1329222Shselasky/* 2329222Shselasky * Copyright (c) 2004 Topspin Communications. All rights reserved. 3329222Shselasky * Copyright (c) 2005 Cisco Systems. All rights reserved. 4329222Shselasky * Copyright (c) 2005 Mellanox Technologies. All rights reserved. 5329222Shselasky * Copyright (c) 2004 Voltaire, Inc. All rights reserved. 6329222Shselasky * 7329222Shselasky * This software is available to you under a choice of one of two 8329222Shselasky * licenses. You may choose to be licensed under the terms of the GNU 9329222Shselasky * General Public License (GPL) Version 2, available from the file 10329222Shselasky * COPYING in the main directory of this source tree, or the 11329222Shselasky * OpenIB.org BSD license below: 12329222Shselasky * 13329222Shselasky * Redistribution and use in source and binary forms, with or 14329222Shselasky * without modification, are permitted provided that the following 15329222Shselasky * conditions are met: 16329222Shselasky * 17329222Shselasky * - Redistributions of source code must retain the above 18329222Shselasky * copyright notice, this list of conditions and the following 19329222Shselasky * disclaimer. 20329222Shselasky * 21329222Shselasky * - Redistributions in binary form must reproduce the above 22329222Shselasky * copyright notice, this list of conditions and the following 23329222Shselasky * disclaimer in the documentation and/or other materials 24329222Shselasky * provided with the distribution. 25329222Shselasky * 26329222Shselasky * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 27329222Shselasky * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 28329222Shselasky * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 29329222Shselasky * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 30329222Shselasky * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 31329222Shselasky * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 32329222Shselasky * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 33329222Shselasky * SOFTWARE. 34329222Shselasky */ 35329222Shselasky 36329222Shselasky#include <linux/string.h> 37329222Shselasky#include <linux/slab.h> 38329222Shselasky#include <linux/sched.h> 39329222Shselasky 40329222Shselasky#include <asm/io.h> 41329222Shselasky 42329222Shselasky#include <rdma/ib_verbs.h> 43329222Shselasky#include <rdma/ib_cache.h> 44329222Shselasky#include <rdma/ib_pack.h> 45329222Shselasky 46329222Shselasky#include "mthca_dev.h" 47329222Shselasky#include "mthca_cmd.h" 48329222Shselasky#include "mthca_memfree.h" 49329222Shselasky#include "mthca_wqe.h" 50329222Shselasky 51329222Shselaskyenum { 52329222Shselasky MTHCA_MAX_DIRECT_QP_SIZE = 4 * PAGE_SIZE, 53329222Shselasky MTHCA_ACK_REQ_FREQ = 10, 54329222Shselasky MTHCA_FLIGHT_LIMIT = 9, 55329222Shselasky MTHCA_UD_HEADER_SIZE = 72, /* largest UD header possible */ 56329222Shselasky MTHCA_INLINE_HEADER_SIZE = 4, /* data segment overhead for inline */ 57329222Shselasky MTHCA_INLINE_CHUNK_SIZE = 16 /* inline data segment chunk */ 58329222Shselasky}; 59329222Shselasky 60329222Shselaskyenum { 61329222Shselasky MTHCA_QP_STATE_RST = 0, 62329222Shselasky MTHCA_QP_STATE_INIT = 1, 63329222Shselasky MTHCA_QP_STATE_RTR = 2, 64329222Shselasky MTHCA_QP_STATE_RTS = 3, 65329222Shselasky MTHCA_QP_STATE_SQE = 4, 66329222Shselasky MTHCA_QP_STATE_SQD = 5, 67329222Shselasky MTHCA_QP_STATE_ERR = 6, 68329222Shselasky MTHCA_QP_STATE_DRAINING = 7 69329222Shselasky}; 70329222Shselasky 71329222Shselaskyenum { 72329222Shselasky MTHCA_QP_ST_RC = 0x0, 73329222Shselasky MTHCA_QP_ST_UC = 0x1, 74329222Shselasky MTHCA_QP_ST_RD = 0x2, 75329222Shselasky MTHCA_QP_ST_UD = 0x3, 76329222Shselasky MTHCA_QP_ST_MLX = 0x7 77329222Shselasky}; 78329222Shselasky 79329222Shselaskyenum { 80329222Shselasky MTHCA_QP_PM_MIGRATED = 0x3, 81329222Shselasky MTHCA_QP_PM_ARMED = 0x0, 82329222Shselasky MTHCA_QP_PM_REARM = 0x1 83329222Shselasky}; 84329222Shselasky 85329222Shselaskyenum { 86329222Shselasky /* qp_context flags */ 87329222Shselasky MTHCA_QP_BIT_DE = 1 << 8, 88329222Shselasky /* params1 */ 89329222Shselasky MTHCA_QP_BIT_SRE = 1 << 15, 90329222Shselasky MTHCA_QP_BIT_SWE = 1 << 14, 91329222Shselasky MTHCA_QP_BIT_SAE = 1 << 13, 92329222Shselasky MTHCA_QP_BIT_SIC = 1 << 4, 93329222Shselasky MTHCA_QP_BIT_SSC = 1 << 3, 94329222Shselasky /* params2 */ 95329222Shselasky MTHCA_QP_BIT_RRE = 1 << 15, 96329222Shselasky MTHCA_QP_BIT_RWE = 1 << 14, 97329222Shselasky MTHCA_QP_BIT_RAE = 1 << 13, 98329222Shselasky MTHCA_QP_BIT_RIC = 1 << 4, 99329222Shselasky MTHCA_QP_BIT_RSC = 1 << 3 100329222Shselasky}; 101329222Shselasky 102329222Shselaskyenum { 103329222Shselasky MTHCA_SEND_DOORBELL_FENCE = 1 << 5 104329222Shselasky}; 105329222Shselasky 106329222Shselaskystruct mthca_qp_path { 107329222Shselasky __be32 port_pkey; 108329222Shselasky u8 rnr_retry; 109329222Shselasky u8 g_mylmc; 110329222Shselasky __be16 rlid; 111329222Shselasky u8 ackto; 112329222Shselasky u8 mgid_index; 113329222Shselasky u8 static_rate; 114329222Shselasky u8 hop_limit; 115329222Shselasky __be32 sl_tclass_flowlabel; 116329222Shselasky u8 rgid[16]; 117329222Shselasky} __attribute__((packed)); 118329222Shselasky 119329222Shselaskystruct mthca_qp_context { 120329222Shselasky __be32 flags; 121329222Shselasky __be32 tavor_sched_queue; /* Reserved on Arbel */ 122329222Shselasky u8 mtu_msgmax; 123329222Shselasky u8 rq_size_stride; /* Reserved on Tavor */ 124329222Shselasky u8 sq_size_stride; /* Reserved on Tavor */ 125329222Shselasky u8 rlkey_arbel_sched_queue; /* Reserved on Tavor */ 126329222Shselasky __be32 usr_page; 127329222Shselasky __be32 local_qpn; 128329222Shselasky __be32 remote_qpn; 129329222Shselasky u32 reserved1[2]; 130329222Shselasky struct mthca_qp_path pri_path; 131329222Shselasky struct mthca_qp_path alt_path; 132329222Shselasky __be32 rdd; 133329222Shselasky __be32 pd; 134329222Shselasky __be32 wqe_base; 135329222Shselasky __be32 wqe_lkey; 136329222Shselasky __be32 params1; 137329222Shselasky __be32 reserved2; 138329222Shselasky __be32 next_send_psn; 139329222Shselasky __be32 cqn_snd; 140329222Shselasky __be32 snd_wqe_base_l; /* Next send WQE on Tavor */ 141329222Shselasky __be32 snd_db_index; /* (debugging only entries) */ 142329222Shselasky __be32 last_acked_psn; 143329222Shselasky __be32 ssn; 144329222Shselasky __be32 params2; 145329222Shselasky __be32 rnr_nextrecvpsn; 146329222Shselasky __be32 ra_buff_indx; 147329222Shselasky __be32 cqn_rcv; 148329222Shselasky __be32 rcv_wqe_base_l; /* Next recv WQE on Tavor */ 149329222Shselasky __be32 rcv_db_index; /* (debugging only entries) */ 150329222Shselasky __be32 qkey; 151329222Shselasky __be32 srqn; 152329222Shselasky __be32 rmsn; 153329222Shselasky __be16 rq_wqe_counter; /* reserved on Tavor */ 154329222Shselasky __be16 sq_wqe_counter; /* reserved on Tavor */ 155329222Shselasky u32 reserved3[18]; 156329222Shselasky} __attribute__((packed)); 157329222Shselasky 158329222Shselaskystruct mthca_qp_param { 159329222Shselasky __be32 opt_param_mask; 160329222Shselasky u32 reserved1; 161329222Shselasky struct mthca_qp_context context; 162329222Shselasky u32 reserved2[62]; 163329222Shselasky} __attribute__((packed)); 164329222Shselasky 165329222Shselaskyenum { 166329222Shselasky MTHCA_QP_OPTPAR_ALT_ADDR_PATH = 1 << 0, 167329222Shselasky MTHCA_QP_OPTPAR_RRE = 1 << 1, 168329222Shselasky MTHCA_QP_OPTPAR_RAE = 1 << 2, 169329222Shselasky MTHCA_QP_OPTPAR_RWE = 1 << 3, 170329222Shselasky MTHCA_QP_OPTPAR_PKEY_INDEX = 1 << 4, 171329222Shselasky MTHCA_QP_OPTPAR_Q_KEY = 1 << 5, 172329222Shselasky MTHCA_QP_OPTPAR_RNR_TIMEOUT = 1 << 6, 173329222Shselasky MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH = 1 << 7, 174329222Shselasky MTHCA_QP_OPTPAR_SRA_MAX = 1 << 8, 175329222Shselasky MTHCA_QP_OPTPAR_RRA_MAX = 1 << 9, 176329222Shselasky MTHCA_QP_OPTPAR_PM_STATE = 1 << 10, 177329222Shselasky MTHCA_QP_OPTPAR_PORT_NUM = 1 << 11, 178329222Shselasky MTHCA_QP_OPTPAR_RETRY_COUNT = 1 << 12, 179329222Shselasky MTHCA_QP_OPTPAR_ALT_RNR_RETRY = 1 << 13, 180329222Shselasky MTHCA_QP_OPTPAR_ACK_TIMEOUT = 1 << 14, 181329222Shselasky MTHCA_QP_OPTPAR_RNR_RETRY = 1 << 15, 182329222Shselasky MTHCA_QP_OPTPAR_SCHED_QUEUE = 1 << 16 183329222Shselasky}; 184329222Shselasky 185329222Shselaskystatic const u8 mthca_opcode[] = { 186329222Shselasky [IB_WR_SEND] = MTHCA_OPCODE_SEND, 187329222Shselasky [IB_WR_SEND_WITH_IMM] = MTHCA_OPCODE_SEND_IMM, 188329222Shselasky [IB_WR_RDMA_WRITE] = MTHCA_OPCODE_RDMA_WRITE, 189329222Shselasky [IB_WR_RDMA_WRITE_WITH_IMM] = MTHCA_OPCODE_RDMA_WRITE_IMM, 190329222Shselasky [IB_WR_RDMA_READ] = MTHCA_OPCODE_RDMA_READ, 191329222Shselasky [IB_WR_ATOMIC_CMP_AND_SWP] = MTHCA_OPCODE_ATOMIC_CS, 192329222Shselasky [IB_WR_ATOMIC_FETCH_AND_ADD] = MTHCA_OPCODE_ATOMIC_FA, 193329222Shselasky}; 194329222Shselasky 195329222Shselaskystatic int is_sqp(struct mthca_dev *dev, struct mthca_qp *qp) 196329222Shselasky{ 197329222Shselasky return qp->qpn >= dev->qp_table.sqp_start && 198329222Shselasky qp->qpn <= dev->qp_table.sqp_start + 3; 199329222Shselasky} 200329222Shselasky 201329222Shselaskystatic int is_qp0(struct mthca_dev *dev, struct mthca_qp *qp) 202329222Shselasky{ 203329222Shselasky return qp->qpn >= dev->qp_table.sqp_start && 204329222Shselasky qp->qpn <= dev->qp_table.sqp_start + 1; 205329222Shselasky} 206329222Shselasky 207329222Shselaskystatic void *get_recv_wqe(struct mthca_qp *qp, int n) 208329222Shselasky{ 209329222Shselasky if (qp->is_direct) 210329222Shselasky return qp->queue.direct.buf + (n << qp->rq.wqe_shift); 211329222Shselasky else 212329222Shselasky return qp->queue.page_list[(n << qp->rq.wqe_shift) >> PAGE_SHIFT].buf + 213329222Shselasky ((n << qp->rq.wqe_shift) & (PAGE_SIZE - 1)); 214329222Shselasky} 215329222Shselasky 216329222Shselaskystatic void *get_send_wqe(struct mthca_qp *qp, int n) 217329222Shselasky{ 218329222Shselasky if (qp->is_direct) 219329222Shselasky return qp->queue.direct.buf + qp->send_wqe_offset + 220329222Shselasky (n << qp->sq.wqe_shift); 221329222Shselasky else 222329222Shselasky return qp->queue.page_list[(qp->send_wqe_offset + 223329222Shselasky (n << qp->sq.wqe_shift)) >> 224329222Shselasky PAGE_SHIFT].buf + 225329222Shselasky ((qp->send_wqe_offset + (n << qp->sq.wqe_shift)) & 226329222Shselasky (PAGE_SIZE - 1)); 227329222Shselasky} 228329222Shselasky 229329222Shselaskystatic void mthca_wq_reset(struct mthca_wq *wq) 230329222Shselasky{ 231329222Shselasky wq->next_ind = 0; 232329222Shselasky wq->last_comp = wq->max - 1; 233329222Shselasky wq->head = 0; 234329222Shselasky wq->tail = 0; 235329222Shselasky} 236329222Shselasky 237329222Shselaskyvoid mthca_qp_event(struct mthca_dev *dev, u32 qpn, 238329222Shselasky enum ib_event_type event_type) 239329222Shselasky{ 240329222Shselasky struct mthca_qp *qp; 241329222Shselasky struct ib_event event; 242329222Shselasky 243329222Shselasky spin_lock(&dev->qp_table.lock); 244329222Shselasky qp = mthca_array_get(&dev->qp_table.qp, qpn & (dev->limits.num_qps - 1)); 245329222Shselasky if (qp) 246329222Shselasky ++qp->refcount; 247329222Shselasky spin_unlock(&dev->qp_table.lock); 248329222Shselasky 249329222Shselasky if (!qp) { 250329222Shselasky mthca_warn(dev, "Async event %d for bogus QP %08x\n", 251329222Shselasky event_type, qpn); 252329222Shselasky return; 253329222Shselasky } 254329222Shselasky 255329222Shselasky if (event_type == IB_EVENT_PATH_MIG) 256329222Shselasky qp->port = qp->alt_port; 257329222Shselasky 258329222Shselasky event.device = &dev->ib_dev; 259329222Shselasky event.event = event_type; 260329222Shselasky event.element.qp = &qp->ibqp; 261329222Shselasky if (qp->ibqp.event_handler) 262329222Shselasky qp->ibqp.event_handler(&event, qp->ibqp.qp_context); 263329222Shselasky 264329222Shselasky spin_lock(&dev->qp_table.lock); 265329222Shselasky if (!--qp->refcount) 266329222Shselasky wake_up(&qp->wait); 267329222Shselasky spin_unlock(&dev->qp_table.lock); 268329222Shselasky} 269329222Shselasky 270329222Shselaskystatic int to_mthca_state(enum ib_qp_state ib_state) 271329222Shselasky{ 272329222Shselasky switch (ib_state) { 273329222Shselasky case IB_QPS_RESET: return MTHCA_QP_STATE_RST; 274329222Shselasky case IB_QPS_INIT: return MTHCA_QP_STATE_INIT; 275329222Shselasky case IB_QPS_RTR: return MTHCA_QP_STATE_RTR; 276329222Shselasky case IB_QPS_RTS: return MTHCA_QP_STATE_RTS; 277329222Shselasky case IB_QPS_SQD: return MTHCA_QP_STATE_SQD; 278329222Shselasky case IB_QPS_SQE: return MTHCA_QP_STATE_SQE; 279329222Shselasky case IB_QPS_ERR: return MTHCA_QP_STATE_ERR; 280329222Shselasky default: return -1; 281329222Shselasky } 282329222Shselasky} 283329222Shselasky 284329222Shselaskyenum { RC, UC, UD, RD, RDEE, MLX, NUM_TRANS }; 285329222Shselasky 286329222Shselaskystatic int to_mthca_st(int transport) 287329222Shselasky{ 288329222Shselasky switch (transport) { 289329222Shselasky case RC: return MTHCA_QP_ST_RC; 290329222Shselasky case UC: return MTHCA_QP_ST_UC; 291329222Shselasky case UD: return MTHCA_QP_ST_UD; 292329222Shselasky case RD: return MTHCA_QP_ST_RD; 293329222Shselasky case MLX: return MTHCA_QP_ST_MLX; 294329222Shselasky default: return -1; 295329222Shselasky } 296329222Shselasky} 297329222Shselasky 298329222Shselaskystatic void store_attrs(struct mthca_sqp *sqp, const struct ib_qp_attr *attr, 299329222Shselasky int attr_mask) 300329222Shselasky{ 301329222Shselasky if (attr_mask & IB_QP_PKEY_INDEX) 302329222Shselasky sqp->pkey_index = attr->pkey_index; 303329222Shselasky if (attr_mask & IB_QP_QKEY) 304329222Shselasky sqp->qkey = attr->qkey; 305329222Shselasky if (attr_mask & IB_QP_SQ_PSN) 306329222Shselasky sqp->send_psn = attr->sq_psn; 307329222Shselasky} 308329222Shselasky 309329222Shselaskystatic void init_port(struct mthca_dev *dev, int port) 310329222Shselasky{ 311329222Shselasky int err; 312329222Shselasky struct mthca_init_ib_param param; 313329222Shselasky 314329222Shselasky memset(¶m, 0, sizeof param); 315329222Shselasky 316329222Shselasky param.port_width = dev->limits.port_width_cap; 317329222Shselasky param.vl_cap = dev->limits.vl_cap; 318329222Shselasky param.mtu_cap = dev->limits.mtu_cap; 319329222Shselasky param.gid_cap = dev->limits.gid_table_len; 320329222Shselasky param.pkey_cap = dev->limits.pkey_table_len; 321329222Shselasky 322329222Shselasky err = mthca_INIT_IB(dev, ¶m, port); 323329222Shselasky if (err) 324329222Shselasky mthca_warn(dev, "INIT_IB failed, return code %d.\n", err); 325329222Shselasky} 326329222Shselasky 327329222Shselaskystatic __be32 get_hw_access_flags(struct mthca_qp *qp, const struct ib_qp_attr *attr, 328329222Shselasky int attr_mask) 329329222Shselasky{ 330329222Shselasky u8 dest_rd_atomic; 331329222Shselasky u32 access_flags; 332329222Shselasky u32 hw_access_flags = 0; 333329222Shselasky 334329222Shselasky if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 335329222Shselasky dest_rd_atomic = attr->max_dest_rd_atomic; 336329222Shselasky else 337329222Shselasky dest_rd_atomic = qp->resp_depth; 338329222Shselasky 339329222Shselasky if (attr_mask & IB_QP_ACCESS_FLAGS) 340329222Shselasky access_flags = attr->qp_access_flags; 341329222Shselasky else 342329222Shselasky access_flags = qp->atomic_rd_en; 343329222Shselasky 344329222Shselasky if (!dest_rd_atomic) 345329222Shselasky access_flags &= IB_ACCESS_REMOTE_WRITE; 346329222Shselasky 347329222Shselasky if (access_flags & IB_ACCESS_REMOTE_READ) 348329222Shselasky hw_access_flags |= MTHCA_QP_BIT_RRE; 349329222Shselasky if (access_flags & IB_ACCESS_REMOTE_ATOMIC) 350329222Shselasky hw_access_flags |= MTHCA_QP_BIT_RAE; 351329222Shselasky if (access_flags & IB_ACCESS_REMOTE_WRITE) 352329222Shselasky hw_access_flags |= MTHCA_QP_BIT_RWE; 353329222Shselasky 354329222Shselasky return cpu_to_be32(hw_access_flags); 355329222Shselasky} 356329222Shselasky 357329222Shselaskystatic inline enum ib_qp_state to_ib_qp_state(int mthca_state) 358329222Shselasky{ 359329222Shselasky switch (mthca_state) { 360329222Shselasky case MTHCA_QP_STATE_RST: return IB_QPS_RESET; 361329222Shselasky case MTHCA_QP_STATE_INIT: return IB_QPS_INIT; 362329222Shselasky case MTHCA_QP_STATE_RTR: return IB_QPS_RTR; 363329222Shselasky case MTHCA_QP_STATE_RTS: return IB_QPS_RTS; 364329222Shselasky case MTHCA_QP_STATE_DRAINING: 365329222Shselasky case MTHCA_QP_STATE_SQD: return IB_QPS_SQD; 366329222Shselasky case MTHCA_QP_STATE_SQE: return IB_QPS_SQE; 367329222Shselasky case MTHCA_QP_STATE_ERR: return IB_QPS_ERR; 368329222Shselasky default: return -1; 369329222Shselasky } 370329222Shselasky} 371329222Shselasky 372329222Shselaskystatic inline enum ib_mig_state to_ib_mig_state(int mthca_mig_state) 373329222Shselasky{ 374329222Shselasky switch (mthca_mig_state) { 375329222Shselasky case 0: return IB_MIG_ARMED; 376329222Shselasky case 1: return IB_MIG_REARM; 377329222Shselasky case 3: return IB_MIG_MIGRATED; 378329222Shselasky default: return -1; 379329222Shselasky } 380329222Shselasky} 381329222Shselasky 382329222Shselaskystatic int to_ib_qp_access_flags(int mthca_flags) 383329222Shselasky{ 384329222Shselasky int ib_flags = 0; 385329222Shselasky 386329222Shselasky if (mthca_flags & MTHCA_QP_BIT_RRE) 387329222Shselasky ib_flags |= IB_ACCESS_REMOTE_READ; 388329222Shselasky if (mthca_flags & MTHCA_QP_BIT_RWE) 389329222Shselasky ib_flags |= IB_ACCESS_REMOTE_WRITE; 390329222Shselasky if (mthca_flags & MTHCA_QP_BIT_RAE) 391329222Shselasky ib_flags |= IB_ACCESS_REMOTE_ATOMIC; 392329222Shselasky 393329222Shselasky return ib_flags; 394329222Shselasky} 395329222Shselasky 396329222Shselaskystatic void to_ib_ah_attr(struct mthca_dev *dev, struct ib_ah_attr *ib_ah_attr, 397329222Shselasky struct mthca_qp_path *path) 398329222Shselasky{ 399329222Shselasky memset(ib_ah_attr, 0, sizeof *ib_ah_attr); 400329222Shselasky ib_ah_attr->port_num = (be32_to_cpu(path->port_pkey) >> 24) & 0x3; 401329222Shselasky 402329222Shselasky if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->limits.num_ports) 403329222Shselasky return; 404329222Shselasky 405329222Shselasky ib_ah_attr->dlid = be16_to_cpu(path->rlid); 406329222Shselasky ib_ah_attr->sl = be32_to_cpu(path->sl_tclass_flowlabel) >> 28; 407329222Shselasky ib_ah_attr->src_path_bits = path->g_mylmc & 0x7f; 408329222Shselasky ib_ah_attr->static_rate = mthca_rate_to_ib(dev, 409329222Shselasky path->static_rate & 0xf, 410329222Shselasky ib_ah_attr->port_num); 411329222Shselasky ib_ah_attr->ah_flags = (path->g_mylmc & (1 << 7)) ? IB_AH_GRH : 0; 412329222Shselasky if (ib_ah_attr->ah_flags) { 413329222Shselasky ib_ah_attr->grh.sgid_index = path->mgid_index & (dev->limits.gid_table_len - 1); 414329222Shselasky ib_ah_attr->grh.hop_limit = path->hop_limit; 415329222Shselasky ib_ah_attr->grh.traffic_class = 416329222Shselasky (be32_to_cpu(path->sl_tclass_flowlabel) >> 20) & 0xff; 417329222Shselasky ib_ah_attr->grh.flow_label = 418329222Shselasky be32_to_cpu(path->sl_tclass_flowlabel) & 0xfffff; 419329222Shselasky memcpy(ib_ah_attr->grh.dgid.raw, 420329222Shselasky path->rgid, sizeof ib_ah_attr->grh.dgid.raw); 421329222Shselasky } 422329222Shselasky} 423329222Shselasky 424329222Shselaskyint mthca_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask, 425329222Shselasky struct ib_qp_init_attr *qp_init_attr) 426329222Shselasky{ 427329222Shselasky struct mthca_dev *dev = to_mdev(ibqp->device); 428329222Shselasky struct mthca_qp *qp = to_mqp(ibqp); 429329222Shselasky int err = 0; 430329222Shselasky struct mthca_mailbox *mailbox = NULL; 431329222Shselasky struct mthca_qp_param *qp_param; 432329222Shselasky struct mthca_qp_context *context; 433329222Shselasky int mthca_state; 434329222Shselasky 435329222Shselasky mutex_lock(&qp->mutex); 436329222Shselasky 437329222Shselasky if (qp->state == IB_QPS_RESET) { 438329222Shselasky qp_attr->qp_state = IB_QPS_RESET; 439329222Shselasky goto done; 440329222Shselasky } 441329222Shselasky 442329222Shselasky mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL); 443329222Shselasky if (IS_ERR(mailbox)) { 444329222Shselasky err = PTR_ERR(mailbox); 445329222Shselasky goto out; 446329222Shselasky } 447329222Shselasky 448329222Shselasky err = mthca_QUERY_QP(dev, qp->qpn, 0, mailbox); 449329222Shselasky if (err) { 450329222Shselasky mthca_warn(dev, "QUERY_QP failed (%d)\n", err); 451329222Shselasky goto out_mailbox; 452329222Shselasky } 453329222Shselasky 454329222Shselasky qp_param = mailbox->buf; 455329222Shselasky context = &qp_param->context; 456329222Shselasky mthca_state = be32_to_cpu(context->flags) >> 28; 457329222Shselasky 458329222Shselasky qp->state = to_ib_qp_state(mthca_state); 459329222Shselasky qp_attr->qp_state = qp->state; 460329222Shselasky qp_attr->path_mtu = context->mtu_msgmax >> 5; 461329222Shselasky qp_attr->path_mig_state = 462329222Shselasky to_ib_mig_state((be32_to_cpu(context->flags) >> 11) & 0x3); 463329222Shselasky qp_attr->qkey = be32_to_cpu(context->qkey); 464329222Shselasky qp_attr->rq_psn = be32_to_cpu(context->rnr_nextrecvpsn) & 0xffffff; 465329222Shselasky qp_attr->sq_psn = be32_to_cpu(context->next_send_psn) & 0xffffff; 466329222Shselasky qp_attr->dest_qp_num = be32_to_cpu(context->remote_qpn) & 0xffffff; 467329222Shselasky qp_attr->qp_access_flags = 468329222Shselasky to_ib_qp_access_flags(be32_to_cpu(context->params2)); 469329222Shselasky 470329222Shselasky if (qp->transport == RC || qp->transport == UC) { 471329222Shselasky to_ib_ah_attr(dev, &qp_attr->ah_attr, &context->pri_path); 472329222Shselasky to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context->alt_path); 473329222Shselasky qp_attr->alt_pkey_index = 474329222Shselasky be32_to_cpu(context->alt_path.port_pkey) & 0x7f; 475329222Shselasky qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num; 476329222Shselasky } 477329222Shselasky 478329222Shselasky qp_attr->pkey_index = be32_to_cpu(context->pri_path.port_pkey) & 0x7f; 479329222Shselasky qp_attr->port_num = 480329222Shselasky (be32_to_cpu(context->pri_path.port_pkey) >> 24) & 0x3; 481329222Shselasky 482329222Shselasky /* qp_attr->en_sqd_async_notify is only applicable in modify qp */ 483329222Shselasky qp_attr->sq_draining = mthca_state == MTHCA_QP_STATE_DRAINING; 484329222Shselasky 485329222Shselasky qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context->params1) >> 21) & 0x7); 486329222Shselasky 487329222Shselasky qp_attr->max_dest_rd_atomic = 488329222Shselasky 1 << ((be32_to_cpu(context->params2) >> 21) & 0x7); 489329222Shselasky qp_attr->min_rnr_timer = 490329222Shselasky (be32_to_cpu(context->rnr_nextrecvpsn) >> 24) & 0x1f; 491329222Shselasky qp_attr->timeout = context->pri_path.ackto >> 3; 492329222Shselasky qp_attr->retry_cnt = (be32_to_cpu(context->params1) >> 16) & 0x7; 493329222Shselasky qp_attr->rnr_retry = context->pri_path.rnr_retry >> 5; 494329222Shselasky qp_attr->alt_timeout = context->alt_path.ackto >> 3; 495329222Shselasky 496329222Shselaskydone: 497329222Shselasky qp_attr->cur_qp_state = qp_attr->qp_state; 498329222Shselasky qp_attr->cap.max_send_wr = qp->sq.max; 499329222Shselasky qp_attr->cap.max_recv_wr = qp->rq.max; 500329222Shselasky qp_attr->cap.max_send_sge = qp->sq.max_gs; 501329222Shselasky qp_attr->cap.max_recv_sge = qp->rq.max_gs; 502329222Shselasky qp_attr->cap.max_inline_data = qp->max_inline_data; 503329222Shselasky 504329222Shselasky qp_init_attr->cap = qp_attr->cap; 505329222Shselasky qp_init_attr->sq_sig_type = qp->sq_policy; 506329222Shselasky 507329222Shselaskyout_mailbox: 508329222Shselasky mthca_free_mailbox(dev, mailbox); 509329222Shselasky 510329222Shselaskyout: 511329222Shselasky mutex_unlock(&qp->mutex); 512329222Shselasky return err; 513329222Shselasky} 514329222Shselasky 515329222Shselaskystatic int mthca_path_set(struct mthca_dev *dev, const struct ib_ah_attr *ah, 516329222Shselasky struct mthca_qp_path *path, u8 port) 517329222Shselasky{ 518329222Shselasky path->g_mylmc = ah->src_path_bits & 0x7f; 519329222Shselasky path->rlid = cpu_to_be16(ah->dlid); 520329222Shselasky path->static_rate = mthca_get_rate(dev, ah->static_rate, port); 521329222Shselasky 522329222Shselasky if (ah->ah_flags & IB_AH_GRH) { 523329222Shselasky if (ah->grh.sgid_index >= dev->limits.gid_table_len) { 524329222Shselasky mthca_dbg(dev, "sgid_index (%u) too large. max is %d\n", 525329222Shselasky ah->grh.sgid_index, dev->limits.gid_table_len-1); 526329222Shselasky return -1; 527329222Shselasky } 528329222Shselasky 529329222Shselasky path->g_mylmc |= 1 << 7; 530329222Shselasky path->mgid_index = ah->grh.sgid_index; 531329222Shselasky path->hop_limit = ah->grh.hop_limit; 532329222Shselasky path->sl_tclass_flowlabel = 533329222Shselasky cpu_to_be32((ah->sl << 28) | 534329222Shselasky (ah->grh.traffic_class << 20) | 535329222Shselasky (ah->grh.flow_label)); 536329222Shselasky memcpy(path->rgid, ah->grh.dgid.raw, 16); 537329222Shselasky } else 538329222Shselasky path->sl_tclass_flowlabel = cpu_to_be32(ah->sl << 28); 539329222Shselasky 540329222Shselasky return 0; 541329222Shselasky} 542329222Shselasky 543329222Shselaskystatic int __mthca_modify_qp(struct ib_qp *ibqp, 544329222Shselasky const struct ib_qp_attr *attr, int attr_mask, 545329222Shselasky enum ib_qp_state cur_state, enum ib_qp_state new_state) 546329222Shselasky{ 547329222Shselasky struct mthca_dev *dev = to_mdev(ibqp->device); 548329222Shselasky struct mthca_qp *qp = to_mqp(ibqp); 549329222Shselasky struct mthca_mailbox *mailbox; 550329222Shselasky struct mthca_qp_param *qp_param; 551329222Shselasky struct mthca_qp_context *qp_context; 552329222Shselasky u32 sqd_event = 0; 553329222Shselasky int err = -EINVAL; 554329222Shselasky 555329222Shselasky mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL); 556329222Shselasky if (IS_ERR(mailbox)) { 557329222Shselasky err = PTR_ERR(mailbox); 558329222Shselasky goto out; 559329222Shselasky } 560329222Shselasky qp_param = mailbox->buf; 561329222Shselasky qp_context = &qp_param->context; 562329222Shselasky memset(qp_param, 0, sizeof *qp_param); 563329222Shselasky 564329222Shselasky qp_context->flags = cpu_to_be32((to_mthca_state(new_state) << 28) | 565329222Shselasky (to_mthca_st(qp->transport) << 16)); 566329222Shselasky qp_context->flags |= cpu_to_be32(MTHCA_QP_BIT_DE); 567329222Shselasky if (!(attr_mask & IB_QP_PATH_MIG_STATE)) 568329222Shselasky qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11); 569329222Shselasky else { 570329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PM_STATE); 571329222Shselasky switch (attr->path_mig_state) { 572329222Shselasky case IB_MIG_MIGRATED: 573329222Shselasky qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_MIGRATED << 11); 574329222Shselasky break; 575329222Shselasky case IB_MIG_REARM: 576329222Shselasky qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_REARM << 11); 577329222Shselasky break; 578329222Shselasky case IB_MIG_ARMED: 579329222Shselasky qp_context->flags |= cpu_to_be32(MTHCA_QP_PM_ARMED << 11); 580329222Shselasky break; 581329222Shselasky } 582329222Shselasky } 583329222Shselasky 584329222Shselasky /* leave tavor_sched_queue as 0 */ 585329222Shselasky 586329222Shselasky if (qp->transport == MLX || qp->transport == UD) 587329222Shselasky qp_context->mtu_msgmax = (IB_MTU_2048 << 5) | 11; 588329222Shselasky else if (attr_mask & IB_QP_PATH_MTU) { 589329222Shselasky if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_2048) { 590329222Shselasky mthca_dbg(dev, "path MTU (%u) is invalid\n", 591329222Shselasky attr->path_mtu); 592329222Shselasky goto out_mailbox; 593329222Shselasky } 594329222Shselasky qp_context->mtu_msgmax = (attr->path_mtu << 5) | 31; 595329222Shselasky } 596329222Shselasky 597329222Shselasky if (mthca_is_memfree(dev)) { 598329222Shselasky if (qp->rq.max) 599329222Shselasky qp_context->rq_size_stride = ilog2(qp->rq.max) << 3; 600329222Shselasky qp_context->rq_size_stride |= qp->rq.wqe_shift - 4; 601329222Shselasky 602329222Shselasky if (qp->sq.max) 603329222Shselasky qp_context->sq_size_stride = ilog2(qp->sq.max) << 3; 604329222Shselasky qp_context->sq_size_stride |= qp->sq.wqe_shift - 4; 605329222Shselasky } 606329222Shselasky 607329222Shselasky /* leave arbel_sched_queue as 0 */ 608329222Shselasky 609329222Shselasky if (qp->ibqp.uobject) 610329222Shselasky qp_context->usr_page = 611329222Shselasky cpu_to_be32(to_mucontext(qp->ibqp.uobject->context)->uar.index); 612329222Shselasky else 613329222Shselasky qp_context->usr_page = cpu_to_be32(dev->driver_uar.index); 614329222Shselasky qp_context->local_qpn = cpu_to_be32(qp->qpn); 615329222Shselasky if (attr_mask & IB_QP_DEST_QPN) { 616329222Shselasky qp_context->remote_qpn = cpu_to_be32(attr->dest_qp_num); 617329222Shselasky } 618329222Shselasky 619329222Shselasky if (qp->transport == MLX) 620329222Shselasky qp_context->pri_path.port_pkey |= 621329222Shselasky cpu_to_be32(qp->port << 24); 622329222Shselasky else { 623329222Shselasky if (attr_mask & IB_QP_PORT) { 624329222Shselasky qp_context->pri_path.port_pkey |= 625329222Shselasky cpu_to_be32(attr->port_num << 24); 626329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PORT_NUM); 627329222Shselasky } 628329222Shselasky } 629329222Shselasky 630329222Shselasky if (attr_mask & IB_QP_PKEY_INDEX) { 631329222Shselasky qp_context->pri_path.port_pkey |= 632329222Shselasky cpu_to_be32(attr->pkey_index); 633329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PKEY_INDEX); 634329222Shselasky } 635329222Shselasky 636329222Shselasky if (attr_mask & IB_QP_RNR_RETRY) { 637329222Shselasky qp_context->alt_path.rnr_retry = qp_context->pri_path.rnr_retry = 638329222Shselasky attr->rnr_retry << 5; 639329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_RETRY | 640329222Shselasky MTHCA_QP_OPTPAR_ALT_RNR_RETRY); 641329222Shselasky } 642329222Shselasky 643329222Shselasky if (attr_mask & IB_QP_AV) { 644329222Shselasky if (mthca_path_set(dev, &attr->ah_attr, &qp_context->pri_path, 645329222Shselasky attr_mask & IB_QP_PORT ? attr->port_num : qp->port)) 646329222Shselasky goto out_mailbox; 647329222Shselasky 648329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_PRIMARY_ADDR_PATH); 649329222Shselasky } 650329222Shselasky 651329222Shselasky if (ibqp->qp_type == IB_QPT_RC && 652329222Shselasky cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) { 653329222Shselasky u8 sched_queue = ibqp->uobject ? 0x2 : 0x1; 654329222Shselasky 655329222Shselasky if (mthca_is_memfree(dev)) 656329222Shselasky qp_context->rlkey_arbel_sched_queue |= sched_queue; 657329222Shselasky else 658329222Shselasky qp_context->tavor_sched_queue |= cpu_to_be32(sched_queue); 659329222Shselasky 660329222Shselasky qp_param->opt_param_mask |= 661329222Shselasky cpu_to_be32(MTHCA_QP_OPTPAR_SCHED_QUEUE); 662329222Shselasky } 663329222Shselasky 664329222Shselasky if (attr_mask & IB_QP_TIMEOUT) { 665329222Shselasky qp_context->pri_path.ackto = attr->timeout << 3; 666329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ACK_TIMEOUT); 667329222Shselasky } 668329222Shselasky 669329222Shselasky if (attr_mask & IB_QP_ALT_PATH) { 670329222Shselasky if (attr->alt_pkey_index >= dev->limits.pkey_table_len) { 671329222Shselasky mthca_dbg(dev, "Alternate P_Key index (%u) too large. max is %d\n", 672329222Shselasky attr->alt_pkey_index, dev->limits.pkey_table_len-1); 673329222Shselasky goto out_mailbox; 674329222Shselasky } 675329222Shselasky 676329222Shselasky if (attr->alt_port_num == 0 || attr->alt_port_num > dev->limits.num_ports) { 677329222Shselasky mthca_dbg(dev, "Alternate port number (%u) is invalid\n", 678329222Shselasky attr->alt_port_num); 679329222Shselasky goto out_mailbox; 680329222Shselasky } 681329222Shselasky 682329222Shselasky if (mthca_path_set(dev, &attr->alt_ah_attr, &qp_context->alt_path, 683329222Shselasky attr->alt_ah_attr.port_num)) 684329222Shselasky goto out_mailbox; 685329222Shselasky 686329222Shselasky qp_context->alt_path.port_pkey |= cpu_to_be32(attr->alt_pkey_index | 687329222Shselasky attr->alt_port_num << 24); 688329222Shselasky qp_context->alt_path.ackto = attr->alt_timeout << 3; 689329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_ALT_ADDR_PATH); 690329222Shselasky } 691329222Shselasky 692329222Shselasky /* leave rdd as 0 */ 693329222Shselasky qp_context->pd = cpu_to_be32(to_mpd(ibqp->pd)->pd_num); 694329222Shselasky /* leave wqe_base as 0 (we always create an MR based at 0 for WQs) */ 695329222Shselasky qp_context->wqe_lkey = cpu_to_be32(qp->mr.ibmr.lkey); 696329222Shselasky qp_context->params1 = cpu_to_be32((MTHCA_ACK_REQ_FREQ << 28) | 697329222Shselasky (MTHCA_FLIGHT_LIMIT << 24) | 698329222Shselasky MTHCA_QP_BIT_SWE); 699329222Shselasky if (qp->sq_policy == IB_SIGNAL_ALL_WR) 700329222Shselasky qp_context->params1 |= cpu_to_be32(MTHCA_QP_BIT_SSC); 701329222Shselasky if (attr_mask & IB_QP_RETRY_CNT) { 702329222Shselasky qp_context->params1 |= cpu_to_be32(attr->retry_cnt << 16); 703329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RETRY_COUNT); 704329222Shselasky } 705329222Shselasky 706329222Shselasky if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) { 707329222Shselasky if (attr->max_rd_atomic) { 708329222Shselasky qp_context->params1 |= 709329222Shselasky cpu_to_be32(MTHCA_QP_BIT_SRE | 710329222Shselasky MTHCA_QP_BIT_SAE); 711329222Shselasky qp_context->params1 |= 712329222Shselasky cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21); 713329222Shselasky } 714329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_SRA_MAX); 715329222Shselasky } 716329222Shselasky 717329222Shselasky if (attr_mask & IB_QP_SQ_PSN) 718329222Shselasky qp_context->next_send_psn = cpu_to_be32(attr->sq_psn); 719329222Shselasky qp_context->cqn_snd = cpu_to_be32(to_mcq(ibqp->send_cq)->cqn); 720329222Shselasky 721329222Shselasky if (mthca_is_memfree(dev)) { 722329222Shselasky qp_context->snd_wqe_base_l = cpu_to_be32(qp->send_wqe_offset); 723329222Shselasky qp_context->snd_db_index = cpu_to_be32(qp->sq.db_index); 724329222Shselasky } 725329222Shselasky 726329222Shselasky if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) { 727329222Shselasky if (attr->max_dest_rd_atomic) 728329222Shselasky qp_context->params2 |= 729329222Shselasky cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21); 730329222Shselasky 731329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RRA_MAX); 732329222Shselasky } 733329222Shselasky 734329222Shselasky if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) { 735329222Shselasky qp_context->params2 |= get_hw_access_flags(qp, attr, attr_mask); 736329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RWE | 737329222Shselasky MTHCA_QP_OPTPAR_RRE | 738329222Shselasky MTHCA_QP_OPTPAR_RAE); 739329222Shselasky } 740329222Shselasky 741329222Shselasky qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RSC); 742329222Shselasky 743329222Shselasky if (ibqp->srq) 744329222Shselasky qp_context->params2 |= cpu_to_be32(MTHCA_QP_BIT_RIC); 745329222Shselasky 746329222Shselasky if (attr_mask & IB_QP_MIN_RNR_TIMER) { 747329222Shselasky qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24); 748329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_RNR_TIMEOUT); 749329222Shselasky } 750329222Shselasky if (attr_mask & IB_QP_RQ_PSN) 751329222Shselasky qp_context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn); 752329222Shselasky 753329222Shselasky qp_context->ra_buff_indx = 754329222Shselasky cpu_to_be32(dev->qp_table.rdb_base + 755329222Shselasky ((qp->qpn & (dev->limits.num_qps - 1)) * MTHCA_RDB_ENTRY_SIZE << 756329222Shselasky dev->qp_table.rdb_shift)); 757329222Shselasky 758329222Shselasky qp_context->cqn_rcv = cpu_to_be32(to_mcq(ibqp->recv_cq)->cqn); 759329222Shselasky 760329222Shselasky if (mthca_is_memfree(dev)) 761329222Shselasky qp_context->rcv_db_index = cpu_to_be32(qp->rq.db_index); 762329222Shselasky 763329222Shselasky if (attr_mask & IB_QP_QKEY) { 764329222Shselasky qp_context->qkey = cpu_to_be32(attr->qkey); 765329222Shselasky qp_param->opt_param_mask |= cpu_to_be32(MTHCA_QP_OPTPAR_Q_KEY); 766329222Shselasky } 767329222Shselasky 768329222Shselasky if (ibqp->srq) 769329222Shselasky qp_context->srqn = cpu_to_be32(1 << 24 | 770329222Shselasky to_msrq(ibqp->srq)->srqn); 771329222Shselasky 772329222Shselasky if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD && 773329222Shselasky attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && 774329222Shselasky attr->en_sqd_async_notify) 775329222Shselasky sqd_event = 1 << 31; 776329222Shselasky 777329222Shselasky err = mthca_MODIFY_QP(dev, cur_state, new_state, qp->qpn, 0, 778329222Shselasky mailbox, sqd_event); 779329222Shselasky if (err) { 780329222Shselasky mthca_warn(dev, "modify QP %d->%d returned %d.\n", 781329222Shselasky cur_state, new_state, err); 782329222Shselasky goto out_mailbox; 783329222Shselasky } 784329222Shselasky 785329222Shselasky qp->state = new_state; 786329222Shselasky if (attr_mask & IB_QP_ACCESS_FLAGS) 787329222Shselasky qp->atomic_rd_en = attr->qp_access_flags; 788329222Shselasky if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) 789329222Shselasky qp->resp_depth = attr->max_dest_rd_atomic; 790329222Shselasky if (attr_mask & IB_QP_PORT) 791329222Shselasky qp->port = attr->port_num; 792329222Shselasky if (attr_mask & IB_QP_ALT_PATH) 793329222Shselasky qp->alt_port = attr->alt_port_num; 794329222Shselasky 795329222Shselasky if (is_sqp(dev, qp)) 796329222Shselasky store_attrs(to_msqp(qp), attr, attr_mask); 797329222Shselasky 798329222Shselasky /* 799329222Shselasky * If we moved QP0 to RTR, bring the IB link up; if we moved 800329222Shselasky * QP0 to RESET or ERROR, bring the link back down. 801329222Shselasky */ 802329222Shselasky if (is_qp0(dev, qp)) { 803329222Shselasky if (cur_state != IB_QPS_RTR && 804329222Shselasky new_state == IB_QPS_RTR) 805329222Shselasky init_port(dev, qp->port); 806329222Shselasky 807329222Shselasky if (cur_state != IB_QPS_RESET && 808329222Shselasky cur_state != IB_QPS_ERR && 809329222Shselasky (new_state == IB_QPS_RESET || 810329222Shselasky new_state == IB_QPS_ERR)) 811329222Shselasky mthca_CLOSE_IB(dev, qp->port); 812329222Shselasky } 813329222Shselasky 814329222Shselasky /* 815329222Shselasky * If we moved a kernel QP to RESET, clean up all old CQ 816329222Shselasky * entries and reinitialize the QP. 817329222Shselasky */ 818329222Shselasky if (new_state == IB_QPS_RESET && !qp->ibqp.uobject) { 819329222Shselasky mthca_cq_clean(dev, to_mcq(qp->ibqp.recv_cq), qp->qpn, 820329222Shselasky qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL); 821329222Shselasky if (qp->ibqp.send_cq != qp->ibqp.recv_cq) 822329222Shselasky mthca_cq_clean(dev, to_mcq(qp->ibqp.send_cq), qp->qpn, NULL); 823329222Shselasky 824329222Shselasky mthca_wq_reset(&qp->sq); 825329222Shselasky qp->sq.last = get_send_wqe(qp, qp->sq.max - 1); 826329222Shselasky 827329222Shselasky mthca_wq_reset(&qp->rq); 828329222Shselasky qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1); 829329222Shselasky 830329222Shselasky if (mthca_is_memfree(dev)) { 831329222Shselasky *qp->sq.db = 0; 832329222Shselasky *qp->rq.db = 0; 833329222Shselasky } 834329222Shselasky } 835329222Shselasky 836329222Shselaskyout_mailbox: 837329222Shselasky mthca_free_mailbox(dev, mailbox); 838329222Shselaskyout: 839329222Shselasky return err; 840329222Shselasky} 841329222Shselasky 842329222Shselaskyint mthca_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask, 843329222Shselasky struct ib_udata *udata) 844329222Shselasky{ 845329222Shselasky struct mthca_dev *dev = to_mdev(ibqp->device); 846329222Shselasky struct mthca_qp *qp = to_mqp(ibqp); 847329222Shselasky enum ib_qp_state cur_state, new_state; 848329222Shselasky int err = -EINVAL; 849329222Shselasky 850329222Shselasky mutex_lock(&qp->mutex); 851329222Shselasky if (attr_mask & IB_QP_CUR_STATE) { 852329222Shselasky cur_state = attr->cur_qp_state; 853329222Shselasky } else { 854329222Shselasky spin_lock_irq(&qp->sq.lock); 855329222Shselasky spin_lock(&qp->rq.lock); 856329222Shselasky cur_state = qp->state; 857329222Shselasky spin_unlock(&qp->rq.lock); 858329222Shselasky spin_unlock_irq(&qp->sq.lock); 859329222Shselasky } 860329222Shselasky 861329222Shselasky new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; 862329222Shselasky 863329222Shselasky if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, attr_mask, 864329222Shselasky IB_LINK_LAYER_UNSPECIFIED)) { 865329222Shselasky mthca_dbg(dev, "Bad QP transition (transport %d) " 866329222Shselasky "%d->%d with attr 0x%08x\n", 867329222Shselasky qp->transport, cur_state, new_state, 868329222Shselasky attr_mask); 869329222Shselasky goto out; 870329222Shselasky } 871329222Shselasky 872329222Shselasky if ((attr_mask & IB_QP_PKEY_INDEX) && 873329222Shselasky attr->pkey_index >= dev->limits.pkey_table_len) { 874329222Shselasky mthca_dbg(dev, "P_Key index (%u) too large. max is %d\n", 875329222Shselasky attr->pkey_index, dev->limits.pkey_table_len-1); 876329222Shselasky goto out; 877329222Shselasky } 878329222Shselasky 879329222Shselasky if ((attr_mask & IB_QP_PORT) && 880329222Shselasky (attr->port_num == 0 || attr->port_num > dev->limits.num_ports)) { 881329222Shselasky mthca_dbg(dev, "Port number (%u) is invalid\n", attr->port_num); 882329222Shselasky goto out; 883329222Shselasky } 884329222Shselasky 885329222Shselasky if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC && 886329222Shselasky attr->max_rd_atomic > dev->limits.max_qp_init_rdma) { 887329222Shselasky mthca_dbg(dev, "Max rdma_atomic as initiator %u too large (max is %d)\n", 888329222Shselasky attr->max_rd_atomic, dev->limits.max_qp_init_rdma); 889329222Shselasky goto out; 890329222Shselasky } 891329222Shselasky 892329222Shselasky if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC && 893329222Shselasky attr->max_dest_rd_atomic > 1 << dev->qp_table.rdb_shift) { 894329222Shselasky mthca_dbg(dev, "Max rdma_atomic as responder %u too large (max %d)\n", 895329222Shselasky attr->max_dest_rd_atomic, 1 << dev->qp_table.rdb_shift); 896329222Shselasky goto out; 897329222Shselasky } 898329222Shselasky 899329222Shselasky if (cur_state == new_state && cur_state == IB_QPS_RESET) { 900329222Shselasky err = 0; 901329222Shselasky goto out; 902329222Shselasky } 903329222Shselasky 904329222Shselasky err = __mthca_modify_qp(ibqp, attr, attr_mask, cur_state, new_state); 905329222Shselasky 906329222Shselaskyout: 907329222Shselasky mutex_unlock(&qp->mutex); 908329222Shselasky return err; 909329222Shselasky} 910329222Shselasky 911329222Shselaskystatic int mthca_max_data_size(struct mthca_dev *dev, struct mthca_qp *qp, int desc_sz) 912329222Shselasky{ 913329222Shselasky /* 914329222Shselasky * Calculate the maximum size of WQE s/g segments, excluding 915329222Shselasky * the next segment and other non-data segments. 916329222Shselasky */ 917329222Shselasky int max_data_size = desc_sz - sizeof (struct mthca_next_seg); 918329222Shselasky 919329222Shselasky switch (qp->transport) { 920329222Shselasky case MLX: 921329222Shselasky max_data_size -= 2 * sizeof (struct mthca_data_seg); 922329222Shselasky break; 923329222Shselasky 924329222Shselasky case UD: 925329222Shselasky if (mthca_is_memfree(dev)) 926329222Shselasky max_data_size -= sizeof (struct mthca_arbel_ud_seg); 927329222Shselasky else 928329222Shselasky max_data_size -= sizeof (struct mthca_tavor_ud_seg); 929329222Shselasky break; 930329222Shselasky 931329222Shselasky default: 932329222Shselasky max_data_size -= sizeof (struct mthca_raddr_seg); 933329222Shselasky break; 934329222Shselasky } 935329222Shselasky 936329222Shselasky return max_data_size; 937329222Shselasky} 938329222Shselasky 939329222Shselaskystatic inline int mthca_max_inline_data(struct mthca_pd *pd, int max_data_size) 940329222Shselasky{ 941329222Shselasky /* We don't support inline data for kernel QPs (yet). */ 942329222Shselasky return pd->ibpd.uobject ? max_data_size - MTHCA_INLINE_HEADER_SIZE : 0; 943329222Shselasky} 944329222Shselasky 945329222Shselaskystatic void mthca_adjust_qp_caps(struct mthca_dev *dev, 946329222Shselasky struct mthca_pd *pd, 947329222Shselasky struct mthca_qp *qp) 948329222Shselasky{ 949329222Shselasky int max_data_size = mthca_max_data_size(dev, qp, 950329222Shselasky min(dev->limits.max_desc_sz, 951329222Shselasky 1 << qp->sq.wqe_shift)); 952329222Shselasky 953329222Shselasky qp->max_inline_data = mthca_max_inline_data(pd, max_data_size); 954329222Shselasky 955329222Shselasky qp->sq.max_gs = min_t(int, dev->limits.max_sg, 956329222Shselasky max_data_size / sizeof (struct mthca_data_seg)); 957329222Shselasky qp->rq.max_gs = min_t(int, dev->limits.max_sg, 958329222Shselasky (min(dev->limits.max_desc_sz, 1 << qp->rq.wqe_shift) - 959329222Shselasky sizeof (struct mthca_next_seg)) / 960329222Shselasky sizeof (struct mthca_data_seg)); 961329222Shselasky} 962329222Shselasky 963329222Shselasky/* 964329222Shselasky * Allocate and register buffer for WQEs. qp->rq.max, sq.max, 965329222Shselasky * rq.max_gs and sq.max_gs must all be assigned. 966329222Shselasky * mthca_alloc_wqe_buf will calculate rq.wqe_shift and 967329222Shselasky * sq.wqe_shift (as well as send_wqe_offset, is_direct, and 968329222Shselasky * queue) 969329222Shselasky */ 970329222Shselaskystatic int mthca_alloc_wqe_buf(struct mthca_dev *dev, 971329222Shselasky struct mthca_pd *pd, 972329222Shselasky struct mthca_qp *qp) 973329222Shselasky{ 974329222Shselasky int size; 975329222Shselasky int err = -ENOMEM; 976329222Shselasky 977329222Shselasky size = sizeof (struct mthca_next_seg) + 978329222Shselasky qp->rq.max_gs * sizeof (struct mthca_data_seg); 979329222Shselasky 980329222Shselasky if (size > dev->limits.max_desc_sz) 981329222Shselasky return -EINVAL; 982329222Shselasky 983329222Shselasky for (qp->rq.wqe_shift = 6; 1 << qp->rq.wqe_shift < size; 984329222Shselasky qp->rq.wqe_shift++) 985329222Shselasky ; /* nothing */ 986329222Shselasky 987329222Shselasky size = qp->sq.max_gs * sizeof (struct mthca_data_seg); 988329222Shselasky switch (qp->transport) { 989329222Shselasky case MLX: 990329222Shselasky size += 2 * sizeof (struct mthca_data_seg); 991329222Shselasky break; 992329222Shselasky 993329222Shselasky case UD: 994329222Shselasky size += mthca_is_memfree(dev) ? 995329222Shselasky sizeof (struct mthca_arbel_ud_seg) : 996329222Shselasky sizeof (struct mthca_tavor_ud_seg); 997329222Shselasky break; 998329222Shselasky 999329222Shselasky case UC: 1000329222Shselasky size += sizeof (struct mthca_raddr_seg); 1001329222Shselasky break; 1002329222Shselasky 1003329222Shselasky case RC: 1004329222Shselasky size += sizeof (struct mthca_raddr_seg); 1005329222Shselasky /* 1006329222Shselasky * An atomic op will require an atomic segment, a 1007329222Shselasky * remote address segment and one scatter entry. 1008329222Shselasky */ 1009329222Shselasky size = max_t(int, size, 1010329222Shselasky sizeof (struct mthca_atomic_seg) + 1011329222Shselasky sizeof (struct mthca_raddr_seg) + 1012329222Shselasky sizeof (struct mthca_data_seg)); 1013329222Shselasky break; 1014329222Shselasky 1015329222Shselasky default: 1016329222Shselasky break; 1017329222Shselasky } 1018329222Shselasky 1019329222Shselasky /* Make sure that we have enough space for a bind request */ 1020329222Shselasky size = max_t(int, size, sizeof (struct mthca_bind_seg)); 1021329222Shselasky 1022329222Shselasky size += sizeof (struct mthca_next_seg); 1023329222Shselasky 1024329222Shselasky if (size > dev->limits.max_desc_sz) 1025329222Shselasky return -EINVAL; 1026329222Shselasky 1027329222Shselasky for (qp->sq.wqe_shift = 6; 1 << qp->sq.wqe_shift < size; 1028329222Shselasky qp->sq.wqe_shift++) 1029329222Shselasky ; /* nothing */ 1030329222Shselasky 1031329222Shselasky qp->send_wqe_offset = ALIGN(qp->rq.max << qp->rq.wqe_shift, 1032329222Shselasky 1 << qp->sq.wqe_shift); 1033329222Shselasky 1034329222Shselasky /* 1035329222Shselasky * If this is a userspace QP, we don't actually have to 1036329222Shselasky * allocate anything. All we need is to calculate the WQE 1037329222Shselasky * sizes and the send_wqe_offset, so we're done now. 1038329222Shselasky */ 1039329222Shselasky if (pd->ibpd.uobject) 1040329222Shselasky return 0; 1041329222Shselasky 1042329222Shselasky size = PAGE_ALIGN(qp->send_wqe_offset + 1043329222Shselasky (qp->sq.max << qp->sq.wqe_shift)); 1044329222Shselasky 1045329222Shselasky qp->wrid = kmalloc((qp->rq.max + qp->sq.max) * sizeof (u64), 1046329222Shselasky GFP_KERNEL); 1047329222Shselasky if (!qp->wrid) 1048329222Shselasky goto err_out; 1049329222Shselasky 1050329222Shselasky err = mthca_buf_alloc(dev, size, MTHCA_MAX_DIRECT_QP_SIZE, 1051329222Shselasky &qp->queue, &qp->is_direct, pd, 0, &qp->mr); 1052329222Shselasky if (err) 1053329222Shselasky goto err_out; 1054329222Shselasky 1055329222Shselasky return 0; 1056329222Shselasky 1057329222Shselaskyerr_out: 1058329222Shselasky kfree(qp->wrid); 1059329222Shselasky return err; 1060329222Shselasky} 1061329222Shselasky 1062329222Shselaskystatic void mthca_free_wqe_buf(struct mthca_dev *dev, 1063329222Shselasky struct mthca_qp *qp) 1064329222Shselasky{ 1065329222Shselasky mthca_buf_free(dev, PAGE_ALIGN(qp->send_wqe_offset + 1066329222Shselasky (qp->sq.max << qp->sq.wqe_shift)), 1067329222Shselasky &qp->queue, qp->is_direct, &qp->mr); 1068329222Shselasky kfree(qp->wrid); 1069329222Shselasky} 1070329222Shselasky 1071329222Shselaskystatic int mthca_map_memfree(struct mthca_dev *dev, 1072329222Shselasky struct mthca_qp *qp) 1073329222Shselasky{ 1074329222Shselasky int ret; 1075329222Shselasky 1076329222Shselasky if (mthca_is_memfree(dev)) { 1077329222Shselasky ret = mthca_table_get(dev, dev->qp_table.qp_table, qp->qpn); 1078329222Shselasky if (ret) 1079329222Shselasky return ret; 1080329222Shselasky 1081329222Shselasky ret = mthca_table_get(dev, dev->qp_table.eqp_table, qp->qpn); 1082329222Shselasky if (ret) 1083329222Shselasky goto err_qpc; 1084329222Shselasky 1085329222Shselasky ret = mthca_table_get(dev, dev->qp_table.rdb_table, 1086329222Shselasky qp->qpn << dev->qp_table.rdb_shift); 1087329222Shselasky if (ret) 1088329222Shselasky goto err_eqpc; 1089329222Shselasky 1090329222Shselasky } 1091329222Shselasky 1092329222Shselasky return 0; 1093329222Shselasky 1094329222Shselaskyerr_eqpc: 1095329222Shselasky mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn); 1096329222Shselasky 1097329222Shselaskyerr_qpc: 1098329222Shselasky mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn); 1099329222Shselasky 1100329222Shselasky return ret; 1101329222Shselasky} 1102329222Shselasky 1103329222Shselaskystatic void mthca_unmap_memfree(struct mthca_dev *dev, 1104329222Shselasky struct mthca_qp *qp) 1105329222Shselasky{ 1106329222Shselasky mthca_table_put(dev, dev->qp_table.rdb_table, 1107329222Shselasky qp->qpn << dev->qp_table.rdb_shift); 1108329222Shselasky mthca_table_put(dev, dev->qp_table.eqp_table, qp->qpn); 1109329222Shselasky mthca_table_put(dev, dev->qp_table.qp_table, qp->qpn); 1110329222Shselasky} 1111329222Shselasky 1112329222Shselaskystatic int mthca_alloc_memfree(struct mthca_dev *dev, 1113329222Shselasky struct mthca_qp *qp) 1114329222Shselasky{ 1115329222Shselasky if (mthca_is_memfree(dev)) { 1116329222Shselasky qp->rq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_RQ, 1117329222Shselasky qp->qpn, &qp->rq.db); 1118329222Shselasky if (qp->rq.db_index < 0) 1119329222Shselasky return -ENOMEM; 1120329222Shselasky 1121329222Shselasky qp->sq.db_index = mthca_alloc_db(dev, MTHCA_DB_TYPE_SQ, 1122329222Shselasky qp->qpn, &qp->sq.db); 1123329222Shselasky if (qp->sq.db_index < 0) { 1124329222Shselasky mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index); 1125329222Shselasky return -ENOMEM; 1126329222Shselasky } 1127329222Shselasky } 1128329222Shselasky 1129329222Shselasky return 0; 1130329222Shselasky} 1131329222Shselasky 1132329222Shselaskystatic void mthca_free_memfree(struct mthca_dev *dev, 1133329222Shselasky struct mthca_qp *qp) 1134329222Shselasky{ 1135329222Shselasky if (mthca_is_memfree(dev)) { 1136329222Shselasky mthca_free_db(dev, MTHCA_DB_TYPE_SQ, qp->sq.db_index); 1137329222Shselasky mthca_free_db(dev, MTHCA_DB_TYPE_RQ, qp->rq.db_index); 1138329222Shselasky } 1139329222Shselasky} 1140329222Shselasky 1141329222Shselaskystatic int mthca_alloc_qp_common(struct mthca_dev *dev, 1142329222Shselasky struct mthca_pd *pd, 1143329222Shselasky struct mthca_cq *send_cq, 1144329222Shselasky struct mthca_cq *recv_cq, 1145329222Shselasky enum ib_sig_type send_policy, 1146329222Shselasky struct mthca_qp *qp) 1147329222Shselasky{ 1148329222Shselasky int ret; 1149329222Shselasky int i; 1150329222Shselasky struct mthca_next_seg *next; 1151329222Shselasky 1152329222Shselasky qp->refcount = 1; 1153329222Shselasky init_waitqueue_head(&qp->wait); 1154329222Shselasky mutex_init(&qp->mutex); 1155329222Shselasky qp->state = IB_QPS_RESET; 1156329222Shselasky qp->atomic_rd_en = 0; 1157329222Shselasky qp->resp_depth = 0; 1158329222Shselasky qp->sq_policy = send_policy; 1159329222Shselasky mthca_wq_reset(&qp->sq); 1160329222Shselasky mthca_wq_reset(&qp->rq); 1161329222Shselasky 1162329222Shselasky spin_lock_init(&qp->sq.lock); 1163329222Shselasky spin_lock_init(&qp->rq.lock); 1164329222Shselasky 1165329222Shselasky ret = mthca_map_memfree(dev, qp); 1166329222Shselasky if (ret) 1167329222Shselasky return ret; 1168329222Shselasky 1169329222Shselasky ret = mthca_alloc_wqe_buf(dev, pd, qp); 1170329222Shselasky if (ret) { 1171329222Shselasky mthca_unmap_memfree(dev, qp); 1172329222Shselasky return ret; 1173329222Shselasky } 1174329222Shselasky 1175329222Shselasky mthca_adjust_qp_caps(dev, pd, qp); 1176329222Shselasky 1177329222Shselasky /* 1178329222Shselasky * If this is a userspace QP, we're done now. The doorbells 1179329222Shselasky * will be allocated and buffers will be initialized in 1180329222Shselasky * userspace. 1181329222Shselasky */ 1182329222Shselasky if (pd->ibpd.uobject) 1183329222Shselasky return 0; 1184329222Shselasky 1185329222Shselasky ret = mthca_alloc_memfree(dev, qp); 1186329222Shselasky if (ret) { 1187329222Shselasky mthca_free_wqe_buf(dev, qp); 1188329222Shselasky mthca_unmap_memfree(dev, qp); 1189329222Shselasky return ret; 1190329222Shselasky } 1191329222Shselasky 1192329222Shselasky if (mthca_is_memfree(dev)) { 1193329222Shselasky struct mthca_data_seg *scatter; 1194329222Shselasky int size = (sizeof (struct mthca_next_seg) + 1195329222Shselasky qp->rq.max_gs * sizeof (struct mthca_data_seg)) / 16; 1196329222Shselasky 1197329222Shselasky for (i = 0; i < qp->rq.max; ++i) { 1198329222Shselasky next = get_recv_wqe(qp, i); 1199329222Shselasky next->nda_op = cpu_to_be32(((i + 1) & (qp->rq.max - 1)) << 1200329222Shselasky qp->rq.wqe_shift); 1201329222Shselasky next->ee_nds = cpu_to_be32(size); 1202329222Shselasky 1203329222Shselasky for (scatter = (void *) (next + 1); 1204329222Shselasky (void *) scatter < (void *) next + (1 << qp->rq.wqe_shift); 1205329222Shselasky ++scatter) 1206329222Shselasky scatter->lkey = cpu_to_be32(MTHCA_INVAL_LKEY); 1207329222Shselasky } 1208329222Shselasky 1209329222Shselasky for (i = 0; i < qp->sq.max; ++i) { 1210329222Shselasky next = get_send_wqe(qp, i); 1211329222Shselasky next->nda_op = cpu_to_be32((((i + 1) & (qp->sq.max - 1)) << 1212329222Shselasky qp->sq.wqe_shift) + 1213329222Shselasky qp->send_wqe_offset); 1214329222Shselasky } 1215329222Shselasky } else { 1216329222Shselasky for (i = 0; i < qp->rq.max; ++i) { 1217329222Shselasky next = get_recv_wqe(qp, i); 1218329222Shselasky next->nda_op = htonl((((i + 1) % qp->rq.max) << 1219329222Shselasky qp->rq.wqe_shift) | 1); 1220329222Shselasky } 1221329222Shselasky 1222329222Shselasky } 1223329222Shselasky 1224329222Shselasky qp->sq.last = get_send_wqe(qp, qp->sq.max - 1); 1225329222Shselasky qp->rq.last = get_recv_wqe(qp, qp->rq.max - 1); 1226329222Shselasky 1227329222Shselasky return 0; 1228329222Shselasky} 1229329222Shselasky 1230329222Shselaskystatic int mthca_set_qp_size(struct mthca_dev *dev, struct ib_qp_cap *cap, 1231329222Shselasky struct mthca_pd *pd, struct mthca_qp *qp) 1232329222Shselasky{ 1233329222Shselasky int max_data_size = mthca_max_data_size(dev, qp, dev->limits.max_desc_sz); 1234329222Shselasky 1235329222Shselasky /* Sanity check QP size before proceeding */ 1236329222Shselasky if (cap->max_send_wr > dev->limits.max_wqes || 1237329222Shselasky cap->max_recv_wr > dev->limits.max_wqes || 1238329222Shselasky cap->max_send_sge > dev->limits.max_sg || 1239329222Shselasky cap->max_recv_sge > dev->limits.max_sg || 1240329222Shselasky cap->max_inline_data > mthca_max_inline_data(pd, max_data_size)) 1241329222Shselasky return -EINVAL; 1242329222Shselasky 1243329222Shselasky /* 1244329222Shselasky * For MLX transport we need 2 extra send gather entries: 1245329222Shselasky * one for the header and one for the checksum at the end 1246329222Shselasky */ 1247329222Shselasky if (qp->transport == MLX && cap->max_send_sge + 2 > dev->limits.max_sg) 1248329222Shselasky return -EINVAL; 1249329222Shselasky 1250329222Shselasky if (mthca_is_memfree(dev)) { 1251329222Shselasky qp->rq.max = cap->max_recv_wr ? 1252329222Shselasky roundup_pow_of_two(cap->max_recv_wr) : 0; 1253329222Shselasky qp->sq.max = cap->max_send_wr ? 1254329222Shselasky roundup_pow_of_two(cap->max_send_wr) : 0; 1255329222Shselasky } else { 1256329222Shselasky qp->rq.max = cap->max_recv_wr; 1257329222Shselasky qp->sq.max = cap->max_send_wr; 1258329222Shselasky } 1259329222Shselasky 1260329222Shselasky qp->rq.max_gs = cap->max_recv_sge; 1261329222Shselasky qp->sq.max_gs = max_t(int, cap->max_send_sge, 1262329222Shselasky ALIGN(cap->max_inline_data + MTHCA_INLINE_HEADER_SIZE, 1263329222Shselasky MTHCA_INLINE_CHUNK_SIZE) / 1264329222Shselasky sizeof (struct mthca_data_seg)); 1265329222Shselasky 1266329222Shselasky return 0; 1267329222Shselasky} 1268329222Shselasky 1269329222Shselaskyint mthca_alloc_qp(struct mthca_dev *dev, 1270329222Shselasky struct mthca_pd *pd, 1271329222Shselasky struct mthca_cq *send_cq, 1272329222Shselasky struct mthca_cq *recv_cq, 1273329222Shselasky enum ib_qp_type type, 1274329222Shselasky enum ib_sig_type send_policy, 1275329222Shselasky struct ib_qp_cap *cap, 1276329222Shselasky struct mthca_qp *qp) 1277329222Shselasky{ 1278329222Shselasky int err; 1279329222Shselasky 1280329222Shselasky switch (type) { 1281329222Shselasky case IB_QPT_RC: qp->transport = RC; break; 1282329222Shselasky case IB_QPT_UC: qp->transport = UC; break; 1283329222Shselasky case IB_QPT_UD: qp->transport = UD; break; 1284329222Shselasky default: return -EINVAL; 1285329222Shselasky } 1286329222Shselasky 1287329222Shselasky err = mthca_set_qp_size(dev, cap, pd, qp); 1288329222Shselasky if (err) 1289329222Shselasky return err; 1290329222Shselasky 1291329222Shselasky qp->qpn = mthca_alloc(&dev->qp_table.alloc); 1292329222Shselasky if (qp->qpn == -1) 1293329222Shselasky return -ENOMEM; 1294329222Shselasky 1295329222Shselasky /* initialize port to zero for error-catching. */ 1296329222Shselasky qp->port = 0; 1297329222Shselasky 1298329222Shselasky err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq, 1299329222Shselasky send_policy, qp); 1300329222Shselasky if (err) { 1301329222Shselasky mthca_free(&dev->qp_table.alloc, qp->qpn); 1302329222Shselasky return err; 1303329222Shselasky } 1304329222Shselasky 1305329222Shselasky spin_lock_irq(&dev->qp_table.lock); 1306329222Shselasky mthca_array_set(&dev->qp_table.qp, 1307329222Shselasky qp->qpn & (dev->limits.num_qps - 1), qp); 1308329222Shselasky spin_unlock_irq(&dev->qp_table.lock); 1309329222Shselasky 1310329222Shselasky return 0; 1311329222Shselasky} 1312329222Shselasky 1313329222Shselaskystatic void mthca_lock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq) 1314329222Shselasky __acquires(&send_cq->lock) __acquires(&recv_cq->lock) 1315329222Shselasky{ 1316329222Shselasky if (send_cq == recv_cq) { 1317329222Shselasky spin_lock_irq(&send_cq->lock); 1318329222Shselasky __acquire(&recv_cq->lock); 1319329222Shselasky } else if (send_cq->cqn < recv_cq->cqn) { 1320329222Shselasky spin_lock_irq(&send_cq->lock); 1321329222Shselasky spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING); 1322329222Shselasky } else { 1323329222Shselasky spin_lock_irq(&recv_cq->lock); 1324329222Shselasky spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING); 1325329222Shselasky } 1326329222Shselasky} 1327329222Shselasky 1328329222Shselaskystatic void mthca_unlock_cqs(struct mthca_cq *send_cq, struct mthca_cq *recv_cq) 1329329222Shselasky __releases(&send_cq->lock) __releases(&recv_cq->lock) 1330329222Shselasky{ 1331329222Shselasky if (send_cq == recv_cq) { 1332329222Shselasky __release(&recv_cq->lock); 1333329222Shselasky spin_unlock_irq(&send_cq->lock); 1334329222Shselasky } else if (send_cq->cqn < recv_cq->cqn) { 1335329222Shselasky spin_unlock(&recv_cq->lock); 1336329222Shselasky spin_unlock_irq(&send_cq->lock); 1337329222Shselasky } else { 1338329222Shselasky spin_unlock(&send_cq->lock); 1339329222Shselasky spin_unlock_irq(&recv_cq->lock); 1340329222Shselasky } 1341329222Shselasky} 1342329222Shselasky 1343329222Shselaskyint mthca_alloc_sqp(struct mthca_dev *dev, 1344329222Shselasky struct mthca_pd *pd, 1345329222Shselasky struct mthca_cq *send_cq, 1346329222Shselasky struct mthca_cq *recv_cq, 1347329222Shselasky enum ib_sig_type send_policy, 1348329222Shselasky struct ib_qp_cap *cap, 1349329222Shselasky int qpn, 1350329222Shselasky int port, 1351329222Shselasky struct mthca_sqp *sqp) 1352329222Shselasky{ 1353329222Shselasky u32 mqpn = qpn * 2 + dev->qp_table.sqp_start + port - 1; 1354329222Shselasky int err; 1355329222Shselasky 1356329222Shselasky sqp->qp.transport = MLX; 1357329222Shselasky err = mthca_set_qp_size(dev, cap, pd, &sqp->qp); 1358329222Shselasky if (err) 1359329222Shselasky return err; 1360329222Shselasky 1361329222Shselasky sqp->header_buf_size = sqp->qp.sq.max * MTHCA_UD_HEADER_SIZE; 1362329222Shselasky sqp->header_buf = dma_alloc_coherent(&dev->pdev->dev, sqp->header_buf_size, 1363329222Shselasky &sqp->header_dma, GFP_KERNEL); 1364329222Shselasky if (!sqp->header_buf) 1365329222Shselasky return -ENOMEM; 1366329222Shselasky 1367329222Shselasky spin_lock_irq(&dev->qp_table.lock); 1368329222Shselasky if (mthca_array_get(&dev->qp_table.qp, mqpn)) 1369329222Shselasky err = -EBUSY; 1370329222Shselasky else 1371329222Shselasky mthca_array_set(&dev->qp_table.qp, mqpn, sqp); 1372329222Shselasky spin_unlock_irq(&dev->qp_table.lock); 1373329222Shselasky 1374329222Shselasky if (err) 1375329222Shselasky goto err_out; 1376329222Shselasky 1377329222Shselasky sqp->qp.port = port; 1378329222Shselasky sqp->qp.qpn = mqpn; 1379329222Shselasky sqp->qp.transport = MLX; 1380329222Shselasky 1381329222Shselasky err = mthca_alloc_qp_common(dev, pd, send_cq, recv_cq, 1382329222Shselasky send_policy, &sqp->qp); 1383329222Shselasky if (err) 1384329222Shselasky goto err_out_free; 1385329222Shselasky 1386329222Shselasky atomic_inc(&pd->sqp_count); 1387329222Shselasky 1388329222Shselasky return 0; 1389329222Shselasky 1390329222Shselasky err_out_free: 1391329222Shselasky /* 1392329222Shselasky * Lock CQs here, so that CQ polling code can do QP lookup 1393329222Shselasky * without taking a lock. 1394329222Shselasky */ 1395329222Shselasky mthca_lock_cqs(send_cq, recv_cq); 1396329222Shselasky 1397329222Shselasky spin_lock(&dev->qp_table.lock); 1398329222Shselasky mthca_array_clear(&dev->qp_table.qp, mqpn); 1399329222Shselasky spin_unlock(&dev->qp_table.lock); 1400329222Shselasky 1401329222Shselasky mthca_unlock_cqs(send_cq, recv_cq); 1402329222Shselasky 1403329222Shselasky err_out: 1404329222Shselasky dma_free_coherent(&dev->pdev->dev, sqp->header_buf_size, 1405329222Shselasky sqp->header_buf, sqp->header_dma); 1406329222Shselasky 1407329222Shselasky return err; 1408329222Shselasky} 1409329222Shselasky 1410329222Shselaskystatic inline int get_qp_refcount(struct mthca_dev *dev, struct mthca_qp *qp) 1411329222Shselasky{ 1412329222Shselasky int c; 1413329222Shselasky 1414329222Shselasky spin_lock_irq(&dev->qp_table.lock); 1415329222Shselasky c = qp->refcount; 1416329222Shselasky spin_unlock_irq(&dev->qp_table.lock); 1417329222Shselasky 1418329222Shselasky return c; 1419329222Shselasky} 1420329222Shselasky 1421329222Shselaskyvoid mthca_free_qp(struct mthca_dev *dev, 1422329222Shselasky struct mthca_qp *qp) 1423329222Shselasky{ 1424329222Shselasky struct mthca_cq *send_cq; 1425329222Shselasky struct mthca_cq *recv_cq; 1426329222Shselasky 1427329222Shselasky send_cq = to_mcq(qp->ibqp.send_cq); 1428329222Shselasky recv_cq = to_mcq(qp->ibqp.recv_cq); 1429329222Shselasky 1430329222Shselasky /* 1431329222Shselasky * Lock CQs here, so that CQ polling code can do QP lookup 1432329222Shselasky * without taking a lock. 1433329222Shselasky */ 1434329222Shselasky mthca_lock_cqs(send_cq, recv_cq); 1435329222Shselasky 1436329222Shselasky spin_lock(&dev->qp_table.lock); 1437329222Shselasky mthca_array_clear(&dev->qp_table.qp, 1438329222Shselasky qp->qpn & (dev->limits.num_qps - 1)); 1439329222Shselasky --qp->refcount; 1440329222Shselasky spin_unlock(&dev->qp_table.lock); 1441329222Shselasky 1442329222Shselasky mthca_unlock_cqs(send_cq, recv_cq); 1443329222Shselasky 1444329222Shselasky wait_event(qp->wait, !get_qp_refcount(dev, qp)); 1445329222Shselasky 1446329222Shselasky if (qp->state != IB_QPS_RESET) 1447329222Shselasky mthca_MODIFY_QP(dev, qp->state, IB_QPS_RESET, qp->qpn, 0, 1448329222Shselasky NULL, 0); 1449329222Shselasky 1450329222Shselasky /* 1451329222Shselasky * If this is a userspace QP, the buffers, MR, CQs and so on 1452329222Shselasky * will be cleaned up in userspace, so all we have to do is 1453329222Shselasky * unref the mem-free tables and free the QPN in our table. 1454329222Shselasky */ 1455329222Shselasky if (!qp->ibqp.uobject) { 1456329222Shselasky mthca_cq_clean(dev, recv_cq, qp->qpn, 1457329222Shselasky qp->ibqp.srq ? to_msrq(qp->ibqp.srq) : NULL); 1458329222Shselasky if (send_cq != recv_cq) 1459329222Shselasky mthca_cq_clean(dev, send_cq, qp->qpn, NULL); 1460329222Shselasky 1461329222Shselasky mthca_free_memfree(dev, qp); 1462329222Shselasky mthca_free_wqe_buf(dev, qp); 1463329222Shselasky } 1464329222Shselasky 1465329222Shselasky mthca_unmap_memfree(dev, qp); 1466329222Shselasky 1467329222Shselasky if (is_sqp(dev, qp)) { 1468329222Shselasky atomic_dec(&(to_mpd(qp->ibqp.pd)->sqp_count)); 1469329222Shselasky dma_free_coherent(&dev->pdev->dev, 1470329222Shselasky to_msqp(qp)->header_buf_size, 1471329222Shselasky to_msqp(qp)->header_buf, 1472329222Shselasky to_msqp(qp)->header_dma); 1473329222Shselasky } else 1474329222Shselasky mthca_free(&dev->qp_table.alloc, qp->qpn); 1475329222Shselasky} 1476329222Shselasky 1477329222Shselasky/* Create UD header for an MLX send and build a data segment for it */ 1478329222Shselaskystatic int build_mlx_header(struct mthca_dev *dev, struct mthca_sqp *sqp, 1479329222Shselasky int ind, struct ib_ud_wr *wr, 1480329222Shselasky struct mthca_mlx_seg *mlx, 1481329222Shselasky struct mthca_data_seg *data) 1482329222Shselasky{ 1483329222Shselasky int header_size; 1484329222Shselasky int err; 1485329222Shselasky u16 pkey; 1486329222Shselasky 1487329222Shselasky ib_ud_header_init(256, /* assume a MAD */ 1, 0, 0, 1488329222Shselasky mthca_ah_grh_present(to_mah(wr->ah)), 0, 0, 0, 1489329222Shselasky &sqp->ud_header); 1490329222Shselasky 1491329222Shselasky err = mthca_read_ah(dev, to_mah(wr->ah), &sqp->ud_header); 1492329222Shselasky if (err) 1493329222Shselasky return err; 1494329222Shselasky mlx->flags &= ~cpu_to_be32(MTHCA_NEXT_SOLICIT | 1); 1495329222Shselasky mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MTHCA_MLX_VL15 : 0) | 1496329222Shselasky (sqp->ud_header.lrh.destination_lid == 1497329222Shselasky IB_LID_PERMISSIVE ? MTHCA_MLX_SLR : 0) | 1498329222Shselasky (sqp->ud_header.lrh.service_level << 8)); 1499329222Shselasky mlx->rlid = sqp->ud_header.lrh.destination_lid; 1500329222Shselasky mlx->vcrc = 0; 1501329222Shselasky 1502329222Shselasky switch (wr->wr.opcode) { 1503329222Shselasky case IB_WR_SEND: 1504329222Shselasky sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY; 1505329222Shselasky sqp->ud_header.immediate_present = 0; 1506329222Shselasky break; 1507329222Shselasky case IB_WR_SEND_WITH_IMM: 1508329222Shselasky sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE; 1509329222Shselasky sqp->ud_header.immediate_present = 1; 1510329222Shselasky sqp->ud_header.immediate_data = wr->wr.ex.imm_data; 1511329222Shselasky break; 1512329222Shselasky default: 1513329222Shselasky return -EINVAL; 1514329222Shselasky } 1515329222Shselasky 1516329222Shselasky sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 : 0; 1517329222Shselasky if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE) 1518329222Shselasky sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE; 1519329222Shselasky sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED); 1520329222Shselasky if (!sqp->qp.ibqp.qp_num) 1521329222Shselasky ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port, 1522329222Shselasky sqp->pkey_index, &pkey); 1523329222Shselasky else 1524329222Shselasky ib_get_cached_pkey(&dev->ib_dev, sqp->qp.port, 1525329222Shselasky wr->pkey_index, &pkey); 1526329222Shselasky sqp->ud_header.bth.pkey = cpu_to_be16(pkey); 1527329222Shselasky sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn); 1528329222Shselasky sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1)); 1529329222Shselasky sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ? 1530329222Shselasky sqp->qkey : wr->remote_qkey); 1531329222Shselasky sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num); 1532329222Shselasky 1533329222Shselasky header_size = ib_ud_header_pack(&sqp->ud_header, 1534329222Shselasky sqp->header_buf + 1535329222Shselasky ind * MTHCA_UD_HEADER_SIZE); 1536329222Shselasky 1537329222Shselasky data->byte_count = cpu_to_be32(header_size); 1538329222Shselasky data->lkey = cpu_to_be32(to_mpd(sqp->qp.ibqp.pd)->ntmr.ibmr.lkey); 1539329222Shselasky data->addr = cpu_to_be64(sqp->header_dma + 1540329222Shselasky ind * MTHCA_UD_HEADER_SIZE); 1541329222Shselasky 1542329222Shselasky return 0; 1543329222Shselasky} 1544329222Shselasky 1545329222Shselaskystatic inline int mthca_wq_overflow(struct mthca_wq *wq, int nreq, 1546329222Shselasky struct ib_cq *ib_cq) 1547329222Shselasky{ 1548329222Shselasky unsigned cur; 1549329222Shselasky struct mthca_cq *cq; 1550329222Shselasky 1551329222Shselasky cur = wq->head - wq->tail; 1552329222Shselasky if (likely(cur + nreq < wq->max)) 1553329222Shselasky return 0; 1554329222Shselasky 1555329222Shselasky cq = to_mcq(ib_cq); 1556329222Shselasky spin_lock(&cq->lock); 1557329222Shselasky cur = wq->head - wq->tail; 1558329222Shselasky spin_unlock(&cq->lock); 1559329222Shselasky 1560329222Shselasky return cur + nreq >= wq->max; 1561329222Shselasky} 1562329222Shselasky 1563329222Shselaskystatic __always_inline void set_raddr_seg(struct mthca_raddr_seg *rseg, 1564329222Shselasky u64 remote_addr, u32 rkey) 1565329222Shselasky{ 1566329222Shselasky rseg->raddr = cpu_to_be64(remote_addr); 1567329222Shselasky rseg->rkey = cpu_to_be32(rkey); 1568329222Shselasky rseg->reserved = 0; 1569329222Shselasky} 1570329222Shselasky 1571329222Shselaskystatic __always_inline void set_atomic_seg(struct mthca_atomic_seg *aseg, 1572329222Shselasky struct ib_atomic_wr *wr) 1573329222Shselasky{ 1574329222Shselasky if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) { 1575329222Shselasky aseg->swap_add = cpu_to_be64(wr->swap); 1576329222Shselasky aseg->compare = cpu_to_be64(wr->compare_add); 1577329222Shselasky } else { 1578329222Shselasky aseg->swap_add = cpu_to_be64(wr->compare_add); 1579329222Shselasky aseg->compare = 0; 1580329222Shselasky } 1581329222Shselasky 1582329222Shselasky} 1583329222Shselasky 1584329222Shselaskystatic void set_tavor_ud_seg(struct mthca_tavor_ud_seg *useg, 1585329222Shselasky struct ib_ud_wr *wr) 1586329222Shselasky{ 1587329222Shselasky useg->lkey = cpu_to_be32(to_mah(wr->ah)->key); 1588329222Shselasky useg->av_addr = cpu_to_be64(to_mah(wr->ah)->avdma); 1589329222Shselasky useg->dqpn = cpu_to_be32(wr->remote_qpn); 1590329222Shselasky useg->qkey = cpu_to_be32(wr->remote_qkey); 1591329222Shselasky 1592329222Shselasky} 1593329222Shselasky 1594329222Shselaskystatic void set_arbel_ud_seg(struct mthca_arbel_ud_seg *useg, 1595329222Shselasky struct ib_ud_wr *wr) 1596329222Shselasky{ 1597329222Shselasky memcpy(useg->av, to_mah(wr->ah)->av, MTHCA_AV_SIZE); 1598329222Shselasky useg->dqpn = cpu_to_be32(wr->remote_qpn); 1599329222Shselasky useg->qkey = cpu_to_be32(wr->remote_qkey); 1600329222Shselasky} 1601329222Shselasky 1602329222Shselaskyint mthca_tavor_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 1603329222Shselasky struct ib_send_wr **bad_wr) 1604329222Shselasky{ 1605329222Shselasky struct mthca_dev *dev = to_mdev(ibqp->device); 1606329222Shselasky struct mthca_qp *qp = to_mqp(ibqp); 1607329222Shselasky void *wqe; 1608329222Shselasky void *prev_wqe; 1609329222Shselasky unsigned long flags; 1610329222Shselasky int err = 0; 1611329222Shselasky int nreq; 1612329222Shselasky int i; 1613329222Shselasky int size; 1614329222Shselasky /* 1615329222Shselasky * f0 and size0 are only used if nreq != 0, and they will 1616329222Shselasky * always be initialized the first time through the main loop 1617329222Shselasky * before nreq is incremented. So nreq cannot become non-zero 1618329222Shselasky * without initializing f0 and size0, and they are in fact 1619329222Shselasky * never used uninitialized. 1620329222Shselasky */ 1621329222Shselasky int uninitialized_var(size0); 1622329222Shselasky u32 uninitialized_var(f0); 1623329222Shselasky int ind; 1624329222Shselasky u8 op0 = 0; 1625329222Shselasky 1626329222Shselasky spin_lock_irqsave(&qp->sq.lock, flags); 1627329222Shselasky 1628329222Shselasky /* XXX check that state is OK to post send */ 1629329222Shselasky 1630329222Shselasky ind = qp->sq.next_ind; 1631329222Shselasky 1632329222Shselasky for (nreq = 0; wr; ++nreq, wr = wr->next) { 1633329222Shselasky if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) { 1634329222Shselasky mthca_err(dev, "SQ %06x full (%u head, %u tail," 1635329222Shselasky " %d max, %d nreq)\n", qp->qpn, 1636329222Shselasky qp->sq.head, qp->sq.tail, 1637329222Shselasky qp->sq.max, nreq); 1638329222Shselasky err = -ENOMEM; 1639329222Shselasky *bad_wr = wr; 1640329222Shselasky goto out; 1641329222Shselasky } 1642329222Shselasky 1643329222Shselasky wqe = get_send_wqe(qp, ind); 1644329222Shselasky prev_wqe = qp->sq.last; 1645329222Shselasky qp->sq.last = wqe; 1646329222Shselasky 1647329222Shselasky ((struct mthca_next_seg *) wqe)->nda_op = 0; 1648329222Shselasky ((struct mthca_next_seg *) wqe)->ee_nds = 0; 1649329222Shselasky ((struct mthca_next_seg *) wqe)->flags = 1650329222Shselasky ((wr->send_flags & IB_SEND_SIGNALED) ? 1651329222Shselasky cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) | 1652329222Shselasky ((wr->send_flags & IB_SEND_SOLICITED) ? 1653329222Shselasky cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) | 1654329222Shselasky cpu_to_be32(1); 1655329222Shselasky if (wr->opcode == IB_WR_SEND_WITH_IMM || 1656329222Shselasky wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) 1657329222Shselasky ((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data; 1658329222Shselasky 1659329222Shselasky wqe += sizeof (struct mthca_next_seg); 1660329222Shselasky size = sizeof (struct mthca_next_seg) / 16; 1661329222Shselasky 1662329222Shselasky switch (qp->transport) { 1663329222Shselasky case RC: 1664329222Shselasky switch (wr->opcode) { 1665329222Shselasky case IB_WR_ATOMIC_CMP_AND_SWP: 1666329222Shselasky case IB_WR_ATOMIC_FETCH_AND_ADD: 1667329222Shselasky set_raddr_seg(wqe, atomic_wr(wr)->remote_addr, 1668329222Shselasky atomic_wr(wr)->rkey); 1669329222Shselasky wqe += sizeof (struct mthca_raddr_seg); 1670329222Shselasky 1671329222Shselasky set_atomic_seg(wqe, atomic_wr(wr)); 1672329222Shselasky wqe += sizeof (struct mthca_atomic_seg); 1673329222Shselasky size += (sizeof (struct mthca_raddr_seg) + 1674329222Shselasky sizeof (struct mthca_atomic_seg)) / 16; 1675329222Shselasky break; 1676329222Shselasky 1677329222Shselasky case IB_WR_RDMA_WRITE: 1678329222Shselasky case IB_WR_RDMA_WRITE_WITH_IMM: 1679329222Shselasky case IB_WR_RDMA_READ: 1680329222Shselasky set_raddr_seg(wqe, rdma_wr(wr)->remote_addr, 1681329222Shselasky rdma_wr(wr)->rkey); 1682329222Shselasky wqe += sizeof (struct mthca_raddr_seg); 1683329222Shselasky size += sizeof (struct mthca_raddr_seg) / 16; 1684329222Shselasky break; 1685329222Shselasky 1686329222Shselasky default: 1687329222Shselasky /* No extra segments required for sends */ 1688329222Shselasky break; 1689329222Shselasky } 1690329222Shselasky 1691329222Shselasky break; 1692329222Shselasky 1693329222Shselasky case UC: 1694329222Shselasky switch (wr->opcode) { 1695329222Shselasky case IB_WR_RDMA_WRITE: 1696329222Shselasky case IB_WR_RDMA_WRITE_WITH_IMM: 1697329222Shselasky set_raddr_seg(wqe, rdma_wr(wr)->remote_addr, 1698329222Shselasky rdma_wr(wr)->rkey); 1699329222Shselasky wqe += sizeof (struct mthca_raddr_seg); 1700329222Shselasky size += sizeof (struct mthca_raddr_seg) / 16; 1701329222Shselasky break; 1702329222Shselasky 1703329222Shselasky default: 1704329222Shselasky /* No extra segments required for sends */ 1705329222Shselasky break; 1706329222Shselasky } 1707329222Shselasky 1708329222Shselasky break; 1709329222Shselasky 1710329222Shselasky case UD: 1711329222Shselasky set_tavor_ud_seg(wqe, ud_wr(wr)); 1712329222Shselasky wqe += sizeof (struct mthca_tavor_ud_seg); 1713329222Shselasky size += sizeof (struct mthca_tavor_ud_seg) / 16; 1714329222Shselasky break; 1715329222Shselasky 1716329222Shselasky case MLX: 1717329222Shselasky err = build_mlx_header(dev, to_msqp(qp), ind, ud_wr(wr), 1718329222Shselasky wqe - sizeof (struct mthca_next_seg), 1719329222Shselasky wqe); 1720329222Shselasky if (err) { 1721329222Shselasky *bad_wr = wr; 1722329222Shselasky goto out; 1723329222Shselasky } 1724329222Shselasky wqe += sizeof (struct mthca_data_seg); 1725329222Shselasky size += sizeof (struct mthca_data_seg) / 16; 1726329222Shselasky break; 1727329222Shselasky } 1728329222Shselasky 1729329222Shselasky if (wr->num_sge > qp->sq.max_gs) { 1730329222Shselasky mthca_err(dev, "too many gathers\n"); 1731329222Shselasky err = -EINVAL; 1732329222Shselasky *bad_wr = wr; 1733329222Shselasky goto out; 1734329222Shselasky } 1735329222Shselasky 1736329222Shselasky for (i = 0; i < wr->num_sge; ++i) { 1737329222Shselasky mthca_set_data_seg(wqe, wr->sg_list + i); 1738329222Shselasky wqe += sizeof (struct mthca_data_seg); 1739329222Shselasky size += sizeof (struct mthca_data_seg) / 16; 1740329222Shselasky } 1741329222Shselasky 1742329222Shselasky /* Add one more inline data segment for ICRC */ 1743329222Shselasky if (qp->transport == MLX) { 1744329222Shselasky ((struct mthca_data_seg *) wqe)->byte_count = 1745329222Shselasky cpu_to_be32((1 << 31) | 4); 1746329222Shselasky ((u32 *) wqe)[1] = 0; 1747329222Shselasky wqe += sizeof (struct mthca_data_seg); 1748329222Shselasky size += sizeof (struct mthca_data_seg) / 16; 1749329222Shselasky } 1750329222Shselasky 1751329222Shselasky qp->wrid[ind + qp->rq.max] = wr->wr_id; 1752329222Shselasky 1753329222Shselasky if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) { 1754329222Shselasky mthca_err(dev, "opcode invalid\n"); 1755329222Shselasky err = -EINVAL; 1756329222Shselasky *bad_wr = wr; 1757329222Shselasky goto out; 1758329222Shselasky } 1759329222Shselasky 1760329222Shselasky ((struct mthca_next_seg *) prev_wqe)->nda_op = 1761329222Shselasky cpu_to_be32(((ind << qp->sq.wqe_shift) + 1762329222Shselasky qp->send_wqe_offset) | 1763329222Shselasky mthca_opcode[wr->opcode]); 1764329222Shselasky wmb(); 1765329222Shselasky ((struct mthca_next_seg *) prev_wqe)->ee_nds = 1766329222Shselasky cpu_to_be32((nreq ? 0 : MTHCA_NEXT_DBD) | size | 1767329222Shselasky ((wr->send_flags & IB_SEND_FENCE) ? 1768329222Shselasky MTHCA_NEXT_FENCE : 0)); 1769329222Shselasky 1770329222Shselasky if (!nreq) { 1771329222Shselasky size0 = size; 1772329222Shselasky op0 = mthca_opcode[wr->opcode]; 1773329222Shselasky f0 = wr->send_flags & IB_SEND_FENCE ? 1774329222Shselasky MTHCA_SEND_DOORBELL_FENCE : 0; 1775329222Shselasky } 1776329222Shselasky 1777329222Shselasky ++ind; 1778329222Shselasky if (unlikely(ind >= qp->sq.max)) 1779329222Shselasky ind -= qp->sq.max; 1780329222Shselasky } 1781329222Shselasky 1782329222Shselaskyout: 1783329222Shselasky if (likely(nreq)) { 1784329222Shselasky wmb(); 1785329222Shselasky 1786329222Shselasky mthca_write64(((qp->sq.next_ind << qp->sq.wqe_shift) + 1787329222Shselasky qp->send_wqe_offset) | f0 | op0, 1788329222Shselasky (qp->qpn << 8) | size0, 1789329222Shselasky dev->kar + MTHCA_SEND_DOORBELL, 1790329222Shselasky MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock)); 1791329222Shselasky /* 1792329222Shselasky * Make sure doorbells don't leak out of SQ spinlock 1793329222Shselasky * and reach the HCA out of order: 1794329222Shselasky */ 1795329222Shselasky mmiowb(); 1796329222Shselasky } 1797329222Shselasky 1798329222Shselasky qp->sq.next_ind = ind; 1799329222Shselasky qp->sq.head += nreq; 1800329222Shselasky 1801329222Shselasky spin_unlock_irqrestore(&qp->sq.lock, flags); 1802329222Shselasky return err; 1803329222Shselasky} 1804329222Shselasky 1805329222Shselaskyint mthca_tavor_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr, 1806329222Shselasky struct ib_recv_wr **bad_wr) 1807329222Shselasky{ 1808329222Shselasky struct mthca_dev *dev = to_mdev(ibqp->device); 1809329222Shselasky struct mthca_qp *qp = to_mqp(ibqp); 1810329222Shselasky unsigned long flags; 1811329222Shselasky int err = 0; 1812329222Shselasky int nreq; 1813329222Shselasky int i; 1814329222Shselasky int size; 1815329222Shselasky /* 1816329222Shselasky * size0 is only used if nreq != 0, and it will always be 1817329222Shselasky * initialized the first time through the main loop before 1818329222Shselasky * nreq is incremented. So nreq cannot become non-zero 1819329222Shselasky * without initializing size0, and it is in fact never used 1820329222Shselasky * uninitialized. 1821329222Shselasky */ 1822329222Shselasky int uninitialized_var(size0); 1823329222Shselasky int ind; 1824329222Shselasky void *wqe; 1825329222Shselasky void *prev_wqe; 1826329222Shselasky 1827329222Shselasky spin_lock_irqsave(&qp->rq.lock, flags); 1828329222Shselasky 1829329222Shselasky /* XXX check that state is OK to post receive */ 1830329222Shselasky 1831329222Shselasky ind = qp->rq.next_ind; 1832329222Shselasky 1833329222Shselasky for (nreq = 0; wr; wr = wr->next) { 1834329222Shselasky if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) { 1835329222Shselasky mthca_err(dev, "RQ %06x full (%u head, %u tail," 1836329222Shselasky " %d max, %d nreq)\n", qp->qpn, 1837329222Shselasky qp->rq.head, qp->rq.tail, 1838329222Shselasky qp->rq.max, nreq); 1839329222Shselasky err = -ENOMEM; 1840329222Shselasky *bad_wr = wr; 1841329222Shselasky goto out; 1842329222Shselasky } 1843329222Shselasky 1844329222Shselasky wqe = get_recv_wqe(qp, ind); 1845329222Shselasky prev_wqe = qp->rq.last; 1846329222Shselasky qp->rq.last = wqe; 1847329222Shselasky 1848329222Shselasky ((struct mthca_next_seg *) wqe)->ee_nds = 1849329222Shselasky cpu_to_be32(MTHCA_NEXT_DBD); 1850329222Shselasky ((struct mthca_next_seg *) wqe)->flags = 0; 1851329222Shselasky 1852329222Shselasky wqe += sizeof (struct mthca_next_seg); 1853329222Shselasky size = sizeof (struct mthca_next_seg) / 16; 1854329222Shselasky 1855329222Shselasky if (unlikely(wr->num_sge > qp->rq.max_gs)) { 1856329222Shselasky err = -EINVAL; 1857329222Shselasky *bad_wr = wr; 1858329222Shselasky goto out; 1859329222Shselasky } 1860329222Shselasky 1861329222Shselasky for (i = 0; i < wr->num_sge; ++i) { 1862329222Shselasky mthca_set_data_seg(wqe, wr->sg_list + i); 1863329222Shselasky wqe += sizeof (struct mthca_data_seg); 1864329222Shselasky size += sizeof (struct mthca_data_seg) / 16; 1865329222Shselasky } 1866329222Shselasky 1867329222Shselasky qp->wrid[ind] = wr->wr_id; 1868329222Shselasky 1869329222Shselasky ((struct mthca_next_seg *) prev_wqe)->ee_nds = 1870329222Shselasky cpu_to_be32(MTHCA_NEXT_DBD | size); 1871329222Shselasky 1872329222Shselasky if (!nreq) 1873329222Shselasky size0 = size; 1874329222Shselasky 1875329222Shselasky ++ind; 1876329222Shselasky if (unlikely(ind >= qp->rq.max)) 1877329222Shselasky ind -= qp->rq.max; 1878329222Shselasky 1879329222Shselasky ++nreq; 1880329222Shselasky if (unlikely(nreq == MTHCA_TAVOR_MAX_WQES_PER_RECV_DB)) { 1881329222Shselasky nreq = 0; 1882329222Shselasky 1883329222Shselasky wmb(); 1884329222Shselasky 1885329222Shselasky mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0, 1886329222Shselasky qp->qpn << 8, dev->kar + MTHCA_RECEIVE_DOORBELL, 1887329222Shselasky MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock)); 1888329222Shselasky 1889329222Shselasky qp->rq.next_ind = ind; 1890329222Shselasky qp->rq.head += MTHCA_TAVOR_MAX_WQES_PER_RECV_DB; 1891329222Shselasky } 1892329222Shselasky } 1893329222Shselasky 1894329222Shselaskyout: 1895329222Shselasky if (likely(nreq)) { 1896329222Shselasky wmb(); 1897329222Shselasky 1898329222Shselasky mthca_write64((qp->rq.next_ind << qp->rq.wqe_shift) | size0, 1899329222Shselasky qp->qpn << 8 | nreq, dev->kar + MTHCA_RECEIVE_DOORBELL, 1900329222Shselasky MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock)); 1901329222Shselasky } 1902329222Shselasky 1903329222Shselasky qp->rq.next_ind = ind; 1904329222Shselasky qp->rq.head += nreq; 1905329222Shselasky 1906329222Shselasky /* 1907329222Shselasky * Make sure doorbells don't leak out of RQ spinlock and reach 1908329222Shselasky * the HCA out of order: 1909329222Shselasky */ 1910329222Shselasky mmiowb(); 1911329222Shselasky 1912329222Shselasky spin_unlock_irqrestore(&qp->rq.lock, flags); 1913329222Shselasky return err; 1914329222Shselasky} 1915329222Shselasky 1916329222Shselaskyint mthca_arbel_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 1917329222Shselasky struct ib_send_wr **bad_wr) 1918329222Shselasky{ 1919329222Shselasky struct mthca_dev *dev = to_mdev(ibqp->device); 1920329222Shselasky struct mthca_qp *qp = to_mqp(ibqp); 1921329222Shselasky u32 dbhi; 1922329222Shselasky void *wqe; 1923329222Shselasky void *prev_wqe; 1924329222Shselasky unsigned long flags; 1925329222Shselasky int err = 0; 1926329222Shselasky int nreq; 1927329222Shselasky int i; 1928329222Shselasky int size; 1929329222Shselasky /* 1930329222Shselasky * f0 and size0 are only used if nreq != 0, and they will 1931329222Shselasky * always be initialized the first time through the main loop 1932329222Shselasky * before nreq is incremented. So nreq cannot become non-zero 1933329222Shselasky * without initializing f0 and size0, and they are in fact 1934329222Shselasky * never used uninitialized. 1935329222Shselasky */ 1936329222Shselasky int uninitialized_var(size0); 1937329222Shselasky u32 uninitialized_var(f0); 1938329222Shselasky int ind; 1939329222Shselasky u8 op0 = 0; 1940329222Shselasky 1941329222Shselasky spin_lock_irqsave(&qp->sq.lock, flags); 1942329222Shselasky 1943329222Shselasky /* XXX check that state is OK to post send */ 1944329222Shselasky 1945329222Shselasky ind = qp->sq.head & (qp->sq.max - 1); 1946329222Shselasky 1947329222Shselasky for (nreq = 0; wr; ++nreq, wr = wr->next) { 1948329222Shselasky if (unlikely(nreq == MTHCA_ARBEL_MAX_WQES_PER_SEND_DB)) { 1949329222Shselasky nreq = 0; 1950329222Shselasky 1951329222Shselasky dbhi = (MTHCA_ARBEL_MAX_WQES_PER_SEND_DB << 24) | 1952329222Shselasky ((qp->sq.head & 0xffff) << 8) | f0 | op0; 1953329222Shselasky 1954329222Shselasky qp->sq.head += MTHCA_ARBEL_MAX_WQES_PER_SEND_DB; 1955329222Shselasky 1956329222Shselasky /* 1957329222Shselasky * Make sure that descriptors are written before 1958329222Shselasky * doorbell record. 1959329222Shselasky */ 1960329222Shselasky wmb(); 1961329222Shselasky *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff); 1962329222Shselasky 1963329222Shselasky /* 1964329222Shselasky * Make sure doorbell record is written before we 1965329222Shselasky * write MMIO send doorbell. 1966329222Shselasky */ 1967329222Shselasky wmb(); 1968329222Shselasky 1969329222Shselasky mthca_write64(dbhi, (qp->qpn << 8) | size0, 1970329222Shselasky dev->kar + MTHCA_SEND_DOORBELL, 1971329222Shselasky MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock)); 1972329222Shselasky } 1973329222Shselasky 1974329222Shselasky if (mthca_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) { 1975329222Shselasky mthca_err(dev, "SQ %06x full (%u head, %u tail," 1976329222Shselasky " %d max, %d nreq)\n", qp->qpn, 1977329222Shselasky qp->sq.head, qp->sq.tail, 1978329222Shselasky qp->sq.max, nreq); 1979329222Shselasky err = -ENOMEM; 1980329222Shselasky *bad_wr = wr; 1981329222Shselasky goto out; 1982329222Shselasky } 1983329222Shselasky 1984329222Shselasky wqe = get_send_wqe(qp, ind); 1985329222Shselasky prev_wqe = qp->sq.last; 1986329222Shselasky qp->sq.last = wqe; 1987329222Shselasky 1988329222Shselasky ((struct mthca_next_seg *) wqe)->flags = 1989329222Shselasky ((wr->send_flags & IB_SEND_SIGNALED) ? 1990329222Shselasky cpu_to_be32(MTHCA_NEXT_CQ_UPDATE) : 0) | 1991329222Shselasky ((wr->send_flags & IB_SEND_SOLICITED) ? 1992329222Shselasky cpu_to_be32(MTHCA_NEXT_SOLICIT) : 0) | 1993329222Shselasky ((wr->send_flags & IB_SEND_IP_CSUM) ? 1994329222Shselasky cpu_to_be32(MTHCA_NEXT_IP_CSUM | MTHCA_NEXT_TCP_UDP_CSUM) : 0) | 1995329222Shselasky cpu_to_be32(1); 1996329222Shselasky if (wr->opcode == IB_WR_SEND_WITH_IMM || 1997329222Shselasky wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) 1998329222Shselasky ((struct mthca_next_seg *) wqe)->imm = wr->ex.imm_data; 1999329222Shselasky 2000329222Shselasky wqe += sizeof (struct mthca_next_seg); 2001329222Shselasky size = sizeof (struct mthca_next_seg) / 16; 2002329222Shselasky 2003329222Shselasky switch (qp->transport) { 2004329222Shselasky case RC: 2005329222Shselasky switch (wr->opcode) { 2006329222Shselasky case IB_WR_ATOMIC_CMP_AND_SWP: 2007329222Shselasky case IB_WR_ATOMIC_FETCH_AND_ADD: 2008329222Shselasky set_raddr_seg(wqe, atomic_wr(wr)->remote_addr, 2009329222Shselasky atomic_wr(wr)->rkey); 2010329222Shselasky wqe += sizeof (struct mthca_raddr_seg); 2011329222Shselasky 2012329222Shselasky set_atomic_seg(wqe, atomic_wr(wr)); 2013329222Shselasky wqe += sizeof (struct mthca_atomic_seg); 2014329222Shselasky size += (sizeof (struct mthca_raddr_seg) + 2015329222Shselasky sizeof (struct mthca_atomic_seg)) / 16; 2016329222Shselasky break; 2017329222Shselasky 2018329222Shselasky case IB_WR_RDMA_READ: 2019329222Shselasky case IB_WR_RDMA_WRITE: 2020329222Shselasky case IB_WR_RDMA_WRITE_WITH_IMM: 2021329222Shselasky set_raddr_seg(wqe, rdma_wr(wr)->remote_addr, 2022329222Shselasky rdma_wr(wr)->rkey); 2023329222Shselasky wqe += sizeof (struct mthca_raddr_seg); 2024329222Shselasky size += sizeof (struct mthca_raddr_seg) / 16; 2025329222Shselasky break; 2026329222Shselasky 2027329222Shselasky default: 2028329222Shselasky /* No extra segments required for sends */ 2029329222Shselasky break; 2030329222Shselasky } 2031329222Shselasky 2032329222Shselasky break; 2033329222Shselasky 2034329222Shselasky case UC: 2035329222Shselasky switch (wr->opcode) { 2036329222Shselasky case IB_WR_RDMA_WRITE: 2037329222Shselasky case IB_WR_RDMA_WRITE_WITH_IMM: 2038329222Shselasky set_raddr_seg(wqe, rdma_wr(wr)->remote_addr, 2039329222Shselasky rdma_wr(wr)->rkey); 2040329222Shselasky wqe += sizeof (struct mthca_raddr_seg); 2041329222Shselasky size += sizeof (struct mthca_raddr_seg) / 16; 2042329222Shselasky break; 2043329222Shselasky 2044329222Shselasky default: 2045329222Shselasky /* No extra segments required for sends */ 2046329222Shselasky break; 2047329222Shselasky } 2048329222Shselasky 2049329222Shselasky break; 2050329222Shselasky 2051329222Shselasky case UD: 2052329222Shselasky set_arbel_ud_seg(wqe, ud_wr(wr)); 2053329222Shselasky wqe += sizeof (struct mthca_arbel_ud_seg); 2054329222Shselasky size += sizeof (struct mthca_arbel_ud_seg) / 16; 2055329222Shselasky break; 2056329222Shselasky 2057329222Shselasky case MLX: 2058329222Shselasky err = build_mlx_header(dev, to_msqp(qp), ind, ud_wr(wr), 2059329222Shselasky wqe - sizeof (struct mthca_next_seg), 2060329222Shselasky wqe); 2061329222Shselasky if (err) { 2062329222Shselasky *bad_wr = wr; 2063329222Shselasky goto out; 2064329222Shselasky } 2065329222Shselasky wqe += sizeof (struct mthca_data_seg); 2066329222Shselasky size += sizeof (struct mthca_data_seg) / 16; 2067329222Shselasky break; 2068329222Shselasky } 2069329222Shselasky 2070329222Shselasky if (wr->num_sge > qp->sq.max_gs) { 2071329222Shselasky mthca_err(dev, "too many gathers\n"); 2072329222Shselasky err = -EINVAL; 2073329222Shselasky *bad_wr = wr; 2074329222Shselasky goto out; 2075329222Shselasky } 2076329222Shselasky 2077329222Shselasky for (i = 0; i < wr->num_sge; ++i) { 2078329222Shselasky mthca_set_data_seg(wqe, wr->sg_list + i); 2079329222Shselasky wqe += sizeof (struct mthca_data_seg); 2080329222Shselasky size += sizeof (struct mthca_data_seg) / 16; 2081329222Shselasky } 2082329222Shselasky 2083329222Shselasky /* Add one more inline data segment for ICRC */ 2084329222Shselasky if (qp->transport == MLX) { 2085329222Shselasky ((struct mthca_data_seg *) wqe)->byte_count = 2086329222Shselasky cpu_to_be32((1 << 31) | 4); 2087329222Shselasky ((u32 *) wqe)[1] = 0; 2088329222Shselasky wqe += sizeof (struct mthca_data_seg); 2089329222Shselasky size += sizeof (struct mthca_data_seg) / 16; 2090329222Shselasky } 2091329222Shselasky 2092329222Shselasky qp->wrid[ind + qp->rq.max] = wr->wr_id; 2093329222Shselasky 2094329222Shselasky if (wr->opcode >= ARRAY_SIZE(mthca_opcode)) { 2095329222Shselasky mthca_err(dev, "opcode invalid\n"); 2096329222Shselasky err = -EINVAL; 2097329222Shselasky *bad_wr = wr; 2098329222Shselasky goto out; 2099329222Shselasky } 2100329222Shselasky 2101329222Shselasky ((struct mthca_next_seg *) prev_wqe)->nda_op = 2102329222Shselasky cpu_to_be32(((ind << qp->sq.wqe_shift) + 2103329222Shselasky qp->send_wqe_offset) | 2104329222Shselasky mthca_opcode[wr->opcode]); 2105329222Shselasky wmb(); 2106329222Shselasky ((struct mthca_next_seg *) prev_wqe)->ee_nds = 2107329222Shselasky cpu_to_be32(MTHCA_NEXT_DBD | size | 2108329222Shselasky ((wr->send_flags & IB_SEND_FENCE) ? 2109329222Shselasky MTHCA_NEXT_FENCE : 0)); 2110329222Shselasky 2111329222Shselasky if (!nreq) { 2112329222Shselasky size0 = size; 2113329222Shselasky op0 = mthca_opcode[wr->opcode]; 2114329222Shselasky f0 = wr->send_flags & IB_SEND_FENCE ? 2115329222Shselasky MTHCA_SEND_DOORBELL_FENCE : 0; 2116329222Shselasky } 2117329222Shselasky 2118329222Shselasky ++ind; 2119329222Shselasky if (unlikely(ind >= qp->sq.max)) 2120329222Shselasky ind -= qp->sq.max; 2121329222Shselasky } 2122329222Shselasky 2123329222Shselaskyout: 2124329222Shselasky if (likely(nreq)) { 2125329222Shselasky dbhi = (nreq << 24) | ((qp->sq.head & 0xffff) << 8) | f0 | op0; 2126329222Shselasky 2127329222Shselasky qp->sq.head += nreq; 2128329222Shselasky 2129329222Shselasky /* 2130329222Shselasky * Make sure that descriptors are written before 2131329222Shselasky * doorbell record. 2132329222Shselasky */ 2133329222Shselasky wmb(); 2134329222Shselasky *qp->sq.db = cpu_to_be32(qp->sq.head & 0xffff); 2135329222Shselasky 2136329222Shselasky /* 2137329222Shselasky * Make sure doorbell record is written before we 2138329222Shselasky * write MMIO send doorbell. 2139329222Shselasky */ 2140329222Shselasky wmb(); 2141329222Shselasky 2142329222Shselasky mthca_write64(dbhi, (qp->qpn << 8) | size0, dev->kar + MTHCA_SEND_DOORBELL, 2143329222Shselasky MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock)); 2144329222Shselasky } 2145329222Shselasky 2146329222Shselasky /* 2147329222Shselasky * Make sure doorbells don't leak out of SQ spinlock and reach 2148329222Shselasky * the HCA out of order: 2149329222Shselasky */ 2150329222Shselasky mmiowb(); 2151329222Shselasky 2152329222Shselasky spin_unlock_irqrestore(&qp->sq.lock, flags); 2153329222Shselasky return err; 2154329222Shselasky} 2155329222Shselasky 2156329222Shselaskyint mthca_arbel_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr, 2157329222Shselasky struct ib_recv_wr **bad_wr) 2158329222Shselasky{ 2159329222Shselasky struct mthca_dev *dev = to_mdev(ibqp->device); 2160329222Shselasky struct mthca_qp *qp = to_mqp(ibqp); 2161329222Shselasky unsigned long flags; 2162329222Shselasky int err = 0; 2163329222Shselasky int nreq; 2164329222Shselasky int ind; 2165329222Shselasky int i; 2166329222Shselasky void *wqe; 2167329222Shselasky 2168329222Shselasky spin_lock_irqsave(&qp->rq.lock, flags); 2169329222Shselasky 2170329222Shselasky /* XXX check that state is OK to post receive */ 2171329222Shselasky 2172329222Shselasky ind = qp->rq.head & (qp->rq.max - 1); 2173329222Shselasky 2174329222Shselasky for (nreq = 0; wr; ++nreq, wr = wr->next) { 2175329222Shselasky if (mthca_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) { 2176329222Shselasky mthca_err(dev, "RQ %06x full (%u head, %u tail," 2177329222Shselasky " %d max, %d nreq)\n", qp->qpn, 2178329222Shselasky qp->rq.head, qp->rq.tail, 2179329222Shselasky qp->rq.max, nreq); 2180329222Shselasky err = -ENOMEM; 2181329222Shselasky *bad_wr = wr; 2182329222Shselasky goto out; 2183329222Shselasky } 2184329222Shselasky 2185329222Shselasky wqe = get_recv_wqe(qp, ind); 2186329222Shselasky 2187329222Shselasky ((struct mthca_next_seg *) wqe)->flags = 0; 2188329222Shselasky 2189329222Shselasky wqe += sizeof (struct mthca_next_seg); 2190329222Shselasky 2191329222Shselasky if (unlikely(wr->num_sge > qp->rq.max_gs)) { 2192329222Shselasky err = -EINVAL; 2193329222Shselasky *bad_wr = wr; 2194329222Shselasky goto out; 2195329222Shselasky } 2196329222Shselasky 2197329222Shselasky for (i = 0; i < wr->num_sge; ++i) { 2198329222Shselasky mthca_set_data_seg(wqe, wr->sg_list + i); 2199329222Shselasky wqe += sizeof (struct mthca_data_seg); 2200329222Shselasky } 2201329222Shselasky 2202329222Shselasky if (i < qp->rq.max_gs) 2203329222Shselasky mthca_set_data_seg_inval(wqe); 2204329222Shselasky 2205329222Shselasky qp->wrid[ind] = wr->wr_id; 2206329222Shselasky 2207329222Shselasky ++ind; 2208329222Shselasky if (unlikely(ind >= qp->rq.max)) 2209329222Shselasky ind -= qp->rq.max; 2210329222Shselasky } 2211329222Shselaskyout: 2212329222Shselasky if (likely(nreq)) { 2213329222Shselasky qp->rq.head += nreq; 2214329222Shselasky 2215329222Shselasky /* 2216329222Shselasky * Make sure that descriptors are written before 2217329222Shselasky * doorbell record. 2218329222Shselasky */ 2219329222Shselasky wmb(); 2220329222Shselasky *qp->rq.db = cpu_to_be32(qp->rq.head & 0xffff); 2221329222Shselasky } 2222329222Shselasky 2223329222Shselasky spin_unlock_irqrestore(&qp->rq.lock, flags); 2224329222Shselasky return err; 2225329222Shselasky} 2226329222Shselasky 2227329222Shselaskyvoid mthca_free_err_wqe(struct mthca_dev *dev, struct mthca_qp *qp, int is_send, 2228329222Shselasky int index, int *dbd, __be32 *new_wqe) 2229329222Shselasky{ 2230329222Shselasky struct mthca_next_seg *next; 2231329222Shselasky 2232329222Shselasky /* 2233329222Shselasky * For SRQs, all receive WQEs generate a CQE, so we're always 2234329222Shselasky * at the end of the doorbell chain. 2235329222Shselasky */ 2236329222Shselasky if (qp->ibqp.srq && !is_send) { 2237329222Shselasky *new_wqe = 0; 2238329222Shselasky return; 2239329222Shselasky } 2240329222Shselasky 2241329222Shselasky if (is_send) 2242329222Shselasky next = get_send_wqe(qp, index); 2243329222Shselasky else 2244329222Shselasky next = get_recv_wqe(qp, index); 2245329222Shselasky 2246329222Shselasky *dbd = !!(next->ee_nds & cpu_to_be32(MTHCA_NEXT_DBD)); 2247329222Shselasky if (next->ee_nds & cpu_to_be32(0x3f)) 2248329222Shselasky *new_wqe = (next->nda_op & cpu_to_be32(~0x3f)) | 2249329222Shselasky (next->ee_nds & cpu_to_be32(0x3f)); 2250329222Shselasky else 2251329222Shselasky *new_wqe = 0; 2252329222Shselasky} 2253329222Shselasky 2254329222Shselaskyint mthca_init_qp_table(struct mthca_dev *dev) 2255329222Shselasky{ 2256329222Shselasky int err; 2257329222Shselasky int i; 2258329222Shselasky 2259329222Shselasky spin_lock_init(&dev->qp_table.lock); 2260329222Shselasky 2261329222Shselasky /* 2262329222Shselasky * We reserve 2 extra QPs per port for the special QPs. The 2263329222Shselasky * special QP for port 1 has to be even, so round up. 2264329222Shselasky */ 2265329222Shselasky dev->qp_table.sqp_start = (dev->limits.reserved_qps + 1) & ~1UL; 2266329222Shselasky err = mthca_alloc_init(&dev->qp_table.alloc, 2267329222Shselasky dev->limits.num_qps, 2268329222Shselasky (1 << 24) - 1, 2269329222Shselasky dev->qp_table.sqp_start + 2270329222Shselasky MTHCA_MAX_PORTS * 2); 2271329222Shselasky if (err) 2272329222Shselasky return err; 2273329222Shselasky 2274329222Shselasky err = mthca_array_init(&dev->qp_table.qp, 2275329222Shselasky dev->limits.num_qps); 2276329222Shselasky if (err) { 2277329222Shselasky mthca_alloc_cleanup(&dev->qp_table.alloc); 2278329222Shselasky return err; 2279329222Shselasky } 2280329222Shselasky 2281329222Shselasky for (i = 0; i < 2; ++i) { 2282329222Shselasky err = mthca_CONF_SPECIAL_QP(dev, i ? IB_QPT_GSI : IB_QPT_SMI, 2283329222Shselasky dev->qp_table.sqp_start + i * 2); 2284329222Shselasky if (err) { 2285329222Shselasky mthca_warn(dev, "CONF_SPECIAL_QP returned " 2286329222Shselasky "%d, aborting.\n", err); 2287329222Shselasky goto err_out; 2288329222Shselasky } 2289329222Shselasky } 2290329222Shselasky return 0; 2291329222Shselasky 2292329222Shselasky err_out: 2293329222Shselasky for (i = 0; i < 2; ++i) 2294329222Shselasky mthca_CONF_SPECIAL_QP(dev, i, 0); 2295329222Shselasky 2296329222Shselasky mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps); 2297329222Shselasky mthca_alloc_cleanup(&dev->qp_table.alloc); 2298329222Shselasky 2299329222Shselasky return err; 2300329222Shselasky} 2301329222Shselasky 2302329222Shselaskyvoid mthca_cleanup_qp_table(struct mthca_dev *dev) 2303329222Shselasky{ 2304329222Shselasky int i; 2305329222Shselasky 2306329222Shselasky for (i = 0; i < 2; ++i) 2307329222Shselasky mthca_CONF_SPECIAL_QP(dev, i, 0); 2308329222Shselasky 2309329222Shselasky mthca_array_cleanup(&dev->qp_table.qp, dev->limits.num_qps); 2310329222Shselasky mthca_alloc_cleanup(&dev->qp_table.alloc); 2311329222Shselasky} 2312