2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 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"
24 #include <netinet/in.h>
25 #include <sys/socket.h>
30 #include <sys/ioctl.h>
37 #include "dp-packet.h"
39 #include "dpif-provider.h"
41 #include "dynamic-string.h"
42 #include "fat-rwlock.h"
50 #include "netdev-dpdk.h"
51 #include "netdev-vport.h"
53 #include "odp-execute.h"
55 #include "ofp-print.h"
60 #include "poll-loop.h"
67 #include "tnl-arp-cache.h"
68 #include "tnl-ports.h"
71 #include "openvswitch/vlog.h"
73 VLOG_DEFINE_THIS_MODULE(dpif_netdev
);
75 #define FLOW_DUMP_MAX_BATCH 50
76 /* Use per thread recirc_depth to prevent recirculation loop. */
77 #define MAX_RECIRC_DEPTH 5
78 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth
, 0)
80 /* Configuration parameters. */
81 enum { MAX_FLOWS
= 65536 }; /* Maximum number of flows in flow table. */
83 /* Protects against changes to 'dp_netdevs'. */
84 static struct ovs_mutex dp_netdev_mutex
= OVS_MUTEX_INITIALIZER
;
86 /* Contains all 'struct dp_netdev's. */
87 static struct shash dp_netdevs
OVS_GUARDED_BY(dp_netdev_mutex
)
88 = SHASH_INITIALIZER(&dp_netdevs
);
90 static struct vlog_rate_limit upcall_rl
= VLOG_RATE_LIMIT_INIT(600, 600);
92 static struct odp_support dp_netdev_support
= {
93 .max_mpls_depth
= SIZE_MAX
,
97 /* Stores a miniflow with inline values */
99 struct netdev_flow_key
{
100 uint32_t hash
; /* Hash function differs for different users. */
101 uint32_t len
; /* Length of the following miniflow (incl. map). */
103 uint64_t buf
[FLOW_MAX_PACKET_U64S
];
106 /* Exact match cache for frequently used flows
108 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
109 * search its entries for a miniflow that matches exactly the miniflow of the
110 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
112 * A cache entry holds a reference to its 'dp_netdev_flow'.
114 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
115 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
116 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
117 * value is the index of a cache entry where the miniflow could be.
123 * Each pmd_thread has its own private exact match cache.
124 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
127 #define EM_FLOW_HASH_SHIFT 13
128 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
129 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
130 #define EM_FLOW_HASH_SEGS 2
133 struct dp_netdev_flow
*flow
;
134 struct netdev_flow_key key
; /* key.hash used for emc hash value. */
138 struct emc_entry entries
[EM_FLOW_HASH_ENTRIES
];
139 int sweep_idx
; /* For emc_cache_slow_sweep(). */
142 /* Iterate in the exact match cache through every entry that might contain a
143 * miniflow with hash 'HASH'. */
144 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
145 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
146 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
147 i__ < EM_FLOW_HASH_SEGS; \
148 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
150 /* Simple non-wildcarding single-priority classifier. */
153 struct cmap subtables_map
;
154 struct pvector subtables
;
157 /* A rule to be inserted to the classifier. */
159 struct cmap_node cmap_node
; /* Within struct dpcls_subtable 'rules'. */
160 struct netdev_flow_key
*mask
; /* Subtable's mask. */
161 struct netdev_flow_key flow
; /* Matching key. */
162 /* 'flow' must be the last field, additional space is allocated here. */
165 static void dpcls_init(struct dpcls
*);
166 static void dpcls_destroy(struct dpcls
*);
167 static void dpcls_insert(struct dpcls
*, struct dpcls_rule
*,
168 const struct netdev_flow_key
*mask
);
169 static void dpcls_remove(struct dpcls
*, struct dpcls_rule
*);
170 static bool dpcls_lookup(const struct dpcls
*cls
,
171 const struct netdev_flow_key keys
[],
172 struct dpcls_rule
**rules
, size_t cnt
);
174 /* Datapath based on the network device interface from netdev.h.
180 * Some members, marked 'const', are immutable. Accessing other members
181 * requires synchronization, as noted in more detail below.
183 * Acquisition order is, from outermost to innermost:
185 * dp_netdev_mutex (global)
189 const struct dpif_class
*const class;
190 const char *const name
;
192 struct ovs_refcount ref_cnt
;
193 atomic_flag destroyed
;
197 * Protected by RCU. Take the mutex to add or remove ports. */
198 struct ovs_mutex port_mutex
;
200 struct seq
*port_seq
; /* Incremented whenever a port changes. */
202 /* Protects access to ofproto-dpif-upcall interface during revalidator
203 * thread synchronization. */
204 struct fat_rwlock upcall_rwlock
;
205 upcall_callback
*upcall_cb
; /* Callback function for executing upcalls. */
208 /* Callback function for notifying the purging of dp flows (during
209 * reseting pmd deletion). */
210 dp_purge_callback
*dp_purge_cb
;
213 /* Stores all 'struct dp_netdev_pmd_thread's. */
214 struct cmap poll_threads
;
216 /* Protects the access of the 'struct dp_netdev_pmd_thread'
217 * instance for non-pmd thread. */
218 struct ovs_mutex non_pmd_mutex
;
220 /* Each pmd thread will store its pointer to
221 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
222 ovsthread_key_t per_pmd_key
;
224 /* Number of rx queues for each dpdk interface and the cpu mask
225 * for pin of pmd threads. */
228 uint64_t last_tnl_conf_seq
;
231 static struct dp_netdev_port
*dp_netdev_lookup_port(const struct dp_netdev
*dp
,
235 DP_STAT_EXACT_HIT
, /* Packets that had an exact match (emc). */
236 DP_STAT_MASKED_HIT
, /* Packets that matched in the flow table. */
237 DP_STAT_MISS
, /* Packets that did not match. */
238 DP_STAT_LOST
, /* Packets not passed up to the client. */
242 enum pmd_cycles_counter_type
{
243 PMD_CYCLES_POLLING
, /* Cycles spent polling NICs. */
244 PMD_CYCLES_PROCESSING
, /* Cycles spent processing packets */
248 /* A port in a netdev-based datapath. */
249 struct dp_netdev_port
{
251 struct netdev
*netdev
;
252 struct cmap_node node
; /* Node in dp_netdev's 'ports'. */
253 struct netdev_saved_flags
*sf
;
254 struct netdev_rxq
**rxq
;
255 struct ovs_refcount ref_cnt
;
256 char *type
; /* Port type as requested by user. */
259 /* Contained by struct dp_netdev_flow's 'stats' member. */
260 struct dp_netdev_flow_stats
{
261 atomic_llong used
; /* Last used time, in monotonic msecs. */
262 atomic_ullong packet_count
; /* Number of packets matched. */
263 atomic_ullong byte_count
; /* Number of bytes matched. */
264 atomic_uint16_t tcp_flags
; /* Bitwise-OR of seen tcp_flags values. */
267 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
273 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
274 * its pmd thread's classifier. The text below calls this classifier 'cls'.
279 * The thread safety rules described here for "struct dp_netdev_flow" are
280 * motivated by two goals:
282 * - Prevent threads that read members of "struct dp_netdev_flow" from
283 * reading bad data due to changes by some thread concurrently modifying
286 * - Prevent two threads making changes to members of a given "struct
287 * dp_netdev_flow" from interfering with each other.
293 * A flow 'flow' may be accessed without a risk of being freed during an RCU
294 * grace period. Code that needs to hold onto a flow for a while
295 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
297 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
298 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
301 * Some members, marked 'const', are immutable. Accessing other members
302 * requires synchronization, as noted in more detail below.
304 struct dp_netdev_flow
{
305 const struct flow flow
; /* Unmasked flow that created this entry. */
306 /* Hash table index by unmasked flow. */
307 const struct cmap_node node
; /* In owning dp_netdev_pmd_thread's */
309 const ovs_u128 ufid
; /* Unique flow identifier. */
310 const unsigned pmd_id
; /* The 'core_id' of pmd thread owning this */
313 /* Number of references.
314 * The classifier owns one reference.
315 * Any thread trying to keep a rule from being freed should hold its own
317 struct ovs_refcount ref_cnt
;
322 struct dp_netdev_flow_stats stats
;
325 OVSRCU_TYPE(struct dp_netdev_actions
*) actions
;
327 /* While processing a group of input packets, the datapath uses the next
328 * member to store a pointer to the output batch for the flow. It is
329 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
330 * packet_batch_init() and packet_batch_execute()). */
331 struct packet_batch
*batch
;
333 /* Packet classification. */
334 struct dpcls_rule cr
; /* In owning dp_netdev's 'cls'. */
335 /* 'cr' must be the last member. */
338 static void dp_netdev_flow_unref(struct dp_netdev_flow
*);
339 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*);
340 static int dpif_netdev_flow_from_nlattrs(const struct nlattr
*, uint32_t,
343 /* A set of datapath actions within a "struct dp_netdev_flow".
349 * A struct dp_netdev_actions 'actions' is protected with RCU. */
350 struct dp_netdev_actions
{
351 /* These members are immutable: they do not change during the struct's
353 unsigned int size
; /* Size of 'actions', in bytes. */
354 struct nlattr actions
[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
357 struct dp_netdev_actions
*dp_netdev_actions_create(const struct nlattr
*,
359 struct dp_netdev_actions
*dp_netdev_flow_get_actions(
360 const struct dp_netdev_flow
*);
361 static void dp_netdev_actions_free(struct dp_netdev_actions
*);
363 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
364 struct dp_netdev_pmd_stats
{
365 /* Indexed by DP_STAT_*. */
366 atomic_ullong n
[DP_N_STATS
];
369 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
370 struct dp_netdev_pmd_cycles
{
371 /* Indexed by PMD_CYCLES_*. */
372 atomic_ullong n
[PMD_N_CYCLES
];
375 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
376 * the performance overhead of interrupt processing. Therefore netdev can
377 * not implement rx-wait for these devices. dpif-netdev needs to poll
378 * these device to check for recv buffer. pmd-thread does polling for
379 * devices assigned to itself.
381 * DPDK used PMD for accessing NIC.
383 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
384 * I/O of all non-pmd threads. There will be no actual thread created
387 * Each struct has its own flow table and classifier. Packets received
388 * from managed ports are looked up in the corresponding pmd thread's
389 * flow table, and are executed with the found actions.
391 struct dp_netdev_pmd_thread
{
392 struct dp_netdev
*dp
;
393 struct ovs_refcount ref_cnt
; /* Every reference must be refcount'ed. */
394 struct cmap_node node
; /* In 'dp->poll_threads'. */
396 pthread_cond_t cond
; /* For synchronizing pmd thread reload. */
397 struct ovs_mutex cond_mutex
; /* Mutex for condition variable. */
399 /* Per thread exact-match cache. Note, the instance for cpu core
400 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
401 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
402 * instances will only be accessed by its own pmd thread. */
403 struct emc_cache flow_cache
;
405 /* Classifier and Flow-Table.
407 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
408 * changes to 'cls' must be made while still holding the 'flow_mutex'.
410 struct ovs_mutex flow_mutex
;
412 struct cmap flow_table OVS_GUARDED
; /* Flow table. */
415 struct dp_netdev_pmd_stats stats
;
417 /* Cycles counters */
418 struct dp_netdev_pmd_cycles cycles
;
420 /* Used to count cicles. See 'cycles_counter_end()' */
421 unsigned long long last_cycles
;
423 struct latch exit_latch
; /* For terminating the pmd thread. */
424 atomic_uint change_seq
; /* For reloading pmd ports. */
426 int index
; /* Idx of this pmd thread among pmd*/
427 /* threads on same numa node. */
428 unsigned core_id
; /* CPU core id of this pmd thread. */
429 int numa_id
; /* numa node id of this pmd thread. */
430 int tx_qid
; /* Queue id used by this pmd thread to
431 * send packets on all netdevs */
433 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
434 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
435 * values and subtracts them from 'stats' and 'cycles' before
436 * reporting to the user */
437 unsigned long long stats_zero
[DP_N_STATS
];
438 uint64_t cycles_zero
[PMD_N_CYCLES
];
441 #define PMD_INITIAL_SEQ 1
443 /* Interface to netdev-based datapath. */
446 struct dp_netdev
*dp
;
447 uint64_t last_port_seq
;
450 static int get_port_by_number(struct dp_netdev
*dp
, odp_port_t port_no
,
451 struct dp_netdev_port
**portp
);
452 static int get_port_by_name(struct dp_netdev
*dp
, const char *devname
,
453 struct dp_netdev_port
**portp
);
454 static void dp_netdev_free(struct dp_netdev
*)
455 OVS_REQUIRES(dp_netdev_mutex
);
456 static int do_add_port(struct dp_netdev
*dp
, const char *devname
,
457 const char *type
, odp_port_t port_no
)
458 OVS_REQUIRES(dp
->port_mutex
);
459 static void do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*)
460 OVS_REQUIRES(dp
->port_mutex
);
461 static int dpif_netdev_open(const struct dpif_class
*, const char *name
,
462 bool create
, struct dpif
**);
463 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
464 struct dp_packet
**, int c
,
466 const struct nlattr
*actions
,
468 static void dp_netdev_input(struct dp_netdev_pmd_thread
*,
469 struct dp_packet
**, int cnt
);
471 static void dp_netdev_disable_upcall(struct dp_netdev
*);
472 void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
);
473 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
,
474 struct dp_netdev
*dp
, int index
,
475 unsigned core_id
, int numa_id
);
476 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
);
477 static void dp_netdev_set_nonpmd(struct dp_netdev
*dp
);
478 static struct dp_netdev_pmd_thread
*dp_netdev_get_pmd(struct dp_netdev
*dp
,
480 static struct dp_netdev_pmd_thread
*
481 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
);
482 static void dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
);
483 static void dp_netdev_del_pmds_on_numa(struct dp_netdev
*dp
, int numa_id
);
484 static void dp_netdev_set_pmds_on_numa(struct dp_netdev
*dp
, int numa_id
);
485 static void dp_netdev_reset_pmd_threads(struct dp_netdev
*dp
);
486 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
);
487 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
);
488 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
);
490 static inline bool emc_entry_alive(struct emc_entry
*ce
);
491 static void emc_clear_entry(struct emc_entry
*ce
);
494 emc_cache_init(struct emc_cache
*flow_cache
)
498 flow_cache
->sweep_idx
= 0;
499 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
500 flow_cache
->entries
[i
].flow
= NULL
;
501 flow_cache
->entries
[i
].key
.hash
= 0;
502 flow_cache
->entries
[i
].key
.len
= sizeof(struct miniflow
);
503 flowmap_init(&flow_cache
->entries
[i
].key
.mf
.map
);
508 emc_cache_uninit(struct emc_cache
*flow_cache
)
512 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
513 emc_clear_entry(&flow_cache
->entries
[i
]);
517 /* Check and clear dead flow references slowly (one entry at each
520 emc_cache_slow_sweep(struct emc_cache
*flow_cache
)
522 struct emc_entry
*entry
= &flow_cache
->entries
[flow_cache
->sweep_idx
];
524 if (!emc_entry_alive(entry
)) {
525 emc_clear_entry(entry
);
527 flow_cache
->sweep_idx
= (flow_cache
->sweep_idx
+ 1) & EM_FLOW_HASH_MASK
;
530 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
532 dpif_is_netdev(const struct dpif
*dpif
)
534 return dpif
->dpif_class
->open
== dpif_netdev_open
;
537 static struct dpif_netdev
*
538 dpif_netdev_cast(const struct dpif
*dpif
)
540 ovs_assert(dpif_is_netdev(dpif
));
541 return CONTAINER_OF(dpif
, struct dpif_netdev
, dpif
);
544 static struct dp_netdev
*
545 get_dp_netdev(const struct dpif
*dpif
)
547 return dpif_netdev_cast(dpif
)->dp
;
551 PMD_INFO_SHOW_STATS
, /* show how cpu cycles are spent */
552 PMD_INFO_CLEAR_STATS
/* set the cycles count to 0 */
556 pmd_info_show_stats(struct ds
*reply
,
557 struct dp_netdev_pmd_thread
*pmd
,
558 unsigned long long stats
[DP_N_STATS
],
559 uint64_t cycles
[PMD_N_CYCLES
])
561 unsigned long long total_packets
= 0;
562 uint64_t total_cycles
= 0;
565 /* These loops subtracts reference values ('*_zero') from the counters.
566 * Since loads and stores are relaxed, it might be possible for a '*_zero'
567 * value to be more recent than the current value we're reading from the
568 * counter. This is not a big problem, since these numbers are not
569 * supposed to be too accurate, but we should at least make sure that
570 * the result is not negative. */
571 for (i
= 0; i
< DP_N_STATS
; i
++) {
572 if (stats
[i
] > pmd
->stats_zero
[i
]) {
573 stats
[i
] -= pmd
->stats_zero
[i
];
578 if (i
!= DP_STAT_LOST
) {
579 /* Lost packets are already included in DP_STAT_MISS */
580 total_packets
+= stats
[i
];
584 for (i
= 0; i
< PMD_N_CYCLES
; i
++) {
585 if (cycles
[i
] > pmd
->cycles_zero
[i
]) {
586 cycles
[i
] -= pmd
->cycles_zero
[i
];
591 total_cycles
+= cycles
[i
];
594 ds_put_cstr(reply
, (pmd
->core_id
== NON_PMD_CORE_ID
)
595 ? "main thread" : "pmd thread");
597 if (pmd
->numa_id
!= OVS_NUMA_UNSPEC
) {
598 ds_put_format(reply
, " numa_id %d", pmd
->numa_id
);
600 if (pmd
->core_id
!= OVS_CORE_UNSPEC
&& pmd
->core_id
!= NON_PMD_CORE_ID
) {
601 ds_put_format(reply
, " core_id %u", pmd
->core_id
);
603 ds_put_cstr(reply
, ":\n");
606 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
607 "\tmiss:%llu\n\tlost:%llu\n",
608 stats
[DP_STAT_EXACT_HIT
], stats
[DP_STAT_MASKED_HIT
],
609 stats
[DP_STAT_MISS
], stats
[DP_STAT_LOST
]);
611 if (total_cycles
== 0) {
616 "\tpolling cycles:%"PRIu64
" (%.02f%%)\n"
617 "\tprocessing cycles:%"PRIu64
" (%.02f%%)\n",
618 cycles
[PMD_CYCLES_POLLING
],
619 cycles
[PMD_CYCLES_POLLING
] / (double)total_cycles
* 100,
620 cycles
[PMD_CYCLES_PROCESSING
],
621 cycles
[PMD_CYCLES_PROCESSING
] / (double)total_cycles
* 100);
623 if (total_packets
== 0) {
628 "\tavg cycles per packet: %.02f (%"PRIu64
"/%llu)\n",
629 total_cycles
/ (double)total_packets
,
630 total_cycles
, total_packets
);
633 "\tavg processing cycles per packet: "
634 "%.02f (%"PRIu64
"/%llu)\n",
635 cycles
[PMD_CYCLES_PROCESSING
] / (double)total_packets
,
636 cycles
[PMD_CYCLES_PROCESSING
], total_packets
);
640 pmd_info_clear_stats(struct ds
*reply OVS_UNUSED
,
641 struct dp_netdev_pmd_thread
*pmd
,
642 unsigned long long stats
[DP_N_STATS
],
643 uint64_t cycles
[PMD_N_CYCLES
])
647 /* We cannot write 'stats' and 'cycles' (because they're written by other
648 * threads) and we shouldn't change 'stats' (because they're used to count
649 * datapath stats, which must not be cleared here). Instead, we save the
650 * current values and subtract them from the values to be displayed in the
652 for (i
= 0; i
< DP_N_STATS
; i
++) {
653 pmd
->stats_zero
[i
] = stats
[i
];
655 for (i
= 0; i
< PMD_N_CYCLES
; i
++) {
656 pmd
->cycles_zero
[i
] = cycles
[i
];
661 dpif_netdev_pmd_info(struct unixctl_conn
*conn
, int argc
, const char *argv
[],
664 struct ds reply
= DS_EMPTY_INITIALIZER
;
665 struct dp_netdev_pmd_thread
*pmd
;
666 struct dp_netdev
*dp
= NULL
;
667 enum pmd_info_type type
= *(enum pmd_info_type
*) aux
;
669 ovs_mutex_lock(&dp_netdev_mutex
);
672 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
673 } else if (shash_count(&dp_netdevs
) == 1) {
674 /* There's only one datapath */
675 dp
= shash_first(&dp_netdevs
)->data
;
679 ovs_mutex_unlock(&dp_netdev_mutex
);
680 unixctl_command_reply_error(conn
,
681 "please specify an existing datapath");
685 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
686 unsigned long long stats
[DP_N_STATS
];
687 uint64_t cycles
[PMD_N_CYCLES
];
690 /* Read current stats and cycle counters */
691 for (i
= 0; i
< ARRAY_SIZE(stats
); i
++) {
692 atomic_read_relaxed(&pmd
->stats
.n
[i
], &stats
[i
]);
694 for (i
= 0; i
< ARRAY_SIZE(cycles
); i
++) {
695 atomic_read_relaxed(&pmd
->cycles
.n
[i
], &cycles
[i
]);
698 if (type
== PMD_INFO_CLEAR_STATS
) {
699 pmd_info_clear_stats(&reply
, pmd
, stats
, cycles
);
700 } else if (type
== PMD_INFO_SHOW_STATS
) {
701 pmd_info_show_stats(&reply
, pmd
, stats
, cycles
);
705 ovs_mutex_unlock(&dp_netdev_mutex
);
707 unixctl_command_reply(conn
, ds_cstr(&reply
));
712 dpif_netdev_init(void)
714 static enum pmd_info_type show_aux
= PMD_INFO_SHOW_STATS
,
715 clear_aux
= PMD_INFO_CLEAR_STATS
;
717 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
718 0, 1, dpif_netdev_pmd_info
,
720 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
721 0, 1, dpif_netdev_pmd_info
,
727 dpif_netdev_enumerate(struct sset
*all_dps
,
728 const struct dpif_class
*dpif_class
)
730 struct shash_node
*node
;
732 ovs_mutex_lock(&dp_netdev_mutex
);
733 SHASH_FOR_EACH(node
, &dp_netdevs
) {
734 struct dp_netdev
*dp
= node
->data
;
735 if (dpif_class
!= dp
->class) {
736 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
737 * If the class doesn't match, skip this dpif. */
740 sset_add(all_dps
, node
->name
);
742 ovs_mutex_unlock(&dp_netdev_mutex
);
748 dpif_netdev_class_is_dummy(const struct dpif_class
*class)
750 return class != &dpif_netdev_class
;
754 dpif_netdev_port_open_type(const struct dpif_class
*class, const char *type
)
756 return strcmp(type
, "internal") ? type
757 : dpif_netdev_class_is_dummy(class) ? "dummy"
762 create_dpif_netdev(struct dp_netdev
*dp
)
764 uint16_t netflow_id
= hash_string(dp
->name
, 0);
765 struct dpif_netdev
*dpif
;
767 ovs_refcount_ref(&dp
->ref_cnt
);
769 dpif
= xmalloc(sizeof *dpif
);
770 dpif_init(&dpif
->dpif
, dp
->class, dp
->name
, netflow_id
>> 8, netflow_id
);
772 dpif
->last_port_seq
= seq_read(dp
->port_seq
);
777 /* Choose an unused, non-zero port number and return it on success.
778 * Return ODPP_NONE on failure. */
780 choose_port(struct dp_netdev
*dp
, const char *name
)
781 OVS_REQUIRES(dp
->port_mutex
)
785 if (dp
->class != &dpif_netdev_class
) {
789 /* If the port name begins with "br", start the number search at
790 * 100 to make writing tests easier. */
791 if (!strncmp(name
, "br", 2)) {
795 /* If the port name contains a number, try to assign that port number.
796 * This can make writing unit tests easier because port numbers are
798 for (p
= name
; *p
!= '\0'; p
++) {
799 if (isdigit((unsigned char) *p
)) {
800 port_no
= start_no
+ strtol(p
, NULL
, 10);
801 if (port_no
> 0 && port_no
!= odp_to_u32(ODPP_NONE
)
802 && !dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
803 return u32_to_odp(port_no
);
810 for (port_no
= 1; port_no
<= UINT16_MAX
; port_no
++) {
811 if (!dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
812 return u32_to_odp(port_no
);
820 create_dp_netdev(const char *name
, const struct dpif_class
*class,
821 struct dp_netdev
**dpp
)
822 OVS_REQUIRES(dp_netdev_mutex
)
824 struct dp_netdev
*dp
;
827 dp
= xzalloc(sizeof *dp
);
828 shash_add(&dp_netdevs
, name
, dp
);
830 *CONST_CAST(const struct dpif_class
**, &dp
->class) = class;
831 *CONST_CAST(const char **, &dp
->name
) = xstrdup(name
);
832 ovs_refcount_init(&dp
->ref_cnt
);
833 atomic_flag_clear(&dp
->destroyed
);
835 ovs_mutex_init(&dp
->port_mutex
);
836 cmap_init(&dp
->ports
);
837 dp
->port_seq
= seq_create();
838 fat_rwlock_init(&dp
->upcall_rwlock
);
840 /* Disable upcalls by default. */
841 dp_netdev_disable_upcall(dp
);
842 dp
->upcall_aux
= NULL
;
843 dp
->upcall_cb
= NULL
;
845 cmap_init(&dp
->poll_threads
);
846 ovs_mutex_init_recursive(&dp
->non_pmd_mutex
);
847 ovsthread_key_create(&dp
->per_pmd_key
, NULL
);
849 dp_netdev_set_nonpmd(dp
);
850 dp
->n_dpdk_rxqs
= NR_QUEUE
;
852 ovs_mutex_lock(&dp
->port_mutex
);
853 error
= do_add_port(dp
, name
, "internal", ODPP_LOCAL
);
854 ovs_mutex_unlock(&dp
->port_mutex
);
860 dp
->last_tnl_conf_seq
= seq_read(tnl_conf_seq
);
866 dpif_netdev_open(const struct dpif_class
*class, const char *name
,
867 bool create
, struct dpif
**dpifp
)
869 struct dp_netdev
*dp
;
872 ovs_mutex_lock(&dp_netdev_mutex
);
873 dp
= shash_find_data(&dp_netdevs
, name
);
875 error
= create
? create_dp_netdev(name
, class, &dp
) : ENODEV
;
877 error
= (dp
->class != class ? EINVAL
882 *dpifp
= create_dpif_netdev(dp
);
885 ovs_mutex_unlock(&dp_netdev_mutex
);
891 dp_netdev_destroy_upcall_lock(struct dp_netdev
*dp
)
892 OVS_NO_THREAD_SAFETY_ANALYSIS
894 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
895 ovs_assert(fat_rwlock_tryrdlock(&dp
->upcall_rwlock
));
897 /* Before freeing a lock we should release it */
898 fat_rwlock_unlock(&dp
->upcall_rwlock
);
899 fat_rwlock_destroy(&dp
->upcall_rwlock
);
902 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
903 * through the 'dp_netdevs' shash while freeing 'dp'. */
905 dp_netdev_free(struct dp_netdev
*dp
)
906 OVS_REQUIRES(dp_netdev_mutex
)
908 struct dp_netdev_port
*port
;
910 shash_find_and_delete(&dp_netdevs
, dp
->name
);
912 dp_netdev_destroy_all_pmds(dp
);
913 cmap_destroy(&dp
->poll_threads
);
914 ovs_mutex_destroy(&dp
->non_pmd_mutex
);
915 ovsthread_key_delete(dp
->per_pmd_key
);
917 ovs_mutex_lock(&dp
->port_mutex
);
918 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
919 do_del_port(dp
, port
);
921 ovs_mutex_unlock(&dp
->port_mutex
);
923 seq_destroy(dp
->port_seq
);
924 cmap_destroy(&dp
->ports
);
926 /* Upcalls must be disabled at this point */
927 dp_netdev_destroy_upcall_lock(dp
);
930 free(CONST_CAST(char *, dp
->name
));
935 dp_netdev_unref(struct dp_netdev
*dp
)
938 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
939 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
940 ovs_mutex_lock(&dp_netdev_mutex
);
941 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
944 ovs_mutex_unlock(&dp_netdev_mutex
);
949 dpif_netdev_close(struct dpif
*dpif
)
951 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
958 dpif_netdev_destroy(struct dpif
*dpif
)
960 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
962 if (!atomic_flag_test_and_set(&dp
->destroyed
)) {
963 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
964 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
972 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
973 * load/store semantics. While the increment is not atomic, the load and
974 * store operations are, making it impossible to read inconsistent values.
976 * This is used to update thread local stats counters. */
978 non_atomic_ullong_add(atomic_ullong
*var
, unsigned long long n
)
980 unsigned long long tmp
;
982 atomic_read_relaxed(var
, &tmp
);
984 atomic_store_relaxed(var
, tmp
);
988 dpif_netdev_get_stats(const struct dpif
*dpif
, struct dpif_dp_stats
*stats
)
990 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
991 struct dp_netdev_pmd_thread
*pmd
;
993 stats
->n_flows
= stats
->n_hit
= stats
->n_missed
= stats
->n_lost
= 0;
994 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
995 unsigned long long n
;
996 stats
->n_flows
+= cmap_count(&pmd
->flow_table
);
998 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_MASKED_HIT
], &n
);
1000 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_EXACT_HIT
], &n
);
1002 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_MISS
], &n
);
1003 stats
->n_missed
+= n
;
1004 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_LOST
], &n
);
1007 stats
->n_masks
= UINT32_MAX
;
1008 stats
->n_mask_hit
= UINT64_MAX
;
1014 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread
*pmd
)
1018 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
1022 ovs_mutex_lock(&pmd
->cond_mutex
);
1023 atomic_add_relaxed(&pmd
->change_seq
, 1, &old_seq
);
1024 ovs_mutex_cond_wait(&pmd
->cond
, &pmd
->cond_mutex
);
1025 ovs_mutex_unlock(&pmd
->cond_mutex
);
1028 /* Causes all pmd threads to reload its tx/rx devices.
1029 * Must be called after adding/removing ports. */
1031 dp_netdev_reload_pmds(struct dp_netdev
*dp
)
1033 struct dp_netdev_pmd_thread
*pmd
;
1035 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1036 dp_netdev_reload_pmd__(pmd
);
1041 hash_port_no(odp_port_t port_no
)
1043 return hash_int(odp_to_u32(port_no
), 0);
1047 do_add_port(struct dp_netdev
*dp
, const char *devname
, const char *type
,
1049 OVS_REQUIRES(dp
->port_mutex
)
1051 struct netdev_saved_flags
*sf
;
1052 struct dp_netdev_port
*port
;
1053 struct netdev
*netdev
;
1054 enum netdev_flags flags
;
1055 const char *open_type
;
1059 /* Reject devices already in 'dp'. */
1060 if (!get_port_by_name(dp
, devname
, &port
)) {
1064 /* Open and validate network device. */
1065 open_type
= dpif_netdev_port_open_type(dp
->class, type
);
1066 error
= netdev_open(devname
, open_type
, &netdev
);
1070 /* XXX reject non-Ethernet devices */
1072 netdev_get_flags(netdev
, &flags
);
1073 if (flags
& NETDEV_LOOPBACK
) {
1074 VLOG_ERR("%s: cannot add a loopback device", devname
);
1075 netdev_close(netdev
);
1079 if (netdev_is_pmd(netdev
)) {
1080 int n_cores
= ovs_numa_get_n_cores();
1082 if (n_cores
== OVS_CORE_UNSPEC
) {
1083 VLOG_ERR("%s, cannot get cpu core info", devname
);
1086 /* There can only be ovs_numa_get_n_cores() pmd threads,
1087 * so creates a txq for each, and one extra for the non
1089 error
= netdev_set_multiq(netdev
, n_cores
+ 1, dp
->n_dpdk_rxqs
);
1090 if (error
&& (error
!= EOPNOTSUPP
)) {
1091 VLOG_ERR("%s, cannot set multiq", devname
);
1095 port
= xzalloc(sizeof *port
);
1096 port
->port_no
= port_no
;
1097 port
->netdev
= netdev
;
1098 port
->rxq
= xmalloc(sizeof *port
->rxq
* netdev_n_rxq(netdev
));
1099 port
->type
= xstrdup(type
);
1100 for (i
= 0; i
< netdev_n_rxq(netdev
); i
++) {
1101 error
= netdev_rxq_open(netdev
, &port
->rxq
[i
], i
);
1103 && !(error
== EOPNOTSUPP
&& dpif_netdev_class_is_dummy(dp
->class))) {
1104 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1105 devname
, ovs_strerror(errno
));
1106 netdev_close(netdev
);
1114 error
= netdev_turn_flags_on(netdev
, NETDEV_PROMISC
, &sf
);
1116 for (i
= 0; i
< netdev_n_rxq(netdev
); i
++) {
1117 netdev_rxq_close(port
->rxq
[i
]);
1119 netdev_close(netdev
);
1127 ovs_refcount_init(&port
->ref_cnt
);
1128 cmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
1130 if (netdev_is_pmd(netdev
)) {
1131 dp_netdev_set_pmds_on_numa(dp
, netdev_get_numa_id(netdev
));
1132 dp_netdev_reload_pmds(dp
);
1134 seq_change(dp
->port_seq
);
1140 dpif_netdev_port_add(struct dpif
*dpif
, struct netdev
*netdev
,
1141 odp_port_t
*port_nop
)
1143 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1144 char namebuf
[NETDEV_VPORT_NAME_BUFSIZE
];
1145 const char *dpif_port
;
1149 ovs_mutex_lock(&dp
->port_mutex
);
1150 dpif_port
= netdev_vport_get_dpif_port(netdev
, namebuf
, sizeof namebuf
);
1151 if (*port_nop
!= ODPP_NONE
) {
1152 port_no
= *port_nop
;
1153 error
= dp_netdev_lookup_port(dp
, *port_nop
) ? EBUSY
: 0;
1155 port_no
= choose_port(dp
, dpif_port
);
1156 error
= port_no
== ODPP_NONE
? EFBIG
: 0;
1159 *port_nop
= port_no
;
1160 error
= do_add_port(dp
, dpif_port
, netdev_get_type(netdev
), port_no
);
1162 ovs_mutex_unlock(&dp
->port_mutex
);
1168 dpif_netdev_port_del(struct dpif
*dpif
, odp_port_t port_no
)
1170 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1173 ovs_mutex_lock(&dp
->port_mutex
);
1174 if (port_no
== ODPP_LOCAL
) {
1177 struct dp_netdev_port
*port
;
1179 error
= get_port_by_number(dp
, port_no
, &port
);
1181 do_del_port(dp
, port
);
1184 ovs_mutex_unlock(&dp
->port_mutex
);
1190 is_valid_port_number(odp_port_t port_no
)
1192 return port_no
!= ODPP_NONE
;
1195 static struct dp_netdev_port
*
1196 dp_netdev_lookup_port(const struct dp_netdev
*dp
, odp_port_t port_no
)
1198 struct dp_netdev_port
*port
;
1200 CMAP_FOR_EACH_WITH_HASH (port
, node
, hash_port_no(port_no
), &dp
->ports
) {
1201 if (port
->port_no
== port_no
) {
1209 get_port_by_number(struct dp_netdev
*dp
,
1210 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1212 if (!is_valid_port_number(port_no
)) {
1216 *portp
= dp_netdev_lookup_port(dp
, port_no
);
1217 return *portp
? 0 : ENOENT
;
1222 port_ref(struct dp_netdev_port
*port
)
1225 ovs_refcount_ref(&port
->ref_cnt
);
1230 port_try_ref(struct dp_netdev_port
*port
)
1233 return ovs_refcount_try_ref_rcu(&port
->ref_cnt
);
1240 port_unref(struct dp_netdev_port
*port
)
1242 if (port
&& ovs_refcount_unref_relaxed(&port
->ref_cnt
) == 1) {
1243 int n_rxq
= netdev_n_rxq(port
->netdev
);
1246 netdev_close(port
->netdev
);
1247 netdev_restore_flags(port
->sf
);
1249 for (i
= 0; i
< n_rxq
; i
++) {
1250 netdev_rxq_close(port
->rxq
[i
]);
1259 get_port_by_name(struct dp_netdev
*dp
,
1260 const char *devname
, struct dp_netdev_port
**portp
)
1261 OVS_REQUIRES(dp
->port_mutex
)
1263 struct dp_netdev_port
*port
;
1265 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1266 if (!strcmp(netdev_get_name(port
->netdev
), devname
)) {
1275 get_n_pmd_threads_on_numa(struct dp_netdev
*dp
, int numa_id
)
1277 struct dp_netdev_pmd_thread
*pmd
;
1280 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1281 if (pmd
->numa_id
== numa_id
) {
1289 /* Returns 'true' if there is a port with pmd netdev and the netdev
1290 * is on numa node 'numa_id'. */
1292 has_pmd_port_for_numa(struct dp_netdev
*dp
, int numa_id
)
1294 struct dp_netdev_port
*port
;
1296 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1297 if (netdev_is_pmd(port
->netdev
)
1298 && netdev_get_numa_id(port
->netdev
) == numa_id
) {
1308 do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*port
)
1309 OVS_REQUIRES(dp
->port_mutex
)
1311 cmap_remove(&dp
->ports
, &port
->node
, hash_odp_port(port
->port_no
));
1312 seq_change(dp
->port_seq
);
1313 if (netdev_is_pmd(port
->netdev
)) {
1314 int numa_id
= netdev_get_numa_id(port
->netdev
);
1316 /* If there is no netdev on the numa node, deletes the pmd threads
1317 * for that numa. Else, just reloads the queues. */
1318 if (!has_pmd_port_for_numa(dp
, numa_id
)) {
1319 dp_netdev_del_pmds_on_numa(dp
, numa_id
);
1321 dp_netdev_reload_pmds(dp
);
1328 answer_port_query(const struct dp_netdev_port
*port
,
1329 struct dpif_port
*dpif_port
)
1331 dpif_port
->name
= xstrdup(netdev_get_name(port
->netdev
));
1332 dpif_port
->type
= xstrdup(port
->type
);
1333 dpif_port
->port_no
= port
->port_no
;
1337 dpif_netdev_port_query_by_number(const struct dpif
*dpif
, odp_port_t port_no
,
1338 struct dpif_port
*dpif_port
)
1340 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1341 struct dp_netdev_port
*port
;
1344 error
= get_port_by_number(dp
, port_no
, &port
);
1345 if (!error
&& dpif_port
) {
1346 answer_port_query(port
, dpif_port
);
1353 dpif_netdev_port_query_by_name(const struct dpif
*dpif
, const char *devname
,
1354 struct dpif_port
*dpif_port
)
1356 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1357 struct dp_netdev_port
*port
;
1360 ovs_mutex_lock(&dp
->port_mutex
);
1361 error
= get_port_by_name(dp
, devname
, &port
);
1362 if (!error
&& dpif_port
) {
1363 answer_port_query(port
, dpif_port
);
1365 ovs_mutex_unlock(&dp
->port_mutex
);
1371 dp_netdev_flow_free(struct dp_netdev_flow
*flow
)
1373 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow
));
1377 static void dp_netdev_flow_unref(struct dp_netdev_flow
*flow
)
1379 if (ovs_refcount_unref_relaxed(&flow
->ref_cnt
) == 1) {
1380 ovsrcu_postpone(dp_netdev_flow_free
, flow
);
1385 dp_netdev_flow_hash(const ovs_u128
*ufid
)
1387 return ufid
->u32
[0];
1391 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread
*pmd
,
1392 struct dp_netdev_flow
*flow
)
1393 OVS_REQUIRES(pmd
->flow_mutex
)
1395 struct cmap_node
*node
= CONST_CAST(struct cmap_node
*, &flow
->node
);
1397 dpcls_remove(&pmd
->cls
, &flow
->cr
);
1398 flow
->cr
.mask
= NULL
; /* Accessing rule's mask after this is not safe. */
1400 cmap_remove(&pmd
->flow_table
, node
, dp_netdev_flow_hash(&flow
->ufid
));
1403 dp_netdev_flow_unref(flow
);
1407 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
)
1409 struct dp_netdev_flow
*netdev_flow
;
1411 ovs_mutex_lock(&pmd
->flow_mutex
);
1412 CMAP_FOR_EACH (netdev_flow
, node
, &pmd
->flow_table
) {
1413 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
1415 ovs_mutex_unlock(&pmd
->flow_mutex
);
1419 dpif_netdev_flow_flush(struct dpif
*dpif
)
1421 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1422 struct dp_netdev_pmd_thread
*pmd
;
1424 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1425 dp_netdev_pmd_flow_flush(pmd
);
1431 struct dp_netdev_port_state
{
1432 struct cmap_position position
;
1437 dpif_netdev_port_dump_start(const struct dpif
*dpif OVS_UNUSED
, void **statep
)
1439 *statep
= xzalloc(sizeof(struct dp_netdev_port_state
));
1444 dpif_netdev_port_dump_next(const struct dpif
*dpif
, void *state_
,
1445 struct dpif_port
*dpif_port
)
1447 struct dp_netdev_port_state
*state
= state_
;
1448 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1449 struct cmap_node
*node
;
1452 node
= cmap_next_position(&dp
->ports
, &state
->position
);
1454 struct dp_netdev_port
*port
;
1456 port
= CONTAINER_OF(node
, struct dp_netdev_port
, node
);
1459 state
->name
= xstrdup(netdev_get_name(port
->netdev
));
1460 dpif_port
->name
= state
->name
;
1461 dpif_port
->type
= port
->type
;
1462 dpif_port
->port_no
= port
->port_no
;
1473 dpif_netdev_port_dump_done(const struct dpif
*dpif OVS_UNUSED
, void *state_
)
1475 struct dp_netdev_port_state
*state
= state_
;
1482 dpif_netdev_port_poll(const struct dpif
*dpif_
, char **devnamep OVS_UNUSED
)
1484 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
1485 uint64_t new_port_seq
;
1488 new_port_seq
= seq_read(dpif
->dp
->port_seq
);
1489 if (dpif
->last_port_seq
!= new_port_seq
) {
1490 dpif
->last_port_seq
= new_port_seq
;
1500 dpif_netdev_port_poll_wait(const struct dpif
*dpif_
)
1502 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
1504 seq_wait(dpif
->dp
->port_seq
, dpif
->last_port_seq
);
1507 static struct dp_netdev_flow
*
1508 dp_netdev_flow_cast(const struct dpcls_rule
*cr
)
1510 return cr
? CONTAINER_OF(cr
, struct dp_netdev_flow
, cr
) : NULL
;
1513 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*flow
)
1515 return ovs_refcount_try_ref_rcu(&flow
->ref_cnt
);
1518 /* netdev_flow_key utilities.
1520 * netdev_flow_key is basically a miniflow. We use these functions
1521 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1522 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1524 * - Since we are dealing exclusively with miniflows created by
1525 * miniflow_extract(), if the map is different the miniflow is different.
1526 * Therefore we can be faster by comparing the map and the miniflow in a
1528 * - These functions can be inlined by the compiler. */
1530 /* Given the number of bits set in miniflow's maps, returns the size of the
1531 * 'netdev_flow_key.mf' */
1532 static inline size_t
1533 netdev_flow_key_size(size_t flow_u64s
)
1535 return sizeof(struct miniflow
) + MINIFLOW_VALUES_SIZE(flow_u64s
);
1539 netdev_flow_key_equal(const struct netdev_flow_key
*a
,
1540 const struct netdev_flow_key
*b
)
1542 /* 'b->len' may be not set yet. */
1543 return a
->hash
== b
->hash
&& !memcmp(&a
->mf
, &b
->mf
, a
->len
);
1546 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1547 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1548 * generated by miniflow_extract. */
1550 netdev_flow_key_equal_mf(const struct netdev_flow_key
*key
,
1551 const struct miniflow
*mf
)
1553 return !memcmp(&key
->mf
, mf
, key
->len
);
1557 netdev_flow_key_clone(struct netdev_flow_key
*dst
,
1558 const struct netdev_flow_key
*src
)
1561 offsetof(struct netdev_flow_key
, mf
) + src
->len
);
1566 netdev_flow_key_from_flow(struct netdev_flow_key
*dst
,
1567 const struct flow
*src
)
1569 struct dp_packet packet
;
1570 uint64_t buf_stub
[512 / 8];
1572 dp_packet_use_stub(&packet
, buf_stub
, sizeof buf_stub
);
1573 pkt_metadata_from_flow(&packet
.md
, src
);
1574 flow_compose(&packet
, src
);
1575 miniflow_extract(&packet
, &dst
->mf
);
1576 dp_packet_uninit(&packet
);
1578 dst
->len
= netdev_flow_key_size(miniflow_n_values(&dst
->mf
));
1579 dst
->hash
= 0; /* Not computed yet. */
1582 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1584 netdev_flow_mask_init(struct netdev_flow_key
*mask
,
1585 const struct match
*match
)
1587 uint64_t *dst
= miniflow_values(&mask
->mf
);
1588 struct flowmap fmap
;
1592 /* Only check masks that make sense for the flow. */
1593 flow_wc_map(&match
->flow
, &fmap
);
1594 flowmap_init(&mask
->mf
.map
);
1596 FLOWMAP_FOR_EACH_INDEX(idx
, fmap
) {
1597 uint64_t mask_u64
= flow_u64_value(&match
->wc
.masks
, idx
);
1600 flowmap_set(&mask
->mf
.map
, idx
, 1);
1602 hash
= hash_add64(hash
, mask_u64
);
1608 FLOWMAP_FOR_EACH_MAP (map
, mask
->mf
.map
) {
1609 hash
= hash_add64(hash
, map
);
1612 size_t n
= dst
- miniflow_get_values(&mask
->mf
);
1614 mask
->hash
= hash_finish(hash
, n
* 8);
1615 mask
->len
= netdev_flow_key_size(n
);
1618 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
1620 netdev_flow_key_init_masked(struct netdev_flow_key
*dst
,
1621 const struct flow
*flow
,
1622 const struct netdev_flow_key
*mask
)
1624 uint64_t *dst_u64
= miniflow_values(&dst
->mf
);
1625 const uint64_t *mask_u64
= miniflow_get_values(&mask
->mf
);
1629 dst
->len
= mask
->len
;
1630 dst
->mf
= mask
->mf
; /* Copy maps. */
1632 FLOW_FOR_EACH_IN_MAPS(value
, flow
, mask
->mf
.map
) {
1633 *dst_u64
= value
& *mask_u64
++;
1634 hash
= hash_add64(hash
, *dst_u64
++);
1636 dst
->hash
= hash_finish(hash
,
1637 (dst_u64
- miniflow_get_values(&dst
->mf
)) * 8);
1640 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
1641 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
1642 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
1644 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1646 static inline uint32_t
1647 netdev_flow_key_hash_in_mask(const struct netdev_flow_key
*key
,
1648 const struct netdev_flow_key
*mask
)
1650 const uint64_t *p
= miniflow_get_values(&mask
->mf
);
1654 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, key
, mask
->mf
.map
) {
1655 hash
= hash_add64(hash
, value
& *p
++);
1658 return hash_finish(hash
, (p
- miniflow_get_values(&mask
->mf
)) * 8);
1662 emc_entry_alive(struct emc_entry
*ce
)
1664 return ce
->flow
&& !ce
->flow
->dead
;
1668 emc_clear_entry(struct emc_entry
*ce
)
1671 dp_netdev_flow_unref(ce
->flow
);
1677 emc_change_entry(struct emc_entry
*ce
, struct dp_netdev_flow
*flow
,
1678 const struct netdev_flow_key
*key
)
1680 if (ce
->flow
!= flow
) {
1682 dp_netdev_flow_unref(ce
->flow
);
1685 if (dp_netdev_flow_ref(flow
)) {
1692 netdev_flow_key_clone(&ce
->key
, key
);
1697 emc_insert(struct emc_cache
*cache
, const struct netdev_flow_key
*key
,
1698 struct dp_netdev_flow
*flow
)
1700 struct emc_entry
*to_be_replaced
= NULL
;
1701 struct emc_entry
*current_entry
;
1703 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
1704 if (netdev_flow_key_equal(¤t_entry
->key
, key
)) {
1705 /* We found the entry with the 'mf' miniflow */
1706 emc_change_entry(current_entry
, flow
, NULL
);
1710 /* Replacement policy: put the flow in an empty (not alive) entry, or
1711 * in the first entry where it can be */
1713 || (emc_entry_alive(to_be_replaced
)
1714 && !emc_entry_alive(current_entry
))
1715 || current_entry
->key
.hash
< to_be_replaced
->key
.hash
) {
1716 to_be_replaced
= current_entry
;
1719 /* We didn't find the miniflow in the cache.
1720 * The 'to_be_replaced' entry is where the new flow will be stored */
1722 emc_change_entry(to_be_replaced
, flow
, key
);
1725 static inline struct dp_netdev_flow
*
1726 emc_lookup(struct emc_cache
*cache
, const struct netdev_flow_key
*key
)
1728 struct emc_entry
*current_entry
;
1730 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
1731 if (current_entry
->key
.hash
== key
->hash
1732 && emc_entry_alive(current_entry
)
1733 && netdev_flow_key_equal_mf(¤t_entry
->key
, &key
->mf
)) {
1735 /* We found the entry with the 'key->mf' miniflow */
1736 return current_entry
->flow
;
1743 static struct dp_netdev_flow
*
1744 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread
*pmd
,
1745 const struct netdev_flow_key
*key
)
1747 struct dp_netdev_flow
*netdev_flow
;
1748 struct dpcls_rule
*rule
;
1750 dpcls_lookup(&pmd
->cls
, key
, &rule
, 1);
1751 netdev_flow
= dp_netdev_flow_cast(rule
);
1756 static struct dp_netdev_flow
*
1757 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread
*pmd
,
1758 const ovs_u128
*ufidp
, const struct nlattr
*key
,
1761 struct dp_netdev_flow
*netdev_flow
;
1765 /* If a UFID is not provided, determine one based on the key. */
1766 if (!ufidp
&& key
&& key_len
1767 && !dpif_netdev_flow_from_nlattrs(key
, key_len
, &flow
)) {
1768 dpif_flow_hash(pmd
->dp
->dpif
, &flow
, sizeof flow
, &ufid
);
1773 CMAP_FOR_EACH_WITH_HASH (netdev_flow
, node
, dp_netdev_flow_hash(ufidp
),
1775 if (ovs_u128_equals(&netdev_flow
->ufid
, ufidp
)) {
1785 get_dpif_flow_stats(const struct dp_netdev_flow
*netdev_flow_
,
1786 struct dpif_flow_stats
*stats
)
1788 struct dp_netdev_flow
*netdev_flow
;
1789 unsigned long long n
;
1793 netdev_flow
= CONST_CAST(struct dp_netdev_flow
*, netdev_flow_
);
1795 atomic_read_relaxed(&netdev_flow
->stats
.packet_count
, &n
);
1796 stats
->n_packets
= n
;
1797 atomic_read_relaxed(&netdev_flow
->stats
.byte_count
, &n
);
1799 atomic_read_relaxed(&netdev_flow
->stats
.used
, &used
);
1801 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
1802 stats
->tcp_flags
= flags
;
1805 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1806 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1807 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1810 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow
*netdev_flow
,
1811 struct ofpbuf
*key_buf
, struct ofpbuf
*mask_buf
,
1812 struct dpif_flow
*flow
, bool terse
)
1815 memset(flow
, 0, sizeof *flow
);
1817 struct flow_wildcards wc
;
1818 struct dp_netdev_actions
*actions
;
1820 struct odp_flow_key_parms odp_parms
= {
1821 .flow
= &netdev_flow
->flow
,
1823 .support
= dp_netdev_support
,
1826 miniflow_expand(&netdev_flow
->cr
.mask
->mf
, &wc
.masks
);
1829 offset
= key_buf
->size
;
1830 flow
->key
= ofpbuf_tail(key_buf
);
1831 odp_parms
.odp_in_port
= netdev_flow
->flow
.in_port
.odp_port
;
1832 odp_flow_key_from_flow(&odp_parms
, key_buf
);
1833 flow
->key_len
= key_buf
->size
- offset
;
1836 offset
= mask_buf
->size
;
1837 flow
->mask
= ofpbuf_tail(mask_buf
);
1838 odp_parms
.odp_in_port
= wc
.masks
.in_port
.odp_port
;
1839 odp_parms
.key_buf
= key_buf
;
1840 odp_flow_key_from_mask(&odp_parms
, mask_buf
);
1841 flow
->mask_len
= mask_buf
->size
- offset
;
1844 actions
= dp_netdev_flow_get_actions(netdev_flow
);
1845 flow
->actions
= actions
->actions
;
1846 flow
->actions_len
= actions
->size
;
1849 flow
->ufid
= netdev_flow
->ufid
;
1850 flow
->ufid_present
= true;
1851 flow
->pmd_id
= netdev_flow
->pmd_id
;
1852 get_dpif_flow_stats(netdev_flow
, &flow
->stats
);
1856 dpif_netdev_mask_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
1857 const struct nlattr
*mask_key
,
1858 uint32_t mask_key_len
, const struct flow
*flow
,
1859 struct flow_wildcards
*wc
)
1862 enum odp_key_fitness fitness
;
1864 fitness
= odp_flow_key_to_mask_udpif(mask_key
, mask_key_len
, key
,
1865 key_len
, &wc
->masks
, flow
);
1867 /* This should not happen: it indicates that
1868 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1869 * disagree on the acceptable form of a mask. Log the problem
1870 * as an error, with enough details to enable debugging. */
1871 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
1873 if (!VLOG_DROP_ERR(&rl
)) {
1877 odp_flow_format(key
, key_len
, mask_key
, mask_key_len
, NULL
, &s
,
1879 VLOG_ERR("internal error parsing flow mask %s (%s)",
1880 ds_cstr(&s
), odp_key_fitness_to_string(fitness
));
1887 flow_wildcards_init_for_packet(wc
, flow
);
1894 dpif_netdev_flow_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
1899 if (odp_flow_key_to_flow_udpif(key
, key_len
, flow
)) {
1900 /* This should not happen: it indicates that odp_flow_key_from_flow()
1901 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1902 * flow. Log the problem as an error, with enough details to enable
1904 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
1906 if (!VLOG_DROP_ERR(&rl
)) {
1910 odp_flow_format(key
, key_len
, NULL
, 0, NULL
, &s
, true);
1911 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s
));
1918 in_port
= flow
->in_port
.odp_port
;
1919 if (!is_valid_port_number(in_port
) && in_port
!= ODPP_NONE
) {
1923 /* Userspace datapath doesn't support conntrack. */
1924 if (flow
->ct_state
|| flow
->ct_zone
|| flow
->ct_mark
1925 || !ovs_u128_is_zero(&flow
->ct_label
)) {
1933 dpif_netdev_flow_get(const struct dpif
*dpif
, const struct dpif_flow_get
*get
)
1935 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1936 struct dp_netdev_flow
*netdev_flow
;
1937 struct dp_netdev_pmd_thread
*pmd
;
1938 unsigned pmd_id
= get
->pmd_id
== PMD_ID_NULL
1939 ? NON_PMD_CORE_ID
: get
->pmd_id
;
1942 pmd
= dp_netdev_get_pmd(dp
, pmd_id
);
1947 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, get
->ufid
, get
->key
,
1950 dp_netdev_flow_to_dpif_flow(netdev_flow
, get
->buffer
, get
->buffer
,
1955 dp_netdev_pmd_unref(pmd
);
1961 static struct dp_netdev_flow
*
1962 dp_netdev_flow_add(struct dp_netdev_pmd_thread
*pmd
,
1963 struct match
*match
, const ovs_u128
*ufid
,
1964 const struct nlattr
*actions
, size_t actions_len
)
1965 OVS_REQUIRES(pmd
->flow_mutex
)
1967 struct dp_netdev_flow
*flow
;
1968 struct netdev_flow_key mask
;
1970 netdev_flow_mask_init(&mask
, match
);
1971 /* Make sure wc does not have metadata. */
1972 ovs_assert(!FLOWMAP_HAS_FIELD(&mask
.mf
.map
, metadata
)
1973 && !FLOWMAP_HAS_FIELD(&mask
.mf
.map
, regs
));
1975 /* Do not allocate extra space. */
1976 flow
= xmalloc(sizeof *flow
- sizeof flow
->cr
.flow
.mf
+ mask
.len
);
1977 memset(&flow
->stats
, 0, sizeof flow
->stats
);
1980 *CONST_CAST(unsigned *, &flow
->pmd_id
) = pmd
->core_id
;
1981 *CONST_CAST(struct flow
*, &flow
->flow
) = match
->flow
;
1982 *CONST_CAST(ovs_u128
*, &flow
->ufid
) = *ufid
;
1983 ovs_refcount_init(&flow
->ref_cnt
);
1984 ovsrcu_set(&flow
->actions
, dp_netdev_actions_create(actions
, actions_len
));
1986 netdev_flow_key_init_masked(&flow
->cr
.flow
, &match
->flow
, &mask
);
1987 dpcls_insert(&pmd
->cls
, &flow
->cr
, &mask
);
1989 cmap_insert(&pmd
->flow_table
, CONST_CAST(struct cmap_node
*, &flow
->node
),
1990 dp_netdev_flow_hash(&flow
->ufid
));
1992 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1994 struct ds ds
= DS_EMPTY_INITIALIZER
;
1996 match
.flow
= flow
->flow
;
1997 miniflow_expand(&flow
->cr
.mask
->mf
, &match
.wc
.masks
);
1999 ds_put_cstr(&ds
, "flow_add: ");
2000 odp_format_ufid(ufid
, &ds
);
2001 ds_put_cstr(&ds
, " ");
2002 match_format(&match
, &ds
, OFP_DEFAULT_PRIORITY
);
2003 ds_put_cstr(&ds
, ", actions:");
2004 format_odp_actions(&ds
, actions
, actions_len
);
2006 VLOG_DBG_RL(&upcall_rl
, "%s", ds_cstr(&ds
));
2015 dpif_netdev_flow_put(struct dpif
*dpif
, const struct dpif_flow_put
*put
)
2017 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2018 struct dp_netdev_flow
*netdev_flow
;
2019 struct netdev_flow_key key
;
2020 struct dp_netdev_pmd_thread
*pmd
;
2023 unsigned pmd_id
= put
->pmd_id
== PMD_ID_NULL
2024 ? NON_PMD_CORE_ID
: put
->pmd_id
;
2027 error
= dpif_netdev_flow_from_nlattrs(put
->key
, put
->key_len
, &match
.flow
);
2031 error
= dpif_netdev_mask_from_nlattrs(put
->key
, put
->key_len
,
2032 put
->mask
, put
->mask_len
,
2033 &match
.flow
, &match
.wc
);
2038 pmd
= dp_netdev_get_pmd(dp
, pmd_id
);
2043 /* Must produce a netdev_flow_key for lookup.
2044 * This interface is no longer performance critical, since it is not used
2045 * for upcall processing any more. */
2046 netdev_flow_key_from_flow(&key
, &match
.flow
);
2051 dpif_flow_hash(dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
2054 ovs_mutex_lock(&pmd
->flow_mutex
);
2055 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, &key
);
2057 if (put
->flags
& DPIF_FP_CREATE
) {
2058 if (cmap_count(&pmd
->flow_table
) < MAX_FLOWS
) {
2060 memset(put
->stats
, 0, sizeof *put
->stats
);
2062 dp_netdev_flow_add(pmd
, &match
, &ufid
, put
->actions
,
2072 if (put
->flags
& DPIF_FP_MODIFY
2073 && flow_equal(&match
.flow
, &netdev_flow
->flow
)) {
2074 struct dp_netdev_actions
*new_actions
;
2075 struct dp_netdev_actions
*old_actions
;
2077 new_actions
= dp_netdev_actions_create(put
->actions
,
2080 old_actions
= dp_netdev_flow_get_actions(netdev_flow
);
2081 ovsrcu_set(&netdev_flow
->actions
, new_actions
);
2084 get_dpif_flow_stats(netdev_flow
, put
->stats
);
2086 if (put
->flags
& DPIF_FP_ZERO_STATS
) {
2087 /* XXX: The userspace datapath uses thread local statistics
2088 * (for flows), which should be updated only by the owning
2089 * thread. Since we cannot write on stats memory here,
2090 * we choose not to support this flag. Please note:
2091 * - This feature is currently used only by dpctl commands with
2093 * - Should the need arise, this operation can be implemented
2094 * by keeping a base value (to be update here) for each
2095 * counter, and subtracting it before outputting the stats */
2099 ovsrcu_postpone(dp_netdev_actions_free
, old_actions
);
2100 } else if (put
->flags
& DPIF_FP_CREATE
) {
2103 /* Overlapping flow. */
2107 ovs_mutex_unlock(&pmd
->flow_mutex
);
2108 dp_netdev_pmd_unref(pmd
);
2114 dpif_netdev_flow_del(struct dpif
*dpif
, const struct dpif_flow_del
*del
)
2116 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2117 struct dp_netdev_flow
*netdev_flow
;
2118 struct dp_netdev_pmd_thread
*pmd
;
2119 unsigned pmd_id
= del
->pmd_id
== PMD_ID_NULL
2120 ? NON_PMD_CORE_ID
: del
->pmd_id
;
2123 pmd
= dp_netdev_get_pmd(dp
, pmd_id
);
2128 ovs_mutex_lock(&pmd
->flow_mutex
);
2129 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, del
->ufid
, del
->key
,
2133 get_dpif_flow_stats(netdev_flow
, del
->stats
);
2135 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
2139 ovs_mutex_unlock(&pmd
->flow_mutex
);
2140 dp_netdev_pmd_unref(pmd
);
2145 struct dpif_netdev_flow_dump
{
2146 struct dpif_flow_dump up
;
2147 struct cmap_position poll_thread_pos
;
2148 struct cmap_position flow_pos
;
2149 struct dp_netdev_pmd_thread
*cur_pmd
;
2151 struct ovs_mutex mutex
;
2154 static struct dpif_netdev_flow_dump
*
2155 dpif_netdev_flow_dump_cast(struct dpif_flow_dump
*dump
)
2157 return CONTAINER_OF(dump
, struct dpif_netdev_flow_dump
, up
);
2160 static struct dpif_flow_dump
*
2161 dpif_netdev_flow_dump_create(const struct dpif
*dpif_
, bool terse
)
2163 struct dpif_netdev_flow_dump
*dump
;
2165 dump
= xzalloc(sizeof *dump
);
2166 dpif_flow_dump_init(&dump
->up
, dpif_
);
2167 dump
->up
.terse
= terse
;
2168 ovs_mutex_init(&dump
->mutex
);
2174 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump
*dump_
)
2176 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
2178 ovs_mutex_destroy(&dump
->mutex
);
2183 struct dpif_netdev_flow_dump_thread
{
2184 struct dpif_flow_dump_thread up
;
2185 struct dpif_netdev_flow_dump
*dump
;
2186 struct odputil_keybuf keybuf
[FLOW_DUMP_MAX_BATCH
];
2187 struct odputil_keybuf maskbuf
[FLOW_DUMP_MAX_BATCH
];
2190 static struct dpif_netdev_flow_dump_thread
*
2191 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread
*thread
)
2193 return CONTAINER_OF(thread
, struct dpif_netdev_flow_dump_thread
, up
);
2196 static struct dpif_flow_dump_thread
*
2197 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump
*dump_
)
2199 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
2200 struct dpif_netdev_flow_dump_thread
*thread
;
2202 thread
= xmalloc(sizeof *thread
);
2203 dpif_flow_dump_thread_init(&thread
->up
, &dump
->up
);
2204 thread
->dump
= dump
;
2209 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread
*thread_
)
2211 struct dpif_netdev_flow_dump_thread
*thread
2212 = dpif_netdev_flow_dump_thread_cast(thread_
);
2218 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread
*thread_
,
2219 struct dpif_flow
*flows
, int max_flows
)
2221 struct dpif_netdev_flow_dump_thread
*thread
2222 = dpif_netdev_flow_dump_thread_cast(thread_
);
2223 struct dpif_netdev_flow_dump
*dump
= thread
->dump
;
2224 struct dp_netdev_flow
*netdev_flows
[FLOW_DUMP_MAX_BATCH
];
2228 ovs_mutex_lock(&dump
->mutex
);
2229 if (!dump
->status
) {
2230 struct dpif_netdev
*dpif
= dpif_netdev_cast(thread
->up
.dpif
);
2231 struct dp_netdev
*dp
= get_dp_netdev(&dpif
->dpif
);
2232 struct dp_netdev_pmd_thread
*pmd
= dump
->cur_pmd
;
2233 int flow_limit
= MIN(max_flows
, FLOW_DUMP_MAX_BATCH
);
2235 /* First call to dump_next(), extracts the first pmd thread.
2236 * If there is no pmd thread, returns immediately. */
2238 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
2240 ovs_mutex_unlock(&dump
->mutex
);
2247 for (n_flows
= 0; n_flows
< flow_limit
; n_flows
++) {
2248 struct cmap_node
*node
;
2250 node
= cmap_next_position(&pmd
->flow_table
, &dump
->flow_pos
);
2254 netdev_flows
[n_flows
] = CONTAINER_OF(node
,
2255 struct dp_netdev_flow
,
2258 /* When finishing dumping the current pmd thread, moves to
2260 if (n_flows
< flow_limit
) {
2261 memset(&dump
->flow_pos
, 0, sizeof dump
->flow_pos
);
2262 dp_netdev_pmd_unref(pmd
);
2263 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
2269 /* Keeps the reference to next caller. */
2270 dump
->cur_pmd
= pmd
;
2272 /* If the current dump is empty, do not exit the loop, since the
2273 * remaining pmds could have flows to be dumped. Just dumps again
2274 * on the new 'pmd'. */
2277 ovs_mutex_unlock(&dump
->mutex
);
2279 for (i
= 0; i
< n_flows
; i
++) {
2280 struct odputil_keybuf
*maskbuf
= &thread
->maskbuf
[i
];
2281 struct odputil_keybuf
*keybuf
= &thread
->keybuf
[i
];
2282 struct dp_netdev_flow
*netdev_flow
= netdev_flows
[i
];
2283 struct dpif_flow
*f
= &flows
[i
];
2284 struct ofpbuf key
, mask
;
2286 ofpbuf_use_stack(&key
, keybuf
, sizeof *keybuf
);
2287 ofpbuf_use_stack(&mask
, maskbuf
, sizeof *maskbuf
);
2288 dp_netdev_flow_to_dpif_flow(netdev_flow
, &key
, &mask
, f
,
2296 dpif_netdev_execute(struct dpif
*dpif
, struct dpif_execute
*execute
)
2297 OVS_NO_THREAD_SAFETY_ANALYSIS
2299 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2300 struct dp_netdev_pmd_thread
*pmd
;
2301 struct dp_packet
*pp
;
2303 if (dp_packet_size(execute
->packet
) < ETH_HEADER_LEN
||
2304 dp_packet_size(execute
->packet
) > UINT16_MAX
) {
2308 /* Tries finding the 'pmd'. If NULL is returned, that means
2309 * the current thread is a non-pmd thread and should use
2310 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2311 pmd
= ovsthread_getspecific(dp
->per_pmd_key
);
2313 pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
2316 /* If the current thread is non-pmd thread, acquires
2317 * the 'non_pmd_mutex'. */
2318 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2319 ovs_mutex_lock(&dp
->non_pmd_mutex
);
2320 ovs_mutex_lock(&dp
->port_mutex
);
2323 pp
= execute
->packet
;
2324 dp_netdev_execute_actions(pmd
, &pp
, 1, false, execute
->actions
,
2325 execute
->actions_len
);
2326 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2327 dp_netdev_pmd_unref(pmd
);
2328 ovs_mutex_unlock(&dp
->port_mutex
);
2329 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
2336 dpif_netdev_operate(struct dpif
*dpif
, struct dpif_op
**ops
, size_t n_ops
)
2340 for (i
= 0; i
< n_ops
; i
++) {
2341 struct dpif_op
*op
= ops
[i
];
2344 case DPIF_OP_FLOW_PUT
:
2345 op
->error
= dpif_netdev_flow_put(dpif
, &op
->u
.flow_put
);
2348 case DPIF_OP_FLOW_DEL
:
2349 op
->error
= dpif_netdev_flow_del(dpif
, &op
->u
.flow_del
);
2352 case DPIF_OP_EXECUTE
:
2353 op
->error
= dpif_netdev_execute(dpif
, &op
->u
.execute
);
2356 case DPIF_OP_FLOW_GET
:
2357 op
->error
= dpif_netdev_flow_get(dpif
, &op
->u
.flow_get
);
2363 /* Returns true if the configuration for rx queues or cpu mask
2366 pmd_config_changed(const struct dp_netdev
*dp
, size_t rxqs
, const char *cmask
)
2368 if (dp
->n_dpdk_rxqs
!= rxqs
) {
2371 if (dp
->pmd_cmask
!= NULL
&& cmask
!= NULL
) {
2372 return strcmp(dp
->pmd_cmask
, cmask
);
2374 return (dp
->pmd_cmask
!= NULL
|| cmask
!= NULL
);
2379 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2381 dpif_netdev_pmd_set(struct dpif
*dpif
, unsigned int n_rxqs
, const char *cmask
)
2383 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2385 if (pmd_config_changed(dp
, n_rxqs
, cmask
)) {
2386 struct dp_netdev_port
*port
;
2388 dp_netdev_destroy_all_pmds(dp
);
2390 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2391 if (netdev_is_pmd(port
->netdev
)) {
2394 /* Closes the existing 'rxq's. */
2395 for (i
= 0; i
< netdev_n_rxq(port
->netdev
); i
++) {
2396 netdev_rxq_close(port
->rxq
[i
]);
2397 port
->rxq
[i
] = NULL
;
2400 /* Sets the new rx queue config. */
2401 err
= netdev_set_multiq(port
->netdev
,
2402 ovs_numa_get_n_cores() + 1,
2404 if (err
&& (err
!= EOPNOTSUPP
)) {
2405 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2406 " %u", netdev_get_name(port
->netdev
),
2411 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2412 port
->rxq
= xrealloc(port
->rxq
, sizeof *port
->rxq
2413 * netdev_n_rxq(port
->netdev
));
2414 for (i
= 0; i
< netdev_n_rxq(port
->netdev
); i
++) {
2415 netdev_rxq_open(port
->netdev
, &port
->rxq
[i
], i
);
2419 dp
->n_dpdk_rxqs
= n_rxqs
;
2421 /* Reconfigures the cpu mask. */
2422 ovs_numa_set_cpu_mask(cmask
);
2423 free(dp
->pmd_cmask
);
2424 dp
->pmd_cmask
= cmask
? xstrdup(cmask
) : NULL
;
2426 /* Restores the non-pmd. */
2427 dp_netdev_set_nonpmd(dp
);
2428 /* Restores all pmd threads. */
2429 dp_netdev_reset_pmd_threads(dp
);
2436 dpif_netdev_queue_to_priority(const struct dpif
*dpif OVS_UNUSED
,
2437 uint32_t queue_id
, uint32_t *priority
)
2439 *priority
= queue_id
;
2444 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2445 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2446 struct dp_netdev_actions
*
2447 dp_netdev_actions_create(const struct nlattr
*actions
, size_t size
)
2449 struct dp_netdev_actions
*netdev_actions
;
2451 netdev_actions
= xmalloc(sizeof *netdev_actions
+ size
);
2452 memcpy(netdev_actions
->actions
, actions
, size
);
2453 netdev_actions
->size
= size
;
2455 return netdev_actions
;
2458 struct dp_netdev_actions
*
2459 dp_netdev_flow_get_actions(const struct dp_netdev_flow
*flow
)
2461 return ovsrcu_get(struct dp_netdev_actions
*, &flow
->actions
);
2465 dp_netdev_actions_free(struct dp_netdev_actions
*actions
)
2470 static inline unsigned long long
2471 cycles_counter(void)
2474 return rte_get_tsc_cycles();
2480 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2481 extern struct ovs_mutex cycles_counter_fake_mutex
;
2483 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2485 cycles_count_start(struct dp_netdev_pmd_thread
*pmd
)
2486 OVS_ACQUIRES(&cycles_counter_fake_mutex
)
2487 OVS_NO_THREAD_SAFETY_ANALYSIS
2489 pmd
->last_cycles
= cycles_counter();
2492 /* Stop counting cycles and add them to the counter 'type' */
2494 cycles_count_end(struct dp_netdev_pmd_thread
*pmd
,
2495 enum pmd_cycles_counter_type type
)
2496 OVS_RELEASES(&cycles_counter_fake_mutex
)
2497 OVS_NO_THREAD_SAFETY_ANALYSIS
2499 unsigned long long interval
= cycles_counter() - pmd
->last_cycles
;
2501 non_atomic_ullong_add(&pmd
->cycles
.n
[type
], interval
);
2505 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread
*pmd
,
2506 struct dp_netdev_port
*port
,
2507 struct netdev_rxq
*rxq
)
2509 struct dp_packet
*packets
[NETDEV_MAX_BURST
];
2512 cycles_count_start(pmd
);
2513 error
= netdev_rxq_recv(rxq
, packets
, &cnt
);
2514 cycles_count_end(pmd
, PMD_CYCLES_POLLING
);
2518 *recirc_depth_get() = 0;
2520 /* XXX: initialize md in netdev implementation. */
2521 for (i
= 0; i
< cnt
; i
++) {
2522 pkt_metadata_init(&packets
[i
]->md
, port
->port_no
);
2524 cycles_count_start(pmd
);
2525 dp_netdev_input(pmd
, packets
, cnt
);
2526 cycles_count_end(pmd
, PMD_CYCLES_PROCESSING
);
2527 } else if (error
!= EAGAIN
&& error
!= EOPNOTSUPP
) {
2528 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
2530 VLOG_ERR_RL(&rl
, "error receiving data from %s: %s",
2531 netdev_get_name(port
->netdev
), ovs_strerror(error
));
2535 /* Return true if needs to revalidate datapath flows. */
2537 dpif_netdev_run(struct dpif
*dpif
)
2539 struct dp_netdev_port
*port
;
2540 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2541 struct dp_netdev_pmd_thread
*non_pmd
= dp_netdev_get_pmd(dp
,
2543 uint64_t new_tnl_seq
;
2545 ovs_mutex_lock(&dp
->non_pmd_mutex
);
2546 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2547 if (!netdev_is_pmd(port
->netdev
)) {
2550 for (i
= 0; i
< netdev_n_rxq(port
->netdev
); i
++) {
2551 dp_netdev_process_rxq_port(non_pmd
, port
, port
->rxq
[i
]);
2555 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
2556 dp_netdev_pmd_unref(non_pmd
);
2558 tnl_arp_cache_run();
2560 new_tnl_seq
= seq_read(tnl_conf_seq
);
2562 if (dp
->last_tnl_conf_seq
!= new_tnl_seq
) {
2563 dp
->last_tnl_conf_seq
= new_tnl_seq
;
2570 dpif_netdev_wait(struct dpif
*dpif
)
2572 struct dp_netdev_port
*port
;
2573 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2575 ovs_mutex_lock(&dp_netdev_mutex
);
2576 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2577 if (!netdev_is_pmd(port
->netdev
)) {
2580 for (i
= 0; i
< netdev_n_rxq(port
->netdev
); i
++) {
2581 netdev_rxq_wait(port
->rxq
[i
]);
2585 ovs_mutex_unlock(&dp_netdev_mutex
);
2586 seq_wait(tnl_conf_seq
, dp
->last_tnl_conf_seq
);
2590 struct dp_netdev_port
*port
;
2591 struct netdev_rxq
*rx
;
2595 pmd_load_queues(struct dp_netdev_pmd_thread
*pmd
,
2596 struct rxq_poll
**ppoll_list
, int poll_cnt
)
2598 struct rxq_poll
*poll_list
= *ppoll_list
;
2599 struct dp_netdev_port
*port
;
2600 int n_pmds_on_numa
, index
, i
;
2602 /* Simple scheduler for netdev rx polling. */
2603 for (i
= 0; i
< poll_cnt
; i
++) {
2604 port_unref(poll_list
[i
].port
);
2608 n_pmds_on_numa
= get_n_pmd_threads_on_numa(pmd
->dp
, pmd
->numa_id
);
2611 CMAP_FOR_EACH (port
, node
, &pmd
->dp
->ports
) {
2612 /* Calls port_try_ref() to prevent the main thread
2613 * from deleting the port. */
2614 if (port_try_ref(port
)) {
2615 if (netdev_is_pmd(port
->netdev
)
2616 && netdev_get_numa_id(port
->netdev
) == pmd
->numa_id
) {
2619 for (i
= 0; i
< netdev_n_rxq(port
->netdev
); i
++) {
2620 if ((index
% n_pmds_on_numa
) == pmd
->index
) {
2621 poll_list
= xrealloc(poll_list
,
2622 sizeof *poll_list
* (poll_cnt
+ 1));
2625 poll_list
[poll_cnt
].port
= port
;
2626 poll_list
[poll_cnt
].rx
= port
->rxq
[i
];
2632 /* Unrefs the port_try_ref(). */
2637 *ppoll_list
= poll_list
;
2642 pmd_thread_main(void *f_
)
2644 struct dp_netdev_pmd_thread
*pmd
= f_
;
2645 unsigned int lc
= 0;
2646 struct rxq_poll
*poll_list
;
2647 unsigned int port_seq
= PMD_INITIAL_SEQ
;
2654 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2655 ovsthread_setspecific(pmd
->dp
->per_pmd_key
, pmd
);
2656 pmd_thread_setaffinity_cpu(pmd
->core_id
);
2658 emc_cache_init(&pmd
->flow_cache
);
2659 poll_cnt
= pmd_load_queues(pmd
, &poll_list
, poll_cnt
);
2661 /* List port/core affinity */
2662 for (i
= 0; i
< poll_cnt
; i
++) {
2663 VLOG_INFO("Core %d processing port \'%s\'\n", pmd
->core_id
, netdev_get_name(poll_list
[i
].port
->netdev
));
2666 /* Signal here to make sure the pmd finishes
2667 * reloading the updated configuration. */
2668 dp_netdev_pmd_reload_done(pmd
);
2673 for (i
= 0; i
< poll_cnt
; i
++) {
2674 dp_netdev_process_rxq_port(pmd
, poll_list
[i
].port
, poll_list
[i
].rx
);
2682 emc_cache_slow_sweep(&pmd
->flow_cache
);
2683 coverage_try_clear();
2686 atomic_read_relaxed(&pmd
->change_seq
, &seq
);
2687 if (seq
!= port_seq
) {
2694 emc_cache_uninit(&pmd
->flow_cache
);
2696 if (!latch_is_set(&pmd
->exit_latch
)){
2700 for (i
= 0; i
< poll_cnt
; i
++) {
2701 port_unref(poll_list
[i
].port
);
2704 dp_netdev_pmd_reload_done(pmd
);
2711 dp_netdev_disable_upcall(struct dp_netdev
*dp
)
2712 OVS_ACQUIRES(dp
->upcall_rwlock
)
2714 fat_rwlock_wrlock(&dp
->upcall_rwlock
);
2718 dpif_netdev_disable_upcall(struct dpif
*dpif
)
2719 OVS_NO_THREAD_SAFETY_ANALYSIS
2721 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2722 dp_netdev_disable_upcall(dp
);
2726 dp_netdev_enable_upcall(struct dp_netdev
*dp
)
2727 OVS_RELEASES(dp
->upcall_rwlock
)
2729 fat_rwlock_unlock(&dp
->upcall_rwlock
);
2733 dpif_netdev_enable_upcall(struct dpif
*dpif
)
2734 OVS_NO_THREAD_SAFETY_ANALYSIS
2736 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2737 dp_netdev_enable_upcall(dp
);
2741 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
)
2743 ovs_mutex_lock(&pmd
->cond_mutex
);
2744 xpthread_cond_signal(&pmd
->cond
);
2745 ovs_mutex_unlock(&pmd
->cond_mutex
);
2748 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2749 * the pointer if succeeds, otherwise, NULL.
2751 * Caller must unrefs the returned reference. */
2752 static struct dp_netdev_pmd_thread
*
2753 dp_netdev_get_pmd(struct dp_netdev
*dp
, unsigned core_id
)
2755 struct dp_netdev_pmd_thread
*pmd
;
2756 const struct cmap_node
*pnode
;
2758 pnode
= cmap_find(&dp
->poll_threads
, hash_int(core_id
, 0));
2762 pmd
= CONTAINER_OF(pnode
, struct dp_netdev_pmd_thread
, node
);
2764 return dp_netdev_pmd_try_ref(pmd
) ? pmd
: NULL
;
2767 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2769 dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
2771 struct dp_netdev_pmd_thread
*non_pmd
;
2773 non_pmd
= xzalloc(sizeof *non_pmd
);
2774 dp_netdev_configure_pmd(non_pmd
, dp
, 0, NON_PMD_CORE_ID
,
2778 /* Caller must have valid pointer to 'pmd'. */
2780 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
)
2782 return ovs_refcount_try_ref_rcu(&pmd
->ref_cnt
);
2786 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
)
2788 if (pmd
&& ovs_refcount_unref(&pmd
->ref_cnt
) == 1) {
2789 ovsrcu_postpone(dp_netdev_destroy_pmd
, pmd
);
2793 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2794 * fails, keeps checking for next node until reaching the end of cmap.
2796 * Caller must unrefs the returned reference. */
2797 static struct dp_netdev_pmd_thread
*
2798 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
)
2800 struct dp_netdev_pmd_thread
*next
;
2803 struct cmap_node
*node
;
2805 node
= cmap_next_position(&dp
->poll_threads
, pos
);
2806 next
= node
? CONTAINER_OF(node
, struct dp_netdev_pmd_thread
, node
)
2808 } while (next
&& !dp_netdev_pmd_try_ref(next
));
2814 core_id_to_qid(unsigned core_id
)
2816 if (core_id
!= NON_PMD_CORE_ID
) {
2819 return ovs_numa_get_n_cores();
2823 /* Configures the 'pmd' based on the input argument. */
2825 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
, struct dp_netdev
*dp
,
2826 int index
, unsigned core_id
, int numa_id
)
2830 pmd
->core_id
= core_id
;
2831 pmd
->tx_qid
= core_id_to_qid(core_id
);
2832 pmd
->numa_id
= numa_id
;
2834 ovs_refcount_init(&pmd
->ref_cnt
);
2835 latch_init(&pmd
->exit_latch
);
2836 atomic_init(&pmd
->change_seq
, PMD_INITIAL_SEQ
);
2837 xpthread_cond_init(&pmd
->cond
, NULL
);
2838 ovs_mutex_init(&pmd
->cond_mutex
);
2839 ovs_mutex_init(&pmd
->flow_mutex
);
2840 dpcls_init(&pmd
->cls
);
2841 cmap_init(&pmd
->flow_table
);
2842 /* init the 'flow_cache' since there is no
2843 * actual thread created for NON_PMD_CORE_ID. */
2844 if (core_id
== NON_PMD_CORE_ID
) {
2845 emc_cache_init(&pmd
->flow_cache
);
2847 cmap_insert(&dp
->poll_threads
, CONST_CAST(struct cmap_node
*, &pmd
->node
),
2848 hash_int(core_id
, 0));
2852 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
)
2854 dp_netdev_pmd_flow_flush(pmd
);
2855 dpcls_destroy(&pmd
->cls
);
2856 cmap_destroy(&pmd
->flow_table
);
2857 ovs_mutex_destroy(&pmd
->flow_mutex
);
2858 latch_destroy(&pmd
->exit_latch
);
2859 xpthread_cond_destroy(&pmd
->cond
);
2860 ovs_mutex_destroy(&pmd
->cond_mutex
);
2864 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2865 * and unrefs the struct. */
2867 dp_netdev_del_pmd(struct dp_netdev
*dp
, struct dp_netdev_pmd_thread
*pmd
)
2869 /* Uninit the 'flow_cache' since there is
2870 * no actual thread uninit it for NON_PMD_CORE_ID. */
2871 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2872 emc_cache_uninit(&pmd
->flow_cache
);
2874 latch_set(&pmd
->exit_latch
);
2875 dp_netdev_reload_pmd__(pmd
);
2876 ovs_numa_unpin_core(pmd
->core_id
);
2877 xpthread_join(pmd
->thread
, NULL
);
2879 /* Purges the 'pmd''s flows after stopping the thread, but before
2880 * destroying the flows, so that the flow stats can be collected. */
2881 if (dp
->dp_purge_cb
) {
2882 dp
->dp_purge_cb(dp
->dp_purge_aux
, pmd
->core_id
);
2884 cmap_remove(&pmd
->dp
->poll_threads
, &pmd
->node
, hash_int(pmd
->core_id
, 0));
2885 dp_netdev_pmd_unref(pmd
);
2888 /* Destroys all pmd threads. */
2890 dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
)
2892 struct dp_netdev_pmd_thread
*pmd
;
2894 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2895 dp_netdev_del_pmd(dp
, pmd
);
2899 /* Deletes all pmd threads on numa node 'numa_id'. */
2901 dp_netdev_del_pmds_on_numa(struct dp_netdev
*dp
, int numa_id
)
2903 struct dp_netdev_pmd_thread
*pmd
;
2905 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2906 if (pmd
->numa_id
== numa_id
) {
2907 dp_netdev_del_pmd(dp
, pmd
);
2912 /* Checks the numa node id of 'netdev' and starts pmd threads for
2915 dp_netdev_set_pmds_on_numa(struct dp_netdev
*dp
, int numa_id
)
2919 if (!ovs_numa_numa_id_is_valid(numa_id
)) {
2920 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2921 "invalid", numa_id
);
2925 n_pmds
= get_n_pmd_threads_on_numa(dp
, numa_id
);
2927 /* If there are already pmd threads created for the numa node
2928 * in which 'netdev' is on, do nothing. Else, creates the
2929 * pmd threads for the numa node. */
2931 int can_have
, n_unpinned
, i
;
2932 struct dp_netdev_pmd_thread
**pmds
;
2934 n_unpinned
= ovs_numa_get_n_unpinned_cores_on_numa(numa_id
);
2936 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2937 "cores on numa node");
2941 /* If cpu mask is specified, uses all unpinned cores, otherwise
2942 * tries creating NR_PMD_THREADS pmd threads. */
2943 can_have
= dp
->pmd_cmask
? n_unpinned
: MIN(n_unpinned
, NR_PMD_THREADS
);
2944 pmds
= xzalloc(can_have
* sizeof *pmds
);
2945 for (i
= 0; i
< can_have
; i
++) {
2946 unsigned core_id
= ovs_numa_get_unpinned_core_on_numa(numa_id
);
2947 pmds
[i
] = xzalloc(sizeof **pmds
);
2948 dp_netdev_configure_pmd(pmds
[i
], dp
, i
, core_id
, numa_id
);
2950 /* The pmd thread code needs to see all the others configured pmd
2951 * threads on the same numa node. That's why we call
2952 * 'dp_netdev_configure_pmd()' on all the threads and then we actually
2954 for (i
= 0; i
< can_have
; i
++) {
2955 /* Each thread will distribute all devices rx-queues among
2957 pmds
[i
]->thread
= ovs_thread_create("pmd", pmd_thread_main
, pmds
[i
]);
2960 VLOG_INFO("Created %d pmd threads on numa node %d", can_have
, numa_id
);
2965 /* Called after pmd threads config change. Restarts pmd threads with
2966 * new configuration. */
2968 dp_netdev_reset_pmd_threads(struct dp_netdev
*dp
)
2970 struct dp_netdev_port
*port
;
2972 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2973 if (netdev_is_pmd(port
->netdev
)) {
2974 int numa_id
= netdev_get_numa_id(port
->netdev
);
2976 dp_netdev_set_pmds_on_numa(dp
, numa_id
);
2982 dpif_netdev_get_datapath_version(void)
2984 return xstrdup("<built-in>");
2988 dp_netdev_flow_used(struct dp_netdev_flow
*netdev_flow
, int cnt
, int size
,
2989 uint16_t tcp_flags
, long long now
)
2993 atomic_store_relaxed(&netdev_flow
->stats
.used
, now
);
2994 non_atomic_ullong_add(&netdev_flow
->stats
.packet_count
, cnt
);
2995 non_atomic_ullong_add(&netdev_flow
->stats
.byte_count
, size
);
2996 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
2998 atomic_store_relaxed(&netdev_flow
->stats
.tcp_flags
, flags
);
3002 dp_netdev_count_packet(struct dp_netdev_pmd_thread
*pmd
,
3003 enum dp_stat_type type
, int cnt
)
3005 non_atomic_ullong_add(&pmd
->stats
.n
[type
], cnt
);
3009 dp_netdev_upcall(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
*packet_
,
3010 struct flow
*flow
, struct flow_wildcards
*wc
, ovs_u128
*ufid
,
3011 enum dpif_upcall_type type
, const struct nlattr
*userdata
,
3012 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
3014 struct dp_netdev
*dp
= pmd
->dp
;
3015 struct flow_tnl orig_tunnel
;
3018 if (OVS_UNLIKELY(!dp
->upcall_cb
)) {
3022 /* Upcall processing expects the Geneve options to be in the translated
3023 * format but we need to retain the raw format for datapath use. */
3024 orig_tunnel
.flags
= flow
->tunnel
.flags
;
3025 if (flow
->tunnel
.flags
& FLOW_TNL_F_UDPIF
) {
3026 orig_tunnel
.metadata
.present
.len
= flow
->tunnel
.metadata
.present
.len
;
3027 memcpy(orig_tunnel
.metadata
.opts
.gnv
, flow
->tunnel
.metadata
.opts
.gnv
,
3028 flow
->tunnel
.metadata
.present
.len
);
3029 err
= tun_metadata_from_geneve_udpif(&orig_tunnel
, &orig_tunnel
,
3036 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl
))) {
3037 struct ds ds
= DS_EMPTY_INITIALIZER
;
3040 struct odp_flow_key_parms odp_parms
= {
3043 .odp_in_port
= flow
->in_port
.odp_port
,
3044 .support
= dp_netdev_support
,
3047 ofpbuf_init(&key
, 0);
3048 odp_flow_key_from_flow(&odp_parms
, &key
);
3049 packet_str
= ofp_packet_to_string(dp_packet_data(packet_
),
3050 dp_packet_size(packet_
));
3052 odp_flow_key_format(key
.data
, key
.size
, &ds
);
3054 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp
->name
,
3055 dpif_upcall_type_to_string(type
), ds_cstr(&ds
), packet_str
);
3057 ofpbuf_uninit(&key
);
3063 err
= dp
->upcall_cb(packet_
, flow
, ufid
, pmd
->core_id
, type
, userdata
,
3064 actions
, wc
, put_actions
, dp
->upcall_aux
);
3065 if (err
&& err
!= ENOSPC
) {
3069 /* Translate tunnel metadata masks to datapath format. */
3071 if (wc
->masks
.tunnel
.metadata
.present
.map
) {
3072 struct geneve_opt opts
[GENEVE_TOT_OPT_SIZE
/
3073 sizeof(struct geneve_opt
)];
3075 tun_metadata_to_geneve_udpif_mask(&flow
->tunnel
,
3077 orig_tunnel
.metadata
.opts
.gnv
,
3078 orig_tunnel
.metadata
.present
.len
,
3081 memset(&wc
->masks
.tunnel
.metadata
, 0,
3082 sizeof wc
->masks
.tunnel
.metadata
);
3083 memcpy(&wc
->masks
.tunnel
.metadata
.opts
.gnv
, opts
,
3084 orig_tunnel
.metadata
.present
.len
);
3086 wc
->masks
.tunnel
.metadata
.present
.len
= 0xff;
3089 /* Restore tunnel metadata. We need to use the saved options to ensure
3090 * that any unknown options are not lost. The generated mask will have
3091 * the same structure, matching on types and lengths but wildcarding
3092 * option data we don't care about. */
3093 if (orig_tunnel
.flags
& FLOW_TNL_F_UDPIF
) {
3094 memcpy(&flow
->tunnel
.metadata
.opts
.gnv
, orig_tunnel
.metadata
.opts
.gnv
,
3095 orig_tunnel
.metadata
.present
.len
);
3096 flow
->tunnel
.metadata
.present
.len
= orig_tunnel
.metadata
.present
.len
;
3097 flow
->tunnel
.flags
|= FLOW_TNL_F_UDPIF
;
3103 static inline uint32_t
3104 dpif_netdev_packet_get_rss_hash(struct dp_packet
*packet
,
3105 const struct miniflow
*mf
)
3107 uint32_t hash
, recirc_depth
;
3109 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
3110 hash
= dp_packet_get_rss_hash(packet
);
3112 hash
= miniflow_hash_5tuple(mf
, 0);
3113 dp_packet_set_rss_hash(packet
, hash
);
3116 /* The RSS hash must account for the recirculation depth to avoid
3117 * collisions in the exact match cache */
3118 recirc_depth
= *recirc_depth_get_unsafe();
3119 if (OVS_UNLIKELY(recirc_depth
)) {
3120 hash
= hash_finish(hash
, recirc_depth
);
3121 dp_packet_set_rss_hash(packet
, hash
);
3126 struct packet_batch
{
3127 unsigned int packet_count
;
3128 unsigned int byte_count
;
3131 struct dp_netdev_flow
*flow
;
3133 struct dp_packet
*packets
[NETDEV_MAX_BURST
];
3137 packet_batch_update(struct packet_batch
*batch
, struct dp_packet
*packet
,
3138 const struct miniflow
*mf
)
3140 batch
->tcp_flags
|= miniflow_get_tcp_flags(mf
);
3141 batch
->packets
[batch
->packet_count
++] = packet
;
3142 batch
->byte_count
+= dp_packet_size(packet
);
3146 packet_batch_init(struct packet_batch
*batch
, struct dp_netdev_flow
*flow
)
3148 flow
->batch
= batch
;
3151 batch
->packet_count
= 0;
3152 batch
->byte_count
= 0;
3153 batch
->tcp_flags
= 0;
3157 packet_batch_execute(struct packet_batch
*batch
,
3158 struct dp_netdev_pmd_thread
*pmd
,
3161 struct dp_netdev_actions
*actions
;
3162 struct dp_netdev_flow
*flow
= batch
->flow
;
3164 dp_netdev_flow_used(flow
, batch
->packet_count
, batch
->byte_count
,
3165 batch
->tcp_flags
, now
);
3167 actions
= dp_netdev_flow_get_actions(flow
);
3169 dp_netdev_execute_actions(pmd
, batch
->packets
, batch
->packet_count
, true,
3170 actions
->actions
, actions
->size
);
3174 dp_netdev_queue_batches(struct dp_packet
*pkt
,
3175 struct dp_netdev_flow
*flow
, const struct miniflow
*mf
,
3176 struct packet_batch
*batches
, size_t *n_batches
)
3178 struct packet_batch
*batch
= flow
->batch
;
3180 if (OVS_LIKELY(batch
)) {
3181 packet_batch_update(batch
, pkt
, mf
);
3185 batch
= &batches
[(*n_batches
)++];
3186 packet_batch_init(batch
, flow
);
3187 packet_batch_update(batch
, pkt
, mf
);
3191 dp_packet_swap(struct dp_packet
**a
, struct dp_packet
**b
)
3193 struct dp_packet
*tmp
= *a
;
3198 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3199 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', the
3200 * miniflow is copied into 'keys' and the packet pointer is moved at the
3201 * beginning of the 'packets' array.
3203 * The function returns the number of packets that needs to be processed in the
3204 * 'packets' array (they have been moved to the beginning of the vector).
3206 static inline size_t
3207 emc_processing(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
**packets
,
3208 size_t cnt
, struct netdev_flow_key
*keys
,
3209 struct packet_batch batches
[], size_t *n_batches
)
3211 struct emc_cache
*flow_cache
= &pmd
->flow_cache
;
3212 struct netdev_flow_key key
;
3213 size_t i
, notfound_cnt
= 0;
3215 for (i
= 0; i
< cnt
; i
++) {
3216 struct dp_netdev_flow
*flow
;
3218 if (OVS_UNLIKELY(dp_packet_size(packets
[i
]) < ETH_HEADER_LEN
)) {
3219 dp_packet_delete(packets
[i
]);
3224 /* Prefetch next packet data */
3225 OVS_PREFETCH(dp_packet_data(packets
[i
+1]));
3228 miniflow_extract(packets
[i
], &key
.mf
);
3229 key
.len
= 0; /* Not computed yet. */
3230 key
.hash
= dpif_netdev_packet_get_rss_hash(packets
[i
], &key
.mf
);
3232 flow
= emc_lookup(flow_cache
, &key
);
3233 if (OVS_LIKELY(flow
)) {
3234 dp_netdev_queue_batches(packets
[i
], flow
, &key
.mf
, batches
,
3237 if (i
!= notfound_cnt
) {
3238 dp_packet_swap(&packets
[i
], &packets
[notfound_cnt
]);
3241 keys
[notfound_cnt
++] = key
;
3245 dp_netdev_count_packet(pmd
, DP_STAT_EXACT_HIT
, cnt
- notfound_cnt
);
3247 return notfound_cnt
;
3251 fast_path_processing(struct dp_netdev_pmd_thread
*pmd
,
3252 struct dp_packet
**packets
, size_t cnt
,
3253 struct netdev_flow_key
*keys
,
3254 struct packet_batch batches
[], size_t *n_batches
)
3256 #if !defined(__CHECKER__) && !defined(_WIN32)
3257 const size_t PKT_ARRAY_SIZE
= cnt
;
3259 /* Sparse or MSVC doesn't like variable length array. */
3260 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
3262 struct dpcls_rule
*rules
[PKT_ARRAY_SIZE
];
3263 struct dp_netdev
*dp
= pmd
->dp
;
3264 struct emc_cache
*flow_cache
= &pmd
->flow_cache
;
3265 int miss_cnt
= 0, lost_cnt
= 0;
3269 for (i
= 0; i
< cnt
; i
++) {
3270 /* Key length is needed in all the cases, hash computed on demand. */
3271 keys
[i
].len
= netdev_flow_key_size(miniflow_n_values(&keys
[i
].mf
));
3273 any_miss
= !dpcls_lookup(&pmd
->cls
, keys
, rules
, cnt
);
3274 if (OVS_UNLIKELY(any_miss
) && !fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
3275 uint64_t actions_stub
[512 / 8], slow_stub
[512 / 8];
3276 struct ofpbuf actions
, put_actions
;
3279 ofpbuf_use_stub(&actions
, actions_stub
, sizeof actions_stub
);
3280 ofpbuf_use_stub(&put_actions
, slow_stub
, sizeof slow_stub
);
3282 for (i
= 0; i
< cnt
; i
++) {
3283 struct dp_netdev_flow
*netdev_flow
;
3284 struct ofpbuf
*add_actions
;
3288 if (OVS_LIKELY(rules
[i
])) {
3292 /* It's possible that an earlier slow path execution installed
3293 * a rule covering this flow. In this case, it's a lot cheaper
3294 * to catch it here than execute a miss. */
3295 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, &keys
[i
]);
3297 rules
[i
] = &netdev_flow
->cr
;
3303 miniflow_expand(&keys
[i
].mf
, &match
.flow
);
3305 ofpbuf_clear(&actions
);
3306 ofpbuf_clear(&put_actions
);
3308 dpif_flow_hash(dp
->dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
3309 error
= dp_netdev_upcall(pmd
, packets
[i
], &match
.flow
, &match
.wc
,
3310 &ufid
, DPIF_UC_MISS
, NULL
, &actions
,
3312 if (OVS_UNLIKELY(error
&& error
!= ENOSPC
)) {
3313 dp_packet_delete(packets
[i
]);
3318 /* The Netlink encoding of datapath flow keys cannot express
3319 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
3320 * tag is interpreted as exact match on the fact that there is no
3321 * VLAN. Unless we refactor a lot of code that translates between
3322 * Netlink and struct flow representations, we have to do the same
3324 if (!match
.wc
.masks
.vlan_tci
) {
3325 match
.wc
.masks
.vlan_tci
= htons(0xffff);
3328 /* We can't allow the packet batching in the next loop to execute
3329 * the actions. Otherwise, if there are any slow path actions,
3330 * we'll send the packet up twice. */
3331 dp_netdev_execute_actions(pmd
, &packets
[i
], 1, true,
3332 actions
.data
, actions
.size
);
3334 add_actions
= put_actions
.size
? &put_actions
: &actions
;
3335 if (OVS_LIKELY(error
!= ENOSPC
)) {
3336 /* XXX: There's a race window where a flow covering this packet
3337 * could have already been installed since we last did the flow
3338 * lookup before upcall. This could be solved by moving the
3339 * mutex lock outside the loop, but that's an awful long time
3340 * to be locking everyone out of making flow installs. If we
3341 * move to a per-core classifier, it would be reasonable. */
3342 ovs_mutex_lock(&pmd
->flow_mutex
);
3343 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, &keys
[i
]);
3344 if (OVS_LIKELY(!netdev_flow
)) {
3345 netdev_flow
= dp_netdev_flow_add(pmd
, &match
, &ufid
,
3349 ovs_mutex_unlock(&pmd
->flow_mutex
);
3351 emc_insert(flow_cache
, &keys
[i
], netdev_flow
);
3355 ofpbuf_uninit(&actions
);
3356 ofpbuf_uninit(&put_actions
);
3357 fat_rwlock_unlock(&dp
->upcall_rwlock
);
3358 dp_netdev_count_packet(pmd
, DP_STAT_LOST
, lost_cnt
);
3359 } else if (OVS_UNLIKELY(any_miss
)) {
3360 for (i
= 0; i
< cnt
; i
++) {
3361 if (OVS_UNLIKELY(!rules
[i
])) {
3362 dp_packet_delete(packets
[i
]);
3369 for (i
= 0; i
< cnt
; i
++) {
3370 struct dp_packet
*packet
= packets
[i
];
3371 struct dp_netdev_flow
*flow
;
3373 if (OVS_UNLIKELY(!rules
[i
])) {
3377 flow
= dp_netdev_flow_cast(rules
[i
]);
3379 emc_insert(flow_cache
, &keys
[i
], flow
);
3380 dp_netdev_queue_batches(packet
, flow
, &keys
[i
].mf
, batches
, n_batches
);
3383 dp_netdev_count_packet(pmd
, DP_STAT_MASKED_HIT
, cnt
- miss_cnt
);
3384 dp_netdev_count_packet(pmd
, DP_STAT_MISS
, miss_cnt
);
3385 dp_netdev_count_packet(pmd
, DP_STAT_LOST
, lost_cnt
);
3389 dp_netdev_input(struct dp_netdev_pmd_thread
*pmd
,
3390 struct dp_packet
**packets
, int cnt
)
3392 #if !defined(__CHECKER__) && !defined(_WIN32)
3393 const size_t PKT_ARRAY_SIZE
= cnt
;
3395 /* Sparse or MSVC doesn't like variable length array. */
3396 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
3398 struct netdev_flow_key keys
[PKT_ARRAY_SIZE
];
3399 struct packet_batch batches
[PKT_ARRAY_SIZE
];
3400 long long now
= time_msec();
3401 size_t newcnt
, n_batches
, i
;
3404 newcnt
= emc_processing(pmd
, packets
, cnt
, keys
, batches
, &n_batches
);
3405 if (OVS_UNLIKELY(newcnt
)) {
3406 fast_path_processing(pmd
, packets
, newcnt
, keys
, batches
, &n_batches
);
3409 for (i
= 0; i
< n_batches
; i
++) {
3410 batches
[i
].flow
->batch
= NULL
;
3413 for (i
= 0; i
< n_batches
; i
++) {
3414 packet_batch_execute(&batches
[i
], pmd
, now
);
3418 struct dp_netdev_execute_aux
{
3419 struct dp_netdev_pmd_thread
*pmd
;
3423 dpif_netdev_register_dp_purge_cb(struct dpif
*dpif
, dp_purge_callback
*cb
,
3426 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3427 dp
->dp_purge_aux
= aux
;
3428 dp
->dp_purge_cb
= cb
;
3432 dpif_netdev_register_upcall_cb(struct dpif
*dpif
, upcall_callback
*cb
,
3435 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3436 dp
->upcall_aux
= aux
;
3441 dp_netdev_drop_packets(struct dp_packet
**packets
, int cnt
, bool may_steal
)
3446 for (i
= 0; i
< cnt
; i
++) {
3447 dp_packet_delete(packets
[i
]);
3453 push_tnl_action(const struct dp_netdev
*dp
,
3454 const struct nlattr
*attr
,
3455 struct dp_packet
**packets
, int cnt
)
3457 struct dp_netdev_port
*tun_port
;
3458 const struct ovs_action_push_tnl
*data
;
3460 data
= nl_attr_get(attr
);
3462 tun_port
= dp_netdev_lookup_port(dp
, u32_to_odp(data
->tnl_port
));
3466 netdev_push_header(tun_port
->netdev
, packets
, cnt
, data
);
3472 dp_netdev_clone_pkt_batch(struct dp_packet
**dst_pkts
,
3473 struct dp_packet
**src_pkts
, int cnt
)
3477 for (i
= 0; i
< cnt
; i
++) {
3478 dst_pkts
[i
] = dp_packet_clone(src_pkts
[i
]);
3483 dp_execute_cb(void *aux_
, struct dp_packet
**packets
, int cnt
,
3484 const struct nlattr
*a
, bool may_steal
)
3485 OVS_NO_THREAD_SAFETY_ANALYSIS
3487 struct dp_netdev_execute_aux
*aux
= aux_
;
3488 uint32_t *depth
= recirc_depth_get();
3489 struct dp_netdev_pmd_thread
*pmd
= aux
->pmd
;
3490 struct dp_netdev
*dp
= pmd
->dp
;
3491 int type
= nl_attr_type(a
);
3492 struct dp_netdev_port
*p
;
3495 switch ((enum ovs_action_attr
)type
) {
3496 case OVS_ACTION_ATTR_OUTPUT
:
3497 p
= dp_netdev_lookup_port(dp
, u32_to_odp(nl_attr_get_u32(a
)));
3498 if (OVS_LIKELY(p
)) {
3499 netdev_send(p
->netdev
, pmd
->tx_qid
, packets
, cnt
, may_steal
);
3504 case OVS_ACTION_ATTR_TUNNEL_PUSH
:
3505 if (*depth
< MAX_RECIRC_DEPTH
) {
3506 struct dp_packet
*tnl_pkt
[NETDEV_MAX_BURST
];
3510 dp_netdev_clone_pkt_batch(tnl_pkt
, packets
, cnt
);
3514 err
= push_tnl_action(dp
, a
, packets
, cnt
);
3517 dp_netdev_input(pmd
, packets
, cnt
);
3520 dp_netdev_drop_packets(tnl_pkt
, cnt
, !may_steal
);
3526 case OVS_ACTION_ATTR_TUNNEL_POP
:
3527 if (*depth
< MAX_RECIRC_DEPTH
) {
3528 odp_port_t portno
= u32_to_odp(nl_attr_get_u32(a
));
3530 p
= dp_netdev_lookup_port(dp
, portno
);
3532 struct dp_packet
*tnl_pkt
[NETDEV_MAX_BURST
];
3536 dp_netdev_clone_pkt_batch(tnl_pkt
, packets
, cnt
);
3540 err
= netdev_pop_header(p
->netdev
, packets
, cnt
);
3543 for (i
= 0; i
< cnt
; i
++) {
3544 packets
[i
]->md
.in_port
.odp_port
= portno
;
3548 dp_netdev_input(pmd
, packets
, cnt
);
3551 dp_netdev_drop_packets(tnl_pkt
, cnt
, !may_steal
);
3558 case OVS_ACTION_ATTR_USERSPACE
:
3559 if (!fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
3560 const struct nlattr
*userdata
;
3561 struct ofpbuf actions
;
3565 userdata
= nl_attr_find_nested(a
, OVS_USERSPACE_ATTR_USERDATA
);
3566 ofpbuf_init(&actions
, 0);
3568 for (i
= 0; i
< cnt
; i
++) {
3571 ofpbuf_clear(&actions
);
3573 flow_extract(packets
[i
], &flow
);
3574 dpif_flow_hash(dp
->dpif
, &flow
, sizeof flow
, &ufid
);
3575 error
= dp_netdev_upcall(pmd
, packets
[i
], &flow
, NULL
, &ufid
,
3576 DPIF_UC_ACTION
, userdata
,&actions
,
3578 if (!error
|| error
== ENOSPC
) {
3579 dp_netdev_execute_actions(pmd
, &packets
[i
], 1, may_steal
,
3580 actions
.data
, actions
.size
);
3581 } else if (may_steal
) {
3582 dp_packet_delete(packets
[i
]);
3585 ofpbuf_uninit(&actions
);
3586 fat_rwlock_unlock(&dp
->upcall_rwlock
);
3592 case OVS_ACTION_ATTR_RECIRC
:
3593 if (*depth
< MAX_RECIRC_DEPTH
) {
3594 struct dp_packet
*recirc_pkts
[NETDEV_MAX_BURST
];
3597 dp_netdev_clone_pkt_batch(recirc_pkts
, packets
, cnt
);
3598 packets
= recirc_pkts
;
3601 for (i
= 0; i
< cnt
; i
++) {
3602 packets
[i
]->md
.recirc_id
= nl_attr_get_u32(a
);
3606 dp_netdev_input(pmd
, packets
, cnt
);
3612 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3615 case OVS_ACTION_ATTR_CT
:
3616 /* If a flow with this action is slow-pathed, datapath assistance is
3617 * required to implement it. However, we don't support this action
3618 * in the userspace datapath. */
3619 VLOG_WARN("Cannot execute conntrack action in userspace.");
3622 case OVS_ACTION_ATTR_PUSH_VLAN
:
3623 case OVS_ACTION_ATTR_POP_VLAN
:
3624 case OVS_ACTION_ATTR_PUSH_MPLS
:
3625 case OVS_ACTION_ATTR_POP_MPLS
:
3626 case OVS_ACTION_ATTR_SET
:
3627 case OVS_ACTION_ATTR_SET_MASKED
:
3628 case OVS_ACTION_ATTR_SAMPLE
:
3629 case OVS_ACTION_ATTR_HASH
:
3630 case OVS_ACTION_ATTR_UNSPEC
:
3631 case __OVS_ACTION_ATTR_MAX
:
3635 dp_netdev_drop_packets(packets
, cnt
, may_steal
);
3639 dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
3640 struct dp_packet
**packets
, int cnt
,
3642 const struct nlattr
*actions
, size_t actions_len
)
3644 struct dp_netdev_execute_aux aux
= { pmd
};
3646 odp_execute_actions(&aux
, packets
, cnt
, may_steal
, actions
,
3647 actions_len
, dp_execute_cb
);
3650 const struct dpif_class dpif_netdev_class
= {
3653 dpif_netdev_enumerate
,
3654 dpif_netdev_port_open_type
,
3657 dpif_netdev_destroy
,
3660 dpif_netdev_get_stats
,
3661 dpif_netdev_port_add
,
3662 dpif_netdev_port_del
,
3663 dpif_netdev_port_query_by_number
,
3664 dpif_netdev_port_query_by_name
,
3665 NULL
, /* port_get_pid */
3666 dpif_netdev_port_dump_start
,
3667 dpif_netdev_port_dump_next
,
3668 dpif_netdev_port_dump_done
,
3669 dpif_netdev_port_poll
,
3670 dpif_netdev_port_poll_wait
,
3671 dpif_netdev_flow_flush
,
3672 dpif_netdev_flow_dump_create
,
3673 dpif_netdev_flow_dump_destroy
,
3674 dpif_netdev_flow_dump_thread_create
,
3675 dpif_netdev_flow_dump_thread_destroy
,
3676 dpif_netdev_flow_dump_next
,
3677 dpif_netdev_operate
,
3678 NULL
, /* recv_set */
3679 NULL
, /* handlers_set */
3680 dpif_netdev_pmd_set
,
3681 dpif_netdev_queue_to_priority
,
3683 NULL
, /* recv_wait */
3684 NULL
, /* recv_purge */
3685 dpif_netdev_register_dp_purge_cb
,
3686 dpif_netdev_register_upcall_cb
,
3687 dpif_netdev_enable_upcall
,
3688 dpif_netdev_disable_upcall
,
3689 dpif_netdev_get_datapath_version
,
3693 dpif_dummy_change_port_number(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
3694 const char *argv
[], void *aux OVS_UNUSED
)
3696 struct dp_netdev_port
*old_port
;
3697 struct dp_netdev_port
*new_port
;
3698 struct dp_netdev
*dp
;
3701 ovs_mutex_lock(&dp_netdev_mutex
);
3702 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
3703 if (!dp
|| !dpif_netdev_class_is_dummy(dp
->class)) {
3704 ovs_mutex_unlock(&dp_netdev_mutex
);
3705 unixctl_command_reply_error(conn
, "unknown datapath or not a dummy");
3708 ovs_refcount_ref(&dp
->ref_cnt
);
3709 ovs_mutex_unlock(&dp_netdev_mutex
);
3711 ovs_mutex_lock(&dp
->port_mutex
);
3712 if (get_port_by_name(dp
, argv
[2], &old_port
)) {
3713 unixctl_command_reply_error(conn
, "unknown port");
3717 port_no
= u32_to_odp(atoi(argv
[3]));
3718 if (!port_no
|| port_no
== ODPP_NONE
) {
3719 unixctl_command_reply_error(conn
, "bad port number");
3722 if (dp_netdev_lookup_port(dp
, port_no
)) {
3723 unixctl_command_reply_error(conn
, "port number already in use");
3727 /* Remove old port. */
3728 cmap_remove(&dp
->ports
, &old_port
->node
, hash_port_no(old_port
->port_no
));
3729 ovsrcu_postpone(free
, old_port
);
3731 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3732 new_port
= xmemdup(old_port
, sizeof *old_port
);
3733 new_port
->port_no
= port_no
;
3734 cmap_insert(&dp
->ports
, &new_port
->node
, hash_port_no(port_no
));
3736 seq_change(dp
->port_seq
);
3737 unixctl_command_reply(conn
, NULL
);
3740 ovs_mutex_unlock(&dp
->port_mutex
);
3741 dp_netdev_unref(dp
);
3745 dpif_dummy_delete_port(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
3746 const char *argv
[], void *aux OVS_UNUSED
)
3748 struct dp_netdev_port
*port
;
3749 struct dp_netdev
*dp
;
3751 ovs_mutex_lock(&dp_netdev_mutex
);
3752 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
3753 if (!dp
|| !dpif_netdev_class_is_dummy(dp
->class)) {
3754 ovs_mutex_unlock(&dp_netdev_mutex
);
3755 unixctl_command_reply_error(conn
, "unknown datapath or not a dummy");
3758 ovs_refcount_ref(&dp
->ref_cnt
);
3759 ovs_mutex_unlock(&dp_netdev_mutex
);
3761 ovs_mutex_lock(&dp
->port_mutex
);
3762 if (get_port_by_name(dp
, argv
[2], &port
)) {
3763 unixctl_command_reply_error(conn
, "unknown port");
3764 } else if (port
->port_no
== ODPP_LOCAL
) {
3765 unixctl_command_reply_error(conn
, "can't delete local port");
3767 do_del_port(dp
, port
);
3768 unixctl_command_reply(conn
, NULL
);
3770 ovs_mutex_unlock(&dp
->port_mutex
);
3772 dp_netdev_unref(dp
);
3776 dpif_dummy_register__(const char *type
)
3778 struct dpif_class
*class;
3780 class = xmalloc(sizeof *class);
3781 *class = dpif_netdev_class
;
3782 class->type
= xstrdup(type
);
3783 dp_register_provider(class);
3787 dpif_dummy_override(const char *type
)
3792 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
3793 * a userland-only build. It's useful for testsuite.
3795 error
= dp_unregister_provider(type
);
3796 if (error
== 0 || error
== EAFNOSUPPORT
) {
3797 dpif_dummy_register__(type
);
3802 dpif_dummy_register(enum dummy_level level
)
3804 if (level
== DUMMY_OVERRIDE_ALL
) {
3809 dp_enumerate_types(&types
);
3810 SSET_FOR_EACH (type
, &types
) {
3811 dpif_dummy_override(type
);
3813 sset_destroy(&types
);
3814 } else if (level
== DUMMY_OVERRIDE_SYSTEM
) {
3815 dpif_dummy_override("system");
3818 dpif_dummy_register__("dummy");
3820 unixctl_command_register("dpif-dummy/change-port-number",
3821 "dp port new-number",
3822 3, 3, dpif_dummy_change_port_number
, NULL
);
3823 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3824 2, 2, dpif_dummy_delete_port
, NULL
);
3827 /* Datapath Classifier. */
3829 /* A set of rules that all have the same fields wildcarded. */
3830 struct dpcls_subtable
{
3831 /* The fields are only used by writers. */
3832 struct cmap_node cmap_node OVS_GUARDED
; /* Within dpcls 'subtables_map'. */
3834 /* These fields are accessed by readers. */
3835 struct cmap rules
; /* Contains "struct dpcls_rule"s. */
3836 struct netdev_flow_key mask
; /* Wildcards for fields (const). */
3837 /* 'mask' must be the last field, additional space is allocated here. */
3840 /* Initializes 'cls' as a classifier that initially contains no classification
3843 dpcls_init(struct dpcls
*cls
)
3845 cmap_init(&cls
->subtables_map
);
3846 pvector_init(&cls
->subtables
);
3850 dpcls_destroy_subtable(struct dpcls
*cls
, struct dpcls_subtable
*subtable
)
3852 pvector_remove(&cls
->subtables
, subtable
);
3853 cmap_remove(&cls
->subtables_map
, &subtable
->cmap_node
,
3854 subtable
->mask
.hash
);
3855 cmap_destroy(&subtable
->rules
);
3856 ovsrcu_postpone(free
, subtable
);
3859 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3860 * caller's responsibility.
3861 * May only be called after all the readers have been terminated. */
3863 dpcls_destroy(struct dpcls
*cls
)
3866 struct dpcls_subtable
*subtable
;
3868 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
3869 ovs_assert(cmap_count(&subtable
->rules
) == 0);
3870 dpcls_destroy_subtable(cls
, subtable
);
3872 cmap_destroy(&cls
->subtables_map
);
3873 pvector_destroy(&cls
->subtables
);
3877 static struct dpcls_subtable
*
3878 dpcls_create_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
3880 struct dpcls_subtable
*subtable
;
3882 /* Need to add one. */
3883 subtable
= xmalloc(sizeof *subtable
3884 - sizeof subtable
->mask
.mf
+ mask
->len
);
3885 cmap_init(&subtable
->rules
);
3886 netdev_flow_key_clone(&subtable
->mask
, mask
);
3887 cmap_insert(&cls
->subtables_map
, &subtable
->cmap_node
, mask
->hash
);
3888 pvector_insert(&cls
->subtables
, subtable
, 0);
3889 pvector_publish(&cls
->subtables
);
3894 static inline struct dpcls_subtable
*
3895 dpcls_find_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
3897 struct dpcls_subtable
*subtable
;
3899 CMAP_FOR_EACH_WITH_HASH (subtable
, cmap_node
, mask
->hash
,
3900 &cls
->subtables_map
) {
3901 if (netdev_flow_key_equal(&subtable
->mask
, mask
)) {
3905 return dpcls_create_subtable(cls
, mask
);
3908 /* Insert 'rule' into 'cls'. */
3910 dpcls_insert(struct dpcls
*cls
, struct dpcls_rule
*rule
,
3911 const struct netdev_flow_key
*mask
)
3913 struct dpcls_subtable
*subtable
= dpcls_find_subtable(cls
, mask
);
3915 rule
->mask
= &subtable
->mask
;
3916 cmap_insert(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
);
3919 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3921 dpcls_remove(struct dpcls
*cls
, struct dpcls_rule
*rule
)
3923 struct dpcls_subtable
*subtable
;
3925 ovs_assert(rule
->mask
);
3927 INIT_CONTAINER(subtable
, rule
->mask
, mask
);
3929 if (cmap_remove(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
)
3931 dpcls_destroy_subtable(cls
, subtable
);
3932 pvector_publish(&cls
->subtables
);
3936 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
3937 * in 'mask' the values in 'key' and 'target' are the same. */
3939 dpcls_rule_matches_key(const struct dpcls_rule
*rule
,
3940 const struct netdev_flow_key
*target
)
3942 const uint64_t *keyp
= miniflow_get_values(&rule
->flow
.mf
);
3943 const uint64_t *maskp
= miniflow_get_values(&rule
->mask
->mf
);
3946 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, target
, rule
->flow
.mf
.map
) {
3947 if (OVS_UNLIKELY((value
& *maskp
++) != *keyp
++)) {
3954 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3955 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3956 * NULL it is skipped.
3958 * This function is optimized for use in the userspace datapath and therefore
3959 * does not implement a lot of features available in the standard
3960 * classifier_lookup() function. Specifically, it does not implement
3961 * priorities, instead returning any rule which matches the flow.
3963 * Returns true if all flows found a corresponding rule. */
3965 dpcls_lookup(const struct dpcls
*cls
, const struct netdev_flow_key keys
[],
3966 struct dpcls_rule
**rules
, const size_t cnt
)
3968 /* The batch size 16 was experimentally found faster than 8 or 32. */
3969 typedef uint16_t map_type
;
3970 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3972 #if !defined(__CHECKER__) && !defined(_WIN32)
3973 const int N_MAPS
= DIV_ROUND_UP(cnt
, MAP_BITS
);
3975 enum { N_MAPS
= DIV_ROUND_UP(NETDEV_MAX_BURST
, MAP_BITS
) };
3977 map_type maps
[N_MAPS
];
3978 struct dpcls_subtable
*subtable
;
3980 memset(maps
, 0xff, sizeof maps
);
3981 if (cnt
% MAP_BITS
) {
3982 maps
[N_MAPS
- 1] >>= MAP_BITS
- cnt
% MAP_BITS
; /* Clear extra bits. */
3984 memset(rules
, 0, cnt
* sizeof *rules
);
3986 PVECTOR_FOR_EACH (subtable
, &cls
->subtables
) {
3987 const struct netdev_flow_key
*mkeys
= keys
;
3988 struct dpcls_rule
**mrules
= rules
;
3989 map_type remains
= 0;
3992 BUILD_ASSERT_DECL(sizeof remains
== sizeof *maps
);
3994 for (m
= 0; m
< N_MAPS
; m
++, mkeys
+= MAP_BITS
, mrules
+= MAP_BITS
) {
3995 uint32_t hashes
[MAP_BITS
];
3996 const struct cmap_node
*nodes
[MAP_BITS
];
3997 unsigned long map
= maps
[m
];
4001 continue; /* Skip empty maps. */
4004 /* Compute hashes for the remaining keys. */
4005 ULLONG_FOR_EACH_1(i
, map
) {
4006 hashes
[i
] = netdev_flow_key_hash_in_mask(&mkeys
[i
],
4010 map
= cmap_find_batch(&subtable
->rules
, map
, hashes
, nodes
);
4011 /* Check results. */
4012 ULLONG_FOR_EACH_1(i
, map
) {
4013 struct dpcls_rule
*rule
;
4015 CMAP_NODE_FOR_EACH (rule
, cmap_node
, nodes
[i
]) {
4016 if (OVS_LIKELY(dpcls_rule_matches_key(rule
, &mkeys
[i
]))) {
4021 ULLONG_SET0(map
, i
); /* Did not match. */
4023 ; /* Keep Sparse happy. */
4025 maps
[m
] &= ~map
; /* Clear the found rules. */
4029 return true; /* All found. */
4032 return false; /* Some misses. */