2 * Copyright (c) 2010 Riccardo Panicucci, Universita` di Pisa
3 * Copyright (c) 2000-2002 Luigi Rizzo, Universita` di Pisa
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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.
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
33 #include <sys/malloc.h>
34 #include <sys/socket.h>
35 #include <sys/socketvar.h>
36 #include <sys/kernel.h>
39 #include <sys/module.h>
40 #include <sys/rwlock.h>
41 #include <net/if.h> /* IFNAMSIZ */
42 #include <netinet/in.h>
43 #include <netinet/ip_var.h> /* ipfw_rule_ref */
44 #include <netinet/ip_fw.h> /* flow_id */
45 #include <netinet/ip_dummynet.h>
46 #include <netpfil/ipfw/ip_fw_private.h>
47 #include <netpfil/ipfw/dn_heap.h>
48 #include <netpfil/ipfw/ip_dn_private.h>
50 #include <netpfil/ipfw/dn_aqm.h>
52 #include <netpfil/ipfw/dn_sched.h>
58 #define MAX64(x,y) (( (int64_t) ( (y)-(x) )) > 0 ) ? (y) : (x)
62 * timestamps are computed on 64 bit using fixed point arithmetic.
63 * LMAX_BITS, WMAX_BITS are the max number of bits for the packet len
64 * and sum of weights, respectively. FRAC_BITS is the number of
65 * fractional bits. We want FRAC_BITS >> WMAX_BITS to avoid too large
66 * errors when computing the inverse, FRAC_BITS < 32 so we can do 1/w
67 * using an unsigned 32-bit division, and to avoid wraparounds we need
68 * LMAX_BITS + WMAX_BITS + FRAC_BITS << 64
70 * FRAC_BITS = 26, LMAX_BITS=14, WMAX_BITS = 19
73 #define FRAC_BITS 28 /* shift for fixed point arithmetic */
74 #define ONE_FP (1UL << FRAC_BITS)
78 * Private information for the scheduler instance:
79 * sch_heap (key is Finish time) returns the next queue to serve
80 * ne_heap (key is Start time) stores not-eligible queues
81 * idle_heap (key=start/finish time) stores idle flows. It must
82 * support extract-from-middle.
83 * A flow is only in 1 of the three heaps.
84 * XXX todo: use a more efficient data structure, e.g. a tree sorted
85 * by F with min_subtree(S) in each node
88 struct dn_heap sch_heap; /* top extract - key Finish time */
89 struct dn_heap ne_heap; /* top extract - key Start time */
90 struct dn_heap idle_heap; /* random extract - key Start=Finish time */
91 uint64_t V; /* virtual time */
92 uint32_t inv_wsum; /* inverse of sum of weights */
93 uint32_t wsum; /* sum of weights */
98 uint64_t S, F; /* start time, finish time */
99 uint32_t inv_w; /* ONE_FP / weight */
100 int32_t heap_pos; /* position (index) of struct in heap */
104 * This file implements a WF2Q+ scheduler as it has been in dummynet
106 * The scheduler supports per-flow queues and has O(log N) complexity.
108 * WF2Q+ needs to drain entries from the idle heap so that we
109 * can keep the sum of weights up to date. We can do it whenever
110 * we get a chance, or periodically, or following some other
111 * strategy. The function idle_check() drains at most N elements
112 * from the idle heap.
115 idle_check(struct wf2qp_si *si, int n, int force)
117 struct dn_heap *h = &si->idle_heap;
118 while (n-- > 0 && h->elements > 0 &&
119 (force || DN_KEY_LT(HEAP_TOP(h)->key, si->V))) {
120 struct dn_queue *q = HEAP_TOP(h)->object;
121 struct wf2qp_queue *alg_fq = (struct wf2qp_queue *)q;
123 heap_extract(h, NULL);
124 /* XXX to let the flowset delete the queue we should
125 * mark it as 'unused' by the scheduler.
127 alg_fq->S = alg_fq->F + 1; /* Mark timestamp as invalid. */
128 si->wsum -= q->fs->fs.par[0]; /* adjust sum of weights */
130 si->inv_wsum = ONE_FP/si->wsum;
135 wf2qp_enqueue(struct dn_sch_inst *_si, struct dn_queue *q, struct mbuf *m)
137 struct dn_fsk *fs = q->fs;
138 struct wf2qp_si *si = (struct wf2qp_si *)(_si + 1);
139 struct wf2qp_queue *alg_fq;
140 uint64_t len = m->m_pkthdr.len;
142 if (m != q->mq.head) {
143 if (dn_enqueue(q, m, 0)) /* packet was dropped */
145 if (m != q->mq.head) /* queue was already busy */
149 /* If reach this point, queue q was idle */
150 alg_fq = (struct wf2qp_queue *)q;
152 if (DN_KEY_LT(alg_fq->F, alg_fq->S)) {
153 /* F<S means timestamps are invalid ->brand new queue. */
154 alg_fq->S = si->V; /* init start time */
155 si->wsum += fs->fs.par[0]; /* add weight of new queue. */
156 si->inv_wsum = ONE_FP/si->wsum;
157 } else { /* if it was idle then it was in the idle heap */
158 heap_extract(&si->idle_heap, q);
159 alg_fq->S = MAX64(alg_fq->F, si->V); /* compute new S */
161 alg_fq->F = alg_fq->S + len * alg_fq->inv_w;
163 /* if nothing is backlogged, make sure this flow is eligible */
164 if (si->ne_heap.elements == 0 && si->sch_heap.elements == 0)
165 si->V = MAX64(alg_fq->S, si->V);
168 * Look at eligibility. A flow is not eligibile if S>V (when
169 * this happens, it means that there is some other flow already
170 * scheduled for the same pipe, so the sch_heap cannot be
171 * empty). If the flow is not eligible we just store it in the
172 * ne_heap. Otherwise, we store in the sch_heap.
173 * Note that for all flows in sch_heap (SCH), S_i <= V,
174 * and for all flows in ne_heap (NEH), S_i > V.
175 * So when we need to compute max(V, min(S_i)) forall i in
176 * SCH+NEH, we only need to look into NEH.
178 if (DN_KEY_LT(si->V, alg_fq->S)) {
179 /* S>V means flow Not eligible. */
180 if (si->sch_heap.elements == 0)
181 D("++ ouch! not eligible but empty scheduler!");
182 heap_insert(&si->ne_heap, alg_fq->S, q);
184 heap_insert(&si->sch_heap, alg_fq->F, q);
189 /* XXX invariant: sch > 0 || V >= min(S in neh) */
191 wf2qp_dequeue(struct dn_sch_inst *_si)
193 /* Access scheduler instance private data */
194 struct wf2qp_si *si = (struct wf2qp_si *)(_si + 1);
197 struct dn_heap *sch = &si->sch_heap;
198 struct dn_heap *neh = &si->ne_heap;
199 struct wf2qp_queue *alg_fq;
201 if (sch->elements == 0 && neh->elements == 0) {
202 /* we have nothing to do. We could kill the idle heap
203 * altogether and reset V
205 idle_check(si, 0x7fffffff, 1);
207 si->wsum = 0; /* should be set already */
208 return NULL; /* quick return if nothing to do */
210 idle_check(si, 1, 0); /* drain something from the idle heap */
212 /* make sure at least one element is eligible, bumping V
213 * and moving entries that have become eligible.
214 * We need to repeat the first part twice, before and
215 * after extracting the candidate, or enqueue() will
216 * find the data structure in a wrong state.
221 * Compute V = max(V, min(S_i)). Remember that all elements
222 * in sch have by definition S_i <= V so if sch is not empty,
223 * V is surely the max and we must not update it. Conversely,
224 * if sch is empty we only need to look at neh.
225 * We don't need to move the queues, as it will be done at the
228 if (sch->elements == 0 && neh->elements > 0) {
229 si->V = MAX64(si->V, HEAP_TOP(neh)->key);
231 while (neh->elements > 0 &&
232 DN_KEY_LEQ(HEAP_TOP(neh)->key, si->V)) {
233 q = HEAP_TOP(neh)->object;
234 alg_fq = (struct wf2qp_queue *)q;
235 heap_extract(neh, NULL);
236 heap_insert(sch, alg_fq->F, q);
238 if (m) /* pkt found in previous iteration */
240 /* ok we have at least one eligible pkt */
241 q = HEAP_TOP(sch)->object;
242 alg_fq = (struct wf2qp_queue *)q;
244 heap_extract(sch, NULL); /* Remove queue from heap. */
245 si->V += (uint64_t)(m->m_pkthdr.len) * si->inv_wsum;
246 alg_fq->S = alg_fq->F; /* Update start time. */
247 if (q->mq.head == 0) { /* not backlogged any more. */
248 heap_insert(&si->idle_heap, alg_fq->F, q);
249 } else { /* Still backlogged. */
250 /* Update F, store in neh or sch */
251 uint64_t len = q->mq.head->m_pkthdr.len;
252 alg_fq->F += len * alg_fq->inv_w;
253 if (DN_KEY_LEQ(alg_fq->S, si->V)) {
254 heap_insert(sch, alg_fq->F, q);
256 heap_insert(neh, alg_fq->S, q);
264 wf2qp_new_sched(struct dn_sch_inst *_si)
266 struct wf2qp_si *si = (struct wf2qp_si *)(_si + 1);
267 int ofs = offsetof(struct wf2qp_queue, heap_pos);
269 /* all heaps support extract from middle */
270 if (heap_init(&si->idle_heap, 16, ofs) ||
271 heap_init(&si->sch_heap, 16, ofs) ||
272 heap_init(&si->ne_heap, 16, ofs)) {
273 heap_free(&si->ne_heap);
274 heap_free(&si->sch_heap);
275 heap_free(&si->idle_heap);
282 wf2qp_free_sched(struct dn_sch_inst *_si)
284 struct wf2qp_si *si = (struct wf2qp_si *)(_si + 1);
286 heap_free(&si->sch_heap);
287 heap_free(&si->ne_heap);
288 heap_free(&si->idle_heap);
294 wf2qp_new_fsk(struct dn_fsk *fs)
296 ipdn_bound_var(&fs->fs.par[0], 1,
297 1, 100, "WF2Q+ weight");
302 wf2qp_new_queue(struct dn_queue *_q)
304 struct wf2qp_queue *q = (struct wf2qp_queue *)_q;
306 _q->ni.oid.subtype = DN_SCHED_WF2QP;
307 q->F = 0; /* not strictly necessary */
308 q->S = q->F + 1; /* mark timestamp as invalid. */
309 q->inv_w = ONE_FP / _q->fs->fs.par[0];
310 if (_q->mq.head != NULL) {
311 wf2qp_enqueue(_q->_si, _q, _q->mq.head);
317 * Called when the infrastructure removes a queue (e.g. flowset
318 * is reconfigured). Nothing to do if we did not 'own' the queue,
319 * otherwise remove it from the right heap and adjust the sum
323 wf2qp_free_queue(struct dn_queue *q)
325 struct wf2qp_queue *alg_fq = (struct wf2qp_queue *)q;
326 struct wf2qp_si *si = (struct wf2qp_si *)(q->_si + 1);
328 if (alg_fq->S >= alg_fq->F + 1)
329 return 0; /* nothing to do, not in any heap */
330 si->wsum -= q->fs->fs.par[0];
332 si->inv_wsum = ONE_FP/si->wsum;
334 /* extract from the heap. XXX TODO we may need to adjust V
335 * to make sure the invariants hold.
337 if (q->mq.head == NULL) {
338 heap_extract(&si->idle_heap, q);
339 } else if (DN_KEY_LT(si->V, alg_fq->S)) {
340 heap_extract(&si->ne_heap, q);
342 heap_extract(&si->sch_heap, q);
348 * WF2Q+ scheduler descriptor
349 * contains the type of the scheduler, the name, the size of the
350 * structures and function pointers.
352 static struct dn_alg wf2qp_desc = {
353 _SI( .type = ) DN_SCHED_WF2QP,
354 _SI( .name = ) "WF2Q+",
355 _SI( .flags = ) DN_MULTIQUEUE,
357 /* we need extra space in the si and the queue */
358 _SI( .schk_datalen = ) 0,
359 _SI( .si_datalen = ) sizeof(struct wf2qp_si),
360 _SI( .q_datalen = ) sizeof(struct wf2qp_queue) -
361 sizeof(struct dn_queue),
363 _SI( .enqueue = ) wf2qp_enqueue,
364 _SI( .dequeue = ) wf2qp_dequeue,
366 _SI( .config = ) NULL,
367 _SI( .destroy = ) NULL,
368 _SI( .new_sched = ) wf2qp_new_sched,
369 _SI( .free_sched = ) wf2qp_free_sched,
371 _SI( .new_fsk = ) wf2qp_new_fsk,
372 _SI( .free_fsk = ) NULL,
374 _SI( .new_queue = ) wf2qp_new_queue,
375 _SI( .free_queue = ) wf2qp_free_queue,
377 _SI( .getconfig = ) NULL,
383 DECLARE_DNSCHED_MODULE(dn_wf2qp, &wf2qp_desc);