| Current Path : /usr/src/contrib/ofed/libmthca/src/ |
FreeBSD hs32.drive.ne.jp 9.1-RELEASE FreeBSD 9.1-RELEASE #1: Wed Jan 14 12:18:08 JST 2015 root@hs32.drive.ne.jp:/sys/amd64/compile/hs32 amd64 |
| Current File : //usr/src/contrib/ofed/libmthca/src/cq.c |
/*
* Copyright (c) 2005 Topspin Communications. All rights reserved.
* Copyright (c) 2005 Mellanox Technologies Ltd. All rights reserved.
* Copyright (c) 2006 Cisco Systems. All rights reserved.
*
* This software is available to you under a choice of one of two
* licenses. You may choose to be licensed under the terms of the GNU
* General Public License (GPL) Version 2, available from the file
* COPYING in the main directory of this source tree, or the
* OpenIB.org BSD license below:
*
* Redistribution and use in source and binary forms, with or
* without modification, are permitted provided that the following
* conditions are met:
*
* - Redistributions of source code must retain the above
* copyright notice, this list of conditions and the following
* disclaimer.
*
* - Redistributions in binary form must reproduce the above
* copyright notice, this list of conditions and the following
* disclaimer in the documentation and/or other materials
* provided with the distribution.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
* BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
* ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
* CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#if HAVE_CONFIG_H
# include <config.h>
#endif /* HAVE_CONFIG_H */
#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>
#include <netinet/in.h>
#include <string.h>
#include <infiniband/opcode.h>
#include "mthca.h"
#include "doorbell.h"
enum {
MTHCA_CQ_DOORBELL = 0x20
};
enum {
CQ_OK = 0,
CQ_EMPTY = -1,
CQ_POLL_ERR = -2
};
#define MTHCA_TAVOR_CQ_DB_INC_CI (1 << 24)
#define MTHCA_TAVOR_CQ_DB_REQ_NOT (2 << 24)
#define MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL (3 << 24)
#define MTHCA_TAVOR_CQ_DB_SET_CI (4 << 24)
#define MTHCA_TAVOR_CQ_DB_REQ_NOT_MULT (5 << 24)
#define MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL (1 << 24)
#define MTHCA_ARBEL_CQ_DB_REQ_NOT (2 << 24)
#define MTHCA_ARBEL_CQ_DB_REQ_NOT_MULT (3 << 24)
enum {
MTHCA_CQ_ENTRY_OWNER_SW = 0x00,
MTHCA_CQ_ENTRY_OWNER_HW = 0x80,
MTHCA_ERROR_CQE_OPCODE_MASK = 0xfe
};
enum {
SYNDROME_LOCAL_LENGTH_ERR = 0x01,
SYNDROME_LOCAL_QP_OP_ERR = 0x02,
SYNDROME_LOCAL_EEC_OP_ERR = 0x03,
SYNDROME_LOCAL_PROT_ERR = 0x04,
SYNDROME_WR_FLUSH_ERR = 0x05,
SYNDROME_MW_BIND_ERR = 0x06,
SYNDROME_BAD_RESP_ERR = 0x10,
SYNDROME_LOCAL_ACCESS_ERR = 0x11,
SYNDROME_REMOTE_INVAL_REQ_ERR = 0x12,
SYNDROME_REMOTE_ACCESS_ERR = 0x13,
SYNDROME_REMOTE_OP_ERR = 0x14,
SYNDROME_RETRY_EXC_ERR = 0x15,
SYNDROME_RNR_RETRY_EXC_ERR = 0x16,
SYNDROME_LOCAL_RDD_VIOL_ERR = 0x20,
SYNDROME_REMOTE_INVAL_RD_REQ_ERR = 0x21,
SYNDROME_REMOTE_ABORTED_ERR = 0x22,
SYNDROME_INVAL_EECN_ERR = 0x23,
SYNDROME_INVAL_EEC_STATE_ERR = 0x24
};
struct mthca_cqe {
uint32_t my_qpn;
uint32_t my_ee;
uint32_t rqpn;
uint16_t sl_g_mlpath;
uint16_t rlid;
uint32_t imm_etype_pkey_eec;
uint32_t byte_cnt;
uint32_t wqe;
uint8_t opcode;
uint8_t is_send;
uint8_t reserved;
uint8_t owner;
};
struct mthca_err_cqe {
uint32_t my_qpn;
uint32_t reserved1[3];
uint8_t syndrome;
uint8_t vendor_err;
uint16_t db_cnt;
uint32_t reserved2;
uint32_t wqe;
uint8_t opcode;
uint8_t reserved3[2];
uint8_t owner;
};
static inline struct mthca_cqe *get_cqe(struct mthca_cq *cq, int entry)
{
return cq->buf.buf + entry * MTHCA_CQ_ENTRY_SIZE;
}
static inline struct mthca_cqe *cqe_sw(struct mthca_cq *cq, int i)
{
struct mthca_cqe *cqe = get_cqe(cq, i);
return MTHCA_CQ_ENTRY_OWNER_HW & cqe->owner ? NULL : cqe;
}
static inline struct mthca_cqe *next_cqe_sw(struct mthca_cq *cq)
{
return cqe_sw(cq, cq->cons_index & cq->ibv_cq.cqe);
}
static inline void set_cqe_hw(struct mthca_cqe *cqe)
{
VALGRIND_MAKE_MEM_UNDEFINED(cqe, sizeof *cqe);
cqe->owner = MTHCA_CQ_ENTRY_OWNER_HW;
}
/*
* incr is ignored in native Arbel (mem-free) mode, so cq->cons_index
* should be correct before calling update_cons_index().
*/
static inline void update_cons_index(struct mthca_cq *cq, int incr)
{
uint32_t doorbell[2];
if (mthca_is_memfree(cq->ibv_cq.context)) {
*cq->set_ci_db = htonl(cq->cons_index);
wmb();
} else {
doorbell[0] = htonl(MTHCA_TAVOR_CQ_DB_INC_CI | cq->cqn);
doorbell[1] = htonl(incr - 1);
mthca_write64(doorbell, to_mctx(cq->ibv_cq.context), MTHCA_CQ_DOORBELL);
}
}
static void dump_cqe(void *cqe_ptr)
{
uint32_t *cqe = cqe_ptr;
int i;
for (i = 0; i < 8; ++i)
printf(" [%2x] %08x\n", i * 4, ntohl(((uint32_t *) cqe)[i]));
}
static int handle_error_cqe(struct mthca_cq *cq,
struct mthca_qp *qp, int wqe_index, int is_send,
struct mthca_err_cqe *cqe,
struct ibv_wc *wc, int *free_cqe)
{
int err;
int dbd;
uint32_t new_wqe;
if (cqe->syndrome == SYNDROME_LOCAL_QP_OP_ERR) {
printf("local QP operation err "
"(QPN %06x, WQE @ %08x, CQN %06x, index %d)\n",
ntohl(cqe->my_qpn), ntohl(cqe->wqe),
cq->cqn, cq->cons_index);
dump_cqe(cqe);
}
/*
* For completions in error, only work request ID, status, vendor error
* (and freed resource count for RD) have to be set.
*/
switch (cqe->syndrome) {
case SYNDROME_LOCAL_LENGTH_ERR:
wc->status = IBV_WC_LOC_LEN_ERR;
break;
case SYNDROME_LOCAL_QP_OP_ERR:
wc->status = IBV_WC_LOC_QP_OP_ERR;
break;
case SYNDROME_LOCAL_EEC_OP_ERR:
wc->status = IBV_WC_LOC_EEC_OP_ERR;
break;
case SYNDROME_LOCAL_PROT_ERR:
wc->status = IBV_WC_LOC_PROT_ERR;
break;
case SYNDROME_WR_FLUSH_ERR:
wc->status = IBV_WC_WR_FLUSH_ERR;
break;
case SYNDROME_MW_BIND_ERR:
wc->status = IBV_WC_MW_BIND_ERR;
break;
case SYNDROME_BAD_RESP_ERR:
wc->status = IBV_WC_BAD_RESP_ERR;
break;
case SYNDROME_LOCAL_ACCESS_ERR:
wc->status = IBV_WC_LOC_ACCESS_ERR;
break;
case SYNDROME_REMOTE_INVAL_REQ_ERR:
wc->status = IBV_WC_REM_INV_REQ_ERR;
break;
case SYNDROME_REMOTE_ACCESS_ERR:
wc->status = IBV_WC_REM_ACCESS_ERR;
break;
case SYNDROME_REMOTE_OP_ERR:
wc->status = IBV_WC_REM_OP_ERR;
break;
case SYNDROME_RETRY_EXC_ERR:
wc->status = IBV_WC_RETRY_EXC_ERR;
break;
case SYNDROME_RNR_RETRY_EXC_ERR:
wc->status = IBV_WC_RNR_RETRY_EXC_ERR;
break;
case SYNDROME_LOCAL_RDD_VIOL_ERR:
wc->status = IBV_WC_LOC_RDD_VIOL_ERR;
break;
case SYNDROME_REMOTE_INVAL_RD_REQ_ERR:
wc->status = IBV_WC_REM_INV_RD_REQ_ERR;
break;
case SYNDROME_REMOTE_ABORTED_ERR:
wc->status = IBV_WC_REM_ABORT_ERR;
break;
case SYNDROME_INVAL_EECN_ERR:
wc->status = IBV_WC_INV_EECN_ERR;
break;
case SYNDROME_INVAL_EEC_STATE_ERR:
wc->status = IBV_WC_INV_EEC_STATE_ERR;
break;
default:
wc->status = IBV_WC_GENERAL_ERR;
break;
}
wc->vendor_err = cqe->vendor_err;
/*
* Mem-free HCAs always generate one CQE per WQE, even in the
* error case, so we don't have to check the doorbell count, etc.
*/
if (mthca_is_memfree(cq->ibv_cq.context))
return 0;
err = mthca_free_err_wqe(qp, is_send, wqe_index, &dbd, &new_wqe);
if (err)
return err;
/*
* If we're at the end of the WQE chain, or we've used up our
* doorbell count, free the CQE. Otherwise just update it for
* the next poll operation.
*
* This doesn't apply to mem-free HCAs, which never use the
* doorbell count field. In that case we always free the CQE.
*/
if (mthca_is_memfree(cq->ibv_cq.context) ||
!(new_wqe & htonl(0x3f)) || (!cqe->db_cnt && dbd))
return 0;
cqe->db_cnt = htons(ntohs(cqe->db_cnt) - dbd);
cqe->wqe = new_wqe;
cqe->syndrome = SYNDROME_WR_FLUSH_ERR;
*free_cqe = 0;
return 0;
}
static inline int mthca_poll_one(struct mthca_cq *cq,
struct mthca_qp **cur_qp,
int *freed,
struct ibv_wc *wc)
{
struct mthca_wq *wq;
struct mthca_cqe *cqe;
struct mthca_srq *srq;
uint32_t qpn;
uint32_t wqe;
int wqe_index;
int is_error;
int is_send;
int free_cqe = 1;
int err = 0;
cqe = next_cqe_sw(cq);
if (!cqe)
return CQ_EMPTY;
VALGRIND_MAKE_MEM_DEFINED(cqe, sizeof *cqe);
/*
* Make sure we read CQ entry contents after we've checked the
* ownership bit.
*/
rmb();
qpn = ntohl(cqe->my_qpn);
is_error = (cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
MTHCA_ERROR_CQE_OPCODE_MASK;
is_send = is_error ? cqe->opcode & 0x01 : cqe->is_send & 0x80;
if (!*cur_qp || ntohl(cqe->my_qpn) != (*cur_qp)->ibv_qp.qp_num) {
/*
* We do not have to take the QP table lock here,
* because CQs will be locked while QPs are removed
* from the table.
*/
*cur_qp = mthca_find_qp(to_mctx(cq->ibv_cq.context), ntohl(cqe->my_qpn));
if (!*cur_qp) {
err = CQ_POLL_ERR;
goto out;
}
}
wc->qp_num = (*cur_qp)->ibv_qp.qp_num;
if (is_send) {
wq = &(*cur_qp)->sq;
wqe_index = ((ntohl(cqe->wqe) - (*cur_qp)->send_wqe_offset) >> wq->wqe_shift);
wc->wr_id = (*cur_qp)->wrid[wqe_index + (*cur_qp)->rq.max];
} else if ((*cur_qp)->ibv_qp.srq) {
srq = to_msrq((*cur_qp)->ibv_qp.srq);
wqe = htonl(cqe->wqe);
wq = NULL;
wqe_index = wqe >> srq->wqe_shift;
wc->wr_id = srq->wrid[wqe_index];
mthca_free_srq_wqe(srq, wqe_index);
} else {
int32_t wqe;
wq = &(*cur_qp)->rq;
wqe = ntohl(cqe->wqe);
wqe_index = wqe >> wq->wqe_shift;
/*
* WQE addr == base - 1 might be reported by Sinai FW
* 1.0.800 and Arbel FW 5.1.400 in receive completion
* with error instead of (rq size - 1). This bug
* should be fixed in later FW revisions.
*/
if (wqe_index < 0)
wqe_index = wq->max - 1;
wc->wr_id = (*cur_qp)->wrid[wqe_index];
}
if (wq) {
if (wq->last_comp < wqe_index)
wq->tail += wqe_index - wq->last_comp;
else
wq->tail += wqe_index + wq->max - wq->last_comp;
wq->last_comp = wqe_index;
}
if (is_error) {
err = handle_error_cqe(cq, *cur_qp, wqe_index, is_send,
(struct mthca_err_cqe *) cqe,
wc, &free_cqe);
goto out;
}
if (is_send) {
wc->wc_flags = 0;
switch (cqe->opcode) {
case MTHCA_OPCODE_RDMA_WRITE:
wc->opcode = IBV_WC_RDMA_WRITE;
break;
case MTHCA_OPCODE_RDMA_WRITE_IMM:
wc->opcode = IBV_WC_RDMA_WRITE;
wc->wc_flags |= IBV_WC_WITH_IMM;
break;
case MTHCA_OPCODE_SEND:
wc->opcode = IBV_WC_SEND;
break;
case MTHCA_OPCODE_SEND_IMM:
wc->opcode = IBV_WC_SEND;
wc->wc_flags |= IBV_WC_WITH_IMM;
break;
case MTHCA_OPCODE_RDMA_READ:
wc->opcode = IBV_WC_RDMA_READ;
wc->byte_len = ntohl(cqe->byte_cnt);
break;
case MTHCA_OPCODE_ATOMIC_CS:
wc->opcode = IBV_WC_COMP_SWAP;
wc->byte_len = ntohl(cqe->byte_cnt);
break;
case MTHCA_OPCODE_ATOMIC_FA:
wc->opcode = IBV_WC_FETCH_ADD;
wc->byte_len = ntohl(cqe->byte_cnt);
break;
case MTHCA_OPCODE_BIND_MW:
wc->opcode = IBV_WC_BIND_MW;
break;
default:
/* assume it's a send completion */
wc->opcode = IBV_WC_SEND;
break;
}
} else {
wc->byte_len = ntohl(cqe->byte_cnt);
switch (cqe->opcode & 0x1f) {
case IBV_OPCODE_SEND_LAST_WITH_IMMEDIATE:
case IBV_OPCODE_SEND_ONLY_WITH_IMMEDIATE:
wc->wc_flags = IBV_WC_WITH_IMM;
wc->imm_data = cqe->imm_etype_pkey_eec;
wc->opcode = IBV_WC_RECV;
break;
case IBV_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE:
case IBV_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE:
wc->wc_flags = IBV_WC_WITH_IMM;
wc->imm_data = cqe->imm_etype_pkey_eec;
wc->opcode = IBV_WC_RECV_RDMA_WITH_IMM;
break;
default:
wc->wc_flags = 0;
wc->opcode = IBV_WC_RECV;
break;
}
wc->slid = ntohs(cqe->rlid);
wc->sl = ntohs(cqe->sl_g_mlpath) >> 12;
wc->src_qp = ntohl(cqe->rqpn) & 0xffffff;
wc->dlid_path_bits = ntohs(cqe->sl_g_mlpath) & 0x7f;
wc->pkey_index = ntohl(cqe->imm_etype_pkey_eec) >> 16;
wc->wc_flags |= ntohs(cqe->sl_g_mlpath) & 0x80 ?
IBV_WC_GRH : 0;
}
wc->status = IBV_WC_SUCCESS;
out:
if (free_cqe) {
set_cqe_hw(cqe);
++(*freed);
++cq->cons_index;
}
return err;
}
int mthca_poll_cq(struct ibv_cq *ibcq, int ne, struct ibv_wc *wc)
{
struct mthca_cq *cq = to_mcq(ibcq);
struct mthca_qp *qp = NULL;
int npolled;
int err = CQ_OK;
int freed = 0;
pthread_spin_lock(&cq->lock);
for (npolled = 0; npolled < ne; ++npolled) {
err = mthca_poll_one(cq, &qp, &freed, wc + npolled);
if (err != CQ_OK)
break;
}
if (freed) {
wmb();
update_cons_index(cq, freed);
}
pthread_spin_unlock(&cq->lock);
return err == CQ_POLL_ERR ? err : npolled;
}
int mthca_tavor_arm_cq(struct ibv_cq *cq, int solicited)
{
uint32_t doorbell[2];
doorbell[0] = htonl((solicited ?
MTHCA_TAVOR_CQ_DB_REQ_NOT_SOL :
MTHCA_TAVOR_CQ_DB_REQ_NOT) |
to_mcq(cq)->cqn);
doorbell[1] = 0xffffffff;
mthca_write64(doorbell, to_mctx(cq->context), MTHCA_CQ_DOORBELL);
return 0;
}
int mthca_arbel_arm_cq(struct ibv_cq *ibvcq, int solicited)
{
struct mthca_cq *cq = to_mcq(ibvcq);
uint32_t doorbell[2];
uint32_t sn;
uint32_t ci;
sn = cq->arm_sn & 3;
ci = htonl(cq->cons_index);
doorbell[0] = ci;
doorbell[1] = htonl((cq->cqn << 8) | (2 << 5) | (sn << 3) |
(solicited ? 1 : 2));
mthca_write_db_rec(doorbell, cq->arm_db);
/*
* Make sure that the doorbell record in host memory is
* written before ringing the doorbell via PCI MMIO.
*/
wmb();
doorbell[0] = htonl((sn << 28) |
(solicited ?
MTHCA_ARBEL_CQ_DB_REQ_NOT_SOL :
MTHCA_ARBEL_CQ_DB_REQ_NOT) |
cq->cqn);
doorbell[1] = ci;
mthca_write64(doorbell, to_mctx(ibvcq->context), MTHCA_CQ_DOORBELL);
return 0;
}
void mthca_arbel_cq_event(struct ibv_cq *cq)
{
to_mcq(cq)->arm_sn++;
}
static inline int is_recv_cqe(struct mthca_cqe *cqe)
{
if ((cqe->opcode & MTHCA_ERROR_CQE_OPCODE_MASK) ==
MTHCA_ERROR_CQE_OPCODE_MASK)
return !(cqe->opcode & 0x01);
else
return !(cqe->is_send & 0x80);
}
void __mthca_cq_clean(struct mthca_cq *cq, uint32_t qpn, struct mthca_srq *srq)
{
struct mthca_cqe *cqe;
uint32_t prod_index;
int i, nfreed = 0;
/*
* First we need to find the current producer index, so we
* know where to start cleaning from. It doesn't matter if HW
* adds new entries after this loop -- the QP we're worried
* about is already in RESET, so the new entries won't come
* from our QP and therefore don't need to be checked.
*/
for (prod_index = cq->cons_index;
cqe_sw(cq, prod_index & cq->ibv_cq.cqe);
++prod_index)
if (prod_index == cq->cons_index + cq->ibv_cq.cqe)
break;
/*
* Now sweep backwards through the CQ, removing CQ entries
* that match our QP by copying older entries on top of them.
*/
while ((int) --prod_index - (int) cq->cons_index >= 0) {
cqe = get_cqe(cq, prod_index & cq->ibv_cq.cqe);
if (cqe->my_qpn == htonl(qpn)) {
if (srq && is_recv_cqe(cqe))
mthca_free_srq_wqe(srq,
ntohl(cqe->wqe) >> srq->wqe_shift);
++nfreed;
} else if (nfreed)
memcpy(get_cqe(cq, (prod_index + nfreed) & cq->ibv_cq.cqe),
cqe, MTHCA_CQ_ENTRY_SIZE);
}
if (nfreed) {
for (i = 0; i < nfreed; ++i)
set_cqe_hw(get_cqe(cq, (cq->cons_index + i) & cq->ibv_cq.cqe));
wmb();
cq->cons_index += nfreed;
update_cons_index(cq, nfreed);
}
}
void mthca_cq_clean(struct mthca_cq *cq, uint32_t qpn, struct mthca_srq *srq)
{
pthread_spin_lock(&cq->lock);
__mthca_cq_clean(cq, qpn, srq);
pthread_spin_unlock(&cq->lock);
}
void mthca_cq_resize_copy_cqes(struct mthca_cq *cq, void *buf, int old_cqe)
{
int i;
/*
* In Tavor mode, the hardware keeps the consumer and producer
* indices mod the CQ size. Since we might be making the CQ
* bigger, we need to deal with the case where the producer
* index wrapped around before the CQ was resized.
*/
if (!mthca_is_memfree(cq->ibv_cq.context) && old_cqe < cq->ibv_cq.cqe) {
cq->cons_index &= old_cqe;
if (cqe_sw(cq, old_cqe))
cq->cons_index -= old_cqe + 1;
}
for (i = cq->cons_index; cqe_sw(cq, i & old_cqe); ++i)
memcpy(buf + (i & cq->ibv_cq.cqe) * MTHCA_CQ_ENTRY_SIZE,
get_cqe(cq, i & old_cqe), MTHCA_CQ_ENTRY_SIZE);
}
int mthca_alloc_cq_buf(struct mthca_device *dev, struct mthca_buf *buf, int nent)
{
int i;
if (mthca_alloc_buf(buf, align(nent * MTHCA_CQ_ENTRY_SIZE, dev->page_size),
dev->page_size))
return -1;
for (i = 0; i < nent; ++i)
((struct mthca_cqe *) buf->buf)[i].owner = MTHCA_CQ_ENTRY_OWNER_HW;
return 0;
}