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 static struct ovsthread_once tsc_freq_check
= OVSTHREAD_ONCE_INITIALIZER
;
1521 struct dp_netdev
*dp
;
1524 /* Avoid estimating TSC frequency for dummy datapath to not slow down
1526 if (!dpif_netdev_class_is_dummy(class)
1527 && ovsthread_once_start(&tsc_freq_check
)) {
1528 pmd_perf_estimate_tsc_frequency();
1529 ovsthread_once_done(&tsc_freq_check
);
1532 dp
= xzalloc(sizeof *dp
);
1533 shash_add(&dp_netdevs
, name
, dp
);
1535 *CONST_CAST(const struct dpif_class
**, &dp
->class) = class;
1536 *CONST_CAST(const char **, &dp
->name
) = xstrdup(name
);
1537 ovs_refcount_init(&dp
->ref_cnt
);
1538 atomic_flag_clear(&dp
->destroyed
);
1540 ovs_mutex_init(&dp
->port_mutex
);
1541 hmap_init(&dp
->ports
);
1542 dp
->port_seq
= seq_create();
1543 fat_rwlock_init(&dp
->upcall_rwlock
);
1545 dp
->reconfigure_seq
= seq_create();
1546 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
1548 for (int i
= 0; i
< N_METER_LOCKS
; ++i
) {
1549 ovs_mutex_init_adaptive(&dp
->meter_locks
[i
]);
1552 /* Disable upcalls by default. */
1553 dp_netdev_disable_upcall(dp
);
1554 dp
->upcall_aux
= NULL
;
1555 dp
->upcall_cb
= NULL
;
1557 dp
->conntrack
= conntrack_init();
1559 atomic_init(&dp
->emc_insert_min
, DEFAULT_EM_FLOW_INSERT_MIN
);
1560 atomic_init(&dp
->tx_flush_interval
, DEFAULT_TX_FLUSH_INTERVAL
);
1562 cmap_init(&dp
->poll_threads
);
1563 dp
->pmd_rxq_assign_cyc
= true;
1565 ovs_mutex_init(&dp
->tx_qid_pool_mutex
);
1566 /* We need 1 Tx queue for each possible core + 1 for non-PMD threads. */
1567 dp
->tx_qid_pool
= id_pool_create(0, ovs_numa_get_n_cores() + 1);
1569 ovs_mutex_init_recursive(&dp
->non_pmd_mutex
);
1570 ovsthread_key_create(&dp
->per_pmd_key
, NULL
);
1572 ovs_mutex_lock(&dp
->port_mutex
);
1573 /* non-PMD will be created before all other threads and will
1574 * allocate static_tx_qid = 0. */
1575 dp_netdev_set_nonpmd(dp
);
1577 error
= do_add_port(dp
, name
, dpif_netdev_port_open_type(dp
->class,
1580 ovs_mutex_unlock(&dp
->port_mutex
);
1586 dp
->last_tnl_conf_seq
= seq_read(tnl_conf_seq
);
1592 dp_netdev_request_reconfigure(struct dp_netdev
*dp
)
1594 seq_change(dp
->reconfigure_seq
);
1598 dp_netdev_is_reconf_required(struct dp_netdev
*dp
)
1600 return seq_read(dp
->reconfigure_seq
) != dp
->last_reconfigure_seq
;
1604 dpif_netdev_open(const struct dpif_class
*class, const char *name
,
1605 bool create
, struct dpif
**dpifp
)
1607 struct dp_netdev
*dp
;
1610 ovs_mutex_lock(&dp_netdev_mutex
);
1611 dp
= shash_find_data(&dp_netdevs
, name
);
1613 error
= create
? create_dp_netdev(name
, class, &dp
) : ENODEV
;
1615 error
= (dp
->class != class ? EINVAL
1620 *dpifp
= create_dpif_netdev(dp
);
1623 ovs_mutex_unlock(&dp_netdev_mutex
);
1629 dp_netdev_destroy_upcall_lock(struct dp_netdev
*dp
)
1630 OVS_NO_THREAD_SAFETY_ANALYSIS
1632 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
1633 ovs_assert(fat_rwlock_tryrdlock(&dp
->upcall_rwlock
));
1635 /* Before freeing a lock we should release it */
1636 fat_rwlock_unlock(&dp
->upcall_rwlock
);
1637 fat_rwlock_destroy(&dp
->upcall_rwlock
);
1641 dp_delete_meter(struct dp_netdev
*dp
, uint32_t meter_id
)
1642 OVS_REQUIRES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
1644 if (dp
->meters
[meter_id
]) {
1645 free(dp
->meters
[meter_id
]);
1646 dp
->meters
[meter_id
] = NULL
;
1650 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
1651 * through the 'dp_netdevs' shash while freeing 'dp'. */
1653 dp_netdev_free(struct dp_netdev
*dp
)
1654 OVS_REQUIRES(dp_netdev_mutex
)
1656 struct dp_netdev_port
*port
, *next
;
1658 shash_find_and_delete(&dp_netdevs
, dp
->name
);
1660 ovs_mutex_lock(&dp
->port_mutex
);
1661 HMAP_FOR_EACH_SAFE (port
, next
, node
, &dp
->ports
) {
1662 do_del_port(dp
, port
);
1664 ovs_mutex_unlock(&dp
->port_mutex
);
1666 dp_netdev_destroy_all_pmds(dp
, true);
1667 cmap_destroy(&dp
->poll_threads
);
1669 ovs_mutex_destroy(&dp
->tx_qid_pool_mutex
);
1670 id_pool_destroy(dp
->tx_qid_pool
);
1672 ovs_mutex_destroy(&dp
->non_pmd_mutex
);
1673 ovsthread_key_delete(dp
->per_pmd_key
);
1675 conntrack_destroy(dp
->conntrack
);
1678 seq_destroy(dp
->reconfigure_seq
);
1680 seq_destroy(dp
->port_seq
);
1681 hmap_destroy(&dp
->ports
);
1682 ovs_mutex_destroy(&dp
->port_mutex
);
1684 /* Upcalls must be disabled at this point */
1685 dp_netdev_destroy_upcall_lock(dp
);
1689 for (i
= 0; i
< MAX_METERS
; ++i
) {
1691 dp_delete_meter(dp
, i
);
1692 meter_unlock(dp
, i
);
1694 for (i
= 0; i
< N_METER_LOCKS
; ++i
) {
1695 ovs_mutex_destroy(&dp
->meter_locks
[i
]);
1698 free(dp
->pmd_cmask
);
1699 free(CONST_CAST(char *, dp
->name
));
1704 dp_netdev_unref(struct dp_netdev
*dp
)
1707 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
1708 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
1709 ovs_mutex_lock(&dp_netdev_mutex
);
1710 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1713 ovs_mutex_unlock(&dp_netdev_mutex
);
1718 dpif_netdev_close(struct dpif
*dpif
)
1720 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1722 dp_netdev_unref(dp
);
1727 dpif_netdev_destroy(struct dpif
*dpif
)
1729 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1731 if (!atomic_flag_test_and_set(&dp
->destroyed
)) {
1732 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1733 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
1741 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
1742 * load/store semantics. While the increment is not atomic, the load and
1743 * store operations are, making it impossible to read inconsistent values.
1745 * This is used to update thread local stats counters. */
1747 non_atomic_ullong_add(atomic_ullong
*var
, unsigned long long n
)
1749 unsigned long long tmp
;
1751 atomic_read_relaxed(var
, &tmp
);
1753 atomic_store_relaxed(var
, tmp
);
1757 dpif_netdev_get_stats(const struct dpif
*dpif
, struct dpif_dp_stats
*stats
)
1759 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1760 struct dp_netdev_pmd_thread
*pmd
;
1761 uint64_t pmd_stats
[PMD_N_STATS
];
1763 stats
->n_flows
= stats
->n_hit
= stats
->n_missed
= stats
->n_lost
= 0;
1764 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1765 stats
->n_flows
+= cmap_count(&pmd
->flow_table
);
1766 pmd_perf_read_counters(&pmd
->perf_stats
, pmd_stats
);
1767 stats
->n_hit
+= pmd_stats
[PMD_STAT_EXACT_HIT
];
1768 stats
->n_hit
+= pmd_stats
[PMD_STAT_SMC_HIT
];
1769 stats
->n_hit
+= pmd_stats
[PMD_STAT_MASKED_HIT
];
1770 stats
->n_missed
+= pmd_stats
[PMD_STAT_MISS
];
1771 stats
->n_lost
+= pmd_stats
[PMD_STAT_LOST
];
1773 stats
->n_masks
= UINT32_MAX
;
1774 stats
->n_mask_hit
= UINT64_MAX
;
1780 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread
*pmd
)
1782 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
1783 ovs_mutex_lock(&pmd
->dp
->non_pmd_mutex
);
1784 ovs_mutex_lock(&pmd
->port_mutex
);
1785 pmd_load_cached_ports(pmd
);
1786 ovs_mutex_unlock(&pmd
->port_mutex
);
1787 ovs_mutex_unlock(&pmd
->dp
->non_pmd_mutex
);
1791 seq_change(pmd
->reload_seq
);
1792 atomic_store_explicit(&pmd
->reload
, true, memory_order_release
);
1796 hash_port_no(odp_port_t port_no
)
1798 return hash_int(odp_to_u32(port_no
), 0);
1802 port_create(const char *devname
, const char *type
,
1803 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1805 struct dp_netdev_port
*port
;
1806 enum netdev_flags flags
;
1807 struct netdev
*netdev
;
1812 /* Open and validate network device. */
1813 error
= netdev_open(devname
, type
, &netdev
);
1817 /* XXX reject non-Ethernet devices */
1819 netdev_get_flags(netdev
, &flags
);
1820 if (flags
& NETDEV_LOOPBACK
) {
1821 VLOG_ERR("%s: cannot add a loopback device", devname
);
1826 port
= xzalloc(sizeof *port
);
1827 port
->port_no
= port_no
;
1828 port
->netdev
= netdev
;
1829 port
->type
= xstrdup(type
);
1831 port
->emc_enabled
= true;
1832 port
->need_reconfigure
= true;
1833 ovs_mutex_init(&port
->txq_used_mutex
);
1840 netdev_close(netdev
);
1845 do_add_port(struct dp_netdev
*dp
, const char *devname
, const char *type
,
1847 OVS_REQUIRES(dp
->port_mutex
)
1849 struct netdev_saved_flags
*sf
;
1850 struct dp_netdev_port
*port
;
1853 /* Reject devices already in 'dp'. */
1854 if (!get_port_by_name(dp
, devname
, &port
)) {
1858 error
= port_create(devname
, type
, port_no
, &port
);
1863 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
1864 seq_change(dp
->port_seq
);
1866 reconfigure_datapath(dp
);
1868 /* Check that port was successfully configured. */
1869 if (!dp_netdev_lookup_port(dp
, port_no
)) {
1873 /* Updating device flags triggers an if_notifier, which triggers a bridge
1874 * reconfiguration and another attempt to add this port, leading to an
1875 * infinite loop if the device is configured incorrectly and cannot be
1876 * added. Setting the promisc mode after a successful reconfiguration,
1877 * since we already know that the device is somehow properly configured. */
1878 error
= netdev_turn_flags_on(port
->netdev
, NETDEV_PROMISC
, &sf
);
1880 VLOG_ERR("%s: cannot set promisc flag", devname
);
1881 do_del_port(dp
, port
);
1890 dpif_netdev_port_add(struct dpif
*dpif
, struct netdev
*netdev
,
1891 odp_port_t
*port_nop
)
1893 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1894 char namebuf
[NETDEV_VPORT_NAME_BUFSIZE
];
1895 const char *dpif_port
;
1899 ovs_mutex_lock(&dp
->port_mutex
);
1900 dpif_port
= netdev_vport_get_dpif_port(netdev
, namebuf
, sizeof namebuf
);
1901 if (*port_nop
!= ODPP_NONE
) {
1902 port_no
= *port_nop
;
1903 error
= dp_netdev_lookup_port(dp
, *port_nop
) ? EBUSY
: 0;
1905 port_no
= choose_port(dp
, dpif_port
);
1906 error
= port_no
== ODPP_NONE
? EFBIG
: 0;
1909 *port_nop
= port_no
;
1910 error
= do_add_port(dp
, dpif_port
, netdev_get_type(netdev
), port_no
);
1912 ovs_mutex_unlock(&dp
->port_mutex
);
1918 dpif_netdev_port_del(struct dpif
*dpif
, odp_port_t port_no
)
1920 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1923 ovs_mutex_lock(&dp
->port_mutex
);
1924 if (port_no
== ODPP_LOCAL
) {
1927 struct dp_netdev_port
*port
;
1929 error
= get_port_by_number(dp
, port_no
, &port
);
1931 do_del_port(dp
, port
);
1934 ovs_mutex_unlock(&dp
->port_mutex
);
1940 is_valid_port_number(odp_port_t port_no
)
1942 return port_no
!= ODPP_NONE
;
1945 static struct dp_netdev_port
*
1946 dp_netdev_lookup_port(const struct dp_netdev
*dp
, odp_port_t port_no
)
1947 OVS_REQUIRES(dp
->port_mutex
)
1949 struct dp_netdev_port
*port
;
1951 HMAP_FOR_EACH_WITH_HASH (port
, node
, hash_port_no(port_no
), &dp
->ports
) {
1952 if (port
->port_no
== port_no
) {
1960 get_port_by_number(struct dp_netdev
*dp
,
1961 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1962 OVS_REQUIRES(dp
->port_mutex
)
1964 if (!is_valid_port_number(port_no
)) {
1968 *portp
= dp_netdev_lookup_port(dp
, port_no
);
1969 return *portp
? 0 : ENODEV
;
1974 port_destroy(struct dp_netdev_port
*port
)
1980 netdev_close(port
->netdev
);
1981 netdev_restore_flags(port
->sf
);
1983 for (unsigned i
= 0; i
< port
->n_rxq
; i
++) {
1984 netdev_rxq_close(port
->rxqs
[i
].rx
);
1986 ovs_mutex_destroy(&port
->txq_used_mutex
);
1987 free(port
->rxq_affinity_list
);
1988 free(port
->txq_used
);
1995 get_port_by_name(struct dp_netdev
*dp
,
1996 const char *devname
, struct dp_netdev_port
**portp
)
1997 OVS_REQUIRES(dp
->port_mutex
)
1999 struct dp_netdev_port
*port
;
2001 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2002 if (!strcmp(netdev_get_name(port
->netdev
), devname
)) {
2008 /* Callers of dpif_netdev_port_query_by_name() expect ENODEV for a non
2013 /* Returns 'true' if there is a port with pmd netdev. */
2015 has_pmd_port(struct dp_netdev
*dp
)
2016 OVS_REQUIRES(dp
->port_mutex
)
2018 struct dp_netdev_port
*port
;
2020 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2021 if (netdev_is_pmd(port
->netdev
)) {
2030 do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*port
)
2031 OVS_REQUIRES(dp
->port_mutex
)
2033 hmap_remove(&dp
->ports
, &port
->node
);
2034 seq_change(dp
->port_seq
);
2036 reconfigure_datapath(dp
);
2042 answer_port_query(const struct dp_netdev_port
*port
,
2043 struct dpif_port
*dpif_port
)
2045 dpif_port
->name
= xstrdup(netdev_get_name(port
->netdev
));
2046 dpif_port
->type
= xstrdup(port
->type
);
2047 dpif_port
->port_no
= port
->port_no
;
2051 dpif_netdev_port_query_by_number(const struct dpif
*dpif
, odp_port_t port_no
,
2052 struct dpif_port
*dpif_port
)
2054 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2055 struct dp_netdev_port
*port
;
2058 ovs_mutex_lock(&dp
->port_mutex
);
2059 error
= get_port_by_number(dp
, port_no
, &port
);
2060 if (!error
&& dpif_port
) {
2061 answer_port_query(port
, dpif_port
);
2063 ovs_mutex_unlock(&dp
->port_mutex
);
2069 dpif_netdev_port_query_by_name(const struct dpif
*dpif
, const char *devname
,
2070 struct dpif_port
*dpif_port
)
2072 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2073 struct dp_netdev_port
*port
;
2076 ovs_mutex_lock(&dp
->port_mutex
);
2077 error
= get_port_by_name(dp
, devname
, &port
);
2078 if (!error
&& dpif_port
) {
2079 answer_port_query(port
, dpif_port
);
2081 ovs_mutex_unlock(&dp
->port_mutex
);
2087 dp_netdev_flow_free(struct dp_netdev_flow
*flow
)
2089 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow
));
2093 static void dp_netdev_flow_unref(struct dp_netdev_flow
*flow
)
2095 if (ovs_refcount_unref_relaxed(&flow
->ref_cnt
) == 1) {
2096 ovsrcu_postpone(dp_netdev_flow_free
, flow
);
2101 dp_netdev_flow_hash(const ovs_u128
*ufid
)
2103 return ufid
->u32
[0];
2106 static inline struct dpcls
*
2107 dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread
*pmd
,
2111 uint32_t hash
= hash_port_no(in_port
);
2112 CMAP_FOR_EACH_WITH_HASH (cls
, node
, hash
, &pmd
->classifiers
) {
2113 if (cls
->in_port
== in_port
) {
2114 /* Port classifier exists already */
2121 static inline struct dpcls
*
2122 dp_netdev_pmd_find_dpcls(struct dp_netdev_pmd_thread
*pmd
,
2124 OVS_REQUIRES(pmd
->flow_mutex
)
2126 struct dpcls
*cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2127 uint32_t hash
= hash_port_no(in_port
);
2130 /* Create new classifier for in_port */
2131 cls
= xmalloc(sizeof(*cls
));
2133 cls
->in_port
= in_port
;
2134 cmap_insert(&pmd
->classifiers
, &cls
->node
, hash
);
2135 VLOG_DBG("Creating dpcls %p for in_port %d", cls
, in_port
);
2140 #define MAX_FLOW_MARK (UINT32_MAX - 1)
2141 #define INVALID_FLOW_MARK (UINT32_MAX)
2143 struct megaflow_to_mark_data
{
2144 const struct cmap_node node
;
2150 struct cmap megaflow_to_mark
;
2151 struct cmap mark_to_flow
;
2152 struct id_pool
*pool
;
2155 static struct flow_mark flow_mark
= {
2156 .megaflow_to_mark
= CMAP_INITIALIZER
,
2157 .mark_to_flow
= CMAP_INITIALIZER
,
2161 flow_mark_alloc(void)
2165 if (!flow_mark
.pool
) {
2166 /* Haven't initiated yet, do it here */
2167 flow_mark
.pool
= id_pool_create(0, MAX_FLOW_MARK
);
2170 if (id_pool_alloc_id(flow_mark
.pool
, &mark
)) {
2174 return INVALID_FLOW_MARK
;
2178 flow_mark_free(uint32_t mark
)
2180 id_pool_free_id(flow_mark
.pool
, mark
);
2183 /* associate megaflow with a mark, which is a 1:1 mapping */
2185 megaflow_to_mark_associate(const ovs_u128
*mega_ufid
, uint32_t mark
)
2187 size_t hash
= dp_netdev_flow_hash(mega_ufid
);
2188 struct megaflow_to_mark_data
*data
= xzalloc(sizeof(*data
));
2190 data
->mega_ufid
= *mega_ufid
;
2193 cmap_insert(&flow_mark
.megaflow_to_mark
,
2194 CONST_CAST(struct cmap_node
*, &data
->node
), hash
);
2197 /* disassociate meagaflow with a mark */
2199 megaflow_to_mark_disassociate(const ovs_u128
*mega_ufid
)
2201 size_t hash
= dp_netdev_flow_hash(mega_ufid
);
2202 struct megaflow_to_mark_data
*data
;
2204 CMAP_FOR_EACH_WITH_HASH (data
, node
, hash
, &flow_mark
.megaflow_to_mark
) {
2205 if (ovs_u128_equals(*mega_ufid
, data
->mega_ufid
)) {
2206 cmap_remove(&flow_mark
.megaflow_to_mark
,
2207 CONST_CAST(struct cmap_node
*, &data
->node
), hash
);
2208 ovsrcu_postpone(free
, data
);
2213 VLOG_WARN("Masked ufid "UUID_FMT
" is not associated with a mark?\n",
2214 UUID_ARGS((struct uuid
*)mega_ufid
));
2217 static inline uint32_t
2218 megaflow_to_mark_find(const ovs_u128
*mega_ufid
)
2220 size_t hash
= dp_netdev_flow_hash(mega_ufid
);
2221 struct megaflow_to_mark_data
*data
;
2223 CMAP_FOR_EACH_WITH_HASH (data
, node
, hash
, &flow_mark
.megaflow_to_mark
) {
2224 if (ovs_u128_equals(*mega_ufid
, data
->mega_ufid
)) {
2229 VLOG_DBG("Mark id for ufid "UUID_FMT
" was not found\n",
2230 UUID_ARGS((struct uuid
*)mega_ufid
));
2231 return INVALID_FLOW_MARK
;
2234 /* associate mark with a flow, which is 1:N mapping */
2236 mark_to_flow_associate(const uint32_t mark
, struct dp_netdev_flow
*flow
)
2238 dp_netdev_flow_ref(flow
);
2240 cmap_insert(&flow_mark
.mark_to_flow
,
2241 CONST_CAST(struct cmap_node
*, &flow
->mark_node
),
2245 VLOG_DBG("Associated dp_netdev flow %p with mark %u\n", flow
, mark
);
2249 flow_mark_has_no_ref(uint32_t mark
)
2251 struct dp_netdev_flow
*flow
;
2253 CMAP_FOR_EACH_WITH_HASH (flow
, mark_node
, hash_int(mark
, 0),
2254 &flow_mark
.mark_to_flow
) {
2255 if (flow
->mark
== mark
) {
2264 mark_to_flow_disassociate(struct dp_netdev_pmd_thread
*pmd
,
2265 struct dp_netdev_flow
*flow
)
2268 uint32_t mark
= flow
->mark
;
2269 struct cmap_node
*mark_node
= CONST_CAST(struct cmap_node
*,
2272 cmap_remove(&flow_mark
.mark_to_flow
, mark_node
, hash_int(mark
, 0));
2273 flow
->mark
= INVALID_FLOW_MARK
;
2276 * no flow is referencing the mark any more? If so, let's
2277 * remove the flow from hardware and free the mark.
2279 if (flow_mark_has_no_ref(mark
)) {
2280 struct netdev
*port
;
2281 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
2283 port
= netdev_ports_get(in_port
, pmd
->dp
->class);
2285 /* Taking a global 'port_mutex' to fulfill thread safety
2286 * restrictions for the netdev-offload-dpdk module. */
2287 ovs_mutex_lock(&pmd
->dp
->port_mutex
);
2288 ret
= netdev_flow_del(port
, &flow
->mega_ufid
, NULL
);
2289 ovs_mutex_unlock(&pmd
->dp
->port_mutex
);
2293 flow_mark_free(mark
);
2294 VLOG_DBG("Freed flow mark %u\n", mark
);
2296 megaflow_to_mark_disassociate(&flow
->mega_ufid
);
2298 dp_netdev_flow_unref(flow
);
2304 flow_mark_flush(struct dp_netdev_pmd_thread
*pmd
)
2306 struct dp_netdev_flow
*flow
;
2308 CMAP_FOR_EACH (flow
, mark_node
, &flow_mark
.mark_to_flow
) {
2309 if (flow
->pmd_id
== pmd
->core_id
) {
2310 queue_netdev_flow_del(pmd
, flow
);
2315 static struct dp_netdev_flow
*
2316 mark_to_flow_find(const struct dp_netdev_pmd_thread
*pmd
,
2317 const uint32_t mark
)
2319 struct dp_netdev_flow
*flow
;
2321 CMAP_FOR_EACH_WITH_HASH (flow
, mark_node
, hash_int(mark
, 0),
2322 &flow_mark
.mark_to_flow
) {
2323 if (flow
->mark
== mark
&& flow
->pmd_id
== pmd
->core_id
&&
2324 flow
->dead
== false) {
2332 static struct dp_flow_offload_item
*
2333 dp_netdev_alloc_flow_offload(struct dp_netdev_pmd_thread
*pmd
,
2334 struct dp_netdev_flow
*flow
,
2337 struct dp_flow_offload_item
*offload
;
2339 offload
= xzalloc(sizeof(*offload
));
2341 offload
->flow
= flow
;
2344 dp_netdev_flow_ref(flow
);
2345 dp_netdev_pmd_try_ref(pmd
);
2351 dp_netdev_free_flow_offload(struct dp_flow_offload_item
*offload
)
2353 dp_netdev_pmd_unref(offload
->pmd
);
2354 dp_netdev_flow_unref(offload
->flow
);
2356 free(offload
->actions
);
2361 dp_netdev_append_flow_offload(struct dp_flow_offload_item
*offload
)
2363 ovs_mutex_lock(&dp_flow_offload
.mutex
);
2364 ovs_list_push_back(&dp_flow_offload
.list
, &offload
->node
);
2365 xpthread_cond_signal(&dp_flow_offload
.cond
);
2366 ovs_mutex_unlock(&dp_flow_offload
.mutex
);
2370 dp_netdev_flow_offload_del(struct dp_flow_offload_item
*offload
)
2372 return mark_to_flow_disassociate(offload
->pmd
, offload
->flow
);
2376 * There are two flow offload operations here: addition and modification.
2378 * For flow addition, this function does:
2379 * - allocate a new flow mark id
2380 * - perform hardware flow offload
2381 * - associate the flow mark with flow and mega flow
2383 * For flow modification, both flow mark and the associations are still
2384 * valid, thus only item 2 needed.
2387 dp_netdev_flow_offload_put(struct dp_flow_offload_item
*offload
)
2389 struct dp_netdev_pmd_thread
*pmd
= offload
->pmd
;
2390 struct dp_netdev_flow
*flow
= offload
->flow
;
2391 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
2392 bool modification
= offload
->op
== DP_NETDEV_FLOW_OFFLOAD_OP_MOD
;
2393 struct offload_info info
;
2394 struct netdev
*port
;
2404 ovs_assert(mark
!= INVALID_FLOW_MARK
);
2407 * If a mega flow has already been offloaded (from other PMD
2408 * instances), do not offload it again.
2410 mark
= megaflow_to_mark_find(&flow
->mega_ufid
);
2411 if (mark
!= INVALID_FLOW_MARK
) {
2412 VLOG_DBG("Flow has already been offloaded with mark %u\n", mark
);
2413 if (flow
->mark
!= INVALID_FLOW_MARK
) {
2414 ovs_assert(flow
->mark
== mark
);
2416 mark_to_flow_associate(mark
, flow
);
2421 mark
= flow_mark_alloc();
2422 if (mark
== INVALID_FLOW_MARK
) {
2423 VLOG_ERR("Failed to allocate flow mark!\n");
2426 info
.flow_mark
= mark
;
2428 port
= netdev_ports_get(in_port
, pmd
->dp
->class);
2429 if (!port
|| netdev_vport_is_vport_class(port
->netdev_class
)) {
2433 /* Taking a global 'port_mutex' to fulfill thread safety restrictions for
2434 * the netdev-offload-dpdk module. */
2435 ovs_mutex_lock(&pmd
->dp
->port_mutex
);
2436 ret
= netdev_flow_put(port
, &offload
->match
,
2437 CONST_CAST(struct nlattr
*, offload
->actions
),
2438 offload
->actions_len
, &flow
->mega_ufid
, &info
,
2440 ovs_mutex_unlock(&pmd
->dp
->port_mutex
);
2447 if (!modification
) {
2448 megaflow_to_mark_associate(&flow
->mega_ufid
, mark
);
2449 mark_to_flow_associate(mark
, flow
);
2454 if (!modification
) {
2455 flow_mark_free(mark
);
2457 mark_to_flow_disassociate(pmd
, flow
);
2463 dp_netdev_flow_offload_main(void *data OVS_UNUSED
)
2465 struct dp_flow_offload_item
*offload
;
2466 struct ovs_list
*list
;
2471 ovs_mutex_lock(&dp_flow_offload
.mutex
);
2472 if (ovs_list_is_empty(&dp_flow_offload
.list
)) {
2473 ovsrcu_quiesce_start();
2474 ovs_mutex_cond_wait(&dp_flow_offload
.cond
,
2475 &dp_flow_offload
.mutex
);
2476 ovsrcu_quiesce_end();
2478 list
= ovs_list_pop_front(&dp_flow_offload
.list
);
2479 offload
= CONTAINER_OF(list
, struct dp_flow_offload_item
, node
);
2480 ovs_mutex_unlock(&dp_flow_offload
.mutex
);
2482 switch (offload
->op
) {
2483 case DP_NETDEV_FLOW_OFFLOAD_OP_ADD
:
2485 ret
= dp_netdev_flow_offload_put(offload
);
2487 case DP_NETDEV_FLOW_OFFLOAD_OP_MOD
:
2489 ret
= dp_netdev_flow_offload_put(offload
);
2491 case DP_NETDEV_FLOW_OFFLOAD_OP_DEL
:
2493 ret
= dp_netdev_flow_offload_del(offload
);
2499 VLOG_DBG("%s to %s netdev flow\n",
2500 ret
== 0 ? "succeed" : "failed", op
);
2501 dp_netdev_free_flow_offload(offload
);
2508 queue_netdev_flow_del(struct dp_netdev_pmd_thread
*pmd
,
2509 struct dp_netdev_flow
*flow
)
2511 struct dp_flow_offload_item
*offload
;
2513 if (ovsthread_once_start(&offload_thread_once
)) {
2514 xpthread_cond_init(&dp_flow_offload
.cond
, NULL
);
2515 ovs_thread_create("dp_netdev_flow_offload",
2516 dp_netdev_flow_offload_main
, NULL
);
2517 ovsthread_once_done(&offload_thread_once
);
2520 offload
= dp_netdev_alloc_flow_offload(pmd
, flow
,
2521 DP_NETDEV_FLOW_OFFLOAD_OP_DEL
);
2522 dp_netdev_append_flow_offload(offload
);
2526 queue_netdev_flow_put(struct dp_netdev_pmd_thread
*pmd
,
2527 struct dp_netdev_flow
*flow
, struct match
*match
,
2528 const struct nlattr
*actions
, size_t actions_len
)
2530 struct dp_flow_offload_item
*offload
;
2533 if (!netdev_is_flow_api_enabled()) {
2537 if (ovsthread_once_start(&offload_thread_once
)) {
2538 xpthread_cond_init(&dp_flow_offload
.cond
, NULL
);
2539 ovs_thread_create("dp_netdev_flow_offload",
2540 dp_netdev_flow_offload_main
, NULL
);
2541 ovsthread_once_done(&offload_thread_once
);
2544 if (flow
->mark
!= INVALID_FLOW_MARK
) {
2545 op
= DP_NETDEV_FLOW_OFFLOAD_OP_MOD
;
2547 op
= DP_NETDEV_FLOW_OFFLOAD_OP_ADD
;
2549 offload
= dp_netdev_alloc_flow_offload(pmd
, flow
, op
);
2550 offload
->match
= *match
;
2551 offload
->actions
= xmalloc(actions_len
);
2552 memcpy(offload
->actions
, actions
, actions_len
);
2553 offload
->actions_len
= actions_len
;
2555 dp_netdev_append_flow_offload(offload
);
2559 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread
*pmd
,
2560 struct dp_netdev_flow
*flow
)
2561 OVS_REQUIRES(pmd
->flow_mutex
)
2563 struct cmap_node
*node
= CONST_CAST(struct cmap_node
*, &flow
->node
);
2565 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
2567 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2568 ovs_assert(cls
!= NULL
);
2569 dpcls_remove(cls
, &flow
->cr
);
2570 cmap_remove(&pmd
->flow_table
, node
, dp_netdev_flow_hash(&flow
->ufid
));
2571 if (flow
->mark
!= INVALID_FLOW_MARK
) {
2572 queue_netdev_flow_del(pmd
, flow
);
2576 dp_netdev_flow_unref(flow
);
2580 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
)
2582 struct dp_netdev_flow
*netdev_flow
;
2584 ovs_mutex_lock(&pmd
->flow_mutex
);
2585 CMAP_FOR_EACH (netdev_flow
, node
, &pmd
->flow_table
) {
2586 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
2588 ovs_mutex_unlock(&pmd
->flow_mutex
);
2592 dpif_netdev_flow_flush(struct dpif
*dpif
)
2594 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2595 struct dp_netdev_pmd_thread
*pmd
;
2597 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2598 dp_netdev_pmd_flow_flush(pmd
);
2604 struct dp_netdev_port_state
{
2605 struct hmap_position position
;
2610 dpif_netdev_port_dump_start(const struct dpif
*dpif OVS_UNUSED
, void **statep
)
2612 *statep
= xzalloc(sizeof(struct dp_netdev_port_state
));
2617 dpif_netdev_port_dump_next(const struct dpif
*dpif
, void *state_
,
2618 struct dpif_port
*dpif_port
)
2620 struct dp_netdev_port_state
*state
= state_
;
2621 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2622 struct hmap_node
*node
;
2625 ovs_mutex_lock(&dp
->port_mutex
);
2626 node
= hmap_at_position(&dp
->ports
, &state
->position
);
2628 struct dp_netdev_port
*port
;
2630 port
= CONTAINER_OF(node
, struct dp_netdev_port
, node
);
2633 state
->name
= xstrdup(netdev_get_name(port
->netdev
));
2634 dpif_port
->name
= state
->name
;
2635 dpif_port
->type
= port
->type
;
2636 dpif_port
->port_no
= port
->port_no
;
2642 ovs_mutex_unlock(&dp
->port_mutex
);
2648 dpif_netdev_port_dump_done(const struct dpif
*dpif OVS_UNUSED
, void *state_
)
2650 struct dp_netdev_port_state
*state
= state_
;
2657 dpif_netdev_port_poll(const struct dpif
*dpif_
, char **devnamep OVS_UNUSED
)
2659 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
2660 uint64_t new_port_seq
;
2663 new_port_seq
= seq_read(dpif
->dp
->port_seq
);
2664 if (dpif
->last_port_seq
!= new_port_seq
) {
2665 dpif
->last_port_seq
= new_port_seq
;
2675 dpif_netdev_port_poll_wait(const struct dpif
*dpif_
)
2677 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
2679 seq_wait(dpif
->dp
->port_seq
, dpif
->last_port_seq
);
2682 static struct dp_netdev_flow
*
2683 dp_netdev_flow_cast(const struct dpcls_rule
*cr
)
2685 return cr
? CONTAINER_OF(cr
, struct dp_netdev_flow
, cr
) : NULL
;
2688 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*flow
)
2690 return ovs_refcount_try_ref_rcu(&flow
->ref_cnt
);
2693 /* netdev_flow_key utilities.
2695 * netdev_flow_key is basically a miniflow. We use these functions
2696 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
2697 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
2699 * - Since we are dealing exclusively with miniflows created by
2700 * miniflow_extract(), if the map is different the miniflow is different.
2701 * Therefore we can be faster by comparing the map and the miniflow in a
2703 * - These functions can be inlined by the compiler. */
2705 /* Given the number of bits set in miniflow's maps, returns the size of the
2706 * 'netdev_flow_key.mf' */
2707 static inline size_t
2708 netdev_flow_key_size(size_t flow_u64s
)
2710 return sizeof(struct miniflow
) + MINIFLOW_VALUES_SIZE(flow_u64s
);
2714 netdev_flow_key_equal(const struct netdev_flow_key
*a
,
2715 const struct netdev_flow_key
*b
)
2717 /* 'b->len' may be not set yet. */
2718 return a
->hash
== b
->hash
&& !memcmp(&a
->mf
, &b
->mf
, a
->len
);
2721 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
2722 * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been
2723 * generated by miniflow_extract. */
2725 netdev_flow_key_equal_mf(const struct netdev_flow_key
*key
,
2726 const struct miniflow
*mf
)
2728 return !memcmp(&key
->mf
, mf
, key
->len
);
2732 netdev_flow_key_clone(struct netdev_flow_key
*dst
,
2733 const struct netdev_flow_key
*src
)
2736 offsetof(struct netdev_flow_key
, mf
) + src
->len
);
2739 /* Initialize a netdev_flow_key 'mask' from 'match'. */
2741 netdev_flow_mask_init(struct netdev_flow_key
*mask
,
2742 const struct match
*match
)
2744 uint64_t *dst
= miniflow_values(&mask
->mf
);
2745 struct flowmap fmap
;
2749 /* Only check masks that make sense for the flow. */
2750 flow_wc_map(&match
->flow
, &fmap
);
2751 flowmap_init(&mask
->mf
.map
);
2753 FLOWMAP_FOR_EACH_INDEX(idx
, fmap
) {
2754 uint64_t mask_u64
= flow_u64_value(&match
->wc
.masks
, idx
);
2757 flowmap_set(&mask
->mf
.map
, idx
, 1);
2759 hash
= hash_add64(hash
, mask_u64
);
2765 FLOWMAP_FOR_EACH_MAP (map
, mask
->mf
.map
) {
2766 hash
= hash_add64(hash
, map
);
2769 size_t n
= dst
- miniflow_get_values(&mask
->mf
);
2771 mask
->hash
= hash_finish(hash
, n
* 8);
2772 mask
->len
= netdev_flow_key_size(n
);
2775 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
2777 netdev_flow_key_init_masked(struct netdev_flow_key
*dst
,
2778 const struct flow
*flow
,
2779 const struct netdev_flow_key
*mask
)
2781 uint64_t *dst_u64
= miniflow_values(&dst
->mf
);
2782 const uint64_t *mask_u64
= miniflow_get_values(&mask
->mf
);
2786 dst
->len
= mask
->len
;
2787 dst
->mf
= mask
->mf
; /* Copy maps. */
2789 FLOW_FOR_EACH_IN_MAPS(value
, flow
, mask
->mf
.map
) {
2790 *dst_u64
= value
& *mask_u64
++;
2791 hash
= hash_add64(hash
, *dst_u64
++);
2793 dst
->hash
= hash_finish(hash
,
2794 (dst_u64
- miniflow_get_values(&dst
->mf
)) * 8);
2798 emc_entry_alive(struct emc_entry
*ce
)
2800 return ce
->flow
&& !ce
->flow
->dead
;
2804 emc_clear_entry(struct emc_entry
*ce
)
2807 dp_netdev_flow_unref(ce
->flow
);
2813 emc_change_entry(struct emc_entry
*ce
, struct dp_netdev_flow
*flow
,
2814 const struct netdev_flow_key
*key
)
2816 if (ce
->flow
!= flow
) {
2818 dp_netdev_flow_unref(ce
->flow
);
2821 if (dp_netdev_flow_ref(flow
)) {
2828 netdev_flow_key_clone(&ce
->key
, key
);
2833 emc_insert(struct emc_cache
*cache
, const struct netdev_flow_key
*key
,
2834 struct dp_netdev_flow
*flow
)
2836 struct emc_entry
*to_be_replaced
= NULL
;
2837 struct emc_entry
*current_entry
;
2839 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2840 if (netdev_flow_key_equal(¤t_entry
->key
, key
)) {
2841 /* We found the entry with the 'mf' miniflow */
2842 emc_change_entry(current_entry
, flow
, NULL
);
2846 /* Replacement policy: put the flow in an empty (not alive) entry, or
2847 * in the first entry where it can be */
2849 || (emc_entry_alive(to_be_replaced
)
2850 && !emc_entry_alive(current_entry
))
2851 || current_entry
->key
.hash
< to_be_replaced
->key
.hash
) {
2852 to_be_replaced
= current_entry
;
2855 /* We didn't find the miniflow in the cache.
2856 * The 'to_be_replaced' entry is where the new flow will be stored */
2858 emc_change_entry(to_be_replaced
, flow
, key
);
2862 emc_probabilistic_insert(struct dp_netdev_pmd_thread
*pmd
,
2863 const struct netdev_flow_key
*key
,
2864 struct dp_netdev_flow
*flow
)
2866 /* Insert an entry into the EMC based on probability value 'min'. By
2867 * default the value is UINT32_MAX / 100 which yields an insertion
2868 * probability of 1/100 ie. 1% */
2870 uint32_t min
= pmd
->ctx
.emc_insert_min
;
2872 if (min
&& random_uint32() <= min
) {
2873 emc_insert(&(pmd
->flow_cache
).emc_cache
, key
, flow
);
2877 static inline struct dp_netdev_flow
*
2878 emc_lookup(struct emc_cache
*cache
, const struct netdev_flow_key
*key
)
2880 struct emc_entry
*current_entry
;
2882 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2883 if (current_entry
->key
.hash
== key
->hash
2884 && emc_entry_alive(current_entry
)
2885 && netdev_flow_key_equal_mf(¤t_entry
->key
, &key
->mf
)) {
2887 /* We found the entry with the 'key->mf' miniflow */
2888 return current_entry
->flow
;
2895 static inline const struct cmap_node
*
2896 smc_entry_get(struct dp_netdev_pmd_thread
*pmd
, const uint32_t hash
)
2898 struct smc_cache
*cache
= &(pmd
->flow_cache
).smc_cache
;
2899 struct smc_bucket
*bucket
= &cache
->buckets
[hash
& SMC_MASK
];
2900 uint16_t sig
= hash
>> 16;
2901 uint16_t index
= UINT16_MAX
;
2903 for (int i
= 0; i
< SMC_ENTRY_PER_BUCKET
; i
++) {
2904 if (bucket
->sig
[i
] == sig
) {
2905 index
= bucket
->flow_idx
[i
];
2909 if (index
!= UINT16_MAX
) {
2910 return cmap_find_by_index(&pmd
->flow_table
, index
);
2916 smc_clear_entry(struct smc_bucket
*b
, int idx
)
2918 b
->flow_idx
[idx
] = UINT16_MAX
;
2921 /* Insert the flow_table index into SMC. Insertion may fail when 1) SMC is
2922 * turned off, 2) the flow_table index is larger than uint16_t can handle.
2923 * If there is already an SMC entry having same signature, the index will be
2924 * updated. If there is no existing entry, but an empty entry is available,
2925 * the empty entry will be taken. If no empty entry or existing same signature,
2926 * a random entry from the hashed bucket will be picked. */
2928 smc_insert(struct dp_netdev_pmd_thread
*pmd
,
2929 const struct netdev_flow_key
*key
,
2932 struct smc_cache
*smc_cache
= &(pmd
->flow_cache
).smc_cache
;
2933 struct smc_bucket
*bucket
= &smc_cache
->buckets
[key
->hash
& SMC_MASK
];
2935 uint32_t cmap_index
;
2939 atomic_read_relaxed(&pmd
->dp
->smc_enable_db
, &smc_enable_db
);
2940 if (!smc_enable_db
) {
2944 cmap_index
= cmap_find_index(&pmd
->flow_table
, hash
);
2945 index
= (cmap_index
>= UINT16_MAX
) ? UINT16_MAX
: (uint16_t)cmap_index
;
2947 /* If the index is larger than SMC can handle (uint16_t), we don't
2949 if (index
== UINT16_MAX
) {
2953 /* If an entry with same signature already exists, update the index */
2954 uint16_t sig
= key
->hash
>> 16;
2955 for (i
= 0; i
< SMC_ENTRY_PER_BUCKET
; i
++) {
2956 if (bucket
->sig
[i
] == sig
) {
2957 bucket
->flow_idx
[i
] = index
;
2961 /* If there is an empty entry, occupy it. */
2962 for (i
= 0; i
< SMC_ENTRY_PER_BUCKET
; i
++) {
2963 if (bucket
->flow_idx
[i
] == UINT16_MAX
) {
2964 bucket
->sig
[i
] = sig
;
2965 bucket
->flow_idx
[i
] = index
;
2969 /* Otherwise, pick a random entry. */
2970 i
= random_uint32() % SMC_ENTRY_PER_BUCKET
;
2971 bucket
->sig
[i
] = sig
;
2972 bucket
->flow_idx
[i
] = index
;
2975 static struct dp_netdev_flow
*
2976 dp_netdev_pmd_lookup_flow(struct dp_netdev_pmd_thread
*pmd
,
2977 const struct netdev_flow_key
*key
,
2981 struct dpcls_rule
*rule
;
2982 odp_port_t in_port
= u32_to_odp(MINIFLOW_GET_U32(&key
->mf
,
2984 struct dp_netdev_flow
*netdev_flow
= NULL
;
2986 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2987 if (OVS_LIKELY(cls
)) {
2988 dpcls_lookup(cls
, &key
, &rule
, 1, lookup_num_p
);
2989 netdev_flow
= dp_netdev_flow_cast(rule
);
2994 static struct dp_netdev_flow
*
2995 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread
*pmd
,
2996 const ovs_u128
*ufidp
, const struct nlattr
*key
,
2999 struct dp_netdev_flow
*netdev_flow
;
3003 /* If a UFID is not provided, determine one based on the key. */
3004 if (!ufidp
&& key
&& key_len
3005 && !dpif_netdev_flow_from_nlattrs(key
, key_len
, &flow
, false)) {
3006 dpif_flow_hash(pmd
->dp
->dpif
, &flow
, sizeof flow
, &ufid
);
3011 CMAP_FOR_EACH_WITH_HASH (netdev_flow
, node
, dp_netdev_flow_hash(ufidp
),
3013 if (ovs_u128_equals(netdev_flow
->ufid
, *ufidp
)) {
3023 get_dpif_flow_stats(const struct dp_netdev_flow
*netdev_flow_
,
3024 struct dpif_flow_stats
*stats
)
3026 struct dp_netdev_flow
*netdev_flow
;
3027 unsigned long long n
;
3031 netdev_flow
= CONST_CAST(struct dp_netdev_flow
*, netdev_flow_
);
3033 atomic_read_relaxed(&netdev_flow
->stats
.packet_count
, &n
);
3034 stats
->n_packets
= n
;
3035 atomic_read_relaxed(&netdev_flow
->stats
.byte_count
, &n
);
3037 atomic_read_relaxed(&netdev_flow
->stats
.used
, &used
);
3039 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
3040 stats
->tcp_flags
= flags
;
3043 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
3044 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
3045 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
3048 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow
*netdev_flow
,
3049 struct ofpbuf
*key_buf
, struct ofpbuf
*mask_buf
,
3050 struct dpif_flow
*flow
, bool terse
)
3053 memset(flow
, 0, sizeof *flow
);
3055 struct flow_wildcards wc
;
3056 struct dp_netdev_actions
*actions
;
3058 struct odp_flow_key_parms odp_parms
= {
3059 .flow
= &netdev_flow
->flow
,
3061 .support
= dp_netdev_support
,
3064 miniflow_expand(&netdev_flow
->cr
.mask
->mf
, &wc
.masks
);
3065 /* in_port is exact matched, but we have left it out from the mask for
3066 * optimnization reasons. Add in_port back to the mask. */
3067 wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
3070 offset
= key_buf
->size
;
3071 flow
->key
= ofpbuf_tail(key_buf
);
3072 odp_flow_key_from_flow(&odp_parms
, key_buf
);
3073 flow
->key_len
= key_buf
->size
- offset
;
3076 offset
= mask_buf
->size
;
3077 flow
->mask
= ofpbuf_tail(mask_buf
);
3078 odp_parms
.key_buf
= key_buf
;
3079 odp_flow_key_from_mask(&odp_parms
, mask_buf
);
3080 flow
->mask_len
= mask_buf
->size
- offset
;
3083 actions
= dp_netdev_flow_get_actions(netdev_flow
);
3084 flow
->actions
= actions
->actions
;
3085 flow
->actions_len
= actions
->size
;
3088 flow
->ufid
= netdev_flow
->ufid
;
3089 flow
->ufid_present
= true;
3090 flow
->pmd_id
= netdev_flow
->pmd_id
;
3091 get_dpif_flow_stats(netdev_flow
, &flow
->stats
);
3093 flow
->attrs
.offloaded
= false;
3094 flow
->attrs
.dp_layer
= "ovs";
3098 dpif_netdev_mask_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
3099 const struct nlattr
*mask_key
,
3100 uint32_t mask_key_len
, const struct flow
*flow
,
3101 struct flow_wildcards
*wc
, bool probe
)
3103 enum odp_key_fitness fitness
;
3105 fitness
= odp_flow_key_to_mask(mask_key
, mask_key_len
, wc
, flow
, NULL
);
3108 /* This should not happen: it indicates that
3109 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
3110 * disagree on the acceptable form of a mask. Log the problem
3111 * as an error, with enough details to enable debugging. */
3112 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
3114 if (!VLOG_DROP_ERR(&rl
)) {
3118 odp_flow_format(key
, key_len
, mask_key
, mask_key_len
, NULL
, &s
,
3120 VLOG_ERR("internal error parsing flow mask %s (%s)",
3121 ds_cstr(&s
), odp_key_fitness_to_string(fitness
));
3133 dpif_netdev_flow_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
3134 struct flow
*flow
, bool probe
)
3136 if (odp_flow_key_to_flow(key
, key_len
, flow
, NULL
)) {
3138 /* This should not happen: it indicates that
3139 * odp_flow_key_from_flow() and odp_flow_key_to_flow() disagree on
3140 * the acceptable form of a flow. Log the problem as an error,
3141 * with enough details to enable debugging. */
3142 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
3144 if (!VLOG_DROP_ERR(&rl
)) {
3148 odp_flow_format(key
, key_len
, NULL
, 0, NULL
, &s
, true);
3149 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s
));
3157 if (flow
->ct_state
& DP_NETDEV_CS_UNSUPPORTED_MASK
) {
3165 dpif_netdev_flow_get(const struct dpif
*dpif
, const struct dpif_flow_get
*get
)
3167 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3168 struct dp_netdev_flow
*netdev_flow
;
3169 struct dp_netdev_pmd_thread
*pmd
;
3170 struct hmapx to_find
= HMAPX_INITIALIZER(&to_find
);
3171 struct hmapx_node
*node
;
3174 if (get
->pmd_id
== PMD_ID_NULL
) {
3175 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3176 if (dp_netdev_pmd_try_ref(pmd
) && !hmapx_add(&to_find
, pmd
)) {
3177 dp_netdev_pmd_unref(pmd
);
3181 pmd
= dp_netdev_get_pmd(dp
, get
->pmd_id
);
3185 hmapx_add(&to_find
, pmd
);
3188 if (!hmapx_count(&to_find
)) {
3192 HMAPX_FOR_EACH (node
, &to_find
) {
3193 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3194 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, get
->ufid
, get
->key
,
3197 dp_netdev_flow_to_dpif_flow(netdev_flow
, get
->buffer
, get
->buffer
,
3206 HMAPX_FOR_EACH (node
, &to_find
) {
3207 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3208 dp_netdev_pmd_unref(pmd
);
3211 hmapx_destroy(&to_find
);
3216 dp_netdev_get_mega_ufid(const struct match
*match
, ovs_u128
*mega_ufid
)
3218 struct flow masked_flow
;
3221 for (i
= 0; i
< sizeof(struct flow
); i
++) {
3222 ((uint8_t *)&masked_flow
)[i
] = ((uint8_t *)&match
->flow
)[i
] &
3223 ((uint8_t *)&match
->wc
)[i
];
3225 dpif_flow_hash(NULL
, &masked_flow
, sizeof(struct flow
), mega_ufid
);
3228 static struct dp_netdev_flow
*
3229 dp_netdev_flow_add(struct dp_netdev_pmd_thread
*pmd
,
3230 struct match
*match
, const ovs_u128
*ufid
,
3231 const struct nlattr
*actions
, size_t actions_len
)
3232 OVS_REQUIRES(pmd
->flow_mutex
)
3234 struct dp_netdev_flow
*flow
;
3235 struct netdev_flow_key mask
;
3238 /* Make sure in_port is exact matched before we read it. */
3239 ovs_assert(match
->wc
.masks
.in_port
.odp_port
== ODPP_NONE
);
3240 odp_port_t in_port
= match
->flow
.in_port
.odp_port
;
3242 /* As we select the dpcls based on the port number, each netdev flow
3243 * belonging to the same dpcls will have the same odp_port value.
3244 * For performance reasons we wildcard odp_port here in the mask. In the
3245 * typical case dp_hash is also wildcarded, and the resulting 8-byte
3246 * chunk {dp_hash, in_port} will be ignored by netdev_flow_mask_init() and
3247 * will not be part of the subtable mask.
3248 * This will speed up the hash computation during dpcls_lookup() because
3249 * there is one less call to hash_add64() in this case. */
3250 match
->wc
.masks
.in_port
.odp_port
= 0;
3251 netdev_flow_mask_init(&mask
, match
);
3252 match
->wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
3254 /* Make sure wc does not have metadata. */
3255 ovs_assert(!FLOWMAP_HAS_FIELD(&mask
.mf
.map
, metadata
)
3256 && !FLOWMAP_HAS_FIELD(&mask
.mf
.map
, regs
));
3258 /* Do not allocate extra space. */
3259 flow
= xmalloc(sizeof *flow
- sizeof flow
->cr
.flow
.mf
+ mask
.len
);
3260 memset(&flow
->stats
, 0, sizeof flow
->stats
);
3263 flow
->mark
= INVALID_FLOW_MARK
;
3264 *CONST_CAST(unsigned *, &flow
->pmd_id
) = pmd
->core_id
;
3265 *CONST_CAST(struct flow
*, &flow
->flow
) = match
->flow
;
3266 *CONST_CAST(ovs_u128
*, &flow
->ufid
) = *ufid
;
3267 ovs_refcount_init(&flow
->ref_cnt
);
3268 ovsrcu_set(&flow
->actions
, dp_netdev_actions_create(actions
, actions_len
));
3270 dp_netdev_get_mega_ufid(match
, CONST_CAST(ovs_u128
*, &flow
->mega_ufid
));
3271 netdev_flow_key_init_masked(&flow
->cr
.flow
, &match
->flow
, &mask
);
3273 /* Select dpcls for in_port. Relies on in_port to be exact match. */
3274 cls
= dp_netdev_pmd_find_dpcls(pmd
, in_port
);
3275 dpcls_insert(cls
, &flow
->cr
, &mask
);
3277 cmap_insert(&pmd
->flow_table
, CONST_CAST(struct cmap_node
*, &flow
->node
),
3278 dp_netdev_flow_hash(&flow
->ufid
));
3280 queue_netdev_flow_put(pmd
, flow
, match
, actions
, actions_len
);
3282 if (OVS_UNLIKELY(!VLOG_DROP_DBG((&upcall_rl
)))) {
3283 struct ds ds
= DS_EMPTY_INITIALIZER
;
3284 struct ofpbuf key_buf
, mask_buf
;
3285 struct odp_flow_key_parms odp_parms
= {
3286 .flow
= &match
->flow
,
3287 .mask
= &match
->wc
.masks
,
3288 .support
= dp_netdev_support
,
3291 ofpbuf_init(&key_buf
, 0);
3292 ofpbuf_init(&mask_buf
, 0);
3294 odp_flow_key_from_flow(&odp_parms
, &key_buf
);
3295 odp_parms
.key_buf
= &key_buf
;
3296 odp_flow_key_from_mask(&odp_parms
, &mask_buf
);
3298 ds_put_cstr(&ds
, "flow_add: ");
3299 odp_format_ufid(ufid
, &ds
);
3300 ds_put_cstr(&ds
, " ");
3301 odp_flow_format(key_buf
.data
, key_buf
.size
,
3302 mask_buf
.data
, mask_buf
.size
,
3304 ds_put_cstr(&ds
, ", actions:");
3305 format_odp_actions(&ds
, actions
, actions_len
, NULL
);
3307 VLOG_DBG("%s", ds_cstr(&ds
));
3309 ofpbuf_uninit(&key_buf
);
3310 ofpbuf_uninit(&mask_buf
);
3312 /* Add a printout of the actual match installed. */
3315 ds_put_cstr(&ds
, "flow match: ");
3316 miniflow_expand(&flow
->cr
.flow
.mf
, &m
.flow
);
3317 miniflow_expand(&flow
->cr
.mask
->mf
, &m
.wc
.masks
);
3318 memset(&m
.tun_md
, 0, sizeof m
.tun_md
);
3319 match_format(&m
, NULL
, &ds
, OFP_DEFAULT_PRIORITY
);
3321 VLOG_DBG("%s", ds_cstr(&ds
));
3330 flow_put_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
3331 struct netdev_flow_key
*key
,
3332 struct match
*match
,
3334 const struct dpif_flow_put
*put
,
3335 struct dpif_flow_stats
*stats
)
3337 struct dp_netdev_flow
*netdev_flow
;
3341 memset(stats
, 0, sizeof *stats
);
3344 ovs_mutex_lock(&pmd
->flow_mutex
);
3345 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
3347 if (put
->flags
& DPIF_FP_CREATE
) {
3348 if (cmap_count(&pmd
->flow_table
) < MAX_FLOWS
) {
3349 dp_netdev_flow_add(pmd
, match
, ufid
, put
->actions
,
3359 if (put
->flags
& DPIF_FP_MODIFY
) {
3360 struct dp_netdev_actions
*new_actions
;
3361 struct dp_netdev_actions
*old_actions
;
3363 new_actions
= dp_netdev_actions_create(put
->actions
,
3366 old_actions
= dp_netdev_flow_get_actions(netdev_flow
);
3367 ovsrcu_set(&netdev_flow
->actions
, new_actions
);
3369 queue_netdev_flow_put(pmd
, netdev_flow
, match
,
3370 put
->actions
, put
->actions_len
);
3373 get_dpif_flow_stats(netdev_flow
, stats
);
3375 if (put
->flags
& DPIF_FP_ZERO_STATS
) {
3376 /* XXX: The userspace datapath uses thread local statistics
3377 * (for flows), which should be updated only by the owning
3378 * thread. Since we cannot write on stats memory here,
3379 * we choose not to support this flag. Please note:
3380 * - This feature is currently used only by dpctl commands with
3382 * - Should the need arise, this operation can be implemented
3383 * by keeping a base value (to be update here) for each
3384 * counter, and subtracting it before outputting the stats */
3388 ovsrcu_postpone(dp_netdev_actions_free
, old_actions
);
3389 } else if (put
->flags
& DPIF_FP_CREATE
) {
3392 /* Overlapping flow. */
3396 ovs_mutex_unlock(&pmd
->flow_mutex
);
3401 dpif_netdev_flow_put(struct dpif
*dpif
, const struct dpif_flow_put
*put
)
3403 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3404 struct netdev_flow_key key
, mask
;
3405 struct dp_netdev_pmd_thread
*pmd
;
3409 bool probe
= put
->flags
& DPIF_FP_PROBE
;
3412 memset(put
->stats
, 0, sizeof *put
->stats
);
3414 error
= dpif_netdev_flow_from_nlattrs(put
->key
, put
->key_len
, &match
.flow
,
3419 error
= dpif_netdev_mask_from_nlattrs(put
->key
, put
->key_len
,
3420 put
->mask
, put
->mask_len
,
3421 &match
.flow
, &match
.wc
, probe
);
3429 dpif_flow_hash(dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
3432 /* The Netlink encoding of datapath flow keys cannot express
3433 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3434 * tag is interpreted as exact match on the fact that there is no
3435 * VLAN. Unless we refactor a lot of code that translates between
3436 * Netlink and struct flow representations, we have to do the same
3437 * here. This must be in sync with 'match' in handle_packet_upcall(). */
3438 if (!match
.wc
.masks
.vlans
[0].tci
) {
3439 match
.wc
.masks
.vlans
[0].tci
= htons(0xffff);
3442 /* Must produce a netdev_flow_key for lookup.
3443 * Use the same method as employed to create the key when adding
3444 * the flow to the dplcs to make sure they match. */
3445 netdev_flow_mask_init(&mask
, &match
);
3446 netdev_flow_key_init_masked(&key
, &match
.flow
, &mask
);
3448 if (put
->pmd_id
== PMD_ID_NULL
) {
3449 if (cmap_count(&dp
->poll_threads
) == 0) {
3452 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3453 struct dpif_flow_stats pmd_stats
;
3456 pmd_error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
,
3460 } else if (put
->stats
) {
3461 put
->stats
->n_packets
+= pmd_stats
.n_packets
;
3462 put
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
3463 put
->stats
->used
= MAX(put
->stats
->used
, pmd_stats
.used
);
3464 put
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
3468 pmd
= dp_netdev_get_pmd(dp
, put
->pmd_id
);
3472 error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
, put
->stats
);
3473 dp_netdev_pmd_unref(pmd
);
3480 flow_del_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
3481 struct dpif_flow_stats
*stats
,
3482 const struct dpif_flow_del
*del
)
3484 struct dp_netdev_flow
*netdev_flow
;
3487 ovs_mutex_lock(&pmd
->flow_mutex
);
3488 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, del
->ufid
, del
->key
,
3492 get_dpif_flow_stats(netdev_flow
, stats
);
3494 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
3498 ovs_mutex_unlock(&pmd
->flow_mutex
);
3504 dpif_netdev_flow_del(struct dpif
*dpif
, const struct dpif_flow_del
*del
)
3506 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3507 struct dp_netdev_pmd_thread
*pmd
;
3511 memset(del
->stats
, 0, sizeof *del
->stats
);
3514 if (del
->pmd_id
== PMD_ID_NULL
) {
3515 if (cmap_count(&dp
->poll_threads
) == 0) {
3518 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3519 struct dpif_flow_stats pmd_stats
;
3522 pmd_error
= flow_del_on_pmd(pmd
, &pmd_stats
, del
);
3525 } else if (del
->stats
) {
3526 del
->stats
->n_packets
+= pmd_stats
.n_packets
;
3527 del
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
3528 del
->stats
->used
= MAX(del
->stats
->used
, pmd_stats
.used
);
3529 del
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
3533 pmd
= dp_netdev_get_pmd(dp
, del
->pmd_id
);
3537 error
= flow_del_on_pmd(pmd
, del
->stats
, del
);
3538 dp_netdev_pmd_unref(pmd
);
3545 struct dpif_netdev_flow_dump
{
3546 struct dpif_flow_dump up
;
3547 struct cmap_position poll_thread_pos
;
3548 struct cmap_position flow_pos
;
3549 struct dp_netdev_pmd_thread
*cur_pmd
;
3551 struct ovs_mutex mutex
;
3554 static struct dpif_netdev_flow_dump
*
3555 dpif_netdev_flow_dump_cast(struct dpif_flow_dump
*dump
)
3557 return CONTAINER_OF(dump
, struct dpif_netdev_flow_dump
, up
);
3560 static struct dpif_flow_dump
*
3561 dpif_netdev_flow_dump_create(const struct dpif
*dpif_
, bool terse
,
3562 struct dpif_flow_dump_types
*types OVS_UNUSED
)
3564 struct dpif_netdev_flow_dump
*dump
;
3566 dump
= xzalloc(sizeof *dump
);
3567 dpif_flow_dump_init(&dump
->up
, dpif_
);
3568 dump
->up
.terse
= terse
;
3569 ovs_mutex_init(&dump
->mutex
);
3575 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump
*dump_
)
3577 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
3579 ovs_mutex_destroy(&dump
->mutex
);
3584 struct dpif_netdev_flow_dump_thread
{
3585 struct dpif_flow_dump_thread up
;
3586 struct dpif_netdev_flow_dump
*dump
;
3587 struct odputil_keybuf keybuf
[FLOW_DUMP_MAX_BATCH
];
3588 struct odputil_keybuf maskbuf
[FLOW_DUMP_MAX_BATCH
];
3591 static struct dpif_netdev_flow_dump_thread
*
3592 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread
*thread
)
3594 return CONTAINER_OF(thread
, struct dpif_netdev_flow_dump_thread
, up
);
3597 static struct dpif_flow_dump_thread
*
3598 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump
*dump_
)
3600 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
3601 struct dpif_netdev_flow_dump_thread
*thread
;
3603 thread
= xmalloc(sizeof *thread
);
3604 dpif_flow_dump_thread_init(&thread
->up
, &dump
->up
);
3605 thread
->dump
= dump
;
3610 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread
*thread_
)
3612 struct dpif_netdev_flow_dump_thread
*thread
3613 = dpif_netdev_flow_dump_thread_cast(thread_
);
3619 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread
*thread_
,
3620 struct dpif_flow
*flows
, int max_flows
)
3622 struct dpif_netdev_flow_dump_thread
*thread
3623 = dpif_netdev_flow_dump_thread_cast(thread_
);
3624 struct dpif_netdev_flow_dump
*dump
= thread
->dump
;
3625 struct dp_netdev_flow
*netdev_flows
[FLOW_DUMP_MAX_BATCH
];
3629 ovs_mutex_lock(&dump
->mutex
);
3630 if (!dump
->status
) {
3631 struct dpif_netdev
*dpif
= dpif_netdev_cast(thread
->up
.dpif
);
3632 struct dp_netdev
*dp
= get_dp_netdev(&dpif
->dpif
);
3633 struct dp_netdev_pmd_thread
*pmd
= dump
->cur_pmd
;
3634 int flow_limit
= MIN(max_flows
, FLOW_DUMP_MAX_BATCH
);
3636 /* First call to dump_next(), extracts the first pmd thread.
3637 * If there is no pmd thread, returns immediately. */
3639 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
3641 ovs_mutex_unlock(&dump
->mutex
);
3648 for (n_flows
= 0; n_flows
< flow_limit
; n_flows
++) {
3649 struct cmap_node
*node
;
3651 node
= cmap_next_position(&pmd
->flow_table
, &dump
->flow_pos
);
3655 netdev_flows
[n_flows
] = CONTAINER_OF(node
,
3656 struct dp_netdev_flow
,
3659 /* When finishing dumping the current pmd thread, moves to
3661 if (n_flows
< flow_limit
) {
3662 memset(&dump
->flow_pos
, 0, sizeof dump
->flow_pos
);
3663 dp_netdev_pmd_unref(pmd
);
3664 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
3670 /* Keeps the reference to next caller. */
3671 dump
->cur_pmd
= pmd
;
3673 /* If the current dump is empty, do not exit the loop, since the
3674 * remaining pmds could have flows to be dumped. Just dumps again
3675 * on the new 'pmd'. */
3678 ovs_mutex_unlock(&dump
->mutex
);
3680 for (i
= 0; i
< n_flows
; i
++) {
3681 struct odputil_keybuf
*maskbuf
= &thread
->maskbuf
[i
];
3682 struct odputil_keybuf
*keybuf
= &thread
->keybuf
[i
];
3683 struct dp_netdev_flow
*netdev_flow
= netdev_flows
[i
];
3684 struct dpif_flow
*f
= &flows
[i
];
3685 struct ofpbuf key
, mask
;
3687 ofpbuf_use_stack(&key
, keybuf
, sizeof *keybuf
);
3688 ofpbuf_use_stack(&mask
, maskbuf
, sizeof *maskbuf
);
3689 dp_netdev_flow_to_dpif_flow(netdev_flow
, &key
, &mask
, f
,
3697 dpif_netdev_execute(struct dpif
*dpif
, struct dpif_execute
*execute
)
3698 OVS_NO_THREAD_SAFETY_ANALYSIS
3700 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3701 struct dp_netdev_pmd_thread
*pmd
;
3702 struct dp_packet_batch pp
;
3704 if (dp_packet_size(execute
->packet
) < ETH_HEADER_LEN
||
3705 dp_packet_size(execute
->packet
) > UINT16_MAX
) {
3709 /* Tries finding the 'pmd'. If NULL is returned, that means
3710 * the current thread is a non-pmd thread and should use
3711 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
3712 pmd
= ovsthread_getspecific(dp
->per_pmd_key
);
3714 pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
3720 if (execute
->probe
) {
3721 /* If this is part of a probe, Drop the packet, since executing
3722 * the action may actually cause spurious packets be sent into
3724 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3725 dp_netdev_pmd_unref(pmd
);
3730 /* If the current thread is non-pmd thread, acquires
3731 * the 'non_pmd_mutex'. */
3732 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3733 ovs_mutex_lock(&dp
->non_pmd_mutex
);
3736 /* Update current time in PMD context. We don't care about EMC insertion
3737 * probability, because we are on a slow path. */
3738 pmd_thread_ctx_time_update(pmd
);
3740 /* The action processing expects the RSS hash to be valid, because
3741 * it's always initialized at the beginning of datapath processing.
3742 * In this case, though, 'execute->packet' may not have gone through
3743 * the datapath at all, it may have been generated by the upper layer
3744 * (OpenFlow packet-out, BFD frame, ...). */
3745 if (!dp_packet_rss_valid(execute
->packet
)) {
3746 dp_packet_set_rss_hash(execute
->packet
,
3747 flow_hash_5tuple(execute
->flow
, 0));
3750 dp_packet_batch_init_packet(&pp
, execute
->packet
);
3751 pp
.do_not_steal
= true;
3752 dp_netdev_execute_actions(pmd
, &pp
, false, execute
->flow
,
3753 execute
->actions
, execute
->actions_len
);
3754 dp_netdev_pmd_flush_output_packets(pmd
, true);
3756 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3757 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
3758 dp_netdev_pmd_unref(pmd
);
3765 dpif_netdev_operate(struct dpif
*dpif
, struct dpif_op
**ops
, size_t n_ops
,
3766 enum dpif_offload_type offload_type OVS_UNUSED
)
3770 for (i
= 0; i
< n_ops
; i
++) {
3771 struct dpif_op
*op
= ops
[i
];
3774 case DPIF_OP_FLOW_PUT
:
3775 op
->error
= dpif_netdev_flow_put(dpif
, &op
->flow_put
);
3778 case DPIF_OP_FLOW_DEL
:
3779 op
->error
= dpif_netdev_flow_del(dpif
, &op
->flow_del
);
3782 case DPIF_OP_EXECUTE
:
3783 op
->error
= dpif_netdev_execute(dpif
, &op
->execute
);
3786 case DPIF_OP_FLOW_GET
:
3787 op
->error
= dpif_netdev_flow_get(dpif
, &op
->flow_get
);
3793 /* Enable or Disable PMD auto load balancing. */
3795 set_pmd_auto_lb(struct dp_netdev
*dp
)
3797 unsigned int cnt
= 0;
3798 struct dp_netdev_pmd_thread
*pmd
;
3799 struct pmd_auto_lb
*pmd_alb
= &dp
->pmd_alb
;
3801 bool enable_alb
= false;
3802 bool multi_rxq
= false;
3803 bool pmd_rxq_assign_cyc
= dp
->pmd_rxq_assign_cyc
;
3805 /* Ensure that there is at least 2 non-isolated PMDs and
3806 * one of them is polling more than one rxq. */
3807 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3808 if (pmd
->core_id
== NON_PMD_CORE_ID
|| pmd
->isolated
) {
3812 if (hmap_count(&pmd
->poll_list
) > 1) {
3815 if (cnt
&& multi_rxq
) {
3822 /* Enable auto LB if it is requested and cycle based assignment is true. */
3823 enable_alb
= enable_alb
&& pmd_rxq_assign_cyc
&&
3824 pmd_alb
->auto_lb_requested
;
3826 if (pmd_alb
->is_enabled
!= enable_alb
) {
3827 pmd_alb
->is_enabled
= enable_alb
;
3828 if (pmd_alb
->is_enabled
) {
3829 VLOG_INFO("PMD auto load balance is enabled "
3830 "(with rebalance interval:%"PRIu64
" msec)",
3831 pmd_alb
->rebalance_intvl
);
3833 pmd_alb
->rebalance_poll_timer
= 0;
3834 VLOG_INFO("PMD auto load balance is disabled");
3840 /* Applies datapath configuration from the database. Some of the changes are
3841 * actually applied in dpif_netdev_run(). */
3843 dpif_netdev_set_config(struct dpif
*dpif
, const struct smap
*other_config
)
3845 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3846 const char *cmask
= smap_get(other_config
, "pmd-cpu-mask");
3847 const char *pmd_rxq_assign
= smap_get_def(other_config
, "pmd-rxq-assign",
3849 unsigned long long insert_prob
=
3850 smap_get_ullong(other_config
, "emc-insert-inv-prob",
3851 DEFAULT_EM_FLOW_INSERT_INV_PROB
);
3852 uint32_t insert_min
, cur_min
;
3853 uint32_t tx_flush_interval
, cur_tx_flush_interval
;
3854 uint64_t rebalance_intvl
;
3856 tx_flush_interval
= smap_get_int(other_config
, "tx-flush-interval",
3857 DEFAULT_TX_FLUSH_INTERVAL
);
3858 atomic_read_relaxed(&dp
->tx_flush_interval
, &cur_tx_flush_interval
);
3859 if (tx_flush_interval
!= cur_tx_flush_interval
) {
3860 atomic_store_relaxed(&dp
->tx_flush_interval
, tx_flush_interval
);
3861 VLOG_INFO("Flushing interval for tx queues set to %"PRIu32
" us",
3865 if (!nullable_string_is_equal(dp
->pmd_cmask
, cmask
)) {
3866 free(dp
->pmd_cmask
);
3867 dp
->pmd_cmask
= nullable_xstrdup(cmask
);
3868 dp_netdev_request_reconfigure(dp
);
3871 atomic_read_relaxed(&dp
->emc_insert_min
, &cur_min
);
3872 if (insert_prob
<= UINT32_MAX
) {
3873 insert_min
= insert_prob
== 0 ? 0 : UINT32_MAX
/ insert_prob
;
3875 insert_min
= DEFAULT_EM_FLOW_INSERT_MIN
;
3876 insert_prob
= DEFAULT_EM_FLOW_INSERT_INV_PROB
;
3879 if (insert_min
!= cur_min
) {
3880 atomic_store_relaxed(&dp
->emc_insert_min
, insert_min
);
3881 if (insert_min
== 0) {
3882 VLOG_INFO("EMC insertion probability changed to zero");
3884 VLOG_INFO("EMC insertion probability changed to 1/%llu (~%.2f%%)",
3885 insert_prob
, (100 / (float)insert_prob
));
3889 bool perf_enabled
= smap_get_bool(other_config
, "pmd-perf-metrics", false);
3890 bool cur_perf_enabled
;
3891 atomic_read_relaxed(&dp
->pmd_perf_metrics
, &cur_perf_enabled
);
3892 if (perf_enabled
!= cur_perf_enabled
) {
3893 atomic_store_relaxed(&dp
->pmd_perf_metrics
, perf_enabled
);
3895 VLOG_INFO("PMD performance metrics collection enabled");
3897 VLOG_INFO("PMD performance metrics collection disabled");
3901 bool smc_enable
= smap_get_bool(other_config
, "smc-enable", false);
3903 atomic_read_relaxed(&dp
->smc_enable_db
, &cur_smc
);
3904 if (smc_enable
!= cur_smc
) {
3905 atomic_store_relaxed(&dp
->smc_enable_db
, smc_enable
);
3907 VLOG_INFO("SMC cache is enabled");
3909 VLOG_INFO("SMC cache is disabled");
3913 bool pmd_rxq_assign_cyc
= !strcmp(pmd_rxq_assign
, "cycles");
3914 if (!pmd_rxq_assign_cyc
&& strcmp(pmd_rxq_assign
, "roundrobin")) {
3915 VLOG_WARN("Unsupported Rxq to PMD assignment mode in pmd-rxq-assign. "
3916 "Defaulting to 'cycles'.");
3917 pmd_rxq_assign_cyc
= true;
3918 pmd_rxq_assign
= "cycles";
3920 if (dp
->pmd_rxq_assign_cyc
!= pmd_rxq_assign_cyc
) {
3921 dp
->pmd_rxq_assign_cyc
= pmd_rxq_assign_cyc
;
3922 VLOG_INFO("Rxq to PMD assignment mode changed to: \'%s\'.",
3924 dp_netdev_request_reconfigure(dp
);
3927 struct pmd_auto_lb
*pmd_alb
= &dp
->pmd_alb
;
3928 pmd_alb
->auto_lb_requested
= smap_get_bool(other_config
, "pmd-auto-lb",
3931 rebalance_intvl
= smap_get_int(other_config
, "pmd-auto-lb-rebal-interval",
3932 ALB_PMD_REBALANCE_POLL_INTERVAL
);
3934 /* Input is in min, convert it to msec. */
3936 rebalance_intvl
? rebalance_intvl
* MIN_TO_MSEC
: MIN_TO_MSEC
;
3938 if (pmd_alb
->rebalance_intvl
!= rebalance_intvl
) {
3939 pmd_alb
->rebalance_intvl
= rebalance_intvl
;
3942 set_pmd_auto_lb(dp
);
3946 /* Parses affinity list and returns result in 'core_ids'. */
3948 parse_affinity_list(const char *affinity_list
, unsigned *core_ids
, int n_rxq
)
3951 char *list
, *copy
, *key
, *value
;
3954 for (i
= 0; i
< n_rxq
; i
++) {
3955 core_ids
[i
] = OVS_CORE_UNSPEC
;
3958 if (!affinity_list
) {
3962 list
= copy
= xstrdup(affinity_list
);
3964 while (ofputil_parse_key_value(&list
, &key
, &value
)) {
3965 int rxq_id
, core_id
;
3967 if (!str_to_int(key
, 0, &rxq_id
) || rxq_id
< 0
3968 || !str_to_int(value
, 0, &core_id
) || core_id
< 0) {
3973 if (rxq_id
< n_rxq
) {
3974 core_ids
[rxq_id
] = core_id
;
3982 /* Parses 'affinity_list' and applies configuration if it is valid. */
3984 dpif_netdev_port_set_rxq_affinity(struct dp_netdev_port
*port
,
3985 const char *affinity_list
)
3987 unsigned *core_ids
, i
;
3990 core_ids
= xmalloc(port
->n_rxq
* sizeof *core_ids
);
3991 if (parse_affinity_list(affinity_list
, core_ids
, port
->n_rxq
)) {
3996 for (i
= 0; i
< port
->n_rxq
; i
++) {
3997 port
->rxqs
[i
].core_id
= core_ids
[i
];
4005 /* Returns 'true' if one of the 'port's RX queues exists in 'poll_list'
4006 * of given PMD thread. */
4008 dpif_netdev_pmd_polls_port(struct dp_netdev_pmd_thread
*pmd
,
4009 struct dp_netdev_port
*port
)
4010 OVS_EXCLUDED(pmd
->port_mutex
)
4012 struct rxq_poll
*poll
;
4015 ovs_mutex_lock(&pmd
->port_mutex
);
4016 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
4017 if (port
== poll
->rxq
->port
) {
4022 ovs_mutex_unlock(&pmd
->port_mutex
);
4026 /* Updates port configuration from the database. The changes are actually
4027 * applied in dpif_netdev_run(). */
4029 dpif_netdev_port_set_config(struct dpif
*dpif
, odp_port_t port_no
,
4030 const struct smap
*cfg
)
4032 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4033 struct dp_netdev_port
*port
;
4035 const char *affinity_list
= smap_get(cfg
, "pmd-rxq-affinity");
4036 bool emc_enabled
= smap_get_bool(cfg
, "emc-enable", true);
4038 ovs_mutex_lock(&dp
->port_mutex
);
4039 error
= get_port_by_number(dp
, port_no
, &port
);
4044 if (emc_enabled
!= port
->emc_enabled
) {
4045 struct dp_netdev_pmd_thread
*pmd
;
4046 struct ds ds
= DS_EMPTY_INITIALIZER
;
4047 uint32_t cur_min
, insert_prob
;
4049 port
->emc_enabled
= emc_enabled
;
4050 /* Mark for reload all the threads that polls this port and request
4051 * for reconfiguration for the actual reloading of threads. */
4052 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4053 if (dpif_netdev_pmd_polls_port(pmd
, port
)) {
4054 pmd
->need_reload
= true;
4057 dp_netdev_request_reconfigure(dp
);
4059 ds_put_format(&ds
, "%s: EMC has been %s.",
4060 netdev_get_name(port
->netdev
),
4061 (emc_enabled
) ? "enabled" : "disabled");
4063 ds_put_cstr(&ds
, " Current insertion probability is ");
4064 atomic_read_relaxed(&dp
->emc_insert_min
, &cur_min
);
4066 ds_put_cstr(&ds
, "zero.");
4068 insert_prob
= UINT32_MAX
/ cur_min
;
4069 ds_put_format(&ds
, "1/%"PRIu32
" (~%.2f%%).",
4070 insert_prob
, 100 / (float) insert_prob
);
4073 VLOG_INFO("%s", ds_cstr(&ds
));
4077 /* Checking for RXq affinity changes. */
4078 if (!netdev_is_pmd(port
->netdev
)
4079 || nullable_string_is_equal(affinity_list
, port
->rxq_affinity_list
)) {
4083 error
= dpif_netdev_port_set_rxq_affinity(port
, affinity_list
);
4087 free(port
->rxq_affinity_list
);
4088 port
->rxq_affinity_list
= nullable_xstrdup(affinity_list
);
4090 dp_netdev_request_reconfigure(dp
);
4092 ovs_mutex_unlock(&dp
->port_mutex
);
4097 dpif_netdev_queue_to_priority(const struct dpif
*dpif OVS_UNUSED
,
4098 uint32_t queue_id
, uint32_t *priority
)
4100 *priority
= queue_id
;
4105 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
4106 * a copy of the 'size' bytes of 'actions' input parameters. */
4107 struct dp_netdev_actions
*
4108 dp_netdev_actions_create(const struct nlattr
*actions
, size_t size
)
4110 struct dp_netdev_actions
*netdev_actions
;
4112 netdev_actions
= xmalloc(sizeof *netdev_actions
+ size
);
4113 memcpy(netdev_actions
->actions
, actions
, size
);
4114 netdev_actions
->size
= size
;
4116 return netdev_actions
;
4119 struct dp_netdev_actions
*
4120 dp_netdev_flow_get_actions(const struct dp_netdev_flow
*flow
)
4122 return ovsrcu_get(struct dp_netdev_actions
*, &flow
->actions
);
4126 dp_netdev_actions_free(struct dp_netdev_actions
*actions
)
4132 dp_netdev_rxq_set_cycles(struct dp_netdev_rxq
*rx
,
4133 enum rxq_cycles_counter_type type
,
4134 unsigned long long cycles
)
4136 atomic_store_relaxed(&rx
->cycles
[type
], cycles
);
4140 dp_netdev_rxq_add_cycles(struct dp_netdev_rxq
*rx
,
4141 enum rxq_cycles_counter_type type
,
4142 unsigned long long cycles
)
4144 non_atomic_ullong_add(&rx
->cycles
[type
], cycles
);
4148 dp_netdev_rxq_get_cycles(struct dp_netdev_rxq
*rx
,
4149 enum rxq_cycles_counter_type type
)
4151 unsigned long long processing_cycles
;
4152 atomic_read_relaxed(&rx
->cycles
[type
], &processing_cycles
);
4153 return processing_cycles
;
4157 dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq
*rx
,
4158 unsigned long long cycles
)
4160 unsigned int idx
= rx
->intrvl_idx
++ % PMD_RXQ_INTERVAL_MAX
;
4161 atomic_store_relaxed(&rx
->cycles_intrvl
[idx
], cycles
);
4165 dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq
*rx
, unsigned idx
)
4167 unsigned long long processing_cycles
;
4168 atomic_read_relaxed(&rx
->cycles_intrvl
[idx
], &processing_cycles
);
4169 return processing_cycles
;
4172 #if ATOMIC_ALWAYS_LOCK_FREE_8B
4174 pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread
*pmd
)
4176 bool pmd_perf_enabled
;
4177 atomic_read_relaxed(&pmd
->dp
->pmd_perf_metrics
, &pmd_perf_enabled
);
4178 return pmd_perf_enabled
;
4181 /* If stores and reads of 64-bit integers are not atomic, the full PMD
4182 * performance metrics are not available as locked access to 64 bit
4183 * integers would be prohibitively expensive. */
4185 pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread
*pmd OVS_UNUSED
)
4192 dp_netdev_pmd_flush_output_on_port(struct dp_netdev_pmd_thread
*pmd
,
4199 struct cycle_timer timer
;
4201 uint32_t tx_flush_interval
;
4203 cycle_timer_start(&pmd
->perf_stats
, &timer
);
4205 dynamic_txqs
= p
->port
->dynamic_txqs
;
4207 tx_qid
= dpif_netdev_xps_get_tx_qid(pmd
, p
);
4209 tx_qid
= pmd
->static_tx_qid
;
4212 output_cnt
= dp_packet_batch_size(&p
->output_pkts
);
4213 ovs_assert(output_cnt
> 0);
4215 netdev_send(p
->port
->netdev
, tx_qid
, &p
->output_pkts
, dynamic_txqs
);
4216 dp_packet_batch_init(&p
->output_pkts
);
4218 /* Update time of the next flush. */
4219 atomic_read_relaxed(&pmd
->dp
->tx_flush_interval
, &tx_flush_interval
);
4220 p
->flush_time
= pmd
->ctx
.now
+ tx_flush_interval
;
4222 ovs_assert(pmd
->n_output_batches
> 0);
4223 pmd
->n_output_batches
--;
4225 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_SENT_PKTS
, output_cnt
);
4226 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_SENT_BATCHES
, 1);
4228 /* Distribute send cycles evenly among transmitted packets and assign to
4229 * their respective rx queues. */
4230 cycles
= cycle_timer_stop(&pmd
->perf_stats
, &timer
) / output_cnt
;
4231 for (i
= 0; i
< output_cnt
; i
++) {
4232 if (p
->output_pkts_rxqs
[i
]) {
4233 dp_netdev_rxq_add_cycles(p
->output_pkts_rxqs
[i
],
4234 RXQ_CYCLES_PROC_CURR
, cycles
);
4242 dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread
*pmd
,
4248 if (!pmd
->n_output_batches
) {
4252 HMAP_FOR_EACH (p
, node
, &pmd
->send_port_cache
) {
4253 if (!dp_packet_batch_is_empty(&p
->output_pkts
)
4254 && (force
|| pmd
->ctx
.now
>= p
->flush_time
)) {
4255 output_cnt
+= dp_netdev_pmd_flush_output_on_port(pmd
, p
);
4262 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread
*pmd
,
4263 struct dp_netdev_rxq
*rxq
,
4266 struct pmd_perf_stats
*s
= &pmd
->perf_stats
;
4267 struct dp_packet_batch batch
;
4268 struct cycle_timer timer
;
4271 int rem_qlen
= 0, *qlen_p
= NULL
;
4274 /* Measure duration for polling and processing rx burst. */
4275 cycle_timer_start(&pmd
->perf_stats
, &timer
);
4277 pmd
->ctx
.last_rxq
= rxq
;
4278 dp_packet_batch_init(&batch
);
4280 /* Fetch the rx queue length only for vhostuser ports. */
4281 if (pmd_perf_metrics_enabled(pmd
) && rxq
->is_vhost
) {
4285 error
= netdev_rxq_recv(rxq
->rx
, &batch
, qlen_p
);
4287 /* At least one packet received. */
4288 *recirc_depth_get() = 0;
4289 pmd_thread_ctx_time_update(pmd
);
4290 batch_cnt
= dp_packet_batch_size(&batch
);
4291 if (pmd_perf_metrics_enabled(pmd
)) {
4292 /* Update batch histogram. */
4293 s
->current
.batches
++;
4294 histogram_add_sample(&s
->pkts_per_batch
, batch_cnt
);
4295 /* Update the maximum vhost rx queue fill level. */
4296 if (rxq
->is_vhost
&& rem_qlen
>= 0) {
4297 uint32_t qfill
= batch_cnt
+ rem_qlen
;
4298 if (qfill
> s
->current
.max_vhost_qfill
) {
4299 s
->current
.max_vhost_qfill
= qfill
;
4303 /* Process packet batch. */
4304 dp_netdev_input(pmd
, &batch
, port_no
);
4306 /* Assign processing cycles to rx queue. */
4307 cycles
= cycle_timer_stop(&pmd
->perf_stats
, &timer
);
4308 dp_netdev_rxq_add_cycles(rxq
, RXQ_CYCLES_PROC_CURR
, cycles
);
4310 dp_netdev_pmd_flush_output_packets(pmd
, false);
4312 /* Discard cycles. */
4313 cycle_timer_stop(&pmd
->perf_stats
, &timer
);
4314 if (error
!= EAGAIN
&& error
!= EOPNOTSUPP
) {
4315 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
4317 VLOG_ERR_RL(&rl
, "error receiving data from %s: %s",
4318 netdev_rxq_get_name(rxq
->rx
), ovs_strerror(error
));
4322 pmd
->ctx
.last_rxq
= NULL
;
4327 static struct tx_port
*
4328 tx_port_lookup(const struct hmap
*hmap
, odp_port_t port_no
)
4332 HMAP_FOR_EACH_IN_BUCKET (tx
, node
, hash_port_no(port_no
), hmap
) {
4333 if (tx
->port
->port_no
== port_no
) {
4342 port_reconfigure(struct dp_netdev_port
*port
)
4344 struct netdev
*netdev
= port
->netdev
;
4347 /* Closes the existing 'rxq's. */
4348 for (i
= 0; i
< port
->n_rxq
; i
++) {
4349 netdev_rxq_close(port
->rxqs
[i
].rx
);
4350 port
->rxqs
[i
].rx
= NULL
;
4352 unsigned last_nrxq
= port
->n_rxq
;
4355 /* Allows 'netdev' to apply the pending configuration changes. */
4356 if (netdev_is_reconf_required(netdev
) || port
->need_reconfigure
) {
4357 err
= netdev_reconfigure(netdev
);
4358 if (err
&& (err
!= EOPNOTSUPP
)) {
4359 VLOG_ERR("Failed to set interface %s new configuration",
4360 netdev_get_name(netdev
));
4364 /* If the netdev_reconfigure() above succeeds, reopens the 'rxq's. */
4365 port
->rxqs
= xrealloc(port
->rxqs
,
4366 sizeof *port
->rxqs
* netdev_n_rxq(netdev
));
4367 /* Realloc 'used' counters for tx queues. */
4368 free(port
->txq_used
);
4369 port
->txq_used
= xcalloc(netdev_n_txq(netdev
), sizeof *port
->txq_used
);
4371 for (i
= 0; i
< netdev_n_rxq(netdev
); i
++) {
4372 bool new_queue
= i
>= last_nrxq
;
4374 memset(&port
->rxqs
[i
], 0, sizeof port
->rxqs
[i
]);
4377 port
->rxqs
[i
].port
= port
;
4378 port
->rxqs
[i
].is_vhost
= !strncmp(port
->type
, "dpdkvhost", 9);
4380 err
= netdev_rxq_open(netdev
, &port
->rxqs
[i
].rx
, i
);
4387 /* Parse affinity list to apply configuration for new queues. */
4388 dpif_netdev_port_set_rxq_affinity(port
, port
->rxq_affinity_list
);
4390 /* If reconfiguration was successful mark it as such, so we can use it */
4391 port
->need_reconfigure
= false;
4396 struct rr_numa_list
{
4397 struct hmap numas
; /* Contains 'struct rr_numa' */
4401 struct hmap_node node
;
4405 /* Non isolated pmds on numa node 'numa_id' */
4406 struct dp_netdev_pmd_thread
**pmds
;
4413 static struct rr_numa
*
4414 rr_numa_list_lookup(struct rr_numa_list
*rr
, int numa_id
)
4416 struct rr_numa
*numa
;
4418 HMAP_FOR_EACH_WITH_HASH (numa
, node
, hash_int(numa_id
, 0), &rr
->numas
) {
4419 if (numa
->numa_id
== numa_id
) {
4427 /* Returns the next node in numa list following 'numa' in round-robin fashion.
4428 * Returns first node if 'numa' is a null pointer or the last node in 'rr'.
4429 * Returns NULL if 'rr' numa list is empty. */
4430 static struct rr_numa
*
4431 rr_numa_list_next(struct rr_numa_list
*rr
, const struct rr_numa
*numa
)
4433 struct hmap_node
*node
= NULL
;
4436 node
= hmap_next(&rr
->numas
, &numa
->node
);
4439 node
= hmap_first(&rr
->numas
);
4442 return (node
) ? CONTAINER_OF(node
, struct rr_numa
, node
) : NULL
;
4446 rr_numa_list_populate(struct dp_netdev
*dp
, struct rr_numa_list
*rr
)
4448 struct dp_netdev_pmd_thread
*pmd
;
4449 struct rr_numa
*numa
;
4451 hmap_init(&rr
->numas
);
4453 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4454 if (pmd
->core_id
== NON_PMD_CORE_ID
|| pmd
->isolated
) {
4458 numa
= rr_numa_list_lookup(rr
, pmd
->numa_id
);
4460 numa
= xzalloc(sizeof *numa
);
4461 numa
->numa_id
= pmd
->numa_id
;
4462 hmap_insert(&rr
->numas
, &numa
->node
, hash_int(pmd
->numa_id
, 0));
4465 numa
->pmds
= xrealloc(numa
->pmds
, numa
->n_pmds
* sizeof *numa
->pmds
);
4466 numa
->pmds
[numa
->n_pmds
- 1] = pmd
;
4467 /* At least one pmd so initialise curr_idx and idx_inc. */
4468 numa
->cur_index
= 0;
4469 numa
->idx_inc
= true;
4474 * Returns the next pmd from the numa node.
4476 * If 'updown' is 'true' it will alternate between selecting the next pmd in
4477 * either an up or down walk, switching between up/down when the first or last
4478 * core is reached. e.g. 1,2,3,3,2,1,1,2...
4480 * If 'updown' is 'false' it will select the next pmd wrapping around when last
4481 * core reached. e.g. 1,2,3,1,2,3,1,2...
4483 static struct dp_netdev_pmd_thread
*
4484 rr_numa_get_pmd(struct rr_numa
*numa
, bool updown
)
4486 int numa_idx
= numa
->cur_index
;
4488 if (numa
->idx_inc
== true) {
4489 /* Incrementing through list of pmds. */
4490 if (numa
->cur_index
== numa
->n_pmds
-1) {
4491 /* Reached the last pmd. */
4493 numa
->idx_inc
= false;
4495 numa
->cur_index
= 0;
4501 /* Decrementing through list of pmds. */
4502 if (numa
->cur_index
== 0) {
4503 /* Reached the first pmd. */
4504 numa
->idx_inc
= true;
4509 return numa
->pmds
[numa_idx
];
4513 rr_numa_list_destroy(struct rr_numa_list
*rr
)
4515 struct rr_numa
*numa
;
4517 HMAP_FOR_EACH_POP (numa
, node
, &rr
->numas
) {
4521 hmap_destroy(&rr
->numas
);
4524 /* Sort Rx Queues by the processing cycles they are consuming. */
4526 compare_rxq_cycles(const void *a
, const void *b
)
4528 struct dp_netdev_rxq
*qa
;
4529 struct dp_netdev_rxq
*qb
;
4530 uint64_t cycles_qa
, cycles_qb
;
4532 qa
= *(struct dp_netdev_rxq
**) a
;
4533 qb
= *(struct dp_netdev_rxq
**) b
;
4535 cycles_qa
= dp_netdev_rxq_get_cycles(qa
, RXQ_CYCLES_PROC_HIST
);
4536 cycles_qb
= dp_netdev_rxq_get_cycles(qb
, RXQ_CYCLES_PROC_HIST
);
4538 if (cycles_qa
!= cycles_qb
) {
4539 return (cycles_qa
< cycles_qb
) ? 1 : -1;
4541 /* Cycles are the same so tiebreak on port/queue id.
4542 * Tiebreaking (as opposed to return 0) ensures consistent
4543 * sort results across multiple OS's. */
4544 uint32_t port_qa
= odp_to_u32(qa
->port
->port_no
);
4545 uint32_t port_qb
= odp_to_u32(qb
->port
->port_no
);
4546 if (port_qa
!= port_qb
) {
4547 return port_qa
> port_qb
? 1 : -1;
4549 return netdev_rxq_get_queue_id(qa
->rx
)
4550 - netdev_rxq_get_queue_id(qb
->rx
);
4555 /* Assign pmds to queues. If 'pinned' is true, assign pmds to pinned
4556 * queues and marks the pmds as isolated. Otherwise, assign non isolated
4557 * pmds to unpinned queues.
4559 * The function doesn't touch the pmd threads, it just stores the assignment
4560 * in the 'pmd' member of each rxq. */
4562 rxq_scheduling(struct dp_netdev
*dp
, bool pinned
) OVS_REQUIRES(dp
->port_mutex
)
4564 struct dp_netdev_port
*port
;
4565 struct rr_numa_list rr
;
4566 struct rr_numa
*non_local_numa
= NULL
;
4567 struct dp_netdev_rxq
** rxqs
= NULL
;
4569 struct rr_numa
*numa
= NULL
;
4571 bool assign_cyc
= dp
->pmd_rxq_assign_cyc
;
4573 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4574 if (!netdev_is_pmd(port
->netdev
)) {
4578 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
4579 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
4581 if (pinned
&& q
->core_id
!= OVS_CORE_UNSPEC
) {
4582 struct dp_netdev_pmd_thread
*pmd
;
4584 pmd
= dp_netdev_get_pmd(dp
, q
->core_id
);
4586 VLOG_WARN("There is no PMD thread on core %d. Queue "
4587 "%d on port \'%s\' will not be polled.",
4588 q
->core_id
, qid
, netdev_get_name(port
->netdev
));
4591 pmd
->isolated
= true;
4592 VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
4593 "rx queue %d.", pmd
->core_id
, pmd
->numa_id
,
4594 netdev_rxq_get_name(q
->rx
),
4595 netdev_rxq_get_queue_id(q
->rx
));
4596 dp_netdev_pmd_unref(pmd
);
4598 } else if (!pinned
&& q
->core_id
== OVS_CORE_UNSPEC
) {
4599 uint64_t cycle_hist
= 0;
4602 rxqs
= xmalloc(sizeof *rxqs
);
4604 rxqs
= xrealloc(rxqs
, sizeof *rxqs
* (n_rxqs
+ 1));
4608 /* Sum the queue intervals and store the cycle history. */
4609 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
4610 cycle_hist
+= dp_netdev_rxq_get_intrvl_cycles(q
, i
);
4612 dp_netdev_rxq_set_cycles(q
, RXQ_CYCLES_PROC_HIST
,
4615 /* Store the queue. */
4621 if (n_rxqs
> 1 && assign_cyc
) {
4622 /* Sort the queues in order of the processing cycles
4623 * they consumed during their last pmd interval. */
4624 qsort(rxqs
, n_rxqs
, sizeof *rxqs
, compare_rxq_cycles
);
4627 rr_numa_list_populate(dp
, &rr
);
4628 /* Assign the sorted queues to pmds in round robin. */
4629 for (int i
= 0; i
< n_rxqs
; i
++) {
4630 numa_id
= netdev_get_numa_id(rxqs
[i
]->port
->netdev
);
4631 numa
= rr_numa_list_lookup(&rr
, numa_id
);
4633 /* There are no pmds on the queue's local NUMA node.
4634 Round robin on the NUMA nodes that do have pmds. */
4635 non_local_numa
= rr_numa_list_next(&rr
, non_local_numa
);
4636 if (!non_local_numa
) {
4637 VLOG_ERR("There is no available (non-isolated) pmd "
4638 "thread for port \'%s\' queue %d. This queue "
4639 "will not be polled. Is pmd-cpu-mask set to "
4640 "zero? Or are all PMDs isolated to other "
4641 "queues?", netdev_rxq_get_name(rxqs
[i
]->rx
),
4642 netdev_rxq_get_queue_id(rxqs
[i
]->rx
));
4645 rxqs
[i
]->pmd
= rr_numa_get_pmd(non_local_numa
, assign_cyc
);
4646 VLOG_WARN("There's no available (non-isolated) pmd thread "
4647 "on numa node %d. Queue %d on port \'%s\' will "
4648 "be assigned to the pmd on core %d "
4649 "(numa node %d). Expect reduced performance.",
4650 numa_id
, netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
4651 netdev_rxq_get_name(rxqs
[i
]->rx
),
4652 rxqs
[i
]->pmd
->core_id
, rxqs
[i
]->pmd
->numa_id
);
4654 rxqs
[i
]->pmd
= rr_numa_get_pmd(numa
, assign_cyc
);
4656 VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
4658 "(measured processing cycles %"PRIu64
").",
4659 rxqs
[i
]->pmd
->core_id
, numa_id
,
4660 netdev_rxq_get_name(rxqs
[i
]->rx
),
4661 netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
4662 dp_netdev_rxq_get_cycles(rxqs
[i
],
4663 RXQ_CYCLES_PROC_HIST
));
4665 VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
4666 "rx queue %d.", rxqs
[i
]->pmd
->core_id
, numa_id
,
4667 netdev_rxq_get_name(rxqs
[i
]->rx
),
4668 netdev_rxq_get_queue_id(rxqs
[i
]->rx
));
4673 rr_numa_list_destroy(&rr
);
4678 reload_affected_pmds(struct dp_netdev
*dp
)
4680 struct dp_netdev_pmd_thread
*pmd
;
4682 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4683 if (pmd
->need_reload
) {
4684 flow_mark_flush(pmd
);
4685 dp_netdev_reload_pmd__(pmd
);
4689 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4690 if (pmd
->need_reload
) {
4691 if (pmd
->core_id
!= NON_PMD_CORE_ID
) {
4695 atomic_read_explicit(&pmd
->reload
, &reload
,
4696 memory_order_acquire
);
4699 pmd
->need_reload
= false;
4705 reconfigure_pmd_threads(struct dp_netdev
*dp
)
4706 OVS_REQUIRES(dp
->port_mutex
)
4708 struct dp_netdev_pmd_thread
*pmd
;
4709 struct ovs_numa_dump
*pmd_cores
;
4710 struct ovs_numa_info_core
*core
;
4711 struct hmapx to_delete
= HMAPX_INITIALIZER(&to_delete
);
4712 struct hmapx_node
*node
;
4713 bool changed
= false;
4714 bool need_to_adjust_static_tx_qids
= false;
4716 /* The pmd threads should be started only if there's a pmd port in the
4717 * datapath. If the user didn't provide any "pmd-cpu-mask", we start
4718 * NR_PMD_THREADS per numa node. */
4719 if (!has_pmd_port(dp
)) {
4720 pmd_cores
= ovs_numa_dump_n_cores_per_numa(0);
4721 } else if (dp
->pmd_cmask
&& dp
->pmd_cmask
[0]) {
4722 pmd_cores
= ovs_numa_dump_cores_with_cmask(dp
->pmd_cmask
);
4724 pmd_cores
= ovs_numa_dump_n_cores_per_numa(NR_PMD_THREADS
);
4727 /* We need to adjust 'static_tx_qid's only if we're reducing number of
4728 * PMD threads. Otherwise, new threads will allocate all the freed ids. */
4729 if (ovs_numa_dump_count(pmd_cores
) < cmap_count(&dp
->poll_threads
) - 1) {
4730 /* Adjustment is required to keep 'static_tx_qid's sequential and
4731 * avoid possible issues, for example, imbalanced tx queue usage
4732 * and unnecessary locking caused by remapping on netdev level. */
4733 need_to_adjust_static_tx_qids
= true;
4736 /* Check for unwanted pmd threads */
4737 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4738 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
4741 if (!ovs_numa_dump_contains_core(pmd_cores
, pmd
->numa_id
,
4743 hmapx_add(&to_delete
, pmd
);
4744 } else if (need_to_adjust_static_tx_qids
) {
4745 atomic_store_relaxed(&pmd
->reload_tx_qid
, true);
4746 pmd
->need_reload
= true;
4750 HMAPX_FOR_EACH (node
, &to_delete
) {
4751 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
4752 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d destroyed.",
4753 pmd
->numa_id
, pmd
->core_id
);
4754 dp_netdev_del_pmd(dp
, pmd
);
4756 changed
= !hmapx_is_empty(&to_delete
);
4757 hmapx_destroy(&to_delete
);
4759 if (need_to_adjust_static_tx_qids
) {
4760 /* 'static_tx_qid's are not sequential now.
4761 * Reload remaining threads to fix this. */
4762 reload_affected_pmds(dp
);
4765 /* Check for required new pmd threads */
4766 FOR_EACH_CORE_ON_DUMP(core
, pmd_cores
) {
4767 pmd
= dp_netdev_get_pmd(dp
, core
->core_id
);
4769 struct ds name
= DS_EMPTY_INITIALIZER
;
4771 pmd
= xzalloc(sizeof *pmd
);
4772 dp_netdev_configure_pmd(pmd
, dp
, core
->core_id
, core
->numa_id
);
4774 ds_put_format(&name
, "pmd-c%02d/id:", core
->core_id
);
4775 pmd
->thread
= ovs_thread_create(ds_cstr(&name
),
4776 pmd_thread_main
, pmd
);
4779 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d created.",
4780 pmd
->numa_id
, pmd
->core_id
);
4783 dp_netdev_pmd_unref(pmd
);
4788 struct ovs_numa_info_numa
*numa
;
4790 /* Log the number of pmd threads per numa node. */
4791 FOR_EACH_NUMA_ON_DUMP (numa
, pmd_cores
) {
4792 VLOG_INFO("There are %"PRIuSIZE
" pmd threads on numa node %d",
4793 numa
->n_cores
, numa
->numa_id
);
4797 ovs_numa_dump_destroy(pmd_cores
);
4801 pmd_remove_stale_ports(struct dp_netdev
*dp
,
4802 struct dp_netdev_pmd_thread
*pmd
)
4803 OVS_EXCLUDED(pmd
->port_mutex
)
4804 OVS_REQUIRES(dp
->port_mutex
)
4806 struct rxq_poll
*poll
, *poll_next
;
4807 struct tx_port
*tx
, *tx_next
;
4809 ovs_mutex_lock(&pmd
->port_mutex
);
4810 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
4811 struct dp_netdev_port
*port
= poll
->rxq
->port
;
4813 if (port
->need_reconfigure
4814 || !hmap_contains(&dp
->ports
, &port
->node
)) {
4815 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
4818 HMAP_FOR_EACH_SAFE (tx
, tx_next
, node
, &pmd
->tx_ports
) {
4819 struct dp_netdev_port
*port
= tx
->port
;
4821 if (port
->need_reconfigure
4822 || !hmap_contains(&dp
->ports
, &port
->node
)) {
4823 dp_netdev_del_port_tx_from_pmd(pmd
, tx
);
4826 ovs_mutex_unlock(&pmd
->port_mutex
);
4829 /* Must be called each time a port is added/removed or the cmask changes.
4830 * This creates and destroys pmd threads, reconfigures ports, opens their
4831 * rxqs and assigns all rxqs/txqs to pmd threads. */
4833 reconfigure_datapath(struct dp_netdev
*dp
)
4834 OVS_REQUIRES(dp
->port_mutex
)
4836 struct hmapx busy_threads
= HMAPX_INITIALIZER(&busy_threads
);
4837 struct dp_netdev_pmd_thread
*pmd
;
4838 struct dp_netdev_port
*port
;
4841 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
4843 /* Step 1: Adjust the pmd threads based on the datapath ports, the cores
4844 * on the system and the user configuration. */
4845 reconfigure_pmd_threads(dp
);
4847 wanted_txqs
= cmap_count(&dp
->poll_threads
);
4849 /* The number of pmd threads might have changed, or a port can be new:
4850 * adjust the txqs. */
4851 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4852 netdev_set_tx_multiq(port
->netdev
, wanted_txqs
);
4855 /* Step 2: Remove from the pmd threads ports that have been removed or
4856 * need reconfiguration. */
4858 /* Check for all the ports that need reconfiguration. We cache this in
4859 * 'port->need_reconfigure', because netdev_is_reconf_required() can
4860 * change at any time. */
4861 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4862 if (netdev_is_reconf_required(port
->netdev
)) {
4863 port
->need_reconfigure
= true;
4867 /* Remove from the pmd threads all the ports that have been deleted or
4868 * need reconfiguration. */
4869 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4870 pmd_remove_stale_ports(dp
, pmd
);
4873 /* Reload affected pmd threads. We must wait for the pmd threads before
4874 * reconfiguring the ports, because a port cannot be reconfigured while
4875 * it's being used. */
4876 reload_affected_pmds(dp
);
4878 /* Step 3: Reconfigure ports. */
4880 /* We only reconfigure the ports that we determined above, because they're
4881 * not being used by any pmd thread at the moment. If a port fails to
4882 * reconfigure we remove it from the datapath. */
4883 struct dp_netdev_port
*next_port
;
4884 HMAP_FOR_EACH_SAFE (port
, next_port
, node
, &dp
->ports
) {
4887 if (!port
->need_reconfigure
) {
4891 err
= port_reconfigure(port
);
4893 hmap_remove(&dp
->ports
, &port
->node
);
4894 seq_change(dp
->port_seq
);
4897 port
->dynamic_txqs
= netdev_n_txq(port
->netdev
) < wanted_txqs
;
4901 /* Step 4: Compute new rxq scheduling. We don't touch the pmd threads
4902 * for now, we just update the 'pmd' pointer in each rxq to point to the
4903 * wanted thread according to the scheduling policy. */
4905 /* Reset all the pmd threads to non isolated. */
4906 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4907 pmd
->isolated
= false;
4910 /* Reset all the queues to unassigned */
4911 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4912 for (int i
= 0; i
< port
->n_rxq
; i
++) {
4913 port
->rxqs
[i
].pmd
= NULL
;
4917 /* Add pinned queues and mark pmd threads isolated. */
4918 rxq_scheduling(dp
, true);
4920 /* Add non-pinned queues. */
4921 rxq_scheduling(dp
, false);
4923 /* Step 5: Remove queues not compliant with new scheduling. */
4925 /* Count all the threads that will have at least one queue to poll. */
4926 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4927 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
4928 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
4931 hmapx_add(&busy_threads
, q
->pmd
);
4936 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4937 struct rxq_poll
*poll
, *poll_next
;
4939 ovs_mutex_lock(&pmd
->port_mutex
);
4940 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
4941 if (poll
->rxq
->pmd
!= pmd
) {
4942 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
4944 /* This pmd might sleep after this step if it has no rxq
4945 * remaining. Tell it to busy wait for new assignment if it
4946 * has at least one scheduled queue. */
4947 if (hmap_count(&pmd
->poll_list
) == 0 &&
4948 hmapx_contains(&busy_threads
, pmd
)) {
4949 atomic_store_relaxed(&pmd
->wait_for_reload
, true);
4953 ovs_mutex_unlock(&pmd
->port_mutex
);
4956 hmapx_destroy(&busy_threads
);
4958 /* Reload affected pmd threads. We must wait for the pmd threads to remove
4959 * the old queues before readding them, otherwise a queue can be polled by
4960 * two threads at the same time. */
4961 reload_affected_pmds(dp
);
4963 /* Step 6: Add queues from scheduling, if they're not there already. */
4964 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4965 if (!netdev_is_pmd(port
->netdev
)) {
4969 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
4970 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
4973 ovs_mutex_lock(&q
->pmd
->port_mutex
);
4974 dp_netdev_add_rxq_to_pmd(q
->pmd
, q
);
4975 ovs_mutex_unlock(&q
->pmd
->port_mutex
);
4980 /* Add every port to the tx cache of every pmd thread, if it's not
4981 * there already and if this pmd has at least one rxq to poll. */
4982 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4983 ovs_mutex_lock(&pmd
->port_mutex
);
4984 if (hmap_count(&pmd
->poll_list
) || pmd
->core_id
== NON_PMD_CORE_ID
) {
4985 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4986 dp_netdev_add_port_tx_to_pmd(pmd
, port
);
4989 ovs_mutex_unlock(&pmd
->port_mutex
);
4992 /* Reload affected pmd threads. */
4993 reload_affected_pmds(dp
);
4995 /* Check if PMD Auto LB is to be enabled */
4996 set_pmd_auto_lb(dp
);
4999 /* Returns true if one of the netdevs in 'dp' requires a reconfiguration */
5001 ports_require_restart(const struct dp_netdev
*dp
)
5002 OVS_REQUIRES(dp
->port_mutex
)
5004 struct dp_netdev_port
*port
;
5006 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5007 if (netdev_is_reconf_required(port
->netdev
)) {
5015 /* Calculates variance in the values stored in array 'a'. 'n' is the number
5016 * of elements in array to be considered for calculating vairance.
5017 * Usage example: data array 'a' contains the processing load of each pmd and
5018 * 'n' is the number of PMDs. It returns the variance in processing load of
5021 variance(uint64_t a
[], int n
)
5023 /* Compute mean (average of elements). */
5026 uint64_t sqDiff
= 0;
5032 for (int i
= 0; i
< n
; i
++) {
5039 /* Compute sum squared differences with mean. */
5040 for (int i
= 0; i
< n
; i
++) {
5041 sqDiff
+= (a
[i
] - mean
)*(a
[i
] - mean
);
5044 return (sqDiff
? (sqDiff
/ n
) : 0);
5048 /* Returns the variance in the PMDs usage as part of dry run of rxqs
5049 * assignment to PMDs. */
5051 get_dry_run_variance(struct dp_netdev
*dp
, uint32_t *core_list
,
5052 uint32_t num_pmds
, uint64_t *predicted_variance
)
5053 OVS_REQUIRES(dp
->port_mutex
)
5055 struct dp_netdev_port
*port
;
5056 struct dp_netdev_pmd_thread
*pmd
;
5057 struct dp_netdev_rxq
**rxqs
= NULL
;
5058 struct rr_numa
*numa
= NULL
;
5059 struct rr_numa_list rr
;
5062 uint64_t *pmd_usage
;
5064 if (!predicted_variance
) {
5068 pmd_usage
= xcalloc(num_pmds
, sizeof(uint64_t));
5070 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5071 if (!netdev_is_pmd(port
->netdev
)) {
5075 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
5076 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
5077 uint64_t cycle_hist
= 0;
5079 if (q
->pmd
->isolated
) {
5084 rxqs
= xmalloc(sizeof *rxqs
);
5086 rxqs
= xrealloc(rxqs
, sizeof *rxqs
* (n_rxqs
+ 1));
5089 /* Sum the queue intervals and store the cycle history. */
5090 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
5091 cycle_hist
+= dp_netdev_rxq_get_intrvl_cycles(q
, i
);
5093 dp_netdev_rxq_set_cycles(q
, RXQ_CYCLES_PROC_HIST
,
5095 /* Store the queue. */
5100 /* Sort the queues in order of the processing cycles
5101 * they consumed during their last pmd interval. */
5102 qsort(rxqs
, n_rxqs
, sizeof *rxqs
, compare_rxq_cycles
);
5104 rr_numa_list_populate(dp
, &rr
);
5106 for (int i
= 0; i
< n_rxqs
; i
++) {
5107 int numa_id
= netdev_get_numa_id(rxqs
[i
]->port
->netdev
);
5108 numa
= rr_numa_list_lookup(&rr
, numa_id
);
5110 /* Abort if cross NUMA polling. */
5111 VLOG_DBG("PMD auto lb dry run."
5112 " Aborting due to cross-numa polling.");
5116 pmd
= rr_numa_get_pmd(numa
, true);
5117 VLOG_DBG("PMD auto lb dry run. Predicted: Core %d on numa node %d "
5118 "to be assigned port \'%s\' rx queue %d "
5119 "(measured processing cycles %"PRIu64
").",
5120 pmd
->core_id
, numa_id
,
5121 netdev_rxq_get_name(rxqs
[i
]->rx
),
5122 netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
5123 dp_netdev_rxq_get_cycles(rxqs
[i
], RXQ_CYCLES_PROC_HIST
));
5125 for (int id
= 0; id
< num_pmds
; id
++) {
5126 if (pmd
->core_id
== core_list
[id
]) {
5127 /* Add the processing cycles of rxq to pmd polling it. */
5128 pmd_usage
[id
] += dp_netdev_rxq_get_cycles(rxqs
[i
],
5129 RXQ_CYCLES_PROC_HIST
);
5134 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
5135 uint64_t total_cycles
= 0;
5137 if ((pmd
->core_id
== NON_PMD_CORE_ID
) || pmd
->isolated
) {
5141 /* Get the total pmd cycles for an interval. */
5142 atomic_read_relaxed(&pmd
->intrvl_cycles
, &total_cycles
);
5143 /* Estimate the cycles to cover all intervals. */
5144 total_cycles
*= PMD_RXQ_INTERVAL_MAX
;
5145 for (int id
= 0; id
< num_pmds
; id
++) {
5146 if (pmd
->core_id
== core_list
[id
]) {
5147 if (pmd_usage
[id
]) {
5148 pmd_usage
[id
] = (pmd_usage
[id
] * 100) / total_cycles
;
5150 VLOG_DBG("PMD auto lb dry run. Predicted: Core %d, "
5151 "usage %"PRIu64
"", pmd
->core_id
, pmd_usage
[id
]);
5155 *predicted_variance
= variance(pmd_usage
, num_pmds
);
5159 rr_numa_list_destroy(&rr
);
5165 /* Does the dry run of Rxq assignment to PMDs and returns true if it gives
5166 * better distribution of load on PMDs. */
5168 pmd_rebalance_dry_run(struct dp_netdev
*dp
)
5169 OVS_REQUIRES(dp
->port_mutex
)
5171 struct dp_netdev_pmd_thread
*pmd
;
5172 uint64_t *curr_pmd_usage
;
5174 uint64_t curr_variance
;
5175 uint64_t new_variance
;
5176 uint64_t improvement
= 0;
5178 uint32_t *pmd_corelist
;
5179 struct rxq_poll
*poll
;
5182 num_pmds
= cmap_count(&dp
->poll_threads
);
5185 curr_pmd_usage
= xcalloc(num_pmds
, sizeof(uint64_t));
5186 pmd_corelist
= xcalloc(num_pmds
, sizeof(uint32_t));
5192 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
5193 uint64_t total_cycles
= 0;
5194 uint64_t total_proc
= 0;
5196 if ((pmd
->core_id
== NON_PMD_CORE_ID
) || pmd
->isolated
) {
5200 /* Get the total pmd cycles for an interval. */
5201 atomic_read_relaxed(&pmd
->intrvl_cycles
, &total_cycles
);
5202 /* Estimate the cycles to cover all intervals. */
5203 total_cycles
*= PMD_RXQ_INTERVAL_MAX
;
5205 ovs_mutex_lock(&pmd
->port_mutex
);
5206 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
5207 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
5208 total_proc
+= dp_netdev_rxq_get_intrvl_cycles(poll
->rxq
, i
);
5211 ovs_mutex_unlock(&pmd
->port_mutex
);
5214 curr_pmd_usage
[num_pmds
] = (total_proc
* 100) / total_cycles
;
5217 VLOG_DBG("PMD auto lb dry run. Current: Core %d, usage %"PRIu64
"",
5218 pmd
->core_id
, curr_pmd_usage
[num_pmds
]);
5220 if (atomic_count_get(&pmd
->pmd_overloaded
)) {
5221 atomic_count_set(&pmd
->pmd_overloaded
, 0);
5224 pmd_corelist
[num_pmds
] = pmd
->core_id
;
5228 curr_variance
= variance(curr_pmd_usage
, num_pmds
);
5229 ret
= get_dry_run_variance(dp
, pmd_corelist
, num_pmds
, &new_variance
);
5232 VLOG_DBG("PMD auto lb dry run. Current PMD variance: %"PRIu64
","
5233 " Predicted PMD variance: %"PRIu64
"",
5234 curr_variance
, new_variance
);
5236 if (new_variance
< curr_variance
) {
5238 ((curr_variance
- new_variance
) * 100) / curr_variance
;
5240 if (improvement
< ALB_ACCEPTABLE_IMPROVEMENT
) {
5245 free(curr_pmd_usage
);
5251 /* Return true if needs to revalidate datapath flows. */
5253 dpif_netdev_run(struct dpif
*dpif
)
5255 struct dp_netdev_port
*port
;
5256 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5257 struct dp_netdev_pmd_thread
*non_pmd
;
5258 uint64_t new_tnl_seq
;
5259 bool need_to_flush
= true;
5260 bool pmd_rebalance
= false;
5261 long long int now
= time_msec();
5262 struct dp_netdev_pmd_thread
*pmd
;
5264 ovs_mutex_lock(&dp
->port_mutex
);
5265 non_pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
5267 ovs_mutex_lock(&dp
->non_pmd_mutex
);
5268 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5269 if (!netdev_is_pmd(port
->netdev
)) {
5272 if (port
->emc_enabled
) {
5273 atomic_read_relaxed(&dp
->emc_insert_min
,
5274 &non_pmd
->ctx
.emc_insert_min
);
5276 non_pmd
->ctx
.emc_insert_min
= 0;
5279 for (i
= 0; i
< port
->n_rxq
; i
++) {
5281 if (!netdev_rxq_enabled(port
->rxqs
[i
].rx
)) {
5285 if (dp_netdev_process_rxq_port(non_pmd
,
5288 need_to_flush
= false;
5293 if (need_to_flush
) {
5294 /* We didn't receive anything in the process loop.
5295 * Check if we need to send something.
5296 * There was no time updates on current iteration. */
5297 pmd_thread_ctx_time_update(non_pmd
);
5298 dp_netdev_pmd_flush_output_packets(non_pmd
, false);
5301 dpif_netdev_xps_revalidate_pmd(non_pmd
, false);
5302 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
5304 dp_netdev_pmd_unref(non_pmd
);
5307 struct pmd_auto_lb
*pmd_alb
= &dp
->pmd_alb
;
5308 if (pmd_alb
->is_enabled
) {
5309 if (!pmd_alb
->rebalance_poll_timer
) {
5310 pmd_alb
->rebalance_poll_timer
= now
;
5311 } else if ((pmd_alb
->rebalance_poll_timer
+
5312 pmd_alb
->rebalance_intvl
) < now
) {
5313 pmd_alb
->rebalance_poll_timer
= now
;
5314 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
5315 if (atomic_count_get(&pmd
->pmd_overloaded
) >=
5316 PMD_RXQ_INTERVAL_MAX
) {
5317 pmd_rebalance
= true;
5322 if (pmd_rebalance
&&
5323 !dp_netdev_is_reconf_required(dp
) &&
5324 !ports_require_restart(dp
) &&
5325 pmd_rebalance_dry_run(dp
)) {
5326 VLOG_INFO("PMD auto lb dry run."
5327 " requesting datapath reconfigure.");
5328 dp_netdev_request_reconfigure(dp
);
5333 if (dp_netdev_is_reconf_required(dp
) || ports_require_restart(dp
)) {
5334 reconfigure_datapath(dp
);
5336 ovs_mutex_unlock(&dp
->port_mutex
);
5338 tnl_neigh_cache_run();
5340 new_tnl_seq
= seq_read(tnl_conf_seq
);
5342 if (dp
->last_tnl_conf_seq
!= new_tnl_seq
) {
5343 dp
->last_tnl_conf_seq
= new_tnl_seq
;
5350 dpif_netdev_wait(struct dpif
*dpif
)
5352 struct dp_netdev_port
*port
;
5353 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5355 ovs_mutex_lock(&dp_netdev_mutex
);
5356 ovs_mutex_lock(&dp
->port_mutex
);
5357 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5358 netdev_wait_reconf_required(port
->netdev
);
5359 if (!netdev_is_pmd(port
->netdev
)) {
5362 for (i
= 0; i
< port
->n_rxq
; i
++) {
5363 netdev_rxq_wait(port
->rxqs
[i
].rx
);
5367 ovs_mutex_unlock(&dp
->port_mutex
);
5368 ovs_mutex_unlock(&dp_netdev_mutex
);
5369 seq_wait(tnl_conf_seq
, dp
->last_tnl_conf_seq
);
5373 pmd_free_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
5375 struct tx_port
*tx_port_cached
;
5377 /* Flush all the queued packets. */
5378 dp_netdev_pmd_flush_output_packets(pmd
, true);
5379 /* Free all used tx queue ids. */
5380 dpif_netdev_xps_revalidate_pmd(pmd
, true);
5382 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->tnl_port_cache
) {
5383 free(tx_port_cached
);
5385 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->send_port_cache
) {
5386 free(tx_port_cached
);
5390 /* Copies ports from 'pmd->tx_ports' (shared with the main thread) to
5391 * thread-local copies. Copy to 'pmd->tnl_port_cache' if it is a tunnel
5392 * device, otherwise to 'pmd->send_port_cache' if the port has at least
5395 pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
5396 OVS_REQUIRES(pmd
->port_mutex
)
5398 struct tx_port
*tx_port
, *tx_port_cached
;
5400 pmd_free_cached_ports(pmd
);
5401 hmap_shrink(&pmd
->send_port_cache
);
5402 hmap_shrink(&pmd
->tnl_port_cache
);
5404 HMAP_FOR_EACH (tx_port
, node
, &pmd
->tx_ports
) {
5405 if (netdev_has_tunnel_push_pop(tx_port
->port
->netdev
)) {
5406 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
5407 hmap_insert(&pmd
->tnl_port_cache
, &tx_port_cached
->node
,
5408 hash_port_no(tx_port_cached
->port
->port_no
));
5411 if (netdev_n_txq(tx_port
->port
->netdev
)) {
5412 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
5413 hmap_insert(&pmd
->send_port_cache
, &tx_port_cached
->node
,
5414 hash_port_no(tx_port_cached
->port
->port_no
));
5420 pmd_alloc_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
5422 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
5423 if (!id_pool_alloc_id(pmd
->dp
->tx_qid_pool
, &pmd
->static_tx_qid
)) {
5424 VLOG_ABORT("static_tx_qid allocation failed for PMD on core %2d"
5425 ", numa_id %d.", pmd
->core_id
, pmd
->numa_id
);
5427 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
5429 VLOG_DBG("static_tx_qid = %d allocated for PMD thread on core %2d"
5430 ", numa_id %d.", pmd
->static_tx_qid
, pmd
->core_id
, pmd
->numa_id
);
5434 pmd_free_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
5436 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
5437 id_pool_free_id(pmd
->dp
->tx_qid_pool
, pmd
->static_tx_qid
);
5438 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
5442 pmd_load_queues_and_ports(struct dp_netdev_pmd_thread
*pmd
,
5443 struct polled_queue
**ppoll_list
)
5445 struct polled_queue
*poll_list
= *ppoll_list
;
5446 struct rxq_poll
*poll
;
5449 ovs_mutex_lock(&pmd
->port_mutex
);
5450 poll_list
= xrealloc(poll_list
, hmap_count(&pmd
->poll_list
)
5451 * sizeof *poll_list
);
5454 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
5455 poll_list
[i
].rxq
= poll
->rxq
;
5456 poll_list
[i
].port_no
= poll
->rxq
->port
->port_no
;
5457 poll_list
[i
].emc_enabled
= poll
->rxq
->port
->emc_enabled
;
5458 poll_list
[i
].rxq_enabled
= netdev_rxq_enabled(poll
->rxq
->rx
);
5459 poll_list
[i
].change_seq
=
5460 netdev_get_change_seq(poll
->rxq
->port
->netdev
);
5464 pmd_load_cached_ports(pmd
);
5466 ovs_mutex_unlock(&pmd
->port_mutex
);
5468 *ppoll_list
= poll_list
;
5473 pmd_thread_main(void *f_
)
5475 struct dp_netdev_pmd_thread
*pmd
= f_
;
5476 struct pmd_perf_stats
*s
= &pmd
->perf_stats
;
5477 unsigned int lc
= 0;
5478 struct polled_queue
*poll_list
;
5479 bool wait_for_reload
= false;
5485 int process_packets
= 0;
5489 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
5490 ovsthread_setspecific(pmd
->dp
->per_pmd_key
, pmd
);
5491 ovs_numa_thread_setaffinity_core(pmd
->core_id
);
5492 dpdk_set_lcore_id(pmd
->core_id
);
5493 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
5494 dfc_cache_init(&pmd
->flow_cache
);
5495 pmd_alloc_static_tx_qid(pmd
);
5498 atomic_count_init(&pmd
->pmd_overloaded
, 0);
5500 /* List port/core affinity */
5501 for (i
= 0; i
< poll_cnt
; i
++) {
5502 VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n",
5503 pmd
->core_id
, netdev_rxq_get_name(poll_list
[i
].rxq
->rx
),
5504 netdev_rxq_get_queue_id(poll_list
[i
].rxq
->rx
));
5505 /* Reset the rxq current cycles counter. */
5506 dp_netdev_rxq_set_cycles(poll_list
[i
].rxq
, RXQ_CYCLES_PROC_CURR
, 0);
5510 if (wait_for_reload
) {
5511 /* Don't sleep, control thread will ask for a reload shortly. */
5513 atomic_read_explicit(&pmd
->reload
, &reload
,
5514 memory_order_acquire
);
5517 while (seq_read(pmd
->reload_seq
) == pmd
->last_reload_seq
) {
5518 seq_wait(pmd
->reload_seq
, pmd
->last_reload_seq
);
5524 pmd
->intrvl_tsc_prev
= 0;
5525 atomic_store_relaxed(&pmd
->intrvl_cycles
, 0);
5526 cycles_counter_update(s
);
5527 /* Protect pmd stats from external clearing while polling. */
5528 ovs_mutex_lock(&pmd
->perf_stats
.stats_mutex
);
5530 uint64_t rx_packets
= 0, tx_packets
= 0;
5532 pmd_perf_start_iteration(s
);
5534 for (i
= 0; i
< poll_cnt
; i
++) {
5536 if (!poll_list
[i
].rxq_enabled
) {
5540 if (poll_list
[i
].emc_enabled
) {
5541 atomic_read_relaxed(&pmd
->dp
->emc_insert_min
,
5542 &pmd
->ctx
.emc_insert_min
);
5544 pmd
->ctx
.emc_insert_min
= 0;
5548 dp_netdev_process_rxq_port(pmd
, poll_list
[i
].rxq
,
5549 poll_list
[i
].port_no
);
5550 rx_packets
+= process_packets
;
5554 /* We didn't receive anything in the process loop.
5555 * Check if we need to send something.
5556 * There was no time updates on current iteration. */
5557 pmd_thread_ctx_time_update(pmd
);
5558 tx_packets
= dp_netdev_pmd_flush_output_packets(pmd
, false);
5564 coverage_try_clear();
5565 dp_netdev_pmd_try_optimize(pmd
, poll_list
, poll_cnt
);
5566 if (!ovsrcu_try_quiesce()) {
5567 emc_cache_slow_sweep(&((pmd
->flow_cache
).emc_cache
));
5570 for (i
= 0; i
< poll_cnt
; i
++) {
5571 uint64_t current_seq
=
5572 netdev_get_change_seq(poll_list
[i
].rxq
->port
->netdev
);
5573 if (poll_list
[i
].change_seq
!= current_seq
) {
5574 poll_list
[i
].change_seq
= current_seq
;
5575 poll_list
[i
].rxq_enabled
=
5576 netdev_rxq_enabled(poll_list
[i
].rxq
->rx
);
5581 atomic_read_explicit(&pmd
->reload
, &reload
, memory_order_acquire
);
5582 if (OVS_UNLIKELY(reload
)) {
5586 pmd_perf_end_iteration(s
, rx_packets
, tx_packets
,
5587 pmd_perf_metrics_enabled(pmd
));
5589 ovs_mutex_unlock(&pmd
->perf_stats
.stats_mutex
);
5591 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
5592 atomic_read_relaxed(&pmd
->wait_for_reload
, &wait_for_reload
);
5593 atomic_read_relaxed(&pmd
->reload_tx_qid
, &reload_tx_qid
);
5594 atomic_read_relaxed(&pmd
->exit
, &exiting
);
5595 /* Signal here to make sure the pmd finishes
5596 * reloading the updated configuration. */
5597 dp_netdev_pmd_reload_done(pmd
);
5599 if (reload_tx_qid
) {
5600 pmd_free_static_tx_qid(pmd
);
5601 pmd_alloc_static_tx_qid(pmd
);
5608 pmd_free_static_tx_qid(pmd
);
5609 dfc_cache_uninit(&pmd
->flow_cache
);
5611 pmd_free_cached_ports(pmd
);
5616 dp_netdev_disable_upcall(struct dp_netdev
*dp
)
5617 OVS_ACQUIRES(dp
->upcall_rwlock
)
5619 fat_rwlock_wrlock(&dp
->upcall_rwlock
);
5625 dpif_netdev_meter_get_features(const struct dpif
* dpif OVS_UNUSED
,
5626 struct ofputil_meter_features
*features
)
5628 features
->max_meters
= MAX_METERS
;
5629 features
->band_types
= DP_SUPPORTED_METER_BAND_TYPES
;
5630 features
->capabilities
= DP_SUPPORTED_METER_FLAGS_MASK
;
5631 features
->max_bands
= MAX_BANDS
;
5632 features
->max_color
= 0;
5635 /* Applies the meter identified by 'meter_id' to 'packets_'. Packets
5636 * that exceed a band are dropped in-place. */
5638 dp_netdev_run_meter(struct dp_netdev
*dp
, struct dp_packet_batch
*packets_
,
5639 uint32_t meter_id
, long long int now
)
5641 struct dp_meter
*meter
;
5642 struct dp_meter_band
*band
;
5643 struct dp_packet
*packet
;
5644 long long int long_delta_t
; /* msec */
5645 uint32_t delta_t
; /* msec */
5646 const size_t cnt
= dp_packet_batch_size(packets_
);
5647 uint32_t bytes
, volume
;
5648 int exceeded_band
[NETDEV_MAX_BURST
];
5649 uint32_t exceeded_rate
[NETDEV_MAX_BURST
];
5650 int exceeded_pkt
= cnt
; /* First packet that exceeded a band rate. */
5652 if (meter_id
>= MAX_METERS
) {
5656 meter_lock(dp
, meter_id
);
5657 meter
= dp
->meters
[meter_id
];
5662 /* Initialize as negative values. */
5663 memset(exceeded_band
, 0xff, cnt
* sizeof *exceeded_band
);
5664 /* Initialize as zeroes. */
5665 memset(exceeded_rate
, 0, cnt
* sizeof *exceeded_rate
);
5667 /* All packets will hit the meter at the same time. */
5668 long_delta_t
= now
/ 1000 - meter
->used
/ 1000; /* msec */
5670 if (long_delta_t
< 0) {
5671 /* This condition means that we have several threads fighting for a
5672 meter lock, and the one who received the packets a bit later wins.
5673 Assuming that all racing threads received packets at the same time
5674 to avoid overflow. */
5678 /* Make sure delta_t will not be too large, so that bucket will not
5679 * wrap around below. */
5680 delta_t
= (long_delta_t
> (long long int)meter
->max_delta_t
)
5681 ? meter
->max_delta_t
: (uint32_t)long_delta_t
;
5683 /* Update meter stats. */
5685 meter
->packet_count
+= cnt
;
5687 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
5688 bytes
+= dp_packet_size(packet
);
5690 meter
->byte_count
+= bytes
;
5692 /* Meters can operate in terms of packets per second or kilobits per
5694 if (meter
->flags
& OFPMF13_PKTPS
) {
5695 /* Rate in packets/second, bucket 1/1000 packets. */
5696 /* msec * packets/sec = 1/1000 packets. */
5697 volume
= cnt
* 1000; /* Take 'cnt' packets from the bucket. */
5699 /* Rate in kbps, bucket in bits. */
5700 /* msec * kbps = bits */
5704 /* Update all bands and find the one hit with the highest rate for each
5705 * packet (if any). */
5706 for (int m
= 0; m
< meter
->n_bands
; ++m
) {
5707 band
= &meter
->bands
[m
];
5709 /* Update band's bucket. */
5710 band
->bucket
+= delta_t
* band
->up
.rate
;
5711 if (band
->bucket
> band
->up
.burst_size
) {
5712 band
->bucket
= band
->up
.burst_size
;
5715 /* Drain the bucket for all the packets, if possible. */
5716 if (band
->bucket
>= volume
) {
5717 band
->bucket
-= volume
;
5719 int band_exceeded_pkt
;
5721 /* Band limit hit, must process packet-by-packet. */
5722 if (meter
->flags
& OFPMF13_PKTPS
) {
5723 band_exceeded_pkt
= band
->bucket
/ 1000;
5724 band
->bucket
%= 1000; /* Remainder stays in bucket. */
5726 /* Update the exceeding band for each exceeding packet.
5727 * (Only one band will be fired by a packet, and that
5728 * can be different for each packet.) */
5729 for (int i
= band_exceeded_pkt
; i
< cnt
; i
++) {
5730 if (band
->up
.rate
> exceeded_rate
[i
]) {
5731 exceeded_rate
[i
] = band
->up
.rate
;
5732 exceeded_band
[i
] = m
;
5736 /* Packet sizes differ, must process one-by-one. */
5737 band_exceeded_pkt
= cnt
;
5738 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
5739 uint32_t bits
= dp_packet_size(packet
) * 8;
5741 if (band
->bucket
>= bits
) {
5742 band
->bucket
-= bits
;
5744 if (i
< band_exceeded_pkt
) {
5745 band_exceeded_pkt
= i
;
5747 /* Update the exceeding band for the exceeding packet.
5748 * (Only one band will be fired by a packet, and that
5749 * can be different for each packet.) */
5750 if (band
->up
.rate
> exceeded_rate
[i
]) {
5751 exceeded_rate
[i
] = band
->up
.rate
;
5752 exceeded_band
[i
] = m
;
5757 /* Remember the first exceeding packet. */
5758 if (exceeded_pkt
> band_exceeded_pkt
) {
5759 exceeded_pkt
= band_exceeded_pkt
;
5764 /* Fire the highest rate band exceeded by each packet, and drop
5765 * packets if needed. */
5767 DP_PACKET_BATCH_REFILL_FOR_EACH (j
, cnt
, packet
, packets_
) {
5768 if (exceeded_band
[j
] >= 0) {
5769 /* Meter drop packet. */
5770 band
= &meter
->bands
[exceeded_band
[j
]];
5771 band
->packet_count
+= 1;
5772 band
->byte_count
+= dp_packet_size(packet
);
5774 dp_packet_delete(packet
);
5776 /* Meter accepts packet. */
5777 dp_packet_batch_refill(packets_
, packet
, j
);
5781 meter_unlock(dp
, meter_id
);
5784 /* Meter set/get/del processing is still single-threaded. */
5786 dpif_netdev_meter_set(struct dpif
*dpif
, ofproto_meter_id meter_id
,
5787 struct ofputil_meter_config
*config
)
5789 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5790 uint32_t mid
= meter_id
.uint32
;
5791 struct dp_meter
*meter
;
5794 if (mid
>= MAX_METERS
) {
5795 return EFBIG
; /* Meter_id out of range. */
5798 if (config
->flags
& ~DP_SUPPORTED_METER_FLAGS_MASK
) {
5799 return EBADF
; /* Unsupported flags set */
5802 if (config
->n_bands
> MAX_BANDS
) {
5806 for (i
= 0; i
< config
->n_bands
; ++i
) {
5807 switch (config
->bands
[i
].type
) {
5811 return ENODEV
; /* Unsupported band type */
5815 /* Allocate meter */
5816 meter
= xzalloc(sizeof *meter
5817 + config
->n_bands
* sizeof(struct dp_meter_band
));
5819 meter
->flags
= config
->flags
;
5820 meter
->n_bands
= config
->n_bands
;
5821 meter
->max_delta_t
= 0;
5822 meter
->used
= time_usec();
5825 for (i
= 0; i
< config
->n_bands
; ++i
) {
5826 uint32_t band_max_delta_t
;
5828 /* Set burst size to a workable value if none specified. */
5829 if (config
->bands
[i
].burst_size
== 0) {
5830 config
->bands
[i
].burst_size
= config
->bands
[i
].rate
;
5833 meter
->bands
[i
].up
= config
->bands
[i
];
5834 /* Convert burst size to the bucket units: */
5835 /* pkts => 1/1000 packets, kilobits => bits. */
5836 meter
->bands
[i
].up
.burst_size
*= 1000;
5837 /* Initialize bucket to empty. */
5838 meter
->bands
[i
].bucket
= 0;
5840 /* Figure out max delta_t that is enough to fill any bucket. */
5842 = meter
->bands
[i
].up
.burst_size
/ meter
->bands
[i
].up
.rate
;
5843 if (band_max_delta_t
> meter
->max_delta_t
) {
5844 meter
->max_delta_t
= band_max_delta_t
;
5848 meter_lock(dp
, mid
);
5849 dp_delete_meter(dp
, mid
); /* Free existing meter, if any */
5850 dp
->meters
[mid
] = meter
;
5851 meter_unlock(dp
, mid
);
5857 dpif_netdev_meter_get(const struct dpif
*dpif
,
5858 ofproto_meter_id meter_id_
,
5859 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
5861 const struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5862 uint32_t meter_id
= meter_id_
.uint32
;
5865 if (meter_id
>= MAX_METERS
) {
5869 meter_lock(dp
, meter_id
);
5870 const struct dp_meter
*meter
= dp
->meters
[meter_id
];
5878 stats
->packet_in_count
= meter
->packet_count
;
5879 stats
->byte_in_count
= meter
->byte_count
;
5881 for (i
= 0; i
< n_bands
&& i
< meter
->n_bands
; ++i
) {
5882 stats
->bands
[i
].packet_count
= meter
->bands
[i
].packet_count
;
5883 stats
->bands
[i
].byte_count
= meter
->bands
[i
].byte_count
;
5890 meter_unlock(dp
, meter_id
);
5895 dpif_netdev_meter_del(struct dpif
*dpif
,
5896 ofproto_meter_id meter_id_
,
5897 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
5899 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5902 error
= dpif_netdev_meter_get(dpif
, meter_id_
, stats
, n_bands
);
5904 uint32_t meter_id
= meter_id_
.uint32
;
5906 meter_lock(dp
, meter_id
);
5907 dp_delete_meter(dp
, meter_id
);
5908 meter_unlock(dp
, meter_id
);
5915 dpif_netdev_disable_upcall(struct dpif
*dpif
)
5916 OVS_NO_THREAD_SAFETY_ANALYSIS
5918 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5919 dp_netdev_disable_upcall(dp
);
5923 dp_netdev_enable_upcall(struct dp_netdev
*dp
)
5924 OVS_RELEASES(dp
->upcall_rwlock
)
5926 fat_rwlock_unlock(&dp
->upcall_rwlock
);
5930 dpif_netdev_enable_upcall(struct dpif
*dpif
)
5931 OVS_NO_THREAD_SAFETY_ANALYSIS
5933 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5934 dp_netdev_enable_upcall(dp
);
5938 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
)
5940 atomic_store_relaxed(&pmd
->wait_for_reload
, false);
5941 atomic_store_relaxed(&pmd
->reload_tx_qid
, false);
5942 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
5943 atomic_store_explicit(&pmd
->reload
, false, memory_order_release
);
5946 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
5947 * the pointer if succeeds, otherwise, NULL (it can return NULL even if
5948 * 'core_id' is NON_PMD_CORE_ID).
5950 * Caller must unrefs the returned reference. */
5951 static struct dp_netdev_pmd_thread
*
5952 dp_netdev_get_pmd(struct dp_netdev
*dp
, unsigned core_id
)
5954 struct dp_netdev_pmd_thread
*pmd
;
5955 const struct cmap_node
*pnode
;
5957 pnode
= cmap_find(&dp
->poll_threads
, hash_int(core_id
, 0));
5961 pmd
= CONTAINER_OF(pnode
, struct dp_netdev_pmd_thread
, node
);
5963 return dp_netdev_pmd_try_ref(pmd
) ? pmd
: NULL
;
5966 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
5968 dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
5969 OVS_REQUIRES(dp
->port_mutex
)
5971 struct dp_netdev_pmd_thread
*non_pmd
;
5973 non_pmd
= xzalloc(sizeof *non_pmd
);
5974 dp_netdev_configure_pmd(non_pmd
, dp
, NON_PMD_CORE_ID
, OVS_NUMA_UNSPEC
);
5977 /* Caller must have valid pointer to 'pmd'. */
5979 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
)
5981 return ovs_refcount_try_ref_rcu(&pmd
->ref_cnt
);
5985 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
)
5987 if (pmd
&& ovs_refcount_unref(&pmd
->ref_cnt
) == 1) {
5988 ovsrcu_postpone(dp_netdev_destroy_pmd
, pmd
);
5992 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
5993 * fails, keeps checking for next node until reaching the end of cmap.
5995 * Caller must unrefs the returned reference. */
5996 static struct dp_netdev_pmd_thread
*
5997 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
)
5999 struct dp_netdev_pmd_thread
*next
;
6002 struct cmap_node
*node
;
6004 node
= cmap_next_position(&dp
->poll_threads
, pos
);
6005 next
= node
? CONTAINER_OF(node
, struct dp_netdev_pmd_thread
, node
)
6007 } while (next
&& !dp_netdev_pmd_try_ref(next
));
6012 /* Configures the 'pmd' based on the input argument. */
6014 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
, struct dp_netdev
*dp
,
6015 unsigned core_id
, int numa_id
)
6018 pmd
->core_id
= core_id
;
6019 pmd
->numa_id
= numa_id
;
6020 pmd
->need_reload
= false;
6021 pmd
->n_output_batches
= 0;
6023 ovs_refcount_init(&pmd
->ref_cnt
);
6024 atomic_init(&pmd
->exit
, false);
6025 pmd
->reload_seq
= seq_create();
6026 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
6027 atomic_init(&pmd
->reload
, false);
6028 ovs_mutex_init(&pmd
->flow_mutex
);
6029 ovs_mutex_init(&pmd
->port_mutex
);
6030 cmap_init(&pmd
->flow_table
);
6031 cmap_init(&pmd
->classifiers
);
6032 pmd
->ctx
.last_rxq
= NULL
;
6033 pmd_thread_ctx_time_update(pmd
);
6034 pmd
->next_optimization
= pmd
->ctx
.now
+ DPCLS_OPTIMIZATION_INTERVAL
;
6035 pmd
->rxq_next_cycle_store
= pmd
->ctx
.now
+ PMD_RXQ_INTERVAL_LEN
;
6036 hmap_init(&pmd
->poll_list
);
6037 hmap_init(&pmd
->tx_ports
);
6038 hmap_init(&pmd
->tnl_port_cache
);
6039 hmap_init(&pmd
->send_port_cache
);
6040 /* init the 'flow_cache' since there is no
6041 * actual thread created for NON_PMD_CORE_ID. */
6042 if (core_id
== NON_PMD_CORE_ID
) {
6043 dfc_cache_init(&pmd
->flow_cache
);
6044 pmd_alloc_static_tx_qid(pmd
);
6046 pmd_perf_stats_init(&pmd
->perf_stats
);
6047 cmap_insert(&dp
->poll_threads
, CONST_CAST(struct cmap_node
*, &pmd
->node
),
6048 hash_int(core_id
, 0));
6052 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
)
6056 dp_netdev_pmd_flow_flush(pmd
);
6057 hmap_destroy(&pmd
->send_port_cache
);
6058 hmap_destroy(&pmd
->tnl_port_cache
);
6059 hmap_destroy(&pmd
->tx_ports
);
6060 hmap_destroy(&pmd
->poll_list
);
6061 /* All flows (including their dpcls_rules) have been deleted already */
6062 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
6064 ovsrcu_postpone(free
, cls
);
6066 cmap_destroy(&pmd
->classifiers
);
6067 cmap_destroy(&pmd
->flow_table
);
6068 ovs_mutex_destroy(&pmd
->flow_mutex
);
6069 seq_destroy(pmd
->reload_seq
);
6070 ovs_mutex_destroy(&pmd
->port_mutex
);
6074 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
6075 * and unrefs the struct. */
6077 dp_netdev_del_pmd(struct dp_netdev
*dp
, struct dp_netdev_pmd_thread
*pmd
)
6079 /* NON_PMD_CORE_ID doesn't have a thread, so we don't have to synchronize,
6080 * but extra cleanup is necessary */
6081 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
6082 ovs_mutex_lock(&dp
->non_pmd_mutex
);
6083 dfc_cache_uninit(&pmd
->flow_cache
);
6084 pmd_free_cached_ports(pmd
);
6085 pmd_free_static_tx_qid(pmd
);
6086 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
6088 atomic_store_relaxed(&pmd
->exit
, true);
6089 dp_netdev_reload_pmd__(pmd
);
6090 xpthread_join(pmd
->thread
, NULL
);
6093 dp_netdev_pmd_clear_ports(pmd
);
6095 /* Purges the 'pmd''s flows after stopping the thread, but before
6096 * destroying the flows, so that the flow stats can be collected. */
6097 if (dp
->dp_purge_cb
) {
6098 dp
->dp_purge_cb(dp
->dp_purge_aux
, pmd
->core_id
);
6100 cmap_remove(&pmd
->dp
->poll_threads
, &pmd
->node
, hash_int(pmd
->core_id
, 0));
6101 dp_netdev_pmd_unref(pmd
);
6104 /* Destroys all pmd threads. If 'non_pmd' is true it also destroys the non pmd
6107 dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
, bool non_pmd
)
6109 struct dp_netdev_pmd_thread
*pmd
;
6110 struct dp_netdev_pmd_thread
**pmd_list
;
6111 size_t k
= 0, n_pmds
;
6113 n_pmds
= cmap_count(&dp
->poll_threads
);
6114 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
6116 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
6117 if (!non_pmd
&& pmd
->core_id
== NON_PMD_CORE_ID
) {
6120 /* We cannot call dp_netdev_del_pmd(), since it alters
6121 * 'dp->poll_threads' (while we're iterating it) and it
6123 ovs_assert(k
< n_pmds
);
6124 pmd_list
[k
++] = pmd
;
6127 for (size_t i
= 0; i
< k
; i
++) {
6128 dp_netdev_del_pmd(dp
, pmd_list
[i
]);
6133 /* Deletes all rx queues from pmd->poll_list and all the ports from
6136 dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
)
6138 struct rxq_poll
*poll
;
6139 struct tx_port
*port
;
6141 ovs_mutex_lock(&pmd
->port_mutex
);
6142 HMAP_FOR_EACH_POP (poll
, node
, &pmd
->poll_list
) {
6145 HMAP_FOR_EACH_POP (port
, node
, &pmd
->tx_ports
) {
6148 ovs_mutex_unlock(&pmd
->port_mutex
);
6151 /* Adds rx queue to poll_list of PMD thread, if it's not there already. */
6153 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
6154 struct dp_netdev_rxq
*rxq
)
6155 OVS_REQUIRES(pmd
->port_mutex
)
6157 int qid
= netdev_rxq_get_queue_id(rxq
->rx
);
6158 uint32_t hash
= hash_2words(odp_to_u32(rxq
->port
->port_no
), qid
);
6159 struct rxq_poll
*poll
;
6161 HMAP_FOR_EACH_WITH_HASH (poll
, node
, hash
, &pmd
->poll_list
) {
6162 if (poll
->rxq
== rxq
) {
6163 /* 'rxq' is already polled by this thread. Do nothing. */
6168 poll
= xmalloc(sizeof *poll
);
6170 hmap_insert(&pmd
->poll_list
, &poll
->node
, hash
);
6172 pmd
->need_reload
= true;
6175 /* Delete 'poll' from poll_list of PMD thread. */
6177 dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
6178 struct rxq_poll
*poll
)
6179 OVS_REQUIRES(pmd
->port_mutex
)
6181 hmap_remove(&pmd
->poll_list
, &poll
->node
);
6184 pmd
->need_reload
= true;
6187 /* Add 'port' to the tx port cache of 'pmd', which must be reloaded for the
6188 * changes to take effect. */
6190 dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
6191 struct dp_netdev_port
*port
)
6192 OVS_REQUIRES(pmd
->port_mutex
)
6196 tx
= tx_port_lookup(&pmd
->tx_ports
, port
->port_no
);
6198 /* 'port' is already on this thread tx cache. Do nothing. */
6202 tx
= xzalloc(sizeof *tx
);
6206 tx
->flush_time
= 0LL;
6207 dp_packet_batch_init(&tx
->output_pkts
);
6209 hmap_insert(&pmd
->tx_ports
, &tx
->node
, hash_port_no(tx
->port
->port_no
));
6210 pmd
->need_reload
= true;
6213 /* Del 'tx' from the tx port cache of 'pmd', which must be reloaded for the
6214 * changes to take effect. */
6216 dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
6218 OVS_REQUIRES(pmd
->port_mutex
)
6220 hmap_remove(&pmd
->tx_ports
, &tx
->node
);
6222 pmd
->need_reload
= true;
6226 dpif_netdev_get_datapath_version(void)
6228 return xstrdup("<built-in>");
6232 dp_netdev_flow_used(struct dp_netdev_flow
*netdev_flow
, int cnt
, int size
,
6233 uint16_t tcp_flags
, long long now
)
6237 atomic_store_relaxed(&netdev_flow
->stats
.used
, now
);
6238 non_atomic_ullong_add(&netdev_flow
->stats
.packet_count
, cnt
);
6239 non_atomic_ullong_add(&netdev_flow
->stats
.byte_count
, size
);
6240 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
6242 atomic_store_relaxed(&netdev_flow
->stats
.tcp_flags
, flags
);
6246 dp_netdev_upcall(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
*packet_
,
6247 struct flow
*flow
, struct flow_wildcards
*wc
, ovs_u128
*ufid
,
6248 enum dpif_upcall_type type
, const struct nlattr
*userdata
,
6249 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
6251 struct dp_netdev
*dp
= pmd
->dp
;
6253 if (OVS_UNLIKELY(!dp
->upcall_cb
)) {
6257 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl
))) {
6258 struct ds ds
= DS_EMPTY_INITIALIZER
;
6261 struct odp_flow_key_parms odp_parms
= {
6263 .mask
= wc
? &wc
->masks
: NULL
,
6264 .support
= dp_netdev_support
,
6267 ofpbuf_init(&key
, 0);
6268 odp_flow_key_from_flow(&odp_parms
, &key
);
6269 packet_str
= ofp_dp_packet_to_string(packet_
);
6271 odp_flow_key_format(key
.data
, key
.size
, &ds
);
6273 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp
->name
,
6274 dpif_upcall_type_to_string(type
), ds_cstr(&ds
), packet_str
);
6276 ofpbuf_uninit(&key
);
6282 return dp
->upcall_cb(packet_
, flow
, ufid
, pmd
->core_id
, type
, userdata
,
6283 actions
, wc
, put_actions
, dp
->upcall_aux
);
6286 static inline uint32_t
6287 dpif_netdev_packet_get_rss_hash_orig_pkt(struct dp_packet
*packet
,
6288 const struct miniflow
*mf
)
6292 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
6293 hash
= dp_packet_get_rss_hash(packet
);
6295 hash
= miniflow_hash_5tuple(mf
, 0);
6296 dp_packet_set_rss_hash(packet
, hash
);
6302 static inline uint32_t
6303 dpif_netdev_packet_get_rss_hash(struct dp_packet
*packet
,
6304 const struct miniflow
*mf
)
6306 uint32_t hash
, recirc_depth
;
6308 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
6309 hash
= dp_packet_get_rss_hash(packet
);
6311 hash
= miniflow_hash_5tuple(mf
, 0);
6312 dp_packet_set_rss_hash(packet
, hash
);
6315 /* The RSS hash must account for the recirculation depth to avoid
6316 * collisions in the exact match cache */
6317 recirc_depth
= *recirc_depth_get_unsafe();
6318 if (OVS_UNLIKELY(recirc_depth
)) {
6319 hash
= hash_finish(hash
, recirc_depth
);
6324 struct packet_batch_per_flow
{
6325 unsigned int byte_count
;
6327 struct dp_netdev_flow
*flow
;
6329 struct dp_packet_batch array
;
6333 packet_batch_per_flow_update(struct packet_batch_per_flow
*batch
,
6334 struct dp_packet
*packet
,
6337 batch
->byte_count
+= dp_packet_size(packet
);
6338 batch
->tcp_flags
|= tcp_flags
;
6339 dp_packet_batch_add(&batch
->array
, packet
);
6343 packet_batch_per_flow_init(struct packet_batch_per_flow
*batch
,
6344 struct dp_netdev_flow
*flow
)
6346 flow
->batch
= batch
;
6349 dp_packet_batch_init(&batch
->array
);
6350 batch
->byte_count
= 0;
6351 batch
->tcp_flags
= 0;
6355 packet_batch_per_flow_execute(struct packet_batch_per_flow
*batch
,
6356 struct dp_netdev_pmd_thread
*pmd
)
6358 struct dp_netdev_actions
*actions
;
6359 struct dp_netdev_flow
*flow
= batch
->flow
;
6361 dp_netdev_flow_used(flow
, dp_packet_batch_size(&batch
->array
),
6363 batch
->tcp_flags
, pmd
->ctx
.now
/ 1000);
6365 actions
= dp_netdev_flow_get_actions(flow
);
6367 dp_netdev_execute_actions(pmd
, &batch
->array
, true, &flow
->flow
,
6368 actions
->actions
, actions
->size
);
6372 dp_netdev_queue_batches(struct dp_packet
*pkt
,
6373 struct dp_netdev_flow
*flow
, uint16_t tcp_flags
,
6374 struct packet_batch_per_flow
*batches
,
6377 struct packet_batch_per_flow
*batch
= flow
->batch
;
6379 if (OVS_UNLIKELY(!batch
)) {
6380 batch
= &batches
[(*n_batches
)++];
6381 packet_batch_per_flow_init(batch
, flow
);
6384 packet_batch_per_flow_update(batch
, pkt
, tcp_flags
);
6388 packet_enqueue_to_flow_map(struct dp_packet
*packet
,
6389 struct dp_netdev_flow
*flow
,
6391 struct dp_packet_flow_map
*flow_map
,
6394 struct dp_packet_flow_map
*map
= &flow_map
[index
];
6396 map
->packet
= packet
;
6397 map
->tcp_flags
= tcp_flags
;
6400 /* SMC lookup function for a batch of packets.
6401 * By doing batching SMC lookup, we can use prefetch
6402 * to hide memory access latency.
6405 smc_lookup_batch(struct dp_netdev_pmd_thread
*pmd
,
6406 struct netdev_flow_key
*keys
,
6407 struct netdev_flow_key
**missed_keys
,
6408 struct dp_packet_batch
*packets_
,
6410 struct dp_packet_flow_map
*flow_map
,
6414 struct dp_packet
*packet
;
6415 size_t n_smc_hit
= 0, n_missed
= 0;
6416 struct dfc_cache
*cache
= &pmd
->flow_cache
;
6417 struct smc_cache
*smc_cache
= &cache
->smc_cache
;
6418 const struct cmap_node
*flow_node
;
6422 /* Prefetch buckets for all packets */
6423 for (i
= 0; i
< cnt
; i
++) {
6424 OVS_PREFETCH(&smc_cache
->buckets
[keys
[i
].hash
& SMC_MASK
]);
6427 DP_PACKET_BATCH_REFILL_FOR_EACH (i
, cnt
, packet
, packets_
) {
6428 struct dp_netdev_flow
*flow
= NULL
;
6429 flow_node
= smc_entry_get(pmd
, keys
[i
].hash
);
6431 /* Get the original order of this packet in received batch. */
6432 recv_idx
= index_map
[i
];
6434 if (OVS_LIKELY(flow_node
!= NULL
)) {
6435 CMAP_NODE_FOR_EACH (flow
, node
, flow_node
) {
6436 /* Since we dont have per-port megaflow to check the port
6437 * number, we need to verify that the input ports match. */
6438 if (OVS_LIKELY(dpcls_rule_matches_key(&flow
->cr
, &keys
[i
]) &&
6439 flow
->flow
.in_port
.odp_port
== packet
->md
.in_port
.odp_port
)) {
6440 tcp_flags
= miniflow_get_tcp_flags(&keys
[i
].mf
);
6442 /* SMC hit and emc miss, we insert into EMC */
6444 netdev_flow_key_size(miniflow_n_values(&keys
[i
].mf
));
6445 emc_probabilistic_insert(pmd
, &keys
[i
], flow
);
6446 /* Add these packets into the flow map in the same order
6449 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6450 flow_map
, recv_idx
);
6461 /* SMC missed. Group missed packets together at
6462 * the beginning of the 'packets' array. */
6463 dp_packet_batch_refill(packets_
, packet
, i
);
6465 /* Preserve the order of packet for flow batching. */
6466 index_map
[n_missed
] = recv_idx
;
6468 /* Put missed keys to the pointer arrays return to the caller */
6469 missed_keys
[n_missed
++] = &keys
[i
];
6472 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_SMC_HIT
, n_smc_hit
);
6475 /* Try to process all ('cnt') the 'packets' using only the datapath flow cache
6476 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
6477 * miniflow is copied into 'keys' and the packet pointer is moved at the
6478 * beginning of the 'packets' array. The pointers of missed keys are put in the
6479 * missed_keys pointer array for future processing.
6481 * The function returns the number of packets that needs to be processed in the
6482 * 'packets' array (they have been moved to the beginning of the vector).
6484 * For performance reasons a caller may choose not to initialize the metadata
6485 * in 'packets_'. If 'md_is_valid' is false, the metadata in 'packets'
6486 * is not valid and must be initialized by this function using 'port_no'.
6487 * If 'md_is_valid' is true, the metadata is already valid and 'port_no'
6490 static inline size_t
6491 dfc_processing(struct dp_netdev_pmd_thread
*pmd
,
6492 struct dp_packet_batch
*packets_
,
6493 struct netdev_flow_key
*keys
,
6494 struct netdev_flow_key
**missed_keys
,
6495 struct packet_batch_per_flow batches
[], size_t *n_batches
,
6496 struct dp_packet_flow_map
*flow_map
,
6497 size_t *n_flows
, uint8_t *index_map
,
6498 bool md_is_valid
, odp_port_t port_no
)
6500 struct netdev_flow_key
*key
= &keys
[0];
6501 size_t n_missed
= 0, n_emc_hit
= 0;
6502 struct dfc_cache
*cache
= &pmd
->flow_cache
;
6503 struct dp_packet
*packet
;
6504 const size_t cnt
= dp_packet_batch_size(packets_
);
6505 uint32_t cur_min
= pmd
->ctx
.emc_insert_min
;
6510 bool batch_enable
= true;
6512 atomic_read_relaxed(&pmd
->dp
->smc_enable_db
, &smc_enable_db
);
6513 pmd_perf_update_counter(&pmd
->perf_stats
,
6514 md_is_valid
? PMD_STAT_RECIRC
: PMD_STAT_RECV
,
6517 DP_PACKET_BATCH_REFILL_FOR_EACH (i
, cnt
, packet
, packets_
) {
6518 struct dp_netdev_flow
*flow
;
6521 if (OVS_UNLIKELY(dp_packet_size(packet
) < ETH_HEADER_LEN
)) {
6522 dp_packet_delete(packet
);
6527 struct dp_packet
**packets
= packets_
->packets
;
6528 /* Prefetch next packet data and metadata. */
6529 OVS_PREFETCH(dp_packet_data(packets
[i
+1]));
6530 pkt_metadata_prefetch_init(&packets
[i
+1]->md
);
6534 pkt_metadata_init(&packet
->md
, port_no
);
6537 if ((*recirc_depth_get() == 0) &&
6538 dp_packet_has_flow_mark(packet
, &mark
)) {
6539 flow
= mark_to_flow_find(pmd
, mark
);
6540 if (OVS_LIKELY(flow
)) {
6541 tcp_flags
= parse_tcp_flags(packet
);
6542 if (OVS_LIKELY(batch_enable
)) {
6543 dp_netdev_queue_batches(packet
, flow
, tcp_flags
, batches
,
6546 /* Flow batching should be performed only after fast-path
6547 * processing is also completed for packets with emc miss
6548 * or else it will result in reordering of packets with
6549 * same datapath flows. */
6550 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6551 flow_map
, map_cnt
++);
6557 miniflow_extract(packet
, &key
->mf
);
6558 key
->len
= 0; /* Not computed yet. */
6560 (md_is_valid
== false)
6561 ? dpif_netdev_packet_get_rss_hash_orig_pkt(packet
, &key
->mf
)
6562 : dpif_netdev_packet_get_rss_hash(packet
, &key
->mf
);
6564 /* If EMC is disabled skip emc_lookup */
6565 flow
= (cur_min
!= 0) ? emc_lookup(&cache
->emc_cache
, key
) : NULL
;
6566 if (OVS_LIKELY(flow
)) {
6567 tcp_flags
= miniflow_get_tcp_flags(&key
->mf
);
6569 if (OVS_LIKELY(batch_enable
)) {
6570 dp_netdev_queue_batches(packet
, flow
, tcp_flags
, batches
,
6573 /* Flow batching should be performed only after fast-path
6574 * processing is also completed for packets with emc miss
6575 * or else it will result in reordering of packets with
6576 * same datapath flows. */
6577 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6578 flow_map
, map_cnt
++);
6581 /* Exact match cache missed. Group missed packets together at
6582 * the beginning of the 'packets' array. */
6583 dp_packet_batch_refill(packets_
, packet
, i
);
6585 /* Preserve the order of packet for flow batching. */
6586 index_map
[n_missed
] = map_cnt
;
6587 flow_map
[map_cnt
++].flow
= NULL
;
6589 /* 'key[n_missed]' contains the key of the current packet and it
6590 * will be passed to SMC lookup. The next key should be extracted
6591 * to 'keys[n_missed + 1]'.
6592 * We also maintain a pointer array to keys missed both SMC and EMC
6593 * which will be returned to the caller for future processing. */
6594 missed_keys
[n_missed
] = key
;
6595 key
= &keys
[++n_missed
];
6597 /* Skip batching for subsequent packets to avoid reordering. */
6598 batch_enable
= false;
6601 /* Count of packets which are not flow batched. */
6604 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_EXACT_HIT
, n_emc_hit
);
6606 if (!smc_enable_db
) {
6607 return dp_packet_batch_size(packets_
);
6610 /* Packets miss EMC will do a batch lookup in SMC if enabled */
6611 smc_lookup_batch(pmd
, keys
, missed_keys
, packets_
,
6612 n_missed
, flow_map
, index_map
);
6614 return dp_packet_batch_size(packets_
);
6618 handle_packet_upcall(struct dp_netdev_pmd_thread
*pmd
,
6619 struct dp_packet
*packet
,
6620 const struct netdev_flow_key
*key
,
6621 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
6623 struct ofpbuf
*add_actions
;
6624 struct dp_packet_batch b
;
6628 uint64_t cycles
= cycles_counter_update(&pmd
->perf_stats
);
6630 match
.tun_md
.valid
= false;
6631 miniflow_expand(&key
->mf
, &match
.flow
);
6632 memset(&match
.wc
, 0, sizeof match
.wc
);
6634 ofpbuf_clear(actions
);
6635 ofpbuf_clear(put_actions
);
6637 dpif_flow_hash(pmd
->dp
->dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
6638 error
= dp_netdev_upcall(pmd
, packet
, &match
.flow
, &match
.wc
,
6639 &ufid
, DPIF_UC_MISS
, NULL
, actions
,
6641 if (OVS_UNLIKELY(error
&& error
!= ENOSPC
)) {
6642 dp_packet_delete(packet
);
6646 /* The Netlink encoding of datapath flow keys cannot express
6647 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
6648 * tag is interpreted as exact match on the fact that there is no
6649 * VLAN. Unless we refactor a lot of code that translates between
6650 * Netlink and struct flow representations, we have to do the same
6651 * here. This must be in sync with 'match' in dpif_netdev_flow_put(). */
6652 if (!match
.wc
.masks
.vlans
[0].tci
) {
6653 match
.wc
.masks
.vlans
[0].tci
= htons(0xffff);
6656 /* We can't allow the packet batching in the next loop to execute
6657 * the actions. Otherwise, if there are any slow path actions,
6658 * we'll send the packet up twice. */
6659 dp_packet_batch_init_packet(&b
, packet
);
6660 dp_netdev_execute_actions(pmd
, &b
, true, &match
.flow
,
6661 actions
->data
, actions
->size
);
6663 add_actions
= put_actions
->size
? put_actions
: actions
;
6664 if (OVS_LIKELY(error
!= ENOSPC
)) {
6665 struct dp_netdev_flow
*netdev_flow
;
6667 /* XXX: There's a race window where a flow covering this packet
6668 * could have already been installed since we last did the flow
6669 * lookup before upcall. This could be solved by moving the
6670 * mutex lock outside the loop, but that's an awful long time
6671 * to be locking revalidators out of making flow modifications. */
6672 ovs_mutex_lock(&pmd
->flow_mutex
);
6673 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
6674 if (OVS_LIKELY(!netdev_flow
)) {
6675 netdev_flow
= dp_netdev_flow_add(pmd
, &match
, &ufid
,
6679 ovs_mutex_unlock(&pmd
->flow_mutex
);
6680 uint32_t hash
= dp_netdev_flow_hash(&netdev_flow
->ufid
);
6681 smc_insert(pmd
, key
, hash
);
6682 emc_probabilistic_insert(pmd
, key
, netdev_flow
);
6684 if (pmd_perf_metrics_enabled(pmd
)) {
6685 /* Update upcall stats. */
6686 cycles
= cycles_counter_update(&pmd
->perf_stats
) - cycles
;
6687 struct pmd_perf_stats
*s
= &pmd
->perf_stats
;
6688 s
->current
.upcalls
++;
6689 s
->current
.upcall_cycles
+= cycles
;
6690 histogram_add_sample(&s
->cycles_per_upcall
, cycles
);
6696 fast_path_processing(struct dp_netdev_pmd_thread
*pmd
,
6697 struct dp_packet_batch
*packets_
,
6698 struct netdev_flow_key
**keys
,
6699 struct dp_packet_flow_map
*flow_map
,
6703 const size_t cnt
= dp_packet_batch_size(packets_
);
6704 #if !defined(__CHECKER__) && !defined(_WIN32)
6705 const size_t PKT_ARRAY_SIZE
= cnt
;
6707 /* Sparse or MSVC doesn't like variable length array. */
6708 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
6710 struct dp_packet
*packet
;
6712 struct dpcls_rule
*rules
[PKT_ARRAY_SIZE
];
6713 struct dp_netdev
*dp
= pmd
->dp
;
6714 int upcall_ok_cnt
= 0, upcall_fail_cnt
= 0;
6715 int lookup_cnt
= 0, add_lookup_cnt
;
6718 for (size_t i
= 0; i
< cnt
; i
++) {
6719 /* Key length is needed in all the cases, hash computed on demand. */
6720 keys
[i
]->len
= netdev_flow_key_size(miniflow_n_values(&keys
[i
]->mf
));
6722 /* Get the classifier for the in_port */
6723 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
6724 if (OVS_LIKELY(cls
)) {
6725 any_miss
= !dpcls_lookup(cls
, (const struct netdev_flow_key
**)keys
,
6726 rules
, cnt
, &lookup_cnt
);
6729 memset(rules
, 0, sizeof(rules
));
6731 if (OVS_UNLIKELY(any_miss
) && !fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
6732 uint64_t actions_stub
[512 / 8], slow_stub
[512 / 8];
6733 struct ofpbuf actions
, put_actions
;
6735 ofpbuf_use_stub(&actions
, actions_stub
, sizeof actions_stub
);
6736 ofpbuf_use_stub(&put_actions
, slow_stub
, sizeof slow_stub
);
6738 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
6739 struct dp_netdev_flow
*netdev_flow
;
6741 if (OVS_LIKELY(rules
[i
])) {
6745 /* It's possible that an earlier slow path execution installed
6746 * a rule covering this flow. In this case, it's a lot cheaper
6747 * to catch it here than execute a miss. */
6748 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, keys
[i
],
6751 lookup_cnt
+= add_lookup_cnt
;
6752 rules
[i
] = &netdev_flow
->cr
;
6756 int error
= handle_packet_upcall(pmd
, packet
, keys
[i
],
6757 &actions
, &put_actions
);
6759 if (OVS_UNLIKELY(error
)) {
6766 ofpbuf_uninit(&actions
);
6767 ofpbuf_uninit(&put_actions
);
6768 fat_rwlock_unlock(&dp
->upcall_rwlock
);
6769 } else if (OVS_UNLIKELY(any_miss
)) {
6770 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
6771 if (OVS_UNLIKELY(!rules
[i
])) {
6772 dp_packet_delete(packet
);
6778 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
6779 struct dp_netdev_flow
*flow
;
6780 /* Get the original order of this packet in received batch. */
6781 int recv_idx
= index_map
[i
];
6784 if (OVS_UNLIKELY(!rules
[i
])) {
6788 flow
= dp_netdev_flow_cast(rules
[i
]);
6789 uint32_t hash
= dp_netdev_flow_hash(&flow
->ufid
);
6790 smc_insert(pmd
, keys
[i
], hash
);
6792 emc_probabilistic_insert(pmd
, keys
[i
], flow
);
6793 /* Add these packets into the flow map in the same order
6796 tcp_flags
= miniflow_get_tcp_flags(&keys
[i
]->mf
);
6797 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6798 flow_map
, recv_idx
);
6801 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_MASKED_HIT
,
6802 cnt
- upcall_ok_cnt
- upcall_fail_cnt
);
6803 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_MASKED_LOOKUP
,
6805 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_MISS
,
6807 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_LOST
,
6811 /* Packets enter the datapath from a port (or from recirculation) here.
6813 * When 'md_is_valid' is true the metadata in 'packets' are already valid.
6814 * When false the metadata in 'packets' need to be initialized. */
6816 dp_netdev_input__(struct dp_netdev_pmd_thread
*pmd
,
6817 struct dp_packet_batch
*packets
,
6818 bool md_is_valid
, odp_port_t port_no
)
6820 #if !defined(__CHECKER__) && !defined(_WIN32)
6821 const size_t PKT_ARRAY_SIZE
= dp_packet_batch_size(packets
);
6823 /* Sparse or MSVC doesn't like variable length array. */
6824 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
6826 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
)
6827 struct netdev_flow_key keys
[PKT_ARRAY_SIZE
];
6828 struct netdev_flow_key
*missed_keys
[PKT_ARRAY_SIZE
];
6829 struct packet_batch_per_flow batches
[PKT_ARRAY_SIZE
];
6831 struct dp_packet_flow_map flow_map
[PKT_ARRAY_SIZE
];
6832 uint8_t index_map
[PKT_ARRAY_SIZE
];
6838 dfc_processing(pmd
, packets
, keys
, missed_keys
, batches
, &n_batches
,
6839 flow_map
, &n_flows
, index_map
, md_is_valid
, port_no
);
6841 if (!dp_packet_batch_is_empty(packets
)) {
6842 /* Get ingress port from first packet's metadata. */
6843 in_port
= packets
->packets
[0]->md
.in_port
.odp_port
;
6844 fast_path_processing(pmd
, packets
, missed_keys
,
6845 flow_map
, index_map
, in_port
);
6848 /* Batch rest of packets which are in flow map. */
6849 for (i
= 0; i
< n_flows
; i
++) {
6850 struct dp_packet_flow_map
*map
= &flow_map
[i
];
6852 if (OVS_UNLIKELY(!map
->flow
)) {
6855 dp_netdev_queue_batches(map
->packet
, map
->flow
, map
->tcp_flags
,
6856 batches
, &n_batches
);
6859 /* All the flow batches need to be reset before any call to
6860 * packet_batch_per_flow_execute() as it could potentially trigger
6861 * recirculation. When a packet matching flow ‘j’ happens to be
6862 * recirculated, the nested call to dp_netdev_input__() could potentially
6863 * classify the packet as matching another flow - say 'k'. It could happen
6864 * that in the previous call to dp_netdev_input__() that same flow 'k' had
6865 * already its own batches[k] still waiting to be served. So if its
6866 * ‘batch’ member is not reset, the recirculated packet would be wrongly
6867 * appended to batches[k] of the 1st call to dp_netdev_input__(). */
6868 for (i
= 0; i
< n_batches
; i
++) {
6869 batches
[i
].flow
->batch
= NULL
;
6872 for (i
= 0; i
< n_batches
; i
++) {
6873 packet_batch_per_flow_execute(&batches
[i
], pmd
);
6878 dp_netdev_input(struct dp_netdev_pmd_thread
*pmd
,
6879 struct dp_packet_batch
*packets
,
6882 dp_netdev_input__(pmd
, packets
, false, port_no
);
6886 dp_netdev_recirculate(struct dp_netdev_pmd_thread
*pmd
,
6887 struct dp_packet_batch
*packets
)
6889 dp_netdev_input__(pmd
, packets
, true, 0);
6892 struct dp_netdev_execute_aux
{
6893 struct dp_netdev_pmd_thread
*pmd
;
6894 const struct flow
*flow
;
6898 dpif_netdev_register_dp_purge_cb(struct dpif
*dpif
, dp_purge_callback
*cb
,
6901 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
6902 dp
->dp_purge_aux
= aux
;
6903 dp
->dp_purge_cb
= cb
;
6907 dpif_netdev_register_upcall_cb(struct dpif
*dpif
, upcall_callback
*cb
,
6910 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
6911 dp
->upcall_aux
= aux
;
6916 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
6920 struct dp_netdev_port
*port
;
6923 HMAP_FOR_EACH (tx
, node
, &pmd
->send_port_cache
) {
6924 if (!tx
->port
->dynamic_txqs
) {
6927 interval
= pmd
->ctx
.now
- tx
->last_used
;
6928 if (tx
->qid
>= 0 && (purge
|| interval
>= XPS_TIMEOUT
)) {
6930 ovs_mutex_lock(&port
->txq_used_mutex
);
6931 port
->txq_used
[tx
->qid
]--;
6932 ovs_mutex_unlock(&port
->txq_used_mutex
);
6939 dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
6942 struct dp_netdev_port
*port
;
6944 int i
, min_cnt
, min_qid
;
6946 interval
= pmd
->ctx
.now
- tx
->last_used
;
6947 tx
->last_used
= pmd
->ctx
.now
;
6949 if (OVS_LIKELY(tx
->qid
>= 0 && interval
< XPS_TIMEOUT
)) {
6955 ovs_mutex_lock(&port
->txq_used_mutex
);
6957 port
->txq_used
[tx
->qid
]--;
6963 for (i
= 0; i
< netdev_n_txq(port
->netdev
); i
++) {
6964 if (port
->txq_used
[i
] < min_cnt
|| min_cnt
== -1) {
6965 min_cnt
= port
->txq_used
[i
];
6970 port
->txq_used
[min_qid
]++;
6973 ovs_mutex_unlock(&port
->txq_used_mutex
);
6975 dpif_netdev_xps_revalidate_pmd(pmd
, false);
6977 VLOG_DBG("Core %d: New TX queue ID %d for port \'%s\'.",
6978 pmd
->core_id
, tx
->qid
, netdev_get_name(tx
->port
->netdev
));
6982 static struct tx_port
*
6983 pmd_tnl_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
6986 return tx_port_lookup(&pmd
->tnl_port_cache
, port_no
);
6989 static struct tx_port
*
6990 pmd_send_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
6993 return tx_port_lookup(&pmd
->send_port_cache
, port_no
);
6997 push_tnl_action(const struct dp_netdev_pmd_thread
*pmd
,
6998 const struct nlattr
*attr
,
6999 struct dp_packet_batch
*batch
)
7001 struct tx_port
*tun_port
;
7002 const struct ovs_action_push_tnl
*data
;
7005 data
= nl_attr_get(attr
);
7007 tun_port
= pmd_tnl_port_cache_lookup(pmd
, data
->tnl_port
);
7012 err
= netdev_push_header(tun_port
->port
->netdev
, batch
, data
);
7017 dp_packet_delete_batch(batch
, true);
7022 dp_execute_userspace_action(struct dp_netdev_pmd_thread
*pmd
,
7023 struct dp_packet
*packet
, bool should_steal
,
7024 struct flow
*flow
, ovs_u128
*ufid
,
7025 struct ofpbuf
*actions
,
7026 const struct nlattr
*userdata
)
7028 struct dp_packet_batch b
;
7031 ofpbuf_clear(actions
);
7033 error
= dp_netdev_upcall(pmd
, packet
, flow
, NULL
, ufid
,
7034 DPIF_UC_ACTION
, userdata
, actions
,
7036 if (!error
|| error
== ENOSPC
) {
7037 dp_packet_batch_init_packet(&b
, packet
);
7038 dp_netdev_execute_actions(pmd
, &b
, should_steal
, flow
,
7039 actions
->data
, actions
->size
);
7040 } else if (should_steal
) {
7041 dp_packet_delete(packet
);
7046 dp_execute_cb(void *aux_
, struct dp_packet_batch
*packets_
,
7047 const struct nlattr
*a
, bool should_steal
)
7048 OVS_NO_THREAD_SAFETY_ANALYSIS
7050 struct dp_netdev_execute_aux
*aux
= aux_
;
7051 uint32_t *depth
= recirc_depth_get();
7052 struct dp_netdev_pmd_thread
*pmd
= aux
->pmd
;
7053 struct dp_netdev
*dp
= pmd
->dp
;
7054 int type
= nl_attr_type(a
);
7057 switch ((enum ovs_action_attr
)type
) {
7058 case OVS_ACTION_ATTR_OUTPUT
:
7059 p
= pmd_send_port_cache_lookup(pmd
, nl_attr_get_odp_port(a
));
7060 if (OVS_LIKELY(p
)) {
7061 struct dp_packet
*packet
;
7062 struct dp_packet_batch out
;
7064 if (!should_steal
) {
7065 dp_packet_batch_clone(&out
, packets_
);
7066 dp_packet_batch_reset_cutlen(packets_
);
7069 dp_packet_batch_apply_cutlen(packets_
);
7072 if (OVS_UNLIKELY(!dp_packet_batch_is_empty(&p
->output_pkts
)
7073 && packets_
->packets
[0]->source
7074 != p
->output_pkts
.packets
[0]->source
)) {
7075 /* XXX: netdev-dpdk assumes that all packets in a single
7076 * output batch has the same source. Flush here to
7077 * avoid memory access issues. */
7078 dp_netdev_pmd_flush_output_on_port(pmd
, p
);
7081 if (dp_packet_batch_size(&p
->output_pkts
)
7082 + dp_packet_batch_size(packets_
) > NETDEV_MAX_BURST
) {
7083 /* Flush here to avoid overflow. */
7084 dp_netdev_pmd_flush_output_on_port(pmd
, p
);
7087 if (dp_packet_batch_is_empty(&p
->output_pkts
)) {
7088 pmd
->n_output_batches
++;
7091 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7092 p
->output_pkts_rxqs
[dp_packet_batch_size(&p
->output_pkts
)] =
7094 dp_packet_batch_add(&p
->output_pkts
, packet
);
7100 case OVS_ACTION_ATTR_TUNNEL_PUSH
:
7102 /* We're requested to push tunnel header, but also we need to take
7103 * the ownership of these packets. Thus, we can avoid performing
7104 * the action, because the caller will not use the result anyway.
7105 * Just break to free the batch. */
7108 dp_packet_batch_apply_cutlen(packets_
);
7109 push_tnl_action(pmd
, a
, packets_
);
7112 case OVS_ACTION_ATTR_TUNNEL_POP
:
7113 if (*depth
< MAX_RECIRC_DEPTH
) {
7114 struct dp_packet_batch
*orig_packets_
= packets_
;
7115 odp_port_t portno
= nl_attr_get_odp_port(a
);
7117 p
= pmd_tnl_port_cache_lookup(pmd
, portno
);
7119 struct dp_packet_batch tnl_pkt
;
7121 if (!should_steal
) {
7122 dp_packet_batch_clone(&tnl_pkt
, packets_
);
7123 packets_
= &tnl_pkt
;
7124 dp_packet_batch_reset_cutlen(orig_packets_
);
7127 dp_packet_batch_apply_cutlen(packets_
);
7129 netdev_pop_header(p
->port
->netdev
, packets_
);
7130 if (dp_packet_batch_is_empty(packets_
)) {
7134 struct dp_packet
*packet
;
7135 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7136 packet
->md
.in_port
.odp_port
= portno
;
7140 dp_netdev_recirculate(pmd
, packets_
);
7147 case OVS_ACTION_ATTR_USERSPACE
:
7148 if (!fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
7149 struct dp_packet_batch
*orig_packets_
= packets_
;
7150 const struct nlattr
*userdata
;
7151 struct dp_packet_batch usr_pkt
;
7152 struct ofpbuf actions
;
7157 userdata
= nl_attr_find_nested(a
, OVS_USERSPACE_ATTR_USERDATA
);
7158 ofpbuf_init(&actions
, 0);
7160 if (packets_
->trunc
) {
7161 if (!should_steal
) {
7162 dp_packet_batch_clone(&usr_pkt
, packets_
);
7163 packets_
= &usr_pkt
;
7165 dp_packet_batch_reset_cutlen(orig_packets_
);
7168 dp_packet_batch_apply_cutlen(packets_
);
7171 struct dp_packet
*packet
;
7172 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7173 flow_extract(packet
, &flow
);
7174 dpif_flow_hash(dp
->dpif
, &flow
, sizeof flow
, &ufid
);
7175 dp_execute_userspace_action(pmd
, packet
, should_steal
, &flow
,
7176 &ufid
, &actions
, userdata
);
7180 dp_packet_delete_batch(packets_
, true);
7183 ofpbuf_uninit(&actions
);
7184 fat_rwlock_unlock(&dp
->upcall_rwlock
);
7190 case OVS_ACTION_ATTR_RECIRC
:
7191 if (*depth
< MAX_RECIRC_DEPTH
) {
7192 struct dp_packet_batch recirc_pkts
;
7194 if (!should_steal
) {
7195 dp_packet_batch_clone(&recirc_pkts
, packets_
);
7196 packets_
= &recirc_pkts
;
7199 struct dp_packet
*packet
;
7200 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7201 packet
->md
.recirc_id
= nl_attr_get_u32(a
);
7205 dp_netdev_recirculate(pmd
, packets_
);
7211 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
7214 case OVS_ACTION_ATTR_CT
: {
7215 const struct nlattr
*b
;
7217 bool commit
= false;
7220 const char *helper
= NULL
;
7221 const uint32_t *setmark
= NULL
;
7222 const struct ovs_key_ct_labels
*setlabel
= NULL
;
7223 struct nat_action_info_t nat_action_info
;
7224 struct nat_action_info_t
*nat_action_info_ref
= NULL
;
7225 bool nat_config
= false;
7227 NL_ATTR_FOR_EACH_UNSAFE (b
, left
, nl_attr_get(a
),
7228 nl_attr_get_size(a
)) {
7229 enum ovs_ct_attr sub_type
= nl_attr_type(b
);
7232 case OVS_CT_ATTR_FORCE_COMMIT
:
7235 case OVS_CT_ATTR_COMMIT
:
7238 case OVS_CT_ATTR_ZONE
:
7239 zone
= nl_attr_get_u16(b
);
7241 case OVS_CT_ATTR_HELPER
:
7242 helper
= nl_attr_get_string(b
);
7244 case OVS_CT_ATTR_MARK
:
7245 setmark
= nl_attr_get(b
);
7247 case OVS_CT_ATTR_LABELS
:
7248 setlabel
= nl_attr_get(b
);
7250 case OVS_CT_ATTR_EVENTMASK
:
7251 /* Silently ignored, as userspace datapath does not generate
7252 * netlink events. */
7254 case OVS_CT_ATTR_TIMEOUT
:
7255 /* Userspace datapath does not support customized timeout
7258 case OVS_CT_ATTR_NAT
: {
7259 const struct nlattr
*b_nest
;
7260 unsigned int left_nest
;
7261 bool ip_min_specified
= false;
7262 bool proto_num_min_specified
= false;
7263 bool ip_max_specified
= false;
7264 bool proto_num_max_specified
= false;
7265 memset(&nat_action_info
, 0, sizeof nat_action_info
);
7266 nat_action_info_ref
= &nat_action_info
;
7268 NL_NESTED_FOR_EACH_UNSAFE (b_nest
, left_nest
, b
) {
7269 enum ovs_nat_attr sub_type_nest
= nl_attr_type(b_nest
);
7271 switch (sub_type_nest
) {
7272 case OVS_NAT_ATTR_SRC
:
7273 case OVS_NAT_ATTR_DST
:
7275 nat_action_info
.nat_action
|=
7276 ((sub_type_nest
== OVS_NAT_ATTR_SRC
)
7277 ? NAT_ACTION_SRC
: NAT_ACTION_DST
);
7279 case OVS_NAT_ATTR_IP_MIN
:
7280 memcpy(&nat_action_info
.min_addr
,
7281 nl_attr_get(b_nest
),
7282 nl_attr_get_size(b_nest
));
7283 ip_min_specified
= true;
7285 case OVS_NAT_ATTR_IP_MAX
:
7286 memcpy(&nat_action_info
.max_addr
,
7287 nl_attr_get(b_nest
),
7288 nl_attr_get_size(b_nest
));
7289 ip_max_specified
= true;
7291 case OVS_NAT_ATTR_PROTO_MIN
:
7292 nat_action_info
.min_port
=
7293 nl_attr_get_u16(b_nest
);
7294 proto_num_min_specified
= true;
7296 case OVS_NAT_ATTR_PROTO_MAX
:
7297 nat_action_info
.max_port
=
7298 nl_attr_get_u16(b_nest
);
7299 proto_num_max_specified
= true;
7301 case OVS_NAT_ATTR_PERSISTENT
:
7302 case OVS_NAT_ATTR_PROTO_HASH
:
7303 case OVS_NAT_ATTR_PROTO_RANDOM
:
7305 case OVS_NAT_ATTR_UNSPEC
:
7306 case __OVS_NAT_ATTR_MAX
:
7311 if (ip_min_specified
&& !ip_max_specified
) {
7312 nat_action_info
.max_addr
= nat_action_info
.min_addr
;
7314 if (proto_num_min_specified
&& !proto_num_max_specified
) {
7315 nat_action_info
.max_port
= nat_action_info
.min_port
;
7317 if (proto_num_min_specified
|| proto_num_max_specified
) {
7318 if (nat_action_info
.nat_action
& NAT_ACTION_SRC
) {
7319 nat_action_info
.nat_action
|= NAT_ACTION_SRC_PORT
;
7320 } else if (nat_action_info
.nat_action
& NAT_ACTION_DST
) {
7321 nat_action_info
.nat_action
|= NAT_ACTION_DST_PORT
;
7326 case OVS_CT_ATTR_UNSPEC
:
7327 case __OVS_CT_ATTR_MAX
:
7332 /* We won't be able to function properly in this case, hence
7333 * complain loudly. */
7334 if (nat_config
&& !commit
) {
7335 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(5, 5);
7336 VLOG_WARN_RL(&rl
, "NAT specified without commit.");
7339 conntrack_execute(dp
->conntrack
, packets_
, aux
->flow
->dl_type
, force
,
7340 commit
, zone
, setmark
, setlabel
, aux
->flow
->tp_src
,
7341 aux
->flow
->tp_dst
, helper
, nat_action_info_ref
,
7342 pmd
->ctx
.now
/ 1000);
7346 case OVS_ACTION_ATTR_METER
:
7347 dp_netdev_run_meter(pmd
->dp
, packets_
, nl_attr_get_u32(a
),
7351 case OVS_ACTION_ATTR_PUSH_VLAN
:
7352 case OVS_ACTION_ATTR_POP_VLAN
:
7353 case OVS_ACTION_ATTR_PUSH_MPLS
:
7354 case OVS_ACTION_ATTR_POP_MPLS
:
7355 case OVS_ACTION_ATTR_SET
:
7356 case OVS_ACTION_ATTR_SET_MASKED
:
7357 case OVS_ACTION_ATTR_SAMPLE
:
7358 case OVS_ACTION_ATTR_HASH
:
7359 case OVS_ACTION_ATTR_UNSPEC
:
7360 case OVS_ACTION_ATTR_TRUNC
:
7361 case OVS_ACTION_ATTR_PUSH_ETH
:
7362 case OVS_ACTION_ATTR_POP_ETH
:
7363 case OVS_ACTION_ATTR_CLONE
:
7364 case OVS_ACTION_ATTR_PUSH_NSH
:
7365 case OVS_ACTION_ATTR_POP_NSH
:
7366 case OVS_ACTION_ATTR_CT_CLEAR
:
7367 case OVS_ACTION_ATTR_CHECK_PKT_LEN
:
7368 case __OVS_ACTION_ATTR_MAX
:
7372 dp_packet_delete_batch(packets_
, should_steal
);
7376 dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
7377 struct dp_packet_batch
*packets
,
7378 bool should_steal
, const struct flow
*flow
,
7379 const struct nlattr
*actions
, size_t actions_len
)
7381 struct dp_netdev_execute_aux aux
= { pmd
, flow
};
7383 odp_execute_actions(&aux
, packets
, should_steal
, actions
,
7384 actions_len
, dp_execute_cb
);
7387 struct dp_netdev_ct_dump
{
7388 struct ct_dpif_dump_state up
;
7389 struct conntrack_dump dump
;
7390 struct conntrack
*ct
;
7391 struct dp_netdev
*dp
;
7395 dpif_netdev_ct_dump_start(struct dpif
*dpif
, struct ct_dpif_dump_state
**dump_
,
7396 const uint16_t *pzone
, int *ptot_bkts
)
7398 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7399 struct dp_netdev_ct_dump
*dump
;
7401 dump
= xzalloc(sizeof *dump
);
7403 dump
->ct
= dp
->conntrack
;
7405 conntrack_dump_start(dp
->conntrack
, &dump
->dump
, pzone
, ptot_bkts
);
7413 dpif_netdev_ct_dump_next(struct dpif
*dpif OVS_UNUSED
,
7414 struct ct_dpif_dump_state
*dump_
,
7415 struct ct_dpif_entry
*entry
)
7417 struct dp_netdev_ct_dump
*dump
;
7419 INIT_CONTAINER(dump
, dump_
, up
);
7421 return conntrack_dump_next(&dump
->dump
, entry
);
7425 dpif_netdev_ct_dump_done(struct dpif
*dpif OVS_UNUSED
,
7426 struct ct_dpif_dump_state
*dump_
)
7428 struct dp_netdev_ct_dump
*dump
;
7431 INIT_CONTAINER(dump
, dump_
, up
);
7433 err
= conntrack_dump_done(&dump
->dump
);
7441 dpif_netdev_ct_flush(struct dpif
*dpif
, const uint16_t *zone
,
7442 const struct ct_dpif_tuple
*tuple
)
7444 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7447 return conntrack_flush_tuple(dp
->conntrack
, tuple
, zone
? *zone
: 0);
7449 return conntrack_flush(dp
->conntrack
, zone
);
7453 dpif_netdev_ct_set_maxconns(struct dpif
*dpif
, uint32_t maxconns
)
7455 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7457 return conntrack_set_maxconns(dp
->conntrack
, maxconns
);
7461 dpif_netdev_ct_get_maxconns(struct dpif
*dpif
, uint32_t *maxconns
)
7463 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7465 return conntrack_get_maxconns(dp
->conntrack
, maxconns
);
7469 dpif_netdev_ct_get_nconns(struct dpif
*dpif
, uint32_t *nconns
)
7471 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7473 return conntrack_get_nconns(dp
->conntrack
, nconns
);
7477 dpif_netdev_ct_set_tcp_seq_chk(struct dpif
*dpif
, bool enabled
)
7479 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7481 return conntrack_set_tcp_seq_chk(dp
->conntrack
, enabled
);
7485 dpif_netdev_ct_get_tcp_seq_chk(struct dpif
*dpif
, bool *enabled
)
7487 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7488 *enabled
= conntrack_get_tcp_seq_chk(dp
->conntrack
);
7493 dpif_netdev_ct_set_limits(struct dpif
*dpif OVS_UNUSED
,
7494 const uint32_t *default_limits
,
7495 const struct ovs_list
*zone_limits
)
7498 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7499 if (default_limits
) {
7500 err
= zone_limit_update(dp
->conntrack
, DEFAULT_ZONE
, *default_limits
);
7506 struct ct_dpif_zone_limit
*zone_limit
;
7507 LIST_FOR_EACH (zone_limit
, node
, zone_limits
) {
7508 err
= zone_limit_update(dp
->conntrack
, zone_limit
->zone
,
7518 dpif_netdev_ct_get_limits(struct dpif
*dpif OVS_UNUSED
,
7519 uint32_t *default_limit
,
7520 const struct ovs_list
*zone_limits_request
,
7521 struct ovs_list
*zone_limits_reply
)
7523 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7524 struct conntrack_zone_limit czl
;
7526 czl
= zone_limit_get(dp
->conntrack
, DEFAULT_ZONE
);
7527 if (czl
.zone
== DEFAULT_ZONE
) {
7528 *default_limit
= czl
.limit
;
7533 if (!ovs_list_is_empty(zone_limits_request
)) {
7534 struct ct_dpif_zone_limit
*zone_limit
;
7535 LIST_FOR_EACH (zone_limit
, node
, zone_limits_request
) {
7536 czl
= zone_limit_get(dp
->conntrack
, zone_limit
->zone
);
7537 if (czl
.zone
== zone_limit
->zone
|| czl
.zone
== DEFAULT_ZONE
) {
7538 ct_dpif_push_zone_limit(zone_limits_reply
, zone_limit
->zone
,
7539 czl
.limit
, czl
.count
);
7545 for (int z
= MIN_ZONE
; z
<= MAX_ZONE
; z
++) {
7546 czl
= zone_limit_get(dp
->conntrack
, z
);
7547 if (czl
.zone
== z
) {
7548 ct_dpif_push_zone_limit(zone_limits_reply
, z
, czl
.limit
,
7558 dpif_netdev_ct_del_limits(struct dpif
*dpif OVS_UNUSED
,
7559 const struct ovs_list
*zone_limits
)
7562 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7563 struct ct_dpif_zone_limit
*zone_limit
;
7564 LIST_FOR_EACH (zone_limit
, node
, zone_limits
) {
7565 err
= zone_limit_delete(dp
->conntrack
, zone_limit
->zone
);
7575 dpif_netdev_ipf_set_enabled(struct dpif
*dpif
, bool v6
, bool enable
)
7577 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7578 return ipf_set_enabled(conntrack_ipf_ctx(dp
->conntrack
), v6
, enable
);
7582 dpif_netdev_ipf_set_min_frag(struct dpif
*dpif
, bool v6
, uint32_t min_frag
)
7584 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7585 return ipf_set_min_frag(conntrack_ipf_ctx(dp
->conntrack
), v6
, min_frag
);
7589 dpif_netdev_ipf_set_max_nfrags(struct dpif
*dpif
, uint32_t max_frags
)
7591 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7592 return ipf_set_max_nfrags(conntrack_ipf_ctx(dp
->conntrack
), max_frags
);
7595 /* Adjust this function if 'dpif_ipf_status' and 'ipf_status' were to
7598 dpif_netdev_ipf_get_status(struct dpif
*dpif
,
7599 struct dpif_ipf_status
*dpif_ipf_status
)
7601 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7602 ipf_get_status(conntrack_ipf_ctx(dp
->conntrack
),
7603 (struct ipf_status
*) dpif_ipf_status
);
7608 dpif_netdev_ipf_dump_start(struct dpif
*dpif OVS_UNUSED
,
7609 struct ipf_dump_ctx
**ipf_dump_ctx
)
7611 return ipf_dump_start(ipf_dump_ctx
);
7615 dpif_netdev_ipf_dump_next(struct dpif
*dpif
, void *ipf_dump_ctx
, char **dump
)
7617 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7618 return ipf_dump_next(conntrack_ipf_ctx(dp
->conntrack
), ipf_dump_ctx
,
7623 dpif_netdev_ipf_dump_done(struct dpif
*dpif OVS_UNUSED
, void *ipf_dump_ctx
)
7625 return ipf_dump_done(ipf_dump_ctx
);
7629 const struct dpif_class dpif_netdev_class
= {
7631 true, /* cleanup_required */
7633 dpif_netdev_enumerate
,
7634 dpif_netdev_port_open_type
,
7637 dpif_netdev_destroy
,
7640 dpif_netdev_get_stats
,
7641 dpif_netdev_port_add
,
7642 dpif_netdev_port_del
,
7643 dpif_netdev_port_set_config
,
7644 dpif_netdev_port_query_by_number
,
7645 dpif_netdev_port_query_by_name
,
7646 NULL
, /* port_get_pid */
7647 dpif_netdev_port_dump_start
,
7648 dpif_netdev_port_dump_next
,
7649 dpif_netdev_port_dump_done
,
7650 dpif_netdev_port_poll
,
7651 dpif_netdev_port_poll_wait
,
7652 dpif_netdev_flow_flush
,
7653 dpif_netdev_flow_dump_create
,
7654 dpif_netdev_flow_dump_destroy
,
7655 dpif_netdev_flow_dump_thread_create
,
7656 dpif_netdev_flow_dump_thread_destroy
,
7657 dpif_netdev_flow_dump_next
,
7658 dpif_netdev_operate
,
7659 NULL
, /* recv_set */
7660 NULL
, /* handlers_set */
7661 dpif_netdev_set_config
,
7662 dpif_netdev_queue_to_priority
,
7664 NULL
, /* recv_wait */
7665 NULL
, /* recv_purge */
7666 dpif_netdev_register_dp_purge_cb
,
7667 dpif_netdev_register_upcall_cb
,
7668 dpif_netdev_enable_upcall
,
7669 dpif_netdev_disable_upcall
,
7670 dpif_netdev_get_datapath_version
,
7671 dpif_netdev_ct_dump_start
,
7672 dpif_netdev_ct_dump_next
,
7673 dpif_netdev_ct_dump_done
,
7674 dpif_netdev_ct_flush
,
7675 dpif_netdev_ct_set_maxconns
,
7676 dpif_netdev_ct_get_maxconns
,
7677 dpif_netdev_ct_get_nconns
,
7678 dpif_netdev_ct_set_tcp_seq_chk
,
7679 dpif_netdev_ct_get_tcp_seq_chk
,
7680 dpif_netdev_ct_set_limits
,
7681 dpif_netdev_ct_get_limits
,
7682 dpif_netdev_ct_del_limits
,
7683 NULL
, /* ct_set_timeout_policy */
7684 NULL
, /* ct_get_timeout_policy */
7685 NULL
, /* ct_del_timeout_policy */
7686 NULL
, /* ct_timeout_policy_dump_start */
7687 NULL
, /* ct_timeout_policy_dump_next */
7688 NULL
, /* ct_timeout_policy_dump_done */
7689 NULL
, /* ct_get_timeout_policy_name */
7690 dpif_netdev_ipf_set_enabled
,
7691 dpif_netdev_ipf_set_min_frag
,
7692 dpif_netdev_ipf_set_max_nfrags
,
7693 dpif_netdev_ipf_get_status
,
7694 dpif_netdev_ipf_dump_start
,
7695 dpif_netdev_ipf_dump_next
,
7696 dpif_netdev_ipf_dump_done
,
7697 dpif_netdev_meter_get_features
,
7698 dpif_netdev_meter_set
,
7699 dpif_netdev_meter_get
,
7700 dpif_netdev_meter_del
,
7704 dpif_dummy_change_port_number(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
7705 const char *argv
[], void *aux OVS_UNUSED
)
7707 struct dp_netdev_port
*port
;
7708 struct dp_netdev
*dp
;
7711 ovs_mutex_lock(&dp_netdev_mutex
);
7712 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
7713 if (!dp
|| !dpif_netdev_class_is_dummy(dp
->class)) {
7714 ovs_mutex_unlock(&dp_netdev_mutex
);
7715 unixctl_command_reply_error(conn
, "unknown datapath or not a dummy");
7718 ovs_refcount_ref(&dp
->ref_cnt
);
7719 ovs_mutex_unlock(&dp_netdev_mutex
);
7721 ovs_mutex_lock(&dp
->port_mutex
);
7722 if (get_port_by_name(dp
, argv
[2], &port
)) {
7723 unixctl_command_reply_error(conn
, "unknown port");
7727 port_no
= u32_to_odp(atoi(argv
[3]));
7728 if (!port_no
|| port_no
== ODPP_NONE
) {
7729 unixctl_command_reply_error(conn
, "bad port number");
7732 if (dp_netdev_lookup_port(dp
, port_no
)) {
7733 unixctl_command_reply_error(conn
, "port number already in use");
7738 hmap_remove(&dp
->ports
, &port
->node
);
7739 reconfigure_datapath(dp
);
7741 /* Reinsert with new port number. */
7742 port
->port_no
= port_no
;
7743 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
7744 reconfigure_datapath(dp
);
7746 seq_change(dp
->port_seq
);
7747 unixctl_command_reply(conn
, NULL
);
7750 ovs_mutex_unlock(&dp
->port_mutex
);
7751 dp_netdev_unref(dp
);
7755 dpif_dummy_register__(const char *type
)
7757 struct dpif_class
*class;
7759 class = xmalloc(sizeof *class);
7760 *class = dpif_netdev_class
;
7761 class->type
= xstrdup(type
);
7762 dp_register_provider(class);
7766 dpif_dummy_override(const char *type
)
7771 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
7772 * a userland-only build. It's useful for testsuite.
7774 error
= dp_unregister_provider(type
);
7775 if (error
== 0 || error
== EAFNOSUPPORT
) {
7776 dpif_dummy_register__(type
);
7781 dpif_dummy_register(enum dummy_level level
)
7783 if (level
== DUMMY_OVERRIDE_ALL
) {
7788 dp_enumerate_types(&types
);
7789 SSET_FOR_EACH (type
, &types
) {
7790 dpif_dummy_override(type
);
7792 sset_destroy(&types
);
7793 } else if (level
== DUMMY_OVERRIDE_SYSTEM
) {
7794 dpif_dummy_override("system");
7797 dpif_dummy_register__("dummy");
7799 unixctl_command_register("dpif-dummy/change-port-number",
7800 "dp port new-number",
7801 3, 3, dpif_dummy_change_port_number
, NULL
);
7804 /* Datapath Classifier. */
7807 dpcls_subtable_destroy_cb(struct dpcls_subtable
*subtable
)
7809 cmap_destroy(&subtable
->rules
);
7810 ovsrcu_postpone(free
, subtable
->mf_masks
);
7811 ovsrcu_postpone(free
, subtable
);
7814 /* Initializes 'cls' as a classifier that initially contains no classification
7817 dpcls_init(struct dpcls
*cls
)
7819 cmap_init(&cls
->subtables_map
);
7820 pvector_init(&cls
->subtables
);
7824 dpcls_destroy_subtable(struct dpcls
*cls
, struct dpcls_subtable
*subtable
)
7826 VLOG_DBG("Destroying subtable %p for in_port %d", subtable
, cls
->in_port
);
7827 pvector_remove(&cls
->subtables
, subtable
);
7828 cmap_remove(&cls
->subtables_map
, &subtable
->cmap_node
,
7829 subtable
->mask
.hash
);
7830 ovsrcu_postpone(dpcls_subtable_destroy_cb
, subtable
);
7833 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
7834 * caller's responsibility.
7835 * May only be called after all the readers have been terminated. */
7837 dpcls_destroy(struct dpcls
*cls
)
7840 struct dpcls_subtable
*subtable
;
7842 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
7843 ovs_assert(cmap_count(&subtable
->rules
) == 0);
7844 dpcls_destroy_subtable(cls
, subtable
);
7846 cmap_destroy(&cls
->subtables_map
);
7847 pvector_destroy(&cls
->subtables
);
7851 static struct dpcls_subtable
*
7852 dpcls_create_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
7854 struct dpcls_subtable
*subtable
;
7856 /* Need to add one. */
7857 subtable
= xmalloc(sizeof *subtable
7858 - sizeof subtable
->mask
.mf
+ mask
->len
);
7859 cmap_init(&subtable
->rules
);
7860 subtable
->hit_cnt
= 0;
7861 netdev_flow_key_clone(&subtable
->mask
, mask
);
7863 /* The count of bits in the mask defines the space required for masks.
7864 * Then call gen_masks() to create the appropriate masks, avoiding the cost
7865 * of doing runtime calculations. */
7866 uint32_t unit0
= count_1bits(mask
->mf
.map
.bits
[0]);
7867 uint32_t unit1
= count_1bits(mask
->mf
.map
.bits
[1]);
7868 subtable
->mf_bits_set_unit0
= unit0
;
7869 subtable
->mf_bits_set_unit1
= unit1
;
7870 subtable
->mf_masks
= xmalloc(sizeof(uint64_t) * (unit0
+ unit1
));
7871 netdev_flow_key_gen_masks(mask
, subtable
->mf_masks
, unit0
, unit1
);
7873 /* Probe for a specialized generic lookup function. */
7874 subtable
->lookup_func
= dpcls_subtable_generic_probe(unit0
, unit1
);
7876 /* If not set, assign generic lookup. Generic works for any miniflow. */
7877 if (!subtable
->lookup_func
) {
7878 subtable
->lookup_func
= dpcls_subtable_lookup_generic
;
7881 cmap_insert(&cls
->subtables_map
, &subtable
->cmap_node
, mask
->hash
);
7882 /* Add the new subtable at the end of the pvector (with no hits yet) */
7883 pvector_insert(&cls
->subtables
, subtable
, 0);
7884 VLOG_DBG("Creating %"PRIuSIZE
". subtable %p for in_port %d",
7885 cmap_count(&cls
->subtables_map
), subtable
, cls
->in_port
);
7886 pvector_publish(&cls
->subtables
);
7891 static inline struct dpcls_subtable
*
7892 dpcls_find_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
7894 struct dpcls_subtable
*subtable
;
7896 CMAP_FOR_EACH_WITH_HASH (subtable
, cmap_node
, mask
->hash
,
7897 &cls
->subtables_map
) {
7898 if (netdev_flow_key_equal(&subtable
->mask
, mask
)) {
7902 return dpcls_create_subtable(cls
, mask
);
7906 /* Periodically sort the dpcls subtable vectors according to hit counts */
7908 dpcls_sort_subtable_vector(struct dpcls
*cls
)
7910 struct pvector
*pvec
= &cls
->subtables
;
7911 struct dpcls_subtable
*subtable
;
7913 PVECTOR_FOR_EACH (subtable
, pvec
) {
7914 pvector_change_priority(pvec
, subtable
, subtable
->hit_cnt
);
7915 subtable
->hit_cnt
= 0;
7917 pvector_publish(pvec
);
7921 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
,
7922 struct polled_queue
*poll_list
, int poll_cnt
)
7925 uint64_t tot_idle
= 0, tot_proc
= 0;
7926 unsigned int pmd_load
= 0;
7928 if (pmd
->ctx
.now
> pmd
->rxq_next_cycle_store
) {
7930 struct pmd_auto_lb
*pmd_alb
= &pmd
->dp
->pmd_alb
;
7931 if (pmd_alb
->is_enabled
&& !pmd
->isolated
7932 && (pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_IDLE
] >=
7933 pmd
->prev_stats
[PMD_CYCLES_ITER_IDLE
])
7934 && (pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_BUSY
] >=
7935 pmd
->prev_stats
[PMD_CYCLES_ITER_BUSY
]))
7937 tot_idle
= pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_IDLE
] -
7938 pmd
->prev_stats
[PMD_CYCLES_ITER_IDLE
];
7939 tot_proc
= pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_BUSY
] -
7940 pmd
->prev_stats
[PMD_CYCLES_ITER_BUSY
];
7943 pmd_load
= ((tot_proc
* 100) / (tot_idle
+ tot_proc
));
7946 if (pmd_load
>= ALB_PMD_LOAD_THRESHOLD
) {
7947 atomic_count_inc(&pmd
->pmd_overloaded
);
7949 atomic_count_set(&pmd
->pmd_overloaded
, 0);
7953 pmd
->prev_stats
[PMD_CYCLES_ITER_IDLE
] =
7954 pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_IDLE
];
7955 pmd
->prev_stats
[PMD_CYCLES_ITER_BUSY
] =
7956 pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_BUSY
];
7958 /* Get the cycles that were used to process each queue and store. */
7959 for (unsigned i
= 0; i
< poll_cnt
; i
++) {
7960 uint64_t rxq_cyc_curr
= dp_netdev_rxq_get_cycles(poll_list
[i
].rxq
,
7961 RXQ_CYCLES_PROC_CURR
);
7962 dp_netdev_rxq_set_intrvl_cycles(poll_list
[i
].rxq
, rxq_cyc_curr
);
7963 dp_netdev_rxq_set_cycles(poll_list
[i
].rxq
, RXQ_CYCLES_PROC_CURR
,
7966 curr_tsc
= cycles_counter_update(&pmd
->perf_stats
);
7967 if (pmd
->intrvl_tsc_prev
) {
7968 /* There is a prev timestamp, store a new intrvl cycle count. */
7969 atomic_store_relaxed(&pmd
->intrvl_cycles
,
7970 curr_tsc
- pmd
->intrvl_tsc_prev
);
7972 pmd
->intrvl_tsc_prev
= curr_tsc
;
7973 /* Start new measuring interval */
7974 pmd
->rxq_next_cycle_store
= pmd
->ctx
.now
+ PMD_RXQ_INTERVAL_LEN
;
7977 if (pmd
->ctx
.now
> pmd
->next_optimization
) {
7978 /* Try to obtain the flow lock to block out revalidator threads.
7979 * If not possible, just try next time. */
7980 if (!ovs_mutex_trylock(&pmd
->flow_mutex
)) {
7981 /* Optimize each classifier */
7982 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
7983 dpcls_sort_subtable_vector(cls
);
7985 ovs_mutex_unlock(&pmd
->flow_mutex
);
7986 /* Start new measuring interval */
7987 pmd
->next_optimization
= pmd
->ctx
.now
7988 + DPCLS_OPTIMIZATION_INTERVAL
;
7993 /* Insert 'rule' into 'cls'. */
7995 dpcls_insert(struct dpcls
*cls
, struct dpcls_rule
*rule
,
7996 const struct netdev_flow_key
*mask
)
7998 struct dpcls_subtable
*subtable
= dpcls_find_subtable(cls
, mask
);
8000 /* Refer to subtable's mask, also for later removal. */
8001 rule
->mask
= &subtable
->mask
;
8002 cmap_insert(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
);
8005 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
8007 dpcls_remove(struct dpcls
*cls
, struct dpcls_rule
*rule
)
8009 struct dpcls_subtable
*subtable
;
8011 ovs_assert(rule
->mask
);
8013 /* Get subtable from reference in rule->mask. */
8014 INIT_CONTAINER(subtable
, rule
->mask
, mask
);
8015 if (cmap_remove(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
)
8017 /* Delete empty subtable. */
8018 dpcls_destroy_subtable(cls
, subtable
);
8019 pvector_publish(&cls
->subtables
);
8023 /* Inner loop for mask generation of a unit, see netdev_flow_key_gen_masks. */
8025 netdev_flow_key_gen_mask_unit(uint64_t iter
,
8026 const uint64_t count
,
8030 for (i
= 0; i
< count
; i
++) {
8031 uint64_t lowest_bit
= (iter
& -iter
);
8032 iter
&= ~lowest_bit
;
8033 mf_masks
[i
] = (lowest_bit
- 1);
8035 /* Checks that count has covered all bits in the iter bitmap. */
8036 ovs_assert(iter
== 0);
8039 /* Generate a mask for each block in the miniflow, based on the bits set. This
8040 * allows easily masking packets with the generated array here, without
8041 * calculations. This replaces runtime-calculating the masks.
8042 * @param key The table to generate the mf_masks for
8043 * @param mf_masks Pointer to a u64 array of at least *mf_bits* in size
8044 * @param mf_bits_total Number of bits set in the whole miniflow (both units)
8045 * @param mf_bits_unit0 Number of bits set in unit0 of the miniflow
8048 netdev_flow_key_gen_masks(const struct netdev_flow_key
*tbl
,
8050 const uint32_t mf_bits_u0
,
8051 const uint32_t mf_bits_u1
)
8053 uint64_t iter_u0
= tbl
->mf
.map
.bits
[0];
8054 uint64_t iter_u1
= tbl
->mf
.map
.bits
[1];
8056 netdev_flow_key_gen_mask_unit(iter_u0
, mf_bits_u0
, &mf_masks
[0]);
8057 netdev_flow_key_gen_mask_unit(iter_u1
, mf_bits_u1
, &mf_masks
[mf_bits_u0
]);
8060 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
8061 * in 'mask' the values in 'key' and 'target' are the same. */
8063 dpcls_rule_matches_key(const struct dpcls_rule
*rule
,
8064 const struct netdev_flow_key
*target
)
8066 const uint64_t *keyp
= miniflow_get_values(&rule
->flow
.mf
);
8067 const uint64_t *maskp
= miniflow_get_values(&rule
->mask
->mf
);
8070 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, target
, rule
->flow
.mf
.map
) {
8071 if (OVS_UNLIKELY((value
& *maskp
++) != *keyp
++)) {
8078 /* For each miniflow in 'keys' performs a classifier lookup writing the result
8079 * into the corresponding slot in 'rules'. If a particular entry in 'keys' is
8080 * NULL it is skipped.
8082 * This function is optimized for use in the userspace datapath and therefore
8083 * does not implement a lot of features available in the standard
8084 * classifier_lookup() function. Specifically, it does not implement
8085 * priorities, instead returning any rule which matches the flow.
8087 * Returns true if all miniflows found a corresponding rule. */
8089 dpcls_lookup(struct dpcls
*cls
, const struct netdev_flow_key
*keys
[],
8090 struct dpcls_rule
**rules
, const size_t cnt
,
8093 /* The received 'cnt' miniflows are the search-keys that will be processed
8094 * to find a matching entry into the available subtables.
8095 * The number of bits in map_type is equal to NETDEV_MAX_BURST. */
8096 #define MAP_BITS (sizeof(uint32_t) * CHAR_BIT)
8097 BUILD_ASSERT_DECL(MAP_BITS
>= NETDEV_MAX_BURST
);
8099 struct dpcls_subtable
*subtable
;
8100 uint32_t keys_map
= TYPE_MAXIMUM(uint32_t); /* Set all bits. */
8102 if (cnt
!= MAP_BITS
) {
8103 keys_map
>>= MAP_BITS
- cnt
; /* Clear extra bits. */
8105 memset(rules
, 0, cnt
* sizeof *rules
);
8107 int lookups_match
= 0, subtable_pos
= 1;
8110 /* The Datapath classifier - aka dpcls - is composed of subtables.
8111 * Subtables are dynamically created as needed when new rules are inserted.
8112 * Each subtable collects rules with matches on a specific subset of packet
8113 * fields as defined by the subtable's mask. We proceed to process every
8114 * search-key against each subtable, but when a match is found for a
8115 * search-key, the search for that key can stop because the rules are
8116 * non-overlapping. */
8117 PVECTOR_FOR_EACH (subtable
, &cls
->subtables
) {
8118 /* Call the subtable specific lookup function. */
8119 found_map
= subtable
->lookup_func(subtable
, keys_map
, keys
, rules
);
8121 /* Count the number of subtables searched for this packet match. This
8122 * estimates the "spread" of subtables looked at per matched packet. */
8123 uint32_t pkts_matched
= count_1bits(found_map
);
8124 lookups_match
+= pkts_matched
* subtable_pos
;
8126 /* Clear the found rules, and return early if all packets are found. */
8127 keys_map
&= ~found_map
;
8129 if (num_lookups_p
) {
8130 *num_lookups_p
= lookups_match
;
8137 if (num_lookups_p
) {
8138 *num_lookups_p
= lookups_match
;