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 netdev_saved_flags
*sf
;
1806 struct dp_netdev_port
*port
;
1807 enum netdev_flags flags
;
1808 struct netdev
*netdev
;
1813 /* Open and validate network device. */
1814 error
= netdev_open(devname
, type
, &netdev
);
1818 /* XXX reject non-Ethernet devices */
1820 netdev_get_flags(netdev
, &flags
);
1821 if (flags
& NETDEV_LOOPBACK
) {
1822 VLOG_ERR("%s: cannot add a loopback device", devname
);
1827 error
= netdev_turn_flags_on(netdev
, NETDEV_PROMISC
, &sf
);
1829 VLOG_ERR("%s: cannot set promisc flag", devname
);
1833 port
= xzalloc(sizeof *port
);
1834 port
->port_no
= port_no
;
1835 port
->netdev
= netdev
;
1836 port
->type
= xstrdup(type
);
1838 port
->emc_enabled
= true;
1839 port
->need_reconfigure
= true;
1840 ovs_mutex_init(&port
->txq_used_mutex
);
1847 netdev_close(netdev
);
1852 do_add_port(struct dp_netdev
*dp
, const char *devname
, const char *type
,
1854 OVS_REQUIRES(dp
->port_mutex
)
1856 struct dp_netdev_port
*port
;
1859 /* Reject devices already in 'dp'. */
1860 if (!get_port_by_name(dp
, devname
, &port
)) {
1864 error
= port_create(devname
, type
, port_no
, &port
);
1869 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
1870 seq_change(dp
->port_seq
);
1872 reconfigure_datapath(dp
);
1874 /* Check that port was successfully configured. */
1875 return dp_netdev_lookup_port(dp
, port_no
) ? 0 : EINVAL
;
1879 dpif_netdev_port_add(struct dpif
*dpif
, struct netdev
*netdev
,
1880 odp_port_t
*port_nop
)
1882 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1883 char namebuf
[NETDEV_VPORT_NAME_BUFSIZE
];
1884 const char *dpif_port
;
1888 ovs_mutex_lock(&dp
->port_mutex
);
1889 dpif_port
= netdev_vport_get_dpif_port(netdev
, namebuf
, sizeof namebuf
);
1890 if (*port_nop
!= ODPP_NONE
) {
1891 port_no
= *port_nop
;
1892 error
= dp_netdev_lookup_port(dp
, *port_nop
) ? EBUSY
: 0;
1894 port_no
= choose_port(dp
, dpif_port
);
1895 error
= port_no
== ODPP_NONE
? EFBIG
: 0;
1898 *port_nop
= port_no
;
1899 error
= do_add_port(dp
, dpif_port
, netdev_get_type(netdev
), port_no
);
1901 ovs_mutex_unlock(&dp
->port_mutex
);
1907 dpif_netdev_port_del(struct dpif
*dpif
, odp_port_t port_no
)
1909 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1912 ovs_mutex_lock(&dp
->port_mutex
);
1913 if (port_no
== ODPP_LOCAL
) {
1916 struct dp_netdev_port
*port
;
1918 error
= get_port_by_number(dp
, port_no
, &port
);
1920 do_del_port(dp
, port
);
1923 ovs_mutex_unlock(&dp
->port_mutex
);
1929 is_valid_port_number(odp_port_t port_no
)
1931 return port_no
!= ODPP_NONE
;
1934 static struct dp_netdev_port
*
1935 dp_netdev_lookup_port(const struct dp_netdev
*dp
, odp_port_t port_no
)
1936 OVS_REQUIRES(dp
->port_mutex
)
1938 struct dp_netdev_port
*port
;
1940 HMAP_FOR_EACH_WITH_HASH (port
, node
, hash_port_no(port_no
), &dp
->ports
) {
1941 if (port
->port_no
== port_no
) {
1949 get_port_by_number(struct dp_netdev
*dp
,
1950 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1951 OVS_REQUIRES(dp
->port_mutex
)
1953 if (!is_valid_port_number(port_no
)) {
1957 *portp
= dp_netdev_lookup_port(dp
, port_no
);
1958 return *portp
? 0 : ENODEV
;
1963 port_destroy(struct dp_netdev_port
*port
)
1969 netdev_close(port
->netdev
);
1970 netdev_restore_flags(port
->sf
);
1972 for (unsigned i
= 0; i
< port
->n_rxq
; i
++) {
1973 netdev_rxq_close(port
->rxqs
[i
].rx
);
1975 ovs_mutex_destroy(&port
->txq_used_mutex
);
1976 free(port
->rxq_affinity_list
);
1977 free(port
->txq_used
);
1984 get_port_by_name(struct dp_netdev
*dp
,
1985 const char *devname
, struct dp_netdev_port
**portp
)
1986 OVS_REQUIRES(dp
->port_mutex
)
1988 struct dp_netdev_port
*port
;
1990 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1991 if (!strcmp(netdev_get_name(port
->netdev
), devname
)) {
1997 /* Callers of dpif_netdev_port_query_by_name() expect ENODEV for a non
2002 /* Returns 'true' if there is a port with pmd netdev. */
2004 has_pmd_port(struct dp_netdev
*dp
)
2005 OVS_REQUIRES(dp
->port_mutex
)
2007 struct dp_netdev_port
*port
;
2009 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2010 if (netdev_is_pmd(port
->netdev
)) {
2019 do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*port
)
2020 OVS_REQUIRES(dp
->port_mutex
)
2022 hmap_remove(&dp
->ports
, &port
->node
);
2023 seq_change(dp
->port_seq
);
2025 reconfigure_datapath(dp
);
2031 answer_port_query(const struct dp_netdev_port
*port
,
2032 struct dpif_port
*dpif_port
)
2034 dpif_port
->name
= xstrdup(netdev_get_name(port
->netdev
));
2035 dpif_port
->type
= xstrdup(port
->type
);
2036 dpif_port
->port_no
= port
->port_no
;
2040 dpif_netdev_port_query_by_number(const struct dpif
*dpif
, odp_port_t port_no
,
2041 struct dpif_port
*dpif_port
)
2043 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2044 struct dp_netdev_port
*port
;
2047 ovs_mutex_lock(&dp
->port_mutex
);
2048 error
= get_port_by_number(dp
, port_no
, &port
);
2049 if (!error
&& dpif_port
) {
2050 answer_port_query(port
, dpif_port
);
2052 ovs_mutex_unlock(&dp
->port_mutex
);
2058 dpif_netdev_port_query_by_name(const struct dpif
*dpif
, const char *devname
,
2059 struct dpif_port
*dpif_port
)
2061 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2062 struct dp_netdev_port
*port
;
2065 ovs_mutex_lock(&dp
->port_mutex
);
2066 error
= get_port_by_name(dp
, devname
, &port
);
2067 if (!error
&& dpif_port
) {
2068 answer_port_query(port
, dpif_port
);
2070 ovs_mutex_unlock(&dp
->port_mutex
);
2076 dp_netdev_flow_free(struct dp_netdev_flow
*flow
)
2078 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow
));
2082 static void dp_netdev_flow_unref(struct dp_netdev_flow
*flow
)
2084 if (ovs_refcount_unref_relaxed(&flow
->ref_cnt
) == 1) {
2085 ovsrcu_postpone(dp_netdev_flow_free
, flow
);
2090 dp_netdev_flow_hash(const ovs_u128
*ufid
)
2092 return ufid
->u32
[0];
2095 static inline struct dpcls
*
2096 dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread
*pmd
,
2100 uint32_t hash
= hash_port_no(in_port
);
2101 CMAP_FOR_EACH_WITH_HASH (cls
, node
, hash
, &pmd
->classifiers
) {
2102 if (cls
->in_port
== in_port
) {
2103 /* Port classifier exists already */
2110 static inline struct dpcls
*
2111 dp_netdev_pmd_find_dpcls(struct dp_netdev_pmd_thread
*pmd
,
2113 OVS_REQUIRES(pmd
->flow_mutex
)
2115 struct dpcls
*cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2116 uint32_t hash
= hash_port_no(in_port
);
2119 /* Create new classifier for in_port */
2120 cls
= xmalloc(sizeof(*cls
));
2122 cls
->in_port
= in_port
;
2123 cmap_insert(&pmd
->classifiers
, &cls
->node
, hash
);
2124 VLOG_DBG("Creating dpcls %p for in_port %d", cls
, in_port
);
2129 #define MAX_FLOW_MARK (UINT32_MAX - 1)
2130 #define INVALID_FLOW_MARK (UINT32_MAX)
2132 struct megaflow_to_mark_data
{
2133 const struct cmap_node node
;
2139 struct cmap megaflow_to_mark
;
2140 struct cmap mark_to_flow
;
2141 struct id_pool
*pool
;
2144 static struct flow_mark flow_mark
= {
2145 .megaflow_to_mark
= CMAP_INITIALIZER
,
2146 .mark_to_flow
= CMAP_INITIALIZER
,
2150 flow_mark_alloc(void)
2154 if (!flow_mark
.pool
) {
2155 /* Haven't initiated yet, do it here */
2156 flow_mark
.pool
= id_pool_create(0, MAX_FLOW_MARK
);
2159 if (id_pool_alloc_id(flow_mark
.pool
, &mark
)) {
2163 return INVALID_FLOW_MARK
;
2167 flow_mark_free(uint32_t mark
)
2169 id_pool_free_id(flow_mark
.pool
, mark
);
2172 /* associate megaflow with a mark, which is a 1:1 mapping */
2174 megaflow_to_mark_associate(const ovs_u128
*mega_ufid
, uint32_t mark
)
2176 size_t hash
= dp_netdev_flow_hash(mega_ufid
);
2177 struct megaflow_to_mark_data
*data
= xzalloc(sizeof(*data
));
2179 data
->mega_ufid
= *mega_ufid
;
2182 cmap_insert(&flow_mark
.megaflow_to_mark
,
2183 CONST_CAST(struct cmap_node
*, &data
->node
), hash
);
2186 /* disassociate meagaflow with a mark */
2188 megaflow_to_mark_disassociate(const ovs_u128
*mega_ufid
)
2190 size_t hash
= dp_netdev_flow_hash(mega_ufid
);
2191 struct megaflow_to_mark_data
*data
;
2193 CMAP_FOR_EACH_WITH_HASH (data
, node
, hash
, &flow_mark
.megaflow_to_mark
) {
2194 if (ovs_u128_equals(*mega_ufid
, data
->mega_ufid
)) {
2195 cmap_remove(&flow_mark
.megaflow_to_mark
,
2196 CONST_CAST(struct cmap_node
*, &data
->node
), hash
);
2197 ovsrcu_postpone(free
, data
);
2202 VLOG_WARN("Masked ufid "UUID_FMT
" is not associated with a mark?\n",
2203 UUID_ARGS((struct uuid
*)mega_ufid
));
2206 static inline uint32_t
2207 megaflow_to_mark_find(const ovs_u128
*mega_ufid
)
2209 size_t hash
= dp_netdev_flow_hash(mega_ufid
);
2210 struct megaflow_to_mark_data
*data
;
2212 CMAP_FOR_EACH_WITH_HASH (data
, node
, hash
, &flow_mark
.megaflow_to_mark
) {
2213 if (ovs_u128_equals(*mega_ufid
, data
->mega_ufid
)) {
2218 VLOG_DBG("Mark id for ufid "UUID_FMT
" was not found\n",
2219 UUID_ARGS((struct uuid
*)mega_ufid
));
2220 return INVALID_FLOW_MARK
;
2223 /* associate mark with a flow, which is 1:N mapping */
2225 mark_to_flow_associate(const uint32_t mark
, struct dp_netdev_flow
*flow
)
2227 dp_netdev_flow_ref(flow
);
2229 cmap_insert(&flow_mark
.mark_to_flow
,
2230 CONST_CAST(struct cmap_node
*, &flow
->mark_node
),
2234 VLOG_DBG("Associated dp_netdev flow %p with mark %u\n", flow
, mark
);
2238 flow_mark_has_no_ref(uint32_t mark
)
2240 struct dp_netdev_flow
*flow
;
2242 CMAP_FOR_EACH_WITH_HASH (flow
, mark_node
, hash_int(mark
, 0),
2243 &flow_mark
.mark_to_flow
) {
2244 if (flow
->mark
== mark
) {
2253 mark_to_flow_disassociate(struct dp_netdev_pmd_thread
*pmd
,
2254 struct dp_netdev_flow
*flow
)
2257 uint32_t mark
= flow
->mark
;
2258 struct cmap_node
*mark_node
= CONST_CAST(struct cmap_node
*,
2261 cmap_remove(&flow_mark
.mark_to_flow
, mark_node
, hash_int(mark
, 0));
2262 flow
->mark
= INVALID_FLOW_MARK
;
2265 * no flow is referencing the mark any more? If so, let's
2266 * remove the flow from hardware and free the mark.
2268 if (flow_mark_has_no_ref(mark
)) {
2269 struct dp_netdev_port
*port
;
2270 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
2272 ovs_mutex_lock(&pmd
->dp
->port_mutex
);
2273 port
= dp_netdev_lookup_port(pmd
->dp
, in_port
);
2275 ret
= netdev_flow_del(port
->netdev
, &flow
->mega_ufid
, NULL
);
2277 ovs_mutex_unlock(&pmd
->dp
->port_mutex
);
2279 flow_mark_free(mark
);
2280 VLOG_DBG("Freed flow mark %u\n", mark
);
2282 megaflow_to_mark_disassociate(&flow
->mega_ufid
);
2284 dp_netdev_flow_unref(flow
);
2290 flow_mark_flush(struct dp_netdev_pmd_thread
*pmd
)
2292 struct dp_netdev_flow
*flow
;
2294 CMAP_FOR_EACH (flow
, mark_node
, &flow_mark
.mark_to_flow
) {
2295 if (flow
->pmd_id
== pmd
->core_id
) {
2296 queue_netdev_flow_del(pmd
, flow
);
2301 static struct dp_netdev_flow
*
2302 mark_to_flow_find(const struct dp_netdev_pmd_thread
*pmd
,
2303 const uint32_t mark
)
2305 struct dp_netdev_flow
*flow
;
2307 CMAP_FOR_EACH_WITH_HASH (flow
, mark_node
, hash_int(mark
, 0),
2308 &flow_mark
.mark_to_flow
) {
2309 if (flow
->mark
== mark
&& flow
->pmd_id
== pmd
->core_id
&&
2310 flow
->dead
== false) {
2318 static struct dp_flow_offload_item
*
2319 dp_netdev_alloc_flow_offload(struct dp_netdev_pmd_thread
*pmd
,
2320 struct dp_netdev_flow
*flow
,
2323 struct dp_flow_offload_item
*offload
;
2325 offload
= xzalloc(sizeof(*offload
));
2327 offload
->flow
= flow
;
2330 dp_netdev_flow_ref(flow
);
2331 dp_netdev_pmd_try_ref(pmd
);
2337 dp_netdev_free_flow_offload(struct dp_flow_offload_item
*offload
)
2339 dp_netdev_pmd_unref(offload
->pmd
);
2340 dp_netdev_flow_unref(offload
->flow
);
2342 free(offload
->actions
);
2347 dp_netdev_append_flow_offload(struct dp_flow_offload_item
*offload
)
2349 ovs_mutex_lock(&dp_flow_offload
.mutex
);
2350 ovs_list_push_back(&dp_flow_offload
.list
, &offload
->node
);
2351 xpthread_cond_signal(&dp_flow_offload
.cond
);
2352 ovs_mutex_unlock(&dp_flow_offload
.mutex
);
2356 dp_netdev_flow_offload_del(struct dp_flow_offload_item
*offload
)
2358 return mark_to_flow_disassociate(offload
->pmd
, offload
->flow
);
2362 * There are two flow offload operations here: addition and modification.
2364 * For flow addition, this function does:
2365 * - allocate a new flow mark id
2366 * - perform hardware flow offload
2367 * - associate the flow mark with flow and mega flow
2369 * For flow modification, both flow mark and the associations are still
2370 * valid, thus only item 2 needed.
2373 dp_netdev_flow_offload_put(struct dp_flow_offload_item
*offload
)
2375 struct dp_netdev_port
*port
;
2376 struct dp_netdev_pmd_thread
*pmd
= offload
->pmd
;
2377 struct dp_netdev_flow
*flow
= offload
->flow
;
2378 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
2379 bool modification
= offload
->op
== DP_NETDEV_FLOW_OFFLOAD_OP_MOD
;
2380 struct offload_info info
;
2390 ovs_assert(mark
!= INVALID_FLOW_MARK
);
2393 * If a mega flow has already been offloaded (from other PMD
2394 * instances), do not offload it again.
2396 mark
= megaflow_to_mark_find(&flow
->mega_ufid
);
2397 if (mark
!= INVALID_FLOW_MARK
) {
2398 VLOG_DBG("Flow has already been offloaded with mark %u\n", mark
);
2399 if (flow
->mark
!= INVALID_FLOW_MARK
) {
2400 ovs_assert(flow
->mark
== mark
);
2402 mark_to_flow_associate(mark
, flow
);
2407 mark
= flow_mark_alloc();
2408 if (mark
== INVALID_FLOW_MARK
) {
2409 VLOG_ERR("Failed to allocate flow mark!\n");
2412 info
.flow_mark
= mark
;
2414 ovs_mutex_lock(&pmd
->dp
->port_mutex
);
2415 port
= dp_netdev_lookup_port(pmd
->dp
, in_port
);
2416 if (!port
|| netdev_vport_is_vport_class(port
->netdev
->netdev_class
)) {
2417 ovs_mutex_unlock(&pmd
->dp
->port_mutex
);
2420 ret
= netdev_flow_put(port
->netdev
, &offload
->match
,
2421 CONST_CAST(struct nlattr
*, offload
->actions
),
2422 offload
->actions_len
, &flow
->mega_ufid
, &info
,
2424 ovs_mutex_unlock(&pmd
->dp
->port_mutex
);
2430 if (!modification
) {
2431 megaflow_to_mark_associate(&flow
->mega_ufid
, mark
);
2432 mark_to_flow_associate(mark
, flow
);
2437 if (!modification
) {
2438 flow_mark_free(mark
);
2440 mark_to_flow_disassociate(pmd
, flow
);
2446 dp_netdev_flow_offload_main(void *data OVS_UNUSED
)
2448 struct dp_flow_offload_item
*offload
;
2449 struct ovs_list
*list
;
2454 ovs_mutex_lock(&dp_flow_offload
.mutex
);
2455 if (ovs_list_is_empty(&dp_flow_offload
.list
)) {
2456 ovsrcu_quiesce_start();
2457 ovs_mutex_cond_wait(&dp_flow_offload
.cond
,
2458 &dp_flow_offload
.mutex
);
2459 ovsrcu_quiesce_end();
2461 list
= ovs_list_pop_front(&dp_flow_offload
.list
);
2462 offload
= CONTAINER_OF(list
, struct dp_flow_offload_item
, node
);
2463 ovs_mutex_unlock(&dp_flow_offload
.mutex
);
2465 switch (offload
->op
) {
2466 case DP_NETDEV_FLOW_OFFLOAD_OP_ADD
:
2468 ret
= dp_netdev_flow_offload_put(offload
);
2470 case DP_NETDEV_FLOW_OFFLOAD_OP_MOD
:
2472 ret
= dp_netdev_flow_offload_put(offload
);
2474 case DP_NETDEV_FLOW_OFFLOAD_OP_DEL
:
2476 ret
= dp_netdev_flow_offload_del(offload
);
2482 VLOG_DBG("%s to %s netdev flow\n",
2483 ret
== 0 ? "succeed" : "failed", op
);
2484 dp_netdev_free_flow_offload(offload
);
2491 queue_netdev_flow_del(struct dp_netdev_pmd_thread
*pmd
,
2492 struct dp_netdev_flow
*flow
)
2494 struct dp_flow_offload_item
*offload
;
2496 if (ovsthread_once_start(&offload_thread_once
)) {
2497 xpthread_cond_init(&dp_flow_offload
.cond
, NULL
);
2498 ovs_thread_create("dp_netdev_flow_offload",
2499 dp_netdev_flow_offload_main
, NULL
);
2500 ovsthread_once_done(&offload_thread_once
);
2503 offload
= dp_netdev_alloc_flow_offload(pmd
, flow
,
2504 DP_NETDEV_FLOW_OFFLOAD_OP_DEL
);
2505 dp_netdev_append_flow_offload(offload
);
2509 queue_netdev_flow_put(struct dp_netdev_pmd_thread
*pmd
,
2510 struct dp_netdev_flow
*flow
, struct match
*match
,
2511 const struct nlattr
*actions
, size_t actions_len
)
2513 struct dp_flow_offload_item
*offload
;
2516 if (!netdev_is_flow_api_enabled()) {
2520 if (ovsthread_once_start(&offload_thread_once
)) {
2521 xpthread_cond_init(&dp_flow_offload
.cond
, NULL
);
2522 ovs_thread_create("dp_netdev_flow_offload",
2523 dp_netdev_flow_offload_main
, NULL
);
2524 ovsthread_once_done(&offload_thread_once
);
2527 if (flow
->mark
!= INVALID_FLOW_MARK
) {
2528 op
= DP_NETDEV_FLOW_OFFLOAD_OP_MOD
;
2530 op
= DP_NETDEV_FLOW_OFFLOAD_OP_ADD
;
2532 offload
= dp_netdev_alloc_flow_offload(pmd
, flow
, op
);
2533 offload
->match
= *match
;
2534 offload
->actions
= xmalloc(actions_len
);
2535 memcpy(offload
->actions
, actions
, actions_len
);
2536 offload
->actions_len
= actions_len
;
2538 dp_netdev_append_flow_offload(offload
);
2542 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread
*pmd
,
2543 struct dp_netdev_flow
*flow
)
2544 OVS_REQUIRES(pmd
->flow_mutex
)
2546 struct cmap_node
*node
= CONST_CAST(struct cmap_node
*, &flow
->node
);
2548 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
2550 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2551 ovs_assert(cls
!= NULL
);
2552 dpcls_remove(cls
, &flow
->cr
);
2553 cmap_remove(&pmd
->flow_table
, node
, dp_netdev_flow_hash(&flow
->ufid
));
2554 if (flow
->mark
!= INVALID_FLOW_MARK
) {
2555 queue_netdev_flow_del(pmd
, flow
);
2559 dp_netdev_flow_unref(flow
);
2563 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
)
2565 struct dp_netdev_flow
*netdev_flow
;
2567 ovs_mutex_lock(&pmd
->flow_mutex
);
2568 CMAP_FOR_EACH (netdev_flow
, node
, &pmd
->flow_table
) {
2569 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
2571 ovs_mutex_unlock(&pmd
->flow_mutex
);
2575 dpif_netdev_flow_flush(struct dpif
*dpif
)
2577 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2578 struct dp_netdev_pmd_thread
*pmd
;
2580 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2581 dp_netdev_pmd_flow_flush(pmd
);
2587 struct dp_netdev_port_state
{
2588 struct hmap_position position
;
2593 dpif_netdev_port_dump_start(const struct dpif
*dpif OVS_UNUSED
, void **statep
)
2595 *statep
= xzalloc(sizeof(struct dp_netdev_port_state
));
2600 dpif_netdev_port_dump_next(const struct dpif
*dpif
, void *state_
,
2601 struct dpif_port
*dpif_port
)
2603 struct dp_netdev_port_state
*state
= state_
;
2604 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2605 struct hmap_node
*node
;
2608 ovs_mutex_lock(&dp
->port_mutex
);
2609 node
= hmap_at_position(&dp
->ports
, &state
->position
);
2611 struct dp_netdev_port
*port
;
2613 port
= CONTAINER_OF(node
, struct dp_netdev_port
, node
);
2616 state
->name
= xstrdup(netdev_get_name(port
->netdev
));
2617 dpif_port
->name
= state
->name
;
2618 dpif_port
->type
= port
->type
;
2619 dpif_port
->port_no
= port
->port_no
;
2625 ovs_mutex_unlock(&dp
->port_mutex
);
2631 dpif_netdev_port_dump_done(const struct dpif
*dpif OVS_UNUSED
, void *state_
)
2633 struct dp_netdev_port_state
*state
= state_
;
2640 dpif_netdev_port_poll(const struct dpif
*dpif_
, char **devnamep OVS_UNUSED
)
2642 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
2643 uint64_t new_port_seq
;
2646 new_port_seq
= seq_read(dpif
->dp
->port_seq
);
2647 if (dpif
->last_port_seq
!= new_port_seq
) {
2648 dpif
->last_port_seq
= new_port_seq
;
2658 dpif_netdev_port_poll_wait(const struct dpif
*dpif_
)
2660 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
2662 seq_wait(dpif
->dp
->port_seq
, dpif
->last_port_seq
);
2665 static struct dp_netdev_flow
*
2666 dp_netdev_flow_cast(const struct dpcls_rule
*cr
)
2668 return cr
? CONTAINER_OF(cr
, struct dp_netdev_flow
, cr
) : NULL
;
2671 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*flow
)
2673 return ovs_refcount_try_ref_rcu(&flow
->ref_cnt
);
2676 /* netdev_flow_key utilities.
2678 * netdev_flow_key is basically a miniflow. We use these functions
2679 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
2680 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
2682 * - Since we are dealing exclusively with miniflows created by
2683 * miniflow_extract(), if the map is different the miniflow is different.
2684 * Therefore we can be faster by comparing the map and the miniflow in a
2686 * - These functions can be inlined by the compiler. */
2688 /* Given the number of bits set in miniflow's maps, returns the size of the
2689 * 'netdev_flow_key.mf' */
2690 static inline size_t
2691 netdev_flow_key_size(size_t flow_u64s
)
2693 return sizeof(struct miniflow
) + MINIFLOW_VALUES_SIZE(flow_u64s
);
2697 netdev_flow_key_equal(const struct netdev_flow_key
*a
,
2698 const struct netdev_flow_key
*b
)
2700 /* 'b->len' may be not set yet. */
2701 return a
->hash
== b
->hash
&& !memcmp(&a
->mf
, &b
->mf
, a
->len
);
2704 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
2705 * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been
2706 * generated by miniflow_extract. */
2708 netdev_flow_key_equal_mf(const struct netdev_flow_key
*key
,
2709 const struct miniflow
*mf
)
2711 return !memcmp(&key
->mf
, mf
, key
->len
);
2715 netdev_flow_key_clone(struct netdev_flow_key
*dst
,
2716 const struct netdev_flow_key
*src
)
2719 offsetof(struct netdev_flow_key
, mf
) + src
->len
);
2722 /* Initialize a netdev_flow_key 'mask' from 'match'. */
2724 netdev_flow_mask_init(struct netdev_flow_key
*mask
,
2725 const struct match
*match
)
2727 uint64_t *dst
= miniflow_values(&mask
->mf
);
2728 struct flowmap fmap
;
2732 /* Only check masks that make sense for the flow. */
2733 flow_wc_map(&match
->flow
, &fmap
);
2734 flowmap_init(&mask
->mf
.map
);
2736 FLOWMAP_FOR_EACH_INDEX(idx
, fmap
) {
2737 uint64_t mask_u64
= flow_u64_value(&match
->wc
.masks
, idx
);
2740 flowmap_set(&mask
->mf
.map
, idx
, 1);
2742 hash
= hash_add64(hash
, mask_u64
);
2748 FLOWMAP_FOR_EACH_MAP (map
, mask
->mf
.map
) {
2749 hash
= hash_add64(hash
, map
);
2752 size_t n
= dst
- miniflow_get_values(&mask
->mf
);
2754 mask
->hash
= hash_finish(hash
, n
* 8);
2755 mask
->len
= netdev_flow_key_size(n
);
2758 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
2760 netdev_flow_key_init_masked(struct netdev_flow_key
*dst
,
2761 const struct flow
*flow
,
2762 const struct netdev_flow_key
*mask
)
2764 uint64_t *dst_u64
= miniflow_values(&dst
->mf
);
2765 const uint64_t *mask_u64
= miniflow_get_values(&mask
->mf
);
2769 dst
->len
= mask
->len
;
2770 dst
->mf
= mask
->mf
; /* Copy maps. */
2772 FLOW_FOR_EACH_IN_MAPS(value
, flow
, mask
->mf
.map
) {
2773 *dst_u64
= value
& *mask_u64
++;
2774 hash
= hash_add64(hash
, *dst_u64
++);
2776 dst
->hash
= hash_finish(hash
,
2777 (dst_u64
- miniflow_get_values(&dst
->mf
)) * 8);
2781 emc_entry_alive(struct emc_entry
*ce
)
2783 return ce
->flow
&& !ce
->flow
->dead
;
2787 emc_clear_entry(struct emc_entry
*ce
)
2790 dp_netdev_flow_unref(ce
->flow
);
2796 emc_change_entry(struct emc_entry
*ce
, struct dp_netdev_flow
*flow
,
2797 const struct netdev_flow_key
*key
)
2799 if (ce
->flow
!= flow
) {
2801 dp_netdev_flow_unref(ce
->flow
);
2804 if (dp_netdev_flow_ref(flow
)) {
2811 netdev_flow_key_clone(&ce
->key
, key
);
2816 emc_insert(struct emc_cache
*cache
, const struct netdev_flow_key
*key
,
2817 struct dp_netdev_flow
*flow
)
2819 struct emc_entry
*to_be_replaced
= NULL
;
2820 struct emc_entry
*current_entry
;
2822 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2823 if (netdev_flow_key_equal(¤t_entry
->key
, key
)) {
2824 /* We found the entry with the 'mf' miniflow */
2825 emc_change_entry(current_entry
, flow
, NULL
);
2829 /* Replacement policy: put the flow in an empty (not alive) entry, or
2830 * in the first entry where it can be */
2832 || (emc_entry_alive(to_be_replaced
)
2833 && !emc_entry_alive(current_entry
))
2834 || current_entry
->key
.hash
< to_be_replaced
->key
.hash
) {
2835 to_be_replaced
= current_entry
;
2838 /* We didn't find the miniflow in the cache.
2839 * The 'to_be_replaced' entry is where the new flow will be stored */
2841 emc_change_entry(to_be_replaced
, flow
, key
);
2845 emc_probabilistic_insert(struct dp_netdev_pmd_thread
*pmd
,
2846 const struct netdev_flow_key
*key
,
2847 struct dp_netdev_flow
*flow
)
2849 /* Insert an entry into the EMC based on probability value 'min'. By
2850 * default the value is UINT32_MAX / 100 which yields an insertion
2851 * probability of 1/100 ie. 1% */
2853 uint32_t min
= pmd
->ctx
.emc_insert_min
;
2855 if (min
&& random_uint32() <= min
) {
2856 emc_insert(&(pmd
->flow_cache
).emc_cache
, key
, flow
);
2860 static inline struct dp_netdev_flow
*
2861 emc_lookup(struct emc_cache
*cache
, const struct netdev_flow_key
*key
)
2863 struct emc_entry
*current_entry
;
2865 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2866 if (current_entry
->key
.hash
== key
->hash
2867 && emc_entry_alive(current_entry
)
2868 && netdev_flow_key_equal_mf(¤t_entry
->key
, &key
->mf
)) {
2870 /* We found the entry with the 'key->mf' miniflow */
2871 return current_entry
->flow
;
2878 static inline const struct cmap_node
*
2879 smc_entry_get(struct dp_netdev_pmd_thread
*pmd
, const uint32_t hash
)
2881 struct smc_cache
*cache
= &(pmd
->flow_cache
).smc_cache
;
2882 struct smc_bucket
*bucket
= &cache
->buckets
[hash
& SMC_MASK
];
2883 uint16_t sig
= hash
>> 16;
2884 uint16_t index
= UINT16_MAX
;
2886 for (int i
= 0; i
< SMC_ENTRY_PER_BUCKET
; i
++) {
2887 if (bucket
->sig
[i
] == sig
) {
2888 index
= bucket
->flow_idx
[i
];
2892 if (index
!= UINT16_MAX
) {
2893 return cmap_find_by_index(&pmd
->flow_table
, index
);
2899 smc_clear_entry(struct smc_bucket
*b
, int idx
)
2901 b
->flow_idx
[idx
] = UINT16_MAX
;
2904 /* Insert the flow_table index into SMC. Insertion may fail when 1) SMC is
2905 * turned off, 2) the flow_table index is larger than uint16_t can handle.
2906 * If there is already an SMC entry having same signature, the index will be
2907 * updated. If there is no existing entry, but an empty entry is available,
2908 * the empty entry will be taken. If no empty entry or existing same signature,
2909 * a random entry from the hashed bucket will be picked. */
2911 smc_insert(struct dp_netdev_pmd_thread
*pmd
,
2912 const struct netdev_flow_key
*key
,
2915 struct smc_cache
*smc_cache
= &(pmd
->flow_cache
).smc_cache
;
2916 struct smc_bucket
*bucket
= &smc_cache
->buckets
[key
->hash
& SMC_MASK
];
2918 uint32_t cmap_index
;
2922 atomic_read_relaxed(&pmd
->dp
->smc_enable_db
, &smc_enable_db
);
2923 if (!smc_enable_db
) {
2927 cmap_index
= cmap_find_index(&pmd
->flow_table
, hash
);
2928 index
= (cmap_index
>= UINT16_MAX
) ? UINT16_MAX
: (uint16_t)cmap_index
;
2930 /* If the index is larger than SMC can handle (uint16_t), we don't
2932 if (index
== UINT16_MAX
) {
2936 /* If an entry with same signature already exists, update the index */
2937 uint16_t sig
= key
->hash
>> 16;
2938 for (i
= 0; i
< SMC_ENTRY_PER_BUCKET
; i
++) {
2939 if (bucket
->sig
[i
] == sig
) {
2940 bucket
->flow_idx
[i
] = index
;
2944 /* If there is an empty entry, occupy it. */
2945 for (i
= 0; i
< SMC_ENTRY_PER_BUCKET
; i
++) {
2946 if (bucket
->flow_idx
[i
] == UINT16_MAX
) {
2947 bucket
->sig
[i
] = sig
;
2948 bucket
->flow_idx
[i
] = index
;
2952 /* Otherwise, pick a random entry. */
2953 i
= random_uint32() % SMC_ENTRY_PER_BUCKET
;
2954 bucket
->sig
[i
] = sig
;
2955 bucket
->flow_idx
[i
] = index
;
2958 static struct dp_netdev_flow
*
2959 dp_netdev_pmd_lookup_flow(struct dp_netdev_pmd_thread
*pmd
,
2960 const struct netdev_flow_key
*key
,
2964 struct dpcls_rule
*rule
;
2965 odp_port_t in_port
= u32_to_odp(MINIFLOW_GET_U32(&key
->mf
,
2967 struct dp_netdev_flow
*netdev_flow
= NULL
;
2969 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2970 if (OVS_LIKELY(cls
)) {
2971 dpcls_lookup(cls
, &key
, &rule
, 1, lookup_num_p
);
2972 netdev_flow
= dp_netdev_flow_cast(rule
);
2977 static struct dp_netdev_flow
*
2978 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread
*pmd
,
2979 const ovs_u128
*ufidp
, const struct nlattr
*key
,
2982 struct dp_netdev_flow
*netdev_flow
;
2986 /* If a UFID is not provided, determine one based on the key. */
2987 if (!ufidp
&& key
&& key_len
2988 && !dpif_netdev_flow_from_nlattrs(key
, key_len
, &flow
, false)) {
2989 dpif_flow_hash(pmd
->dp
->dpif
, &flow
, sizeof flow
, &ufid
);
2994 CMAP_FOR_EACH_WITH_HASH (netdev_flow
, node
, dp_netdev_flow_hash(ufidp
),
2996 if (ovs_u128_equals(netdev_flow
->ufid
, *ufidp
)) {
3006 get_dpif_flow_stats(const struct dp_netdev_flow
*netdev_flow_
,
3007 struct dpif_flow_stats
*stats
)
3009 struct dp_netdev_flow
*netdev_flow
;
3010 unsigned long long n
;
3014 netdev_flow
= CONST_CAST(struct dp_netdev_flow
*, netdev_flow_
);
3016 atomic_read_relaxed(&netdev_flow
->stats
.packet_count
, &n
);
3017 stats
->n_packets
= n
;
3018 atomic_read_relaxed(&netdev_flow
->stats
.byte_count
, &n
);
3020 atomic_read_relaxed(&netdev_flow
->stats
.used
, &used
);
3022 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
3023 stats
->tcp_flags
= flags
;
3026 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
3027 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
3028 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
3031 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow
*netdev_flow
,
3032 struct ofpbuf
*key_buf
, struct ofpbuf
*mask_buf
,
3033 struct dpif_flow
*flow
, bool terse
)
3036 memset(flow
, 0, sizeof *flow
);
3038 struct flow_wildcards wc
;
3039 struct dp_netdev_actions
*actions
;
3041 struct odp_flow_key_parms odp_parms
= {
3042 .flow
= &netdev_flow
->flow
,
3044 .support
= dp_netdev_support
,
3047 miniflow_expand(&netdev_flow
->cr
.mask
->mf
, &wc
.masks
);
3048 /* in_port is exact matched, but we have left it out from the mask for
3049 * optimnization reasons. Add in_port back to the mask. */
3050 wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
3053 offset
= key_buf
->size
;
3054 flow
->key
= ofpbuf_tail(key_buf
);
3055 odp_flow_key_from_flow(&odp_parms
, key_buf
);
3056 flow
->key_len
= key_buf
->size
- offset
;
3059 offset
= mask_buf
->size
;
3060 flow
->mask
= ofpbuf_tail(mask_buf
);
3061 odp_parms
.key_buf
= key_buf
;
3062 odp_flow_key_from_mask(&odp_parms
, mask_buf
);
3063 flow
->mask_len
= mask_buf
->size
- offset
;
3066 actions
= dp_netdev_flow_get_actions(netdev_flow
);
3067 flow
->actions
= actions
->actions
;
3068 flow
->actions_len
= actions
->size
;
3071 flow
->ufid
= netdev_flow
->ufid
;
3072 flow
->ufid_present
= true;
3073 flow
->pmd_id
= netdev_flow
->pmd_id
;
3074 get_dpif_flow_stats(netdev_flow
, &flow
->stats
);
3076 flow
->attrs
.offloaded
= false;
3077 flow
->attrs
.dp_layer
= "ovs";
3081 dpif_netdev_mask_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
3082 const struct nlattr
*mask_key
,
3083 uint32_t mask_key_len
, const struct flow
*flow
,
3084 struct flow_wildcards
*wc
, bool probe
)
3086 enum odp_key_fitness fitness
;
3088 fitness
= odp_flow_key_to_mask(mask_key
, mask_key_len
, wc
, flow
, NULL
);
3091 /* This should not happen: it indicates that
3092 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
3093 * disagree on the acceptable form of a mask. Log the problem
3094 * as an error, with enough details to enable debugging. */
3095 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
3097 if (!VLOG_DROP_ERR(&rl
)) {
3101 odp_flow_format(key
, key_len
, mask_key
, mask_key_len
, NULL
, &s
,
3103 VLOG_ERR("internal error parsing flow mask %s (%s)",
3104 ds_cstr(&s
), odp_key_fitness_to_string(fitness
));
3116 dpif_netdev_flow_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
3117 struct flow
*flow
, bool probe
)
3119 if (odp_flow_key_to_flow(key
, key_len
, flow
, NULL
)) {
3121 /* This should not happen: it indicates that
3122 * odp_flow_key_from_flow() and odp_flow_key_to_flow() disagree on
3123 * the acceptable form of a flow. Log the problem as an error,
3124 * with enough details to enable debugging. */
3125 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
3127 if (!VLOG_DROP_ERR(&rl
)) {
3131 odp_flow_format(key
, key_len
, NULL
, 0, NULL
, &s
, true);
3132 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s
));
3140 if (flow
->ct_state
& DP_NETDEV_CS_UNSUPPORTED_MASK
) {
3148 dpif_netdev_flow_get(const struct dpif
*dpif
, const struct dpif_flow_get
*get
)
3150 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3151 struct dp_netdev_flow
*netdev_flow
;
3152 struct dp_netdev_pmd_thread
*pmd
;
3153 struct hmapx to_find
= HMAPX_INITIALIZER(&to_find
);
3154 struct hmapx_node
*node
;
3157 if (get
->pmd_id
== PMD_ID_NULL
) {
3158 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3159 if (dp_netdev_pmd_try_ref(pmd
) && !hmapx_add(&to_find
, pmd
)) {
3160 dp_netdev_pmd_unref(pmd
);
3164 pmd
= dp_netdev_get_pmd(dp
, get
->pmd_id
);
3168 hmapx_add(&to_find
, pmd
);
3171 if (!hmapx_count(&to_find
)) {
3175 HMAPX_FOR_EACH (node
, &to_find
) {
3176 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3177 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, get
->ufid
, get
->key
,
3180 dp_netdev_flow_to_dpif_flow(netdev_flow
, get
->buffer
, get
->buffer
,
3189 HMAPX_FOR_EACH (node
, &to_find
) {
3190 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3191 dp_netdev_pmd_unref(pmd
);
3194 hmapx_destroy(&to_find
);
3199 dp_netdev_get_mega_ufid(const struct match
*match
, ovs_u128
*mega_ufid
)
3201 struct flow masked_flow
;
3204 for (i
= 0; i
< sizeof(struct flow
); i
++) {
3205 ((uint8_t *)&masked_flow
)[i
] = ((uint8_t *)&match
->flow
)[i
] &
3206 ((uint8_t *)&match
->wc
)[i
];
3208 dpif_flow_hash(NULL
, &masked_flow
, sizeof(struct flow
), mega_ufid
);
3211 static struct dp_netdev_flow
*
3212 dp_netdev_flow_add(struct dp_netdev_pmd_thread
*pmd
,
3213 struct match
*match
, const ovs_u128
*ufid
,
3214 const struct nlattr
*actions
, size_t actions_len
)
3215 OVS_REQUIRES(pmd
->flow_mutex
)
3217 struct dp_netdev_flow
*flow
;
3218 struct netdev_flow_key mask
;
3221 /* Make sure in_port is exact matched before we read it. */
3222 ovs_assert(match
->wc
.masks
.in_port
.odp_port
== ODPP_NONE
);
3223 odp_port_t in_port
= match
->flow
.in_port
.odp_port
;
3225 /* As we select the dpcls based on the port number, each netdev flow
3226 * belonging to the same dpcls will have the same odp_port value.
3227 * For performance reasons we wildcard odp_port here in the mask. In the
3228 * typical case dp_hash is also wildcarded, and the resulting 8-byte
3229 * chunk {dp_hash, in_port} will be ignored by netdev_flow_mask_init() and
3230 * will not be part of the subtable mask.
3231 * This will speed up the hash computation during dpcls_lookup() because
3232 * there is one less call to hash_add64() in this case. */
3233 match
->wc
.masks
.in_port
.odp_port
= 0;
3234 netdev_flow_mask_init(&mask
, match
);
3235 match
->wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
3237 /* Make sure wc does not have metadata. */
3238 ovs_assert(!FLOWMAP_HAS_FIELD(&mask
.mf
.map
, metadata
)
3239 && !FLOWMAP_HAS_FIELD(&mask
.mf
.map
, regs
));
3241 /* Do not allocate extra space. */
3242 flow
= xmalloc(sizeof *flow
- sizeof flow
->cr
.flow
.mf
+ mask
.len
);
3243 memset(&flow
->stats
, 0, sizeof flow
->stats
);
3246 flow
->mark
= INVALID_FLOW_MARK
;
3247 *CONST_CAST(unsigned *, &flow
->pmd_id
) = pmd
->core_id
;
3248 *CONST_CAST(struct flow
*, &flow
->flow
) = match
->flow
;
3249 *CONST_CAST(ovs_u128
*, &flow
->ufid
) = *ufid
;
3250 ovs_refcount_init(&flow
->ref_cnt
);
3251 ovsrcu_set(&flow
->actions
, dp_netdev_actions_create(actions
, actions_len
));
3253 dp_netdev_get_mega_ufid(match
, CONST_CAST(ovs_u128
*, &flow
->mega_ufid
));
3254 netdev_flow_key_init_masked(&flow
->cr
.flow
, &match
->flow
, &mask
);
3256 /* Select dpcls for in_port. Relies on in_port to be exact match. */
3257 cls
= dp_netdev_pmd_find_dpcls(pmd
, in_port
);
3258 dpcls_insert(cls
, &flow
->cr
, &mask
);
3260 cmap_insert(&pmd
->flow_table
, CONST_CAST(struct cmap_node
*, &flow
->node
),
3261 dp_netdev_flow_hash(&flow
->ufid
));
3263 queue_netdev_flow_put(pmd
, flow
, match
, actions
, actions_len
);
3265 if (OVS_UNLIKELY(!VLOG_DROP_DBG((&upcall_rl
)))) {
3266 struct ds ds
= DS_EMPTY_INITIALIZER
;
3267 struct ofpbuf key_buf
, mask_buf
;
3268 struct odp_flow_key_parms odp_parms
= {
3269 .flow
= &match
->flow
,
3270 .mask
= &match
->wc
.masks
,
3271 .support
= dp_netdev_support
,
3274 ofpbuf_init(&key_buf
, 0);
3275 ofpbuf_init(&mask_buf
, 0);
3277 odp_flow_key_from_flow(&odp_parms
, &key_buf
);
3278 odp_parms
.key_buf
= &key_buf
;
3279 odp_flow_key_from_mask(&odp_parms
, &mask_buf
);
3281 ds_put_cstr(&ds
, "flow_add: ");
3282 odp_format_ufid(ufid
, &ds
);
3283 ds_put_cstr(&ds
, " ");
3284 odp_flow_format(key_buf
.data
, key_buf
.size
,
3285 mask_buf
.data
, mask_buf
.size
,
3287 ds_put_cstr(&ds
, ", actions:");
3288 format_odp_actions(&ds
, actions
, actions_len
, NULL
);
3290 VLOG_DBG("%s", ds_cstr(&ds
));
3292 ofpbuf_uninit(&key_buf
);
3293 ofpbuf_uninit(&mask_buf
);
3295 /* Add a printout of the actual match installed. */
3298 ds_put_cstr(&ds
, "flow match: ");
3299 miniflow_expand(&flow
->cr
.flow
.mf
, &m
.flow
);
3300 miniflow_expand(&flow
->cr
.mask
->mf
, &m
.wc
.masks
);
3301 memset(&m
.tun_md
, 0, sizeof m
.tun_md
);
3302 match_format(&m
, NULL
, &ds
, OFP_DEFAULT_PRIORITY
);
3304 VLOG_DBG("%s", ds_cstr(&ds
));
3313 flow_put_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
3314 struct netdev_flow_key
*key
,
3315 struct match
*match
,
3317 const struct dpif_flow_put
*put
,
3318 struct dpif_flow_stats
*stats
)
3320 struct dp_netdev_flow
*netdev_flow
;
3324 memset(stats
, 0, sizeof *stats
);
3327 ovs_mutex_lock(&pmd
->flow_mutex
);
3328 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
3330 if (put
->flags
& DPIF_FP_CREATE
) {
3331 if (cmap_count(&pmd
->flow_table
) < MAX_FLOWS
) {
3332 dp_netdev_flow_add(pmd
, match
, ufid
, put
->actions
,
3342 if (put
->flags
& DPIF_FP_MODIFY
) {
3343 struct dp_netdev_actions
*new_actions
;
3344 struct dp_netdev_actions
*old_actions
;
3346 new_actions
= dp_netdev_actions_create(put
->actions
,
3349 old_actions
= dp_netdev_flow_get_actions(netdev_flow
);
3350 ovsrcu_set(&netdev_flow
->actions
, new_actions
);
3352 queue_netdev_flow_put(pmd
, netdev_flow
, match
,
3353 put
->actions
, put
->actions_len
);
3356 get_dpif_flow_stats(netdev_flow
, stats
);
3358 if (put
->flags
& DPIF_FP_ZERO_STATS
) {
3359 /* XXX: The userspace datapath uses thread local statistics
3360 * (for flows), which should be updated only by the owning
3361 * thread. Since we cannot write on stats memory here,
3362 * we choose not to support this flag. Please note:
3363 * - This feature is currently used only by dpctl commands with
3365 * - Should the need arise, this operation can be implemented
3366 * by keeping a base value (to be update here) for each
3367 * counter, and subtracting it before outputting the stats */
3371 ovsrcu_postpone(dp_netdev_actions_free
, old_actions
);
3372 } else if (put
->flags
& DPIF_FP_CREATE
) {
3375 /* Overlapping flow. */
3379 ovs_mutex_unlock(&pmd
->flow_mutex
);
3384 dpif_netdev_flow_put(struct dpif
*dpif
, const struct dpif_flow_put
*put
)
3386 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3387 struct netdev_flow_key key
, mask
;
3388 struct dp_netdev_pmd_thread
*pmd
;
3392 bool probe
= put
->flags
& DPIF_FP_PROBE
;
3395 memset(put
->stats
, 0, sizeof *put
->stats
);
3397 error
= dpif_netdev_flow_from_nlattrs(put
->key
, put
->key_len
, &match
.flow
,
3402 error
= dpif_netdev_mask_from_nlattrs(put
->key
, put
->key_len
,
3403 put
->mask
, put
->mask_len
,
3404 &match
.flow
, &match
.wc
, probe
);
3412 dpif_flow_hash(dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
3415 /* The Netlink encoding of datapath flow keys cannot express
3416 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3417 * tag is interpreted as exact match on the fact that there is no
3418 * VLAN. Unless we refactor a lot of code that translates between
3419 * Netlink and struct flow representations, we have to do the same
3420 * here. This must be in sync with 'match' in handle_packet_upcall(). */
3421 if (!match
.wc
.masks
.vlans
[0].tci
) {
3422 match
.wc
.masks
.vlans
[0].tci
= htons(0xffff);
3425 /* Must produce a netdev_flow_key for lookup.
3426 * Use the same method as employed to create the key when adding
3427 * the flow to the dplcs to make sure they match. */
3428 netdev_flow_mask_init(&mask
, &match
);
3429 netdev_flow_key_init_masked(&key
, &match
.flow
, &mask
);
3431 if (put
->pmd_id
== PMD_ID_NULL
) {
3432 if (cmap_count(&dp
->poll_threads
) == 0) {
3435 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3436 struct dpif_flow_stats pmd_stats
;
3439 pmd_error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
,
3443 } else if (put
->stats
) {
3444 put
->stats
->n_packets
+= pmd_stats
.n_packets
;
3445 put
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
3446 put
->stats
->used
= MAX(put
->stats
->used
, pmd_stats
.used
);
3447 put
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
3451 pmd
= dp_netdev_get_pmd(dp
, put
->pmd_id
);
3455 error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
, put
->stats
);
3456 dp_netdev_pmd_unref(pmd
);
3463 flow_del_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
3464 struct dpif_flow_stats
*stats
,
3465 const struct dpif_flow_del
*del
)
3467 struct dp_netdev_flow
*netdev_flow
;
3470 ovs_mutex_lock(&pmd
->flow_mutex
);
3471 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, del
->ufid
, del
->key
,
3475 get_dpif_flow_stats(netdev_flow
, stats
);
3477 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
3481 ovs_mutex_unlock(&pmd
->flow_mutex
);
3487 dpif_netdev_flow_del(struct dpif
*dpif
, const struct dpif_flow_del
*del
)
3489 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3490 struct dp_netdev_pmd_thread
*pmd
;
3494 memset(del
->stats
, 0, sizeof *del
->stats
);
3497 if (del
->pmd_id
== PMD_ID_NULL
) {
3498 if (cmap_count(&dp
->poll_threads
) == 0) {
3501 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3502 struct dpif_flow_stats pmd_stats
;
3505 pmd_error
= flow_del_on_pmd(pmd
, &pmd_stats
, del
);
3508 } else if (del
->stats
) {
3509 del
->stats
->n_packets
+= pmd_stats
.n_packets
;
3510 del
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
3511 del
->stats
->used
= MAX(del
->stats
->used
, pmd_stats
.used
);
3512 del
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
3516 pmd
= dp_netdev_get_pmd(dp
, del
->pmd_id
);
3520 error
= flow_del_on_pmd(pmd
, del
->stats
, del
);
3521 dp_netdev_pmd_unref(pmd
);
3528 struct dpif_netdev_flow_dump
{
3529 struct dpif_flow_dump up
;
3530 struct cmap_position poll_thread_pos
;
3531 struct cmap_position flow_pos
;
3532 struct dp_netdev_pmd_thread
*cur_pmd
;
3534 struct ovs_mutex mutex
;
3537 static struct dpif_netdev_flow_dump
*
3538 dpif_netdev_flow_dump_cast(struct dpif_flow_dump
*dump
)
3540 return CONTAINER_OF(dump
, struct dpif_netdev_flow_dump
, up
);
3543 static struct dpif_flow_dump
*
3544 dpif_netdev_flow_dump_create(const struct dpif
*dpif_
, bool terse
,
3545 struct dpif_flow_dump_types
*types OVS_UNUSED
)
3547 struct dpif_netdev_flow_dump
*dump
;
3549 dump
= xzalloc(sizeof *dump
);
3550 dpif_flow_dump_init(&dump
->up
, dpif_
);
3551 dump
->up
.terse
= terse
;
3552 ovs_mutex_init(&dump
->mutex
);
3558 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump
*dump_
)
3560 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
3562 ovs_mutex_destroy(&dump
->mutex
);
3567 struct dpif_netdev_flow_dump_thread
{
3568 struct dpif_flow_dump_thread up
;
3569 struct dpif_netdev_flow_dump
*dump
;
3570 struct odputil_keybuf keybuf
[FLOW_DUMP_MAX_BATCH
];
3571 struct odputil_keybuf maskbuf
[FLOW_DUMP_MAX_BATCH
];
3574 static struct dpif_netdev_flow_dump_thread
*
3575 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread
*thread
)
3577 return CONTAINER_OF(thread
, struct dpif_netdev_flow_dump_thread
, up
);
3580 static struct dpif_flow_dump_thread
*
3581 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump
*dump_
)
3583 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
3584 struct dpif_netdev_flow_dump_thread
*thread
;
3586 thread
= xmalloc(sizeof *thread
);
3587 dpif_flow_dump_thread_init(&thread
->up
, &dump
->up
);
3588 thread
->dump
= dump
;
3593 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread
*thread_
)
3595 struct dpif_netdev_flow_dump_thread
*thread
3596 = dpif_netdev_flow_dump_thread_cast(thread_
);
3602 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread
*thread_
,
3603 struct dpif_flow
*flows
, int max_flows
)
3605 struct dpif_netdev_flow_dump_thread
*thread
3606 = dpif_netdev_flow_dump_thread_cast(thread_
);
3607 struct dpif_netdev_flow_dump
*dump
= thread
->dump
;
3608 struct dp_netdev_flow
*netdev_flows
[FLOW_DUMP_MAX_BATCH
];
3612 ovs_mutex_lock(&dump
->mutex
);
3613 if (!dump
->status
) {
3614 struct dpif_netdev
*dpif
= dpif_netdev_cast(thread
->up
.dpif
);
3615 struct dp_netdev
*dp
= get_dp_netdev(&dpif
->dpif
);
3616 struct dp_netdev_pmd_thread
*pmd
= dump
->cur_pmd
;
3617 int flow_limit
= MIN(max_flows
, FLOW_DUMP_MAX_BATCH
);
3619 /* First call to dump_next(), extracts the first pmd thread.
3620 * If there is no pmd thread, returns immediately. */
3622 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
3624 ovs_mutex_unlock(&dump
->mutex
);
3631 for (n_flows
= 0; n_flows
< flow_limit
; n_flows
++) {
3632 struct cmap_node
*node
;
3634 node
= cmap_next_position(&pmd
->flow_table
, &dump
->flow_pos
);
3638 netdev_flows
[n_flows
] = CONTAINER_OF(node
,
3639 struct dp_netdev_flow
,
3642 /* When finishing dumping the current pmd thread, moves to
3644 if (n_flows
< flow_limit
) {
3645 memset(&dump
->flow_pos
, 0, sizeof dump
->flow_pos
);
3646 dp_netdev_pmd_unref(pmd
);
3647 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
3653 /* Keeps the reference to next caller. */
3654 dump
->cur_pmd
= pmd
;
3656 /* If the current dump is empty, do not exit the loop, since the
3657 * remaining pmds could have flows to be dumped. Just dumps again
3658 * on the new 'pmd'. */
3661 ovs_mutex_unlock(&dump
->mutex
);
3663 for (i
= 0; i
< n_flows
; i
++) {
3664 struct odputil_keybuf
*maskbuf
= &thread
->maskbuf
[i
];
3665 struct odputil_keybuf
*keybuf
= &thread
->keybuf
[i
];
3666 struct dp_netdev_flow
*netdev_flow
= netdev_flows
[i
];
3667 struct dpif_flow
*f
= &flows
[i
];
3668 struct ofpbuf key
, mask
;
3670 ofpbuf_use_stack(&key
, keybuf
, sizeof *keybuf
);
3671 ofpbuf_use_stack(&mask
, maskbuf
, sizeof *maskbuf
);
3672 dp_netdev_flow_to_dpif_flow(netdev_flow
, &key
, &mask
, f
,
3680 dpif_netdev_execute(struct dpif
*dpif
, struct dpif_execute
*execute
)
3681 OVS_NO_THREAD_SAFETY_ANALYSIS
3683 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3684 struct dp_netdev_pmd_thread
*pmd
;
3685 struct dp_packet_batch pp
;
3687 if (dp_packet_size(execute
->packet
) < ETH_HEADER_LEN
||
3688 dp_packet_size(execute
->packet
) > UINT16_MAX
) {
3692 /* Tries finding the 'pmd'. If NULL is returned, that means
3693 * the current thread is a non-pmd thread and should use
3694 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
3695 pmd
= ovsthread_getspecific(dp
->per_pmd_key
);
3697 pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
3703 if (execute
->probe
) {
3704 /* If this is part of a probe, Drop the packet, since executing
3705 * the action may actually cause spurious packets be sent into
3707 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3708 dp_netdev_pmd_unref(pmd
);
3713 /* If the current thread is non-pmd thread, acquires
3714 * the 'non_pmd_mutex'. */
3715 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3716 ovs_mutex_lock(&dp
->non_pmd_mutex
);
3719 /* Update current time in PMD context. We don't care about EMC insertion
3720 * probability, because we are on a slow path. */
3721 pmd_thread_ctx_time_update(pmd
);
3723 /* The action processing expects the RSS hash to be valid, because
3724 * it's always initialized at the beginning of datapath processing.
3725 * In this case, though, 'execute->packet' may not have gone through
3726 * the datapath at all, it may have been generated by the upper layer
3727 * (OpenFlow packet-out, BFD frame, ...). */
3728 if (!dp_packet_rss_valid(execute
->packet
)) {
3729 dp_packet_set_rss_hash(execute
->packet
,
3730 flow_hash_5tuple(execute
->flow
, 0));
3733 dp_packet_batch_init_packet(&pp
, execute
->packet
);
3734 pp
.do_not_steal
= true;
3735 dp_netdev_execute_actions(pmd
, &pp
, false, execute
->flow
,
3736 execute
->actions
, execute
->actions_len
);
3737 dp_netdev_pmd_flush_output_packets(pmd
, true);
3739 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3740 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
3741 dp_netdev_pmd_unref(pmd
);
3748 dpif_netdev_operate(struct dpif
*dpif
, struct dpif_op
**ops
, size_t n_ops
,
3749 enum dpif_offload_type offload_type OVS_UNUSED
)
3753 for (i
= 0; i
< n_ops
; i
++) {
3754 struct dpif_op
*op
= ops
[i
];
3757 case DPIF_OP_FLOW_PUT
:
3758 op
->error
= dpif_netdev_flow_put(dpif
, &op
->flow_put
);
3761 case DPIF_OP_FLOW_DEL
:
3762 op
->error
= dpif_netdev_flow_del(dpif
, &op
->flow_del
);
3765 case DPIF_OP_EXECUTE
:
3766 op
->error
= dpif_netdev_execute(dpif
, &op
->execute
);
3769 case DPIF_OP_FLOW_GET
:
3770 op
->error
= dpif_netdev_flow_get(dpif
, &op
->flow_get
);
3776 /* Enable or Disable PMD auto load balancing. */
3778 set_pmd_auto_lb(struct dp_netdev
*dp
)
3780 unsigned int cnt
= 0;
3781 struct dp_netdev_pmd_thread
*pmd
;
3782 struct pmd_auto_lb
*pmd_alb
= &dp
->pmd_alb
;
3784 bool enable_alb
= false;
3785 bool multi_rxq
= false;
3786 bool pmd_rxq_assign_cyc
= dp
->pmd_rxq_assign_cyc
;
3788 /* Ensure that there is at least 2 non-isolated PMDs and
3789 * one of them is polling more than one rxq. */
3790 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3791 if (pmd
->core_id
== NON_PMD_CORE_ID
|| pmd
->isolated
) {
3795 if (hmap_count(&pmd
->poll_list
) > 1) {
3798 if (cnt
&& multi_rxq
) {
3805 /* Enable auto LB if it is requested and cycle based assignment is true. */
3806 enable_alb
= enable_alb
&& pmd_rxq_assign_cyc
&&
3807 pmd_alb
->auto_lb_requested
;
3809 if (pmd_alb
->is_enabled
!= enable_alb
) {
3810 pmd_alb
->is_enabled
= enable_alb
;
3811 if (pmd_alb
->is_enabled
) {
3812 VLOG_INFO("PMD auto load balance is enabled "
3813 "(with rebalance interval:%"PRIu64
" msec)",
3814 pmd_alb
->rebalance_intvl
);
3816 pmd_alb
->rebalance_poll_timer
= 0;
3817 VLOG_INFO("PMD auto load balance is disabled");
3823 /* Applies datapath configuration from the database. Some of the changes are
3824 * actually applied in dpif_netdev_run(). */
3826 dpif_netdev_set_config(struct dpif
*dpif
, const struct smap
*other_config
)
3828 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3829 const char *cmask
= smap_get(other_config
, "pmd-cpu-mask");
3830 const char *pmd_rxq_assign
= smap_get_def(other_config
, "pmd-rxq-assign",
3832 unsigned long long insert_prob
=
3833 smap_get_ullong(other_config
, "emc-insert-inv-prob",
3834 DEFAULT_EM_FLOW_INSERT_INV_PROB
);
3835 uint32_t insert_min
, cur_min
;
3836 uint32_t tx_flush_interval
, cur_tx_flush_interval
;
3837 uint64_t rebalance_intvl
;
3839 tx_flush_interval
= smap_get_int(other_config
, "tx-flush-interval",
3840 DEFAULT_TX_FLUSH_INTERVAL
);
3841 atomic_read_relaxed(&dp
->tx_flush_interval
, &cur_tx_flush_interval
);
3842 if (tx_flush_interval
!= cur_tx_flush_interval
) {
3843 atomic_store_relaxed(&dp
->tx_flush_interval
, tx_flush_interval
);
3844 VLOG_INFO("Flushing interval for tx queues set to %"PRIu32
" us",
3848 if (!nullable_string_is_equal(dp
->pmd_cmask
, cmask
)) {
3849 free(dp
->pmd_cmask
);
3850 dp
->pmd_cmask
= nullable_xstrdup(cmask
);
3851 dp_netdev_request_reconfigure(dp
);
3854 atomic_read_relaxed(&dp
->emc_insert_min
, &cur_min
);
3855 if (insert_prob
<= UINT32_MAX
) {
3856 insert_min
= insert_prob
== 0 ? 0 : UINT32_MAX
/ insert_prob
;
3858 insert_min
= DEFAULT_EM_FLOW_INSERT_MIN
;
3859 insert_prob
= DEFAULT_EM_FLOW_INSERT_INV_PROB
;
3862 if (insert_min
!= cur_min
) {
3863 atomic_store_relaxed(&dp
->emc_insert_min
, insert_min
);
3864 if (insert_min
== 0) {
3865 VLOG_INFO("EMC insertion probability changed to zero");
3867 VLOG_INFO("EMC insertion probability changed to 1/%llu (~%.2f%%)",
3868 insert_prob
, (100 / (float)insert_prob
));
3872 bool perf_enabled
= smap_get_bool(other_config
, "pmd-perf-metrics", false);
3873 bool cur_perf_enabled
;
3874 atomic_read_relaxed(&dp
->pmd_perf_metrics
, &cur_perf_enabled
);
3875 if (perf_enabled
!= cur_perf_enabled
) {
3876 atomic_store_relaxed(&dp
->pmd_perf_metrics
, perf_enabled
);
3878 VLOG_INFO("PMD performance metrics collection enabled");
3880 VLOG_INFO("PMD performance metrics collection disabled");
3884 bool smc_enable
= smap_get_bool(other_config
, "smc-enable", false);
3886 atomic_read_relaxed(&dp
->smc_enable_db
, &cur_smc
);
3887 if (smc_enable
!= cur_smc
) {
3888 atomic_store_relaxed(&dp
->smc_enable_db
, smc_enable
);
3890 VLOG_INFO("SMC cache is enabled");
3892 VLOG_INFO("SMC cache is disabled");
3896 bool pmd_rxq_assign_cyc
= !strcmp(pmd_rxq_assign
, "cycles");
3897 if (!pmd_rxq_assign_cyc
&& strcmp(pmd_rxq_assign
, "roundrobin")) {
3898 VLOG_WARN("Unsupported Rxq to PMD assignment mode in pmd-rxq-assign. "
3899 "Defaulting to 'cycles'.");
3900 pmd_rxq_assign_cyc
= true;
3901 pmd_rxq_assign
= "cycles";
3903 if (dp
->pmd_rxq_assign_cyc
!= pmd_rxq_assign_cyc
) {
3904 dp
->pmd_rxq_assign_cyc
= pmd_rxq_assign_cyc
;
3905 VLOG_INFO("Rxq to PMD assignment mode changed to: \'%s\'.",
3907 dp_netdev_request_reconfigure(dp
);
3910 struct pmd_auto_lb
*pmd_alb
= &dp
->pmd_alb
;
3911 pmd_alb
->auto_lb_requested
= smap_get_bool(other_config
, "pmd-auto-lb",
3914 rebalance_intvl
= smap_get_int(other_config
, "pmd-auto-lb-rebal-interval",
3915 ALB_PMD_REBALANCE_POLL_INTERVAL
);
3917 /* Input is in min, convert it to msec. */
3919 rebalance_intvl
? rebalance_intvl
* MIN_TO_MSEC
: MIN_TO_MSEC
;
3921 if (pmd_alb
->rebalance_intvl
!= rebalance_intvl
) {
3922 pmd_alb
->rebalance_intvl
= rebalance_intvl
;
3925 set_pmd_auto_lb(dp
);
3929 /* Parses affinity list and returns result in 'core_ids'. */
3931 parse_affinity_list(const char *affinity_list
, unsigned *core_ids
, int n_rxq
)
3934 char *list
, *copy
, *key
, *value
;
3937 for (i
= 0; i
< n_rxq
; i
++) {
3938 core_ids
[i
] = OVS_CORE_UNSPEC
;
3941 if (!affinity_list
) {
3945 list
= copy
= xstrdup(affinity_list
);
3947 while (ofputil_parse_key_value(&list
, &key
, &value
)) {
3948 int rxq_id
, core_id
;
3950 if (!str_to_int(key
, 0, &rxq_id
) || rxq_id
< 0
3951 || !str_to_int(value
, 0, &core_id
) || core_id
< 0) {
3956 if (rxq_id
< n_rxq
) {
3957 core_ids
[rxq_id
] = core_id
;
3965 /* Parses 'affinity_list' and applies configuration if it is valid. */
3967 dpif_netdev_port_set_rxq_affinity(struct dp_netdev_port
*port
,
3968 const char *affinity_list
)
3970 unsigned *core_ids
, i
;
3973 core_ids
= xmalloc(port
->n_rxq
* sizeof *core_ids
);
3974 if (parse_affinity_list(affinity_list
, core_ids
, port
->n_rxq
)) {
3979 for (i
= 0; i
< port
->n_rxq
; i
++) {
3980 port
->rxqs
[i
].core_id
= core_ids
[i
];
3988 /* Returns 'true' if one of the 'port's RX queues exists in 'poll_list'
3989 * of given PMD thread. */
3991 dpif_netdev_pmd_polls_port(struct dp_netdev_pmd_thread
*pmd
,
3992 struct dp_netdev_port
*port
)
3993 OVS_EXCLUDED(pmd
->port_mutex
)
3995 struct rxq_poll
*poll
;
3998 ovs_mutex_lock(&pmd
->port_mutex
);
3999 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
4000 if (port
== poll
->rxq
->port
) {
4005 ovs_mutex_unlock(&pmd
->port_mutex
);
4009 /* Updates port configuration from the database. The changes are actually
4010 * applied in dpif_netdev_run(). */
4012 dpif_netdev_port_set_config(struct dpif
*dpif
, odp_port_t port_no
,
4013 const struct smap
*cfg
)
4015 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4016 struct dp_netdev_port
*port
;
4018 const char *affinity_list
= smap_get(cfg
, "pmd-rxq-affinity");
4019 bool emc_enabled
= smap_get_bool(cfg
, "emc-enable", true);
4021 ovs_mutex_lock(&dp
->port_mutex
);
4022 error
= get_port_by_number(dp
, port_no
, &port
);
4027 if (emc_enabled
!= port
->emc_enabled
) {
4028 struct dp_netdev_pmd_thread
*pmd
;
4029 struct ds ds
= DS_EMPTY_INITIALIZER
;
4030 uint32_t cur_min
, insert_prob
;
4032 port
->emc_enabled
= emc_enabled
;
4033 /* Mark for reload all the threads that polls this port and request
4034 * for reconfiguration for the actual reloading of threads. */
4035 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4036 if (dpif_netdev_pmd_polls_port(pmd
, port
)) {
4037 pmd
->need_reload
= true;
4040 dp_netdev_request_reconfigure(dp
);
4042 ds_put_format(&ds
, "%s: EMC has been %s.",
4043 netdev_get_name(port
->netdev
),
4044 (emc_enabled
) ? "enabled" : "disabled");
4046 ds_put_cstr(&ds
, " Current insertion probability is ");
4047 atomic_read_relaxed(&dp
->emc_insert_min
, &cur_min
);
4049 ds_put_cstr(&ds
, "zero.");
4051 insert_prob
= UINT32_MAX
/ cur_min
;
4052 ds_put_format(&ds
, "1/%"PRIu32
" (~%.2f%%).",
4053 insert_prob
, 100 / (float) insert_prob
);
4056 VLOG_INFO("%s", ds_cstr(&ds
));
4060 /* Checking for RXq affinity changes. */
4061 if (!netdev_is_pmd(port
->netdev
)
4062 || nullable_string_is_equal(affinity_list
, port
->rxq_affinity_list
)) {
4066 error
= dpif_netdev_port_set_rxq_affinity(port
, affinity_list
);
4070 free(port
->rxq_affinity_list
);
4071 port
->rxq_affinity_list
= nullable_xstrdup(affinity_list
);
4073 dp_netdev_request_reconfigure(dp
);
4075 ovs_mutex_unlock(&dp
->port_mutex
);
4080 dpif_netdev_queue_to_priority(const struct dpif
*dpif OVS_UNUSED
,
4081 uint32_t queue_id
, uint32_t *priority
)
4083 *priority
= queue_id
;
4088 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
4089 * a copy of the 'size' bytes of 'actions' input parameters. */
4090 struct dp_netdev_actions
*
4091 dp_netdev_actions_create(const struct nlattr
*actions
, size_t size
)
4093 struct dp_netdev_actions
*netdev_actions
;
4095 netdev_actions
= xmalloc(sizeof *netdev_actions
+ size
);
4096 memcpy(netdev_actions
->actions
, actions
, size
);
4097 netdev_actions
->size
= size
;
4099 return netdev_actions
;
4102 struct dp_netdev_actions
*
4103 dp_netdev_flow_get_actions(const struct dp_netdev_flow
*flow
)
4105 return ovsrcu_get(struct dp_netdev_actions
*, &flow
->actions
);
4109 dp_netdev_actions_free(struct dp_netdev_actions
*actions
)
4115 dp_netdev_rxq_set_cycles(struct dp_netdev_rxq
*rx
,
4116 enum rxq_cycles_counter_type type
,
4117 unsigned long long cycles
)
4119 atomic_store_relaxed(&rx
->cycles
[type
], cycles
);
4123 dp_netdev_rxq_add_cycles(struct dp_netdev_rxq
*rx
,
4124 enum rxq_cycles_counter_type type
,
4125 unsigned long long cycles
)
4127 non_atomic_ullong_add(&rx
->cycles
[type
], cycles
);
4131 dp_netdev_rxq_get_cycles(struct dp_netdev_rxq
*rx
,
4132 enum rxq_cycles_counter_type type
)
4134 unsigned long long processing_cycles
;
4135 atomic_read_relaxed(&rx
->cycles
[type
], &processing_cycles
);
4136 return processing_cycles
;
4140 dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq
*rx
,
4141 unsigned long long cycles
)
4143 unsigned int idx
= rx
->intrvl_idx
++ % PMD_RXQ_INTERVAL_MAX
;
4144 atomic_store_relaxed(&rx
->cycles_intrvl
[idx
], cycles
);
4148 dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq
*rx
, unsigned idx
)
4150 unsigned long long processing_cycles
;
4151 atomic_read_relaxed(&rx
->cycles_intrvl
[idx
], &processing_cycles
);
4152 return processing_cycles
;
4155 #if ATOMIC_ALWAYS_LOCK_FREE_8B
4157 pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread
*pmd
)
4159 bool pmd_perf_enabled
;
4160 atomic_read_relaxed(&pmd
->dp
->pmd_perf_metrics
, &pmd_perf_enabled
);
4161 return pmd_perf_enabled
;
4164 /* If stores and reads of 64-bit integers are not atomic, the full PMD
4165 * performance metrics are not available as locked access to 64 bit
4166 * integers would be prohibitively expensive. */
4168 pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread
*pmd OVS_UNUSED
)
4175 dp_netdev_pmd_flush_output_on_port(struct dp_netdev_pmd_thread
*pmd
,
4182 struct cycle_timer timer
;
4184 uint32_t tx_flush_interval
;
4186 cycle_timer_start(&pmd
->perf_stats
, &timer
);
4188 dynamic_txqs
= p
->port
->dynamic_txqs
;
4190 tx_qid
= dpif_netdev_xps_get_tx_qid(pmd
, p
);
4192 tx_qid
= pmd
->static_tx_qid
;
4195 output_cnt
= dp_packet_batch_size(&p
->output_pkts
);
4196 ovs_assert(output_cnt
> 0);
4198 netdev_send(p
->port
->netdev
, tx_qid
, &p
->output_pkts
, dynamic_txqs
);
4199 dp_packet_batch_init(&p
->output_pkts
);
4201 /* Update time of the next flush. */
4202 atomic_read_relaxed(&pmd
->dp
->tx_flush_interval
, &tx_flush_interval
);
4203 p
->flush_time
= pmd
->ctx
.now
+ tx_flush_interval
;
4205 ovs_assert(pmd
->n_output_batches
> 0);
4206 pmd
->n_output_batches
--;
4208 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_SENT_PKTS
, output_cnt
);
4209 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_SENT_BATCHES
, 1);
4211 /* Distribute send cycles evenly among transmitted packets and assign to
4212 * their respective rx queues. */
4213 cycles
= cycle_timer_stop(&pmd
->perf_stats
, &timer
) / output_cnt
;
4214 for (i
= 0; i
< output_cnt
; i
++) {
4215 if (p
->output_pkts_rxqs
[i
]) {
4216 dp_netdev_rxq_add_cycles(p
->output_pkts_rxqs
[i
],
4217 RXQ_CYCLES_PROC_CURR
, cycles
);
4225 dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread
*pmd
,
4231 if (!pmd
->n_output_batches
) {
4235 HMAP_FOR_EACH (p
, node
, &pmd
->send_port_cache
) {
4236 if (!dp_packet_batch_is_empty(&p
->output_pkts
)
4237 && (force
|| pmd
->ctx
.now
>= p
->flush_time
)) {
4238 output_cnt
+= dp_netdev_pmd_flush_output_on_port(pmd
, p
);
4245 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread
*pmd
,
4246 struct dp_netdev_rxq
*rxq
,
4249 struct pmd_perf_stats
*s
= &pmd
->perf_stats
;
4250 struct dp_packet_batch batch
;
4251 struct cycle_timer timer
;
4254 int rem_qlen
= 0, *qlen_p
= NULL
;
4257 /* Measure duration for polling and processing rx burst. */
4258 cycle_timer_start(&pmd
->perf_stats
, &timer
);
4260 pmd
->ctx
.last_rxq
= rxq
;
4261 dp_packet_batch_init(&batch
);
4263 /* Fetch the rx queue length only for vhostuser ports. */
4264 if (pmd_perf_metrics_enabled(pmd
) && rxq
->is_vhost
) {
4268 error
= netdev_rxq_recv(rxq
->rx
, &batch
, qlen_p
);
4270 /* At least one packet received. */
4271 *recirc_depth_get() = 0;
4272 pmd_thread_ctx_time_update(pmd
);
4273 batch_cnt
= batch
.count
;
4274 if (pmd_perf_metrics_enabled(pmd
)) {
4275 /* Update batch histogram. */
4276 s
->current
.batches
++;
4277 histogram_add_sample(&s
->pkts_per_batch
, batch_cnt
);
4278 /* Update the maximum vhost rx queue fill level. */
4279 if (rxq
->is_vhost
&& rem_qlen
>= 0) {
4280 uint32_t qfill
= batch_cnt
+ rem_qlen
;
4281 if (qfill
> s
->current
.max_vhost_qfill
) {
4282 s
->current
.max_vhost_qfill
= qfill
;
4286 /* Process packet batch. */
4287 dp_netdev_input(pmd
, &batch
, port_no
);
4289 /* Assign processing cycles to rx queue. */
4290 cycles
= cycle_timer_stop(&pmd
->perf_stats
, &timer
);
4291 dp_netdev_rxq_add_cycles(rxq
, RXQ_CYCLES_PROC_CURR
, cycles
);
4293 dp_netdev_pmd_flush_output_packets(pmd
, false);
4295 /* Discard cycles. */
4296 cycle_timer_stop(&pmd
->perf_stats
, &timer
);
4297 if (error
!= EAGAIN
&& error
!= EOPNOTSUPP
) {
4298 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
4300 VLOG_ERR_RL(&rl
, "error receiving data from %s: %s",
4301 netdev_rxq_get_name(rxq
->rx
), ovs_strerror(error
));
4305 pmd
->ctx
.last_rxq
= NULL
;
4310 static struct tx_port
*
4311 tx_port_lookup(const struct hmap
*hmap
, odp_port_t port_no
)
4315 HMAP_FOR_EACH_IN_BUCKET (tx
, node
, hash_port_no(port_no
), hmap
) {
4316 if (tx
->port
->port_no
== port_no
) {
4325 port_reconfigure(struct dp_netdev_port
*port
)
4327 struct netdev
*netdev
= port
->netdev
;
4330 /* Closes the existing 'rxq's. */
4331 for (i
= 0; i
< port
->n_rxq
; i
++) {
4332 netdev_rxq_close(port
->rxqs
[i
].rx
);
4333 port
->rxqs
[i
].rx
= NULL
;
4335 unsigned last_nrxq
= port
->n_rxq
;
4338 /* Allows 'netdev' to apply the pending configuration changes. */
4339 if (netdev_is_reconf_required(netdev
) || port
->need_reconfigure
) {
4340 err
= netdev_reconfigure(netdev
);
4341 if (err
&& (err
!= EOPNOTSUPP
)) {
4342 VLOG_ERR("Failed to set interface %s new configuration",
4343 netdev_get_name(netdev
));
4347 /* If the netdev_reconfigure() above succeeds, reopens the 'rxq's. */
4348 port
->rxqs
= xrealloc(port
->rxqs
,
4349 sizeof *port
->rxqs
* netdev_n_rxq(netdev
));
4350 /* Realloc 'used' counters for tx queues. */
4351 free(port
->txq_used
);
4352 port
->txq_used
= xcalloc(netdev_n_txq(netdev
), sizeof *port
->txq_used
);
4354 for (i
= 0; i
< netdev_n_rxq(netdev
); i
++) {
4355 bool new_queue
= i
>= last_nrxq
;
4357 memset(&port
->rxqs
[i
], 0, sizeof port
->rxqs
[i
]);
4360 port
->rxqs
[i
].port
= port
;
4361 port
->rxqs
[i
].is_vhost
= !strncmp(port
->type
, "dpdkvhost", 9);
4363 err
= netdev_rxq_open(netdev
, &port
->rxqs
[i
].rx
, i
);
4370 /* Parse affinity list to apply configuration for new queues. */
4371 dpif_netdev_port_set_rxq_affinity(port
, port
->rxq_affinity_list
);
4373 /* If reconfiguration was successful mark it as such, so we can use it */
4374 port
->need_reconfigure
= false;
4379 struct rr_numa_list
{
4380 struct hmap numas
; /* Contains 'struct rr_numa' */
4384 struct hmap_node node
;
4388 /* Non isolated pmds on numa node 'numa_id' */
4389 struct dp_netdev_pmd_thread
**pmds
;
4396 static struct rr_numa
*
4397 rr_numa_list_lookup(struct rr_numa_list
*rr
, int numa_id
)
4399 struct rr_numa
*numa
;
4401 HMAP_FOR_EACH_WITH_HASH (numa
, node
, hash_int(numa_id
, 0), &rr
->numas
) {
4402 if (numa
->numa_id
== numa_id
) {
4410 /* Returns the next node in numa list following 'numa' in round-robin fashion.
4411 * Returns first node if 'numa' is a null pointer or the last node in 'rr'.
4412 * Returns NULL if 'rr' numa list is empty. */
4413 static struct rr_numa
*
4414 rr_numa_list_next(struct rr_numa_list
*rr
, const struct rr_numa
*numa
)
4416 struct hmap_node
*node
= NULL
;
4419 node
= hmap_next(&rr
->numas
, &numa
->node
);
4422 node
= hmap_first(&rr
->numas
);
4425 return (node
) ? CONTAINER_OF(node
, struct rr_numa
, node
) : NULL
;
4429 rr_numa_list_populate(struct dp_netdev
*dp
, struct rr_numa_list
*rr
)
4431 struct dp_netdev_pmd_thread
*pmd
;
4432 struct rr_numa
*numa
;
4434 hmap_init(&rr
->numas
);
4436 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4437 if (pmd
->core_id
== NON_PMD_CORE_ID
|| pmd
->isolated
) {
4441 numa
= rr_numa_list_lookup(rr
, pmd
->numa_id
);
4443 numa
= xzalloc(sizeof *numa
);
4444 numa
->numa_id
= pmd
->numa_id
;
4445 hmap_insert(&rr
->numas
, &numa
->node
, hash_int(pmd
->numa_id
, 0));
4448 numa
->pmds
= xrealloc(numa
->pmds
, numa
->n_pmds
* sizeof *numa
->pmds
);
4449 numa
->pmds
[numa
->n_pmds
- 1] = pmd
;
4450 /* At least one pmd so initialise curr_idx and idx_inc. */
4451 numa
->cur_index
= 0;
4452 numa
->idx_inc
= true;
4457 * Returns the next pmd from the numa node.
4459 * If 'updown' is 'true' it will alternate between selecting the next pmd in
4460 * either an up or down walk, switching between up/down when the first or last
4461 * core is reached. e.g. 1,2,3,3,2,1,1,2...
4463 * If 'updown' is 'false' it will select the next pmd wrapping around when last
4464 * core reached. e.g. 1,2,3,1,2,3,1,2...
4466 static struct dp_netdev_pmd_thread
*
4467 rr_numa_get_pmd(struct rr_numa
*numa
, bool updown
)
4469 int numa_idx
= numa
->cur_index
;
4471 if (numa
->idx_inc
== true) {
4472 /* Incrementing through list of pmds. */
4473 if (numa
->cur_index
== numa
->n_pmds
-1) {
4474 /* Reached the last pmd. */
4476 numa
->idx_inc
= false;
4478 numa
->cur_index
= 0;
4484 /* Decrementing through list of pmds. */
4485 if (numa
->cur_index
== 0) {
4486 /* Reached the first pmd. */
4487 numa
->idx_inc
= true;
4492 return numa
->pmds
[numa_idx
];
4496 rr_numa_list_destroy(struct rr_numa_list
*rr
)
4498 struct rr_numa
*numa
;
4500 HMAP_FOR_EACH_POP (numa
, node
, &rr
->numas
) {
4504 hmap_destroy(&rr
->numas
);
4507 /* Sort Rx Queues by the processing cycles they are consuming. */
4509 compare_rxq_cycles(const void *a
, const void *b
)
4511 struct dp_netdev_rxq
*qa
;
4512 struct dp_netdev_rxq
*qb
;
4513 uint64_t cycles_qa
, cycles_qb
;
4515 qa
= *(struct dp_netdev_rxq
**) a
;
4516 qb
= *(struct dp_netdev_rxq
**) b
;
4518 cycles_qa
= dp_netdev_rxq_get_cycles(qa
, RXQ_CYCLES_PROC_HIST
);
4519 cycles_qb
= dp_netdev_rxq_get_cycles(qb
, RXQ_CYCLES_PROC_HIST
);
4521 if (cycles_qa
!= cycles_qb
) {
4522 return (cycles_qa
< cycles_qb
) ? 1 : -1;
4524 /* Cycles are the same so tiebreak on port/queue id.
4525 * Tiebreaking (as opposed to return 0) ensures consistent
4526 * sort results across multiple OS's. */
4527 uint32_t port_qa
= odp_to_u32(qa
->port
->port_no
);
4528 uint32_t port_qb
= odp_to_u32(qb
->port
->port_no
);
4529 if (port_qa
!= port_qb
) {
4530 return port_qa
> port_qb
? 1 : -1;
4532 return netdev_rxq_get_queue_id(qa
->rx
)
4533 - netdev_rxq_get_queue_id(qb
->rx
);
4538 /* Assign pmds to queues. If 'pinned' is true, assign pmds to pinned
4539 * queues and marks the pmds as isolated. Otherwise, assign non isolated
4540 * pmds to unpinned queues.
4542 * The function doesn't touch the pmd threads, it just stores the assignment
4543 * in the 'pmd' member of each rxq. */
4545 rxq_scheduling(struct dp_netdev
*dp
, bool pinned
) OVS_REQUIRES(dp
->port_mutex
)
4547 struct dp_netdev_port
*port
;
4548 struct rr_numa_list rr
;
4549 struct rr_numa
*non_local_numa
= NULL
;
4550 struct dp_netdev_rxq
** rxqs
= NULL
;
4552 struct rr_numa
*numa
= NULL
;
4554 bool assign_cyc
= dp
->pmd_rxq_assign_cyc
;
4556 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4557 if (!netdev_is_pmd(port
->netdev
)) {
4561 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
4562 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
4564 if (pinned
&& q
->core_id
!= OVS_CORE_UNSPEC
) {
4565 struct dp_netdev_pmd_thread
*pmd
;
4567 pmd
= dp_netdev_get_pmd(dp
, q
->core_id
);
4569 VLOG_WARN("There is no PMD thread on core %d. Queue "
4570 "%d on port \'%s\' will not be polled.",
4571 q
->core_id
, qid
, netdev_get_name(port
->netdev
));
4574 pmd
->isolated
= true;
4575 dp_netdev_pmd_unref(pmd
);
4577 } else if (!pinned
&& q
->core_id
== OVS_CORE_UNSPEC
) {
4578 uint64_t cycle_hist
= 0;
4581 rxqs
= xmalloc(sizeof *rxqs
);
4583 rxqs
= xrealloc(rxqs
, sizeof *rxqs
* (n_rxqs
+ 1));
4587 /* Sum the queue intervals and store the cycle history. */
4588 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
4589 cycle_hist
+= dp_netdev_rxq_get_intrvl_cycles(q
, i
);
4591 dp_netdev_rxq_set_cycles(q
, RXQ_CYCLES_PROC_HIST
,
4594 /* Store the queue. */
4600 if (n_rxqs
> 1 && assign_cyc
) {
4601 /* Sort the queues in order of the processing cycles
4602 * they consumed during their last pmd interval. */
4603 qsort(rxqs
, n_rxqs
, sizeof *rxqs
, compare_rxq_cycles
);
4606 rr_numa_list_populate(dp
, &rr
);
4607 /* Assign the sorted queues to pmds in round robin. */
4608 for (int i
= 0; i
< n_rxqs
; i
++) {
4609 numa_id
= netdev_get_numa_id(rxqs
[i
]->port
->netdev
);
4610 numa
= rr_numa_list_lookup(&rr
, numa_id
);
4612 /* There are no pmds on the queue's local NUMA node.
4613 Round robin on the NUMA nodes that do have pmds. */
4614 non_local_numa
= rr_numa_list_next(&rr
, non_local_numa
);
4615 if (!non_local_numa
) {
4616 VLOG_ERR("There is no available (non-isolated) pmd "
4617 "thread for port \'%s\' queue %d. This queue "
4618 "will not be polled. Is pmd-cpu-mask set to "
4619 "zero? Or are all PMDs isolated to other "
4620 "queues?", netdev_rxq_get_name(rxqs
[i
]->rx
),
4621 netdev_rxq_get_queue_id(rxqs
[i
]->rx
));
4624 rxqs
[i
]->pmd
= rr_numa_get_pmd(non_local_numa
, assign_cyc
);
4625 VLOG_WARN("There's no available (non-isolated) pmd thread "
4626 "on numa node %d. Queue %d on port \'%s\' will "
4627 "be assigned to the pmd on core %d "
4628 "(numa node %d). Expect reduced performance.",
4629 numa_id
, netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
4630 netdev_rxq_get_name(rxqs
[i
]->rx
),
4631 rxqs
[i
]->pmd
->core_id
, rxqs
[i
]->pmd
->numa_id
);
4633 rxqs
[i
]->pmd
= rr_numa_get_pmd(numa
, assign_cyc
);
4635 VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
4637 "(measured processing cycles %"PRIu64
").",
4638 rxqs
[i
]->pmd
->core_id
, numa_id
,
4639 netdev_rxq_get_name(rxqs
[i
]->rx
),
4640 netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
4641 dp_netdev_rxq_get_cycles(rxqs
[i
],
4642 RXQ_CYCLES_PROC_HIST
));
4644 VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
4645 "rx queue %d.", rxqs
[i
]->pmd
->core_id
, numa_id
,
4646 netdev_rxq_get_name(rxqs
[i
]->rx
),
4647 netdev_rxq_get_queue_id(rxqs
[i
]->rx
));
4652 rr_numa_list_destroy(&rr
);
4657 reload_affected_pmds(struct dp_netdev
*dp
)
4659 struct dp_netdev_pmd_thread
*pmd
;
4661 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4662 if (pmd
->need_reload
) {
4663 flow_mark_flush(pmd
);
4664 dp_netdev_reload_pmd__(pmd
);
4668 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4669 if (pmd
->need_reload
) {
4670 if (pmd
->core_id
!= NON_PMD_CORE_ID
) {
4674 atomic_read_explicit(&pmd
->reload
, &reload
,
4675 memory_order_acquire
);
4678 pmd
->need_reload
= false;
4684 reconfigure_pmd_threads(struct dp_netdev
*dp
)
4685 OVS_REQUIRES(dp
->port_mutex
)
4687 struct dp_netdev_pmd_thread
*pmd
;
4688 struct ovs_numa_dump
*pmd_cores
;
4689 struct ovs_numa_info_core
*core
;
4690 struct hmapx to_delete
= HMAPX_INITIALIZER(&to_delete
);
4691 struct hmapx_node
*node
;
4692 bool changed
= false;
4693 bool need_to_adjust_static_tx_qids
= false;
4695 /* The pmd threads should be started only if there's a pmd port in the
4696 * datapath. If the user didn't provide any "pmd-cpu-mask", we start
4697 * NR_PMD_THREADS per numa node. */
4698 if (!has_pmd_port(dp
)) {
4699 pmd_cores
= ovs_numa_dump_n_cores_per_numa(0);
4700 } else if (dp
->pmd_cmask
&& dp
->pmd_cmask
[0]) {
4701 pmd_cores
= ovs_numa_dump_cores_with_cmask(dp
->pmd_cmask
);
4703 pmd_cores
= ovs_numa_dump_n_cores_per_numa(NR_PMD_THREADS
);
4706 /* We need to adjust 'static_tx_qid's only if we're reducing number of
4707 * PMD threads. Otherwise, new threads will allocate all the freed ids. */
4708 if (ovs_numa_dump_count(pmd_cores
) < cmap_count(&dp
->poll_threads
) - 1) {
4709 /* Adjustment is required to keep 'static_tx_qid's sequential and
4710 * avoid possible issues, for example, imbalanced tx queue usage
4711 * and unnecessary locking caused by remapping on netdev level. */
4712 need_to_adjust_static_tx_qids
= true;
4715 /* Check for unwanted pmd threads */
4716 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4717 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
4720 if (!ovs_numa_dump_contains_core(pmd_cores
, pmd
->numa_id
,
4722 hmapx_add(&to_delete
, pmd
);
4723 } else if (need_to_adjust_static_tx_qids
) {
4724 atomic_store_relaxed(&pmd
->reload_tx_qid
, true);
4725 pmd
->need_reload
= true;
4729 HMAPX_FOR_EACH (node
, &to_delete
) {
4730 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
4731 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d destroyed.",
4732 pmd
->numa_id
, pmd
->core_id
);
4733 dp_netdev_del_pmd(dp
, pmd
);
4735 changed
= !hmapx_is_empty(&to_delete
);
4736 hmapx_destroy(&to_delete
);
4738 if (need_to_adjust_static_tx_qids
) {
4739 /* 'static_tx_qid's are not sequential now.
4740 * Reload remaining threads to fix this. */
4741 reload_affected_pmds(dp
);
4744 /* Check for required new pmd threads */
4745 FOR_EACH_CORE_ON_DUMP(core
, pmd_cores
) {
4746 pmd
= dp_netdev_get_pmd(dp
, core
->core_id
);
4748 struct ds name
= DS_EMPTY_INITIALIZER
;
4750 pmd
= xzalloc(sizeof *pmd
);
4751 dp_netdev_configure_pmd(pmd
, dp
, core
->core_id
, core
->numa_id
);
4753 ds_put_format(&name
, "pmd-c%02d/id:", core
->core_id
);
4754 pmd
->thread
= ovs_thread_create(ds_cstr(&name
),
4755 pmd_thread_main
, pmd
);
4758 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d created.",
4759 pmd
->numa_id
, pmd
->core_id
);
4762 dp_netdev_pmd_unref(pmd
);
4767 struct ovs_numa_info_numa
*numa
;
4769 /* Log the number of pmd threads per numa node. */
4770 FOR_EACH_NUMA_ON_DUMP (numa
, pmd_cores
) {
4771 VLOG_INFO("There are %"PRIuSIZE
" pmd threads on numa node %d",
4772 numa
->n_cores
, numa
->numa_id
);
4776 ovs_numa_dump_destroy(pmd_cores
);
4780 pmd_remove_stale_ports(struct dp_netdev
*dp
,
4781 struct dp_netdev_pmd_thread
*pmd
)
4782 OVS_EXCLUDED(pmd
->port_mutex
)
4783 OVS_REQUIRES(dp
->port_mutex
)
4785 struct rxq_poll
*poll
, *poll_next
;
4786 struct tx_port
*tx
, *tx_next
;
4788 ovs_mutex_lock(&pmd
->port_mutex
);
4789 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
4790 struct dp_netdev_port
*port
= poll
->rxq
->port
;
4792 if (port
->need_reconfigure
4793 || !hmap_contains(&dp
->ports
, &port
->node
)) {
4794 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
4797 HMAP_FOR_EACH_SAFE (tx
, tx_next
, node
, &pmd
->tx_ports
) {
4798 struct dp_netdev_port
*port
= tx
->port
;
4800 if (port
->need_reconfigure
4801 || !hmap_contains(&dp
->ports
, &port
->node
)) {
4802 dp_netdev_del_port_tx_from_pmd(pmd
, tx
);
4805 ovs_mutex_unlock(&pmd
->port_mutex
);
4808 /* Must be called each time a port is added/removed or the cmask changes.
4809 * This creates and destroys pmd threads, reconfigures ports, opens their
4810 * rxqs and assigns all rxqs/txqs to pmd threads. */
4812 reconfigure_datapath(struct dp_netdev
*dp
)
4813 OVS_REQUIRES(dp
->port_mutex
)
4815 struct hmapx busy_threads
= HMAPX_INITIALIZER(&busy_threads
);
4816 struct dp_netdev_pmd_thread
*pmd
;
4817 struct dp_netdev_port
*port
;
4820 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
4822 /* Step 1: Adjust the pmd threads based on the datapath ports, the cores
4823 * on the system and the user configuration. */
4824 reconfigure_pmd_threads(dp
);
4826 wanted_txqs
= cmap_count(&dp
->poll_threads
);
4828 /* The number of pmd threads might have changed, or a port can be new:
4829 * adjust the txqs. */
4830 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4831 netdev_set_tx_multiq(port
->netdev
, wanted_txqs
);
4834 /* Step 2: Remove from the pmd threads ports that have been removed or
4835 * need reconfiguration. */
4837 /* Check for all the ports that need reconfiguration. We cache this in
4838 * 'port->need_reconfigure', because netdev_is_reconf_required() can
4839 * change at any time. */
4840 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4841 if (netdev_is_reconf_required(port
->netdev
)) {
4842 port
->need_reconfigure
= true;
4846 /* Remove from the pmd threads all the ports that have been deleted or
4847 * need reconfiguration. */
4848 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4849 pmd_remove_stale_ports(dp
, pmd
);
4852 /* Reload affected pmd threads. We must wait for the pmd threads before
4853 * reconfiguring the ports, because a port cannot be reconfigured while
4854 * it's being used. */
4855 reload_affected_pmds(dp
);
4857 /* Step 3: Reconfigure ports. */
4859 /* We only reconfigure the ports that we determined above, because they're
4860 * not being used by any pmd thread at the moment. If a port fails to
4861 * reconfigure we remove it from the datapath. */
4862 struct dp_netdev_port
*next_port
;
4863 HMAP_FOR_EACH_SAFE (port
, next_port
, node
, &dp
->ports
) {
4866 if (!port
->need_reconfigure
) {
4870 err
= port_reconfigure(port
);
4872 hmap_remove(&dp
->ports
, &port
->node
);
4873 seq_change(dp
->port_seq
);
4876 port
->dynamic_txqs
= netdev_n_txq(port
->netdev
) < wanted_txqs
;
4880 /* Step 4: Compute new rxq scheduling. We don't touch the pmd threads
4881 * for now, we just update the 'pmd' pointer in each rxq to point to the
4882 * wanted thread according to the scheduling policy. */
4884 /* Reset all the pmd threads to non isolated. */
4885 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4886 pmd
->isolated
= false;
4889 /* Reset all the queues to unassigned */
4890 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4891 for (int i
= 0; i
< port
->n_rxq
; i
++) {
4892 port
->rxqs
[i
].pmd
= NULL
;
4896 /* Add pinned queues and mark pmd threads isolated. */
4897 rxq_scheduling(dp
, true);
4899 /* Add non-pinned queues. */
4900 rxq_scheduling(dp
, false);
4902 /* Step 5: Remove queues not compliant with new scheduling. */
4904 /* Count all the threads that will have at least one queue to poll. */
4905 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4906 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
4907 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
4910 hmapx_add(&busy_threads
, q
->pmd
);
4915 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4916 struct rxq_poll
*poll
, *poll_next
;
4918 ovs_mutex_lock(&pmd
->port_mutex
);
4919 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
4920 if (poll
->rxq
->pmd
!= pmd
) {
4921 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
4923 /* This pmd might sleep after this step if it has no rxq
4924 * remaining. Tell it to busy wait for new assignment if it
4925 * has at least one scheduled queue. */
4926 if (hmap_count(&pmd
->poll_list
) == 0 &&
4927 hmapx_contains(&busy_threads
, pmd
)) {
4928 atomic_store_relaxed(&pmd
->wait_for_reload
, true);
4932 ovs_mutex_unlock(&pmd
->port_mutex
);
4935 hmapx_destroy(&busy_threads
);
4937 /* Reload affected pmd threads. We must wait for the pmd threads to remove
4938 * the old queues before readding them, otherwise a queue can be polled by
4939 * two threads at the same time. */
4940 reload_affected_pmds(dp
);
4942 /* Step 6: Add queues from scheduling, if they're not there already. */
4943 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4944 if (!netdev_is_pmd(port
->netdev
)) {
4948 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
4949 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
4952 ovs_mutex_lock(&q
->pmd
->port_mutex
);
4953 dp_netdev_add_rxq_to_pmd(q
->pmd
, q
);
4954 ovs_mutex_unlock(&q
->pmd
->port_mutex
);
4959 /* Add every port to the tx cache of every pmd thread, if it's not
4960 * there already and if this pmd has at least one rxq to poll. */
4961 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4962 ovs_mutex_lock(&pmd
->port_mutex
);
4963 if (hmap_count(&pmd
->poll_list
) || pmd
->core_id
== NON_PMD_CORE_ID
) {
4964 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4965 dp_netdev_add_port_tx_to_pmd(pmd
, port
);
4968 ovs_mutex_unlock(&pmd
->port_mutex
);
4971 /* Reload affected pmd threads. */
4972 reload_affected_pmds(dp
);
4974 /* Check if PMD Auto LB is to be enabled */
4975 set_pmd_auto_lb(dp
);
4978 /* Returns true if one of the netdevs in 'dp' requires a reconfiguration */
4980 ports_require_restart(const struct dp_netdev
*dp
)
4981 OVS_REQUIRES(dp
->port_mutex
)
4983 struct dp_netdev_port
*port
;
4985 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
4986 if (netdev_is_reconf_required(port
->netdev
)) {
4994 /* Calculates variance in the values stored in array 'a'. 'n' is the number
4995 * of elements in array to be considered for calculating vairance.
4996 * Usage example: data array 'a' contains the processing load of each pmd and
4997 * 'n' is the number of PMDs. It returns the variance in processing load of
5000 variance(uint64_t a
[], int n
)
5002 /* Compute mean (average of elements). */
5005 uint64_t sqDiff
= 0;
5011 for (int i
= 0; i
< n
; i
++) {
5018 /* Compute sum squared differences with mean. */
5019 for (int i
= 0; i
< n
; i
++) {
5020 sqDiff
+= (a
[i
] - mean
)*(a
[i
] - mean
);
5023 return (sqDiff
? (sqDiff
/ n
) : 0);
5027 /* Returns the variance in the PMDs usage as part of dry run of rxqs
5028 * assignment to PMDs. */
5030 get_dry_run_variance(struct dp_netdev
*dp
, uint32_t *core_list
,
5031 uint32_t num_pmds
, uint64_t *predicted_variance
)
5032 OVS_REQUIRES(dp
->port_mutex
)
5034 struct dp_netdev_port
*port
;
5035 struct dp_netdev_pmd_thread
*pmd
;
5036 struct dp_netdev_rxq
**rxqs
= NULL
;
5037 struct rr_numa
*numa
= NULL
;
5038 struct rr_numa_list rr
;
5041 uint64_t *pmd_usage
;
5043 if (!predicted_variance
) {
5047 pmd_usage
= xcalloc(num_pmds
, sizeof(uint64_t));
5049 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5050 if (!netdev_is_pmd(port
->netdev
)) {
5054 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
5055 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
5056 uint64_t cycle_hist
= 0;
5058 if (q
->pmd
->isolated
) {
5063 rxqs
= xmalloc(sizeof *rxqs
);
5065 rxqs
= xrealloc(rxqs
, sizeof *rxqs
* (n_rxqs
+ 1));
5068 /* Sum the queue intervals and store the cycle history. */
5069 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
5070 cycle_hist
+= dp_netdev_rxq_get_intrvl_cycles(q
, i
);
5072 dp_netdev_rxq_set_cycles(q
, RXQ_CYCLES_PROC_HIST
,
5074 /* Store the queue. */
5079 /* Sort the queues in order of the processing cycles
5080 * they consumed during their last pmd interval. */
5081 qsort(rxqs
, n_rxqs
, sizeof *rxqs
, compare_rxq_cycles
);
5083 rr_numa_list_populate(dp
, &rr
);
5085 for (int i
= 0; i
< n_rxqs
; i
++) {
5086 int numa_id
= netdev_get_numa_id(rxqs
[i
]->port
->netdev
);
5087 numa
= rr_numa_list_lookup(&rr
, numa_id
);
5089 /* Abort if cross NUMA polling. */
5090 VLOG_DBG("PMD auto lb dry run."
5091 " Aborting due to cross-numa polling.");
5095 pmd
= rr_numa_get_pmd(numa
, true);
5096 VLOG_DBG("PMD auto lb dry run. Predicted: Core %d on numa node %d "
5097 "to be assigned port \'%s\' rx queue %d "
5098 "(measured processing cycles %"PRIu64
").",
5099 pmd
->core_id
, numa_id
,
5100 netdev_rxq_get_name(rxqs
[i
]->rx
),
5101 netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
5102 dp_netdev_rxq_get_cycles(rxqs
[i
], RXQ_CYCLES_PROC_HIST
));
5104 for (int id
= 0; id
< num_pmds
; id
++) {
5105 if (pmd
->core_id
== core_list
[id
]) {
5106 /* Add the processing cycles of rxq to pmd polling it. */
5107 pmd_usage
[id
] += dp_netdev_rxq_get_cycles(rxqs
[i
],
5108 RXQ_CYCLES_PROC_HIST
);
5113 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
5114 uint64_t total_cycles
= 0;
5116 if ((pmd
->core_id
== NON_PMD_CORE_ID
) || pmd
->isolated
) {
5120 /* Get the total pmd cycles for an interval. */
5121 atomic_read_relaxed(&pmd
->intrvl_cycles
, &total_cycles
);
5122 /* Estimate the cycles to cover all intervals. */
5123 total_cycles
*= PMD_RXQ_INTERVAL_MAX
;
5124 for (int id
= 0; id
< num_pmds
; id
++) {
5125 if (pmd
->core_id
== core_list
[id
]) {
5126 if (pmd_usage
[id
]) {
5127 pmd_usage
[id
] = (pmd_usage
[id
] * 100) / total_cycles
;
5129 VLOG_DBG("PMD auto lb dry run. Predicted: Core %d, "
5130 "usage %"PRIu64
"", pmd
->core_id
, pmd_usage
[id
]);
5134 *predicted_variance
= variance(pmd_usage
, num_pmds
);
5138 rr_numa_list_destroy(&rr
);
5144 /* Does the dry run of Rxq assignment to PMDs and returns true if it gives
5145 * better distribution of load on PMDs. */
5147 pmd_rebalance_dry_run(struct dp_netdev
*dp
)
5148 OVS_REQUIRES(dp
->port_mutex
)
5150 struct dp_netdev_pmd_thread
*pmd
;
5151 uint64_t *curr_pmd_usage
;
5153 uint64_t curr_variance
;
5154 uint64_t new_variance
;
5155 uint64_t improvement
= 0;
5157 uint32_t *pmd_corelist
;
5158 struct rxq_poll
*poll
;
5161 num_pmds
= cmap_count(&dp
->poll_threads
);
5164 curr_pmd_usage
= xcalloc(num_pmds
, sizeof(uint64_t));
5165 pmd_corelist
= xcalloc(num_pmds
, sizeof(uint32_t));
5171 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
5172 uint64_t total_cycles
= 0;
5173 uint64_t total_proc
= 0;
5175 if ((pmd
->core_id
== NON_PMD_CORE_ID
) || pmd
->isolated
) {
5179 /* Get the total pmd cycles for an interval. */
5180 atomic_read_relaxed(&pmd
->intrvl_cycles
, &total_cycles
);
5181 /* Estimate the cycles to cover all intervals. */
5182 total_cycles
*= PMD_RXQ_INTERVAL_MAX
;
5184 ovs_mutex_lock(&pmd
->port_mutex
);
5185 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
5186 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
5187 total_proc
+= dp_netdev_rxq_get_intrvl_cycles(poll
->rxq
, i
);
5190 ovs_mutex_unlock(&pmd
->port_mutex
);
5193 curr_pmd_usage
[num_pmds
] = (total_proc
* 100) / total_cycles
;
5196 VLOG_DBG("PMD auto lb dry run. Current: Core %d, usage %"PRIu64
"",
5197 pmd
->core_id
, curr_pmd_usage
[num_pmds
]);
5199 if (atomic_count_get(&pmd
->pmd_overloaded
)) {
5200 atomic_count_set(&pmd
->pmd_overloaded
, 0);
5203 pmd_corelist
[num_pmds
] = pmd
->core_id
;
5207 curr_variance
= variance(curr_pmd_usage
, num_pmds
);
5208 ret
= get_dry_run_variance(dp
, pmd_corelist
, num_pmds
, &new_variance
);
5211 VLOG_DBG("PMD auto lb dry run. Current PMD variance: %"PRIu64
","
5212 " Predicted PMD variance: %"PRIu64
"",
5213 curr_variance
, new_variance
);
5215 if (new_variance
< curr_variance
) {
5217 ((curr_variance
- new_variance
) * 100) / curr_variance
;
5219 if (improvement
< ALB_ACCEPTABLE_IMPROVEMENT
) {
5224 free(curr_pmd_usage
);
5230 /* Return true if needs to revalidate datapath flows. */
5232 dpif_netdev_run(struct dpif
*dpif
)
5234 struct dp_netdev_port
*port
;
5235 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5236 struct dp_netdev_pmd_thread
*non_pmd
;
5237 uint64_t new_tnl_seq
;
5238 bool need_to_flush
= true;
5239 bool pmd_rebalance
= false;
5240 long long int now
= time_msec();
5241 struct dp_netdev_pmd_thread
*pmd
;
5243 ovs_mutex_lock(&dp
->port_mutex
);
5244 non_pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
5246 ovs_mutex_lock(&dp
->non_pmd_mutex
);
5247 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5248 if (!netdev_is_pmd(port
->netdev
)) {
5251 if (port
->emc_enabled
) {
5252 atomic_read_relaxed(&dp
->emc_insert_min
,
5253 &non_pmd
->ctx
.emc_insert_min
);
5255 non_pmd
->ctx
.emc_insert_min
= 0;
5258 for (i
= 0; i
< port
->n_rxq
; i
++) {
5260 if (!netdev_rxq_enabled(port
->rxqs
[i
].rx
)) {
5264 if (dp_netdev_process_rxq_port(non_pmd
,
5267 need_to_flush
= false;
5272 if (need_to_flush
) {
5273 /* We didn't receive anything in the process loop.
5274 * Check if we need to send something.
5275 * There was no time updates on current iteration. */
5276 pmd_thread_ctx_time_update(non_pmd
);
5277 dp_netdev_pmd_flush_output_packets(non_pmd
, false);
5280 dpif_netdev_xps_revalidate_pmd(non_pmd
, false);
5281 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
5283 dp_netdev_pmd_unref(non_pmd
);
5286 struct pmd_auto_lb
*pmd_alb
= &dp
->pmd_alb
;
5287 if (pmd_alb
->is_enabled
) {
5288 if (!pmd_alb
->rebalance_poll_timer
) {
5289 pmd_alb
->rebalance_poll_timer
= now
;
5290 } else if ((pmd_alb
->rebalance_poll_timer
+
5291 pmd_alb
->rebalance_intvl
) < now
) {
5292 pmd_alb
->rebalance_poll_timer
= now
;
5293 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
5294 if (atomic_count_get(&pmd
->pmd_overloaded
) >=
5295 PMD_RXQ_INTERVAL_MAX
) {
5296 pmd_rebalance
= true;
5301 if (pmd_rebalance
&&
5302 !dp_netdev_is_reconf_required(dp
) &&
5303 !ports_require_restart(dp
) &&
5304 pmd_rebalance_dry_run(dp
)) {
5305 VLOG_INFO("PMD auto lb dry run."
5306 " requesting datapath reconfigure.");
5307 dp_netdev_request_reconfigure(dp
);
5312 if (dp_netdev_is_reconf_required(dp
) || ports_require_restart(dp
)) {
5313 reconfigure_datapath(dp
);
5315 ovs_mutex_unlock(&dp
->port_mutex
);
5317 tnl_neigh_cache_run();
5319 new_tnl_seq
= seq_read(tnl_conf_seq
);
5321 if (dp
->last_tnl_conf_seq
!= new_tnl_seq
) {
5322 dp
->last_tnl_conf_seq
= new_tnl_seq
;
5329 dpif_netdev_wait(struct dpif
*dpif
)
5331 struct dp_netdev_port
*port
;
5332 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5334 ovs_mutex_lock(&dp_netdev_mutex
);
5335 ovs_mutex_lock(&dp
->port_mutex
);
5336 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
5337 netdev_wait_reconf_required(port
->netdev
);
5338 if (!netdev_is_pmd(port
->netdev
)) {
5341 for (i
= 0; i
< port
->n_rxq
; i
++) {
5342 netdev_rxq_wait(port
->rxqs
[i
].rx
);
5346 ovs_mutex_unlock(&dp
->port_mutex
);
5347 ovs_mutex_unlock(&dp_netdev_mutex
);
5348 seq_wait(tnl_conf_seq
, dp
->last_tnl_conf_seq
);
5352 pmd_free_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
5354 struct tx_port
*tx_port_cached
;
5356 /* Flush all the queued packets. */
5357 dp_netdev_pmd_flush_output_packets(pmd
, true);
5358 /* Free all used tx queue ids. */
5359 dpif_netdev_xps_revalidate_pmd(pmd
, true);
5361 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->tnl_port_cache
) {
5362 free(tx_port_cached
);
5364 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->send_port_cache
) {
5365 free(tx_port_cached
);
5369 /* Copies ports from 'pmd->tx_ports' (shared with the main thread) to
5370 * thread-local copies. Copy to 'pmd->tnl_port_cache' if it is a tunnel
5371 * device, otherwise to 'pmd->send_port_cache' if the port has at least
5374 pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
5375 OVS_REQUIRES(pmd
->port_mutex
)
5377 struct tx_port
*tx_port
, *tx_port_cached
;
5379 pmd_free_cached_ports(pmd
);
5380 hmap_shrink(&pmd
->send_port_cache
);
5381 hmap_shrink(&pmd
->tnl_port_cache
);
5383 HMAP_FOR_EACH (tx_port
, node
, &pmd
->tx_ports
) {
5384 if (netdev_has_tunnel_push_pop(tx_port
->port
->netdev
)) {
5385 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
5386 hmap_insert(&pmd
->tnl_port_cache
, &tx_port_cached
->node
,
5387 hash_port_no(tx_port_cached
->port
->port_no
));
5390 if (netdev_n_txq(tx_port
->port
->netdev
)) {
5391 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
5392 hmap_insert(&pmd
->send_port_cache
, &tx_port_cached
->node
,
5393 hash_port_no(tx_port_cached
->port
->port_no
));
5399 pmd_alloc_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
5401 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
5402 if (!id_pool_alloc_id(pmd
->dp
->tx_qid_pool
, &pmd
->static_tx_qid
)) {
5403 VLOG_ABORT("static_tx_qid allocation failed for PMD on core %2d"
5404 ", numa_id %d.", pmd
->core_id
, pmd
->numa_id
);
5406 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
5408 VLOG_DBG("static_tx_qid = %d allocated for PMD thread on core %2d"
5409 ", numa_id %d.", pmd
->static_tx_qid
, pmd
->core_id
, pmd
->numa_id
);
5413 pmd_free_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
5415 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
5416 id_pool_free_id(pmd
->dp
->tx_qid_pool
, pmd
->static_tx_qid
);
5417 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
5421 pmd_load_queues_and_ports(struct dp_netdev_pmd_thread
*pmd
,
5422 struct polled_queue
**ppoll_list
)
5424 struct polled_queue
*poll_list
= *ppoll_list
;
5425 struct rxq_poll
*poll
;
5428 ovs_mutex_lock(&pmd
->port_mutex
);
5429 poll_list
= xrealloc(poll_list
, hmap_count(&pmd
->poll_list
)
5430 * sizeof *poll_list
);
5433 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
5434 poll_list
[i
].rxq
= poll
->rxq
;
5435 poll_list
[i
].port_no
= poll
->rxq
->port
->port_no
;
5436 poll_list
[i
].emc_enabled
= poll
->rxq
->port
->emc_enabled
;
5437 poll_list
[i
].rxq_enabled
= netdev_rxq_enabled(poll
->rxq
->rx
);
5438 poll_list
[i
].change_seq
=
5439 netdev_get_change_seq(poll
->rxq
->port
->netdev
);
5443 pmd_load_cached_ports(pmd
);
5445 ovs_mutex_unlock(&pmd
->port_mutex
);
5447 *ppoll_list
= poll_list
;
5452 pmd_thread_main(void *f_
)
5454 struct dp_netdev_pmd_thread
*pmd
= f_
;
5455 struct pmd_perf_stats
*s
= &pmd
->perf_stats
;
5456 unsigned int lc
= 0;
5457 struct polled_queue
*poll_list
;
5458 bool wait_for_reload
= false;
5464 int process_packets
= 0;
5468 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
5469 ovsthread_setspecific(pmd
->dp
->per_pmd_key
, pmd
);
5470 ovs_numa_thread_setaffinity_core(pmd
->core_id
);
5471 dpdk_set_lcore_id(pmd
->core_id
);
5472 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
5473 dfc_cache_init(&pmd
->flow_cache
);
5474 pmd_alloc_static_tx_qid(pmd
);
5477 atomic_count_init(&pmd
->pmd_overloaded
, 0);
5479 /* List port/core affinity */
5480 for (i
= 0; i
< poll_cnt
; i
++) {
5481 VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n",
5482 pmd
->core_id
, netdev_rxq_get_name(poll_list
[i
].rxq
->rx
),
5483 netdev_rxq_get_queue_id(poll_list
[i
].rxq
->rx
));
5484 /* Reset the rxq current cycles counter. */
5485 dp_netdev_rxq_set_cycles(poll_list
[i
].rxq
, RXQ_CYCLES_PROC_CURR
, 0);
5489 if (wait_for_reload
) {
5490 /* Don't sleep, control thread will ask for a reload shortly. */
5492 atomic_read_explicit(&pmd
->reload
, &reload
,
5493 memory_order_acquire
);
5496 while (seq_read(pmd
->reload_seq
) == pmd
->last_reload_seq
) {
5497 seq_wait(pmd
->reload_seq
, pmd
->last_reload_seq
);
5503 pmd
->intrvl_tsc_prev
= 0;
5504 atomic_store_relaxed(&pmd
->intrvl_cycles
, 0);
5505 cycles_counter_update(s
);
5506 /* Protect pmd stats from external clearing while polling. */
5507 ovs_mutex_lock(&pmd
->perf_stats
.stats_mutex
);
5509 uint64_t rx_packets
= 0, tx_packets
= 0;
5511 pmd_perf_start_iteration(s
);
5513 for (i
= 0; i
< poll_cnt
; i
++) {
5515 if (!poll_list
[i
].rxq_enabled
) {
5519 if (poll_list
[i
].emc_enabled
) {
5520 atomic_read_relaxed(&pmd
->dp
->emc_insert_min
,
5521 &pmd
->ctx
.emc_insert_min
);
5523 pmd
->ctx
.emc_insert_min
= 0;
5527 dp_netdev_process_rxq_port(pmd
, poll_list
[i
].rxq
,
5528 poll_list
[i
].port_no
);
5529 rx_packets
+= process_packets
;
5533 /* We didn't receive anything in the process loop.
5534 * Check if we need to send something.
5535 * There was no time updates on current iteration. */
5536 pmd_thread_ctx_time_update(pmd
);
5537 tx_packets
= dp_netdev_pmd_flush_output_packets(pmd
, false);
5543 coverage_try_clear();
5544 dp_netdev_pmd_try_optimize(pmd
, poll_list
, poll_cnt
);
5545 if (!ovsrcu_try_quiesce()) {
5546 emc_cache_slow_sweep(&((pmd
->flow_cache
).emc_cache
));
5549 for (i
= 0; i
< poll_cnt
; i
++) {
5550 uint64_t current_seq
=
5551 netdev_get_change_seq(poll_list
[i
].rxq
->port
->netdev
);
5552 if (poll_list
[i
].change_seq
!= current_seq
) {
5553 poll_list
[i
].change_seq
= current_seq
;
5554 poll_list
[i
].rxq_enabled
=
5555 netdev_rxq_enabled(poll_list
[i
].rxq
->rx
);
5560 atomic_read_explicit(&pmd
->reload
, &reload
, memory_order_acquire
);
5561 if (OVS_UNLIKELY(reload
)) {
5565 pmd_perf_end_iteration(s
, rx_packets
, tx_packets
,
5566 pmd_perf_metrics_enabled(pmd
));
5568 ovs_mutex_unlock(&pmd
->perf_stats
.stats_mutex
);
5570 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
5571 atomic_read_relaxed(&pmd
->wait_for_reload
, &wait_for_reload
);
5572 atomic_read_relaxed(&pmd
->reload_tx_qid
, &reload_tx_qid
);
5573 atomic_read_relaxed(&pmd
->exit
, &exiting
);
5574 /* Signal here to make sure the pmd finishes
5575 * reloading the updated configuration. */
5576 dp_netdev_pmd_reload_done(pmd
);
5578 if (reload_tx_qid
) {
5579 pmd_free_static_tx_qid(pmd
);
5580 pmd_alloc_static_tx_qid(pmd
);
5587 pmd_free_static_tx_qid(pmd
);
5588 dfc_cache_uninit(&pmd
->flow_cache
);
5590 pmd_free_cached_ports(pmd
);
5595 dp_netdev_disable_upcall(struct dp_netdev
*dp
)
5596 OVS_ACQUIRES(dp
->upcall_rwlock
)
5598 fat_rwlock_wrlock(&dp
->upcall_rwlock
);
5604 dpif_netdev_meter_get_features(const struct dpif
* dpif OVS_UNUSED
,
5605 struct ofputil_meter_features
*features
)
5607 features
->max_meters
= MAX_METERS
;
5608 features
->band_types
= DP_SUPPORTED_METER_BAND_TYPES
;
5609 features
->capabilities
= DP_SUPPORTED_METER_FLAGS_MASK
;
5610 features
->max_bands
= MAX_BANDS
;
5611 features
->max_color
= 0;
5614 /* Applies the meter identified by 'meter_id' to 'packets_'. Packets
5615 * that exceed a band are dropped in-place. */
5617 dp_netdev_run_meter(struct dp_netdev
*dp
, struct dp_packet_batch
*packets_
,
5618 uint32_t meter_id
, long long int now
)
5620 struct dp_meter
*meter
;
5621 struct dp_meter_band
*band
;
5622 struct dp_packet
*packet
;
5623 long long int long_delta_t
; /* msec */
5624 uint32_t delta_t
; /* msec */
5625 const size_t cnt
= dp_packet_batch_size(packets_
);
5626 uint32_t bytes
, volume
;
5627 int exceeded_band
[NETDEV_MAX_BURST
];
5628 uint32_t exceeded_rate
[NETDEV_MAX_BURST
];
5629 int exceeded_pkt
= cnt
; /* First packet that exceeded a band rate. */
5631 if (meter_id
>= MAX_METERS
) {
5635 meter_lock(dp
, meter_id
);
5636 meter
= dp
->meters
[meter_id
];
5641 /* Initialize as negative values. */
5642 memset(exceeded_band
, 0xff, cnt
* sizeof *exceeded_band
);
5643 /* Initialize as zeroes. */
5644 memset(exceeded_rate
, 0, cnt
* sizeof *exceeded_rate
);
5646 /* All packets will hit the meter at the same time. */
5647 long_delta_t
= now
/ 1000 - meter
->used
/ 1000; /* msec */
5649 /* Make sure delta_t will not be too large, so that bucket will not
5650 * wrap around below. */
5651 delta_t
= (long_delta_t
> (long long int)meter
->max_delta_t
)
5652 ? meter
->max_delta_t
: (uint32_t)long_delta_t
;
5654 /* Update meter stats. */
5656 meter
->packet_count
+= cnt
;
5658 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
5659 bytes
+= dp_packet_size(packet
);
5661 meter
->byte_count
+= bytes
;
5663 /* Meters can operate in terms of packets per second or kilobits per
5665 if (meter
->flags
& OFPMF13_PKTPS
) {
5666 /* Rate in packets/second, bucket 1/1000 packets. */
5667 /* msec * packets/sec = 1/1000 packets. */
5668 volume
= cnt
* 1000; /* Take 'cnt' packets from the bucket. */
5670 /* Rate in kbps, bucket in bits. */
5671 /* msec * kbps = bits */
5675 /* Update all bands and find the one hit with the highest rate for each
5676 * packet (if any). */
5677 for (int m
= 0; m
< meter
->n_bands
; ++m
) {
5678 band
= &meter
->bands
[m
];
5680 /* Update band's bucket. */
5681 band
->bucket
+= delta_t
* band
->up
.rate
;
5682 if (band
->bucket
> band
->up
.burst_size
) {
5683 band
->bucket
= band
->up
.burst_size
;
5686 /* Drain the bucket for all the packets, if possible. */
5687 if (band
->bucket
>= volume
) {
5688 band
->bucket
-= volume
;
5690 int band_exceeded_pkt
;
5692 /* Band limit hit, must process packet-by-packet. */
5693 if (meter
->flags
& OFPMF13_PKTPS
) {
5694 band_exceeded_pkt
= band
->bucket
/ 1000;
5695 band
->bucket
%= 1000; /* Remainder stays in bucket. */
5697 /* Update the exceeding band for each exceeding packet.
5698 * (Only one band will be fired by a packet, and that
5699 * can be different for each packet.) */
5700 for (int i
= band_exceeded_pkt
; i
< cnt
; i
++) {
5701 if (band
->up
.rate
> exceeded_rate
[i
]) {
5702 exceeded_rate
[i
] = band
->up
.rate
;
5703 exceeded_band
[i
] = m
;
5707 /* Packet sizes differ, must process one-by-one. */
5708 band_exceeded_pkt
= cnt
;
5709 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
5710 uint32_t bits
= dp_packet_size(packet
) * 8;
5712 if (band
->bucket
>= bits
) {
5713 band
->bucket
-= bits
;
5715 if (i
< band_exceeded_pkt
) {
5716 band_exceeded_pkt
= i
;
5718 /* Update the exceeding band for the exceeding packet.
5719 * (Only one band will be fired by a packet, and that
5720 * can be different for each packet.) */
5721 if (band
->up
.rate
> exceeded_rate
[i
]) {
5722 exceeded_rate
[i
] = band
->up
.rate
;
5723 exceeded_band
[i
] = m
;
5728 /* Remember the first exceeding packet. */
5729 if (exceeded_pkt
> band_exceeded_pkt
) {
5730 exceeded_pkt
= band_exceeded_pkt
;
5735 /* Fire the highest rate band exceeded by each packet, and drop
5736 * packets if needed. */
5738 DP_PACKET_BATCH_REFILL_FOR_EACH (j
, cnt
, packet
, packets_
) {
5739 if (exceeded_band
[j
] >= 0) {
5740 /* Meter drop packet. */
5741 band
= &meter
->bands
[exceeded_band
[j
]];
5742 band
->packet_count
+= 1;
5743 band
->byte_count
+= dp_packet_size(packet
);
5745 dp_packet_delete(packet
);
5747 /* Meter accepts packet. */
5748 dp_packet_batch_refill(packets_
, packet
, j
);
5752 meter_unlock(dp
, meter_id
);
5755 /* Meter set/get/del processing is still single-threaded. */
5757 dpif_netdev_meter_set(struct dpif
*dpif
, ofproto_meter_id meter_id
,
5758 struct ofputil_meter_config
*config
)
5760 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5761 uint32_t mid
= meter_id
.uint32
;
5762 struct dp_meter
*meter
;
5765 if (mid
>= MAX_METERS
) {
5766 return EFBIG
; /* Meter_id out of range. */
5769 if (config
->flags
& ~DP_SUPPORTED_METER_FLAGS_MASK
) {
5770 return EBADF
; /* Unsupported flags set */
5773 if (config
->n_bands
> MAX_BANDS
) {
5777 for (i
= 0; i
< config
->n_bands
; ++i
) {
5778 switch (config
->bands
[i
].type
) {
5782 return ENODEV
; /* Unsupported band type */
5786 /* Allocate meter */
5787 meter
= xzalloc(sizeof *meter
5788 + config
->n_bands
* sizeof(struct dp_meter_band
));
5790 meter
->flags
= config
->flags
;
5791 meter
->n_bands
= config
->n_bands
;
5792 meter
->max_delta_t
= 0;
5793 meter
->used
= time_usec();
5796 for (i
= 0; i
< config
->n_bands
; ++i
) {
5797 uint32_t band_max_delta_t
;
5799 /* Set burst size to a workable value if none specified. */
5800 if (config
->bands
[i
].burst_size
== 0) {
5801 config
->bands
[i
].burst_size
= config
->bands
[i
].rate
;
5804 meter
->bands
[i
].up
= config
->bands
[i
];
5805 /* Convert burst size to the bucket units: */
5806 /* pkts => 1/1000 packets, kilobits => bits. */
5807 meter
->bands
[i
].up
.burst_size
*= 1000;
5808 /* Initialize bucket to empty. */
5809 meter
->bands
[i
].bucket
= 0;
5811 /* Figure out max delta_t that is enough to fill any bucket. */
5813 = meter
->bands
[i
].up
.burst_size
/ meter
->bands
[i
].up
.rate
;
5814 if (band_max_delta_t
> meter
->max_delta_t
) {
5815 meter
->max_delta_t
= band_max_delta_t
;
5819 meter_lock(dp
, mid
);
5820 dp_delete_meter(dp
, mid
); /* Free existing meter, if any */
5821 dp
->meters
[mid
] = meter
;
5822 meter_unlock(dp
, mid
);
5828 dpif_netdev_meter_get(const struct dpif
*dpif
,
5829 ofproto_meter_id meter_id_
,
5830 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
5832 const struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5833 uint32_t meter_id
= meter_id_
.uint32
;
5836 if (meter_id
>= MAX_METERS
) {
5840 meter_lock(dp
, meter_id
);
5841 const struct dp_meter
*meter
= dp
->meters
[meter_id
];
5849 stats
->packet_in_count
= meter
->packet_count
;
5850 stats
->byte_in_count
= meter
->byte_count
;
5852 for (i
= 0; i
< n_bands
&& i
< meter
->n_bands
; ++i
) {
5853 stats
->bands
[i
].packet_count
= meter
->bands
[i
].packet_count
;
5854 stats
->bands
[i
].byte_count
= meter
->bands
[i
].byte_count
;
5861 meter_unlock(dp
, meter_id
);
5866 dpif_netdev_meter_del(struct dpif
*dpif
,
5867 ofproto_meter_id meter_id_
,
5868 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
5870 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5873 error
= dpif_netdev_meter_get(dpif
, meter_id_
, stats
, n_bands
);
5875 uint32_t meter_id
= meter_id_
.uint32
;
5877 meter_lock(dp
, meter_id
);
5878 dp_delete_meter(dp
, meter_id
);
5879 meter_unlock(dp
, meter_id
);
5886 dpif_netdev_disable_upcall(struct dpif
*dpif
)
5887 OVS_NO_THREAD_SAFETY_ANALYSIS
5889 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5890 dp_netdev_disable_upcall(dp
);
5894 dp_netdev_enable_upcall(struct dp_netdev
*dp
)
5895 OVS_RELEASES(dp
->upcall_rwlock
)
5897 fat_rwlock_unlock(&dp
->upcall_rwlock
);
5901 dpif_netdev_enable_upcall(struct dpif
*dpif
)
5902 OVS_NO_THREAD_SAFETY_ANALYSIS
5904 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5905 dp_netdev_enable_upcall(dp
);
5909 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
)
5911 atomic_store_relaxed(&pmd
->wait_for_reload
, false);
5912 atomic_store_relaxed(&pmd
->reload_tx_qid
, false);
5913 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
5914 atomic_store_explicit(&pmd
->reload
, false, memory_order_release
);
5917 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
5918 * the pointer if succeeds, otherwise, NULL (it can return NULL even if
5919 * 'core_id' is NON_PMD_CORE_ID).
5921 * Caller must unrefs the returned reference. */
5922 static struct dp_netdev_pmd_thread
*
5923 dp_netdev_get_pmd(struct dp_netdev
*dp
, unsigned core_id
)
5925 struct dp_netdev_pmd_thread
*pmd
;
5926 const struct cmap_node
*pnode
;
5928 pnode
= cmap_find(&dp
->poll_threads
, hash_int(core_id
, 0));
5932 pmd
= CONTAINER_OF(pnode
, struct dp_netdev_pmd_thread
, node
);
5934 return dp_netdev_pmd_try_ref(pmd
) ? pmd
: NULL
;
5937 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
5939 dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
5940 OVS_REQUIRES(dp
->port_mutex
)
5942 struct dp_netdev_pmd_thread
*non_pmd
;
5944 non_pmd
= xzalloc(sizeof *non_pmd
);
5945 dp_netdev_configure_pmd(non_pmd
, dp
, NON_PMD_CORE_ID
, OVS_NUMA_UNSPEC
);
5948 /* Caller must have valid pointer to 'pmd'. */
5950 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
)
5952 return ovs_refcount_try_ref_rcu(&pmd
->ref_cnt
);
5956 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
)
5958 if (pmd
&& ovs_refcount_unref(&pmd
->ref_cnt
) == 1) {
5959 ovsrcu_postpone(dp_netdev_destroy_pmd
, pmd
);
5963 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
5964 * fails, keeps checking for next node until reaching the end of cmap.
5966 * Caller must unrefs the returned reference. */
5967 static struct dp_netdev_pmd_thread
*
5968 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
)
5970 struct dp_netdev_pmd_thread
*next
;
5973 struct cmap_node
*node
;
5975 node
= cmap_next_position(&dp
->poll_threads
, pos
);
5976 next
= node
? CONTAINER_OF(node
, struct dp_netdev_pmd_thread
, node
)
5978 } while (next
&& !dp_netdev_pmd_try_ref(next
));
5983 /* Configures the 'pmd' based on the input argument. */
5985 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
, struct dp_netdev
*dp
,
5986 unsigned core_id
, int numa_id
)
5989 pmd
->core_id
= core_id
;
5990 pmd
->numa_id
= numa_id
;
5991 pmd
->need_reload
= false;
5992 pmd
->n_output_batches
= 0;
5994 ovs_refcount_init(&pmd
->ref_cnt
);
5995 atomic_init(&pmd
->exit
, false);
5996 pmd
->reload_seq
= seq_create();
5997 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
5998 atomic_init(&pmd
->reload
, false);
5999 ovs_mutex_init(&pmd
->flow_mutex
);
6000 ovs_mutex_init(&pmd
->port_mutex
);
6001 cmap_init(&pmd
->flow_table
);
6002 cmap_init(&pmd
->classifiers
);
6003 pmd
->ctx
.last_rxq
= NULL
;
6004 pmd_thread_ctx_time_update(pmd
);
6005 pmd
->next_optimization
= pmd
->ctx
.now
+ DPCLS_OPTIMIZATION_INTERVAL
;
6006 pmd
->rxq_next_cycle_store
= pmd
->ctx
.now
+ PMD_RXQ_INTERVAL_LEN
;
6007 hmap_init(&pmd
->poll_list
);
6008 hmap_init(&pmd
->tx_ports
);
6009 hmap_init(&pmd
->tnl_port_cache
);
6010 hmap_init(&pmd
->send_port_cache
);
6011 /* init the 'flow_cache' since there is no
6012 * actual thread created for NON_PMD_CORE_ID. */
6013 if (core_id
== NON_PMD_CORE_ID
) {
6014 dfc_cache_init(&pmd
->flow_cache
);
6015 pmd_alloc_static_tx_qid(pmd
);
6017 pmd_perf_stats_init(&pmd
->perf_stats
);
6018 cmap_insert(&dp
->poll_threads
, CONST_CAST(struct cmap_node
*, &pmd
->node
),
6019 hash_int(core_id
, 0));
6023 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
)
6027 dp_netdev_pmd_flow_flush(pmd
);
6028 hmap_destroy(&pmd
->send_port_cache
);
6029 hmap_destroy(&pmd
->tnl_port_cache
);
6030 hmap_destroy(&pmd
->tx_ports
);
6031 hmap_destroy(&pmd
->poll_list
);
6032 /* All flows (including their dpcls_rules) have been deleted already */
6033 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
6035 ovsrcu_postpone(free
, cls
);
6037 cmap_destroy(&pmd
->classifiers
);
6038 cmap_destroy(&pmd
->flow_table
);
6039 ovs_mutex_destroy(&pmd
->flow_mutex
);
6040 seq_destroy(pmd
->reload_seq
);
6041 ovs_mutex_destroy(&pmd
->port_mutex
);
6045 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
6046 * and unrefs the struct. */
6048 dp_netdev_del_pmd(struct dp_netdev
*dp
, struct dp_netdev_pmd_thread
*pmd
)
6050 /* NON_PMD_CORE_ID doesn't have a thread, so we don't have to synchronize,
6051 * but extra cleanup is necessary */
6052 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
6053 ovs_mutex_lock(&dp
->non_pmd_mutex
);
6054 dfc_cache_uninit(&pmd
->flow_cache
);
6055 pmd_free_cached_ports(pmd
);
6056 pmd_free_static_tx_qid(pmd
);
6057 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
6059 atomic_store_relaxed(&pmd
->exit
, true);
6060 dp_netdev_reload_pmd__(pmd
);
6061 xpthread_join(pmd
->thread
, NULL
);
6064 dp_netdev_pmd_clear_ports(pmd
);
6066 /* Purges the 'pmd''s flows after stopping the thread, but before
6067 * destroying the flows, so that the flow stats can be collected. */
6068 if (dp
->dp_purge_cb
) {
6069 dp
->dp_purge_cb(dp
->dp_purge_aux
, pmd
->core_id
);
6071 cmap_remove(&pmd
->dp
->poll_threads
, &pmd
->node
, hash_int(pmd
->core_id
, 0));
6072 dp_netdev_pmd_unref(pmd
);
6075 /* Destroys all pmd threads. If 'non_pmd' is true it also destroys the non pmd
6078 dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
, bool non_pmd
)
6080 struct dp_netdev_pmd_thread
*pmd
;
6081 struct dp_netdev_pmd_thread
**pmd_list
;
6082 size_t k
= 0, n_pmds
;
6084 n_pmds
= cmap_count(&dp
->poll_threads
);
6085 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
6087 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
6088 if (!non_pmd
&& pmd
->core_id
== NON_PMD_CORE_ID
) {
6091 /* We cannot call dp_netdev_del_pmd(), since it alters
6092 * 'dp->poll_threads' (while we're iterating it) and it
6094 ovs_assert(k
< n_pmds
);
6095 pmd_list
[k
++] = pmd
;
6098 for (size_t i
= 0; i
< k
; i
++) {
6099 dp_netdev_del_pmd(dp
, pmd_list
[i
]);
6104 /* Deletes all rx queues from pmd->poll_list and all the ports from
6107 dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
)
6109 struct rxq_poll
*poll
;
6110 struct tx_port
*port
;
6112 ovs_mutex_lock(&pmd
->port_mutex
);
6113 HMAP_FOR_EACH_POP (poll
, node
, &pmd
->poll_list
) {
6116 HMAP_FOR_EACH_POP (port
, node
, &pmd
->tx_ports
) {
6119 ovs_mutex_unlock(&pmd
->port_mutex
);
6122 /* Adds rx queue to poll_list of PMD thread, if it's not there already. */
6124 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
6125 struct dp_netdev_rxq
*rxq
)
6126 OVS_REQUIRES(pmd
->port_mutex
)
6128 int qid
= netdev_rxq_get_queue_id(rxq
->rx
);
6129 uint32_t hash
= hash_2words(odp_to_u32(rxq
->port
->port_no
), qid
);
6130 struct rxq_poll
*poll
;
6132 HMAP_FOR_EACH_WITH_HASH (poll
, node
, hash
, &pmd
->poll_list
) {
6133 if (poll
->rxq
== rxq
) {
6134 /* 'rxq' is already polled by this thread. Do nothing. */
6139 poll
= xmalloc(sizeof *poll
);
6141 hmap_insert(&pmd
->poll_list
, &poll
->node
, hash
);
6143 pmd
->need_reload
= true;
6146 /* Delete 'poll' from poll_list of PMD thread. */
6148 dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
6149 struct rxq_poll
*poll
)
6150 OVS_REQUIRES(pmd
->port_mutex
)
6152 hmap_remove(&pmd
->poll_list
, &poll
->node
);
6155 pmd
->need_reload
= true;
6158 /* Add 'port' to the tx port cache of 'pmd', which must be reloaded for the
6159 * changes to take effect. */
6161 dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
6162 struct dp_netdev_port
*port
)
6163 OVS_REQUIRES(pmd
->port_mutex
)
6167 tx
= tx_port_lookup(&pmd
->tx_ports
, port
->port_no
);
6169 /* 'port' is already on this thread tx cache. Do nothing. */
6173 tx
= xzalloc(sizeof *tx
);
6177 tx
->flush_time
= 0LL;
6178 dp_packet_batch_init(&tx
->output_pkts
);
6180 hmap_insert(&pmd
->tx_ports
, &tx
->node
, hash_port_no(tx
->port
->port_no
));
6181 pmd
->need_reload
= true;
6184 /* Del 'tx' from the tx port cache of 'pmd', which must be reloaded for the
6185 * changes to take effect. */
6187 dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
6189 OVS_REQUIRES(pmd
->port_mutex
)
6191 hmap_remove(&pmd
->tx_ports
, &tx
->node
);
6193 pmd
->need_reload
= true;
6197 dpif_netdev_get_datapath_version(void)
6199 return xstrdup("<built-in>");
6203 dp_netdev_flow_used(struct dp_netdev_flow
*netdev_flow
, int cnt
, int size
,
6204 uint16_t tcp_flags
, long long now
)
6208 atomic_store_relaxed(&netdev_flow
->stats
.used
, now
);
6209 non_atomic_ullong_add(&netdev_flow
->stats
.packet_count
, cnt
);
6210 non_atomic_ullong_add(&netdev_flow
->stats
.byte_count
, size
);
6211 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
6213 atomic_store_relaxed(&netdev_flow
->stats
.tcp_flags
, flags
);
6217 dp_netdev_upcall(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
*packet_
,
6218 struct flow
*flow
, struct flow_wildcards
*wc
, ovs_u128
*ufid
,
6219 enum dpif_upcall_type type
, const struct nlattr
*userdata
,
6220 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
6222 struct dp_netdev
*dp
= pmd
->dp
;
6224 if (OVS_UNLIKELY(!dp
->upcall_cb
)) {
6228 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl
))) {
6229 struct ds ds
= DS_EMPTY_INITIALIZER
;
6232 struct odp_flow_key_parms odp_parms
= {
6234 .mask
= wc
? &wc
->masks
: NULL
,
6235 .support
= dp_netdev_support
,
6238 ofpbuf_init(&key
, 0);
6239 odp_flow_key_from_flow(&odp_parms
, &key
);
6240 packet_str
= ofp_dp_packet_to_string(packet_
);
6242 odp_flow_key_format(key
.data
, key
.size
, &ds
);
6244 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp
->name
,
6245 dpif_upcall_type_to_string(type
), ds_cstr(&ds
), packet_str
);
6247 ofpbuf_uninit(&key
);
6253 return dp
->upcall_cb(packet_
, flow
, ufid
, pmd
->core_id
, type
, userdata
,
6254 actions
, wc
, put_actions
, dp
->upcall_aux
);
6257 static inline uint32_t
6258 dpif_netdev_packet_get_rss_hash_orig_pkt(struct dp_packet
*packet
,
6259 const struct miniflow
*mf
)
6263 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
6264 hash
= dp_packet_get_rss_hash(packet
);
6266 hash
= miniflow_hash_5tuple(mf
, 0);
6267 dp_packet_set_rss_hash(packet
, hash
);
6273 static inline uint32_t
6274 dpif_netdev_packet_get_rss_hash(struct dp_packet
*packet
,
6275 const struct miniflow
*mf
)
6277 uint32_t hash
, recirc_depth
;
6279 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
6280 hash
= dp_packet_get_rss_hash(packet
);
6282 hash
= miniflow_hash_5tuple(mf
, 0);
6283 dp_packet_set_rss_hash(packet
, hash
);
6286 /* The RSS hash must account for the recirculation depth to avoid
6287 * collisions in the exact match cache */
6288 recirc_depth
= *recirc_depth_get_unsafe();
6289 if (OVS_UNLIKELY(recirc_depth
)) {
6290 hash
= hash_finish(hash
, recirc_depth
);
6291 dp_packet_set_rss_hash(packet
, hash
);
6296 struct packet_batch_per_flow
{
6297 unsigned int byte_count
;
6299 struct dp_netdev_flow
*flow
;
6301 struct dp_packet_batch array
;
6305 packet_batch_per_flow_update(struct packet_batch_per_flow
*batch
,
6306 struct dp_packet
*packet
,
6309 batch
->byte_count
+= dp_packet_size(packet
);
6310 batch
->tcp_flags
|= tcp_flags
;
6311 batch
->array
.packets
[batch
->array
.count
++] = packet
;
6315 packet_batch_per_flow_init(struct packet_batch_per_flow
*batch
,
6316 struct dp_netdev_flow
*flow
)
6318 flow
->batch
= batch
;
6321 dp_packet_batch_init(&batch
->array
);
6322 batch
->byte_count
= 0;
6323 batch
->tcp_flags
= 0;
6327 packet_batch_per_flow_execute(struct packet_batch_per_flow
*batch
,
6328 struct dp_netdev_pmd_thread
*pmd
)
6330 struct dp_netdev_actions
*actions
;
6331 struct dp_netdev_flow
*flow
= batch
->flow
;
6333 dp_netdev_flow_used(flow
, batch
->array
.count
, batch
->byte_count
,
6334 batch
->tcp_flags
, pmd
->ctx
.now
/ 1000);
6336 actions
= dp_netdev_flow_get_actions(flow
);
6338 dp_netdev_execute_actions(pmd
, &batch
->array
, true, &flow
->flow
,
6339 actions
->actions
, actions
->size
);
6343 dp_netdev_queue_batches(struct dp_packet
*pkt
,
6344 struct dp_netdev_flow
*flow
, uint16_t tcp_flags
,
6345 struct packet_batch_per_flow
*batches
,
6348 struct packet_batch_per_flow
*batch
= flow
->batch
;
6350 if (OVS_UNLIKELY(!batch
)) {
6351 batch
= &batches
[(*n_batches
)++];
6352 packet_batch_per_flow_init(batch
, flow
);
6355 packet_batch_per_flow_update(batch
, pkt
, tcp_flags
);
6359 packet_enqueue_to_flow_map(struct dp_packet
*packet
,
6360 struct dp_netdev_flow
*flow
,
6362 struct dp_packet_flow_map
*flow_map
,
6365 struct dp_packet_flow_map
*map
= &flow_map
[index
];
6367 map
->packet
= packet
;
6368 map
->tcp_flags
= tcp_flags
;
6371 /* SMC lookup function for a batch of packets.
6372 * By doing batching SMC lookup, we can use prefetch
6373 * to hide memory access latency.
6376 smc_lookup_batch(struct dp_netdev_pmd_thread
*pmd
,
6377 struct netdev_flow_key
*keys
,
6378 struct netdev_flow_key
**missed_keys
,
6379 struct dp_packet_batch
*packets_
,
6381 struct dp_packet_flow_map
*flow_map
,
6385 struct dp_packet
*packet
;
6386 size_t n_smc_hit
= 0, n_missed
= 0;
6387 struct dfc_cache
*cache
= &pmd
->flow_cache
;
6388 struct smc_cache
*smc_cache
= &cache
->smc_cache
;
6389 const struct cmap_node
*flow_node
;
6393 /* Prefetch buckets for all packets */
6394 for (i
= 0; i
< cnt
; i
++) {
6395 OVS_PREFETCH(&smc_cache
->buckets
[keys
[i
].hash
& SMC_MASK
]);
6398 DP_PACKET_BATCH_REFILL_FOR_EACH (i
, cnt
, packet
, packets_
) {
6399 struct dp_netdev_flow
*flow
= NULL
;
6400 flow_node
= smc_entry_get(pmd
, keys
[i
].hash
);
6402 /* Get the original order of this packet in received batch. */
6403 recv_idx
= index_map
[i
];
6405 if (OVS_LIKELY(flow_node
!= NULL
)) {
6406 CMAP_NODE_FOR_EACH (flow
, node
, flow_node
) {
6407 /* Since we dont have per-port megaflow to check the port
6408 * number, we need to verify that the input ports match. */
6409 if (OVS_LIKELY(dpcls_rule_matches_key(&flow
->cr
, &keys
[i
]) &&
6410 flow
->flow
.in_port
.odp_port
== packet
->md
.in_port
.odp_port
)) {
6411 tcp_flags
= miniflow_get_tcp_flags(&keys
[i
].mf
);
6413 /* SMC hit and emc miss, we insert into EMC */
6415 netdev_flow_key_size(miniflow_n_values(&keys
[i
].mf
));
6416 emc_probabilistic_insert(pmd
, &keys
[i
], flow
);
6417 /* Add these packets into the flow map in the same order
6420 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6421 flow_map
, recv_idx
);
6432 /* SMC missed. Group missed packets together at
6433 * the beginning of the 'packets' array. */
6434 dp_packet_batch_refill(packets_
, packet
, i
);
6436 /* Preserve the order of packet for flow batching. */
6437 index_map
[n_missed
] = recv_idx
;
6439 /* Put missed keys to the pointer arrays return to the caller */
6440 missed_keys
[n_missed
++] = &keys
[i
];
6443 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_SMC_HIT
, n_smc_hit
);
6446 /* Try to process all ('cnt') the 'packets' using only the datapath flow cache
6447 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
6448 * miniflow is copied into 'keys' and the packet pointer is moved at the
6449 * beginning of the 'packets' array. The pointers of missed keys are put in the
6450 * missed_keys pointer array for future processing.
6452 * The function returns the number of packets that needs to be processed in the
6453 * 'packets' array (they have been moved to the beginning of the vector).
6455 * For performance reasons a caller may choose not to initialize the metadata
6456 * in 'packets_'. If 'md_is_valid' is false, the metadata in 'packets'
6457 * is not valid and must be initialized by this function using 'port_no'.
6458 * If 'md_is_valid' is true, the metadata is already valid and 'port_no'
6461 static inline size_t
6462 dfc_processing(struct dp_netdev_pmd_thread
*pmd
,
6463 struct dp_packet_batch
*packets_
,
6464 struct netdev_flow_key
*keys
,
6465 struct netdev_flow_key
**missed_keys
,
6466 struct packet_batch_per_flow batches
[], size_t *n_batches
,
6467 struct dp_packet_flow_map
*flow_map
,
6468 size_t *n_flows
, uint8_t *index_map
,
6469 bool md_is_valid
, odp_port_t port_no
)
6471 struct netdev_flow_key
*key
= &keys
[0];
6472 size_t n_missed
= 0, n_emc_hit
= 0;
6473 struct dfc_cache
*cache
= &pmd
->flow_cache
;
6474 struct dp_packet
*packet
;
6475 const size_t cnt
= dp_packet_batch_size(packets_
);
6476 uint32_t cur_min
= pmd
->ctx
.emc_insert_min
;
6481 bool batch_enable
= true;
6483 atomic_read_relaxed(&pmd
->dp
->smc_enable_db
, &smc_enable_db
);
6484 pmd_perf_update_counter(&pmd
->perf_stats
,
6485 md_is_valid
? PMD_STAT_RECIRC
: PMD_STAT_RECV
,
6488 DP_PACKET_BATCH_REFILL_FOR_EACH (i
, cnt
, packet
, packets_
) {
6489 struct dp_netdev_flow
*flow
;
6492 if (OVS_UNLIKELY(dp_packet_size(packet
) < ETH_HEADER_LEN
)) {
6493 dp_packet_delete(packet
);
6498 struct dp_packet
**packets
= packets_
->packets
;
6499 /* Prefetch next packet data and metadata. */
6500 OVS_PREFETCH(dp_packet_data(packets
[i
+1]));
6501 pkt_metadata_prefetch_init(&packets
[i
+1]->md
);
6505 pkt_metadata_init(&packet
->md
, port_no
);
6508 if ((*recirc_depth_get() == 0) &&
6509 dp_packet_has_flow_mark(packet
, &mark
)) {
6510 flow
= mark_to_flow_find(pmd
, mark
);
6511 if (OVS_LIKELY(flow
)) {
6512 tcp_flags
= parse_tcp_flags(packet
);
6513 if (OVS_LIKELY(batch_enable
)) {
6514 dp_netdev_queue_batches(packet
, flow
, tcp_flags
, batches
,
6517 /* Flow batching should be performed only after fast-path
6518 * processing is also completed for packets with emc miss
6519 * or else it will result in reordering of packets with
6520 * same datapath flows. */
6521 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6522 flow_map
, map_cnt
++);
6528 miniflow_extract(packet
, &key
->mf
);
6529 key
->len
= 0; /* Not computed yet. */
6531 (md_is_valid
== false)
6532 ? dpif_netdev_packet_get_rss_hash_orig_pkt(packet
, &key
->mf
)
6533 : dpif_netdev_packet_get_rss_hash(packet
, &key
->mf
);
6535 /* If EMC is disabled skip emc_lookup */
6536 flow
= (cur_min
!= 0) ? emc_lookup(&cache
->emc_cache
, key
) : NULL
;
6537 if (OVS_LIKELY(flow
)) {
6538 tcp_flags
= miniflow_get_tcp_flags(&key
->mf
);
6540 if (OVS_LIKELY(batch_enable
)) {
6541 dp_netdev_queue_batches(packet
, flow
, tcp_flags
, batches
,
6544 /* Flow batching should be performed only after fast-path
6545 * processing is also completed for packets with emc miss
6546 * or else it will result in reordering of packets with
6547 * same datapath flows. */
6548 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6549 flow_map
, map_cnt
++);
6552 /* Exact match cache missed. Group missed packets together at
6553 * the beginning of the 'packets' array. */
6554 dp_packet_batch_refill(packets_
, packet
, i
);
6556 /* Preserve the order of packet for flow batching. */
6557 index_map
[n_missed
] = map_cnt
;
6558 flow_map
[map_cnt
++].flow
= NULL
;
6560 /* 'key[n_missed]' contains the key of the current packet and it
6561 * will be passed to SMC lookup. The next key should be extracted
6562 * to 'keys[n_missed + 1]'.
6563 * We also maintain a pointer array to keys missed both SMC and EMC
6564 * which will be returned to the caller for future processing. */
6565 missed_keys
[n_missed
] = key
;
6566 key
= &keys
[++n_missed
];
6568 /* Skip batching for subsequent packets to avoid reordering. */
6569 batch_enable
= false;
6572 /* Count of packets which are not flow batched. */
6575 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_EXACT_HIT
, n_emc_hit
);
6577 if (!smc_enable_db
) {
6578 return dp_packet_batch_size(packets_
);
6581 /* Packets miss EMC will do a batch lookup in SMC if enabled */
6582 smc_lookup_batch(pmd
, keys
, missed_keys
, packets_
,
6583 n_missed
, flow_map
, index_map
);
6585 return dp_packet_batch_size(packets_
);
6589 handle_packet_upcall(struct dp_netdev_pmd_thread
*pmd
,
6590 struct dp_packet
*packet
,
6591 const struct netdev_flow_key
*key
,
6592 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
6594 struct ofpbuf
*add_actions
;
6595 struct dp_packet_batch b
;
6599 uint64_t cycles
= cycles_counter_update(&pmd
->perf_stats
);
6601 match
.tun_md
.valid
= false;
6602 miniflow_expand(&key
->mf
, &match
.flow
);
6603 memset(&match
.wc
, 0, sizeof match
.wc
);
6605 ofpbuf_clear(actions
);
6606 ofpbuf_clear(put_actions
);
6608 dpif_flow_hash(pmd
->dp
->dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
6609 error
= dp_netdev_upcall(pmd
, packet
, &match
.flow
, &match
.wc
,
6610 &ufid
, DPIF_UC_MISS
, NULL
, actions
,
6612 if (OVS_UNLIKELY(error
&& error
!= ENOSPC
)) {
6613 dp_packet_delete(packet
);
6617 /* The Netlink encoding of datapath flow keys cannot express
6618 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
6619 * tag is interpreted as exact match on the fact that there is no
6620 * VLAN. Unless we refactor a lot of code that translates between
6621 * Netlink and struct flow representations, we have to do the same
6622 * here. This must be in sync with 'match' in dpif_netdev_flow_put(). */
6623 if (!match
.wc
.masks
.vlans
[0].tci
) {
6624 match
.wc
.masks
.vlans
[0].tci
= htons(0xffff);
6627 /* We can't allow the packet batching in the next loop to execute
6628 * the actions. Otherwise, if there are any slow path actions,
6629 * we'll send the packet up twice. */
6630 dp_packet_batch_init_packet(&b
, packet
);
6631 dp_netdev_execute_actions(pmd
, &b
, true, &match
.flow
,
6632 actions
->data
, actions
->size
);
6634 add_actions
= put_actions
->size
? put_actions
: actions
;
6635 if (OVS_LIKELY(error
!= ENOSPC
)) {
6636 struct dp_netdev_flow
*netdev_flow
;
6638 /* XXX: There's a race window where a flow covering this packet
6639 * could have already been installed since we last did the flow
6640 * lookup before upcall. This could be solved by moving the
6641 * mutex lock outside the loop, but that's an awful long time
6642 * to be locking revalidators out of making flow modifications. */
6643 ovs_mutex_lock(&pmd
->flow_mutex
);
6644 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
6645 if (OVS_LIKELY(!netdev_flow
)) {
6646 netdev_flow
= dp_netdev_flow_add(pmd
, &match
, &ufid
,
6650 ovs_mutex_unlock(&pmd
->flow_mutex
);
6651 uint32_t hash
= dp_netdev_flow_hash(&netdev_flow
->ufid
);
6652 smc_insert(pmd
, key
, hash
);
6653 emc_probabilistic_insert(pmd
, key
, netdev_flow
);
6655 if (pmd_perf_metrics_enabled(pmd
)) {
6656 /* Update upcall stats. */
6657 cycles
= cycles_counter_update(&pmd
->perf_stats
) - cycles
;
6658 struct pmd_perf_stats
*s
= &pmd
->perf_stats
;
6659 s
->current
.upcalls
++;
6660 s
->current
.upcall_cycles
+= cycles
;
6661 histogram_add_sample(&s
->cycles_per_upcall
, cycles
);
6667 fast_path_processing(struct dp_netdev_pmd_thread
*pmd
,
6668 struct dp_packet_batch
*packets_
,
6669 struct netdev_flow_key
**keys
,
6670 struct dp_packet_flow_map
*flow_map
,
6674 const size_t cnt
= dp_packet_batch_size(packets_
);
6675 #if !defined(__CHECKER__) && !defined(_WIN32)
6676 const size_t PKT_ARRAY_SIZE
= cnt
;
6678 /* Sparse or MSVC doesn't like variable length array. */
6679 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
6681 struct dp_packet
*packet
;
6683 struct dpcls_rule
*rules
[PKT_ARRAY_SIZE
];
6684 struct dp_netdev
*dp
= pmd
->dp
;
6685 int upcall_ok_cnt
= 0, upcall_fail_cnt
= 0;
6686 int lookup_cnt
= 0, add_lookup_cnt
;
6689 for (size_t i
= 0; i
< cnt
; i
++) {
6690 /* Key length is needed in all the cases, hash computed on demand. */
6691 keys
[i
]->len
= netdev_flow_key_size(miniflow_n_values(&keys
[i
]->mf
));
6693 /* Get the classifier for the in_port */
6694 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
6695 if (OVS_LIKELY(cls
)) {
6696 any_miss
= !dpcls_lookup(cls
, (const struct netdev_flow_key
**)keys
,
6697 rules
, cnt
, &lookup_cnt
);
6700 memset(rules
, 0, sizeof(rules
));
6702 if (OVS_UNLIKELY(any_miss
) && !fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
6703 uint64_t actions_stub
[512 / 8], slow_stub
[512 / 8];
6704 struct ofpbuf actions
, put_actions
;
6706 ofpbuf_use_stub(&actions
, actions_stub
, sizeof actions_stub
);
6707 ofpbuf_use_stub(&put_actions
, slow_stub
, sizeof slow_stub
);
6709 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
6710 struct dp_netdev_flow
*netdev_flow
;
6712 if (OVS_LIKELY(rules
[i
])) {
6716 /* It's possible that an earlier slow path execution installed
6717 * a rule covering this flow. In this case, it's a lot cheaper
6718 * to catch it here than execute a miss. */
6719 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, keys
[i
],
6722 lookup_cnt
+= add_lookup_cnt
;
6723 rules
[i
] = &netdev_flow
->cr
;
6727 int error
= handle_packet_upcall(pmd
, packet
, keys
[i
],
6728 &actions
, &put_actions
);
6730 if (OVS_UNLIKELY(error
)) {
6737 ofpbuf_uninit(&actions
);
6738 ofpbuf_uninit(&put_actions
);
6739 fat_rwlock_unlock(&dp
->upcall_rwlock
);
6740 } else if (OVS_UNLIKELY(any_miss
)) {
6741 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
6742 if (OVS_UNLIKELY(!rules
[i
])) {
6743 dp_packet_delete(packet
);
6749 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
6750 struct dp_netdev_flow
*flow
;
6751 /* Get the original order of this packet in received batch. */
6752 int recv_idx
= index_map
[i
];
6755 if (OVS_UNLIKELY(!rules
[i
])) {
6759 flow
= dp_netdev_flow_cast(rules
[i
]);
6760 uint32_t hash
= dp_netdev_flow_hash(&flow
->ufid
);
6761 smc_insert(pmd
, keys
[i
], hash
);
6763 emc_probabilistic_insert(pmd
, keys
[i
], flow
);
6764 /* Add these packets into the flow map in the same order
6767 tcp_flags
= miniflow_get_tcp_flags(&keys
[i
]->mf
);
6768 packet_enqueue_to_flow_map(packet
, flow
, tcp_flags
,
6769 flow_map
, recv_idx
);
6772 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_MASKED_HIT
,
6773 cnt
- upcall_ok_cnt
- upcall_fail_cnt
);
6774 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_MASKED_LOOKUP
,
6776 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_MISS
,
6778 pmd_perf_update_counter(&pmd
->perf_stats
, PMD_STAT_LOST
,
6782 /* Packets enter the datapath from a port (or from recirculation) here.
6784 * When 'md_is_valid' is true the metadata in 'packets' are already valid.
6785 * When false the metadata in 'packets' need to be initialized. */
6787 dp_netdev_input__(struct dp_netdev_pmd_thread
*pmd
,
6788 struct dp_packet_batch
*packets
,
6789 bool md_is_valid
, odp_port_t port_no
)
6791 #if !defined(__CHECKER__) && !defined(_WIN32)
6792 const size_t PKT_ARRAY_SIZE
= dp_packet_batch_size(packets
);
6794 /* Sparse or MSVC doesn't like variable length array. */
6795 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
6797 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
)
6798 struct netdev_flow_key keys
[PKT_ARRAY_SIZE
];
6799 struct netdev_flow_key
*missed_keys
[PKT_ARRAY_SIZE
];
6800 struct packet_batch_per_flow batches
[PKT_ARRAY_SIZE
];
6802 struct dp_packet_flow_map flow_map
[PKT_ARRAY_SIZE
];
6803 uint8_t index_map
[PKT_ARRAY_SIZE
];
6809 dfc_processing(pmd
, packets
, keys
, missed_keys
, batches
, &n_batches
,
6810 flow_map
, &n_flows
, index_map
, md_is_valid
, port_no
);
6812 if (!dp_packet_batch_is_empty(packets
)) {
6813 /* Get ingress port from first packet's metadata. */
6814 in_port
= packets
->packets
[0]->md
.in_port
.odp_port
;
6815 fast_path_processing(pmd
, packets
, missed_keys
,
6816 flow_map
, index_map
, in_port
);
6819 /* Batch rest of packets which are in flow map. */
6820 for (i
= 0; i
< n_flows
; i
++) {
6821 struct dp_packet_flow_map
*map
= &flow_map
[i
];
6823 if (OVS_UNLIKELY(!map
->flow
)) {
6826 dp_netdev_queue_batches(map
->packet
, map
->flow
, map
->tcp_flags
,
6827 batches
, &n_batches
);
6830 /* All the flow batches need to be reset before any call to
6831 * packet_batch_per_flow_execute() as it could potentially trigger
6832 * recirculation. When a packet matching flow ‘j’ happens to be
6833 * recirculated, the nested call to dp_netdev_input__() could potentially
6834 * classify the packet as matching another flow - say 'k'. It could happen
6835 * that in the previous call to dp_netdev_input__() that same flow 'k' had
6836 * already its own batches[k] still waiting to be served. So if its
6837 * ‘batch’ member is not reset, the recirculated packet would be wrongly
6838 * appended to batches[k] of the 1st call to dp_netdev_input__(). */
6839 for (i
= 0; i
< n_batches
; i
++) {
6840 batches
[i
].flow
->batch
= NULL
;
6843 for (i
= 0; i
< n_batches
; i
++) {
6844 packet_batch_per_flow_execute(&batches
[i
], pmd
);
6849 dp_netdev_input(struct dp_netdev_pmd_thread
*pmd
,
6850 struct dp_packet_batch
*packets
,
6853 dp_netdev_input__(pmd
, packets
, false, port_no
);
6857 dp_netdev_recirculate(struct dp_netdev_pmd_thread
*pmd
,
6858 struct dp_packet_batch
*packets
)
6860 dp_netdev_input__(pmd
, packets
, true, 0);
6863 struct dp_netdev_execute_aux
{
6864 struct dp_netdev_pmd_thread
*pmd
;
6865 const struct flow
*flow
;
6869 dpif_netdev_register_dp_purge_cb(struct dpif
*dpif
, dp_purge_callback
*cb
,
6872 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
6873 dp
->dp_purge_aux
= aux
;
6874 dp
->dp_purge_cb
= cb
;
6878 dpif_netdev_register_upcall_cb(struct dpif
*dpif
, upcall_callback
*cb
,
6881 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
6882 dp
->upcall_aux
= aux
;
6887 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
6891 struct dp_netdev_port
*port
;
6894 HMAP_FOR_EACH (tx
, node
, &pmd
->send_port_cache
) {
6895 if (!tx
->port
->dynamic_txqs
) {
6898 interval
= pmd
->ctx
.now
- tx
->last_used
;
6899 if (tx
->qid
>= 0 && (purge
|| interval
>= XPS_TIMEOUT
)) {
6901 ovs_mutex_lock(&port
->txq_used_mutex
);
6902 port
->txq_used
[tx
->qid
]--;
6903 ovs_mutex_unlock(&port
->txq_used_mutex
);
6910 dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
6913 struct dp_netdev_port
*port
;
6915 int i
, min_cnt
, min_qid
;
6917 interval
= pmd
->ctx
.now
- tx
->last_used
;
6918 tx
->last_used
= pmd
->ctx
.now
;
6920 if (OVS_LIKELY(tx
->qid
>= 0 && interval
< XPS_TIMEOUT
)) {
6926 ovs_mutex_lock(&port
->txq_used_mutex
);
6928 port
->txq_used
[tx
->qid
]--;
6934 for (i
= 0; i
< netdev_n_txq(port
->netdev
); i
++) {
6935 if (port
->txq_used
[i
] < min_cnt
|| min_cnt
== -1) {
6936 min_cnt
= port
->txq_used
[i
];
6941 port
->txq_used
[min_qid
]++;
6944 ovs_mutex_unlock(&port
->txq_used_mutex
);
6946 dpif_netdev_xps_revalidate_pmd(pmd
, false);
6948 VLOG_DBG("Core %d: New TX queue ID %d for port \'%s\'.",
6949 pmd
->core_id
, tx
->qid
, netdev_get_name(tx
->port
->netdev
));
6953 static struct tx_port
*
6954 pmd_tnl_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
6957 return tx_port_lookup(&pmd
->tnl_port_cache
, port_no
);
6960 static struct tx_port
*
6961 pmd_send_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
6964 return tx_port_lookup(&pmd
->send_port_cache
, port_no
);
6968 push_tnl_action(const struct dp_netdev_pmd_thread
*pmd
,
6969 const struct nlattr
*attr
,
6970 struct dp_packet_batch
*batch
)
6972 struct tx_port
*tun_port
;
6973 const struct ovs_action_push_tnl
*data
;
6976 data
= nl_attr_get(attr
);
6978 tun_port
= pmd_tnl_port_cache_lookup(pmd
, data
->tnl_port
);
6983 err
= netdev_push_header(tun_port
->port
->netdev
, batch
, data
);
6988 dp_packet_delete_batch(batch
, true);
6993 dp_execute_userspace_action(struct dp_netdev_pmd_thread
*pmd
,
6994 struct dp_packet
*packet
, bool should_steal
,
6995 struct flow
*flow
, ovs_u128
*ufid
,
6996 struct ofpbuf
*actions
,
6997 const struct nlattr
*userdata
)
6999 struct dp_packet_batch b
;
7002 ofpbuf_clear(actions
);
7004 error
= dp_netdev_upcall(pmd
, packet
, flow
, NULL
, ufid
,
7005 DPIF_UC_ACTION
, userdata
, actions
,
7007 if (!error
|| error
== ENOSPC
) {
7008 dp_packet_batch_init_packet(&b
, packet
);
7009 dp_netdev_execute_actions(pmd
, &b
, should_steal
, flow
,
7010 actions
->data
, actions
->size
);
7011 } else if (should_steal
) {
7012 dp_packet_delete(packet
);
7017 dp_execute_cb(void *aux_
, struct dp_packet_batch
*packets_
,
7018 const struct nlattr
*a
, bool should_steal
)
7019 OVS_NO_THREAD_SAFETY_ANALYSIS
7021 struct dp_netdev_execute_aux
*aux
= aux_
;
7022 uint32_t *depth
= recirc_depth_get();
7023 struct dp_netdev_pmd_thread
*pmd
= aux
->pmd
;
7024 struct dp_netdev
*dp
= pmd
->dp
;
7025 int type
= nl_attr_type(a
);
7028 switch ((enum ovs_action_attr
)type
) {
7029 case OVS_ACTION_ATTR_OUTPUT
:
7030 p
= pmd_send_port_cache_lookup(pmd
, nl_attr_get_odp_port(a
));
7031 if (OVS_LIKELY(p
)) {
7032 struct dp_packet
*packet
;
7033 struct dp_packet_batch out
;
7035 if (!should_steal
) {
7036 dp_packet_batch_clone(&out
, packets_
);
7037 dp_packet_batch_reset_cutlen(packets_
);
7040 dp_packet_batch_apply_cutlen(packets_
);
7043 if (OVS_UNLIKELY(!dp_packet_batch_is_empty(&p
->output_pkts
)
7044 && packets_
->packets
[0]->source
7045 != p
->output_pkts
.packets
[0]->source
)) {
7046 /* XXX: netdev-dpdk assumes that all packets in a single
7047 * output batch has the same source. Flush here to
7048 * avoid memory access issues. */
7049 dp_netdev_pmd_flush_output_on_port(pmd
, p
);
7052 if (dp_packet_batch_size(&p
->output_pkts
)
7053 + dp_packet_batch_size(packets_
) > NETDEV_MAX_BURST
) {
7054 /* Flush here to avoid overflow. */
7055 dp_netdev_pmd_flush_output_on_port(pmd
, p
);
7058 if (dp_packet_batch_is_empty(&p
->output_pkts
)) {
7059 pmd
->n_output_batches
++;
7062 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7063 p
->output_pkts_rxqs
[dp_packet_batch_size(&p
->output_pkts
)] =
7065 dp_packet_batch_add(&p
->output_pkts
, packet
);
7071 case OVS_ACTION_ATTR_TUNNEL_PUSH
:
7073 /* We're requested to push tunnel header, but also we need to take
7074 * the ownership of these packets. Thus, we can avoid performing
7075 * the action, because the caller will not use the result anyway.
7076 * Just break to free the batch. */
7079 dp_packet_batch_apply_cutlen(packets_
);
7080 push_tnl_action(pmd
, a
, packets_
);
7083 case OVS_ACTION_ATTR_TUNNEL_POP
:
7084 if (*depth
< MAX_RECIRC_DEPTH
) {
7085 struct dp_packet_batch
*orig_packets_
= packets_
;
7086 odp_port_t portno
= nl_attr_get_odp_port(a
);
7088 p
= pmd_tnl_port_cache_lookup(pmd
, portno
);
7090 struct dp_packet_batch tnl_pkt
;
7092 if (!should_steal
) {
7093 dp_packet_batch_clone(&tnl_pkt
, packets_
);
7094 packets_
= &tnl_pkt
;
7095 dp_packet_batch_reset_cutlen(orig_packets_
);
7098 dp_packet_batch_apply_cutlen(packets_
);
7100 netdev_pop_header(p
->port
->netdev
, packets_
);
7101 if (dp_packet_batch_is_empty(packets_
)) {
7105 struct dp_packet
*packet
;
7106 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7107 packet
->md
.in_port
.odp_port
= portno
;
7111 dp_netdev_recirculate(pmd
, packets_
);
7118 case OVS_ACTION_ATTR_USERSPACE
:
7119 if (!fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
7120 struct dp_packet_batch
*orig_packets_
= packets_
;
7121 const struct nlattr
*userdata
;
7122 struct dp_packet_batch usr_pkt
;
7123 struct ofpbuf actions
;
7128 userdata
= nl_attr_find_nested(a
, OVS_USERSPACE_ATTR_USERDATA
);
7129 ofpbuf_init(&actions
, 0);
7131 if (packets_
->trunc
) {
7132 if (!should_steal
) {
7133 dp_packet_batch_clone(&usr_pkt
, packets_
);
7134 packets_
= &usr_pkt
;
7136 dp_packet_batch_reset_cutlen(orig_packets_
);
7139 dp_packet_batch_apply_cutlen(packets_
);
7142 struct dp_packet
*packet
;
7143 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7144 flow_extract(packet
, &flow
);
7145 dpif_flow_hash(dp
->dpif
, &flow
, sizeof flow
, &ufid
);
7146 dp_execute_userspace_action(pmd
, packet
, should_steal
, &flow
,
7147 &ufid
, &actions
, userdata
);
7151 dp_packet_delete_batch(packets_
, true);
7154 ofpbuf_uninit(&actions
);
7155 fat_rwlock_unlock(&dp
->upcall_rwlock
);
7161 case OVS_ACTION_ATTR_RECIRC
:
7162 if (*depth
< MAX_RECIRC_DEPTH
) {
7163 struct dp_packet_batch recirc_pkts
;
7165 if (!should_steal
) {
7166 dp_packet_batch_clone(&recirc_pkts
, packets_
);
7167 packets_
= &recirc_pkts
;
7170 struct dp_packet
*packet
;
7171 DP_PACKET_BATCH_FOR_EACH (i
, packet
, packets_
) {
7172 packet
->md
.recirc_id
= nl_attr_get_u32(a
);
7176 dp_netdev_recirculate(pmd
, packets_
);
7182 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
7185 case OVS_ACTION_ATTR_CT
: {
7186 const struct nlattr
*b
;
7188 bool commit
= false;
7191 const char *helper
= NULL
;
7192 const uint32_t *setmark
= NULL
;
7193 const struct ovs_key_ct_labels
*setlabel
= NULL
;
7194 struct nat_action_info_t nat_action_info
;
7195 struct nat_action_info_t
*nat_action_info_ref
= NULL
;
7196 bool nat_config
= false;
7198 NL_ATTR_FOR_EACH_UNSAFE (b
, left
, nl_attr_get(a
),
7199 nl_attr_get_size(a
)) {
7200 enum ovs_ct_attr sub_type
= nl_attr_type(b
);
7203 case OVS_CT_ATTR_FORCE_COMMIT
:
7206 case OVS_CT_ATTR_COMMIT
:
7209 case OVS_CT_ATTR_ZONE
:
7210 zone
= nl_attr_get_u16(b
);
7212 case OVS_CT_ATTR_HELPER
:
7213 helper
= nl_attr_get_string(b
);
7215 case OVS_CT_ATTR_MARK
:
7216 setmark
= nl_attr_get(b
);
7218 case OVS_CT_ATTR_LABELS
:
7219 setlabel
= nl_attr_get(b
);
7221 case OVS_CT_ATTR_EVENTMASK
:
7222 /* Silently ignored, as userspace datapath does not generate
7223 * netlink events. */
7225 case OVS_CT_ATTR_NAT
: {
7226 const struct nlattr
*b_nest
;
7227 unsigned int left_nest
;
7228 bool ip_min_specified
= false;
7229 bool proto_num_min_specified
= false;
7230 bool ip_max_specified
= false;
7231 bool proto_num_max_specified
= false;
7232 memset(&nat_action_info
, 0, sizeof nat_action_info
);
7233 nat_action_info_ref
= &nat_action_info
;
7235 NL_NESTED_FOR_EACH_UNSAFE (b_nest
, left_nest
, b
) {
7236 enum ovs_nat_attr sub_type_nest
= nl_attr_type(b_nest
);
7238 switch (sub_type_nest
) {
7239 case OVS_NAT_ATTR_SRC
:
7240 case OVS_NAT_ATTR_DST
:
7242 nat_action_info
.nat_action
|=
7243 ((sub_type_nest
== OVS_NAT_ATTR_SRC
)
7244 ? NAT_ACTION_SRC
: NAT_ACTION_DST
);
7246 case OVS_NAT_ATTR_IP_MIN
:
7247 memcpy(&nat_action_info
.min_addr
,
7248 nl_attr_get(b_nest
),
7249 nl_attr_get_size(b_nest
));
7250 ip_min_specified
= true;
7252 case OVS_NAT_ATTR_IP_MAX
:
7253 memcpy(&nat_action_info
.max_addr
,
7254 nl_attr_get(b_nest
),
7255 nl_attr_get_size(b_nest
));
7256 ip_max_specified
= true;
7258 case OVS_NAT_ATTR_PROTO_MIN
:
7259 nat_action_info
.min_port
=
7260 nl_attr_get_u16(b_nest
);
7261 proto_num_min_specified
= true;
7263 case OVS_NAT_ATTR_PROTO_MAX
:
7264 nat_action_info
.max_port
=
7265 nl_attr_get_u16(b_nest
);
7266 proto_num_max_specified
= true;
7268 case OVS_NAT_ATTR_PERSISTENT
:
7269 case OVS_NAT_ATTR_PROTO_HASH
:
7270 case OVS_NAT_ATTR_PROTO_RANDOM
:
7272 case OVS_NAT_ATTR_UNSPEC
:
7273 case __OVS_NAT_ATTR_MAX
:
7278 if (ip_min_specified
&& !ip_max_specified
) {
7279 nat_action_info
.max_addr
= nat_action_info
.min_addr
;
7281 if (proto_num_min_specified
&& !proto_num_max_specified
) {
7282 nat_action_info
.max_port
= nat_action_info
.min_port
;
7284 if (proto_num_min_specified
|| proto_num_max_specified
) {
7285 if (nat_action_info
.nat_action
& NAT_ACTION_SRC
) {
7286 nat_action_info
.nat_action
|= NAT_ACTION_SRC_PORT
;
7287 } else if (nat_action_info
.nat_action
& NAT_ACTION_DST
) {
7288 nat_action_info
.nat_action
|= NAT_ACTION_DST_PORT
;
7293 case OVS_CT_ATTR_UNSPEC
:
7294 case __OVS_CT_ATTR_MAX
:
7299 /* We won't be able to function properly in this case, hence
7300 * complain loudly. */
7301 if (nat_config
&& !commit
) {
7302 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(5, 5);
7303 VLOG_WARN_RL(&rl
, "NAT specified without commit.");
7306 conntrack_execute(dp
->conntrack
, packets_
, aux
->flow
->dl_type
, force
,
7307 commit
, zone
, setmark
, setlabel
, aux
->flow
->tp_src
,
7308 aux
->flow
->tp_dst
, helper
, nat_action_info_ref
,
7309 pmd
->ctx
.now
/ 1000);
7313 case OVS_ACTION_ATTR_METER
:
7314 dp_netdev_run_meter(pmd
->dp
, packets_
, nl_attr_get_u32(a
),
7318 case OVS_ACTION_ATTR_PUSH_VLAN
:
7319 case OVS_ACTION_ATTR_POP_VLAN
:
7320 case OVS_ACTION_ATTR_PUSH_MPLS
:
7321 case OVS_ACTION_ATTR_POP_MPLS
:
7322 case OVS_ACTION_ATTR_SET
:
7323 case OVS_ACTION_ATTR_SET_MASKED
:
7324 case OVS_ACTION_ATTR_SAMPLE
:
7325 case OVS_ACTION_ATTR_HASH
:
7326 case OVS_ACTION_ATTR_UNSPEC
:
7327 case OVS_ACTION_ATTR_TRUNC
:
7328 case OVS_ACTION_ATTR_PUSH_ETH
:
7329 case OVS_ACTION_ATTR_POP_ETH
:
7330 case OVS_ACTION_ATTR_CLONE
:
7331 case OVS_ACTION_ATTR_PUSH_NSH
:
7332 case OVS_ACTION_ATTR_POP_NSH
:
7333 case OVS_ACTION_ATTR_CT_CLEAR
:
7334 case OVS_ACTION_ATTR_CHECK_PKT_LEN
:
7335 case __OVS_ACTION_ATTR_MAX
:
7339 dp_packet_delete_batch(packets_
, should_steal
);
7343 dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
7344 struct dp_packet_batch
*packets
,
7345 bool should_steal
, const struct flow
*flow
,
7346 const struct nlattr
*actions
, size_t actions_len
)
7348 struct dp_netdev_execute_aux aux
= { pmd
, flow
};
7350 odp_execute_actions(&aux
, packets
, should_steal
, actions
,
7351 actions_len
, dp_execute_cb
);
7354 struct dp_netdev_ct_dump
{
7355 struct ct_dpif_dump_state up
;
7356 struct conntrack_dump dump
;
7357 struct conntrack
*ct
;
7358 struct dp_netdev
*dp
;
7362 dpif_netdev_ct_dump_start(struct dpif
*dpif
, struct ct_dpif_dump_state
**dump_
,
7363 const uint16_t *pzone
, int *ptot_bkts
)
7365 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7366 struct dp_netdev_ct_dump
*dump
;
7368 dump
= xzalloc(sizeof *dump
);
7370 dump
->ct
= dp
->conntrack
;
7372 conntrack_dump_start(dp
->conntrack
, &dump
->dump
, pzone
, ptot_bkts
);
7380 dpif_netdev_ct_dump_next(struct dpif
*dpif OVS_UNUSED
,
7381 struct ct_dpif_dump_state
*dump_
,
7382 struct ct_dpif_entry
*entry
)
7384 struct dp_netdev_ct_dump
*dump
;
7386 INIT_CONTAINER(dump
, dump_
, up
);
7388 return conntrack_dump_next(&dump
->dump
, entry
);
7392 dpif_netdev_ct_dump_done(struct dpif
*dpif OVS_UNUSED
,
7393 struct ct_dpif_dump_state
*dump_
)
7395 struct dp_netdev_ct_dump
*dump
;
7398 INIT_CONTAINER(dump
, dump_
, up
);
7400 err
= conntrack_dump_done(&dump
->dump
);
7408 dpif_netdev_ct_flush(struct dpif
*dpif
, const uint16_t *zone
,
7409 const struct ct_dpif_tuple
*tuple
)
7411 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7414 return conntrack_flush_tuple(dp
->conntrack
, tuple
, zone
? *zone
: 0);
7416 return conntrack_flush(dp
->conntrack
, zone
);
7420 dpif_netdev_ct_set_maxconns(struct dpif
*dpif
, uint32_t maxconns
)
7422 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7424 return conntrack_set_maxconns(dp
->conntrack
, maxconns
);
7428 dpif_netdev_ct_get_maxconns(struct dpif
*dpif
, uint32_t *maxconns
)
7430 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7432 return conntrack_get_maxconns(dp
->conntrack
, maxconns
);
7436 dpif_netdev_ct_get_nconns(struct dpif
*dpif
, uint32_t *nconns
)
7438 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7440 return conntrack_get_nconns(dp
->conntrack
, nconns
);
7444 dpif_netdev_ct_set_tcp_seq_chk(struct dpif
*dpif
, bool enabled
)
7446 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7448 return conntrack_set_tcp_seq_chk(dp
->conntrack
, enabled
);
7452 dpif_netdev_ct_get_tcp_seq_chk(struct dpif
*dpif
, bool *enabled
)
7454 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7455 *enabled
= conntrack_get_tcp_seq_chk(dp
->conntrack
);
7460 dpif_netdev_ipf_set_enabled(struct dpif
*dpif
, bool v6
, bool enable
)
7462 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7463 return ipf_set_enabled(conntrack_ipf_ctx(dp
->conntrack
), v6
, enable
);
7467 dpif_netdev_ipf_set_min_frag(struct dpif
*dpif
, bool v6
, uint32_t min_frag
)
7469 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7470 return ipf_set_min_frag(conntrack_ipf_ctx(dp
->conntrack
), v6
, min_frag
);
7474 dpif_netdev_ipf_set_max_nfrags(struct dpif
*dpif
, uint32_t max_frags
)
7476 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7477 return ipf_set_max_nfrags(conntrack_ipf_ctx(dp
->conntrack
), max_frags
);
7480 /* Adjust this function if 'dpif_ipf_status' and 'ipf_status' were to
7483 dpif_netdev_ipf_get_status(struct dpif
*dpif
,
7484 struct dpif_ipf_status
*dpif_ipf_status
)
7486 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7487 ipf_get_status(conntrack_ipf_ctx(dp
->conntrack
),
7488 (struct ipf_status
*) dpif_ipf_status
);
7493 dpif_netdev_ipf_dump_start(struct dpif
*dpif OVS_UNUSED
,
7494 struct ipf_dump_ctx
**ipf_dump_ctx
)
7496 return ipf_dump_start(ipf_dump_ctx
);
7500 dpif_netdev_ipf_dump_next(struct dpif
*dpif
, void *ipf_dump_ctx
, char **dump
)
7502 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
7503 return ipf_dump_next(conntrack_ipf_ctx(dp
->conntrack
), ipf_dump_ctx
,
7508 dpif_netdev_ipf_dump_done(struct dpif
*dpif OVS_UNUSED
, void *ipf_dump_ctx
)
7510 return ipf_dump_done(ipf_dump_ctx
);
7514 const struct dpif_class dpif_netdev_class
= {
7516 true, /* cleanup_required */
7518 dpif_netdev_enumerate
,
7519 dpif_netdev_port_open_type
,
7522 dpif_netdev_destroy
,
7525 dpif_netdev_get_stats
,
7526 dpif_netdev_port_add
,
7527 dpif_netdev_port_del
,
7528 dpif_netdev_port_set_config
,
7529 dpif_netdev_port_query_by_number
,
7530 dpif_netdev_port_query_by_name
,
7531 NULL
, /* port_get_pid */
7532 dpif_netdev_port_dump_start
,
7533 dpif_netdev_port_dump_next
,
7534 dpif_netdev_port_dump_done
,
7535 dpif_netdev_port_poll
,
7536 dpif_netdev_port_poll_wait
,
7537 dpif_netdev_flow_flush
,
7538 dpif_netdev_flow_dump_create
,
7539 dpif_netdev_flow_dump_destroy
,
7540 dpif_netdev_flow_dump_thread_create
,
7541 dpif_netdev_flow_dump_thread_destroy
,
7542 dpif_netdev_flow_dump_next
,
7543 dpif_netdev_operate
,
7544 NULL
, /* recv_set */
7545 NULL
, /* handlers_set */
7546 dpif_netdev_set_config
,
7547 dpif_netdev_queue_to_priority
,
7549 NULL
, /* recv_wait */
7550 NULL
, /* recv_purge */
7551 dpif_netdev_register_dp_purge_cb
,
7552 dpif_netdev_register_upcall_cb
,
7553 dpif_netdev_enable_upcall
,
7554 dpif_netdev_disable_upcall
,
7555 dpif_netdev_get_datapath_version
,
7556 dpif_netdev_ct_dump_start
,
7557 dpif_netdev_ct_dump_next
,
7558 dpif_netdev_ct_dump_done
,
7559 dpif_netdev_ct_flush
,
7560 dpif_netdev_ct_set_maxconns
,
7561 dpif_netdev_ct_get_maxconns
,
7562 dpif_netdev_ct_get_nconns
,
7563 dpif_netdev_ct_set_tcp_seq_chk
,
7564 dpif_netdev_ct_get_tcp_seq_chk
,
7565 NULL
, /* ct_set_limits */
7566 NULL
, /* ct_get_limits */
7567 NULL
, /* ct_del_limits */
7568 dpif_netdev_ipf_set_enabled
,
7569 dpif_netdev_ipf_set_min_frag
,
7570 dpif_netdev_ipf_set_max_nfrags
,
7571 dpif_netdev_ipf_get_status
,
7572 dpif_netdev_ipf_dump_start
,
7573 dpif_netdev_ipf_dump_next
,
7574 dpif_netdev_ipf_dump_done
,
7575 dpif_netdev_meter_get_features
,
7576 dpif_netdev_meter_set
,
7577 dpif_netdev_meter_get
,
7578 dpif_netdev_meter_del
,
7582 dpif_dummy_change_port_number(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
7583 const char *argv
[], void *aux OVS_UNUSED
)
7585 struct dp_netdev_port
*port
;
7586 struct dp_netdev
*dp
;
7589 ovs_mutex_lock(&dp_netdev_mutex
);
7590 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
7591 if (!dp
|| !dpif_netdev_class_is_dummy(dp
->class)) {
7592 ovs_mutex_unlock(&dp_netdev_mutex
);
7593 unixctl_command_reply_error(conn
, "unknown datapath or not a dummy");
7596 ovs_refcount_ref(&dp
->ref_cnt
);
7597 ovs_mutex_unlock(&dp_netdev_mutex
);
7599 ovs_mutex_lock(&dp
->port_mutex
);
7600 if (get_port_by_name(dp
, argv
[2], &port
)) {
7601 unixctl_command_reply_error(conn
, "unknown port");
7605 port_no
= u32_to_odp(atoi(argv
[3]));
7606 if (!port_no
|| port_no
== ODPP_NONE
) {
7607 unixctl_command_reply_error(conn
, "bad port number");
7610 if (dp_netdev_lookup_port(dp
, port_no
)) {
7611 unixctl_command_reply_error(conn
, "port number already in use");
7616 hmap_remove(&dp
->ports
, &port
->node
);
7617 reconfigure_datapath(dp
);
7619 /* Reinsert with new port number. */
7620 port
->port_no
= port_no
;
7621 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
7622 reconfigure_datapath(dp
);
7624 seq_change(dp
->port_seq
);
7625 unixctl_command_reply(conn
, NULL
);
7628 ovs_mutex_unlock(&dp
->port_mutex
);
7629 dp_netdev_unref(dp
);
7633 dpif_dummy_register__(const char *type
)
7635 struct dpif_class
*class;
7637 class = xmalloc(sizeof *class);
7638 *class = dpif_netdev_class
;
7639 class->type
= xstrdup(type
);
7640 dp_register_provider(class);
7644 dpif_dummy_override(const char *type
)
7649 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
7650 * a userland-only build. It's useful for testsuite.
7652 error
= dp_unregister_provider(type
);
7653 if (error
== 0 || error
== EAFNOSUPPORT
) {
7654 dpif_dummy_register__(type
);
7659 dpif_dummy_register(enum dummy_level level
)
7661 if (level
== DUMMY_OVERRIDE_ALL
) {
7666 dp_enumerate_types(&types
);
7667 SSET_FOR_EACH (type
, &types
) {
7668 dpif_dummy_override(type
);
7670 sset_destroy(&types
);
7671 } else if (level
== DUMMY_OVERRIDE_SYSTEM
) {
7672 dpif_dummy_override("system");
7675 dpif_dummy_register__("dummy");
7677 unixctl_command_register("dpif-dummy/change-port-number",
7678 "dp port new-number",
7679 3, 3, dpif_dummy_change_port_number
, NULL
);
7682 /* Datapath Classifier. */
7685 dpcls_subtable_destroy_cb(struct dpcls_subtable
*subtable
)
7687 cmap_destroy(&subtable
->rules
);
7688 ovsrcu_postpone(free
, subtable
->mf_masks
);
7689 ovsrcu_postpone(free
, subtable
);
7692 /* Initializes 'cls' as a classifier that initially contains no classification
7695 dpcls_init(struct dpcls
*cls
)
7697 cmap_init(&cls
->subtables_map
);
7698 pvector_init(&cls
->subtables
);
7702 dpcls_destroy_subtable(struct dpcls
*cls
, struct dpcls_subtable
*subtable
)
7704 VLOG_DBG("Destroying subtable %p for in_port %d", subtable
, cls
->in_port
);
7705 pvector_remove(&cls
->subtables
, subtable
);
7706 cmap_remove(&cls
->subtables_map
, &subtable
->cmap_node
,
7707 subtable
->mask
.hash
);
7708 ovsrcu_postpone(dpcls_subtable_destroy_cb
, subtable
);
7711 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
7712 * caller's responsibility.
7713 * May only be called after all the readers have been terminated. */
7715 dpcls_destroy(struct dpcls
*cls
)
7718 struct dpcls_subtable
*subtable
;
7720 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
7721 ovs_assert(cmap_count(&subtable
->rules
) == 0);
7722 dpcls_destroy_subtable(cls
, subtable
);
7724 cmap_destroy(&cls
->subtables_map
);
7725 pvector_destroy(&cls
->subtables
);
7729 static struct dpcls_subtable
*
7730 dpcls_create_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
7732 struct dpcls_subtable
*subtable
;
7734 /* Need to add one. */
7735 subtable
= xmalloc(sizeof *subtable
7736 - sizeof subtable
->mask
.mf
+ mask
->len
);
7737 cmap_init(&subtable
->rules
);
7738 subtable
->hit_cnt
= 0;
7739 netdev_flow_key_clone(&subtable
->mask
, mask
);
7741 /* The count of bits in the mask defines the space required for masks.
7742 * Then call gen_masks() to create the appropriate masks, avoiding the cost
7743 * of doing runtime calculations. */
7744 uint32_t unit0
= count_1bits(mask
->mf
.map
.bits
[0]);
7745 uint32_t unit1
= count_1bits(mask
->mf
.map
.bits
[1]);
7746 subtable
->mf_bits_set_unit0
= unit0
;
7747 subtable
->mf_bits_set_unit1
= unit1
;
7748 subtable
->mf_masks
= xmalloc(sizeof(uint64_t) * (unit0
+ unit1
));
7749 netdev_flow_key_gen_masks(mask
, subtable
->mf_masks
, unit0
, unit1
);
7751 /* Probe for a specialized generic lookup function. */
7752 subtable
->lookup_func
= dpcls_subtable_generic_probe(unit0
, unit1
);
7754 /* If not set, assign generic lookup. Generic works for any miniflow. */
7755 if (!subtable
->lookup_func
) {
7756 subtable
->lookup_func
= dpcls_subtable_lookup_generic
;
7759 cmap_insert(&cls
->subtables_map
, &subtable
->cmap_node
, mask
->hash
);
7760 /* Add the new subtable at the end of the pvector (with no hits yet) */
7761 pvector_insert(&cls
->subtables
, subtable
, 0);
7762 VLOG_DBG("Creating %"PRIuSIZE
". subtable %p for in_port %d",
7763 cmap_count(&cls
->subtables_map
), subtable
, cls
->in_port
);
7764 pvector_publish(&cls
->subtables
);
7769 static inline struct dpcls_subtable
*
7770 dpcls_find_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
7772 struct dpcls_subtable
*subtable
;
7774 CMAP_FOR_EACH_WITH_HASH (subtable
, cmap_node
, mask
->hash
,
7775 &cls
->subtables_map
) {
7776 if (netdev_flow_key_equal(&subtable
->mask
, mask
)) {
7780 return dpcls_create_subtable(cls
, mask
);
7784 /* Periodically sort the dpcls subtable vectors according to hit counts */
7786 dpcls_sort_subtable_vector(struct dpcls
*cls
)
7788 struct pvector
*pvec
= &cls
->subtables
;
7789 struct dpcls_subtable
*subtable
;
7791 PVECTOR_FOR_EACH (subtable
, pvec
) {
7792 pvector_change_priority(pvec
, subtable
, subtable
->hit_cnt
);
7793 subtable
->hit_cnt
= 0;
7795 pvector_publish(pvec
);
7799 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
,
7800 struct polled_queue
*poll_list
, int poll_cnt
)
7803 uint64_t tot_idle
= 0, tot_proc
= 0;
7804 unsigned int pmd_load
= 0;
7806 if (pmd
->ctx
.now
> pmd
->rxq_next_cycle_store
) {
7808 struct pmd_auto_lb
*pmd_alb
= &pmd
->dp
->pmd_alb
;
7809 if (pmd_alb
->is_enabled
&& !pmd
->isolated
7810 && (pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_IDLE
] >=
7811 pmd
->prev_stats
[PMD_CYCLES_ITER_IDLE
])
7812 && (pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_BUSY
] >=
7813 pmd
->prev_stats
[PMD_CYCLES_ITER_BUSY
]))
7815 tot_idle
= pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_IDLE
] -
7816 pmd
->prev_stats
[PMD_CYCLES_ITER_IDLE
];
7817 tot_proc
= pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_BUSY
] -
7818 pmd
->prev_stats
[PMD_CYCLES_ITER_BUSY
];
7821 pmd_load
= ((tot_proc
* 100) / (tot_idle
+ tot_proc
));
7824 if (pmd_load
>= ALB_PMD_LOAD_THRESHOLD
) {
7825 atomic_count_inc(&pmd
->pmd_overloaded
);
7827 atomic_count_set(&pmd
->pmd_overloaded
, 0);
7831 pmd
->prev_stats
[PMD_CYCLES_ITER_IDLE
] =
7832 pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_IDLE
];
7833 pmd
->prev_stats
[PMD_CYCLES_ITER_BUSY
] =
7834 pmd
->perf_stats
.counters
.n
[PMD_CYCLES_ITER_BUSY
];
7836 /* Get the cycles that were used to process each queue and store. */
7837 for (unsigned i
= 0; i
< poll_cnt
; i
++) {
7838 uint64_t rxq_cyc_curr
= dp_netdev_rxq_get_cycles(poll_list
[i
].rxq
,
7839 RXQ_CYCLES_PROC_CURR
);
7840 dp_netdev_rxq_set_intrvl_cycles(poll_list
[i
].rxq
, rxq_cyc_curr
);
7841 dp_netdev_rxq_set_cycles(poll_list
[i
].rxq
, RXQ_CYCLES_PROC_CURR
,
7844 curr_tsc
= cycles_counter_update(&pmd
->perf_stats
);
7845 if (pmd
->intrvl_tsc_prev
) {
7846 /* There is a prev timestamp, store a new intrvl cycle count. */
7847 atomic_store_relaxed(&pmd
->intrvl_cycles
,
7848 curr_tsc
- pmd
->intrvl_tsc_prev
);
7850 pmd
->intrvl_tsc_prev
= curr_tsc
;
7851 /* Start new measuring interval */
7852 pmd
->rxq_next_cycle_store
= pmd
->ctx
.now
+ PMD_RXQ_INTERVAL_LEN
;
7855 if (pmd
->ctx
.now
> pmd
->next_optimization
) {
7856 /* Try to obtain the flow lock to block out revalidator threads.
7857 * If not possible, just try next time. */
7858 if (!ovs_mutex_trylock(&pmd
->flow_mutex
)) {
7859 /* Optimize each classifier */
7860 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
7861 dpcls_sort_subtable_vector(cls
);
7863 ovs_mutex_unlock(&pmd
->flow_mutex
);
7864 /* Start new measuring interval */
7865 pmd
->next_optimization
= pmd
->ctx
.now
7866 + DPCLS_OPTIMIZATION_INTERVAL
;
7871 /* Insert 'rule' into 'cls'. */
7873 dpcls_insert(struct dpcls
*cls
, struct dpcls_rule
*rule
,
7874 const struct netdev_flow_key
*mask
)
7876 struct dpcls_subtable
*subtable
= dpcls_find_subtable(cls
, mask
);
7878 /* Refer to subtable's mask, also for later removal. */
7879 rule
->mask
= &subtable
->mask
;
7880 cmap_insert(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
);
7883 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
7885 dpcls_remove(struct dpcls
*cls
, struct dpcls_rule
*rule
)
7887 struct dpcls_subtable
*subtable
;
7889 ovs_assert(rule
->mask
);
7891 /* Get subtable from reference in rule->mask. */
7892 INIT_CONTAINER(subtable
, rule
->mask
, mask
);
7893 if (cmap_remove(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
)
7895 /* Delete empty subtable. */
7896 dpcls_destroy_subtable(cls
, subtable
);
7897 pvector_publish(&cls
->subtables
);
7901 /* Inner loop for mask generation of a unit, see netdev_flow_key_gen_masks. */
7903 netdev_flow_key_gen_mask_unit(uint64_t iter
,
7904 const uint64_t count
,
7908 for (i
= 0; i
< count
; i
++) {
7909 uint64_t lowest_bit
= (iter
& -iter
);
7910 iter
&= ~lowest_bit
;
7911 mf_masks
[i
] = (lowest_bit
- 1);
7913 /* Checks that count has covered all bits in the iter bitmap. */
7914 ovs_assert(iter
== 0);
7917 /* Generate a mask for each block in the miniflow, based on the bits set. This
7918 * allows easily masking packets with the generated array here, without
7919 * calculations. This replaces runtime-calculating the masks.
7920 * @param key The table to generate the mf_masks for
7921 * @param mf_masks Pointer to a u64 array of at least *mf_bits* in size
7922 * @param mf_bits_total Number of bits set in the whole miniflow (both units)
7923 * @param mf_bits_unit0 Number of bits set in unit0 of the miniflow
7926 netdev_flow_key_gen_masks(const struct netdev_flow_key
*tbl
,
7928 const uint32_t mf_bits_u0
,
7929 const uint32_t mf_bits_u1
)
7931 uint64_t iter_u0
= tbl
->mf
.map
.bits
[0];
7932 uint64_t iter_u1
= tbl
->mf
.map
.bits
[1];
7934 netdev_flow_key_gen_mask_unit(iter_u0
, mf_bits_u0
, &mf_masks
[0]);
7935 netdev_flow_key_gen_mask_unit(iter_u1
, mf_bits_u1
, &mf_masks
[mf_bits_u0
]);
7938 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
7939 * in 'mask' the values in 'key' and 'target' are the same. */
7941 dpcls_rule_matches_key(const struct dpcls_rule
*rule
,
7942 const struct netdev_flow_key
*target
)
7944 const uint64_t *keyp
= miniflow_get_values(&rule
->flow
.mf
);
7945 const uint64_t *maskp
= miniflow_get_values(&rule
->mask
->mf
);
7948 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, target
, rule
->flow
.mf
.map
) {
7949 if (OVS_UNLIKELY((value
& *maskp
++) != *keyp
++)) {
7956 /* For each miniflow in 'keys' performs a classifier lookup writing the result
7957 * into the corresponding slot in 'rules'. If a particular entry in 'keys' is
7958 * NULL it is skipped.
7960 * This function is optimized for use in the userspace datapath and therefore
7961 * does not implement a lot of features available in the standard
7962 * classifier_lookup() function. Specifically, it does not implement
7963 * priorities, instead returning any rule which matches the flow.
7965 * Returns true if all miniflows found a corresponding rule. */
7967 dpcls_lookup(struct dpcls
*cls
, const struct netdev_flow_key
*keys
[],
7968 struct dpcls_rule
**rules
, const size_t cnt
,
7971 /* The received 'cnt' miniflows are the search-keys that will be processed
7972 * to find a matching entry into the available subtables.
7973 * The number of bits in map_type is equal to NETDEV_MAX_BURST. */
7974 #define MAP_BITS (sizeof(uint32_t) * CHAR_BIT)
7975 BUILD_ASSERT_DECL(MAP_BITS
>= NETDEV_MAX_BURST
);
7977 struct dpcls_subtable
*subtable
;
7978 uint32_t keys_map
= TYPE_MAXIMUM(uint32_t); /* Set all bits. */
7980 if (cnt
!= MAP_BITS
) {
7981 keys_map
>>= MAP_BITS
- cnt
; /* Clear extra bits. */
7983 memset(rules
, 0, cnt
* sizeof *rules
);
7985 int lookups_match
= 0, subtable_pos
= 1;
7988 /* The Datapath classifier - aka dpcls - is composed of subtables.
7989 * Subtables are dynamically created as needed when new rules are inserted.
7990 * Each subtable collects rules with matches on a specific subset of packet
7991 * fields as defined by the subtable's mask. We proceed to process every
7992 * search-key against each subtable, but when a match is found for a
7993 * search-key, the search for that key can stop because the rules are
7994 * non-overlapping. */
7995 PVECTOR_FOR_EACH (subtable
, &cls
->subtables
) {
7996 /* Call the subtable specific lookup function. */
7997 found_map
= subtable
->lookup_func(subtable
, keys_map
, keys
, rules
);
7999 /* Count the number of subtables searched for this packet match. This
8000 * estimates the "spread" of subtables looked at per matched packet. */
8001 uint32_t pkts_matched
= count_1bits(found_map
);
8002 lookups_match
+= pkts_matched
* subtable_pos
;
8004 /* Clear the found rules, and return early if all packets are found. */
8005 keys_map
&= ~found_map
;
8007 if (num_lookups_p
) {
8008 *num_lookups_p
= lookups_match
;
8015 if (num_lookups_p
) {
8016 *num_lookups_p
= lookups_match
;