]> CyberLeo.Net >> Repos - FreeBSD/stable/10.git/blob - sys/dev/cxgbe/t4_netmap.c
MFC 305695,305696,305699,305702,305703,305713,305715,305827,305852,305906,
[FreeBSD/stable/10.git] / sys / dev / cxgbe / t4_netmap.c
1 /*-
2  * Copyright (c) 2014 Chelsio Communications, Inc.
3  * All rights reserved.
4  * Written by: Navdeep Parhar <np@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
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.
14  *
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
25  * SUCH DAMAGE.
26  */
27
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33
34 #ifdef DEV_NETMAP
35 #include <sys/param.h>
36 #include <sys/bus.h>
37 #include <sys/eventhandler.h>
38 #include <sys/lock.h>
39 #include <sys/mbuf.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>
46 #include <net/if.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>
53
54 #include "common/common.h"
55 #include "common/t4_regs.h"
56 #include "common/t4_regs_values.h"
57
58 extern int fl_pad;      /* XXXNM */
59
60 SYSCTL_NODE(_hw, OID_AUTO, cxgbe, CTLFLAG_RD, 0, "cxgbe netmap parameters");
61
62 /*
63  * 0 = normal netmap rx
64  * 1 = black hole
65  * 2 = supermassive black hole (buffer packing enabled)
66  */
67 int black_hole = 0;
68 SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_black_hole, CTLFLAG_RDTUN, &black_hole, 0,
69     "Sink incoming packets.");
70
71 int rx_ndesc = 256;
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.");
74
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.");
78
79 /*
80  * Congestion drops.
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.
84  */
85 static int nm_cong_drop = 1;
86 TUNABLE_INT("hw.cxgbe.nm_cong_drop", &nm_cong_drop);
87
88 static int
89 alloc_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int cong)
90 {
91         int rc, cntxt_id, i;
92         __be32 v;
93         struct adapter *sc = vi->pi->adapter;
94         struct sge_params *sp = &sc->params.sge;
95         struct netmap_adapter *na = NA(vi->ifp);
96         struct fw_iq_cmd c;
97
98         MPASS(na != NULL);
99         MPASS(nm_rxq->iq_desc != NULL);
100         MPASS(nm_rxq->fl_desc != NULL);
101
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);
104
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) |
108             V_FW_IQ_CMD_VFN(0));
109         c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
110             FW_LEN16(c));
111         if (vi->flags & INTR_RXQ) {
112                 KASSERT(nm_rxq->intr_idx < sc->intr_count,
113                     ("%s: invalid direct intr_idx %d", __func__,
114                     nm_rxq->intr_idx));
115                 v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx);
116         } else {
117                 CXGBE_UNIMPLEMENTED(__func__);  /* XXXNM: needs review */
118                 v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx) |
119                     F_FW_IQ_CMD_IQANDST;
120         }
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);
131         if (cong >= 0) {
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);
135         }
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(chip_id(sc) <= CHELSIO_T5 ?
143                 X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
144                 V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
145                 X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
146         c.fl0size = htobe16(na->num_rx_desc / 8 + sp->spg_len / EQ_ESIZE);
147         c.fl0addr = htobe64(nm_rxq->fl_ba);
148
149         rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
150         if (rc != 0) {
151                 device_printf(sc->dev,
152                     "failed to create netmap ingress queue: %d\n", rc);
153                 return (rc);
154         }
155
156         nm_rxq->iq_cidx = 0;
157         MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - sp->spg_len / IQ_ESIZE);
158         nm_rxq->iq_gen = F_RSPD_GEN;
159         nm_rxq->iq_cntxt_id = be16toh(c.iqid);
160         nm_rxq->iq_abs_id = be16toh(c.physiqid);
161         cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start;
162         if (cntxt_id >= sc->sge.niq) {
163                 panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)",
164                     __func__, cntxt_id, sc->sge.niq - 1);
165         }
166         sc->sge.iqmap[cntxt_id] = (void *)nm_rxq;
167
168         nm_rxq->fl_cntxt_id = be16toh(c.fl0id);
169         nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
170         MPASS(nm_rxq->fl_sidx == na->num_rx_desc);
171         cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start;
172         if (cntxt_id >= sc->sge.neq) {
173                 panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)",
174                     __func__, cntxt_id, sc->sge.neq - 1);
175         }
176         sc->sge.eqmap[cntxt_id] = (void *)nm_rxq;
177
178         nm_rxq->fl_db_val = V_QID(nm_rxq->fl_cntxt_id) |
179             sc->chip_params->sge_fl_db;
180
181         if (chip_id(sc) >= CHELSIO_T5 && cong >= 0) {
182                 uint32_t param, val;
183
184                 param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
185                     V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
186                     V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
187                 param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
188                     V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
189                     V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id);
190                 if (cong == 0)
191                         val = 1 << 19;
192                 else {
193                         val = 2 << 19;
194                         for (i = 0; i < 4; i++) {
195                                 if (cong & (1 << i))
196                                         val |= 1 << (i << 2);
197                         }
198                 }
199
200                 rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
201                 if (rc != 0) {
202                         /* report error but carry on */
203                         device_printf(sc->dev,
204                             "failed to set congestion manager context for "
205                             "ingress queue %d: %d\n", nm_rxq->iq_cntxt_id, rc);
206                 }
207         }
208
209         t4_write_reg(sc, sc->sge_gts_reg,
210             V_INGRESSQID(nm_rxq->iq_cntxt_id) |
211             V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
212
213         return (rc);
214 }
215
216 static int
217 free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
218 {
219         struct adapter *sc = vi->pi->adapter;
220         int rc;
221
222         rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
223             nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff);
224         if (rc != 0)
225                 device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n",
226                     __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc);
227         return (rc);
228 }
229
230 static int
231 alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
232 {
233         int rc, cntxt_id;
234         size_t len;
235         struct adapter *sc = vi->pi->adapter;
236         struct netmap_adapter *na = NA(vi->ifp);
237         struct fw_eq_eth_cmd c;
238
239         MPASS(na != NULL);
240         MPASS(nm_txq->desc != NULL);
241
242         len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
243         bzero(nm_txq->desc, len);
244
245         bzero(&c, sizeof(c));
246         c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
247             F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
248             V_FW_EQ_ETH_CMD_VFN(0));
249         c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
250             F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
251         c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
252             F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
253         c.fetchszm_to_iqid =
254             htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
255                 V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
256                 V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id));
257         c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
258                       V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
259                       V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE));
260         c.eqaddr = htobe64(nm_txq->ba);
261
262         rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
263         if (rc != 0) {
264                 device_printf(vi->dev,
265                     "failed to create netmap egress queue: %d\n", rc);
266                 return (rc);
267         }
268
269         nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
270         cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start;
271         if (cntxt_id >= sc->sge.neq)
272             panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__,
273                 cntxt_id, sc->sge.neq - 1);
274         sc->sge.eqmap[cntxt_id] = (void *)nm_txq;
275
276         nm_txq->pidx = nm_txq->cidx = 0;
277         MPASS(nm_txq->sidx == na->num_tx_desc);
278         nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0;
279
280         nm_txq->doorbells = sc->doorbells;
281         if (isset(&nm_txq->doorbells, DOORBELL_UDB) ||
282             isset(&nm_txq->doorbells, DOORBELL_UDBWC) ||
283             isset(&nm_txq->doorbells, DOORBELL_WCWR)) {
284                 uint32_t s_qpp = sc->params.sge.eq_s_qpp;
285                 uint32_t mask = (1 << s_qpp) - 1;
286                 volatile uint8_t *udb;
287
288                 udb = sc->udbs_base + UDBS_DB_OFFSET;
289                 udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT;
290                 nm_txq->udb_qid = nm_txq->cntxt_id & mask;
291                 if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
292                         clrbit(&nm_txq->doorbells, DOORBELL_WCWR);
293                 else {
294                         udb += nm_txq->udb_qid << UDBS_SEG_SHIFT;
295                         nm_txq->udb_qid = 0;
296                 }
297                 nm_txq->udb = (volatile void *)udb;
298         }
299
300         return (rc);
301 }
302
303 static int
304 free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
305 {
306         struct adapter *sc = vi->pi->adapter;
307         int rc;
308
309         rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id);
310         if (rc != 0)
311                 device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__,
312                     nm_txq->cntxt_id, rc);
313         return (rc);
314 }
315
316 static int
317 cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
318     struct netmap_adapter *na)
319 {
320         struct netmap_slot *slot;
321         struct sge_nm_rxq *nm_rxq;
322         struct sge_nm_txq *nm_txq;
323         int rc, i, j, hwidx;
324         struct hw_buf_info *hwb;
325
326         ASSERT_SYNCHRONIZED_OP(sc);
327
328         if ((vi->flags & VI_INIT_DONE) == 0 ||
329             (ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
330                 return (EAGAIN);
331
332         hwb = &sc->sge.hw_buf_info[0];
333         for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
334                 if (hwb->size == NETMAP_BUF_SIZE(na))
335                         break;
336         }
337         if (i >= SGE_FLBUF_SIZES) {
338                 if_printf(ifp, "no hwidx for netmap buffer size %d.\n",
339                     NETMAP_BUF_SIZE(na));
340                 return (ENXIO);
341         }
342         hwidx = i;
343
344         /* Must set caps before calling netmap_reset */
345         nm_set_native_flags(na);
346
347         for_each_nm_rxq(vi, i, nm_rxq) {
348                 struct irq *irq = &sc->irq[vi->first_intr + i];
349
350                 alloc_nm_rxq_hwq(vi, nm_rxq, tnl_cong(vi->pi, nm_cong_drop));
351                 nm_rxq->fl_hwidx = hwidx;
352                 slot = netmap_reset(na, NR_RX, i, 0);
353                 MPASS(slot != NULL);    /* XXXNM: error check, not assert */
354
355                 /* We deal with 8 bufs at a time */
356                 MPASS((na->num_rx_desc & 7) == 0);
357                 MPASS(na->num_rx_desc == nm_rxq->fl_sidx);
358                 for (j = 0; j < nm_rxq->fl_sidx; j++) {
359                         uint64_t ba;
360
361                         PNMB(na, &slot[j], &ba);
362                         MPASS(ba != 0);
363                         nm_rxq->fl_desc[j] = htobe64(ba | hwidx);
364                 }
365                 j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8;
366                 MPASS((j & 7) == 0);
367                 j /= 8; /* driver pidx to hardware pidx */
368                 wmb();
369                 t4_write_reg(sc, sc->sge_kdoorbell_reg,
370                     nm_rxq->fl_db_val | V_PIDX(j));
371
372                 atomic_cmpset_int(&irq->nm_state, NM_OFF, NM_ON);
373         }
374
375         for_each_nm_txq(vi, i, nm_txq) {
376                 alloc_nm_txq_hwq(vi, nm_txq);
377                 slot = netmap_reset(na, NR_TX, i, 0);
378                 MPASS(slot != NULL);    /* XXXNM: error check, not assert */
379         }
380
381         if (vi->nm_rss == NULL) {
382                 vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE,
383                     M_ZERO | M_WAITOK);
384         }
385         for (i = 0; i < vi->rss_size;) {
386                 for_each_nm_rxq(vi, j, nm_rxq) {
387                         vi->nm_rss[i++] = nm_rxq->iq_abs_id;
388                         if (i == vi->rss_size)
389                                 break;
390                 }
391         }
392         rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
393             vi->nm_rss, vi->rss_size);
394         if (rc != 0)
395                 if_printf(ifp, "netmap rss_config failed: %d\n", rc);
396
397         return (rc);
398 }
399
400 static int
401 cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp,
402     struct netmap_adapter *na)
403 {
404         int rc, i;
405         struct sge_nm_txq *nm_txq;
406         struct sge_nm_rxq *nm_rxq;
407
408         ASSERT_SYNCHRONIZED_OP(sc);
409
410         if ((vi->flags & VI_INIT_DONE) == 0)
411                 return (0);
412
413         rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size,
414             vi->rss, vi->rss_size);
415         if (rc != 0)
416                 if_printf(ifp, "failed to restore RSS config: %d\n", rc);
417         nm_clear_native_flags(na);
418
419         for_each_nm_txq(vi, i, nm_txq) {
420                 struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
421
422                 /* Wait for hw pidx to catch up ... */
423                 while (be16toh(nm_txq->pidx) != spg->pidx)
424                         pause("nmpidx", 1);
425
426                 /* ... and then for the cidx. */
427                 while (spg->pidx != spg->cidx)
428                         pause("nmcidx", 1);
429
430                 free_nm_txq_hwq(vi, nm_txq);
431         }
432         for_each_nm_rxq(vi, i, nm_rxq) {
433                 struct irq *irq = &sc->irq[vi->first_intr + i];
434
435                 while (!atomic_cmpset_int(&irq->nm_state, NM_ON, NM_OFF))
436                         pause("nmst", 1);
437
438                 free_nm_rxq_hwq(vi, nm_rxq);
439         }
440
441         return (rc);
442 }
443
444 static int
445 cxgbe_netmap_reg(struct netmap_adapter *na, int on)
446 {
447         struct ifnet *ifp = na->ifp;
448         struct vi_info *vi = ifp->if_softc;
449         struct adapter *sc = vi->pi->adapter;
450         int rc;
451
452         rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg");
453         if (rc != 0)
454                 return (rc);
455         if (on)
456                 rc = cxgbe_netmap_on(sc, vi, ifp, na);
457         else
458                 rc = cxgbe_netmap_off(sc, vi, ifp, na);
459         end_synchronized_op(sc, 0);
460
461         return (rc);
462 }
463
464 /* How many packets can a single type1 WR carry in n descriptors */
465 static inline int
466 ndesc_to_npkt(const int n)
467 {
468
469         MPASS(n > 0 && n <= SGE_MAX_WR_NDESC);
470
471         return (n * 2 - 1);
472 }
473 #define MAX_NPKT_IN_TYPE1_WR    (ndesc_to_npkt(SGE_MAX_WR_NDESC))
474
475 /* Space (in descriptors) needed for a type1 WR that carries n packets */
476 static inline int
477 npkt_to_ndesc(const int n)
478 {
479
480         MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
481
482         return ((n + 2) / 2);
483 }
484
485 /* Space (in 16B units) needed for a type1 WR that carries n packets */
486 static inline int
487 npkt_to_len16(const int n)
488 {
489
490         MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR);
491
492         return (n * 2 + 1);
493 }
494
495 #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx)
496
497 static void
498 ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq)
499 {
500         int n;
501         u_int db = nm_txq->doorbells;
502
503         MPASS(nm_txq->pidx != nm_txq->dbidx);
504
505         n = NMIDXDIFF(nm_txq, dbidx);
506         if (n > 1)
507                 clrbit(&db, DOORBELL_WCWR);
508         wmb();
509
510         switch (ffs(db) - 1) {
511         case DOORBELL_UDB:
512                 *nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
513                 break;
514
515         case DOORBELL_WCWR: {
516                 volatile uint64_t *dst, *src;
517
518                 /*
519                  * Queues whose 128B doorbell segment fits in the page do not
520                  * use relative qid (udb_qid is always 0).  Only queues with
521                  * doorbell segments can do WCWR.
522                  */
523                 KASSERT(nm_txq->udb_qid == 0 && n == 1,
524                     ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p",
525                     __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq));
526
527                 dst = (volatile void *)((uintptr_t)nm_txq->udb +
528                     UDBS_WR_OFFSET - UDBS_DB_OFFSET);
529                 src = (void *)&nm_txq->desc[nm_txq->dbidx];
530                 while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1])
531                         *dst++ = *src++;
532                 wmb();
533                 break;
534         }
535
536         case DOORBELL_UDBWC:
537                 *nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n));
538                 wmb();
539                 break;
540
541         case DOORBELL_KDB:
542                 t4_write_reg(sc, sc->sge_kdoorbell_reg,
543                     V_QID(nm_txq->cntxt_id) | V_PIDX(n));
544                 break;
545         }
546         nm_txq->dbidx = nm_txq->pidx;
547 }
548
549 int lazy_tx_credit_flush = 1;
550
551 /*
552  * Write work requests to send 'npkt' frames and ring the doorbell to send them
553  * on their way.  No need to check for wraparound.
554  */
555 static void
556 cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq,
557     struct netmap_kring *kring, int npkt, int npkt_remaining, int txcsum)
558 {
559         struct netmap_ring *ring = kring->ring;
560         struct netmap_slot *slot;
561         const u_int lim = kring->nkr_num_slots - 1;
562         struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx];
563         uint16_t len;
564         uint64_t ba;
565         struct cpl_tx_pkt_core *cpl;
566         struct ulptx_sgl *usgl;
567         int i, n;
568
569         while (npkt) {
570                 n = min(npkt, MAX_NPKT_IN_TYPE1_WR);
571                 len = 0;
572
573                 wr = (void *)&nm_txq->desc[nm_txq->pidx];
574                 wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
575                 wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n)));
576                 wr->npkt = n;
577                 wr->r3 = 0;
578                 wr->type = 1;
579                 cpl = (void *)(wr + 1);
580
581                 for (i = 0; i < n; i++) {
582                         slot = &ring->slot[kring->nr_hwcur];
583                         PNMB(kring->na, slot, &ba);
584                         MPASS(ba != 0);
585
586                         cpl->ctrl0 = nm_txq->cpl_ctrl0;
587                         cpl->pack = 0;
588                         cpl->len = htobe16(slot->len);
589                         /*
590                          * netmap(4) says "netmap does not use features such as
591                          * checksum offloading, TCP segmentation offloading,
592                          * encryption, VLAN encapsulation/decapsulation, etc."
593                          *
594                          * So the ncxl interfaces have tx hardware checksumming
595                          * disabled by default.  But you can override netmap by
596                          * enabling IFCAP_TXCSUM on the interface manully.
597                          */
598                         cpl->ctrl1 = txcsum ? 0 :
599                             htobe64(F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
600
601                         usgl = (void *)(cpl + 1);
602                         usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
603                             V_ULPTX_NSGE(1));
604                         usgl->len0 = htobe32(slot->len);
605                         usgl->addr0 = htobe64(ba);
606
607                         slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
608                         cpl = (void *)(usgl + 1);
609                         MPASS(slot->len + len <= UINT16_MAX);
610                         len += slot->len;
611                         kring->nr_hwcur = nm_next(kring->nr_hwcur, lim);
612                 }
613                 wr->plen = htobe16(len);
614
615                 npkt -= n;
616                 nm_txq->pidx += npkt_to_ndesc(n);
617                 MPASS(nm_txq->pidx <= nm_txq->sidx);
618                 if (__predict_false(nm_txq->pidx == nm_txq->sidx)) {
619                         /*
620                          * This routine doesn't know how to write WRs that wrap
621                          * around.  Make sure it wasn't asked to.
622                          */
623                         MPASS(npkt == 0);
624                         nm_txq->pidx = 0;
625                 }
626
627                 if (npkt == 0 && npkt_remaining == 0) {
628                         /* All done. */
629                         if (lazy_tx_credit_flush == 0) {
630                                 wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
631                                     F_FW_WR_EQUIQ);
632                                 nm_txq->equeqidx = nm_txq->pidx;
633                                 nm_txq->equiqidx = nm_txq->pidx;
634                         }
635                         ring_nm_txq_db(sc, nm_txq);
636                         return;
637                 }
638
639                 if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) {
640                         wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ |
641                             F_FW_WR_EQUIQ);
642                         nm_txq->equeqidx = nm_txq->pidx;
643                         nm_txq->equiqidx = nm_txq->pidx;
644                 } else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) {
645                         wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
646                         nm_txq->equeqidx = nm_txq->pidx;
647                 }
648                 if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC)
649                         ring_nm_txq_db(sc, nm_txq);
650         }
651
652         /* Will get called again. */
653         MPASS(npkt_remaining);
654 }
655
656 /* How many contiguous free descriptors starting at pidx */
657 static inline int
658 contiguous_ndesc_available(struct sge_nm_txq *nm_txq)
659 {
660
661         if (nm_txq->cidx > nm_txq->pidx)
662                 return (nm_txq->cidx - nm_txq->pidx - 1);
663         else if (nm_txq->cidx > 0)
664                 return (nm_txq->sidx - nm_txq->pidx);
665         else
666                 return (nm_txq->sidx - nm_txq->pidx - 1);
667 }
668
669 static int
670 reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq)
671 {
672         struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx];
673         uint16_t hw_cidx = spg->cidx;   /* snapshot */
674         struct fw_eth_tx_pkts_wr *wr;
675         int n = 0;
676
677         hw_cidx = be16toh(hw_cidx);
678
679         while (nm_txq->cidx != hw_cidx) {
680                 wr = (void *)&nm_txq->desc[nm_txq->cidx];
681
682                 MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)));
683                 MPASS(wr->type == 1);
684                 MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR);
685
686                 n += wr->npkt;
687                 nm_txq->cidx += npkt_to_ndesc(wr->npkt);
688
689                 /*
690                  * We never sent a WR that wrapped around so the credits coming
691                  * back, WR by WR, should never cause the cidx to wrap around
692                  * either.
693                  */
694                 MPASS(nm_txq->cidx <= nm_txq->sidx);
695                 if (__predict_false(nm_txq->cidx == nm_txq->sidx))
696                         nm_txq->cidx = 0;
697         }
698
699         return (n);
700 }
701
702 static int
703 cxgbe_netmap_txsync(struct netmap_kring *kring, int flags)
704 {
705         struct netmap_adapter *na = kring->na;
706         struct ifnet *ifp = na->ifp;
707         struct vi_info *vi = ifp->if_softc;
708         struct adapter *sc = vi->pi->adapter;
709         struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id];
710         const u_int head = kring->rhead;
711         u_int reclaimed = 0;
712         int n, d, npkt_remaining, ndesc_remaining, txcsum;
713
714         /*
715          * Tx was at kring->nr_hwcur last time around and now we need to advance
716          * to kring->rhead.  Note that the driver's pidx moves independent of
717          * netmap's kring->nr_hwcur (pidx counts descriptors and the relation
718          * between descriptors and frames isn't 1:1).
719          */
720
721         npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
722             kring->nkr_num_slots - kring->nr_hwcur + head;
723         txcsum = ifp->if_capenable & (IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6);
724         while (npkt_remaining) {
725                 reclaimed += reclaim_nm_tx_desc(nm_txq);
726                 ndesc_remaining = contiguous_ndesc_available(nm_txq);
727                 /* Can't run out of descriptors with packets still remaining */
728                 MPASS(ndesc_remaining > 0);
729
730                 /* # of desc needed to tx all remaining packets */
731                 d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC;
732                 if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR)
733                         d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR);
734
735                 if (d <= ndesc_remaining)
736                         n = npkt_remaining;
737                 else {
738                         /* Can't send all, calculate how many can be sent */
739                         n = (ndesc_remaining / SGE_MAX_WR_NDESC) *
740                             MAX_NPKT_IN_TYPE1_WR;
741                         if (ndesc_remaining % SGE_MAX_WR_NDESC)
742                                 n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC);
743                 }
744
745                 /* Send n packets and update nm_txq->pidx and kring->nr_hwcur */
746                 npkt_remaining -= n;
747                 cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining, txcsum);
748         }
749         MPASS(npkt_remaining == 0);
750         MPASS(kring->nr_hwcur == head);
751         MPASS(nm_txq->dbidx == nm_txq->pidx);
752
753         /*
754          * Second part: reclaim buffers for completed transmissions.
755          */
756         if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
757                 reclaimed += reclaim_nm_tx_desc(nm_txq);
758                 kring->nr_hwtail += reclaimed;
759                 if (kring->nr_hwtail >= kring->nkr_num_slots)
760                         kring->nr_hwtail -= kring->nkr_num_slots;
761         }
762
763         nm_txsync_finalize(kring);
764
765         return (0);
766 }
767
768 static int
769 cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
770 {
771         struct netmap_adapter *na = kring->na;
772         struct netmap_ring *ring = kring->ring;
773         struct ifnet *ifp = na->ifp;
774         struct vi_info *vi = ifp->if_softc;
775         struct adapter *sc = vi->pi->adapter;
776         struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id];
777         u_int const head = nm_rxsync_prologue(kring);
778         u_int n;
779         int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
780
781         if (black_hole)
782                 return (0);     /* No updates ever. */
783
784         if (netmap_no_pendintr || force_update) {
785                 kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx);
786                 kring->nr_kflags &= ~NKR_PENDINTR;
787         }
788
789         /* Userspace done with buffers from kring->nr_hwcur to head */
790         n = head >= kring->nr_hwcur ? head - kring->nr_hwcur :
791             kring->nkr_num_slots - kring->nr_hwcur + head;
792         n &= ~7U;
793         if (n > 0) {
794                 u_int fl_pidx = nm_rxq->fl_pidx;
795                 struct netmap_slot *slot = &ring->slot[fl_pidx];
796                 uint64_t ba;
797                 int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx;
798
799                 /*
800                  * We always deal with 8 buffers at a time.  We must have
801                  * stopped at an 8B boundary (fl_pidx) last time around and we
802                  * must have a multiple of 8B buffers to give to the freelist.
803                  */
804                 MPASS((fl_pidx & 7) == 0);
805                 MPASS((n & 7) == 0);
806
807                 IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots);
808                 IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx);
809
810                 while (n > 0) {
811                         for (i = 0; i < 8; i++, fl_pidx++, slot++) {
812                                 PNMB(na, slot, &ba);
813                                 MPASS(ba != 0);
814                                 nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx);
815                                 slot->flags &= ~NS_BUF_CHANGED;
816                                 MPASS(fl_pidx <= nm_rxq->fl_sidx);
817                         }
818                         n -= 8;
819                         if (fl_pidx == nm_rxq->fl_sidx) {
820                                 fl_pidx = 0;
821                                 slot = &ring->slot[0];
822                         }
823                         if (++dbinc == 8 && n >= 32) {
824                                 wmb();
825                                 t4_write_reg(sc, sc->sge_kdoorbell_reg,
826                                     nm_rxq->fl_db_val | V_PIDX(dbinc));
827                                 dbinc = 0;
828                         }
829                 }
830                 MPASS(nm_rxq->fl_pidx == fl_pidx);
831
832                 if (dbinc > 0) {
833                         wmb();
834                         t4_write_reg(sc, sc->sge_kdoorbell_reg,
835                             nm_rxq->fl_db_val | V_PIDX(dbinc));
836                 }
837         }
838
839         nm_rxsync_finalize(kring);
840
841         return (0);
842 }
843
844 void
845 cxgbe_nm_attach(struct vi_info *vi)
846 {
847         struct port_info *pi;
848         struct adapter *sc;
849         struct netmap_adapter na;
850
851         MPASS(vi->nnmrxq > 0);
852         MPASS(vi->ifp != NULL);
853
854         pi = vi->pi;
855         sc = pi->adapter;
856
857         bzero(&na, sizeof(na));
858
859         na.ifp = vi->ifp;
860         na.na_flags = NAF_BDG_MAYSLEEP;
861
862         /* Netmap doesn't know about the space reserved for the status page. */
863         na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE;
864
865         /*
866          * The freelist's cidx/pidx drives netmap's rx cidx/pidx.  So
867          * num_rx_desc is based on the number of buffers that can be held in the
868          * freelist, and not the number of entries in the iq.  (These two are
869          * not exactly the same due to the space taken up by the status page).
870          */
871         na.num_rx_desc = (vi->qsize_rxq / 8) * 8;
872         na.nm_txsync = cxgbe_netmap_txsync;
873         na.nm_rxsync = cxgbe_netmap_rxsync;
874         na.nm_register = cxgbe_netmap_reg;
875         na.num_tx_rings = vi->nnmtxq;
876         na.num_rx_rings = vi->nnmrxq;
877         netmap_attach(&na);     /* This adds IFCAP_NETMAP to if_capabilities */
878 }
879
880 void
881 cxgbe_nm_detach(struct vi_info *vi)
882 {
883
884         MPASS(vi->nnmrxq > 0);
885         MPASS(vi->ifp != NULL);
886
887         netmap_detach(vi->ifp);
888 }
889
890 static inline const void *
891 unwrap_nm_fw6_msg(const struct cpl_fw6_msg *cpl)
892 {
893
894         MPASS(cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL);
895
896         /* data[0] is RSS header */
897         return (&cpl->data[1]);
898 }
899
900 static void
901 handle_nm_sge_egr_update(struct adapter *sc, struct ifnet *ifp,
902     const struct cpl_sge_egr_update *egr)
903 {
904         uint32_t oq;
905         struct sge_nm_txq *nm_txq;
906
907         oq = be32toh(egr->opcode_qid);
908         MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE);
909         nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start];
910
911         netmap_tx_irq(ifp, nm_txq->nid);
912 }
913
914 void
915 t4_nm_intr(void *arg)
916 {
917         struct sge_nm_rxq *nm_rxq = arg;
918         struct vi_info *vi = nm_rxq->vi;
919         struct adapter *sc = vi->pi->adapter;
920         struct ifnet *ifp = vi->ifp;
921         struct netmap_adapter *na = NA(ifp);
922         struct netmap_kring *kring = &na->rx_rings[nm_rxq->nid];
923         struct netmap_ring *ring = kring->ring;
924         struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx];
925         const void *cpl;
926         uint32_t lq;
927         u_int n = 0, work = 0;
928         uint8_t opcode;
929         uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx);
930         u_int fl_credits = fl_cidx & 7;
931
932         while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) {
933
934                 rmb();
935
936                 lq = be32toh(d->rsp.pldbuflen_qid);
937                 opcode = d->rss.opcode;
938                 cpl = &d->cpl[0];
939
940                 switch (G_RSPD_TYPE(d->rsp.u.type_gen)) {
941                 case X_RSPD_TYPE_FLBUF:
942                         if (black_hole != 2) {
943                                 /* No buffer packing so new buf every time */
944                                 MPASS(lq & F_RSPD_NEWBUF);
945                         }
946
947                         /* fall through */
948
949                 case X_RSPD_TYPE_CPL:
950                         MPASS(opcode < NUM_CPL_CMDS);
951
952                         switch (opcode) {
953                         case CPL_FW4_MSG:
954                         case CPL_FW6_MSG:
955                                 cpl = unwrap_nm_fw6_msg(cpl);
956                                 /* fall through */
957                         case CPL_SGE_EGR_UPDATE:
958                                 handle_nm_sge_egr_update(sc, ifp, cpl);
959                                 break;
960                         case CPL_RX_PKT:
961                                 ring->slot[fl_cidx].len = G_RSPD_LEN(lq) -
962                                     sc->params.sge.fl_pktshift;
963                                 ring->slot[fl_cidx].flags = kring->nkr_slot_flags;
964                                 fl_cidx += (lq & F_RSPD_NEWBUF) ? 1 : 0;
965                                 fl_credits += (lq & F_RSPD_NEWBUF) ? 1 : 0;
966                                 if (__predict_false(fl_cidx == nm_rxq->fl_sidx))
967                                         fl_cidx = 0;
968                                 break;
969                         default:
970                                 panic("%s: unexpected opcode 0x%x on nm_rxq %p",
971                                     __func__, opcode, nm_rxq);
972                         }
973                         break;
974
975                 case X_RSPD_TYPE_INTR:
976                         /* Not equipped to handle forwarded interrupts. */
977                         panic("%s: netmap queue received interrupt for iq %u\n",
978                             __func__, lq);
979
980                 default:
981                         panic("%s: illegal response type %d on nm_rxq %p",
982                             __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq);
983                 }
984
985                 d++;
986                 if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) {
987                         nm_rxq->iq_cidx = 0;
988                         d = &nm_rxq->iq_desc[0];
989                         nm_rxq->iq_gen ^= F_RSPD_GEN;
990                 }
991
992                 if (__predict_false(++n == rx_ndesc)) {
993                         atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
994                         if (black_hole && fl_credits >= 8) {
995                                 fl_credits /= 8;
996                                 IDXINCR(nm_rxq->fl_pidx, fl_credits * 8,
997                                     nm_rxq->fl_sidx);
998                                 t4_write_reg(sc, sc->sge_kdoorbell_reg,
999                                     nm_rxq->fl_db_val | V_PIDX(fl_credits));
1000                                 fl_credits = fl_cidx & 7;
1001                         } else if (!black_hole) {
1002                                 netmap_rx_irq(ifp, nm_rxq->nid, &work);
1003                                 MPASS(work != 0);
1004                         }
1005                         t4_write_reg(sc, sc->sge_gts_reg,
1006                             V_CIDXINC(n) | V_INGRESSQID(nm_rxq->iq_cntxt_id) |
1007                             V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1008                         n = 0;
1009                 }
1010         }
1011
1012         atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx);
1013         if (black_hole) {
1014                 fl_credits /= 8;
1015                 IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx);
1016                 t4_write_reg(sc, sc->sge_kdoorbell_reg,
1017                     nm_rxq->fl_db_val | V_PIDX(fl_credits));
1018         } else
1019                 netmap_rx_irq(ifp, nm_rxq->nid, &work);
1020
1021         t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(n) |
1022             V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) |
1023             V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx)));
1024 }
1025 #endif