linux/drivers/infiniband/hw/ehca/ehca_reqs.c

955 lines
26 KiB
C
Raw Normal View History

/*
* IBM eServer eHCA Infiniband device driver for Linux on POWER
*
* post_send/recv, poll_cq, req_notify
*
* Authors: Hoang-Nam Nguyen <hnguyen@de.ibm.com>
* Waleri Fomin <fomin@de.ibm.com>
* Joachim Fenkes <fenkes@de.ibm.com>
* Reinhard Ernst <rernst@de.ibm.com>
*
* Copyright (c) 2005 IBM Corporation
*
* All rights reserved.
*
* This source code is distributed under a dual license of GPL v2.0 and OpenIB
* BSD.
*
* OpenIB BSD License
*
* 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.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER
* IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
#include <asm/system.h>
#include "ehca_classes.h"
#include "ehca_tools.h"
#include "ehca_qes.h"
#include "ehca_iverbs.h"
#include "hcp_if.h"
#include "hipz_fns.h"
/* in RC traffic, insert an empty RDMA READ every this many packets */
#define ACK_CIRC_THRESHOLD 2000000
static u64 replace_wr_id(u64 wr_id, u16 idx)
{
u64 ret;
ret = wr_id & ~QMAP_IDX_MASK;
ret |= idx & QMAP_IDX_MASK;
return ret;
}
static u16 get_app_wr_id(u64 wr_id)
{
return wr_id & QMAP_IDX_MASK;
}
static inline int ehca_write_rwqe(struct ipz_queue *ipz_rqueue,
struct ehca_wqe *wqe_p,
struct ib_recv_wr *recv_wr,
u32 rq_map_idx)
{
u8 cnt_ds;
if (unlikely((recv_wr->num_sge < 0) ||
(recv_wr->num_sge > ipz_rqueue->act_nr_of_sg))) {
ehca_gen_err("Invalid number of WQE SGE. "
"num_sqe=%x max_nr_of_sg=%x",
recv_wr->num_sge, ipz_rqueue->act_nr_of_sg);
return -EINVAL; /* invalid SG list length */
}
/* clear wqe header until sglist */
memset(wqe_p, 0, offsetof(struct ehca_wqe, u.ud_av.sg_list));
wqe_p->work_request_id = replace_wr_id(recv_wr->wr_id, rq_map_idx);
wqe_p->nr_of_data_seg = recv_wr->num_sge;
for (cnt_ds = 0; cnt_ds < recv_wr->num_sge; cnt_ds++) {
wqe_p->u.all_rcv.sg_list[cnt_ds].vaddr =
recv_wr->sg_list[cnt_ds].addr;
wqe_p->u.all_rcv.sg_list[cnt_ds].lkey =
recv_wr->sg_list[cnt_ds].lkey;
wqe_p->u.all_rcv.sg_list[cnt_ds].length =
recv_wr->sg_list[cnt_ds].length;
}
if (ehca_debug_level >= 3) {
ehca_gen_dbg("RECEIVE WQE written into ipz_rqueue=%p",
ipz_rqueue);
ehca_dmp(wqe_p, 16*(6 + wqe_p->nr_of_data_seg), "recv wqe");
}
return 0;
}
#if defined(DEBUG_GSI_SEND_WR)
/* need ib_mad struct */
#include <rdma/ib_mad.h>
static void trace_send_wr_ud(const struct ib_send_wr *send_wr)
{
int idx;
int j;
while (send_wr) {
struct ib_mad_hdr *mad_hdr = send_wr->wr.ud.mad_hdr;
struct ib_sge *sge = send_wr->sg_list;
ehca_gen_dbg("send_wr#%x wr_id=%lx num_sge=%x "
"send_flags=%x opcode=%x", idx, send_wr->wr_id,
send_wr->num_sge, send_wr->send_flags,
send_wr->opcode);
if (mad_hdr) {
ehca_gen_dbg("send_wr#%x mad_hdr base_version=%x "
"mgmt_class=%x class_version=%x method=%x "
"status=%x class_specific=%x tid=%lx "
"attr_id=%x resv=%x attr_mod=%x",
idx, mad_hdr->base_version,
mad_hdr->mgmt_class,
mad_hdr->class_version, mad_hdr->method,
mad_hdr->status, mad_hdr->class_specific,
mad_hdr->tid, mad_hdr->attr_id,
mad_hdr->resv,
mad_hdr->attr_mod);
}
for (j = 0; j < send_wr->num_sge; j++) {
u8 *data = (u8 *)abs_to_virt(sge->addr);
ehca_gen_dbg("send_wr#%x sge#%x addr=%p length=%x "
"lkey=%x",
idx, j, data, sge->length, sge->lkey);
/* assume length is n*16 */
ehca_dmp(data, sge->length, "send_wr#%x sge#%x",
idx, j);
sge++;
} /* eof for j */
idx++;
send_wr = send_wr->next;
} /* eof while send_wr */
}
#endif /* DEBUG_GSI_SEND_WR */
static inline int ehca_write_swqe(struct ehca_qp *qp,
struct ehca_wqe *wqe_p,
const struct ib_send_wr *send_wr,
u32 sq_map_idx,
int hidden)
{
u32 idx;
u64 dma_length;
struct ehca_av *my_av;
u32 remote_qkey = send_wr->wr.ud.remote_qkey;
struct ehca_qmap_entry *qmap_entry = &qp->sq_map.map[sq_map_idx];
if (unlikely((send_wr->num_sge < 0) ||
(send_wr->num_sge > qp->ipz_squeue.act_nr_of_sg))) {
ehca_gen_err("Invalid number of WQE SGE. "
"num_sqe=%x max_nr_of_sg=%x",
send_wr->num_sge, qp->ipz_squeue.act_nr_of_sg);
return -EINVAL; /* invalid SG list length */
}
/* clear wqe header until sglist */
memset(wqe_p, 0, offsetof(struct ehca_wqe, u.ud_av.sg_list));
wqe_p->work_request_id = replace_wr_id(send_wr->wr_id, sq_map_idx);
qmap_entry->app_wr_id = get_app_wr_id(send_wr->wr_id);
qmap_entry->reported = 0;
qmap_entry->cqe_req = 0;
switch (send_wr->opcode) {
case IB_WR_SEND:
case IB_WR_SEND_WITH_IMM:
wqe_p->optype = WQE_OPTYPE_SEND;
break;
case IB_WR_RDMA_WRITE:
case IB_WR_RDMA_WRITE_WITH_IMM:
wqe_p->optype = WQE_OPTYPE_RDMAWRITE;
break;
case IB_WR_RDMA_READ:
wqe_p->optype = WQE_OPTYPE_RDMAREAD;
break;
default:
ehca_gen_err("Invalid opcode=%x", send_wr->opcode);
return -EINVAL; /* invalid opcode */
}
wqe_p->wqef = (send_wr->opcode) & WQEF_HIGH_NIBBLE;
wqe_p->wr_flag = 0;
if ((send_wr->send_flags & IB_SEND_SIGNALED ||
qp->init_attr.sq_sig_type == IB_SIGNAL_ALL_WR)
&& !hidden) {
wqe_p->wr_flag |= WQE_WRFLAG_REQ_SIGNAL_COM;
qmap_entry->cqe_req = 1;
}
if (send_wr->opcode == IB_WR_SEND_WITH_IMM ||
send_wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
/* this might not work as long as HW does not support it */
2008-04-17 04:09:32 +00:00
wqe_p->immediate_data = be32_to_cpu(send_wr->ex.imm_data);
wqe_p->wr_flag |= WQE_WRFLAG_IMM_DATA_PRESENT;
}
wqe_p->nr_of_data_seg = send_wr->num_sge;
switch (qp->qp_type) {
case IB_QPT_SMI:
case IB_QPT_GSI:
/* no break is intential here */
case IB_QPT_UD:
/* IB 1.2 spec C10-15 compliance */
if (send_wr->wr.ud.remote_qkey & 0x80000000)
remote_qkey = qp->qkey;
wqe_p->destination_qp_number = send_wr->wr.ud.remote_qpn << 8;
wqe_p->local_ee_context_qkey = remote_qkey;
if (unlikely(!send_wr->wr.ud.ah)) {
ehca_gen_err("wr.ud.ah is NULL. qp=%p", qp);
return -EINVAL;
}
if (unlikely(send_wr->wr.ud.remote_qpn == 0)) {
ehca_gen_err("dest QP# is 0. qp=%x", qp->real_qp_num);
return -EINVAL;
}
my_av = container_of(send_wr->wr.ud.ah, struct ehca_av, ib_ah);
wqe_p->u.ud_av.ud_av = my_av->av;
/*
* omitted check of IB_SEND_INLINE
* since HW does not support it
*/
for (idx = 0; idx < send_wr->num_sge; idx++) {
wqe_p->u.ud_av.sg_list[idx].vaddr =
send_wr->sg_list[idx].addr;
wqe_p->u.ud_av.sg_list[idx].lkey =
send_wr->sg_list[idx].lkey;
wqe_p->u.ud_av.sg_list[idx].length =
send_wr->sg_list[idx].length;
} /* eof for idx */
if (qp->qp_type == IB_QPT_SMI ||
qp->qp_type == IB_QPT_GSI)
wqe_p->u.ud_av.ud_av.pmtu = 1;
if (qp->qp_type == IB_QPT_GSI) {
wqe_p->pkeyi = send_wr->wr.ud.pkey_index;
#ifdef DEBUG_GSI_SEND_WR
trace_send_wr_ud(send_wr);
#endif /* DEBUG_GSI_SEND_WR */
}
break;
case IB_QPT_UC:
if (send_wr->send_flags & IB_SEND_FENCE)
wqe_p->wr_flag |= WQE_WRFLAG_FENCE;
/* no break is intentional here */
case IB_QPT_RC:
/* TODO: atomic not implemented */
wqe_p->u.nud.remote_virtual_address =
send_wr->wr.rdma.remote_addr;
wqe_p->u.nud.rkey = send_wr->wr.rdma.rkey;
/*
* omitted checking of IB_SEND_INLINE
* since HW does not support it
*/
dma_length = 0;
for (idx = 0; idx < send_wr->num_sge; idx++) {
wqe_p->u.nud.sg_list[idx].vaddr =
send_wr->sg_list[idx].addr;
wqe_p->u.nud.sg_list[idx].lkey =
send_wr->sg_list[idx].lkey;
wqe_p->u.nud.sg_list[idx].length =
send_wr->sg_list[idx].length;
dma_length += send_wr->sg_list[idx].length;
} /* eof idx */
wqe_p->u.nud.atomic_1st_op_dma_len = dma_length;
/* unsolicited ack circumvention */
if (send_wr->opcode == IB_WR_RDMA_READ) {
/* on RDMA read, switch on and reset counters */
qp->message_count = qp->packet_count = 0;
qp->unsol_ack_circ = 1;
} else
/* else estimate #packets */
qp->packet_count += (dma_length >> qp->mtu_shift) + 1;
break;
default:
ehca_gen_err("Invalid qptype=%x", qp->qp_type);
return -EINVAL;
}
if (ehca_debug_level >= 3) {
ehca_gen_dbg("SEND WQE written into queue qp=%p ", qp);
ehca_dmp( wqe_p, 16*(6 + wqe_p->nr_of_data_seg), "send wqe");
}
return 0;
}
/* map_ib_wc_status converts raw cqe_status to ib_wc_status */
static inline void map_ib_wc_status(u32 cqe_status,
enum ib_wc_status *wc_status)
{
if (unlikely(cqe_status & WC_STATUS_ERROR_BIT)) {
switch (cqe_status & 0x3F) {
case 0x01:
case 0x21:
*wc_status = IB_WC_LOC_LEN_ERR;
break;
case 0x02:
case 0x22:
*wc_status = IB_WC_LOC_QP_OP_ERR;
break;
case 0x03:
case 0x23:
*wc_status = IB_WC_LOC_EEC_OP_ERR;
break;
case 0x04:
case 0x24:
*wc_status = IB_WC_LOC_PROT_ERR;
break;
case 0x05:
case 0x25:
*wc_status = IB_WC_WR_FLUSH_ERR;
break;
case 0x06:
*wc_status = IB_WC_MW_BIND_ERR;
break;
case 0x07: /* remote error - look into bits 20:24 */
switch ((cqe_status
& WC_STATUS_REMOTE_ERROR_FLAGS) >> 11) {
case 0x0:
/*
* PSN Sequence Error!
* couldn't find a matching status!
*/
*wc_status = IB_WC_GENERAL_ERR;
break;
case 0x1:
*wc_status = IB_WC_REM_INV_REQ_ERR;
break;
case 0x2:
*wc_status = IB_WC_REM_ACCESS_ERR;
break;
case 0x3:
*wc_status = IB_WC_REM_OP_ERR;
break;
case 0x4:
*wc_status = IB_WC_REM_INV_RD_REQ_ERR;
break;
}
break;
case 0x08:
*wc_status = IB_WC_RETRY_EXC_ERR;
break;
case 0x09:
*wc_status = IB_WC_RNR_RETRY_EXC_ERR;
break;
case 0x0A:
case 0x2D:
*wc_status = IB_WC_REM_ABORT_ERR;
break;
case 0x0B:
case 0x2E:
*wc_status = IB_WC_INV_EECN_ERR;
break;
case 0x0C:
case 0x2F:
*wc_status = IB_WC_INV_EEC_STATE_ERR;
break;
case 0x0D:
*wc_status = IB_WC_BAD_RESP_ERR;
break;
case 0x10:
/* WQE purged */
*wc_status = IB_WC_WR_FLUSH_ERR;
break;
default:
*wc_status = IB_WC_FATAL_ERR;
}
} else
*wc_status = IB_WC_SUCCESS;
}
static inline int post_one_send(struct ehca_qp *my_qp,
struct ib_send_wr *cur_send_wr,
int hidden)
{
struct ehca_wqe *wqe_p;
int ret;
u32 sq_map_idx;
u64 start_offset = my_qp->ipz_squeue.current_q_offset;
/* get pointer next to free WQE */
wqe_p = ipz_qeit_get_inc(&my_qp->ipz_squeue);
if (unlikely(!wqe_p)) {
/* too many posted work requests: queue overflow */
ehca_err(my_qp->ib_qp.device, "Too many posted WQEs "
"qp_num=%x", my_qp->ib_qp.qp_num);
return -ENOMEM;
}
/*
* Get the index of the WQE in the send queue. The same index is used
* for writing into the sq_map.
*/
sq_map_idx = start_offset / my_qp->ipz_squeue.qe_size;
/* write a SEND WQE into the QUEUE */
ret = ehca_write_swqe(my_qp, wqe_p, cur_send_wr, sq_map_idx, hidden);
/*
* if something failed,
* reset the free entry pointer to the start value
*/
if (unlikely(ret)) {
my_qp->ipz_squeue.current_q_offset = start_offset;
ehca_err(my_qp->ib_qp.device, "Could not write WQE "
"qp_num=%x", my_qp->ib_qp.qp_num);
return -EINVAL;
}
return 0;
}
int ehca_post_send(struct ib_qp *qp,
struct ib_send_wr *send_wr,
struct ib_send_wr **bad_send_wr)
{
struct ehca_qp *my_qp = container_of(qp, struct ehca_qp, ib_qp);
int wqe_cnt = 0;
int ret = 0;
unsigned long flags;
/* Reject WR if QP is in RESET, INIT or RTR state */
if (unlikely(my_qp->state < IB_QPS_RTS)) {
ehca_err(qp->device, "Invalid QP state qp_state=%d qpn=%x",
my_qp->state, qp->qp_num);
ret = -EINVAL;
goto out;
}
/* LOCK the QUEUE */
spin_lock_irqsave(&my_qp->spinlock_s, flags);
/* Send an empty extra RDMA read if:
* 1) there has been an RDMA read on this connection before
* 2) no RDMA read occurred for ACK_CIRC_THRESHOLD link packets
* 3) we can be sure that any previous extra RDMA read has been
* processed so we don't overflow the SQ
*/
if (unlikely(my_qp->unsol_ack_circ &&
my_qp->packet_count > ACK_CIRC_THRESHOLD &&
my_qp->message_count > my_qp->init_attr.cap.max_send_wr)) {
/* insert an empty RDMA READ to fix up the remote QP state */
struct ib_send_wr circ_wr;
memset(&circ_wr, 0, sizeof(circ_wr));
circ_wr.opcode = IB_WR_RDMA_READ;
post_one_send(my_qp, &circ_wr, 1); /* ignore retcode */
wqe_cnt++;
ehca_dbg(qp->device, "posted circ wr qp_num=%x", qp->qp_num);
my_qp->message_count = my_qp->packet_count = 0;
}
/* loop processes list of send reqs */
while (send_wr) {
ret = post_one_send(my_qp, send_wr, 0);
if (unlikely(ret)) {
goto post_send_exit0;
}
wqe_cnt++;
send_wr = send_wr->next;
}
post_send_exit0:
iosync(); /* serialize GAL register access */
hipz_update_sqa(my_qp, wqe_cnt);
if (unlikely(ret || ehca_debug_level >= 2))
ehca_dbg(qp->device, "ehca_qp=%p qp_num=%x wqe_cnt=%d ret=%i",
my_qp, qp->qp_num, wqe_cnt, ret);
my_qp->message_count += wqe_cnt;
spin_unlock_irqrestore(&my_qp->spinlock_s, flags);
out:
if (ret)
*bad_send_wr = send_wr;
return ret;
}
static int internal_post_recv(struct ehca_qp *my_qp,
struct ib_device *dev,
struct ib_recv_wr *recv_wr,
struct ib_recv_wr **bad_recv_wr)
{
struct ehca_wqe *wqe_p;
int wqe_cnt = 0;
int ret = 0;
u32 rq_map_idx;
unsigned long flags;
struct ehca_qmap_entry *qmap_entry;
if (unlikely(!HAS_RQ(my_qp))) {
ehca_err(dev, "QP has no RQ ehca_qp=%p qp_num=%x ext_type=%d",
my_qp, my_qp->real_qp_num, my_qp->ext_type);
ret = -ENODEV;
goto out;
}
/* LOCK the QUEUE */
spin_lock_irqsave(&my_qp->spinlock_r, flags);
/* loop processes list of recv reqs */
while (recv_wr) {
u64 start_offset = my_qp->ipz_rqueue.current_q_offset;
/* get pointer next to free WQE */
wqe_p = ipz_qeit_get_inc(&my_qp->ipz_rqueue);
if (unlikely(!wqe_p)) {
/* too many posted work requests: queue overflow */
ret = -ENOMEM;
ehca_err(dev, "Too many posted WQEs "
"qp_num=%x", my_qp->real_qp_num);
goto post_recv_exit0;
}
/*
* Get the index of the WQE in the recv queue. The same index
* is used for writing into the rq_map.
*/
rq_map_idx = start_offset / my_qp->ipz_rqueue.qe_size;
/* write a RECV WQE into the QUEUE */
ret = ehca_write_rwqe(&my_qp->ipz_rqueue, wqe_p, recv_wr,
rq_map_idx);
/*
* if something failed,
* reset the free entry pointer to the start value
*/
if (unlikely(ret)) {
my_qp->ipz_rqueue.current_q_offset = start_offset;
ret = -EINVAL;
ehca_err(dev, "Could not write WQE "
"qp_num=%x", my_qp->real_qp_num);
goto post_recv_exit0;
}
qmap_entry = &my_qp->rq_map.map[rq_map_idx];
qmap_entry->app_wr_id = get_app_wr_id(recv_wr->wr_id);
qmap_entry->reported = 0;
qmap_entry->cqe_req = 1;
wqe_cnt++;
recv_wr = recv_wr->next;
} /* eof for recv_wr */
post_recv_exit0:
iosync(); /* serialize GAL register access */
hipz_update_rqa(my_qp, wqe_cnt);
if (unlikely(ret || ehca_debug_level >= 2))
ehca_dbg(dev, "ehca_qp=%p qp_num=%x wqe_cnt=%d ret=%i",
my_qp, my_qp->real_qp_num, wqe_cnt, ret);
spin_unlock_irqrestore(&my_qp->spinlock_r, flags);
out:
if (ret)
*bad_recv_wr = recv_wr;
return ret;
}
int ehca_post_recv(struct ib_qp *qp,
struct ib_recv_wr *recv_wr,
struct ib_recv_wr **bad_recv_wr)
{
struct ehca_qp *my_qp = container_of(qp, struct ehca_qp, ib_qp);
/* Reject WR if QP is in RESET state */
if (unlikely(my_qp->state == IB_QPS_RESET)) {
ehca_err(qp->device, "Invalid QP state qp_state=%d qpn=%x",
my_qp->state, qp->qp_num);
*bad_recv_wr = recv_wr;
return -EINVAL;
}
return internal_post_recv(my_qp, qp->device, recv_wr, bad_recv_wr);
}
int ehca_post_srq_recv(struct ib_srq *srq,
struct ib_recv_wr *recv_wr,
struct ib_recv_wr **bad_recv_wr)
{
return internal_post_recv(container_of(srq, struct ehca_qp, ib_srq),
srq->device, recv_wr, bad_recv_wr);
}
/*
* ib_wc_opcode table converts ehca wc opcode to ib
* Since we use zero to indicate invalid opcode, the actual ib opcode must
* be decremented!!!
*/
static const u8 ib_wc_opcode[255] = {
[0x01] = IB_WC_RECV+1,
[0x02] = IB_WC_RECV_RDMA_WITH_IMM+1,
[0x04] = IB_WC_BIND_MW+1,
[0x08] = IB_WC_FETCH_ADD+1,
[0x10] = IB_WC_COMP_SWAP+1,
[0x20] = IB_WC_RDMA_WRITE+1,
[0x40] = IB_WC_RDMA_READ+1,
[0x80] = IB_WC_SEND+1
};
/* internal function to poll one entry of cq */
static inline int ehca_poll_cq_one(struct ib_cq *cq, struct ib_wc *wc)
{
int ret = 0, qmap_tail_idx;
struct ehca_cq *my_cq = container_of(cq, struct ehca_cq, ib_cq);
struct ehca_cqe *cqe;
struct ehca_qp *my_qp;
struct ehca_qmap_entry *qmap_entry;
struct ehca_queue_map *qmap;
int cqe_count = 0, is_error;
repoll:
cqe = (struct ehca_cqe *)
ipz_qeit_get_inc_valid(&my_cq->ipz_queue);
if (!cqe) {
ret = -EAGAIN;
if (ehca_debug_level >= 3)
ehca_dbg(cq->device, "Completion queue is empty "
"my_cq=%p cq_num=%x", my_cq, my_cq->cq_number);
goto poll_cq_one_exit0;
}
/* prevents loads being reordered across this point */
rmb();
cqe_count++;
if (unlikely(cqe->status & WC_STATUS_PURGE_BIT)) {
struct ehca_qp *qp;
int purgeflag;
unsigned long flags;
qp = ehca_cq_get_qp(my_cq, cqe->local_qp_number);
if (!qp) {
ehca_err(cq->device, "cq_num=%x qp_num=%x "
"could not find qp -> ignore cqe",
my_cq->cq_number, cqe->local_qp_number);
ehca_dmp(cqe, 64, "cq_num=%x qp_num=%x",
my_cq->cq_number, cqe->local_qp_number);
/* ignore this purged cqe */
goto repoll;
}
spin_lock_irqsave(&qp->spinlock_s, flags);
purgeflag = qp->sqerr_purgeflag;
spin_unlock_irqrestore(&qp->spinlock_s, flags);
if (purgeflag) {
ehca_dbg(cq->device,
"Got CQE with purged bit qp_num=%x src_qp=%x",
cqe->local_qp_number, cqe->remote_qp_number);
if (ehca_debug_level >= 2)
ehca_dmp(cqe, 64, "qp_num=%x src_qp=%x",
cqe->local_qp_number,
cqe->remote_qp_number);
/*
* ignore this to avoid double cqes of bad wqe
* that caused sqe and turn off purge flag
*/
qp->sqerr_purgeflag = 0;
goto repoll;
}
}
is_error = cqe->status & WC_STATUS_ERROR_BIT;
/* trace error CQEs if debug_level >= 1, trace all CQEs if >= 3 */
if (unlikely(ehca_debug_level >= 3 || (ehca_debug_level && is_error))) {
ehca_dbg(cq->device,
"Received %sCOMPLETION ehca_cq=%p cq_num=%x -----",
is_error ? "ERROR " : "", my_cq, my_cq->cq_number);
ehca_dmp(cqe, 64, "ehca_cq=%p cq_num=%x",
my_cq, my_cq->cq_number);
ehca_dbg(cq->device,
"ehca_cq=%p cq_num=%x -------------------------",
my_cq, my_cq->cq_number);
}
read_lock(&ehca_qp_idr_lock);
my_qp = idr_find(&ehca_qp_idr, cqe->qp_token);
read_unlock(&ehca_qp_idr_lock);
if (!my_qp)
goto repoll;
wc->qp = &my_qp->ib_qp;
qmap_tail_idx = get_app_wr_id(cqe->work_request_id);
if (!(cqe->w_completion_flags & WC_SEND_RECEIVE_BIT))
/* We got a send completion. */
qmap = &my_qp->sq_map;
else
/* We got a receive completion. */
qmap = &my_qp->rq_map;
/* advance the tail pointer */
qmap->tail = qmap_tail_idx;
if (is_error) {
/*
* set left_to_poll to 0 because in error state, we will not
* get any additional CQEs
*/
my_qp->sq_map.next_wqe_idx = next_index(my_qp->sq_map.tail,
my_qp->sq_map.entries);
my_qp->sq_map.left_to_poll = 0;
ehca_add_to_err_list(my_qp, 1);
my_qp->rq_map.next_wqe_idx = next_index(my_qp->rq_map.tail,
my_qp->rq_map.entries);
my_qp->rq_map.left_to_poll = 0;
if (HAS_RQ(my_qp))
ehca_add_to_err_list(my_qp, 0);
}
qmap_entry = &qmap->map[qmap_tail_idx];
if (qmap_entry->reported) {
ehca_warn(cq->device, "Double cqe on qp_num=%#x",
my_qp->real_qp_num);
/* found a double cqe, discard it and read next one */
goto repoll;
}
wc->wr_id = replace_wr_id(cqe->work_request_id, qmap_entry->app_wr_id);
qmap_entry->reported = 1;
/* if left_to_poll is decremented to 0, add the QP to the error list */
if (qmap->left_to_poll > 0) {
qmap->left_to_poll--;
if ((my_qp->sq_map.left_to_poll == 0) &&
(my_qp->rq_map.left_to_poll == 0)) {
ehca_add_to_err_list(my_qp, 1);
if (HAS_RQ(my_qp))
ehca_add_to_err_list(my_qp, 0);
}
}
/* eval ib_wc_opcode */
wc->opcode = ib_wc_opcode[cqe->optype]-1;
if (unlikely(wc->opcode == -1)) {
ehca_err(cq->device, "Invalid cqe->OPType=%x cqe->status=%x "
"ehca_cq=%p cq_num=%x",
cqe->optype, cqe->status, my_cq, my_cq->cq_number);
/* dump cqe for other infos */
ehca_dmp(cqe, 64, "ehca_cq=%p cq_num=%x",
my_cq, my_cq->cq_number);
/* update also queue adder to throw away this entry!!! */
goto repoll;
}
/* eval ib_wc_status */
if (unlikely(is_error)) {
/* complete with errors */
map_ib_wc_status(cqe->status, &wc->status);
wc->vendor_err = wc->status;
} else
wc->status = IB_WC_SUCCESS;
wc->byte_len = cqe->nr_bytes_transferred;
wc->pkey_index = cqe->pkey_index;
wc->slid = cqe->rlid;
wc->dlid_path_bits = cqe->dlid;
wc->src_qp = cqe->remote_qp_number;
/*
* HW has "Immed data present" and "GRH present" in bits 6 and 5.
* SW defines those in bits 1 and 0, so we can just shift and mask.
*/
wc->wc_flags = (cqe->w_completion_flags >> 5) & 3;
RDMA/core: Add memory management extensions support This patch adds support for the IB "base memory management extension" (BMME) and the equivalent iWARP operations (which the iWARP verbs mandates all devices must implement). The new operations are: - Allocate an ib_mr for use in fast register work requests. - Allocate/free a physical buffer lists for use in fast register work requests. This allows device drivers to allocate this memory as needed for use in posting send requests (eg via dma_alloc_coherent). - New send queue work requests: * send with remote invalidate * fast register memory region * local invalidate memory region * RDMA read with invalidate local memory region (iWARP only) Consumer interface details: - A new device capability flag IB_DEVICE_MEM_MGT_EXTENSIONS is added to indicate device support for these features. - New send work request opcodes IB_WR_FAST_REG_MR, IB_WR_LOCAL_INV, IB_WR_RDMA_READ_WITH_INV are added. - A new consumer API function, ib_alloc_mr() is added to allocate fast register memory regions. - New consumer API functions, ib_alloc_fast_reg_page_list() and ib_free_fast_reg_page_list() are added to allocate and free device-specific memory for fast registration page lists. - A new consumer API function, ib_update_fast_reg_key(), is added to allow the key portion of the R_Key and L_Key of a fast registration MR to be updated. Consumers call this if desired before posting a IB_WR_FAST_REG_MR work request. Consumers can use this as follows: - MR is allocated with ib_alloc_mr(). - Page list memory is allocated with ib_alloc_fast_reg_page_list(). - MR R_Key/L_Key "key" field is updated with ib_update_fast_reg_key(). - MR made VALID and bound to a specific page list via ib_post_send(IB_WR_FAST_REG_MR) - MR made INVALID via ib_post_send(IB_WR_LOCAL_INV), ib_post_send(IB_WR_RDMA_READ_WITH_INV) or an incoming send with invalidate operation. - MR is deallocated with ib_dereg_mr() - page lists dealloced via ib_free_fast_reg_page_list(). Applications can allocate a fast register MR once, and then can repeatedly bind the MR to different physical block lists (PBLs) via posting work requests to a send queue (SQ). For each outstanding MR-to-PBL binding in the SQ pipe, a fast_reg_page_list needs to be allocated (the fast_reg_page_list is owned by the low-level driver from the consumer posting a work request until the request completes). Thus pipelining can be achieved while still allowing device-specific page_list processing. The 32-bit fast register memory key/STag is composed of a 24-bit index and an 8-bit key. The application can change the key each time it fast registers thus allowing more control over the peer's use of the key/STag (ie it can effectively be changed each time the rkey is rebound to a page list). Signed-off-by: Steve Wise <swise@opengridcomputing.com> Signed-off-by: Roland Dreier <rolandd@cisco.com>
2008-07-15 06:48:45 +00:00
wc->ex.imm_data = cpu_to_be32(cqe->immediate_data);
wc->sl = cqe->service_level;
poll_cq_one_exit0:
if (cqe_count > 0)
hipz_update_feca(my_cq, cqe_count);
return ret;
}
static int generate_flush_cqes(struct ehca_qp *my_qp, struct ib_cq *cq,
struct ib_wc *wc, int num_entries,
struct ipz_queue *ipz_queue, int on_sq)
{
int nr = 0;
struct ehca_wqe *wqe;
u64 offset;
struct ehca_queue_map *qmap;
struct ehca_qmap_entry *qmap_entry;
if (on_sq)
qmap = &my_qp->sq_map;
else
qmap = &my_qp->rq_map;
qmap_entry = &qmap->map[qmap->next_wqe_idx];
while ((nr < num_entries) && (qmap_entry->reported == 0)) {
/* generate flush CQE */
memset(wc, 0, sizeof(*wc));
offset = qmap->next_wqe_idx * ipz_queue->qe_size;
wqe = (struct ehca_wqe *)ipz_qeit_calc(ipz_queue, offset);
if (!wqe) {
ehca_err(cq->device, "Invalid wqe offset=%#llx on "
"qp_num=%#x", offset, my_qp->real_qp_num);
return nr;
}
wc->wr_id = replace_wr_id(wqe->work_request_id,
qmap_entry->app_wr_id);
if (on_sq) {
switch (wqe->optype) {
case WQE_OPTYPE_SEND:
wc->opcode = IB_WC_SEND;
break;
case WQE_OPTYPE_RDMAWRITE:
wc->opcode = IB_WC_RDMA_WRITE;
break;
case WQE_OPTYPE_RDMAREAD:
wc->opcode = IB_WC_RDMA_READ;
break;
default:
ehca_err(cq->device, "Invalid optype=%x",
wqe->optype);
return nr;
}
} else
wc->opcode = IB_WC_RECV;
if (wqe->wr_flag & WQE_WRFLAG_IMM_DATA_PRESENT) {
wc->ex.imm_data = wqe->immediate_data;
wc->wc_flags |= IB_WC_WITH_IMM;
}
wc->status = IB_WC_WR_FLUSH_ERR;
wc->qp = &my_qp->ib_qp;
/* mark as reported and advance next_wqe pointer */
qmap_entry->reported = 1;
qmap->next_wqe_idx = next_index(qmap->next_wqe_idx,
qmap->entries);
qmap_entry = &qmap->map[qmap->next_wqe_idx];
wc++; nr++;
}
return nr;
}
int ehca_poll_cq(struct ib_cq *cq, int num_entries, struct ib_wc *wc)
{
struct ehca_cq *my_cq = container_of(cq, struct ehca_cq, ib_cq);
int nr;
struct ehca_qp *err_qp;
struct ib_wc *current_wc = wc;
int ret = 0;
unsigned long flags;
int entries_left = num_entries;
if (num_entries < 1) {
ehca_err(cq->device, "Invalid num_entries=%d ehca_cq=%p "
"cq_num=%x", num_entries, my_cq, my_cq->cq_number);
ret = -EINVAL;
goto poll_cq_exit0;
}
spin_lock_irqsave(&my_cq->spinlock, flags);
/* generate flush cqes for send queues */
list_for_each_entry(err_qp, &my_cq->sqp_err_list, sq_err_node) {
nr = generate_flush_cqes(err_qp, cq, current_wc, entries_left,
&err_qp->ipz_squeue, 1);
entries_left -= nr;
current_wc += nr;
if (entries_left == 0)
break;
}
/* generate flush cqes for receive queues */
list_for_each_entry(err_qp, &my_cq->rqp_err_list, rq_err_node) {
nr = generate_flush_cqes(err_qp, cq, current_wc, entries_left,
&err_qp->ipz_rqueue, 0);
entries_left -= nr;
current_wc += nr;
if (entries_left == 0)
break;
}
for (nr = 0; nr < entries_left; nr++) {
ret = ehca_poll_cq_one(cq, current_wc);
if (ret)
break;
current_wc++;
} /* eof for nr */
entries_left -= nr;
spin_unlock_irqrestore(&my_cq->spinlock, flags);
if (ret == -EAGAIN || !ret)
ret = num_entries - entries_left;
poll_cq_exit0:
return ret;
}
IB: Return "maybe missed event" hint from ib_req_notify_cq() The semantics defined by the InfiniBand specification say that completion events are only generated when a completions is added to a completion queue (CQ) after completion notification is requested. In other words, this means that the following race is possible: while (CQ is not empty) ib_poll_cq(CQ); // new completion is added after while loop is exited ib_req_notify_cq(CQ); // no event is generated for the existing completion To close this race, the IB spec recommends doing another poll of the CQ after requesting notification. However, it is not always possible to arrange code this way (for example, we have found that NAPI for IPoIB cannot poll after requesting notification). Also, some hardware (eg Mellanox HCAs) actually will generate an event for completions added before the call to ib_req_notify_cq() -- which is allowed by the spec, since there's no way for any upper-layer consumer to know exactly when a completion was really added -- so the extra poll of the CQ is just a waste. Motivated by this, we add a new flag "IB_CQ_REPORT_MISSED_EVENTS" for ib_req_notify_cq() so that it can return a hint about whether the a completion may have been added before the request for notification. The return value of ib_req_notify_cq() is extended so: < 0 means an error occurred while requesting notification == 0 means notification was requested successfully, and if IB_CQ_REPORT_MISSED_EVENTS was passed in, then no events were missed and it is safe to wait for another event. > 0 is only returned if IB_CQ_REPORT_MISSED_EVENTS was passed in. It means that the consumer must poll the CQ again to make sure it is empty to avoid the race described above. We add a flag to enable this behavior rather than turning it on unconditionally, because checking for missed events may incur significant overhead for some low-level drivers, and consumers that don't care about the results of this test shouldn't be forced to pay for the test. Signed-off-by: Roland Dreier <rolandd@cisco.com>
2007-05-07 04:02:48 +00:00
int ehca_req_notify_cq(struct ib_cq *cq, enum ib_cq_notify_flags notify_flags)
{
struct ehca_cq *my_cq = container_of(cq, struct ehca_cq, ib_cq);
IB: Return "maybe missed event" hint from ib_req_notify_cq() The semantics defined by the InfiniBand specification say that completion events are only generated when a completions is added to a completion queue (CQ) after completion notification is requested. In other words, this means that the following race is possible: while (CQ is not empty) ib_poll_cq(CQ); // new completion is added after while loop is exited ib_req_notify_cq(CQ); // no event is generated for the existing completion To close this race, the IB spec recommends doing another poll of the CQ after requesting notification. However, it is not always possible to arrange code this way (for example, we have found that NAPI for IPoIB cannot poll after requesting notification). Also, some hardware (eg Mellanox HCAs) actually will generate an event for completions added before the call to ib_req_notify_cq() -- which is allowed by the spec, since there's no way for any upper-layer consumer to know exactly when a completion was really added -- so the extra poll of the CQ is just a waste. Motivated by this, we add a new flag "IB_CQ_REPORT_MISSED_EVENTS" for ib_req_notify_cq() so that it can return a hint about whether the a completion may have been added before the request for notification. The return value of ib_req_notify_cq() is extended so: < 0 means an error occurred while requesting notification == 0 means notification was requested successfully, and if IB_CQ_REPORT_MISSED_EVENTS was passed in, then no events were missed and it is safe to wait for another event. > 0 is only returned if IB_CQ_REPORT_MISSED_EVENTS was passed in. It means that the consumer must poll the CQ again to make sure it is empty to avoid the race described above. We add a flag to enable this behavior rather than turning it on unconditionally, because checking for missed events may incur significant overhead for some low-level drivers, and consumers that don't care about the results of this test shouldn't be forced to pay for the test. Signed-off-by: Roland Dreier <rolandd@cisco.com>
2007-05-07 04:02:48 +00:00
int ret = 0;
IB: Return "maybe missed event" hint from ib_req_notify_cq() The semantics defined by the InfiniBand specification say that completion events are only generated when a completions is added to a completion queue (CQ) after completion notification is requested. In other words, this means that the following race is possible: while (CQ is not empty) ib_poll_cq(CQ); // new completion is added after while loop is exited ib_req_notify_cq(CQ); // no event is generated for the existing completion To close this race, the IB spec recommends doing another poll of the CQ after requesting notification. However, it is not always possible to arrange code this way (for example, we have found that NAPI for IPoIB cannot poll after requesting notification). Also, some hardware (eg Mellanox HCAs) actually will generate an event for completions added before the call to ib_req_notify_cq() -- which is allowed by the spec, since there's no way for any upper-layer consumer to know exactly when a completion was really added -- so the extra poll of the CQ is just a waste. Motivated by this, we add a new flag "IB_CQ_REPORT_MISSED_EVENTS" for ib_req_notify_cq() so that it can return a hint about whether the a completion may have been added before the request for notification. The return value of ib_req_notify_cq() is extended so: < 0 means an error occurred while requesting notification == 0 means notification was requested successfully, and if IB_CQ_REPORT_MISSED_EVENTS was passed in, then no events were missed and it is safe to wait for another event. > 0 is only returned if IB_CQ_REPORT_MISSED_EVENTS was passed in. It means that the consumer must poll the CQ again to make sure it is empty to avoid the race described above. We add a flag to enable this behavior rather than turning it on unconditionally, because checking for missed events may incur significant overhead for some low-level drivers, and consumers that don't care about the results of this test shouldn't be forced to pay for the test. Signed-off-by: Roland Dreier <rolandd@cisco.com>
2007-05-07 04:02:48 +00:00
switch (notify_flags & IB_CQ_SOLICITED_MASK) {
case IB_CQ_SOLICITED:
hipz_set_cqx_n0(my_cq, 1);
break;
case IB_CQ_NEXT_COMP:
hipz_set_cqx_n1(my_cq, 1);
break;
default:
return -EINVAL;
}
IB: Return "maybe missed event" hint from ib_req_notify_cq() The semantics defined by the InfiniBand specification say that completion events are only generated when a completions is added to a completion queue (CQ) after completion notification is requested. In other words, this means that the following race is possible: while (CQ is not empty) ib_poll_cq(CQ); // new completion is added after while loop is exited ib_req_notify_cq(CQ); // no event is generated for the existing completion To close this race, the IB spec recommends doing another poll of the CQ after requesting notification. However, it is not always possible to arrange code this way (for example, we have found that NAPI for IPoIB cannot poll after requesting notification). Also, some hardware (eg Mellanox HCAs) actually will generate an event for completions added before the call to ib_req_notify_cq() -- which is allowed by the spec, since there's no way for any upper-layer consumer to know exactly when a completion was really added -- so the extra poll of the CQ is just a waste. Motivated by this, we add a new flag "IB_CQ_REPORT_MISSED_EVENTS" for ib_req_notify_cq() so that it can return a hint about whether the a completion may have been added before the request for notification. The return value of ib_req_notify_cq() is extended so: < 0 means an error occurred while requesting notification == 0 means notification was requested successfully, and if IB_CQ_REPORT_MISSED_EVENTS was passed in, then no events were missed and it is safe to wait for another event. > 0 is only returned if IB_CQ_REPORT_MISSED_EVENTS was passed in. It means that the consumer must poll the CQ again to make sure it is empty to avoid the race described above. We add a flag to enable this behavior rather than turning it on unconditionally, because checking for missed events may incur significant overhead for some low-level drivers, and consumers that don't care about the results of this test shouldn't be forced to pay for the test. Signed-off-by: Roland Dreier <rolandd@cisco.com>
2007-05-07 04:02:48 +00:00
if (notify_flags & IB_CQ_REPORT_MISSED_EVENTS) {
unsigned long spl_flags;
IB: Return "maybe missed event" hint from ib_req_notify_cq() The semantics defined by the InfiniBand specification say that completion events are only generated when a completions is added to a completion queue (CQ) after completion notification is requested. In other words, this means that the following race is possible: while (CQ is not empty) ib_poll_cq(CQ); // new completion is added after while loop is exited ib_req_notify_cq(CQ); // no event is generated for the existing completion To close this race, the IB spec recommends doing another poll of the CQ after requesting notification. However, it is not always possible to arrange code this way (for example, we have found that NAPI for IPoIB cannot poll after requesting notification). Also, some hardware (eg Mellanox HCAs) actually will generate an event for completions added before the call to ib_req_notify_cq() -- which is allowed by the spec, since there's no way for any upper-layer consumer to know exactly when a completion was really added -- so the extra poll of the CQ is just a waste. Motivated by this, we add a new flag "IB_CQ_REPORT_MISSED_EVENTS" for ib_req_notify_cq() so that it can return a hint about whether the a completion may have been added before the request for notification. The return value of ib_req_notify_cq() is extended so: < 0 means an error occurred while requesting notification == 0 means notification was requested successfully, and if IB_CQ_REPORT_MISSED_EVENTS was passed in, then no events were missed and it is safe to wait for another event. > 0 is only returned if IB_CQ_REPORT_MISSED_EVENTS was passed in. It means that the consumer must poll the CQ again to make sure it is empty to avoid the race described above. We add a flag to enable this behavior rather than turning it on unconditionally, because checking for missed events may incur significant overhead for some low-level drivers, and consumers that don't care about the results of this test shouldn't be forced to pay for the test. Signed-off-by: Roland Dreier <rolandd@cisco.com>
2007-05-07 04:02:48 +00:00
spin_lock_irqsave(&my_cq->spinlock, spl_flags);
ret = ipz_qeit_is_valid(&my_cq->ipz_queue);
spin_unlock_irqrestore(&my_cq->spinlock, spl_flags);
}
return ret;
}