Home | History | Annotate | Line # | Download | only in net
pktqueue.c revision 1.9
      1 /*	$NetBSD: pktqueue.c,v 1.9 2017/06/01 02:45:14 chs Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 2014 The NetBSD Foundation, Inc.
      5  * All rights reserved.
      6  *
      7  * This code is derived from software contributed to The NetBSD Foundation
      8  * by Mindaugas Rasiukevicius.
      9  *
     10  * Redistribution and use in source and binary forms, with or without
     11  * modification, are permitted provided that the following conditions
     12  * are met:
     13  * 1. Redistributions of source code must retain the above copyright
     14  *    notice, this list of conditions and the following disclaimer.
     15  * 2. Redistributions in binary form must reproduce the above copyright
     16  *    notice, this list of conditions and the following disclaimer in the
     17  *    documentation and/or other materials provided with the distribution.
     18  *
     19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     29  * POSSIBILITY OF SUCH DAMAGE.
     30  */
     31 
     32 /*
     33  * The packet queue (pktqueue) interface is a lockless IP input queue
     34  * which also abstracts and handles network ISR scheduling.  It provides
     35  * a mechanism to enable receiver-side packet steering (RPS).
     36  */
     37 
     38 #include <sys/cdefs.h>
     39 __KERNEL_RCSID(0, "$NetBSD: pktqueue.c,v 1.9 2017/06/01 02:45:14 chs Exp $");
     40 
     41 #include <sys/param.h>
     42 #include <sys/types.h>
     43 
     44 #include <sys/atomic.h>
     45 #include <sys/cpu.h>
     46 #include <sys/pcq.h>
     47 #include <sys/intr.h>
     48 #include <sys/mbuf.h>
     49 #include <sys/proc.h>
     50 #include <sys/percpu.h>
     51 
     52 #include <net/pktqueue.h>
     53 
     54 /*
     55  * WARNING: update this if struct pktqueue changes.
     56  */
     57 #define	PKTQ_CLPAD	\
     58     MAX(COHERENCY_UNIT, COHERENCY_UNIT - sizeof(kmutex_t) - sizeof(u_int))
     59 
     60 struct pktqueue {
     61 	/*
     62 	 * The lock used for a barrier mechanism.  The barrier counter,
     63 	 * as well as the drop counter, are managed atomically though.
     64 	 * Ensure this group is in a separate cache line.
     65 	 */
     66 	kmutex_t	pq_lock;
     67 	volatile u_int	pq_barrier;
     68 	uint8_t		_pad[PKTQ_CLPAD];
     69 
     70 	/* The size of the queue, counters and the interrupt handler. */
     71 	u_int		pq_maxlen;
     72 	percpu_t *	pq_counters;
     73 	void *		pq_sih;
     74 
     75 	/* Finally, per-CPU queues. */
     76 	pcq_t *		pq_queue[];
     77 };
     78 
     79 /* The counters of the packet queue. */
     80 #define	PQCNT_ENQUEUE	0
     81 #define	PQCNT_DEQUEUE	1
     82 #define	PQCNT_DROP	2
     83 #define	PQCNT_NCOUNTERS	3
     84 
     85 typedef struct {
     86 	uint64_t	count[PQCNT_NCOUNTERS];
     87 } pktq_counters_t;
     88 
     89 /* Special marker value used by pktq_barrier() mechanism. */
     90 #define	PKTQ_MARKER	((void *)(~0ULL))
     91 
     92 /*
     93  * The total size of pktqueue_t which depends on the number of CPUs.
     94  */
     95 #define	PKTQUEUE_STRUCT_LEN(ncpu)	\
     96     roundup2(offsetof(pktqueue_t, pq_queue[ncpu]), coherency_unit)
     97 
     98 pktqueue_t *
     99 pktq_create(size_t maxlen, void (*intrh)(void *), void *sc)
    100 {
    101 	const u_int sflags = SOFTINT_NET | SOFTINT_MPSAFE | SOFTINT_RCPU;
    102 	const size_t len = PKTQUEUE_STRUCT_LEN(ncpu);
    103 	pktqueue_t *pq;
    104 	percpu_t *pc;
    105 	void *sih;
    106 
    107 	pc = percpu_alloc(sizeof(pktq_counters_t));
    108 	if ((sih = softint_establish(sflags, intrh, sc)) == NULL) {
    109 		percpu_free(pc, sizeof(pktq_counters_t));
    110 		return NULL;
    111 	}
    112 
    113 	pq = kmem_zalloc(len, KM_SLEEP);
    114 	for (u_int i = 0; i < ncpu; i++) {
    115 		pq->pq_queue[i] = pcq_create(maxlen, KM_SLEEP);
    116 	}
    117 	mutex_init(&pq->pq_lock, MUTEX_DEFAULT, IPL_NONE);
    118 	pq->pq_maxlen = maxlen;
    119 	pq->pq_counters = pc;
    120 	pq->pq_sih = sih;
    121 
    122 	return pq;
    123 }
    124 
    125 void
    126 pktq_destroy(pktqueue_t *pq)
    127 {
    128 	const size_t len = PKTQUEUE_STRUCT_LEN(ncpu);
    129 
    130 	for (u_int i = 0; i < ncpu; i++) {
    131 		pcq_t *q = pq->pq_queue[i];
    132 		KASSERT(pcq_peek(q) == NULL);
    133 		pcq_destroy(q);
    134 	}
    135 	percpu_free(pq->pq_counters, sizeof(pktq_counters_t));
    136 	softint_disestablish(pq->pq_sih);
    137 	mutex_destroy(&pq->pq_lock);
    138 	kmem_free(pq, len);
    139 }
    140 
    141 /*
    142  * - pktq_inc_counter: increment the counter given an ID.
    143  * - pktq_collect_counts: handler to sum up the counts from each CPU.
    144  * - pktq_getcount: return the effective count given an ID.
    145  */
    146 
    147 static inline void
    148 pktq_inc_count(pktqueue_t *pq, u_int i)
    149 {
    150 	percpu_t *pc = pq->pq_counters;
    151 	pktq_counters_t *c;
    152 
    153 	c = percpu_getref(pc);
    154 	c->count[i]++;
    155 	percpu_putref(pc);
    156 }
    157 
    158 static void
    159 pktq_collect_counts(void *mem, void *arg, struct cpu_info *ci)
    160 {
    161 	const pktq_counters_t *c = mem;
    162 	pktq_counters_t *sum = arg;
    163 
    164 	for (u_int i = 0; i < PQCNT_NCOUNTERS; i++) {
    165 		sum->count[i] += c->count[i];
    166 	}
    167 }
    168 
    169 uint64_t
    170 pktq_get_count(pktqueue_t *pq, pktq_count_t c)
    171 {
    172 	pktq_counters_t sum;
    173 
    174 	if (c != PKTQ_MAXLEN) {
    175 		memset(&sum, 0, sizeof(sum));
    176 		percpu_foreach(pq->pq_counters, pktq_collect_counts, &sum);
    177 	}
    178 	switch (c) {
    179 	case PKTQ_NITEMS:
    180 		return sum.count[PQCNT_ENQUEUE] - sum.count[PQCNT_DEQUEUE];
    181 	case PKTQ_DROPS:
    182 		return sum.count[PQCNT_DROP];
    183 	case PKTQ_MAXLEN:
    184 		return pq->pq_maxlen;
    185 	}
    186 	return 0;
    187 }
    188 
    189 uint32_t
    190 pktq_rps_hash(const struct mbuf *m __unused)
    191 {
    192 	/*
    193 	 * XXX: No distribution yet; the softnet_lock contention
    194 	 * XXX: must be eliminated first.
    195 	 */
    196 	return 0;
    197 }
    198 
    199 /*
    200  * pktq_enqueue: inject the packet into the end of the queue.
    201  *
    202  * => Must be called from the interrupt or with the preemption disabled.
    203  * => Consumes the packet and returns true on success.
    204  * => Returns false on failure; caller is responsible to free the packet.
    205  */
    206 bool
    207 pktq_enqueue(pktqueue_t *pq, struct mbuf *m, const u_int hash __unused)
    208 {
    209 #if defined(_RUMPKERNEL) || defined(_RUMP_NATIVE_ABI)
    210 	const unsigned cpuid = curcpu()->ci_index;
    211 #else
    212 	const unsigned cpuid = hash % ncpu;
    213 #endif
    214 
    215 	KASSERT(kpreempt_disabled());
    216 
    217 	if (__predict_false(!pcq_put(pq->pq_queue[cpuid], m))) {
    218 		pktq_inc_count(pq, PQCNT_DROP);
    219 		return false;
    220 	}
    221 	softint_schedule_cpu(pq->pq_sih, cpu_lookup(cpuid));
    222 	pktq_inc_count(pq, PQCNT_ENQUEUE);
    223 	return true;
    224 }
    225 
    226 /*
    227  * pktq_dequeue: take a packet from the queue.
    228  *
    229  * => Must be called with preemption disabled.
    230  * => Must ensure there are not concurrent dequeue calls.
    231  */
    232 struct mbuf *
    233 pktq_dequeue(pktqueue_t *pq)
    234 {
    235 	const struct cpu_info *ci = curcpu();
    236 	const unsigned cpuid = cpu_index(ci);
    237 	struct mbuf *m;
    238 
    239 	m = pcq_get(pq->pq_queue[cpuid]);
    240 	if (__predict_false(m == PKTQ_MARKER)) {
    241 		/* Note the marker entry. */
    242 		atomic_inc_uint(&pq->pq_barrier);
    243 		return NULL;
    244 	}
    245 	if (__predict_true(m != NULL)) {
    246 		pktq_inc_count(pq, PQCNT_DEQUEUE);
    247 	}
    248 	return m;
    249 }
    250 
    251 /*
    252  * pktq_barrier: waits for a grace period when all packets enqueued at
    253  * the moment of calling this routine will be processed.  This is used
    254  * to ensure that e.g. packets referencing some interface were drained.
    255  */
    256 void
    257 pktq_barrier(pktqueue_t *pq)
    258 {
    259 	u_int pending = 0;
    260 
    261 	mutex_enter(&pq->pq_lock);
    262 	KASSERT(pq->pq_barrier == 0);
    263 
    264 	for (u_int i = 0; i < ncpu; i++) {
    265 		pcq_t *q = pq->pq_queue[i];
    266 
    267 		/* If the queue is empty - nothing to do. */
    268 		if (pcq_peek(q) == NULL) {
    269 			continue;
    270 		}
    271 		/* Otherwise, put the marker and entry. */
    272 		while (!pcq_put(q, PKTQ_MARKER)) {
    273 			kpause("pktqsync", false, 1, NULL);
    274 		}
    275 		kpreempt_disable();
    276 		softint_schedule_cpu(pq->pq_sih, cpu_lookup(i));
    277 		kpreempt_enable();
    278 		pending++;
    279 	}
    280 
    281 	/* Wait for each queue to process the markers. */
    282 	while (pq->pq_barrier != pending) {
    283 		kpause("pktqsync", false, 1, NULL);
    284 	}
    285 	pq->pq_barrier = 0;
    286 	mutex_exit(&pq->pq_lock);
    287 }
    288 
    289 /*
    290  * pktq_flush: free mbufs in all queues.
    291  *
    292  * => The caller must ensure there are no concurrent writers or flush calls.
    293  */
    294 void
    295 pktq_flush(pktqueue_t *pq)
    296 {
    297 	struct mbuf *m;
    298 
    299 	for (u_int i = 0; i < ncpu; i++) {
    300 		while ((m = pcq_get(pq->pq_queue[i])) != NULL) {
    301 			pktq_inc_count(pq, PQCNT_DEQUEUE);
    302 			m_freem(m);
    303 		}
    304 	}
    305 }
    306 
    307 /*
    308  * pktq_set_maxlen: create per-CPU queues using a new size and replace
    309  * the existing queues without losing any packets.
    310  */
    311 int
    312 pktq_set_maxlen(pktqueue_t *pq, size_t maxlen)
    313 {
    314 	const u_int slotbytes = ncpu * sizeof(pcq_t *);
    315 	pcq_t **qs;
    316 
    317 	if (!maxlen || maxlen > PCQ_MAXLEN)
    318 		return EINVAL;
    319 	if (pq->pq_maxlen == maxlen)
    320 		return 0;
    321 
    322 	/* First, allocate the new queues and replace them. */
    323 	qs = kmem_zalloc(slotbytes, KM_SLEEP);
    324 	for (u_int i = 0; i < ncpu; i++) {
    325 		qs[i] = pcq_create(maxlen, KM_SLEEP);
    326 	}
    327 	mutex_enter(&pq->pq_lock);
    328 	for (u_int i = 0; i < ncpu; i++) {
    329 		/* Swap: store of a word is atomic. */
    330 		pcq_t *q = pq->pq_queue[i];
    331 		pq->pq_queue[i] = qs[i];
    332 		qs[i] = q;
    333 	}
    334 	pq->pq_maxlen = maxlen;
    335 	mutex_exit(&pq->pq_lock);
    336 
    337 	/*
    338 	 * At this point, the new packets are flowing into the new
    339 	 * queues.  However, the old queues may have some packets
    340 	 * present which are no longer being processed.  We are going
    341 	 * to re-enqueue them.  This may change the order of packet
    342 	 * arrival, but it is not considered an issue.
    343 	 *
    344 	 * There may be in-flight interrupts calling pktq_dequeue()
    345 	 * which reference the old queues.  Issue a barrier to ensure
    346 	 * that we are going to be the only pcq_get() callers on the
    347 	 * old queues.
    348 	 */
    349 	pktq_barrier(pq);
    350 
    351 	for (u_int i = 0; i < ncpu; i++) {
    352 		struct mbuf *m;
    353 
    354 		while ((m = pcq_get(qs[i])) != NULL) {
    355 			while (!pcq_put(pq->pq_queue[i], m)) {
    356 				kpause("pktqrenq", false, 1, NULL);
    357 			}
    358 		}
    359 		pcq_destroy(qs[i]);
    360 	}
    361 
    362 	/* Well, that was fun. */
    363 	kmem_free(qs, slotbytes);
    364 	return 0;
    365 }
    366 
    367 int
    368 sysctl_pktq_maxlen(SYSCTLFN_ARGS, pktqueue_t *pq)
    369 {
    370 	u_int nmaxlen = pktq_get_count(pq, PKTQ_MAXLEN);
    371 	struct sysctlnode node = *rnode;
    372 	int error;
    373 
    374 	node.sysctl_data = &nmaxlen;
    375 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
    376 	if (error || newp == NULL)
    377 		return error;
    378 	return pktq_set_maxlen(pq, nmaxlen);
    379 }
    380 
    381 int
    382 sysctl_pktq_count(SYSCTLFN_ARGS, pktqueue_t *pq, u_int count_id)
    383 {
    384 	int count = pktq_get_count(pq, count_id);
    385 	struct sysctlnode node = *rnode;
    386 	node.sysctl_data = &count;
    387 	return sysctl_lookup(SYSCTLFN_CALL(&node));
    388 }
    389