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
;
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
];
817 /* Sum of all the matched and not matched packets gives the total. */
818 total_packets
= stats
[DP_STAT_EXACT_HIT
] + stats
[DP_STAT_MASKED_HIT
]
819 + stats
[DP_STAT_MISS
];
821 for (i
= 0; i
< PMD_N_CYCLES
; i
++) {
822 if (cycles
[i
] > pmd
->cycles_zero
[i
]) {
823 cycles
[i
] -= pmd
->cycles_zero
[i
];
828 total_cycles
+= cycles
[i
];
831 ds_put_cstr(reply
, (pmd
->core_id
== NON_PMD_CORE_ID
)
832 ? "main thread" : "pmd thread");
834 if (pmd
->numa_id
!= OVS_NUMA_UNSPEC
) {
835 ds_put_format(reply
, " numa_id %d", pmd
->numa_id
);
837 if (pmd
->core_id
!= OVS_CORE_UNSPEC
&& pmd
->core_id
!= NON_PMD_CORE_ID
) {
838 ds_put_format(reply
, " core_id %u", pmd
->core_id
);
840 ds_put_cstr(reply
, ":\n");
843 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
844 "\tavg. subtable lookups per hit:%.2f\n"
845 "\tmiss:%llu\n\tlost:%llu\n",
846 stats
[DP_STAT_EXACT_HIT
], stats
[DP_STAT_MASKED_HIT
],
847 stats
[DP_STAT_MASKED_HIT
] > 0
848 ? (1.0*stats
[DP_STAT_LOOKUP_HIT
])/stats
[DP_STAT_MASKED_HIT
]
850 stats
[DP_STAT_MISS
], stats
[DP_STAT_LOST
]);
852 if (total_cycles
== 0) {
857 "\tidle cycles:%"PRIu64
" (%.02f%%)\n"
858 "\tprocessing cycles:%"PRIu64
" (%.02f%%)\n",
859 cycles
[PMD_CYCLES_IDLE
],
860 cycles
[PMD_CYCLES_IDLE
] / (double)total_cycles
* 100,
861 cycles
[PMD_CYCLES_PROCESSING
],
862 cycles
[PMD_CYCLES_PROCESSING
] / (double)total_cycles
* 100);
864 if (total_packets
== 0) {
869 "\tavg cycles per packet: %.02f (%"PRIu64
"/%llu)\n",
870 total_cycles
/ (double)total_packets
,
871 total_cycles
, total_packets
);
874 "\tavg processing cycles per packet: "
875 "%.02f (%"PRIu64
"/%llu)\n",
876 cycles
[PMD_CYCLES_PROCESSING
] / (double)total_packets
,
877 cycles
[PMD_CYCLES_PROCESSING
], total_packets
);
881 pmd_info_clear_stats(struct ds
*reply OVS_UNUSED
,
882 struct dp_netdev_pmd_thread
*pmd
,
883 unsigned long long stats
[DP_N_STATS
],
884 uint64_t cycles
[PMD_N_CYCLES
])
888 /* We cannot write 'stats' and 'cycles' (because they're written by other
889 * threads) and we shouldn't change 'stats' (because they're used to count
890 * datapath stats, which must not be cleared here). Instead, we save the
891 * current values and subtract them from the values to be displayed in the
893 for (i
= 0; i
< DP_N_STATS
; i
++) {
894 pmd
->stats_zero
[i
] = stats
[i
];
896 for (i
= 0; i
< PMD_N_CYCLES
; i
++) {
897 pmd
->cycles_zero
[i
] = cycles
[i
];
902 compare_poll_list(const void *a_
, const void *b_
)
904 const struct rxq_poll
*a
= a_
;
905 const struct rxq_poll
*b
= b_
;
907 const char *namea
= netdev_rxq_get_name(a
->rxq
->rx
);
908 const char *nameb
= netdev_rxq_get_name(b
->rxq
->rx
);
910 int cmp
= strcmp(namea
, nameb
);
912 return netdev_rxq_get_queue_id(a
->rxq
->rx
)
913 - netdev_rxq_get_queue_id(b
->rxq
->rx
);
920 sorted_poll_list(struct dp_netdev_pmd_thread
*pmd
, struct rxq_poll
**list
,
923 struct rxq_poll
*ret
, *poll
;
926 *n
= hmap_count(&pmd
->poll_list
);
930 ret
= xcalloc(*n
, sizeof *ret
);
932 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
937 qsort(ret
, *n
, sizeof *ret
, compare_poll_list
);
944 pmd_info_show_rxq(struct ds
*reply
, struct dp_netdev_pmd_thread
*pmd
)
946 if (pmd
->core_id
!= NON_PMD_CORE_ID
) {
947 const char *prev_name
= NULL
;
948 struct rxq_poll
*list
;
952 "pmd thread numa_id %d core_id %u:\n\tisolated : %s\n",
953 pmd
->numa_id
, pmd
->core_id
, (pmd
->isolated
)
956 ovs_mutex_lock(&pmd
->port_mutex
);
957 sorted_poll_list(pmd
, &list
, &n
);
958 for (i
= 0; i
< n
; i
++) {
959 const char *name
= netdev_rxq_get_name(list
[i
].rxq
->rx
);
961 if (!prev_name
|| strcmp(name
, prev_name
)) {
963 ds_put_cstr(reply
, "\n");
965 ds_put_format(reply
, "\tport: %s\tqueue-id:", name
);
967 ds_put_format(reply
, " %d",
968 netdev_rxq_get_queue_id(list
[i
].rxq
->rx
));
971 ovs_mutex_unlock(&pmd
->port_mutex
);
972 ds_put_cstr(reply
, "\n");
978 compare_poll_thread_list(const void *a_
, const void *b_
)
980 const struct dp_netdev_pmd_thread
*a
, *b
;
982 a
= *(struct dp_netdev_pmd_thread
**)a_
;
983 b
= *(struct dp_netdev_pmd_thread
**)b_
;
985 if (a
->core_id
< b
->core_id
) {
988 if (a
->core_id
> b
->core_id
) {
994 /* Create a sorted list of pmd's from the dp->poll_threads cmap. We can use
995 * this list, as long as we do not go to quiescent state. */
997 sorted_poll_thread_list(struct dp_netdev
*dp
,
998 struct dp_netdev_pmd_thread
***list
,
1001 struct dp_netdev_pmd_thread
*pmd
;
1002 struct dp_netdev_pmd_thread
**pmd_list
;
1003 size_t k
= 0, n_pmds
;
1005 n_pmds
= cmap_count(&dp
->poll_threads
);
1006 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
1008 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1012 pmd_list
[k
++] = pmd
;
1015 qsort(pmd_list
, k
, sizeof *pmd_list
, compare_poll_thread_list
);
1022 dpif_netdev_pmd_rebalance(struct unixctl_conn
*conn
, int argc
,
1023 const char *argv
[], void *aux OVS_UNUSED
)
1025 struct ds reply
= DS_EMPTY_INITIALIZER
;
1026 struct dp_netdev
*dp
= NULL
;
1028 ovs_mutex_lock(&dp_netdev_mutex
);
1031 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
1032 } else if (shash_count(&dp_netdevs
) == 1) {
1033 /* There's only one datapath */
1034 dp
= shash_first(&dp_netdevs
)->data
;
1038 ovs_mutex_unlock(&dp_netdev_mutex
);
1039 unixctl_command_reply_error(conn
,
1040 "please specify an existing datapath");
1044 dp_netdev_request_reconfigure(dp
);
1045 ovs_mutex_unlock(&dp_netdev_mutex
);
1046 ds_put_cstr(&reply
, "pmd rxq rebalance requested.\n");
1047 unixctl_command_reply(conn
, ds_cstr(&reply
));
1052 dpif_netdev_pmd_info(struct unixctl_conn
*conn
, int argc
, const char *argv
[],
1055 struct ds reply
= DS_EMPTY_INITIALIZER
;
1056 struct dp_netdev_pmd_thread
**pmd_list
;
1057 struct dp_netdev
*dp
= NULL
;
1059 enum pmd_info_type type
= *(enum pmd_info_type
*) aux
;
1061 ovs_mutex_lock(&dp_netdev_mutex
);
1064 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
1065 } else if (shash_count(&dp_netdevs
) == 1) {
1066 /* There's only one datapath */
1067 dp
= shash_first(&dp_netdevs
)->data
;
1071 ovs_mutex_unlock(&dp_netdev_mutex
);
1072 unixctl_command_reply_error(conn
,
1073 "please specify an existing datapath");
1077 sorted_poll_thread_list(dp
, &pmd_list
, &n
);
1078 for (size_t i
= 0; i
< n
; i
++) {
1079 struct dp_netdev_pmd_thread
*pmd
= pmd_list
[i
];
1084 if (type
== PMD_INFO_SHOW_RXQ
) {
1085 pmd_info_show_rxq(&reply
, pmd
);
1087 unsigned long long stats
[DP_N_STATS
];
1088 uint64_t cycles
[PMD_N_CYCLES
];
1090 /* Read current stats and cycle counters */
1091 for (size_t j
= 0; j
< ARRAY_SIZE(stats
); j
++) {
1092 atomic_read_relaxed(&pmd
->stats
.n
[j
], &stats
[j
]);
1094 for (size_t j
= 0; j
< ARRAY_SIZE(cycles
); j
++) {
1095 atomic_read_relaxed(&pmd
->cycles
.n
[j
], &cycles
[j
]);
1098 if (type
== PMD_INFO_CLEAR_STATS
) {
1099 pmd_info_clear_stats(&reply
, pmd
, stats
, cycles
);
1100 } else if (type
== PMD_INFO_SHOW_STATS
) {
1101 pmd_info_show_stats(&reply
, pmd
, stats
, cycles
);
1107 ovs_mutex_unlock(&dp_netdev_mutex
);
1109 unixctl_command_reply(conn
, ds_cstr(&reply
));
1114 dpif_netdev_init(void)
1116 static enum pmd_info_type show_aux
= PMD_INFO_SHOW_STATS
,
1117 clear_aux
= PMD_INFO_CLEAR_STATS
,
1118 poll_aux
= PMD_INFO_SHOW_RXQ
;
1120 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
1121 0, 1, dpif_netdev_pmd_info
,
1123 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
1124 0, 1, dpif_netdev_pmd_info
,
1125 (void *)&clear_aux
);
1126 unixctl_command_register("dpif-netdev/pmd-rxq-show", "[dp]",
1127 0, 1, dpif_netdev_pmd_info
,
1129 unixctl_command_register("dpif-netdev/pmd-rxq-rebalance", "[dp]",
1130 0, 1, dpif_netdev_pmd_rebalance
,
1136 dpif_netdev_enumerate(struct sset
*all_dps
,
1137 const struct dpif_class
*dpif_class
)
1139 struct shash_node
*node
;
1141 ovs_mutex_lock(&dp_netdev_mutex
);
1142 SHASH_FOR_EACH(node
, &dp_netdevs
) {
1143 struct dp_netdev
*dp
= node
->data
;
1144 if (dpif_class
!= dp
->class) {
1145 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
1146 * If the class doesn't match, skip this dpif. */
1149 sset_add(all_dps
, node
->name
);
1151 ovs_mutex_unlock(&dp_netdev_mutex
);
1157 dpif_netdev_class_is_dummy(const struct dpif_class
*class)
1159 return class != &dpif_netdev_class
;
1163 dpif_netdev_port_open_type(const struct dpif_class
*class, const char *type
)
1165 return strcmp(type
, "internal") ? type
1166 : dpif_netdev_class_is_dummy(class) ? "dummy-internal"
1170 static struct dpif
*
1171 create_dpif_netdev(struct dp_netdev
*dp
)
1173 uint16_t netflow_id
= hash_string(dp
->name
, 0);
1174 struct dpif_netdev
*dpif
;
1176 ovs_refcount_ref(&dp
->ref_cnt
);
1178 dpif
= xmalloc(sizeof *dpif
);
1179 dpif_init(&dpif
->dpif
, dp
->class, dp
->name
, netflow_id
>> 8, netflow_id
);
1181 dpif
->last_port_seq
= seq_read(dp
->port_seq
);
1186 /* Choose an unused, non-zero port number and return it on success.
1187 * Return ODPP_NONE on failure. */
1189 choose_port(struct dp_netdev
*dp
, const char *name
)
1190 OVS_REQUIRES(dp
->port_mutex
)
1194 if (dp
->class != &dpif_netdev_class
) {
1198 /* If the port name begins with "br", start the number search at
1199 * 100 to make writing tests easier. */
1200 if (!strncmp(name
, "br", 2)) {
1204 /* If the port name contains a number, try to assign that port number.
1205 * This can make writing unit tests easier because port numbers are
1207 for (p
= name
; *p
!= '\0'; p
++) {
1208 if (isdigit((unsigned char) *p
)) {
1209 port_no
= start_no
+ strtol(p
, NULL
, 10);
1210 if (port_no
> 0 && port_no
!= odp_to_u32(ODPP_NONE
)
1211 && !dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
1212 return u32_to_odp(port_no
);
1219 for (port_no
= 1; port_no
<= UINT16_MAX
; port_no
++) {
1220 if (!dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
1221 return u32_to_odp(port_no
);
1229 create_dp_netdev(const char *name
, const struct dpif_class
*class,
1230 struct dp_netdev
**dpp
)
1231 OVS_REQUIRES(dp_netdev_mutex
)
1233 struct dp_netdev
*dp
;
1236 dp
= xzalloc(sizeof *dp
);
1237 shash_add(&dp_netdevs
, name
, dp
);
1239 *CONST_CAST(const struct dpif_class
**, &dp
->class) = class;
1240 *CONST_CAST(const char **, &dp
->name
) = xstrdup(name
);
1241 ovs_refcount_init(&dp
->ref_cnt
);
1242 atomic_flag_clear(&dp
->destroyed
);
1244 ovs_mutex_init(&dp
->port_mutex
);
1245 hmap_init(&dp
->ports
);
1246 dp
->port_seq
= seq_create();
1247 fat_rwlock_init(&dp
->upcall_rwlock
);
1249 dp
->reconfigure_seq
= seq_create();
1250 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
1252 for (int i
= 0; i
< N_METER_LOCKS
; ++i
) {
1253 ovs_mutex_init_adaptive(&dp
->meter_locks
[i
]);
1256 /* Disable upcalls by default. */
1257 dp_netdev_disable_upcall(dp
);
1258 dp
->upcall_aux
= NULL
;
1259 dp
->upcall_cb
= NULL
;
1261 conntrack_init(&dp
->conntrack
);
1263 atomic_init(&dp
->emc_insert_min
, DEFAULT_EM_FLOW_INSERT_MIN
);
1265 cmap_init(&dp
->poll_threads
);
1267 ovs_mutex_init(&dp
->tx_qid_pool_mutex
);
1268 /* We need 1 Tx queue for each possible core + 1 for non-PMD threads. */
1269 dp
->tx_qid_pool
= id_pool_create(0, ovs_numa_get_n_cores() + 1);
1271 ovs_mutex_init_recursive(&dp
->non_pmd_mutex
);
1272 ovsthread_key_create(&dp
->per_pmd_key
, NULL
);
1274 ovs_mutex_lock(&dp
->port_mutex
);
1275 /* non-PMD will be created before all other threads and will
1276 * allocate static_tx_qid = 0. */
1277 dp_netdev_set_nonpmd(dp
);
1279 error
= do_add_port(dp
, name
, dpif_netdev_port_open_type(dp
->class,
1282 ovs_mutex_unlock(&dp
->port_mutex
);
1288 dp
->last_tnl_conf_seq
= seq_read(tnl_conf_seq
);
1294 dp_netdev_request_reconfigure(struct dp_netdev
*dp
)
1296 seq_change(dp
->reconfigure_seq
);
1300 dp_netdev_is_reconf_required(struct dp_netdev
*dp
)
1302 return seq_read(dp
->reconfigure_seq
) != dp
->last_reconfigure_seq
;
1306 dpif_netdev_open(const struct dpif_class
*class, const char *name
,
1307 bool create
, struct dpif
**dpifp
)
1309 struct dp_netdev
*dp
;
1312 ovs_mutex_lock(&dp_netdev_mutex
);
1313 dp
= shash_find_data(&dp_netdevs
, name
);
1315 error
= create
? create_dp_netdev(name
, class, &dp
) : ENODEV
;
1317 error
= (dp
->class != class ? EINVAL
1322 *dpifp
= create_dpif_netdev(dp
);
1325 ovs_mutex_unlock(&dp_netdev_mutex
);
1331 dp_netdev_destroy_upcall_lock(struct dp_netdev
*dp
)
1332 OVS_NO_THREAD_SAFETY_ANALYSIS
1334 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
1335 ovs_assert(fat_rwlock_tryrdlock(&dp
->upcall_rwlock
));
1337 /* Before freeing a lock we should release it */
1338 fat_rwlock_unlock(&dp
->upcall_rwlock
);
1339 fat_rwlock_destroy(&dp
->upcall_rwlock
);
1343 dp_delete_meter(struct dp_netdev
*dp
, uint32_t meter_id
)
1344 OVS_REQUIRES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
1346 if (dp
->meters
[meter_id
]) {
1347 free(dp
->meters
[meter_id
]);
1348 dp
->meters
[meter_id
] = NULL
;
1352 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
1353 * through the 'dp_netdevs' shash while freeing 'dp'. */
1355 dp_netdev_free(struct dp_netdev
*dp
)
1356 OVS_REQUIRES(dp_netdev_mutex
)
1358 struct dp_netdev_port
*port
, *next
;
1360 shash_find_and_delete(&dp_netdevs
, dp
->name
);
1362 ovs_mutex_lock(&dp
->port_mutex
);
1363 HMAP_FOR_EACH_SAFE (port
, next
, node
, &dp
->ports
) {
1364 do_del_port(dp
, port
);
1366 ovs_mutex_unlock(&dp
->port_mutex
);
1368 dp_netdev_destroy_all_pmds(dp
, true);
1369 cmap_destroy(&dp
->poll_threads
);
1371 ovs_mutex_destroy(&dp
->tx_qid_pool_mutex
);
1372 id_pool_destroy(dp
->tx_qid_pool
);
1374 ovs_mutex_destroy(&dp
->non_pmd_mutex
);
1375 ovsthread_key_delete(dp
->per_pmd_key
);
1377 conntrack_destroy(&dp
->conntrack
);
1380 seq_destroy(dp
->reconfigure_seq
);
1382 seq_destroy(dp
->port_seq
);
1383 hmap_destroy(&dp
->ports
);
1384 ovs_mutex_destroy(&dp
->port_mutex
);
1386 /* Upcalls must be disabled at this point */
1387 dp_netdev_destroy_upcall_lock(dp
);
1391 for (i
= 0; i
< MAX_METERS
; ++i
) {
1393 dp_delete_meter(dp
, i
);
1394 meter_unlock(dp
, i
);
1396 for (i
= 0; i
< N_METER_LOCKS
; ++i
) {
1397 ovs_mutex_destroy(&dp
->meter_locks
[i
]);
1400 free(dp
->pmd_cmask
);
1401 free(CONST_CAST(char *, dp
->name
));
1406 dp_netdev_unref(struct dp_netdev
*dp
)
1409 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
1410 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
1411 ovs_mutex_lock(&dp_netdev_mutex
);
1412 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1415 ovs_mutex_unlock(&dp_netdev_mutex
);
1420 dpif_netdev_close(struct dpif
*dpif
)
1422 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1424 dp_netdev_unref(dp
);
1429 dpif_netdev_destroy(struct dpif
*dpif
)
1431 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1433 if (!atomic_flag_test_and_set(&dp
->destroyed
)) {
1434 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1435 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
1443 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
1444 * load/store semantics. While the increment is not atomic, the load and
1445 * store operations are, making it impossible to read inconsistent values.
1447 * This is used to update thread local stats counters. */
1449 non_atomic_ullong_add(atomic_ullong
*var
, unsigned long long n
)
1451 unsigned long long tmp
;
1453 atomic_read_relaxed(var
, &tmp
);
1455 atomic_store_relaxed(var
, tmp
);
1459 dpif_netdev_get_stats(const struct dpif
*dpif
, struct dpif_dp_stats
*stats
)
1461 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1462 struct dp_netdev_pmd_thread
*pmd
;
1464 stats
->n_flows
= stats
->n_hit
= stats
->n_missed
= stats
->n_lost
= 0;
1465 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1466 unsigned long long n
;
1467 stats
->n_flows
+= cmap_count(&pmd
->flow_table
);
1469 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_MASKED_HIT
], &n
);
1471 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_EXACT_HIT
], &n
);
1473 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_MISS
], &n
);
1474 stats
->n_missed
+= n
;
1475 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_LOST
], &n
);
1478 stats
->n_masks
= UINT32_MAX
;
1479 stats
->n_mask_hit
= UINT64_MAX
;
1485 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread
*pmd
)
1487 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
1488 ovs_mutex_lock(&pmd
->dp
->non_pmd_mutex
);
1489 ovs_mutex_lock(&pmd
->port_mutex
);
1490 pmd_load_cached_ports(pmd
);
1491 ovs_mutex_unlock(&pmd
->port_mutex
);
1492 ovs_mutex_unlock(&pmd
->dp
->non_pmd_mutex
);
1496 ovs_mutex_lock(&pmd
->cond_mutex
);
1497 seq_change(pmd
->reload_seq
);
1498 atomic_store_relaxed(&pmd
->reload
, true);
1499 ovs_mutex_cond_wait(&pmd
->cond
, &pmd
->cond_mutex
);
1500 ovs_mutex_unlock(&pmd
->cond_mutex
);
1504 hash_port_no(odp_port_t port_no
)
1506 return hash_int(odp_to_u32(port_no
), 0);
1510 port_create(const char *devname
, const char *type
,
1511 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1513 struct netdev_saved_flags
*sf
;
1514 struct dp_netdev_port
*port
;
1515 enum netdev_flags flags
;
1516 struct netdev
*netdev
;
1521 /* Open and validate network device. */
1522 error
= netdev_open(devname
, type
, &netdev
);
1526 /* XXX reject non-Ethernet devices */
1528 netdev_get_flags(netdev
, &flags
);
1529 if (flags
& NETDEV_LOOPBACK
) {
1530 VLOG_ERR("%s: cannot add a loopback device", devname
);
1535 error
= netdev_turn_flags_on(netdev
, NETDEV_PROMISC
, &sf
);
1537 VLOG_ERR("%s: cannot set promisc flag", devname
);
1541 port
= xzalloc(sizeof *port
);
1542 port
->port_no
= port_no
;
1543 port
->netdev
= netdev
;
1544 port
->type
= xstrdup(type
);
1546 port
->need_reconfigure
= true;
1547 ovs_mutex_init(&port
->txq_used_mutex
);
1554 netdev_close(netdev
);
1559 do_add_port(struct dp_netdev
*dp
, const char *devname
, const char *type
,
1561 OVS_REQUIRES(dp
->port_mutex
)
1563 struct dp_netdev_port
*port
;
1566 /* Reject devices already in 'dp'. */
1567 if (!get_port_by_name(dp
, devname
, &port
)) {
1571 error
= port_create(devname
, type
, port_no
, &port
);
1576 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
1577 seq_change(dp
->port_seq
);
1579 reconfigure_datapath(dp
);
1585 dpif_netdev_port_add(struct dpif
*dpif
, struct netdev
*netdev
,
1586 odp_port_t
*port_nop
)
1588 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1589 char namebuf
[NETDEV_VPORT_NAME_BUFSIZE
];
1590 const char *dpif_port
;
1594 ovs_mutex_lock(&dp
->port_mutex
);
1595 dpif_port
= netdev_vport_get_dpif_port(netdev
, namebuf
, sizeof namebuf
);
1596 if (*port_nop
!= ODPP_NONE
) {
1597 port_no
= *port_nop
;
1598 error
= dp_netdev_lookup_port(dp
, *port_nop
) ? EBUSY
: 0;
1600 port_no
= choose_port(dp
, dpif_port
);
1601 error
= port_no
== ODPP_NONE
? EFBIG
: 0;
1604 *port_nop
= port_no
;
1605 error
= do_add_port(dp
, dpif_port
, netdev_get_type(netdev
), port_no
);
1607 ovs_mutex_unlock(&dp
->port_mutex
);
1613 dpif_netdev_port_del(struct dpif
*dpif
, odp_port_t port_no
)
1615 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1618 ovs_mutex_lock(&dp
->port_mutex
);
1619 if (port_no
== ODPP_LOCAL
) {
1622 struct dp_netdev_port
*port
;
1624 error
= get_port_by_number(dp
, port_no
, &port
);
1626 do_del_port(dp
, port
);
1629 ovs_mutex_unlock(&dp
->port_mutex
);
1635 is_valid_port_number(odp_port_t port_no
)
1637 return port_no
!= ODPP_NONE
;
1640 static struct dp_netdev_port
*
1641 dp_netdev_lookup_port(const struct dp_netdev
*dp
, odp_port_t port_no
)
1642 OVS_REQUIRES(dp
->port_mutex
)
1644 struct dp_netdev_port
*port
;
1646 HMAP_FOR_EACH_WITH_HASH (port
, node
, hash_port_no(port_no
), &dp
->ports
) {
1647 if (port
->port_no
== port_no
) {
1655 get_port_by_number(struct dp_netdev
*dp
,
1656 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1657 OVS_REQUIRES(dp
->port_mutex
)
1659 if (!is_valid_port_number(port_no
)) {
1663 *portp
= dp_netdev_lookup_port(dp
, port_no
);
1664 return *portp
? 0 : ENODEV
;
1669 port_destroy(struct dp_netdev_port
*port
)
1675 netdev_close(port
->netdev
);
1676 netdev_restore_flags(port
->sf
);
1678 for (unsigned i
= 0; i
< port
->n_rxq
; i
++) {
1679 netdev_rxq_close(port
->rxqs
[i
].rx
);
1681 ovs_mutex_destroy(&port
->txq_used_mutex
);
1682 free(port
->rxq_affinity_list
);
1683 free(port
->txq_used
);
1690 get_port_by_name(struct dp_netdev
*dp
,
1691 const char *devname
, struct dp_netdev_port
**portp
)
1692 OVS_REQUIRES(dp
->port_mutex
)
1694 struct dp_netdev_port
*port
;
1696 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1697 if (!strcmp(netdev_get_name(port
->netdev
), devname
)) {
1703 /* Callers of dpif_netdev_port_query_by_name() expect ENODEV for a non
1708 /* Returns 'true' if there is a port with pmd netdev. */
1710 has_pmd_port(struct dp_netdev
*dp
)
1711 OVS_REQUIRES(dp
->port_mutex
)
1713 struct dp_netdev_port
*port
;
1715 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1716 if (netdev_is_pmd(port
->netdev
)) {
1725 do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*port
)
1726 OVS_REQUIRES(dp
->port_mutex
)
1728 hmap_remove(&dp
->ports
, &port
->node
);
1729 seq_change(dp
->port_seq
);
1731 reconfigure_datapath(dp
);
1737 answer_port_query(const struct dp_netdev_port
*port
,
1738 struct dpif_port
*dpif_port
)
1740 dpif_port
->name
= xstrdup(netdev_get_name(port
->netdev
));
1741 dpif_port
->type
= xstrdup(port
->type
);
1742 dpif_port
->port_no
= port
->port_no
;
1746 dpif_netdev_port_query_by_number(const struct dpif
*dpif
, odp_port_t port_no
,
1747 struct dpif_port
*dpif_port
)
1749 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1750 struct dp_netdev_port
*port
;
1753 ovs_mutex_lock(&dp
->port_mutex
);
1754 error
= get_port_by_number(dp
, port_no
, &port
);
1755 if (!error
&& dpif_port
) {
1756 answer_port_query(port
, dpif_port
);
1758 ovs_mutex_unlock(&dp
->port_mutex
);
1764 dpif_netdev_port_query_by_name(const struct dpif
*dpif
, const char *devname
,
1765 struct dpif_port
*dpif_port
)
1767 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1768 struct dp_netdev_port
*port
;
1771 ovs_mutex_lock(&dp
->port_mutex
);
1772 error
= get_port_by_name(dp
, devname
, &port
);
1773 if (!error
&& dpif_port
) {
1774 answer_port_query(port
, dpif_port
);
1776 ovs_mutex_unlock(&dp
->port_mutex
);
1782 dp_netdev_flow_free(struct dp_netdev_flow
*flow
)
1784 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow
));
1788 static void dp_netdev_flow_unref(struct dp_netdev_flow
*flow
)
1790 if (ovs_refcount_unref_relaxed(&flow
->ref_cnt
) == 1) {
1791 ovsrcu_postpone(dp_netdev_flow_free
, flow
);
1796 dp_netdev_flow_hash(const ovs_u128
*ufid
)
1798 return ufid
->u32
[0];
1801 static inline struct dpcls
*
1802 dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread
*pmd
,
1806 uint32_t hash
= hash_port_no(in_port
);
1807 CMAP_FOR_EACH_WITH_HASH (cls
, node
, hash
, &pmd
->classifiers
) {
1808 if (cls
->in_port
== in_port
) {
1809 /* Port classifier exists already */
1816 static inline struct dpcls
*
1817 dp_netdev_pmd_find_dpcls(struct dp_netdev_pmd_thread
*pmd
,
1819 OVS_REQUIRES(pmd
->flow_mutex
)
1821 struct dpcls
*cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
1822 uint32_t hash
= hash_port_no(in_port
);
1825 /* Create new classifier for in_port */
1826 cls
= xmalloc(sizeof(*cls
));
1828 cls
->in_port
= in_port
;
1829 cmap_insert(&pmd
->classifiers
, &cls
->node
, hash
);
1830 VLOG_DBG("Creating dpcls %p for in_port %d", cls
, in_port
);
1836 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread
*pmd
,
1837 struct dp_netdev_flow
*flow
)
1838 OVS_REQUIRES(pmd
->flow_mutex
)
1840 struct cmap_node
*node
= CONST_CAST(struct cmap_node
*, &flow
->node
);
1842 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
1844 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
1845 ovs_assert(cls
!= NULL
);
1846 dpcls_remove(cls
, &flow
->cr
);
1847 cmap_remove(&pmd
->flow_table
, node
, dp_netdev_flow_hash(&flow
->ufid
));
1850 dp_netdev_flow_unref(flow
);
1854 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
)
1856 struct dp_netdev_flow
*netdev_flow
;
1858 ovs_mutex_lock(&pmd
->flow_mutex
);
1859 CMAP_FOR_EACH (netdev_flow
, node
, &pmd
->flow_table
) {
1860 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
1862 ovs_mutex_unlock(&pmd
->flow_mutex
);
1866 dpif_netdev_flow_flush(struct dpif
*dpif
)
1868 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1869 struct dp_netdev_pmd_thread
*pmd
;
1871 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1872 dp_netdev_pmd_flow_flush(pmd
);
1878 struct dp_netdev_port_state
{
1879 struct hmap_position position
;
1884 dpif_netdev_port_dump_start(const struct dpif
*dpif OVS_UNUSED
, void **statep
)
1886 *statep
= xzalloc(sizeof(struct dp_netdev_port_state
));
1891 dpif_netdev_port_dump_next(const struct dpif
*dpif
, void *state_
,
1892 struct dpif_port
*dpif_port
)
1894 struct dp_netdev_port_state
*state
= state_
;
1895 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1896 struct hmap_node
*node
;
1899 ovs_mutex_lock(&dp
->port_mutex
);
1900 node
= hmap_at_position(&dp
->ports
, &state
->position
);
1902 struct dp_netdev_port
*port
;
1904 port
= CONTAINER_OF(node
, struct dp_netdev_port
, node
);
1907 state
->name
= xstrdup(netdev_get_name(port
->netdev
));
1908 dpif_port
->name
= state
->name
;
1909 dpif_port
->type
= port
->type
;
1910 dpif_port
->port_no
= port
->port_no
;
1916 ovs_mutex_unlock(&dp
->port_mutex
);
1922 dpif_netdev_port_dump_done(const struct dpif
*dpif OVS_UNUSED
, void *state_
)
1924 struct dp_netdev_port_state
*state
= state_
;
1931 dpif_netdev_port_poll(const struct dpif
*dpif_
, char **devnamep OVS_UNUSED
)
1933 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
1934 uint64_t new_port_seq
;
1937 new_port_seq
= seq_read(dpif
->dp
->port_seq
);
1938 if (dpif
->last_port_seq
!= new_port_seq
) {
1939 dpif
->last_port_seq
= new_port_seq
;
1949 dpif_netdev_port_poll_wait(const struct dpif
*dpif_
)
1951 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
1953 seq_wait(dpif
->dp
->port_seq
, dpif
->last_port_seq
);
1956 static struct dp_netdev_flow
*
1957 dp_netdev_flow_cast(const struct dpcls_rule
*cr
)
1959 return cr
? CONTAINER_OF(cr
, struct dp_netdev_flow
, cr
) : NULL
;
1962 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*flow
)
1964 return ovs_refcount_try_ref_rcu(&flow
->ref_cnt
);
1967 /* netdev_flow_key utilities.
1969 * netdev_flow_key is basically a miniflow. We use these functions
1970 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1971 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1973 * - Since we are dealing exclusively with miniflows created by
1974 * miniflow_extract(), if the map is different the miniflow is different.
1975 * Therefore we can be faster by comparing the map and the miniflow in a
1977 * - These functions can be inlined by the compiler. */
1979 /* Given the number of bits set in miniflow's maps, returns the size of the
1980 * 'netdev_flow_key.mf' */
1981 static inline size_t
1982 netdev_flow_key_size(size_t flow_u64s
)
1984 return sizeof(struct miniflow
) + MINIFLOW_VALUES_SIZE(flow_u64s
);
1988 netdev_flow_key_equal(const struct netdev_flow_key
*a
,
1989 const struct netdev_flow_key
*b
)
1991 /* 'b->len' may be not set yet. */
1992 return a
->hash
== b
->hash
&& !memcmp(&a
->mf
, &b
->mf
, a
->len
);
1995 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1996 * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been
1997 * generated by miniflow_extract. */
1999 netdev_flow_key_equal_mf(const struct netdev_flow_key
*key
,
2000 const struct miniflow
*mf
)
2002 return !memcmp(&key
->mf
, mf
, key
->len
);
2006 netdev_flow_key_clone(struct netdev_flow_key
*dst
,
2007 const struct netdev_flow_key
*src
)
2010 offsetof(struct netdev_flow_key
, mf
) + src
->len
);
2013 /* Initialize a netdev_flow_key 'mask' from 'match'. */
2015 netdev_flow_mask_init(struct netdev_flow_key
*mask
,
2016 const struct match
*match
)
2018 uint64_t *dst
= miniflow_values(&mask
->mf
);
2019 struct flowmap fmap
;
2023 /* Only check masks that make sense for the flow. */
2024 flow_wc_map(&match
->flow
, &fmap
);
2025 flowmap_init(&mask
->mf
.map
);
2027 FLOWMAP_FOR_EACH_INDEX(idx
, fmap
) {
2028 uint64_t mask_u64
= flow_u64_value(&match
->wc
.masks
, idx
);
2031 flowmap_set(&mask
->mf
.map
, idx
, 1);
2033 hash
= hash_add64(hash
, mask_u64
);
2039 FLOWMAP_FOR_EACH_MAP (map
, mask
->mf
.map
) {
2040 hash
= hash_add64(hash
, map
);
2043 size_t n
= dst
- miniflow_get_values(&mask
->mf
);
2045 mask
->hash
= hash_finish(hash
, n
* 8);
2046 mask
->len
= netdev_flow_key_size(n
);
2049 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
2051 netdev_flow_key_init_masked(struct netdev_flow_key
*dst
,
2052 const struct flow
*flow
,
2053 const struct netdev_flow_key
*mask
)
2055 uint64_t *dst_u64
= miniflow_values(&dst
->mf
);
2056 const uint64_t *mask_u64
= miniflow_get_values(&mask
->mf
);
2060 dst
->len
= mask
->len
;
2061 dst
->mf
= mask
->mf
; /* Copy maps. */
2063 FLOW_FOR_EACH_IN_MAPS(value
, flow
, mask
->mf
.map
) {
2064 *dst_u64
= value
& *mask_u64
++;
2065 hash
= hash_add64(hash
, *dst_u64
++);
2067 dst
->hash
= hash_finish(hash
,
2068 (dst_u64
- miniflow_get_values(&dst
->mf
)) * 8);
2071 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
2072 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
2073 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
2075 /* Returns a hash value for the bits of 'key' where there are 1-bits in
2077 static inline uint32_t
2078 netdev_flow_key_hash_in_mask(const struct netdev_flow_key
*key
,
2079 const struct netdev_flow_key
*mask
)
2081 const uint64_t *p
= miniflow_get_values(&mask
->mf
);
2085 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, key
, mask
->mf
.map
) {
2086 hash
= hash_add64(hash
, value
& *p
++);
2089 return hash_finish(hash
, (p
- miniflow_get_values(&mask
->mf
)) * 8);
2093 emc_entry_alive(struct emc_entry
*ce
)
2095 return ce
->flow
&& !ce
->flow
->dead
;
2099 emc_clear_entry(struct emc_entry
*ce
)
2102 dp_netdev_flow_unref(ce
->flow
);
2108 emc_change_entry(struct emc_entry
*ce
, struct dp_netdev_flow
*flow
,
2109 const struct netdev_flow_key
*key
)
2111 if (ce
->flow
!= flow
) {
2113 dp_netdev_flow_unref(ce
->flow
);
2116 if (dp_netdev_flow_ref(flow
)) {
2123 netdev_flow_key_clone(&ce
->key
, key
);
2128 emc_insert(struct emc_cache
*cache
, const struct netdev_flow_key
*key
,
2129 struct dp_netdev_flow
*flow
)
2131 struct emc_entry
*to_be_replaced
= NULL
;
2132 struct emc_entry
*current_entry
;
2134 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2135 if (netdev_flow_key_equal(¤t_entry
->key
, key
)) {
2136 /* We found the entry with the 'mf' miniflow */
2137 emc_change_entry(current_entry
, flow
, NULL
);
2141 /* Replacement policy: put the flow in an empty (not alive) entry, or
2142 * in the first entry where it can be */
2144 || (emc_entry_alive(to_be_replaced
)
2145 && !emc_entry_alive(current_entry
))
2146 || current_entry
->key
.hash
< to_be_replaced
->key
.hash
) {
2147 to_be_replaced
= current_entry
;
2150 /* We didn't find the miniflow in the cache.
2151 * The 'to_be_replaced' entry is where the new flow will be stored */
2153 emc_change_entry(to_be_replaced
, flow
, key
);
2157 emc_probabilistic_insert(struct dp_netdev_pmd_thread
*pmd
,
2158 const struct netdev_flow_key
*key
,
2159 struct dp_netdev_flow
*flow
)
2161 /* Insert an entry into the EMC based on probability value 'min'. By
2162 * default the value is UINT32_MAX / 100 which yields an insertion
2163 * probability of 1/100 ie. 1% */
2166 atomic_read_relaxed(&pmd
->dp
->emc_insert_min
, &min
);
2168 if (min
&& random_uint32() <= min
) {
2169 emc_insert(&pmd
->flow_cache
, key
, flow
);
2173 static inline struct dp_netdev_flow
*
2174 emc_lookup(struct emc_cache
*cache
, const struct netdev_flow_key
*key
)
2176 struct emc_entry
*current_entry
;
2178 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2179 if (current_entry
->key
.hash
== key
->hash
2180 && emc_entry_alive(current_entry
)
2181 && netdev_flow_key_equal_mf(¤t_entry
->key
, &key
->mf
)) {
2183 /* We found the entry with the 'key->mf' miniflow */
2184 return current_entry
->flow
;
2191 static struct dp_netdev_flow
*
2192 dp_netdev_pmd_lookup_flow(struct dp_netdev_pmd_thread
*pmd
,
2193 const struct netdev_flow_key
*key
,
2197 struct dpcls_rule
*rule
;
2198 odp_port_t in_port
= u32_to_odp(MINIFLOW_GET_U32(&key
->mf
, in_port
));
2199 struct dp_netdev_flow
*netdev_flow
= NULL
;
2201 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2202 if (OVS_LIKELY(cls
)) {
2203 dpcls_lookup(cls
, key
, &rule
, 1, lookup_num_p
);
2204 netdev_flow
= dp_netdev_flow_cast(rule
);
2209 static struct dp_netdev_flow
*
2210 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread
*pmd
,
2211 const ovs_u128
*ufidp
, const struct nlattr
*key
,
2214 struct dp_netdev_flow
*netdev_flow
;
2218 /* If a UFID is not provided, determine one based on the key. */
2219 if (!ufidp
&& key
&& key_len
2220 && !dpif_netdev_flow_from_nlattrs(key
, key_len
, &flow
, false)) {
2221 dpif_flow_hash(pmd
->dp
->dpif
, &flow
, sizeof flow
, &ufid
);
2226 CMAP_FOR_EACH_WITH_HASH (netdev_flow
, node
, dp_netdev_flow_hash(ufidp
),
2228 if (ovs_u128_equals(netdev_flow
->ufid
, *ufidp
)) {
2238 get_dpif_flow_stats(const struct dp_netdev_flow
*netdev_flow_
,
2239 struct dpif_flow_stats
*stats
)
2241 struct dp_netdev_flow
*netdev_flow
;
2242 unsigned long long n
;
2246 netdev_flow
= CONST_CAST(struct dp_netdev_flow
*, netdev_flow_
);
2248 atomic_read_relaxed(&netdev_flow
->stats
.packet_count
, &n
);
2249 stats
->n_packets
= n
;
2250 atomic_read_relaxed(&netdev_flow
->stats
.byte_count
, &n
);
2252 atomic_read_relaxed(&netdev_flow
->stats
.used
, &used
);
2254 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
2255 stats
->tcp_flags
= flags
;
2258 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
2259 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
2260 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
2263 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow
*netdev_flow
,
2264 struct ofpbuf
*key_buf
, struct ofpbuf
*mask_buf
,
2265 struct dpif_flow
*flow
, bool terse
)
2268 memset(flow
, 0, sizeof *flow
);
2270 struct flow_wildcards wc
;
2271 struct dp_netdev_actions
*actions
;
2273 struct odp_flow_key_parms odp_parms
= {
2274 .flow
= &netdev_flow
->flow
,
2276 .support
= dp_netdev_support
,
2279 miniflow_expand(&netdev_flow
->cr
.mask
->mf
, &wc
.masks
);
2280 /* in_port is exact matched, but we have left it out from the mask for
2281 * optimnization reasons. Add in_port back to the mask. */
2282 wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
2285 offset
= key_buf
->size
;
2286 flow
->key
= ofpbuf_tail(key_buf
);
2287 odp_flow_key_from_flow(&odp_parms
, key_buf
);
2288 flow
->key_len
= key_buf
->size
- offset
;
2291 offset
= mask_buf
->size
;
2292 flow
->mask
= ofpbuf_tail(mask_buf
);
2293 odp_parms
.key_buf
= key_buf
;
2294 odp_flow_key_from_mask(&odp_parms
, mask_buf
);
2295 flow
->mask_len
= mask_buf
->size
- offset
;
2298 actions
= dp_netdev_flow_get_actions(netdev_flow
);
2299 flow
->actions
= actions
->actions
;
2300 flow
->actions_len
= actions
->size
;
2303 flow
->ufid
= netdev_flow
->ufid
;
2304 flow
->ufid_present
= true;
2305 flow
->pmd_id
= netdev_flow
->pmd_id
;
2306 get_dpif_flow_stats(netdev_flow
, &flow
->stats
);
2310 dpif_netdev_mask_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
2311 const struct nlattr
*mask_key
,
2312 uint32_t mask_key_len
, const struct flow
*flow
,
2313 struct flow_wildcards
*wc
, bool probe
)
2315 enum odp_key_fitness fitness
;
2317 fitness
= odp_flow_key_to_mask(mask_key
, mask_key_len
, wc
, flow
);
2320 /* This should not happen: it indicates that
2321 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
2322 * disagree on the acceptable form of a mask. Log the problem
2323 * as an error, with enough details to enable debugging. */
2324 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
2326 if (!VLOG_DROP_ERR(&rl
)) {
2330 odp_flow_format(key
, key_len
, mask_key
, mask_key_len
, NULL
, &s
,
2332 VLOG_ERR("internal error parsing flow mask %s (%s)",
2333 ds_cstr(&s
), odp_key_fitness_to_string(fitness
));
2345 dpif_netdev_flow_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
2346 struct flow
*flow
, bool probe
)
2348 if (odp_flow_key_to_flow(key
, key_len
, flow
)) {
2350 /* This should not happen: it indicates that
2351 * odp_flow_key_from_flow() and odp_flow_key_to_flow() disagree on
2352 * the acceptable form of a flow. Log the problem as an error,
2353 * with enough details to enable debugging. */
2354 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
2356 if (!VLOG_DROP_ERR(&rl
)) {
2360 odp_flow_format(key
, key_len
, NULL
, 0, NULL
, &s
, true);
2361 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s
));
2369 if (flow
->ct_state
& DP_NETDEV_CS_UNSUPPORTED_MASK
) {
2377 dpif_netdev_flow_get(const struct dpif
*dpif
, const struct dpif_flow_get
*get
)
2379 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2380 struct dp_netdev_flow
*netdev_flow
;
2381 struct dp_netdev_pmd_thread
*pmd
;
2382 struct hmapx to_find
= HMAPX_INITIALIZER(&to_find
);
2383 struct hmapx_node
*node
;
2386 if (get
->pmd_id
== PMD_ID_NULL
) {
2387 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2388 if (dp_netdev_pmd_try_ref(pmd
) && !hmapx_add(&to_find
, pmd
)) {
2389 dp_netdev_pmd_unref(pmd
);
2393 pmd
= dp_netdev_get_pmd(dp
, get
->pmd_id
);
2397 hmapx_add(&to_find
, pmd
);
2400 if (!hmapx_count(&to_find
)) {
2404 HMAPX_FOR_EACH (node
, &to_find
) {
2405 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
2406 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, get
->ufid
, get
->key
,
2409 dp_netdev_flow_to_dpif_flow(netdev_flow
, get
->buffer
, get
->buffer
,
2418 HMAPX_FOR_EACH (node
, &to_find
) {
2419 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
2420 dp_netdev_pmd_unref(pmd
);
2423 hmapx_destroy(&to_find
);
2427 static struct dp_netdev_flow
*
2428 dp_netdev_flow_add(struct dp_netdev_pmd_thread
*pmd
,
2429 struct match
*match
, const ovs_u128
*ufid
,
2430 const struct nlattr
*actions
, size_t actions_len
)
2431 OVS_REQUIRES(pmd
->flow_mutex
)
2433 struct dp_netdev_flow
*flow
;
2434 struct netdev_flow_key mask
;
2437 /* Make sure in_port is exact matched before we read it. */
2438 ovs_assert(match
->wc
.masks
.in_port
.odp_port
== ODPP_NONE
);
2439 odp_port_t in_port
= match
->flow
.in_port
.odp_port
;
2441 /* As we select the dpcls based on the port number, each netdev flow
2442 * belonging to the same dpcls will have the same odp_port value.
2443 * For performance reasons we wildcard odp_port here in the mask. In the
2444 * typical case dp_hash is also wildcarded, and the resulting 8-byte
2445 * chunk {dp_hash, in_port} will be ignored by netdev_flow_mask_init() and
2446 * will not be part of the subtable mask.
2447 * This will speed up the hash computation during dpcls_lookup() because
2448 * there is one less call to hash_add64() in this case. */
2449 match
->wc
.masks
.in_port
.odp_port
= 0;
2450 netdev_flow_mask_init(&mask
, match
);
2451 match
->wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
2453 /* Make sure wc does not have metadata. */
2454 ovs_assert(!FLOWMAP_HAS_FIELD(&mask
.mf
.map
, metadata
)
2455 && !FLOWMAP_HAS_FIELD(&mask
.mf
.map
, regs
));
2457 /* Do not allocate extra space. */
2458 flow
= xmalloc(sizeof *flow
- sizeof flow
->cr
.flow
.mf
+ mask
.len
);
2459 memset(&flow
->stats
, 0, sizeof flow
->stats
);
2462 *CONST_CAST(unsigned *, &flow
->pmd_id
) = pmd
->core_id
;
2463 *CONST_CAST(struct flow
*, &flow
->flow
) = match
->flow
;
2464 *CONST_CAST(ovs_u128
*, &flow
->ufid
) = *ufid
;
2465 ovs_refcount_init(&flow
->ref_cnt
);
2466 ovsrcu_set(&flow
->actions
, dp_netdev_actions_create(actions
, actions_len
));
2468 netdev_flow_key_init_masked(&flow
->cr
.flow
, &match
->flow
, &mask
);
2470 /* Select dpcls for in_port. Relies on in_port to be exact match. */
2471 cls
= dp_netdev_pmd_find_dpcls(pmd
, in_port
);
2472 dpcls_insert(cls
, &flow
->cr
, &mask
);
2474 cmap_insert(&pmd
->flow_table
, CONST_CAST(struct cmap_node
*, &flow
->node
),
2475 dp_netdev_flow_hash(&flow
->ufid
));
2477 if (OVS_UNLIKELY(!VLOG_DROP_DBG((&upcall_rl
)))) {
2478 struct ds ds
= DS_EMPTY_INITIALIZER
;
2479 struct ofpbuf key_buf
, mask_buf
;
2480 struct odp_flow_key_parms odp_parms
= {
2481 .flow
= &match
->flow
,
2482 .mask
= &match
->wc
.masks
,
2483 .support
= dp_netdev_support
,
2486 ofpbuf_init(&key_buf
, 0);
2487 ofpbuf_init(&mask_buf
, 0);
2489 odp_flow_key_from_flow(&odp_parms
, &key_buf
);
2490 odp_parms
.key_buf
= &key_buf
;
2491 odp_flow_key_from_mask(&odp_parms
, &mask_buf
);
2493 ds_put_cstr(&ds
, "flow_add: ");
2494 odp_format_ufid(ufid
, &ds
);
2495 ds_put_cstr(&ds
, " ");
2496 odp_flow_format(key_buf
.data
, key_buf
.size
,
2497 mask_buf
.data
, mask_buf
.size
,
2499 ds_put_cstr(&ds
, ", actions:");
2500 format_odp_actions(&ds
, actions
, actions_len
, NULL
);
2502 VLOG_DBG("%s", ds_cstr(&ds
));
2504 ofpbuf_uninit(&key_buf
);
2505 ofpbuf_uninit(&mask_buf
);
2507 /* Add a printout of the actual match installed. */
2510 ds_put_cstr(&ds
, "flow match: ");
2511 miniflow_expand(&flow
->cr
.flow
.mf
, &m
.flow
);
2512 miniflow_expand(&flow
->cr
.mask
->mf
, &m
.wc
.masks
);
2513 memset(&m
.tun_md
, 0, sizeof m
.tun_md
);
2514 match_format(&m
, NULL
, &ds
, OFP_DEFAULT_PRIORITY
);
2516 VLOG_DBG("%s", ds_cstr(&ds
));
2525 flow_put_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
2526 struct netdev_flow_key
*key
,
2527 struct match
*match
,
2529 const struct dpif_flow_put
*put
,
2530 struct dpif_flow_stats
*stats
)
2532 struct dp_netdev_flow
*netdev_flow
;
2536 memset(stats
, 0, sizeof *stats
);
2539 ovs_mutex_lock(&pmd
->flow_mutex
);
2540 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
2542 if (put
->flags
& DPIF_FP_CREATE
) {
2543 if (cmap_count(&pmd
->flow_table
) < MAX_FLOWS
) {
2544 dp_netdev_flow_add(pmd
, match
, ufid
, put
->actions
,
2554 if (put
->flags
& DPIF_FP_MODIFY
) {
2555 struct dp_netdev_actions
*new_actions
;
2556 struct dp_netdev_actions
*old_actions
;
2558 new_actions
= dp_netdev_actions_create(put
->actions
,
2561 old_actions
= dp_netdev_flow_get_actions(netdev_flow
);
2562 ovsrcu_set(&netdev_flow
->actions
, new_actions
);
2565 get_dpif_flow_stats(netdev_flow
, stats
);
2567 if (put
->flags
& DPIF_FP_ZERO_STATS
) {
2568 /* XXX: The userspace datapath uses thread local statistics
2569 * (for flows), which should be updated only by the owning
2570 * thread. Since we cannot write on stats memory here,
2571 * we choose not to support this flag. Please note:
2572 * - This feature is currently used only by dpctl commands with
2574 * - Should the need arise, this operation can be implemented
2575 * by keeping a base value (to be update here) for each
2576 * counter, and subtracting it before outputting the stats */
2580 ovsrcu_postpone(dp_netdev_actions_free
, old_actions
);
2581 } else if (put
->flags
& DPIF_FP_CREATE
) {
2584 /* Overlapping flow. */
2588 ovs_mutex_unlock(&pmd
->flow_mutex
);
2593 dpif_netdev_flow_put(struct dpif
*dpif
, const struct dpif_flow_put
*put
)
2595 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2596 struct netdev_flow_key key
, mask
;
2597 struct dp_netdev_pmd_thread
*pmd
;
2601 bool probe
= put
->flags
& DPIF_FP_PROBE
;
2604 memset(put
->stats
, 0, sizeof *put
->stats
);
2606 error
= dpif_netdev_flow_from_nlattrs(put
->key
, put
->key_len
, &match
.flow
,
2611 error
= dpif_netdev_mask_from_nlattrs(put
->key
, put
->key_len
,
2612 put
->mask
, put
->mask_len
,
2613 &match
.flow
, &match
.wc
, probe
);
2621 dpif_flow_hash(dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
2624 /* Must produce a netdev_flow_key for lookup.
2625 * Use the same method as employed to create the key when adding
2626 * the flow to the dplcs to make sure they match. */
2627 netdev_flow_mask_init(&mask
, &match
);
2628 netdev_flow_key_init_masked(&key
, &match
.flow
, &mask
);
2630 if (put
->pmd_id
== PMD_ID_NULL
) {
2631 if (cmap_count(&dp
->poll_threads
) == 0) {
2634 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2635 struct dpif_flow_stats pmd_stats
;
2638 pmd_error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
,
2642 } else if (put
->stats
) {
2643 put
->stats
->n_packets
+= pmd_stats
.n_packets
;
2644 put
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
2645 put
->stats
->used
= MAX(put
->stats
->used
, pmd_stats
.used
);
2646 put
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
2650 pmd
= dp_netdev_get_pmd(dp
, put
->pmd_id
);
2654 error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
, put
->stats
);
2655 dp_netdev_pmd_unref(pmd
);
2662 flow_del_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
2663 struct dpif_flow_stats
*stats
,
2664 const struct dpif_flow_del
*del
)
2666 struct dp_netdev_flow
*netdev_flow
;
2669 ovs_mutex_lock(&pmd
->flow_mutex
);
2670 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, del
->ufid
, del
->key
,
2674 get_dpif_flow_stats(netdev_flow
, stats
);
2676 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
2680 ovs_mutex_unlock(&pmd
->flow_mutex
);
2686 dpif_netdev_flow_del(struct dpif
*dpif
, const struct dpif_flow_del
*del
)
2688 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2689 struct dp_netdev_pmd_thread
*pmd
;
2693 memset(del
->stats
, 0, sizeof *del
->stats
);
2696 if (del
->pmd_id
== PMD_ID_NULL
) {
2697 if (cmap_count(&dp
->poll_threads
) == 0) {
2700 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2701 struct dpif_flow_stats pmd_stats
;
2704 pmd_error
= flow_del_on_pmd(pmd
, &pmd_stats
, del
);
2707 } else if (del
->stats
) {
2708 del
->stats
->n_packets
+= pmd_stats
.n_packets
;
2709 del
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
2710 del
->stats
->used
= MAX(del
->stats
->used
, pmd_stats
.used
);
2711 del
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
2715 pmd
= dp_netdev_get_pmd(dp
, del
->pmd_id
);
2719 error
= flow_del_on_pmd(pmd
, del
->stats
, del
);
2720 dp_netdev_pmd_unref(pmd
);
2727 struct dpif_netdev_flow_dump
{
2728 struct dpif_flow_dump up
;
2729 struct cmap_position poll_thread_pos
;
2730 struct cmap_position flow_pos
;
2731 struct dp_netdev_pmd_thread
*cur_pmd
;
2733 struct ovs_mutex mutex
;
2736 static struct dpif_netdev_flow_dump
*
2737 dpif_netdev_flow_dump_cast(struct dpif_flow_dump
*dump
)
2739 return CONTAINER_OF(dump
, struct dpif_netdev_flow_dump
, up
);
2742 static struct dpif_flow_dump
*
2743 dpif_netdev_flow_dump_create(const struct dpif
*dpif_
, bool terse
,
2744 char *type OVS_UNUSED
)
2746 struct dpif_netdev_flow_dump
*dump
;
2748 dump
= xzalloc(sizeof *dump
);
2749 dpif_flow_dump_init(&dump
->up
, dpif_
);
2750 dump
->up
.terse
= terse
;
2751 ovs_mutex_init(&dump
->mutex
);
2757 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump
*dump_
)
2759 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
2761 ovs_mutex_destroy(&dump
->mutex
);
2766 struct dpif_netdev_flow_dump_thread
{
2767 struct dpif_flow_dump_thread up
;
2768 struct dpif_netdev_flow_dump
*dump
;
2769 struct odputil_keybuf keybuf
[FLOW_DUMP_MAX_BATCH
];
2770 struct odputil_keybuf maskbuf
[FLOW_DUMP_MAX_BATCH
];
2773 static struct dpif_netdev_flow_dump_thread
*
2774 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread
*thread
)
2776 return CONTAINER_OF(thread
, struct dpif_netdev_flow_dump_thread
, up
);
2779 static struct dpif_flow_dump_thread
*
2780 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump
*dump_
)
2782 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
2783 struct dpif_netdev_flow_dump_thread
*thread
;
2785 thread
= xmalloc(sizeof *thread
);
2786 dpif_flow_dump_thread_init(&thread
->up
, &dump
->up
);
2787 thread
->dump
= dump
;
2792 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread
*thread_
)
2794 struct dpif_netdev_flow_dump_thread
*thread
2795 = dpif_netdev_flow_dump_thread_cast(thread_
);
2801 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread
*thread_
,
2802 struct dpif_flow
*flows
, int max_flows
)
2804 struct dpif_netdev_flow_dump_thread
*thread
2805 = dpif_netdev_flow_dump_thread_cast(thread_
);
2806 struct dpif_netdev_flow_dump
*dump
= thread
->dump
;
2807 struct dp_netdev_flow
*netdev_flows
[FLOW_DUMP_MAX_BATCH
];
2811 ovs_mutex_lock(&dump
->mutex
);
2812 if (!dump
->status
) {
2813 struct dpif_netdev
*dpif
= dpif_netdev_cast(thread
->up
.dpif
);
2814 struct dp_netdev
*dp
= get_dp_netdev(&dpif
->dpif
);
2815 struct dp_netdev_pmd_thread
*pmd
= dump
->cur_pmd
;
2816 int flow_limit
= MIN(max_flows
, FLOW_DUMP_MAX_BATCH
);
2818 /* First call to dump_next(), extracts the first pmd thread.
2819 * If there is no pmd thread, returns immediately. */
2821 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
2823 ovs_mutex_unlock(&dump
->mutex
);
2830 for (n_flows
= 0; n_flows
< flow_limit
; n_flows
++) {
2831 struct cmap_node
*node
;
2833 node
= cmap_next_position(&pmd
->flow_table
, &dump
->flow_pos
);
2837 netdev_flows
[n_flows
] = CONTAINER_OF(node
,
2838 struct dp_netdev_flow
,
2841 /* When finishing dumping the current pmd thread, moves to
2843 if (n_flows
< flow_limit
) {
2844 memset(&dump
->flow_pos
, 0, sizeof dump
->flow_pos
);
2845 dp_netdev_pmd_unref(pmd
);
2846 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
2852 /* Keeps the reference to next caller. */
2853 dump
->cur_pmd
= pmd
;
2855 /* If the current dump is empty, do not exit the loop, since the
2856 * remaining pmds could have flows to be dumped. Just dumps again
2857 * on the new 'pmd'. */
2860 ovs_mutex_unlock(&dump
->mutex
);
2862 for (i
= 0; i
< n_flows
; i
++) {
2863 struct odputil_keybuf
*maskbuf
= &thread
->maskbuf
[i
];
2864 struct odputil_keybuf
*keybuf
= &thread
->keybuf
[i
];
2865 struct dp_netdev_flow
*netdev_flow
= netdev_flows
[i
];
2866 struct dpif_flow
*f
= &flows
[i
];
2867 struct ofpbuf key
, mask
;
2869 ofpbuf_use_stack(&key
, keybuf
, sizeof *keybuf
);
2870 ofpbuf_use_stack(&mask
, maskbuf
, sizeof *maskbuf
);
2871 dp_netdev_flow_to_dpif_flow(netdev_flow
, &key
, &mask
, f
,
2879 dpif_netdev_execute(struct dpif
*dpif
, struct dpif_execute
*execute
)
2880 OVS_NO_THREAD_SAFETY_ANALYSIS
2882 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2883 struct dp_netdev_pmd_thread
*pmd
;
2884 struct dp_packet_batch pp
;
2886 if (dp_packet_size(execute
->packet
) < ETH_HEADER_LEN
||
2887 dp_packet_size(execute
->packet
) > UINT16_MAX
) {
2891 /* Tries finding the 'pmd'. If NULL is returned, that means
2892 * the current thread is a non-pmd thread and should use
2893 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2894 pmd
= ovsthread_getspecific(dp
->per_pmd_key
);
2896 pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
2902 if (execute
->probe
) {
2903 /* If this is part of a probe, Drop the packet, since executing
2904 * the action may actually cause spurious packets be sent into
2909 /* If the current thread is non-pmd thread, acquires
2910 * the 'non_pmd_mutex'. */
2911 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2912 ovs_mutex_lock(&dp
->non_pmd_mutex
);
2915 /* The action processing expects the RSS hash to be valid, because
2916 * it's always initialized at the beginning of datapath processing.
2917 * In this case, though, 'execute->packet' may not have gone through
2918 * the datapath at all, it may have been generated by the upper layer
2919 * (OpenFlow packet-out, BFD frame, ...). */
2920 if (!dp_packet_rss_valid(execute
->packet
)) {
2921 dp_packet_set_rss_hash(execute
->packet
,
2922 flow_hash_5tuple(execute
->flow
, 0));
2925 dp_packet_batch_init_packet(&pp
, execute
->packet
);
2926 dp_netdev_execute_actions(pmd
, &pp
, false, execute
->flow
,
2927 execute
->actions
, execute
->actions_len
,
2930 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2931 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
2932 dp_netdev_pmd_unref(pmd
);
2939 dpif_netdev_operate(struct dpif
*dpif
, struct dpif_op
**ops
, size_t n_ops
)
2943 for (i
= 0; i
< n_ops
; i
++) {
2944 struct dpif_op
*op
= ops
[i
];
2947 case DPIF_OP_FLOW_PUT
:
2948 op
->error
= dpif_netdev_flow_put(dpif
, &op
->u
.flow_put
);
2951 case DPIF_OP_FLOW_DEL
:
2952 op
->error
= dpif_netdev_flow_del(dpif
, &op
->u
.flow_del
);
2955 case DPIF_OP_EXECUTE
:
2956 op
->error
= dpif_netdev_execute(dpif
, &op
->u
.execute
);
2959 case DPIF_OP_FLOW_GET
:
2960 op
->error
= dpif_netdev_flow_get(dpif
, &op
->u
.flow_get
);
2966 /* Applies datapath configuration from the database. Some of the changes are
2967 * actually applied in dpif_netdev_run(). */
2969 dpif_netdev_set_config(struct dpif
*dpif
, const struct smap
*other_config
)
2971 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2972 const char *cmask
= smap_get(other_config
, "pmd-cpu-mask");
2973 unsigned long long insert_prob
=
2974 smap_get_ullong(other_config
, "emc-insert-inv-prob",
2975 DEFAULT_EM_FLOW_INSERT_INV_PROB
);
2976 uint32_t insert_min
, cur_min
;
2978 if (!nullable_string_is_equal(dp
->pmd_cmask
, cmask
)) {
2979 free(dp
->pmd_cmask
);
2980 dp
->pmd_cmask
= nullable_xstrdup(cmask
);
2981 dp_netdev_request_reconfigure(dp
);
2984 atomic_read_relaxed(&dp
->emc_insert_min
, &cur_min
);
2985 if (insert_prob
<= UINT32_MAX
) {
2986 insert_min
= insert_prob
== 0 ? 0 : UINT32_MAX
/ insert_prob
;
2988 insert_min
= DEFAULT_EM_FLOW_INSERT_MIN
;
2989 insert_prob
= DEFAULT_EM_FLOW_INSERT_INV_PROB
;
2992 if (insert_min
!= cur_min
) {
2993 atomic_store_relaxed(&dp
->emc_insert_min
, insert_min
);
2994 if (insert_min
== 0) {
2995 VLOG_INFO("EMC has been disabled");
2997 VLOG_INFO("EMC insertion probability changed to 1/%llu (~%.2f%%)",
2998 insert_prob
, (100 / (float)insert_prob
));
3005 /* Parses affinity list and returns result in 'core_ids'. */
3007 parse_affinity_list(const char *affinity_list
, unsigned *core_ids
, int n_rxq
)
3010 char *list
, *copy
, *key
, *value
;
3013 for (i
= 0; i
< n_rxq
; i
++) {
3014 core_ids
[i
] = OVS_CORE_UNSPEC
;
3017 if (!affinity_list
) {
3021 list
= copy
= xstrdup(affinity_list
);
3023 while (ofputil_parse_key_value(&list
, &key
, &value
)) {
3024 int rxq_id
, core_id
;
3026 if (!str_to_int(key
, 0, &rxq_id
) || rxq_id
< 0
3027 || !str_to_int(value
, 0, &core_id
) || core_id
< 0) {
3032 if (rxq_id
< n_rxq
) {
3033 core_ids
[rxq_id
] = core_id
;
3041 /* Parses 'affinity_list' and applies configuration if it is valid. */
3043 dpif_netdev_port_set_rxq_affinity(struct dp_netdev_port
*port
,
3044 const char *affinity_list
)
3046 unsigned *core_ids
, i
;
3049 core_ids
= xmalloc(port
->n_rxq
* sizeof *core_ids
);
3050 if (parse_affinity_list(affinity_list
, core_ids
, port
->n_rxq
)) {
3055 for (i
= 0; i
< port
->n_rxq
; i
++) {
3056 port
->rxqs
[i
].core_id
= core_ids
[i
];
3064 /* Changes the affinity of port's rx queues. The changes are actually applied
3065 * in dpif_netdev_run(). */
3067 dpif_netdev_port_set_config(struct dpif
*dpif
, odp_port_t port_no
,
3068 const struct smap
*cfg
)
3070 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3071 struct dp_netdev_port
*port
;
3073 const char *affinity_list
= smap_get(cfg
, "pmd-rxq-affinity");
3075 ovs_mutex_lock(&dp
->port_mutex
);
3076 error
= get_port_by_number(dp
, port_no
, &port
);
3077 if (error
|| !netdev_is_pmd(port
->netdev
)
3078 || nullable_string_is_equal(affinity_list
, port
->rxq_affinity_list
)) {
3082 error
= dpif_netdev_port_set_rxq_affinity(port
, affinity_list
);
3086 free(port
->rxq_affinity_list
);
3087 port
->rxq_affinity_list
= nullable_xstrdup(affinity_list
);
3089 dp_netdev_request_reconfigure(dp
);
3091 ovs_mutex_unlock(&dp
->port_mutex
);
3096 dpif_netdev_queue_to_priority(const struct dpif
*dpif OVS_UNUSED
,
3097 uint32_t queue_id
, uint32_t *priority
)
3099 *priority
= queue_id
;
3104 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
3105 * a copy of the 'size' bytes of 'actions' input parameters. */
3106 struct dp_netdev_actions
*
3107 dp_netdev_actions_create(const struct nlattr
*actions
, size_t size
)
3109 struct dp_netdev_actions
*netdev_actions
;
3111 netdev_actions
= xmalloc(sizeof *netdev_actions
+ size
);
3112 memcpy(netdev_actions
->actions
, actions
, size
);
3113 netdev_actions
->size
= size
;
3115 return netdev_actions
;
3118 struct dp_netdev_actions
*
3119 dp_netdev_flow_get_actions(const struct dp_netdev_flow
*flow
)
3121 return ovsrcu_get(struct dp_netdev_actions
*, &flow
->actions
);
3125 dp_netdev_actions_free(struct dp_netdev_actions
*actions
)
3130 static inline unsigned long long
3131 cycles_counter(void)
3134 return rte_get_tsc_cycles();
3140 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
3141 extern struct ovs_mutex cycles_counter_fake_mutex
;
3143 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
3145 cycles_count_start(struct dp_netdev_pmd_thread
*pmd
)
3146 OVS_ACQUIRES(&cycles_counter_fake_mutex
)
3147 OVS_NO_THREAD_SAFETY_ANALYSIS
3149 pmd
->last_cycles
= cycles_counter();
3152 /* Stop counting cycles and add them to the counter 'type' */
3154 cycles_count_end(struct dp_netdev_pmd_thread
*pmd
,
3155 enum pmd_cycles_counter_type type
)
3156 OVS_RELEASES(&cycles_counter_fake_mutex
)
3157 OVS_NO_THREAD_SAFETY_ANALYSIS
3159 unsigned long long interval
= cycles_counter() - pmd
->last_cycles
;
3161 non_atomic_ullong_add(&pmd
->cycles
.n
[type
], interval
);
3164 /* Calculate the intermediate cycle result and add to the counter 'type' */
3166 cycles_count_intermediate(struct dp_netdev_pmd_thread
*pmd
,
3167 struct dp_netdev_rxq
*rxq
,
3168 enum pmd_cycles_counter_type type
)
3169 OVS_NO_THREAD_SAFETY_ANALYSIS
3171 unsigned long long new_cycles
= cycles_counter();
3172 unsigned long long interval
= new_cycles
- pmd
->last_cycles
;
3173 pmd
->last_cycles
= new_cycles
;
3175 non_atomic_ullong_add(&pmd
->cycles
.n
[type
], interval
);
3176 if (rxq
&& (type
== PMD_CYCLES_PROCESSING
)) {
3177 /* Add to the amount of current processing cycles. */
3178 non_atomic_ullong_add(&rxq
->cycles
[RXQ_CYCLES_PROC_CURR
], interval
);
3183 dp_netdev_rxq_set_cycles(struct dp_netdev_rxq
*rx
,
3184 enum rxq_cycles_counter_type type
,
3185 unsigned long long cycles
)
3187 atomic_store_relaxed(&rx
->cycles
[type
], cycles
);
3191 dp_netdev_rxq_get_cycles(struct dp_netdev_rxq
*rx
,
3192 enum rxq_cycles_counter_type type
)
3194 unsigned long long processing_cycles
;
3195 atomic_read_relaxed(&rx
->cycles
[type
], &processing_cycles
);
3196 return processing_cycles
;
3200 dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq
*rx
,
3201 unsigned long long cycles
)
3203 unsigned int idx
= rx
->intrvl_idx
++ % PMD_RXQ_INTERVAL_MAX
;
3204 atomic_store_relaxed(&rx
->cycles_intrvl
[idx
], cycles
);
3208 dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq
*rx
, unsigned idx
)
3210 unsigned long long processing_cycles
;
3211 atomic_read_relaxed(&rx
->cycles_intrvl
[idx
], &processing_cycles
);
3212 return processing_cycles
;
3216 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread
*pmd
,
3217 struct netdev_rxq
*rx
,
3220 struct dp_packet_batch batch
;
3224 dp_packet_batch_init(&batch
);
3225 error
= netdev_rxq_recv(rx
, &batch
);
3227 *recirc_depth_get() = 0;
3229 batch_cnt
= batch
.count
;
3230 dp_netdev_input(pmd
, &batch
, port_no
);
3231 } else if (error
!= EAGAIN
&& error
!= EOPNOTSUPP
) {
3232 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
3234 VLOG_ERR_RL(&rl
, "error receiving data from %s: %s",
3235 netdev_rxq_get_name(rx
), ovs_strerror(error
));
3241 static struct tx_port
*
3242 tx_port_lookup(const struct hmap
*hmap
, odp_port_t port_no
)
3246 HMAP_FOR_EACH_IN_BUCKET (tx
, node
, hash_port_no(port_no
), hmap
) {
3247 if (tx
->port
->port_no
== port_no
) {
3256 port_reconfigure(struct dp_netdev_port
*port
)
3258 struct netdev
*netdev
= port
->netdev
;
3261 port
->need_reconfigure
= false;
3263 /* Closes the existing 'rxq's. */
3264 for (i
= 0; i
< port
->n_rxq
; i
++) {
3265 netdev_rxq_close(port
->rxqs
[i
].rx
);
3266 port
->rxqs
[i
].rx
= NULL
;
3268 unsigned last_nrxq
= port
->n_rxq
;
3271 /* Allows 'netdev' to apply the pending configuration changes. */
3272 if (netdev_is_reconf_required(netdev
)) {
3273 err
= netdev_reconfigure(netdev
);
3274 if (err
&& (err
!= EOPNOTSUPP
)) {
3275 VLOG_ERR("Failed to set interface %s new configuration",
3276 netdev_get_name(netdev
));
3280 /* If the netdev_reconfigure() above succeeds, reopens the 'rxq's. */
3281 port
->rxqs
= xrealloc(port
->rxqs
,
3282 sizeof *port
->rxqs
* netdev_n_rxq(netdev
));
3283 /* Realloc 'used' counters for tx queues. */
3284 free(port
->txq_used
);
3285 port
->txq_used
= xcalloc(netdev_n_txq(netdev
), sizeof *port
->txq_used
);
3287 for (i
= 0; i
< netdev_n_rxq(netdev
); i
++) {
3288 bool new_queue
= i
>= last_nrxq
;
3290 memset(&port
->rxqs
[i
], 0, sizeof port
->rxqs
[i
]);
3293 port
->rxqs
[i
].port
= port
;
3296 dp_netdev_rxq_set_cycles(&port
->rxqs
[i
], RXQ_CYCLES_PROC_CURR
, 0);
3297 dp_netdev_rxq_set_cycles(&port
->rxqs
[i
], RXQ_CYCLES_PROC_HIST
, 0);
3298 for (unsigned j
= 0; j
< PMD_RXQ_INTERVAL_MAX
; j
++) {
3299 dp_netdev_rxq_set_intrvl_cycles(&port
->rxqs
[i
], 0);
3302 err
= netdev_rxq_open(netdev
, &port
->rxqs
[i
].rx
, i
);
3309 /* Parse affinity list to apply configuration for new queues. */
3310 dpif_netdev_port_set_rxq_affinity(port
, port
->rxq_affinity_list
);
3315 struct rr_numa_list
{
3316 struct hmap numas
; /* Contains 'struct rr_numa' */
3320 struct hmap_node node
;
3324 /* Non isolated pmds on numa node 'numa_id' */
3325 struct dp_netdev_pmd_thread
**pmds
;
3332 static struct rr_numa
*
3333 rr_numa_list_lookup(struct rr_numa_list
*rr
, int numa_id
)
3335 struct rr_numa
*numa
;
3337 HMAP_FOR_EACH_WITH_HASH (numa
, node
, hash_int(numa_id
, 0), &rr
->numas
) {
3338 if (numa
->numa_id
== numa_id
) {
3346 /* Returns the next node in numa list following 'numa' in round-robin fashion.
3347 * Returns first node if 'numa' is a null pointer or the last node in 'rr'.
3348 * Returns NULL if 'rr' numa list is empty. */
3349 static struct rr_numa
*
3350 rr_numa_list_next(struct rr_numa_list
*rr
, const struct rr_numa
*numa
)
3352 struct hmap_node
*node
= NULL
;
3355 node
= hmap_next(&rr
->numas
, &numa
->node
);
3358 node
= hmap_first(&rr
->numas
);
3361 return (node
) ? CONTAINER_OF(node
, struct rr_numa
, node
) : NULL
;
3365 rr_numa_list_populate(struct dp_netdev
*dp
, struct rr_numa_list
*rr
)
3367 struct dp_netdev_pmd_thread
*pmd
;
3368 struct rr_numa
*numa
;
3370 hmap_init(&rr
->numas
);
3372 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3373 if (pmd
->core_id
== NON_PMD_CORE_ID
|| pmd
->isolated
) {
3377 numa
= rr_numa_list_lookup(rr
, pmd
->numa_id
);
3379 numa
= xzalloc(sizeof *numa
);
3380 numa
->numa_id
= pmd
->numa_id
;
3381 hmap_insert(&rr
->numas
, &numa
->node
, hash_int(pmd
->numa_id
, 0));
3384 numa
->pmds
= xrealloc(numa
->pmds
, numa
->n_pmds
* sizeof *numa
->pmds
);
3385 numa
->pmds
[numa
->n_pmds
- 1] = pmd
;
3386 /* At least one pmd so initialise curr_idx and idx_inc. */
3387 numa
->cur_index
= 0;
3388 numa
->idx_inc
= true;
3392 /* Returns the next pmd from the numa node in
3393 * incrementing or decrementing order. */
3394 static struct dp_netdev_pmd_thread
*
3395 rr_numa_get_pmd(struct rr_numa
*numa
)
3397 int numa_idx
= numa
->cur_index
;
3399 if (numa
->idx_inc
== true) {
3400 /* Incrementing through list of pmds. */
3401 if (numa
->cur_index
== numa
->n_pmds
-1) {
3402 /* Reached the last pmd. */
3403 numa
->idx_inc
= false;
3408 /* Decrementing through list of pmds. */
3409 if (numa
->cur_index
== 0) {
3410 /* Reached the first pmd. */
3411 numa
->idx_inc
= true;
3416 return numa
->pmds
[numa_idx
];
3420 rr_numa_list_destroy(struct rr_numa_list
*rr
)
3422 struct rr_numa
*numa
;
3424 HMAP_FOR_EACH_POP (numa
, node
, &rr
->numas
) {
3428 hmap_destroy(&rr
->numas
);
3431 /* Sort Rx Queues by the processing cycles they are consuming. */
3433 rxq_cycle_sort(const void *a
, const void *b
)
3435 struct dp_netdev_rxq
*qa
;
3436 struct dp_netdev_rxq
*qb
;
3437 uint64_t total_qa
, total_qb
;
3440 qa
= *(struct dp_netdev_rxq
**) a
;
3441 qb
= *(struct dp_netdev_rxq
**) b
;
3443 total_qa
= total_qb
= 0;
3444 for (i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
3445 total_qa
+= dp_netdev_rxq_get_intrvl_cycles(qa
, i
);
3446 total_qb
+= dp_netdev_rxq_get_intrvl_cycles(qb
, i
);
3448 dp_netdev_rxq_set_cycles(qa
, RXQ_CYCLES_PROC_HIST
, total_qa
);
3449 dp_netdev_rxq_set_cycles(qb
, RXQ_CYCLES_PROC_HIST
, total_qb
);
3451 if (total_qa
>= total_qb
) {
3457 /* Assign pmds to queues. If 'pinned' is true, assign pmds to pinned
3458 * queues and marks the pmds as isolated. Otherwise, assign non isolated
3459 * pmds to unpinned queues.
3461 * If 'pinned' is false queues will be sorted by processing cycles they are
3462 * consuming and then assigned to pmds in round robin order.
3464 * The function doesn't touch the pmd threads, it just stores the assignment
3465 * in the 'pmd' member of each rxq. */
3467 rxq_scheduling(struct dp_netdev
*dp
, bool pinned
) OVS_REQUIRES(dp
->port_mutex
)
3469 struct dp_netdev_port
*port
;
3470 struct rr_numa_list rr
;
3471 struct rr_numa
*non_local_numa
= NULL
;
3472 struct dp_netdev_rxq
** rxqs
= NULL
;
3474 struct rr_numa
*numa
= NULL
;
3477 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3478 if (!netdev_is_pmd(port
->netdev
)) {
3482 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
3483 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
3485 if (pinned
&& q
->core_id
!= OVS_CORE_UNSPEC
) {
3486 struct dp_netdev_pmd_thread
*pmd
;
3488 pmd
= dp_netdev_get_pmd(dp
, q
->core_id
);
3490 VLOG_WARN("There is no PMD thread on core %d. Queue "
3491 "%d on port \'%s\' will not be polled.",
3492 q
->core_id
, qid
, netdev_get_name(port
->netdev
));
3495 pmd
->isolated
= true;
3496 dp_netdev_pmd_unref(pmd
);
3498 } else if (!pinned
&& q
->core_id
== OVS_CORE_UNSPEC
) {
3500 rxqs
= xmalloc(sizeof *rxqs
);
3502 rxqs
= xrealloc(rxqs
, sizeof *rxqs
* (n_rxqs
+ 1));
3504 /* Store the queue. */
3511 /* Sort the queues in order of the processing cycles
3512 * they consumed during their last pmd interval. */
3513 qsort(rxqs
, n_rxqs
, sizeof *rxqs
, rxq_cycle_sort
);
3516 rr_numa_list_populate(dp
, &rr
);
3517 /* Assign the sorted queues to pmds in round robin. */
3518 for (i
= 0; i
< n_rxqs
; i
++) {
3519 numa_id
= netdev_get_numa_id(rxqs
[i
]->port
->netdev
);
3520 numa
= rr_numa_list_lookup(&rr
, numa_id
);
3522 /* There are no pmds on the queue's local NUMA node.
3523 Round robin on the NUMA nodes that do have pmds. */
3524 non_local_numa
= rr_numa_list_next(&rr
, non_local_numa
);
3525 if (!non_local_numa
) {
3526 VLOG_ERR("There is no available (non-isolated) pmd "
3527 "thread for port \'%s\' queue %d. This queue "
3528 "will not be polled. Is pmd-cpu-mask set to "
3529 "zero? Or are all PMDs isolated to other "
3530 "queues?", netdev_rxq_get_name(rxqs
[i
]->rx
),
3531 netdev_rxq_get_queue_id(rxqs
[i
]->rx
));
3534 rxqs
[i
]->pmd
= rr_numa_get_pmd(non_local_numa
);
3535 VLOG_WARN("There's no available (non-isolated) pmd thread "
3536 "on numa node %d. Queue %d on port \'%s\' will "
3537 "be assigned to the pmd on core %d "
3538 "(numa node %d). Expect reduced performance.",
3539 numa_id
, netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
3540 netdev_rxq_get_name(rxqs
[i
]->rx
),
3541 rxqs
[i
]->pmd
->core_id
, rxqs
[i
]->pmd
->numa_id
);
3543 rxqs
[i
]->pmd
= rr_numa_get_pmd(numa
);
3544 VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
3545 "rx queue %d (measured processing cycles %"PRIu64
").",
3546 rxqs
[i
]->pmd
->core_id
, numa_id
,
3547 netdev_rxq_get_name(rxqs
[i
]->rx
),
3548 netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
3549 dp_netdev_rxq_get_cycles(rxqs
[i
], RXQ_CYCLES_PROC_HIST
));
3553 rr_numa_list_destroy(&rr
);
3558 reload_affected_pmds(struct dp_netdev
*dp
)
3560 struct dp_netdev_pmd_thread
*pmd
;
3562 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3563 if (pmd
->need_reload
) {
3564 dp_netdev_reload_pmd__(pmd
);
3565 pmd
->need_reload
= false;
3571 reconfigure_pmd_threads(struct dp_netdev
*dp
)
3572 OVS_REQUIRES(dp
->port_mutex
)
3574 struct dp_netdev_pmd_thread
*pmd
;
3575 struct ovs_numa_dump
*pmd_cores
;
3576 struct ovs_numa_info_core
*core
;
3577 struct hmapx to_delete
= HMAPX_INITIALIZER(&to_delete
);
3578 struct hmapx_node
*node
;
3579 bool changed
= false;
3580 bool need_to_adjust_static_tx_qids
= false;
3582 /* The pmd threads should be started only if there's a pmd port in the
3583 * datapath. If the user didn't provide any "pmd-cpu-mask", we start
3584 * NR_PMD_THREADS per numa node. */
3585 if (!has_pmd_port(dp
)) {
3586 pmd_cores
= ovs_numa_dump_n_cores_per_numa(0);
3587 } else if (dp
->pmd_cmask
&& dp
->pmd_cmask
[0]) {
3588 pmd_cores
= ovs_numa_dump_cores_with_cmask(dp
->pmd_cmask
);
3590 pmd_cores
= ovs_numa_dump_n_cores_per_numa(NR_PMD_THREADS
);
3593 /* We need to adjust 'static_tx_qid's only if we're reducing number of
3594 * PMD threads. Otherwise, new threads will allocate all the freed ids. */
3595 if (ovs_numa_dump_count(pmd_cores
) < cmap_count(&dp
->poll_threads
) - 1) {
3596 /* Adjustment is required to keep 'static_tx_qid's sequential and
3597 * avoid possible issues, for example, imbalanced tx queue usage
3598 * and unnecessary locking caused by remapping on netdev level. */
3599 need_to_adjust_static_tx_qids
= true;
3602 /* Check for unwanted pmd threads */
3603 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3604 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3607 if (!ovs_numa_dump_contains_core(pmd_cores
, pmd
->numa_id
,
3609 hmapx_add(&to_delete
, pmd
);
3610 } else if (need_to_adjust_static_tx_qids
) {
3611 pmd
->need_reload
= true;
3615 HMAPX_FOR_EACH (node
, &to_delete
) {
3616 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3617 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d destroyed.",
3618 pmd
->numa_id
, pmd
->core_id
);
3619 dp_netdev_del_pmd(dp
, pmd
);
3621 changed
= !hmapx_is_empty(&to_delete
);
3622 hmapx_destroy(&to_delete
);
3624 if (need_to_adjust_static_tx_qids
) {
3625 /* 'static_tx_qid's are not sequential now.
3626 * Reload remaining threads to fix this. */
3627 reload_affected_pmds(dp
);
3630 /* Check for required new pmd threads */
3631 FOR_EACH_CORE_ON_DUMP(core
, pmd_cores
) {
3632 pmd
= dp_netdev_get_pmd(dp
, core
->core_id
);
3634 pmd
= xzalloc(sizeof *pmd
);
3635 dp_netdev_configure_pmd(pmd
, dp
, core
->core_id
, core
->numa_id
);
3636 pmd
->thread
= ovs_thread_create("pmd", pmd_thread_main
, pmd
);
3637 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d created.",
3638 pmd
->numa_id
, pmd
->core_id
);
3641 dp_netdev_pmd_unref(pmd
);
3646 struct ovs_numa_info_numa
*numa
;
3648 /* Log the number of pmd threads per numa node. */
3649 FOR_EACH_NUMA_ON_DUMP (numa
, pmd_cores
) {
3650 VLOG_INFO("There are %"PRIuSIZE
" pmd threads on numa node %d",
3651 numa
->n_cores
, numa
->numa_id
);
3655 ovs_numa_dump_destroy(pmd_cores
);
3659 pmd_remove_stale_ports(struct dp_netdev
*dp
,
3660 struct dp_netdev_pmd_thread
*pmd
)
3661 OVS_EXCLUDED(pmd
->port_mutex
)
3662 OVS_REQUIRES(dp
->port_mutex
)
3664 struct rxq_poll
*poll
, *poll_next
;
3665 struct tx_port
*tx
, *tx_next
;
3667 ovs_mutex_lock(&pmd
->port_mutex
);
3668 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
3669 struct dp_netdev_port
*port
= poll
->rxq
->port
;
3671 if (port
->need_reconfigure
3672 || !hmap_contains(&dp
->ports
, &port
->node
)) {
3673 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
3676 HMAP_FOR_EACH_SAFE (tx
, tx_next
, node
, &pmd
->tx_ports
) {
3677 struct dp_netdev_port
*port
= tx
->port
;
3679 if (port
->need_reconfigure
3680 || !hmap_contains(&dp
->ports
, &port
->node
)) {
3681 dp_netdev_del_port_tx_from_pmd(pmd
, tx
);
3684 ovs_mutex_unlock(&pmd
->port_mutex
);
3687 /* Must be called each time a port is added/removed or the cmask changes.
3688 * This creates and destroys pmd threads, reconfigures ports, opens their
3689 * rxqs and assigns all rxqs/txqs to pmd threads. */
3691 reconfigure_datapath(struct dp_netdev
*dp
)
3692 OVS_REQUIRES(dp
->port_mutex
)
3694 struct dp_netdev_pmd_thread
*pmd
;
3695 struct dp_netdev_port
*port
;
3698 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
3700 /* Step 1: Adjust the pmd threads based on the datapath ports, the cores
3701 * on the system and the user configuration. */
3702 reconfigure_pmd_threads(dp
);
3704 wanted_txqs
= cmap_count(&dp
->poll_threads
);
3706 /* The number of pmd threads might have changed, or a port can be new:
3707 * adjust the txqs. */
3708 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3709 netdev_set_tx_multiq(port
->netdev
, wanted_txqs
);
3712 /* Step 2: Remove from the pmd threads ports that have been removed or
3713 * need reconfiguration. */
3715 /* Check for all the ports that need reconfiguration. We cache this in
3716 * 'port->need_reconfigure', because netdev_is_reconf_required() can
3717 * change at any time. */
3718 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3719 if (netdev_is_reconf_required(port
->netdev
)) {
3720 port
->need_reconfigure
= true;
3724 /* Remove from the pmd threads all the ports that have been deleted or
3725 * need reconfiguration. */
3726 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3727 pmd_remove_stale_ports(dp
, pmd
);
3730 /* Reload affected pmd threads. We must wait for the pmd threads before
3731 * reconfiguring the ports, because a port cannot be reconfigured while
3732 * it's being used. */
3733 reload_affected_pmds(dp
);
3735 /* Step 3: Reconfigure ports. */
3737 /* We only reconfigure the ports that we determined above, because they're
3738 * not being used by any pmd thread at the moment. If a port fails to
3739 * reconfigure we remove it from the datapath. */
3740 struct dp_netdev_port
*next_port
;
3741 HMAP_FOR_EACH_SAFE (port
, next_port
, node
, &dp
->ports
) {
3744 if (!port
->need_reconfigure
) {
3748 err
= port_reconfigure(port
);
3750 hmap_remove(&dp
->ports
, &port
->node
);
3751 seq_change(dp
->port_seq
);
3754 port
->dynamic_txqs
= netdev_n_txq(port
->netdev
) < wanted_txqs
;
3758 /* Step 4: Compute new rxq scheduling. We don't touch the pmd threads
3759 * for now, we just update the 'pmd' pointer in each rxq to point to the
3760 * wanted thread according to the scheduling policy. */
3762 /* Reset all the pmd threads to non isolated. */
3763 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3764 pmd
->isolated
= false;
3767 /* Reset all the queues to unassigned */
3768 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3769 for (int i
= 0; i
< port
->n_rxq
; i
++) {
3770 port
->rxqs
[i
].pmd
= NULL
;
3774 /* Add pinned queues and mark pmd threads isolated. */
3775 rxq_scheduling(dp
, true);
3777 /* Add non-pinned queues. */
3778 rxq_scheduling(dp
, false);
3780 /* Step 5: Remove queues not compliant with new scheduling. */
3781 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3782 struct rxq_poll
*poll
, *poll_next
;
3784 ovs_mutex_lock(&pmd
->port_mutex
);
3785 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
3786 if (poll
->rxq
->pmd
!= pmd
) {
3787 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
3790 ovs_mutex_unlock(&pmd
->port_mutex
);
3793 /* Reload affected pmd threads. We must wait for the pmd threads to remove
3794 * the old queues before readding them, otherwise a queue can be polled by
3795 * two threads at the same time. */
3796 reload_affected_pmds(dp
);
3798 /* Step 6: Add queues from scheduling, if they're not there already. */
3799 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3800 if (!netdev_is_pmd(port
->netdev
)) {
3804 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
3805 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
3808 ovs_mutex_lock(&q
->pmd
->port_mutex
);
3809 dp_netdev_add_rxq_to_pmd(q
->pmd
, q
);
3810 ovs_mutex_unlock(&q
->pmd
->port_mutex
);
3815 /* Add every port to the tx cache of every pmd thread, if it's not
3816 * there already and if this pmd has at least one rxq to poll. */
3817 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3818 ovs_mutex_lock(&pmd
->port_mutex
);
3819 if (hmap_count(&pmd
->poll_list
) || pmd
->core_id
== NON_PMD_CORE_ID
) {
3820 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3821 dp_netdev_add_port_tx_to_pmd(pmd
, port
);
3824 ovs_mutex_unlock(&pmd
->port_mutex
);
3827 /* Reload affected pmd threads. */
3828 reload_affected_pmds(dp
);
3831 /* Returns true if one of the netdevs in 'dp' requires a reconfiguration */
3833 ports_require_restart(const struct dp_netdev
*dp
)
3834 OVS_REQUIRES(dp
->port_mutex
)
3836 struct dp_netdev_port
*port
;
3838 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3839 if (netdev_is_reconf_required(port
->netdev
)) {
3847 /* Return true if needs to revalidate datapath flows. */
3849 dpif_netdev_run(struct dpif
*dpif
)
3851 struct dp_netdev_port
*port
;
3852 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3853 struct dp_netdev_pmd_thread
*non_pmd
;
3854 uint64_t new_tnl_seq
;
3855 int process_packets
= 0;
3857 ovs_mutex_lock(&dp
->port_mutex
);
3858 non_pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
3860 ovs_mutex_lock(&dp
->non_pmd_mutex
);
3861 cycles_count_start(non_pmd
);
3862 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3863 if (!netdev_is_pmd(port
->netdev
)) {
3866 for (i
= 0; i
< port
->n_rxq
; i
++) {
3868 dp_netdev_process_rxq_port(non_pmd
,
3871 cycles_count_intermediate(non_pmd
, NULL
,
3873 ? PMD_CYCLES_PROCESSING
3878 cycles_count_end(non_pmd
, PMD_CYCLES_IDLE
);
3879 dpif_netdev_xps_revalidate_pmd(non_pmd
, time_msec(), false);
3880 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
3882 dp_netdev_pmd_unref(non_pmd
);
3885 if (dp_netdev_is_reconf_required(dp
) || ports_require_restart(dp
)) {
3886 reconfigure_datapath(dp
);
3888 ovs_mutex_unlock(&dp
->port_mutex
);
3890 tnl_neigh_cache_run();
3892 new_tnl_seq
= seq_read(tnl_conf_seq
);
3894 if (dp
->last_tnl_conf_seq
!= new_tnl_seq
) {
3895 dp
->last_tnl_conf_seq
= new_tnl_seq
;
3902 dpif_netdev_wait(struct dpif
*dpif
)
3904 struct dp_netdev_port
*port
;
3905 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3907 ovs_mutex_lock(&dp_netdev_mutex
);
3908 ovs_mutex_lock(&dp
->port_mutex
);
3909 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3910 netdev_wait_reconf_required(port
->netdev
);
3911 if (!netdev_is_pmd(port
->netdev
)) {
3914 for (i
= 0; i
< port
->n_rxq
; i
++) {
3915 netdev_rxq_wait(port
->rxqs
[i
].rx
);
3919 ovs_mutex_unlock(&dp
->port_mutex
);
3920 ovs_mutex_unlock(&dp_netdev_mutex
);
3921 seq_wait(tnl_conf_seq
, dp
->last_tnl_conf_seq
);
3925 pmd_free_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
3927 struct tx_port
*tx_port_cached
;
3929 /* Free all used tx queue ids. */
3930 dpif_netdev_xps_revalidate_pmd(pmd
, 0, true);
3932 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->tnl_port_cache
) {
3933 free(tx_port_cached
);
3935 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->send_port_cache
) {
3936 free(tx_port_cached
);
3940 /* Copies ports from 'pmd->tx_ports' (shared with the main thread) to
3941 * thread-local copies. Copy to 'pmd->tnl_port_cache' if it is a tunnel
3942 * device, otherwise to 'pmd->send_port_cache' if the port has at least
3945 pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
3946 OVS_REQUIRES(pmd
->port_mutex
)
3948 struct tx_port
*tx_port
, *tx_port_cached
;
3950 pmd_free_cached_ports(pmd
);
3951 hmap_shrink(&pmd
->send_port_cache
);
3952 hmap_shrink(&pmd
->tnl_port_cache
);
3954 HMAP_FOR_EACH (tx_port
, node
, &pmd
->tx_ports
) {
3955 if (netdev_has_tunnel_push_pop(tx_port
->port
->netdev
)) {
3956 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
3957 hmap_insert(&pmd
->tnl_port_cache
, &tx_port_cached
->node
,
3958 hash_port_no(tx_port_cached
->port
->port_no
));
3961 if (netdev_n_txq(tx_port
->port
->netdev
)) {
3962 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
3963 hmap_insert(&pmd
->send_port_cache
, &tx_port_cached
->node
,
3964 hash_port_no(tx_port_cached
->port
->port_no
));
3970 pmd_alloc_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
3972 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
3973 if (!id_pool_alloc_id(pmd
->dp
->tx_qid_pool
, &pmd
->static_tx_qid
)) {
3974 VLOG_ABORT("static_tx_qid allocation failed for PMD on core %2d"
3975 ", numa_id %d.", pmd
->core_id
, pmd
->numa_id
);
3977 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
3979 VLOG_DBG("static_tx_qid = %d allocated for PMD thread on core %2d"
3980 ", numa_id %d.", pmd
->static_tx_qid
, pmd
->core_id
, pmd
->numa_id
);
3984 pmd_free_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
3986 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
3987 id_pool_free_id(pmd
->dp
->tx_qid_pool
, pmd
->static_tx_qid
);
3988 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
3992 pmd_load_queues_and_ports(struct dp_netdev_pmd_thread
*pmd
,
3993 struct polled_queue
**ppoll_list
)
3995 struct polled_queue
*poll_list
= *ppoll_list
;
3996 struct rxq_poll
*poll
;
3999 ovs_mutex_lock(&pmd
->port_mutex
);
4000 poll_list
= xrealloc(poll_list
, hmap_count(&pmd
->poll_list
)
4001 * sizeof *poll_list
);
4004 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
4005 poll_list
[i
].rxq
= poll
->rxq
;
4006 poll_list
[i
].port_no
= poll
->rxq
->port
->port_no
;
4010 pmd_load_cached_ports(pmd
);
4012 ovs_mutex_unlock(&pmd
->port_mutex
);
4014 *ppoll_list
= poll_list
;
4019 pmd_thread_main(void *f_
)
4021 struct dp_netdev_pmd_thread
*pmd
= f_
;
4022 unsigned int lc
= 0;
4023 struct polled_queue
*poll_list
;
4027 int process_packets
= 0;
4031 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
4032 ovsthread_setspecific(pmd
->dp
->per_pmd_key
, pmd
);
4033 ovs_numa_thread_setaffinity_core(pmd
->core_id
);
4034 dpdk_set_lcore_id(pmd
->core_id
);
4035 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
4036 emc_cache_init(&pmd
->flow_cache
);
4038 pmd_alloc_static_tx_qid(pmd
);
4040 /* List port/core affinity */
4041 for (i
= 0; i
< poll_cnt
; i
++) {
4042 VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n",
4043 pmd
->core_id
, netdev_rxq_get_name(poll_list
[i
].rxq
->rx
),
4044 netdev_rxq_get_queue_id(poll_list
[i
].rxq
->rx
));
4048 while (seq_read(pmd
->reload_seq
) == pmd
->last_reload_seq
) {
4049 seq_wait(pmd
->reload_seq
, pmd
->last_reload_seq
);
4055 cycles_count_start(pmd
);
4057 for (i
= 0; i
< poll_cnt
; i
++) {
4059 dp_netdev_process_rxq_port(pmd
, poll_list
[i
].rxq
->rx
,
4060 poll_list
[i
].port_no
);
4061 cycles_count_intermediate(pmd
, poll_list
[i
].rxq
,
4062 process_packets
? PMD_CYCLES_PROCESSING
4071 coverage_try_clear();
4072 dp_netdev_pmd_try_optimize(pmd
, poll_list
, poll_cnt
);
4073 if (!ovsrcu_try_quiesce()) {
4074 emc_cache_slow_sweep(&pmd
->flow_cache
);
4077 atomic_read_relaxed(&pmd
->reload
, &reload
);
4084 cycles_count_end(pmd
, PMD_CYCLES_IDLE
);
4086 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
4087 exiting
= latch_is_set(&pmd
->exit_latch
);
4088 /* Signal here to make sure the pmd finishes
4089 * reloading the updated configuration. */
4090 dp_netdev_pmd_reload_done(pmd
);
4092 pmd_free_static_tx_qid(pmd
);
4098 emc_cache_uninit(&pmd
->flow_cache
);
4100 pmd_free_cached_ports(pmd
);
4105 dp_netdev_disable_upcall(struct dp_netdev
*dp
)
4106 OVS_ACQUIRES(dp
->upcall_rwlock
)
4108 fat_rwlock_wrlock(&dp
->upcall_rwlock
);
4114 dpif_netdev_meter_get_features(const struct dpif
* dpif OVS_UNUSED
,
4115 struct ofputil_meter_features
*features
)
4117 features
->max_meters
= MAX_METERS
;
4118 features
->band_types
= DP_SUPPORTED_METER_BAND_TYPES
;
4119 features
->capabilities
= DP_SUPPORTED_METER_FLAGS_MASK
;
4120 features
->max_bands
= MAX_BANDS
;
4121 features
->max_color
= 0;
4124 /* Returns false when packet needs to be dropped. */
4126 dp_netdev_run_meter(struct dp_netdev
*dp
, struct dp_packet_batch
*packets_
,
4127 uint32_t meter_id
, long long int now
)
4129 struct dp_meter
*meter
;
4130 struct dp_meter_band
*band
;
4131 struct dp_packet
*packet
;
4132 long long int long_delta_t
; /* msec */
4133 uint32_t delta_t
; /* msec */
4135 const size_t cnt
= dp_packet_batch_size(packets_
);
4136 uint32_t bytes
, volume
;
4137 int exceeded_band
[NETDEV_MAX_BURST
];
4138 uint32_t exceeded_rate
[NETDEV_MAX_BURST
];
4139 int exceeded_pkt
= cnt
; /* First packet that exceeded a band rate. */
4141 if (meter_id
>= MAX_METERS
) {
4145 meter_lock(dp
, meter_id
);
4146 meter
= dp
->meters
[meter_id
];
4151 /* Initialize as negative values. */
4152 memset(exceeded_band
, 0xff, cnt
* sizeof *exceeded_band
);
4153 /* Initialize as zeroes. */
4154 memset(exceeded_rate
, 0, cnt
* sizeof *exceeded_rate
);
4156 /* All packets will hit the meter at the same time. */
4157 long_delta_t
= (now
- meter
->used
); /* msec */
4159 /* Make sure delta_t will not be too large, so that bucket will not
4160 * wrap around below. */
4161 delta_t
= (long_delta_t
> (long long int)meter
->max_delta_t
)
4162 ? meter
->max_delta_t
: (uint32_t)long_delta_t
;
4164 /* Update meter stats. */
4166 meter
->packet_count
+= cnt
;
4168 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
4169 bytes
+= dp_packet_size(packet
);
4171 meter
->byte_count
+= bytes
;
4173 /* Meters can operate in terms of packets per second or kilobits per
4175 if (meter
->flags
& OFPMF13_PKTPS
) {
4176 /* Rate in packets/second, bucket 1/1000 packets. */
4177 /* msec * packets/sec = 1/1000 packets. */
4178 volume
= cnt
* 1000; /* Take 'cnt' packets from the bucket. */
4180 /* Rate in kbps, bucket in bits. */
4181 /* msec * kbps = bits */
4185 /* Update all bands and find the one hit with the highest rate for each
4186 * packet (if any). */
4187 for (int m
= 0; m
< meter
->n_bands
; ++m
) {
4188 band
= &meter
->bands
[m
];
4190 /* Update band's bucket. */
4191 band
->bucket
+= delta_t
* band
->up
.rate
;
4192 if (band
->bucket
> band
->up
.burst_size
) {
4193 band
->bucket
= band
->up
.burst_size
;
4196 /* Drain the bucket for all the packets, if possible. */
4197 if (band
->bucket
>= volume
) {
4198 band
->bucket
-= volume
;
4200 int band_exceeded_pkt
;
4202 /* Band limit hit, must process packet-by-packet. */
4203 if (meter
->flags
& OFPMF13_PKTPS
) {
4204 band_exceeded_pkt
= band
->bucket
/ 1000;
4205 band
->bucket
%= 1000; /* Remainder stays in bucket. */
4207 /* Update the exceeding band for each exceeding packet.
4208 * (Only one band will be fired by a packet, and that
4209 * can be different for each packet.) */
4210 for (i
= band_exceeded_pkt
; i
< cnt
; i
++) {
4211 if (band
->up
.rate
> exceeded_rate
[i
]) {
4212 exceeded_rate
[i
] = band
->up
.rate
;
4213 exceeded_band
[i
] = m
;
4217 /* Packet sizes differ, must process one-by-one. */
4218 band_exceeded_pkt
= cnt
;
4219 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
4220 uint32_t bits
= dp_packet_size(packet
) * 8;
4222 if (band
->bucket
>= bits
) {
4223 band
->bucket
-= bits
;
4225 if (i
< band_exceeded_pkt
) {
4226 band_exceeded_pkt
= i
;
4228 /* Update the exceeding band for the exceeding packet.
4229 * (Only one band will be fired by a packet, and that
4230 * can be different for each packet.) */
4231 if (band
->up
.rate
> exceeded_rate
[i
]) {
4232 exceeded_rate
[i
] = band
->up
.rate
;
4233 exceeded_band
[i
] = m
;
4238 /* Remember the first exceeding packet. */
4239 if (exceeded_pkt
> band_exceeded_pkt
) {
4240 exceeded_pkt
= band_exceeded_pkt
;
4245 /* Fire the highest rate band exceeded by each packet.
4246 * Drop packets if needed, by swapping packet to the end that will be
4249 DP_PACKET_BATCH_REFILL_FOR_EACH (j
, cnt
, packet
, packets_
) {
4250 if (exceeded_band
[j
] >= 0) {
4251 /* Meter drop packet. */
4252 band
= &meter
->bands
[exceeded_band
[j
]];
4253 band
->packet_count
+= 1;
4254 band
->byte_count
+= dp_packet_size(packet
);
4256 dp_packet_delete(packet
);
4258 /* Meter accepts packet. */
4259 dp_packet_batch_refill(packets_
, packet
, j
);
4263 meter_unlock(dp
, meter_id
);
4266 /* Meter set/get/del processing is still single-threaded. */
4268 dpif_netdev_meter_set(struct dpif
*dpif
, ofproto_meter_id
*meter_id
,
4269 struct ofputil_meter_config
*config
)
4271 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4272 uint32_t mid
= meter_id
->uint32
;
4273 struct dp_meter
*meter
;
4276 if (mid
>= MAX_METERS
) {
4277 return EFBIG
; /* Meter_id out of range. */
4280 if (config
->flags
& ~DP_SUPPORTED_METER_FLAGS_MASK
||
4281 !(config
->flags
& (OFPMF13_KBPS
| OFPMF13_PKTPS
))) {
4282 return EBADF
; /* Unsupported flags set */
4285 /* Validate bands */
4286 if (config
->n_bands
== 0 || config
->n_bands
> MAX_BANDS
) {
4287 return EINVAL
; /* Too many bands */
4290 /* Validate rates */
4291 for (i
= 0; i
< config
->n_bands
; i
++) {
4292 if (config
->bands
[i
].rate
== 0) {
4293 return EDOM
; /* rate must be non-zero */
4297 for (i
= 0; i
< config
->n_bands
; ++i
) {
4298 switch (config
->bands
[i
].type
) {
4302 return ENODEV
; /* Unsupported band type */
4306 /* Allocate meter */
4307 meter
= xzalloc(sizeof *meter
4308 + config
->n_bands
* sizeof(struct dp_meter_band
));
4310 meter
->flags
= config
->flags
;
4311 meter
->n_bands
= config
->n_bands
;
4312 meter
->max_delta_t
= 0;
4313 meter
->used
= time_msec();
4316 for (i
= 0; i
< config
->n_bands
; ++i
) {
4317 uint32_t band_max_delta_t
;
4319 /* Set burst size to a workable value if none specified. */
4320 if (config
->bands
[i
].burst_size
== 0) {
4321 config
->bands
[i
].burst_size
= config
->bands
[i
].rate
;
4324 meter
->bands
[i
].up
= config
->bands
[i
];
4325 /* Convert burst size to the bucket units: */
4326 /* pkts => 1/1000 packets, kilobits => bits. */
4327 meter
->bands
[i
].up
.burst_size
*= 1000;
4328 /* Initialize bucket to empty. */
4329 meter
->bands
[i
].bucket
= 0;
4331 /* Figure out max delta_t that is enough to fill any bucket. */
4333 = meter
->bands
[i
].up
.burst_size
/ meter
->bands
[i
].up
.rate
;
4334 if (band_max_delta_t
> meter
->max_delta_t
) {
4335 meter
->max_delta_t
= band_max_delta_t
;
4339 meter_lock(dp
, mid
);
4340 dp_delete_meter(dp
, mid
); /* Free existing meter, if any */
4341 dp
->meters
[mid
] = meter
;
4342 meter_unlock(dp
, mid
);
4350 dpif_netdev_meter_get(const struct dpif
*dpif
,
4351 ofproto_meter_id meter_id_
,
4352 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
4354 const struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4355 const struct dp_meter
*meter
;
4356 uint32_t meter_id
= meter_id_
.uint32
;
4358 if (meter_id
>= MAX_METERS
) {
4361 meter
= dp
->meters
[meter_id
];
4368 meter_lock(dp
, meter_id
);
4369 stats
->packet_in_count
= meter
->packet_count
;
4370 stats
->byte_in_count
= meter
->byte_count
;
4372 for (i
= 0; i
< n_bands
&& i
< meter
->n_bands
; ++i
) {
4373 stats
->bands
[i
].packet_count
= meter
->bands
[i
].packet_count
;
4374 stats
->bands
[i
].byte_count
= meter
->bands
[i
].byte_count
;
4376 meter_unlock(dp
, meter_id
);
4384 dpif_netdev_meter_del(struct dpif
*dpif
,
4385 ofproto_meter_id meter_id_
,
4386 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
4388 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4391 error
= dpif_netdev_meter_get(dpif
, meter_id_
, stats
, n_bands
);
4393 uint32_t meter_id
= meter_id_
.uint32
;
4395 meter_lock(dp
, meter_id
);
4396 dp_delete_meter(dp
, meter_id
);
4397 meter_unlock(dp
, meter_id
);
4404 dpif_netdev_disable_upcall(struct dpif
*dpif
)
4405 OVS_NO_THREAD_SAFETY_ANALYSIS
4407 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4408 dp_netdev_disable_upcall(dp
);
4412 dp_netdev_enable_upcall(struct dp_netdev
*dp
)
4413 OVS_RELEASES(dp
->upcall_rwlock
)
4415 fat_rwlock_unlock(&dp
->upcall_rwlock
);
4419 dpif_netdev_enable_upcall(struct dpif
*dpif
)
4420 OVS_NO_THREAD_SAFETY_ANALYSIS
4422 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4423 dp_netdev_enable_upcall(dp
);
4427 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
)
4429 ovs_mutex_lock(&pmd
->cond_mutex
);
4430 atomic_store_relaxed(&pmd
->reload
, false);
4431 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
4432 xpthread_cond_signal(&pmd
->cond
);
4433 ovs_mutex_unlock(&pmd
->cond_mutex
);
4436 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
4437 * the pointer if succeeds, otherwise, NULL (it can return NULL even if
4438 * 'core_id' is NON_PMD_CORE_ID).
4440 * Caller must unrefs the returned reference. */
4441 static struct dp_netdev_pmd_thread
*
4442 dp_netdev_get_pmd(struct dp_netdev
*dp
, unsigned core_id
)
4444 struct dp_netdev_pmd_thread
*pmd
;
4445 const struct cmap_node
*pnode
;
4447 pnode
= cmap_find(&dp
->poll_threads
, hash_int(core_id
, 0));
4451 pmd
= CONTAINER_OF(pnode
, struct dp_netdev_pmd_thread
, node
);
4453 return dp_netdev_pmd_try_ref(pmd
) ? pmd
: NULL
;
4456 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
4458 dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
4459 OVS_REQUIRES(dp
->port_mutex
)
4461 struct dp_netdev_pmd_thread
*non_pmd
;
4463 non_pmd
= xzalloc(sizeof *non_pmd
);
4464 dp_netdev_configure_pmd(non_pmd
, dp
, NON_PMD_CORE_ID
, OVS_NUMA_UNSPEC
);
4467 /* Caller must have valid pointer to 'pmd'. */
4469 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
)
4471 return ovs_refcount_try_ref_rcu(&pmd
->ref_cnt
);
4475 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
)
4477 if (pmd
&& ovs_refcount_unref(&pmd
->ref_cnt
) == 1) {
4478 ovsrcu_postpone(dp_netdev_destroy_pmd
, pmd
);
4482 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
4483 * fails, keeps checking for next node until reaching the end of cmap.
4485 * Caller must unrefs the returned reference. */
4486 static struct dp_netdev_pmd_thread
*
4487 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
)
4489 struct dp_netdev_pmd_thread
*next
;
4492 struct cmap_node
*node
;
4494 node
= cmap_next_position(&dp
->poll_threads
, pos
);
4495 next
= node
? CONTAINER_OF(node
, struct dp_netdev_pmd_thread
, node
)
4497 } while (next
&& !dp_netdev_pmd_try_ref(next
));
4502 /* Configures the 'pmd' based on the input argument. */
4504 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
, struct dp_netdev
*dp
,
4505 unsigned core_id
, int numa_id
)
4508 pmd
->core_id
= core_id
;
4509 pmd
->numa_id
= numa_id
;
4510 pmd
->need_reload
= false;
4512 ovs_refcount_init(&pmd
->ref_cnt
);
4513 latch_init(&pmd
->exit_latch
);
4514 pmd
->reload_seq
= seq_create();
4515 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
4516 atomic_init(&pmd
->reload
, false);
4517 xpthread_cond_init(&pmd
->cond
, NULL
);
4518 ovs_mutex_init(&pmd
->cond_mutex
);
4519 ovs_mutex_init(&pmd
->flow_mutex
);
4520 ovs_mutex_init(&pmd
->port_mutex
);
4521 cmap_init(&pmd
->flow_table
);
4522 cmap_init(&pmd
->classifiers
);
4523 pmd
->next_optimization
= time_msec() + DPCLS_OPTIMIZATION_INTERVAL
;
4524 pmd
->rxq_interval
= time_msec() + PMD_RXQ_INTERVAL_LEN
;
4525 hmap_init(&pmd
->poll_list
);
4526 hmap_init(&pmd
->tx_ports
);
4527 hmap_init(&pmd
->tnl_port_cache
);
4528 hmap_init(&pmd
->send_port_cache
);
4529 /* init the 'flow_cache' since there is no
4530 * actual thread created for NON_PMD_CORE_ID. */
4531 if (core_id
== NON_PMD_CORE_ID
) {
4532 emc_cache_init(&pmd
->flow_cache
);
4533 pmd_alloc_static_tx_qid(pmd
);
4535 cmap_insert(&dp
->poll_threads
, CONST_CAST(struct cmap_node
*, &pmd
->node
),
4536 hash_int(core_id
, 0));
4540 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
)
4544 dp_netdev_pmd_flow_flush(pmd
);
4545 hmap_destroy(&pmd
->send_port_cache
);
4546 hmap_destroy(&pmd
->tnl_port_cache
);
4547 hmap_destroy(&pmd
->tx_ports
);
4548 hmap_destroy(&pmd
->poll_list
);
4549 /* All flows (including their dpcls_rules) have been deleted already */
4550 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
4552 ovsrcu_postpone(free
, cls
);
4554 cmap_destroy(&pmd
->classifiers
);
4555 cmap_destroy(&pmd
->flow_table
);
4556 ovs_mutex_destroy(&pmd
->flow_mutex
);
4557 latch_destroy(&pmd
->exit_latch
);
4558 seq_destroy(pmd
->reload_seq
);
4559 xpthread_cond_destroy(&pmd
->cond
);
4560 ovs_mutex_destroy(&pmd
->cond_mutex
);
4561 ovs_mutex_destroy(&pmd
->port_mutex
);
4565 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
4566 * and unrefs the struct. */
4568 dp_netdev_del_pmd(struct dp_netdev
*dp
, struct dp_netdev_pmd_thread
*pmd
)
4570 /* NON_PMD_CORE_ID doesn't have a thread, so we don't have to synchronize,
4571 * but extra cleanup is necessary */
4572 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
4573 ovs_mutex_lock(&dp
->non_pmd_mutex
);
4574 emc_cache_uninit(&pmd
->flow_cache
);
4575 pmd_free_cached_ports(pmd
);
4576 pmd_free_static_tx_qid(pmd
);
4577 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
4579 latch_set(&pmd
->exit_latch
);
4580 dp_netdev_reload_pmd__(pmd
);
4581 xpthread_join(pmd
->thread
, NULL
);
4584 dp_netdev_pmd_clear_ports(pmd
);
4586 /* Purges the 'pmd''s flows after stopping the thread, but before
4587 * destroying the flows, so that the flow stats can be collected. */
4588 if (dp
->dp_purge_cb
) {
4589 dp
->dp_purge_cb(dp
->dp_purge_aux
, pmd
->core_id
);
4591 cmap_remove(&pmd
->dp
->poll_threads
, &pmd
->node
, hash_int(pmd
->core_id
, 0));
4592 dp_netdev_pmd_unref(pmd
);
4595 /* Destroys all pmd threads. If 'non_pmd' is true it also destroys the non pmd
4598 dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
, bool non_pmd
)
4600 struct dp_netdev_pmd_thread
*pmd
;
4601 struct dp_netdev_pmd_thread
**pmd_list
;
4602 size_t k
= 0, n_pmds
;
4604 n_pmds
= cmap_count(&dp
->poll_threads
);
4605 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
4607 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4608 if (!non_pmd
&& pmd
->core_id
== NON_PMD_CORE_ID
) {
4611 /* We cannot call dp_netdev_del_pmd(), since it alters
4612 * 'dp->poll_threads' (while we're iterating it) and it
4614 ovs_assert(k
< n_pmds
);
4615 pmd_list
[k
++] = pmd
;
4618 for (size_t i
= 0; i
< k
; i
++) {
4619 dp_netdev_del_pmd(dp
, pmd_list
[i
]);
4624 /* Deletes all rx queues from pmd->poll_list and all the ports from
4627 dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
)
4629 struct rxq_poll
*poll
;
4630 struct tx_port
*port
;
4632 ovs_mutex_lock(&pmd
->port_mutex
);
4633 HMAP_FOR_EACH_POP (poll
, node
, &pmd
->poll_list
) {
4636 HMAP_FOR_EACH_POP (port
, node
, &pmd
->tx_ports
) {
4639 ovs_mutex_unlock(&pmd
->port_mutex
);
4642 /* Adds rx queue to poll_list of PMD thread, if it's not there already. */
4644 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
4645 struct dp_netdev_rxq
*rxq
)
4646 OVS_REQUIRES(pmd
->port_mutex
)
4648 int qid
= netdev_rxq_get_queue_id(rxq
->rx
);
4649 uint32_t hash
= hash_2words(odp_to_u32(rxq
->port
->port_no
), qid
);
4650 struct rxq_poll
*poll
;
4652 HMAP_FOR_EACH_WITH_HASH (poll
, node
, hash
, &pmd
->poll_list
) {
4653 if (poll
->rxq
== rxq
) {
4654 /* 'rxq' is already polled by this thread. Do nothing. */
4659 poll
= xmalloc(sizeof *poll
);
4661 hmap_insert(&pmd
->poll_list
, &poll
->node
, hash
);
4663 pmd
->need_reload
= true;
4666 /* Delete 'poll' from poll_list of PMD thread. */
4668 dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
4669 struct rxq_poll
*poll
)
4670 OVS_REQUIRES(pmd
->port_mutex
)
4672 hmap_remove(&pmd
->poll_list
, &poll
->node
);
4675 pmd
->need_reload
= true;
4678 /* Add 'port' to the tx port cache of 'pmd', which must be reloaded for the
4679 * changes to take effect. */
4681 dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
4682 struct dp_netdev_port
*port
)
4683 OVS_REQUIRES(pmd
->port_mutex
)
4687 tx
= tx_port_lookup(&pmd
->tx_ports
, port
->port_no
);
4689 /* 'port' is already on this thread tx cache. Do nothing. */
4693 tx
= xzalloc(sizeof *tx
);
4698 hmap_insert(&pmd
->tx_ports
, &tx
->node
, hash_port_no(tx
->port
->port_no
));
4699 pmd
->need_reload
= true;
4702 /* Del 'tx' from the tx port cache of 'pmd', which must be reloaded for the
4703 * changes to take effect. */
4705 dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
4707 OVS_REQUIRES(pmd
->port_mutex
)
4709 hmap_remove(&pmd
->tx_ports
, &tx
->node
);
4711 pmd
->need_reload
= true;
4715 dpif_netdev_get_datapath_version(void)
4717 return xstrdup("<built-in>");
4721 dp_netdev_flow_used(struct dp_netdev_flow
*netdev_flow
, int cnt
, int size
,
4722 uint16_t tcp_flags
, long long now
)
4726 atomic_store_relaxed(&netdev_flow
->stats
.used
, now
);
4727 non_atomic_ullong_add(&netdev_flow
->stats
.packet_count
, cnt
);
4728 non_atomic_ullong_add(&netdev_flow
->stats
.byte_count
, size
);
4729 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
4731 atomic_store_relaxed(&netdev_flow
->stats
.tcp_flags
, flags
);
4735 dp_netdev_count_packet(struct dp_netdev_pmd_thread
*pmd
,
4736 enum dp_stat_type type
, int cnt
)
4738 non_atomic_ullong_add(&pmd
->stats
.n
[type
], cnt
);
4742 dp_netdev_upcall(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
*packet_
,
4743 struct flow
*flow
, struct flow_wildcards
*wc
, ovs_u128
*ufid
,
4744 enum dpif_upcall_type type
, const struct nlattr
*userdata
,
4745 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
4747 struct dp_netdev
*dp
= pmd
->dp
;
4749 if (OVS_UNLIKELY(!dp
->upcall_cb
)) {
4753 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl
))) {
4754 struct ds ds
= DS_EMPTY_INITIALIZER
;
4757 struct odp_flow_key_parms odp_parms
= {
4759 .mask
= wc
? &wc
->masks
: NULL
,
4760 .support
= dp_netdev_support
,
4763 ofpbuf_init(&key
, 0);
4764 odp_flow_key_from_flow(&odp_parms
, &key
);
4765 packet_str
= ofp_dp_packet_to_string(packet_
);
4767 odp_flow_key_format(key
.data
, key
.size
, &ds
);
4769 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp
->name
,
4770 dpif_upcall_type_to_string(type
), ds_cstr(&ds
), packet_str
);
4772 ofpbuf_uninit(&key
);
4778 return dp
->upcall_cb(packet_
, flow
, ufid
, pmd
->core_id
, type
, userdata
,
4779 actions
, wc
, put_actions
, dp
->upcall_aux
);
4782 static inline uint32_t
4783 dpif_netdev_packet_get_rss_hash_orig_pkt(struct dp_packet
*packet
,
4784 const struct miniflow
*mf
)
4788 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
4789 hash
= dp_packet_get_rss_hash(packet
);
4791 hash
= miniflow_hash_5tuple(mf
, 0);
4792 dp_packet_set_rss_hash(packet
, hash
);
4798 static inline uint32_t
4799 dpif_netdev_packet_get_rss_hash(struct dp_packet
*packet
,
4800 const struct miniflow
*mf
)
4802 uint32_t hash
, recirc_depth
;
4804 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
4805 hash
= dp_packet_get_rss_hash(packet
);
4807 hash
= miniflow_hash_5tuple(mf
, 0);
4808 dp_packet_set_rss_hash(packet
, hash
);
4811 /* The RSS hash must account for the recirculation depth to avoid
4812 * collisions in the exact match cache */
4813 recirc_depth
= *recirc_depth_get_unsafe();
4814 if (OVS_UNLIKELY(recirc_depth
)) {
4815 hash
= hash_finish(hash
, recirc_depth
);
4816 dp_packet_set_rss_hash(packet
, hash
);
4821 struct packet_batch_per_flow
{
4822 unsigned int byte_count
;
4824 struct dp_netdev_flow
*flow
;
4826 struct dp_packet_batch array
;
4830 packet_batch_per_flow_update(struct packet_batch_per_flow
*batch
,
4831 struct dp_packet
*packet
,
4832 const struct miniflow
*mf
)
4834 batch
->byte_count
+= dp_packet_size(packet
);
4835 batch
->tcp_flags
|= miniflow_get_tcp_flags(mf
);
4836 batch
->array
.packets
[batch
->array
.count
++] = packet
;
4840 packet_batch_per_flow_init(struct packet_batch_per_flow
*batch
,
4841 struct dp_netdev_flow
*flow
)
4843 flow
->batch
= batch
;
4846 dp_packet_batch_init(&batch
->array
);
4847 batch
->byte_count
= 0;
4848 batch
->tcp_flags
= 0;
4852 packet_batch_per_flow_execute(struct packet_batch_per_flow
*batch
,
4853 struct dp_netdev_pmd_thread
*pmd
,
4856 struct dp_netdev_actions
*actions
;
4857 struct dp_netdev_flow
*flow
= batch
->flow
;
4859 dp_netdev_flow_used(flow
, batch
->array
.count
, batch
->byte_count
,
4860 batch
->tcp_flags
, now
);
4862 actions
= dp_netdev_flow_get_actions(flow
);
4864 dp_netdev_execute_actions(pmd
, &batch
->array
, true, &flow
->flow
,
4865 actions
->actions
, actions
->size
, now
);
4869 dp_netdev_queue_batches(struct dp_packet
*pkt
,
4870 struct dp_netdev_flow
*flow
, const struct miniflow
*mf
,
4871 struct packet_batch_per_flow
*batches
,
4874 struct packet_batch_per_flow
*batch
= flow
->batch
;
4876 if (OVS_UNLIKELY(!batch
)) {
4877 batch
= &batches
[(*n_batches
)++];
4878 packet_batch_per_flow_init(batch
, flow
);
4881 packet_batch_per_flow_update(batch
, pkt
, mf
);
4884 /* Try to process all ('cnt') the 'packets' using only the exact match cache
4885 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
4886 * miniflow is copied into 'keys' and the packet pointer is moved at the
4887 * beginning of the 'packets' array.
4889 * The function returns the number of packets that needs to be processed in the
4890 * 'packets' array (they have been moved to the beginning of the vector).
4892 * For performance reasons a caller may choose not to initialize the metadata
4893 * in 'packets_'. If 'md_is_valid' is false, the metadata in 'packets'
4894 * is not valid and must be initialized by this function using 'port_no'.
4895 * If 'md_is_valid' is true, the metadata is already valid and 'port_no'
4898 static inline size_t
4899 emc_processing(struct dp_netdev_pmd_thread
*pmd
,
4900 struct dp_packet_batch
*packets_
,
4901 struct netdev_flow_key
*keys
,
4902 struct packet_batch_per_flow batches
[], size_t *n_batches
,
4903 bool md_is_valid
, odp_port_t port_no
)
4905 struct emc_cache
*flow_cache
= &pmd
->flow_cache
;
4906 struct netdev_flow_key
*key
= &keys
[0];
4907 size_t n_missed
= 0, n_dropped
= 0;
4908 struct dp_packet
*packet
;
4909 const size_t cnt
= dp_packet_batch_size(packets_
);
4913 atomic_read_relaxed(&pmd
->dp
->emc_insert_min
, &cur_min
);
4915 DP_PACKET_BATCH_REFILL_FOR_EACH (i
, cnt
, packet
, packets_
) {
4916 struct dp_netdev_flow
*flow
;
4918 if (OVS_UNLIKELY(dp_packet_size(packet
) < ETH_HEADER_LEN
)) {
4919 dp_packet_delete(packet
);
4925 struct dp_packet
**packets
= packets_
->packets
;
4926 /* Prefetch next packet data and metadata. */
4927 OVS_PREFETCH(dp_packet_data(packets
[i
+1]));
4928 pkt_metadata_prefetch_init(&packets
[i
+1]->md
);
4932 pkt_metadata_init(&packet
->md
, port_no
);
4934 miniflow_extract(packet
, &key
->mf
);
4935 key
->len
= 0; /* Not computed yet. */
4936 /* If EMC is disabled skip hash computation and emc_lookup */
4939 key
->hash
= dpif_netdev_packet_get_rss_hash_orig_pkt(packet
,
4942 key
->hash
= dpif_netdev_packet_get_rss_hash(packet
, &key
->mf
);
4944 flow
= emc_lookup(flow_cache
, key
);
4948 if (OVS_LIKELY(flow
)) {
4949 dp_netdev_queue_batches(packet
, flow
, &key
->mf
, batches
,
4952 /* Exact match cache missed. Group missed packets together at
4953 * the beginning of the 'packets' array. */
4954 dp_packet_batch_refill(packets_
, packet
, i
);
4955 /* 'key[n_missed]' contains the key of the current packet and it
4956 * must be returned to the caller. The next key should be extracted
4957 * to 'keys[n_missed + 1]'. */
4958 key
= &keys
[++n_missed
];
4962 dp_netdev_count_packet(pmd
, DP_STAT_EXACT_HIT
,
4963 cnt
- n_dropped
- n_missed
);
4965 return dp_packet_batch_size(packets_
);
4969 handle_packet_upcall(struct dp_netdev_pmd_thread
*pmd
,
4970 struct dp_packet
*packet
,
4971 const struct netdev_flow_key
*key
,
4972 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
,
4973 int *lost_cnt
, long long now
)
4975 struct ofpbuf
*add_actions
;
4976 struct dp_packet_batch b
;
4981 match
.tun_md
.valid
= false;
4982 miniflow_expand(&key
->mf
, &match
.flow
);
4984 ofpbuf_clear(actions
);
4985 ofpbuf_clear(put_actions
);
4987 dpif_flow_hash(pmd
->dp
->dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
4988 error
= dp_netdev_upcall(pmd
, packet
, &match
.flow
, &match
.wc
,
4989 &ufid
, DPIF_UC_MISS
, NULL
, actions
,
4991 if (OVS_UNLIKELY(error
&& error
!= ENOSPC
)) {
4992 dp_packet_delete(packet
);
4997 /* The Netlink encoding of datapath flow keys cannot express
4998 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
4999 * tag is interpreted as exact match on the fact that there is no
5000 * VLAN. Unless we refactor a lot of code that translates between
5001 * Netlink and struct flow representations, we have to do the same
5003 if (!match
.wc
.masks
.vlans
[0].tci
) {
5004 match
.wc
.masks
.vlans
[0].tci
= htons(0xffff);
5007 /* We can't allow the packet batching in the next loop to execute
5008 * the actions. Otherwise, if there are any slow path actions,
5009 * we'll send the packet up twice. */
5010 dp_packet_batch_init_packet(&b
, packet
);
5011 dp_netdev_execute_actions(pmd
, &b
, true, &match
.flow
,
5012 actions
->data
, actions
->size
, now
);
5014 add_actions
= put_actions
->size
? put_actions
: actions
;
5015 if (OVS_LIKELY(error
!= ENOSPC
)) {
5016 struct dp_netdev_flow
*netdev_flow
;
5018 /* XXX: There's a race window where a flow covering this packet
5019 * could have already been installed since we last did the flow
5020 * lookup before upcall. This could be solved by moving the
5021 * mutex lock outside the loop, but that's an awful long time
5022 * to be locking everyone out of making flow installs. If we
5023 * move to a per-core classifier, it would be reasonable. */
5024 ovs_mutex_lock(&pmd
->flow_mutex
);
5025 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
5026 if (OVS_LIKELY(!netdev_flow
)) {
5027 netdev_flow
= dp_netdev_flow_add(pmd
, &match
, &ufid
,
5031 ovs_mutex_unlock(&pmd
->flow_mutex
);
5032 emc_probabilistic_insert(pmd
, key
, netdev_flow
);
5037 fast_path_processing(struct dp_netdev_pmd_thread
*pmd
,
5038 struct dp_packet_batch
*packets_
,
5039 struct netdev_flow_key
*keys
,
5040 struct packet_batch_per_flow batches
[], size_t *n_batches
,
5044 const size_t cnt
= dp_packet_batch_size(packets_
);
5045 #if !defined(__CHECKER__) && !defined(_WIN32)
5046 const size_t PKT_ARRAY_SIZE
= cnt
;
5048 /* Sparse or MSVC doesn't like variable length array. */
5049 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
5051 struct dp_packet
*packet
;
5053 struct dpcls_rule
*rules
[PKT_ARRAY_SIZE
];
5054 struct dp_netdev
*dp
= pmd
->dp
;
5055 int miss_cnt
= 0, lost_cnt
= 0;
5056 int lookup_cnt
= 0, add_lookup_cnt
;
5060 for (i
= 0; i
< cnt
; i
++) {
5061 /* Key length is needed in all the cases, hash computed on demand. */
5062 keys
[i
].len
= netdev_flow_key_size(miniflow_n_values(&keys
[i
].mf
));
5064 /* Get the classifier for the in_port */
5065 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
5066 if (OVS_LIKELY(cls
)) {
5067 any_miss
= !dpcls_lookup(cls
, keys
, rules
, cnt
, &lookup_cnt
);
5070 memset(rules
, 0, sizeof(rules
));
5072 if (OVS_UNLIKELY(any_miss
) && !fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
5073 uint64_t actions_stub
[512 / 8], slow_stub
[512 / 8];
5074 struct ofpbuf actions
, put_actions
;
5076 ofpbuf_use_stub(&actions
, actions_stub
, sizeof actions_stub
);
5077 ofpbuf_use_stub(&put_actions
, slow_stub
, sizeof slow_stub
);
5079 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5080 struct dp_netdev_flow
*netdev_flow
;
5082 if (OVS_LIKELY(rules
[i
])) {
5086 /* It's possible that an earlier slow path execution installed
5087 * a rule covering this flow. In this case, it's a lot cheaper
5088 * to catch it here than execute a miss. */
5089 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, &keys
[i
],
5092 lookup_cnt
+= add_lookup_cnt
;
5093 rules
[i
] = &netdev_flow
->cr
;
5098 handle_packet_upcall(pmd
, packet
, &keys
[i
], &actions
,
5099 &put_actions
, &lost_cnt
, now
);
5102 ofpbuf_uninit(&actions
);
5103 ofpbuf_uninit(&put_actions
);
5104 fat_rwlock_unlock(&dp
->upcall_rwlock
);
5105 } else if (OVS_UNLIKELY(any_miss
)) {
5106 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5107 if (OVS_UNLIKELY(!rules
[i
])) {
5108 dp_packet_delete(packet
);
5115 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5116 struct dp_netdev_flow
*flow
;
5118 if (OVS_UNLIKELY(!rules
[i
])) {
5122 flow
= dp_netdev_flow_cast(rules
[i
]);
5124 emc_probabilistic_insert(pmd
, &keys
[i
], flow
);
5125 dp_netdev_queue_batches(packet
, flow
, &keys
[i
].mf
, batches
, n_batches
);
5128 dp_netdev_count_packet(pmd
, DP_STAT_MASKED_HIT
, cnt
- miss_cnt
);
5129 dp_netdev_count_packet(pmd
, DP_STAT_LOOKUP_HIT
, lookup_cnt
);
5130 dp_netdev_count_packet(pmd
, DP_STAT_MISS
, miss_cnt
);
5131 dp_netdev_count_packet(pmd
, DP_STAT_LOST
, lost_cnt
);
5134 /* Packets enter the datapath from a port (or from recirculation) here.
5136 * When 'md_is_valid' is true the metadata in 'packets' are already valid.
5137 * When false the metadata in 'packets' need to be initialized. */
5139 dp_netdev_input__(struct dp_netdev_pmd_thread
*pmd
,
5140 struct dp_packet_batch
*packets
,
5141 bool md_is_valid
, odp_port_t port_no
)
5143 #if !defined(__CHECKER__) && !defined(_WIN32)
5144 const size_t PKT_ARRAY_SIZE
= dp_packet_batch_size(packets
);
5146 /* Sparse or MSVC doesn't like variable length array. */
5147 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
5149 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
)
5150 struct netdev_flow_key keys
[PKT_ARRAY_SIZE
];
5151 struct packet_batch_per_flow batches
[PKT_ARRAY_SIZE
];
5152 long long now
= time_msec();
5157 emc_processing(pmd
, packets
, keys
, batches
, &n_batches
,
5158 md_is_valid
, port_no
);
5159 if (!dp_packet_batch_is_empty(packets
)) {
5160 /* Get ingress port from first packet's metadata. */
5161 in_port
= packets
->packets
[0]->md
.in_port
.odp_port
;
5162 fast_path_processing(pmd
, packets
, keys
, batches
, &n_batches
,
5166 /* All the flow batches need to be reset before any call to
5167 * packet_batch_per_flow_execute() as it could potentially trigger
5168 * recirculation. When a packet matching flow ‘j’ happens to be
5169 * recirculated, the nested call to dp_netdev_input__() could potentially
5170 * classify the packet as matching another flow - say 'k'. It could happen
5171 * that in the previous call to dp_netdev_input__() that same flow 'k' had
5172 * already its own batches[k] still waiting to be served. So if its
5173 * ‘batch’ member is not reset, the recirculated packet would be wrongly
5174 * appended to batches[k] of the 1st call to dp_netdev_input__(). */
5176 for (i
= 0; i
< n_batches
; i
++) {
5177 batches
[i
].flow
->batch
= NULL
;
5180 for (i
= 0; i
< n_batches
; i
++) {
5181 packet_batch_per_flow_execute(&batches
[i
], pmd
, now
);
5186 dp_netdev_input(struct dp_netdev_pmd_thread
*pmd
,
5187 struct dp_packet_batch
*packets
,
5190 dp_netdev_input__(pmd
, packets
, false, port_no
);
5194 dp_netdev_recirculate(struct dp_netdev_pmd_thread
*pmd
,
5195 struct dp_packet_batch
*packets
)
5197 dp_netdev_input__(pmd
, packets
, true, 0);
5200 struct dp_netdev_execute_aux
{
5201 struct dp_netdev_pmd_thread
*pmd
;
5203 const struct flow
*flow
;
5207 dpif_netdev_register_dp_purge_cb(struct dpif
*dpif
, dp_purge_callback
*cb
,
5210 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5211 dp
->dp_purge_aux
= aux
;
5212 dp
->dp_purge_cb
= cb
;
5216 dpif_netdev_register_upcall_cb(struct dpif
*dpif
, upcall_callback
*cb
,
5219 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5220 dp
->upcall_aux
= aux
;
5225 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
5226 long long now
, bool purge
)
5229 struct dp_netdev_port
*port
;
5232 HMAP_FOR_EACH (tx
, node
, &pmd
->send_port_cache
) {
5233 if (!tx
->port
->dynamic_txqs
) {
5236 interval
= now
- tx
->last_used
;
5237 if (tx
->qid
>= 0 && (purge
|| interval
>= XPS_TIMEOUT_MS
)) {
5239 ovs_mutex_lock(&port
->txq_used_mutex
);
5240 port
->txq_used
[tx
->qid
]--;
5241 ovs_mutex_unlock(&port
->txq_used_mutex
);
5248 dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
5249 struct tx_port
*tx
, long long now
)
5251 struct dp_netdev_port
*port
;
5253 int i
, min_cnt
, min_qid
;
5255 if (OVS_UNLIKELY(!now
)) {
5259 interval
= now
- tx
->last_used
;
5260 tx
->last_used
= now
;
5262 if (OVS_LIKELY(tx
->qid
>= 0 && interval
< XPS_TIMEOUT_MS
)) {
5268 ovs_mutex_lock(&port
->txq_used_mutex
);
5270 port
->txq_used
[tx
->qid
]--;
5276 for (i
= 0; i
< netdev_n_txq(port
->netdev
); i
++) {
5277 if (port
->txq_used
[i
] < min_cnt
|| min_cnt
== -1) {
5278 min_cnt
= port
->txq_used
[i
];
5283 port
->txq_used
[min_qid
]++;
5286 ovs_mutex_unlock(&port
->txq_used_mutex
);
5288 dpif_netdev_xps_revalidate_pmd(pmd
, now
, false);
5290 VLOG_DBG("Core %d: New TX queue ID %d for port \'%s\'.",
5291 pmd
->core_id
, tx
->qid
, netdev_get_name(tx
->port
->netdev
));
5295 static struct tx_port
*
5296 pmd_tnl_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
5299 return tx_port_lookup(&pmd
->tnl_port_cache
, port_no
);
5302 static struct tx_port
*
5303 pmd_send_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
5306 return tx_port_lookup(&pmd
->send_port_cache
, port_no
);
5310 push_tnl_action(const struct dp_netdev_pmd_thread
*pmd
,
5311 const struct nlattr
*attr
,
5312 struct dp_packet_batch
*batch
)
5314 struct tx_port
*tun_port
;
5315 const struct ovs_action_push_tnl
*data
;
5318 data
= nl_attr_get(attr
);
5320 tun_port
= pmd_tnl_port_cache_lookup(pmd
, data
->tnl_port
);
5325 err
= netdev_push_header(tun_port
->port
->netdev
, batch
, data
);
5330 dp_packet_delete_batch(batch
, true);
5335 dp_execute_userspace_action(struct dp_netdev_pmd_thread
*pmd
,
5336 struct dp_packet
*packet
, bool may_steal
,
5337 struct flow
*flow
, ovs_u128
*ufid
,
5338 struct ofpbuf
*actions
,
5339 const struct nlattr
*userdata
, long long now
)
5341 struct dp_packet_batch b
;
5344 ofpbuf_clear(actions
);
5346 error
= dp_netdev_upcall(pmd
, packet
, flow
, NULL
, ufid
,
5347 DPIF_UC_ACTION
, userdata
, actions
,
5349 if (!error
|| error
== ENOSPC
) {
5350 dp_packet_batch_init_packet(&b
, packet
);
5351 dp_netdev_execute_actions(pmd
, &b
, may_steal
, flow
,
5352 actions
->data
, actions
->size
, now
);
5353 } else if (may_steal
) {
5354 dp_packet_delete(packet
);
5359 dp_execute_cb(void *aux_
, struct dp_packet_batch
*packets_
,
5360 const struct nlattr
*a
, bool may_steal
)
5361 OVS_NO_THREAD_SAFETY_ANALYSIS
5363 struct dp_netdev_execute_aux
*aux
= aux_
;
5364 uint32_t *depth
= recirc_depth_get();
5365 struct dp_netdev_pmd_thread
*pmd
= aux
->pmd
;
5366 struct dp_netdev
*dp
= pmd
->dp
;
5367 int type
= nl_attr_type(a
);
5368 long long now
= aux
->now
;
5371 switch ((enum ovs_action_attr
)type
) {
5372 case OVS_ACTION_ATTR_OUTPUT
:
5373 p
= pmd_send_port_cache_lookup(pmd
, nl_attr_get_odp_port(a
));
5374 if (OVS_LIKELY(p
)) {
5378 dynamic_txqs
= p
->port
->dynamic_txqs
;
5380 tx_qid
= dpif_netdev_xps_get_tx_qid(pmd
, p
, now
);
5382 tx_qid
= pmd
->static_tx_qid
;
5385 netdev_send(p
->port
->netdev
, tx_qid
, packets_
, may_steal
,
5391 case OVS_ACTION_ATTR_TUNNEL_PUSH
:
5392 if (*depth
< MAX_RECIRC_DEPTH
) {
5393 dp_packet_batch_apply_cutlen(packets_
);
5394 push_tnl_action(pmd
, a
, packets_
);
5399 case OVS_ACTION_ATTR_TUNNEL_POP
:
5400 if (*depth
< MAX_RECIRC_DEPTH
) {
5401 struct dp_packet_batch
*orig_packets_
= packets_
;
5402 odp_port_t portno
= nl_attr_get_odp_port(a
);
5404 p
= pmd_tnl_port_cache_lookup(pmd
, portno
);
5406 struct dp_packet_batch tnl_pkt
;
5409 dp_packet_batch_clone(&tnl_pkt
, packets_
);
5410 packets_
= &tnl_pkt
;
5411 dp_packet_batch_reset_cutlen(orig_packets_
);
5414 dp_packet_batch_apply_cutlen(packets_
);
5416 netdev_pop_header(p
->port
->netdev
, packets_
);
5417 if (dp_packet_batch_is_empty(packets_
)) {
5421 struct dp_packet
*packet
;
5422 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5423 packet
->md
.in_port
.odp_port
= portno
;
5427 dp_netdev_recirculate(pmd
, packets_
);
5434 case OVS_ACTION_ATTR_USERSPACE
:
5435 if (!fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
5436 struct dp_packet_batch
*orig_packets_
= packets_
;
5437 const struct nlattr
*userdata
;
5438 struct dp_packet_batch usr_pkt
;
5439 struct ofpbuf actions
;
5444 userdata
= nl_attr_find_nested(a
, OVS_USERSPACE_ATTR_USERDATA
);
5445 ofpbuf_init(&actions
, 0);
5447 if (packets_
->trunc
) {
5449 dp_packet_batch_clone(&usr_pkt
, packets_
);
5450 packets_
= &usr_pkt
;
5452 dp_packet_batch_reset_cutlen(orig_packets_
);
5455 dp_packet_batch_apply_cutlen(packets_
);
5458 struct dp_packet
*packet
;
5459 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5460 flow_extract(packet
, &flow
);
5461 dpif_flow_hash(dp
->dpif
, &flow
, sizeof flow
, &ufid
);
5462 dp_execute_userspace_action(pmd
, packet
, may_steal
, &flow
,
5463 &ufid
, &actions
, userdata
, now
);
5467 dp_packet_delete_batch(packets_
, true);
5470 ofpbuf_uninit(&actions
);
5471 fat_rwlock_unlock(&dp
->upcall_rwlock
);
5477 case OVS_ACTION_ATTR_RECIRC
:
5478 if (*depth
< MAX_RECIRC_DEPTH
) {
5479 struct dp_packet_batch recirc_pkts
;
5482 dp_packet_batch_clone(&recirc_pkts
, packets_
);
5483 packets_
= &recirc_pkts
;
5486 struct dp_packet
*packet
;
5487 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5488 packet
->md
.recirc_id
= nl_attr_get_u32(a
);
5492 dp_netdev_recirculate(pmd
, packets_
);
5498 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
5501 case OVS_ACTION_ATTR_CT
: {
5502 const struct nlattr
*b
;
5504 bool commit
= false;
5507 const char *helper
= NULL
;
5508 const uint32_t *setmark
= NULL
;
5509 const struct ovs_key_ct_labels
*setlabel
= NULL
;
5510 struct nat_action_info_t nat_action_info
;
5511 struct nat_action_info_t
*nat_action_info_ref
= NULL
;
5512 bool nat_config
= false;
5514 NL_ATTR_FOR_EACH_UNSAFE (b
, left
, nl_attr_get(a
),
5515 nl_attr_get_size(a
)) {
5516 enum ovs_ct_attr sub_type
= nl_attr_type(b
);
5519 case OVS_CT_ATTR_FORCE_COMMIT
:
5522 case OVS_CT_ATTR_COMMIT
:
5525 case OVS_CT_ATTR_ZONE
:
5526 zone
= nl_attr_get_u16(b
);
5528 case OVS_CT_ATTR_HELPER
:
5529 helper
= nl_attr_get_string(b
);
5531 case OVS_CT_ATTR_MARK
:
5532 setmark
= nl_attr_get(b
);
5534 case OVS_CT_ATTR_LABELS
:
5535 setlabel
= nl_attr_get(b
);
5537 case OVS_CT_ATTR_EVENTMASK
:
5538 /* Silently ignored, as userspace datapath does not generate
5539 * netlink events. */
5541 case OVS_CT_ATTR_NAT
: {
5542 const struct nlattr
*b_nest
;
5543 unsigned int left_nest
;
5544 bool ip_min_specified
= false;
5545 bool proto_num_min_specified
= false;
5546 bool ip_max_specified
= false;
5547 bool proto_num_max_specified
= false;
5548 memset(&nat_action_info
, 0, sizeof nat_action_info
);
5549 nat_action_info_ref
= &nat_action_info
;
5551 NL_NESTED_FOR_EACH_UNSAFE (b_nest
, left_nest
, b
) {
5552 enum ovs_nat_attr sub_type_nest
= nl_attr_type(b_nest
);
5554 switch (sub_type_nest
) {
5555 case OVS_NAT_ATTR_SRC
:
5556 case OVS_NAT_ATTR_DST
:
5558 nat_action_info
.nat_action
|=
5559 ((sub_type_nest
== OVS_NAT_ATTR_SRC
)
5560 ? NAT_ACTION_SRC
: NAT_ACTION_DST
);
5562 case OVS_NAT_ATTR_IP_MIN
:
5563 memcpy(&nat_action_info
.min_addr
,
5564 nl_attr_get(b_nest
),
5565 nl_attr_get_size(b_nest
));
5566 ip_min_specified
= true;
5568 case OVS_NAT_ATTR_IP_MAX
:
5569 memcpy(&nat_action_info
.max_addr
,
5570 nl_attr_get(b_nest
),
5571 nl_attr_get_size(b_nest
));
5572 ip_max_specified
= true;
5574 case OVS_NAT_ATTR_PROTO_MIN
:
5575 nat_action_info
.min_port
=
5576 nl_attr_get_u16(b_nest
);
5577 proto_num_min_specified
= true;
5579 case OVS_NAT_ATTR_PROTO_MAX
:
5580 nat_action_info
.max_port
=
5581 nl_attr_get_u16(b_nest
);
5582 proto_num_max_specified
= true;
5584 case OVS_NAT_ATTR_PERSISTENT
:
5585 case OVS_NAT_ATTR_PROTO_HASH
:
5586 case OVS_NAT_ATTR_PROTO_RANDOM
:
5588 case OVS_NAT_ATTR_UNSPEC
:
5589 case __OVS_NAT_ATTR_MAX
:
5594 if (ip_min_specified
&& !ip_max_specified
) {
5595 nat_action_info
.max_addr
= nat_action_info
.min_addr
;
5597 if (proto_num_min_specified
&& !proto_num_max_specified
) {
5598 nat_action_info
.max_port
= nat_action_info
.min_port
;
5600 if (proto_num_min_specified
|| proto_num_max_specified
) {
5601 if (nat_action_info
.nat_action
& NAT_ACTION_SRC
) {
5602 nat_action_info
.nat_action
|= NAT_ACTION_SRC_PORT
;
5603 } else if (nat_action_info
.nat_action
& NAT_ACTION_DST
) {
5604 nat_action_info
.nat_action
|= NAT_ACTION_DST_PORT
;
5609 case OVS_CT_ATTR_UNSPEC
:
5610 case __OVS_CT_ATTR_MAX
:
5615 /* We won't be able to function properly in this case, hence
5616 * complain loudly. */
5617 if (nat_config
&& !commit
) {
5618 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(5, 5);
5619 VLOG_WARN_RL(&rl
, "NAT specified without commit.");
5622 conntrack_execute(&dp
->conntrack
, packets_
, aux
->flow
->dl_type
, force
,
5623 commit
, zone
, setmark
, setlabel
, helper
,
5624 nat_action_info_ref
, now
);
5628 case OVS_ACTION_ATTR_METER
:
5629 dp_netdev_run_meter(pmd
->dp
, packets_
, nl_attr_get_u32(a
),
5633 case OVS_ACTION_ATTR_PUSH_VLAN
:
5634 case OVS_ACTION_ATTR_POP_VLAN
:
5635 case OVS_ACTION_ATTR_PUSH_MPLS
:
5636 case OVS_ACTION_ATTR_POP_MPLS
:
5637 case OVS_ACTION_ATTR_SET
:
5638 case OVS_ACTION_ATTR_SET_MASKED
:
5639 case OVS_ACTION_ATTR_SAMPLE
:
5640 case OVS_ACTION_ATTR_HASH
:
5641 case OVS_ACTION_ATTR_UNSPEC
:
5642 case OVS_ACTION_ATTR_TRUNC
:
5643 case OVS_ACTION_ATTR_PUSH_ETH
:
5644 case OVS_ACTION_ATTR_POP_ETH
:
5645 case OVS_ACTION_ATTR_CLONE
:
5646 case OVS_ACTION_ATTR_ENCAP_NSH
:
5647 case OVS_ACTION_ATTR_DECAP_NSH
:
5648 case __OVS_ACTION_ATTR_MAX
:
5652 dp_packet_delete_batch(packets_
, may_steal
);
5656 dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
5657 struct dp_packet_batch
*packets
,
5658 bool may_steal
, const struct flow
*flow
,
5659 const struct nlattr
*actions
, size_t actions_len
,
5662 struct dp_netdev_execute_aux aux
= { pmd
, now
, flow
};
5664 odp_execute_actions(&aux
, packets
, may_steal
, actions
,
5665 actions_len
, dp_execute_cb
);
5668 struct dp_netdev_ct_dump
{
5669 struct ct_dpif_dump_state up
;
5670 struct conntrack_dump dump
;
5671 struct conntrack
*ct
;
5672 struct dp_netdev
*dp
;
5676 dpif_netdev_ct_dump_start(struct dpif
*dpif
, struct ct_dpif_dump_state
**dump_
,
5677 const uint16_t *pzone
, int *ptot_bkts
)
5679 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5680 struct dp_netdev_ct_dump
*dump
;
5682 dump
= xzalloc(sizeof *dump
);
5684 dump
->ct
= &dp
->conntrack
;
5686 conntrack_dump_start(&dp
->conntrack
, &dump
->dump
, pzone
, ptot_bkts
);
5694 dpif_netdev_ct_dump_next(struct dpif
*dpif OVS_UNUSED
,
5695 struct ct_dpif_dump_state
*dump_
,
5696 struct ct_dpif_entry
*entry
)
5698 struct dp_netdev_ct_dump
*dump
;
5700 INIT_CONTAINER(dump
, dump_
, up
);
5702 return conntrack_dump_next(&dump
->dump
, entry
);
5706 dpif_netdev_ct_dump_done(struct dpif
*dpif OVS_UNUSED
,
5707 struct ct_dpif_dump_state
*dump_
)
5709 struct dp_netdev_ct_dump
*dump
;
5712 INIT_CONTAINER(dump
, dump_
, up
);
5714 err
= conntrack_dump_done(&dump
->dump
);
5722 dpif_netdev_ct_flush(struct dpif
*dpif
, const uint16_t *zone
)
5724 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5726 return conntrack_flush(&dp
->conntrack
, zone
);
5729 const struct dpif_class dpif_netdev_class
= {
5732 dpif_netdev_enumerate
,
5733 dpif_netdev_port_open_type
,
5736 dpif_netdev_destroy
,
5739 dpif_netdev_get_stats
,
5740 dpif_netdev_port_add
,
5741 dpif_netdev_port_del
,
5742 dpif_netdev_port_set_config
,
5743 dpif_netdev_port_query_by_number
,
5744 dpif_netdev_port_query_by_name
,
5745 NULL
, /* port_get_pid */
5746 dpif_netdev_port_dump_start
,
5747 dpif_netdev_port_dump_next
,
5748 dpif_netdev_port_dump_done
,
5749 dpif_netdev_port_poll
,
5750 dpif_netdev_port_poll_wait
,
5751 dpif_netdev_flow_flush
,
5752 dpif_netdev_flow_dump_create
,
5753 dpif_netdev_flow_dump_destroy
,
5754 dpif_netdev_flow_dump_thread_create
,
5755 dpif_netdev_flow_dump_thread_destroy
,
5756 dpif_netdev_flow_dump_next
,
5757 dpif_netdev_operate
,
5758 NULL
, /* recv_set */
5759 NULL
, /* handlers_set */
5760 dpif_netdev_set_config
,
5761 dpif_netdev_queue_to_priority
,
5763 NULL
, /* recv_wait */
5764 NULL
, /* recv_purge */
5765 dpif_netdev_register_dp_purge_cb
,
5766 dpif_netdev_register_upcall_cb
,
5767 dpif_netdev_enable_upcall
,
5768 dpif_netdev_disable_upcall
,
5769 dpif_netdev_get_datapath_version
,
5770 dpif_netdev_ct_dump_start
,
5771 dpif_netdev_ct_dump_next
,
5772 dpif_netdev_ct_dump_done
,
5773 dpif_netdev_ct_flush
,
5774 dpif_netdev_meter_get_features
,
5775 dpif_netdev_meter_set
,
5776 dpif_netdev_meter_get
,
5777 dpif_netdev_meter_del
,
5781 dpif_dummy_change_port_number(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
5782 const char *argv
[], void *aux OVS_UNUSED
)
5784 struct dp_netdev_port
*port
;
5785 struct dp_netdev
*dp
;
5788 ovs_mutex_lock(&dp_netdev_mutex
);
5789 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
5790 if (!dp
|| !dpif_netdev_class_is_dummy(dp
->class)) {
5791 ovs_mutex_unlock(&dp_netdev_mutex
);
5792 unixctl_command_reply_error(conn
, "unknown datapath or not a dummy");
5795 ovs_refcount_ref(&dp
->ref_cnt
);
5796 ovs_mutex_unlock(&dp_netdev_mutex
);
5798 ovs_mutex_lock(&dp
->port_mutex
);
5799 if (get_port_by_name(dp
, argv
[2], &port
)) {
5800 unixctl_command_reply_error(conn
, "unknown port");
5804 port_no
= u32_to_odp(atoi(argv
[3]));
5805 if (!port_no
|| port_no
== ODPP_NONE
) {
5806 unixctl_command_reply_error(conn
, "bad port number");
5809 if (dp_netdev_lookup_port(dp
, port_no
)) {
5810 unixctl_command_reply_error(conn
, "port number already in use");
5815 hmap_remove(&dp
->ports
, &port
->node
);
5816 reconfigure_datapath(dp
);
5818 /* Reinsert with new port number. */
5819 port
->port_no
= port_no
;
5820 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
5821 reconfigure_datapath(dp
);
5823 seq_change(dp
->port_seq
);
5824 unixctl_command_reply(conn
, NULL
);
5827 ovs_mutex_unlock(&dp
->port_mutex
);
5828 dp_netdev_unref(dp
);
5832 dpif_dummy_register__(const char *type
)
5834 struct dpif_class
*class;
5836 class = xmalloc(sizeof *class);
5837 *class = dpif_netdev_class
;
5838 class->type
= xstrdup(type
);
5839 dp_register_provider(class);
5843 dpif_dummy_override(const char *type
)
5848 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
5849 * a userland-only build. It's useful for testsuite.
5851 error
= dp_unregister_provider(type
);
5852 if (error
== 0 || error
== EAFNOSUPPORT
) {
5853 dpif_dummy_register__(type
);
5858 dpif_dummy_register(enum dummy_level level
)
5860 if (level
== DUMMY_OVERRIDE_ALL
) {
5865 dp_enumerate_types(&types
);
5866 SSET_FOR_EACH (type
, &types
) {
5867 dpif_dummy_override(type
);
5869 sset_destroy(&types
);
5870 } else if (level
== DUMMY_OVERRIDE_SYSTEM
) {
5871 dpif_dummy_override("system");
5874 dpif_dummy_register__("dummy");
5876 unixctl_command_register("dpif-dummy/change-port-number",
5877 "dp port new-number",
5878 3, 3, dpif_dummy_change_port_number
, NULL
);
5881 /* Datapath Classifier. */
5883 /* A set of rules that all have the same fields wildcarded. */
5884 struct dpcls_subtable
{
5885 /* The fields are only used by writers. */
5886 struct cmap_node cmap_node OVS_GUARDED
; /* Within dpcls 'subtables_map'. */
5888 /* These fields are accessed by readers. */
5889 struct cmap rules
; /* Contains "struct dpcls_rule"s. */
5890 uint32_t hit_cnt
; /* Number of match hits in subtable in current
5891 optimization interval. */
5892 struct netdev_flow_key mask
; /* Wildcards for fields (const). */
5893 /* 'mask' must be the last field, additional space is allocated here. */
5896 /* Initializes 'cls' as a classifier that initially contains no classification
5899 dpcls_init(struct dpcls
*cls
)
5901 cmap_init(&cls
->subtables_map
);
5902 pvector_init(&cls
->subtables
);
5906 dpcls_destroy_subtable(struct dpcls
*cls
, struct dpcls_subtable
*subtable
)
5908 VLOG_DBG("Destroying subtable %p for in_port %d", subtable
, cls
->in_port
);
5909 pvector_remove(&cls
->subtables
, subtable
);
5910 cmap_remove(&cls
->subtables_map
, &subtable
->cmap_node
,
5911 subtable
->mask
.hash
);
5912 cmap_destroy(&subtable
->rules
);
5913 ovsrcu_postpone(free
, subtable
);
5916 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
5917 * caller's responsibility.
5918 * May only be called after all the readers have been terminated. */
5920 dpcls_destroy(struct dpcls
*cls
)
5923 struct dpcls_subtable
*subtable
;
5925 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
5926 ovs_assert(cmap_count(&subtable
->rules
) == 0);
5927 dpcls_destroy_subtable(cls
, subtable
);
5929 cmap_destroy(&cls
->subtables_map
);
5930 pvector_destroy(&cls
->subtables
);
5934 static struct dpcls_subtable
*
5935 dpcls_create_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
5937 struct dpcls_subtable
*subtable
;
5939 /* Need to add one. */
5940 subtable
= xmalloc(sizeof *subtable
5941 - sizeof subtable
->mask
.mf
+ mask
->len
);
5942 cmap_init(&subtable
->rules
);
5943 subtable
->hit_cnt
= 0;
5944 netdev_flow_key_clone(&subtable
->mask
, mask
);
5945 cmap_insert(&cls
->subtables_map
, &subtable
->cmap_node
, mask
->hash
);
5946 /* Add the new subtable at the end of the pvector (with no hits yet) */
5947 pvector_insert(&cls
->subtables
, subtable
, 0);
5948 VLOG_DBG("Creating %"PRIuSIZE
". subtable %p for in_port %d",
5949 cmap_count(&cls
->subtables_map
), subtable
, cls
->in_port
);
5950 pvector_publish(&cls
->subtables
);
5955 static inline struct dpcls_subtable
*
5956 dpcls_find_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
5958 struct dpcls_subtable
*subtable
;
5960 CMAP_FOR_EACH_WITH_HASH (subtable
, cmap_node
, mask
->hash
,
5961 &cls
->subtables_map
) {
5962 if (netdev_flow_key_equal(&subtable
->mask
, mask
)) {
5966 return dpcls_create_subtable(cls
, mask
);
5970 /* Periodically sort the dpcls subtable vectors according to hit counts */
5972 dpcls_sort_subtable_vector(struct dpcls
*cls
)
5974 struct pvector
*pvec
= &cls
->subtables
;
5975 struct dpcls_subtable
*subtable
;
5977 PVECTOR_FOR_EACH (subtable
, pvec
) {
5978 pvector_change_priority(pvec
, subtable
, subtable
->hit_cnt
);
5979 subtable
->hit_cnt
= 0;
5981 pvector_publish(pvec
);
5985 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
,
5986 struct polled_queue
*poll_list
, int poll_cnt
)
5989 long long int now
= time_msec();
5991 if (now
> pmd
->rxq_interval
) {
5992 /* Get the cycles that were used to process each queue and store. */
5993 for (unsigned i
= 0; i
< poll_cnt
; i
++) {
5994 uint64_t rxq_cyc_curr
= dp_netdev_rxq_get_cycles(poll_list
[i
].rxq
,
5995 RXQ_CYCLES_PROC_CURR
);
5996 dp_netdev_rxq_set_intrvl_cycles(poll_list
[i
].rxq
, rxq_cyc_curr
);
5997 dp_netdev_rxq_set_cycles(poll_list
[i
].rxq
, RXQ_CYCLES_PROC_CURR
,
6000 /* Start new measuring interval */
6001 pmd
->rxq_interval
= now
+ PMD_RXQ_INTERVAL_LEN
;
6004 if (now
> pmd
->next_optimization
) {
6005 /* Try to obtain the flow lock to block out revalidator threads.
6006 * If not possible, just try next time. */
6007 if (!ovs_mutex_trylock(&pmd
->flow_mutex
)) {
6008 /* Optimize each classifier */
6009 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
6010 dpcls_sort_subtable_vector(cls
);
6012 ovs_mutex_unlock(&pmd
->flow_mutex
);
6013 /* Start new measuring interval */
6014 pmd
->next_optimization
= now
+ DPCLS_OPTIMIZATION_INTERVAL
;
6019 /* Insert 'rule' into 'cls'. */
6021 dpcls_insert(struct dpcls
*cls
, struct dpcls_rule
*rule
,
6022 const struct netdev_flow_key
*mask
)
6024 struct dpcls_subtable
*subtable
= dpcls_find_subtable(cls
, mask
);
6026 /* Refer to subtable's mask, also for later removal. */
6027 rule
->mask
= &subtable
->mask
;
6028 cmap_insert(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
);
6031 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
6033 dpcls_remove(struct dpcls
*cls
, struct dpcls_rule
*rule
)
6035 struct dpcls_subtable
*subtable
;
6037 ovs_assert(rule
->mask
);
6039 /* Get subtable from reference in rule->mask. */
6040 INIT_CONTAINER(subtable
, rule
->mask
, mask
);
6041 if (cmap_remove(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
)
6043 /* Delete empty subtable. */
6044 dpcls_destroy_subtable(cls
, subtable
);
6045 pvector_publish(&cls
->subtables
);
6049 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
6050 * in 'mask' the values in 'key' and 'target' are the same. */
6052 dpcls_rule_matches_key(const struct dpcls_rule
*rule
,
6053 const struct netdev_flow_key
*target
)
6055 const uint64_t *keyp
= miniflow_get_values(&rule
->flow
.mf
);
6056 const uint64_t *maskp
= miniflow_get_values(&rule
->mask
->mf
);
6059 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, target
, rule
->flow
.mf
.map
) {
6060 if (OVS_UNLIKELY((value
& *maskp
++) != *keyp
++)) {
6067 /* For each miniflow in 'keys' performs a classifier lookup writing the result
6068 * into the corresponding slot in 'rules'. If a particular entry in 'keys' is
6069 * NULL it is skipped.
6071 * This function is optimized for use in the userspace datapath and therefore
6072 * does not implement a lot of features available in the standard
6073 * classifier_lookup() function. Specifically, it does not implement
6074 * priorities, instead returning any rule which matches the flow.
6076 * Returns true if all miniflows found a corresponding rule. */
6078 dpcls_lookup(struct dpcls
*cls
, const struct netdev_flow_key keys
[],
6079 struct dpcls_rule
**rules
, const size_t cnt
,
6082 /* The received 'cnt' miniflows are the search-keys that will be processed
6083 * to find a matching entry into the available subtables.
6084 * The number of bits in map_type is equal to NETDEV_MAX_BURST. */
6085 typedef uint32_t map_type
;
6086 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
6087 BUILD_ASSERT_DECL(MAP_BITS
>= NETDEV_MAX_BURST
);
6089 struct dpcls_subtable
*subtable
;
6091 map_type keys_map
= TYPE_MAXIMUM(map_type
); /* Set all bits. */
6093 uint32_t hashes
[MAP_BITS
];
6094 const struct cmap_node
*nodes
[MAP_BITS
];
6096 if (cnt
!= MAP_BITS
) {
6097 keys_map
>>= MAP_BITS
- cnt
; /* Clear extra bits. */
6099 memset(rules
, 0, cnt
* sizeof *rules
);
6101 int lookups_match
= 0, subtable_pos
= 1;
6103 /* The Datapath classifier - aka dpcls - is composed of subtables.
6104 * Subtables are dynamically created as needed when new rules are inserted.
6105 * Each subtable collects rules with matches on a specific subset of packet
6106 * fields as defined by the subtable's mask. We proceed to process every
6107 * search-key against each subtable, but when a match is found for a
6108 * search-key, the search for that key can stop because the rules are
6109 * non-overlapping. */
6110 PVECTOR_FOR_EACH (subtable
, &cls
->subtables
) {
6113 /* Compute hashes for the remaining keys. Each search-key is
6114 * masked with the subtable's mask to avoid hashing the wildcarded
6116 ULLONG_FOR_EACH_1(i
, keys_map
) {
6117 hashes
[i
] = netdev_flow_key_hash_in_mask(&keys
[i
],
6121 found_map
= cmap_find_batch(&subtable
->rules
, keys_map
, hashes
, nodes
);
6122 /* Check results. When the i-th bit of found_map is set, it means
6123 * that a set of nodes with a matching hash value was found for the
6124 * i-th search-key. Due to possible hash collisions we need to check
6125 * which of the found rules, if any, really matches our masked
6127 ULLONG_FOR_EACH_1(i
, found_map
) {
6128 struct dpcls_rule
*rule
;
6130 CMAP_NODE_FOR_EACH (rule
, cmap_node
, nodes
[i
]) {
6131 if (OVS_LIKELY(dpcls_rule_matches_key(rule
, &keys
[i
]))) {
6133 /* Even at 20 Mpps the 32-bit hit_cnt cannot wrap
6134 * within one second optimization interval. */
6135 subtable
->hit_cnt
++;
6136 lookups_match
+= subtable_pos
;
6140 /* None of the found rules was a match. Reset the i-th bit to
6141 * keep searching this key in the next subtable. */
6142 ULLONG_SET0(found_map
, i
); /* Did not match. */
6144 ; /* Keep Sparse happy. */
6146 keys_map
&= ~found_map
; /* Clear the found rules. */
6148 if (num_lookups_p
) {
6149 *num_lookups_p
= lookups_match
;
6151 return true; /* All found. */
6155 if (num_lookups_p
) {
6156 *num_lookups_p
= lookups_match
;
6158 return false; /* Some misses. */