2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2016, 2017 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 #include "dpif-netdev.h"
25 #include <netinet/in.h>
29 #include <sys/ioctl.h>
30 #include <sys/socket.h>
35 #include <rte_cycles.h>
40 #include "conntrack.h"
44 #include "dp-packet.h"
46 #include "dpif-provider.h"
48 #include "fat-rwlock.h"
54 #include "netdev-vport.h"
56 #include "odp-execute.h"
58 #include "openvswitch/dynamic-string.h"
59 #include "openvswitch/list.h"
60 #include "openvswitch/match.h"
61 #include "openvswitch/ofp-print.h"
62 #include "openvswitch/ofp-util.h"
63 #include "openvswitch/ofpbuf.h"
64 #include "openvswitch/shash.h"
65 #include "openvswitch/vlog.h"
69 #include "openvswitch/poll-loop.h"
76 #include "tnl-neigh-cache.h"
77 #include "tnl-ports.h"
81 VLOG_DEFINE_THIS_MODULE(dpif_netdev
);
83 #define FLOW_DUMP_MAX_BATCH 50
84 /* Use per thread recirc_depth to prevent recirculation loop. */
85 #define MAX_RECIRC_DEPTH 6
86 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth
, 0)
88 /* Configuration parameters. */
89 enum { MAX_FLOWS
= 65536 }; /* Maximum number of flows in flow table. */
90 enum { MAX_METERS
= 65536 }; /* Maximum number of meters. */
91 enum { MAX_BANDS
= 8 }; /* Maximum number of bands / meter. */
92 enum { N_METER_LOCKS
= 64 }; /* Maximum number of meters. */
94 /* Protects against changes to 'dp_netdevs'. */
95 static struct ovs_mutex dp_netdev_mutex
= OVS_MUTEX_INITIALIZER
;
97 /* Contains all 'struct dp_netdev's. */
98 static struct shash dp_netdevs
OVS_GUARDED_BY(dp_netdev_mutex
)
99 = SHASH_INITIALIZER(&dp_netdevs
);
101 static struct vlog_rate_limit upcall_rl
= VLOG_RATE_LIMIT_INIT(600, 600);
103 #define DP_NETDEV_CS_SUPPORTED_MASK (CS_NEW | CS_ESTABLISHED | CS_RELATED \
104 | CS_INVALID | CS_REPLY_DIR | CS_TRACKED \
105 | CS_SRC_NAT | CS_DST_NAT)
106 #define DP_NETDEV_CS_UNSUPPORTED_MASK (~(uint32_t)DP_NETDEV_CS_SUPPORTED_MASK)
108 static struct odp_support dp_netdev_support
= {
109 .max_vlan_headers
= SIZE_MAX
,
110 .max_mpls_depth
= SIZE_MAX
,
116 .ct_state_nat
= true,
117 .ct_orig_tuple
= true,
118 .ct_orig_tuple6
= true,
121 /* Stores a miniflow with inline values */
123 struct netdev_flow_key
{
124 uint32_t hash
; /* Hash function differs for different users. */
125 uint32_t len
; /* Length of the following miniflow (incl. map). */
127 uint64_t buf
[FLOW_MAX_PACKET_U64S
];
130 /* Exact match cache for frequently used flows
132 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
133 * search its entries for a miniflow that matches exactly the miniflow of the
134 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
136 * A cache entry holds a reference to its 'dp_netdev_flow'.
138 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
139 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
140 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
141 * value is the index of a cache entry where the miniflow could be.
147 * Each pmd_thread has its own private exact match cache.
148 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
151 #define EM_FLOW_HASH_SHIFT 13
152 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
153 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
154 #define EM_FLOW_HASH_SEGS 2
156 /* Default EMC insert probability is 1 / DEFAULT_EM_FLOW_INSERT_INV_PROB */
157 #define DEFAULT_EM_FLOW_INSERT_INV_PROB 100
158 #define DEFAULT_EM_FLOW_INSERT_MIN (UINT32_MAX / \
159 DEFAULT_EM_FLOW_INSERT_INV_PROB)
162 struct dp_netdev_flow
*flow
;
163 struct netdev_flow_key key
; /* key.hash used for emc hash value. */
167 struct emc_entry entries
[EM_FLOW_HASH_ENTRIES
];
168 int sweep_idx
; /* For emc_cache_slow_sweep(). */
171 /* Iterate in the exact match cache through every entry that might contain a
172 * miniflow with hash 'HASH'. */
173 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
174 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
175 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
176 i__ < EM_FLOW_HASH_SEGS; \
177 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
179 /* Simple non-wildcarding single-priority classifier. */
181 /* Time in ms between successive optimizations of the dpcls subtable vector */
182 #define DPCLS_OPTIMIZATION_INTERVAL 1000
184 /* Time in ms of the interval in which rxq processing cycles used in
185 * rxq to pmd assignments is measured and stored. */
186 #define PMD_RXQ_INTERVAL_LEN 10000
188 /* Number of intervals for which cycles are stored
189 * and used during rxq to pmd assignment. */
190 #define PMD_RXQ_INTERVAL_MAX 6
193 struct cmap_node node
; /* Within dp_netdev_pmd_thread.classifiers */
195 struct cmap subtables_map
;
196 struct pvector subtables
;
199 /* A rule to be inserted to the classifier. */
201 struct cmap_node cmap_node
; /* Within struct dpcls_subtable 'rules'. */
202 struct netdev_flow_key
*mask
; /* Subtable's mask. */
203 struct netdev_flow_key flow
; /* Matching key. */
204 /* 'flow' must be the last field, additional space is allocated here. */
207 static void dpcls_init(struct dpcls
*);
208 static void dpcls_destroy(struct dpcls
*);
209 static void dpcls_sort_subtable_vector(struct dpcls
*);
210 static void dpcls_insert(struct dpcls
*, struct dpcls_rule
*,
211 const struct netdev_flow_key
*mask
);
212 static void dpcls_remove(struct dpcls
*, struct dpcls_rule
*);
213 static bool dpcls_lookup(struct dpcls
*cls
,
214 const struct netdev_flow_key keys
[],
215 struct dpcls_rule
**rules
, size_t cnt
,
218 /* Set of supported meter flags */
219 #define DP_SUPPORTED_METER_FLAGS_MASK \
220 (OFPMF13_STATS | OFPMF13_PKTPS | OFPMF13_KBPS | OFPMF13_BURST)
222 /* Set of supported meter band types */
223 #define DP_SUPPORTED_METER_BAND_TYPES \
224 ( 1 << OFPMBT13_DROP )
226 struct dp_meter_band
{
227 struct ofputil_meter_band up
; /* type, prec_level, pad, rate, burst_size */
228 uint32_t bucket
; /* In 1/1000 packets (for PKTPS), or in bits (for KBPS) */
229 uint64_t packet_count
;
236 uint32_t max_delta_t
;
238 uint64_t packet_count
;
240 struct dp_meter_band bands
[];
243 /* Datapath based on the network device interface from netdev.h.
249 * Some members, marked 'const', are immutable. Accessing other members
250 * requires synchronization, as noted in more detail below.
252 * Acquisition order is, from outermost to innermost:
254 * dp_netdev_mutex (global)
259 const struct dpif_class
*const class;
260 const char *const name
;
262 struct ovs_refcount ref_cnt
;
263 atomic_flag destroyed
;
267 * Any lookup into 'ports' or any access to the dp_netdev_ports found
268 * through 'ports' requires taking 'port_mutex'. */
269 struct ovs_mutex port_mutex
;
271 struct seq
*port_seq
; /* Incremented whenever a port changes. */
274 struct ovs_mutex meter_locks
[N_METER_LOCKS
];
275 struct dp_meter
*meters
[MAX_METERS
]; /* Meter bands. */
277 /* Probability of EMC insertions is a factor of 'emc_insert_min'.*/
278 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
) atomic_uint32_t emc_insert_min
;
280 /* Protects access to ofproto-dpif-upcall interface during revalidator
281 * thread synchronization. */
282 struct fat_rwlock upcall_rwlock
;
283 upcall_callback
*upcall_cb
; /* Callback function for executing upcalls. */
286 /* Callback function for notifying the purging of dp flows (during
287 * reseting pmd deletion). */
288 dp_purge_callback
*dp_purge_cb
;
291 /* Stores all 'struct dp_netdev_pmd_thread's. */
292 struct cmap poll_threads
;
293 /* id pool for per thread static_tx_qid. */
294 struct id_pool
*tx_qid_pool
;
295 struct ovs_mutex tx_qid_pool_mutex
;
297 /* Protects the access of the 'struct dp_netdev_pmd_thread'
298 * instance for non-pmd thread. */
299 struct ovs_mutex non_pmd_mutex
;
301 /* Each pmd thread will store its pointer to
302 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
303 ovsthread_key_t per_pmd_key
;
305 struct seq
*reconfigure_seq
;
306 uint64_t last_reconfigure_seq
;
308 /* Cpu mask for pin of pmd threads. */
311 uint64_t last_tnl_conf_seq
;
313 struct conntrack conntrack
;
316 static void meter_lock(const struct dp_netdev
*dp
, uint32_t meter_id
)
317 OVS_ACQUIRES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
319 ovs_mutex_lock(&dp
->meter_locks
[meter_id
% N_METER_LOCKS
]);
322 static void meter_unlock(const struct dp_netdev
*dp
, uint32_t meter_id
)
323 OVS_RELEASES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
325 ovs_mutex_unlock(&dp
->meter_locks
[meter_id
% N_METER_LOCKS
]);
329 static struct dp_netdev_port
*dp_netdev_lookup_port(const struct dp_netdev
*dp
,
331 OVS_REQUIRES(dp
->port_mutex
);
334 DP_STAT_EXACT_HIT
, /* Packets that had an exact match (emc). */
335 DP_STAT_MASKED_HIT
, /* Packets that matched in the flow table. */
336 DP_STAT_MISS
, /* Packets that did not match. */
337 DP_STAT_LOST
, /* Packets not passed up to the client. */
338 DP_STAT_LOOKUP_HIT
, /* Number of subtable lookups for flow table
343 enum pmd_cycles_counter_type
{
344 PMD_CYCLES_IDLE
, /* Cycles spent idle or unsuccessful polling */
345 PMD_CYCLES_PROCESSING
, /* Cycles spent successfully polling and
346 * processing polled packets */
350 enum rxq_cycles_counter_type
{
351 RXQ_CYCLES_PROC_CURR
, /* Cycles spent successfully polling and
352 processing packets during the current
354 RXQ_CYCLES_PROC_HIST
, /* Total cycles of all intervals that are used
355 during rxq to pmd assignment. */
359 #define XPS_TIMEOUT_MS 500LL
361 /* Contained by struct dp_netdev_port's 'rxqs' member. */
362 struct dp_netdev_rxq
{
363 struct dp_netdev_port
*port
;
364 struct netdev_rxq
*rx
;
365 unsigned core_id
; /* Core to which this queue should be
366 pinned. OVS_CORE_UNSPEC if the
367 queue doesn't need to be pinned to a
369 unsigned intrvl_idx
; /* Write index for 'cycles_intrvl'. */
370 struct dp_netdev_pmd_thread
*pmd
; /* pmd thread that polls this queue. */
372 /* Counters of cycles spent successfully polling and processing pkts. */
373 atomic_ullong cycles
[RXQ_N_CYCLES
];
374 /* We store PMD_RXQ_INTERVAL_MAX intervals of data for an rxq and then
375 sum them to yield the cycles used for an rxq. */
376 atomic_ullong cycles_intrvl
[PMD_RXQ_INTERVAL_MAX
];
379 /* A port in a netdev-based datapath. */
380 struct dp_netdev_port
{
382 bool dynamic_txqs
; /* If true XPS will be used. */
383 bool need_reconfigure
; /* True if we should reconfigure netdev. */
384 struct netdev
*netdev
;
385 struct hmap_node node
; /* Node in dp_netdev's 'ports'. */
386 struct netdev_saved_flags
*sf
;
387 struct dp_netdev_rxq
*rxqs
;
388 unsigned n_rxq
; /* Number of elements in 'rxqs' */
389 unsigned *txq_used
; /* Number of threads that use each tx queue. */
390 struct ovs_mutex txq_used_mutex
;
391 char *type
; /* Port type as requested by user. */
392 char *rxq_affinity_list
; /* Requested affinity of rx queues. */
395 /* Contained by struct dp_netdev_flow's 'stats' member. */
396 struct dp_netdev_flow_stats
{
397 atomic_llong used
; /* Last used time, in monotonic msecs. */
398 atomic_ullong packet_count
; /* Number of packets matched. */
399 atomic_ullong byte_count
; /* Number of bytes matched. */
400 atomic_uint16_t tcp_flags
; /* Bitwise-OR of seen tcp_flags values. */
403 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
409 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
410 * its pmd thread's classifier. The text below calls this classifier 'cls'.
415 * The thread safety rules described here for "struct dp_netdev_flow" are
416 * motivated by two goals:
418 * - Prevent threads that read members of "struct dp_netdev_flow" from
419 * reading bad data due to changes by some thread concurrently modifying
422 * - Prevent two threads making changes to members of a given "struct
423 * dp_netdev_flow" from interfering with each other.
429 * A flow 'flow' may be accessed without a risk of being freed during an RCU
430 * grace period. Code that needs to hold onto a flow for a while
431 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
433 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
434 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
437 * Some members, marked 'const', are immutable. Accessing other members
438 * requires synchronization, as noted in more detail below.
440 struct dp_netdev_flow
{
441 const struct flow flow
; /* Unmasked flow that created this entry. */
442 /* Hash table index by unmasked flow. */
443 const struct cmap_node node
; /* In owning dp_netdev_pmd_thread's */
445 const ovs_u128 ufid
; /* Unique flow identifier. */
446 const unsigned pmd_id
; /* The 'core_id' of pmd thread owning this */
449 /* Number of references.
450 * The classifier owns one reference.
451 * Any thread trying to keep a rule from being freed should hold its own
453 struct ovs_refcount ref_cnt
;
458 struct dp_netdev_flow_stats stats
;
461 OVSRCU_TYPE(struct dp_netdev_actions
*) actions
;
463 /* While processing a group of input packets, the datapath uses the next
464 * member to store a pointer to the output batch for the flow. It is
465 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
466 * packet_batch_per_flow_init() and packet_batch_per_flow_execute()). */
467 struct packet_batch_per_flow
*batch
;
469 /* Packet classification. */
470 struct dpcls_rule cr
; /* In owning dp_netdev's 'cls'. */
471 /* 'cr' must be the last member. */
474 static void dp_netdev_flow_unref(struct dp_netdev_flow
*);
475 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*);
476 static int dpif_netdev_flow_from_nlattrs(const struct nlattr
*, uint32_t,
477 struct flow
*, bool);
479 /* A set of datapath actions within a "struct dp_netdev_flow".
485 * A struct dp_netdev_actions 'actions' is protected with RCU. */
486 struct dp_netdev_actions
{
487 /* These members are immutable: they do not change during the struct's
489 unsigned int size
; /* Size of 'actions', in bytes. */
490 struct nlattr actions
[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
493 struct dp_netdev_actions
*dp_netdev_actions_create(const struct nlattr
*,
495 struct dp_netdev_actions
*dp_netdev_flow_get_actions(
496 const struct dp_netdev_flow
*);
497 static void dp_netdev_actions_free(struct dp_netdev_actions
*);
499 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
500 struct dp_netdev_pmd_stats
{
501 /* Indexed by DP_STAT_*. */
502 atomic_ullong n
[DP_N_STATS
];
505 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
506 struct dp_netdev_pmd_cycles
{
507 /* Indexed by PMD_CYCLES_*. */
508 atomic_ullong n
[PMD_N_CYCLES
];
511 struct polled_queue
{
512 struct dp_netdev_rxq
*rxq
;
516 /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */
518 struct dp_netdev_rxq
*rxq
;
519 struct hmap_node node
;
522 /* Contained by struct dp_netdev_pmd_thread's 'send_port_cache',
523 * 'tnl_port_cache' or 'tx_ports'. */
525 struct dp_netdev_port
*port
;
528 struct hmap_node node
;
529 struct dp_packet_batch output_pkts
;
532 /* A set of properties for the current processing loop that is not directly
533 * associated with the pmd thread itself, but with the packets being
534 * processed or the short-term system configuration (for example, time).
535 * Contained by struct dp_netdev_pmd_thread's 'ctx' member. */
536 struct dp_netdev_pmd_thread_ctx
{
537 /* Latest measured time. See 'pmd_thread_ctx_time_update()'. */
539 /* Used to count cycles. See 'cycles_count_end()' */
540 unsigned long long last_cycles
;
543 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
544 * the performance overhead of interrupt processing. Therefore netdev can
545 * not implement rx-wait for these devices. dpif-netdev needs to poll
546 * these device to check for recv buffer. pmd-thread does polling for
547 * devices assigned to itself.
549 * DPDK used PMD for accessing NIC.
551 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
552 * I/O of all non-pmd threads. There will be no actual thread created
555 * Each struct has its own flow cache and classifier per managed ingress port.
556 * For packets received on ingress port, a look up is done on corresponding PMD
557 * thread's flow cache and in case of a miss, lookup is performed in the
558 * corresponding classifier of port. Packets are executed with the found
559 * actions in either case.
561 struct dp_netdev_pmd_thread
{
562 struct dp_netdev
*dp
;
563 struct ovs_refcount ref_cnt
; /* Every reference must be refcount'ed. */
564 struct cmap_node node
; /* In 'dp->poll_threads'. */
566 pthread_cond_t cond
; /* For synchronizing pmd thread reload. */
567 struct ovs_mutex cond_mutex
; /* Mutex for condition variable. */
569 /* Per thread exact-match cache. Note, the instance for cpu core
570 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
571 * need to be protected by 'non_pmd_mutex'. Every other instance
572 * will only be accessed by its own pmd thread. */
573 struct emc_cache flow_cache
;
575 /* Flow-Table and classifiers
577 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
578 * changes to 'classifiers' must be made while still holding the
581 struct ovs_mutex flow_mutex
;
582 struct cmap flow_table OVS_GUARDED
; /* Flow table. */
584 /* One classifier per in_port polled by the pmd */
585 struct cmap classifiers
;
586 /* Periodically sort subtable vectors according to hit frequencies */
587 long long int next_optimization
;
588 /* End of the next time interval for which processing cycles
589 are stored for each polled rxq. */
590 long long int rxq_next_cycle_store
;
593 struct dp_netdev_pmd_stats stats
;
595 /* Cycles counters */
596 struct dp_netdev_pmd_cycles cycles
;
598 /* Current context of the PMD thread. */
599 struct dp_netdev_pmd_thread_ctx ctx
;
601 struct latch exit_latch
; /* For terminating the pmd thread. */
602 struct seq
*reload_seq
;
603 uint64_t last_reload_seq
;
604 atomic_bool reload
; /* Do we need to reload ports? */
606 unsigned core_id
; /* CPU core id of this pmd thread. */
607 int numa_id
; /* numa node id of this pmd thread. */
610 /* Queue id used by this pmd thread to send packets on all netdevs if
611 * XPS disabled for this netdev. All static_tx_qid's are unique and less
612 * than 'cmap_count(dp->poll_threads)'. */
613 uint32_t static_tx_qid
;
615 struct ovs_mutex port_mutex
; /* Mutex for 'poll_list' and 'tx_ports'. */
616 /* List of rx queues to poll. */
617 struct hmap poll_list OVS_GUARDED
;
618 /* Map of 'tx_port's used for transmission. Written by the main thread,
619 * read by the pmd thread. */
620 struct hmap tx_ports OVS_GUARDED
;
622 /* These are thread-local copies of 'tx_ports'. One contains only tunnel
623 * ports (that support push_tunnel/pop_tunnel), the other contains ports
624 * with at least one txq (that support send). A port can be in both.
626 * There are two separate maps to make sure that we don't try to execute
627 * OUTPUT on a device which has 0 txqs or PUSH/POP on a non-tunnel device.
629 * The instances for cpu core NON_PMD_CORE_ID can be accessed by multiple
630 * threads, and thusly need to be protected by 'non_pmd_mutex'. Every
631 * other instance will only be accessed by its own pmd thread. */
632 struct hmap tnl_port_cache
;
633 struct hmap send_port_cache
;
635 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
636 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
637 * values and subtracts them from 'stats' and 'cycles' before
638 * reporting to the user */
639 unsigned long long stats_zero
[DP_N_STATS
];
640 uint64_t cycles_zero
[PMD_N_CYCLES
];
642 /* Set to true if the pmd thread needs to be reloaded. */
646 /* Interface to netdev-based datapath. */
649 struct dp_netdev
*dp
;
650 uint64_t last_port_seq
;
653 static int get_port_by_number(struct dp_netdev
*dp
, odp_port_t port_no
,
654 struct dp_netdev_port
**portp
)
655 OVS_REQUIRES(dp
->port_mutex
);
656 static int get_port_by_name(struct dp_netdev
*dp
, const char *devname
,
657 struct dp_netdev_port
**portp
)
658 OVS_REQUIRES(dp
->port_mutex
);
659 static void dp_netdev_free(struct dp_netdev
*)
660 OVS_REQUIRES(dp_netdev_mutex
);
661 static int do_add_port(struct dp_netdev
*dp
, const char *devname
,
662 const char *type
, odp_port_t port_no
)
663 OVS_REQUIRES(dp
->port_mutex
);
664 static void do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*)
665 OVS_REQUIRES(dp
->port_mutex
);
666 static int dpif_netdev_open(const struct dpif_class
*, const char *name
,
667 bool create
, struct dpif
**);
668 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
669 struct dp_packet_batch
*,
670 bool may_steal
, const struct flow
*flow
,
671 const struct nlattr
*actions
,
673 static void dp_netdev_input(struct dp_netdev_pmd_thread
*,
674 struct dp_packet_batch
*, odp_port_t port_no
);
675 static void dp_netdev_recirculate(struct dp_netdev_pmd_thread
*,
676 struct dp_packet_batch
*);
678 static void dp_netdev_disable_upcall(struct dp_netdev
*);
679 static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
);
680 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
,
681 struct dp_netdev
*dp
, unsigned core_id
,
683 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
);
684 static void dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
685 OVS_REQUIRES(dp
->port_mutex
);
687 static void *pmd_thread_main(void *);
688 static struct dp_netdev_pmd_thread
*dp_netdev_get_pmd(struct dp_netdev
*dp
,
690 static struct dp_netdev_pmd_thread
*
691 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
);
692 static void dp_netdev_del_pmd(struct dp_netdev
*dp
,
693 struct dp_netdev_pmd_thread
*pmd
);
694 static void dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
, bool non_pmd
);
695 static void dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
);
696 static void dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
697 struct dp_netdev_port
*port
)
698 OVS_REQUIRES(pmd
->port_mutex
);
699 static void dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
701 OVS_REQUIRES(pmd
->port_mutex
);
702 static void dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
703 struct dp_netdev_rxq
*rxq
)
704 OVS_REQUIRES(pmd
->port_mutex
);
705 static void dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
706 struct rxq_poll
*poll
)
707 OVS_REQUIRES(pmd
->port_mutex
);
709 dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread
*pmd
);
711 static void reconfigure_datapath(struct dp_netdev
*dp
)
712 OVS_REQUIRES(dp
->port_mutex
);
713 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
);
714 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
);
715 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
);
716 static void pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
717 OVS_REQUIRES(pmd
->port_mutex
);
719 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
,
720 struct polled_queue
*poll_list
, int poll_cnt
);
722 dp_netdev_rxq_set_cycles(struct dp_netdev_rxq
*rx
,
723 enum rxq_cycles_counter_type type
,
724 unsigned long long cycles
);
726 dp_netdev_rxq_get_cycles(struct dp_netdev_rxq
*rx
,
727 enum rxq_cycles_counter_type type
);
729 dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq
*rx
,
730 unsigned long long cycles
);
732 dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq
*rx
, unsigned idx
);
734 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
736 static int dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
739 static inline bool emc_entry_alive(struct emc_entry
*ce
);
740 static void emc_clear_entry(struct emc_entry
*ce
);
742 static void dp_netdev_request_reconfigure(struct dp_netdev
*dp
);
745 emc_cache_init(struct emc_cache
*flow_cache
)
749 flow_cache
->sweep_idx
= 0;
750 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
751 flow_cache
->entries
[i
].flow
= NULL
;
752 flow_cache
->entries
[i
].key
.hash
= 0;
753 flow_cache
->entries
[i
].key
.len
= sizeof(struct miniflow
);
754 flowmap_init(&flow_cache
->entries
[i
].key
.mf
.map
);
759 emc_cache_uninit(struct emc_cache
*flow_cache
)
763 for (i
= 0; i
< ARRAY_SIZE(flow_cache
->entries
); i
++) {
764 emc_clear_entry(&flow_cache
->entries
[i
]);
768 /* Check and clear dead flow references slowly (one entry at each
771 emc_cache_slow_sweep(struct emc_cache
*flow_cache
)
773 struct emc_entry
*entry
= &flow_cache
->entries
[flow_cache
->sweep_idx
];
775 if (!emc_entry_alive(entry
)) {
776 emc_clear_entry(entry
);
778 flow_cache
->sweep_idx
= (flow_cache
->sweep_idx
+ 1) & EM_FLOW_HASH_MASK
;
781 /* Updates the time in PMD threads context and should be called in three cases:
783 * 1. PMD structure initialization:
784 * - dp_netdev_configure_pmd()
786 * 2. Before processing of the new packet batch:
787 * - dpif_netdev_execute()
788 * - dp_netdev_process_rxq_port()
790 * 3. At least once per polling iteration in main polling threads if no
791 * packets received on current iteration:
792 * - dpif_netdev_run()
793 * - pmd_thread_main()
795 * 'pmd->ctx.now' should be used without update in all other cases if possible.
798 pmd_thread_ctx_time_update(struct dp_netdev_pmd_thread
*pmd
)
800 pmd
->ctx
.now
= time_msec();
803 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
805 dpif_is_netdev(const struct dpif
*dpif
)
807 return dpif
->dpif_class
->open
== dpif_netdev_open
;
810 static struct dpif_netdev
*
811 dpif_netdev_cast(const struct dpif
*dpif
)
813 ovs_assert(dpif_is_netdev(dpif
));
814 return CONTAINER_OF(dpif
, struct dpif_netdev
, dpif
);
817 static struct dp_netdev
*
818 get_dp_netdev(const struct dpif
*dpif
)
820 return dpif_netdev_cast(dpif
)->dp
;
824 PMD_INFO_SHOW_STATS
, /* Show how cpu cycles are spent. */
825 PMD_INFO_CLEAR_STATS
, /* Set the cycles count to 0. */
826 PMD_INFO_SHOW_RXQ
/* Show poll-lists of pmd threads. */
830 pmd_info_show_stats(struct ds
*reply
,
831 struct dp_netdev_pmd_thread
*pmd
,
832 unsigned long long stats
[DP_N_STATS
],
833 uint64_t cycles
[PMD_N_CYCLES
])
835 unsigned long long total_packets
;
836 uint64_t total_cycles
= 0;
839 /* These loops subtracts reference values ('*_zero') from the counters.
840 * Since loads and stores are relaxed, it might be possible for a '*_zero'
841 * value to be more recent than the current value we're reading from the
842 * counter. This is not a big problem, since these numbers are not
843 * supposed to be too accurate, but we should at least make sure that
844 * the result is not negative. */
845 for (i
= 0; i
< DP_N_STATS
; i
++) {
846 if (stats
[i
] > pmd
->stats_zero
[i
]) {
847 stats
[i
] -= pmd
->stats_zero
[i
];
853 /* Sum of all the matched and not matched packets gives the total. */
854 total_packets
= stats
[DP_STAT_EXACT_HIT
] + stats
[DP_STAT_MASKED_HIT
]
855 + stats
[DP_STAT_MISS
];
857 for (i
= 0; i
< PMD_N_CYCLES
; i
++) {
858 if (cycles
[i
] > pmd
->cycles_zero
[i
]) {
859 cycles
[i
] -= pmd
->cycles_zero
[i
];
864 total_cycles
+= cycles
[i
];
867 ds_put_cstr(reply
, (pmd
->core_id
== NON_PMD_CORE_ID
)
868 ? "main thread" : "pmd thread");
870 if (pmd
->numa_id
!= OVS_NUMA_UNSPEC
) {
871 ds_put_format(reply
, " numa_id %d", pmd
->numa_id
);
873 if (pmd
->core_id
!= OVS_CORE_UNSPEC
&& pmd
->core_id
!= NON_PMD_CORE_ID
) {
874 ds_put_format(reply
, " core_id %u", pmd
->core_id
);
876 ds_put_cstr(reply
, ":\n");
879 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
880 "\tavg. subtable lookups per hit:%.2f\n"
881 "\tmiss:%llu\n\tlost:%llu\n",
882 stats
[DP_STAT_EXACT_HIT
], stats
[DP_STAT_MASKED_HIT
],
883 stats
[DP_STAT_MASKED_HIT
] > 0
884 ? (1.0*stats
[DP_STAT_LOOKUP_HIT
])/stats
[DP_STAT_MASKED_HIT
]
886 stats
[DP_STAT_MISS
], stats
[DP_STAT_LOST
]);
888 if (total_cycles
== 0) {
893 "\tidle cycles:%"PRIu64
" (%.02f%%)\n"
894 "\tprocessing cycles:%"PRIu64
" (%.02f%%)\n",
895 cycles
[PMD_CYCLES_IDLE
],
896 cycles
[PMD_CYCLES_IDLE
] / (double)total_cycles
* 100,
897 cycles
[PMD_CYCLES_PROCESSING
],
898 cycles
[PMD_CYCLES_PROCESSING
] / (double)total_cycles
* 100);
900 if (total_packets
== 0) {
905 "\tavg cycles per packet: %.02f (%"PRIu64
"/%llu)\n",
906 total_cycles
/ (double)total_packets
,
907 total_cycles
, total_packets
);
910 "\tavg processing cycles per packet: "
911 "%.02f (%"PRIu64
"/%llu)\n",
912 cycles
[PMD_CYCLES_PROCESSING
] / (double)total_packets
,
913 cycles
[PMD_CYCLES_PROCESSING
], total_packets
);
917 pmd_info_clear_stats(struct ds
*reply OVS_UNUSED
,
918 struct dp_netdev_pmd_thread
*pmd
,
919 unsigned long long stats
[DP_N_STATS
],
920 uint64_t cycles
[PMD_N_CYCLES
])
924 /* We cannot write 'stats' and 'cycles' (because they're written by other
925 * threads) and we shouldn't change 'stats' (because they're used to count
926 * datapath stats, which must not be cleared here). Instead, we save the
927 * current values and subtract them from the values to be displayed in the
929 for (i
= 0; i
< DP_N_STATS
; i
++) {
930 pmd
->stats_zero
[i
] = stats
[i
];
932 for (i
= 0; i
< PMD_N_CYCLES
; i
++) {
933 pmd
->cycles_zero
[i
] = cycles
[i
];
938 compare_poll_list(const void *a_
, const void *b_
)
940 const struct rxq_poll
*a
= a_
;
941 const struct rxq_poll
*b
= b_
;
943 const char *namea
= netdev_rxq_get_name(a
->rxq
->rx
);
944 const char *nameb
= netdev_rxq_get_name(b
->rxq
->rx
);
946 int cmp
= strcmp(namea
, nameb
);
948 return netdev_rxq_get_queue_id(a
->rxq
->rx
)
949 - netdev_rxq_get_queue_id(b
->rxq
->rx
);
956 sorted_poll_list(struct dp_netdev_pmd_thread
*pmd
, struct rxq_poll
**list
,
959 struct rxq_poll
*ret
, *poll
;
962 *n
= hmap_count(&pmd
->poll_list
);
966 ret
= xcalloc(*n
, sizeof *ret
);
968 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
973 qsort(ret
, *n
, sizeof *ret
, compare_poll_list
);
980 pmd_info_show_rxq(struct ds
*reply
, struct dp_netdev_pmd_thread
*pmd
)
982 if (pmd
->core_id
!= NON_PMD_CORE_ID
) {
983 const char *prev_name
= NULL
;
984 struct rxq_poll
*list
;
988 "pmd thread numa_id %d core_id %u:\n\tisolated : %s\n",
989 pmd
->numa_id
, pmd
->core_id
, (pmd
->isolated
)
992 ovs_mutex_lock(&pmd
->port_mutex
);
993 sorted_poll_list(pmd
, &list
, &n
);
994 for (i
= 0; i
< n
; i
++) {
995 const char *name
= netdev_rxq_get_name(list
[i
].rxq
->rx
);
997 if (!prev_name
|| strcmp(name
, prev_name
)) {
999 ds_put_cstr(reply
, "\n");
1001 ds_put_format(reply
, "\tport: %s\tqueue-id:", name
);
1003 ds_put_format(reply
, " %d",
1004 netdev_rxq_get_queue_id(list
[i
].rxq
->rx
));
1007 ovs_mutex_unlock(&pmd
->port_mutex
);
1008 ds_put_cstr(reply
, "\n");
1014 compare_poll_thread_list(const void *a_
, const void *b_
)
1016 const struct dp_netdev_pmd_thread
*a
, *b
;
1018 a
= *(struct dp_netdev_pmd_thread
**)a_
;
1019 b
= *(struct dp_netdev_pmd_thread
**)b_
;
1021 if (a
->core_id
< b
->core_id
) {
1024 if (a
->core_id
> b
->core_id
) {
1030 /* Create a sorted list of pmd's from the dp->poll_threads cmap. We can use
1031 * this list, as long as we do not go to quiescent state. */
1033 sorted_poll_thread_list(struct dp_netdev
*dp
,
1034 struct dp_netdev_pmd_thread
***list
,
1037 struct dp_netdev_pmd_thread
*pmd
;
1038 struct dp_netdev_pmd_thread
**pmd_list
;
1039 size_t k
= 0, n_pmds
;
1041 n_pmds
= cmap_count(&dp
->poll_threads
);
1042 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
1044 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1048 pmd_list
[k
++] = pmd
;
1051 qsort(pmd_list
, k
, sizeof *pmd_list
, compare_poll_thread_list
);
1058 dpif_netdev_pmd_rebalance(struct unixctl_conn
*conn
, int argc
,
1059 const char *argv
[], void *aux OVS_UNUSED
)
1061 struct ds reply
= DS_EMPTY_INITIALIZER
;
1062 struct dp_netdev
*dp
= NULL
;
1064 ovs_mutex_lock(&dp_netdev_mutex
);
1067 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
1068 } else if (shash_count(&dp_netdevs
) == 1) {
1069 /* There's only one datapath */
1070 dp
= shash_first(&dp_netdevs
)->data
;
1074 ovs_mutex_unlock(&dp_netdev_mutex
);
1075 unixctl_command_reply_error(conn
,
1076 "please specify an existing datapath");
1080 dp_netdev_request_reconfigure(dp
);
1081 ovs_mutex_unlock(&dp_netdev_mutex
);
1082 ds_put_cstr(&reply
, "pmd rxq rebalance requested.\n");
1083 unixctl_command_reply(conn
, ds_cstr(&reply
));
1088 dpif_netdev_pmd_info(struct unixctl_conn
*conn
, int argc
, const char *argv
[],
1091 struct ds reply
= DS_EMPTY_INITIALIZER
;
1092 struct dp_netdev_pmd_thread
**pmd_list
;
1093 struct dp_netdev
*dp
= NULL
;
1095 enum pmd_info_type type
= *(enum pmd_info_type
*) aux
;
1097 ovs_mutex_lock(&dp_netdev_mutex
);
1100 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
1101 } else if (shash_count(&dp_netdevs
) == 1) {
1102 /* There's only one datapath */
1103 dp
= shash_first(&dp_netdevs
)->data
;
1107 ovs_mutex_unlock(&dp_netdev_mutex
);
1108 unixctl_command_reply_error(conn
,
1109 "please specify an existing datapath");
1113 sorted_poll_thread_list(dp
, &pmd_list
, &n
);
1114 for (size_t i
= 0; i
< n
; i
++) {
1115 struct dp_netdev_pmd_thread
*pmd
= pmd_list
[i
];
1120 if (type
== PMD_INFO_SHOW_RXQ
) {
1121 pmd_info_show_rxq(&reply
, pmd
);
1123 unsigned long long stats
[DP_N_STATS
];
1124 uint64_t cycles
[PMD_N_CYCLES
];
1126 /* Read current stats and cycle counters */
1127 for (size_t j
= 0; j
< ARRAY_SIZE(stats
); j
++) {
1128 atomic_read_relaxed(&pmd
->stats
.n
[j
], &stats
[j
]);
1130 for (size_t j
= 0; j
< ARRAY_SIZE(cycles
); j
++) {
1131 atomic_read_relaxed(&pmd
->cycles
.n
[j
], &cycles
[j
]);
1134 if (type
== PMD_INFO_CLEAR_STATS
) {
1135 pmd_info_clear_stats(&reply
, pmd
, stats
, cycles
);
1136 } else if (type
== PMD_INFO_SHOW_STATS
) {
1137 pmd_info_show_stats(&reply
, pmd
, stats
, cycles
);
1143 ovs_mutex_unlock(&dp_netdev_mutex
);
1145 unixctl_command_reply(conn
, ds_cstr(&reply
));
1150 dpif_netdev_init(void)
1152 static enum pmd_info_type show_aux
= PMD_INFO_SHOW_STATS
,
1153 clear_aux
= PMD_INFO_CLEAR_STATS
,
1154 poll_aux
= PMD_INFO_SHOW_RXQ
;
1156 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
1157 0, 1, dpif_netdev_pmd_info
,
1159 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
1160 0, 1, dpif_netdev_pmd_info
,
1161 (void *)&clear_aux
);
1162 unixctl_command_register("dpif-netdev/pmd-rxq-show", "[dp]",
1163 0, 1, dpif_netdev_pmd_info
,
1165 unixctl_command_register("dpif-netdev/pmd-rxq-rebalance", "[dp]",
1166 0, 1, dpif_netdev_pmd_rebalance
,
1172 dpif_netdev_enumerate(struct sset
*all_dps
,
1173 const struct dpif_class
*dpif_class
)
1175 struct shash_node
*node
;
1177 ovs_mutex_lock(&dp_netdev_mutex
);
1178 SHASH_FOR_EACH(node
, &dp_netdevs
) {
1179 struct dp_netdev
*dp
= node
->data
;
1180 if (dpif_class
!= dp
->class) {
1181 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
1182 * If the class doesn't match, skip this dpif. */
1185 sset_add(all_dps
, node
->name
);
1187 ovs_mutex_unlock(&dp_netdev_mutex
);
1193 dpif_netdev_class_is_dummy(const struct dpif_class
*class)
1195 return class != &dpif_netdev_class
;
1199 dpif_netdev_port_open_type(const struct dpif_class
*class, const char *type
)
1201 return strcmp(type
, "internal") ? type
1202 : dpif_netdev_class_is_dummy(class) ? "dummy-internal"
1206 static struct dpif
*
1207 create_dpif_netdev(struct dp_netdev
*dp
)
1209 uint16_t netflow_id
= hash_string(dp
->name
, 0);
1210 struct dpif_netdev
*dpif
;
1212 ovs_refcount_ref(&dp
->ref_cnt
);
1214 dpif
= xmalloc(sizeof *dpif
);
1215 dpif_init(&dpif
->dpif
, dp
->class, dp
->name
, netflow_id
>> 8, netflow_id
);
1217 dpif
->last_port_seq
= seq_read(dp
->port_seq
);
1222 /* Choose an unused, non-zero port number and return it on success.
1223 * Return ODPP_NONE on failure. */
1225 choose_port(struct dp_netdev
*dp
, const char *name
)
1226 OVS_REQUIRES(dp
->port_mutex
)
1230 if (dp
->class != &dpif_netdev_class
) {
1234 /* If the port name begins with "br", start the number search at
1235 * 100 to make writing tests easier. */
1236 if (!strncmp(name
, "br", 2)) {
1240 /* If the port name contains a number, try to assign that port number.
1241 * This can make writing unit tests easier because port numbers are
1243 for (p
= name
; *p
!= '\0'; p
++) {
1244 if (isdigit((unsigned char) *p
)) {
1245 port_no
= start_no
+ strtol(p
, NULL
, 10);
1246 if (port_no
> 0 && port_no
!= odp_to_u32(ODPP_NONE
)
1247 && !dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
1248 return u32_to_odp(port_no
);
1255 for (port_no
= 1; port_no
<= UINT16_MAX
; port_no
++) {
1256 if (!dp_netdev_lookup_port(dp
, u32_to_odp(port_no
))) {
1257 return u32_to_odp(port_no
);
1265 create_dp_netdev(const char *name
, const struct dpif_class
*class,
1266 struct dp_netdev
**dpp
)
1267 OVS_REQUIRES(dp_netdev_mutex
)
1269 struct dp_netdev
*dp
;
1272 dp
= xzalloc(sizeof *dp
);
1273 shash_add(&dp_netdevs
, name
, dp
);
1275 *CONST_CAST(const struct dpif_class
**, &dp
->class) = class;
1276 *CONST_CAST(const char **, &dp
->name
) = xstrdup(name
);
1277 ovs_refcount_init(&dp
->ref_cnt
);
1278 atomic_flag_clear(&dp
->destroyed
);
1280 ovs_mutex_init(&dp
->port_mutex
);
1281 hmap_init(&dp
->ports
);
1282 dp
->port_seq
= seq_create();
1283 fat_rwlock_init(&dp
->upcall_rwlock
);
1285 dp
->reconfigure_seq
= seq_create();
1286 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
1288 for (int i
= 0; i
< N_METER_LOCKS
; ++i
) {
1289 ovs_mutex_init_adaptive(&dp
->meter_locks
[i
]);
1292 /* Disable upcalls by default. */
1293 dp_netdev_disable_upcall(dp
);
1294 dp
->upcall_aux
= NULL
;
1295 dp
->upcall_cb
= NULL
;
1297 conntrack_init(&dp
->conntrack
);
1299 atomic_init(&dp
->emc_insert_min
, DEFAULT_EM_FLOW_INSERT_MIN
);
1301 cmap_init(&dp
->poll_threads
);
1303 ovs_mutex_init(&dp
->tx_qid_pool_mutex
);
1304 /* We need 1 Tx queue for each possible core + 1 for non-PMD threads. */
1305 dp
->tx_qid_pool
= id_pool_create(0, ovs_numa_get_n_cores() + 1);
1307 ovs_mutex_init_recursive(&dp
->non_pmd_mutex
);
1308 ovsthread_key_create(&dp
->per_pmd_key
, NULL
);
1310 ovs_mutex_lock(&dp
->port_mutex
);
1311 /* non-PMD will be created before all other threads and will
1312 * allocate static_tx_qid = 0. */
1313 dp_netdev_set_nonpmd(dp
);
1315 error
= do_add_port(dp
, name
, dpif_netdev_port_open_type(dp
->class,
1318 ovs_mutex_unlock(&dp
->port_mutex
);
1324 dp
->last_tnl_conf_seq
= seq_read(tnl_conf_seq
);
1330 dp_netdev_request_reconfigure(struct dp_netdev
*dp
)
1332 seq_change(dp
->reconfigure_seq
);
1336 dp_netdev_is_reconf_required(struct dp_netdev
*dp
)
1338 return seq_read(dp
->reconfigure_seq
) != dp
->last_reconfigure_seq
;
1342 dpif_netdev_open(const struct dpif_class
*class, const char *name
,
1343 bool create
, struct dpif
**dpifp
)
1345 struct dp_netdev
*dp
;
1348 ovs_mutex_lock(&dp_netdev_mutex
);
1349 dp
= shash_find_data(&dp_netdevs
, name
);
1351 error
= create
? create_dp_netdev(name
, class, &dp
) : ENODEV
;
1353 error
= (dp
->class != class ? EINVAL
1358 *dpifp
= create_dpif_netdev(dp
);
1361 ovs_mutex_unlock(&dp_netdev_mutex
);
1367 dp_netdev_destroy_upcall_lock(struct dp_netdev
*dp
)
1368 OVS_NO_THREAD_SAFETY_ANALYSIS
1370 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
1371 ovs_assert(fat_rwlock_tryrdlock(&dp
->upcall_rwlock
));
1373 /* Before freeing a lock we should release it */
1374 fat_rwlock_unlock(&dp
->upcall_rwlock
);
1375 fat_rwlock_destroy(&dp
->upcall_rwlock
);
1379 dp_delete_meter(struct dp_netdev
*dp
, uint32_t meter_id
)
1380 OVS_REQUIRES(dp
->meter_locks
[meter_id
% N_METER_LOCKS
])
1382 if (dp
->meters
[meter_id
]) {
1383 free(dp
->meters
[meter_id
]);
1384 dp
->meters
[meter_id
] = NULL
;
1388 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
1389 * through the 'dp_netdevs' shash while freeing 'dp'. */
1391 dp_netdev_free(struct dp_netdev
*dp
)
1392 OVS_REQUIRES(dp_netdev_mutex
)
1394 struct dp_netdev_port
*port
, *next
;
1396 shash_find_and_delete(&dp_netdevs
, dp
->name
);
1398 ovs_mutex_lock(&dp
->port_mutex
);
1399 HMAP_FOR_EACH_SAFE (port
, next
, node
, &dp
->ports
) {
1400 do_del_port(dp
, port
);
1402 ovs_mutex_unlock(&dp
->port_mutex
);
1404 dp_netdev_destroy_all_pmds(dp
, true);
1405 cmap_destroy(&dp
->poll_threads
);
1407 ovs_mutex_destroy(&dp
->tx_qid_pool_mutex
);
1408 id_pool_destroy(dp
->tx_qid_pool
);
1410 ovs_mutex_destroy(&dp
->non_pmd_mutex
);
1411 ovsthread_key_delete(dp
->per_pmd_key
);
1413 conntrack_destroy(&dp
->conntrack
);
1416 seq_destroy(dp
->reconfigure_seq
);
1418 seq_destroy(dp
->port_seq
);
1419 hmap_destroy(&dp
->ports
);
1420 ovs_mutex_destroy(&dp
->port_mutex
);
1422 /* Upcalls must be disabled at this point */
1423 dp_netdev_destroy_upcall_lock(dp
);
1427 for (i
= 0; i
< MAX_METERS
; ++i
) {
1429 dp_delete_meter(dp
, i
);
1430 meter_unlock(dp
, i
);
1432 for (i
= 0; i
< N_METER_LOCKS
; ++i
) {
1433 ovs_mutex_destroy(&dp
->meter_locks
[i
]);
1436 free(dp
->pmd_cmask
);
1437 free(CONST_CAST(char *, dp
->name
));
1442 dp_netdev_unref(struct dp_netdev
*dp
)
1445 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
1446 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
1447 ovs_mutex_lock(&dp_netdev_mutex
);
1448 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1451 ovs_mutex_unlock(&dp_netdev_mutex
);
1456 dpif_netdev_close(struct dpif
*dpif
)
1458 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1460 dp_netdev_unref(dp
);
1465 dpif_netdev_destroy(struct dpif
*dpif
)
1467 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1469 if (!atomic_flag_test_and_set(&dp
->destroyed
)) {
1470 if (ovs_refcount_unref_relaxed(&dp
->ref_cnt
) == 1) {
1471 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
1479 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
1480 * load/store semantics. While the increment is not atomic, the load and
1481 * store operations are, making it impossible to read inconsistent values.
1483 * This is used to update thread local stats counters. */
1485 non_atomic_ullong_add(atomic_ullong
*var
, unsigned long long n
)
1487 unsigned long long tmp
;
1489 atomic_read_relaxed(var
, &tmp
);
1491 atomic_store_relaxed(var
, tmp
);
1495 dpif_netdev_get_stats(const struct dpif
*dpif
, struct dpif_dp_stats
*stats
)
1497 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1498 struct dp_netdev_pmd_thread
*pmd
;
1500 stats
->n_flows
= stats
->n_hit
= stats
->n_missed
= stats
->n_lost
= 0;
1501 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1502 unsigned long long n
;
1503 stats
->n_flows
+= cmap_count(&pmd
->flow_table
);
1505 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_MASKED_HIT
], &n
);
1507 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_EXACT_HIT
], &n
);
1509 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_MISS
], &n
);
1510 stats
->n_missed
+= n
;
1511 atomic_read_relaxed(&pmd
->stats
.n
[DP_STAT_LOST
], &n
);
1514 stats
->n_masks
= UINT32_MAX
;
1515 stats
->n_mask_hit
= UINT64_MAX
;
1521 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread
*pmd
)
1523 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
1524 ovs_mutex_lock(&pmd
->dp
->non_pmd_mutex
);
1525 ovs_mutex_lock(&pmd
->port_mutex
);
1526 pmd_load_cached_ports(pmd
);
1527 ovs_mutex_unlock(&pmd
->port_mutex
);
1528 ovs_mutex_unlock(&pmd
->dp
->non_pmd_mutex
);
1532 ovs_mutex_lock(&pmd
->cond_mutex
);
1533 seq_change(pmd
->reload_seq
);
1534 atomic_store_relaxed(&pmd
->reload
, true);
1535 ovs_mutex_cond_wait(&pmd
->cond
, &pmd
->cond_mutex
);
1536 ovs_mutex_unlock(&pmd
->cond_mutex
);
1540 hash_port_no(odp_port_t port_no
)
1542 return hash_int(odp_to_u32(port_no
), 0);
1546 port_create(const char *devname
, const char *type
,
1547 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1549 struct netdev_saved_flags
*sf
;
1550 struct dp_netdev_port
*port
;
1551 enum netdev_flags flags
;
1552 struct netdev
*netdev
;
1557 /* Open and validate network device. */
1558 error
= netdev_open(devname
, type
, &netdev
);
1562 /* XXX reject non-Ethernet devices */
1564 netdev_get_flags(netdev
, &flags
);
1565 if (flags
& NETDEV_LOOPBACK
) {
1566 VLOG_ERR("%s: cannot add a loopback device", devname
);
1571 error
= netdev_turn_flags_on(netdev
, NETDEV_PROMISC
, &sf
);
1573 VLOG_ERR("%s: cannot set promisc flag", devname
);
1577 port
= xzalloc(sizeof *port
);
1578 port
->port_no
= port_no
;
1579 port
->netdev
= netdev
;
1580 port
->type
= xstrdup(type
);
1582 port
->need_reconfigure
= true;
1583 ovs_mutex_init(&port
->txq_used_mutex
);
1590 netdev_close(netdev
);
1595 do_add_port(struct dp_netdev
*dp
, const char *devname
, const char *type
,
1597 OVS_REQUIRES(dp
->port_mutex
)
1599 struct dp_netdev_port
*port
;
1602 /* Reject devices already in 'dp'. */
1603 if (!get_port_by_name(dp
, devname
, &port
)) {
1607 error
= port_create(devname
, type
, port_no
, &port
);
1612 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
1613 seq_change(dp
->port_seq
);
1615 reconfigure_datapath(dp
);
1621 dpif_netdev_port_add(struct dpif
*dpif
, struct netdev
*netdev
,
1622 odp_port_t
*port_nop
)
1624 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1625 char namebuf
[NETDEV_VPORT_NAME_BUFSIZE
];
1626 const char *dpif_port
;
1630 ovs_mutex_lock(&dp
->port_mutex
);
1631 dpif_port
= netdev_vport_get_dpif_port(netdev
, namebuf
, sizeof namebuf
);
1632 if (*port_nop
!= ODPP_NONE
) {
1633 port_no
= *port_nop
;
1634 error
= dp_netdev_lookup_port(dp
, *port_nop
) ? EBUSY
: 0;
1636 port_no
= choose_port(dp
, dpif_port
);
1637 error
= port_no
== ODPP_NONE
? EFBIG
: 0;
1640 *port_nop
= port_no
;
1641 error
= do_add_port(dp
, dpif_port
, netdev_get_type(netdev
), port_no
);
1643 ovs_mutex_unlock(&dp
->port_mutex
);
1649 dpif_netdev_port_del(struct dpif
*dpif
, odp_port_t port_no
)
1651 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1654 ovs_mutex_lock(&dp
->port_mutex
);
1655 if (port_no
== ODPP_LOCAL
) {
1658 struct dp_netdev_port
*port
;
1660 error
= get_port_by_number(dp
, port_no
, &port
);
1662 do_del_port(dp
, port
);
1665 ovs_mutex_unlock(&dp
->port_mutex
);
1671 is_valid_port_number(odp_port_t port_no
)
1673 return port_no
!= ODPP_NONE
;
1676 static struct dp_netdev_port
*
1677 dp_netdev_lookup_port(const struct dp_netdev
*dp
, odp_port_t port_no
)
1678 OVS_REQUIRES(dp
->port_mutex
)
1680 struct dp_netdev_port
*port
;
1682 HMAP_FOR_EACH_WITH_HASH (port
, node
, hash_port_no(port_no
), &dp
->ports
) {
1683 if (port
->port_no
== port_no
) {
1691 get_port_by_number(struct dp_netdev
*dp
,
1692 odp_port_t port_no
, struct dp_netdev_port
**portp
)
1693 OVS_REQUIRES(dp
->port_mutex
)
1695 if (!is_valid_port_number(port_no
)) {
1699 *portp
= dp_netdev_lookup_port(dp
, port_no
);
1700 return *portp
? 0 : ENODEV
;
1705 port_destroy(struct dp_netdev_port
*port
)
1711 netdev_close(port
->netdev
);
1712 netdev_restore_flags(port
->sf
);
1714 for (unsigned i
= 0; i
< port
->n_rxq
; i
++) {
1715 netdev_rxq_close(port
->rxqs
[i
].rx
);
1717 ovs_mutex_destroy(&port
->txq_used_mutex
);
1718 free(port
->rxq_affinity_list
);
1719 free(port
->txq_used
);
1726 get_port_by_name(struct dp_netdev
*dp
,
1727 const char *devname
, struct dp_netdev_port
**portp
)
1728 OVS_REQUIRES(dp
->port_mutex
)
1730 struct dp_netdev_port
*port
;
1732 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1733 if (!strcmp(netdev_get_name(port
->netdev
), devname
)) {
1739 /* Callers of dpif_netdev_port_query_by_name() expect ENODEV for a non
1744 /* Returns 'true' if there is a port with pmd netdev. */
1746 has_pmd_port(struct dp_netdev
*dp
)
1747 OVS_REQUIRES(dp
->port_mutex
)
1749 struct dp_netdev_port
*port
;
1751 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
1752 if (netdev_is_pmd(port
->netdev
)) {
1761 do_del_port(struct dp_netdev
*dp
, struct dp_netdev_port
*port
)
1762 OVS_REQUIRES(dp
->port_mutex
)
1764 hmap_remove(&dp
->ports
, &port
->node
);
1765 seq_change(dp
->port_seq
);
1767 reconfigure_datapath(dp
);
1773 answer_port_query(const struct dp_netdev_port
*port
,
1774 struct dpif_port
*dpif_port
)
1776 dpif_port
->name
= xstrdup(netdev_get_name(port
->netdev
));
1777 dpif_port
->type
= xstrdup(port
->type
);
1778 dpif_port
->port_no
= port
->port_no
;
1782 dpif_netdev_port_query_by_number(const struct dpif
*dpif
, odp_port_t port_no
,
1783 struct dpif_port
*dpif_port
)
1785 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1786 struct dp_netdev_port
*port
;
1789 ovs_mutex_lock(&dp
->port_mutex
);
1790 error
= get_port_by_number(dp
, port_no
, &port
);
1791 if (!error
&& dpif_port
) {
1792 answer_port_query(port
, dpif_port
);
1794 ovs_mutex_unlock(&dp
->port_mutex
);
1800 dpif_netdev_port_query_by_name(const struct dpif
*dpif
, const char *devname
,
1801 struct dpif_port
*dpif_port
)
1803 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1804 struct dp_netdev_port
*port
;
1807 ovs_mutex_lock(&dp
->port_mutex
);
1808 error
= get_port_by_name(dp
, devname
, &port
);
1809 if (!error
&& dpif_port
) {
1810 answer_port_query(port
, dpif_port
);
1812 ovs_mutex_unlock(&dp
->port_mutex
);
1818 dp_netdev_flow_free(struct dp_netdev_flow
*flow
)
1820 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow
));
1824 static void dp_netdev_flow_unref(struct dp_netdev_flow
*flow
)
1826 if (ovs_refcount_unref_relaxed(&flow
->ref_cnt
) == 1) {
1827 ovsrcu_postpone(dp_netdev_flow_free
, flow
);
1832 dp_netdev_flow_hash(const ovs_u128
*ufid
)
1834 return ufid
->u32
[0];
1837 static inline struct dpcls
*
1838 dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread
*pmd
,
1842 uint32_t hash
= hash_port_no(in_port
);
1843 CMAP_FOR_EACH_WITH_HASH (cls
, node
, hash
, &pmd
->classifiers
) {
1844 if (cls
->in_port
== in_port
) {
1845 /* Port classifier exists already */
1852 static inline struct dpcls
*
1853 dp_netdev_pmd_find_dpcls(struct dp_netdev_pmd_thread
*pmd
,
1855 OVS_REQUIRES(pmd
->flow_mutex
)
1857 struct dpcls
*cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
1858 uint32_t hash
= hash_port_no(in_port
);
1861 /* Create new classifier for in_port */
1862 cls
= xmalloc(sizeof(*cls
));
1864 cls
->in_port
= in_port
;
1865 cmap_insert(&pmd
->classifiers
, &cls
->node
, hash
);
1866 VLOG_DBG("Creating dpcls %p for in_port %d", cls
, in_port
);
1872 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread
*pmd
,
1873 struct dp_netdev_flow
*flow
)
1874 OVS_REQUIRES(pmd
->flow_mutex
)
1876 struct cmap_node
*node
= CONST_CAST(struct cmap_node
*, &flow
->node
);
1878 odp_port_t in_port
= flow
->flow
.in_port
.odp_port
;
1880 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
1881 ovs_assert(cls
!= NULL
);
1882 dpcls_remove(cls
, &flow
->cr
);
1883 cmap_remove(&pmd
->flow_table
, node
, dp_netdev_flow_hash(&flow
->ufid
));
1886 dp_netdev_flow_unref(flow
);
1890 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread
*pmd
)
1892 struct dp_netdev_flow
*netdev_flow
;
1894 ovs_mutex_lock(&pmd
->flow_mutex
);
1895 CMAP_FOR_EACH (netdev_flow
, node
, &pmd
->flow_table
) {
1896 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
1898 ovs_mutex_unlock(&pmd
->flow_mutex
);
1902 dpif_netdev_flow_flush(struct dpif
*dpif
)
1904 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1905 struct dp_netdev_pmd_thread
*pmd
;
1907 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
1908 dp_netdev_pmd_flow_flush(pmd
);
1914 struct dp_netdev_port_state
{
1915 struct hmap_position position
;
1920 dpif_netdev_port_dump_start(const struct dpif
*dpif OVS_UNUSED
, void **statep
)
1922 *statep
= xzalloc(sizeof(struct dp_netdev_port_state
));
1927 dpif_netdev_port_dump_next(const struct dpif
*dpif
, void *state_
,
1928 struct dpif_port
*dpif_port
)
1930 struct dp_netdev_port_state
*state
= state_
;
1931 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
1932 struct hmap_node
*node
;
1935 ovs_mutex_lock(&dp
->port_mutex
);
1936 node
= hmap_at_position(&dp
->ports
, &state
->position
);
1938 struct dp_netdev_port
*port
;
1940 port
= CONTAINER_OF(node
, struct dp_netdev_port
, node
);
1943 state
->name
= xstrdup(netdev_get_name(port
->netdev
));
1944 dpif_port
->name
= state
->name
;
1945 dpif_port
->type
= port
->type
;
1946 dpif_port
->port_no
= port
->port_no
;
1952 ovs_mutex_unlock(&dp
->port_mutex
);
1958 dpif_netdev_port_dump_done(const struct dpif
*dpif OVS_UNUSED
, void *state_
)
1960 struct dp_netdev_port_state
*state
= state_
;
1967 dpif_netdev_port_poll(const struct dpif
*dpif_
, char **devnamep OVS_UNUSED
)
1969 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
1970 uint64_t new_port_seq
;
1973 new_port_seq
= seq_read(dpif
->dp
->port_seq
);
1974 if (dpif
->last_port_seq
!= new_port_seq
) {
1975 dpif
->last_port_seq
= new_port_seq
;
1985 dpif_netdev_port_poll_wait(const struct dpif
*dpif_
)
1987 struct dpif_netdev
*dpif
= dpif_netdev_cast(dpif_
);
1989 seq_wait(dpif
->dp
->port_seq
, dpif
->last_port_seq
);
1992 static struct dp_netdev_flow
*
1993 dp_netdev_flow_cast(const struct dpcls_rule
*cr
)
1995 return cr
? CONTAINER_OF(cr
, struct dp_netdev_flow
, cr
) : NULL
;
1998 static bool dp_netdev_flow_ref(struct dp_netdev_flow
*flow
)
2000 return ovs_refcount_try_ref_rcu(&flow
->ref_cnt
);
2003 /* netdev_flow_key utilities.
2005 * netdev_flow_key is basically a miniflow. We use these functions
2006 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
2007 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
2009 * - Since we are dealing exclusively with miniflows created by
2010 * miniflow_extract(), if the map is different the miniflow is different.
2011 * Therefore we can be faster by comparing the map and the miniflow in a
2013 * - These functions can be inlined by the compiler. */
2015 /* Given the number of bits set in miniflow's maps, returns the size of the
2016 * 'netdev_flow_key.mf' */
2017 static inline size_t
2018 netdev_flow_key_size(size_t flow_u64s
)
2020 return sizeof(struct miniflow
) + MINIFLOW_VALUES_SIZE(flow_u64s
);
2024 netdev_flow_key_equal(const struct netdev_flow_key
*a
,
2025 const struct netdev_flow_key
*b
)
2027 /* 'b->len' may be not set yet. */
2028 return a
->hash
== b
->hash
&& !memcmp(&a
->mf
, &b
->mf
, a
->len
);
2031 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
2032 * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been
2033 * generated by miniflow_extract. */
2035 netdev_flow_key_equal_mf(const struct netdev_flow_key
*key
,
2036 const struct miniflow
*mf
)
2038 return !memcmp(&key
->mf
, mf
, key
->len
);
2042 netdev_flow_key_clone(struct netdev_flow_key
*dst
,
2043 const struct netdev_flow_key
*src
)
2046 offsetof(struct netdev_flow_key
, mf
) + src
->len
);
2049 /* Initialize a netdev_flow_key 'mask' from 'match'. */
2051 netdev_flow_mask_init(struct netdev_flow_key
*mask
,
2052 const struct match
*match
)
2054 uint64_t *dst
= miniflow_values(&mask
->mf
);
2055 struct flowmap fmap
;
2059 /* Only check masks that make sense for the flow. */
2060 flow_wc_map(&match
->flow
, &fmap
);
2061 flowmap_init(&mask
->mf
.map
);
2063 FLOWMAP_FOR_EACH_INDEX(idx
, fmap
) {
2064 uint64_t mask_u64
= flow_u64_value(&match
->wc
.masks
, idx
);
2067 flowmap_set(&mask
->mf
.map
, idx
, 1);
2069 hash
= hash_add64(hash
, mask_u64
);
2075 FLOWMAP_FOR_EACH_MAP (map
, mask
->mf
.map
) {
2076 hash
= hash_add64(hash
, map
);
2079 size_t n
= dst
- miniflow_get_values(&mask
->mf
);
2081 mask
->hash
= hash_finish(hash
, n
* 8);
2082 mask
->len
= netdev_flow_key_size(n
);
2085 /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */
2087 netdev_flow_key_init_masked(struct netdev_flow_key
*dst
,
2088 const struct flow
*flow
,
2089 const struct netdev_flow_key
*mask
)
2091 uint64_t *dst_u64
= miniflow_values(&dst
->mf
);
2092 const uint64_t *mask_u64
= miniflow_get_values(&mask
->mf
);
2096 dst
->len
= mask
->len
;
2097 dst
->mf
= mask
->mf
; /* Copy maps. */
2099 FLOW_FOR_EACH_IN_MAPS(value
, flow
, mask
->mf
.map
) {
2100 *dst_u64
= value
& *mask_u64
++;
2101 hash
= hash_add64(hash
, *dst_u64
++);
2103 dst
->hash
= hash_finish(hash
,
2104 (dst_u64
- miniflow_get_values(&dst
->mf
)) * 8);
2107 /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */
2108 #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \
2109 MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP)
2111 /* Returns a hash value for the bits of 'key' where there are 1-bits in
2113 static inline uint32_t
2114 netdev_flow_key_hash_in_mask(const struct netdev_flow_key
*key
,
2115 const struct netdev_flow_key
*mask
)
2117 const uint64_t *p
= miniflow_get_values(&mask
->mf
);
2121 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, key
, mask
->mf
.map
) {
2122 hash
= hash_add64(hash
, value
& *p
++);
2125 return hash_finish(hash
, (p
- miniflow_get_values(&mask
->mf
)) * 8);
2129 emc_entry_alive(struct emc_entry
*ce
)
2131 return ce
->flow
&& !ce
->flow
->dead
;
2135 emc_clear_entry(struct emc_entry
*ce
)
2138 dp_netdev_flow_unref(ce
->flow
);
2144 emc_change_entry(struct emc_entry
*ce
, struct dp_netdev_flow
*flow
,
2145 const struct netdev_flow_key
*key
)
2147 if (ce
->flow
!= flow
) {
2149 dp_netdev_flow_unref(ce
->flow
);
2152 if (dp_netdev_flow_ref(flow
)) {
2159 netdev_flow_key_clone(&ce
->key
, key
);
2164 emc_insert(struct emc_cache
*cache
, const struct netdev_flow_key
*key
,
2165 struct dp_netdev_flow
*flow
)
2167 struct emc_entry
*to_be_replaced
= NULL
;
2168 struct emc_entry
*current_entry
;
2170 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2171 if (netdev_flow_key_equal(¤t_entry
->key
, key
)) {
2172 /* We found the entry with the 'mf' miniflow */
2173 emc_change_entry(current_entry
, flow
, NULL
);
2177 /* Replacement policy: put the flow in an empty (not alive) entry, or
2178 * in the first entry where it can be */
2180 || (emc_entry_alive(to_be_replaced
)
2181 && !emc_entry_alive(current_entry
))
2182 || current_entry
->key
.hash
< to_be_replaced
->key
.hash
) {
2183 to_be_replaced
= current_entry
;
2186 /* We didn't find the miniflow in the cache.
2187 * The 'to_be_replaced' entry is where the new flow will be stored */
2189 emc_change_entry(to_be_replaced
, flow
, key
);
2193 emc_probabilistic_insert(struct dp_netdev_pmd_thread
*pmd
,
2194 const struct netdev_flow_key
*key
,
2195 struct dp_netdev_flow
*flow
)
2197 /* Insert an entry into the EMC based on probability value 'min'. By
2198 * default the value is UINT32_MAX / 100 which yields an insertion
2199 * probability of 1/100 ie. 1% */
2202 atomic_read_relaxed(&pmd
->dp
->emc_insert_min
, &min
);
2204 if (min
&& random_uint32() <= min
) {
2205 emc_insert(&pmd
->flow_cache
, key
, flow
);
2209 static inline struct dp_netdev_flow
*
2210 emc_lookup(struct emc_cache
*cache
, const struct netdev_flow_key
*key
)
2212 struct emc_entry
*current_entry
;
2214 EMC_FOR_EACH_POS_WITH_HASH(cache
, current_entry
, key
->hash
) {
2215 if (current_entry
->key
.hash
== key
->hash
2216 && emc_entry_alive(current_entry
)
2217 && netdev_flow_key_equal_mf(¤t_entry
->key
, &key
->mf
)) {
2219 /* We found the entry with the 'key->mf' miniflow */
2220 return current_entry
->flow
;
2227 static struct dp_netdev_flow
*
2228 dp_netdev_pmd_lookup_flow(struct dp_netdev_pmd_thread
*pmd
,
2229 const struct netdev_flow_key
*key
,
2233 struct dpcls_rule
*rule
;
2234 odp_port_t in_port
= u32_to_odp(MINIFLOW_GET_U32(&key
->mf
, in_port
));
2235 struct dp_netdev_flow
*netdev_flow
= NULL
;
2237 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
2238 if (OVS_LIKELY(cls
)) {
2239 dpcls_lookup(cls
, key
, &rule
, 1, lookup_num_p
);
2240 netdev_flow
= dp_netdev_flow_cast(rule
);
2245 static struct dp_netdev_flow
*
2246 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread
*pmd
,
2247 const ovs_u128
*ufidp
, const struct nlattr
*key
,
2250 struct dp_netdev_flow
*netdev_flow
;
2254 /* If a UFID is not provided, determine one based on the key. */
2255 if (!ufidp
&& key
&& key_len
2256 && !dpif_netdev_flow_from_nlattrs(key
, key_len
, &flow
, false)) {
2257 dpif_flow_hash(pmd
->dp
->dpif
, &flow
, sizeof flow
, &ufid
);
2262 CMAP_FOR_EACH_WITH_HASH (netdev_flow
, node
, dp_netdev_flow_hash(ufidp
),
2264 if (ovs_u128_equals(netdev_flow
->ufid
, *ufidp
)) {
2274 get_dpif_flow_stats(const struct dp_netdev_flow
*netdev_flow_
,
2275 struct dpif_flow_stats
*stats
)
2277 struct dp_netdev_flow
*netdev_flow
;
2278 unsigned long long n
;
2282 netdev_flow
= CONST_CAST(struct dp_netdev_flow
*, netdev_flow_
);
2284 atomic_read_relaxed(&netdev_flow
->stats
.packet_count
, &n
);
2285 stats
->n_packets
= n
;
2286 atomic_read_relaxed(&netdev_flow
->stats
.byte_count
, &n
);
2288 atomic_read_relaxed(&netdev_flow
->stats
.used
, &used
);
2290 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
2291 stats
->tcp_flags
= flags
;
2294 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
2295 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
2296 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
2299 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow
*netdev_flow
,
2300 struct ofpbuf
*key_buf
, struct ofpbuf
*mask_buf
,
2301 struct dpif_flow
*flow
, bool terse
)
2304 memset(flow
, 0, sizeof *flow
);
2306 struct flow_wildcards wc
;
2307 struct dp_netdev_actions
*actions
;
2309 struct odp_flow_key_parms odp_parms
= {
2310 .flow
= &netdev_flow
->flow
,
2312 .support
= dp_netdev_support
,
2315 miniflow_expand(&netdev_flow
->cr
.mask
->mf
, &wc
.masks
);
2316 /* in_port is exact matched, but we have left it out from the mask for
2317 * optimnization reasons. Add in_port back to the mask. */
2318 wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
2321 offset
= key_buf
->size
;
2322 flow
->key
= ofpbuf_tail(key_buf
);
2323 odp_flow_key_from_flow(&odp_parms
, key_buf
);
2324 flow
->key_len
= key_buf
->size
- offset
;
2327 offset
= mask_buf
->size
;
2328 flow
->mask
= ofpbuf_tail(mask_buf
);
2329 odp_parms
.key_buf
= key_buf
;
2330 odp_flow_key_from_mask(&odp_parms
, mask_buf
);
2331 flow
->mask_len
= mask_buf
->size
- offset
;
2334 actions
= dp_netdev_flow_get_actions(netdev_flow
);
2335 flow
->actions
= actions
->actions
;
2336 flow
->actions_len
= actions
->size
;
2339 flow
->ufid
= netdev_flow
->ufid
;
2340 flow
->ufid_present
= true;
2341 flow
->pmd_id
= netdev_flow
->pmd_id
;
2342 get_dpif_flow_stats(netdev_flow
, &flow
->stats
);
2346 dpif_netdev_mask_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
2347 const struct nlattr
*mask_key
,
2348 uint32_t mask_key_len
, const struct flow
*flow
,
2349 struct flow_wildcards
*wc
, bool probe
)
2351 enum odp_key_fitness fitness
;
2353 fitness
= odp_flow_key_to_mask(mask_key
, mask_key_len
, wc
, flow
);
2356 /* This should not happen: it indicates that
2357 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
2358 * disagree on the acceptable form of a mask. Log the problem
2359 * as an error, with enough details to enable debugging. */
2360 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
2362 if (!VLOG_DROP_ERR(&rl
)) {
2366 odp_flow_format(key
, key_len
, mask_key
, mask_key_len
, NULL
, &s
,
2368 VLOG_ERR("internal error parsing flow mask %s (%s)",
2369 ds_cstr(&s
), odp_key_fitness_to_string(fitness
));
2381 dpif_netdev_flow_from_nlattrs(const struct nlattr
*key
, uint32_t key_len
,
2382 struct flow
*flow
, bool probe
)
2384 if (odp_flow_key_to_flow(key
, key_len
, flow
)) {
2386 /* This should not happen: it indicates that
2387 * odp_flow_key_from_flow() and odp_flow_key_to_flow() disagree on
2388 * the acceptable form of a flow. Log the problem as an error,
2389 * with enough details to enable debugging. */
2390 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
2392 if (!VLOG_DROP_ERR(&rl
)) {
2396 odp_flow_format(key
, key_len
, NULL
, 0, NULL
, &s
, true);
2397 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s
));
2405 if (flow
->ct_state
& DP_NETDEV_CS_UNSUPPORTED_MASK
) {
2413 dpif_netdev_flow_get(const struct dpif
*dpif
, const struct dpif_flow_get
*get
)
2415 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2416 struct dp_netdev_flow
*netdev_flow
;
2417 struct dp_netdev_pmd_thread
*pmd
;
2418 struct hmapx to_find
= HMAPX_INITIALIZER(&to_find
);
2419 struct hmapx_node
*node
;
2422 if (get
->pmd_id
== PMD_ID_NULL
) {
2423 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2424 if (dp_netdev_pmd_try_ref(pmd
) && !hmapx_add(&to_find
, pmd
)) {
2425 dp_netdev_pmd_unref(pmd
);
2429 pmd
= dp_netdev_get_pmd(dp
, get
->pmd_id
);
2433 hmapx_add(&to_find
, pmd
);
2436 if (!hmapx_count(&to_find
)) {
2440 HMAPX_FOR_EACH (node
, &to_find
) {
2441 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
2442 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, get
->ufid
, get
->key
,
2445 dp_netdev_flow_to_dpif_flow(netdev_flow
, get
->buffer
, get
->buffer
,
2454 HMAPX_FOR_EACH (node
, &to_find
) {
2455 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
2456 dp_netdev_pmd_unref(pmd
);
2459 hmapx_destroy(&to_find
);
2463 static struct dp_netdev_flow
*
2464 dp_netdev_flow_add(struct dp_netdev_pmd_thread
*pmd
,
2465 struct match
*match
, const ovs_u128
*ufid
,
2466 const struct nlattr
*actions
, size_t actions_len
)
2467 OVS_REQUIRES(pmd
->flow_mutex
)
2469 struct dp_netdev_flow
*flow
;
2470 struct netdev_flow_key mask
;
2473 /* Make sure in_port is exact matched before we read it. */
2474 ovs_assert(match
->wc
.masks
.in_port
.odp_port
== ODPP_NONE
);
2475 odp_port_t in_port
= match
->flow
.in_port
.odp_port
;
2477 /* As we select the dpcls based on the port number, each netdev flow
2478 * belonging to the same dpcls will have the same odp_port value.
2479 * For performance reasons we wildcard odp_port here in the mask. In the
2480 * typical case dp_hash is also wildcarded, and the resulting 8-byte
2481 * chunk {dp_hash, in_port} will be ignored by netdev_flow_mask_init() and
2482 * will not be part of the subtable mask.
2483 * This will speed up the hash computation during dpcls_lookup() because
2484 * there is one less call to hash_add64() in this case. */
2485 match
->wc
.masks
.in_port
.odp_port
= 0;
2486 netdev_flow_mask_init(&mask
, match
);
2487 match
->wc
.masks
.in_port
.odp_port
= ODPP_NONE
;
2489 /* Make sure wc does not have metadata. */
2490 ovs_assert(!FLOWMAP_HAS_FIELD(&mask
.mf
.map
, metadata
)
2491 && !FLOWMAP_HAS_FIELD(&mask
.mf
.map
, regs
));
2493 /* Do not allocate extra space. */
2494 flow
= xmalloc(sizeof *flow
- sizeof flow
->cr
.flow
.mf
+ mask
.len
);
2495 memset(&flow
->stats
, 0, sizeof flow
->stats
);
2498 *CONST_CAST(unsigned *, &flow
->pmd_id
) = pmd
->core_id
;
2499 *CONST_CAST(struct flow
*, &flow
->flow
) = match
->flow
;
2500 *CONST_CAST(ovs_u128
*, &flow
->ufid
) = *ufid
;
2501 ovs_refcount_init(&flow
->ref_cnt
);
2502 ovsrcu_set(&flow
->actions
, dp_netdev_actions_create(actions
, actions_len
));
2504 netdev_flow_key_init_masked(&flow
->cr
.flow
, &match
->flow
, &mask
);
2506 /* Select dpcls for in_port. Relies on in_port to be exact match. */
2507 cls
= dp_netdev_pmd_find_dpcls(pmd
, in_port
);
2508 dpcls_insert(cls
, &flow
->cr
, &mask
);
2510 cmap_insert(&pmd
->flow_table
, CONST_CAST(struct cmap_node
*, &flow
->node
),
2511 dp_netdev_flow_hash(&flow
->ufid
));
2513 if (OVS_UNLIKELY(!VLOG_DROP_DBG((&upcall_rl
)))) {
2514 struct ds ds
= DS_EMPTY_INITIALIZER
;
2515 struct ofpbuf key_buf
, mask_buf
;
2516 struct odp_flow_key_parms odp_parms
= {
2517 .flow
= &match
->flow
,
2518 .mask
= &match
->wc
.masks
,
2519 .support
= dp_netdev_support
,
2522 ofpbuf_init(&key_buf
, 0);
2523 ofpbuf_init(&mask_buf
, 0);
2525 odp_flow_key_from_flow(&odp_parms
, &key_buf
);
2526 odp_parms
.key_buf
= &key_buf
;
2527 odp_flow_key_from_mask(&odp_parms
, &mask_buf
);
2529 ds_put_cstr(&ds
, "flow_add: ");
2530 odp_format_ufid(ufid
, &ds
);
2531 ds_put_cstr(&ds
, " ");
2532 odp_flow_format(key_buf
.data
, key_buf
.size
,
2533 mask_buf
.data
, mask_buf
.size
,
2535 ds_put_cstr(&ds
, ", actions:");
2536 format_odp_actions(&ds
, actions
, actions_len
, NULL
);
2538 VLOG_DBG("%s", ds_cstr(&ds
));
2540 ofpbuf_uninit(&key_buf
);
2541 ofpbuf_uninit(&mask_buf
);
2543 /* Add a printout of the actual match installed. */
2546 ds_put_cstr(&ds
, "flow match: ");
2547 miniflow_expand(&flow
->cr
.flow
.mf
, &m
.flow
);
2548 miniflow_expand(&flow
->cr
.mask
->mf
, &m
.wc
.masks
);
2549 memset(&m
.tun_md
, 0, sizeof m
.tun_md
);
2550 match_format(&m
, NULL
, &ds
, OFP_DEFAULT_PRIORITY
);
2552 VLOG_DBG("%s", ds_cstr(&ds
));
2561 flow_put_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
2562 struct netdev_flow_key
*key
,
2563 struct match
*match
,
2565 const struct dpif_flow_put
*put
,
2566 struct dpif_flow_stats
*stats
)
2568 struct dp_netdev_flow
*netdev_flow
;
2572 memset(stats
, 0, sizeof *stats
);
2575 ovs_mutex_lock(&pmd
->flow_mutex
);
2576 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
2578 if (put
->flags
& DPIF_FP_CREATE
) {
2579 if (cmap_count(&pmd
->flow_table
) < MAX_FLOWS
) {
2580 dp_netdev_flow_add(pmd
, match
, ufid
, put
->actions
,
2590 if (put
->flags
& DPIF_FP_MODIFY
) {
2591 struct dp_netdev_actions
*new_actions
;
2592 struct dp_netdev_actions
*old_actions
;
2594 new_actions
= dp_netdev_actions_create(put
->actions
,
2597 old_actions
= dp_netdev_flow_get_actions(netdev_flow
);
2598 ovsrcu_set(&netdev_flow
->actions
, new_actions
);
2601 get_dpif_flow_stats(netdev_flow
, stats
);
2603 if (put
->flags
& DPIF_FP_ZERO_STATS
) {
2604 /* XXX: The userspace datapath uses thread local statistics
2605 * (for flows), which should be updated only by the owning
2606 * thread. Since we cannot write on stats memory here,
2607 * we choose not to support this flag. Please note:
2608 * - This feature is currently used only by dpctl commands with
2610 * - Should the need arise, this operation can be implemented
2611 * by keeping a base value (to be update here) for each
2612 * counter, and subtracting it before outputting the stats */
2616 ovsrcu_postpone(dp_netdev_actions_free
, old_actions
);
2617 } else if (put
->flags
& DPIF_FP_CREATE
) {
2620 /* Overlapping flow. */
2624 ovs_mutex_unlock(&pmd
->flow_mutex
);
2629 dpif_netdev_flow_put(struct dpif
*dpif
, const struct dpif_flow_put
*put
)
2631 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2632 struct netdev_flow_key key
, mask
;
2633 struct dp_netdev_pmd_thread
*pmd
;
2637 bool probe
= put
->flags
& DPIF_FP_PROBE
;
2640 memset(put
->stats
, 0, sizeof *put
->stats
);
2642 error
= dpif_netdev_flow_from_nlattrs(put
->key
, put
->key_len
, &match
.flow
,
2647 error
= dpif_netdev_mask_from_nlattrs(put
->key
, put
->key_len
,
2648 put
->mask
, put
->mask_len
,
2649 &match
.flow
, &match
.wc
, probe
);
2657 dpif_flow_hash(dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
2660 /* Must produce a netdev_flow_key for lookup.
2661 * Use the same method as employed to create the key when adding
2662 * the flow to the dplcs to make sure they match. */
2663 netdev_flow_mask_init(&mask
, &match
);
2664 netdev_flow_key_init_masked(&key
, &match
.flow
, &mask
);
2666 if (put
->pmd_id
== PMD_ID_NULL
) {
2667 if (cmap_count(&dp
->poll_threads
) == 0) {
2670 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2671 struct dpif_flow_stats pmd_stats
;
2674 pmd_error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
,
2678 } else if (put
->stats
) {
2679 put
->stats
->n_packets
+= pmd_stats
.n_packets
;
2680 put
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
2681 put
->stats
->used
= MAX(put
->stats
->used
, pmd_stats
.used
);
2682 put
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
2686 pmd
= dp_netdev_get_pmd(dp
, put
->pmd_id
);
2690 error
= flow_put_on_pmd(pmd
, &key
, &match
, &ufid
, put
, put
->stats
);
2691 dp_netdev_pmd_unref(pmd
);
2698 flow_del_on_pmd(struct dp_netdev_pmd_thread
*pmd
,
2699 struct dpif_flow_stats
*stats
,
2700 const struct dpif_flow_del
*del
)
2702 struct dp_netdev_flow
*netdev_flow
;
2705 ovs_mutex_lock(&pmd
->flow_mutex
);
2706 netdev_flow
= dp_netdev_pmd_find_flow(pmd
, del
->ufid
, del
->key
,
2710 get_dpif_flow_stats(netdev_flow
, stats
);
2712 dp_netdev_pmd_remove_flow(pmd
, netdev_flow
);
2716 ovs_mutex_unlock(&pmd
->flow_mutex
);
2722 dpif_netdev_flow_del(struct dpif
*dpif
, const struct dpif_flow_del
*del
)
2724 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2725 struct dp_netdev_pmd_thread
*pmd
;
2729 memset(del
->stats
, 0, sizeof *del
->stats
);
2732 if (del
->pmd_id
== PMD_ID_NULL
) {
2733 if (cmap_count(&dp
->poll_threads
) == 0) {
2736 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
2737 struct dpif_flow_stats pmd_stats
;
2740 pmd_error
= flow_del_on_pmd(pmd
, &pmd_stats
, del
);
2743 } else if (del
->stats
) {
2744 del
->stats
->n_packets
+= pmd_stats
.n_packets
;
2745 del
->stats
->n_bytes
+= pmd_stats
.n_bytes
;
2746 del
->stats
->used
= MAX(del
->stats
->used
, pmd_stats
.used
);
2747 del
->stats
->tcp_flags
|= pmd_stats
.tcp_flags
;
2751 pmd
= dp_netdev_get_pmd(dp
, del
->pmd_id
);
2755 error
= flow_del_on_pmd(pmd
, del
->stats
, del
);
2756 dp_netdev_pmd_unref(pmd
);
2763 struct dpif_netdev_flow_dump
{
2764 struct dpif_flow_dump up
;
2765 struct cmap_position poll_thread_pos
;
2766 struct cmap_position flow_pos
;
2767 struct dp_netdev_pmd_thread
*cur_pmd
;
2769 struct ovs_mutex mutex
;
2772 static struct dpif_netdev_flow_dump
*
2773 dpif_netdev_flow_dump_cast(struct dpif_flow_dump
*dump
)
2775 return CONTAINER_OF(dump
, struct dpif_netdev_flow_dump
, up
);
2778 static struct dpif_flow_dump
*
2779 dpif_netdev_flow_dump_create(const struct dpif
*dpif_
, bool terse
,
2780 char *type OVS_UNUSED
)
2782 struct dpif_netdev_flow_dump
*dump
;
2784 dump
= xzalloc(sizeof *dump
);
2785 dpif_flow_dump_init(&dump
->up
, dpif_
);
2786 dump
->up
.terse
= terse
;
2787 ovs_mutex_init(&dump
->mutex
);
2793 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump
*dump_
)
2795 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
2797 ovs_mutex_destroy(&dump
->mutex
);
2802 struct dpif_netdev_flow_dump_thread
{
2803 struct dpif_flow_dump_thread up
;
2804 struct dpif_netdev_flow_dump
*dump
;
2805 struct odputil_keybuf keybuf
[FLOW_DUMP_MAX_BATCH
];
2806 struct odputil_keybuf maskbuf
[FLOW_DUMP_MAX_BATCH
];
2809 static struct dpif_netdev_flow_dump_thread
*
2810 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread
*thread
)
2812 return CONTAINER_OF(thread
, struct dpif_netdev_flow_dump_thread
, up
);
2815 static struct dpif_flow_dump_thread
*
2816 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump
*dump_
)
2818 struct dpif_netdev_flow_dump
*dump
= dpif_netdev_flow_dump_cast(dump_
);
2819 struct dpif_netdev_flow_dump_thread
*thread
;
2821 thread
= xmalloc(sizeof *thread
);
2822 dpif_flow_dump_thread_init(&thread
->up
, &dump
->up
);
2823 thread
->dump
= dump
;
2828 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread
*thread_
)
2830 struct dpif_netdev_flow_dump_thread
*thread
2831 = dpif_netdev_flow_dump_thread_cast(thread_
);
2837 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread
*thread_
,
2838 struct dpif_flow
*flows
, int max_flows
)
2840 struct dpif_netdev_flow_dump_thread
*thread
2841 = dpif_netdev_flow_dump_thread_cast(thread_
);
2842 struct dpif_netdev_flow_dump
*dump
= thread
->dump
;
2843 struct dp_netdev_flow
*netdev_flows
[FLOW_DUMP_MAX_BATCH
];
2847 ovs_mutex_lock(&dump
->mutex
);
2848 if (!dump
->status
) {
2849 struct dpif_netdev
*dpif
= dpif_netdev_cast(thread
->up
.dpif
);
2850 struct dp_netdev
*dp
= get_dp_netdev(&dpif
->dpif
);
2851 struct dp_netdev_pmd_thread
*pmd
= dump
->cur_pmd
;
2852 int flow_limit
= MIN(max_flows
, FLOW_DUMP_MAX_BATCH
);
2854 /* First call to dump_next(), extracts the first pmd thread.
2855 * If there is no pmd thread, returns immediately. */
2857 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
2859 ovs_mutex_unlock(&dump
->mutex
);
2866 for (n_flows
= 0; n_flows
< flow_limit
; n_flows
++) {
2867 struct cmap_node
*node
;
2869 node
= cmap_next_position(&pmd
->flow_table
, &dump
->flow_pos
);
2873 netdev_flows
[n_flows
] = CONTAINER_OF(node
,
2874 struct dp_netdev_flow
,
2877 /* When finishing dumping the current pmd thread, moves to
2879 if (n_flows
< flow_limit
) {
2880 memset(&dump
->flow_pos
, 0, sizeof dump
->flow_pos
);
2881 dp_netdev_pmd_unref(pmd
);
2882 pmd
= dp_netdev_pmd_get_next(dp
, &dump
->poll_thread_pos
);
2888 /* Keeps the reference to next caller. */
2889 dump
->cur_pmd
= pmd
;
2891 /* If the current dump is empty, do not exit the loop, since the
2892 * remaining pmds could have flows to be dumped. Just dumps again
2893 * on the new 'pmd'. */
2896 ovs_mutex_unlock(&dump
->mutex
);
2898 for (i
= 0; i
< n_flows
; i
++) {
2899 struct odputil_keybuf
*maskbuf
= &thread
->maskbuf
[i
];
2900 struct odputil_keybuf
*keybuf
= &thread
->keybuf
[i
];
2901 struct dp_netdev_flow
*netdev_flow
= netdev_flows
[i
];
2902 struct dpif_flow
*f
= &flows
[i
];
2903 struct ofpbuf key
, mask
;
2905 ofpbuf_use_stack(&key
, keybuf
, sizeof *keybuf
);
2906 ofpbuf_use_stack(&mask
, maskbuf
, sizeof *maskbuf
);
2907 dp_netdev_flow_to_dpif_flow(netdev_flow
, &key
, &mask
, f
,
2915 dpif_netdev_execute(struct dpif
*dpif
, struct dpif_execute
*execute
)
2916 OVS_NO_THREAD_SAFETY_ANALYSIS
2918 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
2919 struct dp_netdev_pmd_thread
*pmd
;
2920 struct dp_packet_batch pp
;
2922 if (dp_packet_size(execute
->packet
) < ETH_HEADER_LEN
||
2923 dp_packet_size(execute
->packet
) > UINT16_MAX
) {
2927 /* Tries finding the 'pmd'. If NULL is returned, that means
2928 * the current thread is a non-pmd thread and should use
2929 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2930 pmd
= ovsthread_getspecific(dp
->per_pmd_key
);
2932 pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
2938 if (execute
->probe
) {
2939 /* If this is part of a probe, Drop the packet, since executing
2940 * the action may actually cause spurious packets be sent into
2942 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2943 dp_netdev_pmd_unref(pmd
);
2948 /* If the current thread is non-pmd thread, acquires
2949 * the 'non_pmd_mutex'. */
2950 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2951 ovs_mutex_lock(&dp
->non_pmd_mutex
);
2954 /* Update current time in PMD context. */
2955 pmd_thread_ctx_time_update(pmd
);
2957 /* The action processing expects the RSS hash to be valid, because
2958 * it's always initialized at the beginning of datapath processing.
2959 * In this case, though, 'execute->packet' may not have gone through
2960 * the datapath at all, it may have been generated by the upper layer
2961 * (OpenFlow packet-out, BFD frame, ...). */
2962 if (!dp_packet_rss_valid(execute
->packet
)) {
2963 dp_packet_set_rss_hash(execute
->packet
,
2964 flow_hash_5tuple(execute
->flow
, 0));
2967 dp_packet_batch_init_packet(&pp
, execute
->packet
);
2968 dp_netdev_execute_actions(pmd
, &pp
, false, execute
->flow
,
2969 execute
->actions
, execute
->actions_len
);
2970 dp_netdev_pmd_flush_output_packets(pmd
);
2972 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
2973 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
2974 dp_netdev_pmd_unref(pmd
);
2981 dpif_netdev_operate(struct dpif
*dpif
, struct dpif_op
**ops
, size_t n_ops
)
2985 for (i
= 0; i
< n_ops
; i
++) {
2986 struct dpif_op
*op
= ops
[i
];
2989 case DPIF_OP_FLOW_PUT
:
2990 op
->error
= dpif_netdev_flow_put(dpif
, &op
->u
.flow_put
);
2993 case DPIF_OP_FLOW_DEL
:
2994 op
->error
= dpif_netdev_flow_del(dpif
, &op
->u
.flow_del
);
2997 case DPIF_OP_EXECUTE
:
2998 op
->error
= dpif_netdev_execute(dpif
, &op
->u
.execute
);
3001 case DPIF_OP_FLOW_GET
:
3002 op
->error
= dpif_netdev_flow_get(dpif
, &op
->u
.flow_get
);
3008 /* Applies datapath configuration from the database. Some of the changes are
3009 * actually applied in dpif_netdev_run(). */
3011 dpif_netdev_set_config(struct dpif
*dpif
, const struct smap
*other_config
)
3013 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3014 const char *cmask
= smap_get(other_config
, "pmd-cpu-mask");
3015 unsigned long long insert_prob
=
3016 smap_get_ullong(other_config
, "emc-insert-inv-prob",
3017 DEFAULT_EM_FLOW_INSERT_INV_PROB
);
3018 uint32_t insert_min
, cur_min
;
3020 if (!nullable_string_is_equal(dp
->pmd_cmask
, cmask
)) {
3021 free(dp
->pmd_cmask
);
3022 dp
->pmd_cmask
= nullable_xstrdup(cmask
);
3023 dp_netdev_request_reconfigure(dp
);
3026 atomic_read_relaxed(&dp
->emc_insert_min
, &cur_min
);
3027 if (insert_prob
<= UINT32_MAX
) {
3028 insert_min
= insert_prob
== 0 ? 0 : UINT32_MAX
/ insert_prob
;
3030 insert_min
= DEFAULT_EM_FLOW_INSERT_MIN
;
3031 insert_prob
= DEFAULT_EM_FLOW_INSERT_INV_PROB
;
3034 if (insert_min
!= cur_min
) {
3035 atomic_store_relaxed(&dp
->emc_insert_min
, insert_min
);
3036 if (insert_min
== 0) {
3037 VLOG_INFO("EMC has been disabled");
3039 VLOG_INFO("EMC insertion probability changed to 1/%llu (~%.2f%%)",
3040 insert_prob
, (100 / (float)insert_prob
));
3047 /* Parses affinity list and returns result in 'core_ids'. */
3049 parse_affinity_list(const char *affinity_list
, unsigned *core_ids
, int n_rxq
)
3052 char *list
, *copy
, *key
, *value
;
3055 for (i
= 0; i
< n_rxq
; i
++) {
3056 core_ids
[i
] = OVS_CORE_UNSPEC
;
3059 if (!affinity_list
) {
3063 list
= copy
= xstrdup(affinity_list
);
3065 while (ofputil_parse_key_value(&list
, &key
, &value
)) {
3066 int rxq_id
, core_id
;
3068 if (!str_to_int(key
, 0, &rxq_id
) || rxq_id
< 0
3069 || !str_to_int(value
, 0, &core_id
) || core_id
< 0) {
3074 if (rxq_id
< n_rxq
) {
3075 core_ids
[rxq_id
] = core_id
;
3083 /* Parses 'affinity_list' and applies configuration if it is valid. */
3085 dpif_netdev_port_set_rxq_affinity(struct dp_netdev_port
*port
,
3086 const char *affinity_list
)
3088 unsigned *core_ids
, i
;
3091 core_ids
= xmalloc(port
->n_rxq
* sizeof *core_ids
);
3092 if (parse_affinity_list(affinity_list
, core_ids
, port
->n_rxq
)) {
3097 for (i
= 0; i
< port
->n_rxq
; i
++) {
3098 port
->rxqs
[i
].core_id
= core_ids
[i
];
3106 /* Changes the affinity of port's rx queues. The changes are actually applied
3107 * in dpif_netdev_run(). */
3109 dpif_netdev_port_set_config(struct dpif
*dpif
, odp_port_t port_no
,
3110 const struct smap
*cfg
)
3112 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3113 struct dp_netdev_port
*port
;
3115 const char *affinity_list
= smap_get(cfg
, "pmd-rxq-affinity");
3117 ovs_mutex_lock(&dp
->port_mutex
);
3118 error
= get_port_by_number(dp
, port_no
, &port
);
3119 if (error
|| !netdev_is_pmd(port
->netdev
)
3120 || nullable_string_is_equal(affinity_list
, port
->rxq_affinity_list
)) {
3124 error
= dpif_netdev_port_set_rxq_affinity(port
, affinity_list
);
3128 free(port
->rxq_affinity_list
);
3129 port
->rxq_affinity_list
= nullable_xstrdup(affinity_list
);
3131 dp_netdev_request_reconfigure(dp
);
3133 ovs_mutex_unlock(&dp
->port_mutex
);
3138 dpif_netdev_queue_to_priority(const struct dpif
*dpif OVS_UNUSED
,
3139 uint32_t queue_id
, uint32_t *priority
)
3141 *priority
= queue_id
;
3146 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
3147 * a copy of the 'size' bytes of 'actions' input parameters. */
3148 struct dp_netdev_actions
*
3149 dp_netdev_actions_create(const struct nlattr
*actions
, size_t size
)
3151 struct dp_netdev_actions
*netdev_actions
;
3153 netdev_actions
= xmalloc(sizeof *netdev_actions
+ size
);
3154 memcpy(netdev_actions
->actions
, actions
, size
);
3155 netdev_actions
->size
= size
;
3157 return netdev_actions
;
3160 struct dp_netdev_actions
*
3161 dp_netdev_flow_get_actions(const struct dp_netdev_flow
*flow
)
3163 return ovsrcu_get(struct dp_netdev_actions
*, &flow
->actions
);
3167 dp_netdev_actions_free(struct dp_netdev_actions
*actions
)
3172 static inline unsigned long long
3173 cycles_counter(void)
3176 return rte_get_tsc_cycles();
3182 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
3183 extern struct ovs_mutex cycles_counter_fake_mutex
;
3185 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
3187 cycles_count_start(struct dp_netdev_pmd_thread
*pmd
)
3188 OVS_ACQUIRES(&cycles_counter_fake_mutex
)
3189 OVS_NO_THREAD_SAFETY_ANALYSIS
3191 pmd
->ctx
.last_cycles
= cycles_counter();
3194 /* Stop counting cycles and add them to the counter 'type' */
3196 cycles_count_end(struct dp_netdev_pmd_thread
*pmd
,
3197 enum pmd_cycles_counter_type type
)
3198 OVS_RELEASES(&cycles_counter_fake_mutex
)
3199 OVS_NO_THREAD_SAFETY_ANALYSIS
3201 unsigned long long interval
= cycles_counter() - pmd
->ctx
.last_cycles
;
3203 non_atomic_ullong_add(&pmd
->cycles
.n
[type
], interval
);
3206 /* Calculate the intermediate cycle result and add to the counter 'type' */
3208 cycles_count_intermediate(struct dp_netdev_pmd_thread
*pmd
,
3209 struct dp_netdev_rxq
*rxq
,
3210 enum pmd_cycles_counter_type type
)
3211 OVS_NO_THREAD_SAFETY_ANALYSIS
3213 unsigned long long new_cycles
= cycles_counter();
3214 unsigned long long interval
= new_cycles
- pmd
->ctx
.last_cycles
;
3215 pmd
->ctx
.last_cycles
= new_cycles
;
3217 non_atomic_ullong_add(&pmd
->cycles
.n
[type
], interval
);
3218 if (rxq
&& (type
== PMD_CYCLES_PROCESSING
)) {
3219 /* Add to the amount of current processing cycles. */
3220 non_atomic_ullong_add(&rxq
->cycles
[RXQ_CYCLES_PROC_CURR
], interval
);
3225 dp_netdev_rxq_set_cycles(struct dp_netdev_rxq
*rx
,
3226 enum rxq_cycles_counter_type type
,
3227 unsigned long long cycles
)
3229 atomic_store_relaxed(&rx
->cycles
[type
], cycles
);
3233 dp_netdev_rxq_get_cycles(struct dp_netdev_rxq
*rx
,
3234 enum rxq_cycles_counter_type type
)
3236 unsigned long long processing_cycles
;
3237 atomic_read_relaxed(&rx
->cycles
[type
], &processing_cycles
);
3238 return processing_cycles
;
3242 dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq
*rx
,
3243 unsigned long long cycles
)
3245 unsigned int idx
= rx
->intrvl_idx
++ % PMD_RXQ_INTERVAL_MAX
;
3246 atomic_store_relaxed(&rx
->cycles_intrvl
[idx
], cycles
);
3250 dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq
*rx
, unsigned idx
)
3252 unsigned long long processing_cycles
;
3253 atomic_read_relaxed(&rx
->cycles_intrvl
[idx
], &processing_cycles
);
3254 return processing_cycles
;
3258 dp_netdev_pmd_flush_output_on_port(struct dp_netdev_pmd_thread
*pmd
,
3264 dynamic_txqs
= p
->port
->dynamic_txqs
;
3266 tx_qid
= dpif_netdev_xps_get_tx_qid(pmd
, p
);
3268 tx_qid
= pmd
->static_tx_qid
;
3271 netdev_send(p
->port
->netdev
, tx_qid
, &p
->output_pkts
, dynamic_txqs
);
3272 dp_packet_batch_init(&p
->output_pkts
);
3276 dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread
*pmd
)
3280 HMAP_FOR_EACH (p
, node
, &pmd
->send_port_cache
) {
3281 if (!dp_packet_batch_is_empty(&p
->output_pkts
)) {
3282 dp_netdev_pmd_flush_output_on_port(pmd
, p
);
3288 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread
*pmd
,
3289 struct netdev_rxq
*rx
,
3292 struct dp_packet_batch batch
;
3296 dp_packet_batch_init(&batch
);
3297 error
= netdev_rxq_recv(rx
, &batch
);
3299 *recirc_depth_get() = 0;
3300 pmd_thread_ctx_time_update(pmd
);
3302 batch_cnt
= batch
.count
;
3303 dp_netdev_input(pmd
, &batch
, port_no
);
3304 dp_netdev_pmd_flush_output_packets(pmd
);
3305 } else if (error
!= EAGAIN
&& error
!= EOPNOTSUPP
) {
3306 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(1, 5);
3308 VLOG_ERR_RL(&rl
, "error receiving data from %s: %s",
3309 netdev_rxq_get_name(rx
), ovs_strerror(error
));
3315 static struct tx_port
*
3316 tx_port_lookup(const struct hmap
*hmap
, odp_port_t port_no
)
3320 HMAP_FOR_EACH_IN_BUCKET (tx
, node
, hash_port_no(port_no
), hmap
) {
3321 if (tx
->port
->port_no
== port_no
) {
3330 port_reconfigure(struct dp_netdev_port
*port
)
3332 struct netdev
*netdev
= port
->netdev
;
3335 port
->need_reconfigure
= false;
3337 /* Closes the existing 'rxq's. */
3338 for (i
= 0; i
< port
->n_rxq
; i
++) {
3339 netdev_rxq_close(port
->rxqs
[i
].rx
);
3340 port
->rxqs
[i
].rx
= NULL
;
3342 unsigned last_nrxq
= port
->n_rxq
;
3345 /* Allows 'netdev' to apply the pending configuration changes. */
3346 if (netdev_is_reconf_required(netdev
)) {
3347 err
= netdev_reconfigure(netdev
);
3348 if (err
&& (err
!= EOPNOTSUPP
)) {
3349 VLOG_ERR("Failed to set interface %s new configuration",
3350 netdev_get_name(netdev
));
3354 /* If the netdev_reconfigure() above succeeds, reopens the 'rxq's. */
3355 port
->rxqs
= xrealloc(port
->rxqs
,
3356 sizeof *port
->rxqs
* netdev_n_rxq(netdev
));
3357 /* Realloc 'used' counters for tx queues. */
3358 free(port
->txq_used
);
3359 port
->txq_used
= xcalloc(netdev_n_txq(netdev
), sizeof *port
->txq_used
);
3361 for (i
= 0; i
< netdev_n_rxq(netdev
); i
++) {
3362 bool new_queue
= i
>= last_nrxq
;
3364 memset(&port
->rxqs
[i
], 0, sizeof port
->rxqs
[i
]);
3367 port
->rxqs
[i
].port
= port
;
3369 err
= netdev_rxq_open(netdev
, &port
->rxqs
[i
].rx
, i
);
3376 /* Parse affinity list to apply configuration for new queues. */
3377 dpif_netdev_port_set_rxq_affinity(port
, port
->rxq_affinity_list
);
3382 struct rr_numa_list
{
3383 struct hmap numas
; /* Contains 'struct rr_numa' */
3387 struct hmap_node node
;
3391 /* Non isolated pmds on numa node 'numa_id' */
3392 struct dp_netdev_pmd_thread
**pmds
;
3399 static struct rr_numa
*
3400 rr_numa_list_lookup(struct rr_numa_list
*rr
, int numa_id
)
3402 struct rr_numa
*numa
;
3404 HMAP_FOR_EACH_WITH_HASH (numa
, node
, hash_int(numa_id
, 0), &rr
->numas
) {
3405 if (numa
->numa_id
== numa_id
) {
3413 /* Returns the next node in numa list following 'numa' in round-robin fashion.
3414 * Returns first node if 'numa' is a null pointer or the last node in 'rr'.
3415 * Returns NULL if 'rr' numa list is empty. */
3416 static struct rr_numa
*
3417 rr_numa_list_next(struct rr_numa_list
*rr
, const struct rr_numa
*numa
)
3419 struct hmap_node
*node
= NULL
;
3422 node
= hmap_next(&rr
->numas
, &numa
->node
);
3425 node
= hmap_first(&rr
->numas
);
3428 return (node
) ? CONTAINER_OF(node
, struct rr_numa
, node
) : NULL
;
3432 rr_numa_list_populate(struct dp_netdev
*dp
, struct rr_numa_list
*rr
)
3434 struct dp_netdev_pmd_thread
*pmd
;
3435 struct rr_numa
*numa
;
3437 hmap_init(&rr
->numas
);
3439 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3440 if (pmd
->core_id
== NON_PMD_CORE_ID
|| pmd
->isolated
) {
3444 numa
= rr_numa_list_lookup(rr
, pmd
->numa_id
);
3446 numa
= xzalloc(sizeof *numa
);
3447 numa
->numa_id
= pmd
->numa_id
;
3448 hmap_insert(&rr
->numas
, &numa
->node
, hash_int(pmd
->numa_id
, 0));
3451 numa
->pmds
= xrealloc(numa
->pmds
, numa
->n_pmds
* sizeof *numa
->pmds
);
3452 numa
->pmds
[numa
->n_pmds
- 1] = pmd
;
3453 /* At least one pmd so initialise curr_idx and idx_inc. */
3454 numa
->cur_index
= 0;
3455 numa
->idx_inc
= true;
3459 /* Returns the next pmd from the numa node in
3460 * incrementing or decrementing order. */
3461 static struct dp_netdev_pmd_thread
*
3462 rr_numa_get_pmd(struct rr_numa
*numa
)
3464 int numa_idx
= numa
->cur_index
;
3466 if (numa
->idx_inc
== true) {
3467 /* Incrementing through list of pmds. */
3468 if (numa
->cur_index
== numa
->n_pmds
-1) {
3469 /* Reached the last pmd. */
3470 numa
->idx_inc
= false;
3475 /* Decrementing through list of pmds. */
3476 if (numa
->cur_index
== 0) {
3477 /* Reached the first pmd. */
3478 numa
->idx_inc
= true;
3483 return numa
->pmds
[numa_idx
];
3487 rr_numa_list_destroy(struct rr_numa_list
*rr
)
3489 struct rr_numa
*numa
;
3491 HMAP_FOR_EACH_POP (numa
, node
, &rr
->numas
) {
3495 hmap_destroy(&rr
->numas
);
3498 /* Sort Rx Queues by the processing cycles they are consuming. */
3500 compare_rxq_cycles(const void *a
, const void *b
)
3502 struct dp_netdev_rxq
*qa
;
3503 struct dp_netdev_rxq
*qb
;
3504 uint64_t cycles_qa
, cycles_qb
;
3506 qa
= *(struct dp_netdev_rxq
**) a
;
3507 qb
= *(struct dp_netdev_rxq
**) b
;
3509 cycles_qa
= dp_netdev_rxq_get_cycles(qa
, RXQ_CYCLES_PROC_HIST
);
3510 cycles_qb
= dp_netdev_rxq_get_cycles(qb
, RXQ_CYCLES_PROC_HIST
);
3512 if (cycles_qa
!= cycles_qb
) {
3513 return (cycles_qa
< cycles_qb
) ? 1 : -1;
3515 /* Cycles are the same so tiebreak on port/queue id.
3516 * Tiebreaking (as opposed to return 0) ensures consistent
3517 * sort results across multiple OS's. */
3518 uint32_t port_qa
= odp_to_u32(qa
->port
->port_no
);
3519 uint32_t port_qb
= odp_to_u32(qb
->port
->port_no
);
3520 if (port_qa
!= port_qb
) {
3521 return port_qa
> port_qb
? 1 : -1;
3523 return netdev_rxq_get_queue_id(qa
->rx
)
3524 - netdev_rxq_get_queue_id(qb
->rx
);
3529 /* Assign pmds to queues. If 'pinned' is true, assign pmds to pinned
3530 * queues and marks the pmds as isolated. Otherwise, assign non isolated
3531 * pmds to unpinned queues.
3533 * If 'pinned' is false queues will be sorted by processing cycles they are
3534 * consuming and then assigned to pmds in round robin order.
3536 * The function doesn't touch the pmd threads, it just stores the assignment
3537 * in the 'pmd' member of each rxq. */
3539 rxq_scheduling(struct dp_netdev
*dp
, bool pinned
) OVS_REQUIRES(dp
->port_mutex
)
3541 struct dp_netdev_port
*port
;
3542 struct rr_numa_list rr
;
3543 struct rr_numa
*non_local_numa
= NULL
;
3544 struct dp_netdev_rxq
** rxqs
= NULL
;
3546 struct rr_numa
*numa
= NULL
;
3549 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3550 if (!netdev_is_pmd(port
->netdev
)) {
3554 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
3555 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
3557 if (pinned
&& q
->core_id
!= OVS_CORE_UNSPEC
) {
3558 struct dp_netdev_pmd_thread
*pmd
;
3560 pmd
= dp_netdev_get_pmd(dp
, q
->core_id
);
3562 VLOG_WARN("There is no PMD thread on core %d. Queue "
3563 "%d on port \'%s\' will not be polled.",
3564 q
->core_id
, qid
, netdev_get_name(port
->netdev
));
3567 pmd
->isolated
= true;
3568 dp_netdev_pmd_unref(pmd
);
3570 } else if (!pinned
&& q
->core_id
== OVS_CORE_UNSPEC
) {
3571 uint64_t cycle_hist
= 0;
3574 rxqs
= xmalloc(sizeof *rxqs
);
3576 rxqs
= xrealloc(rxqs
, sizeof *rxqs
* (n_rxqs
+ 1));
3578 /* Sum the queue intervals and store the cycle history. */
3579 for (unsigned i
= 0; i
< PMD_RXQ_INTERVAL_MAX
; i
++) {
3580 cycle_hist
+= dp_netdev_rxq_get_intrvl_cycles(q
, i
);
3582 dp_netdev_rxq_set_cycles(q
, RXQ_CYCLES_PROC_HIST
, cycle_hist
);
3584 /* Store the queue. */
3591 /* Sort the queues in order of the processing cycles
3592 * they consumed during their last pmd interval. */
3593 qsort(rxqs
, n_rxqs
, sizeof *rxqs
, compare_rxq_cycles
);
3596 rr_numa_list_populate(dp
, &rr
);
3597 /* Assign the sorted queues to pmds in round robin. */
3598 for (i
= 0; i
< n_rxqs
; i
++) {
3599 numa_id
= netdev_get_numa_id(rxqs
[i
]->port
->netdev
);
3600 numa
= rr_numa_list_lookup(&rr
, numa_id
);
3602 /* There are no pmds on the queue's local NUMA node.
3603 Round robin on the NUMA nodes that do have pmds. */
3604 non_local_numa
= rr_numa_list_next(&rr
, non_local_numa
);
3605 if (!non_local_numa
) {
3606 VLOG_ERR("There is no available (non-isolated) pmd "
3607 "thread for port \'%s\' queue %d. This queue "
3608 "will not be polled. Is pmd-cpu-mask set to "
3609 "zero? Or are all PMDs isolated to other "
3610 "queues?", netdev_rxq_get_name(rxqs
[i
]->rx
),
3611 netdev_rxq_get_queue_id(rxqs
[i
]->rx
));
3614 rxqs
[i
]->pmd
= rr_numa_get_pmd(non_local_numa
);
3615 VLOG_WARN("There's no available (non-isolated) pmd thread "
3616 "on numa node %d. Queue %d on port \'%s\' will "
3617 "be assigned to the pmd on core %d "
3618 "(numa node %d). Expect reduced performance.",
3619 numa_id
, netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
3620 netdev_rxq_get_name(rxqs
[i
]->rx
),
3621 rxqs
[i
]->pmd
->core_id
, rxqs
[i
]->pmd
->numa_id
);
3623 rxqs
[i
]->pmd
= rr_numa_get_pmd(numa
);
3624 VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
3625 "rx queue %d (measured processing cycles %"PRIu64
").",
3626 rxqs
[i
]->pmd
->core_id
, numa_id
,
3627 netdev_rxq_get_name(rxqs
[i
]->rx
),
3628 netdev_rxq_get_queue_id(rxqs
[i
]->rx
),
3629 dp_netdev_rxq_get_cycles(rxqs
[i
], RXQ_CYCLES_PROC_HIST
));
3633 rr_numa_list_destroy(&rr
);
3638 reload_affected_pmds(struct dp_netdev
*dp
)
3640 struct dp_netdev_pmd_thread
*pmd
;
3642 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3643 if (pmd
->need_reload
) {
3644 dp_netdev_reload_pmd__(pmd
);
3645 pmd
->need_reload
= false;
3651 reconfigure_pmd_threads(struct dp_netdev
*dp
)
3652 OVS_REQUIRES(dp
->port_mutex
)
3654 struct dp_netdev_pmd_thread
*pmd
;
3655 struct ovs_numa_dump
*pmd_cores
;
3656 struct ovs_numa_info_core
*core
;
3657 struct hmapx to_delete
= HMAPX_INITIALIZER(&to_delete
);
3658 struct hmapx_node
*node
;
3659 bool changed
= false;
3660 bool need_to_adjust_static_tx_qids
= false;
3662 /* The pmd threads should be started only if there's a pmd port in the
3663 * datapath. If the user didn't provide any "pmd-cpu-mask", we start
3664 * NR_PMD_THREADS per numa node. */
3665 if (!has_pmd_port(dp
)) {
3666 pmd_cores
= ovs_numa_dump_n_cores_per_numa(0);
3667 } else if (dp
->pmd_cmask
&& dp
->pmd_cmask
[0]) {
3668 pmd_cores
= ovs_numa_dump_cores_with_cmask(dp
->pmd_cmask
);
3670 pmd_cores
= ovs_numa_dump_n_cores_per_numa(NR_PMD_THREADS
);
3673 /* We need to adjust 'static_tx_qid's only if we're reducing number of
3674 * PMD threads. Otherwise, new threads will allocate all the freed ids. */
3675 if (ovs_numa_dump_count(pmd_cores
) < cmap_count(&dp
->poll_threads
) - 1) {
3676 /* Adjustment is required to keep 'static_tx_qid's sequential and
3677 * avoid possible issues, for example, imbalanced tx queue usage
3678 * and unnecessary locking caused by remapping on netdev level. */
3679 need_to_adjust_static_tx_qids
= true;
3682 /* Check for unwanted pmd threads */
3683 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3684 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
3687 if (!ovs_numa_dump_contains_core(pmd_cores
, pmd
->numa_id
,
3689 hmapx_add(&to_delete
, pmd
);
3690 } else if (need_to_adjust_static_tx_qids
) {
3691 pmd
->need_reload
= true;
3695 HMAPX_FOR_EACH (node
, &to_delete
) {
3696 pmd
= (struct dp_netdev_pmd_thread
*) node
->data
;
3697 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d destroyed.",
3698 pmd
->numa_id
, pmd
->core_id
);
3699 dp_netdev_del_pmd(dp
, pmd
);
3701 changed
= !hmapx_is_empty(&to_delete
);
3702 hmapx_destroy(&to_delete
);
3704 if (need_to_adjust_static_tx_qids
) {
3705 /* 'static_tx_qid's are not sequential now.
3706 * Reload remaining threads to fix this. */
3707 reload_affected_pmds(dp
);
3710 /* Check for required new pmd threads */
3711 FOR_EACH_CORE_ON_DUMP(core
, pmd_cores
) {
3712 pmd
= dp_netdev_get_pmd(dp
, core
->core_id
);
3714 pmd
= xzalloc(sizeof *pmd
);
3715 dp_netdev_configure_pmd(pmd
, dp
, core
->core_id
, core
->numa_id
);
3716 pmd
->thread
= ovs_thread_create("pmd", pmd_thread_main
, pmd
);
3717 VLOG_INFO("PMD thread on numa_id: %d, core id: %2d created.",
3718 pmd
->numa_id
, pmd
->core_id
);
3721 dp_netdev_pmd_unref(pmd
);
3726 struct ovs_numa_info_numa
*numa
;
3728 /* Log the number of pmd threads per numa node. */
3729 FOR_EACH_NUMA_ON_DUMP (numa
, pmd_cores
) {
3730 VLOG_INFO("There are %"PRIuSIZE
" pmd threads on numa node %d",
3731 numa
->n_cores
, numa
->numa_id
);
3735 ovs_numa_dump_destroy(pmd_cores
);
3739 pmd_remove_stale_ports(struct dp_netdev
*dp
,
3740 struct dp_netdev_pmd_thread
*pmd
)
3741 OVS_EXCLUDED(pmd
->port_mutex
)
3742 OVS_REQUIRES(dp
->port_mutex
)
3744 struct rxq_poll
*poll
, *poll_next
;
3745 struct tx_port
*tx
, *tx_next
;
3747 ovs_mutex_lock(&pmd
->port_mutex
);
3748 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
3749 struct dp_netdev_port
*port
= poll
->rxq
->port
;
3751 if (port
->need_reconfigure
3752 || !hmap_contains(&dp
->ports
, &port
->node
)) {
3753 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
3756 HMAP_FOR_EACH_SAFE (tx
, tx_next
, node
, &pmd
->tx_ports
) {
3757 struct dp_netdev_port
*port
= tx
->port
;
3759 if (port
->need_reconfigure
3760 || !hmap_contains(&dp
->ports
, &port
->node
)) {
3761 dp_netdev_del_port_tx_from_pmd(pmd
, tx
);
3764 ovs_mutex_unlock(&pmd
->port_mutex
);
3767 /* Must be called each time a port is added/removed or the cmask changes.
3768 * This creates and destroys pmd threads, reconfigures ports, opens their
3769 * rxqs and assigns all rxqs/txqs to pmd threads. */
3771 reconfigure_datapath(struct dp_netdev
*dp
)
3772 OVS_REQUIRES(dp
->port_mutex
)
3774 struct dp_netdev_pmd_thread
*pmd
;
3775 struct dp_netdev_port
*port
;
3778 dp
->last_reconfigure_seq
= seq_read(dp
->reconfigure_seq
);
3780 /* Step 1: Adjust the pmd threads based on the datapath ports, the cores
3781 * on the system and the user configuration. */
3782 reconfigure_pmd_threads(dp
);
3784 wanted_txqs
= cmap_count(&dp
->poll_threads
);
3786 /* The number of pmd threads might have changed, or a port can be new:
3787 * adjust the txqs. */
3788 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3789 netdev_set_tx_multiq(port
->netdev
, wanted_txqs
);
3792 /* Step 2: Remove from the pmd threads ports that have been removed or
3793 * need reconfiguration. */
3795 /* Check for all the ports that need reconfiguration. We cache this in
3796 * 'port->need_reconfigure', because netdev_is_reconf_required() can
3797 * change at any time. */
3798 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3799 if (netdev_is_reconf_required(port
->netdev
)) {
3800 port
->need_reconfigure
= true;
3804 /* Remove from the pmd threads all the ports that have been deleted or
3805 * need reconfiguration. */
3806 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3807 pmd_remove_stale_ports(dp
, pmd
);
3810 /* Reload affected pmd threads. We must wait for the pmd threads before
3811 * reconfiguring the ports, because a port cannot be reconfigured while
3812 * it's being used. */
3813 reload_affected_pmds(dp
);
3815 /* Step 3: Reconfigure ports. */
3817 /* We only reconfigure the ports that we determined above, because they're
3818 * not being used by any pmd thread at the moment. If a port fails to
3819 * reconfigure we remove it from the datapath. */
3820 struct dp_netdev_port
*next_port
;
3821 HMAP_FOR_EACH_SAFE (port
, next_port
, node
, &dp
->ports
) {
3824 if (!port
->need_reconfigure
) {
3828 err
= port_reconfigure(port
);
3830 hmap_remove(&dp
->ports
, &port
->node
);
3831 seq_change(dp
->port_seq
);
3834 port
->dynamic_txqs
= netdev_n_txq(port
->netdev
) < wanted_txqs
;
3838 /* Step 4: Compute new rxq scheduling. We don't touch the pmd threads
3839 * for now, we just update the 'pmd' pointer in each rxq to point to the
3840 * wanted thread according to the scheduling policy. */
3842 /* Reset all the pmd threads to non isolated. */
3843 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3844 pmd
->isolated
= false;
3847 /* Reset all the queues to unassigned */
3848 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3849 for (int i
= 0; i
< port
->n_rxq
; i
++) {
3850 port
->rxqs
[i
].pmd
= NULL
;
3854 /* Add pinned queues and mark pmd threads isolated. */
3855 rxq_scheduling(dp
, true);
3857 /* Add non-pinned queues. */
3858 rxq_scheduling(dp
, false);
3860 /* Step 5: Remove queues not compliant with new scheduling. */
3861 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3862 struct rxq_poll
*poll
, *poll_next
;
3864 ovs_mutex_lock(&pmd
->port_mutex
);
3865 HMAP_FOR_EACH_SAFE (poll
, poll_next
, node
, &pmd
->poll_list
) {
3866 if (poll
->rxq
->pmd
!= pmd
) {
3867 dp_netdev_del_rxq_from_pmd(pmd
, poll
);
3870 ovs_mutex_unlock(&pmd
->port_mutex
);
3873 /* Reload affected pmd threads. We must wait for the pmd threads to remove
3874 * the old queues before readding them, otherwise a queue can be polled by
3875 * two threads at the same time. */
3876 reload_affected_pmds(dp
);
3878 /* Step 6: Add queues from scheduling, if they're not there already. */
3879 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3880 if (!netdev_is_pmd(port
->netdev
)) {
3884 for (int qid
= 0; qid
< port
->n_rxq
; qid
++) {
3885 struct dp_netdev_rxq
*q
= &port
->rxqs
[qid
];
3888 ovs_mutex_lock(&q
->pmd
->port_mutex
);
3889 dp_netdev_add_rxq_to_pmd(q
->pmd
, q
);
3890 ovs_mutex_unlock(&q
->pmd
->port_mutex
);
3895 /* Add every port to the tx cache of every pmd thread, if it's not
3896 * there already and if this pmd has at least one rxq to poll. */
3897 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
3898 ovs_mutex_lock(&pmd
->port_mutex
);
3899 if (hmap_count(&pmd
->poll_list
) || pmd
->core_id
== NON_PMD_CORE_ID
) {
3900 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3901 dp_netdev_add_port_tx_to_pmd(pmd
, port
);
3904 ovs_mutex_unlock(&pmd
->port_mutex
);
3907 /* Reload affected pmd threads. */
3908 reload_affected_pmds(dp
);
3911 /* Returns true if one of the netdevs in 'dp' requires a reconfiguration */
3913 ports_require_restart(const struct dp_netdev
*dp
)
3914 OVS_REQUIRES(dp
->port_mutex
)
3916 struct dp_netdev_port
*port
;
3918 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3919 if (netdev_is_reconf_required(port
->netdev
)) {
3927 /* Return true if needs to revalidate datapath flows. */
3929 dpif_netdev_run(struct dpif
*dpif
)
3931 struct dp_netdev_port
*port
;
3932 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3933 struct dp_netdev_pmd_thread
*non_pmd
;
3934 uint64_t new_tnl_seq
;
3935 int process_packets
= 0;
3937 ovs_mutex_lock(&dp
->port_mutex
);
3938 non_pmd
= dp_netdev_get_pmd(dp
, NON_PMD_CORE_ID
);
3940 ovs_mutex_lock(&dp
->non_pmd_mutex
);
3941 cycles_count_start(non_pmd
);
3942 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3943 if (!netdev_is_pmd(port
->netdev
)) {
3946 for (i
= 0; i
< port
->n_rxq
; i
++) {
3948 dp_netdev_process_rxq_port(non_pmd
,
3951 cycles_count_intermediate(non_pmd
, NULL
,
3953 ? PMD_CYCLES_PROCESSING
3958 cycles_count_end(non_pmd
, PMD_CYCLES_IDLE
);
3959 pmd_thread_ctx_time_update(non_pmd
);
3960 dpif_netdev_xps_revalidate_pmd(non_pmd
, false);
3961 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
3963 dp_netdev_pmd_unref(non_pmd
);
3966 if (dp_netdev_is_reconf_required(dp
) || ports_require_restart(dp
)) {
3967 reconfigure_datapath(dp
);
3969 ovs_mutex_unlock(&dp
->port_mutex
);
3971 tnl_neigh_cache_run();
3973 new_tnl_seq
= seq_read(tnl_conf_seq
);
3975 if (dp
->last_tnl_conf_seq
!= new_tnl_seq
) {
3976 dp
->last_tnl_conf_seq
= new_tnl_seq
;
3983 dpif_netdev_wait(struct dpif
*dpif
)
3985 struct dp_netdev_port
*port
;
3986 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
3988 ovs_mutex_lock(&dp_netdev_mutex
);
3989 ovs_mutex_lock(&dp
->port_mutex
);
3990 HMAP_FOR_EACH (port
, node
, &dp
->ports
) {
3991 netdev_wait_reconf_required(port
->netdev
);
3992 if (!netdev_is_pmd(port
->netdev
)) {
3995 for (i
= 0; i
< port
->n_rxq
; i
++) {
3996 netdev_rxq_wait(port
->rxqs
[i
].rx
);
4000 ovs_mutex_unlock(&dp
->port_mutex
);
4001 ovs_mutex_unlock(&dp_netdev_mutex
);
4002 seq_wait(tnl_conf_seq
, dp
->last_tnl_conf_seq
);
4006 pmd_free_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
4008 struct tx_port
*tx_port_cached
;
4010 /* Free all used tx queue ids. */
4011 dpif_netdev_xps_revalidate_pmd(pmd
, true);
4013 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->tnl_port_cache
) {
4014 free(tx_port_cached
);
4016 HMAP_FOR_EACH_POP (tx_port_cached
, node
, &pmd
->send_port_cache
) {
4017 free(tx_port_cached
);
4021 /* Copies ports from 'pmd->tx_ports' (shared with the main thread) to
4022 * thread-local copies. Copy to 'pmd->tnl_port_cache' if it is a tunnel
4023 * device, otherwise to 'pmd->send_port_cache' if the port has at least
4026 pmd_load_cached_ports(struct dp_netdev_pmd_thread
*pmd
)
4027 OVS_REQUIRES(pmd
->port_mutex
)
4029 struct tx_port
*tx_port
, *tx_port_cached
;
4031 pmd_free_cached_ports(pmd
);
4032 hmap_shrink(&pmd
->send_port_cache
);
4033 hmap_shrink(&pmd
->tnl_port_cache
);
4035 HMAP_FOR_EACH (tx_port
, node
, &pmd
->tx_ports
) {
4036 if (netdev_has_tunnel_push_pop(tx_port
->port
->netdev
)) {
4037 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
4038 hmap_insert(&pmd
->tnl_port_cache
, &tx_port_cached
->node
,
4039 hash_port_no(tx_port_cached
->port
->port_no
));
4042 if (netdev_n_txq(tx_port
->port
->netdev
)) {
4043 tx_port_cached
= xmemdup(tx_port
, sizeof *tx_port_cached
);
4044 hmap_insert(&pmd
->send_port_cache
, &tx_port_cached
->node
,
4045 hash_port_no(tx_port_cached
->port
->port_no
));
4051 pmd_alloc_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
4053 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
4054 if (!id_pool_alloc_id(pmd
->dp
->tx_qid_pool
, &pmd
->static_tx_qid
)) {
4055 VLOG_ABORT("static_tx_qid allocation failed for PMD on core %2d"
4056 ", numa_id %d.", pmd
->core_id
, pmd
->numa_id
);
4058 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
4060 VLOG_DBG("static_tx_qid = %d allocated for PMD thread on core %2d"
4061 ", numa_id %d.", pmd
->static_tx_qid
, pmd
->core_id
, pmd
->numa_id
);
4065 pmd_free_static_tx_qid(struct dp_netdev_pmd_thread
*pmd
)
4067 ovs_mutex_lock(&pmd
->dp
->tx_qid_pool_mutex
);
4068 id_pool_free_id(pmd
->dp
->tx_qid_pool
, pmd
->static_tx_qid
);
4069 ovs_mutex_unlock(&pmd
->dp
->tx_qid_pool_mutex
);
4073 pmd_load_queues_and_ports(struct dp_netdev_pmd_thread
*pmd
,
4074 struct polled_queue
**ppoll_list
)
4076 struct polled_queue
*poll_list
= *ppoll_list
;
4077 struct rxq_poll
*poll
;
4080 ovs_mutex_lock(&pmd
->port_mutex
);
4081 poll_list
= xrealloc(poll_list
, hmap_count(&pmd
->poll_list
)
4082 * sizeof *poll_list
);
4085 HMAP_FOR_EACH (poll
, node
, &pmd
->poll_list
) {
4086 poll_list
[i
].rxq
= poll
->rxq
;
4087 poll_list
[i
].port_no
= poll
->rxq
->port
->port_no
;
4091 pmd_load_cached_ports(pmd
);
4093 ovs_mutex_unlock(&pmd
->port_mutex
);
4095 *ppoll_list
= poll_list
;
4100 pmd_thread_main(void *f_
)
4102 struct dp_netdev_pmd_thread
*pmd
= f_
;
4103 unsigned int lc
= 0;
4104 struct polled_queue
*poll_list
;
4108 int process_packets
= 0;
4112 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
4113 ovsthread_setspecific(pmd
->dp
->per_pmd_key
, pmd
);
4114 ovs_numa_thread_setaffinity_core(pmd
->core_id
);
4115 dpdk_set_lcore_id(pmd
->core_id
);
4116 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
4117 emc_cache_init(&pmd
->flow_cache
);
4119 pmd_alloc_static_tx_qid(pmd
);
4121 /* List port/core affinity */
4122 for (i
= 0; i
< poll_cnt
; i
++) {
4123 VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n",
4124 pmd
->core_id
, netdev_rxq_get_name(poll_list
[i
].rxq
->rx
),
4125 netdev_rxq_get_queue_id(poll_list
[i
].rxq
->rx
));
4129 while (seq_read(pmd
->reload_seq
) == pmd
->last_reload_seq
) {
4130 seq_wait(pmd
->reload_seq
, pmd
->last_reload_seq
);
4136 cycles_count_start(pmd
);
4138 for (i
= 0; i
< poll_cnt
; i
++) {
4140 dp_netdev_process_rxq_port(pmd
, poll_list
[i
].rxq
->rx
,
4141 poll_list
[i
].port_no
);
4142 cycles_count_intermediate(pmd
, poll_list
[i
].rxq
,
4143 process_packets
? PMD_CYCLES_PROCESSING
4152 coverage_try_clear();
4153 /* It's possible that the time was not updated on current
4154 * iteration, if there were no received packets. */
4155 pmd_thread_ctx_time_update(pmd
);
4156 dp_netdev_pmd_try_optimize(pmd
, poll_list
, poll_cnt
);
4157 if (!ovsrcu_try_quiesce()) {
4158 emc_cache_slow_sweep(&pmd
->flow_cache
);
4161 atomic_read_relaxed(&pmd
->reload
, &reload
);
4168 cycles_count_end(pmd
, PMD_CYCLES_IDLE
);
4170 poll_cnt
= pmd_load_queues_and_ports(pmd
, &poll_list
);
4171 exiting
= latch_is_set(&pmd
->exit_latch
);
4172 /* Signal here to make sure the pmd finishes
4173 * reloading the updated configuration. */
4174 dp_netdev_pmd_reload_done(pmd
);
4176 pmd_free_static_tx_qid(pmd
);
4182 emc_cache_uninit(&pmd
->flow_cache
);
4184 pmd_free_cached_ports(pmd
);
4189 dp_netdev_disable_upcall(struct dp_netdev
*dp
)
4190 OVS_ACQUIRES(dp
->upcall_rwlock
)
4192 fat_rwlock_wrlock(&dp
->upcall_rwlock
);
4198 dpif_netdev_meter_get_features(const struct dpif
* dpif OVS_UNUSED
,
4199 struct ofputil_meter_features
*features
)
4201 features
->max_meters
= MAX_METERS
;
4202 features
->band_types
= DP_SUPPORTED_METER_BAND_TYPES
;
4203 features
->capabilities
= DP_SUPPORTED_METER_FLAGS_MASK
;
4204 features
->max_bands
= MAX_BANDS
;
4205 features
->max_color
= 0;
4208 /* Returns false when packet needs to be dropped. */
4210 dp_netdev_run_meter(struct dp_netdev
*dp
, struct dp_packet_batch
*packets_
,
4211 uint32_t meter_id
, long long int now
)
4213 struct dp_meter
*meter
;
4214 struct dp_meter_band
*band
;
4215 struct dp_packet
*packet
;
4216 long long int long_delta_t
; /* msec */
4217 uint32_t delta_t
; /* msec */
4219 const size_t cnt
= dp_packet_batch_size(packets_
);
4220 uint32_t bytes
, volume
;
4221 int exceeded_band
[NETDEV_MAX_BURST
];
4222 uint32_t exceeded_rate
[NETDEV_MAX_BURST
];
4223 int exceeded_pkt
= cnt
; /* First packet that exceeded a band rate. */
4225 if (meter_id
>= MAX_METERS
) {
4229 meter_lock(dp
, meter_id
);
4230 meter
= dp
->meters
[meter_id
];
4235 /* Initialize as negative values. */
4236 memset(exceeded_band
, 0xff, cnt
* sizeof *exceeded_band
);
4237 /* Initialize as zeroes. */
4238 memset(exceeded_rate
, 0, cnt
* sizeof *exceeded_rate
);
4240 /* All packets will hit the meter at the same time. */
4241 long_delta_t
= (now
- meter
->used
); /* msec */
4243 /* Make sure delta_t will not be too large, so that bucket will not
4244 * wrap around below. */
4245 delta_t
= (long_delta_t
> (long long int)meter
->max_delta_t
)
4246 ? meter
->max_delta_t
: (uint32_t)long_delta_t
;
4248 /* Update meter stats. */
4250 meter
->packet_count
+= cnt
;
4252 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
4253 bytes
+= dp_packet_size(packet
);
4255 meter
->byte_count
+= bytes
;
4257 /* Meters can operate in terms of packets per second or kilobits per
4259 if (meter
->flags
& OFPMF13_PKTPS
) {
4260 /* Rate in packets/second, bucket 1/1000 packets. */
4261 /* msec * packets/sec = 1/1000 packets. */
4262 volume
= cnt
* 1000; /* Take 'cnt' packets from the bucket. */
4264 /* Rate in kbps, bucket in bits. */
4265 /* msec * kbps = bits */
4269 /* Update all bands and find the one hit with the highest rate for each
4270 * packet (if any). */
4271 for (int m
= 0; m
< meter
->n_bands
; ++m
) {
4272 band
= &meter
->bands
[m
];
4274 /* Update band's bucket. */
4275 band
->bucket
+= delta_t
* band
->up
.rate
;
4276 if (band
->bucket
> band
->up
.burst_size
) {
4277 band
->bucket
= band
->up
.burst_size
;
4280 /* Drain the bucket for all the packets, if possible. */
4281 if (band
->bucket
>= volume
) {
4282 band
->bucket
-= volume
;
4284 int band_exceeded_pkt
;
4286 /* Band limit hit, must process packet-by-packet. */
4287 if (meter
->flags
& OFPMF13_PKTPS
) {
4288 band_exceeded_pkt
= band
->bucket
/ 1000;
4289 band
->bucket
%= 1000; /* Remainder stays in bucket. */
4291 /* Update the exceeding band for each exceeding packet.
4292 * (Only one band will be fired by a packet, and that
4293 * can be different for each packet.) */
4294 for (i
= band_exceeded_pkt
; i
< cnt
; i
++) {
4295 if (band
->up
.rate
> exceeded_rate
[i
]) {
4296 exceeded_rate
[i
] = band
->up
.rate
;
4297 exceeded_band
[i
] = m
;
4301 /* Packet sizes differ, must process one-by-one. */
4302 band_exceeded_pkt
= cnt
;
4303 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
4304 uint32_t bits
= dp_packet_size(packet
) * 8;
4306 if (band
->bucket
>= bits
) {
4307 band
->bucket
-= bits
;
4309 if (i
< band_exceeded_pkt
) {
4310 band_exceeded_pkt
= i
;
4312 /* Update the exceeding band for the exceeding packet.
4313 * (Only one band will be fired by a packet, and that
4314 * can be different for each packet.) */
4315 if (band
->up
.rate
> exceeded_rate
[i
]) {
4316 exceeded_rate
[i
] = band
->up
.rate
;
4317 exceeded_band
[i
] = m
;
4322 /* Remember the first exceeding packet. */
4323 if (exceeded_pkt
> band_exceeded_pkt
) {
4324 exceeded_pkt
= band_exceeded_pkt
;
4329 /* Fire the highest rate band exceeded by each packet.
4330 * Drop packets if needed, by swapping packet to the end that will be
4333 DP_PACKET_BATCH_REFILL_FOR_EACH (j
, cnt
, packet
, packets_
) {
4334 if (exceeded_band
[j
] >= 0) {
4335 /* Meter drop packet. */
4336 band
= &meter
->bands
[exceeded_band
[j
]];
4337 band
->packet_count
+= 1;
4338 band
->byte_count
+= dp_packet_size(packet
);
4340 dp_packet_delete(packet
);
4342 /* Meter accepts packet. */
4343 dp_packet_batch_refill(packets_
, packet
, j
);
4347 meter_unlock(dp
, meter_id
);
4350 /* Meter set/get/del processing is still single-threaded. */
4352 dpif_netdev_meter_set(struct dpif
*dpif
, ofproto_meter_id
*meter_id
,
4353 struct ofputil_meter_config
*config
)
4355 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4356 uint32_t mid
= meter_id
->uint32
;
4357 struct dp_meter
*meter
;
4360 if (mid
>= MAX_METERS
) {
4361 return EFBIG
; /* Meter_id out of range. */
4364 if (config
->flags
& ~DP_SUPPORTED_METER_FLAGS_MASK
||
4365 !(config
->flags
& (OFPMF13_KBPS
| OFPMF13_PKTPS
))) {
4366 return EBADF
; /* Unsupported flags set */
4369 /* Validate bands */
4370 if (config
->n_bands
== 0 || config
->n_bands
> MAX_BANDS
) {
4371 return EINVAL
; /* Too many bands */
4374 /* Validate rates */
4375 for (i
= 0; i
< config
->n_bands
; i
++) {
4376 if (config
->bands
[i
].rate
== 0) {
4377 return EDOM
; /* rate must be non-zero */
4381 for (i
= 0; i
< config
->n_bands
; ++i
) {
4382 switch (config
->bands
[i
].type
) {
4386 return ENODEV
; /* Unsupported band type */
4390 /* Allocate meter */
4391 meter
= xzalloc(sizeof *meter
4392 + config
->n_bands
* sizeof(struct dp_meter_band
));
4394 meter
->flags
= config
->flags
;
4395 meter
->n_bands
= config
->n_bands
;
4396 meter
->max_delta_t
= 0;
4397 meter
->used
= time_msec();
4400 for (i
= 0; i
< config
->n_bands
; ++i
) {
4401 uint32_t band_max_delta_t
;
4403 /* Set burst size to a workable value if none specified. */
4404 if (config
->bands
[i
].burst_size
== 0) {
4405 config
->bands
[i
].burst_size
= config
->bands
[i
].rate
;
4408 meter
->bands
[i
].up
= config
->bands
[i
];
4409 /* Convert burst size to the bucket units: */
4410 /* pkts => 1/1000 packets, kilobits => bits. */
4411 meter
->bands
[i
].up
.burst_size
*= 1000;
4412 /* Initialize bucket to empty. */
4413 meter
->bands
[i
].bucket
= 0;
4415 /* Figure out max delta_t that is enough to fill any bucket. */
4417 = meter
->bands
[i
].up
.burst_size
/ meter
->bands
[i
].up
.rate
;
4418 if (band_max_delta_t
> meter
->max_delta_t
) {
4419 meter
->max_delta_t
= band_max_delta_t
;
4423 meter_lock(dp
, mid
);
4424 dp_delete_meter(dp
, mid
); /* Free existing meter, if any */
4425 dp
->meters
[mid
] = meter
;
4426 meter_unlock(dp
, mid
);
4434 dpif_netdev_meter_get(const struct dpif
*dpif
,
4435 ofproto_meter_id meter_id_
,
4436 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
4438 const struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4439 const struct dp_meter
*meter
;
4440 uint32_t meter_id
= meter_id_
.uint32
;
4442 if (meter_id
>= MAX_METERS
) {
4445 meter
= dp
->meters
[meter_id
];
4452 meter_lock(dp
, meter_id
);
4453 stats
->packet_in_count
= meter
->packet_count
;
4454 stats
->byte_in_count
= meter
->byte_count
;
4456 for (i
= 0; i
< n_bands
&& i
< meter
->n_bands
; ++i
) {
4457 stats
->bands
[i
].packet_count
= meter
->bands
[i
].packet_count
;
4458 stats
->bands
[i
].byte_count
= meter
->bands
[i
].byte_count
;
4460 meter_unlock(dp
, meter_id
);
4468 dpif_netdev_meter_del(struct dpif
*dpif
,
4469 ofproto_meter_id meter_id_
,
4470 struct ofputil_meter_stats
*stats
, uint16_t n_bands
)
4472 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4475 error
= dpif_netdev_meter_get(dpif
, meter_id_
, stats
, n_bands
);
4477 uint32_t meter_id
= meter_id_
.uint32
;
4479 meter_lock(dp
, meter_id
);
4480 dp_delete_meter(dp
, meter_id
);
4481 meter_unlock(dp
, meter_id
);
4488 dpif_netdev_disable_upcall(struct dpif
*dpif
)
4489 OVS_NO_THREAD_SAFETY_ANALYSIS
4491 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4492 dp_netdev_disable_upcall(dp
);
4496 dp_netdev_enable_upcall(struct dp_netdev
*dp
)
4497 OVS_RELEASES(dp
->upcall_rwlock
)
4499 fat_rwlock_unlock(&dp
->upcall_rwlock
);
4503 dpif_netdev_enable_upcall(struct dpif
*dpif
)
4504 OVS_NO_THREAD_SAFETY_ANALYSIS
4506 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
4507 dp_netdev_enable_upcall(dp
);
4511 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread
*pmd
)
4513 ovs_mutex_lock(&pmd
->cond_mutex
);
4514 atomic_store_relaxed(&pmd
->reload
, false);
4515 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
4516 xpthread_cond_signal(&pmd
->cond
);
4517 ovs_mutex_unlock(&pmd
->cond_mutex
);
4520 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
4521 * the pointer if succeeds, otherwise, NULL (it can return NULL even if
4522 * 'core_id' is NON_PMD_CORE_ID).
4524 * Caller must unrefs the returned reference. */
4525 static struct dp_netdev_pmd_thread
*
4526 dp_netdev_get_pmd(struct dp_netdev
*dp
, unsigned core_id
)
4528 struct dp_netdev_pmd_thread
*pmd
;
4529 const struct cmap_node
*pnode
;
4531 pnode
= cmap_find(&dp
->poll_threads
, hash_int(core_id
, 0));
4535 pmd
= CONTAINER_OF(pnode
, struct dp_netdev_pmd_thread
, node
);
4537 return dp_netdev_pmd_try_ref(pmd
) ? pmd
: NULL
;
4540 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
4542 dp_netdev_set_nonpmd(struct dp_netdev
*dp
)
4543 OVS_REQUIRES(dp
->port_mutex
)
4545 struct dp_netdev_pmd_thread
*non_pmd
;
4547 non_pmd
= xzalloc(sizeof *non_pmd
);
4548 dp_netdev_configure_pmd(non_pmd
, dp
, NON_PMD_CORE_ID
, OVS_NUMA_UNSPEC
);
4551 /* Caller must have valid pointer to 'pmd'. */
4553 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread
*pmd
)
4555 return ovs_refcount_try_ref_rcu(&pmd
->ref_cnt
);
4559 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread
*pmd
)
4561 if (pmd
&& ovs_refcount_unref(&pmd
->ref_cnt
) == 1) {
4562 ovsrcu_postpone(dp_netdev_destroy_pmd
, pmd
);
4566 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
4567 * fails, keeps checking for next node until reaching the end of cmap.
4569 * Caller must unrefs the returned reference. */
4570 static struct dp_netdev_pmd_thread
*
4571 dp_netdev_pmd_get_next(struct dp_netdev
*dp
, struct cmap_position
*pos
)
4573 struct dp_netdev_pmd_thread
*next
;
4576 struct cmap_node
*node
;
4578 node
= cmap_next_position(&dp
->poll_threads
, pos
);
4579 next
= node
? CONTAINER_OF(node
, struct dp_netdev_pmd_thread
, node
)
4581 } while (next
&& !dp_netdev_pmd_try_ref(next
));
4586 /* Configures the 'pmd' based on the input argument. */
4588 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread
*pmd
, struct dp_netdev
*dp
,
4589 unsigned core_id
, int numa_id
)
4592 pmd
->core_id
= core_id
;
4593 pmd
->numa_id
= numa_id
;
4594 pmd
->need_reload
= false;
4596 ovs_refcount_init(&pmd
->ref_cnt
);
4597 latch_init(&pmd
->exit_latch
);
4598 pmd
->reload_seq
= seq_create();
4599 pmd
->last_reload_seq
= seq_read(pmd
->reload_seq
);
4600 atomic_init(&pmd
->reload
, false);
4601 xpthread_cond_init(&pmd
->cond
, NULL
);
4602 ovs_mutex_init(&pmd
->cond_mutex
);
4603 ovs_mutex_init(&pmd
->flow_mutex
);
4604 ovs_mutex_init(&pmd
->port_mutex
);
4605 cmap_init(&pmd
->flow_table
);
4606 cmap_init(&pmd
->classifiers
);
4607 pmd_thread_ctx_time_update(pmd
);
4608 pmd
->next_optimization
= pmd
->ctx
.now
+ DPCLS_OPTIMIZATION_INTERVAL
;
4609 pmd
->rxq_next_cycle_store
= pmd
->ctx
.now
+ PMD_RXQ_INTERVAL_LEN
;
4610 hmap_init(&pmd
->poll_list
);
4611 hmap_init(&pmd
->tx_ports
);
4612 hmap_init(&pmd
->tnl_port_cache
);
4613 hmap_init(&pmd
->send_port_cache
);
4614 /* init the 'flow_cache' since there is no
4615 * actual thread created for NON_PMD_CORE_ID. */
4616 if (core_id
== NON_PMD_CORE_ID
) {
4617 emc_cache_init(&pmd
->flow_cache
);
4618 pmd_alloc_static_tx_qid(pmd
);
4620 cmap_insert(&dp
->poll_threads
, CONST_CAST(struct cmap_node
*, &pmd
->node
),
4621 hash_int(core_id
, 0));
4625 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread
*pmd
)
4629 dp_netdev_pmd_flow_flush(pmd
);
4630 hmap_destroy(&pmd
->send_port_cache
);
4631 hmap_destroy(&pmd
->tnl_port_cache
);
4632 hmap_destroy(&pmd
->tx_ports
);
4633 hmap_destroy(&pmd
->poll_list
);
4634 /* All flows (including their dpcls_rules) have been deleted already */
4635 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
4637 ovsrcu_postpone(free
, cls
);
4639 cmap_destroy(&pmd
->classifiers
);
4640 cmap_destroy(&pmd
->flow_table
);
4641 ovs_mutex_destroy(&pmd
->flow_mutex
);
4642 latch_destroy(&pmd
->exit_latch
);
4643 seq_destroy(pmd
->reload_seq
);
4644 xpthread_cond_destroy(&pmd
->cond
);
4645 ovs_mutex_destroy(&pmd
->cond_mutex
);
4646 ovs_mutex_destroy(&pmd
->port_mutex
);
4650 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
4651 * and unrefs the struct. */
4653 dp_netdev_del_pmd(struct dp_netdev
*dp
, struct dp_netdev_pmd_thread
*pmd
)
4655 /* NON_PMD_CORE_ID doesn't have a thread, so we don't have to synchronize,
4656 * but extra cleanup is necessary */
4657 if (pmd
->core_id
== NON_PMD_CORE_ID
) {
4658 ovs_mutex_lock(&dp
->non_pmd_mutex
);
4659 emc_cache_uninit(&pmd
->flow_cache
);
4660 pmd_free_cached_ports(pmd
);
4661 pmd_free_static_tx_qid(pmd
);
4662 ovs_mutex_unlock(&dp
->non_pmd_mutex
);
4664 latch_set(&pmd
->exit_latch
);
4665 dp_netdev_reload_pmd__(pmd
);
4666 xpthread_join(pmd
->thread
, NULL
);
4669 dp_netdev_pmd_clear_ports(pmd
);
4671 /* Purges the 'pmd''s flows after stopping the thread, but before
4672 * destroying the flows, so that the flow stats can be collected. */
4673 if (dp
->dp_purge_cb
) {
4674 dp
->dp_purge_cb(dp
->dp_purge_aux
, pmd
->core_id
);
4676 cmap_remove(&pmd
->dp
->poll_threads
, &pmd
->node
, hash_int(pmd
->core_id
, 0));
4677 dp_netdev_pmd_unref(pmd
);
4680 /* Destroys all pmd threads. If 'non_pmd' is true it also destroys the non pmd
4683 dp_netdev_destroy_all_pmds(struct dp_netdev
*dp
, bool non_pmd
)
4685 struct dp_netdev_pmd_thread
*pmd
;
4686 struct dp_netdev_pmd_thread
**pmd_list
;
4687 size_t k
= 0, n_pmds
;
4689 n_pmds
= cmap_count(&dp
->poll_threads
);
4690 pmd_list
= xcalloc(n_pmds
, sizeof *pmd_list
);
4692 CMAP_FOR_EACH (pmd
, node
, &dp
->poll_threads
) {
4693 if (!non_pmd
&& pmd
->core_id
== NON_PMD_CORE_ID
) {
4696 /* We cannot call dp_netdev_del_pmd(), since it alters
4697 * 'dp->poll_threads' (while we're iterating it) and it
4699 ovs_assert(k
< n_pmds
);
4700 pmd_list
[k
++] = pmd
;
4703 for (size_t i
= 0; i
< k
; i
++) {
4704 dp_netdev_del_pmd(dp
, pmd_list
[i
]);
4709 /* Deletes all rx queues from pmd->poll_list and all the ports from
4712 dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread
*pmd
)
4714 struct rxq_poll
*poll
;
4715 struct tx_port
*port
;
4717 ovs_mutex_lock(&pmd
->port_mutex
);
4718 HMAP_FOR_EACH_POP (poll
, node
, &pmd
->poll_list
) {
4721 HMAP_FOR_EACH_POP (port
, node
, &pmd
->tx_ports
) {
4724 ovs_mutex_unlock(&pmd
->port_mutex
);
4727 /* Adds rx queue to poll_list of PMD thread, if it's not there already. */
4729 dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
4730 struct dp_netdev_rxq
*rxq
)
4731 OVS_REQUIRES(pmd
->port_mutex
)
4733 int qid
= netdev_rxq_get_queue_id(rxq
->rx
);
4734 uint32_t hash
= hash_2words(odp_to_u32(rxq
->port
->port_no
), qid
);
4735 struct rxq_poll
*poll
;
4737 HMAP_FOR_EACH_WITH_HASH (poll
, node
, hash
, &pmd
->poll_list
) {
4738 if (poll
->rxq
== rxq
) {
4739 /* 'rxq' is already polled by this thread. Do nothing. */
4744 poll
= xmalloc(sizeof *poll
);
4746 hmap_insert(&pmd
->poll_list
, &poll
->node
, hash
);
4748 pmd
->need_reload
= true;
4751 /* Delete 'poll' from poll_list of PMD thread. */
4753 dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
4754 struct rxq_poll
*poll
)
4755 OVS_REQUIRES(pmd
->port_mutex
)
4757 hmap_remove(&pmd
->poll_list
, &poll
->node
);
4760 pmd
->need_reload
= true;
4763 /* Add 'port' to the tx port cache of 'pmd', which must be reloaded for the
4764 * changes to take effect. */
4766 dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread
*pmd
,
4767 struct dp_netdev_port
*port
)
4768 OVS_REQUIRES(pmd
->port_mutex
)
4772 tx
= tx_port_lookup(&pmd
->tx_ports
, port
->port_no
);
4774 /* 'port' is already on this thread tx cache. Do nothing. */
4778 tx
= xzalloc(sizeof *tx
);
4782 dp_packet_batch_init(&tx
->output_pkts
);
4784 hmap_insert(&pmd
->tx_ports
, &tx
->node
, hash_port_no(tx
->port
->port_no
));
4785 pmd
->need_reload
= true;
4788 /* Del 'tx' from the tx port cache of 'pmd', which must be reloaded for the
4789 * changes to take effect. */
4791 dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread
*pmd
,
4793 OVS_REQUIRES(pmd
->port_mutex
)
4795 hmap_remove(&pmd
->tx_ports
, &tx
->node
);
4797 pmd
->need_reload
= true;
4801 dpif_netdev_get_datapath_version(void)
4803 return xstrdup("<built-in>");
4807 dp_netdev_flow_used(struct dp_netdev_flow
*netdev_flow
, int cnt
, int size
,
4808 uint16_t tcp_flags
, long long now
)
4812 atomic_store_relaxed(&netdev_flow
->stats
.used
, now
);
4813 non_atomic_ullong_add(&netdev_flow
->stats
.packet_count
, cnt
);
4814 non_atomic_ullong_add(&netdev_flow
->stats
.byte_count
, size
);
4815 atomic_read_relaxed(&netdev_flow
->stats
.tcp_flags
, &flags
);
4817 atomic_store_relaxed(&netdev_flow
->stats
.tcp_flags
, flags
);
4821 dp_netdev_count_packet(struct dp_netdev_pmd_thread
*pmd
,
4822 enum dp_stat_type type
, int cnt
)
4824 non_atomic_ullong_add(&pmd
->stats
.n
[type
], cnt
);
4828 dp_netdev_upcall(struct dp_netdev_pmd_thread
*pmd
, struct dp_packet
*packet_
,
4829 struct flow
*flow
, struct flow_wildcards
*wc
, ovs_u128
*ufid
,
4830 enum dpif_upcall_type type
, const struct nlattr
*userdata
,
4831 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
)
4833 struct dp_netdev
*dp
= pmd
->dp
;
4835 if (OVS_UNLIKELY(!dp
->upcall_cb
)) {
4839 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl
))) {
4840 struct ds ds
= DS_EMPTY_INITIALIZER
;
4843 struct odp_flow_key_parms odp_parms
= {
4845 .mask
= wc
? &wc
->masks
: NULL
,
4846 .support
= dp_netdev_support
,
4849 ofpbuf_init(&key
, 0);
4850 odp_flow_key_from_flow(&odp_parms
, &key
);
4851 packet_str
= ofp_dp_packet_to_string(packet_
);
4853 odp_flow_key_format(key
.data
, key
.size
, &ds
);
4855 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp
->name
,
4856 dpif_upcall_type_to_string(type
), ds_cstr(&ds
), packet_str
);
4858 ofpbuf_uninit(&key
);
4864 return dp
->upcall_cb(packet_
, flow
, ufid
, pmd
->core_id
, type
, userdata
,
4865 actions
, wc
, put_actions
, dp
->upcall_aux
);
4868 static inline uint32_t
4869 dpif_netdev_packet_get_rss_hash_orig_pkt(struct dp_packet
*packet
,
4870 const struct miniflow
*mf
)
4874 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
4875 hash
= dp_packet_get_rss_hash(packet
);
4877 hash
= miniflow_hash_5tuple(mf
, 0);
4878 dp_packet_set_rss_hash(packet
, hash
);
4884 static inline uint32_t
4885 dpif_netdev_packet_get_rss_hash(struct dp_packet
*packet
,
4886 const struct miniflow
*mf
)
4888 uint32_t hash
, recirc_depth
;
4890 if (OVS_LIKELY(dp_packet_rss_valid(packet
))) {
4891 hash
= dp_packet_get_rss_hash(packet
);
4893 hash
= miniflow_hash_5tuple(mf
, 0);
4894 dp_packet_set_rss_hash(packet
, hash
);
4897 /* The RSS hash must account for the recirculation depth to avoid
4898 * collisions in the exact match cache */
4899 recirc_depth
= *recirc_depth_get_unsafe();
4900 if (OVS_UNLIKELY(recirc_depth
)) {
4901 hash
= hash_finish(hash
, recirc_depth
);
4902 dp_packet_set_rss_hash(packet
, hash
);
4907 struct packet_batch_per_flow
{
4908 unsigned int byte_count
;
4910 struct dp_netdev_flow
*flow
;
4912 struct dp_packet_batch array
;
4916 packet_batch_per_flow_update(struct packet_batch_per_flow
*batch
,
4917 struct dp_packet
*packet
,
4918 const struct miniflow
*mf
)
4920 batch
->byte_count
+= dp_packet_size(packet
);
4921 batch
->tcp_flags
|= miniflow_get_tcp_flags(mf
);
4922 batch
->array
.packets
[batch
->array
.count
++] = packet
;
4926 packet_batch_per_flow_init(struct packet_batch_per_flow
*batch
,
4927 struct dp_netdev_flow
*flow
)
4929 flow
->batch
= batch
;
4932 dp_packet_batch_init(&batch
->array
);
4933 batch
->byte_count
= 0;
4934 batch
->tcp_flags
= 0;
4938 packet_batch_per_flow_execute(struct packet_batch_per_flow
*batch
,
4939 struct dp_netdev_pmd_thread
*pmd
)
4941 struct dp_netdev_actions
*actions
;
4942 struct dp_netdev_flow
*flow
= batch
->flow
;
4944 dp_netdev_flow_used(flow
, batch
->array
.count
, batch
->byte_count
,
4945 batch
->tcp_flags
, pmd
->ctx
.now
);
4947 actions
= dp_netdev_flow_get_actions(flow
);
4949 dp_netdev_execute_actions(pmd
, &batch
->array
, true, &flow
->flow
,
4950 actions
->actions
, actions
->size
);
4954 dp_netdev_queue_batches(struct dp_packet
*pkt
,
4955 struct dp_netdev_flow
*flow
, const struct miniflow
*mf
,
4956 struct packet_batch_per_flow
*batches
,
4959 struct packet_batch_per_flow
*batch
= flow
->batch
;
4961 if (OVS_UNLIKELY(!batch
)) {
4962 batch
= &batches
[(*n_batches
)++];
4963 packet_batch_per_flow_init(batch
, flow
);
4966 packet_batch_per_flow_update(batch
, pkt
, mf
);
4969 /* Try to process all ('cnt') the 'packets' using only the exact match cache
4970 * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the
4971 * miniflow is copied into 'keys' and the packet pointer is moved at the
4972 * beginning of the 'packets' array.
4974 * The function returns the number of packets that needs to be processed in the
4975 * 'packets' array (they have been moved to the beginning of the vector).
4977 * For performance reasons a caller may choose not to initialize the metadata
4978 * in 'packets_'. If 'md_is_valid' is false, the metadata in 'packets'
4979 * is not valid and must be initialized by this function using 'port_no'.
4980 * If 'md_is_valid' is true, the metadata is already valid and 'port_no'
4983 static inline size_t
4984 emc_processing(struct dp_netdev_pmd_thread
*pmd
,
4985 struct dp_packet_batch
*packets_
,
4986 struct netdev_flow_key
*keys
,
4987 struct packet_batch_per_flow batches
[], size_t *n_batches
,
4988 bool md_is_valid
, odp_port_t port_no
)
4990 struct emc_cache
*flow_cache
= &pmd
->flow_cache
;
4991 struct netdev_flow_key
*key
= &keys
[0];
4992 size_t n_missed
= 0, n_dropped
= 0;
4993 struct dp_packet
*packet
;
4994 const size_t cnt
= dp_packet_batch_size(packets_
);
4998 atomic_read_relaxed(&pmd
->dp
->emc_insert_min
, &cur_min
);
5000 DP_PACKET_BATCH_REFILL_FOR_EACH (i
, cnt
, packet
, packets_
) {
5001 struct dp_netdev_flow
*flow
;
5003 if (OVS_UNLIKELY(dp_packet_size(packet
) < ETH_HEADER_LEN
)) {
5004 dp_packet_delete(packet
);
5010 struct dp_packet
**packets
= packets_
->packets
;
5011 /* Prefetch next packet data and metadata. */
5012 OVS_PREFETCH(dp_packet_data(packets
[i
+1]));
5013 pkt_metadata_prefetch_init(&packets
[i
+1]->md
);
5017 pkt_metadata_init(&packet
->md
, port_no
);
5019 miniflow_extract(packet
, &key
->mf
);
5020 key
->len
= 0; /* Not computed yet. */
5021 /* If EMC is disabled skip hash computation and emc_lookup */
5024 key
->hash
= dpif_netdev_packet_get_rss_hash_orig_pkt(packet
,
5027 key
->hash
= dpif_netdev_packet_get_rss_hash(packet
, &key
->mf
);
5029 flow
= emc_lookup(flow_cache
, key
);
5033 if (OVS_LIKELY(flow
)) {
5034 dp_netdev_queue_batches(packet
, flow
, &key
->mf
, batches
,
5037 /* Exact match cache missed. Group missed packets together at
5038 * the beginning of the 'packets' array. */
5039 dp_packet_batch_refill(packets_
, packet
, i
);
5040 /* 'key[n_missed]' contains the key of the current packet and it
5041 * must be returned to the caller. The next key should be extracted
5042 * to 'keys[n_missed + 1]'. */
5043 key
= &keys
[++n_missed
];
5047 dp_netdev_count_packet(pmd
, DP_STAT_EXACT_HIT
,
5048 cnt
- n_dropped
- n_missed
);
5050 return dp_packet_batch_size(packets_
);
5054 handle_packet_upcall(struct dp_netdev_pmd_thread
*pmd
,
5055 struct dp_packet
*packet
,
5056 const struct netdev_flow_key
*key
,
5057 struct ofpbuf
*actions
, struct ofpbuf
*put_actions
,
5060 struct ofpbuf
*add_actions
;
5061 struct dp_packet_batch b
;
5066 match
.tun_md
.valid
= false;
5067 miniflow_expand(&key
->mf
, &match
.flow
);
5069 ofpbuf_clear(actions
);
5070 ofpbuf_clear(put_actions
);
5072 dpif_flow_hash(pmd
->dp
->dpif
, &match
.flow
, sizeof match
.flow
, &ufid
);
5073 error
= dp_netdev_upcall(pmd
, packet
, &match
.flow
, &match
.wc
,
5074 &ufid
, DPIF_UC_MISS
, NULL
, actions
,
5076 if (OVS_UNLIKELY(error
&& error
!= ENOSPC
)) {
5077 dp_packet_delete(packet
);
5082 /* The Netlink encoding of datapath flow keys cannot express
5083 * wildcarding the presence of a VLAN tag. Instead, a missing VLAN
5084 * tag is interpreted as exact match on the fact that there is no
5085 * VLAN. Unless we refactor a lot of code that translates between
5086 * Netlink and struct flow representations, we have to do the same
5088 if (!match
.wc
.masks
.vlans
[0].tci
) {
5089 match
.wc
.masks
.vlans
[0].tci
= htons(0xffff);
5092 /* We can't allow the packet batching in the next loop to execute
5093 * the actions. Otherwise, if there are any slow path actions,
5094 * we'll send the packet up twice. */
5095 dp_packet_batch_init_packet(&b
, packet
);
5096 dp_netdev_execute_actions(pmd
, &b
, true, &match
.flow
,
5097 actions
->data
, actions
->size
);
5099 add_actions
= put_actions
->size
? put_actions
: actions
;
5100 if (OVS_LIKELY(error
!= ENOSPC
)) {
5101 struct dp_netdev_flow
*netdev_flow
;
5103 /* XXX: There's a race window where a flow covering this packet
5104 * could have already been installed since we last did the flow
5105 * lookup before upcall. This could be solved by moving the
5106 * mutex lock outside the loop, but that's an awful long time
5107 * to be locking everyone out of making flow installs. If we
5108 * move to a per-core classifier, it would be reasonable. */
5109 ovs_mutex_lock(&pmd
->flow_mutex
);
5110 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, key
, NULL
);
5111 if (OVS_LIKELY(!netdev_flow
)) {
5112 netdev_flow
= dp_netdev_flow_add(pmd
, &match
, &ufid
,
5116 ovs_mutex_unlock(&pmd
->flow_mutex
);
5117 emc_probabilistic_insert(pmd
, key
, netdev_flow
);
5122 fast_path_processing(struct dp_netdev_pmd_thread
*pmd
,
5123 struct dp_packet_batch
*packets_
,
5124 struct netdev_flow_key
*keys
,
5125 struct packet_batch_per_flow batches
[],
5129 const size_t cnt
= dp_packet_batch_size(packets_
);
5130 #if !defined(__CHECKER__) && !defined(_WIN32)
5131 const size_t PKT_ARRAY_SIZE
= cnt
;
5133 /* Sparse or MSVC doesn't like variable length array. */
5134 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
5136 struct dp_packet
*packet
;
5138 struct dpcls_rule
*rules
[PKT_ARRAY_SIZE
];
5139 struct dp_netdev
*dp
= pmd
->dp
;
5140 int miss_cnt
= 0, lost_cnt
= 0;
5141 int lookup_cnt
= 0, add_lookup_cnt
;
5145 for (i
= 0; i
< cnt
; i
++) {
5146 /* Key length is needed in all the cases, hash computed on demand. */
5147 keys
[i
].len
= netdev_flow_key_size(miniflow_n_values(&keys
[i
].mf
));
5149 /* Get the classifier for the in_port */
5150 cls
= dp_netdev_pmd_lookup_dpcls(pmd
, in_port
);
5151 if (OVS_LIKELY(cls
)) {
5152 any_miss
= !dpcls_lookup(cls
, keys
, rules
, cnt
, &lookup_cnt
);
5155 memset(rules
, 0, sizeof(rules
));
5157 if (OVS_UNLIKELY(any_miss
) && !fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
5158 uint64_t actions_stub
[512 / 8], slow_stub
[512 / 8];
5159 struct ofpbuf actions
, put_actions
;
5161 ofpbuf_use_stub(&actions
, actions_stub
, sizeof actions_stub
);
5162 ofpbuf_use_stub(&put_actions
, slow_stub
, sizeof slow_stub
);
5164 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5165 struct dp_netdev_flow
*netdev_flow
;
5167 if (OVS_LIKELY(rules
[i
])) {
5171 /* It's possible that an earlier slow path execution installed
5172 * a rule covering this flow. In this case, it's a lot cheaper
5173 * to catch it here than execute a miss. */
5174 netdev_flow
= dp_netdev_pmd_lookup_flow(pmd
, &keys
[i
],
5177 lookup_cnt
+= add_lookup_cnt
;
5178 rules
[i
] = &netdev_flow
->cr
;
5183 handle_packet_upcall(pmd
, packet
, &keys
[i
], &actions
,
5184 &put_actions
, &lost_cnt
);
5187 ofpbuf_uninit(&actions
);
5188 ofpbuf_uninit(&put_actions
);
5189 fat_rwlock_unlock(&dp
->upcall_rwlock
);
5190 } else if (OVS_UNLIKELY(any_miss
)) {
5191 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5192 if (OVS_UNLIKELY(!rules
[i
])) {
5193 dp_packet_delete(packet
);
5200 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5201 struct dp_netdev_flow
*flow
;
5203 if (OVS_UNLIKELY(!rules
[i
])) {
5207 flow
= dp_netdev_flow_cast(rules
[i
]);
5209 emc_probabilistic_insert(pmd
, &keys
[i
], flow
);
5210 dp_netdev_queue_batches(packet
, flow
, &keys
[i
].mf
, batches
, n_batches
);
5213 dp_netdev_count_packet(pmd
, DP_STAT_MASKED_HIT
, cnt
- miss_cnt
);
5214 dp_netdev_count_packet(pmd
, DP_STAT_LOOKUP_HIT
, lookup_cnt
);
5215 dp_netdev_count_packet(pmd
, DP_STAT_MISS
, miss_cnt
);
5216 dp_netdev_count_packet(pmd
, DP_STAT_LOST
, lost_cnt
);
5219 /* Packets enter the datapath from a port (or from recirculation) here.
5221 * When 'md_is_valid' is true the metadata in 'packets' are already valid.
5222 * When false the metadata in 'packets' need to be initialized. */
5224 dp_netdev_input__(struct dp_netdev_pmd_thread
*pmd
,
5225 struct dp_packet_batch
*packets
,
5226 bool md_is_valid
, odp_port_t port_no
)
5228 #if !defined(__CHECKER__) && !defined(_WIN32)
5229 const size_t PKT_ARRAY_SIZE
= dp_packet_batch_size(packets
);
5231 /* Sparse or MSVC doesn't like variable length array. */
5232 enum { PKT_ARRAY_SIZE
= NETDEV_MAX_BURST
};
5234 OVS_ALIGNED_VAR(CACHE_LINE_SIZE
)
5235 struct netdev_flow_key keys
[PKT_ARRAY_SIZE
];
5236 struct packet_batch_per_flow batches
[PKT_ARRAY_SIZE
];
5241 emc_processing(pmd
, packets
, keys
, batches
, &n_batches
,
5242 md_is_valid
, port_no
);
5243 if (!dp_packet_batch_is_empty(packets
)) {
5244 /* Get ingress port from first packet's metadata. */
5245 in_port
= packets
->packets
[0]->md
.in_port
.odp_port
;
5246 fast_path_processing(pmd
, packets
, keys
,
5247 batches
, &n_batches
, in_port
);
5250 /* All the flow batches need to be reset before any call to
5251 * packet_batch_per_flow_execute() as it could potentially trigger
5252 * recirculation. When a packet matching flow ‘j’ happens to be
5253 * recirculated, the nested call to dp_netdev_input__() could potentially
5254 * classify the packet as matching another flow - say 'k'. It could happen
5255 * that in the previous call to dp_netdev_input__() that same flow 'k' had
5256 * already its own batches[k] still waiting to be served. So if its
5257 * ‘batch’ member is not reset, the recirculated packet would be wrongly
5258 * appended to batches[k] of the 1st call to dp_netdev_input__(). */
5260 for (i
= 0; i
< n_batches
; i
++) {
5261 batches
[i
].flow
->batch
= NULL
;
5264 for (i
= 0; i
< n_batches
; i
++) {
5265 packet_batch_per_flow_execute(&batches
[i
], pmd
);
5270 dp_netdev_input(struct dp_netdev_pmd_thread
*pmd
,
5271 struct dp_packet_batch
*packets
,
5274 dp_netdev_input__(pmd
, packets
, false, port_no
);
5278 dp_netdev_recirculate(struct dp_netdev_pmd_thread
*pmd
,
5279 struct dp_packet_batch
*packets
)
5281 dp_netdev_input__(pmd
, packets
, true, 0);
5284 struct dp_netdev_execute_aux
{
5285 struct dp_netdev_pmd_thread
*pmd
;
5286 const struct flow
*flow
;
5290 dpif_netdev_register_dp_purge_cb(struct dpif
*dpif
, dp_purge_callback
*cb
,
5293 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5294 dp
->dp_purge_aux
= aux
;
5295 dp
->dp_purge_cb
= cb
;
5299 dpif_netdev_register_upcall_cb(struct dpif
*dpif
, upcall_callback
*cb
,
5302 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5303 dp
->upcall_aux
= aux
;
5308 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread
*pmd
,
5312 struct dp_netdev_port
*port
;
5315 HMAP_FOR_EACH (tx
, node
, &pmd
->send_port_cache
) {
5316 if (!tx
->port
->dynamic_txqs
) {
5319 interval
= pmd
->ctx
.now
- tx
->last_used
;
5320 if (tx
->qid
>= 0 && (purge
|| interval
>= XPS_TIMEOUT_MS
)) {
5322 ovs_mutex_lock(&port
->txq_used_mutex
);
5323 port
->txq_used
[tx
->qid
]--;
5324 ovs_mutex_unlock(&port
->txq_used_mutex
);
5331 dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread
*pmd
,
5334 struct dp_netdev_port
*port
;
5336 int i
, min_cnt
, min_qid
;
5338 interval
= pmd
->ctx
.now
- tx
->last_used
;
5339 tx
->last_used
= pmd
->ctx
.now
;
5341 if (OVS_LIKELY(tx
->qid
>= 0 && interval
< XPS_TIMEOUT_MS
)) {
5347 ovs_mutex_lock(&port
->txq_used_mutex
);
5349 port
->txq_used
[tx
->qid
]--;
5355 for (i
= 0; i
< netdev_n_txq(port
->netdev
); i
++) {
5356 if (port
->txq_used
[i
] < min_cnt
|| min_cnt
== -1) {
5357 min_cnt
= port
->txq_used
[i
];
5362 port
->txq_used
[min_qid
]++;
5365 ovs_mutex_unlock(&port
->txq_used_mutex
);
5367 dpif_netdev_xps_revalidate_pmd(pmd
, false);
5369 VLOG_DBG("Core %d: New TX queue ID %d for port \'%s\'.",
5370 pmd
->core_id
, tx
->qid
, netdev_get_name(tx
->port
->netdev
));
5374 static struct tx_port
*
5375 pmd_tnl_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
5378 return tx_port_lookup(&pmd
->tnl_port_cache
, port_no
);
5381 static struct tx_port
*
5382 pmd_send_port_cache_lookup(const struct dp_netdev_pmd_thread
*pmd
,
5385 return tx_port_lookup(&pmd
->send_port_cache
, port_no
);
5389 push_tnl_action(const struct dp_netdev_pmd_thread
*pmd
,
5390 const struct nlattr
*attr
,
5391 struct dp_packet_batch
*batch
)
5393 struct tx_port
*tun_port
;
5394 const struct ovs_action_push_tnl
*data
;
5397 data
= nl_attr_get(attr
);
5399 tun_port
= pmd_tnl_port_cache_lookup(pmd
, data
->tnl_port
);
5404 err
= netdev_push_header(tun_port
->port
->netdev
, batch
, data
);
5409 dp_packet_delete_batch(batch
, true);
5414 dp_execute_userspace_action(struct dp_netdev_pmd_thread
*pmd
,
5415 struct dp_packet
*packet
, bool may_steal
,
5416 struct flow
*flow
, ovs_u128
*ufid
,
5417 struct ofpbuf
*actions
,
5418 const struct nlattr
*userdata
)
5420 struct dp_packet_batch b
;
5423 ofpbuf_clear(actions
);
5425 error
= dp_netdev_upcall(pmd
, packet
, flow
, NULL
, ufid
,
5426 DPIF_UC_ACTION
, userdata
, actions
,
5428 if (!error
|| error
== ENOSPC
) {
5429 dp_packet_batch_init_packet(&b
, packet
);
5430 dp_netdev_execute_actions(pmd
, &b
, may_steal
, flow
,
5431 actions
->data
, actions
->size
);
5432 } else if (may_steal
) {
5433 dp_packet_delete(packet
);
5438 dp_execute_cb(void *aux_
, struct dp_packet_batch
*packets_
,
5439 const struct nlattr
*a
, bool may_steal
)
5440 OVS_NO_THREAD_SAFETY_ANALYSIS
5442 struct dp_netdev_execute_aux
*aux
= aux_
;
5443 uint32_t *depth
= recirc_depth_get();
5444 struct dp_netdev_pmd_thread
*pmd
= aux
->pmd
;
5445 struct dp_netdev
*dp
= pmd
->dp
;
5446 int type
= nl_attr_type(a
);
5449 switch ((enum ovs_action_attr
)type
) {
5450 case OVS_ACTION_ATTR_OUTPUT
:
5451 p
= pmd_send_port_cache_lookup(pmd
, nl_attr_get_odp_port(a
));
5452 if (OVS_LIKELY(p
)) {
5453 struct dp_packet
*packet
;
5454 struct dp_packet_batch out
;
5457 dp_packet_batch_clone(&out
, packets_
);
5458 dp_packet_batch_reset_cutlen(packets_
);
5461 dp_packet_batch_apply_cutlen(packets_
);
5464 if (OVS_UNLIKELY(!dp_packet_batch_is_empty(&p
->output_pkts
)
5465 && packets_
->packets
[0]->source
5466 != p
->output_pkts
.packets
[0]->source
)) {
5467 /* XXX: netdev-dpdk assumes that all packets in a single
5468 * output batch has the same source. Flush here to
5469 * avoid memory access issues. */
5470 dp_netdev_pmd_flush_output_on_port(pmd
, p
);
5473 if (OVS_UNLIKELY(dp_packet_batch_size(&p
->output_pkts
)
5474 + dp_packet_batch_size(packets_
) > NETDEV_MAX_BURST
)) {
5475 /* Some packets was generated while input batch processing.
5476 * Flush here to avoid overflow. */
5477 dp_netdev_pmd_flush_output_on_port(pmd
, p
);
5479 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5480 dp_packet_batch_add(&p
->output_pkts
, packet
);
5486 case OVS_ACTION_ATTR_TUNNEL_PUSH
:
5487 if (*depth
< MAX_RECIRC_DEPTH
) {
5488 dp_packet_batch_apply_cutlen(packets_
);
5489 push_tnl_action(pmd
, a
, packets_
);
5494 case OVS_ACTION_ATTR_TUNNEL_POP
:
5495 if (*depth
< MAX_RECIRC_DEPTH
) {
5496 struct dp_packet_batch
*orig_packets_
= packets_
;
5497 odp_port_t portno
= nl_attr_get_odp_port(a
);
5499 p
= pmd_tnl_port_cache_lookup(pmd
, portno
);
5501 struct dp_packet_batch tnl_pkt
;
5504 dp_packet_batch_clone(&tnl_pkt
, packets_
);
5505 packets_
= &tnl_pkt
;
5506 dp_packet_batch_reset_cutlen(orig_packets_
);
5509 dp_packet_batch_apply_cutlen(packets_
);
5511 netdev_pop_header(p
->port
->netdev
, packets_
);
5512 if (dp_packet_batch_is_empty(packets_
)) {
5516 struct dp_packet
*packet
;
5517 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5518 packet
->md
.in_port
.odp_port
= portno
;
5522 dp_netdev_recirculate(pmd
, packets_
);
5529 case OVS_ACTION_ATTR_USERSPACE
:
5530 if (!fat_rwlock_tryrdlock(&dp
->upcall_rwlock
)) {
5531 struct dp_packet_batch
*orig_packets_
= packets_
;
5532 const struct nlattr
*userdata
;
5533 struct dp_packet_batch usr_pkt
;
5534 struct ofpbuf actions
;
5539 userdata
= nl_attr_find_nested(a
, OVS_USERSPACE_ATTR_USERDATA
);
5540 ofpbuf_init(&actions
, 0);
5542 if (packets_
->trunc
) {
5544 dp_packet_batch_clone(&usr_pkt
, packets_
);
5545 packets_
= &usr_pkt
;
5547 dp_packet_batch_reset_cutlen(orig_packets_
);
5550 dp_packet_batch_apply_cutlen(packets_
);
5553 struct dp_packet
*packet
;
5554 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5555 flow_extract(packet
, &flow
);
5556 dpif_flow_hash(dp
->dpif
, &flow
, sizeof flow
, &ufid
);
5557 dp_execute_userspace_action(pmd
, packet
, may_steal
, &flow
,
5558 &ufid
, &actions
, userdata
);
5562 dp_packet_delete_batch(packets_
, true);
5565 ofpbuf_uninit(&actions
);
5566 fat_rwlock_unlock(&dp
->upcall_rwlock
);
5572 case OVS_ACTION_ATTR_RECIRC
:
5573 if (*depth
< MAX_RECIRC_DEPTH
) {
5574 struct dp_packet_batch recirc_pkts
;
5577 dp_packet_batch_clone(&recirc_pkts
, packets_
);
5578 packets_
= &recirc_pkts
;
5581 struct dp_packet
*packet
;
5582 DP_PACKET_BATCH_FOR_EACH (packet
, packets_
) {
5583 packet
->md
.recirc_id
= nl_attr_get_u32(a
);
5587 dp_netdev_recirculate(pmd
, packets_
);
5593 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
5596 case OVS_ACTION_ATTR_CT
: {
5597 const struct nlattr
*b
;
5599 bool commit
= false;
5602 const char *helper
= NULL
;
5603 const uint32_t *setmark
= NULL
;
5604 const struct ovs_key_ct_labels
*setlabel
= NULL
;
5605 struct nat_action_info_t nat_action_info
;
5606 struct nat_action_info_t
*nat_action_info_ref
= NULL
;
5607 bool nat_config
= false;
5609 NL_ATTR_FOR_EACH_UNSAFE (b
, left
, nl_attr_get(a
),
5610 nl_attr_get_size(a
)) {
5611 enum ovs_ct_attr sub_type
= nl_attr_type(b
);
5614 case OVS_CT_ATTR_FORCE_COMMIT
:
5617 case OVS_CT_ATTR_COMMIT
:
5620 case OVS_CT_ATTR_ZONE
:
5621 zone
= nl_attr_get_u16(b
);
5623 case OVS_CT_ATTR_HELPER
:
5624 helper
= nl_attr_get_string(b
);
5626 case OVS_CT_ATTR_MARK
:
5627 setmark
= nl_attr_get(b
);
5629 case OVS_CT_ATTR_LABELS
:
5630 setlabel
= nl_attr_get(b
);
5632 case OVS_CT_ATTR_EVENTMASK
:
5633 /* Silently ignored, as userspace datapath does not generate
5634 * netlink events. */
5636 case OVS_CT_ATTR_NAT
: {
5637 const struct nlattr
*b_nest
;
5638 unsigned int left_nest
;
5639 bool ip_min_specified
= false;
5640 bool proto_num_min_specified
= false;
5641 bool ip_max_specified
= false;
5642 bool proto_num_max_specified
= false;
5643 memset(&nat_action_info
, 0, sizeof nat_action_info
);
5644 nat_action_info_ref
= &nat_action_info
;
5646 NL_NESTED_FOR_EACH_UNSAFE (b_nest
, left_nest
, b
) {
5647 enum ovs_nat_attr sub_type_nest
= nl_attr_type(b_nest
);
5649 switch (sub_type_nest
) {
5650 case OVS_NAT_ATTR_SRC
:
5651 case OVS_NAT_ATTR_DST
:
5653 nat_action_info
.nat_action
|=
5654 ((sub_type_nest
== OVS_NAT_ATTR_SRC
)
5655 ? NAT_ACTION_SRC
: NAT_ACTION_DST
);
5657 case OVS_NAT_ATTR_IP_MIN
:
5658 memcpy(&nat_action_info
.min_addr
,
5659 nl_attr_get(b_nest
),
5660 nl_attr_get_size(b_nest
));
5661 ip_min_specified
= true;
5663 case OVS_NAT_ATTR_IP_MAX
:
5664 memcpy(&nat_action_info
.max_addr
,
5665 nl_attr_get(b_nest
),
5666 nl_attr_get_size(b_nest
));
5667 ip_max_specified
= true;
5669 case OVS_NAT_ATTR_PROTO_MIN
:
5670 nat_action_info
.min_port
=
5671 nl_attr_get_u16(b_nest
);
5672 proto_num_min_specified
= true;
5674 case OVS_NAT_ATTR_PROTO_MAX
:
5675 nat_action_info
.max_port
=
5676 nl_attr_get_u16(b_nest
);
5677 proto_num_max_specified
= true;
5679 case OVS_NAT_ATTR_PERSISTENT
:
5680 case OVS_NAT_ATTR_PROTO_HASH
:
5681 case OVS_NAT_ATTR_PROTO_RANDOM
:
5683 case OVS_NAT_ATTR_UNSPEC
:
5684 case __OVS_NAT_ATTR_MAX
:
5689 if (ip_min_specified
&& !ip_max_specified
) {
5690 nat_action_info
.max_addr
= nat_action_info
.min_addr
;
5692 if (proto_num_min_specified
&& !proto_num_max_specified
) {
5693 nat_action_info
.max_port
= nat_action_info
.min_port
;
5695 if (proto_num_min_specified
|| proto_num_max_specified
) {
5696 if (nat_action_info
.nat_action
& NAT_ACTION_SRC
) {
5697 nat_action_info
.nat_action
|= NAT_ACTION_SRC_PORT
;
5698 } else if (nat_action_info
.nat_action
& NAT_ACTION_DST
) {
5699 nat_action_info
.nat_action
|= NAT_ACTION_DST_PORT
;
5704 case OVS_CT_ATTR_UNSPEC
:
5705 case __OVS_CT_ATTR_MAX
:
5710 /* We won't be able to function properly in this case, hence
5711 * complain loudly. */
5712 if (nat_config
&& !commit
) {
5713 static struct vlog_rate_limit rl
= VLOG_RATE_LIMIT_INIT(5, 5);
5714 VLOG_WARN_RL(&rl
, "NAT specified without commit.");
5717 conntrack_execute(&dp
->conntrack
, packets_
, aux
->flow
->dl_type
, force
,
5718 commit
, zone
, setmark
, setlabel
, aux
->flow
->tp_src
,
5719 aux
->flow
->tp_dst
, helper
, nat_action_info_ref
,
5724 case OVS_ACTION_ATTR_METER
:
5725 dp_netdev_run_meter(pmd
->dp
, packets_
, nl_attr_get_u32(a
),
5729 case OVS_ACTION_ATTR_PUSH_VLAN
:
5730 case OVS_ACTION_ATTR_POP_VLAN
:
5731 case OVS_ACTION_ATTR_PUSH_MPLS
:
5732 case OVS_ACTION_ATTR_POP_MPLS
:
5733 case OVS_ACTION_ATTR_SET
:
5734 case OVS_ACTION_ATTR_SET_MASKED
:
5735 case OVS_ACTION_ATTR_SAMPLE
:
5736 case OVS_ACTION_ATTR_HASH
:
5737 case OVS_ACTION_ATTR_UNSPEC
:
5738 case OVS_ACTION_ATTR_TRUNC
:
5739 case OVS_ACTION_ATTR_PUSH_ETH
:
5740 case OVS_ACTION_ATTR_POP_ETH
:
5741 case OVS_ACTION_ATTR_CLONE
:
5742 case OVS_ACTION_ATTR_ENCAP_NSH
:
5743 case OVS_ACTION_ATTR_DECAP_NSH
:
5744 case __OVS_ACTION_ATTR_MAX
:
5748 dp_packet_delete_batch(packets_
, may_steal
);
5752 dp_netdev_execute_actions(struct dp_netdev_pmd_thread
*pmd
,
5753 struct dp_packet_batch
*packets
,
5754 bool may_steal
, const struct flow
*flow
,
5755 const struct nlattr
*actions
, size_t actions_len
)
5757 struct dp_netdev_execute_aux aux
= { pmd
, flow
};
5759 odp_execute_actions(&aux
, packets
, may_steal
, actions
,
5760 actions_len
, dp_execute_cb
);
5763 struct dp_netdev_ct_dump
{
5764 struct ct_dpif_dump_state up
;
5765 struct conntrack_dump dump
;
5766 struct conntrack
*ct
;
5767 struct dp_netdev
*dp
;
5771 dpif_netdev_ct_dump_start(struct dpif
*dpif
, struct ct_dpif_dump_state
**dump_
,
5772 const uint16_t *pzone
, int *ptot_bkts
)
5774 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5775 struct dp_netdev_ct_dump
*dump
;
5777 dump
= xzalloc(sizeof *dump
);
5779 dump
->ct
= &dp
->conntrack
;
5781 conntrack_dump_start(&dp
->conntrack
, &dump
->dump
, pzone
, ptot_bkts
);
5789 dpif_netdev_ct_dump_next(struct dpif
*dpif OVS_UNUSED
,
5790 struct ct_dpif_dump_state
*dump_
,
5791 struct ct_dpif_entry
*entry
)
5793 struct dp_netdev_ct_dump
*dump
;
5795 INIT_CONTAINER(dump
, dump_
, up
);
5797 return conntrack_dump_next(&dump
->dump
, entry
);
5801 dpif_netdev_ct_dump_done(struct dpif
*dpif OVS_UNUSED
,
5802 struct ct_dpif_dump_state
*dump_
)
5804 struct dp_netdev_ct_dump
*dump
;
5807 INIT_CONTAINER(dump
, dump_
, up
);
5809 err
= conntrack_dump_done(&dump
->dump
);
5817 dpif_netdev_ct_flush(struct dpif
*dpif
, const uint16_t *zone
,
5818 const struct ct_dpif_tuple
*tuple
)
5820 struct dp_netdev
*dp
= get_dp_netdev(dpif
);
5825 return conntrack_flush(&dp
->conntrack
, zone
);
5828 const struct dpif_class dpif_netdev_class
= {
5831 dpif_netdev_enumerate
,
5832 dpif_netdev_port_open_type
,
5835 dpif_netdev_destroy
,
5838 dpif_netdev_get_stats
,
5839 dpif_netdev_port_add
,
5840 dpif_netdev_port_del
,
5841 dpif_netdev_port_set_config
,
5842 dpif_netdev_port_query_by_number
,
5843 dpif_netdev_port_query_by_name
,
5844 NULL
, /* port_get_pid */
5845 dpif_netdev_port_dump_start
,
5846 dpif_netdev_port_dump_next
,
5847 dpif_netdev_port_dump_done
,
5848 dpif_netdev_port_poll
,
5849 dpif_netdev_port_poll_wait
,
5850 dpif_netdev_flow_flush
,
5851 dpif_netdev_flow_dump_create
,
5852 dpif_netdev_flow_dump_destroy
,
5853 dpif_netdev_flow_dump_thread_create
,
5854 dpif_netdev_flow_dump_thread_destroy
,
5855 dpif_netdev_flow_dump_next
,
5856 dpif_netdev_operate
,
5857 NULL
, /* recv_set */
5858 NULL
, /* handlers_set */
5859 dpif_netdev_set_config
,
5860 dpif_netdev_queue_to_priority
,
5862 NULL
, /* recv_wait */
5863 NULL
, /* recv_purge */
5864 dpif_netdev_register_dp_purge_cb
,
5865 dpif_netdev_register_upcall_cb
,
5866 dpif_netdev_enable_upcall
,
5867 dpif_netdev_disable_upcall
,
5868 dpif_netdev_get_datapath_version
,
5869 dpif_netdev_ct_dump_start
,
5870 dpif_netdev_ct_dump_next
,
5871 dpif_netdev_ct_dump_done
,
5872 dpif_netdev_ct_flush
,
5873 dpif_netdev_meter_get_features
,
5874 dpif_netdev_meter_set
,
5875 dpif_netdev_meter_get
,
5876 dpif_netdev_meter_del
,
5880 dpif_dummy_change_port_number(struct unixctl_conn
*conn
, int argc OVS_UNUSED
,
5881 const char *argv
[], void *aux OVS_UNUSED
)
5883 struct dp_netdev_port
*port
;
5884 struct dp_netdev
*dp
;
5887 ovs_mutex_lock(&dp_netdev_mutex
);
5888 dp
= shash_find_data(&dp_netdevs
, argv
[1]);
5889 if (!dp
|| !dpif_netdev_class_is_dummy(dp
->class)) {
5890 ovs_mutex_unlock(&dp_netdev_mutex
);
5891 unixctl_command_reply_error(conn
, "unknown datapath or not a dummy");
5894 ovs_refcount_ref(&dp
->ref_cnt
);
5895 ovs_mutex_unlock(&dp_netdev_mutex
);
5897 ovs_mutex_lock(&dp
->port_mutex
);
5898 if (get_port_by_name(dp
, argv
[2], &port
)) {
5899 unixctl_command_reply_error(conn
, "unknown port");
5903 port_no
= u32_to_odp(atoi(argv
[3]));
5904 if (!port_no
|| port_no
== ODPP_NONE
) {
5905 unixctl_command_reply_error(conn
, "bad port number");
5908 if (dp_netdev_lookup_port(dp
, port_no
)) {
5909 unixctl_command_reply_error(conn
, "port number already in use");
5914 hmap_remove(&dp
->ports
, &port
->node
);
5915 reconfigure_datapath(dp
);
5917 /* Reinsert with new port number. */
5918 port
->port_no
= port_no
;
5919 hmap_insert(&dp
->ports
, &port
->node
, hash_port_no(port_no
));
5920 reconfigure_datapath(dp
);
5922 seq_change(dp
->port_seq
);
5923 unixctl_command_reply(conn
, NULL
);
5926 ovs_mutex_unlock(&dp
->port_mutex
);
5927 dp_netdev_unref(dp
);
5931 dpif_dummy_register__(const char *type
)
5933 struct dpif_class
*class;
5935 class = xmalloc(sizeof *class);
5936 *class = dpif_netdev_class
;
5937 class->type
= xstrdup(type
);
5938 dp_register_provider(class);
5942 dpif_dummy_override(const char *type
)
5947 * Ignore EAFNOSUPPORT to allow --enable-dummy=system with
5948 * a userland-only build. It's useful for testsuite.
5950 error
= dp_unregister_provider(type
);
5951 if (error
== 0 || error
== EAFNOSUPPORT
) {
5952 dpif_dummy_register__(type
);
5957 dpif_dummy_register(enum dummy_level level
)
5959 if (level
== DUMMY_OVERRIDE_ALL
) {
5964 dp_enumerate_types(&types
);
5965 SSET_FOR_EACH (type
, &types
) {
5966 dpif_dummy_override(type
);
5968 sset_destroy(&types
);
5969 } else if (level
== DUMMY_OVERRIDE_SYSTEM
) {
5970 dpif_dummy_override("system");
5973 dpif_dummy_register__("dummy");
5975 unixctl_command_register("dpif-dummy/change-port-number",
5976 "dp port new-number",
5977 3, 3, dpif_dummy_change_port_number
, NULL
);
5980 /* Datapath Classifier. */
5982 /* A set of rules that all have the same fields wildcarded. */
5983 struct dpcls_subtable
{
5984 /* The fields are only used by writers. */
5985 struct cmap_node cmap_node OVS_GUARDED
; /* Within dpcls 'subtables_map'. */
5987 /* These fields are accessed by readers. */
5988 struct cmap rules
; /* Contains "struct dpcls_rule"s. */
5989 uint32_t hit_cnt
; /* Number of match hits in subtable in current
5990 optimization interval. */
5991 struct netdev_flow_key mask
; /* Wildcards for fields (const). */
5992 /* 'mask' must be the last field, additional space is allocated here. */
5995 /* Initializes 'cls' as a classifier that initially contains no classification
5998 dpcls_init(struct dpcls
*cls
)
6000 cmap_init(&cls
->subtables_map
);
6001 pvector_init(&cls
->subtables
);
6005 dpcls_destroy_subtable(struct dpcls
*cls
, struct dpcls_subtable
*subtable
)
6007 VLOG_DBG("Destroying subtable %p for in_port %d", subtable
, cls
->in_port
);
6008 pvector_remove(&cls
->subtables
, subtable
);
6009 cmap_remove(&cls
->subtables_map
, &subtable
->cmap_node
,
6010 subtable
->mask
.hash
);
6011 cmap_destroy(&subtable
->rules
);
6012 ovsrcu_postpone(free
, subtable
);
6015 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
6016 * caller's responsibility.
6017 * May only be called after all the readers have been terminated. */
6019 dpcls_destroy(struct dpcls
*cls
)
6022 struct dpcls_subtable
*subtable
;
6024 CMAP_FOR_EACH (subtable
, cmap_node
, &cls
->subtables_map
) {
6025 ovs_assert(cmap_count(&subtable
->rules
) == 0);
6026 dpcls_destroy_subtable(cls
, subtable
);
6028 cmap_destroy(&cls
->subtables_map
);
6029 pvector_destroy(&cls
->subtables
);
6033 static struct dpcls_subtable
*
6034 dpcls_create_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
6036 struct dpcls_subtable
*subtable
;
6038 /* Need to add one. */
6039 subtable
= xmalloc(sizeof *subtable
6040 - sizeof subtable
->mask
.mf
+ mask
->len
);
6041 cmap_init(&subtable
->rules
);
6042 subtable
->hit_cnt
= 0;
6043 netdev_flow_key_clone(&subtable
->mask
, mask
);
6044 cmap_insert(&cls
->subtables_map
, &subtable
->cmap_node
, mask
->hash
);
6045 /* Add the new subtable at the end of the pvector (with no hits yet) */
6046 pvector_insert(&cls
->subtables
, subtable
, 0);
6047 VLOG_DBG("Creating %"PRIuSIZE
". subtable %p for in_port %d",
6048 cmap_count(&cls
->subtables_map
), subtable
, cls
->in_port
);
6049 pvector_publish(&cls
->subtables
);
6054 static inline struct dpcls_subtable
*
6055 dpcls_find_subtable(struct dpcls
*cls
, const struct netdev_flow_key
*mask
)
6057 struct dpcls_subtable
*subtable
;
6059 CMAP_FOR_EACH_WITH_HASH (subtable
, cmap_node
, mask
->hash
,
6060 &cls
->subtables_map
) {
6061 if (netdev_flow_key_equal(&subtable
->mask
, mask
)) {
6065 return dpcls_create_subtable(cls
, mask
);
6069 /* Periodically sort the dpcls subtable vectors according to hit counts */
6071 dpcls_sort_subtable_vector(struct dpcls
*cls
)
6073 struct pvector
*pvec
= &cls
->subtables
;
6074 struct dpcls_subtable
*subtable
;
6076 PVECTOR_FOR_EACH (subtable
, pvec
) {
6077 pvector_change_priority(pvec
, subtable
, subtable
->hit_cnt
);
6078 subtable
->hit_cnt
= 0;
6080 pvector_publish(pvec
);
6084 dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread
*pmd
,
6085 struct polled_queue
*poll_list
, int poll_cnt
)
6089 if (pmd
->ctx
.now
> pmd
->rxq_next_cycle_store
) {
6090 /* Get the cycles that were used to process each queue and store. */
6091 for (unsigned i
= 0; i
< poll_cnt
; i
++) {
6092 uint64_t rxq_cyc_curr
= dp_netdev_rxq_get_cycles(poll_list
[i
].rxq
,
6093 RXQ_CYCLES_PROC_CURR
);
6094 dp_netdev_rxq_set_intrvl_cycles(poll_list
[i
].rxq
, rxq_cyc_curr
);
6095 dp_netdev_rxq_set_cycles(poll_list
[i
].rxq
, RXQ_CYCLES_PROC_CURR
,
6098 /* Start new measuring interval */
6099 pmd
->rxq_next_cycle_store
= pmd
->ctx
.now
+ PMD_RXQ_INTERVAL_LEN
;
6102 if (pmd
->ctx
.now
> pmd
->next_optimization
) {
6103 /* Try to obtain the flow lock to block out revalidator threads.
6104 * If not possible, just try next time. */
6105 if (!ovs_mutex_trylock(&pmd
->flow_mutex
)) {
6106 /* Optimize each classifier */
6107 CMAP_FOR_EACH (cls
, node
, &pmd
->classifiers
) {
6108 dpcls_sort_subtable_vector(cls
);
6110 ovs_mutex_unlock(&pmd
->flow_mutex
);
6111 /* Start new measuring interval */
6112 pmd
->next_optimization
= pmd
->ctx
.now
6113 + DPCLS_OPTIMIZATION_INTERVAL
;
6118 /* Insert 'rule' into 'cls'. */
6120 dpcls_insert(struct dpcls
*cls
, struct dpcls_rule
*rule
,
6121 const struct netdev_flow_key
*mask
)
6123 struct dpcls_subtable
*subtable
= dpcls_find_subtable(cls
, mask
);
6125 /* Refer to subtable's mask, also for later removal. */
6126 rule
->mask
= &subtable
->mask
;
6127 cmap_insert(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
);
6130 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
6132 dpcls_remove(struct dpcls
*cls
, struct dpcls_rule
*rule
)
6134 struct dpcls_subtable
*subtable
;
6136 ovs_assert(rule
->mask
);
6138 /* Get subtable from reference in rule->mask. */
6139 INIT_CONTAINER(subtable
, rule
->mask
, mask
);
6140 if (cmap_remove(&subtable
->rules
, &rule
->cmap_node
, rule
->flow
.hash
)
6142 /* Delete empty subtable. */
6143 dpcls_destroy_subtable(cls
, subtable
);
6144 pvector_publish(&cls
->subtables
);
6148 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
6149 * in 'mask' the values in 'key' and 'target' are the same. */
6151 dpcls_rule_matches_key(const struct dpcls_rule
*rule
,
6152 const struct netdev_flow_key
*target
)
6154 const uint64_t *keyp
= miniflow_get_values(&rule
->flow
.mf
);
6155 const uint64_t *maskp
= miniflow_get_values(&rule
->mask
->mf
);
6158 NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value
, target
, rule
->flow
.mf
.map
) {
6159 if (OVS_UNLIKELY((value
& *maskp
++) != *keyp
++)) {
6166 /* For each miniflow in 'keys' performs a classifier lookup writing the result
6167 * into the corresponding slot in 'rules'. If a particular entry in 'keys' is
6168 * NULL it is skipped.
6170 * This function is optimized for use in the userspace datapath and therefore
6171 * does not implement a lot of features available in the standard
6172 * classifier_lookup() function. Specifically, it does not implement
6173 * priorities, instead returning any rule which matches the flow.
6175 * Returns true if all miniflows found a corresponding rule. */
6177 dpcls_lookup(struct dpcls
*cls
, const struct netdev_flow_key keys
[],
6178 struct dpcls_rule
**rules
, const size_t cnt
,
6181 /* The received 'cnt' miniflows are the search-keys that will be processed
6182 * to find a matching entry into the available subtables.
6183 * The number of bits in map_type is equal to NETDEV_MAX_BURST. */
6184 typedef uint32_t map_type
;
6185 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
6186 BUILD_ASSERT_DECL(MAP_BITS
>= NETDEV_MAX_BURST
);
6188 struct dpcls_subtable
*subtable
;
6190 map_type keys_map
= TYPE_MAXIMUM(map_type
); /* Set all bits. */
6192 uint32_t hashes
[MAP_BITS
];
6193 const struct cmap_node
*nodes
[MAP_BITS
];
6195 if (cnt
!= MAP_BITS
) {
6196 keys_map
>>= MAP_BITS
- cnt
; /* Clear extra bits. */
6198 memset(rules
, 0, cnt
* sizeof *rules
);
6200 int lookups_match
= 0, subtable_pos
= 1;
6202 /* The Datapath classifier - aka dpcls - is composed of subtables.
6203 * Subtables are dynamically created as needed when new rules are inserted.
6204 * Each subtable collects rules with matches on a specific subset of packet
6205 * fields as defined by the subtable's mask. We proceed to process every
6206 * search-key against each subtable, but when a match is found for a
6207 * search-key, the search for that key can stop because the rules are
6208 * non-overlapping. */
6209 PVECTOR_FOR_EACH (subtable
, &cls
->subtables
) {
6212 /* Compute hashes for the remaining keys. Each search-key is
6213 * masked with the subtable's mask to avoid hashing the wildcarded
6215 ULLONG_FOR_EACH_1(i
, keys_map
) {
6216 hashes
[i
] = netdev_flow_key_hash_in_mask(&keys
[i
],
6220 found_map
= cmap_find_batch(&subtable
->rules
, keys_map
, hashes
, nodes
);
6221 /* Check results. When the i-th bit of found_map is set, it means
6222 * that a set of nodes with a matching hash value was found for the
6223 * i-th search-key. Due to possible hash collisions we need to check
6224 * which of the found rules, if any, really matches our masked
6226 ULLONG_FOR_EACH_1(i
, found_map
) {
6227 struct dpcls_rule
*rule
;
6229 CMAP_NODE_FOR_EACH (rule
, cmap_node
, nodes
[i
]) {
6230 if (OVS_LIKELY(dpcls_rule_matches_key(rule
, &keys
[i
]))) {
6232 /* Even at 20 Mpps the 32-bit hit_cnt cannot wrap
6233 * within one second optimization interval. */
6234 subtable
->hit_cnt
++;
6235 lookups_match
+= subtable_pos
;
6239 /* None of the found rules was a match. Reset the i-th bit to
6240 * keep searching this key in the next subtable. */
6241 ULLONG_SET0(found_map
, i
); /* Did not match. */
6243 ; /* Keep Sparse happy. */
6245 keys_map
&= ~found_map
; /* Clear the found rules. */
6247 if (num_lookups_p
) {
6248 *num_lookups_p
= lookups_match
;
6250 return true; /* All found. */
6254 if (num_lookups_p
) {
6255 *num_lookups_p
= lookups_match
;
6257 return false; /* Some misses. */