2 * Copyright (c) 2009-2014, 2016-2018 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"
19 #include "dpif-netdev-private.h"
26 #include <sys/types.h>
27 #include <netinet/in.h>
31 #include <sys/ioctl.h>
32 #include <sys/socket.h>
38 #include "conntrack.h"
42 #include "dp-packet.h"
44 #include "dpif-netdev-perf.h"
45 #include "dpif-provider.h"
47 #include "fat-rwlock.h"
53 #include "netdev-offload.h"
54 #include "netdev-provider.h"
55 #include "netdev-vport.h"
57 #include "odp-execute.h"
59 #include "openvswitch/dynamic-string.h"
60 #include "openvswitch/list.h"
61 #include "openvswitch/match.h"
62 #include "openvswitch/ofp-parse.h"
63 #include "openvswitch/ofp-print.h"
64 #include "openvswitch/ofpbuf.h"
65 #include "openvswitch/shash.h"
66 #include "openvswitch/vlog.h"
70 #include "openvswitch/poll-loop.h"
77 #include "tnl-neigh-cache.h"
78 #include "tnl-ports.h"
83 VLOG_DEFINE_THIS_MODULE(dpif_netdev
);
85 /* Auto Load Balancing Defaults */
86 #define ALB_ACCEPTABLE_IMPROVEMENT 25
87 #define ALB_PMD_LOAD_THRESHOLD 95
88 #define ALB_PMD_REBALANCE_POLL_INTERVAL 1 /* 1 Min */
89 #define MIN_TO_MSEC 60000
91 #define FLOW_DUMP_MAX_BATCH 50
92 /* Use per thread recirc_depth to prevent recirculation loop. */
93 #define MAX_RECIRC_DEPTH 6
94 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth
, 0)
96 /* Use instant packet send by default. */
97 #define DEFAULT_TX_FLUSH_INTERVAL 0
99 /* Configuration parameters. */
100 enum { MAX_FLOWS
= 65536 }; /* Maximum number of flows in flow table. */
101 enum { MAX_METERS
= 65536 }; /* Maximum number of meters. */
102 enum { MAX_BANDS
= 8 }; /* Maximum number of bands / meter. */
103 enum { N_METER_LOCKS
= 64 }; /* Maximum number of meters. */
105 /* Protects against changes to 'dp_netdevs'. */
106 static struct ovs_mutex dp_netdev_mutex
= OVS_MUTEX_INITIALIZER
;
108 /* Contains all 'struct dp_netdev's. */
109 static struct shash dp_netdevs
OVS_GUARDED_BY(dp_netdev_mutex
)
110 = SHASH_INITIALIZER(&dp_netdevs
);
112 static struct vlog_rate_limit upcall_rl
= VLOG_RATE_LIMIT_INIT(600, 600);
114 #define DP_NETDEV_CS_SUPPORTED_MASK (CS_NEW | CS_ESTABLISHED | CS_RELATED \
115 | CS_INVALID | CS_REPLY_DIR | CS_TRACKED \
116 | CS_SRC_NAT | CS_DST_NAT)
117 #define DP_NETDEV_CS_UNSUPPORTED_MASK (~(uint32_t)DP_NETDEV_CS_SUPPORTED_MASK)
119 static struct odp_support dp_netdev_support
= {
120 .max_vlan_headers
= SIZE_MAX
,
121 .max_mpls_depth
= SIZE_MAX
,
127 .ct_state_nat
= true,
128 .ct_orig_tuple
= true,
129 .ct_orig_tuple6
= true,
132 /* EMC cache and SMC cache compose the datapath flow cache (DFC)
134 * Exact match cache for frequently used flows
136 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
137 * search its entries for a miniflow that matches exactly the miniflow of the
138 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
140 * A cache entry holds a reference to its 'dp_netdev_flow'.
142 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
143 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
144 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
145 * value is the index of a cache entry where the miniflow could be.
148 * Signature match cache (SMC)
150 * This cache stores a 16-bit signature for each flow without storing keys, and
151 * stores the corresponding 16-bit flow_table index to the 'dp_netdev_flow'.
152 * Each flow thus occupies 32bit which is much more memory efficient than EMC.
153 * SMC uses a set-associative design that each bucket contains
154 * SMC_ENTRY_PER_BUCKET number of entries.
155 * Since 16-bit flow_table index is used, if there are more than 2^16
156 * dp_netdev_flow, SMC will miss them that cannot be indexed by a 16-bit value.
162 * Each pmd_thread has its own private exact match cache.
163 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
166 #define EM_FLOW_HASH_SHIFT 13
167 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
168 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
169 #define EM_FLOW_HASH_SEGS 2
171 /* SMC uses a set-associative design. A bucket contains a set of entries that
172 * a flow item can occupy. For now, it uses one hash function rather than two
173 * as for the EMC design. */
174 #define SMC_ENTRY_PER_BUCKET 4
175 #define SMC_ENTRIES (1u << 20)
176 #define SMC_BUCKET_CNT (SMC_ENTRIES / SMC_ENTRY_PER_BUCKET)
177 #define SMC_MASK (SMC_BUCKET_CNT - 1)
179 /* Default EMC insert probability is 1 / DEFAULT_EM_FLOW_INSERT_INV_PROB */
180 #define DEFAULT_EM_FLOW_INSERT_INV_PROB 100
181 #define DEFAULT_EM_FLOW_INSERT_MIN (UINT32_MAX / \
182 DEFAULT_EM_FLOW_INSERT_INV_PROB)
185 struct dp_netdev_flow
*flow
;
186 struct netdev_flow_key key
; /* key.hash used for emc hash value. */
190 struct emc_entry entries
[EM_FLOW_HASH_ENTRIES
];
191 int sweep_idx
; /* For emc_cache_slow_sweep(). */
195 uint16_t sig
[SMC_ENTRY_PER_BUCKET
];
196 uint16_t flow_idx
[SMC_ENTRY_PER_BUCKET
];
199 /* Signature match cache, differentiate from EMC cache */
201 struct smc_bucket buckets
[SMC_BUCKET_CNT
];
205 struct emc_cache emc_cache
;
206 struct smc_cache smc_cache
;
209 /* Iterate in the exact match cache through every entry that might contain a
210 * miniflow with hash 'HASH'. */
211 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
212 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
213 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
214 i__ < EM_FLOW_HASH_SEGS; \
215 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
217 /* Simple non-wildcarding single-priority classifier. */
219 /* Time in microseconds between successive optimizations of the dpcls
221 #define DPCLS_OPTIMIZATION_INTERVAL 1000000LL
223 /* Time in microseconds of the interval in which rxq processing cycles used
224 * in rxq to pmd assignments is measured and stored. */
225 #define PMD_RXQ_INTERVAL_LEN 10000000LL
227 /* Number of intervals for which cycles are stored
228 * and used during rxq to pmd assignment. */
229 #define PMD_RXQ_INTERVAL_MAX 6
232 struct cmap_node node
; /* Within dp_netdev_pmd_thread.classifiers */
234 struct cmap subtables_map
;
235 struct pvector subtables
;
238 /* Data structure to keep packet order till fastpath processing. */
239 struct dp_packet_flow_map
{
240 struct dp_packet
*packet
;
241 struct dp_netdev_flow
*flow
;
245 static void dpcls_init(struct dpcls
*);
246 static void dpcls_destroy(struct dpcls
*);
247 static void dpcls_sort_subtable_vector(struct dpcls
*);
248 static void dpcls_insert(struct dpcls
*, struct dpcls_rule
*,
249 const struct netdev_flow_key
*mask
);
250 static void dpcls_remove(struct dpcls
*, struct dpcls_rule
*);
251 static bool dpcls_lookup(struct dpcls
*cls
,
252 const struct netdev_flow_key
*keys
[],
253 struct dpcls_rule
**rules
, size_t cnt
,
255 bool dpcls_rule_matches_key(const struct dpcls_rule
*rule
,
256 const struct netdev_flow_key
*target
);
257 /* Set of supported meter flags */
258 #define DP_SUPPORTED_METER_FLAGS_MASK \
259 (OFPMF13_STATS | OFPMF13_PKTPS | OFPMF13_KBPS | OFPMF13_BURST)
261 /* Set of supported meter band types */
262 #define DP_SUPPORTED_METER_BAND_TYPES \
263 ( 1 << OFPMBT13_DROP )
265 struct dp_meter_band
{
266 struct ofputil_meter_band up
; /* type, prec_level, pad, rate, burst_size */
267 uint32_t bucket
; /* In 1/1000 packets (for PKTPS), or in bits (for KBPS) */
268 uint64_t packet_count
;
275 uint32_t max_delta_t
;
277 uint64_t packet_count
;
279 struct dp_meter_band bands
[];
283 bool auto_lb_requested
; /* Auto load balancing requested by user. */
284 bool is_enabled
; /* Current status of Auto load balancing. */
285 uint64_t rebalance_intvl
;
286 uint64_t rebalance_poll_timer
;
289 /* Datapath based on the network device interface from netdev.h.
295 * Some members, marked 'const', are immutable. Accessing other members
296 * requires synchronization, as noted in more detail below.
298 * Acquisition order is, from outermost to innermost:
300 * dp_netdev_mutex (global)
305 const struct dpif_class
*const class;
306 const char *const name
;
308 struct ovs_refcount ref_cnt
;
309 atomic_flag destroyed
;
313 * Any lookup into 'ports' or any access to the dp_netdev_ports found
314 * through 'ports' requires taking 'port_mutex'. */
315 struct ovs_mutex port_mutex
;
317 struct seq
*port_seq
; /* Incremented whenever a port changes. */
319 /* The time that a packet can wait in output batch for sending. */
320 atomic_uint32_t tx_flush_interval
;
323 struct ovs_mutex meter_locks
[N_METER_LOCKS
];
324 struct dp_meter
*meters
[MAX_METERS
]; /* Meter bands. */
326 /* Probability of EMC insertions is a factor of 'emc_insert_min'.*/
327 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
) atomic_uint32_t emc_insert_min
;
328 /* Enable collection of PMD performance metrics. */
329 atomic_bool pmd_perf_metrics
;
330 /* Enable the SMC cache from ovsdb config */
331 atomic_bool smc_enable_db
;
333 /* Protects access to ofproto-dpif-upcall interface during revalidator
334 * thread synchronization. */
335 struct fat_rwlock upcall_rwlock
;
336 upcall_callback
*upcall_cb
; /* Callback function for executing upcalls. */
339 /* Callback function for notifying the purging of dp flows (during
340 * reseting pmd deletion). */
341 dp_purge_callback
*dp_purge_cb
;
344 /* Stores all 'struct dp_netdev_pmd_thread's. */
345 struct cmap poll_threads
;
346 /* id pool for per thread static_tx_qid. */
347 struct id_pool
*tx_qid_pool
;
348 struct ovs_mutex tx_qid_pool_mutex
;
349 /* Use measured cycles for rxq to pmd assignment. */
350 bool pmd_rxq_assign_cyc
;
352 /* Protects the access of the 'struct dp_netdev_pmd_thread'
353 * instance for non-pmd thread. */
354 struct ovs_mutex non_pmd_mutex
;
356 /* Each pmd thread will store its pointer to
357 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
358 ovsthread_key_t per_pmd_key
;
360 struct seq
*reconfigure_seq
;
361 uint64_t last_reconfigure_seq
;
363 /* Cpu mask for pin of pmd threads. */
366 uint64_t last_tnl_conf_seq
;
368 struct conntrack
*conntrack
;
369 struct pmd_auto_lb pmd_alb
;
372 static void meter_lock(const struct dp_netdev
*dp
, uint32_t meter_id
)
373 OVS_ACQUIRES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
375 ovs_mutex_lock(&dp
->meter_locks
[meter_id
% N_METER_LOCKS
]);
378 static void meter_unlock(const struct dp_netdev
*dp
, uint32_t meter_id
)
379 OVS_RELEASES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
381 ovs_mutex_unlock(&dp
->meter_locks
[meter_id
% N_METER_LOCKS
]);
385 static struct dp_netdev_port
*dp_netdev_lookup_port(const struct dp_netdev
*dp
,
387 OVS_REQUIRES(dp
->port_mutex
);
389 enum rxq_cycles_counter_type
{
390 RXQ_CYCLES_PROC_CURR
, /* Cycles spent successfully polling and
391 processing packets during the current
393 RXQ_CYCLES_PROC_HIST
, /* Total cycles of all intervals that are used
394 during rxq to pmd assignment. */
399 DP_NETDEV_FLOW_OFFLOAD_OP_ADD
,
400 DP_NETDEV_FLOW_OFFLOAD_OP_MOD
,
401 DP_NETDEV_FLOW_OFFLOAD_OP_DEL
,
404 struct dp_flow_offload_item
{
405 struct dp_netdev_pmd_thread
*pmd
;
406 struct dp_netdev_flow
*flow
;
409 struct nlattr
*actions
;
412 struct ovs_list node
;
415 struct dp_flow_offload
{
416 struct ovs_mutex mutex
;
417 struct ovs_list list
;
421 static struct dp_flow_offload dp_flow_offload
= {
422 .mutex
= OVS_MUTEX_INITIALIZER
,
423 .list
= OVS_LIST_INITIALIZER(&dp_flow_offload
.list
),
426 static struct ovsthread_once offload_thread_once
427 = OVSTHREAD_ONCE_INITIALIZER
;
429 #define XPS_TIMEOUT 500000LL /* In microseconds. */
431 /* Contained by struct dp_netdev_port's 'rxqs' member. */
432 struct dp_netdev_rxq
{
433 struct dp_netdev_port
*port
;
434 struct netdev_rxq
*rx
;
435 unsigned core_id
; /* Core to which this queue should be
436 pinned. OVS_CORE_UNSPEC if the
437 queue doesn't need to be pinned to a
439 unsigned intrvl_idx
; /* Write index for 'cycles_intrvl'. */
440 struct dp_netdev_pmd_thread
*pmd
; /* pmd thread that polls this queue. */
441 bool is_vhost
; /* Is rxq of a vhost port. */
443 /* Counters of cycles spent successfully polling and processing pkts. */
444 atomic_ullong cycles
[RXQ_N_CYCLES
];
445 /* We store PMD_RXQ_INTERVAL_MAX intervals of data for an rxq and then
446 sum them to yield the cycles used for an rxq. */
447 atomic_ullong cycles_intrvl
[PMD_RXQ_INTERVAL_MAX
];
450 /* A port in a netdev-based datapath. */
451 struct dp_netdev_port
{
453 bool dynamic_txqs
; /* If true XPS will be used. */
454 bool need_reconfigure
; /* True if we should reconfigure netdev. */
455 struct netdev
*netdev
;
456 struct hmap_node node
; /* Node in dp_netdev's 'ports'. */
457 struct netdev_saved_flags
*sf
;
458 struct dp_netdev_rxq
*rxqs
;
459 unsigned n_rxq
; /* Number of elements in 'rxqs' */
460 unsigned *txq_used
; /* Number of threads that use each tx queue. */
461 struct ovs_mutex txq_used_mutex
;
462 bool emc_enabled
; /* If true EMC will be used. */
463 char *type
; /* Port type as requested by user. */
464 char *rxq_affinity_list
; /* Requested affinity of rx queues. */
467 /* Contained by struct dp_netdev_flow's 'stats' member. */
468 struct dp_netdev_flow_stats
{
469 atomic_llong used
; /* Last used time, in monotonic msecs. */
470 atomic_ullong packet_count
; /* Number of packets matched. */
471 atomic_ullong byte_count
; /* Number of bytes matched. */
472 atomic_uint16_t tcp_flags
; /* Bitwise-OR of seen tcp_flags values. */
475 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
481 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
482 * its pmd thread's classifier. The text below calls this classifier 'cls'.
487 * The thread safety rules described here for "struct dp_netdev_flow" are
488 * motivated by two goals:
490 * - Prevent threads that read members of "struct dp_netdev_flow" from
491 * reading bad data due to changes by some thread concurrently modifying
494 * - Prevent two threads making changes to members of a given "struct
495 * dp_netdev_flow" from interfering with each other.
501 * A flow 'flow' may be accessed without a risk of being freed during an RCU
502 * grace period. Code that needs to hold onto a flow for a while
503 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
505 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
506 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
509 * Some members, marked 'const', are immutable. Accessing other members
510 * requires synchronization, as noted in more detail below.
512 struct dp_netdev_flow
{
513 const struct flow flow
; /* Unmasked flow that created this entry. */
514 /* Hash table index by unmasked flow. */
515 const struct cmap_node node
; /* In owning dp_netdev_pmd_thread's */
517 const struct cmap_node mark_node
; /* In owning flow_mark's mark_to_flow */
518 const ovs_u128 ufid
; /* Unique flow identifier. */
519 const ovs_u128 mega_ufid
; /* Unique mega flow identifier. */
520 const unsigned pmd_id
; /* The 'core_id' of pmd thread owning this */
523 /* Number of references.
524 * The classifier owns one reference.
525 * Any thread trying to keep a rule from being freed should hold its own
527 struct ovs_refcount ref_cnt
;
530 uint32_t mark
; /* Unique flow mark assigned to a flow */
533 struct dp_netdev_flow_stats stats
;
536 OVSRCU_TYPE(struct dp_netdev_actions
*) actions
;
538 /* While processing a group of input packets, the datapath uses the next
539 * member to store a pointer to the output batch for the flow. It is
540 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
541 * packet_batch_per_flow_init() and packet_batch_per_flow_execute()). */
542 struct packet_batch_per_flow
*batch
;
544 /* Packet classification. */
545 struct dpcls_rule cr
; /* In owning dp_netdev's 'cls'. */
546 /* 'cr' must be the last member. */
549 static void dp_netdev_flow_unref(struct dp_netdev_flow
*);
550 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*);
551 static int dpif_netdev_flow_from_nlattrs(const struct nlattr
*, uint32_t,
552 struct flow
*, bool);
554 /* A set of datapath actions within a "struct dp_netdev_flow".
560 * A struct dp_netdev_actions 'actions' is protected with RCU. */
561 struct dp_netdev_actions
{
562 /* These members are immutable: they do not change during the struct's
564 unsigned int size
; /* Size of 'actions', in bytes. */
565 struct nlattr actions
[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
568 struct dp_netdev_actions
*dp_netdev_actions_create(const struct nlattr
*,
570 struct dp_netdev_actions
*dp_netdev_flow_get_actions(
571 const struct dp_netdev_flow
*);
572 static void dp_netdev_actions_free(struct dp_netdev_actions
*);
574 struct polled_queue
{
575 struct dp_netdev_rxq
*rxq
;
582 /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */
584 struct dp_netdev_rxq
*rxq
;
585 struct hmap_node node
;
588 /* Contained by struct dp_netdev_pmd_thread's 'send_port_cache',
589 * 'tnl_port_cache' or 'tx_ports'. */
591 struct dp_netdev_port
*port
;
594 struct hmap_node node
;
595 long long flush_time
;
596 struct dp_packet_batch output_pkts
;
597 struct dp_netdev_rxq
*output_pkts_rxqs
[NETDEV_MAX_BURST
];
600 /* A set of properties for the current processing loop that is not directly
601 * associated with the pmd thread itself, but with the packets being
602 * processed or the short-term system configuration (for example, time).
603 * Contained by struct dp_netdev_pmd_thread's 'ctx' member. */
604 struct dp_netdev_pmd_thread_ctx
{
605 /* Latest measured time. See 'pmd_thread_ctx_time_update()'. */
607 /* RX queue from which last packet was received. */
608 struct dp_netdev_rxq
*last_rxq
;
609 /* EMC insertion probability context for the current processing cycle. */
610 uint32_t emc_insert_min
;
613 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
614 * the performance overhead of interrupt processing. Therefore netdev can
615 * not implement rx-wait for these devices. dpif-netdev needs to poll
616 * these device to check for recv buffer. pmd-thread does polling for
617 * devices assigned to itself.
619 * DPDK used PMD for accessing NIC.
621 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
622 * I/O of all non-pmd threads. There will be no actual thread created
625 * Each struct has its own flow cache and classifier per managed ingress port.
626 * For packets received on ingress port, a look up is done on corresponding PMD
627 * thread's flow cache and in case of a miss, lookup is performed in the
628 * corresponding classifier of port. Packets are executed with the found
629 * actions in either case.
631 struct dp_netdev_pmd_thread
{
632 struct dp_netdev
*dp
;
633 struct ovs_refcount ref_cnt
; /* Every reference must be refcount'ed. */
634 struct cmap_node node
; /* In 'dp->poll_threads'. */
636 /* Per thread exact-match cache. Note, the instance for cpu core
637 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
638 * need to be protected by 'non_pmd_mutex'. Every other instance
639 * will only be accessed by its own pmd thread. */
640 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
) struct dfc_cache flow_cache
;
642 /* Flow-Table and classifiers
644 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
645 * changes to 'classifiers' must be made while still holding the
648 struct ovs_mutex flow_mutex
;
649 struct cmap flow_table OVS_GUARDED
; /* Flow table. */
651 /* One classifier per in_port polled by the pmd */
652 struct cmap classifiers
;
653 /* Periodically sort subtable vectors according to hit frequencies */
654 long long int next_optimization
;
655 /* End of the next time interval for which processing cycles
656 are stored for each polled rxq. */
657 long long int rxq_next_cycle_store
;
659 /* Last interval timestamp. */
660 uint64_t intrvl_tsc_prev
;
661 /* Last interval cycles. */
662 atomic_ullong intrvl_cycles
;
664 /* Current context of the PMD thread. */
665 struct dp_netdev_pmd_thread_ctx ctx
;
667 struct seq
*reload_seq
;
668 uint64_t last_reload_seq
;
670 /* These are atomic variables used as a synchronization and configuration
671 * points for thread reload/exit.
673 * 'reload' atomic is the main one and it's used as a memory
674 * synchronization point for all other knobs and data.
676 * For a thread that requests PMD reload:
678 * * All changes that should be visible to the PMD thread must be made
679 * before setting the 'reload'. These changes could use any memory
680 * ordering model including 'relaxed'.
681 * * Setting the 'reload' atomic should occur in the same thread where
682 * all other PMD configuration options updated.
683 * * Setting the 'reload' atomic should be done with 'release' memory
684 * ordering model or stricter. This will guarantee that all previous
685 * changes (including non-atomic and 'relaxed') will be visible to
687 * * To check that reload is done, thread should poll the 'reload' atomic
688 * to become 'false'. Polling should be done with 'acquire' memory
689 * ordering model or stricter. This ensures that PMD thread completed
690 * the reload process.
692 * For the PMD thread:
694 * * PMD thread should read 'reload' atomic with 'acquire' memory
695 * ordering model or stricter. This will guarantee that all changes
696 * made before setting the 'reload' in the requesting thread will be
697 * visible to the PMD thread.
698 * * All other configuration data could be read with any memory
699 * ordering model (including non-atomic and 'relaxed') but *only after*
700 * reading the 'reload' atomic set to 'true'.
701 * * When the PMD reload done, PMD should (optionally) set all the below
702 * knobs except the 'reload' to their default ('false') values and
703 * (mandatory), as the last step, set the 'reload' to 'false' using
704 * 'release' memory ordering model or stricter. This will inform the
705 * requesting thread that PMD has completed a reload cycle.
707 atomic_bool reload
; /* Do we need to reload ports? */
708 atomic_bool wait_for_reload
; /* Can we busy wait for the next reload? */
709 atomic_bool reload_tx_qid
; /* Do we need to reload static_tx_qid? */
710 atomic_bool exit
; /* For terminating the pmd thread. */
713 unsigned core_id
; /* CPU core id of this pmd thread. */
714 int numa_id
; /* numa node id of this pmd thread. */
717 /* Queue id used by this pmd thread to send packets on all netdevs if
718 * XPS disabled for this netdev. All static_tx_qid's are unique and less
719 * than 'cmap_count(dp->poll_threads)'. */
720 uint32_t static_tx_qid
;
722 /* Number of filled output batches. */
723 int n_output_batches
;
725 struct ovs_mutex port_mutex
; /* Mutex for 'poll_list' and 'tx_ports'. */
726 /* List of rx queues to poll. */
727 struct hmap poll_list OVS_GUARDED
;
728 /* Map of 'tx_port's used for transmission. Written by the main thread,
729 * read by the pmd thread. */
730 struct hmap tx_ports OVS_GUARDED
;
732 /* These are thread-local copies of 'tx_ports'. One contains only tunnel
733 * ports (that support push_tunnel/pop_tunnel), the other contains ports
734 * with at least one txq (that support send). A port can be in both.
736 * There are two separate maps to make sure that we don't try to execute
737 * OUTPUT on a device which has 0 txqs or PUSH/POP on a non-tunnel device.
739 * The instances for cpu core NON_PMD_CORE_ID can be accessed by multiple
740 * threads, and thusly need to be protected by 'non_pmd_mutex'. Every
741 * other instance will only be accessed by its own pmd thread. */
742 struct hmap tnl_port_cache
;
743 struct hmap send_port_cache
;
745 /* Keep track of detailed PMD performance statistics. */
746 struct pmd_perf_stats perf_stats
;
748 /* Stats from previous iteration used by automatic pmd
749 * load balance logic. */
750 uint64_t prev_stats
[PMD_N_STATS
];
751 atomic_count pmd_overloaded
;
753 /* Set to true if the pmd thread needs to be reloaded. */
757 /* Interface to netdev-based datapath. */
760 struct dp_netdev
*dp
;
761 uint64_t last_port_seq
;
764 static int get_port_by_number(struct dp_netdev
*dp
, odp_port_t port_no
,
765 struct dp_netdev_port
**portp
)
766 OVS_REQUIRES(dp
->port_mutex
);
767 static int get_port_by_name(struct dp_netdev
*dp
, const char *devname
,
768 struct dp_netdev_port
**portp
)
769 OVS_REQUIRES(dp
->port_mutex
);
770 static void dp_netdev_free(struct dp_netdev
*)
771 OVS_REQUIRES(dp_netdev_mutex
);
772 static int do_add_port(struct dp_netdev
*dp
, const char *devname
,
773 const char *type
, odp_port_t port_no
)
774 OVS_REQUIRES(dp
->port_mutex
);
775 static void do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*)
776 OVS_REQUIRES(dp
->port_mutex
);
777 static int dpif_netdev_open(const struct dpif_class
*, const char *name
,
778 bool create
, struct dpif
**);
779 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
780 struct dp_packet_batch
*,
782 const struct flow
*flow
,
783 const struct nlattr
*actions
,
785 static void dp_netdev_input(struct dp_netdev_pmd_thread
*,
786 struct dp_packet_batch
*, odp_port_t port_no
);
787 static void dp_netdev_recirculate(struct dp_netdev_pmd_thread
*,
788 struct dp_packet_batch
*);
790 static void dp_netdev_disable_upcall(struct dp_netdev
*);
791 static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
);
792 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
,
793 struct dp_netdev
*dp
, unsigned core_id
,
795 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
);
796 static void dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
797 OVS_REQUIRES(dp
->port_mutex
);
799 static void *pmd_thread_main(void *);
800 static struct dp_netdev_pmd_thread
*dp_netdev_get_pmd(struct dp_netdev
*dp
,
802 static struct dp_netdev_pmd_thread
*
803 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
);
804 static void dp_netdev_del_pmd(struct dp_netdev
*dp
,
805 struct dp_netdev_pmd_thread
*pmd
);
806 static void dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
, bool non_pmd
);
807 static void dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
);
808 static void dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
809 struct dp_netdev_port
*port
)
810 OVS_REQUIRES(pmd
->port_mutex
);
811 static void dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
813 OVS_REQUIRES(pmd
->port_mutex
);
814 static void dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
815 struct dp_netdev_rxq
*rxq
)
816 OVS_REQUIRES(pmd
->port_mutex
);
817 static void dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
818 struct rxq_poll
*poll
)
819 OVS_REQUIRES(pmd
->port_mutex
);
821 dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread
*pmd
,
824 static void reconfigure_datapath(struct dp_netdev
*dp
)
825 OVS_REQUIRES(dp
->port_mutex
);
826 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
);
827 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
);
828 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
);
829 static void pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
830 OVS_REQUIRES(pmd
->port_mutex
);
832 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
,
833 struct polled_queue
*poll_list
, int poll_cnt
);
835 dp_netdev_rxq_set_cycles(struct dp_netdev_rxq
*rx
,
836 enum rxq_cycles_counter_type type
,
837 unsigned long long cycles
);
839 dp_netdev_rxq_get_cycles(struct dp_netdev_rxq
*rx
,
840 enum rxq_cycles_counter_type type
);
842 dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq
*rx
,
843 unsigned long long cycles
);
845 dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq
*rx
, unsigned idx
);
847 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
849 static int dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
852 static inline bool emc_entry_alive(struct emc_entry
*ce
);
853 static void emc_clear_entry(struct emc_entry
*ce
);
854 static void smc_clear_entry(struct smc_bucket
*b
, int idx
);
856 static void dp_netdev_request_reconfigure(struct dp_netdev
*dp
);
858 pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread
*pmd
);
859 static void queue_netdev_flow_del(struct dp_netdev_pmd_thread
*pmd
,
860 struct dp_netdev_flow
*flow
);
863 emc_cache_init(struct emc_cache
*flow_cache
)
867 flow_cache
->sweep_idx
= 0;
868 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
869 flow_cache
->entries
[i
].flow
= NULL
;
870 flow_cache
->entries
[i
].key
.hash
= 0;
871 flow_cache
->entries
[i
].key
.len
= sizeof(struct miniflow
);
872 flowmap_init(&flow_cache
->entries
[i
].key
.mf
.map
);
877 smc_cache_init(struct smc_cache
*smc_cache
)
880 for (i
= 0; i
< SMC_BUCKET_CNT
; i
++) {
881 for (j
= 0; j
< SMC_ENTRY_PER_BUCKET
; j
++) {
882 smc_cache
->buckets
[i
].flow_idx
[j
] = UINT16_MAX
;
888 dfc_cache_init(struct dfc_cache
*flow_cache
)
890 emc_cache_init(&flow_cache
->emc_cache
);
891 smc_cache_init(&flow_cache
->smc_cache
);
895 emc_cache_uninit(struct emc_cache
*flow_cache
)
899 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
900 emc_clear_entry(&flow_cache
->entries
[i
]);
905 smc_cache_uninit(struct smc_cache
*smc
)
909 for (i
= 0; i
< SMC_BUCKET_CNT
; i
++) {
910 for (j
= 0; j
< SMC_ENTRY_PER_BUCKET
; j
++) {
911 smc_clear_entry(&(smc
->buckets
[i
]), j
);
917 dfc_cache_uninit(struct dfc_cache
*flow_cache
)
919 smc_cache_uninit(&flow_cache
->smc_cache
);
920 emc_cache_uninit(&flow_cache
->emc_cache
);
923 /* Check and clear dead flow references slowly (one entry at each
926 emc_cache_slow_sweep(struct emc_cache
*flow_cache
)
928 struct emc_entry
*entry
= &flow_cache
->entries
[flow_cache
->sweep_idx
];
930 if (!emc_entry_alive(entry
)) {
931 emc_clear_entry(entry
);
933 flow_cache
->sweep_idx
= (flow_cache
->sweep_idx
+ 1) & EM_FLOW_HASH_MASK
;
936 /* Updates the time in PMD threads context and should be called in three cases:
938 * 1. PMD structure initialization:
939 * - dp_netdev_configure_pmd()
941 * 2. Before processing of the new packet batch:
942 * - dpif_netdev_execute()
943 * - dp_netdev_process_rxq_port()
945 * 3. At least once per polling iteration in main polling threads if no
946 * packets received on current iteration:
947 * - dpif_netdev_run()
948 * - pmd_thread_main()
950 * 'pmd->ctx.now' should be used without update in all other cases if possible.
953 pmd_thread_ctx_time_update(struct dp_netdev_pmd_thread
*pmd
)
955 pmd
->ctx
.now
= time_usec();
958 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
960 dpif_is_netdev(const struct dpif
*dpif
)
962 return dpif
->dpif_class
->open
== dpif_netdev_open
;
965 static struct dpif_netdev
*
966 dpif_netdev_cast(const struct dpif
*dpif
)
968 ovs_assert(dpif_is_netdev(dpif
));
969 return CONTAINER_OF(dpif
, struct dpif_netdev
, dpif
);
972 static struct dp_netdev
*
973 get_dp_netdev(const struct dpif
*dpif
)
975 return dpif_netdev_cast(dpif
)->dp
;
979 PMD_INFO_SHOW_STATS
, /* Show how cpu cycles are spent. */
980 PMD_INFO_CLEAR_STATS
, /* Set the cycles count to 0. */
981 PMD_INFO_SHOW_RXQ
, /* Show poll lists of pmd threads. */
982 PMD_INFO_PERF_SHOW
, /* Show pmd performance details. */
986 format_pmd_thread(struct ds
*reply
, struct dp_netdev_pmd_thread
*pmd
)
988 ds_put_cstr(reply
, (pmd
->core_id
== NON_PMD_CORE_ID
)
989 ? "main thread" : "pmd thread");
990 if (pmd
->numa_id
!= OVS_NUMA_UNSPEC
) {
991 ds_put_format(reply
, " numa_id %d", pmd
->numa_id
);
993 if (pmd
->core_id
!= OVS_CORE_UNSPEC
&& pmd
->core_id
!= NON_PMD_CORE_ID
) {
994 ds_put_format(reply
, " core_id %u", pmd
->core_id
);
996 ds_put_cstr(reply
, ":\n");
1000 pmd_info_show_stats(struct ds
*reply
,
1001 struct dp_netdev_pmd_thread
*pmd
)
1003 uint64_t stats
[PMD_N_STATS
];
1004 uint64_t total_cycles
, total_packets
;
1005 double passes_per_pkt
= 0;
1006 double lookups_per_hit
= 0;
1007 double packets_per_batch
= 0;
1009 pmd_perf_read_counters(&pmd
->perf_stats
, stats
);
1010 total_cycles
= stats
[PMD_CYCLES_ITER_IDLE
]
1011 + stats
[PMD_CYCLES_ITER_BUSY
];
1012 total_packets
= stats
[PMD_STAT_RECV
];
1014 format_pmd_thread(reply
, pmd
);
1016 if (total_packets
> 0) {
1017 passes_per_pkt
= (total_packets
+ stats
[PMD_STAT_RECIRC
])
1018 / (double) total_packets
;
1020 if (stats
[PMD_STAT_MASKED_HIT
] > 0) {
1021 lookups_per_hit
= stats
[PMD_STAT_MASKED_LOOKUP
]
1022 / (double) stats
[PMD_STAT_MASKED_HIT
];
1024 if (stats
[PMD_STAT_SENT_BATCHES
] > 0) {
1025 packets_per_batch
= stats
[PMD_STAT_SENT_PKTS
]
1026 / (double) stats
[PMD_STAT_SENT_BATCHES
];
1029 ds_put_format(reply
,
1030 " packets received: %"PRIu64
"\n"
1031 " packet recirculations: %"PRIu64
"\n"
1032 " avg. datapath passes per packet: %.02f\n"
1033 " emc hits: %"PRIu64
"\n"
1034 " smc hits: %"PRIu64
"\n"
1035 " megaflow hits: %"PRIu64
"\n"
1036 " avg. subtable lookups per megaflow hit: %.02f\n"
1037 " miss with success upcall: %"PRIu64
"\n"
1038 " miss with failed upcall: %"PRIu64
"\n"
1039 " avg. packets per output batch: %.02f\n",
1040 total_packets
, stats
[PMD_STAT_RECIRC
],
1041 passes_per_pkt
, stats
[PMD_STAT_EXACT_HIT
],
1042 stats
[PMD_STAT_SMC_HIT
],
1043 stats
[PMD_STAT_MASKED_HIT
], lookups_per_hit
,
1044 stats
[PMD_STAT_MISS
], stats
[PMD_STAT_LOST
],
1047 if (total_cycles
== 0) {
1051 ds_put_format(reply
,
1052 " idle cycles: %"PRIu64
" (%.02f%%)\n"
1053 " processing cycles: %"PRIu64
" (%.02f%%)\n",
1054 stats
[PMD_CYCLES_ITER_IDLE
],
1055 stats
[PMD_CYCLES_ITER_IDLE
] / (double) total_cycles
* 100,
1056 stats
[PMD_CYCLES_ITER_BUSY
],
1057 stats
[PMD_CYCLES_ITER_BUSY
] / (double) total_cycles
* 100);
1059 if (total_packets
== 0) {
1063 ds_put_format(reply
,
1064 " avg cycles per packet: %.02f (%"PRIu64
"/%"PRIu64
")\n",
1065 total_cycles
/ (double) total_packets
,
1066 total_cycles
, total_packets
);
1068 ds_put_format(reply
,
1069 " avg processing cycles per packet: "
1070 "%.02f (%"PRIu64
"/%"PRIu64
")\n",
1071 stats
[PMD_CYCLES_ITER_BUSY
] / (double) total_packets
,
1072 stats
[PMD_CYCLES_ITER_BUSY
], total_packets
);
1076 pmd_info_show_perf(struct ds
*reply
,
1077 struct dp_netdev_pmd_thread
*pmd
,
1078 struct pmd_perf_params
*par
)
1080 if (pmd
->core_id
!= NON_PMD_CORE_ID
) {
1082 xastrftime_msec("%H:%M:%S.###", time_wall_msec(), true);
1083 long long now
= time_msec();
1084 double duration
= (now
- pmd
->perf_stats
.start_ms
) / 1000.0;
1086 ds_put_cstr(reply
, "\n");
1087 ds_put_format(reply
, "Time: %s\n", time_str
);
1088 ds_put_format(reply
, "Measurement duration: %.3f s\n", duration
);
1089 ds_put_cstr(reply
, "\n");
1090 format_pmd_thread(reply
, pmd
);
1091 ds_put_cstr(reply
, "\n");
1092 pmd_perf_format_overall_stats(reply
, &pmd
->perf_stats
, duration
);
1093 if (pmd_perf_metrics_enabled(pmd
)) {
1094 /* Prevent parallel clearing of perf metrics. */
1095 ovs_mutex_lock(&pmd
->perf_stats
.clear_mutex
);
1096 if (par
->histograms
) {
1097 ds_put_cstr(reply
, "\n");
1098 pmd_perf_format_histograms(reply
, &pmd
->perf_stats
);
1100 if (par
->iter_hist_len
> 0) {
1101 ds_put_cstr(reply
, "\n");
1102 pmd_perf_format_iteration_history(reply
, &pmd
->perf_stats
,
1103 par
->iter_hist_len
);
1105 if (par
->ms_hist_len
> 0) {
1106 ds_put_cstr(reply
, "\n");
1107 pmd_perf_format_ms_history(reply
, &pmd
->perf_stats
,
1110 ovs_mutex_unlock(&pmd
->perf_stats
.clear_mutex
);
1117 compare_poll_list(const void *a_
, const void *b_
)
1119 const struct rxq_poll
*a
= a_
;
1120 const struct rxq_poll
*b
= b_
;
1122 const char *namea
= netdev_rxq_get_name(a
->rxq
->rx
);
1123 const char *nameb
= netdev_rxq_get_name(b
->rxq
->rx
);
1125 int cmp
= strcmp(namea
, nameb
);
1127 return netdev_rxq_get_queue_id(a
->rxq
->rx
)
1128 - netdev_rxq_get_queue_id(b
->rxq
->rx
);
1135 sorted_poll_list(struct dp_netdev_pmd_thread
*pmd
, struct rxq_poll
**list
,
1137 OVS_REQUIRES(pmd
->port_mutex
)
1139 struct rxq_poll
*ret
, *poll
;
1142 *n
= hmap_count(&pmd
->poll_list
);
1146 ret
= xcalloc(*n
, sizeof *ret
);
1148 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
1152 ovs_assert(i
== *n
);
1153 qsort(ret
, *n
, sizeof *ret
, compare_poll_list
);
1160 pmd_info_show_rxq(struct ds
*reply
, struct dp_netdev_pmd_thread
*pmd
)
1162 if (pmd
->core_id
!= NON_PMD_CORE_ID
) {
1163 struct rxq_poll
*list
;
1165 uint64_t total_cycles
= 0;
1167 ds_put_format(reply
,
1168 "pmd thread numa_id %d core_id %u:\n isolated : %s\n",
1169 pmd
->numa_id
, pmd
->core_id
, (pmd
->isolated
)
1170 ? "true" : "false");
1172 ovs_mutex_lock(&pmd
->port_mutex
);
1173 sorted_poll_list(pmd
, &list
, &n_rxq
);
1175 /* Get the total pmd cycles for an interval. */
1176 atomic_read_relaxed(&pmd
->intrvl_cycles
, &total_cycles
);
1177 /* Estimate the cycles to cover all intervals. */
1178 total_cycles
*= PMD_RXQ_INTERVAL_MAX
;
1180 for (int i
= 0; i
< n_rxq
; i
++) {
1181 struct dp_netdev_rxq
*rxq
= list
[i
].rxq
;
1182 const char *name
= netdev_rxq_get_name(rxq
->rx
);
1183 uint64_t proc_cycles
= 0;
1185 for (int j
= 0; j
< PMD_RXQ_INTERVAL_MAX
; j
++) {
1186 proc_cycles
+= dp_netdev_rxq_get_intrvl_cycles(rxq
, j
);
1188 ds_put_format(reply
, " port: %-16s queue-id: %2d", name
,
1189 netdev_rxq_get_queue_id(list
[i
].rxq
->rx
));
1190 ds_put_format(reply
, " %s", netdev_rxq_enabled(list
[i
].rxq
->rx
)
1191 ? "(enabled) " : "(disabled)");
1192 ds_put_format(reply
, " pmd usage: ");
1194 ds_put_format(reply
, "%2"PRIu64
"",
1195 proc_cycles
* 100 / total_cycles
);
1196 ds_put_cstr(reply
, " %");
1198 ds_put_format(reply
, "%s", "NOT AVAIL");
1200 ds_put_cstr(reply
, "\n");
1202 ovs_mutex_unlock(&pmd
->port_mutex
);
1208 compare_poll_thread_list(const void *a_
, const void *b_
)
1210 const struct dp_netdev_pmd_thread
*a
, *b
;
1212 a
= *(struct dp_netdev_pmd_thread
**)a_
;
1213 b
= *(struct dp_netdev_pmd_thread
**)b_
;
1215 if (a
->core_id
< b
->core_id
) {
1218 if (a
->core_id
> b
->core_id
) {
1224 /* Create a sorted list of pmd's from the dp->poll_threads cmap. We can use
1225 * this list, as long as we do not go to quiescent state. */
1227 sorted_poll_thread_list(struct dp_netdev
*dp
,
1228 struct dp_netdev_pmd_thread
***list
,
1231 struct dp_netdev_pmd_thread
*pmd
;
1232 struct dp_netdev_pmd_thread
**pmd_list
;
1233 size_t k
= 0, n_pmds
;
1235 n_pmds
= cmap_count(&dp
->poll_threads
);
1236 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
1238 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1242 pmd_list
[k
++] = pmd
;
1245 qsort(pmd_list
, k
, sizeof *pmd_list
, compare_poll_thread_list
);
1252 dpif_netdev_pmd_rebalance(struct unixctl_conn
*conn
, int argc
,
1253 const char *argv
[], void *aux OVS_UNUSED
)
1255 struct ds reply
= DS_EMPTY_INITIALIZER
;
1256 struct dp_netdev
*dp
= NULL
;
1258 ovs_mutex_lock(&dp_netdev_mutex
);
1261 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
1262 } else if (shash_count(&dp_netdevs
) == 1) {
1263 /* There's only one datapath */
1264 dp
= shash_first(&dp_netdevs
)->data
;
1268 ovs_mutex_unlock(&dp_netdev_mutex
);
1269 unixctl_command_reply_error(conn
,
1270 "please specify an existing datapath");
1274 dp_netdev_request_reconfigure(dp
);
1275 ovs_mutex_unlock(&dp_netdev_mutex
);
1276 ds_put_cstr(&reply
, "pmd rxq rebalance requested.\n");
1277 unixctl_command_reply(conn
, ds_cstr(&reply
));
1282 dpif_netdev_pmd_info(struct unixctl_conn
*conn
, int argc
, const char *argv
[],
1285 struct ds reply
= DS_EMPTY_INITIALIZER
;
1286 struct dp_netdev_pmd_thread
**pmd_list
;
1287 struct dp_netdev
*dp
= NULL
;
1288 enum pmd_info_type type
= *(enum pmd_info_type
*) aux
;
1289 unsigned int core_id
;
1290 bool filter_on_pmd
= false;
1293 ovs_mutex_lock(&dp_netdev_mutex
);
1296 if (!strcmp(argv
[1], "-pmd") && argc
> 2) {
1297 if (str_to_uint(argv
[2], 10, &core_id
)) {
1298 filter_on_pmd
= true;
1303 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
1310 if (shash_count(&dp_netdevs
) == 1) {
1311 /* There's only one datapath */
1312 dp
= shash_first(&dp_netdevs
)->data
;
1314 ovs_mutex_unlock(&dp_netdev_mutex
);
1315 unixctl_command_reply_error(conn
,
1316 "please specify an existing datapath");
1321 sorted_poll_thread_list(dp
, &pmd_list
, &n
);
1322 for (size_t i
= 0; i
< n
; i
++) {
1323 struct dp_netdev_pmd_thread
*pmd
= pmd_list
[i
];
1327 if (filter_on_pmd
&& pmd
->core_id
!= core_id
) {
1330 if (type
== PMD_INFO_SHOW_RXQ
) {
1331 pmd_info_show_rxq(&reply
, pmd
);
1332 } else if (type
== PMD_INFO_CLEAR_STATS
) {
1333 pmd_perf_stats_clear(&pmd
->perf_stats
);
1334 } else if (type
== PMD_INFO_SHOW_STATS
) {
1335 pmd_info_show_stats(&reply
, pmd
);
1336 } else if (type
== PMD_INFO_PERF_SHOW
) {
1337 pmd_info_show_perf(&reply
, pmd
, (struct pmd_perf_params
*)aux
);
1342 ovs_mutex_unlock(&dp_netdev_mutex
);
1344 unixctl_command_reply(conn
, ds_cstr(&reply
));
1349 pmd_perf_show_cmd(struct unixctl_conn
*conn
, int argc
,
1351 void *aux OVS_UNUSED
)
1353 struct pmd_perf_params par
;
1354 long int it_hist
= 0, ms_hist
= 0;
1355 par
.histograms
= true;
1358 if (!strcmp(argv
[1], "-nh")) {
1359 par
.histograms
= false;
1362 } else if (!strcmp(argv
[1], "-it") && argc
> 2) {
1363 it_hist
= strtol(argv
[2], NULL
, 10);
1366 } else if (it_hist
> HISTORY_LEN
) {
1367 it_hist
= HISTORY_LEN
;
1371 } else if (!strcmp(argv
[1], "-ms") && argc
> 2) {
1372 ms_hist
= strtol(argv
[2], NULL
, 10);
1375 } else if (ms_hist
> HISTORY_LEN
) {
1376 ms_hist
= HISTORY_LEN
;
1384 par
.iter_hist_len
= it_hist
;
1385 par
.ms_hist_len
= ms_hist
;
1386 par
.command_type
= PMD_INFO_PERF_SHOW
;
1387 dpif_netdev_pmd_info(conn
, argc
, argv
, &par
);
1391 dpif_netdev_init(void)
1393 static enum pmd_info_type show_aux
= PMD_INFO_SHOW_STATS
,
1394 clear_aux
= PMD_INFO_CLEAR_STATS
,
1395 poll_aux
= PMD_INFO_SHOW_RXQ
;
1397 unixctl_command_register("dpif-netdev/pmd-stats-show", "[-pmd core] [dp]",
1398 0, 3, dpif_netdev_pmd_info
,
1400 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[-pmd core] [dp]",
1401 0, 3, dpif_netdev_pmd_info
,
1402 (void *)&clear_aux
);
1403 unixctl_command_register("dpif-netdev/pmd-rxq-show", "[-pmd core] [dp]",
1404 0, 3, dpif_netdev_pmd_info
,
1406 unixctl_command_register("dpif-netdev/pmd-perf-show",
1407 "[-nh] [-it iter-history-len]"
1408 " [-ms ms-history-len]"
1409 " [-pmd core] [dp]",
1410 0, 8, pmd_perf_show_cmd
,
1412 unixctl_command_register("dpif-netdev/pmd-rxq-rebalance", "[dp]",
1413 0, 1, dpif_netdev_pmd_rebalance
,
1415 unixctl_command_register("dpif-netdev/pmd-perf-log-set",
1416 "on|off [-b before] [-a after] [-e|-ne] "
1417 "[-us usec] [-q qlen]",
1418 0, 10, pmd_perf_log_set_cmd
,
1424 dpif_netdev_enumerate(struct sset
*all_dps
,
1425 const struct dpif_class
*dpif_class
)
1427 struct shash_node
*node
;
1429 ovs_mutex_lock(&dp_netdev_mutex
);
1430 SHASH_FOR_EACH(node
, &dp_netdevs
) {
1431 struct dp_netdev
*dp
= node
->data
;
1432 if (dpif_class
!= dp
->class) {
1433 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
1434 * If the class doesn't match, skip this dpif. */
1437 sset_add(all_dps
, node
->name
);
1439 ovs_mutex_unlock(&dp_netdev_mutex
);
1445 dpif_netdev_class_is_dummy(const struct dpif_class
*class)
1447 return class != &dpif_netdev_class
;
1451 dpif_netdev_port_open_type(const struct dpif_class
*class, const char *type
)
1453 return strcmp(type
, "internal") ? type
1454 : dpif_netdev_class_is_dummy(class) ? "dummy-internal"
1458 static struct dpif
*
1459 create_dpif_netdev(struct dp_netdev
*dp
)
1461 uint16_t netflow_id
= hash_string(dp
->name
, 0);
1462 struct dpif_netdev
*dpif
;
1464 ovs_refcount_ref(&dp
->ref_cnt
);
1466 dpif
= xmalloc(sizeof *dpif
);
1467 dpif_init(&dpif
->dpif
, dp
->class, dp
->name
, netflow_id
>> 8, netflow_id
);
1469 dpif
->last_port_seq
= seq_read(dp
->port_seq
);
1474 /* Choose an unused, non-zero port number and return it on success.
1475 * Return ODPP_NONE on failure. */
1477 choose_port(struct dp_netdev
*dp
, const char *name
)
1478 OVS_REQUIRES(dp
->port_mutex
)
1482 if (dp
->class != &dpif_netdev_class
) {
1486 /* If the port name begins with "br", start the number search at
1487 * 100 to make writing tests easier. */
1488 if (!strncmp(name
, "br", 2)) {
1492 /* If the port name contains a number, try to assign that port number.
1493 * This can make writing unit tests easier because port numbers are
1495 for (p
= name
; *p
!= '\0'; p
++) {
1496 if (isdigit((unsigned char) *p
)) {
1497 port_no
= start_no
+ strtol(p
, NULL
, 10);
1498 if (port_no
> 0 && port_no
!= odp_to_u32(ODPP_NONE
)
1499 && !dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
1500 return u32_to_odp(port_no
);
1507 for (port_no
= 1; port_no
<= UINT16_MAX
; port_no
++) {
1508 if (!dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
1509 return u32_to_odp(port_no
);
1517 create_dp_netdev(const char *name
, const struct dpif_class
*class,
1518 struct dp_netdev
**dpp
)
1519 OVS_REQUIRES(dp_netdev_mutex
)
1521 struct dp_netdev
*dp
;
1524 dp
= xzalloc(sizeof *dp
);
1525 shash_add(&dp_netdevs
, name
, dp
);
1527 *CONST_CAST(const struct dpif_class
**, &dp
->class) = class;
1528 *CONST_CAST(const char **, &dp
->name
) = xstrdup(name
);
1529 ovs_refcount_init(&dp
->ref_cnt
);
1530 atomic_flag_clear(&dp
->destroyed
);
1532 ovs_mutex_init(&dp
->port_mutex
);
1533 hmap_init(&dp
->ports
);
1534 dp
->port_seq
= seq_create();
1535 fat_rwlock_init(&dp
->upcall_rwlock
);
1537 dp
->reconfigure_seq
= seq_create();
1538 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
1540 for (int i
= 0; i
< N_METER_LOCKS
; ++i
) {
1541 ovs_mutex_init_adaptive(&dp
->meter_locks
[i
]);
1544 /* Disable upcalls by default. */
1545 dp_netdev_disable_upcall(dp
);
1546 dp
->upcall_aux
= NULL
;
1547 dp
->upcall_cb
= NULL
;
1549 dp
->conntrack
= conntrack_init();
1551 atomic_init(&dp
->emc_insert_min
, DEFAULT_EM_FLOW_INSERT_MIN
);
1552 atomic_init(&dp
->tx_flush_interval
, DEFAULT_TX_FLUSH_INTERVAL
);
1554 cmap_init(&dp
->poll_threads
);
1555 dp
->pmd_rxq_assign_cyc
= true;
1557 ovs_mutex_init(&dp
->tx_qid_pool_mutex
);
1558 /* We need 1 Tx queue for each possible core + 1 for non-PMD threads. */
1559 dp
->tx_qid_pool
= id_pool_create(0, ovs_numa_get_n_cores() + 1);
1561 ovs_mutex_init_recursive(&dp
->non_pmd_mutex
);
1562 ovsthread_key_create(&dp
->per_pmd_key
, NULL
);
1564 ovs_mutex_lock(&dp
->port_mutex
);
1565 /* non-PMD will be created before all other threads and will
1566 * allocate static_tx_qid = 0. */
1567 dp_netdev_set_nonpmd(dp
);
1569 error
= do_add_port(dp
, name
, dpif_netdev_port_open_type(dp
->class,
1572 ovs_mutex_unlock(&dp
->port_mutex
);
1578 dp
->last_tnl_conf_seq
= seq_read(tnl_conf_seq
);
1584 dp_netdev_request_reconfigure(struct dp_netdev
*dp
)
1586 seq_change(dp
->reconfigure_seq
);
1590 dp_netdev_is_reconf_required(struct dp_netdev
*dp
)
1592 return seq_read(dp
->reconfigure_seq
) != dp
->last_reconfigure_seq
;
1596 dpif_netdev_open(const struct dpif_class
*class, const char *name
,
1597 bool create
, struct dpif
**dpifp
)
1599 struct dp_netdev
*dp
;
1602 ovs_mutex_lock(&dp_netdev_mutex
);
1603 dp
= shash_find_data(&dp_netdevs
, name
);
1605 error
= create
? create_dp_netdev(name
, class, &dp
) : ENODEV
;
1607 error
= (dp
->class != class ? EINVAL
1612 *dpifp
= create_dpif_netdev(dp
);
1615 ovs_mutex_unlock(&dp_netdev_mutex
);
1621 dp_netdev_destroy_upcall_lock(struct dp_netdev
*dp
)
1622 OVS_NO_THREAD_SAFETY_ANALYSIS
1624 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
1625 ovs_assert(fat_rwlock_tryrdlock(&dp
->upcall_rwlock
));
1627 /* Before freeing a lock we should release it */
1628 fat_rwlock_unlock(&dp
->upcall_rwlock
);
1629 fat_rwlock_destroy(&dp
->upcall_rwlock
);
1633 dp_delete_meter(struct dp_netdev
*dp
, uint32_t meter_id
)
1634 OVS_REQUIRES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
1636 if (dp
->meters
[meter_id
]) {
1637 free(dp
->meters
[meter_id
]);
1638 dp
->meters
[meter_id
] = NULL
;
1642 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
1643 * through the 'dp_netdevs' shash while freeing 'dp'. */
1645 dp_netdev_free(struct dp_netdev
*dp
)
1646 OVS_REQUIRES(dp_netdev_mutex
)
1648 struct dp_netdev_port
*port
, *next
;
1650 shash_find_and_delete(&dp_netdevs
, dp
->name
);
1652 ovs_mutex_lock(&dp
->port_mutex
);
1653 HMAP_FOR_EACH_SAFE (port
, next
, node
, &dp
->ports
) {
1654 do_del_port(dp
, port
);
1656 ovs_mutex_unlock(&dp
->port_mutex
);
1658 dp_netdev_destroy_all_pmds(dp
, true);
1659 cmap_destroy(&dp
->poll_threads
);
1661 ovs_mutex_destroy(&dp
->tx_qid_pool_mutex
);
1662 id_pool_destroy(dp
->tx_qid_pool
);
1664 ovs_mutex_destroy(&dp
->non_pmd_mutex
);
1665 ovsthread_key_delete(dp
->per_pmd_key
);
1667 conntrack_destroy(dp
->conntrack
);
1670 seq_destroy(dp
->reconfigure_seq
);
1672 seq_destroy(dp
->port_seq
);
1673 hmap_destroy(&dp
->ports
);
1674 ovs_mutex_destroy(&dp
->port_mutex
);
1676 /* Upcalls must be disabled at this point */
1677 dp_netdev_destroy_upcall_lock(dp
);
1681 for (i
= 0; i
< MAX_METERS
; ++i
) {
1683 dp_delete_meter(dp
, i
);
1684 meter_unlock(dp
, i
);
1686 for (i
= 0; i
< N_METER_LOCKS
; ++i
) {
1687 ovs_mutex_destroy(&dp
->meter_locks
[i
]);
1690 free(dp
->pmd_cmask
);
1691 free(CONST_CAST(char *, dp
->name
));
1696 dp_netdev_unref(struct dp_netdev
*dp
)
1699 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
1700 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
1701 ovs_mutex_lock(&dp_netdev_mutex
);
1702 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1705 ovs_mutex_unlock(&dp_netdev_mutex
);
1710 dpif_netdev_close(struct dpif
*dpif
)
1712 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1714 dp_netdev_unref(dp
);
1719 dpif_netdev_destroy(struct dpif
*dpif
)
1721 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1723 if (!atomic_flag_test_and_set(&dp
->destroyed
)) {
1724 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1725 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
1733 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
1734 * load/store semantics. While the increment is not atomic, the load and
1735 * store operations are, making it impossible to read inconsistent values.
1737 * This is used to update thread local stats counters. */
1739 non_atomic_ullong_add(atomic_ullong
*var
, unsigned long long n
)
1741 unsigned long long tmp
;
1743 atomic_read_relaxed(var
, &tmp
);
1745 atomic_store_relaxed(var
, tmp
);
1749 dpif_netdev_get_stats(const struct dpif
*dpif
, struct dpif_dp_stats
*stats
)
1751 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1752 struct dp_netdev_pmd_thread
*pmd
;
1753 uint64_t pmd_stats
[PMD_N_STATS
];
1755 stats
->n_flows
= stats
->n_hit
= stats
->n_missed
= stats
->n_lost
= 0;
1756 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1757 stats
->n_flows
+= cmap_count(&pmd
->flow_table
);
1758 pmd_perf_read_counters(&pmd
->perf_stats
, pmd_stats
);
1759 stats
->n_hit
+= pmd_stats
[PMD_STAT_EXACT_HIT
];
1760 stats
->n_hit
+= pmd_stats
[PMD_STAT_SMC_HIT
];
1761 stats
->n_hit
+= pmd_stats
[PMD_STAT_MASKED_HIT
];
1762 stats
->n_missed
+= pmd_stats
[PMD_STAT_MISS
];
1763 stats
->n_lost
+= pmd_stats
[PMD_STAT_LOST
];
1765 stats
->n_masks
= UINT32_MAX
;
1766 stats
->n_mask_hit
= UINT64_MAX
;
1772 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread
*pmd
)
1774 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
1775 ovs_mutex_lock(&pmd
->dp
->non_pmd_mutex
);
1776 ovs_mutex_lock(&pmd
->port_mutex
);
1777 pmd_load_cached_ports(pmd
);
1778 ovs_mutex_unlock(&pmd
->port_mutex
);
1779 ovs_mutex_unlock(&pmd
->dp
->non_pmd_mutex
);
1783 seq_change(pmd
->reload_seq
);
1784 atomic_store_explicit(&pmd
->reload
, true, memory_order_release
);
1788 hash_port_no(odp_port_t port_no
)
1790 return hash_int(odp_to_u32(port_no
), 0);
1794 port_create(const char *devname
, const char *type
,
1795 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1797 struct netdev_saved_flags
*sf
;
1798 struct dp_netdev_port
*port
;
1799 enum netdev_flags flags
;
1800 struct netdev
*netdev
;
1805 /* Open and validate network device. */
1806 error
= netdev_open(devname
, type
, &netdev
);
1810 /* XXX reject non-Ethernet devices */
1812 netdev_get_flags(netdev
, &flags
);
1813 if (flags
& NETDEV_LOOPBACK
) {
1814 VLOG_ERR("%s: cannot add a loopback device", devname
);
1819 error
= netdev_turn_flags_on(netdev
, NETDEV_PROMISC
, &sf
);
1821 VLOG_ERR("%s: cannot set promisc flag", devname
);
1825 port
= xzalloc(sizeof *port
);
1826 port
->port_no
= port_no
;
1827 port
->netdev
= netdev
;
1828 port
->type
= xstrdup(type
);
1830 port
->emc_enabled
= true;
1831 port
->need_reconfigure
= true;
1832 ovs_mutex_init(&port
->txq_used_mutex
);
1839 netdev_close(netdev
);
1844 do_add_port(struct dp_netdev
*dp
, const char *devname
, const char *type
,
1846 OVS_REQUIRES(dp
->port_mutex
)
1848 struct dp_netdev_port
*port
;
1851 /* Reject devices already in 'dp'. */
1852 if (!get_port_by_name(dp
, devname
, &port
)) {
1856 error
= port_create(devname
, type
, port_no
, &port
);
1861 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
1862 seq_change(dp
->port_seq
);
1864 reconfigure_datapath(dp
);
1870 dpif_netdev_port_add(struct dpif
*dpif
, struct netdev
*netdev
,
1871 odp_port_t
*port_nop
)
1873 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1874 char namebuf
[NETDEV_VPORT_NAME_BUFSIZE
];
1875 const char *dpif_port
;
1879 ovs_mutex_lock(&dp
->port_mutex
);
1880 dpif_port
= netdev_vport_get_dpif_port(netdev
, namebuf
, sizeof namebuf
);
1881 if (*port_nop
!= ODPP_NONE
) {
1882 port_no
= *port_nop
;
1883 error
= dp_netdev_lookup_port(dp
, *port_nop
) ? EBUSY
: 0;
1885 port_no
= choose_port(dp
, dpif_port
);
1886 error
= port_no
== ODPP_NONE
? EFBIG
: 0;
1889 *port_nop
= port_no
;
1890 error
= do_add_port(dp
, dpif_port
, netdev_get_type(netdev
), port_no
);
1892 ovs_mutex_unlock(&dp
->port_mutex
);
1898 dpif_netdev_port_del(struct dpif
*dpif
, odp_port_t port_no
)
1900 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1903 ovs_mutex_lock(&dp
->port_mutex
);
1904 if (port_no
== ODPP_LOCAL
) {
1907 struct dp_netdev_port
*port
;
1909 error
= get_port_by_number(dp
, port_no
, &port
);
1911 do_del_port(dp
, port
);
1914 ovs_mutex_unlock(&dp
->port_mutex
);
1920 is_valid_port_number(odp_port_t port_no
)
1922 return port_no
!= ODPP_NONE
;
1925 static struct dp_netdev_port
*
1926 dp_netdev_lookup_port(const struct dp_netdev
*dp
, odp_port_t port_no
)
1927 OVS_REQUIRES(dp
->port_mutex
)
1929 struct dp_netdev_port
*port
;
1931 HMAP_FOR_EACH_WITH_HASH (port
, node
, hash_port_no(port_no
), &dp
->ports
) {
1932 if (port
->port_no
== port_no
) {
1940 get_port_by_number(struct dp_netdev
*dp
,
1941 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1942 OVS_REQUIRES(dp
->port_mutex
)
1944 if (!is_valid_port_number(port_no
)) {
1948 *portp
= dp_netdev_lookup_port(dp
, port_no
);
1949 return *portp
? 0 : ENODEV
;
1954 port_destroy(struct dp_netdev_port
*port
)
1960 netdev_close(port
->netdev
);
1961 netdev_restore_flags(port
->sf
);
1963 for (unsigned i
= 0; i
< port
->n_rxq
; i
++) {
1964 netdev_rxq_close(port
->rxqs
[i
].rx
);
1966 ovs_mutex_destroy(&port
->txq_used_mutex
);
1967 free(port
->rxq_affinity_list
);
1968 free(port
->txq_used
);
1975 get_port_by_name(struct dp_netdev
*dp
,
1976 const char *devname
, struct dp_netdev_port
**portp
)
1977 OVS_REQUIRES(dp
->port_mutex
)
1979 struct dp_netdev_port
*port
;
1981 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1982 if (!strcmp(netdev_get_name(port
->netdev
), devname
)) {
1988 /* Callers of dpif_netdev_port_query_by_name() expect ENODEV for a non
1993 /* Returns 'true' if there is a port with pmd netdev. */
1995 has_pmd_port(struct dp_netdev
*dp
)
1996 OVS_REQUIRES(dp
->port_mutex
)
1998 struct dp_netdev_port
*port
;
2000 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2001 if (netdev_is_pmd(port
->netdev
)) {
2010 do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*port
)
2011 OVS_REQUIRES(dp
->port_mutex
)
2013 hmap_remove(&dp
->ports
, &port
->node
);
2014 seq_change(dp
->port_seq
);
2016 reconfigure_datapath(dp
);
2022 answer_port_query(const struct dp_netdev_port
*port
,
2023 struct dpif_port
*dpif_port
)
2025 dpif_port
->name
= xstrdup(netdev_get_name(port
->netdev
));
2026 dpif_port
->type
= xstrdup(port
->type
);
2027 dpif_port
->port_no
= port
->port_no
;
2031 dpif_netdev_port_query_by_number(const struct dpif
*dpif
, odp_port_t port_no
,
2032 struct dpif_port
*dpif_port
)
2034 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2035 struct dp_netdev_port
*port
;
2038 ovs_mutex_lock(&dp
->port_mutex
);
2039 error
= get_port_by_number(dp
, port_no
, &port
);
2040 if (!error
&& dpif_port
) {
2041 answer_port_query(port
, dpif_port
);
2043 ovs_mutex_unlock(&dp
->port_mutex
);
2049 dpif_netdev_port_query_by_name(const struct dpif
*dpif
, const char *devname
,
2050 struct dpif_port
*dpif_port
)
2052 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2053 struct dp_netdev_port
*port
;
2056 ovs_mutex_lock(&dp
->port_mutex
);
2057 error
= get_port_by_name(dp
, devname
, &port
);
2058 if (!error
&& dpif_port
) {
2059 answer_port_query(port
, dpif_port
);
2061 ovs_mutex_unlock(&dp
->port_mutex
);
2067 dp_netdev_flow_free(struct dp_netdev_flow
*flow
)
2069 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow
));
2073 static void dp_netdev_flow_unref(struct dp_netdev_flow
*flow
)
2075 if (ovs_refcount_unref_relaxed(&flow
->ref_cnt
) == 1) {
2076 ovsrcu_postpone(dp_netdev_flow_free
, flow
);
2081 dp_netdev_flow_hash(const ovs_u128
*ufid
)
2083 return ufid
->u32
[0];
2086 static inline struct dpcls
*
2087 dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread
*pmd
,
2091 uint32_t hash
= hash_port_no(in_port
);
2092 CMAP_FOR_EACH_WITH_HASH (cls
, node
, hash
, &pmd
->classifiers
) {
2093 if (cls
->in_port
== in_port
) {
2094 /* Port classifier exists already */
2101 static inline struct dpcls
*
2102 dp_netdev_pmd_find_dpcls(struct dp_netdev_pmd_thread
*pmd
,
2104 OVS_REQUIRES(pmd
->flow_mutex
)
2106 struct dpcls
*cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2107 uint32_t hash
= hash_port_no(in_port
);
2110 /* Create new classifier for in_port */
2111 cls
= xmalloc(sizeof(*cls
));
2113 cls
->in_port
= in_port
;
2114 cmap_insert(&pmd
->classifiers
, &cls
->node
, hash
);
2115 VLOG_DBG("Creating dpcls %p for in_port %d", cls
, in_port
);
2120 #define MAX_FLOW_MARK (UINT32_MAX - 1)
2121 #define INVALID_FLOW_MARK (UINT32_MAX)
2123 struct megaflow_to_mark_data
{
2124 const struct cmap_node node
;
2130 struct cmap megaflow_to_mark
;
2131 struct cmap mark_to_flow
;
2132 struct id_pool
*pool
;
2135 static struct flow_mark flow_mark
= {
2136 .megaflow_to_mark
= CMAP_INITIALIZER
,
2137 .mark_to_flow
= CMAP_INITIALIZER
,
2141 flow_mark_alloc(void)
2145 if (!flow_mark
.pool
) {
2146 /* Haven't initiated yet, do it here */
2147 flow_mark
.pool
= id_pool_create(0, MAX_FLOW_MARK
);
2150 if (id_pool_alloc_id(flow_mark
.pool
, &mark
)) {
2154 return INVALID_FLOW_MARK
;
2158 flow_mark_free(uint32_t mark
)
2160 id_pool_free_id(flow_mark
.pool
, mark
);
2163 /* associate megaflow with a mark, which is a 1:1 mapping */
2165 megaflow_to_mark_associate(const ovs_u128
*mega_ufid
, uint32_t mark
)
2167 size_t hash
= dp_netdev_flow_hash(mega_ufid
);
2168 struct megaflow_to_mark_data
*data
= xzalloc(sizeof(*data
));
2170 data
->mega_ufid
= *mega_ufid
;
2173 cmap_insert(&flow_mark
.megaflow_to_mark
,
2174 CONST_CAST(struct cmap_node
*, &data
->node
), hash
);
2177 /* disassociate meagaflow with a mark */
2179 megaflow_to_mark_disassociate(const ovs_u128
*mega_ufid
)
2181 size_t hash
= dp_netdev_flow_hash(mega_ufid
);
2182 struct megaflow_to_mark_data
*data
;
2184 CMAP_FOR_EACH_WITH_HASH (data
, node
, hash
, &flow_mark
.megaflow_to_mark
) {
2185 if (ovs_u128_equals(*mega_ufid
, data
->mega_ufid
)) {
2186 cmap_remove(&flow_mark
.megaflow_to_mark
,
2187 CONST_CAST(struct cmap_node
*, &data
->node
), hash
);
2188 ovsrcu_postpone(free
, data
);
2193 VLOG_WARN("Masked ufid "UUID_FMT
" is not associated with a mark?\n",
2194 UUID_ARGS((struct uuid
*)mega_ufid
));
2197 static inline uint32_t
2198 megaflow_to_mark_find(const ovs_u128
*mega_ufid
)
2200 size_t hash
= dp_netdev_flow_hash(mega_ufid
);
2201 struct megaflow_to_mark_data
*data
;
2203 CMAP_FOR_EACH_WITH_HASH (data
, node
, hash
, &flow_mark
.megaflow_to_mark
) {
2204 if (ovs_u128_equals(*mega_ufid
, data
->mega_ufid
)) {
2209 VLOG_DBG("Mark id for ufid "UUID_FMT
" was not found\n",
2210 UUID_ARGS((struct uuid
*)mega_ufid
));
2211 return INVALID_FLOW_MARK
;
2214 /* associate mark with a flow, which is 1:N mapping */
2216 mark_to_flow_associate(const uint32_t mark
, struct dp_netdev_flow
*flow
)
2218 dp_netdev_flow_ref(flow
);
2220 cmap_insert(&flow_mark
.mark_to_flow
,
2221 CONST_CAST(struct cmap_node
*, &flow
->mark_node
),
2225 VLOG_DBG("Associated dp_netdev flow %p with mark %u\n", flow
, mark
);
2229 flow_mark_has_no_ref(uint32_t mark
)
2231 struct dp_netdev_flow
*flow
;
2233 CMAP_FOR_EACH_WITH_HASH (flow
, mark_node
, hash_int(mark
, 0),
2234 &flow_mark
.mark_to_flow
) {
2235 if (flow
->mark
== mark
) {
2244 mark_to_flow_disassociate(struct dp_netdev_pmd_thread
*pmd
,
2245 struct dp_netdev_flow
*flow
)
2248 uint32_t mark
= flow
->mark
;
2249 struct cmap_node
*mark_node
= CONST_CAST(struct cmap_node
*,
2252 cmap_remove(&flow_mark
.mark_to_flow
, mark_node
, hash_int(mark
, 0));
2253 flow
->mark
= INVALID_FLOW_MARK
;
2256 * no flow is referencing the mark any more? If so, let's
2257 * remove the flow from hardware and free the mark.
2259 if (flow_mark_has_no_ref(mark
)) {
2260 struct dp_netdev_port
*port
;
2261 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
2263 ovs_mutex_lock(&pmd
->dp
->port_mutex
);
2264 port
= dp_netdev_lookup_port(pmd
->dp
, in_port
);
2266 ret
= netdev_flow_del(port
->netdev
, &flow
->mega_ufid
, NULL
);
2268 ovs_mutex_unlock(&pmd
->dp
->port_mutex
);
2270 flow_mark_free(mark
);
2271 VLOG_DBG("Freed flow mark %u\n", mark
);
2273 megaflow_to_mark_disassociate(&flow
->mega_ufid
);
2275 dp_netdev_flow_unref(flow
);
2281 flow_mark_flush(struct dp_netdev_pmd_thread
*pmd
)
2283 struct dp_netdev_flow
*flow
;
2285 CMAP_FOR_EACH (flow
, mark_node
, &flow_mark
.mark_to_flow
) {
2286 if (flow
->pmd_id
== pmd
->core_id
) {
2287 queue_netdev_flow_del(pmd
, flow
);
2292 static struct dp_netdev_flow
*
2293 mark_to_flow_find(const struct dp_netdev_pmd_thread
*pmd
,
2294 const uint32_t mark
)
2296 struct dp_netdev_flow
*flow
;
2298 CMAP_FOR_EACH_WITH_HASH (flow
, mark_node
, hash_int(mark
, 0),
2299 &flow_mark
.mark_to_flow
) {
2300 if (flow
->mark
== mark
&& flow
->pmd_id
== pmd
->core_id
&&
2301 flow
->dead
== false) {
2309 static struct dp_flow_offload_item
*
2310 dp_netdev_alloc_flow_offload(struct dp_netdev_pmd_thread
*pmd
,
2311 struct dp_netdev_flow
*flow
,
2314 struct dp_flow_offload_item
*offload
;
2316 offload
= xzalloc(sizeof(*offload
));
2318 offload
->flow
= flow
;
2321 dp_netdev_flow_ref(flow
);
2322 dp_netdev_pmd_try_ref(pmd
);
2328 dp_netdev_free_flow_offload(struct dp_flow_offload_item
*offload
)
2330 dp_netdev_pmd_unref(offload
->pmd
);
2331 dp_netdev_flow_unref(offload
->flow
);
2333 free(offload
->actions
);
2338 dp_netdev_append_flow_offload(struct dp_flow_offload_item
*offload
)
2340 ovs_mutex_lock(&dp_flow_offload
.mutex
);
2341 ovs_list_push_back(&dp_flow_offload
.list
, &offload
->node
);
2342 xpthread_cond_signal(&dp_flow_offload
.cond
);
2343 ovs_mutex_unlock(&dp_flow_offload
.mutex
);
2347 dp_netdev_flow_offload_del(struct dp_flow_offload_item
*offload
)
2349 return mark_to_flow_disassociate(offload
->pmd
, offload
->flow
);
2353 * There are two flow offload operations here: addition and modification.
2355 * For flow addition, this function does:
2356 * - allocate a new flow mark id
2357 * - perform hardware flow offload
2358 * - associate the flow mark with flow and mega flow
2360 * For flow modification, both flow mark and the associations are still
2361 * valid, thus only item 2 needed.
2364 dp_netdev_flow_offload_put(struct dp_flow_offload_item
*offload
)
2366 struct dp_netdev_port
*port
;
2367 struct dp_netdev_pmd_thread
*pmd
= offload
->pmd
;
2368 struct dp_netdev_flow
*flow
= offload
->flow
;
2369 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
2370 bool modification
= offload
->op
== DP_NETDEV_FLOW_OFFLOAD_OP_MOD
;
2371 struct offload_info info
;
2381 ovs_assert(mark
!= INVALID_FLOW_MARK
);
2384 * If a mega flow has already been offloaded (from other PMD
2385 * instances), do not offload it again.
2387 mark
= megaflow_to_mark_find(&flow
->mega_ufid
);
2388 if (mark
!= INVALID_FLOW_MARK
) {
2389 VLOG_DBG("Flow has already been offloaded with mark %u\n", mark
);
2390 if (flow
->mark
!= INVALID_FLOW_MARK
) {
2391 ovs_assert(flow
->mark
== mark
);
2393 mark_to_flow_associate(mark
, flow
);
2398 mark
= flow_mark_alloc();
2399 if (mark
== INVALID_FLOW_MARK
) {
2400 VLOG_ERR("Failed to allocate flow mark!\n");
2403 info
.flow_mark
= mark
;
2405 ovs_mutex_lock(&pmd
->dp
->port_mutex
);
2406 port
= dp_netdev_lookup_port(pmd
->dp
, in_port
);
2407 if (!port
|| netdev_vport_is_vport_class(port
->netdev
->netdev_class
)) {
2408 ovs_mutex_unlock(&pmd
->dp
->port_mutex
);
2411 ret
= netdev_flow_put(port
->netdev
, &offload
->match
,
2412 CONST_CAST(struct nlattr
*, offload
->actions
),
2413 offload
->actions_len
, &flow
->mega_ufid
, &info
,
2415 ovs_mutex_unlock(&pmd
->dp
->port_mutex
);
2421 if (!modification
) {
2422 megaflow_to_mark_associate(&flow
->mega_ufid
, mark
);
2423 mark_to_flow_associate(mark
, flow
);
2428 if (!modification
) {
2429 flow_mark_free(mark
);
2431 mark_to_flow_disassociate(pmd
, flow
);
2437 dp_netdev_flow_offload_main(void *data OVS_UNUSED
)
2439 struct dp_flow_offload_item
*offload
;
2440 struct ovs_list
*list
;
2445 ovs_mutex_lock(&dp_flow_offload
.mutex
);
2446 if (ovs_list_is_empty(&dp_flow_offload
.list
)) {
2447 ovsrcu_quiesce_start();
2448 ovs_mutex_cond_wait(&dp_flow_offload
.cond
,
2449 &dp_flow_offload
.mutex
);
2450 ovsrcu_quiesce_end();
2452 list
= ovs_list_pop_front(&dp_flow_offload
.list
);
2453 offload
= CONTAINER_OF(list
, struct dp_flow_offload_item
, node
);
2454 ovs_mutex_unlock(&dp_flow_offload
.mutex
);
2456 switch (offload
->op
) {
2457 case DP_NETDEV_FLOW_OFFLOAD_OP_ADD
:
2459 ret
= dp_netdev_flow_offload_put(offload
);
2461 case DP_NETDEV_FLOW_OFFLOAD_OP_MOD
:
2463 ret
= dp_netdev_flow_offload_put(offload
);
2465 case DP_NETDEV_FLOW_OFFLOAD_OP_DEL
:
2467 ret
= dp_netdev_flow_offload_del(offload
);
2473 VLOG_DBG("%s to %s netdev flow\n",
2474 ret
== 0 ? "succeed" : "failed", op
);
2475 dp_netdev_free_flow_offload(offload
);
2482 queue_netdev_flow_del(struct dp_netdev_pmd_thread
*pmd
,
2483 struct dp_netdev_flow
*flow
)
2485 struct dp_flow_offload_item
*offload
;
2487 if (ovsthread_once_start(&offload_thread_once
)) {
2488 xpthread_cond_init(&dp_flow_offload
.cond
, NULL
);
2489 ovs_thread_create("dp_netdev_flow_offload",
2490 dp_netdev_flow_offload_main
, NULL
);
2491 ovsthread_once_done(&offload_thread_once
);
2494 offload
= dp_netdev_alloc_flow_offload(pmd
, flow
,
2495 DP_NETDEV_FLOW_OFFLOAD_OP_DEL
);
2496 dp_netdev_append_flow_offload(offload
);
2500 queue_netdev_flow_put(struct dp_netdev_pmd_thread
*pmd
,
2501 struct dp_netdev_flow
*flow
, struct match
*match
,
2502 const struct nlattr
*actions
, size_t actions_len
)
2504 struct dp_flow_offload_item
*offload
;
2507 if (!netdev_is_flow_api_enabled()) {
2511 if (ovsthread_once_start(&offload_thread_once
)) {
2512 xpthread_cond_init(&dp_flow_offload
.cond
, NULL
);
2513 ovs_thread_create("dp_netdev_flow_offload",
2514 dp_netdev_flow_offload_main
, NULL
);
2515 ovsthread_once_done(&offload_thread_once
);
2518 if (flow
->mark
!= INVALID_FLOW_MARK
) {
2519 op
= DP_NETDEV_FLOW_OFFLOAD_OP_MOD
;
2521 op
= DP_NETDEV_FLOW_OFFLOAD_OP_ADD
;
2523 offload
= dp_netdev_alloc_flow_offload(pmd
, flow
, op
);
2524 offload
->match
= *match
;
2525 offload
->actions
= xmalloc(actions_len
);
2526 memcpy(offload
->actions
, actions
, actions_len
);
2527 offload
->actions_len
= actions_len
;
2529 dp_netdev_append_flow_offload(offload
);
2533 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread
*pmd
,
2534 struct dp_netdev_flow
*flow
)
2535 OVS_REQUIRES(pmd
->flow_mutex
)
2537 struct cmap_node
*node
= CONST_CAST(struct cmap_node
*, &flow
->node
);
2539 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
2541 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2542 ovs_assert(cls
!= NULL
);
2543 dpcls_remove(cls
, &flow
->cr
);
2544 cmap_remove(&pmd
->flow_table
, node
, dp_netdev_flow_hash(&flow
->ufid
));
2545 if (flow
->mark
!= INVALID_FLOW_MARK
) {
2546 queue_netdev_flow_del(pmd
, flow
);
2550 dp_netdev_flow_unref(flow
);
2554 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
)
2556 struct dp_netdev_flow
*netdev_flow
;
2558 ovs_mutex_lock(&pmd
->flow_mutex
);
2559 CMAP_FOR_EACH (netdev_flow
, node
, &pmd
->flow_table
) {
2560 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
2562 ovs_mutex_unlock(&pmd
->flow_mutex
);
2566 dpif_netdev_flow_flush(struct dpif
*dpif
)
2568 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2569 struct dp_netdev_pmd_thread
*pmd
;
2571 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2572 dp_netdev_pmd_flow_flush(pmd
);
2578 struct dp_netdev_port_state
{
2579 struct hmap_position position
;
2584 dpif_netdev_port_dump_start(const struct dpif
*dpif OVS_UNUSED
, void **statep
)
2586 *statep
= xzalloc(sizeof(struct dp_netdev_port_state
));
2591 dpif_netdev_port_dump_next(const struct dpif
*dpif
, void *state_
,
2592 struct dpif_port
*dpif_port
)
2594 struct dp_netdev_port_state
*state
= state_
;
2595 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2596 struct hmap_node
*node
;
2599 ovs_mutex_lock(&dp
->port_mutex
);
2600 node
= hmap_at_position(&dp
->ports
, &state
->position
);
2602 struct dp_netdev_port
*port
;
2604 port
= CONTAINER_OF(node
, struct dp_netdev_port
, node
);
2607 state
->name
= xstrdup(netdev_get_name(port
->netdev
));
2608 dpif_port
->name
= state
->name
;
2609 dpif_port
->type
= port
->type
;
2610 dpif_port
->port_no
= port
->port_no
;
2616 ovs_mutex_unlock(&dp
->port_mutex
);
2622 dpif_netdev_port_dump_done(const struct dpif
*dpif OVS_UNUSED
, void *state_
)
2624 struct dp_netdev_port_state
*state
= state_
;
2631 dpif_netdev_port_poll(const struct dpif
*dpif_
, char **devnamep OVS_UNUSED
)
2633 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
2634 uint64_t new_port_seq
;
2637 new_port_seq
= seq_read(dpif
->dp
->port_seq
);
2638 if (dpif
->last_port_seq
!= new_port_seq
) {
2639 dpif
->last_port_seq
= new_port_seq
;
2649 dpif_netdev_port_poll_wait(const struct dpif
*dpif_
)
2651 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
2653 seq_wait(dpif
->dp
->port_seq
, dpif
->last_port_seq
);
2656 static struct dp_netdev_flow
*
2657 dp_netdev_flow_cast(const struct dpcls_rule
*cr
)
2659 return cr
? CONTAINER_OF(cr
, struct dp_netdev_flow
, cr
) : NULL
;
2662 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*flow
)
2664 return ovs_refcount_try_ref_rcu(&flow
->ref_cnt
);
2667 /* netdev_flow_key utilities.
2669 * netdev_flow_key is basically a miniflow. We use these functions
2670 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
2671 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
2673 * - Since we are dealing exclusively with miniflows created by
2674 * miniflow_extract(), if the map is different the miniflow is different.
2675 * Therefore we can be faster by comparing the map and the miniflow in a
2677 * - These functions can be inlined by the compiler. */
2679 /* Given the number of bits set in miniflow's maps, returns the size of the
2680 * 'netdev_flow_key.mf' */
2681 static inline size_t
2682 netdev_flow_key_size(size_t flow_u64s
)
2684 return sizeof(struct miniflow
) + MINIFLOW_VALUES_SIZE(flow_u64s
);
2688 netdev_flow_key_equal(const struct netdev_flow_key
*a
,
2689 const struct netdev_flow_key
*b
)
2691 /* 'b->len' may be not set yet. */
2692 return a
->hash
== b
->hash
&& !memcmp(&a
->mf
, &b
->mf
, a
->len
);
2695 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
2696 * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been
2697 * generated by miniflow_extract. */
2699 netdev_flow_key_equal_mf(const struct netdev_flow_key
*key
,
2700 const struct miniflow
*mf
)
2702 return !memcmp(&key
->mf
, mf
, key
->len
);
2706 netdev_flow_key_clone(struct netdev_flow_key
*dst
,
2707 const struct netdev_flow_key
*src
)
2710 offsetof(struct netdev_flow_key
, mf
) + src
->len
);
2713 /* Initialize a netdev_flow_key 'mask' from 'match'. */
2715 netdev_flow_mask_init(struct netdev_flow_key
*mask
,
2716 const struct match
*match
)
2718 uint64_t *dst
= miniflow_values(&mask
->mf
);
2719 struct flowmap fmap
;
2723 /* Only check masks that make sense for the flow. */
2724 flow_wc_map(&match
->flow
, &fmap
);
2725 flowmap_init(&mask
->mf
.map
);
2727 FLOWMAP_FOR_EACH_INDEX(idx
, fmap
) {
2728 uint64_t mask_u64
= flow_u64_value(&match
->wc
.masks
, idx
);
2731 flowmap_set(&mask
->mf
.map
, idx
, 1);
2733 hash
= hash_add64(hash
, mask_u64
);
2739 FLOWMAP_FOR_EACH_MAP (map
, mask
->mf
.map
) {
2740 hash
= hash_add64(hash
, map
);
2743 size_t n
= dst
- miniflow_get_values(&mask
->mf
);
2745 mask
->hash
= hash_finish(hash
, n
* 8);
2746 mask
->len
= netdev_flow_key_size(n
);
2749 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
2751 netdev_flow_key_init_masked(struct netdev_flow_key
*dst
,
2752 const struct flow
*flow
,
2753 const struct netdev_flow_key
*mask
)
2755 uint64_t *dst_u64
= miniflow_values(&dst
->mf
);
2756 const uint64_t *mask_u64
= miniflow_get_values(&mask
->mf
);
2760 dst
->len
= mask
->len
;
2761 dst
->mf
= mask
->mf
; /* Copy maps. */
2763 FLOW_FOR_EACH_IN_MAPS(value
, flow
, mask
->mf
.map
) {
2764 *dst_u64
= value
& *mask_u64
++;
2765 hash
= hash_add64(hash
, *dst_u64
++);
2767 dst
->hash
= hash_finish(hash
,
2768 (dst_u64
- miniflow_get_values(&dst
->mf
)) * 8);
2771 /* Returns a hash value for the bits of 'key' where there are 1-bits in
2773 static inline uint32_t
2774 netdev_flow_key_hash_in_mask(const struct netdev_flow_key
*key
,
2775 const struct netdev_flow_key
*mask
)
2777 const uint64_t *p
= miniflow_get_values(&mask
->mf
);
2781 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, key
, mask
->mf
.map
) {
2782 hash
= hash_add64(hash
, value
& *p
++);
2785 return hash_finish(hash
, (p
- miniflow_get_values(&mask
->mf
)) * 8);
2789 emc_entry_alive(struct emc_entry
*ce
)
2791 return ce
->flow
&& !ce
->flow
->dead
;
2795 emc_clear_entry(struct emc_entry
*ce
)
2798 dp_netdev_flow_unref(ce
->flow
);
2804 emc_change_entry(struct emc_entry
*ce
, struct dp_netdev_flow
*flow
,
2805 const struct netdev_flow_key
*key
)
2807 if (ce
->flow
!= flow
) {
2809 dp_netdev_flow_unref(ce
->flow
);
2812 if (dp_netdev_flow_ref(flow
)) {
2819 netdev_flow_key_clone(&ce
->key
, key
);
2824 emc_insert(struct emc_cache
*cache
, const struct netdev_flow_key
*key
,
2825 struct dp_netdev_flow
*flow
)
2827 struct emc_entry
*to_be_replaced
= NULL
;
2828 struct emc_entry
*current_entry
;
2830 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2831 if (netdev_flow_key_equal(¤t_entry
->key
, key
)) {
2832 /* We found the entry with the 'mf' miniflow */
2833 emc_change_entry(current_entry
, flow
, NULL
);
2837 /* Replacement policy: put the flow in an empty (not alive) entry, or
2838 * in the first entry where it can be */
2840 || (emc_entry_alive(to_be_replaced
)
2841 && !emc_entry_alive(current_entry
))
2842 || current_entry
->key
.hash
< to_be_replaced
->key
.hash
) {
2843 to_be_replaced
= current_entry
;
2846 /* We didn't find the miniflow in the cache.
2847 * The 'to_be_replaced' entry is where the new flow will be stored */
2849 emc_change_entry(to_be_replaced
, flow
, key
);
2853 emc_probabilistic_insert(struct dp_netdev_pmd_thread
*pmd
,
2854 const struct netdev_flow_key
*key
,
2855 struct dp_netdev_flow
*flow
)
2857 /* Insert an entry into the EMC based on probability value 'min'. By
2858 * default the value is UINT32_MAX / 100 which yields an insertion
2859 * probability of 1/100 ie. 1% */
2861 uint32_t min
= pmd
->ctx
.emc_insert_min
;
2863 if (min
&& random_uint32() <= min
) {
2864 emc_insert(&(pmd
->flow_cache
).emc_cache
, key
, flow
);
2868 static inline struct dp_netdev_flow
*
2869 emc_lookup(struct emc_cache
*cache
, const struct netdev_flow_key
*key
)
2871 struct emc_entry
*current_entry
;
2873 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2874 if (current_entry
->key
.hash
== key
->hash
2875 && emc_entry_alive(current_entry
)
2876 && netdev_flow_key_equal_mf(¤t_entry
->key
, &key
->mf
)) {
2878 /* We found the entry with the 'key->mf' miniflow */
2879 return current_entry
->flow
;
2886 static inline const struct cmap_node
*
2887 smc_entry_get(struct dp_netdev_pmd_thread
*pmd
, const uint32_t hash
)
2889 struct smc_cache
*cache
= &(pmd
->flow_cache
).smc_cache
;
2890 struct smc_bucket
*bucket
= &cache
->buckets
[hash
& SMC_MASK
];
2891 uint16_t sig
= hash
>> 16;
2892 uint16_t index
= UINT16_MAX
;
2894 for (int i
= 0; i
< SMC_ENTRY_PER_BUCKET
; i
++) {
2895 if (bucket
->sig
[i
] == sig
) {
2896 index
= bucket
->flow_idx
[i
];
2900 if (index
!= UINT16_MAX
) {
2901 return cmap_find_by_index(&pmd
->flow_table
, index
);
2907 smc_clear_entry(struct smc_bucket
*b
, int idx
)
2909 b
->flow_idx
[idx
] = UINT16_MAX
;
2912 /* Insert the flow_table index into SMC. Insertion may fail when 1) SMC is
2913 * turned off, 2) the flow_table index is larger than uint16_t can handle.
2914 * If there is already an SMC entry having same signature, the index will be
2915 * updated. If there is no existing entry, but an empty entry is available,
2916 * the empty entry will be taken. If no empty entry or existing same signature,
2917 * a random entry from the hashed bucket will be picked. */
2919 smc_insert(struct dp_netdev_pmd_thread
*pmd
,
2920 const struct netdev_flow_key
*key
,
2923 struct smc_cache
*smc_cache
= &(pmd
->flow_cache
).smc_cache
;
2924 struct smc_bucket
*bucket
= &smc_cache
->buckets
[key
->hash
& SMC_MASK
];
2926 uint32_t cmap_index
;
2930 atomic_read_relaxed(&pmd
->dp
->smc_enable_db
, &smc_enable_db
);
2931 if (!smc_enable_db
) {
2935 cmap_index
= cmap_find_index(&pmd
->flow_table
, hash
);
2936 index
= (cmap_index
>= UINT16_MAX
) ? UINT16_MAX
: (uint16_t)cmap_index
;
2938 /* If the index is larger than SMC can handle (uint16_t), we don't
2940 if (index
== UINT16_MAX
) {
2944 /* If an entry with same signature already exists, update the index */
2945 uint16_t sig
= key
->hash
>> 16;
2946 for (i
= 0; i
< SMC_ENTRY_PER_BUCKET
; i
++) {
2947 if (bucket
->sig
[i
] == sig
) {
2948 bucket
->flow_idx
[i
] = index
;
2952 /* If there is an empty entry, occupy it. */
2953 for (i
= 0; i
< SMC_ENTRY_PER_BUCKET
; i
++) {
2954 if (bucket
->flow_idx
[i
] == UINT16_MAX
) {
2955 bucket
->sig
[i
] = sig
;
2956 bucket
->flow_idx
[i
] = index
;
2960 /* Otherwise, pick a random entry. */
2961 i
= random_uint32() % SMC_ENTRY_PER_BUCKET
;
2962 bucket
->sig
[i
] = sig
;
2963 bucket
->flow_idx
[i
] = index
;
2966 static struct dp_netdev_flow
*
2967 dp_netdev_pmd_lookup_flow(struct dp_netdev_pmd_thread
*pmd
,
2968 const struct netdev_flow_key
*key
,
2972 struct dpcls_rule
*rule
;
2973 odp_port_t in_port
= u32_to_odp(MINIFLOW_GET_U32(&key
->mf
,
2975 struct dp_netdev_flow
*netdev_flow
= NULL
;
2977 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2978 if (OVS_LIKELY(cls
)) {
2979 dpcls_lookup(cls
, &key
, &rule
, 1, lookup_num_p
);
2980 netdev_flow
= dp_netdev_flow_cast(rule
);
2985 static struct dp_netdev_flow
*
2986 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread
*pmd
,
2987 const ovs_u128
*ufidp
, const struct nlattr
*key
,
2990 struct dp_netdev_flow
*netdev_flow
;
2994 /* If a UFID is not provided, determine one based on the key. */
2995 if (!ufidp
&& key
&& key_len
2996 && !dpif_netdev_flow_from_nlattrs(key
, key_len
, &flow
, false)) {
2997 dpif_flow_hash(pmd
->dp
->dpif
, &flow
, sizeof flow
, &ufid
);
3002 CMAP_FOR_EACH_WITH_HASH (netdev_flow
, node
, dp_netdev_flow_hash(ufidp
),
3004 if (ovs_u128_equals(netdev_flow
->ufid
, *ufidp
)) {
3014 get_dpif_flow_stats(const struct dp_netdev_flow
*netdev_flow_
,
3015 struct dpif_flow_stats
*stats
)
3017 struct dp_netdev_flow
*netdev_flow
;
3018 unsigned long long n
;
3022 netdev_flow
= CONST_CAST(struct dp_netdev_flow
*, netdev_flow_
);
3024 atomic_read_relaxed(&netdev_flow
->stats
.packet_count
, &n
);
3025 stats
->n_packets
= n
;
3026 atomic_read_relaxed(&netdev_flow
->stats
.byte_count
, &n
);
3028 atomic_read_relaxed(&netdev_flow
->stats
.used
, &used
);
3030 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
3031 stats
->tcp_flags
= flags
;
3034 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
3035 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
3036 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
3039 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow
*netdev_flow
,
3040 struct ofpbuf
*key_buf
, struct ofpbuf
*mask_buf
,
3041 struct dpif_flow
*flow
, bool terse
)
3044 memset(flow
, 0, sizeof *flow
);
3046 struct flow_wildcards wc
;
3047 struct dp_netdev_actions
*actions
;
3049 struct odp_flow_key_parms odp_parms
= {
3050 .flow
= &netdev_flow
->flow
,
3052 .support
= dp_netdev_support
,
3055 miniflow_expand(&netdev_flow
->cr
.mask
->mf
, &wc
.masks
);
3056 /* in_port is exact matched, but we have left it out from the mask for
3057 * optimnization reasons. Add in_port back to the mask. */
3058 wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
3061 offset
= key_buf
->size
;
3062 flow
->key
= ofpbuf_tail(key_buf
);
3063 odp_flow_key_from_flow(&odp_parms
, key_buf
);
3064 flow
->key_len
= key_buf
->size
- offset
;
3067 offset
= mask_buf
->size
;
3068 flow
->mask
= ofpbuf_tail(mask_buf
);
3069 odp_parms
.key_buf
= key_buf
;
3070 odp_flow_key_from_mask(&odp_parms
, mask_buf
);
3071 flow
->mask_len
= mask_buf
->size
- offset
;
3074 actions
= dp_netdev_flow_get_actions(netdev_flow
);
3075 flow
->actions
= actions
->actions
;
3076 flow
->actions_len
= actions
->size
;
3079 flow
->ufid
= netdev_flow
->ufid
;
3080 flow
->ufid_present
= true;
3081 flow
->pmd_id
= netdev_flow
->pmd_id
;
3082 get_dpif_flow_stats(netdev_flow
, &flow
->stats
);
3084 flow
->attrs
.offloaded
= false;
3085 flow
->attrs
.dp_layer
= "ovs";
3089 dpif_netdev_mask_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
3090 const struct nlattr
*mask_key
,
3091 uint32_t mask_key_len
, const struct flow
*flow
,
3092 struct flow_wildcards
*wc
, bool probe
)
3094 enum odp_key_fitness fitness
;
3096 fitness
= odp_flow_key_to_mask(mask_key
, mask_key_len
, wc
, flow
, NULL
);
3099 /* This should not happen: it indicates that
3100 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
3101 * disagree on the acceptable form of a mask. Log the problem
3102 * as an error, with enough details to enable debugging. */
3103 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
3105 if (!VLOG_DROP_ERR(&rl
)) {
3109 odp_flow_format(key
, key_len
, mask_key
, mask_key_len
, NULL
, &s
,
3111 VLOG_ERR("internal error parsing flow mask %s (%s)",
3112 ds_cstr(&s
), odp_key_fitness_to_string(fitness
));
3124 dpif_netdev_flow_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
3125 struct flow
*flow
, bool probe
)
3127 if (odp_flow_key_to_flow(key
, key_len
, flow
, NULL
)) {
3129 /* This should not happen: it indicates that
3130 * odp_flow_key_from_flow() and odp_flow_key_to_flow() disagree on
3131 * the acceptable form of a flow. Log the problem as an error,
3132 * with enough details to enable debugging. */
3133 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
3135 if (!VLOG_DROP_ERR(&rl
)) {
3139 odp_flow_format(key
, key_len
, NULL
, 0, NULL
, &s
, true);
3140 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s
));
3148 if (flow
->ct_state
& DP_NETDEV_CS_UNSUPPORTED_MASK
) {
3156 dpif_netdev_flow_get(const struct dpif
*dpif
, const struct dpif_flow_get
*get
)
3158 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3159 struct dp_netdev_flow
*netdev_flow
;
3160 struct dp_netdev_pmd_thread
*pmd
;
3161 struct hmapx to_find
= HMAPX_INITIALIZER(&to_find
);
3162 struct hmapx_node
*node
;
3165 if (get
->pmd_id
== PMD_ID_NULL
) {
3166 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3167 if (dp_netdev_pmd_try_ref(pmd
) && !hmapx_add(&to_find
, pmd
)) {
3168 dp_netdev_pmd_unref(pmd
);
3172 pmd
= dp_netdev_get_pmd(dp
, get
->pmd_id
);
3176 hmapx_add(&to_find
, pmd
);
3179 if (!hmapx_count(&to_find
)) {
3183 HMAPX_FOR_EACH (node
, &to_find
) {
3184 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3185 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, get
->ufid
, get
->key
,
3188 dp_netdev_flow_to_dpif_flow(netdev_flow
, get
->buffer
, get
->buffer
,
3197 HMAPX_FOR_EACH (node
, &to_find
) {
3198 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3199 dp_netdev_pmd_unref(pmd
);
3202 hmapx_destroy(&to_find
);
3207 dp_netdev_get_mega_ufid(const struct match
*match
, ovs_u128
*mega_ufid
)
3209 struct flow masked_flow
;
3212 for (i
= 0; i
< sizeof(struct flow
); i
++) {
3213 ((uint8_t *)&masked_flow
)[i
] = ((uint8_t *)&match
->flow
)[i
] &
3214 ((uint8_t *)&match
->wc
)[i
];
3216 dpif_flow_hash(NULL
, &masked_flow
, sizeof(struct flow
), mega_ufid
);
3219 static struct dp_netdev_flow
*
3220 dp_netdev_flow_add(struct dp_netdev_pmd_thread
*pmd
,
3221 struct match
*match
, const ovs_u128
*ufid
,
3222 const struct nlattr
*actions
, size_t actions_len
)
3223 OVS_REQUIRES(pmd
->flow_mutex
)
3225 struct dp_netdev_flow
*flow
;
3226 struct netdev_flow_key mask
;
3229 /* Make sure in_port is exact matched before we read it. */
3230 ovs_assert(match
->wc
.masks
.in_port
.odp_port
== ODPP_NONE
);
3231 odp_port_t in_port
= match
->flow
.in_port
.odp_port
;
3233 /* As we select the dpcls based on the port number, each netdev flow
3234 * belonging to the same dpcls will have the same odp_port value.
3235 * For performance reasons we wildcard odp_port here in the mask. In the
3236 * typical case dp_hash is also wildcarded, and the resulting 8-byte
3237 * chunk {dp_hash, in_port} will be ignored by netdev_flow_mask_init() and
3238 * will not be part of the subtable mask.
3239 * This will speed up the hash computation during dpcls_lookup() because
3240 * there is one less call to hash_add64() in this case. */
3241 match
->wc
.masks
.in_port
.odp_port
= 0;
3242 netdev_flow_mask_init(&mask
, match
);
3243 match
->wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
3245 /* Make sure wc does not have metadata. */
3246 ovs_assert(!FLOWMAP_HAS_FIELD(&mask
.mf
.map
, metadata
)
3247 && !FLOWMAP_HAS_FIELD(&mask
.mf
.map
, regs
));
3249 /* Do not allocate extra space. */
3250 flow
= xmalloc(sizeof *flow
- sizeof flow
->cr
.flow
.mf
+ mask
.len
);
3251 memset(&flow
->stats
, 0, sizeof flow
->stats
);
3254 flow
->mark
= INVALID_FLOW_MARK
;
3255 *CONST_CAST(unsigned *, &flow
->pmd_id
) = pmd
->core_id
;
3256 *CONST_CAST(struct flow
*, &flow
->flow
) = match
->flow
;
3257 *CONST_CAST(ovs_u128
*, &flow
->ufid
) = *ufid
;
3258 ovs_refcount_init(&flow
->ref_cnt
);
3259 ovsrcu_set(&flow
->actions
, dp_netdev_actions_create(actions
, actions_len
));
3261 dp_netdev_get_mega_ufid(match
, CONST_CAST(ovs_u128
*, &flow
->mega_ufid
));
3262 netdev_flow_key_init_masked(&flow
->cr
.flow
, &match
->flow
, &mask
);
3264 /* Select dpcls for in_port. Relies on in_port to be exact match. */
3265 cls
= dp_netdev_pmd_find_dpcls(pmd
, in_port
);
3266 dpcls_insert(cls
, &flow
->cr
, &mask
);
3268 cmap_insert(&pmd
->flow_table
, CONST_CAST(struct cmap_node
*, &flow
->node
),
3269 dp_netdev_flow_hash(&flow
->ufid
));
3271 queue_netdev_flow_put(pmd
, flow
, match
, actions
, actions_len
);
3273 if (OVS_UNLIKELY(!VLOG_DROP_DBG((&upcall_rl
)))) {
3274 struct ds ds
= DS_EMPTY_INITIALIZER
;
3275 struct ofpbuf key_buf
, mask_buf
;
3276 struct odp_flow_key_parms odp_parms
= {
3277 .flow
= &match
->flow
,
3278 .mask
= &match
->wc
.masks
,
3279 .support
= dp_netdev_support
,
3282 ofpbuf_init(&key_buf
, 0);
3283 ofpbuf_init(&mask_buf
, 0);
3285 odp_flow_key_from_flow(&odp_parms
, &key_buf
);
3286 odp_parms
.key_buf
= &key_buf
;
3287 odp_flow_key_from_mask(&odp_parms
, &mask_buf
);
3289 ds_put_cstr(&ds
, "flow_add: ");
3290 odp_format_ufid(ufid
, &ds
);
3291 ds_put_cstr(&ds
, " ");
3292 odp_flow_format(key_buf
.data
, key_buf
.size
,
3293 mask_buf
.data
, mask_buf
.size
,
3295 ds_put_cstr(&ds
, ", actions:");
3296 format_odp_actions(&ds
, actions
, actions_len
, NULL
);
3298 VLOG_DBG("%s", ds_cstr(&ds
));
3300 ofpbuf_uninit(&key_buf
);
3301 ofpbuf_uninit(&mask_buf
);
3303 /* Add a printout of the actual match installed. */
3306 ds_put_cstr(&ds
, "flow match: ");
3307 miniflow_expand(&flow
->cr
.flow
.mf
, &m
.flow
);
3308 miniflow_expand(&flow
->cr
.mask
->mf
, &m
.wc
.masks
);
3309 memset(&m
.tun_md
, 0, sizeof m
.tun_md
);
3310 match_format(&m
, NULL
, &ds
, OFP_DEFAULT_PRIORITY
);
3312 VLOG_DBG("%s", ds_cstr(&ds
));
3321 flow_put_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
3322 struct netdev_flow_key
*key
,
3323 struct match
*match
,
3325 const struct dpif_flow_put
*put
,
3326 struct dpif_flow_stats
*stats
)
3328 struct dp_netdev_flow
*netdev_flow
;
3332 memset(stats
, 0, sizeof *stats
);
3335 ovs_mutex_lock(&pmd
->flow_mutex
);
3336 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
3338 if (put
->flags
& DPIF_FP_CREATE
) {
3339 if (cmap_count(&pmd
->flow_table
) < MAX_FLOWS
) {
3340 dp_netdev_flow_add(pmd
, match
, ufid
, put
->actions
,
3350 if (put
->flags
& DPIF_FP_MODIFY
) {
3351 struct dp_netdev_actions
*new_actions
;
3352 struct dp_netdev_actions
*old_actions
;
3354 new_actions
= dp_netdev_actions_create(put
->actions
,
3357 old_actions
= dp_netdev_flow_get_actions(netdev_flow
);
3358 ovsrcu_set(&netdev_flow
->actions
, new_actions
);
3360 queue_netdev_flow_put(pmd
, netdev_flow
, match
,
3361 put
->actions
, put
->actions_len
);
3364 get_dpif_flow_stats(netdev_flow
, stats
);
3366 if (put
->flags
& DPIF_FP_ZERO_STATS
) {
3367 /* XXX: The userspace datapath uses thread local statistics
3368 * (for flows), which should be updated only by the owning
3369 * thread. Since we cannot write on stats memory here,
3370 * we choose not to support this flag. Please note:
3371 * - This feature is currently used only by dpctl commands with
3373 * - Should the need arise, this operation can be implemented
3374 * by keeping a base value (to be update here) for each
3375 * counter, and subtracting it before outputting the stats */
3379 ovsrcu_postpone(dp_netdev_actions_free
, old_actions
);
3380 } else if (put
->flags
& DPIF_FP_CREATE
) {
3383 /* Overlapping flow. */
3387 ovs_mutex_unlock(&pmd
->flow_mutex
);
3392 dpif_netdev_flow_put(struct dpif
*dpif
, const struct dpif_flow_put
*put
)
3394 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3395 struct netdev_flow_key key
, mask
;
3396 struct dp_netdev_pmd_thread
*pmd
;
3400 bool probe
= put
->flags
& DPIF_FP_PROBE
;
3403 memset(put
->stats
, 0, sizeof *put
->stats
);
3405 error
= dpif_netdev_flow_from_nlattrs(put
->key
, put
->key_len
, &match
.flow
,
3410 error
= dpif_netdev_mask_from_nlattrs(put
->key
, put
->key_len
,
3411 put
->mask
, put
->mask_len
,
3412 &match
.flow
, &match
.wc
, probe
);
3420 dpif_flow_hash(dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
3423 /* The Netlink encoding of datapath flow keys cannot express
3424 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3425 * tag is interpreted as exact match on the fact that there is no
3426 * VLAN. Unless we refactor a lot of code that translates between
3427 * Netlink and struct flow representations, we have to do the same
3428 * here. This must be in sync with 'match' in handle_packet_upcall(). */
3429 if (!match
.wc
.masks
.vlans
[0].tci
) {
3430 match
.wc
.masks
.vlans
[0].tci
= htons(0xffff);
3433 /* Must produce a netdev_flow_key for lookup.
3434 * Use the same method as employed to create the key when adding
3435 * the flow to the dplcs to make sure they match. */
3436 netdev_flow_mask_init(&mask
, &match
);
3437 netdev_flow_key_init_masked(&key
, &match
.flow
, &mask
);
3439 if (put
->pmd_id
== PMD_ID_NULL
) {
3440 if (cmap_count(&dp
->poll_threads
) == 0) {
3443 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3444 struct dpif_flow_stats pmd_stats
;
3447 pmd_error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
,
3451 } else if (put
->stats
) {
3452 put
->stats
->n_packets
+= pmd_stats
.n_packets
;
3453 put
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
3454 put
->stats
->used
= MAX(put
->stats
->used
, pmd_stats
.used
);
3455 put
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
3459 pmd
= dp_netdev_get_pmd(dp
, put
->pmd_id
);
3463 error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
, put
->stats
);
3464 dp_netdev_pmd_unref(pmd
);
3471 flow_del_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
3472 struct dpif_flow_stats
*stats
,
3473 const struct dpif_flow_del
*del
)
3475 struct dp_netdev_flow
*netdev_flow
;
3478 ovs_mutex_lock(&pmd
->flow_mutex
);
3479 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, del
->ufid
, del
->key
,
3483 get_dpif_flow_stats(netdev_flow
, stats
);
3485 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
3489 ovs_mutex_unlock(&pmd
->flow_mutex
);
3495 dpif_netdev_flow_del(struct dpif
*dpif
, const struct dpif_flow_del
*del
)
3497 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3498 struct dp_netdev_pmd_thread
*pmd
;
3502 memset(del
->stats
, 0, sizeof *del
->stats
);
3505 if (del
->pmd_id
== PMD_ID_NULL
) {
3506 if (cmap_count(&dp
->poll_threads
) == 0) {
3509 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3510 struct dpif_flow_stats pmd_stats
;
3513 pmd_error
= flow_del_on_pmd(pmd
, &pmd_stats
, del
);
3516 } else if (del
->stats
) {
3517 del
->stats
->n_packets
+= pmd_stats
.n_packets
;
3518 del
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
3519 del
->stats
->used
= MAX(del
->stats
->used
, pmd_stats
.used
);
3520 del
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
3524 pmd
= dp_netdev_get_pmd(dp
, del
->pmd_id
);
3528 error
= flow_del_on_pmd(pmd
, del
->stats
, del
);
3529 dp_netdev_pmd_unref(pmd
);
3536 struct dpif_netdev_flow_dump
{
3537 struct dpif_flow_dump up
;
3538 struct cmap_position poll_thread_pos
;
3539 struct cmap_position flow_pos
;
3540 struct dp_netdev_pmd_thread
*cur_pmd
;
3542 struct ovs_mutex mutex
;
3545 static struct dpif_netdev_flow_dump
*
3546 dpif_netdev_flow_dump_cast(struct dpif_flow_dump
*dump
)
3548 return CONTAINER_OF(dump
, struct dpif_netdev_flow_dump
, up
);
3551 static struct dpif_flow_dump
*
3552 dpif_netdev_flow_dump_create(const struct dpif
*dpif_
, bool terse
,
3553 struct dpif_flow_dump_types
*types OVS_UNUSED
)
3555 struct dpif_netdev_flow_dump
*dump
;
3557 dump
= xzalloc(sizeof *dump
);
3558 dpif_flow_dump_init(&dump
->up
, dpif_
);
3559 dump
->up
.terse
= terse
;
3560 ovs_mutex_init(&dump
->mutex
);
3566 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump
*dump_
)
3568 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
3570 ovs_mutex_destroy(&dump
->mutex
);
3575 struct dpif_netdev_flow_dump_thread
{
3576 struct dpif_flow_dump_thread up
;
3577 struct dpif_netdev_flow_dump
*dump
;
3578 struct odputil_keybuf keybuf
[FLOW_DUMP_MAX_BATCH
];
3579 struct odputil_keybuf maskbuf
[FLOW_DUMP_MAX_BATCH
];
3582 static struct dpif_netdev_flow_dump_thread
*
3583 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread
*thread
)
3585 return CONTAINER_OF(thread
, struct dpif_netdev_flow_dump_thread
, up
);
3588 static struct dpif_flow_dump_thread
*
3589 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump
*dump_
)
3591 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
3592 struct dpif_netdev_flow_dump_thread
*thread
;
3594 thread
= xmalloc(sizeof *thread
);
3595 dpif_flow_dump_thread_init(&thread
->up
, &dump
->up
);
3596 thread
->dump
= dump
;
3601 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread
*thread_
)
3603 struct dpif_netdev_flow_dump_thread
*thread
3604 = dpif_netdev_flow_dump_thread_cast(thread_
);
3610 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread
*thread_
,
3611 struct dpif_flow
*flows
, int max_flows
)
3613 struct dpif_netdev_flow_dump_thread
*thread
3614 = dpif_netdev_flow_dump_thread_cast(thread_
);
3615 struct dpif_netdev_flow_dump
*dump
= thread
->dump
;
3616 struct dp_netdev_flow
*netdev_flows
[FLOW_DUMP_MAX_BATCH
];
3620 ovs_mutex_lock(&dump
->mutex
);
3621 if (!dump
->status
) {
3622 struct dpif_netdev
*dpif
= dpif_netdev_cast(thread
->up
.dpif
);
3623 struct dp_netdev
*dp
= get_dp_netdev(&dpif
->dpif
);
3624 struct dp_netdev_pmd_thread
*pmd
= dump
->cur_pmd
;
3625 int flow_limit
= MIN(max_flows
, FLOW_DUMP_MAX_BATCH
);
3627 /* First call to dump_next(), extracts the first pmd thread.
3628 * If there is no pmd thread, returns immediately. */
3630 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
3632 ovs_mutex_unlock(&dump
->mutex
);
3639 for (n_flows
= 0; n_flows
< flow_limit
; n_flows
++) {
3640 struct cmap_node
*node
;
3642 node
= cmap_next_position(&pmd
->flow_table
, &dump
->flow_pos
);
3646 netdev_flows
[n_flows
] = CONTAINER_OF(node
,
3647 struct dp_netdev_flow
,
3650 /* When finishing dumping the current pmd thread, moves to
3652 if (n_flows
< flow_limit
) {
3653 memset(&dump
->flow_pos
, 0, sizeof dump
->flow_pos
);
3654 dp_netdev_pmd_unref(pmd
);
3655 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
3661 /* Keeps the reference to next caller. */
3662 dump
->cur_pmd
= pmd
;
3664 /* If the current dump is empty, do not exit the loop, since the
3665 * remaining pmds could have flows to be dumped. Just dumps again
3666 * on the new 'pmd'. */
3669 ovs_mutex_unlock(&dump
->mutex
);
3671 for (i
= 0; i
< n_flows
; i
++) {
3672 struct odputil_keybuf
*maskbuf
= &thread
->maskbuf
[i
];
3673 struct odputil_keybuf
*keybuf
= &thread
->keybuf
[i
];
3674 struct dp_netdev_flow
*netdev_flow
= netdev_flows
[i
];
3675 struct dpif_flow
*f
= &flows
[i
];
3676 struct ofpbuf key
, mask
;
3678 ofpbuf_use_stack(&key
, keybuf
, sizeof *keybuf
);
3679 ofpbuf_use_stack(&mask
, maskbuf
, sizeof *maskbuf
);
3680 dp_netdev_flow_to_dpif_flow(netdev_flow
, &key
, &mask
, f
,
3688 dpif_netdev_execute(struct dpif
*dpif
, struct dpif_execute
*execute
)
3689 OVS_NO_THREAD_SAFETY_ANALYSIS
3691 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3692 struct dp_netdev_pmd_thread
*pmd
;
3693 struct dp_packet_batch pp
;
3695 if (dp_packet_size(execute
->packet
) < ETH_HEADER_LEN
||
3696 dp_packet_size(execute
->packet
) > UINT16_MAX
) {
3700 /* Tries finding the 'pmd'. If NULL is returned, that means
3701 * the current thread is a non-pmd thread and should use
3702 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
3703 pmd
= ovsthread_getspecific(dp
->per_pmd_key
);
3705 pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
3711 if (execute
->probe
) {
3712 /* If this is part of a probe, Drop the packet, since executing
3713 * the action may actually cause spurious packets be sent into
3715 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3716 dp_netdev_pmd_unref(pmd
);
3721 /* If the current thread is non-pmd thread, acquires
3722 * the 'non_pmd_mutex'. */
3723 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3724 ovs_mutex_lock(&dp
->non_pmd_mutex
);
3727 /* Update current time in PMD context. We don't care about EMC insertion
3728 * probability, because we are on a slow path. */
3729 pmd_thread_ctx_time_update(pmd
);
3731 /* The action processing expects the RSS hash to be valid, because
3732 * it's always initialized at the beginning of datapath processing.
3733 * In this case, though, 'execute->packet' may not have gone through
3734 * the datapath at all, it may have been generated by the upper layer
3735 * (OpenFlow packet-out, BFD frame, ...). */
3736 if (!dp_packet_rss_valid(execute
->packet
)) {
3737 dp_packet_set_rss_hash(execute
->packet
,
3738 flow_hash_5tuple(execute
->flow
, 0));
3741 dp_packet_batch_init_packet(&pp
, execute
->packet
);
3742 pp
.do_not_steal
= true;
3743 dp_netdev_execute_actions(pmd
, &pp
, false, execute
->flow
,
3744 execute
->actions
, execute
->actions_len
);
3745 dp_netdev_pmd_flush_output_packets(pmd
, true);
3747 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3748 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
3749 dp_netdev_pmd_unref(pmd
);
3756 dpif_netdev_operate(struct dpif
*dpif
, struct dpif_op
**ops
, size_t n_ops
,
3757 enum dpif_offload_type offload_type OVS_UNUSED
)
3761 for (i
= 0; i
< n_ops
; i
++) {
3762 struct dpif_op
*op
= ops
[i
];
3765 case DPIF_OP_FLOW_PUT
:
3766 op
->error
= dpif_netdev_flow_put(dpif
, &op
->flow_put
);
3769 case DPIF_OP_FLOW_DEL
:
3770 op
->error
= dpif_netdev_flow_del(dpif
, &op
->flow_del
);
3773 case DPIF_OP_EXECUTE
:
3774 op
->error
= dpif_netdev_execute(dpif
, &op
->execute
);
3777 case DPIF_OP_FLOW_GET
:
3778 op
->error
= dpif_netdev_flow_get(dpif
, &op
->flow_get
);
3784 /* Enable or Disable PMD auto load balancing. */
3786 set_pmd_auto_lb(struct dp_netdev
*dp
)
3788 unsigned int cnt
= 0;
3789 struct dp_netdev_pmd_thread
*pmd
;
3790 struct pmd_auto_lb
*pmd_alb
= &dp
->pmd_alb
;
3792 bool enable_alb
= false;
3793 bool multi_rxq
= false;
3794 bool pmd_rxq_assign_cyc
= dp
->pmd_rxq_assign_cyc
;
3796 /* Ensure that there is at least 2 non-isolated PMDs and
3797 * one of them is polling more than one rxq. */
3798 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3799 if (pmd
->core_id
== NON_PMD_CORE_ID
|| pmd
->isolated
) {
3803 if (hmap_count(&pmd
->poll_list
) > 1) {
3806 if (cnt
&& multi_rxq
) {
3813 /* Enable auto LB if it is requested and cycle based assignment is true. */
3814 enable_alb
= enable_alb
&& pmd_rxq_assign_cyc
&&
3815 pmd_alb
->auto_lb_requested
;
3817 if (pmd_alb
->is_enabled
!= enable_alb
) {
3818 pmd_alb
->is_enabled
= enable_alb
;
3819 if (pmd_alb
->is_enabled
) {
3820 VLOG_INFO("PMD auto load balance is enabled "
3821 "(with rebalance interval:%"PRIu64
" msec)",
3822 pmd_alb
->rebalance_intvl
);
3824 pmd_alb
->rebalance_poll_timer
= 0;
3825 VLOG_INFO("PMD auto load balance is disabled");
3831 /* Applies datapath configuration from the database. Some of the changes are
3832 * actually applied in dpif_netdev_run(). */
3834 dpif_netdev_set_config(struct dpif
*dpif
, const struct smap
*other_config
)
3836 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3837 const char *cmask
= smap_get(other_config
, "pmd-cpu-mask");
3838 const char *pmd_rxq_assign
= smap_get_def(other_config
, "pmd-rxq-assign",
3840 unsigned long long insert_prob
=
3841 smap_get_ullong(other_config
, "emc-insert-inv-prob",
3842 DEFAULT_EM_FLOW_INSERT_INV_PROB
);
3843 uint32_t insert_min
, cur_min
;
3844 uint32_t tx_flush_interval
, cur_tx_flush_interval
;
3845 uint64_t rebalance_intvl
;
3847 tx_flush_interval
= smap_get_int(other_config
, "tx-flush-interval",
3848 DEFAULT_TX_FLUSH_INTERVAL
);
3849 atomic_read_relaxed(&dp
->tx_flush_interval
, &cur_tx_flush_interval
);
3850 if (tx_flush_interval
!= cur_tx_flush_interval
) {
3851 atomic_store_relaxed(&dp
->tx_flush_interval
, tx_flush_interval
);
3852 VLOG_INFO("Flushing interval for tx queues set to %"PRIu32
" us",
3856 if (!nullable_string_is_equal(dp
->pmd_cmask
, cmask
)) {
3857 free(dp
->pmd_cmask
);
3858 dp
->pmd_cmask
= nullable_xstrdup(cmask
);
3859 dp_netdev_request_reconfigure(dp
);
3862 atomic_read_relaxed(&dp
->emc_insert_min
, &cur_min
);
3863 if (insert_prob
<= UINT32_MAX
) {
3864 insert_min
= insert_prob
== 0 ? 0 : UINT32_MAX
/ insert_prob
;
3866 insert_min
= DEFAULT_EM_FLOW_INSERT_MIN
;
3867 insert_prob
= DEFAULT_EM_FLOW_INSERT_INV_PROB
;
3870 if (insert_min
!= cur_min
) {
3871 atomic_store_relaxed(&dp
->emc_insert_min
, insert_min
);
3872 if (insert_min
== 0) {
3873 VLOG_INFO("EMC insertion probability changed to zero");
3875 VLOG_INFO("EMC insertion probability changed to 1/%llu (~%.2f%%)",
3876 insert_prob
, (100 / (float)insert_prob
));
3880 bool perf_enabled
= smap_get_bool(other_config
, "pmd-perf-metrics", false);
3881 bool cur_perf_enabled
;
3882 atomic_read_relaxed(&dp
->pmd_perf_metrics
, &cur_perf_enabled
);
3883 if (perf_enabled
!= cur_perf_enabled
) {
3884 atomic_store_relaxed(&dp
->pmd_perf_metrics
, perf_enabled
);
3886 VLOG_INFO("PMD performance metrics collection enabled");
3888 VLOG_INFO("PMD performance metrics collection disabled");
3892 bool smc_enable
= smap_get_bool(other_config
, "smc-enable", false);
3894 atomic_read_relaxed(&dp
->smc_enable_db
, &cur_smc
);
3895 if (smc_enable
!= cur_smc
) {
3896 atomic_store_relaxed(&dp
->smc_enable_db
, smc_enable
);
3898 VLOG_INFO("SMC cache is enabled");
3900 VLOG_INFO("SMC cache is disabled");
3904 bool pmd_rxq_assign_cyc
= !strcmp(pmd_rxq_assign
, "cycles");
3905 if (!pmd_rxq_assign_cyc
&& strcmp(pmd_rxq_assign
, "roundrobin")) {
3906 VLOG_WARN("Unsupported Rxq to PMD assignment mode in pmd-rxq-assign. "
3907 "Defaulting to 'cycles'.");
3908 pmd_rxq_assign_cyc
= true;
3909 pmd_rxq_assign
= "cycles";
3911 if (dp
->pmd_rxq_assign_cyc
!= pmd_rxq_assign_cyc
) {
3912 dp
->pmd_rxq_assign_cyc
= pmd_rxq_assign_cyc
;
3913 VLOG_INFO("Rxq to PMD assignment mode changed to: \'%s\'.",
3915 dp_netdev_request_reconfigure(dp
);
3918 struct pmd_auto_lb
*pmd_alb
= &dp
->pmd_alb
;
3919 pmd_alb
->auto_lb_requested
= smap_get_bool(other_config
, "pmd-auto-lb",
3922 rebalance_intvl
= smap_get_int(other_config
, "pmd-auto-lb-rebal-interval",
3923 ALB_PMD_REBALANCE_POLL_INTERVAL
);
3925 /* Input is in min, convert it to msec. */
3927 rebalance_intvl
? rebalance_intvl
* MIN_TO_MSEC
: MIN_TO_MSEC
;
3929 if (pmd_alb
->rebalance_intvl
!= rebalance_intvl
) {
3930 pmd_alb
->rebalance_intvl
= rebalance_intvl
;
3933 set_pmd_auto_lb(dp
);
3937 /* Parses affinity list and returns result in 'core_ids'. */
3939 parse_affinity_list(const char *affinity_list
, unsigned *core_ids
, int n_rxq
)
3942 char *list
, *copy
, *key
, *value
;
3945 for (i
= 0; i
< n_rxq
; i
++) {
3946 core_ids
[i
] = OVS_CORE_UNSPEC
;
3949 if (!affinity_list
) {
3953 list
= copy
= xstrdup(affinity_list
);
3955 while (ofputil_parse_key_value(&list
, &key
, &value
)) {
3956 int rxq_id
, core_id
;
3958 if (!str_to_int(key
, 0, &rxq_id
) || rxq_id
< 0
3959 || !str_to_int(value
, 0, &core_id
) || core_id
< 0) {
3964 if (rxq_id
< n_rxq
) {
3965 core_ids
[rxq_id
] = core_id
;
3973 /* Parses 'affinity_list' and applies configuration if it is valid. */
3975 dpif_netdev_port_set_rxq_affinity(struct dp_netdev_port
*port
,
3976 const char *affinity_list
)
3978 unsigned *core_ids
, i
;
3981 core_ids
= xmalloc(port
->n_rxq
* sizeof *core_ids
);
3982 if (parse_affinity_list(affinity_list
, core_ids
, port
->n_rxq
)) {
3987 for (i
= 0; i
< port
->n_rxq
; i
++) {
3988 port
->rxqs
[i
].core_id
= core_ids
[i
];
3996 /* Returns 'true' if one of the 'port's RX queues exists in 'poll_list'
3997 * of given PMD thread. */
3999 dpif_netdev_pmd_polls_port(struct dp_netdev_pmd_thread
*pmd
,
4000 struct dp_netdev_port
*port
)
4001 OVS_EXCLUDED(pmd
->port_mutex
)
4003 struct rxq_poll
*poll
;
4006 ovs_mutex_lock(&pmd
->port_mutex
);
4007 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
4008 if (port
== poll
->rxq
->port
) {
4013 ovs_mutex_unlock(&pmd
->port_mutex
);
4017 /* Updates port configuration from the database. The changes are actually
4018 * applied in dpif_netdev_run(). */
4020 dpif_netdev_port_set_config(struct dpif
*dpif
, odp_port_t port_no
,
4021 const struct smap
*cfg
)
4023 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4024 struct dp_netdev_port
*port
;
4026 const char *affinity_list
= smap_get(cfg
, "pmd-rxq-affinity");
4027 bool emc_enabled
= smap_get_bool(cfg
, "emc-enable", true);
4029 ovs_mutex_lock(&dp
->port_mutex
);
4030 error
= get_port_by_number(dp
, port_no
, &port
);
4035 if (emc_enabled
!= port
->emc_enabled
) {
4036 struct dp_netdev_pmd_thread
*pmd
;
4037 struct ds ds
= DS_EMPTY_INITIALIZER
;
4038 uint32_t cur_min
, insert_prob
;
4040 port
->emc_enabled
= emc_enabled
;
4041 /* Mark for reload all the threads that polls this port and request
4042 * for reconfiguration for the actual reloading of threads. */
4043 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4044 if (dpif_netdev_pmd_polls_port(pmd
, port
)) {
4045 pmd
->need_reload
= true;
4048 dp_netdev_request_reconfigure(dp
);
4050 ds_put_format(&ds
, "%s: EMC has been %s.",
4051 netdev_get_name(port
->netdev
),
4052 (emc_enabled
) ? "enabled" : "disabled");
4054 ds_put_cstr(&ds
, " Current insertion probability is ");
4055 atomic_read_relaxed(&dp
->emc_insert_min
, &cur_min
);
4057 ds_put_cstr(&ds
, "zero.");
4059 insert_prob
= UINT32_MAX
/ cur_min
;
4060 ds_put_format(&ds
, "1/%"PRIu32
" (~%.2f%%).",
4061 insert_prob
, 100 / (float) insert_prob
);
4064 VLOG_INFO("%s", ds_cstr(&ds
));
4068 /* Checking for RXq affinity changes. */
4069 if (!netdev_is_pmd(port
->netdev
)
4070 || nullable_string_is_equal(affinity_list
, port
->rxq_affinity_list
)) {
4074 error
= dpif_netdev_port_set_rxq_affinity(port
, affinity_list
);
4078 free(port
->rxq_affinity_list
);
4079 port
->rxq_affinity_list
= nullable_xstrdup(affinity_list
);
4081 dp_netdev_request_reconfigure(dp
);
4083 ovs_mutex_unlock(&dp
->port_mutex
);
4088 dpif_netdev_queue_to_priority(const struct dpif
*dpif OVS_UNUSED
,
4089 uint32_t queue_id
, uint32_t *priority
)
4091 *priority
= queue_id
;
4096 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
4097 * a copy of the 'size' bytes of 'actions' input parameters. */
4098 struct dp_netdev_actions
*
4099 dp_netdev_actions_create(const struct nlattr
*actions
, size_t size
)
4101 struct dp_netdev_actions
*netdev_actions
;
4103 netdev_actions
= xmalloc(sizeof *netdev_actions
+ size
);
4104 memcpy(netdev_actions
->actions
, actions
, size
);
4105 netdev_actions
->size
= size
;
4107 return netdev_actions
;
4110 struct dp_netdev_actions
*
4111 dp_netdev_flow_get_actions(const struct dp_netdev_flow
*flow
)
4113 return ovsrcu_get(struct dp_netdev_actions
*, &flow
->actions
);
4117 dp_netdev_actions_free(struct dp_netdev_actions
*actions
)
4123 dp_netdev_rxq_set_cycles(struct dp_netdev_rxq
*rx
,
4124 enum rxq_cycles_counter_type type
,
4125 unsigned long long cycles
)
4127 atomic_store_relaxed(&rx
->cycles
[type
], cycles
);
4131 dp_netdev_rxq_add_cycles(struct dp_netdev_rxq
*rx
,
4132 enum rxq_cycles_counter_type type
,
4133 unsigned long long cycles
)
4135 non_atomic_ullong_add(&rx
->cycles
[type
], cycles
);
4139 dp_netdev_rxq_get_cycles(struct dp_netdev_rxq
*rx
,
4140 enum rxq_cycles_counter_type type
)
4142 unsigned long long processing_cycles
;
4143 atomic_read_relaxed(&rx
->cycles
[type
], &processing_cycles
);
4144 return processing_cycles
;
4148 dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq
*rx
,
4149 unsigned long long cycles
)
4151 unsigned int idx
= rx
->intrvl_idx
++ % PMD_RXQ_INTERVAL_MAX
;
4152 atomic_store_relaxed(&rx
->cycles_intrvl
[idx
], cycles
);
4156 dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq
*rx
, unsigned idx
)
4158 unsigned long long processing_cycles
;
4159 atomic_read_relaxed(&rx
->cycles_intrvl
[idx
], &processing_cycles
);
4160 return processing_cycles
;
4163 #if ATOMIC_ALWAYS_LOCK_FREE_8B
4165 pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread
*pmd
)
4167 bool pmd_perf_enabled
;
4168 atomic_read_relaxed(&pmd
->dp
->pmd_perf_metrics
, &pmd_perf_enabled
);
4169 return pmd_perf_enabled
;
4172 /* If stores and reads of 64-bit integers are not atomic, the full PMD
4173 * performance metrics are not available as locked access to 64 bit
4174 * integers would be prohibitively expensive. */
4176 pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread
*pmd OVS_UNUSED
)
4183 dp_netdev_pmd_flush_output_on_port(struct dp_netdev_pmd_thread
*pmd
,
4190 struct cycle_timer timer
;
4192 uint32_t tx_flush_interval
;
4194 cycle_timer_start(&pmd
->perf_stats
, &timer
);
4196 dynamic_txqs
= p
->port
->dynamic_txqs
;
4198 tx_qid
= dpif_netdev_xps_get_tx_qid(pmd
, p
);
4200 tx_qid
= pmd
->static_tx_qid
;
4203 output_cnt
= dp_packet_batch_size(&p
->output_pkts
);
4204 ovs_assert(output_cnt
> 0);
4206 netdev_send(p
->port
->netdev
, tx_qid
, &p
->output_pkts
, dynamic_txqs
);
4207 dp_packet_batch_init(&p
->output_pkts
);
4209 /* Update time of the next flush. */
4210 atomic_read_relaxed(&pmd
->dp
->tx_flush_interval
, &tx_flush_interval
);
4211 p
->flush_time
= pmd
->ctx
.now
+ tx_flush_interval
;
4213 ovs_assert(pmd
->n_output_batches
> 0);
4214 pmd
->n_output_batches
--;
4216 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_SENT_PKTS
, output_cnt
);
4217 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_SENT_BATCHES
, 1);
4219 /* Distribute send cycles evenly among transmitted packets and assign to
4220 * their respective rx queues. */
4221 cycles
= cycle_timer_stop(&pmd
->perf_stats
, &timer
) / output_cnt
;
4222 for (i
= 0; i
< output_cnt
; i
++) {
4223 if (p
->output_pkts_rxqs
[i
]) {
4224 dp_netdev_rxq_add_cycles(p
->output_pkts_rxqs
[i
],
4225 RXQ_CYCLES_PROC_CURR
, cycles
);
4233 dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread
*pmd
,
4239 if (!pmd
->n_output_batches
) {
4243 HMAP_FOR_EACH (p
, node
, &pmd
->send_port_cache
) {
4244 if (!dp_packet_batch_is_empty(&p
->output_pkts
)
4245 && (force
|| pmd
->ctx
.now
>= p
->flush_time
)) {
4246 output_cnt
+= dp_netdev_pmd_flush_output_on_port(pmd
, p
);
4253 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread
*pmd
,
4254 struct dp_netdev_rxq
*rxq
,
4257 struct pmd_perf_stats
*s
= &pmd
->perf_stats
;
4258 struct dp_packet_batch batch
;
4259 struct cycle_timer timer
;
4262 int rem_qlen
= 0, *qlen_p
= NULL
;
4265 /* Measure duration for polling and processing rx burst. */
4266 cycle_timer_start(&pmd
->perf_stats
, &timer
);
4268 pmd
->ctx
.last_rxq
= rxq
;
4269 dp_packet_batch_init(&batch
);
4271 /* Fetch the rx queue length only for vhostuser ports. */
4272 if (pmd_perf_metrics_enabled(pmd
) && rxq
->is_vhost
) {
4276 error
= netdev_rxq_recv(rxq
->rx
, &batch
, qlen_p
);
4278 /* At least one packet received. */
4279 *recirc_depth_get() = 0;
4280 pmd_thread_ctx_time_update(pmd
);
4281 batch_cnt
= batch
.count
;
4282 if (pmd_perf_metrics_enabled(pmd
)) {
4283 /* Update batch histogram. */
4284 s
->current
.batches
++;
4285 histogram_add_sample(&s
->pkts_per_batch
, batch_cnt
);
4286 /* Update the maximum vhost rx queue fill level. */
4287 if (rxq
->is_vhost
&& rem_qlen
>= 0) {
4288 uint32_t qfill
= batch_cnt
+ rem_qlen
;
4289 if (qfill
> s
->current
.max_vhost_qfill
) {
4290 s
->current
.max_vhost_qfill
= qfill
;
4294 /* Process packet batch. */
4295 dp_netdev_input(pmd
, &batch
, port_no
);
4297 /* Assign processing cycles to rx queue. */
4298 cycles
= cycle_timer_stop(&pmd
->perf_stats
, &timer
);
4299 dp_netdev_rxq_add_cycles(rxq
, RXQ_CYCLES_PROC_CURR
, cycles
);
4301 dp_netdev_pmd_flush_output_packets(pmd
, false);
4303 /* Discard cycles. */
4304 cycle_timer_stop(&pmd
->perf_stats
, &timer
);
4305 if (error
!= EAGAIN
&& error
!= EOPNOTSUPP
) {
4306 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
4308 VLOG_ERR_RL(&rl
, "error receiving data from %s: %s",
4309 netdev_rxq_get_name(rxq
->rx
), ovs_strerror(error
));
4313 pmd
->ctx
.last_rxq
= NULL
;
4318 static struct tx_port
*
4319 tx_port_lookup(const struct hmap
*hmap
, odp_port_t port_no
)
4323 HMAP_FOR_EACH_IN_BUCKET (tx
, node
, hash_port_no(port_no
), hmap
) {
4324 if (tx
->port
->port_no
== port_no
) {
4333 port_reconfigure(struct dp_netdev_port
*port
)
4335 struct netdev
*netdev
= port
->netdev
;
4338 /* Closes the existing 'rxq's. */
4339 for (i
= 0; i
< port
->n_rxq
; i
++) {
4340 netdev_rxq_close(port
->rxqs
[i
].rx
);
4341 port
->rxqs
[i
].rx
= NULL
;
4343 unsigned last_nrxq
= port
->n_rxq
;
4346 /* Allows 'netdev' to apply the pending configuration changes. */
4347 if (netdev_is_reconf_required(netdev
) || port
->need_reconfigure
) {
4348 err
= netdev_reconfigure(netdev
);
4349 if (err
&& (err
!= EOPNOTSUPP
)) {
4350 VLOG_ERR("Failed to set interface %s new configuration",
4351 netdev_get_name(netdev
));
4355 /* If the netdev_reconfigure() above succeeds, reopens the 'rxq's. */
4356 port
->rxqs
= xrealloc(port
->rxqs
,
4357 sizeof *port
->rxqs
* netdev_n_rxq(netdev
));
4358 /* Realloc 'used' counters for tx queues. */
4359 free(port
->txq_used
);
4360 port
->txq_used
= xcalloc(netdev_n_txq(netdev
), sizeof *port
->txq_used
);
4362 for (i
= 0; i
< netdev_n_rxq(netdev
); i
++) {
4363 bool new_queue
= i
>= last_nrxq
;
4365 memset(&port
->rxqs
[i
], 0, sizeof port
->rxqs
[i
]);
4368 port
->rxqs
[i
].port
= port
;
4369 port
->rxqs
[i
].is_vhost
= !strncmp(port
->type
, "dpdkvhost", 9);
4371 err
= netdev_rxq_open(netdev
, &port
->rxqs
[i
].rx
, i
);
4378 /* Parse affinity list to apply configuration for new queues. */
4379 dpif_netdev_port_set_rxq_affinity(port
, port
->rxq_affinity_list
);
4381 /* If reconfiguration was successful mark it as such, so we can use it */
4382 port
->need_reconfigure
= false;
4387 struct rr_numa_list
{
4388 struct hmap numas
; /* Contains 'struct rr_numa' */
4392 struct hmap_node node
;
4396 /* Non isolated pmds on numa node 'numa_id' */
4397 struct dp_netdev_pmd_thread
**pmds
;
4404 static struct rr_numa
*
4405 rr_numa_list_lookup(struct rr_numa_list
*rr
, int numa_id
)
4407 struct rr_numa
*numa
;
4409 HMAP_FOR_EACH_WITH_HASH (numa
, node
, hash_int(numa_id
, 0), &rr
->numas
) {
4410 if (numa
->numa_id
== numa_id
) {
4418 /* Returns the next node in numa list following 'numa' in round-robin fashion.
4419 * Returns first node if 'numa' is a null pointer or the last node in 'rr'.
4420 * Returns NULL if 'rr' numa list is empty. */
4421 static struct rr_numa
*
4422 rr_numa_list_next(struct rr_numa_list
*rr
, const struct rr_numa
*numa
)
4424 struct hmap_node
*node
= NULL
;
4427 node
= hmap_next(&rr
->numas
, &numa
->node
);
4430 node
= hmap_first(&rr
->numas
);
4433 return (node
) ? CONTAINER_OF(node
, struct rr_numa
, node
) : NULL
;
4437 rr_numa_list_populate(struct dp_netdev
*dp
, struct rr_numa_list
*rr
)
4439 struct dp_netdev_pmd_thread
*pmd
;
4440 struct rr_numa
*numa
;
4442 hmap_init(&rr
->numas
);
4444 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4445 if (pmd
->core_id
== NON_PMD_CORE_ID
|| pmd
->isolated
) {
4449 numa
= rr_numa_list_lookup(rr
, pmd
->numa_id
);
4451 numa
= xzalloc(sizeof *numa
);
4452 numa
->numa_id
= pmd
->numa_id
;
4453 hmap_insert(&rr
->numas
, &numa
->node
, hash_int(pmd
->numa_id
, 0));
4456 numa
->pmds
= xrealloc(numa
->pmds
, numa
->n_pmds
* sizeof *numa
->pmds
);
4457 numa
->pmds
[numa
->n_pmds
- 1] = pmd
;
4458 /* At least one pmd so initialise curr_idx and idx_inc. */
4459 numa
->cur_index
= 0;
4460 numa
->idx_inc
= true;
4465 * Returns the next pmd from the numa node.
4467 * If 'updown' is 'true' it will alternate between selecting the next pmd in
4468 * either an up or down walk, switching between up/down when the first or last
4469 * core is reached. e.g. 1,2,3,3,2,1,1,2...
4471 * If 'updown' is 'false' it will select the next pmd wrapping around when last
4472 * core reached. e.g. 1,2,3,1,2,3,1,2...
4474 static struct dp_netdev_pmd_thread
*
4475 rr_numa_get_pmd(struct rr_numa
*numa
, bool updown
)
4477 int numa_idx
= numa
->cur_index
;
4479 if (numa
->idx_inc
== true) {
4480 /* Incrementing through list of pmds. */
4481 if (numa
->cur_index
== numa
->n_pmds
-1) {
4482 /* Reached the last pmd. */
4484 numa
->idx_inc
= false;
4486 numa
->cur_index
= 0;
4492 /* Decrementing through list of pmds. */
4493 if (numa
->cur_index
== 0) {
4494 /* Reached the first pmd. */
4495 numa
->idx_inc
= true;
4500 return numa
->pmds
[numa_idx
];
4504 rr_numa_list_destroy(struct rr_numa_list
*rr
)
4506 struct rr_numa
*numa
;
4508 HMAP_FOR_EACH_POP (numa
, node
, &rr
->numas
) {
4512 hmap_destroy(&rr
->numas
);
4515 /* Sort Rx Queues by the processing cycles they are consuming. */
4517 compare_rxq_cycles(const void *a
, const void *b
)
4519 struct dp_netdev_rxq
*qa
;
4520 struct dp_netdev_rxq
*qb
;
4521 uint64_t cycles_qa
, cycles_qb
;
4523 qa
= *(struct dp_netdev_rxq
**) a
;
4524 qb
= *(struct dp_netdev_rxq
**) b
;
4526 cycles_qa
= dp_netdev_rxq_get_cycles(qa
, RXQ_CYCLES_PROC_HIST
);
4527 cycles_qb
= dp_netdev_rxq_get_cycles(qb
, RXQ_CYCLES_PROC_HIST
);
4529 if (cycles_qa
!= cycles_qb
) {
4530 return (cycles_qa
< cycles_qb
) ? 1 : -1;
4532 /* Cycles are the same so tiebreak on port/queue id.
4533 * Tiebreaking (as opposed to return 0) ensures consistent
4534 * sort results across multiple OS's. */
4535 uint32_t port_qa
= odp_to_u32(qa
->port
->port_no
);
4536 uint32_t port_qb
= odp_to_u32(qb
->port
->port_no
);
4537 if (port_qa
!= port_qb
) {
4538 return port_qa
> port_qb
? 1 : -1;
4540 return netdev_rxq_get_queue_id(qa
->rx
)
4541 - netdev_rxq_get_queue_id(qb
->rx
);
4546 /* Assign pmds to queues. If 'pinned' is true, assign pmds to pinned
4547 * queues and marks the pmds as isolated. Otherwise, assign non isolated
4548 * pmds to unpinned queues.
4550 * The function doesn't touch the pmd threads, it just stores the assignment
4551 * in the 'pmd' member of each rxq. */
4553 rxq_scheduling(struct dp_netdev
*dp
, bool pinned
) OVS_REQUIRES(dp
->port_mutex
)
4555 struct dp_netdev_port
*port
;
4556 struct rr_numa_list rr
;
4557 struct rr_numa
*non_local_numa
= NULL
;
4558 struct dp_netdev_rxq
** rxqs
= NULL
;
4560 struct rr_numa
*numa
= NULL
;
4562 bool assign_cyc
= dp
->pmd_rxq_assign_cyc
;
4564 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4565 if (!netdev_is_pmd(port
->netdev
)) {
4569 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
4570 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
4572 if (pinned
&& q
->core_id
!= OVS_CORE_UNSPEC
) {
4573 struct dp_netdev_pmd_thread
*pmd
;
4575 pmd
= dp_netdev_get_pmd(dp
, q
->core_id
);
4577 VLOG_WARN("There is no PMD thread on core %d. Queue "
4578 "%d on port \'%s\' will not be polled.",
4579 q
->core_id
, qid
, netdev_get_name(port
->netdev
));
4582 pmd
->isolated
= true;
4583 dp_netdev_pmd_unref(pmd
);
4585 } else if (!pinned
&& q
->core_id
== OVS_CORE_UNSPEC
) {
4586 uint64_t cycle_hist
= 0;
4589 rxqs
= xmalloc(sizeof *rxqs
);
4591 rxqs
= xrealloc(rxqs
, sizeof *rxqs
* (n_rxqs
+ 1));
4595 /* Sum the queue intervals and store the cycle history. */
4596 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
4597 cycle_hist
+= dp_netdev_rxq_get_intrvl_cycles(q
, i
);
4599 dp_netdev_rxq_set_cycles(q
, RXQ_CYCLES_PROC_HIST
,
4602 /* Store the queue. */
4608 if (n_rxqs
> 1 && assign_cyc
) {
4609 /* Sort the queues in order of the processing cycles
4610 * they consumed during their last pmd interval. */
4611 qsort(rxqs
, n_rxqs
, sizeof *rxqs
, compare_rxq_cycles
);
4614 rr_numa_list_populate(dp
, &rr
);
4615 /* Assign the sorted queues to pmds in round robin. */
4616 for (int i
= 0; i
< n_rxqs
; i
++) {
4617 numa_id
= netdev_get_numa_id(rxqs
[i
]->port
->netdev
);
4618 numa
= rr_numa_list_lookup(&rr
, numa_id
);
4620 /* There are no pmds on the queue's local NUMA node.
4621 Round robin on the NUMA nodes that do have pmds. */
4622 non_local_numa
= rr_numa_list_next(&rr
, non_local_numa
);
4623 if (!non_local_numa
) {
4624 VLOG_ERR("There is no available (non-isolated) pmd "
4625 "thread for port \'%s\' queue %d. This queue "
4626 "will not be polled. Is pmd-cpu-mask set to "
4627 "zero? Or are all PMDs isolated to other "
4628 "queues?", netdev_rxq_get_name(rxqs
[i
]->rx
),
4629 netdev_rxq_get_queue_id(rxqs
[i
]->rx
));
4632 rxqs
[i
]->pmd
= rr_numa_get_pmd(non_local_numa
, assign_cyc
);
4633 VLOG_WARN("There's no available (non-isolated) pmd thread "
4634 "on numa node %d. Queue %d on port \'%s\' will "
4635 "be assigned to the pmd on core %d "
4636 "(numa node %d). Expect reduced performance.",
4637 numa_id
, netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
4638 netdev_rxq_get_name(rxqs
[i
]->rx
),
4639 rxqs
[i
]->pmd
->core_id
, rxqs
[i
]->pmd
->numa_id
);
4641 rxqs
[i
]->pmd
= rr_numa_get_pmd(numa
, assign_cyc
);
4643 VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
4645 "(measured processing cycles %"PRIu64
").",
4646 rxqs
[i
]->pmd
->core_id
, numa_id
,
4647 netdev_rxq_get_name(rxqs
[i
]->rx
),
4648 netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
4649 dp_netdev_rxq_get_cycles(rxqs
[i
],
4650 RXQ_CYCLES_PROC_HIST
));
4652 VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
4653 "rx queue %d.", rxqs
[i
]->pmd
->core_id
, numa_id
,
4654 netdev_rxq_get_name(rxqs
[i
]->rx
),
4655 netdev_rxq_get_queue_id(rxqs
[i
]->rx
));
4660 rr_numa_list_destroy(&rr
);
4665 reload_affected_pmds(struct dp_netdev
*dp
)
4667 struct dp_netdev_pmd_thread
*pmd
;
4669 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4670 if (pmd
->need_reload
) {
4671 flow_mark_flush(pmd
);
4672 dp_netdev_reload_pmd__(pmd
);
4676 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4677 if (pmd
->need_reload
) {
4678 if (pmd
->core_id
!= NON_PMD_CORE_ID
) {
4682 atomic_read_explicit(&pmd
->reload
, &reload
,
4683 memory_order_acquire
);
4686 pmd
->need_reload
= false;
4692 reconfigure_pmd_threads(struct dp_netdev
*dp
)
4693 OVS_REQUIRES(dp
->port_mutex
)
4695 struct dp_netdev_pmd_thread
*pmd
;
4696 struct ovs_numa_dump
*pmd_cores
;
4697 struct ovs_numa_info_core
*core
;
4698 struct hmapx to_delete
= HMAPX_INITIALIZER(&to_delete
);
4699 struct hmapx_node
*node
;
4700 bool changed
= false;
4701 bool need_to_adjust_static_tx_qids
= false;
4703 /* The pmd threads should be started only if there's a pmd port in the
4704 * datapath. If the user didn't provide any "pmd-cpu-mask", we start
4705 * NR_PMD_THREADS per numa node. */
4706 if (!has_pmd_port(dp
)) {
4707 pmd_cores
= ovs_numa_dump_n_cores_per_numa(0);
4708 } else if (dp
->pmd_cmask
&& dp
->pmd_cmask
[0]) {
4709 pmd_cores
= ovs_numa_dump_cores_with_cmask(dp
->pmd_cmask
);
4711 pmd_cores
= ovs_numa_dump_n_cores_per_numa(NR_PMD_THREADS
);
4714 /* We need to adjust 'static_tx_qid's only if we're reducing number of
4715 * PMD threads. Otherwise, new threads will allocate all the freed ids. */
4716 if (ovs_numa_dump_count(pmd_cores
) < cmap_count(&dp
->poll_threads
) - 1) {
4717 /* Adjustment is required to keep 'static_tx_qid's sequential and
4718 * avoid possible issues, for example, imbalanced tx queue usage
4719 * and unnecessary locking caused by remapping on netdev level. */
4720 need_to_adjust_static_tx_qids
= true;
4723 /* Check for unwanted pmd threads */
4724 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4725 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
4728 if (!ovs_numa_dump_contains_core(pmd_cores
, pmd
->numa_id
,
4730 hmapx_add(&to_delete
, pmd
);
4731 } else if (need_to_adjust_static_tx_qids
) {
4732 atomic_store_relaxed(&pmd
->reload_tx_qid
, true);
4733 pmd
->need_reload
= true;
4737 HMAPX_FOR_EACH (node
, &to_delete
) {
4738 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
4739 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d destroyed.",
4740 pmd
->numa_id
, pmd
->core_id
);
4741 dp_netdev_del_pmd(dp
, pmd
);
4743 changed
= !hmapx_is_empty(&to_delete
);
4744 hmapx_destroy(&to_delete
);
4746 if (need_to_adjust_static_tx_qids
) {
4747 /* 'static_tx_qid's are not sequential now.
4748 * Reload remaining threads to fix this. */
4749 reload_affected_pmds(dp
);
4752 /* Check for required new pmd threads */
4753 FOR_EACH_CORE_ON_DUMP(core
, pmd_cores
) {
4754 pmd
= dp_netdev_get_pmd(dp
, core
->core_id
);
4756 pmd
= xzalloc(sizeof *pmd
);
4757 dp_netdev_configure_pmd(pmd
, dp
, core
->core_id
, core
->numa_id
);
4758 pmd
->thread
= ovs_thread_create("pmd", pmd_thread_main
, pmd
);
4759 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d created.",
4760 pmd
->numa_id
, pmd
->core_id
);
4763 dp_netdev_pmd_unref(pmd
);
4768 struct ovs_numa_info_numa
*numa
;
4770 /* Log the number of pmd threads per numa node. */
4771 FOR_EACH_NUMA_ON_DUMP (numa
, pmd_cores
) {
4772 VLOG_INFO("There are %"PRIuSIZE
" pmd threads on numa node %d",
4773 numa
->n_cores
, numa
->numa_id
);
4777 ovs_numa_dump_destroy(pmd_cores
);
4781 pmd_remove_stale_ports(struct dp_netdev
*dp
,
4782 struct dp_netdev_pmd_thread
*pmd
)
4783 OVS_EXCLUDED(pmd
->port_mutex
)
4784 OVS_REQUIRES(dp
->port_mutex
)
4786 struct rxq_poll
*poll
, *poll_next
;
4787 struct tx_port
*tx
, *tx_next
;
4789 ovs_mutex_lock(&pmd
->port_mutex
);
4790 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
4791 struct dp_netdev_port
*port
= poll
->rxq
->port
;
4793 if (port
->need_reconfigure
4794 || !hmap_contains(&dp
->ports
, &port
->node
)) {
4795 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
4798 HMAP_FOR_EACH_SAFE (tx
, tx_next
, node
, &pmd
->tx_ports
) {
4799 struct dp_netdev_port
*port
= tx
->port
;
4801 if (port
->need_reconfigure
4802 || !hmap_contains(&dp
->ports
, &port
->node
)) {
4803 dp_netdev_del_port_tx_from_pmd(pmd
, tx
);
4806 ovs_mutex_unlock(&pmd
->port_mutex
);
4809 /* Must be called each time a port is added/removed or the cmask changes.
4810 * This creates and destroys pmd threads, reconfigures ports, opens their
4811 * rxqs and assigns all rxqs/txqs to pmd threads. */
4813 reconfigure_datapath(struct dp_netdev
*dp
)
4814 OVS_REQUIRES(dp
->port_mutex
)
4816 struct hmapx busy_threads
= HMAPX_INITIALIZER(&busy_threads
);
4817 struct dp_netdev_pmd_thread
*pmd
;
4818 struct dp_netdev_port
*port
;
4821 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
4823 /* Step 1: Adjust the pmd threads based on the datapath ports, the cores
4824 * on the system and the user configuration. */
4825 reconfigure_pmd_threads(dp
);
4827 wanted_txqs
= cmap_count(&dp
->poll_threads
);
4829 /* The number of pmd threads might have changed, or a port can be new:
4830 * adjust the txqs. */
4831 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4832 netdev_set_tx_multiq(port
->netdev
, wanted_txqs
);
4835 /* Step 2: Remove from the pmd threads ports that have been removed or
4836 * need reconfiguration. */
4838 /* Check for all the ports that need reconfiguration. We cache this in
4839 * 'port->need_reconfigure', because netdev_is_reconf_required() can
4840 * change at any time. */
4841 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4842 if (netdev_is_reconf_required(port
->netdev
)) {
4843 port
->need_reconfigure
= true;
4847 /* Remove from the pmd threads all the ports that have been deleted or
4848 * need reconfiguration. */
4849 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4850 pmd_remove_stale_ports(dp
, pmd
);
4853 /* Reload affected pmd threads. We must wait for the pmd threads before
4854 * reconfiguring the ports, because a port cannot be reconfigured while
4855 * it's being used. */
4856 reload_affected_pmds(dp
);
4858 /* Step 3: Reconfigure ports. */
4860 /* We only reconfigure the ports that we determined above, because they're
4861 * not being used by any pmd thread at the moment. If a port fails to
4862 * reconfigure we remove it from the datapath. */
4863 struct dp_netdev_port
*next_port
;
4864 HMAP_FOR_EACH_SAFE (port
, next_port
, node
, &dp
->ports
) {
4867 if (!port
->need_reconfigure
) {
4871 err
= port_reconfigure(port
);
4873 hmap_remove(&dp
->ports
, &port
->node
);
4874 seq_change(dp
->port_seq
);
4877 port
->dynamic_txqs
= netdev_n_txq(port
->netdev
) < wanted_txqs
;
4881 /* Step 4: Compute new rxq scheduling. We don't touch the pmd threads
4882 * for now, we just update the 'pmd' pointer in each rxq to point to the
4883 * wanted thread according to the scheduling policy. */
4885 /* Reset all the pmd threads to non isolated. */
4886 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4887 pmd
->isolated
= false;
4890 /* Reset all the queues to unassigned */
4891 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4892 for (int i
= 0; i
< port
->n_rxq
; i
++) {
4893 port
->rxqs
[i
].pmd
= NULL
;
4897 /* Add pinned queues and mark pmd threads isolated. */
4898 rxq_scheduling(dp
, true);
4900 /* Add non-pinned queues. */
4901 rxq_scheduling(dp
, false);
4903 /* Step 5: Remove queues not compliant with new scheduling. */
4905 /* Count all the threads that will have at least one queue to poll. */
4906 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4907 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
4908 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
4911 hmapx_add(&busy_threads
, q
->pmd
);
4916 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4917 struct rxq_poll
*poll
, *poll_next
;
4919 ovs_mutex_lock(&pmd
->port_mutex
);
4920 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
4921 if (poll
->rxq
->pmd
!= pmd
) {
4922 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
4924 /* This pmd might sleep after this step if it has no rxq
4925 * remaining. Tell it to busy wait for new assignment if it
4926 * has at least one scheduled queue. */
4927 if (hmap_count(&pmd
->poll_list
) == 0 &&
4928 hmapx_contains(&busy_threads
, pmd
)) {
4929 atomic_store_relaxed(&pmd
->wait_for_reload
, true);
4933 ovs_mutex_unlock(&pmd
->port_mutex
);
4936 hmapx_destroy(&busy_threads
);
4938 /* Reload affected pmd threads. We must wait for the pmd threads to remove
4939 * the old queues before readding them, otherwise a queue can be polled by
4940 * two threads at the same time. */
4941 reload_affected_pmds(dp
);
4943 /* Step 6: Add queues from scheduling, if they're not there already. */
4944 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4945 if (!netdev_is_pmd(port
->netdev
)) {
4949 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
4950 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
4953 ovs_mutex_lock(&q
->pmd
->port_mutex
);
4954 dp_netdev_add_rxq_to_pmd(q
->pmd
, q
);
4955 ovs_mutex_unlock(&q
->pmd
->port_mutex
);
4960 /* Add every port to the tx cache of every pmd thread, if it's not
4961 * there already and if this pmd has at least one rxq to poll. */
4962 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4963 ovs_mutex_lock(&pmd
->port_mutex
);
4964 if (hmap_count(&pmd
->poll_list
) || pmd
->core_id
== NON_PMD_CORE_ID
) {
4965 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4966 dp_netdev_add_port_tx_to_pmd(pmd
, port
);
4969 ovs_mutex_unlock(&pmd
->port_mutex
);
4972 /* Reload affected pmd threads. */
4973 reload_affected_pmds(dp
);
4975 /* Check if PMD Auto LB is to be enabled */
4976 set_pmd_auto_lb(dp
);
4979 /* Returns true if one of the netdevs in 'dp' requires a reconfiguration */
4981 ports_require_restart(const struct dp_netdev
*dp
)
4982 OVS_REQUIRES(dp
->port_mutex
)
4984 struct dp_netdev_port
*port
;
4986 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4987 if (netdev_is_reconf_required(port
->netdev
)) {
4995 /* Calculates variance in the values stored in array 'a'. 'n' is the number
4996 * of elements in array to be considered for calculating vairance.
4997 * Usage example: data array 'a' contains the processing load of each pmd and
4998 * 'n' is the number of PMDs. It returns the variance in processing load of
5001 variance(uint64_t a
[], int n
)
5003 /* Compute mean (average of elements). */
5006 uint64_t sqDiff
= 0;
5012 for (int i
= 0; i
< n
; i
++) {
5019 /* Compute sum squared differences with mean. */
5020 for (int i
= 0; i
< n
; i
++) {
5021 sqDiff
+= (a
[i
] - mean
)*(a
[i
] - mean
);
5024 return (sqDiff
? (sqDiff
/ n
) : 0);
5028 /* Returns the variance in the PMDs usage as part of dry run of rxqs
5029 * assignment to PMDs. */
5031 get_dry_run_variance(struct dp_netdev
*dp
, uint32_t *core_list
,
5032 uint32_t num_pmds
, uint64_t *predicted_variance
)
5033 OVS_REQUIRES(dp
->port_mutex
)
5035 struct dp_netdev_port
*port
;
5036 struct dp_netdev_pmd_thread
*pmd
;
5037 struct dp_netdev_rxq
**rxqs
= NULL
;
5038 struct rr_numa
*numa
= NULL
;
5039 struct rr_numa_list rr
;
5042 uint64_t *pmd_usage
;
5044 if (!predicted_variance
) {
5048 pmd_usage
= xcalloc(num_pmds
, sizeof(uint64_t));
5050 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5051 if (!netdev_is_pmd(port
->netdev
)) {
5055 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
5056 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
5057 uint64_t cycle_hist
= 0;
5059 if (q
->pmd
->isolated
) {
5064 rxqs
= xmalloc(sizeof *rxqs
);
5066 rxqs
= xrealloc(rxqs
, sizeof *rxqs
* (n_rxqs
+ 1));
5069 /* Sum the queue intervals and store the cycle history. */
5070 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
5071 cycle_hist
+= dp_netdev_rxq_get_intrvl_cycles(q
, i
);
5073 dp_netdev_rxq_set_cycles(q
, RXQ_CYCLES_PROC_HIST
,
5075 /* Store the queue. */
5080 /* Sort the queues in order of the processing cycles
5081 * they consumed during their last pmd interval. */
5082 qsort(rxqs
, n_rxqs
, sizeof *rxqs
, compare_rxq_cycles
);
5084 rr_numa_list_populate(dp
, &rr
);
5086 for (int i
= 0; i
< n_rxqs
; i
++) {
5087 int numa_id
= netdev_get_numa_id(rxqs
[i
]->port
->netdev
);
5088 numa
= rr_numa_list_lookup(&rr
, numa_id
);
5090 /* Abort if cross NUMA polling. */
5091 VLOG_DBG("PMD auto lb dry run."
5092 " Aborting due to cross-numa polling.");
5096 pmd
= rr_numa_get_pmd(numa
, true);
5097 VLOG_DBG("PMD auto lb dry run. Predicted: Core %d on numa node %d "
5098 "to be assigned port \'%s\' rx queue %d "
5099 "(measured processing cycles %"PRIu64
").",
5100 pmd
->core_id
, numa_id
,
5101 netdev_rxq_get_name(rxqs
[i
]->rx
),
5102 netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
5103 dp_netdev_rxq_get_cycles(rxqs
[i
], RXQ_CYCLES_PROC_HIST
));
5105 for (int id
= 0; id
< num_pmds
; id
++) {
5106 if (pmd
->core_id
== core_list
[id
]) {
5107 /* Add the processing cycles of rxq to pmd polling it. */
5108 pmd_usage
[id
] += dp_netdev_rxq_get_cycles(rxqs
[i
],
5109 RXQ_CYCLES_PROC_HIST
);
5114 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
5115 uint64_t total_cycles
= 0;
5117 if ((pmd
->core_id
== NON_PMD_CORE_ID
) || pmd
->isolated
) {
5121 /* Get the total pmd cycles for an interval. */
5122 atomic_read_relaxed(&pmd
->intrvl_cycles
, &total_cycles
);
5123 /* Estimate the cycles to cover all intervals. */
5124 total_cycles
*= PMD_RXQ_INTERVAL_MAX
;
5125 for (int id
= 0; id
< num_pmds
; id
++) {
5126 if (pmd
->core_id
== core_list
[id
]) {
5127 if (pmd_usage
[id
]) {
5128 pmd_usage
[id
] = (pmd_usage
[id
] * 100) / total_cycles
;
5130 VLOG_DBG("PMD auto lb dry run. Predicted: Core %d, "
5131 "usage %"PRIu64
"", pmd
->core_id
, pmd_usage
[id
]);
5135 *predicted_variance
= variance(pmd_usage
, num_pmds
);
5139 rr_numa_list_destroy(&rr
);
5145 /* Does the dry run of Rxq assignment to PMDs and returns true if it gives
5146 * better distribution of load on PMDs. */
5148 pmd_rebalance_dry_run(struct dp_netdev
*dp
)
5149 OVS_REQUIRES(dp
->port_mutex
)
5151 struct dp_netdev_pmd_thread
*pmd
;
5152 uint64_t *curr_pmd_usage
;
5154 uint64_t curr_variance
;
5155 uint64_t new_variance
;
5156 uint64_t improvement
= 0;
5158 uint32_t *pmd_corelist
;
5159 struct rxq_poll
*poll
;
5162 num_pmds
= cmap_count(&dp
->poll_threads
);
5165 curr_pmd_usage
= xcalloc(num_pmds
, sizeof(uint64_t));
5166 pmd_corelist
= xcalloc(num_pmds
, sizeof(uint32_t));
5172 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
5173 uint64_t total_cycles
= 0;
5174 uint64_t total_proc
= 0;
5176 if ((pmd
->core_id
== NON_PMD_CORE_ID
) || pmd
->isolated
) {
5180 /* Get the total pmd cycles for an interval. */
5181 atomic_read_relaxed(&pmd
->intrvl_cycles
, &total_cycles
);
5182 /* Estimate the cycles to cover all intervals. */
5183 total_cycles
*= PMD_RXQ_INTERVAL_MAX
;
5185 ovs_mutex_lock(&pmd
->port_mutex
);
5186 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
5187 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
5188 total_proc
+= dp_netdev_rxq_get_intrvl_cycles(poll
->rxq
, i
);
5191 ovs_mutex_unlock(&pmd
->port_mutex
);
5194 curr_pmd_usage
[num_pmds
] = (total_proc
* 100) / total_cycles
;
5197 VLOG_DBG("PMD auto lb dry run. Current: Core %d, usage %"PRIu64
"",
5198 pmd
->core_id
, curr_pmd_usage
[num_pmds
]);
5200 if (atomic_count_get(&pmd
->pmd_overloaded
)) {
5201 atomic_count_set(&pmd
->pmd_overloaded
, 0);
5204 pmd_corelist
[num_pmds
] = pmd
->core_id
;
5208 curr_variance
= variance(curr_pmd_usage
, num_pmds
);
5209 ret
= get_dry_run_variance(dp
, pmd_corelist
, num_pmds
, &new_variance
);
5212 VLOG_DBG("PMD auto lb dry run. Current PMD variance: %"PRIu64
","
5213 " Predicted PMD variance: %"PRIu64
"",
5214 curr_variance
, new_variance
);
5216 if (new_variance
< curr_variance
) {
5218 ((curr_variance
- new_variance
) * 100) / curr_variance
;
5220 if (improvement
< ALB_ACCEPTABLE_IMPROVEMENT
) {
5225 free(curr_pmd_usage
);
5231 /* Return true if needs to revalidate datapath flows. */
5233 dpif_netdev_run(struct dpif
*dpif
)
5235 struct dp_netdev_port
*port
;
5236 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5237 struct dp_netdev_pmd_thread
*non_pmd
;
5238 uint64_t new_tnl_seq
;
5239 bool need_to_flush
= true;
5240 bool pmd_rebalance
= false;
5241 long long int now
= time_msec();
5242 struct dp_netdev_pmd_thread
*pmd
;
5244 ovs_mutex_lock(&dp
->port_mutex
);
5245 non_pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
5247 ovs_mutex_lock(&dp
->non_pmd_mutex
);
5248 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5249 if (!netdev_is_pmd(port
->netdev
)) {
5252 if (port
->emc_enabled
) {
5253 atomic_read_relaxed(&dp
->emc_insert_min
,
5254 &non_pmd
->ctx
.emc_insert_min
);
5256 non_pmd
->ctx
.emc_insert_min
= 0;
5259 for (i
= 0; i
< port
->n_rxq
; i
++) {
5261 if (!netdev_rxq_enabled(port
->rxqs
[i
].rx
)) {
5265 if (dp_netdev_process_rxq_port(non_pmd
,
5268 need_to_flush
= false;
5273 if (need_to_flush
) {
5274 /* We didn't receive anything in the process loop.
5275 * Check if we need to send something.
5276 * There was no time updates on current iteration. */
5277 pmd_thread_ctx_time_update(non_pmd
);
5278 dp_netdev_pmd_flush_output_packets(non_pmd
, false);
5281 dpif_netdev_xps_revalidate_pmd(non_pmd
, false);
5282 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
5284 dp_netdev_pmd_unref(non_pmd
);
5287 struct pmd_auto_lb
*pmd_alb
= &dp
->pmd_alb
;
5288 if (pmd_alb
->is_enabled
) {
5289 if (!pmd_alb
->rebalance_poll_timer
) {
5290 pmd_alb
->rebalance_poll_timer
= now
;
5291 } else if ((pmd_alb
->rebalance_poll_timer
+
5292 pmd_alb
->rebalance_intvl
) < now
) {
5293 pmd_alb
->rebalance_poll_timer
= now
;
5294 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
5295 if (atomic_count_get(&pmd
->pmd_overloaded
) >=
5296 PMD_RXQ_INTERVAL_MAX
) {
5297 pmd_rebalance
= true;
5302 if (pmd_rebalance
&&
5303 !dp_netdev_is_reconf_required(dp
) &&
5304 !ports_require_restart(dp
) &&
5305 pmd_rebalance_dry_run(dp
)) {
5306 VLOG_INFO("PMD auto lb dry run."
5307 " requesting datapath reconfigure.");
5308 dp_netdev_request_reconfigure(dp
);
5313 if (dp_netdev_is_reconf_required(dp
) || ports_require_restart(dp
)) {
5314 reconfigure_datapath(dp
);
5316 ovs_mutex_unlock(&dp
->port_mutex
);
5318 tnl_neigh_cache_run();
5320 new_tnl_seq
= seq_read(tnl_conf_seq
);
5322 if (dp
->last_tnl_conf_seq
!= new_tnl_seq
) {
5323 dp
->last_tnl_conf_seq
= new_tnl_seq
;
5330 dpif_netdev_wait(struct dpif
*dpif
)
5332 struct dp_netdev_port
*port
;
5333 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5335 ovs_mutex_lock(&dp_netdev_mutex
);
5336 ovs_mutex_lock(&dp
->port_mutex
);
5337 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5338 netdev_wait_reconf_required(port
->netdev
);
5339 if (!netdev_is_pmd(port
->netdev
)) {
5342 for (i
= 0; i
< port
->n_rxq
; i
++) {
5343 netdev_rxq_wait(port
->rxqs
[i
].rx
);
5347 ovs_mutex_unlock(&dp
->port_mutex
);
5348 ovs_mutex_unlock(&dp_netdev_mutex
);
5349 seq_wait(tnl_conf_seq
, dp
->last_tnl_conf_seq
);
5353 pmd_free_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
5355 struct tx_port
*tx_port_cached
;
5357 /* Flush all the queued packets. */
5358 dp_netdev_pmd_flush_output_packets(pmd
, true);
5359 /* Free all used tx queue ids. */
5360 dpif_netdev_xps_revalidate_pmd(pmd
, true);
5362 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->tnl_port_cache
) {
5363 free(tx_port_cached
);
5365 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->send_port_cache
) {
5366 free(tx_port_cached
);
5370 /* Copies ports from 'pmd->tx_ports' (shared with the main thread) to
5371 * thread-local copies. Copy to 'pmd->tnl_port_cache' if it is a tunnel
5372 * device, otherwise to 'pmd->send_port_cache' if the port has at least
5375 pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
5376 OVS_REQUIRES(pmd
->port_mutex
)
5378 struct tx_port
*tx_port
, *tx_port_cached
;
5380 pmd_free_cached_ports(pmd
);
5381 hmap_shrink(&pmd
->send_port_cache
);
5382 hmap_shrink(&pmd
->tnl_port_cache
);
5384 HMAP_FOR_EACH (tx_port
, node
, &pmd
->tx_ports
) {
5385 if (netdev_has_tunnel_push_pop(tx_port
->port
->netdev
)) {
5386 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
5387 hmap_insert(&pmd
->tnl_port_cache
, &tx_port_cached
->node
,
5388 hash_port_no(tx_port_cached
->port
->port_no
));
5391 if (netdev_n_txq(tx_port
->port
->netdev
)) {
5392 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
5393 hmap_insert(&pmd
->send_port_cache
, &tx_port_cached
->node
,
5394 hash_port_no(tx_port_cached
->port
->port_no
));
5400 pmd_alloc_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
5402 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
5403 if (!id_pool_alloc_id(pmd
->dp
->tx_qid_pool
, &pmd
->static_tx_qid
)) {
5404 VLOG_ABORT("static_tx_qid allocation failed for PMD on core %2d"
5405 ", numa_id %d.", pmd
->core_id
, pmd
->numa_id
);
5407 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
5409 VLOG_DBG("static_tx_qid = %d allocated for PMD thread on core %2d"
5410 ", numa_id %d.", pmd
->static_tx_qid
, pmd
->core_id
, pmd
->numa_id
);
5414 pmd_free_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
5416 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
5417 id_pool_free_id(pmd
->dp
->tx_qid_pool
, pmd
->static_tx_qid
);
5418 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
5422 pmd_load_queues_and_ports(struct dp_netdev_pmd_thread
*pmd
,
5423 struct polled_queue
**ppoll_list
)
5425 struct polled_queue
*poll_list
= *ppoll_list
;
5426 struct rxq_poll
*poll
;
5429 ovs_mutex_lock(&pmd
->port_mutex
);
5430 poll_list
= xrealloc(poll_list
, hmap_count(&pmd
->poll_list
)
5431 * sizeof *poll_list
);
5434 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
5435 poll_list
[i
].rxq
= poll
->rxq
;
5436 poll_list
[i
].port_no
= poll
->rxq
->port
->port_no
;
5437 poll_list
[i
].emc_enabled
= poll
->rxq
->port
->emc_enabled
;
5438 poll_list
[i
].rxq_enabled
= netdev_rxq_enabled(poll
->rxq
->rx
);
5439 poll_list
[i
].change_seq
=
5440 netdev_get_change_seq(poll
->rxq
->port
->netdev
);
5444 pmd_load_cached_ports(pmd
);
5446 ovs_mutex_unlock(&pmd
->port_mutex
);
5448 *ppoll_list
= poll_list
;
5453 pmd_thread_main(void *f_
)
5455 struct dp_netdev_pmd_thread
*pmd
= f_
;
5456 struct pmd_perf_stats
*s
= &pmd
->perf_stats
;
5457 unsigned int lc
= 0;
5458 struct polled_queue
*poll_list
;
5459 bool wait_for_reload
= false;
5465 int process_packets
= 0;
5469 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
5470 ovsthread_setspecific(pmd
->dp
->per_pmd_key
, pmd
);
5471 ovs_numa_thread_setaffinity_core(pmd
->core_id
);
5472 dpdk_set_lcore_id(pmd
->core_id
);
5473 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
5474 dfc_cache_init(&pmd
->flow_cache
);
5475 pmd_alloc_static_tx_qid(pmd
);
5478 atomic_count_init(&pmd
->pmd_overloaded
, 0);
5480 /* List port/core affinity */
5481 for (i
= 0; i
< poll_cnt
; i
++) {
5482 VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n",
5483 pmd
->core_id
, netdev_rxq_get_name(poll_list
[i
].rxq
->rx
),
5484 netdev_rxq_get_queue_id(poll_list
[i
].rxq
->rx
));
5485 /* Reset the rxq current cycles counter. */
5486 dp_netdev_rxq_set_cycles(poll_list
[i
].rxq
, RXQ_CYCLES_PROC_CURR
, 0);
5490 if (wait_for_reload
) {
5491 /* Don't sleep, control thread will ask for a reload shortly. */
5493 atomic_read_explicit(&pmd
->reload
, &reload
,
5494 memory_order_acquire
);
5497 while (seq_read(pmd
->reload_seq
) == pmd
->last_reload_seq
) {
5498 seq_wait(pmd
->reload_seq
, pmd
->last_reload_seq
);
5504 pmd
->intrvl_tsc_prev
= 0;
5505 atomic_store_relaxed(&pmd
->intrvl_cycles
, 0);
5506 cycles_counter_update(s
);
5507 /* Protect pmd stats from external clearing while polling. */
5508 ovs_mutex_lock(&pmd
->perf_stats
.stats_mutex
);
5510 uint64_t rx_packets
= 0, tx_packets
= 0;
5512 pmd_perf_start_iteration(s
);
5514 for (i
= 0; i
< poll_cnt
; i
++) {
5516 if (!poll_list
[i
].rxq_enabled
) {
5520 if (poll_list
[i
].emc_enabled
) {
5521 atomic_read_relaxed(&pmd
->dp
->emc_insert_min
,
5522 &pmd
->ctx
.emc_insert_min
);
5524 pmd
->ctx
.emc_insert_min
= 0;
5528 dp_netdev_process_rxq_port(pmd
, poll_list
[i
].rxq
,
5529 poll_list
[i
].port_no
);
5530 rx_packets
+= process_packets
;
5534 /* We didn't receive anything in the process loop.
5535 * Check if we need to send something.
5536 * There was no time updates on current iteration. */
5537 pmd_thread_ctx_time_update(pmd
);
5538 tx_packets
= dp_netdev_pmd_flush_output_packets(pmd
, false);
5544 coverage_try_clear();
5545 dp_netdev_pmd_try_optimize(pmd
, poll_list
, poll_cnt
);
5546 if (!ovsrcu_try_quiesce()) {
5547 emc_cache_slow_sweep(&((pmd
->flow_cache
).emc_cache
));
5550 for (i
= 0; i
< poll_cnt
; i
++) {
5551 uint64_t current_seq
=
5552 netdev_get_change_seq(poll_list
[i
].rxq
->port
->netdev
);
5553 if (poll_list
[i
].change_seq
!= current_seq
) {
5554 poll_list
[i
].change_seq
= current_seq
;
5555 poll_list
[i
].rxq_enabled
=
5556 netdev_rxq_enabled(poll_list
[i
].rxq
->rx
);
5561 atomic_read_explicit(&pmd
->reload
, &reload
, memory_order_acquire
);
5562 if (OVS_UNLIKELY(reload
)) {
5566 pmd_perf_end_iteration(s
, rx_packets
, tx_packets
,
5567 pmd_perf_metrics_enabled(pmd
));
5569 ovs_mutex_unlock(&pmd
->perf_stats
.stats_mutex
);
5571 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
5572 atomic_read_relaxed(&pmd
->wait_for_reload
, &wait_for_reload
);
5573 atomic_read_relaxed(&pmd
->reload_tx_qid
, &reload_tx_qid
);
5574 atomic_read_relaxed(&pmd
->exit
, &exiting
);
5575 /* Signal here to make sure the pmd finishes
5576 * reloading the updated configuration. */
5577 dp_netdev_pmd_reload_done(pmd
);
5579 if (reload_tx_qid
) {
5580 pmd_free_static_tx_qid(pmd
);
5581 pmd_alloc_static_tx_qid(pmd
);
5588 pmd_free_static_tx_qid(pmd
);
5589 dfc_cache_uninit(&pmd
->flow_cache
);
5591 pmd_free_cached_ports(pmd
);
5596 dp_netdev_disable_upcall(struct dp_netdev
*dp
)
5597 OVS_ACQUIRES(dp
->upcall_rwlock
)
5599 fat_rwlock_wrlock(&dp
->upcall_rwlock
);
5605 dpif_netdev_meter_get_features(const struct dpif
* dpif OVS_UNUSED
,
5606 struct ofputil_meter_features
*features
)
5608 features
->max_meters
= MAX_METERS
;
5609 features
->band_types
= DP_SUPPORTED_METER_BAND_TYPES
;
5610 features
->capabilities
= DP_SUPPORTED_METER_FLAGS_MASK
;
5611 features
->max_bands
= MAX_BANDS
;
5612 features
->max_color
= 0;
5615 /* Applies the meter identified by 'meter_id' to 'packets_'. Packets
5616 * that exceed a band are dropped in-place. */
5618 dp_netdev_run_meter(struct dp_netdev
*dp
, struct dp_packet_batch
*packets_
,
5619 uint32_t meter_id
, long long int now
)
5621 struct dp_meter
*meter
;
5622 struct dp_meter_band
*band
;
5623 struct dp_packet
*packet
;
5624 long long int long_delta_t
; /* msec */
5625 uint32_t delta_t
; /* msec */
5626 const size_t cnt
= dp_packet_batch_size(packets_
);
5627 uint32_t bytes
, volume
;
5628 int exceeded_band
[NETDEV_MAX_BURST
];
5629 uint32_t exceeded_rate
[NETDEV_MAX_BURST
];
5630 int exceeded_pkt
= cnt
; /* First packet that exceeded a band rate. */
5632 if (meter_id
>= MAX_METERS
) {
5636 meter_lock(dp
, meter_id
);
5637 meter
= dp
->meters
[meter_id
];
5642 /* Initialize as negative values. */
5643 memset(exceeded_band
, 0xff, cnt
* sizeof *exceeded_band
);
5644 /* Initialize as zeroes. */
5645 memset(exceeded_rate
, 0, cnt
* sizeof *exceeded_rate
);
5647 /* All packets will hit the meter at the same time. */
5648 long_delta_t
= now
/ 1000 - meter
->used
/ 1000; /* msec */
5650 /* Make sure delta_t will not be too large, so that bucket will not
5651 * wrap around below. */
5652 delta_t
= (long_delta_t
> (long long int)meter
->max_delta_t
)
5653 ? meter
->max_delta_t
: (uint32_t)long_delta_t
;
5655 /* Update meter stats. */
5657 meter
->packet_count
+= cnt
;
5659 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
5660 bytes
+= dp_packet_size(packet
);
5662 meter
->byte_count
+= bytes
;
5664 /* Meters can operate in terms of packets per second or kilobits per
5666 if (meter
->flags
& OFPMF13_PKTPS
) {
5667 /* Rate in packets/second, bucket 1/1000 packets. */
5668 /* msec * packets/sec = 1/1000 packets. */
5669 volume
= cnt
* 1000; /* Take 'cnt' packets from the bucket. */
5671 /* Rate in kbps, bucket in bits. */
5672 /* msec * kbps = bits */
5676 /* Update all bands and find the one hit with the highest rate for each
5677 * packet (if any). */
5678 for (int m
= 0; m
< meter
->n_bands
; ++m
) {
5679 band
= &meter
->bands
[m
];
5681 /* Update band's bucket. */
5682 band
->bucket
+= delta_t
* band
->up
.rate
;
5683 if (band
->bucket
> band
->up
.burst_size
) {
5684 band
->bucket
= band
->up
.burst_size
;
5687 /* Drain the bucket for all the packets, if possible. */
5688 if (band
->bucket
>= volume
) {
5689 band
->bucket
-= volume
;
5691 int band_exceeded_pkt
;
5693 /* Band limit hit, must process packet-by-packet. */
5694 if (meter
->flags
& OFPMF13_PKTPS
) {
5695 band_exceeded_pkt
= band
->bucket
/ 1000;
5696 band
->bucket
%= 1000; /* Remainder stays in bucket. */
5698 /* Update the exceeding band for each exceeding packet.
5699 * (Only one band will be fired by a packet, and that
5700 * can be different for each packet.) */
5701 for (int i
= band_exceeded_pkt
; i
< cnt
; i
++) {
5702 if (band
->up
.rate
> exceeded_rate
[i
]) {
5703 exceeded_rate
[i
] = band
->up
.rate
;
5704 exceeded_band
[i
] = m
;
5708 /* Packet sizes differ, must process one-by-one. */
5709 band_exceeded_pkt
= cnt
;
5710 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
5711 uint32_t bits
= dp_packet_size(packet
) * 8;
5713 if (band
->bucket
>= bits
) {
5714 band
->bucket
-= bits
;
5716 if (i
< band_exceeded_pkt
) {
5717 band_exceeded_pkt
= i
;
5719 /* Update the exceeding band for the exceeding packet.
5720 * (Only one band will be fired by a packet, and that
5721 * can be different for each packet.) */
5722 if (band
->up
.rate
> exceeded_rate
[i
]) {
5723 exceeded_rate
[i
] = band
->up
.rate
;
5724 exceeded_band
[i
] = m
;
5729 /* Remember the first exceeding packet. */
5730 if (exceeded_pkt
> band_exceeded_pkt
) {
5731 exceeded_pkt
= band_exceeded_pkt
;
5736 /* Fire the highest rate band exceeded by each packet, and drop
5737 * packets if needed. */
5739 DP_PACKET_BATCH_REFILL_FOR_EACH (j
, cnt
, packet
, packets_
) {
5740 if (exceeded_band
[j
] >= 0) {
5741 /* Meter drop packet. */
5742 band
= &meter
->bands
[exceeded_band
[j
]];
5743 band
->packet_count
+= 1;
5744 band
->byte_count
+= dp_packet_size(packet
);
5746 dp_packet_delete(packet
);
5748 /* Meter accepts packet. */
5749 dp_packet_batch_refill(packets_
, packet
, j
);
5753 meter_unlock(dp
, meter_id
);
5756 /* Meter set/get/del processing is still single-threaded. */
5758 dpif_netdev_meter_set(struct dpif
*dpif
, ofproto_meter_id meter_id
,
5759 struct ofputil_meter_config
*config
)
5761 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5762 uint32_t mid
= meter_id
.uint32
;
5763 struct dp_meter
*meter
;
5766 if (mid
>= MAX_METERS
) {
5767 return EFBIG
; /* Meter_id out of range. */
5770 if (config
->flags
& ~DP_SUPPORTED_METER_FLAGS_MASK
) {
5771 return EBADF
; /* Unsupported flags set */
5774 if (config
->n_bands
> MAX_BANDS
) {
5778 for (i
= 0; i
< config
->n_bands
; ++i
) {
5779 switch (config
->bands
[i
].type
) {
5783 return ENODEV
; /* Unsupported band type */
5787 /* Allocate meter */
5788 meter
= xzalloc(sizeof *meter
5789 + config
->n_bands
* sizeof(struct dp_meter_band
));
5791 meter
->flags
= config
->flags
;
5792 meter
->n_bands
= config
->n_bands
;
5793 meter
->max_delta_t
= 0;
5794 meter
->used
= time_usec();
5797 for (i
= 0; i
< config
->n_bands
; ++i
) {
5798 uint32_t band_max_delta_t
;
5800 /* Set burst size to a workable value if none specified. */
5801 if (config
->bands
[i
].burst_size
== 0) {
5802 config
->bands
[i
].burst_size
= config
->bands
[i
].rate
;
5805 meter
->bands
[i
].up
= config
->bands
[i
];
5806 /* Convert burst size to the bucket units: */
5807 /* pkts => 1/1000 packets, kilobits => bits. */
5808 meter
->bands
[i
].up
.burst_size
*= 1000;
5809 /* Initialize bucket to empty. */
5810 meter
->bands
[i
].bucket
= 0;
5812 /* Figure out max delta_t that is enough to fill any bucket. */
5814 = meter
->bands
[i
].up
.burst_size
/ meter
->bands
[i
].up
.rate
;
5815 if (band_max_delta_t
> meter
->max_delta_t
) {
5816 meter
->max_delta_t
= band_max_delta_t
;
5820 meter_lock(dp
, mid
);
5821 dp_delete_meter(dp
, mid
); /* Free existing meter, if any */
5822 dp
->meters
[mid
] = meter
;
5823 meter_unlock(dp
, mid
);
5829 dpif_netdev_meter_get(const struct dpif
*dpif
,
5830 ofproto_meter_id meter_id_
,
5831 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
5833 const struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5834 uint32_t meter_id
= meter_id_
.uint32
;
5837 if (meter_id
>= MAX_METERS
) {
5841 meter_lock(dp
, meter_id
);
5842 const struct dp_meter
*meter
= dp
->meters
[meter_id
];
5850 stats
->packet_in_count
= meter
->packet_count
;
5851 stats
->byte_in_count
= meter
->byte_count
;
5853 for (i
= 0; i
< n_bands
&& i
< meter
->n_bands
; ++i
) {
5854 stats
->bands
[i
].packet_count
= meter
->bands
[i
].packet_count
;
5855 stats
->bands
[i
].byte_count
= meter
->bands
[i
].byte_count
;
5862 meter_unlock(dp
, meter_id
);
5867 dpif_netdev_meter_del(struct dpif
*dpif
,
5868 ofproto_meter_id meter_id_
,
5869 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
5871 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5874 error
= dpif_netdev_meter_get(dpif
, meter_id_
, stats
, n_bands
);
5876 uint32_t meter_id
= meter_id_
.uint32
;
5878 meter_lock(dp
, meter_id
);
5879 dp_delete_meter(dp
, meter_id
);
5880 meter_unlock(dp
, meter_id
);
5887 dpif_netdev_disable_upcall(struct dpif
*dpif
)
5888 OVS_NO_THREAD_SAFETY_ANALYSIS
5890 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5891 dp_netdev_disable_upcall(dp
);
5895 dp_netdev_enable_upcall(struct dp_netdev
*dp
)
5896 OVS_RELEASES(dp
->upcall_rwlock
)
5898 fat_rwlock_unlock(&dp
->upcall_rwlock
);
5902 dpif_netdev_enable_upcall(struct dpif
*dpif
)
5903 OVS_NO_THREAD_SAFETY_ANALYSIS
5905 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5906 dp_netdev_enable_upcall(dp
);
5910 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
)
5912 atomic_store_relaxed(&pmd
->wait_for_reload
, false);
5913 atomic_store_relaxed(&pmd
->reload_tx_qid
, false);
5914 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
5915 atomic_store_explicit(&pmd
->reload
, false, memory_order_release
);
5918 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
5919 * the pointer if succeeds, otherwise, NULL (it can return NULL even if
5920 * 'core_id' is NON_PMD_CORE_ID).
5922 * Caller must unrefs the returned reference. */
5923 static struct dp_netdev_pmd_thread
*
5924 dp_netdev_get_pmd(struct dp_netdev
*dp
, unsigned core_id
)
5926 struct dp_netdev_pmd_thread
*pmd
;
5927 const struct cmap_node
*pnode
;
5929 pnode
= cmap_find(&dp
->poll_threads
, hash_int(core_id
, 0));
5933 pmd
= CONTAINER_OF(pnode
, struct dp_netdev_pmd_thread
, node
);
5935 return dp_netdev_pmd_try_ref(pmd
) ? pmd
: NULL
;
5938 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
5940 dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
5941 OVS_REQUIRES(dp
->port_mutex
)
5943 struct dp_netdev_pmd_thread
*non_pmd
;
5945 non_pmd
= xzalloc(sizeof *non_pmd
);
5946 dp_netdev_configure_pmd(non_pmd
, dp
, NON_PMD_CORE_ID
, OVS_NUMA_UNSPEC
);
5949 /* Caller must have valid pointer to 'pmd'. */
5951 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
)
5953 return ovs_refcount_try_ref_rcu(&pmd
->ref_cnt
);
5957 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
)
5959 if (pmd
&& ovs_refcount_unref(&pmd
->ref_cnt
) == 1) {
5960 ovsrcu_postpone(dp_netdev_destroy_pmd
, pmd
);
5964 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
5965 * fails, keeps checking for next node until reaching the end of cmap.
5967 * Caller must unrefs the returned reference. */
5968 static struct dp_netdev_pmd_thread
*
5969 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
)
5971 struct dp_netdev_pmd_thread
*next
;
5974 struct cmap_node
*node
;
5976 node
= cmap_next_position(&dp
->poll_threads
, pos
);
5977 next
= node
? CONTAINER_OF(node
, struct dp_netdev_pmd_thread
, node
)
5979 } while (next
&& !dp_netdev_pmd_try_ref(next
));
5984 /* Configures the 'pmd' based on the input argument. */
5986 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
, struct dp_netdev
*dp
,
5987 unsigned core_id
, int numa_id
)
5990 pmd
->core_id
= core_id
;
5991 pmd
->numa_id
= numa_id
;
5992 pmd
->need_reload
= false;
5993 pmd
->n_output_batches
= 0;
5995 ovs_refcount_init(&pmd
->ref_cnt
);
5996 atomic_init(&pmd
->exit
, false);
5997 pmd
->reload_seq
= seq_create();
5998 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
5999 atomic_init(&pmd
->reload
, false);
6000 ovs_mutex_init(&pmd
->flow_mutex
);
6001 ovs_mutex_init(&pmd
->port_mutex
);
6002 cmap_init(&pmd
->flow_table
);
6003 cmap_init(&pmd
->classifiers
);
6004 pmd
->ctx
.last_rxq
= NULL
;
6005 pmd_thread_ctx_time_update(pmd
);
6006 pmd
->next_optimization
= pmd
->ctx
.now
+ DPCLS_OPTIMIZATION_INTERVAL
;
6007 pmd
->rxq_next_cycle_store
= pmd
->ctx
.now
+ PMD_RXQ_INTERVAL_LEN
;
6008 hmap_init(&pmd
->poll_list
);
6009 hmap_init(&pmd
->tx_ports
);
6010 hmap_init(&pmd
->tnl_port_cache
);
6011 hmap_init(&pmd
->send_port_cache
);
6012 /* init the 'flow_cache' since there is no
6013 * actual thread created for NON_PMD_CORE_ID. */
6014 if (core_id
== NON_PMD_CORE_ID
) {
6015 dfc_cache_init(&pmd
->flow_cache
);
6016 pmd_alloc_static_tx_qid(pmd
);
6018 pmd_perf_stats_init(&pmd
->perf_stats
);
6019 cmap_insert(&dp
->poll_threads
, CONST_CAST(struct cmap_node
*, &pmd
->node
),
6020 hash_int(core_id
, 0));
6024 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
)
6028 dp_netdev_pmd_flow_flush(pmd
);
6029 hmap_destroy(&pmd
->send_port_cache
);
6030 hmap_destroy(&pmd
->tnl_port_cache
);
6031 hmap_destroy(&pmd
->tx_ports
);
6032 hmap_destroy(&pmd
->poll_list
);
6033 /* All flows (including their dpcls_rules) have been deleted already */
6034 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
6036 ovsrcu_postpone(free
, cls
);
6038 cmap_destroy(&pmd
->classifiers
);
6039 cmap_destroy(&pmd
->flow_table
);
6040 ovs_mutex_destroy(&pmd
->flow_mutex
);
6041 seq_destroy(pmd
->reload_seq
);
6042 ovs_mutex_destroy(&pmd
->port_mutex
);
6046 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
6047 * and unrefs the struct. */
6049 dp_netdev_del_pmd(struct dp_netdev
*dp
, struct dp_netdev_pmd_thread
*pmd
)
6051 /* NON_PMD_CORE_ID doesn't have a thread, so we don't have to synchronize,
6052 * but extra cleanup is necessary */
6053 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
6054 ovs_mutex_lock(&dp
->non_pmd_mutex
);
6055 dfc_cache_uninit(&pmd
->flow_cache
);
6056 pmd_free_cached_ports(pmd
);
6057 pmd_free_static_tx_qid(pmd
);
6058 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
6060 atomic_store_relaxed(&pmd
->exit
, true);
6061 dp_netdev_reload_pmd__(pmd
);
6062 xpthread_join(pmd
->thread
, NULL
);
6065 dp_netdev_pmd_clear_ports(pmd
);
6067 /* Purges the 'pmd''s flows after stopping the thread, but before
6068 * destroying the flows, so that the flow stats can be collected. */
6069 if (dp
->dp_purge_cb
) {
6070 dp
->dp_purge_cb(dp
->dp_purge_aux
, pmd
->core_id
);
6072 cmap_remove(&pmd
->dp
->poll_threads
, &pmd
->node
, hash_int(pmd
->core_id
, 0));
6073 dp_netdev_pmd_unref(pmd
);
6076 /* Destroys all pmd threads. If 'non_pmd' is true it also destroys the non pmd
6079 dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
, bool non_pmd
)
6081 struct dp_netdev_pmd_thread
*pmd
;
6082 struct dp_netdev_pmd_thread
**pmd_list
;
6083 size_t k
= 0, n_pmds
;
6085 n_pmds
= cmap_count(&dp
->poll_threads
);
6086 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
6088 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
6089 if (!non_pmd
&& pmd
->core_id
== NON_PMD_CORE_ID
) {
6092 /* We cannot call dp_netdev_del_pmd(), since it alters
6093 * 'dp->poll_threads' (while we're iterating it) and it
6095 ovs_assert(k
< n_pmds
);
6096 pmd_list
[k
++] = pmd
;
6099 for (size_t i
= 0; i
< k
; i
++) {
6100 dp_netdev_del_pmd(dp
, pmd_list
[i
]);
6105 /* Deletes all rx queues from pmd->poll_list and all the ports from
6108 dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
)
6110 struct rxq_poll
*poll
;
6111 struct tx_port
*port
;
6113 ovs_mutex_lock(&pmd
->port_mutex
);
6114 HMAP_FOR_EACH_POP (poll
, node
, &pmd
->poll_list
) {
6117 HMAP_FOR_EACH_POP (port
, node
, &pmd
->tx_ports
) {
6120 ovs_mutex_unlock(&pmd
->port_mutex
);
6123 /* Adds rx queue to poll_list of PMD thread, if it's not there already. */
6125 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
6126 struct dp_netdev_rxq
*rxq
)
6127 OVS_REQUIRES(pmd
->port_mutex
)
6129 int qid
= netdev_rxq_get_queue_id(rxq
->rx
);
6130 uint32_t hash
= hash_2words(odp_to_u32(rxq
->port
->port_no
), qid
);
6131 struct rxq_poll
*poll
;
6133 HMAP_FOR_EACH_WITH_HASH (poll
, node
, hash
, &pmd
->poll_list
) {
6134 if (poll
->rxq
== rxq
) {
6135 /* 'rxq' is already polled by this thread. Do nothing. */
6140 poll
= xmalloc(sizeof *poll
);
6142 hmap_insert(&pmd
->poll_list
, &poll
->node
, hash
);
6144 pmd
->need_reload
= true;
6147 /* Delete 'poll' from poll_list of PMD thread. */
6149 dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
6150 struct rxq_poll
*poll
)
6151 OVS_REQUIRES(pmd
->port_mutex
)
6153 hmap_remove(&pmd
->poll_list
, &poll
->node
);
6156 pmd
->need_reload
= true;
6159 /* Add 'port' to the tx port cache of 'pmd', which must be reloaded for the
6160 * changes to take effect. */
6162 dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
6163 struct dp_netdev_port
*port
)
6164 OVS_REQUIRES(pmd
->port_mutex
)
6168 tx
= tx_port_lookup(&pmd
->tx_ports
, port
->port_no
);
6170 /* 'port' is already on this thread tx cache. Do nothing. */
6174 tx
= xzalloc(sizeof *tx
);
6178 tx
->flush_time
= 0LL;
6179 dp_packet_batch_init(&tx
->output_pkts
);
6181 hmap_insert(&pmd
->tx_ports
, &tx
->node
, hash_port_no(tx
->port
->port_no
));
6182 pmd
->need_reload
= true;
6185 /* Del 'tx' from the tx port cache of 'pmd', which must be reloaded for the
6186 * changes to take effect. */
6188 dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
6190 OVS_REQUIRES(pmd
->port_mutex
)
6192 hmap_remove(&pmd
->tx_ports
, &tx
->node
);
6194 pmd
->need_reload
= true;
6198 dpif_netdev_get_datapath_version(void)
6200 return xstrdup("<built-in>");
6204 dp_netdev_flow_used(struct dp_netdev_flow
*netdev_flow
, int cnt
, int size
,
6205 uint16_t tcp_flags
, long long now
)
6209 atomic_store_relaxed(&netdev_flow
->stats
.used
, now
);
6210 non_atomic_ullong_add(&netdev_flow
->stats
.packet_count
, cnt
);
6211 non_atomic_ullong_add(&netdev_flow
->stats
.byte_count
, size
);
6212 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
6214 atomic_store_relaxed(&netdev_flow
->stats
.tcp_flags
, flags
);
6218 dp_netdev_upcall(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
*packet_
,
6219 struct flow
*flow
, struct flow_wildcards
*wc
, ovs_u128
*ufid
,
6220 enum dpif_upcall_type type
, const struct nlattr
*userdata
,
6221 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
6223 struct dp_netdev
*dp
= pmd
->dp
;
6225 if (OVS_UNLIKELY(!dp
->upcall_cb
)) {
6229 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl
))) {
6230 struct ds ds
= DS_EMPTY_INITIALIZER
;
6233 struct odp_flow_key_parms odp_parms
= {
6235 .mask
= wc
? &wc
->masks
: NULL
,
6236 .support
= dp_netdev_support
,
6239 ofpbuf_init(&key
, 0);
6240 odp_flow_key_from_flow(&odp_parms
, &key
);
6241 packet_str
= ofp_dp_packet_to_string(packet_
);
6243 odp_flow_key_format(key
.data
, key
.size
, &ds
);
6245 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp
->name
,
6246 dpif_upcall_type_to_string(type
), ds_cstr(&ds
), packet_str
);
6248 ofpbuf_uninit(&key
);
6254 return dp
->upcall_cb(packet_
, flow
, ufid
, pmd
->core_id
, type
, userdata
,
6255 actions
, wc
, put_actions
, dp
->upcall_aux
);
6258 static inline uint32_t
6259 dpif_netdev_packet_get_rss_hash_orig_pkt(struct dp_packet
*packet
,
6260 const struct miniflow
*mf
)
6264 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
6265 hash
= dp_packet_get_rss_hash(packet
);
6267 hash
= miniflow_hash_5tuple(mf
, 0);
6268 dp_packet_set_rss_hash(packet
, hash
);
6274 static inline uint32_t
6275 dpif_netdev_packet_get_rss_hash(struct dp_packet
*packet
,
6276 const struct miniflow
*mf
)
6278 uint32_t hash
, recirc_depth
;
6280 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
6281 hash
= dp_packet_get_rss_hash(packet
);
6283 hash
= miniflow_hash_5tuple(mf
, 0);
6284 dp_packet_set_rss_hash(packet
, hash
);
6287 /* The RSS hash must account for the recirculation depth to avoid
6288 * collisions in the exact match cache */
6289 recirc_depth
= *recirc_depth_get_unsafe();
6290 if (OVS_UNLIKELY(recirc_depth
)) {
6291 hash
= hash_finish(hash
, recirc_depth
);
6292 dp_packet_set_rss_hash(packet
, hash
);
6297 struct packet_batch_per_flow
{
6298 unsigned int byte_count
;
6300 struct dp_netdev_flow
*flow
;
6302 struct dp_packet_batch array
;
6306 packet_batch_per_flow_update(struct packet_batch_per_flow
*batch
,
6307 struct dp_packet
*packet
,
6310 batch
->byte_count
+= dp_packet_size(packet
);
6311 batch
->tcp_flags
|= tcp_flags
;
6312 batch
->array
.packets
[batch
->array
.count
++] = packet
;
6316 packet_batch_per_flow_init(struct packet_batch_per_flow
*batch
,
6317 struct dp_netdev_flow
*flow
)
6319 flow
->batch
= batch
;
6322 dp_packet_batch_init(&batch
->array
);
6323 batch
->byte_count
= 0;
6324 batch
->tcp_flags
= 0;
6328 packet_batch_per_flow_execute(struct packet_batch_per_flow
*batch
,
6329 struct dp_netdev_pmd_thread
*pmd
)
6331 struct dp_netdev_actions
*actions
;
6332 struct dp_netdev_flow
*flow
= batch
->flow
;
6334 dp_netdev_flow_used(flow
, batch
->array
.count
, batch
->byte_count
,
6335 batch
->tcp_flags
, pmd
->ctx
.now
/ 1000);
6337 actions
= dp_netdev_flow_get_actions(flow
);
6339 dp_netdev_execute_actions(pmd
, &batch
->array
, true, &flow
->flow
,
6340 actions
->actions
, actions
->size
);
6344 dp_netdev_queue_batches(struct dp_packet
*pkt
,
6345 struct dp_netdev_flow
*flow
, uint16_t tcp_flags
,
6346 struct packet_batch_per_flow
*batches
,
6349 struct packet_batch_per_flow
*batch
= flow
->batch
;
6351 if (OVS_UNLIKELY(!batch
)) {
6352 batch
= &batches
[(*n_batches
)++];
6353 packet_batch_per_flow_init(batch
, flow
);
6356 packet_batch_per_flow_update(batch
, pkt
, tcp_flags
);
6360 packet_enqueue_to_flow_map(struct dp_packet
*packet
,
6361 struct dp_netdev_flow
*flow
,
6363 struct dp_packet_flow_map
*flow_map
,
6366 struct dp_packet_flow_map
*map
= &flow_map
[index
];
6368 map
->packet
= packet
;
6369 map
->tcp_flags
= tcp_flags
;
6372 /* SMC lookup function for a batch of packets.
6373 * By doing batching SMC lookup, we can use prefetch
6374 * to hide memory access latency.
6377 smc_lookup_batch(struct dp_netdev_pmd_thread
*pmd
,
6378 struct netdev_flow_key
*keys
,
6379 struct netdev_flow_key
**missed_keys
,
6380 struct dp_packet_batch
*packets_
,
6382 struct dp_packet_flow_map
*flow_map
,
6386 struct dp_packet
*packet
;
6387 size_t n_smc_hit
= 0, n_missed
= 0;
6388 struct dfc_cache
*cache
= &pmd
->flow_cache
;
6389 struct smc_cache
*smc_cache
= &cache
->smc_cache
;
6390 const struct cmap_node
*flow_node
;
6394 /* Prefetch buckets for all packets */
6395 for (i
= 0; i
< cnt
; i
++) {
6396 OVS_PREFETCH(&smc_cache
->buckets
[keys
[i
].hash
& SMC_MASK
]);
6399 DP_PACKET_BATCH_REFILL_FOR_EACH (i
, cnt
, packet
, packets_
) {
6400 struct dp_netdev_flow
*flow
= NULL
;
6401 flow_node
= smc_entry_get(pmd
, keys
[i
].hash
);
6403 /* Get the original order of this packet in received batch. */
6404 recv_idx
= index_map
[i
];
6406 if (OVS_LIKELY(flow_node
!= NULL
)) {
6407 CMAP_NODE_FOR_EACH (flow
, node
, flow_node
) {
6408 /* Since we dont have per-port megaflow to check the port
6409 * number, we need to verify that the input ports match. */
6410 if (OVS_LIKELY(dpcls_rule_matches_key(&flow
->cr
, &keys
[i
]) &&
6411 flow
->flow
.in_port
.odp_port
== packet
->md
.in_port
.odp_port
)) {
6412 tcp_flags
= miniflow_get_tcp_flags(&keys
[i
].mf
);
6414 /* SMC hit and emc miss, we insert into EMC */
6416 netdev_flow_key_size(miniflow_n_values(&keys
[i
].mf
));
6417 emc_probabilistic_insert(pmd
, &keys
[i
], flow
);
6418 /* Add these packets into the flow map in the same order
6421 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6422 flow_map
, recv_idx
);
6433 /* SMC missed. Group missed packets together at
6434 * the beginning of the 'packets' array. */
6435 dp_packet_batch_refill(packets_
, packet
, i
);
6437 /* Preserve the order of packet for flow batching. */
6438 index_map
[n_missed
] = recv_idx
;
6440 /* Put missed keys to the pointer arrays return to the caller */
6441 missed_keys
[n_missed
++] = &keys
[i
];
6444 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_SMC_HIT
, n_smc_hit
);
6447 /* Try to process all ('cnt') the 'packets' using only the datapath flow cache
6448 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
6449 * miniflow is copied into 'keys' and the packet pointer is moved at the
6450 * beginning of the 'packets' array. The pointers of missed keys are put in the
6451 * missed_keys pointer array for future processing.
6453 * The function returns the number of packets that needs to be processed in the
6454 * 'packets' array (they have been moved to the beginning of the vector).
6456 * For performance reasons a caller may choose not to initialize the metadata
6457 * in 'packets_'. If 'md_is_valid' is false, the metadata in 'packets'
6458 * is not valid and must be initialized by this function using 'port_no'.
6459 * If 'md_is_valid' is true, the metadata is already valid and 'port_no'
6462 static inline size_t
6463 dfc_processing(struct dp_netdev_pmd_thread
*pmd
,
6464 struct dp_packet_batch
*packets_
,
6465 struct netdev_flow_key
*keys
,
6466 struct netdev_flow_key
**missed_keys
,
6467 struct packet_batch_per_flow batches
[], size_t *n_batches
,
6468 struct dp_packet_flow_map
*flow_map
,
6469 size_t *n_flows
, uint8_t *index_map
,
6470 bool md_is_valid
, odp_port_t port_no
)
6472 struct netdev_flow_key
*key
= &keys
[0];
6473 size_t n_missed
= 0, n_emc_hit
= 0;
6474 struct dfc_cache
*cache
= &pmd
->flow_cache
;
6475 struct dp_packet
*packet
;
6476 const size_t cnt
= dp_packet_batch_size(packets_
);
6477 uint32_t cur_min
= pmd
->ctx
.emc_insert_min
;
6482 bool batch_enable
= true;
6484 atomic_read_relaxed(&pmd
->dp
->smc_enable_db
, &smc_enable_db
);
6485 pmd_perf_update_counter(&pmd
->perf_stats
,
6486 md_is_valid
? PMD_STAT_RECIRC
: PMD_STAT_RECV
,
6489 DP_PACKET_BATCH_REFILL_FOR_EACH (i
, cnt
, packet
, packets_
) {
6490 struct dp_netdev_flow
*flow
;
6493 if (OVS_UNLIKELY(dp_packet_size(packet
) < ETH_HEADER_LEN
)) {
6494 dp_packet_delete(packet
);
6499 struct dp_packet
**packets
= packets_
->packets
;
6500 /* Prefetch next packet data and metadata. */
6501 OVS_PREFETCH(dp_packet_data(packets
[i
+1]));
6502 pkt_metadata_prefetch_init(&packets
[i
+1]->md
);
6506 pkt_metadata_init(&packet
->md
, port_no
);
6509 if ((*recirc_depth_get() == 0) &&
6510 dp_packet_has_flow_mark(packet
, &mark
)) {
6511 flow
= mark_to_flow_find(pmd
, mark
);
6512 if (OVS_LIKELY(flow
)) {
6513 tcp_flags
= parse_tcp_flags(packet
);
6514 if (OVS_LIKELY(batch_enable
)) {
6515 dp_netdev_queue_batches(packet
, flow
, tcp_flags
, batches
,
6518 /* Flow batching should be performed only after fast-path
6519 * processing is also completed for packets with emc miss
6520 * or else it will result in reordering of packets with
6521 * same datapath flows. */
6522 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6523 flow_map
, map_cnt
++);
6529 miniflow_extract(packet
, &key
->mf
);
6530 key
->len
= 0; /* Not computed yet. */
6532 (md_is_valid
== false)
6533 ? dpif_netdev_packet_get_rss_hash_orig_pkt(packet
, &key
->mf
)
6534 : dpif_netdev_packet_get_rss_hash(packet
, &key
->mf
);
6536 /* If EMC is disabled skip emc_lookup */
6537 flow
= (cur_min
!= 0) ? emc_lookup(&cache
->emc_cache
, key
) : NULL
;
6538 if (OVS_LIKELY(flow
)) {
6539 tcp_flags
= miniflow_get_tcp_flags(&key
->mf
);
6541 if (OVS_LIKELY(batch_enable
)) {
6542 dp_netdev_queue_batches(packet
, flow
, tcp_flags
, batches
,
6545 /* Flow batching should be performed only after fast-path
6546 * processing is also completed for packets with emc miss
6547 * or else it will result in reordering of packets with
6548 * same datapath flows. */
6549 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6550 flow_map
, map_cnt
++);
6553 /* Exact match cache missed. Group missed packets together at
6554 * the beginning of the 'packets' array. */
6555 dp_packet_batch_refill(packets_
, packet
, i
);
6557 /* Preserve the order of packet for flow batching. */
6558 index_map
[n_missed
] = map_cnt
;
6559 flow_map
[map_cnt
++].flow
= NULL
;
6561 /* 'key[n_missed]' contains the key of the current packet and it
6562 * will be passed to SMC lookup. The next key should be extracted
6563 * to 'keys[n_missed + 1]'.
6564 * We also maintain a pointer array to keys missed both SMC and EMC
6565 * which will be returned to the caller for future processing. */
6566 missed_keys
[n_missed
] = key
;
6567 key
= &keys
[++n_missed
];
6569 /* Skip batching for subsequent packets to avoid reordering. */
6570 batch_enable
= false;
6573 /* Count of packets which are not flow batched. */
6576 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_EXACT_HIT
, n_emc_hit
);
6578 if (!smc_enable_db
) {
6579 return dp_packet_batch_size(packets_
);
6582 /* Packets miss EMC will do a batch lookup in SMC if enabled */
6583 smc_lookup_batch(pmd
, keys
, missed_keys
, packets_
,
6584 n_missed
, flow_map
, index_map
);
6586 return dp_packet_batch_size(packets_
);
6590 handle_packet_upcall(struct dp_netdev_pmd_thread
*pmd
,
6591 struct dp_packet
*packet
,
6592 const struct netdev_flow_key
*key
,
6593 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
6595 struct ofpbuf
*add_actions
;
6596 struct dp_packet_batch b
;
6600 uint64_t cycles
= cycles_counter_update(&pmd
->perf_stats
);
6602 match
.tun_md
.valid
= false;
6603 miniflow_expand(&key
->mf
, &match
.flow
);
6605 ofpbuf_clear(actions
);
6606 ofpbuf_clear(put_actions
);
6608 dpif_flow_hash(pmd
->dp
->dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
6609 error
= dp_netdev_upcall(pmd
, packet
, &match
.flow
, &match
.wc
,
6610 &ufid
, DPIF_UC_MISS
, NULL
, actions
,
6612 if (OVS_UNLIKELY(error
&& error
!= ENOSPC
)) {
6613 dp_packet_delete(packet
);
6617 /* The Netlink encoding of datapath flow keys cannot express
6618 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
6619 * tag is interpreted as exact match on the fact that there is no
6620 * VLAN. Unless we refactor a lot of code that translates between
6621 * Netlink and struct flow representations, we have to do the same
6622 * here. This must be in sync with 'match' in dpif_netdev_flow_put(). */
6623 if (!match
.wc
.masks
.vlans
[0].tci
) {
6624 match
.wc
.masks
.vlans
[0].tci
= htons(0xffff);
6627 /* We can't allow the packet batching in the next loop to execute
6628 * the actions. Otherwise, if there are any slow path actions,
6629 * we'll send the packet up twice. */
6630 dp_packet_batch_init_packet(&b
, packet
);
6631 dp_netdev_execute_actions(pmd
, &b
, true, &match
.flow
,
6632 actions
->data
, actions
->size
);
6634 add_actions
= put_actions
->size
? put_actions
: actions
;
6635 if (OVS_LIKELY(error
!= ENOSPC
)) {
6636 struct dp_netdev_flow
*netdev_flow
;
6638 /* XXX: There's a race window where a flow covering this packet
6639 * could have already been installed since we last did the flow
6640 * lookup before upcall. This could be solved by moving the
6641 * mutex lock outside the loop, but that's an awful long time
6642 * to be locking revalidators out of making flow modifications. */
6643 ovs_mutex_lock(&pmd
->flow_mutex
);
6644 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
6645 if (OVS_LIKELY(!netdev_flow
)) {
6646 netdev_flow
= dp_netdev_flow_add(pmd
, &match
, &ufid
,
6650 ovs_mutex_unlock(&pmd
->flow_mutex
);
6651 uint32_t hash
= dp_netdev_flow_hash(&netdev_flow
->ufid
);
6652 smc_insert(pmd
, key
, hash
);
6653 emc_probabilistic_insert(pmd
, key
, netdev_flow
);
6655 if (pmd_perf_metrics_enabled(pmd
)) {
6656 /* Update upcall stats. */
6657 cycles
= cycles_counter_update(&pmd
->perf_stats
) - cycles
;
6658 struct pmd_perf_stats
*s
= &pmd
->perf_stats
;
6659 s
->current
.upcalls
++;
6660 s
->current
.upcall_cycles
+= cycles
;
6661 histogram_add_sample(&s
->cycles_per_upcall
, cycles
);
6667 fast_path_processing(struct dp_netdev_pmd_thread
*pmd
,
6668 struct dp_packet_batch
*packets_
,
6669 struct netdev_flow_key
**keys
,
6670 struct dp_packet_flow_map
*flow_map
,
6674 const size_t cnt
= dp_packet_batch_size(packets_
);
6675 #if !defined(__CHECKER__) && !defined(_WIN32)
6676 const size_t PKT_ARRAY_SIZE
= cnt
;
6678 /* Sparse or MSVC doesn't like variable length array. */
6679 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
6681 struct dp_packet
*packet
;
6683 struct dpcls_rule
*rules
[PKT_ARRAY_SIZE
];
6684 struct dp_netdev
*dp
= pmd
->dp
;
6685 int upcall_ok_cnt
= 0, upcall_fail_cnt
= 0;
6686 int lookup_cnt
= 0, add_lookup_cnt
;
6689 for (size_t i
= 0; i
< cnt
; i
++) {
6690 /* Key length is needed in all the cases, hash computed on demand. */
6691 keys
[i
]->len
= netdev_flow_key_size(miniflow_n_values(&keys
[i
]->mf
));
6693 /* Get the classifier for the in_port */
6694 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
6695 if (OVS_LIKELY(cls
)) {
6696 any_miss
= !dpcls_lookup(cls
, (const struct netdev_flow_key
**)keys
,
6697 rules
, cnt
, &lookup_cnt
);
6700 memset(rules
, 0, sizeof(rules
));
6702 if (OVS_UNLIKELY(any_miss
) && !fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
6703 uint64_t actions_stub
[512 / 8], slow_stub
[512 / 8];
6704 struct ofpbuf actions
, put_actions
;
6706 ofpbuf_use_stub(&actions
, actions_stub
, sizeof actions_stub
);
6707 ofpbuf_use_stub(&put_actions
, slow_stub
, sizeof slow_stub
);
6709 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
6710 struct dp_netdev_flow
*netdev_flow
;
6712 if (OVS_LIKELY(rules
[i
])) {
6716 /* It's possible that an earlier slow path execution installed
6717 * a rule covering this flow. In this case, it's a lot cheaper
6718 * to catch it here than execute a miss. */
6719 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, keys
[i
],
6722 lookup_cnt
+= add_lookup_cnt
;
6723 rules
[i
] = &netdev_flow
->cr
;
6727 int error
= handle_packet_upcall(pmd
, packet
, keys
[i
],
6728 &actions
, &put_actions
);
6730 if (OVS_UNLIKELY(error
)) {
6737 ofpbuf_uninit(&actions
);
6738 ofpbuf_uninit(&put_actions
);
6739 fat_rwlock_unlock(&dp
->upcall_rwlock
);
6740 } else if (OVS_UNLIKELY(any_miss
)) {
6741 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
6742 if (OVS_UNLIKELY(!rules
[i
])) {
6743 dp_packet_delete(packet
);
6749 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
6750 struct dp_netdev_flow
*flow
;
6751 /* Get the original order of this packet in received batch. */
6752 int recv_idx
= index_map
[i
];
6755 if (OVS_UNLIKELY(!rules
[i
])) {
6759 flow
= dp_netdev_flow_cast(rules
[i
]);
6760 uint32_t hash
= dp_netdev_flow_hash(&flow
->ufid
);
6761 smc_insert(pmd
, keys
[i
], hash
);
6763 emc_probabilistic_insert(pmd
, keys
[i
], flow
);
6764 /* Add these packets into the flow map in the same order
6767 tcp_flags
= miniflow_get_tcp_flags(&keys
[i
]->mf
);
6768 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6769 flow_map
, recv_idx
);
6772 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_MASKED_HIT
,
6773 cnt
- upcall_ok_cnt
- upcall_fail_cnt
);
6774 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_MASKED_LOOKUP
,
6776 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_MISS
,
6778 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_LOST
,
6782 /* Packets enter the datapath from a port (or from recirculation) here.
6784 * When 'md_is_valid' is true the metadata in 'packets' are already valid.
6785 * When false the metadata in 'packets' need to be initialized. */
6787 dp_netdev_input__(struct dp_netdev_pmd_thread
*pmd
,
6788 struct dp_packet_batch
*packets
,
6789 bool md_is_valid
, odp_port_t port_no
)
6791 #if !defined(__CHECKER__) && !defined(_WIN32)
6792 const size_t PKT_ARRAY_SIZE
= dp_packet_batch_size(packets
);
6794 /* Sparse or MSVC doesn't like variable length array. */
6795 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
6797 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
)
6798 struct netdev_flow_key keys
[PKT_ARRAY_SIZE
];
6799 struct netdev_flow_key
*missed_keys
[PKT_ARRAY_SIZE
];
6800 struct packet_batch_per_flow batches
[PKT_ARRAY_SIZE
];
6802 struct dp_packet_flow_map flow_map
[PKT_ARRAY_SIZE
];
6803 uint8_t index_map
[PKT_ARRAY_SIZE
];
6809 dfc_processing(pmd
, packets
, keys
, missed_keys
, batches
, &n_batches
,
6810 flow_map
, &n_flows
, index_map
, md_is_valid
, port_no
);
6812 if (!dp_packet_batch_is_empty(packets
)) {
6813 /* Get ingress port from first packet's metadata. */
6814 in_port
= packets
->packets
[0]->md
.in_port
.odp_port
;
6815 fast_path_processing(pmd
, packets
, missed_keys
,
6816 flow_map
, index_map
, in_port
);
6819 /* Batch rest of packets which are in flow map. */
6820 for (i
= 0; i
< n_flows
; i
++) {
6821 struct dp_packet_flow_map
*map
= &flow_map
[i
];
6823 if (OVS_UNLIKELY(!map
->flow
)) {
6826 dp_netdev_queue_batches(map
->packet
, map
->flow
, map
->tcp_flags
,
6827 batches
, &n_batches
);
6830 /* All the flow batches need to be reset before any call to
6831 * packet_batch_per_flow_execute() as it could potentially trigger
6832 * recirculation. When a packet matching flow ‘j’ happens to be
6833 * recirculated, the nested call to dp_netdev_input__() could potentially
6834 * classify the packet as matching another flow - say 'k'. It could happen
6835 * that in the previous call to dp_netdev_input__() that same flow 'k' had
6836 * already its own batches[k] still waiting to be served. So if its
6837 * ‘batch’ member is not reset, the recirculated packet would be wrongly
6838 * appended to batches[k] of the 1st call to dp_netdev_input__(). */
6839 for (i
= 0; i
< n_batches
; i
++) {
6840 batches
[i
].flow
->batch
= NULL
;
6843 for (i
= 0; i
< n_batches
; i
++) {
6844 packet_batch_per_flow_execute(&batches
[i
], pmd
);
6849 dp_netdev_input(struct dp_netdev_pmd_thread
*pmd
,
6850 struct dp_packet_batch
*packets
,
6853 dp_netdev_input__(pmd
, packets
, false, port_no
);
6857 dp_netdev_recirculate(struct dp_netdev_pmd_thread
*pmd
,
6858 struct dp_packet_batch
*packets
)
6860 dp_netdev_input__(pmd
, packets
, true, 0);
6863 struct dp_netdev_execute_aux
{
6864 struct dp_netdev_pmd_thread
*pmd
;
6865 const struct flow
*flow
;
6869 dpif_netdev_register_dp_purge_cb(struct dpif
*dpif
, dp_purge_callback
*cb
,
6872 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
6873 dp
->dp_purge_aux
= aux
;
6874 dp
->dp_purge_cb
= cb
;
6878 dpif_netdev_register_upcall_cb(struct dpif
*dpif
, upcall_callback
*cb
,
6881 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
6882 dp
->upcall_aux
= aux
;
6887 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
6891 struct dp_netdev_port
*port
;
6894 HMAP_FOR_EACH (tx
, node
, &pmd
->send_port_cache
) {
6895 if (!tx
->port
->dynamic_txqs
) {
6898 interval
= pmd
->ctx
.now
- tx
->last_used
;
6899 if (tx
->qid
>= 0 && (purge
|| interval
>= XPS_TIMEOUT
)) {
6901 ovs_mutex_lock(&port
->txq_used_mutex
);
6902 port
->txq_used
[tx
->qid
]--;
6903 ovs_mutex_unlock(&port
->txq_used_mutex
);
6910 dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
6913 struct dp_netdev_port
*port
;
6915 int i
, min_cnt
, min_qid
;
6917 interval
= pmd
->ctx
.now
- tx
->last_used
;
6918 tx
->last_used
= pmd
->ctx
.now
;
6920 if (OVS_LIKELY(tx
->qid
>= 0 && interval
< XPS_TIMEOUT
)) {
6926 ovs_mutex_lock(&port
->txq_used_mutex
);
6928 port
->txq_used
[tx
->qid
]--;
6934 for (i
= 0; i
< netdev_n_txq(port
->netdev
); i
++) {
6935 if (port
->txq_used
[i
] < min_cnt
|| min_cnt
== -1) {
6936 min_cnt
= port
->txq_used
[i
];
6941 port
->txq_used
[min_qid
]++;
6944 ovs_mutex_unlock(&port
->txq_used_mutex
);
6946 dpif_netdev_xps_revalidate_pmd(pmd
, false);
6948 VLOG_DBG("Core %d: New TX queue ID %d for port \'%s\'.",
6949 pmd
->core_id
, tx
->qid
, netdev_get_name(tx
->port
->netdev
));
6953 static struct tx_port
*
6954 pmd_tnl_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
6957 return tx_port_lookup(&pmd
->tnl_port_cache
, port_no
);
6960 static struct tx_port
*
6961 pmd_send_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
6964 return tx_port_lookup(&pmd
->send_port_cache
, port_no
);
6968 push_tnl_action(const struct dp_netdev_pmd_thread
*pmd
,
6969 const struct nlattr
*attr
,
6970 struct dp_packet_batch
*batch
)
6972 struct tx_port
*tun_port
;
6973 const struct ovs_action_push_tnl
*data
;
6976 data
= nl_attr_get(attr
);
6978 tun_port
= pmd_tnl_port_cache_lookup(pmd
, data
->tnl_port
);
6983 err
= netdev_push_header(tun_port
->port
->netdev
, batch
, data
);
6988 dp_packet_delete_batch(batch
, true);
6993 dp_execute_userspace_action(struct dp_netdev_pmd_thread
*pmd
,
6994 struct dp_packet
*packet
, bool should_steal
,
6995 struct flow
*flow
, ovs_u128
*ufid
,
6996 struct ofpbuf
*actions
,
6997 const struct nlattr
*userdata
)
6999 struct dp_packet_batch b
;
7002 ofpbuf_clear(actions
);
7004 error
= dp_netdev_upcall(pmd
, packet
, flow
, NULL
, ufid
,
7005 DPIF_UC_ACTION
, userdata
, actions
,
7007 if (!error
|| error
== ENOSPC
) {
7008 dp_packet_batch_init_packet(&b
, packet
);
7009 dp_netdev_execute_actions(pmd
, &b
, should_steal
, flow
,
7010 actions
->data
, actions
->size
);
7011 } else if (should_steal
) {
7012 dp_packet_delete(packet
);
7017 dp_execute_cb(void *aux_
, struct dp_packet_batch
*packets_
,
7018 const struct nlattr
*a
, bool should_steal
)
7019 OVS_NO_THREAD_SAFETY_ANALYSIS
7021 struct dp_netdev_execute_aux
*aux
= aux_
;
7022 uint32_t *depth
= recirc_depth_get();
7023 struct dp_netdev_pmd_thread
*pmd
= aux
->pmd
;
7024 struct dp_netdev
*dp
= pmd
->dp
;
7025 int type
= nl_attr_type(a
);
7028 switch ((enum ovs_action_attr
)type
) {
7029 case OVS_ACTION_ATTR_OUTPUT
:
7030 p
= pmd_send_port_cache_lookup(pmd
, nl_attr_get_odp_port(a
));
7031 if (OVS_LIKELY(p
)) {
7032 struct dp_packet
*packet
;
7033 struct dp_packet_batch out
;
7035 if (!should_steal
) {
7036 dp_packet_batch_clone(&out
, packets_
);
7037 dp_packet_batch_reset_cutlen(packets_
);
7040 dp_packet_batch_apply_cutlen(packets_
);
7043 if (OVS_UNLIKELY(!dp_packet_batch_is_empty(&p
->output_pkts
)
7044 && packets_
->packets
[0]->source
7045 != p
->output_pkts
.packets
[0]->source
)) {
7046 /* XXX: netdev-dpdk assumes that all packets in a single
7047 * output batch has the same source. Flush here to
7048 * avoid memory access issues. */
7049 dp_netdev_pmd_flush_output_on_port(pmd
, p
);
7052 if (dp_packet_batch_size(&p
->output_pkts
)
7053 + dp_packet_batch_size(packets_
) > NETDEV_MAX_BURST
) {
7054 /* Flush here to avoid overflow. */
7055 dp_netdev_pmd_flush_output_on_port(pmd
, p
);
7058 if (dp_packet_batch_is_empty(&p
->output_pkts
)) {
7059 pmd
->n_output_batches
++;
7062 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7063 p
->output_pkts_rxqs
[dp_packet_batch_size(&p
->output_pkts
)] =
7065 dp_packet_batch_add(&p
->output_pkts
, packet
);
7071 case OVS_ACTION_ATTR_TUNNEL_PUSH
:
7073 /* We're requested to push tunnel header, but also we need to take
7074 * the ownership of these packets. Thus, we can avoid performing
7075 * the action, because the caller will not use the result anyway.
7076 * Just break to free the batch. */
7079 dp_packet_batch_apply_cutlen(packets_
);
7080 push_tnl_action(pmd
, a
, packets_
);
7083 case OVS_ACTION_ATTR_TUNNEL_POP
:
7084 if (*depth
< MAX_RECIRC_DEPTH
) {
7085 struct dp_packet_batch
*orig_packets_
= packets_
;
7086 odp_port_t portno
= nl_attr_get_odp_port(a
);
7088 p
= pmd_tnl_port_cache_lookup(pmd
, portno
);
7090 struct dp_packet_batch tnl_pkt
;
7092 if (!should_steal
) {
7093 dp_packet_batch_clone(&tnl_pkt
, packets_
);
7094 packets_
= &tnl_pkt
;
7095 dp_packet_batch_reset_cutlen(orig_packets_
);
7098 dp_packet_batch_apply_cutlen(packets_
);
7100 netdev_pop_header(p
->port
->netdev
, packets_
);
7101 if (dp_packet_batch_is_empty(packets_
)) {
7105 struct dp_packet
*packet
;
7106 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7107 packet
->md
.in_port
.odp_port
= portno
;
7111 dp_netdev_recirculate(pmd
, packets_
);
7118 case OVS_ACTION_ATTR_USERSPACE
:
7119 if (!fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
7120 struct dp_packet_batch
*orig_packets_
= packets_
;
7121 const struct nlattr
*userdata
;
7122 struct dp_packet_batch usr_pkt
;
7123 struct ofpbuf actions
;
7128 userdata
= nl_attr_find_nested(a
, OVS_USERSPACE_ATTR_USERDATA
);
7129 ofpbuf_init(&actions
, 0);
7131 if (packets_
->trunc
) {
7132 if (!should_steal
) {
7133 dp_packet_batch_clone(&usr_pkt
, packets_
);
7134 packets_
= &usr_pkt
;
7136 dp_packet_batch_reset_cutlen(orig_packets_
);
7139 dp_packet_batch_apply_cutlen(packets_
);
7142 struct dp_packet
*packet
;
7143 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7144 flow_extract(packet
, &flow
);
7145 dpif_flow_hash(dp
->dpif
, &flow
, sizeof flow
, &ufid
);
7146 dp_execute_userspace_action(pmd
, packet
, should_steal
, &flow
,
7147 &ufid
, &actions
, userdata
);
7151 dp_packet_delete_batch(packets_
, true);
7154 ofpbuf_uninit(&actions
);
7155 fat_rwlock_unlock(&dp
->upcall_rwlock
);
7161 case OVS_ACTION_ATTR_RECIRC
:
7162 if (*depth
< MAX_RECIRC_DEPTH
) {
7163 struct dp_packet_batch recirc_pkts
;
7165 if (!should_steal
) {
7166 dp_packet_batch_clone(&recirc_pkts
, packets_
);
7167 packets_
= &recirc_pkts
;
7170 struct dp_packet
*packet
;
7171 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7172 packet
->md
.recirc_id
= nl_attr_get_u32(a
);
7176 dp_netdev_recirculate(pmd
, packets_
);
7182 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
7185 case OVS_ACTION_ATTR_CT
: {
7186 const struct nlattr
*b
;
7188 bool commit
= false;
7191 const char *helper
= NULL
;
7192 const uint32_t *setmark
= NULL
;
7193 const struct ovs_key_ct_labels
*setlabel
= NULL
;
7194 struct nat_action_info_t nat_action_info
;
7195 struct nat_action_info_t
*nat_action_info_ref
= NULL
;
7196 bool nat_config
= false;
7198 NL_ATTR_FOR_EACH_UNSAFE (b
, left
, nl_attr_get(a
),
7199 nl_attr_get_size(a
)) {
7200 enum ovs_ct_attr sub_type
= nl_attr_type(b
);
7203 case OVS_CT_ATTR_FORCE_COMMIT
:
7206 case OVS_CT_ATTR_COMMIT
:
7209 case OVS_CT_ATTR_ZONE
:
7210 zone
= nl_attr_get_u16(b
);
7212 case OVS_CT_ATTR_HELPER
:
7213 helper
= nl_attr_get_string(b
);
7215 case OVS_CT_ATTR_MARK
:
7216 setmark
= nl_attr_get(b
);
7218 case OVS_CT_ATTR_LABELS
:
7219 setlabel
= nl_attr_get(b
);
7221 case OVS_CT_ATTR_EVENTMASK
:
7222 /* Silently ignored, as userspace datapath does not generate
7223 * netlink events. */
7225 case OVS_CT_ATTR_NAT
: {
7226 const struct nlattr
*b_nest
;
7227 unsigned int left_nest
;
7228 bool ip_min_specified
= false;
7229 bool proto_num_min_specified
= false;
7230 bool ip_max_specified
= false;
7231 bool proto_num_max_specified
= false;
7232 memset(&nat_action_info
, 0, sizeof nat_action_info
);
7233 nat_action_info_ref
= &nat_action_info
;
7235 NL_NESTED_FOR_EACH_UNSAFE (b_nest
, left_nest
, b
) {
7236 enum ovs_nat_attr sub_type_nest
= nl_attr_type(b_nest
);
7238 switch (sub_type_nest
) {
7239 case OVS_NAT_ATTR_SRC
:
7240 case OVS_NAT_ATTR_DST
:
7242 nat_action_info
.nat_action
|=
7243 ((sub_type_nest
== OVS_NAT_ATTR_SRC
)
7244 ? NAT_ACTION_SRC
: NAT_ACTION_DST
);
7246 case OVS_NAT_ATTR_IP_MIN
:
7247 memcpy(&nat_action_info
.min_addr
,
7248 nl_attr_get(b_nest
),
7249 nl_attr_get_size(b_nest
));
7250 ip_min_specified
= true;
7252 case OVS_NAT_ATTR_IP_MAX
:
7253 memcpy(&nat_action_info
.max_addr
,
7254 nl_attr_get(b_nest
),
7255 nl_attr_get_size(b_nest
));
7256 ip_max_specified
= true;
7258 case OVS_NAT_ATTR_PROTO_MIN
:
7259 nat_action_info
.min_port
=
7260 nl_attr_get_u16(b_nest
);
7261 proto_num_min_specified
= true;
7263 case OVS_NAT_ATTR_PROTO_MAX
:
7264 nat_action_info
.max_port
=
7265 nl_attr_get_u16(b_nest
);
7266 proto_num_max_specified
= true;
7268 case OVS_NAT_ATTR_PERSISTENT
:
7269 case OVS_NAT_ATTR_PROTO_HASH
:
7270 case OVS_NAT_ATTR_PROTO_RANDOM
:
7272 case OVS_NAT_ATTR_UNSPEC
:
7273 case __OVS_NAT_ATTR_MAX
:
7278 if (ip_min_specified
&& !ip_max_specified
) {
7279 nat_action_info
.max_addr
= nat_action_info
.min_addr
;
7281 if (proto_num_min_specified
&& !proto_num_max_specified
) {
7282 nat_action_info
.max_port
= nat_action_info
.min_port
;
7284 if (proto_num_min_specified
|| proto_num_max_specified
) {
7285 if (nat_action_info
.nat_action
& NAT_ACTION_SRC
) {
7286 nat_action_info
.nat_action
|= NAT_ACTION_SRC_PORT
;
7287 } else if (nat_action_info
.nat_action
& NAT_ACTION_DST
) {
7288 nat_action_info
.nat_action
|= NAT_ACTION_DST_PORT
;
7293 case OVS_CT_ATTR_UNSPEC
:
7294 case __OVS_CT_ATTR_MAX
:
7299 /* We won't be able to function properly in this case, hence
7300 * complain loudly. */
7301 if (nat_config
&& !commit
) {
7302 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(5, 5);
7303 VLOG_WARN_RL(&rl
, "NAT specified without commit.");
7306 conntrack_execute(dp
->conntrack
, packets_
, aux
->flow
->dl_type
, force
,
7307 commit
, zone
, setmark
, setlabel
, aux
->flow
->tp_src
,
7308 aux
->flow
->tp_dst
, helper
, nat_action_info_ref
,
7309 pmd
->ctx
.now
/ 1000);
7313 case OVS_ACTION_ATTR_METER
:
7314 dp_netdev_run_meter(pmd
->dp
, packets_
, nl_attr_get_u32(a
),
7318 case OVS_ACTION_ATTR_PUSH_VLAN
:
7319 case OVS_ACTION_ATTR_POP_VLAN
:
7320 case OVS_ACTION_ATTR_PUSH_MPLS
:
7321 case OVS_ACTION_ATTR_POP_MPLS
:
7322 case OVS_ACTION_ATTR_SET
:
7323 case OVS_ACTION_ATTR_SET_MASKED
:
7324 case OVS_ACTION_ATTR_SAMPLE
:
7325 case OVS_ACTION_ATTR_HASH
:
7326 case OVS_ACTION_ATTR_UNSPEC
:
7327 case OVS_ACTION_ATTR_TRUNC
:
7328 case OVS_ACTION_ATTR_PUSH_ETH
:
7329 case OVS_ACTION_ATTR_POP_ETH
:
7330 case OVS_ACTION_ATTR_CLONE
:
7331 case OVS_ACTION_ATTR_PUSH_NSH
:
7332 case OVS_ACTION_ATTR_POP_NSH
:
7333 case OVS_ACTION_ATTR_CT_CLEAR
:
7334 case OVS_ACTION_ATTR_CHECK_PKT_LEN
:
7335 case __OVS_ACTION_ATTR_MAX
:
7339 dp_packet_delete_batch(packets_
, should_steal
);
7343 dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
7344 struct dp_packet_batch
*packets
,
7345 bool should_steal
, const struct flow
*flow
,
7346 const struct nlattr
*actions
, size_t actions_len
)
7348 struct dp_netdev_execute_aux aux
= { pmd
, flow
};
7350 odp_execute_actions(&aux
, packets
, should_steal
, actions
,
7351 actions_len
, dp_execute_cb
);
7354 struct dp_netdev_ct_dump
{
7355 struct ct_dpif_dump_state up
;
7356 struct conntrack_dump dump
;
7357 struct conntrack
*ct
;
7358 struct dp_netdev
*dp
;
7362 dpif_netdev_ct_dump_start(struct dpif
*dpif
, struct ct_dpif_dump_state
**dump_
,
7363 const uint16_t *pzone
, int *ptot_bkts
)
7365 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7366 struct dp_netdev_ct_dump
*dump
;
7368 dump
= xzalloc(sizeof *dump
);
7370 dump
->ct
= dp
->conntrack
;
7372 conntrack_dump_start(dp
->conntrack
, &dump
->dump
, pzone
, ptot_bkts
);
7380 dpif_netdev_ct_dump_next(struct dpif
*dpif OVS_UNUSED
,
7381 struct ct_dpif_dump_state
*dump_
,
7382 struct ct_dpif_entry
*entry
)
7384 struct dp_netdev_ct_dump
*dump
;
7386 INIT_CONTAINER(dump
, dump_
, up
);
7388 return conntrack_dump_next(&dump
->dump
, entry
);
7392 dpif_netdev_ct_dump_done(struct dpif
*dpif OVS_UNUSED
,
7393 struct ct_dpif_dump_state
*dump_
)
7395 struct dp_netdev_ct_dump
*dump
;
7398 INIT_CONTAINER(dump
, dump_
, up
);
7400 err
= conntrack_dump_done(&dump
->dump
);
7408 dpif_netdev_ct_flush(struct dpif
*dpif
, const uint16_t *zone
,
7409 const struct ct_dpif_tuple
*tuple
)
7411 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7414 return conntrack_flush_tuple(dp
->conntrack
, tuple
, zone
? *zone
: 0);
7416 return conntrack_flush(dp
->conntrack
, zone
);
7420 dpif_netdev_ct_set_maxconns(struct dpif
*dpif
, uint32_t maxconns
)
7422 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7424 return conntrack_set_maxconns(dp
->conntrack
, maxconns
);
7428 dpif_netdev_ct_get_maxconns(struct dpif
*dpif
, uint32_t *maxconns
)
7430 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7432 return conntrack_get_maxconns(dp
->conntrack
, maxconns
);
7436 dpif_netdev_ct_get_nconns(struct dpif
*dpif
, uint32_t *nconns
)
7438 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7440 return conntrack_get_nconns(dp
->conntrack
, nconns
);
7444 dpif_netdev_ipf_set_enabled(struct dpif
*dpif
, bool v6
, bool enable
)
7446 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7447 return ipf_set_enabled(conntrack_ipf_ctx(dp
->conntrack
), v6
, enable
);
7451 dpif_netdev_ipf_set_min_frag(struct dpif
*dpif
, bool v6
, uint32_t min_frag
)
7453 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7454 return ipf_set_min_frag(conntrack_ipf_ctx(dp
->conntrack
), v6
, min_frag
);
7458 dpif_netdev_ipf_set_max_nfrags(struct dpif
*dpif
, uint32_t max_frags
)
7460 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7461 return ipf_set_max_nfrags(conntrack_ipf_ctx(dp
->conntrack
), max_frags
);
7464 /* Adjust this function if 'dpif_ipf_status' and 'ipf_status' were to
7467 dpif_netdev_ipf_get_status(struct dpif
*dpif
,
7468 struct dpif_ipf_status
*dpif_ipf_status
)
7470 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7471 ipf_get_status(conntrack_ipf_ctx(dp
->conntrack
),
7472 (struct ipf_status
*) dpif_ipf_status
);
7477 dpif_netdev_ipf_dump_start(struct dpif
*dpif OVS_UNUSED
,
7478 struct ipf_dump_ctx
**ipf_dump_ctx
)
7480 return ipf_dump_start(ipf_dump_ctx
);
7484 dpif_netdev_ipf_dump_next(struct dpif
*dpif
, void *ipf_dump_ctx
, char **dump
)
7486 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7487 return ipf_dump_next(conntrack_ipf_ctx(dp
->conntrack
), ipf_dump_ctx
,
7492 dpif_netdev_ipf_dump_done(struct dpif
*dpif OVS_UNUSED
, void *ipf_dump_ctx
)
7494 return ipf_dump_done(ipf_dump_ctx
);
7498 const struct dpif_class dpif_netdev_class
= {
7500 true, /* cleanup_required */
7502 dpif_netdev_enumerate
,
7503 dpif_netdev_port_open_type
,
7506 dpif_netdev_destroy
,
7509 dpif_netdev_get_stats
,
7510 dpif_netdev_port_add
,
7511 dpif_netdev_port_del
,
7512 dpif_netdev_port_set_config
,
7513 dpif_netdev_port_query_by_number
,
7514 dpif_netdev_port_query_by_name
,
7515 NULL
, /* port_get_pid */
7516 dpif_netdev_port_dump_start
,
7517 dpif_netdev_port_dump_next
,
7518 dpif_netdev_port_dump_done
,
7519 dpif_netdev_port_poll
,
7520 dpif_netdev_port_poll_wait
,
7521 dpif_netdev_flow_flush
,
7522 dpif_netdev_flow_dump_create
,
7523 dpif_netdev_flow_dump_destroy
,
7524 dpif_netdev_flow_dump_thread_create
,
7525 dpif_netdev_flow_dump_thread_destroy
,
7526 dpif_netdev_flow_dump_next
,
7527 dpif_netdev_operate
,
7528 NULL
, /* recv_set */
7529 NULL
, /* handlers_set */
7530 dpif_netdev_set_config
,
7531 dpif_netdev_queue_to_priority
,
7533 NULL
, /* recv_wait */
7534 NULL
, /* recv_purge */
7535 dpif_netdev_register_dp_purge_cb
,
7536 dpif_netdev_register_upcall_cb
,
7537 dpif_netdev_enable_upcall
,
7538 dpif_netdev_disable_upcall
,
7539 dpif_netdev_get_datapath_version
,
7540 dpif_netdev_ct_dump_start
,
7541 dpif_netdev_ct_dump_next
,
7542 dpif_netdev_ct_dump_done
,
7543 dpif_netdev_ct_flush
,
7544 dpif_netdev_ct_set_maxconns
,
7545 dpif_netdev_ct_get_maxconns
,
7546 dpif_netdev_ct_get_nconns
,
7547 NULL
, /* ct_set_limits */
7548 NULL
, /* ct_get_limits */
7549 NULL
, /* ct_del_limits */
7550 dpif_netdev_ipf_set_enabled
,
7551 dpif_netdev_ipf_set_min_frag
,
7552 dpif_netdev_ipf_set_max_nfrags
,
7553 dpif_netdev_ipf_get_status
,
7554 dpif_netdev_ipf_dump_start
,
7555 dpif_netdev_ipf_dump_next
,
7556 dpif_netdev_ipf_dump_done
,
7557 dpif_netdev_meter_get_features
,
7558 dpif_netdev_meter_set
,
7559 dpif_netdev_meter_get
,
7560 dpif_netdev_meter_del
,
7564 dpif_dummy_change_port_number(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
7565 const char *argv
[], void *aux OVS_UNUSED
)
7567 struct dp_netdev_port
*port
;
7568 struct dp_netdev
*dp
;
7571 ovs_mutex_lock(&dp_netdev_mutex
);
7572 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
7573 if (!dp
|| !dpif_netdev_class_is_dummy(dp
->class)) {
7574 ovs_mutex_unlock(&dp_netdev_mutex
);
7575 unixctl_command_reply_error(conn
, "unknown datapath or not a dummy");
7578 ovs_refcount_ref(&dp
->ref_cnt
);
7579 ovs_mutex_unlock(&dp_netdev_mutex
);
7581 ovs_mutex_lock(&dp
->port_mutex
);
7582 if (get_port_by_name(dp
, argv
[2], &port
)) {
7583 unixctl_command_reply_error(conn
, "unknown port");
7587 port_no
= u32_to_odp(atoi(argv
[3]));
7588 if (!port_no
|| port_no
== ODPP_NONE
) {
7589 unixctl_command_reply_error(conn
, "bad port number");
7592 if (dp_netdev_lookup_port(dp
, port_no
)) {
7593 unixctl_command_reply_error(conn
, "port number already in use");
7598 hmap_remove(&dp
->ports
, &port
->node
);
7599 reconfigure_datapath(dp
);
7601 /* Reinsert with new port number. */
7602 port
->port_no
= port_no
;
7603 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
7604 reconfigure_datapath(dp
);
7606 seq_change(dp
->port_seq
);
7607 unixctl_command_reply(conn
, NULL
);
7610 ovs_mutex_unlock(&dp
->port_mutex
);
7611 dp_netdev_unref(dp
);
7615 dpif_dummy_register__(const char *type
)
7617 struct dpif_class
*class;
7619 class = xmalloc(sizeof *class);
7620 *class = dpif_netdev_class
;
7621 class->type
= xstrdup(type
);
7622 dp_register_provider(class);
7626 dpif_dummy_override(const char *type
)
7631 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
7632 * a userland-only build. It's useful for testsuite.
7634 error
= dp_unregister_provider(type
);
7635 if (error
== 0 || error
== EAFNOSUPPORT
) {
7636 dpif_dummy_register__(type
);
7641 dpif_dummy_register(enum dummy_level level
)
7643 if (level
== DUMMY_OVERRIDE_ALL
) {
7648 dp_enumerate_types(&types
);
7649 SSET_FOR_EACH (type
, &types
) {
7650 dpif_dummy_override(type
);
7652 sset_destroy(&types
);
7653 } else if (level
== DUMMY_OVERRIDE_SYSTEM
) {
7654 dpif_dummy_override("system");
7657 dpif_dummy_register__("dummy");
7659 unixctl_command_register("dpif-dummy/change-port-number",
7660 "dp port new-number",
7661 3, 3, dpif_dummy_change_port_number
, NULL
);
7664 /* Datapath Classifier. */
7667 dpcls_subtable_destroy_cb(struct dpcls_subtable
*subtable
)
7669 cmap_destroy(&subtable
->rules
);
7670 ovsrcu_postpone(free
, subtable
);
7673 /* Initializes 'cls' as a classifier that initially contains no classification
7676 dpcls_init(struct dpcls
*cls
)
7678 cmap_init(&cls
->subtables_map
);
7679 pvector_init(&cls
->subtables
);
7683 dpcls_destroy_subtable(struct dpcls
*cls
, struct dpcls_subtable
*subtable
)
7685 VLOG_DBG("Destroying subtable %p for in_port %d", subtable
, cls
->in_port
);
7686 pvector_remove(&cls
->subtables
, subtable
);
7687 cmap_remove(&cls
->subtables_map
, &subtable
->cmap_node
,
7688 subtable
->mask
.hash
);
7689 ovsrcu_postpone(dpcls_subtable_destroy_cb
, subtable
);
7692 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
7693 * caller's responsibility.
7694 * May only be called after all the readers have been terminated. */
7696 dpcls_destroy(struct dpcls
*cls
)
7699 struct dpcls_subtable
*subtable
;
7701 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
7702 ovs_assert(cmap_count(&subtable
->rules
) == 0);
7703 dpcls_destroy_subtable(cls
, subtable
);
7705 cmap_destroy(&cls
->subtables_map
);
7706 pvector_destroy(&cls
->subtables
);
7710 static struct dpcls_subtable
*
7711 dpcls_create_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
7713 struct dpcls_subtable
*subtable
;
7715 /* Need to add one. */
7716 subtable
= xmalloc(sizeof *subtable
7717 - sizeof subtable
->mask
.mf
+ mask
->len
);
7718 cmap_init(&subtable
->rules
);
7719 subtable
->hit_cnt
= 0;
7720 netdev_flow_key_clone(&subtable
->mask
, mask
);
7722 /* Decide which hash/lookup/verify function to use. */
7723 subtable
->lookup_func
= dpcls_subtable_lookup_generic
;
7725 cmap_insert(&cls
->subtables_map
, &subtable
->cmap_node
, mask
->hash
);
7726 /* Add the new subtable at the end of the pvector (with no hits yet) */
7727 pvector_insert(&cls
->subtables
, subtable
, 0);
7728 VLOG_DBG("Creating %"PRIuSIZE
". subtable %p for in_port %d",
7729 cmap_count(&cls
->subtables_map
), subtable
, cls
->in_port
);
7730 pvector_publish(&cls
->subtables
);
7735 static inline struct dpcls_subtable
*
7736 dpcls_find_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
7738 struct dpcls_subtable
*subtable
;
7740 CMAP_FOR_EACH_WITH_HASH (subtable
, cmap_node
, mask
->hash
,
7741 &cls
->subtables_map
) {
7742 if (netdev_flow_key_equal(&subtable
->mask
, mask
)) {
7746 return dpcls_create_subtable(cls
, mask
);
7750 /* Periodically sort the dpcls subtable vectors according to hit counts */
7752 dpcls_sort_subtable_vector(struct dpcls
*cls
)
7754 struct pvector
*pvec
= &cls
->subtables
;
7755 struct dpcls_subtable
*subtable
;
7757 PVECTOR_FOR_EACH (subtable
, pvec
) {
7758 pvector_change_priority(pvec
, subtable
, subtable
->hit_cnt
);
7759 subtable
->hit_cnt
= 0;
7761 pvector_publish(pvec
);
7765 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
,
7766 struct polled_queue
*poll_list
, int poll_cnt
)
7769 uint64_t tot_idle
= 0, tot_proc
= 0;
7770 unsigned int pmd_load
= 0;
7772 if (pmd
->ctx
.now
> pmd
->rxq_next_cycle_store
) {
7774 struct pmd_auto_lb
*pmd_alb
= &pmd
->dp
->pmd_alb
;
7775 if (pmd_alb
->is_enabled
&& !pmd
->isolated
7776 && (pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_IDLE
] >=
7777 pmd
->prev_stats
[PMD_CYCLES_ITER_IDLE
])
7778 && (pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_BUSY
] >=
7779 pmd
->prev_stats
[PMD_CYCLES_ITER_BUSY
]))
7781 tot_idle
= pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_IDLE
] -
7782 pmd
->prev_stats
[PMD_CYCLES_ITER_IDLE
];
7783 tot_proc
= pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_BUSY
] -
7784 pmd
->prev_stats
[PMD_CYCLES_ITER_BUSY
];
7787 pmd_load
= ((tot_proc
* 100) / (tot_idle
+ tot_proc
));
7790 if (pmd_load
>= ALB_PMD_LOAD_THRESHOLD
) {
7791 atomic_count_inc(&pmd
->pmd_overloaded
);
7793 atomic_count_set(&pmd
->pmd_overloaded
, 0);
7797 pmd
->prev_stats
[PMD_CYCLES_ITER_IDLE
] =
7798 pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_IDLE
];
7799 pmd
->prev_stats
[PMD_CYCLES_ITER_BUSY
] =
7800 pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_BUSY
];
7802 /* Get the cycles that were used to process each queue and store. */
7803 for (unsigned i
= 0; i
< poll_cnt
; i
++) {
7804 uint64_t rxq_cyc_curr
= dp_netdev_rxq_get_cycles(poll_list
[i
].rxq
,
7805 RXQ_CYCLES_PROC_CURR
);
7806 dp_netdev_rxq_set_intrvl_cycles(poll_list
[i
].rxq
, rxq_cyc_curr
);
7807 dp_netdev_rxq_set_cycles(poll_list
[i
].rxq
, RXQ_CYCLES_PROC_CURR
,
7810 curr_tsc
= cycles_counter_update(&pmd
->perf_stats
);
7811 if (pmd
->intrvl_tsc_prev
) {
7812 /* There is a prev timestamp, store a new intrvl cycle count. */
7813 atomic_store_relaxed(&pmd
->intrvl_cycles
,
7814 curr_tsc
- pmd
->intrvl_tsc_prev
);
7816 pmd
->intrvl_tsc_prev
= curr_tsc
;
7817 /* Start new measuring interval */
7818 pmd
->rxq_next_cycle_store
= pmd
->ctx
.now
+ PMD_RXQ_INTERVAL_LEN
;
7821 if (pmd
->ctx
.now
> pmd
->next_optimization
) {
7822 /* Try to obtain the flow lock to block out revalidator threads.
7823 * If not possible, just try next time. */
7824 if (!ovs_mutex_trylock(&pmd
->flow_mutex
)) {
7825 /* Optimize each classifier */
7826 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
7827 dpcls_sort_subtable_vector(cls
);
7829 ovs_mutex_unlock(&pmd
->flow_mutex
);
7830 /* Start new measuring interval */
7831 pmd
->next_optimization
= pmd
->ctx
.now
7832 + DPCLS_OPTIMIZATION_INTERVAL
;
7837 /* Insert 'rule' into 'cls'. */
7839 dpcls_insert(struct dpcls
*cls
, struct dpcls_rule
*rule
,
7840 const struct netdev_flow_key
*mask
)
7842 struct dpcls_subtable
*subtable
= dpcls_find_subtable(cls
, mask
);
7844 /* Refer to subtable's mask, also for later removal. */
7845 rule
->mask
= &subtable
->mask
;
7846 cmap_insert(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
);
7849 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
7851 dpcls_remove(struct dpcls
*cls
, struct dpcls_rule
*rule
)
7853 struct dpcls_subtable
*subtable
;
7855 ovs_assert(rule
->mask
);
7857 /* Get subtable from reference in rule->mask. */
7858 INIT_CONTAINER(subtable
, rule
->mask
, mask
);
7859 if (cmap_remove(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
)
7861 /* Delete empty subtable. */
7862 dpcls_destroy_subtable(cls
, subtable
);
7863 pvector_publish(&cls
->subtables
);
7867 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
7868 * in 'mask' the values in 'key' and 'target' are the same. */
7870 dpcls_rule_matches_key(const struct dpcls_rule
*rule
,
7871 const struct netdev_flow_key
*target
)
7873 const uint64_t *keyp
= miniflow_get_values(&rule
->flow
.mf
);
7874 const uint64_t *maskp
= miniflow_get_values(&rule
->mask
->mf
);
7877 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, target
, rule
->flow
.mf
.map
) {
7878 if (OVS_UNLIKELY((value
& *maskp
++) != *keyp
++)) {
7886 dpcls_subtable_lookup_generic(struct dpcls_subtable
*subtable
,
7888 const struct netdev_flow_key
*keys
[],
7889 struct dpcls_rule
**rules
)
7894 /* Compute hashes for the remaining keys. Each search-key is
7895 * masked with the subtable's mask to avoid hashing the wildcarded
7897 uint32_t hashes
[NETDEV_MAX_BURST
];
7898 ULLONG_FOR_EACH_1 (i
, keys_map
) {
7899 hashes
[i
] = netdev_flow_key_hash_in_mask(keys
[i
],
7904 const struct cmap_node
*nodes
[NETDEV_MAX_BURST
];
7905 found_map
= cmap_find_batch(&subtable
->rules
, keys_map
, hashes
, nodes
);
7907 /* Check results. When the i-th bit of found_map is set, it means
7908 * that a set of nodes with a matching hash value was found for the
7909 * i-th search-key. Due to possible hash collisions we need to check
7910 * which of the found rules, if any, really matches our masked
7912 ULLONG_FOR_EACH_1 (i
, found_map
) {
7913 struct dpcls_rule
*rule
;
7915 CMAP_NODE_FOR_EACH (rule
, cmap_node
, nodes
[i
]) {
7916 if (OVS_LIKELY(dpcls_rule_matches_key(rule
, keys
[i
]))) {
7918 /* Even at 20 Mpps the 32-bit hit_cnt cannot wrap
7919 * within one second optimization interval. */
7920 subtable
->hit_cnt
++;
7924 /* None of the found rules was a match. Reset the i-th bit to
7925 * keep searching this key in the next subtable. */
7926 ULLONG_SET0(found_map
, i
); /* Did not match. */
7928 ; /* Keep Sparse happy. */
7934 /* For each miniflow in 'keys' performs a classifier lookup writing the result
7935 * into the corresponding slot in 'rules'. If a particular entry in 'keys' is
7936 * NULL it is skipped.
7938 * This function is optimized for use in the userspace datapath and therefore
7939 * does not implement a lot of features available in the standard
7940 * classifier_lookup() function. Specifically, it does not implement
7941 * priorities, instead returning any rule which matches the flow.
7943 * Returns true if all miniflows found a corresponding rule. */
7945 dpcls_lookup(struct dpcls
*cls
, const struct netdev_flow_key
*keys
[],
7946 struct dpcls_rule
**rules
, const size_t cnt
,
7949 /* The received 'cnt' miniflows are the search-keys that will be processed
7950 * to find a matching entry into the available subtables.
7951 * The number of bits in map_type is equal to NETDEV_MAX_BURST. */
7952 #define MAP_BITS (sizeof(uint32_t) * CHAR_BIT)
7953 BUILD_ASSERT_DECL(MAP_BITS
>= NETDEV_MAX_BURST
);
7955 struct dpcls_subtable
*subtable
;
7957 uint32_t keys_map
= TYPE_MAXIMUM(uint32_t); /* Set all bits. */
7959 if (cnt
!= MAP_BITS
) {
7960 keys_map
>>= MAP_BITS
- cnt
; /* Clear extra bits. */
7962 memset(rules
, 0, cnt
* sizeof *rules
);
7964 int lookups_match
= 0, subtable_pos
= 1;
7967 /* The Datapath classifier - aka dpcls - is composed of subtables.
7968 * Subtables are dynamically created as needed when new rules are inserted.
7969 * Each subtable collects rules with matches on a specific subset of packet
7970 * fields as defined by the subtable's mask. We proceed to process every
7971 * search-key against each subtable, but when a match is found for a
7972 * search-key, the search for that key can stop because the rules are
7973 * non-overlapping. */
7974 PVECTOR_FOR_EACH (subtable
, &cls
->subtables
) {
7975 /* Call the subtable specific lookup function. */
7976 found_map
= subtable
->lookup_func(subtable
, keys_map
, keys
, rules
);
7978 /* Count the number of subtables searched for this packet match. This
7979 * estimates the "spread" of subtables looked at per matched packet. */
7980 uint32_t pkts_matched
= count_1bits(found_map
);
7981 lookups_match
+= pkts_matched
* subtable_pos
;
7983 /* Clear the found rules, and return early if all packets are found. */
7984 keys_map
&= ~found_map
;
7986 if (num_lookups_p
) {
7987 *num_lookups_p
= lookups_match
;
7994 if (num_lookups_p
) {
7995 *num_lookups_p
= lookups_match
;