xref: /illumos-gate/usr/src/uts/common/io/cxgbe/t4nex/t4_sge.c (revision ba52565b00cef6c84d9c58d122f646c8f30b4a38)
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 "version.h"
40 #include "common/common.h"
41 #include "common/t4_msg.h"
42 #include "common/t4_regs.h"
43 #include "common/t4_regs_values.h"
44 
45 /* TODO: Tune. */
46 int rx_buf_size = 8192;
47 int tx_copy_threshold = 256;
48 uint16_t rx_copy_threshold = 256;
49 
50 /* Used to track coalesced tx work request */
51 struct txpkts {
52 	mblk_t *tail;		/* head is in the software descriptor */
53 	uint64_t *flitp;	/* ptr to flit where next pkt should start */
54 	uint8_t npkt;		/* # of packets in this work request */
55 	uint8_t nflits;		/* # of flits used by this work request */
56 	uint16_t plen;		/* total payload (sum of all packets) */
57 };
58 
59 /* All information needed to tx a frame */
60 struct txinfo {
61 	uint32_t len;		/* Total length of frame */
62 	uint32_t flags;		/* Checksum and LSO flags */
63 	uint32_t mss;		/* MSS for LSO */
64 	uint8_t nsegs;		/* # of segments in the SGL, 0 means imm. tx */
65 	uint8_t nflits;		/* # of flits needed for the SGL */
66 	uint8_t hdls_used;	/* # of DMA handles used */
67 	uint32_t txb_used;	/* txb_space used */
68 	mac_ether_offload_info_t meoi;	/* pkt hdr info for offloads */
69 	struct ulptx_sgl sgl __attribute__((aligned(8)));
70 	struct ulptx_sge_pair reserved[TX_SGL_SEGS / 2];
71 };
72 
73 struct mblk_pair {
74 	mblk_t *head, *tail;
75 };
76 
77 struct rxbuf {
78 	kmem_cache_t *cache;		/* the kmem_cache this rxb came from */
79 	ddi_dma_handle_t dhdl;
80 	ddi_acc_handle_t ahdl;
81 	caddr_t va;			/* KVA of buffer */
82 	uint64_t ba;			/* bus address of buffer */
83 	frtn_t freefunc;
84 	uint_t buf_size;
85 	volatile uint_t ref_cnt;
86 };
87 
88 static int service_iq(struct sge_iq *iq, int budget);
89 static inline void init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx,
90     int8_t pktc_idx, int qsize, uint8_t esize);
91 static inline void init_fl(struct sge_fl *fl, uint16_t qsize);
92 static inline void init_eq(struct adapter *sc, struct sge_eq *eq,
93     uint16_t eqtype, uint16_t qsize, uint8_t tx_chan, uint16_t iqid);
94 static int alloc_iq_fl(struct port_info *pi, struct sge_iq *iq,
95     struct sge_fl *fl, int intr_idx, int cong);
96 static int free_iq_fl(struct port_info *pi, struct sge_iq *iq,
97     struct sge_fl *fl);
98 static int alloc_fwq(struct adapter *sc);
99 static int free_fwq(struct adapter *sc);
100 static int alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx,
101     int i);
102 static int free_rxq(struct port_info *pi, struct sge_rxq *rxq);
103 static int ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq);
104 static int eth_eq_alloc(struct adapter *sc, struct port_info *pi,
105     struct sge_eq *eq);
106 static int alloc_eq(struct adapter *sc, struct port_info *pi,
107     struct sge_eq *eq);
108 static int free_eq(struct adapter *sc, struct sge_eq *eq);
109 static int alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx);
110 static int free_txq(struct port_info *pi, struct sge_txq *txq);
111 static int alloc_dma_memory(struct adapter *sc, size_t len, int flags,
112     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
113     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
114     caddr_t *pva);
115 static int free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
116 static int alloc_desc_ring(struct adapter *sc, size_t len, int rw,
117     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
118     caddr_t *pva);
119 static int free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
120 static int alloc_tx_copybuffer(struct adapter *sc, size_t len,
121     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
122     caddr_t *pva);
123 static inline bool is_new_response(const struct sge_iq *iq,
124     struct rsp_ctrl **ctrl);
125 static inline void iq_next(struct sge_iq *iq);
126 static int refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs);
127 static void refill_sfl(void *arg);
128 static void add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl);
129 static void free_fl_bufs(struct sge_fl *fl);
130 static mblk_t *get_fl_payload(struct adapter *sc, struct sge_fl *fl,
131     uint32_t len_newbuf, int *fl_bufs_used);
132 static int get_frame_txinfo(struct sge_txq *txq, mblk_t **fp,
133     struct txinfo *txinfo, int sgl_only);
134 static inline int fits_in_txb(struct sge_txq *txq, int len, int *waste);
135 static inline int copy_into_txb(struct sge_txq *txq, mblk_t *m, int len,
136     struct txinfo *txinfo);
137 static inline void add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len);
138 static inline int add_mblk(struct sge_txq *txq, struct txinfo *txinfo,
139     mblk_t *m, int len);
140 static void free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo);
141 static int add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
142     struct txinfo *txinfo);
143 static void write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts);
144 static int write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
145     struct txinfo *txinfo);
146 static inline void write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
147     struct txpkts *txpkts, struct txinfo *txinfo);
148 static inline void copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to,
149     int len);
150 static inline void ring_tx_db(struct adapter *sc, struct sge_eq *eq);
151 static int reclaim_tx_descs(struct sge_txq *txq, int howmany);
152 static void write_txqflush_wr(struct sge_txq *txq);
153 static int t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss,
154     mblk_t *m);
155 static inline void ring_fl_db(struct adapter *sc, struct sge_fl *fl);
156 static kstat_t *setup_port_config_kstats(struct port_info *pi);
157 static kstat_t *setup_port_info_kstats(struct port_info *pi);
158 static kstat_t *setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq,
159     int idx);
160 static int update_rxq_kstats(kstat_t *ksp, int rw);
161 static int update_port_info_kstats(kstat_t *ksp, int rw);
162 static kstat_t *setup_txq_kstats(struct port_info *pi, struct sge_txq *txq,
163     int idx);
164 static int update_txq_kstats(kstat_t *ksp, int rw);
165 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
166     mblk_t *);
167 static int handle_fw_rpl(struct sge_iq *iq, const struct rss_header *rss,
168     mblk_t *m);
169 
170 static kmem_cache_t *rxbuf_cache_create(struct rxbuf_cache_params *);
171 static struct rxbuf *rxbuf_alloc(kmem_cache_t *, int, uint_t);
172 static void rxbuf_free(struct rxbuf *);
173 static int rxbuf_ctor(void *, void *, int);
174 static void rxbuf_dtor(void *, void *);
175 
176 static inline int
reclaimable(struct sge_eq * eq)177 reclaimable(struct sge_eq *eq)
178 {
179 	unsigned int cidx;
180 
181 	cidx = eq->spg->cidx;   /* stable snapshot */
182 	cidx = be16_to_cpu(cidx);
183 
184 	if (cidx >= eq->cidx)
185 		return (cidx - eq->cidx);
186 	else
187 		return (cidx + eq->cap - eq->cidx);
188 }
189 
190 void
t4_sge_init(struct adapter * sc)191 t4_sge_init(struct adapter *sc)
192 {
193 	struct driver_properties *p = &sc->props;
194 	ddi_dma_attr_t *dma_attr;
195 	ddi_device_acc_attr_t *acc_attr;
196 	uint32_t sge_control, sge_conm_ctrl;
197 	int egress_threshold;
198 
199 	/*
200 	 * Device access and DMA attributes for descriptor rings
201 	 */
202 	acc_attr = &sc->sge.acc_attr_desc;
203 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
204 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
205 	acc_attr->devacc_attr_dataorder = DDI_STRICTORDER_ACC;
206 
207 	dma_attr = &sc->sge.dma_attr_desc;
208 	dma_attr->dma_attr_version = DMA_ATTR_V0;
209 	dma_attr->dma_attr_addr_lo = 0;
210 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
211 	dma_attr->dma_attr_count_max = UINT64_MAX;
212 	dma_attr->dma_attr_align = 512;
213 	dma_attr->dma_attr_burstsizes = 0xfff;
214 	dma_attr->dma_attr_minxfer = 1;
215 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
216 	dma_attr->dma_attr_seg = UINT64_MAX;
217 	dma_attr->dma_attr_sgllen = 1;
218 	dma_attr->dma_attr_granular = 1;
219 	dma_attr->dma_attr_flags = 0;
220 
221 	/*
222 	 * Device access and DMA attributes for tx buffers
223 	 */
224 	acc_attr = &sc->sge.acc_attr_tx;
225 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
226 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
227 
228 	dma_attr = &sc->sge.dma_attr_tx;
229 	dma_attr->dma_attr_version = DMA_ATTR_V0;
230 	dma_attr->dma_attr_addr_lo = 0;
231 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
232 	dma_attr->dma_attr_count_max = UINT64_MAX;
233 	dma_attr->dma_attr_align = 1;
234 	dma_attr->dma_attr_burstsizes = 0xfff;
235 	dma_attr->dma_attr_minxfer = 1;
236 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
237 	dma_attr->dma_attr_seg = UINT64_MAX;
238 	dma_attr->dma_attr_sgllen = TX_SGL_SEGS;
239 	dma_attr->dma_attr_granular = 1;
240 	dma_attr->dma_attr_flags = 0;
241 
242 	/*
243 	 * Ingress Padding Boundary and Egress Status Page Size are set up by
244 	 * t4_fixup_host_params().
245 	 */
246 	sge_control = t4_read_reg(sc, A_SGE_CONTROL);
247 	sc->sge.pktshift = G_PKTSHIFT(sge_control);
248 	sc->sge.stat_len = (sge_control & F_EGRSTATUSPAGESIZE) ? 128 : 64;
249 
250 	/* t4_nex uses FLM packed mode */
251 	sc->sge.fl_align = t4_fl_pkt_align(sc, true);
252 
253 	/*
254 	 * Device access and DMA attributes for rx buffers
255 	 */
256 	sc->sge.rxb_params.dip = sc->dip;
257 	sc->sge.rxb_params.buf_size = rx_buf_size;
258 
259 	acc_attr = &sc->sge.rxb_params.acc_attr_rx;
260 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
261 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
262 
263 	dma_attr = &sc->sge.rxb_params.dma_attr_rx;
264 	dma_attr->dma_attr_version = DMA_ATTR_V0;
265 	dma_attr->dma_attr_addr_lo = 0;
266 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
267 	dma_attr->dma_attr_count_max = UINT64_MAX;
268 	/*
269 	 * Low 4 bits of an rx buffer address have a special meaning to the SGE
270 	 * and an rx buf cannot have an address with any of these bits set.
271 	 * FL_ALIGN is >= 32 so we're sure things are ok.
272 	 */
273 	dma_attr->dma_attr_align = sc->sge.fl_align;
274 	dma_attr->dma_attr_burstsizes = 0xfff;
275 	dma_attr->dma_attr_minxfer = 1;
276 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
277 	dma_attr->dma_attr_seg = UINT64_MAX;
278 	dma_attr->dma_attr_sgllen = 1;
279 	dma_attr->dma_attr_granular = 1;
280 	dma_attr->dma_attr_flags = 0;
281 
282 	sc->sge.rxbuf_cache = rxbuf_cache_create(&sc->sge.rxb_params);
283 
284 	/*
285 	 * A FL with <= fl_starve_thres buffers is starving and a periodic
286 	 * timer will attempt to refill it.  This needs to be larger than the
287 	 * SGE's Egress Congestion Threshold.  If it isn't, then we can get
288 	 * stuck waiting for new packets while the SGE is waiting for us to
289 	 * give it more Free List entries.  (Note that the SGE's Egress
290 	 * Congestion Threshold is in units of 2 Free List pointers.) For T4,
291 	 * there was only a single field to control this.  For T5 there's the
292 	 * original field which now only applies to Unpacked Mode Free List
293 	 * buffers and a new field which only applies to Packed Mode Free List
294 	 * buffers.
295 	 */
296 
297 	sge_conm_ctrl = t4_read_reg(sc, A_SGE_CONM_CTRL);
298 	switch (CHELSIO_CHIP_VERSION(sc->params.chip)) {
299 	case CHELSIO_T4:
300 		egress_threshold = G_EGRTHRESHOLD(sge_conm_ctrl);
301 		break;
302 	case CHELSIO_T5:
303 		egress_threshold = G_EGRTHRESHOLDPACKING(sge_conm_ctrl);
304 		break;
305 	case CHELSIO_T6:
306 	default:
307 		egress_threshold = G_T6_EGRTHRESHOLDPACKING(sge_conm_ctrl);
308 	}
309 	sc->sge.fl_starve_threshold = 2*egress_threshold + 1;
310 
311 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, rx_buf_size);
312 
313 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD,
314 	    V_THRESHOLD_0(p->counter_val[0]) |
315 	    V_THRESHOLD_1(p->counter_val[1]) |
316 	    V_THRESHOLD_2(p->counter_val[2]) |
317 	    V_THRESHOLD_3(p->counter_val[3]));
318 
319 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1,
320 	    V_TIMERVALUE0(us_to_core_ticks(sc, p->timer_val[0])) |
321 	    V_TIMERVALUE1(us_to_core_ticks(sc, p->timer_val[1])));
322 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3,
323 	    V_TIMERVALUE2(us_to_core_ticks(sc, p->timer_val[2])) |
324 	    V_TIMERVALUE3(us_to_core_ticks(sc, p->timer_val[3])));
325 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5,
326 	    V_TIMERVALUE4(us_to_core_ticks(sc, p->timer_val[4])) |
327 	    V_TIMERVALUE5(us_to_core_ticks(sc, p->timer_val[5])));
328 
329 	(void) t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_rpl);
330 	(void) t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_rpl);
331 	(void) t4_register_cpl_handler(sc, CPL_SGE_EGR_UPDATE,
332 	    handle_sge_egr_update);
333 	(void) t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
334 	(void) t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL,
335 	    t4_handle_fw_rpl);
336 }
337 
338 /*
339  * Allocate and initialize the firmware event queue and the forwarded interrupt
340  * queues, if any.  The adapter owns all these queues as they are not associated
341  * with any particular port.
342  *
343  * Returns errno on failure.  Resources allocated up to that point may still be
344  * allocated.  Caller is responsible for cleanup in case this function fails.
345  */
346 int
t4_setup_adapter_queues(struct adapter * sc)347 t4_setup_adapter_queues(struct adapter *sc)
348 {
349 	int rc;
350 
351 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
352 
353 	/*
354 	 * Firmware event queue
355 	 */
356 	rc = alloc_fwq(sc);
357 	if (rc != 0)
358 		return (rc);
359 
360 	return (rc);
361 }
362 
363 /*
364  * Idempotent
365  */
366 int
t4_teardown_adapter_queues(struct adapter * sc)367 t4_teardown_adapter_queues(struct adapter *sc)
368 {
369 
370 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
371 
372 	(void) free_fwq(sc);
373 
374 	return (0);
375 }
376 
377 static inline int
first_vector(struct port_info * pi)378 first_vector(struct port_info *pi)
379 {
380 	struct adapter *sc = pi->adapter;
381 	int rc = T4_EXTRA_INTR, i;
382 
383 	if (sc->intr_count == 1)
384 		return (0);
385 
386 	for_each_port(sc, i) {
387 		struct port_info *p = sc->port[i];
388 
389 		if (i == pi->port_id)
390 			break;
391 
392 		/*
393 		 * Not compiled with offload support and intr_count > 1.  Only
394 		 * NIC queues exist and they'd better be taking direct
395 		 * interrupts.
396 		 */
397 		ASSERT(!(sc->flags & INTR_FWD));
398 		rc += p->nrxq;
399 	}
400 	return (rc);
401 }
402 
403 /*
404  * Given an arbitrary "index," come up with an iq that can be used by other
405  * queues (of this port) for interrupt forwarding, SGE egress updates, etc.
406  * The iq returned is guaranteed to be something that takes direct interrupts.
407  */
408 static struct sge_iq *
port_intr_iq(struct port_info * pi,int idx)409 port_intr_iq(struct port_info *pi, int idx)
410 {
411 	struct adapter *sc = pi->adapter;
412 	struct sge *s = &sc->sge;
413 	struct sge_iq *iq = NULL;
414 
415 	if (sc->intr_count == 1)
416 		return (&sc->sge.fwq);
417 
418 	/*
419 	 * Not compiled with offload support and intr_count > 1.  Only NIC
420 	 * queues exist and they'd better be taking direct interrupts.
421 	 */
422 	ASSERT(!(sc->flags & INTR_FWD));
423 
424 	idx %= pi->nrxq;
425 	iq = &s->rxq[pi->first_rxq + idx].iq;
426 
427 	return (iq);
428 }
429 
430 int
t4_setup_port_queues(struct port_info * pi)431 t4_setup_port_queues(struct port_info *pi)
432 {
433 	int rc = 0, i, intr_idx, j;
434 	struct sge_rxq *rxq;
435 	struct sge_txq *txq;
436 	struct adapter *sc = pi->adapter;
437 	struct driver_properties *p = &sc->props;
438 
439 	pi->ksp_config = setup_port_config_kstats(pi);
440 	pi->ksp_info   = setup_port_info_kstats(pi);
441 
442 	/* Interrupt vector to start from (when using multiple vectors) */
443 	intr_idx = first_vector(pi);
444 
445 	/*
446 	 * First pass over all rx queues (NIC and TOE):
447 	 * a) initialize iq and fl
448 	 * b) allocate queue iff it will take direct interrupts.
449 	 */
450 
451 	for_each_rxq(pi, i, rxq) {
452 
453 		init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, p->qsize_rxq,
454 		    RX_IQ_ESIZE);
455 
456 		init_fl(&rxq->fl, p->qsize_rxq / 8); /* 8 bufs in each entry */
457 
458 		if ((!(sc->flags & INTR_FWD)) ||
459 		    (sc->intr_count > 1 && pi->nrxq)) {
460 			rxq->iq.flags |= IQ_INTR;
461 			rc = alloc_rxq(pi, rxq, intr_idx, i);
462 			if (rc != 0)
463 				goto done;
464 			intr_idx++;
465 		}
466 
467 	}
468 
469 	/*
470 	 * Second pass over all rx queues (NIC and TOE).  The queues forwarding
471 	 * their interrupts are allocated now.
472 	 */
473 	j = 0;
474 	for_each_rxq(pi, i, rxq) {
475 		if (rxq->iq.flags & IQ_INTR)
476 			continue;
477 
478 		intr_idx = port_intr_iq(pi, j)->abs_id;
479 
480 		rc = alloc_rxq(pi, rxq, intr_idx, i);
481 		if (rc != 0)
482 			goto done;
483 		j++;
484 	}
485 
486 	/*
487 	 * Now the tx queues.  Only one pass needed.
488 	 */
489 	j = 0;
490 	for_each_txq(pi, i, txq) {
491 		uint16_t iqid;
492 
493 		iqid = port_intr_iq(pi, j)->cntxt_id;
494 		init_eq(sc, &txq->eq, EQ_ETH, p->qsize_txq, pi->tx_chan, iqid);
495 		rc = alloc_txq(pi, txq, i);
496 		if (rc != 0)
497 			goto done;
498 	}
499 
500 done:
501 	if (rc != 0)
502 		(void) t4_teardown_port_queues(pi);
503 
504 	return (rc);
505 }
506 
507 /*
508  * Idempotent
509  */
510 int
t4_teardown_port_queues(struct port_info * pi)511 t4_teardown_port_queues(struct port_info *pi)
512 {
513 	int i;
514 	struct sge_rxq *rxq;
515 	struct sge_txq *txq;
516 
517 	if (pi->ksp_config != NULL) {
518 		kstat_delete(pi->ksp_config);
519 		pi->ksp_config = NULL;
520 	}
521 	if (pi->ksp_info != NULL) {
522 		kstat_delete(pi->ksp_info);
523 		pi->ksp_info = NULL;
524 	}
525 
526 	for_each_txq(pi, i, txq) {
527 		(void) free_txq(pi, txq);
528 	}
529 
530 	for_each_rxq(pi, i, rxq) {
531 		if ((rxq->iq.flags & IQ_INTR) == 0)
532 			(void) free_rxq(pi, rxq);
533 	}
534 
535 	/*
536 	 * Then take down the rx queues that take direct interrupts.
537 	 */
538 
539 	for_each_rxq(pi, i, rxq) {
540 		if (rxq->iq.flags & IQ_INTR)
541 			(void) free_rxq(pi, rxq);
542 	}
543 
544 	return (0);
545 }
546 
547 /* Deals with errors and forwarded interrupts */
548 uint_t
t4_intr_all(caddr_t arg1,caddr_t arg2)549 t4_intr_all(caddr_t arg1, caddr_t arg2)
550 {
551 
552 	(void) t4_intr_err(arg1, arg2);
553 	(void) t4_intr(arg1, arg2);
554 
555 	return (DDI_INTR_CLAIMED);
556 }
557 
558 static void
t4_intr_rx_work(struct sge_iq * iq)559 t4_intr_rx_work(struct sge_iq *iq)
560 {
561 	mblk_t *mp = NULL;
562 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
563 	RXQ_LOCK(rxq);
564 	if (!iq->polling) {
565 		mp = t4_ring_rx(rxq, iq->qsize/8);
566 		t4_write_reg(iq->adapter, MYPF_REG(A_SGE_PF_GTS),
567 		    V_INGRESSQID((u32)iq->cntxt_id) |
568 		    V_SEINTARM(iq->intr_next));
569 	}
570 	RXQ_UNLOCK(rxq);
571 	if (mp != NULL) {
572 		mac_rx_ring(rxq->port->mh, rxq->ring_handle, mp,
573 		    rxq->ring_gen_num);
574 	}
575 }
576 
577 /* Deals with interrupts on the given ingress queue */
578 /* ARGSUSED */
579 uint_t
t4_intr(caddr_t arg1,caddr_t arg2)580 t4_intr(caddr_t arg1, caddr_t arg2)
581 {
582 	struct sge_iq *iq = (struct sge_iq *)arg2;
583 	int state;
584 
585 	/*
586 	 * Right now receive polling is only enabled for MSI-X and
587 	 * when we have enough msi-x vectors i.e no interrupt forwarding.
588 	 */
589 	if (iq->adapter->props.multi_rings) {
590 		t4_intr_rx_work(iq);
591 	} else {
592 		state = atomic_cas_uint(&iq->state, IQS_IDLE, IQS_BUSY);
593 		if (state == IQS_IDLE) {
594 			(void) service_iq(iq, 0);
595 			(void) atomic_cas_uint(&iq->state, IQS_BUSY, IQS_IDLE);
596 		}
597 	}
598 	return (DDI_INTR_CLAIMED);
599 }
600 
601 /* Deals with error interrupts */
602 /* ARGSUSED */
603 uint_t
t4_intr_err(caddr_t arg1,caddr_t arg2)604 t4_intr_err(caddr_t arg1, caddr_t arg2)
605 {
606 	/* LINTED: E_BAD_PTR_CAST_ALIGN */
607 	struct adapter *sc = (struct adapter *)arg1;
608 
609 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
610 	(void) t4_slow_intr_handler(sc);
611 
612 	return (DDI_INTR_CLAIMED);
613 }
614 
615 /*
616  * t4_ring_rx - Process responses from an SGE response queue.
617  *
618  * This function processes responses from an SGE response queue up to the
619  * supplied budget.  Responses include received packets as well as control
620  * messages from FW or HW.
621  *
622  * It returns a chain of mblks containing the received data, to be
623  * passed up to mac_rx_ring().
624  */
625 mblk_t *
t4_ring_rx(struct sge_rxq * rxq,int budget)626 t4_ring_rx(struct sge_rxq *rxq, int budget)
627 {
628 	struct sge_iq *iq = &rxq->iq;
629 	struct sge_fl *fl = &rxq->fl;		/* Use iff IQ_HAS_FL */
630 	struct adapter *sc = iq->adapter;
631 	struct rsp_ctrl *ctrl;
632 	const struct rss_header *rss;
633 	int ndescs = 0, fl_bufs_used = 0;
634 	int rsp_type;
635 	uint32_t lq;
636 	mblk_t *mblk_head = NULL, **mblk_tail, *m;
637 	struct cpl_rx_pkt *cpl;
638 	uint32_t received_bytes = 0, pkt_len = 0;
639 	bool csum_ok;
640 	uint16_t err_vec;
641 
642 	mblk_tail = &mblk_head;
643 
644 	while (is_new_response(iq, &ctrl)) {
645 
646 		membar_consumer();
647 
648 		m = NULL;
649 		rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
650 		lq = be32_to_cpu(ctrl->pldbuflen_qid);
651 		rss = (const void *)iq->cdesc;
652 
653 		switch (rsp_type) {
654 		case X_RSPD_TYPE_FLBUF:
655 
656 			ASSERT(iq->flags & IQ_HAS_FL);
657 
658 			if (CPL_RX_PKT == rss->opcode) {
659 				cpl = (void *)(rss + 1);
660 				pkt_len = be16_to_cpu(cpl->len);
661 
662 				if (iq->polling &&
663 				    ((received_bytes + pkt_len) > budget))
664 					goto done;
665 
666 				m = get_fl_payload(sc, fl, lq, &fl_bufs_used);
667 				if (m == NULL)
668 					goto done;
669 
670 				iq->intr_next = iq->intr_params;
671 				m->b_rptr += sc->sge.pktshift;
672 				if (sc->params.tp.rx_pkt_encap) {
673 					/* Enabled only in T6 config file */
674 					err_vec = G_T6_COMPR_RXERR_VEC(
675 					    ntohs(cpl->err_vec));
676 				} else {
677 					err_vec = ntohs(cpl->err_vec);
678 				}
679 
680 				csum_ok = cpl->csum_calc && !err_vec;
681 
682 				/* TODO: what about cpl->ip_frag? */
683 				if (csum_ok && !cpl->ip_frag) {
684 					mac_hcksum_set(m, 0, 0, 0, 0xffff,
685 					    HCK_FULLCKSUM_OK | HCK_FULLCKSUM |
686 					    HCK_IPV4_HDRCKSUM_OK);
687 					rxq->rxcsum++;
688 				}
689 				rxq->rxpkts++;
690 				rxq->rxbytes += pkt_len;
691 				received_bytes += pkt_len;
692 
693 				*mblk_tail = m;
694 				mblk_tail = &m->b_next;
695 
696 				break;
697 			}
698 
699 			m = get_fl_payload(sc, fl, lq, &fl_bufs_used);
700 			if (m == NULL)
701 				goto done;
702 			/* FALLTHROUGH */
703 
704 		case X_RSPD_TYPE_CPL:
705 			ASSERT(rss->opcode < NUM_CPL_CMDS);
706 			sc->cpl_handler[rss->opcode](iq, rss, m);
707 			break;
708 
709 		default:
710 			break;
711 		}
712 		iq_next(iq);
713 		++ndescs;
714 		if (!iq->polling && (ndescs == budget))
715 			break;
716 	}
717 
718 done:
719 
720 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
721 	    V_CIDXINC(ndescs) | V_INGRESSQID(iq->cntxt_id) |
722 	    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
723 
724 	if ((fl_bufs_used > 0) || (iq->flags & IQ_HAS_FL)) {
725 		int starved;
726 		FL_LOCK(fl);
727 		fl->needed += fl_bufs_used;
728 		starved = refill_fl(sc, fl, fl->cap / 8);
729 		FL_UNLOCK(fl);
730 		if (starved)
731 			add_fl_to_sfl(sc, fl);
732 	}
733 	return (mblk_head);
734 }
735 
736 /*
737  * Deals with anything and everything on the given ingress queue.
738  */
739 static int
service_iq(struct sge_iq * iq,int budget)740 service_iq(struct sge_iq *iq, int budget)
741 {
742 	struct sge_iq *q;
743 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
744 	struct sge_fl *fl = &rxq->fl;		/* Use iff IQ_HAS_FL */
745 	struct adapter *sc = iq->adapter;
746 	struct rsp_ctrl *ctrl;
747 	const struct rss_header *rss;
748 	int ndescs = 0, limit, fl_bufs_used = 0;
749 	int rsp_type;
750 	uint32_t lq;
751 	int starved;
752 	mblk_t *m;
753 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
754 
755 	limit = budget ? budget : iq->qsize / 8;
756 
757 	/*
758 	 * We always come back and check the descriptor ring for new indirect
759 	 * interrupts and other responses after running a single handler.
760 	 */
761 	for (;;) {
762 		while (is_new_response(iq, &ctrl)) {
763 
764 			membar_consumer();
765 
766 			m = NULL;
767 			rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
768 			lq = be32_to_cpu(ctrl->pldbuflen_qid);
769 			rss = (const void *)iq->cdesc;
770 
771 			switch (rsp_type) {
772 			case X_RSPD_TYPE_FLBUF:
773 
774 				ASSERT(iq->flags & IQ_HAS_FL);
775 
776 				m = get_fl_payload(sc, fl, lq, &fl_bufs_used);
777 				if (m == NULL) {
778 					/*
779 					 * Rearm the iq with a
780 					 * longer-than-default timer
781 					 */
782 					const uint32_t timer_idx =
783 					    V_QINTR_TIMER_IDX(SGE_NTIMERS-1);
784 					t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
785 					    V_CIDXINC(ndescs) |
786 					    V_INGRESSQID((u32)iq->cntxt_id) |
787 					    V_SEINTARM(timer_idx));
788 					if (fl_bufs_used > 0) {
789 						ASSERT(iq->flags & IQ_HAS_FL);
790 						FL_LOCK(fl);
791 						fl->needed += fl_bufs_used;
792 						starved = refill_fl(sc, fl,
793 						    fl->cap / 8);
794 						FL_UNLOCK(fl);
795 						if (starved)
796 							add_fl_to_sfl(sc, fl);
797 					}
798 					return (0);
799 				}
800 
801 			/* FALLTHRU */
802 			case X_RSPD_TYPE_CPL:
803 
804 				ASSERT(rss->opcode < NUM_CPL_CMDS);
805 				sc->cpl_handler[rss->opcode](iq, rss, m);
806 				break;
807 
808 			case X_RSPD_TYPE_INTR:
809 
810 				/*
811 				 * Interrupts should be forwarded only to queues
812 				 * that are not forwarding their interrupts.
813 				 * This means service_iq can recurse but only 1
814 				 * level deep.
815 				 */
816 				ASSERT(budget == 0);
817 
818 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
819 				if (atomic_cas_uint(&q->state, IQS_IDLE,
820 				    IQS_BUSY) == IQS_IDLE) {
821 					if (service_iq(q, q->qsize / 8) == 0) {
822 						(void) atomic_cas_uint(
823 						    &q->state, IQS_BUSY,
824 						    IQS_IDLE);
825 					} else {
826 						STAILQ_INSERT_TAIL(&iql, q,
827 						    link);
828 					}
829 				}
830 				break;
831 
832 			default:
833 				break;
834 			}
835 
836 			iq_next(iq);
837 			if (++ndescs == limit) {
838 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
839 				    V_CIDXINC(ndescs) |
840 				    V_INGRESSQID(iq->cntxt_id) |
841 				    V_SEINTARM(V_QINTR_TIMER_IDX(
842 				    X_TIMERREG_UPDATE_CIDX)));
843 				ndescs = 0;
844 
845 				if (fl_bufs_used > 0) {
846 					ASSERT(iq->flags & IQ_HAS_FL);
847 					FL_LOCK(fl);
848 					fl->needed += fl_bufs_used;
849 					(void) refill_fl(sc, fl, fl->cap / 8);
850 					FL_UNLOCK(fl);
851 					fl_bufs_used = 0;
852 				}
853 
854 				if (budget != 0)
855 					return (EINPROGRESS);
856 			}
857 		}
858 
859 		if (STAILQ_EMPTY(&iql) != 0)
860 			break;
861 
862 		/*
863 		 * Process the head only, and send it to the back of the list if
864 		 * it's still not done.
865 		 */
866 		q = STAILQ_FIRST(&iql);
867 		STAILQ_REMOVE_HEAD(&iql, link);
868 		if (service_iq(q, q->qsize / 8) == 0)
869 			(void) atomic_cas_uint(&q->state, IQS_BUSY, IQS_IDLE);
870 		else
871 			STAILQ_INSERT_TAIL(&iql, q, link);
872 	}
873 
874 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
875 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_next));
876 
877 	if (iq->flags & IQ_HAS_FL) {
878 
879 		FL_LOCK(fl);
880 		fl->needed += fl_bufs_used;
881 		starved = refill_fl(sc, fl, fl->cap / 4);
882 		FL_UNLOCK(fl);
883 		if (starved != 0)
884 			add_fl_to_sfl(sc, fl);
885 	}
886 
887 	return (0);
888 }
889 
890 /* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */
891 #define	TXPKTS_PKT_HDR ((\
892 	sizeof (struct ulp_txpkt) + \
893 	sizeof (struct ulptx_idata) + \
894 	sizeof (struct cpl_tx_pkt_core)) / 8)
895 
896 /* Header of a coalesced tx WR, before SGL of first packet (in flits) */
897 #define	TXPKTS_WR_HDR (\
898 	sizeof (struct fw_eth_tx_pkts_wr) / 8 + \
899 	TXPKTS_PKT_HDR)
900 
901 /* Header of a tx WR, before SGL of first packet (in flits) */
902 #define	TXPKT_WR_HDR ((\
903 	sizeof (struct fw_eth_tx_pkt_wr) + \
904 	sizeof (struct cpl_tx_pkt_core)) / 8)
905 
906 /* Header of a tx LSO WR, before SGL of first packet (in flits) */
907 #define	TXPKT_LSO_WR_HDR ((\
908 	sizeof (struct fw_eth_tx_pkt_wr) + \
909 	sizeof (struct cpl_tx_pkt_lso_core) + \
910 	sizeof (struct cpl_tx_pkt_core)) / 8)
911 
912 mblk_t *
t4_eth_tx(void * arg,mblk_t * frame)913 t4_eth_tx(void *arg, mblk_t *frame)
914 {
915 	struct sge_txq *txq = (struct sge_txq *)arg;
916 	struct port_info *pi = txq->port;
917 	struct adapter *sc = pi->adapter;
918 	struct sge_eq *eq = &txq->eq;
919 	mblk_t *next_frame;
920 	int rc, coalescing;
921 	struct txpkts txpkts;
922 	struct txinfo txinfo;
923 
924 	txpkts.npkt = 0; /* indicates there's nothing in txpkts */
925 	coalescing = 0;
926 
927 	TXQ_LOCK(txq);
928 	if (eq->avail < 8)
929 		(void) reclaim_tx_descs(txq, 8);
930 	for (; frame; frame = next_frame) {
931 
932 		if (eq->avail < 8)
933 			break;
934 
935 		next_frame = frame->b_next;
936 		frame->b_next = NULL;
937 
938 		if (next_frame != NULL)
939 			coalescing = 1;
940 
941 		rc = get_frame_txinfo(txq, &frame, &txinfo, coalescing);
942 		if (rc != 0) {
943 			if (rc == ENOMEM) {
944 
945 				/* Short of resources, suspend tx */
946 
947 				frame->b_next = next_frame;
948 				break;
949 			}
950 
951 			/*
952 			 * Unrecoverable error for this frame, throw it
953 			 * away and move on to the next.
954 			 */
955 
956 			freemsg(frame);
957 			continue;
958 		}
959 
960 		if (coalescing != 0 &&
961 		    add_to_txpkts(txq, &txpkts, frame, &txinfo) == 0) {
962 
963 			/* Successfully absorbed into txpkts */
964 
965 			write_ulp_cpl_sgl(pi, txq, &txpkts, &txinfo);
966 			goto doorbell;
967 		}
968 
969 		/*
970 		 * We weren't coalescing to begin with, or current frame could
971 		 * not be coalesced (add_to_txpkts flushes txpkts if a frame
972 		 * given to it can't be coalesced).  Either way there should be
973 		 * nothing in txpkts.
974 		 */
975 		ASSERT(txpkts.npkt == 0);
976 
977 		/* We're sending out individual frames now */
978 		coalescing = 0;
979 
980 		if (eq->avail < 8)
981 			(void) reclaim_tx_descs(txq, 8);
982 		rc = write_txpkt_wr(pi, txq, frame, &txinfo);
983 		if (rc != 0) {
984 
985 			/* Short of hardware descriptors, suspend tx */
986 
987 			/*
988 			 * This is an unlikely but expensive failure.  We've
989 			 * done all the hard work (DMA bindings etc.) and now we
990 			 * can't send out the frame.  What's worse, we have to
991 			 * spend even more time freeing up everything in txinfo.
992 			 */
993 			txq->qfull++;
994 			free_txinfo_resources(txq, &txinfo);
995 
996 			frame->b_next = next_frame;
997 			break;
998 		}
999 
1000 doorbell:
1001 		/* Fewer and fewer doorbells as the queue fills up */
1002 		if (eq->pending >= (1 << (fls(eq->qsize - eq->avail) / 2))) {
1003 			txq->txbytes += txinfo.len;
1004 			txq->txpkts++;
1005 			ring_tx_db(sc, eq);
1006 		}
1007 		(void) reclaim_tx_descs(txq, 32);
1008 	}
1009 
1010 	if (txpkts.npkt > 0)
1011 		write_txpkts_wr(txq, &txpkts);
1012 
1013 	/*
1014 	 * frame not NULL means there was an error but we haven't thrown it
1015 	 * away.  This can happen when we're short of tx descriptors (qfull) or
1016 	 * maybe even DMA handles (dma_hdl_failed).  Either way, a credit flush
1017 	 * and reclaim will get things going again.
1018 	 *
1019 	 * If eq->avail is already 0 we know a credit flush was requested in the
1020 	 * WR that reduced it to 0 so we don't need another flush (we don't have
1021 	 * any descriptor for a flush WR anyway, duh).
1022 	 */
1023 	if (frame && eq->avail > 0)
1024 		write_txqflush_wr(txq);
1025 
1026 	if (eq->pending != 0)
1027 		ring_tx_db(sc, eq);
1028 
1029 	(void) reclaim_tx_descs(txq, eq->qsize);
1030 	TXQ_UNLOCK(txq);
1031 
1032 	return (frame);
1033 }
1034 
1035 static inline void
init_iq(struct sge_iq * iq,struct adapter * sc,int tmr_idx,int8_t pktc_idx,int qsize,uint8_t esize)1036 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int8_t pktc_idx,
1037     int qsize, uint8_t esize)
1038 {
1039 	ASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS);
1040 	ASSERT(pktc_idx < SGE_NCOUNTERS);	/* -ve is ok, means don't use */
1041 
1042 	iq->flags = 0;
1043 	iq->adapter = sc;
1044 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
1045 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
1046 	if (pktc_idx >= 0) {
1047 		iq->intr_params |= F_QINTR_CNT_EN;
1048 		iq->intr_pktc_idx = pktc_idx;
1049 	}
1050 	iq->qsize = roundup(qsize, 16);		/* See FW_IQ_CMD/iqsize */
1051 	iq->esize = max(esize, 16);		/* See FW_IQ_CMD/iqesize */
1052 }
1053 
1054 static inline void
init_fl(struct sge_fl * fl,uint16_t qsize)1055 init_fl(struct sge_fl *fl, uint16_t qsize)
1056 {
1057 
1058 	fl->qsize = qsize;
1059 	fl->allocb_fail = 0;
1060 }
1061 
1062 static inline void
init_eq(struct adapter * sc,struct sge_eq * eq,uint16_t eqtype,uint16_t qsize,uint8_t tx_chan,uint16_t iqid)1063 init_eq(struct adapter *sc, struct sge_eq *eq, uint16_t eqtype, uint16_t qsize,
1064     uint8_t tx_chan, uint16_t iqid)
1065 {
1066 	struct sge *s = &sc->sge;
1067 	uint32_t r;
1068 
1069 	ASSERT(tx_chan < NCHAN);
1070 	ASSERT(eqtype <= EQ_TYPEMASK);
1071 
1072 	if (is_t5(sc->params.chip)) {
1073 		r = t4_read_reg(sc, A_SGE_EGRESS_QUEUES_PER_PAGE_PF);
1074 		r >>= S_QUEUESPERPAGEPF0 +
1075 		    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
1076 		s->s_qpp = r & M_QUEUESPERPAGEPF0;
1077 	}
1078 
1079 	eq->flags = eqtype & EQ_TYPEMASK;
1080 	eq->tx_chan = tx_chan;
1081 	eq->iqid = iqid;
1082 	eq->qsize = qsize;
1083 }
1084 
1085 /*
1086  * Allocates the ring for an ingress queue and an optional freelist.  If the
1087  * freelist is specified it will be allocated and then associated with the
1088  * ingress queue.
1089  *
1090  * Returns errno on failure.  Resources allocated up to that point may still be
1091  * allocated.  Caller is responsible for cleanup in case this function fails.
1092  *
1093  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
1094  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
1095  * the index of the queue to which its interrupts will be forwarded.
1096  */
1097 static int
alloc_iq_fl(struct port_info * pi,struct sge_iq * iq,struct sge_fl * fl,int intr_idx,int cong)1098 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
1099     int intr_idx, int cong)
1100 {
1101 	int rc, i, cntxt_id;
1102 	size_t len;
1103 	struct fw_iq_cmd c;
1104 	struct adapter *sc = iq->adapter;
1105 	uint32_t v = 0;
1106 
1107 	len = iq->qsize * iq->esize;
1108 	rc = alloc_desc_ring(sc, len, DDI_DMA_READ, &iq->dhdl, &iq->ahdl,
1109 	    &iq->ba, (caddr_t *)&iq->desc);
1110 	if (rc != 0)
1111 		return (rc);
1112 
1113 	bzero(&c, sizeof (c));
1114 	c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
1115 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
1116 	    V_FW_IQ_CMD_VFN(0));
1117 
1118 	c.alloc_to_len16 = cpu_to_be32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
1119 	    FW_LEN16(c));
1120 
1121 	/* Special handling for firmware event queue */
1122 	if (iq == &sc->sge.fwq)
1123 		v |= F_FW_IQ_CMD_IQASYNCH;
1124 
1125 	if (iq->flags & IQ_INTR)
1126 		ASSERT(intr_idx < sc->intr_count);
1127 	else
1128 		v |= F_FW_IQ_CMD_IQANDST;
1129 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
1130 
1131 	c.type_to_iqandstindex = cpu_to_be32(v |
1132 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
1133 	    V_FW_IQ_CMD_VIID(pi->viid) |
1134 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
1135 	c.iqdroprss_to_iqesize = cpu_to_be16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
1136 	    F_FW_IQ_CMD_IQGTSMODE |
1137 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
1138 	    V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4));
1139 	c.iqsize = cpu_to_be16(iq->qsize);
1140 	c.iqaddr = cpu_to_be64(iq->ba);
1141 	if (cong >= 0) {
1142 		const uint32_t iq_type =
1143 		    cong ? FW_IQ_IQTYPE_NIC : FW_IQ_IQTYPE_OFLD;
1144 		c.iqns_to_fl0congen = BE_32(F_FW_IQ_CMD_IQFLINTCONGEN |
1145 		    V_FW_IQ_CMD_IQTYPE(iq_type));
1146 	}
1147 
1148 	if (fl != NULL) {
1149 		unsigned int chip_ver = CHELSIO_CHIP_VERSION(sc->params.chip);
1150 
1151 		mutex_init(&fl->lock, NULL, MUTEX_DRIVER,
1152 		    DDI_INTR_PRI(sc->intr_pri));
1153 		fl->flags |= FL_MTX;
1154 
1155 		len = fl->qsize * RX_FL_ESIZE;
1156 		rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &fl->dhdl,
1157 		    &fl->ahdl, &fl->ba, (caddr_t *)&fl->desc);
1158 		if (rc != 0)
1159 			return (rc);
1160 
1161 		/* Allocate space for one software descriptor per buffer. */
1162 		fl->cap = (fl->qsize - sc->sge.stat_len / RX_FL_ESIZE) * 8;
1163 		fl->sdesc = kmem_zalloc(sizeof (struct fl_sdesc) * fl->cap,
1164 		    KM_SLEEP);
1165 		fl->needed = fl->cap;
1166 		fl->lowat = roundup(sc->sge.fl_starve_threshold, 8);
1167 
1168 		c.iqns_to_fl0congen |=
1169 		    cpu_to_be32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
1170 		    F_FW_IQ_CMD_FL0PACKEN | F_FW_IQ_CMD_FL0PADEN);
1171 		if (cong >= 0) {
1172 			c.iqns_to_fl0congen |=
1173 			    BE_32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
1174 			    F_FW_IQ_CMD_FL0CONGCIF |
1175 			    F_FW_IQ_CMD_FL0CONGEN);
1176 		}
1177 
1178 		/*
1179 		 * In T6, for egress queue type FL there is internal overhead
1180 		 * of 16B for header going into FLM module.  Hence the maximum
1181 		 * allowed burst size is 448 bytes.  For T4/T5, the hardware
1182 		 * doesn't coalesce fetch requests if more than 64 bytes of
1183 		 * Free List pointers are provided, so we use a 128-byte Fetch
1184 		 * Burst Minimum there (T6 implements coalescing so we can use
1185 		 * the smaller 64-byte value there).
1186 		 */
1187 
1188 		c.fl0dcaen_to_fl0cidxfthresh = cpu_to_be16(
1189 		    V_FW_IQ_CMD_FL0FBMIN(chip_ver <= CHELSIO_T5 ?
1190 		    X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
1191 		    V_FW_IQ_CMD_FL0FBMAX(chip_ver <= CHELSIO_T5 ?
1192 		    X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
1193 		c.fl0size = cpu_to_be16(fl->qsize);
1194 		c.fl0addr = cpu_to_be64(fl->ba);
1195 	}
1196 
1197 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1198 	if (rc != 0) {
1199 		cxgb_printf(sc->dip, CE_WARN,
1200 		    "failed to create ingress queue: %d", rc);
1201 		return (rc);
1202 	}
1203 
1204 	iq->cdesc = iq->desc;
1205 	iq->cidx = 0;
1206 	iq->gen = 1;
1207 	iq->intr_next = iq->intr_params;
1208 	iq->adapter = sc;
1209 	iq->cntxt_id = be16_to_cpu(c.iqid);
1210 	iq->abs_id = be16_to_cpu(c.physiqid);
1211 	iq->flags |= IQ_ALLOCATED;
1212 	mutex_init(&iq->lock, NULL, MUTEX_DRIVER,
1213 	    DDI_INTR_PRI(DDI_INTR_PRI(sc->intr_pri)));
1214 	iq->polling = 0;
1215 
1216 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
1217 	if (cntxt_id >= sc->sge.iqmap_sz) {
1218 		panic("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
1219 		    cntxt_id, sc->sge.iqmap_sz - 1);
1220 	}
1221 	sc->sge.iqmap[cntxt_id] = iq;
1222 
1223 	if (fl != NULL) {
1224 		fl->cntxt_id = be16_to_cpu(c.fl0id);
1225 		fl->pidx = fl->cidx = 0;
1226 		fl->copy_threshold = rx_copy_threshold;
1227 
1228 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
1229 		if (cntxt_id >= sc->sge.eqmap_sz) {
1230 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
1231 			    __func__, cntxt_id, sc->sge.eqmap_sz - 1);
1232 		}
1233 		sc->sge.eqmap[cntxt_id] = (void *)fl;
1234 
1235 		FL_LOCK(fl);
1236 		(void) refill_fl(sc, fl, fl->lowat);
1237 		FL_UNLOCK(fl);
1238 
1239 		iq->flags |= IQ_HAS_FL;
1240 	}
1241 
1242 	if (is_t5(sc->params.chip) && cong >= 0) {
1243 		uint32_t param, val;
1244 
1245 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
1246 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
1247 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
1248 		if (cong == 0)
1249 			val = 1 << 19;
1250 		else {
1251 			val = 2 << 19;
1252 			for (i = 0; i < 4; i++) {
1253 				if (cong & (1 << i))
1254 					val |= 1 << (i << 2);
1255 			}
1256 		}
1257 
1258 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
1259 		if (rc != 0) {
1260 			/* report error but carry on */
1261 			cxgb_printf(sc->dip, CE_WARN,
1262 			    "failed to set congestion manager context for "
1263 			    "ingress queue %d: %d", iq->cntxt_id, rc);
1264 		}
1265 	}
1266 
1267 	/* Enable IQ interrupts */
1268 	iq->state = IQS_IDLE;
1269 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
1270 	    V_INGRESSQID(iq->cntxt_id));
1271 
1272 	return (0);
1273 }
1274 
1275 static int
free_iq_fl(struct port_info * pi,struct sge_iq * iq,struct sge_fl * fl)1276 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
1277 {
1278 	int rc;
1279 
1280 	if (iq != NULL) {
1281 		struct adapter *sc = iq->adapter;
1282 		dev_info_t *dip;
1283 
1284 		dip = pi ? pi->dip : sc->dip;
1285 		if (iq->flags & IQ_ALLOCATED) {
1286 			rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
1287 			    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
1288 			    fl ? fl->cntxt_id : 0xffff, 0xffff);
1289 			if (rc != 0) {
1290 				cxgb_printf(dip, CE_WARN,
1291 				    "failed to free queue %p: %d", iq, rc);
1292 				return (rc);
1293 			}
1294 			mutex_destroy(&iq->lock);
1295 			iq->flags &= ~IQ_ALLOCATED;
1296 		}
1297 
1298 		if (iq->desc != NULL) {
1299 			(void) free_desc_ring(&iq->dhdl, &iq->ahdl);
1300 			iq->desc = NULL;
1301 		}
1302 
1303 		bzero(iq, sizeof (*iq));
1304 	}
1305 
1306 	if (fl != NULL) {
1307 		if (fl->sdesc != NULL) {
1308 			FL_LOCK(fl);
1309 			free_fl_bufs(fl);
1310 			FL_UNLOCK(fl);
1311 
1312 			kmem_free(fl->sdesc, sizeof (struct fl_sdesc) *
1313 			    fl->cap);
1314 			fl->sdesc = NULL;
1315 		}
1316 
1317 		if (fl->desc != NULL) {
1318 			(void) free_desc_ring(&fl->dhdl, &fl->ahdl);
1319 			fl->desc = NULL;
1320 		}
1321 
1322 		if (fl->flags & FL_MTX) {
1323 			mutex_destroy(&fl->lock);
1324 			fl->flags &= ~FL_MTX;
1325 		}
1326 
1327 		bzero(fl, sizeof (struct sge_fl));
1328 	}
1329 
1330 	return (0);
1331 }
1332 
1333 static int
alloc_fwq(struct adapter * sc)1334 alloc_fwq(struct adapter *sc)
1335 {
1336 	int rc, intr_idx;
1337 	struct sge_iq *fwq = &sc->sge.fwq;
1338 
1339 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, FW_IQ_ESIZE);
1340 	fwq->flags |= IQ_INTR;	/* always */
1341 	intr_idx = sc->intr_count > 1 ? 1 : 0;
1342 	rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1);
1343 	if (rc != 0) {
1344 		cxgb_printf(sc->dip, CE_WARN,
1345 		    "failed to create firmware event queue: %d.", rc);
1346 		return (rc);
1347 	}
1348 
1349 	return (0);
1350 }
1351 
1352 static int
free_fwq(struct adapter * sc)1353 free_fwq(struct adapter *sc)
1354 {
1355 
1356 	return (free_iq_fl(NULL, &sc->sge.fwq, NULL));
1357 }
1358 
1359 static int
alloc_rxq(struct port_info * pi,struct sge_rxq * rxq,int intr_idx,int i)1360 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int i)
1361 {
1362 	int rc;
1363 
1364 	rxq->port = pi;
1365 	rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx,
1366 	    t4_get_tp_ch_map(pi->adapter, pi->tx_chan));
1367 	if (rc != 0)
1368 		return (rc);
1369 
1370 	rxq->ksp = setup_rxq_kstats(pi, rxq, i);
1371 
1372 	return (rc);
1373 }
1374 
1375 static int
free_rxq(struct port_info * pi,struct sge_rxq * rxq)1376 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
1377 {
1378 	int rc;
1379 
1380 	if (rxq->ksp != NULL) {
1381 		kstat_delete(rxq->ksp);
1382 		rxq->ksp = NULL;
1383 	}
1384 
1385 	rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
1386 	if (rc == 0)
1387 		bzero(&rxq->fl, sizeof (*rxq) - offsetof(struct sge_rxq, fl));
1388 
1389 	return (rc);
1390 }
1391 
1392 static int
ctrl_eq_alloc(struct adapter * sc,struct sge_eq * eq)1393 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
1394 {
1395 	int rc, cntxt_id;
1396 	struct fw_eq_ctrl_cmd c;
1397 
1398 	bzero(&c, sizeof (c));
1399 
1400 	c.op_to_vfn = BE_32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
1401 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
1402 	    V_FW_EQ_CTRL_CMD_VFN(0));
1403 	c.alloc_to_len16 = BE_32(F_FW_EQ_CTRL_CMD_ALLOC |
1404 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
1405 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); /* TODO */
1406 	c.physeqid_pkd = BE_32(0);
1407 	c.fetchszm_to_iqid =
1408 	    BE_32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1409 	    V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
1410 	    F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
1411 	c.dcaen_to_eqsize =
1412 	    BE_32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1413 	    V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1414 	    V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1415 	    V_FW_EQ_CTRL_CMD_EQSIZE(eq->qsize));
1416 	c.eqaddr = BE_64(eq->ba);
1417 
1418 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1419 	if (rc != 0) {
1420 		cxgb_printf(sc->dip, CE_WARN,
1421 		    "failed to create control queue %d: %d", eq->tx_chan, rc);
1422 		return (rc);
1423 	}
1424 	eq->flags |= EQ_ALLOCATED;
1425 
1426 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(BE_32(c.cmpliqid_eqid));
1427 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1428 	if (cntxt_id >= sc->sge.eqmap_sz)
1429 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1430 		    cntxt_id, sc->sge.eqmap_sz - 1);
1431 	sc->sge.eqmap[cntxt_id] = eq;
1432 
1433 	return (rc);
1434 }
1435 
1436 static int
eth_eq_alloc(struct adapter * sc,struct port_info * pi,struct sge_eq * eq)1437 eth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1438 {
1439 	int rc, cntxt_id;
1440 	struct fw_eq_eth_cmd c;
1441 
1442 	bzero(&c, sizeof (c));
1443 
1444 	c.op_to_vfn = BE_32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
1445 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
1446 	    V_FW_EQ_ETH_CMD_VFN(0));
1447 	c.alloc_to_len16 = BE_32(F_FW_EQ_ETH_CMD_ALLOC |
1448 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
1449 	c.autoequiqe_to_viid = BE_32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
1450 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(pi->viid));
1451 	c.fetchszm_to_iqid =
1452 	    BE_32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1453 	    V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
1454 	    V_FW_EQ_ETH_CMD_IQID(eq->iqid));
1455 	c.dcaen_to_eqsize = BE_32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1456 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1457 	    V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1458 	    V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize));
1459 	c.eqaddr = BE_64(eq->ba);
1460 
1461 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1462 	if (rc != 0) {
1463 		cxgb_printf(pi->dip, CE_WARN,
1464 		    "failed to create Ethernet egress queue: %d", rc);
1465 		return (rc);
1466 	}
1467 	eq->flags |= EQ_ALLOCATED;
1468 
1469 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(BE_32(c.eqid_pkd));
1470 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1471 	if (cntxt_id >= sc->sge.eqmap_sz)
1472 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1473 		    cntxt_id, sc->sge.eqmap_sz - 1);
1474 	sc->sge.eqmap[cntxt_id] = eq;
1475 
1476 	return (rc);
1477 }
1478 
1479 static int
alloc_eq(struct adapter * sc,struct port_info * pi,struct sge_eq * eq)1480 alloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1481 {
1482 	int rc;
1483 	size_t len;
1484 
1485 	mutex_init(&eq->lock, NULL, MUTEX_DRIVER, DDI_INTR_PRI(sc->intr_pri));
1486 	eq->flags |= EQ_MTX;
1487 
1488 	len = eq->qsize * EQ_ESIZE;
1489 	rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &eq->desc_dhdl,
1490 	    &eq->desc_ahdl, &eq->ba, (caddr_t *)&eq->desc);
1491 	if (rc != 0)
1492 		return (rc);
1493 
1494 	eq->cap = eq->qsize - sc->sge.stat_len / EQ_ESIZE;
1495 	eq->spg = (void *)&eq->desc[eq->cap];
1496 	eq->avail = eq->cap - 1;	/* one less to avoid cidx = pidx */
1497 	eq->pidx = eq->cidx = 0;
1498 	eq->doorbells = sc->doorbells;
1499 
1500 	switch (eq->flags & EQ_TYPEMASK) {
1501 	case EQ_CTRL:
1502 		rc = ctrl_eq_alloc(sc, eq);
1503 		break;
1504 
1505 	case EQ_ETH:
1506 		rc = eth_eq_alloc(sc, pi, eq);
1507 		break;
1508 
1509 	default:
1510 		panic("%s: invalid eq type %d.", __func__,
1511 		    eq->flags & EQ_TYPEMASK);
1512 	}
1513 
1514 	if (eq->doorbells & (DOORBELL_UDB | DOORBELL_UDBWC | DOORBELL_WCWR)) {
1515 		uint32_t s_qpp = sc->sge.s_qpp;
1516 		uint32_t mask = (1 << s_qpp) - 1;
1517 		volatile uint8_t *udb;
1518 
1519 		udb = (volatile uint8_t *)sc->reg1p + UDBS_DB_OFFSET;
1520 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
1521 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
1522 		if (eq->udb_qid > PAGE_SIZE / UDBS_SEG_SIZE)
1523 			eq->doorbells &= ~DOORBELL_WCWR;
1524 		else {
1525 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
1526 			eq->udb_qid = 0;
1527 		}
1528 		eq->udb = (volatile void *)udb;
1529 	}
1530 
1531 	if (rc != 0) {
1532 		cxgb_printf(sc->dip, CE_WARN,
1533 		    "failed to allocate egress queue(%d): %d",
1534 		    eq->flags & EQ_TYPEMASK, rc);
1535 	}
1536 
1537 	return (rc);
1538 }
1539 
1540 static int
free_eq(struct adapter * sc,struct sge_eq * eq)1541 free_eq(struct adapter *sc, struct sge_eq *eq)
1542 {
1543 	int rc;
1544 
1545 	if (eq->flags & EQ_ALLOCATED) {
1546 		switch (eq->flags & EQ_TYPEMASK) {
1547 		case EQ_CTRL:
1548 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
1549 			    eq->cntxt_id);
1550 			break;
1551 
1552 		case EQ_ETH:
1553 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
1554 			    eq->cntxt_id);
1555 			break;
1556 		default:
1557 			panic("%s: invalid eq type %d.", __func__,
1558 			    eq->flags & EQ_TYPEMASK);
1559 		}
1560 		if (rc != 0) {
1561 			cxgb_printf(sc->dip, CE_WARN,
1562 			    "failed to free egress queue (%d): %d",
1563 			    eq->flags & EQ_TYPEMASK, rc);
1564 			return (rc);
1565 		}
1566 		eq->flags &= ~EQ_ALLOCATED;
1567 	}
1568 
1569 	if (eq->desc != NULL) {
1570 		(void) free_desc_ring(&eq->desc_dhdl, &eq->desc_ahdl);
1571 		eq->desc = NULL;
1572 	}
1573 
1574 	if (eq->flags & EQ_MTX)
1575 		mutex_destroy(&eq->lock);
1576 
1577 	bzero(eq, sizeof (*eq));
1578 	return (0);
1579 }
1580 
1581 static int
alloc_txq(struct port_info * pi,struct sge_txq * txq,int idx)1582 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx)
1583 {
1584 	int rc, i;
1585 	struct adapter *sc = pi->adapter;
1586 	struct sge_eq *eq = &txq->eq;
1587 
1588 	rc = alloc_eq(sc, pi, eq);
1589 	if (rc != 0)
1590 		return (rc);
1591 
1592 	txq->port = pi;
1593 	txq->sdesc = kmem_zalloc(sizeof (struct tx_sdesc) * eq->cap, KM_SLEEP);
1594 	txq->txb_size = eq->qsize * tx_copy_threshold;
1595 	rc = alloc_tx_copybuffer(sc, txq->txb_size, &txq->txb_dhdl,
1596 	    &txq->txb_ahdl, &txq->txb_ba, &txq->txb_va);
1597 	if (rc == 0)
1598 		txq->txb_avail = txq->txb_size;
1599 	else
1600 		txq->txb_avail = txq->txb_size = 0;
1601 
1602 	/*
1603 	 * TODO: is this too low?  Worst case would need around 4 times qsize
1604 	 * (all tx descriptors filled to the brim with SGLs, with each entry in
1605 	 * the SGL coming from a distinct DMA handle).  Increase tx_dhdl_total
1606 	 * if you see too many dma_hdl_failed.
1607 	 */
1608 	txq->tx_dhdl_total = eq->qsize * 2;
1609 	txq->tx_dhdl = kmem_zalloc(sizeof (ddi_dma_handle_t) *
1610 	    txq->tx_dhdl_total, KM_SLEEP);
1611 	for (i = 0; i < txq->tx_dhdl_total; i++) {
1612 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
1613 		    DDI_DMA_SLEEP, 0, &txq->tx_dhdl[i]);
1614 		if (rc != DDI_SUCCESS) {
1615 			cxgb_printf(sc->dip, CE_WARN,
1616 			    "%s: failed to allocate DMA handle (%d)",
1617 			    __func__, rc);
1618 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
1619 		}
1620 		txq->tx_dhdl_avail++;
1621 	}
1622 
1623 	txq->ksp = setup_txq_kstats(pi, txq, idx);
1624 
1625 	return (rc);
1626 }
1627 
1628 static int
free_txq(struct port_info * pi,struct sge_txq * txq)1629 free_txq(struct port_info *pi, struct sge_txq *txq)
1630 {
1631 	int i;
1632 	struct adapter *sc = pi->adapter;
1633 	struct sge_eq *eq = &txq->eq;
1634 
1635 	if (txq->ksp != NULL) {
1636 		kstat_delete(txq->ksp);
1637 		txq->ksp = NULL;
1638 	}
1639 
1640 	if (txq->txb_va != NULL) {
1641 		(void) free_desc_ring(&txq->txb_dhdl, &txq->txb_ahdl);
1642 		txq->txb_va = NULL;
1643 	}
1644 
1645 	if (txq->sdesc != NULL) {
1646 		struct tx_sdesc *sd;
1647 		ddi_dma_handle_t hdl;
1648 
1649 		TXQ_LOCK(txq);
1650 		while (eq->cidx != eq->pidx) {
1651 			sd = &txq->sdesc[eq->cidx];
1652 
1653 			for (i = sd->hdls_used; i; i--) {
1654 				hdl = txq->tx_dhdl[txq->tx_dhdl_cidx];
1655 				(void) ddi_dma_unbind_handle(hdl);
1656 				if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
1657 					txq->tx_dhdl_cidx = 0;
1658 			}
1659 
1660 			ASSERT(sd->m);
1661 			freemsgchain(sd->m);
1662 
1663 			eq->cidx += sd->desc_used;
1664 			if (eq->cidx >= eq->cap)
1665 				eq->cidx -= eq->cap;
1666 
1667 			txq->txb_avail += txq->txb_used;
1668 		}
1669 		ASSERT(txq->tx_dhdl_cidx == txq->tx_dhdl_pidx);
1670 		ASSERT(txq->txb_avail == txq->txb_size);
1671 		TXQ_UNLOCK(txq);
1672 
1673 		kmem_free(txq->sdesc, sizeof (struct tx_sdesc) * eq->cap);
1674 		txq->sdesc = NULL;
1675 	}
1676 
1677 	if (txq->tx_dhdl != NULL) {
1678 		for (i = 0; i < txq->tx_dhdl_total; i++) {
1679 			if (txq->tx_dhdl[i] != NULL)
1680 				ddi_dma_free_handle(&txq->tx_dhdl[i]);
1681 		}
1682 		kmem_free(txq->tx_dhdl,
1683 		    sizeof (ddi_dma_handle_t) * txq->tx_dhdl_total);
1684 		txq->tx_dhdl = NULL;
1685 	}
1686 
1687 	(void) free_eq(sc, &txq->eq);
1688 
1689 	bzero(txq, sizeof (*txq));
1690 	return (0);
1691 }
1692 
1693 /*
1694  * Allocates a block of contiguous memory for DMA.  Can be used to allocate
1695  * memory for descriptor rings or for tx/rx copy buffers.
1696  *
1697  * Caller does not have to clean up anything if this function fails, it cleans
1698  * up after itself.
1699  *
1700  * Caller provides the following:
1701  * len		length of the block of memory to allocate.
1702  * flags	DDI_DMA_* flags to use (CONSISTENT/STREAMING, READ/WRITE/RDWR)
1703  * acc_attr	device access attributes for the allocation.
1704  * dma_attr	DMA attributes for the allocation
1705  *
1706  * If the function is successful it fills up this information:
1707  * dma_hdl	DMA handle for the allocated memory
1708  * acc_hdl	access handle for the allocated memory
1709  * ba		bus address of the allocated memory
1710  * va		KVA of the allocated memory.
1711  */
1712 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)1713 alloc_dma_memory(struct adapter *sc, size_t len, int flags,
1714     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
1715     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
1716     uint64_t *pba, caddr_t *pva)
1717 {
1718 	int rc;
1719 	ddi_dma_handle_t dhdl;
1720 	ddi_acc_handle_t ahdl;
1721 	ddi_dma_cookie_t cookie;
1722 	uint_t ccount;
1723 	caddr_t va;
1724 	size_t real_len;
1725 
1726 	*pva = NULL;
1727 
1728 	/*
1729 	 * DMA handle.
1730 	 */
1731 	rc = ddi_dma_alloc_handle(sc->dip, dma_attr, DDI_DMA_SLEEP, 0, &dhdl);
1732 	if (rc != DDI_SUCCESS) {
1733 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
1734 	}
1735 
1736 	/*
1737 	 * Memory suitable for DMA.
1738 	 */
1739 	rc = ddi_dma_mem_alloc(dhdl, len, acc_attr,
1740 	    flags & DDI_DMA_CONSISTENT ? DDI_DMA_CONSISTENT : DDI_DMA_STREAMING,
1741 	    DDI_DMA_SLEEP, 0, &va, &real_len, &ahdl);
1742 	if (rc != DDI_SUCCESS) {
1743 		ddi_dma_free_handle(&dhdl);
1744 		return (ENOMEM);
1745 	}
1746 
1747 	/*
1748 	 * DMA bindings.
1749 	 */
1750 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, va, real_len, flags, NULL,
1751 	    NULL, &cookie, &ccount);
1752 	if (rc != DDI_DMA_MAPPED) {
1753 		ddi_dma_mem_free(&ahdl);
1754 		ddi_dma_free_handle(&dhdl);
1755 		return (ENOMEM);
1756 	}
1757 	if (ccount != 1) {
1758 		/* unusable DMA mapping */
1759 		(void) free_desc_ring(&dhdl, &ahdl);
1760 		return (ENOMEM);
1761 	}
1762 
1763 	bzero(va, real_len);
1764 	*dma_hdl = dhdl;
1765 	*acc_hdl = ahdl;
1766 	*pba = cookie.dmac_laddress;
1767 	*pva = va;
1768 
1769 	return (0);
1770 }
1771 
1772 static int
free_dma_memory(ddi_dma_handle_t * dhdl,ddi_acc_handle_t * ahdl)1773 free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
1774 {
1775 	(void) ddi_dma_unbind_handle(*dhdl);
1776 	ddi_dma_mem_free(ahdl);
1777 	ddi_dma_free_handle(dhdl);
1778 
1779 	return (0);
1780 }
1781 
1782 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)1783 alloc_desc_ring(struct adapter *sc, size_t len, int rw,
1784     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
1785     uint64_t *pba, caddr_t *pva)
1786 {
1787 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_desc;
1788 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc;
1789 
1790 	return (alloc_dma_memory(sc, len, DDI_DMA_CONSISTENT | rw, acc_attr,
1791 	    dma_attr, dma_hdl, acc_hdl, pba, pva));
1792 }
1793 
1794 static int
free_desc_ring(ddi_dma_handle_t * dhdl,ddi_acc_handle_t * ahdl)1795 free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
1796 {
1797 	return (free_dma_memory(dhdl, ahdl));
1798 }
1799 
1800 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)1801 alloc_tx_copybuffer(struct adapter *sc, size_t len,
1802     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
1803     uint64_t *pba, caddr_t *pva)
1804 {
1805 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_tx;
1806 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc; /* NOT dma_attr_tx */
1807 
1808 	return (alloc_dma_memory(sc, len, DDI_DMA_STREAMING | DDI_DMA_WRITE,
1809 	    acc_attr, dma_attr, dma_hdl, acc_hdl, pba, pva));
1810 }
1811 
1812 static inline bool
is_new_response(const struct sge_iq * iq,struct rsp_ctrl ** ctrl)1813 is_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl)
1814 {
1815 	(void) ddi_dma_sync(iq->dhdl, (uintptr_t)iq->cdesc -
1816 	    (uintptr_t)iq->desc, iq->esize, DDI_DMA_SYNC_FORKERNEL);
1817 
1818 	*ctrl = (void *)((uintptr_t)iq->cdesc +
1819 	    (iq->esize - sizeof (struct rsp_ctrl)));
1820 
1821 	return ((((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen));
1822 }
1823 
1824 static inline void
iq_next(struct sge_iq * iq)1825 iq_next(struct sge_iq *iq)
1826 {
1827 	iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize);
1828 	if (++iq->cidx == iq->qsize - 1) {
1829 		iq->cidx = 0;
1830 		iq->gen ^= 1;
1831 		iq->cdesc = iq->desc;
1832 	}
1833 }
1834 
1835 /*
1836  * Fill up the freelist by upto nbufs and maybe ring its doorbell.
1837  *
1838  * Returns non-zero to indicate that it should be added to the list of starving
1839  * freelists.
1840  */
1841 static int
refill_fl(struct adapter * sc,struct sge_fl * fl,int nbufs)1842 refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs)
1843 {
1844 	uint64_t *d = &fl->desc[fl->pidx];
1845 	struct fl_sdesc *sd = &fl->sdesc[fl->pidx];
1846 
1847 	FL_LOCK_ASSERT_OWNED(fl);
1848 	ASSERT(nbufs >= 0);
1849 
1850 	if (nbufs > fl->needed)
1851 		nbufs = fl->needed;
1852 
1853 	while (nbufs--) {
1854 		if (sd->rxb != NULL) {
1855 			if (sd->rxb->ref_cnt == 1) {
1856 				/*
1857 				 * Buffer is available for recycling.  Two ways
1858 				 * this can happen:
1859 				 *
1860 				 * a) All the packets DMA'd into it last time
1861 				 *    around were within the rx_copy_threshold
1862 				 *    and no part of the buffer was ever passed
1863 				 *    up (ref_cnt never went over 1).
1864 				 *
1865 				 * b) Packets DMA'd into the buffer were passed
1866 				 *    up but have all been freed by the upper
1867 				 *    layers by now (ref_cnt went over 1 but is
1868 				 *    now back to 1).
1869 				 *
1870 				 * Either way the bus address in the descriptor
1871 				 * ring is already valid.
1872 				 */
1873 				ASSERT(*d == cpu_to_be64(sd->rxb->ba));
1874 				d++;
1875 				goto recycled;
1876 			} else {
1877 				/*
1878 				 * Buffer still in use and we need a
1879 				 * replacement. But first release our reference
1880 				 * on the existing buffer.
1881 				 */
1882 				rxbuf_free(sd->rxb);
1883 			}
1884 		}
1885 
1886 		sd->rxb = rxbuf_alloc(sc->sge.rxbuf_cache, KM_NOSLEEP, 1);
1887 		if (sd->rxb == NULL)
1888 			break;
1889 		*d++ = cpu_to_be64(sd->rxb->ba);
1890 
1891 recycled:	fl->pending++;
1892 		sd++;
1893 		fl->needed--;
1894 		if (++fl->pidx == fl->cap) {
1895 			fl->pidx = 0;
1896 			sd = fl->sdesc;
1897 			d = fl->desc;
1898 		}
1899 	}
1900 
1901 	if (fl->pending >= 8)
1902 		ring_fl_db(sc, fl);
1903 
1904 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
1905 }
1906 
1907 #ifndef TAILQ_FOREACH_SAFE
1908 #define	TAILQ_FOREACH_SAFE(var, head, field, tvar)			\
1909 	for ((var) = TAILQ_FIRST((head));				\
1910 	    (var) && ((tvar) = TAILQ_NEXT((var), field), 1);		\
1911 	    (var) = (tvar))
1912 #endif
1913 
1914 /*
1915  * Attempt to refill all starving freelists.
1916  */
1917 static void
refill_sfl(void * arg)1918 refill_sfl(void *arg)
1919 {
1920 	struct adapter *sc = arg;
1921 	struct sge_fl *fl, *fl_temp;
1922 
1923 	mutex_enter(&sc->sfl_lock);
1924 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
1925 		FL_LOCK(fl);
1926 		(void) refill_fl(sc, fl, 64);
1927 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
1928 			TAILQ_REMOVE(&sc->sfl, fl, link);
1929 			fl->flags &= ~FL_STARVING;
1930 		}
1931 		FL_UNLOCK(fl);
1932 	}
1933 
1934 	if (!TAILQ_EMPTY(&sc->sfl) != 0)
1935 		sc->sfl_timer =  timeout(refill_sfl, sc, drv_usectohz(100000));
1936 	mutex_exit(&sc->sfl_lock);
1937 }
1938 
1939 static void
add_fl_to_sfl(struct adapter * sc,struct sge_fl * fl)1940 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
1941 {
1942 	mutex_enter(&sc->sfl_lock);
1943 	FL_LOCK(fl);
1944 	if ((fl->flags & FL_DOOMED) == 0) {
1945 		if (TAILQ_EMPTY(&sc->sfl) != 0) {
1946 			sc->sfl_timer = timeout(refill_sfl, sc,
1947 			    drv_usectohz(100000));
1948 		}
1949 		fl->flags |= FL_STARVING;
1950 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
1951 	}
1952 	FL_UNLOCK(fl);
1953 	mutex_exit(&sc->sfl_lock);
1954 }
1955 
1956 static void
free_fl_bufs(struct sge_fl * fl)1957 free_fl_bufs(struct sge_fl *fl)
1958 {
1959 	struct fl_sdesc *sd;
1960 	unsigned int i;
1961 
1962 	FL_LOCK_ASSERT_OWNED(fl);
1963 
1964 	for (i = 0; i < fl->cap; i++) {
1965 		sd = &fl->sdesc[i];
1966 
1967 		if (sd->rxb != NULL) {
1968 			rxbuf_free(sd->rxb);
1969 			sd->rxb = NULL;
1970 		}
1971 	}
1972 }
1973 
1974 /*
1975  * Note that fl->cidx and fl->offset are left unchanged in case of failure.
1976  */
1977 static mblk_t *
get_fl_payload(struct adapter * sc,struct sge_fl * fl,uint32_t len_newbuf,int * fl_bufs_used)1978 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf,
1979     int *fl_bufs_used)
1980 {
1981 	struct mblk_pair frame = {0};
1982 	struct rxbuf *rxb;
1983 	mblk_t *m = NULL;
1984 	uint_t nbuf = 0, len, copy, n;
1985 	uint32_t cidx, offset, rcidx, roffset;
1986 
1987 	/*
1988 	 * The SGE won't pack a new frame into the current buffer if the entire
1989 	 * payload doesn't fit in the remaining space.  Move on to the next buf
1990 	 * in that case.
1991 	 */
1992 	rcidx = fl->cidx;
1993 	roffset = fl->offset;
1994 	if (fl->offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1995 		fl->offset = 0;
1996 		if (++fl->cidx == fl->cap)
1997 			fl->cidx = 0;
1998 		nbuf++;
1999 	}
2000 	cidx = fl->cidx;
2001 	offset = fl->offset;
2002 
2003 	len = G_RSPD_LEN(len_newbuf);	/* pktshift + payload length */
2004 	copy = (len <= fl->copy_threshold);
2005 	if (copy != 0) {
2006 		frame.head = m = allocb(len, BPRI_HI);
2007 		if (m == NULL) {
2008 			fl->allocb_fail++;
2009 			DTRACE_PROBE1(t4__fl_alloc_fail, struct sge_fl *, fl);
2010 			fl->cidx = rcidx;
2011 			fl->offset = roffset;
2012 			return (NULL);
2013 		}
2014 	}
2015 
2016 	while (len) {
2017 		rxb = fl->sdesc[cidx].rxb;
2018 		n = min(len, rxb->buf_size - offset);
2019 
2020 		(void) ddi_dma_sync(rxb->dhdl, offset, n,
2021 		    DDI_DMA_SYNC_FORKERNEL);
2022 
2023 		if (copy != 0)
2024 			bcopy(rxb->va + offset, m->b_wptr, n);
2025 		else {
2026 			m = desballoc((unsigned char *)rxb->va + offset, n,
2027 			    BPRI_HI, &rxb->freefunc);
2028 			if (m == NULL) {
2029 				fl->allocb_fail++;
2030 				DTRACE_PROBE1(t4__fl_alloc_fail,
2031 				    struct sge_fl *, fl);
2032 				if (frame.head)
2033 					freemsgchain(frame.head);
2034 				fl->cidx = rcidx;
2035 				fl->offset = roffset;
2036 				return (NULL);
2037 			}
2038 			atomic_inc_uint(&rxb->ref_cnt);
2039 			if (frame.head != NULL)
2040 				frame.tail->b_cont = m;
2041 			else
2042 				frame.head = m;
2043 			frame.tail = m;
2044 		}
2045 		m->b_wptr += n;
2046 		len -= n;
2047 		offset += roundup(n, sc->sge.fl_align);
2048 		ASSERT(offset <= rxb->buf_size);
2049 		if (offset == rxb->buf_size) {
2050 			offset = 0;
2051 			if (++cidx == fl->cap)
2052 				cidx = 0;
2053 			nbuf++;
2054 		}
2055 	}
2056 
2057 	fl->cidx = cidx;
2058 	fl->offset = offset;
2059 	(*fl_bufs_used) += nbuf;
2060 
2061 	ASSERT(frame.head != NULL);
2062 	return (frame.head);
2063 }
2064 
2065 /*
2066  * We'll do immediate data tx for non-LSO, but only when not coalescing.  We're
2067  * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes
2068  * of immediate data.
2069  */
2070 #define	IMM_LEN ( \
2071 	2 * EQ_ESIZE \
2072 	- sizeof (struct fw_eth_tx_pkt_wr) \
2073 	- sizeof (struct cpl_tx_pkt_core))
2074 
2075 /*
2076  * Returns non-zero on failure, no need to cleanup anything in that case.
2077  *
2078  * Note 1: We always try to pull up the mblk if required and return E2BIG only
2079  * if this fails.
2080  *
2081  * Note 2: We'll also pullup incoming mblk if HW_LSO is set and the first mblk
2082  * does not have the TCP header in it.
2083  */
2084 static int
get_frame_txinfo(struct sge_txq * txq,mblk_t ** fp,struct txinfo * txinfo,int sgl_only)2085 get_frame_txinfo(struct sge_txq *txq, mblk_t **fp, struct txinfo *txinfo,
2086     int sgl_only)
2087 {
2088 	uint32_t flags = 0, len, n;
2089 	mblk_t *m = *fp;
2090 	int rc;
2091 
2092 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb and dma_hdls */
2093 
2094 	mac_hcksum_get(m, NULL, NULL, NULL, NULL, &flags);
2095 	txinfo->flags = (flags & HCK_TX_FLAGS);
2096 
2097 	mac_lso_get(m, &txinfo->mss, &flags);
2098 	txinfo->flags |= (flags & HW_LSO_FLAGS);
2099 
2100 	if (flags & HW_LSO)
2101 		sgl_only = 1;	/* Do not allow immediate data with LSO */
2102 
2103 	/*
2104 	 * If checksum or segmentation offloads are requested, gather
2105 	 * information about the sizes and types of headers in the packet.
2106 	 */
2107 	if (txinfo->flags != 0) {
2108 		/*
2109 		 * Even if this fails, the meoi_flags field will be capable of
2110 		 * communicating the lack of useful packet information.
2111 		 */
2112 		(void) mac_ether_offload_info(m, &txinfo->meoi);
2113 	} else {
2114 		bzero(&txinfo->meoi, sizeof (txinfo->meoi));
2115 	}
2116 
2117 start:	txinfo->nsegs = 0;
2118 	txinfo->hdls_used = 0;
2119 	txinfo->txb_used = 0;
2120 	txinfo->len = 0;
2121 
2122 	/* total length and a rough estimate of # of segments */
2123 	n = 0;
2124 	for (; m; m = m->b_cont) {
2125 		len = MBLKL(m);
2126 		n += (len / PAGE_SIZE) + 1;
2127 		txinfo->len += len;
2128 	}
2129 	m = *fp;
2130 
2131 	if (n >= TX_SGL_SEGS || (flags & HW_LSO && MBLKL(m) < 50)) {
2132 		txq->pullup_early++;
2133 		m = msgpullup(*fp, -1);
2134 		if (m == NULL) {
2135 			txq->pullup_failed++;
2136 			return (E2BIG);	/* (*fp) left as it was */
2137 		}
2138 		freemsg(*fp);
2139 		*fp = m;
2140 		mac_hcksum_set(m, 0, 0, 0, 0, txinfo->flags);
2141 	}
2142 
2143 	if (txinfo->len <= IMM_LEN && !sgl_only)
2144 		return (0);	/* nsegs = 0 tells caller to use imm. tx */
2145 
2146 	if (txinfo->len <= txq->copy_threshold &&
2147 	    copy_into_txb(txq, m, txinfo->len, txinfo) == 0)
2148 		goto done;
2149 
2150 	for (; m; m = m->b_cont) {
2151 
2152 		len = MBLKL(m);
2153 
2154 		/* Use tx copy buffer if this mblk is small enough */
2155 		if (len <= txq->copy_threshold &&
2156 		    copy_into_txb(txq, m, len, txinfo) == 0)
2157 			continue;
2158 
2159 		/* Add DMA bindings for this mblk to the SGL */
2160 		rc = add_mblk(txq, txinfo, m, len);
2161 
2162 		if (rc == E2BIG ||
2163 		    (txinfo->nsegs == TX_SGL_SEGS && m->b_cont)) {
2164 
2165 			txq->pullup_late++;
2166 			m = msgpullup(*fp, -1);
2167 			if (m != NULL) {
2168 				free_txinfo_resources(txq, txinfo);
2169 				freemsg(*fp);
2170 				*fp = m;
2171 				mac_hcksum_set(m, 0, 0, 0, 0, txinfo->flags);
2172 				goto start;
2173 			}
2174 
2175 			txq->pullup_failed++;
2176 			rc = E2BIG;
2177 		}
2178 
2179 		if (rc != 0) {
2180 			free_txinfo_resources(txq, txinfo);
2181 			return (rc);
2182 		}
2183 	}
2184 
2185 	ASSERT(txinfo->nsegs > 0 && txinfo->nsegs <= TX_SGL_SEGS);
2186 
2187 done:
2188 
2189 	/*
2190 	 * Store the # of flits required to hold this frame's SGL in nflits.  An
2191 	 * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by
2192 	 * multiple (len0 + len1, addr0, addr1) tuples.  If addr1 is not used
2193 	 * then len1 must be set to 0.
2194 	 */
2195 	n = txinfo->nsegs - 1;
2196 	txinfo->nflits = (3 * n) / 2 + (n & 1) + 2;
2197 	if (n & 1)
2198 		txinfo->sgl.sge[n / 2].len[1] = cpu_to_be32(0);
2199 
2200 	txinfo->sgl.cmd_nsge = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_DSGL) |
2201 	    V_ULPTX_NSGE(txinfo->nsegs));
2202 
2203 	return (0);
2204 }
2205 
2206 static inline int
fits_in_txb(struct sge_txq * txq,int len,int * waste)2207 fits_in_txb(struct sge_txq *txq, int len, int *waste)
2208 {
2209 	if (txq->txb_avail < len)
2210 		return (0);
2211 
2212 	if (txq->txb_next + len <= txq->txb_size) {
2213 		*waste = 0;
2214 		return (1);
2215 	}
2216 
2217 	*waste = txq->txb_size - txq->txb_next;
2218 
2219 	return (txq->txb_avail - *waste < len ? 0 : 1);
2220 }
2221 
2222 #define	TXB_CHUNK	64
2223 
2224 /*
2225  * Copies the specified # of bytes into txq's tx copy buffer and updates txinfo
2226  * and txq to indicate resources used.  Caller has to make sure that those many
2227  * bytes are available in the mblk chain (b_cont linked).
2228  */
2229 static inline int
copy_into_txb(struct sge_txq * txq,mblk_t * m,int len,struct txinfo * txinfo)2230 copy_into_txb(struct sge_txq *txq, mblk_t *m, int len, struct txinfo *txinfo)
2231 {
2232 	int waste, n;
2233 
2234 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb */
2235 
2236 	if (!fits_in_txb(txq, len, &waste)) {
2237 		txq->txb_full++;
2238 		return (ENOMEM);
2239 	}
2240 
2241 	if (waste != 0) {
2242 		ASSERT((waste & (TXB_CHUNK - 1)) == 0);
2243 		txinfo->txb_used += waste;
2244 		txq->txb_avail -= waste;
2245 		txq->txb_next = 0;
2246 	}
2247 
2248 	for (n = 0; n < len; m = m->b_cont) {
2249 		bcopy(m->b_rptr, txq->txb_va + txq->txb_next + n, MBLKL(m));
2250 		n += MBLKL(m);
2251 	}
2252 
2253 	add_seg(txinfo, txq->txb_ba + txq->txb_next, len);
2254 
2255 	n = roundup(len, TXB_CHUNK);
2256 	txinfo->txb_used += n;
2257 	txq->txb_avail -= n;
2258 	txq->txb_next += n;
2259 	ASSERT(txq->txb_next <= txq->txb_size);
2260 	if (txq->txb_next == txq->txb_size)
2261 		txq->txb_next = 0;
2262 
2263 	return (0);
2264 }
2265 
2266 static inline void
add_seg(struct txinfo * txinfo,uint64_t ba,uint32_t len)2267 add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len)
2268 {
2269 	ASSERT(txinfo->nsegs < TX_SGL_SEGS);	/* must have room */
2270 
2271 	if (txinfo->nsegs != 0) {
2272 		int idx = txinfo->nsegs - 1;
2273 		txinfo->sgl.sge[idx / 2].len[idx & 1] = cpu_to_be32(len);
2274 		txinfo->sgl.sge[idx / 2].addr[idx & 1] = cpu_to_be64(ba);
2275 	} else {
2276 		txinfo->sgl.len0 = cpu_to_be32(len);
2277 		txinfo->sgl.addr0 = cpu_to_be64(ba);
2278 	}
2279 	txinfo->nsegs++;
2280 }
2281 
2282 /*
2283  * This function cleans up any partially allocated resources when it fails so
2284  * there's nothing for the caller to clean up in that case.
2285  *
2286  * EIO indicates permanent failure.  Caller should drop the frame containing
2287  * this mblk and continue.
2288  *
2289  * E2BIG indicates that the SGL length for this mblk exceeds the hardware
2290  * limit.  Caller should pull up the frame before trying to send it out.
2291  * (This error means our pullup_early heuristic did not work for this frame)
2292  *
2293  * ENOMEM indicates a temporary shortage of resources (DMA handles, other DMA
2294  * resources, etc.).  Caller should suspend the tx queue and wait for reclaim to
2295  * free up resources.
2296  */
2297 static inline int
add_mblk(struct sge_txq * txq,struct txinfo * txinfo,mblk_t * m,int len)2298 add_mblk(struct sge_txq *txq, struct txinfo *txinfo, mblk_t *m, int len)
2299 {
2300 	ddi_dma_handle_t dhdl;
2301 	ddi_dma_cookie_t cookie;
2302 	uint_t ccount = 0;
2303 	int rc;
2304 
2305 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate dhdls */
2306 
2307 	if (txq->tx_dhdl_avail == 0) {
2308 		txq->dma_hdl_failed++;
2309 		return (ENOMEM);
2310 	}
2311 
2312 	dhdl = txq->tx_dhdl[txq->tx_dhdl_pidx];
2313 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, (caddr_t)m->b_rptr, len,
2314 	    DDI_DMA_WRITE | DDI_DMA_STREAMING, DDI_DMA_DONTWAIT, NULL, &cookie,
2315 	    &ccount);
2316 	if (rc != DDI_DMA_MAPPED) {
2317 		txq->dma_map_failed++;
2318 
2319 		ASSERT(rc != DDI_DMA_INUSE && rc != DDI_DMA_PARTIAL_MAP);
2320 
2321 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EIO);
2322 	}
2323 
2324 	if (ccount + txinfo->nsegs > TX_SGL_SEGS) {
2325 		(void) ddi_dma_unbind_handle(dhdl);
2326 		return (E2BIG);
2327 	}
2328 
2329 	add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2330 	while (--ccount) {
2331 		ddi_dma_nextcookie(dhdl, &cookie);
2332 		add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2333 	}
2334 
2335 	if (++txq->tx_dhdl_pidx == txq->tx_dhdl_total)
2336 		txq->tx_dhdl_pidx = 0;
2337 	txq->tx_dhdl_avail--;
2338 	txinfo->hdls_used++;
2339 
2340 	return (0);
2341 }
2342 
2343 /*
2344  * Releases all the txq resources used up in the specified txinfo.
2345  */
2346 static void
free_txinfo_resources(struct sge_txq * txq,struct txinfo * txinfo)2347 free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo)
2348 {
2349 	int n;
2350 
2351 	TXQ_LOCK_ASSERT_OWNED(txq);	/* dhdls, txb */
2352 
2353 	n = txinfo->txb_used;
2354 	if (n > 0) {
2355 		txq->txb_avail += n;
2356 		if (n <= txq->txb_next)
2357 			txq->txb_next -= n;
2358 		else {
2359 			n -= txq->txb_next;
2360 			txq->txb_next = txq->txb_size - n;
2361 		}
2362 	}
2363 
2364 	for (n = txinfo->hdls_used; n > 0; n--) {
2365 		if (txq->tx_dhdl_pidx > 0)
2366 			txq->tx_dhdl_pidx--;
2367 		else
2368 			txq->tx_dhdl_pidx = txq->tx_dhdl_total - 1;
2369 		txq->tx_dhdl_avail++;
2370 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_pidx]);
2371 	}
2372 }
2373 
2374 /*
2375  * Returns 0 to indicate that m has been accepted into a coalesced tx work
2376  * request.  It has either been folded into txpkts or txpkts was flushed and m
2377  * has started a new coalesced work request (as the first frame in a fresh
2378  * txpkts).
2379  *
2380  * Returns non-zero to indicate a failure - caller is responsible for
2381  * transmitting m, if there was anything in txpkts it has been flushed.
2382  */
2383 static int
add_to_txpkts(struct sge_txq * txq,struct txpkts * txpkts,mblk_t * m,struct txinfo * txinfo)2384 add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
2385     struct txinfo *txinfo)
2386 {
2387 	struct sge_eq *eq = &txq->eq;
2388 	int can_coalesce;
2389 	struct tx_sdesc *txsd;
2390 	uint8_t flits;
2391 
2392 	TXQ_LOCK_ASSERT_OWNED(txq);
2393 
2394 	if (txpkts->npkt > 0) {
2395 		flits = TXPKTS_PKT_HDR + txinfo->nflits;
2396 		can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2397 		    txpkts->nflits + flits <= TX_WR_FLITS &&
2398 		    txpkts->nflits + flits <= eq->avail * 8 &&
2399 		    txpkts->plen + txinfo->len < 65536;
2400 
2401 		if (can_coalesce != 0) {
2402 			txpkts->tail->b_next = m;
2403 			txpkts->tail = m;
2404 			txpkts->npkt++;
2405 			txpkts->nflits += flits;
2406 			txpkts->plen += txinfo->len;
2407 
2408 			txsd = &txq->sdesc[eq->pidx];
2409 			txsd->txb_used += txinfo->txb_used;
2410 			txsd->hdls_used += txinfo->hdls_used;
2411 
2412 			return (0);
2413 		}
2414 
2415 		/*
2416 		 * Couldn't coalesce m into txpkts.  The first order of business
2417 		 * is to send txpkts on its way.  Then we'll revisit m.
2418 		 */
2419 		write_txpkts_wr(txq, txpkts);
2420 	}
2421 
2422 	/*
2423 	 * Check if we can start a new coalesced tx work request with m as
2424 	 * the first packet in it.
2425 	 */
2426 
2427 	ASSERT(txpkts->npkt == 0);
2428 	ASSERT(txinfo->len < 65536);
2429 
2430 	flits = TXPKTS_WR_HDR + txinfo->nflits;
2431 	can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2432 	    flits <= eq->avail * 8 && flits <= TX_WR_FLITS;
2433 
2434 	if (can_coalesce == 0)
2435 		return (EINVAL);
2436 
2437 	/*
2438 	 * Start a fresh coalesced tx WR with m as the first frame in it.
2439 	 */
2440 	txpkts->tail = m;
2441 	txpkts->npkt = 1;
2442 	txpkts->nflits = flits;
2443 	txpkts->flitp = &eq->desc[eq->pidx].flit[2];
2444 	txpkts->plen = txinfo->len;
2445 
2446 	txsd = &txq->sdesc[eq->pidx];
2447 	txsd->m = m;
2448 	txsd->txb_used = txinfo->txb_used;
2449 	txsd->hdls_used = txinfo->hdls_used;
2450 
2451 	return (0);
2452 }
2453 
2454 /*
2455  * Note that write_txpkts_wr can never run out of hardware descriptors (but
2456  * write_txpkt_wr can).  add_to_txpkts ensures that a frame is accepted for
2457  * coalescing only if sufficient hardware descriptors are available.
2458  */
2459 static void
write_txpkts_wr(struct sge_txq * txq,struct txpkts * txpkts)2460 write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts)
2461 {
2462 	struct sge_eq *eq = &txq->eq;
2463 	struct fw_eth_tx_pkts_wr *wr;
2464 	struct tx_sdesc *txsd;
2465 	uint32_t ctrl;
2466 	uint16_t ndesc;
2467 
2468 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2469 
2470 	ndesc = howmany(txpkts->nflits, 8);
2471 
2472 	wr = (void *)&eq->desc[eq->pidx];
2473 	wr->op_pkd = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR) |
2474 	    V_FW_WR_IMMDLEN(0)); /* immdlen does not matter in this WR */
2475 	ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2));
2476 	if (eq->avail == ndesc)
2477 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2478 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2479 	wr->plen = cpu_to_be16(txpkts->plen);
2480 	wr->npkt = txpkts->npkt;
2481 	wr->r3 = wr->type = 0;
2482 
2483 	/* Everything else already written */
2484 
2485 	txsd = &txq->sdesc[eq->pidx];
2486 	txsd->desc_used = ndesc;
2487 
2488 	txq->txb_used += txsd->txb_used / TXB_CHUNK;
2489 	txq->hdl_used += txsd->hdls_used;
2490 
2491 	ASSERT(eq->avail >= ndesc);
2492 
2493 	eq->pending += ndesc;
2494 	eq->avail -= ndesc;
2495 	eq->pidx += ndesc;
2496 	if (eq->pidx >= eq->cap)
2497 		eq->pidx -= eq->cap;
2498 
2499 	txq->txpkts_pkts += txpkts->npkt;
2500 	txq->txpkts_wrs++;
2501 	txpkts->npkt = 0;	/* emptied */
2502 }
2503 
2504 typedef enum {
2505 	COS_SUCCESS,	/* ctrl flit contains proper bits for csum offload */
2506 	COS_IGNORE,	/* no csum offload requested */
2507 	COS_FAIL,	/* csum offload requested, but pkt data missing */
2508 } csum_offload_status_t;
2509 /*
2510  * Build a ctrl1 flit for checksum offload in CPL_TX_PKT_XT command
2511  */
2512 static csum_offload_status_t
csum_to_ctrl(const struct txinfo * txinfo,uint32_t chip_version,uint64_t * ctrlp)2513 csum_to_ctrl(const struct txinfo *txinfo, uint32_t chip_version,
2514     uint64_t *ctrlp)
2515 {
2516 	const mac_ether_offload_info_t *meoi = &txinfo->meoi;
2517 	const uint32_t tx_flags = txinfo->flags;
2518 	const boolean_t needs_l3_csum = (tx_flags & HW_LSO) != 0 ||
2519 	    (tx_flags & HCK_IPV4_HDRCKSUM) != 0;
2520 	const boolean_t needs_l4_csum = (tx_flags & HW_LSO) != 0 ||
2521 	    (tx_flags & (HCK_FULLCKSUM | HCK_PARTIALCKSUM)) != 0;
2522 
2523 	/*
2524 	 * Default to disabling any checksumming both for cases where it is not
2525 	 * requested, but also if we cannot appropriately interrogate the
2526 	 * required information from the packet.
2527 	 */
2528 	uint64_t ctrl = F_TXPKT_L4CSUM_DIS | F_TXPKT_IPCSUM_DIS;
2529 	if (!needs_l3_csum && !needs_l4_csum) {
2530 		*ctrlp = ctrl;
2531 		return (COS_IGNORE);
2532 	}
2533 
2534 	if (needs_l3_csum) {
2535 		/* Only IPv4 checksums are supported (for L3) */
2536 		if ((meoi->meoi_flags & MEOI_L3INFO_SET) == 0 ||
2537 		    meoi->meoi_l3proto != ETHERTYPE_IP) {
2538 			*ctrlp = ctrl;
2539 			return (COS_FAIL);
2540 		}
2541 		ctrl &= ~F_TXPKT_IPCSUM_DIS;
2542 	}
2543 
2544 	if (needs_l4_csum) {
2545 		/*
2546 		 * We need at least all of the L3 header to make decisions about
2547 		 * the contained L4 protocol.  If not all of the L4 information
2548 		 * is present, we will leave it to the NIC to checksum all it is
2549 		 * able to.
2550 		 */
2551 		if ((meoi->meoi_flags & MEOI_L3INFO_SET) == 0) {
2552 			*ctrlp = ctrl;
2553 			return (COS_FAIL);
2554 		}
2555 
2556 		/*
2557 		 * Since we are parsing the packet anyways, make the checksum
2558 		 * decision based on the L4 protocol, rather than using the
2559 		 * Generic TCP/UDP checksum using start & end offsets in the
2560 		 * packet (like requested with PARTIALCKSUM).
2561 		 */
2562 		int csum_type = -1;
2563 		if (meoi->meoi_l3proto == ETHERTYPE_IP &&
2564 		    meoi->meoi_l4proto == IPPROTO_TCP) {
2565 			csum_type = TX_CSUM_TCPIP;
2566 		} else if (meoi->meoi_l3proto == ETHERTYPE_IPV6 &&
2567 		    meoi->meoi_l4proto == IPPROTO_TCP) {
2568 			csum_type = TX_CSUM_TCPIP6;
2569 		} else if (meoi->meoi_l3proto == ETHERTYPE_IP &&
2570 		    meoi->meoi_l4proto == IPPROTO_UDP) {
2571 			csum_type = TX_CSUM_UDPIP;
2572 		} else if (meoi->meoi_l3proto == ETHERTYPE_IPV6 &&
2573 		    meoi->meoi_l4proto == IPPROTO_UDP) {
2574 			csum_type = TX_CSUM_UDPIP6;
2575 		} else {
2576 			*ctrlp = ctrl;
2577 			return (COS_FAIL);
2578 		}
2579 
2580 		ASSERT(csum_type != -1);
2581 		ctrl &= ~F_TXPKT_L4CSUM_DIS;
2582 		ctrl |= V_TXPKT_CSUM_TYPE(csum_type);
2583 	}
2584 
2585 	if ((ctrl & F_TXPKT_IPCSUM_DIS) == 0 &&
2586 	    (ctrl & F_TXPKT_L4CSUM_DIS) != 0) {
2587 		/*
2588 		 * If only the IPv4 checksum is requested, we need to set an
2589 		 * appropriate type in the command for it.
2590 		 */
2591 		ctrl |= V_TXPKT_CSUM_TYPE(TX_CSUM_IP);
2592 	}
2593 
2594 	ASSERT(ctrl != (F_TXPKT_L4CSUM_DIS | F_TXPKT_IPCSUM_DIS));
2595 
2596 	/*
2597 	 * Fill in the requisite L2/L3 header length data.
2598 	 *
2599 	 * The Ethernet header length is recorded as 'size - 14 bytes'
2600 	 */
2601 	const uint8_t eth_len = meoi->meoi_l2hlen - 14;
2602 	if (chip_version >= CHELSIO_T6) {
2603 		ctrl |= V_T6_TXPKT_ETHHDR_LEN(eth_len);
2604 	} else {
2605 		ctrl |= V_TXPKT_ETHHDR_LEN(eth_len);
2606 	}
2607 	ctrl |= V_TXPKT_IPHDR_LEN(meoi->meoi_l3hlen);
2608 
2609 	*ctrlp = ctrl;
2610 	return (COS_SUCCESS);
2611 }
2612 
2613 static int
write_txpkt_wr(struct port_info * pi,struct sge_txq * txq,mblk_t * m,struct txinfo * txinfo)2614 write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
2615     struct txinfo *txinfo)
2616 {
2617 	struct sge_eq *eq = &txq->eq;
2618 	struct fw_eth_tx_pkt_wr *wr;
2619 	struct cpl_tx_pkt_core *cpl;
2620 	uint32_t ctrl;	/* used in many unrelated places */
2621 	uint64_t ctrl1;
2622 	int nflits, ndesc;
2623 	struct tx_sdesc *txsd;
2624 	caddr_t dst;
2625 	const mac_ether_offload_info_t *meoi = &txinfo->meoi;
2626 
2627 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2628 
2629 	/*
2630 	 * Do we have enough flits to send this frame out?
2631 	 */
2632 	ctrl = sizeof (struct cpl_tx_pkt_core);
2633 	if (txinfo->flags & HW_LSO) {
2634 		nflits = TXPKT_LSO_WR_HDR;
2635 		ctrl += sizeof (struct cpl_tx_pkt_lso_core);
2636 	} else {
2637 		nflits = TXPKT_WR_HDR;
2638 	}
2639 	if (txinfo->nsegs > 0)
2640 		nflits += txinfo->nflits;
2641 	else {
2642 		nflits += howmany(txinfo->len, 8);
2643 		ctrl += txinfo->len;
2644 	}
2645 	ndesc = howmany(nflits, 8);
2646 	if (ndesc > eq->avail)
2647 		return (ENOMEM);
2648 
2649 	/* Firmware work request header */
2650 	wr = (void *)&eq->desc[eq->pidx];
2651 	wr->op_immdlen = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
2652 	    V_FW_WR_IMMDLEN(ctrl));
2653 	ctrl = V_FW_WR_LEN16(howmany(nflits, 2));
2654 	if (eq->avail == ndesc)
2655 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2656 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2657 	wr->r3 = 0;
2658 
2659 	if (txinfo->flags & HW_LSO &&
2660 	    (meoi->meoi_flags & MEOI_L4INFO_SET) != 0 &&
2661 	    meoi->meoi_l4proto == IPPROTO_TCP) {
2662 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
2663 
2664 		ctrl = V_LSO_OPCODE((u32)CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
2665 		    F_LSO_LAST_SLICE;
2666 
2667 		if (meoi->meoi_l2hlen > sizeof (struct ether_header)) {
2668 			/*
2669 			 * This presently assumes a standard VLAN header,
2670 			 * without support for Q-in-Q.
2671 			 */
2672 			ctrl |= V_LSO_ETHHDR_LEN(1);
2673 		}
2674 
2675 		switch (meoi->meoi_l3proto) {
2676 		case ETHERTYPE_IPV6:
2677 			ctrl |= F_LSO_IPV6;
2678 			/* FALLTHROUGH */
2679 		case ETHERTYPE_IP:
2680 			ctrl |= V_LSO_IPHDR_LEN(meoi->meoi_l3hlen / 4);
2681 			break;
2682 		default:
2683 			break;
2684 		}
2685 
2686 		ctrl |= V_LSO_TCPHDR_LEN(meoi->meoi_l4hlen / 4);
2687 
2688 		lso->lso_ctrl = cpu_to_be32(ctrl);
2689 		lso->ipid_ofst = cpu_to_be16(0);
2690 		lso->mss = cpu_to_be16(txinfo->mss);
2691 		lso->seqno_offset = cpu_to_be32(0);
2692 		if (is_t4(pi->adapter->params.chip))
2693 			lso->len = cpu_to_be32(txinfo->len);
2694 		else
2695 			lso->len = cpu_to_be32(V_LSO_T5_XFER_SIZE(txinfo->len));
2696 
2697 		cpl = (void *)(lso + 1);
2698 
2699 		txq->tso_wrs++;
2700 	} else {
2701 		cpl = (void *)(wr + 1);
2702 	}
2703 
2704 	/* Checksum offload */
2705 	switch (csum_to_ctrl(txinfo,
2706 	    CHELSIO_CHIP_VERSION(pi->adapter->params.chip), &ctrl1)) {
2707 	case COS_SUCCESS:
2708 		txq->txcsum++;
2709 		break;
2710 	case COS_FAIL:
2711 		/*
2712 		 * Packet will be going out with checksums which are probably
2713 		 * wrong but there is little we can do now.
2714 		 */
2715 		txq->csum_failed++;
2716 		break;
2717 	default:
2718 		break;
2719 	}
2720 
2721 	/* CPL header */
2722 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
2723 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
2724 	cpl->pack = 0;
2725 	cpl->len = cpu_to_be16(txinfo->len);
2726 	cpl->ctrl1 = cpu_to_be64(ctrl1);
2727 
2728 	/* Software descriptor */
2729 	txsd = &txq->sdesc[eq->pidx];
2730 	txsd->m = m;
2731 	txsd->txb_used = txinfo->txb_used;
2732 	txsd->hdls_used = txinfo->hdls_used;
2733 	/* LINTED: E_ASSIGN_NARROW_CONV */
2734 	txsd->desc_used = ndesc;
2735 
2736 	txq->txb_used += txinfo->txb_used / TXB_CHUNK;
2737 	txq->hdl_used += txinfo->hdls_used;
2738 
2739 	eq->pending += ndesc;
2740 	eq->avail -= ndesc;
2741 	eq->pidx += ndesc;
2742 	if (eq->pidx >= eq->cap)
2743 		eq->pidx -= eq->cap;
2744 
2745 	/* SGL */
2746 	dst = (void *)(cpl + 1);
2747 	if (txinfo->nsegs > 0) {
2748 		txq->sgl_wrs++;
2749 		copy_to_txd(eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
2750 
2751 		/* Need to zero-pad to a 16 byte boundary if not on one */
2752 		if ((uintptr_t)dst & 0xf)
2753 			/* LINTED: E_BAD_PTR_CAST_ALIGN */
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 inline void
write_ulp_cpl_sgl(struct port_info * pi,struct sge_txq * txq,struct txpkts * txpkts,struct txinfo * txinfo)2775 write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
2776     struct txpkts *txpkts, struct txinfo *txinfo)
2777 {
2778 	struct ulp_txpkt *ulpmc;
2779 	struct ulptx_idata *ulpsc;
2780 	struct cpl_tx_pkt_core *cpl;
2781 	uintptr_t flitp, start, end;
2782 	uint64_t ctrl;
2783 	caddr_t dst;
2784 
2785 	ASSERT(txpkts->npkt > 0);
2786 
2787 	start = (uintptr_t)txq->eq.desc;
2788 	end = (uintptr_t)txq->eq.spg;
2789 
2790 	/* Checksum offload */
2791 	switch (csum_to_ctrl(txinfo,
2792 	    CHELSIO_CHIP_VERSION(pi->adapter->params.chip), &ctrl)) {
2793 	case COS_SUCCESS:
2794 		txq->txcsum++;
2795 		break;
2796 	case COS_FAIL:
2797 		/*
2798 		 * Packet will be going out with checksums which are probably
2799 		 * wrong but there is little we can do now.
2800 		 */
2801 		txq->csum_failed++;
2802 		break;
2803 	default:
2804 		break;
2805 	}
2806 
2807 	/*
2808 	 * The previous packet's SGL must have ended at a 16 byte boundary (this
2809 	 * is required by the firmware/hardware).  It follows that flitp cannot
2810 	 * wrap around between the ULPTX master command and ULPTX subcommand (8
2811 	 * bytes each), and that it can not wrap around in the middle of the
2812 	 * cpl_tx_pkt_core either.
2813 	 */
2814 	flitp = (uintptr_t)txpkts->flitp;
2815 	ASSERT((flitp & 0xf) == 0);
2816 
2817 	/* ULP master command */
2818 	ulpmc = (void *)flitp;
2819 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
2820 	ulpmc->len = htonl(howmany(sizeof (*ulpmc) + sizeof (*ulpsc) +
2821 	    sizeof (*cpl) + 8 * txinfo->nflits, 16));
2822 
2823 	/* ULP subcommand */
2824 	ulpsc = (void *)(ulpmc + 1);
2825 	ulpsc->cmd_more = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) |
2826 	    F_ULP_TX_SC_MORE);
2827 	ulpsc->len = cpu_to_be32(sizeof (struct cpl_tx_pkt_core));
2828 
2829 	flitp += sizeof (*ulpmc) + sizeof (*ulpsc);
2830 	if (flitp == end)
2831 		flitp = start;
2832 
2833 	/* CPL_TX_PKT_XT */
2834 	cpl = (void *)flitp;
2835 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
2836 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
2837 	cpl->pack = 0;
2838 	cpl->len = cpu_to_be16(txinfo->len);
2839 	cpl->ctrl1 = cpu_to_be64(ctrl);
2840 
2841 	flitp += sizeof (*cpl);
2842 	if (flitp == end)
2843 		flitp = start;
2844 
2845 	/* SGL for this frame */
2846 	dst = (caddr_t)flitp;
2847 	copy_to_txd(&txq->eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
2848 	flitp = (uintptr_t)dst;
2849 
2850 	/* Zero pad and advance to a 16 byte boundary if not already at one. */
2851 	if (flitp & 0xf) {
2852 
2853 		/* no matter what, flitp should be on an 8 byte boundary */
2854 		ASSERT((flitp & 0x7) == 0);
2855 
2856 		*(uint64_t *)flitp = 0;
2857 		flitp += sizeof (uint64_t);
2858 		txpkts->nflits++;
2859 	}
2860 
2861 	if (flitp == end)
2862 		flitp = start;
2863 
2864 	txpkts->flitp = (void *)flitp;
2865 }
2866 
2867 static inline void
copy_to_txd(struct sge_eq * eq,caddr_t from,caddr_t * to,int len)2868 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
2869 {
2870 	if ((uintptr_t)(*to) + len <= (uintptr_t)eq->spg) {
2871 		bcopy(from, *to, len);
2872 		(*to) += len;
2873 	} else {
2874 		int portion = (uintptr_t)eq->spg - (uintptr_t)(*to);
2875 
2876 		bcopy(from, *to, portion);
2877 		from += portion;
2878 		portion = len - portion;	/* remaining */
2879 		bcopy(from, (void *)eq->desc, portion);
2880 		(*to) = (caddr_t)eq->desc + portion;
2881 	}
2882 }
2883 
2884 static inline void
ring_tx_db(struct adapter * sc,struct sge_eq * eq)2885 ring_tx_db(struct adapter *sc, struct sge_eq *eq)
2886 {
2887 	int val, db_mode;
2888 	uint_t db = eq->doorbells;
2889 
2890 	if (eq->pending > 1)
2891 		db &= ~DOORBELL_WCWR;
2892 
2893 	if (eq->pending > eq->pidx) {
2894 		int offset = eq->cap - (eq->pending - eq->pidx);
2895 
2896 		/* pidx has wrapped around since last doorbell */
2897 
2898 		(void) ddi_dma_sync(eq->desc_dhdl,
2899 		    offset * sizeof (struct tx_desc), 0,
2900 		    DDI_DMA_SYNC_FORDEV);
2901 		(void) ddi_dma_sync(eq->desc_dhdl,
2902 		    0, eq->pidx * sizeof (struct tx_desc),
2903 		    DDI_DMA_SYNC_FORDEV);
2904 	} else if (eq->pending > 0) {
2905 		(void) ddi_dma_sync(eq->desc_dhdl,
2906 		    (eq->pidx - eq->pending) * sizeof (struct tx_desc),
2907 		    eq->pending * sizeof (struct tx_desc),
2908 		    DDI_DMA_SYNC_FORDEV);
2909 	}
2910 
2911 	membar_producer();
2912 
2913 	if (is_t4(sc->params.chip))
2914 		val = V_PIDX(eq->pending);
2915 	else
2916 		val = V_PIDX_T5(eq->pending);
2917 
2918 	db_mode = (1 << (ffs(db) - 1));
2919 	switch (db_mode) {
2920 		case DOORBELL_UDB:
2921 			*eq->udb = LE_32(V_QID(eq->udb_qid) | val);
2922 			break;
2923 
2924 		case DOORBELL_WCWR:
2925 			{
2926 				volatile uint64_t *dst, *src;
2927 				int i;
2928 				/*
2929 				 * Queues whose 128B doorbell segment fits in
2930 				 * the page do not use relative qid
2931 				 * (udb_qid is always 0).  Only queues with
2932 				 * doorbell segments can do WCWR.
2933 				 */
2934 				ASSERT(eq->udb_qid == 0 && eq->pending == 1);
2935 
2936 				dst = (volatile void *)((uintptr_t)eq->udb +
2937 				    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
2938 				i = eq->pidx ? eq->pidx - 1 : eq->cap - 1;
2939 				src = (void *)&eq->desc[i];
2940 				while (src != (void *)&eq->desc[i + 1]) {
2941 					*dst++ = *src++;
2942 				}
2943 				membar_producer();
2944 				break;
2945 			}
2946 
2947 		case DOORBELL_UDBWC:
2948 			*eq->udb = LE_32(V_QID(eq->udb_qid) | val);
2949 			membar_producer();
2950 			break;
2951 
2952 		case DOORBELL_KDB:
2953 			t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
2954 			    V_QID(eq->cntxt_id) | val);
2955 			break;
2956 	}
2957 
2958 	eq->pending = 0;
2959 }
2960 
2961 static int
reclaim_tx_descs(struct sge_txq * txq,int howmany)2962 reclaim_tx_descs(struct sge_txq *txq, int howmany)
2963 {
2964 	struct tx_sdesc *txsd;
2965 	uint_t cidx, can_reclaim, reclaimed, txb_freed, hdls_freed;
2966 	struct sge_eq *eq = &txq->eq;
2967 
2968 	EQ_LOCK_ASSERT_OWNED(eq);
2969 
2970 	cidx = eq->spg->cidx;	/* stable snapshot */
2971 	cidx = be16_to_cpu(cidx);
2972 
2973 	if (cidx >= eq->cidx)
2974 		can_reclaim = cidx - eq->cidx;
2975 	else
2976 		can_reclaim = cidx + eq->cap - eq->cidx;
2977 
2978 	if (can_reclaim == 0)
2979 		return (0);
2980 
2981 	txb_freed = hdls_freed = reclaimed = 0;
2982 	do {
2983 		int ndesc;
2984 
2985 		txsd = &txq->sdesc[eq->cidx];
2986 		ndesc = txsd->desc_used;
2987 
2988 		/* Firmware doesn't return "partial" credits. */
2989 		ASSERT(can_reclaim >= ndesc);
2990 
2991 		/*
2992 		 * We always keep mblk around, even for immediate data.  If mblk
2993 		 * is NULL, this has to be the software descriptor for a credit
2994 		 * flush work request.
2995 		 */
2996 		if (txsd->m != NULL)
2997 			freemsgchain(txsd->m);
2998 #ifdef DEBUG
2999 		else {
3000 			ASSERT(txsd->txb_used == 0);
3001 			ASSERT(txsd->hdls_used == 0);
3002 			ASSERT(ndesc == 1);
3003 		}
3004 #endif
3005 
3006 		txb_freed += txsd->txb_used;
3007 		hdls_freed += txsd->hdls_used;
3008 		reclaimed += ndesc;
3009 
3010 		eq->cidx += ndesc;
3011 		if (eq->cidx >= eq->cap)
3012 			eq->cidx -= eq->cap;
3013 
3014 		can_reclaim -= ndesc;
3015 
3016 	} while (can_reclaim && reclaimed < howmany);
3017 
3018 	eq->avail += reclaimed;
3019 	ASSERT(eq->avail < eq->cap);	/* avail tops out at (cap - 1) */
3020 
3021 	txq->txb_avail += txb_freed;
3022 
3023 	txq->tx_dhdl_avail += hdls_freed;
3024 	ASSERT(txq->tx_dhdl_avail <= txq->tx_dhdl_total);
3025 	for (; hdls_freed; hdls_freed--) {
3026 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_cidx]);
3027 		if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
3028 			txq->tx_dhdl_cidx = 0;
3029 	}
3030 
3031 	return (reclaimed);
3032 }
3033 
3034 static void
write_txqflush_wr(struct sge_txq * txq)3035 write_txqflush_wr(struct sge_txq *txq)
3036 {
3037 	struct sge_eq *eq = &txq->eq;
3038 	struct fw_eq_flush_wr *wr;
3039 	struct tx_sdesc *txsd;
3040 
3041 	EQ_LOCK_ASSERT_OWNED(eq);
3042 	ASSERT(eq->avail > 0);
3043 
3044 	wr = (void *)&eq->desc[eq->pidx];
3045 	bzero(wr, sizeof (*wr));
3046 	wr->opcode = FW_EQ_FLUSH_WR;
3047 	wr->equiq_to_len16 = cpu_to_be32(V_FW_WR_LEN16(sizeof (*wr) / 16) |
3048 	    F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
3049 
3050 	txsd = &txq->sdesc[eq->pidx];
3051 	txsd->m = NULL;
3052 	txsd->txb_used = 0;
3053 	txsd->hdls_used = 0;
3054 	txsd->desc_used = 1;
3055 
3056 	eq->pending++;
3057 	eq->avail--;
3058 	if (++eq->pidx == eq->cap)
3059 		eq->pidx = 0;
3060 }
3061 
3062 static int
t4_eth_rx(struct sge_iq * iq,const struct rss_header * rss,mblk_t * m)3063 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
3064 {
3065 	bool csum_ok;
3066 	uint16_t err_vec;
3067 	struct sge_rxq *rxq = (void *)iq;
3068 	struct mblk_pair chain = {0};
3069 	struct adapter *sc = iq->adapter;
3070 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
3071 
3072 	iq->intr_next = iq->intr_params;
3073 
3074 	m->b_rptr += sc->sge.pktshift;
3075 
3076 	/* Compressed error vector is enabled for T6 only */
3077 	if (sc->params.tp.rx_pkt_encap)
3078 		/* It is enabled only in T6 config file */
3079 		err_vec = G_T6_COMPR_RXERR_VEC(ntohs(cpl->err_vec));
3080 	else
3081 		err_vec = ntohs(cpl->err_vec);
3082 
3083 	csum_ok = cpl->csum_calc && !err_vec;
3084 	/* TODO: what about cpl->ip_frag? */
3085 	if (csum_ok && !cpl->ip_frag) {
3086 		mac_hcksum_set(m, 0, 0, 0, 0xffff,
3087 		    HCK_FULLCKSUM_OK | HCK_FULLCKSUM |
3088 		    HCK_IPV4_HDRCKSUM_OK);
3089 		rxq->rxcsum++;
3090 	}
3091 
3092 	/* Add to the chain that we'll send up */
3093 	if (chain.head != NULL)
3094 		chain.tail->b_next = m;
3095 	else
3096 		chain.head = m;
3097 	chain.tail = m;
3098 
3099 	t4_mac_rx(rxq->port, rxq, chain.head);
3100 
3101 	rxq->rxpkts++;
3102 	rxq->rxbytes  += be16_to_cpu(cpl->len);
3103 	return (0);
3104 }
3105 
3106 #define	FL_HW_IDX(idx)	((idx) >> 3)
3107 
3108 static inline void
ring_fl_db(struct adapter * sc,struct sge_fl * fl)3109 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3110 {
3111 	int desc_start, desc_last, ndesc;
3112 	uint32_t v = sc->params.arch.sge_fl_db;
3113 
3114 	ndesc = FL_HW_IDX(fl->pending);
3115 
3116 	/* Hold back one credit if pidx = cidx */
3117 	if (FL_HW_IDX(fl->pidx) == FL_HW_IDX(fl->cidx))
3118 		ndesc--;
3119 
3120 	/*
3121 	 * There are chances of ndesc modified above (to avoid pidx = cidx).
3122 	 * If there is nothing to post, return.
3123 	 */
3124 	if (ndesc <= 0)
3125 		return;
3126 
3127 	desc_last = FL_HW_IDX(fl->pidx);
3128 
3129 	if (fl->pidx < fl->pending) {
3130 		/* There was a wrap */
3131 		desc_start = FL_HW_IDX(fl->pidx + fl->cap - fl->pending);
3132 
3133 		/* From desc_start to the end of list */
3134 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE, 0,
3135 		    DDI_DMA_SYNC_FORDEV);
3136 
3137 		/* From start of list to the desc_last */
3138 		if (desc_last != 0)
3139 			(void) ddi_dma_sync(fl->dhdl, 0, desc_last *
3140 			    RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3141 	} else {
3142 		/* There was no wrap, sync from start_desc to last_desc */
3143 		desc_start = FL_HW_IDX(fl->pidx - fl->pending);
3144 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE,
3145 		    ndesc * RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3146 	}
3147 
3148 	if (is_t4(sc->params.chip))
3149 		v |= V_PIDX(ndesc);
3150 	else
3151 		v |= V_PIDX_T5(ndesc);
3152 	v |= V_QID(fl->cntxt_id) | V_PIDX(ndesc);
3153 
3154 	membar_producer();
3155 
3156 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3157 
3158 	/*
3159 	 * Update pending count:
3160 	 * Deduct the number of descriptors posted
3161 	 */
3162 	fl->pending -= ndesc * 8;
3163 }
3164 
3165 static void
tx_reclaim_task(void * arg)3166 tx_reclaim_task(void *arg)
3167 {
3168 	struct sge_txq *txq = arg;
3169 
3170 	TXQ_LOCK(txq);
3171 	reclaim_tx_descs(txq, txq->eq.qsize);
3172 	TXQ_UNLOCK(txq);
3173 }
3174 
3175 /* ARGSUSED */
3176 static int
handle_sge_egr_update(struct sge_iq * iq,const struct rss_header * rss,mblk_t * m)3177 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
3178     mblk_t *m)
3179 {
3180 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
3181 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
3182 	struct adapter *sc = iq->adapter;
3183 	struct sge *s = &sc->sge;
3184 	struct sge_eq *eq;
3185 	struct sge_txq *txq;
3186 
3187 	txq = (void *)s->eqmap[qid - s->eq_start];
3188 	eq = &txq->eq;
3189 	txq->qflush++;
3190 	t4_mac_tx_update(txq->port, txq);
3191 
3192 	ddi_taskq_dispatch(sc->tq[eq->tx_chan], tx_reclaim_task,
3193 	    (void *)txq, DDI_NOSLEEP);
3194 
3195 	return (0);
3196 }
3197 
3198 static int
handle_fw_rpl(struct sge_iq * iq,const struct rss_header * rss,mblk_t * m)3199 handle_fw_rpl(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
3200 {
3201 	struct adapter *sc = iq->adapter;
3202 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
3203 
3204 	ASSERT(m == NULL);
3205 
3206 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
3207 		const struct rss_header *rss2;
3208 
3209 		rss2 = (const struct rss_header *)&cpl->data[0];
3210 		return (sc->cpl_handler[rss2->opcode](iq, rss2, m));
3211 	}
3212 	return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0]));
3213 }
3214 
3215 int
t4_alloc_tx_maps(struct adapter * sc,struct tx_maps * txmaps,int count,int flags)3216 t4_alloc_tx_maps(struct adapter *sc, struct tx_maps *txmaps, int count,
3217     int flags)
3218 {
3219 	int i, rc;
3220 
3221 	txmaps->map_total =  count;
3222 	txmaps->map_avail = txmaps->map_cidx = txmaps->map_pidx = 0;
3223 
3224 	txmaps->map =  kmem_zalloc(sizeof (ddi_dma_handle_t) *
3225 	    txmaps->map_total, flags);
3226 
3227 	for (i = 0; i < count; i++) {
3228 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
3229 		    DDI_DMA_SLEEP, 0, &txmaps->map[i]);
3230 		if (rc != DDI_SUCCESS) {
3231 			cxgb_printf(sc->dip, CE_WARN,
3232 			    "%s: failed to allocate DMA handle (%d)",
3233 			    __func__, rc);
3234 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
3235 		}
3236 		txmaps->map_avail++;
3237 	}
3238 
3239 	return (0);
3240 }
3241 
3242 #define	KS_UINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_ULONG)
3243 #define	KS_CINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_CHAR)
3244 #define	KS_U_SET(x, y)	kstatp->x.value.ul = (y)
3245 #define	KS_U_FROM(x, y)	kstatp->x.value.ul = (y)->x
3246 #define	KS_C_SET(x, ...)	\
3247 			(void) snprintf(kstatp->x.value.c, 16,  __VA_ARGS__)
3248 
3249 /*
3250  * cxgbe:X:config
3251  */
3252 struct cxgbe_port_config_kstats {
3253 	kstat_named_t idx;
3254 	kstat_named_t nrxq;
3255 	kstat_named_t ntxq;
3256 	kstat_named_t first_rxq;
3257 	kstat_named_t first_txq;
3258 	kstat_named_t controller;
3259 	kstat_named_t factory_mac_address;
3260 };
3261 
3262 /*
3263  * cxgbe:X:info
3264  */
3265 struct cxgbe_port_info_kstats {
3266 	kstat_named_t transceiver;
3267 	kstat_named_t rx_ovflow0;
3268 	kstat_named_t rx_ovflow1;
3269 	kstat_named_t rx_ovflow2;
3270 	kstat_named_t rx_ovflow3;
3271 	kstat_named_t rx_trunc0;
3272 	kstat_named_t rx_trunc1;
3273 	kstat_named_t rx_trunc2;
3274 	kstat_named_t rx_trunc3;
3275 	kstat_named_t tx_pause;
3276 	kstat_named_t rx_pause;
3277 };
3278 
3279 static kstat_t *
setup_port_config_kstats(struct port_info * pi)3280 setup_port_config_kstats(struct port_info *pi)
3281 {
3282 	kstat_t *ksp;
3283 	struct cxgbe_port_config_kstats *kstatp;
3284 	int ndata;
3285 	dev_info_t *pdip = ddi_get_parent(pi->dip);
3286 	uint8_t *ma = &pi->hw_addr[0];
3287 
3288 	ndata = sizeof (struct cxgbe_port_config_kstats) /
3289 	    sizeof (kstat_named_t);
3290 
3291 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "config",
3292 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3293 	if (ksp == NULL) {
3294 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3295 		return (NULL);
3296 	}
3297 
3298 	kstatp = (struct cxgbe_port_config_kstats *)ksp->ks_data;
3299 
3300 	KS_UINIT(idx);
3301 	KS_UINIT(nrxq);
3302 	KS_UINIT(ntxq);
3303 	KS_UINIT(first_rxq);
3304 	KS_UINIT(first_txq);
3305 	KS_CINIT(controller);
3306 	KS_CINIT(factory_mac_address);
3307 
3308 	KS_U_SET(idx, pi->port_id);
3309 	KS_U_SET(nrxq, pi->nrxq);
3310 	KS_U_SET(ntxq, pi->ntxq);
3311 	KS_U_SET(first_rxq, pi->first_rxq);
3312 	KS_U_SET(first_txq, pi->first_txq);
3313 	KS_C_SET(controller, "%s%d", ddi_driver_name(pdip),
3314 	    ddi_get_instance(pdip));
3315 	KS_C_SET(factory_mac_address, "%02X%02X%02X%02X%02X%02X",
3316 	    ma[0], ma[1], ma[2], ma[3], ma[4], ma[5]);
3317 
3318 	/* Do NOT set ksp->ks_update.  These kstats do not change. */
3319 
3320 	/* Install the kstat */
3321 	ksp->ks_private = (void *)pi;
3322 	kstat_install(ksp);
3323 
3324 	return (ksp);
3325 }
3326 
3327 static kstat_t *
setup_port_info_kstats(struct port_info * pi)3328 setup_port_info_kstats(struct port_info *pi)
3329 {
3330 	kstat_t *ksp;
3331 	struct cxgbe_port_info_kstats *kstatp;
3332 	int ndata;
3333 
3334 	ndata = sizeof (struct cxgbe_port_info_kstats) / sizeof (kstat_named_t);
3335 
3336 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "info",
3337 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3338 	if (ksp == NULL) {
3339 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3340 		return (NULL);
3341 	}
3342 
3343 	kstatp = (struct cxgbe_port_info_kstats *)ksp->ks_data;
3344 
3345 	KS_CINIT(transceiver);
3346 	KS_UINIT(rx_ovflow0);
3347 	KS_UINIT(rx_ovflow1);
3348 	KS_UINIT(rx_ovflow2);
3349 	KS_UINIT(rx_ovflow3);
3350 	KS_UINIT(rx_trunc0);
3351 	KS_UINIT(rx_trunc1);
3352 	KS_UINIT(rx_trunc2);
3353 	KS_UINIT(rx_trunc3);
3354 	KS_UINIT(tx_pause);
3355 	KS_UINIT(rx_pause);
3356 
3357 	/* Install the kstat */
3358 	ksp->ks_update = update_port_info_kstats;
3359 	ksp->ks_private = (void *)pi;
3360 	kstat_install(ksp);
3361 
3362 	return (ksp);
3363 }
3364 
3365 static int
update_port_info_kstats(kstat_t * ksp,int rw)3366 update_port_info_kstats(kstat_t *ksp, int rw)
3367 {
3368 	struct cxgbe_port_info_kstats *kstatp =
3369 	    (struct cxgbe_port_info_kstats *)ksp->ks_data;
3370 	struct port_info *pi = ksp->ks_private;
3371 	static const char *mod_str[] = { NULL, "LR", "SR", "ER", "TWINAX",
3372 	    "active TWINAX", "LRM" };
3373 	uint32_t bgmap;
3374 
3375 	if (rw == KSTAT_WRITE)
3376 		return (0);
3377 
3378 	if (pi->mod_type == FW_PORT_MOD_TYPE_NONE)
3379 		KS_C_SET(transceiver, "unplugged");
3380 	else if (pi->mod_type == FW_PORT_MOD_TYPE_UNKNOWN)
3381 		KS_C_SET(transceiver, "unknown");
3382 	else if (pi->mod_type == FW_PORT_MOD_TYPE_NOTSUPPORTED)
3383 		KS_C_SET(transceiver, "unsupported");
3384 	else if (pi->mod_type > 0 && pi->mod_type < ARRAY_SIZE(mod_str))
3385 		KS_C_SET(transceiver, "%s", mod_str[pi->mod_type]);
3386 	else
3387 		KS_C_SET(transceiver, "type %d", pi->mod_type);
3388 
3389 #define	GET_STAT(name) t4_read_reg64(pi->adapter, \
3390 	    PORT_REG(pi->port_id, A_MPS_PORT_STAT_##name##_L))
3391 #define	GET_STAT_COM(name) t4_read_reg64(pi->adapter, \
3392 	    A_MPS_STAT_##name##_L)
3393 
3394 	bgmap = G_NUMPORTS(t4_read_reg(pi->adapter, A_MPS_CMN_CTL));
3395 	if (bgmap == 0)
3396 		bgmap = (pi->port_id == 0) ? 0xf : 0;
3397 	else if (bgmap == 1)
3398 		bgmap = (pi->port_id < 2) ? (3 << (2 * pi->port_id)) : 0;
3399 	else
3400 		bgmap = 1;
3401 
3402 	KS_U_SET(rx_ovflow0, (bgmap & 1) ?
3403 	    GET_STAT_COM(RX_BG_0_MAC_DROP_FRAME) : 0);
3404 	KS_U_SET(rx_ovflow1, (bgmap & 2) ?
3405 	    GET_STAT_COM(RX_BG_1_MAC_DROP_FRAME) : 0);
3406 	KS_U_SET(rx_ovflow2, (bgmap & 4) ?
3407 	    GET_STAT_COM(RX_BG_2_MAC_DROP_FRAME) : 0);
3408 	KS_U_SET(rx_ovflow3, (bgmap & 8) ?
3409 	    GET_STAT_COM(RX_BG_3_MAC_DROP_FRAME) : 0);
3410 	KS_U_SET(rx_trunc0,  (bgmap & 1) ?
3411 	    GET_STAT_COM(RX_BG_0_MAC_TRUNC_FRAME) : 0);
3412 	KS_U_SET(rx_trunc1,  (bgmap & 2) ?
3413 	    GET_STAT_COM(RX_BG_1_MAC_TRUNC_FRAME) : 0);
3414 	KS_U_SET(rx_trunc2,  (bgmap & 4) ?
3415 	    GET_STAT_COM(RX_BG_2_MAC_TRUNC_FRAME) : 0);
3416 	KS_U_SET(rx_trunc3,  (bgmap & 8) ?
3417 	    GET_STAT_COM(RX_BG_3_MAC_TRUNC_FRAME) : 0);
3418 
3419 	KS_U_SET(tx_pause, GET_STAT(TX_PORT_PAUSE));
3420 	KS_U_SET(rx_pause, GET_STAT(RX_PORT_PAUSE));
3421 
3422 	return (0);
3423 
3424 }
3425 
3426 /*
3427  * cxgbe:X:rxqY
3428  */
3429 struct rxq_kstats {
3430 	kstat_named_t rxcsum;
3431 	kstat_named_t rxpkts;
3432 	kstat_named_t rxbytes;
3433 	kstat_named_t nomem;
3434 };
3435 
3436 static kstat_t *
setup_rxq_kstats(struct port_info * pi,struct sge_rxq * rxq,int idx)3437 setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq, int idx)
3438 {
3439 	struct kstat *ksp;
3440 	struct rxq_kstats *kstatp;
3441 	int ndata;
3442 	char str[16];
3443 
3444 	ndata = sizeof (struct rxq_kstats) / sizeof (kstat_named_t);
3445 	(void) snprintf(str, sizeof (str), "rxq%u", idx);
3446 
3447 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "rxq",
3448 	    KSTAT_TYPE_NAMED, ndata, 0);
3449 	if (ksp == NULL) {
3450 		cxgb_printf(pi->dip, CE_WARN,
3451 		    "%s: failed to initialize rxq kstats for queue %d.",
3452 		    __func__, idx);
3453 		return (NULL);
3454 	}
3455 
3456 	kstatp = (struct rxq_kstats *)ksp->ks_data;
3457 
3458 	KS_UINIT(rxcsum);
3459 	KS_UINIT(rxpkts);
3460 	KS_UINIT(rxbytes);
3461 	KS_UINIT(nomem);
3462 
3463 	ksp->ks_update = update_rxq_kstats;
3464 	ksp->ks_private = (void *)rxq;
3465 	kstat_install(ksp);
3466 
3467 	return (ksp);
3468 }
3469 
3470 static int
update_rxq_kstats(kstat_t * ksp,int rw)3471 update_rxq_kstats(kstat_t *ksp, int rw)
3472 {
3473 	struct rxq_kstats *kstatp = (struct rxq_kstats *)ksp->ks_data;
3474 	struct sge_rxq *rxq = ksp->ks_private;
3475 
3476 	if (rw == KSTAT_WRITE)
3477 		return (0);
3478 
3479 	KS_U_FROM(rxcsum, rxq);
3480 	KS_U_FROM(rxpkts, rxq);
3481 	KS_U_FROM(rxbytes, rxq);
3482 	KS_U_FROM(nomem, rxq);
3483 
3484 	return (0);
3485 }
3486 
3487 /*
3488  * cxgbe:X:txqY
3489  */
3490 struct txq_kstats {
3491 	kstat_named_t txcsum;
3492 	kstat_named_t tso_wrs;
3493 	kstat_named_t imm_wrs;
3494 	kstat_named_t sgl_wrs;
3495 	kstat_named_t txpkt_wrs;
3496 	kstat_named_t txpkts_wrs;
3497 	kstat_named_t txpkts_pkts;
3498 	kstat_named_t txb_used;
3499 	kstat_named_t hdl_used;
3500 	kstat_named_t txb_full;
3501 	kstat_named_t dma_hdl_failed;
3502 	kstat_named_t dma_map_failed;
3503 	kstat_named_t qfull;
3504 	kstat_named_t qflush;
3505 	kstat_named_t pullup_early;
3506 	kstat_named_t pullup_late;
3507 	kstat_named_t pullup_failed;
3508 	kstat_named_t csum_failed;
3509 };
3510 
3511 static kstat_t *
setup_txq_kstats(struct port_info * pi,struct sge_txq * txq,int idx)3512 setup_txq_kstats(struct port_info *pi, struct sge_txq *txq, int idx)
3513 {
3514 	struct kstat *ksp;
3515 	struct txq_kstats *kstatp;
3516 	int ndata;
3517 	char str[16];
3518 
3519 	ndata = sizeof (struct txq_kstats) / sizeof (kstat_named_t);
3520 	(void) snprintf(str, sizeof (str), "txq%u", idx);
3521 
3522 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "txq",
3523 	    KSTAT_TYPE_NAMED, ndata, 0);
3524 	if (ksp == NULL) {
3525 		cxgb_printf(pi->dip, CE_WARN,
3526 		    "%s: failed to initialize txq kstats for queue %d.",
3527 		    __func__, idx);
3528 		return (NULL);
3529 	}
3530 
3531 	kstatp = (struct txq_kstats *)ksp->ks_data;
3532 
3533 	KS_UINIT(txcsum);
3534 	KS_UINIT(tso_wrs);
3535 	KS_UINIT(imm_wrs);
3536 	KS_UINIT(sgl_wrs);
3537 	KS_UINIT(txpkt_wrs);
3538 	KS_UINIT(txpkts_wrs);
3539 	KS_UINIT(txpkts_pkts);
3540 	KS_UINIT(txb_used);
3541 	KS_UINIT(hdl_used);
3542 	KS_UINIT(txb_full);
3543 	KS_UINIT(dma_hdl_failed);
3544 	KS_UINIT(dma_map_failed);
3545 	KS_UINIT(qfull);
3546 	KS_UINIT(qflush);
3547 	KS_UINIT(pullup_early);
3548 	KS_UINIT(pullup_late);
3549 	KS_UINIT(pullup_failed);
3550 	KS_UINIT(csum_failed);
3551 
3552 	ksp->ks_update = update_txq_kstats;
3553 	ksp->ks_private = (void *)txq;
3554 	kstat_install(ksp);
3555 
3556 	return (ksp);
3557 }
3558 
3559 static int
update_txq_kstats(kstat_t * ksp,int rw)3560 update_txq_kstats(kstat_t *ksp, int rw)
3561 {
3562 	struct txq_kstats *kstatp = (struct txq_kstats *)ksp->ks_data;
3563 	struct sge_txq *txq = ksp->ks_private;
3564 
3565 	if (rw == KSTAT_WRITE)
3566 		return (0);
3567 
3568 	KS_U_FROM(txcsum, txq);
3569 	KS_U_FROM(tso_wrs, txq);
3570 	KS_U_FROM(imm_wrs, txq);
3571 	KS_U_FROM(sgl_wrs, txq);
3572 	KS_U_FROM(txpkt_wrs, txq);
3573 	KS_U_FROM(txpkts_wrs, txq);
3574 	KS_U_FROM(txpkts_pkts, txq);
3575 	KS_U_FROM(txb_used, txq);
3576 	KS_U_FROM(hdl_used, txq);
3577 	KS_U_FROM(txb_full, txq);
3578 	KS_U_FROM(dma_hdl_failed, txq);
3579 	KS_U_FROM(dma_map_failed, txq);
3580 	KS_U_FROM(qfull, txq);
3581 	KS_U_FROM(qflush, txq);
3582 	KS_U_FROM(pullup_early, txq);
3583 	KS_U_FROM(pullup_late, txq);
3584 	KS_U_FROM(pullup_failed, txq);
3585 	KS_U_FROM(csum_failed, txq);
3586 
3587 	return (0);
3588 }
3589 
3590 static int rxbuf_ctor(void *, void *, int);
3591 static void rxbuf_dtor(void *, void *);
3592 
3593 static kmem_cache_t *
rxbuf_cache_create(struct rxbuf_cache_params * p)3594 rxbuf_cache_create(struct rxbuf_cache_params *p)
3595 {
3596 	char name[32];
3597 
3598 	(void) snprintf(name, sizeof (name), "%s%d_rxbuf_cache",
3599 	    ddi_driver_name(p->dip), ddi_get_instance(p->dip));
3600 
3601 	return kmem_cache_create(name, sizeof (struct rxbuf), _CACHE_LINE_SIZE,
3602 	    rxbuf_ctor, rxbuf_dtor, NULL, p, NULL, 0);
3603 }
3604 
3605 /*
3606  * If ref_cnt is more than 1 then those many calls to rxbuf_free will
3607  * have to be made before the rxb is released back to the kmem_cache.
3608  */
3609 static struct rxbuf *
rxbuf_alloc(kmem_cache_t * cache,int kmflags,uint_t ref_cnt)3610 rxbuf_alloc(kmem_cache_t *cache, int kmflags, uint_t ref_cnt)
3611 {
3612 	struct rxbuf *rxb;
3613 
3614 	ASSERT(ref_cnt > 0);
3615 
3616 	rxb = kmem_cache_alloc(cache, kmflags);
3617 	if (rxb != NULL) {
3618 		rxb->ref_cnt = ref_cnt;
3619 		rxb->cache = cache;
3620 	}
3621 
3622 	return (rxb);
3623 }
3624 
3625 /*
3626  * This is normally called via the rxb's freefunc, when an mblk referencing the
3627  * rxb is freed.
3628  */
3629 static void
rxbuf_free(struct rxbuf * rxb)3630 rxbuf_free(struct rxbuf *rxb)
3631 {
3632 	if (atomic_dec_uint_nv(&rxb->ref_cnt) == 0)
3633 		kmem_cache_free(rxb->cache, rxb);
3634 }
3635 
3636 static int
rxbuf_ctor(void * arg1,void * arg2,int kmflag)3637 rxbuf_ctor(void *arg1, void *arg2, int kmflag)
3638 {
3639 	struct rxbuf *rxb = arg1;
3640 	struct rxbuf_cache_params *p = arg2;
3641 	size_t real_len;
3642 	ddi_dma_cookie_t cookie;
3643 	uint_t ccount = 0;
3644 	int (*callback)(caddr_t);
3645 	int rc = ENOMEM;
3646 
3647 	if ((kmflag & KM_NOSLEEP) != 0)
3648 		callback = DDI_DMA_DONTWAIT;
3649 	else
3650 		callback = DDI_DMA_SLEEP;
3651 
3652 	rc = ddi_dma_alloc_handle(p->dip, &p->dma_attr_rx, callback, 0,
3653 	    &rxb->dhdl);
3654 	if (rc != DDI_SUCCESS)
3655 		return (rc == DDI_DMA_BADATTR ? EINVAL : ENOMEM);
3656 
3657 	rc = ddi_dma_mem_alloc(rxb->dhdl, p->buf_size, &p->acc_attr_rx,
3658 	    DDI_DMA_STREAMING, callback, 0, &rxb->va, &real_len, &rxb->ahdl);
3659 	if (rc != DDI_SUCCESS) {
3660 		rc = ENOMEM;
3661 		goto fail1;
3662 	}
3663 
3664 	rc = ddi_dma_addr_bind_handle(rxb->dhdl, NULL, rxb->va, p->buf_size,
3665 	    DDI_DMA_READ | DDI_DMA_STREAMING, NULL, NULL, &cookie, &ccount);
3666 	if (rc != DDI_DMA_MAPPED) {
3667 		if (rc == DDI_DMA_INUSE)
3668 			rc = EBUSY;
3669 		else if (rc == DDI_DMA_TOOBIG)
3670 			rc = E2BIG;
3671 		else
3672 			rc = ENOMEM;
3673 		goto fail2;
3674 	}
3675 
3676 	if (ccount != 1) {
3677 		rc = E2BIG;
3678 		goto fail3;
3679 	}
3680 
3681 	rxb->ref_cnt = 0;
3682 	rxb->buf_size = p->buf_size;
3683 	rxb->freefunc.free_arg = (caddr_t)rxb;
3684 	rxb->freefunc.free_func = rxbuf_free;
3685 	rxb->ba = cookie.dmac_laddress;
3686 
3687 	return (0);
3688 
3689 fail3:	(void) ddi_dma_unbind_handle(rxb->dhdl);
3690 fail2:	ddi_dma_mem_free(&rxb->ahdl);
3691 fail1:	ddi_dma_free_handle(&rxb->dhdl);
3692 	return (rc);
3693 }
3694 
3695 static void
rxbuf_dtor(void * arg1,void * arg2)3696 rxbuf_dtor(void *arg1, void *arg2)
3697 {
3698 	struct rxbuf *rxb = arg1;
3699 
3700 	(void) ddi_dma_unbind_handle(rxb->dhdl);
3701 	ddi_dma_mem_free(&rxb->ahdl);
3702 	ddi_dma_free_handle(&rxb->dhdl);
3703 }
3704