2 * Copyright (c) 2011 Chelsio Communications, Inc.
4 * Written by: Navdeep Parhar <np@FreeBSD.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
32 #include "opt_inet6.h"
34 #include <sys/types.h>
36 #include <sys/socket.h>
37 #include <sys/kernel.h>
39 #include <sys/malloc.h>
40 #include <sys/queue.h>
41 #include <sys/taskqueue.h>
42 #include <sys/sysctl.h>
45 #include <net/ethernet.h>
47 #include <net/if_vlan_var.h>
48 #include <netinet/in.h>
49 #include <netinet/ip.h>
50 #include <netinet/ip6.h>
51 #include <netinet/tcp.h>
53 #include "common/common.h"
54 #include "common/t4_regs.h"
55 #include "common/t4_regs_values.h"
56 #include "common/t4_msg.h"
64 /* Filled up by t4_sge_modload */
65 static struct fl_buf_info fl_buf_info[FL_BUF_SIZES];
67 #define FL_BUF_SIZE(x) (fl_buf_info[x].size)
68 #define FL_BUF_TYPE(x) (fl_buf_info[x].type)
69 #define FL_BUF_ZONE(x) (fl_buf_info[x].zone)
71 #ifdef T4_PKT_TIMESTAMP
72 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
74 #define RX_COPY_THRESHOLD MINCLSIZE
78 * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
79 * 0-7 are valid values.
81 static int fl_pktshift = 2;
82 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
85 * Pad ethernet payload up to this boundary.
86 * -1: driver should figure out a good value.
87 * Any power of 2, from 32 to 4096 (both inclusive) is a valid value.
89 static int fl_pad = -1;
90 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
94 * -1: driver should figure out a good value.
95 * 64 or 128 are the only other valid values.
97 static int spg_len = -1;
98 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
102 * -1: no congestion feedback (not recommended).
103 * 0: backpressure the channel instead of dropping packets right away.
104 * 1: no backpressure, drop packets for the congested queue immediately.
106 static int cong_drop = 0;
107 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
109 /* Used to track coalesced tx work request */
111 uint64_t *flitp; /* ptr to flit where next pkt should start */
112 uint8_t npkt; /* # of packets in this work request */
113 uint8_t nflits; /* # of flits used by this work request */
114 uint16_t plen; /* total payload (sum of all packets) */
117 /* A packet's SGL. This + m_pkthdr has all info needed for tx */
119 int nsegs; /* # of segments in the SGL, 0 means imm. tx */
120 int nflits; /* # of flits needed for the SGL */
121 bus_dma_segment_t seg[TX_SGL_SEGS];
124 static int service_iq(struct sge_iq *, int);
125 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t,
127 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
128 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int,
130 static inline void init_fl(struct sge_fl *, int, int, char *);
131 static inline void init_eq(struct sge_eq *, int, int, uint8_t, uint16_t,
133 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
134 bus_addr_t *, void **);
135 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
137 static int alloc_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *,
139 static int free_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *);
140 static int alloc_fwq(struct adapter *);
141 static int free_fwq(struct adapter *);
142 static int alloc_mgmtq(struct adapter *);
143 static int free_mgmtq(struct adapter *);
144 static int alloc_rxq(struct port_info *, struct sge_rxq *, int, int,
145 struct sysctl_oid *);
146 static int free_rxq(struct port_info *, struct sge_rxq *);
148 static int alloc_ofld_rxq(struct port_info *, struct sge_ofld_rxq *, int, int,
149 struct sysctl_oid *);
150 static int free_ofld_rxq(struct port_info *, struct sge_ofld_rxq *);
152 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
153 static int eth_eq_alloc(struct adapter *, struct port_info *, struct sge_eq *);
155 static int ofld_eq_alloc(struct adapter *, struct port_info *, struct sge_eq *);
157 static int alloc_eq(struct adapter *, struct port_info *, struct sge_eq *);
158 static int free_eq(struct adapter *, struct sge_eq *);
159 static int alloc_wrq(struct adapter *, struct port_info *, struct sge_wrq *,
160 struct sysctl_oid *);
161 static int free_wrq(struct adapter *, struct sge_wrq *);
162 static int alloc_txq(struct port_info *, struct sge_txq *, int,
163 struct sysctl_oid *);
164 static int free_txq(struct port_info *, struct sge_txq *);
165 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
166 static inline bool is_new_response(const struct sge_iq *, struct rsp_ctrl **);
167 static inline void iq_next(struct sge_iq *);
168 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
169 static int refill_fl(struct adapter *, struct sge_fl *, int);
170 static void refill_sfl(void *);
171 static int alloc_fl_sdesc(struct sge_fl *);
172 static void free_fl_sdesc(struct sge_fl *);
173 static void set_fl_tag_idx(struct sge_fl *, int);
174 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
176 static int get_pkt_sgl(struct sge_txq *, struct mbuf **, struct sgl *, int);
177 static int free_pkt_sgl(struct sge_txq *, struct sgl *);
178 static int write_txpkt_wr(struct port_info *, struct sge_txq *, struct mbuf *,
180 static int add_to_txpkts(struct port_info *, struct sge_txq *, struct txpkts *,
181 struct mbuf *, struct sgl *);
182 static void write_txpkts_wr(struct sge_txq *, struct txpkts *);
183 static inline void write_ulp_cpl_sgl(struct port_info *, struct sge_txq *,
184 struct txpkts *, struct mbuf *, struct sgl *);
185 static int write_sgl_to_txd(struct sge_eq *, struct sgl *, caddr_t *);
186 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
187 static inline void ring_eq_db(struct adapter *, struct sge_eq *);
188 static inline int reclaimable(struct sge_eq *);
189 static int reclaim_tx_descs(struct sge_txq *, int, int);
190 static void write_eqflush_wr(struct sge_eq *);
191 static __be64 get_flit(bus_dma_segment_t *, int, int);
192 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
194 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
197 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
199 #if defined(__i386__) || defined(__amd64__)
200 extern u_int cpu_clflush_line_size;
204 * Called on MOD_LOAD. Fills up fl_buf_info[] and validates/calculates the SGE
211 int bufsize[FL_BUF_SIZES] = {
213 #if MJUMPAGESIZE != MCLBYTES
220 for (i = 0; i < FL_BUF_SIZES; i++) {
221 FL_BUF_SIZE(i) = bufsize[i];
222 FL_BUF_TYPE(i) = m_gettype(bufsize[i]);
223 FL_BUF_ZONE(i) = m_getzone(bufsize[i]);
226 if (fl_pktshift < 0 || fl_pktshift > 7) {
227 printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
228 " using 2 instead.\n", fl_pktshift);
232 if (fl_pad < 32 || fl_pad > 4096 || !powerof2(fl_pad)) {
235 #if defined(__i386__) || defined(__amd64__)
236 pad = max(cpu_clflush_line_size, 32);
238 pad = max(CACHE_LINE_SIZE, 32);
240 pad = min(pad, 4096);
243 printf("Invalid hw.cxgbe.fl_pad value (%d),"
244 " using %d instead.\n", fl_pad, pad);
249 if (spg_len != 64 && spg_len != 128) {
252 #if defined(__i386__) || defined(__amd64__)
253 len = cpu_clflush_line_size > 64 ? 128 : 64;
258 printf("Invalid hw.cxgbe.spg_len value (%d),"
259 " using %d instead.\n", spg_len, len);
264 if (cong_drop < -1 || cong_drop > 1) {
265 printf("Invalid hw.cxgbe.cong_drop value (%d),"
266 " using 0 instead.\n", cong_drop);
272 t4_init_sge_cpl_handlers(struct adapter *sc)
275 t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_msg);
276 t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_msg);
277 t4_register_cpl_handler(sc, CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
278 t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
279 t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
283 * adap->params.vpd.cclk must be set up before this is called.
286 t4_tweak_chip_settings(struct adapter *sc)
290 int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
291 int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
292 int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
293 uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
295 KASSERT(sc->flags & MASTER_PF,
296 ("%s: trying to change chip settings when not master.", __func__));
298 m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE |
299 V_INGPADBOUNDARY(M_INGPADBOUNDARY) | F_EGRSTATUSPAGESIZE;
300 v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
301 V_INGPADBOUNDARY(ilog2(fl_pad) - 5) |
302 V_EGRSTATUSPAGESIZE(spg_len == 128);
303 t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
305 v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
306 V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
307 V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
308 V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
309 V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
310 V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
311 V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
312 V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
313 t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
315 for (i = 0; i < FL_BUF_SIZES; i++) {
316 t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
320 v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
321 V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
322 t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
324 KASSERT(intr_timer[0] <= timer_max,
325 ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
327 for (i = 1; i < nitems(intr_timer); i++) {
328 KASSERT(intr_timer[i] >= intr_timer[i - 1],
329 ("%s: timers not listed in increasing order (%d)",
332 while (intr_timer[i] > timer_max) {
333 if (i == nitems(intr_timer) - 1) {
334 intr_timer[i] = timer_max;
337 intr_timer[i] += intr_timer[i - 1];
342 v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
343 V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
344 t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
345 v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
346 V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
347 t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
348 v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
349 V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
350 t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
352 if (cong_drop == 0) {
353 m = F_TUNNELCNGDROP0 | F_TUNNELCNGDROP1 | F_TUNNELCNGDROP2 |
355 t4_set_reg_field(sc, A_TP_PARA_REG3, m, 0);
358 /* 4K, 16K, 64K, 256K DDP "page sizes" */
359 v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
360 t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
362 m = v = F_TDDPTAGTCB;
363 t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
365 m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
367 v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
368 t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
372 * XXX: driver really should be able to deal with unexpected settings.
375 t4_read_chip_settings(struct adapter *sc)
377 struct sge *s = &sc->sge;
380 uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
382 m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE |
383 V_INGPADBOUNDARY(M_INGPADBOUNDARY) | F_EGRSTATUSPAGESIZE;
384 v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
385 V_INGPADBOUNDARY(ilog2(fl_pad) - 5) |
386 V_EGRSTATUSPAGESIZE(spg_len == 128);
387 r = t4_read_reg(sc, A_SGE_CONTROL);
389 device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
393 v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
394 V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
395 V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
396 V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
397 V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
398 V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
399 V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
400 V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
401 r = t4_read_reg(sc, A_SGE_HOST_PAGE_SIZE);
403 device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
407 for (i = 0; i < FL_BUF_SIZES; i++) {
408 v = t4_read_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i));
409 if (v != FL_BUF_SIZE(i)) {
410 device_printf(sc->dev,
411 "invalid SGE_FL_BUFFER_SIZE[%d](0x%x)\n", i, v);
416 r = t4_read_reg(sc, A_SGE_INGRESS_RX_THRESHOLD);
417 s->counter_val[0] = G_THRESHOLD_0(r);
418 s->counter_val[1] = G_THRESHOLD_1(r);
419 s->counter_val[2] = G_THRESHOLD_2(r);
420 s->counter_val[3] = G_THRESHOLD_3(r);
422 r = t4_read_reg(sc, A_SGE_TIMER_VALUE_0_AND_1);
423 s->timer_val[0] = G_TIMERVALUE0(r) / core_ticks_per_usec(sc);
424 s->timer_val[1] = G_TIMERVALUE1(r) / core_ticks_per_usec(sc);
425 r = t4_read_reg(sc, A_SGE_TIMER_VALUE_2_AND_3);
426 s->timer_val[2] = G_TIMERVALUE2(r) / core_ticks_per_usec(sc);
427 s->timer_val[3] = G_TIMERVALUE3(r) / core_ticks_per_usec(sc);
428 r = t4_read_reg(sc, A_SGE_TIMER_VALUE_4_AND_5);
429 s->timer_val[4] = G_TIMERVALUE4(r) / core_ticks_per_usec(sc);
430 s->timer_val[5] = G_TIMERVALUE5(r) / core_ticks_per_usec(sc);
432 if (cong_drop == 0) {
433 m = F_TUNNELCNGDROP0 | F_TUNNELCNGDROP1 | F_TUNNELCNGDROP2 |
435 r = t4_read_reg(sc, A_TP_PARA_REG3);
437 device_printf(sc->dev,
438 "invalid TP_PARA_REG3(0x%x)\n", r);
443 v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
444 r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
446 device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
450 m = v = F_TDDPTAGTCB;
451 r = t4_read_reg(sc, A_ULP_RX_CTL);
453 device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
457 m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
459 v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
460 r = t4_read_reg(sc, A_TP_PARA_REG5);
462 device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
466 r = t4_read_reg(sc, A_SGE_CONM_CTRL);
467 s->fl_starve_threshold = G_EGRTHRESHOLD(r) * 2 + 1;
470 r = t4_read_reg(sc, A_SGE_EGRESS_QUEUES_PER_PAGE_PF);
471 r >>= S_QUEUESPERPAGEPF0 +
472 (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
473 s->s_qpp = r & M_QUEUESPERPAGEPF0;
476 t4_init_tp_params(sc);
478 t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
479 t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
485 t4_create_dma_tag(struct adapter *sc)
489 rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
490 BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
491 BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
494 device_printf(sc->dev,
495 "failed to create main DMA tag: %d\n", rc);
502 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
503 struct sysctl_oid_list *children)
506 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
507 NULL, fl_pktshift, "payload DMA offset in rx buffer (bytes)");
509 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
510 NULL, fl_pad, "payload pad boundary (bytes)");
512 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
513 NULL, spg_len, "status page size (bytes)");
515 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
516 NULL, cong_drop, "congestion drop setting");
520 t4_destroy_dma_tag(struct adapter *sc)
523 bus_dma_tag_destroy(sc->dmat);
529 * Allocate and initialize the firmware event queue and the management queue.
531 * Returns errno on failure. Resources allocated up to that point may still be
532 * allocated. Caller is responsible for cleanup in case this function fails.
535 t4_setup_adapter_queues(struct adapter *sc)
539 ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
541 sysctl_ctx_init(&sc->ctx);
542 sc->flags |= ADAP_SYSCTL_CTX;
545 * Firmware event queue
552 * Management queue. This is just a control queue that uses the fwq as
555 rc = alloc_mgmtq(sc);
564 t4_teardown_adapter_queues(struct adapter *sc)
567 ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
569 /* Do this before freeing the queue */
570 if (sc->flags & ADAP_SYSCTL_CTX) {
571 sysctl_ctx_free(&sc->ctx);
572 sc->flags &= ~ADAP_SYSCTL_CTX;
582 first_vector(struct port_info *pi)
584 struct adapter *sc = pi->adapter;
585 int rc = T4_EXTRA_INTR, i;
587 if (sc->intr_count == 1)
590 for_each_port(sc, i) {
591 struct port_info *p = sc->port[i];
593 if (i == pi->port_id)
597 if (sc->flags & INTR_DIRECT)
598 rc += p->nrxq + p->nofldrxq;
600 rc += max(p->nrxq, p->nofldrxq);
603 * Not compiled with offload support and intr_count > 1. Only
604 * NIC queues exist and they'd better be taking direct
607 KASSERT(sc->flags & INTR_DIRECT,
608 ("%s: intr_count %d, !INTR_DIRECT", __func__,
619 * Given an arbitrary "index," come up with an iq that can be used by other
620 * queues (of this port) for interrupt forwarding, SGE egress updates, etc.
621 * The iq returned is guaranteed to be something that takes direct interrupts.
623 static struct sge_iq *
624 port_intr_iq(struct port_info *pi, int idx)
626 struct adapter *sc = pi->adapter;
627 struct sge *s = &sc->sge;
628 struct sge_iq *iq = NULL;
630 if (sc->intr_count == 1)
631 return (&sc->sge.fwq);
634 if (sc->flags & INTR_DIRECT) {
635 idx %= pi->nrxq + pi->nofldrxq;
637 if (idx >= pi->nrxq) {
639 iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
641 iq = &s->rxq[pi->first_rxq + idx].iq;
644 idx %= max(pi->nrxq, pi->nofldrxq);
646 if (pi->nrxq >= pi->nofldrxq)
647 iq = &s->rxq[pi->first_rxq + idx].iq;
649 iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
653 * Not compiled with offload support and intr_count > 1. Only NIC
654 * queues exist and they'd better be taking direct interrupts.
656 KASSERT(sc->flags & INTR_DIRECT,
657 ("%s: intr_count %d, !INTR_DIRECT", __func__, sc->intr_count));
660 iq = &s->rxq[pi->first_rxq + idx].iq;
663 KASSERT(iq->flags & IQ_INTR, ("%s: EDOOFUS", __func__));
668 mtu_to_bufsize(int mtu)
672 /* large enough for a frame even when VLAN extraction is disabled */
673 bufsize = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN + mtu;
674 bufsize = roundup2(bufsize + fl_pktshift, fl_pad);
681 mtu_to_bufsize_toe(struct adapter *sc, int mtu)
684 if (sc->tt.rx_coalesce)
685 return (G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)));
692 t4_setup_port_queues(struct port_info *pi)
694 int rc = 0, i, j, intr_idx, iqid;
697 struct sge_wrq *ctrlq;
699 struct sge_ofld_rxq *ofld_rxq;
700 struct sge_wrq *ofld_txq;
701 struct sysctl_oid *oid2 = NULL;
704 struct adapter *sc = pi->adapter;
705 struct ifnet *ifp = pi->ifp;
706 struct sysctl_oid *oid = device_get_sysctl_tree(pi->dev);
707 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
710 oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "rxq", CTLFLAG_RD,
714 if (is_offload(sc)) {
715 oid2 = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_rxq",
717 "rx queues for offloaded TCP connections");
721 /* Interrupt vector to start from (when using multiple vectors) */
722 intr_idx = first_vector(pi);
725 * First pass over all rx queues (NIC and TOE):
726 * a) initialize iq and fl
727 * b) allocate queue iff it will take direct interrupts.
729 bufsize = mtu_to_bufsize(ifp->if_mtu);
730 for_each_rxq(pi, i, rxq) {
732 init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, pi->qsize_rxq,
735 snprintf(name, sizeof(name), "%s rxq%d-fl",
736 device_get_nameunit(pi->dev), i);
737 init_fl(&rxq->fl, pi->qsize_rxq / 8, bufsize, name);
739 if (sc->flags & INTR_DIRECT
741 || (sc->intr_count > 1 && pi->nrxq >= pi->nofldrxq)
744 rxq->iq.flags |= IQ_INTR;
745 rc = alloc_rxq(pi, rxq, intr_idx, i, oid);
753 bufsize = mtu_to_bufsize_toe(sc, ifp->if_mtu);
754 for_each_ofld_rxq(pi, i, ofld_rxq) {
756 init_iq(&ofld_rxq->iq, sc, pi->tmr_idx, pi->pktc_idx,
757 pi->qsize_rxq, RX_IQ_ESIZE);
759 snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
760 device_get_nameunit(pi->dev), i);
761 init_fl(&ofld_rxq->fl, pi->qsize_rxq / 8, bufsize, name);
763 if (sc->flags & INTR_DIRECT ||
764 (sc->intr_count > 1 && pi->nofldrxq > pi->nrxq)) {
765 ofld_rxq->iq.flags |= IQ_INTR;
766 rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx, i, oid2);
775 * Second pass over all rx queues (NIC and TOE). The queues forwarding
776 * their interrupts are allocated now.
779 for_each_rxq(pi, i, rxq) {
780 if (rxq->iq.flags & IQ_INTR)
783 intr_idx = port_intr_iq(pi, j)->abs_id;
785 rc = alloc_rxq(pi, rxq, intr_idx, i, oid);
792 for_each_ofld_rxq(pi, i, ofld_rxq) {
793 if (ofld_rxq->iq.flags & IQ_INTR)
796 intr_idx = port_intr_iq(pi, j)->abs_id;
798 rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx, i, oid2);
806 * Now the tx queues. Only one pass needed.
808 oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
811 for_each_txq(pi, i, txq) {
814 iqid = port_intr_iq(pi, j)->cntxt_id;
816 snprintf(name, sizeof(name), "%s txq%d",
817 device_get_nameunit(pi->dev), i);
818 init_eq(&txq->eq, EQ_ETH, pi->qsize_txq, pi->tx_chan, iqid,
821 rc = alloc_txq(pi, txq, i, oid);
828 oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_txq",
829 CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
830 for_each_ofld_txq(pi, i, ofld_txq) {
833 iqid = port_intr_iq(pi, j)->cntxt_id;
835 snprintf(name, sizeof(name), "%s ofld_txq%d",
836 device_get_nameunit(pi->dev), i);
837 init_eq(&ofld_txq->eq, EQ_OFLD, pi->qsize_txq, pi->tx_chan,
840 snprintf(name, sizeof(name), "%d", i);
841 oid2 = SYSCTL_ADD_NODE(&pi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
842 name, CTLFLAG_RD, NULL, "offload tx queue");
844 rc = alloc_wrq(sc, pi, ofld_txq, oid2);
852 * Finally, the control queue.
854 oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
856 ctrlq = &sc->sge.ctrlq[pi->port_id];
857 iqid = port_intr_iq(pi, 0)->cntxt_id;
858 snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(pi->dev));
859 init_eq(&ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid, name);
860 rc = alloc_wrq(sc, pi, ctrlq, oid);
864 t4_teardown_port_queues(pi);
873 t4_teardown_port_queues(struct port_info *pi)
876 struct adapter *sc = pi->adapter;
880 struct sge_ofld_rxq *ofld_rxq;
881 struct sge_wrq *ofld_txq;
884 /* Do this before freeing the queues */
885 if (pi->flags & PORT_SYSCTL_CTX) {
886 sysctl_ctx_free(&pi->ctx);
887 pi->flags &= ~PORT_SYSCTL_CTX;
891 * Take down all the tx queues first, as they reference the rx queues
892 * (for egress updates, etc.).
895 free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
897 for_each_txq(pi, i, txq) {
902 for_each_ofld_txq(pi, i, ofld_txq) {
903 free_wrq(sc, ofld_txq);
908 * Then take down the rx queues that forward their interrupts, as they
909 * reference other rx queues.
912 for_each_rxq(pi, i, rxq) {
913 if ((rxq->iq.flags & IQ_INTR) == 0)
918 for_each_ofld_rxq(pi, i, ofld_rxq) {
919 if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
920 free_ofld_rxq(pi, ofld_rxq);
925 * Then take down the rx queues that take direct interrupts.
928 for_each_rxq(pi, i, rxq) {
929 if (rxq->iq.flags & IQ_INTR)
934 for_each_ofld_rxq(pi, i, ofld_rxq) {
935 if (ofld_rxq->iq.flags & IQ_INTR)
936 free_ofld_rxq(pi, ofld_rxq);
944 * Deals with errors and the firmware event queue. All data rx queues forward
945 * their interrupt to the firmware event queue.
948 t4_intr_all(void *arg)
950 struct adapter *sc = arg;
951 struct sge_iq *fwq = &sc->sge.fwq;
954 if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
956 atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
960 /* Deals with error interrupts */
962 t4_intr_err(void *arg)
964 struct adapter *sc = arg;
966 t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
967 t4_slow_intr_handler(sc);
971 t4_intr_evt(void *arg)
973 struct sge_iq *iq = arg;
975 if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
977 atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
984 struct sge_iq *iq = arg;
986 if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
988 atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
993 * Deals with anything and everything on the given ingress queue.
996 service_iq(struct sge_iq *iq, int budget)
999 struct sge_rxq *rxq = iq_to_rxq(iq); /* Use iff iq is part of rxq */
1000 struct sge_fl *fl = &rxq->fl; /* Use iff IQ_HAS_FL */
1001 struct adapter *sc = iq->adapter;
1002 struct rsp_ctrl *ctrl;
1003 const struct rss_header *rss;
1004 int ndescs = 0, limit, fl_bufs_used = 0;
1008 STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1010 limit = budget ? budget : iq->qsize / 8;
1012 KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1015 * We always come back and check the descriptor ring for new indirect
1016 * interrupts and other responses after running a single handler.
1019 while (is_new_response(iq, &ctrl)) {
1024 rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
1025 lq = be32toh(ctrl->pldbuflen_qid);
1026 rss = (const void *)iq->cdesc;
1029 case X_RSPD_TYPE_FLBUF:
1031 KASSERT(iq->flags & IQ_HAS_FL,
1032 ("%s: data for an iq (%p) with no freelist",
1035 m0 = get_fl_payload(sc, fl, lq, &fl_bufs_used);
1036 #ifdef T4_PKT_TIMESTAMP
1038 * 60 bit timestamp for the payload is
1039 * *(uint64_t *)m0->m_pktdat. Note that it is
1040 * in the leading free-space in the mbuf. The
1041 * kernel can clobber it during a pullup,
1042 * m_copymdata, etc. You need to make sure that
1043 * the mbuf reaches you unmolested if you care
1044 * about the timestamp.
1046 *(uint64_t *)m0->m_pktdat =
1047 be64toh(ctrl->u.last_flit) &
1053 case X_RSPD_TYPE_CPL:
1054 KASSERT(rss->opcode < NUM_CPL_CMDS,
1055 ("%s: bad opcode %02x.", __func__,
1057 sc->cpl_handler[rss->opcode](iq, rss, m0);
1060 case X_RSPD_TYPE_INTR:
1063 * Interrupts should be forwarded only to queues
1064 * that are not forwarding their interrupts.
1065 * This means service_iq can recurse but only 1
1068 KASSERT(budget == 0,
1069 ("%s: budget %u, rsp_type %u", __func__,
1073 * There are 1K interrupt-capable queues (qids 0
1074 * through 1023). A response type indicating a
1075 * forwarded interrupt with a qid >= 1K is an
1076 * iWARP async notification.
1079 sc->an_handler(iq, ctrl);
1083 q = sc->sge.iqmap[lq - sc->sge.iq_start];
1084 if (atomic_cmpset_int(&q->state, IQS_IDLE,
1086 if (service_iq(q, q->qsize / 8) == 0) {
1087 atomic_cmpset_int(&q->state,
1088 IQS_BUSY, IQS_IDLE);
1090 STAILQ_INSERT_TAIL(&iql, q,
1098 ("%s: illegal response type %d on iq %p",
1099 __func__, rsp_type, iq));
1101 "%s: illegal response type %d on iq %p",
1102 device_get_nameunit(sc->dev), rsp_type, iq);
1107 if (++ndescs == limit) {
1108 t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
1110 V_INGRESSQID(iq->cntxt_id) |
1111 V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1114 if (fl_bufs_used > 0) {
1116 fl->needed += fl_bufs_used;
1117 refill_fl(sc, fl, fl->cap / 8);
1123 return (EINPROGRESS);
1127 if (STAILQ_EMPTY(&iql))
1131 * Process the head only, and send it to the back of the list if
1132 * it's still not done.
1134 q = STAILQ_FIRST(&iql);
1135 STAILQ_REMOVE_HEAD(&iql, link);
1136 if (service_iq(q, q->qsize / 8) == 0)
1137 atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1139 STAILQ_INSERT_TAIL(&iql, q, link);
1142 #if defined(INET) || defined(INET6)
1143 if (iq->flags & IQ_LRO_ENABLED) {
1144 struct lro_ctrl *lro = &rxq->lro;
1145 struct lro_entry *l;
1147 while (!SLIST_EMPTY(&lro->lro_active)) {
1148 l = SLIST_FIRST(&lro->lro_active);
1149 SLIST_REMOVE_HEAD(&lro->lro_active, next);
1150 tcp_lro_flush(lro, l);
1155 t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
1156 V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1158 if (iq->flags & IQ_HAS_FL) {
1162 fl->needed += fl_bufs_used;
1163 starved = refill_fl(sc, fl, fl->cap / 4);
1165 if (__predict_false(starved != 0))
1166 add_fl_to_sfl(sc, fl);
1172 static struct mbuf *
1173 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf,
1176 struct mbuf *m0, *m;
1177 struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1178 unsigned int nbuf, len;
1181 * No assertion for the fl lock because we don't need it. This routine
1182 * is called only from the rx interrupt handler and it only updates
1183 * fl->cidx. (Contrast that with fl->pidx/fl->needed which could be
1184 * updated in the rx interrupt handler or the starvation helper routine.
1185 * That's why code that manipulates fl->pidx/fl->needed needs the fl
1186 * lock but this routine does not).
1189 if (__predict_false((len_newbuf & F_RSPD_NEWBUF) == 0))
1190 panic("%s: cannot handle packed frames", __func__);
1191 len = G_RSPD_LEN(len_newbuf);
1194 sd->m = NULL; /* consumed */
1196 bus_dmamap_sync(fl->tag[sd->tag_idx], sd->map, BUS_DMASYNC_POSTREAD);
1197 m_init(m0, NULL, 0, M_NOWAIT, MT_DATA, M_PKTHDR);
1198 #ifdef T4_PKT_TIMESTAMP
1199 /* Leave room for a timestamp */
1203 if (len < RX_COPY_THRESHOLD) {
1204 /* copy data to mbuf, buffer will be recycled */
1205 bcopy(sd->cl, mtod(m0, caddr_t), len);
1208 bus_dmamap_unload(fl->tag[sd->tag_idx], sd->map);
1209 m_cljset(m0, sd->cl, FL_BUF_TYPE(sd->tag_idx));
1210 sd->cl = NULL; /* consumed */
1211 m0->m_len = min(len, FL_BUF_SIZE(sd->tag_idx));
1213 m0->m_pkthdr.len = len;
1216 if (__predict_false(++fl->cidx == fl->cap)) {
1223 nbuf = 1; /* # of fl buffers used */
1227 sd->m = NULL; /* consumed */
1230 bus_dmamap_sync(fl->tag[sd->tag_idx], sd->map,
1231 BUS_DMASYNC_POSTREAD);
1233 m_init(m, NULL, 0, M_NOWAIT, MT_DATA, 0);
1235 bcopy(sd->cl, mtod(m, caddr_t), len);
1238 bus_dmamap_unload(fl->tag[sd->tag_idx],
1240 m_cljset(m, sd->cl, FL_BUF_TYPE(sd->tag_idx));
1241 sd->cl = NULL; /* consumed */
1242 m->m_len = min(len, FL_BUF_SIZE(sd->tag_idx));
1246 if (__predict_false(++fl->cidx == fl->cap)) {
1255 (*fl_bufs_used) += nbuf;
1261 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1263 struct sge_rxq *rxq = iq_to_rxq(iq);
1264 struct ifnet *ifp = rxq->ifp;
1265 const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1266 #if defined(INET) || defined(INET6)
1267 struct lro_ctrl *lro = &rxq->lro;
1270 KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1273 m0->m_pkthdr.len -= fl_pktshift;
1274 m0->m_len -= fl_pktshift;
1275 m0->m_data += fl_pktshift;
1277 m0->m_pkthdr.rcvif = ifp;
1278 m0->m_flags |= M_FLOWID;
1279 m0->m_pkthdr.flowid = rss->hash_val;
1281 if (cpl->csum_calc && !cpl->err_vec) {
1282 if (ifp->if_capenable & IFCAP_RXCSUM &&
1283 cpl->l2info & htobe32(F_RXF_IP)) {
1284 m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1285 CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1287 } else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1288 cpl->l2info & htobe32(F_RXF_IP6)) {
1289 m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1294 if (__predict_false(cpl->ip_frag))
1295 m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1297 m0->m_pkthdr.csum_data = 0xffff;
1301 m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1302 m0->m_flags |= M_VLANTAG;
1303 rxq->vlan_extraction++;
1306 #if defined(INET) || defined(INET6)
1307 if (cpl->l2info & htobe32(F_RXF_LRO) &&
1308 iq->flags & IQ_LRO_ENABLED &&
1309 tcp_lro_rx(lro, m0, 0) == 0) {
1310 /* queued for LRO */
1313 ifp->if_input(ifp, m0);
1319 * Doesn't fail. Holds on to work requests it can't send right away.
1322 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
1324 struct sge_eq *eq = &wrq->eq;
1328 TXQ_LOCK_ASSERT_OWNED(wrq);
1330 KASSERT((eq->flags & EQ_TYPEMASK) == EQ_OFLD ||
1331 (eq->flags & EQ_TYPEMASK) == EQ_CTRL,
1332 ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK));
1334 KASSERT((eq->flags & EQ_TYPEMASK) == EQ_CTRL,
1335 ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK));
1338 if (__predict_true(wr != NULL))
1339 STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
1341 can_reclaim = reclaimable(eq);
1342 if (__predict_false(eq->flags & EQ_STALLED)) {
1343 if (can_reclaim < tx_resume_threshold(eq))
1345 eq->flags &= ~EQ_STALLED;
1348 eq->cidx += can_reclaim;
1349 eq->avail += can_reclaim;
1350 if (__predict_false(eq->cidx >= eq->cap))
1351 eq->cidx -= eq->cap;
1353 while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL) {
1356 if (__predict_false(wr->wr_len < 0 ||
1357 wr->wr_len > SGE_MAX_WR_LEN || (wr->wr_len & 0x7))) {
1360 panic("%s: work request with length %d", __func__,
1366 log(LOG_ERR, "%s: %s work request with length %d",
1367 device_get_nameunit(sc->dev), __func__, wr->wr_len);
1368 STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1373 ndesc = howmany(wr->wr_len, EQ_ESIZE);
1374 if (eq->avail < ndesc) {
1379 dst = (void *)&eq->desc[eq->pidx];
1380 copy_to_txd(eq, wrtod(wr), &dst, wr->wr_len);
1384 if (__predict_false(eq->pidx >= eq->cap))
1385 eq->pidx -= eq->cap;
1387 eq->pending += ndesc;
1388 if (eq->pending >= 8)
1392 STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1395 if (eq->avail < 8) {
1396 can_reclaim = reclaimable(eq);
1397 eq->cidx += can_reclaim;
1398 eq->avail += can_reclaim;
1399 if (__predict_false(eq->cidx >= eq->cap))
1400 eq->cidx -= eq->cap;
1408 eq->flags |= EQ_STALLED;
1409 if (callout_pending(&eq->tx_callout) == 0)
1410 callout_reset(&eq->tx_callout, 1, t4_tx_callout, eq);
1414 /* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */
1415 #define TXPKTS_PKT_HDR ((\
1416 sizeof(struct ulp_txpkt) + \
1417 sizeof(struct ulptx_idata) + \
1418 sizeof(struct cpl_tx_pkt_core) \
1421 /* Header of a coalesced tx WR, before SGL of first packet (in flits) */
1422 #define TXPKTS_WR_HDR (\
1423 sizeof(struct fw_eth_tx_pkts_wr) / 8 + \
1426 /* Header of a tx WR, before SGL of first packet (in flits) */
1427 #define TXPKT_WR_HDR ((\
1428 sizeof(struct fw_eth_tx_pkt_wr) + \
1429 sizeof(struct cpl_tx_pkt_core) \
1432 /* Header of a tx LSO WR, before SGL of first packet (in flits) */
1433 #define TXPKT_LSO_WR_HDR ((\
1434 sizeof(struct fw_eth_tx_pkt_wr) + \
1435 sizeof(struct cpl_tx_pkt_lso_core) + \
1436 sizeof(struct cpl_tx_pkt_core) \
1440 t4_eth_tx(struct ifnet *ifp, struct sge_txq *txq, struct mbuf *m)
1442 struct port_info *pi = (void *)ifp->if_softc;
1443 struct adapter *sc = pi->adapter;
1444 struct sge_eq *eq = &txq->eq;
1445 struct buf_ring *br = txq->br;
1447 int rc, coalescing, can_reclaim;
1448 struct txpkts txpkts;
1451 TXQ_LOCK_ASSERT_OWNED(txq);
1452 KASSERT(m, ("%s: called with nothing to do.", __func__));
1453 KASSERT((eq->flags & EQ_TYPEMASK) == EQ_ETH,
1454 ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK));
1456 prefetch(&eq->desc[eq->pidx]);
1457 prefetch(&txq->sdesc[eq->pidx]);
1459 txpkts.npkt = 0;/* indicates there's nothing in txpkts */
1462 can_reclaim = reclaimable(eq);
1463 if (__predict_false(eq->flags & EQ_STALLED)) {
1464 if (can_reclaim < tx_resume_threshold(eq)) {
1468 eq->flags &= ~EQ_STALLED;
1472 if (__predict_false(eq->flags & EQ_DOOMED)) {
1474 while ((m = buf_ring_dequeue_sc(txq->br)) != NULL)
1479 if (eq->avail < 8 && can_reclaim)
1480 reclaim_tx_descs(txq, can_reclaim, 32);
1482 for (; m; m = next ? next : drbr_dequeue(ifp, br)) {
1487 next = m->m_nextpkt;
1488 m->m_nextpkt = NULL;
1490 if (next || buf_ring_peek(br))
1493 rc = get_pkt_sgl(txq, &m, &sgl, coalescing);
1497 /* Short of resources, suspend tx */
1499 m->m_nextpkt = next;
1504 * Unrecoverable error for this packet, throw it away
1505 * and move on to the next. get_pkt_sgl may already
1506 * have freed m (it will be NULL in that case and the
1507 * m_freem here is still safe).
1515 add_to_txpkts(pi, txq, &txpkts, m, &sgl) == 0) {
1517 /* Successfully absorbed into txpkts */
1519 write_ulp_cpl_sgl(pi, txq, &txpkts, m, &sgl);
1524 * We weren't coalescing to begin with, or current frame could
1525 * not be coalesced (add_to_txpkts flushes txpkts if a frame
1526 * given to it can't be coalesced). Either way there should be
1527 * nothing in txpkts.
1529 KASSERT(txpkts.npkt == 0,
1530 ("%s: txpkts not empty: %d", __func__, txpkts.npkt));
1532 /* We're sending out individual packets now */
1536 reclaim_tx_descs(txq, 0, 8);
1537 rc = write_txpkt_wr(pi, txq, m, &sgl);
1540 /* Short of hardware descriptors, suspend tx */
1543 * This is an unlikely but expensive failure. We've
1544 * done all the hard work (DMA mappings etc.) and now we
1545 * can't send out the packet. What's worse, we have to
1546 * spend even more time freeing up everything in sgl.
1549 free_pkt_sgl(txq, &sgl);
1551 m->m_nextpkt = next;
1555 ETHER_BPF_MTAP(ifp, m);
1559 if (eq->pending >= 8)
1562 can_reclaim = reclaimable(eq);
1563 if (can_reclaim >= 32)
1564 reclaim_tx_descs(txq, can_reclaim, 64);
1567 if (txpkts.npkt > 0)
1568 write_txpkts_wr(txq, &txpkts);
1571 * m not NULL means there was an error but we haven't thrown it away.
1572 * This can happen when we're short of tx descriptors (no_desc) or maybe
1573 * even DMA maps (no_dmamap). Either way, a credit flush and reclaim
1574 * will get things going again.
1576 if (m && !(eq->flags & EQ_CRFLUSHED)) {
1577 struct tx_sdesc *txsd = &txq->sdesc[eq->pidx];
1580 * If EQ_CRFLUSHED is not set then we know we have at least one
1581 * available descriptor because any WR that reduces eq->avail to
1582 * 0 also sets EQ_CRFLUSHED.
1584 KASSERT(eq->avail > 0, ("%s: no space for eqflush.", __func__));
1586 txsd->desc_used = 1;
1588 write_eqflush_wr(eq);
1595 reclaim_tx_descs(txq, 0, 128);
1597 if (eq->flags & EQ_STALLED && callout_pending(&eq->tx_callout) == 0)
1598 callout_reset(&eq->tx_callout, 1, t4_tx_callout, eq);
1604 t4_update_fl_bufsize(struct ifnet *ifp)
1606 struct port_info *pi = ifp->if_softc;
1607 struct sge_rxq *rxq;
1609 struct sge_ofld_rxq *ofld_rxq;
1614 bufsize = mtu_to_bufsize(ifp->if_mtu);
1615 for_each_rxq(pi, i, rxq) {
1619 set_fl_tag_idx(fl, bufsize);
1623 bufsize = mtu_to_bufsize_toe(pi->adapter, ifp->if_mtu);
1624 for_each_ofld_rxq(pi, i, ofld_rxq) {
1628 set_fl_tag_idx(fl, bufsize);
1635 can_resume_tx(struct sge_eq *eq)
1637 return (reclaimable(eq) >= tx_resume_threshold(eq));
1641 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
1642 int qsize, int esize)
1644 KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
1645 ("%s: bad tmr_idx %d", __func__, tmr_idx));
1646 KASSERT(pktc_idx < SGE_NCOUNTERS, /* -ve is ok, means don't use */
1647 ("%s: bad pktc_idx %d", __func__, pktc_idx));
1651 iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
1652 iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
1653 if (pktc_idx >= 0) {
1654 iq->intr_params |= F_QINTR_CNT_EN;
1655 iq->intr_pktc_idx = pktc_idx;
1657 iq->qsize = roundup2(qsize, 16); /* See FW_IQ_CMD/iqsize */
1658 iq->esize = max(esize, 16); /* See FW_IQ_CMD/iqesize */
1662 init_fl(struct sge_fl *fl, int qsize, int bufsize, char *name)
1665 strlcpy(fl->lockname, name, sizeof(fl->lockname));
1666 set_fl_tag_idx(fl, bufsize);
1670 init_eq(struct sge_eq *eq, int eqtype, int qsize, uint8_t tx_chan,
1671 uint16_t iqid, char *name)
1673 KASSERT(tx_chan < NCHAN, ("%s: bad tx channel %d", __func__, tx_chan));
1674 KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
1676 eq->flags = eqtype & EQ_TYPEMASK;
1677 eq->tx_chan = tx_chan;
1680 strlcpy(eq->lockname, name, sizeof(eq->lockname));
1682 TASK_INIT(&eq->tx_task, 0, t4_tx_task, eq);
1683 callout_init(&eq->tx_callout, CALLOUT_MPSAFE);
1687 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
1688 bus_dmamap_t *map, bus_addr_t *pa, void **va)
1692 rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
1693 BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
1695 device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
1699 rc = bus_dmamem_alloc(*tag, va,
1700 BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
1702 device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
1706 rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
1708 device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
1713 free_ring(sc, *tag, *map, *pa, *va);
1719 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
1720 bus_addr_t pa, void *va)
1723 bus_dmamap_unload(tag, map);
1725 bus_dmamem_free(tag, va, map);
1727 bus_dma_tag_destroy(tag);
1733 * Allocates the ring for an ingress queue and an optional freelist. If the
1734 * freelist is specified it will be allocated and then associated with the
1737 * Returns errno on failure. Resources allocated up to that point may still be
1738 * allocated. Caller is responsible for cleanup in case this function fails.
1740 * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
1741 * the intr_idx specifies the vector, starting from 0. Otherwise it specifies
1742 * the abs_id of the ingress queue to which its interrupts should be forwarded.
1745 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
1746 int intr_idx, int cong)
1748 int rc, i, cntxt_id;
1751 struct adapter *sc = iq->adapter;
1754 len = iq->qsize * iq->esize;
1755 rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
1756 (void **)&iq->desc);
1760 bzero(&c, sizeof(c));
1761 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
1762 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
1763 V_FW_IQ_CMD_VFN(0));
1765 c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
1768 /* Special handling for firmware event queue */
1769 if (iq == &sc->sge.fwq)
1770 v |= F_FW_IQ_CMD_IQASYNCH;
1772 if (iq->flags & IQ_INTR) {
1773 KASSERT(intr_idx < sc->intr_count,
1774 ("%s: invalid direct intr_idx %d", __func__, intr_idx));
1776 v |= F_FW_IQ_CMD_IQANDST;
1777 v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
1779 c.type_to_iqandstindex = htobe32(v |
1780 V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
1781 V_FW_IQ_CMD_VIID(pi->viid) |
1782 V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
1783 c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
1784 F_FW_IQ_CMD_IQGTSMODE |
1785 V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
1786 V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4));
1787 c.iqsize = htobe16(iq->qsize);
1788 c.iqaddr = htobe64(iq->ba);
1790 c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
1793 mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
1795 for (i = 0; i < FL_BUF_SIZES; i++) {
1798 * A freelist buffer must be 16 byte aligned as the SGE
1799 * uses the low 4 bits of the bus addr to figure out the
1802 rc = bus_dma_tag_create(sc->dmat, 16, 0,
1803 BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
1804 FL_BUF_SIZE(i), 1, FL_BUF_SIZE(i), BUS_DMA_ALLOCNOW,
1805 NULL, NULL, &fl->tag[i]);
1807 device_printf(sc->dev,
1808 "failed to create fl DMA tag[%d]: %d\n",
1813 len = fl->qsize * RX_FL_ESIZE;
1814 rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
1815 &fl->ba, (void **)&fl->desc);
1819 /* Allocate space for one software descriptor per buffer. */
1820 fl->cap = (fl->qsize - spg_len / RX_FL_ESIZE) * 8;
1821 rc = alloc_fl_sdesc(fl);
1823 device_printf(sc->dev,
1824 "failed to setup fl software descriptors: %d\n",
1828 fl->needed = fl->cap;
1829 fl->lowat = roundup2(sc->sge.fl_starve_threshold, 8);
1831 c.iqns_to_fl0congen |=
1832 htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
1833 F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
1834 F_FW_IQ_CMD_FL0PADEN);
1836 c.iqns_to_fl0congen |=
1837 htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
1838 F_FW_IQ_CMD_FL0CONGCIF |
1839 F_FW_IQ_CMD_FL0CONGEN);
1841 c.fl0dcaen_to_fl0cidxfthresh =
1842 htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_64B) |
1843 V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
1844 c.fl0size = htobe16(fl->qsize);
1845 c.fl0addr = htobe64(fl->ba);
1848 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
1850 device_printf(sc->dev,
1851 "failed to create ingress queue: %d\n", rc);
1855 iq->cdesc = iq->desc;
1858 iq->intr_next = iq->intr_params;
1859 iq->cntxt_id = be16toh(c.iqid);
1860 iq->abs_id = be16toh(c.physiqid);
1861 iq->flags |= IQ_ALLOCATED;
1863 cntxt_id = iq->cntxt_id - sc->sge.iq_start;
1864 if (cntxt_id >= sc->sge.niq) {
1865 panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
1866 cntxt_id, sc->sge.niq - 1);
1868 sc->sge.iqmap[cntxt_id] = iq;
1871 fl->cntxt_id = be16toh(c.fl0id);
1872 fl->pidx = fl->cidx = 0;
1874 cntxt_id = fl->cntxt_id - sc->sge.eq_start;
1875 if (cntxt_id >= sc->sge.neq) {
1876 panic("%s: fl->cntxt_id (%d) more than the max (%d)",
1877 __func__, cntxt_id, sc->sge.neq - 1);
1879 sc->sge.eqmap[cntxt_id] = (void *)fl;
1882 /* Enough to make sure the SGE doesn't think it's starved */
1883 refill_fl(sc, fl, fl->lowat);
1886 iq->flags |= IQ_HAS_FL;
1889 if (is_t5(sc) && cong >= 0) {
1890 uint32_t param, val;
1892 param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
1893 V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
1894 V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
1899 for (i = 0; i < 4; i++) {
1900 if (cong & (1 << i))
1901 val |= 1 << (i << 2);
1905 rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, ¶m, &val);
1907 /* report error but carry on */
1908 device_printf(sc->dev,
1909 "failed to set congestion manager context for "
1910 "ingress queue %d: %d\n", iq->cntxt_id, rc);
1914 /* Enable IQ interrupts */
1915 atomic_store_rel_int(&iq->state, IQS_IDLE);
1916 t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
1917 V_INGRESSQID(iq->cntxt_id));
1923 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
1926 struct adapter *sc = iq->adapter;
1930 return (0); /* nothing to do */
1932 dev = pi ? pi->dev : sc->dev;
1934 if (iq->flags & IQ_ALLOCATED) {
1935 rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
1936 FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
1937 fl ? fl->cntxt_id : 0xffff, 0xffff);
1940 "failed to free queue %p: %d\n", iq, rc);
1943 iq->flags &= ~IQ_ALLOCATED;
1946 free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
1948 bzero(iq, sizeof(*iq));
1951 free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
1957 if (mtx_initialized(&fl->fl_lock))
1958 mtx_destroy(&fl->fl_lock);
1960 for (i = 0; i < FL_BUF_SIZES; i++) {
1962 bus_dma_tag_destroy(fl->tag[i]);
1965 bzero(fl, sizeof(*fl));
1972 alloc_fwq(struct adapter *sc)
1975 struct sge_iq *fwq = &sc->sge.fwq;
1976 struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
1977 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1979 init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, FW_IQ_ESIZE);
1980 fwq->flags |= IQ_INTR; /* always */
1981 intr_idx = sc->intr_count > 1 ? 1 : 0;
1982 rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1);
1984 device_printf(sc->dev,
1985 "failed to create firmware event queue: %d\n", rc);
1989 oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
1990 NULL, "firmware event queue");
1991 children = SYSCTL_CHILDREN(oid);
1993 SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id",
1994 CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I",
1995 "absolute id of the queue");
1996 SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id",
1997 CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I",
1998 "SGE context id of the queue");
1999 SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx",
2000 CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I",
2007 free_fwq(struct adapter *sc)
2009 return free_iq_fl(NULL, &sc->sge.fwq, NULL);
2013 alloc_mgmtq(struct adapter *sc)
2016 struct sge_wrq *mgmtq = &sc->sge.mgmtq;
2018 struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2019 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2021 oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
2022 NULL, "management queue");
2024 snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
2025 init_eq(&mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
2026 sc->sge.fwq.cntxt_id, name);
2027 rc = alloc_wrq(sc, NULL, mgmtq, oid);
2029 device_printf(sc->dev,
2030 "failed to create management queue: %d\n", rc);
2038 free_mgmtq(struct adapter *sc)
2041 return free_wrq(sc, &sc->sge.mgmtq);
2045 tnl_cong(struct port_info *pi)
2048 if (cong_drop == -1)
2050 else if (cong_drop == 1)
2053 return (1 << pi->tx_chan);
2057 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int idx,
2058 struct sysctl_oid *oid)
2061 struct sysctl_oid_list *children;
2064 rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx, tnl_cong(pi));
2069 refill_fl(pi->adapter, &rxq->fl, rxq->fl.needed / 8);
2070 FL_UNLOCK(&rxq->fl);
2072 #if defined(INET) || defined(INET6)
2073 rc = tcp_lro_init(&rxq->lro);
2076 rxq->lro.ifp = pi->ifp; /* also indicates LRO init'ed */
2078 if (pi->ifp->if_capenable & IFCAP_LRO)
2079 rxq->iq.flags |= IQ_LRO_ENABLED;
2083 children = SYSCTL_CHILDREN(oid);
2085 snprintf(name, sizeof(name), "%d", idx);
2086 oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2088 children = SYSCTL_CHILDREN(oid);
2090 SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "abs_id",
2091 CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I",
2092 "absolute id of the queue");
2093 SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2094 CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I",
2095 "SGE context id of the queue");
2096 SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2097 CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I",
2099 #if defined(INET) || defined(INET6)
2100 SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
2101 &rxq->lro.lro_queued, 0, NULL);
2102 SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
2103 &rxq->lro.lro_flushed, 0, NULL);
2105 SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
2106 &rxq->rxcsum, "# of times hardware assisted with checksum");
2107 SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_extraction",
2108 CTLFLAG_RD, &rxq->vlan_extraction,
2109 "# of times hardware extracted 802.1Q tag");
2111 children = SYSCTL_CHILDREN(oid);
2112 oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "fl", CTLFLAG_RD,
2114 children = SYSCTL_CHILDREN(oid);
2116 SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2117 CTLTYPE_INT | CTLFLAG_RD, &rxq->fl.cntxt_id, 0, sysctl_uint16, "I",
2118 "SGE context id of the queue");
2119 SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
2120 &rxq->fl.cidx, 0, "consumer index");
2121 SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
2122 &rxq->fl.pidx, 0, "producer index");
2128 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
2132 #if defined(INET) || defined(INET6)
2134 tcp_lro_free(&rxq->lro);
2135 rxq->lro.ifp = NULL;
2139 rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
2141 bzero(rxq, sizeof(*rxq));
2148 alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
2149 int intr_idx, int idx, struct sysctl_oid *oid)
2152 struct sysctl_oid_list *children;
2155 rc = alloc_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
2160 children = SYSCTL_CHILDREN(oid);
2162 snprintf(name, sizeof(name), "%d", idx);
2163 oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2165 children = SYSCTL_CHILDREN(oid);
2167 SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "abs_id",
2168 CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16,
2169 "I", "absolute id of the queue");
2170 SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2171 CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16,
2172 "I", "SGE context id of the queue");
2173 SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2174 CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I",
2177 children = SYSCTL_CHILDREN(oid);
2178 oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "fl", CTLFLAG_RD,
2180 children = SYSCTL_CHILDREN(oid);
2182 SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2183 CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->fl.cntxt_id, 0, sysctl_uint16,
2184 "I", "SGE context id of the queue");
2185 SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
2186 &ofld_rxq->fl.cidx, 0, "consumer index");
2187 SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
2188 &ofld_rxq->fl.pidx, 0, "producer index");
2194 free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq)
2198 rc = free_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl);
2200 bzero(ofld_rxq, sizeof(*ofld_rxq));
2207 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
2210 struct fw_eq_ctrl_cmd c;
2212 bzero(&c, sizeof(c));
2214 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
2215 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
2216 V_FW_EQ_CTRL_CMD_VFN(0));
2217 c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
2218 F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
2219 c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); /* XXX */
2220 c.physeqid_pkd = htobe32(0);
2221 c.fetchszm_to_iqid =
2222 htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
2223 V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
2224 F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
2226 htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2227 V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2228 V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
2229 V_FW_EQ_CTRL_CMD_EQSIZE(eq->qsize));
2230 c.eqaddr = htobe64(eq->ba);
2232 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2234 device_printf(sc->dev,
2235 "failed to create control queue %d: %d\n", eq->tx_chan, rc);
2238 eq->flags |= EQ_ALLOCATED;
2240 eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
2241 cntxt_id = eq->cntxt_id - sc->sge.eq_start;
2242 if (cntxt_id >= sc->sge.neq)
2243 panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
2244 cntxt_id, sc->sge.neq - 1);
2245 sc->sge.eqmap[cntxt_id] = eq;
2251 eth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
2254 struct fw_eq_eth_cmd c;
2256 bzero(&c, sizeof(c));
2258 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
2259 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
2260 V_FW_EQ_ETH_CMD_VFN(0));
2261 c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
2262 F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
2263 c.viid_pkd = htobe32(V_FW_EQ_ETH_CMD_VIID(pi->viid));
2264 c.fetchszm_to_iqid =
2265 htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
2266 V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
2267 V_FW_EQ_ETH_CMD_IQID(eq->iqid));
2268 c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2269 V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2270 V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
2271 V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize));
2272 c.eqaddr = htobe64(eq->ba);
2274 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2276 device_printf(pi->dev,
2277 "failed to create Ethernet egress queue: %d\n", rc);
2280 eq->flags |= EQ_ALLOCATED;
2282 eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
2283 cntxt_id = eq->cntxt_id - sc->sge.eq_start;
2284 if (cntxt_id >= sc->sge.neq)
2285 panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
2286 cntxt_id, sc->sge.neq - 1);
2287 sc->sge.eqmap[cntxt_id] = eq;
2294 ofld_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
2297 struct fw_eq_ofld_cmd c;
2299 bzero(&c, sizeof(c));
2301 c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
2302 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
2303 V_FW_EQ_OFLD_CMD_VFN(0));
2304 c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
2305 F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
2306 c.fetchszm_to_iqid =
2307 htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
2308 V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
2309 F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
2311 htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2312 V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2313 V_FW_EQ_OFLD_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
2314 V_FW_EQ_OFLD_CMD_EQSIZE(eq->qsize));
2315 c.eqaddr = htobe64(eq->ba);
2317 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2319 device_printf(pi->dev,
2320 "failed to create egress queue for TCP offload: %d\n", rc);
2323 eq->flags |= EQ_ALLOCATED;
2325 eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
2326 cntxt_id = eq->cntxt_id - sc->sge.eq_start;
2327 if (cntxt_id >= sc->sge.neq)
2328 panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
2329 cntxt_id, sc->sge.neq - 1);
2330 sc->sge.eqmap[cntxt_id] = eq;
2337 alloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
2342 mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
2344 len = eq->qsize * EQ_ESIZE;
2345 rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
2346 &eq->ba, (void **)&eq->desc);
2350 eq->cap = eq->qsize - spg_len / EQ_ESIZE;
2351 eq->spg = (void *)&eq->desc[eq->cap];
2352 eq->avail = eq->cap - 1; /* one less to avoid cidx = pidx */
2353 eq->pidx = eq->cidx = 0;
2354 eq->doorbells = sc->doorbells;
2356 switch (eq->flags & EQ_TYPEMASK) {
2358 rc = ctrl_eq_alloc(sc, eq);
2362 rc = eth_eq_alloc(sc, pi, eq);
2367 rc = ofld_eq_alloc(sc, pi, eq);
2372 panic("%s: invalid eq type %d.", __func__,
2373 eq->flags & EQ_TYPEMASK);
2376 device_printf(sc->dev,
2377 "failed to allocate egress queue(%d): %d",
2378 eq->flags & EQ_TYPEMASK, rc);
2381 eq->tx_callout.c_cpu = eq->cntxt_id % mp_ncpus;
2383 if (isset(&eq->doorbells, DOORBELL_UDB) ||
2384 isset(&eq->doorbells, DOORBELL_UDBWC) ||
2385 isset(&eq->doorbells, DOORBELL_WCWR)) {
2386 uint32_t s_qpp = sc->sge.s_qpp;
2387 uint32_t mask = (1 << s_qpp) - 1;
2388 volatile uint8_t *udb;
2390 udb = sc->udbs_base + UDBS_DB_OFFSET;
2391 udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT; /* pg offset */
2392 eq->udb_qid = eq->cntxt_id & mask; /* id in page */
2393 if (eq->udb_qid > PAGE_SIZE / UDBS_SEG_SIZE)
2394 clrbit(&eq->doorbells, DOORBELL_WCWR);
2396 udb += eq->udb_qid << UDBS_SEG_SHIFT; /* seg offset */
2399 eq->udb = (volatile void *)udb;
2406 free_eq(struct adapter *sc, struct sge_eq *eq)
2410 if (eq->flags & EQ_ALLOCATED) {
2411 switch (eq->flags & EQ_TYPEMASK) {
2413 rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
2418 rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
2424 rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
2430 panic("%s: invalid eq type %d.", __func__,
2431 eq->flags & EQ_TYPEMASK);
2434 device_printf(sc->dev,
2435 "failed to free egress queue (%d): %d\n",
2436 eq->flags & EQ_TYPEMASK, rc);
2439 eq->flags &= ~EQ_ALLOCATED;
2442 free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
2444 if (mtx_initialized(&eq->eq_lock))
2445 mtx_destroy(&eq->eq_lock);
2447 bzero(eq, sizeof(*eq));
2452 alloc_wrq(struct adapter *sc, struct port_info *pi, struct sge_wrq *wrq,
2453 struct sysctl_oid *oid)
2456 struct sysctl_ctx_list *ctx = pi ? &pi->ctx : &sc->ctx;
2457 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2459 rc = alloc_eq(sc, pi, &wrq->eq);
2464 STAILQ_INIT(&wrq->wr_list);
2466 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
2467 &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
2468 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
2469 CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
2471 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
2472 CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
2474 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs", CTLFLAG_RD,
2475 &wrq->tx_wrs, "# of work requests");
2476 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "no_desc", CTLFLAG_RD,
2478 "# of times queue ran out of hardware descriptors");
2479 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "unstalled", CTLFLAG_RD,
2480 &wrq->eq.unstalled, 0, "# of times queue recovered after stall");
2487 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
2491 rc = free_eq(sc, &wrq->eq);
2495 bzero(wrq, sizeof(*wrq));
2500 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx,
2501 struct sysctl_oid *oid)
2504 struct adapter *sc = pi->adapter;
2505 struct sge_eq *eq = &txq->eq;
2507 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2509 rc = alloc_eq(sc, pi, eq);
2515 txq->sdesc = malloc(eq->cap * sizeof(struct tx_sdesc), M_CXGBE,
2517 txq->br = buf_ring_alloc(eq->qsize, M_CXGBE, M_WAITOK, &eq->eq_lock);
2519 rc = bus_dma_tag_create(sc->dmat, 1, 0, BUS_SPACE_MAXADDR,
2520 BUS_SPACE_MAXADDR, NULL, NULL, 64 * 1024, TX_SGL_SEGS,
2521 BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL, NULL, &txq->tx_tag);
2523 device_printf(sc->dev,
2524 "failed to create tx DMA tag: %d\n", rc);
2529 * We can stuff ~10 frames in an 8-descriptor txpkts WR (8 is the SGE
2530 * limit for any WR). txq->no_dmamap events shouldn't occur if maps is
2531 * sized for the worst case.
2533 rc = t4_alloc_tx_maps(&txq->txmaps, txq->tx_tag, eq->qsize * 10 / 8,
2536 device_printf(sc->dev, "failed to setup tx DMA maps: %d\n", rc);
2540 snprintf(name, sizeof(name), "%d", idx);
2541 oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2543 children = SYSCTL_CHILDREN(oid);
2545 SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
2546 &eq->cntxt_id, 0, "SGE context id of the queue");
2547 SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2548 CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
2550 SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "pidx",
2551 CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
2554 SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
2555 &txq->txcsum, "# of times hardware assisted with checksum");
2556 SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_insertion",
2557 CTLFLAG_RD, &txq->vlan_insertion,
2558 "# of times hardware inserted 802.1Q tag");
2559 SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
2560 &txq->tso_wrs, "# of TSO work requests");
2561 SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
2562 &txq->imm_wrs, "# of work requests with immediate data");
2563 SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
2564 &txq->sgl_wrs, "# of work requests with direct SGL");
2565 SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
2566 &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
2567 SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts_wrs", CTLFLAG_RD,
2568 &txq->txpkts_wrs, "# of txpkts work requests (multiple pkts/WR)");
2569 SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts_pkts", CTLFLAG_RD,
2570 &txq->txpkts_pkts, "# of frames tx'd using txpkts work requests");
2572 SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "br_drops", CTLFLAG_RD,
2573 &txq->br->br_drops, "# of drops in the buf_ring for this queue");
2574 SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "no_dmamap", CTLFLAG_RD,
2575 &txq->no_dmamap, 0, "# of times txq ran out of DMA maps");
2576 SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "no_desc", CTLFLAG_RD,
2577 &txq->no_desc, 0, "# of times txq ran out of hardware descriptors");
2578 SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "egr_update", CTLFLAG_RD,
2579 &eq->egr_update, 0, "egress update notifications from the SGE");
2580 SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "unstalled", CTLFLAG_RD,
2581 &eq->unstalled, 0, "# of times txq recovered after stall");
2587 free_txq(struct port_info *pi, struct sge_txq *txq)
2590 struct adapter *sc = pi->adapter;
2591 struct sge_eq *eq = &txq->eq;
2593 rc = free_eq(sc, eq);
2597 free(txq->sdesc, M_CXGBE);
2599 if (txq->txmaps.maps)
2600 t4_free_tx_maps(&txq->txmaps, txq->tx_tag);
2602 buf_ring_free(txq->br, M_CXGBE);
2605 bus_dma_tag_destroy(txq->tx_tag);
2607 bzero(txq, sizeof(*txq));
2612 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2614 bus_addr_t *ba = arg;
2617 ("%s meant for single segment mappings only.", __func__));
2619 *ba = error ? 0 : segs->ds_addr;
2623 is_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl)
2625 *ctrl = (void *)((uintptr_t)iq->cdesc +
2626 (iq->esize - sizeof(struct rsp_ctrl)));
2628 return (((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen);
2632 iq_next(struct sge_iq *iq)
2634 iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize);
2635 if (__predict_false(++iq->cidx == iq->qsize - 1)) {
2638 iq->cdesc = iq->desc;
2642 #define FL_HW_IDX(x) ((x) >> 3)
2644 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
2646 int ndesc = fl->pending / 8;
2649 if (FL_HW_IDX(fl->pidx) == FL_HW_IDX(fl->cidx))
2650 ndesc--; /* hold back one credit */
2653 return; /* nothing to do */
2655 v = F_DBPRIO | V_QID(fl->cntxt_id) | V_PIDX(ndesc);
2661 t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
2662 fl->pending -= ndesc * 8;
2666 * Fill up the freelist by upto nbufs and maybe ring its doorbell.
2668 * Returns non-zero to indicate that it should be added to the list of starving
2672 refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs)
2674 __be64 *d = &fl->desc[fl->pidx];
2675 struct fl_sdesc *sd = &fl->sdesc[fl->pidx];
2681 FL_LOCK_ASSERT_OWNED(fl);
2683 if (nbufs > fl->needed)
2688 if (sd->cl != NULL) {
2691 * This happens when a frame small enough to fit
2692 * entirely in an mbuf was received in cl last time.
2693 * We'd held on to cl and can reuse it now. Note that
2694 * we reuse a cluster of the old size if fl->tag_idx is
2695 * no longer the same as sd->tag_idx.
2698 KASSERT(*d == sd->ba_tag,
2699 ("%s: recyling problem at pidx %d",
2700 __func__, fl->pidx));
2707 if (fl->tag_idx != sd->tag_idx) {
2709 bus_dma_tag_t newtag = fl->tag[fl->tag_idx];
2710 bus_dma_tag_t oldtag = fl->tag[sd->tag_idx];
2713 * An MTU change can get us here. Discard the old map
2714 * which was created with the old tag, but only if
2715 * we're able to get a new one.
2717 rc = bus_dmamap_create(newtag, 0, &map);
2719 bus_dmamap_destroy(oldtag, sd->map);
2721 sd->tag_idx = fl->tag_idx;
2725 tag = fl->tag[sd->tag_idx];
2727 cl = m_cljget(NULL, M_NOWAIT, FL_BUF_SIZE(sd->tag_idx));
2731 rc = bus_dmamap_load(tag, sd->map, cl, FL_BUF_SIZE(sd->tag_idx),
2732 oneseg_dma_callback, &pa, 0);
2733 if (rc != 0 || pa == 0) {
2734 fl->dmamap_failed++;
2735 uma_zfree(FL_BUF_ZONE(sd->tag_idx), cl);
2740 *d++ = htobe64(pa | sd->tag_idx);
2743 sd->ba_tag = htobe64(pa | sd->tag_idx);
2747 /* sd->m is never recycled, should always be NULL */
2748 KASSERT(sd->m == NULL, ("%s: stray mbuf", __func__));
2750 sd->m = m_gethdr(M_NOWAIT, MT_NOINIT);
2757 if (++fl->pidx == fl->cap) {
2764 if (fl->pending >= 8)
2767 return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
2771 * Attempt to refill all starving freelists.
2774 refill_sfl(void *arg)
2776 struct adapter *sc = arg;
2777 struct sge_fl *fl, *fl_temp;
2779 mtx_lock(&sc->sfl_lock);
2780 TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
2782 refill_fl(sc, fl, 64);
2783 if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
2784 TAILQ_REMOVE(&sc->sfl, fl, link);
2785 fl->flags &= ~FL_STARVING;
2790 if (!TAILQ_EMPTY(&sc->sfl))
2791 callout_schedule(&sc->sfl_callout, hz / 5);
2792 mtx_unlock(&sc->sfl_lock);
2796 alloc_fl_sdesc(struct sge_fl *fl)
2798 struct fl_sdesc *sd;
2802 fl->sdesc = malloc(fl->cap * sizeof(struct fl_sdesc), M_CXGBE,
2805 tag = fl->tag[fl->tag_idx];
2807 for (i = 0; i < fl->cap; i++, sd++) {
2809 sd->tag_idx = fl->tag_idx;
2810 rc = bus_dmamap_create(tag, 0, &sd->map);
2819 bus_dmamap_destroy(tag, sd->map);
2821 m_init(sd->m, NULL, 0, M_NOWAIT, MT_DATA, 0);
2826 KASSERT(sd == fl->sdesc, ("%s: EDOOFUS", __func__));
2828 free(fl->sdesc, M_CXGBE);
2835 free_fl_sdesc(struct sge_fl *fl)
2837 struct fl_sdesc *sd;
2841 for (i = 0; i < fl->cap; i++, sd++) {
2844 m_init(sd->m, NULL, 0, M_NOWAIT, MT_DATA, 0);
2850 bus_dmamap_unload(fl->tag[sd->tag_idx], sd->map);
2851 uma_zfree(FL_BUF_ZONE(sd->tag_idx), sd->cl);
2855 bus_dmamap_destroy(fl->tag[sd->tag_idx], sd->map);
2858 free(fl->sdesc, M_CXGBE);
2863 t4_alloc_tx_maps(struct tx_maps *txmaps, bus_dma_tag_t tx_tag, int count,
2869 txmaps->map_total = txmaps->map_avail = count;
2870 txmaps->map_cidx = txmaps->map_pidx = 0;
2872 txmaps->maps = malloc(count * sizeof(struct tx_map), M_CXGBE,
2876 for (i = 0; i < count; i++, txm++) {
2877 rc = bus_dmamap_create(tx_tag, 0, &txm->map);
2886 bus_dmamap_destroy(tx_tag, txm->map);
2888 KASSERT(txm == txmaps->maps, ("%s: EDOOFUS", __func__));
2890 free(txmaps->maps, M_CXGBE);
2891 txmaps->maps = NULL;
2897 t4_free_tx_maps(struct tx_maps *txmaps, bus_dma_tag_t tx_tag)
2903 for (i = 0; i < txmaps->map_total; i++, txm++) {
2906 bus_dmamap_unload(tx_tag, txm->map);
2911 bus_dmamap_destroy(tx_tag, txm->map);
2914 free(txmaps->maps, M_CXGBE);
2915 txmaps->maps = NULL;
2919 * We'll do immediate data tx for non-TSO, but only when not coalescing. We're
2920 * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes
2921 * of immediate data.
2925 - sizeof(struct fw_eth_tx_pkt_wr) \
2926 - sizeof(struct cpl_tx_pkt_core))
2929 * Returns non-zero on failure, no need to cleanup anything in that case.
2931 * Note 1: We always try to defrag the mbuf if required and return EFBIG only
2932 * if the resulting chain still won't fit in a tx descriptor.
2934 * Note 2: We'll pullup the mbuf chain if TSO is requested and the first mbuf
2935 * does not have the TCP header in it.
2938 get_pkt_sgl(struct sge_txq *txq, struct mbuf **fp, struct sgl *sgl,
2941 struct mbuf *m = *fp;
2942 struct tx_maps *txmaps;
2944 int rc, defragged = 0, n;
2946 TXQ_LOCK_ASSERT_OWNED(txq);
2948 if (m->m_pkthdr.tso_segsz)
2949 sgl_only = 1; /* Do not allow immediate data with LSO */
2951 start: sgl->nsegs = 0;
2953 if (m->m_pkthdr.len <= IMM_LEN && !sgl_only)
2954 return (0); /* nsegs = 0 tells caller to use imm. tx */
2956 txmaps = &txq->txmaps;
2957 if (txmaps->map_avail == 0) {
2961 txm = &txmaps->maps[txmaps->map_pidx];
2963 if (m->m_pkthdr.tso_segsz && m->m_len < 50) {
2964 *fp = m_pullup(m, 50);
2970 rc = bus_dmamap_load_mbuf_sg(txq->tx_tag, txm->map, m, sgl->seg,
2971 &sgl->nsegs, BUS_DMA_NOWAIT);
2972 if (rc == EFBIG && defragged == 0) {
2973 m = m_defrag(m, M_NOWAIT);
2985 txmaps->map_avail--;
2986 if (++txmaps->map_pidx == txmaps->map_total)
2987 txmaps->map_pidx = 0;
2989 KASSERT(sgl->nsegs > 0 && sgl->nsegs <= TX_SGL_SEGS,
2990 ("%s: bad DMA mapping (%d segments)", __func__, sgl->nsegs));
2993 * Store the # of flits required to hold this frame's SGL in nflits. An
2994 * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by
2995 * multiple (len0 + len1, addr0, addr1) tuples. If addr1 is not used
2996 * then len1 must be set to 0.
2999 sgl->nflits = (3 * n) / 2 + (n & 1) + 2;
3006 * Releases all the txq resources used up in the specified sgl.
3009 free_pkt_sgl(struct sge_txq *txq, struct sgl *sgl)
3011 struct tx_maps *txmaps;
3014 TXQ_LOCK_ASSERT_OWNED(txq);
3016 if (sgl->nsegs == 0)
3017 return (0); /* didn't use any map */
3019 txmaps = &txq->txmaps;
3021 /* 1 pkt uses exactly 1 map, back it out */
3023 txmaps->map_avail++;
3024 if (txmaps->map_pidx > 0)
3027 txmaps->map_pidx = txmaps->map_total - 1;
3029 txm = &txmaps->maps[txmaps->map_pidx];
3030 bus_dmamap_unload(txq->tx_tag, txm->map);
3037 write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, struct mbuf *m,
3040 struct sge_eq *eq = &txq->eq;
3041 struct fw_eth_tx_pkt_wr *wr;
3042 struct cpl_tx_pkt_core *cpl;
3043 uint32_t ctrl; /* used in many unrelated places */
3045 int nflits, ndesc, pktlen;
3046 struct tx_sdesc *txsd;
3049 TXQ_LOCK_ASSERT_OWNED(txq);
3051 pktlen = m->m_pkthdr.len;
3054 * Do we have enough flits to send this frame out?
3056 ctrl = sizeof(struct cpl_tx_pkt_core);
3057 if (m->m_pkthdr.tso_segsz) {
3058 nflits = TXPKT_LSO_WR_HDR;
3059 ctrl += sizeof(struct cpl_tx_pkt_lso_core);
3061 nflits = TXPKT_WR_HDR;
3063 nflits += sgl->nflits;
3065 nflits += howmany(pktlen, 8);
3068 ndesc = howmany(nflits, 8);
3069 if (ndesc > eq->avail)
3072 /* Firmware work request header */
3073 wr = (void *)&eq->desc[eq->pidx];
3074 wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
3075 V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
3076 ctrl = V_FW_WR_LEN16(howmany(nflits, 2));
3077 if (eq->avail == ndesc) {
3078 if (!(eq->flags & EQ_CRFLUSHED)) {
3079 ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
3080 eq->flags |= EQ_CRFLUSHED;
3082 eq->flags |= EQ_STALLED;
3085 wr->equiq_to_len16 = htobe32(ctrl);
3088 if (m->m_pkthdr.tso_segsz) {
3089 struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
3090 struct ether_header *eh;
3092 #if defined(INET) || defined(INET6)
3097 ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
3100 eh = mtod(m, struct ether_header *);
3101 eh_type = ntohs(eh->ether_type);
3102 if (eh_type == ETHERTYPE_VLAN) {
3103 struct ether_vlan_header *evh = (void *)eh;
3105 ctrl |= V_LSO_ETHHDR_LEN(1);
3107 eh_type = ntohs(evh->evl_proto);
3113 case ETHERTYPE_IPV6:
3115 struct ip6_hdr *ip6 = l3hdr;
3118 * XXX-BZ For now we do not pretend to support
3119 * IPv6 extension headers.
3121 KASSERT(ip6->ip6_nxt == IPPROTO_TCP, ("%s: CSUM_TSO "
3122 "with ip6_nxt != TCP: %u", __func__, ip6->ip6_nxt));
3123 tcp = (struct tcphdr *)(ip6 + 1);
3125 ctrl |= V_LSO_IPHDR_LEN(sizeof(*ip6) >> 2) |
3126 V_LSO_TCPHDR_LEN(tcp->th_off);
3133 struct ip *ip = l3hdr;
3135 tcp = (void *)((uintptr_t)ip + ip->ip_hl * 4);
3136 ctrl |= V_LSO_IPHDR_LEN(ip->ip_hl) |
3137 V_LSO_TCPHDR_LEN(tcp->th_off);
3142 panic("%s: CSUM_TSO but no supported IP version "
3143 "(0x%04x)", __func__, eh_type);
3146 lso->lso_ctrl = htobe32(ctrl);
3147 lso->ipid_ofst = htobe16(0);
3148 lso->mss = htobe16(m->m_pkthdr.tso_segsz);
3149 lso->seqno_offset = htobe32(0);
3150 lso->len = htobe32(pktlen);
3152 cpl = (void *)(lso + 1);
3156 cpl = (void *)(wr + 1);
3158 /* Checksum offload */
3160 if (!(m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO)))
3161 ctrl1 |= F_TXPKT_IPCSUM_DIS;
3162 if (!(m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
3163 CSUM_TCP_IPV6 | CSUM_TSO)))
3164 ctrl1 |= F_TXPKT_L4CSUM_DIS;
3165 if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
3166 CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
3167 txq->txcsum++; /* some hardware assistance provided */
3169 /* VLAN tag insertion */
3170 if (m->m_flags & M_VLANTAG) {
3171 ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
3172 txq->vlan_insertion++;
3176 cpl->ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3177 V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
3179 cpl->len = htobe16(pktlen);
3180 cpl->ctrl1 = htobe64(ctrl1);
3182 /* Software descriptor */
3183 txsd = &txq->sdesc[eq->pidx];
3184 txsd->desc_used = ndesc;
3186 eq->pending += ndesc;
3189 if (eq->pidx >= eq->cap)
3190 eq->pidx -= eq->cap;
3193 dst = (void *)(cpl + 1);
3194 if (sgl->nsegs > 0) {
3197 write_sgl_to_txd(eq, sgl, &dst);
3201 for (; m; m = m->m_next) {
3202 copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
3208 KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
3218 * Returns 0 to indicate that m has been accepted into a coalesced tx work
3219 * request. It has either been folded into txpkts or txpkts was flushed and m
3220 * has started a new coalesced work request (as the first frame in a fresh
3223 * Returns non-zero to indicate a failure - caller is responsible for
3224 * transmitting m, if there was anything in txpkts it has been flushed.
3227 add_to_txpkts(struct port_info *pi, struct sge_txq *txq, struct txpkts *txpkts,
3228 struct mbuf *m, struct sgl *sgl)
3230 struct sge_eq *eq = &txq->eq;
3232 struct tx_sdesc *txsd;
3235 TXQ_LOCK_ASSERT_OWNED(txq);
3237 KASSERT(sgl->nsegs, ("%s: can't coalesce imm data", __func__));
3239 if (txpkts->npkt > 0) {
3240 flits = TXPKTS_PKT_HDR + sgl->nflits;
3241 can_coalesce = m->m_pkthdr.tso_segsz == 0 &&
3242 txpkts->nflits + flits <= TX_WR_FLITS &&
3243 txpkts->nflits + flits <= eq->avail * 8 &&
3244 txpkts->plen + m->m_pkthdr.len < 65536;
3248 txpkts->nflits += flits;
3249 txpkts->plen += m->m_pkthdr.len;
3251 txsd = &txq->sdesc[eq->pidx];
3258 * Couldn't coalesce m into txpkts. The first order of business
3259 * is to send txpkts on its way. Then we'll revisit m.
3261 write_txpkts_wr(txq, txpkts);
3265 * Check if we can start a new coalesced tx work request with m as
3266 * the first packet in it.
3269 KASSERT(txpkts->npkt == 0, ("%s: txpkts not empty", __func__));
3271 flits = TXPKTS_WR_HDR + sgl->nflits;
3272 can_coalesce = m->m_pkthdr.tso_segsz == 0 &&
3273 flits <= eq->avail * 8 && flits <= TX_WR_FLITS;
3275 if (can_coalesce == 0)
3279 * Start a fresh coalesced tx WR with m as the first frame in it.
3282 txpkts->nflits = flits;
3283 txpkts->flitp = &eq->desc[eq->pidx].flit[2];
3284 txpkts->plen = m->m_pkthdr.len;
3286 txsd = &txq->sdesc[eq->pidx];
3293 * Note that write_txpkts_wr can never run out of hardware descriptors (but
3294 * write_txpkt_wr can). add_to_txpkts ensures that a frame is accepted for
3295 * coalescing only if sufficient hardware descriptors are available.
3298 write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts)
3300 struct sge_eq *eq = &txq->eq;
3301 struct fw_eth_tx_pkts_wr *wr;
3302 struct tx_sdesc *txsd;
3306 TXQ_LOCK_ASSERT_OWNED(txq);
3308 ndesc = howmany(txpkts->nflits, 8);
3310 wr = (void *)&eq->desc[eq->pidx];
3311 wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
3312 ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2));
3313 if (eq->avail == ndesc) {
3314 if (!(eq->flags & EQ_CRFLUSHED)) {
3315 ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
3316 eq->flags |= EQ_CRFLUSHED;
3318 eq->flags |= EQ_STALLED;
3320 wr->equiq_to_len16 = htobe32(ctrl);
3321 wr->plen = htobe16(txpkts->plen);
3322 wr->npkt = txpkts->npkt;
3323 wr->r3 = wr->type = 0;
3325 /* Everything else already written */
3327 txsd = &txq->sdesc[eq->pidx];
3328 txsd->desc_used = ndesc;
3330 KASSERT(eq->avail >= ndesc, ("%s: out of descriptors", __func__));
3332 eq->pending += ndesc;
3335 if (eq->pidx >= eq->cap)
3336 eq->pidx -= eq->cap;
3338 txq->txpkts_pkts += txpkts->npkt;
3340 txpkts->npkt = 0; /* emptied */
3344 write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
3345 struct txpkts *txpkts, struct mbuf *m, struct sgl *sgl)
3347 struct ulp_txpkt *ulpmc;
3348 struct ulptx_idata *ulpsc;
3349 struct cpl_tx_pkt_core *cpl;
3350 struct sge_eq *eq = &txq->eq;
3351 uintptr_t flitp, start, end;
3355 KASSERT(txpkts->npkt > 0, ("%s: txpkts is empty", __func__));
3357 start = (uintptr_t)eq->desc;
3358 end = (uintptr_t)eq->spg;
3360 /* Checksum offload */
3362 if (!(m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO)))
3363 ctrl |= F_TXPKT_IPCSUM_DIS;
3364 if (!(m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
3365 CSUM_TCP_IPV6 | CSUM_TSO)))
3366 ctrl |= F_TXPKT_L4CSUM_DIS;
3367 if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
3368 CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
3369 txq->txcsum++; /* some hardware assistance provided */
3371 /* VLAN tag insertion */
3372 if (m->m_flags & M_VLANTAG) {
3373 ctrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
3374 txq->vlan_insertion++;
3378 * The previous packet's SGL must have ended at a 16 byte boundary (this
3379 * is required by the firmware/hardware). It follows that flitp cannot
3380 * wrap around between the ULPTX master command and ULPTX subcommand (8
3381 * bytes each), and that it can not wrap around in the middle of the
3382 * cpl_tx_pkt_core either.
3384 flitp = (uintptr_t)txpkts->flitp;
3385 KASSERT((flitp & 0xf) == 0,
3386 ("%s: last SGL did not end at 16 byte boundary: %p",
3387 __func__, txpkts->flitp));
3389 /* ULP master command */
3390 ulpmc = (void *)flitp;
3391 ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0) |
3392 V_ULP_TXPKT_FID(eq->iqid));
3393 ulpmc->len = htonl(howmany(sizeof(*ulpmc) + sizeof(*ulpsc) +
3394 sizeof(*cpl) + 8 * sgl->nflits, 16));
3396 /* ULP subcommand */
3397 ulpsc = (void *)(ulpmc + 1);
3398 ulpsc->cmd_more = htobe32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) |
3400 ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
3402 flitp += sizeof(*ulpmc) + sizeof(*ulpsc);
3407 cpl = (void *)flitp;
3408 cpl->ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3409 V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
3411 cpl->len = htobe16(m->m_pkthdr.len);
3412 cpl->ctrl1 = htobe64(ctrl);
3414 flitp += sizeof(*cpl);
3418 /* SGL for this frame */
3419 dst = (caddr_t)flitp;
3420 txpkts->nflits += write_sgl_to_txd(eq, sgl, &dst);
3421 txpkts->flitp = (void *)dst;
3423 KASSERT(((uintptr_t)dst & 0xf) == 0,
3424 ("%s: SGL ends at %p (not a 16 byte boundary)", __func__, dst));
3428 * If the SGL ends on an address that is not 16 byte aligned, this function will
3429 * add a 0 filled flit at the end. It returns 1 in that case.
3432 write_sgl_to_txd(struct sge_eq *eq, struct sgl *sgl, caddr_t *to)
3434 __be64 *flitp, *end;
3435 struct ulptx_sgl *usgl;
3436 bus_dma_segment_t *seg;
3439 KASSERT(sgl->nsegs > 0 && sgl->nflits > 0,
3440 ("%s: bad SGL - nsegs=%d, nflits=%d",
3441 __func__, sgl->nsegs, sgl->nflits));
3443 KASSERT(((uintptr_t)(*to) & 0xf) == 0,
3444 ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
3446 flitp = (__be64 *)(*to);
3447 end = flitp + sgl->nflits;
3449 usgl = (void *)flitp;
3452 * We start at a 16 byte boundary somewhere inside the tx descriptor
3453 * ring, so we're at least 16 bytes away from the status page. There is
3454 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
3457 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
3458 V_ULPTX_NSGE(sgl->nsegs));
3459 usgl->len0 = htobe32(seg->ds_len);
3460 usgl->addr0 = htobe64(seg->ds_addr);
3463 if ((uintptr_t)end <= (uintptr_t)eq->spg) {
3465 /* Won't wrap around at all */
3467 for (i = 0; i < sgl->nsegs - 1; i++, seg++) {
3468 usgl->sge[i / 2].len[i & 1] = htobe32(seg->ds_len);
3469 usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ds_addr);
3472 usgl->sge[i / 2].len[1] = htobe32(0);
3475 /* Will wrap somewhere in the rest of the SGL */
3477 /* 2 flits already written, write the rest flit by flit */
3478 flitp = (void *)(usgl + 1);
3479 for (i = 0; i < sgl->nflits - 2; i++) {
3480 if ((uintptr_t)flitp == (uintptr_t)eq->spg)
3481 flitp = (void *)eq->desc;
3482 *flitp++ = get_flit(seg, sgl->nsegs - 1, i);
3487 if ((uintptr_t)end & 0xf) {
3488 *(uint64_t *)end = 0;
3494 if ((uintptr_t)end == (uintptr_t)eq->spg)
3495 *to = (void *)eq->desc;
3503 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
3505 if (__predict_true((uintptr_t)(*to) + len <= (uintptr_t)eq->spg)) {
3506 bcopy(from, *to, len);
3509 int portion = (uintptr_t)eq->spg - (uintptr_t)(*to);
3511 bcopy(from, *to, portion);
3513 portion = len - portion; /* remaining */
3514 bcopy(from, (void *)eq->desc, portion);
3515 (*to) = (caddr_t)eq->desc + portion;
3520 ring_eq_db(struct adapter *sc, struct sge_eq *eq)
3525 pending = eq->pending;
3527 clrbit(&db, DOORBELL_WCWR);
3531 switch (ffs(db) - 1) {
3533 *eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(pending));
3536 case DOORBELL_WCWR: {
3537 volatile uint64_t *dst, *src;
3541 * Queues whose 128B doorbell segment fits in the page do not
3542 * use relative qid (udb_qid is always 0). Only queues with
3543 * doorbell segments can do WCWR.
3545 KASSERT(eq->udb_qid == 0 && pending == 1,
3546 ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
3547 __func__, eq->doorbells, pending, eq->pidx, eq));
3549 dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
3551 i = eq->pidx ? eq->pidx - 1 : eq->cap - 1;
3552 src = (void *)&eq->desc[i];
3553 while (src != (void *)&eq->desc[i + 1])
3559 case DOORBELL_UDBWC:
3560 *eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(pending));
3565 t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
3566 V_QID(eq->cntxt_id) | V_PIDX(pending));
3572 reclaimable(struct sge_eq *eq)
3576 cidx = eq->spg->cidx; /* stable snapshot */
3577 cidx = be16toh(cidx);
3579 if (cidx >= eq->cidx)
3580 return (cidx - eq->cidx);
3582 return (cidx + eq->cap - eq->cidx);
3586 * There are "can_reclaim" tx descriptors ready to be reclaimed. Reclaim as
3587 * many as possible but stop when there are around "n" mbufs to free.
3589 * The actual number reclaimed is provided as the return value.
3592 reclaim_tx_descs(struct sge_txq *txq, int can_reclaim, int n)
3594 struct tx_sdesc *txsd;
3595 struct tx_maps *txmaps;
3597 unsigned int reclaimed, maps;
3598 struct sge_eq *eq = &txq->eq;
3600 TXQ_LOCK_ASSERT_OWNED(txq);
3602 if (can_reclaim == 0)
3603 can_reclaim = reclaimable(eq);
3605 maps = reclaimed = 0;
3606 while (can_reclaim && maps < n) {
3609 txsd = &txq->sdesc[eq->cidx];
3610 ndesc = txsd->desc_used;
3612 /* Firmware doesn't return "partial" credits. */
3613 KASSERT(can_reclaim >= ndesc,
3614 ("%s: unexpected number of credits: %d, %d",
3615 __func__, can_reclaim, ndesc));
3617 maps += txsd->credits;
3620 can_reclaim -= ndesc;
3623 if (__predict_false(eq->cidx >= eq->cap))
3624 eq->cidx -= eq->cap;
3627 txmaps = &txq->txmaps;
3628 txm = &txmaps->maps[txmaps->map_cidx];
3632 eq->avail += reclaimed;
3633 KASSERT(eq->avail < eq->cap, /* avail tops out at (cap - 1) */
3634 ("%s: too many descriptors available", __func__));
3636 txmaps->map_avail += maps;
3637 KASSERT(txmaps->map_avail <= txmaps->map_total,
3638 ("%s: too many maps available", __func__));
3641 struct tx_map *next;
3644 if (__predict_false(txmaps->map_cidx + 1 == txmaps->map_total))
3645 next = txmaps->maps;
3648 bus_dmamap_unload(txq->tx_tag, txm->map);
3653 if (__predict_false(++txmaps->map_cidx == txmaps->map_total))
3654 txmaps->map_cidx = 0;
3661 write_eqflush_wr(struct sge_eq *eq)
3663 struct fw_eq_flush_wr *wr;
3665 EQ_LOCK_ASSERT_OWNED(eq);
3666 KASSERT(eq->avail > 0, ("%s: no descriptors left.", __func__));
3667 KASSERT(!(eq->flags & EQ_CRFLUSHED), ("%s: flushed already", __func__));
3669 wr = (void *)&eq->desc[eq->pidx];
3670 bzero(wr, sizeof(*wr));
3671 wr->opcode = FW_EQ_FLUSH_WR;
3672 wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(sizeof(*wr) / 16) |
3673 F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
3675 eq->flags |= (EQ_CRFLUSHED | EQ_STALLED);
3678 if (++eq->pidx == eq->cap)
3683 get_flit(bus_dma_segment_t *sgl, int nsegs, int idx)
3685 int i = (idx / 3) * 2;
3691 rc = htobe32(sgl[i].ds_len);
3693 rc |= (uint64_t)htobe32(sgl[i + 1].ds_len) << 32;
3698 return htobe64(sgl[i].ds_addr);
3700 return htobe64(sgl[i + 1].ds_addr);
3707 set_fl_tag_idx(struct sge_fl *fl, int bufsize)
3711 for (i = 0; i < FL_BUF_SIZES - 1; i++) {
3712 if (FL_BUF_SIZE(i) >= bufsize)
3720 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
3722 mtx_lock(&sc->sfl_lock);
3724 if ((fl->flags & FL_DOOMED) == 0) {
3725 fl->flags |= FL_STARVING;
3726 TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
3727 callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
3730 mtx_unlock(&sc->sfl_lock);
3734 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
3737 const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
3738 unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
3739 struct adapter *sc = iq->adapter;
3740 struct sge *s = &sc->sge;
3743 KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
3746 eq = s->eqmap[qid - s->eq_start];
3748 KASSERT(eq->flags & EQ_CRFLUSHED,
3749 ("%s: unsolicited egress update", __func__));
3750 eq->flags &= ~EQ_CRFLUSHED;
3753 if (__predict_false(eq->flags & EQ_DOOMED))
3755 else if (eq->flags & EQ_STALLED && can_resume_tx(eq))
3756 taskqueue_enqueue(sc->tq[eq->tx_chan], &eq->tx_task);
3762 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
3763 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
3764 offsetof(struct cpl_fw6_msg, data));
3767 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
3769 struct adapter *sc = iq->adapter;
3770 const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
3772 KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
3775 if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
3776 const struct rss_header *rss2;
3778 rss2 = (const struct rss_header *)&cpl->data[0];
3779 return (sc->cpl_handler[rss2->opcode](iq, rss2, m));
3782 return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0]));
3786 sysctl_uint16(SYSCTL_HANDLER_ARGS)
3788 uint16_t *id = arg1;
3791 return sysctl_handle_int(oidp, &i, 0, req);