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
,
256 /* Set of supported meter flags */
257 #define DP_SUPPORTED_METER_FLAGS_MASK \
258 (OFPMF13_STATS | OFPMF13_PKTPS | OFPMF13_KBPS | OFPMF13_BURST)
260 /* Set of supported meter band types */
261 #define DP_SUPPORTED_METER_BAND_TYPES \
262 ( 1 << OFPMBT13_DROP )
264 struct dp_meter_band
{
265 struct ofputil_meter_band up
; /* type, prec_level, pad, rate, burst_size */
266 uint32_t bucket
; /* In 1/1000 packets (for PKTPS), or in bits (for KBPS) */
267 uint64_t packet_count
;
274 uint32_t max_delta_t
;
276 uint64_t packet_count
;
278 struct dp_meter_band bands
[];
282 bool auto_lb_requested
; /* Auto load balancing requested by user. */
283 bool is_enabled
; /* Current status of Auto load balancing. */
284 uint64_t rebalance_intvl
;
285 uint64_t rebalance_poll_timer
;
288 /* Datapath based on the network device interface from netdev.h.
294 * Some members, marked 'const', are immutable. Accessing other members
295 * requires synchronization, as noted in more detail below.
297 * Acquisition order is, from outermost to innermost:
299 * dp_netdev_mutex (global)
304 const struct dpif_class
*const class;
305 const char *const name
;
307 struct ovs_refcount ref_cnt
;
308 atomic_flag destroyed
;
312 * Any lookup into 'ports' or any access to the dp_netdev_ports found
313 * through 'ports' requires taking 'port_mutex'. */
314 struct ovs_mutex port_mutex
;
316 struct seq
*port_seq
; /* Incremented whenever a port changes. */
318 /* The time that a packet can wait in output batch for sending. */
319 atomic_uint32_t tx_flush_interval
;
322 struct ovs_mutex meter_locks
[N_METER_LOCKS
];
323 struct dp_meter
*meters
[MAX_METERS
]; /* Meter bands. */
325 /* Probability of EMC insertions is a factor of 'emc_insert_min'.*/
326 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
) atomic_uint32_t emc_insert_min
;
327 /* Enable collection of PMD performance metrics. */
328 atomic_bool pmd_perf_metrics
;
329 /* Enable the SMC cache from ovsdb config */
330 atomic_bool smc_enable_db
;
332 /* Protects access to ofproto-dpif-upcall interface during revalidator
333 * thread synchronization. */
334 struct fat_rwlock upcall_rwlock
;
335 upcall_callback
*upcall_cb
; /* Callback function for executing upcalls. */
338 /* Callback function for notifying the purging of dp flows (during
339 * reseting pmd deletion). */
340 dp_purge_callback
*dp_purge_cb
;
343 /* Stores all 'struct dp_netdev_pmd_thread's. */
344 struct cmap poll_threads
;
345 /* id pool for per thread static_tx_qid. */
346 struct id_pool
*tx_qid_pool
;
347 struct ovs_mutex tx_qid_pool_mutex
;
348 /* Use measured cycles for rxq to pmd assignment. */
349 bool pmd_rxq_assign_cyc
;
351 /* Protects the access of the 'struct dp_netdev_pmd_thread'
352 * instance for non-pmd thread. */
353 struct ovs_mutex non_pmd_mutex
;
355 /* Each pmd thread will store its pointer to
356 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
357 ovsthread_key_t per_pmd_key
;
359 struct seq
*reconfigure_seq
;
360 uint64_t last_reconfigure_seq
;
362 /* Cpu mask for pin of pmd threads. */
365 uint64_t last_tnl_conf_seq
;
367 struct conntrack
*conntrack
;
368 struct pmd_auto_lb pmd_alb
;
371 static void meter_lock(const struct dp_netdev
*dp
, uint32_t meter_id
)
372 OVS_ACQUIRES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
374 ovs_mutex_lock(&dp
->meter_locks
[meter_id
% N_METER_LOCKS
]);
377 static void meter_unlock(const struct dp_netdev
*dp
, uint32_t meter_id
)
378 OVS_RELEASES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
380 ovs_mutex_unlock(&dp
->meter_locks
[meter_id
% N_METER_LOCKS
]);
384 static struct dp_netdev_port
*dp_netdev_lookup_port(const struct dp_netdev
*dp
,
386 OVS_REQUIRES(dp
->port_mutex
);
388 enum rxq_cycles_counter_type
{
389 RXQ_CYCLES_PROC_CURR
, /* Cycles spent successfully polling and
390 processing packets during the current
392 RXQ_CYCLES_PROC_HIST
, /* Total cycles of all intervals that are used
393 during rxq to pmd assignment. */
398 DP_NETDEV_FLOW_OFFLOAD_OP_ADD
,
399 DP_NETDEV_FLOW_OFFLOAD_OP_MOD
,
400 DP_NETDEV_FLOW_OFFLOAD_OP_DEL
,
403 struct dp_flow_offload_item
{
404 struct dp_netdev_pmd_thread
*pmd
;
405 struct dp_netdev_flow
*flow
;
408 struct nlattr
*actions
;
411 struct ovs_list node
;
414 struct dp_flow_offload
{
415 struct ovs_mutex mutex
;
416 struct ovs_list list
;
420 static struct dp_flow_offload dp_flow_offload
= {
421 .mutex
= OVS_MUTEX_INITIALIZER
,
422 .list
= OVS_LIST_INITIALIZER(&dp_flow_offload
.list
),
425 static struct ovsthread_once offload_thread_once
426 = OVSTHREAD_ONCE_INITIALIZER
;
428 #define XPS_TIMEOUT 500000LL /* In microseconds. */
430 /* Contained by struct dp_netdev_port's 'rxqs' member. */
431 struct dp_netdev_rxq
{
432 struct dp_netdev_port
*port
;
433 struct netdev_rxq
*rx
;
434 unsigned core_id
; /* Core to which this queue should be
435 pinned. OVS_CORE_UNSPEC if the
436 queue doesn't need to be pinned to a
438 unsigned intrvl_idx
; /* Write index for 'cycles_intrvl'. */
439 struct dp_netdev_pmd_thread
*pmd
; /* pmd thread that polls this queue. */
440 bool is_vhost
; /* Is rxq of a vhost port. */
442 /* Counters of cycles spent successfully polling and processing pkts. */
443 atomic_ullong cycles
[RXQ_N_CYCLES
];
444 /* We store PMD_RXQ_INTERVAL_MAX intervals of data for an rxq and then
445 sum them to yield the cycles used for an rxq. */
446 atomic_ullong cycles_intrvl
[PMD_RXQ_INTERVAL_MAX
];
449 /* A port in a netdev-based datapath. */
450 struct dp_netdev_port
{
452 bool dynamic_txqs
; /* If true XPS will be used. */
453 bool need_reconfigure
; /* True if we should reconfigure netdev. */
454 struct netdev
*netdev
;
455 struct hmap_node node
; /* Node in dp_netdev's 'ports'. */
456 struct netdev_saved_flags
*sf
;
457 struct dp_netdev_rxq
*rxqs
;
458 unsigned n_rxq
; /* Number of elements in 'rxqs' */
459 unsigned *txq_used
; /* Number of threads that use each tx queue. */
460 struct ovs_mutex txq_used_mutex
;
461 bool emc_enabled
; /* If true EMC will be used. */
462 char *type
; /* Port type as requested by user. */
463 char *rxq_affinity_list
; /* Requested affinity of rx queues. */
466 /* Contained by struct dp_netdev_flow's 'stats' member. */
467 struct dp_netdev_flow_stats
{
468 atomic_llong used
; /* Last used time, in monotonic msecs. */
469 atomic_ullong packet_count
; /* Number of packets matched. */
470 atomic_ullong byte_count
; /* Number of bytes matched. */
471 atomic_uint16_t tcp_flags
; /* Bitwise-OR of seen tcp_flags values. */
474 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
480 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
481 * its pmd thread's classifier. The text below calls this classifier 'cls'.
486 * The thread safety rules described here for "struct dp_netdev_flow" are
487 * motivated by two goals:
489 * - Prevent threads that read members of "struct dp_netdev_flow" from
490 * reading bad data due to changes by some thread concurrently modifying
493 * - Prevent two threads making changes to members of a given "struct
494 * dp_netdev_flow" from interfering with each other.
500 * A flow 'flow' may be accessed without a risk of being freed during an RCU
501 * grace period. Code that needs to hold onto a flow for a while
502 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
504 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
505 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
508 * Some members, marked 'const', are immutable. Accessing other members
509 * requires synchronization, as noted in more detail below.
511 struct dp_netdev_flow
{
512 const struct flow flow
; /* Unmasked flow that created this entry. */
513 /* Hash table index by unmasked flow. */
514 const struct cmap_node node
; /* In owning dp_netdev_pmd_thread's */
516 const struct cmap_node mark_node
; /* In owning flow_mark's mark_to_flow */
517 const ovs_u128 ufid
; /* Unique flow identifier. */
518 const ovs_u128 mega_ufid
; /* Unique mega flow identifier. */
519 const unsigned pmd_id
; /* The 'core_id' of pmd thread owning this */
522 /* Number of references.
523 * The classifier owns one reference.
524 * Any thread trying to keep a rule from being freed should hold its own
526 struct ovs_refcount ref_cnt
;
529 uint32_t mark
; /* Unique flow mark assigned to a flow */
532 struct dp_netdev_flow_stats stats
;
535 OVSRCU_TYPE(struct dp_netdev_actions
*) actions
;
537 /* While processing a group of input packets, the datapath uses the next
538 * member to store a pointer to the output batch for the flow. It is
539 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
540 * packet_batch_per_flow_init() and packet_batch_per_flow_execute()). */
541 struct packet_batch_per_flow
*batch
;
543 /* Packet classification. */
544 struct dpcls_rule cr
; /* In owning dp_netdev's 'cls'. */
545 /* 'cr' must be the last member. */
548 static void dp_netdev_flow_unref(struct dp_netdev_flow
*);
549 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*);
550 static int dpif_netdev_flow_from_nlattrs(const struct nlattr
*, uint32_t,
551 struct flow
*, bool);
553 /* A set of datapath actions within a "struct dp_netdev_flow".
559 * A struct dp_netdev_actions 'actions' is protected with RCU. */
560 struct dp_netdev_actions
{
561 /* These members are immutable: they do not change during the struct's
563 unsigned int size
; /* Size of 'actions', in bytes. */
564 struct nlattr actions
[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
567 struct dp_netdev_actions
*dp_netdev_actions_create(const struct nlattr
*,
569 struct dp_netdev_actions
*dp_netdev_flow_get_actions(
570 const struct dp_netdev_flow
*);
571 static void dp_netdev_actions_free(struct dp_netdev_actions
*);
573 struct polled_queue
{
574 struct dp_netdev_rxq
*rxq
;
581 /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */
583 struct dp_netdev_rxq
*rxq
;
584 struct hmap_node node
;
587 /* Contained by struct dp_netdev_pmd_thread's 'send_port_cache',
588 * 'tnl_port_cache' or 'tx_ports'. */
590 struct dp_netdev_port
*port
;
593 struct hmap_node node
;
594 long long flush_time
;
595 struct dp_packet_batch output_pkts
;
596 struct dp_netdev_rxq
*output_pkts_rxqs
[NETDEV_MAX_BURST
];
599 /* A set of properties for the current processing loop that is not directly
600 * associated with the pmd thread itself, but with the packets being
601 * processed or the short-term system configuration (for example, time).
602 * Contained by struct dp_netdev_pmd_thread's 'ctx' member. */
603 struct dp_netdev_pmd_thread_ctx
{
604 /* Latest measured time. See 'pmd_thread_ctx_time_update()'. */
606 /* RX queue from which last packet was received. */
607 struct dp_netdev_rxq
*last_rxq
;
608 /* EMC insertion probability context for the current processing cycle. */
609 uint32_t emc_insert_min
;
612 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
613 * the performance overhead of interrupt processing. Therefore netdev can
614 * not implement rx-wait for these devices. dpif-netdev needs to poll
615 * these device to check for recv buffer. pmd-thread does polling for
616 * devices assigned to itself.
618 * DPDK used PMD for accessing NIC.
620 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
621 * I/O of all non-pmd threads. There will be no actual thread created
624 * Each struct has its own flow cache and classifier per managed ingress port.
625 * For packets received on ingress port, a look up is done on corresponding PMD
626 * thread's flow cache and in case of a miss, lookup is performed in the
627 * corresponding classifier of port. Packets are executed with the found
628 * actions in either case.
630 struct dp_netdev_pmd_thread
{
631 struct dp_netdev
*dp
;
632 struct ovs_refcount ref_cnt
; /* Every reference must be refcount'ed. */
633 struct cmap_node node
; /* In 'dp->poll_threads'. */
635 /* Per thread exact-match cache. Note, the instance for cpu core
636 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
637 * need to be protected by 'non_pmd_mutex'. Every other instance
638 * will only be accessed by its own pmd thread. */
639 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
) struct dfc_cache flow_cache
;
641 /* Flow-Table and classifiers
643 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
644 * changes to 'classifiers' must be made while still holding the
647 struct ovs_mutex flow_mutex
;
648 struct cmap flow_table OVS_GUARDED
; /* Flow table. */
650 /* One classifier per in_port polled by the pmd */
651 struct cmap classifiers
;
652 /* Periodically sort subtable vectors according to hit frequencies */
653 long long int next_optimization
;
654 /* End of the next time interval for which processing cycles
655 are stored for each polled rxq. */
656 long long int rxq_next_cycle_store
;
658 /* Last interval timestamp. */
659 uint64_t intrvl_tsc_prev
;
660 /* Last interval cycles. */
661 atomic_ullong intrvl_cycles
;
663 /* Current context of the PMD thread. */
664 struct dp_netdev_pmd_thread_ctx ctx
;
666 struct seq
*reload_seq
;
667 uint64_t last_reload_seq
;
669 /* These are atomic variables used as a synchronization and configuration
670 * points for thread reload/exit.
672 * 'reload' atomic is the main one and it's used as a memory
673 * synchronization point for all other knobs and data.
675 * For a thread that requests PMD reload:
677 * * All changes that should be visible to the PMD thread must be made
678 * before setting the 'reload'. These changes could use any memory
679 * ordering model including 'relaxed'.
680 * * Setting the 'reload' atomic should occur in the same thread where
681 * all other PMD configuration options updated.
682 * * Setting the 'reload' atomic should be done with 'release' memory
683 * ordering model or stricter. This will guarantee that all previous
684 * changes (including non-atomic and 'relaxed') will be visible to
686 * * To check that reload is done, thread should poll the 'reload' atomic
687 * to become 'false'. Polling should be done with 'acquire' memory
688 * ordering model or stricter. This ensures that PMD thread completed
689 * the reload process.
691 * For the PMD thread:
693 * * PMD thread should read 'reload' atomic with 'acquire' memory
694 * ordering model or stricter. This will guarantee that all changes
695 * made before setting the 'reload' in the requesting thread will be
696 * visible to the PMD thread.
697 * * All other configuration data could be read with any memory
698 * ordering model (including non-atomic and 'relaxed') but *only after*
699 * reading the 'reload' atomic set to 'true'.
700 * * When the PMD reload done, PMD should (optionally) set all the below
701 * knobs except the 'reload' to their default ('false') values and
702 * (mandatory), as the last step, set the 'reload' to 'false' using
703 * 'release' memory ordering model or stricter. This will inform the
704 * requesting thread that PMD has completed a reload cycle.
706 atomic_bool reload
; /* Do we need to reload ports? */
707 atomic_bool wait_for_reload
; /* Can we busy wait for the next reload? */
708 atomic_bool reload_tx_qid
; /* Do we need to reload static_tx_qid? */
709 atomic_bool exit
; /* For terminating the pmd thread. */
712 unsigned core_id
; /* CPU core id of this pmd thread. */
713 int numa_id
; /* numa node id of this pmd thread. */
716 /* Queue id used by this pmd thread to send packets on all netdevs if
717 * XPS disabled for this netdev. All static_tx_qid's are unique and less
718 * than 'cmap_count(dp->poll_threads)'. */
719 uint32_t static_tx_qid
;
721 /* Number of filled output batches. */
722 int n_output_batches
;
724 struct ovs_mutex port_mutex
; /* Mutex for 'poll_list' and 'tx_ports'. */
725 /* List of rx queues to poll. */
726 struct hmap poll_list OVS_GUARDED
;
727 /* Map of 'tx_port's used for transmission. Written by the main thread,
728 * read by the pmd thread. */
729 struct hmap tx_ports OVS_GUARDED
;
731 /* These are thread-local copies of 'tx_ports'. One contains only tunnel
732 * ports (that support push_tunnel/pop_tunnel), the other contains ports
733 * with at least one txq (that support send). A port can be in both.
735 * There are two separate maps to make sure that we don't try to execute
736 * OUTPUT on a device which has 0 txqs or PUSH/POP on a non-tunnel device.
738 * The instances for cpu core NON_PMD_CORE_ID can be accessed by multiple
739 * threads, and thusly need to be protected by 'non_pmd_mutex'. Every
740 * other instance will only be accessed by its own pmd thread. */
741 struct hmap tnl_port_cache
;
742 struct hmap send_port_cache
;
744 /* Keep track of detailed PMD performance statistics. */
745 struct pmd_perf_stats perf_stats
;
747 /* Stats from previous iteration used by automatic pmd
748 * load balance logic. */
749 uint64_t prev_stats
[PMD_N_STATS
];
750 atomic_count pmd_overloaded
;
752 /* Set to true if the pmd thread needs to be reloaded. */
756 /* Interface to netdev-based datapath. */
759 struct dp_netdev
*dp
;
760 uint64_t last_port_seq
;
763 static int get_port_by_number(struct dp_netdev
*dp
, odp_port_t port_no
,
764 struct dp_netdev_port
**portp
)
765 OVS_REQUIRES(dp
->port_mutex
);
766 static int get_port_by_name(struct dp_netdev
*dp
, const char *devname
,
767 struct dp_netdev_port
**portp
)
768 OVS_REQUIRES(dp
->port_mutex
);
769 static void dp_netdev_free(struct dp_netdev
*)
770 OVS_REQUIRES(dp_netdev_mutex
);
771 static int do_add_port(struct dp_netdev
*dp
, const char *devname
,
772 const char *type
, odp_port_t port_no
)
773 OVS_REQUIRES(dp
->port_mutex
);
774 static void do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*)
775 OVS_REQUIRES(dp
->port_mutex
);
776 static int dpif_netdev_open(const struct dpif_class
*, const char *name
,
777 bool create
, struct dpif
**);
778 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
779 struct dp_packet_batch
*,
781 const struct flow
*flow
,
782 const struct nlattr
*actions
,
784 static void dp_netdev_input(struct dp_netdev_pmd_thread
*,
785 struct dp_packet_batch
*, odp_port_t port_no
);
786 static void dp_netdev_recirculate(struct dp_netdev_pmd_thread
*,
787 struct dp_packet_batch
*);
789 static void dp_netdev_disable_upcall(struct dp_netdev
*);
790 static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
);
791 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
,
792 struct dp_netdev
*dp
, unsigned core_id
,
794 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
);
795 static void dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
796 OVS_REQUIRES(dp
->port_mutex
);
798 static void *pmd_thread_main(void *);
799 static struct dp_netdev_pmd_thread
*dp_netdev_get_pmd(struct dp_netdev
*dp
,
801 static struct dp_netdev_pmd_thread
*
802 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
);
803 static void dp_netdev_del_pmd(struct dp_netdev
*dp
,
804 struct dp_netdev_pmd_thread
*pmd
);
805 static void dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
, bool non_pmd
);
806 static void dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
);
807 static void dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
808 struct dp_netdev_port
*port
)
809 OVS_REQUIRES(pmd
->port_mutex
);
810 static void dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
812 OVS_REQUIRES(pmd
->port_mutex
);
813 static void dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
814 struct dp_netdev_rxq
*rxq
)
815 OVS_REQUIRES(pmd
->port_mutex
);
816 static void dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
817 struct rxq_poll
*poll
)
818 OVS_REQUIRES(pmd
->port_mutex
);
820 dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread
*pmd
,
823 static void reconfigure_datapath(struct dp_netdev
*dp
)
824 OVS_REQUIRES(dp
->port_mutex
);
825 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
);
826 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
);
827 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
);
828 static void pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
829 OVS_REQUIRES(pmd
->port_mutex
);
831 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
,
832 struct polled_queue
*poll_list
, int poll_cnt
);
834 dp_netdev_rxq_set_cycles(struct dp_netdev_rxq
*rx
,
835 enum rxq_cycles_counter_type type
,
836 unsigned long long cycles
);
838 dp_netdev_rxq_get_cycles(struct dp_netdev_rxq
*rx
,
839 enum rxq_cycles_counter_type type
);
841 dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq
*rx
,
842 unsigned long long cycles
);
844 dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq
*rx
, unsigned idx
);
846 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
848 static int dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
851 static inline bool emc_entry_alive(struct emc_entry
*ce
);
852 static void emc_clear_entry(struct emc_entry
*ce
);
853 static void smc_clear_entry(struct smc_bucket
*b
, int idx
);
855 static void dp_netdev_request_reconfigure(struct dp_netdev
*dp
);
857 pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread
*pmd
);
858 static void queue_netdev_flow_del(struct dp_netdev_pmd_thread
*pmd
,
859 struct dp_netdev_flow
*flow
);
862 emc_cache_init(struct emc_cache
*flow_cache
)
866 flow_cache
->sweep_idx
= 0;
867 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
868 flow_cache
->entries
[i
].flow
= NULL
;
869 flow_cache
->entries
[i
].key
.hash
= 0;
870 flow_cache
->entries
[i
].key
.len
= sizeof(struct miniflow
);
871 flowmap_init(&flow_cache
->entries
[i
].key
.mf
.map
);
876 smc_cache_init(struct smc_cache
*smc_cache
)
879 for (i
= 0; i
< SMC_BUCKET_CNT
; i
++) {
880 for (j
= 0; j
< SMC_ENTRY_PER_BUCKET
; j
++) {
881 smc_cache
->buckets
[i
].flow_idx
[j
] = UINT16_MAX
;
887 dfc_cache_init(struct dfc_cache
*flow_cache
)
889 emc_cache_init(&flow_cache
->emc_cache
);
890 smc_cache_init(&flow_cache
->smc_cache
);
894 emc_cache_uninit(struct emc_cache
*flow_cache
)
898 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
899 emc_clear_entry(&flow_cache
->entries
[i
]);
904 smc_cache_uninit(struct smc_cache
*smc
)
908 for (i
= 0; i
< SMC_BUCKET_CNT
; i
++) {
909 for (j
= 0; j
< SMC_ENTRY_PER_BUCKET
; j
++) {
910 smc_clear_entry(&(smc
->buckets
[i
]), j
);
916 dfc_cache_uninit(struct dfc_cache
*flow_cache
)
918 smc_cache_uninit(&flow_cache
->smc_cache
);
919 emc_cache_uninit(&flow_cache
->emc_cache
);
922 /* Check and clear dead flow references slowly (one entry at each
925 emc_cache_slow_sweep(struct emc_cache
*flow_cache
)
927 struct emc_entry
*entry
= &flow_cache
->entries
[flow_cache
->sweep_idx
];
929 if (!emc_entry_alive(entry
)) {
930 emc_clear_entry(entry
);
932 flow_cache
->sweep_idx
= (flow_cache
->sweep_idx
+ 1) & EM_FLOW_HASH_MASK
;
935 /* Updates the time in PMD threads context and should be called in three cases:
937 * 1. PMD structure initialization:
938 * - dp_netdev_configure_pmd()
940 * 2. Before processing of the new packet batch:
941 * - dpif_netdev_execute()
942 * - dp_netdev_process_rxq_port()
944 * 3. At least once per polling iteration in main polling threads if no
945 * packets received on current iteration:
946 * - dpif_netdev_run()
947 * - pmd_thread_main()
949 * 'pmd->ctx.now' should be used without update in all other cases if possible.
952 pmd_thread_ctx_time_update(struct dp_netdev_pmd_thread
*pmd
)
954 pmd
->ctx
.now
= time_usec();
957 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
959 dpif_is_netdev(const struct dpif
*dpif
)
961 return dpif
->dpif_class
->open
== dpif_netdev_open
;
964 static struct dpif_netdev
*
965 dpif_netdev_cast(const struct dpif
*dpif
)
967 ovs_assert(dpif_is_netdev(dpif
));
968 return CONTAINER_OF(dpif
, struct dpif_netdev
, dpif
);
971 static struct dp_netdev
*
972 get_dp_netdev(const struct dpif
*dpif
)
974 return dpif_netdev_cast(dpif
)->dp
;
978 PMD_INFO_SHOW_STATS
, /* Show how cpu cycles are spent. */
979 PMD_INFO_CLEAR_STATS
, /* Set the cycles count to 0. */
980 PMD_INFO_SHOW_RXQ
, /* Show poll lists of pmd threads. */
981 PMD_INFO_PERF_SHOW
, /* Show pmd performance details. */
985 format_pmd_thread(struct ds
*reply
, struct dp_netdev_pmd_thread
*pmd
)
987 ds_put_cstr(reply
, (pmd
->core_id
== NON_PMD_CORE_ID
)
988 ? "main thread" : "pmd thread");
989 if (pmd
->numa_id
!= OVS_NUMA_UNSPEC
) {
990 ds_put_format(reply
, " numa_id %d", pmd
->numa_id
);
992 if (pmd
->core_id
!= OVS_CORE_UNSPEC
&& pmd
->core_id
!= NON_PMD_CORE_ID
) {
993 ds_put_format(reply
, " core_id %u", pmd
->core_id
);
995 ds_put_cstr(reply
, ":\n");
999 pmd_info_show_stats(struct ds
*reply
,
1000 struct dp_netdev_pmd_thread
*pmd
)
1002 uint64_t stats
[PMD_N_STATS
];
1003 uint64_t total_cycles
, total_packets
;
1004 double passes_per_pkt
= 0;
1005 double lookups_per_hit
= 0;
1006 double packets_per_batch
= 0;
1008 pmd_perf_read_counters(&pmd
->perf_stats
, stats
);
1009 total_cycles
= stats
[PMD_CYCLES_ITER_IDLE
]
1010 + stats
[PMD_CYCLES_ITER_BUSY
];
1011 total_packets
= stats
[PMD_STAT_RECV
];
1013 format_pmd_thread(reply
, pmd
);
1015 if (total_packets
> 0) {
1016 passes_per_pkt
= (total_packets
+ stats
[PMD_STAT_RECIRC
])
1017 / (double) total_packets
;
1019 if (stats
[PMD_STAT_MASKED_HIT
] > 0) {
1020 lookups_per_hit
= stats
[PMD_STAT_MASKED_LOOKUP
]
1021 / (double) stats
[PMD_STAT_MASKED_HIT
];
1023 if (stats
[PMD_STAT_SENT_BATCHES
] > 0) {
1024 packets_per_batch
= stats
[PMD_STAT_SENT_PKTS
]
1025 / (double) stats
[PMD_STAT_SENT_BATCHES
];
1028 ds_put_format(reply
,
1029 " packets received: %"PRIu64
"\n"
1030 " packet recirculations: %"PRIu64
"\n"
1031 " avg. datapath passes per packet: %.02f\n"
1032 " emc hits: %"PRIu64
"\n"
1033 " smc hits: %"PRIu64
"\n"
1034 " megaflow hits: %"PRIu64
"\n"
1035 " avg. subtable lookups per megaflow hit: %.02f\n"
1036 " miss with success upcall: %"PRIu64
"\n"
1037 " miss with failed upcall: %"PRIu64
"\n"
1038 " avg. packets per output batch: %.02f\n",
1039 total_packets
, stats
[PMD_STAT_RECIRC
],
1040 passes_per_pkt
, stats
[PMD_STAT_EXACT_HIT
],
1041 stats
[PMD_STAT_SMC_HIT
],
1042 stats
[PMD_STAT_MASKED_HIT
], lookups_per_hit
,
1043 stats
[PMD_STAT_MISS
], stats
[PMD_STAT_LOST
],
1046 if (total_cycles
== 0) {
1050 ds_put_format(reply
,
1051 " idle cycles: %"PRIu64
" (%.02f%%)\n"
1052 " processing cycles: %"PRIu64
" (%.02f%%)\n",
1053 stats
[PMD_CYCLES_ITER_IDLE
],
1054 stats
[PMD_CYCLES_ITER_IDLE
] / (double) total_cycles
* 100,
1055 stats
[PMD_CYCLES_ITER_BUSY
],
1056 stats
[PMD_CYCLES_ITER_BUSY
] / (double) total_cycles
* 100);
1058 if (total_packets
== 0) {
1062 ds_put_format(reply
,
1063 " avg cycles per packet: %.02f (%"PRIu64
"/%"PRIu64
")\n",
1064 total_cycles
/ (double) total_packets
,
1065 total_cycles
, total_packets
);
1067 ds_put_format(reply
,
1068 " avg processing cycles per packet: "
1069 "%.02f (%"PRIu64
"/%"PRIu64
")\n",
1070 stats
[PMD_CYCLES_ITER_BUSY
] / (double) total_packets
,
1071 stats
[PMD_CYCLES_ITER_BUSY
], total_packets
);
1075 pmd_info_show_perf(struct ds
*reply
,
1076 struct dp_netdev_pmd_thread
*pmd
,
1077 struct pmd_perf_params
*par
)
1079 if (pmd
->core_id
!= NON_PMD_CORE_ID
) {
1081 xastrftime_msec("%H:%M:%S.###", time_wall_msec(), true);
1082 long long now
= time_msec();
1083 double duration
= (now
- pmd
->perf_stats
.start_ms
) / 1000.0;
1085 ds_put_cstr(reply
, "\n");
1086 ds_put_format(reply
, "Time: %s\n", time_str
);
1087 ds_put_format(reply
, "Measurement duration: %.3f s\n", duration
);
1088 ds_put_cstr(reply
, "\n");
1089 format_pmd_thread(reply
, pmd
);
1090 ds_put_cstr(reply
, "\n");
1091 pmd_perf_format_overall_stats(reply
, &pmd
->perf_stats
, duration
);
1092 if (pmd_perf_metrics_enabled(pmd
)) {
1093 /* Prevent parallel clearing of perf metrics. */
1094 ovs_mutex_lock(&pmd
->perf_stats
.clear_mutex
);
1095 if (par
->histograms
) {
1096 ds_put_cstr(reply
, "\n");
1097 pmd_perf_format_histograms(reply
, &pmd
->perf_stats
);
1099 if (par
->iter_hist_len
> 0) {
1100 ds_put_cstr(reply
, "\n");
1101 pmd_perf_format_iteration_history(reply
, &pmd
->perf_stats
,
1102 par
->iter_hist_len
);
1104 if (par
->ms_hist_len
> 0) {
1105 ds_put_cstr(reply
, "\n");
1106 pmd_perf_format_ms_history(reply
, &pmd
->perf_stats
,
1109 ovs_mutex_unlock(&pmd
->perf_stats
.clear_mutex
);
1116 compare_poll_list(const void *a_
, const void *b_
)
1118 const struct rxq_poll
*a
= a_
;
1119 const struct rxq_poll
*b
= b_
;
1121 const char *namea
= netdev_rxq_get_name(a
->rxq
->rx
);
1122 const char *nameb
= netdev_rxq_get_name(b
->rxq
->rx
);
1124 int cmp
= strcmp(namea
, nameb
);
1126 return netdev_rxq_get_queue_id(a
->rxq
->rx
)
1127 - netdev_rxq_get_queue_id(b
->rxq
->rx
);
1134 sorted_poll_list(struct dp_netdev_pmd_thread
*pmd
, struct rxq_poll
**list
,
1136 OVS_REQUIRES(pmd
->port_mutex
)
1138 struct rxq_poll
*ret
, *poll
;
1141 *n
= hmap_count(&pmd
->poll_list
);
1145 ret
= xcalloc(*n
, sizeof *ret
);
1147 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
1151 ovs_assert(i
== *n
);
1152 qsort(ret
, *n
, sizeof *ret
, compare_poll_list
);
1159 pmd_info_show_rxq(struct ds
*reply
, struct dp_netdev_pmd_thread
*pmd
)
1161 if (pmd
->core_id
!= NON_PMD_CORE_ID
) {
1162 struct rxq_poll
*list
;
1164 uint64_t total_cycles
= 0;
1166 ds_put_format(reply
,
1167 "pmd thread numa_id %d core_id %u:\n isolated : %s\n",
1168 pmd
->numa_id
, pmd
->core_id
, (pmd
->isolated
)
1169 ? "true" : "false");
1171 ovs_mutex_lock(&pmd
->port_mutex
);
1172 sorted_poll_list(pmd
, &list
, &n_rxq
);
1174 /* Get the total pmd cycles for an interval. */
1175 atomic_read_relaxed(&pmd
->intrvl_cycles
, &total_cycles
);
1176 /* Estimate the cycles to cover all intervals. */
1177 total_cycles
*= PMD_RXQ_INTERVAL_MAX
;
1179 for (int i
= 0; i
< n_rxq
; i
++) {
1180 struct dp_netdev_rxq
*rxq
= list
[i
].rxq
;
1181 const char *name
= netdev_rxq_get_name(rxq
->rx
);
1182 uint64_t proc_cycles
= 0;
1184 for (int j
= 0; j
< PMD_RXQ_INTERVAL_MAX
; j
++) {
1185 proc_cycles
+= dp_netdev_rxq_get_intrvl_cycles(rxq
, j
);
1187 ds_put_format(reply
, " port: %-16s queue-id: %2d", name
,
1188 netdev_rxq_get_queue_id(list
[i
].rxq
->rx
));
1189 ds_put_format(reply
, " %s", netdev_rxq_enabled(list
[i
].rxq
->rx
)
1190 ? "(enabled) " : "(disabled)");
1191 ds_put_format(reply
, " pmd usage: ");
1193 ds_put_format(reply
, "%2"PRIu64
"",
1194 proc_cycles
* 100 / total_cycles
);
1195 ds_put_cstr(reply
, " %");
1197 ds_put_format(reply
, "%s", "NOT AVAIL");
1199 ds_put_cstr(reply
, "\n");
1201 ovs_mutex_unlock(&pmd
->port_mutex
);
1207 compare_poll_thread_list(const void *a_
, const void *b_
)
1209 const struct dp_netdev_pmd_thread
*a
, *b
;
1211 a
= *(struct dp_netdev_pmd_thread
**)a_
;
1212 b
= *(struct dp_netdev_pmd_thread
**)b_
;
1214 if (a
->core_id
< b
->core_id
) {
1217 if (a
->core_id
> b
->core_id
) {
1223 /* Create a sorted list of pmd's from the dp->poll_threads cmap. We can use
1224 * this list, as long as we do not go to quiescent state. */
1226 sorted_poll_thread_list(struct dp_netdev
*dp
,
1227 struct dp_netdev_pmd_thread
***list
,
1230 struct dp_netdev_pmd_thread
*pmd
;
1231 struct dp_netdev_pmd_thread
**pmd_list
;
1232 size_t k
= 0, n_pmds
;
1234 n_pmds
= cmap_count(&dp
->poll_threads
);
1235 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
1237 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1241 pmd_list
[k
++] = pmd
;
1244 qsort(pmd_list
, k
, sizeof *pmd_list
, compare_poll_thread_list
);
1251 dpif_netdev_pmd_rebalance(struct unixctl_conn
*conn
, int argc
,
1252 const char *argv
[], void *aux OVS_UNUSED
)
1254 struct ds reply
= DS_EMPTY_INITIALIZER
;
1255 struct dp_netdev
*dp
= NULL
;
1257 ovs_mutex_lock(&dp_netdev_mutex
);
1260 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
1261 } else if (shash_count(&dp_netdevs
) == 1) {
1262 /* There's only one datapath */
1263 dp
= shash_first(&dp_netdevs
)->data
;
1267 ovs_mutex_unlock(&dp_netdev_mutex
);
1268 unixctl_command_reply_error(conn
,
1269 "please specify an existing datapath");
1273 dp_netdev_request_reconfigure(dp
);
1274 ovs_mutex_unlock(&dp_netdev_mutex
);
1275 ds_put_cstr(&reply
, "pmd rxq rebalance requested.\n");
1276 unixctl_command_reply(conn
, ds_cstr(&reply
));
1281 dpif_netdev_pmd_info(struct unixctl_conn
*conn
, int argc
, const char *argv
[],
1284 struct ds reply
= DS_EMPTY_INITIALIZER
;
1285 struct dp_netdev_pmd_thread
**pmd_list
;
1286 struct dp_netdev
*dp
= NULL
;
1287 enum pmd_info_type type
= *(enum pmd_info_type
*) aux
;
1288 unsigned int core_id
;
1289 bool filter_on_pmd
= false;
1292 ovs_mutex_lock(&dp_netdev_mutex
);
1295 if (!strcmp(argv
[1], "-pmd") && argc
> 2) {
1296 if (str_to_uint(argv
[2], 10, &core_id
)) {
1297 filter_on_pmd
= true;
1302 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
1309 if (shash_count(&dp_netdevs
) == 1) {
1310 /* There's only one datapath */
1311 dp
= shash_first(&dp_netdevs
)->data
;
1313 ovs_mutex_unlock(&dp_netdev_mutex
);
1314 unixctl_command_reply_error(conn
,
1315 "please specify an existing datapath");
1320 sorted_poll_thread_list(dp
, &pmd_list
, &n
);
1321 for (size_t i
= 0; i
< n
; i
++) {
1322 struct dp_netdev_pmd_thread
*pmd
= pmd_list
[i
];
1326 if (filter_on_pmd
&& pmd
->core_id
!= core_id
) {
1329 if (type
== PMD_INFO_SHOW_RXQ
) {
1330 pmd_info_show_rxq(&reply
, pmd
);
1331 } else if (type
== PMD_INFO_CLEAR_STATS
) {
1332 pmd_perf_stats_clear(&pmd
->perf_stats
);
1333 } else if (type
== PMD_INFO_SHOW_STATS
) {
1334 pmd_info_show_stats(&reply
, pmd
);
1335 } else if (type
== PMD_INFO_PERF_SHOW
) {
1336 pmd_info_show_perf(&reply
, pmd
, (struct pmd_perf_params
*)aux
);
1341 ovs_mutex_unlock(&dp_netdev_mutex
);
1343 unixctl_command_reply(conn
, ds_cstr(&reply
));
1348 pmd_perf_show_cmd(struct unixctl_conn
*conn
, int argc
,
1350 void *aux OVS_UNUSED
)
1352 struct pmd_perf_params par
;
1353 long int it_hist
= 0, ms_hist
= 0;
1354 par
.histograms
= true;
1357 if (!strcmp(argv
[1], "-nh")) {
1358 par
.histograms
= false;
1361 } else if (!strcmp(argv
[1], "-it") && argc
> 2) {
1362 it_hist
= strtol(argv
[2], NULL
, 10);
1365 } else if (it_hist
> HISTORY_LEN
) {
1366 it_hist
= HISTORY_LEN
;
1370 } else if (!strcmp(argv
[1], "-ms") && argc
> 2) {
1371 ms_hist
= strtol(argv
[2], NULL
, 10);
1374 } else if (ms_hist
> HISTORY_LEN
) {
1375 ms_hist
= HISTORY_LEN
;
1383 par
.iter_hist_len
= it_hist
;
1384 par
.ms_hist_len
= ms_hist
;
1385 par
.command_type
= PMD_INFO_PERF_SHOW
;
1386 dpif_netdev_pmd_info(conn
, argc
, argv
, &par
);
1390 dpif_netdev_init(void)
1392 static enum pmd_info_type show_aux
= PMD_INFO_SHOW_STATS
,
1393 clear_aux
= PMD_INFO_CLEAR_STATS
,
1394 poll_aux
= PMD_INFO_SHOW_RXQ
;
1396 unixctl_command_register("dpif-netdev/pmd-stats-show", "[-pmd core] [dp]",
1397 0, 3, dpif_netdev_pmd_info
,
1399 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[-pmd core] [dp]",
1400 0, 3, dpif_netdev_pmd_info
,
1401 (void *)&clear_aux
);
1402 unixctl_command_register("dpif-netdev/pmd-rxq-show", "[-pmd core] [dp]",
1403 0, 3, dpif_netdev_pmd_info
,
1405 unixctl_command_register("dpif-netdev/pmd-perf-show",
1406 "[-nh] [-it iter-history-len]"
1407 " [-ms ms-history-len]"
1408 " [-pmd core] [dp]",
1409 0, 8, pmd_perf_show_cmd
,
1411 unixctl_command_register("dpif-netdev/pmd-rxq-rebalance", "[dp]",
1412 0, 1, dpif_netdev_pmd_rebalance
,
1414 unixctl_command_register("dpif-netdev/pmd-perf-log-set",
1415 "on|off [-b before] [-a after] [-e|-ne] "
1416 "[-us usec] [-q qlen]",
1417 0, 10, pmd_perf_log_set_cmd
,
1423 dpif_netdev_enumerate(struct sset
*all_dps
,
1424 const struct dpif_class
*dpif_class
)
1426 struct shash_node
*node
;
1428 ovs_mutex_lock(&dp_netdev_mutex
);
1429 SHASH_FOR_EACH(node
, &dp_netdevs
) {
1430 struct dp_netdev
*dp
= node
->data
;
1431 if (dpif_class
!= dp
->class) {
1432 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
1433 * If the class doesn't match, skip this dpif. */
1436 sset_add(all_dps
, node
->name
);
1438 ovs_mutex_unlock(&dp_netdev_mutex
);
1444 dpif_netdev_class_is_dummy(const struct dpif_class
*class)
1446 return class != &dpif_netdev_class
;
1450 dpif_netdev_port_open_type(const struct dpif_class
*class, const char *type
)
1452 return strcmp(type
, "internal") ? type
1453 : dpif_netdev_class_is_dummy(class) ? "dummy-internal"
1457 static struct dpif
*
1458 create_dpif_netdev(struct dp_netdev
*dp
)
1460 uint16_t netflow_id
= hash_string(dp
->name
, 0);
1461 struct dpif_netdev
*dpif
;
1463 ovs_refcount_ref(&dp
->ref_cnt
);
1465 dpif
= xmalloc(sizeof *dpif
);
1466 dpif_init(&dpif
->dpif
, dp
->class, dp
->name
, netflow_id
>> 8, netflow_id
);
1468 dpif
->last_port_seq
= seq_read(dp
->port_seq
);
1473 /* Choose an unused, non-zero port number and return it on success.
1474 * Return ODPP_NONE on failure. */
1476 choose_port(struct dp_netdev
*dp
, const char *name
)
1477 OVS_REQUIRES(dp
->port_mutex
)
1481 if (dp
->class != &dpif_netdev_class
) {
1485 /* If the port name begins with "br", start the number search at
1486 * 100 to make writing tests easier. */
1487 if (!strncmp(name
, "br", 2)) {
1491 /* If the port name contains a number, try to assign that port number.
1492 * This can make writing unit tests easier because port numbers are
1494 for (p
= name
; *p
!= '\0'; p
++) {
1495 if (isdigit((unsigned char) *p
)) {
1496 port_no
= start_no
+ strtol(p
, NULL
, 10);
1497 if (port_no
> 0 && port_no
!= odp_to_u32(ODPP_NONE
)
1498 && !dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
1499 return u32_to_odp(port_no
);
1506 for (port_no
= 1; port_no
<= UINT16_MAX
; port_no
++) {
1507 if (!dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
1508 return u32_to_odp(port_no
);
1516 create_dp_netdev(const char *name
, const struct dpif_class
*class,
1517 struct dp_netdev
**dpp
)
1518 OVS_REQUIRES(dp_netdev_mutex
)
1520 struct dp_netdev
*dp
;
1523 dp
= xzalloc(sizeof *dp
);
1524 shash_add(&dp_netdevs
, name
, dp
);
1526 *CONST_CAST(const struct dpif_class
**, &dp
->class) = class;
1527 *CONST_CAST(const char **, &dp
->name
) = xstrdup(name
);
1528 ovs_refcount_init(&dp
->ref_cnt
);
1529 atomic_flag_clear(&dp
->destroyed
);
1531 ovs_mutex_init(&dp
->port_mutex
);
1532 hmap_init(&dp
->ports
);
1533 dp
->port_seq
= seq_create();
1534 fat_rwlock_init(&dp
->upcall_rwlock
);
1536 dp
->reconfigure_seq
= seq_create();
1537 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
1539 for (int i
= 0; i
< N_METER_LOCKS
; ++i
) {
1540 ovs_mutex_init_adaptive(&dp
->meter_locks
[i
]);
1543 /* Disable upcalls by default. */
1544 dp_netdev_disable_upcall(dp
);
1545 dp
->upcall_aux
= NULL
;
1546 dp
->upcall_cb
= NULL
;
1548 dp
->conntrack
= conntrack_init();
1550 atomic_init(&dp
->emc_insert_min
, DEFAULT_EM_FLOW_INSERT_MIN
);
1551 atomic_init(&dp
->tx_flush_interval
, DEFAULT_TX_FLUSH_INTERVAL
);
1553 cmap_init(&dp
->poll_threads
);
1554 dp
->pmd_rxq_assign_cyc
= true;
1556 ovs_mutex_init(&dp
->tx_qid_pool_mutex
);
1557 /* We need 1 Tx queue for each possible core + 1 for non-PMD threads. */
1558 dp
->tx_qid_pool
= id_pool_create(0, ovs_numa_get_n_cores() + 1);
1560 ovs_mutex_init_recursive(&dp
->non_pmd_mutex
);
1561 ovsthread_key_create(&dp
->per_pmd_key
, NULL
);
1563 ovs_mutex_lock(&dp
->port_mutex
);
1564 /* non-PMD will be created before all other threads and will
1565 * allocate static_tx_qid = 0. */
1566 dp_netdev_set_nonpmd(dp
);
1568 error
= do_add_port(dp
, name
, dpif_netdev_port_open_type(dp
->class,
1571 ovs_mutex_unlock(&dp
->port_mutex
);
1577 dp
->last_tnl_conf_seq
= seq_read(tnl_conf_seq
);
1583 dp_netdev_request_reconfigure(struct dp_netdev
*dp
)
1585 seq_change(dp
->reconfigure_seq
);
1589 dp_netdev_is_reconf_required(struct dp_netdev
*dp
)
1591 return seq_read(dp
->reconfigure_seq
) != dp
->last_reconfigure_seq
;
1595 dpif_netdev_open(const struct dpif_class
*class, const char *name
,
1596 bool create
, struct dpif
**dpifp
)
1598 struct dp_netdev
*dp
;
1601 ovs_mutex_lock(&dp_netdev_mutex
);
1602 dp
= shash_find_data(&dp_netdevs
, name
);
1604 error
= create
? create_dp_netdev(name
, class, &dp
) : ENODEV
;
1606 error
= (dp
->class != class ? EINVAL
1611 *dpifp
= create_dpif_netdev(dp
);
1614 ovs_mutex_unlock(&dp_netdev_mutex
);
1620 dp_netdev_destroy_upcall_lock(struct dp_netdev
*dp
)
1621 OVS_NO_THREAD_SAFETY_ANALYSIS
1623 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
1624 ovs_assert(fat_rwlock_tryrdlock(&dp
->upcall_rwlock
));
1626 /* Before freeing a lock we should release it */
1627 fat_rwlock_unlock(&dp
->upcall_rwlock
);
1628 fat_rwlock_destroy(&dp
->upcall_rwlock
);
1632 dp_delete_meter(struct dp_netdev
*dp
, uint32_t meter_id
)
1633 OVS_REQUIRES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
1635 if (dp
->meters
[meter_id
]) {
1636 free(dp
->meters
[meter_id
]);
1637 dp
->meters
[meter_id
] = NULL
;
1641 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
1642 * through the 'dp_netdevs' shash while freeing 'dp'. */
1644 dp_netdev_free(struct dp_netdev
*dp
)
1645 OVS_REQUIRES(dp_netdev_mutex
)
1647 struct dp_netdev_port
*port
, *next
;
1649 shash_find_and_delete(&dp_netdevs
, dp
->name
);
1651 ovs_mutex_lock(&dp
->port_mutex
);
1652 HMAP_FOR_EACH_SAFE (port
, next
, node
, &dp
->ports
) {
1653 do_del_port(dp
, port
);
1655 ovs_mutex_unlock(&dp
->port_mutex
);
1657 dp_netdev_destroy_all_pmds(dp
, true);
1658 cmap_destroy(&dp
->poll_threads
);
1660 ovs_mutex_destroy(&dp
->tx_qid_pool_mutex
);
1661 id_pool_destroy(dp
->tx_qid_pool
);
1663 ovs_mutex_destroy(&dp
->non_pmd_mutex
);
1664 ovsthread_key_delete(dp
->per_pmd_key
);
1666 conntrack_destroy(dp
->conntrack
);
1669 seq_destroy(dp
->reconfigure_seq
);
1671 seq_destroy(dp
->port_seq
);
1672 hmap_destroy(&dp
->ports
);
1673 ovs_mutex_destroy(&dp
->port_mutex
);
1675 /* Upcalls must be disabled at this point */
1676 dp_netdev_destroy_upcall_lock(dp
);
1680 for (i
= 0; i
< MAX_METERS
; ++i
) {
1682 dp_delete_meter(dp
, i
);
1683 meter_unlock(dp
, i
);
1685 for (i
= 0; i
< N_METER_LOCKS
; ++i
) {
1686 ovs_mutex_destroy(&dp
->meter_locks
[i
]);
1689 free(dp
->pmd_cmask
);
1690 free(CONST_CAST(char *, dp
->name
));
1695 dp_netdev_unref(struct dp_netdev
*dp
)
1698 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
1699 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
1700 ovs_mutex_lock(&dp_netdev_mutex
);
1701 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1704 ovs_mutex_unlock(&dp_netdev_mutex
);
1709 dpif_netdev_close(struct dpif
*dpif
)
1711 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1713 dp_netdev_unref(dp
);
1718 dpif_netdev_destroy(struct dpif
*dpif
)
1720 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1722 if (!atomic_flag_test_and_set(&dp
->destroyed
)) {
1723 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1724 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
1732 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
1733 * load/store semantics. While the increment is not atomic, the load and
1734 * store operations are, making it impossible to read inconsistent values.
1736 * This is used to update thread local stats counters. */
1738 non_atomic_ullong_add(atomic_ullong
*var
, unsigned long long n
)
1740 unsigned long long tmp
;
1742 atomic_read_relaxed(var
, &tmp
);
1744 atomic_store_relaxed(var
, tmp
);
1748 dpif_netdev_get_stats(const struct dpif
*dpif
, struct dpif_dp_stats
*stats
)
1750 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1751 struct dp_netdev_pmd_thread
*pmd
;
1752 uint64_t pmd_stats
[PMD_N_STATS
];
1754 stats
->n_flows
= stats
->n_hit
= stats
->n_missed
= stats
->n_lost
= 0;
1755 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1756 stats
->n_flows
+= cmap_count(&pmd
->flow_table
);
1757 pmd_perf_read_counters(&pmd
->perf_stats
, pmd_stats
);
1758 stats
->n_hit
+= pmd_stats
[PMD_STAT_EXACT_HIT
];
1759 stats
->n_hit
+= pmd_stats
[PMD_STAT_SMC_HIT
];
1760 stats
->n_hit
+= pmd_stats
[PMD_STAT_MASKED_HIT
];
1761 stats
->n_missed
+= pmd_stats
[PMD_STAT_MISS
];
1762 stats
->n_lost
+= pmd_stats
[PMD_STAT_LOST
];
1764 stats
->n_masks
= UINT32_MAX
;
1765 stats
->n_mask_hit
= UINT64_MAX
;
1771 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread
*pmd
)
1773 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
1774 ovs_mutex_lock(&pmd
->dp
->non_pmd_mutex
);
1775 ovs_mutex_lock(&pmd
->port_mutex
);
1776 pmd_load_cached_ports(pmd
);
1777 ovs_mutex_unlock(&pmd
->port_mutex
);
1778 ovs_mutex_unlock(&pmd
->dp
->non_pmd_mutex
);
1782 seq_change(pmd
->reload_seq
);
1783 atomic_store_explicit(&pmd
->reload
, true, memory_order_release
);
1787 hash_port_no(odp_port_t port_no
)
1789 return hash_int(odp_to_u32(port_no
), 0);
1793 port_create(const char *devname
, const char *type
,
1794 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1796 struct netdev_saved_flags
*sf
;
1797 struct dp_netdev_port
*port
;
1798 enum netdev_flags flags
;
1799 struct netdev
*netdev
;
1804 /* Open and validate network device. */
1805 error
= netdev_open(devname
, type
, &netdev
);
1809 /* XXX reject non-Ethernet devices */
1811 netdev_get_flags(netdev
, &flags
);
1812 if (flags
& NETDEV_LOOPBACK
) {
1813 VLOG_ERR("%s: cannot add a loopback device", devname
);
1818 error
= netdev_turn_flags_on(netdev
, NETDEV_PROMISC
, &sf
);
1820 VLOG_ERR("%s: cannot set promisc flag", devname
);
1824 port
= xzalloc(sizeof *port
);
1825 port
->port_no
= port_no
;
1826 port
->netdev
= netdev
;
1827 port
->type
= xstrdup(type
);
1829 port
->emc_enabled
= true;
1830 port
->need_reconfigure
= true;
1831 ovs_mutex_init(&port
->txq_used_mutex
);
1838 netdev_close(netdev
);
1843 do_add_port(struct dp_netdev
*dp
, const char *devname
, const char *type
,
1845 OVS_REQUIRES(dp
->port_mutex
)
1847 struct dp_netdev_port
*port
;
1850 /* Reject devices already in 'dp'. */
1851 if (!get_port_by_name(dp
, devname
, &port
)) {
1855 error
= port_create(devname
, type
, port_no
, &port
);
1860 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
1861 seq_change(dp
->port_seq
);
1863 reconfigure_datapath(dp
);
1869 dpif_netdev_port_add(struct dpif
*dpif
, struct netdev
*netdev
,
1870 odp_port_t
*port_nop
)
1872 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1873 char namebuf
[NETDEV_VPORT_NAME_BUFSIZE
];
1874 const char *dpif_port
;
1878 ovs_mutex_lock(&dp
->port_mutex
);
1879 dpif_port
= netdev_vport_get_dpif_port(netdev
, namebuf
, sizeof namebuf
);
1880 if (*port_nop
!= ODPP_NONE
) {
1881 port_no
= *port_nop
;
1882 error
= dp_netdev_lookup_port(dp
, *port_nop
) ? EBUSY
: 0;
1884 port_no
= choose_port(dp
, dpif_port
);
1885 error
= port_no
== ODPP_NONE
? EFBIG
: 0;
1888 *port_nop
= port_no
;
1889 error
= do_add_port(dp
, dpif_port
, netdev_get_type(netdev
), port_no
);
1891 ovs_mutex_unlock(&dp
->port_mutex
);
1897 dpif_netdev_port_del(struct dpif
*dpif
, odp_port_t port_no
)
1899 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1902 ovs_mutex_lock(&dp
->port_mutex
);
1903 if (port_no
== ODPP_LOCAL
) {
1906 struct dp_netdev_port
*port
;
1908 error
= get_port_by_number(dp
, port_no
, &port
);
1910 do_del_port(dp
, port
);
1913 ovs_mutex_unlock(&dp
->port_mutex
);
1919 is_valid_port_number(odp_port_t port_no
)
1921 return port_no
!= ODPP_NONE
;
1924 static struct dp_netdev_port
*
1925 dp_netdev_lookup_port(const struct dp_netdev
*dp
, odp_port_t port_no
)
1926 OVS_REQUIRES(dp
->port_mutex
)
1928 struct dp_netdev_port
*port
;
1930 HMAP_FOR_EACH_WITH_HASH (port
, node
, hash_port_no(port_no
), &dp
->ports
) {
1931 if (port
->port_no
== port_no
) {
1939 get_port_by_number(struct dp_netdev
*dp
,
1940 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1941 OVS_REQUIRES(dp
->port_mutex
)
1943 if (!is_valid_port_number(port_no
)) {
1947 *portp
= dp_netdev_lookup_port(dp
, port_no
);
1948 return *portp
? 0 : ENODEV
;
1953 port_destroy(struct dp_netdev_port
*port
)
1959 netdev_close(port
->netdev
);
1960 netdev_restore_flags(port
->sf
);
1962 for (unsigned i
= 0; i
< port
->n_rxq
; i
++) {
1963 netdev_rxq_close(port
->rxqs
[i
].rx
);
1965 ovs_mutex_destroy(&port
->txq_used_mutex
);
1966 free(port
->rxq_affinity_list
);
1967 free(port
->txq_used
);
1974 get_port_by_name(struct dp_netdev
*dp
,
1975 const char *devname
, struct dp_netdev_port
**portp
)
1976 OVS_REQUIRES(dp
->port_mutex
)
1978 struct dp_netdev_port
*port
;
1980 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1981 if (!strcmp(netdev_get_name(port
->netdev
), devname
)) {
1987 /* Callers of dpif_netdev_port_query_by_name() expect ENODEV for a non
1992 /* Returns 'true' if there is a port with pmd netdev. */
1994 has_pmd_port(struct dp_netdev
*dp
)
1995 OVS_REQUIRES(dp
->port_mutex
)
1997 struct dp_netdev_port
*port
;
1999 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2000 if (netdev_is_pmd(port
->netdev
)) {
2009 do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*port
)
2010 OVS_REQUIRES(dp
->port_mutex
)
2012 hmap_remove(&dp
->ports
, &port
->node
);
2013 seq_change(dp
->port_seq
);
2015 reconfigure_datapath(dp
);
2021 answer_port_query(const struct dp_netdev_port
*port
,
2022 struct dpif_port
*dpif_port
)
2024 dpif_port
->name
= xstrdup(netdev_get_name(port
->netdev
));
2025 dpif_port
->type
= xstrdup(port
->type
);
2026 dpif_port
->port_no
= port
->port_no
;
2030 dpif_netdev_port_query_by_number(const struct dpif
*dpif
, odp_port_t port_no
,
2031 struct dpif_port
*dpif_port
)
2033 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2034 struct dp_netdev_port
*port
;
2037 ovs_mutex_lock(&dp
->port_mutex
);
2038 error
= get_port_by_number(dp
, port_no
, &port
);
2039 if (!error
&& dpif_port
) {
2040 answer_port_query(port
, dpif_port
);
2042 ovs_mutex_unlock(&dp
->port_mutex
);
2048 dpif_netdev_port_query_by_name(const struct dpif
*dpif
, const char *devname
,
2049 struct dpif_port
*dpif_port
)
2051 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2052 struct dp_netdev_port
*port
;
2055 ovs_mutex_lock(&dp
->port_mutex
);
2056 error
= get_port_by_name(dp
, devname
, &port
);
2057 if (!error
&& dpif_port
) {
2058 answer_port_query(port
, dpif_port
);
2060 ovs_mutex_unlock(&dp
->port_mutex
);
2066 dp_netdev_flow_free(struct dp_netdev_flow
*flow
)
2068 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow
));
2072 static void dp_netdev_flow_unref(struct dp_netdev_flow
*flow
)
2074 if (ovs_refcount_unref_relaxed(&flow
->ref_cnt
) == 1) {
2075 ovsrcu_postpone(dp_netdev_flow_free
, flow
);
2080 dp_netdev_flow_hash(const ovs_u128
*ufid
)
2082 return ufid
->u32
[0];
2085 static inline struct dpcls
*
2086 dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread
*pmd
,
2090 uint32_t hash
= hash_port_no(in_port
);
2091 CMAP_FOR_EACH_WITH_HASH (cls
, node
, hash
, &pmd
->classifiers
) {
2092 if (cls
->in_port
== in_port
) {
2093 /* Port classifier exists already */
2100 static inline struct dpcls
*
2101 dp_netdev_pmd_find_dpcls(struct dp_netdev_pmd_thread
*pmd
,
2103 OVS_REQUIRES(pmd
->flow_mutex
)
2105 struct dpcls
*cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2106 uint32_t hash
= hash_port_no(in_port
);
2109 /* Create new classifier for in_port */
2110 cls
= xmalloc(sizeof(*cls
));
2112 cls
->in_port
= in_port
;
2113 cmap_insert(&pmd
->classifiers
, &cls
->node
, hash
);
2114 VLOG_DBG("Creating dpcls %p for in_port %d", cls
, in_port
);
2119 #define MAX_FLOW_MARK (UINT32_MAX - 1)
2120 #define INVALID_FLOW_MARK (UINT32_MAX)
2122 struct megaflow_to_mark_data
{
2123 const struct cmap_node node
;
2129 struct cmap megaflow_to_mark
;
2130 struct cmap mark_to_flow
;
2131 struct id_pool
*pool
;
2134 static struct flow_mark flow_mark
= {
2135 .megaflow_to_mark
= CMAP_INITIALIZER
,
2136 .mark_to_flow
= CMAP_INITIALIZER
,
2140 flow_mark_alloc(void)
2144 if (!flow_mark
.pool
) {
2145 /* Haven't initiated yet, do it here */
2146 flow_mark
.pool
= id_pool_create(0, MAX_FLOW_MARK
);
2149 if (id_pool_alloc_id(flow_mark
.pool
, &mark
)) {
2153 return INVALID_FLOW_MARK
;
2157 flow_mark_free(uint32_t mark
)
2159 id_pool_free_id(flow_mark
.pool
, mark
);
2162 /* associate megaflow with a mark, which is a 1:1 mapping */
2164 megaflow_to_mark_associate(const ovs_u128
*mega_ufid
, uint32_t mark
)
2166 size_t hash
= dp_netdev_flow_hash(mega_ufid
);
2167 struct megaflow_to_mark_data
*data
= xzalloc(sizeof(*data
));
2169 data
->mega_ufid
= *mega_ufid
;
2172 cmap_insert(&flow_mark
.megaflow_to_mark
,
2173 CONST_CAST(struct cmap_node
*, &data
->node
), hash
);
2176 /* disassociate meagaflow with a mark */
2178 megaflow_to_mark_disassociate(const ovs_u128
*mega_ufid
)
2180 size_t hash
= dp_netdev_flow_hash(mega_ufid
);
2181 struct megaflow_to_mark_data
*data
;
2183 CMAP_FOR_EACH_WITH_HASH (data
, node
, hash
, &flow_mark
.megaflow_to_mark
) {
2184 if (ovs_u128_equals(*mega_ufid
, data
->mega_ufid
)) {
2185 cmap_remove(&flow_mark
.megaflow_to_mark
,
2186 CONST_CAST(struct cmap_node
*, &data
->node
), hash
);
2187 ovsrcu_postpone(free
, data
);
2192 VLOG_WARN("Masked ufid "UUID_FMT
" is not associated with a mark?\n",
2193 UUID_ARGS((struct uuid
*)mega_ufid
));
2196 static inline uint32_t
2197 megaflow_to_mark_find(const ovs_u128
*mega_ufid
)
2199 size_t hash
= dp_netdev_flow_hash(mega_ufid
);
2200 struct megaflow_to_mark_data
*data
;
2202 CMAP_FOR_EACH_WITH_HASH (data
, node
, hash
, &flow_mark
.megaflow_to_mark
) {
2203 if (ovs_u128_equals(*mega_ufid
, data
->mega_ufid
)) {
2208 VLOG_DBG("Mark id for ufid "UUID_FMT
" was not found\n",
2209 UUID_ARGS((struct uuid
*)mega_ufid
));
2210 return INVALID_FLOW_MARK
;
2213 /* associate mark with a flow, which is 1:N mapping */
2215 mark_to_flow_associate(const uint32_t mark
, struct dp_netdev_flow
*flow
)
2217 dp_netdev_flow_ref(flow
);
2219 cmap_insert(&flow_mark
.mark_to_flow
,
2220 CONST_CAST(struct cmap_node
*, &flow
->mark_node
),
2224 VLOG_DBG("Associated dp_netdev flow %p with mark %u\n", flow
, mark
);
2228 flow_mark_has_no_ref(uint32_t mark
)
2230 struct dp_netdev_flow
*flow
;
2232 CMAP_FOR_EACH_WITH_HASH (flow
, mark_node
, hash_int(mark
, 0),
2233 &flow_mark
.mark_to_flow
) {
2234 if (flow
->mark
== mark
) {
2243 mark_to_flow_disassociate(struct dp_netdev_pmd_thread
*pmd
,
2244 struct dp_netdev_flow
*flow
)
2247 uint32_t mark
= flow
->mark
;
2248 struct cmap_node
*mark_node
= CONST_CAST(struct cmap_node
*,
2251 cmap_remove(&flow_mark
.mark_to_flow
, mark_node
, hash_int(mark
, 0));
2252 flow
->mark
= INVALID_FLOW_MARK
;
2255 * no flow is referencing the mark any more? If so, let's
2256 * remove the flow from hardware and free the mark.
2258 if (flow_mark_has_no_ref(mark
)) {
2259 struct dp_netdev_port
*port
;
2260 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
2262 ovs_mutex_lock(&pmd
->dp
->port_mutex
);
2263 port
= dp_netdev_lookup_port(pmd
->dp
, in_port
);
2265 ret
= netdev_flow_del(port
->netdev
, &flow
->mega_ufid
, NULL
);
2267 ovs_mutex_unlock(&pmd
->dp
->port_mutex
);
2269 flow_mark_free(mark
);
2270 VLOG_DBG("Freed flow mark %u\n", mark
);
2272 megaflow_to_mark_disassociate(&flow
->mega_ufid
);
2274 dp_netdev_flow_unref(flow
);
2280 flow_mark_flush(struct dp_netdev_pmd_thread
*pmd
)
2282 struct dp_netdev_flow
*flow
;
2284 CMAP_FOR_EACH (flow
, mark_node
, &flow_mark
.mark_to_flow
) {
2285 if (flow
->pmd_id
== pmd
->core_id
) {
2286 queue_netdev_flow_del(pmd
, flow
);
2291 static struct dp_netdev_flow
*
2292 mark_to_flow_find(const struct dp_netdev_pmd_thread
*pmd
,
2293 const uint32_t mark
)
2295 struct dp_netdev_flow
*flow
;
2297 CMAP_FOR_EACH_WITH_HASH (flow
, mark_node
, hash_int(mark
, 0),
2298 &flow_mark
.mark_to_flow
) {
2299 if (flow
->mark
== mark
&& flow
->pmd_id
== pmd
->core_id
&&
2300 flow
->dead
== false) {
2308 static struct dp_flow_offload_item
*
2309 dp_netdev_alloc_flow_offload(struct dp_netdev_pmd_thread
*pmd
,
2310 struct dp_netdev_flow
*flow
,
2313 struct dp_flow_offload_item
*offload
;
2315 offload
= xzalloc(sizeof(*offload
));
2317 offload
->flow
= flow
;
2320 dp_netdev_flow_ref(flow
);
2321 dp_netdev_pmd_try_ref(pmd
);
2327 dp_netdev_free_flow_offload(struct dp_flow_offload_item
*offload
)
2329 dp_netdev_pmd_unref(offload
->pmd
);
2330 dp_netdev_flow_unref(offload
->flow
);
2332 free(offload
->actions
);
2337 dp_netdev_append_flow_offload(struct dp_flow_offload_item
*offload
)
2339 ovs_mutex_lock(&dp_flow_offload
.mutex
);
2340 ovs_list_push_back(&dp_flow_offload
.list
, &offload
->node
);
2341 xpthread_cond_signal(&dp_flow_offload
.cond
);
2342 ovs_mutex_unlock(&dp_flow_offload
.mutex
);
2346 dp_netdev_flow_offload_del(struct dp_flow_offload_item
*offload
)
2348 return mark_to_flow_disassociate(offload
->pmd
, offload
->flow
);
2352 * There are two flow offload operations here: addition and modification.
2354 * For flow addition, this function does:
2355 * - allocate a new flow mark id
2356 * - perform hardware flow offload
2357 * - associate the flow mark with flow and mega flow
2359 * For flow modification, both flow mark and the associations are still
2360 * valid, thus only item 2 needed.
2363 dp_netdev_flow_offload_put(struct dp_flow_offload_item
*offload
)
2365 struct dp_netdev_port
*port
;
2366 struct dp_netdev_pmd_thread
*pmd
= offload
->pmd
;
2367 struct dp_netdev_flow
*flow
= offload
->flow
;
2368 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
2369 bool modification
= offload
->op
== DP_NETDEV_FLOW_OFFLOAD_OP_MOD
;
2370 struct offload_info info
;
2380 ovs_assert(mark
!= INVALID_FLOW_MARK
);
2383 * If a mega flow has already been offloaded (from other PMD
2384 * instances), do not offload it again.
2386 mark
= megaflow_to_mark_find(&flow
->mega_ufid
);
2387 if (mark
!= INVALID_FLOW_MARK
) {
2388 VLOG_DBG("Flow has already been offloaded with mark %u\n", mark
);
2389 if (flow
->mark
!= INVALID_FLOW_MARK
) {
2390 ovs_assert(flow
->mark
== mark
);
2392 mark_to_flow_associate(mark
, flow
);
2397 mark
= flow_mark_alloc();
2398 if (mark
== INVALID_FLOW_MARK
) {
2399 VLOG_ERR("Failed to allocate flow mark!\n");
2402 info
.flow_mark
= mark
;
2404 ovs_mutex_lock(&pmd
->dp
->port_mutex
);
2405 port
= dp_netdev_lookup_port(pmd
->dp
, in_port
);
2406 if (!port
|| netdev_vport_is_vport_class(port
->netdev
->netdev_class
)) {
2407 ovs_mutex_unlock(&pmd
->dp
->port_mutex
);
2410 ret
= netdev_flow_put(port
->netdev
, &offload
->match
,
2411 CONST_CAST(struct nlattr
*, offload
->actions
),
2412 offload
->actions_len
, &flow
->mega_ufid
, &info
,
2414 ovs_mutex_unlock(&pmd
->dp
->port_mutex
);
2420 if (!modification
) {
2421 megaflow_to_mark_associate(&flow
->mega_ufid
, mark
);
2422 mark_to_flow_associate(mark
, flow
);
2427 if (!modification
) {
2428 flow_mark_free(mark
);
2430 mark_to_flow_disassociate(pmd
, flow
);
2436 dp_netdev_flow_offload_main(void *data OVS_UNUSED
)
2438 struct dp_flow_offload_item
*offload
;
2439 struct ovs_list
*list
;
2444 ovs_mutex_lock(&dp_flow_offload
.mutex
);
2445 if (ovs_list_is_empty(&dp_flow_offload
.list
)) {
2446 ovsrcu_quiesce_start();
2447 ovs_mutex_cond_wait(&dp_flow_offload
.cond
,
2448 &dp_flow_offload
.mutex
);
2449 ovsrcu_quiesce_end();
2451 list
= ovs_list_pop_front(&dp_flow_offload
.list
);
2452 offload
= CONTAINER_OF(list
, struct dp_flow_offload_item
, node
);
2453 ovs_mutex_unlock(&dp_flow_offload
.mutex
);
2455 switch (offload
->op
) {
2456 case DP_NETDEV_FLOW_OFFLOAD_OP_ADD
:
2458 ret
= dp_netdev_flow_offload_put(offload
);
2460 case DP_NETDEV_FLOW_OFFLOAD_OP_MOD
:
2462 ret
= dp_netdev_flow_offload_put(offload
);
2464 case DP_NETDEV_FLOW_OFFLOAD_OP_DEL
:
2466 ret
= dp_netdev_flow_offload_del(offload
);
2472 VLOG_DBG("%s to %s netdev flow\n",
2473 ret
== 0 ? "succeed" : "failed", op
);
2474 dp_netdev_free_flow_offload(offload
);
2481 queue_netdev_flow_del(struct dp_netdev_pmd_thread
*pmd
,
2482 struct dp_netdev_flow
*flow
)
2484 struct dp_flow_offload_item
*offload
;
2486 if (ovsthread_once_start(&offload_thread_once
)) {
2487 xpthread_cond_init(&dp_flow_offload
.cond
, NULL
);
2488 ovs_thread_create("dp_netdev_flow_offload",
2489 dp_netdev_flow_offload_main
, NULL
);
2490 ovsthread_once_done(&offload_thread_once
);
2493 offload
= dp_netdev_alloc_flow_offload(pmd
, flow
,
2494 DP_NETDEV_FLOW_OFFLOAD_OP_DEL
);
2495 dp_netdev_append_flow_offload(offload
);
2499 queue_netdev_flow_put(struct dp_netdev_pmd_thread
*pmd
,
2500 struct dp_netdev_flow
*flow
, struct match
*match
,
2501 const struct nlattr
*actions
, size_t actions_len
)
2503 struct dp_flow_offload_item
*offload
;
2506 if (!netdev_is_flow_api_enabled()) {
2510 if (ovsthread_once_start(&offload_thread_once
)) {
2511 xpthread_cond_init(&dp_flow_offload
.cond
, NULL
);
2512 ovs_thread_create("dp_netdev_flow_offload",
2513 dp_netdev_flow_offload_main
, NULL
);
2514 ovsthread_once_done(&offload_thread_once
);
2517 if (flow
->mark
!= INVALID_FLOW_MARK
) {
2518 op
= DP_NETDEV_FLOW_OFFLOAD_OP_MOD
;
2520 op
= DP_NETDEV_FLOW_OFFLOAD_OP_ADD
;
2522 offload
= dp_netdev_alloc_flow_offload(pmd
, flow
, op
);
2523 offload
->match
= *match
;
2524 offload
->actions
= xmalloc(actions_len
);
2525 memcpy(offload
->actions
, actions
, actions_len
);
2526 offload
->actions_len
= actions_len
;
2528 dp_netdev_append_flow_offload(offload
);
2532 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread
*pmd
,
2533 struct dp_netdev_flow
*flow
)
2534 OVS_REQUIRES(pmd
->flow_mutex
)
2536 struct cmap_node
*node
= CONST_CAST(struct cmap_node
*, &flow
->node
);
2538 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
2540 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2541 ovs_assert(cls
!= NULL
);
2542 dpcls_remove(cls
, &flow
->cr
);
2543 cmap_remove(&pmd
->flow_table
, node
, dp_netdev_flow_hash(&flow
->ufid
));
2544 if (flow
->mark
!= INVALID_FLOW_MARK
) {
2545 queue_netdev_flow_del(pmd
, flow
);
2549 dp_netdev_flow_unref(flow
);
2553 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
)
2555 struct dp_netdev_flow
*netdev_flow
;
2557 ovs_mutex_lock(&pmd
->flow_mutex
);
2558 CMAP_FOR_EACH (netdev_flow
, node
, &pmd
->flow_table
) {
2559 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
2561 ovs_mutex_unlock(&pmd
->flow_mutex
);
2565 dpif_netdev_flow_flush(struct dpif
*dpif
)
2567 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2568 struct dp_netdev_pmd_thread
*pmd
;
2570 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2571 dp_netdev_pmd_flow_flush(pmd
);
2577 struct dp_netdev_port_state
{
2578 struct hmap_position position
;
2583 dpif_netdev_port_dump_start(const struct dpif
*dpif OVS_UNUSED
, void **statep
)
2585 *statep
= xzalloc(sizeof(struct dp_netdev_port_state
));
2590 dpif_netdev_port_dump_next(const struct dpif
*dpif
, void *state_
,
2591 struct dpif_port
*dpif_port
)
2593 struct dp_netdev_port_state
*state
= state_
;
2594 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2595 struct hmap_node
*node
;
2598 ovs_mutex_lock(&dp
->port_mutex
);
2599 node
= hmap_at_position(&dp
->ports
, &state
->position
);
2601 struct dp_netdev_port
*port
;
2603 port
= CONTAINER_OF(node
, struct dp_netdev_port
, node
);
2606 state
->name
= xstrdup(netdev_get_name(port
->netdev
));
2607 dpif_port
->name
= state
->name
;
2608 dpif_port
->type
= port
->type
;
2609 dpif_port
->port_no
= port
->port_no
;
2615 ovs_mutex_unlock(&dp
->port_mutex
);
2621 dpif_netdev_port_dump_done(const struct dpif
*dpif OVS_UNUSED
, void *state_
)
2623 struct dp_netdev_port_state
*state
= state_
;
2630 dpif_netdev_port_poll(const struct dpif
*dpif_
, char **devnamep OVS_UNUSED
)
2632 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
2633 uint64_t new_port_seq
;
2636 new_port_seq
= seq_read(dpif
->dp
->port_seq
);
2637 if (dpif
->last_port_seq
!= new_port_seq
) {
2638 dpif
->last_port_seq
= new_port_seq
;
2648 dpif_netdev_port_poll_wait(const struct dpif
*dpif_
)
2650 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
2652 seq_wait(dpif
->dp
->port_seq
, dpif
->last_port_seq
);
2655 static struct dp_netdev_flow
*
2656 dp_netdev_flow_cast(const struct dpcls_rule
*cr
)
2658 return cr
? CONTAINER_OF(cr
, struct dp_netdev_flow
, cr
) : NULL
;
2661 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*flow
)
2663 return ovs_refcount_try_ref_rcu(&flow
->ref_cnt
);
2666 /* netdev_flow_key utilities.
2668 * netdev_flow_key is basically a miniflow. We use these functions
2669 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
2670 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
2672 * - Since we are dealing exclusively with miniflows created by
2673 * miniflow_extract(), if the map is different the miniflow is different.
2674 * Therefore we can be faster by comparing the map and the miniflow in a
2676 * - These functions can be inlined by the compiler. */
2678 /* Given the number of bits set in miniflow's maps, returns the size of the
2679 * 'netdev_flow_key.mf' */
2680 static inline size_t
2681 netdev_flow_key_size(size_t flow_u64s
)
2683 return sizeof(struct miniflow
) + MINIFLOW_VALUES_SIZE(flow_u64s
);
2687 netdev_flow_key_equal(const struct netdev_flow_key
*a
,
2688 const struct netdev_flow_key
*b
)
2690 /* 'b->len' may be not set yet. */
2691 return a
->hash
== b
->hash
&& !memcmp(&a
->mf
, &b
->mf
, a
->len
);
2694 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
2695 * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been
2696 * generated by miniflow_extract. */
2698 netdev_flow_key_equal_mf(const struct netdev_flow_key
*key
,
2699 const struct miniflow
*mf
)
2701 return !memcmp(&key
->mf
, mf
, key
->len
);
2705 netdev_flow_key_clone(struct netdev_flow_key
*dst
,
2706 const struct netdev_flow_key
*src
)
2709 offsetof(struct netdev_flow_key
, mf
) + src
->len
);
2712 /* Initialize a netdev_flow_key 'mask' from 'match'. */
2714 netdev_flow_mask_init(struct netdev_flow_key
*mask
,
2715 const struct match
*match
)
2717 uint64_t *dst
= miniflow_values(&mask
->mf
);
2718 struct flowmap fmap
;
2722 /* Only check masks that make sense for the flow. */
2723 flow_wc_map(&match
->flow
, &fmap
);
2724 flowmap_init(&mask
->mf
.map
);
2726 FLOWMAP_FOR_EACH_INDEX(idx
, fmap
) {
2727 uint64_t mask_u64
= flow_u64_value(&match
->wc
.masks
, idx
);
2730 flowmap_set(&mask
->mf
.map
, idx
, 1);
2732 hash
= hash_add64(hash
, mask_u64
);
2738 FLOWMAP_FOR_EACH_MAP (map
, mask
->mf
.map
) {
2739 hash
= hash_add64(hash
, map
);
2742 size_t n
= dst
- miniflow_get_values(&mask
->mf
);
2744 mask
->hash
= hash_finish(hash
, n
* 8);
2745 mask
->len
= netdev_flow_key_size(n
);
2748 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
2750 netdev_flow_key_init_masked(struct netdev_flow_key
*dst
,
2751 const struct flow
*flow
,
2752 const struct netdev_flow_key
*mask
)
2754 uint64_t *dst_u64
= miniflow_values(&dst
->mf
);
2755 const uint64_t *mask_u64
= miniflow_get_values(&mask
->mf
);
2759 dst
->len
= mask
->len
;
2760 dst
->mf
= mask
->mf
; /* Copy maps. */
2762 FLOW_FOR_EACH_IN_MAPS(value
, flow
, mask
->mf
.map
) {
2763 *dst_u64
= value
& *mask_u64
++;
2764 hash
= hash_add64(hash
, *dst_u64
++);
2766 dst
->hash
= hash_finish(hash
,
2767 (dst_u64
- miniflow_get_values(&dst
->mf
)) * 8);
2771 emc_entry_alive(struct emc_entry
*ce
)
2773 return ce
->flow
&& !ce
->flow
->dead
;
2777 emc_clear_entry(struct emc_entry
*ce
)
2780 dp_netdev_flow_unref(ce
->flow
);
2786 emc_change_entry(struct emc_entry
*ce
, struct dp_netdev_flow
*flow
,
2787 const struct netdev_flow_key
*key
)
2789 if (ce
->flow
!= flow
) {
2791 dp_netdev_flow_unref(ce
->flow
);
2794 if (dp_netdev_flow_ref(flow
)) {
2801 netdev_flow_key_clone(&ce
->key
, key
);
2806 emc_insert(struct emc_cache
*cache
, const struct netdev_flow_key
*key
,
2807 struct dp_netdev_flow
*flow
)
2809 struct emc_entry
*to_be_replaced
= NULL
;
2810 struct emc_entry
*current_entry
;
2812 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2813 if (netdev_flow_key_equal(¤t_entry
->key
, key
)) {
2814 /* We found the entry with the 'mf' miniflow */
2815 emc_change_entry(current_entry
, flow
, NULL
);
2819 /* Replacement policy: put the flow in an empty (not alive) entry, or
2820 * in the first entry where it can be */
2822 || (emc_entry_alive(to_be_replaced
)
2823 && !emc_entry_alive(current_entry
))
2824 || current_entry
->key
.hash
< to_be_replaced
->key
.hash
) {
2825 to_be_replaced
= current_entry
;
2828 /* We didn't find the miniflow in the cache.
2829 * The 'to_be_replaced' entry is where the new flow will be stored */
2831 emc_change_entry(to_be_replaced
, flow
, key
);
2835 emc_probabilistic_insert(struct dp_netdev_pmd_thread
*pmd
,
2836 const struct netdev_flow_key
*key
,
2837 struct dp_netdev_flow
*flow
)
2839 /* Insert an entry into the EMC based on probability value 'min'. By
2840 * default the value is UINT32_MAX / 100 which yields an insertion
2841 * probability of 1/100 ie. 1% */
2843 uint32_t min
= pmd
->ctx
.emc_insert_min
;
2845 if (min
&& random_uint32() <= min
) {
2846 emc_insert(&(pmd
->flow_cache
).emc_cache
, key
, flow
);
2850 static inline struct dp_netdev_flow
*
2851 emc_lookup(struct emc_cache
*cache
, const struct netdev_flow_key
*key
)
2853 struct emc_entry
*current_entry
;
2855 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2856 if (current_entry
->key
.hash
== key
->hash
2857 && emc_entry_alive(current_entry
)
2858 && netdev_flow_key_equal_mf(¤t_entry
->key
, &key
->mf
)) {
2860 /* We found the entry with the 'key->mf' miniflow */
2861 return current_entry
->flow
;
2868 static inline const struct cmap_node
*
2869 smc_entry_get(struct dp_netdev_pmd_thread
*pmd
, const uint32_t hash
)
2871 struct smc_cache
*cache
= &(pmd
->flow_cache
).smc_cache
;
2872 struct smc_bucket
*bucket
= &cache
->buckets
[hash
& SMC_MASK
];
2873 uint16_t sig
= hash
>> 16;
2874 uint16_t index
= UINT16_MAX
;
2876 for (int i
= 0; i
< SMC_ENTRY_PER_BUCKET
; i
++) {
2877 if (bucket
->sig
[i
] == sig
) {
2878 index
= bucket
->flow_idx
[i
];
2882 if (index
!= UINT16_MAX
) {
2883 return cmap_find_by_index(&pmd
->flow_table
, index
);
2889 smc_clear_entry(struct smc_bucket
*b
, int idx
)
2891 b
->flow_idx
[idx
] = UINT16_MAX
;
2894 /* Insert the flow_table index into SMC. Insertion may fail when 1) SMC is
2895 * turned off, 2) the flow_table index is larger than uint16_t can handle.
2896 * If there is already an SMC entry having same signature, the index will be
2897 * updated. If there is no existing entry, but an empty entry is available,
2898 * the empty entry will be taken. If no empty entry or existing same signature,
2899 * a random entry from the hashed bucket will be picked. */
2901 smc_insert(struct dp_netdev_pmd_thread
*pmd
,
2902 const struct netdev_flow_key
*key
,
2905 struct smc_cache
*smc_cache
= &(pmd
->flow_cache
).smc_cache
;
2906 struct smc_bucket
*bucket
= &smc_cache
->buckets
[key
->hash
& SMC_MASK
];
2908 uint32_t cmap_index
;
2912 atomic_read_relaxed(&pmd
->dp
->smc_enable_db
, &smc_enable_db
);
2913 if (!smc_enable_db
) {
2917 cmap_index
= cmap_find_index(&pmd
->flow_table
, hash
);
2918 index
= (cmap_index
>= UINT16_MAX
) ? UINT16_MAX
: (uint16_t)cmap_index
;
2920 /* If the index is larger than SMC can handle (uint16_t), we don't
2922 if (index
== UINT16_MAX
) {
2926 /* If an entry with same signature already exists, update the index */
2927 uint16_t sig
= key
->hash
>> 16;
2928 for (i
= 0; i
< SMC_ENTRY_PER_BUCKET
; i
++) {
2929 if (bucket
->sig
[i
] == sig
) {
2930 bucket
->flow_idx
[i
] = index
;
2934 /* If there is an empty entry, occupy it. */
2935 for (i
= 0; i
< SMC_ENTRY_PER_BUCKET
; i
++) {
2936 if (bucket
->flow_idx
[i
] == UINT16_MAX
) {
2937 bucket
->sig
[i
] = sig
;
2938 bucket
->flow_idx
[i
] = index
;
2942 /* Otherwise, pick a random entry. */
2943 i
= random_uint32() % SMC_ENTRY_PER_BUCKET
;
2944 bucket
->sig
[i
] = sig
;
2945 bucket
->flow_idx
[i
] = index
;
2948 static struct dp_netdev_flow
*
2949 dp_netdev_pmd_lookup_flow(struct dp_netdev_pmd_thread
*pmd
,
2950 const struct netdev_flow_key
*key
,
2954 struct dpcls_rule
*rule
;
2955 odp_port_t in_port
= u32_to_odp(MINIFLOW_GET_U32(&key
->mf
,
2957 struct dp_netdev_flow
*netdev_flow
= NULL
;
2959 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2960 if (OVS_LIKELY(cls
)) {
2961 dpcls_lookup(cls
, &key
, &rule
, 1, lookup_num_p
);
2962 netdev_flow
= dp_netdev_flow_cast(rule
);
2967 static struct dp_netdev_flow
*
2968 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread
*pmd
,
2969 const ovs_u128
*ufidp
, const struct nlattr
*key
,
2972 struct dp_netdev_flow
*netdev_flow
;
2976 /* If a UFID is not provided, determine one based on the key. */
2977 if (!ufidp
&& key
&& key_len
2978 && !dpif_netdev_flow_from_nlattrs(key
, key_len
, &flow
, false)) {
2979 dpif_flow_hash(pmd
->dp
->dpif
, &flow
, sizeof flow
, &ufid
);
2984 CMAP_FOR_EACH_WITH_HASH (netdev_flow
, node
, dp_netdev_flow_hash(ufidp
),
2986 if (ovs_u128_equals(netdev_flow
->ufid
, *ufidp
)) {
2996 get_dpif_flow_stats(const struct dp_netdev_flow
*netdev_flow_
,
2997 struct dpif_flow_stats
*stats
)
2999 struct dp_netdev_flow
*netdev_flow
;
3000 unsigned long long n
;
3004 netdev_flow
= CONST_CAST(struct dp_netdev_flow
*, netdev_flow_
);
3006 atomic_read_relaxed(&netdev_flow
->stats
.packet_count
, &n
);
3007 stats
->n_packets
= n
;
3008 atomic_read_relaxed(&netdev_flow
->stats
.byte_count
, &n
);
3010 atomic_read_relaxed(&netdev_flow
->stats
.used
, &used
);
3012 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
3013 stats
->tcp_flags
= flags
;
3016 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
3017 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
3018 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
3021 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow
*netdev_flow
,
3022 struct ofpbuf
*key_buf
, struct ofpbuf
*mask_buf
,
3023 struct dpif_flow
*flow
, bool terse
)
3026 memset(flow
, 0, sizeof *flow
);
3028 struct flow_wildcards wc
;
3029 struct dp_netdev_actions
*actions
;
3031 struct odp_flow_key_parms odp_parms
= {
3032 .flow
= &netdev_flow
->flow
,
3034 .support
= dp_netdev_support
,
3037 miniflow_expand(&netdev_flow
->cr
.mask
->mf
, &wc
.masks
);
3038 /* in_port is exact matched, but we have left it out from the mask for
3039 * optimnization reasons. Add in_port back to the mask. */
3040 wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
3043 offset
= key_buf
->size
;
3044 flow
->key
= ofpbuf_tail(key_buf
);
3045 odp_flow_key_from_flow(&odp_parms
, key_buf
);
3046 flow
->key_len
= key_buf
->size
- offset
;
3049 offset
= mask_buf
->size
;
3050 flow
->mask
= ofpbuf_tail(mask_buf
);
3051 odp_parms
.key_buf
= key_buf
;
3052 odp_flow_key_from_mask(&odp_parms
, mask_buf
);
3053 flow
->mask_len
= mask_buf
->size
- offset
;
3056 actions
= dp_netdev_flow_get_actions(netdev_flow
);
3057 flow
->actions
= actions
->actions
;
3058 flow
->actions_len
= actions
->size
;
3061 flow
->ufid
= netdev_flow
->ufid
;
3062 flow
->ufid_present
= true;
3063 flow
->pmd_id
= netdev_flow
->pmd_id
;
3064 get_dpif_flow_stats(netdev_flow
, &flow
->stats
);
3066 flow
->attrs
.offloaded
= false;
3067 flow
->attrs
.dp_layer
= "ovs";
3071 dpif_netdev_mask_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
3072 const struct nlattr
*mask_key
,
3073 uint32_t mask_key_len
, const struct flow
*flow
,
3074 struct flow_wildcards
*wc
, bool probe
)
3076 enum odp_key_fitness fitness
;
3078 fitness
= odp_flow_key_to_mask(mask_key
, mask_key_len
, wc
, flow
, NULL
);
3081 /* This should not happen: it indicates that
3082 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
3083 * disagree on the acceptable form of a mask. Log the problem
3084 * as an error, with enough details to enable debugging. */
3085 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
3087 if (!VLOG_DROP_ERR(&rl
)) {
3091 odp_flow_format(key
, key_len
, mask_key
, mask_key_len
, NULL
, &s
,
3093 VLOG_ERR("internal error parsing flow mask %s (%s)",
3094 ds_cstr(&s
), odp_key_fitness_to_string(fitness
));
3106 dpif_netdev_flow_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
3107 struct flow
*flow
, bool probe
)
3109 if (odp_flow_key_to_flow(key
, key_len
, flow
, NULL
)) {
3111 /* This should not happen: it indicates that
3112 * odp_flow_key_from_flow() and odp_flow_key_to_flow() disagree on
3113 * the acceptable form of a flow. Log the problem as an error,
3114 * with enough details to enable debugging. */
3115 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
3117 if (!VLOG_DROP_ERR(&rl
)) {
3121 odp_flow_format(key
, key_len
, NULL
, 0, NULL
, &s
, true);
3122 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s
));
3130 if (flow
->ct_state
& DP_NETDEV_CS_UNSUPPORTED_MASK
) {
3138 dpif_netdev_flow_get(const struct dpif
*dpif
, const struct dpif_flow_get
*get
)
3140 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3141 struct dp_netdev_flow
*netdev_flow
;
3142 struct dp_netdev_pmd_thread
*pmd
;
3143 struct hmapx to_find
= HMAPX_INITIALIZER(&to_find
);
3144 struct hmapx_node
*node
;
3147 if (get
->pmd_id
== PMD_ID_NULL
) {
3148 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3149 if (dp_netdev_pmd_try_ref(pmd
) && !hmapx_add(&to_find
, pmd
)) {
3150 dp_netdev_pmd_unref(pmd
);
3154 pmd
= dp_netdev_get_pmd(dp
, get
->pmd_id
);
3158 hmapx_add(&to_find
, pmd
);
3161 if (!hmapx_count(&to_find
)) {
3165 HMAPX_FOR_EACH (node
, &to_find
) {
3166 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3167 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, get
->ufid
, get
->key
,
3170 dp_netdev_flow_to_dpif_flow(netdev_flow
, get
->buffer
, get
->buffer
,
3179 HMAPX_FOR_EACH (node
, &to_find
) {
3180 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3181 dp_netdev_pmd_unref(pmd
);
3184 hmapx_destroy(&to_find
);
3189 dp_netdev_get_mega_ufid(const struct match
*match
, ovs_u128
*mega_ufid
)
3191 struct flow masked_flow
;
3194 for (i
= 0; i
< sizeof(struct flow
); i
++) {
3195 ((uint8_t *)&masked_flow
)[i
] = ((uint8_t *)&match
->flow
)[i
] &
3196 ((uint8_t *)&match
->wc
)[i
];
3198 dpif_flow_hash(NULL
, &masked_flow
, sizeof(struct flow
), mega_ufid
);
3201 static struct dp_netdev_flow
*
3202 dp_netdev_flow_add(struct dp_netdev_pmd_thread
*pmd
,
3203 struct match
*match
, const ovs_u128
*ufid
,
3204 const struct nlattr
*actions
, size_t actions_len
)
3205 OVS_REQUIRES(pmd
->flow_mutex
)
3207 struct dp_netdev_flow
*flow
;
3208 struct netdev_flow_key mask
;
3211 /* Make sure in_port is exact matched before we read it. */
3212 ovs_assert(match
->wc
.masks
.in_port
.odp_port
== ODPP_NONE
);
3213 odp_port_t in_port
= match
->flow
.in_port
.odp_port
;
3215 /* As we select the dpcls based on the port number, each netdev flow
3216 * belonging to the same dpcls will have the same odp_port value.
3217 * For performance reasons we wildcard odp_port here in the mask. In the
3218 * typical case dp_hash is also wildcarded, and the resulting 8-byte
3219 * chunk {dp_hash, in_port} will be ignored by netdev_flow_mask_init() and
3220 * will not be part of the subtable mask.
3221 * This will speed up the hash computation during dpcls_lookup() because
3222 * there is one less call to hash_add64() in this case. */
3223 match
->wc
.masks
.in_port
.odp_port
= 0;
3224 netdev_flow_mask_init(&mask
, match
);
3225 match
->wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
3227 /* Make sure wc does not have metadata. */
3228 ovs_assert(!FLOWMAP_HAS_FIELD(&mask
.mf
.map
, metadata
)
3229 && !FLOWMAP_HAS_FIELD(&mask
.mf
.map
, regs
));
3231 /* Do not allocate extra space. */
3232 flow
= xmalloc(sizeof *flow
- sizeof flow
->cr
.flow
.mf
+ mask
.len
);
3233 memset(&flow
->stats
, 0, sizeof flow
->stats
);
3236 flow
->mark
= INVALID_FLOW_MARK
;
3237 *CONST_CAST(unsigned *, &flow
->pmd_id
) = pmd
->core_id
;
3238 *CONST_CAST(struct flow
*, &flow
->flow
) = match
->flow
;
3239 *CONST_CAST(ovs_u128
*, &flow
->ufid
) = *ufid
;
3240 ovs_refcount_init(&flow
->ref_cnt
);
3241 ovsrcu_set(&flow
->actions
, dp_netdev_actions_create(actions
, actions_len
));
3243 dp_netdev_get_mega_ufid(match
, CONST_CAST(ovs_u128
*, &flow
->mega_ufid
));
3244 netdev_flow_key_init_masked(&flow
->cr
.flow
, &match
->flow
, &mask
);
3246 /* Select dpcls for in_port. Relies on in_port to be exact match. */
3247 cls
= dp_netdev_pmd_find_dpcls(pmd
, in_port
);
3248 dpcls_insert(cls
, &flow
->cr
, &mask
);
3250 cmap_insert(&pmd
->flow_table
, CONST_CAST(struct cmap_node
*, &flow
->node
),
3251 dp_netdev_flow_hash(&flow
->ufid
));
3253 queue_netdev_flow_put(pmd
, flow
, match
, actions
, actions_len
);
3255 if (OVS_UNLIKELY(!VLOG_DROP_DBG((&upcall_rl
)))) {
3256 struct ds ds
= DS_EMPTY_INITIALIZER
;
3257 struct ofpbuf key_buf
, mask_buf
;
3258 struct odp_flow_key_parms odp_parms
= {
3259 .flow
= &match
->flow
,
3260 .mask
= &match
->wc
.masks
,
3261 .support
= dp_netdev_support
,
3264 ofpbuf_init(&key_buf
, 0);
3265 ofpbuf_init(&mask_buf
, 0);
3267 odp_flow_key_from_flow(&odp_parms
, &key_buf
);
3268 odp_parms
.key_buf
= &key_buf
;
3269 odp_flow_key_from_mask(&odp_parms
, &mask_buf
);
3271 ds_put_cstr(&ds
, "flow_add: ");
3272 odp_format_ufid(ufid
, &ds
);
3273 ds_put_cstr(&ds
, " ");
3274 odp_flow_format(key_buf
.data
, key_buf
.size
,
3275 mask_buf
.data
, mask_buf
.size
,
3277 ds_put_cstr(&ds
, ", actions:");
3278 format_odp_actions(&ds
, actions
, actions_len
, NULL
);
3280 VLOG_DBG("%s", ds_cstr(&ds
));
3282 ofpbuf_uninit(&key_buf
);
3283 ofpbuf_uninit(&mask_buf
);
3285 /* Add a printout of the actual match installed. */
3288 ds_put_cstr(&ds
, "flow match: ");
3289 miniflow_expand(&flow
->cr
.flow
.mf
, &m
.flow
);
3290 miniflow_expand(&flow
->cr
.mask
->mf
, &m
.wc
.masks
);
3291 memset(&m
.tun_md
, 0, sizeof m
.tun_md
);
3292 match_format(&m
, NULL
, &ds
, OFP_DEFAULT_PRIORITY
);
3294 VLOG_DBG("%s", ds_cstr(&ds
));
3303 flow_put_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
3304 struct netdev_flow_key
*key
,
3305 struct match
*match
,
3307 const struct dpif_flow_put
*put
,
3308 struct dpif_flow_stats
*stats
)
3310 struct dp_netdev_flow
*netdev_flow
;
3314 memset(stats
, 0, sizeof *stats
);
3317 ovs_mutex_lock(&pmd
->flow_mutex
);
3318 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
3320 if (put
->flags
& DPIF_FP_CREATE
) {
3321 if (cmap_count(&pmd
->flow_table
) < MAX_FLOWS
) {
3322 dp_netdev_flow_add(pmd
, match
, ufid
, put
->actions
,
3332 if (put
->flags
& DPIF_FP_MODIFY
) {
3333 struct dp_netdev_actions
*new_actions
;
3334 struct dp_netdev_actions
*old_actions
;
3336 new_actions
= dp_netdev_actions_create(put
->actions
,
3339 old_actions
= dp_netdev_flow_get_actions(netdev_flow
);
3340 ovsrcu_set(&netdev_flow
->actions
, new_actions
);
3342 queue_netdev_flow_put(pmd
, netdev_flow
, match
,
3343 put
->actions
, put
->actions_len
);
3346 get_dpif_flow_stats(netdev_flow
, stats
);
3348 if (put
->flags
& DPIF_FP_ZERO_STATS
) {
3349 /* XXX: The userspace datapath uses thread local statistics
3350 * (for flows), which should be updated only by the owning
3351 * thread. Since we cannot write on stats memory here,
3352 * we choose not to support this flag. Please note:
3353 * - This feature is currently used only by dpctl commands with
3355 * - Should the need arise, this operation can be implemented
3356 * by keeping a base value (to be update here) for each
3357 * counter, and subtracting it before outputting the stats */
3361 ovsrcu_postpone(dp_netdev_actions_free
, old_actions
);
3362 } else if (put
->flags
& DPIF_FP_CREATE
) {
3365 /* Overlapping flow. */
3369 ovs_mutex_unlock(&pmd
->flow_mutex
);
3374 dpif_netdev_flow_put(struct dpif
*dpif
, const struct dpif_flow_put
*put
)
3376 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3377 struct netdev_flow_key key
, mask
;
3378 struct dp_netdev_pmd_thread
*pmd
;
3382 bool probe
= put
->flags
& DPIF_FP_PROBE
;
3385 memset(put
->stats
, 0, sizeof *put
->stats
);
3387 error
= dpif_netdev_flow_from_nlattrs(put
->key
, put
->key_len
, &match
.flow
,
3392 error
= dpif_netdev_mask_from_nlattrs(put
->key
, put
->key_len
,
3393 put
->mask
, put
->mask_len
,
3394 &match
.flow
, &match
.wc
, probe
);
3402 dpif_flow_hash(dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
3405 /* The Netlink encoding of datapath flow keys cannot express
3406 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3407 * tag is interpreted as exact match on the fact that there is no
3408 * VLAN. Unless we refactor a lot of code that translates between
3409 * Netlink and struct flow representations, we have to do the same
3410 * here. This must be in sync with 'match' in handle_packet_upcall(). */
3411 if (!match
.wc
.masks
.vlans
[0].tci
) {
3412 match
.wc
.masks
.vlans
[0].tci
= htons(0xffff);
3415 /* Must produce a netdev_flow_key for lookup.
3416 * Use the same method as employed to create the key when adding
3417 * the flow to the dplcs to make sure they match. */
3418 netdev_flow_mask_init(&mask
, &match
);
3419 netdev_flow_key_init_masked(&key
, &match
.flow
, &mask
);
3421 if (put
->pmd_id
== PMD_ID_NULL
) {
3422 if (cmap_count(&dp
->poll_threads
) == 0) {
3425 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3426 struct dpif_flow_stats pmd_stats
;
3429 pmd_error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
,
3433 } else if (put
->stats
) {
3434 put
->stats
->n_packets
+= pmd_stats
.n_packets
;
3435 put
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
3436 put
->stats
->used
= MAX(put
->stats
->used
, pmd_stats
.used
);
3437 put
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
3441 pmd
= dp_netdev_get_pmd(dp
, put
->pmd_id
);
3445 error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
, put
->stats
);
3446 dp_netdev_pmd_unref(pmd
);
3453 flow_del_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
3454 struct dpif_flow_stats
*stats
,
3455 const struct dpif_flow_del
*del
)
3457 struct dp_netdev_flow
*netdev_flow
;
3460 ovs_mutex_lock(&pmd
->flow_mutex
);
3461 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, del
->ufid
, del
->key
,
3465 get_dpif_flow_stats(netdev_flow
, stats
);
3467 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
3471 ovs_mutex_unlock(&pmd
->flow_mutex
);
3477 dpif_netdev_flow_del(struct dpif
*dpif
, const struct dpif_flow_del
*del
)
3479 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3480 struct dp_netdev_pmd_thread
*pmd
;
3484 memset(del
->stats
, 0, sizeof *del
->stats
);
3487 if (del
->pmd_id
== PMD_ID_NULL
) {
3488 if (cmap_count(&dp
->poll_threads
) == 0) {
3491 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3492 struct dpif_flow_stats pmd_stats
;
3495 pmd_error
= flow_del_on_pmd(pmd
, &pmd_stats
, del
);
3498 } else if (del
->stats
) {
3499 del
->stats
->n_packets
+= pmd_stats
.n_packets
;
3500 del
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
3501 del
->stats
->used
= MAX(del
->stats
->used
, pmd_stats
.used
);
3502 del
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
3506 pmd
= dp_netdev_get_pmd(dp
, del
->pmd_id
);
3510 error
= flow_del_on_pmd(pmd
, del
->stats
, del
);
3511 dp_netdev_pmd_unref(pmd
);
3518 struct dpif_netdev_flow_dump
{
3519 struct dpif_flow_dump up
;
3520 struct cmap_position poll_thread_pos
;
3521 struct cmap_position flow_pos
;
3522 struct dp_netdev_pmd_thread
*cur_pmd
;
3524 struct ovs_mutex mutex
;
3527 static struct dpif_netdev_flow_dump
*
3528 dpif_netdev_flow_dump_cast(struct dpif_flow_dump
*dump
)
3530 return CONTAINER_OF(dump
, struct dpif_netdev_flow_dump
, up
);
3533 static struct dpif_flow_dump
*
3534 dpif_netdev_flow_dump_create(const struct dpif
*dpif_
, bool terse
,
3535 struct dpif_flow_dump_types
*types OVS_UNUSED
)
3537 struct dpif_netdev_flow_dump
*dump
;
3539 dump
= xzalloc(sizeof *dump
);
3540 dpif_flow_dump_init(&dump
->up
, dpif_
);
3541 dump
->up
.terse
= terse
;
3542 ovs_mutex_init(&dump
->mutex
);
3548 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump
*dump_
)
3550 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
3552 ovs_mutex_destroy(&dump
->mutex
);
3557 struct dpif_netdev_flow_dump_thread
{
3558 struct dpif_flow_dump_thread up
;
3559 struct dpif_netdev_flow_dump
*dump
;
3560 struct odputil_keybuf keybuf
[FLOW_DUMP_MAX_BATCH
];
3561 struct odputil_keybuf maskbuf
[FLOW_DUMP_MAX_BATCH
];
3564 static struct dpif_netdev_flow_dump_thread
*
3565 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread
*thread
)
3567 return CONTAINER_OF(thread
, struct dpif_netdev_flow_dump_thread
, up
);
3570 static struct dpif_flow_dump_thread
*
3571 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump
*dump_
)
3573 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
3574 struct dpif_netdev_flow_dump_thread
*thread
;
3576 thread
= xmalloc(sizeof *thread
);
3577 dpif_flow_dump_thread_init(&thread
->up
, &dump
->up
);
3578 thread
->dump
= dump
;
3583 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread
*thread_
)
3585 struct dpif_netdev_flow_dump_thread
*thread
3586 = dpif_netdev_flow_dump_thread_cast(thread_
);
3592 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread
*thread_
,
3593 struct dpif_flow
*flows
, int max_flows
)
3595 struct dpif_netdev_flow_dump_thread
*thread
3596 = dpif_netdev_flow_dump_thread_cast(thread_
);
3597 struct dpif_netdev_flow_dump
*dump
= thread
->dump
;
3598 struct dp_netdev_flow
*netdev_flows
[FLOW_DUMP_MAX_BATCH
];
3602 ovs_mutex_lock(&dump
->mutex
);
3603 if (!dump
->status
) {
3604 struct dpif_netdev
*dpif
= dpif_netdev_cast(thread
->up
.dpif
);
3605 struct dp_netdev
*dp
= get_dp_netdev(&dpif
->dpif
);
3606 struct dp_netdev_pmd_thread
*pmd
= dump
->cur_pmd
;
3607 int flow_limit
= MIN(max_flows
, FLOW_DUMP_MAX_BATCH
);
3609 /* First call to dump_next(), extracts the first pmd thread.
3610 * If there is no pmd thread, returns immediately. */
3612 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
3614 ovs_mutex_unlock(&dump
->mutex
);
3621 for (n_flows
= 0; n_flows
< flow_limit
; n_flows
++) {
3622 struct cmap_node
*node
;
3624 node
= cmap_next_position(&pmd
->flow_table
, &dump
->flow_pos
);
3628 netdev_flows
[n_flows
] = CONTAINER_OF(node
,
3629 struct dp_netdev_flow
,
3632 /* When finishing dumping the current pmd thread, moves to
3634 if (n_flows
< flow_limit
) {
3635 memset(&dump
->flow_pos
, 0, sizeof dump
->flow_pos
);
3636 dp_netdev_pmd_unref(pmd
);
3637 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
3643 /* Keeps the reference to next caller. */
3644 dump
->cur_pmd
= pmd
;
3646 /* If the current dump is empty, do not exit the loop, since the
3647 * remaining pmds could have flows to be dumped. Just dumps again
3648 * on the new 'pmd'. */
3651 ovs_mutex_unlock(&dump
->mutex
);
3653 for (i
= 0; i
< n_flows
; i
++) {
3654 struct odputil_keybuf
*maskbuf
= &thread
->maskbuf
[i
];
3655 struct odputil_keybuf
*keybuf
= &thread
->keybuf
[i
];
3656 struct dp_netdev_flow
*netdev_flow
= netdev_flows
[i
];
3657 struct dpif_flow
*f
= &flows
[i
];
3658 struct ofpbuf key
, mask
;
3660 ofpbuf_use_stack(&key
, keybuf
, sizeof *keybuf
);
3661 ofpbuf_use_stack(&mask
, maskbuf
, sizeof *maskbuf
);
3662 dp_netdev_flow_to_dpif_flow(netdev_flow
, &key
, &mask
, f
,
3670 dpif_netdev_execute(struct dpif
*dpif
, struct dpif_execute
*execute
)
3671 OVS_NO_THREAD_SAFETY_ANALYSIS
3673 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3674 struct dp_netdev_pmd_thread
*pmd
;
3675 struct dp_packet_batch pp
;
3677 if (dp_packet_size(execute
->packet
) < ETH_HEADER_LEN
||
3678 dp_packet_size(execute
->packet
) > UINT16_MAX
) {
3682 /* Tries finding the 'pmd'. If NULL is returned, that means
3683 * the current thread is a non-pmd thread and should use
3684 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
3685 pmd
= ovsthread_getspecific(dp
->per_pmd_key
);
3687 pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
3693 if (execute
->probe
) {
3694 /* If this is part of a probe, Drop the packet, since executing
3695 * the action may actually cause spurious packets be sent into
3697 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3698 dp_netdev_pmd_unref(pmd
);
3703 /* If the current thread is non-pmd thread, acquires
3704 * the 'non_pmd_mutex'. */
3705 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3706 ovs_mutex_lock(&dp
->non_pmd_mutex
);
3709 /* Update current time in PMD context. We don't care about EMC insertion
3710 * probability, because we are on a slow path. */
3711 pmd_thread_ctx_time_update(pmd
);
3713 /* The action processing expects the RSS hash to be valid, because
3714 * it's always initialized at the beginning of datapath processing.
3715 * In this case, though, 'execute->packet' may not have gone through
3716 * the datapath at all, it may have been generated by the upper layer
3717 * (OpenFlow packet-out, BFD frame, ...). */
3718 if (!dp_packet_rss_valid(execute
->packet
)) {
3719 dp_packet_set_rss_hash(execute
->packet
,
3720 flow_hash_5tuple(execute
->flow
, 0));
3723 dp_packet_batch_init_packet(&pp
, execute
->packet
);
3724 pp
.do_not_steal
= true;
3725 dp_netdev_execute_actions(pmd
, &pp
, false, execute
->flow
,
3726 execute
->actions
, execute
->actions_len
);
3727 dp_netdev_pmd_flush_output_packets(pmd
, true);
3729 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3730 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
3731 dp_netdev_pmd_unref(pmd
);
3738 dpif_netdev_operate(struct dpif
*dpif
, struct dpif_op
**ops
, size_t n_ops
,
3739 enum dpif_offload_type offload_type OVS_UNUSED
)
3743 for (i
= 0; i
< n_ops
; i
++) {
3744 struct dpif_op
*op
= ops
[i
];
3747 case DPIF_OP_FLOW_PUT
:
3748 op
->error
= dpif_netdev_flow_put(dpif
, &op
->flow_put
);
3751 case DPIF_OP_FLOW_DEL
:
3752 op
->error
= dpif_netdev_flow_del(dpif
, &op
->flow_del
);
3755 case DPIF_OP_EXECUTE
:
3756 op
->error
= dpif_netdev_execute(dpif
, &op
->execute
);
3759 case DPIF_OP_FLOW_GET
:
3760 op
->error
= dpif_netdev_flow_get(dpif
, &op
->flow_get
);
3766 /* Enable or Disable PMD auto load balancing. */
3768 set_pmd_auto_lb(struct dp_netdev
*dp
)
3770 unsigned int cnt
= 0;
3771 struct dp_netdev_pmd_thread
*pmd
;
3772 struct pmd_auto_lb
*pmd_alb
= &dp
->pmd_alb
;
3774 bool enable_alb
= false;
3775 bool multi_rxq
= false;
3776 bool pmd_rxq_assign_cyc
= dp
->pmd_rxq_assign_cyc
;
3778 /* Ensure that there is at least 2 non-isolated PMDs and
3779 * one of them is polling more than one rxq. */
3780 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3781 if (pmd
->core_id
== NON_PMD_CORE_ID
|| pmd
->isolated
) {
3785 if (hmap_count(&pmd
->poll_list
) > 1) {
3788 if (cnt
&& multi_rxq
) {
3795 /* Enable auto LB if it is requested and cycle based assignment is true. */
3796 enable_alb
= enable_alb
&& pmd_rxq_assign_cyc
&&
3797 pmd_alb
->auto_lb_requested
;
3799 if (pmd_alb
->is_enabled
!= enable_alb
) {
3800 pmd_alb
->is_enabled
= enable_alb
;
3801 if (pmd_alb
->is_enabled
) {
3802 VLOG_INFO("PMD auto load balance is enabled "
3803 "(with rebalance interval:%"PRIu64
" msec)",
3804 pmd_alb
->rebalance_intvl
);
3806 pmd_alb
->rebalance_poll_timer
= 0;
3807 VLOG_INFO("PMD auto load balance is disabled");
3813 /* Applies datapath configuration from the database. Some of the changes are
3814 * actually applied in dpif_netdev_run(). */
3816 dpif_netdev_set_config(struct dpif
*dpif
, const struct smap
*other_config
)
3818 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3819 const char *cmask
= smap_get(other_config
, "pmd-cpu-mask");
3820 const char *pmd_rxq_assign
= smap_get_def(other_config
, "pmd-rxq-assign",
3822 unsigned long long insert_prob
=
3823 smap_get_ullong(other_config
, "emc-insert-inv-prob",
3824 DEFAULT_EM_FLOW_INSERT_INV_PROB
);
3825 uint32_t insert_min
, cur_min
;
3826 uint32_t tx_flush_interval
, cur_tx_flush_interval
;
3827 uint64_t rebalance_intvl
;
3829 tx_flush_interval
= smap_get_int(other_config
, "tx-flush-interval",
3830 DEFAULT_TX_FLUSH_INTERVAL
);
3831 atomic_read_relaxed(&dp
->tx_flush_interval
, &cur_tx_flush_interval
);
3832 if (tx_flush_interval
!= cur_tx_flush_interval
) {
3833 atomic_store_relaxed(&dp
->tx_flush_interval
, tx_flush_interval
);
3834 VLOG_INFO("Flushing interval for tx queues set to %"PRIu32
" us",
3838 if (!nullable_string_is_equal(dp
->pmd_cmask
, cmask
)) {
3839 free(dp
->pmd_cmask
);
3840 dp
->pmd_cmask
= nullable_xstrdup(cmask
);
3841 dp_netdev_request_reconfigure(dp
);
3844 atomic_read_relaxed(&dp
->emc_insert_min
, &cur_min
);
3845 if (insert_prob
<= UINT32_MAX
) {
3846 insert_min
= insert_prob
== 0 ? 0 : UINT32_MAX
/ insert_prob
;
3848 insert_min
= DEFAULT_EM_FLOW_INSERT_MIN
;
3849 insert_prob
= DEFAULT_EM_FLOW_INSERT_INV_PROB
;
3852 if (insert_min
!= cur_min
) {
3853 atomic_store_relaxed(&dp
->emc_insert_min
, insert_min
);
3854 if (insert_min
== 0) {
3855 VLOG_INFO("EMC insertion probability changed to zero");
3857 VLOG_INFO("EMC insertion probability changed to 1/%llu (~%.2f%%)",
3858 insert_prob
, (100 / (float)insert_prob
));
3862 bool perf_enabled
= smap_get_bool(other_config
, "pmd-perf-metrics", false);
3863 bool cur_perf_enabled
;
3864 atomic_read_relaxed(&dp
->pmd_perf_metrics
, &cur_perf_enabled
);
3865 if (perf_enabled
!= cur_perf_enabled
) {
3866 atomic_store_relaxed(&dp
->pmd_perf_metrics
, perf_enabled
);
3868 VLOG_INFO("PMD performance metrics collection enabled");
3870 VLOG_INFO("PMD performance metrics collection disabled");
3874 bool smc_enable
= smap_get_bool(other_config
, "smc-enable", false);
3876 atomic_read_relaxed(&dp
->smc_enable_db
, &cur_smc
);
3877 if (smc_enable
!= cur_smc
) {
3878 atomic_store_relaxed(&dp
->smc_enable_db
, smc_enable
);
3880 VLOG_INFO("SMC cache is enabled");
3882 VLOG_INFO("SMC cache is disabled");
3886 bool pmd_rxq_assign_cyc
= !strcmp(pmd_rxq_assign
, "cycles");
3887 if (!pmd_rxq_assign_cyc
&& strcmp(pmd_rxq_assign
, "roundrobin")) {
3888 VLOG_WARN("Unsupported Rxq to PMD assignment mode in pmd-rxq-assign. "
3889 "Defaulting to 'cycles'.");
3890 pmd_rxq_assign_cyc
= true;
3891 pmd_rxq_assign
= "cycles";
3893 if (dp
->pmd_rxq_assign_cyc
!= pmd_rxq_assign_cyc
) {
3894 dp
->pmd_rxq_assign_cyc
= pmd_rxq_assign_cyc
;
3895 VLOG_INFO("Rxq to PMD assignment mode changed to: \'%s\'.",
3897 dp_netdev_request_reconfigure(dp
);
3900 struct pmd_auto_lb
*pmd_alb
= &dp
->pmd_alb
;
3901 pmd_alb
->auto_lb_requested
= smap_get_bool(other_config
, "pmd-auto-lb",
3904 rebalance_intvl
= smap_get_int(other_config
, "pmd-auto-lb-rebal-interval",
3905 ALB_PMD_REBALANCE_POLL_INTERVAL
);
3907 /* Input is in min, convert it to msec. */
3909 rebalance_intvl
? rebalance_intvl
* MIN_TO_MSEC
: MIN_TO_MSEC
;
3911 if (pmd_alb
->rebalance_intvl
!= rebalance_intvl
) {
3912 pmd_alb
->rebalance_intvl
= rebalance_intvl
;
3915 set_pmd_auto_lb(dp
);
3919 /* Parses affinity list and returns result in 'core_ids'. */
3921 parse_affinity_list(const char *affinity_list
, unsigned *core_ids
, int n_rxq
)
3924 char *list
, *copy
, *key
, *value
;
3927 for (i
= 0; i
< n_rxq
; i
++) {
3928 core_ids
[i
] = OVS_CORE_UNSPEC
;
3931 if (!affinity_list
) {
3935 list
= copy
= xstrdup(affinity_list
);
3937 while (ofputil_parse_key_value(&list
, &key
, &value
)) {
3938 int rxq_id
, core_id
;
3940 if (!str_to_int(key
, 0, &rxq_id
) || rxq_id
< 0
3941 || !str_to_int(value
, 0, &core_id
) || core_id
< 0) {
3946 if (rxq_id
< n_rxq
) {
3947 core_ids
[rxq_id
] = core_id
;
3955 /* Parses 'affinity_list' and applies configuration if it is valid. */
3957 dpif_netdev_port_set_rxq_affinity(struct dp_netdev_port
*port
,
3958 const char *affinity_list
)
3960 unsigned *core_ids
, i
;
3963 core_ids
= xmalloc(port
->n_rxq
* sizeof *core_ids
);
3964 if (parse_affinity_list(affinity_list
, core_ids
, port
->n_rxq
)) {
3969 for (i
= 0; i
< port
->n_rxq
; i
++) {
3970 port
->rxqs
[i
].core_id
= core_ids
[i
];
3978 /* Returns 'true' if one of the 'port's RX queues exists in 'poll_list'
3979 * of given PMD thread. */
3981 dpif_netdev_pmd_polls_port(struct dp_netdev_pmd_thread
*pmd
,
3982 struct dp_netdev_port
*port
)
3983 OVS_EXCLUDED(pmd
->port_mutex
)
3985 struct rxq_poll
*poll
;
3988 ovs_mutex_lock(&pmd
->port_mutex
);
3989 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
3990 if (port
== poll
->rxq
->port
) {
3995 ovs_mutex_unlock(&pmd
->port_mutex
);
3999 /* Updates port configuration from the database. The changes are actually
4000 * applied in dpif_netdev_run(). */
4002 dpif_netdev_port_set_config(struct dpif
*dpif
, odp_port_t port_no
,
4003 const struct smap
*cfg
)
4005 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4006 struct dp_netdev_port
*port
;
4008 const char *affinity_list
= smap_get(cfg
, "pmd-rxq-affinity");
4009 bool emc_enabled
= smap_get_bool(cfg
, "emc-enable", true);
4011 ovs_mutex_lock(&dp
->port_mutex
);
4012 error
= get_port_by_number(dp
, port_no
, &port
);
4017 if (emc_enabled
!= port
->emc_enabled
) {
4018 struct dp_netdev_pmd_thread
*pmd
;
4019 struct ds ds
= DS_EMPTY_INITIALIZER
;
4020 uint32_t cur_min
, insert_prob
;
4022 port
->emc_enabled
= emc_enabled
;
4023 /* Mark for reload all the threads that polls this port and request
4024 * for reconfiguration for the actual reloading of threads. */
4025 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4026 if (dpif_netdev_pmd_polls_port(pmd
, port
)) {
4027 pmd
->need_reload
= true;
4030 dp_netdev_request_reconfigure(dp
);
4032 ds_put_format(&ds
, "%s: EMC has been %s.",
4033 netdev_get_name(port
->netdev
),
4034 (emc_enabled
) ? "enabled" : "disabled");
4036 ds_put_cstr(&ds
, " Current insertion probability is ");
4037 atomic_read_relaxed(&dp
->emc_insert_min
, &cur_min
);
4039 ds_put_cstr(&ds
, "zero.");
4041 insert_prob
= UINT32_MAX
/ cur_min
;
4042 ds_put_format(&ds
, "1/%"PRIu32
" (~%.2f%%).",
4043 insert_prob
, 100 / (float) insert_prob
);
4046 VLOG_INFO("%s", ds_cstr(&ds
));
4050 /* Checking for RXq affinity changes. */
4051 if (!netdev_is_pmd(port
->netdev
)
4052 || nullable_string_is_equal(affinity_list
, port
->rxq_affinity_list
)) {
4056 error
= dpif_netdev_port_set_rxq_affinity(port
, affinity_list
);
4060 free(port
->rxq_affinity_list
);
4061 port
->rxq_affinity_list
= nullable_xstrdup(affinity_list
);
4063 dp_netdev_request_reconfigure(dp
);
4065 ovs_mutex_unlock(&dp
->port_mutex
);
4070 dpif_netdev_queue_to_priority(const struct dpif
*dpif OVS_UNUSED
,
4071 uint32_t queue_id
, uint32_t *priority
)
4073 *priority
= queue_id
;
4078 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
4079 * a copy of the 'size' bytes of 'actions' input parameters. */
4080 struct dp_netdev_actions
*
4081 dp_netdev_actions_create(const struct nlattr
*actions
, size_t size
)
4083 struct dp_netdev_actions
*netdev_actions
;
4085 netdev_actions
= xmalloc(sizeof *netdev_actions
+ size
);
4086 memcpy(netdev_actions
->actions
, actions
, size
);
4087 netdev_actions
->size
= size
;
4089 return netdev_actions
;
4092 struct dp_netdev_actions
*
4093 dp_netdev_flow_get_actions(const struct dp_netdev_flow
*flow
)
4095 return ovsrcu_get(struct dp_netdev_actions
*, &flow
->actions
);
4099 dp_netdev_actions_free(struct dp_netdev_actions
*actions
)
4105 dp_netdev_rxq_set_cycles(struct dp_netdev_rxq
*rx
,
4106 enum rxq_cycles_counter_type type
,
4107 unsigned long long cycles
)
4109 atomic_store_relaxed(&rx
->cycles
[type
], cycles
);
4113 dp_netdev_rxq_add_cycles(struct dp_netdev_rxq
*rx
,
4114 enum rxq_cycles_counter_type type
,
4115 unsigned long long cycles
)
4117 non_atomic_ullong_add(&rx
->cycles
[type
], cycles
);
4121 dp_netdev_rxq_get_cycles(struct dp_netdev_rxq
*rx
,
4122 enum rxq_cycles_counter_type type
)
4124 unsigned long long processing_cycles
;
4125 atomic_read_relaxed(&rx
->cycles
[type
], &processing_cycles
);
4126 return processing_cycles
;
4130 dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq
*rx
,
4131 unsigned long long cycles
)
4133 unsigned int idx
= rx
->intrvl_idx
++ % PMD_RXQ_INTERVAL_MAX
;
4134 atomic_store_relaxed(&rx
->cycles_intrvl
[idx
], cycles
);
4138 dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq
*rx
, unsigned idx
)
4140 unsigned long long processing_cycles
;
4141 atomic_read_relaxed(&rx
->cycles_intrvl
[idx
], &processing_cycles
);
4142 return processing_cycles
;
4145 #if ATOMIC_ALWAYS_LOCK_FREE_8B
4147 pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread
*pmd
)
4149 bool pmd_perf_enabled
;
4150 atomic_read_relaxed(&pmd
->dp
->pmd_perf_metrics
, &pmd_perf_enabled
);
4151 return pmd_perf_enabled
;
4154 /* If stores and reads of 64-bit integers are not atomic, the full PMD
4155 * performance metrics are not available as locked access to 64 bit
4156 * integers would be prohibitively expensive. */
4158 pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread
*pmd OVS_UNUSED
)
4165 dp_netdev_pmd_flush_output_on_port(struct dp_netdev_pmd_thread
*pmd
,
4172 struct cycle_timer timer
;
4174 uint32_t tx_flush_interval
;
4176 cycle_timer_start(&pmd
->perf_stats
, &timer
);
4178 dynamic_txqs
= p
->port
->dynamic_txqs
;
4180 tx_qid
= dpif_netdev_xps_get_tx_qid(pmd
, p
);
4182 tx_qid
= pmd
->static_tx_qid
;
4185 output_cnt
= dp_packet_batch_size(&p
->output_pkts
);
4186 ovs_assert(output_cnt
> 0);
4188 netdev_send(p
->port
->netdev
, tx_qid
, &p
->output_pkts
, dynamic_txqs
);
4189 dp_packet_batch_init(&p
->output_pkts
);
4191 /* Update time of the next flush. */
4192 atomic_read_relaxed(&pmd
->dp
->tx_flush_interval
, &tx_flush_interval
);
4193 p
->flush_time
= pmd
->ctx
.now
+ tx_flush_interval
;
4195 ovs_assert(pmd
->n_output_batches
> 0);
4196 pmd
->n_output_batches
--;
4198 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_SENT_PKTS
, output_cnt
);
4199 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_SENT_BATCHES
, 1);
4201 /* Distribute send cycles evenly among transmitted packets and assign to
4202 * their respective rx queues. */
4203 cycles
= cycle_timer_stop(&pmd
->perf_stats
, &timer
) / output_cnt
;
4204 for (i
= 0; i
< output_cnt
; i
++) {
4205 if (p
->output_pkts_rxqs
[i
]) {
4206 dp_netdev_rxq_add_cycles(p
->output_pkts_rxqs
[i
],
4207 RXQ_CYCLES_PROC_CURR
, cycles
);
4215 dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread
*pmd
,
4221 if (!pmd
->n_output_batches
) {
4225 HMAP_FOR_EACH (p
, node
, &pmd
->send_port_cache
) {
4226 if (!dp_packet_batch_is_empty(&p
->output_pkts
)
4227 && (force
|| pmd
->ctx
.now
>= p
->flush_time
)) {
4228 output_cnt
+= dp_netdev_pmd_flush_output_on_port(pmd
, p
);
4235 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread
*pmd
,
4236 struct dp_netdev_rxq
*rxq
,
4239 struct pmd_perf_stats
*s
= &pmd
->perf_stats
;
4240 struct dp_packet_batch batch
;
4241 struct cycle_timer timer
;
4244 int rem_qlen
= 0, *qlen_p
= NULL
;
4247 /* Measure duration for polling and processing rx burst. */
4248 cycle_timer_start(&pmd
->perf_stats
, &timer
);
4250 pmd
->ctx
.last_rxq
= rxq
;
4251 dp_packet_batch_init(&batch
);
4253 /* Fetch the rx queue length only for vhostuser ports. */
4254 if (pmd_perf_metrics_enabled(pmd
) && rxq
->is_vhost
) {
4258 error
= netdev_rxq_recv(rxq
->rx
, &batch
, qlen_p
);
4260 /* At least one packet received. */
4261 *recirc_depth_get() = 0;
4262 pmd_thread_ctx_time_update(pmd
);
4263 batch_cnt
= batch
.count
;
4264 if (pmd_perf_metrics_enabled(pmd
)) {
4265 /* Update batch histogram. */
4266 s
->current
.batches
++;
4267 histogram_add_sample(&s
->pkts_per_batch
, batch_cnt
);
4268 /* Update the maximum vhost rx queue fill level. */
4269 if (rxq
->is_vhost
&& rem_qlen
>= 0) {
4270 uint32_t qfill
= batch_cnt
+ rem_qlen
;
4271 if (qfill
> s
->current
.max_vhost_qfill
) {
4272 s
->current
.max_vhost_qfill
= qfill
;
4276 /* Process packet batch. */
4277 dp_netdev_input(pmd
, &batch
, port_no
);
4279 /* Assign processing cycles to rx queue. */
4280 cycles
= cycle_timer_stop(&pmd
->perf_stats
, &timer
);
4281 dp_netdev_rxq_add_cycles(rxq
, RXQ_CYCLES_PROC_CURR
, cycles
);
4283 dp_netdev_pmd_flush_output_packets(pmd
, false);
4285 /* Discard cycles. */
4286 cycle_timer_stop(&pmd
->perf_stats
, &timer
);
4287 if (error
!= EAGAIN
&& error
!= EOPNOTSUPP
) {
4288 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
4290 VLOG_ERR_RL(&rl
, "error receiving data from %s: %s",
4291 netdev_rxq_get_name(rxq
->rx
), ovs_strerror(error
));
4295 pmd
->ctx
.last_rxq
= NULL
;
4300 static struct tx_port
*
4301 tx_port_lookup(const struct hmap
*hmap
, odp_port_t port_no
)
4305 HMAP_FOR_EACH_IN_BUCKET (tx
, node
, hash_port_no(port_no
), hmap
) {
4306 if (tx
->port
->port_no
== port_no
) {
4315 port_reconfigure(struct dp_netdev_port
*port
)
4317 struct netdev
*netdev
= port
->netdev
;
4320 /* Closes the existing 'rxq's. */
4321 for (i
= 0; i
< port
->n_rxq
; i
++) {
4322 netdev_rxq_close(port
->rxqs
[i
].rx
);
4323 port
->rxqs
[i
].rx
= NULL
;
4325 unsigned last_nrxq
= port
->n_rxq
;
4328 /* Allows 'netdev' to apply the pending configuration changes. */
4329 if (netdev_is_reconf_required(netdev
) || port
->need_reconfigure
) {
4330 err
= netdev_reconfigure(netdev
);
4331 if (err
&& (err
!= EOPNOTSUPP
)) {
4332 VLOG_ERR("Failed to set interface %s new configuration",
4333 netdev_get_name(netdev
));
4337 /* If the netdev_reconfigure() above succeeds, reopens the 'rxq's. */
4338 port
->rxqs
= xrealloc(port
->rxqs
,
4339 sizeof *port
->rxqs
* netdev_n_rxq(netdev
));
4340 /* Realloc 'used' counters for tx queues. */
4341 free(port
->txq_used
);
4342 port
->txq_used
= xcalloc(netdev_n_txq(netdev
), sizeof *port
->txq_used
);
4344 for (i
= 0; i
< netdev_n_rxq(netdev
); i
++) {
4345 bool new_queue
= i
>= last_nrxq
;
4347 memset(&port
->rxqs
[i
], 0, sizeof port
->rxqs
[i
]);
4350 port
->rxqs
[i
].port
= port
;
4351 port
->rxqs
[i
].is_vhost
= !strncmp(port
->type
, "dpdkvhost", 9);
4353 err
= netdev_rxq_open(netdev
, &port
->rxqs
[i
].rx
, i
);
4360 /* Parse affinity list to apply configuration for new queues. */
4361 dpif_netdev_port_set_rxq_affinity(port
, port
->rxq_affinity_list
);
4363 /* If reconfiguration was successful mark it as such, so we can use it */
4364 port
->need_reconfigure
= false;
4369 struct rr_numa_list
{
4370 struct hmap numas
; /* Contains 'struct rr_numa' */
4374 struct hmap_node node
;
4378 /* Non isolated pmds on numa node 'numa_id' */
4379 struct dp_netdev_pmd_thread
**pmds
;
4386 static struct rr_numa
*
4387 rr_numa_list_lookup(struct rr_numa_list
*rr
, int numa_id
)
4389 struct rr_numa
*numa
;
4391 HMAP_FOR_EACH_WITH_HASH (numa
, node
, hash_int(numa_id
, 0), &rr
->numas
) {
4392 if (numa
->numa_id
== numa_id
) {
4400 /* Returns the next node in numa list following 'numa' in round-robin fashion.
4401 * Returns first node if 'numa' is a null pointer or the last node in 'rr'.
4402 * Returns NULL if 'rr' numa list is empty. */
4403 static struct rr_numa
*
4404 rr_numa_list_next(struct rr_numa_list
*rr
, const struct rr_numa
*numa
)
4406 struct hmap_node
*node
= NULL
;
4409 node
= hmap_next(&rr
->numas
, &numa
->node
);
4412 node
= hmap_first(&rr
->numas
);
4415 return (node
) ? CONTAINER_OF(node
, struct rr_numa
, node
) : NULL
;
4419 rr_numa_list_populate(struct dp_netdev
*dp
, struct rr_numa_list
*rr
)
4421 struct dp_netdev_pmd_thread
*pmd
;
4422 struct rr_numa
*numa
;
4424 hmap_init(&rr
->numas
);
4426 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4427 if (pmd
->core_id
== NON_PMD_CORE_ID
|| pmd
->isolated
) {
4431 numa
= rr_numa_list_lookup(rr
, pmd
->numa_id
);
4433 numa
= xzalloc(sizeof *numa
);
4434 numa
->numa_id
= pmd
->numa_id
;
4435 hmap_insert(&rr
->numas
, &numa
->node
, hash_int(pmd
->numa_id
, 0));
4438 numa
->pmds
= xrealloc(numa
->pmds
, numa
->n_pmds
* sizeof *numa
->pmds
);
4439 numa
->pmds
[numa
->n_pmds
- 1] = pmd
;
4440 /* At least one pmd so initialise curr_idx and idx_inc. */
4441 numa
->cur_index
= 0;
4442 numa
->idx_inc
= true;
4447 * Returns the next pmd from the numa node.
4449 * If 'updown' is 'true' it will alternate between selecting the next pmd in
4450 * either an up or down walk, switching between up/down when the first or last
4451 * core is reached. e.g. 1,2,3,3,2,1,1,2...
4453 * If 'updown' is 'false' it will select the next pmd wrapping around when last
4454 * core reached. e.g. 1,2,3,1,2,3,1,2...
4456 static struct dp_netdev_pmd_thread
*
4457 rr_numa_get_pmd(struct rr_numa
*numa
, bool updown
)
4459 int numa_idx
= numa
->cur_index
;
4461 if (numa
->idx_inc
== true) {
4462 /* Incrementing through list of pmds. */
4463 if (numa
->cur_index
== numa
->n_pmds
-1) {
4464 /* Reached the last pmd. */
4466 numa
->idx_inc
= false;
4468 numa
->cur_index
= 0;
4474 /* Decrementing through list of pmds. */
4475 if (numa
->cur_index
== 0) {
4476 /* Reached the first pmd. */
4477 numa
->idx_inc
= true;
4482 return numa
->pmds
[numa_idx
];
4486 rr_numa_list_destroy(struct rr_numa_list
*rr
)
4488 struct rr_numa
*numa
;
4490 HMAP_FOR_EACH_POP (numa
, node
, &rr
->numas
) {
4494 hmap_destroy(&rr
->numas
);
4497 /* Sort Rx Queues by the processing cycles they are consuming. */
4499 compare_rxq_cycles(const void *a
, const void *b
)
4501 struct dp_netdev_rxq
*qa
;
4502 struct dp_netdev_rxq
*qb
;
4503 uint64_t cycles_qa
, cycles_qb
;
4505 qa
= *(struct dp_netdev_rxq
**) a
;
4506 qb
= *(struct dp_netdev_rxq
**) b
;
4508 cycles_qa
= dp_netdev_rxq_get_cycles(qa
, RXQ_CYCLES_PROC_HIST
);
4509 cycles_qb
= dp_netdev_rxq_get_cycles(qb
, RXQ_CYCLES_PROC_HIST
);
4511 if (cycles_qa
!= cycles_qb
) {
4512 return (cycles_qa
< cycles_qb
) ? 1 : -1;
4514 /* Cycles are the same so tiebreak on port/queue id.
4515 * Tiebreaking (as opposed to return 0) ensures consistent
4516 * sort results across multiple OS's. */
4517 uint32_t port_qa
= odp_to_u32(qa
->port
->port_no
);
4518 uint32_t port_qb
= odp_to_u32(qb
->port
->port_no
);
4519 if (port_qa
!= port_qb
) {
4520 return port_qa
> port_qb
? 1 : -1;
4522 return netdev_rxq_get_queue_id(qa
->rx
)
4523 - netdev_rxq_get_queue_id(qb
->rx
);
4528 /* Assign pmds to queues. If 'pinned' is true, assign pmds to pinned
4529 * queues and marks the pmds as isolated. Otherwise, assign non isolated
4530 * pmds to unpinned queues.
4532 * The function doesn't touch the pmd threads, it just stores the assignment
4533 * in the 'pmd' member of each rxq. */
4535 rxq_scheduling(struct dp_netdev
*dp
, bool pinned
) OVS_REQUIRES(dp
->port_mutex
)
4537 struct dp_netdev_port
*port
;
4538 struct rr_numa_list rr
;
4539 struct rr_numa
*non_local_numa
= NULL
;
4540 struct dp_netdev_rxq
** rxqs
= NULL
;
4542 struct rr_numa
*numa
= NULL
;
4544 bool assign_cyc
= dp
->pmd_rxq_assign_cyc
;
4546 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4547 if (!netdev_is_pmd(port
->netdev
)) {
4551 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
4552 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
4554 if (pinned
&& q
->core_id
!= OVS_CORE_UNSPEC
) {
4555 struct dp_netdev_pmd_thread
*pmd
;
4557 pmd
= dp_netdev_get_pmd(dp
, q
->core_id
);
4559 VLOG_WARN("There is no PMD thread on core %d. Queue "
4560 "%d on port \'%s\' will not be polled.",
4561 q
->core_id
, qid
, netdev_get_name(port
->netdev
));
4564 pmd
->isolated
= true;
4565 dp_netdev_pmd_unref(pmd
);
4567 } else if (!pinned
&& q
->core_id
== OVS_CORE_UNSPEC
) {
4568 uint64_t cycle_hist
= 0;
4571 rxqs
= xmalloc(sizeof *rxqs
);
4573 rxqs
= xrealloc(rxqs
, sizeof *rxqs
* (n_rxqs
+ 1));
4577 /* Sum the queue intervals and store the cycle history. */
4578 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
4579 cycle_hist
+= dp_netdev_rxq_get_intrvl_cycles(q
, i
);
4581 dp_netdev_rxq_set_cycles(q
, RXQ_CYCLES_PROC_HIST
,
4584 /* Store the queue. */
4590 if (n_rxqs
> 1 && assign_cyc
) {
4591 /* Sort the queues in order of the processing cycles
4592 * they consumed during their last pmd interval. */
4593 qsort(rxqs
, n_rxqs
, sizeof *rxqs
, compare_rxq_cycles
);
4596 rr_numa_list_populate(dp
, &rr
);
4597 /* Assign the sorted queues to pmds in round robin. */
4598 for (int i
= 0; i
< n_rxqs
; i
++) {
4599 numa_id
= netdev_get_numa_id(rxqs
[i
]->port
->netdev
);
4600 numa
= rr_numa_list_lookup(&rr
, numa_id
);
4602 /* There are no pmds on the queue's local NUMA node.
4603 Round robin on the NUMA nodes that do have pmds. */
4604 non_local_numa
= rr_numa_list_next(&rr
, non_local_numa
);
4605 if (!non_local_numa
) {
4606 VLOG_ERR("There is no available (non-isolated) pmd "
4607 "thread for port \'%s\' queue %d. This queue "
4608 "will not be polled. Is pmd-cpu-mask set to "
4609 "zero? Or are all PMDs isolated to other "
4610 "queues?", netdev_rxq_get_name(rxqs
[i
]->rx
),
4611 netdev_rxq_get_queue_id(rxqs
[i
]->rx
));
4614 rxqs
[i
]->pmd
= rr_numa_get_pmd(non_local_numa
, assign_cyc
);
4615 VLOG_WARN("There's no available (non-isolated) pmd thread "
4616 "on numa node %d. Queue %d on port \'%s\' will "
4617 "be assigned to the pmd on core %d "
4618 "(numa node %d). Expect reduced performance.",
4619 numa_id
, netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
4620 netdev_rxq_get_name(rxqs
[i
]->rx
),
4621 rxqs
[i
]->pmd
->core_id
, rxqs
[i
]->pmd
->numa_id
);
4623 rxqs
[i
]->pmd
= rr_numa_get_pmd(numa
, assign_cyc
);
4625 VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
4627 "(measured processing cycles %"PRIu64
").",
4628 rxqs
[i
]->pmd
->core_id
, numa_id
,
4629 netdev_rxq_get_name(rxqs
[i
]->rx
),
4630 netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
4631 dp_netdev_rxq_get_cycles(rxqs
[i
],
4632 RXQ_CYCLES_PROC_HIST
));
4634 VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
4635 "rx queue %d.", rxqs
[i
]->pmd
->core_id
, numa_id
,
4636 netdev_rxq_get_name(rxqs
[i
]->rx
),
4637 netdev_rxq_get_queue_id(rxqs
[i
]->rx
));
4642 rr_numa_list_destroy(&rr
);
4647 reload_affected_pmds(struct dp_netdev
*dp
)
4649 struct dp_netdev_pmd_thread
*pmd
;
4651 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4652 if (pmd
->need_reload
) {
4653 flow_mark_flush(pmd
);
4654 dp_netdev_reload_pmd__(pmd
);
4658 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4659 if (pmd
->need_reload
) {
4660 if (pmd
->core_id
!= NON_PMD_CORE_ID
) {
4664 atomic_read_explicit(&pmd
->reload
, &reload
,
4665 memory_order_acquire
);
4668 pmd
->need_reload
= false;
4674 reconfigure_pmd_threads(struct dp_netdev
*dp
)
4675 OVS_REQUIRES(dp
->port_mutex
)
4677 struct dp_netdev_pmd_thread
*pmd
;
4678 struct ovs_numa_dump
*pmd_cores
;
4679 struct ovs_numa_info_core
*core
;
4680 struct hmapx to_delete
= HMAPX_INITIALIZER(&to_delete
);
4681 struct hmapx_node
*node
;
4682 bool changed
= false;
4683 bool need_to_adjust_static_tx_qids
= false;
4685 /* The pmd threads should be started only if there's a pmd port in the
4686 * datapath. If the user didn't provide any "pmd-cpu-mask", we start
4687 * NR_PMD_THREADS per numa node. */
4688 if (!has_pmd_port(dp
)) {
4689 pmd_cores
= ovs_numa_dump_n_cores_per_numa(0);
4690 } else if (dp
->pmd_cmask
&& dp
->pmd_cmask
[0]) {
4691 pmd_cores
= ovs_numa_dump_cores_with_cmask(dp
->pmd_cmask
);
4693 pmd_cores
= ovs_numa_dump_n_cores_per_numa(NR_PMD_THREADS
);
4696 /* We need to adjust 'static_tx_qid's only if we're reducing number of
4697 * PMD threads. Otherwise, new threads will allocate all the freed ids. */
4698 if (ovs_numa_dump_count(pmd_cores
) < cmap_count(&dp
->poll_threads
) - 1) {
4699 /* Adjustment is required to keep 'static_tx_qid's sequential and
4700 * avoid possible issues, for example, imbalanced tx queue usage
4701 * and unnecessary locking caused by remapping on netdev level. */
4702 need_to_adjust_static_tx_qids
= true;
4705 /* Check for unwanted pmd threads */
4706 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4707 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
4710 if (!ovs_numa_dump_contains_core(pmd_cores
, pmd
->numa_id
,
4712 hmapx_add(&to_delete
, pmd
);
4713 } else if (need_to_adjust_static_tx_qids
) {
4714 atomic_store_relaxed(&pmd
->reload_tx_qid
, true);
4715 pmd
->need_reload
= true;
4719 HMAPX_FOR_EACH (node
, &to_delete
) {
4720 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
4721 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d destroyed.",
4722 pmd
->numa_id
, pmd
->core_id
);
4723 dp_netdev_del_pmd(dp
, pmd
);
4725 changed
= !hmapx_is_empty(&to_delete
);
4726 hmapx_destroy(&to_delete
);
4728 if (need_to_adjust_static_tx_qids
) {
4729 /* 'static_tx_qid's are not sequential now.
4730 * Reload remaining threads to fix this. */
4731 reload_affected_pmds(dp
);
4734 /* Check for required new pmd threads */
4735 FOR_EACH_CORE_ON_DUMP(core
, pmd_cores
) {
4736 pmd
= dp_netdev_get_pmd(dp
, core
->core_id
);
4738 pmd
= xzalloc(sizeof *pmd
);
4739 dp_netdev_configure_pmd(pmd
, dp
, core
->core_id
, core
->numa_id
);
4740 pmd
->thread
= ovs_thread_create("pmd", pmd_thread_main
, pmd
);
4741 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d created.",
4742 pmd
->numa_id
, pmd
->core_id
);
4745 dp_netdev_pmd_unref(pmd
);
4750 struct ovs_numa_info_numa
*numa
;
4752 /* Log the number of pmd threads per numa node. */
4753 FOR_EACH_NUMA_ON_DUMP (numa
, pmd_cores
) {
4754 VLOG_INFO("There are %"PRIuSIZE
" pmd threads on numa node %d",
4755 numa
->n_cores
, numa
->numa_id
);
4759 ovs_numa_dump_destroy(pmd_cores
);
4763 pmd_remove_stale_ports(struct dp_netdev
*dp
,
4764 struct dp_netdev_pmd_thread
*pmd
)
4765 OVS_EXCLUDED(pmd
->port_mutex
)
4766 OVS_REQUIRES(dp
->port_mutex
)
4768 struct rxq_poll
*poll
, *poll_next
;
4769 struct tx_port
*tx
, *tx_next
;
4771 ovs_mutex_lock(&pmd
->port_mutex
);
4772 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
4773 struct dp_netdev_port
*port
= poll
->rxq
->port
;
4775 if (port
->need_reconfigure
4776 || !hmap_contains(&dp
->ports
, &port
->node
)) {
4777 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
4780 HMAP_FOR_EACH_SAFE (tx
, tx_next
, node
, &pmd
->tx_ports
) {
4781 struct dp_netdev_port
*port
= tx
->port
;
4783 if (port
->need_reconfigure
4784 || !hmap_contains(&dp
->ports
, &port
->node
)) {
4785 dp_netdev_del_port_tx_from_pmd(pmd
, tx
);
4788 ovs_mutex_unlock(&pmd
->port_mutex
);
4791 /* Must be called each time a port is added/removed or the cmask changes.
4792 * This creates and destroys pmd threads, reconfigures ports, opens their
4793 * rxqs and assigns all rxqs/txqs to pmd threads. */
4795 reconfigure_datapath(struct dp_netdev
*dp
)
4796 OVS_REQUIRES(dp
->port_mutex
)
4798 struct hmapx busy_threads
= HMAPX_INITIALIZER(&busy_threads
);
4799 struct dp_netdev_pmd_thread
*pmd
;
4800 struct dp_netdev_port
*port
;
4803 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
4805 /* Step 1: Adjust the pmd threads based on the datapath ports, the cores
4806 * on the system and the user configuration. */
4807 reconfigure_pmd_threads(dp
);
4809 wanted_txqs
= cmap_count(&dp
->poll_threads
);
4811 /* The number of pmd threads might have changed, or a port can be new:
4812 * adjust the txqs. */
4813 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4814 netdev_set_tx_multiq(port
->netdev
, wanted_txqs
);
4817 /* Step 2: Remove from the pmd threads ports that have been removed or
4818 * need reconfiguration. */
4820 /* Check for all the ports that need reconfiguration. We cache this in
4821 * 'port->need_reconfigure', because netdev_is_reconf_required() can
4822 * change at any time. */
4823 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4824 if (netdev_is_reconf_required(port
->netdev
)) {
4825 port
->need_reconfigure
= true;
4829 /* Remove from the pmd threads all the ports that have been deleted or
4830 * need reconfiguration. */
4831 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4832 pmd_remove_stale_ports(dp
, pmd
);
4835 /* Reload affected pmd threads. We must wait for the pmd threads before
4836 * reconfiguring the ports, because a port cannot be reconfigured while
4837 * it's being used. */
4838 reload_affected_pmds(dp
);
4840 /* Step 3: Reconfigure ports. */
4842 /* We only reconfigure the ports that we determined above, because they're
4843 * not being used by any pmd thread at the moment. If a port fails to
4844 * reconfigure we remove it from the datapath. */
4845 struct dp_netdev_port
*next_port
;
4846 HMAP_FOR_EACH_SAFE (port
, next_port
, node
, &dp
->ports
) {
4849 if (!port
->need_reconfigure
) {
4853 err
= port_reconfigure(port
);
4855 hmap_remove(&dp
->ports
, &port
->node
);
4856 seq_change(dp
->port_seq
);
4859 port
->dynamic_txqs
= netdev_n_txq(port
->netdev
) < wanted_txqs
;
4863 /* Step 4: Compute new rxq scheduling. We don't touch the pmd threads
4864 * for now, we just update the 'pmd' pointer in each rxq to point to the
4865 * wanted thread according to the scheduling policy. */
4867 /* Reset all the pmd threads to non isolated. */
4868 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4869 pmd
->isolated
= false;
4872 /* Reset all the queues to unassigned */
4873 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4874 for (int i
= 0; i
< port
->n_rxq
; i
++) {
4875 port
->rxqs
[i
].pmd
= NULL
;
4879 /* Add pinned queues and mark pmd threads isolated. */
4880 rxq_scheduling(dp
, true);
4882 /* Add non-pinned queues. */
4883 rxq_scheduling(dp
, false);
4885 /* Step 5: Remove queues not compliant with new scheduling. */
4887 /* Count all the threads that will have at least one queue to poll. */
4888 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4889 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
4890 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
4893 hmapx_add(&busy_threads
, q
->pmd
);
4898 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4899 struct rxq_poll
*poll
, *poll_next
;
4901 ovs_mutex_lock(&pmd
->port_mutex
);
4902 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
4903 if (poll
->rxq
->pmd
!= pmd
) {
4904 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
4906 /* This pmd might sleep after this step if it has no rxq
4907 * remaining. Tell it to busy wait for new assignment if it
4908 * has at least one scheduled queue. */
4909 if (hmap_count(&pmd
->poll_list
) == 0 &&
4910 hmapx_contains(&busy_threads
, pmd
)) {
4911 atomic_store_relaxed(&pmd
->wait_for_reload
, true);
4915 ovs_mutex_unlock(&pmd
->port_mutex
);
4918 hmapx_destroy(&busy_threads
);
4920 /* Reload affected pmd threads. We must wait for the pmd threads to remove
4921 * the old queues before readding them, otherwise a queue can be polled by
4922 * two threads at the same time. */
4923 reload_affected_pmds(dp
);
4925 /* Step 6: Add queues from scheduling, if they're not there already. */
4926 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4927 if (!netdev_is_pmd(port
->netdev
)) {
4931 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
4932 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
4935 ovs_mutex_lock(&q
->pmd
->port_mutex
);
4936 dp_netdev_add_rxq_to_pmd(q
->pmd
, q
);
4937 ovs_mutex_unlock(&q
->pmd
->port_mutex
);
4942 /* Add every port to the tx cache of every pmd thread, if it's not
4943 * there already and if this pmd has at least one rxq to poll. */
4944 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4945 ovs_mutex_lock(&pmd
->port_mutex
);
4946 if (hmap_count(&pmd
->poll_list
) || pmd
->core_id
== NON_PMD_CORE_ID
) {
4947 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4948 dp_netdev_add_port_tx_to_pmd(pmd
, port
);
4951 ovs_mutex_unlock(&pmd
->port_mutex
);
4954 /* Reload affected pmd threads. */
4955 reload_affected_pmds(dp
);
4957 /* Check if PMD Auto LB is to be enabled */
4958 set_pmd_auto_lb(dp
);
4961 /* Returns true if one of the netdevs in 'dp' requires a reconfiguration */
4963 ports_require_restart(const struct dp_netdev
*dp
)
4964 OVS_REQUIRES(dp
->port_mutex
)
4966 struct dp_netdev_port
*port
;
4968 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4969 if (netdev_is_reconf_required(port
->netdev
)) {
4977 /* Calculates variance in the values stored in array 'a'. 'n' is the number
4978 * of elements in array to be considered for calculating vairance.
4979 * Usage example: data array 'a' contains the processing load of each pmd and
4980 * 'n' is the number of PMDs. It returns the variance in processing load of
4983 variance(uint64_t a
[], int n
)
4985 /* Compute mean (average of elements). */
4988 uint64_t sqDiff
= 0;
4994 for (int i
= 0; i
< n
; i
++) {
5001 /* Compute sum squared differences with mean. */
5002 for (int i
= 0; i
< n
; i
++) {
5003 sqDiff
+= (a
[i
] - mean
)*(a
[i
] - mean
);
5006 return (sqDiff
? (sqDiff
/ n
) : 0);
5010 /* Returns the variance in the PMDs usage as part of dry run of rxqs
5011 * assignment to PMDs. */
5013 get_dry_run_variance(struct dp_netdev
*dp
, uint32_t *core_list
,
5014 uint32_t num_pmds
, uint64_t *predicted_variance
)
5015 OVS_REQUIRES(dp
->port_mutex
)
5017 struct dp_netdev_port
*port
;
5018 struct dp_netdev_pmd_thread
*pmd
;
5019 struct dp_netdev_rxq
**rxqs
= NULL
;
5020 struct rr_numa
*numa
= NULL
;
5021 struct rr_numa_list rr
;
5024 uint64_t *pmd_usage
;
5026 if (!predicted_variance
) {
5030 pmd_usage
= xcalloc(num_pmds
, sizeof(uint64_t));
5032 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5033 if (!netdev_is_pmd(port
->netdev
)) {
5037 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
5038 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
5039 uint64_t cycle_hist
= 0;
5041 if (q
->pmd
->isolated
) {
5046 rxqs
= xmalloc(sizeof *rxqs
);
5048 rxqs
= xrealloc(rxqs
, sizeof *rxqs
* (n_rxqs
+ 1));
5051 /* Sum the queue intervals and store the cycle history. */
5052 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
5053 cycle_hist
+= dp_netdev_rxq_get_intrvl_cycles(q
, i
);
5055 dp_netdev_rxq_set_cycles(q
, RXQ_CYCLES_PROC_HIST
,
5057 /* Store the queue. */
5062 /* Sort the queues in order of the processing cycles
5063 * they consumed during their last pmd interval. */
5064 qsort(rxqs
, n_rxqs
, sizeof *rxqs
, compare_rxq_cycles
);
5066 rr_numa_list_populate(dp
, &rr
);
5068 for (int i
= 0; i
< n_rxqs
; i
++) {
5069 int numa_id
= netdev_get_numa_id(rxqs
[i
]->port
->netdev
);
5070 numa
= rr_numa_list_lookup(&rr
, numa_id
);
5072 /* Abort if cross NUMA polling. */
5073 VLOG_DBG("PMD auto lb dry run."
5074 " Aborting due to cross-numa polling.");
5078 pmd
= rr_numa_get_pmd(numa
, true);
5079 VLOG_DBG("PMD auto lb dry run. Predicted: Core %d on numa node %d "
5080 "to be assigned port \'%s\' rx queue %d "
5081 "(measured processing cycles %"PRIu64
").",
5082 pmd
->core_id
, numa_id
,
5083 netdev_rxq_get_name(rxqs
[i
]->rx
),
5084 netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
5085 dp_netdev_rxq_get_cycles(rxqs
[i
], RXQ_CYCLES_PROC_HIST
));
5087 for (int id
= 0; id
< num_pmds
; id
++) {
5088 if (pmd
->core_id
== core_list
[id
]) {
5089 /* Add the processing cycles of rxq to pmd polling it. */
5090 pmd_usage
[id
] += dp_netdev_rxq_get_cycles(rxqs
[i
],
5091 RXQ_CYCLES_PROC_HIST
);
5096 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
5097 uint64_t total_cycles
= 0;
5099 if ((pmd
->core_id
== NON_PMD_CORE_ID
) || pmd
->isolated
) {
5103 /* Get the total pmd cycles for an interval. */
5104 atomic_read_relaxed(&pmd
->intrvl_cycles
, &total_cycles
);
5105 /* Estimate the cycles to cover all intervals. */
5106 total_cycles
*= PMD_RXQ_INTERVAL_MAX
;
5107 for (int id
= 0; id
< num_pmds
; id
++) {
5108 if (pmd
->core_id
== core_list
[id
]) {
5109 if (pmd_usage
[id
]) {
5110 pmd_usage
[id
] = (pmd_usage
[id
] * 100) / total_cycles
;
5112 VLOG_DBG("PMD auto lb dry run. Predicted: Core %d, "
5113 "usage %"PRIu64
"", pmd
->core_id
, pmd_usage
[id
]);
5117 *predicted_variance
= variance(pmd_usage
, num_pmds
);
5121 rr_numa_list_destroy(&rr
);
5127 /* Does the dry run of Rxq assignment to PMDs and returns true if it gives
5128 * better distribution of load on PMDs. */
5130 pmd_rebalance_dry_run(struct dp_netdev
*dp
)
5131 OVS_REQUIRES(dp
->port_mutex
)
5133 struct dp_netdev_pmd_thread
*pmd
;
5134 uint64_t *curr_pmd_usage
;
5136 uint64_t curr_variance
;
5137 uint64_t new_variance
;
5138 uint64_t improvement
= 0;
5140 uint32_t *pmd_corelist
;
5141 struct rxq_poll
*poll
;
5144 num_pmds
= cmap_count(&dp
->poll_threads
);
5147 curr_pmd_usage
= xcalloc(num_pmds
, sizeof(uint64_t));
5148 pmd_corelist
= xcalloc(num_pmds
, sizeof(uint32_t));
5154 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
5155 uint64_t total_cycles
= 0;
5156 uint64_t total_proc
= 0;
5158 if ((pmd
->core_id
== NON_PMD_CORE_ID
) || pmd
->isolated
) {
5162 /* Get the total pmd cycles for an interval. */
5163 atomic_read_relaxed(&pmd
->intrvl_cycles
, &total_cycles
);
5164 /* Estimate the cycles to cover all intervals. */
5165 total_cycles
*= PMD_RXQ_INTERVAL_MAX
;
5167 ovs_mutex_lock(&pmd
->port_mutex
);
5168 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
5169 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
5170 total_proc
+= dp_netdev_rxq_get_intrvl_cycles(poll
->rxq
, i
);
5173 ovs_mutex_unlock(&pmd
->port_mutex
);
5176 curr_pmd_usage
[num_pmds
] = (total_proc
* 100) / total_cycles
;
5179 VLOG_DBG("PMD auto lb dry run. Current: Core %d, usage %"PRIu64
"",
5180 pmd
->core_id
, curr_pmd_usage
[num_pmds
]);
5182 if (atomic_count_get(&pmd
->pmd_overloaded
)) {
5183 atomic_count_set(&pmd
->pmd_overloaded
, 0);
5186 pmd_corelist
[num_pmds
] = pmd
->core_id
;
5190 curr_variance
= variance(curr_pmd_usage
, num_pmds
);
5191 ret
= get_dry_run_variance(dp
, pmd_corelist
, num_pmds
, &new_variance
);
5194 VLOG_DBG("PMD auto lb dry run. Current PMD variance: %"PRIu64
","
5195 " Predicted PMD variance: %"PRIu64
"",
5196 curr_variance
, new_variance
);
5198 if (new_variance
< curr_variance
) {
5200 ((curr_variance
- new_variance
) * 100) / curr_variance
;
5202 if (improvement
< ALB_ACCEPTABLE_IMPROVEMENT
) {
5207 free(curr_pmd_usage
);
5213 /* Return true if needs to revalidate datapath flows. */
5215 dpif_netdev_run(struct dpif
*dpif
)
5217 struct dp_netdev_port
*port
;
5218 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5219 struct dp_netdev_pmd_thread
*non_pmd
;
5220 uint64_t new_tnl_seq
;
5221 bool need_to_flush
= true;
5222 bool pmd_rebalance
= false;
5223 long long int now
= time_msec();
5224 struct dp_netdev_pmd_thread
*pmd
;
5226 ovs_mutex_lock(&dp
->port_mutex
);
5227 non_pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
5229 ovs_mutex_lock(&dp
->non_pmd_mutex
);
5230 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5231 if (!netdev_is_pmd(port
->netdev
)) {
5234 if (port
->emc_enabled
) {
5235 atomic_read_relaxed(&dp
->emc_insert_min
,
5236 &non_pmd
->ctx
.emc_insert_min
);
5238 non_pmd
->ctx
.emc_insert_min
= 0;
5241 for (i
= 0; i
< port
->n_rxq
; i
++) {
5243 if (!netdev_rxq_enabled(port
->rxqs
[i
].rx
)) {
5247 if (dp_netdev_process_rxq_port(non_pmd
,
5250 need_to_flush
= false;
5255 if (need_to_flush
) {
5256 /* We didn't receive anything in the process loop.
5257 * Check if we need to send something.
5258 * There was no time updates on current iteration. */
5259 pmd_thread_ctx_time_update(non_pmd
);
5260 dp_netdev_pmd_flush_output_packets(non_pmd
, false);
5263 dpif_netdev_xps_revalidate_pmd(non_pmd
, false);
5264 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
5266 dp_netdev_pmd_unref(non_pmd
);
5269 struct pmd_auto_lb
*pmd_alb
= &dp
->pmd_alb
;
5270 if (pmd_alb
->is_enabled
) {
5271 if (!pmd_alb
->rebalance_poll_timer
) {
5272 pmd_alb
->rebalance_poll_timer
= now
;
5273 } else if ((pmd_alb
->rebalance_poll_timer
+
5274 pmd_alb
->rebalance_intvl
) < now
) {
5275 pmd_alb
->rebalance_poll_timer
= now
;
5276 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
5277 if (atomic_count_get(&pmd
->pmd_overloaded
) >=
5278 PMD_RXQ_INTERVAL_MAX
) {
5279 pmd_rebalance
= true;
5284 if (pmd_rebalance
&&
5285 !dp_netdev_is_reconf_required(dp
) &&
5286 !ports_require_restart(dp
) &&
5287 pmd_rebalance_dry_run(dp
)) {
5288 VLOG_INFO("PMD auto lb dry run."
5289 " requesting datapath reconfigure.");
5290 dp_netdev_request_reconfigure(dp
);
5295 if (dp_netdev_is_reconf_required(dp
) || ports_require_restart(dp
)) {
5296 reconfigure_datapath(dp
);
5298 ovs_mutex_unlock(&dp
->port_mutex
);
5300 tnl_neigh_cache_run();
5302 new_tnl_seq
= seq_read(tnl_conf_seq
);
5304 if (dp
->last_tnl_conf_seq
!= new_tnl_seq
) {
5305 dp
->last_tnl_conf_seq
= new_tnl_seq
;
5312 dpif_netdev_wait(struct dpif
*dpif
)
5314 struct dp_netdev_port
*port
;
5315 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5317 ovs_mutex_lock(&dp_netdev_mutex
);
5318 ovs_mutex_lock(&dp
->port_mutex
);
5319 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5320 netdev_wait_reconf_required(port
->netdev
);
5321 if (!netdev_is_pmd(port
->netdev
)) {
5324 for (i
= 0; i
< port
->n_rxq
; i
++) {
5325 netdev_rxq_wait(port
->rxqs
[i
].rx
);
5329 ovs_mutex_unlock(&dp
->port_mutex
);
5330 ovs_mutex_unlock(&dp_netdev_mutex
);
5331 seq_wait(tnl_conf_seq
, dp
->last_tnl_conf_seq
);
5335 pmd_free_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
5337 struct tx_port
*tx_port_cached
;
5339 /* Flush all the queued packets. */
5340 dp_netdev_pmd_flush_output_packets(pmd
, true);
5341 /* Free all used tx queue ids. */
5342 dpif_netdev_xps_revalidate_pmd(pmd
, true);
5344 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->tnl_port_cache
) {
5345 free(tx_port_cached
);
5347 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->send_port_cache
) {
5348 free(tx_port_cached
);
5352 /* Copies ports from 'pmd->tx_ports' (shared with the main thread) to
5353 * thread-local copies. Copy to 'pmd->tnl_port_cache' if it is a tunnel
5354 * device, otherwise to 'pmd->send_port_cache' if the port has at least
5357 pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
5358 OVS_REQUIRES(pmd
->port_mutex
)
5360 struct tx_port
*tx_port
, *tx_port_cached
;
5362 pmd_free_cached_ports(pmd
);
5363 hmap_shrink(&pmd
->send_port_cache
);
5364 hmap_shrink(&pmd
->tnl_port_cache
);
5366 HMAP_FOR_EACH (tx_port
, node
, &pmd
->tx_ports
) {
5367 if (netdev_has_tunnel_push_pop(tx_port
->port
->netdev
)) {
5368 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
5369 hmap_insert(&pmd
->tnl_port_cache
, &tx_port_cached
->node
,
5370 hash_port_no(tx_port_cached
->port
->port_no
));
5373 if (netdev_n_txq(tx_port
->port
->netdev
)) {
5374 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
5375 hmap_insert(&pmd
->send_port_cache
, &tx_port_cached
->node
,
5376 hash_port_no(tx_port_cached
->port
->port_no
));
5382 pmd_alloc_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
5384 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
5385 if (!id_pool_alloc_id(pmd
->dp
->tx_qid_pool
, &pmd
->static_tx_qid
)) {
5386 VLOG_ABORT("static_tx_qid allocation failed for PMD on core %2d"
5387 ", numa_id %d.", pmd
->core_id
, pmd
->numa_id
);
5389 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
5391 VLOG_DBG("static_tx_qid = %d allocated for PMD thread on core %2d"
5392 ", numa_id %d.", pmd
->static_tx_qid
, pmd
->core_id
, pmd
->numa_id
);
5396 pmd_free_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
5398 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
5399 id_pool_free_id(pmd
->dp
->tx_qid_pool
, pmd
->static_tx_qid
);
5400 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
5404 pmd_load_queues_and_ports(struct dp_netdev_pmd_thread
*pmd
,
5405 struct polled_queue
**ppoll_list
)
5407 struct polled_queue
*poll_list
= *ppoll_list
;
5408 struct rxq_poll
*poll
;
5411 ovs_mutex_lock(&pmd
->port_mutex
);
5412 poll_list
= xrealloc(poll_list
, hmap_count(&pmd
->poll_list
)
5413 * sizeof *poll_list
);
5416 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
5417 poll_list
[i
].rxq
= poll
->rxq
;
5418 poll_list
[i
].port_no
= poll
->rxq
->port
->port_no
;
5419 poll_list
[i
].emc_enabled
= poll
->rxq
->port
->emc_enabled
;
5420 poll_list
[i
].rxq_enabled
= netdev_rxq_enabled(poll
->rxq
->rx
);
5421 poll_list
[i
].change_seq
=
5422 netdev_get_change_seq(poll
->rxq
->port
->netdev
);
5426 pmd_load_cached_ports(pmd
);
5428 ovs_mutex_unlock(&pmd
->port_mutex
);
5430 *ppoll_list
= poll_list
;
5435 pmd_thread_main(void *f_
)
5437 struct dp_netdev_pmd_thread
*pmd
= f_
;
5438 struct pmd_perf_stats
*s
= &pmd
->perf_stats
;
5439 unsigned int lc
= 0;
5440 struct polled_queue
*poll_list
;
5441 bool wait_for_reload
= false;
5447 int process_packets
= 0;
5451 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
5452 ovsthread_setspecific(pmd
->dp
->per_pmd_key
, pmd
);
5453 ovs_numa_thread_setaffinity_core(pmd
->core_id
);
5454 dpdk_set_lcore_id(pmd
->core_id
);
5455 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
5456 dfc_cache_init(&pmd
->flow_cache
);
5457 pmd_alloc_static_tx_qid(pmd
);
5460 atomic_count_init(&pmd
->pmd_overloaded
, 0);
5462 /* List port/core affinity */
5463 for (i
= 0; i
< poll_cnt
; i
++) {
5464 VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n",
5465 pmd
->core_id
, netdev_rxq_get_name(poll_list
[i
].rxq
->rx
),
5466 netdev_rxq_get_queue_id(poll_list
[i
].rxq
->rx
));
5467 /* Reset the rxq current cycles counter. */
5468 dp_netdev_rxq_set_cycles(poll_list
[i
].rxq
, RXQ_CYCLES_PROC_CURR
, 0);
5472 if (wait_for_reload
) {
5473 /* Don't sleep, control thread will ask for a reload shortly. */
5475 atomic_read_explicit(&pmd
->reload
, &reload
,
5476 memory_order_acquire
);
5479 while (seq_read(pmd
->reload_seq
) == pmd
->last_reload_seq
) {
5480 seq_wait(pmd
->reload_seq
, pmd
->last_reload_seq
);
5486 pmd
->intrvl_tsc_prev
= 0;
5487 atomic_store_relaxed(&pmd
->intrvl_cycles
, 0);
5488 cycles_counter_update(s
);
5489 /* Protect pmd stats from external clearing while polling. */
5490 ovs_mutex_lock(&pmd
->perf_stats
.stats_mutex
);
5492 uint64_t rx_packets
= 0, tx_packets
= 0;
5494 pmd_perf_start_iteration(s
);
5496 for (i
= 0; i
< poll_cnt
; i
++) {
5498 if (!poll_list
[i
].rxq_enabled
) {
5502 if (poll_list
[i
].emc_enabled
) {
5503 atomic_read_relaxed(&pmd
->dp
->emc_insert_min
,
5504 &pmd
->ctx
.emc_insert_min
);
5506 pmd
->ctx
.emc_insert_min
= 0;
5510 dp_netdev_process_rxq_port(pmd
, poll_list
[i
].rxq
,
5511 poll_list
[i
].port_no
);
5512 rx_packets
+= process_packets
;
5516 /* We didn't receive anything in the process loop.
5517 * Check if we need to send something.
5518 * There was no time updates on current iteration. */
5519 pmd_thread_ctx_time_update(pmd
);
5520 tx_packets
= dp_netdev_pmd_flush_output_packets(pmd
, false);
5526 coverage_try_clear();
5527 dp_netdev_pmd_try_optimize(pmd
, poll_list
, poll_cnt
);
5528 if (!ovsrcu_try_quiesce()) {
5529 emc_cache_slow_sweep(&((pmd
->flow_cache
).emc_cache
));
5532 for (i
= 0; i
< poll_cnt
; i
++) {
5533 uint64_t current_seq
=
5534 netdev_get_change_seq(poll_list
[i
].rxq
->port
->netdev
);
5535 if (poll_list
[i
].change_seq
!= current_seq
) {
5536 poll_list
[i
].change_seq
= current_seq
;
5537 poll_list
[i
].rxq_enabled
=
5538 netdev_rxq_enabled(poll_list
[i
].rxq
->rx
);
5543 atomic_read_explicit(&pmd
->reload
, &reload
, memory_order_acquire
);
5544 if (OVS_UNLIKELY(reload
)) {
5548 pmd_perf_end_iteration(s
, rx_packets
, tx_packets
,
5549 pmd_perf_metrics_enabled(pmd
));
5551 ovs_mutex_unlock(&pmd
->perf_stats
.stats_mutex
);
5553 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
5554 atomic_read_relaxed(&pmd
->wait_for_reload
, &wait_for_reload
);
5555 atomic_read_relaxed(&pmd
->reload_tx_qid
, &reload_tx_qid
);
5556 atomic_read_relaxed(&pmd
->exit
, &exiting
);
5557 /* Signal here to make sure the pmd finishes
5558 * reloading the updated configuration. */
5559 dp_netdev_pmd_reload_done(pmd
);
5561 if (reload_tx_qid
) {
5562 pmd_free_static_tx_qid(pmd
);
5563 pmd_alloc_static_tx_qid(pmd
);
5570 pmd_free_static_tx_qid(pmd
);
5571 dfc_cache_uninit(&pmd
->flow_cache
);
5573 pmd_free_cached_ports(pmd
);
5578 dp_netdev_disable_upcall(struct dp_netdev
*dp
)
5579 OVS_ACQUIRES(dp
->upcall_rwlock
)
5581 fat_rwlock_wrlock(&dp
->upcall_rwlock
);
5587 dpif_netdev_meter_get_features(const struct dpif
* dpif OVS_UNUSED
,
5588 struct ofputil_meter_features
*features
)
5590 features
->max_meters
= MAX_METERS
;
5591 features
->band_types
= DP_SUPPORTED_METER_BAND_TYPES
;
5592 features
->capabilities
= DP_SUPPORTED_METER_FLAGS_MASK
;
5593 features
->max_bands
= MAX_BANDS
;
5594 features
->max_color
= 0;
5597 /* Applies the meter identified by 'meter_id' to 'packets_'. Packets
5598 * that exceed a band are dropped in-place. */
5600 dp_netdev_run_meter(struct dp_netdev
*dp
, struct dp_packet_batch
*packets_
,
5601 uint32_t meter_id
, long long int now
)
5603 struct dp_meter
*meter
;
5604 struct dp_meter_band
*band
;
5605 struct dp_packet
*packet
;
5606 long long int long_delta_t
; /* msec */
5607 uint32_t delta_t
; /* msec */
5608 const size_t cnt
= dp_packet_batch_size(packets_
);
5609 uint32_t bytes
, volume
;
5610 int exceeded_band
[NETDEV_MAX_BURST
];
5611 uint32_t exceeded_rate
[NETDEV_MAX_BURST
];
5612 int exceeded_pkt
= cnt
; /* First packet that exceeded a band rate. */
5614 if (meter_id
>= MAX_METERS
) {
5618 meter_lock(dp
, meter_id
);
5619 meter
= dp
->meters
[meter_id
];
5624 /* Initialize as negative values. */
5625 memset(exceeded_band
, 0xff, cnt
* sizeof *exceeded_band
);
5626 /* Initialize as zeroes. */
5627 memset(exceeded_rate
, 0, cnt
* sizeof *exceeded_rate
);
5629 /* All packets will hit the meter at the same time. */
5630 long_delta_t
= now
/ 1000 - meter
->used
/ 1000; /* msec */
5632 /* Make sure delta_t will not be too large, so that bucket will not
5633 * wrap around below. */
5634 delta_t
= (long_delta_t
> (long long int)meter
->max_delta_t
)
5635 ? meter
->max_delta_t
: (uint32_t)long_delta_t
;
5637 /* Update meter stats. */
5639 meter
->packet_count
+= cnt
;
5641 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
5642 bytes
+= dp_packet_size(packet
);
5644 meter
->byte_count
+= bytes
;
5646 /* Meters can operate in terms of packets per second or kilobits per
5648 if (meter
->flags
& OFPMF13_PKTPS
) {
5649 /* Rate in packets/second, bucket 1/1000 packets. */
5650 /* msec * packets/sec = 1/1000 packets. */
5651 volume
= cnt
* 1000; /* Take 'cnt' packets from the bucket. */
5653 /* Rate in kbps, bucket in bits. */
5654 /* msec * kbps = bits */
5658 /* Update all bands and find the one hit with the highest rate for each
5659 * packet (if any). */
5660 for (int m
= 0; m
< meter
->n_bands
; ++m
) {
5661 band
= &meter
->bands
[m
];
5663 /* Update band's bucket. */
5664 band
->bucket
+= delta_t
* band
->up
.rate
;
5665 if (band
->bucket
> band
->up
.burst_size
) {
5666 band
->bucket
= band
->up
.burst_size
;
5669 /* Drain the bucket for all the packets, if possible. */
5670 if (band
->bucket
>= volume
) {
5671 band
->bucket
-= volume
;
5673 int band_exceeded_pkt
;
5675 /* Band limit hit, must process packet-by-packet. */
5676 if (meter
->flags
& OFPMF13_PKTPS
) {
5677 band_exceeded_pkt
= band
->bucket
/ 1000;
5678 band
->bucket
%= 1000; /* Remainder stays in bucket. */
5680 /* Update the exceeding band for each exceeding packet.
5681 * (Only one band will be fired by a packet, and that
5682 * can be different for each packet.) */
5683 for (int i
= band_exceeded_pkt
; i
< cnt
; i
++) {
5684 if (band
->up
.rate
> exceeded_rate
[i
]) {
5685 exceeded_rate
[i
] = band
->up
.rate
;
5686 exceeded_band
[i
] = m
;
5690 /* Packet sizes differ, must process one-by-one. */
5691 band_exceeded_pkt
= cnt
;
5692 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
5693 uint32_t bits
= dp_packet_size(packet
) * 8;
5695 if (band
->bucket
>= bits
) {
5696 band
->bucket
-= bits
;
5698 if (i
< band_exceeded_pkt
) {
5699 band_exceeded_pkt
= i
;
5701 /* Update the exceeding band for the exceeding packet.
5702 * (Only one band will be fired by a packet, and that
5703 * can be different for each packet.) */
5704 if (band
->up
.rate
> exceeded_rate
[i
]) {
5705 exceeded_rate
[i
] = band
->up
.rate
;
5706 exceeded_band
[i
] = m
;
5711 /* Remember the first exceeding packet. */
5712 if (exceeded_pkt
> band_exceeded_pkt
) {
5713 exceeded_pkt
= band_exceeded_pkt
;
5718 /* Fire the highest rate band exceeded by each packet, and drop
5719 * packets if needed. */
5721 DP_PACKET_BATCH_REFILL_FOR_EACH (j
, cnt
, packet
, packets_
) {
5722 if (exceeded_band
[j
] >= 0) {
5723 /* Meter drop packet. */
5724 band
= &meter
->bands
[exceeded_band
[j
]];
5725 band
->packet_count
+= 1;
5726 band
->byte_count
+= dp_packet_size(packet
);
5728 dp_packet_delete(packet
);
5730 /* Meter accepts packet. */
5731 dp_packet_batch_refill(packets_
, packet
, j
);
5735 meter_unlock(dp
, meter_id
);
5738 /* Meter set/get/del processing is still single-threaded. */
5740 dpif_netdev_meter_set(struct dpif
*dpif
, ofproto_meter_id meter_id
,
5741 struct ofputil_meter_config
*config
)
5743 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5744 uint32_t mid
= meter_id
.uint32
;
5745 struct dp_meter
*meter
;
5748 if (mid
>= MAX_METERS
) {
5749 return EFBIG
; /* Meter_id out of range. */
5752 if (config
->flags
& ~DP_SUPPORTED_METER_FLAGS_MASK
) {
5753 return EBADF
; /* Unsupported flags set */
5756 if (config
->n_bands
> MAX_BANDS
) {
5760 for (i
= 0; i
< config
->n_bands
; ++i
) {
5761 switch (config
->bands
[i
].type
) {
5765 return ENODEV
; /* Unsupported band type */
5769 /* Allocate meter */
5770 meter
= xzalloc(sizeof *meter
5771 + config
->n_bands
* sizeof(struct dp_meter_band
));
5773 meter
->flags
= config
->flags
;
5774 meter
->n_bands
= config
->n_bands
;
5775 meter
->max_delta_t
= 0;
5776 meter
->used
= time_usec();
5779 for (i
= 0; i
< config
->n_bands
; ++i
) {
5780 uint32_t band_max_delta_t
;
5782 /* Set burst size to a workable value if none specified. */
5783 if (config
->bands
[i
].burst_size
== 0) {
5784 config
->bands
[i
].burst_size
= config
->bands
[i
].rate
;
5787 meter
->bands
[i
].up
= config
->bands
[i
];
5788 /* Convert burst size to the bucket units: */
5789 /* pkts => 1/1000 packets, kilobits => bits. */
5790 meter
->bands
[i
].up
.burst_size
*= 1000;
5791 /* Initialize bucket to empty. */
5792 meter
->bands
[i
].bucket
= 0;
5794 /* Figure out max delta_t that is enough to fill any bucket. */
5796 = meter
->bands
[i
].up
.burst_size
/ meter
->bands
[i
].up
.rate
;
5797 if (band_max_delta_t
> meter
->max_delta_t
) {
5798 meter
->max_delta_t
= band_max_delta_t
;
5802 meter_lock(dp
, mid
);
5803 dp_delete_meter(dp
, mid
); /* Free existing meter, if any */
5804 dp
->meters
[mid
] = meter
;
5805 meter_unlock(dp
, mid
);
5811 dpif_netdev_meter_get(const struct dpif
*dpif
,
5812 ofproto_meter_id meter_id_
,
5813 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
5815 const struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5816 uint32_t meter_id
= meter_id_
.uint32
;
5819 if (meter_id
>= MAX_METERS
) {
5823 meter_lock(dp
, meter_id
);
5824 const struct dp_meter
*meter
= dp
->meters
[meter_id
];
5832 stats
->packet_in_count
= meter
->packet_count
;
5833 stats
->byte_in_count
= meter
->byte_count
;
5835 for (i
= 0; i
< n_bands
&& i
< meter
->n_bands
; ++i
) {
5836 stats
->bands
[i
].packet_count
= meter
->bands
[i
].packet_count
;
5837 stats
->bands
[i
].byte_count
= meter
->bands
[i
].byte_count
;
5844 meter_unlock(dp
, meter_id
);
5849 dpif_netdev_meter_del(struct dpif
*dpif
,
5850 ofproto_meter_id meter_id_
,
5851 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
5853 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5856 error
= dpif_netdev_meter_get(dpif
, meter_id_
, stats
, n_bands
);
5858 uint32_t meter_id
= meter_id_
.uint32
;
5860 meter_lock(dp
, meter_id
);
5861 dp_delete_meter(dp
, meter_id
);
5862 meter_unlock(dp
, meter_id
);
5869 dpif_netdev_disable_upcall(struct dpif
*dpif
)
5870 OVS_NO_THREAD_SAFETY_ANALYSIS
5872 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5873 dp_netdev_disable_upcall(dp
);
5877 dp_netdev_enable_upcall(struct dp_netdev
*dp
)
5878 OVS_RELEASES(dp
->upcall_rwlock
)
5880 fat_rwlock_unlock(&dp
->upcall_rwlock
);
5884 dpif_netdev_enable_upcall(struct dpif
*dpif
)
5885 OVS_NO_THREAD_SAFETY_ANALYSIS
5887 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5888 dp_netdev_enable_upcall(dp
);
5892 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
)
5894 atomic_store_relaxed(&pmd
->wait_for_reload
, false);
5895 atomic_store_relaxed(&pmd
->reload_tx_qid
, false);
5896 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
5897 atomic_store_explicit(&pmd
->reload
, false, memory_order_release
);
5900 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
5901 * the pointer if succeeds, otherwise, NULL (it can return NULL even if
5902 * 'core_id' is NON_PMD_CORE_ID).
5904 * Caller must unrefs the returned reference. */
5905 static struct dp_netdev_pmd_thread
*
5906 dp_netdev_get_pmd(struct dp_netdev
*dp
, unsigned core_id
)
5908 struct dp_netdev_pmd_thread
*pmd
;
5909 const struct cmap_node
*pnode
;
5911 pnode
= cmap_find(&dp
->poll_threads
, hash_int(core_id
, 0));
5915 pmd
= CONTAINER_OF(pnode
, struct dp_netdev_pmd_thread
, node
);
5917 return dp_netdev_pmd_try_ref(pmd
) ? pmd
: NULL
;
5920 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
5922 dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
5923 OVS_REQUIRES(dp
->port_mutex
)
5925 struct dp_netdev_pmd_thread
*non_pmd
;
5927 non_pmd
= xzalloc(sizeof *non_pmd
);
5928 dp_netdev_configure_pmd(non_pmd
, dp
, NON_PMD_CORE_ID
, OVS_NUMA_UNSPEC
);
5931 /* Caller must have valid pointer to 'pmd'. */
5933 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
)
5935 return ovs_refcount_try_ref_rcu(&pmd
->ref_cnt
);
5939 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
)
5941 if (pmd
&& ovs_refcount_unref(&pmd
->ref_cnt
) == 1) {
5942 ovsrcu_postpone(dp_netdev_destroy_pmd
, pmd
);
5946 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
5947 * fails, keeps checking for next node until reaching the end of cmap.
5949 * Caller must unrefs the returned reference. */
5950 static struct dp_netdev_pmd_thread
*
5951 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
)
5953 struct dp_netdev_pmd_thread
*next
;
5956 struct cmap_node
*node
;
5958 node
= cmap_next_position(&dp
->poll_threads
, pos
);
5959 next
= node
? CONTAINER_OF(node
, struct dp_netdev_pmd_thread
, node
)
5961 } while (next
&& !dp_netdev_pmd_try_ref(next
));
5966 /* Configures the 'pmd' based on the input argument. */
5968 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
, struct dp_netdev
*dp
,
5969 unsigned core_id
, int numa_id
)
5972 pmd
->core_id
= core_id
;
5973 pmd
->numa_id
= numa_id
;
5974 pmd
->need_reload
= false;
5975 pmd
->n_output_batches
= 0;
5977 ovs_refcount_init(&pmd
->ref_cnt
);
5978 atomic_init(&pmd
->exit
, false);
5979 pmd
->reload_seq
= seq_create();
5980 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
5981 atomic_init(&pmd
->reload
, false);
5982 ovs_mutex_init(&pmd
->flow_mutex
);
5983 ovs_mutex_init(&pmd
->port_mutex
);
5984 cmap_init(&pmd
->flow_table
);
5985 cmap_init(&pmd
->classifiers
);
5986 pmd
->ctx
.last_rxq
= NULL
;
5987 pmd_thread_ctx_time_update(pmd
);
5988 pmd
->next_optimization
= pmd
->ctx
.now
+ DPCLS_OPTIMIZATION_INTERVAL
;
5989 pmd
->rxq_next_cycle_store
= pmd
->ctx
.now
+ PMD_RXQ_INTERVAL_LEN
;
5990 hmap_init(&pmd
->poll_list
);
5991 hmap_init(&pmd
->tx_ports
);
5992 hmap_init(&pmd
->tnl_port_cache
);
5993 hmap_init(&pmd
->send_port_cache
);
5994 /* init the 'flow_cache' since there is no
5995 * actual thread created for NON_PMD_CORE_ID. */
5996 if (core_id
== NON_PMD_CORE_ID
) {
5997 dfc_cache_init(&pmd
->flow_cache
);
5998 pmd_alloc_static_tx_qid(pmd
);
6000 pmd_perf_stats_init(&pmd
->perf_stats
);
6001 cmap_insert(&dp
->poll_threads
, CONST_CAST(struct cmap_node
*, &pmd
->node
),
6002 hash_int(core_id
, 0));
6006 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
)
6010 dp_netdev_pmd_flow_flush(pmd
);
6011 hmap_destroy(&pmd
->send_port_cache
);
6012 hmap_destroy(&pmd
->tnl_port_cache
);
6013 hmap_destroy(&pmd
->tx_ports
);
6014 hmap_destroy(&pmd
->poll_list
);
6015 /* All flows (including their dpcls_rules) have been deleted already */
6016 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
6018 ovsrcu_postpone(free
, cls
);
6020 cmap_destroy(&pmd
->classifiers
);
6021 cmap_destroy(&pmd
->flow_table
);
6022 ovs_mutex_destroy(&pmd
->flow_mutex
);
6023 seq_destroy(pmd
->reload_seq
);
6024 ovs_mutex_destroy(&pmd
->port_mutex
);
6028 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
6029 * and unrefs the struct. */
6031 dp_netdev_del_pmd(struct dp_netdev
*dp
, struct dp_netdev_pmd_thread
*pmd
)
6033 /* NON_PMD_CORE_ID doesn't have a thread, so we don't have to synchronize,
6034 * but extra cleanup is necessary */
6035 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
6036 ovs_mutex_lock(&dp
->non_pmd_mutex
);
6037 dfc_cache_uninit(&pmd
->flow_cache
);
6038 pmd_free_cached_ports(pmd
);
6039 pmd_free_static_tx_qid(pmd
);
6040 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
6042 atomic_store_relaxed(&pmd
->exit
, true);
6043 dp_netdev_reload_pmd__(pmd
);
6044 xpthread_join(pmd
->thread
, NULL
);
6047 dp_netdev_pmd_clear_ports(pmd
);
6049 /* Purges the 'pmd''s flows after stopping the thread, but before
6050 * destroying the flows, so that the flow stats can be collected. */
6051 if (dp
->dp_purge_cb
) {
6052 dp
->dp_purge_cb(dp
->dp_purge_aux
, pmd
->core_id
);
6054 cmap_remove(&pmd
->dp
->poll_threads
, &pmd
->node
, hash_int(pmd
->core_id
, 0));
6055 dp_netdev_pmd_unref(pmd
);
6058 /* Destroys all pmd threads. If 'non_pmd' is true it also destroys the non pmd
6061 dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
, bool non_pmd
)
6063 struct dp_netdev_pmd_thread
*pmd
;
6064 struct dp_netdev_pmd_thread
**pmd_list
;
6065 size_t k
= 0, n_pmds
;
6067 n_pmds
= cmap_count(&dp
->poll_threads
);
6068 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
6070 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
6071 if (!non_pmd
&& pmd
->core_id
== NON_PMD_CORE_ID
) {
6074 /* We cannot call dp_netdev_del_pmd(), since it alters
6075 * 'dp->poll_threads' (while we're iterating it) and it
6077 ovs_assert(k
< n_pmds
);
6078 pmd_list
[k
++] = pmd
;
6081 for (size_t i
= 0; i
< k
; i
++) {
6082 dp_netdev_del_pmd(dp
, pmd_list
[i
]);
6087 /* Deletes all rx queues from pmd->poll_list and all the ports from
6090 dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
)
6092 struct rxq_poll
*poll
;
6093 struct tx_port
*port
;
6095 ovs_mutex_lock(&pmd
->port_mutex
);
6096 HMAP_FOR_EACH_POP (poll
, node
, &pmd
->poll_list
) {
6099 HMAP_FOR_EACH_POP (port
, node
, &pmd
->tx_ports
) {
6102 ovs_mutex_unlock(&pmd
->port_mutex
);
6105 /* Adds rx queue to poll_list of PMD thread, if it's not there already. */
6107 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
6108 struct dp_netdev_rxq
*rxq
)
6109 OVS_REQUIRES(pmd
->port_mutex
)
6111 int qid
= netdev_rxq_get_queue_id(rxq
->rx
);
6112 uint32_t hash
= hash_2words(odp_to_u32(rxq
->port
->port_no
), qid
);
6113 struct rxq_poll
*poll
;
6115 HMAP_FOR_EACH_WITH_HASH (poll
, node
, hash
, &pmd
->poll_list
) {
6116 if (poll
->rxq
== rxq
) {
6117 /* 'rxq' is already polled by this thread. Do nothing. */
6122 poll
= xmalloc(sizeof *poll
);
6124 hmap_insert(&pmd
->poll_list
, &poll
->node
, hash
);
6126 pmd
->need_reload
= true;
6129 /* Delete 'poll' from poll_list of PMD thread. */
6131 dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
6132 struct rxq_poll
*poll
)
6133 OVS_REQUIRES(pmd
->port_mutex
)
6135 hmap_remove(&pmd
->poll_list
, &poll
->node
);
6138 pmd
->need_reload
= true;
6141 /* Add 'port' to the tx port cache of 'pmd', which must be reloaded for the
6142 * changes to take effect. */
6144 dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
6145 struct dp_netdev_port
*port
)
6146 OVS_REQUIRES(pmd
->port_mutex
)
6150 tx
= tx_port_lookup(&pmd
->tx_ports
, port
->port_no
);
6152 /* 'port' is already on this thread tx cache. Do nothing. */
6156 tx
= xzalloc(sizeof *tx
);
6160 tx
->flush_time
= 0LL;
6161 dp_packet_batch_init(&tx
->output_pkts
);
6163 hmap_insert(&pmd
->tx_ports
, &tx
->node
, hash_port_no(tx
->port
->port_no
));
6164 pmd
->need_reload
= true;
6167 /* Del 'tx' from the tx port cache of 'pmd', which must be reloaded for the
6168 * changes to take effect. */
6170 dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
6172 OVS_REQUIRES(pmd
->port_mutex
)
6174 hmap_remove(&pmd
->tx_ports
, &tx
->node
);
6176 pmd
->need_reload
= true;
6180 dpif_netdev_get_datapath_version(void)
6182 return xstrdup("<built-in>");
6186 dp_netdev_flow_used(struct dp_netdev_flow
*netdev_flow
, int cnt
, int size
,
6187 uint16_t tcp_flags
, long long now
)
6191 atomic_store_relaxed(&netdev_flow
->stats
.used
, now
);
6192 non_atomic_ullong_add(&netdev_flow
->stats
.packet_count
, cnt
);
6193 non_atomic_ullong_add(&netdev_flow
->stats
.byte_count
, size
);
6194 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
6196 atomic_store_relaxed(&netdev_flow
->stats
.tcp_flags
, flags
);
6200 dp_netdev_upcall(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
*packet_
,
6201 struct flow
*flow
, struct flow_wildcards
*wc
, ovs_u128
*ufid
,
6202 enum dpif_upcall_type type
, const struct nlattr
*userdata
,
6203 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
6205 struct dp_netdev
*dp
= pmd
->dp
;
6207 if (OVS_UNLIKELY(!dp
->upcall_cb
)) {
6211 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl
))) {
6212 struct ds ds
= DS_EMPTY_INITIALIZER
;
6215 struct odp_flow_key_parms odp_parms
= {
6217 .mask
= wc
? &wc
->masks
: NULL
,
6218 .support
= dp_netdev_support
,
6221 ofpbuf_init(&key
, 0);
6222 odp_flow_key_from_flow(&odp_parms
, &key
);
6223 packet_str
= ofp_dp_packet_to_string(packet_
);
6225 odp_flow_key_format(key
.data
, key
.size
, &ds
);
6227 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp
->name
,
6228 dpif_upcall_type_to_string(type
), ds_cstr(&ds
), packet_str
);
6230 ofpbuf_uninit(&key
);
6236 return dp
->upcall_cb(packet_
, flow
, ufid
, pmd
->core_id
, type
, userdata
,
6237 actions
, wc
, put_actions
, dp
->upcall_aux
);
6240 static inline uint32_t
6241 dpif_netdev_packet_get_rss_hash_orig_pkt(struct dp_packet
*packet
,
6242 const struct miniflow
*mf
)
6246 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
6247 hash
= dp_packet_get_rss_hash(packet
);
6249 hash
= miniflow_hash_5tuple(mf
, 0);
6250 dp_packet_set_rss_hash(packet
, hash
);
6256 static inline uint32_t
6257 dpif_netdev_packet_get_rss_hash(struct dp_packet
*packet
,
6258 const struct miniflow
*mf
)
6260 uint32_t hash
, recirc_depth
;
6262 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
6263 hash
= dp_packet_get_rss_hash(packet
);
6265 hash
= miniflow_hash_5tuple(mf
, 0);
6266 dp_packet_set_rss_hash(packet
, hash
);
6269 /* The RSS hash must account for the recirculation depth to avoid
6270 * collisions in the exact match cache */
6271 recirc_depth
= *recirc_depth_get_unsafe();
6272 if (OVS_UNLIKELY(recirc_depth
)) {
6273 hash
= hash_finish(hash
, recirc_depth
);
6274 dp_packet_set_rss_hash(packet
, hash
);
6279 struct packet_batch_per_flow
{
6280 unsigned int byte_count
;
6282 struct dp_netdev_flow
*flow
;
6284 struct dp_packet_batch array
;
6288 packet_batch_per_flow_update(struct packet_batch_per_flow
*batch
,
6289 struct dp_packet
*packet
,
6292 batch
->byte_count
+= dp_packet_size(packet
);
6293 batch
->tcp_flags
|= tcp_flags
;
6294 batch
->array
.packets
[batch
->array
.count
++] = packet
;
6298 packet_batch_per_flow_init(struct packet_batch_per_flow
*batch
,
6299 struct dp_netdev_flow
*flow
)
6301 flow
->batch
= batch
;
6304 dp_packet_batch_init(&batch
->array
);
6305 batch
->byte_count
= 0;
6306 batch
->tcp_flags
= 0;
6310 packet_batch_per_flow_execute(struct packet_batch_per_flow
*batch
,
6311 struct dp_netdev_pmd_thread
*pmd
)
6313 struct dp_netdev_actions
*actions
;
6314 struct dp_netdev_flow
*flow
= batch
->flow
;
6316 dp_netdev_flow_used(flow
, batch
->array
.count
, batch
->byte_count
,
6317 batch
->tcp_flags
, pmd
->ctx
.now
/ 1000);
6319 actions
= dp_netdev_flow_get_actions(flow
);
6321 dp_netdev_execute_actions(pmd
, &batch
->array
, true, &flow
->flow
,
6322 actions
->actions
, actions
->size
);
6326 dp_netdev_queue_batches(struct dp_packet
*pkt
,
6327 struct dp_netdev_flow
*flow
, uint16_t tcp_flags
,
6328 struct packet_batch_per_flow
*batches
,
6331 struct packet_batch_per_flow
*batch
= flow
->batch
;
6333 if (OVS_UNLIKELY(!batch
)) {
6334 batch
= &batches
[(*n_batches
)++];
6335 packet_batch_per_flow_init(batch
, flow
);
6338 packet_batch_per_flow_update(batch
, pkt
, tcp_flags
);
6342 packet_enqueue_to_flow_map(struct dp_packet
*packet
,
6343 struct dp_netdev_flow
*flow
,
6345 struct dp_packet_flow_map
*flow_map
,
6348 struct dp_packet_flow_map
*map
= &flow_map
[index
];
6350 map
->packet
= packet
;
6351 map
->tcp_flags
= tcp_flags
;
6354 /* SMC lookup function for a batch of packets.
6355 * By doing batching SMC lookup, we can use prefetch
6356 * to hide memory access latency.
6359 smc_lookup_batch(struct dp_netdev_pmd_thread
*pmd
,
6360 struct netdev_flow_key
*keys
,
6361 struct netdev_flow_key
**missed_keys
,
6362 struct dp_packet_batch
*packets_
,
6364 struct dp_packet_flow_map
*flow_map
,
6368 struct dp_packet
*packet
;
6369 size_t n_smc_hit
= 0, n_missed
= 0;
6370 struct dfc_cache
*cache
= &pmd
->flow_cache
;
6371 struct smc_cache
*smc_cache
= &cache
->smc_cache
;
6372 const struct cmap_node
*flow_node
;
6376 /* Prefetch buckets for all packets */
6377 for (i
= 0; i
< cnt
; i
++) {
6378 OVS_PREFETCH(&smc_cache
->buckets
[keys
[i
].hash
& SMC_MASK
]);
6381 DP_PACKET_BATCH_REFILL_FOR_EACH (i
, cnt
, packet
, packets_
) {
6382 struct dp_netdev_flow
*flow
= NULL
;
6383 flow_node
= smc_entry_get(pmd
, keys
[i
].hash
);
6385 /* Get the original order of this packet in received batch. */
6386 recv_idx
= index_map
[i
];
6388 if (OVS_LIKELY(flow_node
!= NULL
)) {
6389 CMAP_NODE_FOR_EACH (flow
, node
, flow_node
) {
6390 /* Since we dont have per-port megaflow to check the port
6391 * number, we need to verify that the input ports match. */
6392 if (OVS_LIKELY(dpcls_rule_matches_key(&flow
->cr
, &keys
[i
]) &&
6393 flow
->flow
.in_port
.odp_port
== packet
->md
.in_port
.odp_port
)) {
6394 tcp_flags
= miniflow_get_tcp_flags(&keys
[i
].mf
);
6396 /* SMC hit and emc miss, we insert into EMC */
6398 netdev_flow_key_size(miniflow_n_values(&keys
[i
].mf
));
6399 emc_probabilistic_insert(pmd
, &keys
[i
], flow
);
6400 /* Add these packets into the flow map in the same order
6403 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6404 flow_map
, recv_idx
);
6415 /* SMC missed. Group missed packets together at
6416 * the beginning of the 'packets' array. */
6417 dp_packet_batch_refill(packets_
, packet
, i
);
6419 /* Preserve the order of packet for flow batching. */
6420 index_map
[n_missed
] = recv_idx
;
6422 /* Put missed keys to the pointer arrays return to the caller */
6423 missed_keys
[n_missed
++] = &keys
[i
];
6426 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_SMC_HIT
, n_smc_hit
);
6429 /* Try to process all ('cnt') the 'packets' using only the datapath flow cache
6430 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
6431 * miniflow is copied into 'keys' and the packet pointer is moved at the
6432 * beginning of the 'packets' array. The pointers of missed keys are put in the
6433 * missed_keys pointer array for future processing.
6435 * The function returns the number of packets that needs to be processed in the
6436 * 'packets' array (they have been moved to the beginning of the vector).
6438 * For performance reasons a caller may choose not to initialize the metadata
6439 * in 'packets_'. If 'md_is_valid' is false, the metadata in 'packets'
6440 * is not valid and must be initialized by this function using 'port_no'.
6441 * If 'md_is_valid' is true, the metadata is already valid and 'port_no'
6444 static inline size_t
6445 dfc_processing(struct dp_netdev_pmd_thread
*pmd
,
6446 struct dp_packet_batch
*packets_
,
6447 struct netdev_flow_key
*keys
,
6448 struct netdev_flow_key
**missed_keys
,
6449 struct packet_batch_per_flow batches
[], size_t *n_batches
,
6450 struct dp_packet_flow_map
*flow_map
,
6451 size_t *n_flows
, uint8_t *index_map
,
6452 bool md_is_valid
, odp_port_t port_no
)
6454 struct netdev_flow_key
*key
= &keys
[0];
6455 size_t n_missed
= 0, n_emc_hit
= 0;
6456 struct dfc_cache
*cache
= &pmd
->flow_cache
;
6457 struct dp_packet
*packet
;
6458 const size_t cnt
= dp_packet_batch_size(packets_
);
6459 uint32_t cur_min
= pmd
->ctx
.emc_insert_min
;
6464 bool batch_enable
= true;
6466 atomic_read_relaxed(&pmd
->dp
->smc_enable_db
, &smc_enable_db
);
6467 pmd_perf_update_counter(&pmd
->perf_stats
,
6468 md_is_valid
? PMD_STAT_RECIRC
: PMD_STAT_RECV
,
6471 DP_PACKET_BATCH_REFILL_FOR_EACH (i
, cnt
, packet
, packets_
) {
6472 struct dp_netdev_flow
*flow
;
6475 if (OVS_UNLIKELY(dp_packet_size(packet
) < ETH_HEADER_LEN
)) {
6476 dp_packet_delete(packet
);
6481 struct dp_packet
**packets
= packets_
->packets
;
6482 /* Prefetch next packet data and metadata. */
6483 OVS_PREFETCH(dp_packet_data(packets
[i
+1]));
6484 pkt_metadata_prefetch_init(&packets
[i
+1]->md
);
6488 pkt_metadata_init(&packet
->md
, port_no
);
6491 if ((*recirc_depth_get() == 0) &&
6492 dp_packet_has_flow_mark(packet
, &mark
)) {
6493 flow
= mark_to_flow_find(pmd
, mark
);
6494 if (OVS_LIKELY(flow
)) {
6495 tcp_flags
= parse_tcp_flags(packet
);
6496 if (OVS_LIKELY(batch_enable
)) {
6497 dp_netdev_queue_batches(packet
, flow
, tcp_flags
, batches
,
6500 /* Flow batching should be performed only after fast-path
6501 * processing is also completed for packets with emc miss
6502 * or else it will result in reordering of packets with
6503 * same datapath flows. */
6504 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6505 flow_map
, map_cnt
++);
6511 miniflow_extract(packet
, &key
->mf
);
6512 key
->len
= 0; /* Not computed yet. */
6514 (md_is_valid
== false)
6515 ? dpif_netdev_packet_get_rss_hash_orig_pkt(packet
, &key
->mf
)
6516 : dpif_netdev_packet_get_rss_hash(packet
, &key
->mf
);
6518 /* If EMC is disabled skip emc_lookup */
6519 flow
= (cur_min
!= 0) ? emc_lookup(&cache
->emc_cache
, key
) : NULL
;
6520 if (OVS_LIKELY(flow
)) {
6521 tcp_flags
= miniflow_get_tcp_flags(&key
->mf
);
6523 if (OVS_LIKELY(batch_enable
)) {
6524 dp_netdev_queue_batches(packet
, flow
, tcp_flags
, batches
,
6527 /* Flow batching should be performed only after fast-path
6528 * processing is also completed for packets with emc miss
6529 * or else it will result in reordering of packets with
6530 * same datapath flows. */
6531 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6532 flow_map
, map_cnt
++);
6535 /* Exact match cache missed. Group missed packets together at
6536 * the beginning of the 'packets' array. */
6537 dp_packet_batch_refill(packets_
, packet
, i
);
6539 /* Preserve the order of packet for flow batching. */
6540 index_map
[n_missed
] = map_cnt
;
6541 flow_map
[map_cnt
++].flow
= NULL
;
6543 /* 'key[n_missed]' contains the key of the current packet and it
6544 * will be passed to SMC lookup. The next key should be extracted
6545 * to 'keys[n_missed + 1]'.
6546 * We also maintain a pointer array to keys missed both SMC and EMC
6547 * which will be returned to the caller for future processing. */
6548 missed_keys
[n_missed
] = key
;
6549 key
= &keys
[++n_missed
];
6551 /* Skip batching for subsequent packets to avoid reordering. */
6552 batch_enable
= false;
6555 /* Count of packets which are not flow batched. */
6558 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_EXACT_HIT
, n_emc_hit
);
6560 if (!smc_enable_db
) {
6561 return dp_packet_batch_size(packets_
);
6564 /* Packets miss EMC will do a batch lookup in SMC if enabled */
6565 smc_lookup_batch(pmd
, keys
, missed_keys
, packets_
,
6566 n_missed
, flow_map
, index_map
);
6568 return dp_packet_batch_size(packets_
);
6572 handle_packet_upcall(struct dp_netdev_pmd_thread
*pmd
,
6573 struct dp_packet
*packet
,
6574 const struct netdev_flow_key
*key
,
6575 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
6577 struct ofpbuf
*add_actions
;
6578 struct dp_packet_batch b
;
6582 uint64_t cycles
= cycles_counter_update(&pmd
->perf_stats
);
6584 match
.tun_md
.valid
= false;
6585 miniflow_expand(&key
->mf
, &match
.flow
);
6587 ofpbuf_clear(actions
);
6588 ofpbuf_clear(put_actions
);
6590 dpif_flow_hash(pmd
->dp
->dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
6591 error
= dp_netdev_upcall(pmd
, packet
, &match
.flow
, &match
.wc
,
6592 &ufid
, DPIF_UC_MISS
, NULL
, actions
,
6594 if (OVS_UNLIKELY(error
&& error
!= ENOSPC
)) {
6595 dp_packet_delete(packet
);
6599 /* The Netlink encoding of datapath flow keys cannot express
6600 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
6601 * tag is interpreted as exact match on the fact that there is no
6602 * VLAN. Unless we refactor a lot of code that translates between
6603 * Netlink and struct flow representations, we have to do the same
6604 * here. This must be in sync with 'match' in dpif_netdev_flow_put(). */
6605 if (!match
.wc
.masks
.vlans
[0].tci
) {
6606 match
.wc
.masks
.vlans
[0].tci
= htons(0xffff);
6609 /* We can't allow the packet batching in the next loop to execute
6610 * the actions. Otherwise, if there are any slow path actions,
6611 * we'll send the packet up twice. */
6612 dp_packet_batch_init_packet(&b
, packet
);
6613 dp_netdev_execute_actions(pmd
, &b
, true, &match
.flow
,
6614 actions
->data
, actions
->size
);
6616 add_actions
= put_actions
->size
? put_actions
: actions
;
6617 if (OVS_LIKELY(error
!= ENOSPC
)) {
6618 struct dp_netdev_flow
*netdev_flow
;
6620 /* XXX: There's a race window where a flow covering this packet
6621 * could have already been installed since we last did the flow
6622 * lookup before upcall. This could be solved by moving the
6623 * mutex lock outside the loop, but that's an awful long time
6624 * to be locking revalidators out of making flow modifications. */
6625 ovs_mutex_lock(&pmd
->flow_mutex
);
6626 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
6627 if (OVS_LIKELY(!netdev_flow
)) {
6628 netdev_flow
= dp_netdev_flow_add(pmd
, &match
, &ufid
,
6632 ovs_mutex_unlock(&pmd
->flow_mutex
);
6633 uint32_t hash
= dp_netdev_flow_hash(&netdev_flow
->ufid
);
6634 smc_insert(pmd
, key
, hash
);
6635 emc_probabilistic_insert(pmd
, key
, netdev_flow
);
6637 if (pmd_perf_metrics_enabled(pmd
)) {
6638 /* Update upcall stats. */
6639 cycles
= cycles_counter_update(&pmd
->perf_stats
) - cycles
;
6640 struct pmd_perf_stats
*s
= &pmd
->perf_stats
;
6641 s
->current
.upcalls
++;
6642 s
->current
.upcall_cycles
+= cycles
;
6643 histogram_add_sample(&s
->cycles_per_upcall
, cycles
);
6649 fast_path_processing(struct dp_netdev_pmd_thread
*pmd
,
6650 struct dp_packet_batch
*packets_
,
6651 struct netdev_flow_key
**keys
,
6652 struct dp_packet_flow_map
*flow_map
,
6656 const size_t cnt
= dp_packet_batch_size(packets_
);
6657 #if !defined(__CHECKER__) && !defined(_WIN32)
6658 const size_t PKT_ARRAY_SIZE
= cnt
;
6660 /* Sparse or MSVC doesn't like variable length array. */
6661 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
6663 struct dp_packet
*packet
;
6665 struct dpcls_rule
*rules
[PKT_ARRAY_SIZE
];
6666 struct dp_netdev
*dp
= pmd
->dp
;
6667 int upcall_ok_cnt
= 0, upcall_fail_cnt
= 0;
6668 int lookup_cnt
= 0, add_lookup_cnt
;
6671 for (size_t i
= 0; i
< cnt
; i
++) {
6672 /* Key length is needed in all the cases, hash computed on demand. */
6673 keys
[i
]->len
= netdev_flow_key_size(miniflow_n_values(&keys
[i
]->mf
));
6675 /* Get the classifier for the in_port */
6676 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
6677 if (OVS_LIKELY(cls
)) {
6678 any_miss
= !dpcls_lookup(cls
, (const struct netdev_flow_key
**)keys
,
6679 rules
, cnt
, &lookup_cnt
);
6682 memset(rules
, 0, sizeof(rules
));
6684 if (OVS_UNLIKELY(any_miss
) && !fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
6685 uint64_t actions_stub
[512 / 8], slow_stub
[512 / 8];
6686 struct ofpbuf actions
, put_actions
;
6688 ofpbuf_use_stub(&actions
, actions_stub
, sizeof actions_stub
);
6689 ofpbuf_use_stub(&put_actions
, slow_stub
, sizeof slow_stub
);
6691 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
6692 struct dp_netdev_flow
*netdev_flow
;
6694 if (OVS_LIKELY(rules
[i
])) {
6698 /* It's possible that an earlier slow path execution installed
6699 * a rule covering this flow. In this case, it's a lot cheaper
6700 * to catch it here than execute a miss. */
6701 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, keys
[i
],
6704 lookup_cnt
+= add_lookup_cnt
;
6705 rules
[i
] = &netdev_flow
->cr
;
6709 int error
= handle_packet_upcall(pmd
, packet
, keys
[i
],
6710 &actions
, &put_actions
);
6712 if (OVS_UNLIKELY(error
)) {
6719 ofpbuf_uninit(&actions
);
6720 ofpbuf_uninit(&put_actions
);
6721 fat_rwlock_unlock(&dp
->upcall_rwlock
);
6722 } else if (OVS_UNLIKELY(any_miss
)) {
6723 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
6724 if (OVS_UNLIKELY(!rules
[i
])) {
6725 dp_packet_delete(packet
);
6731 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
6732 struct dp_netdev_flow
*flow
;
6733 /* Get the original order of this packet in received batch. */
6734 int recv_idx
= index_map
[i
];
6737 if (OVS_UNLIKELY(!rules
[i
])) {
6741 flow
= dp_netdev_flow_cast(rules
[i
]);
6742 uint32_t hash
= dp_netdev_flow_hash(&flow
->ufid
);
6743 smc_insert(pmd
, keys
[i
], hash
);
6745 emc_probabilistic_insert(pmd
, keys
[i
], flow
);
6746 /* Add these packets into the flow map in the same order
6749 tcp_flags
= miniflow_get_tcp_flags(&keys
[i
]->mf
);
6750 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6751 flow_map
, recv_idx
);
6754 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_MASKED_HIT
,
6755 cnt
- upcall_ok_cnt
- upcall_fail_cnt
);
6756 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_MASKED_LOOKUP
,
6758 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_MISS
,
6760 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_LOST
,
6764 /* Packets enter the datapath from a port (or from recirculation) here.
6766 * When 'md_is_valid' is true the metadata in 'packets' are already valid.
6767 * When false the metadata in 'packets' need to be initialized. */
6769 dp_netdev_input__(struct dp_netdev_pmd_thread
*pmd
,
6770 struct dp_packet_batch
*packets
,
6771 bool md_is_valid
, odp_port_t port_no
)
6773 #if !defined(__CHECKER__) && !defined(_WIN32)
6774 const size_t PKT_ARRAY_SIZE
= dp_packet_batch_size(packets
);
6776 /* Sparse or MSVC doesn't like variable length array. */
6777 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
6779 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
)
6780 struct netdev_flow_key keys
[PKT_ARRAY_SIZE
];
6781 struct netdev_flow_key
*missed_keys
[PKT_ARRAY_SIZE
];
6782 struct packet_batch_per_flow batches
[PKT_ARRAY_SIZE
];
6784 struct dp_packet_flow_map flow_map
[PKT_ARRAY_SIZE
];
6785 uint8_t index_map
[PKT_ARRAY_SIZE
];
6791 dfc_processing(pmd
, packets
, keys
, missed_keys
, batches
, &n_batches
,
6792 flow_map
, &n_flows
, index_map
, md_is_valid
, port_no
);
6794 if (!dp_packet_batch_is_empty(packets
)) {
6795 /* Get ingress port from first packet's metadata. */
6796 in_port
= packets
->packets
[0]->md
.in_port
.odp_port
;
6797 fast_path_processing(pmd
, packets
, missed_keys
,
6798 flow_map
, index_map
, in_port
);
6801 /* Batch rest of packets which are in flow map. */
6802 for (i
= 0; i
< n_flows
; i
++) {
6803 struct dp_packet_flow_map
*map
= &flow_map
[i
];
6805 if (OVS_UNLIKELY(!map
->flow
)) {
6808 dp_netdev_queue_batches(map
->packet
, map
->flow
, map
->tcp_flags
,
6809 batches
, &n_batches
);
6812 /* All the flow batches need to be reset before any call to
6813 * packet_batch_per_flow_execute() as it could potentially trigger
6814 * recirculation. When a packet matching flow ‘j’ happens to be
6815 * recirculated, the nested call to dp_netdev_input__() could potentially
6816 * classify the packet as matching another flow - say 'k'. It could happen
6817 * that in the previous call to dp_netdev_input__() that same flow 'k' had
6818 * already its own batches[k] still waiting to be served. So if its
6819 * ‘batch’ member is not reset, the recirculated packet would be wrongly
6820 * appended to batches[k] of the 1st call to dp_netdev_input__(). */
6821 for (i
= 0; i
< n_batches
; i
++) {
6822 batches
[i
].flow
->batch
= NULL
;
6825 for (i
= 0; i
< n_batches
; i
++) {
6826 packet_batch_per_flow_execute(&batches
[i
], pmd
);
6831 dp_netdev_input(struct dp_netdev_pmd_thread
*pmd
,
6832 struct dp_packet_batch
*packets
,
6835 dp_netdev_input__(pmd
, packets
, false, port_no
);
6839 dp_netdev_recirculate(struct dp_netdev_pmd_thread
*pmd
,
6840 struct dp_packet_batch
*packets
)
6842 dp_netdev_input__(pmd
, packets
, true, 0);
6845 struct dp_netdev_execute_aux
{
6846 struct dp_netdev_pmd_thread
*pmd
;
6847 const struct flow
*flow
;
6851 dpif_netdev_register_dp_purge_cb(struct dpif
*dpif
, dp_purge_callback
*cb
,
6854 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
6855 dp
->dp_purge_aux
= aux
;
6856 dp
->dp_purge_cb
= cb
;
6860 dpif_netdev_register_upcall_cb(struct dpif
*dpif
, upcall_callback
*cb
,
6863 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
6864 dp
->upcall_aux
= aux
;
6869 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
6873 struct dp_netdev_port
*port
;
6876 HMAP_FOR_EACH (tx
, node
, &pmd
->send_port_cache
) {
6877 if (!tx
->port
->dynamic_txqs
) {
6880 interval
= pmd
->ctx
.now
- tx
->last_used
;
6881 if (tx
->qid
>= 0 && (purge
|| interval
>= XPS_TIMEOUT
)) {
6883 ovs_mutex_lock(&port
->txq_used_mutex
);
6884 port
->txq_used
[tx
->qid
]--;
6885 ovs_mutex_unlock(&port
->txq_used_mutex
);
6892 dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
6895 struct dp_netdev_port
*port
;
6897 int i
, min_cnt
, min_qid
;
6899 interval
= pmd
->ctx
.now
- tx
->last_used
;
6900 tx
->last_used
= pmd
->ctx
.now
;
6902 if (OVS_LIKELY(tx
->qid
>= 0 && interval
< XPS_TIMEOUT
)) {
6908 ovs_mutex_lock(&port
->txq_used_mutex
);
6910 port
->txq_used
[tx
->qid
]--;
6916 for (i
= 0; i
< netdev_n_txq(port
->netdev
); i
++) {
6917 if (port
->txq_used
[i
] < min_cnt
|| min_cnt
== -1) {
6918 min_cnt
= port
->txq_used
[i
];
6923 port
->txq_used
[min_qid
]++;
6926 ovs_mutex_unlock(&port
->txq_used_mutex
);
6928 dpif_netdev_xps_revalidate_pmd(pmd
, false);
6930 VLOG_DBG("Core %d: New TX queue ID %d for port \'%s\'.",
6931 pmd
->core_id
, tx
->qid
, netdev_get_name(tx
->port
->netdev
));
6935 static struct tx_port
*
6936 pmd_tnl_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
6939 return tx_port_lookup(&pmd
->tnl_port_cache
, port_no
);
6942 static struct tx_port
*
6943 pmd_send_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
6946 return tx_port_lookup(&pmd
->send_port_cache
, port_no
);
6950 push_tnl_action(const struct dp_netdev_pmd_thread
*pmd
,
6951 const struct nlattr
*attr
,
6952 struct dp_packet_batch
*batch
)
6954 struct tx_port
*tun_port
;
6955 const struct ovs_action_push_tnl
*data
;
6958 data
= nl_attr_get(attr
);
6960 tun_port
= pmd_tnl_port_cache_lookup(pmd
, data
->tnl_port
);
6965 err
= netdev_push_header(tun_port
->port
->netdev
, batch
, data
);
6970 dp_packet_delete_batch(batch
, true);
6975 dp_execute_userspace_action(struct dp_netdev_pmd_thread
*pmd
,
6976 struct dp_packet
*packet
, bool should_steal
,
6977 struct flow
*flow
, ovs_u128
*ufid
,
6978 struct ofpbuf
*actions
,
6979 const struct nlattr
*userdata
)
6981 struct dp_packet_batch b
;
6984 ofpbuf_clear(actions
);
6986 error
= dp_netdev_upcall(pmd
, packet
, flow
, NULL
, ufid
,
6987 DPIF_UC_ACTION
, userdata
, actions
,
6989 if (!error
|| error
== ENOSPC
) {
6990 dp_packet_batch_init_packet(&b
, packet
);
6991 dp_netdev_execute_actions(pmd
, &b
, should_steal
, flow
,
6992 actions
->data
, actions
->size
);
6993 } else if (should_steal
) {
6994 dp_packet_delete(packet
);
6999 dp_execute_cb(void *aux_
, struct dp_packet_batch
*packets_
,
7000 const struct nlattr
*a
, bool should_steal
)
7001 OVS_NO_THREAD_SAFETY_ANALYSIS
7003 struct dp_netdev_execute_aux
*aux
= aux_
;
7004 uint32_t *depth
= recirc_depth_get();
7005 struct dp_netdev_pmd_thread
*pmd
= aux
->pmd
;
7006 struct dp_netdev
*dp
= pmd
->dp
;
7007 int type
= nl_attr_type(a
);
7010 switch ((enum ovs_action_attr
)type
) {
7011 case OVS_ACTION_ATTR_OUTPUT
:
7012 p
= pmd_send_port_cache_lookup(pmd
, nl_attr_get_odp_port(a
));
7013 if (OVS_LIKELY(p
)) {
7014 struct dp_packet
*packet
;
7015 struct dp_packet_batch out
;
7017 if (!should_steal
) {
7018 dp_packet_batch_clone(&out
, packets_
);
7019 dp_packet_batch_reset_cutlen(packets_
);
7022 dp_packet_batch_apply_cutlen(packets_
);
7025 if (OVS_UNLIKELY(!dp_packet_batch_is_empty(&p
->output_pkts
)
7026 && packets_
->packets
[0]->source
7027 != p
->output_pkts
.packets
[0]->source
)) {
7028 /* XXX: netdev-dpdk assumes that all packets in a single
7029 * output batch has the same source. Flush here to
7030 * avoid memory access issues. */
7031 dp_netdev_pmd_flush_output_on_port(pmd
, p
);
7034 if (dp_packet_batch_size(&p
->output_pkts
)
7035 + dp_packet_batch_size(packets_
) > NETDEV_MAX_BURST
) {
7036 /* Flush here to avoid overflow. */
7037 dp_netdev_pmd_flush_output_on_port(pmd
, p
);
7040 if (dp_packet_batch_is_empty(&p
->output_pkts
)) {
7041 pmd
->n_output_batches
++;
7044 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7045 p
->output_pkts_rxqs
[dp_packet_batch_size(&p
->output_pkts
)] =
7047 dp_packet_batch_add(&p
->output_pkts
, packet
);
7053 case OVS_ACTION_ATTR_TUNNEL_PUSH
:
7055 /* We're requested to push tunnel header, but also we need to take
7056 * the ownership of these packets. Thus, we can avoid performing
7057 * the action, because the caller will not use the result anyway.
7058 * Just break to free the batch. */
7061 dp_packet_batch_apply_cutlen(packets_
);
7062 push_tnl_action(pmd
, a
, packets_
);
7065 case OVS_ACTION_ATTR_TUNNEL_POP
:
7066 if (*depth
< MAX_RECIRC_DEPTH
) {
7067 struct dp_packet_batch
*orig_packets_
= packets_
;
7068 odp_port_t portno
= nl_attr_get_odp_port(a
);
7070 p
= pmd_tnl_port_cache_lookup(pmd
, portno
);
7072 struct dp_packet_batch tnl_pkt
;
7074 if (!should_steal
) {
7075 dp_packet_batch_clone(&tnl_pkt
, packets_
);
7076 packets_
= &tnl_pkt
;
7077 dp_packet_batch_reset_cutlen(orig_packets_
);
7080 dp_packet_batch_apply_cutlen(packets_
);
7082 netdev_pop_header(p
->port
->netdev
, packets_
);
7083 if (dp_packet_batch_is_empty(packets_
)) {
7087 struct dp_packet
*packet
;
7088 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7089 packet
->md
.in_port
.odp_port
= portno
;
7093 dp_netdev_recirculate(pmd
, packets_
);
7100 case OVS_ACTION_ATTR_USERSPACE
:
7101 if (!fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
7102 struct dp_packet_batch
*orig_packets_
= packets_
;
7103 const struct nlattr
*userdata
;
7104 struct dp_packet_batch usr_pkt
;
7105 struct ofpbuf actions
;
7110 userdata
= nl_attr_find_nested(a
, OVS_USERSPACE_ATTR_USERDATA
);
7111 ofpbuf_init(&actions
, 0);
7113 if (packets_
->trunc
) {
7114 if (!should_steal
) {
7115 dp_packet_batch_clone(&usr_pkt
, packets_
);
7116 packets_
= &usr_pkt
;
7118 dp_packet_batch_reset_cutlen(orig_packets_
);
7121 dp_packet_batch_apply_cutlen(packets_
);
7124 struct dp_packet
*packet
;
7125 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7126 flow_extract(packet
, &flow
);
7127 dpif_flow_hash(dp
->dpif
, &flow
, sizeof flow
, &ufid
);
7128 dp_execute_userspace_action(pmd
, packet
, should_steal
, &flow
,
7129 &ufid
, &actions
, userdata
);
7133 dp_packet_delete_batch(packets_
, true);
7136 ofpbuf_uninit(&actions
);
7137 fat_rwlock_unlock(&dp
->upcall_rwlock
);
7143 case OVS_ACTION_ATTR_RECIRC
:
7144 if (*depth
< MAX_RECIRC_DEPTH
) {
7145 struct dp_packet_batch recirc_pkts
;
7147 if (!should_steal
) {
7148 dp_packet_batch_clone(&recirc_pkts
, packets_
);
7149 packets_
= &recirc_pkts
;
7152 struct dp_packet
*packet
;
7153 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7154 packet
->md
.recirc_id
= nl_attr_get_u32(a
);
7158 dp_netdev_recirculate(pmd
, packets_
);
7164 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
7167 case OVS_ACTION_ATTR_CT
: {
7168 const struct nlattr
*b
;
7170 bool commit
= false;
7173 const char *helper
= NULL
;
7174 const uint32_t *setmark
= NULL
;
7175 const struct ovs_key_ct_labels
*setlabel
= NULL
;
7176 struct nat_action_info_t nat_action_info
;
7177 struct nat_action_info_t
*nat_action_info_ref
= NULL
;
7178 bool nat_config
= false;
7180 NL_ATTR_FOR_EACH_UNSAFE (b
, left
, nl_attr_get(a
),
7181 nl_attr_get_size(a
)) {
7182 enum ovs_ct_attr sub_type
= nl_attr_type(b
);
7185 case OVS_CT_ATTR_FORCE_COMMIT
:
7188 case OVS_CT_ATTR_COMMIT
:
7191 case OVS_CT_ATTR_ZONE
:
7192 zone
= nl_attr_get_u16(b
);
7194 case OVS_CT_ATTR_HELPER
:
7195 helper
= nl_attr_get_string(b
);
7197 case OVS_CT_ATTR_MARK
:
7198 setmark
= nl_attr_get(b
);
7200 case OVS_CT_ATTR_LABELS
:
7201 setlabel
= nl_attr_get(b
);
7203 case OVS_CT_ATTR_EVENTMASK
:
7204 /* Silently ignored, as userspace datapath does not generate
7205 * netlink events. */
7207 case OVS_CT_ATTR_NAT
: {
7208 const struct nlattr
*b_nest
;
7209 unsigned int left_nest
;
7210 bool ip_min_specified
= false;
7211 bool proto_num_min_specified
= false;
7212 bool ip_max_specified
= false;
7213 bool proto_num_max_specified
= false;
7214 memset(&nat_action_info
, 0, sizeof nat_action_info
);
7215 nat_action_info_ref
= &nat_action_info
;
7217 NL_NESTED_FOR_EACH_UNSAFE (b_nest
, left_nest
, b
) {
7218 enum ovs_nat_attr sub_type_nest
= nl_attr_type(b_nest
);
7220 switch (sub_type_nest
) {
7221 case OVS_NAT_ATTR_SRC
:
7222 case OVS_NAT_ATTR_DST
:
7224 nat_action_info
.nat_action
|=
7225 ((sub_type_nest
== OVS_NAT_ATTR_SRC
)
7226 ? NAT_ACTION_SRC
: NAT_ACTION_DST
);
7228 case OVS_NAT_ATTR_IP_MIN
:
7229 memcpy(&nat_action_info
.min_addr
,
7230 nl_attr_get(b_nest
),
7231 nl_attr_get_size(b_nest
));
7232 ip_min_specified
= true;
7234 case OVS_NAT_ATTR_IP_MAX
:
7235 memcpy(&nat_action_info
.max_addr
,
7236 nl_attr_get(b_nest
),
7237 nl_attr_get_size(b_nest
));
7238 ip_max_specified
= true;
7240 case OVS_NAT_ATTR_PROTO_MIN
:
7241 nat_action_info
.min_port
=
7242 nl_attr_get_u16(b_nest
);
7243 proto_num_min_specified
= true;
7245 case OVS_NAT_ATTR_PROTO_MAX
:
7246 nat_action_info
.max_port
=
7247 nl_attr_get_u16(b_nest
);
7248 proto_num_max_specified
= true;
7250 case OVS_NAT_ATTR_PERSISTENT
:
7251 case OVS_NAT_ATTR_PROTO_HASH
:
7252 case OVS_NAT_ATTR_PROTO_RANDOM
:
7254 case OVS_NAT_ATTR_UNSPEC
:
7255 case __OVS_NAT_ATTR_MAX
:
7260 if (ip_min_specified
&& !ip_max_specified
) {
7261 nat_action_info
.max_addr
= nat_action_info
.min_addr
;
7263 if (proto_num_min_specified
&& !proto_num_max_specified
) {
7264 nat_action_info
.max_port
= nat_action_info
.min_port
;
7266 if (proto_num_min_specified
|| proto_num_max_specified
) {
7267 if (nat_action_info
.nat_action
& NAT_ACTION_SRC
) {
7268 nat_action_info
.nat_action
|= NAT_ACTION_SRC_PORT
;
7269 } else if (nat_action_info
.nat_action
& NAT_ACTION_DST
) {
7270 nat_action_info
.nat_action
|= NAT_ACTION_DST_PORT
;
7275 case OVS_CT_ATTR_UNSPEC
:
7276 case __OVS_CT_ATTR_MAX
:
7281 /* We won't be able to function properly in this case, hence
7282 * complain loudly. */
7283 if (nat_config
&& !commit
) {
7284 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(5, 5);
7285 VLOG_WARN_RL(&rl
, "NAT specified without commit.");
7288 conntrack_execute(dp
->conntrack
, packets_
, aux
->flow
->dl_type
, force
,
7289 commit
, zone
, setmark
, setlabel
, aux
->flow
->tp_src
,
7290 aux
->flow
->tp_dst
, helper
, nat_action_info_ref
,
7291 pmd
->ctx
.now
/ 1000);
7295 case OVS_ACTION_ATTR_METER
:
7296 dp_netdev_run_meter(pmd
->dp
, packets_
, nl_attr_get_u32(a
),
7300 case OVS_ACTION_ATTR_PUSH_VLAN
:
7301 case OVS_ACTION_ATTR_POP_VLAN
:
7302 case OVS_ACTION_ATTR_PUSH_MPLS
:
7303 case OVS_ACTION_ATTR_POP_MPLS
:
7304 case OVS_ACTION_ATTR_SET
:
7305 case OVS_ACTION_ATTR_SET_MASKED
:
7306 case OVS_ACTION_ATTR_SAMPLE
:
7307 case OVS_ACTION_ATTR_HASH
:
7308 case OVS_ACTION_ATTR_UNSPEC
:
7309 case OVS_ACTION_ATTR_TRUNC
:
7310 case OVS_ACTION_ATTR_PUSH_ETH
:
7311 case OVS_ACTION_ATTR_POP_ETH
:
7312 case OVS_ACTION_ATTR_CLONE
:
7313 case OVS_ACTION_ATTR_PUSH_NSH
:
7314 case OVS_ACTION_ATTR_POP_NSH
:
7315 case OVS_ACTION_ATTR_CT_CLEAR
:
7316 case OVS_ACTION_ATTR_CHECK_PKT_LEN
:
7317 case __OVS_ACTION_ATTR_MAX
:
7321 dp_packet_delete_batch(packets_
, should_steal
);
7325 dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
7326 struct dp_packet_batch
*packets
,
7327 bool should_steal
, const struct flow
*flow
,
7328 const struct nlattr
*actions
, size_t actions_len
)
7330 struct dp_netdev_execute_aux aux
= { pmd
, flow
};
7332 odp_execute_actions(&aux
, packets
, should_steal
, actions
,
7333 actions_len
, dp_execute_cb
);
7336 struct dp_netdev_ct_dump
{
7337 struct ct_dpif_dump_state up
;
7338 struct conntrack_dump dump
;
7339 struct conntrack
*ct
;
7340 struct dp_netdev
*dp
;
7344 dpif_netdev_ct_dump_start(struct dpif
*dpif
, struct ct_dpif_dump_state
**dump_
,
7345 const uint16_t *pzone
, int *ptot_bkts
)
7347 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7348 struct dp_netdev_ct_dump
*dump
;
7350 dump
= xzalloc(sizeof *dump
);
7352 dump
->ct
= dp
->conntrack
;
7354 conntrack_dump_start(dp
->conntrack
, &dump
->dump
, pzone
, ptot_bkts
);
7362 dpif_netdev_ct_dump_next(struct dpif
*dpif OVS_UNUSED
,
7363 struct ct_dpif_dump_state
*dump_
,
7364 struct ct_dpif_entry
*entry
)
7366 struct dp_netdev_ct_dump
*dump
;
7368 INIT_CONTAINER(dump
, dump_
, up
);
7370 return conntrack_dump_next(&dump
->dump
, entry
);
7374 dpif_netdev_ct_dump_done(struct dpif
*dpif OVS_UNUSED
,
7375 struct ct_dpif_dump_state
*dump_
)
7377 struct dp_netdev_ct_dump
*dump
;
7380 INIT_CONTAINER(dump
, dump_
, up
);
7382 err
= conntrack_dump_done(&dump
->dump
);
7390 dpif_netdev_ct_flush(struct dpif
*dpif
, const uint16_t *zone
,
7391 const struct ct_dpif_tuple
*tuple
)
7393 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7396 return conntrack_flush_tuple(dp
->conntrack
, tuple
, zone
? *zone
: 0);
7398 return conntrack_flush(dp
->conntrack
, zone
);
7402 dpif_netdev_ct_set_maxconns(struct dpif
*dpif
, uint32_t maxconns
)
7404 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7406 return conntrack_set_maxconns(dp
->conntrack
, maxconns
);
7410 dpif_netdev_ct_get_maxconns(struct dpif
*dpif
, uint32_t *maxconns
)
7412 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7414 return conntrack_get_maxconns(dp
->conntrack
, maxconns
);
7418 dpif_netdev_ct_get_nconns(struct dpif
*dpif
, uint32_t *nconns
)
7420 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7422 return conntrack_get_nconns(dp
->conntrack
, nconns
);
7426 dpif_netdev_ipf_set_enabled(struct dpif
*dpif
, bool v6
, bool enable
)
7428 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7429 return ipf_set_enabled(conntrack_ipf_ctx(dp
->conntrack
), v6
, enable
);
7433 dpif_netdev_ipf_set_min_frag(struct dpif
*dpif
, bool v6
, uint32_t min_frag
)
7435 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7436 return ipf_set_min_frag(conntrack_ipf_ctx(dp
->conntrack
), v6
, min_frag
);
7440 dpif_netdev_ipf_set_max_nfrags(struct dpif
*dpif
, uint32_t max_frags
)
7442 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7443 return ipf_set_max_nfrags(conntrack_ipf_ctx(dp
->conntrack
), max_frags
);
7446 /* Adjust this function if 'dpif_ipf_status' and 'ipf_status' were to
7449 dpif_netdev_ipf_get_status(struct dpif
*dpif
,
7450 struct dpif_ipf_status
*dpif_ipf_status
)
7452 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7453 ipf_get_status(conntrack_ipf_ctx(dp
->conntrack
),
7454 (struct ipf_status
*) dpif_ipf_status
);
7459 dpif_netdev_ipf_dump_start(struct dpif
*dpif OVS_UNUSED
,
7460 struct ipf_dump_ctx
**ipf_dump_ctx
)
7462 return ipf_dump_start(ipf_dump_ctx
);
7466 dpif_netdev_ipf_dump_next(struct dpif
*dpif
, void *ipf_dump_ctx
, char **dump
)
7468 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7469 return ipf_dump_next(conntrack_ipf_ctx(dp
->conntrack
), ipf_dump_ctx
,
7474 dpif_netdev_ipf_dump_done(struct dpif
*dpif OVS_UNUSED
, void *ipf_dump_ctx
)
7476 return ipf_dump_done(ipf_dump_ctx
);
7480 const struct dpif_class dpif_netdev_class
= {
7482 true, /* cleanup_required */
7484 dpif_netdev_enumerate
,
7485 dpif_netdev_port_open_type
,
7488 dpif_netdev_destroy
,
7491 dpif_netdev_get_stats
,
7492 dpif_netdev_port_add
,
7493 dpif_netdev_port_del
,
7494 dpif_netdev_port_set_config
,
7495 dpif_netdev_port_query_by_number
,
7496 dpif_netdev_port_query_by_name
,
7497 NULL
, /* port_get_pid */
7498 dpif_netdev_port_dump_start
,
7499 dpif_netdev_port_dump_next
,
7500 dpif_netdev_port_dump_done
,
7501 dpif_netdev_port_poll
,
7502 dpif_netdev_port_poll_wait
,
7503 dpif_netdev_flow_flush
,
7504 dpif_netdev_flow_dump_create
,
7505 dpif_netdev_flow_dump_destroy
,
7506 dpif_netdev_flow_dump_thread_create
,
7507 dpif_netdev_flow_dump_thread_destroy
,
7508 dpif_netdev_flow_dump_next
,
7509 dpif_netdev_operate
,
7510 NULL
, /* recv_set */
7511 NULL
, /* handlers_set */
7512 dpif_netdev_set_config
,
7513 dpif_netdev_queue_to_priority
,
7515 NULL
, /* recv_wait */
7516 NULL
, /* recv_purge */
7517 dpif_netdev_register_dp_purge_cb
,
7518 dpif_netdev_register_upcall_cb
,
7519 dpif_netdev_enable_upcall
,
7520 dpif_netdev_disable_upcall
,
7521 dpif_netdev_get_datapath_version
,
7522 dpif_netdev_ct_dump_start
,
7523 dpif_netdev_ct_dump_next
,
7524 dpif_netdev_ct_dump_done
,
7525 dpif_netdev_ct_flush
,
7526 dpif_netdev_ct_set_maxconns
,
7527 dpif_netdev_ct_get_maxconns
,
7528 dpif_netdev_ct_get_nconns
,
7529 NULL
, /* ct_set_limits */
7530 NULL
, /* ct_get_limits */
7531 NULL
, /* ct_del_limits */
7532 dpif_netdev_ipf_set_enabled
,
7533 dpif_netdev_ipf_set_min_frag
,
7534 dpif_netdev_ipf_set_max_nfrags
,
7535 dpif_netdev_ipf_get_status
,
7536 dpif_netdev_ipf_dump_start
,
7537 dpif_netdev_ipf_dump_next
,
7538 dpif_netdev_ipf_dump_done
,
7539 dpif_netdev_meter_get_features
,
7540 dpif_netdev_meter_set
,
7541 dpif_netdev_meter_get
,
7542 dpif_netdev_meter_del
,
7546 dpif_dummy_change_port_number(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
7547 const char *argv
[], void *aux OVS_UNUSED
)
7549 struct dp_netdev_port
*port
;
7550 struct dp_netdev
*dp
;
7553 ovs_mutex_lock(&dp_netdev_mutex
);
7554 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
7555 if (!dp
|| !dpif_netdev_class_is_dummy(dp
->class)) {
7556 ovs_mutex_unlock(&dp_netdev_mutex
);
7557 unixctl_command_reply_error(conn
, "unknown datapath or not a dummy");
7560 ovs_refcount_ref(&dp
->ref_cnt
);
7561 ovs_mutex_unlock(&dp_netdev_mutex
);
7563 ovs_mutex_lock(&dp
->port_mutex
);
7564 if (get_port_by_name(dp
, argv
[2], &port
)) {
7565 unixctl_command_reply_error(conn
, "unknown port");
7569 port_no
= u32_to_odp(atoi(argv
[3]));
7570 if (!port_no
|| port_no
== ODPP_NONE
) {
7571 unixctl_command_reply_error(conn
, "bad port number");
7574 if (dp_netdev_lookup_port(dp
, port_no
)) {
7575 unixctl_command_reply_error(conn
, "port number already in use");
7580 hmap_remove(&dp
->ports
, &port
->node
);
7581 reconfigure_datapath(dp
);
7583 /* Reinsert with new port number. */
7584 port
->port_no
= port_no
;
7585 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
7586 reconfigure_datapath(dp
);
7588 seq_change(dp
->port_seq
);
7589 unixctl_command_reply(conn
, NULL
);
7592 ovs_mutex_unlock(&dp
->port_mutex
);
7593 dp_netdev_unref(dp
);
7597 dpif_dummy_register__(const char *type
)
7599 struct dpif_class
*class;
7601 class = xmalloc(sizeof *class);
7602 *class = dpif_netdev_class
;
7603 class->type
= xstrdup(type
);
7604 dp_register_provider(class);
7608 dpif_dummy_override(const char *type
)
7613 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
7614 * a userland-only build. It's useful for testsuite.
7616 error
= dp_unregister_provider(type
);
7617 if (error
== 0 || error
== EAFNOSUPPORT
) {
7618 dpif_dummy_register__(type
);
7623 dpif_dummy_register(enum dummy_level level
)
7625 if (level
== DUMMY_OVERRIDE_ALL
) {
7630 dp_enumerate_types(&types
);
7631 SSET_FOR_EACH (type
, &types
) {
7632 dpif_dummy_override(type
);
7634 sset_destroy(&types
);
7635 } else if (level
== DUMMY_OVERRIDE_SYSTEM
) {
7636 dpif_dummy_override("system");
7639 dpif_dummy_register__("dummy");
7641 unixctl_command_register("dpif-dummy/change-port-number",
7642 "dp port new-number",
7643 3, 3, dpif_dummy_change_port_number
, NULL
);
7646 /* Datapath Classifier. */
7649 dpcls_subtable_destroy_cb(struct dpcls_subtable
*subtable
)
7651 cmap_destroy(&subtable
->rules
);
7652 ovsrcu_postpone(free
, subtable
->mf_masks
);
7653 ovsrcu_postpone(free
, subtable
);
7656 /* Initializes 'cls' as a classifier that initially contains no classification
7659 dpcls_init(struct dpcls
*cls
)
7661 cmap_init(&cls
->subtables_map
);
7662 pvector_init(&cls
->subtables
);
7666 dpcls_destroy_subtable(struct dpcls
*cls
, struct dpcls_subtable
*subtable
)
7668 VLOG_DBG("Destroying subtable %p for in_port %d", subtable
, cls
->in_port
);
7669 pvector_remove(&cls
->subtables
, subtable
);
7670 cmap_remove(&cls
->subtables_map
, &subtable
->cmap_node
,
7671 subtable
->mask
.hash
);
7672 ovsrcu_postpone(dpcls_subtable_destroy_cb
, subtable
);
7675 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
7676 * caller's responsibility.
7677 * May only be called after all the readers have been terminated. */
7679 dpcls_destroy(struct dpcls
*cls
)
7682 struct dpcls_subtable
*subtable
;
7684 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
7685 ovs_assert(cmap_count(&subtable
->rules
) == 0);
7686 dpcls_destroy_subtable(cls
, subtable
);
7688 cmap_destroy(&cls
->subtables_map
);
7689 pvector_destroy(&cls
->subtables
);
7693 static struct dpcls_subtable
*
7694 dpcls_create_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
7696 struct dpcls_subtable
*subtable
;
7698 /* Need to add one. */
7699 subtable
= xmalloc(sizeof *subtable
7700 - sizeof subtable
->mask
.mf
+ mask
->len
);
7701 cmap_init(&subtable
->rules
);
7702 subtable
->hit_cnt
= 0;
7703 netdev_flow_key_clone(&subtable
->mask
, mask
);
7705 /* The count of bits in the mask defines the space required for masks.
7706 * Then call gen_masks() to create the appropriate masks, avoiding the cost
7707 * of doing runtime calculations. */
7708 uint32_t unit0
= count_1bits(mask
->mf
.map
.bits
[0]);
7709 uint32_t unit1
= count_1bits(mask
->mf
.map
.bits
[1]);
7710 subtable
->mf_bits_set_unit0
= unit0
;
7711 subtable
->mf_bits_set_unit1
= unit1
;
7712 subtable
->mf_masks
= xmalloc(sizeof(uint64_t) * (unit0
+ unit1
));
7713 netdev_flow_key_gen_masks(mask
, subtable
->mf_masks
, unit0
, unit1
);
7715 /* Probe for a specialized generic lookup function. */
7716 subtable
->lookup_func
= dpcls_subtable_generic_probe(unit0
, unit1
);
7718 /* If not set, assign generic lookup. Generic works for any miniflow. */
7719 if (!subtable
->lookup_func
) {
7720 subtable
->lookup_func
= dpcls_subtable_lookup_generic
;
7723 cmap_insert(&cls
->subtables_map
, &subtable
->cmap_node
, mask
->hash
);
7724 /* Add the new subtable at the end of the pvector (with no hits yet) */
7725 pvector_insert(&cls
->subtables
, subtable
, 0);
7726 VLOG_DBG("Creating %"PRIuSIZE
". subtable %p for in_port %d",
7727 cmap_count(&cls
->subtables_map
), subtable
, cls
->in_port
);
7728 pvector_publish(&cls
->subtables
);
7733 static inline struct dpcls_subtable
*
7734 dpcls_find_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
7736 struct dpcls_subtable
*subtable
;
7738 CMAP_FOR_EACH_WITH_HASH (subtable
, cmap_node
, mask
->hash
,
7739 &cls
->subtables_map
) {
7740 if (netdev_flow_key_equal(&subtable
->mask
, mask
)) {
7744 return dpcls_create_subtable(cls
, mask
);
7748 /* Periodically sort the dpcls subtable vectors according to hit counts */
7750 dpcls_sort_subtable_vector(struct dpcls
*cls
)
7752 struct pvector
*pvec
= &cls
->subtables
;
7753 struct dpcls_subtable
*subtable
;
7755 PVECTOR_FOR_EACH (subtable
, pvec
) {
7756 pvector_change_priority(pvec
, subtable
, subtable
->hit_cnt
);
7757 subtable
->hit_cnt
= 0;
7759 pvector_publish(pvec
);
7763 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
,
7764 struct polled_queue
*poll_list
, int poll_cnt
)
7767 uint64_t tot_idle
= 0, tot_proc
= 0;
7768 unsigned int pmd_load
= 0;
7770 if (pmd
->ctx
.now
> pmd
->rxq_next_cycle_store
) {
7772 struct pmd_auto_lb
*pmd_alb
= &pmd
->dp
->pmd_alb
;
7773 if (pmd_alb
->is_enabled
&& !pmd
->isolated
7774 && (pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_IDLE
] >=
7775 pmd
->prev_stats
[PMD_CYCLES_ITER_IDLE
])
7776 && (pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_BUSY
] >=
7777 pmd
->prev_stats
[PMD_CYCLES_ITER_BUSY
]))
7779 tot_idle
= pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_IDLE
] -
7780 pmd
->prev_stats
[PMD_CYCLES_ITER_IDLE
];
7781 tot_proc
= pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_BUSY
] -
7782 pmd
->prev_stats
[PMD_CYCLES_ITER_BUSY
];
7785 pmd_load
= ((tot_proc
* 100) / (tot_idle
+ tot_proc
));
7788 if (pmd_load
>= ALB_PMD_LOAD_THRESHOLD
) {
7789 atomic_count_inc(&pmd
->pmd_overloaded
);
7791 atomic_count_set(&pmd
->pmd_overloaded
, 0);
7795 pmd
->prev_stats
[PMD_CYCLES_ITER_IDLE
] =
7796 pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_IDLE
];
7797 pmd
->prev_stats
[PMD_CYCLES_ITER_BUSY
] =
7798 pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_BUSY
];
7800 /* Get the cycles that were used to process each queue and store. */
7801 for (unsigned i
= 0; i
< poll_cnt
; i
++) {
7802 uint64_t rxq_cyc_curr
= dp_netdev_rxq_get_cycles(poll_list
[i
].rxq
,
7803 RXQ_CYCLES_PROC_CURR
);
7804 dp_netdev_rxq_set_intrvl_cycles(poll_list
[i
].rxq
, rxq_cyc_curr
);
7805 dp_netdev_rxq_set_cycles(poll_list
[i
].rxq
, RXQ_CYCLES_PROC_CURR
,
7808 curr_tsc
= cycles_counter_update(&pmd
->perf_stats
);
7809 if (pmd
->intrvl_tsc_prev
) {
7810 /* There is a prev timestamp, store a new intrvl cycle count. */
7811 atomic_store_relaxed(&pmd
->intrvl_cycles
,
7812 curr_tsc
- pmd
->intrvl_tsc_prev
);
7814 pmd
->intrvl_tsc_prev
= curr_tsc
;
7815 /* Start new measuring interval */
7816 pmd
->rxq_next_cycle_store
= pmd
->ctx
.now
+ PMD_RXQ_INTERVAL_LEN
;
7819 if (pmd
->ctx
.now
> pmd
->next_optimization
) {
7820 /* Try to obtain the flow lock to block out revalidator threads.
7821 * If not possible, just try next time. */
7822 if (!ovs_mutex_trylock(&pmd
->flow_mutex
)) {
7823 /* Optimize each classifier */
7824 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
7825 dpcls_sort_subtable_vector(cls
);
7827 ovs_mutex_unlock(&pmd
->flow_mutex
);
7828 /* Start new measuring interval */
7829 pmd
->next_optimization
= pmd
->ctx
.now
7830 + DPCLS_OPTIMIZATION_INTERVAL
;
7835 /* Insert 'rule' into 'cls'. */
7837 dpcls_insert(struct dpcls
*cls
, struct dpcls_rule
*rule
,
7838 const struct netdev_flow_key
*mask
)
7840 struct dpcls_subtable
*subtable
= dpcls_find_subtable(cls
, mask
);
7842 /* Refer to subtable's mask, also for later removal. */
7843 rule
->mask
= &subtable
->mask
;
7844 cmap_insert(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
);
7847 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
7849 dpcls_remove(struct dpcls
*cls
, struct dpcls_rule
*rule
)
7851 struct dpcls_subtable
*subtable
;
7853 ovs_assert(rule
->mask
);
7855 /* Get subtable from reference in rule->mask. */
7856 INIT_CONTAINER(subtable
, rule
->mask
, mask
);
7857 if (cmap_remove(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
)
7859 /* Delete empty subtable. */
7860 dpcls_destroy_subtable(cls
, subtable
);
7861 pvector_publish(&cls
->subtables
);
7865 /* Inner loop for mask generation of a unit, see netdev_flow_key_gen_masks. */
7867 netdev_flow_key_gen_mask_unit(uint64_t iter
,
7868 const uint64_t count
,
7872 for (i
= 0; i
< count
; i
++) {
7873 uint64_t lowest_bit
= (iter
& -iter
);
7874 iter
&= ~lowest_bit
;
7875 mf_masks
[i
] = (lowest_bit
- 1);
7877 /* Checks that count has covered all bits in the iter bitmap. */
7878 ovs_assert(iter
== 0);
7881 /* Generate a mask for each block in the miniflow, based on the bits set. This
7882 * allows easily masking packets with the generated array here, without
7883 * calculations. This replaces runtime-calculating the masks.
7884 * @param key The table to generate the mf_masks for
7885 * @param mf_masks Pointer to a u64 array of at least *mf_bits* in size
7886 * @param mf_bits_total Number of bits set in the whole miniflow (both units)
7887 * @param mf_bits_unit0 Number of bits set in unit0 of the miniflow
7890 netdev_flow_key_gen_masks(const struct netdev_flow_key
*tbl
,
7892 const uint32_t mf_bits_u0
,
7893 const uint32_t mf_bits_u1
)
7895 uint64_t iter_u0
= tbl
->mf
.map
.bits
[0];
7896 uint64_t iter_u1
= tbl
->mf
.map
.bits
[1];
7898 netdev_flow_key_gen_mask_unit(iter_u0
, mf_bits_u0
, &mf_masks
[0]);
7899 netdev_flow_key_gen_mask_unit(iter_u1
, mf_bits_u1
, &mf_masks
[mf_bits_u0
]);
7902 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
7903 * in 'mask' the values in 'key' and 'target' are the same. */
7905 dpcls_rule_matches_key(const struct dpcls_rule
*rule
,
7906 const struct netdev_flow_key
*target
)
7908 const uint64_t *keyp
= miniflow_get_values(&rule
->flow
.mf
);
7909 const uint64_t *maskp
= miniflow_get_values(&rule
->mask
->mf
);
7912 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, target
, rule
->flow
.mf
.map
) {
7913 if (OVS_UNLIKELY((value
& *maskp
++) != *keyp
++)) {
7920 /* For each miniflow in 'keys' performs a classifier lookup writing the result
7921 * into the corresponding slot in 'rules'. If a particular entry in 'keys' is
7922 * NULL it is skipped.
7924 * This function is optimized for use in the userspace datapath and therefore
7925 * does not implement a lot of features available in the standard
7926 * classifier_lookup() function. Specifically, it does not implement
7927 * priorities, instead returning any rule which matches the flow.
7929 * Returns true if all miniflows found a corresponding rule. */
7931 dpcls_lookup(struct dpcls
*cls
, const struct netdev_flow_key
*keys
[],
7932 struct dpcls_rule
**rules
, const size_t cnt
,
7935 /* The received 'cnt' miniflows are the search-keys that will be processed
7936 * to find a matching entry into the available subtables.
7937 * The number of bits in map_type is equal to NETDEV_MAX_BURST. */
7938 #define MAP_BITS (sizeof(uint32_t) * CHAR_BIT)
7939 BUILD_ASSERT_DECL(MAP_BITS
>= NETDEV_MAX_BURST
);
7941 struct dpcls_subtable
*subtable
;
7942 uint32_t keys_map
= TYPE_MAXIMUM(uint32_t); /* Set all bits. */
7944 if (cnt
!= MAP_BITS
) {
7945 keys_map
>>= MAP_BITS
- cnt
; /* Clear extra bits. */
7947 memset(rules
, 0, cnt
* sizeof *rules
);
7949 int lookups_match
= 0, subtable_pos
= 1;
7952 /* The Datapath classifier - aka dpcls - is composed of subtables.
7953 * Subtables are dynamically created as needed when new rules are inserted.
7954 * Each subtable collects rules with matches on a specific subset of packet
7955 * fields as defined by the subtable's mask. We proceed to process every
7956 * search-key against each subtable, but when a match is found for a
7957 * search-key, the search for that key can stop because the rules are
7958 * non-overlapping. */
7959 PVECTOR_FOR_EACH (subtable
, &cls
->subtables
) {
7960 /* Call the subtable specific lookup function. */
7961 found_map
= subtable
->lookup_func(subtable
, keys_map
, keys
, rules
);
7963 /* Count the number of subtables searched for this packet match. This
7964 * estimates the "spread" of subtables looked at per matched packet. */
7965 uint32_t pkts_matched
= count_1bits(found_map
);
7966 lookups_match
+= pkts_matched
* subtable_pos
;
7968 /* Clear the found rules, and return early if all packets are found. */
7969 keys_map
&= ~found_map
;
7971 if (num_lookups_p
) {
7972 *num_lookups_p
= lookups_match
;
7979 if (num_lookups_p
) {
7980 *num_lookups_p
= lookups_match
;