2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2011 Chelsio Communications, Inc.
6 * Written by: Navdeep Parhar <np@FreeBSD.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
34 #include "opt_inet6.h"
35 #include "opt_ratelimit.h"
37 #include <sys/types.h>
38 #include <sys/eventhandler.h>
40 #include <sys/socket.h>
41 #include <sys/kernel.h>
42 #include <sys/malloc.h>
43 #include <sys/queue.h>
45 #include <sys/taskqueue.h>
47 #include <sys/sglist.h>
48 #include <sys/sysctl.h>
50 #include <sys/counter.h>
52 #include <net/ethernet.h>
54 #include <net/if_vlan_var.h>
55 #include <netinet/in.h>
56 #include <netinet/ip.h>
57 #include <netinet/ip6.h>
58 #include <netinet/tcp.h>
59 #include <netinet/udp.h>
60 #include <machine/in_cksum.h>
61 #include <machine/md_var.h>
65 #include <machine/bus.h>
66 #include <sys/selinfo.h>
67 #include <net/if_var.h>
68 #include <net/netmap.h>
69 #include <dev/netmap/netmap_kern.h>
72 #include "common/common.h"
73 #include "common/t4_regs.h"
74 #include "common/t4_regs_values.h"
75 #include "common/t4_msg.h"
77 #include "t4_mp_ring.h"
79 #ifdef T4_PKT_TIMESTAMP
80 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
82 #define RX_COPY_THRESHOLD MINCLSIZE
85 /* Internal mbuf flags stored in PH_loc.eight[1]. */
86 #define MC_RAW_WR 0x02
89 * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
90 * 0-7 are valid values.
92 static int fl_pktshift = 0;
93 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pktshift, CTLFLAG_RDTUN, &fl_pktshift, 0,
94 "payload DMA offset in rx buffer (bytes)");
97 * Pad ethernet payload up to this boundary.
98 * -1: driver should figure out a good value.
100 * Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
103 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pad, CTLFLAG_RDTUN, &fl_pad, 0,
104 "payload pad boundary (bytes)");
107 * Status page length.
108 * -1: driver should figure out a good value.
109 * 64 or 128 are the only other valid values.
111 static int spg_len = -1;
112 SYSCTL_INT(_hw_cxgbe, OID_AUTO, spg_len, CTLFLAG_RDTUN, &spg_len, 0,
113 "status page size (bytes)");
117 * -1: no congestion feedback (not recommended).
118 * 0: backpressure the channel instead of dropping packets right away.
119 * 1: no backpressure, drop packets for the congested queue immediately.
121 static int cong_drop = 0;
122 SYSCTL_INT(_hw_cxgbe, OID_AUTO, cong_drop, CTLFLAG_RDTUN, &cong_drop, 0,
123 "Congestion control for RX queues (0 = backpressure, 1 = drop");
126 * Deliver multiple frames in the same free list buffer if they fit.
127 * -1: let the driver decide whether to enable buffer packing or not.
128 * 0: disable buffer packing.
129 * 1: enable buffer packing.
131 static int buffer_packing = -1;
132 SYSCTL_INT(_hw_cxgbe, OID_AUTO, buffer_packing, CTLFLAG_RDTUN, &buffer_packing,
133 0, "Enable buffer packing");
136 * Start next frame in a packed buffer at this boundary.
137 * -1: driver should figure out a good value.
138 * T4: driver will ignore this and use the same value as fl_pad above.
139 * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
141 static int fl_pack = -1;
142 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pack, CTLFLAG_RDTUN, &fl_pack, 0,
143 "payload pack boundary (bytes)");
146 * Allow the driver to create mbuf(s) in a cluster allocated for rx.
147 * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
148 * 1: ok to create mbuf(s) within a cluster if there is room.
150 static int allow_mbufs_in_cluster = 1;
151 SYSCTL_INT(_hw_cxgbe, OID_AUTO, allow_mbufs_in_cluster, CTLFLAG_RDTUN,
152 &allow_mbufs_in_cluster, 0,
153 "Allow driver to create mbufs within a rx cluster");
156 * Largest rx cluster size that the driver is allowed to allocate.
158 static int largest_rx_cluster = MJUM16BYTES;
159 SYSCTL_INT(_hw_cxgbe, OID_AUTO, largest_rx_cluster, CTLFLAG_RDTUN,
160 &largest_rx_cluster, 0, "Largest rx cluster (bytes)");
163 * Size of cluster allocation that's most likely to succeed. The driver will
164 * fall back to this size if it fails to allocate clusters larger than this.
166 static int safest_rx_cluster = PAGE_SIZE;
167 SYSCTL_INT(_hw_cxgbe, OID_AUTO, safest_rx_cluster, CTLFLAG_RDTUN,
168 &safest_rx_cluster, 0, "Safe rx cluster (bytes)");
172 * Knob to control TCP timestamp rewriting, and the granularity of the tick used
173 * for rewriting. -1 and 0-3 are all valid values.
174 * -1: hardware should leave the TCP timestamps alone.
180 static int tsclk = -1;
181 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tsclk, CTLFLAG_RDTUN, &tsclk, 0,
182 "Control TCP timestamp rewriting when using pacing");
184 static int eo_max_backlog = 1024 * 1024;
185 SYSCTL_INT(_hw_cxgbe, OID_AUTO, eo_max_backlog, CTLFLAG_RDTUN, &eo_max_backlog,
186 0, "Maximum backlog of ratelimited data per flow");
190 * The interrupt holdoff timers are multiplied by this value on T6+.
191 * 1 and 3-17 (both inclusive) are legal values.
193 static int tscale = 1;
194 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tscale, CTLFLAG_RDTUN, &tscale, 0,
195 "Interrupt holdoff timer scale on T6+");
198 * Number of LRO entries in the lro_ctrl structure per rx queue.
200 static int lro_entries = TCP_LRO_ENTRIES;
201 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_entries, CTLFLAG_RDTUN, &lro_entries, 0,
202 "Number of LRO entries per RX queue");
205 * This enables presorting of frames before they're fed into tcp_lro_rx.
207 static int lro_mbufs = 0;
208 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_mbufs, CTLFLAG_RDTUN, &lro_mbufs, 0,
209 "Enable presorting of LRO frames");
212 u_int wr_type; /* type 0 or type 1 */
213 u_int npkt; /* # of packets in this work request */
214 u_int plen; /* total payload (sum of all packets) */
215 u_int len16; /* # of 16B pieces used by this work request */
218 /* A packet's SGL. This + m_pkthdr has all info needed for tx */
221 struct sglist_seg seg[TX_SGL_SEGS];
224 static int service_iq(struct sge_iq *, int);
225 static int service_iq_fl(struct sge_iq *, int);
226 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
227 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
228 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
229 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
230 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
232 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
233 bus_addr_t *, void **);
234 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
236 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
238 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
239 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
241 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
242 struct sysctl_oid *, struct sge_fl *);
243 static int alloc_fwq(struct adapter *);
244 static int free_fwq(struct adapter *);
245 static int alloc_ctrlq(struct adapter *, struct sge_wrq *, int,
246 struct sysctl_oid *);
247 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
248 struct sysctl_oid *);
249 static int free_rxq(struct vi_info *, struct sge_rxq *);
251 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
252 struct sysctl_oid *);
253 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
256 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
257 struct sysctl_oid *);
258 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
259 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
260 struct sysctl_oid *);
261 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
263 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
264 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
265 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
266 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
268 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
269 static int free_eq(struct adapter *, struct sge_eq *);
270 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
271 struct sysctl_oid *);
272 static int free_wrq(struct adapter *, struct sge_wrq *);
273 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
274 struct sysctl_oid *);
275 static int free_txq(struct vi_info *, struct sge_txq *);
276 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
277 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
278 static int refill_fl(struct adapter *, struct sge_fl *, int);
279 static void refill_sfl(void *);
280 static int alloc_fl_sdesc(struct sge_fl *);
281 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
282 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
283 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
284 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
286 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
287 static inline u_int txpkt_len16(u_int, u_int);
288 static inline u_int txpkt_vm_len16(u_int, u_int);
289 static inline u_int txpkts0_len16(u_int);
290 static inline u_int txpkts1_len16(void);
291 static u_int write_raw_wr(struct sge_txq *, void *, struct mbuf *, u_int);
292 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
293 struct mbuf *, u_int);
294 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
295 struct fw_eth_tx_pkt_vm_wr *, struct mbuf *, u_int);
296 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
297 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
298 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
299 struct mbuf *, const struct txpkts *, u_int);
300 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
301 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
302 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
303 static inline uint16_t read_hw_cidx(struct sge_eq *);
304 static inline u_int reclaimable_tx_desc(struct sge_eq *);
305 static inline u_int total_available_tx_desc(struct sge_eq *);
306 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
307 static void tx_reclaim(void *, int);
308 static __be64 get_flit(struct sglist_seg *, int, int);
309 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
311 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
313 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
314 static void wrq_tx_drain(void *, int);
315 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
317 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
318 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
320 static inline u_int txpkt_eo_len16(u_int, u_int, u_int);
321 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *,
325 static counter_u64_t extfree_refs;
326 static counter_u64_t extfree_rels;
328 an_handler_t t4_an_handler;
329 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
330 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
331 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
332 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
333 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
334 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES];
335 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES];
338 t4_register_an_handler(an_handler_t h)
342 MPASS(h == NULL || t4_an_handler == NULL);
344 loc = (uintptr_t *)&t4_an_handler;
345 atomic_store_rel_ptr(loc, (uintptr_t)h);
349 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
353 MPASS(type < nitems(t4_fw_msg_handler));
354 MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
356 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
357 * handler dispatch table. Reject any attempt to install a handler for
360 MPASS(type != FW_TYPE_RSSCPL);
361 MPASS(type != FW6_TYPE_RSSCPL);
363 loc = (uintptr_t *)&t4_fw_msg_handler[type];
364 atomic_store_rel_ptr(loc, (uintptr_t)h);
368 t4_register_cpl_handler(int opcode, cpl_handler_t h)
372 MPASS(opcode < nitems(t4_cpl_handler));
373 MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
375 loc = (uintptr_t *)&t4_cpl_handler[opcode];
376 atomic_store_rel_ptr(loc, (uintptr_t)h);
380 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
383 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
390 if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) {
392 * The return code for filter-write is put in the CPL cookie so
393 * we have to rely on the hardware tid (is_ftid) to determine
394 * that this is a response to a filter.
396 cookie = CPL_COOKIE_FILTER;
398 cookie = G_COOKIE(cpl->cookie);
400 MPASS(cookie > CPL_COOKIE_RESERVED);
401 MPASS(cookie < nitems(set_tcb_rpl_handlers));
403 return (set_tcb_rpl_handlers[cookie](iq, rss, m));
407 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
410 const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
415 cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
416 return (l2t_write_rpl_handlers[cookie](iq, rss, m));
420 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
423 const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
424 u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
427 MPASS(cookie != CPL_COOKIE_RESERVED);
429 return (act_open_rpl_handlers[cookie](iq, rss, m));
433 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss,
436 struct adapter *sc = iq->adapter;
440 if (is_hashfilter(sc))
441 cookie = CPL_COOKIE_HASHFILTER;
443 cookie = CPL_COOKIE_TOM;
445 return (abort_rpl_rss_handlers[cookie](iq, rss, m));
449 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
451 struct adapter *sc = iq->adapter;
452 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
453 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
457 if (is_etid(sc, tid))
458 cookie = CPL_COOKIE_ETHOFLD;
460 cookie = CPL_COOKIE_TOM;
462 return (fw4_ack_handlers[cookie](iq, rss, m));
466 t4_init_shared_cpl_handlers(void)
469 t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
470 t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
471 t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
472 t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler);
473 t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler);
477 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
481 MPASS(opcode < nitems(t4_cpl_handler));
482 MPASS(cookie > CPL_COOKIE_RESERVED);
483 MPASS(cookie < NUM_CPL_COOKIES);
484 MPASS(t4_cpl_handler[opcode] != NULL);
487 case CPL_SET_TCB_RPL:
488 loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
490 case CPL_L2T_WRITE_RPL:
491 loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
493 case CPL_ACT_OPEN_RPL:
494 loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
496 case CPL_ABORT_RPL_RSS:
497 loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie];
500 loc = (uintptr_t *)&fw4_ack_handlers[cookie];
506 MPASS(h == NULL || *loc == (uintptr_t)NULL);
507 atomic_store_rel_ptr(loc, (uintptr_t)h);
511 * Called on MOD_LOAD. Validates and calculates the SGE tunables.
517 if (fl_pktshift < 0 || fl_pktshift > 7) {
518 printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
519 " using 0 instead.\n", fl_pktshift);
523 if (spg_len != 64 && spg_len != 128) {
526 #if defined(__i386__) || defined(__amd64__)
527 len = cpu_clflush_line_size > 64 ? 128 : 64;
532 printf("Invalid hw.cxgbe.spg_len value (%d),"
533 " using %d instead.\n", spg_len, len);
538 if (cong_drop < -1 || cong_drop > 1) {
539 printf("Invalid hw.cxgbe.cong_drop value (%d),"
540 " using 0 instead.\n", cong_drop);
544 if (tscale != 1 && (tscale < 3 || tscale > 17)) {
545 printf("Invalid hw.cxgbe.tscale value (%d),"
546 " using 1 instead.\n", tscale);
550 extfree_refs = counter_u64_alloc(M_WAITOK);
551 extfree_rels = counter_u64_alloc(M_WAITOK);
552 counter_u64_zero(extfree_refs);
553 counter_u64_zero(extfree_rels);
555 t4_init_shared_cpl_handlers();
556 t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
557 t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
558 t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
559 t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
561 t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack,
564 t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
565 t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
569 t4_sge_modunload(void)
572 counter_u64_free(extfree_refs);
573 counter_u64_free(extfree_rels);
577 t4_sge_extfree_refs(void)
581 rels = counter_u64_fetch(extfree_rels);
582 refs = counter_u64_fetch(extfree_refs);
584 return (refs - rels);
588 setup_pad_and_pack_boundaries(struct adapter *sc)
591 int pad, pack, pad_shift;
593 pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
594 X_INGPADBOUNDARY_SHIFT;
596 if (fl_pad < (1 << pad_shift) ||
597 fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
600 * If there is any chance that we might use buffer packing and
601 * the chip is a T4, then pick 64 as the pad/pack boundary. Set
602 * it to the minimum allowed in all other cases.
604 pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
607 * For fl_pad = 0 we'll still write a reasonable value to the
608 * register but all the freelists will opt out of padding.
609 * We'll complain here only if the user tried to set it to a
610 * value greater than 0 that was invalid.
613 device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
614 " (%d), using %d instead.\n", fl_pad, pad);
617 m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
618 v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
619 t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
622 if (fl_pack != -1 && fl_pack != pad) {
623 /* Complain but carry on. */
624 device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
625 " using %d instead.\n", fl_pack, pad);
631 if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
632 !powerof2(fl_pack)) {
633 pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
634 MPASS(powerof2(pack));
642 device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
643 " (%d), using %d instead.\n", fl_pack, pack);
646 m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
648 v = V_INGPACKBOUNDARY(0);
650 v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
652 MPASS(!is_t4(sc)); /* T4 doesn't have SGE_CONTROL2 */
653 t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
657 * adap->params.vpd.cclk must be set up before this is called.
660 t4_tweak_chip_settings(struct adapter *sc)
664 int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
665 int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
666 int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
667 uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
668 static int sge_flbuf_sizes[] = {
670 #if MJUMPAGESIZE != MCLBYTES
672 MJUMPAGESIZE - CL_METADATA_SIZE,
673 MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
677 MCLBYTES - MSIZE - CL_METADATA_SIZE,
678 MJUM9BYTES - CL_METADATA_SIZE,
679 MJUM16BYTES - CL_METADATA_SIZE,
682 KASSERT(sc->flags & MASTER_PF,
683 ("%s: trying to change chip settings when not master.", __func__));
685 m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
686 v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
687 V_EGRSTATUSPAGESIZE(spg_len == 128);
688 t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
690 setup_pad_and_pack_boundaries(sc);
692 v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
693 V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
694 V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
695 V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
696 V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
697 V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
698 V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
699 V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
700 t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
702 KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
703 ("%s: hw buffer size table too big", __func__));
704 t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, 4096);
705 t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE1, 65536);
706 for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
707 t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE15 - (4 * i),
711 v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
712 V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
713 t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
715 KASSERT(intr_timer[0] <= timer_max,
716 ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
718 for (i = 1; i < nitems(intr_timer); i++) {
719 KASSERT(intr_timer[i] >= intr_timer[i - 1],
720 ("%s: timers not listed in increasing order (%d)",
723 while (intr_timer[i] > timer_max) {
724 if (i == nitems(intr_timer) - 1) {
725 intr_timer[i] = timer_max;
728 intr_timer[i] += intr_timer[i - 1];
733 v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
734 V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
735 t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
736 v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
737 V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
738 t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
739 v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
740 V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
741 t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
743 if (chip_id(sc) >= CHELSIO_T6) {
744 m = V_TSCALE(M_TSCALE);
748 v = V_TSCALE(tscale - 2);
749 t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
751 if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
752 m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
753 V_WRTHRTHRESH(M_WRTHRTHRESH);
754 t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
756 v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
758 t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
762 /* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
763 v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
764 t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
767 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP. These have been
768 * chosen with MAXPHYS = 128K in mind. The largest DDP buffer that we
769 * may have to deal with is MAXPHYS + 1 page.
771 v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
772 t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
774 /* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
775 m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
776 t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
778 m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
780 v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
781 t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
785 * SGE wants the buffer to be at least 64B and then a multiple of 16. If
786 * padding is in use, the buffer's start and end need to be aligned to the pad
787 * boundary as well. We'll just make sure that the size is a multiple of the
788 * boundary here, it is up to the buffer allocation code to make sure the start
789 * of the buffer is aligned as well.
792 hwsz_ok(struct adapter *sc, int hwsz)
794 int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
796 return (hwsz >= 64 && (hwsz & mask) == 0);
800 * XXX: driver really should be able to deal with unexpected settings.
803 t4_read_chip_settings(struct adapter *sc)
805 struct sge *s = &sc->sge;
806 struct sge_params *sp = &sc->params.sge;
809 uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
810 static int sw_buf_sizes[] = { /* Sorted by size */
812 #if MJUMPAGESIZE != MCLBYTES
818 struct sw_zone_info *swz, *safe_swz;
819 struct hw_buf_info *hwb;
823 r = sc->params.sge.sge_control;
825 device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
830 * If this changes then every single use of PAGE_SHIFT in the driver
831 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
833 if (sp->page_shift != PAGE_SHIFT) {
834 device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
838 /* Filter out unusable hw buffer sizes entirely (mark with -2). */
839 hwb = &s->hw_buf_info[0];
840 for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
841 r = sc->params.sge.sge_fl_buffer_size[i];
843 hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
848 * Create a sorted list in decreasing order of hw buffer sizes (and so
849 * increasing order of spare area) for each software zone.
851 * If padding is enabled then the start and end of the buffer must align
852 * to the pad boundary; if packing is enabled then they must align with
853 * the pack boundary as well. Allocations from the cluster zones are
854 * aligned to min(size, 4K), so the buffer starts at that alignment and
855 * ends at hwb->size alignment. If mbuf inlining is allowed the
856 * starting alignment will be reduced to MSIZE and the driver will
857 * exercise appropriate caution when deciding on the best buffer layout
860 n = 0; /* no usable buffer size to begin with */
861 swz = &s->sw_zone_info[0];
863 for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
864 int8_t head = -1, tail = -1;
866 swz->size = sw_buf_sizes[i];
867 swz->zone = m_getzone(swz->size);
868 swz->type = m_gettype(swz->size);
870 if (swz->size < PAGE_SIZE) {
871 MPASS(powerof2(swz->size));
872 if (fl_pad && (swz->size % sp->pad_boundary != 0))
876 if (swz->size == safest_rx_cluster)
879 hwb = &s->hw_buf_info[0];
880 for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
881 if (hwb->zidx != -1 || hwb->size > swz->size)
885 MPASS(hwb->size % sp->pad_boundary == 0);
890 else if (hwb->size < s->hw_buf_info[tail].size) {
891 s->hw_buf_info[tail].next = j;
895 struct hw_buf_info *t;
897 for (cur = &head; *cur != -1; cur = &t->next) {
898 t = &s->hw_buf_info[*cur];
899 if (hwb->size == t->size) {
903 if (hwb->size > t->size) {
911 swz->head_hwidx = head;
912 swz->tail_hwidx = tail;
916 if (swz->size - s->hw_buf_info[tail].size >=
918 sc->flags |= BUF_PACKING_OK;
922 device_printf(sc->dev, "no usable SGE FL buffer size.\n");
928 if (safe_swz != NULL) {
929 s->safe_hwidx1 = safe_swz->head_hwidx;
930 for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
933 hwb = &s->hw_buf_info[i];
936 MPASS(hwb->size % sp->pad_boundary == 0);
938 spare = safe_swz->size - hwb->size;
939 if (spare >= CL_METADATA_SIZE) {
946 if (sc->flags & IS_VF)
949 v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
950 r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
952 device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
956 m = v = F_TDDPTAGTCB;
957 r = t4_read_reg(sc, A_ULP_RX_CTL);
959 device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
963 m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
965 v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
966 r = t4_read_reg(sc, A_TP_PARA_REG5);
968 device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
972 t4_init_tp_params(sc, 1);
974 t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
975 t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
981 t4_create_dma_tag(struct adapter *sc)
985 rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
986 BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
987 BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
990 device_printf(sc->dev,
991 "failed to create main DMA tag: %d\n", rc);
998 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
999 struct sysctl_oid_list *children)
1001 struct sge_params *sp = &sc->params.sge;
1003 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
1004 CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
1005 "freelist buffer sizes");
1007 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
1008 NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
1010 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
1011 NULL, sp->pad_boundary, "payload pad boundary (bytes)");
1013 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
1014 NULL, sp->spg_len, "status page size (bytes)");
1016 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
1017 NULL, cong_drop, "congestion drop setting");
1019 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
1020 NULL, sp->pack_boundary, "payload pack boundary (bytes)");
1024 t4_destroy_dma_tag(struct adapter *sc)
1027 bus_dma_tag_destroy(sc->dmat);
1033 * Allocate and initialize the firmware event queue, control queues, and special
1034 * purpose rx queues owned by the adapter.
1036 * Returns errno on failure. Resources allocated up to that point may still be
1037 * allocated. Caller is responsible for cleanup in case this function fails.
1040 t4_setup_adapter_queues(struct adapter *sc)
1042 struct sysctl_oid *oid;
1043 struct sysctl_oid_list *children;
1046 ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1048 sysctl_ctx_init(&sc->ctx);
1049 sc->flags |= ADAP_SYSCTL_CTX;
1052 * Firmware event queue
1059 * That's all for the VF driver.
1061 if (sc->flags & IS_VF)
1064 oid = device_get_sysctl_tree(sc->dev);
1065 children = SYSCTL_CHILDREN(oid);
1068 * XXX: General purpose rx queues, one per port.
1072 * Control queues, one per port.
1074 oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "ctrlq",
1075 CTLFLAG_RD, NULL, "control queues");
1076 for_each_port(sc, i) {
1077 struct sge_wrq *ctrlq = &sc->sge.ctrlq[i];
1079 rc = alloc_ctrlq(sc, ctrlq, i, oid);
1091 t4_teardown_adapter_queues(struct adapter *sc)
1095 ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1097 /* Do this before freeing the queue */
1098 if (sc->flags & ADAP_SYSCTL_CTX) {
1099 sysctl_ctx_free(&sc->ctx);
1100 sc->flags &= ~ADAP_SYSCTL_CTX;
1103 if (!(sc->flags & IS_VF)) {
1104 for_each_port(sc, i)
1105 free_wrq(sc, &sc->sge.ctrlq[i]);
1112 /* Maximum payload that can be delivered with a single iq descriptor */
1114 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
1120 int rxcs = G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2));
1122 /* Note that COP can set rx_coalesce on/off per connection. */
1123 payload = max(mtu, rxcs);
1126 /* large enough even when hw VLAN extraction is disabled */
1127 payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1128 ETHER_VLAN_ENCAP_LEN + mtu;
1137 t4_setup_vi_queues(struct vi_info *vi)
1139 int rc = 0, i, intr_idx, iqidx;
1140 struct sge_rxq *rxq;
1141 struct sge_txq *txq;
1143 struct sge_ofld_rxq *ofld_rxq;
1145 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1146 struct sge_wrq *ofld_txq;
1150 struct sge_nm_rxq *nm_rxq;
1151 struct sge_nm_txq *nm_txq;
1154 struct port_info *pi = vi->pi;
1155 struct adapter *sc = pi->adapter;
1156 struct ifnet *ifp = vi->ifp;
1157 struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1158 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1159 int maxp, mtu = ifp->if_mtu;
1161 /* Interrupt vector to start from (when using multiple vectors) */
1162 intr_idx = vi->first_intr;
1165 saved_idx = intr_idx;
1166 if (ifp->if_capabilities & IFCAP_NETMAP) {
1168 /* netmap is supported with direct interrupts only. */
1169 MPASS(!forwarding_intr_to_fwq(sc));
1172 * We don't have buffers to back the netmap rx queues
1173 * right now so we create the queues in a way that
1174 * doesn't set off any congestion signal in the chip.
1176 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1177 CTLFLAG_RD, NULL, "rx queues");
1178 for_each_nm_rxq(vi, i, nm_rxq) {
1179 rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1185 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1186 CTLFLAG_RD, NULL, "tx queues");
1187 for_each_nm_txq(vi, i, nm_txq) {
1188 iqidx = vi->first_nm_rxq + (i % vi->nnmrxq);
1189 rc = alloc_nm_txq(vi, nm_txq, iqidx, i, oid);
1195 /* Normal rx queues and netmap rx queues share the same interrupts. */
1196 intr_idx = saved_idx;
1200 * Allocate rx queues first because a default iqid is required when
1201 * creating a tx queue.
1203 maxp = mtu_to_max_payload(sc, mtu, 0);
1204 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1205 CTLFLAG_RD, NULL, "rx queues");
1206 for_each_rxq(vi, i, rxq) {
1208 init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1210 snprintf(name, sizeof(name), "%s rxq%d-fl",
1211 device_get_nameunit(vi->dev), i);
1212 init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1214 rc = alloc_rxq(vi, rxq,
1215 forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1221 if (ifp->if_capabilities & IFCAP_NETMAP)
1222 intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1225 maxp = mtu_to_max_payload(sc, mtu, 1);
1226 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1227 CTLFLAG_RD, NULL, "rx queues for offloaded TCP connections");
1228 for_each_ofld_rxq(vi, i, ofld_rxq) {
1230 init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
1233 snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1234 device_get_nameunit(vi->dev), i);
1235 init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1237 rc = alloc_ofld_rxq(vi, ofld_rxq,
1238 forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1246 * Now the tx queues.
1248 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1250 for_each_txq(vi, i, txq) {
1251 iqidx = vi->first_rxq + (i % vi->nrxq);
1252 snprintf(name, sizeof(name), "%s txq%d",
1253 device_get_nameunit(vi->dev), i);
1254 init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
1255 sc->sge.rxq[iqidx].iq.cntxt_id, name);
1257 rc = alloc_txq(vi, txq, i, oid);
1261 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1262 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1263 CTLFLAG_RD, NULL, "tx queues for TOE/ETHOFLD");
1264 for_each_ofld_txq(vi, i, ofld_txq) {
1265 struct sysctl_oid *oid2;
1267 snprintf(name, sizeof(name), "%s ofld_txq%d",
1268 device_get_nameunit(vi->dev), i);
1269 if (vi->nofldrxq > 0) {
1270 iqidx = vi->first_ofld_rxq + (i % vi->nofldrxq);
1271 init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1272 pi->tx_chan, sc->sge.ofld_rxq[iqidx].iq.cntxt_id,
1275 iqidx = vi->first_rxq + (i % vi->nrxq);
1276 init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1277 pi->tx_chan, sc->sge.rxq[iqidx].iq.cntxt_id, name);
1280 snprintf(name, sizeof(name), "%d", i);
1281 oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1282 name, CTLFLAG_RD, NULL, "offload tx queue");
1284 rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1291 t4_teardown_vi_queues(vi);
1300 t4_teardown_vi_queues(struct vi_info *vi)
1303 struct sge_rxq *rxq;
1304 struct sge_txq *txq;
1305 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1306 struct port_info *pi = vi->pi;
1307 struct adapter *sc = pi->adapter;
1308 struct sge_wrq *ofld_txq;
1311 struct sge_ofld_rxq *ofld_rxq;
1314 struct sge_nm_rxq *nm_rxq;
1315 struct sge_nm_txq *nm_txq;
1318 /* Do this before freeing the queues */
1319 if (vi->flags & VI_SYSCTL_CTX) {
1320 sysctl_ctx_free(&vi->ctx);
1321 vi->flags &= ~VI_SYSCTL_CTX;
1325 if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1326 for_each_nm_txq(vi, i, nm_txq) {
1327 free_nm_txq(vi, nm_txq);
1330 for_each_nm_rxq(vi, i, nm_rxq) {
1331 free_nm_rxq(vi, nm_rxq);
1337 * Take down all the tx queues first, as they reference the rx queues
1338 * (for egress updates, etc.).
1341 for_each_txq(vi, i, txq) {
1344 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1345 for_each_ofld_txq(vi, i, ofld_txq) {
1346 free_wrq(sc, ofld_txq);
1351 * Then take down the rx queues.
1354 for_each_rxq(vi, i, rxq) {
1358 for_each_ofld_rxq(vi, i, ofld_rxq) {
1359 free_ofld_rxq(vi, ofld_rxq);
1367 * Interrupt handler when the driver is using only 1 interrupt. This is a very
1370 * a) Deals with errors, if any.
1371 * b) Services firmware event queue, which is taking interrupts for all other
1375 t4_intr_all(void *arg)
1377 struct adapter *sc = arg;
1378 struct sge_iq *fwq = &sc->sge.fwq;
1380 MPASS(sc->intr_count == 1);
1382 if (sc->intr_type == INTR_INTX)
1383 t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1390 * Interrupt handler for errors (installed directly when multiple interrupts are
1391 * being used, or called by t4_intr_all).
1394 t4_intr_err(void *arg)
1396 struct adapter *sc = arg;
1398 const bool verbose = (sc->debug_flags & DF_VERBOSE_SLOWINTR) != 0;
1400 if (sc->flags & ADAP_ERR)
1403 v = t4_read_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE));
1406 t4_write_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE), v);
1409 t4_slow_intr_handler(sc, verbose);
1413 * Interrupt handler for iq-only queues. The firmware event queue is the only
1414 * such queue right now.
1417 t4_intr_evt(void *arg)
1419 struct sge_iq *iq = arg;
1421 if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1423 (void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1428 * Interrupt handler for iq+fl queues.
1433 struct sge_iq *iq = arg;
1435 if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1436 service_iq_fl(iq, 0);
1437 (void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1443 * Interrupt handler for netmap rx queues.
1446 t4_nm_intr(void *arg)
1448 struct sge_nm_rxq *nm_rxq = arg;
1450 if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1451 service_nm_rxq(nm_rxq);
1452 (void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1457 * Interrupt handler for vectors shared between NIC and netmap rx queues.
1460 t4_vi_intr(void *arg)
1462 struct irq *irq = arg;
1464 MPASS(irq->nm_rxq != NULL);
1465 t4_nm_intr(irq->nm_rxq);
1467 MPASS(irq->rxq != NULL);
1473 * Deals with interrupts on an iq-only (no freelist) queue.
1476 service_iq(struct sge_iq *iq, int budget)
1479 struct adapter *sc = iq->adapter;
1480 struct iq_desc *d = &iq->desc[iq->cidx];
1481 int ndescs = 0, limit;
1484 STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1486 KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1487 KASSERT((iq->flags & IQ_HAS_FL) == 0,
1488 ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1490 MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1491 MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1493 limit = budget ? budget : iq->qsize / 16;
1496 * We always come back and check the descriptor ring for new indirect
1497 * interrupts and other responses after running a single handler.
1500 while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1504 rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1505 lq = be32toh(d->rsp.pldbuflen_qid);
1508 case X_RSPD_TYPE_FLBUF:
1509 panic("%s: data for an iq (%p) with no freelist",
1514 case X_RSPD_TYPE_CPL:
1515 KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1516 ("%s: bad opcode %02x.", __func__,
1518 t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1521 case X_RSPD_TYPE_INTR:
1523 * There are 1K interrupt-capable queues (qids 0
1524 * through 1023). A response type indicating a
1525 * forwarded interrupt with a qid >= 1K is an
1526 * iWARP async notification.
1528 if (__predict_true(lq >= 1024)) {
1529 t4_an_handler(iq, &d->rsp);
1533 q = sc->sge.iqmap[lq - sc->sge.iq_start -
1535 if (atomic_cmpset_int(&q->state, IQS_IDLE,
1537 if (service_iq_fl(q, q->qsize / 16) == 0) {
1538 (void) atomic_cmpset_int(&q->state,
1539 IQS_BUSY, IQS_IDLE);
1541 STAILQ_INSERT_TAIL(&iql, q,
1549 ("%s: illegal response type %d on iq %p",
1550 __func__, rsp_type, iq));
1552 "%s: illegal response type %d on iq %p",
1553 device_get_nameunit(sc->dev), rsp_type, iq);
1558 if (__predict_false(++iq->cidx == iq->sidx)) {
1560 iq->gen ^= F_RSPD_GEN;
1563 if (__predict_false(++ndescs == limit)) {
1564 t4_write_reg(sc, sc->sge_gts_reg,
1566 V_INGRESSQID(iq->cntxt_id) |
1567 V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1571 return (EINPROGRESS);
1576 if (STAILQ_EMPTY(&iql))
1580 * Process the head only, and send it to the back of the list if
1581 * it's still not done.
1583 q = STAILQ_FIRST(&iql);
1584 STAILQ_REMOVE_HEAD(&iql, link);
1585 if (service_iq_fl(q, q->qsize / 8) == 0)
1586 (void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1588 STAILQ_INSERT_TAIL(&iql, q, link);
1591 t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1592 V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1598 sort_before_lro(struct lro_ctrl *lro)
1601 return (lro->lro_mbuf_max != 0);
1604 static inline uint64_t
1605 last_flit_to_ns(struct adapter *sc, uint64_t lf)
1607 uint64_t n = be64toh(lf) & 0xfffffffffffffff; /* 60b, not 64b. */
1609 if (n > UINT64_MAX / 1000000)
1610 return (n / sc->params.vpd.cclk * 1000000);
1612 return (n * 1000000 / sc->params.vpd.cclk);
1616 * Deals with interrupts on an iq+fl queue.
1619 service_iq_fl(struct sge_iq *iq, int budget)
1621 struct sge_rxq *rxq = iq_to_rxq(iq);
1623 struct adapter *sc = iq->adapter;
1624 struct iq_desc *d = &iq->desc[iq->cidx];
1625 int ndescs = 0, limit;
1626 int rsp_type, refill, starved;
1628 uint16_t fl_hw_cidx;
1630 #if defined(INET) || defined(INET6)
1631 const struct timeval lro_timeout = {0, sc->lro_timeout};
1632 struct lro_ctrl *lro = &rxq->lro;
1635 KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1636 MPASS(iq->flags & IQ_HAS_FL);
1638 limit = budget ? budget : iq->qsize / 16;
1640 fl_hw_cidx = fl->hw_cidx; /* stable snapshot */
1642 #if defined(INET) || defined(INET6)
1643 if (iq->flags & IQ_ADJ_CREDIT) {
1644 MPASS(sort_before_lro(lro));
1645 iq->flags &= ~IQ_ADJ_CREDIT;
1646 if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1647 tcp_lro_flush_all(lro);
1648 t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1649 V_INGRESSQID((u32)iq->cntxt_id) |
1650 V_SEINTARM(iq->intr_params));
1656 MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1659 while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1665 rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1666 lq = be32toh(d->rsp.pldbuflen_qid);
1669 case X_RSPD_TYPE_FLBUF:
1671 m0 = get_fl_payload(sc, fl, lq);
1672 if (__predict_false(m0 == NULL))
1674 refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1676 if (iq->flags & IQ_RX_TIMESTAMP) {
1678 * Fill up rcv_tstmp but do not set M_TSTMP.
1679 * rcv_tstmp is not in the format that the
1680 * kernel expects and we don't want to mislead
1681 * it. For now this is only for custom code
1682 * that knows how to interpret cxgbe's stamp.
1684 m0->m_pkthdr.rcv_tstmp =
1685 last_flit_to_ns(sc, d->rsp.u.last_flit);
1687 m0->m_flags |= M_TSTMP;
1693 case X_RSPD_TYPE_CPL:
1694 KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1695 ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1696 t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1699 case X_RSPD_TYPE_INTR:
1702 * There are 1K interrupt-capable queues (qids 0
1703 * through 1023). A response type indicating a
1704 * forwarded interrupt with a qid >= 1K is an
1705 * iWARP async notification. That is the only
1706 * acceptable indirect interrupt on this queue.
1708 if (__predict_false(lq < 1024)) {
1709 panic("%s: indirect interrupt on iq_fl %p "
1710 "with qid %u", __func__, iq, lq);
1713 t4_an_handler(iq, &d->rsp);
1717 KASSERT(0, ("%s: illegal response type %d on iq %p",
1718 __func__, rsp_type, iq));
1719 log(LOG_ERR, "%s: illegal response type %d on iq %p",
1720 device_get_nameunit(sc->dev), rsp_type, iq);
1725 if (__predict_false(++iq->cidx == iq->sidx)) {
1727 iq->gen ^= F_RSPD_GEN;
1730 if (__predict_false(++ndescs == limit)) {
1731 t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1732 V_INGRESSQID(iq->cntxt_id) |
1733 V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1736 #if defined(INET) || defined(INET6)
1737 if (iq->flags & IQ_LRO_ENABLED &&
1738 !sort_before_lro(lro) &&
1739 sc->lro_timeout != 0) {
1740 tcp_lro_flush_inactive(lro, &lro_timeout);
1745 refill_fl(sc, fl, 32);
1748 return (EINPROGRESS);
1753 refill_fl(sc, fl, 32);
1755 fl_hw_cidx = fl->hw_cidx;
1759 #if defined(INET) || defined(INET6)
1760 if (iq->flags & IQ_LRO_ENABLED) {
1761 if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1762 MPASS(sort_before_lro(lro));
1763 /* hold back one credit and don't flush LRO state */
1764 iq->flags |= IQ_ADJ_CREDIT;
1767 tcp_lro_flush_all(lro);
1772 t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1773 V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1776 starved = refill_fl(sc, fl, 64);
1778 if (__predict_false(starved != 0))
1779 add_fl_to_sfl(sc, fl);
1785 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1787 int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1790 MPASS(cll->region3 >= CL_METADATA_SIZE);
1795 static inline struct cluster_metadata *
1796 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1800 if (cl_has_metadata(fl, cll)) {
1801 struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1803 return ((struct cluster_metadata *)(cl + swz->size) - 1);
1809 rxb_free(struct mbuf *m)
1811 uma_zone_t zone = m->m_ext.ext_arg1;
1812 void *cl = m->m_ext.ext_arg2;
1814 uma_zfree(zone, cl);
1815 counter_u64_add(extfree_rels, 1);
1819 * The mbuf returned by this function could be allocated from zone_mbuf or
1820 * constructed in spare room in the cluster.
1822 * The mbuf carries the payload in one of these ways
1823 * a) frame inside the mbuf (mbuf from zone_mbuf)
1824 * b) m_cljset (for clusters without metadata) zone_mbuf
1825 * c) m_extaddref (cluster with metadata) inline mbuf
1826 * d) m_extaddref (cluster with metadata) zone_mbuf
1828 static struct mbuf *
1829 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1833 struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1834 struct cluster_layout *cll = &sd->cll;
1835 struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1836 struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1837 struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1841 blen = hwb->size - fl->rx_offset; /* max possible in this buf */
1842 len = min(remaining, blen);
1843 payload = sd->cl + cll->region1 + fl->rx_offset;
1844 if (fl->flags & FL_BUF_PACKING) {
1845 const u_int l = fr_offset + len;
1846 const u_int pad = roundup2(l, fl->buf_boundary) - l;
1848 if (fl->rx_offset + len + pad < hwb->size)
1850 MPASS(fl->rx_offset + blen <= hwb->size);
1852 MPASS(fl->rx_offset == 0); /* not packing */
1856 if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1859 * Copy payload into a freshly allocated mbuf.
1862 m = fr_offset == 0 ?
1863 m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1866 fl->mbuf_allocated++;
1868 /* copy data to mbuf */
1869 bcopy(payload, mtod(m, caddr_t), len);
1871 } else if (sd->nmbuf * MSIZE < cll->region1) {
1874 * There's spare room in the cluster for an mbuf. Create one
1875 * and associate it with the payload that's in the cluster.
1879 m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1880 /* No bzero required */
1881 if (m_init(m, M_NOWAIT, MT_DATA,
1882 fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1885 m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1887 if (sd->nmbuf++ == 0)
1888 counter_u64_add(extfree_refs, 1);
1893 * Grab an mbuf from zone_mbuf and associate it with the
1894 * payload in the cluster.
1897 m = fr_offset == 0 ?
1898 m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1901 fl->mbuf_allocated++;
1903 m_extaddref(m, payload, blen, &clm->refcount,
1904 rxb_free, swz->zone, sd->cl);
1905 if (sd->nmbuf++ == 0)
1906 counter_u64_add(extfree_refs, 1);
1908 m_cljset(m, sd->cl, swz->type);
1909 sd->cl = NULL; /* consumed, not a recycle candidate */
1913 m->m_pkthdr.len = remaining;
1916 if (fl->flags & FL_BUF_PACKING) {
1917 fl->rx_offset += blen;
1918 MPASS(fl->rx_offset <= hwb->size);
1919 if (fl->rx_offset < hwb->size)
1920 return (m); /* without advancing the cidx */
1923 if (__predict_false(++fl->cidx % 8 == 0)) {
1924 uint16_t cidx = fl->cidx / 8;
1926 if (__predict_false(cidx == fl->sidx))
1927 fl->cidx = cidx = 0;
1935 static struct mbuf *
1936 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1938 struct mbuf *m0, *m, **pnext;
1940 const u_int total = G_RSPD_LEN(len_newbuf);
1942 if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1943 M_ASSERTPKTHDR(fl->m0);
1944 MPASS(fl->m0->m_pkthdr.len == total);
1945 MPASS(fl->remaining < total);
1949 remaining = fl->remaining;
1950 fl->flags &= ~FL_BUF_RESUME;
1954 if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1956 if (__predict_false(++fl->cidx % 8 == 0)) {
1957 uint16_t cidx = fl->cidx / 8;
1959 if (__predict_false(cidx == fl->sidx))
1960 fl->cidx = cidx = 0;
1966 * Payload starts at rx_offset in the current hw buffer. Its length is
1967 * 'len' and it may span multiple hw buffers.
1970 m0 = get_scatter_segment(sc, fl, 0, total);
1973 remaining = total - m0->m_len;
1974 pnext = &m0->m_next;
1975 while (remaining > 0) {
1977 MPASS(fl->rx_offset == 0);
1978 m = get_scatter_segment(sc, fl, total - remaining, remaining);
1979 if (__predict_false(m == NULL)) {
1982 fl->remaining = remaining;
1983 fl->flags |= FL_BUF_RESUME;
1988 remaining -= m->m_len;
1997 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1999 struct sge_rxq *rxq = iq_to_rxq(iq);
2000 struct ifnet *ifp = rxq->ifp;
2001 struct adapter *sc = iq->adapter;
2002 const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
2003 #if defined(INET) || defined(INET6)
2004 struct lro_ctrl *lro = &rxq->lro;
2006 static const int sw_hashtype[4][2] = {
2007 {M_HASHTYPE_NONE, M_HASHTYPE_NONE},
2008 {M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
2009 {M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
2010 {M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
2013 KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
2016 m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
2017 m0->m_len -= sc->params.sge.fl_pktshift;
2018 m0->m_data += sc->params.sge.fl_pktshift;
2020 m0->m_pkthdr.rcvif = ifp;
2021 M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
2022 m0->m_pkthdr.flowid = be32toh(rss->hash_val);
2024 if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
2025 if (ifp->if_capenable & IFCAP_RXCSUM &&
2026 cpl->l2info & htobe32(F_RXF_IP)) {
2027 m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
2028 CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
2030 } else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
2031 cpl->l2info & htobe32(F_RXF_IP6)) {
2032 m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
2037 if (__predict_false(cpl->ip_frag))
2038 m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2040 m0->m_pkthdr.csum_data = 0xffff;
2044 m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2045 m0->m_flags |= M_VLANTAG;
2046 rxq->vlan_extraction++;
2050 m0->m_pkthdr.numa_domain = ifp->if_numa_domain;
2052 #if defined(INET) || defined(INET6)
2053 if (iq->flags & IQ_LRO_ENABLED) {
2054 if (sort_before_lro(lro)) {
2055 tcp_lro_queue_mbuf(lro, m0);
2056 return (0); /* queued for sort, then LRO */
2058 if (tcp_lro_rx(lro, m0, 0) == 0)
2059 return (0); /* queued for LRO */
2062 ifp->if_input(ifp, m0);
2068 * Must drain the wrq or make sure that someone else will.
2071 wrq_tx_drain(void *arg, int n)
2073 struct sge_wrq *wrq = arg;
2074 struct sge_eq *eq = &wrq->eq;
2077 if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2078 drain_wrq_wr_list(wrq->adapter, wrq);
2083 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2085 struct sge_eq *eq = &wrq->eq;
2086 u_int available, dbdiff; /* # of hardware descriptors */
2089 struct fw_eth_tx_pkt_wr *dst; /* any fw WR struct will do */
2091 EQ_LOCK_ASSERT_OWNED(eq);
2092 MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2093 wr = STAILQ_FIRST(&wrq->wr_list);
2094 MPASS(wr != NULL); /* Must be called with something useful to do */
2095 MPASS(eq->pidx == eq->dbidx);
2099 eq->cidx = read_hw_cidx(eq);
2100 if (eq->pidx == eq->cidx)
2101 available = eq->sidx - 1;
2103 available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2105 MPASS(wr->wrq == wrq);
2106 n = howmany(wr->wr_len, EQ_ESIZE);
2110 dst = (void *)&eq->desc[eq->pidx];
2111 if (__predict_true(eq->sidx - eq->pidx > n)) {
2112 /* Won't wrap, won't end exactly at the status page. */
2113 bcopy(&wr->wr[0], dst, wr->wr_len);
2116 int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2118 bcopy(&wr->wr[0], dst, first_portion);
2119 if (wr->wr_len > first_portion) {
2120 bcopy(&wr->wr[first_portion], &eq->desc[0],
2121 wr->wr_len - first_portion);
2123 eq->pidx = n - (eq->sidx - eq->pidx);
2125 wrq->tx_wrs_copied++;
2127 if (available < eq->sidx / 4 &&
2128 atomic_cmpset_int(&eq->equiq, 0, 1)) {
2130 * XXX: This is not 100% reliable with some
2131 * types of WRs. But this is a very unusual
2132 * situation for an ofld/ctrl queue anyway.
2134 dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2140 ring_eq_db(sc, eq, dbdiff);
2144 STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2146 MPASS(wrq->nwr_pending > 0);
2148 MPASS(wrq->ndesc_needed >= n);
2149 wrq->ndesc_needed -= n;
2150 } while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2153 ring_eq_db(sc, eq, dbdiff);
2157 * Doesn't fail. Holds on to work requests it can't send right away.
2160 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2163 struct sge_eq *eq = &wrq->eq;
2166 EQ_LOCK_ASSERT_OWNED(eq);
2168 MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2169 MPASS((wr->wr_len & 0x7) == 0);
2171 STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2173 wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2175 if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2176 return; /* commit_wrq_wr will drain wr_list as well. */
2178 drain_wrq_wr_list(sc, wrq);
2180 /* Doorbell must have caught up to the pidx. */
2181 MPASS(eq->pidx == eq->dbidx);
2185 t4_update_fl_bufsize(struct ifnet *ifp)
2187 struct vi_info *vi = ifp->if_softc;
2188 struct adapter *sc = vi->pi->adapter;
2189 struct sge_rxq *rxq;
2191 struct sge_ofld_rxq *ofld_rxq;
2194 int i, maxp, mtu = ifp->if_mtu;
2196 maxp = mtu_to_max_payload(sc, mtu, 0);
2197 for_each_rxq(vi, i, rxq) {
2201 find_best_refill_source(sc, fl, maxp);
2205 maxp = mtu_to_max_payload(sc, mtu, 1);
2206 for_each_ofld_rxq(vi, i, ofld_rxq) {
2210 find_best_refill_source(sc, fl, maxp);
2217 mbuf_nsegs(struct mbuf *m)
2221 KASSERT(m->m_pkthdr.l5hlen > 0,
2222 ("%s: mbuf %p missing information on # of segments.", __func__, m));
2224 return (m->m_pkthdr.l5hlen);
2228 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2232 m->m_pkthdr.l5hlen = nsegs;
2236 mbuf_cflags(struct mbuf *m)
2240 return (m->m_pkthdr.PH_loc.eight[4]);
2244 set_mbuf_cflags(struct mbuf *m, uint8_t flags)
2248 m->m_pkthdr.PH_loc.eight[4] = flags;
2252 mbuf_len16(struct mbuf *m)
2257 n = m->m_pkthdr.PH_loc.eight[0];
2258 MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2264 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2268 m->m_pkthdr.PH_loc.eight[0] = len16;
2273 mbuf_eo_nsegs(struct mbuf *m)
2277 return (m->m_pkthdr.PH_loc.eight[1]);
2281 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2285 m->m_pkthdr.PH_loc.eight[1] = nsegs;
2289 mbuf_eo_len16(struct mbuf *m)
2294 n = m->m_pkthdr.PH_loc.eight[2];
2295 MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2301 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2305 m->m_pkthdr.PH_loc.eight[2] = len16;
2309 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2313 return (m->m_pkthdr.PH_loc.eight[3]);
2317 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2321 m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2325 needs_eo(struct mbuf *m)
2328 return (m->m_pkthdr.csum_flags & CSUM_SND_TAG);
2333 * Try to allocate an mbuf to contain a raw work request. To make it
2334 * easy to construct the work request, don't allocate a chain but a
2338 alloc_wr_mbuf(int len, int how)
2343 m = m_gethdr(how, MT_DATA);
2344 else if (len <= MCLBYTES)
2345 m = m_getcl(how, MT_DATA, M_PKTHDR);
2350 m->m_pkthdr.len = len;
2352 set_mbuf_cflags(m, MC_RAW_WR);
2353 set_mbuf_len16(m, howmany(len, 16));
2358 needs_tso(struct mbuf *m)
2363 return (m->m_pkthdr.csum_flags & CSUM_TSO);
2367 needs_l3_csum(struct mbuf *m)
2372 return (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO));
2376 needs_l4_csum(struct mbuf *m)
2381 return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2382 CSUM_TCP_IPV6 | CSUM_TSO));
2386 needs_tcp_csum(struct mbuf *m)
2390 return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_TCP_IPV6 | CSUM_TSO));
2395 needs_udp_csum(struct mbuf *m)
2399 return (m->m_pkthdr.csum_flags & (CSUM_UDP | CSUM_UDP_IPV6));
2404 needs_vlan_insertion(struct mbuf *m)
2409 return (m->m_flags & M_VLANTAG);
2413 m_advance(struct mbuf **pm, int *poffset, int len)
2415 struct mbuf *m = *pm;
2416 int offset = *poffset;
2422 if (offset + len < m->m_len) {
2424 p = mtod(m, uintptr_t) + offset;
2427 len -= m->m_len - offset;
2438 * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2439 * must have at least one mbuf that's not empty. It is possible for this
2440 * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2443 count_mbuf_nsegs(struct mbuf *m, int skip)
2445 vm_paddr_t lastb, next;
2450 MPASS(m->m_pkthdr.len > 0);
2451 MPASS(m->m_pkthdr.len >= skip);
2455 for (; m; m = m->m_next) {
2458 if (__predict_false(len == 0))
2464 va = mtod(m, vm_offset_t) + skip;
2467 next = pmap_kextract(va);
2468 nsegs += sglist_count((void *)(uintptr_t)va, len);
2469 if (lastb + 1 == next)
2471 lastb = pmap_kextract(va + len - 1);
2478 * Analyze the mbuf to determine its tx needs. The mbuf passed in may change:
2479 * a) caller can assume it's been freed if this function returns with an error.
2480 * b) it may get defragged up if the gather list is too long for the hardware.
2483 parse_pkt(struct adapter *sc, struct mbuf **mp)
2485 struct mbuf *m0 = *mp, *m;
2486 int rc, nsegs, defragged = 0, offset;
2487 struct ether_header *eh;
2489 #if defined(INET) || defined(INET6)
2495 if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2504 * First count the number of gather list segments in the payload.
2505 * Defrag the mbuf if nsegs exceeds the hardware limit.
2508 MPASS(m0->m_pkthdr.len > 0);
2509 nsegs = count_mbuf_nsegs(m0, 0);
2510 if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2511 if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2515 *mp = m0 = m; /* update caller's copy after defrag */
2519 if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2520 m0 = m_pullup(m0, m0->m_pkthdr.len);
2522 /* Should have left well enough alone. */
2526 *mp = m0; /* update caller's copy after pullup */
2529 set_mbuf_nsegs(m0, nsegs);
2530 set_mbuf_cflags(m0, 0);
2531 if (sc->flags & IS_VF)
2532 set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2534 set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2538 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2539 * checksumming is enabled. needs_l4_csum happens to check for all the
2542 if (__predict_false(needs_eo(m0) && !needs_l4_csum(m0))) {
2543 m_snd_tag_rele(m0->m_pkthdr.snd_tag);
2544 m0->m_pkthdr.snd_tag = NULL;
2545 m0->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
2549 if (!needs_tso(m0) &&
2553 !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2557 eh = mtod(m, struct ether_header *);
2558 eh_type = ntohs(eh->ether_type);
2559 if (eh_type == ETHERTYPE_VLAN) {
2560 struct ether_vlan_header *evh = (void *)eh;
2562 eh_type = ntohs(evh->evl_proto);
2563 m0->m_pkthdr.l2hlen = sizeof(*evh);
2565 m0->m_pkthdr.l2hlen = sizeof(*eh);
2568 l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2572 case ETHERTYPE_IPV6:
2574 struct ip6_hdr *ip6 = l3hdr;
2576 MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2578 m0->m_pkthdr.l3hlen = sizeof(*ip6);
2585 struct ip *ip = l3hdr;
2587 m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2592 panic("%s: ethertype 0x%04x unknown. if_cxgbe must be compiled"
2593 " with the same INET/INET6 options as the kernel.",
2597 #if defined(INET) || defined(INET6)
2598 if (needs_tcp_csum(m0)) {
2599 tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2600 m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2602 if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2603 set_mbuf_eo_tsclk_tsoff(m0,
2604 V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2605 V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2607 set_mbuf_eo_tsclk_tsoff(m0, 0);
2608 } else if (needs_udp_csum(m)) {
2609 m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2616 /* EO WRs have the headers in the WR and not the GL. */
2617 immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2618 m0->m_pkthdr.l4hlen;
2619 nsegs = count_mbuf_nsegs(m0, immhdrs);
2620 set_mbuf_eo_nsegs(m0, nsegs);
2621 set_mbuf_eo_len16(m0,
2622 txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2631 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2633 struct sge_eq *eq = &wrq->eq;
2634 struct adapter *sc = wrq->adapter;
2635 int ndesc, available;
2640 ndesc = howmany(len16, EQ_ESIZE / 16);
2641 MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2645 if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2646 drain_wrq_wr_list(sc, wrq);
2648 if (!STAILQ_EMPTY(&wrq->wr_list)) {
2651 wr = alloc_wrqe(len16 * 16, wrq);
2652 if (__predict_false(wr == NULL))
2655 cookie->ndesc = ndesc;
2659 eq->cidx = read_hw_cidx(eq);
2660 if (eq->pidx == eq->cidx)
2661 available = eq->sidx - 1;
2663 available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2664 if (available < ndesc)
2667 cookie->pidx = eq->pidx;
2668 cookie->ndesc = ndesc;
2669 TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2671 w = &eq->desc[eq->pidx];
2672 IDXINCR(eq->pidx, ndesc, eq->sidx);
2673 if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2675 wrq->ss_pidx = cookie->pidx;
2676 wrq->ss_len = len16 * 16;
2685 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2687 struct sge_eq *eq = &wrq->eq;
2688 struct adapter *sc = wrq->adapter;
2690 struct wrq_cookie *prev, *next;
2692 if (cookie->pidx == -1) {
2693 struct wrqe *wr = __containerof(w, struct wrqe, wr);
2699 if (__predict_false(w == &wrq->ss[0])) {
2700 int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2702 MPASS(wrq->ss_len > n); /* WR had better wrap around. */
2703 bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2704 bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2707 wrq->tx_wrs_direct++;
2710 ndesc = cookie->ndesc; /* Can be more than SGE_MAX_WR_NDESC here. */
2711 pidx = cookie->pidx;
2712 MPASS(pidx >= 0 && pidx < eq->sidx);
2713 prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2714 next = TAILQ_NEXT(cookie, link);
2716 MPASS(pidx == eq->dbidx);
2717 if (next == NULL || ndesc >= 16) {
2719 struct fw_eth_tx_pkt_wr *dst; /* any fw WR struct will do */
2722 * Note that the WR via which we'll request tx updates
2723 * is at pidx and not eq->pidx, which has moved on
2726 dst = (void *)&eq->desc[pidx];
2727 available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2728 if (available < eq->sidx / 4 &&
2729 atomic_cmpset_int(&eq->equiq, 0, 1)) {
2731 * XXX: This is not 100% reliable with some
2732 * types of WRs. But this is a very unusual
2733 * situation for an ofld/ctrl queue anyway.
2735 dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2739 ring_eq_db(wrq->adapter, eq, ndesc);
2741 MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2743 next->ndesc += ndesc;
2746 MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2747 prev->ndesc += ndesc;
2749 TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2751 if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2752 drain_wrq_wr_list(sc, wrq);
2755 if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2756 /* Doorbell must have caught up to the pidx. */
2757 MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2764 can_resume_eth_tx(struct mp_ring *r)
2766 struct sge_eq *eq = r->cookie;
2768 return (total_available_tx_desc(eq) > eq->sidx / 8);
2772 cannot_use_txpkts(struct mbuf *m)
2774 /* maybe put a GL limit too, to avoid silliness? */
2776 return (needs_tso(m) || (mbuf_cflags(m) & MC_RAW_WR) != 0);
2780 discard_tx(struct sge_eq *eq)
2783 return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2787 wr_can_update_eq(struct fw_eth_tx_pkts_wr *wr)
2790 switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
2792 case FW_ETH_TX_PKT_WR:
2793 case FW_ETH_TX_PKTS_WR:
2794 case FW_ETH_TX_PKT_VM_WR:
2802 * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2803 * be consumed. Return the actual number consumed. 0 indicates a stall.
2806 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2808 struct sge_txq *txq = r->cookie;
2809 struct sge_eq *eq = &txq->eq;
2810 struct ifnet *ifp = txq->ifp;
2811 struct vi_info *vi = ifp->if_softc;
2812 struct port_info *pi = vi->pi;
2813 struct adapter *sc = pi->adapter;
2814 u_int total, remaining; /* # of packets */
2815 u_int available, dbdiff; /* # of hardware descriptors */
2817 struct mbuf *m0, *tail;
2819 struct fw_eth_tx_pkts_wr *wr; /* any fw WR struct will do */
2821 remaining = IDXDIFF(pidx, cidx, r->size);
2822 MPASS(remaining > 0); /* Must not be called without work to do. */
2826 if (__predict_false(discard_tx(eq))) {
2827 while (cidx != pidx) {
2828 m0 = r->items[cidx];
2830 if (++cidx == r->size)
2833 reclaim_tx_descs(txq, 2048);
2838 /* How many hardware descriptors do we have readily available. */
2839 if (eq->pidx == eq->cidx)
2840 available = eq->sidx - 1;
2842 available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2843 dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2845 while (remaining > 0) {
2847 m0 = r->items[cidx];
2849 MPASS(m0->m_nextpkt == NULL);
2851 if (available < SGE_MAX_WR_NDESC) {
2852 available += reclaim_tx_descs(txq, 64);
2853 if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2854 break; /* out of descriptors */
2857 next_cidx = cidx + 1;
2858 if (__predict_false(next_cidx == r->size))
2861 wr = (void *)&eq->desc[eq->pidx];
2862 if (sc->flags & IS_VF) {
2865 ETHER_BPF_MTAP(ifp, m0);
2866 n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2868 } else if (remaining > 1 &&
2869 try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2871 /* pkts at cidx, next_cidx should both be in txp. */
2872 MPASS(txp.npkt == 2);
2873 tail = r->items[next_cidx];
2874 MPASS(tail->m_nextpkt == NULL);
2875 ETHER_BPF_MTAP(ifp, m0);
2876 ETHER_BPF_MTAP(ifp, tail);
2877 m0->m_nextpkt = tail;
2879 if (__predict_false(++next_cidx == r->size))
2882 while (next_cidx != pidx) {
2883 if (add_to_txpkts(r->items[next_cidx], &txp,
2886 tail->m_nextpkt = r->items[next_cidx];
2887 tail = tail->m_nextpkt;
2888 ETHER_BPF_MTAP(ifp, tail);
2889 if (__predict_false(++next_cidx == r->size))
2893 n = write_txpkts_wr(txq, wr, m0, &txp, available);
2895 remaining -= txp.npkt;
2896 } else if (mbuf_cflags(m0) & MC_RAW_WR) {
2899 n = write_raw_wr(txq, (void *)wr, m0, available);
2903 ETHER_BPF_MTAP(ifp, m0);
2904 n = write_txpkt_wr(txq, (void *)wr, m0, available);
2906 MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2910 IDXINCR(eq->pidx, n, eq->sidx);
2912 if (wr_can_update_eq(wr)) {
2913 if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2914 atomic_cmpset_int(&eq->equiq, 0, 1)) {
2915 wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2917 eq->equeqidx = eq->pidx;
2918 } else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >=
2920 wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2921 eq->equeqidx = eq->pidx;
2925 if (dbdiff >= 16 && remaining >= 4) {
2926 ring_eq_db(sc, eq, dbdiff);
2927 available += reclaim_tx_descs(txq, 4 * dbdiff);
2934 ring_eq_db(sc, eq, dbdiff);
2935 reclaim_tx_descs(txq, 32);
2944 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2948 KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2949 ("%s: bad tmr_idx %d", __func__, tmr_idx));
2950 KASSERT(pktc_idx < SGE_NCOUNTERS, /* -ve is ok, means don't use */
2951 ("%s: bad pktc_idx %d", __func__, pktc_idx));
2955 iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2956 iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2957 if (pktc_idx >= 0) {
2958 iq->intr_params |= F_QINTR_CNT_EN;
2959 iq->intr_pktc_idx = pktc_idx;
2961 iq->qsize = roundup2(qsize, 16); /* See FW_IQ_CMD/iqsize */
2962 iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2966 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2970 fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2971 strlcpy(fl->lockname, name, sizeof(fl->lockname));
2972 if (sc->flags & BUF_PACKING_OK &&
2973 ((!is_t4(sc) && buffer_packing) || /* T5+: enabled unless 0 */
2974 (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2975 fl->flags |= FL_BUF_PACKING;
2976 find_best_refill_source(sc, fl, maxp);
2977 find_safe_refill_source(sc, fl);
2981 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2982 uint8_t tx_chan, uint16_t iqid, char *name)
2984 KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2986 eq->flags = eqtype & EQ_TYPEMASK;
2987 eq->tx_chan = tx_chan;
2989 eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2990 strlcpy(eq->lockname, name, sizeof(eq->lockname));
2994 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2995 bus_dmamap_t *map, bus_addr_t *pa, void **va)
2999 rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
3000 BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
3002 device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
3006 rc = bus_dmamem_alloc(*tag, va,
3007 BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
3009 device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
3013 rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
3015 device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
3020 free_ring(sc, *tag, *map, *pa, *va);
3026 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3027 bus_addr_t pa, void *va)
3030 bus_dmamap_unload(tag, map);
3032 bus_dmamem_free(tag, va, map);
3034 bus_dma_tag_destroy(tag);
3040 * Allocates the ring for an ingress queue and an optional freelist. If the
3041 * freelist is specified it will be allocated and then associated with the
3044 * Returns errno on failure. Resources allocated up to that point may still be
3045 * allocated. Caller is responsible for cleanup in case this function fails.
3047 * If the ingress queue will take interrupts directly then the intr_idx
3048 * specifies the vector, starting from 0. -1 means the interrupts for this
3049 * queue should be forwarded to the fwq.
3052 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3053 int intr_idx, int cong)
3055 int rc, i, cntxt_id;
3058 struct port_info *pi = vi->pi;
3059 struct adapter *sc = iq->adapter;
3060 struct sge_params *sp = &sc->params.sge;
3063 len = iq->qsize * IQ_ESIZE;
3064 rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3065 (void **)&iq->desc);
3069 bzero(&c, sizeof(c));
3070 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3071 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3072 V_FW_IQ_CMD_VFN(0));
3074 c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3077 /* Special handling for firmware event queue */
3078 if (iq == &sc->sge.fwq)
3079 v |= F_FW_IQ_CMD_IQASYNCH;
3082 /* Forwarded interrupts, all headed to fwq */
3083 v |= F_FW_IQ_CMD_IQANDST;
3084 v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3086 KASSERT(intr_idx < sc->intr_count,
3087 ("%s: invalid direct intr_idx %d", __func__, intr_idx));
3088 v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
3091 c.type_to_iqandstindex = htobe32(v |
3092 V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3093 V_FW_IQ_CMD_VIID(vi->viid) |
3094 V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3095 c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3096 F_FW_IQ_CMD_IQGTSMODE |
3097 V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3098 V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3099 c.iqsize = htobe16(iq->qsize);
3100 c.iqaddr = htobe64(iq->ba);
3102 c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3105 mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3107 len = fl->qsize * EQ_ESIZE;
3108 rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3109 &fl->ba, (void **)&fl->desc);
3113 /* Allocate space for one software descriptor per buffer. */
3114 rc = alloc_fl_sdesc(fl);
3116 device_printf(sc->dev,
3117 "failed to setup fl software descriptors: %d\n",
3122 if (fl->flags & FL_BUF_PACKING) {
3123 fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3124 fl->buf_boundary = sp->pack_boundary;
3126 fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3127 fl->buf_boundary = 16;
3129 if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3130 fl->buf_boundary = sp->pad_boundary;
3132 c.iqns_to_fl0congen |=
3133 htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3134 F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3135 (fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3136 (fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3139 c.iqns_to_fl0congen |=
3140 htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3141 F_FW_IQ_CMD_FL0CONGCIF |
3142 F_FW_IQ_CMD_FL0CONGEN);
3144 c.fl0dcaen_to_fl0cidxfthresh =
3145 htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3146 X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
3147 V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3148 X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3149 c.fl0size = htobe16(fl->qsize);
3150 c.fl0addr = htobe64(fl->ba);
3153 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3155 device_printf(sc->dev,
3156 "failed to create ingress queue: %d\n", rc);
3161 iq->gen = F_RSPD_GEN;
3162 iq->intr_next = iq->intr_params;
3163 iq->cntxt_id = be16toh(c.iqid);
3164 iq->abs_id = be16toh(c.physiqid);
3165 iq->flags |= IQ_ALLOCATED;
3167 cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3168 if (cntxt_id >= sc->sge.niq) {
3169 panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3170 cntxt_id, sc->sge.niq - 1);
3172 sc->sge.iqmap[cntxt_id] = iq;
3177 iq->flags |= IQ_HAS_FL;
3178 fl->cntxt_id = be16toh(c.fl0id);
3179 fl->pidx = fl->cidx = 0;
3181 cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3182 if (cntxt_id >= sc->sge.neq) {
3183 panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3184 __func__, cntxt_id, sc->sge.neq - 1);
3186 sc->sge.eqmap[cntxt_id] = (void *)fl;
3189 if (isset(&sc->doorbells, DOORBELL_UDB)) {
3190 uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3191 uint32_t mask = (1 << s_qpp) - 1;
3192 volatile uint8_t *udb;
3194 udb = sc->udbs_base + UDBS_DB_OFFSET;
3195 udb += (qid >> s_qpp) << PAGE_SHIFT;
3197 if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3198 udb += qid << UDBS_SEG_SHIFT;
3201 fl->udb = (volatile void *)udb;
3203 fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3206 /* Enough to make sure the SGE doesn't think it's starved */
3207 refill_fl(sc, fl, fl->lowat);
3211 if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3212 uint32_t param, val;
3214 param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3215 V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3216 V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3221 for (i = 0; i < 4; i++) {
3222 if (cong & (1 << i))
3223 val |= 1 << (i << 2);
3227 rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, ¶m, &val);
3229 /* report error but carry on */
3230 device_printf(sc->dev,
3231 "failed to set congestion manager context for "
3232 "ingress queue %d: %d\n", iq->cntxt_id, rc);
3236 /* Enable IQ interrupts */
3237 atomic_store_rel_int(&iq->state, IQS_IDLE);
3238 t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3239 V_INGRESSQID(iq->cntxt_id));
3245 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3248 struct adapter *sc = iq->adapter;
3252 return (0); /* nothing to do */
3254 dev = vi ? vi->dev : sc->dev;
3256 if (iq->flags & IQ_ALLOCATED) {
3257 rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3258 FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3259 fl ? fl->cntxt_id : 0xffff, 0xffff);
3262 "failed to free queue %p: %d\n", iq, rc);
3265 iq->flags &= ~IQ_ALLOCATED;
3268 free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3270 bzero(iq, sizeof(*iq));
3273 free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3277 free_fl_sdesc(sc, fl);
3279 if (mtx_initialized(&fl->fl_lock))
3280 mtx_destroy(&fl->fl_lock);
3282 bzero(fl, sizeof(*fl));
3289 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3292 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3294 SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3295 "bus address of descriptor ring");
3296 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3297 iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3298 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3299 CTLTYPE_INT | CTLFLAG_RD, &iq->abs_id, 0, sysctl_uint16, "I",
3300 "absolute id of the queue");
3301 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3302 CTLTYPE_INT | CTLFLAG_RD, &iq->cntxt_id, 0, sysctl_uint16, "I",
3303 "SGE context id of the queue");
3304 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3305 CTLTYPE_INT | CTLFLAG_RD, &iq->cidx, 0, sysctl_uint16, "I",
3310 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3311 struct sysctl_oid *oid, struct sge_fl *fl)
3313 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3315 oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3317 children = SYSCTL_CHILDREN(oid);
3319 SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3320 &fl->ba, "bus address of descriptor ring");
3321 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3322 fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3323 "desc ring size in bytes");
3324 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3325 CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
3326 "SGE context id of the freelist");
3327 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3328 fl_pad ? 1 : 0, "padding enabled");
3329 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3330 fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3331 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3332 0, "consumer index");
3333 if (fl->flags & FL_BUF_PACKING) {
3334 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3335 CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3337 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3338 0, "producer index");
3339 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
3340 CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
3341 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
3342 CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
3343 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3344 CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3345 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3346 CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3347 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3348 CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3352 alloc_fwq(struct adapter *sc)
3355 struct sge_iq *fwq = &sc->sge.fwq;
3356 struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3357 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3359 init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3360 if (sc->flags & IS_VF)
3363 intr_idx = sc->intr_count > 1 ? 1 : 0;
3364 rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3366 device_printf(sc->dev,
3367 "failed to create firmware event queue: %d\n", rc);
3371 oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3372 NULL, "firmware event queue");
3373 add_iq_sysctls(&sc->ctx, oid, fwq);
3379 free_fwq(struct adapter *sc)
3381 return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3385 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3386 struct sysctl_oid *oid)
3390 struct sysctl_oid_list *children;
3392 snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3394 init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3395 sc->sge.fwq.cntxt_id, name);
3397 children = SYSCTL_CHILDREN(oid);
3398 snprintf(name, sizeof(name), "%d", idx);
3399 oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3400 NULL, "ctrl queue");
3401 rc = alloc_wrq(sc, NULL, ctrlq, oid);
3407 tnl_cong(struct port_info *pi, int drop)
3415 return (pi->rx_e_chan_map);
3419 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3420 struct sysctl_oid *oid)
3423 struct adapter *sc = vi->pi->adapter;
3424 struct sysctl_oid_list *children;
3427 rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3428 tnl_cong(vi->pi, cong_drop));
3433 sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3435 KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3436 ("iq_base mismatch"));
3437 KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3438 ("PF with non-zero iq_base"));
3441 * The freelist is just barely above the starvation threshold right now,
3442 * fill it up a bit more.
3445 refill_fl(sc, &rxq->fl, 128);
3446 FL_UNLOCK(&rxq->fl);
3448 #if defined(INET) || defined(INET6)
3449 rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3452 MPASS(rxq->lro.ifp == vi->ifp); /* also indicates LRO init'ed */
3454 if (vi->ifp->if_capenable & IFCAP_LRO)
3455 rxq->iq.flags |= IQ_LRO_ENABLED;
3457 if (vi->ifp->if_capenable & IFCAP_HWRXTSTMP)
3458 rxq->iq.flags |= IQ_RX_TIMESTAMP;
3461 children = SYSCTL_CHILDREN(oid);
3463 snprintf(name, sizeof(name), "%d", idx);
3464 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3466 children = SYSCTL_CHILDREN(oid);
3468 add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3469 #if defined(INET) || defined(INET6)
3470 SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3471 &rxq->lro.lro_queued, 0, NULL);
3472 SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3473 &rxq->lro.lro_flushed, 0, NULL);
3475 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3476 &rxq->rxcsum, "# of times hardware assisted with checksum");
3477 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3478 CTLFLAG_RD, &rxq->vlan_extraction,
3479 "# of times hardware extracted 802.1Q tag");
3481 add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3487 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3491 #if defined(INET) || defined(INET6)
3493 tcp_lro_free(&rxq->lro);
3494 rxq->lro.ifp = NULL;
3498 rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3500 bzero(rxq, sizeof(*rxq));
3507 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3508 int intr_idx, int idx, struct sysctl_oid *oid)
3510 struct port_info *pi = vi->pi;
3512 struct sysctl_oid_list *children;
3515 rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3519 children = SYSCTL_CHILDREN(oid);
3521 snprintf(name, sizeof(name), "%d", idx);
3522 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3524 add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3525 add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3531 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3535 rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3537 bzero(ofld_rxq, sizeof(*ofld_rxq));
3545 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3546 int idx, struct sysctl_oid *oid)
3549 struct sysctl_oid_list *children;
3550 struct sysctl_ctx_list *ctx;
3553 struct adapter *sc = vi->pi->adapter;
3554 struct netmap_adapter *na = NA(vi->ifp);
3558 len = vi->qsize_rxq * IQ_ESIZE;
3559 rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3560 &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3564 len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3565 rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3566 &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3572 nm_rxq->iq_cidx = 0;
3573 nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3574 nm_rxq->iq_gen = F_RSPD_GEN;
3575 nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3576 nm_rxq->fl_sidx = na->num_rx_desc;
3577 nm_rxq->intr_idx = intr_idx;
3578 nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3581 children = SYSCTL_CHILDREN(oid);
3583 snprintf(name, sizeof(name), "%d", idx);
3584 oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3586 children = SYSCTL_CHILDREN(oid);
3588 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3589 CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3590 "I", "absolute id of the queue");
3591 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3592 CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3593 "I", "SGE context id of the queue");
3594 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3595 CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3598 children = SYSCTL_CHILDREN(oid);
3599 oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3601 children = SYSCTL_CHILDREN(oid);
3603 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3604 CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3605 "I", "SGE context id of the freelist");
3606 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3607 &nm_rxq->fl_cidx, 0, "consumer index");
3608 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3609 &nm_rxq->fl_pidx, 0, "producer index");
3616 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3618 struct adapter *sc = vi->pi->adapter;
3620 if (vi->flags & VI_INIT_DONE)
3621 MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3623 MPASS(nm_rxq->iq_cntxt_id == 0);
3625 free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3627 free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3634 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3635 struct sysctl_oid *oid)
3639 struct port_info *pi = vi->pi;
3640 struct adapter *sc = pi->adapter;
3641 struct netmap_adapter *na = NA(vi->ifp);
3643 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3645 len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3646 rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3647 &nm_txq->ba, (void **)&nm_txq->desc);
3651 nm_txq->pidx = nm_txq->cidx = 0;
3652 nm_txq->sidx = na->num_tx_desc;
3654 nm_txq->iqidx = iqidx;
3655 nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3656 V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
3657 V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
3658 nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3660 snprintf(name, sizeof(name), "%d", idx);
3661 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3662 NULL, "netmap tx queue");
3663 children = SYSCTL_CHILDREN(oid);
3665 SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3666 &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3667 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3668 CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3670 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3671 CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3678 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3680 struct adapter *sc = vi->pi->adapter;
3682 if (vi->flags & VI_INIT_DONE)
3683 MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3685 MPASS(nm_txq->cntxt_id == 0);
3687 free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3695 * Returns a reasonable automatic cidx flush threshold for a given queue size.
3698 qsize_to_fthresh(int qsize)
3702 while (!powerof2(qsize))
3704 fthresh = ilog2(qsize);
3705 if (fthresh > X_CIDXFLUSHTHRESH_128)
3706 fthresh = X_CIDXFLUSHTHRESH_128;
3712 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3715 struct fw_eq_ctrl_cmd c;
3716 int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3718 bzero(&c, sizeof(c));
3720 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3721 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3722 V_FW_EQ_CTRL_CMD_VFN(0));
3723 c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3724 F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3725 c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3726 c.physeqid_pkd = htobe32(0);
3727 c.fetchszm_to_iqid =
3728 htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3729 V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3730 F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3732 htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3733 V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3734 V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3735 V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3736 c.eqaddr = htobe64(eq->ba);
3738 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3740 device_printf(sc->dev,
3741 "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3744 eq->flags |= EQ_ALLOCATED;
3746 eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3747 cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3748 if (cntxt_id >= sc->sge.neq)
3749 panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3750 cntxt_id, sc->sge.neq - 1);
3751 sc->sge.eqmap[cntxt_id] = eq;
3757 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3760 struct fw_eq_eth_cmd c;
3761 int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3763 bzero(&c, sizeof(c));
3765 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3766 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3767 V_FW_EQ_ETH_CMD_VFN(0));
3768 c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3769 F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3770 c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3771 F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3772 c.fetchszm_to_iqid =
3773 htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3774 V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3775 V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3776 c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3777 V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3778 V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3779 c.eqaddr = htobe64(eq->ba);
3781 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3783 device_printf(vi->dev,
3784 "failed to create Ethernet egress queue: %d\n", rc);
3787 eq->flags |= EQ_ALLOCATED;
3789 eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3790 eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3791 cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3792 if (cntxt_id >= sc->sge.neq)
3793 panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3794 cntxt_id, sc->sge.neq - 1);
3795 sc->sge.eqmap[cntxt_id] = eq;
3800 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3802 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3805 struct fw_eq_ofld_cmd c;
3806 int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3808 bzero(&c, sizeof(c));
3810 c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3811 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3812 V_FW_EQ_OFLD_CMD_VFN(0));
3813 c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3814 F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3815 c.fetchszm_to_iqid =
3816 htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3817 V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3818 F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3820 htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3821 V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3822 V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3823 V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3824 c.eqaddr = htobe64(eq->ba);
3826 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3828 device_printf(vi->dev,
3829 "failed to create egress queue for TCP offload: %d\n", rc);
3832 eq->flags |= EQ_ALLOCATED;
3834 eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3835 cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3836 if (cntxt_id >= sc->sge.neq)
3837 panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3838 cntxt_id, sc->sge.neq - 1);
3839 sc->sge.eqmap[cntxt_id] = eq;
3846 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3851 mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3853 qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3854 len = qsize * EQ_ESIZE;
3855 rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3856 &eq->ba, (void **)&eq->desc);
3860 eq->pidx = eq->cidx = eq->dbidx = 0;
3861 /* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
3863 eq->doorbells = sc->doorbells;
3865 switch (eq->flags & EQ_TYPEMASK) {
3867 rc = ctrl_eq_alloc(sc, eq);
3871 rc = eth_eq_alloc(sc, vi, eq);
3874 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3876 rc = ofld_eq_alloc(sc, vi, eq);
3881 panic("%s: invalid eq type %d.", __func__,
3882 eq->flags & EQ_TYPEMASK);
3885 device_printf(sc->dev,
3886 "failed to allocate egress queue(%d): %d\n",
3887 eq->flags & EQ_TYPEMASK, rc);
3890 if (isset(&eq->doorbells, DOORBELL_UDB) ||
3891 isset(&eq->doorbells, DOORBELL_UDBWC) ||
3892 isset(&eq->doorbells, DOORBELL_WCWR)) {
3893 uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3894 uint32_t mask = (1 << s_qpp) - 1;
3895 volatile uint8_t *udb;
3897 udb = sc->udbs_base + UDBS_DB_OFFSET;
3898 udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT; /* pg offset */
3899 eq->udb_qid = eq->cntxt_id & mask; /* id in page */
3900 if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3901 clrbit(&eq->doorbells, DOORBELL_WCWR);
3903 udb += eq->udb_qid << UDBS_SEG_SHIFT; /* seg offset */
3906 eq->udb = (volatile void *)udb;
3913 free_eq(struct adapter *sc, struct sge_eq *eq)
3917 if (eq->flags & EQ_ALLOCATED) {
3918 switch (eq->flags & EQ_TYPEMASK) {
3920 rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3925 rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3929 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3931 rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3937 panic("%s: invalid eq type %d.", __func__,
3938 eq->flags & EQ_TYPEMASK);
3941 device_printf(sc->dev,
3942 "failed to free egress queue (%d): %d\n",
3943 eq->flags & EQ_TYPEMASK, rc);
3946 eq->flags &= ~EQ_ALLOCATED;
3949 free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3951 if (mtx_initialized(&eq->eq_lock))
3952 mtx_destroy(&eq->eq_lock);
3954 bzero(eq, sizeof(*eq));
3959 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3960 struct sysctl_oid *oid)
3963 struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3964 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3966 rc = alloc_eq(sc, vi, &wrq->eq);
3971 TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3972 TAILQ_INIT(&wrq->incomplete_wrs);
3973 STAILQ_INIT(&wrq->wr_list);
3974 wrq->nwr_pending = 0;
3975 wrq->ndesc_needed = 0;
3977 SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3978 &wrq->eq.ba, "bus address of descriptor ring");
3979 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3980 wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
3981 "desc ring size in bytes");
3982 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3983 &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3984 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3985 CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3987 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3988 CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3990 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3991 wrq->eq.sidx, "status page index");
3992 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3993 &wrq->tx_wrs_direct, "# of work requests (direct)");
3994 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3995 &wrq->tx_wrs_copied, "# of work requests (copied)");
3996 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
3997 &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
4003 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
4007 rc = free_eq(sc, &wrq->eq);
4011 bzero(wrq, sizeof(*wrq));
4016 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
4017 struct sysctl_oid *oid)
4020 struct port_info *pi = vi->pi;
4021 struct adapter *sc = pi->adapter;
4022 struct sge_eq *eq = &txq->eq;
4024 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4026 rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
4029 device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
4033 rc = alloc_eq(sc, vi, eq);
4035 mp_ring_free(txq->r);
4040 /* Can't fail after this point. */
4043 sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4045 KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4046 ("eq_base mismatch"));
4047 KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4048 ("PF with non-zero eq_base"));
4050 TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4052 txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4053 if (sc->flags & IS_VF)
4054 txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4055 V_TXPKT_INTF(pi->tx_chan));
4057 txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
4058 V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
4059 V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
4061 txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4064 snprintf(name, sizeof(name), "%d", idx);
4065 oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
4067 children = SYSCTL_CHILDREN(oid);
4069 SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4070 &eq->ba, "bus address of descriptor ring");
4071 SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4072 eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4073 "desc ring size in bytes");
4074 SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4075 &eq->abs_id, 0, "absolute id of the queue");
4076 SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4077 &eq->cntxt_id, 0, "SGE context id of the queue");
4078 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
4079 CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
4081 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
4082 CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
4084 SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4085 eq->sidx, "status page index");
4087 SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
4088 CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
4089 "traffic class (-1 means none)");
4091 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4092 &txq->txcsum, "# of times hardware assisted with checksum");
4093 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
4094 CTLFLAG_RD, &txq->vlan_insertion,
4095 "# of times hardware inserted 802.1Q tag");
4096 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4097 &txq->tso_wrs, "# of TSO work requests");
4098 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4099 &txq->imm_wrs, "# of work requests with immediate data");
4100 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4101 &txq->sgl_wrs, "# of work requests with direct SGL");
4102 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4103 &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4104 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
4105 CTLFLAG_RD, &txq->txpkts0_wrs,
4106 "# of txpkts (type 0) work requests");
4107 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
4108 CTLFLAG_RD, &txq->txpkts1_wrs,
4109 "# of txpkts (type 1) work requests");
4110 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
4111 CTLFLAG_RD, &txq->txpkts0_pkts,
4112 "# of frames tx'd using type0 txpkts work requests");
4113 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
4114 CTLFLAG_RD, &txq->txpkts1_pkts,
4115 "# of frames tx'd using type1 txpkts work requests");
4116 SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4117 &txq->raw_wrs, "# of raw work requests (non-packets)");
4119 SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
4120 CTLFLAG_RD, &txq->r->enqueues,
4121 "# of enqueues to the mp_ring for this queue");
4122 SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
4123 CTLFLAG_RD, &txq->r->drops,
4124 "# of drops in the mp_ring for this queue");
4125 SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
4126 CTLFLAG_RD, &txq->r->starts,
4127 "# of normal consumer starts in the mp_ring for this queue");
4128 SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
4129 CTLFLAG_RD, &txq->r->stalls,
4130 "# of consumer stalls in the mp_ring for this queue");
4131 SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
4132 CTLFLAG_RD, &txq->r->restarts,
4133 "# of consumer restarts in the mp_ring for this queue");
4134 SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
4135 CTLFLAG_RD, &txq->r->abdications,
4136 "# of consumer abdications in the mp_ring for this queue");
4142 free_txq(struct vi_info *vi, struct sge_txq *txq)
4145 struct adapter *sc = vi->pi->adapter;
4146 struct sge_eq *eq = &txq->eq;
4148 rc = free_eq(sc, eq);
4152 sglist_free(txq->gl);
4153 free(txq->sdesc, M_CXGBE);
4154 mp_ring_free(txq->r);
4156 bzero(txq, sizeof(*txq));
4161 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4163 bus_addr_t *ba = arg;
4166 ("%s meant for single segment mappings only.", __func__));
4168 *ba = error ? 0 : segs->ds_addr;
4172 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4176 n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
4180 v = fl->dbval | V_PIDX(n);
4182 *fl->udb = htole32(v);
4184 t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4185 IDXINCR(fl->dbidx, n, fl->sidx);
4189 * Fills up the freelist by allocating up to 'n' buffers. Buffers that are
4190 * recycled do not count towards this allocation budget.
4192 * Returns non-zero to indicate that this freelist should be added to the list
4193 * of starving freelists.
4196 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4199 struct fl_sdesc *sd;
4202 struct cluster_layout *cll;
4203 struct sw_zone_info *swz;
4204 struct cluster_metadata *clm;
4206 uint16_t hw_cidx = fl->hw_cidx; /* stable snapshot */
4208 FL_LOCK_ASSERT_OWNED(fl);
4211 * We always stop at the beginning of the hardware descriptor that's just
4212 * before the one with the hw cidx. This is to avoid hw pidx = hw cidx,
4213 * which would mean an empty freelist to the chip.
4215 max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4216 if (fl->pidx == max_pidx * 8)
4219 d = &fl->desc[fl->pidx];
4220 sd = &fl->sdesc[fl->pidx];
4221 cll = &fl->cll_def; /* default layout */
4222 swz = &sc->sge.sw_zone_info[cll->zidx];
4226 if (sd->cl != NULL) {
4228 if (sd->nmbuf == 0) {
4230 * Fast recycle without involving any atomics on
4231 * the cluster's metadata (if the cluster has
4232 * metadata). This happens when all frames
4233 * received in the cluster were small enough to
4234 * fit within a single mbuf each.
4236 fl->cl_fast_recycled++;
4238 clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4240 MPASS(clm->refcount == 1);
4246 * Cluster is guaranteed to have metadata. Clusters
4247 * without metadata always take the fast recycle path
4248 * when they're recycled.
4250 clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4253 if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4255 counter_u64_add(extfree_rels, 1);
4258 sd->cl = NULL; /* gave up my reference */
4260 MPASS(sd->cl == NULL);
4262 cl = uma_zalloc(swz->zone, M_NOWAIT);
4263 if (__predict_false(cl == NULL)) {
4264 if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
4265 fl->cll_def.zidx == fl->cll_alt.zidx)
4268 /* fall back to the safe zone */
4270 swz = &sc->sge.sw_zone_info[cll->zidx];
4276 pa = pmap_kextract((vm_offset_t)cl);
4280 *d = htobe64(pa | cll->hwidx);
4281 clm = cl_metadata(sc, fl, cll, cl);
4293 if (__predict_false(++fl->pidx % 8 == 0)) {
4294 uint16_t pidx = fl->pidx / 8;
4296 if (__predict_false(pidx == fl->sidx)) {
4302 if (pidx == max_pidx)
4305 if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4310 if (fl->pidx / 8 != fl->dbidx)
4313 return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4317 * Attempt to refill all starving freelists.
4320 refill_sfl(void *arg)
4322 struct adapter *sc = arg;
4323 struct sge_fl *fl, *fl_temp;
4325 mtx_assert(&sc->sfl_lock, MA_OWNED);
4326 TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4328 refill_fl(sc, fl, 64);
4329 if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4330 TAILQ_REMOVE(&sc->sfl, fl, link);
4331 fl->flags &= ~FL_STARVING;
4336 if (!TAILQ_EMPTY(&sc->sfl))
4337 callout_schedule(&sc->sfl_callout, hz / 5);
4341 alloc_fl_sdesc(struct sge_fl *fl)
4344 fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4351 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4353 struct fl_sdesc *sd;
4354 struct cluster_metadata *clm;
4355 struct cluster_layout *cll;
4359 for (i = 0; i < fl->sidx * 8; i++, sd++) {
4364 clm = cl_metadata(sc, fl, cll, sd->cl);
4366 uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4367 else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4368 uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4369 counter_u64_add(extfree_rels, 1);
4374 free(fl->sdesc, M_CXGBE);
4379 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4386 rc = sglist_append_mbuf(gl, m);
4387 if (__predict_false(rc != 0)) {
4388 panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4389 "with %d.", __func__, m, mbuf_nsegs(m), rc);
4392 KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4393 ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4394 mbuf_nsegs(m), gl->sg_nseg));
4395 KASSERT(gl->sg_nseg > 0 &&
4396 gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4397 ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4398 gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4402 * len16 for a txpkt WR with a GL. Includes the firmware work request header.
4405 txpkt_len16(u_int nsegs, u_int tso)
4411 nsegs--; /* first segment is part of ulptx_sgl */
4412 n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4413 sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4415 n += sizeof(struct cpl_tx_pkt_lso_core);
4417 return (howmany(n, 16));
4421 * len16 for a txpkt_vm WR with a GL. Includes the firmware work
4425 txpkt_vm_len16(u_int nsegs, u_int tso)
4431 nsegs--; /* first segment is part of ulptx_sgl */
4432 n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4433 sizeof(struct cpl_tx_pkt_core) +
4434 sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4436 n += sizeof(struct cpl_tx_pkt_lso_core);
4438 return (howmany(n, 16));
4442 * len16 for a txpkts type 0 WR with a GL. Does not include the firmware work
4446 txpkts0_len16(u_int nsegs)
4452 nsegs--; /* first segment is part of ulptx_sgl */
4453 n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4454 sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4455 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4457 return (howmany(n, 16));
4461 * len16 for a txpkts type 1 WR with a GL. Does not include the firmware work
4469 n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4471 return (howmany(n, 16));
4475 imm_payload(u_int ndesc)
4479 n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4480 sizeof(struct cpl_tx_pkt_core);
4486 * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4487 * software descriptor, and advance the pidx. It is guaranteed that enough
4488 * descriptors are available.
4490 * The return value is the # of hardware descriptors used.
4493 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4494 struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4496 struct sge_eq *eq = &txq->eq;
4497 struct tx_sdesc *txsd;
4498 struct cpl_tx_pkt_core *cpl;
4499 uint32_t ctrl; /* used in many unrelated places */
4501 int csum_type, len16, ndesc, pktlen, nsegs;
4504 TXQ_LOCK_ASSERT_OWNED(txq);
4506 MPASS(available > 0 && available < eq->sidx);
4508 len16 = mbuf_len16(m0);
4509 nsegs = mbuf_nsegs(m0);
4510 pktlen = m0->m_pkthdr.len;
4511 ctrl = sizeof(struct cpl_tx_pkt_core);
4513 ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4514 ndesc = howmany(len16, EQ_ESIZE / 16);
4515 MPASS(ndesc <= available);
4517 /* Firmware work request header */
4518 MPASS(wr == (void *)&eq->desc[eq->pidx]);
4519 wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4520 V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4522 ctrl = V_FW_WR_LEN16(len16);
4523 wr->equiq_to_len16 = htobe32(ctrl);
4528 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4529 * vlantci is ignored unless the ethtype is 0x8100, so it's
4530 * simpler to always copy it rather than making it
4531 * conditional. Also, it seems that we do not have to set
4532 * vlantci or fake the ethtype when doing VLAN tag insertion.
4534 m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4537 if (needs_tso(m0)) {
4538 struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4540 KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4541 m0->m_pkthdr.l4hlen > 0,
4542 ("%s: mbuf %p needs TSO but missing header lengths",
4545 ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4546 F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4547 | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4548 if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4549 ctrl |= V_LSO_ETHHDR_LEN(1);
4550 if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4553 lso->lso_ctrl = htobe32(ctrl);
4554 lso->ipid_ofst = htobe16(0);
4555 lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4556 lso->seqno_offset = htobe32(0);
4557 lso->len = htobe32(pktlen);
4559 if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4560 csum_type = TX_CSUM_TCPIP6;
4562 csum_type = TX_CSUM_TCPIP;
4564 cpl = (void *)(lso + 1);
4568 if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4569 csum_type = TX_CSUM_TCPIP;
4570 else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4571 csum_type = TX_CSUM_UDPIP;
4572 else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4573 csum_type = TX_CSUM_TCPIP6;
4574 else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4575 csum_type = TX_CSUM_UDPIP6;
4577 else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4579 * XXX: The firmware appears to stomp on the
4580 * fragment/flags field of the IP header when
4581 * using TX_CSUM_IP. Fall back to doing
4582 * software checksums.
4590 sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4591 offsetof(struct ip, ip_sum));
4592 *sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4593 m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4594 m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4598 cpl = (void *)(wr + 1);
4601 /* Checksum offload */
4603 if (needs_l3_csum(m0) == 0)
4604 ctrl1 |= F_TXPKT_IPCSUM_DIS;
4605 if (csum_type >= 0) {
4606 KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4607 ("%s: mbuf %p needs checksum offload but missing header lengths",
4610 if (chip_id(sc) <= CHELSIO_T5) {
4611 ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4614 ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4617 ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4618 ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4620 ctrl1 |= F_TXPKT_L4CSUM_DIS;
4621 if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4622 CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4623 txq->txcsum++; /* some hardware assistance provided */
4625 /* VLAN tag insertion */
4626 if (needs_vlan_insertion(m0)) {
4627 ctrl1 |= F_TXPKT_VLAN_VLD |
4628 V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4629 txq->vlan_insertion++;
4633 cpl->ctrl0 = txq->cpl_ctrl0;
4635 cpl->len = htobe16(pktlen);
4636 cpl->ctrl1 = htobe64(ctrl1);
4639 dst = (void *)(cpl + 1);
4642 * A packet using TSO will use up an entire descriptor for the
4643 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4644 * If this descriptor is the last descriptor in the ring, wrap
4645 * around to the front of the ring explicitly for the start of
4648 if (dst == (void *)&eq->desc[eq->sidx]) {
4649 dst = (void *)&eq->desc[0];
4650 write_gl_to_txd(txq, m0, &dst, 0);
4652 write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4657 txsd = &txq->sdesc[eq->pidx];
4659 txsd->desc_used = ndesc;
4665 * Write a raw WR to the hardware descriptors, update the software
4666 * descriptor, and advance the pidx. It is guaranteed that enough
4667 * descriptors are available.
4669 * The return value is the # of hardware descriptors used.
4672 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
4674 struct sge_eq *eq = &txq->eq;
4675 struct tx_sdesc *txsd;
4680 len16 = mbuf_len16(m0);
4681 ndesc = howmany(len16, EQ_ESIZE / 16);
4682 MPASS(ndesc <= available);
4685 for (m = m0; m != NULL; m = m->m_next)
4686 copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4690 txsd = &txq->sdesc[eq->pidx];
4692 txsd->desc_used = ndesc;
4698 * Write a txpkt WR for this packet to the hardware descriptors, update the
4699 * software descriptor, and advance the pidx. It is guaranteed that enough
4700 * descriptors are available.
4702 * The return value is the # of hardware descriptors used.
4705 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4706 struct mbuf *m0, u_int available)
4708 struct sge_eq *eq = &txq->eq;
4709 struct tx_sdesc *txsd;
4710 struct cpl_tx_pkt_core *cpl;
4711 uint32_t ctrl; /* used in many unrelated places */
4713 int len16, ndesc, pktlen, nsegs;
4716 TXQ_LOCK_ASSERT_OWNED(txq);
4718 MPASS(available > 0 && available < eq->sidx);
4720 len16 = mbuf_len16(m0);
4721 nsegs = mbuf_nsegs(m0);
4722 pktlen = m0->m_pkthdr.len;
4723 ctrl = sizeof(struct cpl_tx_pkt_core);
4725 ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4726 else if (pktlen <= imm_payload(2) && available >= 2) {
4727 /* Immediate data. Recalculate len16 and set nsegs to 0. */
4729 len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4730 sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4733 ndesc = howmany(len16, EQ_ESIZE / 16);
4734 MPASS(ndesc <= available);
4736 /* Firmware work request header */
4737 MPASS(wr == (void *)&eq->desc[eq->pidx]);
4738 wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4739 V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4741 ctrl = V_FW_WR_LEN16(len16);
4742 wr->equiq_to_len16 = htobe32(ctrl);
4745 if (needs_tso(m0)) {
4746 struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4748 KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4749 m0->m_pkthdr.l4hlen > 0,
4750 ("%s: mbuf %p needs TSO but missing header lengths",
4753 ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4754 F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4755 | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4756 if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4757 ctrl |= V_LSO_ETHHDR_LEN(1);
4758 if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4761 lso->lso_ctrl = htobe32(ctrl);
4762 lso->ipid_ofst = htobe16(0);
4763 lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4764 lso->seqno_offset = htobe32(0);
4765 lso->len = htobe32(pktlen);
4767 cpl = (void *)(lso + 1);
4771 cpl = (void *)(wr + 1);
4773 /* Checksum offload */
4775 if (needs_l3_csum(m0) == 0)
4776 ctrl1 |= F_TXPKT_IPCSUM_DIS;
4777 if (needs_l4_csum(m0) == 0)
4778 ctrl1 |= F_TXPKT_L4CSUM_DIS;
4779 if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4780 CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4781 txq->txcsum++; /* some hardware assistance provided */
4783 /* VLAN tag insertion */
4784 if (needs_vlan_insertion(m0)) {
4785 ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4786 txq->vlan_insertion++;
4790 cpl->ctrl0 = txq->cpl_ctrl0;
4792 cpl->len = htobe16(pktlen);
4793 cpl->ctrl1 = htobe64(ctrl1);
4796 dst = (void *)(cpl + 1);
4799 write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4804 for (m = m0; m != NULL; m = m->m_next) {
4805 copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4811 KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4818 txsd = &txq->sdesc[eq->pidx];
4820 txsd->desc_used = ndesc;
4826 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4828 u_int needed, nsegs1, nsegs2, l1, l2;
4830 if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4833 nsegs1 = mbuf_nsegs(m);
4834 nsegs2 = mbuf_nsegs(n);
4835 if (nsegs1 + nsegs2 == 2) {
4837 l1 = l2 = txpkts1_len16();
4840 l1 = txpkts0_len16(nsegs1);
4841 l2 = txpkts0_len16(nsegs2);
4843 txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4844 needed = howmany(txp->len16, EQ_ESIZE / 16);
4845 if (needed > SGE_MAX_WR_NDESC || needed > available)
4848 txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4849 if (txp->plen > 65535)
4853 set_mbuf_len16(m, l1);
4854 set_mbuf_len16(n, l2);
4860 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4862 u_int plen, len16, needed, nsegs;
4864 MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4866 if (cannot_use_txpkts(m))
4869 nsegs = mbuf_nsegs(m);
4870 if (txp->wr_type == 1 && nsegs != 1)
4873 plen = txp->plen + m->m_pkthdr.len;
4877 if (txp->wr_type == 0)
4878 len16 = txpkts0_len16(nsegs);
4880 len16 = txpkts1_len16();
4881 needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4882 if (needed > SGE_MAX_WR_NDESC || needed > available)
4887 txp->len16 += len16;
4888 set_mbuf_len16(m, len16);
4894 * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4895 * the software descriptor, and advance the pidx. It is guaranteed that enough
4896 * descriptors are available.
4898 * The return value is the # of hardware descriptors used.
4901 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4902 struct mbuf *m0, const struct txpkts *txp, u_int available)
4904 struct sge_eq *eq = &txq->eq;
4905 struct tx_sdesc *txsd;
4906 struct cpl_tx_pkt_core *cpl;
4909 int ndesc, checkwrap;
4913 TXQ_LOCK_ASSERT_OWNED(txq);
4914 MPASS(txp->npkt > 0);
4915 MPASS(txp->plen < 65536);
4917 MPASS(m0->m_nextpkt != NULL);
4918 MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4919 MPASS(available > 0 && available < eq->sidx);
4921 ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4922 MPASS(ndesc <= available);
4924 MPASS(wr == (void *)&eq->desc[eq->pidx]);
4925 wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4926 ctrl = V_FW_WR_LEN16(txp->len16);
4927 wr->equiq_to_len16 = htobe32(ctrl);
4928 wr->plen = htobe16(txp->plen);
4929 wr->npkt = txp->npkt;
4931 wr->type = txp->wr_type;
4935 * At this point we are 16B into a hardware descriptor. If checkwrap is
4936 * set then we know the WR is going to wrap around somewhere. We'll
4937 * check for that at appropriate points.
4939 checkwrap = eq->sidx - ndesc < eq->pidx;
4940 for (m = m0; m != NULL; m = m->m_nextpkt) {
4941 if (txp->wr_type == 0) {
4942 struct ulp_txpkt *ulpmc;
4943 struct ulptx_idata *ulpsc;
4945 /* ULP master command */
4947 ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4948 V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4949 ulpmc->len = htobe32(mbuf_len16(m));
4951 /* ULP subcommand */
4952 ulpsc = (void *)(ulpmc + 1);
4953 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4955 ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4957 cpl = (void *)(ulpsc + 1);
4959 (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4960 cpl = (void *)&eq->desc[0];
4965 /* Checksum offload */
4967 if (needs_l3_csum(m) == 0)
4968 ctrl1 |= F_TXPKT_IPCSUM_DIS;
4969 if (needs_l4_csum(m) == 0)
4970 ctrl1 |= F_TXPKT_L4CSUM_DIS;
4971 if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4972 CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4973 txq->txcsum++; /* some hardware assistance provided */
4975 /* VLAN tag insertion */
4976 if (needs_vlan_insertion(m)) {
4977 ctrl1 |= F_TXPKT_VLAN_VLD |
4978 V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4979 txq->vlan_insertion++;
4983 cpl->ctrl0 = txq->cpl_ctrl0;
4985 cpl->len = htobe16(m->m_pkthdr.len);
4986 cpl->ctrl1 = htobe64(ctrl1);
4990 (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4991 flitp = (void *)&eq->desc[0];
4993 write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4997 if (txp->wr_type == 0) {
4998 txq->txpkts0_pkts += txp->npkt;
5001 txq->txpkts1_pkts += txp->npkt;
5005 txsd = &txq->sdesc[eq->pidx];
5007 txsd->desc_used = ndesc;
5013 * If the SGL ends on an address that is not 16 byte aligned, this function will
5014 * add a 0 filled flit at the end.
5017 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
5019 struct sge_eq *eq = &txq->eq;
5020 struct sglist *gl = txq->gl;
5021 struct sglist_seg *seg;
5022 __be64 *flitp, *wrap;
5023 struct ulptx_sgl *usgl;
5024 int i, nflits, nsegs;
5026 KASSERT(((uintptr_t)(*to) & 0xf) == 0,
5027 ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
5028 MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5029 MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5032 nsegs = gl->sg_nseg;
5035 nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
5036 flitp = (__be64 *)(*to);
5037 wrap = (__be64 *)(&eq->desc[eq->sidx]);
5038 seg = &gl->sg_segs[0];
5039 usgl = (void *)flitp;
5042 * We start at a 16 byte boundary somewhere inside the tx descriptor
5043 * ring, so we're at least 16 bytes away from the status page. There is
5044 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
5047 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5048 V_ULPTX_NSGE(nsegs));
5049 usgl->len0 = htobe32(seg->ss_len);
5050 usgl->addr0 = htobe64(seg->ss_paddr);
5053 if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
5055 /* Won't wrap around at all */
5057 for (i = 0; i < nsegs - 1; i++, seg++) {
5058 usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
5059 usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
5062 usgl->sge[i / 2].len[1] = htobe32(0);
5066 /* Will wrap somewhere in the rest of the SGL */
5068 /* 2 flits already written, write the rest flit by flit */
5069 flitp = (void *)(usgl + 1);
5070 for (i = 0; i < nflits - 2; i++) {
5072 flitp = (void *)eq->desc;
5073 *flitp++ = get_flit(seg, nsegs - 1, i);
5078 MPASS(((uintptr_t)flitp) & 0xf);
5082 MPASS((((uintptr_t)flitp) & 0xf) == 0);
5083 if (__predict_false(flitp == wrap))
5084 *to = (void *)eq->desc;
5086 *to = (void *)flitp;
5090 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
5093 MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5094 MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5096 if (__predict_true((uintptr_t)(*to) + len <=
5097 (uintptr_t)&eq->desc[eq->sidx])) {
5098 bcopy(from, *to, len);
5101 int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
5103 bcopy(from, *to, portion);
5105 portion = len - portion; /* remaining */
5106 bcopy(from, (void *)eq->desc, portion);
5107 (*to) = (caddr_t)eq->desc + portion;
5112 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
5120 clrbit(&db, DOORBELL_WCWR);
5123 switch (ffs(db) - 1) {
5125 *eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5128 case DOORBELL_WCWR: {
5129 volatile uint64_t *dst, *src;
5133 * Queues whose 128B doorbell segment fits in the page do not
5134 * use relative qid (udb_qid is always 0). Only queues with
5135 * doorbell segments can do WCWR.
5137 KASSERT(eq->udb_qid == 0 && n == 1,
5138 ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
5139 __func__, eq->doorbells, n, eq->dbidx, eq));
5141 dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
5144 src = (void *)&eq->desc[i];
5145 while (src != (void *)&eq->desc[i + 1])
5151 case DOORBELL_UDBWC:
5152 *eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5157 t4_write_reg(sc, sc->sge_kdoorbell_reg,
5158 V_QID(eq->cntxt_id) | V_PIDX(n));
5162 IDXINCR(eq->dbidx, n, eq->sidx);
5166 reclaimable_tx_desc(struct sge_eq *eq)
5170 hw_cidx = read_hw_cidx(eq);
5171 return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
5175 total_available_tx_desc(struct sge_eq *eq)
5177 uint16_t hw_cidx, pidx;
5179 hw_cidx = read_hw_cidx(eq);
5182 if (pidx == hw_cidx)
5183 return (eq->sidx - 1);
5185 return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5188 static inline uint16_t
5189 read_hw_cidx(struct sge_eq *eq)
5191 struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5192 uint16_t cidx = spg->cidx; /* stable snapshot */
5194 return (be16toh(cidx));
5198 * Reclaim 'n' descriptors approximately.
5201 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5203 struct tx_sdesc *txsd;
5204 struct sge_eq *eq = &txq->eq;
5205 u_int can_reclaim, reclaimed;
5207 TXQ_LOCK_ASSERT_OWNED(txq);
5211 can_reclaim = reclaimable_tx_desc(eq);
5212 while (can_reclaim && reclaimed < n) {
5214 struct mbuf *m, *nextpkt;
5216 txsd = &txq->sdesc[eq->cidx];
5217 ndesc = txsd->desc_used;
5219 /* Firmware doesn't return "partial" credits. */
5220 KASSERT(can_reclaim >= ndesc,
5221 ("%s: unexpected number of credits: %d, %d",
5222 __func__, can_reclaim, ndesc));
5224 ("%s: descriptor with no credits: cidx %d",
5225 __func__, eq->cidx));
5227 for (m = txsd->m; m != NULL; m = nextpkt) {
5228 nextpkt = m->m_nextpkt;
5229 m->m_nextpkt = NULL;
5233 can_reclaim -= ndesc;
5234 IDXINCR(eq->cidx, ndesc, eq->sidx);
5241 tx_reclaim(void *arg, int n)
5243 struct sge_txq *txq = arg;
5244 struct sge_eq *eq = &txq->eq;
5247 if (TXQ_TRYLOCK(txq) == 0)
5249 n = reclaim_tx_descs(txq, 32);
5250 if (eq->cidx == eq->pidx)
5251 eq->equeqidx = eq->pidx;
5257 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5259 int i = (idx / 3) * 2;
5265 rc = (uint64_t)segs[i].ss_len << 32;
5267 rc |= (uint64_t)(segs[i + 1].ss_len);
5269 return (htobe64(rc));
5272 return (htobe64(segs[i].ss_paddr));
5274 return (htobe64(segs[i + 1].ss_paddr));
5281 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
5283 int8_t zidx, hwidx, idx;
5284 uint16_t region1, region3;
5285 int spare, spare_needed, n;
5286 struct sw_zone_info *swz;
5287 struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
5290 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
5291 * large enough for the max payload and cluster metadata. Otherwise
5292 * settle for the largest bufsize that leaves enough room in the cluster
5295 * Without buffer packing: Look for the smallest zone which has a
5296 * bufsize large enough for the max payload. Settle for the largest
5297 * bufsize available if there's nothing big enough for max payload.
5299 spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
5300 swz = &sc->sge.sw_zone_info[0];
5302 for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
5303 if (swz->size > largest_rx_cluster) {
5304 if (__predict_true(hwidx != -1))
5308 * This is a misconfiguration. largest_rx_cluster is
5309 * preventing us from finding a refill source. See
5310 * dev.t5nex.<n>.buffer_sizes to figure out why.
5312 device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
5313 " refill source for fl %p (dma %u). Ignored.\n",
5314 largest_rx_cluster, fl, maxp);
5316 for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
5317 hwb = &hwb_list[idx];
5318 spare = swz->size - hwb->size;
5319 if (spare < spare_needed)
5322 hwidx = idx; /* best option so far */
5323 if (hwb->size >= maxp) {
5325 if ((fl->flags & FL_BUF_PACKING) == 0)
5326 goto done; /* stop looking (not packing) */
5328 if (swz->size >= safest_rx_cluster)
5329 goto done; /* stop looking (packing) */
5331 break; /* keep looking, next zone */
5335 /* A usable hwidx has been located. */
5337 hwb = &hwb_list[hwidx];
5339 swz = &sc->sge.sw_zone_info[zidx];
5341 region3 = swz->size - hwb->size;
5344 * Stay within this zone and see if there is a better match when mbuf
5345 * inlining is allowed. Remember that the hwidx's are sorted in
5346 * decreasing order of size (so in increasing order of spare area).
5348 for (idx = hwidx; idx != -1; idx = hwb->next) {
5349 hwb = &hwb_list[idx];
5350 spare = swz->size - hwb->size;
5352 if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
5356 * Do not inline mbufs if doing so would violate the pad/pack
5357 * boundary alignment requirement.
5359 if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
5361 if (fl->flags & FL_BUF_PACKING &&
5362 (MSIZE % sc->params.sge.pack_boundary) != 0)
5365 if (spare < CL_METADATA_SIZE + MSIZE)
5367 n = (spare - CL_METADATA_SIZE) / MSIZE;
5368 if (n > howmany(hwb->size, maxp))
5372 if (fl->flags & FL_BUF_PACKING) {
5373 region1 = n * MSIZE;
5374 region3 = spare - region1;
5377 region3 = spare - region1;
5382 KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5383 ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5384 KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5385 ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5386 KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5387 sc->sge.sw_zone_info[zidx].size,
5388 ("%s: bad buffer layout for fl %p, maxp %d. "
5389 "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5390 sc->sge.sw_zone_info[zidx].size, region1,
5391 sc->sge.hw_buf_info[hwidx].size, region3));
5392 if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5393 KASSERT(region3 >= CL_METADATA_SIZE,
5394 ("%s: no room for metadata. fl %p, maxp %d; "
5395 "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5396 sc->sge.sw_zone_info[zidx].size, region1,
5397 sc->sge.hw_buf_info[hwidx].size, region3));
5398 KASSERT(region1 % MSIZE == 0,
5399 ("%s: bad mbuf region for fl %p, maxp %d. "
5400 "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5401 sc->sge.sw_zone_info[zidx].size, region1,
5402 sc->sge.hw_buf_info[hwidx].size, region3));
5405 fl->cll_def.zidx = zidx;
5406 fl->cll_def.hwidx = hwidx;
5407 fl->cll_def.region1 = region1;
5408 fl->cll_def.region3 = region3;
5412 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5414 struct sge *s = &sc->sge;
5415 struct hw_buf_info *hwb;
5416 struct sw_zone_info *swz;
5420 if (fl->flags & FL_BUF_PACKING)
5421 hwidx = s->safe_hwidx2; /* with room for metadata */
5422 else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5423 hwidx = s->safe_hwidx2;
5424 hwb = &s->hw_buf_info[hwidx];
5425 swz = &s->sw_zone_info[hwb->zidx];
5426 spare = swz->size - hwb->size;
5428 /* no good if there isn't room for an mbuf as well */
5429 if (spare < CL_METADATA_SIZE + MSIZE)
5430 hwidx = s->safe_hwidx1;
5432 hwidx = s->safe_hwidx1;
5435 /* No fallback source */
5436 fl->cll_alt.hwidx = -1;
5437 fl->cll_alt.zidx = -1;
5442 hwb = &s->hw_buf_info[hwidx];
5443 swz = &s->sw_zone_info[hwb->zidx];
5444 spare = swz->size - hwb->size;
5445 fl->cll_alt.hwidx = hwidx;
5446 fl->cll_alt.zidx = hwb->zidx;
5447 if (allow_mbufs_in_cluster &&
5448 (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5449 fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5451 fl->cll_alt.region1 = 0;
5452 fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5456 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5458 mtx_lock(&sc->sfl_lock);
5460 if ((fl->flags & FL_DOOMED) == 0) {
5461 fl->flags |= FL_STARVING;
5462 TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5463 callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5466 mtx_unlock(&sc->sfl_lock);
5470 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5472 struct sge_wrq *wrq = (void *)eq;
5474 atomic_readandclear_int(&eq->equiq);
5475 taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5479 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5481 struct sge_txq *txq = (void *)eq;
5483 MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5485 atomic_readandclear_int(&eq->equiq);
5486 mp_ring_check_drainage(txq->r, 0);
5487 taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5491 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5494 const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5495 unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5496 struct adapter *sc = iq->adapter;
5497 struct sge *s = &sc->sge;
5499 static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5500 &handle_wrq_egr_update, &handle_eth_egr_update,
5501 &handle_wrq_egr_update};
5503 KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5506 eq = s->eqmap[qid - s->eq_start - s->eq_base];
5507 (*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5512 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5513 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5514 offsetof(struct cpl_fw6_msg, data));
5517 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5519 struct adapter *sc = iq->adapter;
5520 const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5522 KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5525 if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5526 const struct rss_header *rss2;
5528 rss2 = (const struct rss_header *)&cpl->data[0];
5529 return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5532 return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5536 * t4_handle_wrerr_rpl - process a FW work request error message
5537 * @adap: the adapter
5538 * @rpl: start of the FW message
5541 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5543 u8 opcode = *(const u8 *)rpl;
5544 const struct fw_error_cmd *e = (const void *)rpl;
5547 if (opcode != FW_ERROR_CMD) {
5549 "%s: Received WRERR_RPL message with opcode %#x\n",
5550 device_get_nameunit(adap->dev), opcode);
5553 log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5554 G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5556 switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5557 case FW_ERROR_TYPE_EXCEPTION:
5558 log(LOG_ERR, "exception info:\n");
5559 for (i = 0; i < nitems(e->u.exception.info); i++)
5560 log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5561 be32toh(e->u.exception.info[i]));
5564 case FW_ERROR_TYPE_HWMODULE:
5565 log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5566 be32toh(e->u.hwmodule.regaddr),
5567 be32toh(e->u.hwmodule.regval));
5569 case FW_ERROR_TYPE_WR:
5570 log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5571 be16toh(e->u.wr.cidx),
5572 G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5573 G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5574 be32toh(e->u.wr.eqid));
5575 for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5576 log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5580 case FW_ERROR_TYPE_ACL:
5581 log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5582 be16toh(e->u.acl.cidx),
5583 G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5584 G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5585 be32toh(e->u.acl.eqid),
5586 G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5588 for (i = 0; i < nitems(e->u.acl.val); i++)
5589 log(LOG_ERR, " %02x", e->u.acl.val[i]);
5593 log(LOG_ERR, "type %#x\n",
5594 G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5601 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5603 uint16_t *id = arg1;
5606 return sysctl_handle_int(oidp, &i, 0, req);
5610 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5612 struct sge *s = arg1;
5613 struct hw_buf_info *hwb = &s->hw_buf_info[0];
5614 struct sw_zone_info *swz = &s->sw_zone_info[0];
5619 sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5620 for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5621 if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5626 sbuf_printf(&sb, "%u%c ", hwb->size, c);
5630 rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5637 * len16 for a txpkt WR with a GL. Includes the firmware work request header.
5640 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
5646 n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
5647 sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
5648 if (__predict_false(nsegs == 0))
5651 nsegs--; /* first segment is part of ulptx_sgl */
5652 n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5654 n += sizeof(struct cpl_tx_pkt_lso_core);
5657 return (howmany(n, 16));
5660 #define ETID_FLOWC_NPARAMS 6
5661 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
5662 ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
5663 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
5666 send_etid_flowc_wr(struct cxgbe_snd_tag *cst, struct port_info *pi,
5669 struct wrq_cookie cookie;
5670 u_int pfvf = pi->adapter->pf << S_FW_VIID_PFN;
5671 struct fw_flowc_wr *flowc;
5673 mtx_assert(&cst->lock, MA_OWNED);
5674 MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
5677 flowc = start_wrq_wr(cst->eo_txq, ETID_FLOWC_LEN16, &cookie);
5678 if (__predict_false(flowc == NULL))
5681 bzero(flowc, ETID_FLOWC_LEN);
5682 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5683 V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
5684 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
5685 V_FW_WR_FLOWID(cst->etid));
5686 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
5687 flowc->mnemval[0].val = htobe32(pfvf);
5688 flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
5689 flowc->mnemval[1].val = htobe32(pi->tx_chan);
5690 flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
5691 flowc->mnemval[2].val = htobe32(pi->tx_chan);
5692 flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
5693 flowc->mnemval[3].val = htobe32(cst->iqid);
5694 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
5695 flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
5696 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
5697 flowc->mnemval[5].val = htobe32(cst->schedcl);
5699 commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5701 cst->flags &= ~EO_FLOWC_PENDING;
5702 cst->flags |= EO_FLOWC_RPL_PENDING;
5703 MPASS(cst->tx_credits >= ETID_FLOWC_LEN16); /* flowc is first WR. */
5704 cst->tx_credits -= ETID_FLOWC_LEN16;
5709 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
5712 send_etid_flush_wr(struct cxgbe_snd_tag *cst)
5714 struct fw_flowc_wr *flowc;
5715 struct wrq_cookie cookie;
5717 mtx_assert(&cst->lock, MA_OWNED);
5719 flowc = start_wrq_wr(cst->eo_txq, ETID_FLUSH_LEN16, &cookie);
5720 if (__predict_false(flowc == NULL))
5721 CXGBE_UNIMPLEMENTED(__func__);
5723 bzero(flowc, ETID_FLUSH_LEN16 * 16);
5724 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5725 V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
5726 flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
5727 V_FW_WR_FLOWID(cst->etid));
5729 commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5731 cst->flags |= EO_FLUSH_RPL_PENDING;
5732 MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
5733 cst->tx_credits -= ETID_FLUSH_LEN16;
5738 write_ethofld_wr(struct cxgbe_snd_tag *cst, struct fw_eth_tx_eo_wr *wr,
5739 struct mbuf *m0, int compl)
5741 struct cpl_tx_pkt_core *cpl;
5743 uint32_t ctrl; /* used in many unrelated places */
5744 int len16, pktlen, nsegs, immhdrs;
5747 struct ulptx_sgl *usgl;
5749 struct sglist_seg segs[38]; /* XXX: find real limit. XXX: get off the stack */
5751 mtx_assert(&cst->lock, MA_OWNED);
5753 KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5754 m0->m_pkthdr.l4hlen > 0,
5755 ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
5757 len16 = mbuf_eo_len16(m0);
5758 nsegs = mbuf_eo_nsegs(m0);
5759 pktlen = m0->m_pkthdr.len;
5760 ctrl = sizeof(struct cpl_tx_pkt_core);
5762 ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5763 immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
5766 wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
5767 V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
5768 wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
5769 V_FW_WR_FLOWID(cst->etid));
5771 if (needs_udp_csum(m0)) {
5772 wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
5773 wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
5774 wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5775 wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
5776 wr->u.udpseg.rtplen = 0;
5777 wr->u.udpseg.r4 = 0;
5778 wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
5779 wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
5780 wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
5781 cpl = (void *)(wr + 1);
5783 MPASS(needs_tcp_csum(m0));
5784 wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
5785 wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
5786 wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5787 wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
5788 wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
5789 wr->u.tcpseg.r4 = 0;
5790 wr->u.tcpseg.r5 = 0;
5791 wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
5793 if (needs_tso(m0)) {
5794 struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
5796 wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
5798 ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
5799 F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
5800 V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
5801 V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5802 if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
5803 ctrl |= V_LSO_ETHHDR_LEN(1);
5804 if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5806 lso->lso_ctrl = htobe32(ctrl);
5807 lso->ipid_ofst = htobe16(0);
5808 lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5809 lso->seqno_offset = htobe32(0);
5810 lso->len = htobe32(pktlen);
5812 cpl = (void *)(lso + 1);
5814 wr->u.tcpseg.mss = htobe16(0xffff);
5815 cpl = (void *)(wr + 1);
5819 /* Checksum offload must be requested for ethofld. */
5821 MPASS(needs_l4_csum(m0));
5823 /* VLAN tag insertion */
5824 if (needs_vlan_insertion(m0)) {
5825 ctrl1 |= F_TXPKT_VLAN_VLD |
5826 V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5830 cpl->ctrl0 = cst->ctrl0;
5832 cpl->len = htobe16(pktlen);
5833 cpl->ctrl1 = htobe64(ctrl1);
5835 /* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
5836 p = (uintptr_t)(cpl + 1);
5837 m_copydata(m0, 0, immhdrs, (void *)p);
5840 dst = (void *)(cpl + 1);
5844 /* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
5846 pad = 16 - (immhdrs & 0xf);
5847 bzero((void *)p, pad);
5849 usgl = (void *)(p + pad);
5850 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5851 V_ULPTX_NSGE(nsegs));
5853 sglist_init(&sg, nitems(segs), segs);
5854 for (; m0 != NULL; m0 = m0->m_next) {
5855 if (__predict_false(m0->m_len == 0))
5857 if (immhdrs >= m0->m_len) {
5858 immhdrs -= m0->m_len;
5862 sglist_append(&sg, mtod(m0, char *) + immhdrs,
5863 m0->m_len - immhdrs);
5866 MPASS(sg.sg_nseg == nsegs);
5869 * Zero pad last 8B in case the WR doesn't end on a 16B
5872 *(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
5874 usgl->len0 = htobe32(segs[0].ss_len);
5875 usgl->addr0 = htobe64(segs[0].ss_paddr);
5876 for (i = 0; i < nsegs - 1; i++) {
5877 usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
5878 usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
5881 usgl->sge[i / 2].len[1] = htobe32(0);
5887 ethofld_tx(struct cxgbe_snd_tag *cst)
5890 struct wrq_cookie cookie;
5891 int next_credits, compl;
5892 struct fw_eth_tx_eo_wr *wr;
5894 mtx_assert(&cst->lock, MA_OWNED);
5896 while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
5899 /* How many len16 credits do we need to send this mbuf. */
5900 next_credits = mbuf_eo_len16(m);
5901 MPASS(next_credits > 0);
5902 if (next_credits > cst->tx_credits) {
5904 * Tx will make progress eventually because there is at
5905 * least one outstanding fw4_ack that will return
5906 * credits and kick the tx.
5908 MPASS(cst->ncompl > 0);
5911 wr = start_wrq_wr(cst->eo_txq, next_credits, &cookie);
5912 if (__predict_false(wr == NULL)) {
5913 /* XXX: wishful thinking, not a real assertion. */
5914 MPASS(cst->ncompl > 0);
5917 cst->tx_credits -= next_credits;
5918 cst->tx_nocompl += next_credits;
5919 compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
5920 ETHER_BPF_MTAP(cst->com.ifp, m);
5921 write_ethofld_wr(cst, wr, m, compl);
5922 commit_wrq_wr(cst->eo_txq, wr, &cookie);
5925 cst->tx_nocompl = 0;
5927 (void) mbufq_dequeue(&cst->pending_tx);
5930 * Drop the mbuf's reference on the tag now rather
5931 * than waiting until m_freem(). This ensures that
5932 * cxgbe_snd_tag_free gets called when the inp drops
5933 * its reference on the tag and there are no more
5934 * mbufs in the pending_tx queue and can flush any
5935 * pending requests. Otherwise if the last mbuf
5936 * doesn't request a completion the etid will never be
5939 m->m_pkthdr.snd_tag = NULL;
5940 m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
5941 m_snd_tag_rele(&cst->com);
5943 mbufq_enqueue(&cst->pending_fwack, m);
5948 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
5950 struct cxgbe_snd_tag *cst;
5953 MPASS(m0->m_nextpkt == NULL);
5954 MPASS(m0->m_pkthdr.csum_flags & CSUM_SND_TAG);
5955 MPASS(m0->m_pkthdr.snd_tag != NULL);
5956 cst = mst_to_cst(m0->m_pkthdr.snd_tag);
5958 mtx_lock(&cst->lock);
5959 MPASS(cst->flags & EO_SND_TAG_REF);
5961 if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
5962 struct vi_info *vi = ifp->if_softc;
5963 struct port_info *pi = vi->pi;
5964 struct adapter *sc = pi->adapter;
5965 const uint32_t rss_mask = vi->rss_size - 1;
5968 cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
5969 if (M_HASHTYPE_ISHASH(m0))
5970 rss_hash = m0->m_pkthdr.flowid;
5972 rss_hash = arc4random();
5973 /* We assume RSS hashing */
5974 cst->iqid = vi->rss[rss_hash & rss_mask];
5975 cst->eo_txq += rss_hash % vi->nofldtxq;
5976 rc = send_etid_flowc_wr(cst, pi, vi);
5981 if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
5986 mbufq_enqueue(&cst->pending_tx, m0);
5987 cst->plen += m0->m_pkthdr.len;
5990 * Hold an extra reference on the tag while generating work
5991 * requests to ensure that we don't try to free the tag during
5992 * ethofld_tx() in case we are sending the final mbuf after
5993 * the inp was freed.
5995 m_snd_tag_ref(&cst->com);
5997 mtx_unlock(&cst->lock);
5998 m_snd_tag_rele(&cst->com);
6002 mtx_unlock(&cst->lock);
6003 if (__predict_false(rc != 0))
6009 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
6011 struct adapter *sc = iq->adapter;
6012 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
6014 u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
6015 struct cxgbe_snd_tag *cst;
6016 uint8_t credits = cpl->credits;
6018 cst = lookup_etid(sc, etid);
6019 mtx_lock(&cst->lock);
6020 if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
6021 MPASS(credits >= ETID_FLOWC_LEN16);
6022 credits -= ETID_FLOWC_LEN16;
6023 cst->flags &= ~EO_FLOWC_RPL_PENDING;
6026 KASSERT(cst->ncompl > 0,
6027 ("%s: etid %u (%p) wasn't expecting completion.",
6028 __func__, etid, cst));
6031 while (credits > 0) {
6032 m = mbufq_dequeue(&cst->pending_fwack);
6033 if (__predict_false(m == NULL)) {
6035 * The remaining credits are for the final flush that
6036 * was issued when the tag was freed by the kernel.
6039 (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
6040 EO_FLUSH_RPL_PENDING);
6041 MPASS(credits == ETID_FLUSH_LEN16);
6042 MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
6043 MPASS(cst->ncompl == 0);
6045 cst->flags &= ~EO_FLUSH_RPL_PENDING;
6046 cst->tx_credits += cpl->credits;
6047 cxgbe_snd_tag_free_locked(cst);
6048 return (0); /* cst is gone. */
6051 ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6053 KASSERT(credits >= mbuf_eo_len16(m),
6054 ("%s: too few credits (%u, %u, %u)", __func__,
6055 cpl->credits, credits, mbuf_eo_len16(m)));
6056 credits -= mbuf_eo_len16(m);
6057 cst->plen -= m->m_pkthdr.len;
6061 cst->tx_credits += cpl->credits;
6062 MPASS(cst->tx_credits <= cst->tx_total);
6064 if (cst->flags & EO_SND_TAG_REF) {
6066 * As with ethofld_transmit(), hold an extra reference
6067 * so that the tag is stable across ethold_tx().
6069 m_snd_tag_ref(&cst->com);
6070 m = mbufq_first(&cst->pending_tx);
6071 if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6073 mtx_unlock(&cst->lock);
6074 m_snd_tag_rele(&cst->com);
6077 * There shouldn't be any pending packets if the tag
6078 * was freed by the kernel since any pending packet
6079 * should hold a reference to the tag.
6081 MPASS(mbufq_first(&cst->pending_tx) == NULL);
6082 mtx_unlock(&cst->lock);