AuxXxilium/linux_dsm_epyc7002
1
0
Fork 0
mirror of https://github.com/AuxXxilium/linux_dsm_epyc7002.git synced 2025-02-18 19:07:51 +07:00
Code Issues Actions Packages Projects Releases Wiki Activity

[PATCH] IB/ipath: share more common code between RC and UC protocols

Browse Source 
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com>
Cc: "Michael S. Tsirkin" <mst@mellanox.co.il>
Cc: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This commit is contained in:
Bryan O'Sullivan 2006-07-01 04:35:50 -07:00 committed by Linus Torvalds
parent 759d57686d
commit ddd4bb2210
6 changed files with 270 additions and 362 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
drivers/infiniband/hw/ipath/ipath_qp.c
View File

@ -709,9 +709,7 @@ struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
spin_lock_init(&qp->r_rq.lock); spin_lock_init(&qp->r_rq.lock);
atomic_set(&qp->refcount, 0); atomic_set(&qp->refcount, 0);
init_waitqueue_head(&qp->wait); init_waitqueue_head(&qp->wait);
tasklet_init(&qp->s_task, tasklet_init(&qp->s_task, ipath_do_ruc_send,
init_attr->qp_type == IB_QPT_RC ?
ipath_do_rc_send : ipath_do_uc_send,
(unsigned long)qp); (unsigned long)qp);
INIT_LIST_HEAD(&qp->piowait); INIT_LIST_HEAD(&qp->piowait);
INIT_LIST_HEAD(&qp->timerwait); INIT_LIST_HEAD(&qp->timerwait);
@ -896,9 +894,9 @@ void ipath_get_credit(struct ipath_qp *qp, u32 aeth)
* as many packets as we like. Otherwise, we have to * as many packets as we like. Otherwise, we have to
* honor the credit field. * honor the credit field.
*/ */
if (credit == IPS_AETH_CREDIT_INVAL) { if (credit == IPS_AETH_CREDIT_INVAL)
qp->s_lsn = (u32) -1; qp->s_lsn = (u32) -1;
} else if (qp->s_lsn != (u32) -1) { else if (qp->s_lsn != (u32) -1) {
/* Compute new LSN (i.e., MSN + credit) */ /* Compute new LSN (i.e., MSN + credit) */
credit = (aeth + credit_table[credit]) & IPS_MSN_MASK; credit = (aeth + credit_table[credit]) & IPS_MSN_MASK;
if (ipath_cmp24(credit, qp->s_lsn) > 0) if (ipath_cmp24(credit, qp->s_lsn) > 0)

204
drivers/infiniband/hw/ipath/ipath_rc.c
View File

@ -73,9 +73,9 @@ static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
* Return bth0 if constructed; otherwise, return 0. * Return bth0 if constructed; otherwise, return 0.
* Note the QP s_lock must be held. * Note the QP s_lock must be held.
*/ */
static inline u32 ipath_make_rc_ack(struct ipath_qp *qp, u32 ipath_make_rc_ack(struct ipath_qp *qp,
struct ipath_other_headers *ohdr, struct ipath_other_headers *ohdr,
u32 pmtu) u32 pmtu)
{ {
struct ipath_sge_state *ss; struct ipath_sge_state *ss;
u32 hwords; u32 hwords;
@ -96,8 +96,7 @@ static inline u32 ipath_make_rc_ack(struct ipath_qp *qp,
if (len > pmtu) { if (len > pmtu) {
len = pmtu; len = pmtu;
qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST); qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
} } else
else
qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY); qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
qp->s_rdma_len -= len; qp->s_rdma_len -= len;
bth0 = qp->s_ack_state << 24; bth0 = qp->s_ack_state << 24;
@ -177,9 +176,9 @@ static inline u32 ipath_make_rc_ack(struct ipath_qp *qp,
* Return 1 if constructed; otherwise, return 0. * Return 1 if constructed; otherwise, return 0.
* Note the QP s_lock must be held. * Note the QP s_lock must be held.
*/ */
static inline int ipath_make_rc_req(struct ipath_qp *qp, int ipath_make_rc_req(struct ipath_qp *qp,
struct ipath_other_headers *ohdr, struct ipath_other_headers *ohdr,
u32 pmtu, u32 *bth0p, u32 *bth2p) u32 pmtu, u32 *bth0p, u32 *bth2p)
{ {
struct ipath_ibdev *dev = to_idev(qp->ibqp.device); struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
struct ipath_sge_state *ss; struct ipath_sge_state *ss;
@ -497,160 +496,33 @@ static inline int ipath_make_rc_req(struct ipath_qp *qp,
return 0; return 0;
} }
static inline void ipath_make_rc_grh(struct ipath_qp *qp,
struct ib_global_route *grh,
u32 nwords)
{
struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
/* GRH header size in 32-bit words. */
qp->s_hdrwords += 10;
qp->s_hdr.u.l.grh.version_tclass_flow =
cpu_to_be32((6 << 28) |
(grh->traffic_class << 20) |
grh->flow_label);
qp->s_hdr.u.l.grh.paylen =
cpu_to_be16(((qp->s_hdrwords - 12) + nwords +
SIZE_OF_CRC) << 2);
/* next_hdr is defined by C8-7 in ch. 8.4.1 */
qp->s_hdr.u.l.grh.next_hdr = 0x1B;
qp->s_hdr.u.l.grh.hop_limit = grh->hop_limit;
/* The SGID is 32-bit aligned. */
qp->s_hdr.u.l.grh.sgid.global.subnet_prefix = dev->gid_prefix;
qp->s_hdr.u.l.grh.sgid.global.interface_id =
ipath_layer_get_guid(dev->dd);
qp->s_hdr.u.l.grh.dgid = grh->dgid;
}
/** /**
* ipath_do_rc_send - perform a send on an RC QP * send_rc_ack - Construct an ACK packet and send it
* @data: contains a pointer to the QP * @qp: a pointer to the QP
* *
* Process entries in the send work queue until credit or queue is * This is called from ipath_rc_rcv() and only uses the receive
* exhausted. Only allow one CPU to send a packet per QP (tasklet). * side QP state.
* Otherwise, after we drop the QP s_lock, two threads could send * Note that RDMA reads are handled in the send side QP state and tasklet.
* packets out of order.
*/ */
void ipath_do_rc_send(unsigned long data)
{
struct ipath_qp *qp = (struct ipath_qp *)data;
struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
unsigned long flags;
u16 lrh0;
u32 nwords;
u32 extra_bytes;
u32 bth0;
u32 bth2;
u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
struct ipath_other_headers *ohdr;
if (test_and_set_bit(IPATH_S_BUSY, &qp->s_flags))
goto bail;
if (unlikely(qp->remote_ah_attr.dlid ==
ipath_layer_get_lid(dev->dd))) {
struct ib_wc wc;
/*
* Pass in an uninitialized ib_wc to be consistent with
* other places where ipath_ruc_loopback() is called.
*/
ipath_ruc_loopback(qp, &wc);
goto clear;
}
ohdr = &qp->s_hdr.u.oth;
if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
ohdr = &qp->s_hdr.u.l.oth;
again:
/* Check for a constructed packet to be sent. */
if (qp->s_hdrwords != 0) {
/*
* If no PIO bufs are available, return. An interrupt will
* call ipath_ib_piobufavail() when one is available.
*/
_VERBS_INFO("h %u %p\n", qp->s_hdrwords, &qp->s_hdr);
_VERBS_INFO("d %u %p %u %p %u %u %u %u\n", qp->s_cur_size,
qp->s_cur_sge->sg_list,
qp->s_cur_sge->num_sge,
qp->s_cur_sge->sge.vaddr,
qp->s_cur_sge->sge.sge_length,
qp->s_cur_sge->sge.length,
qp->s_cur_sge->sge.m,
qp->s_cur_sge->sge.n);
if (ipath_verbs_send(dev->dd, qp->s_hdrwords,
(u32 *) &qp->s_hdr, qp->s_cur_size,
qp->s_cur_sge)) {
ipath_no_bufs_available(qp, dev);
goto bail;
}
dev->n_unicast_xmit++;
/* Record that we sent the packet and s_hdr is empty. */
qp->s_hdrwords = 0;
}
/*
* The lock is needed to synchronize between setting
* qp->s_ack_state, resend timer, and post_send().
*/
spin_lock_irqsave(&qp->s_lock, flags);
/* Sending responses has higher priority over sending requests. */
if (qp->s_ack_state != OP(ACKNOWLEDGE) &&
(bth0 = ipath_make_rc_ack(qp, ohdr, pmtu)) != 0)
bth2 = qp->s_ack_psn++ & IPS_PSN_MASK;
else if (!ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2))
goto done;
spin_unlock_irqrestore(&qp->s_lock, flags);
/* Construct the header. */
extra_bytes = (4 - qp->s_cur_size) & 3;
nwords = (qp->s_cur_size + extra_bytes) >> 2;
lrh0 = IPS_LRH_BTH;
if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
ipath_make_rc_grh(qp, &qp->remote_ah_attr.grh, nwords);
lrh0 = IPS_LRH_GRH;
}
lrh0 |= qp->remote_ah_attr.sl << 4;
qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
SIZE_OF_CRC);
qp->s_hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd));
bth0 |= ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
bth0 |= extra_bytes << 20;
ohdr->bth[0] = cpu_to_be32(bth0);
ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
ohdr->bth[2] = cpu_to_be32(bth2);
/* Check for more work to do. */
goto again;
done:
spin_unlock_irqrestore(&qp->s_lock, flags);
clear:
clear_bit(IPATH_S_BUSY, &qp->s_flags);
bail:
return;
}
static void send_rc_ack(struct ipath_qp *qp) static void send_rc_ack(struct ipath_qp *qp)
{ {
struct ipath_ibdev *dev = to_idev(qp->ibqp.device); struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
u16 lrh0; u16 lrh0;
u32 bth0; u32 bth0;
u32 hwords;
struct ipath_ib_header hdr;
struct ipath_other_headers *ohdr; struct ipath_other_headers *ohdr;
/* Construct the header. */ /* Construct the header. */
ohdr = &qp->s_hdr.u.oth; ohdr = &hdr.u.oth;
lrh0 = IPS_LRH_BTH; lrh0 = IPS_LRH_BTH;
/* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */ /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
qp->s_hdrwords = 6; hwords = 6;
if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) { if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
ipath_make_rc_grh(qp, &qp->remote_ah_attr.grh, 0); hwords += ipath_make_grh(dev, &hdr.u.l.grh,
ohdr = &qp->s_hdr.u.l.oth; &qp->remote_ah_attr.grh,
hwords, 0);
ohdr = &hdr.u.l.oth;
lrh0 = IPS_LRH_GRH; lrh0 = IPS_LRH_GRH;
} }
bth0 = ipath_layer_get_pkey(dev->dd, qp->s_pkey_index); bth0 = ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
@ -658,15 +530,14 @@ static void send_rc_ack(struct ipath_qp *qp)
if (qp->s_ack_state >= OP(COMPARE_SWAP)) { if (qp->s_ack_state >= OP(COMPARE_SWAP)) {
bth0 |= IB_OPCODE_ATOMIC_ACKNOWLEDGE << 24; bth0 |= IB_OPCODE_ATOMIC_ACKNOWLEDGE << 24;
ohdr->u.at.atomic_ack_eth = cpu_to_be64(qp->s_ack_atomic); ohdr->u.at.atomic_ack_eth = cpu_to_be64(qp->s_ack_atomic);
qp->s_hdrwords += sizeof(ohdr->u.at.atomic_ack_eth) / 4; hwords += sizeof(ohdr->u.at.atomic_ack_eth) / 4;
} } else
else
bth0 |= OP(ACKNOWLEDGE) << 24; bth0 |= OP(ACKNOWLEDGE) << 24;
lrh0 |= qp->remote_ah_attr.sl << 4; lrh0 |= qp->remote_ah_attr.sl << 4;
qp->s_hdr.lrh[0] = cpu_to_be16(lrh0); hdr.lrh[0] = cpu_to_be16(lrh0);
qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid); hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + SIZE_OF_CRC); hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
qp->s_hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd)); hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd));
ohdr->bth[0] = cpu_to_be32(bth0); ohdr->bth[0] = cpu_to_be32(bth0);
ohdr->bth[1] = cpu_to_be32(qp->remote_qpn); ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
ohdr->bth[2] = cpu_to_be32(qp->s_ack_psn & IPS_PSN_MASK); ohdr->bth[2] = cpu_to_be32(qp->s_ack_psn & IPS_PSN_MASK);
@ -674,8 +545,7 @@ static void send_rc_ack(struct ipath_qp *qp)
/* /*
* If we can send the ACK, clear the ACK state. * If we can send the ACK, clear the ACK state.
*/ */
if (ipath_verbs_send(dev->dd, qp->s_hdrwords, (u32 *) &qp->s_hdr, if (ipath_verbs_send(dev->dd, hwords, (u32 *) &hdr, 0, NULL) == 0) {
0, NULL) == 0) {
qp->s_ack_state = OP(ACKNOWLEDGE); qp->s_ack_state = OP(ACKNOWLEDGE);
dev->n_rc_qacks++; dev->n_rc_qacks++;
dev->n_unicast_xmit++; dev->n_unicast_xmit++;
@ -805,7 +675,7 @@ void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
* @qp: the QP * @qp: the QP
* @psn: the packet sequence number to restart at * @psn: the packet sequence number to restart at
* *
* This is called from ipath_rc_rcv() to process an incoming RC ACK * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
* for the given QP. * for the given QP.
* Called at interrupt level with the QP s_lock held. * Called at interrupt level with the QP s_lock held.
*/ */
@ -1231,18 +1101,12 @@ static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
* ICRC (4). * ICRC (4).
*/ */
if (unlikely(tlen <= (hdrsize + pad + 8))) { if (unlikely(tlen <= (hdrsize + pad + 8))) {
/* /* XXX Need to generate an error CQ entry. */
* XXX Need to generate an error CQ
* entry.
*/
goto ack_done; goto ack_done;
} }
tlen -= hdrsize + pad + 8; tlen -= hdrsize + pad + 8;
if (unlikely(tlen != qp->s_len)) { if (unlikely(tlen != qp->s_len)) {
/* /* XXX Need to generate an error CQ entry. */
* XXX Need to generate an error CQ
* entry.
*/
goto ack_done; goto ack_done;
} }
if (!header_in_data) if (!header_in_data)
@ -1384,7 +1248,7 @@ static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
case OP(COMPARE_SWAP): case OP(COMPARE_SWAP):
case OP(FETCH_ADD): case OP(FETCH_ADD):
/* /*
* Check for the PSN of the last atomic operations * Check for the PSN of the last atomic operation
* performed and resend the result if found. * performed and resend the result if found.
*/ */
if ((psn & IPS_PSN_MASK) != qp->r_atomic_psn) { if ((psn & IPS_PSN_MASK) != qp->r_atomic_psn) {
@ -1454,11 +1318,6 @@ void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
} else } else
psn = be32_to_cpu(ohdr->bth[2]); psn = be32_to_cpu(ohdr->bth[2]);
} }
/*
* The opcode is in the low byte when its in network order
* (top byte when in host order).
*/
opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
/* /*
* Process responses (ACKs) before anything else. Note that the * Process responses (ACKs) before anything else. Note that the
@ -1466,6 +1325,7 @@ void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
* queue rather than the expected receive packet sequence number. * queue rather than the expected receive packet sequence number.
* In other words, this QP is the requester. * In other words, this QP is the requester.
*/ */
opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) && if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
opcode <= OP(ATOMIC_ACKNOWLEDGE)) { opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn, ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,

245
drivers/infiniband/hw/ipath/ipath_ruc.c
View File

@ -32,6 +32,7 @@
*/ */
#include "ipath_verbs.h" #include "ipath_verbs.h"
#include "ips_common.h"
/* /*
* Convert the AETH RNR timeout code into the number of milliseconds. * Convert the AETH RNR timeout code into the number of milliseconds.
@ -188,7 +189,6 @@ int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
/** /**
* ipath_ruc_loopback - handle UC and RC lookback requests * ipath_ruc_loopback - handle UC and RC lookback requests
* @sqp: the loopback QP * @sqp: the loopback QP
* @wc: the work completion entry
* *
* This is called from ipath_do_uc_send() or ipath_do_rc_send() to * This is called from ipath_do_uc_send() or ipath_do_rc_send() to
* forward a WQE addressed to the same HCA. * forward a WQE addressed to the same HCA.
@ -197,13 +197,14 @@ int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
* receive interrupts since this is a connected protocol and all packets * receive interrupts since this is a connected protocol and all packets
* will pass through here. * will pass through here.
*/ */
void ipath_ruc_loopback(struct ipath_qp *sqp, struct ib_wc *wc) static void ipath_ruc_loopback(struct ipath_qp *sqp)
{ {
struct ipath_ibdev *dev = to_idev(sqp->ibqp.device); struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
struct ipath_qp *qp; struct ipath_qp *qp;
struct ipath_swqe *wqe; struct ipath_swqe *wqe;
struct ipath_sge *sge; struct ipath_sge *sge;
unsigned long flags; unsigned long flags;
struct ib_wc wc;
u64 sdata; u64 sdata;
qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn); qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
@ -234,8 +235,8 @@ void ipath_ruc_loopback(struct ipath_qp *sqp, struct ib_wc *wc)
wqe = get_swqe_ptr(sqp, sqp->s_last); wqe = get_swqe_ptr(sqp, sqp->s_last);
spin_unlock_irqrestore(&sqp->s_lock, flags); spin_unlock_irqrestore(&sqp->s_lock, flags);
wc->wc_flags = 0; wc.wc_flags = 0;
wc->imm_data = 0; wc.imm_data = 0;
sqp->s_sge.sge = wqe->sg_list[0]; sqp->s_sge.sge = wqe->sg_list[0];
sqp->s_sge.sg_list = wqe->sg_list + 1; sqp->s_sge.sg_list = wqe->sg_list + 1;
@ -243,8 +244,8 @@ void ipath_ruc_loopback(struct ipath_qp *sqp, struct ib_wc *wc)
sqp->s_len = wqe->length; sqp->s_len = wqe->length;
switch (wqe->wr.opcode) { switch (wqe->wr.opcode) {
case IB_WR_SEND_WITH_IMM: case IB_WR_SEND_WITH_IMM:
wc->wc_flags = IB_WC_WITH_IMM; wc.wc_flags = IB_WC_WITH_IMM;
wc->imm_data = wqe->wr.imm_data; wc.imm_data = wqe->wr.imm_data;
/* FALLTHROUGH */ /* FALLTHROUGH */
case IB_WR_SEND: case IB_WR_SEND:
spin_lock_irqsave(&qp->r_rq.lock, flags); spin_lock_irqsave(&qp->r_rq.lock, flags);
@ -255,7 +256,7 @@ void ipath_ruc_loopback(struct ipath_qp *sqp, struct ib_wc *wc)
if (qp->ibqp.qp_type == IB_QPT_UC) if (qp->ibqp.qp_type == IB_QPT_UC)
goto send_comp; goto send_comp;
if (sqp->s_rnr_retry == 0) { if (sqp->s_rnr_retry == 0) {
wc->status = IB_WC_RNR_RETRY_EXC_ERR; wc.status = IB_WC_RNR_RETRY_EXC_ERR;
goto err; goto err;
} }
if (sqp->s_rnr_retry_cnt < 7) if (sqp->s_rnr_retry_cnt < 7)
@ -270,8 +271,8 @@ void ipath_ruc_loopback(struct ipath_qp *sqp, struct ib_wc *wc)
break; break;
case IB_WR_RDMA_WRITE_WITH_IMM: case IB_WR_RDMA_WRITE_WITH_IMM:
wc->wc_flags = IB_WC_WITH_IMM; wc.wc_flags = IB_WC_WITH_IMM;
wc->imm_data = wqe->wr.imm_data; wc.imm_data = wqe->wr.imm_data;
spin_lock_irqsave(&qp->r_rq.lock, flags); spin_lock_irqsave(&qp->r_rq.lock, flags);
if (!ipath_get_rwqe(qp, 1)) if (!ipath_get_rwqe(qp, 1))
goto rnr_nak; goto rnr_nak;
@ -285,20 +286,20 @@ void ipath_ruc_loopback(struct ipath_qp *sqp, struct ib_wc *wc)
wqe->wr.wr.rdma.rkey, wqe->wr.wr.rdma.rkey,
IB_ACCESS_REMOTE_WRITE))) { IB_ACCESS_REMOTE_WRITE))) {
acc_err: acc_err:
wc->status = IB_WC_REM_ACCESS_ERR; wc.status = IB_WC_REM_ACCESS_ERR;
err: err:
wc->wr_id = wqe->wr.wr_id; wc.wr_id = wqe->wr.wr_id;
wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode]; wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
wc->vendor_err = 0; wc.vendor_err = 0;
wc->byte_len = 0; wc.byte_len = 0;
wc->qp_num = sqp->ibqp.qp_num; wc.qp_num = sqp->ibqp.qp_num;
wc->src_qp = sqp->remote_qpn; wc.src_qp = sqp->remote_qpn;
wc->pkey_index = 0; wc.pkey_index = 0;
wc->slid = sqp->remote_ah_attr.dlid; wc.slid = sqp->remote_ah_attr.dlid;
wc->sl = sqp->remote_ah_attr.sl; wc.sl = sqp->remote_ah_attr.sl;
wc->dlid_path_bits = 0; wc.dlid_path_bits = 0;
wc->port_num = 0; wc.port_num = 0;
ipath_sqerror_qp(sqp, wc); ipath_sqerror_qp(sqp, &wc);
goto done; goto done;
} }
break; break;
@ -374,22 +375,22 @@ void ipath_ruc_loopback(struct ipath_qp *sqp, struct ib_wc *wc)
goto send_comp; goto send_comp;
if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM) if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
wc->opcode = IB_WC_RECV_RDMA_WITH_IMM; wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
else else
wc->opcode = IB_WC_RECV; wc.opcode = IB_WC_RECV;
wc->wr_id = qp->r_wr_id; wc.wr_id = qp->r_wr_id;
wc->status = IB_WC_SUCCESS; wc.status = IB_WC_SUCCESS;
wc->vendor_err = 0; wc.vendor_err = 0;
wc->byte_len = wqe->length; wc.byte_len = wqe->length;
wc->qp_num = qp->ibqp.qp_num; wc.qp_num = qp->ibqp.qp_num;
wc->src_qp = qp->remote_qpn; wc.src_qp = qp->remote_qpn;
/* XXX do we know which pkey matched? Only needed for GSI. */ /* XXX do we know which pkey matched? Only needed for GSI. */
wc->pkey_index = 0; wc.pkey_index = 0;
wc->slid = qp->remote_ah_attr.dlid; wc.slid = qp->remote_ah_attr.dlid;
wc->sl = qp->remote_ah_attr.sl; wc.sl = qp->remote_ah_attr.sl;
wc->dlid_path_bits = 0; wc.dlid_path_bits = 0;
/* Signal completion event if the solicited bit is set. */ /* Signal completion event if the solicited bit is set. */
ipath_cq_enter(to_icq(qp->ibqp.recv_cq), wc, ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
wqe->wr.send_flags & IB_SEND_SOLICITED); wqe->wr.send_flags & IB_SEND_SOLICITED);
send_comp: send_comp:
@ -397,19 +398,19 @@ void ipath_ruc_loopback(struct ipath_qp *sqp, struct ib_wc *wc)
if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &sqp->s_flags) || if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &sqp->s_flags) ||
(wqe->wr.send_flags & IB_SEND_SIGNALED)) { (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
wc->wr_id = wqe->wr.wr_id; wc.wr_id = wqe->wr.wr_id;
wc->status = IB_WC_SUCCESS; wc.status = IB_WC_SUCCESS;
wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode]; wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
wc->vendor_err = 0; wc.vendor_err = 0;
wc->byte_len = wqe->length; wc.byte_len = wqe->length;
wc->qp_num = sqp->ibqp.qp_num; wc.qp_num = sqp->ibqp.qp_num;
wc->src_qp = 0; wc.src_qp = 0;
wc->pkey_index = 0; wc.pkey_index = 0;
wc->slid = 0; wc.slid = 0;
wc->sl = 0; wc.sl = 0;
wc->dlid_path_bits = 0; wc.dlid_path_bits = 0;
wc->port_num = 0; wc.port_num = 0;
ipath_cq_enter(to_icq(sqp->ibqp.send_cq), wc, 0); ipath_cq_enter(to_icq(sqp->ibqp.send_cq), &wc, 0);
} }
/* Update s_last now that we are finished with the SWQE */ /* Update s_last now that we are finished with the SWQE */
@ -455,11 +456,11 @@ void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev)
} }
/** /**
* ipath_post_rc_send - post RC and UC sends * ipath_post_ruc_send - post RC and UC sends
* @qp: the QP to post on * @qp: the QP to post on
* @wr: the work request to send * @wr: the work request to send
*/ */
int ipath_post_rc_send(struct ipath_qp *qp, struct ib_send_wr *wr) int ipath_post_ruc_send(struct ipath_qp *qp, struct ib_send_wr *wr)
{ {
struct ipath_swqe *wqe; struct ipath_swqe *wqe;
unsigned long flags; unsigned long flags;
@ -534,13 +535,149 @@ int ipath_post_rc_send(struct ipath_qp *qp, struct ib_send_wr *wr)
qp->s_head = next; qp->s_head = next;
spin_unlock_irqrestore(&qp->s_lock, flags); spin_unlock_irqrestore(&qp->s_lock, flags);
if (qp->ibqp.qp_type == IB_QPT_UC) ipath_do_ruc_send((unsigned long) qp);
ipath_do_uc_send((unsigned long) qp);
else
ipath_do_rc_send((unsigned long) qp);
ret = 0; ret = 0;
bail: bail:
return ret; return ret;
} }
/**
* ipath_make_grh - construct a GRH header
* @dev: a pointer to the ipath device
* @hdr: a pointer to the GRH header being constructed
* @grh: the global route address to send to
* @hwords: the number of 32 bit words of header being sent
* @nwords: the number of 32 bit words of data being sent
*
* Return the size of the header in 32 bit words.
*/
u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
struct ib_global_route *grh, u32 hwords, u32 nwords)
{
hdr->version_tclass_flow =
cpu_to_be32((6 << 28) |
(grh->traffic_class << 20) |
grh->flow_label);
hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
/* next_hdr is defined by C8-7 in ch. 8.4.1 */
hdr->next_hdr = 0x1B;
hdr->hop_limit = grh->hop_limit;
/* The SGID is 32-bit aligned. */
hdr->sgid.global.subnet_prefix = dev->gid_prefix;
hdr->sgid.global.interface_id = ipath_layer_get_guid(dev->dd);
hdr->dgid = grh->dgid;
/* GRH header size in 32-bit words. */
return sizeof(struct ib_grh) / sizeof(u32);
}
/**
* ipath_do_ruc_send - perform a send on an RC or UC QP
* @data: contains a pointer to the QP
*
* Process entries in the send work queue until credit or queue is
* exhausted. Only allow one CPU to send a packet per QP (tasklet).
* Otherwise, after we drop the QP s_lock, two threads could send
* packets out of order.
*/
void ipath_do_ruc_send(unsigned long data)
{
struct ipath_qp *qp = (struct ipath_qp *)data;
struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
unsigned long flags;
u16 lrh0;
u32 nwords;
u32 extra_bytes;
u32 bth0;
u32 bth2;
u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
struct ipath_other_headers *ohdr;
if (test_and_set_bit(IPATH_S_BUSY, &qp->s_flags))
goto bail;
if (unlikely(qp->remote_ah_attr.dlid ==
ipath_layer_get_lid(dev->dd))) {
ipath_ruc_loopback(qp);
goto clear;
}
ohdr = &qp->s_hdr.u.oth;
if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
ohdr = &qp->s_hdr.u.l.oth;
again:
/* Check for a constructed packet to be sent. */
if (qp->s_hdrwords != 0) {
/*
* If no PIO bufs are available, return. An interrupt will
* call ipath_ib_piobufavail() when one is available.
*/
if (ipath_verbs_send(dev->dd, qp->s_hdrwords,
(u32 *) &qp->s_hdr, qp->s_cur_size,
qp->s_cur_sge)) {
ipath_no_bufs_available(qp, dev);
goto bail;
}
dev->n_unicast_xmit++;
/* Record that we sent the packet and s_hdr is empty. */
qp->s_hdrwords = 0;
}
/*
* The lock is needed to synchronize between setting
* qp->s_ack_state, resend timer, and post_send().
*/
spin_lock_irqsave(&qp->s_lock, flags);
/* Sending responses has higher priority over sending requests. */
if (qp->s_ack_state != IB_OPCODE_RC_ACKNOWLEDGE &&
(bth0 = ipath_make_rc_ack(qp, ohdr, pmtu)) != 0)
bth2 = qp->s_ack_psn++ & IPS_PSN_MASK;
else if (!((qp->ibqp.qp_type == IB_QPT_RC) ?
ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2) :
ipath_make_uc_req(qp, ohdr, pmtu, &bth0, &bth2))) {
/*
* Clear the busy bit before unlocking to avoid races with
* adding new work queue items and then failing to process
* them.
*/
clear_bit(IPATH_S_BUSY, &qp->s_flags);
spin_unlock_irqrestore(&qp->s_lock, flags);
goto bail;
}
spin_unlock_irqrestore(&qp->s_lock, flags);
/* Construct the header. */
extra_bytes = (4 - qp->s_cur_size) & 3;
nwords = (qp->s_cur_size + extra_bytes) >> 2;
lrh0 = IPS_LRH_BTH;
if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
&qp->remote_ah_attr.grh,
qp->s_hdrwords, nwords);
lrh0 = IPS_LRH_GRH;
}
lrh0 |= qp->remote_ah_attr.sl << 4;
qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
SIZE_OF_CRC);
qp->s_hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd));
bth0 |= ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
bth0 |= extra_bytes << 20;
ohdr->bth[0] = cpu_to_be32(bth0);
ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
ohdr->bth[2] = cpu_to_be32(bth2);
/* Check for more work to do. */
goto again;
clear:
clear_bit(IPATH_S_BUSY, &qp->s_flags);
bail:
return;
}

153
drivers/infiniband/hw/ipath/ipath_uc.c
View File

@ -62,90 +62,40 @@ static void complete_last_send(struct ipath_qp *qp, struct ipath_swqe *wqe,
} }
/** /**
* ipath_do_uc_send - do a send on a UC queue * ipath_make_uc_req - construct a request packet (SEND, RDMA write)
* @data: contains a pointer to the QP to send on * @qp: a pointer to the QP
* @ohdr: a pointer to the IB header being constructed
* @pmtu: the path MTU
* @bth0p: pointer to the BTH opcode word
* @bth2p: pointer to the BTH PSN word
* *
* Process entries in the send work queue until the queue is exhausted. * Return 1 if constructed; otherwise, return 0.
* Only allow one CPU to send a packet per QP (tasklet). * Note the QP s_lock must be held and interrupts disabled.
* Otherwise, after we drop the QP lock, two threads could send
* packets out of order.
* This is similar to ipath_do_rc_send() below except we don't have
* timeouts or resends.
*/ */
void ipath_do_uc_send(unsigned long data) int ipath_make_uc_req(struct ipath_qp *qp,
struct ipath_other_headers *ohdr,
u32 pmtu, u32 *bth0p, u32 *bth2p)
{ {
struct ipath_qp *qp = (struct ipath_qp *)data;
struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
struct ipath_swqe *wqe; struct ipath_swqe *wqe;
unsigned long flags;
u16 lrh0;
u32 hwords; u32 hwords;
u32 nwords;
u32 extra_bytes;
u32 bth0; u32 bth0;
u32 bth2;
u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
u32 len; u32 len;
struct ipath_other_headers *ohdr;
struct ib_wc wc; struct ib_wc wc;
if (test_and_set_bit(IPATH_S_BUSY, &qp->s_flags))
goto bail;
if (unlikely(qp->remote_ah_attr.dlid ==
ipath_layer_get_lid(dev->dd))) {
/* Pass in an uninitialized ib_wc to save stack space. */
ipath_ruc_loopback(qp, &wc);
clear_bit(IPATH_S_BUSY, &qp->s_flags);
goto bail;
}
ohdr = &qp->s_hdr.u.oth;
if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
ohdr = &qp->s_hdr.u.l.oth;
again:
/* Check for a constructed packet to be sent. */
if (qp->s_hdrwords != 0) {
/*
* If no PIO bufs are available, return.
* An interrupt will call ipath_ib_piobufavail()
* when one is available.
*/
if (ipath_verbs_send(dev->dd, qp->s_hdrwords,
(u32 *) &qp->s_hdr,
qp->s_cur_size,
qp->s_cur_sge)) {
ipath_no_bufs_available(qp, dev);
goto bail;
}
dev->n_unicast_xmit++;
/* Record that we sent the packet and s_hdr is empty. */
qp->s_hdrwords = 0;
}
lrh0 = IPS_LRH_BTH;
/* header size in 32-bit words LRH+BTH = (8+12)/4. */
hwords = 5;
/*
* The lock is needed to synchronize between
* setting qp->s_ack_state and post_send().
*/
spin_lock_irqsave(&qp->s_lock, flags);
if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)) if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK))
goto done; goto done;
bth0 = ipath_layer_get_pkey(dev->dd, qp->s_pkey_index); /* header size in 32-bit words LRH+BTH = (8+12)/4. */
hwords = 5;
bth0 = 0;
/* Send a request. */ /* Get the next send request. */
wqe = get_swqe_ptr(qp, qp->s_last); wqe = get_swqe_ptr(qp, qp->s_last);
switch (qp->s_state) { switch (qp->s_state) {
default: default:
/* /*
* Signal the completion of the last send (if there is * Signal the completion of the last send
* one). * (if there is one).
*/ */
if (qp->s_last != qp->s_tail) if (qp->s_last != qp->s_tail)
complete_last_send(qp, wqe, &wc); complete_last_send(qp, wqe, &wc);
@ -258,61 +208,16 @@ void ipath_do_uc_send(unsigned long data)
} }
break; break;
} }
bth2 = qp->s_next_psn++ & IPS_PSN_MASK;
qp->s_len -= len; qp->s_len -= len;
bth0 |= qp->s_state << 24;
spin_unlock_irqrestore(&qp->s_lock, flags);
/* Construct the header. */
extra_bytes = (4 - len) & 3;
nwords = (len + extra_bytes) >> 2;
if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
/* Header size in 32-bit words. */
hwords += 10;
lrh0 = IPS_LRH_GRH;
qp->s_hdr.u.l.grh.version_tclass_flow =
cpu_to_be32((6 << 28) |
(qp->remote_ah_attr.grh.traffic_class
<< 20) |
qp->remote_ah_attr.grh.flow_label);
qp->s_hdr.u.l.grh.paylen =
cpu_to_be16(((hwords - 12) + nwords +
SIZE_OF_CRC) << 2);
/* next_hdr is defined by C8-7 in ch. 8.4.1 */
qp->s_hdr.u.l.grh.next_hdr = 0x1B;
qp->s_hdr.u.l.grh.hop_limit =
qp->remote_ah_attr.grh.hop_limit;
/* The SGID is 32-bit aligned. */
qp->s_hdr.u.l.grh.sgid.global.subnet_prefix =
dev->gid_prefix;
qp->s_hdr.u.l.grh.sgid.global.interface_id =
ipath_layer_get_guid(dev->dd);
qp->s_hdr.u.l.grh.dgid = qp->remote_ah_attr.grh.dgid;
}
qp->s_hdrwords = hwords; qp->s_hdrwords = hwords;
qp->s_cur_sge = &qp->s_sge; qp->s_cur_sge = &qp->s_sge;
qp->s_cur_size = len; qp->s_cur_size = len;
lrh0 |= qp->remote_ah_attr.sl << 4; *bth0p = bth0 | (qp->s_state << 24);
qp->s_hdr.lrh[0] = cpu_to_be16(lrh0); *bth2p = qp->s_next_psn++ & IPS_PSN_MASK;
/* DEST LID */ return 1;
qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
qp->s_hdr.lrh[2] = cpu_to_be16(hwords + nwords + SIZE_OF_CRC);
qp->s_hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd));
bth0 |= extra_bytes << 20;
ohdr->bth[0] = cpu_to_be32(bth0);
ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
ohdr->bth[2] = cpu_to_be32(bth2);
/* Check for more work to do. */
goto again;
done: done:
spin_unlock_irqrestore(&qp->s_lock, flags); return 0;
clear_bit(IPATH_S_BUSY, &qp->s_flags);
bail:
return;
} }
/** /**
@ -536,12 +441,13 @@ void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
if (qp->r_len != 0) { if (qp->r_len != 0) {
u32 rkey = be32_to_cpu(reth->rkey); u32 rkey = be32_to_cpu(reth->rkey);
u64 vaddr = be64_to_cpu(reth->vaddr); u64 vaddr = be64_to_cpu(reth->vaddr);
int ok;
/* Check rkey */ /* Check rkey */
if (unlikely(!ipath_rkey_ok( ok = ipath_rkey_ok(dev, &qp->r_sge, qp->r_len,
dev, &qp->r_sge, qp->r_len, vaddr, rkey,
vaddr, rkey, IB_ACCESS_REMOTE_WRITE);
IB_ACCESS_REMOTE_WRITE))) { if (unlikely(!ok)) {
dev->n_pkt_drops++; dev->n_pkt_drops++;
goto done; goto done;
} }
@ -559,8 +465,7 @@ void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
} }
if (opcode == OP(RDMA_WRITE_ONLY)) if (opcode == OP(RDMA_WRITE_ONLY))
goto rdma_last; goto rdma_last;
else if (opcode == else if (opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
goto rdma_last_imm; goto rdma_last_imm;
/* FALLTHROUGH */ /* FALLTHROUGH */
case OP(RDMA_WRITE_MIDDLE): case OP(RDMA_WRITE_MIDDLE):
@ -593,9 +498,9 @@ void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
dev->n_pkt_drops++; dev->n_pkt_drops++;
goto done; goto done;
} }
if (qp->r_reuse_sge) { if (qp->r_reuse_sge)
qp->r_reuse_sge = 0; qp->r_reuse_sge = 0;
} else if (!ipath_get_rwqe(qp, 1)) { else if (!ipath_get_rwqe(qp, 1)) {
dev->n_pkt_drops++; dev->n_pkt_drops++;
goto done; goto done;
} }

2
drivers/infiniband/hw/ipath/ipath_verbs.c
View File

@ -194,7 +194,7 @@ static int ipath_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
switch (qp->ibqp.qp_type) { switch (qp->ibqp.qp_type) {
case IB_QPT_UC: case IB_QPT_UC:
case IB_QPT_RC: case IB_QPT_RC:
err = ipath_post_rc_send(qp, wr); err = ipath_post_ruc_send(qp, wr);
break; break;
case IB_QPT_SMI: case IB_QPT_SMI:

20
drivers/infiniband/hw/ipath/ipath_verbs.h
View File

@ -581,10 +581,6 @@ void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc);
void ipath_get_credit(struct ipath_qp *qp, u32 aeth); void ipath_get_credit(struct ipath_qp *qp, u32 aeth);
void ipath_do_rc_send(unsigned long data);
void ipath_do_uc_send(unsigned long data);
void ipath_cq_enter(struct ipath_cq *cq, struct ib_wc *entry, int sig); void ipath_cq_enter(struct ipath_cq *cq, struct ib_wc *entry, int sig);
int ipath_rkey_ok(struct ipath_ibdev *dev, struct ipath_sge_state *ss, int ipath_rkey_ok(struct ipath_ibdev *dev, struct ipath_sge_state *ss,
@ -597,7 +593,7 @@ void ipath_copy_sge(struct ipath_sge_state *ss, void *data, u32 length);
void ipath_skip_sge(struct ipath_sge_state *ss, u32 length); void ipath_skip_sge(struct ipath_sge_state *ss, u32 length);
int ipath_post_rc_send(struct ipath_qp *qp, struct ib_send_wr *wr); int ipath_post_ruc_send(struct ipath_qp *qp, struct ib_send_wr *wr);
void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr, void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
int has_grh, void *data, u32 tlen, struct ipath_qp *qp); int has_grh, void *data, u32 tlen, struct ipath_qp *qp);
@ -679,7 +675,19 @@ void ipath_insert_rnr_queue(struct ipath_qp *qp);
int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only); int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only);
void ipath_ruc_loopback(struct ipath_qp *sqp, struct ib_wc *wc); u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
struct ib_global_route *grh, u32 hwords, u32 nwords);
void ipath_do_ruc_send(unsigned long data);
u32 ipath_make_rc_ack(struct ipath_qp *qp, struct ipath_other_headers *ohdr,
u32 pmtu);
int ipath_make_rc_req(struct ipath_qp *qp, struct ipath_other_headers *ohdr,
u32 pmtu, u32 *bth0p, u32 *bth2p);
int ipath_make_uc_req(struct ipath_qp *qp, struct ipath_other_headers *ohdr,
u32 pmtu, u32 *bth0p, u32 *bth2p);
extern const enum ib_wc_opcode ib_ipath_wc_opcode[]; extern const enum ib_wc_opcode ib_ipath_wc_opcode[];