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>
36 #include "dp-packet.h"
38 #include "dpif-provider.h"
40 #include "dynamic-string.h"
41 #include "fat-rwlock.h"
47 #include "meta-flow.h"
49 #include "netdev-dpdk.h"
50 #include "netdev-vport.h"
52 #include "odp-execute.h"
54 #include "ofp-print.h"
59 #include "poll-loop.h"
66 #include "tnl-arp-cache.h"
69 #include "openvswitch/vlog.h"
71 VLOG_DEFINE_THIS_MODULE(dpif_netdev
);
73 #define FLOW_DUMP_MAX_BATCH 50
74 /* Use per thread recirc_depth to prevent recirculation loop. */
75 #define MAX_RECIRC_DEPTH 5
76 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth
, 0)
78 /* Configuration parameters. */
79 enum { MAX_FLOWS
= 65536 }; /* Maximum number of flows in flow table. */
81 /* Protects against changes to 'dp_netdevs'. */
82 static struct ovs_mutex dp_netdev_mutex
= OVS_MUTEX_INITIALIZER
;
84 /* Contains all 'struct dp_netdev's. */
85 static struct shash dp_netdevs
OVS_GUARDED_BY(dp_netdev_mutex
)
86 = SHASH_INITIALIZER(&dp_netdevs
);
88 static struct vlog_rate_limit upcall_rl
= VLOG_RATE_LIMIT_INIT(600, 600);
90 /* Stores a miniflow with inline values */
92 struct netdev_flow_key
{
93 uint32_t hash
; /* Hash function differs for different users. */
94 uint32_t len
; /* Length of the following miniflow (incl. map). */
96 uint64_t buf
[FLOW_MAX_PACKET_U64S
- MINI_N_INLINE
];
99 /* Exact match cache for frequently used flows
101 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
102 * search its entries for a miniflow that matches exactly the miniflow of the
103 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
105 * A cache entry holds a reference to its 'dp_netdev_flow'.
107 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
108 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
109 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
110 * value is the index of a cache entry where the miniflow could be.
116 * Each pmd_thread has its own private exact match cache.
117 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
120 #define EM_FLOW_HASH_SHIFT 10
121 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
122 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
123 #define EM_FLOW_HASH_SEGS 2
126 struct dp_netdev_flow
*flow
;
127 struct netdev_flow_key key
; /* key.hash used for emc hash value. */
131 struct emc_entry entries
[EM_FLOW_HASH_ENTRIES
];
132 int sweep_idx
; /* For emc_cache_slow_sweep(). */
135 /* Iterate in the exact match cache through every entry that might contain a
136 * miniflow with hash 'HASH'. */
137 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
138 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
139 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
140 i__ < EM_FLOW_HASH_SEGS; \
141 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
143 /* Simple non-wildcarding single-priority classifier. */
146 struct cmap subtables_map
;
147 struct pvector subtables
;
150 /* A rule to be inserted to the classifier. */
152 struct cmap_node cmap_node
; /* Within struct dpcls_subtable 'rules'. */
153 struct netdev_flow_key
*mask
; /* Subtable's mask. */
154 struct netdev_flow_key flow
; /* Matching key. */
155 /* 'flow' must be the last field, additional space is allocated here. */
158 static void dpcls_init(struct dpcls
*);
159 static void dpcls_destroy(struct dpcls
*);
160 static void dpcls_insert(struct dpcls
*, struct dpcls_rule
*,
161 const struct netdev_flow_key
*mask
);
162 static void dpcls_remove(struct dpcls
*, struct dpcls_rule
*);
163 static bool dpcls_lookup(const struct dpcls
*cls
,
164 const struct netdev_flow_key keys
[],
165 struct dpcls_rule
**rules
, size_t cnt
);
167 /* Datapath based on the network device interface from netdev.h.
173 * Some members, marked 'const', are immutable. Accessing other members
174 * requires synchronization, as noted in more detail below.
176 * Acquisition order is, from outermost to innermost:
178 * dp_netdev_mutex (global)
182 const struct dpif_class
*const class;
183 const char *const name
;
185 struct ovs_refcount ref_cnt
;
186 atomic_flag destroyed
;
190 * Protected by RCU. Take the mutex to add or remove ports. */
191 struct ovs_mutex port_mutex
;
193 struct seq
*port_seq
; /* Incremented whenever a port changes. */
195 /* Protects access to ofproto-dpif-upcall interface during revalidator
196 * thread synchronization. */
197 struct fat_rwlock upcall_rwlock
;
198 upcall_callback
*upcall_cb
; /* Callback function for executing upcalls. */
201 /* Stores all 'struct dp_netdev_pmd_thread's. */
202 struct cmap poll_threads
;
204 /* Protects the access of the 'struct dp_netdev_pmd_thread'
205 * instance for non-pmd thread. */
206 struct ovs_mutex non_pmd_mutex
;
208 /* Each pmd thread will store its pointer to
209 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
210 ovsthread_key_t per_pmd_key
;
212 /* Number of rx queues for each dpdk interface and the cpu mask
213 * for pin of pmd threads. */
216 uint64_t last_tnl_conf_seq
;
219 static struct dp_netdev_port
*dp_netdev_lookup_port(const struct dp_netdev
*dp
,
223 DP_STAT_EXACT_HIT
, /* Packets that had an exact match (emc). */
224 DP_STAT_MASKED_HIT
, /* Packets that matched in the flow table. */
225 DP_STAT_MISS
, /* Packets that did not match. */
226 DP_STAT_LOST
, /* Packets not passed up to the client. */
230 enum pmd_cycles_counter_type
{
231 PMD_CYCLES_POLLING
, /* Cycles spent polling NICs. */
232 PMD_CYCLES_PROCESSING
, /* Cycles spent processing packets */
236 /* A port in a netdev-based datapath. */
237 struct dp_netdev_port
{
238 struct cmap_node node
; /* Node in dp_netdev's 'ports'. */
240 struct netdev
*netdev
;
241 struct netdev_saved_flags
*sf
;
242 struct netdev_rxq
**rxq
;
243 struct ovs_refcount ref_cnt
;
244 char *type
; /* Port type as requested by user. */
247 /* Contained by struct dp_netdev_flow's 'stats' member. */
248 struct dp_netdev_flow_stats
{
249 atomic_llong used
; /* Last used time, in monotonic msecs. */
250 atomic_ullong packet_count
; /* Number of packets matched. */
251 atomic_ullong byte_count
; /* Number of bytes matched. */
252 atomic_uint16_t tcp_flags
; /* Bitwise-OR of seen tcp_flags values. */
255 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
261 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
262 * its pmd thread's classifier. The text below calls this classifier 'cls'.
267 * The thread safety rules described here for "struct dp_netdev_flow" are
268 * motivated by two goals:
270 * - Prevent threads that read members of "struct dp_netdev_flow" from
271 * reading bad data due to changes by some thread concurrently modifying
274 * - Prevent two threads making changes to members of a given "struct
275 * dp_netdev_flow" from interfering with each other.
281 * A flow 'flow' may be accessed without a risk of being freed during an RCU
282 * grace period. Code that needs to hold onto a flow for a while
283 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
285 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
286 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
289 * Some members, marked 'const', are immutable. Accessing other members
290 * requires synchronization, as noted in more detail below.
292 struct dp_netdev_flow
{
295 /* Hash table index by unmasked flow. */
296 const struct cmap_node node
; /* In owning dp_netdev_pmd_thread's */
298 const ovs_u128 ufid
; /* Unique flow identifier. */
299 const struct flow flow
; /* Unmasked flow that created this entry. */
300 const int pmd_id
; /* The 'core_id' of pmd thread owning this */
303 /* Number of references.
304 * The classifier owns one reference.
305 * Any thread trying to keep a rule from being freed should hold its own
307 struct ovs_refcount ref_cnt
;
310 struct dp_netdev_flow_stats stats
;
313 OVSRCU_TYPE(struct dp_netdev_actions
*) actions
;
315 /* Packet classification. */
316 struct dpcls_rule cr
; /* In owning dp_netdev's 'cls'. */
317 /* 'cr' must be the last member. */
320 static void dp_netdev_flow_unref(struct dp_netdev_flow
*);
321 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*);
322 static int dpif_netdev_flow_from_nlattrs(const struct nlattr
*, uint32_t,
325 /* A set of datapath actions within a "struct dp_netdev_flow".
331 * A struct dp_netdev_actions 'actions' is protected with RCU. */
332 struct dp_netdev_actions
{
333 /* These members are immutable: they do not change during the struct's
335 unsigned int size
; /* Size of 'actions', in bytes. */
336 struct nlattr actions
[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
339 struct dp_netdev_actions
*dp_netdev_actions_create(const struct nlattr
*,
341 struct dp_netdev_actions
*dp_netdev_flow_get_actions(
342 const struct dp_netdev_flow
*);
343 static void dp_netdev_actions_free(struct dp_netdev_actions
*);
345 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
346 struct dp_netdev_pmd_stats
{
347 /* Indexed by DP_STAT_*. */
348 atomic_ullong n
[DP_N_STATS
];
351 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
352 struct dp_netdev_pmd_cycles
{
353 /* Indexed by PMD_CYCLES_*. */
354 atomic_ullong n
[PMD_N_CYCLES
];
357 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
358 * the performance overhead of interrupt processing. Therefore netdev can
359 * not implement rx-wait for these devices. dpif-netdev needs to poll
360 * these device to check for recv buffer. pmd-thread does polling for
361 * devices assigned to itself.
363 * DPDK used PMD for accessing NIC.
365 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
366 * I/O of all non-pmd threads. There will be no actual thread created
369 * Each struct has its own flow table and classifier. Packets received
370 * from managed ports are looked up in the corresponding pmd thread's
371 * flow table, and are executed with the found actions.
373 struct dp_netdev_pmd_thread
{
374 struct dp_netdev
*dp
;
375 struct ovs_refcount ref_cnt
; /* Every reference must be refcount'ed. */
376 struct cmap_node node
; /* In 'dp->poll_threads'. */
378 pthread_cond_t cond
; /* For synchronizing pmd thread reload. */
379 struct ovs_mutex cond_mutex
; /* Mutex for condition variable. */
381 /* Per thread exact-match cache. Note, the instance for cpu core
382 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
383 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
384 * instances will only be accessed by its own pmd thread. */
385 struct emc_cache flow_cache
;
387 /* Classifier and Flow-Table.
389 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
390 * changes to 'cls' must be made while still holding the 'flow_mutex'.
392 struct ovs_mutex flow_mutex
;
394 struct cmap flow_table OVS_GUARDED
; /* Flow table. */
397 struct dp_netdev_pmd_stats stats
;
399 /* Cycles counters */
400 struct dp_netdev_pmd_cycles cycles
;
402 /* Used to count cicles. See 'cycles_counter_end()' */
403 unsigned long long last_cycles
;
405 struct latch exit_latch
; /* For terminating the pmd thread. */
406 atomic_uint change_seq
; /* For reloading pmd ports. */
408 int index
; /* Idx of this pmd thread among pmd*/
409 /* threads on same numa node. */
410 int core_id
; /* CPU core id of this pmd thread. */
411 int numa_id
; /* numa node id of this pmd thread. */
413 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
414 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
415 * values and subtracts them from 'stats' and 'cycles' before
416 * reporting to the user */
417 unsigned long long stats_zero
[DP_N_STATS
];
418 uint64_t cycles_zero
[PMD_N_CYCLES
];
421 #define PMD_INITIAL_SEQ 1
423 /* Interface to netdev-based datapath. */
426 struct dp_netdev
*dp
;
427 uint64_t last_port_seq
;
430 static int get_port_by_number(struct dp_netdev
*dp
, odp_port_t port_no
,
431 struct dp_netdev_port
**portp
);
432 static int get_port_by_name(struct dp_netdev
*dp
, const char *devname
,
433 struct dp_netdev_port
**portp
);
434 static void dp_netdev_free(struct dp_netdev
*)
435 OVS_REQUIRES(dp_netdev_mutex
);
436 static int do_add_port(struct dp_netdev
*dp
, const char *devname
,
437 const char *type
, odp_port_t port_no
)
438 OVS_REQUIRES(dp
->port_mutex
);
439 static void do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*)
440 OVS_REQUIRES(dp
->port_mutex
);
441 static int dpif_netdev_open(const struct dpif_class
*, const char *name
,
442 bool create
, struct dpif
**);
443 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
444 struct dp_packet
**, int c
,
446 const struct nlattr
*actions
,
448 static void dp_netdev_input(struct dp_netdev_pmd_thread
*,
449 struct dp_packet
**, int cnt
);
451 static void dp_netdev_disable_upcall(struct dp_netdev
*);
452 void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
);
453 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
,
454 struct dp_netdev
*dp
, int index
,
455 int core_id
, int numa_id
);
456 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
);
457 static void dp_netdev_set_nonpmd(struct dp_netdev
*dp
);
458 static struct dp_netdev_pmd_thread
*dp_netdev_get_pmd(struct dp_netdev
*dp
,
460 static struct dp_netdev_pmd_thread
*
461 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
);
462 static void dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
);
463 static void dp_netdev_del_pmds_on_numa(struct dp_netdev
*dp
, int numa_id
);
464 static void dp_netdev_set_pmds_on_numa(struct dp_netdev
*dp
, int numa_id
);
465 static void dp_netdev_reset_pmd_threads(struct dp_netdev
*dp
);
466 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
);
467 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
);
468 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
);
470 static inline bool emc_entry_alive(struct emc_entry
*ce
);
471 static void emc_clear_entry(struct emc_entry
*ce
);
474 emc_cache_init(struct emc_cache
*flow_cache
)
478 BUILD_ASSERT(offsetof(struct miniflow
, inline_values
) == sizeof(uint64_t));
480 flow_cache
->sweep_idx
= 0;
481 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
482 flow_cache
->entries
[i
].flow
= NULL
;
483 flow_cache
->entries
[i
].key
.hash
= 0;
484 flow_cache
->entries
[i
].key
.len
485 = offsetof(struct miniflow
, inline_values
);
486 miniflow_initialize(&flow_cache
->entries
[i
].key
.mf
,
487 flow_cache
->entries
[i
].key
.buf
);
492 emc_cache_uninit(struct emc_cache
*flow_cache
)
496 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
497 emc_clear_entry(&flow_cache
->entries
[i
]);
501 /* Check and clear dead flow references slowly (one entry at each
504 emc_cache_slow_sweep(struct emc_cache
*flow_cache
)
506 struct emc_entry
*entry
= &flow_cache
->entries
[flow_cache
->sweep_idx
];
508 if (!emc_entry_alive(entry
)) {
509 emc_clear_entry(entry
);
511 flow_cache
->sweep_idx
= (flow_cache
->sweep_idx
+ 1) & EM_FLOW_HASH_MASK
;
514 static struct dpif_netdev
*
515 dpif_netdev_cast(const struct dpif
*dpif
)
517 ovs_assert(dpif
->dpif_class
->open
== dpif_netdev_open
);
518 return CONTAINER_OF(dpif
, struct dpif_netdev
, dpif
);
521 static struct dp_netdev
*
522 get_dp_netdev(const struct dpif
*dpif
)
524 return dpif_netdev_cast(dpif
)->dp
;
528 PMD_INFO_SHOW_STATS
, /* show how cpu cycles are spent */
529 PMD_INFO_CLEAR_STATS
/* set the cycles count to 0 */
533 pmd_info_show_stats(struct ds
*reply
,
534 struct dp_netdev_pmd_thread
*pmd
,
535 unsigned long long stats
[DP_N_STATS
],
536 uint64_t cycles
[PMD_N_CYCLES
])
538 unsigned long long total_packets
= 0;
539 uint64_t total_cycles
= 0;
542 /* These loops subtracts reference values ('*_zero') from the counters.
543 * Since loads and stores are relaxed, it might be possible for a '*_zero'
544 * value to be more recent than the current value we're reading from the
545 * counter. This is not a big problem, since these numbers are not
546 * supposed to be too accurate, but we should at least make sure that
547 * the result is not negative. */
548 for (i
= 0; i
< DP_N_STATS
; i
++) {
549 if (stats
[i
] > pmd
->stats_zero
[i
]) {
550 stats
[i
] -= pmd
->stats_zero
[i
];
555 if (i
!= DP_STAT_LOST
) {
556 /* Lost packets are already included in DP_STAT_MISS */
557 total_packets
+= stats
[i
];
561 for (i
= 0; i
< PMD_N_CYCLES
; i
++) {
562 if (cycles
[i
] > pmd
->cycles_zero
[i
]) {
563 cycles
[i
] -= pmd
->cycles_zero
[i
];
568 total_cycles
+= cycles
[i
];
571 ds_put_cstr(reply
, (pmd
->core_id
== NON_PMD_CORE_ID
)
572 ? "main thread" : "pmd thread");
574 if (pmd
->numa_id
!= OVS_NUMA_UNSPEC
) {
575 ds_put_format(reply
, " numa_id %d", pmd
->numa_id
);
577 if (pmd
->core_id
!= OVS_CORE_UNSPEC
) {
578 ds_put_format(reply
, " core_id %d", pmd
->core_id
);
580 ds_put_cstr(reply
, ":\n");
583 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
584 "\tmiss:%llu\n\tlost:%llu\n",
585 stats
[DP_STAT_EXACT_HIT
], stats
[DP_STAT_MASKED_HIT
],
586 stats
[DP_STAT_MISS
], stats
[DP_STAT_LOST
]);
588 if (total_cycles
== 0) {
593 "\tpolling cycles:%"PRIu64
" (%.02f%%)\n"
594 "\tprocessing cycles:%"PRIu64
" (%.02f%%)\n",
595 cycles
[PMD_CYCLES_POLLING
],
596 cycles
[PMD_CYCLES_POLLING
] / (double)total_cycles
* 100,
597 cycles
[PMD_CYCLES_PROCESSING
],
598 cycles
[PMD_CYCLES_PROCESSING
] / (double)total_cycles
* 100);
600 if (total_packets
== 0) {
605 "\tavg cycles per packet: %.02f (%"PRIu64
"/%llu)\n",
606 total_cycles
/ (double)total_packets
,
607 total_cycles
, total_packets
);
610 "\tavg processing cycles per packet: "
611 "%.02f (%"PRIu64
"/%llu)\n",
612 cycles
[PMD_CYCLES_PROCESSING
] / (double)total_packets
,
613 cycles
[PMD_CYCLES_PROCESSING
], total_packets
);
617 pmd_info_clear_stats(struct ds
*reply OVS_UNUSED
,
618 struct dp_netdev_pmd_thread
*pmd
,
619 unsigned long long stats
[DP_N_STATS
],
620 uint64_t cycles
[PMD_N_CYCLES
])
624 /* We cannot write 'stats' and 'cycles' (because they're written by other
625 * threads) and we shouldn't change 'stats' (because they're used to count
626 * datapath stats, which must not be cleared here). Instead, we save the
627 * current values and subtract them from the values to be displayed in the
629 for (i
= 0; i
< DP_N_STATS
; i
++) {
630 pmd
->stats_zero
[i
] = stats
[i
];
632 for (i
= 0; i
< PMD_N_CYCLES
; i
++) {
633 pmd
->cycles_zero
[i
] = cycles
[i
];
638 dpif_netdev_pmd_info(struct unixctl_conn
*conn
, int argc
, const char *argv
[],
641 struct ds reply
= DS_EMPTY_INITIALIZER
;
642 struct dp_netdev_pmd_thread
*pmd
;
643 struct dp_netdev
*dp
= NULL
;
644 enum pmd_info_type type
= *(enum pmd_info_type
*) aux
;
646 ovs_mutex_lock(&dp_netdev_mutex
);
649 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
650 } else if (shash_count(&dp_netdevs
) == 1) {
651 /* There's only one datapath */
652 dp
= shash_first(&dp_netdevs
)->data
;
656 ovs_mutex_unlock(&dp_netdev_mutex
);
657 unixctl_command_reply_error(conn
,
658 "please specify an existing datapath");
662 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
663 unsigned long long stats
[DP_N_STATS
];
664 uint64_t cycles
[PMD_N_CYCLES
];
667 /* Read current stats and cycle counters */
668 for (i
= 0; i
< ARRAY_SIZE(stats
); i
++) {
669 atomic_read_relaxed(&pmd
->stats
.n
[i
], &stats
[i
]);
671 for (i
= 0; i
< ARRAY_SIZE(cycles
); i
++) {
672 atomic_read_relaxed(&pmd
->cycles
.n
[i
], &cycles
[i
]);
675 if (type
== PMD_INFO_CLEAR_STATS
) {
676 pmd_info_clear_stats(&reply
, pmd
, stats
, cycles
);
677 } else if (type
== PMD_INFO_SHOW_STATS
) {
678 pmd_info_show_stats(&reply
, pmd
, stats
, cycles
);
682 ovs_mutex_unlock(&dp_netdev_mutex
);
684 unixctl_command_reply(conn
, ds_cstr(&reply
));
689 dpif_netdev_init(void)
691 static enum pmd_info_type show_aux
= PMD_INFO_SHOW_STATS
,
692 clear_aux
= PMD_INFO_CLEAR_STATS
;
694 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
695 0, 1, dpif_netdev_pmd_info
,
697 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
698 0, 1, dpif_netdev_pmd_info
,
704 dpif_netdev_enumerate(struct sset
*all_dps
,
705 const struct dpif_class
*dpif_class
)
707 struct shash_node
*node
;
709 ovs_mutex_lock(&dp_netdev_mutex
);
710 SHASH_FOR_EACH(node
, &dp_netdevs
) {
711 struct dp_netdev
*dp
= node
->data
;
712 if (dpif_class
!= dp
->class) {
713 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
714 * If the class doesn't match, skip this dpif. */
717 sset_add(all_dps
, node
->name
);
719 ovs_mutex_unlock(&dp_netdev_mutex
);
725 dpif_netdev_class_is_dummy(const struct dpif_class
*class)
727 return class != &dpif_netdev_class
;
731 dpif_netdev_port_open_type(const struct dpif_class
*class, const char *type
)
733 return strcmp(type
, "internal") ? type
734 : dpif_netdev_class_is_dummy(class) ? "dummy"
739 create_dpif_netdev(struct dp_netdev
*dp
)
741 uint16_t netflow_id
= hash_string(dp
->name
, 0);
742 struct dpif_netdev
*dpif
;
744 ovs_refcount_ref(&dp
->ref_cnt
);
746 dpif
= xmalloc(sizeof *dpif
);
747 dpif_init(&dpif
->dpif
, dp
->class, dp
->name
, netflow_id
>> 8, netflow_id
);
749 dpif
->last_port_seq
= seq_read(dp
->port_seq
);
754 /* Choose an unused, non-zero port number and return it on success.
755 * Return ODPP_NONE on failure. */
757 choose_port(struct dp_netdev
*dp
, const char *name
)
758 OVS_REQUIRES(dp
->port_mutex
)
762 if (dp
->class != &dpif_netdev_class
) {
766 /* If the port name begins with "br", start the number search at
767 * 100 to make writing tests easier. */
768 if (!strncmp(name
, "br", 2)) {
772 /* If the port name contains a number, try to assign that port number.
773 * This can make writing unit tests easier because port numbers are
775 for (p
= name
; *p
!= '\0'; p
++) {
776 if (isdigit((unsigned char) *p
)) {
777 port_no
= start_no
+ strtol(p
, NULL
, 10);
778 if (port_no
> 0 && port_no
!= odp_to_u32(ODPP_NONE
)
779 && !dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
780 return u32_to_odp(port_no
);
787 for (port_no
= 1; port_no
<= UINT16_MAX
; port_no
++) {
788 if (!dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
789 return u32_to_odp(port_no
);
797 create_dp_netdev(const char *name
, const struct dpif_class
*class,
798 struct dp_netdev
**dpp
)
799 OVS_REQUIRES(dp_netdev_mutex
)
801 struct dp_netdev
*dp
;
804 dp
= xzalloc(sizeof *dp
);
805 shash_add(&dp_netdevs
, name
, dp
);
807 *CONST_CAST(const struct dpif_class
**, &dp
->class) = class;
808 *CONST_CAST(const char **, &dp
->name
) = xstrdup(name
);
809 ovs_refcount_init(&dp
->ref_cnt
);
810 atomic_flag_clear(&dp
->destroyed
);
812 ovs_mutex_init(&dp
->port_mutex
);
813 cmap_init(&dp
->ports
);
814 dp
->port_seq
= seq_create();
815 fat_rwlock_init(&dp
->upcall_rwlock
);
817 /* Disable upcalls by default. */
818 dp_netdev_disable_upcall(dp
);
819 dp
->upcall_aux
= NULL
;
820 dp
->upcall_cb
= NULL
;
822 cmap_init(&dp
->poll_threads
);
823 ovs_mutex_init_recursive(&dp
->non_pmd_mutex
);
824 ovsthread_key_create(&dp
->per_pmd_key
, NULL
);
826 /* Reserves the core NON_PMD_CORE_ID for all non-pmd threads. */
827 ovs_numa_try_pin_core_specific(NON_PMD_CORE_ID
);
828 dp_netdev_set_nonpmd(dp
);
829 dp
->n_dpdk_rxqs
= NR_QUEUE
;
831 ovs_mutex_lock(&dp
->port_mutex
);
832 error
= do_add_port(dp
, name
, "internal", ODPP_LOCAL
);
833 ovs_mutex_unlock(&dp
->port_mutex
);
839 dp
->last_tnl_conf_seq
= seq_read(tnl_conf_seq
);
845 dpif_netdev_open(const struct dpif_class
*class, const char *name
,
846 bool create
, struct dpif
**dpifp
)
848 struct dp_netdev
*dp
;
851 ovs_mutex_lock(&dp_netdev_mutex
);
852 dp
= shash_find_data(&dp_netdevs
, name
);
854 error
= create
? create_dp_netdev(name
, class, &dp
) : ENODEV
;
856 error
= (dp
->class != class ? EINVAL
861 *dpifp
= create_dpif_netdev(dp
);
864 ovs_mutex_unlock(&dp_netdev_mutex
);
870 dp_netdev_destroy_upcall_lock(struct dp_netdev
*dp
)
871 OVS_NO_THREAD_SAFETY_ANALYSIS
873 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
874 ovs_assert(fat_rwlock_tryrdlock(&dp
->upcall_rwlock
));
876 /* Before freeing a lock we should release it */
877 fat_rwlock_unlock(&dp
->upcall_rwlock
);
878 fat_rwlock_destroy(&dp
->upcall_rwlock
);
881 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
882 * through the 'dp_netdevs' shash while freeing 'dp'. */
884 dp_netdev_free(struct dp_netdev
*dp
)
885 OVS_REQUIRES(dp_netdev_mutex
)
887 struct dp_netdev_port
*port
;
889 shash_find_and_delete(&dp_netdevs
, dp
->name
);
891 dp_netdev_destroy_all_pmds(dp
);
892 cmap_destroy(&dp
->poll_threads
);
893 ovs_mutex_destroy(&dp
->non_pmd_mutex
);
894 ovsthread_key_delete(dp
->per_pmd_key
);
896 ovs_mutex_lock(&dp
->port_mutex
);
897 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
898 do_del_port(dp
, port
);
900 ovs_mutex_unlock(&dp
->port_mutex
);
902 seq_destroy(dp
->port_seq
);
903 cmap_destroy(&dp
->ports
);
905 /* Upcalls must be disabled at this point */
906 dp_netdev_destroy_upcall_lock(dp
);
909 free(CONST_CAST(char *, dp
->name
));
914 dp_netdev_unref(struct dp_netdev
*dp
)
917 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
918 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
919 ovs_mutex_lock(&dp_netdev_mutex
);
920 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
923 ovs_mutex_unlock(&dp_netdev_mutex
);
928 dpif_netdev_close(struct dpif
*dpif
)
930 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
937 dpif_netdev_destroy(struct dpif
*dpif
)
939 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
941 if (!atomic_flag_test_and_set(&dp
->destroyed
)) {
942 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
943 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
951 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
952 * load/store semantics. While the increment is not atomic, the load and
953 * store operations are, making it impossible to read inconsistent values.
955 * This is used to update thread local stats counters. */
957 non_atomic_ullong_add(atomic_ullong
*var
, unsigned long long n
)
959 unsigned long long tmp
;
961 atomic_read_relaxed(var
, &tmp
);
963 atomic_store_relaxed(var
, tmp
);
967 dpif_netdev_get_stats(const struct dpif
*dpif
, struct dpif_dp_stats
*stats
)
969 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
970 struct dp_netdev_pmd_thread
*pmd
;
972 stats
->n_flows
= stats
->n_hit
= stats
->n_missed
= stats
->n_lost
= 0;
973 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
974 unsigned long long n
;
975 stats
->n_flows
+= cmap_count(&pmd
->flow_table
);
977 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_MASKED_HIT
], &n
);
979 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_EXACT_HIT
], &n
);
981 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_MISS
], &n
);
982 stats
->n_missed
+= n
;
983 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_LOST
], &n
);
986 stats
->n_masks
= UINT32_MAX
;
987 stats
->n_mask_hit
= UINT64_MAX
;
993 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread
*pmd
)
997 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
1001 ovs_mutex_lock(&pmd
->cond_mutex
);
1002 atomic_add_relaxed(&pmd
->change_seq
, 1, &old_seq
);
1003 ovs_mutex_cond_wait(&pmd
->cond
, &pmd
->cond_mutex
);
1004 ovs_mutex_unlock(&pmd
->cond_mutex
);
1007 /* Causes all pmd threads to reload its tx/rx devices.
1008 * Must be called after adding/removing ports. */
1010 dp_netdev_reload_pmds(struct dp_netdev
*dp
)
1012 struct dp_netdev_pmd_thread
*pmd
;
1014 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1015 dp_netdev_reload_pmd__(pmd
);
1020 hash_port_no(odp_port_t port_no
)
1022 return hash_int(odp_to_u32(port_no
), 0);
1026 do_add_port(struct dp_netdev
*dp
, const char *devname
, const char *type
,
1028 OVS_REQUIRES(dp
->port_mutex
)
1030 struct netdev_saved_flags
*sf
;
1031 struct dp_netdev_port
*port
;
1032 struct netdev
*netdev
;
1033 enum netdev_flags flags
;
1034 const char *open_type
;
1038 /* Reject devices already in 'dp'. */
1039 if (!get_port_by_name(dp
, devname
, &port
)) {
1043 /* Open and validate network device. */
1044 open_type
= dpif_netdev_port_open_type(dp
->class, type
);
1045 error
= netdev_open(devname
, open_type
, &netdev
);
1049 /* XXX reject non-Ethernet devices */
1051 netdev_get_flags(netdev
, &flags
);
1052 if (flags
& NETDEV_LOOPBACK
) {
1053 VLOG_ERR("%s: cannot add a loopback device", devname
);
1054 netdev_close(netdev
);
1058 if (netdev_is_pmd(netdev
)) {
1059 int n_cores
= ovs_numa_get_n_cores();
1061 if (n_cores
== OVS_CORE_UNSPEC
) {
1062 VLOG_ERR("%s, cannot get cpu core info", devname
);
1065 /* There can only be ovs_numa_get_n_cores() pmd threads,
1066 * so creates a txq for each. */
1067 error
= netdev_set_multiq(netdev
, n_cores
, dp
->n_dpdk_rxqs
);
1068 if (error
&& (error
!= EOPNOTSUPP
)) {
1069 VLOG_ERR("%s, cannot set multiq", devname
);
1073 port
= xzalloc(sizeof *port
);
1074 port
->port_no
= port_no
;
1075 port
->netdev
= netdev
;
1076 port
->rxq
= xmalloc(sizeof *port
->rxq
* netdev_n_rxq(netdev
));
1077 port
->type
= xstrdup(type
);
1078 for (i
= 0; i
< netdev_n_rxq(netdev
); i
++) {
1079 error
= netdev_rxq_open(netdev
, &port
->rxq
[i
], i
);
1081 && !(error
== EOPNOTSUPP
&& dpif_netdev_class_is_dummy(dp
->class))) {
1082 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1083 devname
, ovs_strerror(errno
));
1084 netdev_close(netdev
);
1092 error
= netdev_turn_flags_on(netdev
, NETDEV_PROMISC
, &sf
);
1094 for (i
= 0; i
< netdev_n_rxq(netdev
); i
++) {
1095 netdev_rxq_close(port
->rxq
[i
]);
1097 netdev_close(netdev
);
1105 ovs_refcount_init(&port
->ref_cnt
);
1106 cmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
1108 if (netdev_is_pmd(netdev
)) {
1109 dp_netdev_set_pmds_on_numa(dp
, netdev_get_numa_id(netdev
));
1110 dp_netdev_reload_pmds(dp
);
1112 seq_change(dp
->port_seq
);
1118 dpif_netdev_port_add(struct dpif
*dpif
, struct netdev
*netdev
,
1119 odp_port_t
*port_nop
)
1121 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1122 char namebuf
[NETDEV_VPORT_NAME_BUFSIZE
];
1123 const char *dpif_port
;
1127 ovs_mutex_lock(&dp
->port_mutex
);
1128 dpif_port
= netdev_vport_get_dpif_port(netdev
, namebuf
, sizeof namebuf
);
1129 if (*port_nop
!= ODPP_NONE
) {
1130 port_no
= *port_nop
;
1131 error
= dp_netdev_lookup_port(dp
, *port_nop
) ? EBUSY
: 0;
1133 port_no
= choose_port(dp
, dpif_port
);
1134 error
= port_no
== ODPP_NONE
? EFBIG
: 0;
1137 *port_nop
= port_no
;
1138 error
= do_add_port(dp
, dpif_port
, netdev_get_type(netdev
), port_no
);
1140 ovs_mutex_unlock(&dp
->port_mutex
);
1146 dpif_netdev_port_del(struct dpif
*dpif
, odp_port_t port_no
)
1148 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1151 ovs_mutex_lock(&dp
->port_mutex
);
1152 if (port_no
== ODPP_LOCAL
) {
1155 struct dp_netdev_port
*port
;
1157 error
= get_port_by_number(dp
, port_no
, &port
);
1159 do_del_port(dp
, port
);
1162 ovs_mutex_unlock(&dp
->port_mutex
);
1168 is_valid_port_number(odp_port_t port_no
)
1170 return port_no
!= ODPP_NONE
;
1173 static struct dp_netdev_port
*
1174 dp_netdev_lookup_port(const struct dp_netdev
*dp
, odp_port_t port_no
)
1176 struct dp_netdev_port
*port
;
1178 CMAP_FOR_EACH_WITH_HASH (port
, node
, hash_port_no(port_no
), &dp
->ports
) {
1179 if (port
->port_no
== port_no
) {
1187 get_port_by_number(struct dp_netdev
*dp
,
1188 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1190 if (!is_valid_port_number(port_no
)) {
1194 *portp
= dp_netdev_lookup_port(dp
, port_no
);
1195 return *portp
? 0 : ENOENT
;
1200 port_ref(struct dp_netdev_port
*port
)
1203 ovs_refcount_ref(&port
->ref_cnt
);
1208 port_try_ref(struct dp_netdev_port
*port
)
1211 return ovs_refcount_try_ref_rcu(&port
->ref_cnt
);
1218 port_unref(struct dp_netdev_port
*port
)
1220 if (port
&& ovs_refcount_unref_relaxed(&port
->ref_cnt
) == 1) {
1221 int n_rxq
= netdev_n_rxq(port
->netdev
);
1224 netdev_close(port
->netdev
);
1225 netdev_restore_flags(port
->sf
);
1227 for (i
= 0; i
< n_rxq
; i
++) {
1228 netdev_rxq_close(port
->rxq
[i
]);
1237 get_port_by_name(struct dp_netdev
*dp
,
1238 const char *devname
, struct dp_netdev_port
**portp
)
1239 OVS_REQUIRES(dp
->port_mutex
)
1241 struct dp_netdev_port
*port
;
1243 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1244 if (!strcmp(netdev_get_name(port
->netdev
), devname
)) {
1253 get_n_pmd_threads_on_numa(struct dp_netdev
*dp
, int numa_id
)
1255 struct dp_netdev_pmd_thread
*pmd
;
1258 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1259 if (pmd
->numa_id
== numa_id
) {
1267 /* Returns 'true' if there is a port with pmd netdev and the netdev
1268 * is on numa node 'numa_id'. */
1270 has_pmd_port_for_numa(struct dp_netdev
*dp
, int numa_id
)
1272 struct dp_netdev_port
*port
;
1274 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1275 if (netdev_is_pmd(port
->netdev
)
1276 && netdev_get_numa_id(port
->netdev
) == numa_id
) {
1286 do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*port
)
1287 OVS_REQUIRES(dp
->port_mutex
)
1289 cmap_remove(&dp
->ports
, &port
->node
, hash_odp_port(port
->port_no
));
1290 seq_change(dp
->port_seq
);
1291 if (netdev_is_pmd(port
->netdev
)) {
1292 int numa_id
= netdev_get_numa_id(port
->netdev
);
1294 /* If there is no netdev on the numa node, deletes the pmd threads
1295 * for that numa. Else, just reloads the queues. */
1296 if (!has_pmd_port_for_numa(dp
, numa_id
)) {
1297 dp_netdev_del_pmds_on_numa(dp
, numa_id
);
1299 dp_netdev_reload_pmds(dp
);
1306 answer_port_query(const struct dp_netdev_port
*port
,
1307 struct dpif_port
*dpif_port
)
1309 dpif_port
->name
= xstrdup(netdev_get_name(port
->netdev
));
1310 dpif_port
->type
= xstrdup(port
->type
);
1311 dpif_port
->port_no
= port
->port_no
;
1315 dpif_netdev_port_query_by_number(const struct dpif
*dpif
, odp_port_t port_no
,
1316 struct dpif_port
*dpif_port
)
1318 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1319 struct dp_netdev_port
*port
;
1322 error
= get_port_by_number(dp
, port_no
, &port
);
1323 if (!error
&& dpif_port
) {
1324 answer_port_query(port
, dpif_port
);
1331 dpif_netdev_port_query_by_name(const struct dpif
*dpif
, const char *devname
,
1332 struct dpif_port
*dpif_port
)
1334 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1335 struct dp_netdev_port
*port
;
1338 ovs_mutex_lock(&dp
->port_mutex
);
1339 error
= get_port_by_name(dp
, devname
, &port
);
1340 if (!error
&& dpif_port
) {
1341 answer_port_query(port
, dpif_port
);
1343 ovs_mutex_unlock(&dp
->port_mutex
);
1349 dp_netdev_flow_free(struct dp_netdev_flow
*flow
)
1351 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow
));
1355 static void dp_netdev_flow_unref(struct dp_netdev_flow
*flow
)
1357 if (ovs_refcount_unref_relaxed(&flow
->ref_cnt
) == 1) {
1358 ovsrcu_postpone(dp_netdev_flow_free
, flow
);
1363 dp_netdev_flow_hash(const ovs_u128
*ufid
)
1365 return ufid
->u32
[0];
1369 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread
*pmd
,
1370 struct dp_netdev_flow
*flow
)
1371 OVS_REQUIRES(pmd
->flow_mutex
)
1373 struct cmap_node
*node
= CONST_CAST(struct cmap_node
*, &flow
->node
);
1375 dpcls_remove(&pmd
->cls
, &flow
->cr
);
1376 cmap_remove(&pmd
->flow_table
, node
, dp_netdev_flow_hash(&flow
->ufid
));
1379 dp_netdev_flow_unref(flow
);
1383 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
)
1385 struct dp_netdev_flow
*netdev_flow
;
1387 ovs_mutex_lock(&pmd
->flow_mutex
);
1388 CMAP_FOR_EACH (netdev_flow
, node
, &pmd
->flow_table
) {
1389 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
1391 ovs_mutex_unlock(&pmd
->flow_mutex
);
1395 dpif_netdev_flow_flush(struct dpif
*dpif
)
1397 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1398 struct dp_netdev_pmd_thread
*pmd
;
1400 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1401 dp_netdev_pmd_flow_flush(pmd
);
1407 struct dp_netdev_port_state
{
1408 struct cmap_position position
;
1413 dpif_netdev_port_dump_start(const struct dpif
*dpif OVS_UNUSED
, void **statep
)
1415 *statep
= xzalloc(sizeof(struct dp_netdev_port_state
));
1420 dpif_netdev_port_dump_next(const struct dpif
*dpif
, void *state_
,
1421 struct dpif_port
*dpif_port
)
1423 struct dp_netdev_port_state
*state
= state_
;
1424 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1425 struct cmap_node
*node
;
1428 node
= cmap_next_position(&dp
->ports
, &state
->position
);
1430 struct dp_netdev_port
*port
;
1432 port
= CONTAINER_OF(node
, struct dp_netdev_port
, node
);
1435 state
->name
= xstrdup(netdev_get_name(port
->netdev
));
1436 dpif_port
->name
= state
->name
;
1437 dpif_port
->type
= port
->type
;
1438 dpif_port
->port_no
= port
->port_no
;
1449 dpif_netdev_port_dump_done(const struct dpif
*dpif OVS_UNUSED
, void *state_
)
1451 struct dp_netdev_port_state
*state
= state_
;
1458 dpif_netdev_port_poll(const struct dpif
*dpif_
, char **devnamep OVS_UNUSED
)
1460 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
1461 uint64_t new_port_seq
;
1464 new_port_seq
= seq_read(dpif
->dp
->port_seq
);
1465 if (dpif
->last_port_seq
!= new_port_seq
) {
1466 dpif
->last_port_seq
= new_port_seq
;
1476 dpif_netdev_port_poll_wait(const struct dpif
*dpif_
)
1478 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
1480 seq_wait(dpif
->dp
->port_seq
, dpif
->last_port_seq
);
1483 static struct dp_netdev_flow
*
1484 dp_netdev_flow_cast(const struct dpcls_rule
*cr
)
1486 return cr
? CONTAINER_OF(cr
, struct dp_netdev_flow
, cr
) : NULL
;
1489 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*flow
)
1491 return ovs_refcount_try_ref_rcu(&flow
->ref_cnt
);
1494 /* netdev_flow_key utilities.
1496 * netdev_flow_key is basically a miniflow. We use these functions
1497 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1498 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1500 * - Since we are dealing exclusively with miniflows created by
1501 * miniflow_extract(), if the map is different the miniflow is different.
1502 * Therefore we can be faster by comparing the map and the miniflow in a
1504 * _ netdev_flow_key's miniflow has always inline values.
1505 * - These functions can be inlined by the compiler.
1507 * The following assertions make sure that what we're doing with miniflow is
1510 BUILD_ASSERT_DECL(offsetof(struct miniflow
, inline_values
)
1511 == sizeof(uint64_t));
1513 /* Given the number of bits set in the miniflow map, returns the size of the
1514 * 'netdev_flow_key.mf' */
1515 static inline uint32_t
1516 netdev_flow_key_size(uint32_t flow_u32s
)
1518 return offsetof(struct miniflow
, inline_values
) +
1519 MINIFLOW_VALUES_SIZE(flow_u32s
);
1523 netdev_flow_key_equal(const struct netdev_flow_key
*a
,
1524 const struct netdev_flow_key
*b
)
1526 /* 'b->len' may be not set yet. */
1527 return a
->hash
== b
->hash
&& !memcmp(&a
->mf
, &b
->mf
, a
->len
);
1530 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1531 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1532 * generated by miniflow_extract. */
1534 netdev_flow_key_equal_mf(const struct netdev_flow_key
*key
,
1535 const struct miniflow
*mf
)
1537 return !memcmp(&key
->mf
, mf
, key
->len
);
1541 netdev_flow_key_clone(struct netdev_flow_key
*dst
,
1542 const struct netdev_flow_key
*src
)
1545 offsetof(struct netdev_flow_key
, mf
) + src
->len
);
1550 netdev_flow_key_from_flow(struct netdev_flow_key
*dst
,
1551 const struct flow
*src
)
1553 struct dp_packet packet
;
1554 uint64_t buf_stub
[512 / 8];
1556 miniflow_initialize(&dst
->mf
, dst
->buf
);
1558 dp_packet_use_stub(&packet
, buf_stub
, sizeof buf_stub
);
1559 pkt_metadata_from_flow(&packet
.md
, src
);
1560 flow_compose(&packet
, src
);
1561 miniflow_extract(&packet
, &dst
->mf
);
1562 dp_packet_uninit(&packet
);
1564 dst
->len
= netdev_flow_key_size(count_1bits(dst
->mf
.map
));
1565 dst
->hash
= 0; /* Not computed yet. */
1568 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1570 netdev_flow_mask_init(struct netdev_flow_key
*mask
,
1571 const struct match
*match
)
1573 const uint64_t *mask_u64
= (const uint64_t *) &match
->wc
.masks
;
1574 uint64_t *dst
= mask
->mf
.inline_values
;
1575 uint64_t map
, mask_map
= 0;
1579 /* Only check masks that make sense for the flow. */
1580 map
= flow_wc_map(&match
->flow
);
1583 uint64_t rm1bit
= rightmost_1bit(map
);
1584 int i
= raw_ctz(map
);
1588 *dst
++ = mask_u64
[i
];
1589 hash
= hash_add64(hash
, mask_u64
[i
]);
1594 mask
->mf
.values_inline
= true;
1595 mask
->mf
.map
= mask_map
;
1597 hash
= hash_add64(hash
, mask_map
);
1599 n
= dst
- mask
->mf
.inline_values
;
1601 mask
->hash
= hash_finish(hash
, n
* 8);
1602 mask
->len
= netdev_flow_key_size(n
);
1605 /* Initializes 'dst' as a copy of 'src' masked with 'mask'. */
1607 netdev_flow_key_init_masked(struct netdev_flow_key
*dst
,
1608 const struct flow
*flow
,
1609 const struct netdev_flow_key
*mask
)
1611 uint64_t *dst_u64
= dst
->mf
.inline_values
;
1612 const uint64_t *mask_u64
= mask
->mf
.inline_values
;
1616 dst
->len
= mask
->len
;
1617 dst
->mf
.values_inline
= true;
1618 dst
->mf
.map
= mask
->mf
.map
;
1620 FLOW_FOR_EACH_IN_MAP(value
, flow
, mask
->mf
.map
) {
1621 *dst_u64
= value
& *mask_u64
++;
1622 hash
= hash_add64(hash
, *dst_u64
++);
1624 dst
->hash
= hash_finish(hash
, (dst_u64
- dst
->mf
.inline_values
) * 8);
1627 /* Iterate through all netdev_flow_key u64 values specified by 'MAP' */
1628 #define NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(VALUE, KEY, MAP) \
1629 for (struct mf_for_each_in_map_aux aux__ \
1630 = { (KEY)->mf.inline_values, (KEY)->mf.map, MAP }; \
1631 mf_get_next_in_map(&aux__, &(VALUE)); \
1634 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1636 static inline uint32_t
1637 netdev_flow_key_hash_in_mask(const struct netdev_flow_key
*key
,
1638 const struct netdev_flow_key
*mask
)
1640 const uint64_t *p
= mask
->mf
.inline_values
;
1644 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(key_u64
, key
, mask
->mf
.map
) {
1645 hash
= hash_add64(hash
, key_u64
& *p
++);
1648 return hash_finish(hash
, (p
- mask
->mf
.inline_values
) * 8);
1652 emc_entry_alive(struct emc_entry
*ce
)
1654 return ce
->flow
&& !ce
->flow
->dead
;
1658 emc_clear_entry(struct emc_entry
*ce
)
1661 dp_netdev_flow_unref(ce
->flow
);
1667 emc_change_entry(struct emc_entry
*ce
, struct dp_netdev_flow
*flow
,
1668 const struct netdev_flow_key
*key
)
1670 if (ce
->flow
!= flow
) {
1672 dp_netdev_flow_unref(ce
->flow
);
1675 if (dp_netdev_flow_ref(flow
)) {
1682 netdev_flow_key_clone(&ce
->key
, key
);
1687 emc_insert(struct emc_cache
*cache
, const struct netdev_flow_key
*key
,
1688 struct dp_netdev_flow
*flow
)
1690 struct emc_entry
*to_be_replaced
= NULL
;
1691 struct emc_entry
*current_entry
;
1693 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
1694 if (netdev_flow_key_equal(¤t_entry
->key
, key
)) {
1695 /* We found the entry with the 'mf' miniflow */
1696 emc_change_entry(current_entry
, flow
, NULL
);
1700 /* Replacement policy: put the flow in an empty (not alive) entry, or
1701 * in the first entry where it can be */
1703 || (emc_entry_alive(to_be_replaced
)
1704 && !emc_entry_alive(current_entry
))
1705 || current_entry
->key
.hash
< to_be_replaced
->key
.hash
) {
1706 to_be_replaced
= current_entry
;
1709 /* We didn't find the miniflow in the cache.
1710 * The 'to_be_replaced' entry is where the new flow will be stored */
1712 emc_change_entry(to_be_replaced
, flow
, key
);
1715 static inline struct dp_netdev_flow
*
1716 emc_lookup(struct emc_cache
*cache
, const struct netdev_flow_key
*key
)
1718 struct emc_entry
*current_entry
;
1720 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
1721 if (current_entry
->key
.hash
== key
->hash
1722 && emc_entry_alive(current_entry
)
1723 && netdev_flow_key_equal_mf(¤t_entry
->key
, &key
->mf
)) {
1725 /* We found the entry with the 'key->mf' miniflow */
1726 return current_entry
->flow
;
1733 static struct dp_netdev_flow
*
1734 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread
*pmd
,
1735 const struct netdev_flow_key
*key
)
1737 struct dp_netdev_flow
*netdev_flow
;
1738 struct dpcls_rule
*rule
;
1740 dpcls_lookup(&pmd
->cls
, key
, &rule
, 1);
1741 netdev_flow
= dp_netdev_flow_cast(rule
);
1746 static struct dp_netdev_flow
*
1747 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread
*pmd
,
1748 const ovs_u128
*ufidp
, const struct nlattr
*key
,
1751 struct dp_netdev_flow
*netdev_flow
;
1755 /* If a UFID is not provided, determine one based on the key. */
1756 if (!ufidp
&& key
&& key_len
1757 && !dpif_netdev_flow_from_nlattrs(key
, key_len
, &flow
)) {
1758 dpif_flow_hash(pmd
->dp
->dpif
, &flow
, sizeof flow
, &ufid
);
1763 CMAP_FOR_EACH_WITH_HASH (netdev_flow
, node
, dp_netdev_flow_hash(ufidp
),
1765 if (ovs_u128_equal(&netdev_flow
->ufid
, ufidp
)) {
1775 get_dpif_flow_stats(const struct dp_netdev_flow
*netdev_flow_
,
1776 struct dpif_flow_stats
*stats
)
1778 struct dp_netdev_flow
*netdev_flow
;
1779 unsigned long long n
;
1783 netdev_flow
= CONST_CAST(struct dp_netdev_flow
*, netdev_flow_
);
1785 atomic_read_relaxed(&netdev_flow
->stats
.packet_count
, &n
);
1786 stats
->n_packets
= n
;
1787 atomic_read_relaxed(&netdev_flow
->stats
.byte_count
, &n
);
1789 atomic_read_relaxed(&netdev_flow
->stats
.used
, &used
);
1791 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
1792 stats
->tcp_flags
= flags
;
1795 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1796 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1797 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1800 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow
*netdev_flow
,
1801 struct ofpbuf
*key_buf
, struct ofpbuf
*mask_buf
,
1802 struct dpif_flow
*flow
, bool terse
)
1805 memset(flow
, 0, sizeof *flow
);
1807 struct flow_wildcards wc
;
1808 struct dp_netdev_actions
*actions
;
1811 miniflow_expand(&netdev_flow
->cr
.mask
->mf
, &wc
.masks
);
1814 offset
= key_buf
->size
;
1815 flow
->key
= ofpbuf_tail(key_buf
);
1816 odp_flow_key_from_flow(key_buf
, &netdev_flow
->flow
, &wc
.masks
,
1817 netdev_flow
->flow
.in_port
.odp_port
, true);
1818 flow
->key_len
= key_buf
->size
- offset
;
1821 offset
= mask_buf
->size
;
1822 flow
->mask
= ofpbuf_tail(mask_buf
);
1823 odp_flow_key_from_mask(mask_buf
, &wc
.masks
, &netdev_flow
->flow
,
1824 odp_to_u32(wc
.masks
.in_port
.odp_port
),
1826 flow
->mask_len
= mask_buf
->size
- offset
;
1829 actions
= dp_netdev_flow_get_actions(netdev_flow
);
1830 flow
->actions
= actions
->actions
;
1831 flow
->actions_len
= actions
->size
;
1834 flow
->ufid
= netdev_flow
->ufid
;
1835 flow
->ufid_present
= true;
1836 flow
->pmd_id
= netdev_flow
->pmd_id
;
1837 get_dpif_flow_stats(netdev_flow
, &flow
->stats
);
1841 dpif_netdev_mask_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
1842 const struct nlattr
*mask_key
,
1843 uint32_t mask_key_len
, const struct flow
*flow
,
1847 enum odp_key_fitness fitness
;
1849 fitness
= odp_flow_key_to_mask(mask_key
, mask_key_len
, mask
, flow
);
1851 /* This should not happen: it indicates that
1852 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1853 * disagree on the acceptable form of a mask. Log the problem
1854 * as an error, with enough details to enable debugging. */
1855 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
1857 if (!VLOG_DROP_ERR(&rl
)) {
1861 odp_flow_format(key
, key_len
, mask_key
, mask_key_len
, NULL
, &s
,
1863 VLOG_ERR("internal error parsing flow mask %s (%s)",
1864 ds_cstr(&s
), odp_key_fitness_to_string(fitness
));
1871 enum mf_field_id id
;
1872 /* No mask key, unwildcard everything except fields whose
1873 * prerequisities are not met. */
1874 memset(mask
, 0x0, sizeof *mask
);
1876 for (id
= 0; id
< MFF_N_IDS
; ++id
) {
1877 /* Skip registers and metadata. */
1878 if (!(id
>= MFF_REG0
&& id
< MFF_REG0
+ FLOW_N_REGS
)
1879 && id
!= MFF_METADATA
) {
1880 const struct mf_field
*mf
= mf_from_id(id
);
1881 if (mf_are_prereqs_ok(mf
, flow
)) {
1882 mf_mask_field(mf
, mask
);
1888 /* Force unwildcard the in_port.
1890 * We need to do this even in the case where we unwildcard "everything"
1891 * above because "everything" only includes the 16-bit OpenFlow port number
1892 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1893 * port number mask->in_port.odp_port. */
1894 mask
->in_port
.odp_port
= u32_to_odp(UINT32_MAX
);
1900 dpif_netdev_flow_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
1905 if (odp_flow_key_to_flow(key
, key_len
, flow
)) {
1906 /* This should not happen: it indicates that odp_flow_key_from_flow()
1907 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1908 * flow. Log the problem as an error, with enough details to enable
1910 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
1912 if (!VLOG_DROP_ERR(&rl
)) {
1916 odp_flow_format(key
, key_len
, NULL
, 0, NULL
, &s
, true);
1917 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s
));
1924 in_port
= flow
->in_port
.odp_port
;
1925 if (!is_valid_port_number(in_port
) && in_port
!= ODPP_NONE
) {
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 int pmd_id
= get
->pmd_id
== PMD_ID_NULL
? NON_PMD_CORE_ID
: get
->pmd_id
;
1941 pmd
= dp_netdev_get_pmd(dp
, pmd_id
);
1946 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, get
->ufid
, get
->key
,
1949 dp_netdev_flow_to_dpif_flow(netdev_flow
, get
->buffer
, get
->buffer
,
1954 dp_netdev_pmd_unref(pmd
);
1960 static struct dp_netdev_flow
*
1961 dp_netdev_flow_add(struct dp_netdev_pmd_thread
*pmd
,
1962 struct match
*match
, const ovs_u128
*ufid
,
1963 const struct nlattr
*actions
, size_t actions_len
)
1964 OVS_REQUIRES(pmd
->flow_mutex
)
1966 struct dp_netdev_flow
*flow
;
1967 struct netdev_flow_key mask
;
1969 netdev_flow_mask_init(&mask
, match
);
1970 /* Make sure wc does not have metadata. */
1971 ovs_assert(!(mask
.mf
.map
& (MINIFLOW_MAP(metadata
) | MINIFLOW_MAP(regs
))));
1973 /* Do not allocate extra space. */
1974 flow
= xmalloc(sizeof *flow
- sizeof flow
->cr
.flow
.mf
+ mask
.len
);
1975 memset(&flow
->stats
, 0, sizeof flow
->stats
);
1977 *CONST_CAST(int *, &flow
->pmd_id
) = pmd
->core_id
;
1978 *CONST_CAST(struct flow
*, &flow
->flow
) = match
->flow
;
1979 *CONST_CAST(ovs_u128
*, &flow
->ufid
) = *ufid
;
1980 ovs_refcount_init(&flow
->ref_cnt
);
1981 ovsrcu_set(&flow
->actions
, dp_netdev_actions_create(actions
, actions_len
));
1983 netdev_flow_key_init_masked(&flow
->cr
.flow
, &match
->flow
, &mask
);
1984 dpcls_insert(&pmd
->cls
, &flow
->cr
, &mask
);
1986 cmap_insert(&pmd
->flow_table
, CONST_CAST(struct cmap_node
*, &flow
->node
),
1987 dp_netdev_flow_hash(&flow
->ufid
));
1989 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1991 struct ds ds
= DS_EMPTY_INITIALIZER
;
1993 match
.flow
= flow
->flow
;
1994 miniflow_expand(&flow
->cr
.mask
->mf
, &match
.wc
.masks
);
1996 ds_put_cstr(&ds
, "flow_add: ");
1997 odp_format_ufid(ufid
, &ds
);
1998 ds_put_cstr(&ds
, " ");
1999 match_format(&match
, &ds
, OFP_DEFAULT_PRIORITY
);
2000 ds_put_cstr(&ds
, ", actions:");
2001 format_odp_actions(&ds
, actions
, actions_len
);
2003 VLOG_DBG_RL(&upcall_rl
, "%s", ds_cstr(&ds
));
2012 dpif_netdev_flow_put(struct dpif
*dpif
, const struct dpif_flow_put
*put
)
2014 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2015 struct dp_netdev_flow
*netdev_flow
;
2016 struct netdev_flow_key key
;
2017 struct dp_netdev_pmd_thread
*pmd
;
2020 int pmd_id
= put
->pmd_id
== PMD_ID_NULL
? NON_PMD_CORE_ID
: put
->pmd_id
;
2023 error
= dpif_netdev_flow_from_nlattrs(put
->key
, put
->key_len
, &match
.flow
);
2027 error
= dpif_netdev_mask_from_nlattrs(put
->key
, put
->key_len
,
2028 put
->mask
, put
->mask_len
,
2029 &match
.flow
, &match
.wc
.masks
);
2034 pmd
= dp_netdev_get_pmd(dp
, pmd_id
);
2039 /* Must produce a netdev_flow_key for lookup.
2040 * This interface is no longer performance critical, since it is not used
2041 * for upcall processing any more. */
2042 netdev_flow_key_from_flow(&key
, &match
.flow
);
2047 dpif_flow_hash(dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
2050 ovs_mutex_lock(&pmd
->flow_mutex
);
2051 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, &key
);
2053 if (put
->flags
& DPIF_FP_CREATE
) {
2054 if (cmap_count(&pmd
->flow_table
) < MAX_FLOWS
) {
2056 memset(put
->stats
, 0, sizeof *put
->stats
);
2058 dp_netdev_flow_add(pmd
, &match
, &ufid
, put
->actions
,
2068 if (put
->flags
& DPIF_FP_MODIFY
2069 && flow_equal(&match
.flow
, &netdev_flow
->flow
)) {
2070 struct dp_netdev_actions
*new_actions
;
2071 struct dp_netdev_actions
*old_actions
;
2073 new_actions
= dp_netdev_actions_create(put
->actions
,
2076 old_actions
= dp_netdev_flow_get_actions(netdev_flow
);
2077 ovsrcu_set(&netdev_flow
->actions
, new_actions
);
2080 get_dpif_flow_stats(netdev_flow
, put
->stats
);
2082 if (put
->flags
& DPIF_FP_ZERO_STATS
) {
2083 /* XXX: The userspace datapath uses thread local statistics
2084 * (for flows), which should be updated only by the owning
2085 * thread. Since we cannot write on stats memory here,
2086 * we choose not to support this flag. Please note:
2087 * - This feature is currently used only by dpctl commands with
2089 * - Should the need arise, this operation can be implemented
2090 * by keeping a base value (to be update here) for each
2091 * counter, and subtracting it before outputting the stats */
2095 ovsrcu_postpone(dp_netdev_actions_free
, old_actions
);
2096 } else if (put
->flags
& DPIF_FP_CREATE
) {
2099 /* Overlapping flow. */
2103 ovs_mutex_unlock(&pmd
->flow_mutex
);
2104 dp_netdev_pmd_unref(pmd
);
2110 dpif_netdev_flow_del(struct dpif
*dpif
, const struct dpif_flow_del
*del
)
2112 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2113 struct dp_netdev_flow
*netdev_flow
;
2114 struct dp_netdev_pmd_thread
*pmd
;
2115 int pmd_id
= del
->pmd_id
== PMD_ID_NULL
? NON_PMD_CORE_ID
: del
->pmd_id
;
2118 pmd
= dp_netdev_get_pmd(dp
, pmd_id
);
2123 ovs_mutex_lock(&pmd
->flow_mutex
);
2124 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, del
->ufid
, del
->key
,
2128 get_dpif_flow_stats(netdev_flow
, del
->stats
);
2130 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
2134 ovs_mutex_unlock(&pmd
->flow_mutex
);
2135 dp_netdev_pmd_unref(pmd
);
2140 struct dpif_netdev_flow_dump
{
2141 struct dpif_flow_dump up
;
2142 struct cmap_position poll_thread_pos
;
2143 struct cmap_position flow_pos
;
2144 struct dp_netdev_pmd_thread
*cur_pmd
;
2146 struct ovs_mutex mutex
;
2149 static struct dpif_netdev_flow_dump
*
2150 dpif_netdev_flow_dump_cast(struct dpif_flow_dump
*dump
)
2152 return CONTAINER_OF(dump
, struct dpif_netdev_flow_dump
, up
);
2155 static struct dpif_flow_dump
*
2156 dpif_netdev_flow_dump_create(const struct dpif
*dpif_
, bool terse
)
2158 struct dpif_netdev_flow_dump
*dump
;
2160 dump
= xzalloc(sizeof *dump
);
2161 dpif_flow_dump_init(&dump
->up
, dpif_
);
2162 dump
->up
.terse
= terse
;
2163 ovs_mutex_init(&dump
->mutex
);
2169 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump
*dump_
)
2171 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
2173 ovs_mutex_destroy(&dump
->mutex
);
2178 struct dpif_netdev_flow_dump_thread
{
2179 struct dpif_flow_dump_thread up
;
2180 struct dpif_netdev_flow_dump
*dump
;
2181 struct odputil_keybuf keybuf
[FLOW_DUMP_MAX_BATCH
];
2182 struct odputil_keybuf maskbuf
[FLOW_DUMP_MAX_BATCH
];
2185 static struct dpif_netdev_flow_dump_thread
*
2186 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread
*thread
)
2188 return CONTAINER_OF(thread
, struct dpif_netdev_flow_dump_thread
, up
);
2191 static struct dpif_flow_dump_thread
*
2192 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump
*dump_
)
2194 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
2195 struct dpif_netdev_flow_dump_thread
*thread
;
2197 thread
= xmalloc(sizeof *thread
);
2198 dpif_flow_dump_thread_init(&thread
->up
, &dump
->up
);
2199 thread
->dump
= dump
;
2204 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread
*thread_
)
2206 struct dpif_netdev_flow_dump_thread
*thread
2207 = dpif_netdev_flow_dump_thread_cast(thread_
);
2213 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread
*thread_
,
2214 struct dpif_flow
*flows
, int max_flows
)
2216 struct dpif_netdev_flow_dump_thread
*thread
2217 = dpif_netdev_flow_dump_thread_cast(thread_
);
2218 struct dpif_netdev_flow_dump
*dump
= thread
->dump
;
2219 struct dp_netdev_flow
*netdev_flows
[FLOW_DUMP_MAX_BATCH
];
2223 ovs_mutex_lock(&dump
->mutex
);
2224 if (!dump
->status
) {
2225 struct dpif_netdev
*dpif
= dpif_netdev_cast(thread
->up
.dpif
);
2226 struct dp_netdev
*dp
= get_dp_netdev(&dpif
->dpif
);
2227 struct dp_netdev_pmd_thread
*pmd
= dump
->cur_pmd
;
2228 int flow_limit
= MIN(max_flows
, FLOW_DUMP_MAX_BATCH
);
2230 /* First call to dump_next(), extracts the first pmd thread.
2231 * If there is no pmd thread, returns immediately. */
2233 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
2235 ovs_mutex_unlock(&dump
->mutex
);
2242 for (n_flows
= 0; n_flows
< flow_limit
; n_flows
++) {
2243 struct cmap_node
*node
;
2245 node
= cmap_next_position(&pmd
->flow_table
, &dump
->flow_pos
);
2249 netdev_flows
[n_flows
] = CONTAINER_OF(node
,
2250 struct dp_netdev_flow
,
2253 /* When finishing dumping the current pmd thread, moves to
2255 if (n_flows
< flow_limit
) {
2256 memset(&dump
->flow_pos
, 0, sizeof dump
->flow_pos
);
2257 dp_netdev_pmd_unref(pmd
);
2258 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
2264 /* Keeps the reference to next caller. */
2265 dump
->cur_pmd
= pmd
;
2267 /* If the current dump is empty, do not exit the loop, since the
2268 * remaining pmds could have flows to be dumped. Just dumps again
2269 * on the new 'pmd'. */
2272 ovs_mutex_unlock(&dump
->mutex
);
2274 for (i
= 0; i
< n_flows
; i
++) {
2275 struct odputil_keybuf
*maskbuf
= &thread
->maskbuf
[i
];
2276 struct odputil_keybuf
*keybuf
= &thread
->keybuf
[i
];
2277 struct dp_netdev_flow
*netdev_flow
= netdev_flows
[i
];
2278 struct dpif_flow
*f
= &flows
[i
];
2279 struct ofpbuf key
, mask
;
2281 ofpbuf_use_stack(&key
, keybuf
, sizeof *keybuf
);
2282 ofpbuf_use_stack(&mask
, maskbuf
, sizeof *maskbuf
);
2283 dp_netdev_flow_to_dpif_flow(netdev_flow
, &key
, &mask
, f
,
2291 dpif_netdev_execute(struct dpif
*dpif
, struct dpif_execute
*execute
)
2292 OVS_NO_THREAD_SAFETY_ANALYSIS
2294 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2295 struct dp_netdev_pmd_thread
*pmd
;
2296 struct dp_packet
*pp
;
2298 if (dp_packet_size(execute
->packet
) < ETH_HEADER_LEN
||
2299 dp_packet_size(execute
->packet
) > UINT16_MAX
) {
2303 /* Tries finding the 'pmd'. If NULL is returned, that means
2304 * the current thread is a non-pmd thread and should use
2305 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2306 pmd
= ovsthread_getspecific(dp
->per_pmd_key
);
2308 pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
2311 /* If the current thread is non-pmd thread, acquires
2312 * the 'non_pmd_mutex'. */
2313 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2314 ovs_mutex_lock(&dp
->non_pmd_mutex
);
2315 ovs_mutex_lock(&dp
->port_mutex
);
2318 pp
= execute
->packet
;
2319 dp_netdev_execute_actions(pmd
, &pp
, 1, false, execute
->actions
,
2320 execute
->actions_len
);
2321 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2322 dp_netdev_pmd_unref(pmd
);
2323 ovs_mutex_unlock(&dp
->port_mutex
);
2324 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
2331 dpif_netdev_operate(struct dpif
*dpif
, struct dpif_op
**ops
, size_t n_ops
)
2335 for (i
= 0; i
< n_ops
; i
++) {
2336 struct dpif_op
*op
= ops
[i
];
2339 case DPIF_OP_FLOW_PUT
:
2340 op
->error
= dpif_netdev_flow_put(dpif
, &op
->u
.flow_put
);
2343 case DPIF_OP_FLOW_DEL
:
2344 op
->error
= dpif_netdev_flow_del(dpif
, &op
->u
.flow_del
);
2347 case DPIF_OP_EXECUTE
:
2348 op
->error
= dpif_netdev_execute(dpif
, &op
->u
.execute
);
2351 case DPIF_OP_FLOW_GET
:
2352 op
->error
= dpif_netdev_flow_get(dpif
, &op
->u
.flow_get
);
2358 /* Returns true if the configuration for rx queues or cpu mask
2361 pmd_config_changed(const struct dp_netdev
*dp
, size_t rxqs
, const char *cmask
)
2363 if (dp
->n_dpdk_rxqs
!= rxqs
) {
2366 if (dp
->pmd_cmask
!= NULL
&& cmask
!= NULL
) {
2367 return strcmp(dp
->pmd_cmask
, cmask
);
2369 return (dp
->pmd_cmask
!= NULL
|| cmask
!= NULL
);
2374 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2376 dpif_netdev_pmd_set(struct dpif
*dpif
, unsigned int n_rxqs
, const char *cmask
)
2378 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2380 if (pmd_config_changed(dp
, n_rxqs
, cmask
)) {
2381 struct dp_netdev_port
*port
;
2383 dp_netdev_destroy_all_pmds(dp
);
2385 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2386 if (netdev_is_pmd(port
->netdev
)) {
2389 /* Closes the existing 'rxq's. */
2390 for (i
= 0; i
< netdev_n_rxq(port
->netdev
); i
++) {
2391 netdev_rxq_close(port
->rxq
[i
]);
2392 port
->rxq
[i
] = NULL
;
2395 /* Sets the new rx queue config. */
2396 err
= netdev_set_multiq(port
->netdev
, ovs_numa_get_n_cores(),
2398 if (err
&& (err
!= EOPNOTSUPP
)) {
2399 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2400 " %u", netdev_get_name(port
->netdev
),
2405 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2406 port
->rxq
= xrealloc(port
->rxq
, sizeof *port
->rxq
2407 * netdev_n_rxq(port
->netdev
));
2408 for (i
= 0; i
< netdev_n_rxq(port
->netdev
); i
++) {
2409 netdev_rxq_open(port
->netdev
, &port
->rxq
[i
], i
);
2413 dp
->n_dpdk_rxqs
= n_rxqs
;
2415 /* Reconfigures the cpu mask. */
2416 ovs_numa_set_cpu_mask(cmask
);
2417 free(dp
->pmd_cmask
);
2418 dp
->pmd_cmask
= cmask
? xstrdup(cmask
) : NULL
;
2420 /* Restores the non-pmd. */
2421 dp_netdev_set_nonpmd(dp
);
2422 /* Restores all pmd threads. */
2423 dp_netdev_reset_pmd_threads(dp
);
2430 dpif_netdev_queue_to_priority(const struct dpif
*dpif OVS_UNUSED
,
2431 uint32_t queue_id
, uint32_t *priority
)
2433 *priority
= queue_id
;
2438 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2439 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2440 struct dp_netdev_actions
*
2441 dp_netdev_actions_create(const struct nlattr
*actions
, size_t size
)
2443 struct dp_netdev_actions
*netdev_actions
;
2445 netdev_actions
= xmalloc(sizeof *netdev_actions
+ size
);
2446 memcpy(netdev_actions
->actions
, actions
, size
);
2447 netdev_actions
->size
= size
;
2449 return netdev_actions
;
2452 struct dp_netdev_actions
*
2453 dp_netdev_flow_get_actions(const struct dp_netdev_flow
*flow
)
2455 return ovsrcu_get(struct dp_netdev_actions
*, &flow
->actions
);
2459 dp_netdev_actions_free(struct dp_netdev_actions
*actions
)
2464 static inline unsigned long long
2465 cycles_counter(void)
2468 return rte_get_tsc_cycles();
2474 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2475 extern struct ovs_mutex cycles_counter_fake_mutex
;
2477 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2479 cycles_count_start(struct dp_netdev_pmd_thread
*pmd
)
2480 OVS_ACQUIRES(&cycles_counter_fake_mutex
)
2481 OVS_NO_THREAD_SAFETY_ANALYSIS
2483 pmd
->last_cycles
= cycles_counter();
2486 /* Stop counting cycles and add them to the counter 'type' */
2488 cycles_count_end(struct dp_netdev_pmd_thread
*pmd
,
2489 enum pmd_cycles_counter_type type
)
2490 OVS_RELEASES(&cycles_counter_fake_mutex
)
2491 OVS_NO_THREAD_SAFETY_ANALYSIS
2493 unsigned long long interval
= cycles_counter() - pmd
->last_cycles
;
2495 non_atomic_ullong_add(&pmd
->cycles
.n
[type
], interval
);
2499 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread
*pmd
,
2500 struct dp_netdev_port
*port
,
2501 struct netdev_rxq
*rxq
)
2503 struct dp_packet
*packets
[NETDEV_MAX_RX_BATCH
];
2506 cycles_count_start(pmd
);
2507 error
= netdev_rxq_recv(rxq
, packets
, &cnt
);
2508 cycles_count_end(pmd
, PMD_CYCLES_POLLING
);
2512 *recirc_depth_get() = 0;
2514 /* XXX: initialize md in netdev implementation. */
2515 for (i
= 0; i
< cnt
; i
++) {
2516 packets
[i
]->md
= PKT_METADATA_INITIALIZER(port
->port_no
);
2518 cycles_count_start(pmd
);
2519 dp_netdev_input(pmd
, packets
, cnt
);
2520 cycles_count_end(pmd
, PMD_CYCLES_PROCESSING
);
2521 } else if (error
!= EAGAIN
&& error
!= EOPNOTSUPP
) {
2522 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
2524 VLOG_ERR_RL(&rl
, "error receiving data from %s: %s",
2525 netdev_get_name(port
->netdev
), ovs_strerror(error
));
2529 /* Return true if needs to revalidate datapath flows. */
2531 dpif_netdev_run(struct dpif
*dpif
)
2533 struct dp_netdev_port
*port
;
2534 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2535 struct dp_netdev_pmd_thread
*non_pmd
= dp_netdev_get_pmd(dp
,
2537 uint64_t new_tnl_seq
;
2539 ovs_mutex_lock(&dp
->non_pmd_mutex
);
2540 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2541 if (!netdev_is_pmd(port
->netdev
)) {
2544 for (i
= 0; i
< netdev_n_rxq(port
->netdev
); i
++) {
2545 dp_netdev_process_rxq_port(non_pmd
, port
, port
->rxq
[i
]);
2549 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
2550 dp_netdev_pmd_unref(non_pmd
);
2552 tnl_arp_cache_run();
2553 new_tnl_seq
= seq_read(tnl_conf_seq
);
2555 if (dp
->last_tnl_conf_seq
!= new_tnl_seq
) {
2556 dp
->last_tnl_conf_seq
= new_tnl_seq
;
2563 dpif_netdev_wait(struct dpif
*dpif
)
2565 struct dp_netdev_port
*port
;
2566 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2568 ovs_mutex_lock(&dp_netdev_mutex
);
2569 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2570 if (!netdev_is_pmd(port
->netdev
)) {
2573 for (i
= 0; i
< netdev_n_rxq(port
->netdev
); i
++) {
2574 netdev_rxq_wait(port
->rxq
[i
]);
2578 ovs_mutex_unlock(&dp_netdev_mutex
);
2579 seq_wait(tnl_conf_seq
, dp
->last_tnl_conf_seq
);
2583 struct dp_netdev_port
*port
;
2584 struct netdev_rxq
*rx
;
2588 pmd_load_queues(struct dp_netdev_pmd_thread
*pmd
,
2589 struct rxq_poll
**ppoll_list
, int poll_cnt
)
2591 struct rxq_poll
*poll_list
= *ppoll_list
;
2592 struct dp_netdev_port
*port
;
2593 int n_pmds_on_numa
, index
, i
;
2595 /* Simple scheduler for netdev rx polling. */
2596 for (i
= 0; i
< poll_cnt
; i
++) {
2597 port_unref(poll_list
[i
].port
);
2601 n_pmds_on_numa
= get_n_pmd_threads_on_numa(pmd
->dp
, pmd
->numa_id
);
2604 CMAP_FOR_EACH (port
, node
, &pmd
->dp
->ports
) {
2605 /* Calls port_try_ref() to prevent the main thread
2606 * from deleting the port. */
2607 if (port_try_ref(port
)) {
2608 if (netdev_is_pmd(port
->netdev
)
2609 && netdev_get_numa_id(port
->netdev
) == pmd
->numa_id
) {
2612 for (i
= 0; i
< netdev_n_rxq(port
->netdev
); i
++) {
2613 if ((index
% n_pmds_on_numa
) == pmd
->index
) {
2614 poll_list
= xrealloc(poll_list
,
2615 sizeof *poll_list
* (poll_cnt
+ 1));
2618 poll_list
[poll_cnt
].port
= port
;
2619 poll_list
[poll_cnt
].rx
= port
->rxq
[i
];
2625 /* Unrefs the port_try_ref(). */
2630 *ppoll_list
= poll_list
;
2635 pmd_thread_main(void *f_
)
2637 struct dp_netdev_pmd_thread
*pmd
= f_
;
2638 unsigned int lc
= 0;
2639 struct rxq_poll
*poll_list
;
2640 unsigned int port_seq
= PMD_INITIAL_SEQ
;
2647 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2648 ovsthread_setspecific(pmd
->dp
->per_pmd_key
, pmd
);
2649 pmd_thread_setaffinity_cpu(pmd
->core_id
);
2651 emc_cache_init(&pmd
->flow_cache
);
2652 poll_cnt
= pmd_load_queues(pmd
, &poll_list
, poll_cnt
);
2654 /* Signal here to make sure the pmd finishes
2655 * reloading the updated configuration. */
2656 dp_netdev_pmd_reload_done(pmd
);
2661 for (i
= 0; i
< poll_cnt
; i
++) {
2662 dp_netdev_process_rxq_port(pmd
, poll_list
[i
].port
, poll_list
[i
].rx
);
2670 emc_cache_slow_sweep(&pmd
->flow_cache
);
2673 atomic_read_relaxed(&pmd
->change_seq
, &seq
);
2674 if (seq
!= port_seq
) {
2681 emc_cache_uninit(&pmd
->flow_cache
);
2683 if (!latch_is_set(&pmd
->exit_latch
)){
2687 for (i
= 0; i
< poll_cnt
; i
++) {
2688 port_unref(poll_list
[i
].port
);
2691 dp_netdev_pmd_reload_done(pmd
);
2698 dp_netdev_disable_upcall(struct dp_netdev
*dp
)
2699 OVS_ACQUIRES(dp
->upcall_rwlock
)
2701 fat_rwlock_wrlock(&dp
->upcall_rwlock
);
2705 dpif_netdev_disable_upcall(struct dpif
*dpif
)
2706 OVS_NO_THREAD_SAFETY_ANALYSIS
2708 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2709 dp_netdev_disable_upcall(dp
);
2713 dp_netdev_enable_upcall(struct dp_netdev
*dp
)
2714 OVS_RELEASES(dp
->upcall_rwlock
)
2716 fat_rwlock_unlock(&dp
->upcall_rwlock
);
2720 dpif_netdev_enable_upcall(struct dpif
*dpif
)
2721 OVS_NO_THREAD_SAFETY_ANALYSIS
2723 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2724 dp_netdev_enable_upcall(dp
);
2728 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
)
2730 ovs_mutex_lock(&pmd
->cond_mutex
);
2731 xpthread_cond_signal(&pmd
->cond
);
2732 ovs_mutex_unlock(&pmd
->cond_mutex
);
2735 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2736 * the pointer if succeeds, otherwise, NULL.
2738 * Caller must unrefs the returned reference. */
2739 static struct dp_netdev_pmd_thread
*
2740 dp_netdev_get_pmd(struct dp_netdev
*dp
, int core_id
)
2742 struct dp_netdev_pmd_thread
*pmd
;
2743 const struct cmap_node
*pnode
;
2745 pnode
= cmap_find(&dp
->poll_threads
, hash_int(core_id
, 0));
2749 pmd
= CONTAINER_OF(pnode
, struct dp_netdev_pmd_thread
, node
);
2751 return dp_netdev_pmd_try_ref(pmd
) ? pmd
: NULL
;
2754 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2756 dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
2758 struct dp_netdev_pmd_thread
*non_pmd
;
2760 non_pmd
= xzalloc(sizeof *non_pmd
);
2761 dp_netdev_configure_pmd(non_pmd
, dp
, 0, NON_PMD_CORE_ID
,
2765 /* Caller must have valid pointer to 'pmd'. */
2767 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
)
2769 return ovs_refcount_try_ref_rcu(&pmd
->ref_cnt
);
2773 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
)
2775 if (pmd
&& ovs_refcount_unref(&pmd
->ref_cnt
) == 1) {
2776 ovsrcu_postpone(dp_netdev_destroy_pmd
, pmd
);
2780 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2781 * fails, keeps checking for next node until reaching the end of cmap.
2783 * Caller must unrefs the returned reference. */
2784 static struct dp_netdev_pmd_thread
*
2785 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
)
2787 struct dp_netdev_pmd_thread
*next
;
2790 struct cmap_node
*node
;
2792 node
= cmap_next_position(&dp
->poll_threads
, pos
);
2793 next
= node
? CONTAINER_OF(node
, struct dp_netdev_pmd_thread
, node
)
2795 } while (next
&& !dp_netdev_pmd_try_ref(next
));
2800 /* Configures the 'pmd' based on the input argument. */
2802 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
, struct dp_netdev
*dp
,
2803 int index
, int core_id
, int numa_id
)
2807 pmd
->core_id
= core_id
;
2808 pmd
->numa_id
= numa_id
;
2810 ovs_refcount_init(&pmd
->ref_cnt
);
2811 latch_init(&pmd
->exit_latch
);
2812 atomic_init(&pmd
->change_seq
, PMD_INITIAL_SEQ
);
2813 xpthread_cond_init(&pmd
->cond
, NULL
);
2814 ovs_mutex_init(&pmd
->cond_mutex
);
2815 ovs_mutex_init(&pmd
->flow_mutex
);
2816 dpcls_init(&pmd
->cls
);
2817 cmap_init(&pmd
->flow_table
);
2818 /* init the 'flow_cache' since there is no
2819 * actual thread created for NON_PMD_CORE_ID. */
2820 if (core_id
== NON_PMD_CORE_ID
) {
2821 emc_cache_init(&pmd
->flow_cache
);
2823 cmap_insert(&dp
->poll_threads
, CONST_CAST(struct cmap_node
*, &pmd
->node
),
2824 hash_int(core_id
, 0));
2828 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
)
2830 dp_netdev_pmd_flow_flush(pmd
);
2831 dpcls_destroy(&pmd
->cls
);
2832 cmap_destroy(&pmd
->flow_table
);
2833 ovs_mutex_destroy(&pmd
->flow_mutex
);
2834 latch_destroy(&pmd
->exit_latch
);
2835 xpthread_cond_destroy(&pmd
->cond
);
2836 ovs_mutex_destroy(&pmd
->cond_mutex
);
2840 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2841 * and unrefs the struct. */
2843 dp_netdev_del_pmd(struct dp_netdev_pmd_thread
*pmd
)
2845 /* Uninit the 'flow_cache' since there is
2846 * no actual thread uninit it for NON_PMD_CORE_ID. */
2847 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2848 emc_cache_uninit(&pmd
->flow_cache
);
2850 latch_set(&pmd
->exit_latch
);
2851 dp_netdev_reload_pmd__(pmd
);
2852 ovs_numa_unpin_core(pmd
->core_id
);
2853 xpthread_join(pmd
->thread
, NULL
);
2855 cmap_remove(&pmd
->dp
->poll_threads
, &pmd
->node
, hash_int(pmd
->core_id
, 0));
2856 dp_netdev_pmd_unref(pmd
);
2859 /* Destroys all pmd threads. */
2861 dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
)
2863 struct dp_netdev_pmd_thread
*pmd
;
2865 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2866 dp_netdev_del_pmd(pmd
);
2870 /* Deletes all pmd threads on numa node 'numa_id'. */
2872 dp_netdev_del_pmds_on_numa(struct dp_netdev
*dp
, int numa_id
)
2874 struct dp_netdev_pmd_thread
*pmd
;
2876 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2877 if (pmd
->numa_id
== numa_id
) {
2878 dp_netdev_del_pmd(pmd
);
2883 /* Checks the numa node id of 'netdev' and starts pmd threads for
2886 dp_netdev_set_pmds_on_numa(struct dp_netdev
*dp
, int numa_id
)
2890 if (!ovs_numa_numa_id_is_valid(numa_id
)) {
2891 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2892 "invalid", numa_id
);
2896 n_pmds
= get_n_pmd_threads_on_numa(dp
, numa_id
);
2898 /* If there are already pmd threads created for the numa node
2899 * in which 'netdev' is on, do nothing. Else, creates the
2900 * pmd threads for the numa node. */
2902 int can_have
, n_unpinned
, i
;
2904 n_unpinned
= ovs_numa_get_n_unpinned_cores_on_numa(numa_id
);
2906 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2907 "cores on numa node");
2911 /* If cpu mask is specified, uses all unpinned cores, otherwise
2912 * tries creating NR_PMD_THREADS pmd threads. */
2913 can_have
= dp
->pmd_cmask
? n_unpinned
: MIN(n_unpinned
, NR_PMD_THREADS
);
2914 for (i
= 0; i
< can_have
; i
++) {
2915 struct dp_netdev_pmd_thread
*pmd
= xzalloc(sizeof *pmd
);
2916 int core_id
= ovs_numa_get_unpinned_core_on_numa(numa_id
);
2918 dp_netdev_configure_pmd(pmd
, dp
, i
, core_id
, numa_id
);
2919 /* Each thread will distribute all devices rx-queues among
2921 pmd
->thread
= ovs_thread_create("pmd", pmd_thread_main
, pmd
);
2923 VLOG_INFO("Created %d pmd threads on numa node %d", can_have
, numa_id
);
2928 /* Called after pmd threads config change. Restarts pmd threads with
2929 * new configuration. */
2931 dp_netdev_reset_pmd_threads(struct dp_netdev
*dp
)
2933 struct dp_netdev_port
*port
;
2935 CMAP_FOR_EACH (port
, node
, &dp
->ports
) {
2936 if (netdev_is_pmd(port
->netdev
)) {
2937 int numa_id
= netdev_get_numa_id(port
->netdev
);
2939 dp_netdev_set_pmds_on_numa(dp
, numa_id
);
2945 dpif_netdev_get_datapath_version(void)
2947 return xstrdup("<built-in>");
2951 dp_netdev_flow_used(struct dp_netdev_flow
*netdev_flow
, int cnt
, int size
,
2952 uint16_t tcp_flags
, long long now
)
2956 atomic_store_relaxed(&netdev_flow
->stats
.used
, now
);
2957 non_atomic_ullong_add(&netdev_flow
->stats
.packet_count
, cnt
);
2958 non_atomic_ullong_add(&netdev_flow
->stats
.byte_count
, size
);
2959 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
2961 atomic_store_relaxed(&netdev_flow
->stats
.tcp_flags
, flags
);
2965 dp_netdev_count_packet(struct dp_netdev_pmd_thread
*pmd
,
2966 enum dp_stat_type type
, int cnt
)
2968 non_atomic_ullong_add(&pmd
->stats
.n
[type
], cnt
);
2972 dp_netdev_upcall(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
*packet_
,
2973 struct flow
*flow
, struct flow_wildcards
*wc
, ovs_u128
*ufid
,
2974 enum dpif_upcall_type type
, const struct nlattr
*userdata
,
2975 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
2977 struct dp_netdev
*dp
= pmd
->dp
;
2979 if (OVS_UNLIKELY(!dp
->upcall_cb
)) {
2983 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl
))) {
2984 struct ds ds
= DS_EMPTY_INITIALIZER
;
2988 ofpbuf_init(&key
, 0);
2989 odp_flow_key_from_flow(&key
, flow
, &wc
->masks
, flow
->in_port
.odp_port
,
2991 packet_str
= ofp_packet_to_string(dp_packet_data(packet_
),
2992 dp_packet_size(packet_
));
2994 odp_flow_key_format(key
.data
, key
.size
, &ds
);
2996 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp
->name
,
2997 dpif_upcall_type_to_string(type
), ds_cstr(&ds
), packet_str
);
2999 ofpbuf_uninit(&key
);
3005 return dp
->upcall_cb(packet_
, flow
, ufid
, pmd
->core_id
, type
, userdata
,
3006 actions
, wc
, put_actions
, dp
->upcall_aux
);
3009 static inline uint32_t
3010 dpif_netdev_packet_get_dp_hash(struct dp_packet
*packet
,
3011 const struct miniflow
*mf
)
3015 hash
= dp_packet_get_rss_hash(packet
);
3016 if (OVS_UNLIKELY(!hash
)) {
3017 hash
= miniflow_hash_5tuple(mf
, 0);
3018 dp_packet_set_rss_hash(packet
, hash
);
3023 struct packet_batch
{
3024 unsigned int packet_count
;
3025 unsigned int byte_count
;
3028 struct dp_netdev_flow
*flow
;
3030 struct dp_packet
*packets
[NETDEV_MAX_RX_BATCH
];
3034 packet_batch_update(struct packet_batch
*batch
, struct dp_packet
*packet
,
3035 const struct miniflow
*mf
)
3037 batch
->tcp_flags
|= miniflow_get_tcp_flags(mf
);
3038 batch
->packets
[batch
->packet_count
++] = packet
;
3039 batch
->byte_count
+= dp_packet_size(packet
);
3043 packet_batch_init(struct packet_batch
*batch
, struct dp_netdev_flow
*flow
)
3047 batch
->packet_count
= 0;
3048 batch
->byte_count
= 0;
3049 batch
->tcp_flags
= 0;
3053 packet_batch_execute(struct packet_batch
*batch
,
3054 struct dp_netdev_pmd_thread
*pmd
,
3055 enum dp_stat_type hit_type
,
3058 struct dp_netdev_actions
*actions
;
3059 struct dp_netdev_flow
*flow
= batch
->flow
;
3061 dp_netdev_flow_used(batch
->flow
, batch
->packet_count
, batch
->byte_count
,
3062 batch
->tcp_flags
, now
);
3064 actions
= dp_netdev_flow_get_actions(flow
);
3066 dp_netdev_execute_actions(pmd
, batch
->packets
, batch
->packet_count
, true,
3067 actions
->actions
, actions
->size
);
3069 dp_netdev_count_packet(pmd
, hit_type
, batch
->packet_count
);
3073 dp_netdev_queue_batches(struct dp_packet
*pkt
,
3074 struct dp_netdev_flow
*flow
, const struct miniflow
*mf
,
3075 struct packet_batch
*batches
, size_t *n_batches
,
3078 struct packet_batch
*batch
= NULL
;
3081 if (OVS_UNLIKELY(!flow
)) {
3084 /* XXX: This O(n^2) algortihm makes sense if we're operating under the
3085 * assumption that the number of distinct flows (and therefore the
3086 * number of distinct batches) is quite small. If this turns out not
3087 * to be the case, it may make sense to pre sort based on the
3088 * netdev_flow pointer. That done we can get the appropriate batching
3089 * in O(n * log(n)) instead. */
3090 for (j
= *n_batches
- 1; j
>= 0; j
--) {
3091 if (batches
[j
].flow
== flow
) {
3092 batch
= &batches
[j
];
3093 packet_batch_update(batch
, pkt
, mf
);
3097 if (OVS_UNLIKELY(*n_batches
>= max_batches
)) {
3101 batch
= &batches
[(*n_batches
)++];
3102 packet_batch_init(batch
, flow
);
3103 packet_batch_update(batch
, pkt
, mf
);
3108 dp_packet_swap(struct dp_packet
**a
, struct dp_packet
**b
)
3110 struct dp_packet
*tmp
= *a
;
3115 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3116 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', or if there
3117 * is no matching batch for a packet's flow, the miniflow is copied into 'keys'
3118 * and the packet pointer is moved at the beginning of the 'packets' array.
3120 * The function returns the number of packets that needs to be processed in the
3121 * 'packets' array (they have been moved to the beginning of the vector).
3123 static inline size_t
3124 emc_processing(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
**packets
,
3125 size_t cnt
, struct netdev_flow_key
*keys
, long long now
)
3127 struct netdev_flow_key key
;
3128 struct packet_batch batches
[4];
3129 struct emc_cache
*flow_cache
= &pmd
->flow_cache
;
3130 size_t n_batches
, i
;
3131 size_t notfound_cnt
= 0;
3134 miniflow_initialize(&key
.mf
, key
.buf
);
3135 for (i
= 0; i
< cnt
; i
++) {
3136 struct dp_netdev_flow
*flow
;
3138 if (OVS_UNLIKELY(dp_packet_size(packets
[i
]) < ETH_HEADER_LEN
)) {
3139 dp_packet_delete(packets
[i
]);
3143 miniflow_extract(packets
[i
], &key
.mf
);
3144 key
.len
= 0; /* Not computed yet. */
3145 key
.hash
= dpif_netdev_packet_get_dp_hash(packets
[i
], &key
.mf
);
3147 flow
= emc_lookup(flow_cache
, &key
);
3148 if (OVS_UNLIKELY(!dp_netdev_queue_batches(packets
[i
], flow
, &key
.mf
,
3149 batches
, &n_batches
,
3150 ARRAY_SIZE(batches
)))) {
3151 if (i
!= notfound_cnt
) {
3152 dp_packet_swap(&packets
[i
], &packets
[notfound_cnt
]);
3155 keys
[notfound_cnt
++] = key
;
3159 for (i
= 0; i
< n_batches
; i
++) {
3160 packet_batch_execute(&batches
[i
], pmd
, DP_STAT_EXACT_HIT
, now
);
3163 return notfound_cnt
;
3167 fast_path_processing(struct dp_netdev_pmd_thread
*pmd
,
3168 struct dp_packet
**packets
, size_t cnt
,
3169 struct netdev_flow_key
*keys
, long long now
)
3171 #if !defined(__CHECKER__) && !defined(_WIN32)
3172 const size_t PKT_ARRAY_SIZE
= cnt
;
3174 /* Sparse or MSVC doesn't like variable length array. */
3175 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_RX_BATCH
};
3177 struct packet_batch batches
[PKT_ARRAY_SIZE
];
3178 struct dpcls_rule
*rules
[PKT_ARRAY_SIZE
];
3179 struct dp_netdev
*dp
= pmd
->dp
;
3180 struct emc_cache
*flow_cache
= &pmd
->flow_cache
;
3181 size_t n_batches
, i
;
3184 for (i
= 0; i
< cnt
; i
++) {
3185 /* Key length is needed in all the cases, hash computed on demand. */
3186 keys
[i
].len
= netdev_flow_key_size(count_1bits(keys
[i
].mf
.map
));
3188 any_miss
= !dpcls_lookup(&pmd
->cls
, keys
, rules
, cnt
);
3189 if (OVS_UNLIKELY(any_miss
) && !fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
3190 uint64_t actions_stub
[512 / 8], slow_stub
[512 / 8];
3191 struct ofpbuf actions
, put_actions
;
3192 int miss_cnt
= 0, lost_cnt
= 0;
3195 ofpbuf_use_stub(&actions
, actions_stub
, sizeof actions_stub
);
3196 ofpbuf_use_stub(&put_actions
, slow_stub
, sizeof slow_stub
);
3198 for (i
= 0; i
< cnt
; i
++) {
3199 struct dp_netdev_flow
*netdev_flow
;
3200 struct ofpbuf
*add_actions
;
3204 if (OVS_LIKELY(rules
[i
])) {
3208 /* It's possible that an earlier slow path execution installed
3209 * a rule covering this flow. In this case, it's a lot cheaper
3210 * to catch it here than execute a miss. */
3211 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, &keys
[i
]);
3213 rules
[i
] = &netdev_flow
->cr
;
3219 miniflow_expand(&keys
[i
].mf
, &match
.flow
);
3221 ofpbuf_clear(&actions
);
3222 ofpbuf_clear(&put_actions
);
3224 dpif_flow_hash(dp
->dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
3225 error
= dp_netdev_upcall(pmd
, packets
[i
], &match
.flow
, &match
.wc
,
3226 &ufid
, DPIF_UC_MISS
, NULL
, &actions
,
3228 if (OVS_UNLIKELY(error
&& error
!= ENOSPC
)) {
3229 dp_packet_delete(packets
[i
]);
3234 /* We can't allow the packet batching in the next loop to execute
3235 * the actions. Otherwise, if there are any slow path actions,
3236 * we'll send the packet up twice. */
3237 dp_netdev_execute_actions(pmd
, &packets
[i
], 1, true,
3238 actions
.data
, actions
.size
);
3240 add_actions
= put_actions
.size
? &put_actions
: &actions
;
3241 if (OVS_LIKELY(error
!= ENOSPC
)) {
3242 /* XXX: There's a race window where a flow covering this packet
3243 * could have already been installed since we last did the flow
3244 * lookup before upcall. This could be solved by moving the
3245 * mutex lock outside the loop, but that's an awful long time
3246 * to be locking everyone out of making flow installs. If we
3247 * move to a per-core classifier, it would be reasonable. */
3248 ovs_mutex_lock(&pmd
->flow_mutex
);
3249 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, &keys
[i
]);
3250 if (OVS_LIKELY(!netdev_flow
)) {
3251 netdev_flow
= dp_netdev_flow_add(pmd
, &match
, &ufid
,
3255 ovs_mutex_unlock(&pmd
->flow_mutex
);
3257 emc_insert(flow_cache
, &keys
[i
], netdev_flow
);
3261 ofpbuf_uninit(&actions
);
3262 ofpbuf_uninit(&put_actions
);
3263 fat_rwlock_unlock(&dp
->upcall_rwlock
);
3264 dp_netdev_count_packet(pmd
, DP_STAT_MISS
, miss_cnt
);
3265 dp_netdev_count_packet(pmd
, DP_STAT_LOST
, lost_cnt
);
3266 } else if (OVS_UNLIKELY(any_miss
)) {
3267 int dropped_cnt
= 0;
3269 for (i
= 0; i
< cnt
; i
++) {
3270 if (OVS_UNLIKELY(!rules
[i
])) {
3271 dp_packet_delete(packets
[i
]);
3276 dp_netdev_count_packet(pmd
, DP_STAT_MISS
, dropped_cnt
);
3277 dp_netdev_count_packet(pmd
, DP_STAT_LOST
, dropped_cnt
);
3281 for (i
= 0; i
< cnt
; i
++) {
3282 struct dp_packet
*packet
= packets
[i
];
3283 struct dp_netdev_flow
*flow
;
3285 if (OVS_UNLIKELY(!rules
[i
])) {
3289 flow
= dp_netdev_flow_cast(rules
[i
]);
3291 emc_insert(flow_cache
, &keys
[i
], flow
);
3292 dp_netdev_queue_batches(packet
, flow
, &keys
[i
].mf
, batches
,
3293 &n_batches
, ARRAY_SIZE(batches
));
3296 for (i
= 0; i
< n_batches
; i
++) {
3297 packet_batch_execute(&batches
[i
], pmd
, DP_STAT_MASKED_HIT
, now
);
3302 dp_netdev_input(struct dp_netdev_pmd_thread
*pmd
,
3303 struct dp_packet
**packets
, int cnt
)
3305 #if !defined(__CHECKER__) && !defined(_WIN32)
3306 const size_t PKT_ARRAY_SIZE
= cnt
;
3308 /* Sparse or MSVC doesn't like variable length array. */
3309 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_RX_BATCH
};
3311 struct netdev_flow_key keys
[PKT_ARRAY_SIZE
];
3312 long long now
= time_msec();
3315 newcnt
= emc_processing(pmd
, packets
, cnt
, keys
, now
);
3316 if (OVS_UNLIKELY(newcnt
)) {
3317 fast_path_processing(pmd
, packets
, newcnt
, keys
, now
);
3321 struct dp_netdev_execute_aux
{
3322 struct dp_netdev_pmd_thread
*pmd
;
3326 dpif_netdev_register_upcall_cb(struct dpif
*dpif
, upcall_callback
*cb
,
3329 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3330 dp
->upcall_aux
= aux
;
3335 dp_netdev_drop_packets(struct dp_packet
** packets
, int cnt
, bool may_steal
)
3340 for (i
= 0; i
< cnt
; i
++) {
3341 dp_packet_delete(packets
[i
]);
3347 push_tnl_action(const struct dp_netdev
*dp
,
3348 const struct nlattr
*attr
,
3349 struct dp_packet
**packets
, int cnt
)
3351 struct dp_netdev_port
*tun_port
;
3352 const struct ovs_action_push_tnl
*data
;
3354 data
= nl_attr_get(attr
);
3356 tun_port
= dp_netdev_lookup_port(dp
, u32_to_odp(data
->tnl_port
));
3360 netdev_push_header(tun_port
->netdev
, packets
, cnt
, data
);
3366 dp_netdev_clone_pkt_batch(struct dp_packet
**tnl_pkt
,
3367 struct dp_packet
**packets
, int cnt
)
3371 for (i
= 0; i
< cnt
; i
++) {
3372 tnl_pkt
[i
] = dp_packet_clone(packets
[i
]);
3377 dp_execute_cb(void *aux_
, struct dp_packet
**packets
, int cnt
,
3378 const struct nlattr
*a
, bool may_steal
)
3379 OVS_NO_THREAD_SAFETY_ANALYSIS
3381 struct dp_netdev_execute_aux
*aux
= aux_
;
3382 uint32_t *depth
= recirc_depth_get();
3383 struct dp_netdev_pmd_thread
*pmd
= aux
->pmd
;
3384 struct dp_netdev
*dp
= pmd
->dp
;
3385 int type
= nl_attr_type(a
);
3386 struct dp_netdev_port
*p
;
3389 switch ((enum ovs_action_attr
)type
) {
3390 case OVS_ACTION_ATTR_OUTPUT
:
3391 p
= dp_netdev_lookup_port(dp
, u32_to_odp(nl_attr_get_u32(a
)));
3392 if (OVS_LIKELY(p
)) {
3393 netdev_send(p
->netdev
, pmd
->core_id
, packets
, cnt
, may_steal
);
3398 case OVS_ACTION_ATTR_TUNNEL_PUSH
:
3399 if (*depth
< MAX_RECIRC_DEPTH
) {
3400 struct dp_packet
*tnl_pkt
[NETDEV_MAX_RX_BATCH
];
3404 dp_netdev_clone_pkt_batch(tnl_pkt
, packets
, cnt
);
3408 err
= push_tnl_action(dp
, a
, packets
, cnt
);
3411 dp_netdev_input(pmd
, packets
, cnt
);
3414 dp_netdev_drop_packets(tnl_pkt
, cnt
, !may_steal
);
3420 case OVS_ACTION_ATTR_TUNNEL_POP
:
3421 if (*depth
< MAX_RECIRC_DEPTH
) {
3422 odp_port_t portno
= u32_to_odp(nl_attr_get_u32(a
));
3424 p
= dp_netdev_lookup_port(dp
, portno
);
3426 struct dp_packet
*tnl_pkt
[NETDEV_MAX_RX_BATCH
];
3430 dp_netdev_clone_pkt_batch(tnl_pkt
, packets
, cnt
);
3434 err
= netdev_pop_header(p
->netdev
, packets
, cnt
);
3437 for (i
= 0; i
< cnt
; i
++) {
3438 packets
[i
]->md
.in_port
.odp_port
= portno
;
3442 dp_netdev_input(pmd
, packets
, cnt
);
3445 dp_netdev_drop_packets(tnl_pkt
, cnt
, !may_steal
);
3452 case OVS_ACTION_ATTR_USERSPACE
:
3453 if (!fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
3454 const struct nlattr
*userdata
;
3455 struct ofpbuf actions
;
3459 userdata
= nl_attr_find_nested(a
, OVS_USERSPACE_ATTR_USERDATA
);
3460 ofpbuf_init(&actions
, 0);
3462 for (i
= 0; i
< cnt
; i
++) {
3465 ofpbuf_clear(&actions
);
3467 flow_extract(packets
[i
], &flow
);
3468 dpif_flow_hash(dp
->dpif
, &flow
, sizeof flow
, &ufid
);
3469 error
= dp_netdev_upcall(pmd
, packets
[i
], &flow
, NULL
, &ufid
,
3470 DPIF_UC_ACTION
, userdata
,&actions
,
3472 if (!error
|| error
== ENOSPC
) {
3473 dp_netdev_execute_actions(pmd
, &packets
[i
], 1, may_steal
,
3474 actions
.data
, actions
.size
);
3475 } else if (may_steal
) {
3476 dp_packet_delete(packets
[i
]);
3479 ofpbuf_uninit(&actions
);
3480 fat_rwlock_unlock(&dp
->upcall_rwlock
);
3486 case OVS_ACTION_ATTR_RECIRC
:
3487 if (*depth
< MAX_RECIRC_DEPTH
) {
3490 for (i
= 0; i
< cnt
; i
++) {
3491 struct dp_packet
*recirc_pkt
;
3493 recirc_pkt
= (may_steal
) ? packets
[i
]
3494 : dp_packet_clone(packets
[i
]);
3496 recirc_pkt
->md
.recirc_id
= nl_attr_get_u32(a
);
3498 dp_netdev_input(pmd
, &recirc_pkt
, 1);
3505 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3508 case OVS_ACTION_ATTR_PUSH_VLAN
:
3509 case OVS_ACTION_ATTR_POP_VLAN
:
3510 case OVS_ACTION_ATTR_PUSH_MPLS
:
3511 case OVS_ACTION_ATTR_POP_MPLS
:
3512 case OVS_ACTION_ATTR_SET
:
3513 case OVS_ACTION_ATTR_SET_MASKED
:
3514 case OVS_ACTION_ATTR_SAMPLE
:
3515 case OVS_ACTION_ATTR_HASH
:
3516 case OVS_ACTION_ATTR_UNSPEC
:
3517 case __OVS_ACTION_ATTR_MAX
:
3521 dp_netdev_drop_packets(packets
, cnt
, may_steal
);
3525 dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
3526 struct dp_packet
**packets
, int cnt
,
3528 const struct nlattr
*actions
, size_t actions_len
)
3530 struct dp_netdev_execute_aux aux
= { pmd
};
3532 odp_execute_actions(&aux
, packets
, cnt
, may_steal
, actions
,
3533 actions_len
, dp_execute_cb
);
3536 const struct dpif_class dpif_netdev_class
= {
3539 dpif_netdev_enumerate
,
3540 dpif_netdev_port_open_type
,
3543 dpif_netdev_destroy
,
3546 dpif_netdev_get_stats
,
3547 dpif_netdev_port_add
,
3548 dpif_netdev_port_del
,
3549 dpif_netdev_port_query_by_number
,
3550 dpif_netdev_port_query_by_name
,
3551 NULL
, /* port_get_pid */
3552 dpif_netdev_port_dump_start
,
3553 dpif_netdev_port_dump_next
,
3554 dpif_netdev_port_dump_done
,
3555 dpif_netdev_port_poll
,
3556 dpif_netdev_port_poll_wait
,
3557 dpif_netdev_flow_flush
,
3558 dpif_netdev_flow_dump_create
,
3559 dpif_netdev_flow_dump_destroy
,
3560 dpif_netdev_flow_dump_thread_create
,
3561 dpif_netdev_flow_dump_thread_destroy
,
3562 dpif_netdev_flow_dump_next
,
3563 dpif_netdev_operate
,
3564 NULL
, /* recv_set */
3565 NULL
, /* handlers_set */
3566 dpif_netdev_pmd_set
,
3567 dpif_netdev_queue_to_priority
,
3569 NULL
, /* recv_wait */
3570 NULL
, /* recv_purge */
3571 dpif_netdev_register_upcall_cb
,
3572 dpif_netdev_enable_upcall
,
3573 dpif_netdev_disable_upcall
,
3574 dpif_netdev_get_datapath_version
,
3578 dpif_dummy_change_port_number(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
3579 const char *argv
[], void *aux OVS_UNUSED
)
3581 struct dp_netdev_port
*old_port
;
3582 struct dp_netdev_port
*new_port
;
3583 struct dp_netdev
*dp
;
3586 ovs_mutex_lock(&dp_netdev_mutex
);
3587 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
3588 if (!dp
|| !dpif_netdev_class_is_dummy(dp
->class)) {
3589 ovs_mutex_unlock(&dp_netdev_mutex
);
3590 unixctl_command_reply_error(conn
, "unknown datapath or not a dummy");
3593 ovs_refcount_ref(&dp
->ref_cnt
);
3594 ovs_mutex_unlock(&dp_netdev_mutex
);
3596 ovs_mutex_lock(&dp
->port_mutex
);
3597 if (get_port_by_name(dp
, argv
[2], &old_port
)) {
3598 unixctl_command_reply_error(conn
, "unknown port");
3602 port_no
= u32_to_odp(atoi(argv
[3]));
3603 if (!port_no
|| port_no
== ODPP_NONE
) {
3604 unixctl_command_reply_error(conn
, "bad port number");
3607 if (dp_netdev_lookup_port(dp
, port_no
)) {
3608 unixctl_command_reply_error(conn
, "port number already in use");
3612 /* Remove old port. */
3613 cmap_remove(&dp
->ports
, &old_port
->node
, hash_port_no(old_port
->port_no
));
3614 ovsrcu_postpone(free
, old_port
);
3616 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3617 new_port
= xmemdup(old_port
, sizeof *old_port
);
3618 new_port
->port_no
= port_no
;
3619 cmap_insert(&dp
->ports
, &new_port
->node
, hash_port_no(port_no
));
3621 seq_change(dp
->port_seq
);
3622 unixctl_command_reply(conn
, NULL
);
3625 ovs_mutex_unlock(&dp
->port_mutex
);
3626 dp_netdev_unref(dp
);
3630 dpif_dummy_delete_port(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
3631 const char *argv
[], void *aux OVS_UNUSED
)
3633 struct dp_netdev_port
*port
;
3634 struct dp_netdev
*dp
;
3636 ovs_mutex_lock(&dp_netdev_mutex
);
3637 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
3638 if (!dp
|| !dpif_netdev_class_is_dummy(dp
->class)) {
3639 ovs_mutex_unlock(&dp_netdev_mutex
);
3640 unixctl_command_reply_error(conn
, "unknown datapath or not a dummy");
3643 ovs_refcount_ref(&dp
->ref_cnt
);
3644 ovs_mutex_unlock(&dp_netdev_mutex
);
3646 ovs_mutex_lock(&dp
->port_mutex
);
3647 if (get_port_by_name(dp
, argv
[2], &port
)) {
3648 unixctl_command_reply_error(conn
, "unknown port");
3649 } else if (port
->port_no
== ODPP_LOCAL
) {
3650 unixctl_command_reply_error(conn
, "can't delete local port");
3652 do_del_port(dp
, port
);
3653 unixctl_command_reply(conn
, NULL
);
3655 ovs_mutex_unlock(&dp
->port_mutex
);
3657 dp_netdev_unref(dp
);
3661 dpif_dummy_register__(const char *type
)
3663 struct dpif_class
*class;
3665 class = xmalloc(sizeof *class);
3666 *class = dpif_netdev_class
;
3667 class->type
= xstrdup(type
);
3668 dp_register_provider(class);
3672 dpif_dummy_register(bool override
)
3679 dp_enumerate_types(&types
);
3680 SSET_FOR_EACH (type
, &types
) {
3681 if (!dp_unregister_provider(type
)) {
3682 dpif_dummy_register__(type
);
3685 sset_destroy(&types
);
3688 dpif_dummy_register__("dummy");
3690 unixctl_command_register("dpif-dummy/change-port-number",
3691 "dp port new-number",
3692 3, 3, dpif_dummy_change_port_number
, NULL
);
3693 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3694 2, 2, dpif_dummy_delete_port
, NULL
);
3697 /* Datapath Classifier. */
3699 /* A set of rules that all have the same fields wildcarded. */
3700 struct dpcls_subtable
{
3701 /* The fields are only used by writers. */
3702 struct cmap_node cmap_node OVS_GUARDED
; /* Within dpcls 'subtables_map'. */
3704 /* These fields are accessed by readers. */
3705 struct cmap rules
; /* Contains "struct dpcls_rule"s. */
3706 struct netdev_flow_key mask
; /* Wildcards for fields (const). */
3707 /* 'mask' must be the last field, additional space is allocated here. */
3710 /* Initializes 'cls' as a classifier that initially contains no classification
3713 dpcls_init(struct dpcls
*cls
)
3715 cmap_init(&cls
->subtables_map
);
3716 pvector_init(&cls
->subtables
);
3720 dpcls_destroy_subtable(struct dpcls
*cls
, struct dpcls_subtable
*subtable
)
3722 pvector_remove(&cls
->subtables
, subtable
);
3723 cmap_remove(&cls
->subtables_map
, &subtable
->cmap_node
,
3724 subtable
->mask
.hash
);
3725 cmap_destroy(&subtable
->rules
);
3726 ovsrcu_postpone(free
, subtable
);
3729 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3730 * caller's responsibility.
3731 * May only be called after all the readers have been terminated. */
3733 dpcls_destroy(struct dpcls
*cls
)
3736 struct dpcls_subtable
*subtable
;
3738 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
3739 dpcls_destroy_subtable(cls
, subtable
);
3741 cmap_destroy(&cls
->subtables_map
);
3742 pvector_destroy(&cls
->subtables
);
3746 static struct dpcls_subtable
*
3747 dpcls_create_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
3749 struct dpcls_subtable
*subtable
;
3751 /* Need to add one. */
3752 subtable
= xmalloc(sizeof *subtable
3753 - sizeof subtable
->mask
.mf
+ mask
->len
);
3754 cmap_init(&subtable
->rules
);
3755 netdev_flow_key_clone(&subtable
->mask
, mask
);
3756 cmap_insert(&cls
->subtables_map
, &subtable
->cmap_node
, mask
->hash
);
3757 pvector_insert(&cls
->subtables
, subtable
, 0);
3758 pvector_publish(&cls
->subtables
);
3763 static inline struct dpcls_subtable
*
3764 dpcls_find_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
3766 struct dpcls_subtable
*subtable
;
3768 CMAP_FOR_EACH_WITH_HASH (subtable
, cmap_node
, mask
->hash
,
3769 &cls
->subtables_map
) {
3770 if (netdev_flow_key_equal(&subtable
->mask
, mask
)) {
3774 return dpcls_create_subtable(cls
, mask
);
3777 /* Insert 'rule' into 'cls'. */
3779 dpcls_insert(struct dpcls
*cls
, struct dpcls_rule
*rule
,
3780 const struct netdev_flow_key
*mask
)
3782 struct dpcls_subtable
*subtable
= dpcls_find_subtable(cls
, mask
);
3784 rule
->mask
= &subtable
->mask
;
3785 cmap_insert(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
);
3788 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3790 dpcls_remove(struct dpcls
*cls
, struct dpcls_rule
*rule
)
3792 struct dpcls_subtable
*subtable
;
3794 ovs_assert(rule
->mask
);
3796 INIT_CONTAINER(subtable
, rule
->mask
, mask
);
3798 if (cmap_remove(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
)
3800 dpcls_destroy_subtable(cls
, subtable
);
3801 pvector_publish(&cls
->subtables
);
3805 /* Returns true if 'target' satisifies 'key' in 'mask', that is, if each 1-bit
3806 * in 'mask' the values in 'key' and 'target' are the same.
3808 * Note: 'key' and 'mask' have the same mask, and 'key' is already masked. */
3810 dpcls_rule_matches_key(const struct dpcls_rule
*rule
,
3811 const struct netdev_flow_key
*target
)
3813 const uint64_t *keyp
= rule
->flow
.mf
.inline_values
;
3814 const uint64_t *maskp
= rule
->mask
->mf
.inline_values
;
3815 uint64_t target_u64
;
3817 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(target_u64
, target
, rule
->flow
.mf
.map
) {
3818 if (OVS_UNLIKELY((target_u64
& *maskp
++) != *keyp
++)) {
3825 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3826 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3827 * NULL it is skipped.
3829 * This function is optimized for use in the userspace datapath and therefore
3830 * does not implement a lot of features available in the standard
3831 * classifier_lookup() function. Specifically, it does not implement
3832 * priorities, instead returning any rule which matches the flow.
3834 * Returns true if all flows found a corresponding rule. */
3836 dpcls_lookup(const struct dpcls
*cls
, const struct netdev_flow_key keys
[],
3837 struct dpcls_rule
**rules
, const size_t cnt
)
3839 /* The batch size 16 was experimentally found faster than 8 or 32. */
3840 typedef uint16_t map_type
;
3841 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3843 #if !defined(__CHECKER__) && !defined(_WIN32)
3844 const int N_MAPS
= DIV_ROUND_UP(cnt
, MAP_BITS
);
3846 enum { N_MAPS
= DIV_ROUND_UP(NETDEV_MAX_RX_BATCH
, MAP_BITS
) };
3848 map_type maps
[N_MAPS
];
3849 struct dpcls_subtable
*subtable
;
3851 memset(maps
, 0xff, sizeof maps
);
3852 if (cnt
% MAP_BITS
) {
3853 maps
[N_MAPS
- 1] >>= MAP_BITS
- cnt
% MAP_BITS
; /* Clear extra bits. */
3855 memset(rules
, 0, cnt
* sizeof *rules
);
3857 PVECTOR_FOR_EACH (subtable
, &cls
->subtables
) {
3858 const struct netdev_flow_key
*mkeys
= keys
;
3859 struct dpcls_rule
**mrules
= rules
;
3860 map_type remains
= 0;
3863 BUILD_ASSERT_DECL(sizeof remains
== sizeof *maps
);
3865 for (m
= 0; m
< N_MAPS
; m
++, mkeys
+= MAP_BITS
, mrules
+= MAP_BITS
) {
3866 uint32_t hashes
[MAP_BITS
];
3867 const struct cmap_node
*nodes
[MAP_BITS
];
3868 unsigned long map
= maps
[m
];
3872 continue; /* Skip empty maps. */
3875 /* Compute hashes for the remaining keys. */
3876 ULONG_FOR_EACH_1(i
, map
) {
3877 hashes
[i
] = netdev_flow_key_hash_in_mask(&mkeys
[i
],
3881 map
= cmap_find_batch(&subtable
->rules
, map
, hashes
, nodes
);
3882 /* Check results. */
3883 ULONG_FOR_EACH_1(i
, map
) {
3884 struct dpcls_rule
*rule
;
3886 CMAP_NODE_FOR_EACH (rule
, cmap_node
, nodes
[i
]) {
3887 if (OVS_LIKELY(dpcls_rule_matches_key(rule
, &mkeys
[i
]))) {
3892 ULONG_SET0(map
, i
); /* Did not match. */
3894 ; /* Keep Sparse happy. */
3896 maps
[m
] &= ~map
; /* Clear the found rules. */
3900 return true; /* All found. */
3903 return false; /* Some misses. */