt->pktsched_pkt, pkt->pktsched_tail, cnt,
1072 		    pkt->pktsched_ptype);
1073 
1074 		fq->fq_bytes += chain_len;
1075 		fq_cl->fcl_stat.fcl_byte_cnt += chain_len;
1076 		fq_cl->fcl_stat.fcl_pkt_cnt += cnt;
1077 
1078 		/*
1079 		 * check if this queue will qualify to be the next
1080 		 * victim queue
1081 		 */
1082 		fq_if_is_flow_heavy(fqs, fq);
1083 
1084 		if (FQ_CONGESTION_FEEDBACK_CAPABLE(fq) && fq->fq_flowsrc == FLOWSRC_INPCB &&
1085 		    fq->fq_congestion_cnt > fq->fq_last_congestion_cnt) {
1086 			KDBG(AQM_KTRACE_CONGESTION_NOTIFIED,
1087 			    fq->fq_flowhash, AQM_KTRACE_FQ_GRP_SC_IDX(fq),
1088 			    fq->fq_bytes, fq->fq_congestion_cnt);
1089 			fq_cl->fcl_stat.fcl_congestion_feedback++;
1090 			ret = CLASSQEQ_CONGESTED;
1091 		}
1092 	} else {
1093 		DTRACE_IP3(fq_drop, fq_if_t *, fqs, int, droptype, int, ret);
1094 		return (ret != CLASSQEQ_SUCCESS) ? ret : CLASSQEQ_DROP;
1095 	}
1096 
1097 	/*
1098 	 * If the queue is not currently active, add it to the end of new
1099 	 * flows list for that service class.
1100 	 */
1101 	if ((fq->fq_flags & (FQF_NEW_FLOW | FQF_OLD_FLOW)) == 0) {
1102 		VERIFY(STAILQ_NEXT(fq, fq_actlink) == NULL);
1103 		STAILQ_INSERT_TAIL(&fq_cl->fcl_new_flows, fq, fq_actlink);
1104 		fq->fq_flags |= FQF_NEW_FLOW;
1105 
1106 		fq_cl->fcl_stat.fcl_newflows_cnt++;
1107 
1108 		fq->fq_deficit = fq_cl->fcl_quantum;
1109 	}
1110 	fq->fq_last_congestion_cnt = fq->fq_congestion_cnt;
1111 
1112 	return ret;
1113 }
1114 
1115 static boolean_t
codel_ok_to_drop(pktsched_pkt_t * pkt,fq_t * fq,struct codel_status * codel_status,uint64_t now)1116 codel_ok_to_drop(pktsched_pkt_t *pkt, fq_t *fq,
1117     struct codel_status *codel_status, uint64_t now)
1118 {
1119 	uint64_t *__single pkt_enq_time, sojourn_time;
1120 	boolean_t ok_to_drop = false;
1121 
1122 	if (pkt->pktsched_pcnt == 0) {
1123 		codel_status->first_above_time = 0;
1124 		return false;
1125 	}
1126 
1127 	pktsched_get_pkt_vars(pkt, NULL, &pkt_enq_time, NULL, NULL, NULL, NULL, NULL);
1128 	sojourn_time = 0;
1129 	// TODO: handle pacing
1130 	VERIFY(now >= *pkt_enq_time);
1131 	if (__probable(now > *pkt_enq_time)) {
1132 		sojourn_time = now - *pkt_enq_time;
1133 	}
1134 
1135 	if (sojourn_time <= FQ_TARGET_DELAY(fq)) {
1136 		codel_status->first_above_time = 0;
1137 		goto end;
1138 	}
1139 	if (codel_status->first_above_time == 0) {
1140 		codel_status->first_above_time = now += FQ_UPDATE_INTERVAL(fq);
1141 	} else if (now >= codel_status->first_above_time) {
1142 		ok_to_drop = true;
1143 		KDBG(AQM_KTRACE_OK_TO_DROP, fq->fq_flowhash, sojourn_time,
1144 		    *pkt_enq_time, now);
1145 	}
1146 
1147 end:
1148 	/* we shouldn't need to access pkt_enq_time again, clear it now */
1149 	*pkt_enq_time = 0;
1150 	return ok_to_drop;
1151 }
1152 
1153 static float
fast_inv_sqrt(float number)1154 fast_inv_sqrt(float number)
1155 {
1156 	union {
1157 		float    f;
1158 		uint32_t i;
1159 	} conv = { .f = number };
1160 	conv.i  = 0x5f3759df - (conv.i >> 1);
1161 	conv.f *= 1.5F - (number * 0.5F * conv.f * conv.f);
1162 	return conv.f;
1163 }
1164 
1165 static uint64_t
codel_control_law(float inv_sqrt,uint64_t t,uint64_t interval)1166 codel_control_law(float inv_sqrt, uint64_t t, uint64_t interval)
1167 {
1168 	/* Drop becomes more frequeut as drop count increases */
1169 	uint64_t val = (uint64_t)(inv_sqrt * interval);
1170 	uint64_t result = t + val;
1171 	VERIFY(val > 0 && val <= interval);
1172 	return result;
1173 }
1174 
1175 static void
fq_drop_pkt(struct ifclassq * ifcq,struct ifnet * ifp,pktsched_pkt_t * pkt)1176 fq_drop_pkt(struct ifclassq *ifcq, struct ifnet *ifp, pktsched_pkt_t *pkt)
1177 {
1178 	IFCQ_DROP_ADD(ifcq, 1, pktsched_get_pkt_len(pkt));
1179 	if (__improbable(droptap_verbose > 0)) {
1180 		pktsched_drop_pkt(pkt, ifp, DROP_REASON_AQM_HIGH_DELAY,
1181 		    __func__, __LINE__, 0);
1182 	} else {
1183 		pktsched_free_pkt(pkt);
1184 	}
1185 }
1186 
1187 void
fq_codel_dq(void * fqs_p,void * fq_p,pktsched_pkt_t * pkt,uint64_t now)1188 fq_codel_dq(void *fqs_p, void *fq_p, pktsched_pkt_t *pkt, uint64_t now)
1189 {
1190 	fq_if_t *fqs = (fq_if_t *)fqs_p;
1191 	fq_t *fq = (fq_t *)fq_p;
1192 	fq_if_classq_t *fq_cl = &FQ_CLASSQ(fq);
1193 	struct ifnet *ifp = fqs->fqs_ifq->ifcq_ifp;
1194 	struct codel_status *status = &fq->codel_status;
1195 	struct ifclassq *ifcq = fqs->fqs_ifq;
1196 	volatile uint32_t *__single pkt_flags;
1197 	uint64_t *__single pkt_timestamp, pkt_tx_time = 0, pacing_delay = 0, pkt_enq_time = 0;
1198 	int64_t qdelay = 0;
1199 	boolean_t l4s_pkt;
1200 	boolean_t ok_to_drop = false;
1201 	uint32_t delta = 0;
1202 
1203 	fq_getq_flow_internal(fqs, fq, pkt);
1204 	if (pkt->pktsched_ptype == QP_INVALID) {
1205 		VERIFY(pkt->pktsched_pkt_mbuf == NULL);
1206 		return;
1207 	}
1208 
1209 	pktsched_get_pkt_vars(pkt, &pkt_flags, &pkt_timestamp, NULL, NULL,
1210 	    NULL, NULL, &pkt_tx_time);
1211 	l4s_pkt = pktsched_is_pkt_l4s(pkt);
1212 	pkt_enq_time = *pkt_timestamp;
1213 
1214 	if (fq_codel_enable_pacing && fq_codel_enable_l4s) {
1215 		if (pkt_tx_time > pkt_enq_time) {
1216 			pacing_delay = pkt_tx_time - pkt_enq_time;
1217 			fq_cl->fcl_stat.fcl_paced_pkts++;
1218 			DTRACE_SKYWALK3(aqm__pacing__delta, uint64_t, now - pkt_tx_time,
1219 			    fq_if_t *, fqs, fq_t *, fq);
1220 		}
1221 #if (DEVELOPMENT || DEBUG)
1222 		else if (pkt_tx_time != 0) {
1223 			DTRACE_SKYWALK5(aqm__miss__pacing__delay, uint64_t, pkt_enq_time,
1224 			    uint64_t, pkt_tx_time, uint64_t, now, fq_if_t *,
1225 			    fqs, fq_t *, fq);
1226 		}
1227 #endif // (DEVELOPMENT || DEBUG)
1228 	}
1229 
1230 	if (fq_codel_enable_l4s && l4s_pkt) {
1231 		/* this will compute qdelay in nanoseconds */
1232 		if (now > pkt_enq_time) {
1233 			qdelay = now - pkt_enq_time;
1234 		}
1235 
1236 		fq->fq_pkts_since_last_report++;
1237 
1238 		if ((l4s_ce_threshold != 0 && qdelay > l4s_ce_threshold + pacing_delay) ||
1239 		    (l4s_ce_threshold == 0 && qdelay > FQ_TARGET_DELAY(fq) + pacing_delay)) {
1240 			DTRACE_SKYWALK4(aqm__mark__ce, uint64_t, qdelay, uint64_t, pacing_delay,
1241 			    fq_if_t *, fqs, fq_t *, fq);
1242 			KDBG(AQM_KTRACE_STATS_FLOW_REPORT_CE, fq->fq_flowhash,
1243 			    AQM_KTRACE_FQ_GRP_SC_IDX(fq), qdelay, pacing_delay);
1244 			/*
1245 			 * The packet buffer that pktsched_mark_ecn writes to can be pageable.
1246 			 * Since it is not safe to write to pageable memory while preemption
1247 			 * is disabled, convert the spin lock into mutex.
1248 			 */
1249 			IFCQ_CONVERT_LOCK(fqs->fqs_ifq);
1250 			if (__improbable(l4s_local_ce_report != 0) &&
1251 			    (*pkt_flags & PKTF_FLOW_ADV) != 0 &&
1252 			    fq_if_report_congestion(fqs, pkt, 0, 1, fq->fq_pkts_since_last_report)) {
1253 				fq->fq_pkts_since_last_report = 0;
1254 				fq_cl->fcl_stat.fcl_ce_reported++;
1255 			} else if (pktsched_mark_ecn(pkt) == 0) {
1256 				fq_cl->fcl_stat.fcl_ce_marked++;
1257 			} else {
1258 				fq_cl->fcl_stat.fcl_ce_mark_failures++;
1259 			}
1260 		}
1261 		*pkt_timestamp = 0;
1262 	} else {
1263 		ok_to_drop = codel_ok_to_drop(pkt, fq, status, now);
1264 
1265 		if (status->dropping) {
1266 			if (!ok_to_drop) {
1267 				status->dropping = false;
1268 			}
1269 			/*
1270 			 * Time for the next drop.  Drop current packet and dequeue
1271 			 * next.  If the dequeue doesn't take us out of dropping
1272 			 * state, schedule the next drop.  A large backlog might
1273 			 * result in drop rates so high that the next drop should
1274 			 * happen now, hence the 'while' loop.
1275 			 */
1276 			while (now >= status->drop_next && status->dropping) {
1277 				status->count++;
1278 				IFCQ_CONVERT_LOCK(fqs->fqs_ifq);
1279 				KDBG(AQM_KTRACE_CONGESTION_INC, fq->fq_flowhash,
1280 				    AQM_KTRACE_FQ_GRP_SC_IDX(fq), fq->fq_bytes, 0);
1281 				if (FQ_CONGESTION_FEEDBACK_CAPABLE(fq) && (*pkt_flags & PKTF_FLOW_ADV)) {
1282 					fq->fq_congestion_cnt++;
1283 					/* Like in the case of ECN, return now and dequeue again. */
1284 					if (fq->fq_flowsrc == FLOWSRC_CHANNEL) {
1285 						fq_if_report_congestion(fqs, pkt, 1, 0, fq->fq_pkts_since_last_report);
1286 						fq_cl->fcl_stat.fcl_congestion_feedback++;
1287 					}
1288 					goto end;
1289 					/*
1290 					 * If congestion feedback is not supported,
1291 					 * try ECN first, if fail, drop the packet.
1292 					 */
1293 				} else if (FQ_IS_ECN_CAPABLE(fq) && pktsched_mark_ecn(pkt) == 0) {
1294 					fq_cl->fcl_stat.fcl_ce_marked++;
1295 					status->drop_next =
1296 					    codel_control_law(fast_inv_sqrt(status->count), status->drop_next, FQ_UPDATE_INTERVAL(fq));
1297 					KDBG(AQM_KTRACE_PKT_DROP, fq->fq_flowhash,
1298 					    status->count, status->drop_next, now);
1299 					/*
1300 					 * Since we are not dropping, return now and let the
1301 					 * caller dequeue again
1302 					 */
1303 					goto end;
1304 				} else {
1305 					/* Disable ECN marking for this flow if marking fails once */
1306 					FQ_CLEAR_ECN_CAPABLE(fq);
1307 					fq_drop_pkt(ifcq, ifp, pkt);
1308 					fq_cl->fcl_stat.fcl_ce_mark_failures++;
1309 					fq_cl->fcl_stat.fcl_high_delay_drop++;
1310 
1311 					fq_getq_flow_internal(fqs, fq, pkt);
1312 					if (!codel_ok_to_drop(pkt, fq, status, now)) {
1313 						/* exit dropping state when queuing delay falls below threshold */
1314 						status->dropping = false;
1315 					} else {
1316 						status->drop_next =
1317 						    codel_control_law(fast_inv_sqrt(status->count), status->drop_next, FQ_UPDATE_INTERVAL(fq));
1318 					}
1319 				}
1320 				KDBG(AQM_KTRACE_PKT_DROP, fq->fq_flowhash,
1321 				    status->count, status->drop_next, now);
1322 			}
1323 			/*
1324 			 * If we get here, we're not in drop state.  The 'ok_to_drop'
1325 			 * return from dodequeue means that the sojourn time has been
1326 			 * above 'TARGET' for 'INTERVAL', so enter drop state.
1327 			 */
1328 		} else if (ok_to_drop) {
1329 			IFCQ_CONVERT_LOCK(fqs->fqs_ifq);
1330 			KDBG(AQM_KTRACE_CONGESTION_INC, fq->fq_flowhash,
1331 			    AQM_KTRACE_FQ_GRP_SC_IDX(fq), fq->fq_bytes, 0);
1332 			if (FQ_CONGESTION_FEEDBACK_CAPABLE(fq) && (*pkt_flags & PKTF_FLOW_ADV)) {
1333 				fq->fq_congestion_cnt++;
1334 				if (fq->fq_flowsrc == FLOWSRC_CHANNEL) {
1335 					fq_if_report_congestion(fqs, pkt, 1, 0, fq->fq_pkts_since_last_report);
1336 					fq_cl->fcl_stat.fcl_congestion_feedback++;
1337 				}
1338 			} else if (FQ_IS_ECN_CAPABLE(fq) && pktsched_mark_ecn(pkt) == 0) {
1339 				fq_cl->fcl_stat.fcl_ce_marked++;
1340 			} else {
1341 				FQ_CLEAR_ECN_CAPABLE(fq);
1342 				fq_drop_pkt(ifcq, ifp, pkt);
1343 				fq_cl->fcl_stat.fcl_ce_mark_failures++;
1344 				fq_cl->fcl_stat.fcl_high_delay_drop++;
1345 
1346 				fq_getq_flow_internal(fqs, fq, pkt);
1347 				ok_to_drop = codel_ok_to_drop(pkt, fq, status, now);
1348 			}
1349 			status->dropping = true;
1350 
1351 			/*
1352 			 * If min went above TARGET close to when it last went
1353 			 * below, assume that the drop rate that controlled the
1354 			 * queue on the last cycle is a good starting point to
1355 			 * control it now.  ('drop_next' will be at most 'INTERVAL'
1356 			 * later than the time of the last drop, so 'now - drop_next'
1357 			 * is a good approximation of the time from the last drop
1358 			 * until now.)
1359 			 */
1360 			delta = status->count - status->lastcnt;
1361 			if (delta > 0 && now - status->drop_next <= 16 * FQ_UPDATE_INTERVAL(fq)) {
1362 				status->count = MAX(status->count - 1, 1);
1363 			} else {
1364 				status->count = 1;
1365 			}
1366 
1367 			status->drop_next =
1368 			    codel_control_law(fast_inv_sqrt(status->count), now, FQ_UPDATE_INTERVAL(fq));
1369 			status->lastcnt = status->count;
1370 
1371 			KDBG(AQM_KTRACE_PKT_DROP, fq->fq_flowhash,
1372 			    status->count, status->drop_next, now);
1373 		}
1374 	}
1375 
1376 end:
1377 	if (fqs->fqs_large_flow != fq || !fq_if_almost_at_drop_limit(fqs) ||
1378 	    fq_empty(fq, fqs->fqs_ptype)) {
1379 		FQ_CLEAR_OVERWHELMING(fq);
1380 	}
1381 	if ((fq->fq_flags & FQF_FLOWCTL_ON) && !FQ_IS_OVERWHELMING(fq)) {
1382 		fq_if_flow_feedback(fqs, fq, fq_cl);
1383 	}
1384 
1385 	if (fq_empty(fq, fqs->fqs_ptype)) {
1386 		/* Reset getqtime so that we don't count idle times */
1387 		fq->fq_getqtime = 0;
1388 		fq->codel_status.dropping = false;
1389 	} else {
1390 		fq->fq_getqtime = now;
1391 	}
1392 	fq_if_is_flow_heavy(fqs, fq);
1393 }
1394