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>
38 #include <sys/malloc.h>
39 #include <sys/queue.h>
41 #include <sys/taskqueue.h>
43 #include <sys/sglist.h>
44 #include <sys/sysctl.h>
46 #include <sys/counter.h>
48 #include <net/ethernet.h>
50 #include <net/if_vlan_var.h>
51 #include <netinet/in.h>
52 #include <netinet/ip.h>
53 #include <netinet/ip6.h>
54 #include <netinet/tcp.h>
55 #include <machine/md_var.h>
59 #include <machine/bus.h>
60 #include <sys/selinfo.h>
61 #include <net/if_var.h>
62 #include <net/netmap.h>
63 #include <dev/netmap/netmap_kern.h>
66 #include "common/common.h"
67 #include "common/t4_regs.h"
68 #include "common/t4_regs_values.h"
69 #include "common/t4_msg.h"
71 #include "t4_mp_ring.h"
73 #ifdef T4_PKT_TIMESTAMP
74 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
76 #define RX_COPY_THRESHOLD MINCLSIZE
80 * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
81 * 0-7 are valid values.
84 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
87 * Pad ethernet payload up to this boundary.
88 * -1: driver should figure out a good value.
90 * Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
93 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
97 * -1: driver should figure out a good value.
98 * 64 or 128 are the only other valid values.
101 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
105 * -1: no congestion feedback (not recommended).
106 * 0: backpressure the channel instead of dropping packets right away.
107 * 1: no backpressure, drop packets for the congested queue immediately.
109 static int cong_drop = 0;
110 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
113 * Deliver multiple frames in the same free list buffer if they fit.
114 * -1: let the driver decide whether to enable buffer packing or not.
115 * 0: disable buffer packing.
116 * 1: enable buffer packing.
118 static int buffer_packing = -1;
119 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
122 * Start next frame in a packed buffer at this boundary.
123 * -1: driver should figure out a good value.
124 * T4: driver will ignore this and use the same value as fl_pad above.
125 * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
127 static int fl_pack = -1;
128 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
131 * Allow the driver to create mbuf(s) in a cluster allocated for rx.
132 * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
133 * 1: ok to create mbuf(s) within a cluster if there is room.
135 static int allow_mbufs_in_cluster = 1;
136 TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster);
139 * Largest rx cluster size that the driver is allowed to allocate.
141 static int largest_rx_cluster = MJUM16BYTES;
142 TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster);
145 * Size of cluster allocation that's most likely to succeed. The driver will
146 * fall back to this size if it fails to allocate clusters larger than this.
148 static int safest_rx_cluster = PAGE_SIZE;
149 TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster);
152 u_int wr_type; /* type 0 or type 1 */
153 u_int npkt; /* # of packets in this work request */
154 u_int plen; /* total payload (sum of all packets) */
155 u_int len16; /* # of 16B pieces used by this work request */
158 /* A packet's SGL. This + m_pkthdr has all info needed for tx */
161 struct sglist_seg seg[TX_SGL_SEGS];
164 static int service_iq(struct sge_iq *, int);
165 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
166 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
167 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
168 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
169 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
171 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
172 bus_addr_t *, void **);
173 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
175 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
177 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
178 static void add_fl_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
180 static int alloc_fwq(struct adapter *);
181 static int free_fwq(struct adapter *);
182 static int alloc_mgmtq(struct adapter *);
183 static int free_mgmtq(struct adapter *);
184 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
185 struct sysctl_oid *);
186 static int free_rxq(struct vi_info *, struct sge_rxq *);
188 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
189 struct sysctl_oid *);
190 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
193 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
194 struct sysctl_oid *);
195 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
196 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
197 struct sysctl_oid *);
198 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
200 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
201 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
203 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
205 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
206 static int free_eq(struct adapter *, struct sge_eq *);
207 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
208 struct sysctl_oid *);
209 static int free_wrq(struct adapter *, struct sge_wrq *);
210 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
211 struct sysctl_oid *);
212 static int free_txq(struct vi_info *, struct sge_txq *);
213 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
214 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
215 static int refill_fl(struct adapter *, struct sge_fl *, int);
216 static void refill_sfl(void *);
217 static int alloc_fl_sdesc(struct sge_fl *);
218 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
219 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
220 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
221 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
223 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
224 static inline u_int txpkt_len16(u_int, u_int);
225 static inline u_int txpkts0_len16(u_int);
226 static inline u_int txpkts1_len16(void);
227 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
228 struct mbuf *, u_int);
229 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
230 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
231 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
232 struct mbuf *, const struct txpkts *, u_int);
233 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
234 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
235 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
236 static inline uint16_t read_hw_cidx(struct sge_eq *);
237 static inline u_int reclaimable_tx_desc(struct sge_eq *);
238 static inline u_int total_available_tx_desc(struct sge_eq *);
239 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
240 static void tx_reclaim(void *, int);
241 static __be64 get_flit(struct sglist_seg *, int, int);
242 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
244 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
246 static void wrq_tx_drain(void *, int);
247 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
249 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
250 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
251 static int sysctl_tc(SYSCTL_HANDLER_ARGS);
253 static counter_u64_t extfree_refs;
254 static counter_u64_t extfree_rels;
256 an_handler_t t4_an_handler;
257 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
258 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
262 an_not_handled(struct sge_iq *iq, const struct rsp_ctrl *ctrl)
266 panic("%s: async notification on iq %p (ctrl %p)", __func__, iq, ctrl);
268 log(LOG_ERR, "%s: async notification on iq %p (ctrl %p)\n",
275 t4_register_an_handler(an_handler_t h)
279 new = h ? (uintptr_t)h : (uintptr_t)an_not_handled;
280 loc = (uintptr_t *) &t4_an_handler;
281 atomic_store_rel_ptr(loc, new);
287 fw_msg_not_handled(struct adapter *sc, const __be64 *rpl)
289 const struct cpl_fw6_msg *cpl =
290 __containerof(rpl, struct cpl_fw6_msg, data[0]);
293 panic("%s: fw_msg type %d", __func__, cpl->type);
295 log(LOG_ERR, "%s: fw_msg type %d\n", __func__, cpl->type);
301 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
305 if (type >= nitems(t4_fw_msg_handler))
309 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
310 * handler dispatch table. Reject any attempt to install a handler for
313 if (type == FW_TYPE_RSSCPL || type == FW6_TYPE_RSSCPL)
316 new = h ? (uintptr_t)h : (uintptr_t)fw_msg_not_handled;
317 loc = (uintptr_t *) &t4_fw_msg_handler[type];
318 atomic_store_rel_ptr(loc, new);
324 cpl_not_handled(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
328 panic("%s: opcode 0x%02x on iq %p with payload %p",
329 __func__, rss->opcode, iq, m);
331 log(LOG_ERR, "%s: opcode 0x%02x on iq %p with payload %p\n",
332 __func__, rss->opcode, iq, m);
339 t4_register_cpl_handler(int opcode, cpl_handler_t h)
343 if (opcode >= nitems(t4_cpl_handler))
346 new = h ? (uintptr_t)h : (uintptr_t)cpl_not_handled;
347 loc = (uintptr_t *) &t4_cpl_handler[opcode];
348 atomic_store_rel_ptr(loc, new);
354 * Called on MOD_LOAD. Validates and calculates the SGE tunables.
361 if (fl_pktshift < 0 || fl_pktshift > 7) {
362 printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
363 " using 2 instead.\n", fl_pktshift);
367 if (spg_len != 64 && spg_len != 128) {
370 #if defined(__i386__) || defined(__amd64__)
371 len = cpu_clflush_line_size > 64 ? 128 : 64;
376 printf("Invalid hw.cxgbe.spg_len value (%d),"
377 " using %d instead.\n", spg_len, len);
382 if (cong_drop < -1 || cong_drop > 1) {
383 printf("Invalid hw.cxgbe.cong_drop value (%d),"
384 " using 0 instead.\n", cong_drop);
388 extfree_refs = counter_u64_alloc(M_WAITOK);
389 extfree_rels = counter_u64_alloc(M_WAITOK);
390 counter_u64_zero(extfree_refs);
391 counter_u64_zero(extfree_rels);
393 t4_an_handler = an_not_handled;
394 for (i = 0; i < nitems(t4_fw_msg_handler); i++)
395 t4_fw_msg_handler[i] = fw_msg_not_handled;
396 for (i = 0; i < nitems(t4_cpl_handler); i++)
397 t4_cpl_handler[i] = cpl_not_handled;
399 t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
400 t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
401 t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
402 t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
403 t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
407 t4_sge_modunload(void)
410 counter_u64_free(extfree_refs);
411 counter_u64_free(extfree_rels);
415 t4_sge_extfree_refs(void)
419 rels = counter_u64_fetch(extfree_rels);
420 refs = counter_u64_fetch(extfree_refs);
422 return (refs - rels);
426 setup_pad_and_pack_boundaries(struct adapter *sc)
432 if (fl_pad < 32 || fl_pad > 4096 || !powerof2(fl_pad)) {
434 * If there is any chance that we might use buffer packing and
435 * the chip is a T4, then pick 64 as the pad/pack boundary. Set
436 * it to 32 in all other cases.
438 pad = is_t4(sc) && buffer_packing ? 64 : 32;
441 * For fl_pad = 0 we'll still write a reasonable value to the
442 * register but all the freelists will opt out of padding.
443 * We'll complain here only if the user tried to set it to a
444 * value greater than 0 that was invalid.
447 device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
448 " (%d), using %d instead.\n", fl_pad, pad);
451 m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
452 v = V_INGPADBOUNDARY(ilog2(pad) - 5);
453 t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
456 if (fl_pack != -1 && fl_pack != pad) {
457 /* Complain but carry on. */
458 device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
459 " using %d instead.\n", fl_pack, pad);
465 if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
466 !powerof2(fl_pack)) {
467 pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
468 MPASS(powerof2(pack));
476 device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
477 " (%d), using %d instead.\n", fl_pack, pack);
480 m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
482 v = V_INGPACKBOUNDARY(0);
484 v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
486 MPASS(!is_t4(sc)); /* T4 doesn't have SGE_CONTROL2 */
487 t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
491 * adap->params.vpd.cclk must be set up before this is called.
494 t4_tweak_chip_settings(struct adapter *sc)
498 int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
499 int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
500 int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
501 uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
502 static int sge_flbuf_sizes[] = {
504 #if MJUMPAGESIZE != MCLBYTES
506 MJUMPAGESIZE - CL_METADATA_SIZE,
507 MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
511 MCLBYTES - MSIZE - CL_METADATA_SIZE,
512 MJUM9BYTES - CL_METADATA_SIZE,
513 MJUM16BYTES - CL_METADATA_SIZE,
516 KASSERT(sc->flags & MASTER_PF,
517 ("%s: trying to change chip settings when not master.", __func__));
519 m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
520 v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
521 V_EGRSTATUSPAGESIZE(spg_len == 128);
522 t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
524 setup_pad_and_pack_boundaries(sc);
526 v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
527 V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
528 V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
529 V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
530 V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
531 V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
532 V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
533 V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
534 t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
536 KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
537 ("%s: hw buffer size table too big", __func__));
538 for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
539 t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
543 v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
544 V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
545 t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
547 KASSERT(intr_timer[0] <= timer_max,
548 ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
550 for (i = 1; i < nitems(intr_timer); i++) {
551 KASSERT(intr_timer[i] >= intr_timer[i - 1],
552 ("%s: timers not listed in increasing order (%d)",
555 while (intr_timer[i] > timer_max) {
556 if (i == nitems(intr_timer) - 1) {
557 intr_timer[i] = timer_max;
560 intr_timer[i] += intr_timer[i - 1];
565 v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
566 V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
567 t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
568 v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
569 V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
570 t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
571 v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
572 V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
573 t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
575 /* 4K, 16K, 64K, 256K DDP "page sizes" */
576 v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
577 t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
579 m = v = F_TDDPTAGTCB;
580 t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
582 m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
584 v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
585 t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
589 * SGE wants the buffer to be at least 64B and then a multiple of 16. If
590 * padding is is use the buffer's start and end need to be aligned to the pad
591 * boundary as well. We'll just make sure that the size is a multiple of the
592 * boundary here, it is up to the buffer allocation code to make sure the start
593 * of the buffer is aligned as well.
596 hwsz_ok(struct adapter *sc, int hwsz)
598 int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
600 return (hwsz >= 64 && (hwsz & mask) == 0);
604 * XXX: driver really should be able to deal with unexpected settings.
607 t4_read_chip_settings(struct adapter *sc)
609 struct sge *s = &sc->sge;
610 struct sge_params *sp = &sc->params.sge;
613 uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
614 static int sw_buf_sizes[] = { /* Sorted by size */
616 #if MJUMPAGESIZE != MCLBYTES
622 struct sw_zone_info *swz, *safe_swz;
623 struct hw_buf_info *hwb;
625 t4_init_sge_params(sc);
629 r = t4_read_reg(sc, A_SGE_CONTROL);
631 device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
636 * If this changes then every single use of PAGE_SHIFT in the driver
637 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
639 if (sp->page_shift != PAGE_SHIFT) {
640 device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
644 /* Filter out unusable hw buffer sizes entirely (mark with -2). */
645 hwb = &s->hw_buf_info[0];
646 for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
647 r = t4_read_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i));
649 hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
654 * Create a sorted list in decreasing order of hw buffer sizes (and so
655 * increasing order of spare area) for each software zone.
657 * If padding is enabled then the start and end of the buffer must align
658 * to the pad boundary; if packing is enabled then they must align with
659 * the pack boundary as well. Allocations from the cluster zones are
660 * aligned to min(size, 4K), so the buffer starts at that alignment and
661 * ends at hwb->size alignment. If mbuf inlining is allowed the
662 * starting alignment will be reduced to MSIZE and the driver will
663 * exercise appropriate caution when deciding on the best buffer layout
666 n = 0; /* no usable buffer size to begin with */
667 swz = &s->sw_zone_info[0];
669 for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
670 int8_t head = -1, tail = -1;
672 swz->size = sw_buf_sizes[i];
673 swz->zone = m_getzone(swz->size);
674 swz->type = m_gettype(swz->size);
676 if (swz->size < PAGE_SIZE) {
677 MPASS(powerof2(swz->size));
678 if (fl_pad && (swz->size % sp->pad_boundary != 0))
682 if (swz->size == safest_rx_cluster)
685 hwb = &s->hw_buf_info[0];
686 for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
687 if (hwb->zidx != -1 || hwb->size > swz->size)
691 MPASS(hwb->size % sp->pad_boundary == 0);
696 else if (hwb->size < s->hw_buf_info[tail].size) {
697 s->hw_buf_info[tail].next = j;
701 struct hw_buf_info *t;
703 for (cur = &head; *cur != -1; cur = &t->next) {
704 t = &s->hw_buf_info[*cur];
705 if (hwb->size == t->size) {
709 if (hwb->size > t->size) {
717 swz->head_hwidx = head;
718 swz->tail_hwidx = tail;
722 if (swz->size - s->hw_buf_info[tail].size >=
724 sc->flags |= BUF_PACKING_OK;
728 device_printf(sc->dev, "no usable SGE FL buffer size.\n");
734 if (safe_swz != NULL) {
735 s->safe_hwidx1 = safe_swz->head_hwidx;
736 for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
739 hwb = &s->hw_buf_info[i];
742 MPASS(hwb->size % sp->pad_boundary == 0);
744 spare = safe_swz->size - hwb->size;
745 if (spare >= CL_METADATA_SIZE) {
752 v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
753 r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
755 device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
759 m = v = F_TDDPTAGTCB;
760 r = t4_read_reg(sc, A_ULP_RX_CTL);
762 device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
766 m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
768 v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
769 r = t4_read_reg(sc, A_TP_PARA_REG5);
771 device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
775 t4_init_tp_params(sc);
777 t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
778 t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
784 t4_create_dma_tag(struct adapter *sc)
788 rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
789 BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
790 BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
793 device_printf(sc->dev,
794 "failed to create main DMA tag: %d\n", rc);
801 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
802 struct sysctl_oid_list *children)
804 struct sge_params *sp = &sc->params.sge;
806 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
807 CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
808 "freelist buffer sizes");
810 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
811 NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
813 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
814 NULL, sp->pad_boundary, "payload pad boundary (bytes)");
816 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
817 NULL, sp->spg_len, "status page size (bytes)");
819 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
820 NULL, cong_drop, "congestion drop setting");
822 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
823 NULL, sp->pack_boundary, "payload pack boundary (bytes)");
827 t4_destroy_dma_tag(struct adapter *sc)
830 bus_dma_tag_destroy(sc->dmat);
836 * Allocate and initialize the firmware event queue and the management queue.
838 * Returns errno on failure. Resources allocated up to that point may still be
839 * allocated. Caller is responsible for cleanup in case this function fails.
842 t4_setup_adapter_queues(struct adapter *sc)
846 ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
848 sysctl_ctx_init(&sc->ctx);
849 sc->flags |= ADAP_SYSCTL_CTX;
852 * Firmware event queue
859 * Management queue. This is just a control queue that uses the fwq as
862 rc = alloc_mgmtq(sc);
871 t4_teardown_adapter_queues(struct adapter *sc)
874 ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
876 /* Do this before freeing the queue */
877 if (sc->flags & ADAP_SYSCTL_CTX) {
878 sysctl_ctx_free(&sc->ctx);
879 sc->flags &= ~ADAP_SYSCTL_CTX;
889 first_vector(struct vi_info *vi)
891 struct adapter *sc = vi->pi->adapter;
893 if (sc->intr_count == 1)
896 return (vi->first_intr);
900 * Given an arbitrary "index," come up with an iq that can be used by other
901 * queues (of this VI) for interrupt forwarding, SGE egress updates, etc.
902 * The iq returned is guaranteed to be something that takes direct interrupts.
904 static struct sge_iq *
905 vi_intr_iq(struct vi_info *vi, int idx)
907 struct adapter *sc = vi->pi->adapter;
908 struct sge *s = &sc->sge;
909 struct sge_iq *iq = NULL;
912 if (sc->intr_count == 1)
913 return (&sc->sge.fwq);
917 ("%s: vi %p has no exclusive interrupts, total interrupts = %d",
918 __func__, vi, sc->intr_count));
921 if (vi->flags & INTR_RXQ) {
923 iq = &s->rxq[vi->first_rxq + i].iq;
929 if (vi->flags & INTR_OFLD_RXQ) {
930 if (i < vi->nofldrxq) {
931 iq = &s->ofld_rxq[vi->first_ofld_rxq + i].iq;
937 panic("%s: vi %p, intr_flags 0x%lx, idx %d, total intr %d\n", __func__,
938 vi, vi->flags & INTR_ALL, idx, nintr);
941 KASSERT(iq->flags & IQ_INTR,
942 ("%s: iq %p (vi %p, intr_flags 0x%lx, idx %d)", __func__, iq, vi,
943 vi->flags & INTR_ALL, idx));
947 /* Maximum payload that can be delivered with a single iq descriptor */
949 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
955 payload = sc->tt.rx_coalesce ?
956 G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)) : mtu;
959 /* large enough even when hw VLAN extraction is disabled */
960 payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
961 ETHER_VLAN_ENCAP_LEN + mtu;
970 t4_setup_vi_queues(struct vi_info *vi)
972 int rc = 0, i, j, intr_idx, iqid;
975 struct sge_wrq *ctrlq;
977 struct sge_ofld_rxq *ofld_rxq;
978 struct sge_wrq *ofld_txq;
982 struct sge_nm_rxq *nm_rxq;
983 struct sge_nm_txq *nm_txq;
986 struct port_info *pi = vi->pi;
987 struct adapter *sc = pi->adapter;
988 struct ifnet *ifp = vi->ifp;
989 struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
990 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
991 int maxp, mtu = ifp->if_mtu;
993 /* Interrupt vector to start from (when using multiple vectors) */
994 intr_idx = first_vector(vi);
997 saved_idx = intr_idx;
998 if (ifp->if_capabilities & IFCAP_NETMAP) {
1000 /* netmap is supported with direct interrupts only. */
1001 MPASS(vi->flags & INTR_RXQ);
1004 * We don't have buffers to back the netmap rx queues
1005 * right now so we create the queues in a way that
1006 * doesn't set off any congestion signal in the chip.
1008 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1009 CTLFLAG_RD, NULL, "rx queues");
1010 for_each_nm_rxq(vi, i, nm_rxq) {
1011 rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1017 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1018 CTLFLAG_RD, NULL, "tx queues");
1019 for_each_nm_txq(vi, i, nm_txq) {
1020 iqid = vi->first_nm_rxq + (i % vi->nnmrxq);
1021 rc = alloc_nm_txq(vi, nm_txq, iqid, i, oid);
1027 /* Normal rx queues and netmap rx queues share the same interrupts. */
1028 intr_idx = saved_idx;
1032 * First pass over all NIC and TOE rx queues:
1033 * a) initialize iq and fl
1034 * b) allocate queue iff it will take direct interrupts.
1036 maxp = mtu_to_max_payload(sc, mtu, 0);
1037 if (vi->flags & INTR_RXQ) {
1038 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1039 CTLFLAG_RD, NULL, "rx queues");
1041 for_each_rxq(vi, i, rxq) {
1043 init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1045 snprintf(name, sizeof(name), "%s rxq%d-fl",
1046 device_get_nameunit(vi->dev), i);
1047 init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1049 if (vi->flags & INTR_RXQ) {
1050 rxq->iq.flags |= IQ_INTR;
1051 rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
1058 if (ifp->if_capabilities & IFCAP_NETMAP)
1059 intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1062 maxp = mtu_to_max_payload(sc, mtu, 1);
1063 if (vi->flags & INTR_OFLD_RXQ) {
1064 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1066 "rx queues for offloaded TCP connections");
1068 for_each_ofld_rxq(vi, i, ofld_rxq) {
1070 init_iq(&ofld_rxq->iq, sc, vi->tmr_idx, vi->pktc_idx,
1073 snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1074 device_get_nameunit(vi->dev), i);
1075 init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1077 if (vi->flags & INTR_OFLD_RXQ) {
1078 ofld_rxq->iq.flags |= IQ_INTR;
1079 rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1088 * Second pass over all NIC and TOE rx queues. The queues forwarding
1089 * their interrupts are allocated now.
1092 if (!(vi->flags & INTR_RXQ)) {
1093 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1094 CTLFLAG_RD, NULL, "rx queues");
1095 for_each_rxq(vi, i, rxq) {
1096 MPASS(!(rxq->iq.flags & IQ_INTR));
1098 intr_idx = vi_intr_iq(vi, j)->abs_id;
1100 rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
1107 if (vi->nofldrxq != 0 && !(vi->flags & INTR_OFLD_RXQ)) {
1108 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1110 "rx queues for offloaded TCP connections");
1111 for_each_ofld_rxq(vi, i, ofld_rxq) {
1112 MPASS(!(ofld_rxq->iq.flags & IQ_INTR));
1114 intr_idx = vi_intr_iq(vi, j)->abs_id;
1116 rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1125 * Now the tx queues. Only one pass needed.
1127 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1130 for_each_txq(vi, i, txq) {
1131 iqid = vi_intr_iq(vi, j)->cntxt_id;
1132 snprintf(name, sizeof(name), "%s txq%d",
1133 device_get_nameunit(vi->dev), i);
1134 init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan, iqid,
1137 rc = alloc_txq(vi, txq, i, oid);
1143 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1144 CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
1145 for_each_ofld_txq(vi, i, ofld_txq) {
1146 struct sysctl_oid *oid2;
1148 iqid = vi_intr_iq(vi, j)->cntxt_id;
1149 snprintf(name, sizeof(name), "%s ofld_txq%d",
1150 device_get_nameunit(vi->dev), i);
1151 init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
1154 snprintf(name, sizeof(name), "%d", i);
1155 oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1156 name, CTLFLAG_RD, NULL, "offload tx queue");
1158 rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1166 * Finally, the control queue.
1168 if (!IS_MAIN_VI(vi))
1170 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
1171 NULL, "ctrl queue");
1172 ctrlq = &sc->sge.ctrlq[pi->port_id];
1173 iqid = vi_intr_iq(vi, 0)->cntxt_id;
1174 snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(vi->dev));
1175 init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid,
1177 rc = alloc_wrq(sc, vi, ctrlq, oid);
1181 t4_teardown_vi_queues(vi);
1190 t4_teardown_vi_queues(struct vi_info *vi)
1193 struct port_info *pi = vi->pi;
1194 struct adapter *sc = pi->adapter;
1195 struct sge_rxq *rxq;
1196 struct sge_txq *txq;
1198 struct sge_ofld_rxq *ofld_rxq;
1199 struct sge_wrq *ofld_txq;
1202 struct sge_nm_rxq *nm_rxq;
1203 struct sge_nm_txq *nm_txq;
1206 /* Do this before freeing the queues */
1207 if (vi->flags & VI_SYSCTL_CTX) {
1208 sysctl_ctx_free(&vi->ctx);
1209 vi->flags &= ~VI_SYSCTL_CTX;
1213 if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1214 for_each_nm_txq(vi, i, nm_txq) {
1215 free_nm_txq(vi, nm_txq);
1218 for_each_nm_rxq(vi, i, nm_rxq) {
1219 free_nm_rxq(vi, nm_rxq);
1225 * Take down all the tx queues first, as they reference the rx queues
1226 * (for egress updates, etc.).
1230 free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
1232 for_each_txq(vi, i, txq) {
1236 for_each_ofld_txq(vi, i, ofld_txq) {
1237 free_wrq(sc, ofld_txq);
1242 * Then take down the rx queues that forward their interrupts, as they
1243 * reference other rx queues.
1246 for_each_rxq(vi, i, rxq) {
1247 if ((rxq->iq.flags & IQ_INTR) == 0)
1251 for_each_ofld_rxq(vi, i, ofld_rxq) {
1252 if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
1253 free_ofld_rxq(vi, ofld_rxq);
1258 * Then take down the rx queues that take direct interrupts.
1261 for_each_rxq(vi, i, rxq) {
1262 if (rxq->iq.flags & IQ_INTR)
1266 for_each_ofld_rxq(vi, i, ofld_rxq) {
1267 if (ofld_rxq->iq.flags & IQ_INTR)
1268 free_ofld_rxq(vi, ofld_rxq);
1276 * Deals with errors and the firmware event queue. All data rx queues forward
1277 * their interrupt to the firmware event queue.
1280 t4_intr_all(void *arg)
1282 struct adapter *sc = arg;
1283 struct sge_iq *fwq = &sc->sge.fwq;
1286 if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
1288 atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
1292 /* Deals with error interrupts */
1294 t4_intr_err(void *arg)
1296 struct adapter *sc = arg;
1298 t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1299 t4_slow_intr_handler(sc);
1303 t4_intr_evt(void *arg)
1305 struct sge_iq *iq = arg;
1307 if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1309 atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1316 struct sge_iq *iq = arg;
1318 if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1320 atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1325 t4_vi_intr(void *arg)
1327 struct irq *irq = arg;
1330 if (atomic_cmpset_int(&irq->nm_state, NM_ON, NM_BUSY)) {
1331 t4_nm_intr(irq->nm_rxq);
1332 atomic_cmpset_int(&irq->nm_state, NM_BUSY, NM_ON);
1335 if (irq->rxq != NULL)
1340 * Deals with anything and everything on the given ingress queue.
1343 service_iq(struct sge_iq *iq, int budget)
1346 struct sge_rxq *rxq = iq_to_rxq(iq); /* Use iff iq is part of rxq */
1347 struct sge_fl *fl; /* Use iff IQ_HAS_FL */
1348 struct adapter *sc = iq->adapter;
1349 struct iq_desc *d = &iq->desc[iq->cidx];
1350 int ndescs = 0, limit;
1351 int rsp_type, refill;
1353 uint16_t fl_hw_cidx;
1355 STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1356 #if defined(INET) || defined(INET6)
1357 const struct timeval lro_timeout = {0, sc->lro_timeout};
1360 KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1362 limit = budget ? budget : iq->qsize / 16;
1364 if (iq->flags & IQ_HAS_FL) {
1366 fl_hw_cidx = fl->hw_cidx; /* stable snapshot */
1369 fl_hw_cidx = 0; /* to silence gcc warning */
1373 * We always come back and check the descriptor ring for new indirect
1374 * interrupts and other responses after running a single handler.
1377 while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1383 rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1384 lq = be32toh(d->rsp.pldbuflen_qid);
1387 case X_RSPD_TYPE_FLBUF:
1389 KASSERT(iq->flags & IQ_HAS_FL,
1390 ("%s: data for an iq (%p) with no freelist",
1393 m0 = get_fl_payload(sc, fl, lq);
1394 if (__predict_false(m0 == NULL))
1396 refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1397 #ifdef T4_PKT_TIMESTAMP
1399 * 60 bit timestamp for the payload is
1400 * *(uint64_t *)m0->m_pktdat. Note that it is
1401 * in the leading free-space in the mbuf. The
1402 * kernel can clobber it during a pullup,
1403 * m_copymdata, etc. You need to make sure that
1404 * the mbuf reaches you unmolested if you care
1405 * about the timestamp.
1407 *(uint64_t *)m0->m_pktdat =
1408 be64toh(ctrl->u.last_flit) &
1414 case X_RSPD_TYPE_CPL:
1415 KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1416 ("%s: bad opcode %02x.", __func__,
1418 t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1421 case X_RSPD_TYPE_INTR:
1424 * Interrupts should be forwarded only to queues
1425 * that are not forwarding their interrupts.
1426 * This means service_iq can recurse but only 1
1429 KASSERT(budget == 0,
1430 ("%s: budget %u, rsp_type %u", __func__,
1434 * There are 1K interrupt-capable queues (qids 0
1435 * through 1023). A response type indicating a
1436 * forwarded interrupt with a qid >= 1K is an
1437 * iWARP async notification.
1440 t4_an_handler(iq, &d->rsp);
1444 q = sc->sge.iqmap[lq - sc->sge.iq_start];
1445 if (atomic_cmpset_int(&q->state, IQS_IDLE,
1447 if (service_iq(q, q->qsize / 16) == 0) {
1448 atomic_cmpset_int(&q->state,
1449 IQS_BUSY, IQS_IDLE);
1451 STAILQ_INSERT_TAIL(&iql, q,
1459 ("%s: illegal response type %d on iq %p",
1460 __func__, rsp_type, iq));
1462 "%s: illegal response type %d on iq %p",
1463 device_get_nameunit(sc->dev), rsp_type, iq);
1468 if (__predict_false(++iq->cidx == iq->sidx)) {
1470 iq->gen ^= F_RSPD_GEN;
1473 if (__predict_false(++ndescs == limit)) {
1474 t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
1476 V_INGRESSQID(iq->cntxt_id) |
1477 V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1480 #if defined(INET) || defined(INET6)
1481 if (iq->flags & IQ_LRO_ENABLED &&
1482 sc->lro_timeout != 0) {
1483 tcp_lro_flush_inactive(&rxq->lro,
1489 if (iq->flags & IQ_HAS_FL) {
1491 refill_fl(sc, fl, 32);
1494 return (EINPROGRESS);
1499 refill_fl(sc, fl, 32);
1501 fl_hw_cidx = fl->hw_cidx;
1506 if (STAILQ_EMPTY(&iql))
1510 * Process the head only, and send it to the back of the list if
1511 * it's still not done.
1513 q = STAILQ_FIRST(&iql);
1514 STAILQ_REMOVE_HEAD(&iql, link);
1515 if (service_iq(q, q->qsize / 8) == 0)
1516 atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1518 STAILQ_INSERT_TAIL(&iql, q, link);
1521 #if defined(INET) || defined(INET6)
1522 if (iq->flags & IQ_LRO_ENABLED) {
1523 struct lro_ctrl *lro = &rxq->lro;
1524 struct lro_entry *l;
1526 while (!SLIST_EMPTY(&lro->lro_active)) {
1527 l = SLIST_FIRST(&lro->lro_active);
1528 SLIST_REMOVE_HEAD(&lro->lro_active, next);
1529 tcp_lro_flush(lro, l);
1534 t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
1535 V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1537 if (iq->flags & IQ_HAS_FL) {
1541 starved = refill_fl(sc, fl, 64);
1543 if (__predict_false(starved != 0))
1544 add_fl_to_sfl(sc, fl);
1551 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1553 int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1556 MPASS(cll->region3 >= CL_METADATA_SIZE);
1561 static inline struct cluster_metadata *
1562 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1566 if (cl_has_metadata(fl, cll)) {
1567 struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1569 return ((struct cluster_metadata *)(cl + swz->size) - 1);
1575 rxb_free(struct mbuf *m, void *arg1, void *arg2)
1577 uma_zone_t zone = arg1;
1580 uma_zfree(zone, cl);
1581 counter_u64_add(extfree_rels, 1);
1583 return (EXT_FREE_OK);
1587 * The mbuf returned by this function could be allocated from zone_mbuf or
1588 * constructed in spare room in the cluster.
1590 * The mbuf carries the payload in one of these ways
1591 * a) frame inside the mbuf (mbuf from zone_mbuf)
1592 * b) m_cljset (for clusters without metadata) zone_mbuf
1593 * c) m_extaddref (cluster with metadata) inline mbuf
1594 * d) m_extaddref (cluster with metadata) zone_mbuf
1596 static struct mbuf *
1597 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1601 struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1602 struct cluster_layout *cll = &sd->cll;
1603 struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1604 struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1605 struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1609 blen = hwb->size - fl->rx_offset; /* max possible in this buf */
1610 len = min(remaining, blen);
1611 payload = sd->cl + cll->region1 + fl->rx_offset;
1612 if (fl->flags & FL_BUF_PACKING) {
1613 const u_int l = fr_offset + len;
1614 const u_int pad = roundup2(l, fl->buf_boundary) - l;
1616 if (fl->rx_offset + len + pad < hwb->size)
1618 MPASS(fl->rx_offset + blen <= hwb->size);
1620 MPASS(fl->rx_offset == 0); /* not packing */
1624 if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1627 * Copy payload into a freshly allocated mbuf.
1630 m = fr_offset == 0 ?
1631 m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1634 fl->mbuf_allocated++;
1635 #ifdef T4_PKT_TIMESTAMP
1636 /* Leave room for a timestamp */
1639 /* copy data to mbuf */
1640 bcopy(payload, mtod(m, caddr_t), len);
1642 } else if (sd->nmbuf * MSIZE < cll->region1) {
1645 * There's spare room in the cluster for an mbuf. Create one
1646 * and associate it with the payload that's in the cluster.
1650 m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1651 /* No bzero required */
1652 if (m_init(m, NULL, 0, M_NOWAIT, MT_DATA,
1653 fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1656 m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1658 if (sd->nmbuf++ == 0)
1659 counter_u64_add(extfree_refs, 1);
1664 * Grab an mbuf from zone_mbuf and associate it with the
1665 * payload in the cluster.
1668 m = fr_offset == 0 ?
1669 m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1672 fl->mbuf_allocated++;
1674 m_extaddref(m, payload, blen, &clm->refcount,
1675 rxb_free, swz->zone, sd->cl);
1676 if (sd->nmbuf++ == 0)
1677 counter_u64_add(extfree_refs, 1);
1679 m_cljset(m, sd->cl, swz->type);
1680 sd->cl = NULL; /* consumed, not a recycle candidate */
1684 m->m_pkthdr.len = remaining;
1687 if (fl->flags & FL_BUF_PACKING) {
1688 fl->rx_offset += blen;
1689 MPASS(fl->rx_offset <= hwb->size);
1690 if (fl->rx_offset < hwb->size)
1691 return (m); /* without advancing the cidx */
1694 if (__predict_false(++fl->cidx % 8 == 0)) {
1695 uint16_t cidx = fl->cidx / 8;
1697 if (__predict_false(cidx == fl->sidx))
1698 fl->cidx = cidx = 0;
1706 static struct mbuf *
1707 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1709 struct mbuf *m0, *m, **pnext;
1711 const u_int total = G_RSPD_LEN(len_newbuf);
1713 if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1714 M_ASSERTPKTHDR(fl->m0);
1715 MPASS(fl->m0->m_pkthdr.len == total);
1716 MPASS(fl->remaining < total);
1720 remaining = fl->remaining;
1721 fl->flags &= ~FL_BUF_RESUME;
1725 if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1727 if (__predict_false(++fl->cidx % 8 == 0)) {
1728 uint16_t cidx = fl->cidx / 8;
1730 if (__predict_false(cidx == fl->sidx))
1731 fl->cidx = cidx = 0;
1737 * Payload starts at rx_offset in the current hw buffer. Its length is
1738 * 'len' and it may span multiple hw buffers.
1741 m0 = get_scatter_segment(sc, fl, 0, total);
1744 remaining = total - m0->m_len;
1745 pnext = &m0->m_next;
1746 while (remaining > 0) {
1748 MPASS(fl->rx_offset == 0);
1749 m = get_scatter_segment(sc, fl, total - remaining, remaining);
1750 if (__predict_false(m == NULL)) {
1753 fl->remaining = remaining;
1754 fl->flags |= FL_BUF_RESUME;
1759 remaining -= m->m_len;
1768 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1770 struct sge_rxq *rxq = iq_to_rxq(iq);
1771 struct ifnet *ifp = rxq->ifp;
1772 struct adapter *sc = iq->adapter;
1773 const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1774 #if defined(INET) || defined(INET6)
1775 struct lro_ctrl *lro = &rxq->lro;
1778 KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1781 m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
1782 m0->m_len -= sc->params.sge.fl_pktshift;
1783 m0->m_data += sc->params.sge.fl_pktshift;
1785 m0->m_pkthdr.rcvif = ifp;
1786 M_HASHTYPE_SET(m0, M_HASHTYPE_OPAQUE);
1787 m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1789 if (cpl->csum_calc && !cpl->err_vec) {
1790 if (ifp->if_capenable & IFCAP_RXCSUM &&
1791 cpl->l2info & htobe32(F_RXF_IP)) {
1792 m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1793 CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1795 } else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1796 cpl->l2info & htobe32(F_RXF_IP6)) {
1797 m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1802 if (__predict_false(cpl->ip_frag))
1803 m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1805 m0->m_pkthdr.csum_data = 0xffff;
1809 m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1810 m0->m_flags |= M_VLANTAG;
1811 rxq->vlan_extraction++;
1814 #if defined(INET) || defined(INET6)
1815 if (cpl->l2info & htobe32(F_RXF_LRO) &&
1816 iq->flags & IQ_LRO_ENABLED &&
1817 tcp_lro_rx(lro, m0, 0) == 0) {
1818 /* queued for LRO */
1821 ifp->if_input(ifp, m0);
1827 * Must drain the wrq or make sure that someone else will.
1830 wrq_tx_drain(void *arg, int n)
1832 struct sge_wrq *wrq = arg;
1833 struct sge_eq *eq = &wrq->eq;
1836 if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
1837 drain_wrq_wr_list(wrq->adapter, wrq);
1842 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
1844 struct sge_eq *eq = &wrq->eq;
1845 u_int available, dbdiff; /* # of hardware descriptors */
1848 struct fw_eth_tx_pkt_wr *dst; /* any fw WR struct will do */
1850 EQ_LOCK_ASSERT_OWNED(eq);
1851 MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
1852 wr = STAILQ_FIRST(&wrq->wr_list);
1853 MPASS(wr != NULL); /* Must be called with something useful to do */
1854 MPASS(eq->pidx == eq->dbidx);
1858 eq->cidx = read_hw_cidx(eq);
1859 if (eq->pidx == eq->cidx)
1860 available = eq->sidx - 1;
1862 available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
1864 MPASS(wr->wrq == wrq);
1865 n = howmany(wr->wr_len, EQ_ESIZE);
1869 dst = (void *)&eq->desc[eq->pidx];
1870 if (__predict_true(eq->sidx - eq->pidx > n)) {
1871 /* Won't wrap, won't end exactly at the status page. */
1872 bcopy(&wr->wr[0], dst, wr->wr_len);
1875 int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
1877 bcopy(&wr->wr[0], dst, first_portion);
1878 if (wr->wr_len > first_portion) {
1879 bcopy(&wr->wr[first_portion], &eq->desc[0],
1880 wr->wr_len - first_portion);
1882 eq->pidx = n - (eq->sidx - eq->pidx);
1885 if (available < eq->sidx / 4 &&
1886 atomic_cmpset_int(&eq->equiq, 0, 1)) {
1887 dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
1889 eq->equeqidx = eq->pidx;
1890 } else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
1891 dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1892 eq->equeqidx = eq->pidx;
1897 ring_eq_db(sc, eq, dbdiff);
1901 STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1903 MPASS(wrq->nwr_pending > 0);
1905 MPASS(wrq->ndesc_needed >= n);
1906 wrq->ndesc_needed -= n;
1907 } while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
1910 ring_eq_db(sc, eq, dbdiff);
1914 * Doesn't fail. Holds on to work requests it can't send right away.
1917 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
1920 struct sge_eq *eq = &wrq->eq;
1923 EQ_LOCK_ASSERT_OWNED(eq);
1925 MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
1926 MPASS((wr->wr_len & 0x7) == 0);
1928 STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
1930 wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
1932 if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
1933 return; /* commit_wrq_wr will drain wr_list as well. */
1935 drain_wrq_wr_list(sc, wrq);
1937 /* Doorbell must have caught up to the pidx. */
1938 MPASS(eq->pidx == eq->dbidx);
1942 t4_update_fl_bufsize(struct ifnet *ifp)
1944 struct vi_info *vi = ifp->if_softc;
1945 struct adapter *sc = vi->pi->adapter;
1946 struct sge_rxq *rxq;
1948 struct sge_ofld_rxq *ofld_rxq;
1951 int i, maxp, mtu = ifp->if_mtu;
1953 maxp = mtu_to_max_payload(sc, mtu, 0);
1954 for_each_rxq(vi, i, rxq) {
1958 find_best_refill_source(sc, fl, maxp);
1962 maxp = mtu_to_max_payload(sc, mtu, 1);
1963 for_each_ofld_rxq(vi, i, ofld_rxq) {
1967 find_best_refill_source(sc, fl, maxp);
1974 mbuf_nsegs(struct mbuf *m)
1978 KASSERT(m->m_pkthdr.l5hlen > 0,
1979 ("%s: mbuf %p missing information on # of segments.", __func__, m));
1981 return (m->m_pkthdr.l5hlen);
1985 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
1989 m->m_pkthdr.l5hlen = nsegs;
1993 mbuf_len16(struct mbuf *m)
1998 n = m->m_pkthdr.PH_loc.eigth[0];
1999 MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2005 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2009 m->m_pkthdr.PH_loc.eigth[0] = len16;
2013 needs_tso(struct mbuf *m)
2018 if (m->m_pkthdr.csum_flags & CSUM_TSO) {
2019 KASSERT(m->m_pkthdr.tso_segsz > 0,
2020 ("%s: TSO requested in mbuf %p but MSS not provided",
2029 needs_l3_csum(struct mbuf *m)
2034 if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO))
2040 needs_l4_csum(struct mbuf *m)
2045 if (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2046 CSUM_TCP_IPV6 | CSUM_TSO))
2052 needs_vlan_insertion(struct mbuf *m)
2057 if (m->m_flags & M_VLANTAG) {
2058 KASSERT(m->m_pkthdr.ether_vtag != 0,
2059 ("%s: HWVLAN requested in mbuf %p but tag not provided",
2067 m_advance(struct mbuf **pm, int *poffset, int len)
2069 struct mbuf *m = *pm;
2070 int offset = *poffset;
2076 if (offset + len < m->m_len) {
2078 p = mtod(m, uintptr_t) + offset;
2081 len -= m->m_len - offset;
2092 same_paddr(char *a, char *b)
2097 else if (a != NULL && b != NULL) {
2098 vm_offset_t x = (vm_offset_t)a;
2099 vm_offset_t y = (vm_offset_t)b;
2101 if ((x & PAGE_MASK) == (y & PAGE_MASK) &&
2102 pmap_kextract(x) == pmap_kextract(y))
2110 * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2111 * must have at least one mbuf that's not empty.
2114 count_mbuf_nsegs(struct mbuf *m)
2116 char *prev_end, *start;
2123 for (; m; m = m->m_next) {
2126 if (__predict_false(len == 0))
2128 start = mtod(m, char *);
2130 nsegs += sglist_count(start, len);
2131 if (same_paddr(prev_end, start))
2133 prev_end = start + len;
2141 * Analyze the mbuf to determine its tx needs. The mbuf passed in may change:
2142 * a) caller can assume it's been freed if this function returns with an error.
2143 * b) it may get defragged up if the gather list is too long for the hardware.
2146 parse_pkt(struct mbuf **mp)
2148 struct mbuf *m0 = *mp, *m;
2149 int rc, nsegs, defragged = 0, offset;
2150 struct ether_header *eh;
2152 #if defined(INET) || defined(INET6)
2158 if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2167 * First count the number of gather list segments in the payload.
2168 * Defrag the mbuf if nsegs exceeds the hardware limit.
2171 MPASS(m0->m_pkthdr.len > 0);
2172 nsegs = count_mbuf_nsegs(m0);
2173 if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2174 if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2178 *mp = m0 = m; /* update caller's copy after defrag */
2182 if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2183 m0 = m_pullup(m0, m0->m_pkthdr.len);
2185 /* Should have left well enough alone. */
2189 *mp = m0; /* update caller's copy after pullup */
2192 set_mbuf_nsegs(m0, nsegs);
2193 set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2199 eh = mtod(m, struct ether_header *);
2200 eh_type = ntohs(eh->ether_type);
2201 if (eh_type == ETHERTYPE_VLAN) {
2202 struct ether_vlan_header *evh = (void *)eh;
2204 eh_type = ntohs(evh->evl_proto);
2205 m0->m_pkthdr.l2hlen = sizeof(*evh);
2207 m0->m_pkthdr.l2hlen = sizeof(*eh);
2210 l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2214 case ETHERTYPE_IPV6:
2216 struct ip6_hdr *ip6 = l3hdr;
2218 MPASS(ip6->ip6_nxt == IPPROTO_TCP);
2220 m0->m_pkthdr.l3hlen = sizeof(*ip6);
2227 struct ip *ip = l3hdr;
2229 m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2234 panic("%s: ethertype 0x%04x unknown. if_cxgbe must be compiled"
2235 " with the same INET/INET6 options as the kernel.",
2239 #if defined(INET) || defined(INET6)
2240 tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2241 m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2248 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2250 struct sge_eq *eq = &wrq->eq;
2251 struct adapter *sc = wrq->adapter;
2252 int ndesc, available;
2257 ndesc = howmany(len16, EQ_ESIZE / 16);
2258 MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2262 if (!STAILQ_EMPTY(&wrq->wr_list))
2263 drain_wrq_wr_list(sc, wrq);
2265 if (!STAILQ_EMPTY(&wrq->wr_list)) {
2268 wr = alloc_wrqe(len16 * 16, wrq);
2269 if (__predict_false(wr == NULL))
2272 cookie->ndesc = ndesc;
2276 eq->cidx = read_hw_cidx(eq);
2277 if (eq->pidx == eq->cidx)
2278 available = eq->sidx - 1;
2280 available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2281 if (available < ndesc)
2284 cookie->pidx = eq->pidx;
2285 cookie->ndesc = ndesc;
2286 TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2288 w = &eq->desc[eq->pidx];
2289 IDXINCR(eq->pidx, ndesc, eq->sidx);
2290 if (__predict_false(eq->pidx < ndesc - 1)) {
2292 wrq->ss_pidx = cookie->pidx;
2293 wrq->ss_len = len16 * 16;
2302 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2304 struct sge_eq *eq = &wrq->eq;
2305 struct adapter *sc = wrq->adapter;
2307 struct wrq_cookie *prev, *next;
2309 if (cookie->pidx == -1) {
2310 struct wrqe *wr = __containerof(w, struct wrqe, wr);
2316 ndesc = cookie->ndesc; /* Can be more than SGE_MAX_WR_NDESC here. */
2317 pidx = cookie->pidx;
2318 MPASS(pidx >= 0 && pidx < eq->sidx);
2319 if (__predict_false(w == &wrq->ss[0])) {
2320 int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2322 MPASS(wrq->ss_len > n); /* WR had better wrap around. */
2323 bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2324 bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2327 wrq->tx_wrs_direct++;
2330 prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2331 next = TAILQ_NEXT(cookie, link);
2333 MPASS(pidx == eq->dbidx);
2334 if (next == NULL || ndesc >= 16)
2335 ring_eq_db(wrq->adapter, eq, ndesc);
2337 MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2339 next->ndesc += ndesc;
2342 MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2343 prev->ndesc += ndesc;
2345 TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2347 if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2348 drain_wrq_wr_list(sc, wrq);
2351 if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2352 /* Doorbell must have caught up to the pidx. */
2353 MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2360 can_resume_eth_tx(struct mp_ring *r)
2362 struct sge_eq *eq = r->cookie;
2364 return (total_available_tx_desc(eq) > eq->sidx / 8);
2368 cannot_use_txpkts(struct mbuf *m)
2370 /* maybe put a GL limit too, to avoid silliness? */
2372 return (needs_tso(m));
2376 * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2377 * be consumed. Return the actual number consumed. 0 indicates a stall.
2380 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2382 struct sge_txq *txq = r->cookie;
2383 struct sge_eq *eq = &txq->eq;
2384 struct ifnet *ifp = txq->ifp;
2385 struct vi_info *vi = ifp->if_softc;
2386 struct port_info *pi = vi->pi;
2387 struct adapter *sc = pi->adapter;
2388 u_int total, remaining; /* # of packets */
2389 u_int available, dbdiff; /* # of hardware descriptors */
2391 struct mbuf *m0, *tail;
2393 struct fw_eth_tx_pkts_wr *wr; /* any fw WR struct will do */
2395 remaining = IDXDIFF(pidx, cidx, r->size);
2396 MPASS(remaining > 0); /* Must not be called without work to do. */
2400 if (__predict_false((eq->flags & EQ_ENABLED) == 0)) {
2401 while (cidx != pidx) {
2402 m0 = r->items[cidx];
2404 if (++cidx == r->size)
2407 reclaim_tx_descs(txq, 2048);
2412 /* How many hardware descriptors do we have readily available. */
2413 if (eq->pidx == eq->cidx)
2414 available = eq->sidx - 1;
2416 available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2417 dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2419 while (remaining > 0) {
2421 m0 = r->items[cidx];
2423 MPASS(m0->m_nextpkt == NULL);
2425 if (available < SGE_MAX_WR_NDESC) {
2426 available += reclaim_tx_descs(txq, 64);
2427 if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2428 break; /* out of descriptors */
2431 next_cidx = cidx + 1;
2432 if (__predict_false(next_cidx == r->size))
2435 wr = (void *)&eq->desc[eq->pidx];
2436 if (remaining > 1 &&
2437 try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2439 /* pkts at cidx, next_cidx should both be in txp. */
2440 MPASS(txp.npkt == 2);
2441 tail = r->items[next_cidx];
2442 MPASS(tail->m_nextpkt == NULL);
2443 ETHER_BPF_MTAP(ifp, m0);
2444 ETHER_BPF_MTAP(ifp, tail);
2445 m0->m_nextpkt = tail;
2447 if (__predict_false(++next_cidx == r->size))
2450 while (next_cidx != pidx) {
2451 if (add_to_txpkts(r->items[next_cidx], &txp,
2454 tail->m_nextpkt = r->items[next_cidx];
2455 tail = tail->m_nextpkt;
2456 ETHER_BPF_MTAP(ifp, tail);
2457 if (__predict_false(++next_cidx == r->size))
2461 n = write_txpkts_wr(txq, wr, m0, &txp, available);
2463 remaining -= txp.npkt;
2467 ETHER_BPF_MTAP(ifp, m0);
2468 n = write_txpkt_wr(txq, (void *)wr, m0, available);
2470 MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2474 IDXINCR(eq->pidx, n, eq->sidx);
2476 if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2477 atomic_cmpset_int(&eq->equiq, 0, 1)) {
2478 wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2480 eq->equeqidx = eq->pidx;
2481 } else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2482 wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2483 eq->equeqidx = eq->pidx;
2486 if (dbdiff >= 16 && remaining >= 4) {
2487 ring_eq_db(sc, eq, dbdiff);
2488 available += reclaim_tx_descs(txq, 4 * dbdiff);
2495 ring_eq_db(sc, eq, dbdiff);
2496 reclaim_tx_descs(txq, 32);
2505 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2509 KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2510 ("%s: bad tmr_idx %d", __func__, tmr_idx));
2511 KASSERT(pktc_idx < SGE_NCOUNTERS, /* -ve is ok, means don't use */
2512 ("%s: bad pktc_idx %d", __func__, pktc_idx));
2516 iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2517 iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2518 if (pktc_idx >= 0) {
2519 iq->intr_params |= F_QINTR_CNT_EN;
2520 iq->intr_pktc_idx = pktc_idx;
2522 iq->qsize = roundup2(qsize, 16); /* See FW_IQ_CMD/iqsize */
2523 iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2527 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2531 fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2532 strlcpy(fl->lockname, name, sizeof(fl->lockname));
2533 if (sc->flags & BUF_PACKING_OK &&
2534 ((!is_t4(sc) && buffer_packing) || /* T5+: enabled unless 0 */
2535 (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2536 fl->flags |= FL_BUF_PACKING;
2537 find_best_refill_source(sc, fl, maxp);
2538 find_safe_refill_source(sc, fl);
2542 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2543 uint8_t tx_chan, uint16_t iqid, char *name)
2545 KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2547 eq->flags = eqtype & EQ_TYPEMASK;
2548 eq->tx_chan = tx_chan;
2550 eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2551 strlcpy(eq->lockname, name, sizeof(eq->lockname));
2555 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2556 bus_dmamap_t *map, bus_addr_t *pa, void **va)
2560 rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2561 BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2563 device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2567 rc = bus_dmamem_alloc(*tag, va,
2568 BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2570 device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2574 rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2576 device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2581 free_ring(sc, *tag, *map, *pa, *va);
2587 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2588 bus_addr_t pa, void *va)
2591 bus_dmamap_unload(tag, map);
2593 bus_dmamem_free(tag, va, map);
2595 bus_dma_tag_destroy(tag);
2601 * Allocates the ring for an ingress queue and an optional freelist. If the
2602 * freelist is specified it will be allocated and then associated with the
2605 * Returns errno on failure. Resources allocated up to that point may still be
2606 * allocated. Caller is responsible for cleanup in case this function fails.
2608 * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
2609 * the intr_idx specifies the vector, starting from 0. Otherwise it specifies
2610 * the abs_id of the ingress queue to which its interrupts should be forwarded.
2613 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
2614 int intr_idx, int cong)
2616 int rc, i, cntxt_id;
2619 struct port_info *pi = vi->pi;
2620 struct adapter *sc = iq->adapter;
2621 struct sge_params *sp = &sc->params.sge;
2624 len = iq->qsize * IQ_ESIZE;
2625 rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2626 (void **)&iq->desc);
2630 bzero(&c, sizeof(c));
2631 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2632 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2633 V_FW_IQ_CMD_VFN(0));
2635 c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2638 /* Special handling for firmware event queue */
2639 if (iq == &sc->sge.fwq)
2640 v |= F_FW_IQ_CMD_IQASYNCH;
2642 if (iq->flags & IQ_INTR) {
2643 KASSERT(intr_idx < sc->intr_count,
2644 ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2646 v |= F_FW_IQ_CMD_IQANDST;
2647 v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2649 c.type_to_iqandstindex = htobe32(v |
2650 V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2651 V_FW_IQ_CMD_VIID(vi->viid) |
2652 V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2653 c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2654 F_FW_IQ_CMD_IQGTSMODE |
2655 V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2656 V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
2657 c.iqsize = htobe16(iq->qsize);
2658 c.iqaddr = htobe64(iq->ba);
2660 c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2663 mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
2665 len = fl->qsize * EQ_ESIZE;
2666 rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
2667 &fl->ba, (void **)&fl->desc);
2671 /* Allocate space for one software descriptor per buffer. */
2672 rc = alloc_fl_sdesc(fl);
2674 device_printf(sc->dev,
2675 "failed to setup fl software descriptors: %d\n",
2680 if (fl->flags & FL_BUF_PACKING) {
2681 fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
2682 fl->buf_boundary = sp->pack_boundary;
2684 fl->lowat = roundup2(sp->fl_starve_threshold, 8);
2685 fl->buf_boundary = 16;
2687 if (fl_pad && fl->buf_boundary < sp->pad_boundary)
2688 fl->buf_boundary = sp->pad_boundary;
2690 c.iqns_to_fl0congen |=
2691 htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
2692 F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
2693 (fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
2694 (fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
2697 c.iqns_to_fl0congen |=
2698 htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
2699 F_FW_IQ_CMD_FL0CONGCIF |
2700 F_FW_IQ_CMD_FL0CONGEN);
2702 c.fl0dcaen_to_fl0cidxfthresh =
2703 htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_128B) |
2704 V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
2705 c.fl0size = htobe16(fl->qsize);
2706 c.fl0addr = htobe64(fl->ba);
2709 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2711 device_printf(sc->dev,
2712 "failed to create ingress queue: %d\n", rc);
2717 iq->gen = F_RSPD_GEN;
2718 iq->intr_next = iq->intr_params;
2719 iq->cntxt_id = be16toh(c.iqid);
2720 iq->abs_id = be16toh(c.physiqid);
2721 iq->flags |= IQ_ALLOCATED;
2723 cntxt_id = iq->cntxt_id - sc->sge.iq_start;
2724 if (cntxt_id >= sc->sge.niq) {
2725 panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
2726 cntxt_id, sc->sge.niq - 1);
2728 sc->sge.iqmap[cntxt_id] = iq;
2733 iq->flags |= IQ_HAS_FL;
2734 fl->cntxt_id = be16toh(c.fl0id);
2735 fl->pidx = fl->cidx = 0;
2737 cntxt_id = fl->cntxt_id - sc->sge.eq_start;
2738 if (cntxt_id >= sc->sge.neq) {
2739 panic("%s: fl->cntxt_id (%d) more than the max (%d)",
2740 __func__, cntxt_id, sc->sge.neq - 1);
2742 sc->sge.eqmap[cntxt_id] = (void *)fl;
2745 if (isset(&sc->doorbells, DOORBELL_UDB)) {
2746 uint32_t s_qpp = sc->params.sge.eq_s_qpp;
2747 uint32_t mask = (1 << s_qpp) - 1;
2748 volatile uint8_t *udb;
2750 udb = sc->udbs_base + UDBS_DB_OFFSET;
2751 udb += (qid >> s_qpp) << PAGE_SHIFT;
2753 if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
2754 udb += qid << UDBS_SEG_SHIFT;
2757 fl->udb = (volatile void *)udb;
2759 fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
2762 /* Enough to make sure the SGE doesn't think it's starved */
2763 refill_fl(sc, fl, fl->lowat);
2767 if (is_t5(sc) && cong >= 0) {
2768 uint32_t param, val;
2770 param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
2771 V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
2772 V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
2777 for (i = 0; i < 4; i++) {
2778 if (cong & (1 << i))
2779 val |= 1 << (i << 2);
2783 rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, ¶m, &val);
2785 /* report error but carry on */
2786 device_printf(sc->dev,
2787 "failed to set congestion manager context for "
2788 "ingress queue %d: %d\n", iq->cntxt_id, rc);
2792 /* Enable IQ interrupts */
2793 atomic_store_rel_int(&iq->state, IQS_IDLE);
2794 t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
2795 V_INGRESSQID(iq->cntxt_id));
2801 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
2804 struct adapter *sc = iq->adapter;
2808 return (0); /* nothing to do */
2810 dev = vi ? vi->dev : sc->dev;
2812 if (iq->flags & IQ_ALLOCATED) {
2813 rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
2814 FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
2815 fl ? fl->cntxt_id : 0xffff, 0xffff);
2818 "failed to free queue %p: %d\n", iq, rc);
2821 iq->flags &= ~IQ_ALLOCATED;
2824 free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
2826 bzero(iq, sizeof(*iq));
2829 free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
2833 free_fl_sdesc(sc, fl);
2835 if (mtx_initialized(&fl->fl_lock))
2836 mtx_destroy(&fl->fl_lock);
2838 bzero(fl, sizeof(*fl));
2845 add_fl_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
2848 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2850 oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
2852 children = SYSCTL_CHILDREN(oid);
2854 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2855 CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
2856 "SGE context id of the freelist");
2857 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
2858 fl_pad ? 1 : 0, "padding enabled");
2859 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
2860 fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
2861 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
2862 0, "consumer index");
2863 if (fl->flags & FL_BUF_PACKING) {
2864 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
2865 CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
2867 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
2868 0, "producer index");
2869 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
2870 CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
2871 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
2872 CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
2873 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
2874 CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
2875 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
2876 CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
2877 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
2878 CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
2882 alloc_fwq(struct adapter *sc)
2885 struct sge_iq *fwq = &sc->sge.fwq;
2886 struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2887 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2889 init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
2890 fwq->flags |= IQ_INTR; /* always */
2891 intr_idx = sc->intr_count > 1 ? 1 : 0;
2892 fwq->set_tcb_rpl = t4_filter_rpl;
2893 fwq->l2t_write_rpl = do_l2t_write_rpl;
2894 rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
2896 device_printf(sc->dev,
2897 "failed to create firmware event queue: %d\n", rc);
2901 oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
2902 NULL, "firmware event queue");
2903 children = SYSCTL_CHILDREN(oid);
2905 SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id",
2906 CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I",
2907 "absolute id of the queue");
2908 SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id",
2909 CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I",
2910 "SGE context id of the queue");
2911 SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx",
2912 CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I",
2919 free_fwq(struct adapter *sc)
2921 return free_iq_fl(NULL, &sc->sge.fwq, NULL);
2925 alloc_mgmtq(struct adapter *sc)
2928 struct sge_wrq *mgmtq = &sc->sge.mgmtq;
2930 struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2931 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2933 oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
2934 NULL, "management queue");
2936 snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
2937 init_eq(sc, &mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
2938 sc->sge.fwq.cntxt_id, name);
2939 rc = alloc_wrq(sc, NULL, mgmtq, oid);
2941 device_printf(sc->dev,
2942 "failed to create management queue: %d\n", rc);
2950 free_mgmtq(struct adapter *sc)
2953 return free_wrq(sc, &sc->sge.mgmtq);
2957 tnl_cong(struct port_info *pi, int drop)
2965 return (pi->rx_chan_map);
2969 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
2970 struct sysctl_oid *oid)
2973 struct sysctl_oid_list *children;
2976 rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
2977 tnl_cong(vi->pi, cong_drop));
2982 * The freelist is just barely above the starvation threshold right now,
2983 * fill it up a bit more.
2986 refill_fl(vi->pi->adapter, &rxq->fl, 128);
2987 FL_UNLOCK(&rxq->fl);
2989 #if defined(INET) || defined(INET6)
2990 rc = tcp_lro_init(&rxq->lro);
2993 rxq->lro.ifp = vi->ifp; /* also indicates LRO init'ed */
2995 if (vi->ifp->if_capenable & IFCAP_LRO)
2996 rxq->iq.flags |= IQ_LRO_ENABLED;
3000 children = SYSCTL_CHILDREN(oid);
3002 snprintf(name, sizeof(name), "%d", idx);
3003 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3005 children = SYSCTL_CHILDREN(oid);
3007 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
3008 CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I",
3009 "absolute id of the queue");
3010 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
3011 CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I",
3012 "SGE context id of the queue");
3013 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3014 CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I",
3016 #if defined(INET) || defined(INET6)
3017 SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3018 &rxq->lro.lro_queued, 0, NULL);
3019 SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3020 &rxq->lro.lro_flushed, 0, NULL);
3022 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3023 &rxq->rxcsum, "# of times hardware assisted with checksum");
3024 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3025 CTLFLAG_RD, &rxq->vlan_extraction,
3026 "# of times hardware extracted 802.1Q tag");
3028 add_fl_sysctls(&vi->ctx, oid, &rxq->fl);
3034 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3038 #if defined(INET) || defined(INET6)
3040 tcp_lro_free(&rxq->lro);
3041 rxq->lro.ifp = NULL;
3045 rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3047 bzero(rxq, sizeof(*rxq));
3054 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3055 int intr_idx, int idx, struct sysctl_oid *oid)
3058 struct sysctl_oid_list *children;
3061 rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
3062 vi->pi->rx_chan_map);
3066 children = SYSCTL_CHILDREN(oid);
3068 snprintf(name, sizeof(name), "%d", idx);
3069 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3071 children = SYSCTL_CHILDREN(oid);
3073 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
3074 CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16,
3075 "I", "absolute id of the queue");
3076 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
3077 CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16,
3078 "I", "SGE context id of the queue");
3079 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3080 CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I",
3083 add_fl_sysctls(&vi->ctx, oid, &ofld_rxq->fl);
3089 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3093 rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3095 bzero(ofld_rxq, sizeof(*ofld_rxq));
3103 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3104 int idx, struct sysctl_oid *oid)
3107 struct sysctl_oid_list *children;
3108 struct sysctl_ctx_list *ctx;
3111 struct adapter *sc = vi->pi->adapter;
3112 struct netmap_adapter *na = NA(vi->ifp);
3116 len = vi->qsize_rxq * IQ_ESIZE;
3117 rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3118 &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3122 len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3123 rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3124 &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3130 nm_rxq->iq_cidx = 0;
3131 nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3132 nm_rxq->iq_gen = F_RSPD_GEN;
3133 nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3134 nm_rxq->fl_sidx = na->num_rx_desc;
3135 nm_rxq->intr_idx = intr_idx;
3138 children = SYSCTL_CHILDREN(oid);
3140 snprintf(name, sizeof(name), "%d", idx);
3141 oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3143 children = SYSCTL_CHILDREN(oid);
3145 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3146 CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3147 "I", "absolute id of the queue");
3148 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3149 CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3150 "I", "SGE context id of the queue");
3151 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3152 CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3155 children = SYSCTL_CHILDREN(oid);
3156 oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3158 children = SYSCTL_CHILDREN(oid);
3160 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3161 CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3162 "I", "SGE context id of the freelist");
3163 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3164 &nm_rxq->fl_cidx, 0, "consumer index");
3165 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3166 &nm_rxq->fl_pidx, 0, "producer index");
3173 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3175 struct adapter *sc = vi->pi->adapter;
3177 free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3179 free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3186 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3187 struct sysctl_oid *oid)
3191 struct port_info *pi = vi->pi;
3192 struct adapter *sc = pi->adapter;
3193 struct netmap_adapter *na = NA(vi->ifp);
3195 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3197 len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3198 rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3199 &nm_txq->ba, (void **)&nm_txq->desc);
3203 nm_txq->pidx = nm_txq->cidx = 0;
3204 nm_txq->sidx = na->num_tx_desc;
3206 nm_txq->iqidx = iqidx;
3207 nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3208 V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_VF_VLD(1) |
3209 V_TXPKT_VF(vi->viid));
3211 snprintf(name, sizeof(name), "%d", idx);
3212 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3213 NULL, "netmap tx queue");
3214 children = SYSCTL_CHILDREN(oid);
3216 SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3217 &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3218 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3219 CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3221 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3222 CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3229 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3231 struct adapter *sc = vi->pi->adapter;
3233 free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3241 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3244 struct fw_eq_ctrl_cmd c;
3245 int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3247 bzero(&c, sizeof(c));
3249 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3250 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3251 V_FW_EQ_CTRL_CMD_VFN(0));
3252 c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3253 F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3254 c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3255 c.physeqid_pkd = htobe32(0);
3256 c.fetchszm_to_iqid =
3257 htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3258 V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3259 F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3261 htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3262 V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3263 V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3264 c.eqaddr = htobe64(eq->ba);
3266 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3268 device_printf(sc->dev,
3269 "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3272 eq->flags |= EQ_ALLOCATED;
3274 eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3275 cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3276 if (cntxt_id >= sc->sge.neq)
3277 panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3278 cntxt_id, sc->sge.neq - 1);
3279 sc->sge.eqmap[cntxt_id] = eq;
3285 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3288 struct fw_eq_eth_cmd c;
3289 int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3291 bzero(&c, sizeof(c));
3293 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3294 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3295 V_FW_EQ_ETH_CMD_VFN(0));
3296 c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3297 F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3298 c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3299 F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3300 c.fetchszm_to_iqid =
3301 htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3302 V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3303 V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3304 c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3305 V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3306 V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3307 c.eqaddr = htobe64(eq->ba);
3309 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3311 device_printf(vi->dev,
3312 "failed to create Ethernet egress queue: %d\n", rc);
3315 eq->flags |= EQ_ALLOCATED;
3317 eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3318 cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3319 if (cntxt_id >= sc->sge.neq)
3320 panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3321 cntxt_id, sc->sge.neq - 1);
3322 sc->sge.eqmap[cntxt_id] = eq;
3329 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3332 struct fw_eq_ofld_cmd c;
3333 int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3335 bzero(&c, sizeof(c));
3337 c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3338 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3339 V_FW_EQ_OFLD_CMD_VFN(0));
3340 c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3341 F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3342 c.fetchszm_to_iqid =
3343 htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3344 V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3345 F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3347 htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3348 V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3349 V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3350 c.eqaddr = htobe64(eq->ba);
3352 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3354 device_printf(vi->dev,
3355 "failed to create egress queue for TCP offload: %d\n", rc);
3358 eq->flags |= EQ_ALLOCATED;
3360 eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3361 cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3362 if (cntxt_id >= sc->sge.neq)
3363 panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3364 cntxt_id, sc->sge.neq - 1);
3365 sc->sge.eqmap[cntxt_id] = eq;
3372 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3377 mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3379 qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3380 len = qsize * EQ_ESIZE;
3381 rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3382 &eq->ba, (void **)&eq->desc);
3386 eq->pidx = eq->cidx = 0;
3387 eq->equeqidx = eq->dbidx = 0;
3388 eq->doorbells = sc->doorbells;
3390 switch (eq->flags & EQ_TYPEMASK) {
3392 rc = ctrl_eq_alloc(sc, eq);
3396 rc = eth_eq_alloc(sc, vi, eq);
3401 rc = ofld_eq_alloc(sc, vi, eq);
3406 panic("%s: invalid eq type %d.", __func__,
3407 eq->flags & EQ_TYPEMASK);
3410 device_printf(sc->dev,
3411 "failed to allocate egress queue(%d): %d\n",
3412 eq->flags & EQ_TYPEMASK, rc);
3415 if (isset(&eq->doorbells, DOORBELL_UDB) ||
3416 isset(&eq->doorbells, DOORBELL_UDBWC) ||
3417 isset(&eq->doorbells, DOORBELL_WCWR)) {
3418 uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3419 uint32_t mask = (1 << s_qpp) - 1;
3420 volatile uint8_t *udb;
3422 udb = sc->udbs_base + UDBS_DB_OFFSET;
3423 udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT; /* pg offset */
3424 eq->udb_qid = eq->cntxt_id & mask; /* id in page */
3425 if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3426 clrbit(&eq->doorbells, DOORBELL_WCWR);
3428 udb += eq->udb_qid << UDBS_SEG_SHIFT; /* seg offset */
3431 eq->udb = (volatile void *)udb;
3438 free_eq(struct adapter *sc, struct sge_eq *eq)
3442 if (eq->flags & EQ_ALLOCATED) {
3443 switch (eq->flags & EQ_TYPEMASK) {
3445 rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3450 rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3456 rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3462 panic("%s: invalid eq type %d.", __func__,
3463 eq->flags & EQ_TYPEMASK);
3466 device_printf(sc->dev,
3467 "failed to free egress queue (%d): %d\n",
3468 eq->flags & EQ_TYPEMASK, rc);
3471 eq->flags &= ~EQ_ALLOCATED;
3474 free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3476 if (mtx_initialized(&eq->eq_lock))
3477 mtx_destroy(&eq->eq_lock);
3479 bzero(eq, sizeof(*eq));
3484 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3485 struct sysctl_oid *oid)
3488 struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3489 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3491 rc = alloc_eq(sc, vi, &wrq->eq);
3496 TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3497 TAILQ_INIT(&wrq->incomplete_wrs);
3498 STAILQ_INIT(&wrq->wr_list);
3499 wrq->nwr_pending = 0;
3500 wrq->ndesc_needed = 0;
3502 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3503 &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3504 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3505 CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3507 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3508 CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3510 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3511 &wrq->tx_wrs_direct, "# of work requests (direct)");
3512 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3513 &wrq->tx_wrs_copied, "# of work requests (copied)");
3519 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3523 rc = free_eq(sc, &wrq->eq);
3527 bzero(wrq, sizeof(*wrq));
3532 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
3533 struct sysctl_oid *oid)
3536 struct port_info *pi = vi->pi;
3537 struct adapter *sc = pi->adapter;
3538 struct sge_eq *eq = &txq->eq;
3540 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3542 rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
3545 device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
3549 rc = alloc_eq(sc, vi, eq);
3551 mp_ring_free(txq->r);
3556 /* Can't fail after this point. */
3558 TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
3560 txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
3561 txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3562 V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_VF_VLD(1) |
3563 V_TXPKT_VF(vi->viid));
3565 txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
3568 snprintf(name, sizeof(name), "%d", idx);
3569 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3571 children = SYSCTL_CHILDREN(oid);
3573 SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3574 &eq->cntxt_id, 0, "SGE context id of the queue");
3575 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3576 CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
3578 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3579 CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
3582 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
3583 CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
3584 "traffic class (-1 means none)");
3586 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
3587 &txq->txcsum, "# of times hardware assisted with checksum");
3588 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
3589 CTLFLAG_RD, &txq->vlan_insertion,
3590 "# of times hardware inserted 802.1Q tag");
3591 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
3592 &txq->tso_wrs, "# of TSO work requests");
3593 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
3594 &txq->imm_wrs, "# of work requests with immediate data");
3595 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
3596 &txq->sgl_wrs, "# of work requests with direct SGL");
3597 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
3598 &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
3599 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
3600 CTLFLAG_RD, &txq->txpkts0_wrs,
3601 "# of txpkts (type 0) work requests");
3602 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
3603 CTLFLAG_RD, &txq->txpkts1_wrs,
3604 "# of txpkts (type 1) work requests");
3605 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
3606 CTLFLAG_RD, &txq->txpkts0_pkts,
3607 "# of frames tx'd using type0 txpkts work requests");
3608 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
3609 CTLFLAG_RD, &txq->txpkts1_pkts,
3610 "# of frames tx'd using type1 txpkts work requests");
3612 SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
3613 CTLFLAG_RD, &txq->r->enqueues,
3614 "# of enqueues to the mp_ring for this queue");
3615 SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
3616 CTLFLAG_RD, &txq->r->drops,
3617 "# of drops in the mp_ring for this queue");
3618 SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
3619 CTLFLAG_RD, &txq->r->starts,
3620 "# of normal consumer starts in the mp_ring for this queue");
3621 SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
3622 CTLFLAG_RD, &txq->r->stalls,
3623 "# of consumer stalls in the mp_ring for this queue");
3624 SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
3625 CTLFLAG_RD, &txq->r->restarts,
3626 "# of consumer restarts in the mp_ring for this queue");
3627 SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
3628 CTLFLAG_RD, &txq->r->abdications,
3629 "# of consumer abdications in the mp_ring for this queue");
3635 free_txq(struct vi_info *vi, struct sge_txq *txq)
3638 struct adapter *sc = vi->pi->adapter;
3639 struct sge_eq *eq = &txq->eq;
3641 rc = free_eq(sc, eq);
3645 sglist_free(txq->gl);
3646 free(txq->sdesc, M_CXGBE);
3647 mp_ring_free(txq->r);
3649 bzero(txq, sizeof(*txq));
3654 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
3656 bus_addr_t *ba = arg;
3659 ("%s meant for single segment mappings only.", __func__));
3661 *ba = error ? 0 : segs->ds_addr;
3665 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3669 n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
3673 v = fl->dbval | V_PIDX(n);
3675 *fl->udb = htole32(v);
3677 t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3678 IDXINCR(fl->dbidx, n, fl->sidx);
3682 * Fills up the freelist by allocating upto 'n' buffers. Buffers that are
3683 * recycled do not count towards this allocation budget.
3685 * Returns non-zero to indicate that this freelist should be added to the list
3686 * of starving freelists.
3689 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
3692 struct fl_sdesc *sd;
3695 struct cluster_layout *cll;
3696 struct sw_zone_info *swz;
3697 struct cluster_metadata *clm;
3699 uint16_t hw_cidx = fl->hw_cidx; /* stable snapshot */
3701 FL_LOCK_ASSERT_OWNED(fl);
3704 * We always stop at the begining of the hardware descriptor that's just
3705 * before the one with the hw cidx. This is to avoid hw pidx = hw cidx,
3706 * which would mean an empty freelist to the chip.
3708 max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
3709 if (fl->pidx == max_pidx * 8)
3712 d = &fl->desc[fl->pidx];
3713 sd = &fl->sdesc[fl->pidx];
3714 cll = &fl->cll_def; /* default layout */
3715 swz = &sc->sge.sw_zone_info[cll->zidx];
3719 if (sd->cl != NULL) {
3721 if (sd->nmbuf == 0) {
3723 * Fast recycle without involving any atomics on
3724 * the cluster's metadata (if the cluster has
3725 * metadata). This happens when all frames
3726 * received in the cluster were small enough to
3727 * fit within a single mbuf each.
3729 fl->cl_fast_recycled++;
3731 clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3733 MPASS(clm->refcount == 1);
3739 * Cluster is guaranteed to have metadata. Clusters
3740 * without metadata always take the fast recycle path
3741 * when they're recycled.
3743 clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3746 if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3748 counter_u64_add(extfree_rels, 1);
3751 sd->cl = NULL; /* gave up my reference */
3753 MPASS(sd->cl == NULL);
3755 cl = uma_zalloc(swz->zone, M_NOWAIT);
3756 if (__predict_false(cl == NULL)) {
3757 if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
3758 fl->cll_def.zidx == fl->cll_alt.zidx)
3761 /* fall back to the safe zone */
3763 swz = &sc->sge.sw_zone_info[cll->zidx];
3769 pa = pmap_kextract((vm_offset_t)cl);
3773 *d = htobe64(pa | cll->hwidx);
3774 clm = cl_metadata(sc, fl, cll, cl);
3786 if (__predict_false(++fl->pidx % 8 == 0)) {
3787 uint16_t pidx = fl->pidx / 8;
3789 if (__predict_false(pidx == fl->sidx)) {
3795 if (pidx == max_pidx)
3798 if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
3803 if (fl->pidx / 8 != fl->dbidx)
3806 return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
3810 * Attempt to refill all starving freelists.
3813 refill_sfl(void *arg)
3815 struct adapter *sc = arg;
3816 struct sge_fl *fl, *fl_temp;
3818 mtx_assert(&sc->sfl_lock, MA_OWNED);
3819 TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
3821 refill_fl(sc, fl, 64);
3822 if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
3823 TAILQ_REMOVE(&sc->sfl, fl, link);
3824 fl->flags &= ~FL_STARVING;
3829 if (!TAILQ_EMPTY(&sc->sfl))
3830 callout_schedule(&sc->sfl_callout, hz / 5);
3834 alloc_fl_sdesc(struct sge_fl *fl)
3837 fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
3844 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
3846 struct fl_sdesc *sd;
3847 struct cluster_metadata *clm;
3848 struct cluster_layout *cll;
3852 for (i = 0; i < fl->sidx * 8; i++, sd++) {
3857 clm = cl_metadata(sc, fl, cll, sd->cl);
3859 uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3860 else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3861 uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3862 counter_u64_add(extfree_rels, 1);
3867 free(fl->sdesc, M_CXGBE);
3872 get_pkt_gl(struct mbuf *m, struct sglist *gl)
3879 rc = sglist_append_mbuf(gl, m);
3880 if (__predict_false(rc != 0)) {
3881 panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
3882 "with %d.", __func__, m, mbuf_nsegs(m), rc);
3885 KASSERT(gl->sg_nseg == mbuf_nsegs(m),
3886 ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
3887 mbuf_nsegs(m), gl->sg_nseg));
3888 KASSERT(gl->sg_nseg > 0 &&
3889 gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
3890 ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
3891 gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
3895 * len16 for a txpkt WR with a GL. Includes the firmware work request header.
3898 txpkt_len16(u_int nsegs, u_int tso)
3904 nsegs--; /* first segment is part of ulptx_sgl */
3905 n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
3906 sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
3908 n += sizeof(struct cpl_tx_pkt_lso_core);
3910 return (howmany(n, 16));
3914 * len16 for a txpkts type 0 WR with a GL. Does not include the firmware work
3918 txpkts0_len16(u_int nsegs)
3924 nsegs--; /* first segment is part of ulptx_sgl */
3925 n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
3926 sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
3927 8 * ((3 * nsegs) / 2 + (nsegs & 1));
3929 return (howmany(n, 16));
3933 * len16 for a txpkts type 1 WR with a GL. Does not include the firmware work
3941 n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
3943 return (howmany(n, 16));
3947 imm_payload(u_int ndesc)
3951 n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
3952 sizeof(struct cpl_tx_pkt_core);
3958 * Write a txpkt WR for this packet to the hardware descriptors, update the
3959 * software descriptor, and advance the pidx. It is guaranteed that enough
3960 * descriptors are available.
3962 * The return value is the # of hardware descriptors used.
3965 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
3966 struct mbuf *m0, u_int available)
3968 struct sge_eq *eq = &txq->eq;
3969 struct tx_sdesc *txsd;
3970 struct cpl_tx_pkt_core *cpl;
3971 uint32_t ctrl; /* used in many unrelated places */
3973 int len16, ndesc, pktlen, nsegs;
3976 TXQ_LOCK_ASSERT_OWNED(txq);
3978 MPASS(available > 0 && available < eq->sidx);
3980 len16 = mbuf_len16(m0);
3981 nsegs = mbuf_nsegs(m0);
3982 pktlen = m0->m_pkthdr.len;
3983 ctrl = sizeof(struct cpl_tx_pkt_core);
3985 ctrl += sizeof(struct cpl_tx_pkt_lso_core);
3986 else if (pktlen <= imm_payload(2) && available >= 2) {
3987 /* Immediate data. Recalculate len16 and set nsegs to 0. */
3989 len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
3990 sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
3993 ndesc = howmany(len16, EQ_ESIZE / 16);
3994 MPASS(ndesc <= available);
3996 /* Firmware work request header */
3997 MPASS(wr == (void *)&eq->desc[eq->pidx]);
3998 wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
3999 V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4001 ctrl = V_FW_WR_LEN16(len16);
4002 wr->equiq_to_len16 = htobe32(ctrl);
4005 if (needs_tso(m0)) {
4006 struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4008 KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4009 m0->m_pkthdr.l4hlen > 0,
4010 ("%s: mbuf %p needs TSO but missing header lengths",
4013 ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4014 F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4015 | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4016 if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4017 ctrl |= V_LSO_ETHHDR_LEN(1);
4018 if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4021 lso->lso_ctrl = htobe32(ctrl);
4022 lso->ipid_ofst = htobe16(0);
4023 lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4024 lso->seqno_offset = htobe32(0);
4025 lso->len = htobe32(pktlen);
4027 cpl = (void *)(lso + 1);
4031 cpl = (void *)(wr + 1);
4033 /* Checksum offload */
4035 if (needs_l3_csum(m0) == 0)
4036 ctrl1 |= F_TXPKT_IPCSUM_DIS;
4037 if (needs_l4_csum(m0) == 0)
4038 ctrl1 |= F_TXPKT_L4CSUM_DIS;
4039 if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4040 CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4041 txq->txcsum++; /* some hardware assistance provided */
4043 /* VLAN tag insertion */
4044 if (needs_vlan_insertion(m0)) {
4045 ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4046 txq->vlan_insertion++;
4050 cpl->ctrl0 = txq->cpl_ctrl0;
4052 cpl->len = htobe16(pktlen);
4053 cpl->ctrl1 = htobe64(ctrl1);
4056 dst = (void *)(cpl + 1);
4059 write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4064 for (m = m0; m != NULL; m = m->m_next) {
4065 copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4071 KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4078 txsd = &txq->sdesc[eq->pidx];
4080 txsd->desc_used = ndesc;
4086 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4088 u_int needed, nsegs1, nsegs2, l1, l2;
4090 if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4093 nsegs1 = mbuf_nsegs(m);
4094 nsegs2 = mbuf_nsegs(n);
4095 if (nsegs1 + nsegs2 == 2) {
4097 l1 = l2 = txpkts1_len16();
4100 l1 = txpkts0_len16(nsegs1);
4101 l2 = txpkts0_len16(nsegs2);
4103 txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4104 needed = howmany(txp->len16, EQ_ESIZE / 16);
4105 if (needed > SGE_MAX_WR_NDESC || needed > available)
4108 txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4109 if (txp->plen > 65535)
4113 set_mbuf_len16(m, l1);
4114 set_mbuf_len16(n, l2);
4120 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4122 u_int plen, len16, needed, nsegs;
4124 MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4126 nsegs = mbuf_nsegs(m);
4127 if (needs_tso(m) || (txp->wr_type == 1 && nsegs != 1))
4130 plen = txp->plen + m->m_pkthdr.len;
4134 if (txp->wr_type == 0)
4135 len16 = txpkts0_len16(nsegs);
4137 len16 = txpkts1_len16();
4138 needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4139 if (needed > SGE_MAX_WR_NDESC || needed > available)
4144 txp->len16 += len16;
4145 set_mbuf_len16(m, len16);
4151 * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4152 * the software descriptor, and advance the pidx. It is guaranteed that enough
4153 * descriptors are available.
4155 * The return value is the # of hardware descriptors used.
4158 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4159 struct mbuf *m0, const struct txpkts *txp, u_int available)
4161 struct sge_eq *eq = &txq->eq;
4162 struct tx_sdesc *txsd;
4163 struct cpl_tx_pkt_core *cpl;
4166 int ndesc, checkwrap;
4170 TXQ_LOCK_ASSERT_OWNED(txq);
4171 MPASS(txp->npkt > 0);
4172 MPASS(txp->plen < 65536);
4174 MPASS(m0->m_nextpkt != NULL);
4175 MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4176 MPASS(available > 0 && available < eq->sidx);
4178 ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4179 MPASS(ndesc <= available);
4181 MPASS(wr == (void *)&eq->desc[eq->pidx]);
4182 wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4183 ctrl = V_FW_WR_LEN16(txp->len16);
4184 wr->equiq_to_len16 = htobe32(ctrl);
4185 wr->plen = htobe16(txp->plen);
4186 wr->npkt = txp->npkt;
4188 wr->type = txp->wr_type;
4192 * At this point we are 16B into a hardware descriptor. If checkwrap is
4193 * set then we know the WR is going to wrap around somewhere. We'll
4194 * check for that at appropriate points.
4196 checkwrap = eq->sidx - ndesc < eq->pidx;
4197 for (m = m0; m != NULL; m = m->m_nextpkt) {
4198 if (txp->wr_type == 0) {
4199 struct ulp_txpkt *ulpmc;
4200 struct ulptx_idata *ulpsc;
4202 /* ULP master command */
4204 ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4205 V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4206 ulpmc->len = htobe32(mbuf_len16(m));
4208 /* ULP subcommand */
4209 ulpsc = (void *)(ulpmc + 1);
4210 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4212 ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4214 cpl = (void *)(ulpsc + 1);
4216 (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4217 cpl = (void *)&eq->desc[0];
4218 txq->txpkts0_pkts += txp->npkt;
4222 txq->txpkts1_pkts += txp->npkt;
4226 /* Checksum offload */
4228 if (needs_l3_csum(m) == 0)
4229 ctrl1 |= F_TXPKT_IPCSUM_DIS;
4230 if (needs_l4_csum(m) == 0)
4231 ctrl1 |= F_TXPKT_L4CSUM_DIS;
4232 if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4233 CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4234 txq->txcsum++; /* some hardware assistance provided */
4236 /* VLAN tag insertion */
4237 if (needs_vlan_insertion(m)) {
4238 ctrl1 |= F_TXPKT_VLAN_VLD |
4239 V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4240 txq->vlan_insertion++;
4244 cpl->ctrl0 = txq->cpl_ctrl0;
4246 cpl->len = htobe16(m->m_pkthdr.len);
4247 cpl->ctrl1 = htobe64(ctrl1);
4251 (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4252 flitp = (void *)&eq->desc[0];
4254 write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4258 txsd = &txq->sdesc[eq->pidx];
4260 txsd->desc_used = ndesc;
4266 * If the SGL ends on an address that is not 16 byte aligned, this function will
4267 * add a 0 filled flit at the end.
4270 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
4272 struct sge_eq *eq = &txq->eq;
4273 struct sglist *gl = txq->gl;
4274 struct sglist_seg *seg;
4275 __be64 *flitp, *wrap;
4276 struct ulptx_sgl *usgl;
4277 int i, nflits, nsegs;
4279 KASSERT(((uintptr_t)(*to) & 0xf) == 0,
4280 ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
4281 MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4282 MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4285 nsegs = gl->sg_nseg;
4288 nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
4289 flitp = (__be64 *)(*to);
4290 wrap = (__be64 *)(&eq->desc[eq->sidx]);
4291 seg = &gl->sg_segs[0];
4292 usgl = (void *)flitp;
4295 * We start at a 16 byte boundary somewhere inside the tx descriptor
4296 * ring, so we're at least 16 bytes away from the status page. There is
4297 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
4300 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
4301 V_ULPTX_NSGE(nsegs));
4302 usgl->len0 = htobe32(seg->ss_len);
4303 usgl->addr0 = htobe64(seg->ss_paddr);
4306 if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
4308 /* Won't wrap around at all */
4310 for (i = 0; i < nsegs - 1; i++, seg++) {
4311 usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
4312 usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
4315 usgl->sge[i / 2].len[1] = htobe32(0);
4319 /* Will wrap somewhere in the rest of the SGL */
4321 /* 2 flits already written, write the rest flit by flit */
4322 flitp = (void *)(usgl + 1);
4323 for (i = 0; i < nflits - 2; i++) {
4325 flitp = (void *)eq->desc;
4326 *flitp++ = get_flit(seg, nsegs - 1, i);
4331 MPASS(((uintptr_t)flitp) & 0xf);
4335 MPASS((((uintptr_t)flitp) & 0xf) == 0);
4336 if (__predict_false(flitp == wrap))
4337 *to = (void *)eq->desc;
4339 *to = (void *)flitp;
4343 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
4346 MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4347 MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4349 if (__predict_true((uintptr_t)(*to) + len <=
4350 (uintptr_t)&eq->desc[eq->sidx])) {
4351 bcopy(from, *to, len);
4354 int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
4356 bcopy(from, *to, portion);
4358 portion = len - portion; /* remaining */
4359 bcopy(from, (void *)eq->desc, portion);
4360 (*to) = (caddr_t)eq->desc + portion;
4365 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
4373 clrbit(&db, DOORBELL_WCWR);
4376 switch (ffs(db) - 1) {
4378 *eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4381 case DOORBELL_WCWR: {
4382 volatile uint64_t *dst, *src;
4386 * Queues whose 128B doorbell segment fits in the page do not
4387 * use relative qid (udb_qid is always 0). Only queues with
4388 * doorbell segments can do WCWR.
4390 KASSERT(eq->udb_qid == 0 && n == 1,
4391 ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
4392 __func__, eq->doorbells, n, eq->dbidx, eq));
4394 dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
4397 src = (void *)&eq->desc[i];
4398 while (src != (void *)&eq->desc[i + 1])
4404 case DOORBELL_UDBWC:
4405 *eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4410 t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
4411 V_QID(eq->cntxt_id) | V_PIDX(n));
4415 IDXINCR(eq->dbidx, n, eq->sidx);
4419 reclaimable_tx_desc(struct sge_eq *eq)
4423 hw_cidx = read_hw_cidx(eq);
4424 return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
4428 total_available_tx_desc(struct sge_eq *eq)
4430 uint16_t hw_cidx, pidx;
4432 hw_cidx = read_hw_cidx(eq);
4435 if (pidx == hw_cidx)
4436 return (eq->sidx - 1);
4438 return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
4441 static inline uint16_t
4442 read_hw_cidx(struct sge_eq *eq)
4444 struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
4445 uint16_t cidx = spg->cidx; /* stable snapshot */
4447 return (be16toh(cidx));
4451 * Reclaim 'n' descriptors approximately.
4454 reclaim_tx_descs(struct sge_txq *txq, u_int n)
4456 struct tx_sdesc *txsd;
4457 struct sge_eq *eq = &txq->eq;
4458 u_int can_reclaim, reclaimed;
4460 TXQ_LOCK_ASSERT_OWNED(txq);
4464 can_reclaim = reclaimable_tx_desc(eq);
4465 while (can_reclaim && reclaimed < n) {
4467 struct mbuf *m, *nextpkt;
4469 txsd = &txq->sdesc[eq->cidx];
4470 ndesc = txsd->desc_used;
4472 /* Firmware doesn't return "partial" credits. */
4473 KASSERT(can_reclaim >= ndesc,
4474 ("%s: unexpected number of credits: %d, %d",
4475 __func__, can_reclaim, ndesc));
4477 for (m = txsd->m; m != NULL; m = nextpkt) {
4478 nextpkt = m->m_nextpkt;
4479 m->m_nextpkt = NULL;
4483 can_reclaim -= ndesc;
4484 IDXINCR(eq->cidx, ndesc, eq->sidx);
4491 tx_reclaim(void *arg, int n)
4493 struct sge_txq *txq = arg;
4494 struct sge_eq *eq = &txq->eq;
4497 if (TXQ_TRYLOCK(txq) == 0)
4499 n = reclaim_tx_descs(txq, 32);
4500 if (eq->cidx == eq->pidx)
4501 eq->equeqidx = eq->pidx;
4507 get_flit(struct sglist_seg *segs, int nsegs, int idx)
4509 int i = (idx / 3) * 2;
4515 rc = htobe32(segs[i].ss_len);
4517 rc |= (uint64_t)htobe32(segs[i + 1].ss_len) << 32;
4522 return (htobe64(segs[i].ss_paddr));
4524 return (htobe64(segs[i + 1].ss_paddr));
4531 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
4533 int8_t zidx, hwidx, idx;
4534 uint16_t region1, region3;
4535 int spare, spare_needed, n;
4536 struct sw_zone_info *swz;
4537 struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
4540 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
4541 * large enough for the max payload and cluster metadata. Otherwise
4542 * settle for the largest bufsize that leaves enough room in the cluster
4545 * Without buffer packing: Look for the smallest zone which has a
4546 * bufsize large enough for the max payload. Settle for the largest
4547 * bufsize available if there's nothing big enough for max payload.
4549 spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
4550 swz = &sc->sge.sw_zone_info[0];
4552 for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
4553 if (swz->size > largest_rx_cluster) {
4554 if (__predict_true(hwidx != -1))
4558 * This is a misconfiguration. largest_rx_cluster is
4559 * preventing us from finding a refill source. See
4560 * dev.t5nex.<n>.buffer_sizes to figure out why.
4562 device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
4563 " refill source for fl %p (dma %u). Ignored.\n",
4564 largest_rx_cluster, fl, maxp);
4566 for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
4567 hwb = &hwb_list[idx];
4568 spare = swz->size - hwb->size;
4569 if (spare < spare_needed)
4572 hwidx = idx; /* best option so far */
4573 if (hwb->size >= maxp) {
4575 if ((fl->flags & FL_BUF_PACKING) == 0)
4576 goto done; /* stop looking (not packing) */
4578 if (swz->size >= safest_rx_cluster)
4579 goto done; /* stop looking (packing) */
4581 break; /* keep looking, next zone */
4585 /* A usable hwidx has been located. */
4587 hwb = &hwb_list[hwidx];
4589 swz = &sc->sge.sw_zone_info[zidx];
4591 region3 = swz->size - hwb->size;
4594 * Stay within this zone and see if there is a better match when mbuf
4595 * inlining is allowed. Remember that the hwidx's are sorted in
4596 * decreasing order of size (so in increasing order of spare area).
4598 for (idx = hwidx; idx != -1; idx = hwb->next) {
4599 hwb = &hwb_list[idx];
4600 spare = swz->size - hwb->size;
4602 if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
4606 * Do not inline mbufs if doing so would violate the pad/pack
4607 * boundary alignment requirement.
4609 if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
4611 if (fl->flags & FL_BUF_PACKING &&
4612 (MSIZE % sc->params.sge.pack_boundary) != 0)
4615 if (spare < CL_METADATA_SIZE + MSIZE)
4617 n = (spare - CL_METADATA_SIZE) / MSIZE;
4618 if (n > howmany(hwb->size, maxp))
4622 if (fl->flags & FL_BUF_PACKING) {
4623 region1 = n * MSIZE;
4624 region3 = spare - region1;
4627 region3 = spare - region1;
4632 KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
4633 ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
4634 KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
4635 ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
4636 KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
4637 sc->sge.sw_zone_info[zidx].size,
4638 ("%s: bad buffer layout for fl %p, maxp %d. "
4639 "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4640 sc->sge.sw_zone_info[zidx].size, region1,
4641 sc->sge.hw_buf_info[hwidx].size, region3));
4642 if (fl->flags & FL_BUF_PACKING || region1 > 0) {
4643 KASSERT(region3 >= CL_METADATA_SIZE,
4644 ("%s: no room for metadata. fl %p, maxp %d; "
4645 "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4646 sc->sge.sw_zone_info[zidx].size, region1,
4647 sc->sge.hw_buf_info[hwidx].size, region3));
4648 KASSERT(region1 % MSIZE == 0,
4649 ("%s: bad mbuf region for fl %p, maxp %d. "
4650 "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4651 sc->sge.sw_zone_info[zidx].size, region1,
4652 sc->sge.hw_buf_info[hwidx].size, region3));
4655 fl->cll_def.zidx = zidx;
4656 fl->cll_def.hwidx = hwidx;
4657 fl->cll_def.region1 = region1;
4658 fl->cll_def.region3 = region3;
4662 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
4664 struct sge *s = &sc->sge;
4665 struct hw_buf_info *hwb;
4666 struct sw_zone_info *swz;
4670 if (fl->flags & FL_BUF_PACKING)
4671 hwidx = s->safe_hwidx2; /* with room for metadata */
4672 else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
4673 hwidx = s->safe_hwidx2;
4674 hwb = &s->hw_buf_info[hwidx];
4675 swz = &s->sw_zone_info[hwb->zidx];
4676 spare = swz->size - hwb->size;
4678 /* no good if there isn't room for an mbuf as well */
4679 if (spare < CL_METADATA_SIZE + MSIZE)
4680 hwidx = s->safe_hwidx1;
4682 hwidx = s->safe_hwidx1;
4685 /* No fallback source */
4686 fl->cll_alt.hwidx = -1;
4687 fl->cll_alt.zidx = -1;
4692 hwb = &s->hw_buf_info[hwidx];
4693 swz = &s->sw_zone_info[hwb->zidx];
4694 spare = swz->size - hwb->size;
4695 fl->cll_alt.hwidx = hwidx;
4696 fl->cll_alt.zidx = hwb->zidx;
4697 if (allow_mbufs_in_cluster &&
4698 (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
4699 fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
4701 fl->cll_alt.region1 = 0;
4702 fl->cll_alt.region3 = spare - fl->cll_alt.region1;
4706 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
4708 mtx_lock(&sc->sfl_lock);
4710 if ((fl->flags & FL_DOOMED) == 0) {
4711 fl->flags |= FL_STARVING;
4712 TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
4713 callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
4716 mtx_unlock(&sc->sfl_lock);
4720 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
4722 struct sge_wrq *wrq = (void *)eq;
4724 atomic_readandclear_int(&eq->equiq);
4725 taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
4729 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
4731 struct sge_txq *txq = (void *)eq;
4733 MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
4735 atomic_readandclear_int(&eq->equiq);
4736 mp_ring_check_drainage(txq->r, 0);
4737 taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
4741 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
4744 const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
4745 unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
4746 struct adapter *sc = iq->adapter;
4747 struct sge *s = &sc->sge;
4749 static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
4750 &handle_wrq_egr_update, &handle_eth_egr_update,
4751 &handle_wrq_egr_update};
4753 KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4756 eq = s->eqmap[qid - s->eq_start];
4757 (*h[eq->flags & EQ_TYPEMASK])(sc, eq);
4762 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
4763 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
4764 offsetof(struct cpl_fw6_msg, data));
4767 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
4769 struct adapter *sc = iq->adapter;
4770 const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
4772 KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4775 if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
4776 const struct rss_header *rss2;
4778 rss2 = (const struct rss_header *)&cpl->data[0];
4779 return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
4782 return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
4786 sysctl_uint16(SYSCTL_HANDLER_ARGS)
4788 uint16_t *id = arg1;
4791 return sysctl_handle_int(oidp, &i, 0, req);
4795 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
4797 struct sge *s = arg1;
4798 struct hw_buf_info *hwb = &s->hw_buf_info[0];
4799 struct sw_zone_info *swz = &s->sw_zone_info[0];
4804 sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
4805 for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
4806 if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
4811 sbuf_printf(&sb, "%u%c ", hwb->size, c);
4815 rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
4821 sysctl_tc(SYSCTL_HANDLER_ARGS)
4823 struct vi_info *vi = arg1;
4824 struct port_info *pi;
4826 struct sge_txq *txq;
4827 struct tx_sched_class *tc;
4828 int qidx = arg2, rc, tc_idx;
4829 uint32_t fw_queue, fw_class;
4831 MPASS(qidx >= 0 && qidx < vi->ntxq);
4834 txq = &sc->sge.txq[vi->first_txq + qidx];
4836 tc_idx = txq->tc_idx;
4837 rc = sysctl_handle_int(oidp, &tc_idx, 0, req);
4838 if (rc != 0 || req->newptr == NULL)
4841 /* Note that -1 is legitimate input (it means unbind). */
4842 if (tc_idx < -1 || tc_idx >= sc->chip_params->nsched_cls)
4845 rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4stc");
4849 if (tc_idx == txq->tc_idx) {
4850 rc = 0; /* No change, nothing to do. */
4854 fw_queue = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
4855 V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH) |
4856 V_FW_PARAMS_PARAM_YZ(txq->eq.cntxt_id);
4859 fw_class = 0xffffffff; /* Unbind. */
4862 * Bind to a different class. Ethernet txq's are only allowed
4863 * to bind to cl-rl mode-class for now. XXX: too restrictive.
4865 tc = &pi->tc[tc_idx];
4866 if (tc->flags & TX_SC_OK &&
4867 tc->params.level == SCHED_CLASS_LEVEL_CL_RL &&
4868 tc->params.mode == SCHED_CLASS_MODE_CLASS) {
4869 /* Ok to proceed. */
4872 rc = tc->flags & TX_SC_OK ? EBUSY : ENXIO;
4877 rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &fw_queue, &fw_class);
4879 if (txq->tc_idx != -1) {
4880 tc = &pi->tc[txq->tc_idx];
4881 MPASS(tc->refcount > 0);
4885 tc = &pi->tc[tc_idx];
4888 txq->tc_idx = tc_idx;
4891 end_synchronized_op(sc, 0);