2 * Copyright (c) 2014-2018, Matthew Macy <mmacy@mattmacy.io>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions are met:
8 * 1. Redistributions of source code must retain the above copyright notice,
9 * this list of conditions and the following disclaimer.
11 * 2. Neither the name of Matthew Macy nor the names of its
12 * contributors may be used to endorse or promote products derived from
13 * this software without specific prior written permission.
15 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25 * POSSIBILITY OF SUCH DAMAGE.
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
32 #include "opt_inet6.h"
34 #include "opt_sched.h"
36 #include <sys/param.h>
37 #include <sys/types.h>
39 #include <sys/eventhandler.h>
41 #include <sys/kernel.h>
44 #include <sys/mutex.h>
45 #include <sys/module.h>
51 #include <sys/socket.h>
52 #include <sys/sockio.h>
53 #include <sys/sysctl.h>
54 #include <sys/syslog.h>
55 #include <sys/taskqueue.h>
56 #include <sys/limits.h>
59 #include <net/if_var.h>
60 #include <net/if_types.h>
61 #include <net/if_media.h>
63 #include <net/ethernet.h>
64 #include <net/mp_ring.h>
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp_lro.h>
70 #include <netinet/in_systm.h>
71 #include <netinet/if_ether.h>
72 #include <netinet/ip.h>
73 #include <netinet/ip6.h>
74 #include <netinet/tcp.h>
75 #include <netinet/ip_var.h>
76 #include <netinet/netdump/netdump.h>
77 #include <netinet6/ip6_var.h>
79 #include <machine/bus.h>
80 #include <machine/in_cksum.h>
85 #include <dev/led/led.h>
86 #include <dev/pci/pcireg.h>
87 #include <dev/pci/pcivar.h>
88 #include <dev/pci/pci_private.h>
90 #include <net/iflib.h>
91 #include <net/iflib_private.h>
96 #include <dev/pci/pci_iov.h>
99 #include <sys/bitstring.h>
101 * enable accounting of every mbuf as it comes in to and goes out of
102 * iflib's software descriptor references
104 #define MEMORY_LOGGING 0
106 * Enable mbuf vectors for compressing long mbuf chains
111 * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
112 * we prefetch needs to be determined by the time spent in m_free vis a vis
113 * the cost of a prefetch. This will of course vary based on the workload:
114 * - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
115 * is quite expensive, thus suggesting very little prefetch.
116 * - small packet forwarding which is just returning a single mbuf to
117 * UMA will typically be very fast vis a vis the cost of a memory
124 * - private structures
125 * - iflib private utility functions
127 * - vlan registry and other exported functions
128 * - iflib public core functions
132 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
135 typedef struct iflib_txq *iflib_txq_t;
137 typedef struct iflib_rxq *iflib_rxq_t;
139 typedef struct iflib_fl *iflib_fl_t;
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
146 typedef struct iflib_filter_info {
147 driver_filter_t *ifi_filter;
148 void *ifi_filter_arg;
149 struct grouptask *ifi_task;
151 } *iflib_filter_info_t;
156 * Pointer to hardware driver's softc
163 if_shared_ctx_t ifc_sctx;
164 struct if_softc_ctx ifc_softc_ctx;
166 struct sx ifc_ctx_sx;
167 struct mtx ifc_state_mtx;
169 iflib_txq_t ifc_txqs;
170 iflib_rxq_t ifc_rxqs;
171 uint32_t ifc_if_flags;
173 uint32_t ifc_max_fl_buf_size;
177 int ifc_watchdog_events;
178 struct cdev *ifc_led_dev;
179 struct resource *ifc_msix_mem;
181 struct if_irq ifc_legacy_irq;
182 struct grouptask ifc_admin_task;
183 struct grouptask ifc_vflr_task;
184 struct iflib_filter_info ifc_filter_info;
185 struct ifmedia ifc_media;
187 struct sysctl_oid *ifc_sysctl_node;
188 uint16_t ifc_sysctl_ntxqs;
189 uint16_t ifc_sysctl_nrxqs;
190 uint16_t ifc_sysctl_qs_eq_override;
191 uint16_t ifc_sysctl_rx_budget;
192 uint16_t ifc_sysctl_tx_abdicate;
194 qidx_t ifc_sysctl_ntxds[8];
195 qidx_t ifc_sysctl_nrxds[8];
196 struct if_txrx ifc_txrx;
197 #define isc_txd_encap ifc_txrx.ift_txd_encap
198 #define isc_txd_flush ifc_txrx.ift_txd_flush
199 #define isc_txd_credits_update ifc_txrx.ift_txd_credits_update
200 #define isc_rxd_available ifc_txrx.ift_rxd_available
201 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
202 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
203 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
204 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
205 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
206 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
207 eventhandler_tag ifc_vlan_attach_event;
208 eventhandler_tag ifc_vlan_detach_event;
209 uint8_t ifc_mac[ETHER_ADDR_LEN];
210 char ifc_mtx_name[16];
215 iflib_get_softc(if_ctx_t ctx)
218 return (ctx->ifc_softc);
222 iflib_get_dev(if_ctx_t ctx)
225 return (ctx->ifc_dev);
229 iflib_get_ifp(if_ctx_t ctx)
232 return (ctx->ifc_ifp);
236 iflib_get_media(if_ctx_t ctx)
239 return (&ctx->ifc_media);
243 iflib_get_flags(if_ctx_t ctx)
245 return (ctx->ifc_flags);
249 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
252 bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
256 iflib_get_softc_ctx(if_ctx_t ctx)
259 return (&ctx->ifc_softc_ctx);
263 iflib_get_sctx(if_ctx_t ctx)
266 return (ctx->ifc_sctx);
269 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
270 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
271 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
273 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
274 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
276 typedef struct iflib_sw_rx_desc_array {
277 bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */
278 struct mbuf **ifsd_m; /* pkthdr mbufs */
279 caddr_t *ifsd_cl; /* direct cluster pointer for rx */
280 bus_addr_t *ifsd_ba; /* bus addr of cluster for rx */
281 } iflib_rxsd_array_t;
283 typedef struct iflib_sw_tx_desc_array {
284 bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */
285 bus_dmamap_t *ifsd_tso_map; /* bus_dma maps for TSO packet */
286 struct mbuf **ifsd_m; /* pkthdr mbufs */
290 /* magic number that should be high enough for any hardware */
291 #define IFLIB_MAX_TX_SEGS 128
292 /* bnxt supports 64 with hardware LRO enabled */
293 #define IFLIB_MAX_RX_SEGS 64
294 #define IFLIB_RX_COPY_THRESH 128
295 #define IFLIB_MAX_RX_REFRESH 32
296 /* The minimum descriptors per second before we start coalescing */
297 #define IFLIB_MIN_DESC_SEC 16384
298 #define IFLIB_DEFAULT_TX_UPDATE_FREQ 16
299 #define IFLIB_QUEUE_IDLE 0
300 #define IFLIB_QUEUE_HUNG 1
301 #define IFLIB_QUEUE_WORKING 2
302 /* maximum number of txqs that can share an rx interrupt */
303 #define IFLIB_MAX_TX_SHARED_INTR 4
305 /* this should really scale with ring size - this is a fairly arbitrary value */
306 #define TX_BATCH_SIZE 32
308 #define IFLIB_RESTART_BUDGET 8
311 #define CSUM_OFFLOAD (CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
312 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
313 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
317 qidx_t ift_cidx_processed;
320 uint8_t ift_br_offset;
321 uint16_t ift_npending;
322 uint16_t ift_db_pending;
323 uint16_t ift_rs_pending;
325 uint8_t ift_txd_size[8];
326 uint64_t ift_processed;
327 uint64_t ift_cleaned;
328 uint64_t ift_cleaned_prev;
330 uint64_t ift_enqueued;
331 uint64_t ift_dequeued;
333 uint64_t ift_no_tx_dma_setup;
334 uint64_t ift_no_desc_avail;
335 uint64_t ift_mbuf_defrag_failed;
336 uint64_t ift_mbuf_defrag;
337 uint64_t ift_map_failed;
338 uint64_t ift_txd_encap_efbig;
339 uint64_t ift_pullups;
340 uint64_t ift_last_timer_tick;
343 struct mtx ift_db_mtx;
345 /* constant values */
347 struct ifmp_ring *ift_br;
348 struct grouptask ift_task;
351 struct callout ift_timer;
353 if_txsd_vec_t ift_sds;
356 uint8_t ift_update_freq;
357 struct iflib_filter_info ift_filter_info;
358 bus_dma_tag_t ift_desc_tag;
359 bus_dma_tag_t ift_tso_desc_tag;
360 iflib_dma_info_t ift_ifdi;
361 #define MTX_NAME_LEN 16
362 char ift_mtx_name[MTX_NAME_LEN];
363 char ift_db_mtx_name[MTX_NAME_LEN];
364 bus_dma_segment_t ift_segs[IFLIB_MAX_TX_SEGS] __aligned(CACHE_LINE_SIZE);
365 #ifdef IFLIB_DIAGNOSTICS
366 uint64_t ift_cpu_exec_count[256];
368 } __aligned(CACHE_LINE_SIZE);
375 uint8_t ifl_rxd_size;
377 uint64_t ifl_m_enqueued;
378 uint64_t ifl_m_dequeued;
379 uint64_t ifl_cl_enqueued;
380 uint64_t ifl_cl_dequeued;
384 bitstr_t *ifl_rx_bitmap;
388 uint16_t ifl_buf_size;
391 iflib_rxsd_array_t ifl_sds;
394 bus_dma_tag_t ifl_desc_tag;
395 iflib_dma_info_t ifl_ifdi;
396 uint64_t ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
397 caddr_t ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
398 qidx_t ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
399 } __aligned(CACHE_LINE_SIZE);
402 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
408 else if (pidx < cidx)
409 used = size - cidx + pidx;
410 else if (gen == 0 && pidx == cidx)
412 else if (gen == 1 && pidx == cidx)
420 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
422 #define IDXDIFF(head, tail, wrap) \
423 ((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
426 /* If there is a separate completion queue -
427 * these are the cq cidx and pidx. Otherwise
434 uint8_t ifr_fl_offset;
440 uint8_t ifr_lro_enabled;
443 uint8_t ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
444 struct lro_ctrl ifr_lc;
445 struct grouptask ifr_task;
446 struct iflib_filter_info ifr_filter_info;
447 iflib_dma_info_t ifr_ifdi;
449 /* dynamically allocate if any drivers need a value substantially larger than this */
450 struct if_rxd_frag ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
451 #ifdef IFLIB_DIAGNOSTICS
452 uint64_t ifr_cpu_exec_count[256];
454 } __aligned(CACHE_LINE_SIZE);
456 typedef struct if_rxsd {
458 struct mbuf **ifsd_m;
463 /* multiple of word size */
465 #define PKT_INFO_SIZE 6
466 #define RXD_INFO_SIZE 5
467 #define PKT_TYPE uint64_t
469 #define PKT_INFO_SIZE 11
470 #define RXD_INFO_SIZE 8
471 #define PKT_TYPE uint32_t
473 #define PKT_LOOP_BOUND ((PKT_INFO_SIZE/3)*3)
474 #define RXD_LOOP_BOUND ((RXD_INFO_SIZE/4)*4)
476 typedef struct if_pkt_info_pad {
477 PKT_TYPE pkt_val[PKT_INFO_SIZE];
478 } *if_pkt_info_pad_t;
479 typedef struct if_rxd_info_pad {
480 PKT_TYPE rxd_val[RXD_INFO_SIZE];
481 } *if_rxd_info_pad_t;
483 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
484 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
488 pkt_info_zero(if_pkt_info_t pi)
490 if_pkt_info_pad_t pi_pad;
492 pi_pad = (if_pkt_info_pad_t)pi;
493 pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
494 pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
496 pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
497 pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
501 static device_method_t iflib_pseudo_methods[] = {
502 DEVMETHOD(device_attach, noop_attach),
503 DEVMETHOD(device_detach, iflib_pseudo_detach),
507 driver_t iflib_pseudodriver = {
508 "iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
512 rxd_info_zero(if_rxd_info_t ri)
514 if_rxd_info_pad_t ri_pad;
517 ri_pad = (if_rxd_info_pad_t)ri;
518 for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
519 ri_pad->rxd_val[i] = 0;
520 ri_pad->rxd_val[i+1] = 0;
521 ri_pad->rxd_val[i+2] = 0;
522 ri_pad->rxd_val[i+3] = 0;
525 ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
530 * Only allow a single packet to take up most 1/nth of the tx ring
532 #define MAX_SINGLE_PACKET_FRACTION 12
533 #define IF_BAD_DMA (bus_addr_t)-1
535 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
537 #define CTX_LOCK_INIT(_sc) sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
538 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
539 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
540 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
543 #define STATE_LOCK_INIT(_sc, _name) mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
544 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
545 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
546 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
550 #define CALLOUT_LOCK(txq) mtx_lock(&txq->ift_mtx)
551 #define CALLOUT_UNLOCK(txq) mtx_unlock(&txq->ift_mtx)
554 iflib_set_detach(if_ctx_t ctx)
557 ctx->ifc_flags |= IFC_IN_DETACH;
561 /* Our boot-time initialization hook */
562 static int iflib_module_event_handler(module_t, int, void *);
564 static moduledata_t iflib_moduledata = {
566 iflib_module_event_handler,
570 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
571 MODULE_VERSION(iflib, 1);
573 MODULE_DEPEND(iflib, pci, 1, 1, 1);
574 MODULE_DEPEND(iflib, ether, 1, 1, 1);
576 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
577 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
579 #ifndef IFLIB_DEBUG_COUNTERS
581 #define IFLIB_DEBUG_COUNTERS 1
583 #define IFLIB_DEBUG_COUNTERS 0
584 #endif /* !INVARIANTS */
587 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
588 "iflib driver parameters");
591 * XXX need to ensure that this can't accidentally cause the head to be moved backwards
593 static int iflib_min_tx_latency = 0;
594 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
595 &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
596 static int iflib_no_tx_batch = 0;
597 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
598 &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
601 #if IFLIB_DEBUG_COUNTERS
603 static int iflib_tx_seen;
604 static int iflib_tx_sent;
605 static int iflib_tx_encap;
606 static int iflib_rx_allocs;
607 static int iflib_fl_refills;
608 static int iflib_fl_refills_large;
609 static int iflib_tx_frees;
611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
612 &iflib_tx_seen, 0, "# tx mbufs seen");
613 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
614 &iflib_tx_sent, 0, "# tx mbufs sent");
615 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
616 &iflib_tx_encap, 0, "# tx mbufs encapped");
617 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
618 &iflib_tx_frees, 0, "# tx frees");
619 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
620 &iflib_rx_allocs, 0, "# rx allocations");
621 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
622 &iflib_fl_refills, 0, "# refills");
623 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
624 &iflib_fl_refills_large, 0, "# large refills");
627 static int iflib_txq_drain_flushing;
628 static int iflib_txq_drain_oactive;
629 static int iflib_txq_drain_notready;
631 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
632 &iflib_txq_drain_flushing, 0, "# drain flushes");
633 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
634 &iflib_txq_drain_oactive, 0, "# drain oactives");
635 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
636 &iflib_txq_drain_notready, 0, "# drain notready");
639 static int iflib_encap_load_mbuf_fail;
640 static int iflib_encap_pad_mbuf_fail;
641 static int iflib_encap_txq_avail_fail;
642 static int iflib_encap_txd_encap_fail;
644 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
645 &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
646 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
647 &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
648 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
649 &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
650 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
651 &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
653 static int iflib_task_fn_rxs;
654 static int iflib_rx_intr_enables;
655 static int iflib_fast_intrs;
656 static int iflib_rx_unavail;
657 static int iflib_rx_ctx_inactive;
658 static int iflib_rx_if_input;
659 static int iflib_rx_mbuf_null;
660 static int iflib_rxd_flush;
662 static int iflib_verbose_debug;
664 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
665 &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
666 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
667 &iflib_rx_intr_enables, 0, "# rx intr enables");
668 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
669 &iflib_fast_intrs, 0, "# fast_intr calls");
670 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
671 &iflib_rx_unavail, 0, "# times rxeof called with no available data");
672 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
673 &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
674 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
675 &iflib_rx_if_input, 0, "# times rxeof called if_input");
676 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
677 &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
678 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
679 &iflib_rxd_flush, 0, "# times rxd_flush called");
680 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
681 &iflib_verbose_debug, 0, "enable verbose debugging");
683 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
685 iflib_debug_reset(void)
687 iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
688 iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
689 iflib_txq_drain_flushing = iflib_txq_drain_oactive =
690 iflib_txq_drain_notready =
691 iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
692 iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
693 iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
695 iflib_rx_ctx_inactive = iflib_rx_if_input =
696 iflib_rx_mbuf_null = iflib_rxd_flush = 0;
700 #define DBG_COUNTER_INC(name)
701 static void iflib_debug_reset(void) {}
704 #define IFLIB_DEBUG 0
706 static void iflib_tx_structures_free(if_ctx_t ctx);
707 static void iflib_rx_structures_free(if_ctx_t ctx);
708 static int iflib_queues_alloc(if_ctx_t ctx);
709 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
710 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
711 static int iflib_qset_structures_setup(if_ctx_t ctx);
712 static int iflib_msix_init(if_ctx_t ctx);
713 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
714 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
715 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
717 static void iflib_altq_if_start(if_t ifp);
718 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
720 static int iflib_register(if_ctx_t);
721 static void iflib_init_locked(if_ctx_t ctx);
722 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
723 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
724 static void iflib_ifmp_purge(iflib_txq_t txq);
725 static void _iflib_pre_assert(if_softc_ctx_t scctx);
726 static void iflib_if_init_locked(if_ctx_t ctx);
727 static void iflib_free_intr_mem(if_ctx_t ctx);
728 #ifndef __NO_STRICT_ALIGNMENT
729 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
732 NETDUMP_DEFINE(iflib);
735 #include <sys/selinfo.h>
736 #include <net/netmap.h>
737 #include <dev/netmap/netmap_kern.h>
739 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
741 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
744 * device-specific sysctl variables:
746 * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
747 * During regular operations the CRC is stripped, but on some
748 * hardware reception of frames not multiple of 64 is slower,
749 * so using crcstrip=0 helps in benchmarks.
751 * iflib_rx_miss, iflib_rx_miss_bufs:
752 * count packets that might be missed due to lost interrupts.
754 SYSCTL_DECL(_dev_netmap);
756 * The xl driver by default strips CRCs and we do not override it.
759 int iflib_crcstrip = 1;
760 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
761 CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
763 int iflib_rx_miss, iflib_rx_miss_bufs;
764 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
765 CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
766 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
767 CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
770 * Register/unregister. We are already under netmap lock.
771 * Only called on the first register or the last unregister.
774 iflib_netmap_register(struct netmap_adapter *na, int onoff)
776 struct ifnet *ifp = na->ifp;
777 if_ctx_t ctx = ifp->if_softc;
781 IFDI_INTR_DISABLE(ctx);
783 /* Tell the stack that the interface is no longer active */
784 ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
787 IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
789 /* enable or disable flags and callbacks in na and ifp */
791 nm_set_native_flags(na);
793 nm_clear_native_flags(na);
796 iflib_init_locked(ctx);
797 IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
798 status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
800 nm_clear_native_flags(na);
806 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
808 struct netmap_adapter *na = kring->na;
809 u_int const lim = kring->nkr_num_slots - 1;
810 u_int head = kring->rhead;
811 struct netmap_ring *ring = kring->ring;
813 struct if_rxd_update iru;
814 if_ctx_t ctx = rxq->ifr_ctx;
815 iflib_fl_t fl = &rxq->ifr_fl[0];
816 uint32_t refill_pidx, nic_i;
817 #if IFLIB_DEBUG_COUNTERS
821 if (nm_i == head && __predict_true(!init))
823 iru_init(&iru, rxq, 0 /* flid */);
824 map = fl->ifl_sds.ifsd_map;
825 refill_pidx = netmap_idx_k2n(kring, nm_i);
827 * IMPORTANT: we must leave one free slot in the ring,
828 * so move head back by one unit
830 head = nm_prev(head, lim);
832 DBG_COUNTER_INC(fl_refills);
833 while (nm_i != head) {
834 #if IFLIB_DEBUG_COUNTERS
836 DBG_COUNTER_INC(fl_refills_large);
838 for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
839 struct netmap_slot *slot = &ring->slot[nm_i];
840 void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
841 uint32_t nic_i_dma = refill_pidx;
842 nic_i = netmap_idx_k2n(kring, nm_i);
844 MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
846 if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
847 return netmap_ring_reinit(kring);
849 fl->ifl_vm_addrs[tmp_pidx] = addr;
850 if (__predict_false(init) && map) {
851 netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
852 } else if (map && (slot->flags & NS_BUF_CHANGED)) {
853 /* buffer has changed, reload map */
854 netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
856 slot->flags &= ~NS_BUF_CHANGED;
858 nm_i = nm_next(nm_i, lim);
859 fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
860 if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
863 iru.iru_pidx = refill_pidx;
864 iru.iru_count = tmp_pidx+1;
865 ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
871 for (int n = 0; n < iru.iru_count; n++) {
872 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
873 BUS_DMASYNC_PREREAD);
874 /* XXX - change this to not use the netmap func*/
875 nic_i_dma = nm_next(nic_i_dma, lim);
879 kring->nr_hwcur = head;
882 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
883 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
884 if (__predict_true(nic_i != UINT_MAX)) {
885 ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
886 DBG_COUNTER_INC(rxd_flush);
892 * Reconcile kernel and user view of the transmit ring.
894 * All information is in the kring.
895 * Userspace wants to send packets up to the one before kring->rhead,
896 * kernel knows kring->nr_hwcur is the first unsent packet.
898 * Here we push packets out (as many as possible), and possibly
899 * reclaim buffers from previously completed transmission.
901 * The caller (netmap) guarantees that there is only one instance
902 * running at any time. Any interference with other driver
903 * methods should be handled by the individual drivers.
906 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
908 struct netmap_adapter *na = kring->na;
909 struct ifnet *ifp = na->ifp;
910 struct netmap_ring *ring = kring->ring;
911 u_int nm_i; /* index into the netmap kring */
912 u_int nic_i; /* index into the NIC ring */
914 u_int const lim = kring->nkr_num_slots - 1;
915 u_int const head = kring->rhead;
916 struct if_pkt_info pi;
919 * interrupts on every tx packet are expensive so request
920 * them every half ring, or where NS_REPORT is set
922 u_int report_frequency = kring->nkr_num_slots >> 1;
923 /* device-specific */
924 if_ctx_t ctx = ifp->if_softc;
925 iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
927 bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
928 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
932 * First part: process new packets to send.
933 * nm_i is the current index in the netmap kring,
934 * nic_i is the corresponding index in the NIC ring.
936 * If we have packets to send (nm_i != head)
937 * iterate over the netmap ring, fetch length and update
938 * the corresponding slot in the NIC ring. Some drivers also
939 * need to update the buffer's physical address in the NIC slot
940 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
942 * The netmap_reload_map() calls is especially expensive,
943 * even when (as in this case) the tag is 0, so do only
944 * when the buffer has actually changed.
946 * If possible do not set the report/intr bit on all slots,
947 * but only a few times per ring or when NS_REPORT is set.
949 * Finally, on 10G and faster drivers, it might be useful
950 * to prefetch the next slot and txr entry.
953 nm_i = kring->nr_hwcur;
954 if (nm_i != head) { /* we have new packets to send */
956 pi.ipi_segs = txq->ift_segs;
957 pi.ipi_qsidx = kring->ring_id;
958 nic_i = netmap_idx_k2n(kring, nm_i);
960 __builtin_prefetch(&ring->slot[nm_i]);
961 __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
962 if (txq->ift_sds.ifsd_map)
963 __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
965 for (n = 0; nm_i != head; n++) {
966 struct netmap_slot *slot = &ring->slot[nm_i];
967 u_int len = slot->len;
969 void *addr = PNMB(na, slot, &paddr);
970 int flags = (slot->flags & NS_REPORT ||
971 nic_i == 0 || nic_i == report_frequency) ?
974 /* device-specific */
976 pi.ipi_segs[0].ds_addr = paddr;
977 pi.ipi_segs[0].ds_len = len;
981 pi.ipi_flags = flags;
983 /* Fill the slot in the NIC ring. */
984 ctx->isc_txd_encap(ctx->ifc_softc, &pi);
985 DBG_COUNTER_INC(tx_encap);
987 /* prefetch for next round */
988 __builtin_prefetch(&ring->slot[nm_i + 1]);
989 __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
990 if (txq->ift_sds.ifsd_map) {
991 __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
993 NM_CHECK_ADDR_LEN(na, addr, len);
995 if (slot->flags & NS_BUF_CHANGED) {
996 /* buffer has changed, reload map */
997 netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
999 /* make sure changes to the buffer are synced */
1000 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
1001 BUS_DMASYNC_PREWRITE);
1003 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
1004 nm_i = nm_next(nm_i, lim);
1005 nic_i = nm_next(nic_i, lim);
1007 kring->nr_hwcur = nm_i;
1009 /* synchronize the NIC ring */
1010 bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
1011 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1013 /* (re)start the tx unit up to slot nic_i (excluded) */
1014 ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1018 * Second part: reclaim buffers for completed transmissions.
1020 * If there are unclaimed buffers, attempt to reclaim them.
1021 * If none are reclaimed, and TX IRQs are not in use, do an initial
1022 * minimal delay, then trigger the tx handler which will spin in the
1025 if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1026 if (iflib_tx_credits_update(ctx, txq)) {
1027 /* some tx completed, increment avail */
1028 nic_i = txq->ift_cidx_processed;
1029 kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1032 if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1033 if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1034 callout_reset_on(&txq->ift_timer, hz < 2000 ? 1 : hz / 1000,
1035 iflib_timer, txq, txq->ift_timer.c_cpu);
1041 * Reconcile kernel and user view of the receive ring.
1042 * Same as for the txsync, this routine must be efficient.
1043 * The caller guarantees a single invocations, but races against
1044 * the rest of the driver should be handled here.
1046 * On call, kring->rhead is the first packet that userspace wants
1047 * to keep, and kring->rcur is the wakeup point.
1048 * The kernel has previously reported packets up to kring->rtail.
1050 * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1051 * of whether or not we received an interrupt.
1054 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1056 struct netmap_adapter *na = kring->na;
1057 struct netmap_ring *ring = kring->ring;
1058 uint32_t nm_i; /* index into the netmap ring */
1059 uint32_t nic_i; /* index into the NIC ring */
1061 u_int const lim = kring->nkr_num_slots - 1;
1062 u_int const head = kring->rhead;
1063 int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1064 struct if_rxd_info ri;
1066 struct ifnet *ifp = na->ifp;
1067 if_ctx_t ctx = ifp->if_softc;
1068 iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1069 iflib_fl_t fl = rxq->ifr_fl;
1071 return netmap_ring_reinit(kring);
1073 /* XXX check sync modes */
1074 for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1075 if (fl->ifl_sds.ifsd_map == NULL)
1077 bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1078 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1081 * First part: import newly received packets.
1083 * nm_i is the index of the next free slot in the netmap ring,
1084 * nic_i is the index of the next received packet in the NIC ring,
1085 * and they may differ in case if_init() has been called while
1086 * in netmap mode. For the receive ring we have
1088 * nic_i = rxr->next_check;
1089 * nm_i = kring->nr_hwtail (previous)
1091 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1093 * rxr->next_check is set to 0 on a ring reinit
1095 if (netmap_no_pendintr || force_update) {
1096 int crclen = iflib_crcstrip ? 0 : 4;
1099 for (i = 0; i < rxq->ifr_nfl; i++) {
1100 fl = &rxq->ifr_fl[i];
1101 nic_i = fl->ifl_cidx;
1102 nm_i = netmap_idx_n2k(kring, nic_i);
1103 avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1104 for (n = 0; avail > 0; n++, avail--) {
1106 ri.iri_frags = rxq->ifr_frags;
1107 ri.iri_qsidx = kring->ring_id;
1108 ri.iri_ifp = ctx->ifc_ifp;
1109 ri.iri_cidx = nic_i;
1111 error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1112 ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1113 ring->slot[nm_i].flags = 0;
1114 bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1115 fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1116 nm_i = nm_next(nm_i, lim);
1117 nic_i = nm_next(nic_i, lim);
1119 if (n) { /* update the state variables */
1120 if (netmap_no_pendintr && !force_update) {
1123 iflib_rx_miss_bufs += n;
1125 fl->ifl_cidx = nic_i;
1126 kring->nr_hwtail = nm_i;
1128 kring->nr_kflags &= ~NKR_PENDINTR;
1132 * Second part: skip past packets that userspace has released.
1133 * (kring->nr_hwcur to head excluded),
1134 * and make the buffers available for reception.
1135 * As usual nm_i is the index in the netmap ring,
1136 * nic_i is the index in the NIC ring, and
1137 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1139 /* XXX not sure how this will work with multiple free lists */
1140 nm_i = kring->nr_hwcur;
1142 return (netmap_fl_refill(rxq, kring, nm_i, false));
1146 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1148 struct ifnet *ifp = na->ifp;
1149 if_ctx_t ctx = ifp->if_softc;
1153 IFDI_INTR_ENABLE(ctx);
1155 IFDI_INTR_DISABLE(ctx);
1162 iflib_netmap_attach(if_ctx_t ctx)
1164 struct netmap_adapter na;
1165 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1167 bzero(&na, sizeof(na));
1169 na.ifp = ctx->ifc_ifp;
1170 na.na_flags = NAF_BDG_MAYSLEEP;
1171 MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1172 MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1174 na.num_tx_desc = scctx->isc_ntxd[0];
1175 na.num_rx_desc = scctx->isc_nrxd[0];
1176 na.nm_txsync = iflib_netmap_txsync;
1177 na.nm_rxsync = iflib_netmap_rxsync;
1178 na.nm_register = iflib_netmap_register;
1179 na.nm_intr = iflib_netmap_intr;
1180 na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1181 na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1182 return (netmap_attach(&na));
1186 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1188 struct netmap_adapter *na = NA(ctx->ifc_ifp);
1189 struct netmap_slot *slot;
1191 slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1194 for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1197 * In netmap mode, set the map for the packet buffer.
1198 * NOTE: Some drivers (not this one) also need to set
1199 * the physical buffer address in the NIC ring.
1200 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1201 * netmap slot index, si
1203 int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1204 netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1209 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1211 struct netmap_adapter *na = NA(ctx->ifc_ifp);
1212 struct netmap_kring *kring = na->rx_rings[rxq->ifr_id];
1213 struct netmap_slot *slot;
1216 slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1219 nm_i = netmap_idx_n2k(kring, 0);
1220 netmap_fl_refill(rxq, kring, nm_i, true);
1224 iflib_netmap_timer_adjust(if_ctx_t ctx, uint16_t txqid, uint32_t *reset_on)
1226 struct netmap_kring *kring;
1228 kring = NA(ctx->ifc_ifp)->tx_rings[txqid];
1230 if (kring->nr_hwcur != nm_next(kring->nr_hwtail, kring->nkr_num_slots - 1)) {
1231 if (ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false))
1232 netmap_tx_irq(ctx->ifc_ifp, txqid);
1233 if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) {
1237 *reset_on = hz / 1000;
1242 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1245 #define iflib_netmap_txq_init(ctx, txq)
1246 #define iflib_netmap_rxq_init(ctx, rxq)
1247 #define iflib_netmap_detach(ifp)
1249 #define iflib_netmap_attach(ctx) (0)
1250 #define netmap_rx_irq(ifp, qid, budget) (0)
1251 #define netmap_tx_irq(ifp, qid) do {} while (0)
1252 #define iflib_netmap_timer_adjust(ctx, txqid, reset_on)
1256 #if defined(__i386__) || defined(__amd64__)
1257 static __inline void
1260 __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1262 static __inline void
1263 prefetch2cachelines(void *x)
1265 __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1266 #if (CACHE_LINE_SIZE < 128)
1267 __asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1272 #define prefetch2cachelines(x)
1276 iflib_gen_mac(if_ctx_t ctx)
1280 char uuid[HOSTUUIDLEN+1];
1281 char buf[HOSTUUIDLEN+16];
1283 unsigned char digest[16];
1287 uuid[HOSTUUIDLEN] = 0;
1288 bcopy(td->td_ucred->cr_prison->pr_hostuuid, uuid, HOSTUUIDLEN);
1289 snprintf(buf, HOSTUUIDLEN+16, "%s-%s", uuid, device_get_nameunit(ctx->ifc_dev));
1291 * Generate a pseudo-random, deterministic MAC
1292 * address based on the UUID and unit number.
1293 * The FreeBSD Foundation OUI of 58-9C-FC is used.
1296 MD5Update(&mdctx, buf, strlen(buf));
1297 MD5Final(digest, &mdctx);
1308 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1312 fl = &rxq->ifr_fl[flid];
1313 iru->iru_paddrs = fl->ifl_bus_addrs;
1314 iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1315 iru->iru_idxs = fl->ifl_rxd_idxs;
1316 iru->iru_qsidx = rxq->ifr_id;
1317 iru->iru_buf_size = fl->ifl_buf_size;
1318 iru->iru_flidx = fl->ifl_id;
1322 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1326 *(bus_addr_t *) arg = segs[0].ds_addr;
1330 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1333 if_shared_ctx_t sctx = ctx->ifc_sctx;
1334 device_t dev = ctx->ifc_dev;
1336 KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1338 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1339 sctx->isc_q_align, 0, /* alignment, bounds */
1340 BUS_SPACE_MAXADDR, /* lowaddr */
1341 BUS_SPACE_MAXADDR, /* highaddr */
1342 NULL, NULL, /* filter, filterarg */
1345 size, /* maxsegsize */
1346 BUS_DMA_ALLOCNOW, /* flags */
1347 NULL, /* lockfunc */
1352 "%s: bus_dma_tag_create failed: %d\n",
1357 err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1358 BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1361 "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1362 __func__, (uintmax_t)size, err);
1366 dma->idi_paddr = IF_BAD_DMA;
1367 err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1368 size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1369 if (err || dma->idi_paddr == IF_BAD_DMA) {
1371 "%s: bus_dmamap_load failed: %d\n",
1376 dma->idi_size = size;
1380 bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1382 bus_dma_tag_destroy(dma->idi_tag);
1384 dma->idi_tag = NULL;
1390 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1393 iflib_dma_info_t *dmaiter;
1396 for (i = 0; i < count; i++, dmaiter++) {
1397 if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1401 iflib_dma_free_multi(dmalist, i);
1406 iflib_dma_free(iflib_dma_info_t dma)
1408 if (dma->idi_tag == NULL)
1410 if (dma->idi_paddr != IF_BAD_DMA) {
1411 bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1412 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1413 bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1414 dma->idi_paddr = IF_BAD_DMA;
1416 if (dma->idi_vaddr != NULL) {
1417 bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1418 dma->idi_vaddr = NULL;
1420 bus_dma_tag_destroy(dma->idi_tag);
1421 dma->idi_tag = NULL;
1425 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1428 iflib_dma_info_t *dmaiter = dmalist;
1430 for (i = 0; i < count; i++, dmaiter++)
1431 iflib_dma_free(*dmaiter);
1434 #ifdef EARLY_AP_STARTUP
1435 static const int iflib_started = 1;
1438 * We used to abuse the smp_started flag to decide if the queues have been
1439 * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1440 * That gave bad races, since the SYSINIT() runs strictly after smp_started
1441 * is set. Run a SYSINIT() strictly after that to just set a usable
1445 static int iflib_started;
1448 iflib_record_started(void *arg)
1453 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1454 iflib_record_started, NULL);
1458 iflib_fast_intr(void *arg)
1460 iflib_filter_info_t info = arg;
1461 struct grouptask *gtask = info->ifi_task;
1463 return (FILTER_HANDLED);
1465 DBG_COUNTER_INC(fast_intrs);
1466 if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1467 return (FILTER_HANDLED);
1469 GROUPTASK_ENQUEUE(gtask);
1470 return (FILTER_HANDLED);
1474 iflib_fast_intr_rxtx(void *arg)
1476 iflib_filter_info_t info = arg;
1477 struct grouptask *gtask = info->ifi_task;
1478 iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1479 if_ctx_t ctx = NULL;;
1483 return (FILTER_HANDLED);
1485 DBG_COUNTER_INC(fast_intrs);
1486 if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1487 return (FILTER_HANDLED);
1489 MPASS(rxq->ifr_ntxqirq);
1490 for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1491 qidx_t txqid = rxq->ifr_txqid[i];
1495 bus_dmamap_sync(rxq->ifr_ifdi->idi_tag, rxq->ifr_ifdi->idi_map,
1496 BUS_DMASYNC_POSTREAD);
1497 if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1498 IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1501 GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1503 if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1504 cidx = rxq->ifr_cq_cidx;
1506 cidx = rxq->ifr_fl[0].ifl_cidx;
1507 if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1508 GROUPTASK_ENQUEUE(gtask);
1510 IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1511 DBG_COUNTER_INC(rx_intr_enables);
1513 return (FILTER_HANDLED);
1518 iflib_fast_intr_ctx(void *arg)
1520 iflib_filter_info_t info = arg;
1521 struct grouptask *gtask = info->ifi_task;
1524 return (FILTER_HANDLED);
1526 DBG_COUNTER_INC(fast_intrs);
1527 if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1528 return (FILTER_HANDLED);
1530 GROUPTASK_ENQUEUE(gtask);
1531 return (FILTER_HANDLED);
1535 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1536 driver_filter_t filter, driver_intr_t handler, void *arg,
1540 struct resource *res;
1542 device_t dev = ctx->ifc_dev;
1545 if (ctx->ifc_flags & IFC_LEGACY)
1546 flags |= RF_SHAREABLE;
1549 res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1552 "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1556 KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1557 rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1558 filter, handler, arg, &tag);
1561 "failed to setup interrupt for rid %d, name %s: %d\n",
1562 rid, name ? name : "unknown", rc);
1565 bus_describe_intr(dev, res, tag, "%s", name);
1572 /*********************************************************************
1574 * Allocate memory for tx_buffer structures. The tx_buffer stores all
1575 * the information needed to transmit a packet on the wire. This is
1576 * called only once at attach, setup is done every reset.
1578 **********************************************************************/
1581 iflib_txsd_alloc(iflib_txq_t txq)
1583 if_ctx_t ctx = txq->ift_ctx;
1584 if_shared_ctx_t sctx = ctx->ifc_sctx;
1585 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1586 device_t dev = ctx->ifc_dev;
1587 bus_size_t tsomaxsize;
1588 int err, nsegments, ntsosegments;
1591 nsegments = scctx->isc_tx_nsegments;
1592 ntsosegments = scctx->isc_tx_tso_segments_max;
1593 tsomaxsize = scctx->isc_tx_tso_size_max;
1594 if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1595 tsomaxsize += sizeof(struct ether_vlan_header);
1596 MPASS(scctx->isc_ntxd[0] > 0);
1597 MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1598 MPASS(nsegments > 0);
1599 if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1600 MPASS(ntsosegments > 0);
1601 MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1605 * Setup DMA descriptor areas.
1607 if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1608 1, 0, /* alignment, bounds */
1609 BUS_SPACE_MAXADDR, /* lowaddr */
1610 BUS_SPACE_MAXADDR, /* highaddr */
1611 NULL, NULL, /* filter, filterarg */
1612 sctx->isc_tx_maxsize, /* maxsize */
1613 nsegments, /* nsegments */
1614 sctx->isc_tx_maxsegsize, /* maxsegsize */
1616 NULL, /* lockfunc */
1617 NULL, /* lockfuncarg */
1618 &txq->ift_desc_tag))) {
1619 device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1620 device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1621 (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1624 tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1625 if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1626 1, 0, /* alignment, bounds */
1627 BUS_SPACE_MAXADDR, /* lowaddr */
1628 BUS_SPACE_MAXADDR, /* highaddr */
1629 NULL, NULL, /* filter, filterarg */
1630 tsomaxsize, /* maxsize */
1631 ntsosegments, /* nsegments */
1632 sctx->isc_tso_maxsegsize,/* maxsegsize */
1634 NULL, /* lockfunc */
1635 NULL, /* lockfuncarg */
1636 &txq->ift_tso_desc_tag))) {
1637 device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1640 if (!(txq->ift_sds.ifsd_m =
1641 (struct mbuf **) malloc(sizeof(struct mbuf *) *
1642 scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1643 device_printf(dev, "Unable to allocate tx_buffer memory\n");
1648 /* Create the descriptor buffer dma maps */
1649 if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1650 sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1651 M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1652 device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1657 if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1658 sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1659 M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1660 device_printf(dev, "Unable to allocate TSO tx_buffer "
1666 for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1667 err = bus_dmamap_create(txq->ift_desc_tag, 0,
1668 &txq->ift_sds.ifsd_map[i]);
1670 device_printf(dev, "Unable to create TX DMA map\n");
1675 err = bus_dmamap_create(txq->ift_tso_desc_tag, 0,
1676 &txq->ift_sds.ifsd_tso_map[i]);
1678 device_printf(dev, "Unable to create TSO TX DMA map\n");
1684 /* We free all, it handles case where we are in the middle */
1685 iflib_tx_structures_free(ctx);
1690 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1695 if (txq->ift_sds.ifsd_map != NULL)
1696 map = txq->ift_sds.ifsd_map[i];
1698 bus_dmamap_sync(txq->ift_desc_tag, map, BUS_DMASYNC_POSTWRITE);
1699 bus_dmamap_unload(txq->ift_desc_tag, map);
1700 bus_dmamap_destroy(txq->ift_desc_tag, map);
1701 txq->ift_sds.ifsd_map[i] = NULL;
1705 if (txq->ift_sds.ifsd_tso_map != NULL)
1706 map = txq->ift_sds.ifsd_tso_map[i];
1708 bus_dmamap_sync(txq->ift_tso_desc_tag, map,
1709 BUS_DMASYNC_POSTWRITE);
1710 bus_dmamap_unload(txq->ift_tso_desc_tag, map);
1711 bus_dmamap_destroy(txq->ift_tso_desc_tag, map);
1712 txq->ift_sds.ifsd_tso_map[i] = NULL;
1717 iflib_txq_destroy(iflib_txq_t txq)
1719 if_ctx_t ctx = txq->ift_ctx;
1721 for (int i = 0; i < txq->ift_size; i++)
1722 iflib_txsd_destroy(ctx, txq, i);
1723 if (txq->ift_sds.ifsd_map != NULL) {
1724 free(txq->ift_sds.ifsd_map, M_IFLIB);
1725 txq->ift_sds.ifsd_map = NULL;
1727 if (txq->ift_sds.ifsd_tso_map != NULL) {
1728 free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1729 txq->ift_sds.ifsd_tso_map = NULL;
1731 if (txq->ift_sds.ifsd_m != NULL) {
1732 free(txq->ift_sds.ifsd_m, M_IFLIB);
1733 txq->ift_sds.ifsd_m = NULL;
1735 if (txq->ift_desc_tag != NULL) {
1736 bus_dma_tag_destroy(txq->ift_desc_tag);
1737 txq->ift_desc_tag = NULL;
1739 if (txq->ift_tso_desc_tag != NULL) {
1740 bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1741 txq->ift_tso_desc_tag = NULL;
1746 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1750 mp = &txq->ift_sds.ifsd_m[i];
1754 if (txq->ift_sds.ifsd_map != NULL) {
1755 bus_dmamap_sync(txq->ift_desc_tag,
1756 txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1757 bus_dmamap_unload(txq->ift_desc_tag, txq->ift_sds.ifsd_map[i]);
1759 if (txq->ift_sds.ifsd_tso_map != NULL) {
1760 bus_dmamap_sync(txq->ift_tso_desc_tag,
1761 txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1762 bus_dmamap_unload(txq->ift_tso_desc_tag,
1763 txq->ift_sds.ifsd_tso_map[i]);
1766 DBG_COUNTER_INC(tx_frees);
1771 iflib_txq_setup(iflib_txq_t txq)
1773 if_ctx_t ctx = txq->ift_ctx;
1774 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1775 if_shared_ctx_t sctx = ctx->ifc_sctx;
1776 iflib_dma_info_t di;
1779 /* Set number of descriptors available */
1780 txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1781 /* XXX make configurable */
1782 txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1785 txq->ift_cidx_processed = 0;
1786 txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1787 txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1789 for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1790 bzero((void *)di->idi_vaddr, di->idi_size);
1792 IFDI_TXQ_SETUP(ctx, txq->ift_id);
1793 for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1794 bus_dmamap_sync(di->idi_tag, di->idi_map,
1795 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1799 /*********************************************************************
1801 * Allocate memory for rx_buffer structures. Since we use one
1802 * rx_buffer per received packet, the maximum number of rx_buffer's
1803 * that we'll need is equal to the number of receive descriptors
1804 * that we've allocated.
1806 **********************************************************************/
1808 iflib_rxsd_alloc(iflib_rxq_t rxq)
1810 if_ctx_t ctx = rxq->ifr_ctx;
1811 if_shared_ctx_t sctx = ctx->ifc_sctx;
1812 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1813 device_t dev = ctx->ifc_dev;
1817 MPASS(scctx->isc_nrxd[0] > 0);
1818 MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1821 for (int i = 0; i < rxq->ifr_nfl; i++, fl++) {
1822 fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1823 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1824 1, 0, /* alignment, bounds */
1825 BUS_SPACE_MAXADDR, /* lowaddr */
1826 BUS_SPACE_MAXADDR, /* highaddr */
1827 NULL, NULL, /* filter, filterarg */
1828 sctx->isc_rx_maxsize, /* maxsize */
1829 sctx->isc_rx_nsegments, /* nsegments */
1830 sctx->isc_rx_maxsegsize, /* maxsegsize */
1832 NULL, /* lockfunc */
1836 device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1840 if (!(fl->ifl_sds.ifsd_m =
1841 (struct mbuf **) malloc(sizeof(struct mbuf *) *
1842 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1843 device_printf(dev, "Unable to allocate tx_buffer memory\n");
1847 if (!(fl->ifl_sds.ifsd_cl =
1848 (caddr_t *) malloc(sizeof(caddr_t) *
1849 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1850 device_printf(dev, "Unable to allocate tx_buffer memory\n");
1855 if (!(fl->ifl_sds.ifsd_ba =
1856 (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1857 scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1858 device_printf(dev, "Unable to allocate rx bus addr memory\n");
1863 /* Create the descriptor buffer dma maps */
1864 if (!(fl->ifl_sds.ifsd_map =
1865 (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1866 device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1871 for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1872 err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1874 device_printf(dev, "Unable to create RX buffer DMA map\n");
1882 iflib_rx_structures_free(ctx);
1888 * Internal service routines
1891 struct rxq_refill_cb_arg {
1893 bus_dma_segment_t seg;
1898 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1900 struct rxq_refill_cb_arg *cb_arg = arg;
1902 cb_arg->error = error;
1903 cb_arg->seg = segs[0];
1904 cb_arg->nseg = nseg;
1908 * rxq_refill - refill an rxq free-buffer list
1909 * @ctx: the iflib context
1910 * @rxq: the free-list to refill
1911 * @n: the number of new buffers to allocate
1913 * (Re)populate an rxq free-buffer list with up to @n new packet buffers.
1914 * The caller must assure that @n does not exceed the queue's capacity.
1917 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1919 struct if_rxd_update iru;
1920 struct rxq_refill_cb_arg cb_arg;
1924 bus_dmamap_t *sd_map;
1925 bus_addr_t bus_addr, *sd_ba;
1926 int err, frag_idx, i, idx, n, pidx;
1929 sd_m = fl->ifl_sds.ifsd_m;
1930 sd_map = fl->ifl_sds.ifsd_map;
1931 sd_cl = fl->ifl_sds.ifsd_cl;
1932 sd_ba = fl->ifl_sds.ifsd_ba;
1933 pidx = fl->ifl_pidx;
1935 frag_idx = fl->ifl_fragidx;
1936 credits = fl->ifl_credits;
1941 MPASS(credits + n <= fl->ifl_size);
1943 if (pidx < fl->ifl_cidx)
1944 MPASS(pidx + n <= fl->ifl_cidx);
1945 if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1946 MPASS(fl->ifl_gen == 0);
1947 if (pidx > fl->ifl_cidx)
1948 MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1950 DBG_COUNTER_INC(fl_refills);
1952 DBG_COUNTER_INC(fl_refills_large);
1953 iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1956 * We allocate an uninitialized mbuf + cluster, mbuf is
1957 * initialized after rx.
1959 * If the cluster is still set then we know a minimum sized packet was received
1961 bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
1964 bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1965 MPASS(frag_idx >= 0);
1966 if ((cl = sd_cl[frag_idx]) == NULL) {
1967 if ((cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1971 MPASS(sd_map != NULL);
1972 err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1973 cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
1975 if (err != 0 || cb_arg.error) {
1979 if (fl->ifl_zone == zone_pack)
1980 uma_zfree(fl->ifl_zone, cl);
1984 bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1985 BUS_DMASYNC_PREREAD);
1986 sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
1987 sd_cl[frag_idx] = cl;
1989 fl->ifl_cl_enqueued++;
1992 bus_addr = sd_ba[frag_idx];
1995 MPASS(sd_m[frag_idx] == NULL);
1996 if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
2000 bit_set(fl->ifl_rx_bitmap, frag_idx);
2002 fl->ifl_m_enqueued++;
2005 DBG_COUNTER_INC(rx_allocs);
2006 fl->ifl_rxd_idxs[i] = frag_idx;
2007 fl->ifl_bus_addrs[i] = bus_addr;
2008 fl->ifl_vm_addrs[i] = cl;
2011 MPASS(credits <= fl->ifl_size);
2012 if (++idx == fl->ifl_size) {
2016 if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2017 iru.iru_pidx = pidx;
2019 ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2023 fl->ifl_credits = credits;
2028 iru.iru_pidx = pidx;
2030 ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2032 fl->ifl_credits = credits;
2034 DBG_COUNTER_INC(rxd_flush);
2035 if (fl->ifl_pidx == 0)
2036 pidx = fl->ifl_size - 1;
2038 pidx = fl->ifl_pidx - 1;
2040 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2041 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2042 ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
2043 fl->ifl_fragidx = frag_idx;
2046 static __inline void
2047 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
2049 /* we avoid allowing pidx to catch up with cidx as it confuses ixl */
2050 int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2052 int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2055 MPASS(fl->ifl_credits <= fl->ifl_size);
2056 MPASS(reclaimable == delta);
2058 if (reclaimable > 0)
2059 _iflib_fl_refill(ctx, fl, min(max, reclaimable));
2063 iflib_in_detach(if_ctx_t ctx)
2067 in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2073 iflib_fl_bufs_free(iflib_fl_t fl)
2075 iflib_dma_info_t idi = fl->ifl_ifdi;
2076 bus_dmamap_t sd_map;
2079 for (i = 0; i < fl->ifl_size; i++) {
2080 struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2081 caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2083 if (*sd_cl != NULL) {
2084 sd_map = fl->ifl_sds.ifsd_map[i];
2085 bus_dmamap_sync(fl->ifl_desc_tag, sd_map,
2086 BUS_DMASYNC_POSTREAD);
2087 bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
2089 uma_zfree(fl->ifl_zone, *sd_cl);
2090 // XXX: Should this get moved out?
2091 if (iflib_in_detach(fl->ifl_rxq->ifr_ctx))
2092 bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
2093 if (*sd_m != NULL) {
2094 m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2095 uma_zfree(zone_mbuf, *sd_m);
2098 MPASS(*sd_cl == NULL);
2099 MPASS(*sd_m == NULL);
2102 fl->ifl_m_dequeued++;
2103 fl->ifl_cl_dequeued++;
2109 for (i = 0; i < fl->ifl_size; i++) {
2110 MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2111 MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2115 * Reset free list values
2117 fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2118 bzero(idi->idi_vaddr, idi->idi_size);
2121 /*********************************************************************
2123 * Initialize a receive ring and its buffers.
2125 **********************************************************************/
2127 iflib_fl_setup(iflib_fl_t fl)
2129 iflib_rxq_t rxq = fl->ifl_rxq;
2130 if_ctx_t ctx = rxq->ifr_ctx;
2131 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2133 bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2135 ** Free current RX buffer structs and their mbufs
2137 iflib_fl_bufs_free(fl);
2138 /* Now replenish the mbufs */
2139 MPASS(fl->ifl_credits == 0);
2141 * XXX don't set the max_frame_size to larger
2142 * than the hardware can handle
2144 if (sctx->isc_max_frame_size <= 2048)
2145 fl->ifl_buf_size = MCLBYTES;
2146 #ifndef CONTIGMALLOC_WORKS
2148 fl->ifl_buf_size = MJUMPAGESIZE;
2150 else if (sctx->isc_max_frame_size <= 4096)
2151 fl->ifl_buf_size = MJUMPAGESIZE;
2152 else if (sctx->isc_max_frame_size <= 9216)
2153 fl->ifl_buf_size = MJUM9BYTES;
2155 fl->ifl_buf_size = MJUM16BYTES;
2157 if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2158 ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2159 fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2160 fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2163 /* avoid pre-allocating zillions of clusters to an idle card
2164 * potentially speeding up attach
2166 _iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2167 MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2168 if (min(128, fl->ifl_size) != fl->ifl_credits)
2174 MPASS(fl->ifl_ifdi != NULL);
2175 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2176 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2180 /*********************************************************************
2182 * Free receive ring data structures
2184 **********************************************************************/
2186 iflib_rx_sds_free(iflib_rxq_t rxq)
2191 if (rxq->ifr_fl != NULL) {
2192 for (i = 0; i < rxq->ifr_nfl; i++) {
2193 fl = &rxq->ifr_fl[i];
2194 if (fl->ifl_desc_tag != NULL) {
2195 if (fl->ifl_sds.ifsd_map != NULL) {
2196 for (j = 0; j < fl->ifl_size; j++) {
2197 if (fl->ifl_sds.ifsd_map[j] ==
2202 fl->ifl_sds.ifsd_map[j],
2203 BUS_DMASYNC_POSTREAD);
2206 fl->ifl_sds.ifsd_map[j]);
2209 bus_dma_tag_destroy(fl->ifl_desc_tag);
2210 fl->ifl_desc_tag = NULL;
2212 free(fl->ifl_sds.ifsd_m, M_IFLIB);
2213 free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2214 free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2215 free(fl->ifl_sds.ifsd_map, M_IFLIB);
2216 fl->ifl_sds.ifsd_m = NULL;
2217 fl->ifl_sds.ifsd_cl = NULL;
2218 fl->ifl_sds.ifsd_ba = NULL;
2219 fl->ifl_sds.ifsd_map = NULL;
2221 free(rxq->ifr_fl, M_IFLIB);
2223 rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2228 * MI independent logic
2232 iflib_timer(void *arg)
2234 iflib_txq_t txq = arg;
2235 if_ctx_t ctx = txq->ift_ctx;
2236 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2237 uint64_t this_tick = ticks;
2238 uint32_t reset_on = hz / 2;
2240 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2243 ** Check on the state of the TX queue(s), this
2244 ** can be done without the lock because its RO
2245 ** and the HUNG state will be static if set.
2247 if (this_tick - txq->ift_last_timer_tick >= hz / 2) {
2248 txq->ift_last_timer_tick = this_tick;
2249 IFDI_TIMER(ctx, txq->ift_id);
2250 if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2251 ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2252 (sctx->isc_pause_frames == 0)))
2255 if (ifmp_ring_is_stalled(txq->ift_br))
2256 txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2257 txq->ift_cleaned_prev = txq->ift_cleaned;
2260 if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
2261 iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
2263 /* handle any laggards */
2264 if (txq->ift_db_pending)
2265 GROUPTASK_ENQUEUE(&txq->ift_task);
2267 sctx->isc_pause_frames = 0;
2268 if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2269 callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
2272 device_printf(ctx->ifc_dev, "TX(%d) desc avail = %d, pidx = %d\n",
2273 txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2275 if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2276 ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2277 iflib_admin_intr_deferred(ctx);
2282 iflib_init_locked(if_ctx_t ctx)
2284 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2285 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2286 if_t ifp = ctx->ifc_ifp;
2290 int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2293 if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2294 IFDI_INTR_DISABLE(ctx);
2296 tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2297 tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2298 /* Set hardware offload abilities */
2299 if_clearhwassist(ifp);
2300 if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2301 if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2302 if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2303 if_sethwassistbits(ifp, tx_ip6_csum_flags, 0);
2304 if (if_getcapenable(ifp) & IFCAP_TSO4)
2305 if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2306 if (if_getcapenable(ifp) & IFCAP_TSO6)
2307 if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2309 for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2311 callout_stop(&txq->ift_timer);
2312 CALLOUT_UNLOCK(txq);
2313 iflib_netmap_txq_init(ctx, txq);
2316 i = if_getdrvflags(ifp);
2319 MPASS(if_getdrvflags(ifp) == i);
2320 for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2321 /* XXX this should really be done on a per-queue basis */
2322 if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2323 MPASS(rxq->ifr_id == i);
2324 iflib_netmap_rxq_init(ctx, rxq);
2327 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2328 if (iflib_fl_setup(fl)) {
2329 device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2335 if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2336 IFDI_INTR_ENABLE(ctx);
2337 txq = ctx->ifc_txqs;
2338 for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2339 callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2340 txq->ift_timer.c_cpu);
2344 iflib_media_change(if_t ifp)
2346 if_ctx_t ctx = if_getsoftc(ifp);
2350 if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2351 iflib_init_locked(ctx);
2357 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2359 if_ctx_t ctx = if_getsoftc(ifp);
2362 IFDI_UPDATE_ADMIN_STATUS(ctx);
2363 IFDI_MEDIA_STATUS(ctx, ifmr);
2368 iflib_stop(if_ctx_t ctx)
2370 iflib_txq_t txq = ctx->ifc_txqs;
2371 iflib_rxq_t rxq = ctx->ifc_rxqs;
2372 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2373 if_shared_ctx_t sctx = ctx->ifc_sctx;
2374 iflib_dma_info_t di;
2378 /* Tell the stack that the interface is no longer active */
2379 if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2381 IFDI_INTR_DISABLE(ctx);
2386 iflib_debug_reset();
2387 /* Wait for current tx queue users to exit to disarm watchdog timer. */
2388 for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2389 /* make sure all transmitters have completed before proceeding XXX */
2392 callout_stop(&txq->ift_timer);
2393 CALLOUT_UNLOCK(txq);
2395 /* clean any enqueued buffers */
2396 iflib_ifmp_purge(txq);
2397 /* Free any existing tx buffers. */
2398 for (j = 0; j < txq->ift_size; j++) {
2399 iflib_txsd_free(ctx, txq, j);
2401 txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2402 txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2403 txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2404 txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2405 txq->ift_pullups = 0;
2406 ifmp_ring_reset_stats(txq->ift_br);
2407 for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2408 bzero((void *)di->idi_vaddr, di->idi_size);
2410 for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2411 /* make sure all transmitters have completed before proceeding XXX */
2413 rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2414 for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2415 bzero((void *)di->idi_vaddr, di->idi_size);
2416 /* also resets the free lists pidx/cidx */
2417 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2418 iflib_fl_bufs_free(fl);
2422 static inline caddr_t
2423 calc_next_rxd(iflib_fl_t fl, int cidx)
2427 caddr_t start, end, cur, next;
2429 nrxd = fl->ifl_size;
2430 size = fl->ifl_rxd_size;
2431 start = fl->ifl_ifdi->idi_vaddr;
2433 if (__predict_false(size == 0))
2435 cur = start + size*cidx;
2436 end = start + size*nrxd;
2437 next = CACHE_PTR_NEXT(cur);
2438 return (next < end ? next : start);
2442 prefetch_pkts(iflib_fl_t fl, int cidx)
2445 int nrxd = fl->ifl_size;
2449 nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2450 prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2451 prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2452 next_rxd = calc_next_rxd(fl, cidx);
2454 prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2455 prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2456 prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2457 prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2458 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2459 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2460 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2461 prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2465 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2470 iflib_dma_info_t di;
2474 flid = irf->irf_flid;
2475 cidx = irf->irf_idx;
2476 fl = &rxq->ifr_fl[flid];
2478 sd->ifsd_cidx = cidx;
2479 sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2480 sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2483 fl->ifl_m_dequeued++;
2485 if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2486 prefetch_pkts(fl, cidx);
2487 next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2488 prefetch(&fl->ifl_sds.ifsd_map[next]);
2489 map = fl->ifl_sds.ifsd_map[cidx];
2491 next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2493 /* not valid assert if bxe really does SGE from non-contiguous elements */
2494 MPASS(fl->ifl_cidx == cidx);
2495 bus_dmamap_sync(fl->ifl_desc_tag, map, BUS_DMASYNC_POSTREAD);
2497 bus_dmamap_unload(fl->ifl_desc_tag, map);
2498 fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2499 if (__predict_false(fl->ifl_cidx == 0))
2501 bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2502 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2503 bit_clear(fl->ifl_rx_bitmap, cidx);
2506 static struct mbuf *
2507 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2509 int i, padlen , flags;
2510 struct mbuf *m, *mh, *mt;
2516 rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2518 MPASS(*sd->ifsd_cl != NULL);
2519 MPASS(*sd->ifsd_m != NULL);
2521 /* Don't include zero-length frags */
2522 if (ri->iri_frags[i].irf_len == 0) {
2523 /* XXX we can save the cluster here, but not the mbuf */
2524 m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2525 m_free(*sd->ifsd_m);
2532 flags = M_PKTHDR|M_EXT;
2534 padlen = ri->iri_pad;
2539 /* assuming padding is only on the first fragment */
2543 *sd->ifsd_cl = NULL;
2545 /* Can these two be made one ? */
2546 m_init(m, M_NOWAIT, MT_DATA, flags);
2547 m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2549 * These must follow m_init and m_cljset
2551 m->m_data += padlen;
2552 ri->iri_len -= padlen;
2553 m->m_len = ri->iri_frags[i].irf_len;
2554 } while (++i < ri->iri_nfrags);
2560 * Process one software descriptor
2562 static struct mbuf *
2563 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2568 /* should I merge this back in now that the two paths are basically duplicated? */
2569 if (ri->iri_nfrags == 1 &&
2570 ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2571 rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2574 m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2575 #ifndef __NO_STRICT_ALIGNMENT
2579 memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2580 bus_dmamap_sync(rxq->ifr_fl->ifl_desc_tag,
2581 rxq->ifr_fl->ifl_sds.ifsd_map[ri->iri_frags[0].irf_idx],
2582 BUS_DMASYNC_PREREAD);
2583 m->m_len = ri->iri_frags[0].irf_len;
2585 m = assemble_segments(rxq, ri, &sd);
2587 m->m_pkthdr.len = ri->iri_len;
2588 m->m_pkthdr.rcvif = ri->iri_ifp;
2589 m->m_flags |= ri->iri_flags;
2590 m->m_pkthdr.ether_vtag = ri->iri_vtag;
2591 m->m_pkthdr.flowid = ri->iri_flowid;
2592 M_HASHTYPE_SET(m, ri->iri_rsstype);
2593 m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2594 m->m_pkthdr.csum_data = ri->iri_csum_data;
2598 #if defined(INET6) || defined(INET)
2600 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2602 CURVNET_SET(lc->ifp->if_vnet);
2604 *v6 = VNET(ip6_forwarding);
2607 *v4 = VNET(ipforwarding);
2613 * Returns true if it's possible this packet could be LROed.
2614 * if it returns false, it is guaranteed that tcp_lro_rx()
2615 * would not return zero.
2618 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2620 struct ether_header *eh;
2623 eh = mtod(m, struct ether_header *);
2624 eh_type = ntohs(eh->ether_type);
2627 case ETHERTYPE_IPV6:
2628 return !v6_forwarding;
2632 return !v4_forwarding;
2640 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2646 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2648 if_ctx_t ctx = rxq->ifr_ctx;
2649 if_shared_ctx_t sctx = ctx->ifc_sctx;
2650 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2651 iflib_dma_info_t di;
2654 struct if_rxd_info ri;
2655 int err, budget_left, rx_bytes, rx_pkts;
2659 bool v4_forwarding, v6_forwarding, lro_possible;
2662 * XXX early demux data packets so that if_input processing only handles
2663 * acks in interrupt context
2665 struct mbuf *m, *mh, *mt, *mf;
2667 lro_possible = v4_forwarding = v6_forwarding = false;
2671 rx_pkts = rx_bytes = 0;
2672 if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2673 cidxp = &rxq->ifr_cq_cidx;
2675 cidxp = &rxq->ifr_fl[0].ifl_cidx;
2676 if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2677 for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2678 __iflib_fl_refill_lt(ctx, fl, budget + 8);
2679 DBG_COUNTER_INC(rx_unavail);
2683 for (budget_left = budget; budget_left > 0 && avail > 0;) {
2684 if (__predict_false(!CTX_ACTIVE(ctx))) {
2685 DBG_COUNTER_INC(rx_ctx_inactive);
2689 * Reset client set fields to their default values
2692 ri.iri_qsidx = rxq->ifr_id;
2693 ri.iri_cidx = *cidxp;
2695 ri.iri_frags = rxq->ifr_frags;
2696 di = rxq->ifr_fl[rxq->ifr_frags[0].irf_flid].ifl_ifdi;
2697 bus_dmamap_sync(di->idi_tag, di->idi_map,
2698 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2699 err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2703 if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2704 *cidxp = ri.iri_cidx;
2705 /* Update our consumer index */
2706 /* XXX NB: shurd - check if this is still safe */
2707 while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2708 rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2709 rxq->ifr_cq_gen = 0;
2711 /* was this only a completion queue message? */
2712 if (__predict_false(ri.iri_nfrags == 0))
2715 MPASS(ri.iri_nfrags != 0);
2716 MPASS(ri.iri_len != 0);
2718 /* will advance the cidx on the corresponding free lists */
2719 m = iflib_rxd_pkt_get(rxq, &ri);
2722 if (avail == 0 && budget_left)
2723 avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2725 if (__predict_false(m == NULL)) {
2726 DBG_COUNTER_INC(rx_mbuf_null);
2729 /* imm_pkt: -- cxgb */
2737 /* make sure that we can refill faster than drain */
2738 for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2739 __iflib_fl_refill_lt(ctx, fl, budget + 8);
2741 lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2743 iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2745 while (mh != NULL) {
2748 m->m_nextpkt = NULL;
2749 #ifndef __NO_STRICT_ALIGNMENT
2750 if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2753 rx_bytes += m->m_pkthdr.len;
2755 #if defined(INET6) || defined(INET)
2757 if (!lro_possible) {
2758 lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2759 if (lro_possible && mf != NULL) {
2760 ifp->if_input(ifp, mf);
2761 DBG_COUNTER_INC(rx_if_input);
2765 if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2766 (CSUM_L4_CALC|CSUM_L4_VALID)) {
2767 if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2773 ifp->if_input(ifp, m);
2774 DBG_COUNTER_INC(rx_if_input);
2785 ifp->if_input(ifp, mf);
2786 DBG_COUNTER_INC(rx_if_input);
2789 if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2790 if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2793 * Flush any outstanding LRO work
2795 #if defined(INET6) || defined(INET)
2796 tcp_lro_flush_all(&rxq->ifr_lc);
2800 return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2803 ctx->ifc_flags |= IFC_DO_RESET;
2804 iflib_admin_intr_deferred(ctx);
2809 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2810 static inline qidx_t
2811 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2813 qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2814 qidx_t minthresh = txq->ift_size / 8;
2815 if (in_use > 4*minthresh)
2816 return (notify_count);
2817 if (in_use > 2*minthresh)
2818 return (notify_count >> 1);
2819 if (in_use > minthresh)
2820 return (notify_count >> 3);
2824 static inline qidx_t
2825 txq_max_rs_deferred(iflib_txq_t txq)
2827 qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2828 qidx_t minthresh = txq->ift_size / 8;
2829 if (txq->ift_in_use > 4*minthresh)
2830 return (notify_count);
2831 if (txq->ift_in_use > 2*minthresh)
2832 return (notify_count >> 1);
2833 if (txq->ift_in_use > minthresh)
2834 return (notify_count >> 2);
2838 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2839 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2841 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2842 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2843 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2845 /* forward compatibility for cxgb */
2846 #define FIRST_QSET(ctx) 0
2847 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2848 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2849 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2850 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2852 /* XXX we should be setting this to something other than zero */
2853 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2854 #define MAX_TX_DESC(ctx) max((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
2855 (ctx)->ifc_softc_ctx.isc_tx_nsegments)
2858 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2864 max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2865 if (ring || txq->ift_db_pending >= max) {
2866 dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2867 ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2868 txq->ift_db_pending = txq->ift_npending = 0;
2876 print_pkt(if_pkt_info_t pi)
2878 printf("pi len: %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2879 pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2880 printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2881 pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2882 printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2883 pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2887 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2888 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
2889 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2890 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
2893 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2895 if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2896 struct ether_vlan_header *eh;
2900 if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2901 M_WRITABLE(m) == 0) {
2902 if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2906 DBG_COUNTER_INC(tx_frees);
2912 * Determine where frame payload starts.
2913 * Jump over vlan headers if already present,
2914 * helpful for QinQ too.
2916 if (__predict_false(m->m_len < sizeof(*eh))) {
2918 if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2921 eh = mtod(m, struct ether_vlan_header *);
2922 if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2923 pi->ipi_etype = ntohs(eh->evl_proto);
2924 pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2926 pi->ipi_etype = ntohs(eh->evl_encap_proto);
2927 pi->ipi_ehdrlen = ETHER_HDR_LEN;
2930 switch (pi->ipi_etype) {
2935 struct ip *ip = NULL;
2936 struct tcphdr *th = NULL;
2939 minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2940 if (__predict_false(m->m_len < minthlen)) {
2942 * if this code bloat is causing too much of a hit
2943 * move it to a separate function and mark it noinline
2945 if (m->m_len == pi->ipi_ehdrlen) {
2948 if (n->m_len >= sizeof(*ip)) {
2949 ip = (struct ip *)n->m_data;
2950 if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2951 th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2954 if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2956 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2960 if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2962 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2963 if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2964 th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2967 ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2968 if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2969 th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2971 pi->ipi_ip_hlen = ip->ip_hl << 2;
2972 pi->ipi_ipproto = ip->ip_p;
2973 pi->ipi_flags |= IPI_TX_IPV4;
2975 /* TCP checksum offload may require TCP header length */
2976 if (IS_TX_OFFLOAD4(pi)) {
2977 if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
2978 if (__predict_false(th == NULL)) {
2980 if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2982 th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2984 pi->ipi_tcp_hflags = th->th_flags;
2985 pi->ipi_tcp_hlen = th->th_off << 2;
2986 pi->ipi_tcp_seq = th->th_seq;
2989 if (__predict_false(ip->ip_p != IPPROTO_TCP))
2992 * TSO always requires hardware checksum offload.
2994 pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
2995 th->th_sum = in_pseudo(ip->ip_src.s_addr,
2996 ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2997 pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2998 if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3000 ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3004 if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3011 case ETHERTYPE_IPV6:
3013 struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3015 pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3017 if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3019 if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3022 th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3024 /* XXX-BZ this will go badly in case of ext hdrs. */
3025 pi->ipi_ipproto = ip6->ip6_nxt;
3026 pi->ipi_flags |= IPI_TX_IPV6;
3028 /* TCP checksum offload may require TCP header length */
3029 if (IS_TX_OFFLOAD6(pi)) {
3030 if (pi->ipi_ipproto == IPPROTO_TCP) {
3031 if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3033 if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3036 pi->ipi_tcp_hflags = th->th_flags;
3037 pi->ipi_tcp_hlen = th->th_off << 2;
3038 pi->ipi_tcp_seq = th->th_seq;
3041 if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3044 * TSO always requires hardware checksum offload.
3046 pi->ipi_csum_flags |= CSUM_IP6_TCP;
3047 th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3048 pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3055 pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3056 pi->ipi_ip_hlen = 0;
3065 * If dodgy hardware rejects the scatter gather chain we've handed it
3066 * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3069 static __noinline struct mbuf *
3070 iflib_remove_mbuf(iflib_txq_t txq)
3073 struct mbuf *m, **ifsd_m;
3075 ifsd_m = txq->ift_sds.ifsd_m;
3076 ntxd = txq->ift_size;
3077 pidx = txq->ift_pidx & (ntxd - 1);
3078 ifsd_m = txq->ift_sds.ifsd_m;
3080 ifsd_m[pidx] = NULL;
3081 bus_dmamap_unload(txq->ift_desc_tag, txq->ift_sds.ifsd_map[pidx]);
3082 if (txq->ift_sds.ifsd_tso_map != NULL)
3083 bus_dmamap_unload(txq->ift_tso_desc_tag,
3084 txq->ift_sds.ifsd_tso_map[pidx]);
3086 txq->ift_dequeued++;
3091 static inline caddr_t
3092 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3096 caddr_t start, end, cur, next;
3098 ntxd = txq->ift_size;
3099 size = txq->ift_txd_size[qid];
3100 start = txq->ift_ifdi[qid].idi_vaddr;
3102 if (__predict_false(size == 0))
3104 cur = start + size*cidx;
3105 end = start + size*ntxd;
3106 next = CACHE_PTR_NEXT(cur);
3107 return (next < end ? next : start);
3111 * Pad an mbuf to ensure a minimum ethernet frame size.
3112 * min_frame_size is the frame size (less CRC) to pad the mbuf to
3114 static __noinline int
3115 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3118 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3119 * and ARP message is the smallest common payload I can think of
3121 static char pad[18]; /* just zeros */
3123 struct mbuf *new_head;
3125 if (!M_WRITABLE(*m_head)) {
3126 new_head = m_dup(*m_head, M_NOWAIT);
3127 if (new_head == NULL) {
3129 device_printf(dev, "cannot pad short frame, m_dup() failed");
3130 DBG_COUNTER_INC(encap_pad_mbuf_fail);
3131 DBG_COUNTER_INC(tx_frees);
3138 for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3139 n > 0; n -= sizeof(pad))
3140 if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3145 device_printf(dev, "cannot pad short frame\n");
3146 DBG_COUNTER_INC(encap_pad_mbuf_fail);
3147 DBG_COUNTER_INC(tx_frees);
3155 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3158 if_shared_ctx_t sctx;
3159 if_softc_ctx_t scctx;
3160 bus_dma_segment_t *segs;
3161 struct mbuf *m_head, **ifsd_m;
3164 struct if_pkt_info pi;
3166 int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3167 bus_dma_tag_t desc_tag;
3170 sctx = ctx->ifc_sctx;
3171 scctx = &ctx->ifc_softc_ctx;
3172 segs = txq->ift_segs;
3173 ntxd = txq->ift_size;
3178 * If we're doing TSO the next descriptor to clean may be quite far ahead
3180 cidx = txq->ift_cidx;
3181 pidx = txq->ift_pidx;
3182 if (ctx->ifc_flags & IFC_PREFETCH) {
3183 next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3184 if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3185 next_txd = calc_next_txd(txq, cidx, 0);
3189 /* prefetch the next cache line of mbuf pointers and flags */
3190 prefetch(&txq->ift_sds.ifsd_m[next]);
3191 prefetch(&txq->ift_sds.ifsd_map[next]);
3192 next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3194 map = txq->ift_sds.ifsd_map[pidx];
3195 ifsd_m = txq->ift_sds.ifsd_m;
3197 if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3198 desc_tag = txq->ift_tso_desc_tag;
3199 max_segs = scctx->isc_tx_tso_segments_max;
3200 map = txq->ift_sds.ifsd_tso_map[pidx];
3201 MPASS(desc_tag != NULL);
3202 MPASS(max_segs > 0);
3204 desc_tag = txq->ift_desc_tag;
3205 max_segs = scctx->isc_tx_nsegments;
3206 map = txq->ift_sds.ifsd_map[pidx];
3208 if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3209 __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3210 err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3212 DBG_COUNTER_INC(encap_txd_encap_fail);
3219 pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3221 pi.ipi_qsidx = txq->ift_id;
3222 pi.ipi_len = m_head->m_pkthdr.len;
3223 pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3224 pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3226 /* deliberate bitwise OR to make one condition */
3227 if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3228 if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3229 DBG_COUNTER_INC(encap_txd_encap_fail);
3236 err = bus_dmamap_load_mbuf_sg(desc_tag, map, m_head, segs, &nsegs,
3239 if (__predict_false(err)) {
3242 /* try collapse once and defrag once */
3244 m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3245 /* try defrag if collapsing fails */
3250 txq->ift_mbuf_defrag++;
3251 m_head = m_defrag(*m_headp, M_NOWAIT);
3254 if (__predict_false(m_head == NULL))
3260 txq->ift_no_tx_dma_setup++;
3263 txq->ift_no_tx_dma_setup++;
3265 DBG_COUNTER_INC(tx_frees);
3269 txq->ift_map_failed++;
3270 DBG_COUNTER_INC(encap_load_mbuf_fail);
3271 DBG_COUNTER_INC(encap_txd_encap_fail);
3274 ifsd_m[pidx] = m_head;
3276 * XXX assumes a 1 to 1 relationship between segments and
3277 * descriptors - this does not hold true on all drivers, e.g.
3280 if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3281 txq->ift_no_desc_avail++;
3282 bus_dmamap_unload(desc_tag, map);
3283 DBG_COUNTER_INC(encap_txq_avail_fail);
3284 DBG_COUNTER_INC(encap_txd_encap_fail);
3285 if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3286 GROUPTASK_ENQUEUE(&txq->ift_task);
3290 * On Intel cards we can greatly reduce the number of TX interrupts
3291 * we see by only setting report status on every Nth descriptor.
3292 * However, this also means that the driver will need to keep track
3293 * of the descriptors that RS was set on to check them for the DD bit.
3295 txq->ift_rs_pending += nsegs + 1;
3296 if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3297 iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3298 pi.ipi_flags |= IPI_TX_INTR;
3299 txq->ift_rs_pending = 0;
3303 pi.ipi_nsegs = nsegs;
3305 MPASS(pidx >= 0 && pidx < txq->ift_size);
3309 bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3310 if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3311 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3312 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3313 DBG_COUNTER_INC(tx_encap);
3314 MPASS(pi.ipi_new_pidx < txq->ift_size);
3316 ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3317 if (pi.ipi_new_pidx < pi.ipi_pidx) {
3318 ndesc += txq->ift_size;
3322 * drivers can need as many as
3325 MPASS(ndesc <= pi.ipi_nsegs + 2);
3326 MPASS(pi.ipi_new_pidx != pidx);
3328 txq->ift_in_use += ndesc;
3331 * We update the last software descriptor again here because there may
3332 * be a sentinel and/or there may be more mbufs than segments
3334 txq->ift_pidx = pi.ipi_new_pidx;
3335 txq->ift_npending += pi.ipi_ndescs;
3337 *m_headp = m_head = iflib_remove_mbuf(txq);
3339 txq->ift_txd_encap_efbig++;
3348 * err can't possibly be non-zero here, so we don't neet to test it
3349 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3354 txq->ift_mbuf_defrag_failed++;
3355 txq->ift_map_failed++;
3357 DBG_COUNTER_INC(tx_frees);
3359 DBG_COUNTER_INC(encap_txd_encap_fail);
3364 iflib_tx_desc_free(iflib_txq_t txq, int n)
3366 uint32_t qsize, cidx, mask, gen;
3367 struct mbuf *m, **ifsd_m;
3370 cidx = txq->ift_cidx;
3372 qsize = txq->ift_size;
3374 ifsd_m = txq->ift_sds.ifsd_m;
3375 do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3379 prefetch(ifsd_m[(cidx + 3) & mask]);
3380 prefetch(ifsd_m[(cidx + 4) & mask]);
3382 if ((m = ifsd_m[cidx]) != NULL) {
3383 prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3384 if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3385 bus_dmamap_sync(txq->ift_tso_desc_tag,
3386 txq->ift_sds.ifsd_tso_map[cidx],
3387 BUS_DMASYNC_POSTWRITE);
3388 bus_dmamap_unload(txq->ift_tso_desc_tag,
3389 txq->ift_sds.ifsd_tso_map[cidx]);
3391 bus_dmamap_sync(txq->ift_desc_tag,
3392 txq->ift_sds.ifsd_map[cidx],
3393 BUS_DMASYNC_POSTWRITE);
3394 bus_dmamap_unload(txq->ift_desc_tag,
3395 txq->ift_sds.ifsd_map[cidx]);
3397 /* XXX we don't support any drivers that batch packets yet */
3398 MPASS(m->m_nextpkt == NULL);
3400 ifsd_m[cidx] = NULL;
3402 txq->ift_dequeued++;
3404 DBG_COUNTER_INC(tx_frees);
3406 if (__predict_false(++cidx == qsize)) {
3411 txq->ift_cidx = cidx;
3416 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3419 if_ctx_t ctx = txq->ift_ctx;
3421 KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3422 MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3425 * Need a rate-limiting check so that this isn't called every time
3427 iflib_tx_credits_update(ctx, txq);
3428 reclaim = DESC_RECLAIMABLE(txq);
3430 if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3432 if (iflib_verbose_debug) {
3433 printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3434 txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3441 iflib_tx_desc_free(txq, reclaim);
3442 txq->ift_cleaned += reclaim;
3443 txq->ift_in_use -= reclaim;
3448 static struct mbuf **
3449 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3452 struct mbuf **items;
3455 next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3456 items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3458 prefetch(items[(cidx + offset) & (size-1)]);
3459 if (remaining > 1) {
3460 prefetch2cachelines(&items[next]);
3461 prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3462 prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3463 prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3465 return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3469 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3472 ifmp_ring_check_drainage(txq->ift_br, budget);
3476 iflib_txq_can_drain(struct ifmp_ring *r)
3478 iflib_txq_t txq = r->cookie;
3479 if_ctx_t ctx = txq->ift_ctx;
3481 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3482 BUS_DMASYNC_POSTREAD);
3483 return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3484 ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3488 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3490 iflib_txq_t txq = r->cookie;
3491 if_ctx_t ctx = txq->ift_ctx;
3492 struct ifnet *ifp = ctx->ifc_ifp;
3493 struct mbuf **mp, *m;
3494 int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3495 int reclaimed, err, in_use_prev, desc_used;
3496 bool do_prefetch, ring, rang;
3498 if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3499 !LINK_ACTIVE(ctx))) {
3500 DBG_COUNTER_INC(txq_drain_notready);
3503 reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3504 rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3505 avail = IDXDIFF(pidx, cidx, r->size);
3506 if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3507 DBG_COUNTER_INC(txq_drain_flushing);
3508 for (i = 0; i < avail; i++) {
3509 if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3510 m_free(r->items[(cidx + i) & (r->size-1)]);
3511 r->items[(cidx + i) & (r->size-1)] = NULL;
3516 if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3517 txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3519 callout_stop(&txq->ift_timer);
3520 CALLOUT_UNLOCK(txq);
3521 DBG_COUNTER_INC(txq_drain_oactive);
3525 txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3526 consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3527 count = MIN(avail, TX_BATCH_SIZE);
3529 if (iflib_verbose_debug)
3530 printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3531 avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3533 do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3534 avail = TXQ_AVAIL(txq);
3536 for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3537 int rem = do_prefetch ? count - i : 0;
3539 mp = _ring_peek_one(r, cidx, i, rem);
3540 MPASS(mp != NULL && *mp != NULL);
3541 if (__predict_false(*mp == (struct mbuf *)txq)) {
3546 in_use_prev = txq->ift_in_use;
3547 err = iflib_encap(txq, mp);
3548 if (__predict_false(err)) {
3549 /* no room - bail out */
3553 /* we can't send this packet - skip it */
3559 DBG_COUNTER_INC(tx_sent);
3560 bytes_sent += m->m_pkthdr.len;
3561 mcast_sent += !!(m->m_flags & M_MCAST);
3562 avail = TXQ_AVAIL(txq);
3564 txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3565 desc_used += (txq->ift_in_use - in_use_prev);
3566 ETHER_BPF_MTAP(ifp, m);
3567 if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3569 rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3572 /* deliberate use of bitwise or to avoid gratuitous short-circuit */
3573 ring = rang ? false : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3574 iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3575 if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3576 if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3578 if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3580 if (iflib_verbose_debug)
3581 printf("consumed=%d\n", consumed);
3587 iflib_txq_drain_always(struct ifmp_ring *r)
3593 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3601 txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3603 callout_stop(&txq->ift_timer);
3604 CALLOUT_UNLOCK(txq);
3606 avail = IDXDIFF(pidx, cidx, r->size);
3607 for (i = 0; i < avail; i++) {
3608 mp = _ring_peek_one(r, cidx, i, avail - i);
3609 if (__predict_false(*mp == (struct mbuf *)txq))
3612 DBG_COUNTER_INC(tx_frees);
3614 MPASS(ifmp_ring_is_stalled(r) == 0);
3619 iflib_ifmp_purge(iflib_txq_t txq)
3621 struct ifmp_ring *r;
3624 r->drain = iflib_txq_drain_free;
3625 r->can_drain = iflib_txq_drain_always;
3627 ifmp_ring_check_drainage(r, r->size);
3629 r->drain = iflib_txq_drain;
3630 r->can_drain = iflib_txq_can_drain;
3634 _task_fn_tx(void *context)
3636 iflib_txq_t txq = context;
3637 if_ctx_t ctx = txq->ift_ctx;
3638 struct ifnet *ifp = ctx->ifc_ifp;
3639 int abdicate = ctx->ifc_sysctl_tx_abdicate;
3641 #ifdef IFLIB_DIAGNOSTICS
3642 txq->ift_cpu_exec_count[curcpu]++;
3644 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3646 if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3647 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3648 BUS_DMASYNC_POSTREAD);
3649 if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3650 netmap_tx_irq(ifp, txq->ift_id);
3651 IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3655 if (ALTQ_IS_ENABLED(&ifp->if_snd))
3656 iflib_altq_if_start(ifp);
3658 if (txq->ift_db_pending)
3659 ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3661 ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3663 * When abdicating, we always need to check drainage, not just when we don't enqueue
3666 ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3667 if (ctx->ifc_flags & IFC_LEGACY)
3668 IFDI_INTR_ENABLE(ctx);
3673 IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3674 KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3679 _task_fn_rx(void *context)
3681 iflib_rxq_t rxq = context;
3682 if_ctx_t ctx = rxq->ifr_ctx;
3686 #ifdef IFLIB_DIAGNOSTICS
3687 rxq->ifr_cpu_exec_count[curcpu]++;
3689 DBG_COUNTER_INC(task_fn_rxs);
3690 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3694 if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3696 if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3701 budget = ctx->ifc_sysctl_rx_budget;
3703 budget = 16; /* XXX */
3704 if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3705 if (ctx->ifc_flags & IFC_LEGACY)
3706 IFDI_INTR_ENABLE(ctx);
3711 IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3712 KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3713 DBG_COUNTER_INC(rx_intr_enables);
3716 if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3719 GROUPTASK_ENQUEUE(&rxq->ifr_task);
3723 _task_fn_admin(void *context)
3725 if_ctx_t ctx = context;
3726 if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3729 bool oactive, running, do_reset, do_watchdog, in_detach;
3730 uint32_t reset_on = hz / 2;
3733 running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
3734 oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
3735 do_reset = (ctx->ifc_flags & IFC_DO_RESET);
3736 do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
3737 in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
3738 ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
3741 if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3747 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3749 callout_stop(&txq->ift_timer);
3750 CALLOUT_UNLOCK(txq);
3753 ctx->ifc_watchdog_events++;
3754 IFDI_WATCHDOG_RESET(ctx);
3756 IFDI_UPDATE_ADMIN_STATUS(ctx);
3757 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3760 if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP)
3761 iflib_netmap_timer_adjust(ctx, txq->ift_id, &reset_on);
3763 callout_reset_on(&txq->ift_timer, reset_on, iflib_timer, txq, txq->ift_timer.c_cpu);
3765 IFDI_LINK_INTR_ENABLE(ctx);
3767 iflib_if_init_locked(ctx);
3770 if (LINK_ACTIVE(ctx) == 0)
3772 for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3773 iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3778 _task_fn_iov(void *context)
3780 if_ctx_t ctx = context;
3782 if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
3783 !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
3787 IFDI_VFLR_HANDLE(ctx);
3792 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3795 if_int_delay_info_t info;
3798 info = (if_int_delay_info_t)arg1;
3799 ctx = info->iidi_ctx;
3800 info->iidi_req = req;
3801 info->iidi_oidp = oidp;
3803 err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3808 /*********************************************************************
3812 **********************************************************************/
3815 iflib_if_init_locked(if_ctx_t ctx)
3818 iflib_init_locked(ctx);
3823 iflib_if_init(void *arg)
3828 iflib_if_init_locked(ctx);
3833 iflib_if_transmit(if_t ifp, struct mbuf *m)
3835 if_ctx_t ctx = if_getsoftc(ifp);
3839 int abdicate = ctx->ifc_sysctl_tx_abdicate;
3841 if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3842 DBG_COUNTER_INC(tx_frees);
3847 MPASS(m->m_nextpkt == NULL);
3848 /* ALTQ-enabled interfaces always use queue 0. */
3850 if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
3851 qidx = QIDX(ctx, m);
3853 * XXX calculate buf_ring based on flowid (divvy up bits?)
3855 txq = &ctx->ifc_txqs[qidx];
3857 #ifdef DRIVER_BACKPRESSURE
3858 if (txq->ift_closed) {
3860 next = m->m_nextpkt;
3861 m->m_nextpkt = NULL;
3863 DBG_COUNTER_INC(tx_frees);
3875 next = next->m_nextpkt;
3876 } while (next != NULL);
3878 if (count > nitems(marr))
3879 if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3880 /* XXX check nextpkt */
3882 /* XXX simplify for now */
3883 DBG_COUNTER_INC(tx_frees);
3886 for (next = m, i = 0; next != NULL; i++) {
3888 next = next->m_nextpkt;
3889 mp[i]->m_nextpkt = NULL;
3892 DBG_COUNTER_INC(tx_seen);
3893 err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
3896 GROUPTASK_ENQUEUE(&txq->ift_task);
3899 GROUPTASK_ENQUEUE(&txq->ift_task);
3900 /* support forthcoming later */
3901 #ifdef DRIVER_BACKPRESSURE
3902 txq->ift_closed = TRUE;
3904 ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3906 DBG_COUNTER_INC(tx_frees);
3914 * The overall approach to integrating iflib with ALTQ is to continue to use
3915 * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
3916 * ring. Technically, when using ALTQ, queueing to an intermediate mp_ring
3917 * is redundant/unnecessary, but doing so minimizes the amount of
3918 * ALTQ-specific code required in iflib. It is assumed that the overhead of
3919 * redundantly queueing to an intermediate mp_ring is swamped by the
3920 * performance limitations inherent in using ALTQ.
3922 * When ALTQ support is compiled in, all iflib drivers will use a transmit
3923 * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
3924 * given interface. If ALTQ is enabled for an interface, then all
3925 * transmitted packets for that interface will be submitted to the ALTQ
3926 * subsystem via IFQ_ENQUEUE(). We don't use the legacy if_transmit()
3927 * implementation because it uses IFQ_HANDOFF(), which will duplicatively
3928 * update stats that the iflib machinery handles, and which is sensitve to
3929 * the disused IFF_DRV_OACTIVE flag. Additionally, iflib_altq_if_start()
3930 * will be installed as the start routine for use by ALTQ facilities that
3931 * need to trigger queue drains on a scheduled basis.
3935 iflib_altq_if_start(if_t ifp)
3937 struct ifaltq *ifq = &ifp->if_snd;
3941 IFQ_DEQUEUE_NOLOCK(ifq, m);
3943 iflib_if_transmit(ifp, m);
3944 IFQ_DEQUEUE_NOLOCK(ifq, m);
3950 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
3954 if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
3955 IFQ_ENQUEUE(&ifp->if_snd, m, err);
3957 iflib_altq_if_start(ifp);
3959 err = iflib_if_transmit(ifp, m);
3966 iflib_if_qflush(if_t ifp)
3968 if_ctx_t ctx = if_getsoftc(ifp);
3969 iflib_txq_t txq = ctx->ifc_txqs;
3973 ctx->ifc_flags |= IFC_QFLUSH;
3975 for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3976 while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3977 iflib_txq_check_drain(txq, 0);
3979 ctx->ifc_flags &= ~IFC_QFLUSH;
3983 * When ALTQ is enabled, this will also take care of purging the
3990 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3991 IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3992 IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
3993 IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM)
3996 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3998 if_ctx_t ctx = if_getsoftc(ifp);
3999 struct ifreq *ifr = (struct ifreq *)data;
4000 #if defined(INET) || defined(INET6)
4001 struct ifaddr *ifa = (struct ifaddr *)data;
4003 bool avoid_reset = FALSE;
4004 int err = 0, reinit = 0, bits;
4009 if (ifa->ifa_addr->sa_family == AF_INET)
4013 if (ifa->ifa_addr->sa_family == AF_INET6)
4017 ** Calling init results in link renegotiation,
4018 ** so we avoid doing it when possible.
4021 if_setflagbits(ifp, IFF_UP,0);
4022 if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4025 if (!(if_getflags(ifp) & IFF_NOARP))
4026 arp_ifinit(ifp, ifa);
4029 err = ether_ioctl(ifp, command, data);
4033 if (ifr->ifr_mtu == if_getmtu(ifp)) {
4037 bits = if_getdrvflags(ifp);
4038 /* stop the driver and free any clusters before proceeding */
4041 if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4043 if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4044 ctx->ifc_flags |= IFC_MULTISEG;
4046 ctx->ifc_flags &= ~IFC_MULTISEG;
4048 err = if_setmtu(ifp, ifr->ifr_mtu);
4050 iflib_init_locked(ctx);
4052 if_setdrvflags(ifp, bits);
4058 if (if_getflags(ifp) & IFF_UP) {
4059 if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4060 if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4061 (IFF_PROMISC | IFF_ALLMULTI)) {
4062 err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4066 } else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4069 ctx->ifc_if_flags = if_getflags(ifp);
4074 if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4076 IFDI_INTR_DISABLE(ctx);
4077 IFDI_MULTI_SET(ctx);
4078 IFDI_INTR_ENABLE(ctx);
4084 IFDI_MEDIA_SET(ctx);
4089 err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
4093 struct ifi2creq i2c;
4095 err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4098 if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4102 if (i2c.len > sizeof(i2c.data)) {
4107 if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4108 err = copyout(&i2c, ifr_data_get_ptr(ifr),
4114 int mask, setmask, oldmask;
4116 oldmask = if_getcapenable(ifp);
4117 mask = ifr->ifr_reqcap ^ oldmask;
4118 mask &= ctx->ifc_softc_ctx.isc_capabilities;
4121 setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4123 setmask |= (mask & IFCAP_FLAGS);
4124 setmask |= (mask & IFCAP_WOL);
4127 * If any RX csum has changed, change all the ones that
4128 * are supported by the driver.
4130 if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4131 setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4132 (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4136 * want to ensure that traffic has stopped before we change any of the flags
4140 bits = if_getdrvflags(ifp);
4141 if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4144 if_togglecapenable(ifp, setmask);
4146 if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4147 iflib_init_locked(ctx);
4149 if_setdrvflags(ifp, bits);
4156 case SIOCGPRIVATE_0:
4160 err = IFDI_PRIV_IOCTL(ctx, command, data);
4164 err = ether_ioctl(ifp, command, data);
4173 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4175 if_ctx_t ctx = if_getsoftc(ifp);
4177 return (IFDI_GET_COUNTER(ctx, cnt));
4180 /*********************************************************************
4182 * OTHER FUNCTIONS EXPORTED TO THE STACK
4184 **********************************************************************/
4187 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4189 if_ctx_t ctx = if_getsoftc(ifp);
4191 if ((void *)ctx != arg)
4194 if ((vtag == 0) || (vtag > 4095))
4198 IFDI_VLAN_REGISTER(ctx, vtag);
4199 /* Re-init to load the changes */
4200 if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4201 iflib_if_init_locked(ctx);
4206 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4208 if_ctx_t ctx = if_getsoftc(ifp);
4210 if ((void *)ctx != arg)
4213 if ((vtag == 0) || (vtag > 4095))
4217 IFDI_VLAN_UNREGISTER(ctx, vtag);
4218 /* Re-init to load the changes */
4219 if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4220 iflib_if_init_locked(ctx);
4225 iflib_led_func(void *arg, int onoff)
4230 IFDI_LED_FUNC(ctx, onoff);
4234 /*********************************************************************
4236 * BUS FUNCTION DEFINITIONS
4238 **********************************************************************/
4241 iflib_device_probe(device_t dev)
4243 pci_vendor_info_t *ent;
4245 uint16_t pci_vendor_id, pci_device_id;
4246 uint16_t pci_subvendor_id, pci_subdevice_id;
4247 uint16_t pci_rev_id;
4248 if_shared_ctx_t sctx;
4250 if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4253 pci_vendor_id = pci_get_vendor(dev);
4254 pci_device_id = pci_get_device(dev);
4255 pci_subvendor_id = pci_get_subvendor(dev);
4256 pci_subdevice_id = pci_get_subdevice(dev);
4257 pci_rev_id = pci_get_revid(dev);
4258 if (sctx->isc_parse_devinfo != NULL)
4259 sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4261 ent = sctx->isc_vendor_info;
4262 while (ent->pvi_vendor_id != 0) {
4263 if (pci_vendor_id != ent->pvi_vendor_id) {
4267 if ((pci_device_id == ent->pvi_device_id) &&
4268 ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4269 (ent->pvi_subvendor_id == 0)) &&
4270 ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4271 (ent->pvi_subdevice_id == 0)) &&
4272 ((pci_rev_id == ent->pvi_rev_id) ||
4273 (ent->pvi_rev_id == 0))) {
4275 device_set_desc_copy(dev, ent->pvi_name);
4276 /* this needs to be changed to zero if the bus probing code
4277 * ever stops re-probing on best match because the sctx
4278 * may have its values over written by register calls
4279 * in subsequent probes
4281 return (BUS_PROBE_DEFAULT);
4289 iflib_reset_qvalues(if_ctx_t ctx)
4291 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4292 if_shared_ctx_t sctx = ctx->ifc_sctx;
4293 device_t dev = ctx->ifc_dev;
4296 scctx->isc_txrx_budget_bytes_max = IFLIB_MAX_TX_BYTES;
4297 scctx->isc_tx_qdepth = IFLIB_DEFAULT_TX_QDEPTH;
4299 * XXX sanity check that ntxd & nrxd are a power of 2
4301 if (ctx->ifc_sysctl_ntxqs != 0)
4302 scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4303 if (ctx->ifc_sysctl_nrxqs != 0)
4304 scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4306 for (i = 0; i < sctx->isc_ntxqs; i++) {
4307 if (ctx->ifc_sysctl_ntxds[i] != 0)
4308 scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4310 scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4313 for (i = 0; i < sctx->isc_nrxqs; i++) {
4314 if (ctx->ifc_sysctl_nrxds[i] != 0)
4315 scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4317 scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4320 for (i = 0; i < sctx->isc_nrxqs; i++) {
4321 if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4322 device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4323 i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4324 scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4326 if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4327 device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4328 i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4329 scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4333 for (i = 0; i < sctx->isc_ntxqs; i++) {
4334 if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4335 device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4336 i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4337 scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4339 if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4340 device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4341 i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4342 scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4348 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4353 if_softc_ctx_t scctx;
4359 ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4362 sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4363 device_set_softc(dev, ctx);
4364 ctx->ifc_flags |= IFC_SC_ALLOCATED;
4367 ctx->ifc_sctx = sctx;
4369 ctx->ifc_softc = sc;
4371 if ((err = iflib_register(ctx)) != 0) {
4372 device_printf(dev, "iflib_register failed %d\n", err);
4375 iflib_add_device_sysctl_pre(ctx);
4377 scctx = &ctx->ifc_softc_ctx;
4380 iflib_reset_qvalues(ctx);
4382 if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4383 device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4386 _iflib_pre_assert(scctx);
4387 ctx->ifc_txrx = *scctx->isc_txrx;
4390 MPASS(scctx->isc_capabilities);
4391 if (scctx->isc_capabilities & IFCAP_TXCSUM)
4392 MPASS(scctx->isc_tx_csum_flags);
4395 if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS);
4396 if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4398 if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4399 scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4400 if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4401 scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4403 main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4404 main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4406 /* XXX change for per-queue sizes */
4407 device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4408 scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4409 for (i = 0; i < sctx->isc_nrxqs; i++) {
4410 if (!powerof2(scctx->isc_nrxd[i])) {
4411 /* round down instead? */
4412 device_printf(dev, "# rx descriptors must be a power of 2\n");
4414 goto fail_iflib_detach;
4417 for (i = 0; i < sctx->isc_ntxqs; i++) {
4418 if (!powerof2(scctx->isc_ntxd[i])) {
4420 "# tx descriptors must be a power of 2");
4422 goto fail_iflib_detach;
4426 if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4427 MAX_SINGLE_PACKET_FRACTION)
4428 scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4429 MAX_SINGLE_PACKET_FRACTION);
4430 if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4431 MAX_SINGLE_PACKET_FRACTION)
4432 scctx->isc_tx_tso_segments_max = max(1,
4433 scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4435 /* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4436 if (if_getcapabilities(ifp) & IFCAP_TSO) {
4438 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4441 if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4444 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4445 * into account. In the worst case, each of these calls will
4446 * add another mbuf and, thus, the requirement for another DMA
4447 * segment. So for best performance, it doesn't make sense to
4448 * advertize a maximum of TSO segments that typically will
4449 * require defragmentation in iflib_encap().
4451 if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4452 if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4454 if (scctx->isc_rss_table_size == 0)
4455 scctx->isc_rss_table_size = 64;
4456 scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4458 GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4459 /* XXX format name */
4460 taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4462 /* Set up cpu set. If it fails, use the set of all CPUs. */
4463 if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4464 device_printf(dev, "Unable to fetch CPU list\n");
4465 CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4467 MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4470 ** Now set up MSI or MSI-X, should return us the number of supported
4471 ** vectors (will be 1 for a legacy interrupt and MSI).
4473 if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4474 msix = scctx->isc_vectors;
4475 } else if (scctx->isc_msix_bar != 0)
4477 * The simple fact that isc_msix_bar is not 0 does not mean we
4478 * we have a good value there that is known to work.
4480 msix = iflib_msix_init(ctx);
4482 scctx->isc_vectors = 1;
4483 scctx->isc_ntxqsets = 1;
4484 scctx->isc_nrxqsets = 1;
4485 scctx->isc_intr = IFLIB_INTR_LEGACY;
4488 /* Get memory for the station queues */
4489 if ((err = iflib_queues_alloc(ctx))) {
4490 device_printf(dev, "Unable to allocate queue memory\n");
4491 goto fail_intr_free;
4494 if ((err = iflib_qset_structures_setup(ctx)))
4498 * Group taskqueues aren't properly set up until SMP is started,
4499 * so we disable interrupts until we can handle them post
4502 * XXX: disabling interrupts doesn't actually work, at least for
4503 * the non-MSI case. When they occur before SI_SUB_SMP completes,
4504 * we do null handling and depend on this not causing too large an
4507 IFDI_INTR_DISABLE(ctx);
4508 if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4509 device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4514 if (scctx->isc_intr == IFLIB_INTR_MSI) {
4518 if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4519 device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4524 ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4526 if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4527 device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4532 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4533 * This must appear after the call to ether_ifattach() because
4534 * ether_ifattach() sets if_hdrlen to the default value.
4536 if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4537 if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4539 if ((err = iflib_netmap_attach(ctx))) {
4540 device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4545 NETDUMP_SET(ctx->ifc_ifp, iflib);
4547 if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4548 iflib_add_device_sysctl_post(ctx);
4549 ctx->ifc_flags |= IFC_INIT_DONE;
4554 ether_ifdetach(ctx->ifc_ifp);
4556 iflib_free_intr_mem(ctx);
4558 iflib_tx_structures_free(ctx);
4559 iflib_rx_structures_free(ctx);
4565 if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4566 free(ctx->ifc_softc, M_IFLIB);
4572 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
4573 struct iflib_cloneattach_ctx *clctx)
4578 if_softc_ctx_t scctx;
4584 ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
4585 sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4586 ctx->ifc_flags |= IFC_SC_ALLOCATED;
4587 if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
4588 ctx->ifc_flags |= IFC_PSEUDO;
4590 ctx->ifc_sctx = sctx;
4591 ctx->ifc_softc = sc;
4594 if ((err = iflib_register(ctx)) != 0) {
4595 device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
4598 iflib_add_device_sysctl_pre(ctx);
4600 scctx = &ctx->ifc_softc_ctx;
4604 * XXX sanity check that ntxd & nrxd are a power of 2
4606 iflib_reset_qvalues(ctx);
4608 if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4609 device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4612 if (sctx->isc_flags & IFLIB_GEN_MAC)
4614 if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
4615 clctx->cc_params)) != 0) {
4616 device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
4619 ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
4620 ifmedia_add(&ctx->ifc_media, IFM_ETHER | IFM_AUTO, 0, NULL);
4621 ifmedia_set(&ctx->ifc_media, IFM_ETHER | IFM_AUTO);
4624 MPASS(scctx->isc_capabilities);
4625 if (scctx->isc_capabilities & IFCAP_TXCSUM)
4626 MPASS(scctx->isc_tx_csum_flags);
4629 if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4630 if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
4632 ifp->if_flags |= IFF_NOGROUP;
4633 if (sctx->isc_flags & IFLIB_PSEUDO) {
4634 ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4636 if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4637 device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4643 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4644 * This must appear after the call to ether_ifattach() because
4645 * ether_ifattach() sets if_hdrlen to the default value.
4647 if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4648 if_setifheaderlen(ifp,
4649 sizeof(struct ether_vlan_header));
4651 if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4652 iflib_add_device_sysctl_post(ctx);
4653 ctx->ifc_flags |= IFC_INIT_DONE;
4656 _iflib_pre_assert(scctx);
4657 ctx->ifc_txrx = *scctx->isc_txrx;
4659 if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4660 scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4661 if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4662 scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4664 main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4665 main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4667 /* XXX change for per-queue sizes */
4668 device_printf(dev, "Using %d tx descriptors and %d rx descriptors\n",
4669 scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4670 for (i = 0; i < sctx->isc_nrxqs; i++) {
4671 if (!powerof2(scctx->isc_nrxd[i])) {
4672 /* round down instead? */
4673 device_printf(dev, "# rx descriptors must be a power of 2\n");
4675 goto fail_iflib_detach;
4678 for (i = 0; i < sctx->isc_ntxqs; i++) {
4679 if (!powerof2(scctx->isc_ntxd[i])) {
4681 "# tx descriptors must be a power of 2");
4683 goto fail_iflib_detach;
4687 if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4688 MAX_SINGLE_PACKET_FRACTION)
4689 scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4690 MAX_SINGLE_PACKET_FRACTION);
4691 if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4692 MAX_SINGLE_PACKET_FRACTION)
4693 scctx->isc_tx_tso_segments_max = max(1,
4694 scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4696 /* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4697 if (if_getcapabilities(ifp) & IFCAP_TSO) {
4699 * The stack can't handle a TSO size larger than IP_MAXPACKET,
4702 if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
4705 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
4706 * into account. In the worst case, each of these calls will
4707 * add another mbuf and, thus, the requirement for another DMA
4708 * segment. So for best performance, it doesn't make sense to
4709 * advertize a maximum of TSO segments that typically will
4710 * require defragmentation in iflib_encap().
4712 if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
4713 if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
4715 if (scctx->isc_rss_table_size == 0)
4716 scctx->isc_rss_table_size = 64;
4717 scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4719 GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4720 /* XXX format name */
4721 taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4723 /* XXX --- can support > 1 -- but keep it simple for now */
4724 scctx->isc_intr = IFLIB_INTR_LEGACY;
4726 /* Get memory for the station queues */
4727 if ((err = iflib_queues_alloc(ctx))) {
4728 device_printf(dev, "Unable to allocate queue memory\n");
4729 goto fail_iflib_detach;
4732 if ((err = iflib_qset_structures_setup(ctx))) {
4733 device_printf(dev, "qset structure setup failed %d\n", err);
4738 * XXX What if anything do we want to do about interrupts?
4740 ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4741 if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4742 device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4747 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
4748 * This must appear after the call to ether_ifattach() because
4749 * ether_ifattach() sets if_hdrlen to the default value.
4751 if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
4752 if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
4754 /* XXX handle more than one queue */
4755 for (i = 0; i < scctx->isc_nrxqsets; i++)
4756 IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
4760 if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4761 iflib_add_device_sysctl_post(ctx);
4762 ctx->ifc_flags |= IFC_INIT_DONE;
4765 ether_ifdetach(ctx->ifc_ifp);
4767 iflib_tx_structures_free(ctx);
4768 iflib_rx_structures_free(ctx);
4772 free(ctx->ifc_softc, M_IFLIB);
4778 iflib_pseudo_deregister(if_ctx_t ctx)
4780 if_t ifp = ctx->ifc_ifp;
4784 struct taskqgroup *tqg;
4787 /* Unregister VLAN events */
4788 if (ctx->ifc_vlan_attach_event != NULL)
4789 EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4790 if (ctx->ifc_vlan_detach_event != NULL)
4791 EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4793 ether_ifdetach(ifp);
4794 /* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4795 CTX_LOCK_DESTROY(ctx);
4796 /* XXX drain any dependent tasks */
4797 tqg = qgroup_if_io_tqg;
4798 for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4799 callout_drain(&txq->ift_timer);
4800 if (txq->ift_task.gt_uniq != NULL)
4801 taskqgroup_detach(tqg, &txq->ift_task);
4803 for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4804 if (rxq->ifr_task.gt_uniq != NULL)
4805 taskqgroup_detach(tqg, &rxq->ifr_task);
4807 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4808 free(fl->ifl_rx_bitmap, M_IFLIB);
4810 tqg = qgroup_if_config_tqg;
4811 if (ctx->ifc_admin_task.gt_uniq != NULL)
4812 taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4813 if (ctx->ifc_vflr_task.gt_uniq != NULL)
4814 taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4818 iflib_tx_structures_free(ctx);
4819 iflib_rx_structures_free(ctx);
4820 if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4821 free(ctx->ifc_softc, M_IFLIB);
4827 iflib_device_attach(device_t dev)
4830 if_shared_ctx_t sctx;
4832 if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4835 pci_enable_busmaster(dev);
4837 return (iflib_device_register(dev, NULL, sctx, &ctx));
4841 iflib_device_deregister(if_ctx_t ctx)
4843 if_t ifp = ctx->ifc_ifp;
4846 device_t dev = ctx->ifc_dev;
4848 struct taskqgroup *tqg;
4851 /* Make sure VLANS are not using driver */
4852 if (if_vlantrunkinuse(ifp)) {
4853 device_printf(dev, "Vlan in use, detach first\n");
4857 if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
4858 device_printf(dev, "SR-IOV in use; detach first.\n");
4864 ctx->ifc_flags |= IFC_IN_DETACH;
4871 /* Unregister VLAN events */
4872 if (ctx->ifc_vlan_attach_event != NULL)
4873 EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4874 if (ctx->ifc_vlan_detach_event != NULL)
4875 EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4877 iflib_netmap_detach(ifp);
4878 ether_ifdetach(ifp);
4879 if (ctx->ifc_led_dev != NULL)
4880 led_destroy(ctx->ifc_led_dev);
4881 /* XXX drain any dependent tasks */
4882 tqg = qgroup_if_io_tqg;
4883 for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4884 callout_drain(&txq->ift_timer);
4885 if (txq->ift_task.gt_uniq != NULL)
4886 taskqgroup_detach(tqg, &txq->ift_task);
4888 for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4889 if (rxq->ifr_task.gt_uniq != NULL)
4890 taskqgroup_detach(tqg, &rxq->ifr_task);
4892 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4893 free(fl->ifl_rx_bitmap, M_IFLIB);
4895 tqg = qgroup_if_config_tqg;
4896 if (ctx->ifc_admin_task.gt_uniq != NULL)
4897 taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4898 if (ctx->ifc_vflr_task.gt_uniq != NULL)
4899 taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4904 /* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4905 CTX_LOCK_DESTROY(ctx);
4906 device_set_softc(ctx->ifc_dev, NULL);
4907 iflib_free_intr_mem(ctx);
4909 bus_generic_detach(dev);
4912 iflib_tx_structures_free(ctx);
4913 iflib_rx_structures_free(ctx);
4914 if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4915 free(ctx->ifc_softc, M_IFLIB);
4916 STATE_LOCK_DESTROY(ctx);
4922 iflib_free_intr_mem(if_ctx_t ctx)
4925 if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4926 iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4928 if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4929 pci_release_msi(ctx->ifc_dev);
4931 if (ctx->ifc_msix_mem != NULL) {
4932 bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4933 rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
4934 ctx->ifc_msix_mem = NULL;
4939 iflib_device_detach(device_t dev)
4941 if_ctx_t ctx = device_get_softc(dev);
4943 return (iflib_device_deregister(ctx));
4947 iflib_device_suspend(device_t dev)
4949 if_ctx_t ctx = device_get_softc(dev);
4955 return bus_generic_suspend(dev);
4958 iflib_device_shutdown(device_t dev)
4960 if_ctx_t ctx = device_get_softc(dev);
4966 return bus_generic_suspend(dev);
4971 iflib_device_resume(device_t dev)
4973 if_ctx_t ctx = device_get_softc(dev);
4974 iflib_txq_t txq = ctx->ifc_txqs;
4978 iflib_if_init_locked(ctx);
4980 for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4981 iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4983 return (bus_generic_resume(dev));
4987 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4990 if_ctx_t ctx = device_get_softc(dev);
4993 error = IFDI_IOV_INIT(ctx, num_vfs, params);
5000 iflib_device_iov_uninit(device_t dev)
5002 if_ctx_t ctx = device_get_softc(dev);
5005 IFDI_IOV_UNINIT(ctx);
5010 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5013 if_ctx_t ctx = device_get_softc(dev);
5016 error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5022 /*********************************************************************
5024 * MODULE FUNCTION DEFINITIONS
5026 **********************************************************************/
5029 * - Start a fast taskqueue thread for each core
5030 * - Start a taskqueue for control operations
5033 iflib_module_init(void)
5039 iflib_module_event_handler(module_t mod, int what, void *arg)
5045 if ((err = iflib_module_init()) != 0)
5051 return (EOPNOTSUPP);
5057 /*********************************************************************
5059 * PUBLIC FUNCTION DEFINITIONS
5060 * ordered as in iflib.h
5062 **********************************************************************/
5066 _iflib_assert(if_shared_ctx_t sctx)
5068 MPASS(sctx->isc_tx_maxsize);
5069 MPASS(sctx->isc_tx_maxsegsize);
5071 MPASS(sctx->isc_rx_maxsize);
5072 MPASS(sctx->isc_rx_nsegments);
5073 MPASS(sctx->isc_rx_maxsegsize);
5075 MPASS(sctx->isc_nrxd_min[0]);
5076 MPASS(sctx->isc_nrxd_max[0]);
5077 MPASS(sctx->isc_nrxd_default[0]);
5078 MPASS(sctx->isc_ntxd_min[0]);
5079 MPASS(sctx->isc_ntxd_max[0]);
5080 MPASS(sctx->isc_ntxd_default[0]);
5084 _iflib_pre_assert(if_softc_ctx_t scctx)
5087 MPASS(scctx->isc_txrx->ift_txd_encap);
5088 MPASS(scctx->isc_txrx->ift_txd_flush);
5089 MPASS(scctx->isc_txrx->ift_txd_credits_update);
5090 MPASS(scctx->isc_txrx->ift_rxd_available);
5091 MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5092 MPASS(scctx->isc_txrx->ift_rxd_refill);
5093 MPASS(scctx->isc_txrx->ift_rxd_flush);
5097 iflib_register(if_ctx_t ctx)
5099 if_shared_ctx_t sctx = ctx->ifc_sctx;
5100 driver_t *driver = sctx->isc_driver;
5101 device_t dev = ctx->ifc_dev;
5104 _iflib_assert(sctx);
5107 STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5108 ifp = ctx->ifc_ifp = if_alloc(IFT_ETHER);
5110 device_printf(dev, "can not allocate ifnet structure\n");
5115 * Initialize our context's device specific methods
5117 kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5118 kobj_class_compile((kobj_class_t) driver);
5121 if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5122 if_setsoftc(ifp, ctx);
5123 if_setdev(ifp, dev);
5124 if_setinitfn(ifp, iflib_if_init);
5125 if_setioctlfn(ifp, iflib_if_ioctl);
5127 if_setstartfn(ifp, iflib_altq_if_start);
5128 if_settransmitfn(ifp, iflib_altq_if_transmit);
5129 if_setsendqready(ifp);
5131 if_settransmitfn(ifp, iflib_if_transmit);
5133 if_setqflushfn(ifp, iflib_if_qflush);
5134 if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
5136 ctx->ifc_vlan_attach_event =
5137 EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5138 EVENTHANDLER_PRI_FIRST);
5139 ctx->ifc_vlan_detach_event =
5140 EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5141 EVENTHANDLER_PRI_FIRST);
5143 ifmedia_init(&ctx->ifc_media, IFM_IMASK,
5144 iflib_media_change, iflib_media_status);
5151 iflib_queues_alloc(if_ctx_t ctx)
5153 if_shared_ctx_t sctx = ctx->ifc_sctx;
5154 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5155 device_t dev = ctx->ifc_dev;
5156 int nrxqsets = scctx->isc_nrxqsets;
5157 int ntxqsets = scctx->isc_ntxqsets;
5160 iflib_fl_t fl = NULL;
5161 int i, j, cpu, err, txconf, rxconf;
5162 iflib_dma_info_t ifdip;
5163 uint32_t *rxqsizes = scctx->isc_rxqsizes;
5164 uint32_t *txqsizes = scctx->isc_txqsizes;
5165 uint8_t nrxqs = sctx->isc_nrxqs;
5166 uint8_t ntxqs = sctx->isc_ntxqs;
5167 int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5171 KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5172 KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5174 /* Allocate the TX ring struct memory */
5175 if (!(ctx->ifc_txqs =
5176 (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5177 ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5178 device_printf(dev, "Unable to allocate TX ring memory\n");
5183 /* Now allocate the RX */
5184 if (!(ctx->ifc_rxqs =
5185 (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5186 nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5187 device_printf(dev, "Unable to allocate RX ring memory\n");
5192 txq = ctx->ifc_txqs;
5193 rxq = ctx->ifc_rxqs;
5196 * XXX handle allocation failure
5198 for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5199 /* Set up some basics */
5201 if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5202 device_printf(dev, "failed to allocate iflib_dma_info\n");
5206 txq->ift_ifdi = ifdip;
5207 for (j = 0; j < ntxqs; j++, ifdip++) {
5208 if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5209 device_printf(dev, "Unable to allocate Descriptor memory\n");
5213 txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5214 bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5218 if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5219 txq->ift_br_offset = 1;
5221 txq->ift_br_offset = 0;
5224 txq->ift_timer.c_cpu = cpu;
5226 if (iflib_txsd_alloc(txq)) {
5227 device_printf(dev, "Critical Failure setting up TX buffers\n");
5232 /* Initialize the TX lock */
5233 snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
5234 device_get_nameunit(dev), txq->ift_id);
5235 mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5236 callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5238 snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
5239 device_get_nameunit(dev), txq->ift_id);
5241 err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5242 iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5244 /* XXX free any allocated rings */
5245 device_printf(dev, "Unable to allocate buf_ring\n");
5250 for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5251 /* Set up some basics */
5253 if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
5254 device_printf(dev, "failed to allocate iflib_dma_info\n");
5259 rxq->ifr_ifdi = ifdip;
5260 /* XXX this needs to be changed if #rx queues != #tx queues */
5261 rxq->ifr_ntxqirq = 1;
5262 rxq->ifr_txqid[0] = i;
5263 for (j = 0; j < nrxqs; j++, ifdip++) {
5264 if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
5265 device_printf(dev, "Unable to allocate Descriptor memory\n");
5269 bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
5273 if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5274 rxq->ifr_fl_offset = 1;
5276 rxq->ifr_fl_offset = 0;
5278 rxq->ifr_nfl = nfree_lists;
5280 (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
5281 device_printf(dev, "Unable to allocate free list memory\n");
5286 for (j = 0; j < nfree_lists; j++) {
5287 fl[j].ifl_rxq = rxq;
5289 fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
5290 fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
5292 /* Allocate receive buffers for the ring */
5293 if (iflib_rxsd_alloc(rxq)) {
5295 "Critical Failure setting up receive buffers\n");
5300 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
5301 fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
5306 vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5307 paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
5308 for (i = 0; i < ntxqsets; i++) {
5309 iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
5311 for (j = 0; j < ntxqs; j++, di++) {
5312 vaddrs[i*ntxqs + j] = di->idi_vaddr;
5313 paddrs[i*ntxqs + j] = di->idi_paddr;
5316 if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
5317 device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5318 iflib_tx_structures_free(ctx);
5319 free(vaddrs, M_IFLIB);
5320 free(paddrs, M_IFLIB);
5323 free(vaddrs, M_IFLIB);
5324 free(paddrs, M_IFLIB);
5327 vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5328 paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
5329 for (i = 0; i < nrxqsets; i++) {
5330 iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
5332 for (j = 0; j < nrxqs; j++, di++) {
5333 vaddrs[i*nrxqs + j] = di->idi_vaddr;
5334 paddrs[i*nrxqs + j] = di->idi_paddr;
5337 if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
5338 device_printf(ctx->ifc_dev, "device queue allocation failed\n");
5339 iflib_tx_structures_free(ctx);
5340 free(vaddrs, M_IFLIB);
5341 free(paddrs, M_IFLIB);
5344 free(vaddrs, M_IFLIB);
5345 free(paddrs, M_IFLIB);
5349 /* XXX handle allocation failure changes */
5353 if (ctx->ifc_rxqs != NULL)
5354 free(ctx->ifc_rxqs, M_IFLIB);
5355 ctx->ifc_rxqs = NULL;
5356 if (ctx->ifc_txqs != NULL)
5357 free(ctx->ifc_txqs, M_IFLIB);
5358 ctx->ifc_txqs = NULL;
5364 iflib_tx_structures_setup(if_ctx_t ctx)
5366 iflib_txq_t txq = ctx->ifc_txqs;
5369 for (i = 0; i < NTXQSETS(ctx); i++, txq++)
5370 iflib_txq_setup(txq);
5376 iflib_tx_structures_free(if_ctx_t ctx)
5378 iflib_txq_t txq = ctx->ifc_txqs;
5379 if_shared_ctx_t sctx = ctx->ifc_sctx;
5382 for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
5383 iflib_txq_destroy(txq);
5384 for (j = 0; j < sctx->isc_ntxqs; j++)
5385 iflib_dma_free(&txq->ift_ifdi[j]);
5387 free(ctx->ifc_txqs, M_IFLIB);
5388 ctx->ifc_txqs = NULL;
5389 IFDI_QUEUES_FREE(ctx);
5392 /*********************************************************************
5394 * Initialize all receive rings.
5396 **********************************************************************/
5398 iflib_rx_structures_setup(if_ctx_t ctx)
5400 iflib_rxq_t rxq = ctx->ifc_rxqs;
5402 #if defined(INET6) || defined(INET)
5406 for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
5407 #if defined(INET6) || defined(INET)
5408 tcp_lro_free(&rxq->ifr_lc);
5409 if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
5410 TCP_LRO_ENTRIES, min(1024,
5411 ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
5412 device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
5415 rxq->ifr_lro_enabled = TRUE;
5417 IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
5420 #if defined(INET6) || defined(INET)
5423 * Free RX software descriptors allocated so far, we will only handle
5424 * the rings that completed, the failing case will have
5425 * cleaned up for itself. 'q' failed, so its the terminus.
5427 rxq = ctx->ifc_rxqs;
5428 for (i = 0; i < q; ++i, rxq++) {
5429 iflib_rx_sds_free(rxq);
5430 rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5436 /*********************************************************************
5438 * Free all receive rings.
5440 **********************************************************************/
5442 iflib_rx_structures_free(if_ctx_t ctx)
5444 iflib_rxq_t rxq = ctx->ifc_rxqs;
5446 for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5447 iflib_rx_sds_free(rxq);
5449 free(ctx->ifc_rxqs, M_IFLIB);
5450 ctx->ifc_rxqs = NULL;
5454 iflib_qset_structures_setup(if_ctx_t ctx)
5459 * It is expected that the caller takes care of freeing queues if this
5462 if ((err = iflib_tx_structures_setup(ctx)) != 0) {
5463 device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
5467 if ((err = iflib_rx_structures_setup(ctx)) != 0)
5468 device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5474 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5475 driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
5478 return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5483 find_nth(if_ctx_t ctx, int qid)
5486 int i, cpuid, eqid, count;
5488 CPU_COPY(&ctx->ifc_cpus, &cpus);
5489 count = CPU_COUNT(&cpus);
5491 /* clear up to the qid'th bit */
5492 for (i = 0; i < eqid; i++) {
5493 cpuid = CPU_FFS(&cpus);
5495 CPU_CLR(cpuid-1, &cpus);
5497 cpuid = CPU_FFS(&cpus);
5503 extern struct cpu_group *cpu_top; /* CPU topology */
5506 find_child_with_core(int cpu, struct cpu_group *grp)
5510 if (grp->cg_children == 0)
5513 MPASS(grp->cg_child);
5514 for (i = 0; i < grp->cg_children; i++) {
5515 if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5523 * Find the nth "close" core to the specified core
5524 * "close" is defined as the deepest level that shares
5525 * at least an L2 cache. With threads, this will be
5526 * threads on the same core. If the sahred cache is L3
5527 * or higher, simply returns the same core.
5530 find_close_core(int cpu, int core_offset)
5532 struct cpu_group *grp;
5541 while ((i = find_child_with_core(cpu, grp)) != -1) {
5542 /* If the child only has one cpu, don't descend */
5543 if (grp->cg_child[i].cg_count <= 1)
5545 grp = &grp->cg_child[i];
5548 /* If they don't share at least an L2 cache, use the same CPU */
5549 if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5553 CPU_COPY(&grp->cg_mask, &cs);
5555 /* Add the selected CPU offset to core offset. */
5556 for (i = 0; (fcpu = CPU_FFS(&cs)) != 0; i++) {
5557 if (fcpu - 1 == cpu)
5559 CPU_CLR(fcpu - 1, &cs);
5565 CPU_COPY(&grp->cg_mask, &cs);
5566 for (i = core_offset % grp->cg_count; i > 0; i--) {
5567 MPASS(CPU_FFS(&cs));
5568 CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5570 MPASS(CPU_FFS(&cs));
5571 return CPU_FFS(&cs) - 1;
5575 find_close_core(int cpu, int core_offset __unused)
5582 get_core_offset(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5586 /* TX queues get cores which share at least an L2 cache with the corresponding RX queue */
5587 /* XXX handle multiple RX threads per core and more than two core per L2 group */
5588 return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5590 case IFLIB_INTR_RXTX:
5591 /* RX queues get the specified core */
5592 return qid / CPU_COUNT(&ctx->ifc_cpus);
5598 #define get_core_offset(ctx, type, qid) CPU_FIRST()
5599 #define find_close_core(cpuid, tid) CPU_FIRST()
5600 #define find_nth(ctx, gid) CPU_FIRST()
5603 /* Just to avoid copy/paste */
5605 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5606 struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name)
5611 cpuid = find_nth(ctx, qid);
5612 tid = get_core_offset(ctx, type, qid);
5614 cpuid = find_close_core(cpuid, tid);
5615 err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5617 device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5621 if (cpuid > ctx->ifc_cpuid_highest)
5622 ctx->ifc_cpuid_highest = cpuid;
5628 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5629 iflib_intr_type_t type, driver_filter_t *filter,
5630 void *filter_arg, int qid, const char *name)
5632 struct grouptask *gtask;
5633 struct taskqgroup *tqg;
5634 iflib_filter_info_t info;
5637 driver_filter_t *intr_fast;
5640 info = &ctx->ifc_filter_info;
5644 /* XXX merge tx/rx for netmap? */
5646 q = &ctx->ifc_txqs[qid];
5647 info = &ctx->ifc_txqs[qid].ift_filter_info;
5648 gtask = &ctx->ifc_txqs[qid].ift_task;
5649 tqg = qgroup_if_io_tqg;
5651 intr_fast = iflib_fast_intr;
5652 GROUPTASK_INIT(gtask, 0, fn, q);
5653 ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
5656 q = &ctx->ifc_rxqs[qid];
5657 info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5658 gtask = &ctx->ifc_rxqs[qid].ifr_task;
5659 tqg = qgroup_if_io_tqg;
5661 intr_fast = iflib_fast_intr;
5662 GROUPTASK_INIT(gtask, 0, fn, q);
5664 case IFLIB_INTR_RXTX:
5665 q = &ctx->ifc_rxqs[qid];
5666 info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5667 gtask = &ctx->ifc_rxqs[qid].ifr_task;
5668 tqg = qgroup_if_io_tqg;
5670 intr_fast = iflib_fast_intr_rxtx;
5671 GROUPTASK_INIT(gtask, 0, fn, q);
5673 case IFLIB_INTR_ADMIN:
5676 info = &ctx->ifc_filter_info;
5677 gtask = &ctx->ifc_admin_task;
5678 tqg = qgroup_if_config_tqg;
5679 fn = _task_fn_admin;
5680 intr_fast = iflib_fast_intr_ctx;
5683 panic("unknown net intr type");
5686 info->ifi_filter = filter;
5687 info->ifi_filter_arg = filter_arg;
5688 info->ifi_task = gtask;
5691 err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info, name);
5693 device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5696 if (type == IFLIB_INTR_ADMIN)
5700 err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5704 taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5711 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
5713 struct grouptask *gtask;
5714 struct taskqgroup *tqg;
5722 q = &ctx->ifc_txqs[qid];
5723 gtask = &ctx->ifc_txqs[qid].ift_task;
5724 tqg = qgroup_if_io_tqg;
5727 irq_num = rman_get_start(irq->ii_res);
5730 q = &ctx->ifc_rxqs[qid];
5731 gtask = &ctx->ifc_rxqs[qid].ifr_task;
5732 tqg = qgroup_if_io_tqg;
5735 irq_num = rman_get_start(irq->ii_res);
5737 case IFLIB_INTR_IOV:
5739 gtask = &ctx->ifc_vflr_task;
5740 tqg = qgroup_if_config_tqg;
5744 panic("unknown net intr type");
5746 GROUPTASK_INIT(gtask, 0, fn, q);
5747 if (irq_num != -1) {
5748 err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5750 taskqgroup_attach(tqg, gtask, q, irq_num, name);
5753 taskqgroup_attach(tqg, gtask, q, irq_num, name);
5758 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5762 bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5765 bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
5766 rman_get_rid(irq->ii_res), irq->ii_res);
5770 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
5772 iflib_txq_t txq = ctx->ifc_txqs;
5773 iflib_rxq_t rxq = ctx->ifc_rxqs;
5774 if_irq_t irq = &ctx->ifc_legacy_irq;
5775 iflib_filter_info_t info;
5776 struct grouptask *gtask;
5777 struct taskqgroup *tqg;
5783 q = &ctx->ifc_rxqs[0];
5784 info = &rxq[0].ifr_filter_info;
5785 gtask = &rxq[0].ifr_task;
5786 tqg = qgroup_if_io_tqg;
5787 tqrid = irq->ii_rid = *rid;
5790 ctx->ifc_flags |= IFC_LEGACY;
5791 info->ifi_filter = filter;
5792 info->ifi_filter_arg = filter_arg;
5793 info->ifi_task = gtask;
5794 info->ifi_ctx = ctx;
5796 /* We allocate a single interrupt resource */
5797 if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5799 GROUPTASK_INIT(gtask, 0, fn, q);
5800 taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5802 GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5803 taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5808 iflib_led_create(if_ctx_t ctx)
5811 ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5812 device_get_nameunit(ctx->ifc_dev));
5816 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5819 GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5823 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5826 GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5830 iflib_admin_intr_deferred(if_ctx_t ctx)
5833 struct grouptask *gtask;
5835 gtask = &ctx->ifc_admin_task;
5836 MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5839 GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5843 iflib_iov_intr_deferred(if_ctx_t ctx)
5846 GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5850 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5853 taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5857 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
5861 GROUPTASK_INIT(gtask, 0, fn, ctx);
5862 taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5866 iflib_config_gtask_deinit(struct grouptask *gtask)
5869 taskqgroup_detach(qgroup_if_config_tqg, gtask);
5873 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5875 if_t ifp = ctx->ifc_ifp;
5876 iflib_txq_t txq = ctx->ifc_txqs;
5878 if_setbaudrate(ifp, baudrate);
5879 if (baudrate >= IF_Gbps(10)) {
5881 ctx->ifc_flags |= IFC_PREFETCH;
5884 /* If link down, disable watchdog */
5885 if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5886 for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5887 txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5889 ctx->ifc_link_state = link_state;
5890 if_link_state_change(ifp, link_state);
5894 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5898 int credits_pre = txq->ift_cidx_processed;
5901 if (ctx->isc_txd_credits_update == NULL)
5904 bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
5905 BUS_DMASYNC_POSTREAD);
5906 if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5909 txq->ift_processed += credits;
5910 txq->ift_cidx_processed += credits;
5912 MPASS(credits_pre + credits == txq->ift_cidx_processed);
5913 if (txq->ift_cidx_processed >= txq->ift_size)
5914 txq->ift_cidx_processed -= txq->ift_size;
5919 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5922 return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5927 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5928 const char *description, if_int_delay_info_t info,
5929 int offset, int value)
5931 info->iidi_ctx = ctx;
5932 info->iidi_offset = offset;
5933 info->iidi_value = value;
5934 SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5935 SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5936 OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5937 info, 0, iflib_sysctl_int_delay, "I", description);
5941 iflib_ctx_lock_get(if_ctx_t ctx)
5944 return (&ctx->ifc_ctx_sx);
5948 iflib_msix_init(if_ctx_t ctx)
5950 device_t dev = ctx->ifc_dev;
5951 if_shared_ctx_t sctx = ctx->ifc_sctx;
5952 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5953 int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5954 int iflib_num_tx_queues, iflib_num_rx_queues;
5955 int err, admincnt, bar;
5957 iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5958 iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5961 device_printf(dev, "msix_init qsets capped at %d\n",
5962 imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5964 bar = ctx->ifc_softc_ctx.isc_msix_bar;
5965 admincnt = sctx->isc_admin_intrcnt;
5966 /* Override by tuneable */
5967 if (scctx->isc_disable_msix)
5970 /* First try MSI-X */
5971 if ((msgs = pci_msix_count(dev)) == 0) {
5973 device_printf(dev, "MSI-X not supported or disabled\n");
5977 * bar == -1 => "trust me I know what I'm doing"
5978 * Some drivers are for hardware that is so shoddily
5979 * documented that no one knows which bars are which
5980 * so the developer has to map all bars. This hack
5981 * allows shoddy garbage to use MSI-X in this framework.
5984 ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5985 SYS_RES_MEMORY, &bar, RF_ACTIVE);
5986 if (ctx->ifc_msix_mem == NULL) {
5987 device_printf(dev, "Unable to map MSI-X table\n");
5992 /* use only 1 qset in debug mode */
5993 queuemsgs = min(msgs - admincnt, 1);
5995 queuemsgs = msgs - admincnt;
5998 queues = imin(queuemsgs, rss_getnumbuckets());
6002 queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6005 "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6006 CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6008 /* If we're doing RSS, clamp at the number of RSS buckets */
6009 if (queues > rss_getnumbuckets())
6010 queues = rss_getnumbuckets();
6012 if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6013 rx_queues = iflib_num_rx_queues;
6017 if (rx_queues > scctx->isc_nrxqsets)
6018 rx_queues = scctx->isc_nrxqsets;
6021 * We want this to be all logical CPUs by default
6023 if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6024 tx_queues = iflib_num_tx_queues;
6026 tx_queues = mp_ncpus;
6028 if (tx_queues > scctx->isc_ntxqsets)
6029 tx_queues = scctx->isc_ntxqsets;
6031 if (ctx->ifc_sysctl_qs_eq_override == 0) {
6033 if (tx_queues != rx_queues)
6035 "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6036 min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6038 tx_queues = min(rx_queues, tx_queues);
6039 rx_queues = min(rx_queues, tx_queues);
6042 device_printf(dev, "Using %d rx queues %d tx queues\n",
6043 rx_queues, tx_queues);
6045 vectors = rx_queues + admincnt;
6046 if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6047 device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6049 scctx->isc_vectors = vectors;
6050 scctx->isc_nrxqsets = rx_queues;
6051 scctx->isc_ntxqsets = tx_queues;
6052 scctx->isc_intr = IFLIB_INTR_MSIX;
6057 "failed to allocate %d MSI-X vectors, err: %d - using MSI\n",
6059 bus_release_resource(dev, SYS_RES_MEMORY, bar,
6061 ctx->ifc_msix_mem = NULL;
6064 vectors = pci_msi_count(dev);
6065 scctx->isc_nrxqsets = 1;
6066 scctx->isc_ntxqsets = 1;
6067 scctx->isc_vectors = vectors;
6068 if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6069 device_printf(dev,"Using an MSI interrupt\n");
6070 scctx->isc_intr = IFLIB_INTR_MSI;
6072 scctx->isc_vectors = 1;
6073 device_printf(dev,"Using a Legacy interrupt\n");
6074 scctx->isc_intr = IFLIB_INTR_LEGACY;
6080 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6083 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6086 uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6088 const char *ring_state = "UNKNOWN";
6091 rc = sysctl_wire_old_buffer(req, 0);
6095 sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6100 ring_state = ring_states[state[3]];
6102 sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6103 state[0], state[1], state[2], ring_state);
6104 rc = sbuf_finish(sb);
6109 enum iflib_ndesc_handler {
6115 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6117 if_ctx_t ctx = (void *)arg1;
6118 enum iflib_ndesc_handler type = arg2;
6119 char buf[256] = {0};
6124 MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
6128 case IFLIB_NTXD_HANDLER:
6129 ndesc = ctx->ifc_sysctl_ntxds;
6131 nqs = ctx->ifc_sctx->isc_ntxqs;
6133 case IFLIB_NRXD_HANDLER:
6134 ndesc = ctx->ifc_sysctl_nrxds;
6136 nqs = ctx->ifc_sctx->isc_nrxqs;
6139 panic("unhandled type");
6144 for (i=0; i<8; i++) {
6149 sprintf(strchr(buf, 0), "%d", ndesc[i]);
6152 rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6153 if (rc || req->newptr == NULL)
6156 for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6157 i++, p = strsep(&next, " ,")) {
6158 ndesc[i] = strtoul(p, NULL, 10);
6164 #define NAME_BUFLEN 32
6166 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6168 device_t dev = iflib_get_dev(ctx);
6169 struct sysctl_oid_list *child, *oid_list;
6170 struct sysctl_ctx_list *ctx_list;
6171 struct sysctl_oid *node;
6173 ctx_list = device_get_sysctl_ctx(dev);
6174 child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6175 ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6176 CTLFLAG_RD, NULL, "IFLIB fields");
6177 oid_list = SYSCTL_CHILDREN(node);
6179 SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6180 CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
6183 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6184 CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6185 "# of txqs to use, 0 => use default #");
6186 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6187 CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6188 "# of rxqs to use, 0 => use default #");
6189 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6190 CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6191 "permit #txq != #rxq");
6192 SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6193 CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6194 "disable MSI-X (default 0)");
6195 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6196 CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6197 "set the rx budget");
6198 SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6199 CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6200 "cause tx to abdicate instead of running to completion");
6202 /* XXX change for per-queue sizes */
6203 SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6204 CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
6205 mp_ndesc_handler, "A",
6206 "list of # of tx descriptors to use, 0 = use default #");
6207 SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6208 CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
6209 mp_ndesc_handler, "A",
6210 "list of # of rx descriptors to use, 0 = use default #");
6214 iflib_add_device_sysctl_post(if_ctx_t ctx)
6216 if_shared_ctx_t sctx = ctx->ifc_sctx;
6217 if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6218 device_t dev = iflib_get_dev(ctx);
6219 struct sysctl_oid_list *child;
6220 struct sysctl_ctx_list *ctx_list;
6225 char namebuf[NAME_BUFLEN];
6227 struct sysctl_oid *queue_node, *fl_node, *node;
6228 struct sysctl_oid_list *queue_list, *fl_list;
6229 ctx_list = device_get_sysctl_ctx(dev);
6231 node = ctx->ifc_sysctl_node;
6232 child = SYSCTL_CHILDREN(node);
6234 if (scctx->isc_ntxqsets > 100)
6236 else if (scctx->isc_ntxqsets > 10)
6240 for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6241 snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6242 queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6243 CTLFLAG_RD, NULL, "Queue Name");
6244 queue_list = SYSCTL_CHILDREN(queue_node);
6246 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6248 &txq->ift_dequeued, "total mbufs freed");
6249 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6251 &txq->ift_enqueued, "total mbufs enqueued");
6253 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6255 &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6256 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6258 &txq->ift_pullups, "# of times m_pullup was called");
6259 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6261 &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6262 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6264 &txq->ift_no_desc_avail, "# of times no descriptors were available");
6265 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6267 &txq->ift_map_failed, "# of times dma map failed");
6268 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6270 &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6271 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6273 &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6274 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6276 &txq->ift_pidx, 1, "Producer Index");
6277 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6279 &txq->ift_cidx, 1, "Consumer Index");
6280 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6282 &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6283 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6285 &txq->ift_in_use, 1, "descriptors in use");
6286 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6288 &txq->ift_processed, "descriptors procesed for clean");
6289 SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6291 &txq->ift_cleaned, "total cleaned");
6292 SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6293 CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
6294 0, mp_ring_state_handler, "A", "soft ring state");
6295 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6296 CTLFLAG_RD, &txq->ift_br->enqueues,
6297 "# of enqueues to the mp_ring for this queue");
6298 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6299 CTLFLAG_RD, &txq->ift_br->drops,
6300 "# of drops in the mp_ring for this queue");
6301 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6302 CTLFLAG_RD, &txq->ift_br->starts,
6303 "# of normal consumer starts in the mp_ring for this queue");
6304 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6305 CTLFLAG_RD, &txq->ift_br->stalls,
6306 "# of consumer stalls in the mp_ring for this queue");
6307 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6308 CTLFLAG_RD, &txq->ift_br->restarts,
6309 "# of consumer restarts in the mp_ring for this queue");
6310 SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6311 CTLFLAG_RD, &txq->ift_br->abdications,
6312 "# of consumer abdications in the mp_ring for this queue");
6315 if (scctx->isc_nrxqsets > 100)
6317 else if (scctx->isc_nrxqsets > 10)
6321 for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6322 snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6323 queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6324 CTLFLAG_RD, NULL, "Queue Name");
6325 queue_list = SYSCTL_CHILDREN(queue_node);
6326 if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6327 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
6329 &rxq->ifr_cq_pidx, 1, "Producer Index");
6330 SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
6332 &rxq->ifr_cq_cidx, 1, "Consumer Index");
6335 for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
6336 snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
6337 fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
6338 CTLFLAG_RD, NULL, "freelist Name");
6339 fl_list = SYSCTL_CHILDREN(fl_node);
6340 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
6342 &fl->ifl_pidx, 1, "Producer Index");
6343 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
6345 &fl->ifl_cidx, 1, "Consumer Index");
6346 SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
6348 &fl->ifl_credits, 1, "credits available");
6350 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
6352 &fl->ifl_m_enqueued, "mbufs allocated");
6353 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
6355 &fl->ifl_m_dequeued, "mbufs freed");
6356 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
6358 &fl->ifl_cl_enqueued, "clusters allocated");
6359 SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
6361 &fl->ifl_cl_dequeued, "clusters freed");
6370 iflib_request_reset(if_ctx_t ctx)
6374 ctx->ifc_flags |= IFC_DO_RESET;
6378 #ifndef __NO_STRICT_ALIGNMENT
6379 static struct mbuf *
6380 iflib_fixup_rx(struct mbuf *m)
6384 if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
6385 bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
6386 m->m_data += ETHER_HDR_LEN;
6389 MGETHDR(n, M_NOWAIT, MT_DATA);
6394 bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
6395 m->m_data += ETHER_HDR_LEN;
6396 m->m_len -= ETHER_HDR_LEN;
6397 n->m_len = ETHER_HDR_LEN;
6398 M_MOVE_PKTHDR(n, m);
6407 iflib_netdump_init(struct ifnet *ifp, int *nrxr, int *ncl, int *clsize)
6411 ctx = if_getsoftc(ifp);
6413 *nrxr = NRXQSETS(ctx);
6414 *ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
6415 *clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
6420 iflib_netdump_event(struct ifnet *ifp, enum netdump_ev event)
6423 if_softc_ctx_t scctx;
6428 ctx = if_getsoftc(ifp);
6429 scctx = &ctx->ifc_softc_ctx;
6433 for (i = 0; i < scctx->isc_nrxqsets; i++) {
6434 rxq = &ctx->ifc_rxqs[i];
6435 for (j = 0; j < rxq->ifr_nfl; j++) {
6437 fl->ifl_zone = m_getzone(fl->ifl_buf_size);
6440 iflib_no_tx_batch = 1;
6448 iflib_netdump_transmit(struct ifnet *ifp, struct mbuf *m)
6454 ctx = if_getsoftc(ifp);
6455 if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6459 txq = &ctx->ifc_txqs[0];
6460 error = iflib_encap(txq, &m);
6462 (void)iflib_txd_db_check(ctx, txq, true, txq->ift_in_use);
6467 iflib_netdump_poll(struct ifnet *ifp, int count)
6470 if_softc_ctx_t scctx;
6474 ctx = if_getsoftc(ifp);
6475 scctx = &ctx->ifc_softc_ctx;
6477 if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
6481 txq = &ctx->ifc_txqs[0];
6482 (void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
6484 for (i = 0; i < scctx->isc_nrxqsets; i++)
6485 (void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
6488 #endif /* NETDUMP */