2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2016, 2017 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "dpif-netdev.h"
25 #include <netinet/in.h>
29 #include <sys/ioctl.h>
30 #include <sys/socket.h>
35 #include <rte_cycles.h>
40 #include "conntrack.h"
44 #include "dp-packet.h"
46 #include "dpif-provider.h"
48 #include "fat-rwlock.h"
54 #include "netdev-vport.h"
56 #include "odp-execute.h"
58 #include "openvswitch/dynamic-string.h"
59 #include "openvswitch/list.h"
60 #include "openvswitch/match.h"
61 #include "openvswitch/ofp-print.h"
62 #include "openvswitch/ofp-util.h"
63 #include "openvswitch/ofpbuf.h"
64 #include "openvswitch/shash.h"
65 #include "openvswitch/vlog.h"
69 #include "poll-loop.h"
76 #include "tnl-neigh-cache.h"
77 #include "tnl-ports.h"
81 VLOG_DEFINE_THIS_MODULE(dpif_netdev
);
83 #define FLOW_DUMP_MAX_BATCH 50
84 /* Use per thread recirc_depth to prevent recirculation loop. */
85 #define MAX_RECIRC_DEPTH 5
86 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth
, 0)
88 /* Configuration parameters. */
89 enum { MAX_FLOWS
= 65536 }; /* Maximum number of flows in flow table. */
90 enum { MAX_METERS
= 65536 }; /* Maximum number of meters. */
91 enum { MAX_BANDS
= 8 }; /* Maximum number of bands / meter. */
92 enum { N_METER_LOCKS
= 64 }; /* Maximum number of meters. */
94 /* Protects against changes to 'dp_netdevs'. */
95 static struct ovs_mutex dp_netdev_mutex
= OVS_MUTEX_INITIALIZER
;
97 /* Contains all 'struct dp_netdev's. */
98 static struct shash dp_netdevs
OVS_GUARDED_BY(dp_netdev_mutex
)
99 = SHASH_INITIALIZER(&dp_netdevs
);
101 static struct vlog_rate_limit upcall_rl
= VLOG_RATE_LIMIT_INIT(600, 600);
103 #define DP_NETDEV_CS_SUPPORTED_MASK (CS_NEW | CS_ESTABLISHED | CS_RELATED \
104 | CS_INVALID | CS_REPLY_DIR | CS_TRACKED \
105 | CS_SRC_NAT | CS_DST_NAT)
106 #define DP_NETDEV_CS_UNSUPPORTED_MASK (~(uint32_t)DP_NETDEV_CS_SUPPORTED_MASK)
108 static struct odp_support dp_netdev_support
= {
109 .max_vlan_headers
= SIZE_MAX
,
110 .max_mpls_depth
= SIZE_MAX
,
116 .ct_state_nat
= true,
117 .ct_orig_tuple
= true,
118 .ct_orig_tuple6
= true,
121 /* Stores a miniflow with inline values */
123 struct netdev_flow_key
{
124 uint32_t hash
; /* Hash function differs for different users. */
125 uint32_t len
; /* Length of the following miniflow (incl. map). */
127 uint64_t buf
[FLOW_MAX_PACKET_U64S
];
130 /* Exact match cache for frequently used flows
132 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
133 * search its entries for a miniflow that matches exactly the miniflow of the
134 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
136 * A cache entry holds a reference to its 'dp_netdev_flow'.
138 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
139 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
140 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
141 * value is the index of a cache entry where the miniflow could be.
147 * Each pmd_thread has its own private exact match cache.
148 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
151 #define EM_FLOW_HASH_SHIFT 13
152 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
153 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
154 #define EM_FLOW_HASH_SEGS 2
156 /* Default EMC insert probability is 1 / DEFAULT_EM_FLOW_INSERT_INV_PROB */
157 #define DEFAULT_EM_FLOW_INSERT_INV_PROB 100
158 #define DEFAULT_EM_FLOW_INSERT_MIN (UINT32_MAX / \
159 DEFAULT_EM_FLOW_INSERT_INV_PROB)
162 struct dp_netdev_flow
*flow
;
163 struct netdev_flow_key key
; /* key.hash used for emc hash value. */
167 struct emc_entry entries
[EM_FLOW_HASH_ENTRIES
];
168 int sweep_idx
; /* For emc_cache_slow_sweep(). */
171 /* Iterate in the exact match cache through every entry that might contain a
172 * miniflow with hash 'HASH'. */
173 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
174 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
175 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
176 i__ < EM_FLOW_HASH_SEGS; \
177 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
179 /* Simple non-wildcarding single-priority classifier. */
181 /* Time in ms between successive optimizations of the dpcls subtable vector */
182 #define DPCLS_OPTIMIZATION_INTERVAL 1000
184 /* Time in ms of the interval in which rxq processing cycles used in
185 * rxq to pmd assignments is measured and stored. */
186 #define PMD_RXQ_INTERVAL_LEN 10000
188 /* Number of intervals for which cycles are stored
189 * and used during rxq to pmd assignment. */
190 #define PMD_RXQ_INTERVAL_MAX 6
193 struct cmap_node node
; /* Within dp_netdev_pmd_thread.classifiers */
195 struct cmap subtables_map
;
196 struct pvector subtables
;
199 /* A rule to be inserted to the classifier. */
201 struct cmap_node cmap_node
; /* Within struct dpcls_subtable 'rules'. */
202 struct netdev_flow_key
*mask
; /* Subtable's mask. */
203 struct netdev_flow_key flow
; /* Matching key. */
204 /* 'flow' must be the last field, additional space is allocated here. */
207 static void dpcls_init(struct dpcls
*);
208 static void dpcls_destroy(struct dpcls
*);
209 static void dpcls_sort_subtable_vector(struct dpcls
*);
210 static void dpcls_insert(struct dpcls
*, struct dpcls_rule
*,
211 const struct netdev_flow_key
*mask
);
212 static void dpcls_remove(struct dpcls
*, struct dpcls_rule
*);
213 static bool dpcls_lookup(struct dpcls
*cls
,
214 const struct netdev_flow_key keys
[],
215 struct dpcls_rule
**rules
, size_t cnt
,
218 /* Set of supported meter flags */
219 #define DP_SUPPORTED_METER_FLAGS_MASK \
220 (OFPMF13_STATS | OFPMF13_PKTPS | OFPMF13_KBPS | OFPMF13_BURST)
222 /* Set of supported meter band types */
223 #define DP_SUPPORTED_METER_BAND_TYPES \
224 ( 1 << OFPMBT13_DROP )
226 struct dp_meter_band
{
227 struct ofputil_meter_band up
; /* type, prec_level, pad, rate, burst_size */
228 uint32_t bucket
; /* In 1/1000 packets (for PKTPS), or in bits (for KBPS) */
229 uint64_t packet_count
;
236 uint32_t max_delta_t
;
238 uint64_t packet_count
;
240 struct dp_meter_band bands
[];
243 /* Datapath based on the network device interface from netdev.h.
249 * Some members, marked 'const', are immutable. Accessing other members
250 * requires synchronization, as noted in more detail below.
252 * Acquisition order is, from outermost to innermost:
254 * dp_netdev_mutex (global)
259 const struct dpif_class
*const class;
260 const char *const name
;
262 struct ovs_refcount ref_cnt
;
263 atomic_flag destroyed
;
267 * Any lookup into 'ports' or any access to the dp_netdev_ports found
268 * through 'ports' requires taking 'port_mutex'. */
269 struct ovs_mutex port_mutex
;
271 struct seq
*port_seq
; /* Incremented whenever a port changes. */
274 struct ovs_mutex meter_locks
[N_METER_LOCKS
];
275 struct dp_meter
*meters
[MAX_METERS
]; /* Meter bands. */
277 /* Probability of EMC insertions is a factor of 'emc_insert_min'.*/
278 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
) atomic_uint32_t emc_insert_min
;
280 /* Protects access to ofproto-dpif-upcall interface during revalidator
281 * thread synchronization. */
282 struct fat_rwlock upcall_rwlock
;
283 upcall_callback
*upcall_cb
; /* Callback function for executing upcalls. */
286 /* Callback function for notifying the purging of dp flows (during
287 * reseting pmd deletion). */
288 dp_purge_callback
*dp_purge_cb
;
291 /* Stores all 'struct dp_netdev_pmd_thread's. */
292 struct cmap poll_threads
;
293 /* id pool for per thread static_tx_qid. */
294 struct id_pool
*tx_qid_pool
;
295 struct ovs_mutex tx_qid_pool_mutex
;
297 /* Protects the access of the 'struct dp_netdev_pmd_thread'
298 * instance for non-pmd thread. */
299 struct ovs_mutex non_pmd_mutex
;
301 /* Each pmd thread will store its pointer to
302 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
303 ovsthread_key_t per_pmd_key
;
305 struct seq
*reconfigure_seq
;
306 uint64_t last_reconfigure_seq
;
308 /* Cpu mask for pin of pmd threads. */
311 uint64_t last_tnl_conf_seq
;
313 struct conntrack conntrack
;
316 static void meter_lock(const struct dp_netdev
*dp
, uint32_t meter_id
)
317 OVS_ACQUIRES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
319 ovs_mutex_lock(&dp
->meter_locks
[meter_id
% N_METER_LOCKS
]);
322 static void meter_unlock(const struct dp_netdev
*dp
, uint32_t meter_id
)
323 OVS_RELEASES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
325 ovs_mutex_unlock(&dp
->meter_locks
[meter_id
% N_METER_LOCKS
]);
329 static struct dp_netdev_port
*dp_netdev_lookup_port(const struct dp_netdev
*dp
,
331 OVS_REQUIRES(dp
->port_mutex
);
334 DP_STAT_EXACT_HIT
, /* Packets that had an exact match (emc). */
335 DP_STAT_MASKED_HIT
, /* Packets that matched in the flow table. */
336 DP_STAT_MISS
, /* Packets that did not match. */
337 DP_STAT_LOST
, /* Packets not passed up to the client. */
338 DP_STAT_LOOKUP_HIT
, /* Number of subtable lookups for flow table
343 enum pmd_cycles_counter_type
{
344 PMD_CYCLES_IDLE
, /* Cycles spent idle or unsuccessful polling */
345 PMD_CYCLES_PROCESSING
, /* Cycles spent successfully polling and
346 * processing polled packets */
350 enum rxq_cycles_counter_type
{
351 RXQ_CYCLES_PROC_CURR
, /* Cycles spent successfully polling and
352 processing packets during the current
354 RXQ_CYCLES_PROC_HIST
, /* Total cycles of all intervals that are used
355 during rxq to pmd assignment. */
359 #define XPS_TIMEOUT_MS 500LL
361 /* Contained by struct dp_netdev_port's 'rxqs' member. */
362 struct dp_netdev_rxq
{
363 struct dp_netdev_port
*port
;
364 struct netdev_rxq
*rx
;
365 unsigned core_id
; /* Core to which this queue should be
366 pinned. OVS_CORE_UNSPEC if the
367 queue doesn't need to be pinned to a
369 struct dp_netdev_pmd_thread
*pmd
; /* pmd thread that polls this queue. */
371 /* Counters of cycles spent successfully polling and processing pkts. */
372 atomic_ullong cycles
[RXQ_N_CYCLES
];
373 /* We store PMD_RXQ_INTERVAL_MAX intervals of data for an rxq and then
374 sum them to yield the cycles used for an rxq. */
375 atomic_ullong cycles_intrvl
[PMD_RXQ_INTERVAL_MAX
];
376 unsigned intrvl_idx
; /* Write index for 'cycles_intrvl'. */
379 /* A port in a netdev-based datapath. */
380 struct dp_netdev_port
{
382 bool dynamic_txqs
; /* If true XPS will be used. */
383 bool need_reconfigure
; /* True if we should reconfigure netdev. */
384 struct netdev
*netdev
;
385 struct hmap_node node
; /* Node in dp_netdev's 'ports'. */
386 struct netdev_saved_flags
*sf
;
387 struct dp_netdev_rxq
*rxqs
;
388 unsigned n_rxq
; /* Number of elements in 'rxqs' */
389 unsigned *txq_used
; /* Number of threads that use each tx queue. */
390 struct ovs_mutex txq_used_mutex
;
391 char *type
; /* Port type as requested by user. */
392 char *rxq_affinity_list
; /* Requested affinity of rx queues. */
395 /* Contained by struct dp_netdev_flow's 'stats' member. */
396 struct dp_netdev_flow_stats
{
397 atomic_llong used
; /* Last used time, in monotonic msecs. */
398 atomic_ullong packet_count
; /* Number of packets matched. */
399 atomic_ullong byte_count
; /* Number of bytes matched. */
400 atomic_uint16_t tcp_flags
; /* Bitwise-OR of seen tcp_flags values. */
403 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
409 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
410 * its pmd thread's classifier. The text below calls this classifier 'cls'.
415 * The thread safety rules described here for "struct dp_netdev_flow" are
416 * motivated by two goals:
418 * - Prevent threads that read members of "struct dp_netdev_flow" from
419 * reading bad data due to changes by some thread concurrently modifying
422 * - Prevent two threads making changes to members of a given "struct
423 * dp_netdev_flow" from interfering with each other.
429 * A flow 'flow' may be accessed without a risk of being freed during an RCU
430 * grace period. Code that needs to hold onto a flow for a while
431 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
433 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
434 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
437 * Some members, marked 'const', are immutable. Accessing other members
438 * requires synchronization, as noted in more detail below.
440 struct dp_netdev_flow
{
441 const struct flow flow
; /* Unmasked flow that created this entry. */
442 /* Hash table index by unmasked flow. */
443 const struct cmap_node node
; /* In owning dp_netdev_pmd_thread's */
445 const ovs_u128 ufid
; /* Unique flow identifier. */
446 const unsigned pmd_id
; /* The 'core_id' of pmd thread owning this */
449 /* Number of references.
450 * The classifier owns one reference.
451 * Any thread trying to keep a rule from being freed should hold its own
453 struct ovs_refcount ref_cnt
;
458 struct dp_netdev_flow_stats stats
;
461 OVSRCU_TYPE(struct dp_netdev_actions
*) actions
;
463 /* While processing a group of input packets, the datapath uses the next
464 * member to store a pointer to the output batch for the flow. It is
465 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
466 * packet_batch_per_flow_init() and packet_batch_per_flow_execute()). */
467 struct packet_batch_per_flow
*batch
;
469 /* Packet classification. */
470 struct dpcls_rule cr
; /* In owning dp_netdev's 'cls'. */
471 /* 'cr' must be the last member. */
474 static void dp_netdev_flow_unref(struct dp_netdev_flow
*);
475 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*);
476 static int dpif_netdev_flow_from_nlattrs(const struct nlattr
*, uint32_t,
477 struct flow
*, bool);
479 /* A set of datapath actions within a "struct dp_netdev_flow".
485 * A struct dp_netdev_actions 'actions' is protected with RCU. */
486 struct dp_netdev_actions
{
487 /* These members are immutable: they do not change during the struct's
489 unsigned int size
; /* Size of 'actions', in bytes. */
490 struct nlattr actions
[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
493 struct dp_netdev_actions
*dp_netdev_actions_create(const struct nlattr
*,
495 struct dp_netdev_actions
*dp_netdev_flow_get_actions(
496 const struct dp_netdev_flow
*);
497 static void dp_netdev_actions_free(struct dp_netdev_actions
*);
499 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
500 struct dp_netdev_pmd_stats
{
501 /* Indexed by DP_STAT_*. */
502 atomic_ullong n
[DP_N_STATS
];
505 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
506 struct dp_netdev_pmd_cycles
{
507 /* Indexed by PMD_CYCLES_*. */
508 atomic_ullong n
[PMD_N_CYCLES
];
511 struct polled_queue
{
512 struct dp_netdev_rxq
*rxq
;
516 /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */
518 struct dp_netdev_rxq
*rxq
;
519 struct hmap_node node
;
522 /* Contained by struct dp_netdev_pmd_thread's 'send_port_cache',
523 * 'tnl_port_cache' or 'tx_ports'. */
525 struct dp_netdev_port
*port
;
528 struct hmap_node node
;
531 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
532 * the performance overhead of interrupt processing. Therefore netdev can
533 * not implement rx-wait for these devices. dpif-netdev needs to poll
534 * these device to check for recv buffer. pmd-thread does polling for
535 * devices assigned to itself.
537 * DPDK used PMD for accessing NIC.
539 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
540 * I/O of all non-pmd threads. There will be no actual thread created
543 * Each struct has its own flow cache and classifier per managed ingress port.
544 * For packets received on ingress port, a look up is done on corresponding PMD
545 * thread's flow cache and in case of a miss, lookup is performed in the
546 * corresponding classifier of port. Packets are executed with the found
547 * actions in either case.
549 struct dp_netdev_pmd_thread
{
550 struct dp_netdev
*dp
;
551 struct ovs_refcount ref_cnt
; /* Every reference must be refcount'ed. */
552 struct cmap_node node
; /* In 'dp->poll_threads'. */
554 pthread_cond_t cond
; /* For synchronizing pmd thread reload. */
555 struct ovs_mutex cond_mutex
; /* Mutex for condition variable. */
557 /* Per thread exact-match cache. Note, the instance for cpu core
558 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
559 * need to be protected by 'non_pmd_mutex'. Every other instance
560 * will only be accessed by its own pmd thread. */
561 struct emc_cache flow_cache
;
563 /* Flow-Table and classifiers
565 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
566 * changes to 'classifiers' must be made while still holding the
569 struct ovs_mutex flow_mutex
;
570 struct cmap flow_table OVS_GUARDED
; /* Flow table. */
572 /* One classifier per in_port polled by the pmd */
573 struct cmap classifiers
;
574 /* Periodically sort subtable vectors according to hit frequencies */
575 long long int next_optimization
;
576 /* End of the next time interval for which processing cycles
577 are stored for each polled rxq. */
578 long long int rxq_interval
;
581 struct dp_netdev_pmd_stats stats
;
583 /* Cycles counters */
584 struct dp_netdev_pmd_cycles cycles
;
586 /* Used to count cicles. See 'cycles_counter_end()' */
587 unsigned long long last_cycles
;
589 struct latch exit_latch
; /* For terminating the pmd thread. */
590 struct seq
*reload_seq
;
591 uint64_t last_reload_seq
;
592 atomic_bool reload
; /* Do we need to reload ports? */
594 unsigned core_id
; /* CPU core id of this pmd thread. */
595 int numa_id
; /* numa node id of this pmd thread. */
598 /* Queue id used by this pmd thread to send packets on all netdevs if
599 * XPS disabled for this netdev. All static_tx_qid's are unique and less
600 * than 'cmap_count(dp->poll_threads)'. */
601 uint32_t static_tx_qid
;
603 struct ovs_mutex port_mutex
; /* Mutex for 'poll_list' and 'tx_ports'. */
604 /* List of rx queues to poll. */
605 struct hmap poll_list OVS_GUARDED
;
606 /* Map of 'tx_port's used for transmission. Written by the main thread,
607 * read by the pmd thread. */
608 struct hmap tx_ports OVS_GUARDED
;
610 /* These are thread-local copies of 'tx_ports'. One contains only tunnel
611 * ports (that support push_tunnel/pop_tunnel), the other contains ports
612 * with at least one txq (that support send). A port can be in both.
614 * There are two separate maps to make sure that we don't try to execute
615 * OUTPUT on a device which has 0 txqs or PUSH/POP on a non-tunnel device.
617 * The instances for cpu core NON_PMD_CORE_ID can be accessed by multiple
618 * threads, and thusly need to be protected by 'non_pmd_mutex'. Every
619 * other instance will only be accessed by its own pmd thread. */
620 struct hmap tnl_port_cache
;
621 struct hmap send_port_cache
;
623 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
624 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
625 * values and subtracts them from 'stats' and 'cycles' before
626 * reporting to the user */
627 unsigned long long stats_zero
[DP_N_STATS
];
628 uint64_t cycles_zero
[PMD_N_CYCLES
];
630 /* Set to true if the pmd thread needs to be reloaded. */
634 /* Interface to netdev-based datapath. */
637 struct dp_netdev
*dp
;
638 uint64_t last_port_seq
;
641 static int get_port_by_number(struct dp_netdev
*dp
, odp_port_t port_no
,
642 struct dp_netdev_port
**portp
)
643 OVS_REQUIRES(dp
->port_mutex
);
644 static int get_port_by_name(struct dp_netdev
*dp
, const char *devname
,
645 struct dp_netdev_port
**portp
)
646 OVS_REQUIRES(dp
->port_mutex
);
647 static void dp_netdev_free(struct dp_netdev
*)
648 OVS_REQUIRES(dp_netdev_mutex
);
649 static int do_add_port(struct dp_netdev
*dp
, const char *devname
,
650 const char *type
, odp_port_t port_no
)
651 OVS_REQUIRES(dp
->port_mutex
);
652 static void do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*)
653 OVS_REQUIRES(dp
->port_mutex
);
654 static int dpif_netdev_open(const struct dpif_class
*, const char *name
,
655 bool create
, struct dpif
**);
656 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
657 struct dp_packet_batch
*,
658 bool may_steal
, const struct flow
*flow
,
659 const struct nlattr
*actions
,
662 static void dp_netdev_input(struct dp_netdev_pmd_thread
*,
663 struct dp_packet_batch
*, odp_port_t port_no
);
664 static void dp_netdev_recirculate(struct dp_netdev_pmd_thread
*,
665 struct dp_packet_batch
*);
667 static void dp_netdev_disable_upcall(struct dp_netdev
*);
668 static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
);
669 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
,
670 struct dp_netdev
*dp
, unsigned core_id
,
672 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
);
673 static void dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
674 OVS_REQUIRES(dp
->port_mutex
);
676 static void *pmd_thread_main(void *);
677 static struct dp_netdev_pmd_thread
*dp_netdev_get_pmd(struct dp_netdev
*dp
,
679 static struct dp_netdev_pmd_thread
*
680 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
);
681 static void dp_netdev_del_pmd(struct dp_netdev
*dp
,
682 struct dp_netdev_pmd_thread
*pmd
);
683 static void dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
, bool non_pmd
);
684 static void dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
);
685 static void dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
686 struct dp_netdev_port
*port
)
687 OVS_REQUIRES(pmd
->port_mutex
);
688 static void dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
690 OVS_REQUIRES(pmd
->port_mutex
);
691 static void dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
692 struct dp_netdev_rxq
*rxq
)
693 OVS_REQUIRES(pmd
->port_mutex
);
694 static void dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
695 struct rxq_poll
*poll
)
696 OVS_REQUIRES(pmd
->port_mutex
);
697 static void reconfigure_datapath(struct dp_netdev
*dp
)
698 OVS_REQUIRES(dp
->port_mutex
);
699 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
);
700 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
);
701 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
);
702 static void pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
703 OVS_REQUIRES(pmd
->port_mutex
);
705 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
,
706 struct polled_queue
*poll_list
, int poll_cnt
);
708 dp_netdev_rxq_set_cycles(struct dp_netdev_rxq
*rx
,
709 enum rxq_cycles_counter_type type
,
710 unsigned long long cycles
);
712 dp_netdev_rxq_get_cycles(struct dp_netdev_rxq
*rx
,
713 enum rxq_cycles_counter_type type
);
715 dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq
*rx
,
716 unsigned long long cycles
);
718 dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq
*rx
, unsigned idx
);
720 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
721 long long now
, bool purge
);
722 static int dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
723 struct tx_port
*tx
, long long now
);
725 static inline bool emc_entry_alive(struct emc_entry
*ce
);
726 static void emc_clear_entry(struct emc_entry
*ce
);
728 static void dp_netdev_request_reconfigure(struct dp_netdev
*dp
);
731 emc_cache_init(struct emc_cache
*flow_cache
)
735 flow_cache
->sweep_idx
= 0;
736 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
737 flow_cache
->entries
[i
].flow
= NULL
;
738 flow_cache
->entries
[i
].key
.hash
= 0;
739 flow_cache
->entries
[i
].key
.len
= sizeof(struct miniflow
);
740 flowmap_init(&flow_cache
->entries
[i
].key
.mf
.map
);
745 emc_cache_uninit(struct emc_cache
*flow_cache
)
749 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
750 emc_clear_entry(&flow_cache
->entries
[i
]);
754 /* Check and clear dead flow references slowly (one entry at each
757 emc_cache_slow_sweep(struct emc_cache
*flow_cache
)
759 struct emc_entry
*entry
= &flow_cache
->entries
[flow_cache
->sweep_idx
];
761 if (!emc_entry_alive(entry
)) {
762 emc_clear_entry(entry
);
764 flow_cache
->sweep_idx
= (flow_cache
->sweep_idx
+ 1) & EM_FLOW_HASH_MASK
;
767 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
769 dpif_is_netdev(const struct dpif
*dpif
)
771 return dpif
->dpif_class
->open
== dpif_netdev_open
;
774 static struct dpif_netdev
*
775 dpif_netdev_cast(const struct dpif
*dpif
)
777 ovs_assert(dpif_is_netdev(dpif
));
778 return CONTAINER_OF(dpif
, struct dpif_netdev
, dpif
);
781 static struct dp_netdev
*
782 get_dp_netdev(const struct dpif
*dpif
)
784 return dpif_netdev_cast(dpif
)->dp
;
788 PMD_INFO_SHOW_STATS
, /* Show how cpu cycles are spent. */
789 PMD_INFO_CLEAR_STATS
, /* Set the cycles count to 0. */
790 PMD_INFO_SHOW_RXQ
/* Show poll-lists of pmd threads. */
794 pmd_info_show_stats(struct ds
*reply
,
795 struct dp_netdev_pmd_thread
*pmd
,
796 unsigned long long stats
[DP_N_STATS
],
797 uint64_t cycles
[PMD_N_CYCLES
])
799 unsigned long long total_packets
= 0;
800 uint64_t total_cycles
= 0;
803 /* These loops subtracts reference values ('*_zero') from the counters.
804 * Since loads and stores are relaxed, it might be possible for a '*_zero'
805 * value to be more recent than the current value we're reading from the
806 * counter. This is not a big problem, since these numbers are not
807 * supposed to be too accurate, but we should at least make sure that
808 * the result is not negative. */
809 for (i
= 0; i
< DP_N_STATS
; i
++) {
810 if (stats
[i
] > pmd
->stats_zero
[i
]) {
811 stats
[i
] -= pmd
->stats_zero
[i
];
816 if (i
!= DP_STAT_LOST
) {
817 /* Lost packets are already included in DP_STAT_MISS */
818 total_packets
+= stats
[i
];
822 for (i
= 0; i
< PMD_N_CYCLES
; i
++) {
823 if (cycles
[i
] > pmd
->cycles_zero
[i
]) {
824 cycles
[i
] -= pmd
->cycles_zero
[i
];
829 total_cycles
+= cycles
[i
];
832 ds_put_cstr(reply
, (pmd
->core_id
== NON_PMD_CORE_ID
)
833 ? "main thread" : "pmd thread");
835 if (pmd
->numa_id
!= OVS_NUMA_UNSPEC
) {
836 ds_put_format(reply
, " numa_id %d", pmd
->numa_id
);
838 if (pmd
->core_id
!= OVS_CORE_UNSPEC
&& pmd
->core_id
!= NON_PMD_CORE_ID
) {
839 ds_put_format(reply
, " core_id %u", pmd
->core_id
);
841 ds_put_cstr(reply
, ":\n");
844 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
845 "\tavg. subtable lookups per hit:%.2f\n"
846 "\tmiss:%llu\n\tlost:%llu\n",
847 stats
[DP_STAT_EXACT_HIT
], stats
[DP_STAT_MASKED_HIT
],
848 stats
[DP_STAT_MASKED_HIT
] > 0
849 ? (1.0*stats
[DP_STAT_LOOKUP_HIT
])/stats
[DP_STAT_MASKED_HIT
]
851 stats
[DP_STAT_MISS
], stats
[DP_STAT_LOST
]);
853 if (total_cycles
== 0) {
858 "\tidle cycles:%"PRIu64
" (%.02f%%)\n"
859 "\tprocessing cycles:%"PRIu64
" (%.02f%%)\n",
860 cycles
[PMD_CYCLES_IDLE
],
861 cycles
[PMD_CYCLES_IDLE
] / (double)total_cycles
* 100,
862 cycles
[PMD_CYCLES_PROCESSING
],
863 cycles
[PMD_CYCLES_PROCESSING
] / (double)total_cycles
* 100);
865 if (total_packets
== 0) {
870 "\tavg cycles per packet: %.02f (%"PRIu64
"/%llu)\n",
871 total_cycles
/ (double)total_packets
,
872 total_cycles
, total_packets
);
875 "\tavg processing cycles per packet: "
876 "%.02f (%"PRIu64
"/%llu)\n",
877 cycles
[PMD_CYCLES_PROCESSING
] / (double)total_packets
,
878 cycles
[PMD_CYCLES_PROCESSING
], total_packets
);
882 pmd_info_clear_stats(struct ds
*reply OVS_UNUSED
,
883 struct dp_netdev_pmd_thread
*pmd
,
884 unsigned long long stats
[DP_N_STATS
],
885 uint64_t cycles
[PMD_N_CYCLES
])
889 /* We cannot write 'stats' and 'cycles' (because they're written by other
890 * threads) and we shouldn't change 'stats' (because they're used to count
891 * datapath stats, which must not be cleared here). Instead, we save the
892 * current values and subtract them from the values to be displayed in the
894 for (i
= 0; i
< DP_N_STATS
; i
++) {
895 pmd
->stats_zero
[i
] = stats
[i
];
897 for (i
= 0; i
< PMD_N_CYCLES
; i
++) {
898 pmd
->cycles_zero
[i
] = cycles
[i
];
903 compare_poll_list(const void *a_
, const void *b_
)
905 const struct rxq_poll
*a
= a_
;
906 const struct rxq_poll
*b
= b_
;
908 const char *namea
= netdev_rxq_get_name(a
->rxq
->rx
);
909 const char *nameb
= netdev_rxq_get_name(b
->rxq
->rx
);
911 int cmp
= strcmp(namea
, nameb
);
913 return netdev_rxq_get_queue_id(a
->rxq
->rx
)
914 - netdev_rxq_get_queue_id(b
->rxq
->rx
);
921 sorted_poll_list(struct dp_netdev_pmd_thread
*pmd
, struct rxq_poll
**list
,
924 struct rxq_poll
*ret
, *poll
;
927 *n
= hmap_count(&pmd
->poll_list
);
931 ret
= xcalloc(*n
, sizeof *ret
);
933 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
938 qsort(ret
, *n
, sizeof *ret
, compare_poll_list
);
945 pmd_info_show_rxq(struct ds
*reply
, struct dp_netdev_pmd_thread
*pmd
)
947 if (pmd
->core_id
!= NON_PMD_CORE_ID
) {
948 const char *prev_name
= NULL
;
949 struct rxq_poll
*list
;
953 "pmd thread numa_id %d core_id %u:\n\tisolated : %s\n",
954 pmd
->numa_id
, pmd
->core_id
, (pmd
->isolated
)
957 ovs_mutex_lock(&pmd
->port_mutex
);
958 sorted_poll_list(pmd
, &list
, &n
);
959 for (i
= 0; i
< n
; i
++) {
960 const char *name
= netdev_rxq_get_name(list
[i
].rxq
->rx
);
962 if (!prev_name
|| strcmp(name
, prev_name
)) {
964 ds_put_cstr(reply
, "\n");
966 ds_put_format(reply
, "\tport: %s\tqueue-id:", name
);
968 ds_put_format(reply
, " %d",
969 netdev_rxq_get_queue_id(list
[i
].rxq
->rx
));
972 ovs_mutex_unlock(&pmd
->port_mutex
);
973 ds_put_cstr(reply
, "\n");
979 compare_poll_thread_list(const void *a_
, const void *b_
)
981 const struct dp_netdev_pmd_thread
*a
, *b
;
983 a
= *(struct dp_netdev_pmd_thread
**)a_
;
984 b
= *(struct dp_netdev_pmd_thread
**)b_
;
986 if (a
->core_id
< b
->core_id
) {
989 if (a
->core_id
> b
->core_id
) {
995 /* Create a sorted list of pmd's from the dp->poll_threads cmap. We can use
996 * this list, as long as we do not go to quiescent state. */
998 sorted_poll_thread_list(struct dp_netdev
*dp
,
999 struct dp_netdev_pmd_thread
***list
,
1002 struct dp_netdev_pmd_thread
*pmd
;
1003 struct dp_netdev_pmd_thread
**pmd_list
;
1004 size_t k
= 0, n_pmds
;
1006 n_pmds
= cmap_count(&dp
->poll_threads
);
1007 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
1009 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1013 pmd_list
[k
++] = pmd
;
1016 qsort(pmd_list
, k
, sizeof *pmd_list
, compare_poll_thread_list
);
1023 dpif_netdev_pmd_rebalance(struct unixctl_conn
*conn
, int argc
,
1024 const char *argv
[], void *aux OVS_UNUSED
)
1026 struct ds reply
= DS_EMPTY_INITIALIZER
;
1027 struct dp_netdev
*dp
= NULL
;
1029 ovs_mutex_lock(&dp_netdev_mutex
);
1032 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
1033 } else if (shash_count(&dp_netdevs
) == 1) {
1034 /* There's only one datapath */
1035 dp
= shash_first(&dp_netdevs
)->data
;
1039 ovs_mutex_unlock(&dp_netdev_mutex
);
1040 unixctl_command_reply_error(conn
,
1041 "please specify an existing datapath");
1045 dp_netdev_request_reconfigure(dp
);
1046 ovs_mutex_unlock(&dp_netdev_mutex
);
1047 ds_put_cstr(&reply
, "pmd rxq rebalance requested.\n");
1048 unixctl_command_reply(conn
, ds_cstr(&reply
));
1053 dpif_netdev_pmd_info(struct unixctl_conn
*conn
, int argc
, const char *argv
[],
1056 struct ds reply
= DS_EMPTY_INITIALIZER
;
1057 struct dp_netdev_pmd_thread
**pmd_list
;
1058 struct dp_netdev
*dp
= NULL
;
1060 enum pmd_info_type type
= *(enum pmd_info_type
*) aux
;
1062 ovs_mutex_lock(&dp_netdev_mutex
);
1065 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
1066 } else if (shash_count(&dp_netdevs
) == 1) {
1067 /* There's only one datapath */
1068 dp
= shash_first(&dp_netdevs
)->data
;
1072 ovs_mutex_unlock(&dp_netdev_mutex
);
1073 unixctl_command_reply_error(conn
,
1074 "please specify an existing datapath");
1078 sorted_poll_thread_list(dp
, &pmd_list
, &n
);
1079 for (size_t i
= 0; i
< n
; i
++) {
1080 struct dp_netdev_pmd_thread
*pmd
= pmd_list
[i
];
1085 if (type
== PMD_INFO_SHOW_RXQ
) {
1086 pmd_info_show_rxq(&reply
, pmd
);
1088 unsigned long long stats
[DP_N_STATS
];
1089 uint64_t cycles
[PMD_N_CYCLES
];
1091 /* Read current stats and cycle counters */
1092 for (size_t j
= 0; j
< ARRAY_SIZE(stats
); j
++) {
1093 atomic_read_relaxed(&pmd
->stats
.n
[j
], &stats
[j
]);
1095 for (size_t j
= 0; j
< ARRAY_SIZE(cycles
); j
++) {
1096 atomic_read_relaxed(&pmd
->cycles
.n
[j
], &cycles
[j
]);
1099 if (type
== PMD_INFO_CLEAR_STATS
) {
1100 pmd_info_clear_stats(&reply
, pmd
, stats
, cycles
);
1101 } else if (type
== PMD_INFO_SHOW_STATS
) {
1102 pmd_info_show_stats(&reply
, pmd
, stats
, cycles
);
1108 ovs_mutex_unlock(&dp_netdev_mutex
);
1110 unixctl_command_reply(conn
, ds_cstr(&reply
));
1115 dpif_netdev_init(void)
1117 static enum pmd_info_type show_aux
= PMD_INFO_SHOW_STATS
,
1118 clear_aux
= PMD_INFO_CLEAR_STATS
,
1119 poll_aux
= PMD_INFO_SHOW_RXQ
;
1121 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
1122 0, 1, dpif_netdev_pmd_info
,
1124 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
1125 0, 1, dpif_netdev_pmd_info
,
1126 (void *)&clear_aux
);
1127 unixctl_command_register("dpif-netdev/pmd-rxq-show", "[dp]",
1128 0, 1, dpif_netdev_pmd_info
,
1130 unixctl_command_register("dpif-netdev/pmd-rxq-rebalance", "[dp]",
1131 0, 1, dpif_netdev_pmd_rebalance
,
1137 dpif_netdev_enumerate(struct sset
*all_dps
,
1138 const struct dpif_class
*dpif_class
)
1140 struct shash_node
*node
;
1142 ovs_mutex_lock(&dp_netdev_mutex
);
1143 SHASH_FOR_EACH(node
, &dp_netdevs
) {
1144 struct dp_netdev
*dp
= node
->data
;
1145 if (dpif_class
!= dp
->class) {
1146 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
1147 * If the class doesn't match, skip this dpif. */
1150 sset_add(all_dps
, node
->name
);
1152 ovs_mutex_unlock(&dp_netdev_mutex
);
1158 dpif_netdev_class_is_dummy(const struct dpif_class
*class)
1160 return class != &dpif_netdev_class
;
1164 dpif_netdev_port_open_type(const struct dpif_class
*class, const char *type
)
1166 return strcmp(type
, "internal") ? type
1167 : dpif_netdev_class_is_dummy(class) ? "dummy-internal"
1171 static struct dpif
*
1172 create_dpif_netdev(struct dp_netdev
*dp
)
1174 uint16_t netflow_id
= hash_string(dp
->name
, 0);
1175 struct dpif_netdev
*dpif
;
1177 ovs_refcount_ref(&dp
->ref_cnt
);
1179 dpif
= xmalloc(sizeof *dpif
);
1180 dpif_init(&dpif
->dpif
, dp
->class, dp
->name
, netflow_id
>> 8, netflow_id
);
1182 dpif
->last_port_seq
= seq_read(dp
->port_seq
);
1187 /* Choose an unused, non-zero port number and return it on success.
1188 * Return ODPP_NONE on failure. */
1190 choose_port(struct dp_netdev
*dp
, const char *name
)
1191 OVS_REQUIRES(dp
->port_mutex
)
1195 if (dp
->class != &dpif_netdev_class
) {
1199 /* If the port name begins with "br", start the number search at
1200 * 100 to make writing tests easier. */
1201 if (!strncmp(name
, "br", 2)) {
1205 /* If the port name contains a number, try to assign that port number.
1206 * This can make writing unit tests easier because port numbers are
1208 for (p
= name
; *p
!= '\0'; p
++) {
1209 if (isdigit((unsigned char) *p
)) {
1210 port_no
= start_no
+ strtol(p
, NULL
, 10);
1211 if (port_no
> 0 && port_no
!= odp_to_u32(ODPP_NONE
)
1212 && !dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
1213 return u32_to_odp(port_no
);
1220 for (port_no
= 1; port_no
<= UINT16_MAX
; port_no
++) {
1221 if (!dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
1222 return u32_to_odp(port_no
);
1230 create_dp_netdev(const char *name
, const struct dpif_class
*class,
1231 struct dp_netdev
**dpp
)
1232 OVS_REQUIRES(dp_netdev_mutex
)
1234 struct dp_netdev
*dp
;
1237 dp
= xzalloc(sizeof *dp
);
1238 shash_add(&dp_netdevs
, name
, dp
);
1240 *CONST_CAST(const struct dpif_class
**, &dp
->class) = class;
1241 *CONST_CAST(const char **, &dp
->name
) = xstrdup(name
);
1242 ovs_refcount_init(&dp
->ref_cnt
);
1243 atomic_flag_clear(&dp
->destroyed
);
1245 ovs_mutex_init(&dp
->port_mutex
);
1246 hmap_init(&dp
->ports
);
1247 dp
->port_seq
= seq_create();
1248 fat_rwlock_init(&dp
->upcall_rwlock
);
1250 dp
->reconfigure_seq
= seq_create();
1251 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
1253 for (int i
= 0; i
< N_METER_LOCKS
; ++i
) {
1254 ovs_mutex_init_adaptive(&dp
->meter_locks
[i
]);
1257 /* Disable upcalls by default. */
1258 dp_netdev_disable_upcall(dp
);
1259 dp
->upcall_aux
= NULL
;
1260 dp
->upcall_cb
= NULL
;
1262 conntrack_init(&dp
->conntrack
);
1264 atomic_init(&dp
->emc_insert_min
, DEFAULT_EM_FLOW_INSERT_MIN
);
1266 cmap_init(&dp
->poll_threads
);
1268 ovs_mutex_init(&dp
->tx_qid_pool_mutex
);
1269 /* We need 1 Tx queue for each possible core + 1 for non-PMD threads. */
1270 dp
->tx_qid_pool
= id_pool_create(0, ovs_numa_get_n_cores() + 1);
1272 ovs_mutex_init_recursive(&dp
->non_pmd_mutex
);
1273 ovsthread_key_create(&dp
->per_pmd_key
, NULL
);
1275 ovs_mutex_lock(&dp
->port_mutex
);
1276 /* non-PMD will be created before all other threads and will
1277 * allocate static_tx_qid = 0. */
1278 dp_netdev_set_nonpmd(dp
);
1280 error
= do_add_port(dp
, name
, dpif_netdev_port_open_type(dp
->class,
1283 ovs_mutex_unlock(&dp
->port_mutex
);
1289 dp
->last_tnl_conf_seq
= seq_read(tnl_conf_seq
);
1295 dp_netdev_request_reconfigure(struct dp_netdev
*dp
)
1297 seq_change(dp
->reconfigure_seq
);
1301 dp_netdev_is_reconf_required(struct dp_netdev
*dp
)
1303 return seq_read(dp
->reconfigure_seq
) != dp
->last_reconfigure_seq
;
1307 dpif_netdev_open(const struct dpif_class
*class, const char *name
,
1308 bool create
, struct dpif
**dpifp
)
1310 struct dp_netdev
*dp
;
1313 ovs_mutex_lock(&dp_netdev_mutex
);
1314 dp
= shash_find_data(&dp_netdevs
, name
);
1316 error
= create
? create_dp_netdev(name
, class, &dp
) : ENODEV
;
1318 error
= (dp
->class != class ? EINVAL
1323 *dpifp
= create_dpif_netdev(dp
);
1326 ovs_mutex_unlock(&dp_netdev_mutex
);
1332 dp_netdev_destroy_upcall_lock(struct dp_netdev
*dp
)
1333 OVS_NO_THREAD_SAFETY_ANALYSIS
1335 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
1336 ovs_assert(fat_rwlock_tryrdlock(&dp
->upcall_rwlock
));
1338 /* Before freeing a lock we should release it */
1339 fat_rwlock_unlock(&dp
->upcall_rwlock
);
1340 fat_rwlock_destroy(&dp
->upcall_rwlock
);
1344 dp_delete_meter(struct dp_netdev
*dp
, uint32_t meter_id
)
1345 OVS_REQUIRES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
1347 if (dp
->meters
[meter_id
]) {
1348 free(dp
->meters
[meter_id
]);
1349 dp
->meters
[meter_id
] = NULL
;
1353 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
1354 * through the 'dp_netdevs' shash while freeing 'dp'. */
1356 dp_netdev_free(struct dp_netdev
*dp
)
1357 OVS_REQUIRES(dp_netdev_mutex
)
1359 struct dp_netdev_port
*port
, *next
;
1361 shash_find_and_delete(&dp_netdevs
, dp
->name
);
1363 ovs_mutex_lock(&dp
->port_mutex
);
1364 HMAP_FOR_EACH_SAFE (port
, next
, node
, &dp
->ports
) {
1365 do_del_port(dp
, port
);
1367 ovs_mutex_unlock(&dp
->port_mutex
);
1369 dp_netdev_destroy_all_pmds(dp
, true);
1370 cmap_destroy(&dp
->poll_threads
);
1372 ovs_mutex_destroy(&dp
->tx_qid_pool_mutex
);
1373 id_pool_destroy(dp
->tx_qid_pool
);
1375 ovs_mutex_destroy(&dp
->non_pmd_mutex
);
1376 ovsthread_key_delete(dp
->per_pmd_key
);
1378 conntrack_destroy(&dp
->conntrack
);
1381 seq_destroy(dp
->reconfigure_seq
);
1383 seq_destroy(dp
->port_seq
);
1384 hmap_destroy(&dp
->ports
);
1385 ovs_mutex_destroy(&dp
->port_mutex
);
1387 /* Upcalls must be disabled at this point */
1388 dp_netdev_destroy_upcall_lock(dp
);
1392 for (i
= 0; i
< MAX_METERS
; ++i
) {
1394 dp_delete_meter(dp
, i
);
1395 meter_unlock(dp
, i
);
1397 for (i
= 0; i
< N_METER_LOCKS
; ++i
) {
1398 ovs_mutex_destroy(&dp
->meter_locks
[i
]);
1401 free(dp
->pmd_cmask
);
1402 free(CONST_CAST(char *, dp
->name
));
1407 dp_netdev_unref(struct dp_netdev
*dp
)
1410 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
1411 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
1412 ovs_mutex_lock(&dp_netdev_mutex
);
1413 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1416 ovs_mutex_unlock(&dp_netdev_mutex
);
1421 dpif_netdev_close(struct dpif
*dpif
)
1423 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1425 dp_netdev_unref(dp
);
1430 dpif_netdev_destroy(struct dpif
*dpif
)
1432 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1434 if (!atomic_flag_test_and_set(&dp
->destroyed
)) {
1435 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1436 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
1444 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
1445 * load/store semantics. While the increment is not atomic, the load and
1446 * store operations are, making it impossible to read inconsistent values.
1448 * This is used to update thread local stats counters. */
1450 non_atomic_ullong_add(atomic_ullong
*var
, unsigned long long n
)
1452 unsigned long long tmp
;
1454 atomic_read_relaxed(var
, &tmp
);
1456 atomic_store_relaxed(var
, tmp
);
1460 dpif_netdev_get_stats(const struct dpif
*dpif
, struct dpif_dp_stats
*stats
)
1462 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1463 struct dp_netdev_pmd_thread
*pmd
;
1465 stats
->n_flows
= stats
->n_hit
= stats
->n_missed
= stats
->n_lost
= 0;
1466 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1467 unsigned long long n
;
1468 stats
->n_flows
+= cmap_count(&pmd
->flow_table
);
1470 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_MASKED_HIT
], &n
);
1472 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_EXACT_HIT
], &n
);
1474 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_MISS
], &n
);
1475 stats
->n_missed
+= n
;
1476 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_LOST
], &n
);
1479 stats
->n_masks
= UINT32_MAX
;
1480 stats
->n_mask_hit
= UINT64_MAX
;
1486 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread
*pmd
)
1488 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
1489 ovs_mutex_lock(&pmd
->dp
->non_pmd_mutex
);
1490 ovs_mutex_lock(&pmd
->port_mutex
);
1491 pmd_load_cached_ports(pmd
);
1492 ovs_mutex_unlock(&pmd
->port_mutex
);
1493 ovs_mutex_unlock(&pmd
->dp
->non_pmd_mutex
);
1497 ovs_mutex_lock(&pmd
->cond_mutex
);
1498 seq_change(pmd
->reload_seq
);
1499 atomic_store_relaxed(&pmd
->reload
, true);
1500 ovs_mutex_cond_wait(&pmd
->cond
, &pmd
->cond_mutex
);
1501 ovs_mutex_unlock(&pmd
->cond_mutex
);
1505 hash_port_no(odp_port_t port_no
)
1507 return hash_int(odp_to_u32(port_no
), 0);
1511 port_create(const char *devname
, const char *type
,
1512 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1514 struct netdev_saved_flags
*sf
;
1515 struct dp_netdev_port
*port
;
1516 enum netdev_flags flags
;
1517 struct netdev
*netdev
;
1522 /* Open and validate network device. */
1523 error
= netdev_open(devname
, type
, &netdev
);
1527 /* XXX reject non-Ethernet devices */
1529 netdev_get_flags(netdev
, &flags
);
1530 if (flags
& NETDEV_LOOPBACK
) {
1531 VLOG_ERR("%s: cannot add a loopback device", devname
);
1536 error
= netdev_turn_flags_on(netdev
, NETDEV_PROMISC
, &sf
);
1538 VLOG_ERR("%s: cannot set promisc flag", devname
);
1542 port
= xzalloc(sizeof *port
);
1543 port
->port_no
= port_no
;
1544 port
->netdev
= netdev
;
1545 port
->type
= xstrdup(type
);
1547 port
->need_reconfigure
= true;
1548 ovs_mutex_init(&port
->txq_used_mutex
);
1555 netdev_close(netdev
);
1560 do_add_port(struct dp_netdev
*dp
, const char *devname
, const char *type
,
1562 OVS_REQUIRES(dp
->port_mutex
)
1564 struct dp_netdev_port
*port
;
1567 /* Reject devices already in 'dp'. */
1568 if (!get_port_by_name(dp
, devname
, &port
)) {
1572 error
= port_create(devname
, type
, port_no
, &port
);
1577 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
1578 seq_change(dp
->port_seq
);
1580 reconfigure_datapath(dp
);
1586 dpif_netdev_port_add(struct dpif
*dpif
, struct netdev
*netdev
,
1587 odp_port_t
*port_nop
)
1589 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1590 char namebuf
[NETDEV_VPORT_NAME_BUFSIZE
];
1591 const char *dpif_port
;
1595 ovs_mutex_lock(&dp
->port_mutex
);
1596 dpif_port
= netdev_vport_get_dpif_port(netdev
, namebuf
, sizeof namebuf
);
1597 if (*port_nop
!= ODPP_NONE
) {
1598 port_no
= *port_nop
;
1599 error
= dp_netdev_lookup_port(dp
, *port_nop
) ? EBUSY
: 0;
1601 port_no
= choose_port(dp
, dpif_port
);
1602 error
= port_no
== ODPP_NONE
? EFBIG
: 0;
1605 *port_nop
= port_no
;
1606 error
= do_add_port(dp
, dpif_port
, netdev_get_type(netdev
), port_no
);
1608 ovs_mutex_unlock(&dp
->port_mutex
);
1614 dpif_netdev_port_del(struct dpif
*dpif
, odp_port_t port_no
)
1616 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1619 ovs_mutex_lock(&dp
->port_mutex
);
1620 if (port_no
== ODPP_LOCAL
) {
1623 struct dp_netdev_port
*port
;
1625 error
= get_port_by_number(dp
, port_no
, &port
);
1627 do_del_port(dp
, port
);
1630 ovs_mutex_unlock(&dp
->port_mutex
);
1636 is_valid_port_number(odp_port_t port_no
)
1638 return port_no
!= ODPP_NONE
;
1641 static struct dp_netdev_port
*
1642 dp_netdev_lookup_port(const struct dp_netdev
*dp
, odp_port_t port_no
)
1643 OVS_REQUIRES(dp
->port_mutex
)
1645 struct dp_netdev_port
*port
;
1647 HMAP_FOR_EACH_WITH_HASH (port
, node
, hash_port_no(port_no
), &dp
->ports
) {
1648 if (port
->port_no
== port_no
) {
1656 get_port_by_number(struct dp_netdev
*dp
,
1657 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1658 OVS_REQUIRES(dp
->port_mutex
)
1660 if (!is_valid_port_number(port_no
)) {
1664 *portp
= dp_netdev_lookup_port(dp
, port_no
);
1665 return *portp
? 0 : ENODEV
;
1670 port_destroy(struct dp_netdev_port
*port
)
1676 netdev_close(port
->netdev
);
1677 netdev_restore_flags(port
->sf
);
1679 for (unsigned i
= 0; i
< port
->n_rxq
; i
++) {
1680 netdev_rxq_close(port
->rxqs
[i
].rx
);
1682 ovs_mutex_destroy(&port
->txq_used_mutex
);
1683 free(port
->rxq_affinity_list
);
1684 free(port
->txq_used
);
1691 get_port_by_name(struct dp_netdev
*dp
,
1692 const char *devname
, struct dp_netdev_port
**portp
)
1693 OVS_REQUIRES(dp
->port_mutex
)
1695 struct dp_netdev_port
*port
;
1697 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1698 if (!strcmp(netdev_get_name(port
->netdev
), devname
)) {
1704 /* Callers of dpif_netdev_port_query_by_name() expect ENODEV for a non
1709 /* Returns 'true' if there is a port with pmd netdev. */
1711 has_pmd_port(struct dp_netdev
*dp
)
1712 OVS_REQUIRES(dp
->port_mutex
)
1714 struct dp_netdev_port
*port
;
1716 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1717 if (netdev_is_pmd(port
->netdev
)) {
1726 do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*port
)
1727 OVS_REQUIRES(dp
->port_mutex
)
1729 hmap_remove(&dp
->ports
, &port
->node
);
1730 seq_change(dp
->port_seq
);
1732 reconfigure_datapath(dp
);
1738 answer_port_query(const struct dp_netdev_port
*port
,
1739 struct dpif_port
*dpif_port
)
1741 dpif_port
->name
= xstrdup(netdev_get_name(port
->netdev
));
1742 dpif_port
->type
= xstrdup(port
->type
);
1743 dpif_port
->port_no
= port
->port_no
;
1747 dpif_netdev_port_query_by_number(const struct dpif
*dpif
, odp_port_t port_no
,
1748 struct dpif_port
*dpif_port
)
1750 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1751 struct dp_netdev_port
*port
;
1754 ovs_mutex_lock(&dp
->port_mutex
);
1755 error
= get_port_by_number(dp
, port_no
, &port
);
1756 if (!error
&& dpif_port
) {
1757 answer_port_query(port
, dpif_port
);
1759 ovs_mutex_unlock(&dp
->port_mutex
);
1765 dpif_netdev_port_query_by_name(const struct dpif
*dpif
, const char *devname
,
1766 struct dpif_port
*dpif_port
)
1768 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1769 struct dp_netdev_port
*port
;
1772 ovs_mutex_lock(&dp
->port_mutex
);
1773 error
= get_port_by_name(dp
, devname
, &port
);
1774 if (!error
&& dpif_port
) {
1775 answer_port_query(port
, dpif_port
);
1777 ovs_mutex_unlock(&dp
->port_mutex
);
1783 dp_netdev_flow_free(struct dp_netdev_flow
*flow
)
1785 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow
));
1789 static void dp_netdev_flow_unref(struct dp_netdev_flow
*flow
)
1791 if (ovs_refcount_unref_relaxed(&flow
->ref_cnt
) == 1) {
1792 ovsrcu_postpone(dp_netdev_flow_free
, flow
);
1797 dp_netdev_flow_hash(const ovs_u128
*ufid
)
1799 return ufid
->u32
[0];
1802 static inline struct dpcls
*
1803 dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread
*pmd
,
1807 uint32_t hash
= hash_port_no(in_port
);
1808 CMAP_FOR_EACH_WITH_HASH (cls
, node
, hash
, &pmd
->classifiers
) {
1809 if (cls
->in_port
== in_port
) {
1810 /* Port classifier exists already */
1817 static inline struct dpcls
*
1818 dp_netdev_pmd_find_dpcls(struct dp_netdev_pmd_thread
*pmd
,
1820 OVS_REQUIRES(pmd
->flow_mutex
)
1822 struct dpcls
*cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
1823 uint32_t hash
= hash_port_no(in_port
);
1826 /* Create new classifier for in_port */
1827 cls
= xmalloc(sizeof(*cls
));
1829 cls
->in_port
= in_port
;
1830 cmap_insert(&pmd
->classifiers
, &cls
->node
, hash
);
1831 VLOG_DBG("Creating dpcls %p for in_port %d", cls
, in_port
);
1837 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread
*pmd
,
1838 struct dp_netdev_flow
*flow
)
1839 OVS_REQUIRES(pmd
->flow_mutex
)
1841 struct cmap_node
*node
= CONST_CAST(struct cmap_node
*, &flow
->node
);
1843 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
1845 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
1846 ovs_assert(cls
!= NULL
);
1847 dpcls_remove(cls
, &flow
->cr
);
1848 cmap_remove(&pmd
->flow_table
, node
, dp_netdev_flow_hash(&flow
->ufid
));
1851 dp_netdev_flow_unref(flow
);
1855 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
)
1857 struct dp_netdev_flow
*netdev_flow
;
1859 ovs_mutex_lock(&pmd
->flow_mutex
);
1860 CMAP_FOR_EACH (netdev_flow
, node
, &pmd
->flow_table
) {
1861 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
1863 ovs_mutex_unlock(&pmd
->flow_mutex
);
1867 dpif_netdev_flow_flush(struct dpif
*dpif
)
1869 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1870 struct dp_netdev_pmd_thread
*pmd
;
1872 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1873 dp_netdev_pmd_flow_flush(pmd
);
1879 struct dp_netdev_port_state
{
1880 struct hmap_position position
;
1885 dpif_netdev_port_dump_start(const struct dpif
*dpif OVS_UNUSED
, void **statep
)
1887 *statep
= xzalloc(sizeof(struct dp_netdev_port_state
));
1892 dpif_netdev_port_dump_next(const struct dpif
*dpif
, void *state_
,
1893 struct dpif_port
*dpif_port
)
1895 struct dp_netdev_port_state
*state
= state_
;
1896 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1897 struct hmap_node
*node
;
1900 ovs_mutex_lock(&dp
->port_mutex
);
1901 node
= hmap_at_position(&dp
->ports
, &state
->position
);
1903 struct dp_netdev_port
*port
;
1905 port
= CONTAINER_OF(node
, struct dp_netdev_port
, node
);
1908 state
->name
= xstrdup(netdev_get_name(port
->netdev
));
1909 dpif_port
->name
= state
->name
;
1910 dpif_port
->type
= port
->type
;
1911 dpif_port
->port_no
= port
->port_no
;
1917 ovs_mutex_unlock(&dp
->port_mutex
);
1923 dpif_netdev_port_dump_done(const struct dpif
*dpif OVS_UNUSED
, void *state_
)
1925 struct dp_netdev_port_state
*state
= state_
;
1932 dpif_netdev_port_poll(const struct dpif
*dpif_
, char **devnamep OVS_UNUSED
)
1934 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
1935 uint64_t new_port_seq
;
1938 new_port_seq
= seq_read(dpif
->dp
->port_seq
);
1939 if (dpif
->last_port_seq
!= new_port_seq
) {
1940 dpif
->last_port_seq
= new_port_seq
;
1950 dpif_netdev_port_poll_wait(const struct dpif
*dpif_
)
1952 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
1954 seq_wait(dpif
->dp
->port_seq
, dpif
->last_port_seq
);
1957 static struct dp_netdev_flow
*
1958 dp_netdev_flow_cast(const struct dpcls_rule
*cr
)
1960 return cr
? CONTAINER_OF(cr
, struct dp_netdev_flow
, cr
) : NULL
;
1963 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*flow
)
1965 return ovs_refcount_try_ref_rcu(&flow
->ref_cnt
);
1968 /* netdev_flow_key utilities.
1970 * netdev_flow_key is basically a miniflow. We use these functions
1971 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1972 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1974 * - Since we are dealing exclusively with miniflows created by
1975 * miniflow_extract(), if the map is different the miniflow is different.
1976 * Therefore we can be faster by comparing the map and the miniflow in a
1978 * - These functions can be inlined by the compiler. */
1980 /* Given the number of bits set in miniflow's maps, returns the size of the
1981 * 'netdev_flow_key.mf' */
1982 static inline size_t
1983 netdev_flow_key_size(size_t flow_u64s
)
1985 return sizeof(struct miniflow
) + MINIFLOW_VALUES_SIZE(flow_u64s
);
1989 netdev_flow_key_equal(const struct netdev_flow_key
*a
,
1990 const struct netdev_flow_key
*b
)
1992 /* 'b->len' may be not set yet. */
1993 return a
->hash
== b
->hash
&& !memcmp(&a
->mf
, &b
->mf
, a
->len
);
1996 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1997 * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been
1998 * generated by miniflow_extract. */
2000 netdev_flow_key_equal_mf(const struct netdev_flow_key
*key
,
2001 const struct miniflow
*mf
)
2003 return !memcmp(&key
->mf
, mf
, key
->len
);
2007 netdev_flow_key_clone(struct netdev_flow_key
*dst
,
2008 const struct netdev_flow_key
*src
)
2011 offsetof(struct netdev_flow_key
, mf
) + src
->len
);
2014 /* Initialize a netdev_flow_key 'mask' from 'match'. */
2016 netdev_flow_mask_init(struct netdev_flow_key
*mask
,
2017 const struct match
*match
)
2019 uint64_t *dst
= miniflow_values(&mask
->mf
);
2020 struct flowmap fmap
;
2024 /* Only check masks that make sense for the flow. */
2025 flow_wc_map(&match
->flow
, &fmap
);
2026 flowmap_init(&mask
->mf
.map
);
2028 FLOWMAP_FOR_EACH_INDEX(idx
, fmap
) {
2029 uint64_t mask_u64
= flow_u64_value(&match
->wc
.masks
, idx
);
2032 flowmap_set(&mask
->mf
.map
, idx
, 1);
2034 hash
= hash_add64(hash
, mask_u64
);
2040 FLOWMAP_FOR_EACH_MAP (map
, mask
->mf
.map
) {
2041 hash
= hash_add64(hash
, map
);
2044 size_t n
= dst
- miniflow_get_values(&mask
->mf
);
2046 mask
->hash
= hash_finish(hash
, n
* 8);
2047 mask
->len
= netdev_flow_key_size(n
);
2050 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
2052 netdev_flow_key_init_masked(struct netdev_flow_key
*dst
,
2053 const struct flow
*flow
,
2054 const struct netdev_flow_key
*mask
)
2056 uint64_t *dst_u64
= miniflow_values(&dst
->mf
);
2057 const uint64_t *mask_u64
= miniflow_get_values(&mask
->mf
);
2061 dst
->len
= mask
->len
;
2062 dst
->mf
= mask
->mf
; /* Copy maps. */
2064 FLOW_FOR_EACH_IN_MAPS(value
, flow
, mask
->mf
.map
) {
2065 *dst_u64
= value
& *mask_u64
++;
2066 hash
= hash_add64(hash
, *dst_u64
++);
2068 dst
->hash
= hash_finish(hash
,
2069 (dst_u64
- miniflow_get_values(&dst
->mf
)) * 8);
2072 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
2073 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
2074 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
2076 /* Returns a hash value for the bits of 'key' where there are 1-bits in
2078 static inline uint32_t
2079 netdev_flow_key_hash_in_mask(const struct netdev_flow_key
*key
,
2080 const struct netdev_flow_key
*mask
)
2082 const uint64_t *p
= miniflow_get_values(&mask
->mf
);
2086 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, key
, mask
->mf
.map
) {
2087 hash
= hash_add64(hash
, value
& *p
++);
2090 return hash_finish(hash
, (p
- miniflow_get_values(&mask
->mf
)) * 8);
2094 emc_entry_alive(struct emc_entry
*ce
)
2096 return ce
->flow
&& !ce
->flow
->dead
;
2100 emc_clear_entry(struct emc_entry
*ce
)
2103 dp_netdev_flow_unref(ce
->flow
);
2109 emc_change_entry(struct emc_entry
*ce
, struct dp_netdev_flow
*flow
,
2110 const struct netdev_flow_key
*key
)
2112 if (ce
->flow
!= flow
) {
2114 dp_netdev_flow_unref(ce
->flow
);
2117 if (dp_netdev_flow_ref(flow
)) {
2124 netdev_flow_key_clone(&ce
->key
, key
);
2129 emc_insert(struct emc_cache
*cache
, const struct netdev_flow_key
*key
,
2130 struct dp_netdev_flow
*flow
)
2132 struct emc_entry
*to_be_replaced
= NULL
;
2133 struct emc_entry
*current_entry
;
2135 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2136 if (netdev_flow_key_equal(¤t_entry
->key
, key
)) {
2137 /* We found the entry with the 'mf' miniflow */
2138 emc_change_entry(current_entry
, flow
, NULL
);
2142 /* Replacement policy: put the flow in an empty (not alive) entry, or
2143 * in the first entry where it can be */
2145 || (emc_entry_alive(to_be_replaced
)
2146 && !emc_entry_alive(current_entry
))
2147 || current_entry
->key
.hash
< to_be_replaced
->key
.hash
) {
2148 to_be_replaced
= current_entry
;
2151 /* We didn't find the miniflow in the cache.
2152 * The 'to_be_replaced' entry is where the new flow will be stored */
2154 emc_change_entry(to_be_replaced
, flow
, key
);
2158 emc_probabilistic_insert(struct dp_netdev_pmd_thread
*pmd
,
2159 const struct netdev_flow_key
*key
,
2160 struct dp_netdev_flow
*flow
)
2162 /* Insert an entry into the EMC based on probability value 'min'. By
2163 * default the value is UINT32_MAX / 100 which yields an insertion
2164 * probability of 1/100 ie. 1% */
2167 atomic_read_relaxed(&pmd
->dp
->emc_insert_min
, &min
);
2169 if (min
&& random_uint32() <= min
) {
2170 emc_insert(&pmd
->flow_cache
, key
, flow
);
2174 static inline struct dp_netdev_flow
*
2175 emc_lookup(struct emc_cache
*cache
, const struct netdev_flow_key
*key
)
2177 struct emc_entry
*current_entry
;
2179 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2180 if (current_entry
->key
.hash
== key
->hash
2181 && emc_entry_alive(current_entry
)
2182 && netdev_flow_key_equal_mf(¤t_entry
->key
, &key
->mf
)) {
2184 /* We found the entry with the 'key->mf' miniflow */
2185 return current_entry
->flow
;
2192 static struct dp_netdev_flow
*
2193 dp_netdev_pmd_lookup_flow(struct dp_netdev_pmd_thread
*pmd
,
2194 const struct netdev_flow_key
*key
,
2198 struct dpcls_rule
*rule
;
2199 odp_port_t in_port
= u32_to_odp(MINIFLOW_GET_U32(&key
->mf
, in_port
));
2200 struct dp_netdev_flow
*netdev_flow
= NULL
;
2202 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2203 if (OVS_LIKELY(cls
)) {
2204 dpcls_lookup(cls
, key
, &rule
, 1, lookup_num_p
);
2205 netdev_flow
= dp_netdev_flow_cast(rule
);
2210 static struct dp_netdev_flow
*
2211 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread
*pmd
,
2212 const ovs_u128
*ufidp
, const struct nlattr
*key
,
2215 struct dp_netdev_flow
*netdev_flow
;
2219 /* If a UFID is not provided, determine one based on the key. */
2220 if (!ufidp
&& key
&& key_len
2221 && !dpif_netdev_flow_from_nlattrs(key
, key_len
, &flow
, false)) {
2222 dpif_flow_hash(pmd
->dp
->dpif
, &flow
, sizeof flow
, &ufid
);
2227 CMAP_FOR_EACH_WITH_HASH (netdev_flow
, node
, dp_netdev_flow_hash(ufidp
),
2229 if (ovs_u128_equals(netdev_flow
->ufid
, *ufidp
)) {
2239 get_dpif_flow_stats(const struct dp_netdev_flow
*netdev_flow_
,
2240 struct dpif_flow_stats
*stats
)
2242 struct dp_netdev_flow
*netdev_flow
;
2243 unsigned long long n
;
2247 netdev_flow
= CONST_CAST(struct dp_netdev_flow
*, netdev_flow_
);
2249 atomic_read_relaxed(&netdev_flow
->stats
.packet_count
, &n
);
2250 stats
->n_packets
= n
;
2251 atomic_read_relaxed(&netdev_flow
->stats
.byte_count
, &n
);
2253 atomic_read_relaxed(&netdev_flow
->stats
.used
, &used
);
2255 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
2256 stats
->tcp_flags
= flags
;
2259 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
2260 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
2261 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
2264 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow
*netdev_flow
,
2265 struct ofpbuf
*key_buf
, struct ofpbuf
*mask_buf
,
2266 struct dpif_flow
*flow
, bool terse
)
2269 memset(flow
, 0, sizeof *flow
);
2271 struct flow_wildcards wc
;
2272 struct dp_netdev_actions
*actions
;
2274 struct odp_flow_key_parms odp_parms
= {
2275 .flow
= &netdev_flow
->flow
,
2277 .support
= dp_netdev_support
,
2280 miniflow_expand(&netdev_flow
->cr
.mask
->mf
, &wc
.masks
);
2281 /* in_port is exact matched, but we have left it out from the mask for
2282 * optimnization reasons. Add in_port back to the mask. */
2283 wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
2286 offset
= key_buf
->size
;
2287 flow
->key
= ofpbuf_tail(key_buf
);
2288 odp_flow_key_from_flow(&odp_parms
, key_buf
);
2289 flow
->key_len
= key_buf
->size
- offset
;
2292 offset
= mask_buf
->size
;
2293 flow
->mask
= ofpbuf_tail(mask_buf
);
2294 odp_parms
.key_buf
= key_buf
;
2295 odp_flow_key_from_mask(&odp_parms
, mask_buf
);
2296 flow
->mask_len
= mask_buf
->size
- offset
;
2299 actions
= dp_netdev_flow_get_actions(netdev_flow
);
2300 flow
->actions
= actions
->actions
;
2301 flow
->actions_len
= actions
->size
;
2304 flow
->ufid
= netdev_flow
->ufid
;
2305 flow
->ufid_present
= true;
2306 flow
->pmd_id
= netdev_flow
->pmd_id
;
2307 get_dpif_flow_stats(netdev_flow
, &flow
->stats
);
2311 dpif_netdev_mask_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
2312 const struct nlattr
*mask_key
,
2313 uint32_t mask_key_len
, const struct flow
*flow
,
2314 struct flow_wildcards
*wc
, bool probe
)
2316 enum odp_key_fitness fitness
;
2318 fitness
= odp_flow_key_to_mask(mask_key
, mask_key_len
, wc
, flow
);
2321 /* This should not happen: it indicates that
2322 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
2323 * disagree on the acceptable form of a mask. Log the problem
2324 * as an error, with enough details to enable debugging. */
2325 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
2327 if (!VLOG_DROP_ERR(&rl
)) {
2331 odp_flow_format(key
, key_len
, mask_key
, mask_key_len
, NULL
, &s
,
2333 VLOG_ERR("internal error parsing flow mask %s (%s)",
2334 ds_cstr(&s
), odp_key_fitness_to_string(fitness
));
2346 dpif_netdev_flow_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
2347 struct flow
*flow
, bool probe
)
2349 if (odp_flow_key_to_flow(key
, key_len
, flow
)) {
2351 /* This should not happen: it indicates that
2352 * odp_flow_key_from_flow() and odp_flow_key_to_flow() disagree on
2353 * the acceptable form of a flow. Log the problem as an error,
2354 * with enough details to enable debugging. */
2355 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
2357 if (!VLOG_DROP_ERR(&rl
)) {
2361 odp_flow_format(key
, key_len
, NULL
, 0, NULL
, &s
, true);
2362 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s
));
2370 if (flow
->ct_state
& DP_NETDEV_CS_UNSUPPORTED_MASK
) {
2378 dpif_netdev_flow_get(const struct dpif
*dpif
, const struct dpif_flow_get
*get
)
2380 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2381 struct dp_netdev_flow
*netdev_flow
;
2382 struct dp_netdev_pmd_thread
*pmd
;
2383 struct hmapx to_find
= HMAPX_INITIALIZER(&to_find
);
2384 struct hmapx_node
*node
;
2387 if (get
->pmd_id
== PMD_ID_NULL
) {
2388 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2389 if (dp_netdev_pmd_try_ref(pmd
) && !hmapx_add(&to_find
, pmd
)) {
2390 dp_netdev_pmd_unref(pmd
);
2394 pmd
= dp_netdev_get_pmd(dp
, get
->pmd_id
);
2398 hmapx_add(&to_find
, pmd
);
2401 if (!hmapx_count(&to_find
)) {
2405 HMAPX_FOR_EACH (node
, &to_find
) {
2406 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
2407 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, get
->ufid
, get
->key
,
2410 dp_netdev_flow_to_dpif_flow(netdev_flow
, get
->buffer
, get
->buffer
,
2419 HMAPX_FOR_EACH (node
, &to_find
) {
2420 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
2421 dp_netdev_pmd_unref(pmd
);
2424 hmapx_destroy(&to_find
);
2428 static struct dp_netdev_flow
*
2429 dp_netdev_flow_add(struct dp_netdev_pmd_thread
*pmd
,
2430 struct match
*match
, const ovs_u128
*ufid
,
2431 const struct nlattr
*actions
, size_t actions_len
)
2432 OVS_REQUIRES(pmd
->flow_mutex
)
2434 struct dp_netdev_flow
*flow
;
2435 struct netdev_flow_key mask
;
2438 /* Make sure in_port is exact matched before we read it. */
2439 ovs_assert(match
->wc
.masks
.in_port
.odp_port
== ODPP_NONE
);
2440 odp_port_t in_port
= match
->flow
.in_port
.odp_port
;
2442 /* As we select the dpcls based on the port number, each netdev flow
2443 * belonging to the same dpcls will have the same odp_port value.
2444 * For performance reasons we wildcard odp_port here in the mask. In the
2445 * typical case dp_hash is also wildcarded, and the resulting 8-byte
2446 * chunk {dp_hash, in_port} will be ignored by netdev_flow_mask_init() and
2447 * will not be part of the subtable mask.
2448 * This will speed up the hash computation during dpcls_lookup() because
2449 * there is one less call to hash_add64() in this case. */
2450 match
->wc
.masks
.in_port
.odp_port
= 0;
2451 netdev_flow_mask_init(&mask
, match
);
2452 match
->wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
2454 /* Make sure wc does not have metadata. */
2455 ovs_assert(!FLOWMAP_HAS_FIELD(&mask
.mf
.map
, metadata
)
2456 && !FLOWMAP_HAS_FIELD(&mask
.mf
.map
, regs
));
2458 /* Do not allocate extra space. */
2459 flow
= xmalloc(sizeof *flow
- sizeof flow
->cr
.flow
.mf
+ mask
.len
);
2460 memset(&flow
->stats
, 0, sizeof flow
->stats
);
2463 *CONST_CAST(unsigned *, &flow
->pmd_id
) = pmd
->core_id
;
2464 *CONST_CAST(struct flow
*, &flow
->flow
) = match
->flow
;
2465 *CONST_CAST(ovs_u128
*, &flow
->ufid
) = *ufid
;
2466 ovs_refcount_init(&flow
->ref_cnt
);
2467 ovsrcu_set(&flow
->actions
, dp_netdev_actions_create(actions
, actions_len
));
2469 netdev_flow_key_init_masked(&flow
->cr
.flow
, &match
->flow
, &mask
);
2471 /* Select dpcls for in_port. Relies on in_port to be exact match. */
2472 cls
= dp_netdev_pmd_find_dpcls(pmd
, in_port
);
2473 dpcls_insert(cls
, &flow
->cr
, &mask
);
2475 cmap_insert(&pmd
->flow_table
, CONST_CAST(struct cmap_node
*, &flow
->node
),
2476 dp_netdev_flow_hash(&flow
->ufid
));
2478 if (OVS_UNLIKELY(!VLOG_DROP_DBG((&upcall_rl
)))) {
2479 struct ds ds
= DS_EMPTY_INITIALIZER
;
2480 struct ofpbuf key_buf
, mask_buf
;
2481 struct odp_flow_key_parms odp_parms
= {
2482 .flow
= &match
->flow
,
2483 .mask
= &match
->wc
.masks
,
2484 .support
= dp_netdev_support
,
2487 ofpbuf_init(&key_buf
, 0);
2488 ofpbuf_init(&mask_buf
, 0);
2490 odp_flow_key_from_flow(&odp_parms
, &key_buf
);
2491 odp_parms
.key_buf
= &key_buf
;
2492 odp_flow_key_from_mask(&odp_parms
, &mask_buf
);
2494 ds_put_cstr(&ds
, "flow_add: ");
2495 odp_format_ufid(ufid
, &ds
);
2496 ds_put_cstr(&ds
, " ");
2497 odp_flow_format(key_buf
.data
, key_buf
.size
,
2498 mask_buf
.data
, mask_buf
.size
,
2500 ds_put_cstr(&ds
, ", actions:");
2501 format_odp_actions(&ds
, actions
, actions_len
, NULL
);
2503 VLOG_DBG("%s", ds_cstr(&ds
));
2505 ofpbuf_uninit(&key_buf
);
2506 ofpbuf_uninit(&mask_buf
);
2508 /* Add a printout of the actual match installed. */
2511 ds_put_cstr(&ds
, "flow match: ");
2512 miniflow_expand(&flow
->cr
.flow
.mf
, &m
.flow
);
2513 miniflow_expand(&flow
->cr
.mask
->mf
, &m
.wc
.masks
);
2514 memset(&m
.tun_md
, 0, sizeof m
.tun_md
);
2515 match_format(&m
, NULL
, &ds
, OFP_DEFAULT_PRIORITY
);
2517 VLOG_DBG("%s", ds_cstr(&ds
));
2526 flow_put_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
2527 struct netdev_flow_key
*key
,
2528 struct match
*match
,
2530 const struct dpif_flow_put
*put
,
2531 struct dpif_flow_stats
*stats
)
2533 struct dp_netdev_flow
*netdev_flow
;
2537 memset(stats
, 0, sizeof *stats
);
2540 ovs_mutex_lock(&pmd
->flow_mutex
);
2541 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
2543 if (put
->flags
& DPIF_FP_CREATE
) {
2544 if (cmap_count(&pmd
->flow_table
) < MAX_FLOWS
) {
2545 dp_netdev_flow_add(pmd
, match
, ufid
, put
->actions
,
2555 if (put
->flags
& DPIF_FP_MODIFY
) {
2556 struct dp_netdev_actions
*new_actions
;
2557 struct dp_netdev_actions
*old_actions
;
2559 new_actions
= dp_netdev_actions_create(put
->actions
,
2562 old_actions
= dp_netdev_flow_get_actions(netdev_flow
);
2563 ovsrcu_set(&netdev_flow
->actions
, new_actions
);
2566 get_dpif_flow_stats(netdev_flow
, stats
);
2568 if (put
->flags
& DPIF_FP_ZERO_STATS
) {
2569 /* XXX: The userspace datapath uses thread local statistics
2570 * (for flows), which should be updated only by the owning
2571 * thread. Since we cannot write on stats memory here,
2572 * we choose not to support this flag. Please note:
2573 * - This feature is currently used only by dpctl commands with
2575 * - Should the need arise, this operation can be implemented
2576 * by keeping a base value (to be update here) for each
2577 * counter, and subtracting it before outputting the stats */
2581 ovsrcu_postpone(dp_netdev_actions_free
, old_actions
);
2582 } else if (put
->flags
& DPIF_FP_CREATE
) {
2585 /* Overlapping flow. */
2589 ovs_mutex_unlock(&pmd
->flow_mutex
);
2594 dpif_netdev_flow_put(struct dpif
*dpif
, const struct dpif_flow_put
*put
)
2596 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2597 struct netdev_flow_key key
, mask
;
2598 struct dp_netdev_pmd_thread
*pmd
;
2602 bool probe
= put
->flags
& DPIF_FP_PROBE
;
2605 memset(put
->stats
, 0, sizeof *put
->stats
);
2607 error
= dpif_netdev_flow_from_nlattrs(put
->key
, put
->key_len
, &match
.flow
,
2612 error
= dpif_netdev_mask_from_nlattrs(put
->key
, put
->key_len
,
2613 put
->mask
, put
->mask_len
,
2614 &match
.flow
, &match
.wc
, probe
);
2622 dpif_flow_hash(dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
2625 /* Must produce a netdev_flow_key for lookup.
2626 * Use the same method as employed to create the key when adding
2627 * the flow to the dplcs to make sure they match. */
2628 netdev_flow_mask_init(&mask
, &match
);
2629 netdev_flow_key_init_masked(&key
, &match
.flow
, &mask
);
2631 if (put
->pmd_id
== PMD_ID_NULL
) {
2632 if (cmap_count(&dp
->poll_threads
) == 0) {
2635 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2636 struct dpif_flow_stats pmd_stats
;
2639 pmd_error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
,
2643 } else if (put
->stats
) {
2644 put
->stats
->n_packets
+= pmd_stats
.n_packets
;
2645 put
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
2646 put
->stats
->used
= MAX(put
->stats
->used
, pmd_stats
.used
);
2647 put
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
2651 pmd
= dp_netdev_get_pmd(dp
, put
->pmd_id
);
2655 error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
, put
->stats
);
2656 dp_netdev_pmd_unref(pmd
);
2663 flow_del_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
2664 struct dpif_flow_stats
*stats
,
2665 const struct dpif_flow_del
*del
)
2667 struct dp_netdev_flow
*netdev_flow
;
2670 ovs_mutex_lock(&pmd
->flow_mutex
);
2671 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, del
->ufid
, del
->key
,
2675 get_dpif_flow_stats(netdev_flow
, stats
);
2677 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
2681 ovs_mutex_unlock(&pmd
->flow_mutex
);
2687 dpif_netdev_flow_del(struct dpif
*dpif
, const struct dpif_flow_del
*del
)
2689 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2690 struct dp_netdev_pmd_thread
*pmd
;
2694 memset(del
->stats
, 0, sizeof *del
->stats
);
2697 if (del
->pmd_id
== PMD_ID_NULL
) {
2698 if (cmap_count(&dp
->poll_threads
) == 0) {
2701 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2702 struct dpif_flow_stats pmd_stats
;
2705 pmd_error
= flow_del_on_pmd(pmd
, &pmd_stats
, del
);
2708 } else if (del
->stats
) {
2709 del
->stats
->n_packets
+= pmd_stats
.n_packets
;
2710 del
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
2711 del
->stats
->used
= MAX(del
->stats
->used
, pmd_stats
.used
);
2712 del
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
2716 pmd
= dp_netdev_get_pmd(dp
, del
->pmd_id
);
2720 error
= flow_del_on_pmd(pmd
, del
->stats
, del
);
2721 dp_netdev_pmd_unref(pmd
);
2728 struct dpif_netdev_flow_dump
{
2729 struct dpif_flow_dump up
;
2730 struct cmap_position poll_thread_pos
;
2731 struct cmap_position flow_pos
;
2732 struct dp_netdev_pmd_thread
*cur_pmd
;
2734 struct ovs_mutex mutex
;
2737 static struct dpif_netdev_flow_dump
*
2738 dpif_netdev_flow_dump_cast(struct dpif_flow_dump
*dump
)
2740 return CONTAINER_OF(dump
, struct dpif_netdev_flow_dump
, up
);
2743 static struct dpif_flow_dump
*
2744 dpif_netdev_flow_dump_create(const struct dpif
*dpif_
, bool terse
,
2745 char *type OVS_UNUSED
)
2747 struct dpif_netdev_flow_dump
*dump
;
2749 dump
= xzalloc(sizeof *dump
);
2750 dpif_flow_dump_init(&dump
->up
, dpif_
);
2751 dump
->up
.terse
= terse
;
2752 ovs_mutex_init(&dump
->mutex
);
2758 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump
*dump_
)
2760 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
2762 ovs_mutex_destroy(&dump
->mutex
);
2767 struct dpif_netdev_flow_dump_thread
{
2768 struct dpif_flow_dump_thread up
;
2769 struct dpif_netdev_flow_dump
*dump
;
2770 struct odputil_keybuf keybuf
[FLOW_DUMP_MAX_BATCH
];
2771 struct odputil_keybuf maskbuf
[FLOW_DUMP_MAX_BATCH
];
2774 static struct dpif_netdev_flow_dump_thread
*
2775 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread
*thread
)
2777 return CONTAINER_OF(thread
, struct dpif_netdev_flow_dump_thread
, up
);
2780 static struct dpif_flow_dump_thread
*
2781 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump
*dump_
)
2783 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
2784 struct dpif_netdev_flow_dump_thread
*thread
;
2786 thread
= xmalloc(sizeof *thread
);
2787 dpif_flow_dump_thread_init(&thread
->up
, &dump
->up
);
2788 thread
->dump
= dump
;
2793 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread
*thread_
)
2795 struct dpif_netdev_flow_dump_thread
*thread
2796 = dpif_netdev_flow_dump_thread_cast(thread_
);
2802 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread
*thread_
,
2803 struct dpif_flow
*flows
, int max_flows
)
2805 struct dpif_netdev_flow_dump_thread
*thread
2806 = dpif_netdev_flow_dump_thread_cast(thread_
);
2807 struct dpif_netdev_flow_dump
*dump
= thread
->dump
;
2808 struct dp_netdev_flow
*netdev_flows
[FLOW_DUMP_MAX_BATCH
];
2812 ovs_mutex_lock(&dump
->mutex
);
2813 if (!dump
->status
) {
2814 struct dpif_netdev
*dpif
= dpif_netdev_cast(thread
->up
.dpif
);
2815 struct dp_netdev
*dp
= get_dp_netdev(&dpif
->dpif
);
2816 struct dp_netdev_pmd_thread
*pmd
= dump
->cur_pmd
;
2817 int flow_limit
= MIN(max_flows
, FLOW_DUMP_MAX_BATCH
);
2819 /* First call to dump_next(), extracts the first pmd thread.
2820 * If there is no pmd thread, returns immediately. */
2822 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
2824 ovs_mutex_unlock(&dump
->mutex
);
2831 for (n_flows
= 0; n_flows
< flow_limit
; n_flows
++) {
2832 struct cmap_node
*node
;
2834 node
= cmap_next_position(&pmd
->flow_table
, &dump
->flow_pos
);
2838 netdev_flows
[n_flows
] = CONTAINER_OF(node
,
2839 struct dp_netdev_flow
,
2842 /* When finishing dumping the current pmd thread, moves to
2844 if (n_flows
< flow_limit
) {
2845 memset(&dump
->flow_pos
, 0, sizeof dump
->flow_pos
);
2846 dp_netdev_pmd_unref(pmd
);
2847 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
2853 /* Keeps the reference to next caller. */
2854 dump
->cur_pmd
= pmd
;
2856 /* If the current dump is empty, do not exit the loop, since the
2857 * remaining pmds could have flows to be dumped. Just dumps again
2858 * on the new 'pmd'. */
2861 ovs_mutex_unlock(&dump
->mutex
);
2863 for (i
= 0; i
< n_flows
; i
++) {
2864 struct odputil_keybuf
*maskbuf
= &thread
->maskbuf
[i
];
2865 struct odputil_keybuf
*keybuf
= &thread
->keybuf
[i
];
2866 struct dp_netdev_flow
*netdev_flow
= netdev_flows
[i
];
2867 struct dpif_flow
*f
= &flows
[i
];
2868 struct ofpbuf key
, mask
;
2870 ofpbuf_use_stack(&key
, keybuf
, sizeof *keybuf
);
2871 ofpbuf_use_stack(&mask
, maskbuf
, sizeof *maskbuf
);
2872 dp_netdev_flow_to_dpif_flow(netdev_flow
, &key
, &mask
, f
,
2880 dpif_netdev_execute(struct dpif
*dpif
, struct dpif_execute
*execute
)
2881 OVS_NO_THREAD_SAFETY_ANALYSIS
2883 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2884 struct dp_netdev_pmd_thread
*pmd
;
2885 struct dp_packet_batch pp
;
2887 if (dp_packet_size(execute
->packet
) < ETH_HEADER_LEN
||
2888 dp_packet_size(execute
->packet
) > UINT16_MAX
) {
2892 /* Tries finding the 'pmd'. If NULL is returned, that means
2893 * the current thread is a non-pmd thread and should use
2894 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2895 pmd
= ovsthread_getspecific(dp
->per_pmd_key
);
2897 pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
2903 if (execute
->probe
) {
2904 /* If this is part of a probe, Drop the packet, since executing
2905 * the action may actually cause spurious packets be sent into
2910 /* If the current thread is non-pmd thread, acquires
2911 * the 'non_pmd_mutex'. */
2912 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2913 ovs_mutex_lock(&dp
->non_pmd_mutex
);
2916 /* The action processing expects the RSS hash to be valid, because
2917 * it's always initialized at the beginning of datapath processing.
2918 * In this case, though, 'execute->packet' may not have gone through
2919 * the datapath at all, it may have been generated by the upper layer
2920 * (OpenFlow packet-out, BFD frame, ...). */
2921 if (!dp_packet_rss_valid(execute
->packet
)) {
2922 dp_packet_set_rss_hash(execute
->packet
,
2923 flow_hash_5tuple(execute
->flow
, 0));
2926 dp_packet_batch_init_packet(&pp
, execute
->packet
);
2927 dp_netdev_execute_actions(pmd
, &pp
, false, execute
->flow
,
2928 execute
->actions
, execute
->actions_len
,
2931 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2932 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
2933 dp_netdev_pmd_unref(pmd
);
2940 dpif_netdev_operate(struct dpif
*dpif
, struct dpif_op
**ops
, size_t n_ops
)
2944 for (i
= 0; i
< n_ops
; i
++) {
2945 struct dpif_op
*op
= ops
[i
];
2948 case DPIF_OP_FLOW_PUT
:
2949 op
->error
= dpif_netdev_flow_put(dpif
, &op
->u
.flow_put
);
2952 case DPIF_OP_FLOW_DEL
:
2953 op
->error
= dpif_netdev_flow_del(dpif
, &op
->u
.flow_del
);
2956 case DPIF_OP_EXECUTE
:
2957 op
->error
= dpif_netdev_execute(dpif
, &op
->u
.execute
);
2960 case DPIF_OP_FLOW_GET
:
2961 op
->error
= dpif_netdev_flow_get(dpif
, &op
->u
.flow_get
);
2967 /* Applies datapath configuration from the database. Some of the changes are
2968 * actually applied in dpif_netdev_run(). */
2970 dpif_netdev_set_config(struct dpif
*dpif
, const struct smap
*other_config
)
2972 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2973 const char *cmask
= smap_get(other_config
, "pmd-cpu-mask");
2974 unsigned long long insert_prob
=
2975 smap_get_ullong(other_config
, "emc-insert-inv-prob",
2976 DEFAULT_EM_FLOW_INSERT_INV_PROB
);
2977 uint32_t insert_min
, cur_min
;
2979 if (!nullable_string_is_equal(dp
->pmd_cmask
, cmask
)) {
2980 free(dp
->pmd_cmask
);
2981 dp
->pmd_cmask
= nullable_xstrdup(cmask
);
2982 dp_netdev_request_reconfigure(dp
);
2985 atomic_read_relaxed(&dp
->emc_insert_min
, &cur_min
);
2986 if (insert_prob
<= UINT32_MAX
) {
2987 insert_min
= insert_prob
== 0 ? 0 : UINT32_MAX
/ insert_prob
;
2989 insert_min
= DEFAULT_EM_FLOW_INSERT_MIN
;
2990 insert_prob
= DEFAULT_EM_FLOW_INSERT_INV_PROB
;
2993 if (insert_min
!= cur_min
) {
2994 atomic_store_relaxed(&dp
->emc_insert_min
, insert_min
);
2995 if (insert_min
== 0) {
2996 VLOG_INFO("EMC has been disabled");
2998 VLOG_INFO("EMC insertion probability changed to 1/%llu (~%.2f%%)",
2999 insert_prob
, (100 / (float)insert_prob
));
3006 /* Parses affinity list and returns result in 'core_ids'. */
3008 parse_affinity_list(const char *affinity_list
, unsigned *core_ids
, int n_rxq
)
3011 char *list
, *copy
, *key
, *value
;
3014 for (i
= 0; i
< n_rxq
; i
++) {
3015 core_ids
[i
] = OVS_CORE_UNSPEC
;
3018 if (!affinity_list
) {
3022 list
= copy
= xstrdup(affinity_list
);
3024 while (ofputil_parse_key_value(&list
, &key
, &value
)) {
3025 int rxq_id
, core_id
;
3027 if (!str_to_int(key
, 0, &rxq_id
) || rxq_id
< 0
3028 || !str_to_int(value
, 0, &core_id
) || core_id
< 0) {
3033 if (rxq_id
< n_rxq
) {
3034 core_ids
[rxq_id
] = core_id
;
3042 /* Parses 'affinity_list' and applies configuration if it is valid. */
3044 dpif_netdev_port_set_rxq_affinity(struct dp_netdev_port
*port
,
3045 const char *affinity_list
)
3047 unsigned *core_ids
, i
;
3050 core_ids
= xmalloc(port
->n_rxq
* sizeof *core_ids
);
3051 if (parse_affinity_list(affinity_list
, core_ids
, port
->n_rxq
)) {
3056 for (i
= 0; i
< port
->n_rxq
; i
++) {
3057 port
->rxqs
[i
].core_id
= core_ids
[i
];
3065 /* Changes the affinity of port's rx queues. The changes are actually applied
3066 * in dpif_netdev_run(). */
3068 dpif_netdev_port_set_config(struct dpif
*dpif
, odp_port_t port_no
,
3069 const struct smap
*cfg
)
3071 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3072 struct dp_netdev_port
*port
;
3074 const char *affinity_list
= smap_get(cfg
, "pmd-rxq-affinity");
3076 ovs_mutex_lock(&dp
->port_mutex
);
3077 error
= get_port_by_number(dp
, port_no
, &port
);
3078 if (error
|| !netdev_is_pmd(port
->netdev
)
3079 || nullable_string_is_equal(affinity_list
, port
->rxq_affinity_list
)) {
3083 error
= dpif_netdev_port_set_rxq_affinity(port
, affinity_list
);
3087 free(port
->rxq_affinity_list
);
3088 port
->rxq_affinity_list
= nullable_xstrdup(affinity_list
);
3090 dp_netdev_request_reconfigure(dp
);
3092 ovs_mutex_unlock(&dp
->port_mutex
);
3097 dpif_netdev_queue_to_priority(const struct dpif
*dpif OVS_UNUSED
,
3098 uint32_t queue_id
, uint32_t *priority
)
3100 *priority
= queue_id
;
3105 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
3106 * a copy of the 'size' bytes of 'actions' input parameters. */
3107 struct dp_netdev_actions
*
3108 dp_netdev_actions_create(const struct nlattr
*actions
, size_t size
)
3110 struct dp_netdev_actions
*netdev_actions
;
3112 netdev_actions
= xmalloc(sizeof *netdev_actions
+ size
);
3113 memcpy(netdev_actions
->actions
, actions
, size
);
3114 netdev_actions
->size
= size
;
3116 return netdev_actions
;
3119 struct dp_netdev_actions
*
3120 dp_netdev_flow_get_actions(const struct dp_netdev_flow
*flow
)
3122 return ovsrcu_get(struct dp_netdev_actions
*, &flow
->actions
);
3126 dp_netdev_actions_free(struct dp_netdev_actions
*actions
)
3131 static inline unsigned long long
3132 cycles_counter(void)
3135 return rte_get_tsc_cycles();
3141 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
3142 extern struct ovs_mutex cycles_counter_fake_mutex
;
3144 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
3146 cycles_count_start(struct dp_netdev_pmd_thread
*pmd
)
3147 OVS_ACQUIRES(&cycles_counter_fake_mutex
)
3148 OVS_NO_THREAD_SAFETY_ANALYSIS
3150 pmd
->last_cycles
= cycles_counter();
3153 /* Stop counting cycles and add them to the counter 'type' */
3155 cycles_count_end(struct dp_netdev_pmd_thread
*pmd
,
3156 enum pmd_cycles_counter_type type
)
3157 OVS_RELEASES(&cycles_counter_fake_mutex
)
3158 OVS_NO_THREAD_SAFETY_ANALYSIS
3160 unsigned long long interval
= cycles_counter() - pmd
->last_cycles
;
3162 non_atomic_ullong_add(&pmd
->cycles
.n
[type
], interval
);
3165 /* Calculate the intermediate cycle result and add to the counter 'type' */
3167 cycles_count_intermediate(struct dp_netdev_pmd_thread
*pmd
,
3168 struct dp_netdev_rxq
*rxq
,
3169 enum pmd_cycles_counter_type type
)
3170 OVS_NO_THREAD_SAFETY_ANALYSIS
3172 unsigned long long new_cycles
= cycles_counter();
3173 unsigned long long interval
= new_cycles
- pmd
->last_cycles
;
3174 pmd
->last_cycles
= new_cycles
;
3176 non_atomic_ullong_add(&pmd
->cycles
.n
[type
], interval
);
3177 if (rxq
&& (type
== PMD_CYCLES_PROCESSING
)) {
3178 /* Add to the amount of current processing cycles. */
3179 non_atomic_ullong_add(&rxq
->cycles
[RXQ_CYCLES_PROC_CURR
], interval
);
3184 dp_netdev_rxq_set_cycles(struct dp_netdev_rxq
*rx
,
3185 enum rxq_cycles_counter_type type
,
3186 unsigned long long cycles
)
3188 atomic_store_relaxed(&rx
->cycles
[type
], cycles
);
3192 dp_netdev_rxq_get_cycles(struct dp_netdev_rxq
*rx
,
3193 enum rxq_cycles_counter_type type
)
3195 unsigned long long processing_cycles
;
3196 atomic_read_relaxed(&rx
->cycles
[type
], &processing_cycles
);
3197 return processing_cycles
;
3201 dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq
*rx
,
3202 unsigned long long cycles
)
3204 atomic_store_relaxed(&rx
->cycles_intrvl
[rx
->intrvl_idx
++
3205 % PMD_RXQ_INTERVAL_MAX
], cycles
);
3209 dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq
*rx
, unsigned idx
)
3211 unsigned long long processing_cycles
;
3212 atomic_read_relaxed(&rx
->cycles_intrvl
[idx
], &processing_cycles
);
3213 return processing_cycles
;
3217 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread
*pmd
,
3218 struct netdev_rxq
*rx
,
3221 struct dp_packet_batch batch
;
3225 dp_packet_batch_init(&batch
);
3226 error
= netdev_rxq_recv(rx
, &batch
);
3228 *recirc_depth_get() = 0;
3230 batch_cnt
= batch
.count
;
3231 dp_netdev_input(pmd
, &batch
, port_no
);
3232 } else if (error
!= EAGAIN
&& error
!= EOPNOTSUPP
) {
3233 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
3235 VLOG_ERR_RL(&rl
, "error receiving data from %s: %s",
3236 netdev_rxq_get_name(rx
), ovs_strerror(error
));
3242 static struct tx_port
*
3243 tx_port_lookup(const struct hmap
*hmap
, odp_port_t port_no
)
3247 HMAP_FOR_EACH_IN_BUCKET (tx
, node
, hash_port_no(port_no
), hmap
) {
3248 if (tx
->port
->port_no
== port_no
) {
3257 port_reconfigure(struct dp_netdev_port
*port
)
3259 struct netdev
*netdev
= port
->netdev
;
3262 port
->need_reconfigure
= false;
3264 /* Closes the existing 'rxq's. */
3265 for (i
= 0; i
< port
->n_rxq
; i
++) {
3266 netdev_rxq_close(port
->rxqs
[i
].rx
);
3267 port
->rxqs
[i
].rx
= NULL
;
3269 unsigned last_nrxq
= port
->n_rxq
;
3272 /* Allows 'netdev' to apply the pending configuration changes. */
3273 if (netdev_is_reconf_required(netdev
)) {
3274 err
= netdev_reconfigure(netdev
);
3275 if (err
&& (err
!= EOPNOTSUPP
)) {
3276 VLOG_ERR("Failed to set interface %s new configuration",
3277 netdev_get_name(netdev
));
3281 /* If the netdev_reconfigure() above succeeds, reopens the 'rxq's. */
3282 port
->rxqs
= xrealloc(port
->rxqs
,
3283 sizeof *port
->rxqs
* netdev_n_rxq(netdev
));
3284 /* Realloc 'used' counters for tx queues. */
3285 free(port
->txq_used
);
3286 port
->txq_used
= xcalloc(netdev_n_txq(netdev
), sizeof *port
->txq_used
);
3288 for (i
= 0; i
< netdev_n_rxq(netdev
); i
++) {
3289 port
->rxqs
[i
].port
= port
;
3290 if (i
>= last_nrxq
) {
3291 /* Only reset cycle stats for new queues */
3292 dp_netdev_rxq_set_cycles(&port
->rxqs
[i
], RXQ_CYCLES_PROC_CURR
, 0);
3293 dp_netdev_rxq_set_cycles(&port
->rxqs
[i
], RXQ_CYCLES_PROC_HIST
, 0);
3294 for (unsigned j
= 0; j
< PMD_RXQ_INTERVAL_MAX
; j
++) {
3295 dp_netdev_rxq_set_intrvl_cycles(&port
->rxqs
[i
], 0);
3298 err
= netdev_rxq_open(netdev
, &port
->rxqs
[i
].rx
, i
);
3305 /* Parse affinity list to apply configuration for new queues. */
3306 dpif_netdev_port_set_rxq_affinity(port
, port
->rxq_affinity_list
);
3311 struct rr_numa_list
{
3312 struct hmap numas
; /* Contains 'struct rr_numa' */
3316 struct hmap_node node
;
3320 /* Non isolated pmds on numa node 'numa_id' */
3321 struct dp_netdev_pmd_thread
**pmds
;
3328 static struct rr_numa
*
3329 rr_numa_list_lookup(struct rr_numa_list
*rr
, int numa_id
)
3331 struct rr_numa
*numa
;
3333 HMAP_FOR_EACH_WITH_HASH (numa
, node
, hash_int(numa_id
, 0), &rr
->numas
) {
3334 if (numa
->numa_id
== numa_id
) {
3342 /* Returns the next node in numa list following 'numa' in round-robin fashion.
3343 * Returns first node if 'numa' is a null pointer or the last node in 'rr'.
3344 * Returns NULL if 'rr' numa list is empty. */
3345 static struct rr_numa
*
3346 rr_numa_list_next(struct rr_numa_list
*rr
, const struct rr_numa
*numa
)
3348 struct hmap_node
*node
= NULL
;
3351 node
= hmap_next(&rr
->numas
, &numa
->node
);
3354 node
= hmap_first(&rr
->numas
);
3357 return (node
) ? CONTAINER_OF(node
, struct rr_numa
, node
) : NULL
;
3361 rr_numa_list_populate(struct dp_netdev
*dp
, struct rr_numa_list
*rr
)
3363 struct dp_netdev_pmd_thread
*pmd
;
3364 struct rr_numa
*numa
;
3366 hmap_init(&rr
->numas
);
3368 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3369 if (pmd
->core_id
== NON_PMD_CORE_ID
|| pmd
->isolated
) {
3373 numa
= rr_numa_list_lookup(rr
, pmd
->numa_id
);
3375 numa
= xzalloc(sizeof *numa
);
3376 numa
->numa_id
= pmd
->numa_id
;
3377 hmap_insert(&rr
->numas
, &numa
->node
, hash_int(pmd
->numa_id
, 0));
3380 numa
->pmds
= xrealloc(numa
->pmds
, numa
->n_pmds
* sizeof *numa
->pmds
);
3381 numa
->pmds
[numa
->n_pmds
- 1] = pmd
;
3382 /* At least one pmd so initialise curr_idx and idx_inc. */
3383 numa
->cur_index
= 0;
3384 numa
->idx_inc
= true;
3388 /* Returns the next pmd from the numa node in
3389 * incrementing or decrementing order. */
3390 static struct dp_netdev_pmd_thread
*
3391 rr_numa_get_pmd(struct rr_numa
*numa
)
3393 int numa_idx
= numa
->cur_index
;
3395 if (numa
->idx_inc
== true) {
3396 /* Incrementing through list of pmds. */
3397 if (numa
->cur_index
== numa
->n_pmds
-1) {
3398 /* Reached the last pmd. */
3399 numa
->idx_inc
= false;
3404 /* Decrementing through list of pmds. */
3405 if (numa
->cur_index
== 0) {
3406 /* Reached the first pmd. */
3407 numa
->idx_inc
= true;
3412 return numa
->pmds
[numa_idx
];
3416 rr_numa_list_destroy(struct rr_numa_list
*rr
)
3418 struct rr_numa
*numa
;
3420 HMAP_FOR_EACH_POP (numa
, node
, &rr
->numas
) {
3424 hmap_destroy(&rr
->numas
);
3427 /* Sort Rx Queues by the processing cycles they are consuming. */
3429 rxq_cycle_sort(const void *a
, const void *b
)
3431 struct dp_netdev_rxq
* qa
;
3432 struct dp_netdev_rxq
* qb
;
3433 uint64_t total_qa
, total_qb
;
3436 qa
= *(struct dp_netdev_rxq
**) a
;
3437 qb
= *(struct dp_netdev_rxq
**) b
;
3439 total_qa
= total_qb
= 0;
3440 for (i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
3441 total_qa
+= dp_netdev_rxq_get_intrvl_cycles(qa
, i
);
3442 total_qb
+= dp_netdev_rxq_get_intrvl_cycles(qb
, i
);
3444 dp_netdev_rxq_set_cycles(qa
, RXQ_CYCLES_PROC_HIST
, total_qa
);
3445 dp_netdev_rxq_set_cycles(qb
, RXQ_CYCLES_PROC_HIST
, total_qb
);
3447 if (total_qa
>= total_qb
) {
3453 /* Assign pmds to queues. If 'pinned' is true, assign pmds to pinned
3454 * queues and marks the pmds as isolated. Otherwise, assign non isolated
3455 * pmds to unpinned queues.
3457 * If 'pinned' is false queues will be sorted by processing cycles they are
3458 * consuming and then assigned to pmds in round robin order.
3460 * The function doesn't touch the pmd threads, it just stores the assignment
3461 * in the 'pmd' member of each rxq. */
3463 rxq_scheduling(struct dp_netdev
*dp
, bool pinned
) OVS_REQUIRES(dp
->port_mutex
)
3465 struct dp_netdev_port
*port
;
3466 struct rr_numa_list rr
;
3467 struct rr_numa
*non_local_numa
= NULL
;
3468 struct dp_netdev_rxq
** rxqs
= NULL
;
3470 struct rr_numa
*numa
= NULL
;
3473 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3474 if (!netdev_is_pmd(port
->netdev
)) {
3478 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
3479 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
3481 if (pinned
&& q
->core_id
!= OVS_CORE_UNSPEC
) {
3482 struct dp_netdev_pmd_thread
*pmd
;
3484 pmd
= dp_netdev_get_pmd(dp
, q
->core_id
);
3486 VLOG_WARN("There is no PMD thread on core %d. Queue "
3487 "%d on port \'%s\' will not be polled.",
3488 q
->core_id
, qid
, netdev_get_name(port
->netdev
));
3491 pmd
->isolated
= true;
3492 dp_netdev_pmd_unref(pmd
);
3494 } else if (!pinned
&& q
->core_id
== OVS_CORE_UNSPEC
) {
3496 rxqs
= xmalloc(sizeof *rxqs
);
3498 rxqs
= xrealloc(rxqs
, sizeof *rxqs
* (n_rxqs
+ 1));
3500 /* Store the queue. */
3507 /* Sort the queues in order of the processing cycles
3508 * they consumed during their last pmd interval. */
3509 qsort(rxqs
, n_rxqs
, sizeof *rxqs
, rxq_cycle_sort
);
3512 rr_numa_list_populate(dp
, &rr
);
3513 /* Assign the sorted queues to pmds in round robin. */
3514 for (i
= 0; i
< n_rxqs
; i
++) {
3515 numa_id
= netdev_get_numa_id(rxqs
[i
]->port
->netdev
);
3516 numa
= rr_numa_list_lookup(&rr
, numa_id
);
3518 /* There are no pmds on the queue's local NUMA node.
3519 Round robin on the NUMA nodes that do have pmds. */
3520 non_local_numa
= rr_numa_list_next(&rr
, non_local_numa
);
3521 if (!non_local_numa
) {
3522 VLOG_ERR("There is no available (non-isolated) pmd "
3523 "thread for port \'%s\' queue %d. This queue "
3524 "will not be polled. Is pmd-cpu-mask set to "
3525 "zero? Or are all PMDs isolated to other "
3526 "queues?", netdev_rxq_get_name(rxqs
[i
]->rx
),
3527 netdev_rxq_get_queue_id(rxqs
[i
]->rx
));
3530 rxqs
[i
]->pmd
= rr_numa_get_pmd(non_local_numa
);
3531 VLOG_WARN("There's no available (non-isolated) pmd thread "
3532 "on numa node %d. Queue %d on port \'%s\' will "
3533 "be assigned to the pmd on core %d "
3534 "(numa node %d). Expect reduced performance.",
3535 numa_id
, netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
3536 netdev_rxq_get_name(rxqs
[i
]->rx
),
3537 rxqs
[i
]->pmd
->core_id
, rxqs
[i
]->pmd
->numa_id
);
3539 rxqs
[i
]->pmd
= rr_numa_get_pmd(numa
);
3540 VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
3541 "rx queue %d (measured processing cycles %"PRIu64
").",
3542 rxqs
[i
]->pmd
->core_id
, numa_id
,
3543 netdev_rxq_get_name(rxqs
[i
]->rx
),
3544 netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
3545 dp_netdev_rxq_get_cycles(rxqs
[i
], RXQ_CYCLES_PROC_HIST
));
3549 rr_numa_list_destroy(&rr
);
3554 reload_affected_pmds(struct dp_netdev
*dp
)
3556 struct dp_netdev_pmd_thread
*pmd
;
3558 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3559 if (pmd
->need_reload
) {
3560 dp_netdev_reload_pmd__(pmd
);
3561 pmd
->need_reload
= false;
3567 reconfigure_pmd_threads(struct dp_netdev
*dp
)
3568 OVS_REQUIRES(dp
->port_mutex
)
3570 struct dp_netdev_pmd_thread
*pmd
;
3571 struct ovs_numa_dump
*pmd_cores
;
3572 struct ovs_numa_info_core
*core
;
3573 struct hmapx to_delete
= HMAPX_INITIALIZER(&to_delete
);
3574 struct hmapx_node
*node
;
3575 bool changed
= false;
3576 bool need_to_adjust_static_tx_qids
= false;
3578 /* The pmd threads should be started only if there's a pmd port in the
3579 * datapath. If the user didn't provide any "pmd-cpu-mask", we start
3580 * NR_PMD_THREADS per numa node. */
3581 if (!has_pmd_port(dp
)) {
3582 pmd_cores
= ovs_numa_dump_n_cores_per_numa(0);
3583 } else if (dp
->pmd_cmask
&& dp
->pmd_cmask
[0]) {
3584 pmd_cores
= ovs_numa_dump_cores_with_cmask(dp
->pmd_cmask
);
3586 pmd_cores
= ovs_numa_dump_n_cores_per_numa(NR_PMD_THREADS
);
3589 /* We need to adjust 'static_tx_qid's only if we're reducing number of
3590 * PMD threads. Otherwise, new threads will allocate all the freed ids. */
3591 if (ovs_numa_dump_count(pmd_cores
) < cmap_count(&dp
->poll_threads
) - 1) {
3592 /* Adjustment is required to keep 'static_tx_qid's sequential and
3593 * avoid possible issues, for example, imbalanced tx queue usage
3594 * and unnecessary locking caused by remapping on netdev level. */
3595 need_to_adjust_static_tx_qids
= true;
3598 /* Check for unwanted pmd threads */
3599 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3600 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3603 if (!ovs_numa_dump_contains_core(pmd_cores
, pmd
->numa_id
,
3605 hmapx_add(&to_delete
, pmd
);
3606 } else if (need_to_adjust_static_tx_qids
) {
3607 pmd
->need_reload
= true;
3611 HMAPX_FOR_EACH (node
, &to_delete
) {
3612 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3613 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d destroyed.",
3614 pmd
->numa_id
, pmd
->core_id
);
3615 dp_netdev_del_pmd(dp
, pmd
);
3617 changed
= !hmapx_is_empty(&to_delete
);
3618 hmapx_destroy(&to_delete
);
3620 if (need_to_adjust_static_tx_qids
) {
3621 /* 'static_tx_qid's are not sequential now.
3622 * Reload remaining threads to fix this. */
3623 reload_affected_pmds(dp
);
3626 /* Check for required new pmd threads */
3627 FOR_EACH_CORE_ON_DUMP(core
, pmd_cores
) {
3628 pmd
= dp_netdev_get_pmd(dp
, core
->core_id
);
3630 pmd
= xzalloc(sizeof *pmd
);
3631 dp_netdev_configure_pmd(pmd
, dp
, core
->core_id
, core
->numa_id
);
3632 pmd
->thread
= ovs_thread_create("pmd", pmd_thread_main
, pmd
);
3633 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d created.",
3634 pmd
->numa_id
, pmd
->core_id
);
3637 dp_netdev_pmd_unref(pmd
);
3642 struct ovs_numa_info_numa
*numa
;
3644 /* Log the number of pmd threads per numa node. */
3645 FOR_EACH_NUMA_ON_DUMP (numa
, pmd_cores
) {
3646 VLOG_INFO("There are %"PRIuSIZE
" pmd threads on numa node %d",
3647 numa
->n_cores
, numa
->numa_id
);
3651 ovs_numa_dump_destroy(pmd_cores
);
3655 pmd_remove_stale_ports(struct dp_netdev
*dp
,
3656 struct dp_netdev_pmd_thread
*pmd
)
3657 OVS_EXCLUDED(pmd
->port_mutex
)
3658 OVS_REQUIRES(dp
->port_mutex
)
3660 struct rxq_poll
*poll
, *poll_next
;
3661 struct tx_port
*tx
, *tx_next
;
3663 ovs_mutex_lock(&pmd
->port_mutex
);
3664 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
3665 struct dp_netdev_port
*port
= poll
->rxq
->port
;
3667 if (port
->need_reconfigure
3668 || !hmap_contains(&dp
->ports
, &port
->node
)) {
3669 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
3672 HMAP_FOR_EACH_SAFE (tx
, tx_next
, node
, &pmd
->tx_ports
) {
3673 struct dp_netdev_port
*port
= tx
->port
;
3675 if (port
->need_reconfigure
3676 || !hmap_contains(&dp
->ports
, &port
->node
)) {
3677 dp_netdev_del_port_tx_from_pmd(pmd
, tx
);
3680 ovs_mutex_unlock(&pmd
->port_mutex
);
3683 /* Must be called each time a port is added/removed or the cmask changes.
3684 * This creates and destroys pmd threads, reconfigures ports, opens their
3685 * rxqs and assigns all rxqs/txqs to pmd threads. */
3687 reconfigure_datapath(struct dp_netdev
*dp
)
3688 OVS_REQUIRES(dp
->port_mutex
)
3690 struct dp_netdev_pmd_thread
*pmd
;
3691 struct dp_netdev_port
*port
;
3694 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
3696 /* Step 1: Adjust the pmd threads based on the datapath ports, the cores
3697 * on the system and the user configuration. */
3698 reconfigure_pmd_threads(dp
);
3700 wanted_txqs
= cmap_count(&dp
->poll_threads
);
3702 /* The number of pmd threads might have changed, or a port can be new:
3703 * adjust the txqs. */
3704 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3705 netdev_set_tx_multiq(port
->netdev
, wanted_txqs
);
3708 /* Step 2: Remove from the pmd threads ports that have been removed or
3709 * need reconfiguration. */
3711 /* Check for all the ports that need reconfiguration. We cache this in
3712 * 'port->need_reconfigure', because netdev_is_reconf_required() can
3713 * change at any time. */
3714 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3715 if (netdev_is_reconf_required(port
->netdev
)) {
3716 port
->need_reconfigure
= true;
3720 /* Remove from the pmd threads all the ports that have been deleted or
3721 * need reconfiguration. */
3722 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3723 pmd_remove_stale_ports(dp
, pmd
);
3726 /* Reload affected pmd threads. We must wait for the pmd threads before
3727 * reconfiguring the ports, because a port cannot be reconfigured while
3728 * it's being used. */
3729 reload_affected_pmds(dp
);
3731 /* Step 3: Reconfigure ports. */
3733 /* We only reconfigure the ports that we determined above, because they're
3734 * not being used by any pmd thread at the moment. If a port fails to
3735 * reconfigure we remove it from the datapath. */
3736 struct dp_netdev_port
*next_port
;
3737 HMAP_FOR_EACH_SAFE (port
, next_port
, node
, &dp
->ports
) {
3740 if (!port
->need_reconfigure
) {
3744 err
= port_reconfigure(port
);
3746 hmap_remove(&dp
->ports
, &port
->node
);
3747 seq_change(dp
->port_seq
);
3750 port
->dynamic_txqs
= netdev_n_txq(port
->netdev
) < wanted_txqs
;
3754 /* Step 4: Compute new rxq scheduling. We don't touch the pmd threads
3755 * for now, we just update the 'pmd' pointer in each rxq to point to the
3756 * wanted thread according to the scheduling policy. */
3758 /* Reset all the pmd threads to non isolated. */
3759 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3760 pmd
->isolated
= false;
3763 /* Reset all the queues to unassigned */
3764 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3765 for (int i
= 0; i
< port
->n_rxq
; i
++) {
3766 port
->rxqs
[i
].pmd
= NULL
;
3770 /* Add pinned queues and mark pmd threads isolated. */
3771 rxq_scheduling(dp
, true);
3773 /* Add non-pinned queues. */
3774 rxq_scheduling(dp
, false);
3776 /* Step 5: Remove queues not compliant with new scheduling. */
3777 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3778 struct rxq_poll
*poll
, *poll_next
;
3780 ovs_mutex_lock(&pmd
->port_mutex
);
3781 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
3782 if (poll
->rxq
->pmd
!= pmd
) {
3783 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
3786 ovs_mutex_unlock(&pmd
->port_mutex
);
3789 /* Reload affected pmd threads. We must wait for the pmd threads to remove
3790 * the old queues before readding them, otherwise a queue can be polled by
3791 * two threads at the same time. */
3792 reload_affected_pmds(dp
);
3794 /* Step 6: Add queues from scheduling, if they're not there already. */
3795 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3796 if (!netdev_is_pmd(port
->netdev
)) {
3800 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
3801 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
3804 ovs_mutex_lock(&q
->pmd
->port_mutex
);
3805 dp_netdev_add_rxq_to_pmd(q
->pmd
, q
);
3806 ovs_mutex_unlock(&q
->pmd
->port_mutex
);
3811 /* Add every port to the tx cache of every pmd thread, if it's not
3812 * there already and if this pmd has at least one rxq to poll. */
3813 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3814 ovs_mutex_lock(&pmd
->port_mutex
);
3815 if (hmap_count(&pmd
->poll_list
) || pmd
->core_id
== NON_PMD_CORE_ID
) {
3816 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3817 dp_netdev_add_port_tx_to_pmd(pmd
, port
);
3820 ovs_mutex_unlock(&pmd
->port_mutex
);
3823 /* Reload affected pmd threads. */
3824 reload_affected_pmds(dp
);
3827 /* Returns true if one of the netdevs in 'dp' requires a reconfiguration */
3829 ports_require_restart(const struct dp_netdev
*dp
)
3830 OVS_REQUIRES(dp
->port_mutex
)
3832 struct dp_netdev_port
*port
;
3834 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3835 if (netdev_is_reconf_required(port
->netdev
)) {
3843 /* Return true if needs to revalidate datapath flows. */
3845 dpif_netdev_run(struct dpif
*dpif
)
3847 struct dp_netdev_port
*port
;
3848 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3849 struct dp_netdev_pmd_thread
*non_pmd
;
3850 uint64_t new_tnl_seq
;
3851 int process_packets
= 0;
3853 ovs_mutex_lock(&dp
->port_mutex
);
3854 non_pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
3856 ovs_mutex_lock(&dp
->non_pmd_mutex
);
3857 cycles_count_start(non_pmd
);
3858 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3859 if (!netdev_is_pmd(port
->netdev
)) {
3862 for (i
= 0; i
< port
->n_rxq
; i
++) {
3864 dp_netdev_process_rxq_port(non_pmd
,
3867 cycles_count_intermediate(non_pmd
, NULL
, process_packets
?
3868 PMD_CYCLES_PROCESSING
3873 cycles_count_end(non_pmd
, PMD_CYCLES_IDLE
);
3874 dpif_netdev_xps_revalidate_pmd(non_pmd
, time_msec(), false);
3875 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
3877 dp_netdev_pmd_unref(non_pmd
);
3880 if (dp_netdev_is_reconf_required(dp
) || ports_require_restart(dp
)) {
3881 reconfigure_datapath(dp
);
3883 ovs_mutex_unlock(&dp
->port_mutex
);
3885 tnl_neigh_cache_run();
3887 new_tnl_seq
= seq_read(tnl_conf_seq
);
3889 if (dp
->last_tnl_conf_seq
!= new_tnl_seq
) {
3890 dp
->last_tnl_conf_seq
= new_tnl_seq
;
3897 dpif_netdev_wait(struct dpif
*dpif
)
3899 struct dp_netdev_port
*port
;
3900 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3902 ovs_mutex_lock(&dp_netdev_mutex
);
3903 ovs_mutex_lock(&dp
->port_mutex
);
3904 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3905 netdev_wait_reconf_required(port
->netdev
);
3906 if (!netdev_is_pmd(port
->netdev
)) {
3909 for (i
= 0; i
< port
->n_rxq
; i
++) {
3910 netdev_rxq_wait(port
->rxqs
[i
].rx
);
3914 ovs_mutex_unlock(&dp
->port_mutex
);
3915 ovs_mutex_unlock(&dp_netdev_mutex
);
3916 seq_wait(tnl_conf_seq
, dp
->last_tnl_conf_seq
);
3920 pmd_free_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
3922 struct tx_port
*tx_port_cached
;
3924 /* Free all used tx queue ids. */
3925 dpif_netdev_xps_revalidate_pmd(pmd
, 0, true);
3927 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->tnl_port_cache
) {
3928 free(tx_port_cached
);
3930 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->send_port_cache
) {
3931 free(tx_port_cached
);
3935 /* Copies ports from 'pmd->tx_ports' (shared with the main thread) to
3936 * 'pmd->port_cache' (thread local) */
3938 pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
3939 OVS_REQUIRES(pmd
->port_mutex
)
3941 struct tx_port
*tx_port
, *tx_port_cached
;
3943 pmd_free_cached_ports(pmd
);
3944 hmap_shrink(&pmd
->send_port_cache
);
3945 hmap_shrink(&pmd
->tnl_port_cache
);
3947 HMAP_FOR_EACH (tx_port
, node
, &pmd
->tx_ports
) {
3948 if (netdev_has_tunnel_push_pop(tx_port
->port
->netdev
)) {
3949 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
3950 hmap_insert(&pmd
->tnl_port_cache
, &tx_port_cached
->node
,
3951 hash_port_no(tx_port_cached
->port
->port_no
));
3954 if (netdev_n_txq(tx_port
->port
->netdev
)) {
3955 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
3956 hmap_insert(&pmd
->send_port_cache
, &tx_port_cached
->node
,
3957 hash_port_no(tx_port_cached
->port
->port_no
));
3963 pmd_alloc_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
3965 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
3966 if (!id_pool_alloc_id(pmd
->dp
->tx_qid_pool
, &pmd
->static_tx_qid
)) {
3967 VLOG_ABORT("static_tx_qid allocation failed for PMD on core %2d"
3968 ", numa_id %d.", pmd
->core_id
, pmd
->numa_id
);
3970 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
3972 VLOG_DBG("static_tx_qid = %d allocated for PMD thread on core %2d"
3973 ", numa_id %d.", pmd
->static_tx_qid
, pmd
->core_id
, pmd
->numa_id
);
3977 pmd_free_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
3979 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
3980 id_pool_free_id(pmd
->dp
->tx_qid_pool
, pmd
->static_tx_qid
);
3981 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
3985 pmd_load_queues_and_ports(struct dp_netdev_pmd_thread
*pmd
,
3986 struct polled_queue
**ppoll_list
)
3988 struct polled_queue
*poll_list
= *ppoll_list
;
3989 struct rxq_poll
*poll
;
3992 ovs_mutex_lock(&pmd
->port_mutex
);
3993 poll_list
= xrealloc(poll_list
, hmap_count(&pmd
->poll_list
)
3994 * sizeof *poll_list
);
3997 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
3998 poll_list
[i
].rxq
= poll
->rxq
;
3999 poll_list
[i
].port_no
= poll
->rxq
->port
->port_no
;
4003 pmd_load_cached_ports(pmd
);
4005 ovs_mutex_unlock(&pmd
->port_mutex
);
4007 *ppoll_list
= poll_list
;
4012 pmd_thread_main(void *f_
)
4014 struct dp_netdev_pmd_thread
*pmd
= f_
;
4015 unsigned int lc
= 0;
4016 struct polled_queue
*poll_list
;
4020 int process_packets
= 0;
4024 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
4025 ovsthread_setspecific(pmd
->dp
->per_pmd_key
, pmd
);
4026 ovs_numa_thread_setaffinity_core(pmd
->core_id
);
4027 dpdk_set_lcore_id(pmd
->core_id
);
4028 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
4029 emc_cache_init(&pmd
->flow_cache
);
4031 pmd_alloc_static_tx_qid(pmd
);
4033 /* List port/core affinity */
4034 for (i
= 0; i
< poll_cnt
; i
++) {
4035 VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n",
4036 pmd
->core_id
, netdev_rxq_get_name(poll_list
[i
].rxq
->rx
),
4037 netdev_rxq_get_queue_id(poll_list
[i
].rxq
->rx
));
4041 while (seq_read(pmd
->reload_seq
) == pmd
->last_reload_seq
) {
4042 seq_wait(pmd
->reload_seq
, pmd
->last_reload_seq
);
4048 cycles_count_start(pmd
);
4050 for (i
= 0; i
< poll_cnt
; i
++) {
4052 dp_netdev_process_rxq_port(pmd
, poll_list
[i
].rxq
->rx
,
4053 poll_list
[i
].port_no
);
4054 cycles_count_intermediate(pmd
, poll_list
[i
].rxq
,
4055 process_packets
? PMD_CYCLES_PROCESSING
4064 coverage_try_clear();
4065 dp_netdev_pmd_try_optimize(pmd
, poll_list
, poll_cnt
);
4066 if (!ovsrcu_try_quiesce()) {
4067 emc_cache_slow_sweep(&pmd
->flow_cache
);
4070 atomic_read_relaxed(&pmd
->reload
, &reload
);
4077 cycles_count_end(pmd
, PMD_CYCLES_IDLE
);
4079 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
4080 exiting
= latch_is_set(&pmd
->exit_latch
);
4081 /* Signal here to make sure the pmd finishes
4082 * reloading the updated configuration. */
4083 dp_netdev_pmd_reload_done(pmd
);
4085 pmd_free_static_tx_qid(pmd
);
4091 emc_cache_uninit(&pmd
->flow_cache
);
4093 pmd_free_cached_ports(pmd
);
4098 dp_netdev_disable_upcall(struct dp_netdev
*dp
)
4099 OVS_ACQUIRES(dp
->upcall_rwlock
)
4101 fat_rwlock_wrlock(&dp
->upcall_rwlock
);
4107 dpif_netdev_meter_get_features(const struct dpif
* dpif OVS_UNUSED
,
4108 struct ofputil_meter_features
*features
)
4110 features
->max_meters
= MAX_METERS
;
4111 features
->band_types
= DP_SUPPORTED_METER_BAND_TYPES
;
4112 features
->capabilities
= DP_SUPPORTED_METER_FLAGS_MASK
;
4113 features
->max_bands
= MAX_BANDS
;
4114 features
->max_color
= 0;
4117 /* Returns false when packet needs to be dropped. */
4119 dp_netdev_run_meter(struct dp_netdev
*dp
, struct dp_packet_batch
*packets_
,
4120 uint32_t meter_id
, long long int now
)
4122 struct dp_meter
*meter
;
4123 struct dp_meter_band
*band
;
4124 long long int long_delta_t
; /* msec */
4125 uint32_t delta_t
; /* msec */
4127 int cnt
= packets_
->count
;
4128 uint32_t bytes
, volume
;
4129 int exceeded_band
[NETDEV_MAX_BURST
];
4130 uint32_t exceeded_rate
[NETDEV_MAX_BURST
];
4131 int exceeded_pkt
= cnt
; /* First packet that exceeded a band rate. */
4133 if (meter_id
>= MAX_METERS
) {
4137 meter_lock(dp
, meter_id
);
4138 meter
= dp
->meters
[meter_id
];
4143 /* Initialize as negative values. */
4144 memset(exceeded_band
, 0xff, cnt
* sizeof *exceeded_band
);
4145 /* Initialize as zeroes. */
4146 memset(exceeded_rate
, 0, cnt
* sizeof *exceeded_rate
);
4148 /* All packets will hit the meter at the same time. */
4149 long_delta_t
= (now
- meter
->used
); /* msec */
4151 /* Make sure delta_t will not be too large, so that bucket will not
4152 * wrap around below. */
4153 delta_t
= (long_delta_t
> (long long int)meter
->max_delta_t
)
4154 ? meter
->max_delta_t
: (uint32_t)long_delta_t
;
4156 /* Update meter stats. */
4158 meter
->packet_count
+= cnt
;
4160 for (i
= 0; i
< cnt
; i
++) {
4161 bytes
+= dp_packet_size(packets_
->packets
[i
]);
4163 meter
->byte_count
+= bytes
;
4165 /* Meters can operate in terms of packets per second or kilobits per
4167 if (meter
->flags
& OFPMF13_PKTPS
) {
4168 /* Rate in packets/second, bucket 1/1000 packets. */
4169 /* msec * packets/sec = 1/1000 packets. */
4170 volume
= cnt
* 1000; /* Take 'cnt' packets from the bucket. */
4172 /* Rate in kbps, bucket in bits. */
4173 /* msec * kbps = bits */
4177 /* Update all bands and find the one hit with the highest rate for each
4178 * packet (if any). */
4179 for (int m
= 0; m
< meter
->n_bands
; ++m
) {
4180 band
= &meter
->bands
[m
];
4182 /* Update band's bucket. */
4183 band
->bucket
+= delta_t
* band
->up
.rate
;
4184 if (band
->bucket
> band
->up
.burst_size
) {
4185 band
->bucket
= band
->up
.burst_size
;
4188 /* Drain the bucket for all the packets, if possible. */
4189 if (band
->bucket
>= volume
) {
4190 band
->bucket
-= volume
;
4192 int band_exceeded_pkt
;
4194 /* Band limit hit, must process packet-by-packet. */
4195 if (meter
->flags
& OFPMF13_PKTPS
) {
4196 band_exceeded_pkt
= band
->bucket
/ 1000;
4197 band
->bucket
%= 1000; /* Remainder stays in bucket. */
4199 /* Update the exceeding band for each exceeding packet.
4200 * (Only one band will be fired by a packet, and that
4201 * can be different for each packet.) */
4202 for (i
= band_exceeded_pkt
; i
< cnt
; i
++) {
4203 if (band
->up
.rate
> exceeded_rate
[i
]) {
4204 exceeded_rate
[i
] = band
->up
.rate
;
4205 exceeded_band
[i
] = m
;
4209 /* Packet sizes differ, must process one-by-one. */
4210 band_exceeded_pkt
= cnt
;
4211 for (i
= 0; i
< cnt
; i
++) {
4212 uint32_t bits
= dp_packet_size(packets_
->packets
[i
]) * 8;
4214 if (band
->bucket
>= bits
) {
4215 band
->bucket
-= bits
;
4217 if (i
< band_exceeded_pkt
) {
4218 band_exceeded_pkt
= i
;
4220 /* Update the exceeding band for the exceeding packet.
4221 * (Only one band will be fired by a packet, and that
4222 * can be different for each packet.) */
4223 if (band
->up
.rate
> exceeded_rate
[i
]) {
4224 exceeded_rate
[i
] = band
->up
.rate
;
4225 exceeded_band
[i
] = m
;
4230 /* Remember the first exceeding packet. */
4231 if (exceeded_pkt
> band_exceeded_pkt
) {
4232 exceeded_pkt
= band_exceeded_pkt
;
4237 /* Fire the highest rate band exceeded by each packet.
4238 * Drop packets if needed, by swapping packet to the end that will be
4240 const size_t size
= dp_packet_batch_size(packets_
);
4241 struct dp_packet
*packet
;
4243 DP_PACKET_BATCH_REFILL_FOR_EACH (j
, size
, packet
, packets_
) {
4244 if (exceeded_band
[j
] >= 0) {
4245 /* Meter drop packet. */
4246 band
= &meter
->bands
[exceeded_band
[j
]];
4247 band
->packet_count
+= 1;
4248 band
->byte_count
+= dp_packet_size(packet
);
4250 dp_packet_delete(packet
);
4252 /* Meter accepts packet. */
4253 dp_packet_batch_refill(packets_
, packet
, j
);
4257 meter_unlock(dp
, meter_id
);
4260 /* Meter set/get/del processing is still single-threaded. */
4262 dpif_netdev_meter_set(struct dpif
*dpif
, ofproto_meter_id
*meter_id
,
4263 struct ofputil_meter_config
*config
)
4265 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4266 uint32_t mid
= meter_id
->uint32
;
4267 struct dp_meter
*meter
;
4270 if (mid
>= MAX_METERS
) {
4271 return EFBIG
; /* Meter_id out of range. */
4274 if (config
->flags
& ~DP_SUPPORTED_METER_FLAGS_MASK
||
4275 !(config
->flags
& (OFPMF13_KBPS
| OFPMF13_PKTPS
))) {
4276 return EBADF
; /* Unsupported flags set */
4278 /* Validate bands */
4279 if (config
->n_bands
== 0 || config
->n_bands
> MAX_BANDS
) {
4280 return EINVAL
; /* Too many bands */
4282 for (i
= 0; i
< config
->n_bands
; ++i
) {
4283 switch (config
->bands
[i
].type
) {
4287 return ENODEV
; /* Unsupported band type */
4291 /* Allocate meter */
4292 meter
= xzalloc(sizeof *meter
4293 + config
->n_bands
* sizeof(struct dp_meter_band
));
4295 meter
->flags
= config
->flags
;
4296 meter
->n_bands
= config
->n_bands
;
4297 meter
->max_delta_t
= 0;
4298 meter
->used
= time_msec();
4301 for (i
= 0; i
< config
->n_bands
; ++i
) {
4302 uint32_t band_max_delta_t
;
4304 /* Set burst size to a workable value if none specified. */
4305 if (config
->bands
[i
].burst_size
== 0) {
4306 config
->bands
[i
].burst_size
= config
->bands
[i
].rate
;
4309 meter
->bands
[i
].up
= config
->bands
[i
];
4310 /* Convert burst size to the bucket units: */
4311 /* pkts => 1/1000 packets, kilobits => bits. */
4312 meter
->bands
[i
].up
.burst_size
*= 1000;
4313 /* Initialize bucket to empty. */
4314 meter
->bands
[i
].bucket
= 0;
4316 /* Figure out max delta_t that is enough to fill any bucket. */
4318 = meter
->bands
[i
].up
.burst_size
/ meter
->bands
[i
].up
.rate
;
4319 if (band_max_delta_t
> meter
->max_delta_t
) {
4320 meter
->max_delta_t
= band_max_delta_t
;
4324 meter_lock(dp
, mid
);
4325 dp_delete_meter(dp
, mid
); /* Free existing meter, if any */
4326 dp
->meters
[mid
] = meter
;
4327 meter_unlock(dp
, mid
);
4335 dpif_netdev_meter_get(const struct dpif
*dpif
,
4336 ofproto_meter_id meter_id_
,
4337 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
4339 const struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4340 const struct dp_meter
*meter
;
4341 uint32_t meter_id
= meter_id_
.uint32
;
4343 if (meter_id
>= MAX_METERS
) {
4346 meter
= dp
->meters
[meter_id
];
4353 meter_lock(dp
, meter_id
);
4354 stats
->packet_in_count
= meter
->packet_count
;
4355 stats
->byte_in_count
= meter
->byte_count
;
4357 for (i
= 0; i
< n_bands
&& i
< meter
->n_bands
; ++i
) {
4358 stats
->bands
[i
].packet_count
= meter
->bands
[i
].packet_count
;
4359 stats
->bands
[i
].byte_count
= meter
->bands
[i
].byte_count
;
4361 meter_unlock(dp
, meter_id
);
4369 dpif_netdev_meter_del(struct dpif
*dpif
,
4370 ofproto_meter_id meter_id_
,
4371 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
4373 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4376 error
= dpif_netdev_meter_get(dpif
, meter_id_
, stats
, n_bands
);
4378 uint32_t meter_id
= meter_id_
.uint32
;
4380 meter_lock(dp
, meter_id
);
4381 dp_delete_meter(dp
, meter_id
);
4382 meter_unlock(dp
, meter_id
);
4389 dpif_netdev_disable_upcall(struct dpif
*dpif
)
4390 OVS_NO_THREAD_SAFETY_ANALYSIS
4392 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4393 dp_netdev_disable_upcall(dp
);
4397 dp_netdev_enable_upcall(struct dp_netdev
*dp
)
4398 OVS_RELEASES(dp
->upcall_rwlock
)
4400 fat_rwlock_unlock(&dp
->upcall_rwlock
);
4404 dpif_netdev_enable_upcall(struct dpif
*dpif
)
4405 OVS_NO_THREAD_SAFETY_ANALYSIS
4407 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4408 dp_netdev_enable_upcall(dp
);
4412 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
)
4414 ovs_mutex_lock(&pmd
->cond_mutex
);
4415 atomic_store_relaxed(&pmd
->reload
, false);
4416 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
4417 xpthread_cond_signal(&pmd
->cond
);
4418 ovs_mutex_unlock(&pmd
->cond_mutex
);
4421 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
4422 * the pointer if succeeds, otherwise, NULL (it can return NULL even if
4423 * 'core_id' is NON_PMD_CORE_ID).
4425 * Caller must unrefs the returned reference. */
4426 static struct dp_netdev_pmd_thread
*
4427 dp_netdev_get_pmd(struct dp_netdev
*dp
, unsigned core_id
)
4429 struct dp_netdev_pmd_thread
*pmd
;
4430 const struct cmap_node
*pnode
;
4432 pnode
= cmap_find(&dp
->poll_threads
, hash_int(core_id
, 0));
4436 pmd
= CONTAINER_OF(pnode
, struct dp_netdev_pmd_thread
, node
);
4438 return dp_netdev_pmd_try_ref(pmd
) ? pmd
: NULL
;
4441 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
4443 dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
4444 OVS_REQUIRES(dp
->port_mutex
)
4446 struct dp_netdev_pmd_thread
*non_pmd
;
4448 non_pmd
= xzalloc(sizeof *non_pmd
);
4449 dp_netdev_configure_pmd(non_pmd
, dp
, NON_PMD_CORE_ID
, OVS_NUMA_UNSPEC
);
4452 /* Caller must have valid pointer to 'pmd'. */
4454 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
)
4456 return ovs_refcount_try_ref_rcu(&pmd
->ref_cnt
);
4460 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
)
4462 if (pmd
&& ovs_refcount_unref(&pmd
->ref_cnt
) == 1) {
4463 ovsrcu_postpone(dp_netdev_destroy_pmd
, pmd
);
4467 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
4468 * fails, keeps checking for next node until reaching the end of cmap.
4470 * Caller must unrefs the returned reference. */
4471 static struct dp_netdev_pmd_thread
*
4472 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
)
4474 struct dp_netdev_pmd_thread
*next
;
4477 struct cmap_node
*node
;
4479 node
= cmap_next_position(&dp
->poll_threads
, pos
);
4480 next
= node
? CONTAINER_OF(node
, struct dp_netdev_pmd_thread
, node
)
4482 } while (next
&& !dp_netdev_pmd_try_ref(next
));
4487 /* Configures the 'pmd' based on the input argument. */
4489 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
, struct dp_netdev
*dp
,
4490 unsigned core_id
, int numa_id
)
4493 pmd
->core_id
= core_id
;
4494 pmd
->numa_id
= numa_id
;
4495 pmd
->need_reload
= false;
4497 ovs_refcount_init(&pmd
->ref_cnt
);
4498 latch_init(&pmd
->exit_latch
);
4499 pmd
->reload_seq
= seq_create();
4500 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
4501 atomic_init(&pmd
->reload
, false);
4502 xpthread_cond_init(&pmd
->cond
, NULL
);
4503 ovs_mutex_init(&pmd
->cond_mutex
);
4504 ovs_mutex_init(&pmd
->flow_mutex
);
4505 ovs_mutex_init(&pmd
->port_mutex
);
4506 cmap_init(&pmd
->flow_table
);
4507 cmap_init(&pmd
->classifiers
);
4508 pmd
->next_optimization
= time_msec() + DPCLS_OPTIMIZATION_INTERVAL
;
4509 pmd
->rxq_interval
= time_msec() + PMD_RXQ_INTERVAL_LEN
;
4510 hmap_init(&pmd
->poll_list
);
4511 hmap_init(&pmd
->tx_ports
);
4512 hmap_init(&pmd
->tnl_port_cache
);
4513 hmap_init(&pmd
->send_port_cache
);
4514 /* init the 'flow_cache' since there is no
4515 * actual thread created for NON_PMD_CORE_ID. */
4516 if (core_id
== NON_PMD_CORE_ID
) {
4517 emc_cache_init(&pmd
->flow_cache
);
4518 pmd_alloc_static_tx_qid(pmd
);
4520 cmap_insert(&dp
->poll_threads
, CONST_CAST(struct cmap_node
*, &pmd
->node
),
4521 hash_int(core_id
, 0));
4525 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
)
4529 dp_netdev_pmd_flow_flush(pmd
);
4530 hmap_destroy(&pmd
->send_port_cache
);
4531 hmap_destroy(&pmd
->tnl_port_cache
);
4532 hmap_destroy(&pmd
->tx_ports
);
4533 hmap_destroy(&pmd
->poll_list
);
4534 /* All flows (including their dpcls_rules) have been deleted already */
4535 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
4537 ovsrcu_postpone(free
, cls
);
4539 cmap_destroy(&pmd
->classifiers
);
4540 cmap_destroy(&pmd
->flow_table
);
4541 ovs_mutex_destroy(&pmd
->flow_mutex
);
4542 latch_destroy(&pmd
->exit_latch
);
4543 seq_destroy(pmd
->reload_seq
);
4544 xpthread_cond_destroy(&pmd
->cond
);
4545 ovs_mutex_destroy(&pmd
->cond_mutex
);
4546 ovs_mutex_destroy(&pmd
->port_mutex
);
4550 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
4551 * and unrefs the struct. */
4553 dp_netdev_del_pmd(struct dp_netdev
*dp
, struct dp_netdev_pmd_thread
*pmd
)
4555 /* NON_PMD_CORE_ID doesn't have a thread, so we don't have to synchronize,
4556 * but extra cleanup is necessary */
4557 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
4558 ovs_mutex_lock(&dp
->non_pmd_mutex
);
4559 emc_cache_uninit(&pmd
->flow_cache
);
4560 pmd_free_cached_ports(pmd
);
4561 pmd_free_static_tx_qid(pmd
);
4562 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
4564 latch_set(&pmd
->exit_latch
);
4565 dp_netdev_reload_pmd__(pmd
);
4566 xpthread_join(pmd
->thread
, NULL
);
4569 dp_netdev_pmd_clear_ports(pmd
);
4571 /* Purges the 'pmd''s flows after stopping the thread, but before
4572 * destroying the flows, so that the flow stats can be collected. */
4573 if (dp
->dp_purge_cb
) {
4574 dp
->dp_purge_cb(dp
->dp_purge_aux
, pmd
->core_id
);
4576 cmap_remove(&pmd
->dp
->poll_threads
, &pmd
->node
, hash_int(pmd
->core_id
, 0));
4577 dp_netdev_pmd_unref(pmd
);
4580 /* Destroys all pmd threads. If 'non_pmd' is true it also destroys the non pmd
4583 dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
, bool non_pmd
)
4585 struct dp_netdev_pmd_thread
*pmd
;
4586 struct dp_netdev_pmd_thread
**pmd_list
;
4587 size_t k
= 0, n_pmds
;
4589 n_pmds
= cmap_count(&dp
->poll_threads
);
4590 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
4592 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4593 if (!non_pmd
&& pmd
->core_id
== NON_PMD_CORE_ID
) {
4596 /* We cannot call dp_netdev_del_pmd(), since it alters
4597 * 'dp->poll_threads' (while we're iterating it) and it
4599 ovs_assert(k
< n_pmds
);
4600 pmd_list
[k
++] = pmd
;
4603 for (size_t i
= 0; i
< k
; i
++) {
4604 dp_netdev_del_pmd(dp
, pmd_list
[i
]);
4609 /* Deletes all rx queues from pmd->poll_list and all the ports from
4612 dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
)
4614 struct rxq_poll
*poll
;
4615 struct tx_port
*port
;
4617 ovs_mutex_lock(&pmd
->port_mutex
);
4618 HMAP_FOR_EACH_POP (poll
, node
, &pmd
->poll_list
) {
4621 HMAP_FOR_EACH_POP (port
, node
, &pmd
->tx_ports
) {
4624 ovs_mutex_unlock(&pmd
->port_mutex
);
4627 /* Adds rx queue to poll_list of PMD thread, if it's not there already. */
4629 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
4630 struct dp_netdev_rxq
*rxq
)
4631 OVS_REQUIRES(pmd
->port_mutex
)
4633 int qid
= netdev_rxq_get_queue_id(rxq
->rx
);
4634 uint32_t hash
= hash_2words(odp_to_u32(rxq
->port
->port_no
), qid
);
4635 struct rxq_poll
*poll
;
4637 HMAP_FOR_EACH_WITH_HASH (poll
, node
, hash
, &pmd
->poll_list
) {
4638 if (poll
->rxq
== rxq
) {
4639 /* 'rxq' is already polled by this thread. Do nothing. */
4644 poll
= xmalloc(sizeof *poll
);
4646 hmap_insert(&pmd
->poll_list
, &poll
->node
, hash
);
4648 pmd
->need_reload
= true;
4651 /* Delete 'poll' from poll_list of PMD thread. */
4653 dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
4654 struct rxq_poll
*poll
)
4655 OVS_REQUIRES(pmd
->port_mutex
)
4657 hmap_remove(&pmd
->poll_list
, &poll
->node
);
4660 pmd
->need_reload
= true;
4663 /* Add 'port' to the tx port cache of 'pmd', which must be reloaded for the
4664 * changes to take effect. */
4666 dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
4667 struct dp_netdev_port
*port
)
4668 OVS_REQUIRES(pmd
->port_mutex
)
4672 tx
= tx_port_lookup(&pmd
->tx_ports
, port
->port_no
);
4674 /* 'port' is already on this thread tx cache. Do nothing. */
4678 tx
= xzalloc(sizeof *tx
);
4683 hmap_insert(&pmd
->tx_ports
, &tx
->node
, hash_port_no(tx
->port
->port_no
));
4684 pmd
->need_reload
= true;
4687 /* Del 'tx' from the tx port cache of 'pmd', which must be reloaded for the
4688 * changes to take effect. */
4690 dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
4692 OVS_REQUIRES(pmd
->port_mutex
)
4694 hmap_remove(&pmd
->tx_ports
, &tx
->node
);
4696 pmd
->need_reload
= true;
4700 dpif_netdev_get_datapath_version(void)
4702 return xstrdup("<built-in>");
4706 dp_netdev_flow_used(struct dp_netdev_flow
*netdev_flow
, int cnt
, int size
,
4707 uint16_t tcp_flags
, long long now
)
4711 atomic_store_relaxed(&netdev_flow
->stats
.used
, now
);
4712 non_atomic_ullong_add(&netdev_flow
->stats
.packet_count
, cnt
);
4713 non_atomic_ullong_add(&netdev_flow
->stats
.byte_count
, size
);
4714 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
4716 atomic_store_relaxed(&netdev_flow
->stats
.tcp_flags
, flags
);
4720 dp_netdev_count_packet(struct dp_netdev_pmd_thread
*pmd
,
4721 enum dp_stat_type type
, int cnt
)
4723 non_atomic_ullong_add(&pmd
->stats
.n
[type
], cnt
);
4727 dp_netdev_upcall(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
*packet_
,
4728 struct flow
*flow
, struct flow_wildcards
*wc
, ovs_u128
*ufid
,
4729 enum dpif_upcall_type type
, const struct nlattr
*userdata
,
4730 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
4732 struct dp_netdev
*dp
= pmd
->dp
;
4734 if (OVS_UNLIKELY(!dp
->upcall_cb
)) {
4738 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl
))) {
4739 struct ds ds
= DS_EMPTY_INITIALIZER
;
4742 struct odp_flow_key_parms odp_parms
= {
4744 .mask
= wc
? &wc
->masks
: NULL
,
4745 .support
= dp_netdev_support
,
4748 ofpbuf_init(&key
, 0);
4749 odp_flow_key_from_flow(&odp_parms
, &key
);
4750 packet_str
= ofp_dp_packet_to_string(packet_
);
4752 odp_flow_key_format(key
.data
, key
.size
, &ds
);
4754 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp
->name
,
4755 dpif_upcall_type_to_string(type
), ds_cstr(&ds
), packet_str
);
4757 ofpbuf_uninit(&key
);
4763 return dp
->upcall_cb(packet_
, flow
, ufid
, pmd
->core_id
, type
, userdata
,
4764 actions
, wc
, put_actions
, dp
->upcall_aux
);
4767 static inline uint32_t
4768 dpif_netdev_packet_get_rss_hash(struct dp_packet
*packet
,
4769 const struct miniflow
*mf
)
4771 uint32_t hash
, recirc_depth
;
4773 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
4774 hash
= dp_packet_get_rss_hash(packet
);
4776 hash
= miniflow_hash_5tuple(mf
, 0);
4777 dp_packet_set_rss_hash(packet
, hash
);
4780 /* The RSS hash must account for the recirculation depth to avoid
4781 * collisions in the exact match cache */
4782 recirc_depth
= *recirc_depth_get_unsafe();
4783 if (OVS_UNLIKELY(recirc_depth
)) {
4784 hash
= hash_finish(hash
, recirc_depth
);
4785 dp_packet_set_rss_hash(packet
, hash
);
4790 struct packet_batch_per_flow
{
4791 unsigned int byte_count
;
4793 struct dp_netdev_flow
*flow
;
4795 struct dp_packet_batch array
;
4799 packet_batch_per_flow_update(struct packet_batch_per_flow
*batch
,
4800 struct dp_packet
*packet
,
4801 const struct miniflow
*mf
)
4803 batch
->byte_count
+= dp_packet_size(packet
);
4804 batch
->tcp_flags
|= miniflow_get_tcp_flags(mf
);
4805 batch
->array
.packets
[batch
->array
.count
++] = packet
;
4809 packet_batch_per_flow_init(struct packet_batch_per_flow
*batch
,
4810 struct dp_netdev_flow
*flow
)
4812 flow
->batch
= batch
;
4815 dp_packet_batch_init(&batch
->array
);
4816 batch
->byte_count
= 0;
4817 batch
->tcp_flags
= 0;
4821 packet_batch_per_flow_execute(struct packet_batch_per_flow
*batch
,
4822 struct dp_netdev_pmd_thread
*pmd
,
4825 struct dp_netdev_actions
*actions
;
4826 struct dp_netdev_flow
*flow
= batch
->flow
;
4828 dp_netdev_flow_used(flow
, batch
->array
.count
, batch
->byte_count
,
4829 batch
->tcp_flags
, now
);
4831 actions
= dp_netdev_flow_get_actions(flow
);
4833 dp_netdev_execute_actions(pmd
, &batch
->array
, true, &flow
->flow
,
4834 actions
->actions
, actions
->size
, now
);
4838 dp_netdev_queue_batches(struct dp_packet
*pkt
,
4839 struct dp_netdev_flow
*flow
, const struct miniflow
*mf
,
4840 struct packet_batch_per_flow
*batches
,
4843 struct packet_batch_per_flow
*batch
= flow
->batch
;
4845 if (OVS_UNLIKELY(!batch
)) {
4846 batch
= &batches
[(*n_batches
)++];
4847 packet_batch_per_flow_init(batch
, flow
);
4850 packet_batch_per_flow_update(batch
, pkt
, mf
);
4853 /* Try to process all ('cnt') the 'packets' using only the exact match cache
4854 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
4855 * miniflow is copied into 'keys' and the packet pointer is moved at the
4856 * beginning of the 'packets' array.
4858 * The function returns the number of packets that needs to be processed in the
4859 * 'packets' array (they have been moved to the beginning of the vector).
4861 * If 'md_is_valid' is false, the metadata in 'packets' is not valid and must
4862 * be initialized by this function using 'port_no'.
4864 static inline size_t
4865 emc_processing(struct dp_netdev_pmd_thread
*pmd
,
4866 struct dp_packet_batch
*packets_
,
4867 struct netdev_flow_key
*keys
,
4868 struct packet_batch_per_flow batches
[], size_t *n_batches
,
4869 bool md_is_valid
, odp_port_t port_no
)
4871 struct emc_cache
*flow_cache
= &pmd
->flow_cache
;
4872 struct netdev_flow_key
*key
= &keys
[0];
4873 size_t n_missed
= 0, n_dropped
= 0;
4874 struct dp_packet
*packet
;
4875 const size_t size
= dp_packet_batch_size(packets_
);
4879 atomic_read_relaxed(&pmd
->dp
->emc_insert_min
, &cur_min
);
4881 DP_PACKET_BATCH_REFILL_FOR_EACH (i
, size
, packet
, packets_
) {
4882 struct dp_netdev_flow
*flow
;
4884 if (OVS_UNLIKELY(dp_packet_size(packet
) < ETH_HEADER_LEN
)) {
4885 dp_packet_delete(packet
);
4890 if (i
!= size
- 1) {
4891 struct dp_packet
**packets
= packets_
->packets
;
4892 /* Prefetch next packet data and metadata. */
4893 OVS_PREFETCH(dp_packet_data(packets
[i
+1]));
4894 pkt_metadata_prefetch_init(&packets
[i
+1]->md
);
4898 pkt_metadata_init(&packet
->md
, port_no
);
4900 miniflow_extract(packet
, &key
->mf
);
4901 key
->len
= 0; /* Not computed yet. */
4902 key
->hash
= dpif_netdev_packet_get_rss_hash(packet
, &key
->mf
);
4904 /* If EMC is disabled skip emc_lookup */
4905 flow
= (cur_min
== 0) ? NULL
: emc_lookup(flow_cache
, key
);
4906 if (OVS_LIKELY(flow
)) {
4907 dp_netdev_queue_batches(packet
, flow
, &key
->mf
, batches
,
4910 /* Exact match cache missed. Group missed packets together at
4911 * the beginning of the 'packets' array. */
4912 dp_packet_batch_refill(packets_
, packet
, i
);
4913 /* 'key[n_missed]' contains the key of the current packet and it
4914 * must be returned to the caller. The next key should be extracted
4915 * to 'keys[n_missed + 1]'. */
4916 key
= &keys
[++n_missed
];
4920 dp_netdev_count_packet(pmd
, DP_STAT_EXACT_HIT
,
4921 size
- n_dropped
- n_missed
);
4923 return dp_packet_batch_size(packets_
);
4927 handle_packet_upcall(struct dp_netdev_pmd_thread
*pmd
,
4928 struct dp_packet
*packet
,
4929 const struct netdev_flow_key
*key
,
4930 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
,
4931 int *lost_cnt
, long long now
)
4933 struct ofpbuf
*add_actions
;
4934 struct dp_packet_batch b
;
4939 match
.tun_md
.valid
= false;
4940 miniflow_expand(&key
->mf
, &match
.flow
);
4942 ofpbuf_clear(actions
);
4943 ofpbuf_clear(put_actions
);
4945 dpif_flow_hash(pmd
->dp
->dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
4946 error
= dp_netdev_upcall(pmd
, packet
, &match
.flow
, &match
.wc
,
4947 &ufid
, DPIF_UC_MISS
, NULL
, actions
,
4949 if (OVS_UNLIKELY(error
&& error
!= ENOSPC
)) {
4950 dp_packet_delete(packet
);
4955 /* The Netlink encoding of datapath flow keys cannot express
4956 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
4957 * tag is interpreted as exact match on the fact that there is no
4958 * VLAN. Unless we refactor a lot of code that translates between
4959 * Netlink and struct flow representations, we have to do the same
4961 if (!match
.wc
.masks
.vlans
[0].tci
) {
4962 match
.wc
.masks
.vlans
[0].tci
= htons(0xffff);
4965 /* We can't allow the packet batching in the next loop to execute
4966 * the actions. Otherwise, if there are any slow path actions,
4967 * we'll send the packet up twice. */
4968 dp_packet_batch_init_packet(&b
, packet
);
4969 dp_netdev_execute_actions(pmd
, &b
, true, &match
.flow
,
4970 actions
->data
, actions
->size
, now
);
4972 add_actions
= put_actions
->size
? put_actions
: actions
;
4973 if (OVS_LIKELY(error
!= ENOSPC
)) {
4974 struct dp_netdev_flow
*netdev_flow
;
4976 /* XXX: There's a race window where a flow covering this packet
4977 * could have already been installed since we last did the flow
4978 * lookup before upcall. This could be solved by moving the
4979 * mutex lock outside the loop, but that's an awful long time
4980 * to be locking everyone out of making flow installs. If we
4981 * move to a per-core classifier, it would be reasonable. */
4982 ovs_mutex_lock(&pmd
->flow_mutex
);
4983 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
4984 if (OVS_LIKELY(!netdev_flow
)) {
4985 netdev_flow
= dp_netdev_flow_add(pmd
, &match
, &ufid
,
4989 ovs_mutex_unlock(&pmd
->flow_mutex
);
4990 emc_probabilistic_insert(pmd
, key
, netdev_flow
);
4995 fast_path_processing(struct dp_netdev_pmd_thread
*pmd
,
4996 struct dp_packet_batch
*packets_
,
4997 struct netdev_flow_key
*keys
,
4998 struct packet_batch_per_flow batches
[], size_t *n_batches
,
5002 int cnt
= packets_
->count
;
5003 #if !defined(__CHECKER__) && !defined(_WIN32)
5004 const size_t PKT_ARRAY_SIZE
= cnt
;
5006 /* Sparse or MSVC doesn't like variable length array. */
5007 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
5009 struct dp_packet
**packets
= packets_
->packets
;
5011 struct dpcls_rule
*rules
[PKT_ARRAY_SIZE
];
5012 struct dp_netdev
*dp
= pmd
->dp
;
5013 int miss_cnt
= 0, lost_cnt
= 0;
5014 int lookup_cnt
= 0, add_lookup_cnt
;
5018 for (i
= 0; i
< cnt
; i
++) {
5019 /* Key length is needed in all the cases, hash computed on demand. */
5020 keys
[i
].len
= netdev_flow_key_size(miniflow_n_values(&keys
[i
].mf
));
5022 /* Get the classifier for the in_port */
5023 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
5024 if (OVS_LIKELY(cls
)) {
5025 any_miss
= !dpcls_lookup(cls
, keys
, rules
, cnt
, &lookup_cnt
);
5028 memset(rules
, 0, sizeof(rules
));
5030 if (OVS_UNLIKELY(any_miss
) && !fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
5031 uint64_t actions_stub
[512 / 8], slow_stub
[512 / 8];
5032 struct ofpbuf actions
, put_actions
;
5034 ofpbuf_use_stub(&actions
, actions_stub
, sizeof actions_stub
);
5035 ofpbuf_use_stub(&put_actions
, slow_stub
, sizeof slow_stub
);
5037 for (i
= 0; i
< cnt
; i
++) {
5038 struct dp_netdev_flow
*netdev_flow
;
5040 if (OVS_LIKELY(rules
[i
])) {
5044 /* It's possible that an earlier slow path execution installed
5045 * a rule covering this flow. In this case, it's a lot cheaper
5046 * to catch it here than execute a miss. */
5047 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, &keys
[i
],
5050 lookup_cnt
+= add_lookup_cnt
;
5051 rules
[i
] = &netdev_flow
->cr
;
5056 handle_packet_upcall(pmd
, packets
[i
], &keys
[i
], &actions
,
5057 &put_actions
, &lost_cnt
, now
);
5060 ofpbuf_uninit(&actions
);
5061 ofpbuf_uninit(&put_actions
);
5062 fat_rwlock_unlock(&dp
->upcall_rwlock
);
5063 } else if (OVS_UNLIKELY(any_miss
)) {
5064 for (i
= 0; i
< cnt
; i
++) {
5065 if (OVS_UNLIKELY(!rules
[i
])) {
5066 dp_packet_delete(packets
[i
]);
5073 for (i
= 0; i
< cnt
; i
++) {
5074 struct dp_packet
*packet
= packets
[i
];
5075 struct dp_netdev_flow
*flow
;
5077 if (OVS_UNLIKELY(!rules
[i
])) {
5081 flow
= dp_netdev_flow_cast(rules
[i
]);
5083 emc_probabilistic_insert(pmd
, &keys
[i
], flow
);
5084 dp_netdev_queue_batches(packet
, flow
, &keys
[i
].mf
, batches
, n_batches
);
5087 dp_netdev_count_packet(pmd
, DP_STAT_MASKED_HIT
, cnt
- miss_cnt
);
5088 dp_netdev_count_packet(pmd
, DP_STAT_LOOKUP_HIT
, lookup_cnt
);
5089 dp_netdev_count_packet(pmd
, DP_STAT_MISS
, miss_cnt
);
5090 dp_netdev_count_packet(pmd
, DP_STAT_LOST
, lost_cnt
);
5093 /* Packets enter the datapath from a port (or from recirculation) here.
5095 * For performance reasons a caller may choose not to initialize the metadata
5096 * in 'packets': in this case 'mdinit' is false and this function needs to
5097 * initialize it using 'port_no'. If the metadata in 'packets' is already
5098 * valid, 'md_is_valid' must be true and 'port_no' will be ignored. */
5100 dp_netdev_input__(struct dp_netdev_pmd_thread
*pmd
,
5101 struct dp_packet_batch
*packets
,
5102 bool md_is_valid
, odp_port_t port_no
)
5104 int cnt
= packets
->count
;
5105 #if !defined(__CHECKER__) && !defined(_WIN32)
5106 const size_t PKT_ARRAY_SIZE
= cnt
;
5108 /* Sparse or MSVC doesn't like variable length array. */
5109 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
5111 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
)
5112 struct netdev_flow_key keys
[PKT_ARRAY_SIZE
];
5113 struct packet_batch_per_flow batches
[PKT_ARRAY_SIZE
];
5114 long long now
= time_msec();
5119 emc_processing(pmd
, packets
, keys
, batches
, &n_batches
,
5120 md_is_valid
, port_no
);
5121 if (!dp_packet_batch_is_empty(packets
)) {
5122 /* Get ingress port from first packet's metadata. */
5123 in_port
= packets
->packets
[0]->md
.in_port
.odp_port
;
5124 fast_path_processing(pmd
, packets
, keys
, batches
, &n_batches
,
5128 /* All the flow batches need to be reset before any call to
5129 * packet_batch_per_flow_execute() as it could potentially trigger
5130 * recirculation. When a packet matching flow ‘j’ happens to be
5131 * recirculated, the nested call to dp_netdev_input__() could potentially
5132 * classify the packet as matching another flow - say 'k'. It could happen
5133 * that in the previous call to dp_netdev_input__() that same flow 'k' had
5134 * already its own batches[k] still waiting to be served. So if its
5135 * ‘batch’ member is not reset, the recirculated packet would be wrongly
5136 * appended to batches[k] of the 1st call to dp_netdev_input__(). */
5138 for (i
= 0; i
< n_batches
; i
++) {
5139 batches
[i
].flow
->batch
= NULL
;
5142 for (i
= 0; i
< n_batches
; i
++) {
5143 packet_batch_per_flow_execute(&batches
[i
], pmd
, now
);
5148 dp_netdev_input(struct dp_netdev_pmd_thread
*pmd
,
5149 struct dp_packet_batch
*packets
,
5152 dp_netdev_input__(pmd
, packets
, false, port_no
);
5156 dp_netdev_recirculate(struct dp_netdev_pmd_thread
*pmd
,
5157 struct dp_packet_batch
*packets
)
5159 dp_netdev_input__(pmd
, packets
, true, 0);
5162 struct dp_netdev_execute_aux
{
5163 struct dp_netdev_pmd_thread
*pmd
;
5165 const struct flow
*flow
;
5169 dpif_netdev_register_dp_purge_cb(struct dpif
*dpif
, dp_purge_callback
*cb
,
5172 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5173 dp
->dp_purge_aux
= aux
;
5174 dp
->dp_purge_cb
= cb
;
5178 dpif_netdev_register_upcall_cb(struct dpif
*dpif
, upcall_callback
*cb
,
5181 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5182 dp
->upcall_aux
= aux
;
5187 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
5188 long long now
, bool purge
)
5191 struct dp_netdev_port
*port
;
5194 HMAP_FOR_EACH (tx
, node
, &pmd
->send_port_cache
) {
5195 if (!tx
->port
->dynamic_txqs
) {
5198 interval
= now
- tx
->last_used
;
5199 if (tx
->qid
>= 0 && (purge
|| interval
>= XPS_TIMEOUT_MS
)) {
5201 ovs_mutex_lock(&port
->txq_used_mutex
);
5202 port
->txq_used
[tx
->qid
]--;
5203 ovs_mutex_unlock(&port
->txq_used_mutex
);
5210 dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
5211 struct tx_port
*tx
, long long now
)
5213 struct dp_netdev_port
*port
;
5215 int i
, min_cnt
, min_qid
;
5217 if (OVS_UNLIKELY(!now
)) {
5221 interval
= now
- tx
->last_used
;
5222 tx
->last_used
= now
;
5224 if (OVS_LIKELY(tx
->qid
>= 0 && interval
< XPS_TIMEOUT_MS
)) {
5230 ovs_mutex_lock(&port
->txq_used_mutex
);
5232 port
->txq_used
[tx
->qid
]--;
5238 for (i
= 0; i
< netdev_n_txq(port
->netdev
); i
++) {
5239 if (port
->txq_used
[i
] < min_cnt
|| min_cnt
== -1) {
5240 min_cnt
= port
->txq_used
[i
];
5245 port
->txq_used
[min_qid
]++;
5248 ovs_mutex_unlock(&port
->txq_used_mutex
);
5250 dpif_netdev_xps_revalidate_pmd(pmd
, now
, false);
5252 VLOG_DBG("Core %d: New TX queue ID %d for port \'%s\'.",
5253 pmd
->core_id
, tx
->qid
, netdev_get_name(tx
->port
->netdev
));
5257 static struct tx_port
*
5258 pmd_tnl_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
5261 return tx_port_lookup(&pmd
->tnl_port_cache
, port_no
);
5264 static struct tx_port
*
5265 pmd_send_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
5268 return tx_port_lookup(&pmd
->send_port_cache
, port_no
);
5272 push_tnl_action(const struct dp_netdev_pmd_thread
*pmd
,
5273 const struct nlattr
*attr
,
5274 struct dp_packet_batch
*batch
)
5276 struct tx_port
*tun_port
;
5277 const struct ovs_action_push_tnl
*data
;
5280 data
= nl_attr_get(attr
);
5282 tun_port
= pmd_tnl_port_cache_lookup(pmd
, data
->tnl_port
);
5287 err
= netdev_push_header(tun_port
->port
->netdev
, batch
, data
);
5292 dp_packet_delete_batch(batch
, true);
5297 dp_execute_userspace_action(struct dp_netdev_pmd_thread
*pmd
,
5298 struct dp_packet
*packet
, bool may_steal
,
5299 struct flow
*flow
, ovs_u128
*ufid
,
5300 struct ofpbuf
*actions
,
5301 const struct nlattr
*userdata
, long long now
)
5303 struct dp_packet_batch b
;
5306 ofpbuf_clear(actions
);
5308 error
= dp_netdev_upcall(pmd
, packet
, flow
, NULL
, ufid
,
5309 DPIF_UC_ACTION
, userdata
, actions
,
5311 if (!error
|| error
== ENOSPC
) {
5312 dp_packet_batch_init_packet(&b
, packet
);
5313 dp_netdev_execute_actions(pmd
, &b
, may_steal
, flow
,
5314 actions
->data
, actions
->size
, now
);
5315 } else if (may_steal
) {
5316 dp_packet_delete(packet
);
5321 dp_execute_cb(void *aux_
, struct dp_packet_batch
*packets_
,
5322 const struct nlattr
*a
, bool may_steal
)
5323 OVS_NO_THREAD_SAFETY_ANALYSIS
5325 struct dp_netdev_execute_aux
*aux
= aux_
;
5326 uint32_t *depth
= recirc_depth_get();
5327 struct dp_netdev_pmd_thread
*pmd
= aux
->pmd
;
5328 struct dp_netdev
*dp
= pmd
->dp
;
5329 int type
= nl_attr_type(a
);
5330 long long now
= aux
->now
;
5333 switch ((enum ovs_action_attr
)type
) {
5334 case OVS_ACTION_ATTR_OUTPUT
:
5335 p
= pmd_send_port_cache_lookup(pmd
, nl_attr_get_odp_port(a
));
5336 if (OVS_LIKELY(p
)) {
5340 dynamic_txqs
= p
->port
->dynamic_txqs
;
5342 tx_qid
= dpif_netdev_xps_get_tx_qid(pmd
, p
, now
);
5344 tx_qid
= pmd
->static_tx_qid
;
5347 netdev_send(p
->port
->netdev
, tx_qid
, packets_
, may_steal
,
5353 case OVS_ACTION_ATTR_TUNNEL_PUSH
:
5354 if (*depth
< MAX_RECIRC_DEPTH
) {
5355 dp_packet_batch_apply_cutlen(packets_
);
5356 push_tnl_action(pmd
, a
, packets_
);
5361 case OVS_ACTION_ATTR_TUNNEL_POP
:
5362 if (*depth
< MAX_RECIRC_DEPTH
) {
5363 struct dp_packet_batch
*orig_packets_
= packets_
;
5364 odp_port_t portno
= nl_attr_get_odp_port(a
);
5366 p
= pmd_tnl_port_cache_lookup(pmd
, portno
);
5368 struct dp_packet_batch tnl_pkt
;
5371 dp_packet_batch_clone(&tnl_pkt
, packets_
);
5372 packets_
= &tnl_pkt
;
5373 dp_packet_batch_reset_cutlen(orig_packets_
);
5376 dp_packet_batch_apply_cutlen(packets_
);
5378 netdev_pop_header(p
->port
->netdev
, packets_
);
5379 if (dp_packet_batch_is_empty(packets_
)) {
5383 struct dp_packet
*packet
;
5384 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5385 packet
->md
.in_port
.odp_port
= portno
;
5389 dp_netdev_recirculate(pmd
, packets_
);
5396 case OVS_ACTION_ATTR_USERSPACE
:
5397 if (!fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
5398 struct dp_packet_batch
*orig_packets_
= packets_
;
5399 const struct nlattr
*userdata
;
5400 struct dp_packet_batch usr_pkt
;
5401 struct ofpbuf actions
;
5406 userdata
= nl_attr_find_nested(a
, OVS_USERSPACE_ATTR_USERDATA
);
5407 ofpbuf_init(&actions
, 0);
5409 if (packets_
->trunc
) {
5411 dp_packet_batch_clone(&usr_pkt
, packets_
);
5412 packets_
= &usr_pkt
;
5414 dp_packet_batch_reset_cutlen(orig_packets_
);
5417 dp_packet_batch_apply_cutlen(packets_
);
5420 struct dp_packet
*packet
;
5421 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5422 flow_extract(packet
, &flow
);
5423 dpif_flow_hash(dp
->dpif
, &flow
, sizeof flow
, &ufid
);
5424 dp_execute_userspace_action(pmd
, packet
, may_steal
, &flow
,
5425 &ufid
, &actions
, userdata
, now
);
5429 dp_packet_delete_batch(packets_
, true);
5432 ofpbuf_uninit(&actions
);
5433 fat_rwlock_unlock(&dp
->upcall_rwlock
);
5439 case OVS_ACTION_ATTR_RECIRC
:
5440 if (*depth
< MAX_RECIRC_DEPTH
) {
5441 struct dp_packet_batch recirc_pkts
;
5444 dp_packet_batch_clone(&recirc_pkts
, packets_
);
5445 packets_
= &recirc_pkts
;
5448 struct dp_packet
*packet
;
5449 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5450 packet
->md
.recirc_id
= nl_attr_get_u32(a
);
5454 dp_netdev_recirculate(pmd
, packets_
);
5460 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
5463 case OVS_ACTION_ATTR_CT
: {
5464 const struct nlattr
*b
;
5466 bool commit
= false;
5469 const char *helper
= NULL
;
5470 const uint32_t *setmark
= NULL
;
5471 const struct ovs_key_ct_labels
*setlabel
= NULL
;
5472 struct nat_action_info_t nat_action_info
;
5473 struct nat_action_info_t
*nat_action_info_ref
= NULL
;
5474 bool nat_config
= false;
5476 NL_ATTR_FOR_EACH_UNSAFE (b
, left
, nl_attr_get(a
),
5477 nl_attr_get_size(a
)) {
5478 enum ovs_ct_attr sub_type
= nl_attr_type(b
);
5481 case OVS_CT_ATTR_FORCE_COMMIT
:
5484 case OVS_CT_ATTR_COMMIT
:
5487 case OVS_CT_ATTR_ZONE
:
5488 zone
= nl_attr_get_u16(b
);
5490 case OVS_CT_ATTR_HELPER
:
5491 helper
= nl_attr_get_string(b
);
5493 case OVS_CT_ATTR_MARK
:
5494 setmark
= nl_attr_get(b
);
5496 case OVS_CT_ATTR_LABELS
:
5497 setlabel
= nl_attr_get(b
);
5499 case OVS_CT_ATTR_EVENTMASK
:
5500 /* Silently ignored, as userspace datapath does not generate
5501 * netlink events. */
5503 case OVS_CT_ATTR_NAT
: {
5504 const struct nlattr
*b_nest
;
5505 unsigned int left_nest
;
5506 bool ip_min_specified
= false;
5507 bool proto_num_min_specified
= false;
5508 bool ip_max_specified
= false;
5509 bool proto_num_max_specified
= false;
5510 memset(&nat_action_info
, 0, sizeof nat_action_info
);
5511 nat_action_info_ref
= &nat_action_info
;
5513 NL_NESTED_FOR_EACH_UNSAFE (b_nest
, left_nest
, b
) {
5514 enum ovs_nat_attr sub_type_nest
= nl_attr_type(b_nest
);
5516 switch (sub_type_nest
) {
5517 case OVS_NAT_ATTR_SRC
:
5518 case OVS_NAT_ATTR_DST
:
5520 nat_action_info
.nat_action
|=
5521 ((sub_type_nest
== OVS_NAT_ATTR_SRC
)
5522 ? NAT_ACTION_SRC
: NAT_ACTION_DST
);
5524 case OVS_NAT_ATTR_IP_MIN
:
5525 memcpy(&nat_action_info
.min_addr
,
5526 nl_attr_get(b_nest
),
5527 nl_attr_get_size(b_nest
));
5528 ip_min_specified
= true;
5530 case OVS_NAT_ATTR_IP_MAX
:
5531 memcpy(&nat_action_info
.max_addr
,
5532 nl_attr_get(b_nest
),
5533 nl_attr_get_size(b_nest
));
5534 ip_max_specified
= true;
5536 case OVS_NAT_ATTR_PROTO_MIN
:
5537 nat_action_info
.min_port
=
5538 nl_attr_get_u16(b_nest
);
5539 proto_num_min_specified
= true;
5541 case OVS_NAT_ATTR_PROTO_MAX
:
5542 nat_action_info
.max_port
=
5543 nl_attr_get_u16(b_nest
);
5544 proto_num_max_specified
= true;
5546 case OVS_NAT_ATTR_PERSISTENT
:
5547 case OVS_NAT_ATTR_PROTO_HASH
:
5548 case OVS_NAT_ATTR_PROTO_RANDOM
:
5550 case OVS_NAT_ATTR_UNSPEC
:
5551 case __OVS_NAT_ATTR_MAX
:
5556 if (ip_min_specified
&& !ip_max_specified
) {
5557 nat_action_info
.max_addr
= nat_action_info
.min_addr
;
5559 if (proto_num_min_specified
&& !proto_num_max_specified
) {
5560 nat_action_info
.max_port
= nat_action_info
.min_port
;
5562 if (proto_num_min_specified
|| proto_num_max_specified
) {
5563 if (nat_action_info
.nat_action
& NAT_ACTION_SRC
) {
5564 nat_action_info
.nat_action
|= NAT_ACTION_SRC_PORT
;
5565 } else if (nat_action_info
.nat_action
& NAT_ACTION_DST
) {
5566 nat_action_info
.nat_action
|= NAT_ACTION_DST_PORT
;
5571 case OVS_CT_ATTR_UNSPEC
:
5572 case __OVS_CT_ATTR_MAX
:
5577 /* We won't be able to function properly in this case, hence
5578 * complain loudly. */
5579 if (nat_config
&& !commit
) {
5580 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(5, 5);
5581 VLOG_WARN_RL(&rl
, "NAT specified without commit.");
5584 conntrack_execute(&dp
->conntrack
, packets_
, aux
->flow
->dl_type
, force
,
5585 commit
, zone
, setmark
, setlabel
, helper
,
5586 nat_action_info_ref
, now
);
5590 case OVS_ACTION_ATTR_METER
:
5591 dp_netdev_run_meter(pmd
->dp
, packets_
, nl_attr_get_u32(a
),
5595 case OVS_ACTION_ATTR_PUSH_VLAN
:
5596 case OVS_ACTION_ATTR_POP_VLAN
:
5597 case OVS_ACTION_ATTR_PUSH_MPLS
:
5598 case OVS_ACTION_ATTR_POP_MPLS
:
5599 case OVS_ACTION_ATTR_SET
:
5600 case OVS_ACTION_ATTR_SET_MASKED
:
5601 case OVS_ACTION_ATTR_SAMPLE
:
5602 case OVS_ACTION_ATTR_HASH
:
5603 case OVS_ACTION_ATTR_UNSPEC
:
5604 case OVS_ACTION_ATTR_TRUNC
:
5605 case OVS_ACTION_ATTR_PUSH_ETH
:
5606 case OVS_ACTION_ATTR_POP_ETH
:
5607 case OVS_ACTION_ATTR_CLONE
:
5608 case OVS_ACTION_ATTR_ENCAP_NSH
:
5609 case OVS_ACTION_ATTR_DECAP_NSH
:
5610 case __OVS_ACTION_ATTR_MAX
:
5614 dp_packet_delete_batch(packets_
, may_steal
);
5618 dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
5619 struct dp_packet_batch
*packets
,
5620 bool may_steal
, const struct flow
*flow
,
5621 const struct nlattr
*actions
, size_t actions_len
,
5624 struct dp_netdev_execute_aux aux
= { pmd
, now
, flow
};
5626 odp_execute_actions(&aux
, packets
, may_steal
, actions
,
5627 actions_len
, dp_execute_cb
);
5630 struct dp_netdev_ct_dump
{
5631 struct ct_dpif_dump_state up
;
5632 struct conntrack_dump dump
;
5633 struct conntrack
*ct
;
5634 struct dp_netdev
*dp
;
5638 dpif_netdev_ct_dump_start(struct dpif
*dpif
, struct ct_dpif_dump_state
**dump_
,
5639 const uint16_t *pzone
, int *ptot_bkts
)
5641 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5642 struct dp_netdev_ct_dump
*dump
;
5644 dump
= xzalloc(sizeof *dump
);
5646 dump
->ct
= &dp
->conntrack
;
5648 conntrack_dump_start(&dp
->conntrack
, &dump
->dump
, pzone
, ptot_bkts
);
5656 dpif_netdev_ct_dump_next(struct dpif
*dpif OVS_UNUSED
,
5657 struct ct_dpif_dump_state
*dump_
,
5658 struct ct_dpif_entry
*entry
)
5660 struct dp_netdev_ct_dump
*dump
;
5662 INIT_CONTAINER(dump
, dump_
, up
);
5664 return conntrack_dump_next(&dump
->dump
, entry
);
5668 dpif_netdev_ct_dump_done(struct dpif
*dpif OVS_UNUSED
,
5669 struct ct_dpif_dump_state
*dump_
)
5671 struct dp_netdev_ct_dump
*dump
;
5674 INIT_CONTAINER(dump
, dump_
, up
);
5676 err
= conntrack_dump_done(&dump
->dump
);
5684 dpif_netdev_ct_flush(struct dpif
*dpif
, const uint16_t *zone
)
5686 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5688 return conntrack_flush(&dp
->conntrack
, zone
);
5691 const struct dpif_class dpif_netdev_class
= {
5694 dpif_netdev_enumerate
,
5695 dpif_netdev_port_open_type
,
5698 dpif_netdev_destroy
,
5701 dpif_netdev_get_stats
,
5702 dpif_netdev_port_add
,
5703 dpif_netdev_port_del
,
5704 dpif_netdev_port_set_config
,
5705 dpif_netdev_port_query_by_number
,
5706 dpif_netdev_port_query_by_name
,
5707 NULL
, /* port_get_pid */
5708 dpif_netdev_port_dump_start
,
5709 dpif_netdev_port_dump_next
,
5710 dpif_netdev_port_dump_done
,
5711 dpif_netdev_port_poll
,
5712 dpif_netdev_port_poll_wait
,
5713 dpif_netdev_flow_flush
,
5714 dpif_netdev_flow_dump_create
,
5715 dpif_netdev_flow_dump_destroy
,
5716 dpif_netdev_flow_dump_thread_create
,
5717 dpif_netdev_flow_dump_thread_destroy
,
5718 dpif_netdev_flow_dump_next
,
5719 dpif_netdev_operate
,
5720 NULL
, /* recv_set */
5721 NULL
, /* handlers_set */
5722 dpif_netdev_set_config
,
5723 dpif_netdev_queue_to_priority
,
5725 NULL
, /* recv_wait */
5726 NULL
, /* recv_purge */
5727 dpif_netdev_register_dp_purge_cb
,
5728 dpif_netdev_register_upcall_cb
,
5729 dpif_netdev_enable_upcall
,
5730 dpif_netdev_disable_upcall
,
5731 dpif_netdev_get_datapath_version
,
5732 dpif_netdev_ct_dump_start
,
5733 dpif_netdev_ct_dump_next
,
5734 dpif_netdev_ct_dump_done
,
5735 dpif_netdev_ct_flush
,
5736 dpif_netdev_meter_get_features
,
5737 dpif_netdev_meter_set
,
5738 dpif_netdev_meter_get
,
5739 dpif_netdev_meter_del
,
5743 dpif_dummy_change_port_number(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
5744 const char *argv
[], void *aux OVS_UNUSED
)
5746 struct dp_netdev_port
*port
;
5747 struct dp_netdev
*dp
;
5750 ovs_mutex_lock(&dp_netdev_mutex
);
5751 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
5752 if (!dp
|| !dpif_netdev_class_is_dummy(dp
->class)) {
5753 ovs_mutex_unlock(&dp_netdev_mutex
);
5754 unixctl_command_reply_error(conn
, "unknown datapath or not a dummy");
5757 ovs_refcount_ref(&dp
->ref_cnt
);
5758 ovs_mutex_unlock(&dp_netdev_mutex
);
5760 ovs_mutex_lock(&dp
->port_mutex
);
5761 if (get_port_by_name(dp
, argv
[2], &port
)) {
5762 unixctl_command_reply_error(conn
, "unknown port");
5766 port_no
= u32_to_odp(atoi(argv
[3]));
5767 if (!port_no
|| port_no
== ODPP_NONE
) {
5768 unixctl_command_reply_error(conn
, "bad port number");
5771 if (dp_netdev_lookup_port(dp
, port_no
)) {
5772 unixctl_command_reply_error(conn
, "port number already in use");
5777 hmap_remove(&dp
->ports
, &port
->node
);
5778 reconfigure_datapath(dp
);
5780 /* Reinsert with new port number. */
5781 port
->port_no
= port_no
;
5782 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
5783 reconfigure_datapath(dp
);
5785 seq_change(dp
->port_seq
);
5786 unixctl_command_reply(conn
, NULL
);
5789 ovs_mutex_unlock(&dp
->port_mutex
);
5790 dp_netdev_unref(dp
);
5794 dpif_dummy_register__(const char *type
)
5796 struct dpif_class
*class;
5798 class = xmalloc(sizeof *class);
5799 *class = dpif_netdev_class
;
5800 class->type
= xstrdup(type
);
5801 dp_register_provider(class);
5805 dpif_dummy_override(const char *type
)
5810 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
5811 * a userland-only build. It's useful for testsuite.
5813 error
= dp_unregister_provider(type
);
5814 if (error
== 0 || error
== EAFNOSUPPORT
) {
5815 dpif_dummy_register__(type
);
5820 dpif_dummy_register(enum dummy_level level
)
5822 if (level
== DUMMY_OVERRIDE_ALL
) {
5827 dp_enumerate_types(&types
);
5828 SSET_FOR_EACH (type
, &types
) {
5829 dpif_dummy_override(type
);
5831 sset_destroy(&types
);
5832 } else if (level
== DUMMY_OVERRIDE_SYSTEM
) {
5833 dpif_dummy_override("system");
5836 dpif_dummy_register__("dummy");
5838 unixctl_command_register("dpif-dummy/change-port-number",
5839 "dp port new-number",
5840 3, 3, dpif_dummy_change_port_number
, NULL
);
5843 /* Datapath Classifier. */
5845 /* A set of rules that all have the same fields wildcarded. */
5846 struct dpcls_subtable
{
5847 /* The fields are only used by writers. */
5848 struct cmap_node cmap_node OVS_GUARDED
; /* Within dpcls 'subtables_map'. */
5850 /* These fields are accessed by readers. */
5851 struct cmap rules
; /* Contains "struct dpcls_rule"s. */
5852 uint32_t hit_cnt
; /* Number of match hits in subtable in current
5853 optimization interval. */
5854 struct netdev_flow_key mask
; /* Wildcards for fields (const). */
5855 /* 'mask' must be the last field, additional space is allocated here. */
5858 /* Initializes 'cls' as a classifier that initially contains no classification
5861 dpcls_init(struct dpcls
*cls
)
5863 cmap_init(&cls
->subtables_map
);
5864 pvector_init(&cls
->subtables
);
5868 dpcls_destroy_subtable(struct dpcls
*cls
, struct dpcls_subtable
*subtable
)
5870 VLOG_DBG("Destroying subtable %p for in_port %d", subtable
, cls
->in_port
);
5871 pvector_remove(&cls
->subtables
, subtable
);
5872 cmap_remove(&cls
->subtables_map
, &subtable
->cmap_node
,
5873 subtable
->mask
.hash
);
5874 cmap_destroy(&subtable
->rules
);
5875 ovsrcu_postpone(free
, subtable
);
5878 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
5879 * caller's responsibility.
5880 * May only be called after all the readers have been terminated. */
5882 dpcls_destroy(struct dpcls
*cls
)
5885 struct dpcls_subtable
*subtable
;
5887 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
5888 ovs_assert(cmap_count(&subtable
->rules
) == 0);
5889 dpcls_destroy_subtable(cls
, subtable
);
5891 cmap_destroy(&cls
->subtables_map
);
5892 pvector_destroy(&cls
->subtables
);
5896 static struct dpcls_subtable
*
5897 dpcls_create_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
5899 struct dpcls_subtable
*subtable
;
5901 /* Need to add one. */
5902 subtable
= xmalloc(sizeof *subtable
5903 - sizeof subtable
->mask
.mf
+ mask
->len
);
5904 cmap_init(&subtable
->rules
);
5905 subtable
->hit_cnt
= 0;
5906 netdev_flow_key_clone(&subtable
->mask
, mask
);
5907 cmap_insert(&cls
->subtables_map
, &subtable
->cmap_node
, mask
->hash
);
5908 /* Add the new subtable at the end of the pvector (with no hits yet) */
5909 pvector_insert(&cls
->subtables
, subtable
, 0);
5910 VLOG_DBG("Creating %"PRIuSIZE
". subtable %p for in_port %d",
5911 cmap_count(&cls
->subtables_map
), subtable
, cls
->in_port
);
5912 pvector_publish(&cls
->subtables
);
5917 static inline struct dpcls_subtable
*
5918 dpcls_find_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
5920 struct dpcls_subtable
*subtable
;
5922 CMAP_FOR_EACH_WITH_HASH (subtable
, cmap_node
, mask
->hash
,
5923 &cls
->subtables_map
) {
5924 if (netdev_flow_key_equal(&subtable
->mask
, mask
)) {
5928 return dpcls_create_subtable(cls
, mask
);
5932 /* Periodically sort the dpcls subtable vectors according to hit counts */
5934 dpcls_sort_subtable_vector(struct dpcls
*cls
)
5936 struct pvector
*pvec
= &cls
->subtables
;
5937 struct dpcls_subtable
*subtable
;
5939 PVECTOR_FOR_EACH (subtable
, pvec
) {
5940 pvector_change_priority(pvec
, subtable
, subtable
->hit_cnt
);
5941 subtable
->hit_cnt
= 0;
5943 pvector_publish(pvec
);
5947 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
,
5948 struct polled_queue
*poll_list
, int poll_cnt
)
5951 long long int now
= time_msec();
5953 if (now
> pmd
->rxq_interval
) {
5954 /* Get the cycles that were used to process each queue and store. */
5955 for (unsigned i
= 0; i
< poll_cnt
; i
++) {
5956 uint64_t rxq_cyc_curr
= dp_netdev_rxq_get_cycles(poll_list
[i
].rxq
,
5957 RXQ_CYCLES_PROC_CURR
);
5958 dp_netdev_rxq_set_intrvl_cycles(poll_list
[i
].rxq
, rxq_cyc_curr
);
5959 dp_netdev_rxq_set_cycles(poll_list
[i
].rxq
, RXQ_CYCLES_PROC_CURR
,
5962 /* Start new measuring interval */
5963 pmd
->rxq_interval
= now
+ PMD_RXQ_INTERVAL_LEN
;
5966 if (now
> pmd
->next_optimization
) {
5967 /* Try to obtain the flow lock to block out revalidator threads.
5968 * If not possible, just try next time. */
5969 if (!ovs_mutex_trylock(&pmd
->flow_mutex
)) {
5970 /* Optimize each classifier */
5971 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
5972 dpcls_sort_subtable_vector(cls
);
5974 ovs_mutex_unlock(&pmd
->flow_mutex
);
5975 /* Start new measuring interval */
5976 pmd
->next_optimization
= now
+ DPCLS_OPTIMIZATION_INTERVAL
;
5981 /* Insert 'rule' into 'cls'. */
5983 dpcls_insert(struct dpcls
*cls
, struct dpcls_rule
*rule
,
5984 const struct netdev_flow_key
*mask
)
5986 struct dpcls_subtable
*subtable
= dpcls_find_subtable(cls
, mask
);
5988 /* Refer to subtable's mask, also for later removal. */
5989 rule
->mask
= &subtable
->mask
;
5990 cmap_insert(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
);
5993 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
5995 dpcls_remove(struct dpcls
*cls
, struct dpcls_rule
*rule
)
5997 struct dpcls_subtable
*subtable
;
5999 ovs_assert(rule
->mask
);
6001 /* Get subtable from reference in rule->mask. */
6002 INIT_CONTAINER(subtable
, rule
->mask
, mask
);
6003 if (cmap_remove(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
)
6005 /* Delete empty subtable. */
6006 dpcls_destroy_subtable(cls
, subtable
);
6007 pvector_publish(&cls
->subtables
);
6011 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
6012 * in 'mask' the values in 'key' and 'target' are the same. */
6014 dpcls_rule_matches_key(const struct dpcls_rule
*rule
,
6015 const struct netdev_flow_key
*target
)
6017 const uint64_t *keyp
= miniflow_get_values(&rule
->flow
.mf
);
6018 const uint64_t *maskp
= miniflow_get_values(&rule
->mask
->mf
);
6021 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, target
, rule
->flow
.mf
.map
) {
6022 if (OVS_UNLIKELY((value
& *maskp
++) != *keyp
++)) {
6029 /* For each miniflow in 'keys' performs a classifier lookup writing the result
6030 * into the corresponding slot in 'rules'. If a particular entry in 'keys' is
6031 * NULL it is skipped.
6033 * This function is optimized for use in the userspace datapath and therefore
6034 * does not implement a lot of features available in the standard
6035 * classifier_lookup() function. Specifically, it does not implement
6036 * priorities, instead returning any rule which matches the flow.
6038 * Returns true if all miniflows found a corresponding rule. */
6040 dpcls_lookup(struct dpcls
*cls
, const struct netdev_flow_key keys
[],
6041 struct dpcls_rule
**rules
, const size_t cnt
,
6044 /* The received 'cnt' miniflows are the search-keys that will be processed
6045 * to find a matching entry into the available subtables.
6046 * The number of bits in map_type is equal to NETDEV_MAX_BURST. */
6047 typedef uint32_t map_type
;
6048 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
6049 BUILD_ASSERT_DECL(MAP_BITS
>= NETDEV_MAX_BURST
);
6051 struct dpcls_subtable
*subtable
;
6053 map_type keys_map
= TYPE_MAXIMUM(map_type
); /* Set all bits. */
6055 uint32_t hashes
[MAP_BITS
];
6056 const struct cmap_node
*nodes
[MAP_BITS
];
6058 if (cnt
!= MAP_BITS
) {
6059 keys_map
>>= MAP_BITS
- cnt
; /* Clear extra bits. */
6061 memset(rules
, 0, cnt
* sizeof *rules
);
6063 int lookups_match
= 0, subtable_pos
= 1;
6065 /* The Datapath classifier - aka dpcls - is composed of subtables.
6066 * Subtables are dynamically created as needed when new rules are inserted.
6067 * Each subtable collects rules with matches on a specific subset of packet
6068 * fields as defined by the subtable's mask. We proceed to process every
6069 * search-key against each subtable, but when a match is found for a
6070 * search-key, the search for that key can stop because the rules are
6071 * non-overlapping. */
6072 PVECTOR_FOR_EACH (subtable
, &cls
->subtables
) {
6075 /* Compute hashes for the remaining keys. Each search-key is
6076 * masked with the subtable's mask to avoid hashing the wildcarded
6078 ULLONG_FOR_EACH_1(i
, keys_map
) {
6079 hashes
[i
] = netdev_flow_key_hash_in_mask(&keys
[i
],
6083 found_map
= cmap_find_batch(&subtable
->rules
, keys_map
, hashes
, nodes
);
6084 /* Check results. When the i-th bit of found_map is set, it means
6085 * that a set of nodes with a matching hash value was found for the
6086 * i-th search-key. Due to possible hash collisions we need to check
6087 * which of the found rules, if any, really matches our masked
6089 ULLONG_FOR_EACH_1(i
, found_map
) {
6090 struct dpcls_rule
*rule
;
6092 CMAP_NODE_FOR_EACH (rule
, cmap_node
, nodes
[i
]) {
6093 if (OVS_LIKELY(dpcls_rule_matches_key(rule
, &keys
[i
]))) {
6095 /* Even at 20 Mpps the 32-bit hit_cnt cannot wrap
6096 * within one second optimization interval. */
6097 subtable
->hit_cnt
++;
6098 lookups_match
+= subtable_pos
;
6102 /* None of the found rules was a match. Reset the i-th bit to
6103 * keep searching this key in the next subtable. */
6104 ULLONG_SET0(found_map
, i
); /* Did not match. */
6106 ; /* Keep Sparse happy. */
6108 keys_map
&= ~found_map
; /* Clear the found rules. */
6110 if (num_lookups_p
) {
6111 *num_lookups_p
= lookups_match
;
6113 return true; /* All found. */
6117 if (num_lookups_p
) {
6118 *num_lookups_p
= lookups_match
;
6120 return false; /* Some misses. */