2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2016 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "dpif-netdev.h"
25 #include <netinet/in.h>
29 #include <sys/ioctl.h>
30 #include <sys/socket.h>
35 #include <rte_cycles.h>
40 #include "conntrack.h"
44 #include "dp-packet.h"
46 #include "dpif-provider.h"
48 #include "fat-rwlock.h"
53 #include "netdev-vport.h"
55 #include "odp-execute.h"
57 #include "openvswitch/dynamic-string.h"
58 #include "openvswitch/list.h"
59 #include "openvswitch/match.h"
60 #include "openvswitch/ofp-print.h"
61 #include "openvswitch/ofp-util.h"
62 #include "openvswitch/ofpbuf.h"
63 #include "openvswitch/shash.h"
64 #include "openvswitch/vlog.h"
68 #include "poll-loop.h"
75 #include "tnl-neigh-cache.h"
76 #include "tnl-ports.h"
80 VLOG_DEFINE_THIS_MODULE(dpif_netdev
);
82 #define FLOW_DUMP_MAX_BATCH 50
83 /* Use per thread recirc_depth to prevent recirculation loop. */
84 #define MAX_RECIRC_DEPTH 5
85 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth
, 0)
87 /* Configuration parameters. */
88 enum { MAX_FLOWS
= 65536 }; /* Maximum number of flows in flow table. */
90 /* Protects against changes to 'dp_netdevs'. */
91 static struct ovs_mutex dp_netdev_mutex
= OVS_MUTEX_INITIALIZER
;
93 /* Contains all 'struct dp_netdev's. */
94 static struct shash dp_netdevs
OVS_GUARDED_BY(dp_netdev_mutex
)
95 = SHASH_INITIALIZER(&dp_netdevs
);
97 static struct vlog_rate_limit upcall_rl
= VLOG_RATE_LIMIT_INIT(600, 600);
99 #define DP_NETDEV_CS_SUPPORTED_MASK (CS_NEW | CS_ESTABLISHED | CS_RELATED \
100 | CS_INVALID | CS_REPLY_DIR | CS_TRACKED)
101 #define DP_NETDEV_CS_UNSUPPORTED_MASK (~(uint32_t)DP_NETDEV_CS_SUPPORTED_MASK)
103 static struct odp_support dp_netdev_support
= {
104 .max_mpls_depth
= SIZE_MAX
,
112 /* Stores a miniflow with inline values */
114 struct netdev_flow_key
{
115 uint32_t hash
; /* Hash function differs for different users. */
116 uint32_t len
; /* Length of the following miniflow (incl. map). */
118 uint64_t buf
[FLOW_MAX_PACKET_U64S
];
121 /* Exact match cache for frequently used flows
123 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
124 * search its entries for a miniflow that matches exactly the miniflow of the
125 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
127 * A cache entry holds a reference to its 'dp_netdev_flow'.
129 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
130 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
131 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
132 * value is the index of a cache entry where the miniflow could be.
138 * Each pmd_thread has its own private exact match cache.
139 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
142 #define EM_FLOW_HASH_SHIFT 13
143 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
144 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
145 #define EM_FLOW_HASH_SEGS 2
148 struct dp_netdev_flow
*flow
;
149 struct netdev_flow_key key
; /* key.hash used for emc hash value. */
153 struct emc_entry entries
[EM_FLOW_HASH_ENTRIES
];
154 int sweep_idx
; /* For emc_cache_slow_sweep(). */
157 /* Iterate in the exact match cache through every entry that might contain a
158 * miniflow with hash 'HASH'. */
159 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
160 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
161 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
162 i__ < EM_FLOW_HASH_SEGS; \
163 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
165 /* Simple non-wildcarding single-priority classifier. */
167 /* Time in ms between successive optimizations of the dpcls subtable vector */
168 #define DPCLS_OPTIMIZATION_INTERVAL 1000
171 struct cmap_node node
; /* Within dp_netdev_pmd_thread.classifiers */
173 struct cmap subtables_map
;
174 struct pvector subtables
;
177 /* A rule to be inserted to the classifier. */
179 struct cmap_node cmap_node
; /* Within struct dpcls_subtable 'rules'. */
180 struct netdev_flow_key
*mask
; /* Subtable's mask. */
181 struct netdev_flow_key flow
; /* Matching key. */
182 /* 'flow' must be the last field, additional space is allocated here. */
185 static void dpcls_init(struct dpcls
*);
186 static void dpcls_destroy(struct dpcls
*);
187 static void dpcls_sort_subtable_vector(struct dpcls
*);
188 static void dpcls_insert(struct dpcls
*, struct dpcls_rule
*,
189 const struct netdev_flow_key
*mask
);
190 static void dpcls_remove(struct dpcls
*, struct dpcls_rule
*);
191 static bool dpcls_lookup(struct dpcls
*cls
,
192 const struct netdev_flow_key keys
[],
193 struct dpcls_rule
**rules
, size_t cnt
,
196 /* Datapath based on the network device interface from netdev.h.
202 * Some members, marked 'const', are immutable. Accessing other members
203 * requires synchronization, as noted in more detail below.
205 * Acquisition order is, from outermost to innermost:
207 * dp_netdev_mutex (global)
212 const struct dpif_class
*const class;
213 const char *const name
;
215 struct ovs_refcount ref_cnt
;
216 atomic_flag destroyed
;
220 * Any lookup into 'ports' or any access to the dp_netdev_ports found
221 * through 'ports' requires taking 'port_mutex'. */
222 struct ovs_mutex port_mutex
;
224 struct seq
*port_seq
; /* Incremented whenever a port changes. */
226 /* Protects access to ofproto-dpif-upcall interface during revalidator
227 * thread synchronization. */
228 struct fat_rwlock upcall_rwlock
;
229 upcall_callback
*upcall_cb
; /* Callback function for executing upcalls. */
232 /* Callback function for notifying the purging of dp flows (during
233 * reseting pmd deletion). */
234 dp_purge_callback
*dp_purge_cb
;
237 /* Stores all 'struct dp_netdev_pmd_thread's. */
238 struct cmap poll_threads
;
240 /* Protects the access of the 'struct dp_netdev_pmd_thread'
241 * instance for non-pmd thread. */
242 struct ovs_mutex non_pmd_mutex
;
244 /* Each pmd thread will store its pointer to
245 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
246 ovsthread_key_t per_pmd_key
;
248 struct seq
*reconfigure_seq
;
249 uint64_t last_reconfigure_seq
;
251 /* Cpu mask for pin of pmd threads. */
254 uint64_t last_tnl_conf_seq
;
256 struct conntrack conntrack
;
259 static struct dp_netdev_port
*dp_netdev_lookup_port(const struct dp_netdev
*dp
,
261 OVS_REQUIRES(dp
->port_mutex
);
264 DP_STAT_EXACT_HIT
, /* Packets that had an exact match (emc). */
265 DP_STAT_MASKED_HIT
, /* Packets that matched in the flow table. */
266 DP_STAT_MISS
, /* Packets that did not match. */
267 DP_STAT_LOST
, /* Packets not passed up to the client. */
268 DP_STAT_LOOKUP_HIT
, /* Number of subtable lookups for flow table
273 enum pmd_cycles_counter_type
{
274 PMD_CYCLES_POLLING
, /* Cycles spent polling NICs. */
275 PMD_CYCLES_PROCESSING
, /* Cycles spent processing packets */
279 #define XPS_TIMEOUT_MS 500LL
281 /* Contained by struct dp_netdev_port's 'rxqs' member. */
282 struct dp_netdev_rxq
{
283 struct netdev_rxq
*rxq
;
284 unsigned core_id
; /* Сore to which this queue is pinned. */
287 /* A port in a netdev-based datapath. */
288 struct dp_netdev_port
{
290 struct netdev
*netdev
;
291 struct hmap_node node
; /* Node in dp_netdev's 'ports'. */
292 struct netdev_saved_flags
*sf
;
293 struct dp_netdev_rxq
*rxqs
;
294 unsigned n_rxq
; /* Number of elements in 'rxq' */
295 bool dynamic_txqs
; /* If true XPS will be used. */
296 unsigned *txq_used
; /* Number of threads that uses each tx queue. */
297 struct ovs_mutex txq_used_mutex
;
298 char *type
; /* Port type as requested by user. */
299 char *rxq_affinity_list
; /* Requested affinity of rx queues. */
302 /* Contained by struct dp_netdev_flow's 'stats' member. */
303 struct dp_netdev_flow_stats
{
304 atomic_llong used
; /* Last used time, in monotonic msecs. */
305 atomic_ullong packet_count
; /* Number of packets matched. */
306 atomic_ullong byte_count
; /* Number of bytes matched. */
307 atomic_uint16_t tcp_flags
; /* Bitwise-OR of seen tcp_flags values. */
310 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
316 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
317 * its pmd thread's classifier. The text below calls this classifier 'cls'.
322 * The thread safety rules described here for "struct dp_netdev_flow" are
323 * motivated by two goals:
325 * - Prevent threads that read members of "struct dp_netdev_flow" from
326 * reading bad data due to changes by some thread concurrently modifying
329 * - Prevent two threads making changes to members of a given "struct
330 * dp_netdev_flow" from interfering with each other.
336 * A flow 'flow' may be accessed without a risk of being freed during an RCU
337 * grace period. Code that needs to hold onto a flow for a while
338 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
340 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
341 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
344 * Some members, marked 'const', are immutable. Accessing other members
345 * requires synchronization, as noted in more detail below.
347 struct dp_netdev_flow
{
348 const struct flow flow
; /* Unmasked flow that created this entry. */
349 /* Hash table index by unmasked flow. */
350 const struct cmap_node node
; /* In owning dp_netdev_pmd_thread's */
352 const ovs_u128 ufid
; /* Unique flow identifier. */
353 const unsigned pmd_id
; /* The 'core_id' of pmd thread owning this */
356 /* Number of references.
357 * The classifier owns one reference.
358 * Any thread trying to keep a rule from being freed should hold its own
360 struct ovs_refcount ref_cnt
;
365 struct dp_netdev_flow_stats stats
;
368 OVSRCU_TYPE(struct dp_netdev_actions
*) actions
;
370 /* While processing a group of input packets, the datapath uses the next
371 * member to store a pointer to the output batch for the flow. It is
372 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
373 * packet_batch_per_flow_init() and packet_batch_per_flow_execute()). */
374 struct packet_batch_per_flow
*batch
;
376 /* Packet classification. */
377 struct dpcls_rule cr
; /* In owning dp_netdev's 'cls'. */
378 /* 'cr' must be the last member. */
381 static void dp_netdev_flow_unref(struct dp_netdev_flow
*);
382 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*);
383 static int dpif_netdev_flow_from_nlattrs(const struct nlattr
*, uint32_t,
386 /* A set of datapath actions within a "struct dp_netdev_flow".
392 * A struct dp_netdev_actions 'actions' is protected with RCU. */
393 struct dp_netdev_actions
{
394 /* These members are immutable: they do not change during the struct's
396 unsigned int size
; /* Size of 'actions', in bytes. */
397 struct nlattr actions
[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
400 struct dp_netdev_actions
*dp_netdev_actions_create(const struct nlattr
*,
402 struct dp_netdev_actions
*dp_netdev_flow_get_actions(
403 const struct dp_netdev_flow
*);
404 static void dp_netdev_actions_free(struct dp_netdev_actions
*);
406 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
407 struct dp_netdev_pmd_stats
{
408 /* Indexed by DP_STAT_*. */
409 atomic_ullong n
[DP_N_STATS
];
412 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
413 struct dp_netdev_pmd_cycles
{
414 /* Indexed by PMD_CYCLES_*. */
415 atomic_ullong n
[PMD_N_CYCLES
];
418 /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */
420 struct dp_netdev_port
*port
;
421 struct netdev_rxq
*rx
;
422 struct ovs_list node
;
425 /* Contained by struct dp_netdev_pmd_thread's 'port_cache' or 'tx_ports'. */
427 struct dp_netdev_port
*port
;
430 struct hmap_node node
;
433 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
434 * the performance overhead of interrupt processing. Therefore netdev can
435 * not implement rx-wait for these devices. dpif-netdev needs to poll
436 * these device to check for recv buffer. pmd-thread does polling for
437 * devices assigned to itself.
439 * DPDK used PMD for accessing NIC.
441 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
442 * I/O of all non-pmd threads. There will be no actual thread created
445 * Each struct has its own flow table and classifier. Packets received
446 * from managed ports are looked up in the corresponding pmd thread's
447 * flow table, and are executed with the found actions.
449 struct dp_netdev_pmd_thread
{
450 struct dp_netdev
*dp
;
451 struct ovs_refcount ref_cnt
; /* Every reference must be refcount'ed. */
452 struct cmap_node node
; /* In 'dp->poll_threads'. */
454 pthread_cond_t cond
; /* For synchronizing pmd thread reload. */
455 struct ovs_mutex cond_mutex
; /* Mutex for condition variable. */
457 /* Per thread exact-match cache. Note, the instance for cpu core
458 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
459 * need to be protected by 'non_pmd_mutex'. Every other instance
460 * will only be accessed by its own pmd thread. */
461 struct emc_cache flow_cache
;
463 /* Flow-Table and classifiers
465 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
466 * changes to 'classifiers' must be made while still holding the
469 struct ovs_mutex flow_mutex
;
470 struct cmap flow_table OVS_GUARDED
; /* Flow table. */
472 /* One classifier per in_port polled by the pmd */
473 struct cmap classifiers
;
474 /* Periodically sort subtable vectors according to hit frequencies */
475 long long int next_optimization
;
478 struct dp_netdev_pmd_stats stats
;
480 /* Cycles counters */
481 struct dp_netdev_pmd_cycles cycles
;
483 /* Used to count cicles. See 'cycles_counter_end()' */
484 unsigned long long last_cycles
;
486 struct latch exit_latch
; /* For terminating the pmd thread. */
487 atomic_uint change_seq
; /* For reloading pmd ports. */
489 unsigned core_id
; /* CPU core id of this pmd thread. */
490 int numa_id
; /* numa node id of this pmd thread. */
493 /* Queue id used by this pmd thread to send packets on all netdevs if
494 * XPS disabled for this netdev. All static_tx_qid's are unique and less
495 * than 'ovs_numa_get_n_cores() + 1'. */
496 atomic_int static_tx_qid
;
498 struct ovs_mutex port_mutex
; /* Mutex for 'poll_list' and 'tx_ports'. */
499 /* List of rx queues to poll. */
500 struct ovs_list poll_list OVS_GUARDED
;
501 /* Number of elements in 'poll_list' */
503 /* Map of 'tx_port's used for transmission. Written by the main thread,
504 * read by the pmd thread. */
505 struct hmap tx_ports OVS_GUARDED
;
507 /* Map of 'tx_port' used in the fast path. This is a thread-local copy of
508 * 'tx_ports'. The instance for cpu core NON_PMD_CORE_ID can be accessed
509 * by multiple threads, and thusly need to be protected by 'non_pmd_mutex'.
510 * Every other instance will only be accessed by its own pmd thread. */
511 struct hmap port_cache
;
513 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
514 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
515 * values and subtracts them from 'stats' and 'cycles' before
516 * reporting to the user */
517 unsigned long long stats_zero
[DP_N_STATS
];
518 uint64_t cycles_zero
[PMD_N_CYCLES
];
521 #define PMD_INITIAL_SEQ 1
523 /* Interface to netdev-based datapath. */
526 struct dp_netdev
*dp
;
527 uint64_t last_port_seq
;
530 static int get_port_by_number(struct dp_netdev
*dp
, odp_port_t port_no
,
531 struct dp_netdev_port
**portp
)
532 OVS_REQUIRES(dp
->port_mutex
);
533 static int get_port_by_name(struct dp_netdev
*dp
, const char *devname
,
534 struct dp_netdev_port
**portp
)
535 OVS_REQUIRES(dp
->port_mutex
);
536 static void dp_netdev_free(struct dp_netdev
*)
537 OVS_REQUIRES(dp_netdev_mutex
);
538 static int do_add_port(struct dp_netdev
*dp
, const char *devname
,
539 const char *type
, odp_port_t port_no
)
540 OVS_REQUIRES(dp
->port_mutex
);
541 static void do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*)
542 OVS_REQUIRES(dp
->port_mutex
);
543 static int dpif_netdev_open(const struct dpif_class
*, const char *name
,
544 bool create
, struct dpif
**);
545 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
546 struct dp_packet_batch
*,
547 bool may_steal
, const struct flow
*flow
,
548 const struct nlattr
*actions
,
551 static void dp_netdev_input(struct dp_netdev_pmd_thread
*,
552 struct dp_packet_batch
*, odp_port_t port_no
);
553 static void dp_netdev_recirculate(struct dp_netdev_pmd_thread
*,
554 struct dp_packet_batch
*);
556 static void dp_netdev_disable_upcall(struct dp_netdev
*);
557 static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
);
558 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
,
559 struct dp_netdev
*dp
, unsigned core_id
,
561 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
);
562 static void dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
563 OVS_REQUIRES(dp
->port_mutex
);
565 static struct dp_netdev_pmd_thread
*dp_netdev_get_pmd(struct dp_netdev
*dp
,
567 static struct dp_netdev_pmd_thread
*
568 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
);
569 static void dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
);
570 static void dp_netdev_del_pmds_on_numa(struct dp_netdev
*dp
, int numa_id
);
571 static void dp_netdev_set_pmds_on_numa(struct dp_netdev
*dp
, int numa_id
)
572 OVS_REQUIRES(dp
->port_mutex
);
573 static void dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
);
574 static void dp_netdev_del_port_from_all_pmds(struct dp_netdev
*dp
,
575 struct dp_netdev_port
*port
);
576 static void dp_netdev_add_port_to_pmds(struct dp_netdev
*dp
,
577 struct dp_netdev_port
*port
);
578 static void dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
579 struct dp_netdev_port
*port
);
580 static void dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
581 struct dp_netdev_port
*port
,
582 struct netdev_rxq
*rx
);
583 static struct dp_netdev_pmd_thread
*
584 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev
*dp
, int numa_id
);
585 static void dp_netdev_reset_pmd_threads(struct dp_netdev
*dp
)
586 OVS_REQUIRES(dp
->port_mutex
);
587 static void reconfigure_pmd_threads(struct dp_netdev
*dp
)
588 OVS_REQUIRES(dp
->port_mutex
);
589 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
);
590 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
);
591 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
);
592 static void pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
593 OVS_REQUIRES(pmd
->port_mutex
);
595 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
);
598 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
599 long long now
, bool purge
);
600 static int dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
601 struct tx_port
*tx
, long long now
);
603 static inline bool emc_entry_alive(struct emc_entry
*ce
);
604 static void emc_clear_entry(struct emc_entry
*ce
);
607 emc_cache_init(struct emc_cache
*flow_cache
)
611 flow_cache
->sweep_idx
= 0;
612 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
613 flow_cache
->entries
[i
].flow
= NULL
;
614 flow_cache
->entries
[i
].key
.hash
= 0;
615 flow_cache
->entries
[i
].key
.len
= sizeof(struct miniflow
);
616 flowmap_init(&flow_cache
->entries
[i
].key
.mf
.map
);
621 emc_cache_uninit(struct emc_cache
*flow_cache
)
625 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
626 emc_clear_entry(&flow_cache
->entries
[i
]);
630 /* Check and clear dead flow references slowly (one entry at each
633 emc_cache_slow_sweep(struct emc_cache
*flow_cache
)
635 struct emc_entry
*entry
= &flow_cache
->entries
[flow_cache
->sweep_idx
];
637 if (!emc_entry_alive(entry
)) {
638 emc_clear_entry(entry
);
640 flow_cache
->sweep_idx
= (flow_cache
->sweep_idx
+ 1) & EM_FLOW_HASH_MASK
;
643 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
645 dpif_is_netdev(const struct dpif
*dpif
)
647 return dpif
->dpif_class
->open
== dpif_netdev_open
;
650 static struct dpif_netdev
*
651 dpif_netdev_cast(const struct dpif
*dpif
)
653 ovs_assert(dpif_is_netdev(dpif
));
654 return CONTAINER_OF(dpif
, struct dpif_netdev
, dpif
);
657 static struct dp_netdev
*
658 get_dp_netdev(const struct dpif
*dpif
)
660 return dpif_netdev_cast(dpif
)->dp
;
664 PMD_INFO_SHOW_STATS
, /* Show how cpu cycles are spent. */
665 PMD_INFO_CLEAR_STATS
, /* Set the cycles count to 0. */
666 PMD_INFO_SHOW_RXQ
/* Show poll-lists of pmd threads. */
670 pmd_info_show_stats(struct ds
*reply
,
671 struct dp_netdev_pmd_thread
*pmd
,
672 unsigned long long stats
[DP_N_STATS
],
673 uint64_t cycles
[PMD_N_CYCLES
])
675 unsigned long long total_packets
= 0;
676 uint64_t total_cycles
= 0;
679 /* These loops subtracts reference values ('*_zero') from the counters.
680 * Since loads and stores are relaxed, it might be possible for a '*_zero'
681 * value to be more recent than the current value we're reading from the
682 * counter. This is not a big problem, since these numbers are not
683 * supposed to be too accurate, but we should at least make sure that
684 * the result is not negative. */
685 for (i
= 0; i
< DP_N_STATS
; i
++) {
686 if (stats
[i
] > pmd
->stats_zero
[i
]) {
687 stats
[i
] -= pmd
->stats_zero
[i
];
692 if (i
!= DP_STAT_LOST
) {
693 /* Lost packets are already included in DP_STAT_MISS */
694 total_packets
+= stats
[i
];
698 for (i
= 0; i
< PMD_N_CYCLES
; i
++) {
699 if (cycles
[i
] > pmd
->cycles_zero
[i
]) {
700 cycles
[i
] -= pmd
->cycles_zero
[i
];
705 total_cycles
+= cycles
[i
];
708 ds_put_cstr(reply
, (pmd
->core_id
== NON_PMD_CORE_ID
)
709 ? "main thread" : "pmd thread");
711 if (pmd
->numa_id
!= OVS_NUMA_UNSPEC
) {
712 ds_put_format(reply
, " numa_id %d", pmd
->numa_id
);
714 if (pmd
->core_id
!= OVS_CORE_UNSPEC
&& pmd
->core_id
!= NON_PMD_CORE_ID
) {
715 ds_put_format(reply
, " core_id %u", pmd
->core_id
);
717 ds_put_cstr(reply
, ":\n");
720 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
721 "\tavg. subtable lookups per hit:%.2f\n"
722 "\tmiss:%llu\n\tlost:%llu\n",
723 stats
[DP_STAT_EXACT_HIT
], stats
[DP_STAT_MASKED_HIT
],
724 stats
[DP_STAT_MASKED_HIT
] > 0
725 ? (1.0*stats
[DP_STAT_LOOKUP_HIT
])/stats
[DP_STAT_MASKED_HIT
]
727 stats
[DP_STAT_MISS
], stats
[DP_STAT_LOST
]);
729 if (total_cycles
== 0) {
734 "\tpolling cycles:%"PRIu64
" (%.02f%%)\n"
735 "\tprocessing cycles:%"PRIu64
" (%.02f%%)\n",
736 cycles
[PMD_CYCLES_POLLING
],
737 cycles
[PMD_CYCLES_POLLING
] / (double)total_cycles
* 100,
738 cycles
[PMD_CYCLES_PROCESSING
],
739 cycles
[PMD_CYCLES_PROCESSING
] / (double)total_cycles
* 100);
741 if (total_packets
== 0) {
746 "\tavg cycles per packet: %.02f (%"PRIu64
"/%llu)\n",
747 total_cycles
/ (double)total_packets
,
748 total_cycles
, total_packets
);
751 "\tavg processing cycles per packet: "
752 "%.02f (%"PRIu64
"/%llu)\n",
753 cycles
[PMD_CYCLES_PROCESSING
] / (double)total_packets
,
754 cycles
[PMD_CYCLES_PROCESSING
], total_packets
);
758 pmd_info_clear_stats(struct ds
*reply OVS_UNUSED
,
759 struct dp_netdev_pmd_thread
*pmd
,
760 unsigned long long stats
[DP_N_STATS
],
761 uint64_t cycles
[PMD_N_CYCLES
])
765 /* We cannot write 'stats' and 'cycles' (because they're written by other
766 * threads) and we shouldn't change 'stats' (because they're used to count
767 * datapath stats, which must not be cleared here). Instead, we save the
768 * current values and subtract them from the values to be displayed in the
770 for (i
= 0; i
< DP_N_STATS
; i
++) {
771 pmd
->stats_zero
[i
] = stats
[i
];
773 for (i
= 0; i
< PMD_N_CYCLES
; i
++) {
774 pmd
->cycles_zero
[i
] = cycles
[i
];
779 pmd_info_show_rxq(struct ds
*reply
, struct dp_netdev_pmd_thread
*pmd
)
781 if (pmd
->core_id
!= NON_PMD_CORE_ID
) {
782 struct rxq_poll
*poll
;
783 const char *prev_name
= NULL
;
786 "pmd thread numa_id %d core_id %u:\n\tisolated : %s\n",
787 pmd
->numa_id
, pmd
->core_id
, (pmd
->isolated
)
790 ovs_mutex_lock(&pmd
->port_mutex
);
791 LIST_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
792 const char *name
= netdev_get_name(poll
->port
->netdev
);
794 if (!prev_name
|| strcmp(name
, prev_name
)) {
796 ds_put_cstr(reply
, "\n");
798 ds_put_format(reply
, "\tport: %s\tqueue-id:",
799 netdev_get_name(poll
->port
->netdev
));
801 ds_put_format(reply
, " %d", netdev_rxq_get_queue_id(poll
->rx
));
804 ovs_mutex_unlock(&pmd
->port_mutex
);
805 ds_put_cstr(reply
, "\n");
810 dpif_netdev_pmd_info(struct unixctl_conn
*conn
, int argc
, const char *argv
[],
813 struct ds reply
= DS_EMPTY_INITIALIZER
;
814 struct dp_netdev_pmd_thread
*pmd
;
815 struct dp_netdev
*dp
= NULL
;
816 enum pmd_info_type type
= *(enum pmd_info_type
*) aux
;
818 ovs_mutex_lock(&dp_netdev_mutex
);
821 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
822 } else if (shash_count(&dp_netdevs
) == 1) {
823 /* There's only one datapath */
824 dp
= shash_first(&dp_netdevs
)->data
;
828 ovs_mutex_unlock(&dp_netdev_mutex
);
829 unixctl_command_reply_error(conn
,
830 "please specify an existing datapath");
834 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
835 if (type
== PMD_INFO_SHOW_RXQ
) {
836 pmd_info_show_rxq(&reply
, pmd
);
838 unsigned long long stats
[DP_N_STATS
];
839 uint64_t cycles
[PMD_N_CYCLES
];
842 /* Read current stats and cycle counters */
843 for (i
= 0; i
< ARRAY_SIZE(stats
); i
++) {
844 atomic_read_relaxed(&pmd
->stats
.n
[i
], &stats
[i
]);
846 for (i
= 0; i
< ARRAY_SIZE(cycles
); i
++) {
847 atomic_read_relaxed(&pmd
->cycles
.n
[i
], &cycles
[i
]);
850 if (type
== PMD_INFO_CLEAR_STATS
) {
851 pmd_info_clear_stats(&reply
, pmd
, stats
, cycles
);
852 } else if (type
== PMD_INFO_SHOW_STATS
) {
853 pmd_info_show_stats(&reply
, pmd
, stats
, cycles
);
858 ovs_mutex_unlock(&dp_netdev_mutex
);
860 unixctl_command_reply(conn
, ds_cstr(&reply
));
865 dpif_netdev_init(void)
867 static enum pmd_info_type show_aux
= PMD_INFO_SHOW_STATS
,
868 clear_aux
= PMD_INFO_CLEAR_STATS
,
869 poll_aux
= PMD_INFO_SHOW_RXQ
;
871 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
872 0, 1, dpif_netdev_pmd_info
,
874 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
875 0, 1, dpif_netdev_pmd_info
,
877 unixctl_command_register("dpif-netdev/pmd-rxq-show", "[dp]",
878 0, 1, dpif_netdev_pmd_info
,
884 dpif_netdev_enumerate(struct sset
*all_dps
,
885 const struct dpif_class
*dpif_class
)
887 struct shash_node
*node
;
889 ovs_mutex_lock(&dp_netdev_mutex
);
890 SHASH_FOR_EACH(node
, &dp_netdevs
) {
891 struct dp_netdev
*dp
= node
->data
;
892 if (dpif_class
!= dp
->class) {
893 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
894 * If the class doesn't match, skip this dpif. */
897 sset_add(all_dps
, node
->name
);
899 ovs_mutex_unlock(&dp_netdev_mutex
);
905 dpif_netdev_class_is_dummy(const struct dpif_class
*class)
907 return class != &dpif_netdev_class
;
911 dpif_netdev_port_open_type(const struct dpif_class
*class, const char *type
)
913 return strcmp(type
, "internal") ? type
914 : dpif_netdev_class_is_dummy(class) ? "dummy-internal"
919 create_dpif_netdev(struct dp_netdev
*dp
)
921 uint16_t netflow_id
= hash_string(dp
->name
, 0);
922 struct dpif_netdev
*dpif
;
924 ovs_refcount_ref(&dp
->ref_cnt
);
926 dpif
= xmalloc(sizeof *dpif
);
927 dpif_init(&dpif
->dpif
, dp
->class, dp
->name
, netflow_id
>> 8, netflow_id
);
929 dpif
->last_port_seq
= seq_read(dp
->port_seq
);
934 /* Choose an unused, non-zero port number and return it on success.
935 * Return ODPP_NONE on failure. */
937 choose_port(struct dp_netdev
*dp
, const char *name
)
938 OVS_REQUIRES(dp
->port_mutex
)
942 if (dp
->class != &dpif_netdev_class
) {
946 /* If the port name begins with "br", start the number search at
947 * 100 to make writing tests easier. */
948 if (!strncmp(name
, "br", 2)) {
952 /* If the port name contains a number, try to assign that port number.
953 * This can make writing unit tests easier because port numbers are
955 for (p
= name
; *p
!= '\0'; p
++) {
956 if (isdigit((unsigned char) *p
)) {
957 port_no
= start_no
+ strtol(p
, NULL
, 10);
958 if (port_no
> 0 && port_no
!= odp_to_u32(ODPP_NONE
)
959 && !dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
960 return u32_to_odp(port_no
);
967 for (port_no
= 1; port_no
<= UINT16_MAX
; port_no
++) {
968 if (!dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
969 return u32_to_odp(port_no
);
977 create_dp_netdev(const char *name
, const struct dpif_class
*class,
978 struct dp_netdev
**dpp
)
979 OVS_REQUIRES(dp_netdev_mutex
)
981 struct dp_netdev
*dp
;
984 dp
= xzalloc(sizeof *dp
);
985 shash_add(&dp_netdevs
, name
, dp
);
987 *CONST_CAST(const struct dpif_class
**, &dp
->class) = class;
988 *CONST_CAST(const char **, &dp
->name
) = xstrdup(name
);
989 ovs_refcount_init(&dp
->ref_cnt
);
990 atomic_flag_clear(&dp
->destroyed
);
992 ovs_mutex_init(&dp
->port_mutex
);
993 hmap_init(&dp
->ports
);
994 dp
->port_seq
= seq_create();
995 fat_rwlock_init(&dp
->upcall_rwlock
);
997 dp
->reconfigure_seq
= seq_create();
998 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
1000 /* Disable upcalls by default. */
1001 dp_netdev_disable_upcall(dp
);
1002 dp
->upcall_aux
= NULL
;
1003 dp
->upcall_cb
= NULL
;
1005 conntrack_init(&dp
->conntrack
);
1007 cmap_init(&dp
->poll_threads
);
1008 ovs_mutex_init_recursive(&dp
->non_pmd_mutex
);
1009 ovsthread_key_create(&dp
->per_pmd_key
, NULL
);
1011 ovs_mutex_lock(&dp
->port_mutex
);
1012 dp_netdev_set_nonpmd(dp
);
1014 error
= do_add_port(dp
, name
, dpif_netdev_port_open_type(dp
->class,
1017 ovs_mutex_unlock(&dp
->port_mutex
);
1023 dp
->last_tnl_conf_seq
= seq_read(tnl_conf_seq
);
1029 dp_netdev_request_reconfigure(struct dp_netdev
*dp
)
1031 seq_change(dp
->reconfigure_seq
);
1035 dp_netdev_is_reconf_required(struct dp_netdev
*dp
)
1037 return seq_read(dp
->reconfigure_seq
) != dp
->last_reconfigure_seq
;
1041 dpif_netdev_open(const struct dpif_class
*class, const char *name
,
1042 bool create
, struct dpif
**dpifp
)
1044 struct dp_netdev
*dp
;
1047 ovs_mutex_lock(&dp_netdev_mutex
);
1048 dp
= shash_find_data(&dp_netdevs
, name
);
1050 error
= create
? create_dp_netdev(name
, class, &dp
) : ENODEV
;
1052 error
= (dp
->class != class ? EINVAL
1057 *dpifp
= create_dpif_netdev(dp
);
1060 ovs_mutex_unlock(&dp_netdev_mutex
);
1066 dp_netdev_destroy_upcall_lock(struct dp_netdev
*dp
)
1067 OVS_NO_THREAD_SAFETY_ANALYSIS
1069 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
1070 ovs_assert(fat_rwlock_tryrdlock(&dp
->upcall_rwlock
));
1072 /* Before freeing a lock we should release it */
1073 fat_rwlock_unlock(&dp
->upcall_rwlock
);
1074 fat_rwlock_destroy(&dp
->upcall_rwlock
);
1077 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
1078 * through the 'dp_netdevs' shash while freeing 'dp'. */
1080 dp_netdev_free(struct dp_netdev
*dp
)
1081 OVS_REQUIRES(dp_netdev_mutex
)
1083 struct dp_netdev_port
*port
, *next
;
1085 shash_find_and_delete(&dp_netdevs
, dp
->name
);
1087 dp_netdev_destroy_all_pmds(dp
);
1088 ovs_mutex_destroy(&dp
->non_pmd_mutex
);
1089 ovsthread_key_delete(dp
->per_pmd_key
);
1091 conntrack_destroy(&dp
->conntrack
);
1093 ovs_mutex_lock(&dp
->port_mutex
);
1094 HMAP_FOR_EACH_SAFE (port
, next
, node
, &dp
->ports
) {
1095 do_del_port(dp
, port
);
1097 ovs_mutex_unlock(&dp
->port_mutex
);
1098 cmap_destroy(&dp
->poll_threads
);
1100 seq_destroy(dp
->reconfigure_seq
);
1102 seq_destroy(dp
->port_seq
);
1103 hmap_destroy(&dp
->ports
);
1104 ovs_mutex_destroy(&dp
->port_mutex
);
1106 /* Upcalls must be disabled at this point */
1107 dp_netdev_destroy_upcall_lock(dp
);
1109 free(dp
->pmd_cmask
);
1110 free(CONST_CAST(char *, dp
->name
));
1115 dp_netdev_unref(struct dp_netdev
*dp
)
1118 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
1119 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
1120 ovs_mutex_lock(&dp_netdev_mutex
);
1121 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1124 ovs_mutex_unlock(&dp_netdev_mutex
);
1129 dpif_netdev_close(struct dpif
*dpif
)
1131 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1133 dp_netdev_unref(dp
);
1138 dpif_netdev_destroy(struct dpif
*dpif
)
1140 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1142 if (!atomic_flag_test_and_set(&dp
->destroyed
)) {
1143 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1144 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
1152 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
1153 * load/store semantics. While the increment is not atomic, the load and
1154 * store operations are, making it impossible to read inconsistent values.
1156 * This is used to update thread local stats counters. */
1158 non_atomic_ullong_add(atomic_ullong
*var
, unsigned long long n
)
1160 unsigned long long tmp
;
1162 atomic_read_relaxed(var
, &tmp
);
1164 atomic_store_relaxed(var
, tmp
);
1168 dpif_netdev_get_stats(const struct dpif
*dpif
, struct dpif_dp_stats
*stats
)
1170 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1171 struct dp_netdev_pmd_thread
*pmd
;
1173 stats
->n_flows
= stats
->n_hit
= stats
->n_missed
= stats
->n_lost
= 0;
1174 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1175 unsigned long long n
;
1176 stats
->n_flows
+= cmap_count(&pmd
->flow_table
);
1178 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_MASKED_HIT
], &n
);
1180 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_EXACT_HIT
], &n
);
1182 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_MISS
], &n
);
1183 stats
->n_missed
+= n
;
1184 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_LOST
], &n
);
1187 stats
->n_masks
= UINT32_MAX
;
1188 stats
->n_mask_hit
= UINT64_MAX
;
1194 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread
*pmd
)
1198 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
1199 ovs_mutex_lock(&pmd
->dp
->non_pmd_mutex
);
1200 ovs_mutex_lock(&pmd
->port_mutex
);
1201 pmd_load_cached_ports(pmd
);
1202 ovs_mutex_unlock(&pmd
->port_mutex
);
1203 ovs_mutex_unlock(&pmd
->dp
->non_pmd_mutex
);
1207 ovs_mutex_lock(&pmd
->cond_mutex
);
1208 atomic_add_relaxed(&pmd
->change_seq
, 1, &old_seq
);
1209 ovs_mutex_cond_wait(&pmd
->cond
, &pmd
->cond_mutex
);
1210 ovs_mutex_unlock(&pmd
->cond_mutex
);
1214 hash_port_no(odp_port_t port_no
)
1216 return hash_int(odp_to_u32(port_no
), 0);
1220 port_create(const char *devname
, const char *type
,
1221 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1223 struct netdev_saved_flags
*sf
;
1224 struct dp_netdev_port
*port
;
1225 enum netdev_flags flags
;
1226 struct netdev
*netdev
;
1227 int n_open_rxqs
= 0;
1230 bool dynamic_txqs
= false;
1234 /* Open and validate network device. */
1235 error
= netdev_open(devname
, type
, &netdev
);
1239 /* XXX reject non-Ethernet devices */
1241 netdev_get_flags(netdev
, &flags
);
1242 if (flags
& NETDEV_LOOPBACK
) {
1243 VLOG_ERR("%s: cannot add a loopback device", devname
);
1248 if (netdev_is_pmd(netdev
)) {
1249 n_cores
= ovs_numa_get_n_cores();
1251 if (n_cores
== OVS_CORE_UNSPEC
) {
1252 VLOG_ERR("%s, cannot get cpu core info", devname
);
1256 /* There can only be ovs_numa_get_n_cores() pmd threads,
1257 * so creates a txq for each, and one extra for the non
1259 error
= netdev_set_tx_multiq(netdev
, n_cores
+ 1);
1260 if (error
&& (error
!= EOPNOTSUPP
)) {
1261 VLOG_ERR("%s, cannot set multiq", devname
);
1266 if (netdev_is_reconf_required(netdev
)) {
1267 error
= netdev_reconfigure(netdev
);
1273 if (netdev_is_pmd(netdev
)) {
1274 if (netdev_n_txq(netdev
) < n_cores
+ 1) {
1275 dynamic_txqs
= true;
1279 port
= xzalloc(sizeof *port
);
1280 port
->port_no
= port_no
;
1281 port
->netdev
= netdev
;
1282 port
->n_rxq
= netdev_n_rxq(netdev
);
1283 port
->rxqs
= xcalloc(port
->n_rxq
, sizeof *port
->rxqs
);
1284 port
->txq_used
= xcalloc(netdev_n_txq(netdev
), sizeof *port
->txq_used
);
1285 port
->type
= xstrdup(type
);
1286 ovs_mutex_init(&port
->txq_used_mutex
);
1287 port
->dynamic_txqs
= dynamic_txqs
;
1289 for (i
= 0; i
< port
->n_rxq
; i
++) {
1290 error
= netdev_rxq_open(netdev
, &port
->rxqs
[i
].rxq
, i
);
1292 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1293 devname
, ovs_strerror(errno
));
1296 port
->rxqs
[i
].core_id
= -1;
1300 error
= netdev_turn_flags_on(netdev
, NETDEV_PROMISC
, &sf
);
1311 for (i
= 0; i
< n_open_rxqs
; i
++) {
1312 netdev_rxq_close(port
->rxqs
[i
].rxq
);
1314 ovs_mutex_destroy(&port
->txq_used_mutex
);
1316 free(port
->txq_used
);
1321 netdev_close(netdev
);
1326 do_add_port(struct dp_netdev
*dp
, const char *devname
, const char *type
,
1328 OVS_REQUIRES(dp
->port_mutex
)
1330 struct dp_netdev_port
*port
;
1333 /* Reject devices already in 'dp'. */
1334 if (!get_port_by_name(dp
, devname
, &port
)) {
1338 error
= port_create(devname
, type
, port_no
, &port
);
1343 if (netdev_is_pmd(port
->netdev
)) {
1344 int numa_id
= netdev_get_numa_id(port
->netdev
);
1346 ovs_assert(ovs_numa_numa_id_is_valid(numa_id
));
1347 dp_netdev_set_pmds_on_numa(dp
, numa_id
);
1350 dp_netdev_add_port_to_pmds(dp
, port
);
1352 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
1353 seq_change(dp
->port_seq
);
1359 dpif_netdev_port_add(struct dpif
*dpif
, struct netdev
*netdev
,
1360 odp_port_t
*port_nop
)
1362 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1363 char namebuf
[NETDEV_VPORT_NAME_BUFSIZE
];
1364 const char *dpif_port
;
1368 ovs_mutex_lock(&dp
->port_mutex
);
1369 dpif_port
= netdev_vport_get_dpif_port(netdev
, namebuf
, sizeof namebuf
);
1370 if (*port_nop
!= ODPP_NONE
) {
1371 port_no
= *port_nop
;
1372 error
= dp_netdev_lookup_port(dp
, *port_nop
) ? EBUSY
: 0;
1374 port_no
= choose_port(dp
, dpif_port
);
1375 error
= port_no
== ODPP_NONE
? EFBIG
: 0;
1378 *port_nop
= port_no
;
1379 error
= do_add_port(dp
, dpif_port
, netdev_get_type(netdev
), port_no
);
1381 ovs_mutex_unlock(&dp
->port_mutex
);
1387 dpif_netdev_port_del(struct dpif
*dpif
, odp_port_t port_no
)
1389 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1392 ovs_mutex_lock(&dp
->port_mutex
);
1393 if (port_no
== ODPP_LOCAL
) {
1396 struct dp_netdev_port
*port
;
1398 error
= get_port_by_number(dp
, port_no
, &port
);
1400 do_del_port(dp
, port
);
1403 ovs_mutex_unlock(&dp
->port_mutex
);
1409 is_valid_port_number(odp_port_t port_no
)
1411 return port_no
!= ODPP_NONE
;
1414 static struct dp_netdev_port
*
1415 dp_netdev_lookup_port(const struct dp_netdev
*dp
, odp_port_t port_no
)
1416 OVS_REQUIRES(dp
->port_mutex
)
1418 struct dp_netdev_port
*port
;
1420 HMAP_FOR_EACH_WITH_HASH (port
, node
, hash_port_no(port_no
), &dp
->ports
) {
1421 if (port
->port_no
== port_no
) {
1429 get_port_by_number(struct dp_netdev
*dp
,
1430 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1431 OVS_REQUIRES(dp
->port_mutex
)
1433 if (!is_valid_port_number(port_no
)) {
1437 *portp
= dp_netdev_lookup_port(dp
, port_no
);
1438 return *portp
? 0 : ENODEV
;
1443 port_destroy(struct dp_netdev_port
*port
)
1449 netdev_close(port
->netdev
);
1450 netdev_restore_flags(port
->sf
);
1452 for (unsigned i
= 0; i
< port
->n_rxq
; i
++) {
1453 netdev_rxq_close(port
->rxqs
[i
].rxq
);
1455 ovs_mutex_destroy(&port
->txq_used_mutex
);
1456 free(port
->rxq_affinity_list
);
1457 free(port
->txq_used
);
1464 get_port_by_name(struct dp_netdev
*dp
,
1465 const char *devname
, struct dp_netdev_port
**portp
)
1466 OVS_REQUIRES(dp
->port_mutex
)
1468 struct dp_netdev_port
*port
;
1470 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1471 if (!strcmp(netdev_get_name(port
->netdev
), devname
)) {
1477 /* Callers of dpif_netdev_port_query_by_name() expect ENODEV for a non
1483 get_n_pmd_threads(struct dp_netdev
*dp
)
1485 /* There is one non pmd thread in dp->poll_threads */
1486 return cmap_count(&dp
->poll_threads
) - 1;
1490 get_n_pmd_threads_on_numa(struct dp_netdev
*dp
, int numa_id
)
1492 struct dp_netdev_pmd_thread
*pmd
;
1495 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1496 if (pmd
->numa_id
== numa_id
) {
1504 /* Returns 'true' if there is a port with pmd netdev and the netdev is on
1505 * numa node 'numa_id' or its rx queue assigned to core on that numa node . */
1507 has_pmd_rxq_for_numa(struct dp_netdev
*dp
, int numa_id
)
1508 OVS_REQUIRES(dp
->port_mutex
)
1510 struct dp_netdev_port
*port
;
1512 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1513 if (netdev_is_pmd(port
->netdev
)) {
1516 if (netdev_get_numa_id(port
->netdev
) == numa_id
) {
1520 for (i
= 0; i
< port
->n_rxq
; i
++) {
1521 unsigned core_id
= port
->rxqs
[i
].core_id
;
1524 && ovs_numa_get_numa_id(core_id
) == numa_id
) {
1536 do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*port
)
1537 OVS_REQUIRES(dp
->port_mutex
)
1539 hmap_remove(&dp
->ports
, &port
->node
);
1540 seq_change(dp
->port_seq
);
1542 dp_netdev_del_port_from_all_pmds(dp
, port
);
1544 if (netdev_is_pmd(port
->netdev
)) {
1545 int numa_id
= netdev_get_numa_id(port
->netdev
);
1547 /* PMD threads can not be on invalid numa node. */
1548 ovs_assert(ovs_numa_numa_id_is_valid(numa_id
));
1549 /* If there is no netdev on the numa node, deletes the pmd threads
1551 if (!has_pmd_rxq_for_numa(dp
, numa_id
)) {
1552 dp_netdev_del_pmds_on_numa(dp
, numa_id
);
1560 answer_port_query(const struct dp_netdev_port
*port
,
1561 struct dpif_port
*dpif_port
)
1563 dpif_port
->name
= xstrdup(netdev_get_name(port
->netdev
));
1564 dpif_port
->type
= xstrdup(port
->type
);
1565 dpif_port
->port_no
= port
->port_no
;
1569 dpif_netdev_port_query_by_number(const struct dpif
*dpif
, odp_port_t port_no
,
1570 struct dpif_port
*dpif_port
)
1572 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1573 struct dp_netdev_port
*port
;
1576 ovs_mutex_lock(&dp
->port_mutex
);
1577 error
= get_port_by_number(dp
, port_no
, &port
);
1578 if (!error
&& dpif_port
) {
1579 answer_port_query(port
, dpif_port
);
1581 ovs_mutex_unlock(&dp
->port_mutex
);
1587 dpif_netdev_port_query_by_name(const struct dpif
*dpif
, const char *devname
,
1588 struct dpif_port
*dpif_port
)
1590 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1591 struct dp_netdev_port
*port
;
1594 ovs_mutex_lock(&dp
->port_mutex
);
1595 error
= get_port_by_name(dp
, devname
, &port
);
1596 if (!error
&& dpif_port
) {
1597 answer_port_query(port
, dpif_port
);
1599 ovs_mutex_unlock(&dp
->port_mutex
);
1605 dp_netdev_flow_free(struct dp_netdev_flow
*flow
)
1607 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow
));
1611 static void dp_netdev_flow_unref(struct dp_netdev_flow
*flow
)
1613 if (ovs_refcount_unref_relaxed(&flow
->ref_cnt
) == 1) {
1614 ovsrcu_postpone(dp_netdev_flow_free
, flow
);
1619 dp_netdev_flow_hash(const ovs_u128
*ufid
)
1621 return ufid
->u32
[0];
1624 static inline struct dpcls
*
1625 dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread
*pmd
,
1629 uint32_t hash
= hash_port_no(in_port
);
1630 CMAP_FOR_EACH_WITH_HASH (cls
, node
, hash
, &pmd
->classifiers
) {
1631 if (cls
->in_port
== in_port
) {
1632 /* Port classifier exists already */
1639 static inline struct dpcls
*
1640 dp_netdev_pmd_find_dpcls(struct dp_netdev_pmd_thread
*pmd
,
1642 OVS_REQUIRES(pmd
->flow_mutex
)
1644 struct dpcls
*cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
1645 uint32_t hash
= hash_port_no(in_port
);
1648 /* Create new classifier for in_port */
1649 cls
= xmalloc(sizeof(*cls
));
1651 cls
->in_port
= in_port
;
1652 cmap_insert(&pmd
->classifiers
, &cls
->node
, hash
);
1653 VLOG_DBG("Creating dpcls %p for in_port %d", cls
, in_port
);
1659 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread
*pmd
,
1660 struct dp_netdev_flow
*flow
)
1661 OVS_REQUIRES(pmd
->flow_mutex
)
1663 struct cmap_node
*node
= CONST_CAST(struct cmap_node
*, &flow
->node
);
1665 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
1667 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
1668 ovs_assert(cls
!= NULL
);
1669 dpcls_remove(cls
, &flow
->cr
);
1670 cmap_remove(&pmd
->flow_table
, node
, dp_netdev_flow_hash(&flow
->ufid
));
1673 dp_netdev_flow_unref(flow
);
1677 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
)
1679 struct dp_netdev_flow
*netdev_flow
;
1681 ovs_mutex_lock(&pmd
->flow_mutex
);
1682 CMAP_FOR_EACH (netdev_flow
, node
, &pmd
->flow_table
) {
1683 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
1685 ovs_mutex_unlock(&pmd
->flow_mutex
);
1689 dpif_netdev_flow_flush(struct dpif
*dpif
)
1691 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1692 struct dp_netdev_pmd_thread
*pmd
;
1694 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1695 dp_netdev_pmd_flow_flush(pmd
);
1701 struct dp_netdev_port_state
{
1702 struct hmap_position position
;
1707 dpif_netdev_port_dump_start(const struct dpif
*dpif OVS_UNUSED
, void **statep
)
1709 *statep
= xzalloc(sizeof(struct dp_netdev_port_state
));
1714 dpif_netdev_port_dump_next(const struct dpif
*dpif
, void *state_
,
1715 struct dpif_port
*dpif_port
)
1717 struct dp_netdev_port_state
*state
= state_
;
1718 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1719 struct hmap_node
*node
;
1722 ovs_mutex_lock(&dp
->port_mutex
);
1723 node
= hmap_at_position(&dp
->ports
, &state
->position
);
1725 struct dp_netdev_port
*port
;
1727 port
= CONTAINER_OF(node
, struct dp_netdev_port
, node
);
1730 state
->name
= xstrdup(netdev_get_name(port
->netdev
));
1731 dpif_port
->name
= state
->name
;
1732 dpif_port
->type
= port
->type
;
1733 dpif_port
->port_no
= port
->port_no
;
1739 ovs_mutex_unlock(&dp
->port_mutex
);
1745 dpif_netdev_port_dump_done(const struct dpif
*dpif OVS_UNUSED
, void *state_
)
1747 struct dp_netdev_port_state
*state
= state_
;
1754 dpif_netdev_port_poll(const struct dpif
*dpif_
, char **devnamep OVS_UNUSED
)
1756 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
1757 uint64_t new_port_seq
;
1760 new_port_seq
= seq_read(dpif
->dp
->port_seq
);
1761 if (dpif
->last_port_seq
!= new_port_seq
) {
1762 dpif
->last_port_seq
= new_port_seq
;
1772 dpif_netdev_port_poll_wait(const struct dpif
*dpif_
)
1774 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
1776 seq_wait(dpif
->dp
->port_seq
, dpif
->last_port_seq
);
1779 static struct dp_netdev_flow
*
1780 dp_netdev_flow_cast(const struct dpcls_rule
*cr
)
1782 return cr
? CONTAINER_OF(cr
, struct dp_netdev_flow
, cr
) : NULL
;
1785 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*flow
)
1787 return ovs_refcount_try_ref_rcu(&flow
->ref_cnt
);
1790 /* netdev_flow_key utilities.
1792 * netdev_flow_key is basically a miniflow. We use these functions
1793 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1794 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1796 * - Since we are dealing exclusively with miniflows created by
1797 * miniflow_extract(), if the map is different the miniflow is different.
1798 * Therefore we can be faster by comparing the map and the miniflow in a
1800 * - These functions can be inlined by the compiler. */
1802 /* Given the number of bits set in miniflow's maps, returns the size of the
1803 * 'netdev_flow_key.mf' */
1804 static inline size_t
1805 netdev_flow_key_size(size_t flow_u64s
)
1807 return sizeof(struct miniflow
) + MINIFLOW_VALUES_SIZE(flow_u64s
);
1811 netdev_flow_key_equal(const struct netdev_flow_key
*a
,
1812 const struct netdev_flow_key
*b
)
1814 /* 'b->len' may be not set yet. */
1815 return a
->hash
== b
->hash
&& !memcmp(&a
->mf
, &b
->mf
, a
->len
);
1818 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1819 * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been
1820 * generated by miniflow_extract. */
1822 netdev_flow_key_equal_mf(const struct netdev_flow_key
*key
,
1823 const struct miniflow
*mf
)
1825 return !memcmp(&key
->mf
, mf
, key
->len
);
1829 netdev_flow_key_clone(struct netdev_flow_key
*dst
,
1830 const struct netdev_flow_key
*src
)
1833 offsetof(struct netdev_flow_key
, mf
) + src
->len
);
1838 netdev_flow_key_from_flow(struct netdev_flow_key
*dst
,
1839 const struct flow
*src
)
1841 struct dp_packet packet
;
1842 uint64_t buf_stub
[512 / 8];
1844 dp_packet_use_stub(&packet
, buf_stub
, sizeof buf_stub
);
1845 pkt_metadata_from_flow(&packet
.md
, src
);
1846 flow_compose(&packet
, src
);
1847 miniflow_extract(&packet
, &dst
->mf
);
1848 dp_packet_uninit(&packet
);
1850 dst
->len
= netdev_flow_key_size(miniflow_n_values(&dst
->mf
));
1851 dst
->hash
= 0; /* Not computed yet. */
1854 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1856 netdev_flow_mask_init(struct netdev_flow_key
*mask
,
1857 const struct match
*match
)
1859 uint64_t *dst
= miniflow_values(&mask
->mf
);
1860 struct flowmap fmap
;
1864 /* Only check masks that make sense for the flow. */
1865 flow_wc_map(&match
->flow
, &fmap
);
1866 flowmap_init(&mask
->mf
.map
);
1868 FLOWMAP_FOR_EACH_INDEX(idx
, fmap
) {
1869 uint64_t mask_u64
= flow_u64_value(&match
->wc
.masks
, idx
);
1872 flowmap_set(&mask
->mf
.map
, idx
, 1);
1874 hash
= hash_add64(hash
, mask_u64
);
1880 FLOWMAP_FOR_EACH_MAP (map
, mask
->mf
.map
) {
1881 hash
= hash_add64(hash
, map
);
1884 size_t n
= dst
- miniflow_get_values(&mask
->mf
);
1886 mask
->hash
= hash_finish(hash
, n
* 8);
1887 mask
->len
= netdev_flow_key_size(n
);
1890 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1892 netdev_flow_key_init_masked(struct netdev_flow_key
*dst
,
1893 const struct flow
*flow
,
1894 const struct netdev_flow_key
*mask
)
1896 uint64_t *dst_u64
= miniflow_values(&dst
->mf
);
1897 const uint64_t *mask_u64
= miniflow_get_values(&mask
->mf
);
1901 dst
->len
= mask
->len
;
1902 dst
->mf
= mask
->mf
; /* Copy maps. */
1904 FLOW_FOR_EACH_IN_MAPS(value
, flow
, mask
->mf
.map
) {
1905 *dst_u64
= value
& *mask_u64
++;
1906 hash
= hash_add64(hash
, *dst_u64
++);
1908 dst
->hash
= hash_finish(hash
,
1909 (dst_u64
- miniflow_get_values(&dst
->mf
)) * 8);
1912 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
1913 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
1914 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
1916 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1918 static inline uint32_t
1919 netdev_flow_key_hash_in_mask(const struct netdev_flow_key
*key
,
1920 const struct netdev_flow_key
*mask
)
1922 const uint64_t *p
= miniflow_get_values(&mask
->mf
);
1926 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, key
, mask
->mf
.map
) {
1927 hash
= hash_add64(hash
, value
& *p
++);
1930 return hash_finish(hash
, (p
- miniflow_get_values(&mask
->mf
)) * 8);
1934 emc_entry_alive(struct emc_entry
*ce
)
1936 return ce
->flow
&& !ce
->flow
->dead
;
1940 emc_clear_entry(struct emc_entry
*ce
)
1943 dp_netdev_flow_unref(ce
->flow
);
1949 emc_change_entry(struct emc_entry
*ce
, struct dp_netdev_flow
*flow
,
1950 const struct netdev_flow_key
*key
)
1952 if (ce
->flow
!= flow
) {
1954 dp_netdev_flow_unref(ce
->flow
);
1957 if (dp_netdev_flow_ref(flow
)) {
1964 netdev_flow_key_clone(&ce
->key
, key
);
1969 emc_insert(struct emc_cache
*cache
, const struct netdev_flow_key
*key
,
1970 struct dp_netdev_flow
*flow
)
1972 struct emc_entry
*to_be_replaced
= NULL
;
1973 struct emc_entry
*current_entry
;
1975 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
1976 if (netdev_flow_key_equal(¤t_entry
->key
, key
)) {
1977 /* We found the entry with the 'mf' miniflow */
1978 emc_change_entry(current_entry
, flow
, NULL
);
1982 /* Replacement policy: put the flow in an empty (not alive) entry, or
1983 * in the first entry where it can be */
1985 || (emc_entry_alive(to_be_replaced
)
1986 && !emc_entry_alive(current_entry
))
1987 || current_entry
->key
.hash
< to_be_replaced
->key
.hash
) {
1988 to_be_replaced
= current_entry
;
1991 /* We didn't find the miniflow in the cache.
1992 * The 'to_be_replaced' entry is where the new flow will be stored */
1994 emc_change_entry(to_be_replaced
, flow
, key
);
1997 static inline struct dp_netdev_flow
*
1998 emc_lookup(struct emc_cache
*cache
, const struct netdev_flow_key
*key
)
2000 struct emc_entry
*current_entry
;
2002 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2003 if (current_entry
->key
.hash
== key
->hash
2004 && emc_entry_alive(current_entry
)
2005 && netdev_flow_key_equal_mf(¤t_entry
->key
, &key
->mf
)) {
2007 /* We found the entry with the 'key->mf' miniflow */
2008 return current_entry
->flow
;
2015 static struct dp_netdev_flow
*
2016 dp_netdev_pmd_lookup_flow(struct dp_netdev_pmd_thread
*pmd
,
2017 const struct netdev_flow_key
*key
,
2021 struct dpcls_rule
*rule
;
2022 odp_port_t in_port
= u32_to_odp(MINIFLOW_GET_U32(&key
->mf
, in_port
));
2023 struct dp_netdev_flow
*netdev_flow
= NULL
;
2025 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2026 if (OVS_LIKELY(cls
)) {
2027 dpcls_lookup(cls
, key
, &rule
, 1, lookup_num_p
);
2028 netdev_flow
= dp_netdev_flow_cast(rule
);
2033 static struct dp_netdev_flow
*
2034 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread
*pmd
,
2035 const ovs_u128
*ufidp
, const struct nlattr
*key
,
2038 struct dp_netdev_flow
*netdev_flow
;
2042 /* If a UFID is not provided, determine one based on the key. */
2043 if (!ufidp
&& key
&& key_len
2044 && !dpif_netdev_flow_from_nlattrs(key
, key_len
, &flow
)) {
2045 dpif_flow_hash(pmd
->dp
->dpif
, &flow
, sizeof flow
, &ufid
);
2050 CMAP_FOR_EACH_WITH_HASH (netdev_flow
, node
, dp_netdev_flow_hash(ufidp
),
2052 if (ovs_u128_equals(netdev_flow
->ufid
, *ufidp
)) {
2062 get_dpif_flow_stats(const struct dp_netdev_flow
*netdev_flow_
,
2063 struct dpif_flow_stats
*stats
)
2065 struct dp_netdev_flow
*netdev_flow
;
2066 unsigned long long n
;
2070 netdev_flow
= CONST_CAST(struct dp_netdev_flow
*, netdev_flow_
);
2072 atomic_read_relaxed(&netdev_flow
->stats
.packet_count
, &n
);
2073 stats
->n_packets
= n
;
2074 atomic_read_relaxed(&netdev_flow
->stats
.byte_count
, &n
);
2076 atomic_read_relaxed(&netdev_flow
->stats
.used
, &used
);
2078 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
2079 stats
->tcp_flags
= flags
;
2082 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
2083 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
2084 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
2087 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow
*netdev_flow
,
2088 struct ofpbuf
*key_buf
, struct ofpbuf
*mask_buf
,
2089 struct dpif_flow
*flow
, bool terse
)
2092 memset(flow
, 0, sizeof *flow
);
2094 struct flow_wildcards wc
;
2095 struct dp_netdev_actions
*actions
;
2097 struct odp_flow_key_parms odp_parms
= {
2098 .flow
= &netdev_flow
->flow
,
2100 .support
= dp_netdev_support
,
2103 miniflow_expand(&netdev_flow
->cr
.mask
->mf
, &wc
.masks
);
2106 offset
= key_buf
->size
;
2107 flow
->key
= ofpbuf_tail(key_buf
);
2108 odp_flow_key_from_flow(&odp_parms
, key_buf
);
2109 flow
->key_len
= key_buf
->size
- offset
;
2112 offset
= mask_buf
->size
;
2113 flow
->mask
= ofpbuf_tail(mask_buf
);
2114 odp_parms
.key_buf
= key_buf
;
2115 odp_flow_key_from_mask(&odp_parms
, mask_buf
);
2116 flow
->mask_len
= mask_buf
->size
- offset
;
2119 actions
= dp_netdev_flow_get_actions(netdev_flow
);
2120 flow
->actions
= actions
->actions
;
2121 flow
->actions_len
= actions
->size
;
2124 flow
->ufid
= netdev_flow
->ufid
;
2125 flow
->ufid_present
= true;
2126 flow
->pmd_id
= netdev_flow
->pmd_id
;
2127 get_dpif_flow_stats(netdev_flow
, &flow
->stats
);
2131 dpif_netdev_mask_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
2132 const struct nlattr
*mask_key
,
2133 uint32_t mask_key_len
, const struct flow
*flow
,
2134 struct flow_wildcards
*wc
)
2136 enum odp_key_fitness fitness
;
2138 fitness
= odp_flow_key_to_mask(mask_key
, mask_key_len
, wc
, flow
);
2140 /* This should not happen: it indicates that
2141 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
2142 * disagree on the acceptable form of a mask. Log the problem
2143 * as an error, with enough details to enable debugging. */
2144 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
2146 if (!VLOG_DROP_ERR(&rl
)) {
2150 odp_flow_format(key
, key_len
, mask_key
, mask_key_len
, NULL
, &s
,
2152 VLOG_ERR("internal error parsing flow mask %s (%s)",
2153 ds_cstr(&s
), odp_key_fitness_to_string(fitness
));
2164 dpif_netdev_flow_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
2169 if (odp_flow_key_to_flow(key
, key_len
, flow
)) {
2170 /* This should not happen: it indicates that odp_flow_key_from_flow()
2171 * and odp_flow_key_to_flow() disagree on the acceptable form of a
2172 * flow. Log the problem as an error, with enough details to enable
2174 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
2176 if (!VLOG_DROP_ERR(&rl
)) {
2180 odp_flow_format(key
, key_len
, NULL
, 0, NULL
, &s
, true);
2181 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s
));
2188 in_port
= flow
->in_port
.odp_port
;
2189 if (!is_valid_port_number(in_port
) && in_port
!= ODPP_NONE
) {
2193 if (flow
->ct_state
& DP_NETDEV_CS_UNSUPPORTED_MASK
) {
2201 dpif_netdev_flow_get(const struct dpif
*dpif
, const struct dpif_flow_get
*get
)
2203 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2204 struct dp_netdev_flow
*netdev_flow
;
2205 struct dp_netdev_pmd_thread
*pmd
;
2206 struct hmapx to_find
= HMAPX_INITIALIZER(&to_find
);
2207 struct hmapx_node
*node
;
2210 if (get
->pmd_id
== PMD_ID_NULL
) {
2211 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2212 if (dp_netdev_pmd_try_ref(pmd
) && !hmapx_add(&to_find
, pmd
)) {
2213 dp_netdev_pmd_unref(pmd
);
2217 pmd
= dp_netdev_get_pmd(dp
, get
->pmd_id
);
2221 hmapx_add(&to_find
, pmd
);
2224 if (!hmapx_count(&to_find
)) {
2228 HMAPX_FOR_EACH (node
, &to_find
) {
2229 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
2230 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, get
->ufid
, get
->key
,
2233 dp_netdev_flow_to_dpif_flow(netdev_flow
, get
->buffer
, get
->buffer
,
2242 HMAPX_FOR_EACH (node
, &to_find
) {
2243 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
2244 dp_netdev_pmd_unref(pmd
);
2247 hmapx_destroy(&to_find
);
2251 static struct dp_netdev_flow
*
2252 dp_netdev_flow_add(struct dp_netdev_pmd_thread
*pmd
,
2253 struct match
*match
, const ovs_u128
*ufid
,
2254 const struct nlattr
*actions
, size_t actions_len
)
2255 OVS_REQUIRES(pmd
->flow_mutex
)
2257 struct dp_netdev_flow
*flow
;
2258 struct netdev_flow_key mask
;
2260 odp_port_t in_port
= match
->flow
.in_port
.odp_port
;
2262 netdev_flow_mask_init(&mask
, match
);
2263 /* Make sure wc does not have metadata. */
2264 ovs_assert(!FLOWMAP_HAS_FIELD(&mask
.mf
.map
, metadata
)
2265 && !FLOWMAP_HAS_FIELD(&mask
.mf
.map
, regs
));
2267 /* Do not allocate extra space. */
2268 flow
= xmalloc(sizeof *flow
- sizeof flow
->cr
.flow
.mf
+ mask
.len
);
2269 memset(&flow
->stats
, 0, sizeof flow
->stats
);
2272 *CONST_CAST(unsigned *, &flow
->pmd_id
) = pmd
->core_id
;
2273 *CONST_CAST(struct flow
*, &flow
->flow
) = match
->flow
;
2274 *CONST_CAST(ovs_u128
*, &flow
->ufid
) = *ufid
;
2275 ovs_refcount_init(&flow
->ref_cnt
);
2276 ovsrcu_set(&flow
->actions
, dp_netdev_actions_create(actions
, actions_len
));
2278 netdev_flow_key_init_masked(&flow
->cr
.flow
, &match
->flow
, &mask
);
2280 /* Select dpcls for in_port. Relies on in_port to be exact match */
2281 ovs_assert(match
->wc
.masks
.in_port
.odp_port
== ODPP_NONE
);
2282 cls
= dp_netdev_pmd_find_dpcls(pmd
, in_port
);
2283 dpcls_insert(cls
, &flow
->cr
, &mask
);
2285 cmap_insert(&pmd
->flow_table
, CONST_CAST(struct cmap_node
*, &flow
->node
),
2286 dp_netdev_flow_hash(&flow
->ufid
));
2288 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2289 struct ds ds
= DS_EMPTY_INITIALIZER
;
2290 struct ofpbuf key_buf
, mask_buf
;
2291 struct odp_flow_key_parms odp_parms
= {
2292 .flow
= &match
->flow
,
2293 .mask
= &match
->wc
.masks
,
2294 .support
= dp_netdev_support
,
2297 ofpbuf_init(&key_buf
, 0);
2298 ofpbuf_init(&mask_buf
, 0);
2300 odp_flow_key_from_flow(&odp_parms
, &key_buf
);
2301 odp_parms
.key_buf
= &key_buf
;
2302 odp_flow_key_from_mask(&odp_parms
, &mask_buf
);
2304 ds_put_cstr(&ds
, "flow_add: ");
2305 odp_format_ufid(ufid
, &ds
);
2306 ds_put_cstr(&ds
, " ");
2307 odp_flow_format(key_buf
.data
, key_buf
.size
,
2308 mask_buf
.data
, mask_buf
.size
,
2310 ds_put_cstr(&ds
, ", actions:");
2311 format_odp_actions(&ds
, actions
, actions_len
);
2313 VLOG_DBG_RL(&upcall_rl
, "%s", ds_cstr(&ds
));
2315 ofpbuf_uninit(&key_buf
);
2316 ofpbuf_uninit(&mask_buf
);
2324 dpif_netdev_flow_put(struct dpif
*dpif
, const struct dpif_flow_put
*put
)
2326 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2327 struct dp_netdev_flow
*netdev_flow
;
2328 struct netdev_flow_key key
;
2329 struct dp_netdev_pmd_thread
*pmd
;
2332 unsigned pmd_id
= put
->pmd_id
== PMD_ID_NULL
2333 ? NON_PMD_CORE_ID
: put
->pmd_id
;
2336 error
= dpif_netdev_flow_from_nlattrs(put
->key
, put
->key_len
, &match
.flow
);
2340 error
= dpif_netdev_mask_from_nlattrs(put
->key
, put
->key_len
,
2341 put
->mask
, put
->mask_len
,
2342 &match
.flow
, &match
.wc
);
2347 pmd
= dp_netdev_get_pmd(dp
, pmd_id
);
2352 /* Must produce a netdev_flow_key for lookup.
2353 * This interface is no longer performance critical, since it is not used
2354 * for upcall processing any more. */
2355 netdev_flow_key_from_flow(&key
, &match
.flow
);
2360 dpif_flow_hash(dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
2363 ovs_mutex_lock(&pmd
->flow_mutex
);
2364 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, &key
, NULL
);
2366 if (put
->flags
& DPIF_FP_CREATE
) {
2367 if (cmap_count(&pmd
->flow_table
) < MAX_FLOWS
) {
2369 memset(put
->stats
, 0, sizeof *put
->stats
);
2371 dp_netdev_flow_add(pmd
, &match
, &ufid
, put
->actions
,
2381 if (put
->flags
& DPIF_FP_MODIFY
2382 && flow_equal(&match
.flow
, &netdev_flow
->flow
)) {
2383 struct dp_netdev_actions
*new_actions
;
2384 struct dp_netdev_actions
*old_actions
;
2386 new_actions
= dp_netdev_actions_create(put
->actions
,
2389 old_actions
= dp_netdev_flow_get_actions(netdev_flow
);
2390 ovsrcu_set(&netdev_flow
->actions
, new_actions
);
2393 get_dpif_flow_stats(netdev_flow
, put
->stats
);
2395 if (put
->flags
& DPIF_FP_ZERO_STATS
) {
2396 /* XXX: The userspace datapath uses thread local statistics
2397 * (for flows), which should be updated only by the owning
2398 * thread. Since we cannot write on stats memory here,
2399 * we choose not to support this flag. Please note:
2400 * - This feature is currently used only by dpctl commands with
2402 * - Should the need arise, this operation can be implemented
2403 * by keeping a base value (to be update here) for each
2404 * counter, and subtracting it before outputting the stats */
2408 ovsrcu_postpone(dp_netdev_actions_free
, old_actions
);
2409 } else if (put
->flags
& DPIF_FP_CREATE
) {
2412 /* Overlapping flow. */
2416 ovs_mutex_unlock(&pmd
->flow_mutex
);
2417 dp_netdev_pmd_unref(pmd
);
2423 dpif_netdev_flow_del(struct dpif
*dpif
, const struct dpif_flow_del
*del
)
2425 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2426 struct dp_netdev_flow
*netdev_flow
;
2427 struct dp_netdev_pmd_thread
*pmd
;
2428 unsigned pmd_id
= del
->pmd_id
== PMD_ID_NULL
2429 ? NON_PMD_CORE_ID
: del
->pmd_id
;
2432 pmd
= dp_netdev_get_pmd(dp
, pmd_id
);
2437 ovs_mutex_lock(&pmd
->flow_mutex
);
2438 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, del
->ufid
, del
->key
,
2442 get_dpif_flow_stats(netdev_flow
, del
->stats
);
2444 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
2448 ovs_mutex_unlock(&pmd
->flow_mutex
);
2449 dp_netdev_pmd_unref(pmd
);
2454 struct dpif_netdev_flow_dump
{
2455 struct dpif_flow_dump up
;
2456 struct cmap_position poll_thread_pos
;
2457 struct cmap_position flow_pos
;
2458 struct dp_netdev_pmd_thread
*cur_pmd
;
2460 struct ovs_mutex mutex
;
2463 static struct dpif_netdev_flow_dump
*
2464 dpif_netdev_flow_dump_cast(struct dpif_flow_dump
*dump
)
2466 return CONTAINER_OF(dump
, struct dpif_netdev_flow_dump
, up
);
2469 static struct dpif_flow_dump
*
2470 dpif_netdev_flow_dump_create(const struct dpif
*dpif_
, bool terse
)
2472 struct dpif_netdev_flow_dump
*dump
;
2474 dump
= xzalloc(sizeof *dump
);
2475 dpif_flow_dump_init(&dump
->up
, dpif_
);
2476 dump
->up
.terse
= terse
;
2477 ovs_mutex_init(&dump
->mutex
);
2483 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump
*dump_
)
2485 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
2487 ovs_mutex_destroy(&dump
->mutex
);
2492 struct dpif_netdev_flow_dump_thread
{
2493 struct dpif_flow_dump_thread up
;
2494 struct dpif_netdev_flow_dump
*dump
;
2495 struct odputil_keybuf keybuf
[FLOW_DUMP_MAX_BATCH
];
2496 struct odputil_keybuf maskbuf
[FLOW_DUMP_MAX_BATCH
];
2499 static struct dpif_netdev_flow_dump_thread
*
2500 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread
*thread
)
2502 return CONTAINER_OF(thread
, struct dpif_netdev_flow_dump_thread
, up
);
2505 static struct dpif_flow_dump_thread
*
2506 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump
*dump_
)
2508 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
2509 struct dpif_netdev_flow_dump_thread
*thread
;
2511 thread
= xmalloc(sizeof *thread
);
2512 dpif_flow_dump_thread_init(&thread
->up
, &dump
->up
);
2513 thread
->dump
= dump
;
2518 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread
*thread_
)
2520 struct dpif_netdev_flow_dump_thread
*thread
2521 = dpif_netdev_flow_dump_thread_cast(thread_
);
2527 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread
*thread_
,
2528 struct dpif_flow
*flows
, int max_flows
)
2530 struct dpif_netdev_flow_dump_thread
*thread
2531 = dpif_netdev_flow_dump_thread_cast(thread_
);
2532 struct dpif_netdev_flow_dump
*dump
= thread
->dump
;
2533 struct dp_netdev_flow
*netdev_flows
[FLOW_DUMP_MAX_BATCH
];
2537 ovs_mutex_lock(&dump
->mutex
);
2538 if (!dump
->status
) {
2539 struct dpif_netdev
*dpif
= dpif_netdev_cast(thread
->up
.dpif
);
2540 struct dp_netdev
*dp
= get_dp_netdev(&dpif
->dpif
);
2541 struct dp_netdev_pmd_thread
*pmd
= dump
->cur_pmd
;
2542 int flow_limit
= MIN(max_flows
, FLOW_DUMP_MAX_BATCH
);
2544 /* First call to dump_next(), extracts the first pmd thread.
2545 * If there is no pmd thread, returns immediately. */
2547 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
2549 ovs_mutex_unlock(&dump
->mutex
);
2556 for (n_flows
= 0; n_flows
< flow_limit
; n_flows
++) {
2557 struct cmap_node
*node
;
2559 node
= cmap_next_position(&pmd
->flow_table
, &dump
->flow_pos
);
2563 netdev_flows
[n_flows
] = CONTAINER_OF(node
,
2564 struct dp_netdev_flow
,
2567 /* When finishing dumping the current pmd thread, moves to
2569 if (n_flows
< flow_limit
) {
2570 memset(&dump
->flow_pos
, 0, sizeof dump
->flow_pos
);
2571 dp_netdev_pmd_unref(pmd
);
2572 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
2578 /* Keeps the reference to next caller. */
2579 dump
->cur_pmd
= pmd
;
2581 /* If the current dump is empty, do not exit the loop, since the
2582 * remaining pmds could have flows to be dumped. Just dumps again
2583 * on the new 'pmd'. */
2586 ovs_mutex_unlock(&dump
->mutex
);
2588 for (i
= 0; i
< n_flows
; i
++) {
2589 struct odputil_keybuf
*maskbuf
= &thread
->maskbuf
[i
];
2590 struct odputil_keybuf
*keybuf
= &thread
->keybuf
[i
];
2591 struct dp_netdev_flow
*netdev_flow
= netdev_flows
[i
];
2592 struct dpif_flow
*f
= &flows
[i
];
2593 struct ofpbuf key
, mask
;
2595 ofpbuf_use_stack(&key
, keybuf
, sizeof *keybuf
);
2596 ofpbuf_use_stack(&mask
, maskbuf
, sizeof *maskbuf
);
2597 dp_netdev_flow_to_dpif_flow(netdev_flow
, &key
, &mask
, f
,
2605 dpif_netdev_execute(struct dpif
*dpif
, struct dpif_execute
*execute
)
2606 OVS_NO_THREAD_SAFETY_ANALYSIS
2608 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2609 struct dp_netdev_pmd_thread
*pmd
;
2610 struct dp_packet_batch pp
;
2612 if (dp_packet_size(execute
->packet
) < ETH_HEADER_LEN
||
2613 dp_packet_size(execute
->packet
) > UINT16_MAX
) {
2617 /* Tries finding the 'pmd'. If NULL is returned, that means
2618 * the current thread is a non-pmd thread and should use
2619 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2620 pmd
= ovsthread_getspecific(dp
->per_pmd_key
);
2622 pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
2628 /* If the current thread is non-pmd thread, acquires
2629 * the 'non_pmd_mutex'. */
2630 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2631 ovs_mutex_lock(&dp
->non_pmd_mutex
);
2634 /* The action processing expects the RSS hash to be valid, because
2635 * it's always initialized at the beginning of datapath processing.
2636 * In this case, though, 'execute->packet' may not have gone through
2637 * the datapath at all, it may have been generated by the upper layer
2638 * (OpenFlow packet-out, BFD frame, ...). */
2639 if (!dp_packet_rss_valid(execute
->packet
)) {
2640 dp_packet_set_rss_hash(execute
->packet
,
2641 flow_hash_5tuple(execute
->flow
, 0));
2644 packet_batch_init_packet(&pp
, execute
->packet
);
2645 dp_netdev_execute_actions(pmd
, &pp
, false, execute
->flow
,
2646 execute
->actions
, execute
->actions_len
,
2649 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2650 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
2651 dp_netdev_pmd_unref(pmd
);
2658 dpif_netdev_operate(struct dpif
*dpif
, struct dpif_op
**ops
, size_t n_ops
)
2662 for (i
= 0; i
< n_ops
; i
++) {
2663 struct dpif_op
*op
= ops
[i
];
2666 case DPIF_OP_FLOW_PUT
:
2667 op
->error
= dpif_netdev_flow_put(dpif
, &op
->u
.flow_put
);
2670 case DPIF_OP_FLOW_DEL
:
2671 op
->error
= dpif_netdev_flow_del(dpif
, &op
->u
.flow_del
);
2674 case DPIF_OP_EXECUTE
:
2675 op
->error
= dpif_netdev_execute(dpif
, &op
->u
.execute
);
2678 case DPIF_OP_FLOW_GET
:
2679 op
->error
= dpif_netdev_flow_get(dpif
, &op
->u
.flow_get
);
2685 /* Changes the number or the affinity of pmd threads. The changes are actually
2686 * applied in dpif_netdev_run(). */
2688 dpif_netdev_pmd_set(struct dpif
*dpif
, const char *cmask
)
2690 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2692 if (!nullable_string_is_equal(dp
->pmd_cmask
, cmask
)) {
2693 free(dp
->pmd_cmask
);
2694 dp
->pmd_cmask
= nullable_xstrdup(cmask
);
2695 dp_netdev_request_reconfigure(dp
);
2701 /* Parses affinity list and returns result in 'core_ids'. */
2703 parse_affinity_list(const char *affinity_list
, unsigned *core_ids
, int n_rxq
)
2706 char *list
, *copy
, *key
, *value
;
2709 for (i
= 0; i
< n_rxq
; i
++) {
2713 if (!affinity_list
) {
2717 list
= copy
= xstrdup(affinity_list
);
2719 while (ofputil_parse_key_value(&list
, &key
, &value
)) {
2720 int rxq_id
, core_id
;
2722 if (!str_to_int(key
, 0, &rxq_id
) || rxq_id
< 0
2723 || !str_to_int(value
, 0, &core_id
) || core_id
< 0) {
2728 if (rxq_id
< n_rxq
) {
2729 core_ids
[rxq_id
] = core_id
;
2737 /* Parses 'affinity_list' and applies configuration if it is valid. */
2739 dpif_netdev_port_set_rxq_affinity(struct dp_netdev_port
*port
,
2740 const char *affinity_list
)
2742 unsigned *core_ids
, i
;
2745 core_ids
= xmalloc(port
->n_rxq
* sizeof *core_ids
);
2746 if (parse_affinity_list(affinity_list
, core_ids
, port
->n_rxq
)) {
2751 for (i
= 0; i
< port
->n_rxq
; i
++) {
2752 port
->rxqs
[i
].core_id
= core_ids
[i
];
2760 /* Changes the affinity of port's rx queues. The changes are actually applied
2761 * in dpif_netdev_run(). */
2763 dpif_netdev_port_set_config(struct dpif
*dpif
, odp_port_t port_no
,
2764 const struct smap
*cfg
)
2766 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2767 struct dp_netdev_port
*port
;
2769 const char *affinity_list
= smap_get(cfg
, "pmd-rxq-affinity");
2771 ovs_mutex_lock(&dp
->port_mutex
);
2772 error
= get_port_by_number(dp
, port_no
, &port
);
2773 if (error
|| !netdev_is_pmd(port
->netdev
)
2774 || nullable_string_is_equal(affinity_list
, port
->rxq_affinity_list
)) {
2778 error
= dpif_netdev_port_set_rxq_affinity(port
, affinity_list
);
2782 free(port
->rxq_affinity_list
);
2783 port
->rxq_affinity_list
= nullable_xstrdup(affinity_list
);
2785 dp_netdev_request_reconfigure(dp
);
2787 ovs_mutex_unlock(&dp
->port_mutex
);
2792 dpif_netdev_queue_to_priority(const struct dpif
*dpif OVS_UNUSED
,
2793 uint32_t queue_id
, uint32_t *priority
)
2795 *priority
= queue_id
;
2800 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2801 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2802 struct dp_netdev_actions
*
2803 dp_netdev_actions_create(const struct nlattr
*actions
, size_t size
)
2805 struct dp_netdev_actions
*netdev_actions
;
2807 netdev_actions
= xmalloc(sizeof *netdev_actions
+ size
);
2808 memcpy(netdev_actions
->actions
, actions
, size
);
2809 netdev_actions
->size
= size
;
2811 return netdev_actions
;
2814 struct dp_netdev_actions
*
2815 dp_netdev_flow_get_actions(const struct dp_netdev_flow
*flow
)
2817 return ovsrcu_get(struct dp_netdev_actions
*, &flow
->actions
);
2821 dp_netdev_actions_free(struct dp_netdev_actions
*actions
)
2826 static inline unsigned long long
2827 cycles_counter(void)
2830 return rte_get_tsc_cycles();
2836 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2837 extern struct ovs_mutex cycles_counter_fake_mutex
;
2839 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2841 cycles_count_start(struct dp_netdev_pmd_thread
*pmd
)
2842 OVS_ACQUIRES(&cycles_counter_fake_mutex
)
2843 OVS_NO_THREAD_SAFETY_ANALYSIS
2845 pmd
->last_cycles
= cycles_counter();
2848 /* Stop counting cycles and add them to the counter 'type' */
2850 cycles_count_end(struct dp_netdev_pmd_thread
*pmd
,
2851 enum pmd_cycles_counter_type type
)
2852 OVS_RELEASES(&cycles_counter_fake_mutex
)
2853 OVS_NO_THREAD_SAFETY_ANALYSIS
2855 unsigned long long interval
= cycles_counter() - pmd
->last_cycles
;
2857 non_atomic_ullong_add(&pmd
->cycles
.n
[type
], interval
);
2861 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread
*pmd
,
2862 struct dp_netdev_port
*port
,
2863 struct netdev_rxq
*rxq
)
2865 struct dp_packet_batch batch
;
2868 dp_packet_batch_init(&batch
);
2869 cycles_count_start(pmd
);
2870 error
= netdev_rxq_recv(rxq
, &batch
);
2871 cycles_count_end(pmd
, PMD_CYCLES_POLLING
);
2873 *recirc_depth_get() = 0;
2875 cycles_count_start(pmd
);
2876 dp_netdev_input(pmd
, &batch
, port
->port_no
);
2877 cycles_count_end(pmd
, PMD_CYCLES_PROCESSING
);
2878 } else if (error
!= EAGAIN
&& error
!= EOPNOTSUPP
) {
2879 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
2881 VLOG_ERR_RL(&rl
, "error receiving data from %s: %s",
2882 netdev_get_name(port
->netdev
), ovs_strerror(error
));
2887 port_reconfigure(struct dp_netdev_port
*port
)
2889 struct netdev
*netdev
= port
->netdev
;
2892 if (!netdev_is_reconf_required(netdev
)) {
2896 /* Closes the existing 'rxq's. */
2897 for (i
= 0; i
< port
->n_rxq
; i
++) {
2898 netdev_rxq_close(port
->rxqs
[i
].rxq
);
2899 port
->rxqs
[i
].rxq
= NULL
;
2903 /* Allows 'netdev' to apply the pending configuration changes. */
2904 err
= netdev_reconfigure(netdev
);
2905 if (err
&& (err
!= EOPNOTSUPP
)) {
2906 VLOG_ERR("Failed to set interface %s new configuration",
2907 netdev_get_name(netdev
));
2910 /* If the netdev_reconfigure() above succeeds, reopens the 'rxq's. */
2911 port
->rxqs
= xrealloc(port
->rxqs
,
2912 sizeof *port
->rxqs
* netdev_n_rxq(netdev
));
2913 /* Realloc 'used' counters for tx queues. */
2914 free(port
->txq_used
);
2915 port
->txq_used
= xcalloc(netdev_n_txq(netdev
), sizeof *port
->txq_used
);
2917 for (i
= 0; i
< netdev_n_rxq(netdev
); i
++) {
2918 err
= netdev_rxq_open(netdev
, &port
->rxqs
[i
].rxq
, i
);
2925 /* Parse affinity list to apply configuration for new queues. */
2926 dpif_netdev_port_set_rxq_affinity(port
, port
->rxq_affinity_list
);
2932 reconfigure_pmd_threads(struct dp_netdev
*dp
)
2933 OVS_REQUIRES(dp
->port_mutex
)
2935 struct dp_netdev_port
*port
, *next
;
2938 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
2940 dp_netdev_destroy_all_pmds(dp
);
2942 /* Reconfigures the cpu mask. */
2943 ovs_numa_set_cpu_mask(dp
->pmd_cmask
);
2945 n_cores
= ovs_numa_get_n_cores();
2946 if (n_cores
== OVS_CORE_UNSPEC
) {
2947 VLOG_ERR("Cannot get cpu core info");
2951 HMAP_FOR_EACH_SAFE (port
, next
, node
, &dp
->ports
) {
2954 err
= port_reconfigure(port
);
2956 hmap_remove(&dp
->ports
, &port
->node
);
2957 seq_change(dp
->port_seq
);
2960 port
->dynamic_txqs
= netdev_n_txq(port
->netdev
) < n_cores
+ 1;
2963 /* Restores the non-pmd. */
2964 dp_netdev_set_nonpmd(dp
);
2965 /* Restores all pmd threads. */
2966 dp_netdev_reset_pmd_threads(dp
);
2969 /* Returns true if one of the netdevs in 'dp' requires a reconfiguration */
2971 ports_require_restart(const struct dp_netdev
*dp
)
2972 OVS_REQUIRES(dp
->port_mutex
)
2974 struct dp_netdev_port
*port
;
2976 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2977 if (netdev_is_reconf_required(port
->netdev
)) {
2985 /* Return true if needs to revalidate datapath flows. */
2987 dpif_netdev_run(struct dpif
*dpif
)
2989 struct dp_netdev_port
*port
;
2990 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2991 struct dp_netdev_pmd_thread
*non_pmd
;
2992 uint64_t new_tnl_seq
;
2994 ovs_mutex_lock(&dp
->port_mutex
);
2995 non_pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
2997 ovs_mutex_lock(&dp
->non_pmd_mutex
);
2998 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2999 if (!netdev_is_pmd(port
->netdev
)) {
3002 for (i
= 0; i
< port
->n_rxq
; i
++) {
3003 dp_netdev_process_rxq_port(non_pmd
, port
,
3008 dpif_netdev_xps_revalidate_pmd(non_pmd
, time_msec(), false);
3009 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
3011 dp_netdev_pmd_unref(non_pmd
);
3014 if (dp_netdev_is_reconf_required(dp
) || ports_require_restart(dp
)) {
3015 reconfigure_pmd_threads(dp
);
3017 ovs_mutex_unlock(&dp
->port_mutex
);
3019 tnl_neigh_cache_run();
3021 new_tnl_seq
= seq_read(tnl_conf_seq
);
3023 if (dp
->last_tnl_conf_seq
!= new_tnl_seq
) {
3024 dp
->last_tnl_conf_seq
= new_tnl_seq
;
3031 dpif_netdev_wait(struct dpif
*dpif
)
3033 struct dp_netdev_port
*port
;
3034 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3036 ovs_mutex_lock(&dp_netdev_mutex
);
3037 ovs_mutex_lock(&dp
->port_mutex
);
3038 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3039 netdev_wait_reconf_required(port
->netdev
);
3040 if (!netdev_is_pmd(port
->netdev
)) {
3043 for (i
= 0; i
< port
->n_rxq
; i
++) {
3044 netdev_rxq_wait(port
->rxqs
[i
].rxq
);
3048 ovs_mutex_unlock(&dp
->port_mutex
);
3049 ovs_mutex_unlock(&dp_netdev_mutex
);
3050 seq_wait(tnl_conf_seq
, dp
->last_tnl_conf_seq
);
3054 pmd_free_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
3056 struct tx_port
*tx_port_cached
;
3058 /* Free all used tx queue ids. */
3059 dpif_netdev_xps_revalidate_pmd(pmd
, 0, true);
3061 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->port_cache
) {
3062 free(tx_port_cached
);
3066 /* Copies ports from 'pmd->tx_ports' (shared with the main thread) to
3067 * 'pmd->port_cache' (thread local) */
3069 pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
3070 OVS_REQUIRES(pmd
->port_mutex
)
3072 struct tx_port
*tx_port
, *tx_port_cached
;
3074 pmd_free_cached_ports(pmd
);
3075 hmap_shrink(&pmd
->port_cache
);
3077 HMAP_FOR_EACH (tx_port
, node
, &pmd
->tx_ports
) {
3078 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
3079 hmap_insert(&pmd
->port_cache
, &tx_port_cached
->node
,
3080 hash_port_no(tx_port_cached
->port
->port_no
));
3085 pmd_load_queues_and_ports(struct dp_netdev_pmd_thread
*pmd
,
3086 struct rxq_poll
**ppoll_list
)
3088 struct rxq_poll
*poll_list
= *ppoll_list
;
3089 struct rxq_poll
*poll
;
3092 ovs_mutex_lock(&pmd
->port_mutex
);
3093 poll_list
= xrealloc(poll_list
, pmd
->poll_cnt
* sizeof *poll_list
);
3096 LIST_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
3097 poll_list
[i
++] = *poll
;
3100 pmd_load_cached_ports(pmd
);
3102 ovs_mutex_unlock(&pmd
->port_mutex
);
3104 *ppoll_list
= poll_list
;
3109 pmd_thread_main(void *f_
)
3111 struct dp_netdev_pmd_thread
*pmd
= f_
;
3112 unsigned int lc
= 0;
3113 struct rxq_poll
*poll_list
;
3114 unsigned int port_seq
= PMD_INITIAL_SEQ
;
3121 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
3122 ovsthread_setspecific(pmd
->dp
->per_pmd_key
, pmd
);
3123 ovs_numa_thread_setaffinity_core(pmd
->core_id
);
3124 dpdk_set_lcore_id(pmd
->core_id
);
3125 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
3127 emc_cache_init(&pmd
->flow_cache
);
3129 /* List port/core affinity */
3130 for (i
= 0; i
< poll_cnt
; i
++) {
3131 VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n",
3132 pmd
->core_id
, netdev_get_name(poll_list
[i
].port
->netdev
),
3133 netdev_rxq_get_queue_id(poll_list
[i
].rx
));
3137 for (i
= 0; i
< poll_cnt
; i
++) {
3138 dp_netdev_process_rxq_port(pmd
, poll_list
[i
].port
, poll_list
[i
].rx
);
3146 coverage_try_clear();
3147 dp_netdev_pmd_try_optimize(pmd
);
3148 if (!ovsrcu_try_quiesce()) {
3149 emc_cache_slow_sweep(&pmd
->flow_cache
);
3152 atomic_read_relaxed(&pmd
->change_seq
, &seq
);
3153 if (seq
!= port_seq
) {
3160 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
3161 exiting
= latch_is_set(&pmd
->exit_latch
);
3162 /* Signal here to make sure the pmd finishes
3163 * reloading the updated configuration. */
3164 dp_netdev_pmd_reload_done(pmd
);
3166 emc_cache_uninit(&pmd
->flow_cache
);
3173 pmd_free_cached_ports(pmd
);
3178 dp_netdev_disable_upcall(struct dp_netdev
*dp
)
3179 OVS_ACQUIRES(dp
->upcall_rwlock
)
3181 fat_rwlock_wrlock(&dp
->upcall_rwlock
);
3185 dpif_netdev_disable_upcall(struct dpif
*dpif
)
3186 OVS_NO_THREAD_SAFETY_ANALYSIS
3188 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3189 dp_netdev_disable_upcall(dp
);
3193 dp_netdev_enable_upcall(struct dp_netdev
*dp
)
3194 OVS_RELEASES(dp
->upcall_rwlock
)
3196 fat_rwlock_unlock(&dp
->upcall_rwlock
);
3200 dpif_netdev_enable_upcall(struct dpif
*dpif
)
3201 OVS_NO_THREAD_SAFETY_ANALYSIS
3203 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3204 dp_netdev_enable_upcall(dp
);
3208 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
)
3210 ovs_mutex_lock(&pmd
->cond_mutex
);
3211 xpthread_cond_signal(&pmd
->cond
);
3212 ovs_mutex_unlock(&pmd
->cond_mutex
);
3215 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
3216 * the pointer if succeeds, otherwise, NULL (it can return NULL even if
3217 * 'core_id' is NON_PMD_CORE_ID).
3219 * Caller must unrefs the returned reference. */
3220 static struct dp_netdev_pmd_thread
*
3221 dp_netdev_get_pmd(struct dp_netdev
*dp
, unsigned core_id
)
3223 struct dp_netdev_pmd_thread
*pmd
;
3224 const struct cmap_node
*pnode
;
3226 pnode
= cmap_find(&dp
->poll_threads
, hash_int(core_id
, 0));
3230 pmd
= CONTAINER_OF(pnode
, struct dp_netdev_pmd_thread
, node
);
3232 return dp_netdev_pmd_try_ref(pmd
) ? pmd
: NULL
;
3235 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
3237 dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
3238 OVS_REQUIRES(dp
->port_mutex
)
3240 struct dp_netdev_pmd_thread
*non_pmd
;
3241 struct dp_netdev_port
*port
;
3243 non_pmd
= xzalloc(sizeof *non_pmd
);
3244 dp_netdev_configure_pmd(non_pmd
, dp
, NON_PMD_CORE_ID
, OVS_NUMA_UNSPEC
);
3246 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3247 dp_netdev_add_port_tx_to_pmd(non_pmd
, port
);
3250 dp_netdev_reload_pmd__(non_pmd
);
3253 /* Caller must have valid pointer to 'pmd'. */
3255 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
)
3257 return ovs_refcount_try_ref_rcu(&pmd
->ref_cnt
);
3261 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
)
3263 if (pmd
&& ovs_refcount_unref(&pmd
->ref_cnt
) == 1) {
3264 ovsrcu_postpone(dp_netdev_destroy_pmd
, pmd
);
3268 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
3269 * fails, keeps checking for next node until reaching the end of cmap.
3271 * Caller must unrefs the returned reference. */
3272 static struct dp_netdev_pmd_thread
*
3273 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
)
3275 struct dp_netdev_pmd_thread
*next
;
3278 struct cmap_node
*node
;
3280 node
= cmap_next_position(&dp
->poll_threads
, pos
);
3281 next
= node
? CONTAINER_OF(node
, struct dp_netdev_pmd_thread
, node
)
3283 } while (next
&& !dp_netdev_pmd_try_ref(next
));
3288 /* Configures the 'pmd' based on the input argument. */
3290 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
, struct dp_netdev
*dp
,
3291 unsigned core_id
, int numa_id
)
3294 pmd
->core_id
= core_id
;
3295 pmd
->numa_id
= numa_id
;
3298 atomic_init(&pmd
->static_tx_qid
,
3299 (core_id
== NON_PMD_CORE_ID
)
3300 ? ovs_numa_get_n_cores()
3301 : get_n_pmd_threads(dp
));
3303 ovs_refcount_init(&pmd
->ref_cnt
);
3304 latch_init(&pmd
->exit_latch
);
3305 atomic_init(&pmd
->change_seq
, PMD_INITIAL_SEQ
);
3306 xpthread_cond_init(&pmd
->cond
, NULL
);
3307 ovs_mutex_init(&pmd
->cond_mutex
);
3308 ovs_mutex_init(&pmd
->flow_mutex
);
3309 ovs_mutex_init(&pmd
->port_mutex
);
3310 cmap_init(&pmd
->flow_table
);
3311 cmap_init(&pmd
->classifiers
);
3312 pmd
->next_optimization
= time_msec() + DPCLS_OPTIMIZATION_INTERVAL
;
3313 ovs_list_init(&pmd
->poll_list
);
3314 hmap_init(&pmd
->tx_ports
);
3315 hmap_init(&pmd
->port_cache
);
3316 /* init the 'flow_cache' since there is no
3317 * actual thread created for NON_PMD_CORE_ID. */
3318 if (core_id
== NON_PMD_CORE_ID
) {
3319 emc_cache_init(&pmd
->flow_cache
);
3321 cmap_insert(&dp
->poll_threads
, CONST_CAST(struct cmap_node
*, &pmd
->node
),
3322 hash_int(core_id
, 0));
3326 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
)
3330 dp_netdev_pmd_flow_flush(pmd
);
3331 hmap_destroy(&pmd
->port_cache
);
3332 hmap_destroy(&pmd
->tx_ports
);
3333 /* All flows (including their dpcls_rules) have been deleted already */
3334 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
3337 cmap_destroy(&pmd
->classifiers
);
3338 cmap_destroy(&pmd
->flow_table
);
3339 ovs_mutex_destroy(&pmd
->flow_mutex
);
3340 latch_destroy(&pmd
->exit_latch
);
3341 xpthread_cond_destroy(&pmd
->cond
);
3342 ovs_mutex_destroy(&pmd
->cond_mutex
);
3343 ovs_mutex_destroy(&pmd
->port_mutex
);
3347 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
3348 * and unrefs the struct. */
3350 dp_netdev_del_pmd(struct dp_netdev
*dp
, struct dp_netdev_pmd_thread
*pmd
)
3352 /* NON_PMD_CORE_ID doesn't have a thread, so we don't have to synchronize,
3353 * but extra cleanup is necessary */
3354 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3355 emc_cache_uninit(&pmd
->flow_cache
);
3356 pmd_free_cached_ports(pmd
);
3358 latch_set(&pmd
->exit_latch
);
3359 dp_netdev_reload_pmd__(pmd
);
3360 ovs_numa_unpin_core(pmd
->core_id
);
3361 xpthread_join(pmd
->thread
, NULL
);
3364 dp_netdev_pmd_clear_ports(pmd
);
3366 /* Purges the 'pmd''s flows after stopping the thread, but before
3367 * destroying the flows, so that the flow stats can be collected. */
3368 if (dp
->dp_purge_cb
) {
3369 dp
->dp_purge_cb(dp
->dp_purge_aux
, pmd
->core_id
);
3371 cmap_remove(&pmd
->dp
->poll_threads
, &pmd
->node
, hash_int(pmd
->core_id
, 0));
3372 dp_netdev_pmd_unref(pmd
);
3375 /* Destroys all pmd threads. */
3377 dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
)
3379 struct dp_netdev_pmd_thread
*pmd
;
3380 struct dp_netdev_pmd_thread
**pmd_list
;
3381 size_t k
= 0, n_pmds
;
3383 n_pmds
= cmap_count(&dp
->poll_threads
);
3384 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
3386 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3387 /* We cannot call dp_netdev_del_pmd(), since it alters
3388 * 'dp->poll_threads' (while we're iterating it) and it
3390 ovs_assert(k
< n_pmds
);
3391 pmd_list
[k
++] = pmd
;
3394 for (size_t i
= 0; i
< k
; i
++) {
3395 dp_netdev_del_pmd(dp
, pmd_list
[i
]);
3400 /* Deletes all pmd threads on numa node 'numa_id' and
3401 * fixes static_tx_qids of other threads to keep them sequential. */
3403 dp_netdev_del_pmds_on_numa(struct dp_netdev
*dp
, int numa_id
)
3405 struct dp_netdev_pmd_thread
*pmd
;
3406 int n_pmds_on_numa
, n_pmds
;
3407 int *free_idx
, k
= 0;
3408 struct dp_netdev_pmd_thread
**pmd_list
;
3410 n_pmds_on_numa
= get_n_pmd_threads_on_numa(dp
, numa_id
);
3411 free_idx
= xcalloc(n_pmds_on_numa
, sizeof *free_idx
);
3412 pmd_list
= xcalloc(n_pmds_on_numa
, sizeof *pmd_list
);
3414 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3415 /* We cannot call dp_netdev_del_pmd(), since it alters
3416 * 'dp->poll_threads' (while we're iterating it) and it
3418 if (pmd
->numa_id
== numa_id
&& pmd
->core_id
!= NON_PMD_CORE_ID
) {
3419 atomic_read_relaxed(&pmd
->static_tx_qid
, &free_idx
[k
]);
3421 ovs_assert(k
< n_pmds_on_numa
);
3426 for (int i
= 0; i
< k
; i
++) {
3427 dp_netdev_del_pmd(dp
, pmd_list
[i
]);
3430 n_pmds
= get_n_pmd_threads(dp
);
3431 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3434 atomic_read_relaxed(&pmd
->static_tx_qid
, &old_tx_qid
);
3436 if (old_tx_qid
>= n_pmds
&& pmd
->core_id
!= NON_PMD_CORE_ID
) {
3437 int new_tx_qid
= free_idx
[--k
];
3439 atomic_store_relaxed(&pmd
->static_tx_qid
, new_tx_qid
);
3447 /* Deletes all rx queues from pmd->poll_list and all the ports from
3450 dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
)
3452 struct rxq_poll
*poll
;
3453 struct tx_port
*port
;
3455 ovs_mutex_lock(&pmd
->port_mutex
);
3456 LIST_FOR_EACH_POP (poll
, node
, &pmd
->poll_list
) {
3460 HMAP_FOR_EACH_POP (port
, node
, &pmd
->tx_ports
) {
3463 ovs_mutex_unlock(&pmd
->port_mutex
);
3466 static struct tx_port
*
3467 tx_port_lookup(const struct hmap
*hmap
, odp_port_t port_no
)
3471 HMAP_FOR_EACH_IN_BUCKET (tx
, node
, hash_port_no(port_no
), hmap
) {
3472 if (tx
->port
->port_no
== port_no
) {
3480 /* Deletes all rx queues of 'port' from 'poll_list', and the 'port' from
3481 * 'tx_ports' of 'pmd' thread. Returns true if 'port' was found in 'pmd'
3482 * (therefore a restart is required). */
3484 dp_netdev_del_port_from_pmd__(struct dp_netdev_port
*port
,
3485 struct dp_netdev_pmd_thread
*pmd
)
3487 struct rxq_poll
*poll
, *next
;
3491 ovs_mutex_lock(&pmd
->port_mutex
);
3492 LIST_FOR_EACH_SAFE (poll
, next
, node
, &pmd
->poll_list
) {
3493 if (poll
->port
== port
) {
3495 ovs_list_remove(&poll
->node
);
3501 tx
= tx_port_lookup(&pmd
->tx_ports
, port
->port_no
);
3503 hmap_remove(&pmd
->tx_ports
, &tx
->node
);
3507 ovs_mutex_unlock(&pmd
->port_mutex
);
3512 /* Deletes 'port' from the 'poll_list' and from the 'tx_ports' of all the pmd
3513 * threads. The pmd threads that need to be restarted are inserted in
3516 dp_netdev_del_port_from_all_pmds__(struct dp_netdev
*dp
,
3517 struct dp_netdev_port
*port
,
3518 struct hmapx
*to_reload
)
3520 struct dp_netdev_pmd_thread
*pmd
;
3522 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3525 found
= dp_netdev_del_port_from_pmd__(port
, pmd
);
3528 hmapx_add(to_reload
, pmd
);
3533 /* Deletes 'port' from the 'poll_list' and from the 'tx_ports' of all the pmd
3534 * threads. Reloads the threads if needed. */
3536 dp_netdev_del_port_from_all_pmds(struct dp_netdev
*dp
,
3537 struct dp_netdev_port
*port
)
3539 struct dp_netdev_pmd_thread
*pmd
;
3540 struct hmapx to_reload
= HMAPX_INITIALIZER(&to_reload
);
3541 struct hmapx_node
*node
;
3543 dp_netdev_del_port_from_all_pmds__(dp
, port
, &to_reload
);
3545 HMAPX_FOR_EACH (node
, &to_reload
) {
3546 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3547 dp_netdev_reload_pmd__(pmd
);
3550 hmapx_destroy(&to_reload
);
3554 /* Returns non-isolated PMD thread from this numa node with fewer
3555 * rx queues to poll. Returns NULL if there is no non-isolated PMD threads
3556 * on this numa node. Can be called safely only by main thread. */
3557 static struct dp_netdev_pmd_thread
*
3558 dp_netdev_less_loaded_pmd_on_numa(struct dp_netdev
*dp
, int numa_id
)
3561 struct dp_netdev_pmd_thread
*pmd
, *res
= NULL
;
3563 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3564 if (!pmd
->isolated
&& pmd
->numa_id
== numa_id
3565 && (min_cnt
> pmd
->poll_cnt
|| res
== NULL
)) {
3566 min_cnt
= pmd
->poll_cnt
;
3574 /* Adds rx queue to poll_list of PMD thread. */
3576 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
3577 struct dp_netdev_port
*port
, struct netdev_rxq
*rx
)
3578 OVS_REQUIRES(pmd
->port_mutex
)
3580 struct rxq_poll
*poll
= xmalloc(sizeof *poll
);
3585 ovs_list_push_back(&pmd
->poll_list
, &poll
->node
);
3589 /* Add 'port' to the tx port cache of 'pmd', which must be reloaded for the
3590 * changes to take effect. */
3592 dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
3593 struct dp_netdev_port
*port
)
3595 struct tx_port
*tx
= xzalloc(sizeof *tx
);
3600 ovs_mutex_lock(&pmd
->port_mutex
);
3601 hmap_insert(&pmd
->tx_ports
, &tx
->node
, hash_port_no(tx
->port
->port_no
));
3602 ovs_mutex_unlock(&pmd
->port_mutex
);
3605 /* Distribute all {pinned|non-pinned} rx queues of 'port' between PMD
3606 * threads in 'dp'. The pmd threads that need to be restarted are inserted
3607 * in 'to_reload'. PMD threads with pinned queues marked as isolated. */
3609 dp_netdev_add_port_rx_to_pmds(struct dp_netdev
*dp
,
3610 struct dp_netdev_port
*port
,
3611 struct hmapx
*to_reload
, bool pinned
)
3613 int numa_id
= netdev_get_numa_id(port
->netdev
);
3614 struct dp_netdev_pmd_thread
*pmd
;
3617 if (!netdev_is_pmd(port
->netdev
)) {
3621 for (i
= 0; i
< port
->n_rxq
; i
++) {
3623 if (port
->rxqs
[i
].core_id
== -1) {
3626 pmd
= dp_netdev_get_pmd(dp
, port
->rxqs
[i
].core_id
);
3628 VLOG_WARN("There is no PMD thread on core %d. "
3629 "Queue %d on port \'%s\' will not be polled.",
3630 port
->rxqs
[i
].core_id
, i
,
3631 netdev_get_name(port
->netdev
));
3634 pmd
->isolated
= true;
3635 dp_netdev_pmd_unref(pmd
);
3637 if (port
->rxqs
[i
].core_id
!= -1) {
3640 pmd
= dp_netdev_less_loaded_pmd_on_numa(dp
, numa_id
);
3642 VLOG_WARN("There's no available pmd thread on numa node %d",
3648 ovs_mutex_lock(&pmd
->port_mutex
);
3649 dp_netdev_add_rxq_to_pmd(pmd
, port
, port
->rxqs
[i
].rxq
);
3650 ovs_mutex_unlock(&pmd
->port_mutex
);
3652 hmapx_add(to_reload
, pmd
);
3656 /* Distributes all non-pinned rx queues of 'port' between all PMD threads
3657 * in 'dp' and inserts 'port' in the PMD threads 'tx_ports'. The pmd threads
3658 * that need to be restarted are inserted in 'to_reload'. */
3660 dp_netdev_add_port_to_pmds__(struct dp_netdev
*dp
, struct dp_netdev_port
*port
,
3661 struct hmapx
*to_reload
)
3663 struct dp_netdev_pmd_thread
*pmd
;
3665 dp_netdev_add_port_rx_to_pmds(dp
, port
, to_reload
, false);
3667 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3668 dp_netdev_add_port_tx_to_pmd(pmd
, port
);
3669 hmapx_add(to_reload
, pmd
);
3673 /* Distributes all non-pinned rx queues of 'port' between all PMD threads
3674 * in 'dp', inserts 'port' in the PMD threads 'tx_ports' and reloads them,
3677 dp_netdev_add_port_to_pmds(struct dp_netdev
*dp
, struct dp_netdev_port
*port
)
3679 struct dp_netdev_pmd_thread
*pmd
;
3680 struct hmapx to_reload
= HMAPX_INITIALIZER(&to_reload
);
3681 struct hmapx_node
*node
;
3683 dp_netdev_add_port_to_pmds__(dp
, port
, &to_reload
);
3685 HMAPX_FOR_EACH (node
, &to_reload
) {
3686 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3687 dp_netdev_reload_pmd__(pmd
);
3690 hmapx_destroy(&to_reload
);
3693 /* Starts pmd threads for the numa node 'numa_id', if not already started.
3694 * The function takes care of filling the threads tx port cache. */
3696 dp_netdev_set_pmds_on_numa(struct dp_netdev
*dp
, int numa_id
)
3697 OVS_REQUIRES(dp
->port_mutex
)
3701 if (!ovs_numa_numa_id_is_valid(numa_id
)) {
3702 VLOG_WARN("Cannot create pmd threads due to numa id (%d) invalid",
3707 n_pmds
= get_n_pmd_threads_on_numa(dp
, numa_id
);
3709 /* If there are already pmd threads created for the numa node
3710 * in which 'netdev' is on, do nothing. Else, creates the
3711 * pmd threads for the numa node. */
3713 int can_have
, n_unpinned
, i
;
3715 n_unpinned
= ovs_numa_get_n_unpinned_cores_on_numa(numa_id
);
3717 VLOG_WARN("Cannot create pmd threads due to out of unpinned "
3718 "cores on numa node %d", numa_id
);
3722 /* If cpu mask is specified, uses all unpinned cores, otherwise
3723 * tries creating NR_PMD_THREADS pmd threads. */
3724 can_have
= dp
->pmd_cmask
? n_unpinned
: MIN(n_unpinned
, NR_PMD_THREADS
);
3725 for (i
= 0; i
< can_have
; i
++) {
3726 unsigned core_id
= ovs_numa_get_unpinned_core_on_numa(numa_id
);
3727 struct dp_netdev_pmd_thread
*pmd
= xzalloc(sizeof *pmd
);
3728 struct dp_netdev_port
*port
;
3730 dp_netdev_configure_pmd(pmd
, dp
, core_id
, numa_id
);
3732 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3733 dp_netdev_add_port_tx_to_pmd(pmd
, port
);
3736 pmd
->thread
= ovs_thread_create("pmd", pmd_thread_main
, pmd
);
3738 VLOG_INFO("Created %d pmd threads on numa node %d", can_have
, numa_id
);
3743 /* Called after pmd threads config change. Restarts pmd threads with
3744 * new configuration. */
3746 dp_netdev_reset_pmd_threads(struct dp_netdev
*dp
)
3747 OVS_REQUIRES(dp
->port_mutex
)
3749 struct hmapx to_reload
= HMAPX_INITIALIZER(&to_reload
);
3750 struct dp_netdev_pmd_thread
*pmd
;
3751 struct dp_netdev_port
*port
;
3752 struct hmapx_node
*node
;
3754 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3755 if (netdev_is_pmd(port
->netdev
)) {
3756 struct hmapx numas
= HMAPX_INITIALIZER(&numas
);
3757 struct hmapx_node
*numa_node
;
3761 numa_id
= netdev_get_numa_id(port
->netdev
);
3762 hmapx_add(&numas
, (void *) numa_id
);
3763 for (i
= 0; i
< port
->n_rxq
; i
++) {
3764 unsigned core_id
= port
->rxqs
[i
].core_id
;
3766 if (core_id
!= -1) {
3767 numa_id
= ovs_numa_get_numa_id(core_id
);
3768 hmapx_add(&numas
, (void *) numa_id
);
3772 HMAPX_FOR_EACH (numa_node
, &numas
) {
3773 dp_netdev_set_pmds_on_numa(dp
, (uintptr_t) numa_node
->data
);
3776 hmapx_destroy(&numas
);
3778 /* Distribute only pinned rx queues first to mark threads as isolated */
3779 dp_netdev_add_port_rx_to_pmds(dp
, port
, &to_reload
, true);
3782 /* Distribute remaining non-pinned rx queues to non-isolated PMD threads. */
3783 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3784 dp_netdev_add_port_rx_to_pmds(dp
, port
, &to_reload
, false);
3787 HMAPX_FOR_EACH (node
, &to_reload
) {
3788 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3789 dp_netdev_reload_pmd__(pmd
);
3792 hmapx_destroy(&to_reload
);
3796 dpif_netdev_get_datapath_version(void)
3798 return xstrdup("<built-in>");
3802 dp_netdev_flow_used(struct dp_netdev_flow
*netdev_flow
, int cnt
, int size
,
3803 uint16_t tcp_flags
, long long now
)
3807 atomic_store_relaxed(&netdev_flow
->stats
.used
, now
);
3808 non_atomic_ullong_add(&netdev_flow
->stats
.packet_count
, cnt
);
3809 non_atomic_ullong_add(&netdev_flow
->stats
.byte_count
, size
);
3810 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
3812 atomic_store_relaxed(&netdev_flow
->stats
.tcp_flags
, flags
);
3816 dp_netdev_count_packet(struct dp_netdev_pmd_thread
*pmd
,
3817 enum dp_stat_type type
, int cnt
)
3819 non_atomic_ullong_add(&pmd
->stats
.n
[type
], cnt
);
3823 dp_netdev_upcall(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
*packet_
,
3824 struct flow
*flow
, struct flow_wildcards
*wc
, ovs_u128
*ufid
,
3825 enum dpif_upcall_type type
, const struct nlattr
*userdata
,
3826 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
3828 struct dp_netdev
*dp
= pmd
->dp
;
3830 if (OVS_UNLIKELY(!dp
->upcall_cb
)) {
3834 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl
))) {
3835 struct ds ds
= DS_EMPTY_INITIALIZER
;
3838 struct odp_flow_key_parms odp_parms
= {
3840 .mask
= wc
? &wc
->masks
: NULL
,
3841 .support
= dp_netdev_support
,
3844 ofpbuf_init(&key
, 0);
3845 odp_flow_key_from_flow(&odp_parms
, &key
);
3846 packet_str
= ofp_packet_to_string(dp_packet_data(packet_
),
3847 dp_packet_size(packet_
));
3849 odp_flow_key_format(key
.data
, key
.size
, &ds
);
3851 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp
->name
,
3852 dpif_upcall_type_to_string(type
), ds_cstr(&ds
), packet_str
);
3854 ofpbuf_uninit(&key
);
3860 return dp
->upcall_cb(packet_
, flow
, ufid
, pmd
->core_id
, type
, userdata
,
3861 actions
, wc
, put_actions
, dp
->upcall_aux
);
3864 static inline uint32_t
3865 dpif_netdev_packet_get_rss_hash(struct dp_packet
*packet
,
3866 const struct miniflow
*mf
)
3868 uint32_t hash
, recirc_depth
;
3870 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
3871 hash
= dp_packet_get_rss_hash(packet
);
3873 hash
= miniflow_hash_5tuple(mf
, 0);
3874 dp_packet_set_rss_hash(packet
, hash
);
3877 /* The RSS hash must account for the recirculation depth to avoid
3878 * collisions in the exact match cache */
3879 recirc_depth
= *recirc_depth_get_unsafe();
3880 if (OVS_UNLIKELY(recirc_depth
)) {
3881 hash
= hash_finish(hash
, recirc_depth
);
3882 dp_packet_set_rss_hash(packet
, hash
);
3887 struct packet_batch_per_flow
{
3888 unsigned int byte_count
;
3890 struct dp_netdev_flow
*flow
;
3892 struct dp_packet_batch array
;
3896 packet_batch_per_flow_update(struct packet_batch_per_flow
*batch
,
3897 struct dp_packet
*packet
,
3898 const struct miniflow
*mf
)
3900 batch
->byte_count
+= dp_packet_size(packet
);
3901 batch
->tcp_flags
|= miniflow_get_tcp_flags(mf
);
3902 batch
->array
.packets
[batch
->array
.count
++] = packet
;
3906 packet_batch_per_flow_init(struct packet_batch_per_flow
*batch
,
3907 struct dp_netdev_flow
*flow
)
3909 flow
->batch
= batch
;
3912 dp_packet_batch_init(&batch
->array
);
3913 batch
->byte_count
= 0;
3914 batch
->tcp_flags
= 0;
3918 packet_batch_per_flow_execute(struct packet_batch_per_flow
*batch
,
3919 struct dp_netdev_pmd_thread
*pmd
,
3922 struct dp_netdev_actions
*actions
;
3923 struct dp_netdev_flow
*flow
= batch
->flow
;
3925 dp_netdev_flow_used(flow
, batch
->array
.count
, batch
->byte_count
,
3926 batch
->tcp_flags
, now
);
3928 actions
= dp_netdev_flow_get_actions(flow
);
3930 dp_netdev_execute_actions(pmd
, &batch
->array
, true, &flow
->flow
,
3931 actions
->actions
, actions
->size
, now
);
3935 dp_netdev_queue_batches(struct dp_packet
*pkt
,
3936 struct dp_netdev_flow
*flow
, const struct miniflow
*mf
,
3937 struct packet_batch_per_flow
*batches
, size_t *n_batches
)
3939 struct packet_batch_per_flow
*batch
= flow
->batch
;
3941 if (OVS_UNLIKELY(!batch
)) {
3942 batch
= &batches
[(*n_batches
)++];
3943 packet_batch_per_flow_init(batch
, flow
);
3946 packet_batch_per_flow_update(batch
, pkt
, mf
);
3949 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3950 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
3951 * miniflow is copied into 'keys' and the packet pointer is moved at the
3952 * beginning of the 'packets' array.
3954 * The function returns the number of packets that needs to be processed in the
3955 * 'packets' array (they have been moved to the beginning of the vector).
3957 * If 'md_is_valid' is false, the metadata in 'packets' is not valid and must be
3958 * initialized by this function using 'port_no'.
3960 static inline size_t
3961 emc_processing(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet_batch
*packets_
,
3962 struct netdev_flow_key
*keys
,
3963 struct packet_batch_per_flow batches
[], size_t *n_batches
,
3964 bool md_is_valid
, odp_port_t port_no
)
3966 struct emc_cache
*flow_cache
= &pmd
->flow_cache
;
3967 struct netdev_flow_key
*key
= &keys
[0];
3968 size_t i
, n_missed
= 0, n_dropped
= 0;
3969 struct dp_packet
**packets
= packets_
->packets
;
3970 int cnt
= packets_
->count
;
3972 for (i
= 0; i
< cnt
; i
++) {
3973 struct dp_netdev_flow
*flow
;
3974 struct dp_packet
*packet
= packets
[i
];
3976 if (OVS_UNLIKELY(dp_packet_size(packet
) < ETH_HEADER_LEN
)) {
3977 dp_packet_delete(packet
);
3983 /* Prefetch next packet data and metadata. */
3984 OVS_PREFETCH(dp_packet_data(packets
[i
+1]));
3985 pkt_metadata_prefetch_init(&packets
[i
+1]->md
);
3989 pkt_metadata_init(&packet
->md
, port_no
);
3991 miniflow_extract(packet
, &key
->mf
);
3992 key
->len
= 0; /* Not computed yet. */
3993 key
->hash
= dpif_netdev_packet_get_rss_hash(packet
, &key
->mf
);
3995 flow
= emc_lookup(flow_cache
, key
);
3996 if (OVS_LIKELY(flow
)) {
3997 dp_netdev_queue_batches(packet
, flow
, &key
->mf
, batches
,
4000 /* Exact match cache missed. Group missed packets together at
4001 * the beginning of the 'packets' array. */
4002 packets
[n_missed
] = packet
;
4003 /* 'key[n_missed]' contains the key of the current packet and it
4004 * must be returned to the caller. The next key should be extracted
4005 * to 'keys[n_missed + 1]'. */
4006 key
= &keys
[++n_missed
];
4010 dp_netdev_count_packet(pmd
, DP_STAT_EXACT_HIT
, cnt
- n_dropped
- n_missed
);
4016 handle_packet_upcall(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
*packet
,
4017 const struct netdev_flow_key
*key
,
4018 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
,
4019 int *lost_cnt
, long long now
)
4021 struct ofpbuf
*add_actions
;
4022 struct dp_packet_batch b
;
4027 match
.tun_md
.valid
= false;
4028 miniflow_expand(&key
->mf
, &match
.flow
);
4030 ofpbuf_clear(actions
);
4031 ofpbuf_clear(put_actions
);
4033 dpif_flow_hash(pmd
->dp
->dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
4034 error
= dp_netdev_upcall(pmd
, packet
, &match
.flow
, &match
.wc
,
4035 &ufid
, DPIF_UC_MISS
, NULL
, actions
,
4037 if (OVS_UNLIKELY(error
&& error
!= ENOSPC
)) {
4038 dp_packet_delete(packet
);
4043 /* The Netlink encoding of datapath flow keys cannot express
4044 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
4045 * tag is interpreted as exact match on the fact that there is no
4046 * VLAN. Unless we refactor a lot of code that translates between
4047 * Netlink and struct flow representations, we have to do the same
4049 if (!match
.wc
.masks
.vlan_tci
) {
4050 match
.wc
.masks
.vlan_tci
= htons(0xffff);
4053 /* We can't allow the packet batching in the next loop to execute
4054 * the actions. Otherwise, if there are any slow path actions,
4055 * we'll send the packet up twice. */
4056 packet_batch_init_packet(&b
, packet
);
4057 dp_netdev_execute_actions(pmd
, &b
, true, &match
.flow
,
4058 actions
->data
, actions
->size
, now
);
4060 add_actions
= put_actions
->size
? put_actions
: actions
;
4061 if (OVS_LIKELY(error
!= ENOSPC
)) {
4062 struct dp_netdev_flow
*netdev_flow
;
4064 /* XXX: There's a race window where a flow covering this packet
4065 * could have already been installed since we last did the flow
4066 * lookup before upcall. This could be solved by moving the
4067 * mutex lock outside the loop, but that's an awful long time
4068 * to be locking everyone out of making flow installs. If we
4069 * move to a per-core classifier, it would be reasonable. */
4070 ovs_mutex_lock(&pmd
->flow_mutex
);
4071 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
4072 if (OVS_LIKELY(!netdev_flow
)) {
4073 netdev_flow
= dp_netdev_flow_add(pmd
, &match
, &ufid
,
4077 ovs_mutex_unlock(&pmd
->flow_mutex
);
4079 emc_insert(&pmd
->flow_cache
, key
, netdev_flow
);
4084 fast_path_processing(struct dp_netdev_pmd_thread
*pmd
,
4085 struct dp_packet_batch
*packets_
,
4086 struct netdev_flow_key
*keys
,
4087 struct packet_batch_per_flow batches
[], size_t *n_batches
,
4091 int cnt
= packets_
->count
;
4092 #if !defined(__CHECKER__) && !defined(_WIN32)
4093 const size_t PKT_ARRAY_SIZE
= cnt
;
4095 /* Sparse or MSVC doesn't like variable length array. */
4096 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
4098 struct dp_packet
**packets
= packets_
->packets
;
4100 struct dpcls_rule
*rules
[PKT_ARRAY_SIZE
];
4101 struct dp_netdev
*dp
= pmd
->dp
;
4102 struct emc_cache
*flow_cache
= &pmd
->flow_cache
;
4103 int miss_cnt
= 0, lost_cnt
= 0;
4104 int lookup_cnt
= 0, add_lookup_cnt
;
4108 for (i
= 0; i
< cnt
; i
++) {
4109 /* Key length is needed in all the cases, hash computed on demand. */
4110 keys
[i
].len
= netdev_flow_key_size(miniflow_n_values(&keys
[i
].mf
));
4112 /* Get the classifier for the in_port */
4113 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
4114 if (OVS_LIKELY(cls
)) {
4115 any_miss
= !dpcls_lookup(cls
, keys
, rules
, cnt
, &lookup_cnt
);
4118 memset(rules
, 0, sizeof(rules
));
4120 if (OVS_UNLIKELY(any_miss
) && !fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
4121 uint64_t actions_stub
[512 / 8], slow_stub
[512 / 8];
4122 struct ofpbuf actions
, put_actions
;
4124 ofpbuf_use_stub(&actions
, actions_stub
, sizeof actions_stub
);
4125 ofpbuf_use_stub(&put_actions
, slow_stub
, sizeof slow_stub
);
4127 for (i
= 0; i
< cnt
; i
++) {
4128 struct dp_netdev_flow
*netdev_flow
;
4130 if (OVS_LIKELY(rules
[i
])) {
4134 /* It's possible that an earlier slow path execution installed
4135 * a rule covering this flow. In this case, it's a lot cheaper
4136 * to catch it here than execute a miss. */
4137 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, &keys
[i
],
4140 lookup_cnt
+= add_lookup_cnt
;
4141 rules
[i
] = &netdev_flow
->cr
;
4146 handle_packet_upcall(pmd
, packets
[i
], &keys
[i
], &actions
,
4147 &put_actions
, &lost_cnt
, now
);
4150 ofpbuf_uninit(&actions
);
4151 ofpbuf_uninit(&put_actions
);
4152 fat_rwlock_unlock(&dp
->upcall_rwlock
);
4153 dp_netdev_count_packet(pmd
, DP_STAT_LOST
, lost_cnt
);
4154 } else if (OVS_UNLIKELY(any_miss
)) {
4155 for (i
= 0; i
< cnt
; i
++) {
4156 if (OVS_UNLIKELY(!rules
[i
])) {
4157 dp_packet_delete(packets
[i
]);
4164 for (i
= 0; i
< cnt
; i
++) {
4165 struct dp_packet
*packet
= packets
[i
];
4166 struct dp_netdev_flow
*flow
;
4168 if (OVS_UNLIKELY(!rules
[i
])) {
4172 flow
= dp_netdev_flow_cast(rules
[i
]);
4174 emc_insert(flow_cache
, &keys
[i
], flow
);
4175 dp_netdev_queue_batches(packet
, flow
, &keys
[i
].mf
, batches
, n_batches
);
4178 dp_netdev_count_packet(pmd
, DP_STAT_MASKED_HIT
, cnt
- miss_cnt
);
4179 dp_netdev_count_packet(pmd
, DP_STAT_LOOKUP_HIT
, lookup_cnt
);
4180 dp_netdev_count_packet(pmd
, DP_STAT_MISS
, miss_cnt
);
4181 dp_netdev_count_packet(pmd
, DP_STAT_LOST
, lost_cnt
);
4184 /* Packets enter the datapath from a port (or from recirculation) here.
4186 * For performance reasons a caller may choose not to initialize the metadata
4187 * in 'packets': in this case 'mdinit' is false and this function needs to
4188 * initialize it using 'port_no'. If the metadata in 'packets' is already
4189 * valid, 'md_is_valid' must be true and 'port_no' will be ignored. */
4191 dp_netdev_input__(struct dp_netdev_pmd_thread
*pmd
,
4192 struct dp_packet_batch
*packets
,
4193 bool md_is_valid
, odp_port_t port_no
)
4195 int cnt
= packets
->count
;
4196 #if !defined(__CHECKER__) && !defined(_WIN32)
4197 const size_t PKT_ARRAY_SIZE
= cnt
;
4199 /* Sparse or MSVC doesn't like variable length array. */
4200 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
4202 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
) struct netdev_flow_key keys
[PKT_ARRAY_SIZE
];
4203 struct packet_batch_per_flow batches
[PKT_ARRAY_SIZE
];
4204 long long now
= time_msec();
4205 size_t newcnt
, n_batches
, i
;
4209 newcnt
= emc_processing(pmd
, packets
, keys
, batches
, &n_batches
,
4210 md_is_valid
, port_no
);
4211 if (OVS_UNLIKELY(newcnt
)) {
4212 packets
->count
= newcnt
;
4213 /* Get ingress port from first packet's metadata. */
4214 in_port
= packets
->packets
[0]->md
.in_port
.odp_port
;
4215 fast_path_processing(pmd
, packets
, keys
, batches
, &n_batches
, in_port
, now
);
4218 /* All the flow batches need to be reset before any call to
4219 * packet_batch_per_flow_execute() as it could potentially trigger
4220 * recirculation. When a packet matching flow ‘j’ happens to be
4221 * recirculated, the nested call to dp_netdev_input__() could potentially
4222 * classify the packet as matching another flow - say 'k'. It could happen
4223 * that in the previous call to dp_netdev_input__() that same flow 'k' had
4224 * already its own batches[k] still waiting to be served. So if its
4225 * ‘batch’ member is not reset, the recirculated packet would be wrongly
4226 * appended to batches[k] of the 1st call to dp_netdev_input__(). */
4227 for (i
= 0; i
< n_batches
; i
++) {
4228 batches
[i
].flow
->batch
= NULL
;
4231 for (i
= 0; i
< n_batches
; i
++) {
4232 packet_batch_per_flow_execute(&batches
[i
], pmd
, now
);
4237 dp_netdev_input(struct dp_netdev_pmd_thread
*pmd
,
4238 struct dp_packet_batch
*packets
,
4241 dp_netdev_input__(pmd
, packets
, false, port_no
);
4245 dp_netdev_recirculate(struct dp_netdev_pmd_thread
*pmd
,
4246 struct dp_packet_batch
*packets
)
4248 dp_netdev_input__(pmd
, packets
, true, 0);
4251 struct dp_netdev_execute_aux
{
4252 struct dp_netdev_pmd_thread
*pmd
;
4254 const struct flow
*flow
;
4258 dpif_netdev_register_dp_purge_cb(struct dpif
*dpif
, dp_purge_callback
*cb
,
4261 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4262 dp
->dp_purge_aux
= aux
;
4263 dp
->dp_purge_cb
= cb
;
4267 dpif_netdev_register_upcall_cb(struct dpif
*dpif
, upcall_callback
*cb
,
4270 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4271 dp
->upcall_aux
= aux
;
4276 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
4277 long long now
, bool purge
)
4280 struct dp_netdev_port
*port
;
4283 HMAP_FOR_EACH (tx
, node
, &pmd
->port_cache
) {
4284 if (!tx
->port
->dynamic_txqs
) {
4287 interval
= now
- tx
->last_used
;
4288 if (tx
->qid
>= 0 && (purge
|| interval
>= XPS_TIMEOUT_MS
)) {
4290 ovs_mutex_lock(&port
->txq_used_mutex
);
4291 port
->txq_used
[tx
->qid
]--;
4292 ovs_mutex_unlock(&port
->txq_used_mutex
);
4299 dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
4300 struct tx_port
*tx
, long long now
)
4302 struct dp_netdev_port
*port
;
4304 int i
, min_cnt
, min_qid
;
4306 if (OVS_UNLIKELY(!now
)) {
4310 interval
= now
- tx
->last_used
;
4311 tx
->last_used
= now
;
4313 if (OVS_LIKELY(tx
->qid
>= 0 && interval
< XPS_TIMEOUT_MS
)) {
4319 ovs_mutex_lock(&port
->txq_used_mutex
);
4321 port
->txq_used
[tx
->qid
]--;
4327 for (i
= 0; i
< netdev_n_txq(port
->netdev
); i
++) {
4328 if (port
->txq_used
[i
] < min_cnt
|| min_cnt
== -1) {
4329 min_cnt
= port
->txq_used
[i
];
4334 port
->txq_used
[min_qid
]++;
4337 ovs_mutex_unlock(&port
->txq_used_mutex
);
4339 dpif_netdev_xps_revalidate_pmd(pmd
, now
, false);
4341 VLOG_DBG("Core %d: New TX queue ID %d for port \'%s\'.",
4342 pmd
->core_id
, tx
->qid
, netdev_get_name(tx
->port
->netdev
));
4346 static struct tx_port
*
4347 pmd_tx_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
4350 return tx_port_lookup(&pmd
->port_cache
, port_no
);
4354 push_tnl_action(const struct dp_netdev_pmd_thread
*pmd
,
4355 const struct nlattr
*attr
,
4356 struct dp_packet_batch
*batch
)
4358 struct tx_port
*tun_port
;
4359 const struct ovs_action_push_tnl
*data
;
4362 data
= nl_attr_get(attr
);
4364 tun_port
= pmd_tx_port_cache_lookup(pmd
, u32_to_odp(data
->tnl_port
));
4369 err
= netdev_push_header(tun_port
->port
->netdev
, batch
, data
);
4374 dp_packet_delete_batch(batch
, true);
4379 dp_execute_userspace_action(struct dp_netdev_pmd_thread
*pmd
,
4380 struct dp_packet
*packet
, bool may_steal
,
4381 struct flow
*flow
, ovs_u128
*ufid
,
4382 struct ofpbuf
*actions
,
4383 const struct nlattr
*userdata
, long long now
)
4385 struct dp_packet_batch b
;
4388 ofpbuf_clear(actions
);
4390 error
= dp_netdev_upcall(pmd
, packet
, flow
, NULL
, ufid
,
4391 DPIF_UC_ACTION
, userdata
, actions
,
4393 if (!error
|| error
== ENOSPC
) {
4394 packet_batch_init_packet(&b
, packet
);
4395 dp_netdev_execute_actions(pmd
, &b
, may_steal
, flow
,
4396 actions
->data
, actions
->size
, now
);
4397 } else if (may_steal
) {
4398 dp_packet_delete(packet
);
4403 dp_execute_cb(void *aux_
, struct dp_packet_batch
*packets_
,
4404 const struct nlattr
*a
, bool may_steal
)
4406 struct dp_netdev_execute_aux
*aux
= aux_
;
4407 uint32_t *depth
= recirc_depth_get();
4408 struct dp_netdev_pmd_thread
*pmd
= aux
->pmd
;
4409 struct dp_netdev
*dp
= pmd
->dp
;
4410 int type
= nl_attr_type(a
);
4411 long long now
= aux
->now
;
4414 switch ((enum ovs_action_attr
)type
) {
4415 case OVS_ACTION_ATTR_OUTPUT
:
4416 p
= pmd_tx_port_cache_lookup(pmd
, nl_attr_get_odp_port(a
));
4417 if (OVS_LIKELY(p
)) {
4421 dynamic_txqs
= p
->port
->dynamic_txqs
;
4423 tx_qid
= dpif_netdev_xps_get_tx_qid(pmd
, p
, now
);
4425 atomic_read_relaxed(&pmd
->static_tx_qid
, &tx_qid
);
4428 netdev_send(p
->port
->netdev
, tx_qid
, packets_
, may_steal
,
4434 case OVS_ACTION_ATTR_TUNNEL_PUSH
:
4435 if (*depth
< MAX_RECIRC_DEPTH
) {
4436 struct dp_packet_batch tnl_pkt
;
4437 struct dp_packet_batch
*orig_packets_
= packets_
;
4441 dp_packet_batch_clone(&tnl_pkt
, packets_
);
4442 packets_
= &tnl_pkt
;
4443 dp_packet_batch_reset_cutlen(orig_packets_
);
4446 dp_packet_batch_apply_cutlen(packets_
);
4448 err
= push_tnl_action(pmd
, a
, packets_
);
4451 dp_netdev_recirculate(pmd
, packets_
);
4458 case OVS_ACTION_ATTR_TUNNEL_POP
:
4459 if (*depth
< MAX_RECIRC_DEPTH
) {
4460 struct dp_packet_batch
*orig_packets_
= packets_
;
4461 odp_port_t portno
= nl_attr_get_odp_port(a
);
4463 p
= pmd_tx_port_cache_lookup(pmd
, portno
);
4465 struct dp_packet_batch tnl_pkt
;
4469 dp_packet_batch_clone(&tnl_pkt
, packets_
);
4470 packets_
= &tnl_pkt
;
4471 dp_packet_batch_reset_cutlen(orig_packets_
);
4474 dp_packet_batch_apply_cutlen(packets_
);
4476 netdev_pop_header(p
->port
->netdev
, packets_
);
4477 if (!packets_
->count
) {
4481 for (i
= 0; i
< packets_
->count
; i
++) {
4482 packets_
->packets
[i
]->md
.in_port
.odp_port
= portno
;
4486 dp_netdev_recirculate(pmd
, packets_
);
4493 case OVS_ACTION_ATTR_USERSPACE
:
4494 if (!fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
4495 struct dp_packet_batch
*orig_packets_
= packets_
;
4496 struct dp_packet
**packets
= packets_
->packets
;
4497 const struct nlattr
*userdata
;
4498 struct dp_packet_batch usr_pkt
;
4499 struct ofpbuf actions
;
4505 userdata
= nl_attr_find_nested(a
, OVS_USERSPACE_ATTR_USERDATA
);
4506 ofpbuf_init(&actions
, 0);
4508 if (packets_
->trunc
) {
4510 dp_packet_batch_clone(&usr_pkt
, packets_
);
4511 packets_
= &usr_pkt
;
4512 packets
= packets_
->packets
;
4514 dp_packet_batch_reset_cutlen(orig_packets_
);
4517 dp_packet_batch_apply_cutlen(packets_
);
4520 for (i
= 0; i
< packets_
->count
; i
++) {
4521 flow_extract(packets
[i
], &flow
);
4522 dpif_flow_hash(dp
->dpif
, &flow
, sizeof flow
, &ufid
);
4523 dp_execute_userspace_action(pmd
, packets
[i
], may_steal
, &flow
,
4524 &ufid
, &actions
, userdata
, now
);
4528 dp_packet_delete_batch(packets_
, true);
4531 ofpbuf_uninit(&actions
);
4532 fat_rwlock_unlock(&dp
->upcall_rwlock
);
4538 case OVS_ACTION_ATTR_RECIRC
:
4539 if (*depth
< MAX_RECIRC_DEPTH
) {
4540 struct dp_packet_batch recirc_pkts
;
4544 dp_packet_batch_clone(&recirc_pkts
, packets_
);
4545 packets_
= &recirc_pkts
;
4548 for (i
= 0; i
< packets_
->count
; i
++) {
4549 packets_
->packets
[i
]->md
.recirc_id
= nl_attr_get_u32(a
);
4553 dp_netdev_recirculate(pmd
, packets_
);
4559 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
4562 case OVS_ACTION_ATTR_CT
: {
4563 const struct nlattr
*b
;
4564 bool commit
= false;
4567 const char *helper
= NULL
;
4568 const uint32_t *setmark
= NULL
;
4569 const struct ovs_key_ct_labels
*setlabel
= NULL
;
4571 NL_ATTR_FOR_EACH_UNSAFE (b
, left
, nl_attr_get(a
),
4572 nl_attr_get_size(a
)) {
4573 enum ovs_ct_attr sub_type
= nl_attr_type(b
);
4576 case OVS_CT_ATTR_COMMIT
:
4579 case OVS_CT_ATTR_ZONE
:
4580 zone
= nl_attr_get_u16(b
);
4582 case OVS_CT_ATTR_HELPER
:
4583 helper
= nl_attr_get_string(b
);
4585 case OVS_CT_ATTR_MARK
:
4586 setmark
= nl_attr_get(b
);
4588 case OVS_CT_ATTR_LABELS
:
4589 setlabel
= nl_attr_get(b
);
4591 case OVS_CT_ATTR_NAT
:
4592 case OVS_CT_ATTR_UNSPEC
:
4593 case __OVS_CT_ATTR_MAX
:
4598 conntrack_execute(&dp
->conntrack
, packets_
, aux
->flow
->dl_type
, commit
,
4599 zone
, setmark
, setlabel
, helper
);
4603 case OVS_ACTION_ATTR_PUSH_VLAN
:
4604 case OVS_ACTION_ATTR_POP_VLAN
:
4605 case OVS_ACTION_ATTR_PUSH_MPLS
:
4606 case OVS_ACTION_ATTR_POP_MPLS
:
4607 case OVS_ACTION_ATTR_SET
:
4608 case OVS_ACTION_ATTR_SET_MASKED
:
4609 case OVS_ACTION_ATTR_SAMPLE
:
4610 case OVS_ACTION_ATTR_HASH
:
4611 case OVS_ACTION_ATTR_UNSPEC
:
4612 case OVS_ACTION_ATTR_TRUNC
:
4613 case __OVS_ACTION_ATTR_MAX
:
4617 dp_packet_delete_batch(packets_
, may_steal
);
4621 dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
4622 struct dp_packet_batch
*packets
,
4623 bool may_steal
, const struct flow
*flow
,
4624 const struct nlattr
*actions
, size_t actions_len
,
4627 struct dp_netdev_execute_aux aux
= { pmd
, now
, flow
};
4629 odp_execute_actions(&aux
, packets
, may_steal
, actions
,
4630 actions_len
, dp_execute_cb
);
4633 struct dp_netdev_ct_dump
{
4634 struct ct_dpif_dump_state up
;
4635 struct conntrack_dump dump
;
4636 struct conntrack
*ct
;
4637 struct dp_netdev
*dp
;
4641 dpif_netdev_ct_dump_start(struct dpif
*dpif
, struct ct_dpif_dump_state
**dump_
,
4642 const uint16_t *pzone
)
4644 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4645 struct dp_netdev_ct_dump
*dump
;
4647 dump
= xzalloc(sizeof *dump
);
4649 dump
->ct
= &dp
->conntrack
;
4651 conntrack_dump_start(&dp
->conntrack
, &dump
->dump
, pzone
);
4659 dpif_netdev_ct_dump_next(struct dpif
*dpif OVS_UNUSED
,
4660 struct ct_dpif_dump_state
*dump_
,
4661 struct ct_dpif_entry
*entry
)
4663 struct dp_netdev_ct_dump
*dump
;
4665 INIT_CONTAINER(dump
, dump_
, up
);
4667 return conntrack_dump_next(&dump
->dump
, entry
);
4671 dpif_netdev_ct_dump_done(struct dpif
*dpif OVS_UNUSED
,
4672 struct ct_dpif_dump_state
*dump_
)
4674 struct dp_netdev_ct_dump
*dump
;
4677 INIT_CONTAINER(dump
, dump_
, up
);
4679 err
= conntrack_dump_done(&dump
->dump
);
4687 dpif_netdev_ct_flush(struct dpif
*dpif
, const uint16_t *zone
)
4689 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4691 return conntrack_flush(&dp
->conntrack
, zone
);
4694 const struct dpif_class dpif_netdev_class
= {
4697 dpif_netdev_enumerate
,
4698 dpif_netdev_port_open_type
,
4701 dpif_netdev_destroy
,
4704 dpif_netdev_get_stats
,
4705 dpif_netdev_port_add
,
4706 dpif_netdev_port_del
,
4707 dpif_netdev_port_set_config
,
4708 dpif_netdev_port_query_by_number
,
4709 dpif_netdev_port_query_by_name
,
4710 NULL
, /* port_get_pid */
4711 dpif_netdev_port_dump_start
,
4712 dpif_netdev_port_dump_next
,
4713 dpif_netdev_port_dump_done
,
4714 dpif_netdev_port_poll
,
4715 dpif_netdev_port_poll_wait
,
4716 dpif_netdev_flow_flush
,
4717 dpif_netdev_flow_dump_create
,
4718 dpif_netdev_flow_dump_destroy
,
4719 dpif_netdev_flow_dump_thread_create
,
4720 dpif_netdev_flow_dump_thread_destroy
,
4721 dpif_netdev_flow_dump_next
,
4722 dpif_netdev_operate
,
4723 NULL
, /* recv_set */
4724 NULL
, /* handlers_set */
4725 dpif_netdev_pmd_set
,
4726 dpif_netdev_queue_to_priority
,
4728 NULL
, /* recv_wait */
4729 NULL
, /* recv_purge */
4730 dpif_netdev_register_dp_purge_cb
,
4731 dpif_netdev_register_upcall_cb
,
4732 dpif_netdev_enable_upcall
,
4733 dpif_netdev_disable_upcall
,
4734 dpif_netdev_get_datapath_version
,
4735 dpif_netdev_ct_dump_start
,
4736 dpif_netdev_ct_dump_next
,
4737 dpif_netdev_ct_dump_done
,
4738 dpif_netdev_ct_flush
,
4742 dpif_dummy_change_port_number(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
4743 const char *argv
[], void *aux OVS_UNUSED
)
4745 struct dp_netdev_port
*port
;
4746 struct dp_netdev
*dp
;
4749 ovs_mutex_lock(&dp_netdev_mutex
);
4750 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
4751 if (!dp
|| !dpif_netdev_class_is_dummy(dp
->class)) {
4752 ovs_mutex_unlock(&dp_netdev_mutex
);
4753 unixctl_command_reply_error(conn
, "unknown datapath or not a dummy");
4756 ovs_refcount_ref(&dp
->ref_cnt
);
4757 ovs_mutex_unlock(&dp_netdev_mutex
);
4759 ovs_mutex_lock(&dp
->port_mutex
);
4760 if (get_port_by_name(dp
, argv
[2], &port
)) {
4761 unixctl_command_reply_error(conn
, "unknown port");
4765 port_no
= u32_to_odp(atoi(argv
[3]));
4766 if (!port_no
|| port_no
== ODPP_NONE
) {
4767 unixctl_command_reply_error(conn
, "bad port number");
4770 if (dp_netdev_lookup_port(dp
, port_no
)) {
4771 unixctl_command_reply_error(conn
, "port number already in use");
4776 hmap_remove(&dp
->ports
, &port
->node
);
4777 dp_netdev_del_port_from_all_pmds(dp
, port
);
4779 /* Reinsert with new port number. */
4780 port
->port_no
= port_no
;
4781 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
4782 dp_netdev_add_port_to_pmds(dp
, port
);
4784 seq_change(dp
->port_seq
);
4785 unixctl_command_reply(conn
, NULL
);
4788 ovs_mutex_unlock(&dp
->port_mutex
);
4789 dp_netdev_unref(dp
);
4793 dpif_dummy_register__(const char *type
)
4795 struct dpif_class
*class;
4797 class = xmalloc(sizeof *class);
4798 *class = dpif_netdev_class
;
4799 class->type
= xstrdup(type
);
4800 dp_register_provider(class);
4804 dpif_dummy_override(const char *type
)
4809 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
4810 * a userland-only build. It's useful for testsuite.
4812 error
= dp_unregister_provider(type
);
4813 if (error
== 0 || error
== EAFNOSUPPORT
) {
4814 dpif_dummy_register__(type
);
4819 dpif_dummy_register(enum dummy_level level
)
4821 if (level
== DUMMY_OVERRIDE_ALL
) {
4826 dp_enumerate_types(&types
);
4827 SSET_FOR_EACH (type
, &types
) {
4828 dpif_dummy_override(type
);
4830 sset_destroy(&types
);
4831 } else if (level
== DUMMY_OVERRIDE_SYSTEM
) {
4832 dpif_dummy_override("system");
4835 dpif_dummy_register__("dummy");
4837 unixctl_command_register("dpif-dummy/change-port-number",
4838 "dp port new-number",
4839 3, 3, dpif_dummy_change_port_number
, NULL
);
4842 /* Datapath Classifier. */
4844 /* A set of rules that all have the same fields wildcarded. */
4845 struct dpcls_subtable
{
4846 /* The fields are only used by writers. */
4847 struct cmap_node cmap_node OVS_GUARDED
; /* Within dpcls 'subtables_map'. */
4849 /* These fields are accessed by readers. */
4850 struct cmap rules
; /* Contains "struct dpcls_rule"s. */
4851 uint32_t hit_cnt
; /* Number of match hits in subtable in current
4852 optimization interval. */
4853 struct netdev_flow_key mask
; /* Wildcards for fields (const). */
4854 /* 'mask' must be the last field, additional space is allocated here. */
4857 /* Initializes 'cls' as a classifier that initially contains no classification
4860 dpcls_init(struct dpcls
*cls
)
4862 cmap_init(&cls
->subtables_map
);
4863 pvector_init(&cls
->subtables
);
4867 dpcls_destroy_subtable(struct dpcls
*cls
, struct dpcls_subtable
*subtable
)
4869 VLOG_DBG("Destroying subtable %p for in_port %d", subtable
, cls
->in_port
);
4870 pvector_remove(&cls
->subtables
, subtable
);
4871 cmap_remove(&cls
->subtables_map
, &subtable
->cmap_node
,
4872 subtable
->mask
.hash
);
4873 cmap_destroy(&subtable
->rules
);
4874 ovsrcu_postpone(free
, subtable
);
4877 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
4878 * caller's responsibility.
4879 * May only be called after all the readers have been terminated. */
4881 dpcls_destroy(struct dpcls
*cls
)
4884 struct dpcls_subtable
*subtable
;
4886 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
4887 ovs_assert(cmap_count(&subtable
->rules
) == 0);
4888 dpcls_destroy_subtable(cls
, subtable
);
4890 cmap_destroy(&cls
->subtables_map
);
4891 pvector_destroy(&cls
->subtables
);
4895 static struct dpcls_subtable
*
4896 dpcls_create_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
4898 struct dpcls_subtable
*subtable
;
4900 /* Need to add one. */
4901 subtable
= xmalloc(sizeof *subtable
4902 - sizeof subtable
->mask
.mf
+ mask
->len
);
4903 cmap_init(&subtable
->rules
);
4904 subtable
->hit_cnt
= 0;
4905 netdev_flow_key_clone(&subtable
->mask
, mask
);
4906 cmap_insert(&cls
->subtables_map
, &subtable
->cmap_node
, mask
->hash
);
4907 /* Add the new subtable at the end of the pvector (with no hits yet) */
4908 pvector_insert(&cls
->subtables
, subtable
, 0);
4909 VLOG_DBG("Creating %"PRIuSIZE
". subtable %p for in_port %d",
4910 cmap_count(&cls
->subtables_map
), subtable
, cls
->in_port
);
4911 pvector_publish(&cls
->subtables
);
4916 static inline struct dpcls_subtable
*
4917 dpcls_find_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
4919 struct dpcls_subtable
*subtable
;
4921 CMAP_FOR_EACH_WITH_HASH (subtable
, cmap_node
, mask
->hash
,
4922 &cls
->subtables_map
) {
4923 if (netdev_flow_key_equal(&subtable
->mask
, mask
)) {
4927 return dpcls_create_subtable(cls
, mask
);
4931 /* Periodically sort the dpcls subtable vectors according to hit counts */
4933 dpcls_sort_subtable_vector(struct dpcls
*cls
)
4935 struct pvector
*pvec
= &cls
->subtables
;
4936 struct dpcls_subtable
*subtable
;
4938 PVECTOR_FOR_EACH (subtable
, pvec
) {
4939 pvector_change_priority(pvec
, subtable
, subtable
->hit_cnt
);
4940 subtable
->hit_cnt
= 0;
4942 pvector_publish(pvec
);
4946 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
)
4949 long long int now
= time_msec();
4951 if (now
> pmd
->next_optimization
) {
4952 /* Try to obtain the flow lock to block out revalidator threads.
4953 * If not possible, just try next time. */
4954 if (!ovs_mutex_trylock(&pmd
->flow_mutex
)) {
4955 /* Optimize each classifier */
4956 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
4957 dpcls_sort_subtable_vector(cls
);
4959 ovs_mutex_unlock(&pmd
->flow_mutex
);
4960 /* Start new measuring interval */
4961 pmd
->next_optimization
= now
+ DPCLS_OPTIMIZATION_INTERVAL
;
4966 /* Insert 'rule' into 'cls'. */
4968 dpcls_insert(struct dpcls
*cls
, struct dpcls_rule
*rule
,
4969 const struct netdev_flow_key
*mask
)
4971 struct dpcls_subtable
*subtable
= dpcls_find_subtable(cls
, mask
);
4973 /* Refer to subtable's mask, also for later removal. */
4974 rule
->mask
= &subtable
->mask
;
4975 cmap_insert(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
);
4978 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
4980 dpcls_remove(struct dpcls
*cls
, struct dpcls_rule
*rule
)
4982 struct dpcls_subtable
*subtable
;
4984 ovs_assert(rule
->mask
);
4986 /* Get subtable from reference in rule->mask. */
4987 INIT_CONTAINER(subtable
, rule
->mask
, mask
);
4988 if (cmap_remove(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
)
4990 /* Delete empty subtable. */
4991 dpcls_destroy_subtable(cls
, subtable
);
4992 pvector_publish(&cls
->subtables
);
4996 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
4997 * in 'mask' the values in 'key' and 'target' are the same. */
4999 dpcls_rule_matches_key(const struct dpcls_rule
*rule
,
5000 const struct netdev_flow_key
*target
)
5002 const uint64_t *keyp
= miniflow_get_values(&rule
->flow
.mf
);
5003 const uint64_t *maskp
= miniflow_get_values(&rule
->mask
->mf
);
5006 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, target
, rule
->flow
.mf
.map
) {
5007 if (OVS_UNLIKELY((value
& *maskp
++) != *keyp
++)) {
5014 /* For each miniflow in 'keys' performs a classifier lookup writing the result
5015 * into the corresponding slot in 'rules'. If a particular entry in 'keys' is
5016 * NULL it is skipped.
5018 * This function is optimized for use in the userspace datapath and therefore
5019 * does not implement a lot of features available in the standard
5020 * classifier_lookup() function. Specifically, it does not implement
5021 * priorities, instead returning any rule which matches the flow.
5023 * Returns true if all miniflows found a corresponding rule. */
5025 dpcls_lookup(struct dpcls
*cls
, const struct netdev_flow_key keys
[],
5026 struct dpcls_rule
**rules
, const size_t cnt
,
5029 /* The received 'cnt' miniflows are the search-keys that will be processed
5030 * to find a matching entry into the available subtables.
5031 * The number of bits in map_type is equal to NETDEV_MAX_BURST. */
5032 typedef uint32_t map_type
;
5033 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
5034 BUILD_ASSERT_DECL(MAP_BITS
>= NETDEV_MAX_BURST
);
5036 struct dpcls_subtable
*subtable
;
5038 map_type keys_map
= TYPE_MAXIMUM(map_type
); /* Set all bits. */
5040 uint32_t hashes
[MAP_BITS
];
5041 const struct cmap_node
*nodes
[MAP_BITS
];
5043 if (cnt
!= MAP_BITS
) {
5044 keys_map
>>= MAP_BITS
- cnt
; /* Clear extra bits. */
5046 memset(rules
, 0, cnt
* sizeof *rules
);
5048 int lookups_match
= 0, subtable_pos
= 1;
5050 /* The Datapath classifier - aka dpcls - is composed of subtables.
5051 * Subtables are dynamically created as needed when new rules are inserted.
5052 * Each subtable collects rules with matches on a specific subset of packet
5053 * fields as defined by the subtable's mask. We proceed to process every
5054 * search-key against each subtable, but when a match is found for a
5055 * search-key, the search for that key can stop because the rules are
5056 * non-overlapping. */
5057 PVECTOR_FOR_EACH (subtable
, &cls
->subtables
) {
5060 /* Compute hashes for the remaining keys. Each search-key is
5061 * masked with the subtable's mask to avoid hashing the wildcarded
5063 ULLONG_FOR_EACH_1(i
, keys_map
) {
5064 hashes
[i
] = netdev_flow_key_hash_in_mask(&keys
[i
],
5068 found_map
= cmap_find_batch(&subtable
->rules
, keys_map
, hashes
, nodes
);
5069 /* Check results. When the i-th bit of found_map is set, it means
5070 * that a set of nodes with a matching hash value was found for the
5071 * i-th search-key. Due to possible hash collisions we need to check
5072 * which of the found rules, if any, really matches our masked
5074 ULLONG_FOR_EACH_1(i
, found_map
) {
5075 struct dpcls_rule
*rule
;
5077 CMAP_NODE_FOR_EACH (rule
, cmap_node
, nodes
[i
]) {
5078 if (OVS_LIKELY(dpcls_rule_matches_key(rule
, &keys
[i
]))) {
5080 /* Even at 20 Mpps the 32-bit hit_cnt cannot wrap
5081 * within one second optimization interval. */
5082 subtable
->hit_cnt
++;
5083 lookups_match
+= subtable_pos
;
5087 /* None of the found rules was a match. Reset the i-th bit to
5088 * keep searching this key in the next subtable. */
5089 ULLONG_SET0(found_map
, i
); /* Did not match. */
5091 ; /* Keep Sparse happy. */
5093 keys_map
&= ~found_map
; /* Clear the found rules. */
5095 if (num_lookups_p
) {
5096 *num_lookups_p
= lookups_match
;
5098 return true; /* All found. */
5102 if (num_lookups_p
) {
5103 *num_lookups_p
= lookups_match
;
5105 return false; /* Some misses. */