1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (c) 2011, 2025 Chelsio Communications. 5 * Written by: Navdeep Parhar <np@FreeBSD.org> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 */ 28 29 #include <sys/cdefs.h> 30 #include "opt_inet.h" 31 #include "opt_inet6.h" 32 #include "opt_kern_tls.h" 33 #include "opt_ratelimit.h" 34 35 #include <sys/types.h> 36 #include <sys/eventhandler.h> 37 #include <sys/mbuf.h> 38 #include <sys/socket.h> 39 #include <sys/kernel.h> 40 #include <sys/ktls.h> 41 #include <sys/malloc.h> 42 #include <sys/msan.h> 43 #include <sys/queue.h> 44 #include <sys/sbuf.h> 45 #include <sys/taskqueue.h> 46 #include <sys/time.h> 47 #include <sys/sglist.h> 48 #include <sys/sysctl.h> 49 #include <sys/smp.h> 50 #include <sys/socketvar.h> 51 #include <sys/counter.h> 52 #include <net/bpf.h> 53 #include <net/ethernet.h> 54 #include <net/if.h> 55 #include <net/if_vlan_var.h> 56 #include <net/if_vxlan.h> 57 #include <netinet/in.h> 58 #include <netinet/ip.h> 59 #include <netinet/ip6.h> 60 #include <netinet/tcp.h> 61 #include <netinet/udp.h> 62 #include <machine/in_cksum.h> 63 #include <machine/md_var.h> 64 #include <vm/vm.h> 65 #include <vm/pmap.h> 66 #ifdef DEV_NETMAP 67 #include <machine/bus.h> 68 #include <sys/selinfo.h> 69 #include <net/if_var.h> 70 #include <net/netmap.h> 71 #include <dev/netmap/netmap_kern.h> 72 #endif 73 74 #include "common/common.h" 75 #include "common/t4_regs.h" 76 #include "common/t4_regs_values.h" 77 #include "common/t4_msg.h" 78 #include "t4_l2t.h" 79 #include "t4_mp_ring.h" 80 81 #define RX_COPY_THRESHOLD MINCLSIZE 82 83 /* 84 * Ethernet frames are DMA'd at this byte offset into the freelist buffer. 85 * 0-7 are valid values. 86 */ 87 static int fl_pktshift = 0; 88 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pktshift, CTLFLAG_RDTUN, &fl_pktshift, 0, 89 "payload DMA offset in rx buffer (bytes)"); 90 91 /* 92 * Pad ethernet payload up to this boundary. 93 * -1: driver should figure out a good value. 94 * 0: disable padding. 95 * Any power of 2 from 32 to 4096 (both inclusive) is also a valid value. 96 */ 97 int fl_pad = -1; 98 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pad, CTLFLAG_RDTUN, &fl_pad, 0, 99 "payload pad boundary (bytes)"); 100 101 /* 102 * Status page length. 103 * -1: driver should figure out a good value. 104 * 64 or 128 are the only other valid values. 105 */ 106 static int spg_len = -1; 107 SYSCTL_INT(_hw_cxgbe, OID_AUTO, spg_len, CTLFLAG_RDTUN, &spg_len, 0, 108 "status page size (bytes)"); 109 110 /* 111 * Congestion drops. 112 * -1: no congestion feedback (not recommended). 113 * 0: backpressure the channel instead of dropping packets right away. 114 * 1: no backpressure, drop packets for the congested queue immediately. 115 * 2: both backpressure and drop. 116 */ 117 static int cong_drop = 0; 118 SYSCTL_INT(_hw_cxgbe, OID_AUTO, cong_drop, CTLFLAG_RDTUN, &cong_drop, 0, 119 "Congestion control for NIC RX queues (0 = backpressure, 1 = drop, 2 = both"); 120 #ifdef TCP_OFFLOAD 121 static int ofld_cong_drop = 0; 122 SYSCTL_INT(_hw_cxgbe, OID_AUTO, ofld_cong_drop, CTLFLAG_RDTUN, &ofld_cong_drop, 0, 123 "Congestion control for TOE RX queues (0 = backpressure, 1 = drop, 2 = both"); 124 #endif 125 126 /* 127 * Deliver multiple frames in the same free list buffer if they fit. 128 * -1: let the driver decide whether to enable buffer packing or not. 129 * 0: disable buffer packing. 130 * 1: enable buffer packing. 131 */ 132 static int buffer_packing = -1; 133 SYSCTL_INT(_hw_cxgbe, OID_AUTO, buffer_packing, CTLFLAG_RDTUN, &buffer_packing, 134 0, "Enable buffer packing"); 135 136 /* 137 * Start next frame in a packed buffer at this boundary. 138 * -1: driver should figure out a good value. 139 * T4: driver will ignore this and use the same value as fl_pad above. 140 * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value. 141 */ 142 static int fl_pack = -1; 143 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pack, CTLFLAG_RDTUN, &fl_pack, 0, 144 "payload pack boundary (bytes)"); 145 146 /* 147 * Largest rx cluster size that the driver is allowed to allocate. 148 */ 149 static int largest_rx_cluster = MJUM16BYTES; 150 SYSCTL_INT(_hw_cxgbe, OID_AUTO, largest_rx_cluster, CTLFLAG_RDTUN, 151 &largest_rx_cluster, 0, "Largest rx cluster (bytes)"); 152 153 /* 154 * Size of cluster allocation that's most likely to succeed. The driver will 155 * fall back to this size if it fails to allocate clusters larger than this. 156 */ 157 static int safest_rx_cluster = PAGE_SIZE; 158 SYSCTL_INT(_hw_cxgbe, OID_AUTO, safest_rx_cluster, CTLFLAG_RDTUN, 159 &safest_rx_cluster, 0, "Safe rx cluster (bytes)"); 160 161 #ifdef RATELIMIT 162 /* 163 * Knob to control TCP timestamp rewriting, and the granularity of the tick used 164 * for rewriting. -1 and 0-3 are all valid values. 165 * -1: hardware should leave the TCP timestamps alone. 166 * 0: 1ms 167 * 1: 100us 168 * 2: 10us 169 * 3: 1us 170 */ 171 static int tsclk = -1; 172 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tsclk, CTLFLAG_RDTUN, &tsclk, 0, 173 "Control TCP timestamp rewriting when using pacing"); 174 175 static int eo_max_backlog = 1024 * 1024; 176 SYSCTL_INT(_hw_cxgbe, OID_AUTO, eo_max_backlog, CTLFLAG_RDTUN, &eo_max_backlog, 177 0, "Maximum backlog of ratelimited data per flow"); 178 #endif 179 180 /* 181 * The interrupt holdoff timers are multiplied by this value on T6+. 182 * 1 and 3-17 (both inclusive) are legal values. 183 */ 184 static int tscale = 1; 185 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tscale, CTLFLAG_RDTUN, &tscale, 0, 186 "Interrupt holdoff timer scale on T6+"); 187 188 /* 189 * Number of LRO entries in the lro_ctrl structure per rx queue. 190 */ 191 static int lro_entries = TCP_LRO_ENTRIES; 192 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_entries, CTLFLAG_RDTUN, &lro_entries, 0, 193 "Number of LRO entries per RX queue"); 194 195 /* 196 * This enables presorting of frames before they're fed into tcp_lro_rx. 197 */ 198 static int lro_mbufs = 0; 199 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_mbufs, CTLFLAG_RDTUN, &lro_mbufs, 0, 200 "Enable presorting of LRO frames"); 201 202 static counter_u64_t pullups; 203 SYSCTL_COUNTER_U64(_hw_cxgbe, OID_AUTO, pullups, CTLFLAG_RD, &pullups, 204 "Number of mbuf pullups performed"); 205 206 static counter_u64_t defrags; 207 SYSCTL_COUNTER_U64(_hw_cxgbe, OID_AUTO, defrags, CTLFLAG_RD, &defrags, 208 "Number of mbuf defrags performed"); 209 210 static int t4_tx_coalesce = 1; 211 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce, CTLFLAG_RWTUN, &t4_tx_coalesce, 0, 212 "tx coalescing allowed"); 213 214 /* 215 * The driver will make aggressive attempts at tx coalescing if it sees these 216 * many packets eligible for coalescing in quick succession, with no more than 217 * the specified gap in between the eth_tx calls that delivered the packets. 218 */ 219 static int t4_tx_coalesce_pkts = 32; 220 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce_pkts, CTLFLAG_RWTUN, 221 &t4_tx_coalesce_pkts, 0, 222 "# of consecutive packets (1 - 255) that will trigger tx coalescing"); 223 static int t4_tx_coalesce_gap = 5; 224 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce_gap, CTLFLAG_RWTUN, 225 &t4_tx_coalesce_gap, 0, "tx gap (in microseconds)"); 226 227 static int service_iq(struct sge_iq *, int); 228 static int service_iq_fl(struct sge_iq *, int); 229 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t); 230 static int eth_rx(struct adapter *, struct sge_rxq *, const struct iq_desc *, 231 u_int); 232 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int, 233 int, int, int); 234 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *); 235 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t, 236 struct sge_iq *, char *); 237 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *, 238 struct sysctl_ctx_list *, struct sysctl_oid *); 239 static void free_iq_fl(struct adapter *, struct sge_iq *, struct sge_fl *); 240 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *, 241 struct sge_iq *); 242 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *, 243 struct sysctl_oid *, struct sge_fl *); 244 static int alloc_iq_fl_hwq(struct vi_info *, struct sge_iq *, struct sge_fl *); 245 static int free_iq_fl_hwq(struct adapter *, struct sge_iq *, struct sge_fl *); 246 static int alloc_fwq(struct adapter *); 247 static void free_fwq(struct adapter *); 248 static int alloc_ctrlq(struct adapter *, int); 249 static void free_ctrlq(struct adapter *, int); 250 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int, int); 251 static void free_rxq(struct vi_info *, struct sge_rxq *); 252 static void add_rxq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *, 253 struct sge_rxq *); 254 #ifdef TCP_OFFLOAD 255 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int, 256 int); 257 static void free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *); 258 static void add_ofld_rxq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *, 259 struct sge_ofld_rxq *); 260 #endif 261 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *, int); 262 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *, 263 int); 264 #if defined(TCP_OFFLOAD) || defined(RATELIMIT) 265 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *, 266 int); 267 #endif 268 static int alloc_eq(struct adapter *, struct sge_eq *, struct sysctl_ctx_list *, 269 struct sysctl_oid *); 270 static void free_eq(struct adapter *, struct sge_eq *); 271 static void add_eq_sysctls(struct adapter *, struct sysctl_ctx_list *, 272 struct sysctl_oid *, struct sge_eq *); 273 static int alloc_eq_hwq(struct adapter *, struct vi_info *, struct sge_eq *, 274 int); 275 static int free_eq_hwq(struct adapter *, struct vi_info *, struct sge_eq *); 276 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *, 277 struct sysctl_ctx_list *, struct sysctl_oid *); 278 static void free_wrq(struct adapter *, struct sge_wrq *); 279 static void add_wrq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *, 280 struct sge_wrq *); 281 static int alloc_txq(struct vi_info *, struct sge_txq *, int); 282 static void free_txq(struct vi_info *, struct sge_txq *); 283 static void add_txq_sysctls(struct vi_info *, struct sysctl_ctx_list *, 284 struct sysctl_oid *, struct sge_txq *); 285 #if defined(TCP_OFFLOAD) || defined(RATELIMIT) 286 static int alloc_ofld_txq(struct vi_info *, struct sge_ofld_txq *, int); 287 static void free_ofld_txq(struct vi_info *, struct sge_ofld_txq *); 288 static void add_ofld_txq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *, 289 struct sge_ofld_txq *); 290 #endif 291 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int); 292 static inline void ring_fl_db(struct adapter *, struct sge_fl *); 293 static int refill_fl(struct adapter *, struct sge_fl *, int); 294 static void refill_sfl(void *); 295 static int find_refill_source(struct adapter *, int, bool); 296 static void add_fl_to_sfl(struct adapter *, struct sge_fl *); 297 298 static inline void get_pkt_gl(struct mbuf *, struct sglist *); 299 static inline u_int txpkt_len16(u_int, const u_int); 300 static inline u_int txpkt_vm_len16(u_int, const u_int); 301 static inline void calculate_mbuf_len16(struct mbuf *, bool); 302 static inline u_int txpkts0_len16(u_int); 303 static inline u_int txpkts1_len16(void); 304 static u_int write_raw_wr(struct sge_txq *, void *, struct mbuf *, u_int); 305 static u_int write_txpkt_wr(struct adapter *, struct sge_txq *, struct mbuf *, 306 u_int); 307 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *, 308 struct mbuf *); 309 static int add_to_txpkts_vf(struct adapter *, struct sge_txq *, struct mbuf *, 310 int, bool *); 311 static int add_to_txpkts_pf(struct adapter *, struct sge_txq *, struct mbuf *, 312 int, bool *); 313 static u_int write_txpkts_wr(struct adapter *, struct sge_txq *); 314 static u_int write_txpkts_vm_wr(struct adapter *, struct sge_txq *); 315 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int); 316 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int); 317 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int); 318 static inline uint16_t read_hw_cidx(struct sge_eq *); 319 static inline u_int reclaimable_tx_desc(struct sge_eq *); 320 static inline u_int total_available_tx_desc(struct sge_eq *); 321 static u_int reclaim_tx_descs(struct sge_txq *, u_int); 322 static void tx_reclaim(void *, int); 323 static __be64 get_flit(struct sglist_seg *, int, int); 324 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *, 325 struct mbuf *); 326 static int handle_fw_msg(struct sge_iq *, const struct rss_header *, 327 struct mbuf *); 328 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *); 329 static void wrq_tx_drain(void *, int); 330 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *); 331 332 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS); 333 #ifdef RATELIMIT 334 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *, 335 struct mbuf *); 336 #if defined(INET) || defined(INET6) 337 static inline u_int txpkt_eo_len16(u_int, u_int, u_int); 338 static int ethofld_transmit(if_t, struct mbuf *); 339 #endif 340 #endif 341 342 static counter_u64_t extfree_refs; 343 static counter_u64_t extfree_rels; 344 345 an_handler_t t4_an_handler; 346 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES]; 347 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS]; 348 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES]; 349 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES]; 350 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES]; 351 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES]; 352 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES]; 353 cpl_handler_t fw6_pld_handlers[NUM_CPL_FW6_COOKIES]; 354 355 void 356 t4_register_an_handler(an_handler_t h) 357 { 358 uintptr_t *loc; 359 360 MPASS(h == NULL || t4_an_handler == NULL); 361 362 loc = (uintptr_t *)&t4_an_handler; 363 atomic_store_rel_ptr(loc, (uintptr_t)h); 364 } 365 366 void 367 t4_register_fw_msg_handler(int type, fw_msg_handler_t h) 368 { 369 uintptr_t *loc; 370 371 MPASS(type < nitems(t4_fw_msg_handler)); 372 MPASS(h == NULL || t4_fw_msg_handler[type] == NULL); 373 /* 374 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL 375 * handler dispatch table. Reject any attempt to install a handler for 376 * this subtype. 377 */ 378 MPASS(type != FW_TYPE_RSSCPL); 379 MPASS(type != FW6_TYPE_RSSCPL); 380 381 loc = (uintptr_t *)&t4_fw_msg_handler[type]; 382 atomic_store_rel_ptr(loc, (uintptr_t)h); 383 } 384 385 void 386 t4_register_cpl_handler(int opcode, cpl_handler_t h) 387 { 388 uintptr_t *loc; 389 390 MPASS(opcode < nitems(t4_cpl_handler)); 391 MPASS(h == NULL || t4_cpl_handler[opcode] == NULL); 392 393 loc = (uintptr_t *)&t4_cpl_handler[opcode]; 394 atomic_store_rel_ptr(loc, (uintptr_t)h); 395 } 396 397 static int 398 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss, 399 struct mbuf *m) 400 { 401 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1); 402 u_int tid; 403 int cookie; 404 405 MPASS(m == NULL); 406 407 tid = GET_TID(cpl); 408 if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) { 409 /* 410 * The return code for filter-write is put in the CPL cookie so 411 * we have to rely on the hardware tid (is_ftid) to determine 412 * that this is a response to a filter. 413 */ 414 cookie = CPL_COOKIE_FILTER; 415 } else { 416 cookie = G_COOKIE(cpl->cookie); 417 } 418 MPASS(cookie > CPL_COOKIE_RESERVED); 419 MPASS(cookie < nitems(set_tcb_rpl_handlers)); 420 421 return (set_tcb_rpl_handlers[cookie](iq, rss, m)); 422 } 423 424 static int 425 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss, 426 struct mbuf *m) 427 { 428 const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1); 429 unsigned int cookie; 430 431 MPASS(m == NULL); 432 433 cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER; 434 return (l2t_write_rpl_handlers[cookie](iq, rss, m)); 435 } 436 437 static int 438 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss, 439 struct mbuf *m) 440 { 441 const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1); 442 u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status))); 443 444 MPASS(m == NULL); 445 MPASS(cookie != CPL_COOKIE_RESERVED); 446 447 return (act_open_rpl_handlers[cookie](iq, rss, m)); 448 } 449 450 static int 451 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss, 452 struct mbuf *m) 453 { 454 struct adapter *sc = iq->adapter; 455 u_int cookie; 456 457 MPASS(m == NULL); 458 if (is_hashfilter(sc)) 459 cookie = CPL_COOKIE_HASHFILTER; 460 else 461 cookie = CPL_COOKIE_TOM; 462 463 return (abort_rpl_rss_handlers[cookie](iq, rss, m)); 464 } 465 466 static int 467 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 468 { 469 struct adapter *sc = iq->adapter; 470 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1); 471 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl))); 472 u_int cookie; 473 474 MPASS(m == NULL); 475 if (is_etid(sc, tid)) 476 cookie = CPL_COOKIE_ETHOFLD; 477 else 478 cookie = CPL_COOKIE_TOM; 479 480 return (fw4_ack_handlers[cookie](iq, rss, m)); 481 } 482 483 static int 484 fw6_pld_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 485 { 486 const struct cpl_fw6_pld *cpl; 487 uint64_t cookie; 488 489 if (m != NULL) 490 cpl = mtod(m, const void *); 491 else 492 cpl = (const void *)(rss + 1); 493 cookie = be64toh(cpl->data[1]) & CPL_FW6_COOKIE_MASK; 494 495 return (fw6_pld_handlers[cookie](iq, rss, m)); 496 } 497 498 static void 499 t4_init_shared_cpl_handlers(void) 500 { 501 502 t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler); 503 t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler); 504 t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler); 505 t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler); 506 t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler); 507 t4_register_cpl_handler(CPL_FW6_PLD, fw6_pld_handler); 508 } 509 510 void 511 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie) 512 { 513 uintptr_t *loc; 514 515 MPASS(opcode < nitems(t4_cpl_handler)); 516 if (opcode == CPL_FW6_PLD) { 517 MPASS(cookie < NUM_CPL_FW6_COOKIES); 518 } else { 519 MPASS(cookie > CPL_COOKIE_RESERVED); 520 MPASS(cookie < NUM_CPL_COOKIES); 521 } 522 MPASS(t4_cpl_handler[opcode] != NULL); 523 524 switch (opcode) { 525 case CPL_SET_TCB_RPL: 526 loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie]; 527 break; 528 case CPL_L2T_WRITE_RPL: 529 loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie]; 530 break; 531 case CPL_ACT_OPEN_RPL: 532 loc = (uintptr_t *)&act_open_rpl_handlers[cookie]; 533 break; 534 case CPL_ABORT_RPL_RSS: 535 loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie]; 536 break; 537 case CPL_FW4_ACK: 538 loc = (uintptr_t *)&fw4_ack_handlers[cookie]; 539 break; 540 case CPL_FW6_PLD: 541 loc = (uintptr_t *)&fw6_pld_handlers[cookie]; 542 break; 543 default: 544 MPASS(0); 545 return; 546 } 547 MPASS(h == NULL || *loc == (uintptr_t)NULL); 548 atomic_store_rel_ptr(loc, (uintptr_t)h); 549 } 550 551 /* 552 * Called on MOD_LOAD. Validates and calculates the SGE tunables. 553 */ 554 void 555 t4_sge_modload(void) 556 { 557 558 if (fl_pktshift < 0 || fl_pktshift > 7) { 559 printf("Invalid hw.cxgbe.fl_pktshift value (%d)," 560 " using 0 instead.\n", fl_pktshift); 561 fl_pktshift = 0; 562 } 563 564 if (spg_len != 64 && spg_len != 128) { 565 int len; 566 567 #if defined(__i386__) || defined(__amd64__) 568 len = cpu_clflush_line_size > 64 ? 128 : 64; 569 #else 570 len = 64; 571 #endif 572 if (spg_len != -1) { 573 printf("Invalid hw.cxgbe.spg_len value (%d)," 574 " using %d instead.\n", spg_len, len); 575 } 576 spg_len = len; 577 } 578 579 if (cong_drop < -1 || cong_drop > 2) { 580 printf("Invalid hw.cxgbe.cong_drop value (%d)," 581 " using 0 instead.\n", cong_drop); 582 cong_drop = 0; 583 } 584 #ifdef TCP_OFFLOAD 585 if (ofld_cong_drop < -1 || ofld_cong_drop > 2) { 586 printf("Invalid hw.cxgbe.ofld_cong_drop value (%d)," 587 " using 0 instead.\n", ofld_cong_drop); 588 ofld_cong_drop = 0; 589 } 590 #endif 591 592 if (tscale != 1 && (tscale < 3 || tscale > 17)) { 593 printf("Invalid hw.cxgbe.tscale value (%d)," 594 " using 1 instead.\n", tscale); 595 tscale = 1; 596 } 597 598 if (largest_rx_cluster != MCLBYTES && 599 #if MJUMPAGESIZE != MCLBYTES 600 largest_rx_cluster != MJUMPAGESIZE && 601 #endif 602 largest_rx_cluster != MJUM9BYTES && 603 largest_rx_cluster != MJUM16BYTES) { 604 printf("Invalid hw.cxgbe.largest_rx_cluster value (%d)," 605 " using %d instead.\n", largest_rx_cluster, MJUM16BYTES); 606 largest_rx_cluster = MJUM16BYTES; 607 } 608 609 if (safest_rx_cluster != MCLBYTES && 610 #if MJUMPAGESIZE != MCLBYTES 611 safest_rx_cluster != MJUMPAGESIZE && 612 #endif 613 safest_rx_cluster != MJUM9BYTES && 614 safest_rx_cluster != MJUM16BYTES) { 615 printf("Invalid hw.cxgbe.safest_rx_cluster value (%d)," 616 " using %d instead.\n", safest_rx_cluster, MJUMPAGESIZE); 617 safest_rx_cluster = MJUMPAGESIZE; 618 } 619 620 extfree_refs = counter_u64_alloc(M_WAITOK); 621 extfree_rels = counter_u64_alloc(M_WAITOK); 622 pullups = counter_u64_alloc(M_WAITOK); 623 defrags = counter_u64_alloc(M_WAITOK); 624 counter_u64_zero(extfree_refs); 625 counter_u64_zero(extfree_rels); 626 counter_u64_zero(pullups); 627 counter_u64_zero(defrags); 628 629 t4_init_shared_cpl_handlers(); 630 t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg); 631 t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg); 632 t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update); 633 #ifdef RATELIMIT 634 t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack, 635 CPL_COOKIE_ETHOFLD); 636 #endif 637 t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl); 638 t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl); 639 } 640 641 void 642 t4_sge_modunload(void) 643 { 644 645 counter_u64_free(extfree_refs); 646 counter_u64_free(extfree_rels); 647 counter_u64_free(pullups); 648 counter_u64_free(defrags); 649 } 650 651 uint64_t 652 t4_sge_extfree_refs(void) 653 { 654 uint64_t refs, rels; 655 656 rels = counter_u64_fetch(extfree_rels); 657 refs = counter_u64_fetch(extfree_refs); 658 659 return (refs - rels); 660 } 661 662 /* max 4096 */ 663 #define MAX_PACK_BOUNDARY 512 664 665 static inline void 666 setup_pad_and_pack_boundaries(struct adapter *sc) 667 { 668 uint32_t v, m; 669 int pad, pack, pad_shift; 670 671 pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT : 672 X_INGPADBOUNDARY_SHIFT; 673 pad = fl_pad; 674 if (fl_pad < (1 << pad_shift) || 675 fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) || 676 !powerof2(fl_pad)) { 677 /* 678 * If there is any chance that we might use buffer packing and 679 * the chip is a T4, then pick 64 as the pad/pack boundary. Set 680 * it to the minimum allowed in all other cases. 681 */ 682 pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift; 683 684 /* 685 * For fl_pad = 0 we'll still write a reasonable value to the 686 * register but all the freelists will opt out of padding. 687 * We'll complain here only if the user tried to set it to a 688 * value greater than 0 that was invalid. 689 */ 690 if (fl_pad > 0) { 691 device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value" 692 " (%d), using %d instead.\n", fl_pad, pad); 693 } 694 } 695 m = V_INGPADBOUNDARY(M_INGPADBOUNDARY); 696 v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift); 697 t4_set_reg_field(sc, A_SGE_CONTROL, m, v); 698 699 if (is_t4(sc)) { 700 if (fl_pack != -1 && fl_pack != pad) { 701 /* Complain but carry on. */ 702 device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored," 703 " using %d instead.\n", fl_pack, pad); 704 } 705 return; 706 } 707 708 pack = fl_pack; 709 if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 || 710 !powerof2(fl_pack)) { 711 if (sc->params.pci.mps > MAX_PACK_BOUNDARY) 712 pack = MAX_PACK_BOUNDARY; 713 else 714 pack = max(sc->params.pci.mps, CACHE_LINE_SIZE); 715 MPASS(powerof2(pack)); 716 if (pack < 16) 717 pack = 16; 718 if (pack == 32) 719 pack = 64; 720 if (pack > 4096) 721 pack = 4096; 722 if (fl_pack != -1) { 723 device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value" 724 " (%d), using %d instead.\n", fl_pack, pack); 725 } 726 } 727 m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY); 728 if (pack == 16) 729 v = V_INGPACKBOUNDARY(0); 730 else 731 v = V_INGPACKBOUNDARY(ilog2(pack) - 5); 732 733 MPASS(!is_t4(sc)); /* T4 doesn't have SGE_CONTROL2 */ 734 t4_set_reg_field(sc, A_SGE_CONTROL2, m, v); 735 } 736 737 /* 738 * adap->params.vpd.cclk must be set up before this is called. 739 */ 740 void 741 t4_tweak_chip_settings(struct adapter *sc) 742 { 743 int i, reg; 744 uint32_t v, m; 745 int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200}; 746 int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk; 747 int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */ 748 uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE); 749 static int sw_buf_sizes[] = { 750 MCLBYTES, 751 #if MJUMPAGESIZE != MCLBYTES 752 MJUMPAGESIZE, 753 #endif 754 MJUM9BYTES, 755 MJUM16BYTES 756 }; 757 758 KASSERT(sc->flags & MASTER_PF, 759 ("%s: trying to change chip settings when not master.", __func__)); 760 761 m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE; 762 v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE | 763 V_EGRSTATUSPAGESIZE(spg_len == 128); 764 t4_set_reg_field(sc, A_SGE_CONTROL, m, v); 765 766 setup_pad_and_pack_boundaries(sc); 767 768 v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) | 769 V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) | 770 V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) | 771 V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) | 772 V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) | 773 V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) | 774 V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) | 775 V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10); 776 t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v); 777 778 t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, 4096); 779 t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE1, 65536); 780 reg = A_SGE_FL_BUFFER_SIZE2; 781 for (i = 0; i < nitems(sw_buf_sizes); i++) { 782 MPASS(reg <= A_SGE_FL_BUFFER_SIZE15); 783 t4_write_reg(sc, reg, sw_buf_sizes[i]); 784 reg += 4; 785 MPASS(reg <= A_SGE_FL_BUFFER_SIZE15); 786 t4_write_reg(sc, reg, sw_buf_sizes[i] - CL_METADATA_SIZE); 787 reg += 4; 788 } 789 790 v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) | 791 V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]); 792 t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v); 793 794 KASSERT(intr_timer[0] <= timer_max, 795 ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0], 796 timer_max)); 797 for (i = 1; i < nitems(intr_timer); i++) { 798 KASSERT(intr_timer[i] >= intr_timer[i - 1], 799 ("%s: timers not listed in increasing order (%d)", 800 __func__, i)); 801 802 while (intr_timer[i] > timer_max) { 803 if (i == nitems(intr_timer) - 1) { 804 intr_timer[i] = timer_max; 805 break; 806 } 807 intr_timer[i] += intr_timer[i - 1]; 808 intr_timer[i] /= 2; 809 } 810 } 811 812 v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) | 813 V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1])); 814 t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v); 815 v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) | 816 V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3])); 817 t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v); 818 v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) | 819 V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5])); 820 t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v); 821 822 if (chip_id(sc) >= CHELSIO_T6) { 823 m = V_TSCALE(M_TSCALE); 824 if (tscale == 1) 825 v = 0; 826 else 827 v = V_TSCALE(tscale - 2); 828 t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v); 829 830 if (sc->debug_flags & DF_DISABLE_TCB_CACHE) { 831 t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1); 832 if (chip_id(sc) >= CHELSIO_T7) { 833 v |= F_GLFL; 834 } else { 835 m = V_RDTHRESHOLD(M_RDTHRESHOLD) | 836 F_WRTHRTHRESHEN | 837 V_WRTHRTHRESH(M_WRTHRTHRESH); 838 v &= ~m; 839 v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN | 840 V_WRTHRTHRESH(16); 841 } 842 t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1); 843 } 844 } 845 846 /* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */ 847 v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6); 848 t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v); 849 850 /* 851 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP. These have been 852 * chosen with MAXPHYS = 128K in mind. The largest DDP buffer that we 853 * may have to deal with is MAXPHYS + 1 page. 854 */ 855 v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4); 856 t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v); 857 858 /* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */ 859 m = v = F_TDDPTAGTCB | F_ISCSITAGTCB; 860 if (sc->nvmecaps != 0) { 861 /* Request DDP status bit for NVMe PDU completions. */ 862 m |= F_NVME_TCP_DDP_VAL_EN; 863 v |= F_NVME_TCP_DDP_VAL_EN; 864 } 865 t4_set_reg_field(sc, A_ULP_RX_CTL, m, v); 866 867 m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET | 868 F_RESETDDPOFFSET; 869 v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET; 870 t4_set_reg_field(sc, A_TP_PARA_REG5, m, v); 871 } 872 873 /* 874 * SGE wants the buffer to be at least 64B and then a multiple of 16. Its 875 * address mut be 16B aligned. If padding is in use the buffer's start and end 876 * need to be aligned to the pad boundary as well. We'll just make sure that 877 * the size is a multiple of the pad boundary here, it is up to the buffer 878 * allocation code to make sure the start of the buffer is aligned. 879 */ 880 static inline int 881 hwsz_ok(struct adapter *sc, int hwsz) 882 { 883 int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1; 884 885 return (hwsz >= 64 && (hwsz & mask) == 0); 886 } 887 888 /* 889 * Initialize the rx buffer sizes and figure out which zones the buffers will 890 * be allocated from. 891 */ 892 void 893 t4_init_rx_buf_info(struct adapter *sc) 894 { 895 struct sge *s = &sc->sge; 896 struct sge_params *sp = &sc->params.sge; 897 int i, j, n; 898 static int sw_buf_sizes[] = { /* Sorted by size */ 899 MCLBYTES, 900 #if MJUMPAGESIZE != MCLBYTES 901 MJUMPAGESIZE, 902 #endif 903 MJUM9BYTES, 904 MJUM16BYTES 905 }; 906 struct rx_buf_info *rxb; 907 908 s->safe_zidx = -1; 909 rxb = &s->rx_buf_info[0]; 910 for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) { 911 rxb->size1 = sw_buf_sizes[i]; 912 rxb->zone = m_getzone(rxb->size1); 913 rxb->type = m_gettype(rxb->size1); 914 rxb->size2 = 0; 915 rxb->hwidx1 = -1; 916 rxb->hwidx2 = -1; 917 for (j = 0; j < SGE_FLBUF_SIZES; j++) { 918 int hwsize = sp->sge_fl_buffer_size[j]; 919 920 if (!hwsz_ok(sc, hwsize)) 921 continue; 922 923 /* hwidx for size1 */ 924 if (rxb->hwidx1 == -1 && rxb->size1 == hwsize) 925 rxb->hwidx1 = j; 926 927 /* hwidx for size2 (buffer packing) */ 928 if (rxb->size1 - CL_METADATA_SIZE < hwsize) 929 continue; 930 n = rxb->size1 - hwsize - CL_METADATA_SIZE; 931 if (n == 0) { 932 rxb->hwidx2 = j; 933 rxb->size2 = hwsize; 934 break; /* stop looking */ 935 } 936 if (rxb->hwidx2 != -1) { 937 if (n < sp->sge_fl_buffer_size[rxb->hwidx2] - 938 hwsize - CL_METADATA_SIZE) { 939 rxb->hwidx2 = j; 940 rxb->size2 = hwsize; 941 } 942 } else if (n <= 2 * CL_METADATA_SIZE) { 943 rxb->hwidx2 = j; 944 rxb->size2 = hwsize; 945 } 946 } 947 if (rxb->hwidx2 != -1) 948 sc->flags |= BUF_PACKING_OK; 949 if (s->safe_zidx == -1 && rxb->size1 == safest_rx_cluster) 950 s->safe_zidx = i; 951 } 952 } 953 954 /* 955 * Verify some basic SGE settings for the PF and VF driver, and other 956 * miscellaneous settings for the PF driver. 957 */ 958 int 959 t4_verify_chip_settings(struct adapter *sc) 960 { 961 struct sge_params *sp = &sc->params.sge; 962 uint32_t m, v, r; 963 int rc = 0; 964 const uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE); 965 966 m = F_RXPKTCPLMODE; 967 v = F_RXPKTCPLMODE; 968 r = sp->sge_control; 969 if ((r & m) != v) { 970 device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r); 971 rc = EINVAL; 972 } 973 974 /* 975 * If this changes then every single use of PAGE_SHIFT in the driver 976 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift. 977 */ 978 if (sp->page_shift != PAGE_SHIFT) { 979 device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r); 980 rc = EINVAL; 981 } 982 983 if (sc->flags & IS_VF) 984 return (0); 985 986 v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6); 987 r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ); 988 if (r != v) { 989 device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r); 990 if (sc->vres.ddp.size != 0) 991 rc = EINVAL; 992 } 993 994 m = v = F_TDDPTAGTCB; 995 r = t4_read_reg(sc, A_ULP_RX_CTL); 996 if ((r & m) != v) { 997 device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r); 998 if (sc->vres.ddp.size != 0) 999 rc = EINVAL; 1000 } 1001 1002 m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET | 1003 F_RESETDDPOFFSET; 1004 v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET; 1005 r = t4_read_reg(sc, A_TP_PARA_REG5); 1006 if ((r & m) != v) { 1007 device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r); 1008 if (sc->vres.ddp.size != 0) 1009 rc = EINVAL; 1010 } 1011 1012 return (rc); 1013 } 1014 1015 int 1016 t4_create_dma_tag(struct adapter *sc) 1017 { 1018 int rc; 1019 1020 rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0, 1021 BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE, 1022 BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL, 1023 NULL, &sc->dmat); 1024 if (rc != 0) { 1025 device_printf(sc->dev, 1026 "failed to create main DMA tag: %d\n", rc); 1027 } 1028 1029 return (rc); 1030 } 1031 1032 void 1033 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx, 1034 struct sysctl_oid_list *children) 1035 { 1036 struct sge_params *sp = &sc->params.sge; 1037 1038 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes", 1039 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, sc, 0, 1040 sysctl_bufsizes, "A", "freelist buffer sizes"); 1041 1042 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD, 1043 NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)"); 1044 1045 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD, 1046 NULL, sp->pad_boundary, "payload pad boundary (bytes)"); 1047 1048 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD, 1049 NULL, sp->spg_len, "status page size (bytes)"); 1050 1051 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD, 1052 NULL, cong_drop, "congestion drop setting"); 1053 #ifdef TCP_OFFLOAD 1054 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "ofld_cong_drop", CTLFLAG_RD, 1055 NULL, ofld_cong_drop, "congestion drop setting"); 1056 #endif 1057 1058 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD, 1059 NULL, sp->pack_boundary, "payload pack boundary (bytes)"); 1060 } 1061 1062 int 1063 t4_destroy_dma_tag(struct adapter *sc) 1064 { 1065 if (sc->dmat) 1066 bus_dma_tag_destroy(sc->dmat); 1067 1068 return (0); 1069 } 1070 1071 /* 1072 * Allocate and initialize the firmware event queue, control queues, and special 1073 * purpose rx queues owned by the adapter. 1074 * 1075 * Returns errno on failure. Resources allocated up to that point may still be 1076 * allocated. Caller is responsible for cleanup in case this function fails. 1077 */ 1078 int 1079 t4_setup_adapter_queues(struct adapter *sc) 1080 { 1081 int rc, i; 1082 1083 ADAPTER_LOCK_ASSERT_NOTOWNED(sc); 1084 1085 /* 1086 * Firmware event queue 1087 */ 1088 rc = alloc_fwq(sc); 1089 if (rc != 0) 1090 return (rc); 1091 1092 /* 1093 * That's all for the VF driver. 1094 */ 1095 if (sc->flags & IS_VF) 1096 return (rc); 1097 1098 /* 1099 * XXX: General purpose rx queues, one per port. 1100 */ 1101 1102 /* 1103 * Control queues. At least one per port and per internal core. 1104 */ 1105 for (i = 0; i < sc->sge.nctrlq; i++) { 1106 rc = alloc_ctrlq(sc, i); 1107 if (rc != 0) 1108 return (rc); 1109 } 1110 1111 return (rc); 1112 } 1113 1114 /* 1115 * Idempotent 1116 */ 1117 int 1118 t4_teardown_adapter_queues(struct adapter *sc) 1119 { 1120 int i; 1121 1122 ADAPTER_LOCK_ASSERT_NOTOWNED(sc); 1123 1124 if (sc->sge.ctrlq != NULL) { 1125 MPASS(!(sc->flags & IS_VF)); /* VFs don't allocate ctrlq. */ 1126 for (i = 0; i < sc->sge.nctrlq; i++) 1127 free_ctrlq(sc, i); 1128 } 1129 free_fwq(sc); 1130 1131 return (0); 1132 } 1133 1134 /* Maximum payload that could arrive with a single iq descriptor. */ 1135 static inline int 1136 max_rx_payload(struct adapter *sc, if_t ifp, const bool ofld) 1137 { 1138 int maxp; 1139 1140 /* large enough even when hw VLAN extraction is disabled */ 1141 maxp = sc->params.sge.fl_pktshift + ETHER_HDR_LEN + 1142 ETHER_VLAN_ENCAP_LEN + if_getmtu(ifp); 1143 if (ofld && sc->tt.tls && sc->cryptocaps & FW_CAPS_CONFIG_TLSKEYS && 1144 maxp < sc->params.tp.max_rx_pdu) 1145 maxp = sc->params.tp.max_rx_pdu; 1146 return (maxp); 1147 } 1148 1149 int 1150 t4_setup_vi_queues(struct vi_info *vi) 1151 { 1152 int rc = 0, i, intr_idx; 1153 struct sge_rxq *rxq; 1154 struct sge_txq *txq; 1155 #ifdef TCP_OFFLOAD 1156 struct sge_ofld_rxq *ofld_rxq; 1157 #endif 1158 #if defined(TCP_OFFLOAD) || defined(RATELIMIT) 1159 struct sge_ofld_txq *ofld_txq; 1160 #endif 1161 #ifdef DEV_NETMAP 1162 int saved_idx, iqidx; 1163 struct sge_nm_rxq *nm_rxq; 1164 struct sge_nm_txq *nm_txq; 1165 #endif 1166 struct adapter *sc = vi->adapter; 1167 if_t ifp = vi->ifp; 1168 int maxp; 1169 1170 /* Interrupt vector to start from (when using multiple vectors) */ 1171 intr_idx = vi->first_intr; 1172 1173 #ifdef DEV_NETMAP 1174 saved_idx = intr_idx; 1175 if (if_getcapabilities(ifp) & IFCAP_NETMAP) { 1176 1177 /* netmap is supported with direct interrupts only. */ 1178 MPASS(!forwarding_intr_to_fwq(sc)); 1179 MPASS(vi->first_intr >= 0); 1180 1181 /* 1182 * We don't have buffers to back the netmap rx queues 1183 * right now so we create the queues in a way that 1184 * doesn't set off any congestion signal in the chip. 1185 */ 1186 for_each_nm_rxq(vi, i, nm_rxq) { 1187 rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i); 1188 if (rc != 0) 1189 goto done; 1190 intr_idx++; 1191 } 1192 1193 for_each_nm_txq(vi, i, nm_txq) { 1194 iqidx = vi->first_nm_rxq + (i % vi->nnmrxq); 1195 rc = alloc_nm_txq(vi, nm_txq, iqidx, i); 1196 if (rc != 0) 1197 goto done; 1198 } 1199 } 1200 1201 /* Normal rx queues and netmap rx queues share the same interrupts. */ 1202 intr_idx = saved_idx; 1203 #endif 1204 1205 /* 1206 * Allocate rx queues first because a default iqid is required when 1207 * creating a tx queue. 1208 */ 1209 maxp = max_rx_payload(sc, ifp, false); 1210 for_each_rxq(vi, i, rxq) { 1211 rc = alloc_rxq(vi, rxq, i, intr_idx, maxp); 1212 if (rc != 0) 1213 goto done; 1214 if (!forwarding_intr_to_fwq(sc)) 1215 intr_idx++; 1216 } 1217 #ifdef DEV_NETMAP 1218 if (if_getcapabilities(ifp) & IFCAP_NETMAP) 1219 intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq); 1220 #endif 1221 #ifdef TCP_OFFLOAD 1222 maxp = max_rx_payload(sc, ifp, true); 1223 for_each_ofld_rxq(vi, i, ofld_rxq) { 1224 rc = alloc_ofld_rxq(vi, ofld_rxq, i, intr_idx, maxp); 1225 if (rc != 0) 1226 goto done; 1227 if (!forwarding_intr_to_fwq(sc)) 1228 intr_idx++; 1229 } 1230 #endif 1231 1232 /* 1233 * Now the tx queues. 1234 */ 1235 for_each_txq(vi, i, txq) { 1236 rc = alloc_txq(vi, txq, i); 1237 if (rc != 0) 1238 goto done; 1239 } 1240 #if defined(TCP_OFFLOAD) || defined(RATELIMIT) 1241 for_each_ofld_txq(vi, i, ofld_txq) { 1242 rc = alloc_ofld_txq(vi, ofld_txq, i); 1243 if (rc != 0) 1244 goto done; 1245 } 1246 #endif 1247 done: 1248 if (rc) 1249 t4_teardown_vi_queues(vi); 1250 1251 return (rc); 1252 } 1253 1254 /* 1255 * Idempotent 1256 */ 1257 int 1258 t4_teardown_vi_queues(struct vi_info *vi) 1259 { 1260 int i; 1261 struct sge_rxq *rxq; 1262 struct sge_txq *txq; 1263 #if defined(TCP_OFFLOAD) || defined(RATELIMIT) 1264 struct sge_ofld_txq *ofld_txq; 1265 #endif 1266 #ifdef TCP_OFFLOAD 1267 struct sge_ofld_rxq *ofld_rxq; 1268 #endif 1269 #ifdef DEV_NETMAP 1270 struct sge_nm_rxq *nm_rxq; 1271 struct sge_nm_txq *nm_txq; 1272 #endif 1273 1274 #ifdef DEV_NETMAP 1275 if (if_getcapabilities(vi->ifp) & IFCAP_NETMAP) { 1276 for_each_nm_txq(vi, i, nm_txq) { 1277 free_nm_txq(vi, nm_txq); 1278 } 1279 1280 for_each_nm_rxq(vi, i, nm_rxq) { 1281 free_nm_rxq(vi, nm_rxq); 1282 } 1283 } 1284 #endif 1285 1286 /* 1287 * Take down all the tx queues first, as they reference the rx queues 1288 * (for egress updates, etc.). 1289 */ 1290 1291 for_each_txq(vi, i, txq) { 1292 free_txq(vi, txq); 1293 } 1294 #if defined(TCP_OFFLOAD) || defined(RATELIMIT) 1295 for_each_ofld_txq(vi, i, ofld_txq) { 1296 free_ofld_txq(vi, ofld_txq); 1297 } 1298 #endif 1299 1300 /* 1301 * Then take down the rx queues. 1302 */ 1303 1304 for_each_rxq(vi, i, rxq) { 1305 free_rxq(vi, rxq); 1306 } 1307 #ifdef TCP_OFFLOAD 1308 for_each_ofld_rxq(vi, i, ofld_rxq) { 1309 free_ofld_rxq(vi, ofld_rxq); 1310 } 1311 #endif 1312 1313 return (0); 1314 } 1315 1316 /* 1317 * Interrupt handler when the driver is using only 1 interrupt. This is a very 1318 * unusual scenario. 1319 * 1320 * a) Deals with errors, if any. 1321 * b) Services firmware event queue, which is taking interrupts for all other 1322 * queues. 1323 */ 1324 void 1325 t4_intr_all(void *arg) 1326 { 1327 struct adapter *sc = arg; 1328 struct sge_iq *fwq = &sc->sge.fwq; 1329 1330 MPASS(sc->intr_count == 1); 1331 1332 if (sc->intr_type == INTR_INTX) 1333 t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0); 1334 1335 t4_intr_err(arg); 1336 t4_intr_evt(fwq); 1337 } 1338 1339 /* 1340 * Interrupt handler for errors (installed directly when multiple interrupts are 1341 * being used, or called by t4_intr_all). 1342 */ 1343 void 1344 t4_intr_err(void *arg) 1345 { 1346 struct adapter *sc = arg; 1347 uint32_t v; 1348 1349 if (atomic_load_int(&sc->error_flags) & ADAP_FATAL_ERR) 1350 return; 1351 1352 v = t4_read_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE)); 1353 if (v & F_PFSW) { 1354 sc->swintr++; 1355 t4_write_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE), v); 1356 } 1357 1358 if (t4_slow_intr_handler(sc, sc->intr_flags)) 1359 t4_fatal_err(sc, false); 1360 } 1361 1362 /* 1363 * Interrupt handler for iq-only queues. The firmware event queue is the only 1364 * such queue right now. 1365 */ 1366 void 1367 t4_intr_evt(void *arg) 1368 { 1369 struct sge_iq *iq = arg; 1370 1371 if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) { 1372 service_iq(iq, 0); 1373 (void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE); 1374 } 1375 } 1376 1377 /* 1378 * Interrupt handler for iq+fl queues. 1379 */ 1380 void 1381 t4_intr(void *arg) 1382 { 1383 struct sge_iq *iq = arg; 1384 1385 if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) { 1386 service_iq_fl(iq, 0); 1387 (void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE); 1388 } 1389 } 1390 1391 #ifdef DEV_NETMAP 1392 /* 1393 * Interrupt handler for netmap rx queues. 1394 */ 1395 void 1396 t4_nm_intr(void *arg) 1397 { 1398 struct sge_nm_rxq *nm_rxq = arg; 1399 1400 if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) { 1401 service_nm_rxq(nm_rxq); 1402 (void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON); 1403 } 1404 } 1405 1406 /* 1407 * Interrupt handler for vectors shared between NIC and netmap rx queues. 1408 */ 1409 void 1410 t4_vi_intr(void *arg) 1411 { 1412 struct irq *irq = arg; 1413 1414 MPASS(irq->nm_rxq != NULL); 1415 t4_nm_intr(irq->nm_rxq); 1416 1417 MPASS(irq->rxq != NULL); 1418 t4_intr(irq->rxq); 1419 } 1420 #endif 1421 1422 /* 1423 * Deals with interrupts on an iq-only (no freelist) queue. 1424 */ 1425 static int 1426 service_iq(struct sge_iq *iq, int budget) 1427 { 1428 struct sge_iq *q; 1429 struct adapter *sc = iq->adapter; 1430 struct iq_desc *d = &iq->desc[iq->cidx]; 1431 int ndescs = 0, limit; 1432 int rsp_type; 1433 uint32_t lq; 1434 STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql); 1435 1436 KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq)); 1437 KASSERT((iq->flags & IQ_HAS_FL) == 0, 1438 ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq, 1439 iq->flags)); 1440 MPASS((iq->flags & IQ_ADJ_CREDIT) == 0); 1441 MPASS((iq->flags & IQ_LRO_ENABLED) == 0); 1442 1443 limit = budget ? budget : iq->qsize / 16; 1444 1445 /* 1446 * We always come back and check the descriptor ring for new indirect 1447 * interrupts and other responses after running a single handler. 1448 */ 1449 for (;;) { 1450 while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) { 1451 1452 rmb(); 1453 1454 rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen); 1455 lq = be32toh(d->rsp.pldbuflen_qid); 1456 1457 switch (rsp_type) { 1458 case X_RSPD_TYPE_FLBUF: 1459 panic("%s: data for an iq (%p) with no freelist", 1460 __func__, iq); 1461 1462 /* NOTREACHED */ 1463 1464 case X_RSPD_TYPE_CPL: 1465 KASSERT(d->rss.opcode < NUM_CPL_CMDS, 1466 ("%s: bad opcode %02x.", __func__, 1467 d->rss.opcode)); 1468 t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL); 1469 break; 1470 1471 case X_RSPD_TYPE_INTR: 1472 /* 1473 * There are 1K interrupt-capable queues (qids 0 1474 * through 1023). A response type indicating a 1475 * forwarded interrupt with a qid >= 1K is an 1476 * iWARP async notification. 1477 */ 1478 if (__predict_true(lq >= 1024)) { 1479 t4_an_handler(iq, &d->rsp); 1480 break; 1481 } 1482 1483 q = sc->sge.iqmap[lq - sc->sge.iq_start - 1484 sc->sge.iq_base]; 1485 if (atomic_cmpset_int(&q->state, IQS_IDLE, 1486 IQS_BUSY)) { 1487 if (service_iq_fl(q, q->qsize / 16) == 0) { 1488 (void) atomic_cmpset_int(&q->state, 1489 IQS_BUSY, IQS_IDLE); 1490 } else { 1491 STAILQ_INSERT_TAIL(&iql, q, 1492 link); 1493 } 1494 } 1495 break; 1496 1497 default: 1498 KASSERT(0, 1499 ("%s: illegal response type %d on iq %p", 1500 __func__, rsp_type, iq)); 1501 log(LOG_ERR, 1502 "%s: illegal response type %d on iq %p", 1503 device_get_nameunit(sc->dev), rsp_type, iq); 1504 break; 1505 } 1506 1507 d++; 1508 if (__predict_false(++iq->cidx == iq->sidx)) { 1509 iq->cidx = 0; 1510 iq->gen ^= F_RSPD_GEN; 1511 d = &iq->desc[0]; 1512 } 1513 if (__predict_false(++ndescs == limit)) { 1514 t4_write_reg(sc, sc->sge_gts_reg, 1515 V_CIDXINC(ndescs) | 1516 V_INGRESSQID(iq->cntxt_id) | 1517 V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX))); 1518 ndescs = 0; 1519 1520 if (budget) { 1521 return (EINPROGRESS); 1522 } 1523 } 1524 } 1525 1526 if (STAILQ_EMPTY(&iql)) 1527 break; 1528 1529 /* 1530 * Process the head only, and send it to the back of the list if 1531 * it's still not done. 1532 */ 1533 q = STAILQ_FIRST(&iql); 1534 STAILQ_REMOVE_HEAD(&iql, link); 1535 if (service_iq_fl(q, q->qsize / 8) == 0) 1536 (void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE); 1537 else 1538 STAILQ_INSERT_TAIL(&iql, q, link); 1539 } 1540 1541 t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) | 1542 V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params)); 1543 1544 return (0); 1545 } 1546 1547 #if defined(INET) || defined(INET6) 1548 static inline int 1549 sort_before_lro(struct lro_ctrl *lro) 1550 { 1551 1552 return (lro->lro_mbuf_max != 0); 1553 } 1554 #endif 1555 1556 static inline uint64_t 1557 t4_tstmp_to_ns(struct adapter *sc, uint64_t hw_tstmp) 1558 { 1559 struct clock_sync *cur, dcur; 1560 uint64_t hw_clocks; 1561 uint64_t hw_clk_div; 1562 sbintime_t sbt_cur_to_prev, sbt; 1563 seqc_t gen; 1564 1565 for (;;) { 1566 cur = &sc->cal_info[sc->cal_current]; 1567 gen = seqc_read(&cur->gen); 1568 if (gen == 0) 1569 return (0); 1570 dcur = *cur; 1571 if (seqc_consistent(&cur->gen, gen)) 1572 break; 1573 } 1574 1575 /* 1576 * Our goal here is to have a result that is: 1577 * 1578 * ( (cur_time - prev_time) ) 1579 * ((hw_tstmp - hw_prev) * ----------------------------- ) + prev_time 1580 * ( (hw_cur - hw_prev) ) 1581 * 1582 * With the constraints that we cannot use float and we 1583 * don't want to overflow the uint64_t numbers we are using. 1584 */ 1585 hw_clocks = hw_tstmp - dcur.hw_prev; 1586 sbt_cur_to_prev = (dcur.sbt_cur - dcur.sbt_prev); 1587 hw_clk_div = dcur.hw_cur - dcur.hw_prev; 1588 sbt = hw_clocks * sbt_cur_to_prev / hw_clk_div + dcur.sbt_prev; 1589 return (sbttons(sbt)); 1590 } 1591 1592 static inline void 1593 move_to_next_rxbuf(struct sge_fl *fl) 1594 { 1595 1596 fl->rx_offset = 0; 1597 if (__predict_false((++fl->cidx & 7) == 0)) { 1598 uint16_t cidx = fl->cidx >> 3; 1599 1600 if (__predict_false(cidx == fl->sidx)) 1601 fl->cidx = cidx = 0; 1602 fl->hw_cidx = cidx; 1603 } 1604 } 1605 1606 /* 1607 * Deals with interrupts on an iq+fl queue. 1608 */ 1609 static int 1610 service_iq_fl(struct sge_iq *iq, int budget) 1611 { 1612 struct sge_rxq *rxq = iq_to_rxq(iq); 1613 struct sge_fl *fl; 1614 struct adapter *sc = iq->adapter; 1615 struct iq_desc *d = &iq->desc[iq->cidx]; 1616 int ndescs, limit; 1617 int rsp_type, starved; 1618 uint32_t lq; 1619 uint16_t fl_hw_cidx; 1620 struct mbuf *m0; 1621 #if defined(INET) || defined(INET6) 1622 const struct timeval lro_timeout = {0, sc->lro_timeout}; 1623 struct lro_ctrl *lro = &rxq->lro; 1624 #endif 1625 1626 KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq)); 1627 MPASS(iq->flags & IQ_HAS_FL); 1628 1629 ndescs = 0; 1630 #if defined(INET) || defined(INET6) 1631 if (iq->flags & IQ_ADJ_CREDIT) { 1632 MPASS(sort_before_lro(lro)); 1633 iq->flags &= ~IQ_ADJ_CREDIT; 1634 if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) { 1635 tcp_lro_flush_all(lro); 1636 t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) | 1637 V_INGRESSQID((u32)iq->cntxt_id) | 1638 V_SEINTARM(iq->intr_params)); 1639 return (0); 1640 } 1641 ndescs = 1; 1642 } 1643 #else 1644 MPASS((iq->flags & IQ_ADJ_CREDIT) == 0); 1645 #endif 1646 1647 limit = budget ? budget : iq->qsize / 16; 1648 fl = &rxq->fl; 1649 fl_hw_cidx = fl->hw_cidx; /* stable snapshot */ 1650 while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) { 1651 1652 rmb(); 1653 1654 m0 = NULL; 1655 rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen); 1656 lq = be32toh(d->rsp.pldbuflen_qid); 1657 1658 switch (rsp_type) { 1659 case X_RSPD_TYPE_FLBUF: 1660 if (lq & F_RSPD_NEWBUF) { 1661 if (fl->rx_offset > 0) 1662 move_to_next_rxbuf(fl); 1663 lq = G_RSPD_LEN(lq); 1664 } 1665 if (IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 4) { 1666 FL_LOCK(fl); 1667 refill_fl(sc, fl, 64); 1668 FL_UNLOCK(fl); 1669 fl_hw_cidx = fl->hw_cidx; 1670 } 1671 1672 if (d->rss.opcode == CPL_RX_PKT) { 1673 if (__predict_true(eth_rx(sc, rxq, d, lq) == 0)) 1674 break; 1675 goto out; 1676 } 1677 m0 = get_fl_payload(sc, fl, lq); 1678 if (__predict_false(m0 == NULL)) 1679 goto out; 1680 1681 /* fall through */ 1682 1683 case X_RSPD_TYPE_CPL: 1684 KASSERT(d->rss.opcode < NUM_CPL_CMDS, 1685 ("%s: bad opcode %02x.", __func__, d->rss.opcode)); 1686 t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0); 1687 break; 1688 1689 case X_RSPD_TYPE_INTR: 1690 1691 /* 1692 * There are 1K interrupt-capable queues (qids 0 1693 * through 1023). A response type indicating a 1694 * forwarded interrupt with a qid >= 1K is an 1695 * iWARP async notification. That is the only 1696 * acceptable indirect interrupt on this queue. 1697 */ 1698 if (__predict_false(lq < 1024)) { 1699 panic("%s: indirect interrupt on iq_fl %p " 1700 "with qid %u", __func__, iq, lq); 1701 } 1702 1703 t4_an_handler(iq, &d->rsp); 1704 break; 1705 1706 default: 1707 KASSERT(0, ("%s: illegal response type %d on iq %p", 1708 __func__, rsp_type, iq)); 1709 log(LOG_ERR, "%s: illegal response type %d on iq %p", 1710 device_get_nameunit(sc->dev), rsp_type, iq); 1711 break; 1712 } 1713 1714 d++; 1715 if (__predict_false(++iq->cidx == iq->sidx)) { 1716 iq->cidx = 0; 1717 iq->gen ^= F_RSPD_GEN; 1718 d = &iq->desc[0]; 1719 } 1720 if (__predict_false(++ndescs == limit)) { 1721 t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) | 1722 V_INGRESSQID(iq->cntxt_id) | 1723 V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX))); 1724 1725 #if defined(INET) || defined(INET6) 1726 if (iq->flags & IQ_LRO_ENABLED && 1727 !sort_before_lro(lro) && 1728 sc->lro_timeout != 0) { 1729 tcp_lro_flush_inactive(lro, &lro_timeout); 1730 } 1731 #endif 1732 if (budget) 1733 return (EINPROGRESS); 1734 ndescs = 0; 1735 } 1736 } 1737 out: 1738 #if defined(INET) || defined(INET6) 1739 if (iq->flags & IQ_LRO_ENABLED) { 1740 if (ndescs > 0 && lro->lro_mbuf_count > 8) { 1741 MPASS(sort_before_lro(lro)); 1742 /* hold back one credit and don't flush LRO state */ 1743 iq->flags |= IQ_ADJ_CREDIT; 1744 ndescs--; 1745 } else { 1746 tcp_lro_flush_all(lro); 1747 } 1748 } 1749 #endif 1750 1751 t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) | 1752 V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params)); 1753 1754 FL_LOCK(fl); 1755 starved = refill_fl(sc, fl, 64); 1756 FL_UNLOCK(fl); 1757 if (__predict_false(starved != 0)) 1758 add_fl_to_sfl(sc, fl); 1759 1760 return (0); 1761 } 1762 1763 static inline struct cluster_metadata * 1764 cl_metadata(struct fl_sdesc *sd) 1765 { 1766 1767 return ((void *)(sd->cl + sd->moff)); 1768 } 1769 1770 static void 1771 rxb_free(struct mbuf *m) 1772 { 1773 struct cluster_metadata *clm = m->m_ext.ext_arg1; 1774 1775 uma_zfree(clm->zone, clm->cl); 1776 counter_u64_add(extfree_rels, 1); 1777 } 1778 1779 /* 1780 * The mbuf returned comes from zone_muf and carries the payload in one of these 1781 * ways 1782 * a) complete frame inside the mbuf 1783 * b) m_cljset (for clusters without metadata) 1784 * d) m_extaddref (cluster with metadata) 1785 */ 1786 static struct mbuf * 1787 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset, 1788 int remaining) 1789 { 1790 struct mbuf *m; 1791 struct fl_sdesc *sd = &fl->sdesc[fl->cidx]; 1792 struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx]; 1793 struct cluster_metadata *clm; 1794 int len, blen; 1795 caddr_t payload; 1796 1797 if (fl->flags & FL_BUF_PACKING) { 1798 u_int l, pad; 1799 1800 blen = rxb->size2 - fl->rx_offset; /* max possible in this buf */ 1801 len = min(remaining, blen); 1802 payload = sd->cl + fl->rx_offset; 1803 1804 l = fr_offset + len; 1805 pad = roundup2(l, fl->buf_boundary) - l; 1806 if (fl->rx_offset + len + pad < rxb->size2) 1807 blen = len + pad; 1808 MPASS(fl->rx_offset + blen <= rxb->size2); 1809 } else { 1810 MPASS(fl->rx_offset == 0); /* not packing */ 1811 blen = rxb->size1; 1812 len = min(remaining, blen); 1813 payload = sd->cl; 1814 } 1815 1816 if (fr_offset == 0) { 1817 m = m_gethdr(M_NOWAIT, MT_DATA); 1818 if (__predict_false(m == NULL)) 1819 return (NULL); 1820 m->m_pkthdr.len = remaining; 1821 } else { 1822 m = m_get(M_NOWAIT, MT_DATA); 1823 if (__predict_false(m == NULL)) 1824 return (NULL); 1825 } 1826 m->m_len = len; 1827 kmsan_mark(payload, len, KMSAN_STATE_INITED); 1828 1829 if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) { 1830 /* copy data to mbuf */ 1831 bcopy(payload, mtod(m, caddr_t), len); 1832 if (fl->flags & FL_BUF_PACKING) { 1833 fl->rx_offset += blen; 1834 MPASS(fl->rx_offset <= rxb->size2); 1835 if (fl->rx_offset < rxb->size2) 1836 return (m); /* without advancing the cidx */ 1837 } 1838 } else if (fl->flags & FL_BUF_PACKING) { 1839 clm = cl_metadata(sd); 1840 if (sd->nmbuf++ == 0) { 1841 clm->refcount = 1; 1842 clm->zone = rxb->zone; 1843 clm->cl = sd->cl; 1844 counter_u64_add(extfree_refs, 1); 1845 } 1846 m_extaddref(m, payload, blen, &clm->refcount, rxb_free, clm, 1847 NULL); 1848 1849 fl->rx_offset += blen; 1850 MPASS(fl->rx_offset <= rxb->size2); 1851 if (fl->rx_offset < rxb->size2) 1852 return (m); /* without advancing the cidx */ 1853 } else { 1854 m_cljset(m, sd->cl, rxb->type); 1855 sd->cl = NULL; /* consumed, not a recycle candidate */ 1856 } 1857 1858 move_to_next_rxbuf(fl); 1859 1860 return (m); 1861 } 1862 1863 static struct mbuf * 1864 get_fl_payload(struct adapter *sc, struct sge_fl *fl, const u_int plen) 1865 { 1866 struct mbuf *m0, *m, **pnext; 1867 u_int remaining; 1868 1869 if (__predict_false(fl->flags & FL_BUF_RESUME)) { 1870 M_ASSERTPKTHDR(fl->m0); 1871 MPASS(fl->m0->m_pkthdr.len == plen); 1872 MPASS(fl->remaining < plen); 1873 1874 m0 = fl->m0; 1875 pnext = fl->pnext; 1876 remaining = fl->remaining; 1877 fl->flags &= ~FL_BUF_RESUME; 1878 goto get_segment; 1879 } 1880 1881 /* 1882 * Payload starts at rx_offset in the current hw buffer. Its length is 1883 * 'len' and it may span multiple hw buffers. 1884 */ 1885 1886 m0 = get_scatter_segment(sc, fl, 0, plen); 1887 if (m0 == NULL) 1888 return (NULL); 1889 remaining = plen - m0->m_len; 1890 pnext = &m0->m_next; 1891 while (remaining > 0) { 1892 get_segment: 1893 MPASS(fl->rx_offset == 0); 1894 m = get_scatter_segment(sc, fl, plen - remaining, remaining); 1895 if (__predict_false(m == NULL)) { 1896 fl->m0 = m0; 1897 fl->pnext = pnext; 1898 fl->remaining = remaining; 1899 fl->flags |= FL_BUF_RESUME; 1900 return (NULL); 1901 } 1902 *pnext = m; 1903 pnext = &m->m_next; 1904 remaining -= m->m_len; 1905 } 1906 *pnext = NULL; 1907 1908 M_ASSERTPKTHDR(m0); 1909 return (m0); 1910 } 1911 1912 static int 1913 skip_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset, 1914 int remaining) 1915 { 1916 struct fl_sdesc *sd = &fl->sdesc[fl->cidx]; 1917 struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx]; 1918 int len, blen; 1919 1920 if (fl->flags & FL_BUF_PACKING) { 1921 u_int l, pad; 1922 1923 blen = rxb->size2 - fl->rx_offset; /* max possible in this buf */ 1924 len = min(remaining, blen); 1925 1926 l = fr_offset + len; 1927 pad = roundup2(l, fl->buf_boundary) - l; 1928 if (fl->rx_offset + len + pad < rxb->size2) 1929 blen = len + pad; 1930 fl->rx_offset += blen; 1931 MPASS(fl->rx_offset <= rxb->size2); 1932 if (fl->rx_offset < rxb->size2) 1933 return (len); /* without advancing the cidx */ 1934 } else { 1935 MPASS(fl->rx_offset == 0); /* not packing */ 1936 blen = rxb->size1; 1937 len = min(remaining, blen); 1938 } 1939 move_to_next_rxbuf(fl); 1940 return (len); 1941 } 1942 1943 static inline void 1944 skip_fl_payload(struct adapter *sc, struct sge_fl *fl, int plen) 1945 { 1946 int remaining, fr_offset, len; 1947 1948 fr_offset = 0; 1949 remaining = plen; 1950 while (remaining > 0) { 1951 len = skip_scatter_segment(sc, fl, fr_offset, remaining); 1952 fr_offset += len; 1953 remaining -= len; 1954 } 1955 } 1956 1957 static inline int 1958 get_segment_len(struct adapter *sc, struct sge_fl *fl, int plen) 1959 { 1960 int len; 1961 struct fl_sdesc *sd = &fl->sdesc[fl->cidx]; 1962 struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx]; 1963 1964 if (fl->flags & FL_BUF_PACKING) 1965 len = rxb->size2 - fl->rx_offset; 1966 else 1967 len = rxb->size1; 1968 1969 return (min(plen, len)); 1970 } 1971 1972 static void 1973 handle_cpl_rx_pkt(struct adapter *sc, struct sge_rxq *rxq, 1974 const struct cpl_rx_pkt *cpl, struct mbuf *m0) 1975 { 1976 if_t ifp = rxq->ifp; 1977 uint16_t err_vec, tnl_type, tnlhdr_len; 1978 static const int sw_csum_flags[2][2] = { 1979 { 1980 /* IP, inner IP */ 1981 CSUM_ENCAP_VXLAN | 1982 CSUM_L3_CALC | CSUM_L3_VALID | 1983 CSUM_L4_CALC | CSUM_L4_VALID | 1984 CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID | 1985 CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID, 1986 1987 /* IP, inner IP6 */ 1988 CSUM_ENCAP_VXLAN | 1989 CSUM_L3_CALC | CSUM_L3_VALID | 1990 CSUM_L4_CALC | CSUM_L4_VALID | 1991 CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID, 1992 }, 1993 { 1994 /* IP6, inner IP */ 1995 CSUM_ENCAP_VXLAN | 1996 CSUM_L4_CALC | CSUM_L4_VALID | 1997 CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID | 1998 CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID, 1999 2000 /* IP6, inner IP6 */ 2001 CSUM_ENCAP_VXLAN | 2002 CSUM_L4_CALC | CSUM_L4_VALID | 2003 CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID, 2004 }, 2005 }; 2006 2007 if (sc->params.tp.rx_pkt_encap) { 2008 const uint16_t ev = be16toh(cpl->err_vec); 2009 2010 err_vec = G_T6_COMPR_RXERR_VEC(ev); 2011 tnl_type = G_T6_RX_TNL_TYPE(ev); 2012 tnlhdr_len = G_T6_RX_TNLHDR_LEN(ev); 2013 } else { 2014 err_vec = be16toh(cpl->err_vec); 2015 tnl_type = 0; 2016 tnlhdr_len = 0; 2017 } 2018 if (cpl->csum_calc && err_vec == 0) { 2019 int ipv6 = !!(cpl->l2info & htobe32(F_RXF_IP6)); 2020 2021 /* checksum(s) calculated and found to be correct. */ 2022 2023 MPASS((cpl->l2info & htobe32(F_RXF_IP)) ^ 2024 (cpl->l2info & htobe32(F_RXF_IP6))); 2025 m0->m_pkthdr.csum_data = be16toh(cpl->csum); 2026 if (tnl_type == 0) { 2027 if (!ipv6 && if_getcapenable(ifp) & IFCAP_RXCSUM) { 2028 m0->m_pkthdr.csum_flags = CSUM_L3_CALC | 2029 CSUM_L3_VALID | CSUM_L4_CALC | 2030 CSUM_L4_VALID; 2031 } else if (ipv6 && if_getcapenable(ifp) & IFCAP_RXCSUM_IPV6) { 2032 m0->m_pkthdr.csum_flags = CSUM_L4_CALC | 2033 CSUM_L4_VALID; 2034 } 2035 rxq->rxcsum++; 2036 } else { 2037 MPASS(tnl_type == RX_PKT_TNL_TYPE_VXLAN); 2038 2039 M_HASHTYPE_SETINNER(m0); 2040 if (__predict_false(cpl->ip_frag)) { 2041 /* 2042 * csum_data is for the inner frame (which is an 2043 * IP fragment) and is not 0xffff. There is no 2044 * way to pass the inner csum_data to the stack. 2045 * We don't want the stack to use the inner 2046 * csum_data to validate the outer frame or it 2047 * will get rejected. So we fix csum_data here 2048 * and let sw do the checksum of inner IP 2049 * fragments. 2050 * 2051 * XXX: Need 32b for csum_data2 in an rx mbuf. 2052 * Maybe stuff it into rcv_tstmp? 2053 */ 2054 m0->m_pkthdr.csum_data = 0xffff; 2055 if (ipv6) { 2056 m0->m_pkthdr.csum_flags = CSUM_L4_CALC | 2057 CSUM_L4_VALID; 2058 } else { 2059 m0->m_pkthdr.csum_flags = CSUM_L3_CALC | 2060 CSUM_L3_VALID | CSUM_L4_CALC | 2061 CSUM_L4_VALID; 2062 } 2063 } else { 2064 int outer_ipv6; 2065 2066 MPASS(m0->m_pkthdr.csum_data == 0xffff); 2067 2068 outer_ipv6 = tnlhdr_len >= 2069 sizeof(struct ether_header) + 2070 sizeof(struct ip6_hdr); 2071 m0->m_pkthdr.csum_flags = 2072 sw_csum_flags[outer_ipv6][ipv6]; 2073 } 2074 rxq->vxlan_rxcsum++; 2075 } 2076 } 2077 2078 if (cpl->vlan_ex) { 2079 if (sc->flags & IS_VF && sc->vlan_id) { 2080 /* 2081 * HW is not setup correctly if extracted vlan_id does 2082 * not match the VF's setting. 2083 */ 2084 MPASS(be16toh(cpl->vlan) == sc->vlan_id); 2085 } else { 2086 m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan); 2087 m0->m_flags |= M_VLANTAG; 2088 rxq->vlan_extraction++; 2089 } 2090 } 2091 } 2092 2093 static int 2094 eth_rx(struct adapter *sc, struct sge_rxq *rxq, const struct iq_desc *d, 2095 u_int plen) 2096 { 2097 struct mbuf *m0; 2098 if_t ifp = rxq->ifp; 2099 struct sge_fl *fl = &rxq->fl; 2100 struct vi_info *vi = if_getsoftc(ifp); 2101 #if defined(INET) || defined(INET6) 2102 struct lro_ctrl *lro = &rxq->lro; 2103 #endif 2104 int rc; 2105 const uint8_t fl_pktshift = sc->params.sge.fl_pktshift; 2106 static const uint8_t sw_hashtype[4][2] = { 2107 {M_HASHTYPE_NONE, M_HASHTYPE_NONE}, 2108 {M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6}, 2109 {M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6}, 2110 {M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6}, 2111 }; 2112 2113 MPASS(plen > fl_pktshift); 2114 if (vi->pfil != NULL && PFIL_HOOKED_IN(vi->pfil) && 2115 __predict_true((fl->flags & FL_BUF_RESUME) == 0)) { 2116 struct fl_sdesc *sd = &fl->sdesc[fl->cidx]; 2117 caddr_t frame; 2118 const int slen = get_segment_len(sc, fl, plen) - fl_pktshift; 2119 2120 frame = sd->cl + fl->rx_offset + fl_pktshift; 2121 CURVNET_SET_QUIET(if_getvnet(ifp)); 2122 rc = pfil_mem_in(vi->pfil, frame, slen, ifp, &m0); 2123 CURVNET_RESTORE(); 2124 if (rc == PFIL_DROPPED || rc == PFIL_CONSUMED) { 2125 skip_fl_payload(sc, fl, plen); 2126 return (0); 2127 } 2128 if (rc == PFIL_REALLOCED) { 2129 skip_fl_payload(sc, fl, plen); 2130 goto have_mbuf; 2131 } 2132 } 2133 2134 m0 = get_fl_payload(sc, fl, plen); 2135 if (__predict_false(m0 == NULL)) 2136 return (ENOMEM); 2137 m0->m_pkthdr.len -= fl_pktshift; 2138 m0->m_len -= fl_pktshift; 2139 m0->m_data += fl_pktshift; 2140 2141 have_mbuf: 2142 m0->m_pkthdr.rcvif = ifp; 2143 M_HASHTYPE_SET(m0, sw_hashtype[d->rss.hash_type][d->rss.ipv6]); 2144 m0->m_pkthdr.flowid = be32toh(d->rss.hash_val); 2145 #ifdef NUMA 2146 m0->m_pkthdr.numa_domain = if_getnumadomain(ifp); 2147 #endif 2148 if (rxq->iq.flags & IQ_RX_TIMESTAMP) { 2149 /* 2150 * Fill up rcv_tstmp and set M_TSTMP if we get a a non-zero back 2151 * from t4_tstmp_to_ns(). The descriptor has a 60b timestamp. 2152 */ 2153 m0->m_pkthdr.rcv_tstmp = t4_tstmp_to_ns(sc, 2154 be64toh(d->rsp.u.last_flit) & 0x0fffffffffffffffULL); 2155 if (m0->m_pkthdr.rcv_tstmp != 0) 2156 m0->m_flags |= M_TSTMP; 2157 } 2158 2159 handle_cpl_rx_pkt(sc, rxq, (const void *)(&d->rss + 1), m0); 2160 2161 #if defined(INET) || defined(INET6) 2162 if (rxq->iq.flags & IQ_LRO_ENABLED && 2163 (m0->m_pkthdr.rsstype & M_HASHTYPE_INNER) == 0 && 2164 (M_HASHTYPE_GET(m0) == M_HASHTYPE_RSS_TCP_IPV4 || 2165 M_HASHTYPE_GET(m0) == M_HASHTYPE_RSS_TCP_IPV6)) { 2166 if (sort_before_lro(lro)) { 2167 tcp_lro_queue_mbuf(lro, m0); 2168 return (0); /* queued for sort, then LRO */ 2169 } 2170 if (tcp_lro_rx(lro, m0, 0) == 0) 2171 return (0); /* queued for LRO */ 2172 } 2173 #endif 2174 if_input(ifp, m0); 2175 2176 return (0); 2177 } 2178 2179 /* 2180 * Must drain the wrq or make sure that someone else will. 2181 */ 2182 static void 2183 wrq_tx_drain(void *arg, int n) 2184 { 2185 struct sge_wrq *wrq = arg; 2186 struct sge_eq *eq = &wrq->eq; 2187 2188 EQ_LOCK(eq); 2189 if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list)) 2190 drain_wrq_wr_list(wrq->adapter, wrq); 2191 EQ_UNLOCK(eq); 2192 } 2193 2194 static void 2195 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq) 2196 { 2197 struct sge_eq *eq = &wrq->eq; 2198 u_int available, dbdiff; /* # of hardware descriptors */ 2199 u_int n; 2200 struct wrqe *wr; 2201 struct fw_eth_tx_pkt_wr *dst; /* any fw WR struct will do */ 2202 2203 EQ_LOCK_ASSERT_OWNED(eq); 2204 MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs)); 2205 wr = STAILQ_FIRST(&wrq->wr_list); 2206 MPASS(wr != NULL); /* Must be called with something useful to do */ 2207 MPASS(eq->pidx == eq->dbidx); 2208 dbdiff = 0; 2209 2210 do { 2211 eq->cidx = read_hw_cidx(eq); 2212 if (eq->pidx == eq->cidx) 2213 available = eq->sidx - 1; 2214 else 2215 available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1; 2216 2217 MPASS(wr->wrq == wrq); 2218 n = howmany(wr->wr_len, EQ_ESIZE); 2219 if (available < n) 2220 break; 2221 2222 dst = (void *)&eq->desc[eq->pidx]; 2223 if (__predict_true(eq->sidx - eq->pidx > n)) { 2224 /* Won't wrap, won't end exactly at the status page. */ 2225 bcopy(&wr->wr[0], dst, wr->wr_len); 2226 eq->pidx += n; 2227 } else { 2228 int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE; 2229 2230 bcopy(&wr->wr[0], dst, first_portion); 2231 if (wr->wr_len > first_portion) { 2232 bcopy(&wr->wr[first_portion], &eq->desc[0], 2233 wr->wr_len - first_portion); 2234 } 2235 eq->pidx = n - (eq->sidx - eq->pidx); 2236 } 2237 wrq->tx_wrs_copied++; 2238 2239 if (available < eq->sidx / 4 && 2240 atomic_cmpset_int(&eq->equiq, 0, 1)) { 2241 /* 2242 * XXX: This is not 100% reliable with some 2243 * types of WRs. But this is a very unusual 2244 * situation for an ofld/ctrl queue anyway. 2245 */ 2246 dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ | 2247 F_FW_WR_EQUEQ); 2248 } 2249 2250 dbdiff += n; 2251 if (dbdiff >= 16) { 2252 ring_eq_db(sc, eq, dbdiff); 2253 dbdiff = 0; 2254 } 2255 2256 STAILQ_REMOVE_HEAD(&wrq->wr_list, link); 2257 free_wrqe(wr); 2258 MPASS(wrq->nwr_pending > 0); 2259 wrq->nwr_pending--; 2260 MPASS(wrq->ndesc_needed >= n); 2261 wrq->ndesc_needed -= n; 2262 } while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL); 2263 2264 if (dbdiff) 2265 ring_eq_db(sc, eq, dbdiff); 2266 } 2267 2268 /* 2269 * Doesn't fail. Holds on to work requests it can't send right away. 2270 */ 2271 void 2272 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr) 2273 { 2274 #ifdef INVARIANTS 2275 struct sge_eq *eq = &wrq->eq; 2276 #endif 2277 2278 EQ_LOCK_ASSERT_OWNED(eq); 2279 MPASS(wr != NULL); 2280 MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN); 2281 MPASS((wr->wr_len & 0x7) == 0); 2282 2283 STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link); 2284 wrq->nwr_pending++; 2285 wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE); 2286 2287 if (!TAILQ_EMPTY(&wrq->incomplete_wrs)) 2288 return; /* commit_wrq_wr will drain wr_list as well. */ 2289 2290 drain_wrq_wr_list(sc, wrq); 2291 2292 /* Doorbell must have caught up to the pidx. */ 2293 MPASS(eq->pidx == eq->dbidx); 2294 } 2295 2296 void 2297 t4_update_fl_bufsize(if_t ifp) 2298 { 2299 struct vi_info *vi = if_getsoftc(ifp); 2300 struct adapter *sc = vi->adapter; 2301 struct sge_rxq *rxq; 2302 #ifdef TCP_OFFLOAD 2303 struct sge_ofld_rxq *ofld_rxq; 2304 #endif 2305 struct sge_fl *fl; 2306 int i, maxp; 2307 2308 maxp = max_rx_payload(sc, ifp, false); 2309 for_each_rxq(vi, i, rxq) { 2310 fl = &rxq->fl; 2311 2312 FL_LOCK(fl); 2313 fl->zidx = find_refill_source(sc, maxp, 2314 fl->flags & FL_BUF_PACKING); 2315 FL_UNLOCK(fl); 2316 } 2317 #ifdef TCP_OFFLOAD 2318 maxp = max_rx_payload(sc, ifp, true); 2319 for_each_ofld_rxq(vi, i, ofld_rxq) { 2320 fl = &ofld_rxq->fl; 2321 2322 FL_LOCK(fl); 2323 fl->zidx = find_refill_source(sc, maxp, 2324 fl->flags & FL_BUF_PACKING); 2325 FL_UNLOCK(fl); 2326 } 2327 #endif 2328 } 2329 2330 #ifdef RATELIMIT 2331 static inline int 2332 mbuf_eo_nsegs(struct mbuf *m) 2333 { 2334 2335 M_ASSERTPKTHDR(m); 2336 return (m->m_pkthdr.PH_loc.eight[1]); 2337 } 2338 2339 #if defined(INET) || defined(INET6) 2340 static inline void 2341 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs) 2342 { 2343 2344 M_ASSERTPKTHDR(m); 2345 m->m_pkthdr.PH_loc.eight[1] = nsegs; 2346 } 2347 #endif 2348 2349 static inline int 2350 mbuf_eo_len16(struct mbuf *m) 2351 { 2352 int n; 2353 2354 M_ASSERTPKTHDR(m); 2355 n = m->m_pkthdr.PH_loc.eight[2]; 2356 MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16); 2357 2358 return (n); 2359 } 2360 2361 #if defined(INET) || defined(INET6) 2362 static inline void 2363 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16) 2364 { 2365 2366 M_ASSERTPKTHDR(m); 2367 m->m_pkthdr.PH_loc.eight[2] = len16; 2368 } 2369 #endif 2370 2371 static inline int 2372 mbuf_eo_tsclk_tsoff(struct mbuf *m) 2373 { 2374 2375 M_ASSERTPKTHDR(m); 2376 return (m->m_pkthdr.PH_loc.eight[3]); 2377 } 2378 2379 #if defined(INET) || defined(INET6) 2380 static inline void 2381 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff) 2382 { 2383 2384 M_ASSERTPKTHDR(m); 2385 m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff; 2386 } 2387 #endif 2388 2389 static inline int 2390 needs_eo(struct m_snd_tag *mst) 2391 { 2392 2393 return (mst != NULL && mst->sw->type == IF_SND_TAG_TYPE_RATE_LIMIT); 2394 } 2395 #endif 2396 2397 /* 2398 * Try to allocate an mbuf to contain a raw work request. To make it 2399 * easy to construct the work request, don't allocate a chain but a 2400 * single mbuf. 2401 */ 2402 struct mbuf * 2403 alloc_wr_mbuf(int len, int how) 2404 { 2405 struct mbuf *m; 2406 2407 if (len <= MHLEN) 2408 m = m_gethdr(how, MT_DATA); 2409 else if (len <= MCLBYTES) 2410 m = m_getcl(how, MT_DATA, M_PKTHDR); 2411 else 2412 m = NULL; 2413 if (m == NULL) 2414 return (NULL); 2415 m->m_pkthdr.len = len; 2416 m->m_len = len; 2417 set_mbuf_cflags(m, MC_RAW_WR); 2418 set_mbuf_len16(m, howmany(len, 16)); 2419 return (m); 2420 } 2421 2422 static inline bool 2423 needs_hwcsum(struct mbuf *m) 2424 { 2425 const uint32_t csum_flags = CSUM_IP | CSUM_IP_UDP | CSUM_IP_TCP | 2426 CSUM_IP_TSO | CSUM_INNER_IP | CSUM_INNER_IP_UDP | 2427 CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO | CSUM_IP6_UDP | 2428 CSUM_IP6_TCP | CSUM_IP6_TSO | CSUM_INNER_IP6_UDP | 2429 CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_TSO; 2430 2431 M_ASSERTPKTHDR(m); 2432 2433 return (m->m_pkthdr.csum_flags & csum_flags); 2434 } 2435 2436 static inline bool 2437 needs_tso(struct mbuf *m) 2438 { 2439 const uint32_t csum_flags = CSUM_IP_TSO | CSUM_IP6_TSO | 2440 CSUM_INNER_IP_TSO | CSUM_INNER_IP6_TSO; 2441 2442 M_ASSERTPKTHDR(m); 2443 2444 return (m->m_pkthdr.csum_flags & csum_flags); 2445 } 2446 2447 static inline bool 2448 needs_vxlan_csum(struct mbuf *m) 2449 { 2450 2451 M_ASSERTPKTHDR(m); 2452 2453 return (m->m_pkthdr.csum_flags & CSUM_ENCAP_VXLAN); 2454 } 2455 2456 static inline bool 2457 needs_vxlan_tso(struct mbuf *m) 2458 { 2459 const uint32_t csum_flags = CSUM_ENCAP_VXLAN | CSUM_INNER_IP_TSO | 2460 CSUM_INNER_IP6_TSO; 2461 2462 M_ASSERTPKTHDR(m); 2463 2464 return ((m->m_pkthdr.csum_flags & csum_flags) != 0 && 2465 (m->m_pkthdr.csum_flags & csum_flags) != CSUM_ENCAP_VXLAN); 2466 } 2467 2468 #if defined(INET) || defined(INET6) 2469 static inline bool 2470 needs_inner_tcp_csum(struct mbuf *m) 2471 { 2472 const uint32_t csum_flags = CSUM_INNER_IP_TSO | CSUM_INNER_IP6_TSO; 2473 2474 M_ASSERTPKTHDR(m); 2475 2476 return (m->m_pkthdr.csum_flags & csum_flags); 2477 } 2478 #endif 2479 2480 static inline bool 2481 needs_l3_csum(struct mbuf *m) 2482 { 2483 const uint32_t csum_flags = CSUM_IP | CSUM_IP_TSO | CSUM_INNER_IP | 2484 CSUM_INNER_IP_TSO; 2485 2486 M_ASSERTPKTHDR(m); 2487 2488 return (m->m_pkthdr.csum_flags & csum_flags); 2489 } 2490 2491 static inline bool 2492 needs_outer_tcp_csum(struct mbuf *m) 2493 { 2494 const uint32_t csum_flags = CSUM_IP_TCP | CSUM_IP_TSO | CSUM_IP6_TCP | 2495 CSUM_IP6_TSO; 2496 2497 M_ASSERTPKTHDR(m); 2498 2499 return (m->m_pkthdr.csum_flags & csum_flags); 2500 } 2501 2502 #ifdef RATELIMIT 2503 static inline bool 2504 needs_outer_l4_csum(struct mbuf *m) 2505 { 2506 const uint32_t csum_flags = CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_TSO | 2507 CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_TSO; 2508 2509 M_ASSERTPKTHDR(m); 2510 2511 return (m->m_pkthdr.csum_flags & csum_flags); 2512 } 2513 2514 static inline bool 2515 needs_outer_udp_csum(struct mbuf *m) 2516 { 2517 const uint32_t csum_flags = CSUM_IP_UDP | CSUM_IP6_UDP; 2518 2519 M_ASSERTPKTHDR(m); 2520 2521 return (m->m_pkthdr.csum_flags & csum_flags); 2522 } 2523 #endif 2524 2525 static inline bool 2526 needs_vlan_insertion(struct mbuf *m) 2527 { 2528 2529 M_ASSERTPKTHDR(m); 2530 2531 return (m->m_flags & M_VLANTAG); 2532 } 2533 2534 #if defined(INET) || defined(INET6) 2535 static void * 2536 m_advance(struct mbuf **pm, int *poffset, int len) 2537 { 2538 struct mbuf *m = *pm; 2539 int offset = *poffset; 2540 uintptr_t p = 0; 2541 2542 MPASS(len > 0); 2543 2544 for (;;) { 2545 if (offset + len < m->m_len) { 2546 offset += len; 2547 p = mtod(m, uintptr_t) + offset; 2548 break; 2549 } 2550 len -= m->m_len - offset; 2551 m = m->m_next; 2552 offset = 0; 2553 MPASS(m != NULL); 2554 } 2555 *poffset = offset; 2556 *pm = m; 2557 return ((void *)p); 2558 } 2559 #endif 2560 2561 static inline int 2562 count_mbuf_ext_pgs(struct mbuf *m, int skip, vm_paddr_t *nextaddr) 2563 { 2564 vm_paddr_t paddr; 2565 int i, len, off, pglen, pgoff, seglen, segoff; 2566 int nsegs = 0; 2567 2568 M_ASSERTEXTPG(m); 2569 off = mtod(m, vm_offset_t); 2570 len = m->m_len; 2571 off += skip; 2572 len -= skip; 2573 2574 if (m->m_epg_hdrlen != 0) { 2575 if (off >= m->m_epg_hdrlen) { 2576 off -= m->m_epg_hdrlen; 2577 } else { 2578 seglen = m->m_epg_hdrlen - off; 2579 segoff = off; 2580 seglen = min(seglen, len); 2581 off = 0; 2582 len -= seglen; 2583 paddr = pmap_kextract( 2584 (vm_offset_t)&m->m_epg_hdr[segoff]); 2585 if (*nextaddr != paddr) 2586 nsegs++; 2587 *nextaddr = paddr + seglen; 2588 } 2589 } 2590 pgoff = m->m_epg_1st_off; 2591 for (i = 0; i < m->m_epg_npgs && len > 0; i++) { 2592 pglen = m_epg_pagelen(m, i, pgoff); 2593 if (off >= pglen) { 2594 off -= pglen; 2595 pgoff = 0; 2596 continue; 2597 } 2598 seglen = pglen - off; 2599 segoff = pgoff + off; 2600 off = 0; 2601 seglen = min(seglen, len); 2602 len -= seglen; 2603 paddr = m->m_epg_pa[i] + segoff; 2604 if (*nextaddr != paddr) 2605 nsegs++; 2606 *nextaddr = paddr + seglen; 2607 pgoff = 0; 2608 }; 2609 if (len != 0) { 2610 seglen = min(len, m->m_epg_trllen - off); 2611 len -= seglen; 2612 paddr = pmap_kextract((vm_offset_t)&m->m_epg_trail[off]); 2613 if (*nextaddr != paddr) 2614 nsegs++; 2615 *nextaddr = paddr + seglen; 2616 } 2617 2618 return (nsegs); 2619 } 2620 2621 2622 /* 2623 * Can deal with empty mbufs in the chain that have m_len = 0, but the chain 2624 * must have at least one mbuf that's not empty. It is possible for this 2625 * routine to return 0 if skip accounts for all the contents of the mbuf chain. 2626 */ 2627 static inline int 2628 count_mbuf_nsegs(struct mbuf *m, int skip, uint8_t *cflags) 2629 { 2630 vm_paddr_t nextaddr, paddr; 2631 vm_offset_t va; 2632 int len, nsegs; 2633 2634 M_ASSERTPKTHDR(m); 2635 MPASS(m->m_pkthdr.len > 0); 2636 MPASS(m->m_pkthdr.len >= skip); 2637 2638 nsegs = 0; 2639 nextaddr = 0; 2640 for (; m; m = m->m_next) { 2641 len = m->m_len; 2642 if (__predict_false(len == 0)) 2643 continue; 2644 if (skip >= len) { 2645 skip -= len; 2646 continue; 2647 } 2648 if ((m->m_flags & M_EXTPG) != 0) { 2649 *cflags |= MC_NOMAP; 2650 nsegs += count_mbuf_ext_pgs(m, skip, &nextaddr); 2651 skip = 0; 2652 continue; 2653 } 2654 va = mtod(m, vm_offset_t) + skip; 2655 len -= skip; 2656 skip = 0; 2657 paddr = pmap_kextract(va); 2658 nsegs += sglist_count((void *)(uintptr_t)va, len); 2659 if (paddr == nextaddr) 2660 nsegs--; 2661 nextaddr = pmap_kextract(va + len - 1) + 1; 2662 } 2663 2664 return (nsegs); 2665 } 2666 2667 /* 2668 * The maximum number of segments that can fit in a WR. 2669 */ 2670 static int 2671 max_nsegs_allowed(struct mbuf *m, bool vm_wr) 2672 { 2673 2674 if (vm_wr) { 2675 if (needs_tso(m)) 2676 return (TX_SGL_SEGS_VM_TSO); 2677 return (TX_SGL_SEGS_VM); 2678 } 2679 2680 if (needs_tso(m)) { 2681 if (needs_vxlan_tso(m)) 2682 return (TX_SGL_SEGS_VXLAN_TSO); 2683 else 2684 return (TX_SGL_SEGS_TSO); 2685 } 2686 2687 return (TX_SGL_SEGS); 2688 } 2689 2690 static struct timeval txerr_ratecheck = {0}; 2691 static const struct timeval txerr_interval = {3, 0}; 2692 2693 /* 2694 * Analyze the mbuf to determine its tx needs. The mbuf passed in may change: 2695 * a) caller can assume it's been freed if this function returns with an error. 2696 * b) it may get defragged up if the gather list is too long for the hardware. 2697 */ 2698 int 2699 parse_pkt(struct mbuf **mp, bool vm_wr) 2700 { 2701 struct mbuf *m0 = *mp, *m; 2702 int rc, nsegs, defragged = 0; 2703 struct ether_header *eh; 2704 #ifdef INET 2705 void *l3hdr; 2706 #endif 2707 #if defined(INET) || defined(INET6) 2708 int offset; 2709 struct tcphdr *tcp; 2710 #endif 2711 #if defined(KERN_TLS) || defined(RATELIMIT) 2712 struct m_snd_tag *mst; 2713 #endif 2714 uint16_t eh_type; 2715 uint8_t cflags; 2716 2717 cflags = 0; 2718 M_ASSERTPKTHDR(m0); 2719 if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) { 2720 rc = EINVAL; 2721 fail: 2722 m_freem(m0); 2723 *mp = NULL; 2724 return (rc); 2725 } 2726 restart: 2727 /* 2728 * First count the number of gather list segments in the payload. 2729 * Defrag the mbuf if nsegs exceeds the hardware limit. 2730 */ 2731 M_ASSERTPKTHDR(m0); 2732 MPASS(m0->m_pkthdr.len > 0); 2733 nsegs = count_mbuf_nsegs(m0, 0, &cflags); 2734 #if defined(KERN_TLS) || defined(RATELIMIT) 2735 if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG) 2736 mst = m0->m_pkthdr.snd_tag; 2737 else 2738 mst = NULL; 2739 #endif 2740 #ifdef KERN_TLS 2741 if (mst != NULL && mst->sw->type == IF_SND_TAG_TYPE_TLS) { 2742 struct vi_info *vi = if_getsoftc(mst->ifp); 2743 2744 cflags |= MC_TLS; 2745 set_mbuf_cflags(m0, cflags); 2746 if (is_t6(vi->pi->adapter)) 2747 rc = t6_ktls_parse_pkt(m0); 2748 else 2749 rc = t7_ktls_parse_pkt(m0); 2750 if (rc != 0) 2751 goto fail; 2752 return (EINPROGRESS); 2753 } 2754 #endif 2755 if (nsegs > max_nsegs_allowed(m0, vm_wr)) { 2756 if (defragged++ > 0) { 2757 rc = EFBIG; 2758 goto fail; 2759 } 2760 counter_u64_add(defrags, 1); 2761 if ((m = m_defrag(m0, M_NOWAIT)) == NULL) { 2762 rc = ENOMEM; 2763 goto fail; 2764 } 2765 *mp = m0 = m; /* update caller's copy after defrag */ 2766 goto restart; 2767 } 2768 2769 if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN && 2770 !(cflags & MC_NOMAP))) { 2771 counter_u64_add(pullups, 1); 2772 m0 = m_pullup(m0, m0->m_pkthdr.len); 2773 if (m0 == NULL) { 2774 /* Should have left well enough alone. */ 2775 rc = EFBIG; 2776 goto fail; 2777 } 2778 *mp = m0; /* update caller's copy after pullup */ 2779 goto restart; 2780 } 2781 set_mbuf_nsegs(m0, nsegs); 2782 set_mbuf_cflags(m0, cflags); 2783 calculate_mbuf_len16(m0, vm_wr); 2784 2785 #ifdef RATELIMIT 2786 /* 2787 * Ethofld is limited to TCP and UDP for now, and only when L4 hw 2788 * checksumming is enabled. needs_outer_l4_csum happens to check for 2789 * all the right things. 2790 */ 2791 if (__predict_false(needs_eo(mst) && !needs_outer_l4_csum(m0))) { 2792 m_snd_tag_rele(m0->m_pkthdr.snd_tag); 2793 m0->m_pkthdr.snd_tag = NULL; 2794 m0->m_pkthdr.csum_flags &= ~CSUM_SND_TAG; 2795 mst = NULL; 2796 } 2797 #endif 2798 2799 if (!needs_hwcsum(m0) 2800 #ifdef RATELIMIT 2801 && !needs_eo(mst) 2802 #endif 2803 ) 2804 return (0); 2805 2806 m = m0; 2807 eh = mtod(m, struct ether_header *); 2808 eh_type = ntohs(eh->ether_type); 2809 if (eh_type == ETHERTYPE_VLAN) { 2810 struct ether_vlan_header *evh = (void *)eh; 2811 2812 eh_type = ntohs(evh->evl_proto); 2813 m0->m_pkthdr.l2hlen = sizeof(*evh); 2814 } else 2815 m0->m_pkthdr.l2hlen = sizeof(*eh); 2816 2817 #if defined(INET) || defined(INET6) 2818 offset = 0; 2819 #ifdef INET 2820 l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen); 2821 #else 2822 m_advance(&m, &offset, m0->m_pkthdr.l2hlen); 2823 #endif 2824 #endif 2825 2826 switch (eh_type) { 2827 #ifdef INET6 2828 case ETHERTYPE_IPV6: 2829 m0->m_pkthdr.l3hlen = sizeof(struct ip6_hdr); 2830 break; 2831 #endif 2832 #ifdef INET 2833 case ETHERTYPE_IP: 2834 { 2835 struct ip *ip = l3hdr; 2836 2837 if (needs_vxlan_csum(m0)) { 2838 /* Driver will do the outer IP hdr checksum. */ 2839 ip->ip_sum = 0; 2840 if (needs_vxlan_tso(m0)) { 2841 const uint16_t ipl = ip->ip_len; 2842 2843 ip->ip_len = 0; 2844 ip->ip_sum = ~in_cksum_hdr(ip); 2845 ip->ip_len = ipl; 2846 } else 2847 ip->ip_sum = in_cksum_hdr(ip); 2848 } 2849 m0->m_pkthdr.l3hlen = ip->ip_hl << 2; 2850 break; 2851 } 2852 #endif 2853 default: 2854 if (ratecheck(&txerr_ratecheck, &txerr_interval)) { 2855 log(LOG_ERR, "%s: ethertype 0x%04x unknown. " 2856 "if_cxgbe must be compiled with the same " 2857 "INET/INET6 options as the kernel.\n", __func__, 2858 eh_type); 2859 } 2860 rc = EINVAL; 2861 goto fail; 2862 } 2863 2864 #if defined(INET) || defined(INET6) 2865 if (needs_vxlan_csum(m0)) { 2866 m0->m_pkthdr.l4hlen = sizeof(struct udphdr); 2867 m0->m_pkthdr.l5hlen = sizeof(struct vxlan_header); 2868 2869 /* Inner headers. */ 2870 eh = m_advance(&m, &offset, m0->m_pkthdr.l3hlen + 2871 sizeof(struct udphdr) + sizeof(struct vxlan_header)); 2872 eh_type = ntohs(eh->ether_type); 2873 if (eh_type == ETHERTYPE_VLAN) { 2874 struct ether_vlan_header *evh = (void *)eh; 2875 2876 eh_type = ntohs(evh->evl_proto); 2877 m0->m_pkthdr.inner_l2hlen = sizeof(*evh); 2878 } else 2879 m0->m_pkthdr.inner_l2hlen = sizeof(*eh); 2880 #ifdef INET 2881 l3hdr = m_advance(&m, &offset, m0->m_pkthdr.inner_l2hlen); 2882 #else 2883 m_advance(&m, &offset, m0->m_pkthdr.inner_l2hlen); 2884 #endif 2885 2886 switch (eh_type) { 2887 #ifdef INET6 2888 case ETHERTYPE_IPV6: 2889 m0->m_pkthdr.inner_l3hlen = sizeof(struct ip6_hdr); 2890 break; 2891 #endif 2892 #ifdef INET 2893 case ETHERTYPE_IP: 2894 { 2895 struct ip *ip = l3hdr; 2896 2897 m0->m_pkthdr.inner_l3hlen = ip->ip_hl << 2; 2898 break; 2899 } 2900 #endif 2901 default: 2902 if (ratecheck(&txerr_ratecheck, &txerr_interval)) { 2903 log(LOG_ERR, "%s: VXLAN hw offload requested" 2904 "with unknown ethertype 0x%04x. if_cxgbe " 2905 "must be compiled with the same INET/INET6 " 2906 "options as the kernel.\n", __func__, 2907 eh_type); 2908 } 2909 rc = EINVAL; 2910 goto fail; 2911 } 2912 if (needs_inner_tcp_csum(m0)) { 2913 tcp = m_advance(&m, &offset, m0->m_pkthdr.inner_l3hlen); 2914 m0->m_pkthdr.inner_l4hlen = tcp->th_off * 4; 2915 } 2916 MPASS((m0->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0); 2917 m0->m_pkthdr.csum_flags &= CSUM_INNER_IP6_UDP | 2918 CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_TSO | CSUM_INNER_IP | 2919 CSUM_INNER_IP_UDP | CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO | 2920 CSUM_ENCAP_VXLAN; 2921 } 2922 2923 if (needs_outer_tcp_csum(m0)) { 2924 tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen); 2925 m0->m_pkthdr.l4hlen = tcp->th_off * 4; 2926 #ifdef RATELIMIT 2927 if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) { 2928 set_mbuf_eo_tsclk_tsoff(m0, 2929 V_FW_ETH_TX_EO_WR_TSCLK(tsclk) | 2930 V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1)); 2931 } else 2932 set_mbuf_eo_tsclk_tsoff(m0, 0); 2933 } else if (needs_outer_udp_csum(m0)) { 2934 m0->m_pkthdr.l4hlen = sizeof(struct udphdr); 2935 #endif 2936 } 2937 #ifdef RATELIMIT 2938 if (needs_eo(mst)) { 2939 u_int immhdrs; 2940 2941 /* EO WRs have the headers in the WR and not the GL. */ 2942 immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + 2943 m0->m_pkthdr.l4hlen; 2944 cflags = 0; 2945 nsegs = count_mbuf_nsegs(m0, immhdrs, &cflags); 2946 MPASS(cflags == mbuf_cflags(m0)); 2947 set_mbuf_eo_nsegs(m0, nsegs); 2948 set_mbuf_eo_len16(m0, 2949 txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0))); 2950 rc = ethofld_transmit(mst->ifp, m0); 2951 if (rc != 0) 2952 goto fail; 2953 return (EINPROGRESS); 2954 } 2955 #endif 2956 #endif 2957 MPASS(m0 == *mp); 2958 return (0); 2959 } 2960 2961 void * 2962 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie) 2963 { 2964 struct sge_eq *eq = &wrq->eq; 2965 struct adapter *sc = wrq->adapter; 2966 int ndesc, available; 2967 struct wrqe *wr; 2968 void *w; 2969 2970 MPASS(len16 > 0); 2971 ndesc = tx_len16_to_desc(len16); 2972 MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC); 2973 2974 EQ_LOCK(eq); 2975 if (__predict_false((eq->flags & EQ_HW_ALLOCATED) == 0)) { 2976 EQ_UNLOCK(eq); 2977 return (NULL); 2978 } 2979 2980 if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list)) 2981 drain_wrq_wr_list(sc, wrq); 2982 2983 if (!STAILQ_EMPTY(&wrq->wr_list)) { 2984 slowpath: 2985 EQ_UNLOCK(eq); 2986 wr = alloc_wrqe(len16 * 16, wrq); 2987 if (__predict_false(wr == NULL)) 2988 return (NULL); 2989 cookie->pidx = -1; 2990 cookie->ndesc = ndesc; 2991 return (&wr->wr); 2992 } 2993 2994 eq->cidx = read_hw_cidx(eq); 2995 if (eq->pidx == eq->cidx) 2996 available = eq->sidx - 1; 2997 else 2998 available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1; 2999 if (available < ndesc) 3000 goto slowpath; 3001 3002 cookie->pidx = eq->pidx; 3003 cookie->ndesc = ndesc; 3004 TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link); 3005 3006 w = &eq->desc[eq->pidx]; 3007 IDXINCR(eq->pidx, ndesc, eq->sidx); 3008 if (__predict_false(cookie->pidx + ndesc > eq->sidx)) { 3009 w = &wrq->ss[0]; 3010 wrq->ss_pidx = cookie->pidx; 3011 wrq->ss_len = len16 * 16; 3012 } 3013 3014 EQ_UNLOCK(eq); 3015 3016 return (w); 3017 } 3018 3019 void 3020 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie) 3021 { 3022 struct sge_eq *eq = &wrq->eq; 3023 struct adapter *sc = wrq->adapter; 3024 int ndesc, pidx; 3025 struct wrq_cookie *prev, *next; 3026 3027 if (cookie->pidx == -1) { 3028 struct wrqe *wr = __containerof(w, struct wrqe, wr); 3029 3030 t4_wrq_tx(sc, wr); 3031 return; 3032 } 3033 3034 if (__predict_false(w == &wrq->ss[0])) { 3035 int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE; 3036 3037 MPASS(wrq->ss_len > n); /* WR had better wrap around. */ 3038 bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n); 3039 bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n); 3040 wrq->tx_wrs_ss++; 3041 } else 3042 wrq->tx_wrs_direct++; 3043 3044 EQ_LOCK(eq); 3045 ndesc = cookie->ndesc; /* Can be more than SGE_MAX_WR_NDESC here. */ 3046 pidx = cookie->pidx; 3047 MPASS(pidx >= 0 && pidx < eq->sidx); 3048 prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link); 3049 next = TAILQ_NEXT(cookie, link); 3050 if (prev == NULL) { 3051 MPASS(pidx == eq->dbidx); 3052 if (next == NULL || ndesc >= 16) { 3053 int available; 3054 struct fw_eth_tx_pkt_wr *dst; /* any fw WR struct will do */ 3055 3056 /* 3057 * Note that the WR via which we'll request tx updates 3058 * is at pidx and not eq->pidx, which has moved on 3059 * already. 3060 */ 3061 dst = (void *)&eq->desc[pidx]; 3062 available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1; 3063 if (available < eq->sidx / 4 && 3064 atomic_cmpset_int(&eq->equiq, 0, 1)) { 3065 /* 3066 * XXX: This is not 100% reliable with some 3067 * types of WRs. But this is a very unusual 3068 * situation for an ofld/ctrl queue anyway. 3069 */ 3070 dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ | 3071 F_FW_WR_EQUEQ); 3072 } 3073 3074 if (__predict_true(eq->flags & EQ_HW_ALLOCATED)) 3075 ring_eq_db(wrq->adapter, eq, ndesc); 3076 else 3077 IDXINCR(eq->dbidx, ndesc, eq->sidx); 3078 } else { 3079 MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc); 3080 next->pidx = pidx; 3081 next->ndesc += ndesc; 3082 } 3083 } else { 3084 MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc); 3085 prev->ndesc += ndesc; 3086 } 3087 TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link); 3088 3089 if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list)) 3090 drain_wrq_wr_list(sc, wrq); 3091 3092 #ifdef INVARIANTS 3093 if (TAILQ_EMPTY(&wrq->incomplete_wrs)) { 3094 /* Doorbell must have caught up to the pidx. */ 3095 MPASS(wrq->eq.pidx == wrq->eq.dbidx); 3096 } 3097 #endif 3098 EQ_UNLOCK(eq); 3099 } 3100 3101 static u_int 3102 can_resume_eth_tx(struct mp_ring *r) 3103 { 3104 struct sge_eq *eq = r->cookie; 3105 3106 return (total_available_tx_desc(eq) > eq->sidx / 8); 3107 } 3108 3109 static inline bool 3110 cannot_use_txpkts(struct mbuf *m) 3111 { 3112 /* maybe put a GL limit too, to avoid silliness? */ 3113 3114 return (needs_tso(m) || (mbuf_cflags(m) & (MC_RAW_WR | MC_TLS)) != 0); 3115 } 3116 3117 static inline int 3118 discard_tx(struct sge_eq *eq) 3119 { 3120 3121 return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED); 3122 } 3123 3124 static inline int 3125 wr_can_update_eq(void *p) 3126 { 3127 struct fw_eth_tx_pkts_wr *wr = p; 3128 3129 switch (G_FW_WR_OP(be32toh(wr->op_pkd))) { 3130 case FW_ULPTX_WR: 3131 case FW_ETH_TX_PKT_WR: 3132 case FW_ETH_TX_PKTS_WR: 3133 case FW_ETH_TX_PKTS2_WR: 3134 case FW_ETH_TX_PKT_VM_WR: 3135 case FW_ETH_TX_PKTS_VM_WR: 3136 return (1); 3137 default: 3138 return (0); 3139 } 3140 } 3141 3142 static inline void 3143 set_txupdate_flags(struct sge_txq *txq, u_int avail, 3144 struct fw_eth_tx_pkt_wr *wr) 3145 { 3146 struct sge_eq *eq = &txq->eq; 3147 struct txpkts *txp = &txq->txp; 3148 3149 if ((txp->npkt > 0 || avail < eq->sidx / 2) && 3150 atomic_cmpset_int(&eq->equiq, 0, 1)) { 3151 wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ | F_FW_WR_EQUIQ); 3152 eq->equeqidx = eq->pidx; 3153 } else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) { 3154 wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ); 3155 eq->equeqidx = eq->pidx; 3156 } 3157 } 3158 3159 #if defined(__i386__) || defined(__amd64__) 3160 extern uint64_t tsc_freq; 3161 #endif 3162 3163 static inline bool 3164 record_eth_tx_time(struct sge_txq *txq) 3165 { 3166 const uint64_t cycles = get_cyclecount(); 3167 const uint64_t last_tx = txq->last_tx; 3168 #if defined(__i386__) || defined(__amd64__) 3169 const uint64_t itg = tsc_freq * t4_tx_coalesce_gap / 1000000; 3170 #else 3171 const uint64_t itg = 0; 3172 #endif 3173 3174 MPASS(cycles >= last_tx); 3175 txq->last_tx = cycles; 3176 return (cycles - last_tx < itg); 3177 } 3178 3179 /* 3180 * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to 3181 * be consumed. Return the actual number consumed. 0 indicates a stall. 3182 */ 3183 static u_int 3184 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx, bool *coalescing) 3185 { 3186 struct sge_txq *txq = r->cookie; 3187 if_t ifp = txq->ifp; 3188 struct sge_eq *eq = &txq->eq; 3189 struct txpkts *txp = &txq->txp; 3190 struct vi_info *vi = if_getsoftc(ifp); 3191 struct adapter *sc = vi->adapter; 3192 u_int total, remaining; /* # of packets */ 3193 u_int n, avail, dbdiff; /* # of hardware descriptors */ 3194 int i, rc; 3195 struct mbuf *m0; 3196 bool snd, recent_tx; 3197 void *wr; /* start of the last WR written to the ring */ 3198 3199 TXQ_LOCK_ASSERT_OWNED(txq); 3200 recent_tx = record_eth_tx_time(txq); 3201 3202 remaining = IDXDIFF(pidx, cidx, r->size); 3203 if (__predict_false(discard_tx(eq))) { 3204 for (i = 0; i < txp->npkt; i++) 3205 m_freem(txp->mb[i]); 3206 txp->npkt = 0; 3207 while (cidx != pidx) { 3208 m0 = r->items[cidx]; 3209 m_freem(m0); 3210 if (++cidx == r->size) 3211 cidx = 0; 3212 } 3213 reclaim_tx_descs(txq, eq->sidx); 3214 *coalescing = false; 3215 return (remaining); /* emptied */ 3216 } 3217 3218 /* How many hardware descriptors do we have readily available. */ 3219 if (eq->pidx == eq->cidx) 3220 avail = eq->sidx - 1; 3221 else 3222 avail = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1; 3223 3224 total = 0; 3225 if (remaining == 0) { 3226 txp->score = 0; 3227 txq->txpkts_flush++; 3228 goto send_txpkts; 3229 } 3230 3231 dbdiff = 0; 3232 MPASS(remaining > 0); 3233 while (remaining > 0) { 3234 m0 = r->items[cidx]; 3235 M_ASSERTPKTHDR(m0); 3236 MPASS(m0->m_nextpkt == NULL); 3237 3238 if (avail < 2 * SGE_MAX_WR_NDESC) 3239 avail += reclaim_tx_descs(txq, 64); 3240 3241 if (t4_tx_coalesce == 0 && txp->npkt == 0) 3242 goto skip_coalescing; 3243 if (cannot_use_txpkts(m0)) 3244 txp->score = 0; 3245 else if (recent_tx) { 3246 if (++txp->score == 0) 3247 txp->score = UINT8_MAX; 3248 } else 3249 txp->score = 1; 3250 if (txp->npkt > 0 || remaining > 1 || 3251 txp->score >= t4_tx_coalesce_pkts || 3252 atomic_load_int(&txq->eq.equiq) != 0) { 3253 if (vi->flags & TX_USES_VM_WR) 3254 rc = add_to_txpkts_vf(sc, txq, m0, avail, &snd); 3255 else 3256 rc = add_to_txpkts_pf(sc, txq, m0, avail, &snd); 3257 } else { 3258 snd = false; 3259 rc = EINVAL; 3260 } 3261 if (snd) { 3262 MPASS(txp->npkt > 0); 3263 for (i = 0; i < txp->npkt; i++) 3264 ETHER_BPF_MTAP(ifp, txp->mb[i]); 3265 if (txp->npkt > 1) { 3266 MPASS(avail >= tx_len16_to_desc(txp->len16)); 3267 if (vi->flags & TX_USES_VM_WR) 3268 n = write_txpkts_vm_wr(sc, txq); 3269 else 3270 n = write_txpkts_wr(sc, txq); 3271 } else { 3272 MPASS(avail >= 3273 tx_len16_to_desc(mbuf_len16(txp->mb[0]))); 3274 if (vi->flags & TX_USES_VM_WR) 3275 n = write_txpkt_vm_wr(sc, txq, 3276 txp->mb[0]); 3277 else 3278 n = write_txpkt_wr(sc, txq, txp->mb[0], 3279 avail); 3280 } 3281 MPASS(n <= SGE_MAX_WR_NDESC); 3282 avail -= n; 3283 dbdiff += n; 3284 wr = &eq->desc[eq->pidx]; 3285 IDXINCR(eq->pidx, n, eq->sidx); 3286 txp->npkt = 0; /* emptied */ 3287 } 3288 if (rc == 0) { 3289 /* m0 was coalesced into txq->txpkts. */ 3290 goto next_mbuf; 3291 } 3292 if (rc == EAGAIN) { 3293 /* 3294 * m0 is suitable for tx coalescing but could not be 3295 * combined with the existing txq->txpkts, which has now 3296 * been transmitted. Start a new txpkts with m0. 3297 */ 3298 MPASS(snd); 3299 MPASS(txp->npkt == 0); 3300 continue; 3301 } 3302 3303 MPASS(rc != 0 && rc != EAGAIN); 3304 MPASS(txp->npkt == 0); 3305 skip_coalescing: 3306 n = tx_len16_to_desc(mbuf_len16(m0)); 3307 if (__predict_false(avail < n)) { 3308 avail += reclaim_tx_descs(txq, min(n, 32)); 3309 if (avail < n) 3310 break; /* out of descriptors */ 3311 } 3312 3313 wr = &eq->desc[eq->pidx]; 3314 if (mbuf_cflags(m0) & MC_RAW_WR) { 3315 n = write_raw_wr(txq, wr, m0, avail); 3316 #ifdef KERN_TLS 3317 } else if (mbuf_cflags(m0) & MC_TLS) { 3318 ETHER_BPF_MTAP(ifp, m0); 3319 if (is_t6(sc)) 3320 n = t6_ktls_write_wr(txq, wr, m0, avail); 3321 else 3322 n = t7_ktls_write_wr(txq, wr, m0, avail); 3323 #endif 3324 } else { 3325 ETHER_BPF_MTAP(ifp, m0); 3326 if (vi->flags & TX_USES_VM_WR) 3327 n = write_txpkt_vm_wr(sc, txq, m0); 3328 else 3329 n = write_txpkt_wr(sc, txq, m0, avail); 3330 } 3331 MPASS(n >= 1 && n <= avail); 3332 if (!(mbuf_cflags(m0) & MC_TLS)) 3333 MPASS(n <= SGE_MAX_WR_NDESC); 3334 3335 avail -= n; 3336 dbdiff += n; 3337 IDXINCR(eq->pidx, n, eq->sidx); 3338 3339 if (dbdiff >= 512 / EQ_ESIZE) { /* X_FETCHBURSTMAX_512B */ 3340 if (wr_can_update_eq(wr)) 3341 set_txupdate_flags(txq, avail, wr); 3342 ring_eq_db(sc, eq, dbdiff); 3343 avail += reclaim_tx_descs(txq, 32); 3344 dbdiff = 0; 3345 } 3346 next_mbuf: 3347 total++; 3348 remaining--; 3349 if (__predict_false(++cidx == r->size)) 3350 cidx = 0; 3351 } 3352 if (dbdiff != 0) { 3353 if (wr_can_update_eq(wr)) 3354 set_txupdate_flags(txq, avail, wr); 3355 ring_eq_db(sc, eq, dbdiff); 3356 reclaim_tx_descs(txq, 32); 3357 } else if (eq->pidx == eq->cidx && txp->npkt > 0 && 3358 atomic_load_int(&txq->eq.equiq) == 0) { 3359 /* 3360 * If nothing was submitted to the chip for tx (it was coalesced 3361 * into txpkts instead) and there is no tx update outstanding 3362 * then we need to send txpkts now. 3363 */ 3364 send_txpkts: 3365 MPASS(txp->npkt > 0); 3366 for (i = 0; i < txp->npkt; i++) 3367 ETHER_BPF_MTAP(ifp, txp->mb[i]); 3368 if (txp->npkt > 1) { 3369 MPASS(avail >= tx_len16_to_desc(txp->len16)); 3370 if (vi->flags & TX_USES_VM_WR) 3371 n = write_txpkts_vm_wr(sc, txq); 3372 else 3373 n = write_txpkts_wr(sc, txq); 3374 } else { 3375 MPASS(avail >= 3376 tx_len16_to_desc(mbuf_len16(txp->mb[0]))); 3377 if (vi->flags & TX_USES_VM_WR) 3378 n = write_txpkt_vm_wr(sc, txq, txp->mb[0]); 3379 else 3380 n = write_txpkt_wr(sc, txq, txp->mb[0], avail); 3381 } 3382 MPASS(n <= SGE_MAX_WR_NDESC); 3383 wr = &eq->desc[eq->pidx]; 3384 IDXINCR(eq->pidx, n, eq->sidx); 3385 txp->npkt = 0; /* emptied */ 3386 3387 MPASS(wr_can_update_eq(wr)); 3388 set_txupdate_flags(txq, avail - n, wr); 3389 ring_eq_db(sc, eq, n); 3390 reclaim_tx_descs(txq, 32); 3391 } 3392 *coalescing = txp->npkt > 0; 3393 3394 return (total); 3395 } 3396 3397 static inline void 3398 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx, 3399 int qsize, int intr_idx, int cong, int qtype) 3400 { 3401 3402 KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS, 3403 ("%s: bad tmr_idx %d", __func__, tmr_idx)); 3404 KASSERT(pktc_idx < SGE_NCOUNTERS, /* -ve is ok, means don't use */ 3405 ("%s: bad pktc_idx %d", __func__, pktc_idx)); 3406 KASSERT(intr_idx >= -1 && intr_idx < sc->intr_count, 3407 ("%s: bad intr_idx %d", __func__, intr_idx)); 3408 KASSERT(qtype == FW_IQ_IQTYPE_OTHER || qtype == FW_IQ_IQTYPE_NIC || 3409 qtype == FW_IQ_IQTYPE_OFLD, ("%s: bad qtype %d", __func__, qtype)); 3410 3411 iq->flags = 0; 3412 iq->state = IQS_DISABLED; 3413 iq->adapter = sc; 3414 iq->qtype = qtype; 3415 iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx); 3416 iq->intr_pktc_idx = SGE_NCOUNTERS - 1; 3417 if (pktc_idx >= 0) { 3418 iq->intr_params |= F_QINTR_CNT_EN; 3419 iq->intr_pktc_idx = pktc_idx; 3420 } 3421 iq->qsize = roundup2(qsize, 16); /* See FW_IQ_CMD/iqsize */ 3422 iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE; 3423 iq->intr_idx = intr_idx; 3424 iq->cong_drop = cong; 3425 } 3426 3427 static inline void 3428 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name) 3429 { 3430 struct sge_params *sp = &sc->params.sge; 3431 3432 fl->qsize = qsize; 3433 fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE; 3434 strlcpy(fl->lockname, name, sizeof(fl->lockname)); 3435 mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF); 3436 if (sc->flags & BUF_PACKING_OK && 3437 ((!is_t4(sc) && buffer_packing) || /* T5+: enabled unless 0 */ 3438 (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */ 3439 fl->flags |= FL_BUF_PACKING; 3440 fl->zidx = find_refill_source(sc, maxp, fl->flags & FL_BUF_PACKING); 3441 fl->safe_zidx = sc->sge.safe_zidx; 3442 if (fl->flags & FL_BUF_PACKING) { 3443 fl->lowat = roundup2(sp->fl_starve_threshold2, 8); 3444 fl->buf_boundary = sp->pack_boundary; 3445 } else { 3446 fl->lowat = roundup2(sp->fl_starve_threshold, 8); 3447 fl->buf_boundary = 16; 3448 } 3449 if (fl_pad && fl->buf_boundary < sp->pad_boundary) 3450 fl->buf_boundary = sp->pad_boundary; 3451 } 3452 3453 static inline void 3454 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize, 3455 uint8_t port_id, struct sge_iq *iq, char *name) 3456 { 3457 KASSERT(eqtype >= EQ_CTRL && eqtype <= EQ_OFLD, 3458 ("%s: bad qtype %d", __func__, eqtype)); 3459 3460 eq->type = eqtype; 3461 eq->port_id = port_id; 3462 eq->tx_chan = sc->port[port_id]->tx_chan; 3463 eq->hw_port = sc->port[port_id]->hw_port; 3464 eq->iq = iq; 3465 eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE; 3466 strlcpy(eq->lockname, name, sizeof(eq->lockname)); 3467 mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF); 3468 } 3469 3470 int 3471 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag, 3472 bus_dmamap_t *map, bus_addr_t *pa, void **va) 3473 { 3474 int rc; 3475 3476 rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR, 3477 BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag); 3478 if (rc != 0) { 3479 CH_ERR(sc, "cannot allocate DMA tag: %d\n", rc); 3480 goto done; 3481 } 3482 3483 rc = bus_dmamem_alloc(*tag, va, 3484 BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map); 3485 if (rc != 0) { 3486 CH_ERR(sc, "cannot allocate DMA memory: %d\n", rc); 3487 goto done; 3488 } 3489 3490 rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0); 3491 if (rc != 0) { 3492 CH_ERR(sc, "cannot load DMA map: %d\n", rc); 3493 goto done; 3494 } 3495 done: 3496 if (rc) 3497 free_ring(sc, *tag, *map, *pa, *va); 3498 3499 return (rc); 3500 } 3501 3502 int 3503 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map, 3504 bus_addr_t pa, void *va) 3505 { 3506 if (pa) 3507 bus_dmamap_unload(tag, map); 3508 if (va) 3509 bus_dmamem_free(tag, va, map); 3510 if (tag) 3511 bus_dma_tag_destroy(tag); 3512 3513 return (0); 3514 } 3515 3516 /* 3517 * Allocates the software resources (mainly memory and sysctl nodes) for an 3518 * ingress queue and an optional freelist. 3519 * 3520 * Sets IQ_SW_ALLOCATED and returns 0 on success. 3521 */ 3522 static int 3523 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl, 3524 struct sysctl_ctx_list *ctx, struct sysctl_oid *oid) 3525 { 3526 int rc; 3527 size_t len; 3528 struct adapter *sc = vi->adapter; 3529 3530 MPASS(!(iq->flags & IQ_SW_ALLOCATED)); 3531 3532 len = iq->qsize * IQ_ESIZE; 3533 rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba, 3534 (void **)&iq->desc); 3535 if (rc != 0) 3536 return (rc); 3537 3538 if (fl) { 3539 len = fl->qsize * EQ_ESIZE; 3540 rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map, 3541 &fl->ba, (void **)&fl->desc); 3542 if (rc) { 3543 free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, 3544 iq->desc); 3545 return (rc); 3546 } 3547 3548 /* Allocate space for one software descriptor per buffer. */ 3549 fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), 3550 M_CXGBE, M_ZERO | M_WAITOK); 3551 3552 add_fl_sysctls(sc, ctx, oid, fl); 3553 iq->flags |= IQ_HAS_FL; 3554 } 3555 add_iq_sysctls(ctx, oid, iq); 3556 iq->flags |= IQ_SW_ALLOCATED; 3557 3558 return (0); 3559 } 3560 3561 /* 3562 * Frees all software resources (memory and locks) associated with an ingress 3563 * queue and an optional freelist. 3564 */ 3565 static void 3566 free_iq_fl(struct adapter *sc, struct sge_iq *iq, struct sge_fl *fl) 3567 { 3568 MPASS(iq->flags & IQ_SW_ALLOCATED); 3569 3570 if (fl) { 3571 MPASS(iq->flags & IQ_HAS_FL); 3572 free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba, fl->desc); 3573 free_fl_buffers(sc, fl); 3574 free(fl->sdesc, M_CXGBE); 3575 mtx_destroy(&fl->fl_lock); 3576 bzero(fl, sizeof(*fl)); 3577 } 3578 free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc); 3579 bzero(iq, sizeof(*iq)); 3580 } 3581 3582 /* 3583 * Allocates a hardware ingress queue and an optional freelist that will be 3584 * associated with it. 3585 * 3586 * Returns errno on failure. Resources allocated up to that point may still be 3587 * allocated. Caller is responsible for cleanup in case this function fails. 3588 */ 3589 static int 3590 alloc_iq_fl_hwq(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl) 3591 { 3592 int rc, cntxt_id, cong_map; 3593 struct fw_iq_cmd c; 3594 struct adapter *sc = vi->adapter; 3595 struct port_info *pi = vi->pi; 3596 __be32 v = 0; 3597 3598 MPASS (!(iq->flags & IQ_HW_ALLOCATED)); 3599 3600 bzero(&c, sizeof(c)); 3601 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST | 3602 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) | 3603 V_FW_IQ_CMD_VFN(0)); 3604 3605 c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART | 3606 FW_LEN16(c)); 3607 3608 /* Special handling for firmware event queue */ 3609 if (iq == &sc->sge.fwq) 3610 v |= F_FW_IQ_CMD_IQASYNCH; 3611 3612 if (iq->intr_idx < 0) { 3613 /* Forwarded interrupts, all headed to fwq */ 3614 v |= F_FW_IQ_CMD_IQANDST; 3615 v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id); 3616 } else { 3617 KASSERT(iq->intr_idx < sc->intr_count, 3618 ("%s: invalid direct intr_idx %d", __func__, iq->intr_idx)); 3619 v |= V_FW_IQ_CMD_IQANDSTINDEX(iq->intr_idx); 3620 } 3621 3622 bzero(iq->desc, iq->qsize * IQ_ESIZE); 3623 c.type_to_iqandstindex = htobe32(v | 3624 V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) | 3625 V_FW_IQ_CMD_VIID(vi->viid) | 3626 V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT)); 3627 c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->hw_port) | 3628 F_FW_IQ_CMD_IQGTSMODE | 3629 V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) | 3630 V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4)); 3631 c.iqsize = htobe16(iq->qsize); 3632 c.iqaddr = htobe64(iq->ba); 3633 c.iqns_to_fl0congen = htobe32(V_FW_IQ_CMD_IQTYPE(iq->qtype)); 3634 if (iq->cong_drop != -1) { 3635 if (iq->qtype == IQ_ETH) { 3636 if (chip_id(sc) >= CHELSIO_T7) 3637 cong_map = 1 << pi->hw_port; 3638 else 3639 cong_map = pi->rx_e_chan_map; 3640 } else 3641 cong_map = 0; 3642 c.iqns_to_fl0congen |= htobe32(F_FW_IQ_CMD_IQFLINTCONGEN); 3643 } 3644 3645 if (fl) { 3646 bzero(fl->desc, fl->sidx * EQ_ESIZE + sc->params.sge.spg_len); 3647 c.iqns_to_fl0congen |= 3648 htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) | 3649 F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO | 3650 (fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) | 3651 (fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN : 3652 0)); 3653 if (iq->cong_drop != -1) { 3654 c.iqns_to_fl0congen |= 3655 htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong_map) | 3656 F_FW_IQ_CMD_FL0CONGCIF | 3657 F_FW_IQ_CMD_FL0CONGEN); 3658 } 3659 c.fl0dcaen_to_fl0cidxfthresh = 3660 htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ? 3661 X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) | 3662 V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ? 3663 X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B)); 3664 c.fl0size = htobe16(fl->qsize); 3665 c.fl0addr = htobe64(fl->ba); 3666 } 3667 3668 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); 3669 if (rc != 0) { 3670 CH_ERR(sc, "failed to create hw ingress queue: %d\n", rc); 3671 return (rc); 3672 } 3673 3674 iq->cidx = 0; 3675 iq->gen = F_RSPD_GEN; 3676 iq->cntxt_id = be16toh(c.iqid); 3677 iq->abs_id = be16toh(c.physiqid); 3678 3679 cntxt_id = iq->cntxt_id - sc->sge.iq_start; 3680 if (cntxt_id >= sc->sge.iqmap_sz) { 3681 panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__, 3682 cntxt_id, sc->sge.iqmap_sz - 1); 3683 } 3684 sc->sge.iqmap[cntxt_id] = iq; 3685 3686 if (fl) { 3687 u_int qid; 3688 #ifdef INVARIANTS 3689 int i; 3690 3691 MPASS(!(fl->flags & FL_BUF_RESUME)); 3692 for (i = 0; i < fl->sidx * 8; i++) 3693 MPASS(fl->sdesc[i].cl == NULL); 3694 #endif 3695 fl->cntxt_id = be16toh(c.fl0id); 3696 fl->pidx = fl->cidx = fl->hw_cidx = fl->dbidx = 0; 3697 fl->rx_offset = 0; 3698 fl->flags &= ~(FL_STARVING | FL_DOOMED); 3699 3700 cntxt_id = fl->cntxt_id - sc->sge.eq_start; 3701 if (cntxt_id >= sc->sge.eqmap_sz) { 3702 panic("%s: fl->cntxt_id (%d) more than the max (%d)", 3703 __func__, cntxt_id, sc->sge.eqmap_sz - 1); 3704 } 3705 sc->sge.eqmap[cntxt_id] = (void *)fl; 3706 3707 qid = fl->cntxt_id; 3708 if (isset(&sc->doorbells, DOORBELL_UDB)) { 3709 uint32_t s_qpp = sc->params.sge.eq_s_qpp; 3710 uint32_t mask = (1 << s_qpp) - 1; 3711 volatile uint8_t *udb; 3712 3713 udb = sc->udbs_base + UDBS_DB_OFFSET; 3714 udb += (qid >> s_qpp) << PAGE_SHIFT; 3715 qid &= mask; 3716 if (qid < PAGE_SIZE / UDBS_SEG_SIZE) { 3717 udb += qid << UDBS_SEG_SHIFT; 3718 qid = 0; 3719 } 3720 fl->udb = (volatile void *)udb; 3721 } 3722 fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db; 3723 3724 FL_LOCK(fl); 3725 /* Enough to make sure the SGE doesn't think it's starved */ 3726 refill_fl(sc, fl, fl->lowat); 3727 FL_UNLOCK(fl); 3728 } 3729 3730 if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && 3731 iq->cong_drop != -1) { 3732 t4_sge_set_conm_context(sc, iq->cntxt_id, iq->cong_drop, 3733 cong_map); 3734 } 3735 3736 /* Enable IQ interrupts */ 3737 atomic_store_rel_int(&iq->state, IQS_IDLE); 3738 t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) | 3739 V_INGRESSQID(iq->cntxt_id)); 3740 3741 iq->flags |= IQ_HW_ALLOCATED; 3742 3743 return (0); 3744 } 3745 3746 static int 3747 free_iq_fl_hwq(struct adapter *sc, struct sge_iq *iq, struct sge_fl *fl) 3748 { 3749 int rc; 3750 3751 MPASS(iq->flags & IQ_HW_ALLOCATED); 3752 rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP, 3753 iq->cntxt_id, fl ? fl->cntxt_id : 0xffff, 0xffff); 3754 if (rc != 0) { 3755 CH_ERR(sc, "failed to free iq %p: %d\n", iq, rc); 3756 return (rc); 3757 } 3758 iq->flags &= ~IQ_HW_ALLOCATED; 3759 3760 return (0); 3761 } 3762 3763 static void 3764 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid, 3765 struct sge_iq *iq) 3766 { 3767 struct sysctl_oid_list *children; 3768 3769 if (ctx == NULL || oid == NULL) 3770 return; 3771 3772 children = SYSCTL_CHILDREN(oid); 3773 SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba, 3774 "bus address of descriptor ring"); 3775 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL, 3776 iq->qsize * IQ_ESIZE, "descriptor ring size in bytes"); 3777 SYSCTL_ADD_U16(ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD, 3778 &iq->abs_id, 0, "absolute id of the queue"); 3779 SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD, 3780 &iq->cntxt_id, 0, "SGE context id of the queue"); 3781 SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &iq->cidx, 3782 0, "consumer index"); 3783 } 3784 3785 static void 3786 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx, 3787 struct sysctl_oid *oid, struct sge_fl *fl) 3788 { 3789 struct sysctl_oid_list *children; 3790 3791 if (ctx == NULL || oid == NULL) 3792 return; 3793 3794 children = SYSCTL_CHILDREN(oid); 3795 oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", 3796 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist"); 3797 children = SYSCTL_CHILDREN(oid); 3798 3799 SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, 3800 &fl->ba, "bus address of descriptor ring"); 3801 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL, 3802 fl->sidx * EQ_ESIZE + sc->params.sge.spg_len, 3803 "desc ring size in bytes"); 3804 SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD, 3805 &fl->cntxt_id, 0, "SGE context id of the freelist"); 3806 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL, 3807 fl_pad ? 1 : 0, "padding enabled"); 3808 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL, 3809 fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled"); 3810 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx, 3811 0, "consumer index"); 3812 if (fl->flags & FL_BUF_PACKING) { 3813 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset", 3814 CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset"); 3815 } 3816 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx, 3817 0, "producer index"); 3818 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated", 3819 CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated"); 3820 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled", 3821 CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled"); 3822 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled", 3823 CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)"); 3824 } 3825 3826 /* 3827 * Idempotent. 3828 */ 3829 static int 3830 alloc_fwq(struct adapter *sc) 3831 { 3832 int rc, intr_idx; 3833 struct sge_iq *fwq = &sc->sge.fwq; 3834 struct vi_info *vi = &sc->port[0]->vi[0]; 3835 3836 if (!(fwq->flags & IQ_SW_ALLOCATED)) { 3837 MPASS(!(fwq->flags & IQ_HW_ALLOCATED)); 3838 3839 if (sc->flags & IS_VF) 3840 intr_idx = 0; 3841 else 3842 intr_idx = sc->intr_count > 1 ? 1 : 0; 3843 init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, intr_idx, -1, IQ_OTHER); 3844 rc = alloc_iq_fl(vi, fwq, NULL, &sc->ctx, sc->fwq_oid); 3845 if (rc != 0) { 3846 CH_ERR(sc, "failed to allocate fwq: %d\n", rc); 3847 return (rc); 3848 } 3849 MPASS(fwq->flags & IQ_SW_ALLOCATED); 3850 } 3851 3852 if (!(fwq->flags & IQ_HW_ALLOCATED)) { 3853 MPASS(fwq->flags & IQ_SW_ALLOCATED); 3854 3855 rc = alloc_iq_fl_hwq(vi, fwq, NULL); 3856 if (rc != 0) { 3857 CH_ERR(sc, "failed to create hw fwq: %d\n", rc); 3858 return (rc); 3859 } 3860 MPASS(fwq->flags & IQ_HW_ALLOCATED); 3861 } 3862 3863 return (0); 3864 } 3865 3866 /* 3867 * Idempotent. 3868 */ 3869 static void 3870 free_fwq(struct adapter *sc) 3871 { 3872 struct sge_iq *fwq = &sc->sge.fwq; 3873 3874 if (fwq->flags & IQ_HW_ALLOCATED) { 3875 MPASS(fwq->flags & IQ_SW_ALLOCATED); 3876 free_iq_fl_hwq(sc, fwq, NULL); 3877 MPASS(!(fwq->flags & IQ_HW_ALLOCATED)); 3878 } 3879 3880 if (fwq->flags & IQ_SW_ALLOCATED) { 3881 MPASS(!(fwq->flags & IQ_HW_ALLOCATED)); 3882 free_iq_fl(sc, fwq, NULL); 3883 MPASS(!(fwq->flags & IQ_SW_ALLOCATED)); 3884 } 3885 } 3886 3887 /* 3888 * Idempotent. 3889 */ 3890 static int 3891 alloc_ctrlq(struct adapter *sc, int idx) 3892 { 3893 int rc; 3894 char name[16]; 3895 struct sysctl_oid *oid; 3896 struct sge_wrq *ctrlq = &sc->sge.ctrlq[idx]; 3897 3898 MPASS(idx < sc->sge.nctrlq); 3899 3900 if (!(ctrlq->eq.flags & EQ_SW_ALLOCATED)) { 3901 MPASS(!(ctrlq->eq.flags & EQ_HW_ALLOCATED)); 3902 3903 snprintf(name, sizeof(name), "%d", idx); 3904 oid = SYSCTL_ADD_NODE(&sc->ctx, SYSCTL_CHILDREN(sc->ctrlq_oid), 3905 OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 3906 "ctrl queue"); 3907 3908 snprintf(name, sizeof(name), "%s ctrlq%d", 3909 device_get_nameunit(sc->dev), idx); 3910 init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, 3911 idx % sc->params.nports, &sc->sge.fwq, name); 3912 rc = alloc_wrq(sc, NULL, ctrlq, &sc->ctx, oid); 3913 if (rc != 0) { 3914 CH_ERR(sc, "failed to allocate ctrlq%d: %d\n", idx, rc); 3915 sysctl_remove_oid(oid, 1, 1); 3916 return (rc); 3917 } 3918 MPASS(ctrlq->eq.flags & EQ_SW_ALLOCATED); 3919 } 3920 3921 if (!(ctrlq->eq.flags & EQ_HW_ALLOCATED)) { 3922 MPASS(ctrlq->eq.flags & EQ_SW_ALLOCATED); 3923 MPASS(ctrlq->nwr_pending == 0); 3924 MPASS(ctrlq->ndesc_needed == 0); 3925 3926 rc = alloc_eq_hwq(sc, NULL, &ctrlq->eq, idx); 3927 if (rc != 0) { 3928 CH_ERR(sc, "failed to create hw ctrlq%d: %d\n", idx, rc); 3929 return (rc); 3930 } 3931 MPASS(ctrlq->eq.flags & EQ_HW_ALLOCATED); 3932 } 3933 3934 return (0); 3935 } 3936 3937 /* 3938 * Idempotent. 3939 */ 3940 static void 3941 free_ctrlq(struct adapter *sc, int idx) 3942 { 3943 struct sge_wrq *ctrlq = &sc->sge.ctrlq[idx]; 3944 3945 if (ctrlq->eq.flags & EQ_HW_ALLOCATED) { 3946 MPASS(ctrlq->eq.flags & EQ_SW_ALLOCATED); 3947 free_eq_hwq(sc, NULL, &ctrlq->eq); 3948 MPASS(!(ctrlq->eq.flags & EQ_HW_ALLOCATED)); 3949 } 3950 3951 if (ctrlq->eq.flags & EQ_SW_ALLOCATED) { 3952 MPASS(!(ctrlq->eq.flags & EQ_HW_ALLOCATED)); 3953 free_wrq(sc, ctrlq); 3954 MPASS(!(ctrlq->eq.flags & EQ_SW_ALLOCATED)); 3955 } 3956 } 3957 3958 int 3959 t4_sge_set_conm_context(struct adapter *sc, int cntxt_id, int cong_drop, 3960 int cong_map) 3961 { 3962 const int cng_ch_bits_log = sc->chip_params->cng_ch_bits_log; 3963 uint32_t param, val; 3964 uint16_t ch_map; 3965 int cong_mode, rc, i; 3966 3967 if (chip_id(sc) < CHELSIO_T5) 3968 return (ENOTSUP); 3969 3970 /* Convert the driver knob to the mode understood by the firmware. */ 3971 switch (cong_drop) { 3972 case -1: 3973 cong_mode = X_CONMCTXT_CNGTPMODE_DISABLE; 3974 break; 3975 case 0: 3976 cong_mode = X_CONMCTXT_CNGTPMODE_CHANNEL; 3977 break; 3978 case 1: 3979 cong_mode = X_CONMCTXT_CNGTPMODE_QUEUE; 3980 break; 3981 case 2: 3982 cong_mode = X_CONMCTXT_CNGTPMODE_BOTH; 3983 break; 3984 default: 3985 MPASS(0); 3986 CH_ERR(sc, "cong_drop = %d is invalid (ingress queue %d).\n", 3987 cong_drop, cntxt_id); 3988 return (EINVAL); 3989 } 3990 3991 param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) | 3992 V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) | 3993 V_FW_PARAMS_PARAM_YZ(cntxt_id); 3994 if (chip_id(sc) >= CHELSIO_T7) { 3995 val = V_T7_DMAQ_CONM_CTXT_CNGTPMODE(cong_mode) | 3996 V_T7_DMAQ_CONM_CTXT_CH_VEC(cong_map); 3997 } else { 3998 val = V_CONMCTXT_CNGTPMODE(cong_mode); 3999 if (cong_mode == X_CONMCTXT_CNGTPMODE_CHANNEL || 4000 cong_mode == X_CONMCTXT_CNGTPMODE_BOTH) { 4001 for (i = 0, ch_map = 0; i < 4; i++) { 4002 if (cong_map & (1 << i)) 4003 ch_map |= 1 << (i << cng_ch_bits_log); 4004 } 4005 val |= V_CONMCTXT_CNGCHMAP(ch_map); 4006 } 4007 } 4008 rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, ¶m, &val); 4009 if (rc != 0) { 4010 CH_ERR(sc, "failed to set congestion manager context " 4011 "for ingress queue %d: %d\n", cntxt_id, rc); 4012 } 4013 4014 return (rc); 4015 } 4016 4017 /* 4018 * Idempotent. 4019 */ 4020 static int 4021 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int idx, int intr_idx, 4022 int maxp) 4023 { 4024 int rc; 4025 struct adapter *sc = vi->adapter; 4026 if_t ifp = vi->ifp; 4027 struct sysctl_oid *oid; 4028 char name[16]; 4029 4030 if (!(rxq->iq.flags & IQ_SW_ALLOCATED)) { 4031 MPASS(!(rxq->iq.flags & IQ_HW_ALLOCATED)); 4032 #if defined(INET) || defined(INET6) 4033 rc = tcp_lro_init_args(&rxq->lro, ifp, lro_entries, lro_mbufs); 4034 if (rc != 0) 4035 return (rc); 4036 MPASS(rxq->lro.ifp == ifp); /* also indicates LRO init'ed */ 4037 #endif 4038 rxq->ifp = ifp; 4039 4040 snprintf(name, sizeof(name), "%d", idx); 4041 oid = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(vi->rxq_oid), 4042 OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 4043 "rx queue"); 4044 4045 init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq, 4046 intr_idx, cong_drop, IQ_ETH); 4047 #if defined(INET) || defined(INET6) 4048 if (if_getcapenable(ifp) & IFCAP_LRO) 4049 rxq->iq.flags |= IQ_LRO_ENABLED; 4050 #endif 4051 if (if_getcapenable(ifp) & IFCAP_HWRXTSTMP) 4052 rxq->iq.flags |= IQ_RX_TIMESTAMP; 4053 snprintf(name, sizeof(name), "%s rxq%d-fl", 4054 device_get_nameunit(vi->dev), idx); 4055 init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name); 4056 rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, &vi->ctx, oid); 4057 if (rc != 0) { 4058 CH_ERR(vi, "failed to allocate rxq%d: %d\n", idx, rc); 4059 sysctl_remove_oid(oid, 1, 1); 4060 #if defined(INET) || defined(INET6) 4061 tcp_lro_free(&rxq->lro); 4062 rxq->lro.ifp = NULL; 4063 #endif 4064 return (rc); 4065 } 4066 MPASS(rxq->iq.flags & IQ_SW_ALLOCATED); 4067 add_rxq_sysctls(&vi->ctx, oid, rxq); 4068 } 4069 4070 if (!(rxq->iq.flags & IQ_HW_ALLOCATED)) { 4071 MPASS(rxq->iq.flags & IQ_SW_ALLOCATED); 4072 rc = alloc_iq_fl_hwq(vi, &rxq->iq, &rxq->fl); 4073 if (rc != 0) { 4074 CH_ERR(vi, "failed to create hw rxq%d: %d\n", idx, rc); 4075 return (rc); 4076 } 4077 MPASS(rxq->iq.flags & IQ_HW_ALLOCATED); 4078 4079 if (idx == 0) 4080 sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id; 4081 else 4082 KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id, 4083 ("iq_base mismatch")); 4084 KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF, 4085 ("PF with non-zero iq_base")); 4086 4087 /* 4088 * The freelist is just barely above the starvation threshold 4089 * right now, fill it up a bit more. 4090 */ 4091 FL_LOCK(&rxq->fl); 4092 refill_fl(sc, &rxq->fl, 128); 4093 FL_UNLOCK(&rxq->fl); 4094 } 4095 4096 return (0); 4097 } 4098 4099 /* 4100 * Idempotent. 4101 */ 4102 static void 4103 free_rxq(struct vi_info *vi, struct sge_rxq *rxq) 4104 { 4105 if (rxq->iq.flags & IQ_HW_ALLOCATED) { 4106 MPASS(rxq->iq.flags & IQ_SW_ALLOCATED); 4107 free_iq_fl_hwq(vi->adapter, &rxq->iq, &rxq->fl); 4108 MPASS(!(rxq->iq.flags & IQ_HW_ALLOCATED)); 4109 } 4110 4111 if (rxq->iq.flags & IQ_SW_ALLOCATED) { 4112 MPASS(!(rxq->iq.flags & IQ_HW_ALLOCATED)); 4113 #if defined(INET) || defined(INET6) 4114 tcp_lro_free(&rxq->lro); 4115 #endif 4116 free_iq_fl(vi->adapter, &rxq->iq, &rxq->fl); 4117 MPASS(!(rxq->iq.flags & IQ_SW_ALLOCATED)); 4118 bzero(rxq, sizeof(*rxq)); 4119 } 4120 } 4121 4122 static void 4123 add_rxq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid, 4124 struct sge_rxq *rxq) 4125 { 4126 struct sysctl_oid_list *children; 4127 4128 if (ctx == NULL || oid == NULL) 4129 return; 4130 4131 children = SYSCTL_CHILDREN(oid); 4132 #if defined(INET) || defined(INET6) 4133 SYSCTL_ADD_U64(ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD, 4134 &rxq->lro.lro_queued, 0, NULL); 4135 SYSCTL_ADD_U64(ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD, 4136 &rxq->lro.lro_flushed, 0, NULL); 4137 #endif 4138 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD, 4139 &rxq->rxcsum, "# of times hardware assisted with checksum"); 4140 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vlan_extraction", CTLFLAG_RD, 4141 &rxq->vlan_extraction, "# of times hardware extracted 802.1Q tag"); 4142 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vxlan_rxcsum", CTLFLAG_RD, 4143 &rxq->vxlan_rxcsum, 4144 "# of times hardware assisted with inner checksum (VXLAN)"); 4145 } 4146 4147 #ifdef TCP_OFFLOAD 4148 /* 4149 * Idempotent. 4150 */ 4151 static int 4152 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq, int idx, 4153 int intr_idx, int maxp) 4154 { 4155 int rc; 4156 struct adapter *sc = vi->adapter; 4157 struct sysctl_oid *oid; 4158 char name[16]; 4159 4160 if (!(ofld_rxq->iq.flags & IQ_SW_ALLOCATED)) { 4161 MPASS(!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED)); 4162 4163 snprintf(name, sizeof(name), "%d", idx); 4164 oid = SYSCTL_ADD_NODE(&vi->ctx, 4165 SYSCTL_CHILDREN(vi->ofld_rxq_oid), OID_AUTO, name, 4166 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "offload rx queue"); 4167 4168 init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx, 4169 vi->qsize_rxq, intr_idx, ofld_cong_drop, IQ_OFLD); 4170 snprintf(name, sizeof(name), "%s ofld_rxq%d-fl", 4171 device_get_nameunit(vi->dev), idx); 4172 init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name); 4173 rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, &vi->ctx, 4174 oid); 4175 if (rc != 0) { 4176 CH_ERR(vi, "failed to allocate ofld_rxq%d: %d\n", idx, 4177 rc); 4178 sysctl_remove_oid(oid, 1, 1); 4179 return (rc); 4180 } 4181 MPASS(ofld_rxq->iq.flags & IQ_SW_ALLOCATED); 4182 ofld_rxq->rx_iscsi_ddp_setup_ok = counter_u64_alloc(M_WAITOK); 4183 ofld_rxq->rx_iscsi_ddp_setup_error = 4184 counter_u64_alloc(M_WAITOK); 4185 ofld_rxq->rx_nvme_ddp_setup_ok = counter_u64_alloc(M_WAITOK); 4186 ofld_rxq->rx_nvme_ddp_setup_no_stag = 4187 counter_u64_alloc(M_WAITOK); 4188 ofld_rxq->rx_nvme_ddp_setup_error = 4189 counter_u64_alloc(M_WAITOK); 4190 ofld_rxq->rx_nvme_ddp_octets = counter_u64_alloc(M_WAITOK); 4191 ofld_rxq->rx_nvme_ddp_pdus = counter_u64_alloc(M_WAITOK); 4192 ofld_rxq->rx_nvme_fl_octets = counter_u64_alloc(M_WAITOK); 4193 ofld_rxq->rx_nvme_fl_pdus = counter_u64_alloc(M_WAITOK); 4194 ofld_rxq->rx_nvme_invalid_headers = counter_u64_alloc(M_WAITOK); 4195 ofld_rxq->rx_nvme_header_digest_errors = 4196 counter_u64_alloc(M_WAITOK); 4197 ofld_rxq->rx_nvme_data_digest_errors = 4198 counter_u64_alloc(M_WAITOK); 4199 ofld_rxq->ddp_buffer_alloc = counter_u64_alloc(M_WAITOK); 4200 ofld_rxq->ddp_buffer_reuse = counter_u64_alloc(M_WAITOK); 4201 ofld_rxq->ddp_buffer_free = counter_u64_alloc(M_WAITOK); 4202 add_ofld_rxq_sysctls(&vi->ctx, oid, ofld_rxq); 4203 } 4204 4205 if (!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED)) { 4206 MPASS(ofld_rxq->iq.flags & IQ_SW_ALLOCATED); 4207 rc = alloc_iq_fl_hwq(vi, &ofld_rxq->iq, &ofld_rxq->fl); 4208 if (rc != 0) { 4209 CH_ERR(vi, "failed to create hw ofld_rxq%d: %d\n", idx, 4210 rc); 4211 return (rc); 4212 } 4213 MPASS(ofld_rxq->iq.flags & IQ_HW_ALLOCATED); 4214 } 4215 return (rc); 4216 } 4217 4218 /* 4219 * Idempotent. 4220 */ 4221 static void 4222 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq) 4223 { 4224 if (ofld_rxq->iq.flags & IQ_HW_ALLOCATED) { 4225 MPASS(ofld_rxq->iq.flags & IQ_SW_ALLOCATED); 4226 free_iq_fl_hwq(vi->adapter, &ofld_rxq->iq, &ofld_rxq->fl); 4227 MPASS(!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED)); 4228 } 4229 4230 if (ofld_rxq->iq.flags & IQ_SW_ALLOCATED) { 4231 MPASS(!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED)); 4232 free_iq_fl(vi->adapter, &ofld_rxq->iq, &ofld_rxq->fl); 4233 MPASS(!(ofld_rxq->iq.flags & IQ_SW_ALLOCATED)); 4234 counter_u64_free(ofld_rxq->rx_iscsi_ddp_setup_ok); 4235 counter_u64_free(ofld_rxq->rx_iscsi_ddp_setup_error); 4236 counter_u64_free(ofld_rxq->rx_nvme_ddp_setup_ok); 4237 counter_u64_free(ofld_rxq->rx_nvme_ddp_setup_no_stag); 4238 counter_u64_free(ofld_rxq->rx_nvme_ddp_setup_error); 4239 counter_u64_free(ofld_rxq->rx_nvme_ddp_octets); 4240 counter_u64_free(ofld_rxq->rx_nvme_ddp_pdus); 4241 counter_u64_free(ofld_rxq->rx_nvme_fl_octets); 4242 counter_u64_free(ofld_rxq->rx_nvme_fl_pdus); 4243 counter_u64_free(ofld_rxq->rx_nvme_invalid_headers); 4244 counter_u64_free(ofld_rxq->rx_nvme_header_digest_errors); 4245 counter_u64_free(ofld_rxq->rx_nvme_data_digest_errors); 4246 counter_u64_free(ofld_rxq->ddp_buffer_alloc); 4247 counter_u64_free(ofld_rxq->ddp_buffer_reuse); 4248 counter_u64_free(ofld_rxq->ddp_buffer_free); 4249 bzero(ofld_rxq, sizeof(*ofld_rxq)); 4250 } 4251 } 4252 4253 static void 4254 add_ofld_rxq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid, 4255 struct sge_ofld_rxq *ofld_rxq) 4256 { 4257 struct sysctl_oid_list *children, *top; 4258 4259 if (ctx == NULL || oid == NULL) 4260 return; 4261 4262 top = children = SYSCTL_CHILDREN(oid); 4263 SYSCTL_ADD_U64(ctx, children, OID_AUTO, "rx_aio_ddp_jobs", 4264 CTLFLAG_RD, &ofld_rxq->rx_aio_ddp_jobs, 0, 4265 "# of aio_read(2) jobs completed via DDP"); 4266 SYSCTL_ADD_U64(ctx, children, OID_AUTO, "rx_aio_ddp_octets", 4267 CTLFLAG_RD, &ofld_rxq->rx_aio_ddp_octets, 0, 4268 "# of octets placed directly for aio_read(2) jobs"); 4269 SYSCTL_ADD_ULONG(ctx, children, OID_AUTO, 4270 "rx_toe_tls_records", CTLFLAG_RD, &ofld_rxq->rx_toe_tls_records, 4271 "# of TOE TLS records received"); 4272 SYSCTL_ADD_ULONG(ctx, children, OID_AUTO, 4273 "rx_toe_tls_octets", CTLFLAG_RD, &ofld_rxq->rx_toe_tls_octets, 4274 "# of payload octets in received TOE TLS records"); 4275 SYSCTL_ADD_ULONG(ctx, children, OID_AUTO, 4276 "rx_toe_ddp_octets", CTLFLAG_RD, &ofld_rxq->rx_toe_ddp_octets, 4277 "# of payload octets received via TCP DDP"); 4278 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, 4279 "ddp_buffer_alloc", CTLFLAG_RD, &ofld_rxq->ddp_buffer_alloc, 4280 "# of DDP RCV buffers allocated"); 4281 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, 4282 "ddp_buffer_reuse", CTLFLAG_RD, &ofld_rxq->ddp_buffer_reuse, 4283 "# of DDP RCV buffers reused"); 4284 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, 4285 "ddp_buffer_free", CTLFLAG_RD, &ofld_rxq->ddp_buffer_free, 4286 "# of DDP RCV buffers freed"); 4287 4288 oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "iscsi", 4289 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "TOE iSCSI statistics"); 4290 children = SYSCTL_CHILDREN(oid); 4291 4292 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_setup_ok", 4293 CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_setup_ok, 4294 "# of times DDP buffer was setup successfully."); 4295 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_setup_error", 4296 CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_setup_error, 4297 "# of times DDP buffer setup failed."); 4298 SYSCTL_ADD_U64(ctx, children, OID_AUTO, "ddp_octets", 4299 CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_octets, 0, 4300 "# of octets placed directly"); 4301 SYSCTL_ADD_U64(ctx, children, OID_AUTO, "ddp_pdus", 4302 CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_pdus, 0, 4303 "# of PDUs with data placed directly."); 4304 SYSCTL_ADD_U64(ctx, children, OID_AUTO, "fl_octets", 4305 CTLFLAG_RD, &ofld_rxq->rx_iscsi_fl_octets, 0, 4306 "# of data octets delivered in freelist"); 4307 SYSCTL_ADD_U64(ctx, children, OID_AUTO, "fl_pdus", 4308 CTLFLAG_RD, &ofld_rxq->rx_iscsi_fl_pdus, 0, 4309 "# of PDUs with data delivered in freelist"); 4310 SYSCTL_ADD_U64(ctx, children, OID_AUTO, "padding_errors", 4311 CTLFLAG_RD, &ofld_rxq->rx_iscsi_padding_errors, 0, 4312 "# of PDUs with invalid padding"); 4313 SYSCTL_ADD_U64(ctx, children, OID_AUTO, "header_digest_errors", 4314 CTLFLAG_RD, &ofld_rxq->rx_iscsi_header_digest_errors, 0, 4315 "# of PDUs with invalid header digests"); 4316 SYSCTL_ADD_U64(ctx, children, OID_AUTO, "data_digest_errors", 4317 CTLFLAG_RD, &ofld_rxq->rx_iscsi_data_digest_errors, 0, 4318 "# of PDUs with invalid data digests"); 4319 4320 oid = SYSCTL_ADD_NODE(ctx, top, OID_AUTO, "nvme", 4321 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "TOE NVMe statistics"); 4322 children = SYSCTL_CHILDREN(oid); 4323 4324 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_setup_ok", 4325 CTLFLAG_RD, &ofld_rxq->rx_nvme_ddp_setup_ok, 4326 "# of times DDP buffer was setup successfully"); 4327 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_setup_no_stag", 4328 CTLFLAG_RD, &ofld_rxq->rx_nvme_ddp_setup_no_stag, 4329 "# of times STAG was not available for DDP buffer setup"); 4330 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_setup_error", 4331 CTLFLAG_RD, &ofld_rxq->rx_nvme_ddp_setup_error, 4332 "# of times DDP buffer setup failed"); 4333 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_octets", 4334 CTLFLAG_RD, &ofld_rxq->rx_nvme_ddp_octets, 4335 "# of octets placed directly"); 4336 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_pdus", 4337 CTLFLAG_RD, &ofld_rxq->rx_nvme_ddp_pdus, 4338 "# of PDUs with data placed directly"); 4339 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "fl_octets", 4340 CTLFLAG_RD, &ofld_rxq->rx_nvme_fl_octets, 4341 "# of data octets delivered in freelist"); 4342 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "fl_pdus", 4343 CTLFLAG_RD, &ofld_rxq->rx_nvme_fl_pdus, 4344 "# of PDUs with data delivered in freelist"); 4345 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "invalid_headers", 4346 CTLFLAG_RD, &ofld_rxq->rx_nvme_invalid_headers, 4347 "# of PDUs with invalid header field"); 4348 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "header_digest_errors", 4349 CTLFLAG_RD, &ofld_rxq->rx_nvme_header_digest_errors, 4350 "# of PDUs with invalid header digests"); 4351 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "data_digest_errors", 4352 CTLFLAG_RD, &ofld_rxq->rx_nvme_data_digest_errors, 4353 "# of PDUs with invalid data digests"); 4354 } 4355 #endif 4356 4357 /* 4358 * Returns a reasonable automatic cidx flush threshold for a given queue size. 4359 */ 4360 static u_int 4361 qsize_to_fthresh(int qsize) 4362 { 4363 u_int fthresh; 4364 4365 fthresh = qsize == 0 ? 0 : order_base_2(qsize); 4366 if (fthresh > X_CIDXFLUSHTHRESH_128) 4367 fthresh = X_CIDXFLUSHTHRESH_128; 4368 4369 return (fthresh); 4370 } 4371 4372 static int 4373 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq, int idx) 4374 { 4375 int rc, cntxt_id; 4376 struct fw_eq_ctrl_cmd c; 4377 int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE; 4378 4379 bzero(&c, sizeof(c)); 4380 4381 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST | 4382 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) | 4383 V_FW_EQ_CTRL_CMD_VFN(0)); 4384 c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC | 4385 V_FW_EQ_CTRL_CMD_COREGROUP(idx % sc->params.ncores) | 4386 F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c)); 4387 c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); 4388 c.physeqid_pkd = htobe32(0); 4389 c.fetchszm_to_iqid = 4390 htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) | 4391 V_FW_EQ_CTRL_CMD_PCIECHN(eq->hw_port) | 4392 F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid)); 4393 c.dcaen_to_eqsize = 4394 htobe32(V_FW_EQ_CTRL_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ? 4395 X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) | 4396 V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) | 4397 V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) | 4398 V_FW_EQ_CTRL_CMD_EQSIZE(qsize)); 4399 c.eqaddr = htobe64(eq->ba); 4400 4401 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); 4402 if (rc != 0) { 4403 CH_ERR(sc, "failed to create hw ctrlq for port %d: %d\n", 4404 eq->port_id, rc); 4405 return (rc); 4406 } 4407 4408 eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid)); 4409 eq->abs_id = G_FW_EQ_CTRL_CMD_PHYSEQID(be32toh(c.physeqid_pkd)); 4410 cntxt_id = eq->cntxt_id - sc->sge.eq_start; 4411 if (cntxt_id >= sc->sge.eqmap_sz) 4412 panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__, 4413 cntxt_id, sc->sge.eqmap_sz - 1); 4414 sc->sge.eqmap[cntxt_id] = eq; 4415 4416 return (rc); 4417 } 4418 4419 static int 4420 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq, int idx) 4421 { 4422 int rc, cntxt_id; 4423 struct fw_eq_eth_cmd c; 4424 int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE; 4425 4426 bzero(&c, sizeof(c)); 4427 4428 c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST | 4429 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) | 4430 V_FW_EQ_ETH_CMD_VFN(0)); 4431 c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC | 4432 V_FW_EQ_ETH_CMD_COREGROUP(idx % sc->params.ncores) | 4433 F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c)); 4434 c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE | 4435 F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid)); 4436 c.fetchszm_to_iqid = 4437 htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) | 4438 V_FW_EQ_ETH_CMD_PCIECHN(eq->hw_port) | F_FW_EQ_ETH_CMD_FETCHRO | 4439 V_FW_EQ_ETH_CMD_IQID(eq->iqid)); 4440 c.dcaen_to_eqsize = 4441 htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ? 4442 X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) | 4443 V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) | 4444 V_FW_EQ_ETH_CMD_EQSIZE(qsize)); 4445 c.eqaddr = htobe64(eq->ba); 4446 4447 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); 4448 if (rc != 0) { 4449 device_printf(vi->dev, 4450 "failed to create Ethernet egress queue: %d\n", rc); 4451 return (rc); 4452 } 4453 4454 eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd)); 4455 eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd)); 4456 cntxt_id = eq->cntxt_id - sc->sge.eq_start; 4457 if (cntxt_id >= sc->sge.eqmap_sz) 4458 panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__, 4459 cntxt_id, sc->sge.eqmap_sz - 1); 4460 sc->sge.eqmap[cntxt_id] = eq; 4461 4462 return (rc); 4463 } 4464 4465 #if defined(TCP_OFFLOAD) || defined(RATELIMIT) 4466 /* 4467 * ncores number of uP cores. 4468 * nq number of queues for this VI 4469 * idx queue index 4470 */ 4471 static inline int 4472 qidx_to_core(int ncores, int nq, int idx) 4473 { 4474 MPASS(nq % ncores == 0); 4475 MPASS(idx >= 0 && idx < nq); 4476 4477 return (idx * ncores / nq); 4478 } 4479 4480 static int 4481 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq, 4482 int idx) 4483 { 4484 int rc, cntxt_id, core; 4485 struct fw_eq_ofld_cmd c; 4486 int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE; 4487 4488 if (sc->params.tid_qid_sel_mask != 0) 4489 core = qidx_to_core(sc->params.ncores, vi->nofldtxq, idx); 4490 else 4491 core = 0; 4492 4493 bzero(&c, sizeof(c)); 4494 4495 c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST | 4496 F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) | 4497 V_FW_EQ_OFLD_CMD_VFN(0)); 4498 c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC | 4499 V_FW_EQ_OFLD_CMD_COREGROUP(core) | 4500 F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c)); 4501 c.fetchszm_to_iqid = 4502 htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) | 4503 V_FW_EQ_OFLD_CMD_PCIECHN(eq->hw_port) | 4504 F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid)); 4505 c.dcaen_to_eqsize = 4506 htobe32(V_FW_EQ_OFLD_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ? 4507 X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) | 4508 V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) | 4509 V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) | 4510 V_FW_EQ_OFLD_CMD_EQSIZE(qsize)); 4511 c.eqaddr = htobe64(eq->ba); 4512 4513 rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); 4514 if (rc != 0) { 4515 device_printf(vi->dev, 4516 "failed to create egress queue for TCP offload: %d\n", rc); 4517 return (rc); 4518 } 4519 4520 eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd)); 4521 eq->abs_id = G_FW_EQ_OFLD_CMD_PHYSEQID(be32toh(c.physeqid_pkd)); 4522 cntxt_id = eq->cntxt_id - sc->sge.eq_start; 4523 if (cntxt_id >= sc->sge.eqmap_sz) 4524 panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__, 4525 cntxt_id, sc->sge.eqmap_sz - 1); 4526 sc->sge.eqmap[cntxt_id] = eq; 4527 4528 return (rc); 4529 } 4530 #endif 4531 4532 /* SW only */ 4533 static int 4534 alloc_eq(struct adapter *sc, struct sge_eq *eq, struct sysctl_ctx_list *ctx, 4535 struct sysctl_oid *oid) 4536 { 4537 int rc, qsize; 4538 size_t len; 4539 4540 MPASS(!(eq->flags & EQ_SW_ALLOCATED)); 4541 4542 qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE; 4543 len = qsize * EQ_ESIZE; 4544 rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map, &eq->ba, 4545 (void **)&eq->desc); 4546 if (rc) 4547 return (rc); 4548 if (ctx != NULL && oid != NULL) 4549 add_eq_sysctls(sc, ctx, oid, eq); 4550 eq->flags |= EQ_SW_ALLOCATED; 4551 4552 return (0); 4553 } 4554 4555 /* SW only */ 4556 static void 4557 free_eq(struct adapter *sc, struct sge_eq *eq) 4558 { 4559 MPASS(eq->flags & EQ_SW_ALLOCATED); 4560 if (eq->type == EQ_ETH) 4561 MPASS(eq->pidx == eq->cidx); 4562 4563 free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc); 4564 mtx_destroy(&eq->eq_lock); 4565 bzero(eq, sizeof(*eq)); 4566 } 4567 4568 static void 4569 add_eq_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx, 4570 struct sysctl_oid *oid, struct sge_eq *eq) 4571 { 4572 struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); 4573 4574 SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &eq->ba, 4575 "bus address of descriptor ring"); 4576 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL, 4577 eq->sidx * EQ_ESIZE + sc->params.sge.spg_len, 4578 "desc ring size in bytes"); 4579 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD, 4580 &eq->abs_id, 0, "absolute id of the queue"); 4581 SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD, 4582 &eq->cntxt_id, 0, "SGE context id of the queue"); 4583 SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &eq->cidx, 4584 0, "consumer index"); 4585 SYSCTL_ADD_U16(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &eq->pidx, 4586 0, "producer index"); 4587 SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL, 4588 eq->sidx, "status page index"); 4589 } 4590 4591 static int 4592 alloc_eq_hwq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq, int idx) 4593 { 4594 int rc; 4595 4596 MPASS(!(eq->flags & EQ_HW_ALLOCATED)); 4597 4598 eq->iqid = eq->iq->cntxt_id; 4599 eq->pidx = eq->cidx = eq->dbidx = 0; 4600 /* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */ 4601 eq->equeqidx = 0; 4602 eq->doorbells = sc->doorbells; 4603 bzero(eq->desc, eq->sidx * EQ_ESIZE + sc->params.sge.spg_len); 4604 4605 switch (eq->type) { 4606 case EQ_CTRL: 4607 rc = ctrl_eq_alloc(sc, eq, idx); 4608 break; 4609 4610 case EQ_ETH: 4611 rc = eth_eq_alloc(sc, vi, eq, idx); 4612 break; 4613 4614 #if defined(TCP_OFFLOAD) || defined(RATELIMIT) 4615 case EQ_OFLD: 4616 rc = ofld_eq_alloc(sc, vi, eq, idx); 4617 break; 4618 #endif 4619 4620 default: 4621 panic("%s: invalid eq type %d.", __func__, eq->type); 4622 } 4623 if (rc != 0) { 4624 CH_ERR(sc, "failed to allocate egress queue(%d): %d\n", 4625 eq->type, rc); 4626 return (rc); 4627 } 4628 4629 if (isset(&eq->doorbells, DOORBELL_UDB) || 4630 isset(&eq->doorbells, DOORBELL_UDBWC) || 4631 isset(&eq->doorbells, DOORBELL_WCWR)) { 4632 uint32_t s_qpp = sc->params.sge.eq_s_qpp; 4633 uint32_t mask = (1 << s_qpp) - 1; 4634 volatile uint8_t *udb; 4635 4636 udb = sc->udbs_base + UDBS_DB_OFFSET; 4637 udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT; /* pg offset */ 4638 eq->udb_qid = eq->cntxt_id & mask; /* id in page */ 4639 if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE) 4640 clrbit(&eq->doorbells, DOORBELL_WCWR); 4641 else { 4642 udb += eq->udb_qid << UDBS_SEG_SHIFT; /* seg offset */ 4643 eq->udb_qid = 0; 4644 } 4645 eq->udb = (volatile void *)udb; 4646 } 4647 4648 eq->flags |= EQ_HW_ALLOCATED; 4649 return (0); 4650 } 4651 4652 static int 4653 free_eq_hwq(struct adapter *sc, struct vi_info *vi __unused, struct sge_eq *eq) 4654 { 4655 int rc; 4656 4657 MPASS(eq->flags & EQ_HW_ALLOCATED); 4658 4659 switch (eq->type) { 4660 case EQ_CTRL: 4661 rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id); 4662 break; 4663 case EQ_ETH: 4664 rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id); 4665 break; 4666 #if defined(TCP_OFFLOAD) || defined(RATELIMIT) 4667 case EQ_OFLD: 4668 rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id); 4669 break; 4670 #endif 4671 default: 4672 panic("%s: invalid eq type %d.", __func__, eq->type); 4673 } 4674 if (rc != 0) { 4675 CH_ERR(sc, "failed to free eq (type %d): %d\n", eq->type, rc); 4676 return (rc); 4677 } 4678 eq->flags &= ~EQ_HW_ALLOCATED; 4679 4680 return (0); 4681 } 4682 4683 static int 4684 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq, 4685 struct sysctl_ctx_list *ctx, struct sysctl_oid *oid) 4686 { 4687 struct sge_eq *eq = &wrq->eq; 4688 int rc; 4689 4690 MPASS(!(eq->flags & EQ_SW_ALLOCATED)); 4691 4692 rc = alloc_eq(sc, eq, ctx, oid); 4693 if (rc) 4694 return (rc); 4695 MPASS(eq->flags & EQ_SW_ALLOCATED); 4696 /* Can't fail after this. */ 4697 4698 wrq->adapter = sc; 4699 TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq); 4700 TAILQ_INIT(&wrq->incomplete_wrs); 4701 STAILQ_INIT(&wrq->wr_list); 4702 wrq->nwr_pending = 0; 4703 wrq->ndesc_needed = 0; 4704 add_wrq_sysctls(ctx, oid, wrq); 4705 4706 return (0); 4707 } 4708 4709 static void 4710 free_wrq(struct adapter *sc, struct sge_wrq *wrq) 4711 { 4712 free_eq(sc, &wrq->eq); 4713 MPASS(wrq->nwr_pending == 0); 4714 MPASS(wrq->ndesc_needed == 0); 4715 MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs)); 4716 MPASS(STAILQ_EMPTY(&wrq->wr_list)); 4717 bzero(wrq, sizeof(*wrq)); 4718 } 4719 4720 static void 4721 add_wrq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid, 4722 struct sge_wrq *wrq) 4723 { 4724 struct sysctl_oid_list *children; 4725 4726 if (ctx == NULL || oid == NULL) 4727 return; 4728 4729 children = SYSCTL_CHILDREN(oid); 4730 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD, 4731 &wrq->tx_wrs_direct, "# of work requests (direct)"); 4732 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD, 4733 &wrq->tx_wrs_copied, "# of work requests (copied)"); 4734 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD, 4735 &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)"); 4736 } 4737 4738 /* 4739 * Idempotent. 4740 */ 4741 static int 4742 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx) 4743 { 4744 int rc, iqidx; 4745 struct port_info *pi = vi->pi; 4746 struct adapter *sc = vi->adapter; 4747 struct sge_eq *eq = &txq->eq; 4748 struct txpkts *txp; 4749 char name[16]; 4750 struct sysctl_oid *oid; 4751 4752 if (!(eq->flags & EQ_SW_ALLOCATED)) { 4753 MPASS(!(eq->flags & EQ_HW_ALLOCATED)); 4754 4755 snprintf(name, sizeof(name), "%d", idx); 4756 oid = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(vi->txq_oid), 4757 OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 4758 "tx queue"); 4759 4760 iqidx = vi->first_rxq + (idx % vi->nrxq); 4761 snprintf(name, sizeof(name), "%s txq%d", 4762 device_get_nameunit(vi->dev), idx); 4763 init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->port_id, 4764 &sc->sge.rxq[iqidx].iq, name); 4765 4766 rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, 4767 can_resume_eth_tx, M_CXGBE, &eq->eq_lock, M_WAITOK); 4768 if (rc != 0) { 4769 CH_ERR(vi, "failed to allocate mp_ring for txq%d: %d\n", 4770 idx, rc); 4771 failed: 4772 sysctl_remove_oid(oid, 1, 1); 4773 return (rc); 4774 } 4775 4776 rc = alloc_eq(sc, eq, &vi->ctx, oid); 4777 if (rc) { 4778 CH_ERR(vi, "failed to allocate txq%d: %d\n", idx, rc); 4779 mp_ring_free(txq->r); 4780 goto failed; 4781 } 4782 MPASS(eq->flags & EQ_SW_ALLOCATED); 4783 /* Can't fail after this point. */ 4784 4785 TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq); 4786 txq->ifp = vi->ifp; 4787 txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK); 4788 txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE, 4789 M_ZERO | M_WAITOK); 4790 4791 add_txq_sysctls(vi, &vi->ctx, oid, txq); 4792 } 4793 4794 if (!(eq->flags & EQ_HW_ALLOCATED)) { 4795 MPASS(eq->flags & EQ_SW_ALLOCATED); 4796 rc = alloc_eq_hwq(sc, vi, eq, idx); 4797 if (rc != 0) { 4798 CH_ERR(vi, "failed to create hw txq%d: %d\n", idx, rc); 4799 return (rc); 4800 } 4801 MPASS(eq->flags & EQ_HW_ALLOCATED); 4802 /* Can't fail after this point. */ 4803 4804 if (idx == 0) 4805 sc->sge.eq_base = eq->abs_id - eq->cntxt_id; 4806 else 4807 KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id, 4808 ("eq_base mismatch")); 4809 KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF, 4810 ("PF with non-zero eq_base")); 4811 4812 txp = &txq->txp; 4813 MPASS(nitems(txp->mb) >= sc->params.max_pkts_per_eth_tx_pkts_wr); 4814 txq->txp.max_npkt = min(nitems(txp->mb), 4815 sc->params.max_pkts_per_eth_tx_pkts_wr); 4816 if (vi->flags & TX_USES_VM_WR && !(sc->flags & IS_VF)) 4817 txq->txp.max_npkt--; 4818 4819 if (vi->flags & TX_USES_VM_WR) 4820 txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) | 4821 V_TXPKT_INTF(pi->hw_port)); 4822 else 4823 txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) | 4824 V_TXPKT_INTF(pi->hw_port) | V_TXPKT_PF(sc->pf) | 4825 V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld)); 4826 4827 txq->tc_idx = -1; 4828 } 4829 4830 return (0); 4831 } 4832 4833 /* 4834 * Idempotent. 4835 */ 4836 static void 4837 free_txq(struct vi_info *vi, struct sge_txq *txq) 4838 { 4839 struct adapter *sc = vi->adapter; 4840 struct sge_eq *eq = &txq->eq; 4841 4842 if (eq->flags & EQ_HW_ALLOCATED) { 4843 MPASS(eq->flags & EQ_SW_ALLOCATED); 4844 free_eq_hwq(sc, NULL, eq); 4845 MPASS(!(eq->flags & EQ_HW_ALLOCATED)); 4846 } 4847 4848 if (eq->flags & EQ_SW_ALLOCATED) { 4849 MPASS(!(eq->flags & EQ_HW_ALLOCATED)); 4850 sglist_free(txq->gl); 4851 free(txq->sdesc, M_CXGBE); 4852 mp_ring_free(txq->r); 4853 free_eq(sc, eq); 4854 MPASS(!(eq->flags & EQ_SW_ALLOCATED)); 4855 bzero(txq, sizeof(*txq)); 4856 } 4857 } 4858 4859 static void 4860 add_txq_sysctls(struct vi_info *vi, struct sysctl_ctx_list *ctx, 4861 struct sysctl_oid *oid, struct sge_txq *txq) 4862 { 4863 struct adapter *sc; 4864 struct sysctl_oid_list *children; 4865 4866 if (ctx == NULL || oid == NULL) 4867 return; 4868 4869 sc = vi->adapter; 4870 children = SYSCTL_CHILDREN(oid); 4871 4872 mp_ring_sysctls(txq->r, ctx, children); 4873 4874 SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "tc", 4875 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, vi, txq - sc->sge.txq, 4876 sysctl_tc, "I", "traffic class (-1 means none)"); 4877 4878 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD, 4879 &txq->txcsum, "# of times hardware assisted with checksum"); 4880 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vlan_insertion", CTLFLAG_RD, 4881 &txq->vlan_insertion, "# of times hardware inserted 802.1Q tag"); 4882 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD, 4883 &txq->tso_wrs, "# of TSO work requests"); 4884 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD, 4885 &txq->imm_wrs, "# of work requests with immediate data"); 4886 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD, 4887 &txq->sgl_wrs, "# of work requests with direct SGL"); 4888 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD, 4889 &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)"); 4890 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts0_wrs", CTLFLAG_RD, 4891 &txq->txpkts0_wrs, "# of txpkts (type 0) work requests"); 4892 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts1_wrs", CTLFLAG_RD, 4893 &txq->txpkts1_wrs, "# of txpkts (type 1) work requests"); 4894 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts0_pkts", CTLFLAG_RD, 4895 &txq->txpkts0_pkts, 4896 "# of frames tx'd using type0 txpkts work requests"); 4897 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts1_pkts", CTLFLAG_RD, 4898 &txq->txpkts1_pkts, 4899 "# of frames tx'd using type1 txpkts work requests"); 4900 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts_flush", CTLFLAG_RD, 4901 &txq->txpkts_flush, 4902 "# of times txpkts had to be flushed out by an egress-update"); 4903 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD, 4904 &txq->raw_wrs, "# of raw work requests (non-packets)"); 4905 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vxlan_tso_wrs", CTLFLAG_RD, 4906 &txq->vxlan_tso_wrs, "# of VXLAN TSO work requests"); 4907 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vxlan_txcsum", CTLFLAG_RD, 4908 &txq->vxlan_txcsum, 4909 "# of times hardware assisted with inner checksums (VXLAN)"); 4910 4911 #ifdef KERN_TLS 4912 if (is_ktls(sc)) { 4913 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_records", 4914 CTLFLAG_RD, &txq->kern_tls_records, 4915 "# of NIC TLS records transmitted"); 4916 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_short", 4917 CTLFLAG_RD, &txq->kern_tls_short, 4918 "# of short NIC TLS records transmitted"); 4919 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_partial", 4920 CTLFLAG_RD, &txq->kern_tls_partial, 4921 "# of partial NIC TLS records transmitted"); 4922 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_full", 4923 CTLFLAG_RD, &txq->kern_tls_full, 4924 "# of full NIC TLS records transmitted"); 4925 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_octets", 4926 CTLFLAG_RD, &txq->kern_tls_octets, 4927 "# of payload octets in transmitted NIC TLS records"); 4928 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_waste", 4929 CTLFLAG_RD, &txq->kern_tls_waste, 4930 "# of octets DMAd but not transmitted in NIC TLS records"); 4931 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_header", 4932 CTLFLAG_RD, &txq->kern_tls_header, 4933 "# of NIC TLS header-only packets transmitted"); 4934 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_fin_short", 4935 CTLFLAG_RD, &txq->kern_tls_fin_short, 4936 "# of NIC TLS padded FIN packets on short TLS records"); 4937 if (is_t6(sc)) { 4938 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, 4939 "kern_tls_options", CTLFLAG_RD, 4940 &txq->kern_tls_options, 4941 "# of NIC TLS options-only packets transmitted"); 4942 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, 4943 "kern_tls_fin", CTLFLAG_RD, &txq->kern_tls_fin, 4944 "# of NIC TLS FIN-only packets transmitted"); 4945 } else { 4946 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, 4947 "kern_tls_ghash_received", CTLFLAG_RD, 4948 &txq->kern_tls_ghash_received, 4949 "# of NIC TLS GHASHes received"); 4950 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, 4951 "kern_tls_ghash_requested", CTLFLAG_RD, 4952 &txq->kern_tls_ghash_requested, 4953 "# of NIC TLS GHASHes requested"); 4954 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, 4955 "kern_tls_lso", CTLFLAG_RD, 4956 &txq->kern_tls_lso, 4957 "# of NIC TLS records transmitted using LSO"); 4958 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, 4959 "kern_tls_partial_ghash", CTLFLAG_RD, 4960 &txq->kern_tls_partial_ghash, 4961 "# of NIC TLS records encrypted using a partial GHASH"); 4962 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, 4963 "kern_tls_splitmode", CTLFLAG_RD, 4964 &txq->kern_tls_splitmode, 4965 "# of NIC TLS records using SplitMode"); 4966 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, 4967 "kern_tls_trailer", CTLFLAG_RD, 4968 &txq->kern_tls_trailer, 4969 "# of NIC TLS trailer-only packets transmitted"); 4970 } 4971 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_cbc", 4972 CTLFLAG_RD, &txq->kern_tls_cbc, 4973 "# of NIC TLS sessions using AES-CBC"); 4974 SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_gcm", 4975 CTLFLAG_RD, &txq->kern_tls_gcm, 4976 "# of NIC TLS sessions using AES-GCM"); 4977 } 4978 #endif 4979 } 4980 4981 #if defined(TCP_OFFLOAD) || defined(RATELIMIT) 4982 /* 4983 * Idempotent. 4984 */ 4985 static int 4986 alloc_ofld_txq(struct vi_info *vi, struct sge_ofld_txq *ofld_txq, int idx) 4987 { 4988 struct sysctl_oid *oid; 4989 struct port_info *pi = vi->pi; 4990 struct adapter *sc = vi->adapter; 4991 struct sge_eq *eq = &ofld_txq->wrq.eq; 4992 int rc, iqidx; 4993 char name[16]; 4994 4995 MPASS(idx >= 0); 4996 MPASS(idx < vi->nofldtxq); 4997 4998 if (!(eq->flags & EQ_SW_ALLOCATED)) { 4999 snprintf(name, sizeof(name), "%d", idx); 5000 oid = SYSCTL_ADD_NODE(&vi->ctx, 5001 SYSCTL_CHILDREN(vi->ofld_txq_oid), OID_AUTO, name, 5002 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "offload tx queue"); 5003 5004 snprintf(name, sizeof(name), "%s ofld_txq%d", 5005 device_get_nameunit(vi->dev), idx); 5006 if (vi->nofldrxq > 0) { 5007 iqidx = vi->first_ofld_rxq + (idx % vi->nofldrxq); 5008 init_eq(sc, eq, EQ_OFLD, vi->qsize_txq, pi->port_id, 5009 &sc->sge.ofld_rxq[iqidx].iq, name); 5010 } else { 5011 iqidx = vi->first_rxq + (idx % vi->nrxq); 5012 init_eq(sc, eq, EQ_OFLD, vi->qsize_txq, pi->port_id, 5013 &sc->sge.rxq[iqidx].iq, name); 5014 } 5015 5016 rc = alloc_wrq(sc, vi, &ofld_txq->wrq, &vi->ctx, oid); 5017 if (rc != 0) { 5018 CH_ERR(vi, "failed to allocate ofld_txq%d: %d\n", idx, 5019 rc); 5020 sysctl_remove_oid(oid, 1, 1); 5021 return (rc); 5022 } 5023 MPASS(eq->flags & EQ_SW_ALLOCATED); 5024 /* Can't fail after this point. */ 5025 5026 ofld_txq->tx_iscsi_pdus = counter_u64_alloc(M_WAITOK); 5027 ofld_txq->tx_iscsi_octets = counter_u64_alloc(M_WAITOK); 5028 ofld_txq->tx_iscsi_iso_wrs = counter_u64_alloc(M_WAITOK); 5029 ofld_txq->tx_nvme_pdus = counter_u64_alloc(M_WAITOK); 5030 ofld_txq->tx_nvme_octets = counter_u64_alloc(M_WAITOK); 5031 ofld_txq->tx_nvme_iso_wrs = counter_u64_alloc(M_WAITOK); 5032 ofld_txq->tx_aio_jobs = counter_u64_alloc(M_WAITOK); 5033 ofld_txq->tx_aio_octets = counter_u64_alloc(M_WAITOK); 5034 ofld_txq->tx_toe_tls_records = counter_u64_alloc(M_WAITOK); 5035 ofld_txq->tx_toe_tls_octets = counter_u64_alloc(M_WAITOK); 5036 add_ofld_txq_sysctls(&vi->ctx, oid, ofld_txq); 5037 } 5038 5039 if (!(eq->flags & EQ_HW_ALLOCATED)) { 5040 MPASS(eq->flags & EQ_SW_ALLOCATED); 5041 MPASS(ofld_txq->wrq.nwr_pending == 0); 5042 MPASS(ofld_txq->wrq.ndesc_needed == 0); 5043 rc = alloc_eq_hwq(sc, vi, eq, idx); 5044 if (rc != 0) { 5045 CH_ERR(vi, "failed to create hw ofld_txq%d: %d\n", idx, 5046 rc); 5047 return (rc); 5048 } 5049 MPASS(eq->flags & EQ_HW_ALLOCATED); 5050 } 5051 5052 return (0); 5053 } 5054 5055 /* 5056 * Idempotent. 5057 */ 5058 static void 5059 free_ofld_txq(struct vi_info *vi, struct sge_ofld_txq *ofld_txq) 5060 { 5061 struct adapter *sc = vi->adapter; 5062 struct sge_eq *eq = &ofld_txq->wrq.eq; 5063 5064 if (eq->flags & EQ_HW_ALLOCATED) { 5065 MPASS(eq->flags & EQ_SW_ALLOCATED); 5066 free_eq_hwq(sc, NULL, eq); 5067 MPASS(!(eq->flags & EQ_HW_ALLOCATED)); 5068 } 5069 5070 if (eq->flags & EQ_SW_ALLOCATED) { 5071 MPASS(!(eq->flags & EQ_HW_ALLOCATED)); 5072 counter_u64_free(ofld_txq->tx_iscsi_pdus); 5073 counter_u64_free(ofld_txq->tx_iscsi_octets); 5074 counter_u64_free(ofld_txq->tx_iscsi_iso_wrs); 5075 counter_u64_free(ofld_txq->tx_nvme_pdus); 5076 counter_u64_free(ofld_txq->tx_nvme_octets); 5077 counter_u64_free(ofld_txq->tx_nvme_iso_wrs); 5078 counter_u64_free(ofld_txq->tx_aio_jobs); 5079 counter_u64_free(ofld_txq->tx_aio_octets); 5080 counter_u64_free(ofld_txq->tx_toe_tls_records); 5081 counter_u64_free(ofld_txq->tx_toe_tls_octets); 5082 free_wrq(sc, &ofld_txq->wrq); 5083 MPASS(!(eq->flags & EQ_SW_ALLOCATED)); 5084 bzero(ofld_txq, sizeof(*ofld_txq)); 5085 } 5086 } 5087 5088 static void 5089 add_ofld_txq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid, 5090 struct sge_ofld_txq *ofld_txq) 5091 { 5092 struct sysctl_oid_list *children; 5093 5094 if (ctx == NULL || oid == NULL) 5095 return; 5096 5097 children = SYSCTL_CHILDREN(oid); 5098 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_iscsi_pdus", 5099 CTLFLAG_RD, &ofld_txq->tx_iscsi_pdus, 5100 "# of iSCSI PDUs transmitted"); 5101 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_iscsi_octets", 5102 CTLFLAG_RD, &ofld_txq->tx_iscsi_octets, 5103 "# of payload octets in transmitted iSCSI PDUs"); 5104 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_iscsi_iso_wrs", 5105 CTLFLAG_RD, &ofld_txq->tx_iscsi_iso_wrs, 5106 "# of iSCSI segmentation offload work requests"); 5107 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_nvme_pdus", 5108 CTLFLAG_RD, &ofld_txq->tx_nvme_pdus, 5109 "# of NVMe PDUs transmitted"); 5110 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_nvme_octets", 5111 CTLFLAG_RD, &ofld_txq->tx_nvme_octets, 5112 "# of payload octets in transmitted NVMe PDUs"); 5113 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_nvme_iso_wrs", 5114 CTLFLAG_RD, &ofld_txq->tx_nvme_iso_wrs, 5115 "# of NVMe segmentation offload work requests"); 5116 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_aio_jobs", 5117 CTLFLAG_RD, &ofld_txq->tx_aio_jobs, 5118 "# of zero-copy aio_write(2) jobs transmitted"); 5119 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_aio_octets", 5120 CTLFLAG_RD, &ofld_txq->tx_aio_octets, 5121 "# of payload octets in transmitted zero-copy aio_write(2) jobs"); 5122 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_toe_tls_records", 5123 CTLFLAG_RD, &ofld_txq->tx_toe_tls_records, 5124 "# of TOE TLS records transmitted"); 5125 SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_toe_tls_octets", 5126 CTLFLAG_RD, &ofld_txq->tx_toe_tls_octets, 5127 "# of payload octets in transmitted TOE TLS records"); 5128 } 5129 #endif 5130 5131 static void 5132 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error) 5133 { 5134 bus_addr_t *ba = arg; 5135 5136 KASSERT(nseg == 1, 5137 ("%s meant for single segment mappings only.", __func__)); 5138 5139 *ba = error ? 0 : segs->ds_addr; 5140 } 5141 5142 static inline void 5143 ring_fl_db(struct adapter *sc, struct sge_fl *fl) 5144 { 5145 uint32_t n, v; 5146 5147 n = IDXDIFF(fl->pidx >> 3, fl->dbidx, fl->sidx); 5148 MPASS(n > 0); 5149 5150 wmb(); 5151 v = fl->dbval | V_PIDX(n); 5152 if (fl->udb) 5153 *fl->udb = htole32(v); 5154 else 5155 t4_write_reg(sc, sc->sge_kdoorbell_reg, v); 5156 IDXINCR(fl->dbidx, n, fl->sidx); 5157 } 5158 5159 /* 5160 * Fills up the freelist by allocating up to 'n' buffers. Buffers that are 5161 * recycled do not count towards this allocation budget. 5162 * 5163 * Returns non-zero to indicate that this freelist should be added to the list 5164 * of starving freelists. 5165 */ 5166 static int 5167 refill_fl(struct adapter *sc, struct sge_fl *fl, int n) 5168 { 5169 __be64 *d; 5170 struct fl_sdesc *sd; 5171 uintptr_t pa; 5172 caddr_t cl; 5173 struct rx_buf_info *rxb; 5174 struct cluster_metadata *clm; 5175 uint16_t max_pidx, zidx = fl->zidx; 5176 uint16_t hw_cidx = fl->hw_cidx; /* stable snapshot */ 5177 5178 FL_LOCK_ASSERT_OWNED(fl); 5179 5180 /* 5181 * We always stop at the beginning of the hardware descriptor that's just 5182 * before the one with the hw cidx. This is to avoid hw pidx = hw cidx, 5183 * which would mean an empty freelist to the chip. 5184 */ 5185 max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1; 5186 if (fl->pidx == max_pidx * 8) 5187 return (0); 5188 5189 d = &fl->desc[fl->pidx]; 5190 sd = &fl->sdesc[fl->pidx]; 5191 rxb = &sc->sge.rx_buf_info[zidx]; 5192 5193 while (n > 0) { 5194 5195 if (sd->cl != NULL) { 5196 5197 if (sd->nmbuf == 0) { 5198 /* 5199 * Fast recycle without involving any atomics on 5200 * the cluster's metadata (if the cluster has 5201 * metadata). This happens when all frames 5202 * received in the cluster were small enough to 5203 * fit within a single mbuf each. 5204 */ 5205 fl->cl_fast_recycled++; 5206 goto recycled; 5207 } 5208 5209 /* 5210 * Cluster is guaranteed to have metadata. Clusters 5211 * without metadata always take the fast recycle path 5212 * when they're recycled. 5213 */ 5214 clm = cl_metadata(sd); 5215 MPASS(clm != NULL); 5216 5217 if (atomic_fetchadd_int(&clm->refcount, -1) == 1) { 5218 fl->cl_recycled++; 5219 counter_u64_add(extfree_rels, 1); 5220 goto recycled; 5221 } 5222 sd->cl = NULL; /* gave up my reference */ 5223 } 5224 MPASS(sd->cl == NULL); 5225 cl = uma_zalloc(rxb->zone, M_NOWAIT); 5226 if (__predict_false(cl == NULL)) { 5227 if (zidx != fl->safe_zidx) { 5228 zidx = fl->safe_zidx; 5229 rxb = &sc->sge.rx_buf_info[zidx]; 5230 cl = uma_zalloc(rxb->zone, M_NOWAIT); 5231 } 5232 if (cl == NULL) 5233 break; 5234 } 5235 fl->cl_allocated++; 5236 n--; 5237 5238 pa = pmap_kextract((vm_offset_t)cl); 5239 sd->cl = cl; 5240 sd->zidx = zidx; 5241 5242 if (fl->flags & FL_BUF_PACKING) { 5243 *d = htobe64(pa | rxb->hwidx2); 5244 sd->moff = rxb->size2; 5245 } else { 5246 *d = htobe64(pa | rxb->hwidx1); 5247 sd->moff = 0; 5248 } 5249 recycled: 5250 sd->nmbuf = 0; 5251 d++; 5252 sd++; 5253 if (__predict_false((++fl->pidx & 7) == 0)) { 5254 uint16_t pidx = fl->pidx >> 3; 5255 5256 if (__predict_false(pidx == fl->sidx)) { 5257 fl->pidx = 0; 5258 pidx = 0; 5259 sd = fl->sdesc; 5260 d = fl->desc; 5261 } 5262 if (n < 8 || pidx == max_pidx) 5263 break; 5264 5265 if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4) 5266 ring_fl_db(sc, fl); 5267 } 5268 } 5269 5270 if ((fl->pidx >> 3) != fl->dbidx) 5271 ring_fl_db(sc, fl); 5272 5273 return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING)); 5274 } 5275 5276 /* 5277 * Attempt to refill all starving freelists. 5278 */ 5279 static void 5280 refill_sfl(void *arg) 5281 { 5282 struct adapter *sc = arg; 5283 struct sge_fl *fl, *fl_temp; 5284 5285 mtx_assert(&sc->sfl_lock, MA_OWNED); 5286 TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) { 5287 FL_LOCK(fl); 5288 refill_fl(sc, fl, 64); 5289 if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) { 5290 TAILQ_REMOVE(&sc->sfl, fl, link); 5291 fl->flags &= ~FL_STARVING; 5292 } 5293 FL_UNLOCK(fl); 5294 } 5295 5296 if (!TAILQ_EMPTY(&sc->sfl)) 5297 callout_schedule(&sc->sfl_callout, hz / 5); 5298 } 5299 5300 /* 5301 * Release the driver's reference on all buffers in the given freelist. Buffers 5302 * with kernel references cannot be freed and will prevent the driver from being 5303 * unloaded safely. 5304 */ 5305 void 5306 free_fl_buffers(struct adapter *sc, struct sge_fl *fl) 5307 { 5308 struct fl_sdesc *sd; 5309 struct cluster_metadata *clm; 5310 int i; 5311 5312 sd = fl->sdesc; 5313 for (i = 0; i < fl->sidx * 8; i++, sd++) { 5314 if (sd->cl == NULL) 5315 continue; 5316 5317 if (sd->nmbuf == 0) 5318 uma_zfree(sc->sge.rx_buf_info[sd->zidx].zone, sd->cl); 5319 else if (fl->flags & FL_BUF_PACKING) { 5320 clm = cl_metadata(sd); 5321 if (atomic_fetchadd_int(&clm->refcount, -1) == 1) { 5322 uma_zfree(sc->sge.rx_buf_info[sd->zidx].zone, 5323 sd->cl); 5324 counter_u64_add(extfree_rels, 1); 5325 } 5326 } 5327 sd->cl = NULL; 5328 } 5329 5330 if (fl->flags & FL_BUF_RESUME) { 5331 m_freem(fl->m0); 5332 fl->flags &= ~FL_BUF_RESUME; 5333 } 5334 } 5335 5336 static inline void 5337 get_pkt_gl(struct mbuf *m, struct sglist *gl) 5338 { 5339 int rc; 5340 5341 M_ASSERTPKTHDR(m); 5342 5343 sglist_reset(gl); 5344 rc = sglist_append_mbuf(gl, m); 5345 if (__predict_false(rc != 0)) { 5346 panic("%s: mbuf %p (%d segs) was vetted earlier but now fails " 5347 "with %d.", __func__, m, mbuf_nsegs(m), rc); 5348 } 5349 5350 KASSERT(gl->sg_nseg == mbuf_nsegs(m), 5351 ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m, 5352 mbuf_nsegs(m), gl->sg_nseg)); 5353 #if 0 /* vm_wr not readily available here. */ 5354 KASSERT(gl->sg_nseg > 0 && gl->sg_nseg <= max_nsegs_allowed(m, vm_wr), 5355 ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__, 5356 gl->sg_nseg, max_nsegs_allowed(m, vm_wr))); 5357 #endif 5358 } 5359 5360 /* 5361 * len16 for a txpkt WR with a GL. Includes the firmware work request header. 5362 */ 5363 static inline u_int 5364 txpkt_len16(u_int nsegs, const u_int extra) 5365 { 5366 u_int n; 5367 5368 MPASS(nsegs > 0); 5369 5370 nsegs--; /* first segment is part of ulptx_sgl */ 5371 n = extra + sizeof(struct fw_eth_tx_pkt_wr) + 5372 sizeof(struct cpl_tx_pkt_core) + 5373 sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1)); 5374 5375 return (howmany(n, 16)); 5376 } 5377 5378 /* 5379 * len16 for a txpkt_vm WR with a GL. Includes the firmware work 5380 * request header. 5381 */ 5382 static inline u_int 5383 txpkt_vm_len16(u_int nsegs, const u_int extra) 5384 { 5385 u_int n; 5386 5387 MPASS(nsegs > 0); 5388 5389 nsegs--; /* first segment is part of ulptx_sgl */ 5390 n = extra + sizeof(struct fw_eth_tx_pkt_vm_wr) + 5391 sizeof(struct cpl_tx_pkt_core) + 5392 sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1)); 5393 5394 return (howmany(n, 16)); 5395 } 5396 5397 static inline void 5398 calculate_mbuf_len16(struct mbuf *m, bool vm_wr) 5399 { 5400 const int lso = sizeof(struct cpl_tx_pkt_lso_core); 5401 const int tnl_lso = sizeof(struct cpl_tx_tnl_lso); 5402 5403 if (vm_wr) { 5404 if (needs_tso(m)) 5405 set_mbuf_len16(m, txpkt_vm_len16(mbuf_nsegs(m), lso)); 5406 else 5407 set_mbuf_len16(m, txpkt_vm_len16(mbuf_nsegs(m), 0)); 5408 return; 5409 } 5410 5411 if (needs_tso(m)) { 5412 if (needs_vxlan_tso(m)) 5413 set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), tnl_lso)); 5414 else 5415 set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), lso)); 5416 } else 5417 set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), 0)); 5418 } 5419 5420 /* 5421 * len16 for a txpkts type 0 WR with a GL. Does not include the firmware work 5422 * request header. 5423 */ 5424 static inline u_int 5425 txpkts0_len16(u_int nsegs) 5426 { 5427 u_int n; 5428 5429 MPASS(nsegs > 0); 5430 5431 nsegs--; /* first segment is part of ulptx_sgl */ 5432 n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) + 5433 sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) + 5434 8 * ((3 * nsegs) / 2 + (nsegs & 1)); 5435 5436 return (howmany(n, 16)); 5437 } 5438 5439 /* 5440 * len16 for a txpkts type 1 WR with a GL. Does not include the firmware work 5441 * request header. 5442 */ 5443 static inline u_int 5444 txpkts1_len16(void) 5445 { 5446 u_int n; 5447 5448 n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl); 5449 5450 return (howmany(n, 16)); 5451 } 5452 5453 static inline u_int 5454 imm_payload(u_int ndesc) 5455 { 5456 u_int n; 5457 5458 n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) - 5459 sizeof(struct cpl_tx_pkt_core); 5460 5461 return (n); 5462 } 5463 5464 static inline uint64_t 5465 csum_to_ctrl(struct adapter *sc, struct mbuf *m) 5466 { 5467 uint64_t ctrl; 5468 int csum_type, l2hlen, l3hlen; 5469 int x, y; 5470 static const uint8_t csum_types[3][2] = { 5471 {TX_CSUM_TCPIP, TX_CSUM_TCPIP6}, 5472 {TX_CSUM_UDPIP, TX_CSUM_UDPIP6}, 5473 {TX_CSUM_IP, 0} 5474 }; 5475 5476 M_ASSERTPKTHDR(m); 5477 5478 if (!needs_hwcsum(m)) 5479 return (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS); 5480 5481 MPASS(m->m_pkthdr.l2hlen >= ETHER_HDR_LEN); 5482 MPASS(m->m_pkthdr.l3hlen >= sizeof(struct ip)); 5483 5484 if (needs_vxlan_csum(m)) { 5485 MPASS(m->m_pkthdr.l4hlen > 0); 5486 MPASS(m->m_pkthdr.l5hlen > 0); 5487 MPASS(m->m_pkthdr.inner_l2hlen >= ETHER_HDR_LEN); 5488 MPASS(m->m_pkthdr.inner_l3hlen >= sizeof(struct ip)); 5489 5490 l2hlen = m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen + 5491 m->m_pkthdr.l4hlen + m->m_pkthdr.l5hlen + 5492 m->m_pkthdr.inner_l2hlen - ETHER_HDR_LEN; 5493 l3hlen = m->m_pkthdr.inner_l3hlen; 5494 } else { 5495 l2hlen = m->m_pkthdr.l2hlen - ETHER_HDR_LEN; 5496 l3hlen = m->m_pkthdr.l3hlen; 5497 } 5498 5499 ctrl = 0; 5500 if (!needs_l3_csum(m)) 5501 ctrl |= F_TXPKT_IPCSUM_DIS; 5502 5503 if (m->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_INNER_IP_TCP | 5504 CSUM_IP6_TCP | CSUM_INNER_IP6_TCP)) 5505 x = 0; /* TCP */ 5506 else if (m->m_pkthdr.csum_flags & (CSUM_IP_UDP | CSUM_INNER_IP_UDP | 5507 CSUM_IP6_UDP | CSUM_INNER_IP6_UDP)) 5508 x = 1; /* UDP */ 5509 else 5510 x = 2; 5511 5512 if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_IP_TCP | CSUM_IP_UDP | 5513 CSUM_INNER_IP | CSUM_INNER_IP_TCP | CSUM_INNER_IP_UDP)) 5514 y = 0; /* IPv4 */ 5515 else { 5516 MPASS(m->m_pkthdr.csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | 5517 CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_UDP)); 5518 y = 1; /* IPv6 */ 5519 } 5520 /* 5521 * needs_hwcsum returned true earlier so there must be some kind of 5522 * checksum to calculate. 5523 */ 5524 csum_type = csum_types[x][y]; 5525 MPASS(csum_type != 0); 5526 if (csum_type == TX_CSUM_IP) 5527 ctrl |= F_TXPKT_L4CSUM_DIS; 5528 ctrl |= V_TXPKT_CSUM_TYPE(csum_type) | V_TXPKT_IPHDR_LEN(l3hlen); 5529 if (chip_id(sc) <= CHELSIO_T5) 5530 ctrl |= V_TXPKT_ETHHDR_LEN(l2hlen); 5531 else 5532 ctrl |= V_T6_TXPKT_ETHHDR_LEN(l2hlen); 5533 5534 return (ctrl); 5535 } 5536 5537 static inline void * 5538 write_lso_cpl(void *cpl, struct mbuf *m0) 5539 { 5540 struct cpl_tx_pkt_lso_core *lso; 5541 uint32_t ctrl; 5542 5543 KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 && 5544 m0->m_pkthdr.l4hlen > 0, 5545 ("%s: mbuf %p needs TSO but missing header lengths", 5546 __func__, m0)); 5547 5548 ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | 5549 F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE | 5550 V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) | 5551 V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) | 5552 V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2); 5553 if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr)) 5554 ctrl |= F_LSO_IPV6; 5555 5556 lso = cpl; 5557 lso->lso_ctrl = htobe32(ctrl); 5558 lso->ipid_ofst = htobe16(0); 5559 lso->mss = htobe16(m0->m_pkthdr.tso_segsz); 5560 lso->seqno_offset = htobe32(0); 5561 lso->len = htobe32(m0->m_pkthdr.len); 5562 5563 return (lso + 1); 5564 } 5565 5566 static void * 5567 write_tnl_lso_cpl(void *cpl, struct mbuf *m0) 5568 { 5569 struct cpl_tx_tnl_lso *tnl_lso = cpl; 5570 uint32_t ctrl; 5571 5572 KASSERT(m0->m_pkthdr.inner_l2hlen > 0 && 5573 m0->m_pkthdr.inner_l3hlen > 0 && m0->m_pkthdr.inner_l4hlen > 0 && 5574 m0->m_pkthdr.inner_l5hlen > 0, 5575 ("%s: mbuf %p needs VXLAN_TSO but missing inner header lengths", 5576 __func__, m0)); 5577 KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 && 5578 m0->m_pkthdr.l4hlen > 0 && m0->m_pkthdr.l5hlen > 0, 5579 ("%s: mbuf %p needs VXLAN_TSO but missing outer header lengths", 5580 __func__, m0)); 5581 5582 /* Outer headers. */ 5583 ctrl = V_CPL_TX_TNL_LSO_OPCODE(CPL_TX_TNL_LSO) | 5584 F_CPL_TX_TNL_LSO_FIRST | F_CPL_TX_TNL_LSO_LAST | 5585 V_CPL_TX_TNL_LSO_ETHHDRLENOUT( 5586 (m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) | 5587 V_CPL_TX_TNL_LSO_IPHDRLENOUT(m0->m_pkthdr.l3hlen >> 2) | 5588 F_CPL_TX_TNL_LSO_IPLENSETOUT; 5589 if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr)) 5590 ctrl |= F_CPL_TX_TNL_LSO_IPV6OUT; 5591 else { 5592 ctrl |= F_CPL_TX_TNL_LSO_IPHDRCHKOUT | 5593 F_CPL_TX_TNL_LSO_IPIDINCOUT; 5594 } 5595 tnl_lso->op_to_IpIdSplitOut = htobe32(ctrl); 5596 tnl_lso->IpIdOffsetOut = 0; 5597 tnl_lso->UdpLenSetOut_to_TnlHdrLen = 5598 htobe16(F_CPL_TX_TNL_LSO_UDPCHKCLROUT | 5599 F_CPL_TX_TNL_LSO_UDPLENSETOUT | 5600 V_CPL_TX_TNL_LSO_TNLHDRLEN(m0->m_pkthdr.l2hlen + 5601 m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen + 5602 m0->m_pkthdr.l5hlen) | 5603 V_CPL_TX_TNL_LSO_TNLTYPE(TX_TNL_TYPE_VXLAN)); 5604 tnl_lso->ipsecen_to_rocev2 = 0; 5605 tnl_lso->roce_eth = 0; 5606 5607 /* Inner headers. */ 5608 ctrl = V_CPL_TX_TNL_LSO_ETHHDRLEN( 5609 (m0->m_pkthdr.inner_l2hlen - ETHER_HDR_LEN) >> 2) | 5610 V_CPL_TX_TNL_LSO_IPHDRLEN(m0->m_pkthdr.inner_l3hlen >> 2) | 5611 V_CPL_TX_TNL_LSO_TCPHDRLEN(m0->m_pkthdr.inner_l4hlen >> 2); 5612 if (m0->m_pkthdr.inner_l3hlen == sizeof(struct ip6_hdr)) 5613 ctrl |= F_CPL_TX_TNL_LSO_IPV6; 5614 tnl_lso->Flow_to_TcpHdrLen = htobe32(ctrl); 5615 tnl_lso->IpIdOffset = 0; 5616 tnl_lso->IpIdSplit_to_Mss = 5617 htobe16(V_CPL_TX_TNL_LSO_MSS(m0->m_pkthdr.tso_segsz)); 5618 tnl_lso->TCPSeqOffset = 0; 5619 tnl_lso->EthLenOffset_Size = 5620 htobe32(V_CPL_TX_TNL_LSO_SIZE(m0->m_pkthdr.len)); 5621 5622 return (tnl_lso + 1); 5623 } 5624 5625 #define VM_TX_L2HDR_LEN 16 /* ethmacdst to vlantci */ 5626 5627 /* 5628 * Write a VM txpkt WR for this packet to the hardware descriptors, update the 5629 * software descriptor, and advance the pidx. It is guaranteed that enough 5630 * descriptors are available. 5631 * 5632 * The return value is the # of hardware descriptors used. 5633 */ 5634 static u_int 5635 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq, struct mbuf *m0) 5636 { 5637 struct sge_eq *eq; 5638 struct fw_eth_tx_pkt_vm_wr *wr; 5639 struct tx_sdesc *txsd; 5640 struct cpl_tx_pkt_core *cpl; 5641 uint32_t ctrl; /* used in many unrelated places */ 5642 uint64_t ctrl1; 5643 int len16, ndesc, pktlen; 5644 caddr_t dst; 5645 5646 TXQ_LOCK_ASSERT_OWNED(txq); 5647 M_ASSERTPKTHDR(m0); 5648 5649 len16 = mbuf_len16(m0); 5650 pktlen = m0->m_pkthdr.len; 5651 ctrl = sizeof(struct cpl_tx_pkt_core); 5652 if (needs_tso(m0)) 5653 ctrl += sizeof(struct cpl_tx_pkt_lso_core); 5654 ndesc = tx_len16_to_desc(len16); 5655 5656 /* Firmware work request header */ 5657 eq = &txq->eq; 5658 wr = (void *)&eq->desc[eq->pidx]; 5659 wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) | 5660 V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl)); 5661 5662 ctrl = V_FW_WR_LEN16(len16); 5663 wr->equiq_to_len16 = htobe32(ctrl); 5664 wr->r3[0] = 0; 5665 wr->r3[1] = 0; 5666 5667 /* 5668 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci. 5669 * vlantci is ignored unless the ethtype is 0x8100, so it's 5670 * simpler to always copy it rather than making it 5671 * conditional. Also, it seems that we do not have to set 5672 * vlantci or fake the ethtype when doing VLAN tag insertion. 5673 */ 5674 m_copydata(m0, 0, VM_TX_L2HDR_LEN, wr->ethmacdst); 5675 5676 if (needs_tso(m0)) { 5677 cpl = write_lso_cpl(wr + 1, m0); 5678 txq->tso_wrs++; 5679 } else 5680 cpl = (void *)(wr + 1); 5681 5682 /* Checksum offload */ 5683 ctrl1 = csum_to_ctrl(sc, m0); 5684 if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) 5685 txq->txcsum++; /* some hardware assistance provided */ 5686 5687 /* VLAN tag insertion */ 5688 if (needs_vlan_insertion(m0)) { 5689 ctrl1 |= F_TXPKT_VLAN_VLD | 5690 V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag); 5691 txq->vlan_insertion++; 5692 } else if (sc->vlan_id) 5693 ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(sc->vlan_id); 5694 5695 /* CPL header */ 5696 cpl->ctrl0 = txq->cpl_ctrl0; 5697 cpl->pack = 0; 5698 cpl->len = htobe16(pktlen); 5699 cpl->ctrl1 = htobe64(ctrl1); 5700 5701 /* SGL */ 5702 dst = (void *)(cpl + 1); 5703 5704 /* 5705 * A packet using TSO will use up an entire descriptor for the 5706 * firmware work request header, LSO CPL, and TX_PKT_XT CPL. 5707 * If this descriptor is the last descriptor in the ring, wrap 5708 * around to the front of the ring explicitly for the start of 5709 * the sgl. 5710 */ 5711 if (dst == (void *)&eq->desc[eq->sidx]) { 5712 dst = (void *)&eq->desc[0]; 5713 write_gl_to_txd(txq, m0, &dst, 0); 5714 } else 5715 write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx); 5716 txq->sgl_wrs++; 5717 txq->txpkt_wrs++; 5718 5719 txsd = &txq->sdesc[eq->pidx]; 5720 txsd->m = m0; 5721 txsd->desc_used = ndesc; 5722 5723 return (ndesc); 5724 } 5725 5726 /* 5727 * Write a raw WR to the hardware descriptors, update the software 5728 * descriptor, and advance the pidx. It is guaranteed that enough 5729 * descriptors are available. 5730 * 5731 * The return value is the # of hardware descriptors used. 5732 */ 5733 static u_int 5734 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available) 5735 { 5736 struct sge_eq *eq = &txq->eq; 5737 struct tx_sdesc *txsd; 5738 struct mbuf *m; 5739 caddr_t dst; 5740 int len16, ndesc; 5741 5742 len16 = mbuf_len16(m0); 5743 ndesc = tx_len16_to_desc(len16); 5744 MPASS(ndesc <= available); 5745 5746 dst = wr; 5747 for (m = m0; m != NULL; m = m->m_next) 5748 copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len); 5749 5750 txq->raw_wrs++; 5751 5752 txsd = &txq->sdesc[eq->pidx]; 5753 txsd->m = m0; 5754 txsd->desc_used = ndesc; 5755 5756 return (ndesc); 5757 } 5758 5759 /* 5760 * Write a txpkt WR for this packet to the hardware descriptors, update the 5761 * software descriptor, and advance the pidx. It is guaranteed that enough 5762 * descriptors are available. 5763 * 5764 * The return value is the # of hardware descriptors used. 5765 */ 5766 static u_int 5767 write_txpkt_wr(struct adapter *sc, struct sge_txq *txq, struct mbuf *m0, 5768 u_int available) 5769 { 5770 struct sge_eq *eq; 5771 struct fw_eth_tx_pkt_wr *wr; 5772 struct tx_sdesc *txsd; 5773 struct cpl_tx_pkt_core *cpl; 5774 uint32_t ctrl; /* used in many unrelated places */ 5775 uint64_t ctrl1; 5776 int len16, ndesc, pktlen, nsegs; 5777 caddr_t dst; 5778 5779 TXQ_LOCK_ASSERT_OWNED(txq); 5780 M_ASSERTPKTHDR(m0); 5781 5782 len16 = mbuf_len16(m0); 5783 nsegs = mbuf_nsegs(m0); 5784 pktlen = m0->m_pkthdr.len; 5785 ctrl = sizeof(struct cpl_tx_pkt_core); 5786 if (needs_tso(m0)) { 5787 if (needs_vxlan_tso(m0)) 5788 ctrl += sizeof(struct cpl_tx_tnl_lso); 5789 else 5790 ctrl += sizeof(struct cpl_tx_pkt_lso_core); 5791 } else if (!(mbuf_cflags(m0) & MC_NOMAP) && pktlen <= imm_payload(2) && 5792 available >= 2) { 5793 /* Immediate data. Recalculate len16 and set nsegs to 0. */ 5794 ctrl += pktlen; 5795 len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) + 5796 sizeof(struct cpl_tx_pkt_core) + pktlen, 16); 5797 nsegs = 0; 5798 } 5799 ndesc = tx_len16_to_desc(len16); 5800 MPASS(ndesc <= available); 5801 5802 /* Firmware work request header */ 5803 eq = &txq->eq; 5804 wr = (void *)&eq->desc[eq->pidx]; 5805 wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) | 5806 V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl)); 5807 5808 ctrl = V_FW_WR_LEN16(len16); 5809 wr->equiq_to_len16 = htobe32(ctrl); 5810 wr->r3 = 0; 5811 5812 if (needs_tso(m0)) { 5813 if (needs_vxlan_tso(m0)) { 5814 cpl = write_tnl_lso_cpl(wr + 1, m0); 5815 txq->vxlan_tso_wrs++; 5816 } else { 5817 cpl = write_lso_cpl(wr + 1, m0); 5818 txq->tso_wrs++; 5819 } 5820 } else 5821 cpl = (void *)(wr + 1); 5822 5823 /* Checksum offload */ 5824 ctrl1 = csum_to_ctrl(sc, m0); 5825 if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) { 5826 /* some hardware assistance provided */ 5827 if (needs_vxlan_csum(m0)) 5828 txq->vxlan_txcsum++; 5829 else 5830 txq->txcsum++; 5831 } 5832 5833 /* VLAN tag insertion */ 5834 if (needs_vlan_insertion(m0)) { 5835 ctrl1 |= F_TXPKT_VLAN_VLD | 5836 V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag); 5837 txq->vlan_insertion++; 5838 } 5839 5840 /* CPL header */ 5841 cpl->ctrl0 = txq->cpl_ctrl0; 5842 cpl->pack = 0; 5843 cpl->len = htobe16(pktlen); 5844 cpl->ctrl1 = htobe64(ctrl1); 5845 5846 /* SGL */ 5847 dst = (void *)(cpl + 1); 5848 if (__predict_false((uintptr_t)dst == (uintptr_t)&eq->desc[eq->sidx])) 5849 dst = (caddr_t)&eq->desc[0]; 5850 if (nsegs > 0) { 5851 5852 write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx); 5853 txq->sgl_wrs++; 5854 } else { 5855 struct mbuf *m; 5856 5857 for (m = m0; m != NULL; m = m->m_next) { 5858 copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len); 5859 #ifdef INVARIANTS 5860 pktlen -= m->m_len; 5861 #endif 5862 } 5863 #ifdef INVARIANTS 5864 KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen)); 5865 #endif 5866 txq->imm_wrs++; 5867 } 5868 5869 txq->txpkt_wrs++; 5870 5871 txsd = &txq->sdesc[eq->pidx]; 5872 txsd->m = m0; 5873 txsd->desc_used = ndesc; 5874 5875 return (ndesc); 5876 } 5877 5878 static inline bool 5879 cmp_l2hdr(struct txpkts *txp, struct mbuf *m) 5880 { 5881 int len; 5882 5883 MPASS(txp->npkt > 0); 5884 MPASS(m->m_len >= VM_TX_L2HDR_LEN); 5885 5886 if (txp->ethtype == be16toh(ETHERTYPE_VLAN)) 5887 len = VM_TX_L2HDR_LEN; 5888 else 5889 len = sizeof(struct ether_header); 5890 5891 return (memcmp(m->m_data, &txp->ethmacdst[0], len) != 0); 5892 } 5893 5894 static inline void 5895 save_l2hdr(struct txpkts *txp, struct mbuf *m) 5896 { 5897 MPASS(m->m_len >= VM_TX_L2HDR_LEN); 5898 5899 memcpy(&txp->ethmacdst[0], mtod(m, const void *), VM_TX_L2HDR_LEN); 5900 } 5901 5902 static int 5903 add_to_txpkts_vf(struct adapter *sc, struct sge_txq *txq, struct mbuf *m, 5904 int avail, bool *send) 5905 { 5906 struct txpkts *txp = &txq->txp; 5907 5908 /* Cannot have TSO and coalesce at the same time. */ 5909 if (cannot_use_txpkts(m)) { 5910 cannot_coalesce: 5911 *send = txp->npkt > 0; 5912 return (EINVAL); 5913 } 5914 5915 /* VF allows coalescing of type 1 (1 GL) only */ 5916 if (mbuf_nsegs(m) > 1) 5917 goto cannot_coalesce; 5918 5919 *send = false; 5920 if (txp->npkt > 0) { 5921 MPASS(tx_len16_to_desc(txp->len16) <= avail); 5922 MPASS(txp->npkt < txp->max_npkt); 5923 MPASS(txp->wr_type == 1); /* VF supports type 1 only */ 5924 5925 if (tx_len16_to_desc(txp->len16 + txpkts1_len16()) > avail) { 5926 retry_after_send: 5927 *send = true; 5928 return (EAGAIN); 5929 } 5930 if (m->m_pkthdr.len + txp->plen > 65535) 5931 goto retry_after_send; 5932 if (cmp_l2hdr(txp, m)) 5933 goto retry_after_send; 5934 5935 txp->len16 += txpkts1_len16(); 5936 txp->plen += m->m_pkthdr.len; 5937 txp->mb[txp->npkt++] = m; 5938 if (txp->npkt == txp->max_npkt) 5939 *send = true; 5940 } else { 5941 txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_vm_wr), 16) + 5942 txpkts1_len16(); 5943 if (tx_len16_to_desc(txp->len16) > avail) 5944 goto cannot_coalesce; 5945 txp->npkt = 1; 5946 txp->wr_type = 1; 5947 txp->plen = m->m_pkthdr.len; 5948 txp->mb[0] = m; 5949 save_l2hdr(txp, m); 5950 } 5951 return (0); 5952 } 5953 5954 static int 5955 add_to_txpkts_pf(struct adapter *sc, struct sge_txq *txq, struct mbuf *m, 5956 int avail, bool *send) 5957 { 5958 struct txpkts *txp = &txq->txp; 5959 int nsegs; 5960 5961 MPASS(!(sc->flags & IS_VF)); 5962 5963 /* Cannot have TSO and coalesce at the same time. */ 5964 if (cannot_use_txpkts(m)) { 5965 cannot_coalesce: 5966 *send = txp->npkt > 0; 5967 return (EINVAL); 5968 } 5969 5970 *send = false; 5971 nsegs = mbuf_nsegs(m); 5972 if (txp->npkt == 0) { 5973 if (m->m_pkthdr.len > 65535) 5974 goto cannot_coalesce; 5975 if (nsegs > 1) { 5976 txp->wr_type = 0; 5977 txp->len16 = 5978 howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + 5979 txpkts0_len16(nsegs); 5980 } else { 5981 txp->wr_type = 1; 5982 txp->len16 = 5983 howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + 5984 txpkts1_len16(); 5985 } 5986 if (tx_len16_to_desc(txp->len16) > avail) 5987 goto cannot_coalesce; 5988 txp->npkt = 1; 5989 txp->plen = m->m_pkthdr.len; 5990 txp->mb[0] = m; 5991 } else { 5992 MPASS(tx_len16_to_desc(txp->len16) <= avail); 5993 MPASS(txp->npkt < txp->max_npkt); 5994 5995 if (m->m_pkthdr.len + txp->plen > 65535) { 5996 retry_after_send: 5997 *send = true; 5998 return (EAGAIN); 5999 } 6000 6001 MPASS(txp->wr_type == 0 || txp->wr_type == 1); 6002 if (txp->wr_type == 0) { 6003 if (tx_len16_to_desc(txp->len16 + 6004 txpkts0_len16(nsegs)) > min(avail, SGE_MAX_WR_NDESC)) 6005 goto retry_after_send; 6006 txp->len16 += txpkts0_len16(nsegs); 6007 } else { 6008 if (nsegs != 1) 6009 goto retry_after_send; 6010 if (tx_len16_to_desc(txp->len16 + txpkts1_len16()) > 6011 avail) 6012 goto retry_after_send; 6013 txp->len16 += txpkts1_len16(); 6014 } 6015 6016 txp->plen += m->m_pkthdr.len; 6017 txp->mb[txp->npkt++] = m; 6018 if (txp->npkt == txp->max_npkt) 6019 *send = true; 6020 } 6021 return (0); 6022 } 6023 6024 /* 6025 * Write a txpkts WR for the packets in txp to the hardware descriptors, update 6026 * the software descriptor, and advance the pidx. It is guaranteed that enough 6027 * descriptors are available. 6028 * 6029 * The return value is the # of hardware descriptors used. 6030 */ 6031 static u_int 6032 write_txpkts_wr(struct adapter *sc, struct sge_txq *txq) 6033 { 6034 const struct txpkts *txp = &txq->txp; 6035 struct sge_eq *eq = &txq->eq; 6036 struct fw_eth_tx_pkts_wr *wr; 6037 struct tx_sdesc *txsd; 6038 struct cpl_tx_pkt_core *cpl; 6039 uint64_t ctrl1; 6040 int ndesc, i, checkwrap; 6041 struct mbuf *m, *last; 6042 void *flitp; 6043 6044 TXQ_LOCK_ASSERT_OWNED(txq); 6045 MPASS(txp->npkt > 0); 6046 MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16)); 6047 6048 wr = (void *)&eq->desc[eq->pidx]; 6049 wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)); 6050 wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(txp->len16)); 6051 wr->plen = htobe16(txp->plen); 6052 wr->npkt = txp->npkt; 6053 wr->r3 = 0; 6054 wr->type = txp->wr_type; 6055 flitp = wr + 1; 6056 6057 /* 6058 * At this point we are 16B into a hardware descriptor. If checkwrap is 6059 * set then we know the WR is going to wrap around somewhere. We'll 6060 * check for that at appropriate points. 6061 */ 6062 ndesc = tx_len16_to_desc(txp->len16); 6063 last = NULL; 6064 checkwrap = eq->sidx - ndesc < eq->pidx; 6065 for (i = 0; i < txp->npkt; i++) { 6066 m = txp->mb[i]; 6067 if (txp->wr_type == 0) { 6068 struct ulp_txpkt *ulpmc; 6069 struct ulptx_idata *ulpsc; 6070 6071 /* ULP master command */ 6072 ulpmc = flitp; 6073 ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) | 6074 V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid)); 6075 ulpmc->len = htobe32(txpkts0_len16(mbuf_nsegs(m))); 6076 6077 /* ULP subcommand */ 6078 ulpsc = (void *)(ulpmc + 1); 6079 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) | 6080 F_ULP_TX_SC_MORE); 6081 ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core)); 6082 6083 cpl = (void *)(ulpsc + 1); 6084 if (checkwrap && 6085 (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx]) 6086 cpl = (void *)&eq->desc[0]; 6087 } else { 6088 cpl = flitp; 6089 } 6090 6091 /* Checksum offload */ 6092 ctrl1 = csum_to_ctrl(sc, m); 6093 if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) { 6094 /* some hardware assistance provided */ 6095 if (needs_vxlan_csum(m)) 6096 txq->vxlan_txcsum++; 6097 else 6098 txq->txcsum++; 6099 } 6100 6101 /* VLAN tag insertion */ 6102 if (needs_vlan_insertion(m)) { 6103 ctrl1 |= F_TXPKT_VLAN_VLD | 6104 V_TXPKT_VLAN(m->m_pkthdr.ether_vtag); 6105 txq->vlan_insertion++; 6106 } 6107 6108 /* CPL header */ 6109 cpl->ctrl0 = txq->cpl_ctrl0; 6110 cpl->pack = 0; 6111 cpl->len = htobe16(m->m_pkthdr.len); 6112 cpl->ctrl1 = htobe64(ctrl1); 6113 6114 flitp = cpl + 1; 6115 if (checkwrap && 6116 (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx]) 6117 flitp = (void *)&eq->desc[0]; 6118 6119 write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap); 6120 6121 if (last != NULL) 6122 last->m_nextpkt = m; 6123 last = m; 6124 } 6125 6126 txq->sgl_wrs++; 6127 if (txp->wr_type == 0) { 6128 txq->txpkts0_pkts += txp->npkt; 6129 txq->txpkts0_wrs++; 6130 } else { 6131 txq->txpkts1_pkts += txp->npkt; 6132 txq->txpkts1_wrs++; 6133 } 6134 6135 txsd = &txq->sdesc[eq->pidx]; 6136 txsd->m = txp->mb[0]; 6137 txsd->desc_used = ndesc; 6138 6139 return (ndesc); 6140 } 6141 6142 static u_int 6143 write_txpkts_vm_wr(struct adapter *sc, struct sge_txq *txq) 6144 { 6145 const struct txpkts *txp = &txq->txp; 6146 struct sge_eq *eq = &txq->eq; 6147 struct fw_eth_tx_pkts_vm_wr *wr; 6148 struct tx_sdesc *txsd; 6149 struct cpl_tx_pkt_core *cpl; 6150 uint64_t ctrl1; 6151 int ndesc, i; 6152 struct mbuf *m, *last; 6153 void *flitp; 6154 6155 TXQ_LOCK_ASSERT_OWNED(txq); 6156 MPASS(txp->npkt > 0); 6157 MPASS(txp->wr_type == 1); /* VF supports type 1 only */ 6158 MPASS(txp->mb[0] != NULL); 6159 MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16)); 6160 6161 wr = (void *)&eq->desc[eq->pidx]; 6162 wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_VM_WR)); 6163 wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(txp->len16)); 6164 wr->r3 = 0; 6165 wr->plen = htobe16(txp->plen); 6166 wr->npkt = txp->npkt; 6167 wr->r4 = 0; 6168 memcpy(&wr->ethmacdst[0], &txp->ethmacdst[0], 16); 6169 flitp = wr + 1; 6170 6171 /* 6172 * At this point we are 32B into a hardware descriptor. Each mbuf in 6173 * the WR will take 32B so we check for the end of the descriptor ring 6174 * before writing odd mbufs (mb[1], 3, 5, ..) 6175 */ 6176 ndesc = tx_len16_to_desc(txp->len16); 6177 last = NULL; 6178 for (i = 0; i < txp->npkt; i++) { 6179 m = txp->mb[i]; 6180 if (i & 1 && (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx]) 6181 flitp = &eq->desc[0]; 6182 cpl = flitp; 6183 6184 /* Checksum offload */ 6185 ctrl1 = csum_to_ctrl(sc, m); 6186 if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) 6187 txq->txcsum++; /* some hardware assistance provided */ 6188 6189 /* VLAN tag insertion */ 6190 if (needs_vlan_insertion(m)) { 6191 ctrl1 |= F_TXPKT_VLAN_VLD | 6192 V_TXPKT_VLAN(m->m_pkthdr.ether_vtag); 6193 txq->vlan_insertion++; 6194 } else if (sc->vlan_id) 6195 ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(sc->vlan_id); 6196 6197 /* CPL header */ 6198 cpl->ctrl0 = txq->cpl_ctrl0; 6199 cpl->pack = 0; 6200 cpl->len = htobe16(m->m_pkthdr.len); 6201 cpl->ctrl1 = htobe64(ctrl1); 6202 6203 flitp = cpl + 1; 6204 MPASS(mbuf_nsegs(m) == 1); 6205 write_gl_to_txd(txq, m, (caddr_t *)(&flitp), 0); 6206 6207 if (last != NULL) 6208 last->m_nextpkt = m; 6209 last = m; 6210 } 6211 6212 txq->sgl_wrs++; 6213 txq->txpkts1_pkts += txp->npkt; 6214 txq->txpkts1_wrs++; 6215 6216 txsd = &txq->sdesc[eq->pidx]; 6217 txsd->m = txp->mb[0]; 6218 txsd->desc_used = ndesc; 6219 6220 return (ndesc); 6221 } 6222 6223 /* 6224 * If the SGL ends on an address that is not 16 byte aligned, this function will 6225 * add a 0 filled flit at the end. 6226 */ 6227 static void 6228 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap) 6229 { 6230 struct sge_eq *eq = &txq->eq; 6231 struct sglist *gl = txq->gl; 6232 struct sglist_seg *seg; 6233 __be64 *flitp, *wrap; 6234 struct ulptx_sgl *usgl; 6235 int i, nflits, nsegs; 6236 6237 KASSERT(((uintptr_t)(*to) & 0xf) == 0, 6238 ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to)); 6239 MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]); 6240 MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]); 6241 6242 get_pkt_gl(m, gl); 6243 nsegs = gl->sg_nseg; 6244 MPASS(nsegs > 0); 6245 6246 nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2; 6247 flitp = (__be64 *)(*to); 6248 wrap = (__be64 *)(&eq->desc[eq->sidx]); 6249 seg = &gl->sg_segs[0]; 6250 usgl = (void *)flitp; 6251 6252 /* 6253 * We start at a 16 byte boundary somewhere inside the tx descriptor 6254 * ring, so we're at least 16 bytes away from the status page. There is 6255 * no chance of a wrap around in the middle of usgl (which is 16 bytes). 6256 */ 6257 6258 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | 6259 V_ULPTX_NSGE(nsegs)); 6260 usgl->len0 = htobe32(seg->ss_len); 6261 usgl->addr0 = htobe64(seg->ss_paddr); 6262 seg++; 6263 6264 if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) { 6265 6266 /* Won't wrap around at all */ 6267 6268 for (i = 0; i < nsegs - 1; i++, seg++) { 6269 usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len); 6270 usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr); 6271 } 6272 if (i & 1) 6273 usgl->sge[i / 2].len[1] = htobe32(0); 6274 flitp += nflits; 6275 } else { 6276 6277 /* Will wrap somewhere in the rest of the SGL */ 6278 6279 /* 2 flits already written, write the rest flit by flit */ 6280 flitp = (void *)(usgl + 1); 6281 for (i = 0; i < nflits - 2; i++) { 6282 if (flitp == wrap) 6283 flitp = (void *)eq->desc; 6284 *flitp++ = get_flit(seg, nsegs - 1, i); 6285 } 6286 } 6287 6288 if (nflits & 1) { 6289 MPASS(((uintptr_t)flitp) & 0xf); 6290 *flitp++ = 0; 6291 } 6292 6293 MPASS((((uintptr_t)flitp) & 0xf) == 0); 6294 if (__predict_false(flitp == wrap)) 6295 *to = (void *)eq->desc; 6296 else 6297 *to = (void *)flitp; 6298 } 6299 6300 static inline void 6301 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len) 6302 { 6303 6304 MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]); 6305 MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]); 6306 6307 if (__predict_true((uintptr_t)(*to) + len <= 6308 (uintptr_t)&eq->desc[eq->sidx])) { 6309 bcopy(from, *to, len); 6310 (*to) += len; 6311 } else { 6312 int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to); 6313 6314 bcopy(from, *to, portion); 6315 from += portion; 6316 portion = len - portion; /* remaining */ 6317 bcopy(from, (void *)eq->desc, portion); 6318 (*to) = (caddr_t)eq->desc + portion; 6319 } 6320 } 6321 6322 static inline void 6323 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n) 6324 { 6325 u_int db; 6326 6327 MPASS(n > 0); 6328 6329 db = eq->doorbells; 6330 if (n > 1) 6331 clrbit(&db, DOORBELL_WCWR); 6332 wmb(); 6333 6334 switch (ffs(db) - 1) { 6335 case DOORBELL_UDB: 6336 *eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n)); 6337 break; 6338 6339 case DOORBELL_WCWR: { 6340 volatile uint64_t *dst, *src; 6341 int i; 6342 6343 /* 6344 * Queues whose 128B doorbell segment fits in the page do not 6345 * use relative qid (udb_qid is always 0). Only queues with 6346 * doorbell segments can do WCWR. 6347 */ 6348 KASSERT(eq->udb_qid == 0 && n == 1, 6349 ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p", 6350 __func__, eq->doorbells, n, eq->dbidx, eq)); 6351 6352 dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET - 6353 UDBS_DB_OFFSET); 6354 i = eq->dbidx; 6355 src = (void *)&eq->desc[i]; 6356 while (src != (void *)&eq->desc[i + 1]) 6357 *dst++ = *src++; 6358 wmb(); 6359 break; 6360 } 6361 6362 case DOORBELL_UDBWC: 6363 *eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n)); 6364 wmb(); 6365 break; 6366 6367 case DOORBELL_KDB: 6368 t4_write_reg(sc, sc->sge_kdoorbell_reg, 6369 V_QID(eq->cntxt_id) | V_PIDX(n)); 6370 break; 6371 } 6372 6373 IDXINCR(eq->dbidx, n, eq->sidx); 6374 } 6375 6376 static inline u_int 6377 reclaimable_tx_desc(struct sge_eq *eq) 6378 { 6379 uint16_t hw_cidx; 6380 6381 hw_cidx = read_hw_cidx(eq); 6382 return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx)); 6383 } 6384 6385 static inline u_int 6386 total_available_tx_desc(struct sge_eq *eq) 6387 { 6388 uint16_t hw_cidx, pidx; 6389 6390 hw_cidx = read_hw_cidx(eq); 6391 pidx = eq->pidx; 6392 6393 if (pidx == hw_cidx) 6394 return (eq->sidx - 1); 6395 else 6396 return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1); 6397 } 6398 6399 static inline uint16_t 6400 read_hw_cidx(struct sge_eq *eq) 6401 { 6402 struct sge_qstat *spg = (void *)&eq->desc[eq->sidx]; 6403 uint16_t cidx = spg->cidx; /* stable snapshot */ 6404 6405 return (be16toh(cidx)); 6406 } 6407 6408 /* 6409 * Reclaim 'n' descriptors approximately. 6410 */ 6411 static u_int 6412 reclaim_tx_descs(struct sge_txq *txq, u_int n) 6413 { 6414 struct tx_sdesc *txsd; 6415 struct sge_eq *eq = &txq->eq; 6416 u_int can_reclaim, reclaimed; 6417 6418 TXQ_LOCK_ASSERT_OWNED(txq); 6419 MPASS(n > 0); 6420 6421 reclaimed = 0; 6422 can_reclaim = reclaimable_tx_desc(eq); 6423 while (can_reclaim && reclaimed < n) { 6424 int ndesc; 6425 struct mbuf *m, *nextpkt; 6426 6427 txsd = &txq->sdesc[eq->cidx]; 6428 ndesc = txsd->desc_used; 6429 6430 /* Firmware doesn't return "partial" credits. */ 6431 KASSERT(can_reclaim >= ndesc, 6432 ("%s: unexpected number of credits: %d, %d", 6433 __func__, can_reclaim, ndesc)); 6434 KASSERT(ndesc != 0, 6435 ("%s: descriptor with no credits: cidx %d", 6436 __func__, eq->cidx)); 6437 6438 for (m = txsd->m; m != NULL; m = nextpkt) { 6439 nextpkt = m->m_nextpkt; 6440 m->m_nextpkt = NULL; 6441 m_freem(m); 6442 } 6443 reclaimed += ndesc; 6444 can_reclaim -= ndesc; 6445 IDXINCR(eq->cidx, ndesc, eq->sidx); 6446 } 6447 6448 return (reclaimed); 6449 } 6450 6451 static void 6452 tx_reclaim(void *arg, int n) 6453 { 6454 struct sge_txq *txq = arg; 6455 struct sge_eq *eq = &txq->eq; 6456 6457 do { 6458 if (TXQ_TRYLOCK(txq) == 0) 6459 break; 6460 n = reclaim_tx_descs(txq, 32); 6461 if (eq->cidx == eq->pidx) 6462 eq->equeqidx = eq->pidx; 6463 TXQ_UNLOCK(txq); 6464 } while (n > 0); 6465 } 6466 6467 static __be64 6468 get_flit(struct sglist_seg *segs, int nsegs, int idx) 6469 { 6470 int i = (idx / 3) * 2; 6471 6472 switch (idx % 3) { 6473 case 0: { 6474 uint64_t rc; 6475 6476 rc = (uint64_t)segs[i].ss_len << 32; 6477 if (i + 1 < nsegs) 6478 rc |= (uint64_t)(segs[i + 1].ss_len); 6479 6480 return (htobe64(rc)); 6481 } 6482 case 1: 6483 return (htobe64(segs[i].ss_paddr)); 6484 case 2: 6485 return (htobe64(segs[i + 1].ss_paddr)); 6486 } 6487 6488 return (0); 6489 } 6490 6491 static int 6492 find_refill_source(struct adapter *sc, int maxp, bool packing) 6493 { 6494 int i, zidx = -1; 6495 struct rx_buf_info *rxb = &sc->sge.rx_buf_info[0]; 6496 6497 if (packing) { 6498 for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) { 6499 if (rxb->hwidx2 == -1) 6500 continue; 6501 if (rxb->size1 < PAGE_SIZE && 6502 rxb->size1 < largest_rx_cluster) 6503 continue; 6504 if (rxb->size1 > largest_rx_cluster) 6505 break; 6506 MPASS(rxb->size1 - rxb->size2 >= CL_METADATA_SIZE); 6507 if (rxb->size2 >= maxp) 6508 return (i); 6509 zidx = i; 6510 } 6511 } else { 6512 for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) { 6513 if (rxb->hwidx1 == -1) 6514 continue; 6515 if (rxb->size1 > largest_rx_cluster) 6516 break; 6517 if (rxb->size1 >= maxp) 6518 return (i); 6519 zidx = i; 6520 } 6521 } 6522 6523 return (zidx); 6524 } 6525 6526 static void 6527 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl) 6528 { 6529 mtx_lock(&sc->sfl_lock); 6530 FL_LOCK(fl); 6531 if ((fl->flags & FL_DOOMED) == 0) { 6532 fl->flags |= FL_STARVING; 6533 TAILQ_INSERT_TAIL(&sc->sfl, fl, link); 6534 callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc); 6535 } 6536 FL_UNLOCK(fl); 6537 mtx_unlock(&sc->sfl_lock); 6538 } 6539 6540 static void 6541 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq) 6542 { 6543 struct sge_wrq *wrq = (void *)eq; 6544 6545 atomic_readandclear_int(&eq->equiq); 6546 taskqueue_enqueue(sc->tq[eq->port_id], &wrq->wrq_tx_task); 6547 } 6548 6549 static void 6550 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq) 6551 { 6552 struct sge_txq *txq = (void *)eq; 6553 6554 MPASS(eq->type == EQ_ETH); 6555 6556 atomic_readandclear_int(&eq->equiq); 6557 if (mp_ring_is_idle(txq->r)) 6558 taskqueue_enqueue(sc->tq[eq->port_id], &txq->tx_reclaim_task); 6559 else 6560 mp_ring_check_drainage(txq->r, 64); 6561 } 6562 6563 static int 6564 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss, 6565 struct mbuf *m) 6566 { 6567 const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1); 6568 unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid)); 6569 struct adapter *sc = iq->adapter; 6570 struct sge *s = &sc->sge; 6571 struct sge_eq *eq; 6572 static void (*h[])(struct adapter *, struct sge_eq *) = {NULL, 6573 &handle_wrq_egr_update, &handle_eth_egr_update, 6574 &handle_wrq_egr_update}; 6575 6576 KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__, 6577 rss->opcode)); 6578 6579 eq = s->eqmap[qid - s->eq_start - s->eq_base]; 6580 (*h[eq->type])(sc, eq); 6581 6582 return (0); 6583 } 6584 6585 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */ 6586 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \ 6587 offsetof(struct cpl_fw6_msg, data)); 6588 6589 static int 6590 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 6591 { 6592 struct adapter *sc = iq->adapter; 6593 const struct cpl_fw6_msg *cpl = (const void *)(rss + 1); 6594 6595 KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__, 6596 rss->opcode)); 6597 6598 if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) { 6599 const struct rss_header *rss2; 6600 6601 rss2 = (const struct rss_header *)&cpl->data[0]; 6602 return (t4_cpl_handler[rss2->opcode](iq, rss2, m)); 6603 } 6604 6605 return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0])); 6606 } 6607 6608 /** 6609 * t4_handle_wrerr_rpl - process a FW work request error message 6610 * @adap: the adapter 6611 * @rpl: start of the FW message 6612 */ 6613 static int 6614 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl) 6615 { 6616 u8 opcode = *(const u8 *)rpl; 6617 const struct fw_error_cmd *e = (const void *)rpl; 6618 unsigned int i; 6619 6620 if (opcode != FW_ERROR_CMD) { 6621 log(LOG_ERR, 6622 "%s: Received WRERR_RPL message with opcode %#x\n", 6623 device_get_nameunit(adap->dev), opcode); 6624 return (EINVAL); 6625 } 6626 log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev), 6627 G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" : 6628 "non-fatal"); 6629 switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) { 6630 case FW_ERROR_TYPE_EXCEPTION: 6631 log(LOG_ERR, "exception info:\n"); 6632 for (i = 0; i < nitems(e->u.exception.info); i++) 6633 log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ", 6634 be32toh(e->u.exception.info[i])); 6635 log(LOG_ERR, "\n"); 6636 break; 6637 case FW_ERROR_TYPE_HWMODULE: 6638 log(LOG_ERR, "HW module regaddr %08x regval %08x\n", 6639 be32toh(e->u.hwmodule.regaddr), 6640 be32toh(e->u.hwmodule.regval)); 6641 break; 6642 case FW_ERROR_TYPE_WR: 6643 log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n", 6644 be16toh(e->u.wr.cidx), 6645 G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)), 6646 G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)), 6647 be32toh(e->u.wr.eqid)); 6648 for (i = 0; i < nitems(e->u.wr.wrhdr); i++) 6649 log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ", 6650 e->u.wr.wrhdr[i]); 6651 log(LOG_ERR, "\n"); 6652 break; 6653 case FW_ERROR_TYPE_ACL: 6654 log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s", 6655 be16toh(e->u.acl.cidx), 6656 G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)), 6657 G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)), 6658 be32toh(e->u.acl.eqid), 6659 G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" : 6660 "MAC"); 6661 for (i = 0; i < nitems(e->u.acl.val); i++) 6662 log(LOG_ERR, " %02x", e->u.acl.val[i]); 6663 log(LOG_ERR, "\n"); 6664 break; 6665 default: 6666 log(LOG_ERR, "type %#x\n", 6667 G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))); 6668 return (EINVAL); 6669 } 6670 return (0); 6671 } 6672 6673 static inline bool 6674 bufidx_used(struct adapter *sc, int idx) 6675 { 6676 struct rx_buf_info *rxb = &sc->sge.rx_buf_info[0]; 6677 int i; 6678 6679 for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) { 6680 if (rxb->size1 > largest_rx_cluster) 6681 continue; 6682 if (rxb->hwidx1 == idx || rxb->hwidx2 == idx) 6683 return (true); 6684 } 6685 6686 return (false); 6687 } 6688 6689 static int 6690 sysctl_bufsizes(SYSCTL_HANDLER_ARGS) 6691 { 6692 struct adapter *sc = arg1; 6693 struct sge_params *sp = &sc->params.sge; 6694 int i, rc; 6695 struct sbuf sb; 6696 char c; 6697 6698 sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND); 6699 for (i = 0; i < SGE_FLBUF_SIZES; i++) { 6700 if (bufidx_used(sc, i)) 6701 c = '*'; 6702 else 6703 c = '\0'; 6704 6705 sbuf_printf(&sb, "%u%c ", sp->sge_fl_buffer_size[i], c); 6706 } 6707 sbuf_trim(&sb); 6708 sbuf_finish(&sb); 6709 rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req); 6710 sbuf_delete(&sb); 6711 return (rc); 6712 } 6713 6714 #ifdef RATELIMIT 6715 #if defined(INET) || defined(INET6) 6716 /* 6717 * len16 for a txpkt WR with a GL. Includes the firmware work request header. 6718 */ 6719 static inline u_int 6720 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso) 6721 { 6722 u_int n; 6723 6724 MPASS(immhdrs > 0); 6725 6726 n = roundup2(sizeof(struct fw_eth_tx_eo_wr) + 6727 sizeof(struct cpl_tx_pkt_core) + immhdrs, 16); 6728 if (__predict_false(nsegs == 0)) 6729 goto done; 6730 6731 nsegs--; /* first segment is part of ulptx_sgl */ 6732 n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1)); 6733 if (tso) 6734 n += sizeof(struct cpl_tx_pkt_lso_core); 6735 6736 done: 6737 return (howmany(n, 16)); 6738 } 6739 #endif 6740 6741 #define ETID_FLOWC_NPARAMS 6 6742 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \ 6743 ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16)) 6744 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16)) 6745 6746 #if defined(INET) || defined(INET6) 6747 static int 6748 send_etid_flowc_wr(struct cxgbe_rate_tag *cst, struct port_info *pi, 6749 struct vi_info *vi) 6750 { 6751 struct wrq_cookie cookie; 6752 u_int pfvf = pi->adapter->pf << S_FW_VIID_PFN; 6753 struct fw_flowc_wr *flowc; 6754 6755 mtx_assert(&cst->lock, MA_OWNED); 6756 MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) == 6757 EO_FLOWC_PENDING); 6758 6759 flowc = start_wrq_wr(&cst->eo_txq->wrq, ETID_FLOWC_LEN16, &cookie); 6760 if (__predict_false(flowc == NULL)) 6761 return (ENOMEM); 6762 6763 bzero(flowc, ETID_FLOWC_LEN); 6764 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | 6765 V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0)); 6766 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) | 6767 V_FW_WR_FLOWID(cst->etid)); 6768 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN; 6769 flowc->mnemval[0].val = htobe32(pfvf); 6770 /* Firmware expects hw port and will translate to channel itself. */ 6771 flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH; 6772 flowc->mnemval[1].val = htobe32(pi->hw_port); 6773 flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT; 6774 flowc->mnemval[2].val = htobe32(pi->hw_port); 6775 flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID; 6776 flowc->mnemval[3].val = htobe32(cst->iqid); 6777 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE; 6778 flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED); 6779 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS; 6780 flowc->mnemval[5].val = htobe32(cst->schedcl); 6781 6782 commit_wrq_wr(&cst->eo_txq->wrq, flowc, &cookie); 6783 6784 cst->flags &= ~EO_FLOWC_PENDING; 6785 cst->flags |= EO_FLOWC_RPL_PENDING; 6786 MPASS(cst->tx_credits >= ETID_FLOWC_LEN16); /* flowc is first WR. */ 6787 cst->tx_credits -= ETID_FLOWC_LEN16; 6788 6789 return (0); 6790 } 6791 #endif 6792 6793 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16)) 6794 6795 void 6796 send_etid_flush_wr(struct cxgbe_rate_tag *cst) 6797 { 6798 struct fw_flowc_wr *flowc; 6799 struct wrq_cookie cookie; 6800 6801 mtx_assert(&cst->lock, MA_OWNED); 6802 6803 flowc = start_wrq_wr(&cst->eo_txq->wrq, ETID_FLUSH_LEN16, &cookie); 6804 if (__predict_false(flowc == NULL)) 6805 CXGBE_UNIMPLEMENTED(__func__); 6806 6807 bzero(flowc, ETID_FLUSH_LEN16 * 16); 6808 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | 6809 V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL); 6810 flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) | 6811 V_FW_WR_FLOWID(cst->etid)); 6812 6813 commit_wrq_wr(&cst->eo_txq->wrq, flowc, &cookie); 6814 6815 cst->flags |= EO_FLUSH_RPL_PENDING; 6816 MPASS(cst->tx_credits >= ETID_FLUSH_LEN16); 6817 cst->tx_credits -= ETID_FLUSH_LEN16; 6818 cst->ncompl++; 6819 } 6820 6821 static void 6822 write_ethofld_wr(struct cxgbe_rate_tag *cst, struct fw_eth_tx_eo_wr *wr, 6823 struct mbuf *m0, int compl) 6824 { 6825 struct cpl_tx_pkt_core *cpl; 6826 uint64_t ctrl1; 6827 uint32_t ctrl; /* used in many unrelated places */ 6828 int len16, pktlen, nsegs, immhdrs; 6829 uintptr_t p; 6830 struct ulptx_sgl *usgl; 6831 struct sglist sg; 6832 struct sglist_seg segs[38]; /* XXX: find real limit. XXX: get off the stack */ 6833 6834 mtx_assert(&cst->lock, MA_OWNED); 6835 M_ASSERTPKTHDR(m0); 6836 KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 && 6837 m0->m_pkthdr.l4hlen > 0, 6838 ("%s: ethofld mbuf %p is missing header lengths", __func__, m0)); 6839 6840 len16 = mbuf_eo_len16(m0); 6841 nsegs = mbuf_eo_nsegs(m0); 6842 pktlen = m0->m_pkthdr.len; 6843 ctrl = sizeof(struct cpl_tx_pkt_core); 6844 if (needs_tso(m0)) 6845 ctrl += sizeof(struct cpl_tx_pkt_lso_core); 6846 immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen; 6847 ctrl += immhdrs; 6848 6849 wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) | 6850 V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl)); 6851 wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) | 6852 V_FW_WR_FLOWID(cst->etid)); 6853 wr->r3 = 0; 6854 if (needs_outer_udp_csum(m0)) { 6855 wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG; 6856 wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen; 6857 wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen); 6858 wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen; 6859 wr->u.udpseg.rtplen = 0; 6860 wr->u.udpseg.r4 = 0; 6861 wr->u.udpseg.mss = htobe16(pktlen - immhdrs); 6862 wr->u.udpseg.schedpktsize = wr->u.udpseg.mss; 6863 wr->u.udpseg.plen = htobe32(pktlen - immhdrs); 6864 cpl = (void *)(wr + 1); 6865 } else { 6866 MPASS(needs_outer_tcp_csum(m0)); 6867 wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG; 6868 wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen; 6869 wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen); 6870 wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen; 6871 wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0); 6872 wr->u.tcpseg.r4 = 0; 6873 wr->u.tcpseg.r5 = 0; 6874 wr->u.tcpseg.plen = htobe32(pktlen - immhdrs); 6875 6876 if (needs_tso(m0)) { 6877 struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1); 6878 6879 wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz); 6880 6881 ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | 6882 F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE | 6883 V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen - 6884 ETHER_HDR_LEN) >> 2) | 6885 V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) | 6886 V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2); 6887 if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr)) 6888 ctrl |= F_LSO_IPV6; 6889 lso->lso_ctrl = htobe32(ctrl); 6890 lso->ipid_ofst = htobe16(0); 6891 lso->mss = htobe16(m0->m_pkthdr.tso_segsz); 6892 lso->seqno_offset = htobe32(0); 6893 lso->len = htobe32(pktlen); 6894 6895 cpl = (void *)(lso + 1); 6896 } else { 6897 wr->u.tcpseg.mss = htobe16(0xffff); 6898 cpl = (void *)(wr + 1); 6899 } 6900 } 6901 6902 /* Checksum offload must be requested for ethofld. */ 6903 MPASS(needs_outer_l4_csum(m0)); 6904 ctrl1 = csum_to_ctrl(cst->adapter, m0); 6905 6906 /* VLAN tag insertion */ 6907 if (needs_vlan_insertion(m0)) { 6908 ctrl1 |= F_TXPKT_VLAN_VLD | 6909 V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag); 6910 } 6911 6912 /* CPL header */ 6913 cpl->ctrl0 = cst->ctrl0; 6914 cpl->pack = 0; 6915 cpl->len = htobe16(pktlen); 6916 cpl->ctrl1 = htobe64(ctrl1); 6917 6918 /* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */ 6919 p = (uintptr_t)(cpl + 1); 6920 m_copydata(m0, 0, immhdrs, (void *)p); 6921 6922 /* SGL */ 6923 if (nsegs > 0) { 6924 int i, pad; 6925 6926 /* zero-pad upto next 16Byte boundary, if not 16Byte aligned */ 6927 p += immhdrs; 6928 pad = 16 - (immhdrs & 0xf); 6929 bzero((void *)p, pad); 6930 6931 usgl = (void *)(p + pad); 6932 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | 6933 V_ULPTX_NSGE(nsegs)); 6934 6935 sglist_init(&sg, nitems(segs), segs); 6936 for (; m0 != NULL; m0 = m0->m_next) { 6937 if (__predict_false(m0->m_len == 0)) 6938 continue; 6939 if (immhdrs >= m0->m_len) { 6940 immhdrs -= m0->m_len; 6941 continue; 6942 } 6943 if (m0->m_flags & M_EXTPG) 6944 sglist_append_mbuf_epg(&sg, m0, 6945 mtod(m0, vm_offset_t), m0->m_len); 6946 else 6947 sglist_append(&sg, mtod(m0, char *) + immhdrs, 6948 m0->m_len - immhdrs); 6949 immhdrs = 0; 6950 } 6951 MPASS(sg.sg_nseg == nsegs); 6952 6953 /* 6954 * Zero pad last 8B in case the WR doesn't end on a 16B 6955 * boundary. 6956 */ 6957 *(uint64_t *)((char *)wr + len16 * 16 - 8) = 0; 6958 6959 usgl->len0 = htobe32(segs[0].ss_len); 6960 usgl->addr0 = htobe64(segs[0].ss_paddr); 6961 for (i = 0; i < nsegs - 1; i++) { 6962 usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len); 6963 usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr); 6964 } 6965 if (i & 1) 6966 usgl->sge[i / 2].len[1] = htobe32(0); 6967 } 6968 6969 } 6970 6971 static void 6972 ethofld_tx(struct cxgbe_rate_tag *cst) 6973 { 6974 struct mbuf *m; 6975 struct wrq_cookie cookie; 6976 int next_credits, compl; 6977 struct fw_eth_tx_eo_wr *wr; 6978 6979 mtx_assert(&cst->lock, MA_OWNED); 6980 6981 while ((m = mbufq_first(&cst->pending_tx)) != NULL) { 6982 M_ASSERTPKTHDR(m); 6983 6984 /* How many len16 credits do we need to send this mbuf. */ 6985 next_credits = mbuf_eo_len16(m); 6986 MPASS(next_credits > 0); 6987 if (next_credits > cst->tx_credits) { 6988 /* 6989 * Tx will make progress eventually because there is at 6990 * least one outstanding fw4_ack that will return 6991 * credits and kick the tx. 6992 */ 6993 MPASS(cst->ncompl > 0); 6994 return; 6995 } 6996 wr = start_wrq_wr(&cst->eo_txq->wrq, next_credits, &cookie); 6997 if (__predict_false(wr == NULL)) { 6998 /* XXX: wishful thinking, not a real assertion. */ 6999 MPASS(cst->ncompl > 0); 7000 return; 7001 } 7002 cst->tx_credits -= next_credits; 7003 cst->tx_nocompl += next_credits; 7004 compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2; 7005 ETHER_BPF_MTAP(cst->com.ifp, m); 7006 write_ethofld_wr(cst, wr, m, compl); 7007 commit_wrq_wr(&cst->eo_txq->wrq, wr, &cookie); 7008 if (compl) { 7009 cst->ncompl++; 7010 cst->tx_nocompl = 0; 7011 } 7012 (void) mbufq_dequeue(&cst->pending_tx); 7013 7014 /* 7015 * Drop the mbuf's reference on the tag now rather 7016 * than waiting until m_freem(). This ensures that 7017 * cxgbe_rate_tag_free gets called when the inp drops 7018 * its reference on the tag and there are no more 7019 * mbufs in the pending_tx queue and can flush any 7020 * pending requests. Otherwise if the last mbuf 7021 * doesn't request a completion the etid will never be 7022 * released. 7023 */ 7024 m->m_pkthdr.snd_tag = NULL; 7025 m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG; 7026 m_snd_tag_rele(&cst->com); 7027 7028 mbufq_enqueue(&cst->pending_fwack, m); 7029 } 7030 } 7031 7032 #if defined(INET) || defined(INET6) 7033 static int 7034 ethofld_transmit(if_t ifp, struct mbuf *m0) 7035 { 7036 struct cxgbe_rate_tag *cst; 7037 int rc; 7038 7039 MPASS(m0->m_nextpkt == NULL); 7040 MPASS(m0->m_pkthdr.csum_flags & CSUM_SND_TAG); 7041 MPASS(m0->m_pkthdr.snd_tag != NULL); 7042 cst = mst_to_crt(m0->m_pkthdr.snd_tag); 7043 7044 mtx_lock(&cst->lock); 7045 MPASS(cst->flags & EO_SND_TAG_REF); 7046 7047 if (__predict_false(cst->flags & EO_FLOWC_PENDING)) { 7048 struct vi_info *vi = if_getsoftc(ifp); 7049 struct port_info *pi = vi->pi; 7050 struct adapter *sc = pi->adapter; 7051 const uint32_t rss_mask = vi->rss_size - 1; 7052 uint32_t rss_hash; 7053 7054 cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq]; 7055 if (M_HASHTYPE_ISHASH(m0)) 7056 rss_hash = m0->m_pkthdr.flowid; 7057 else 7058 rss_hash = arc4random(); 7059 /* We assume RSS hashing */ 7060 cst->iqid = vi->rss[rss_hash & rss_mask]; 7061 cst->eo_txq += rss_hash % vi->nofldtxq; 7062 rc = send_etid_flowc_wr(cst, pi, vi); 7063 if (rc != 0) 7064 goto done; 7065 } 7066 7067 if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) { 7068 rc = ENOBUFS; 7069 goto done; 7070 } 7071 7072 mbufq_enqueue(&cst->pending_tx, m0); 7073 cst->plen += m0->m_pkthdr.len; 7074 7075 /* 7076 * Hold an extra reference on the tag while generating work 7077 * requests to ensure that we don't try to free the tag during 7078 * ethofld_tx() in case we are sending the final mbuf after 7079 * the inp was freed. 7080 */ 7081 m_snd_tag_ref(&cst->com); 7082 ethofld_tx(cst); 7083 mtx_unlock(&cst->lock); 7084 m_snd_tag_rele(&cst->com); 7085 return (0); 7086 7087 done: 7088 mtx_unlock(&cst->lock); 7089 return (rc); 7090 } 7091 #endif 7092 7093 static int 7094 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0) 7095 { 7096 struct adapter *sc = iq->adapter; 7097 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1); 7098 struct mbuf *m; 7099 u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl))); 7100 struct cxgbe_rate_tag *cst; 7101 uint8_t credits = cpl->credits; 7102 7103 cst = lookup_etid(sc, etid); 7104 mtx_lock(&cst->lock); 7105 if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) { 7106 MPASS(credits >= ETID_FLOWC_LEN16); 7107 credits -= ETID_FLOWC_LEN16; 7108 cst->flags &= ~EO_FLOWC_RPL_PENDING; 7109 } 7110 7111 KASSERT(cst->ncompl > 0, 7112 ("%s: etid %u (%p) wasn't expecting completion.", 7113 __func__, etid, cst)); 7114 cst->ncompl--; 7115 7116 while (credits > 0) { 7117 m = mbufq_dequeue(&cst->pending_fwack); 7118 if (__predict_false(m == NULL)) { 7119 /* 7120 * The remaining credits are for the final flush that 7121 * was issued when the tag was freed by the kernel. 7122 */ 7123 MPASS((cst->flags & 7124 (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) == 7125 EO_FLUSH_RPL_PENDING); 7126 MPASS(credits == ETID_FLUSH_LEN16); 7127 MPASS(cst->tx_credits + cpl->credits == cst->tx_total); 7128 MPASS(cst->ncompl == 0); 7129 7130 cst->flags &= ~EO_FLUSH_RPL_PENDING; 7131 cst->tx_credits += cpl->credits; 7132 cxgbe_rate_tag_free_locked(cst); 7133 return (0); /* cst is gone. */ 7134 } 7135 KASSERT(m != NULL, 7136 ("%s: too many credits (%u, %u)", __func__, cpl->credits, 7137 credits)); 7138 KASSERT(credits >= mbuf_eo_len16(m), 7139 ("%s: too few credits (%u, %u, %u)", __func__, 7140 cpl->credits, credits, mbuf_eo_len16(m))); 7141 credits -= mbuf_eo_len16(m); 7142 cst->plen -= m->m_pkthdr.len; 7143 m_freem(m); 7144 } 7145 7146 cst->tx_credits += cpl->credits; 7147 MPASS(cst->tx_credits <= cst->tx_total); 7148 7149 if (cst->flags & EO_SND_TAG_REF) { 7150 /* 7151 * As with ethofld_transmit(), hold an extra reference 7152 * so that the tag is stable across ethold_tx(). 7153 */ 7154 m_snd_tag_ref(&cst->com); 7155 m = mbufq_first(&cst->pending_tx); 7156 if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m)) 7157 ethofld_tx(cst); 7158 mtx_unlock(&cst->lock); 7159 m_snd_tag_rele(&cst->com); 7160 } else { 7161 /* 7162 * There shouldn't be any pending packets if the tag 7163 * was freed by the kernel since any pending packet 7164 * should hold a reference to the tag. 7165 */ 7166 MPASS(mbufq_first(&cst->pending_tx) == NULL); 7167 mtx_unlock(&cst->lock); 7168 } 7169 7170 return (0); 7171 } 7172 #endif 7173