t4_sge.c revision 266965
1218792Snp/*- 2218792Snp * Copyright (c) 2011 Chelsio Communications, Inc. 3218792Snp * All rights reserved. 4218792Snp * Written by: Navdeep Parhar <np@FreeBSD.org> 5218792Snp * 6218792Snp * Redistribution and use in source and binary forms, with or without 7218792Snp * modification, are permitted provided that the following conditions 8218792Snp * are met: 9218792Snp * 1. Redistributions of source code must retain the above copyright 10218792Snp * notice, this list of conditions and the following disclaimer. 11218792Snp * 2. Redistributions in binary form must reproduce the above copyright 12218792Snp * notice, this list of conditions and the following disclaimer in the 13218792Snp * documentation and/or other materials provided with the distribution. 14218792Snp * 15218792Snp * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16218792Snp * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17218792Snp * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18218792Snp * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19218792Snp * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20218792Snp * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21218792Snp * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22218792Snp * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23218792Snp * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24218792Snp * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25218792Snp * SUCH DAMAGE. 26218792Snp */ 27218792Snp 28218792Snp#include <sys/cdefs.h> 29218792Snp__FBSDID("$FreeBSD: stable/10/sys/dev/cxgbe/t4_sge.c 266965 2014-06-02 05:01:08Z np $"); 30218792Snp 31218792Snp#include "opt_inet.h" 32237819Snp#include "opt_inet6.h" 33218792Snp 34218792Snp#include <sys/types.h> 35218792Snp#include <sys/mbuf.h> 36218792Snp#include <sys/socket.h> 37218792Snp#include <sys/kernel.h> 38237263Snp#include <sys/kdb.h> 39219286Snp#include <sys/malloc.h> 40219286Snp#include <sys/queue.h> 41265425Snp#include <sys/sbuf.h> 42219286Snp#include <sys/taskqueue.h> 43255015Snp#include <sys/time.h> 44218792Snp#include <sys/sysctl.h> 45228561Snp#include <sys/smp.h> 46218792Snp#include <net/bpf.h> 47218792Snp#include <net/ethernet.h> 48218792Snp#include <net/if.h> 49218792Snp#include <net/if_vlan_var.h> 50218792Snp#include <netinet/in.h> 51218792Snp#include <netinet/ip.h> 52237819Snp#include <netinet/ip6.h> 53218792Snp#include <netinet/tcp.h> 54256131Sdim#include <machine/md_var.h> 55265425Snp#include <vm/vm.h> 56265425Snp#include <vm/pmap.h> 57218792Snp 58218792Snp#include "common/common.h" 59218792Snp#include "common/t4_regs.h" 60218792Snp#include "common/t4_regs_values.h" 61218792Snp#include "common/t4_msg.h" 62218792Snp 63248925Snp#ifdef T4_PKT_TIMESTAMP 64248925Snp#define RX_COPY_THRESHOLD (MINCLSIZE - 8) 65248925Snp#else 66248925Snp#define RX_COPY_THRESHOLD MINCLSIZE 67248925Snp#endif 68248925Snp 69239258Snp/* 70239258Snp * Ethernet frames are DMA'd at this byte offset into the freelist buffer. 71239258Snp * 0-7 are valid values. 72239258Snp */ 73239258Snpstatic int fl_pktshift = 2; 74239258SnpTUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift); 75218792Snp 76239258Snp/* 77239258Snp * Pad ethernet payload up to this boundary. 78239258Snp * -1: driver should figure out a good value. 79255050Snp * 0: disable padding. 80255050Snp * Any power of 2 from 32 to 4096 (both inclusive) is also a valid value. 81239258Snp */ 82239258Snpstatic int fl_pad = -1; 83239258SnpTUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad); 84218792Snp 85239258Snp/* 86239258Snp * Status page length. 87239258Snp * -1: driver should figure out a good value. 88239258Snp * 64 or 128 are the only other valid values. 89239258Snp */ 90239258Snpstatic int spg_len = -1; 91239258SnpTUNABLE_INT("hw.cxgbe.spg_len", &spg_len); 92239258Snp 93239258Snp/* 94239258Snp * Congestion drops. 95239258Snp * -1: no congestion feedback (not recommended). 96239258Snp * 0: backpressure the channel instead of dropping packets right away. 97239258Snp * 1: no backpressure, drop packets for the congested queue immediately. 98239258Snp */ 99239258Snpstatic int cong_drop = 0; 100239258SnpTUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop); 101239258Snp 102255050Snp/* 103255050Snp * Deliver multiple frames in the same free list buffer if they fit. 104255050Snp * -1: let the driver decide whether to enable buffer packing or not. 105255050Snp * 0: disable buffer packing. 106255050Snp * 1: enable buffer packing. 107255050Snp */ 108255050Snpstatic int buffer_packing = -1; 109255050SnpTUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing); 110255050Snp 111255050Snp/* 112255050Snp * Start next frame in a packed buffer at this boundary. 113255050Snp * -1: driver should figure out a good value. 114255050Snp * T4: 115255050Snp * --- 116255050Snp * if fl_pad != 0 117255050Snp * value specified here will be overridden by fl_pad. 118255050Snp * else 119255050Snp * power of 2 from 32 to 4096 (both inclusive) is a valid value here. 120255050Snp * T5: 121255050Snp * --- 122255050Snp * 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value. 123255050Snp */ 124255050Snpstatic int fl_pack = -1; 125255050Snpstatic int t4_fl_pack; 126255050Snpstatic int t5_fl_pack; 127255050SnpTUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack); 128255050Snp 129265425Snp/* 130265425Snp * Allow the driver to create mbuf(s) in a cluster allocated for rx. 131265425Snp * 0: never; always allocate mbufs from the zone_mbuf UMA zone. 132265425Snp * 1: ok to create mbuf(s) within a cluster if there is room. 133265425Snp */ 134265425Snpstatic int allow_mbufs_in_cluster = 1; 135265425SnpTUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster); 136265425Snp 137265425Snp/* 138265425Snp * Largest rx cluster size that the driver is allowed to allocate. 139265425Snp */ 140265425Snpstatic int largest_rx_cluster = MJUM16BYTES; 141265425SnpTUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster); 142265425Snp 143265425Snp/* 144265425Snp * Size of cluster allocation that's most likely to succeed. The driver will 145265425Snp * fall back to this size if it fails to allocate clusters larger than this. 146265425Snp */ 147265425Snpstatic int safest_rx_cluster = PAGE_SIZE; 148265425SnpTUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster); 149265425Snp 150218792Snp/* Used to track coalesced tx work request */ 151218792Snpstruct txpkts { 152218792Snp uint64_t *flitp; /* ptr to flit where next pkt should start */ 153218792Snp uint8_t npkt; /* # of packets in this work request */ 154218792Snp uint8_t nflits; /* # of flits used by this work request */ 155218792Snp uint16_t plen; /* total payload (sum of all packets) */ 156218792Snp}; 157218792Snp 158218792Snp/* A packet's SGL. This + m_pkthdr has all info needed for tx */ 159218792Snpstruct sgl { 160218792Snp int nsegs; /* # of segments in the SGL, 0 means imm. tx */ 161218792Snp int nflits; /* # of flits needed for the SGL */ 162218792Snp bus_dma_segment_t seg[TX_SGL_SEGS]; 163218792Snp}; 164218792Snp 165228561Snpstatic int service_iq(struct sge_iq *, int); 166265425Snpstatic struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t, 167228561Snp int *); 168228561Snpstatic int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *); 169218792Snpstatic inline void init_iq(struct sge_iq *, struct adapter *, int, int, int, 170241397Snp int); 171255050Snpstatic inline void init_fl(struct adapter *, struct sge_fl *, int, int, int, 172255050Snp char *); 173228561Snpstatic inline void init_eq(struct sge_eq *, int, int, uint8_t, uint16_t, 174228561Snp char *); 175218792Snpstatic int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *, 176218792Snp bus_addr_t *, void **); 177218792Snpstatic int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t, 178218792Snp void *); 179218792Snpstatic int alloc_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *, 180222085Snp int, int); 181218792Snpstatic int free_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *); 182265425Snpstatic void add_fl_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *, 183265425Snp struct sge_fl *); 184228561Snpstatic int alloc_fwq(struct adapter *); 185228561Snpstatic int free_fwq(struct adapter *); 186228561Snpstatic int alloc_mgmtq(struct adapter *); 187228561Snpstatic int free_mgmtq(struct adapter *); 188228561Snpstatic int alloc_rxq(struct port_info *, struct sge_rxq *, int, int, 189228561Snp struct sysctl_oid *); 190218792Snpstatic int free_rxq(struct port_info *, struct sge_rxq *); 191237263Snp#ifdef TCP_OFFLOAD 192228561Snpstatic int alloc_ofld_rxq(struct port_info *, struct sge_ofld_rxq *, int, int, 193228561Snp struct sysctl_oid *); 194228561Snpstatic int free_ofld_rxq(struct port_info *, struct sge_ofld_rxq *); 195228561Snp#endif 196228561Snpstatic int ctrl_eq_alloc(struct adapter *, struct sge_eq *); 197228561Snpstatic int eth_eq_alloc(struct adapter *, struct port_info *, struct sge_eq *); 198237263Snp#ifdef TCP_OFFLOAD 199228561Snpstatic int ofld_eq_alloc(struct adapter *, struct port_info *, struct sge_eq *); 200228561Snp#endif 201228561Snpstatic int alloc_eq(struct adapter *, struct port_info *, struct sge_eq *); 202228561Snpstatic int free_eq(struct adapter *, struct sge_eq *); 203228561Snpstatic int alloc_wrq(struct adapter *, struct port_info *, struct sge_wrq *, 204228561Snp struct sysctl_oid *); 205228561Snpstatic int free_wrq(struct adapter *, struct sge_wrq *); 206228561Snpstatic int alloc_txq(struct port_info *, struct sge_txq *, int, 207228561Snp struct sysctl_oid *); 208218792Snpstatic int free_txq(struct port_info *, struct sge_txq *); 209218792Snpstatic void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int); 210218792Snpstatic inline bool is_new_response(const struct sge_iq *, struct rsp_ctrl **); 211218792Snpstatic inline void iq_next(struct sge_iq *); 212218792Snpstatic inline void ring_fl_db(struct adapter *, struct sge_fl *); 213228561Snpstatic int refill_fl(struct adapter *, struct sge_fl *, int); 214228561Snpstatic void refill_sfl(void *); 215218792Snpstatic int alloc_fl_sdesc(struct sge_fl *); 216255050Snpstatic void free_fl_sdesc(struct adapter *, struct sge_fl *); 217265425Snpstatic void find_best_refill_source(struct adapter *, struct sge_fl *, int); 218265425Snpstatic void find_safe_refill_source(struct adapter *, struct sge_fl *); 219228561Snpstatic void add_fl_to_sfl(struct adapter *, struct sge_fl *); 220218792Snp 221218792Snpstatic int get_pkt_sgl(struct sge_txq *, struct mbuf **, struct sgl *, int); 222218792Snpstatic int free_pkt_sgl(struct sge_txq *, struct sgl *); 223218792Snpstatic int write_txpkt_wr(struct port_info *, struct sge_txq *, struct mbuf *, 224218792Snp struct sgl *); 225218792Snpstatic int add_to_txpkts(struct port_info *, struct sge_txq *, struct txpkts *, 226218792Snp struct mbuf *, struct sgl *); 227218792Snpstatic void write_txpkts_wr(struct sge_txq *, struct txpkts *); 228218792Snpstatic inline void write_ulp_cpl_sgl(struct port_info *, struct sge_txq *, 229218792Snp struct txpkts *, struct mbuf *, struct sgl *); 230218792Snpstatic int write_sgl_to_txd(struct sge_eq *, struct sgl *, caddr_t *); 231218792Snpstatic inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int); 232220873Snpstatic inline void ring_eq_db(struct adapter *, struct sge_eq *); 233219292Snpstatic inline int reclaimable(struct sge_eq *); 234220873Snpstatic int reclaim_tx_descs(struct sge_txq *, int, int); 235218792Snpstatic void write_eqflush_wr(struct sge_eq *); 236218792Snpstatic __be64 get_flit(bus_dma_segment_t *, int, int); 237228561Snpstatic int handle_sge_egr_update(struct sge_iq *, const struct rss_header *, 238228561Snp struct mbuf *); 239239336Snpstatic int handle_fw_msg(struct sge_iq *, const struct rss_header *, 240228561Snp struct mbuf *); 241218792Snp 242222510Snpstatic int sysctl_uint16(SYSCTL_HANDLER_ARGS); 243265425Snpstatic int sysctl_bufsizes(SYSCTL_HANDLER_ARGS); 244220873Snp 245219392Snp/* 246255050Snp * Called on MOD_LOAD. Validates and calculates the SGE tunables. 247219392Snp */ 248219392Snpvoid 249219392Snpt4_sge_modload(void) 250219392Snp{ 251255050Snp int pad; 252255050Snp 253255050Snp /* set pad to a reasonable powerof2 between 16 and 4096 (inclusive) */ 254255050Snp#if defined(__i386__) || defined(__amd64__) 255255050Snp pad = max(cpu_clflush_line_size, 16); 256255050Snp#else 257255050Snp pad = max(CACHE_LINE_SIZE, 16); 258219392Snp#endif 259255050Snp pad = min(pad, 4096); 260219392Snp 261239258Snp if (fl_pktshift < 0 || fl_pktshift > 7) { 262239258Snp printf("Invalid hw.cxgbe.fl_pktshift value (%d)," 263239258Snp " using 2 instead.\n", fl_pktshift); 264239258Snp fl_pktshift = 2; 265239258Snp } 266239258Snp 267255050Snp if (fl_pad != 0 && 268255050Snp (fl_pad < 32 || fl_pad > 4096 || !powerof2(fl_pad))) { 269239258Snp 270239258Snp if (fl_pad != -1) { 271239258Snp printf("Invalid hw.cxgbe.fl_pad value (%d)," 272255050Snp " using %d instead.\n", fl_pad, max(pad, 32)); 273239258Snp } 274255050Snp fl_pad = max(pad, 32); 275239258Snp } 276239258Snp 277255050Snp /* 278255050Snp * T4 has the same pad and pack boundary. If a pad boundary is set, 279255050Snp * pack boundary must be set to the same value. Otherwise take the 280255050Snp * specified value or auto-calculate something reasonable. 281255050Snp */ 282255050Snp if (fl_pad) 283255050Snp t4_fl_pack = fl_pad; 284255050Snp else if (fl_pack < 32 || fl_pack > 4096 || !powerof2(fl_pack)) 285255050Snp t4_fl_pack = max(pad, 32); 286255050Snp else 287255050Snp t4_fl_pack = fl_pack; 288255050Snp 289255050Snp /* T5's pack boundary is independent of the pad boundary. */ 290255050Snp if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 || 291255050Snp !powerof2(fl_pack)) 292265425Snp t5_fl_pack = max(pad, CACHE_LINE_SIZE); 293255050Snp else 294255050Snp t5_fl_pack = fl_pack; 295255050Snp 296239258Snp if (spg_len != 64 && spg_len != 128) { 297239258Snp int len; 298239258Snp 299239258Snp#if defined(__i386__) || defined(__amd64__) 300239258Snp len = cpu_clflush_line_size > 64 ? 128 : 64; 301239258Snp#else 302239258Snp len = 64; 303239258Snp#endif 304239258Snp if (spg_len != -1) { 305239258Snp printf("Invalid hw.cxgbe.spg_len value (%d)," 306239258Snp " using %d instead.\n", spg_len, len); 307239258Snp } 308239258Snp spg_len = len; 309239258Snp } 310239258Snp 311239258Snp if (cong_drop < -1 || cong_drop > 1) { 312239258Snp printf("Invalid hw.cxgbe.cong_drop value (%d)," 313239258Snp " using 0 instead.\n", cong_drop); 314239258Snp cong_drop = 0; 315239258Snp } 316219392Snp} 317219392Snp 318248925Snpvoid 319248925Snpt4_init_sge_cpl_handlers(struct adapter *sc) 320218792Snp{ 321218792Snp 322248925Snp t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_msg); 323248925Snp t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_msg); 324248925Snp t4_register_cpl_handler(sc, CPL_SGE_EGR_UPDATE, handle_sge_egr_update); 325248925Snp t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx); 326248925Snp t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL, t4_handle_fw_rpl); 327248925Snp} 328248925Snp 329249391Snp/* 330249391Snp * adap->params.vpd.cclk must be set up before this is called. 331249391Snp */ 332248925Snpvoid 333248925Snpt4_tweak_chip_settings(struct adapter *sc) 334248925Snp{ 335248925Snp int i; 336248925Snp uint32_t v, m; 337248925Snp int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200}; 338249391Snp int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk; 339248925Snp int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */ 340248925Snp uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE); 341265425Snp static int sge_flbuf_sizes[] = { 342255050Snp MCLBYTES, 343255050Snp#if MJUMPAGESIZE != MCLBYTES 344255050Snp MJUMPAGESIZE, 345265425Snp MJUMPAGESIZE - CL_METADATA_SIZE, 346265425Snp MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE, 347255050Snp#endif 348255050Snp MJUM9BYTES, 349255050Snp MJUM16BYTES, 350265425Snp MCLBYTES - MSIZE - CL_METADATA_SIZE, 351265425Snp MJUM9BYTES - CL_METADATA_SIZE, 352265425Snp MJUM16BYTES - CL_METADATA_SIZE, 353255050Snp }; 354248925Snp 355248925Snp KASSERT(sc->flags & MASTER_PF, 356248925Snp ("%s: trying to change chip settings when not master.", __func__)); 357248925Snp 358255050Snp m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE; 359248925Snp v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE | 360237512Snp V_EGRSTATUSPAGESIZE(spg_len == 128); 361255050Snp if (is_t4(sc) && (fl_pad || buffer_packing)) { 362255050Snp /* t4_fl_pack has the correct value even when fl_pad = 0 */ 363255050Snp m |= V_INGPADBOUNDARY(M_INGPADBOUNDARY); 364255050Snp v |= V_INGPADBOUNDARY(ilog2(t4_fl_pack) - 5); 365255050Snp } else if (is_t5(sc) && fl_pad) { 366255050Snp m |= V_INGPADBOUNDARY(M_INGPADBOUNDARY); 367255050Snp v |= V_INGPADBOUNDARY(ilog2(fl_pad) - 5); 368255050Snp } 369248925Snp t4_set_reg_field(sc, A_SGE_CONTROL, m, v); 370218792Snp 371255050Snp if (is_t5(sc) && buffer_packing) { 372255050Snp m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY); 373255050Snp if (t5_fl_pack == 16) 374255050Snp v = V_INGPACKBOUNDARY(0); 375255050Snp else 376255050Snp v = V_INGPACKBOUNDARY(ilog2(t5_fl_pack) - 5); 377255050Snp t4_set_reg_field(sc, A_SGE_CONTROL2, m, v); 378255050Snp } 379255050Snp 380248925Snp v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) | 381228561Snp V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) | 382228561Snp V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) | 383228561Snp V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) | 384228561Snp V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) | 385228561Snp V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) | 386228561Snp V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) | 387228561Snp V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10); 388248925Snp t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v); 389228561Snp 390265425Snp KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES, 391265425Snp ("%s: hw buffer size table too big", __func__)); 392265425Snp for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) { 393248925Snp t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i), 394265425Snp sge_flbuf_sizes[i]); 395248925Snp } 396228561Snp 397248925Snp v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) | 398248925Snp V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]); 399248925Snp t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v); 400228561Snp 401249391Snp KASSERT(intr_timer[0] <= timer_max, 402249391Snp ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0], 403249391Snp timer_max)); 404249391Snp for (i = 1; i < nitems(intr_timer); i++) { 405249391Snp KASSERT(intr_timer[i] >= intr_timer[i - 1], 406249391Snp ("%s: timers not listed in increasing order (%d)", 407249391Snp __func__, i)); 408249391Snp 409249391Snp while (intr_timer[i] > timer_max) { 410249391Snp if (i == nitems(intr_timer) - 1) { 411249391Snp intr_timer[i] = timer_max; 412249391Snp break; 413249391Snp } 414249391Snp intr_timer[i] += intr_timer[i - 1]; 415249391Snp intr_timer[i] /= 2; 416249391Snp } 417249391Snp } 418249391Snp 419248925Snp v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) | 420248925Snp V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1])); 421248925Snp t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v); 422248925Snp v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) | 423248925Snp V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3])); 424248925Snp t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v); 425248925Snp v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) | 426248925Snp V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5])); 427248925Snp t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v); 428228561Snp 429248925Snp if (cong_drop == 0) { 430248925Snp m = F_TUNNELCNGDROP0 | F_TUNNELCNGDROP1 | F_TUNNELCNGDROP2 | 431248925Snp F_TUNNELCNGDROP3; 432248925Snp t4_set_reg_field(sc, A_TP_PARA_REG3, m, 0); 433228561Snp } 434228561Snp 435248925Snp /* 4K, 16K, 64K, 256K DDP "page sizes" */ 436248925Snp v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6); 437248925Snp t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v); 438248925Snp 439248925Snp m = v = F_TDDPTAGTCB; 440248925Snp t4_set_reg_field(sc, A_ULP_RX_CTL, m, v); 441248925Snp 442248925Snp m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET | 443248925Snp F_RESETDDPOFFSET; 444248925Snp v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET; 445248925Snp t4_set_reg_field(sc, A_TP_PARA_REG5, m, v); 446248925Snp} 447248925Snp 448248925Snp/* 449265425Snp * SGE wants the buffer to be at least 64B and then a multiple of the pad 450265425Snp * boundary or 16, whichever is greater. 451265425Snp */ 452265425Snpstatic inline int 453265425Snphwsz_ok(int hwsz) 454265425Snp{ 455265425Snp int mask = max(fl_pad, 16) - 1; 456265425Snp 457265425Snp return (hwsz >= 64 && (hwsz & mask) == 0); 458265425Snp} 459265425Snp 460265425Snp/* 461248925Snp * XXX: driver really should be able to deal with unexpected settings. 462248925Snp */ 463248925Snpint 464248925Snpt4_read_chip_settings(struct adapter *sc) 465248925Snp{ 466248925Snp struct sge *s = &sc->sge; 467255050Snp int i, j, n, rc = 0; 468248925Snp uint32_t m, v, r; 469248925Snp uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE); 470265425Snp static int sw_buf_sizes[] = { /* Sorted by size */ 471255050Snp MCLBYTES, 472255050Snp#if MJUMPAGESIZE != MCLBYTES 473255050Snp MJUMPAGESIZE, 474255050Snp#endif 475255050Snp MJUM9BYTES, 476255050Snp MJUM16BYTES 477255050Snp }; 478265425Snp struct sw_zone_info *swz, *safe_swz; 479265425Snp struct hw_buf_info *hwb; 480248925Snp 481255050Snp m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE; 482248925Snp v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE | 483248925Snp V_EGRSTATUSPAGESIZE(spg_len == 128); 484255050Snp if (is_t4(sc) && (fl_pad || buffer_packing)) { 485255050Snp m |= V_INGPADBOUNDARY(M_INGPADBOUNDARY); 486255050Snp v |= V_INGPADBOUNDARY(ilog2(t4_fl_pack) - 5); 487255050Snp } else if (is_t5(sc) && fl_pad) { 488255050Snp m |= V_INGPADBOUNDARY(M_INGPADBOUNDARY); 489255050Snp v |= V_INGPADBOUNDARY(ilog2(fl_pad) - 5); 490255050Snp } 491248925Snp r = t4_read_reg(sc, A_SGE_CONTROL); 492248925Snp if ((r & m) != v) { 493248925Snp device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r); 494228561Snp rc = EINVAL; 495228561Snp } 496228561Snp 497255050Snp if (is_t5(sc) && buffer_packing) { 498255050Snp m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY); 499255050Snp if (t5_fl_pack == 16) 500255050Snp v = V_INGPACKBOUNDARY(0); 501255050Snp else 502255050Snp v = V_INGPACKBOUNDARY(ilog2(t5_fl_pack) - 5); 503255050Snp r = t4_read_reg(sc, A_SGE_CONTROL2); 504255050Snp if ((r & m) != v) { 505255050Snp device_printf(sc->dev, 506255050Snp "invalid SGE_CONTROL2(0x%x)\n", r); 507255050Snp rc = EINVAL; 508255050Snp } 509255050Snp } 510265425Snp s->pack_boundary = is_t4(sc) ? t4_fl_pack : t5_fl_pack; 511255050Snp 512248925Snp v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) | 513248925Snp V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) | 514248925Snp V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) | 515248925Snp V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) | 516248925Snp V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) | 517248925Snp V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) | 518248925Snp V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) | 519248925Snp V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10); 520248925Snp r = t4_read_reg(sc, A_SGE_HOST_PAGE_SIZE); 521248925Snp if (r != v) { 522248925Snp device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r); 523228561Snp rc = EINVAL; 524228561Snp } 525228561Snp 526265425Snp /* Filter out unusable hw buffer sizes entirely (mark with -2). */ 527265425Snp hwb = &s->hw_buf_info[0]; 528265425Snp for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) { 529265425Snp r = t4_read_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i)); 530265425Snp hwb->size = r; 531265425Snp hwb->zidx = hwsz_ok(r) ? -1 : -2; 532265425Snp hwb->next = -1; 533265425Snp } 534265425Snp 535255050Snp /* 536265425Snp * Create a sorted list in decreasing order of hw buffer sizes (and so 537265425Snp * increasing order of spare area) for each software zone. 538255050Snp */ 539265425Snp n = 0; /* no usable buffer size to begin with */ 540265425Snp swz = &s->sw_zone_info[0]; 541265425Snp safe_swz = NULL; 542265425Snp for (i = 0; i < SW_ZONE_SIZES; i++, swz++) { 543265425Snp int8_t head = -1, tail = -1; 544265425Snp 545265425Snp swz->size = sw_buf_sizes[i]; 546265425Snp swz->zone = m_getzone(swz->size); 547265425Snp swz->type = m_gettype(swz->size); 548265425Snp 549265425Snp if (swz->size == safest_rx_cluster) 550265425Snp safe_swz = swz; 551265425Snp 552265425Snp hwb = &s->hw_buf_info[0]; 553265425Snp for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) { 554265425Snp if (hwb->zidx != -1 || hwb->size > swz->size) 555265425Snp continue; 556265425Snp hwb->zidx = i; 557265425Snp if (head == -1) 558265425Snp head = tail = j; 559265425Snp else if (hwb->size < s->hw_buf_info[tail].size) { 560265425Snp s->hw_buf_info[tail].next = j; 561265425Snp tail = j; 562265425Snp } else { 563265425Snp int8_t *cur; 564265425Snp struct hw_buf_info *t; 565265425Snp 566265425Snp for (cur = &head; *cur != -1; cur = &t->next) { 567265425Snp t = &s->hw_buf_info[*cur]; 568265425Snp if (hwb->size == t->size) { 569265425Snp hwb->zidx = -2; 570265425Snp break; 571265425Snp } 572265425Snp if (hwb->size > t->size) { 573265425Snp hwb->next = *cur; 574265425Snp *cur = j; 575265425Snp break; 576265425Snp } 577265425Snp } 578265425Snp } 579228561Snp } 580265425Snp swz->head_hwidx = head; 581265425Snp swz->tail_hwidx = tail; 582265425Snp 583265425Snp if (tail != -1) { 584255050Snp n++; 585265425Snp if (swz->size - s->hw_buf_info[tail].size >= 586265425Snp CL_METADATA_SIZE) 587265425Snp sc->flags |= BUF_PACKING_OK; 588255050Snp } 589255050Snp } 590255050Snp if (n == 0) { 591255050Snp device_printf(sc->dev, "no usable SGE FL buffer size.\n"); 592255050Snp rc = EINVAL; 593255050Snp } 594218792Snp 595265425Snp s->safe_hwidx1 = -1; 596265425Snp s->safe_hwidx2 = -1; 597265425Snp if (safe_swz != NULL) { 598265425Snp s->safe_hwidx1 = safe_swz->head_hwidx; 599265425Snp for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) { 600265425Snp int spare; 601265425Snp 602265425Snp hwb = &s->hw_buf_info[i]; 603265425Snp spare = safe_swz->size - hwb->size; 604265425Snp if (spare < CL_METADATA_SIZE) 605265425Snp continue; 606265425Snp if (s->safe_hwidx2 == -1 || 607265425Snp spare == CL_METADATA_SIZE + MSIZE) 608265425Snp s->safe_hwidx2 = i; 609265425Snp if (spare >= CL_METADATA_SIZE + MSIZE) 610265425Snp break; 611265425Snp } 612265425Snp } 613265425Snp 614248925Snp r = t4_read_reg(sc, A_SGE_INGRESS_RX_THRESHOLD); 615248925Snp s->counter_val[0] = G_THRESHOLD_0(r); 616248925Snp s->counter_val[1] = G_THRESHOLD_1(r); 617248925Snp s->counter_val[2] = G_THRESHOLD_2(r); 618248925Snp s->counter_val[3] = G_THRESHOLD_3(r); 619222701Snp 620248925Snp r = t4_read_reg(sc, A_SGE_TIMER_VALUE_0_AND_1); 621248925Snp s->timer_val[0] = G_TIMERVALUE0(r) / core_ticks_per_usec(sc); 622248925Snp s->timer_val[1] = G_TIMERVALUE1(r) / core_ticks_per_usec(sc); 623248925Snp r = t4_read_reg(sc, A_SGE_TIMER_VALUE_2_AND_3); 624248925Snp s->timer_val[2] = G_TIMERVALUE2(r) / core_ticks_per_usec(sc); 625248925Snp s->timer_val[3] = G_TIMERVALUE3(r) / core_ticks_per_usec(sc); 626248925Snp r = t4_read_reg(sc, A_SGE_TIMER_VALUE_4_AND_5); 627248925Snp s->timer_val[4] = G_TIMERVALUE4(r) / core_ticks_per_usec(sc); 628248925Snp s->timer_val[5] = G_TIMERVALUE5(r) / core_ticks_per_usec(sc); 629218792Snp 630248925Snp if (cong_drop == 0) { 631248925Snp m = F_TUNNELCNGDROP0 | F_TUNNELCNGDROP1 | F_TUNNELCNGDROP2 | 632248925Snp F_TUNNELCNGDROP3; 633248925Snp r = t4_read_reg(sc, A_TP_PARA_REG3); 634248925Snp if (r & m) { 635248925Snp device_printf(sc->dev, 636248925Snp "invalid TP_PARA_REG3(0x%x)\n", r); 637248925Snp rc = EINVAL; 638248925Snp } 639248925Snp } 640228561Snp 641248925Snp v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6); 642248925Snp r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ); 643248925Snp if (r != v) { 644248925Snp device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r); 645248925Snp rc = EINVAL; 646248925Snp } 647228561Snp 648248925Snp m = v = F_TDDPTAGTCB; 649248925Snp r = t4_read_reg(sc, A_ULP_RX_CTL); 650248925Snp if ((r & m) != v) { 651248925Snp device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r); 652248925Snp rc = EINVAL; 653248925Snp } 654239336Snp 655248925Snp m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET | 656248925Snp F_RESETDDPOFFSET; 657248925Snp v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET; 658248925Snp r = t4_read_reg(sc, A_TP_PARA_REG5); 659248925Snp if ((r & m) != v) { 660248925Snp device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r); 661248925Snp rc = EINVAL; 662248925Snp } 663248925Snp 664248925Snp r = t4_read_reg(sc, A_SGE_CONM_CTRL); 665248925Snp s->fl_starve_threshold = G_EGRTHRESHOLD(r) * 2 + 1; 666265410Snp if (is_t4(sc)) 667265410Snp s->fl_starve_threshold2 = s->fl_starve_threshold; 668265410Snp else 669265410Snp s->fl_starve_threshold2 = G_EGRTHRESHOLDPACKING(r) * 2 + 1; 670248925Snp 671256794Snp /* egress queues: log2 of # of doorbells per BAR2 page */ 672256794Snp r = t4_read_reg(sc, A_SGE_EGRESS_QUEUES_PER_PAGE_PF); 673256794Snp r >>= S_QUEUESPERPAGEPF0 + 674256794Snp (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf; 675256794Snp s->eq_s_qpp = r & M_QUEUESPERPAGEPF0; 676248925Snp 677256794Snp /* ingress queues: log2 of # of doorbells per BAR2 page */ 678256794Snp r = t4_read_reg(sc, A_SGE_INGRESS_QUEUES_PER_PAGE_PF); 679256794Snp r >>= S_QUEUESPERPAGEPF0 + 680256794Snp (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf; 681256794Snp s->iq_s_qpp = r & M_QUEUESPERPAGEPF0; 682256794Snp 683252705Snp t4_init_tp_params(sc); 684248925Snp 685248925Snp t4_read_mtu_tbl(sc, sc->params.mtus, NULL); 686248925Snp t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd); 687248925Snp 688228561Snp return (rc); 689218792Snp} 690218792Snp 691218792Snpint 692218792Snpt4_create_dma_tag(struct adapter *sc) 693218792Snp{ 694218792Snp int rc; 695218792Snp 696218792Snp rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0, 697218792Snp BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE, 698218792Snp BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL, 699218792Snp NULL, &sc->dmat); 700218792Snp if (rc != 0) { 701218792Snp device_printf(sc->dev, 702218792Snp "failed to create main DMA tag: %d\n", rc); 703218792Snp } 704218792Snp 705218792Snp return (rc); 706218792Snp} 707218792Snp 708255052Snpstatic inline int 709255052Snpenable_buffer_packing(struct adapter *sc) 710255052Snp{ 711255052Snp 712255052Snp if (sc->flags & BUF_PACKING_OK && 713255052Snp ((is_t5(sc) && buffer_packing) || /* 1 or -1 both ok for T5 */ 714255052Snp (is_t4(sc) && buffer_packing == 1))) 715255052Snp return (1); 716255052Snp return (0); 717255052Snp} 718255052Snp 719253829Snpvoid 720253829Snpt4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx, 721253829Snp struct sysctl_oid_list *children) 722253829Snp{ 723253829Snp 724265425Snp SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes", 725265425Snp CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A", 726265425Snp "freelist buffer sizes"); 727265425Snp 728253829Snp SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD, 729253829Snp NULL, fl_pktshift, "payload DMA offset in rx buffer (bytes)"); 730253829Snp 731253829Snp SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD, 732253829Snp NULL, fl_pad, "payload pad boundary (bytes)"); 733253829Snp 734253829Snp SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD, 735253829Snp NULL, spg_len, "status page size (bytes)"); 736253829Snp 737253829Snp SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD, 738253829Snp NULL, cong_drop, "congestion drop setting"); 739255050Snp 740255050Snp SYSCTL_ADD_INT(ctx, children, OID_AUTO, "buffer_packing", CTLFLAG_RD, 741255052Snp NULL, enable_buffer_packing(sc), 742255050Snp "pack multiple frames in one fl buffer"); 743255050Snp 744255050Snp SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD, 745265425Snp NULL, sc->sge.pack_boundary, "payload pack boundary (bytes)"); 746253829Snp} 747253829Snp 748218792Snpint 749218792Snpt4_destroy_dma_tag(struct adapter *sc) 750218792Snp{ 751218792Snp if (sc->dmat) 752218792Snp bus_dma_tag_destroy(sc->dmat); 753218792Snp 754218792Snp return (0); 755218792Snp} 756218792Snp 757218792Snp/* 758228561Snp * Allocate and initialize the firmware event queue and the management queue. 759218792Snp * 760218792Snp * Returns errno on failure. Resources allocated up to that point may still be 761218792Snp * allocated. Caller is responsible for cleanup in case this function fails. 762218792Snp */ 763218792Snpint 764220873Snpt4_setup_adapter_queues(struct adapter *sc) 765218792Snp{ 766228561Snp int rc; 767218792Snp 768218792Snp ADAPTER_LOCK_ASSERT_NOTOWNED(sc); 769218792Snp 770228561Snp sysctl_ctx_init(&sc->ctx); 771228561Snp sc->flags |= ADAP_SYSCTL_CTX; 772220873Snp 773222510Snp /* 774222510Snp * Firmware event queue 775222510Snp */ 776228561Snp rc = alloc_fwq(sc); 777241398Snp if (rc != 0) 778220873Snp return (rc); 779218792Snp 780220873Snp /* 781228561Snp * Management queue. This is just a control queue that uses the fwq as 782228561Snp * its associated iq. 783220873Snp */ 784228561Snp rc = alloc_mgmtq(sc); 785220873Snp 786218792Snp return (rc); 787218792Snp} 788218792Snp 789218792Snp/* 790218792Snp * Idempotent 791218792Snp */ 792218792Snpint 793220873Snpt4_teardown_adapter_queues(struct adapter *sc) 794218792Snp{ 795218792Snp 796218792Snp ADAPTER_LOCK_ASSERT_NOTOWNED(sc); 797218792Snp 798228561Snp /* Do this before freeing the queue */ 799228561Snp if (sc->flags & ADAP_SYSCTL_CTX) { 800220873Snp sysctl_ctx_free(&sc->ctx); 801228561Snp sc->flags &= ~ADAP_SYSCTL_CTX; 802220873Snp } 803220873Snp 804228561Snp free_mgmtq(sc); 805228561Snp free_fwq(sc); 806220873Snp 807228561Snp return (0); 808228561Snp} 809222510Snp 810228561Snpstatic inline int 811228561Snpfirst_vector(struct port_info *pi) 812228561Snp{ 813228561Snp struct adapter *sc = pi->adapter; 814228561Snp int rc = T4_EXTRA_INTR, i; 815228561Snp 816228561Snp if (sc->intr_count == 1) 817228561Snp return (0); 818228561Snp 819228561Snp for_each_port(sc, i) { 820238313Snp struct port_info *p = sc->port[i]; 821238313Snp 822228561Snp if (i == pi->port_id) 823228561Snp break; 824228561Snp 825237263Snp#ifdef TCP_OFFLOAD 826228561Snp if (sc->flags & INTR_DIRECT) 827238313Snp rc += p->nrxq + p->nofldrxq; 828228561Snp else 829238313Snp rc += max(p->nrxq, p->nofldrxq); 830228561Snp#else 831228561Snp /* 832228561Snp * Not compiled with offload support and intr_count > 1. Only 833228561Snp * NIC queues exist and they'd better be taking direct 834228561Snp * interrupts. 835228561Snp */ 836228561Snp KASSERT(sc->flags & INTR_DIRECT, 837228561Snp ("%s: intr_count %d, !INTR_DIRECT", __func__, 838228561Snp sc->intr_count)); 839228561Snp 840238313Snp rc += p->nrxq; 841228561Snp#endif 842218792Snp } 843218792Snp 844228561Snp return (rc); 845218792Snp} 846218792Snp 847228561Snp/* 848228561Snp * Given an arbitrary "index," come up with an iq that can be used by other 849228561Snp * queues (of this port) for interrupt forwarding, SGE egress updates, etc. 850228561Snp * The iq returned is guaranteed to be something that takes direct interrupts. 851228561Snp */ 852228561Snpstatic struct sge_iq * 853228561Snpport_intr_iq(struct port_info *pi, int idx) 854228561Snp{ 855228561Snp struct adapter *sc = pi->adapter; 856228561Snp struct sge *s = &sc->sge; 857228561Snp struct sge_iq *iq = NULL; 858228561Snp 859228561Snp if (sc->intr_count == 1) 860228561Snp return (&sc->sge.fwq); 861228561Snp 862237263Snp#ifdef TCP_OFFLOAD 863228561Snp if (sc->flags & INTR_DIRECT) { 864228561Snp idx %= pi->nrxq + pi->nofldrxq; 865265425Snp 866228561Snp if (idx >= pi->nrxq) { 867228561Snp idx -= pi->nrxq; 868228561Snp iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq; 869228561Snp } else 870228561Snp iq = &s->rxq[pi->first_rxq + idx].iq; 871228561Snp 872228561Snp } else { 873228561Snp idx %= max(pi->nrxq, pi->nofldrxq); 874228561Snp 875228561Snp if (pi->nrxq >= pi->nofldrxq) 876228561Snp iq = &s->rxq[pi->first_rxq + idx].iq; 877228561Snp else 878228561Snp iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq; 879228561Snp } 880228561Snp#else 881228561Snp /* 882228561Snp * Not compiled with offload support and intr_count > 1. Only NIC 883228561Snp * queues exist and they'd better be taking direct interrupts. 884228561Snp */ 885228561Snp KASSERT(sc->flags & INTR_DIRECT, 886228561Snp ("%s: intr_count %d, !INTR_DIRECT", __func__, sc->intr_count)); 887228561Snp 888228561Snp idx %= pi->nrxq; 889228561Snp iq = &s->rxq[pi->first_rxq + idx].iq; 890228561Snp#endif 891228561Snp 892228561Snp KASSERT(iq->flags & IQ_INTR, ("%s: EDOOFUS", __func__)); 893228561Snp return (iq); 894228561Snp} 895228561Snp 896265425Snp/* Maximum payload that can be delivered with a single iq descriptor */ 897239266Snpstatic inline int 898265425Snpmtu_to_max_payload(struct adapter *sc, int mtu, const int toe) 899239266Snp{ 900265425Snp int payload; 901239266Snp 902252728Snp#ifdef TCP_OFFLOAD 903265425Snp if (toe) { 904265425Snp payload = sc->tt.rx_coalesce ? 905265425Snp G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)) : mtu; 906265425Snp } else { 907265425Snp#endif 908265425Snp /* large enough even when hw VLAN extraction is disabled */ 909265425Snp payload = fl_pktshift + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN + 910265425Snp mtu; 911265425Snp#ifdef TCP_OFFLOAD 912265425Snp } 913265425Snp#endif 914265425Snp payload = roundup2(payload, fl_pad); 915252728Snp 916265425Snp return (payload); 917252728Snp} 918252728Snp 919218792Snpint 920228561Snpt4_setup_port_queues(struct port_info *pi) 921218792Snp{ 922228561Snp int rc = 0, i, j, intr_idx, iqid; 923218792Snp struct sge_rxq *rxq; 924218792Snp struct sge_txq *txq; 925228561Snp struct sge_wrq *ctrlq; 926237263Snp#ifdef TCP_OFFLOAD 927228561Snp struct sge_ofld_rxq *ofld_rxq; 928228561Snp struct sge_wrq *ofld_txq; 929237263Snp struct sysctl_oid *oid2 = NULL; 930228561Snp#endif 931218792Snp char name[16]; 932218792Snp struct adapter *sc = pi->adapter; 933252728Snp struct ifnet *ifp = pi->ifp; 934237263Snp struct sysctl_oid *oid = device_get_sysctl_tree(pi->dev); 935228561Snp struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); 936265425Snp int maxp, pack, mtu = ifp->if_mtu; 937218792Snp 938228561Snp oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "rxq", CTLFLAG_RD, 939228561Snp NULL, "rx queues"); 940218792Snp 941237263Snp#ifdef TCP_OFFLOAD 942228561Snp if (is_offload(sc)) { 943228561Snp oid2 = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_rxq", 944228561Snp CTLFLAG_RD, NULL, 945228561Snp "rx queues for offloaded TCP connections"); 946218792Snp } 947228561Snp#endif 948218792Snp 949228561Snp /* Interrupt vector to start from (when using multiple vectors) */ 950228561Snp intr_idx = first_vector(pi); 951228561Snp 952228561Snp /* 953228561Snp * First pass over all rx queues (NIC and TOE): 954228561Snp * a) initialize iq and fl 955228561Snp * b) allocate queue iff it will take direct interrupts. 956228561Snp */ 957265425Snp maxp = mtu_to_max_payload(sc, mtu, 0); 958255052Snp pack = enable_buffer_packing(sc); 959218792Snp for_each_rxq(pi, i, rxq) { 960218792Snp 961228561Snp init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, pi->qsize_rxq, 962241397Snp RX_IQ_ESIZE); 963218792Snp 964218792Snp snprintf(name, sizeof(name), "%s rxq%d-fl", 965218792Snp device_get_nameunit(pi->dev), i); 966265425Snp init_fl(sc, &rxq->fl, pi->qsize_rxq / 8, maxp, pack, name); 967218792Snp 968228561Snp if (sc->flags & INTR_DIRECT 969237263Snp#ifdef TCP_OFFLOAD 970228561Snp || (sc->intr_count > 1 && pi->nrxq >= pi->nofldrxq) 971228561Snp#endif 972228561Snp ) { 973228561Snp rxq->iq.flags |= IQ_INTR; 974228561Snp rc = alloc_rxq(pi, rxq, intr_idx, i, oid); 975228561Snp if (rc != 0) 976228561Snp goto done; 977228561Snp intr_idx++; 978228561Snp } 979228561Snp } 980218792Snp 981237263Snp#ifdef TCP_OFFLOAD 982265425Snp maxp = mtu_to_max_payload(sc, mtu, 1); 983228561Snp for_each_ofld_rxq(pi, i, ofld_rxq) { 984228561Snp 985228561Snp init_iq(&ofld_rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, 986241397Snp pi->qsize_rxq, RX_IQ_ESIZE); 987228561Snp 988228561Snp snprintf(name, sizeof(name), "%s ofld_rxq%d-fl", 989228561Snp device_get_nameunit(pi->dev), i); 990265425Snp init_fl(sc, &ofld_rxq->fl, pi->qsize_rxq / 8, maxp, pack, name); 991228561Snp 992228561Snp if (sc->flags & INTR_DIRECT || 993228561Snp (sc->intr_count > 1 && pi->nofldrxq > pi->nrxq)) { 994228561Snp ofld_rxq->iq.flags |= IQ_INTR; 995228561Snp rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx, i, oid2); 996228561Snp if (rc != 0) 997228561Snp goto done; 998228561Snp intr_idx++; 999228561Snp } 1000228561Snp } 1001228561Snp#endif 1002228561Snp 1003228561Snp /* 1004228561Snp * Second pass over all rx queues (NIC and TOE). The queues forwarding 1005228561Snp * their interrupts are allocated now. 1006228561Snp */ 1007228561Snp j = 0; 1008228561Snp for_each_rxq(pi, i, rxq) { 1009228561Snp if (rxq->iq.flags & IQ_INTR) 1010228561Snp continue; 1011228561Snp 1012228561Snp intr_idx = port_intr_iq(pi, j)->abs_id; 1013228561Snp 1014228561Snp rc = alloc_rxq(pi, rxq, intr_idx, i, oid); 1015218792Snp if (rc != 0) 1016218792Snp goto done; 1017228561Snp j++; 1018218792Snp } 1019218792Snp 1020237263Snp#ifdef TCP_OFFLOAD 1021228561Snp for_each_ofld_rxq(pi, i, ofld_rxq) { 1022228561Snp if (ofld_rxq->iq.flags & IQ_INTR) 1023228561Snp continue; 1024228561Snp 1025228561Snp intr_idx = port_intr_iq(pi, j)->abs_id; 1026228561Snp 1027228561Snp rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx, i, oid2); 1028228561Snp if (rc != 0) 1029228561Snp goto done; 1030228561Snp j++; 1031228561Snp } 1032228561Snp#endif 1033228561Snp 1034228561Snp /* 1035228561Snp * Now the tx queues. Only one pass needed. 1036228561Snp */ 1037228561Snp oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD, 1038228561Snp NULL, "tx queues"); 1039228561Snp j = 0; 1040218792Snp for_each_txq(pi, i, txq) { 1041228561Snp uint16_t iqid; 1042218792Snp 1043228561Snp iqid = port_intr_iq(pi, j)->cntxt_id; 1044228561Snp 1045218792Snp snprintf(name, sizeof(name), "%s txq%d", 1046218792Snp device_get_nameunit(pi->dev), i); 1047228561Snp init_eq(&txq->eq, EQ_ETH, pi->qsize_txq, pi->tx_chan, iqid, 1048228561Snp name); 1049218792Snp 1050228561Snp rc = alloc_txq(pi, txq, i, oid); 1051218792Snp if (rc != 0) 1052218792Snp goto done; 1053228561Snp j++; 1054218792Snp } 1055218792Snp 1056237263Snp#ifdef TCP_OFFLOAD 1057228561Snp oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_txq", 1058228561Snp CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections"); 1059228561Snp for_each_ofld_txq(pi, i, ofld_txq) { 1060228561Snp uint16_t iqid; 1061228561Snp 1062228561Snp iqid = port_intr_iq(pi, j)->cntxt_id; 1063228561Snp 1064228561Snp snprintf(name, sizeof(name), "%s ofld_txq%d", 1065228561Snp device_get_nameunit(pi->dev), i); 1066228561Snp init_eq(&ofld_txq->eq, EQ_OFLD, pi->qsize_txq, pi->tx_chan, 1067228561Snp iqid, name); 1068228561Snp 1069228561Snp snprintf(name, sizeof(name), "%d", i); 1070228561Snp oid2 = SYSCTL_ADD_NODE(&pi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO, 1071228561Snp name, CTLFLAG_RD, NULL, "offload tx queue"); 1072228561Snp 1073228561Snp rc = alloc_wrq(sc, pi, ofld_txq, oid2); 1074228561Snp if (rc != 0) 1075228561Snp goto done; 1076228561Snp j++; 1077228561Snp } 1078228561Snp#endif 1079228561Snp 1080228561Snp /* 1081228561Snp * Finally, the control queue. 1082228561Snp */ 1083228561Snp oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD, 1084228561Snp NULL, "ctrl queue"); 1085228561Snp ctrlq = &sc->sge.ctrlq[pi->port_id]; 1086228561Snp iqid = port_intr_iq(pi, 0)->cntxt_id; 1087228561Snp snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(pi->dev)); 1088228561Snp init_eq(&ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid, name); 1089228561Snp rc = alloc_wrq(sc, pi, ctrlq, oid); 1090228561Snp 1091218792Snpdone: 1092218792Snp if (rc) 1093228561Snp t4_teardown_port_queues(pi); 1094218792Snp 1095218792Snp return (rc); 1096218792Snp} 1097218792Snp 1098218792Snp/* 1099218792Snp * Idempotent 1100218792Snp */ 1101218792Snpint 1102228561Snpt4_teardown_port_queues(struct port_info *pi) 1103218792Snp{ 1104218792Snp int i; 1105228561Snp struct adapter *sc = pi->adapter; 1106218792Snp struct sge_rxq *rxq; 1107218792Snp struct sge_txq *txq; 1108237263Snp#ifdef TCP_OFFLOAD 1109228561Snp struct sge_ofld_rxq *ofld_rxq; 1110228561Snp struct sge_wrq *ofld_txq; 1111228561Snp#endif 1112218792Snp 1113218792Snp /* Do this before freeing the queues */ 1114228561Snp if (pi->flags & PORT_SYSCTL_CTX) { 1115218792Snp sysctl_ctx_free(&pi->ctx); 1116228561Snp pi->flags &= ~PORT_SYSCTL_CTX; 1117218792Snp } 1118218792Snp 1119228561Snp /* 1120228561Snp * Take down all the tx queues first, as they reference the rx queues 1121228561Snp * (for egress updates, etc.). 1122228561Snp */ 1123228561Snp 1124228561Snp free_wrq(sc, &sc->sge.ctrlq[pi->port_id]); 1125228561Snp 1126218792Snp for_each_txq(pi, i, txq) { 1127218792Snp free_txq(pi, txq); 1128218792Snp } 1129218792Snp 1130237263Snp#ifdef TCP_OFFLOAD 1131228561Snp for_each_ofld_txq(pi, i, ofld_txq) { 1132228561Snp free_wrq(sc, ofld_txq); 1133228561Snp } 1134228561Snp#endif 1135228561Snp 1136228561Snp /* 1137228561Snp * Then take down the rx queues that forward their interrupts, as they 1138228561Snp * reference other rx queues. 1139228561Snp */ 1140228561Snp 1141218792Snp for_each_rxq(pi, i, rxq) { 1142228561Snp if ((rxq->iq.flags & IQ_INTR) == 0) 1143228561Snp free_rxq(pi, rxq); 1144218792Snp } 1145218792Snp 1146237263Snp#ifdef TCP_OFFLOAD 1147228561Snp for_each_ofld_rxq(pi, i, ofld_rxq) { 1148228561Snp if ((ofld_rxq->iq.flags & IQ_INTR) == 0) 1149228561Snp free_ofld_rxq(pi, ofld_rxq); 1150228561Snp } 1151228561Snp#endif 1152228561Snp 1153228561Snp /* 1154228561Snp * Then take down the rx queues that take direct interrupts. 1155228561Snp */ 1156228561Snp 1157228561Snp for_each_rxq(pi, i, rxq) { 1158228561Snp if (rxq->iq.flags & IQ_INTR) 1159228561Snp free_rxq(pi, rxq); 1160228561Snp } 1161228561Snp 1162237263Snp#ifdef TCP_OFFLOAD 1163228561Snp for_each_ofld_rxq(pi, i, ofld_rxq) { 1164228561Snp if (ofld_rxq->iq.flags & IQ_INTR) 1165228561Snp free_ofld_rxq(pi, ofld_rxq); 1166228561Snp } 1167228561Snp#endif 1168228561Snp 1169218792Snp return (0); 1170218792Snp} 1171218792Snp 1172228561Snp/* 1173228561Snp * Deals with errors and the firmware event queue. All data rx queues forward 1174228561Snp * their interrupt to the firmware event queue. 1175228561Snp */ 1176218792Snpvoid 1177218792Snpt4_intr_all(void *arg) 1178218792Snp{ 1179218792Snp struct adapter *sc = arg; 1180228561Snp struct sge_iq *fwq = &sc->sge.fwq; 1181218792Snp 1182218792Snp t4_intr_err(arg); 1183228561Snp if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) { 1184228561Snp service_iq(fwq, 0); 1185228561Snp atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE); 1186218792Snp } 1187218792Snp} 1188218792Snp 1189218792Snp/* Deals with error interrupts */ 1190218792Snpvoid 1191218792Snpt4_intr_err(void *arg) 1192218792Snp{ 1193218792Snp struct adapter *sc = arg; 1194218792Snp 1195222510Snp t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0); 1196218792Snp t4_slow_intr_handler(sc); 1197218792Snp} 1198218792Snp 1199218792Snpvoid 1200218792Snpt4_intr_evt(void *arg) 1201218792Snp{ 1202218792Snp struct sge_iq *iq = arg; 1203220649Snp 1204228561Snp if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) { 1205228561Snp service_iq(iq, 0); 1206228561Snp atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE); 1207222510Snp } 1208220649Snp} 1209220649Snp 1210228561Snpvoid 1211228561Snpt4_intr(void *arg) 1212220649Snp{ 1213220649Snp struct sge_iq *iq = arg; 1214228561Snp 1215228561Snp if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) { 1216228561Snp service_iq(iq, 0); 1217228561Snp atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE); 1218228561Snp } 1219228561Snp} 1220228561Snp 1221228561Snp/* 1222228561Snp * Deals with anything and everything on the given ingress queue. 1223228561Snp */ 1224228561Snpstatic int 1225228561Snpservice_iq(struct sge_iq *iq, int budget) 1226228561Snp{ 1227228561Snp struct sge_iq *q; 1228237263Snp struct sge_rxq *rxq = iq_to_rxq(iq); /* Use iff iq is part of rxq */ 1229228561Snp struct sge_fl *fl = &rxq->fl; /* Use iff IQ_HAS_FL */ 1230218792Snp struct adapter *sc = iq->adapter; 1231218792Snp struct rsp_ctrl *ctrl; 1232228561Snp const struct rss_header *rss; 1233228561Snp int ndescs = 0, limit, fl_bufs_used = 0; 1234228561Snp int rsp_type; 1235228561Snp uint32_t lq; 1236228561Snp struct mbuf *m0; 1237228561Snp STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql); 1238255015Snp#if defined(INET) || defined(INET6) 1239255015Snp const struct timeval lro_timeout = {0, sc->lro_timeout}; 1240255015Snp#endif 1241218792Snp 1242228561Snp limit = budget ? budget : iq->qsize / 8; 1243218792Snp 1244228561Snp KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq)); 1245218792Snp 1246228561Snp /* 1247228561Snp * We always come back and check the descriptor ring for new indirect 1248228561Snp * interrupts and other responses after running a single handler. 1249228561Snp */ 1250228561Snp for (;;) { 1251228561Snp while (is_new_response(iq, &ctrl)) { 1252218792Snp 1253228561Snp rmb(); 1254218792Snp 1255228561Snp m0 = NULL; 1256228561Snp rsp_type = G_RSPD_TYPE(ctrl->u.type_gen); 1257228561Snp lq = be32toh(ctrl->pldbuflen_qid); 1258228561Snp rss = (const void *)iq->cdesc; 1259218792Snp 1260228561Snp switch (rsp_type) { 1261228561Snp case X_RSPD_TYPE_FLBUF: 1262228561Snp 1263228561Snp KASSERT(iq->flags & IQ_HAS_FL, 1264228561Snp ("%s: data for an iq (%p) with no freelist", 1265228561Snp __func__, iq)); 1266228561Snp 1267265425Snp m0 = get_fl_payload(sc, fl, lq, &fl_bufs_used); 1268255050Snp if (__predict_false(m0 == NULL)) 1269255050Snp goto process_iql; 1270228561Snp#ifdef T4_PKT_TIMESTAMP 1271228561Snp /* 1272228561Snp * 60 bit timestamp for the payload is 1273228561Snp * *(uint64_t *)m0->m_pktdat. Note that it is 1274228561Snp * in the leading free-space in the mbuf. The 1275228561Snp * kernel can clobber it during a pullup, 1276228561Snp * m_copymdata, etc. You need to make sure that 1277228561Snp * the mbuf reaches you unmolested if you care 1278228561Snp * about the timestamp. 1279228561Snp */ 1280228561Snp *(uint64_t *)m0->m_pktdat = 1281228561Snp be64toh(ctrl->u.last_flit) & 1282228561Snp 0xfffffffffffffff; 1283228561Snp#endif 1284228561Snp 1285228561Snp /* fall through */ 1286228561Snp 1287228561Snp case X_RSPD_TYPE_CPL: 1288228561Snp KASSERT(rss->opcode < NUM_CPL_CMDS, 1289228561Snp ("%s: bad opcode %02x.", __func__, 1290228561Snp rss->opcode)); 1291228561Snp sc->cpl_handler[rss->opcode](iq, rss, m0); 1292228561Snp break; 1293228561Snp 1294228561Snp case X_RSPD_TYPE_INTR: 1295228561Snp 1296228561Snp /* 1297228561Snp * Interrupts should be forwarded only to queues 1298228561Snp * that are not forwarding their interrupts. 1299228561Snp * This means service_iq can recurse but only 1 1300228561Snp * level deep. 1301228561Snp */ 1302228561Snp KASSERT(budget == 0, 1303228561Snp ("%s: budget %u, rsp_type %u", __func__, 1304228561Snp budget, rsp_type)); 1305228561Snp 1306255005Snp /* 1307255005Snp * There are 1K interrupt-capable queues (qids 0 1308255005Snp * through 1023). A response type indicating a 1309255005Snp * forwarded interrupt with a qid >= 1K is an 1310255005Snp * iWARP async notification. 1311255005Snp */ 1312255005Snp if (lq >= 1024) { 1313255005Snp sc->an_handler(iq, ctrl); 1314255005Snp break; 1315255005Snp } 1316255005Snp 1317228561Snp q = sc->sge.iqmap[lq - sc->sge.iq_start]; 1318228561Snp if (atomic_cmpset_int(&q->state, IQS_IDLE, 1319228561Snp IQS_BUSY)) { 1320228561Snp if (service_iq(q, q->qsize / 8) == 0) { 1321228561Snp atomic_cmpset_int(&q->state, 1322228561Snp IQS_BUSY, IQS_IDLE); 1323228561Snp } else { 1324228561Snp STAILQ_INSERT_TAIL(&iql, q, 1325228561Snp link); 1326228561Snp } 1327228561Snp } 1328228561Snp break; 1329228561Snp 1330228561Snp default: 1331255005Snp KASSERT(0, 1332255005Snp ("%s: illegal response type %d on iq %p", 1333255005Snp __func__, rsp_type, iq)); 1334255005Snp log(LOG_ERR, 1335255005Snp "%s: illegal response type %d on iq %p", 1336255005Snp device_get_nameunit(sc->dev), rsp_type, iq); 1337237263Snp break; 1338228561Snp } 1339228561Snp 1340265425Snp if (fl_bufs_used >= 16) { 1341265425Snp FL_LOCK(fl); 1342265425Snp fl->needed += fl_bufs_used; 1343265425Snp refill_fl(sc, fl, 32); 1344265425Snp FL_UNLOCK(fl); 1345265425Snp fl_bufs_used = 0; 1346265425Snp } 1347265425Snp 1348228561Snp iq_next(iq); 1349228561Snp if (++ndescs == limit) { 1350228561Snp t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), 1351228561Snp V_CIDXINC(ndescs) | 1352228561Snp V_INGRESSQID(iq->cntxt_id) | 1353228561Snp V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX))); 1354228561Snp ndescs = 0; 1355228561Snp 1356255015Snp#if defined(INET) || defined(INET6) 1357255015Snp if (iq->flags & IQ_LRO_ENABLED && 1358255015Snp sc->lro_timeout != 0) { 1359255015Snp tcp_lro_flush_inactive(&rxq->lro, 1360255015Snp &lro_timeout); 1361255015Snp } 1362255015Snp#endif 1363255015Snp 1364228561Snp if (budget) 1365228561Snp return (EINPROGRESS); 1366228561Snp } 1367218792Snp } 1368222510Snp 1369255050Snpprocess_iql: 1370228561Snp if (STAILQ_EMPTY(&iql)) 1371228561Snp break; 1372228561Snp 1373228561Snp /* 1374228561Snp * Process the head only, and send it to the back of the list if 1375228561Snp * it's still not done. 1376228561Snp */ 1377228561Snp q = STAILQ_FIRST(&iql); 1378228561Snp STAILQ_REMOVE_HEAD(&iql, link); 1379228561Snp if (service_iq(q, q->qsize / 8) == 0) 1380228561Snp atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE); 1381228561Snp else 1382228561Snp STAILQ_INSERT_TAIL(&iql, q, link); 1383218792Snp } 1384218792Snp 1385237819Snp#if defined(INET) || defined(INET6) 1386228561Snp if (iq->flags & IQ_LRO_ENABLED) { 1387228561Snp struct lro_ctrl *lro = &rxq->lro; 1388228561Snp struct lro_entry *l; 1389228561Snp 1390228561Snp while (!SLIST_EMPTY(&lro->lro_active)) { 1391228561Snp l = SLIST_FIRST(&lro->lro_active); 1392228561Snp SLIST_REMOVE_HEAD(&lro->lro_active, next); 1393228561Snp tcp_lro_flush(lro, l); 1394228561Snp } 1395228561Snp } 1396228561Snp#endif 1397228561Snp 1398228561Snp t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) | 1399228561Snp V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params)); 1400228561Snp 1401228561Snp if (iq->flags & IQ_HAS_FL) { 1402228561Snp int starved; 1403228561Snp 1404228561Snp FL_LOCK(fl); 1405228561Snp fl->needed += fl_bufs_used; 1406265425Snp starved = refill_fl(sc, fl, 64); 1407228561Snp FL_UNLOCK(fl); 1408228561Snp if (__predict_false(starved != 0)) 1409228561Snp add_fl_to_sfl(sc, fl); 1410228561Snp } 1411228561Snp 1412228561Snp return (0); 1413218792Snp} 1414218792Snp 1415265425Snpstatic inline int 1416265425Snpcl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll) 1417255050Snp{ 1418265425Snp int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0; 1419255050Snp 1420265425Snp if (rc) 1421265425Snp MPASS(cll->region3 >= CL_METADATA_SIZE); 1422255050Snp 1423265425Snp return (rc); 1424255050Snp} 1425255050Snp 1426265425Snpstatic inline struct cluster_metadata * 1427265425Snpcl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll, 1428265425Snp caddr_t cl) 1429255050Snp{ 1430255050Snp 1431265425Snp if (cl_has_metadata(fl, cll)) { 1432265425Snp struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx]; 1433255050Snp 1434265425Snp return ((struct cluster_metadata *)(cl + swz->size) - 1); 1435255050Snp } 1436265425Snp return (NULL); 1437255050Snp} 1438255050Snp 1439255050Snpstatic int 1440255050Snprxb_free(struct mbuf *m, void *arg1, void *arg2) 1441255050Snp{ 1442255050Snp uma_zone_t zone = arg1; 1443255050Snp caddr_t cl = arg2; 1444255050Snp 1445255050Snp uma_zfree(zone, cl); 1446255050Snp 1447255050Snp return (EXT_FREE_OK); 1448255050Snp} 1449255050Snp 1450265425Snp/* 1451265425Snp * The mbuf returned by this function could be allocated from zone_mbuf or 1452265425Snp * constructed in spare room in the cluster. 1453265425Snp * 1454265425Snp * The mbuf carries the payload in one of these ways 1455265425Snp * a) frame inside the mbuf (mbuf from zone_mbuf) 1456265425Snp * b) m_cljset (for clusters without metadata) zone_mbuf 1457265425Snp * c) m_extaddref (cluster with metadata) inline mbuf 1458265425Snp * d) m_extaddref (cluster with metadata) zone_mbuf 1459265425Snp */ 1460255050Snpstatic struct mbuf * 1461265425Snpget_scatter_segment(struct adapter *sc, struct sge_fl *fl, int total, int flags) 1462218792Snp{ 1463265425Snp struct mbuf *m; 1464228561Snp struct fl_sdesc *sd = &fl->sdesc[fl->cidx]; 1465265425Snp struct cluster_layout *cll = &sd->cll; 1466265425Snp struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx]; 1467265425Snp struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx]; 1468265425Snp struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl); 1469265425Snp int len, padded_len; 1470265425Snp caddr_t payload; 1471218792Snp 1472265425Snp len = min(total, hwb->size - fl->rx_offset); 1473265425Snp padded_len = roundup2(len, fl_pad); 1474265425Snp payload = sd->cl + cll->region1 + fl->rx_offset; 1475219290Snp 1476265425Snp if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) { 1477255050Snp 1478265425Snp /* 1479265425Snp * Copy payload into a freshly allocated mbuf. 1480265425Snp */ 1481255050Snp 1482265425Snp m = flags & M_PKTHDR ? 1483265425Snp m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA); 1484265425Snp if (m == NULL) 1485255050Snp return (NULL); 1486265425Snp fl->mbuf_allocated++; 1487255050Snp#ifdef T4_PKT_TIMESTAMP 1488265425Snp /* Leave room for a timestamp */ 1489265425Snp m->m_data += 8; 1490255050Snp#endif 1491265425Snp /* copy data to mbuf */ 1492265425Snp bcopy(payload, mtod(m, caddr_t), len); 1493255050Snp 1494265425Snp } else if (sd->nmbuf * MSIZE < cll->region1) { 1495255050Snp 1496265425Snp /* 1497265425Snp * There's spare room in the cluster for an mbuf. Create one 1498265425Snp * and associate it with the payload that's in the cluster too. 1499265425Snp */ 1500255050Snp 1501265425Snp MPASS(clm != NULL); 1502265425Snp m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE); 1503265425Snp /* No bzero required */ 1504265425Snp if (m_init(m, NULL, 0, M_NOWAIT, MT_DATA, flags | M_NOFREE)) 1505265425Snp return (NULL); 1506265425Snp fl->mbuf_inlined++; 1507265425Snp m_extaddref(m, payload, padded_len, &clm->refcount, rxb_free, 1508265425Snp swz->zone, sd->cl); 1509265425Snp sd->nmbuf++; 1510255050Snp 1511265425Snp } else { 1512255050Snp 1513265425Snp /* 1514265425Snp * Grab an mbuf from zone_mbuf and associate it with the 1515265425Snp * payload in the cluster. 1516265425Snp */ 1517255050Snp 1518265425Snp m = flags & M_PKTHDR ? 1519265425Snp m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA); 1520265425Snp if (m == NULL) 1521265425Snp return (NULL); 1522265425Snp fl->mbuf_allocated++; 1523265425Snp if (clm != NULL) 1524265425Snp m_extaddref(m, payload, padded_len, &clm->refcount, 1525265425Snp rxb_free, swz->zone, sd->cl); 1526265425Snp else { 1527265425Snp m_cljset(m, sd->cl, swz->type); 1528265425Snp sd->cl = NULL; /* consumed, not a recycle candidate */ 1529255050Snp } 1530255050Snp } 1531265425Snp if (flags & M_PKTHDR) 1532265425Snp m->m_pkthdr.len = total; 1533265425Snp m->m_len = len; 1534255050Snp 1535265425Snp if (fl->flags & FL_BUF_PACKING) { 1536265425Snp fl->rx_offset += roundup2(padded_len, sc->sge.pack_boundary); 1537265425Snp MPASS(fl->rx_offset <= hwb->size); 1538265425Snp if (fl->rx_offset < hwb->size) 1539265425Snp return (m); /* without advancing the cidx */ 1540265425Snp } 1541255050Snp 1542265425Snp if (__predict_false(++fl->cidx == fl->cap)) 1543265425Snp fl->cidx = 0; 1544265425Snp fl->rx_offset = 0; 1545255050Snp 1546265425Snp return (m); 1547255050Snp} 1548255050Snp 1549255050Snpstatic struct mbuf * 1550265425Snpget_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf, 1551255050Snp int *fl_bufs_used) 1552255050Snp{ 1553265425Snp struct mbuf *m0, *m, **pnext; 1554265425Snp u_int nbuf, len; 1555255050Snp 1556255050Snp /* 1557255050Snp * No assertion for the fl lock because we don't need it. This routine 1558255050Snp * is called only from the rx interrupt handler and it only updates 1559255050Snp * fl->cidx. (Contrast that with fl->pidx/fl->needed which could be 1560255050Snp * updated in the rx interrupt handler or the starvation helper routine. 1561255050Snp * That's why code that manipulates fl->pidx/fl->needed needs the fl 1562255050Snp * lock but this routine does not). 1563255050Snp */ 1564255050Snp 1565265425Snp nbuf = 0; 1566228561Snp len = G_RSPD_LEN(len_newbuf); 1567265425Snp if (__predict_false(fl->m0 != NULL)) { 1568266965Snp M_ASSERTPKTHDR(fl->m0); 1569265425Snp MPASS(len == fl->m0->m_pkthdr.len); 1570265425Snp MPASS(fl->remaining < len); 1571218792Snp 1572265425Snp m0 = fl->m0; 1573265425Snp pnext = fl->pnext; 1574265425Snp len = fl->remaining; 1575265425Snp fl->m0 = NULL; 1576265425Snp goto get_segment; 1577255050Snp } 1578255050Snp 1579265425Snp if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) { 1580265425Snp nbuf++; 1581265425Snp fl->rx_offset = 0; 1582265425Snp if (__predict_false(++fl->cidx == fl->cap)) 1583265425Snp fl->cidx = 0; 1584228561Snp } 1585218792Snp 1586265425Snp /* 1587265425Snp * Payload starts at rx_offset in the current hw buffer. Its length is 1588265425Snp * 'len' and it may span multiple hw buffers. 1589265425Snp */ 1590218792Snp 1591265425Snp m0 = get_scatter_segment(sc, fl, len, M_PKTHDR); 1592266965Snp if (m0 == NULL) 1593266965Snp goto done; 1594265425Snp len -= m0->m_len; 1595265425Snp pnext = &m0->m_next; 1596228561Snp while (len > 0) { 1597265425Snp nbuf++; 1598265425Snpget_segment: 1599265425Snp MPASS(fl->rx_offset == 0); 1600265425Snp m = get_scatter_segment(sc, fl, len, 0); 1601265425Snp if (m == NULL) { 1602265425Snp fl->m0 = m0; 1603265425Snp fl->pnext = pnext; 1604265425Snp fl->remaining = len; 1605266965Snp m0 = NULL; 1606266965Snp goto done; 1607218792Snp } 1608265425Snp *pnext = m; 1609265425Snp pnext = &m->m_next; 1610228561Snp len -= m->m_len; 1611265425Snp } 1612265425Snp *pnext = NULL; 1613265425Snp if (fl->rx_offset == 0) 1614228561Snp nbuf++; 1615266965Snpdone: 1616228561Snp (*fl_bufs_used) += nbuf; 1617228561Snp return (m0); 1618228561Snp} 1619218792Snp 1620228561Snpstatic int 1621228561Snpt4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0) 1622228561Snp{ 1623237463Snp struct sge_rxq *rxq = iq_to_rxq(iq); 1624228561Snp struct ifnet *ifp = rxq->ifp; 1625228561Snp const struct cpl_rx_pkt *cpl = (const void *)(rss + 1); 1626237819Snp#if defined(INET) || defined(INET6) 1627228561Snp struct lro_ctrl *lro = &rxq->lro; 1628228561Snp#endif 1629219290Snp 1630228561Snp KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__, 1631228561Snp rss->opcode)); 1632219290Snp 1633239258Snp m0->m_pkthdr.len -= fl_pktshift; 1634239258Snp m0->m_len -= fl_pktshift; 1635239258Snp m0->m_data += fl_pktshift; 1636219290Snp 1637228561Snp m0->m_pkthdr.rcvif = ifp; 1638228561Snp m0->m_flags |= M_FLOWID; 1639259142Snp m0->m_pkthdr.flowid = be32toh(rss->hash_val); 1640219290Snp 1641237799Snp if (cpl->csum_calc && !cpl->err_vec) { 1642237799Snp if (ifp->if_capenable & IFCAP_RXCSUM && 1643237799Snp cpl->l2info & htobe32(F_RXF_IP)) { 1644237831Snp m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED | 1645237799Snp CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR); 1646237799Snp rxq->rxcsum++; 1647237799Snp } else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 && 1648237799Snp cpl->l2info & htobe32(F_RXF_IP6)) { 1649237831Snp m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 | 1650237799Snp CSUM_PSEUDO_HDR); 1651237799Snp rxq->rxcsum++; 1652237799Snp } 1653237799Snp 1654237799Snp if (__predict_false(cpl->ip_frag)) 1655228561Snp m0->m_pkthdr.csum_data = be16toh(cpl->csum); 1656228561Snp else 1657228561Snp m0->m_pkthdr.csum_data = 0xffff; 1658228561Snp } 1659219290Snp 1660228561Snp if (cpl->vlan_ex) { 1661228561Snp m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan); 1662228561Snp m0->m_flags |= M_VLANTAG; 1663228561Snp rxq->vlan_extraction++; 1664228561Snp } 1665219290Snp 1666237819Snp#if defined(INET) || defined(INET6) 1667228561Snp if (cpl->l2info & htobe32(F_RXF_LRO) && 1668228561Snp iq->flags & IQ_LRO_ENABLED && 1669228561Snp tcp_lro_rx(lro, m0, 0) == 0) { 1670228561Snp /* queued for LRO */ 1671228561Snp } else 1672218792Snp#endif 1673228561Snp ifp->if_input(ifp, m0); 1674218792Snp 1675228561Snp return (0); 1676228561Snp} 1677218792Snp 1678228561Snp/* 1679228561Snp * Doesn't fail. Holds on to work requests it can't send right away. 1680228561Snp */ 1681237263Snpvoid 1682237263Snpt4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr) 1683228561Snp{ 1684228561Snp struct sge_eq *eq = &wrq->eq; 1685228561Snp int can_reclaim; 1686228561Snp caddr_t dst; 1687228561Snp 1688228561Snp TXQ_LOCK_ASSERT_OWNED(wrq); 1689237263Snp#ifdef TCP_OFFLOAD 1690228561Snp KASSERT((eq->flags & EQ_TYPEMASK) == EQ_OFLD || 1691228561Snp (eq->flags & EQ_TYPEMASK) == EQ_CTRL, 1692228561Snp ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK)); 1693237263Snp#else 1694237263Snp KASSERT((eq->flags & EQ_TYPEMASK) == EQ_CTRL, 1695237263Snp ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK)); 1696237263Snp#endif 1697228561Snp 1698237263Snp if (__predict_true(wr != NULL)) 1699237263Snp STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link); 1700218792Snp 1701228561Snp can_reclaim = reclaimable(eq); 1702228561Snp if (__predict_false(eq->flags & EQ_STALLED)) { 1703228561Snp if (can_reclaim < tx_resume_threshold(eq)) 1704237263Snp return; 1705228561Snp eq->flags &= ~EQ_STALLED; 1706228561Snp eq->unstalled++; 1707218792Snp } 1708228561Snp eq->cidx += can_reclaim; 1709228561Snp eq->avail += can_reclaim; 1710228561Snp if (__predict_false(eq->cidx >= eq->cap)) 1711228561Snp eq->cidx -= eq->cap; 1712228561Snp 1713237263Snp while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL) { 1714228561Snp int ndesc; 1715228561Snp 1716237263Snp if (__predict_false(wr->wr_len < 0 || 1717237263Snp wr->wr_len > SGE_MAX_WR_LEN || (wr->wr_len & 0x7))) { 1718228561Snp 1719228561Snp#ifdef INVARIANTS 1720237263Snp panic("%s: work request with length %d", __func__, 1721237263Snp wr->wr_len); 1722237263Snp#endif 1723237263Snp#ifdef KDB 1724237263Snp kdb_backtrace(); 1725237263Snp#endif 1726237263Snp log(LOG_ERR, "%s: %s work request with length %d", 1727237263Snp device_get_nameunit(sc->dev), __func__, wr->wr_len); 1728237263Snp STAILQ_REMOVE_HEAD(&wrq->wr_list, link); 1729237263Snp free_wrqe(wr); 1730228561Snp continue; 1731228561Snp } 1732218792Snp 1733237263Snp ndesc = howmany(wr->wr_len, EQ_ESIZE); 1734228561Snp if (eq->avail < ndesc) { 1735228561Snp wrq->no_desc++; 1736228561Snp break; 1737228561Snp } 1738218792Snp 1739228561Snp dst = (void *)&eq->desc[eq->pidx]; 1740237263Snp copy_to_txd(eq, wrtod(wr), &dst, wr->wr_len); 1741218792Snp 1742228561Snp eq->pidx += ndesc; 1743228561Snp eq->avail -= ndesc; 1744228561Snp if (__predict_false(eq->pidx >= eq->cap)) 1745228561Snp eq->pidx -= eq->cap; 1746228561Snp 1747228561Snp eq->pending += ndesc; 1748252715Snp if (eq->pending >= 8) 1749228561Snp ring_eq_db(sc, eq); 1750228561Snp 1751228561Snp wrq->tx_wrs++; 1752237263Snp STAILQ_REMOVE_HEAD(&wrq->wr_list, link); 1753237263Snp free_wrqe(wr); 1754228561Snp 1755228561Snp if (eq->avail < 8) { 1756228561Snp can_reclaim = reclaimable(eq); 1757228561Snp eq->cidx += can_reclaim; 1758228561Snp eq->avail += can_reclaim; 1759228561Snp if (__predict_false(eq->cidx >= eq->cap)) 1760228561Snp eq->cidx -= eq->cap; 1761228561Snp } 1762228561Snp } 1763228561Snp 1764228561Snp if (eq->pending) 1765228561Snp ring_eq_db(sc, eq); 1766228561Snp 1767237263Snp if (wr != NULL) { 1768228561Snp eq->flags |= EQ_STALLED; 1769228561Snp if (callout_pending(&eq->tx_callout) == 0) 1770228561Snp callout_reset(&eq->tx_callout, 1, t4_tx_callout, eq); 1771228561Snp } 1772220873Snp} 1773220873Snp 1774218792Snp/* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */ 1775218792Snp#define TXPKTS_PKT_HDR ((\ 1776218792Snp sizeof(struct ulp_txpkt) + \ 1777218792Snp sizeof(struct ulptx_idata) + \ 1778218792Snp sizeof(struct cpl_tx_pkt_core) \ 1779218792Snp ) / 8) 1780218792Snp 1781218792Snp/* Header of a coalesced tx WR, before SGL of first packet (in flits) */ 1782218792Snp#define TXPKTS_WR_HDR (\ 1783218792Snp sizeof(struct fw_eth_tx_pkts_wr) / 8 + \ 1784218792Snp TXPKTS_PKT_HDR) 1785218792Snp 1786218792Snp/* Header of a tx WR, before SGL of first packet (in flits) */ 1787218792Snp#define TXPKT_WR_HDR ((\ 1788218792Snp sizeof(struct fw_eth_tx_pkt_wr) + \ 1789218792Snp sizeof(struct cpl_tx_pkt_core) \ 1790218792Snp ) / 8 ) 1791218792Snp 1792218792Snp/* Header of a tx LSO WR, before SGL of first packet (in flits) */ 1793218792Snp#define TXPKT_LSO_WR_HDR ((\ 1794218792Snp sizeof(struct fw_eth_tx_pkt_wr) + \ 1795237436Snp sizeof(struct cpl_tx_pkt_lso_core) + \ 1796218792Snp sizeof(struct cpl_tx_pkt_core) \ 1797218792Snp ) / 8 ) 1798218792Snp 1799218792Snpint 1800218792Snpt4_eth_tx(struct ifnet *ifp, struct sge_txq *txq, struct mbuf *m) 1801218792Snp{ 1802218792Snp struct port_info *pi = (void *)ifp->if_softc; 1803218792Snp struct adapter *sc = pi->adapter; 1804218792Snp struct sge_eq *eq = &txq->eq; 1805220873Snp struct buf_ring *br = txq->br; 1806218792Snp struct mbuf *next; 1807219292Snp int rc, coalescing, can_reclaim; 1808218792Snp struct txpkts txpkts; 1809218792Snp struct sgl sgl; 1810218792Snp 1811218792Snp TXQ_LOCK_ASSERT_OWNED(txq); 1812218792Snp KASSERT(m, ("%s: called with nothing to do.", __func__)); 1813228561Snp KASSERT((eq->flags & EQ_TYPEMASK) == EQ_ETH, 1814228561Snp ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK)); 1815218792Snp 1816219292Snp prefetch(&eq->desc[eq->pidx]); 1817220873Snp prefetch(&txq->sdesc[eq->pidx]); 1818219292Snp 1819218792Snp txpkts.npkt = 0;/* indicates there's nothing in txpkts */ 1820218792Snp coalescing = 0; 1821218792Snp 1822228561Snp can_reclaim = reclaimable(eq); 1823228561Snp if (__predict_false(eq->flags & EQ_STALLED)) { 1824228561Snp if (can_reclaim < tx_resume_threshold(eq)) { 1825228561Snp txq->m = m; 1826228561Snp return (0); 1827228561Snp } 1828228561Snp eq->flags &= ~EQ_STALLED; 1829228561Snp eq->unstalled++; 1830228561Snp } 1831218792Snp 1832228561Snp if (__predict_false(eq->flags & EQ_DOOMED)) { 1833228561Snp m_freem(m); 1834228561Snp while ((m = buf_ring_dequeue_sc(txq->br)) != NULL) 1835228561Snp m_freem(m); 1836228561Snp return (ENETDOWN); 1837228561Snp } 1838228561Snp 1839228561Snp if (eq->avail < 8 && can_reclaim) 1840228561Snp reclaim_tx_descs(txq, can_reclaim, 32); 1841228561Snp 1842218792Snp for (; m; m = next ? next : drbr_dequeue(ifp, br)) { 1843218792Snp 1844218792Snp if (eq->avail < 8) 1845218792Snp break; 1846218792Snp 1847218792Snp next = m->m_nextpkt; 1848218792Snp m->m_nextpkt = NULL; 1849218792Snp 1850218792Snp if (next || buf_ring_peek(br)) 1851218792Snp coalescing = 1; 1852218792Snp 1853218792Snp rc = get_pkt_sgl(txq, &m, &sgl, coalescing); 1854218792Snp if (rc != 0) { 1855218792Snp if (rc == ENOMEM) { 1856218792Snp 1857218792Snp /* Short of resources, suspend tx */ 1858218792Snp 1859218792Snp m->m_nextpkt = next; 1860218792Snp break; 1861218792Snp } 1862218792Snp 1863218792Snp /* 1864218792Snp * Unrecoverable error for this packet, throw it away 1865218792Snp * and move on to the next. get_pkt_sgl may already 1866218792Snp * have freed m (it will be NULL in that case and the 1867218792Snp * m_freem here is still safe). 1868218792Snp */ 1869218792Snp 1870218792Snp m_freem(m); 1871218792Snp continue; 1872218792Snp } 1873218792Snp 1874218792Snp if (coalescing && 1875218792Snp add_to_txpkts(pi, txq, &txpkts, m, &sgl) == 0) { 1876218792Snp 1877218792Snp /* Successfully absorbed into txpkts */ 1878218792Snp 1879218792Snp write_ulp_cpl_sgl(pi, txq, &txpkts, m, &sgl); 1880218792Snp goto doorbell; 1881218792Snp } 1882218792Snp 1883218792Snp /* 1884218792Snp * We weren't coalescing to begin with, or current frame could 1885218792Snp * not be coalesced (add_to_txpkts flushes txpkts if a frame 1886218792Snp * given to it can't be coalesced). Either way there should be 1887218792Snp * nothing in txpkts. 1888218792Snp */ 1889218792Snp KASSERT(txpkts.npkt == 0, 1890218792Snp ("%s: txpkts not empty: %d", __func__, txpkts.npkt)); 1891218792Snp 1892218792Snp /* We're sending out individual packets now */ 1893218792Snp coalescing = 0; 1894218792Snp 1895218792Snp if (eq->avail < 8) 1896220873Snp reclaim_tx_descs(txq, 0, 8); 1897218792Snp rc = write_txpkt_wr(pi, txq, m, &sgl); 1898218792Snp if (rc != 0) { 1899218792Snp 1900218792Snp /* Short of hardware descriptors, suspend tx */ 1901218792Snp 1902218792Snp /* 1903218792Snp * This is an unlikely but expensive failure. We've 1904218792Snp * done all the hard work (DMA mappings etc.) and now we 1905218792Snp * can't send out the packet. What's worse, we have to 1906218792Snp * spend even more time freeing up everything in sgl. 1907218792Snp */ 1908218792Snp txq->no_desc++; 1909218792Snp free_pkt_sgl(txq, &sgl); 1910218792Snp 1911218792Snp m->m_nextpkt = next; 1912218792Snp break; 1913218792Snp } 1914218792Snp 1915218792Snp ETHER_BPF_MTAP(ifp, m); 1916218792Snp if (sgl.nsegs == 0) 1917218792Snp m_freem(m); 1918218792Snpdoorbell: 1919252715Snp if (eq->pending >= 8) 1920252715Snp ring_eq_db(sc, eq); 1921219292Snp 1922219292Snp can_reclaim = reclaimable(eq); 1923219292Snp if (can_reclaim >= 32) 1924228561Snp reclaim_tx_descs(txq, can_reclaim, 64); 1925218792Snp } 1926218792Snp 1927218792Snp if (txpkts.npkt > 0) 1928218792Snp write_txpkts_wr(txq, &txpkts); 1929218792Snp 1930218792Snp /* 1931218792Snp * m not NULL means there was an error but we haven't thrown it away. 1932218792Snp * This can happen when we're short of tx descriptors (no_desc) or maybe 1933218792Snp * even DMA maps (no_dmamap). Either way, a credit flush and reclaim 1934218792Snp * will get things going again. 1935218792Snp */ 1936228561Snp if (m && !(eq->flags & EQ_CRFLUSHED)) { 1937220873Snp struct tx_sdesc *txsd = &txq->sdesc[eq->pidx]; 1938220873Snp 1939228561Snp /* 1940228561Snp * If EQ_CRFLUSHED is not set then we know we have at least one 1941228561Snp * available descriptor because any WR that reduces eq->avail to 1942228561Snp * 0 also sets EQ_CRFLUSHED. 1943228561Snp */ 1944228561Snp KASSERT(eq->avail > 0, ("%s: no space for eqflush.", __func__)); 1945228561Snp 1946220873Snp txsd->desc_used = 1; 1947220873Snp txsd->credits = 0; 1948218792Snp write_eqflush_wr(eq); 1949220873Snp } 1950218792Snp txq->m = m; 1951218792Snp 1952218792Snp if (eq->pending) 1953220873Snp ring_eq_db(sc, eq); 1954218792Snp 1955228561Snp reclaim_tx_descs(txq, 0, 128); 1956218792Snp 1957228561Snp if (eq->flags & EQ_STALLED && callout_pending(&eq->tx_callout) == 0) 1958228561Snp callout_reset(&eq->tx_callout, 1, t4_tx_callout, eq); 1959228561Snp 1960218792Snp return (0); 1961218792Snp} 1962218792Snp 1963218792Snpvoid 1964218792Snpt4_update_fl_bufsize(struct ifnet *ifp) 1965218792Snp{ 1966218792Snp struct port_info *pi = ifp->if_softc; 1967255050Snp struct adapter *sc = pi->adapter; 1968218792Snp struct sge_rxq *rxq; 1969252728Snp#ifdef TCP_OFFLOAD 1970252728Snp struct sge_ofld_rxq *ofld_rxq; 1971252728Snp#endif 1972218792Snp struct sge_fl *fl; 1973265425Snp int i, maxp, mtu = ifp->if_mtu; 1974218792Snp 1975265425Snp maxp = mtu_to_max_payload(sc, mtu, 0); 1976218792Snp for_each_rxq(pi, i, rxq) { 1977218792Snp fl = &rxq->fl; 1978218792Snp 1979218792Snp FL_LOCK(fl); 1980265425Snp find_best_refill_source(sc, fl, maxp); 1981218792Snp FL_UNLOCK(fl); 1982218792Snp } 1983252728Snp#ifdef TCP_OFFLOAD 1984265425Snp maxp = mtu_to_max_payload(sc, mtu, 1); 1985252728Snp for_each_ofld_rxq(pi, i, ofld_rxq) { 1986252728Snp fl = &ofld_rxq->fl; 1987252728Snp 1988252728Snp FL_LOCK(fl); 1989265425Snp find_best_refill_source(sc, fl, maxp); 1990252728Snp FL_UNLOCK(fl); 1991252728Snp } 1992252728Snp#endif 1993218792Snp} 1994218792Snp 1995228561Snpint 1996228561Snpcan_resume_tx(struct sge_eq *eq) 1997228561Snp{ 1998228561Snp return (reclaimable(eq) >= tx_resume_threshold(eq)); 1999228561Snp} 2000228561Snp 2001218792Snpstatic inline void 2002218792Snpinit_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx, 2003241397Snp int qsize, int esize) 2004218792Snp{ 2005218792Snp KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS, 2006218792Snp ("%s: bad tmr_idx %d", __func__, tmr_idx)); 2007218792Snp KASSERT(pktc_idx < SGE_NCOUNTERS, /* -ve is ok, means don't use */ 2008218792Snp ("%s: bad pktc_idx %d", __func__, pktc_idx)); 2009218792Snp 2010218792Snp iq->flags = 0; 2011218792Snp iq->adapter = sc; 2012234833Snp iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx); 2013234833Snp iq->intr_pktc_idx = SGE_NCOUNTERS - 1; 2014234833Snp if (pktc_idx >= 0) { 2015234833Snp iq->intr_params |= F_QINTR_CNT_EN; 2016234833Snp iq->intr_pktc_idx = pktc_idx; 2017234833Snp } 2018248925Snp iq->qsize = roundup2(qsize, 16); /* See FW_IQ_CMD/iqsize */ 2019218792Snp iq->esize = max(esize, 16); /* See FW_IQ_CMD/iqesize */ 2020218792Snp} 2021218792Snp 2022218792Snpstatic inline void 2023265425Snpinit_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, int pack, 2024255050Snp char *name) 2025218792Snp{ 2026255050Snp 2027218792Snp fl->qsize = qsize; 2028218792Snp strlcpy(fl->lockname, name, sizeof(fl->lockname)); 2029255050Snp if (pack) 2030255050Snp fl->flags |= FL_BUF_PACKING; 2031265425Snp find_best_refill_source(sc, fl, maxp); 2032265425Snp find_safe_refill_source(sc, fl); 2033218792Snp} 2034218792Snp 2035218792Snpstatic inline void 2036228561Snpinit_eq(struct sge_eq *eq, int eqtype, int qsize, uint8_t tx_chan, 2037228561Snp uint16_t iqid, char *name) 2038218792Snp{ 2039228561Snp KASSERT(tx_chan < NCHAN, ("%s: bad tx channel %d", __func__, tx_chan)); 2040228561Snp KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype)); 2041228561Snp 2042228561Snp eq->flags = eqtype & EQ_TYPEMASK; 2043228561Snp eq->tx_chan = tx_chan; 2044228561Snp eq->iqid = iqid; 2045220873Snp eq->qsize = qsize; 2046220873Snp strlcpy(eq->lockname, name, sizeof(eq->lockname)); 2047228561Snp 2048228561Snp TASK_INIT(&eq->tx_task, 0, t4_tx_task, eq); 2049228561Snp callout_init(&eq->tx_callout, CALLOUT_MPSAFE); 2050218792Snp} 2051218792Snp 2052218792Snpstatic int 2053218792Snpalloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag, 2054218792Snp bus_dmamap_t *map, bus_addr_t *pa, void **va) 2055218792Snp{ 2056218792Snp int rc; 2057218792Snp 2058218792Snp rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR, 2059218792Snp BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag); 2060218792Snp if (rc != 0) { 2061218792Snp device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc); 2062218792Snp goto done; 2063218792Snp } 2064218792Snp 2065218792Snp rc = bus_dmamem_alloc(*tag, va, 2066218792Snp BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map); 2067218792Snp if (rc != 0) { 2068218792Snp device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc); 2069218792Snp goto done; 2070218792Snp } 2071218792Snp 2072218792Snp rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0); 2073218792Snp if (rc != 0) { 2074218792Snp device_printf(sc->dev, "cannot load DMA map: %d\n", rc); 2075218792Snp goto done; 2076218792Snp } 2077218792Snpdone: 2078218792Snp if (rc) 2079218792Snp free_ring(sc, *tag, *map, *pa, *va); 2080218792Snp 2081218792Snp return (rc); 2082218792Snp} 2083218792Snp 2084218792Snpstatic int 2085218792Snpfree_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map, 2086218792Snp bus_addr_t pa, void *va) 2087218792Snp{ 2088218792Snp if (pa) 2089218792Snp bus_dmamap_unload(tag, map); 2090218792Snp if (va) 2091218792Snp bus_dmamem_free(tag, va, map); 2092218792Snp if (tag) 2093218792Snp bus_dma_tag_destroy(tag); 2094218792Snp 2095218792Snp return (0); 2096218792Snp} 2097218792Snp 2098218792Snp/* 2099218792Snp * Allocates the ring for an ingress queue and an optional freelist. If the 2100218792Snp * freelist is specified it will be allocated and then associated with the 2101218792Snp * ingress queue. 2102218792Snp * 2103218792Snp * Returns errno on failure. Resources allocated up to that point may still be 2104218792Snp * allocated. Caller is responsible for cleanup in case this function fails. 2105218792Snp * 2106228561Snp * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then 2107218792Snp * the intr_idx specifies the vector, starting from 0. Otherwise it specifies 2108228561Snp * the abs_id of the ingress queue to which its interrupts should be forwarded. 2109218792Snp */ 2110218792Snpstatic int 2111218792Snpalloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl, 2112222085Snp int intr_idx, int cong) 2113218792Snp{ 2114218792Snp int rc, i, cntxt_id; 2115218792Snp size_t len; 2116218792Snp struct fw_iq_cmd c; 2117218792Snp struct adapter *sc = iq->adapter; 2118218792Snp __be32 v = 0; 2119218792Snp 2120218792Snp len = iq->qsize * iq->esize; 2121218792Snp rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba, 2122218792Snp (void **)&iq->desc); 2123218792Snp if (rc != 0) 2124218792Snp return (rc); 2125218792Snp 2126218792Snp bzero(&c, sizeof(c)); 2127218792Snp c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST | 2128218792Snp F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) | 2129218792Snp V_FW_IQ_CMD_VFN(0)); 2130218792Snp 2131218792Snp c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART | 2132218792Snp FW_LEN16(c)); 2133218792Snp 2134218792Snp /* Special handling for firmware event queue */ 2135218792Snp if (iq == &sc->sge.fwq) 2136218792Snp v |= F_FW_IQ_CMD_IQASYNCH; 2137218792Snp 2138228561Snp if (iq->flags & IQ_INTR) { 2139218792Snp KASSERT(intr_idx < sc->intr_count, 2140218792Snp ("%s: invalid direct intr_idx %d", __func__, intr_idx)); 2141228561Snp } else 2142228561Snp v |= F_FW_IQ_CMD_IQANDST; 2143228561Snp v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx); 2144218792Snp 2145218792Snp c.type_to_iqandstindex = htobe32(v | 2146218792Snp V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) | 2147218792Snp V_FW_IQ_CMD_VIID(pi->viid) | 2148218792Snp V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT)); 2149218792Snp c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) | 2150218792Snp F_FW_IQ_CMD_IQGTSMODE | 2151218792Snp V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) | 2152218792Snp V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4)); 2153218792Snp c.iqsize = htobe16(iq->qsize); 2154218792Snp c.iqaddr = htobe64(iq->ba); 2155222085Snp if (cong >= 0) 2156222085Snp c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN); 2157218792Snp 2158218792Snp if (fl) { 2159218792Snp mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF); 2160218792Snp 2161218792Snp len = fl->qsize * RX_FL_ESIZE; 2162218792Snp rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map, 2163218792Snp &fl->ba, (void **)&fl->desc); 2164218792Snp if (rc) 2165218792Snp return (rc); 2166218792Snp 2167218792Snp /* Allocate space for one software descriptor per buffer. */ 2168237512Snp fl->cap = (fl->qsize - spg_len / RX_FL_ESIZE) * 8; 2169218792Snp rc = alloc_fl_sdesc(fl); 2170218792Snp if (rc != 0) { 2171218792Snp device_printf(sc->dev, 2172218792Snp "failed to setup fl software descriptors: %d\n", 2173218792Snp rc); 2174218792Snp return (rc); 2175218792Snp } 2176220905Snp fl->needed = fl->cap; 2177265410Snp fl->lowat = fl->flags & FL_BUF_PACKING ? 2178265410Snp roundup2(sc->sge.fl_starve_threshold2, 8) : 2179265410Snp roundup2(sc->sge.fl_starve_threshold, 8); 2180218792Snp 2181228491Snp c.iqns_to_fl0congen |= 2182222085Snp htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) | 2183222085Snp F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO | 2184255050Snp (fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) | 2185255050Snp (fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN : 2186255050Snp 0)); 2187222085Snp if (cong >= 0) { 2188222085Snp c.iqns_to_fl0congen |= 2189222085Snp htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) | 2190222085Snp F_FW_IQ_CMD_FL0CONGCIF | 2191222085Snp F_FW_IQ_CMD_FL0CONGEN); 2192222085Snp } 2193218792Snp c.fl0dcaen_to_fl0cidxfthresh = 2194218792Snp htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_64B) | 2195218792Snp V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B)); 2196218792Snp c.fl0size = htobe16(fl->qsize); 2197218792Snp c.fl0addr = htobe64(fl->ba); 2198218792Snp } 2199218792Snp 2200218792Snp rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); 2201218792Snp if (rc != 0) { 2202218792Snp device_printf(sc->dev, 2203218792Snp "failed to create ingress queue: %d\n", rc); 2204218792Snp return (rc); 2205218792Snp } 2206218792Snp 2207218792Snp iq->cdesc = iq->desc; 2208218792Snp iq->cidx = 0; 2209218792Snp iq->gen = 1; 2210218792Snp iq->intr_next = iq->intr_params; 2211218792Snp iq->cntxt_id = be16toh(c.iqid); 2212218792Snp iq->abs_id = be16toh(c.physiqid); 2213228561Snp iq->flags |= IQ_ALLOCATED; 2214218792Snp 2215218792Snp cntxt_id = iq->cntxt_id - sc->sge.iq_start; 2216228561Snp if (cntxt_id >= sc->sge.niq) { 2217228561Snp panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__, 2218228561Snp cntxt_id, sc->sge.niq - 1); 2219228561Snp } 2220218792Snp sc->sge.iqmap[cntxt_id] = iq; 2221218792Snp 2222218792Snp if (fl) { 2223218792Snp fl->cntxt_id = be16toh(c.fl0id); 2224218792Snp fl->pidx = fl->cidx = 0; 2225218792Snp 2226219883Snp cntxt_id = fl->cntxt_id - sc->sge.eq_start; 2227228561Snp if (cntxt_id >= sc->sge.neq) { 2228228561Snp panic("%s: fl->cntxt_id (%d) more than the max (%d)", 2229228561Snp __func__, cntxt_id, sc->sge.neq - 1); 2230228561Snp } 2231218792Snp sc->sge.eqmap[cntxt_id] = (void *)fl; 2232218792Snp 2233218792Snp FL_LOCK(fl); 2234228561Snp /* Enough to make sure the SGE doesn't think it's starved */ 2235228561Snp refill_fl(sc, fl, fl->lowat); 2236218792Snp FL_UNLOCK(fl); 2237228561Snp 2238228561Snp iq->flags |= IQ_HAS_FL; 2239218792Snp } 2240218792Snp 2241253873Snp if (is_t5(sc) && cong >= 0) { 2242253873Snp uint32_t param, val; 2243253873Snp 2244253873Snp param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) | 2245253873Snp V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) | 2246253873Snp V_FW_PARAMS_PARAM_YZ(iq->cntxt_id); 2247253889Snp if (cong == 0) 2248253889Snp val = 1 << 19; 2249253889Snp else { 2250253889Snp val = 2 << 19; 2251253889Snp for (i = 0; i < 4; i++) { 2252253889Snp if (cong & (1 << i)) 2253253889Snp val |= 1 << (i << 2); 2254253889Snp } 2255253889Snp } 2256253889Snp 2257253873Snp rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, ¶m, &val); 2258253873Snp if (rc != 0) { 2259253873Snp /* report error but carry on */ 2260253873Snp device_printf(sc->dev, 2261253873Snp "failed to set congestion manager context for " 2262253873Snp "ingress queue %d: %d\n", iq->cntxt_id, rc); 2263253873Snp } 2264253873Snp } 2265253873Snp 2266218792Snp /* Enable IQ interrupts */ 2267228561Snp atomic_store_rel_int(&iq->state, IQS_IDLE); 2268218792Snp t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) | 2269218792Snp V_INGRESSQID(iq->cntxt_id)); 2270218792Snp 2271218792Snp return (0); 2272218792Snp} 2273218792Snp 2274218792Snpstatic int 2275218792Snpfree_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl) 2276218792Snp{ 2277265425Snp int rc; 2278218792Snp struct adapter *sc = iq->adapter; 2279218792Snp device_t dev; 2280218792Snp 2281218792Snp if (sc == NULL) 2282218792Snp return (0); /* nothing to do */ 2283218792Snp 2284218792Snp dev = pi ? pi->dev : sc->dev; 2285218792Snp 2286218792Snp if (iq->flags & IQ_ALLOCATED) { 2287218792Snp rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, 2288218792Snp FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id, 2289218792Snp fl ? fl->cntxt_id : 0xffff, 0xffff); 2290218792Snp if (rc != 0) { 2291218792Snp device_printf(dev, 2292218792Snp "failed to free queue %p: %d\n", iq, rc); 2293218792Snp return (rc); 2294218792Snp } 2295218792Snp iq->flags &= ~IQ_ALLOCATED; 2296218792Snp } 2297218792Snp 2298218792Snp free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc); 2299218792Snp 2300218792Snp bzero(iq, sizeof(*iq)); 2301218792Snp 2302218792Snp if (fl) { 2303218792Snp free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba, 2304218792Snp fl->desc); 2305218792Snp 2306254727Snp if (fl->sdesc) 2307255050Snp free_fl_sdesc(sc, fl); 2308218792Snp 2309218792Snp if (mtx_initialized(&fl->fl_lock)) 2310218792Snp mtx_destroy(&fl->fl_lock); 2311218792Snp 2312218792Snp bzero(fl, sizeof(*fl)); 2313218792Snp } 2314218792Snp 2315218792Snp return (0); 2316218792Snp} 2317218792Snp 2318265425Snpstatic void 2319265425Snpadd_fl_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid, 2320265425Snp struct sge_fl *fl) 2321265425Snp{ 2322265425Snp struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); 2323265425Snp 2324265425Snp oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL, 2325265425Snp "freelist"); 2326265425Snp children = SYSCTL_CHILDREN(oid); 2327265425Snp 2328265425Snp SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id", 2329265425Snp CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I", 2330265425Snp "SGE context id of the freelist"); 2331265425Snp SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx, 2332265425Snp 0, "consumer index"); 2333265425Snp if (fl->flags & FL_BUF_PACKING) { 2334265425Snp SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset", 2335265425Snp CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset"); 2336265425Snp } 2337265425Snp SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx, 2338265425Snp 0, "producer index"); 2339265425Snp SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated", 2340265425Snp CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated"); 2341265425Snp SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined", 2342265425Snp CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters"); 2343265425Snp SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated", 2344265425Snp CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated"); 2345265425Snp SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled", 2346265425Snp CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled"); 2347265425Snp SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled", 2348265425Snp CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)"); 2349265425Snp} 2350265425Snp 2351218792Snpstatic int 2352228561Snpalloc_fwq(struct adapter *sc) 2353218792Snp{ 2354228561Snp int rc, intr_idx; 2355228561Snp struct sge_iq *fwq = &sc->sge.fwq; 2356228561Snp struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev); 2357228561Snp struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); 2358222510Snp 2359241397Snp init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, FW_IQ_ESIZE); 2360228561Snp fwq->flags |= IQ_INTR; /* always */ 2361228561Snp intr_idx = sc->intr_count > 1 ? 1 : 0; 2362228561Snp rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1); 2363228561Snp if (rc != 0) { 2364228561Snp device_printf(sc->dev, 2365228561Snp "failed to create firmware event queue: %d\n", rc); 2366222510Snp return (rc); 2367228561Snp } 2368222510Snp 2369228561Snp oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD, 2370228561Snp NULL, "firmware event queue"); 2371222510Snp children = SYSCTL_CHILDREN(oid); 2372222510Snp 2373228561Snp SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id", 2374228561Snp CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I", 2375228561Snp "absolute id of the queue"); 2376228561Snp SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id", 2377228561Snp CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I", 2378228561Snp "SGE context id of the queue"); 2379222510Snp SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx", 2380228561Snp CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I", 2381222510Snp "consumer index"); 2382222510Snp 2383228561Snp return (0); 2384218792Snp} 2385218792Snp 2386218792Snpstatic int 2387228561Snpfree_fwq(struct adapter *sc) 2388218792Snp{ 2389228561Snp return free_iq_fl(NULL, &sc->sge.fwq, NULL); 2390218792Snp} 2391218792Snp 2392218792Snpstatic int 2393228561Snpalloc_mgmtq(struct adapter *sc) 2394222510Snp{ 2395222510Snp int rc; 2396228561Snp struct sge_wrq *mgmtq = &sc->sge.mgmtq; 2397228561Snp char name[16]; 2398228561Snp struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev); 2399228561Snp struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); 2400222510Snp 2401228561Snp oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD, 2402228561Snp NULL, "management queue"); 2403228561Snp 2404228561Snp snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev)); 2405228561Snp init_eq(&mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan, 2406228561Snp sc->sge.fwq.cntxt_id, name); 2407228561Snp rc = alloc_wrq(sc, NULL, mgmtq, oid); 2408228561Snp if (rc != 0) { 2409228561Snp device_printf(sc->dev, 2410228561Snp "failed to create management queue: %d\n", rc); 2411222510Snp return (rc); 2412228561Snp } 2413222510Snp 2414228561Snp return (0); 2415222510Snp} 2416222510Snp 2417222510Snpstatic int 2418228561Snpfree_mgmtq(struct adapter *sc) 2419222510Snp{ 2420237263Snp 2421228561Snp return free_wrq(sc, &sc->sge.mgmtq); 2422222510Snp} 2423222510Snp 2424239258Snpstatic inline int 2425239258Snptnl_cong(struct port_info *pi) 2426239258Snp{ 2427239258Snp 2428239258Snp if (cong_drop == -1) 2429239258Snp return (-1); 2430239258Snp else if (cong_drop == 1) 2431239258Snp return (0); 2432239258Snp else 2433265410Snp return (pi->rx_chan_map); 2434239258Snp} 2435239258Snp 2436222510Snpstatic int 2437228561Snpalloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int idx, 2438228561Snp struct sysctl_oid *oid) 2439218792Snp{ 2440218792Snp int rc; 2441218792Snp struct sysctl_oid_list *children; 2442218792Snp char name[16]; 2443218792Snp 2444239258Snp rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx, tnl_cong(pi)); 2445218792Snp if (rc != 0) 2446218792Snp return (rc); 2447218792Snp 2448222701Snp FL_LOCK(&rxq->fl); 2449228561Snp refill_fl(pi->adapter, &rxq->fl, rxq->fl.needed / 8); 2450222701Snp FL_UNLOCK(&rxq->fl); 2451222701Snp 2452237819Snp#if defined(INET) || defined(INET6) 2453218792Snp rc = tcp_lro_init(&rxq->lro); 2454218792Snp if (rc != 0) 2455218792Snp return (rc); 2456218792Snp rxq->lro.ifp = pi->ifp; /* also indicates LRO init'ed */ 2457218792Snp 2458218792Snp if (pi->ifp->if_capenable & IFCAP_LRO) 2459228561Snp rxq->iq.flags |= IQ_LRO_ENABLED; 2460218792Snp#endif 2461219289Snp rxq->ifp = pi->ifp; 2462218792Snp 2463228561Snp children = SYSCTL_CHILDREN(oid); 2464218792Snp 2465218792Snp snprintf(name, sizeof(name), "%d", idx); 2466218792Snp oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD, 2467218792Snp NULL, "rx queue"); 2468218792Snp children = SYSCTL_CHILDREN(oid); 2469218792Snp 2470221911Snp SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "abs_id", 2471222510Snp CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I", 2472221911Snp "absolute id of the queue"); 2473222973Snp SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id", 2474222973Snp CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I", 2475222973Snp "SGE context id of the queue"); 2476222973Snp SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx", 2477222973Snp CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I", 2478222973Snp "consumer index"); 2479237819Snp#if defined(INET) || defined(INET6) 2480218792Snp SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD, 2481218792Snp &rxq->lro.lro_queued, 0, NULL); 2482218792Snp SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD, 2483218792Snp &rxq->lro.lro_flushed, 0, NULL); 2484219290Snp#endif 2485218792Snp SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD, 2486218792Snp &rxq->rxcsum, "# of times hardware assisted with checksum"); 2487218792Snp SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_extraction", 2488218792Snp CTLFLAG_RD, &rxq->vlan_extraction, 2489218792Snp "# of times hardware extracted 802.1Q tag"); 2490218792Snp 2491265425Snp add_fl_sysctls(&pi->ctx, oid, &rxq->fl); 2492222973Snp 2493218792Snp return (rc); 2494218792Snp} 2495218792Snp 2496218792Snpstatic int 2497218792Snpfree_rxq(struct port_info *pi, struct sge_rxq *rxq) 2498218792Snp{ 2499218792Snp int rc; 2500218792Snp 2501237819Snp#if defined(INET) || defined(INET6) 2502218792Snp if (rxq->lro.ifp) { 2503218792Snp tcp_lro_free(&rxq->lro); 2504218792Snp rxq->lro.ifp = NULL; 2505218792Snp } 2506218792Snp#endif 2507218792Snp 2508218792Snp rc = free_iq_fl(pi, &rxq->iq, &rxq->fl); 2509218792Snp if (rc == 0) 2510218792Snp bzero(rxq, sizeof(*rxq)); 2511218792Snp 2512218792Snp return (rc); 2513218792Snp} 2514218792Snp 2515237263Snp#ifdef TCP_OFFLOAD 2516218792Snpstatic int 2517228561Snpalloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq, 2518228561Snp int intr_idx, int idx, struct sysctl_oid *oid) 2519220873Snp{ 2520228561Snp int rc; 2521228561Snp struct sysctl_oid_list *children; 2522220873Snp char name[16]; 2523220873Snp 2524228561Snp rc = alloc_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 2525265410Snp pi->rx_chan_map); 2526228561Snp if (rc != 0) 2527220873Snp return (rc); 2528220873Snp 2529228561Snp children = SYSCTL_CHILDREN(oid); 2530220873Snp 2531228561Snp snprintf(name, sizeof(name), "%d", idx); 2532228561Snp oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD, 2533228561Snp NULL, "rx queue"); 2534228561Snp children = SYSCTL_CHILDREN(oid); 2535228561Snp 2536228561Snp SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "abs_id", 2537228561Snp CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16, 2538228561Snp "I", "absolute id of the queue"); 2539228561Snp SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id", 2540228561Snp CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16, 2541228561Snp "I", "SGE context id of the queue"); 2542228561Snp SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx", 2543228561Snp CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I", 2544228561Snp "consumer index"); 2545228561Snp 2546265425Snp add_fl_sysctls(&pi->ctx, oid, &ofld_rxq->fl); 2547228561Snp 2548228561Snp return (rc); 2549228561Snp} 2550228561Snp 2551228561Snpstatic int 2552228561Snpfree_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq) 2553228561Snp{ 2554228561Snp int rc; 2555228561Snp 2556228561Snp rc = free_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl); 2557228561Snp if (rc == 0) 2558228561Snp bzero(ofld_rxq, sizeof(*ofld_rxq)); 2559228561Snp 2560228561Snp return (rc); 2561228561Snp} 2562228561Snp#endif 2563228561Snp 2564228561Snpstatic int 2565228561Snpctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq) 2566228561Snp{ 2567228561Snp int rc, cntxt_id; 2568228561Snp struct fw_eq_ctrl_cmd c; 2569228561Snp 2570220873Snp bzero(&c, sizeof(c)); 2571220873Snp 2572220873Snp c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST | 2573220873Snp F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) | 2574220873Snp V_FW_EQ_CTRL_CMD_VFN(0)); 2575220873Snp c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC | 2576220873Snp F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c)); 2577220873Snp c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); /* XXX */ 2578220873Snp c.physeqid_pkd = htobe32(0); 2579220873Snp c.fetchszm_to_iqid = 2580220873Snp htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) | 2581228561Snp V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) | 2582222510Snp F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid)); 2583220873Snp c.dcaen_to_eqsize = 2584220873Snp htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) | 2585220873Snp V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) | 2586220873Snp V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) | 2587220873Snp V_FW_EQ_CTRL_CMD_EQSIZE(eq->qsize)); 2588220873Snp c.eqaddr = htobe64(eq->ba); 2589220873Snp 2590220873Snp rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); 2591220873Snp if (rc != 0) { 2592220873Snp device_printf(sc->dev, 2593228561Snp "failed to create control queue %d: %d\n", eq->tx_chan, rc); 2594220873Snp return (rc); 2595220873Snp } 2596228561Snp eq->flags |= EQ_ALLOCATED; 2597220873Snp 2598220873Snp eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid)); 2599228561Snp cntxt_id = eq->cntxt_id - sc->sge.eq_start; 2600228561Snp if (cntxt_id >= sc->sge.neq) 2601228561Snp panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__, 2602228561Snp cntxt_id, sc->sge.neq - 1); 2603228561Snp sc->sge.eqmap[cntxt_id] = eq; 2604220873Snp 2605228561Snp return (rc); 2606228561Snp} 2607228561Snp 2608228561Snpstatic int 2609228561Snpeth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq) 2610228561Snp{ 2611228561Snp int rc, cntxt_id; 2612228561Snp struct fw_eq_eth_cmd c; 2613228561Snp 2614228561Snp bzero(&c, sizeof(c)); 2615228561Snp 2616228561Snp c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST | 2617228561Snp F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) | 2618228561Snp V_FW_EQ_ETH_CMD_VFN(0)); 2619228561Snp c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC | 2620228561Snp F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c)); 2621228561Snp c.viid_pkd = htobe32(V_FW_EQ_ETH_CMD_VIID(pi->viid)); 2622228561Snp c.fetchszm_to_iqid = 2623228561Snp htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) | 2624228561Snp V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO | 2625228561Snp V_FW_EQ_ETH_CMD_IQID(eq->iqid)); 2626228561Snp c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) | 2627228561Snp V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) | 2628228561Snp V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) | 2629228561Snp V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize)); 2630228561Snp c.eqaddr = htobe64(eq->ba); 2631228561Snp 2632228561Snp rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); 2633228561Snp if (rc != 0) { 2634228561Snp device_printf(pi->dev, 2635228561Snp "failed to create Ethernet egress queue: %d\n", rc); 2636228561Snp return (rc); 2637228561Snp } 2638228561Snp eq->flags |= EQ_ALLOCATED; 2639228561Snp 2640228561Snp eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd)); 2641220873Snp cntxt_id = eq->cntxt_id - sc->sge.eq_start; 2642228561Snp if (cntxt_id >= sc->sge.neq) 2643228561Snp panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__, 2644228561Snp cntxt_id, sc->sge.neq - 1); 2645220873Snp sc->sge.eqmap[cntxt_id] = eq; 2646220873Snp 2647228561Snp return (rc); 2648228561Snp} 2649220873Snp 2650237263Snp#ifdef TCP_OFFLOAD 2651228561Snpstatic int 2652228561Snpofld_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq) 2653228561Snp{ 2654228561Snp int rc, cntxt_id; 2655228561Snp struct fw_eq_ofld_cmd c; 2656220873Snp 2657228561Snp bzero(&c, sizeof(c)); 2658220873Snp 2659228561Snp c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST | 2660228561Snp F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) | 2661228561Snp V_FW_EQ_OFLD_CMD_VFN(0)); 2662228561Snp c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC | 2663228561Snp F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c)); 2664228561Snp c.fetchszm_to_iqid = 2665228561Snp htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) | 2666228561Snp V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) | 2667228561Snp F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid)); 2668228561Snp c.dcaen_to_eqsize = 2669228561Snp htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) | 2670228561Snp V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) | 2671228561Snp V_FW_EQ_OFLD_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) | 2672228561Snp V_FW_EQ_OFLD_CMD_EQSIZE(eq->qsize)); 2673228561Snp c.eqaddr = htobe64(eq->ba); 2674228561Snp 2675228561Snp rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); 2676228561Snp if (rc != 0) { 2677228561Snp device_printf(pi->dev, 2678228561Snp "failed to create egress queue for TCP offload: %d\n", rc); 2679228561Snp return (rc); 2680228561Snp } 2681228561Snp eq->flags |= EQ_ALLOCATED; 2682228561Snp 2683228561Snp eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd)); 2684228561Snp cntxt_id = eq->cntxt_id - sc->sge.eq_start; 2685228561Snp if (cntxt_id >= sc->sge.neq) 2686228561Snp panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__, 2687228561Snp cntxt_id, sc->sge.neq - 1); 2688228561Snp sc->sge.eqmap[cntxt_id] = eq; 2689228561Snp 2690220873Snp return (rc); 2691220873Snp} 2692228561Snp#endif 2693220873Snp 2694220873Snpstatic int 2695228561Snpalloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq) 2696220873Snp{ 2697220873Snp int rc; 2698228561Snp size_t len; 2699220873Snp 2700228561Snp mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF); 2701228561Snp 2702228561Snp len = eq->qsize * EQ_ESIZE; 2703228561Snp rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map, 2704228561Snp &eq->ba, (void **)&eq->desc); 2705228561Snp if (rc) 2706228561Snp return (rc); 2707228561Snp 2708237512Snp eq->cap = eq->qsize - spg_len / EQ_ESIZE; 2709228561Snp eq->spg = (void *)&eq->desc[eq->cap]; 2710228561Snp eq->avail = eq->cap - 1; /* one less to avoid cidx = pidx */ 2711228561Snp eq->pidx = eq->cidx = 0; 2712248925Snp eq->doorbells = sc->doorbells; 2713228561Snp 2714228561Snp switch (eq->flags & EQ_TYPEMASK) { 2715228561Snp case EQ_CTRL: 2716228561Snp rc = ctrl_eq_alloc(sc, eq); 2717228561Snp break; 2718228561Snp 2719228561Snp case EQ_ETH: 2720228561Snp rc = eth_eq_alloc(sc, pi, eq); 2721228561Snp break; 2722228561Snp 2723237263Snp#ifdef TCP_OFFLOAD 2724228561Snp case EQ_OFLD: 2725228561Snp rc = ofld_eq_alloc(sc, pi, eq); 2726228561Snp break; 2727228561Snp#endif 2728228561Snp 2729228561Snp default: 2730228561Snp panic("%s: invalid eq type %d.", __func__, 2731228561Snp eq->flags & EQ_TYPEMASK); 2732228561Snp } 2733228561Snp if (rc != 0) { 2734228561Snp device_printf(sc->dev, 2735228561Snp "failed to allocate egress queue(%d): %d", 2736228561Snp eq->flags & EQ_TYPEMASK, rc); 2737228561Snp } 2738228561Snp 2739228561Snp eq->tx_callout.c_cpu = eq->cntxt_id % mp_ncpus; 2740228561Snp 2741248925Snp if (isset(&eq->doorbells, DOORBELL_UDB) || 2742248925Snp isset(&eq->doorbells, DOORBELL_UDBWC) || 2743249392Snp isset(&eq->doorbells, DOORBELL_WCWR)) { 2744256794Snp uint32_t s_qpp = sc->sge.eq_s_qpp; 2745248925Snp uint32_t mask = (1 << s_qpp) - 1; 2746248925Snp volatile uint8_t *udb; 2747248925Snp 2748248925Snp udb = sc->udbs_base + UDBS_DB_OFFSET; 2749248925Snp udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT; /* pg offset */ 2750248925Snp eq->udb_qid = eq->cntxt_id & mask; /* id in page */ 2751248925Snp if (eq->udb_qid > PAGE_SIZE / UDBS_SEG_SIZE) 2752249392Snp clrbit(&eq->doorbells, DOORBELL_WCWR); 2753248925Snp else { 2754248925Snp udb += eq->udb_qid << UDBS_SEG_SHIFT; /* seg offset */ 2755248925Snp eq->udb_qid = 0; 2756248925Snp } 2757248925Snp eq->udb = (volatile void *)udb; 2758248925Snp } 2759248925Snp 2760228561Snp return (rc); 2761228561Snp} 2762228561Snp 2763228561Snpstatic int 2764228561Snpfree_eq(struct adapter *sc, struct sge_eq *eq) 2765228561Snp{ 2766228561Snp int rc; 2767228561Snp 2768228561Snp if (eq->flags & EQ_ALLOCATED) { 2769228561Snp switch (eq->flags & EQ_TYPEMASK) { 2770228561Snp case EQ_CTRL: 2771228561Snp rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0, 2772228561Snp eq->cntxt_id); 2773228561Snp break; 2774228561Snp 2775228561Snp case EQ_ETH: 2776228561Snp rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, 2777228561Snp eq->cntxt_id); 2778228561Snp break; 2779228561Snp 2780237263Snp#ifdef TCP_OFFLOAD 2781228561Snp case EQ_OFLD: 2782228561Snp rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0, 2783228561Snp eq->cntxt_id); 2784228561Snp break; 2785228561Snp#endif 2786228561Snp 2787228561Snp default: 2788228561Snp panic("%s: invalid eq type %d.", __func__, 2789228561Snp eq->flags & EQ_TYPEMASK); 2790228561Snp } 2791220873Snp if (rc != 0) { 2792220873Snp device_printf(sc->dev, 2793228561Snp "failed to free egress queue (%d): %d\n", 2794228561Snp eq->flags & EQ_TYPEMASK, rc); 2795220873Snp return (rc); 2796220873Snp } 2797228561Snp eq->flags &= ~EQ_ALLOCATED; 2798220873Snp } 2799220873Snp 2800220873Snp free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc); 2801220873Snp 2802220873Snp if (mtx_initialized(&eq->eq_lock)) 2803220873Snp mtx_destroy(&eq->eq_lock); 2804220873Snp 2805228561Snp bzero(eq, sizeof(*eq)); 2806220873Snp return (0); 2807220873Snp} 2808220873Snp 2809220873Snpstatic int 2810228561Snpalloc_wrq(struct adapter *sc, struct port_info *pi, struct sge_wrq *wrq, 2811228561Snp struct sysctl_oid *oid) 2812218792Snp{ 2813228561Snp int rc; 2814228561Snp struct sysctl_ctx_list *ctx = pi ? &pi->ctx : &sc->ctx; 2815228561Snp struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); 2816228561Snp 2817228561Snp rc = alloc_eq(sc, pi, &wrq->eq); 2818228561Snp if (rc) 2819228561Snp return (rc); 2820228561Snp 2821228561Snp wrq->adapter = sc; 2822237263Snp STAILQ_INIT(&wrq->wr_list); 2823228561Snp 2824228561Snp SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD, 2825228561Snp &wrq->eq.cntxt_id, 0, "SGE context id of the queue"); 2826228561Snp SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx", 2827228561Snp CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I", 2828228561Snp "consumer index"); 2829228561Snp SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx", 2830228561Snp CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I", 2831228561Snp "producer index"); 2832228561Snp SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs", CTLFLAG_RD, 2833228561Snp &wrq->tx_wrs, "# of work requests"); 2834228561Snp SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "no_desc", CTLFLAG_RD, 2835228561Snp &wrq->no_desc, 0, 2836228561Snp "# of times queue ran out of hardware descriptors"); 2837228561Snp SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "unstalled", CTLFLAG_RD, 2838228561Snp &wrq->eq.unstalled, 0, "# of times queue recovered after stall"); 2839228561Snp 2840228561Snp return (rc); 2841228561Snp} 2842228561Snp 2843228561Snpstatic int 2844228561Snpfree_wrq(struct adapter *sc, struct sge_wrq *wrq) 2845228561Snp{ 2846228561Snp int rc; 2847228561Snp 2848228561Snp rc = free_eq(sc, &wrq->eq); 2849228561Snp if (rc) 2850228561Snp return (rc); 2851228561Snp 2852228561Snp bzero(wrq, sizeof(*wrq)); 2853228561Snp return (0); 2854228561Snp} 2855228561Snp 2856228561Snpstatic int 2857228561Snpalloc_txq(struct port_info *pi, struct sge_txq *txq, int idx, 2858228561Snp struct sysctl_oid *oid) 2859228561Snp{ 2860228561Snp int rc; 2861218792Snp struct adapter *sc = pi->adapter; 2862218792Snp struct sge_eq *eq = &txq->eq; 2863218792Snp char name[16]; 2864228561Snp struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); 2865218792Snp 2866228561Snp rc = alloc_eq(sc, pi, eq); 2867218792Snp if (rc) 2868218792Snp return (rc); 2869218792Snp 2870228561Snp txq->ifp = pi->ifp; 2871228561Snp 2872220873Snp txq->sdesc = malloc(eq->cap * sizeof(struct tx_sdesc), M_CXGBE, 2873218792Snp M_ZERO | M_WAITOK); 2874220873Snp txq->br = buf_ring_alloc(eq->qsize, M_CXGBE, M_WAITOK, &eq->eq_lock); 2875218792Snp 2876218792Snp rc = bus_dma_tag_create(sc->dmat, 1, 0, BUS_SPACE_MAXADDR, 2877218792Snp BUS_SPACE_MAXADDR, NULL, NULL, 64 * 1024, TX_SGL_SEGS, 2878220873Snp BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL, NULL, &txq->tx_tag); 2879218792Snp if (rc != 0) { 2880218792Snp device_printf(sc->dev, 2881218792Snp "failed to create tx DMA tag: %d\n", rc); 2882218792Snp return (rc); 2883218792Snp } 2884218792Snp 2885228561Snp /* 2886228561Snp * We can stuff ~10 frames in an 8-descriptor txpkts WR (8 is the SGE 2887228561Snp * limit for any WR). txq->no_dmamap events shouldn't occur if maps is 2888228561Snp * sized for the worst case. 2889228561Snp */ 2890228561Snp rc = t4_alloc_tx_maps(&txq->txmaps, txq->tx_tag, eq->qsize * 10 / 8, 2891228561Snp M_WAITOK); 2892218792Snp if (rc != 0) { 2893218792Snp device_printf(sc->dev, "failed to setup tx DMA maps: %d\n", rc); 2894218792Snp return (rc); 2895218792Snp } 2896218792Snp 2897218792Snp snprintf(name, sizeof(name), "%d", idx); 2898218792Snp oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD, 2899218792Snp NULL, "tx queue"); 2900218792Snp children = SYSCTL_CHILDREN(oid); 2901218792Snp 2902222973Snp SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD, 2903222973Snp &eq->cntxt_id, 0, "SGE context id of the queue"); 2904222973Snp SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx", 2905222973Snp CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I", 2906222973Snp "consumer index"); 2907222973Snp SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "pidx", 2908222973Snp CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I", 2909222973Snp "producer index"); 2910222973Snp 2911218792Snp SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD, 2912218792Snp &txq->txcsum, "# of times hardware assisted with checksum"); 2913218792Snp SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_insertion", 2914218792Snp CTLFLAG_RD, &txq->vlan_insertion, 2915218792Snp "# of times hardware inserted 802.1Q tag"); 2916218792Snp SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD, 2917237819Snp &txq->tso_wrs, "# of TSO work requests"); 2918218792Snp SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD, 2919218792Snp &txq->imm_wrs, "# of work requests with immediate data"); 2920218792Snp SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD, 2921218792Snp &txq->sgl_wrs, "# of work requests with direct SGL"); 2922218792Snp SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD, 2923218792Snp &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)"); 2924218792Snp SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts_wrs", CTLFLAG_RD, 2925218792Snp &txq->txpkts_wrs, "# of txpkts work requests (multiple pkts/WR)"); 2926218792Snp SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts_pkts", CTLFLAG_RD, 2927218792Snp &txq->txpkts_pkts, "# of frames tx'd using txpkts work requests"); 2928218792Snp 2929246093Snp SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "br_drops", CTLFLAG_RD, 2930246093Snp &txq->br->br_drops, "# of drops in the buf_ring for this queue"); 2931218792Snp SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "no_dmamap", CTLFLAG_RD, 2932218792Snp &txq->no_dmamap, 0, "# of times txq ran out of DMA maps"); 2933218792Snp SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "no_desc", CTLFLAG_RD, 2934218792Snp &txq->no_desc, 0, "# of times txq ran out of hardware descriptors"); 2935218792Snp SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "egr_update", CTLFLAG_RD, 2936228561Snp &eq->egr_update, 0, "egress update notifications from the SGE"); 2937228561Snp SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "unstalled", CTLFLAG_RD, 2938228561Snp &eq->unstalled, 0, "# of times txq recovered after stall"); 2939218792Snp 2940218792Snp return (rc); 2941218792Snp} 2942218792Snp 2943218792Snpstatic int 2944218792Snpfree_txq(struct port_info *pi, struct sge_txq *txq) 2945218792Snp{ 2946218792Snp int rc; 2947218792Snp struct adapter *sc = pi->adapter; 2948218792Snp struct sge_eq *eq = &txq->eq; 2949218792Snp 2950228561Snp rc = free_eq(sc, eq); 2951228561Snp if (rc) 2952228561Snp return (rc); 2953220649Snp 2954220873Snp free(txq->sdesc, M_CXGBE); 2955218792Snp 2956228561Snp if (txq->txmaps.maps) 2957228561Snp t4_free_tx_maps(&txq->txmaps, txq->tx_tag); 2958218792Snp 2959220873Snp buf_ring_free(txq->br, M_CXGBE); 2960218792Snp 2961220873Snp if (txq->tx_tag) 2962220873Snp bus_dma_tag_destroy(txq->tx_tag); 2963218792Snp 2964218792Snp bzero(txq, sizeof(*txq)); 2965218792Snp return (0); 2966218792Snp} 2967218792Snp 2968218792Snpstatic void 2969218792Snponeseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error) 2970218792Snp{ 2971218792Snp bus_addr_t *ba = arg; 2972218792Snp 2973218792Snp KASSERT(nseg == 1, 2974218792Snp ("%s meant for single segment mappings only.", __func__)); 2975218792Snp 2976218792Snp *ba = error ? 0 : segs->ds_addr; 2977218792Snp} 2978218792Snp 2979218792Snpstatic inline bool 2980218792Snpis_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl) 2981218792Snp{ 2982218792Snp *ctrl = (void *)((uintptr_t)iq->cdesc + 2983218792Snp (iq->esize - sizeof(struct rsp_ctrl))); 2984218792Snp 2985218792Snp return (((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen); 2986218792Snp} 2987218792Snp 2988218792Snpstatic inline void 2989218792Snpiq_next(struct sge_iq *iq) 2990218792Snp{ 2991218792Snp iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize); 2992218792Snp if (__predict_false(++iq->cidx == iq->qsize - 1)) { 2993218792Snp iq->cidx = 0; 2994218792Snp iq->gen ^= 1; 2995218792Snp iq->cdesc = iq->desc; 2996218792Snp } 2997218792Snp} 2998218792Snp 2999220905Snp#define FL_HW_IDX(x) ((x) >> 3) 3000218792Snpstatic inline void 3001218792Snpring_fl_db(struct adapter *sc, struct sge_fl *fl) 3002218792Snp{ 3003218792Snp int ndesc = fl->pending / 8; 3004248925Snp uint32_t v; 3005218792Snp 3006220905Snp if (FL_HW_IDX(fl->pidx) == FL_HW_IDX(fl->cidx)) 3007220905Snp ndesc--; /* hold back one credit */ 3008218792Snp 3009220905Snp if (ndesc <= 0) 3010220905Snp return; /* nothing to do */ 3011220905Snp 3012248925Snp v = F_DBPRIO | V_QID(fl->cntxt_id) | V_PIDX(ndesc); 3013248925Snp if (is_t5(sc)) 3014248925Snp v |= F_DBTYPE; 3015248925Snp 3016218792Snp wmb(); 3017218792Snp 3018248925Snp t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v); 3019220905Snp fl->pending -= ndesc * 8; 3020218792Snp} 3021218792Snp 3022220905Snp/* 3023228561Snp * Fill up the freelist by upto nbufs and maybe ring its doorbell. 3024228561Snp * 3025228561Snp * Returns non-zero to indicate that it should be added to the list of starving 3026228561Snp * freelists. 3027220905Snp */ 3028228561Snpstatic int 3029228561Snprefill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs) 3030218792Snp{ 3031218792Snp __be64 *d = &fl->desc[fl->pidx]; 3032218792Snp struct fl_sdesc *sd = &fl->sdesc[fl->pidx]; 3033265425Snp uintptr_t pa; 3034218792Snp caddr_t cl; 3035265425Snp struct cluster_layout *cll = &fl->cll_def; /* default layout */ 3036265425Snp struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx]; 3037265425Snp struct cluster_metadata *clm; 3038218792Snp 3039218792Snp FL_LOCK_ASSERT_OWNED(fl); 3040218792Snp 3041228561Snp if (nbufs > fl->needed) 3042218792Snp nbufs = fl->needed; 3043265425Snp nbufs -= (fl->pidx + nbufs) % 8; 3044218792Snp 3045218792Snp while (nbufs--) { 3046218792Snp 3047218792Snp if (sd->cl != NULL) { 3048218792Snp 3049265425Snp if (sd->nmbuf == 0) { 3050255050Snp /* 3051265425Snp * Fast recycle without involving any atomics on 3052265425Snp * the cluster's metadata (if the cluster has 3053265425Snp * metadata). This happens when all frames 3054265425Snp * received in the cluster were small enough to 3055265425Snp * fit within a single mbuf each. 3056255050Snp */ 3057265425Snp fl->cl_fast_recycled++; 3058265425Snp goto recycled_fast; 3059255050Snp } 3060218792Snp 3061218792Snp /* 3062265425Snp * Cluster is guaranteed to have metadata. Clusters 3063265425Snp * without metadata always take the fast recycle path 3064265425Snp * when they're recycled. 3065218792Snp */ 3066265425Snp clm = cl_metadata(sc, fl, &sd->cll, sd->cl); 3067265425Snp MPASS(clm != NULL); 3068265425Snp 3069265425Snp if (atomic_fetchadd_int(&clm->refcount, -1) == 1) { 3070265425Snp fl->cl_recycled++; 3071265425Snp goto recycled; 3072218792Snp } 3073265425Snp sd->cl = NULL; /* gave up my reference */ 3074218792Snp } 3075265425Snp MPASS(sd->cl == NULL); 3076265425Snpalloc: 3077265425Snp cl = uma_zalloc(swz->zone, M_NOWAIT); 3078265425Snp if (__predict_false(cl == NULL)) { 3079265425Snp if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 || 3080265425Snp fl->cll_def.zidx == fl->cll_alt.zidx) 3081265425Snp break; 3082218792Snp 3083265425Snp /* fall back to the safe zone */ 3084265425Snp cll = &fl->cll_alt; 3085265425Snp swz = &sc->sge.sw_zone_info[cll->zidx]; 3086265425Snp goto alloc; 3087255050Snp } 3088265425Snp fl->cl_allocated++; 3089218792Snp 3090265425Snp pa = pmap_kextract((vm_offset_t)cl); 3091265425Snp pa += cll->region1; 3092218792Snp sd->cl = cl; 3093265425Snp sd->cll = *cll; 3094265425Snp *d = htobe64(pa | cll->hwidx); 3095265425Snp clm = cl_metadata(sc, fl, cll, cl); 3096265425Snp if (clm != NULL) { 3097265425Snprecycled: 3098218792Snp#ifdef INVARIANTS 3099265425Snp clm->sd = sd; 3100218792Snp#endif 3101265425Snp clm->refcount = 1; 3102265425Snp } 3103265425Snp sd->nmbuf = 0; 3104265425Snprecycled_fast: 3105219290Snp fl->pending++; 3106218792Snp fl->needed--; 3107265425Snp d++; 3108218792Snp sd++; 3109265425Snp if (__predict_false(++fl->pidx == fl->cap)) { 3110218792Snp fl->pidx = 0; 3111218792Snp sd = fl->sdesc; 3112218792Snp d = fl->desc; 3113218792Snp } 3114218792Snp } 3115220905Snp 3116228561Snp if (fl->pending >= 8) 3117220905Snp ring_fl_db(sc, fl); 3118228561Snp 3119228561Snp return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING)); 3120218792Snp} 3121218792Snp 3122228561Snp/* 3123228561Snp * Attempt to refill all starving freelists. 3124228561Snp */ 3125228561Snpstatic void 3126228561Snprefill_sfl(void *arg) 3127228561Snp{ 3128228561Snp struct adapter *sc = arg; 3129228561Snp struct sge_fl *fl, *fl_temp; 3130228561Snp 3131228561Snp mtx_lock(&sc->sfl_lock); 3132228561Snp TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) { 3133228561Snp FL_LOCK(fl); 3134228561Snp refill_fl(sc, fl, 64); 3135228561Snp if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) { 3136228561Snp TAILQ_REMOVE(&sc->sfl, fl, link); 3137228561Snp fl->flags &= ~FL_STARVING; 3138228561Snp } 3139228561Snp FL_UNLOCK(fl); 3140228561Snp } 3141228561Snp 3142228561Snp if (!TAILQ_EMPTY(&sc->sfl)) 3143228561Snp callout_schedule(&sc->sfl_callout, hz / 5); 3144228561Snp mtx_unlock(&sc->sfl_lock); 3145228561Snp} 3146228561Snp 3147218792Snpstatic int 3148218792Snpalloc_fl_sdesc(struct sge_fl *fl) 3149218792Snp{ 3150218792Snp 3151218792Snp fl->sdesc = malloc(fl->cap * sizeof(struct fl_sdesc), M_CXGBE, 3152218792Snp M_ZERO | M_WAITOK); 3153218792Snp 3154218792Snp return (0); 3155218792Snp} 3156218792Snp 3157218792Snpstatic void 3158255050Snpfree_fl_sdesc(struct adapter *sc, struct sge_fl *fl) 3159218792Snp{ 3160218792Snp struct fl_sdesc *sd; 3161265425Snp struct cluster_metadata *clm; 3162265425Snp struct cluster_layout *cll; 3163218792Snp int i; 3164218792Snp 3165218792Snp sd = fl->sdesc; 3166218792Snp for (i = 0; i < fl->cap; i++, sd++) { 3167265425Snp if (sd->cl == NULL) 3168265425Snp continue; 3169218792Snp 3170265425Snp cll = &sd->cll; 3171265425Snp clm = cl_metadata(sc, fl, cll, sd->cl); 3172265425Snp if (sd->nmbuf == 0 || 3173265425Snp (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1)) { 3174265425Snp uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl); 3175218792Snp } 3176265425Snp sd->cl = NULL; 3177218792Snp } 3178218792Snp 3179218792Snp free(fl->sdesc, M_CXGBE); 3180218792Snp fl->sdesc = NULL; 3181218792Snp} 3182218792Snp 3183228561Snpint 3184228561Snpt4_alloc_tx_maps(struct tx_maps *txmaps, bus_dma_tag_t tx_tag, int count, 3185228561Snp int flags) 3186218792Snp{ 3187218792Snp struct tx_map *txm; 3188228561Snp int i, rc; 3189218792Snp 3190228561Snp txmaps->map_total = txmaps->map_avail = count; 3191228561Snp txmaps->map_cidx = txmaps->map_pidx = 0; 3192218792Snp 3193228561Snp txmaps->maps = malloc(count * sizeof(struct tx_map), M_CXGBE, 3194228561Snp M_ZERO | flags); 3195218792Snp 3196228561Snp txm = txmaps->maps; 3197218792Snp for (i = 0; i < count; i++, txm++) { 3198228561Snp rc = bus_dmamap_create(tx_tag, 0, &txm->map); 3199218792Snp if (rc != 0) 3200218792Snp goto failed; 3201218792Snp } 3202218792Snp 3203218792Snp return (0); 3204218792Snpfailed: 3205218792Snp while (--i >= 0) { 3206218792Snp txm--; 3207228561Snp bus_dmamap_destroy(tx_tag, txm->map); 3208218792Snp } 3209228561Snp KASSERT(txm == txmaps->maps, ("%s: EDOOFUS", __func__)); 3210218792Snp 3211228561Snp free(txmaps->maps, M_CXGBE); 3212228561Snp txmaps->maps = NULL; 3213218792Snp 3214218792Snp return (rc); 3215218792Snp} 3216218792Snp 3217228561Snpvoid 3218228561Snpt4_free_tx_maps(struct tx_maps *txmaps, bus_dma_tag_t tx_tag) 3219218792Snp{ 3220218792Snp struct tx_map *txm; 3221218792Snp int i; 3222218792Snp 3223228561Snp txm = txmaps->maps; 3224228561Snp for (i = 0; i < txmaps->map_total; i++, txm++) { 3225218792Snp 3226218792Snp if (txm->m) { 3227228561Snp bus_dmamap_unload(tx_tag, txm->map); 3228218792Snp m_freem(txm->m); 3229218792Snp txm->m = NULL; 3230218792Snp } 3231218792Snp 3232228561Snp bus_dmamap_destroy(tx_tag, txm->map); 3233218792Snp } 3234218792Snp 3235228561Snp free(txmaps->maps, M_CXGBE); 3236228561Snp txmaps->maps = NULL; 3237218792Snp} 3238218792Snp 3239218792Snp/* 3240218792Snp * We'll do immediate data tx for non-TSO, but only when not coalescing. We're 3241218792Snp * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes 3242218792Snp * of immediate data. 3243218792Snp */ 3244218792Snp#define IMM_LEN ( \ 3245228561Snp 2 * EQ_ESIZE \ 3246218792Snp - sizeof(struct fw_eth_tx_pkt_wr) \ 3247218792Snp - sizeof(struct cpl_tx_pkt_core)) 3248218792Snp 3249218792Snp/* 3250218792Snp * Returns non-zero on failure, no need to cleanup anything in that case. 3251218792Snp * 3252218792Snp * Note 1: We always try to defrag the mbuf if required and return EFBIG only 3253218792Snp * if the resulting chain still won't fit in a tx descriptor. 3254218792Snp * 3255218792Snp * Note 2: We'll pullup the mbuf chain if TSO is requested and the first mbuf 3256218792Snp * does not have the TCP header in it. 3257218792Snp */ 3258218792Snpstatic int 3259218792Snpget_pkt_sgl(struct sge_txq *txq, struct mbuf **fp, struct sgl *sgl, 3260218792Snp int sgl_only) 3261218792Snp{ 3262218792Snp struct mbuf *m = *fp; 3263228561Snp struct tx_maps *txmaps; 3264218792Snp struct tx_map *txm; 3265218792Snp int rc, defragged = 0, n; 3266218792Snp 3267218792Snp TXQ_LOCK_ASSERT_OWNED(txq); 3268218792Snp 3269218792Snp if (m->m_pkthdr.tso_segsz) 3270218792Snp sgl_only = 1; /* Do not allow immediate data with LSO */ 3271218792Snp 3272218792Snpstart: sgl->nsegs = 0; 3273218792Snp 3274218792Snp if (m->m_pkthdr.len <= IMM_LEN && !sgl_only) 3275218792Snp return (0); /* nsegs = 0 tells caller to use imm. tx */ 3276218792Snp 3277228561Snp txmaps = &txq->txmaps; 3278228561Snp if (txmaps->map_avail == 0) { 3279218792Snp txq->no_dmamap++; 3280218792Snp return (ENOMEM); 3281218792Snp } 3282228561Snp txm = &txmaps->maps[txmaps->map_pidx]; 3283218792Snp 3284218792Snp if (m->m_pkthdr.tso_segsz && m->m_len < 50) { 3285218792Snp *fp = m_pullup(m, 50); 3286218792Snp m = *fp; 3287218792Snp if (m == NULL) 3288218792Snp return (ENOBUFS); 3289218792Snp } 3290218792Snp 3291220873Snp rc = bus_dmamap_load_mbuf_sg(txq->tx_tag, txm->map, m, sgl->seg, 3292218792Snp &sgl->nsegs, BUS_DMA_NOWAIT); 3293218792Snp if (rc == EFBIG && defragged == 0) { 3294243857Sglebius m = m_defrag(m, M_NOWAIT); 3295218792Snp if (m == NULL) 3296218792Snp return (EFBIG); 3297218792Snp 3298218792Snp defragged = 1; 3299218792Snp *fp = m; 3300218792Snp goto start; 3301218792Snp } 3302218792Snp if (rc != 0) 3303218792Snp return (rc); 3304218792Snp 3305218792Snp txm->m = m; 3306228561Snp txmaps->map_avail--; 3307228561Snp if (++txmaps->map_pidx == txmaps->map_total) 3308228561Snp txmaps->map_pidx = 0; 3309218792Snp 3310218792Snp KASSERT(sgl->nsegs > 0 && sgl->nsegs <= TX_SGL_SEGS, 3311218792Snp ("%s: bad DMA mapping (%d segments)", __func__, sgl->nsegs)); 3312218792Snp 3313218792Snp /* 3314218792Snp * Store the # of flits required to hold this frame's SGL in nflits. An 3315218792Snp * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by 3316218792Snp * multiple (len0 + len1, addr0, addr1) tuples. If addr1 is not used 3317218792Snp * then len1 must be set to 0. 3318218792Snp */ 3319218792Snp n = sgl->nsegs - 1; 3320218792Snp sgl->nflits = (3 * n) / 2 + (n & 1) + 2; 3321218792Snp 3322218792Snp return (0); 3323218792Snp} 3324218792Snp 3325218792Snp 3326218792Snp/* 3327218792Snp * Releases all the txq resources used up in the specified sgl. 3328218792Snp */ 3329218792Snpstatic int 3330218792Snpfree_pkt_sgl(struct sge_txq *txq, struct sgl *sgl) 3331218792Snp{ 3332228561Snp struct tx_maps *txmaps; 3333218792Snp struct tx_map *txm; 3334218792Snp 3335218792Snp TXQ_LOCK_ASSERT_OWNED(txq); 3336218792Snp 3337218792Snp if (sgl->nsegs == 0) 3338218792Snp return (0); /* didn't use any map */ 3339218792Snp 3340228561Snp txmaps = &txq->txmaps; 3341228561Snp 3342218792Snp /* 1 pkt uses exactly 1 map, back it out */ 3343218792Snp 3344228561Snp txmaps->map_avail++; 3345228561Snp if (txmaps->map_pidx > 0) 3346228561Snp txmaps->map_pidx--; 3347218792Snp else 3348228561Snp txmaps->map_pidx = txmaps->map_total - 1; 3349218792Snp 3350228561Snp txm = &txmaps->maps[txmaps->map_pidx]; 3351220873Snp bus_dmamap_unload(txq->tx_tag, txm->map); 3352218792Snp txm->m = NULL; 3353218792Snp 3354218792Snp return (0); 3355218792Snp} 3356218792Snp 3357218792Snpstatic int 3358218792Snpwrite_txpkt_wr(struct port_info *pi, struct sge_txq *txq, struct mbuf *m, 3359218792Snp struct sgl *sgl) 3360218792Snp{ 3361218792Snp struct sge_eq *eq = &txq->eq; 3362218792Snp struct fw_eth_tx_pkt_wr *wr; 3363218792Snp struct cpl_tx_pkt_core *cpl; 3364218792Snp uint32_t ctrl; /* used in many unrelated places */ 3365218792Snp uint64_t ctrl1; 3366219286Snp int nflits, ndesc, pktlen; 3367218792Snp struct tx_sdesc *txsd; 3368218792Snp caddr_t dst; 3369218792Snp 3370218792Snp TXQ_LOCK_ASSERT_OWNED(txq); 3371218792Snp 3372219286Snp pktlen = m->m_pkthdr.len; 3373219286Snp 3374218792Snp /* 3375218792Snp * Do we have enough flits to send this frame out? 3376218792Snp */ 3377218792Snp ctrl = sizeof(struct cpl_tx_pkt_core); 3378218792Snp if (m->m_pkthdr.tso_segsz) { 3379218792Snp nflits = TXPKT_LSO_WR_HDR; 3380237436Snp ctrl += sizeof(struct cpl_tx_pkt_lso_core); 3381218792Snp } else 3382218792Snp nflits = TXPKT_WR_HDR; 3383218792Snp if (sgl->nsegs > 0) 3384218792Snp nflits += sgl->nflits; 3385218792Snp else { 3386219286Snp nflits += howmany(pktlen, 8); 3387219286Snp ctrl += pktlen; 3388218792Snp } 3389218792Snp ndesc = howmany(nflits, 8); 3390218792Snp if (ndesc > eq->avail) 3391218792Snp return (ENOMEM); 3392218792Snp 3393218792Snp /* Firmware work request header */ 3394218792Snp wr = (void *)&eq->desc[eq->pidx]; 3395218792Snp wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) | 3396228561Snp V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl)); 3397218792Snp ctrl = V_FW_WR_LEN16(howmany(nflits, 2)); 3398228561Snp if (eq->avail == ndesc) { 3399228561Snp if (!(eq->flags & EQ_CRFLUSHED)) { 3400228561Snp ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ; 3401228561Snp eq->flags |= EQ_CRFLUSHED; 3402228561Snp } 3403228561Snp eq->flags |= EQ_STALLED; 3404220643Snp } 3405220643Snp 3406218792Snp wr->equiq_to_len16 = htobe32(ctrl); 3407218792Snp wr->r3 = 0; 3408218792Snp 3409218792Snp if (m->m_pkthdr.tso_segsz) { 3410237436Snp struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1); 3411218792Snp struct ether_header *eh; 3412237819Snp void *l3hdr; 3413237819Snp#if defined(INET) || defined(INET6) 3414218792Snp struct tcphdr *tcp; 3415237819Snp#endif 3416237819Snp uint16_t eh_type; 3417218792Snp 3418218792Snp ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE | 3419218792Snp F_LSO_LAST_SLICE; 3420218792Snp 3421218792Snp eh = mtod(m, struct ether_header *); 3422237819Snp eh_type = ntohs(eh->ether_type); 3423237819Snp if (eh_type == ETHERTYPE_VLAN) { 3424237819Snp struct ether_vlan_header *evh = (void *)eh; 3425237819Snp 3426218792Snp ctrl |= V_LSO_ETHHDR_LEN(1); 3427237819Snp l3hdr = evh + 1; 3428237819Snp eh_type = ntohs(evh->evl_proto); 3429218792Snp } else 3430237819Snp l3hdr = eh + 1; 3431218792Snp 3432237819Snp switch (eh_type) { 3433237819Snp#ifdef INET6 3434237819Snp case ETHERTYPE_IPV6: 3435237819Snp { 3436237819Snp struct ip6_hdr *ip6 = l3hdr; 3437218792Snp 3438237819Snp /* 3439237819Snp * XXX-BZ For now we do not pretend to support 3440237819Snp * IPv6 extension headers. 3441237819Snp */ 3442237819Snp KASSERT(ip6->ip6_nxt == IPPROTO_TCP, ("%s: CSUM_TSO " 3443237819Snp "with ip6_nxt != TCP: %u", __func__, ip6->ip6_nxt)); 3444237819Snp tcp = (struct tcphdr *)(ip6 + 1); 3445237819Snp ctrl |= F_LSO_IPV6; 3446237819Snp ctrl |= V_LSO_IPHDR_LEN(sizeof(*ip6) >> 2) | 3447237819Snp V_LSO_TCPHDR_LEN(tcp->th_off); 3448237819Snp break; 3449237819Snp } 3450237819Snp#endif 3451237819Snp#ifdef INET 3452237819Snp case ETHERTYPE_IP: 3453237819Snp { 3454237819Snp struct ip *ip = l3hdr; 3455237819Snp 3456237819Snp tcp = (void *)((uintptr_t)ip + ip->ip_hl * 4); 3457237819Snp ctrl |= V_LSO_IPHDR_LEN(ip->ip_hl) | 3458237819Snp V_LSO_TCPHDR_LEN(tcp->th_off); 3459237819Snp break; 3460237819Snp } 3461237819Snp#endif 3462237819Snp default: 3463237819Snp panic("%s: CSUM_TSO but no supported IP version " 3464237819Snp "(0x%04x)", __func__, eh_type); 3465237819Snp } 3466237819Snp 3467218792Snp lso->lso_ctrl = htobe32(ctrl); 3468218792Snp lso->ipid_ofst = htobe16(0); 3469218792Snp lso->mss = htobe16(m->m_pkthdr.tso_segsz); 3470218792Snp lso->seqno_offset = htobe32(0); 3471219286Snp lso->len = htobe32(pktlen); 3472218792Snp 3473218792Snp cpl = (void *)(lso + 1); 3474218792Snp 3475218792Snp txq->tso_wrs++; 3476218792Snp } else 3477218792Snp cpl = (void *)(wr + 1); 3478218792Snp 3479218792Snp /* Checksum offload */ 3480218792Snp ctrl1 = 0; 3481247062Snp if (!(m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO))) 3482218792Snp ctrl1 |= F_TXPKT_IPCSUM_DIS; 3483237799Snp if (!(m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 | 3484247062Snp CSUM_TCP_IPV6 | CSUM_TSO))) 3485218792Snp ctrl1 |= F_TXPKT_L4CSUM_DIS; 3486237799Snp if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | 3487247062Snp CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO)) 3488218792Snp txq->txcsum++; /* some hardware assistance provided */ 3489218792Snp 3490218792Snp /* VLAN tag insertion */ 3491218792Snp if (m->m_flags & M_VLANTAG) { 3492218792Snp ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag); 3493218792Snp txq->vlan_insertion++; 3494218792Snp } 3495218792Snp 3496218792Snp /* CPL header */ 3497218792Snp cpl->ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) | 3498218792Snp V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf)); 3499218792Snp cpl->pack = 0; 3500219286Snp cpl->len = htobe16(pktlen); 3501218792Snp cpl->ctrl1 = htobe64(ctrl1); 3502218792Snp 3503218792Snp /* Software descriptor */ 3504220873Snp txsd = &txq->sdesc[eq->pidx]; 3505218792Snp txsd->desc_used = ndesc; 3506218792Snp 3507218792Snp eq->pending += ndesc; 3508218792Snp eq->avail -= ndesc; 3509218792Snp eq->pidx += ndesc; 3510218792Snp if (eq->pidx >= eq->cap) 3511218792Snp eq->pidx -= eq->cap; 3512218792Snp 3513218792Snp /* SGL */ 3514218792Snp dst = (void *)(cpl + 1); 3515218792Snp if (sgl->nsegs > 0) { 3516220873Snp txsd->credits = 1; 3517218792Snp txq->sgl_wrs++; 3518218792Snp write_sgl_to_txd(eq, sgl, &dst); 3519218792Snp } else { 3520220873Snp txsd->credits = 0; 3521218792Snp txq->imm_wrs++; 3522218792Snp for (; m; m = m->m_next) { 3523218792Snp copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len); 3524219286Snp#ifdef INVARIANTS 3525219286Snp pktlen -= m->m_len; 3526219286Snp#endif 3527218792Snp } 3528219286Snp#ifdef INVARIANTS 3529219286Snp KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen)); 3530219286Snp#endif 3531219286Snp 3532218792Snp } 3533218792Snp 3534218792Snp txq->txpkt_wrs++; 3535218792Snp return (0); 3536218792Snp} 3537218792Snp 3538218792Snp/* 3539218792Snp * Returns 0 to indicate that m has been accepted into a coalesced tx work 3540218792Snp * request. It has either been folded into txpkts or txpkts was flushed and m 3541218792Snp * has started a new coalesced work request (as the first frame in a fresh 3542218792Snp * txpkts). 3543218792Snp * 3544218792Snp * Returns non-zero to indicate a failure - caller is responsible for 3545218792Snp * transmitting m, if there was anything in txpkts it has been flushed. 3546218792Snp */ 3547218792Snpstatic int 3548218792Snpadd_to_txpkts(struct port_info *pi, struct sge_txq *txq, struct txpkts *txpkts, 3549218792Snp struct mbuf *m, struct sgl *sgl) 3550218792Snp{ 3551218792Snp struct sge_eq *eq = &txq->eq; 3552218792Snp int can_coalesce; 3553218792Snp struct tx_sdesc *txsd; 3554218792Snp int flits; 3555218792Snp 3556218792Snp TXQ_LOCK_ASSERT_OWNED(txq); 3557218792Snp 3558228561Snp KASSERT(sgl->nsegs, ("%s: can't coalesce imm data", __func__)); 3559228561Snp 3560218792Snp if (txpkts->npkt > 0) { 3561218792Snp flits = TXPKTS_PKT_HDR + sgl->nflits; 3562218792Snp can_coalesce = m->m_pkthdr.tso_segsz == 0 && 3563218792Snp txpkts->nflits + flits <= TX_WR_FLITS && 3564218792Snp txpkts->nflits + flits <= eq->avail * 8 && 3565218792Snp txpkts->plen + m->m_pkthdr.len < 65536; 3566218792Snp 3567218792Snp if (can_coalesce) { 3568218792Snp txpkts->npkt++; 3569218792Snp txpkts->nflits += flits; 3570218792Snp txpkts->plen += m->m_pkthdr.len; 3571218792Snp 3572220873Snp txsd = &txq->sdesc[eq->pidx]; 3573220873Snp txsd->credits++; 3574218792Snp 3575218792Snp return (0); 3576218792Snp } 3577218792Snp 3578218792Snp /* 3579218792Snp * Couldn't coalesce m into txpkts. The first order of business 3580218792Snp * is to send txpkts on its way. Then we'll revisit m. 3581218792Snp */ 3582218792Snp write_txpkts_wr(txq, txpkts); 3583218792Snp } 3584218792Snp 3585218792Snp /* 3586218792Snp * Check if we can start a new coalesced tx work request with m as 3587218792Snp * the first packet in it. 3588218792Snp */ 3589218792Snp 3590218792Snp KASSERT(txpkts->npkt == 0, ("%s: txpkts not empty", __func__)); 3591218792Snp 3592218792Snp flits = TXPKTS_WR_HDR + sgl->nflits; 3593218792Snp can_coalesce = m->m_pkthdr.tso_segsz == 0 && 3594218792Snp flits <= eq->avail * 8 && flits <= TX_WR_FLITS; 3595218792Snp 3596218792Snp if (can_coalesce == 0) 3597218792Snp return (EINVAL); 3598218792Snp 3599218792Snp /* 3600218792Snp * Start a fresh coalesced tx WR with m as the first frame in it. 3601218792Snp */ 3602218792Snp txpkts->npkt = 1; 3603218792Snp txpkts->nflits = flits; 3604218792Snp txpkts->flitp = &eq->desc[eq->pidx].flit[2]; 3605218792Snp txpkts->plen = m->m_pkthdr.len; 3606218792Snp 3607220873Snp txsd = &txq->sdesc[eq->pidx]; 3608220873Snp txsd->credits = 1; 3609218792Snp 3610218792Snp return (0); 3611218792Snp} 3612218792Snp 3613218792Snp/* 3614218792Snp * Note that write_txpkts_wr can never run out of hardware descriptors (but 3615218792Snp * write_txpkt_wr can). add_to_txpkts ensures that a frame is accepted for 3616218792Snp * coalescing only if sufficient hardware descriptors are available. 3617218792Snp */ 3618218792Snpstatic void 3619218792Snpwrite_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts) 3620218792Snp{ 3621218792Snp struct sge_eq *eq = &txq->eq; 3622218792Snp struct fw_eth_tx_pkts_wr *wr; 3623218792Snp struct tx_sdesc *txsd; 3624218792Snp uint32_t ctrl; 3625218792Snp int ndesc; 3626218792Snp 3627218792Snp TXQ_LOCK_ASSERT_OWNED(txq); 3628218792Snp 3629218792Snp ndesc = howmany(txpkts->nflits, 8); 3630218792Snp 3631218792Snp wr = (void *)&eq->desc[eq->pidx]; 3632228561Snp wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)); 3633218792Snp ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2)); 3634228561Snp if (eq->avail == ndesc) { 3635228561Snp if (!(eq->flags & EQ_CRFLUSHED)) { 3636228561Snp ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ; 3637228561Snp eq->flags |= EQ_CRFLUSHED; 3638228561Snp } 3639228561Snp eq->flags |= EQ_STALLED; 3640220643Snp } 3641218792Snp wr->equiq_to_len16 = htobe32(ctrl); 3642218792Snp wr->plen = htobe16(txpkts->plen); 3643218792Snp wr->npkt = txpkts->npkt; 3644222513Snp wr->r3 = wr->type = 0; 3645218792Snp 3646218792Snp /* Everything else already written */ 3647218792Snp 3648220873Snp txsd = &txq->sdesc[eq->pidx]; 3649218792Snp txsd->desc_used = ndesc; 3650218792Snp 3651220643Snp KASSERT(eq->avail >= ndesc, ("%s: out of descriptors", __func__)); 3652218792Snp 3653218792Snp eq->pending += ndesc; 3654218792Snp eq->avail -= ndesc; 3655218792Snp eq->pidx += ndesc; 3656218792Snp if (eq->pidx >= eq->cap) 3657218792Snp eq->pidx -= eq->cap; 3658218792Snp 3659218792Snp txq->txpkts_pkts += txpkts->npkt; 3660218792Snp txq->txpkts_wrs++; 3661218792Snp txpkts->npkt = 0; /* emptied */ 3662218792Snp} 3663218792Snp 3664218792Snpstatic inline void 3665218792Snpwrite_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq, 3666218792Snp struct txpkts *txpkts, struct mbuf *m, struct sgl *sgl) 3667218792Snp{ 3668218792Snp struct ulp_txpkt *ulpmc; 3669218792Snp struct ulptx_idata *ulpsc; 3670218792Snp struct cpl_tx_pkt_core *cpl; 3671218792Snp struct sge_eq *eq = &txq->eq; 3672218792Snp uintptr_t flitp, start, end; 3673218792Snp uint64_t ctrl; 3674218792Snp caddr_t dst; 3675218792Snp 3676218792Snp KASSERT(txpkts->npkt > 0, ("%s: txpkts is empty", __func__)); 3677218792Snp 3678218792Snp start = (uintptr_t)eq->desc; 3679218792Snp end = (uintptr_t)eq->spg; 3680218792Snp 3681218792Snp /* Checksum offload */ 3682218792Snp ctrl = 0; 3683247062Snp if (!(m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO))) 3684218792Snp ctrl |= F_TXPKT_IPCSUM_DIS; 3685247062Snp if (!(m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 | 3686247062Snp CSUM_TCP_IPV6 | CSUM_TSO))) 3687218792Snp ctrl |= F_TXPKT_L4CSUM_DIS; 3688247062Snp if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | 3689247062Snp CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO)) 3690218792Snp txq->txcsum++; /* some hardware assistance provided */ 3691218792Snp 3692218792Snp /* VLAN tag insertion */ 3693218792Snp if (m->m_flags & M_VLANTAG) { 3694218792Snp ctrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag); 3695218792Snp txq->vlan_insertion++; 3696218792Snp } 3697218792Snp 3698218792Snp /* 3699218792Snp * The previous packet's SGL must have ended at a 16 byte boundary (this 3700218792Snp * is required by the firmware/hardware). It follows that flitp cannot 3701218792Snp * wrap around between the ULPTX master command and ULPTX subcommand (8 3702218792Snp * bytes each), and that it can not wrap around in the middle of the 3703218792Snp * cpl_tx_pkt_core either. 3704218792Snp */ 3705218792Snp flitp = (uintptr_t)txpkts->flitp; 3706218792Snp KASSERT((flitp & 0xf) == 0, 3707218792Snp ("%s: last SGL did not end at 16 byte boundary: %p", 3708218792Snp __func__, txpkts->flitp)); 3709218792Snp 3710218792Snp /* ULP master command */ 3711218792Snp ulpmc = (void *)flitp; 3712219288Snp ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0) | 3713219288Snp V_ULP_TXPKT_FID(eq->iqid)); 3714218792Snp ulpmc->len = htonl(howmany(sizeof(*ulpmc) + sizeof(*ulpsc) + 3715218792Snp sizeof(*cpl) + 8 * sgl->nflits, 16)); 3716218792Snp 3717218792Snp /* ULP subcommand */ 3718218792Snp ulpsc = (void *)(ulpmc + 1); 3719218792Snp ulpsc->cmd_more = htobe32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) | 3720218792Snp F_ULP_TX_SC_MORE); 3721218792Snp ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core)); 3722218792Snp 3723218792Snp flitp += sizeof(*ulpmc) + sizeof(*ulpsc); 3724218792Snp if (flitp == end) 3725218792Snp flitp = start; 3726218792Snp 3727218792Snp /* CPL_TX_PKT */ 3728218792Snp cpl = (void *)flitp; 3729218792Snp cpl->ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) | 3730218792Snp V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf)); 3731218792Snp cpl->pack = 0; 3732218792Snp cpl->len = htobe16(m->m_pkthdr.len); 3733218792Snp cpl->ctrl1 = htobe64(ctrl); 3734218792Snp 3735218792Snp flitp += sizeof(*cpl); 3736218792Snp if (flitp == end) 3737218792Snp flitp = start; 3738218792Snp 3739218792Snp /* SGL for this frame */ 3740218792Snp dst = (caddr_t)flitp; 3741218792Snp txpkts->nflits += write_sgl_to_txd(eq, sgl, &dst); 3742218792Snp txpkts->flitp = (void *)dst; 3743218792Snp 3744218792Snp KASSERT(((uintptr_t)dst & 0xf) == 0, 3745218792Snp ("%s: SGL ends at %p (not a 16 byte boundary)", __func__, dst)); 3746218792Snp} 3747218792Snp 3748218792Snp/* 3749218792Snp * If the SGL ends on an address that is not 16 byte aligned, this function will 3750218792Snp * add a 0 filled flit at the end. It returns 1 in that case. 3751218792Snp */ 3752218792Snpstatic int 3753218792Snpwrite_sgl_to_txd(struct sge_eq *eq, struct sgl *sgl, caddr_t *to) 3754218792Snp{ 3755218792Snp __be64 *flitp, *end; 3756218792Snp struct ulptx_sgl *usgl; 3757218792Snp bus_dma_segment_t *seg; 3758218792Snp int i, padded; 3759218792Snp 3760218792Snp KASSERT(sgl->nsegs > 0 && sgl->nflits > 0, 3761218792Snp ("%s: bad SGL - nsegs=%d, nflits=%d", 3762218792Snp __func__, sgl->nsegs, sgl->nflits)); 3763218792Snp 3764218792Snp KASSERT(((uintptr_t)(*to) & 0xf) == 0, 3765218792Snp ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to)); 3766218792Snp 3767218792Snp flitp = (__be64 *)(*to); 3768218792Snp end = flitp + sgl->nflits; 3769218792Snp seg = &sgl->seg[0]; 3770218792Snp usgl = (void *)flitp; 3771218792Snp 3772218792Snp /* 3773218792Snp * We start at a 16 byte boundary somewhere inside the tx descriptor 3774218792Snp * ring, so we're at least 16 bytes away from the status page. There is 3775218792Snp * no chance of a wrap around in the middle of usgl (which is 16 bytes). 3776218792Snp */ 3777218792Snp 3778218792Snp usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | 3779218792Snp V_ULPTX_NSGE(sgl->nsegs)); 3780218792Snp usgl->len0 = htobe32(seg->ds_len); 3781218792Snp usgl->addr0 = htobe64(seg->ds_addr); 3782218792Snp seg++; 3783218792Snp 3784218792Snp if ((uintptr_t)end <= (uintptr_t)eq->spg) { 3785218792Snp 3786218792Snp /* Won't wrap around at all */ 3787218792Snp 3788218792Snp for (i = 0; i < sgl->nsegs - 1; i++, seg++) { 3789218792Snp usgl->sge[i / 2].len[i & 1] = htobe32(seg->ds_len); 3790218792Snp usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ds_addr); 3791218792Snp } 3792218792Snp if (i & 1) 3793218792Snp usgl->sge[i / 2].len[1] = htobe32(0); 3794218792Snp } else { 3795218792Snp 3796218792Snp /* Will wrap somewhere in the rest of the SGL */ 3797218792Snp 3798218792Snp /* 2 flits already written, write the rest flit by flit */ 3799218792Snp flitp = (void *)(usgl + 1); 3800218792Snp for (i = 0; i < sgl->nflits - 2; i++) { 3801218792Snp if ((uintptr_t)flitp == (uintptr_t)eq->spg) 3802218792Snp flitp = (void *)eq->desc; 3803218792Snp *flitp++ = get_flit(seg, sgl->nsegs - 1, i); 3804218792Snp } 3805218792Snp end = flitp; 3806218792Snp } 3807218792Snp 3808218792Snp if ((uintptr_t)end & 0xf) { 3809218792Snp *(uint64_t *)end = 0; 3810218792Snp end++; 3811218792Snp padded = 1; 3812218792Snp } else 3813218792Snp padded = 0; 3814218792Snp 3815218792Snp if ((uintptr_t)end == (uintptr_t)eq->spg) 3816218792Snp *to = (void *)eq->desc; 3817218792Snp else 3818218792Snp *to = (void *)end; 3819218792Snp 3820218792Snp return (padded); 3821218792Snp} 3822218792Snp 3823218792Snpstatic inline void 3824218792Snpcopy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len) 3825218792Snp{ 3826237263Snp if (__predict_true((uintptr_t)(*to) + len <= (uintptr_t)eq->spg)) { 3827218792Snp bcopy(from, *to, len); 3828218792Snp (*to) += len; 3829218792Snp } else { 3830218792Snp int portion = (uintptr_t)eq->spg - (uintptr_t)(*to); 3831218792Snp 3832218792Snp bcopy(from, *to, portion); 3833218792Snp from += portion; 3834218792Snp portion = len - portion; /* remaining */ 3835218792Snp bcopy(from, (void *)eq->desc, portion); 3836218792Snp (*to) = (caddr_t)eq->desc + portion; 3837218792Snp } 3838218792Snp} 3839218792Snp 3840218792Snpstatic inline void 3841220873Snpring_eq_db(struct adapter *sc, struct sge_eq *eq) 3842218792Snp{ 3843248925Snp u_int db, pending; 3844248925Snp 3845248925Snp db = eq->doorbells; 3846248925Snp pending = eq->pending; 3847248925Snp if (pending > 1) 3848249392Snp clrbit(&db, DOORBELL_WCWR); 3849248925Snp eq->pending = 0; 3850218792Snp wmb(); 3851248925Snp 3852248925Snp switch (ffs(db) - 1) { 3853248925Snp case DOORBELL_UDB: 3854248925Snp *eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(pending)); 3855248925Snp return; 3856248925Snp 3857249392Snp case DOORBELL_WCWR: { 3858248925Snp volatile uint64_t *dst, *src; 3859248925Snp int i; 3860248925Snp 3861248925Snp /* 3862248925Snp * Queues whose 128B doorbell segment fits in the page do not 3863248925Snp * use relative qid (udb_qid is always 0). Only queues with 3864249392Snp * doorbell segments can do WCWR. 3865248925Snp */ 3866248925Snp KASSERT(eq->udb_qid == 0 && pending == 1, 3867248925Snp ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p", 3868248925Snp __func__, eq->doorbells, pending, eq->pidx, eq)); 3869248925Snp 3870248925Snp dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET - 3871248925Snp UDBS_DB_OFFSET); 3872248925Snp i = eq->pidx ? eq->pidx - 1 : eq->cap - 1; 3873248925Snp src = (void *)&eq->desc[i]; 3874248925Snp while (src != (void *)&eq->desc[i + 1]) 3875248925Snp *dst++ = *src++; 3876248925Snp wmb(); 3877248925Snp return; 3878248925Snp } 3879248925Snp 3880248925Snp case DOORBELL_UDBWC: 3881248925Snp *eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(pending)); 3882248925Snp wmb(); 3883248925Snp return; 3884248925Snp 3885248925Snp case DOORBELL_KDB: 3886248925Snp t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), 3887248925Snp V_QID(eq->cntxt_id) | V_PIDX(pending)); 3888248925Snp return; 3889248925Snp } 3890218792Snp} 3891218792Snp 3892219292Snpstatic inline int 3893219292Snpreclaimable(struct sge_eq *eq) 3894218792Snp{ 3895219292Snp unsigned int cidx; 3896218792Snp 3897218792Snp cidx = eq->spg->cidx; /* stable snapshot */ 3898228561Snp cidx = be16toh(cidx); 3899218792Snp 3900218792Snp if (cidx >= eq->cidx) 3901219292Snp return (cidx - eq->cidx); 3902218792Snp else 3903219292Snp return (cidx + eq->cap - eq->cidx); 3904219292Snp} 3905218792Snp 3906219292Snp/* 3907219292Snp * There are "can_reclaim" tx descriptors ready to be reclaimed. Reclaim as 3908219292Snp * many as possible but stop when there are around "n" mbufs to free. 3909219292Snp * 3910219292Snp * The actual number reclaimed is provided as the return value. 3911219292Snp */ 3912219292Snpstatic int 3913220873Snpreclaim_tx_descs(struct sge_txq *txq, int can_reclaim, int n) 3914219292Snp{ 3915219292Snp struct tx_sdesc *txsd; 3916228561Snp struct tx_maps *txmaps; 3917219292Snp struct tx_map *txm; 3918219292Snp unsigned int reclaimed, maps; 3919220873Snp struct sge_eq *eq = &txq->eq; 3920218792Snp 3921228561Snp TXQ_LOCK_ASSERT_OWNED(txq); 3922218792Snp 3923219292Snp if (can_reclaim == 0) 3924219292Snp can_reclaim = reclaimable(eq); 3925219292Snp 3926218792Snp maps = reclaimed = 0; 3927219292Snp while (can_reclaim && maps < n) { 3928218792Snp int ndesc; 3929218792Snp 3930220873Snp txsd = &txq->sdesc[eq->cidx]; 3931218792Snp ndesc = txsd->desc_used; 3932218792Snp 3933218792Snp /* Firmware doesn't return "partial" credits. */ 3934218792Snp KASSERT(can_reclaim >= ndesc, 3935218792Snp ("%s: unexpected number of credits: %d, %d", 3936218792Snp __func__, can_reclaim, ndesc)); 3937218792Snp 3938220873Snp maps += txsd->credits; 3939219292Snp 3940218792Snp reclaimed += ndesc; 3941219292Snp can_reclaim -= ndesc; 3942218792Snp 3943218792Snp eq->cidx += ndesc; 3944219292Snp if (__predict_false(eq->cidx >= eq->cap)) 3945218792Snp eq->cidx -= eq->cap; 3946219292Snp } 3947218792Snp 3948228561Snp txmaps = &txq->txmaps; 3949228561Snp txm = &txmaps->maps[txmaps->map_cidx]; 3950219292Snp if (maps) 3951219292Snp prefetch(txm->m); 3952218792Snp 3953218792Snp eq->avail += reclaimed; 3954218792Snp KASSERT(eq->avail < eq->cap, /* avail tops out at (cap - 1) */ 3955218792Snp ("%s: too many descriptors available", __func__)); 3956218792Snp 3957228561Snp txmaps->map_avail += maps; 3958228561Snp KASSERT(txmaps->map_avail <= txmaps->map_total, 3959218792Snp ("%s: too many maps available", __func__)); 3960218792Snp 3961218792Snp while (maps--) { 3962219292Snp struct tx_map *next; 3963218792Snp 3964219292Snp next = txm + 1; 3965228561Snp if (__predict_false(txmaps->map_cidx + 1 == txmaps->map_total)) 3966228561Snp next = txmaps->maps; 3967219292Snp prefetch(next->m); 3968219292Snp 3969220873Snp bus_dmamap_unload(txq->tx_tag, txm->map); 3970218792Snp m_freem(txm->m); 3971218792Snp txm->m = NULL; 3972218792Snp 3973219292Snp txm = next; 3974228561Snp if (__predict_false(++txmaps->map_cidx == txmaps->map_total)) 3975228561Snp txmaps->map_cidx = 0; 3976218792Snp } 3977218792Snp 3978218792Snp return (reclaimed); 3979218792Snp} 3980218792Snp 3981218792Snpstatic void 3982218792Snpwrite_eqflush_wr(struct sge_eq *eq) 3983218792Snp{ 3984218792Snp struct fw_eq_flush_wr *wr; 3985218792Snp 3986218792Snp EQ_LOCK_ASSERT_OWNED(eq); 3987218792Snp KASSERT(eq->avail > 0, ("%s: no descriptors left.", __func__)); 3988228561Snp KASSERT(!(eq->flags & EQ_CRFLUSHED), ("%s: flushed already", __func__)); 3989218792Snp 3990218792Snp wr = (void *)&eq->desc[eq->pidx]; 3991218792Snp bzero(wr, sizeof(*wr)); 3992218792Snp wr->opcode = FW_EQ_FLUSH_WR; 3993218792Snp wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(sizeof(*wr) / 16) | 3994218792Snp F_FW_WR_EQUEQ | F_FW_WR_EQUIQ); 3995218792Snp 3996228561Snp eq->flags |= (EQ_CRFLUSHED | EQ_STALLED); 3997218792Snp eq->pending++; 3998218792Snp eq->avail--; 3999218792Snp if (++eq->pidx == eq->cap) 4000265425Snp eq->pidx = 0; 4001218792Snp} 4002218792Snp 4003218792Snpstatic __be64 4004218792Snpget_flit(bus_dma_segment_t *sgl, int nsegs, int idx) 4005218792Snp{ 4006218792Snp int i = (idx / 3) * 2; 4007218792Snp 4008218792Snp switch (idx % 3) { 4009218792Snp case 0: { 4010218792Snp __be64 rc; 4011218792Snp 4012218792Snp rc = htobe32(sgl[i].ds_len); 4013218792Snp if (i + 1 < nsegs) 4014218792Snp rc |= (uint64_t)htobe32(sgl[i + 1].ds_len) << 32; 4015218792Snp 4016218792Snp return (rc); 4017218792Snp } 4018218792Snp case 1: 4019218792Snp return htobe64(sgl[i].ds_addr); 4020218792Snp case 2: 4021218792Snp return htobe64(sgl[i + 1].ds_addr); 4022218792Snp } 4023218792Snp 4024218792Snp return (0); 4025218792Snp} 4026218792Snp 4027218792Snpstatic void 4028265425Snpfind_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp) 4029218792Snp{ 4030265425Snp int8_t zidx, hwidx, idx; 4031265425Snp uint16_t region1, region3; 4032265425Snp int spare, spare_needed, n; 4033265425Snp struct sw_zone_info *swz; 4034265425Snp struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0]; 4035218792Snp 4036265425Snp /* 4037265425Snp * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize 4038265425Snp * large enough for the max payload and cluster metadata. Otherwise 4039265425Snp * settle for the largest bufsize that leaves enough room in the cluster 4040265425Snp * for metadata. 4041265425Snp * 4042265425Snp * Without buffer packing: Look for the smallest zone which has a 4043265425Snp * bufsize large enough for the max payload. Settle for the largest 4044265425Snp * bufsize available if there's nothing big enough for max payload. 4045265425Snp */ 4046265425Snp spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0; 4047265425Snp swz = &sc->sge.sw_zone_info[0]; 4048265425Snp hwidx = -1; 4049265425Snp for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) { 4050265425Snp if (swz->size > largest_rx_cluster) { 4051265425Snp if (__predict_true(hwidx != -1)) 4052265425Snp break; 4053218792Snp 4054265425Snp /* 4055265425Snp * This is a misconfiguration. largest_rx_cluster is 4056265425Snp * preventing us from finding a refill source. See 4057265425Snp * dev.t5nex.<n>.buffer_sizes to figure out why. 4058265425Snp */ 4059265425Snp device_printf(sc->dev, "largest_rx_cluster=%u leaves no" 4060265425Snp " refill source for fl %p (dma %u). Ignored.\n", 4061265425Snp largest_rx_cluster, fl, maxp); 4062265425Snp } 4063265425Snp for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) { 4064265425Snp hwb = &hwb_list[idx]; 4065265425Snp spare = swz->size - hwb->size; 4066265425Snp if (spare < spare_needed) 4067265425Snp continue; 4068265425Snp 4069265425Snp hwidx = idx; /* best option so far */ 4070265425Snp if (hwb->size >= maxp) { 4071265425Snp 4072265425Snp if ((fl->flags & FL_BUF_PACKING) == 0) 4073265425Snp goto done; /* stop looking (not packing) */ 4074265425Snp 4075265425Snp if (swz->size >= safest_rx_cluster) 4076265425Snp goto done; /* stop looking (packing) */ 4077265425Snp } 4078265425Snp break; /* keep looking, next zone */ 4079265425Snp } 4080255050Snp } 4081265425Snpdone: 4082265425Snp /* A usable hwidx has been located. */ 4083265425Snp MPASS(hwidx != -1); 4084265425Snp hwb = &hwb_list[hwidx]; 4085265425Snp zidx = hwb->zidx; 4086265425Snp swz = &sc->sge.sw_zone_info[zidx]; 4087265425Snp region1 = 0; 4088265425Snp region3 = swz->size - hwb->size; 4089255050Snp 4090265425Snp /* 4091265425Snp * Stay within this zone and see if there is a better match when mbuf 4092265425Snp * inlining is allowed. Remember that the hwidx's are sorted in 4093265425Snp * decreasing order of size (so in increasing order of spare area). 4094265425Snp */ 4095265425Snp for (idx = hwidx; idx != -1; idx = hwb->next) { 4096265425Snp hwb = &hwb_list[idx]; 4097265425Snp spare = swz->size - hwb->size; 4098255050Snp 4099265425Snp if (allow_mbufs_in_cluster == 0 || hwb->size < maxp) 4100265425Snp break; 4101265425Snp if (spare < CL_METADATA_SIZE + MSIZE) 4102265425Snp continue; 4103265425Snp n = (spare - CL_METADATA_SIZE) / MSIZE; 4104265425Snp if (n > howmany(hwb->size, maxp)) 4105265425Snp break; 4106255050Snp 4107265425Snp hwidx = idx; 4108265425Snp if (fl->flags & FL_BUF_PACKING) { 4109265425Snp region1 = n * MSIZE; 4110265425Snp region3 = spare - region1; 4111265425Snp } else { 4112265425Snp region1 = MSIZE; 4113265425Snp region3 = spare - region1; 4114265425Snp break; 4115255050Snp } 4116255050Snp } 4117255050Snp 4118265425Snp KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES, 4119265425Snp ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp)); 4120265425Snp KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES, 4121265425Snp ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp)); 4122265425Snp KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 == 4123265425Snp sc->sge.sw_zone_info[zidx].size, 4124265425Snp ("%s: bad buffer layout for fl %p, maxp %d. " 4125265425Snp "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp, 4126265425Snp sc->sge.sw_zone_info[zidx].size, region1, 4127265425Snp sc->sge.hw_buf_info[hwidx].size, region3)); 4128265425Snp if (fl->flags & FL_BUF_PACKING || region1 > 0) { 4129265425Snp KASSERT(region3 >= CL_METADATA_SIZE, 4130265425Snp ("%s: no room for metadata. fl %p, maxp %d; " 4131265425Snp "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp, 4132265425Snp sc->sge.sw_zone_info[zidx].size, region1, 4133265425Snp sc->sge.hw_buf_info[hwidx].size, region3)); 4134265425Snp KASSERT(region1 % MSIZE == 0, 4135265425Snp ("%s: bad mbuf region for fl %p, maxp %d. " 4136265425Snp "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp, 4137265425Snp sc->sge.sw_zone_info[zidx].size, region1, 4138265425Snp sc->sge.hw_buf_info[hwidx].size, region3)); 4139265425Snp } 4140265425Snp 4141265425Snp fl->cll_def.zidx = zidx; 4142265425Snp fl->cll_def.hwidx = hwidx; 4143265425Snp fl->cll_def.region1 = region1; 4144265425Snp fl->cll_def.region3 = region3; 4145265425Snp} 4146265425Snp 4147265425Snpstatic void 4148265425Snpfind_safe_refill_source(struct adapter *sc, struct sge_fl *fl) 4149265425Snp{ 4150265425Snp struct sge *s = &sc->sge; 4151265425Snp struct hw_buf_info *hwb; 4152265425Snp struct sw_zone_info *swz; 4153265425Snp int spare; 4154265425Snp int8_t hwidx; 4155265425Snp 4156265425Snp if (fl->flags & FL_BUF_PACKING) 4157265425Snp hwidx = s->safe_hwidx2; /* with room for metadata */ 4158265425Snp else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) { 4159265425Snp hwidx = s->safe_hwidx2; 4160265425Snp hwb = &s->hw_buf_info[hwidx]; 4161265425Snp swz = &s->sw_zone_info[hwb->zidx]; 4162265425Snp spare = swz->size - hwb->size; 4163265425Snp 4164265425Snp /* no good if there isn't room for an mbuf as well */ 4165265425Snp if (spare < CL_METADATA_SIZE + MSIZE) 4166265425Snp hwidx = s->safe_hwidx1; 4167265425Snp } else 4168265425Snp hwidx = s->safe_hwidx1; 4169265425Snp 4170265425Snp if (hwidx == -1) { 4171265425Snp /* No fallback source */ 4172265425Snp fl->cll_alt.hwidx = -1; 4173265425Snp fl->cll_alt.zidx = -1; 4174265425Snp 4175265425Snp return; 4176265425Snp } 4177265425Snp 4178265425Snp hwb = &s->hw_buf_info[hwidx]; 4179265425Snp swz = &s->sw_zone_info[hwb->zidx]; 4180265425Snp spare = swz->size - hwb->size; 4181265425Snp fl->cll_alt.hwidx = hwidx; 4182265425Snp fl->cll_alt.zidx = hwb->zidx; 4183265425Snp if (allow_mbufs_in_cluster) 4184265425Snp fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE; 4185255050Snp else 4186265425Snp fl->cll_alt.region1 = 0; 4187265425Snp fl->cll_alt.region3 = spare - fl->cll_alt.region1; 4188218792Snp} 4189219286Snp 4190222510Snpstatic void 4191228561Snpadd_fl_to_sfl(struct adapter *sc, struct sge_fl *fl) 4192222510Snp{ 4193228561Snp mtx_lock(&sc->sfl_lock); 4194228561Snp FL_LOCK(fl); 4195228561Snp if ((fl->flags & FL_DOOMED) == 0) { 4196228561Snp fl->flags |= FL_STARVING; 4197228561Snp TAILQ_INSERT_TAIL(&sc->sfl, fl, link); 4198228561Snp callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc); 4199222510Snp } 4200228561Snp FL_UNLOCK(fl); 4201228561Snp mtx_unlock(&sc->sfl_lock); 4202222510Snp} 4203222510Snp 4204220873Snpstatic int 4205228561Snphandle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss, 4206228561Snp struct mbuf *m) 4207220873Snp{ 4208228561Snp const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1); 4209228561Snp unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid)); 4210228561Snp struct adapter *sc = iq->adapter; 4211228561Snp struct sge *s = &sc->sge; 4212228561Snp struct sge_eq *eq; 4213220873Snp 4214228561Snp KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__, 4215228561Snp rss->opcode)); 4216220873Snp 4217228561Snp eq = s->eqmap[qid - s->eq_start]; 4218220873Snp EQ_LOCK(eq); 4219228561Snp KASSERT(eq->flags & EQ_CRFLUSHED, 4220228561Snp ("%s: unsolicited egress update", __func__)); 4221228561Snp eq->flags &= ~EQ_CRFLUSHED; 4222228561Snp eq->egr_update++; 4223220873Snp 4224228561Snp if (__predict_false(eq->flags & EQ_DOOMED)) 4225228561Snp wakeup_one(eq); 4226228561Snp else if (eq->flags & EQ_STALLED && can_resume_tx(eq)) 4227228561Snp taskqueue_enqueue(sc->tq[eq->tx_chan], &eq->tx_task); 4228228561Snp EQ_UNLOCK(eq); 4229220873Snp 4230228561Snp return (0); 4231228561Snp} 4232220873Snp 4233247291Snp/* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */ 4234247291SnpCTASSERT(offsetof(struct cpl_fw4_msg, data) == \ 4235247291Snp offsetof(struct cpl_fw6_msg, data)); 4236247291Snp 4237228561Snpstatic int 4238239336Snphandle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 4239228561Snp{ 4240239336Snp struct adapter *sc = iq->adapter; 4241228561Snp const struct cpl_fw6_msg *cpl = (const void *)(rss + 1); 4242220873Snp 4243228561Snp KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__, 4244228561Snp rss->opcode)); 4245220873Snp 4246247291Snp if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) { 4247247291Snp const struct rss_header *rss2; 4248247291Snp 4249247291Snp rss2 = (const struct rss_header *)&cpl->data[0]; 4250247291Snp return (sc->cpl_handler[rss2->opcode](iq, rss2, m)); 4251247291Snp } 4252247291Snp 4253239336Snp return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0])); 4254220873Snp} 4255221911Snp 4256221911Snpstatic int 4257222510Snpsysctl_uint16(SYSCTL_HANDLER_ARGS) 4258221911Snp{ 4259221911Snp uint16_t *id = arg1; 4260221911Snp int i = *id; 4261221911Snp 4262221911Snp return sysctl_handle_int(oidp, &i, 0, req); 4263221911Snp} 4264265425Snp 4265265425Snpstatic int 4266265425Snpsysctl_bufsizes(SYSCTL_HANDLER_ARGS) 4267265425Snp{ 4268265425Snp struct sge *s = arg1; 4269265425Snp struct hw_buf_info *hwb = &s->hw_buf_info[0]; 4270265425Snp struct sw_zone_info *swz = &s->sw_zone_info[0]; 4271265425Snp int i, rc; 4272265425Snp struct sbuf sb; 4273265425Snp char c; 4274265425Snp 4275265425Snp sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND); 4276265425Snp for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) { 4277265425Snp if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster) 4278265425Snp c = '*'; 4279265425Snp else 4280265425Snp c = '\0'; 4281265425Snp 4282265425Snp sbuf_printf(&sb, "%u%c ", hwb->size, c); 4283265425Snp } 4284265425Snp sbuf_trim(&sb); 4285265425Snp sbuf_finish(&sb); 4286265425Snp rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req); 4287265425Snp sbuf_delete(&sb); 4288265425Snp return (rc); 4289265425Snp} 4290