xref: /illumos-gate/usr/src/uts/common/io/cxgbe/t4nex/t4_sge.c (revision d35eb49b13b481dcf2e0b66df3d6b2fad04f3572)
1 /*
2  * This file and its contents are supplied under the terms of the
3  * Common Development and Distribution License ("CDDL"), version 1.0.
4  * You may only use this file in accordance with the terms of version
5  * 1.0 of the CDDL.
6  *
7  * A full copy of the text of the CDDL should have accompanied this
8  * source. A copy of the CDDL is also available via the Internet at
9  * http://www.illumos.org/license/CDDL.
10  */
11 
12 /*
13  * This file is part of the Chelsio T4 support code.
14  *
15  * Copyright (C) 2010-2013 Chelsio Communications.  All rights reserved.
16  *
17  * This program is distributed in the hope that it will be useful, but WITHOUT
18  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
19  * FITNESS FOR A PARTICULAR PURPOSE.  See the LICENSE file included in this
20  * release for licensing terms and conditions.
21  */
22 
23 /*
24  * Copyright 2025 Oxide Computer Company
25  */
26 
27 #include <sys/ddi.h>
28 #include <sys/sunddi.h>
29 #include <sys/sunndi.h>
30 #include <sys/atomic.h>
31 #include <sys/dlpi.h>
32 #include <sys/pattr.h>
33 #include <sys/strsubr.h>
34 #include <sys/stream.h>
35 #include <sys/strsun.h>
36 #include <inet/ip.h>
37 #include <inet/tcp.h>
38 
39 #include "common/common.h"
40 #include "common/t4_msg.h"
41 #include "common/t4_regs.h"
42 #include "common/t4_regs_values.h"
43 
44 /* TODO: Tune. */
45 int rx_buf_size = 8192;
46 int tx_copy_threshold = 256;
47 uint16_t rx_copy_threshold = 256;
48 
49 /* Used to track coalesced tx work request */
50 struct txpkts {
51 	mblk_t *tail;		/* head is in the software descriptor */
52 	uint64_t *flitp;	/* ptr to flit where next pkt should start */
53 	uint8_t npkt;		/* # of packets in this work request */
54 	uint8_t nflits;		/* # of flits used by this work request */
55 	uint16_t plen;		/* total payload (sum of all packets) */
56 };
57 
58 /* All information needed to tx a frame */
59 struct txinfo {
60 	uint32_t len;		/* Total length of frame */
61 	uint32_t flags;		/* Checksum and LSO flags */
62 	uint32_t mss;		/* MSS for LSO */
63 	uint8_t nsegs;		/* # of segments in the SGL, 0 means imm. tx */
64 	uint8_t nflits;		/* # of flits needed for the SGL */
65 	uint8_t hdls_used;	/* # of DMA handles used */
66 	uint32_t txb_used;	/* txb_space used */
67 	mac_ether_offload_info_t meoi;	/* pkt hdr info for offloads */
68 	struct ulptx_sgl sgl __attribute__((aligned(8)));
69 	struct ulptx_sge_pair reserved[TX_SGL_SEGS / 2];
70 };
71 
72 struct mblk_pair {
73 	mblk_t *head, *tail;
74 };
75 
76 struct rxbuf {
77 	kmem_cache_t *cache;		/* the kmem_cache this rxb came from */
78 	ddi_dma_handle_t dhdl;
79 	ddi_acc_handle_t ahdl;
80 	caddr_t va;			/* KVA of buffer */
81 	uint64_t ba;			/* bus address of buffer */
82 	frtn_t freefunc;
83 	uint_t buf_size;
84 	volatile uint_t ref_cnt;
85 };
86 
87 static int service_iq(struct sge_iq *iq, int budget);
88 static inline void init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx,
89     int8_t pktc_idx, int qsize, uint8_t esize);
90 static inline void init_fl(struct sge_fl *fl, uint16_t qsize);
91 static int alloc_iq_fl(struct port_info *pi, struct sge_iq *iq,
92     struct sge_fl *fl, int intr_idx, int cong);
93 static int free_iq_fl(struct port_info *pi, struct sge_iq *iq,
94     struct sge_fl *fl);
95 static int alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx,
96     int i);
97 static int free_rxq(struct port_info *pi, struct sge_rxq *rxq);
98 static int eth_eq_alloc(struct adapter *sc, struct port_info *pi,
99     struct sge_eq *eq);
100 static int alloc_eq(struct adapter *sc, struct port_info *pi,
101     struct sge_eq *eq);
102 static int free_eq(struct adapter *sc, struct sge_eq *eq);
103 static int alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx);
104 static int free_txq(struct port_info *pi, struct sge_txq *txq);
105 static int alloc_dma_memory(struct adapter *sc, size_t len, int flags,
106     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
107     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
108     caddr_t *pva);
109 static int free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
110 static int alloc_desc_ring(struct adapter *sc, size_t len, int rw,
111     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
112     caddr_t *pva);
113 static int free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
114 static int alloc_tx_copybuffer(struct adapter *sc, size_t len,
115     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
116     caddr_t *pva);
117 static inline bool is_new_response(const struct sge_iq *iq,
118     struct rsp_ctrl **ctrl);
119 static inline void iq_next(struct sge_iq *iq);
120 static int refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs);
121 static void refill_sfl(void *arg);
122 static void add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl);
123 static void free_fl_bufs(struct sge_fl *fl);
124 static mblk_t *get_fl_payload(struct adapter *sc, struct sge_fl *fl,
125     uint32_t len_newbuf, int *fl_bufs_used);
126 static int get_frame_txinfo(struct sge_txq *txq, mblk_t **fp,
127     struct txinfo *txinfo, int sgl_only);
128 static inline int fits_in_txb(struct sge_txq *txq, int len, int *waste);
129 static inline int copy_into_txb(struct sge_txq *txq, mblk_t *m, int len,
130     struct txinfo *txinfo);
131 static inline void add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len);
132 static inline int add_mblk(struct sge_txq *txq, struct txinfo *txinfo,
133     mblk_t *m, int len);
134 static void free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo);
135 static int add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
136     struct txinfo *txinfo);
137 static void write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts);
138 static int write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
139     struct txinfo *txinfo);
140 static void t4_write_flush_wr(struct sge_txq *);
141 static inline void write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
142     struct txpkts *txpkts, struct txinfo *txinfo);
143 static inline void copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to,
144     int len);
145 static void t4_tx_ring_db(struct sge_txq *);
146 static uint_t t4_tx_reclaim_descs(struct sge_txq *, uint_t, mblk_t **);
147 static int t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss,
148     mblk_t *m);
149 static inline void ring_fl_db(struct adapter *sc, struct sge_fl *fl);
150 static kstat_t *setup_port_config_kstats(struct port_info *pi);
151 static kstat_t *setup_port_info_kstats(struct port_info *pi);
152 static kstat_t *setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq,
153     int idx);
154 static int update_rxq_kstats(kstat_t *ksp, int rw);
155 static int update_port_info_kstats(kstat_t *ksp, int rw);
156 static kstat_t *setup_txq_kstats(struct port_info *pi, struct sge_txq *txq,
157     int idx);
158 static int update_txq_kstats(kstat_t *ksp, int rw);
159 static void t4_sge_egr_update(struct sge_iq *, const struct rss_header *);
160 static int t4_handle_cpl_msg(struct sge_iq *, const struct rss_header *,
161     mblk_t *);
162 static int t4_handle_fw_msg(struct sge_iq *, const struct rss_header *);
163 
164 static kmem_cache_t *rxbuf_cache_create(struct rxbuf_cache_params *);
165 static struct rxbuf *rxbuf_alloc(kmem_cache_t *, int, uint_t);
166 static void rxbuf_free(struct rxbuf *);
167 static int rxbuf_ctor(void *, void *, int);
168 static void rxbuf_dtor(void *, void *);
169 
170 static inline void *
t4_rss_payload(const struct rss_header * rss)171 t4_rss_payload(const struct rss_header *rss)
172 {
173 	return ((void *)(&rss[1]));
174 }
175 
176 static inline struct sge_iq **
t4_iqmap_slot(struct adapter * sc,uint_t cntxt_id)177 t4_iqmap_slot(struct adapter *sc, uint_t cntxt_id)
178 {
179 	const uint_t idx = cntxt_id - sc->sge.iq_start;
180 	VERIFY3U(idx, <, sc->sge.iqmap_sz);
181 	return (&sc->sge.iqmap[idx]);
182 }
183 
184 static inline struct sge_eq **
t4_eqmap_slot(struct adapter * sc,uint_t cntxt_id)185 t4_eqmap_slot(struct adapter *sc, uint_t cntxt_id)
186 {
187 	const uint_t idx = cntxt_id - sc->sge.eq_start;
188 	VERIFY3U(idx, <, sc->sge.eqmap_sz);
189 	return (&sc->sge.eqmap[idx]);
190 }
191 
192 static inline int
reclaimable(struct sge_eq * eq)193 reclaimable(struct sge_eq *eq)
194 {
195 	unsigned int cidx;
196 
197 	cidx = eq->spg->cidx;   /* stable snapshot */
198 	cidx = be16_to_cpu(cidx);
199 
200 	if (cidx >= eq->cidx)
201 		return (cidx - eq->cidx);
202 	else
203 		return (cidx + eq->cap - eq->cidx);
204 }
205 
206 void
t4_sge_init(struct adapter * sc)207 t4_sge_init(struct adapter *sc)
208 {
209 	struct driver_properties *p = &sc->props;
210 	ddi_dma_attr_t *dma_attr;
211 	ddi_device_acc_attr_t *acc_attr;
212 	uint32_t sge_control, sge_conm_ctrl;
213 	int egress_threshold;
214 
215 	/*
216 	 * Device access and DMA attributes for descriptor rings
217 	 */
218 	acc_attr = &sc->sge.acc_attr_desc;
219 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
220 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
221 	acc_attr->devacc_attr_dataorder = DDI_STRICTORDER_ACC;
222 
223 	dma_attr = &sc->sge.dma_attr_desc;
224 	dma_attr->dma_attr_version = DMA_ATTR_V0;
225 	dma_attr->dma_attr_addr_lo = 0;
226 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
227 	dma_attr->dma_attr_count_max = UINT64_MAX;
228 	dma_attr->dma_attr_align = 512;
229 	dma_attr->dma_attr_burstsizes = 0xfff;
230 	dma_attr->dma_attr_minxfer = 1;
231 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
232 	dma_attr->dma_attr_seg = UINT64_MAX;
233 	dma_attr->dma_attr_sgllen = 1;
234 	dma_attr->dma_attr_granular = 1;
235 	dma_attr->dma_attr_flags = 0;
236 
237 	/*
238 	 * Device access and DMA attributes for tx buffers
239 	 */
240 	acc_attr = &sc->sge.acc_attr_tx;
241 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
242 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
243 
244 	dma_attr = &sc->sge.dma_attr_tx;
245 	dma_attr->dma_attr_version = DMA_ATTR_V0;
246 	dma_attr->dma_attr_addr_lo = 0;
247 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
248 	dma_attr->dma_attr_count_max = UINT64_MAX;
249 	dma_attr->dma_attr_align = 1;
250 	dma_attr->dma_attr_burstsizes = 0xfff;
251 	dma_attr->dma_attr_minxfer = 1;
252 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
253 	dma_attr->dma_attr_seg = UINT64_MAX;
254 	dma_attr->dma_attr_sgllen = TX_SGL_SEGS;
255 	dma_attr->dma_attr_granular = 1;
256 	dma_attr->dma_attr_flags = 0;
257 
258 	/*
259 	 * Ingress Padding Boundary and Egress Status Page Size are set up by
260 	 * t4_fixup_host_params().
261 	 */
262 	sge_control = t4_read_reg(sc, A_SGE_CONTROL);
263 	sc->sge.pktshift = G_PKTSHIFT(sge_control);
264 	sc->sge.stat_len = (sge_control & F_EGRSTATUSPAGESIZE) ? 128 : 64;
265 
266 	/* t4_nex uses FLM packed mode */
267 	sc->sge.fl_align = t4_fl_pkt_align(sc, true);
268 
269 	/*
270 	 * Device access and DMA attributes for rx buffers
271 	 */
272 	sc->sge.rxb_params.dip = sc->dip;
273 	sc->sge.rxb_params.buf_size = rx_buf_size;
274 
275 	acc_attr = &sc->sge.rxb_params.acc_attr_rx;
276 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
277 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
278 
279 	dma_attr = &sc->sge.rxb_params.dma_attr_rx;
280 	dma_attr->dma_attr_version = DMA_ATTR_V0;
281 	dma_attr->dma_attr_addr_lo = 0;
282 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
283 	dma_attr->dma_attr_count_max = UINT64_MAX;
284 	/*
285 	 * Low 4 bits of an rx buffer address have a special meaning to the SGE
286 	 * and an rx buf cannot have an address with any of these bits set.
287 	 * FL_ALIGN is >= 32 so we're sure things are ok.
288 	 */
289 	dma_attr->dma_attr_align = sc->sge.fl_align;
290 	dma_attr->dma_attr_burstsizes = 0xfff;
291 	dma_attr->dma_attr_minxfer = 1;
292 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
293 	dma_attr->dma_attr_seg = UINT64_MAX;
294 	dma_attr->dma_attr_sgllen = 1;
295 	dma_attr->dma_attr_granular = 1;
296 	dma_attr->dma_attr_flags = 0;
297 
298 	sc->sge.rxbuf_cache = rxbuf_cache_create(&sc->sge.rxb_params);
299 
300 	/*
301 	 * A FL with <= fl_starve_thres buffers is starving and a periodic
302 	 * timer will attempt to refill it.  This needs to be larger than the
303 	 * SGE's Egress Congestion Threshold.  If it isn't, then we can get
304 	 * stuck waiting for new packets while the SGE is waiting for us to
305 	 * give it more Free List entries.  (Note that the SGE's Egress
306 	 * Congestion Threshold is in units of 2 Free List pointers.) For T4,
307 	 * there was only a single field to control this.  For T5 there's the
308 	 * original field which now only applies to Unpacked Mode Free List
309 	 * buffers and a new field which only applies to Packed Mode Free List
310 	 * buffers.
311 	 */
312 
313 	sge_conm_ctrl = t4_read_reg(sc, A_SGE_CONM_CTRL);
314 	switch (CHELSIO_CHIP_VERSION(sc->params.chip)) {
315 	case CHELSIO_T4:
316 		egress_threshold = G_EGRTHRESHOLD(sge_conm_ctrl);
317 		break;
318 	case CHELSIO_T5:
319 		egress_threshold = G_EGRTHRESHOLDPACKING(sge_conm_ctrl);
320 		break;
321 	case CHELSIO_T6:
322 	default:
323 		egress_threshold = G_T6_EGRTHRESHOLDPACKING(sge_conm_ctrl);
324 	}
325 	sc->sge.fl_starve_threshold = 2*egress_threshold + 1;
326 
327 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, rx_buf_size);
328 
329 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD,
330 	    V_THRESHOLD_0(p->holdoff_pktcnt[0]) |
331 	    V_THRESHOLD_1(p->holdoff_pktcnt[1]) |
332 	    V_THRESHOLD_2(p->holdoff_pktcnt[2]) |
333 	    V_THRESHOLD_3(p->holdoff_pktcnt[3]));
334 
335 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1,
336 	    V_TIMERVALUE0(us_to_core_ticks(sc, p->holdoff_timer_us[0])) |
337 	    V_TIMERVALUE1(us_to_core_ticks(sc, p->holdoff_timer_us[1])));
338 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3,
339 	    V_TIMERVALUE2(us_to_core_ticks(sc, p->holdoff_timer_us[2])) |
340 	    V_TIMERVALUE3(us_to_core_ticks(sc, p->holdoff_timer_us[3])));
341 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5,
342 	    V_TIMERVALUE4(us_to_core_ticks(sc, p->holdoff_timer_us[4])) |
343 	    V_TIMERVALUE5(us_to_core_ticks(sc, p->holdoff_timer_us[5])));
344 }
345 
346 static inline int
first_vector(struct port_info * pi)347 first_vector(struct port_info *pi)
348 {
349 	struct adapter *sc = pi->adapter;
350 	int rc = T4_EXTRA_INTR, i;
351 
352 	if (sc->intr_count == 1)
353 		return (0);
354 
355 	for_each_port(sc, i) {
356 		struct port_info *p = sc->port[i];
357 
358 		if (i == pi->port_id)
359 			break;
360 
361 		/*
362 		 * Not compiled with offload support and intr_count > 1.  Only
363 		 * NIC queues exist and they'd better be taking direct
364 		 * interrupts.
365 		 */
366 		ASSERT(!(sc->flags & TAF_INTR_FWD));
367 		rc += p->nrxq;
368 	}
369 	return (rc);
370 }
371 
372 /*
373  * Given an arbitrary "index," come up with an iq that can be used by other
374  * queues (of this port) for interrupt forwarding, SGE egress updates, etc.
375  * The iq returned is guaranteed to be something that takes direct interrupts.
376  */
377 static struct sge_iq *
port_intr_iq(struct port_info * pi,int idx)378 port_intr_iq(struct port_info *pi, int idx)
379 {
380 	struct adapter *sc = pi->adapter;
381 	struct sge *s = &sc->sge;
382 	struct sge_iq *iq = NULL;
383 
384 	if (sc->intr_count == 1)
385 		return (&sc->sge.fwq);
386 
387 	/*
388 	 * Not compiled with offload support and intr_count > 1.  Only NIC
389 	 * queues exist and they'd better be taking direct interrupts.
390 	 */
391 	ASSERT(!(sc->flags & TAF_INTR_FWD));
392 
393 	idx %= pi->nrxq;
394 	iq = &s->rxq[pi->first_rxq + idx].iq;
395 
396 	return (iq);
397 }
398 
399 int
t4_setup_port_queues(struct port_info * pi)400 t4_setup_port_queues(struct port_info *pi)
401 {
402 	int rc = 0, i, intr_idx, j;
403 	struct sge_rxq *rxq;
404 	struct sge_txq *txq;
405 	struct adapter *sc = pi->adapter;
406 	struct driver_properties *p = &sc->props;
407 
408 	pi->ksp_config = setup_port_config_kstats(pi);
409 	pi->ksp_info   = setup_port_info_kstats(pi);
410 
411 	/* Interrupt vector to start from (when using multiple vectors) */
412 	intr_idx = first_vector(pi);
413 
414 	/*
415 	 * First pass over all rx queues (NIC and TOE):
416 	 * a) initialize iq and fl
417 	 * b) allocate queue iff it will take direct interrupts.
418 	 */
419 
420 	for_each_rxq(pi, i, rxq) {
421 
422 		init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, p->qsize_rxq,
423 		    RX_IQ_ESIZE);
424 
425 		init_fl(&rxq->fl, p->qsize_rxq / 8); /* 8 bufs in each entry */
426 
427 		if ((!(sc->flags & TAF_INTR_FWD)) ||
428 		    (sc->intr_count > 1 && pi->nrxq)) {
429 			rxq->iq.flags |= IQ_INTR;
430 			rc = alloc_rxq(pi, rxq, intr_idx, i);
431 			if (rc != 0)
432 				goto done;
433 			intr_idx++;
434 		}
435 
436 	}
437 
438 	/*
439 	 * Second pass over all rx queues (NIC and TOE).  The queues forwarding
440 	 * their interrupts are allocated now.
441 	 */
442 	j = 0;
443 	for_each_rxq(pi, i, rxq) {
444 		if (rxq->iq.flags & IQ_INTR)
445 			continue;
446 
447 		intr_idx = port_intr_iq(pi, j)->abs_id;
448 
449 		rc = alloc_rxq(pi, rxq, intr_idx, i);
450 		if (rc != 0)
451 			goto done;
452 		j++;
453 	}
454 
455 	/*
456 	 * Now the tx queues.  Only one pass needed.
457 	 */
458 	j = 0;
459 	for_each_txq(pi, i, txq) {
460 		txq->eq.flags = 0;
461 		txq->eq.tx_chan = pi->tx_chan;
462 		txq->eq.qsize = p->qsize_txq;
463 
464 		/* For now, direct all TX queue notifications to the FW IQ. */
465 		txq->eq.iqid = sc->sge.fwq.cntxt_id;
466 
467 		rc = alloc_txq(pi, txq, i);
468 		if (rc != 0)
469 			goto done;
470 	}
471 
472 done:
473 	if (rc != 0)
474 		(void) t4_teardown_port_queues(pi);
475 
476 	return (rc);
477 }
478 
479 /*
480  * Idempotent
481  */
482 int
t4_teardown_port_queues(struct port_info * pi)483 t4_teardown_port_queues(struct port_info *pi)
484 {
485 	int i;
486 	struct sge_rxq *rxq;
487 	struct sge_txq *txq;
488 
489 	if (pi->ksp_config != NULL) {
490 		kstat_delete(pi->ksp_config);
491 		pi->ksp_config = NULL;
492 	}
493 	if (pi->ksp_info != NULL) {
494 		kstat_delete(pi->ksp_info);
495 		pi->ksp_info = NULL;
496 	}
497 
498 	for_each_txq(pi, i, txq) {
499 		(void) free_txq(pi, txq);
500 	}
501 
502 	for_each_rxq(pi, i, rxq) {
503 		if ((rxq->iq.flags & IQ_INTR) == 0)
504 			(void) free_rxq(pi, rxq);
505 	}
506 
507 	/*
508 	 * Then take down the rx queues that take direct interrupts.
509 	 */
510 
511 	for_each_rxq(pi, i, rxq) {
512 		if (rxq->iq.flags & IQ_INTR)
513 			(void) free_rxq(pi, rxq);
514 	}
515 
516 	return (0);
517 }
518 
519 /* Deals with errors and forwarded interrupts */
520 uint_t
t4_intr_all(caddr_t arg1,caddr_t arg2)521 t4_intr_all(caddr_t arg1, caddr_t arg2)
522 {
523 
524 	(void) t4_intr_err(arg1, arg2);
525 	(void) t4_intr(arg1, arg2);
526 
527 	return (DDI_INTR_CLAIMED);
528 }
529 
530 /*
531  * We are counting on the values of t4_intr_config_t matching the register
532  * definitions from the shared code.
533  */
534 CTASSERT(TIC_SE_INTR_ARM == F_QINTR_CNT_EN);
535 CTASSERT(TIC_TIMER0 == V_QINTR_TIMER_IDX(X_TIMERREG_COUNTER0));
536 CTASSERT(TIC_TIMER5 == V_QINTR_TIMER_IDX(X_TIMERREG_COUNTER5));
537 CTASSERT(TIC_START_COUNTER == V_QINTR_TIMER_IDX(X_TIMERREG_RESTART_COUNTER));
538 
539 void
t4_iq_update_intr_cfg(struct sge_iq * iq,uint8_t tmr_idx,int8_t pktc_idx)540 t4_iq_update_intr_cfg(struct sge_iq *iq, uint8_t tmr_idx, int8_t pktc_idx)
541 {
542 	ASSERT((pktc_idx >= 0 && pktc_idx < SGE_NCOUNTERS) || pktc_idx == -1);
543 	IQ_LOCK_ASSERT_OWNED(iq);
544 	/*
545 	 * Strictly speaking, the IQ could be programmed with a TimerReg value
546 	 * of 6 (TICK_START_COUNTER), which is outside the range of SGE_NTIMERS.
547 	 *
548 	 * Since we do not currently offer an interface to configure such
549 	 * behavior, we assert its absence here for now.
550 	 */
551 	ASSERT3U(tmr_idx, <, SGE_NTIMERS);
552 
553 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx) |
554 	    ((pktc_idx != -1) ? TIC_SE_INTR_ARM : 0);
555 
556 	/* Update IQ for new packet count threshold, but only if enabled */
557 	if (pktc_idx != iq->intr_pktc_idx && pktc_idx >= 0) {
558 		const uint32_t param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
559 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_IQ_INTCNTTHRESH) |
560 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
561 		const uint32_t val = pktc_idx;
562 
563 		struct adapter *sc = iq->adapter;
564 		int rc =
565 		    -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
566 		if (rc != 0) {
567 			/* report error but carry on */
568 			cxgb_printf(sc->dip, CE_WARN,
569 			    "failed to set intr pktcnt index for IQ %d: %d",
570 			    iq->cntxt_id, rc);
571 		}
572 	}
573 	iq->intr_pktc_idx = pktc_idx;
574 }
575 
576 void
t4_eq_update_dbq_timer(struct sge_eq * eq,struct port_info * pi)577 t4_eq_update_dbq_timer(struct sge_eq *eq, struct port_info *pi)
578 {
579 	struct adapter *sc = pi->adapter;
580 
581 	const uint32_t param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
582 	    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_TIMERIX) |
583 	    V_FW_PARAMS_PARAM_YZ(eq->cntxt_id);
584 	const uint32_t val = pi->dbq_timer_idx;
585 
586 	int rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
587 	if (rc != 0) {
588 		/* report error but carry on */
589 		cxgb_printf(sc->dip, CE_WARN,
590 		    "failed to set DBQ timer index for EQ %d: %d",
591 		    eq->cntxt_id, rc);
592 	}
593 }
594 
595 /*
596  * Update (via GTS) the interrupt/timer config and CIDX value for a specified
597  * ingress queue.
598  */
599 void
t4_iq_gts_update(struct sge_iq * iq,t4_intr_config_t cfg,uint16_t cidx_incr)600 t4_iq_gts_update(struct sge_iq *iq, t4_intr_config_t cfg, uint16_t cidx_incr)
601 {
602 	const uint32_t value =
603 	    V_INGRESSQID((uint32_t)iq->cntxt_id) |
604 	    V_CIDXINC((uint32_t)cidx_incr) |
605 	    V_SEINTARM((uint32_t)cfg);
606 	t4_write_reg(iq->adapter, MYPF_REG(A_SGE_PF_GTS), value);
607 }
608 
609 /*
610  * Update (via GTS) the CIDX value for a specified ingress queue.
611  *
612  * This _only_ increments CIDX and does not alter any other timer related state
613  * associated with the IQ.
614  */
615 static void
t4_iq_gts_incr(struct sge_iq * iq,uint16_t cidx_incr)616 t4_iq_gts_incr(struct sge_iq *iq, uint16_t cidx_incr)
617 {
618 	if (cidx_incr == 0) {
619 		return;
620 	}
621 
622 	const uint32_t value =
623 	    V_INGRESSQID((uint32_t)iq->cntxt_id) |
624 	    V_CIDXINC((uint32_t)cidx_incr) |
625 	    V_SEINTARM((uint32_t)V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX));
626 	t4_write_reg(iq->adapter, MYPF_REG(A_SGE_PF_GTS), value);
627 }
628 
629 static void
t4_intr_rx_work(struct sge_iq * iq)630 t4_intr_rx_work(struct sge_iq *iq)
631 {
632 	mblk_t *mp = NULL;
633 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
634 	RXQ_LOCK(rxq);
635 	if (!iq->polling) {
636 		mp = t4_ring_rx(rxq, iq->qsize/8);
637 		t4_iq_gts_update(iq, iq->intr_params, 0);
638 	}
639 	RXQ_UNLOCK(rxq);
640 	if (mp != NULL) {
641 		mac_rx_ring(rxq->port->mh, rxq->ring_handle, mp,
642 		    rxq->ring_gen_num);
643 	}
644 }
645 
646 /* Deals with interrupts on the given ingress queue */
647 /* ARGSUSED */
648 uint_t
t4_intr(caddr_t arg1,caddr_t arg2)649 t4_intr(caddr_t arg1, caddr_t arg2)
650 {
651 	struct sge_iq *iq = (struct sge_iq *)arg2;
652 	int state;
653 
654 	/*
655 	 * Right now receive polling is only enabled for MSI-X and
656 	 * when we have enough msi-x vectors i.e no interrupt forwarding.
657 	 */
658 	if (iq->adapter->props.multi_rings) {
659 		t4_intr_rx_work(iq);
660 	} else {
661 		state = atomic_cas_uint(&iq->state, IQS_IDLE, IQS_BUSY);
662 		if (state == IQS_IDLE) {
663 			(void) service_iq(iq, 0);
664 			(void) atomic_cas_uint(&iq->state, IQS_BUSY, IQS_IDLE);
665 		}
666 	}
667 	return (DDI_INTR_CLAIMED);
668 }
669 
670 /* Deals with error interrupts */
671 /* ARGSUSED */
672 uint_t
t4_intr_err(caddr_t arg1,caddr_t arg2)673 t4_intr_err(caddr_t arg1, caddr_t arg2)
674 {
675 	struct adapter *sc = (struct adapter *)arg1;
676 
677 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
678 	(void) t4_slow_intr_handler(sc);
679 
680 	return (DDI_INTR_CLAIMED);
681 }
682 
683 /*
684  * t4_ring_rx - Process responses from an SGE response queue.
685  *
686  * This function processes responses from an SGE response queue up to the
687  * supplied budget.  Responses include received packets as well as control
688  * messages from FW or HW.
689  *
690  * It returns a chain of mblks containing the received data, to be
691  * passed up to mac_rx_ring().
692  */
693 mblk_t *
t4_ring_rx(struct sge_rxq * rxq,int budget)694 t4_ring_rx(struct sge_rxq *rxq, int budget)
695 {
696 	struct sge_iq *iq = &rxq->iq;
697 	struct sge_fl *fl = &rxq->fl;		/* Use iff IQ_HAS_FL */
698 	struct adapter *sc = iq->adapter;
699 	struct rsp_ctrl *ctrl;
700 	int ndescs = 0, fl_bufs_used = 0;
701 	mblk_t *mblk_head = NULL, **mblk_tail = &mblk_head;
702 	uint32_t received_bytes = 0, pkt_len = 0;
703 	uint16_t err_vec;
704 
705 	while (is_new_response(iq, &ctrl)) {
706 		membar_consumer();
707 
708 		const uint8_t type_gen = ctrl->u.type_gen;
709 		const uint8_t rsp_type = G_RSPD_TYPE(type_gen);
710 		const bool overflowed = (type_gen & F_RSPD_QOVFL) != 0;
711 		const uint32_t data_len = BE_32(ctrl->pldbuflen_qid);
712 
713 		iq->stats.sis_processed++;
714 		if (overflowed) {
715 			iq->stats.sis_overflow++;
716 		}
717 
718 		const struct rss_header *rss =
719 		    (const struct rss_header *)iq->cdesc;
720 		mblk_t *m = NULL;
721 
722 		switch (rsp_type) {
723 		case X_RSPD_TYPE_FLBUF:
724 
725 			ASSERT(iq->flags & IQ_HAS_FL);
726 
727 			if (CPL_RX_PKT == rss->opcode) {
728 				const struct cpl_rx_pkt *cpl =
729 				    t4_rss_payload(rss);
730 				pkt_len = be16_to_cpu(cpl->len);
731 
732 				if (iq->polling &&
733 				    ((received_bytes + pkt_len) > budget))
734 					goto done;
735 
736 				m = get_fl_payload(sc, fl, data_len,
737 				    &fl_bufs_used);
738 				if (m == NULL)
739 					goto done;
740 
741 				m->b_rptr += sc->sge.pktshift;
742 				if (sc->params.tp.rx_pkt_encap) {
743 					/* Enabled only in T6 config file */
744 					err_vec = G_T6_COMPR_RXERR_VEC(
745 					    ntohs(cpl->err_vec));
746 				} else {
747 					err_vec = ntohs(cpl->err_vec);
748 				}
749 
750 				const bool csum_ok = cpl->csum_calc && !err_vec;
751 
752 				/* TODO: what about cpl->ip_frag? */
753 				if (csum_ok && !cpl->ip_frag) {
754 					mac_hcksum_set(m, 0, 0, 0, 0xffff,
755 					    HCK_FULLCKSUM_OK | HCK_FULLCKSUM |
756 					    HCK_IPV4_HDRCKSUM_OK);
757 					rxq->rxcsum++;
758 				}
759 				rxq->rxpkts++;
760 				rxq->rxbytes += pkt_len;
761 				received_bytes += pkt_len;
762 
763 				*mblk_tail = m;
764 				mblk_tail = &m->b_next;
765 
766 				break;
767 			}
768 
769 			m = get_fl_payload(sc, fl, data_len, &fl_bufs_used);
770 			if (m == NULL)
771 				goto done;
772 			/* FALLTHROUGH */
773 
774 		case X_RSPD_TYPE_CPL:
775 			(void) t4_handle_cpl_msg(iq, rss, m);
776 			break;
777 
778 		default:
779 			break;
780 		}
781 		iq_next(iq);
782 		++ndescs;
783 		if (!iq->polling && (ndescs == budget))
784 			break;
785 	}
786 
787 done:
788 
789 	t4_iq_gts_incr(iq, ndescs);
790 
791 	if ((fl_bufs_used > 0) || (iq->flags & IQ_HAS_FL)) {
792 		int starved;
793 		FL_LOCK(fl);
794 		fl->needed += fl_bufs_used;
795 		starved = refill_fl(sc, fl, fl->cap / 8);
796 		FL_UNLOCK(fl);
797 		if (starved)
798 			add_fl_to_sfl(sc, fl);
799 	}
800 	return (mblk_head);
801 }
802 
803 /*
804  * Deals with anything and everything on the given ingress queue.
805  */
806 static int
service_iq(struct sge_iq * iq,int budget)807 service_iq(struct sge_iq *iq, int budget)
808 {
809 	struct sge_iq *q;
810 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
811 	struct sge_fl *fl = &rxq->fl;		/* Use iff IQ_HAS_FL */
812 	struct adapter *sc = iq->adapter;
813 	struct rsp_ctrl *ctrl;
814 	int ndescs = 0, fl_bufs_used = 0;
815 	int starved;
816 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
817 
818 	const uint_t limit = (budget != 0) ? budget : iq->qsize / 8;
819 
820 	/*
821 	 * We always come back and check the descriptor ring for new indirect
822 	 * interrupts and other responses after running a single handler.
823 	 */
824 	for (;;) {
825 		while (is_new_response(iq, &ctrl)) {
826 			membar_consumer();
827 
828 			const uint8_t type_gen = ctrl->u.type_gen;
829 			const uint8_t rsp_type = G_RSPD_TYPE(type_gen);
830 			const uint32_t dlen_qid = BE_32(ctrl->pldbuflen_qid);
831 
832 			mblk_t *m = NULL;
833 			const struct rss_header *rss =
834 			    (const struct rss_header *)iq->cdesc;
835 
836 			switch (rsp_type) {
837 			case X_RSPD_TYPE_FLBUF:
838 
839 				ASSERT(iq->flags & IQ_HAS_FL);
840 
841 				m = get_fl_payload(sc, fl, dlen_qid,
842 				    &fl_bufs_used);
843 				if (m == NULL) {
844 					/*
845 					 * Rearm the iq with a
846 					 * longer-than-default timer
847 					 */
848 					t4_iq_gts_update(iq, TIC_TIMER5,
849 					    ndescs);
850 					if (fl_bufs_used > 0) {
851 						ASSERT(iq->flags & IQ_HAS_FL);
852 						FL_LOCK(fl);
853 						fl->needed += fl_bufs_used;
854 						starved = refill_fl(sc, fl,
855 						    fl->cap / 8);
856 						FL_UNLOCK(fl);
857 						if (starved)
858 							add_fl_to_sfl(sc, fl);
859 					}
860 					return (0);
861 				}
862 
863 			/* FALLTHRU */
864 			case X_RSPD_TYPE_CPL:
865 				(void) t4_handle_cpl_msg(iq, rss, m);
866 				break;
867 
868 			case X_RSPD_TYPE_INTR:
869 
870 				/*
871 				 * Interrupts should be forwarded only to queues
872 				 * that are not forwarding their interrupts.
873 				 * This means service_iq can recurse but only 1
874 				 * level deep.
875 				 */
876 				ASSERT(budget == 0);
877 
878 				q = *t4_iqmap_slot(sc, dlen_qid);
879 				if (atomic_cas_uint(&q->state, IQS_IDLE,
880 				    IQS_BUSY) == IQS_IDLE) {
881 					if (service_iq(q, q->qsize / 8) == 0) {
882 						(void) atomic_cas_uint(
883 						    &q->state, IQS_BUSY,
884 						    IQS_IDLE);
885 					} else {
886 						STAILQ_INSERT_TAIL(&iql, q,
887 						    link);
888 					}
889 				}
890 				break;
891 
892 			default:
893 				break;
894 			}
895 
896 			iq_next(iq);
897 			if (++ndescs == limit) {
898 				t4_iq_gts_incr(iq, ndescs);
899 				ndescs = 0;
900 
901 				if (fl_bufs_used > 0) {
902 					ASSERT(iq->flags & IQ_HAS_FL);
903 					FL_LOCK(fl);
904 					fl->needed += fl_bufs_used;
905 					(void) refill_fl(sc, fl, fl->cap / 8);
906 					FL_UNLOCK(fl);
907 					fl_bufs_used = 0;
908 				}
909 
910 				if (budget != 0)
911 					return (EINPROGRESS);
912 			}
913 		}
914 
915 		if (STAILQ_EMPTY(&iql) != 0)
916 			break;
917 
918 		/*
919 		 * Process the head only, and send it to the back of the list if
920 		 * it's still not done.
921 		 */
922 		q = STAILQ_FIRST(&iql);
923 		STAILQ_REMOVE_HEAD(&iql, link);
924 		if (service_iq(q, q->qsize / 8) == 0)
925 			(void) atomic_cas_uint(&q->state, IQS_BUSY, IQS_IDLE);
926 		else
927 			STAILQ_INSERT_TAIL(&iql, q, link);
928 	}
929 
930 	t4_iq_gts_update(iq, iq->intr_params, ndescs);
931 
932 	if (iq->flags & IQ_HAS_FL) {
933 		FL_LOCK(fl);
934 		fl->needed += fl_bufs_used;
935 		starved = refill_fl(sc, fl, fl->cap / 4);
936 		FL_UNLOCK(fl);
937 		if (starved != 0)
938 			add_fl_to_sfl(sc, fl);
939 	}
940 
941 	return (0);
942 }
943 
944 /* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */
945 #define	TXPKTS_PKT_HDR ((\
946 	sizeof (struct ulp_txpkt) + \
947 	sizeof (struct ulptx_idata) + \
948 	sizeof (struct cpl_tx_pkt_core)) / 8)
949 
950 /* Header of a coalesced tx WR, before SGL of first packet (in flits) */
951 #define	TXPKTS_WR_HDR (\
952 	sizeof (struct fw_eth_tx_pkts_wr) / 8 + \
953 	TXPKTS_PKT_HDR)
954 
955 /* Header of a tx WR, before SGL of first packet (in flits) */
956 #define	TXPKT_WR_HDR ((\
957 	sizeof (struct fw_eth_tx_pkt_wr) + \
958 	sizeof (struct cpl_tx_pkt_core)) / 8)
959 
960 /* Header of a tx LSO WR, before SGL of first packet (in flits) */
961 #define	TXPKT_LSO_WR_HDR ((\
962 	sizeof (struct fw_eth_tx_pkt_wr) + \
963 	sizeof (struct cpl_tx_pkt_lso_core) + \
964 	sizeof (struct cpl_tx_pkt_core)) / 8)
965 
966 mblk_t *
t4_eth_tx(void * arg,mblk_t * frame)967 t4_eth_tx(void *arg, mblk_t *frame)
968 {
969 	struct sge_txq *txq = (struct sge_txq *)arg;
970 	struct port_info *pi = txq->port;
971 	struct sge_eq *eq = &txq->eq;
972 	mblk_t *next_frame;
973 	int rc, coalescing;
974 	struct txpkts txpkts;
975 	struct txinfo txinfo;
976 
977 	txpkts.npkt = 0; /* indicates there's nothing in txpkts */
978 	coalescing = 0;
979 
980 	TXQ_LOCK(txq);
981 	if (eq->avail < 8)
982 		(void) t4_tx_reclaim_descs(txq, 8, NULL);
983 	for (; frame; frame = next_frame) {
984 
985 		if (eq->avail < 8)
986 			break;
987 
988 		next_frame = frame->b_next;
989 		frame->b_next = NULL;
990 
991 		if (next_frame != NULL)
992 			coalescing = 1;
993 
994 		rc = get_frame_txinfo(txq, &frame, &txinfo, coalescing);
995 		if (rc != 0) {
996 			if (rc == ENOMEM) {
997 				/* Short of resources, suspend tx */
998 				frame->b_next = next_frame;
999 
1000 				/*
1001 				 * Since we are out of memory for this packet,
1002 				 * rather than TX descriptors, enqueue an
1003 				 * flush work request.  This will ensure that a
1004 				 * completion notification is delivered for this
1005 				 * EQ which will trigger a call to update the
1006 				 * state in mac to continue transmissions.
1007 				 */
1008 				t4_write_flush_wr(txq);
1009 
1010 				break;
1011 			}
1012 
1013 			/*
1014 			 * Unrecoverable error for this frame, throw it away and
1015 			 * move on to the next.
1016 			 */
1017 			freemsg(frame);
1018 			continue;
1019 		}
1020 
1021 		if (coalescing != 0 &&
1022 		    add_to_txpkts(txq, &txpkts, frame, &txinfo) == 0) {
1023 
1024 			/* Successfully absorbed into txpkts */
1025 
1026 			write_ulp_cpl_sgl(pi, txq, &txpkts, &txinfo);
1027 			goto doorbell;
1028 		}
1029 
1030 		/*
1031 		 * We weren't coalescing to begin with, or current frame could
1032 		 * not be coalesced (add_to_txpkts flushes txpkts if a frame
1033 		 * given to it can't be coalesced).  Either way there should be
1034 		 * nothing in txpkts.
1035 		 */
1036 		ASSERT(txpkts.npkt == 0);
1037 
1038 		/* We're sending out individual frames now */
1039 		coalescing = 0;
1040 
1041 		if (eq->avail < 8)
1042 			(void) t4_tx_reclaim_descs(txq, 8, NULL);
1043 		rc = write_txpkt_wr(pi, txq, frame, &txinfo);
1044 		if (rc != 0) {
1045 
1046 			/* Short of hardware descriptors, suspend tx */
1047 
1048 			/*
1049 			 * This is an unlikely but expensive failure.  We've
1050 			 * done all the hard work (DMA bindings etc.) and now we
1051 			 * can't send out the frame.  What's worse, we have to
1052 			 * spend even more time freeing up everything in txinfo.
1053 			 */
1054 			txq->qfull++;
1055 			free_txinfo_resources(txq, &txinfo);
1056 
1057 			frame->b_next = next_frame;
1058 			break;
1059 		}
1060 
1061 doorbell:
1062 		/* Fewer and fewer doorbells as the queue fills up */
1063 		if (eq->pending >= (1 << (fls(eq->qsize - eq->avail) / 2))) {
1064 			txq->txbytes += txinfo.len;
1065 			txq->txpkts++;
1066 			t4_tx_ring_db(txq);
1067 		}
1068 		(void) t4_tx_reclaim_descs(txq, 32, NULL);
1069 	}
1070 
1071 	if (txpkts.npkt > 0) {
1072 		write_txpkts_wr(txq, &txpkts);
1073 	}
1074 
1075 	if (eq->pending != 0) {
1076 		t4_tx_ring_db(txq);
1077 	}
1078 
1079 	if (frame != NULL) {
1080 		eq->flags |= EQ_CORKED;
1081 	}
1082 
1083 	(void) t4_tx_reclaim_descs(txq, eq->qsize, NULL);
1084 	TXQ_UNLOCK(txq);
1085 
1086 	return (frame);
1087 }
1088 
1089 static inline void
init_iq(struct sge_iq * iq,struct adapter * sc,int tmr_idx,int8_t pktc_idx,int qsize,uint8_t esize)1090 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int8_t pktc_idx,
1091     int qsize, uint8_t esize)
1092 {
1093 	ASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS);
1094 	ASSERT(pktc_idx < SGE_NCOUNTERS);	/* -ve is ok, means don't use */
1095 
1096 	iq->flags = 0;
1097 	iq->adapter = sc;
1098 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
1099 	iq->intr_pktc_idx = -1;
1100 	if (pktc_idx >= 0) {
1101 		iq->intr_params |= TIC_SE_INTR_ARM;
1102 		iq->intr_pktc_idx = pktc_idx;
1103 	}
1104 	iq->qsize = roundup(qsize, 16);		/* See FW_IQ_CMD/iqsize */
1105 	iq->esize = max(esize, 16);		/* See FW_IQ_CMD/iqesize */
1106 }
1107 
1108 static inline void
init_fl(struct sge_fl * fl,uint16_t qsize)1109 init_fl(struct sge_fl *fl, uint16_t qsize)
1110 {
1111 
1112 	fl->qsize = qsize;
1113 	fl->allocb_fail = 0;
1114 }
1115 
1116 /*
1117  * Allocates the ring for an ingress queue and an optional freelist.  If the
1118  * freelist is specified it will be allocated and then associated with the
1119  * ingress queue.
1120  *
1121  * Returns errno on failure.  Resources allocated up to that point may still be
1122  * allocated.  Caller is responsible for cleanup in case this function fails.
1123  *
1124  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
1125  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
1126  * the index of the queue to which its interrupts will be forwarded.
1127  */
1128 static int
alloc_iq_fl(struct port_info * pi,struct sge_iq * iq,struct sge_fl * fl,int intr_idx,int cong)1129 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
1130     int intr_idx, int cong)
1131 {
1132 	int rc, i;
1133 	size_t len;
1134 	struct fw_iq_cmd c;
1135 	struct adapter *sc = iq->adapter;
1136 	uint32_t v = 0;
1137 
1138 	len = iq->qsize * iq->esize;
1139 	rc = alloc_desc_ring(sc, len, DDI_DMA_READ, &iq->dhdl, &iq->ahdl,
1140 	    &iq->ba, (caddr_t *)&iq->desc);
1141 	if (rc != 0)
1142 		return (rc);
1143 
1144 	bzero(&c, sizeof (c));
1145 	c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
1146 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
1147 	    V_FW_IQ_CMD_VFN(0));
1148 
1149 	c.alloc_to_len16 = cpu_to_be32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
1150 	    FW_LEN16(c));
1151 
1152 	/* Special handling for firmware event queue */
1153 	if (iq == &sc->sge.fwq)
1154 		v |= F_FW_IQ_CMD_IQASYNCH;
1155 
1156 	if (iq->flags & IQ_INTR)
1157 		ASSERT(intr_idx < sc->intr_count);
1158 	else
1159 		v |= F_FW_IQ_CMD_IQANDST;
1160 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
1161 
1162 	/*
1163 	 * If the coalescing counter is not enabled for this IQ, use the 0
1164 	 * index, rather than populating it with the invalid -1 value.
1165 	 *
1166 	 * The selected index does not matter when the counter is not enabled
1167 	 * through the GTS flags.
1168 	 */
1169 	const uint_t pktc_idx = (iq->intr_pktc_idx < 0) ? 0 : iq->intr_pktc_idx;
1170 
1171 	c.type_to_iqandstindex = cpu_to_be32(v |
1172 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
1173 	    V_FW_IQ_CMD_VIID(pi->viid) |
1174 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
1175 	c.iqdroprss_to_iqesize = cpu_to_be16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
1176 	    F_FW_IQ_CMD_IQGTSMODE |
1177 	    V_FW_IQ_CMD_IQINTCNTTHRESH(pktc_idx) |
1178 	    V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4));
1179 	c.iqsize = cpu_to_be16(iq->qsize);
1180 	c.iqaddr = cpu_to_be64(iq->ba);
1181 	if (cong >= 0) {
1182 		const uint32_t iq_type =
1183 		    cong ? FW_IQ_IQTYPE_NIC : FW_IQ_IQTYPE_OFLD;
1184 		c.iqns_to_fl0congen = BE_32(F_FW_IQ_CMD_IQFLINTCONGEN |
1185 		    V_FW_IQ_CMD_IQTYPE(iq_type));
1186 	}
1187 
1188 	if (fl != NULL) {
1189 		mutex_init(&fl->lock, NULL, MUTEX_DRIVER,
1190 		    DDI_INTR_PRI(sc->intr_pri));
1191 		fl->flags |= FL_MTX;
1192 
1193 		len = fl->qsize * RX_FL_ESIZE;
1194 		rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &fl->dhdl,
1195 		    &fl->ahdl, &fl->ba, (caddr_t *)&fl->desc);
1196 		if (rc != 0)
1197 			return (rc);
1198 
1199 		/* Allocate space for one software descriptor per buffer. */
1200 		fl->cap = (fl->qsize - sc->sge.stat_len / RX_FL_ESIZE) * 8;
1201 		fl->sdesc = kmem_zalloc(sizeof (struct fl_sdesc) * fl->cap,
1202 		    KM_SLEEP);
1203 		fl->needed = fl->cap;
1204 		fl->lowat = roundup(sc->sge.fl_starve_threshold, 8);
1205 
1206 		c.iqns_to_fl0congen |=
1207 		    cpu_to_be32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
1208 		    F_FW_IQ_CMD_FL0PACKEN | F_FW_IQ_CMD_FL0PADEN);
1209 		if (cong >= 0) {
1210 			c.iqns_to_fl0congen |=
1211 			    BE_32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
1212 			    F_FW_IQ_CMD_FL0CONGCIF |
1213 			    F_FW_IQ_CMD_FL0CONGEN);
1214 		}
1215 
1216 		/*
1217 		 * In T6, for egress queue type FL there is internal overhead
1218 		 * of 16B for header going into FLM module.  Hence the maximum
1219 		 * allowed burst size is 448 bytes.  For T4/T5, the hardware
1220 		 * doesn't coalesce fetch requests if more than 64 bytes of
1221 		 * Free List pointers are provided, so we use a 128-byte Fetch
1222 		 * Burst Minimum there (T6 implements coalescing so we can use
1223 		 * the smaller 64-byte value there).
1224 		 */
1225 		const uint_t fbmin = t4_cver_ge(sc, CHELSIO_T6) ?
1226 		    X_FETCHBURSTMIN_64B_T6: X_FETCHBURSTMIN_128B;
1227 		const uint_t fbmax = t4_cver_ge(sc, CHELSIO_T6) ?
1228 		    X_FETCHBURSTMAX_256B : X_FETCHBURSTMAX_512B;
1229 		c.fl0dcaen_to_fl0cidxfthresh = cpu_to_be16(
1230 		    V_FW_IQ_CMD_FL0FBMIN(fbmin) |
1231 		    V_FW_IQ_CMD_FL0FBMAX(fbmax));
1232 		c.fl0size = cpu_to_be16(fl->qsize);
1233 		c.fl0addr = cpu_to_be64(fl->ba);
1234 	}
1235 
1236 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1237 	if (rc != 0) {
1238 		cxgb_printf(sc->dip, CE_WARN,
1239 		    "failed to create ingress queue: %d", rc);
1240 		return (rc);
1241 	}
1242 
1243 	iq->cdesc = iq->desc;
1244 	iq->cidx = 0;
1245 	iq->gen = 1;
1246 	iq->adapter = sc;
1247 	iq->cntxt_id = be16_to_cpu(c.iqid);
1248 	iq->abs_id = be16_to_cpu(c.physiqid);
1249 	iq->flags |= IQ_ALLOCATED;
1250 	mutex_init(&iq->lock, NULL, MUTEX_DRIVER,
1251 	    DDI_INTR_PRI(DDI_INTR_PRI(sc->intr_pri)));
1252 	iq->polling = 0;
1253 
1254 	*t4_iqmap_slot(sc, iq->cntxt_id) = iq;
1255 
1256 	if (fl != NULL) {
1257 		fl->cntxt_id = be16_to_cpu(c.fl0id);
1258 		fl->pidx = fl->cidx = 0;
1259 		fl->copy_threshold = rx_copy_threshold;
1260 
1261 		*t4_eqmap_slot(sc, fl->cntxt_id) = (struct sge_eq *)fl;
1262 
1263 		FL_LOCK(fl);
1264 		(void) refill_fl(sc, fl, fl->lowat);
1265 		FL_UNLOCK(fl);
1266 
1267 		iq->flags |= IQ_HAS_FL;
1268 	}
1269 
1270 	if (t4_cver_ge(sc, CHELSIO_T5) && cong >= 0) {
1271 		uint32_t param, val;
1272 
1273 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
1274 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
1275 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
1276 		if (cong == 0)
1277 			val = 1 << 19;
1278 		else {
1279 			val = 2 << 19;
1280 			for (i = 0; i < 4; i++) {
1281 				if (cong & (1 << i))
1282 					val |= 1 << (i << 2);
1283 			}
1284 		}
1285 
1286 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
1287 		if (rc != 0) {
1288 			/* report error but carry on */
1289 			cxgb_printf(sc->dip, CE_WARN,
1290 			    "failed to set congestion manager context for "
1291 			    "ingress queue %d: %d", iq->cntxt_id, rc);
1292 		}
1293 	}
1294 
1295 	/* Enable IQ interrupts */
1296 	iq->state = IQS_IDLE;
1297 	t4_iq_gts_update(iq, iq->intr_params, 0);
1298 
1299 	return (0);
1300 }
1301 
1302 static int
free_iq_fl(struct port_info * pi,struct sge_iq * iq,struct sge_fl * fl)1303 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
1304 {
1305 	int rc;
1306 
1307 	if (iq != NULL) {
1308 		struct adapter *sc = iq->adapter;
1309 		dev_info_t *dip;
1310 
1311 		dip = pi ? pi->dip : sc->dip;
1312 		if (iq->flags & IQ_ALLOCATED) {
1313 			rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
1314 			    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
1315 			    fl ? fl->cntxt_id : 0xffff, 0xffff);
1316 			if (rc != 0) {
1317 				cxgb_printf(dip, CE_WARN,
1318 				    "failed to free queue %p: %d", iq, rc);
1319 				return (rc);
1320 			}
1321 			mutex_destroy(&iq->lock);
1322 			iq->flags &= ~IQ_ALLOCATED;
1323 		}
1324 
1325 		if (iq->desc != NULL) {
1326 			(void) free_desc_ring(&iq->dhdl, &iq->ahdl);
1327 			iq->desc = NULL;
1328 		}
1329 
1330 		bzero(iq, sizeof (*iq));
1331 	}
1332 
1333 	if (fl != NULL) {
1334 		if (fl->sdesc != NULL) {
1335 			FL_LOCK(fl);
1336 			free_fl_bufs(fl);
1337 			FL_UNLOCK(fl);
1338 
1339 			kmem_free(fl->sdesc, sizeof (struct fl_sdesc) *
1340 			    fl->cap);
1341 			fl->sdesc = NULL;
1342 		}
1343 
1344 		if (fl->desc != NULL) {
1345 			(void) free_desc_ring(&fl->dhdl, &fl->ahdl);
1346 			fl->desc = NULL;
1347 		}
1348 
1349 		if (fl->flags & FL_MTX) {
1350 			mutex_destroy(&fl->lock);
1351 			fl->flags &= ~FL_MTX;
1352 		}
1353 
1354 		bzero(fl, sizeof (struct sge_fl));
1355 	}
1356 
1357 	return (0);
1358 }
1359 
1360 int
t4_alloc_fwq(struct adapter * sc)1361 t4_alloc_fwq(struct adapter *sc)
1362 {
1363 	int rc, intr_idx;
1364 	struct sge_iq *fwq = &sc->sge.fwq;
1365 
1366 	init_iq(fwq, sc, sc->sge.fwq_tmr_idx, sc->sge.fwq_pktc_idx,
1367 	    FW_IQ_QSIZE, FW_IQ_ESIZE);
1368 	fwq->flags |= IQ_INTR;	/* always */
1369 	intr_idx = sc->intr_count > 1 ? 1 : 0;
1370 	rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1);
1371 	if (rc != 0) {
1372 		cxgb_printf(sc->dip, CE_WARN,
1373 		    "failed to create firmware event queue: %d.", rc);
1374 		return (rc);
1375 	}
1376 
1377 	return (0);
1378 }
1379 
1380 int
t4_free_fwq(struct adapter * sc)1381 t4_free_fwq(struct adapter *sc)
1382 {
1383 	return (free_iq_fl(NULL, &sc->sge.fwq, NULL));
1384 }
1385 
1386 static int
alloc_rxq(struct port_info * pi,struct sge_rxq * rxq,int intr_idx,int i)1387 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int i)
1388 {
1389 	int rc;
1390 
1391 	rxq->port = pi;
1392 	rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx,
1393 	    t4_get_tp_ch_map(pi->adapter, pi->tx_chan));
1394 	if (rc != 0)
1395 		return (rc);
1396 
1397 	rxq->ksp = setup_rxq_kstats(pi, rxq, i);
1398 
1399 	return (rc);
1400 }
1401 
1402 static int
free_rxq(struct port_info * pi,struct sge_rxq * rxq)1403 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
1404 {
1405 	int rc;
1406 
1407 	if (rxq->ksp != NULL) {
1408 		kstat_delete(rxq->ksp);
1409 		rxq->ksp = NULL;
1410 	}
1411 
1412 	rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
1413 	if (rc == 0)
1414 		bzero(&rxq->fl, sizeof (*rxq) - offsetof(struct sge_rxq, fl));
1415 
1416 	return (rc);
1417 }
1418 
1419 static int
eth_eq_alloc(struct adapter * sc,struct port_info * pi,struct sge_eq * eq)1420 eth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1421 {
1422 	struct fw_eq_eth_cmd c = {
1423 		.op_to_vfn = BE_32(
1424 		    V_FW_CMD_OP(FW_EQ_ETH_CMD) |
1425 		    F_FW_CMD_REQUEST | F_FW_CMD_WRITE | F_FW_CMD_EXEC |
1426 		    V_FW_EQ_ETH_CMD_PFN(sc->pf) |
1427 		    V_FW_EQ_ETH_CMD_VFN(0)),
1428 		.alloc_to_len16 = BE_32(
1429 		    F_FW_EQ_ETH_CMD_ALLOC |
1430 		    F_FW_EQ_ETH_CMD_EQSTART |
1431 		    FW_LEN16(struct fw_eq_eth_cmd)),
1432 		.autoequiqe_to_viid = BE_32(
1433 		    F_FW_EQ_ETH_CMD_AUTOEQUIQE |
1434 		    F_FW_EQ_ETH_CMD_AUTOEQUEQE |
1435 		    V_FW_EQ_ETH_CMD_VIID(pi->viid)),
1436 		.fetchszm_to_iqid = BE_32(
1437 		    V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_BOTH) |
1438 		    V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) |
1439 		    F_FW_EQ_ETH_CMD_FETCHRO |
1440 		    V_FW_EQ_ETH_CMD_IQID(eq->iqid)),
1441 		.dcaen_to_eqsize = BE_32(
1442 		    V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1443 		    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1444 		    V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1445 		    V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize)),
1446 		.eqaddr = BE_64(eq->ba),
1447 	};
1448 
1449 	/*
1450 	 * The EQ is configured to send a notification for every 32 consumed
1451 	 * entries (X_CIDXFLUSHTHRESH_32).  In order to ensure timely
1452 	 * notification of entry consumption during slow periods when that
1453 	 * threshold may not be reached with regularity, two mechanisms exist:
1454 	 *
1455 	 * 1. The DBQ timer can be configured to fire (and send a notification)
1456 	 *    after a period when the EQ has gone idle.  This is available on T6
1457 	 *    and later adapters.
1458 	 *
1459 	 * 2. The CIDXFlushThresholdOverride flag will send a notification
1460 	 *    whenever a consumed entry causes CDIX==PIDX, even if the
1461 	 *    CIDXFlushThreshold has not been reached.
1462 	 *
1463 	 * The DBQ timer is preferred, as it results in no additional
1464 	 * notifications when the EQ is kept busy with small transmissions.
1465 	 * Comparatively, flows of many short packets (like frequent ACKs) can
1466 	 * cause the CIDXFlushThresholdOverride mechanism to induce a
1467 	 * notification for every transmitted packet.
1468 	 */
1469 	if (sc->flags & TAF_DBQ_TIMER) {
1470 		/* Configure the DBQ timer when it is available */
1471 		c.timeren_timerix = BE_32(
1472 		    F_FW_EQ_ETH_CMD_TIMEREN |
1473 		    V_FW_EQ_ETH_CMD_TIMERIX(pi->dbq_timer_idx));
1474 	} else {
1475 		/* Otherwise fall back to CIDXFlushThresholdOverride */
1476 		c.dcaen_to_eqsize |= BE_32(F_FW_EQ_ETH_CMD_CIDXFTHRESHO);
1477 	}
1478 
1479 	int rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1480 	if (rc != 0) {
1481 		cxgb_printf(pi->dip, CE_WARN,
1482 		    "failed to create Ethernet egress queue: %d", rc);
1483 		return (rc);
1484 	}
1485 	eq->flags |= EQ_ALLOCATED;
1486 
1487 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(BE_32(c.eqid_pkd));
1488 
1489 	*t4_eqmap_slot(sc, eq->cntxt_id) = eq;
1490 
1491 	return (rc);
1492 }
1493 
1494 static int
alloc_eq(struct adapter * sc,struct port_info * pi,struct sge_eq * eq)1495 alloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1496 {
1497 	int rc;
1498 	size_t len;
1499 
1500 	mutex_init(&eq->lock, NULL, MUTEX_DRIVER, DDI_INTR_PRI(sc->intr_pri));
1501 	eq->flags |= EQ_MTX;
1502 
1503 	len = eq->qsize * EQ_ESIZE;
1504 	rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &eq->desc_dhdl,
1505 	    &eq->desc_ahdl, &eq->ba, (caddr_t *)&eq->desc);
1506 	if (rc != 0)
1507 		return (rc);
1508 
1509 	eq->cap = eq->qsize - sc->sge.stat_len / EQ_ESIZE;
1510 	eq->spg = (void *)&eq->desc[eq->cap];
1511 	eq->avail = eq->cap - 1;	/* one less to avoid cidx = pidx */
1512 	eq->pidx = eq->cidx = 0;
1513 	eq->doorbells = sc->doorbells;
1514 
1515 	rc = eth_eq_alloc(sc, pi, eq);
1516 	if (rc != 0) {
1517 		cxgb_printf(sc->dip, CE_WARN,
1518 		    "failed to allocate egress queue: %d", rc);
1519 	}
1520 
1521 	if (eq->doorbells & (DOORBELL_UDB | DOORBELL_UDBWC | DOORBELL_WCWR)) {
1522 		uint64_t udb_offset;
1523 		uint_t udb_qid;
1524 
1525 		rc = t4_bar2_sge_qregs(sc, eq->cntxt_id, T4_BAR2_QTYPE_EGRESS,
1526 		    0, &udb_offset, &udb_qid);
1527 
1528 		if (rc == 0) {
1529 			eq->udb = sc->bar2_ptr + udb_offset;
1530 			eq->udb_qid = udb_qid;
1531 		} else {
1532 			eq->doorbells &=
1533 			    ~(DOORBELL_UDB | DOORBELL_UDBWC | DOORBELL_WCWR);
1534 			eq->udb = NULL;
1535 			eq->udb_qid = 0;
1536 		}
1537 	}
1538 
1539 	return (rc);
1540 }
1541 
1542 static int
free_eq(struct adapter * sc,struct sge_eq * eq)1543 free_eq(struct adapter *sc, struct sge_eq *eq)
1544 {
1545 	int rc;
1546 
1547 	if (eq->flags & EQ_ALLOCATED) {
1548 		rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id);
1549 		if (rc != 0) {
1550 			cxgb_printf(sc->dip, CE_WARN,
1551 			    "failed to free egress queue: %d", rc);
1552 			return (rc);
1553 		}
1554 		eq->flags &= ~EQ_ALLOCATED;
1555 	}
1556 
1557 	if (eq->desc != NULL) {
1558 		(void) free_desc_ring(&eq->desc_dhdl, &eq->desc_ahdl);
1559 		eq->desc = NULL;
1560 	}
1561 
1562 	if (eq->flags & EQ_MTX)
1563 		mutex_destroy(&eq->lock);
1564 
1565 	bzero(eq, sizeof (*eq));
1566 	return (0);
1567 }
1568 
1569 static int
alloc_txq(struct port_info * pi,struct sge_txq * txq,int idx)1570 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx)
1571 {
1572 	int rc, i;
1573 	struct adapter *sc = pi->adapter;
1574 	struct sge_eq *eq = &txq->eq;
1575 
1576 	rc = alloc_eq(sc, pi, eq);
1577 	if (rc != 0)
1578 		return (rc);
1579 
1580 	txq->port = pi;
1581 	txq->sdesc = kmem_zalloc(sizeof (struct tx_sdesc) * eq->cap, KM_SLEEP);
1582 	txq->copy_threshold = tx_copy_threshold;
1583 	txq->txb_size = eq->qsize * txq->copy_threshold;
1584 	rc = alloc_tx_copybuffer(sc, txq->txb_size, &txq->txb_dhdl,
1585 	    &txq->txb_ahdl, &txq->txb_ba, &txq->txb_va);
1586 	if (rc == 0)
1587 		txq->txb_avail = txq->txb_size;
1588 	else
1589 		txq->txb_avail = txq->txb_size = 0;
1590 
1591 	/*
1592 	 * TODO: is this too low?  Worst case would need around 4 times qsize
1593 	 * (all tx descriptors filled to the brim with SGLs, with each entry in
1594 	 * the SGL coming from a distinct DMA handle).  Increase tx_dhdl_total
1595 	 * if you see too many dma_hdl_failed.
1596 	 */
1597 	txq->tx_dhdl_total = eq->qsize * 2;
1598 	txq->tx_dhdl = kmem_zalloc(sizeof (ddi_dma_handle_t) *
1599 	    txq->tx_dhdl_total, KM_SLEEP);
1600 	for (i = 0; i < txq->tx_dhdl_total; i++) {
1601 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
1602 		    DDI_DMA_SLEEP, 0, &txq->tx_dhdl[i]);
1603 		if (rc != DDI_SUCCESS) {
1604 			cxgb_printf(sc->dip, CE_WARN,
1605 			    "%s: failed to allocate DMA handle (%d)",
1606 			    __func__, rc);
1607 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
1608 		}
1609 		txq->tx_dhdl_avail++;
1610 	}
1611 
1612 	txq->ksp = setup_txq_kstats(pi, txq, idx);
1613 
1614 	return (rc);
1615 }
1616 
1617 static int
free_txq(struct port_info * pi,struct sge_txq * txq)1618 free_txq(struct port_info *pi, struct sge_txq *txq)
1619 {
1620 	int i;
1621 	struct adapter *sc = pi->adapter;
1622 	struct sge_eq *eq = &txq->eq;
1623 
1624 	if (txq->ksp != NULL) {
1625 		kstat_delete(txq->ksp);
1626 		txq->ksp = NULL;
1627 	}
1628 
1629 	if (txq->txb_va != NULL) {
1630 		(void) free_desc_ring(&txq->txb_dhdl, &txq->txb_ahdl);
1631 		txq->txb_va = NULL;
1632 	}
1633 
1634 	if (txq->sdesc != NULL) {
1635 		struct tx_sdesc *sd;
1636 		ddi_dma_handle_t hdl;
1637 
1638 		TXQ_LOCK(txq);
1639 		while (eq->cidx != eq->pidx) {
1640 			sd = &txq->sdesc[eq->cidx];
1641 
1642 			for (i = sd->hdls_used; i; i--) {
1643 				hdl = txq->tx_dhdl[txq->tx_dhdl_cidx];
1644 				(void) ddi_dma_unbind_handle(hdl);
1645 				if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
1646 					txq->tx_dhdl_cidx = 0;
1647 			}
1648 
1649 			ASSERT(sd->mp_head);
1650 			freemsgchain(sd->mp_head);
1651 			sd->mp_head = sd->mp_tail = NULL;
1652 
1653 			eq->cidx += sd->desc_used;
1654 			if (eq->cidx >= eq->cap)
1655 				eq->cidx -= eq->cap;
1656 
1657 			txq->txb_avail += txq->txb_used;
1658 		}
1659 		ASSERT(txq->tx_dhdl_cidx == txq->tx_dhdl_pidx);
1660 		ASSERT(txq->txb_avail == txq->txb_size);
1661 		TXQ_UNLOCK(txq);
1662 
1663 		kmem_free(txq->sdesc, sizeof (struct tx_sdesc) * eq->cap);
1664 		txq->sdesc = NULL;
1665 	}
1666 
1667 	if (txq->tx_dhdl != NULL) {
1668 		for (i = 0; i < txq->tx_dhdl_total; i++) {
1669 			if (txq->tx_dhdl[i] != NULL)
1670 				ddi_dma_free_handle(&txq->tx_dhdl[i]);
1671 		}
1672 		kmem_free(txq->tx_dhdl,
1673 		    sizeof (ddi_dma_handle_t) * txq->tx_dhdl_total);
1674 		txq->tx_dhdl = NULL;
1675 	}
1676 
1677 	(void) free_eq(sc, &txq->eq);
1678 
1679 	bzero(txq, sizeof (*txq));
1680 	return (0);
1681 }
1682 
1683 /*
1684  * Allocates a block of contiguous memory for DMA.  Can be used to allocate
1685  * memory for descriptor rings or for tx/rx copy buffers.
1686  *
1687  * Caller does not have to clean up anything if this function fails, it cleans
1688  * up after itself.
1689  *
1690  * Caller provides the following:
1691  * len		length of the block of memory to allocate.
1692  * flags	DDI_DMA_* flags to use (CONSISTENT/STREAMING, READ/WRITE/RDWR)
1693  * acc_attr	device access attributes for the allocation.
1694  * dma_attr	DMA attributes for the allocation
1695  *
1696  * If the function is successful it fills up this information:
1697  * dma_hdl	DMA handle for the allocated memory
1698  * acc_hdl	access handle for the allocated memory
1699  * ba		bus address of the allocated memory
1700  * va		KVA of the allocated memory.
1701  */
1702 static int
alloc_dma_memory(struct adapter * sc,size_t len,int flags,ddi_device_acc_attr_t * acc_attr,ddi_dma_attr_t * dma_attr,ddi_dma_handle_t * dma_hdl,ddi_acc_handle_t * acc_hdl,uint64_t * pba,caddr_t * pva)1703 alloc_dma_memory(struct adapter *sc, size_t len, int flags,
1704     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
1705     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
1706     uint64_t *pba, caddr_t *pva)
1707 {
1708 	int rc;
1709 	ddi_dma_handle_t dhdl;
1710 	ddi_acc_handle_t ahdl;
1711 	ddi_dma_cookie_t cookie;
1712 	uint_t ccount;
1713 	caddr_t va;
1714 	size_t real_len;
1715 
1716 	*pva = NULL;
1717 
1718 	/*
1719 	 * DMA handle.
1720 	 */
1721 	rc = ddi_dma_alloc_handle(sc->dip, dma_attr, DDI_DMA_SLEEP, 0, &dhdl);
1722 	if (rc != DDI_SUCCESS) {
1723 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
1724 	}
1725 
1726 	/*
1727 	 * Memory suitable for DMA.
1728 	 */
1729 	rc = ddi_dma_mem_alloc(dhdl, len, acc_attr,
1730 	    flags & DDI_DMA_CONSISTENT ? DDI_DMA_CONSISTENT : DDI_DMA_STREAMING,
1731 	    DDI_DMA_SLEEP, 0, &va, &real_len, &ahdl);
1732 	if (rc != DDI_SUCCESS) {
1733 		ddi_dma_free_handle(&dhdl);
1734 		return (ENOMEM);
1735 	}
1736 
1737 	/*
1738 	 * DMA bindings.
1739 	 */
1740 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, va, real_len, flags, NULL,
1741 	    NULL, &cookie, &ccount);
1742 	if (rc != DDI_DMA_MAPPED) {
1743 		ddi_dma_mem_free(&ahdl);
1744 		ddi_dma_free_handle(&dhdl);
1745 		return (ENOMEM);
1746 	}
1747 	if (ccount != 1) {
1748 		/* unusable DMA mapping */
1749 		(void) free_desc_ring(&dhdl, &ahdl);
1750 		return (ENOMEM);
1751 	}
1752 
1753 	bzero(va, real_len);
1754 	*dma_hdl = dhdl;
1755 	*acc_hdl = ahdl;
1756 	*pba = cookie.dmac_laddress;
1757 	*pva = va;
1758 
1759 	return (0);
1760 }
1761 
1762 static int
free_dma_memory(ddi_dma_handle_t * dhdl,ddi_acc_handle_t * ahdl)1763 free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
1764 {
1765 	(void) ddi_dma_unbind_handle(*dhdl);
1766 	ddi_dma_mem_free(ahdl);
1767 	ddi_dma_free_handle(dhdl);
1768 
1769 	return (0);
1770 }
1771 
1772 static int
alloc_desc_ring(struct adapter * sc,size_t len,int rw,ddi_dma_handle_t * dma_hdl,ddi_acc_handle_t * acc_hdl,uint64_t * pba,caddr_t * pva)1773 alloc_desc_ring(struct adapter *sc, size_t len, int rw,
1774     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
1775     uint64_t *pba, caddr_t *pva)
1776 {
1777 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_desc;
1778 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc;
1779 
1780 	return (alloc_dma_memory(sc, len, DDI_DMA_CONSISTENT | rw, acc_attr,
1781 	    dma_attr, dma_hdl, acc_hdl, pba, pva));
1782 }
1783 
1784 static int
free_desc_ring(ddi_dma_handle_t * dhdl,ddi_acc_handle_t * ahdl)1785 free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
1786 {
1787 	return (free_dma_memory(dhdl, ahdl));
1788 }
1789 
1790 static int
alloc_tx_copybuffer(struct adapter * sc,size_t len,ddi_dma_handle_t * dma_hdl,ddi_acc_handle_t * acc_hdl,uint64_t * pba,caddr_t * pva)1791 alloc_tx_copybuffer(struct adapter *sc, size_t len,
1792     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
1793     uint64_t *pba, caddr_t *pva)
1794 {
1795 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_tx;
1796 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc; /* NOT dma_attr_tx */
1797 
1798 	return (alloc_dma_memory(sc, len, DDI_DMA_STREAMING | DDI_DMA_WRITE,
1799 	    acc_attr, dma_attr, dma_hdl, acc_hdl, pba, pva));
1800 }
1801 
1802 static inline bool
is_new_response(const struct sge_iq * iq,struct rsp_ctrl ** ctrl)1803 is_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl)
1804 {
1805 	(void) ddi_dma_sync(iq->dhdl, (uintptr_t)iq->cdesc -
1806 	    (uintptr_t)iq->desc, iq->esize, DDI_DMA_SYNC_FORKERNEL);
1807 
1808 	*ctrl = (void *)((uintptr_t)iq->cdesc +
1809 	    (iq->esize - sizeof (struct rsp_ctrl)));
1810 
1811 	return ((((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen));
1812 }
1813 
1814 static inline void
iq_next(struct sge_iq * iq)1815 iq_next(struct sge_iq *iq)
1816 {
1817 	iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize);
1818 	if (++iq->cidx == iq->qsize - 1) {
1819 		iq->cidx = 0;
1820 		iq->gen ^= 1;
1821 		iq->cdesc = iq->desc;
1822 	}
1823 }
1824 
1825 /*
1826  * Fill up the freelist by upto nbufs and maybe ring its doorbell.
1827  *
1828  * Returns non-zero to indicate that it should be added to the list of starving
1829  * freelists.
1830  */
1831 static int
refill_fl(struct adapter * sc,struct sge_fl * fl,int nbufs)1832 refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs)
1833 {
1834 	uint64_t *d = &fl->desc[fl->pidx];
1835 	struct fl_sdesc *sd = &fl->sdesc[fl->pidx];
1836 
1837 	FL_LOCK_ASSERT_OWNED(fl);
1838 	ASSERT(nbufs >= 0);
1839 
1840 	if (nbufs > fl->needed)
1841 		nbufs = fl->needed;
1842 
1843 	while (nbufs--) {
1844 		if (sd->rxb != NULL) {
1845 			if (sd->rxb->ref_cnt == 1) {
1846 				/*
1847 				 * Buffer is available for recycling.  Two ways
1848 				 * this can happen:
1849 				 *
1850 				 * a) All the packets DMA'd into it last time
1851 				 *    around were within the rx_copy_threshold
1852 				 *    and no part of the buffer was ever passed
1853 				 *    up (ref_cnt never went over 1).
1854 				 *
1855 				 * b) Packets DMA'd into the buffer were passed
1856 				 *    up but have all been freed by the upper
1857 				 *    layers by now (ref_cnt went over 1 but is
1858 				 *    now back to 1).
1859 				 *
1860 				 * Either way the bus address in the descriptor
1861 				 * ring is already valid.
1862 				 */
1863 				ASSERT(*d == cpu_to_be64(sd->rxb->ba));
1864 				d++;
1865 				goto recycled;
1866 			} else {
1867 				/*
1868 				 * Buffer still in use and we need a
1869 				 * replacement. But first release our reference
1870 				 * on the existing buffer.
1871 				 */
1872 				rxbuf_free(sd->rxb);
1873 			}
1874 		}
1875 
1876 		sd->rxb = rxbuf_alloc(sc->sge.rxbuf_cache, KM_NOSLEEP, 1);
1877 		if (sd->rxb == NULL)
1878 			break;
1879 		*d++ = cpu_to_be64(sd->rxb->ba);
1880 
1881 recycled:	fl->pending++;
1882 		sd++;
1883 		fl->needed--;
1884 		if (++fl->pidx == fl->cap) {
1885 			fl->pidx = 0;
1886 			sd = fl->sdesc;
1887 			d = fl->desc;
1888 		}
1889 	}
1890 
1891 	if (fl->pending >= 8)
1892 		ring_fl_db(sc, fl);
1893 
1894 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
1895 }
1896 
1897 #ifndef TAILQ_FOREACH_SAFE
1898 #define	TAILQ_FOREACH_SAFE(var, head, field, tvar)			\
1899 	for ((var) = TAILQ_FIRST((head));				\
1900 	    (var) && ((tvar) = TAILQ_NEXT((var), field), 1);		\
1901 	    (var) = (tvar))
1902 #endif
1903 
1904 /*
1905  * Attempt to refill all starving freelists.
1906  */
1907 static void
refill_sfl(void * arg)1908 refill_sfl(void *arg)
1909 {
1910 	struct adapter *sc = arg;
1911 	struct sge_fl *fl, *fl_temp;
1912 
1913 	mutex_enter(&sc->sfl_lock);
1914 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
1915 		FL_LOCK(fl);
1916 		(void) refill_fl(sc, fl, 64);
1917 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
1918 			TAILQ_REMOVE(&sc->sfl, fl, link);
1919 			fl->flags &= ~FL_STARVING;
1920 		}
1921 		FL_UNLOCK(fl);
1922 	}
1923 
1924 	if (!TAILQ_EMPTY(&sc->sfl) != 0)
1925 		sc->sfl_timer =  timeout(refill_sfl, sc, drv_usectohz(100000));
1926 	mutex_exit(&sc->sfl_lock);
1927 }
1928 
1929 static void
add_fl_to_sfl(struct adapter * sc,struct sge_fl * fl)1930 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
1931 {
1932 	mutex_enter(&sc->sfl_lock);
1933 	FL_LOCK(fl);
1934 	if ((fl->flags & FL_DOOMED) == 0) {
1935 		if (TAILQ_EMPTY(&sc->sfl) != 0) {
1936 			sc->sfl_timer = timeout(refill_sfl, sc,
1937 			    drv_usectohz(100000));
1938 		}
1939 		fl->flags |= FL_STARVING;
1940 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
1941 	}
1942 	FL_UNLOCK(fl);
1943 	mutex_exit(&sc->sfl_lock);
1944 }
1945 
1946 static void
free_fl_bufs(struct sge_fl * fl)1947 free_fl_bufs(struct sge_fl *fl)
1948 {
1949 	struct fl_sdesc *sd;
1950 	unsigned int i;
1951 
1952 	FL_LOCK_ASSERT_OWNED(fl);
1953 
1954 	for (i = 0; i < fl->cap; i++) {
1955 		sd = &fl->sdesc[i];
1956 
1957 		if (sd->rxb != NULL) {
1958 			rxbuf_free(sd->rxb);
1959 			sd->rxb = NULL;
1960 		}
1961 	}
1962 }
1963 
1964 /*
1965  * Note that fl->cidx and fl->offset are left unchanged in case of failure.
1966  */
1967 static mblk_t *
get_fl_payload(struct adapter * sc,struct sge_fl * fl,uint32_t len_newbuf,int * fl_bufs_used)1968 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf,
1969     int *fl_bufs_used)
1970 {
1971 	struct mblk_pair frame = {0};
1972 	struct rxbuf *rxb;
1973 	mblk_t *m = NULL;
1974 	uint_t nbuf = 0, len, copy, n;
1975 	uint32_t cidx, offset, rcidx, roffset;
1976 
1977 	/*
1978 	 * The SGE won't pack a new frame into the current buffer if the entire
1979 	 * payload doesn't fit in the remaining space.  Move on to the next buf
1980 	 * in that case.
1981 	 */
1982 	rcidx = fl->cidx;
1983 	roffset = fl->offset;
1984 	if (fl->offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1985 		fl->offset = 0;
1986 		if (++fl->cidx == fl->cap)
1987 			fl->cidx = 0;
1988 		nbuf++;
1989 	}
1990 	cidx = fl->cidx;
1991 	offset = fl->offset;
1992 
1993 	len = G_RSPD_LEN(len_newbuf);	/* pktshift + payload length */
1994 	copy = (len <= fl->copy_threshold);
1995 	if (copy != 0) {
1996 		frame.head = m = allocb(len, BPRI_HI);
1997 		if (m == NULL) {
1998 			fl->allocb_fail++;
1999 			DTRACE_PROBE1(t4__fl_alloc_fail, struct sge_fl *, fl);
2000 			fl->cidx = rcidx;
2001 			fl->offset = roffset;
2002 			return (NULL);
2003 		}
2004 	}
2005 
2006 	while (len) {
2007 		rxb = fl->sdesc[cidx].rxb;
2008 		n = min(len, rxb->buf_size - offset);
2009 
2010 		(void) ddi_dma_sync(rxb->dhdl, offset, n,
2011 		    DDI_DMA_SYNC_FORKERNEL);
2012 
2013 		if (copy != 0)
2014 			bcopy(rxb->va + offset, m->b_wptr, n);
2015 		else {
2016 			m = desballoc((unsigned char *)rxb->va + offset, n,
2017 			    BPRI_HI, &rxb->freefunc);
2018 			if (m == NULL) {
2019 				fl->allocb_fail++;
2020 				DTRACE_PROBE1(t4__fl_alloc_fail,
2021 				    struct sge_fl *, fl);
2022 				if (frame.head)
2023 					freemsgchain(frame.head);
2024 				fl->cidx = rcidx;
2025 				fl->offset = roffset;
2026 				return (NULL);
2027 			}
2028 			atomic_inc_uint(&rxb->ref_cnt);
2029 			if (frame.head != NULL)
2030 				frame.tail->b_cont = m;
2031 			else
2032 				frame.head = m;
2033 			frame.tail = m;
2034 		}
2035 		m->b_wptr += n;
2036 		len -= n;
2037 		offset += roundup(n, sc->sge.fl_align);
2038 		ASSERT(offset <= rxb->buf_size);
2039 		if (offset == rxb->buf_size) {
2040 			offset = 0;
2041 			if (++cidx == fl->cap)
2042 				cidx = 0;
2043 			nbuf++;
2044 		}
2045 	}
2046 
2047 	fl->cidx = cidx;
2048 	fl->offset = offset;
2049 	(*fl_bufs_used) += nbuf;
2050 
2051 	ASSERT(frame.head != NULL);
2052 	return (frame.head);
2053 }
2054 
2055 /*
2056  * We'll do immediate data tx for non-LSO, but only when not coalescing.  We're
2057  * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes
2058  * of immediate data.
2059  */
2060 #define	IMM_LEN ( \
2061 	2 * EQ_ESIZE \
2062 	- sizeof (struct fw_eth_tx_pkt_wr) \
2063 	- sizeof (struct cpl_tx_pkt_core))
2064 
2065 /*
2066  * Returns non-zero on failure, no need to cleanup anything in that case.
2067  *
2068  * Note 1: We always try to pull up the mblk if required and return E2BIG only
2069  * if this fails.
2070  *
2071  * Note 2: We'll also pullup incoming mblk if HW_LSO is set and the first mblk
2072  * does not have the TCP header in it.
2073  */
2074 static int
get_frame_txinfo(struct sge_txq * txq,mblk_t ** fp,struct txinfo * txinfo,int sgl_only)2075 get_frame_txinfo(struct sge_txq *txq, mblk_t **fp, struct txinfo *txinfo,
2076     int sgl_only)
2077 {
2078 	uint32_t flags = 0, len, n;
2079 	mblk_t *m = *fp;
2080 	int rc;
2081 
2082 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb and dma_hdls */
2083 
2084 	mac_hcksum_get(m, NULL, NULL, NULL, NULL, &flags);
2085 	txinfo->flags = (flags & HCK_TX_FLAGS);
2086 
2087 	mac_lso_get(m, &txinfo->mss, &flags);
2088 	txinfo->flags |= (flags & HW_LSO_FLAGS);
2089 
2090 	if (flags & HW_LSO)
2091 		sgl_only = 1;	/* Do not allow immediate data with LSO */
2092 
2093 	/*
2094 	 * If checksum or segmentation offloads are requested, gather
2095 	 * information about the sizes and types of headers in the packet.
2096 	 */
2097 	if (txinfo->flags != 0) {
2098 		mac_ether_offload_info(m, &txinfo->meoi);
2099 	} else {
2100 		bzero(&txinfo->meoi, sizeof (txinfo->meoi));
2101 	}
2102 
2103 start:	txinfo->nsegs = 0;
2104 	txinfo->hdls_used = 0;
2105 	txinfo->txb_used = 0;
2106 	txinfo->len = 0;
2107 
2108 	/* total length and a rough estimate of # of segments */
2109 	n = 0;
2110 	for (; m; m = m->b_cont) {
2111 		len = MBLKL(m);
2112 		n += (len / PAGE_SIZE) + 1;
2113 		txinfo->len += len;
2114 	}
2115 	m = *fp;
2116 
2117 	if (n >= TX_SGL_SEGS || (flags & HW_LSO && MBLKL(m) < 50)) {
2118 		txq->pullup_early++;
2119 		m = msgpullup(*fp, -1);
2120 		if (m == NULL) {
2121 			txq->pullup_failed++;
2122 			return (E2BIG);	/* (*fp) left as it was */
2123 		}
2124 		freemsg(*fp);
2125 		*fp = m;
2126 		mac_hcksum_set(m, 0, 0, 0, 0, txinfo->flags);
2127 	}
2128 
2129 	if (txinfo->len <= IMM_LEN && !sgl_only)
2130 		return (0);	/* nsegs = 0 tells caller to use imm. tx */
2131 
2132 	if (txinfo->len <= txq->copy_threshold &&
2133 	    copy_into_txb(txq, m, txinfo->len, txinfo) == 0)
2134 		goto done;
2135 
2136 	for (; m; m = m->b_cont) {
2137 
2138 		len = MBLKL(m);
2139 
2140 		/* Use tx copy buffer if this mblk is small enough */
2141 		if (len <= txq->copy_threshold &&
2142 		    copy_into_txb(txq, m, len, txinfo) == 0)
2143 			continue;
2144 
2145 		/* Add DMA bindings for this mblk to the SGL */
2146 		rc = add_mblk(txq, txinfo, m, len);
2147 
2148 		if (rc == E2BIG ||
2149 		    (txinfo->nsegs == TX_SGL_SEGS && m->b_cont)) {
2150 
2151 			txq->pullup_late++;
2152 			m = msgpullup(*fp, -1);
2153 			if (m != NULL) {
2154 				free_txinfo_resources(txq, txinfo);
2155 				freemsg(*fp);
2156 				*fp = m;
2157 				mac_hcksum_set(m, 0, 0, 0, 0, txinfo->flags);
2158 				goto start;
2159 			}
2160 
2161 			txq->pullup_failed++;
2162 			rc = E2BIG;
2163 		}
2164 
2165 		if (rc != 0) {
2166 			free_txinfo_resources(txq, txinfo);
2167 			return (rc);
2168 		}
2169 	}
2170 
2171 	ASSERT(txinfo->nsegs > 0 && txinfo->nsegs <= TX_SGL_SEGS);
2172 
2173 done:
2174 
2175 	/*
2176 	 * Store the # of flits required to hold this frame's SGL in nflits.  An
2177 	 * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by
2178 	 * multiple (len0 + len1, addr0, addr1) tuples.  If addr1 is not used
2179 	 * then len1 must be set to 0.
2180 	 */
2181 	n = txinfo->nsegs - 1;
2182 	txinfo->nflits = (3 * n) / 2 + (n & 1) + 2;
2183 	if (n & 1)
2184 		txinfo->sgl.sge[n / 2].len[1] = cpu_to_be32(0);
2185 
2186 	txinfo->sgl.cmd_nsge = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_DSGL) |
2187 	    V_ULPTX_NSGE(txinfo->nsegs));
2188 
2189 	return (0);
2190 }
2191 
2192 static inline int
fits_in_txb(struct sge_txq * txq,int len,int * waste)2193 fits_in_txb(struct sge_txq *txq, int len, int *waste)
2194 {
2195 	if (txq->txb_avail < len)
2196 		return (0);
2197 
2198 	if (txq->txb_next + len <= txq->txb_size) {
2199 		*waste = 0;
2200 		return (1);
2201 	}
2202 
2203 	*waste = txq->txb_size - txq->txb_next;
2204 
2205 	return (txq->txb_avail - *waste < len ? 0 : 1);
2206 }
2207 
2208 #define	TXB_CHUNK	64
2209 
2210 /*
2211  * Copies the specified # of bytes into txq's tx copy buffer and updates txinfo
2212  * and txq to indicate resources used.  Caller has to make sure that those many
2213  * bytes are available in the mblk chain (b_cont linked).
2214  */
2215 static inline int
copy_into_txb(struct sge_txq * txq,mblk_t * m,int len,struct txinfo * txinfo)2216 copy_into_txb(struct sge_txq *txq, mblk_t *m, int len, struct txinfo *txinfo)
2217 {
2218 	int waste, n;
2219 
2220 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb */
2221 
2222 	if (!fits_in_txb(txq, len, &waste)) {
2223 		txq->txb_full++;
2224 		return (ENOMEM);
2225 	}
2226 
2227 	if (waste != 0) {
2228 		ASSERT((waste & (TXB_CHUNK - 1)) == 0);
2229 		txinfo->txb_used += waste;
2230 		txq->txb_avail -= waste;
2231 		txq->txb_next = 0;
2232 	}
2233 
2234 	for (n = 0; n < len; m = m->b_cont) {
2235 		bcopy(m->b_rptr, txq->txb_va + txq->txb_next + n, MBLKL(m));
2236 		n += MBLKL(m);
2237 	}
2238 
2239 	add_seg(txinfo, txq->txb_ba + txq->txb_next, len);
2240 
2241 	n = roundup(len, TXB_CHUNK);
2242 	txinfo->txb_used += n;
2243 	txq->txb_avail -= n;
2244 	txq->txb_next += n;
2245 	ASSERT(txq->txb_next <= txq->txb_size);
2246 	if (txq->txb_next == txq->txb_size)
2247 		txq->txb_next = 0;
2248 
2249 	return (0);
2250 }
2251 
2252 static inline void
add_seg(struct txinfo * txinfo,uint64_t ba,uint32_t len)2253 add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len)
2254 {
2255 	ASSERT(txinfo->nsegs < TX_SGL_SEGS);	/* must have room */
2256 
2257 	if (txinfo->nsegs != 0) {
2258 		int idx = txinfo->nsegs - 1;
2259 		txinfo->sgl.sge[idx / 2].len[idx & 1] = cpu_to_be32(len);
2260 		txinfo->sgl.sge[idx / 2].addr[idx & 1] = cpu_to_be64(ba);
2261 	} else {
2262 		txinfo->sgl.len0 = cpu_to_be32(len);
2263 		txinfo->sgl.addr0 = cpu_to_be64(ba);
2264 	}
2265 	txinfo->nsegs++;
2266 }
2267 
2268 /*
2269  * This function cleans up any partially allocated resources when it fails so
2270  * there's nothing for the caller to clean up in that case.
2271  *
2272  * EIO indicates permanent failure.  Caller should drop the frame containing
2273  * this mblk and continue.
2274  *
2275  * E2BIG indicates that the SGL length for this mblk exceeds the hardware
2276  * limit.  Caller should pull up the frame before trying to send it out.
2277  * (This error means our pullup_early heuristic did not work for this frame)
2278  *
2279  * ENOMEM indicates a temporary shortage of resources (DMA handles, other DMA
2280  * resources, etc.).  Caller should suspend the tx queue and wait for reclaim to
2281  * free up resources.
2282  */
2283 static inline int
add_mblk(struct sge_txq * txq,struct txinfo * txinfo,mblk_t * m,int len)2284 add_mblk(struct sge_txq *txq, struct txinfo *txinfo, mblk_t *m, int len)
2285 {
2286 	ddi_dma_handle_t dhdl;
2287 	ddi_dma_cookie_t cookie;
2288 	uint_t ccount = 0;
2289 	int rc;
2290 
2291 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate dhdls */
2292 
2293 	if (txq->tx_dhdl_avail == 0) {
2294 		txq->dma_hdl_failed++;
2295 		return (ENOMEM);
2296 	}
2297 
2298 	dhdl = txq->tx_dhdl[txq->tx_dhdl_pidx];
2299 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, (caddr_t)m->b_rptr, len,
2300 	    DDI_DMA_WRITE | DDI_DMA_STREAMING, DDI_DMA_DONTWAIT, NULL, &cookie,
2301 	    &ccount);
2302 	if (rc != DDI_DMA_MAPPED) {
2303 		txq->dma_map_failed++;
2304 
2305 		ASSERT(rc != DDI_DMA_INUSE && rc != DDI_DMA_PARTIAL_MAP);
2306 
2307 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EIO);
2308 	}
2309 
2310 	if (ccount + txinfo->nsegs > TX_SGL_SEGS) {
2311 		(void) ddi_dma_unbind_handle(dhdl);
2312 		return (E2BIG);
2313 	}
2314 
2315 	add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2316 	while (--ccount) {
2317 		ddi_dma_nextcookie(dhdl, &cookie);
2318 		add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2319 	}
2320 
2321 	if (++txq->tx_dhdl_pidx == txq->tx_dhdl_total)
2322 		txq->tx_dhdl_pidx = 0;
2323 	txq->tx_dhdl_avail--;
2324 	txinfo->hdls_used++;
2325 
2326 	return (0);
2327 }
2328 
2329 /*
2330  * Releases all the txq resources used up in the specified txinfo.
2331  */
2332 static void
free_txinfo_resources(struct sge_txq * txq,struct txinfo * txinfo)2333 free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo)
2334 {
2335 	int n;
2336 
2337 	TXQ_LOCK_ASSERT_OWNED(txq);	/* dhdls, txb */
2338 
2339 	n = txinfo->txb_used;
2340 	if (n > 0) {
2341 		txq->txb_avail += n;
2342 		if (n <= txq->txb_next)
2343 			txq->txb_next -= n;
2344 		else {
2345 			n -= txq->txb_next;
2346 			txq->txb_next = txq->txb_size - n;
2347 		}
2348 	}
2349 
2350 	for (n = txinfo->hdls_used; n > 0; n--) {
2351 		if (txq->tx_dhdl_pidx > 0)
2352 			txq->tx_dhdl_pidx--;
2353 		else
2354 			txq->tx_dhdl_pidx = txq->tx_dhdl_total - 1;
2355 		txq->tx_dhdl_avail++;
2356 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_pidx]);
2357 	}
2358 }
2359 
2360 /*
2361  * Returns 0 to indicate that m has been accepted into a coalesced tx work
2362  * request.  It has either been folded into txpkts or txpkts was flushed and m
2363  * has started a new coalesced work request (as the first frame in a fresh
2364  * txpkts).
2365  *
2366  * Returns non-zero to indicate a failure - caller is responsible for
2367  * transmitting m, if there was anything in txpkts it has been flushed.
2368  */
2369 static int
add_to_txpkts(struct sge_txq * txq,struct txpkts * txpkts,mblk_t * m,struct txinfo * txinfo)2370 add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
2371     struct txinfo *txinfo)
2372 {
2373 	struct sge_eq *eq = &txq->eq;
2374 	int can_coalesce;
2375 	struct tx_sdesc *txsd;
2376 	uint8_t flits;
2377 
2378 	TXQ_LOCK_ASSERT_OWNED(txq);
2379 	ASSERT(m->b_next == NULL);
2380 
2381 	if (txpkts->npkt > 0) {
2382 		flits = TXPKTS_PKT_HDR + txinfo->nflits;
2383 		can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2384 		    txpkts->nflits + flits <= TX_WR_FLITS &&
2385 		    txpkts->nflits + flits <= eq->avail * 8 &&
2386 		    txpkts->plen + txinfo->len < 65536;
2387 
2388 		if (can_coalesce != 0) {
2389 			txpkts->tail->b_next = m;
2390 			txpkts->tail = m;
2391 			txpkts->npkt++;
2392 			txpkts->nflits += flits;
2393 			txpkts->plen += txinfo->len;
2394 
2395 			txsd = &txq->sdesc[eq->pidx];
2396 			txsd->txb_used += txinfo->txb_used;
2397 			txsd->hdls_used += txinfo->hdls_used;
2398 
2399 			/*
2400 			 * The txpkts chaining above has already placed `m` at
2401 			 * the end with b_next.  Keep the txsd notion of this
2402 			 * new tail up to date.
2403 			 */
2404 			ASSERT3P(txsd->mp_tail->b_next, ==, m);
2405 			txsd->mp_tail = m;
2406 
2407 			return (0);
2408 		}
2409 
2410 		/*
2411 		 * Couldn't coalesce m into txpkts.  The first order of business
2412 		 * is to send txpkts on its way.  Then we'll revisit m.
2413 		 */
2414 		write_txpkts_wr(txq, txpkts);
2415 	}
2416 
2417 	/*
2418 	 * Check if we can start a new coalesced tx work request with m as
2419 	 * the first packet in it.
2420 	 */
2421 
2422 	ASSERT(txpkts->npkt == 0);
2423 	ASSERT(txinfo->len < 65536);
2424 
2425 	flits = TXPKTS_WR_HDR + txinfo->nflits;
2426 	can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2427 	    flits <= eq->avail * 8 && flits <= TX_WR_FLITS;
2428 
2429 	if (can_coalesce == 0)
2430 		return (EINVAL);
2431 
2432 	/*
2433 	 * Start a fresh coalesced tx WR with m as the first frame in it.
2434 	 */
2435 	txpkts->tail = m;
2436 	txpkts->npkt = 1;
2437 	txpkts->nflits = flits;
2438 	txpkts->flitp = &eq->desc[eq->pidx].flit[2];
2439 	txpkts->plen = txinfo->len;
2440 
2441 	txsd = &txq->sdesc[eq->pidx];
2442 	txsd->mp_head = txsd->mp_tail = m;
2443 	txsd->txb_used = txinfo->txb_used;
2444 	txsd->hdls_used = txinfo->hdls_used;
2445 
2446 	return (0);
2447 }
2448 
2449 static inline void
t4_tx_incr_pending(struct sge_txq * txq,uint_t ndesc)2450 t4_tx_incr_pending(struct sge_txq *txq, uint_t ndesc)
2451 {
2452 	struct sge_eq *eq = &txq->eq;
2453 
2454 	TXQ_LOCK_ASSERT_OWNED(txq);
2455 	ASSERT3U(ndesc, !=, 0);
2456 	ASSERT3U(eq->avail, >=, ndesc);
2457 
2458 	eq->pending += ndesc;
2459 	eq->avail -= ndesc;
2460 	eq->pidx += ndesc;
2461 	if (eq->pidx >= eq->cap) {
2462 		eq->pidx -= eq->cap;
2463 	}
2464 }
2465 
2466 /*
2467  * Note that write_txpkts_wr can never run out of hardware descriptors (but
2468  * write_txpkt_wr can).  add_to_txpkts ensures that a frame is accepted for
2469  * coalescing only if sufficient hardware descriptors are available.
2470  */
2471 static void
write_txpkts_wr(struct sge_txq * txq,struct txpkts * txpkts)2472 write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts)
2473 {
2474 	struct sge_eq *eq = &txq->eq;
2475 	struct fw_eth_tx_pkts_wr *wr;
2476 	struct tx_sdesc *txsd;
2477 	uint32_t ctrl;
2478 	uint16_t ndesc;
2479 
2480 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2481 
2482 	ndesc = howmany(txpkts->nflits, 8);
2483 
2484 	wr = (void *)&eq->desc[eq->pidx];
2485 	wr->op_pkd = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR) |
2486 	    V_FW_WR_IMMDLEN(0)); /* immdlen does not matter in this WR */
2487 	ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2));
2488 	if (eq->avail == ndesc)
2489 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2490 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2491 	wr->plen = cpu_to_be16(txpkts->plen);
2492 	wr->npkt = txpkts->npkt;
2493 	wr->r3 = wr->type = 0;
2494 
2495 	/* Everything else already written */
2496 
2497 	txsd = &txq->sdesc[eq->pidx];
2498 	txsd->desc_used = ndesc;
2499 
2500 	txq->txb_used += txsd->txb_used / TXB_CHUNK;
2501 	txq->hdl_used += txsd->hdls_used;
2502 
2503 	t4_tx_incr_pending(txq, ndesc);
2504 
2505 	txq->txpkts_pkts += txpkts->npkt;
2506 	txq->txpkts_wrs++;
2507 	txpkts->npkt = 0;	/* emptied */
2508 }
2509 
2510 typedef enum {
2511 	COS_SUCCESS,	/* ctrl flit contains proper bits for csum offload */
2512 	COS_IGNORE,	/* no csum offload requested */
2513 	COS_FAIL,	/* csum offload requested, but pkt data missing */
2514 } csum_offload_status_t;
2515 /*
2516  * Build a ctrl1 flit for checksum offload in CPL_TX_PKT_XT command
2517  */
2518 static csum_offload_status_t
csum_to_ctrl(const struct txinfo * txinfo,uint32_t chip_version,uint64_t * ctrlp)2519 csum_to_ctrl(const struct txinfo *txinfo, uint32_t chip_version,
2520     uint64_t *ctrlp)
2521 {
2522 	const mac_ether_offload_info_t *meoi = &txinfo->meoi;
2523 	const uint32_t tx_flags = txinfo->flags;
2524 	const boolean_t needs_l3_csum = (tx_flags & HW_LSO) != 0 ||
2525 	    (tx_flags & HCK_IPV4_HDRCKSUM) != 0;
2526 	const boolean_t needs_l4_csum = (tx_flags & HW_LSO) != 0 ||
2527 	    (tx_flags & (HCK_FULLCKSUM | HCK_PARTIALCKSUM)) != 0;
2528 
2529 	/*
2530 	 * Default to disabling any checksumming both for cases where it is not
2531 	 * requested, but also if we cannot appropriately interrogate the
2532 	 * required information from the packet.
2533 	 */
2534 	uint64_t ctrl = F_TXPKT_L4CSUM_DIS | F_TXPKT_IPCSUM_DIS;
2535 	if (!needs_l3_csum && !needs_l4_csum) {
2536 		*ctrlp = ctrl;
2537 		return (COS_IGNORE);
2538 	}
2539 
2540 	if (needs_l3_csum) {
2541 		/* Only IPv4 checksums are supported (for L3) */
2542 		if ((meoi->meoi_flags & MEOI_L3INFO_SET) == 0 ||
2543 		    meoi->meoi_l3proto != ETHERTYPE_IP) {
2544 			*ctrlp = ctrl;
2545 			return (COS_FAIL);
2546 		}
2547 		ctrl &= ~F_TXPKT_IPCSUM_DIS;
2548 	}
2549 
2550 	if (needs_l4_csum) {
2551 		/*
2552 		 * We need at least all of the L3 header to make decisions about
2553 		 * the contained L4 protocol.  If not all of the L4 information
2554 		 * is present, we will leave it to the NIC to checksum all it is
2555 		 * able to.
2556 		 */
2557 		if ((meoi->meoi_flags & MEOI_L3INFO_SET) == 0) {
2558 			*ctrlp = ctrl;
2559 			return (COS_FAIL);
2560 		}
2561 
2562 		/*
2563 		 * Since we are parsing the packet anyways, make the checksum
2564 		 * decision based on the L4 protocol, rather than using the
2565 		 * Generic TCP/UDP checksum using start & end offsets in the
2566 		 * packet (like requested with PARTIALCKSUM).
2567 		 */
2568 		int csum_type = -1;
2569 		if (meoi->meoi_l3proto == ETHERTYPE_IP &&
2570 		    meoi->meoi_l4proto == IPPROTO_TCP) {
2571 			csum_type = TX_CSUM_TCPIP;
2572 		} else if (meoi->meoi_l3proto == ETHERTYPE_IPV6 &&
2573 		    meoi->meoi_l4proto == IPPROTO_TCP) {
2574 			csum_type = TX_CSUM_TCPIP6;
2575 		} else if (meoi->meoi_l3proto == ETHERTYPE_IP &&
2576 		    meoi->meoi_l4proto == IPPROTO_UDP) {
2577 			csum_type = TX_CSUM_UDPIP;
2578 		} else if (meoi->meoi_l3proto == ETHERTYPE_IPV6 &&
2579 		    meoi->meoi_l4proto == IPPROTO_UDP) {
2580 			csum_type = TX_CSUM_UDPIP6;
2581 		} else {
2582 			*ctrlp = ctrl;
2583 			return (COS_FAIL);
2584 		}
2585 
2586 		ASSERT(csum_type != -1);
2587 		ctrl &= ~F_TXPKT_L4CSUM_DIS;
2588 		ctrl |= V_TXPKT_CSUM_TYPE(csum_type);
2589 	}
2590 
2591 	if ((ctrl & F_TXPKT_IPCSUM_DIS) == 0 &&
2592 	    (ctrl & F_TXPKT_L4CSUM_DIS) != 0) {
2593 		/*
2594 		 * If only the IPv4 checksum is requested, we need to set an
2595 		 * appropriate type in the command for it.
2596 		 */
2597 		ctrl |= V_TXPKT_CSUM_TYPE(TX_CSUM_IP);
2598 	}
2599 
2600 	ASSERT(ctrl != (F_TXPKT_L4CSUM_DIS | F_TXPKT_IPCSUM_DIS));
2601 
2602 	/*
2603 	 * Fill in the requisite L2/L3 header length data.
2604 	 *
2605 	 * The Ethernet header length is recorded as 'size - 14 bytes'
2606 	 */
2607 	const uint8_t eth_len = meoi->meoi_l2hlen - 14;
2608 	if (chip_version >= CHELSIO_T6) {
2609 		ctrl |= V_T6_TXPKT_ETHHDR_LEN(eth_len);
2610 	} else {
2611 		ctrl |= V_TXPKT_ETHHDR_LEN(eth_len);
2612 	}
2613 	ctrl |= V_TXPKT_IPHDR_LEN(meoi->meoi_l3hlen);
2614 
2615 	*ctrlp = ctrl;
2616 	return (COS_SUCCESS);
2617 }
2618 
2619 static int
write_txpkt_wr(struct port_info * pi,struct sge_txq * txq,mblk_t * m,struct txinfo * txinfo)2620 write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
2621     struct txinfo *txinfo)
2622 {
2623 	struct sge_eq *eq = &txq->eq;
2624 	struct fw_eth_tx_pkt_wr *wr;
2625 	struct cpl_tx_pkt_core *cpl;
2626 	uint32_t ctrl;	/* used in many unrelated places */
2627 	uint64_t ctrl1;
2628 	int nflits, ndesc;
2629 	struct tx_sdesc *txsd;
2630 	caddr_t dst;
2631 	const mac_ether_offload_info_t *meoi = &txinfo->meoi;
2632 
2633 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2634 
2635 	/*
2636 	 * Do we have enough flits to send this frame out?
2637 	 */
2638 	ctrl = sizeof (struct cpl_tx_pkt_core);
2639 	if (txinfo->flags & HW_LSO) {
2640 		nflits = TXPKT_LSO_WR_HDR;
2641 		ctrl += sizeof (struct cpl_tx_pkt_lso_core);
2642 	} else {
2643 		nflits = TXPKT_WR_HDR;
2644 	}
2645 	if (txinfo->nsegs > 0)
2646 		nflits += txinfo->nflits;
2647 	else {
2648 		nflits += howmany(txinfo->len, 8);
2649 		ctrl += txinfo->len;
2650 	}
2651 	ndesc = howmany(nflits, 8);
2652 	if (ndesc > eq->avail)
2653 		return (ENOMEM);
2654 
2655 	/* Firmware work request header */
2656 	wr = (void *)&eq->desc[eq->pidx];
2657 	wr->op_immdlen = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
2658 	    V_FW_WR_IMMDLEN(ctrl));
2659 	ctrl = V_FW_WR_LEN16(howmany(nflits, 2));
2660 	if (eq->avail == ndesc)
2661 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2662 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2663 	wr->r3 = 0;
2664 
2665 	if (txinfo->flags & HW_LSO &&
2666 	    (meoi->meoi_flags & MEOI_L4INFO_SET) != 0 &&
2667 	    meoi->meoi_l4proto == IPPROTO_TCP) {
2668 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
2669 
2670 		ctrl = V_LSO_OPCODE((u32)CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
2671 		    F_LSO_LAST_SLICE;
2672 
2673 		if (meoi->meoi_l2hlen > sizeof (struct ether_header)) {
2674 			/*
2675 			 * This presently assumes a standard VLAN header,
2676 			 * without support for Q-in-Q.
2677 			 */
2678 			ctrl |= V_LSO_ETHHDR_LEN(1);
2679 		}
2680 
2681 		switch (meoi->meoi_l3proto) {
2682 		case ETHERTYPE_IPV6:
2683 			ctrl |= F_LSO_IPV6;
2684 			/* FALLTHROUGH */
2685 		case ETHERTYPE_IP:
2686 			ctrl |= V_LSO_IPHDR_LEN(meoi->meoi_l3hlen / 4);
2687 			break;
2688 		default:
2689 			break;
2690 		}
2691 
2692 		ctrl |= V_LSO_TCPHDR_LEN(meoi->meoi_l4hlen / 4);
2693 
2694 		lso->lso_ctrl = cpu_to_be32(ctrl);
2695 		lso->ipid_ofst = cpu_to_be16(0);
2696 		lso->mss = cpu_to_be16(txinfo->mss);
2697 		lso->seqno_offset = cpu_to_be32(0);
2698 		if (t4_cver_eq(pi->adapter, CHELSIO_T4))
2699 			lso->len = cpu_to_be32(txinfo->len);
2700 		else
2701 			lso->len = cpu_to_be32(V_LSO_T5_XFER_SIZE(txinfo->len));
2702 
2703 		cpl = (void *)(lso + 1);
2704 
2705 		txq->tso_wrs++;
2706 	} else {
2707 		cpl = (void *)(wr + 1);
2708 	}
2709 
2710 	/* Checksum offload */
2711 	switch (csum_to_ctrl(txinfo,
2712 	    CHELSIO_CHIP_VERSION(pi->adapter->params.chip), &ctrl1)) {
2713 	case COS_SUCCESS:
2714 		txq->txcsum++;
2715 		break;
2716 	case COS_FAIL:
2717 		/*
2718 		 * Packet will be going out with checksums which are probably
2719 		 * wrong but there is little we can do now.
2720 		 */
2721 		txq->csum_failed++;
2722 		break;
2723 	default:
2724 		break;
2725 	}
2726 
2727 	/* CPL header */
2728 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
2729 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
2730 	cpl->pack = 0;
2731 	cpl->len = cpu_to_be16(txinfo->len);
2732 	cpl->ctrl1 = cpu_to_be64(ctrl1);
2733 
2734 	/* Software descriptor */
2735 	txsd = &txq->sdesc[eq->pidx];
2736 	txsd->mp_head = txsd->mp_tail = m;
2737 	txsd->txb_used = txinfo->txb_used;
2738 	txsd->hdls_used = txinfo->hdls_used;
2739 	txsd->desc_used = ndesc;
2740 
2741 	txq->txb_used += txinfo->txb_used / TXB_CHUNK;
2742 	txq->hdl_used += txinfo->hdls_used;
2743 
2744 	t4_tx_incr_pending(txq, ndesc);
2745 
2746 	/* SGL */
2747 	dst = (void *)(cpl + 1);
2748 	if (txinfo->nsegs > 0) {
2749 		txq->sgl_wrs++;
2750 		copy_to_txd(eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
2751 
2752 		/* Need to zero-pad to a 16 byte boundary if not on one */
2753 		if ((uintptr_t)dst & 0xf)
2754 			*(uint64_t *)dst = 0;
2755 
2756 	} else {
2757 		txq->imm_wrs++;
2758 #ifdef DEBUG
2759 		ctrl = txinfo->len;
2760 #endif
2761 		for (; m; m = m->b_cont) {
2762 			copy_to_txd(eq, (void *)m->b_rptr, &dst, MBLKL(m));
2763 #ifdef DEBUG
2764 			ctrl -= MBLKL(m);
2765 #endif
2766 		}
2767 		ASSERT(ctrl == 0);
2768 	}
2769 
2770 	txq->txpkt_wrs++;
2771 	return (0);
2772 }
2773 
2774 static void
t4_write_flush_wr(struct sge_txq * txq)2775 t4_write_flush_wr(struct sge_txq *txq)
2776 {
2777 	struct sge_eq *eq = &txq->eq;
2778 
2779 	EQ_LOCK_ASSERT_OWNED(eq);
2780 	ASSERT(eq->avail > 0);
2781 
2782 	const struct fw_eq_flush_wr wr = {
2783 		.opcode = FW_EQ_FLUSH_WR,
2784 		.equiq_to_len16 = BE_32(
2785 		    V_FW_WR_LEN16(sizeof (struct fw_eq_flush_wr) / 16) |
2786 		    F_FW_WR_EQUEQ | F_FW_WR_EQUIQ),
2787 	};
2788 	*(struct fw_eq_flush_wr *)&eq->desc[eq->pidx] = wr;
2789 
2790 	const struct tx_sdesc txsd = {
2791 		.mp_head = NULL,
2792 		.mp_tail = NULL,
2793 		.txb_used = 0,
2794 		.hdls_used = 0,
2795 		.desc_used = 1,
2796 	};
2797 	txq->sdesc[eq->pidx] = txsd;
2798 
2799 	t4_tx_incr_pending(txq, 1);
2800 }
2801 
2802 static inline void
write_ulp_cpl_sgl(struct port_info * pi,struct sge_txq * txq,struct txpkts * txpkts,struct txinfo * txinfo)2803 write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
2804     struct txpkts *txpkts, struct txinfo *txinfo)
2805 {
2806 	struct ulp_txpkt *ulpmc;
2807 	struct ulptx_idata *ulpsc;
2808 	struct cpl_tx_pkt_core *cpl;
2809 	uintptr_t flitp, start, end;
2810 	uint64_t ctrl;
2811 	caddr_t dst;
2812 
2813 	ASSERT(txpkts->npkt > 0);
2814 
2815 	start = (uintptr_t)txq->eq.desc;
2816 	end = (uintptr_t)txq->eq.spg;
2817 
2818 	/* Checksum offload */
2819 	switch (csum_to_ctrl(txinfo,
2820 	    CHELSIO_CHIP_VERSION(pi->adapter->params.chip), &ctrl)) {
2821 	case COS_SUCCESS:
2822 		txq->txcsum++;
2823 		break;
2824 	case COS_FAIL:
2825 		/*
2826 		 * Packet will be going out with checksums which are probably
2827 		 * wrong but there is little we can do now.
2828 		 */
2829 		txq->csum_failed++;
2830 		break;
2831 	default:
2832 		break;
2833 	}
2834 
2835 	/*
2836 	 * The previous packet's SGL must have ended at a 16 byte boundary (this
2837 	 * is required by the firmware/hardware).  It follows that flitp cannot
2838 	 * wrap around between the ULPTX master command and ULPTX subcommand (8
2839 	 * bytes each), and that it can not wrap around in the middle of the
2840 	 * cpl_tx_pkt_core either.
2841 	 */
2842 	flitp = (uintptr_t)txpkts->flitp;
2843 	ASSERT((flitp & 0xf) == 0);
2844 
2845 	/* ULP master command */
2846 	ulpmc = (void *)flitp;
2847 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
2848 	ulpmc->len = htonl(howmany(sizeof (*ulpmc) + sizeof (*ulpsc) +
2849 	    sizeof (*cpl) + 8 * txinfo->nflits, 16));
2850 
2851 	/* ULP subcommand */
2852 	ulpsc = (void *)(ulpmc + 1);
2853 	ulpsc->cmd_more = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) |
2854 	    F_ULP_TX_SC_MORE);
2855 	ulpsc->len = cpu_to_be32(sizeof (struct cpl_tx_pkt_core));
2856 
2857 	flitp += sizeof (*ulpmc) + sizeof (*ulpsc);
2858 	if (flitp == end)
2859 		flitp = start;
2860 
2861 	/* CPL_TX_PKT_XT */
2862 	cpl = (void *)flitp;
2863 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
2864 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
2865 	cpl->pack = 0;
2866 	cpl->len = cpu_to_be16(txinfo->len);
2867 	cpl->ctrl1 = cpu_to_be64(ctrl);
2868 
2869 	flitp += sizeof (*cpl);
2870 	if (flitp == end)
2871 		flitp = start;
2872 
2873 	/* SGL for this frame */
2874 	dst = (caddr_t)flitp;
2875 	copy_to_txd(&txq->eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
2876 	flitp = (uintptr_t)dst;
2877 
2878 	/* Zero pad and advance to a 16 byte boundary if not already at one. */
2879 	if (flitp & 0xf) {
2880 
2881 		/* no matter what, flitp should be on an 8 byte boundary */
2882 		ASSERT((flitp & 0x7) == 0);
2883 
2884 		*(uint64_t *)flitp = 0;
2885 		flitp += sizeof (uint64_t);
2886 		txpkts->nflits++;
2887 	}
2888 
2889 	if (flitp == end)
2890 		flitp = start;
2891 
2892 	txpkts->flitp = (void *)flitp;
2893 }
2894 
2895 static inline void
copy_to_txd(struct sge_eq * eq,caddr_t from,caddr_t * to,int len)2896 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
2897 {
2898 	if ((uintptr_t)(*to) + len <= (uintptr_t)eq->spg) {
2899 		bcopy(from, *to, len);
2900 		(*to) += len;
2901 	} else {
2902 		int portion = (uintptr_t)eq->spg - (uintptr_t)(*to);
2903 
2904 		bcopy(from, *to, portion);
2905 		from += portion;
2906 		portion = len - portion;	/* remaining */
2907 		bcopy(from, (void *)eq->desc, portion);
2908 		(*to) = (caddr_t)eq->desc + portion;
2909 	}
2910 }
2911 
2912 static void
t4_tx_ring_db(struct sge_txq * txq)2913 t4_tx_ring_db(struct sge_txq *txq)
2914 {
2915 	struct sge_eq *eq = &txq->eq;
2916 	struct adapter *sc = txq->port->adapter;
2917 	int val, db_mode;
2918 	t4_doorbells_t db = eq->doorbells;
2919 
2920 	EQ_LOCK_ASSERT_OWNED(eq);
2921 
2922 	if (eq->pending > 1)
2923 		db &= ~DOORBELL_WCWR;
2924 
2925 	if (eq->pending > eq->pidx) {
2926 		int offset = eq->cap - (eq->pending - eq->pidx);
2927 
2928 		/* pidx has wrapped around since last doorbell */
2929 
2930 		(void) ddi_dma_sync(eq->desc_dhdl,
2931 		    offset * sizeof (struct tx_desc), 0,
2932 		    DDI_DMA_SYNC_FORDEV);
2933 		(void) ddi_dma_sync(eq->desc_dhdl,
2934 		    0, eq->pidx * sizeof (struct tx_desc),
2935 		    DDI_DMA_SYNC_FORDEV);
2936 	} else if (eq->pending > 0) {
2937 		(void) ddi_dma_sync(eq->desc_dhdl,
2938 		    (eq->pidx - eq->pending) * sizeof (struct tx_desc),
2939 		    eq->pending * sizeof (struct tx_desc),
2940 		    DDI_DMA_SYNC_FORDEV);
2941 	}
2942 
2943 	membar_producer();
2944 
2945 	if (t4_cver_eq(sc, CHELSIO_T4))
2946 		val = V_PIDX(eq->pending);
2947 	else
2948 		val = V_PIDX_T5(eq->pending);
2949 
2950 	db_mode = (1 << (ffs(db) - 1));
2951 	switch (db_mode) {
2952 		case DOORBELL_WCWR: {
2953 			/*
2954 			 * Queues whose 128B doorbell segment fits in
2955 			 * the page do not use relative qid
2956 			 * (udb_qid is always 0).  Only queues with
2957 			 * doorbell segments can do WCWR.
2958 			 */
2959 			ASSERT(eq->udb_qid == 0 && eq->pending == 1);
2960 
2961 			const uint_t desc_idx =
2962 			    eq->pidx != 0 ? eq->pidx - 1 : eq->cap - 1;
2963 			uint64_t *src = (uint64_t *)&eq->desc[desc_idx];
2964 			volatile uint64_t *dst =
2965 			    (uint64_t *)(eq->udb + UDBS_WR_OFFSET);
2966 
2967 			/* Copy the 8 flits of the TX descriptor to the DB */
2968 			const uint_t flit_count =
2969 			    sizeof (struct tx_desc) / sizeof (uint64_t);
2970 			for (uint_t i = 0; i < flit_count; i++) {
2971 				/*
2972 				 * Perform the copy directly through the BAR
2973 				 * mapping, rather than using ddi_put64().
2974 				 *
2975 				 * The latter was found to impose a significant
2976 				 * performance burden when called in this loop.
2977 				 */
2978 				dst[i] = src[i];
2979 			}
2980 
2981 			membar_producer();
2982 			break;
2983 		}
2984 
2985 		case DOORBELL_UDB:
2986 		case DOORBELL_UDBWC:
2987 			ddi_put32(sc->bar2_hdl,
2988 			    (uint32_t *)(eq->udb + UDBS_DB_OFFSET),
2989 			    LE_32(V_QID(eq->udb_qid) | val));
2990 			membar_producer();
2991 			break;
2992 
2993 		case DOORBELL_KDB:
2994 			t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
2995 			    V_QID(eq->cntxt_id) | val);
2996 			break;
2997 	}
2998 
2999 	eq->pending = 0;
3000 }
3001 
3002 /*
3003  * Reclaim consumed descriptors from egress queue.  This will be capped at an
3004  * upper bound of `howmany`.  The corresponding mblks will be freed inline,
3005  * unless a non-NULL `defer_freemp` is provided, in which case the to-be-freed
3006  * mblk chain will be provided to the caller.
3007  *
3008  * Returns the number of descriptors which underwent reclamation.
3009  */
3010 static uint_t
t4_tx_reclaim_descs(struct sge_txq * txq,uint_t howmany,mblk_t ** defer_freemp)3011 t4_tx_reclaim_descs(struct sge_txq *txq, uint_t howmany, mblk_t **defer_freemp)
3012 {
3013 	struct sge_eq *eq = &txq->eq;
3014 
3015 	EQ_LOCK_ASSERT_OWNED(eq);
3016 
3017 	const uint_t cur_cidx = BE_16(eq->spg->cidx);
3018 	const uint_t reclaim_avail = (cur_cidx >= eq->cidx) ?
3019 	    (cur_cidx - eq->cidx) : (cur_cidx + eq->cap - eq->cidx);
3020 
3021 	if (reclaim_avail == 0) {
3022 		return (0);
3023 	}
3024 
3025 	uint_t txb_freed = 0, hdl_freed = 0, reclaimed = 0;
3026 	do {
3027 		struct tx_sdesc *txsd = &txq->sdesc[eq->cidx];
3028 		const uint_t ndesc = txsd->desc_used;
3029 
3030 		/* Firmware doesn't return "partial" credits. */
3031 		ASSERT3U(reclaimed + ndesc, <=, reclaim_avail);
3032 
3033 		if (txsd->mp_head != NULL) {
3034 			/*
3035 			 * Even when packet content fits entirely in immediate
3036 			 * buffer, the mblk is kept around until the
3037 			 * transmission completes.
3038 			 */
3039 			if (defer_freemp != NULL) {
3040 				/*
3041 				 * Append the mblk chain from this descriptor
3042 				 * onto the end of the defer list.
3043 				 *
3044 				 * In the case that this is the first mblk we
3045 				 * have processed, the below assignment will
3046 				 * communicate the head of the chain to the
3047 				 * caller.
3048 				 */
3049 				*defer_freemp = txsd->mp_head;
3050 				defer_freemp = &txsd->mp_tail->b_next;
3051 			} else {
3052 				freemsgchain(txsd->mp_head);
3053 			}
3054 			txsd->mp_head = txsd->mp_tail = NULL;
3055 		} else {
3056 			/*
3057 			 * If mblk is NULL, this has to be the software
3058 			 * descriptor for a credit flush work request.
3059 			 */
3060 			ASSERT0(txsd->txb_used);
3061 			ASSERT0(txsd->hdls_used);
3062 			ASSERT3U(ndesc, ==, 1);
3063 		}
3064 
3065 		txb_freed += txsd->txb_used;
3066 		hdl_freed += txsd->hdls_used;
3067 		reclaimed += ndesc;
3068 
3069 		eq->cidx += ndesc;
3070 		if (eq->cidx >= eq->cap) {
3071 			eq->cidx -= eq->cap;
3072 		}
3073 	} while (reclaimed < reclaim_avail && reclaimed < howmany);
3074 
3075 	eq->avail += reclaimed;
3076 	txq->txb_avail += txb_freed;
3077 	txq->tx_dhdl_avail += hdl_freed;
3078 
3079 	ASSERT3U(eq->avail, <, eq->cap);
3080 	ASSERT3U(txq->tx_dhdl_avail, <=, txq->tx_dhdl_total);
3081 
3082 	for (; hdl_freed; hdl_freed--) {
3083 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_cidx]);
3084 		if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
3085 			txq->tx_dhdl_cidx = 0;
3086 	}
3087 
3088 	return (reclaimed);
3089 }
3090 
3091 static int
t4_handle_cpl_msg(struct sge_iq * iq,const struct rss_header * rss,mblk_t * mp)3092 t4_handle_cpl_msg(struct sge_iq *iq, const struct rss_header *rss, mblk_t *mp)
3093 {
3094 	const uint8_t opcode = rss->opcode;
3095 
3096 	DTRACE_PROBE4(t4__cpl_msg, struct sge_iq *, iq, uint8_t, opcode,
3097 	    const struct rss_header *, rss, mblk_t *, mp);
3098 
3099 	switch (opcode) {
3100 	case CPL_FW4_MSG:
3101 	case CPL_FW6_MSG:
3102 		ASSERT3P(mp, ==, NULL);
3103 		return (t4_handle_fw_msg(iq, rss));
3104 	case CPL_SGE_EGR_UPDATE:
3105 		ASSERT3P(mp, ==, NULL);
3106 		t4_sge_egr_update(iq, rss);
3107 		return (0);
3108 	case CPL_RX_PKT:
3109 		return (t4_eth_rx(iq, rss, mp));
3110 	default:
3111 		cxgb_printf(iq->adapter->dip, CE_WARN,
3112 		    "unhandled CPL opcode 0x%02x", opcode);
3113 		if (mp != NULL) {
3114 			freemsg(mp);
3115 		}
3116 		return (0);
3117 	}
3118 }
3119 
3120 static int
t4_handle_fw_msg(struct sge_iq * iq,const struct rss_header * rss)3121 t4_handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss)
3122 {
3123 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
3124 	const uint8_t msg_type = cpl->type;
3125 	const struct rss_header *rss2;
3126 	struct adapter *sc = iq->adapter;
3127 
3128 	DTRACE_PROBE3(t4__fw_msg, struct sge_iq *, iq, uint8_t, msg_type,
3129 	    const struct rss_header *, rss);
3130 
3131 	switch (msg_type) {
3132 	case FW_TYPE_RSSCPL:	/* also synonym for FW6_TYPE_RSSCPL */
3133 		rss2 = (const struct rss_header *)&cpl->data[0];
3134 		return (t4_handle_cpl_msg(iq, rss2, NULL));
3135 	case FW6_TYPE_CMD_RPL:
3136 		return (t4_handle_fw_rpl(sc, &cpl->data[0]));
3137 	default:
3138 		cxgb_printf(sc->dip, CE_WARN,
3139 		    "unhandled fw_msg type 0x%02x", msg_type);
3140 		return (0);
3141 	}
3142 }
3143 
3144 static int
t4_eth_rx(struct sge_iq * iq,const struct rss_header * rss,mblk_t * m)3145 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
3146 {
3147 	bool csum_ok;
3148 	uint16_t err_vec;
3149 	struct sge_rxq *rxq = (void *)iq;
3150 	struct mblk_pair chain = {0};
3151 	struct adapter *sc = iq->adapter;
3152 	const struct cpl_rx_pkt *cpl = t4_rss_payload(rss);
3153 
3154 	m->b_rptr += sc->sge.pktshift;
3155 
3156 	/* Compressed error vector is enabled for T6 only */
3157 	if (sc->params.tp.rx_pkt_encap)
3158 		/* It is enabled only in T6 config file */
3159 		err_vec = G_T6_COMPR_RXERR_VEC(ntohs(cpl->err_vec));
3160 	else
3161 		err_vec = ntohs(cpl->err_vec);
3162 
3163 	csum_ok = cpl->csum_calc && !err_vec;
3164 	/* TODO: what about cpl->ip_frag? */
3165 	if (csum_ok && !cpl->ip_frag) {
3166 		mac_hcksum_set(m, 0, 0, 0, 0xffff,
3167 		    HCK_FULLCKSUM_OK | HCK_FULLCKSUM |
3168 		    HCK_IPV4_HDRCKSUM_OK);
3169 		rxq->rxcsum++;
3170 	}
3171 
3172 	/* Add to the chain that we'll send up */
3173 	if (chain.head != NULL)
3174 		chain.tail->b_next = m;
3175 	else
3176 		chain.head = m;
3177 	chain.tail = m;
3178 
3179 	t4_mac_rx(rxq->port, rxq, chain.head);
3180 
3181 	rxq->rxpkts++;
3182 	rxq->rxbytes  += be16_to_cpu(cpl->len);
3183 	return (0);
3184 }
3185 
3186 #define	FL_HW_IDX(idx)	((idx) >> 3)
3187 
3188 static inline void
ring_fl_db(struct adapter * sc,struct sge_fl * fl)3189 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3190 {
3191 	int desc_start, desc_last, ndesc;
3192 	uint32_t v = sc->params.arch.sge_fl_db;
3193 
3194 	ndesc = FL_HW_IDX(fl->pending);
3195 
3196 	/* Hold back one credit if pidx = cidx */
3197 	if (FL_HW_IDX(fl->pidx) == FL_HW_IDX(fl->cidx))
3198 		ndesc--;
3199 
3200 	/*
3201 	 * There are chances of ndesc modified above (to avoid pidx = cidx).
3202 	 * If there is nothing to post, return.
3203 	 */
3204 	if (ndesc <= 0)
3205 		return;
3206 
3207 	desc_last = FL_HW_IDX(fl->pidx);
3208 
3209 	if (fl->pidx < fl->pending) {
3210 		/* There was a wrap */
3211 		desc_start = FL_HW_IDX(fl->pidx + fl->cap - fl->pending);
3212 
3213 		/* From desc_start to the end of list */
3214 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE, 0,
3215 		    DDI_DMA_SYNC_FORDEV);
3216 
3217 		/* From start of list to the desc_last */
3218 		if (desc_last != 0)
3219 			(void) ddi_dma_sync(fl->dhdl, 0, desc_last *
3220 			    RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3221 	} else {
3222 		/* There was no wrap, sync from start_desc to last_desc */
3223 		desc_start = FL_HW_IDX(fl->pidx - fl->pending);
3224 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE,
3225 		    ndesc * RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3226 	}
3227 
3228 	if (t4_cver_eq(sc, CHELSIO_T4))
3229 		v |= V_PIDX(ndesc);
3230 	else
3231 		v |= V_PIDX_T5(ndesc);
3232 	v |= V_QID(fl->cntxt_id) | V_PIDX(ndesc);
3233 
3234 	membar_producer();
3235 
3236 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3237 
3238 	/*
3239 	 * Update pending count:
3240 	 * Deduct the number of descriptors posted
3241 	 */
3242 	fl->pending -= ndesc * 8;
3243 }
3244 
3245 static void
t4_sge_egr_update(struct sge_iq * iq,const struct rss_header * rss)3246 t4_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss)
3247 {
3248 	struct adapter *sc = iq->adapter;
3249 	const struct cpl_sge_egr_update *cpl = t4_rss_payload(rss);
3250 	const uint_t qid = G_EGR_QID(BE_32(cpl->opcode_qid));
3251 	struct sge_txq *txq = (struct sge_txq *)(*t4_eqmap_slot(sc, qid));
3252 	struct sge_eq *eq = &txq->eq;
3253 
3254 	/*
3255 	 * Get a "live" snapshot of the flags and PIDX state from the TXQ.
3256 	 *
3257 	 * This is done without the protection of the TXQ/EQ lock, since the
3258 	 * gathered information is used to avoid contending on that lock for the
3259 	 * reclaim.
3260 	 */
3261 	membar_consumer();
3262 	const uint16_t live_pidx = BE_16(eq->pidx);
3263 	const t4_eq_flags_t live_flags = eq->flags;
3264 
3265 	if ((live_flags & EQ_CORKED) == 0 &&
3266 	    (cpl->pidx != cpl->cidx || live_pidx != cpl->cidx)) {
3267 		/*
3268 		 * A reclaim of the ring can be skipped if:
3269 		 *
3270 		 * 1. The EQ is not in the "corked" state, where it was unable
3271 		 *    allocate descriptors (or memory) while attempting to place
3272 		 *    a packet in the TXQ.
3273 		 *
3274 		 * 2. There are additional transmit descriptors in the EQ which
3275 		 *    will trigger a subsequent SGE_EGR_UPDATE notification.
3276 		 *
3277 		 * When those conditions are met, it is safe to skip performing
3278 		 * a reclaim here, reducing the chance that we contend with
3279 		 * other transmission activity against the TXQ.
3280 		 */
3281 		DTRACE_PROBE2(t4__elide__reclaim,
3282 		    struct sge_txq *, txq, struct cpl_sge_egr_update *, cpl);
3283 		return;
3284 	}
3285 
3286 	mblk_t *freemp = NULL;
3287 	bool do_mac_update = false;
3288 
3289 	TXQ_LOCK(txq);
3290 	(void) t4_tx_reclaim_descs(txq, eq->qsize, &freemp);
3291 	if (eq->flags & EQ_CORKED && eq->avail != 0) {
3292 		do_mac_update = true;
3293 		eq->flags &= ~EQ_CORKED;
3294 	}
3295 	TXQ_UNLOCK(txq);
3296 
3297 	freemsgchain(freemp);
3298 	if (do_mac_update) {
3299 		t4_mac_tx_update(txq->port, txq);
3300 	}
3301 }
3302 
3303 #define	KS_UINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_ULONG)
3304 #define	KS_CINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_CHAR)
3305 #define	KS_U_SET(x, y)	kstatp->x.value.ul = (y)
3306 #define	KS_U_FROM(x, y)	kstatp->x.value.ul = (y)->x
3307 #define	KS_C_SET(x, ...)	\
3308 			(void) snprintf(kstatp->x.value.c, 16,  __VA_ARGS__)
3309 
3310 /*
3311  * cxgbe:X:config
3312  */
3313 struct cxgbe_port_config_kstats {
3314 	kstat_named_t idx;
3315 	kstat_named_t nrxq;
3316 	kstat_named_t ntxq;
3317 	kstat_named_t first_rxq;
3318 	kstat_named_t first_txq;
3319 	kstat_named_t controller;
3320 	kstat_named_t factory_mac_address;
3321 };
3322 
3323 /*
3324  * cxgbe:X:info
3325  */
3326 struct cxgbe_port_info_kstats {
3327 	kstat_named_t transceiver;
3328 	kstat_named_t rx_ovflow0;
3329 	kstat_named_t rx_ovflow1;
3330 	kstat_named_t rx_ovflow2;
3331 	kstat_named_t rx_ovflow3;
3332 	kstat_named_t rx_trunc0;
3333 	kstat_named_t rx_trunc1;
3334 	kstat_named_t rx_trunc2;
3335 	kstat_named_t rx_trunc3;
3336 	kstat_named_t tx_pause;
3337 	kstat_named_t rx_pause;
3338 };
3339 
3340 static kstat_t *
setup_port_config_kstats(struct port_info * pi)3341 setup_port_config_kstats(struct port_info *pi)
3342 {
3343 	kstat_t *ksp;
3344 	struct cxgbe_port_config_kstats *kstatp;
3345 	int ndata;
3346 	dev_info_t *pdip = ddi_get_parent(pi->dip);
3347 	uint8_t *ma = &pi->hw_addr[0];
3348 
3349 	ndata = sizeof (struct cxgbe_port_config_kstats) /
3350 	    sizeof (kstat_named_t);
3351 
3352 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "config",
3353 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3354 	if (ksp == NULL) {
3355 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3356 		return (NULL);
3357 	}
3358 
3359 	kstatp = (struct cxgbe_port_config_kstats *)ksp->ks_data;
3360 
3361 	KS_UINIT(idx);
3362 	KS_UINIT(nrxq);
3363 	KS_UINIT(ntxq);
3364 	KS_UINIT(first_rxq);
3365 	KS_UINIT(first_txq);
3366 	KS_CINIT(controller);
3367 	KS_CINIT(factory_mac_address);
3368 
3369 	KS_U_SET(idx, pi->port_id);
3370 	KS_U_SET(nrxq, pi->nrxq);
3371 	KS_U_SET(ntxq, pi->ntxq);
3372 	KS_U_SET(first_rxq, pi->first_rxq);
3373 	KS_U_SET(first_txq, pi->first_txq);
3374 	KS_C_SET(controller, "%s%d", ddi_driver_name(pdip),
3375 	    ddi_get_instance(pdip));
3376 	KS_C_SET(factory_mac_address, "%02X%02X%02X%02X%02X%02X",
3377 	    ma[0], ma[1], ma[2], ma[3], ma[4], ma[5]);
3378 
3379 	/* Do NOT set ksp->ks_update.  These kstats do not change. */
3380 
3381 	/* Install the kstat */
3382 	ksp->ks_private = (void *)pi;
3383 	kstat_install(ksp);
3384 
3385 	return (ksp);
3386 }
3387 
3388 static kstat_t *
setup_port_info_kstats(struct port_info * pi)3389 setup_port_info_kstats(struct port_info *pi)
3390 {
3391 	kstat_t *ksp;
3392 	struct cxgbe_port_info_kstats *kstatp;
3393 	int ndata;
3394 
3395 	ndata = sizeof (struct cxgbe_port_info_kstats) / sizeof (kstat_named_t);
3396 
3397 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "info",
3398 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3399 	if (ksp == NULL) {
3400 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3401 		return (NULL);
3402 	}
3403 
3404 	kstatp = (struct cxgbe_port_info_kstats *)ksp->ks_data;
3405 
3406 	KS_CINIT(transceiver);
3407 	KS_UINIT(rx_ovflow0);
3408 	KS_UINIT(rx_ovflow1);
3409 	KS_UINIT(rx_ovflow2);
3410 	KS_UINIT(rx_ovflow3);
3411 	KS_UINIT(rx_trunc0);
3412 	KS_UINIT(rx_trunc1);
3413 	KS_UINIT(rx_trunc2);
3414 	KS_UINIT(rx_trunc3);
3415 	KS_UINIT(tx_pause);
3416 	KS_UINIT(rx_pause);
3417 
3418 	/* Install the kstat */
3419 	ksp->ks_update = update_port_info_kstats;
3420 	ksp->ks_private = (void *)pi;
3421 	kstat_install(ksp);
3422 
3423 	return (ksp);
3424 }
3425 
3426 static int
update_port_info_kstats(kstat_t * ksp,int rw)3427 update_port_info_kstats(kstat_t *ksp, int rw)
3428 {
3429 	struct cxgbe_port_info_kstats *kstatp =
3430 	    (struct cxgbe_port_info_kstats *)ksp->ks_data;
3431 	struct port_info *pi = ksp->ks_private;
3432 	static const char *mod_str[] = { NULL, "LR", "SR", "ER", "TWINAX",
3433 	    "active TWINAX", "LRM" };
3434 	uint32_t bgmap;
3435 
3436 	if (rw == KSTAT_WRITE)
3437 		return (0);
3438 
3439 	if (pi->mod_type == FW_PORT_MOD_TYPE_NONE)
3440 		KS_C_SET(transceiver, "unplugged");
3441 	else if (pi->mod_type == FW_PORT_MOD_TYPE_UNKNOWN)
3442 		KS_C_SET(transceiver, "unknown");
3443 	else if (pi->mod_type == FW_PORT_MOD_TYPE_NOTSUPPORTED)
3444 		KS_C_SET(transceiver, "unsupported");
3445 	else if (pi->mod_type > 0 && pi->mod_type < ARRAY_SIZE(mod_str))
3446 		KS_C_SET(transceiver, "%s", mod_str[pi->mod_type]);
3447 	else
3448 		KS_C_SET(transceiver, "type %d", pi->mod_type);
3449 
3450 #define	GET_STAT(name) t4_read_reg64(pi->adapter, \
3451 	    PORT_REG(pi->port_id, A_MPS_PORT_STAT_##name##_L))
3452 #define	GET_STAT_COM(name) t4_read_reg64(pi->adapter, \
3453 	    A_MPS_STAT_##name##_L)
3454 
3455 	bgmap = G_NUMPORTS(t4_read_reg(pi->adapter, A_MPS_CMN_CTL));
3456 	if (bgmap == 0)
3457 		bgmap = (pi->port_id == 0) ? 0xf : 0;
3458 	else if (bgmap == 1)
3459 		bgmap = (pi->port_id < 2) ? (3 << (2 * pi->port_id)) : 0;
3460 	else
3461 		bgmap = 1;
3462 
3463 	KS_U_SET(rx_ovflow0, (bgmap & 1) ?
3464 	    GET_STAT_COM(RX_BG_0_MAC_DROP_FRAME) : 0);
3465 	KS_U_SET(rx_ovflow1, (bgmap & 2) ?
3466 	    GET_STAT_COM(RX_BG_1_MAC_DROP_FRAME) : 0);
3467 	KS_U_SET(rx_ovflow2, (bgmap & 4) ?
3468 	    GET_STAT_COM(RX_BG_2_MAC_DROP_FRAME) : 0);
3469 	KS_U_SET(rx_ovflow3, (bgmap & 8) ?
3470 	    GET_STAT_COM(RX_BG_3_MAC_DROP_FRAME) : 0);
3471 	KS_U_SET(rx_trunc0,  (bgmap & 1) ?
3472 	    GET_STAT_COM(RX_BG_0_MAC_TRUNC_FRAME) : 0);
3473 	KS_U_SET(rx_trunc1,  (bgmap & 2) ?
3474 	    GET_STAT_COM(RX_BG_1_MAC_TRUNC_FRAME) : 0);
3475 	KS_U_SET(rx_trunc2,  (bgmap & 4) ?
3476 	    GET_STAT_COM(RX_BG_2_MAC_TRUNC_FRAME) : 0);
3477 	KS_U_SET(rx_trunc3,  (bgmap & 8) ?
3478 	    GET_STAT_COM(RX_BG_3_MAC_TRUNC_FRAME) : 0);
3479 
3480 	KS_U_SET(tx_pause, GET_STAT(TX_PORT_PAUSE));
3481 	KS_U_SET(rx_pause, GET_STAT(RX_PORT_PAUSE));
3482 
3483 	return (0);
3484 
3485 }
3486 
3487 /*
3488  * cxgbe:X:rxqY
3489  */
3490 struct rxq_kstats {
3491 	kstat_named_t rxcsum;
3492 	kstat_named_t rxpkts;
3493 	kstat_named_t rxbytes;
3494 	kstat_named_t nomem;
3495 };
3496 
3497 static kstat_t *
setup_rxq_kstats(struct port_info * pi,struct sge_rxq * rxq,int idx)3498 setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq, int idx)
3499 {
3500 	struct kstat *ksp;
3501 	struct rxq_kstats *kstatp;
3502 	int ndata;
3503 	char str[16];
3504 
3505 	ndata = sizeof (struct rxq_kstats) / sizeof (kstat_named_t);
3506 	(void) snprintf(str, sizeof (str), "rxq%u", idx);
3507 
3508 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "rxq",
3509 	    KSTAT_TYPE_NAMED, ndata, 0);
3510 	if (ksp == NULL) {
3511 		cxgb_printf(pi->dip, CE_WARN,
3512 		    "%s: failed to initialize rxq kstats for queue %d.",
3513 		    __func__, idx);
3514 		return (NULL);
3515 	}
3516 
3517 	kstatp = (struct rxq_kstats *)ksp->ks_data;
3518 
3519 	KS_UINIT(rxcsum);
3520 	KS_UINIT(rxpkts);
3521 	KS_UINIT(rxbytes);
3522 	KS_UINIT(nomem);
3523 
3524 	ksp->ks_update = update_rxq_kstats;
3525 	ksp->ks_private = (void *)rxq;
3526 	kstat_install(ksp);
3527 
3528 	return (ksp);
3529 }
3530 
3531 static int
update_rxq_kstats(kstat_t * ksp,int rw)3532 update_rxq_kstats(kstat_t *ksp, int rw)
3533 {
3534 	struct rxq_kstats *kstatp = (struct rxq_kstats *)ksp->ks_data;
3535 	struct sge_rxq *rxq = ksp->ks_private;
3536 
3537 	if (rw == KSTAT_WRITE)
3538 		return (0);
3539 
3540 	KS_U_FROM(rxcsum, rxq);
3541 	KS_U_FROM(rxpkts, rxq);
3542 	KS_U_FROM(rxbytes, rxq);
3543 	KS_U_FROM(nomem, rxq);
3544 
3545 	return (0);
3546 }
3547 
3548 /*
3549  * cxgbe:X:txqY
3550  */
3551 struct txq_kstats {
3552 	kstat_named_t txcsum;
3553 	kstat_named_t tso_wrs;
3554 	kstat_named_t imm_wrs;
3555 	kstat_named_t sgl_wrs;
3556 	kstat_named_t txpkt_wrs;
3557 	kstat_named_t txpkts_wrs;
3558 	kstat_named_t txpkts_pkts;
3559 	kstat_named_t txb_used;
3560 	kstat_named_t hdl_used;
3561 	kstat_named_t txb_full;
3562 	kstat_named_t dma_hdl_failed;
3563 	kstat_named_t dma_map_failed;
3564 	kstat_named_t qfull;
3565 	kstat_named_t pullup_early;
3566 	kstat_named_t pullup_late;
3567 	kstat_named_t pullup_failed;
3568 	kstat_named_t csum_failed;
3569 };
3570 
3571 static kstat_t *
setup_txq_kstats(struct port_info * pi,struct sge_txq * txq,int idx)3572 setup_txq_kstats(struct port_info *pi, struct sge_txq *txq, int idx)
3573 {
3574 	struct kstat *ksp;
3575 	struct txq_kstats *kstatp;
3576 	int ndata;
3577 	char str[16];
3578 
3579 	ndata = sizeof (struct txq_kstats) / sizeof (kstat_named_t);
3580 	(void) snprintf(str, sizeof (str), "txq%u", idx);
3581 
3582 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "txq",
3583 	    KSTAT_TYPE_NAMED, ndata, 0);
3584 	if (ksp == NULL) {
3585 		cxgb_printf(pi->dip, CE_WARN,
3586 		    "%s: failed to initialize txq kstats for queue %d.",
3587 		    __func__, idx);
3588 		return (NULL);
3589 	}
3590 
3591 	kstatp = (struct txq_kstats *)ksp->ks_data;
3592 
3593 	KS_UINIT(txcsum);
3594 	KS_UINIT(tso_wrs);
3595 	KS_UINIT(imm_wrs);
3596 	KS_UINIT(sgl_wrs);
3597 	KS_UINIT(txpkt_wrs);
3598 	KS_UINIT(txpkts_wrs);
3599 	KS_UINIT(txpkts_pkts);
3600 	KS_UINIT(txb_used);
3601 	KS_UINIT(hdl_used);
3602 	KS_UINIT(txb_full);
3603 	KS_UINIT(dma_hdl_failed);
3604 	KS_UINIT(dma_map_failed);
3605 	KS_UINIT(qfull);
3606 	KS_UINIT(pullup_early);
3607 	KS_UINIT(pullup_late);
3608 	KS_UINIT(pullup_failed);
3609 	KS_UINIT(csum_failed);
3610 
3611 	ksp->ks_update = update_txq_kstats;
3612 	ksp->ks_private = (void *)txq;
3613 	kstat_install(ksp);
3614 
3615 	return (ksp);
3616 }
3617 
3618 static int
update_txq_kstats(kstat_t * ksp,int rw)3619 update_txq_kstats(kstat_t *ksp, int rw)
3620 {
3621 	struct txq_kstats *kstatp = (struct txq_kstats *)ksp->ks_data;
3622 	struct sge_txq *txq = ksp->ks_private;
3623 
3624 	if (rw == KSTAT_WRITE)
3625 		return (0);
3626 
3627 	KS_U_FROM(txcsum, txq);
3628 	KS_U_FROM(tso_wrs, txq);
3629 	KS_U_FROM(imm_wrs, txq);
3630 	KS_U_FROM(sgl_wrs, txq);
3631 	KS_U_FROM(txpkt_wrs, txq);
3632 	KS_U_FROM(txpkts_wrs, txq);
3633 	KS_U_FROM(txpkts_pkts, txq);
3634 	KS_U_FROM(txb_used, txq);
3635 	KS_U_FROM(hdl_used, txq);
3636 	KS_U_FROM(txb_full, txq);
3637 	KS_U_FROM(dma_hdl_failed, txq);
3638 	KS_U_FROM(dma_map_failed, txq);
3639 	KS_U_FROM(qfull, txq);
3640 	KS_U_FROM(pullup_early, txq);
3641 	KS_U_FROM(pullup_late, txq);
3642 	KS_U_FROM(pullup_failed, txq);
3643 	KS_U_FROM(csum_failed, txq);
3644 
3645 	return (0);
3646 }
3647 
3648 static int rxbuf_ctor(void *, void *, int);
3649 static void rxbuf_dtor(void *, void *);
3650 
3651 static kmem_cache_t *
rxbuf_cache_create(struct rxbuf_cache_params * p)3652 rxbuf_cache_create(struct rxbuf_cache_params *p)
3653 {
3654 	char name[32];
3655 
3656 	(void) snprintf(name, sizeof (name), "%s%d_rxbuf_cache",
3657 	    ddi_driver_name(p->dip), ddi_get_instance(p->dip));
3658 
3659 	return kmem_cache_create(name, sizeof (struct rxbuf), _CACHE_LINE_SIZE,
3660 	    rxbuf_ctor, rxbuf_dtor, NULL, p, NULL, 0);
3661 }
3662 
3663 /*
3664  * If ref_cnt is more than 1 then those many calls to rxbuf_free will
3665  * have to be made before the rxb is released back to the kmem_cache.
3666  */
3667 static struct rxbuf *
rxbuf_alloc(kmem_cache_t * cache,int kmflags,uint_t ref_cnt)3668 rxbuf_alloc(kmem_cache_t *cache, int kmflags, uint_t ref_cnt)
3669 {
3670 	struct rxbuf *rxb;
3671 
3672 	ASSERT(ref_cnt > 0);
3673 
3674 	rxb = kmem_cache_alloc(cache, kmflags);
3675 	if (rxb != NULL) {
3676 		rxb->ref_cnt = ref_cnt;
3677 		rxb->cache = cache;
3678 	}
3679 
3680 	return (rxb);
3681 }
3682 
3683 /*
3684  * This is normally called via the rxb's freefunc, when an mblk referencing the
3685  * rxb is freed.
3686  */
3687 static void
rxbuf_free(struct rxbuf * rxb)3688 rxbuf_free(struct rxbuf *rxb)
3689 {
3690 	if (atomic_dec_uint_nv(&rxb->ref_cnt) == 0)
3691 		kmem_cache_free(rxb->cache, rxb);
3692 }
3693 
3694 static int
rxbuf_ctor(void * arg1,void * arg2,int kmflag)3695 rxbuf_ctor(void *arg1, void *arg2, int kmflag)
3696 {
3697 	struct rxbuf *rxb = arg1;
3698 	struct rxbuf_cache_params *p = arg2;
3699 	size_t real_len;
3700 	ddi_dma_cookie_t cookie;
3701 	uint_t ccount = 0;
3702 	int (*callback)(caddr_t);
3703 	int rc = ENOMEM;
3704 
3705 	if ((kmflag & KM_NOSLEEP) != 0)
3706 		callback = DDI_DMA_DONTWAIT;
3707 	else
3708 		callback = DDI_DMA_SLEEP;
3709 
3710 	rc = ddi_dma_alloc_handle(p->dip, &p->dma_attr_rx, callback, 0,
3711 	    &rxb->dhdl);
3712 	if (rc != DDI_SUCCESS)
3713 		return (rc == DDI_DMA_BADATTR ? EINVAL : ENOMEM);
3714 
3715 	rc = ddi_dma_mem_alloc(rxb->dhdl, p->buf_size, &p->acc_attr_rx,
3716 	    DDI_DMA_STREAMING, callback, 0, &rxb->va, &real_len, &rxb->ahdl);
3717 	if (rc != DDI_SUCCESS) {
3718 		rc = ENOMEM;
3719 		goto fail1;
3720 	}
3721 
3722 	rc = ddi_dma_addr_bind_handle(rxb->dhdl, NULL, rxb->va, p->buf_size,
3723 	    DDI_DMA_READ | DDI_DMA_STREAMING, NULL, NULL, &cookie, &ccount);
3724 	if (rc != DDI_DMA_MAPPED) {
3725 		if (rc == DDI_DMA_INUSE)
3726 			rc = EBUSY;
3727 		else if (rc == DDI_DMA_TOOBIG)
3728 			rc = E2BIG;
3729 		else
3730 			rc = ENOMEM;
3731 		goto fail2;
3732 	}
3733 
3734 	if (ccount != 1) {
3735 		rc = E2BIG;
3736 		goto fail3;
3737 	}
3738 
3739 	rxb->ref_cnt = 0;
3740 	rxb->buf_size = p->buf_size;
3741 	rxb->freefunc.free_arg = (caddr_t)rxb;
3742 	rxb->freefunc.free_func = rxbuf_free;
3743 	rxb->ba = cookie.dmac_laddress;
3744 
3745 	return (0);
3746 
3747 fail3:	(void) ddi_dma_unbind_handle(rxb->dhdl);
3748 fail2:	ddi_dma_mem_free(&rxb->ahdl);
3749 fail1:	ddi_dma_free_handle(&rxb->dhdl);
3750 	return (rc);
3751 }
3752 
3753 static void
rxbuf_dtor(void * arg1,void * arg2)3754 rxbuf_dtor(void *arg1, void *arg2)
3755 {
3756 	struct rxbuf *rxb = arg1;
3757 
3758 	(void) ddi_dma_unbind_handle(rxb->dhdl);
3759 	ddi_dma_mem_free(&rxb->ahdl);
3760 	ddi_dma_free_handle(&rxb->dhdl);
3761 }
3762