2 * Copyright (c) 2014 Chelsio Communications, Inc.
4 * Written by: Navdeep Parhar <np@FreeBSD.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * 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 AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
32 #include "opt_inet6.h"
35 #include <sys/param.h>
37 #include <sys/eventhandler.h>
40 #include <sys/module.h>
41 #include <sys/selinfo.h>
42 #include <sys/socket.h>
43 #include <sys/sockio.h>
44 #include <machine/bus.h>
45 #include <net/ethernet.h>
47 #include <net/if_media.h>
48 #include <net/if_var.h>
49 #include <net/if_clone.h>
50 #include <net/if_types.h>
51 #include <net/netmap.h>
52 #include <dev/netmap/netmap_kern.h>
54 #include "common/common.h"
55 #include "common/t4_regs.h"
56 #include "common/t4_regs_values.h"
58 extern int fl_pad; /* XXXNM */
60 SYSCTL_NODE(_hw, OID_AUTO, cxgbe, CTLFLAG_RD, 0, "cxgbe netmap parameters");
63 * 0 = normal netmap rx
65 * 2 = supermassive black hole (buffer packing enabled)
68 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_black_hole, CTLFLAG_RDTUN, &black_hole, 0,
69 "Sink incoming packets.");
72 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_rx_ndesc, CTLFLAG_RWTUN,
73 &rx_ndesc, 0, "# of rx descriptors after which the hw cidx is updated.");
75 int holdoff_tmr_idx = 2;
76 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_holdoff_tmr_idx, CTLFLAG_RWTUN,
77 &holdoff_tmr_idx, 0, "Holdoff timer index for netmap rx queues.");
81 * -1: no congestion feedback (not recommended).
82 * 0: backpressure the channel instead of dropping packets right away.
83 * 1: no backpressure, drop packets for the congested queue immediately.
85 static int nm_cong_drop = 1;
86 TUNABLE_INT("hw.cxgbe.nm_cong_drop", &nm_cong_drop);
89 alloc_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int cong)
93 struct adapter *sc = vi->pi->adapter;
94 struct sge_params *sp = &sc->params.sge;
95 struct netmap_adapter *na = NA(vi->ifp);
99 MPASS(nm_rxq->iq_desc != NULL);
100 MPASS(nm_rxq->fl_desc != NULL);
102 bzero(nm_rxq->iq_desc, vi->qsize_rxq * IQ_ESIZE);
103 bzero(nm_rxq->fl_desc, na->num_rx_desc * EQ_ESIZE + sp->spg_len);
105 bzero(&c, sizeof(c));
106 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
107 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
109 c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
111 if (vi->flags & INTR_RXQ) {
112 KASSERT(nm_rxq->intr_idx < sc->intr_count,
113 ("%s: invalid direct intr_idx %d", __func__,
115 v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx);
117 CXGBE_UNIMPLEMENTED(__func__); /* XXXNM: needs review */
118 v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx) |
121 c.type_to_iqandstindex = htobe32(v |
122 V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
123 V_FW_IQ_CMD_VIID(vi->viid) |
124 V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
125 c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(vi->pi->tx_chan) |
126 F_FW_IQ_CMD_IQGTSMODE |
127 V_FW_IQ_CMD_IQINTCNTTHRESH(0) |
128 V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
129 c.iqsize = htobe16(vi->qsize_rxq);
130 c.iqaddr = htobe64(nm_rxq->iq_ba);
132 c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN |
133 V_FW_IQ_CMD_FL0CNGCHMAP(cong) | F_FW_IQ_CMD_FL0CONGCIF |
134 F_FW_IQ_CMD_FL0CONGEN);
136 c.iqns_to_fl0congen |=
137 htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
138 F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
139 (fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
140 (black_hole == 2 ? F_FW_IQ_CMD_FL0PACKEN : 0));
141 c.fl0dcaen_to_fl0cidxfthresh =
142 htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_128B) |
143 V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
144 c.fl0size = htobe16(na->num_rx_desc / 8 + sp->spg_len / EQ_ESIZE);
145 c.fl0addr = htobe64(nm_rxq->fl_ba);
147 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
149 device_printf(sc->dev,
150 "failed to create netmap ingress queue: %d\n", rc);
155 MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - sp->spg_len / IQ_ESIZE);
156 nm_rxq->iq_gen = F_RSPD_GEN;
157 nm_rxq->iq_cntxt_id = be16toh(c.iqid);
158 nm_rxq->iq_abs_id = be16toh(c.physiqid);
159 cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start;
160 if (cntxt_id >= sc->sge.niq) {
161 panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)",
162 __func__, cntxt_id, sc->sge.niq - 1);
164 sc->sge.iqmap[cntxt_id] = (void *)nm_rxq;
166 nm_rxq->fl_cntxt_id = be16toh(c.fl0id);
167 nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
168 MPASS(nm_rxq->fl_sidx == na->num_rx_desc);
169 cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start;
170 if (cntxt_id >= sc->sge.neq) {
171 panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)",
172 __func__, cntxt_id, sc->sge.neq - 1);
174 sc->sge.eqmap[cntxt_id] = (void *)nm_rxq;
176 nm_rxq->fl_db_val = V_QID(nm_rxq->fl_cntxt_id) |
177 sc->chip_params->sge_fl_db;
179 if (is_t5(sc) && cong >= 0) {
182 param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
183 V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
184 V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
185 param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
186 V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
187 V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
192 for (i = 0; i < 4; i++) {
194 val |= 1 << (i << 2);
198 rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, ¶m, &val);
200 /* report error but carry on */
201 device_printf(sc->dev,
202 "failed to set congestion manager context for "
203 "ingress queue %d: %d\n", nm_rxq->iq_cntxt_id, rc);
207 t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
208 V_INGRESSQID(nm_rxq->iq_cntxt_id) |
209 V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
215 free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
217 struct adapter *sc = vi->pi->adapter;
220 rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
221 nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
223 device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n",
224 __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc);
229 alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
233 struct adapter *sc = vi->pi->adapter;
234 struct netmap_adapter *na = NA(vi->ifp);
235 struct fw_eq_eth_cmd c;
238 MPASS(nm_txq->desc != NULL);
240 len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
241 bzero(nm_txq->desc, len);
243 bzero(&c, sizeof(c));
244 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
245 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
246 V_FW_EQ_ETH_CMD_VFN(0));
247 c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
248 F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
249 c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
250 F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
252 htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
253 V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
254 V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id));
255 c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
256 V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
257 V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE));
258 c.eqaddr = htobe64(nm_txq->ba);
260 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
262 device_printf(vi->dev,
263 "failed to create netmap egress queue: %d\n", rc);
267 nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
268 cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start;
269 if (cntxt_id >= sc->sge.neq)
270 panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__,
271 cntxt_id, sc->sge.neq - 1);
272 sc->sge.eqmap[cntxt_id] = (void *)nm_txq;
274 nm_txq->pidx = nm_txq->cidx = 0;
275 MPASS(nm_txq->sidx == na->num_tx_desc);
276 nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0;
278 nm_txq->doorbells = sc->doorbells;
279 if (isset(&nm_txq->doorbells, DOORBELL_UDB) ||
280 isset(&nm_txq->doorbells, DOORBELL_UDBWC) ||
281 isset(&nm_txq->doorbells, DOORBELL_WCWR)) {
282 uint32_t s_qpp = sc->params.sge.eq_s_qpp;
283 uint32_t mask = (1 << s_qpp) - 1;
284 volatile uint8_t *udb;
286 udb = sc->udbs_base + UDBS_DB_OFFSET;
287 udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT;
288 nm_txq->udb_qid = nm_txq->cntxt_id & mask;
289 if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
290 clrbit(&nm_txq->doorbells, DOORBELL_WCWR);
292 udb += nm_txq->udb_qid << UDBS_SEG_SHIFT;
295 nm_txq->udb = (volatile void *)udb;
302 free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
304 struct adapter *sc = vi->pi->adapter;
307 rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
309 device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__,
310 nm_txq->cntxt_id, rc);
315 cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
316 struct netmap_adapter *na)
318 struct netmap_slot *slot;
319 struct sge_nm_rxq *nm_rxq;
320 struct sge_nm_txq *nm_txq;
322 struct hw_buf_info *hwb;
324 ASSERT_SYNCHRONIZED_OP(sc);
326 if ((vi->flags & VI_INIT_DONE) == 0 ||
327 (ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
330 hwb = &sc->sge.hw_buf_info[0];
331 for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
332 if (hwb->size == NETMAP_BUF_SIZE(na))
335 if (i >= SGE_FLBUF_SIZES) {
336 if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
337 NETMAP_BUF_SIZE(na));
342 /* Must set caps before calling netmap_reset */
343 nm_set_native_flags(na);
345 for_each_nm_rxq(vi, i, nm_rxq) {
346 struct irq *irq = &sc->irq[vi->first_intr + i];
348 alloc_nm_rxq_hwq(vi, nm_rxq, tnl_cong(vi->pi, nm_cong_drop));
349 nm_rxq->fl_hwidx = hwidx;
350 slot = netmap_reset(na, NR_RX, i, 0);
351 MPASS(slot != NULL); /* XXXNM: error check, not assert */
353 /* We deal with 8 bufs at a time */
354 MPASS((na->num_rx_desc & 7) == 0);
355 MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
356 for (j = 0; j < nm_rxq->fl_sidx; j++) {
359 PNMB(na, &slot[j], &ba);
361 nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
363 j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
365 j /= 8; /* driver pidx to hardware pidx */
367 t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
368 nm_rxq->fl_db_val | V_PIDX(j));
370 atomic_cmpset_int(&irq->nm_state, NM_OFF, NM_ON);
373 for_each_nm_txq(vi, i, nm_txq) {
374 alloc_nm_txq_hwq(vi, nm_txq);
375 slot = netmap_reset(na, NR_TX, i, 0);
376 MPASS(slot != NULL); /* XXXNM: error check, not assert */
379 if (vi->nm_rss == NULL) {
380 vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE,
383 for (i = 0; i < vi->rss_size;) {
384 for_each_nm_rxq(vi, j, nm_rxq) {
385 vi->nm_rss[i++] = nm_rxq->iq_abs_id;
386 if (i == vi->rss_size)
390 rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
391 vi->nm_rss, vi->rss_size);
393 if_printf(ifp, "netmap rss_config failed: %d\n", rc);
399 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
400 struct netmap_adapter *na)
403 struct sge_nm_txq *nm_txq;
404 struct sge_nm_rxq *nm_rxq;
406 ASSERT_SYNCHRONIZED_OP(sc);
408 if ((vi->flags & VI_INIT_DONE) == 0)
411 rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
412 vi->rss, vi->rss_size);
414 if_printf(ifp, "failed to restore RSS config: %d\n", rc);
415 nm_clear_native_flags(na);
417 for_each_nm_txq(vi, i, nm_txq) {
418 struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
420 /* Wait for hw pidx to catch up ... */
421 while (be16toh(nm_txq->pidx) != spg->pidx)
424 /* ... and then for the cidx. */
425 while (spg->pidx != spg->cidx)
428 free_nm_txq_hwq(vi, nm_txq);
430 for_each_nm_rxq(vi, i, nm_rxq) {
431 struct irq *irq = &sc->irq[vi->first_intr + i];
433 while (!atomic_cmpset_int(&irq->nm_state, NM_ON, NM_OFF))
436 free_nm_rxq_hwq(vi, nm_rxq);
443 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
445 struct ifnet *ifp = na->ifp;
446 struct vi_info *vi = ifp->if_softc;
447 struct adapter *sc = vi->pi->adapter;
450 rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
454 rc = cxgbe_netmap_on(sc, vi, ifp, na);
456 rc = cxgbe_netmap_off(sc, vi, ifp, na);
457 end_synchronized_op(sc, 0);
462 /* How many packets can a single type1 WR carry in n descriptors */
464 ndesc_to_npkt(const int n)
467 MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
471 #define MAX_NPKT_IN_TYPE1_WR (ndesc_to_npkt(SGE_MAX_WR_NDESC))
473 /* Space (in descriptors) needed for a type1 WR that carries n packets */
475 npkt_to_ndesc(const int n)
478 MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
480 return ((n + 2) / 2);
483 /* Space (in 16B units) needed for a type1 WR that carries n packets */
485 npkt_to_len16(const int n)
488 MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
493 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
496 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
499 u_int db = nm_txq->doorbells;
501 MPASS(nm_txq->pidx != nm_txq->dbidx);
503 n = NMIDXDIFF(nm_txq, dbidx);
505 clrbit(&db, DOORBELL_WCWR);
508 switch (ffs(db) - 1) {
510 *nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
513 case DOORBELL_WCWR: {
514 volatile uint64_t *dst, *src;
517 * Queues whose 128B doorbell segment fits in the page do not
518 * use relative qid (udb_qid is always 0). Only queues with
519 * doorbell segments can do WCWR.
521 KASSERT(nm_txq->udb_qid == 0 && n == 1,
522 ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
523 __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
525 dst = (volatile void *)((uintptr_t)nm_txq->udb +
526 UDBS_WR_OFFSET - UDBS_DB_OFFSET);
527 src = (void *)&nm_txq->desc[nm_txq->dbidx];
528 while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
535 *nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
540 t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
541 V_QID(nm_txq->cntxt_id) | V_PIDX(n));
544 nm_txq->dbidx = nm_txq->pidx;
547 int lazy_tx_credit_flush = 1;
550 * Write work requests to send 'npkt' frames and ring the doorbell to send them
551 * on their way. No need to check for wraparound.
554 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
555 struct netmap_kring *kring, int npkt, int npkt_remaining, int txcsum)
557 struct netmap_ring *ring = kring->ring;
558 struct netmap_slot *slot;
559 const u_int lim = kring->nkr_num_slots - 1;
560 struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
563 struct cpl_tx_pkt_core *cpl;
564 struct ulptx_sgl *usgl;
568 n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
571 wr = (void *)&nm_txq->desc[nm_txq->pidx];
572 wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
573 wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
577 cpl = (void *)(wr + 1);
579 for (i = 0; i < n; i++) {
580 slot = &ring->slot[kring->nr_hwcur];
581 PNMB(kring->na, slot, &ba);
584 cpl->ctrl0 = nm_txq->cpl_ctrl0;
586 cpl->len = htobe16(slot->len);
588 * netmap(4) says "netmap does not use features such as
589 * checksum offloading, TCP segmentation offloading,
590 * encryption, VLAN encapsulation/decapsulation, etc."
592 * So the ncxl interfaces have tx hardware checksumming
593 * disabled by default. But you can override netmap by
594 * enabling IFCAP_TXCSUM on the interface manully.
596 cpl->ctrl1 = txcsum ? 0 :
597 htobe64(F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
599 usgl = (void *)(cpl + 1);
600 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
602 usgl->len0 = htobe32(slot->len);
603 usgl->addr0 = htobe64(ba);
605 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
606 cpl = (void *)(usgl + 1);
607 MPASS(slot->len + len <= UINT16_MAX);
609 kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
611 wr->plen = htobe16(len);
614 nm_txq->pidx += npkt_to_ndesc(n);
615 MPASS(nm_txq->pidx <= nm_txq->sidx);
616 if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
618 * This routine doesn't know how to write WRs that wrap
619 * around. Make sure it wasn't asked to.
625 if (npkt == 0 && npkt_remaining == 0) {
627 if (lazy_tx_credit_flush == 0) {
628 wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
630 nm_txq->equeqidx = nm_txq->pidx;
631 nm_txq->equiqidx = nm_txq->pidx;
633 ring_nm_txq_db(sc, nm_txq);
637 if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
638 wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
640 nm_txq->equeqidx = nm_txq->pidx;
641 nm_txq->equiqidx = nm_txq->pidx;
642 } else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
643 wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
644 nm_txq->equeqidx = nm_txq->pidx;
646 if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
647 ring_nm_txq_db(sc, nm_txq);
650 /* Will get called again. */
651 MPASS(npkt_remaining);
654 /* How many contiguous free descriptors starting at pidx */
656 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
659 if (nm_txq->cidx > nm_txq->pidx)
660 return (nm_txq->cidx - nm_txq->pidx - 1);
661 else if (nm_txq->cidx > 0)
662 return (nm_txq->sidx - nm_txq->pidx);
664 return (nm_txq->sidx - nm_txq->pidx - 1);
668 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
670 struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
671 uint16_t hw_cidx = spg->cidx; /* snapshot */
672 struct fw_eth_tx_pkts_wr *wr;
675 hw_cidx = be16toh(hw_cidx);
677 while (nm_txq->cidx != hw_cidx) {
678 wr = (void *)&nm_txq->desc[nm_txq->cidx];
680 MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)));
681 MPASS(wr->type == 1);
682 MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
685 nm_txq->cidx += npkt_to_ndesc(wr->npkt);
688 * We never sent a WR that wrapped around so the credits coming
689 * back, WR by WR, should never cause the cidx to wrap around
692 MPASS(nm_txq->cidx <= nm_txq->sidx);
693 if (__predict_false(nm_txq->cidx == nm_txq->sidx))
701 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
703 struct netmap_adapter *na = kring->na;
704 struct ifnet *ifp = na->ifp;
705 struct vi_info *vi = ifp->if_softc;
706 struct adapter *sc = vi->pi->adapter;
707 struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id];
708 const u_int head = kring->rhead;
710 int n, d, npkt_remaining, ndesc_remaining, txcsum;
713 * Tx was at kring->nr_hwcur last time around and now we need to advance
714 * to kring->rhead. Note that the driver's pidx moves independent of
715 * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
716 * between descriptors and frames isn't 1:1).
719 npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
720 kring->nkr_num_slots - kring->nr_hwcur + head;
721 txcsum = ifp->if_capenable & (IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6);
722 while (npkt_remaining) {
723 reclaimed += reclaim_nm_tx_desc(nm_txq);
724 ndesc_remaining = contiguous_ndesc_available(nm_txq);
725 /* Can't run out of descriptors with packets still remaining */
726 MPASS(ndesc_remaining > 0);
728 /* # of desc needed to tx all remaining packets */
729 d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
730 if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
731 d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
733 if (d <= ndesc_remaining)
736 /* Can't send all, calculate how many can be sent */
737 n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
738 MAX_NPKT_IN_TYPE1_WR;
739 if (ndesc_remaining % SGE_MAX_WR_NDESC)
740 n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
743 /* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
745 cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining, txcsum);
747 MPASS(npkt_remaining == 0);
748 MPASS(kring->nr_hwcur == head);
749 MPASS(nm_txq->dbidx == nm_txq->pidx);
752 * Second part: reclaim buffers for completed transmissions.
754 if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
755 reclaimed += reclaim_nm_tx_desc(nm_txq);
756 kring->nr_hwtail += reclaimed;
757 if (kring->nr_hwtail >= kring->nkr_num_slots)
758 kring->nr_hwtail -= kring->nkr_num_slots;
761 nm_txsync_finalize(kring);
767 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
769 struct netmap_adapter *na = kring->na;
770 struct netmap_ring *ring = kring->ring;
771 struct ifnet *ifp = na->ifp;
772 struct vi_info *vi = ifp->if_softc;
773 struct adapter *sc = vi->pi->adapter;
774 struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id];
775 u_int const head = nm_rxsync_prologue(kring);
777 int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
780 return (0); /* No updates ever. */
782 if (netmap_no_pendintr || force_update) {
783 kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
784 kring->nr_kflags &= ~NKR_PENDINTR;
787 /* Userspace done with buffers from kring->nr_hwcur to head */
788 n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
789 kring->nkr_num_slots - kring->nr_hwcur + head;
792 u_int fl_pidx = nm_rxq->fl_pidx;
793 struct netmap_slot *slot = &ring->slot[fl_pidx];
795 int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
798 * We always deal with 8 buffers at a time. We must have
799 * stopped at an 8B boundary (fl_pidx) last time around and we
800 * must have a multiple of 8B buffers to give to the freelist.
802 MPASS((fl_pidx & 7) == 0);
805 IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
806 IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx);
809 for (i = 0; i < 8; i++, fl_pidx++, slot++) {
812 nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
813 slot->flags &= ~NS_BUF_CHANGED;
814 MPASS(fl_pidx <= nm_rxq->fl_sidx);
817 if (fl_pidx == nm_rxq->fl_sidx) {
819 slot = &ring->slot[0];
821 if (++dbinc == 8 && n >= 32) {
823 t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
824 nm_rxq->fl_db_val | V_PIDX(dbinc));
828 MPASS(nm_rxq->fl_pidx == fl_pidx);
832 t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
833 nm_rxq->fl_db_val | V_PIDX(dbinc));
837 nm_rxsync_finalize(kring);
843 cxgbe_nm_attach(struct vi_info *vi)
845 struct port_info *pi;
847 struct netmap_adapter na;
849 MPASS(vi->nnmrxq > 0);
850 MPASS(vi->ifp != NULL);
855 bzero(&na, sizeof(na));
858 na.na_flags = NAF_BDG_MAYSLEEP;
860 /* Netmap doesn't know about the space reserved for the status page. */
861 na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE;
864 * The freelist's cidx/pidx drives netmap's rx cidx/pidx. So
865 * num_rx_desc is based on the number of buffers that can be held in the
866 * freelist, and not the number of entries in the iq. (These two are
867 * not exactly the same due to the space taken up by the status page).
869 na.num_rx_desc = (vi->qsize_rxq / 8) * 8;
870 na.nm_txsync = cxgbe_netmap_txsync;
871 na.nm_rxsync = cxgbe_netmap_rxsync;
872 na.nm_register = cxgbe_netmap_reg;
873 na.num_tx_rings = vi->nnmtxq;
874 na.num_rx_rings = vi->nnmrxq;
875 netmap_attach(&na); /* This adds IFCAP_NETMAP to if_capabilities */
879 cxgbe_nm_detach(struct vi_info *vi)
882 MPASS(vi->nnmrxq > 0);
883 MPASS(vi->ifp != NULL);
885 netmap_detach(vi->ifp);
889 handle_nm_fw6_msg(struct adapter *sc, struct ifnet *ifp,
890 const struct cpl_fw6_msg *cpl)
892 const struct cpl_sge_egr_update *egr;
894 struct sge_nm_txq *nm_txq;
896 if (cpl->type != FW_TYPE_RSSCPL && cpl->type != FW6_TYPE_RSSCPL)
897 panic("%s: FW_TYPE 0x%x on nm_rxq.", __func__, cpl->type);
899 /* data[0] is RSS header */
900 egr = (const void *)&cpl->data[1];
901 oq = be32toh(egr->opcode_qid);
902 MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
903 nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
905 netmap_tx_irq(ifp, nm_txq->nid);
909 t4_nm_intr(void *arg)
911 struct sge_nm_rxq *nm_rxq = arg;
912 struct vi_info *vi = nm_rxq->vi;
913 struct adapter *sc = vi->pi->adapter;
914 struct ifnet *ifp = vi->ifp;
915 struct netmap_adapter *na = NA(ifp);
916 struct netmap_kring *kring = &na->rx_rings[nm_rxq->nid];
917 struct netmap_ring *ring = kring->ring;
918 struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
920 u_int n = 0, work = 0;
922 uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
923 u_int fl_credits = fl_cidx & 7;
925 while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
929 lq = be32toh(d->rsp.pldbuflen_qid);
930 opcode = d->rss.opcode;
932 switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
933 case X_RSPD_TYPE_FLBUF:
934 if (black_hole != 2) {
935 /* No buffer packing so new buf every time */
936 MPASS(lq & F_RSPD_NEWBUF);
941 case X_RSPD_TYPE_CPL:
942 MPASS(opcode < NUM_CPL_CMDS);
947 handle_nm_fw6_msg(sc, ifp,
948 (const void *)&d->cpl[0]);
951 ring->slot[fl_cidx].len = G_RSPD_LEN(lq) -
952 sc->params.sge.fl_pktshift;
953 ring->slot[fl_cidx].flags = kring->nkr_slot_flags;
954 fl_cidx += (lq & F_RSPD_NEWBUF) ? 1 : 0;
955 fl_credits += (lq & F_RSPD_NEWBUF) ? 1 : 0;
956 if (__predict_false(fl_cidx == nm_rxq->fl_sidx))
960 panic("%s: unexpected opcode 0x%x on nm_rxq %p",
961 __func__, opcode, nm_rxq);
965 case X_RSPD_TYPE_INTR:
966 /* Not equipped to handle forwarded interrupts. */
967 panic("%s: netmap queue received interrupt for iq %u\n",
971 panic("%s: illegal response type %d on nm_rxq %p",
972 __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
976 if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
978 d = &nm_rxq->iq_desc[0];
979 nm_rxq->iq_gen ^= F_RSPD_GEN;
982 if (__predict_false(++n == rx_ndesc)) {
983 atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
984 if (black_hole && fl_credits >= 8) {
986 IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
988 t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
989 nm_rxq->fl_db_val | V_PIDX(fl_credits));
990 fl_credits = fl_cidx & 7;
991 } else if (!black_hole) {
992 netmap_rx_irq(ifp, nm_rxq->nid, &work);
995 t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
996 V_CIDXINC(n) | V_INGRESSQID(nm_rxq->iq_cntxt_id) |
997 V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1002 atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1005 IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1006 t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
1007 nm_rxq->fl_db_val | V_PIDX(fl_credits));
1009 netmap_rx_irq(ifp, nm_rxq->nid, &work);
1011 t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(n) |
1012 V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1013 V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));