diff options
Diffstat (limited to 'drivers/infiniband/hw/qib/qib_verbs.c')
| -rw-r--r-- | drivers/infiniband/hw/qib/qib_verbs.c | 2248 |
1 files changed, 2248 insertions, 0 deletions
diff --git a/drivers/infiniband/hw/qib/qib_verbs.c b/drivers/infiniband/hw/qib/qib_verbs.c new file mode 100644 index 00000000000..cda8f4173d2 --- /dev/null +++ b/drivers/infiniband/hw/qib/qib_verbs.c @@ -0,0 +1,2248 @@ +/* + * Copyright (c) 2006, 2007, 2008, 2009, 2010 QLogic Corporation. + * All rights reserved. + * Copyright (c) 2005, 2006 PathScale, Inc. 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. + */ + +#include <rdma/ib_mad.h> +#include <rdma/ib_user_verbs.h> +#include <linux/io.h> +#include <linux/utsname.h> +#include <linux/rculist.h> +#include <linux/mm.h> + +#include "qib.h" +#include "qib_common.h" + +static unsigned int ib_qib_qp_table_size = 251; +module_param_named(qp_table_size, ib_qib_qp_table_size, uint, S_IRUGO); +MODULE_PARM_DESC(qp_table_size, "QP table size"); + +unsigned int ib_qib_lkey_table_size = 16; +module_param_named(lkey_table_size, ib_qib_lkey_table_size, uint, + S_IRUGO); +MODULE_PARM_DESC(lkey_table_size, + "LKEY table size in bits (2^n, 1 <= n <= 23)"); + +static unsigned int ib_qib_max_pds = 0xFFFF; +module_param_named(max_pds, ib_qib_max_pds, uint, S_IRUGO); +MODULE_PARM_DESC(max_pds, + "Maximum number of protection domains to support"); + +static unsigned int ib_qib_max_ahs = 0xFFFF; +module_param_named(max_ahs, ib_qib_max_ahs, uint, S_IRUGO); +MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support"); + +unsigned int ib_qib_max_cqes = 0x2FFFF; +module_param_named(max_cqes, ib_qib_max_cqes, uint, S_IRUGO); +MODULE_PARM_DESC(max_cqes, + "Maximum number of completion queue entries to support"); + +unsigned int ib_qib_max_cqs = 0x1FFFF; +module_param_named(max_cqs, ib_qib_max_cqs, uint, S_IRUGO); +MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support"); + +unsigned int ib_qib_max_qp_wrs = 0x3FFF; +module_param_named(max_qp_wrs, ib_qib_max_qp_wrs, uint, S_IRUGO); +MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support"); + +unsigned int ib_qib_max_qps = 16384; +module_param_named(max_qps, ib_qib_max_qps, uint, S_IRUGO); +MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support"); + +unsigned int ib_qib_max_sges = 0x60; +module_param_named(max_sges, ib_qib_max_sges, uint, S_IRUGO); +MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support"); + +unsigned int ib_qib_max_mcast_grps = 16384; +module_param_named(max_mcast_grps, ib_qib_max_mcast_grps, uint, S_IRUGO); +MODULE_PARM_DESC(max_mcast_grps, + "Maximum number of multicast groups to support"); + +unsigned int ib_qib_max_mcast_qp_attached = 16; +module_param_named(max_mcast_qp_attached, ib_qib_max_mcast_qp_attached, + uint, S_IRUGO); +MODULE_PARM_DESC(max_mcast_qp_attached, + "Maximum number of attached QPs to support"); + +unsigned int ib_qib_max_srqs = 1024; +module_param_named(max_srqs, ib_qib_max_srqs, uint, S_IRUGO); +MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support"); + +unsigned int ib_qib_max_srq_sges = 128; +module_param_named(max_srq_sges, ib_qib_max_srq_sges, uint, S_IRUGO); +MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support"); + +unsigned int ib_qib_max_srq_wrs = 0x1FFFF; +module_param_named(max_srq_wrs, ib_qib_max_srq_wrs, uint, S_IRUGO); +MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support"); + +static unsigned int ib_qib_disable_sma; +module_param_named(disable_sma, ib_qib_disable_sma, uint, S_IWUSR | S_IRUGO); +MODULE_PARM_DESC(disable_sma, "Disable the SMA"); + +/* + * Note that it is OK to post send work requests in the SQE and ERR + * states; qib_do_send() will process them and generate error + * completions as per IB 1.2 C10-96. + */ +const int ib_qib_state_ops[IB_QPS_ERR + 1] = { + [IB_QPS_RESET] = 0, + [IB_QPS_INIT] = QIB_POST_RECV_OK, + [IB_QPS_RTR] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK, + [IB_QPS_RTS] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK | + QIB_POST_SEND_OK | QIB_PROCESS_SEND_OK | + QIB_PROCESS_NEXT_SEND_OK, + [IB_QPS_SQD] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK | + QIB_POST_SEND_OK | QIB_PROCESS_SEND_OK, + [IB_QPS_SQE] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK | + QIB_POST_SEND_OK | QIB_FLUSH_SEND, + [IB_QPS_ERR] = QIB_POST_RECV_OK | QIB_FLUSH_RECV | + QIB_POST_SEND_OK | QIB_FLUSH_SEND, +}; + +struct qib_ucontext { + struct ib_ucontext ibucontext; +}; + +static inline struct qib_ucontext *to_iucontext(struct ib_ucontext + *ibucontext) +{ + return container_of(ibucontext, struct qib_ucontext, ibucontext); +} + +/* + * Translate ib_wr_opcode into ib_wc_opcode. + */ +const enum ib_wc_opcode ib_qib_wc_opcode[] = { + [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE, + [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE, + [IB_WR_SEND] = IB_WC_SEND, + [IB_WR_SEND_WITH_IMM] = IB_WC_SEND, + [IB_WR_RDMA_READ] = IB_WC_RDMA_READ, + [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP, + [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD +}; + +/* + * System image GUID. + */ +__be64 ib_qib_sys_image_guid; + +/** + * qib_copy_sge - copy data to SGE memory + * @ss: the SGE state + * @data: the data to copy + * @length: the length of the data + */ +void qib_copy_sge(struct qib_sge_state *ss, void *data, u32 length, int release) +{ + struct qib_sge *sge = &ss->sge; + + while (length) { + u32 len = sge->length; + + if (len > length) + len = length; + if (len > sge->sge_length) + len = sge->sge_length; + BUG_ON(len == 0); + memcpy(sge->vaddr, data, len); + sge->vaddr += len; + sge->length -= len; + sge->sge_length -= len; + if (sge->sge_length == 0) { + if (release) + atomic_dec(&sge->mr->refcount); + if (--ss->num_sge) + *sge = *ss->sg_list++; + } else if (sge->length == 0 && sge->mr->lkey) { + if (++sge->n >= QIB_SEGSZ) { + if (++sge->m >= sge->mr->mapsz) + break; + sge->n = 0; + } + sge->vaddr = + sge->mr->map[sge->m]->segs[sge->n].vaddr; + sge->length = + sge->mr->map[sge->m]->segs[sge->n].length; + } + data += len; + length -= len; + } +} + +/** + * qib_skip_sge - skip over SGE memory - XXX almost dup of prev func + * @ss: the SGE state + * @length: the number of bytes to skip + */ +void qib_skip_sge(struct qib_sge_state *ss, u32 length, int release) +{ + struct qib_sge *sge = &ss->sge; + + while (length) { + u32 len = sge->length; + + if (len > length) + len = length; + if (len > sge->sge_length) + len = sge->sge_length; + BUG_ON(len == 0); + sge->vaddr += len; + sge->length -= len; + sge->sge_length -= len; + if (sge->sge_length == 0) { + if (release) + atomic_dec(&sge->mr->refcount); + if (--ss->num_sge) + *sge = *ss->sg_list++; + } else if (sge->length == 0 && sge->mr->lkey) { + if (++sge->n >= QIB_SEGSZ) { + if (++sge->m >= sge->mr->mapsz) + break; + sge->n = 0; + } + sge->vaddr = + sge->mr->map[sge->m]->segs[sge->n].vaddr; + sge->length = + sge->mr->map[sge->m]->segs[sge->n].length; + } + length -= len; + } +} + +/* + * Count the number of DMA descriptors needed to send length bytes of data. + * Don't modify the qib_sge_state to get the count. + * Return zero if any of the segments is not aligned. + */ +static u32 qib_count_sge(struct qib_sge_state *ss, u32 length) +{ + struct qib_sge *sg_list = ss->sg_list; + struct qib_sge sge = ss->sge; + u8 num_sge = ss->num_sge; + u32 ndesc = 1; /* count the header */ + + while (length) { + u32 len = sge.length; + + if (len > length) + len = length; + if (len > sge.sge_length) + len = sge.sge_length; + BUG_ON(len == 0); + if (((long) sge.vaddr & (sizeof(u32) - 1)) || + (len != length && (len & (sizeof(u32) - 1)))) { + ndesc = 0; + break; + } + ndesc++; + sge.vaddr += len; + sge.length -= len; + sge.sge_length -= len; + if (sge.sge_length == 0) { + if (--num_sge) + sge = *sg_list++; + } else if (sge.length == 0 && sge.mr->lkey) { + if (++sge.n >= QIB_SEGSZ) { + if (++sge.m >= sge.mr->mapsz) + break; + sge.n = 0; + } + sge.vaddr = + sge.mr->map[sge.m]->segs[sge.n].vaddr; + sge.length = + sge.mr->map[sge.m]->segs[sge.n].length; + } + length -= len; + } + return ndesc; +} + +/* + * Copy from the SGEs to the data buffer. + */ +static void qib_copy_from_sge(void *data, struct qib_sge_state *ss, u32 length) +{ + struct qib_sge *sge = &ss->sge; + + while (length) { + u32 len = sge->length; + + if (len > length) + len = length; + if (len > sge->sge_length) + len = sge->sge_length; + BUG_ON(len == 0); + memcpy(data, sge->vaddr, len); + sge->vaddr += len; + sge->length -= len; + sge->sge_length -= len; + if (sge->sge_length == 0) { + if (--ss->num_sge) + *sge = *ss->sg_list++; + } else if (sge->length == 0 && sge->mr->lkey) { + if (++sge->n >= QIB_SEGSZ) { + if (++sge->m >= sge->mr->mapsz) + break; + sge->n = 0; + } + sge->vaddr = + sge->mr->map[sge->m]->segs[sge->n].vaddr; + sge->length = + sge->mr->map[sge->m]->segs[sge->n].length; + } + data += len; + length -= len; + } +} + +/** + * qib_post_one_send - post one RC, UC, or UD send work request + * @qp: the QP to post on + * @wr: the work request to send + */ +static int qib_post_one_send(struct qib_qp *qp, struct ib_send_wr *wr) +{ + struct qib_swqe *wqe; + u32 next; + int i; + int j; + int acc; + int ret; + unsigned long flags; + struct qib_lkey_table *rkt; + struct qib_pd *pd; + + spin_lock_irqsave(&qp->s_lock, flags); + + /* Check that state is OK to post send. */ + if (unlikely(!(ib_qib_state_ops[qp->state] & QIB_POST_SEND_OK))) + goto bail_inval; + + /* IB spec says that num_sge == 0 is OK. */ + if (wr->num_sge > qp->s_max_sge) + goto bail_inval; + + /* + * Don't allow RDMA reads or atomic operations on UC or + * undefined operations. + * Make sure buffer is large enough to hold the result for atomics. + */ + if (wr->opcode == IB_WR_FAST_REG_MR) { + if (qib_fast_reg_mr(qp, wr)) + goto bail_inval; + } else if (qp->ibqp.qp_type == IB_QPT_UC) { + if ((unsigned) wr->opcode >= IB_WR_RDMA_READ) + goto bail_inval; + } else if (qp->ibqp.qp_type != IB_QPT_RC) { + /* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */ + if (wr->opcode != IB_WR_SEND && + wr->opcode != IB_WR_SEND_WITH_IMM) + goto bail_inval; + /* Check UD destination address PD */ + if (qp->ibqp.pd != wr->wr.ud.ah->pd) + goto bail_inval; + } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) + goto bail_inval; + else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP && + (wr->num_sge == 0 || + wr->sg_list[0].length < sizeof(u64) || + wr->sg_list[0].addr & (sizeof(u64) - 1))) + goto bail_inval; + else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) + goto bail_inval; + + next = qp->s_head + 1; + if (next >= qp->s_size) + next = 0; + if (next == qp->s_last) { + ret = -ENOMEM; + goto bail; + } + + rkt = &to_idev(qp->ibqp.device)->lk_table; + pd = to_ipd(qp->ibqp.pd); + wqe = get_swqe_ptr(qp, qp->s_head); + wqe->wr = *wr; + wqe->length = 0; + j = 0; + if (wr->num_sge) { + acc = wr->opcode >= IB_WR_RDMA_READ ? + IB_ACCESS_LOCAL_WRITE : 0; + for (i = 0; i < wr->num_sge; i++) { + u32 length = wr->sg_list[i].length; + int ok; + + if (length == 0) + continue; + ok = qib_lkey_ok(rkt, pd, &wqe->sg_list[j], + &wr->sg_list[i], acc); + if (!ok) + goto bail_inval_free; + wqe->length += length; + j++; + } + wqe->wr.num_sge = j; + } + if (qp->ibqp.qp_type == IB_QPT_UC || + qp->ibqp.qp_type == IB_QPT_RC) { + if (wqe->length > 0x80000000U) + goto bail_inval_free; + } else if (wqe->length > (dd_from_ibdev(qp->ibqp.device)->pport + + qp->port_num - 1)->ibmtu) + goto bail_inval_free; + else + atomic_inc(&to_iah(wr->wr.ud.ah)->refcount); + wqe->ssn = qp->s_ssn++; + qp->s_head = next; + + ret = 0; + goto bail; + +bail_inval_free: + while (j) { + struct qib_sge *sge = &wqe->sg_list[--j]; + + atomic_dec(&sge->mr->refcount); + } +bail_inval: + ret = -EINVAL; +bail: + spin_unlock_irqrestore(&qp->s_lock, flags); + return ret; +} + +/** + * qib_post_send - post a send on a QP + * @ibqp: the QP to post the send on + * @wr: the list of work requests to post + * @bad_wr: the first bad WR is put here + * + * This may be called from interrupt context. + */ +static int qib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, + struct ib_send_wr **bad_wr) +{ + struct qib_qp *qp = to_iqp(ibqp); + int err = 0; + + for (; wr; wr = wr->next) { + err = qib_post_one_send(qp, wr); + if (err) { + *bad_wr = wr; + goto bail; + } + } + + /* Try to do the send work in the caller's context. */ + qib_do_send(&qp->s_work); + +bail: + return err; +} + +/** + * qib_post_receive - post a receive on a QP + * @ibqp: the QP to post the receive on + * @wr: the WR to post + * @bad_wr: the first bad WR is put here + * + * This may be called from interrupt context. + */ +static int qib_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr, + struct ib_recv_wr **bad_wr) +{ + struct qib_qp *qp = to_iqp(ibqp); + struct qib_rwq *wq = qp->r_rq.wq; + unsigned long flags; + int ret; + + /* Check that state is OK to post receive. */ + if (!(ib_qib_state_ops[qp->state] & QIB_POST_RECV_OK) || !wq) { + *bad_wr = wr; + ret = -EINVAL; + goto bail; + } + + for (; wr; wr = wr->next) { + struct qib_rwqe *wqe; + u32 next; + int i; + + if ((unsigned) wr->num_sge > qp->r_rq.max_sge) { + *bad_wr = wr; + ret = -EINVAL; + goto bail; + } + + spin_lock_irqsave(&qp->r_rq.lock, flags); + next = wq->head + 1; + if (next >= qp->r_rq.size) + next = 0; + if (next == wq->tail) { + spin_unlock_irqrestore(&qp->r_rq.lock, flags); + *bad_wr = wr; + ret = -ENOMEM; + goto bail; + } + + wqe = get_rwqe_ptr(&qp->r_rq, wq->head); + wqe->wr_id = wr->wr_id; + wqe->num_sge = wr->num_sge; + for (i = 0; i < wr->num_sge; i++) + wqe->sg_list[i] = wr->sg_list[i]; + /* Make sure queue entry is written before the head index. */ + smp_wmb(); + wq->head = next; + spin_unlock_irqrestore(&qp->r_rq.lock, flags); + } + ret = 0; + +bail: + return ret; +} + +/** + * qib_qp_rcv - processing an incoming packet on a QP + * @rcd: the context pointer + * @hdr: the packet header + * @has_grh: true if the packet has a GRH + * @data: the packet data + * @tlen: the packet length + * @qp: the QP the packet came on + * + * This is called from qib_ib_rcv() to process an incoming packet + * for the given QP. + * Called at interrupt level. + */ +static void qib_qp_rcv(struct qib_ctxtdata *rcd, struct qib_ib_header *hdr, + int has_grh, void *data, u32 tlen, struct qib_qp *qp) +{ + struct qib_ibport *ibp = &rcd->ppd->ibport_data; + + /* Check for valid receive state. */ + if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK)) { + ibp->n_pkt_drops++; + return; + } + + switch (qp->ibqp.qp_type) { + case IB_QPT_SMI: + case IB_QPT_GSI: + if (ib_qib_disable_sma) + break; + /* FALLTHROUGH */ + case IB_QPT_UD: + qib_ud_rcv(ibp, hdr, has_grh, data, tlen, qp); + break; + + case IB_QPT_RC: + qib_rc_rcv(rcd, hdr, has_grh, data, tlen, qp); + break; + + case IB_QPT_UC: + qib_uc_rcv(ibp, hdr, has_grh, data, tlen, qp); + break; + + default: + break; + } +} + +/** + * qib_ib_rcv - process an incoming packet + * @rcd: the context pointer + * @rhdr: the header of the packet + * @data: the packet payload + * @tlen: the packet length + * + * This is called from qib_kreceive() to process an incoming packet at + * interrupt level. Tlen is the length of the header + data + CRC in bytes. + */ +void qib_ib_rcv(struct qib_ctxtdata *rcd, void *rhdr, void *data, u32 tlen) +{ + struct qib_pportdata *ppd = rcd->ppd; + struct qib_ibport *ibp = &ppd->ibport_data; + struct qib_ib_header *hdr = rhdr; + struct qib_other_headers *ohdr; + struct qib_qp *qp; + u32 qp_num; + int lnh; + u8 opcode; + u16 lid; + + /* 24 == LRH+BTH+CRC */ + if (unlikely(tlen < 24)) + goto drop; + + /* Check for a valid destination LID (see ch. 7.11.1). */ + lid = be16_to_cpu(hdr->lrh[1]); + if (lid < QIB_MULTICAST_LID_BASE) { + lid &= ~((1 << ppd->lmc) - 1); + if (unlikely(lid != ppd->lid)) + goto drop; + } + + /* Check for GRH */ + lnh = be16_to_cpu(hdr->lrh[0]) & 3; + if (lnh == QIB_LRH_BTH) + ohdr = &hdr->u.oth; + else if (lnh == QIB_LRH_GRH) { + u32 vtf; + + ohdr = &hdr->u.l.oth; + if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR) + goto drop; + vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow); + if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION) + goto drop; + } else + goto drop; + + opcode = be32_to_cpu(ohdr->bth[0]) >> 24; + ibp->opstats[opcode & 0x7f].n_bytes += tlen; + ibp->opstats[opcode & 0x7f].n_packets++; + + /* Get the destination QP number. */ + qp_num = be32_to_cpu(ohdr->bth[1]) & QIB_QPN_MASK; + if (qp_num == QIB_MULTICAST_QPN) { + struct qib_mcast *mcast; + struct qib_mcast_qp *p; + + if (lnh != QIB_LRH_GRH) + goto drop; + mcast = qib_mcast_find(ibp, &hdr->u.l.grh.dgid); + if (mcast == NULL) + goto drop; + ibp->n_multicast_rcv++; + list_for_each_entry_rcu(p, &mcast->qp_list, list) + qib_qp_rcv(rcd, hdr, 1, data, tlen, p->qp); + /* + * Notify qib_multicast_detach() if it is waiting for us + * to finish. + */ + if (atomic_dec_return(&mcast->refcount) <= 1) + wake_up(&mcast->wait); + } else { + qp = qib_lookup_qpn(ibp, qp_num); + if (!qp) + goto drop; + ibp->n_unicast_rcv++; + qib_qp_rcv(rcd, hdr, lnh == QIB_LRH_GRH, data, tlen, qp); + /* + * Notify qib_destroy_qp() if it is waiting + * for us to finish. + */ + if (atomic_dec_and_test(&qp->refcount)) + wake_up(&qp->wait); + } + return; + +drop: + ibp->n_pkt_drops++; +} + +/* + * This is called from a timer to check for QPs + * which need kernel memory in order to send a packet. + */ +static void mem_timer(unsigned long data) +{ + struct qib_ibdev *dev = (struct qib_ibdev *) data; + struct list_head *list = &dev->memwait; + struct qib_qp *qp = NULL; + unsigned long flags; + + spin_lock_irqsave(&dev->pending_lock, flags); + if (!list_empty(list)) { + qp = list_entry(list->next, struct qib_qp, iowait); + list_del_init(&qp->iowait); + atomic_inc(&qp->refcount); + if (!list_empty(list)) + mod_timer(&dev->mem_timer, jiffies + 1); + } + spin_unlock_irqrestore(&dev->pending_lock, flags); + + if (qp) { + spin_lock_irqsave(&qp->s_lock, flags); + if (qp->s_flags & QIB_S_WAIT_KMEM) { + qp->s_flags &= ~QIB_S_WAIT_KMEM; + qib_schedule_send(qp); + } + spin_unlock_irqrestore(&qp->s_lock, flags); + if (atomic_dec_and_test(&qp->refcount)) + wake_up(&qp->wait); + } +} + +static void update_sge(struct qib_sge_state *ss, u32 length) +{ + struct qib_sge *sge = &ss->sge; + + sge->vaddr += length; + sge->length -= length; + sge->sge_length -= length; + if (sge->sge_length == 0) { + if (--ss->num_sge) + *sge = *ss->sg_list++; + } else if (sge->length == 0 && sge->mr->lkey) { + if (++sge->n >= QIB_SEGSZ) { + if (++sge->m >= sge->mr->mapsz) + return; + sge->n = 0; + } + sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr; + sge->length = sge->mr->map[sge->m]->segs[sge->n].length; + } +} + +#ifdef __LITTLE_ENDIAN +static inline u32 get_upper_bits(u32 data, u32 shift) +{ + return data >> shift; +} + +static inline u32 set_upper_bits(u32 data, u32 shift) +{ + return data << shift; +} + +static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off) +{ + data <<= ((sizeof(u32) - n) * BITS_PER_BYTE); + data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE); + return data; +} +#else +static inline u32 get_upper_bits(u32 data, u32 shift) +{ + return data << shift; +} + +static inline u32 set_upper_bits(u32 data, u32 shift) +{ + return data >> shift; +} + +static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off) +{ + data >>= ((sizeof(u32) - n) * BITS_PER_BYTE); + data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE); + return data; +} +#endif + +static void copy_io(u32 __iomem *piobuf, struct qib_sge_state *ss, + u32 length, unsigned flush_wc) +{ + u32 extra = 0; + u32 data = 0; + u32 last; + + while (1) { + u32 len = ss->sge.length; + u32 off; + + if (len > length) + len = length; + if (len > ss->sge.sge_length) + len = ss->sge.sge_length; + BUG_ON(len == 0); + /* If the source address is not aligned, try to align it. */ + off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1); + if (off) { + u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr & + ~(sizeof(u32) - 1)); + u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE); + u32 y; + + y = sizeof(u32) - off; + if (len > y) + len = y; + if (len + extra >= sizeof(u32)) { + data |= set_upper_bits(v, extra * + BITS_PER_BYTE); + len = sizeof(u32) - extra; + if (len == length) { + last = data; + break; + } + __raw_writel(data, piobuf); + piobuf++; + extra = 0; + data = 0; + } else { + /* Clear unused upper bytes */ + data |= clear_upper_bytes(v, len, extra); + if (len == length) { + last = data; + break; + } + extra += len; + } + } else if (extra) { + /* Source address is aligned. */ + u32 *addr = (u32 *) ss->sge.vaddr; + int shift = extra * BITS_PER_BYTE; + int ushift = 32 - shift; + u32 l = len; + + while (l >= sizeof(u32)) { + u32 v = *addr; + + data |= set_upper_bits(v, shift); + __raw_writel(data, piobuf); + data = get_upper_bits(v, ushift); + piobuf++; + addr++; + l -= sizeof(u32); + } + /* + * We still have 'extra' number of bytes leftover. + */ + if (l) { + u32 v = *addr; + + if (l + extra >= sizeof(u32)) { + data |= set_upper_bits(v, shift); + len -= l + extra - sizeof(u32); + if (len == length) { + last = data; + break; + } + __raw_writel(data, piobuf); + piobuf++; + extra = 0; + data = 0; + } else { + /* Clear unused upper bytes */ + data |= clear_upper_bytes(v, l, extra); + if (len == length) { + last = data; + break; + } + extra += l; + } + } else if (len == length) { + last = data; + break; + } + } else if (len == length) { + u32 w; + + /* + * Need to round up for the last dword in the + * packet. + */ + w = (len + 3) >> 2; + qib_pio_copy(piobuf, ss->sge.vaddr, w - 1); + piobuf += w - 1; + last = ((u32 *) ss->sge.vaddr)[w - 1]; + break; + } else { + u32 w = len >> 2; + + qib_pio_copy(piobuf, ss->sge.vaddr, w); + piobuf += w; + + extra = len & (sizeof(u32) - 1); + if (extra) { + u32 v = ((u32 *) ss->sge.vaddr)[w]; + + /* Clear unused upper bytes */ + data = clear_upper_bytes(v, extra, 0); + } + } + update_sge(ss, len); + length -= len; + } + /* Update address before sending packet. */ + update_sge(ss, length); + if (flush_wc) { + /* must flush early everything before trigger word */ + qib_flush_wc(); + __raw_writel(last, piobuf); + /* be sure trigger word is written */ + qib_flush_wc(); + } else + __raw_writel(last, piobuf); +} + +static struct qib_verbs_txreq *get_txreq(struct qib_ibdev *dev, + struct qib_qp *qp, int *retp) +{ + struct qib_verbs_txreq *tx; + unsigned long flags; + + spin_lock_irqsave(&qp->s_lock, flags); + spin_lock(&dev->pending_lock); + + if (!list_empty(&dev->txreq_free)) { + struct list_head *l = dev->txreq_free.next; + + list_del(l); + tx = list_entry(l, struct qib_verbs_txreq, txreq.list); + *retp = 0; + } else { + if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK && + list_empty(&qp->iowait)) { + dev->n_txwait++; + qp->s_flags |= QIB_S_WAIT_TX; + list_add_tail(&qp->iowait, &dev->txwait); + } + tx = NULL; + qp->s_flags &= ~QIB_S_BUSY; + *retp = -EBUSY; + } + + spin_unlock(&dev->pending_lock); + spin_unlock_irqrestore(&qp->s_lock, flags); + + return tx; +} + +void qib_put_txreq(struct qib_verbs_txreq *tx) +{ + struct qib_ibdev *dev; + struct qib_qp *qp; + unsigned long flags; + + qp = tx->qp; + dev = to_idev(qp->ibqp.device); + + if (atomic_dec_and_test(&qp->refcount)) + wake_up(&qp->wait); + if (tx->mr) { + atomic_dec(&tx->mr->refcount); + tx->mr = NULL; + } + if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF) { + tx->txreq.flags &= ~QIB_SDMA_TXREQ_F_FREEBUF; + dma_unmap_single(&dd_from_dev(dev)->pcidev->dev, + tx->txreq.addr, tx->hdr_dwords << 2, + DMA_TO_DEVICE); + kfree(tx->align_buf); + } + + spin_lock_irqsave(&dev->pending_lock, flags); + + /* Put struct back on free list */ + list_add(&tx->txreq.list, &dev->txreq_free); + + if (!list_empty(&dev->txwait)) { + /* Wake up first QP wanting a free struct */ + qp = list_entry(dev->txwait.next, struct qib_qp, iowait); + list_del_init(&qp->iowait); + atomic_inc(&qp->refcount); + spin_unlock_irqrestore(&dev->pending_lock, flags); + + spin_lock_irqsave(&qp->s_lock, flags); + if (qp->s_flags & QIB_S_WAIT_TX) { + qp->s_flags &= ~QIB_S_WAIT_TX; + qib_schedule_send(qp); + } + spin_unlock_irqrestore(&qp->s_lock, flags); + + if (atomic_dec_and_test(&qp->refcount)) + wake_up(&qp->wait); + } else + spin_unlock_irqrestore(&dev->pending_lock, flags); +} + +/* + * This is called when there are send DMA descriptors that might be + * available. + * + * This is called with ppd->sdma_lock held. + */ +void qib_verbs_sdma_desc_avail(struct qib_pportdata *ppd, unsigned avail) +{ + struct qib_qp *qp, *nqp; + struct qib_qp *qps[20]; + struct qib_ibdev *dev; + unsigned i, n; + + n = 0; + dev = &ppd->dd->verbs_dev; + spin_lock(&dev->pending_lock); + + /* Search wait list for first QP wanting DMA descriptors. */ + list_for_each_entry_safe(qp, nqp, &dev->dmawait, iowait) { + if (qp->port_num != ppd->port) + continue; + if (n == ARRAY_SIZE(qps)) + break; + if (qp->s_tx->txreq.sg_count > avail) + break; + avail -= qp->s_tx->txreq.sg_count; + list_del_init(&qp->iowait); + atomic_inc(&qp->refcount); + qps[n++] = qp; + } + + spin_unlock(&dev->pending_lock); + + for (i = 0; i < n; i++) { + qp = qps[i]; + spin_lock(&qp->s_lock); + if (qp->s_flags & QIB_S_WAIT_DMA_DESC) { + qp->s_flags &= ~QIB_S_WAIT_DMA_DESC; + qib_schedule_send(qp); + } + spin_unlock(&qp->s_lock); + if (atomic_dec_and_test(&qp->refcount)) + wake_up(&qp->wait); + } +} + +/* + * This is called with ppd->sdma_lock held. + */ +static void sdma_complete(struct qib_sdma_txreq *cookie, int status) +{ + struct qib_verbs_txreq *tx = + container_of(cookie, struct qib_verbs_txreq, txreq); + struct qib_qp *qp = tx->qp; + + spin_lock(&qp->s_lock); + if (tx->wqe) + qib_send_complete(qp, tx->wqe, IB_WC_SUCCESS); + else if (qp->ibqp.qp_type == IB_QPT_RC) { + struct qib_ib_header *hdr; + + if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF) + hdr = &tx->align_buf->hdr; + else { + struct qib_ibdev *dev = to_idev(qp->ibqp.device); + + hdr = &dev->pio_hdrs[tx->hdr_inx].hdr; + } + qib_rc_send_complete(qp, hdr); + } + if (atomic_dec_and_test(&qp->s_dma_busy)) { + if (qp->state == IB_QPS_RESET) + wake_up(&qp->wait_dma); + else if (qp->s_flags & QIB_S_WAIT_DMA) { + qp->s_flags &= ~QIB_S_WAIT_DMA; + qib_schedule_send(qp); + } + } + spin_unlock(&qp->s_lock); + + qib_put_txreq(tx); +} + +static int wait_kmem(struct qib_ibdev *dev, struct qib_qp *qp) +{ + unsigned long flags; + int ret = 0; + + spin_lock_irqsave(&qp->s_lock, flags); + if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) { + spin_lock(&dev->pending_lock); + if (list_empty(&qp->iowait)) { + if (list_empty(&dev->memwait)) + mod_timer(&dev->mem_timer, jiffies + 1); + qp->s_flags |= QIB_S_WAIT_KMEM; + list_add_tail(&qp->iowait, &dev->memwait); + } + spin_unlock(&dev->pending_lock); + qp->s_flags &= ~QIB_S_BUSY; + ret = -EBUSY; + } + spin_unlock_irqrestore(&qp->s_lock, flags); + + return ret; +} + +static int qib_verbs_send_dma(struct qib_qp *qp, struct qib_ib_header *hdr, + u32 hdrwords, struct qib_sge_state *ss, u32 len, + u32 plen, u32 dwords) +{ + struct qib_ibdev *dev = to_idev(qp->ibqp.device); + struct qib_devdata *dd = dd_from_dev(dev); + struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); + struct qib_pportdata *ppd = ppd_from_ibp(ibp); + struct qib_verbs_txreq *tx; + struct qib_pio_header *phdr; + u32 control; + u32 ndesc; + int ret; + + tx = qp->s_tx; + if (tx) { + qp->s_tx = NULL; + /* resend previously constructed packet */ + ret = qib_sdma_verbs_send(ppd, tx->ss, tx->dwords, tx); + goto bail; + } + + tx = get_txreq(dev, qp, &ret); + if (!tx) + goto bail; + + control = dd->f_setpbc_control(ppd, plen, qp->s_srate, + be16_to_cpu(hdr->lrh[0]) >> 12); + tx->qp = qp; + atomic_inc(&qp->refcount); + tx->wqe = qp->s_wqe; + tx->mr = qp->s_rdma_mr; + if (qp->s_rdma_mr) + qp->s_rdma_mr = NULL; + tx->txreq.callback = sdma_complete; + if (dd->flags & QIB_HAS_SDMA_TIMEOUT) + tx->txreq.flags = QIB_SDMA_TXREQ_F_HEADTOHOST; + else + tx->txreq.flags = QIB_SDMA_TXREQ_F_INTREQ; + if (plen + 1 > dd->piosize2kmax_dwords) + tx->txreq.flags |= QIB_SDMA_TXREQ_F_USELARGEBUF; + + if (len) { + /* + * Don't try to DMA if it takes more descriptors than + * the queue holds. + */ + ndesc = qib_count_sge(ss, len); + if (ndesc >= ppd->sdma_descq_cnt) + ndesc = 0; + } else + ndesc = 1; + if (ndesc) { + phdr = &dev->pio_hdrs[tx->hdr_inx]; + phdr->pbc[0] = cpu_to_le32(plen); + phdr->pbc[1] = cpu_to_le32(control); + memcpy(&phdr->hdr, hdr, hdrwords << 2); + tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEDESC; + tx->txreq.sg_count = ndesc; + tx->txreq.addr = dev->pio_hdrs_phys + + tx->hdr_inx * sizeof(struct qib_pio_header); + tx->hdr_dwords = hdrwords + 2; /* add PBC length */ + ret = qib_sdma_verbs_send(ppd, ss, dwords, tx); + goto bail; + } + + /* Allocate a buffer and copy the header and payload to it. */ + tx->hdr_dwords = plen + 1; + phdr = kmalloc(tx->hdr_dwords << 2, GFP_ATOMIC); + if (!phdr) + goto err_tx; + phdr->pbc[0] = cpu_to_le32(plen); + phdr->pbc[1] = cpu_to_le32(control); + memcpy(&phdr->hdr, hdr, hdrwords << 2); + qib_copy_from_sge((u32 *) &phdr->hdr + hdrwords, ss, len); + + tx->txreq.addr = dma_map_single(&dd->pcidev->dev, phdr, + tx->hdr_dwords << 2, DMA_TO_DEVICE); + if (dma_mapping_error(&dd->pcidev->dev, tx-> |