void ipath_do_send(unsigned long data)
{
struct ipath_qp *qp = (struct ipath_qp *)data;
struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
int (*make_req)(struct ipath_qp *qp);
unsigned long flags;
if ((qp->ibqp.qp_type == IB_QPT_RC ||
qp->ibqp.qp_type == IB_QPT_UC) &&
qp->remote_ah_attr.dlid == dev->dd->ipath_lid) {
ipath_ruc_loopback(qp);
goto bail;
}
if (qp->ibqp.qp_type == IB_QPT_RC)
make_req = ipath_make_rc_req;
else if (qp->ibqp.qp_type == IB_QPT_UC)
make_req = ipath_make_uc_req;
else
make_req = ipath_make_ud_req;
spin_lock_irqsave(&qp->s_lock, flags);
if ((qp->s_flags & (IPATH_S_BUSY | IPATH_S_ANY_WAIT)) ||
!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_OR_FLUSH_SEND)) {
spin_unlock_irqrestore(&qp->s_lock, flags);
goto bail;
}
qp->s_flags |= IPATH_S_BUSY;
spin_unlock_irqrestore(&qp->s_lock, flags);
again:
if (qp->s_hdrwords != 0) {
if (ipath_verbs_send(qp, &qp->s_hdr, qp->s_hdrwords,
qp->s_cur_sge, qp->s_cur_size)) {
if (ipath_no_bufs_available(qp, dev))
goto bail;
}
dev->n_unicast_xmit++;
qp->s_hdrwords = 0;
}
if (make_req(qp))
goto again;
bail:;
}
int ipath_destroy_qp(struct ib_qp *ibqp)
{
struct ipath_qp *qp = to_iqp(ibqp);
struct ipath_ibdev *dev = to_idev(ibqp->device);
/* */
spin_lock_irq(&qp->s_lock);
if (qp->state != IB_QPS_RESET) {
qp->state = IB_QPS_RESET;
spin_lock(&dev->pending_lock);
if (!list_empty(&qp->timerwait))
list_del_init(&qp->timerwait);
if (!list_empty(&qp->piowait))
list_del_init(&qp->piowait);
spin_unlock(&dev->pending_lock);
qp->s_flags &= ~IPATH_S_ANY_WAIT;
spin_unlock_irq(&qp->s_lock);
/* */
tasklet_kill(&qp->s_task);
wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
} else
spin_unlock_irq(&qp->s_lock);
ipath_free_qp(&dev->qp_table, qp);
if (qp->s_tx) {
atomic_dec(&qp->refcount);
if (qp->s_tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEBUF)
kfree(qp->s_tx->txreq.map_addr);
spin_lock_irq(&dev->pending_lock);
list_add(&qp->s_tx->txreq.list, &dev->txreq_free);
spin_unlock_irq(&dev->pending_lock);
qp->s_tx = NULL;
}
wait_event(qp->wait, !atomic_read(&qp->refcount));
/* */
free_qpn(&dev->qp_table, qp->ibqp.qp_num);
spin_lock(&dev->n_qps_lock);
dev->n_qps_allocated--;
spin_unlock(&dev->n_qps_lock);
if (qp->ip)
kref_put(&qp->ip->ref, ipath_release_mmap_info);
else
vfree(qp->r_rq.wq);
kfree(qp->r_ud_sg_list);
vfree(qp->s_wq);
kfree(qp);
return 0;
}
static void flush_iowait(struct hfi1_qp *qp)
{
struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
unsigned long flags;
write_seqlock_irqsave(&dev->iowait_lock, flags);
if (!list_empty(&qp->s_iowait.list)) {
list_del_init(&qp->s_iowait.list);
if (atomic_dec_and_test(&qp->refcount))
wake_up(&qp->wait);
}
write_sequnlock_irqrestore(&dev->iowait_lock, flags);
}
/**
* ipath_poll_cq - poll for work completion entries
* @ibcq: the completion queue to poll
* @num_entries: the maximum number of entries to return
* @entry: pointer to array where work completions are placed
*
* Returns the number of completion entries polled.
*
* This may be called from interrupt context. Also called by ib_poll_cq()
* in the generic verbs code.
*/
int ipath_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
{
struct ipath_cq *cq = to_icq(ibcq);
struct ipath_cq_wc *wc;
unsigned long flags;
int npolled;
u32 tail;
spin_lock_irqsave(&cq->lock, flags);
wc = cq->queue;
tail = wc->tail;
if (tail > (u32) cq->ibcq.cqe)
tail = (u32) cq->ibcq.cqe;
for (npolled = 0; npolled < num_entries; ++npolled, ++entry) {
struct ipath_qp *qp;
if (tail == wc->head)
break;
qp = ipath_lookup_qpn(&to_idev(cq->ibcq.device)->qp_table,
wc->queue[tail].qp_num);
entry->qp = &qp->ibqp;
if (atomic_dec_and_test(&qp->refcount))
wake_up(&qp->wait);
entry->wr_id = wc->queue[tail].wr_id;
entry->status = wc->queue[tail].status;
entry->opcode = wc->queue[tail].opcode;
entry->vendor_err = wc->queue[tail].vendor_err;
entry->byte_len = wc->queue[tail].byte_len;
entry->imm_data = wc->queue[tail].imm_data;
entry->src_qp = wc->queue[tail].src_qp;
entry->wc_flags = wc->queue[tail].wc_flags;
entry->pkey_index = wc->queue[tail].pkey_index;
entry->slid = wc->queue[tail].slid;
entry->sl = wc->queue[tail].sl;
entry->dlid_path_bits = wc->queue[tail].dlid_path_bits;
entry->port_num = wc->queue[tail].port_num;
if (tail >= cq->ibcq.cqe)
tail = 0;
else
tail++;
}
wc->tail = tail;
spin_unlock_irqrestore(&cq->lock, flags);
return npolled;
}
/*
* Validate a RWQE and fill in the SGE state.
* Return 1 if OK.
*/
static int qib_init_sge(struct rvt_qp *qp, struct rvt_rwqe *wqe)
{
int i, j, ret;
struct ib_wc wc;
struct rvt_lkey_table *rkt;
struct rvt_pd *pd;
struct rvt_sge_state *ss;
rkt = &to_idev(qp->ibqp.device)->rdi.lkey_table;
pd = ibpd_to_rvtpd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd);
ss = &qp->r_sge;
ss->sg_list = qp->r_sg_list;
qp->r_len = 0;
for (i = j = 0; i < wqe->num_sge; i++) {
if (wqe->sg_list[i].length == 0)
continue;
/* Check LKEY */
ret = rvt_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge,
NULL, &wqe->sg_list[i],
IB_ACCESS_LOCAL_WRITE);
if (unlikely(ret <= 0))
goto bad_lkey;
qp->r_len += wqe->sg_list[i].length;
j++;
}
ss->num_sge = j;
ss->total_len = qp->r_len;
ret = 1;
goto bail;
bad_lkey:
while (j) {
struct rvt_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge;
rvt_put_mr(sge->mr);
}
ss->num_sge = 0;
memset(&wc, 0, sizeof(wc));
wc.wr_id = wqe->wr_id;
wc.status = IB_WC_LOC_PROT_ERR;
wc.opcode = IB_WC_RECV;
wc.qp = &qp->ibqp;
/* Signal solicited completion event. */
rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
ret = 0;
bail:
return ret;
}
/**
* qib_destroy_srq - destroy a shared receive queue
* @ibsrq: the SRQ to destroy
*/
int qib_destroy_srq(struct ib_srq *ibsrq)
{
struct qib_srq *srq = to_isrq(ibsrq);
struct qib_ibdev *dev = to_idev(ibsrq->device);
spin_lock(&dev->n_srqs_lock);
dev->n_srqs_allocated--;
spin_unlock(&dev->n_srqs_lock);
if (srq->ip)
kref_put(&srq->ip->ref, qib_release_mmap_info);
else
vfree(srq->rq.wq);
kfree(srq);
return 0;
}
/*
* Validate a RWQE and fill in the SGE state.
* Return 1 if OK.
*/
static int qib_init_sge(struct qib_qp *qp, struct qib_rwqe *wqe)
{
int i, j, ret;
struct ib_wc wc;
struct qib_lkey_table *rkt;
struct qib_pd *pd;
struct qib_sge_state *ss;
rkt = &to_idev(qp->ibqp.device)->lk_table;
pd = to_ipd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd);
ss = &qp->r_sge;
ss->sg_list = qp->r_sg_list;
qp->r_len = 0;
for (i = j = 0; i < wqe->num_sge; i++) {
if (wqe->sg_list[i].length == 0)
continue;
/* Check LKEY */
if (!qib_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge,
&wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
goto bad_lkey;
qp->r_len += wqe->sg_list[i].length;
j++;
}
ss->num_sge = j;
ss->total_len = qp->r_len;
ret = 1;
goto bail;
bad_lkey:
while (j) {
struct qib_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge;
atomic_dec(&sge->mr->refcount);
}
ss->num_sge = 0;
memset(&wc, 0, sizeof(wc));
wc.wr_id = wqe->wr_id;
wc.status = IB_WC_LOC_PROT_ERR;
wc.opcode = IB_WC_RECV;
wc.qp = &qp->ibqp;
/* Signal solicited completion event. */
qib_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
ret = 0;
bail:
return ret;
}
/**
* qib_destroy_cq - destroy a completion queue
* @ibcq: the completion queue to destroy.
*
* Returns 0 for success.
*
* Called by ib_destroy_cq() in the generic verbs code.
*/
int qib_destroy_cq(struct ib_cq *ibcq)
{
struct qib_ibdev *dev = to_idev(ibcq->device);
struct qib_cq *cq = to_icq(ibcq);
flush_work(&cq->comptask);
spin_lock(&dev->n_cqs_lock);
dev->n_cqs_allocated--;
spin_unlock(&dev->n_cqs_lock);
if (cq->ip)
kref_put(&cq->ip->ref, qib_release_mmap_info);
else
vfree(cq->queue);
kfree(cq);
return 0;
}
static void qp_pio_drain(struct rvt_qp *qp)
{
struct hfi1_ibdev *dev;
struct hfi1_qp_priv *priv = qp->priv;
if (!priv->s_sendcontext)
return;
dev = to_idev(qp->ibqp.device);
while (iowait_pio_pending(&priv->s_iowait)) {
write_seqlock_irq(&dev->iowait_lock);
hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
write_sequnlock_irq(&dev->iowait_lock);
iowait_pio_drain(&priv->s_iowait);
write_seqlock_irq(&dev->iowait_lock);
hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
write_sequnlock_irq(&dev->iowait_lock);
}
}
/**
* hfi1_destroy_qp - destroy a queue pair
* @ibqp: the queue pair to destroy
*
* Returns 0 on success.
*
* Note that this can be called while the QP is actively sending or
* receiving!
*/
int hfi1_destroy_qp(struct ib_qp *ibqp)
{
struct hfi1_qp *qp = to_iqp(ibqp);
struct hfi1_ibdev *dev = to_idev(ibqp->device);
/* Make sure HW and driver activity is stopped. */
spin_lock_irq(&qp->r_lock);
spin_lock(&qp->s_lock);
if (qp->state != IB_QPS_RESET) {
qp->state = IB_QPS_RESET;
flush_iowait(qp);
qp->s_flags &= ~(HFI1_S_TIMER | HFI1_S_ANY_WAIT);
spin_unlock(&qp->s_lock);
spin_unlock_irq(&qp->r_lock);
cancel_work_sync(&qp->s_iowait.iowork);
del_timer_sync(&qp->s_timer);
iowait_sdma_drain(&qp->s_iowait);
flush_tx_list(qp);
remove_qp(dev, qp);
wait_event(qp->wait, !atomic_read(&qp->refcount));
spin_lock_irq(&qp->r_lock);
spin_lock(&qp->s_lock);
clear_mr_refs(qp, 1);
clear_ahg(qp);
}
spin_unlock(&qp->s_lock);
spin_unlock_irq(&qp->r_lock);
/* all user's cleaned up, mark it available */
free_qpn(&dev->qp_dev->qpn_table, qp->ibqp.qp_num);
spin_lock(&dev->n_qps_lock);
dev->n_qps_allocated--;
spin_unlock(&dev->n_qps_lock);
if (qp->ip)
kref_put(&qp->ip->ref, hfi1_release_mmap_info);
else
vfree(qp->r_rq.wq);
vfree(qp->s_wq);
kfree(qp->s_hdr);
kfree(qp);
return 0;
}
/**
* ipath_destroy_qp - destroy a queue pair
* @ibqp: the queue pair to destroy
*
* Returns 0 on success.
*
* Note that this can be called while the QP is actively sending or
* receiving!
*/
int ipath_destroy_qp(struct ib_qp *ibqp)
{
struct ipath_qp *qp = to_iqp(ibqp);
struct ipath_ibdev *dev = to_idev(ibqp->device);
unsigned long flags;
spin_lock_irqsave(&qp->s_lock, flags);
qp->state = IB_QPS_ERR;
spin_unlock_irqrestore(&qp->s_lock, flags);
spin_lock(&dev->n_qps_lock);
dev->n_qps_allocated--;
spin_unlock(&dev->n_qps_lock);
/* Stop the sending tasklet. */
tasklet_kill(&qp->s_task);
/* Make sure the QP isn't on the timeout list. */
spin_lock_irqsave(&dev->pending_lock, flags);
if (!list_empty(&qp->timerwait))
list_del_init(&qp->timerwait);
if (!list_empty(&qp->piowait))
list_del_init(&qp->piowait);
spin_unlock_irqrestore(&dev->pending_lock, flags);
/*
* Make sure that the QP is not in the QPN table so receive
* interrupts will discard packets for this QP. XXX Also remove QP
* from multicast table.
*/
if (atomic_read(&qp->refcount) != 0)
ipath_free_qp(&dev->qp_table, qp);
if (qp->ip)
kref_put(&qp->ip->ref, ipath_release_mmap_info);
else
vfree(qp->r_rq.wq);
vfree(qp->s_wq);
kfree(qp);
return 0;
}
/**
* qib_free_lkey - free an lkey
* @mr: mr to free from tables
*/
void qib_free_lkey(struct qib_mregion *mr)
{
unsigned long flags;
u32 lkey = mr->lkey;
u32 r;
struct qib_ibdev *dev = to_idev(mr->pd->device);
struct qib_lkey_table *rkt = &dev->lk_table;
spin_lock_irqsave(&rkt->lock, flags);
if (!mr->lkey_published)
goto out;
if (lkey == 0)
rcu_assign_pointer(dev->dma_mr, NULL);
else {
r = lkey >> (32 - ib_qib_lkey_table_size);
rcu_assign_pointer(rkt->table[r], NULL);
}
qib_put_mr(mr);
mr->lkey_published = 0;
out:
spin_unlock_irqrestore(&rkt->lock, flags);
}
void hfi1_put_txreq(struct verbs_txreq *tx)
{
struct hfi1_ibdev *dev;
struct rvt_qp *qp;
unsigned long flags;
unsigned int seq;
struct hfi1_qp_priv *priv;
qp = tx->qp;
dev = to_idev(qp->ibqp.device);
if (tx->mr)
rvt_put_mr(tx->mr);
sdma_txclean(dd_from_dev(dev), &tx->txreq);
/* Free verbs_txreq and return to slab cache */
kmem_cache_free(dev->verbs_txreq_cache, tx);
do {
seq = read_seqbegin(&dev->iowait_lock);
if (!list_empty(&dev->txwait)) {
struct iowait *wait;
write_seqlock_irqsave(&dev->iowait_lock, flags);
wait = list_first_entry(&dev->txwait, struct iowait,
list);
qp = iowait_to_qp(wait);
priv = qp->priv;
list_del_init(&priv->s_iowait.list);
/* refcount held until actual wake up */
write_sequnlock_irqrestore(&dev->iowait_lock, flags);
hfi1_qp_wakeup(qp, RVT_S_WAIT_TX);
break;
}
} while (read_seqretry(&dev->iowait_lock, seq));
}
int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
int attr_mask, struct ib_udata *udata)
{
struct ib_qp *ibqp = &qp->ibqp;
struct hfi1_ibdev *dev = to_idev(ibqp->device);
struct hfi1_devdata *dd = dd_from_dev(dev);
u8 sc;
if (attr_mask & IB_QP_AV) {
sc = ah_to_sc(ibqp->device, &attr->ah_attr);
if (sc == 0xf)
return -EINVAL;
if (!qp_to_sdma_engine(qp, sc) &&
dd->flags & HFI1_HAS_SEND_DMA)
return -EINVAL;
if (!qp_to_send_context(qp, sc))
return -EINVAL;
}
if (attr_mask & IB_QP_ALT_PATH) {
sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
if (sc == 0xf)
return -EINVAL;
if (!qp_to_sdma_engine(qp, sc) &&
dd->flags & HFI1_HAS_SEND_DMA)
return -EINVAL;
if (!qp_to_send_context(qp, sc))
return -EINVAL;
}
return 0;
}
/**
* qib_resize_cq - change the size of the CQ
* @ibcq: the completion queue
*
* Returns 0 for success.
*/
int qib_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
{
struct qib_cq *cq = to_icq(ibcq);
struct qib_cq_wc *old_wc;
struct qib_cq_wc *wc;
u32 head, tail, n;
int ret;
u32 sz;
if (cqe < 1 || cqe > ib_qib_max_cqes) {
ret = -EINVAL;
goto bail;
}
/*
* Need to use vmalloc() if we want to support large #s of entries.
*/
sz = sizeof(*wc);
if (udata && udata->outlen >= sizeof(__u64))
sz += sizeof(struct ib_uverbs_wc) * (cqe + 1);
else
sz += sizeof(struct ib_wc) * (cqe + 1);
wc = vmalloc_user(sz);
if (!wc) {
ret = -ENOMEM;
goto bail;
}
/* Check that we can write the offset to mmap. */
if (udata && udata->outlen >= sizeof(__u64)) {
__u64 offset = 0;
ret = ib_copy_to_udata(udata, &offset, sizeof(offset));
if (ret)
goto bail_free;
}
spin_lock_irq(&cq->lock);
/*
* Make sure head and tail are sane since they
* might be user writable.
*/
old_wc = cq->queue;
head = old_wc->head;
if (head > (u32) cq->ibcq.cqe)
head = (u32) cq->ibcq.cqe;
tail = old_wc->tail;
if (tail > (u32) cq->ibcq.cqe)
tail = (u32) cq->ibcq.cqe;
if (head < tail)
n = cq->ibcq.cqe + 1 + head - tail;
else
n = head - tail;
if (unlikely((u32)cqe < n)) {
ret = -EINVAL;
goto bail_unlock;
}
for (n = 0; tail != head; n++) {
if (cq->ip)
wc->uqueue[n] = old_wc->uqueue[tail];
else
wc->kqueue[n] = old_wc->kqueue[tail];
if (tail == (u32) cq->ibcq.cqe)
tail = 0;
else
tail++;
}
cq->ibcq.cqe = cqe;
wc->head = n;
wc->tail = 0;
cq->queue = wc;
spin_unlock_irq(&cq->lock);
vfree(old_wc);
if (cq->ip) {
struct qib_ibdev *dev = to_idev(ibcq->device);
struct qib_mmap_info *ip = cq->ip;
qib_update_mmap_info(dev, ip, sz, wc);
/*
* Return the offset to mmap.
* See qib_mmap() for details.
*/
if (udata && udata->outlen >= sizeof(__u64)) {
ret = ib_copy_to_udata(udata, &ip->offset,
sizeof(ip->offset));
if (ret)
goto bail;
}
spin_lock_irq(&dev->pending_lock);
if (list_empty(&ip->pending_mmaps))
list_add(&ip->pending_mmaps, &dev->pending_mmaps);
spin_unlock_irq(&dev->pending_lock);
}
ret = 0;
goto bail;
bail_unlock:
spin_unlock_irq(&cq->lock);
bail_free:
vfree(wc);
bail:
return ret;
}
/**
* qib_create_cq - create a completion queue
* @ibdev: the device this completion queue is attached to
* @entries: the minimum size of the completion queue
* @context: unused by the QLogic_IB driver
* @udata: user data for libibverbs.so
*
* Returns a pointer to the completion queue or negative errno values
* for failure.
*
* Called by ib_create_cq() in the generic verbs code.
*/
struct ib_cq *qib_create_cq(struct ib_device *ibdev, int entries,
int comp_vector, struct ib_ucontext *context,
struct ib_udata *udata)
{
struct qib_ibdev *dev = to_idev(ibdev);
struct qib_cq *cq;
struct qib_cq_wc *wc;
struct ib_cq *ret;
u32 sz;
if (entries < 1 || entries > ib_qib_max_cqes) {
ret = ERR_PTR(-EINVAL);
goto done;
}
/* Allocate the completion queue structure. */
cq = kmalloc(sizeof(*cq), GFP_KERNEL);
if (!cq) {
ret = ERR_PTR(-ENOMEM);
goto done;
}
/*
* Allocate the completion queue entries and head/tail pointers.
* This is allocated separately so that it can be resized and
* also mapped into user space.
* We need to use vmalloc() in order to support mmap and large
* numbers of entries.
*/
sz = sizeof(*wc);
if (udata && udata->outlen >= sizeof(__u64))
sz += sizeof(struct ib_uverbs_wc) * (entries + 1);
else
sz += sizeof(struct ib_wc) * (entries + 1);
wc = vmalloc_user(sz);
if (!wc) {
ret = ERR_PTR(-ENOMEM);
goto bail_cq;
}
/*
* Return the address of the WC as the offset to mmap.
* See qib_mmap() for details.
*/
if (udata && udata->outlen >= sizeof(__u64)) {
int err;
cq->ip = qib_create_mmap_info(dev, sz, context, wc);
if (!cq->ip) {
ret = ERR_PTR(-ENOMEM);
goto bail_wc;
}
err = ib_copy_to_udata(udata, &cq->ip->offset,
sizeof(cq->ip->offset));
if (err) {
ret = ERR_PTR(err);
goto bail_ip;
}
} else
cq->ip = NULL;
spin_lock(&dev->n_cqs_lock);
if (dev->n_cqs_allocated == ib_qib_max_cqs) {
spin_unlock(&dev->n_cqs_lock);
ret = ERR_PTR(-ENOMEM);
goto bail_ip;
}
dev->n_cqs_allocated++;
spin_unlock(&dev->n_cqs_lock);
if (cq->ip) {
spin_lock_irq(&dev->pending_lock);
list_add(&cq->ip->pending_mmaps, &dev->pending_mmaps);
spin_unlock_irq(&dev->pending_lock);
}
/*
* ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe.
* The number of entries should be >= the number requested or return
* an error.
*/
cq->ibcq.cqe = entries;
cq->notify = IB_CQ_NONE;
cq->triggered = 0;
spin_lock_init(&cq->lock);
INIT_WORK(&cq->comptask, send_complete);
wc->head = 0;
wc->tail = 0;
cq->queue = wc;
ret = &cq->ibcq;
goto done;
bail_ip:
kfree(cq->ip);
bail_wc:
vfree(wc);
bail_cq:
kfree(cq);
done:
return ret;
}
/**
* hfi1_modify_qp - modify the attributes of a queue pair
* @ibqp: the queue pair who's attributes we're modifying
* @attr: the new attributes
* @attr_mask: the mask of attributes to modify
* @udata: user data for libibverbs.so
*
* Returns 0 on success, otherwise returns an errno.
*/
int hfi1_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
int attr_mask, struct ib_udata *udata)
{
struct hfi1_ibdev *dev = to_idev(ibqp->device);
struct hfi1_qp *qp = to_iqp(ibqp);
enum ib_qp_state cur_state, new_state;
struct ib_event ev;
int lastwqe = 0;
int mig = 0;
int ret;
u32 pmtu = 0; /* for gcc warning only */
struct hfi1_devdata *dd;
spin_lock_irq(&qp->r_lock);
spin_lock(&qp->s_lock);
cur_state = attr_mask & IB_QP_CUR_STATE ?
attr->cur_qp_state : qp->state;
new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
attr_mask, IB_LINK_LAYER_UNSPECIFIED))
goto inval;
if (attr_mask & IB_QP_AV) {
if (attr->ah_attr.dlid >= HFI1_MULTICAST_LID_BASE)
goto inval;
if (hfi1_check_ah(qp->ibqp.device, &attr->ah_attr))
goto inval;
}
if (attr_mask & IB_QP_ALT_PATH) {
if (attr->alt_ah_attr.dlid >= HFI1_MULTICAST_LID_BASE)
goto inval;
if (hfi1_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
goto inval;
if (attr->alt_pkey_index >= hfi1_get_npkeys(dd_from_dev(dev)))
goto inval;
}
if (attr_mask & IB_QP_PKEY_INDEX)
if (attr->pkey_index >= hfi1_get_npkeys(dd_from_dev(dev)))
goto inval;
if (attr_mask & IB_QP_MIN_RNR_TIMER)
if (attr->min_rnr_timer > 31)
goto inval;
if (attr_mask & IB_QP_PORT)
if (qp->ibqp.qp_type == IB_QPT_SMI ||
qp->ibqp.qp_type == IB_QPT_GSI ||
attr->port_num == 0 ||
attr->port_num > ibqp->device->phys_port_cnt)
goto inval;
if (attr_mask & IB_QP_DEST_QPN)
if (attr->dest_qp_num > HFI1_QPN_MASK)
goto inval;
if (attr_mask & IB_QP_RETRY_CNT)
if (attr->retry_cnt > 7)
goto inval;
if (attr_mask & IB_QP_RNR_RETRY)
if (attr->rnr_retry > 7)
goto inval;
/*
* Don't allow invalid path_mtu values. OK to set greater
* than the active mtu (or even the max_cap, if we have tuned
* that to a small mtu. We'll set qp->path_mtu
* to the lesser of requested attribute mtu and active,
* for packetizing messages.
* Note that the QP port has to be set in INIT and MTU in RTR.
*/
if (attr_mask & IB_QP_PATH_MTU) {
int mtu, pidx = qp->port_num - 1;
dd = dd_from_dev(dev);
mtu = verbs_mtu_enum_to_int(ibqp->device, attr->path_mtu);
if (mtu == -1)
goto inval;
if (mtu > dd->pport[pidx].ibmtu)
pmtu = mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
else
pmtu = attr->path_mtu;
}
if (attr_mask & IB_QP_PATH_MIG_STATE) {
if (attr->path_mig_state == IB_MIG_REARM) {
if (qp->s_mig_state == IB_MIG_ARMED)
goto inval;
if (new_state != IB_QPS_RTS)
goto inval;
} else if (attr->path_mig_state == IB_MIG_MIGRATED) {
if (qp->s_mig_state == IB_MIG_REARM)
goto inval;
if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
goto inval;
if (qp->s_mig_state == IB_MIG_ARMED)
mig = 1;
} else
goto inval;
}
if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
if (attr->max_dest_rd_atomic > HFI1_MAX_RDMA_ATOMIC)
goto inval;
switch (new_state) {
case IB_QPS_RESET:
if (qp->state != IB_QPS_RESET) {
qp->state = IB_QPS_RESET;
flush_iowait(qp);
qp->s_flags &= ~(HFI1_S_TIMER | HFI1_S_ANY_WAIT);
spin_unlock(&qp->s_lock);
spin_unlock_irq(&qp->r_lock);
/* Stop the sending work queue and retry timer */
cancel_work_sync(&qp->s_iowait.iowork);
del_timer_sync(&qp->s_timer);
iowait_sdma_drain(&qp->s_iowait);
flush_tx_list(qp);
remove_qp(dev, qp);
wait_event(qp->wait, !atomic_read(&qp->refcount));
spin_lock_irq(&qp->r_lock);
spin_lock(&qp->s_lock);
clear_mr_refs(qp, 1);
clear_ahg(qp);
reset_qp(qp, ibqp->qp_type);
}
break;
case IB_QPS_RTR:
/* Allow event to re-trigger if QP set to RTR more than once */
qp->r_flags &= ~HFI1_R_COMM_EST;
qp->state = new_state;
break;
case IB_QPS_SQD:
qp->s_draining = qp->s_last != qp->s_cur;
qp->state = new_state;
break;
case IB_QPS_SQE:
if (qp->ibqp.qp_type == IB_QPT_RC)
goto inval;
qp->state = new_state;
break;
case IB_QPS_ERR:
lastwqe = hfi1_error_qp(qp, IB_WC_WR_FLUSH_ERR);
break;
default:
qp->state = new_state;
break;
}
if (attr_mask & IB_QP_PKEY_INDEX)
qp->s_pkey_index = attr->pkey_index;
if (attr_mask & IB_QP_PORT)
qp->port_num = attr->port_num;
if (attr_mask & IB_QP_DEST_QPN)
qp->remote_qpn = attr->dest_qp_num;
if (attr_mask & IB_QP_SQ_PSN) {
qp->s_next_psn = attr->sq_psn & PSN_MODIFY_MASK;
qp->s_psn = qp->s_next_psn;
qp->s_sending_psn = qp->s_next_psn;
qp->s_last_psn = qp->s_next_psn - 1;
qp->s_sending_hpsn = qp->s_last_psn;
}
if (attr_mask & IB_QP_RQ_PSN)
qp->r_psn = attr->rq_psn & PSN_MODIFY_MASK;
if (attr_mask & IB_QP_ACCESS_FLAGS)
qp->qp_access_flags = attr->qp_access_flags;
if (attr_mask & IB_QP_AV) {
qp->remote_ah_attr = attr->ah_attr;
qp->s_srate = attr->ah_attr.static_rate;
qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
}
if (attr_mask & IB_QP_ALT_PATH) {
qp->alt_ah_attr = attr->alt_ah_attr;
qp->s_alt_pkey_index = attr->alt_pkey_index;
}
if (attr_mask & IB_QP_PATH_MIG_STATE) {
qp->s_mig_state = attr->path_mig_state;
if (mig) {
qp->remote_ah_attr = qp->alt_ah_attr;
qp->port_num = qp->alt_ah_attr.port_num;
qp->s_pkey_index = qp->s_alt_pkey_index;
qp->s_flags |= HFI1_S_AHG_CLEAR;
}
}
if (attr_mask & IB_QP_PATH_MTU) {
struct hfi1_ibport *ibp;
u8 sc, vl;
u32 mtu;
dd = dd_from_dev(dev);
ibp = &dd->pport[qp->port_num - 1].ibport_data;
sc = ibp->sl_to_sc[qp->remote_ah_attr.sl];
vl = sc_to_vlt(dd, sc);
mtu = verbs_mtu_enum_to_int(ibqp->device, pmtu);
if (vl < PER_VL_SEND_CONTEXTS)
mtu = min_t(u32, mtu, dd->vld[vl].mtu);
pmtu = mtu_to_enum(mtu, OPA_MTU_8192);
qp->path_mtu = pmtu;
qp->pmtu = mtu;
}
if (attr_mask & IB_QP_RETRY_CNT) {
qp->s_retry_cnt = attr->retry_cnt;
qp->s_retry = attr->retry_cnt;
}
if (attr_mask & IB_QP_RNR_RETRY) {
qp->s_rnr_retry_cnt = attr->rnr_retry;
qp->s_rnr_retry = attr->rnr_retry;
}
if (attr_mask & IB_QP_MIN_RNR_TIMER)
qp->r_min_rnr_timer = attr->min_rnr_timer;
if (attr_mask & IB_QP_TIMEOUT) {
qp->timeout = attr->timeout;
qp->timeout_jiffies =
usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1000UL);
}
if (attr_mask & IB_QP_QKEY)
qp->qkey = attr->qkey;
if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
qp->s_max_rd_atomic = attr->max_rd_atomic;
spin_unlock(&qp->s_lock);
spin_unlock_irq(&qp->r_lock);
if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
insert_qp(dev, qp);
if (lastwqe) {
ev.device = qp->ibqp.device;
ev.element.qp = &qp->ibqp;
ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
}
if (mig) {
ev.device = qp->ibqp.device;
ev.element.qp = &qp->ibqp;
ev.event = IB_EVENT_PATH_MIG;
qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
}
ret = 0;
goto bail;
inval:
spin_unlock(&qp->s_lock);
spin_unlock_irq(&qp->r_lock);
ret = -EINVAL;
bail:
return ret;
}
/**
* qib_modify_srq - modify a shared receive queue
* @ibsrq: the SRQ to modify
* @attr: the new attributes of the SRQ
* @attr_mask: indicates which attributes to modify
* @udata: user data for libibverbs.so
*/
int qib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
enum ib_srq_attr_mask attr_mask,
struct ib_udata *udata)
{
struct qib_srq *srq = to_isrq(ibsrq);
struct qib_rwq *wq;
int ret = 0;
if (attr_mask & IB_SRQ_MAX_WR) {
struct qib_rwq *owq;
struct qib_rwqe *p;
u32 sz, size, n, head, tail;
/* Check that the requested sizes are below the limits. */
if ((attr->max_wr > ib_qib_max_srq_wrs) ||
((attr_mask & IB_SRQ_LIMIT) ?
attr->srq_limit : srq->limit) > attr->max_wr) {
ret = -EINVAL;
goto bail;
}
sz = sizeof(struct qib_rwqe) +
srq->rq.max_sge * sizeof(struct ib_sge);
size = attr->max_wr + 1;
wq = vmalloc_user(sizeof(struct qib_rwq) + size * sz);
if (!wq) {
ret = -ENOMEM;
goto bail;
}
/* Check that we can write the offset to mmap. */
if (udata && udata->inlen >= sizeof(__u64)) {
__u64 offset_addr;
__u64 offset = 0;
ret = ib_copy_from_udata(&offset_addr, udata,
sizeof(offset_addr));
if (ret)
goto bail_free;
udata->outbuf =
(void __user *) (unsigned long) offset_addr;
ret = ib_copy_to_udata(udata, &offset,
sizeof(offset));
if (ret)
goto bail_free;
}
spin_lock_irq(&srq->rq.lock);
/*
* validate head and tail pointer values and compute
* the number of remaining WQEs.
*/
owq = srq->rq.wq;
head = owq->head;
tail = owq->tail;
if (head >= srq->rq.size || tail >= srq->rq.size) {
ret = -EINVAL;
goto bail_unlock;
}
n = head;
if (n < tail)
n += srq->rq.size - tail;
else
n -= tail;
if (size <= n) {
ret = -EINVAL;
goto bail_unlock;
}
n = 0;
p = wq->wq;
while (tail != head) {
struct qib_rwqe *wqe;
int i;
wqe = get_rwqe_ptr(&srq->rq, tail);
p->wr_id = wqe->wr_id;
p->num_sge = wqe->num_sge;
for (i = 0; i < wqe->num_sge; i++)
p->sg_list[i] = wqe->sg_list[i];
n++;
p = (struct qib_rwqe *)((char *) p + sz);
if (++tail >= srq->rq.size)
tail = 0;
}
srq->rq.wq = wq;
srq->rq.size = size;
wq->head = n;
wq->tail = 0;
if (attr_mask & IB_SRQ_LIMIT)
srq->limit = attr->srq_limit;
spin_unlock_irq(&srq->rq.lock);
vfree(owq);
if (srq->ip) {
struct qib_mmap_info *ip = srq->ip;
struct qib_ibdev *dev = to_idev(srq->ibsrq.device);
u32 s = sizeof(struct qib_rwq) + size * sz;
qib_update_mmap_info(dev, ip, s, wq);
/*
* Return the offset to mmap.
* See qib_mmap() for details.
*/
if (udata && udata->inlen >= sizeof(__u64)) {
ret = ib_copy_to_udata(udata, &ip->offset,
sizeof(ip->offset));
if (ret)
goto bail;
}
/*
* Put user mapping info onto the pending list
* unless it already is on the list.
*/
spin_lock_irq(&dev->pending_lock);
if (list_empty(&ip->pending_mmaps))
list_add(&ip->pending_mmaps,
&dev->pending_mmaps);
spin_unlock_irq(&dev->pending_lock);
}
} else if (attr_mask & IB_SRQ_LIMIT) {
spin_lock_irq(&srq->rq.lock);
if (attr->srq_limit >= srq->rq.size)
ret = -EINVAL;
else
srq->limit = attr->srq_limit;
spin_unlock_irq(&srq->rq.lock);
}
goto bail;
bail_unlock:
spin_unlock_irq(&srq->rq.lock);
bail_free:
vfree(wq);
bail:
return ret;
}
static void ipath_ruc_loopback(struct ipath_qp *sqp)
{
struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
struct ipath_qp *qp;
struct ipath_swqe *wqe;
struct ipath_sge *sge;
unsigned long flags;
struct ib_wc wc;
u64 sdata;
atomic64_t *maddr;
enum ib_wc_status send_status;
qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
spin_lock_irqsave(&sqp->s_lock, flags);
if ((sqp->s_flags & (IPATH_S_BUSY | IPATH_S_ANY_WAIT)) ||
!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_OR_FLUSH_SEND))
goto unlock;
sqp->s_flags |= IPATH_S_BUSY;
again:
if (sqp->s_last == sqp->s_head)
goto clr_busy;
wqe = get_swqe_ptr(sqp, sqp->s_last);
if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_NEXT_SEND_OK)) {
if (!(ib_ipath_state_ops[sqp->state] & IPATH_FLUSH_SEND))
goto clr_busy;
send_status = IB_WC_WR_FLUSH_ERR;
goto flush_send;
}
if (sqp->s_last == sqp->s_cur) {
if (++sqp->s_cur >= sqp->s_size)
sqp->s_cur = 0;
}
spin_unlock_irqrestore(&sqp->s_lock, flags);
if (!qp || !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
dev->n_pkt_drops++;
if (sqp->ibqp.qp_type == IB_QPT_RC)
send_status = IB_WC_RETRY_EXC_ERR;
else
send_status = IB_WC_SUCCESS;
goto serr;
}
memset(&wc, 0, sizeof wc);
send_status = IB_WC_SUCCESS;
sqp->s_sge.sge = wqe->sg_list[0];
sqp->s_sge.sg_list = wqe->sg_list + 1;
sqp->s_sge.num_sge = wqe->wr.num_sge;
sqp->s_len = wqe->length;
switch (wqe->wr.opcode) {
case IB_WR_SEND_WITH_IMM:
wc.wc_flags = IB_WC_WITH_IMM;
wc.ex.imm_data = wqe->wr.ex.imm_data;
case IB_WR_SEND:
if (!ipath_get_rwqe(qp, 0))
goto rnr_nak;
break;
case IB_WR_RDMA_WRITE_WITH_IMM:
if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
goto inv_err;
wc.wc_flags = IB_WC_WITH_IMM;
wc.ex.imm_data = wqe->wr.ex.imm_data;
if (!ipath_get_rwqe(qp, 1))
goto rnr_nak;
case IB_WR_RDMA_WRITE:
if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
goto inv_err;
if (wqe->length == 0)
break;
if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, wqe->length,
wqe->wr.wr.rdma.remote_addr,
wqe->wr.wr.rdma.rkey,
IB_ACCESS_REMOTE_WRITE)))
goto acc_err;
break;
case IB_WR_RDMA_READ:
if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
goto inv_err;
if (unlikely(!ipath_rkey_ok(qp, &sqp->s_sge, wqe->length,
wqe->wr.wr.rdma.remote_addr,
wqe->wr.wr.rdma.rkey,
IB_ACCESS_REMOTE_READ)))
goto acc_err;
qp->r_sge.sge = wqe->sg_list[0];
qp->r_sge.sg_list = wqe->sg_list + 1;
qp->r_sge.num_sge = wqe->wr.num_sge;
break;
case IB_WR_ATOMIC_CMP_AND_SWP:
case IB_WR_ATOMIC_FETCH_AND_ADD:
if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
goto inv_err;
if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, sizeof(u64),
wqe->wr.wr.atomic.remote_addr,
wqe->wr.wr.atomic.rkey,
IB_ACCESS_REMOTE_ATOMIC)))
goto acc_err;
maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
sdata = wqe->wr.wr.atomic.compare_add;
*(u64 *) sqp->s_sge.sge.vaddr =
(wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
(u64) atomic64_add_return(sdata, maddr) - sdata :
(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
sdata, wqe->wr.wr.atomic.swap);
goto send_comp;
default:
send_status = IB_WC_LOC_QP_OP_ERR;
goto serr;
}
sge = &sqp->s_sge.sge;
while (sqp->s_len) {
u32 len = sqp->s_len;
if (len > sge->length)
len = sge->length;
if (len > sge->sge_length)
len = sge->sge_length;
BUG_ON(len == 0);
ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
sge->vaddr += len;
sge->length -= len;
sge->sge_length -= len;
if (sge->sge_length == 0) {
if (--sqp->s_sge.num_sge)
*sge = *sqp->s_sge.sg_list++;
} else if (sge->length == 0 && sge->mr != NULL) {
if (++sge->n >= IPATH_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;
}
sqp->s_len -= len;
}
if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
goto send_comp;
if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
else
wc.opcode = IB_WC_RECV;
wc.wr_id = qp->r_wr_id;
wc.status = IB_WC_SUCCESS;
wc.byte_len = wqe->length;
wc.qp = &qp->ibqp;
wc.src_qp = qp->remote_qpn;
wc.slid = qp->remote_ah_attr.dlid;
wc.sl = qp->remote_ah_attr.sl;
wc.port_num = 1;
ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
wqe->wr.send_flags & IB_SEND_SOLICITED);
send_comp:
spin_lock_irqsave(&sqp->s_lock, flags);
flush_send:
sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
ipath_send_complete(sqp, wqe, send_status);
goto again;
rnr_nak:
if (qp->ibqp.qp_type == IB_QPT_UC)
goto send_comp;
if (sqp->s_rnr_retry == 0) {
send_status = IB_WC_RNR_RETRY_EXC_ERR;
goto serr;
}
if (sqp->s_rnr_retry_cnt < 7)
sqp->s_rnr_retry--;
spin_lock_irqsave(&sqp->s_lock, flags);
if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_RECV_OK))
goto clr_busy;
sqp->s_flags |= IPATH_S_WAITING;
dev->n_rnr_naks++;
sqp->s_rnr_timeout = ib_ipath_rnr_table[qp->r_min_rnr_timer];
ipath_insert_rnr_queue(sqp);
goto clr_busy;
inv_err:
send_status = IB_WC_REM_INV_REQ_ERR;
wc.status = IB_WC_LOC_QP_OP_ERR;
goto err;
acc_err:
send_status = IB_WC_REM_ACCESS_ERR;
wc.status = IB_WC_LOC_PROT_ERR;
err:
ipath_rc_error(qp, wc.status);
serr:
spin_lock_irqsave(&sqp->s_lock, flags);
ipath_send_complete(sqp, wqe, send_status);
if (sqp->ibqp.qp_type == IB_QPT_RC) {
int lastwqe = ipath_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
sqp->s_flags &= ~IPATH_S_BUSY;
spin_unlock_irqrestore(&sqp->s_lock, flags);
if (lastwqe) {
struct ib_event ev;
ev.device = sqp->ibqp.device;
ev.element.qp = &sqp->ibqp;
ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
}
goto done;
}
clr_busy:
sqp->s_flags &= ~IPATH_S_BUSY;
unlock:
spin_unlock_irqrestore(&sqp->s_lock, flags);
done:
if (qp && atomic_dec_and_test(&qp->refcount))
wake_up(&qp->wait);
}
int qib_alloc_lkey(struct qib_mregion *mr, int dma_region)
{
unsigned long flags;
u32 r;
u32 n;
int ret = 0;
struct qib_ibdev *dev = to_idev(mr->pd->device);
struct qib_lkey_table *rkt = &dev->lk_table;
spin_lock_irqsave(&rkt->lock, flags);
/* special case for dma_mr lkey == 0 */
if (dma_region) {
struct qib_mregion *tmr;
tmr = rcu_dereference(dev->dma_mr);
if (!tmr) {
qib_get_mr(mr);
rcu_assign_pointer(dev->dma_mr, mr);
mr->lkey_published = 1;
}
goto success;
}
/* Find the next available LKEY */
r = rkt->next;
n = r;
for (;;) {
if (rkt->table[r] == NULL)
break;
r = (r + 1) & (rkt->max - 1);
if (r == n) {
qib_dbg("LKEY table full\n");
goto bail;
}
}
rkt->next = (r + 1) & (rkt->max - 1);
/*
* Make sure lkey is never zero which is reserved to indicate an
* unrestricted LKEY.
*/
rkt->gen++;
mr->lkey = (r << (32 - ib_qib_lkey_table_size)) |
((((1 << (24 - ib_qib_lkey_table_size)) - 1) & rkt->gen)
<< 8);
if (mr->lkey == 0) {
mr->lkey |= 1 << 8;
rkt->gen++;
}
qib_get_mr(mr);
rcu_assign_pointer(rkt->table[r], mr);
mr->lkey_published = 1;
success:
spin_unlock_irqrestore(&rkt->lock, flags);
out:
return ret;
bail:
spin_unlock_irqrestore(&rkt->lock, flags);
ret = -ENOMEM;
goto out;
}
/**
* qib_create_srq - create a shared receive queue
* @ibpd: the protection domain of the SRQ to create
* @srq_init_attr: the attributes of the SRQ
* @udata: data from libibverbs when creating a user SRQ
*/
struct ib_srq *qib_create_srq(struct ib_pd *ibpd,
struct ib_srq_init_attr *srq_init_attr,
struct ib_udata *udata)
{
struct qib_ibdev *dev = to_idev(ibpd->device);
struct qib_srq *srq;
u32 sz;
struct ib_srq *ret;
if (srq_init_attr->attr.max_sge == 0 ||
srq_init_attr->attr.max_sge > ib_qib_max_srq_sges ||
srq_init_attr->attr.max_wr == 0 ||
srq_init_attr->attr.max_wr > ib_qib_max_srq_wrs) {
ret = ERR_PTR(-EINVAL);
goto done;
}
srq = kmalloc(sizeof(*srq), GFP_KERNEL);
if (!srq) {
ret = ERR_PTR(-ENOMEM);
goto done;
}
/*
* Need to use vmalloc() if we want to support large #s of entries.
*/
srq->rq.size = srq_init_attr->attr.max_wr + 1;
srq->rq.max_sge = srq_init_attr->attr.max_sge;
sz = sizeof(struct ib_sge) * srq->rq.max_sge +
sizeof(struct qib_rwqe);
srq->rq.wq = vmalloc_user(sizeof(struct qib_rwq) + srq->rq.size * sz);
if (!srq->rq.wq) {
ret = ERR_PTR(-ENOMEM);
goto bail_srq;
}
/*
* Return the address of the RWQ as the offset to mmap.
* See qib_mmap() for details.
*/
if (udata && udata->outlen >= sizeof(__u64)) {
int err;
u32 s = sizeof(struct qib_rwq) + srq->rq.size * sz;
srq->ip =
qib_create_mmap_info(dev, s, ibpd->uobject->context,
srq->rq.wq);
if (!srq->ip) {
ret = ERR_PTR(-ENOMEM);
goto bail_wq;
}
err = ib_copy_to_udata(udata, &srq->ip->offset,
sizeof(srq->ip->offset));
if (err) {
ret = ERR_PTR(err);
goto bail_ip;
}
} else
srq->ip = NULL;
/*
* ib_create_srq() will initialize srq->ibsrq.
*/
spin_lock_init(&srq->rq.lock);
srq->rq.wq->head = 0;
srq->rq.wq->tail = 0;
srq->limit = srq_init_attr->attr.srq_limit;
spin_lock(&dev->n_srqs_lock);
if (dev->n_srqs_allocated == ib_qib_max_srqs) {
spin_unlock(&dev->n_srqs_lock);
ret = ERR_PTR(-ENOMEM);
goto bail_ip;
}
dev->n_srqs_allocated++;
spin_unlock(&dev->n_srqs_lock);
if (srq->ip) {
spin_lock_irq(&dev->pending_lock);
list_add(&srq->ip->pending_mmaps, &dev->pending_mmaps);
spin_unlock_irq(&dev->pending_lock);
}
ret = &srq->ibsrq;
goto done;
bail_ip:
kfree(srq->ip);
bail_wq:
vfree(srq->rq.wq);
bail_srq:
kfree(srq);
done:
return ret;
}
/**
* ipath_ruc_loopback - handle UC and RC lookback requests
* @sqp: the loopback QP
* @wc: the work completion entry
*
* This is called from ipath_do_uc_send() or ipath_do_rc_send() to
* forward a WQE addressed to the same HCA.
* Note that although we are single threaded due to the tasklet, we still
* have to protect against post_send(). We don't have to worry about
* receive interrupts since this is a connected protocol and all packets
* will pass through here.
*/
void ipath_ruc_loopback(struct ipath_qp *sqp, struct ib_wc *wc)
{
struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
struct ipath_qp *qp;
struct ipath_swqe *wqe;
struct ipath_sge *sge;
unsigned long flags;
u64 sdata;
qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
if (!qp) {
dev->n_pkt_drops++;
return;
}
again:
spin_lock_irqsave(&sqp->s_lock, flags);
if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_SEND_OK)) {
spin_unlock_irqrestore(&sqp->s_lock, flags);
goto done;
}
/* Get the next send request. */
if (sqp->s_last == sqp->s_head) {
/* Send work queue is empty. */
spin_unlock_irqrestore(&sqp->s_lock, flags);
goto done;
}
/*
* We can rely on the entry not changing without the s_lock
* being held until we update s_last.
*/
wqe = get_swqe_ptr(sqp, sqp->s_last);
spin_unlock_irqrestore(&sqp->s_lock, flags);
wc->wc_flags = 0;
wc->imm_data = 0;
sqp->s_sge.sge = wqe->sg_list[0];
sqp->s_sge.sg_list = wqe->sg_list + 1;
sqp->s_sge.num_sge = wqe->wr.num_sge;
sqp->s_len = wqe->length;
switch (wqe->wr.opcode) {
case IB_WR_SEND_WITH_IMM:
wc->wc_flags = IB_WC_WITH_IMM;
wc->imm_data = wqe->wr.imm_data;
/* FALLTHROUGH */
case IB_WR_SEND:
spin_lock_irqsave(&qp->r_rq.lock, flags);
if (!ipath_get_rwqe(qp, 0)) {
rnr_nak:
spin_unlock_irqrestore(&qp->r_rq.lock, flags);
/* Handle RNR NAK */
if (qp->ibqp.qp_type == IB_QPT_UC)
goto send_comp;
if (sqp->s_rnr_retry == 0) {
wc->status = IB_WC_RNR_RETRY_EXC_ERR;
goto err;
}
if (sqp->s_rnr_retry_cnt < 7)
sqp->s_rnr_retry--;
dev->n_rnr_naks++;
sqp->s_rnr_timeout =
ib_ipath_rnr_table[sqp->s_min_rnr_timer];
ipath_insert_rnr_queue(sqp);
goto done;
}
spin_unlock_irqrestore(&qp->r_rq.lock, flags);
break;
case IB_WR_RDMA_WRITE_WITH_IMM:
wc->wc_flags = IB_WC_WITH_IMM;
wc->imm_data = wqe->wr.imm_data;
spin_lock_irqsave(&qp->r_rq.lock, flags);
if (!ipath_get_rwqe(qp, 1))
goto rnr_nak;
spin_unlock_irqrestore(&qp->r_rq.lock, flags);
/* FALLTHROUGH */
case IB_WR_RDMA_WRITE:
if (wqe->length == 0)
break;
if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge, wqe->length,
wqe->wr.wr.rdma.remote_addr,
wqe->wr.wr.rdma.rkey,
IB_ACCESS_REMOTE_WRITE))) {
acc_err:
wc->status = IB_WC_REM_ACCESS_ERR;
err:
wc->wr_id = wqe->wr.wr_id;
wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
wc->vendor_err = 0;
wc->byte_len = 0;
wc->qp_num = sqp->ibqp.qp_num;
wc->src_qp = sqp->remote_qpn;
wc->pkey_index = 0;
wc->slid = sqp->remote_ah_attr.dlid;
wc->sl = sqp->remote_ah_attr.sl;
wc->dlid_path_bits = 0;
wc->port_num = 0;
ipath_sqerror_qp(sqp, wc);
goto done;
}
break;
case IB_WR_RDMA_READ:
if (unlikely(!ipath_rkey_ok(dev, &sqp->s_sge, wqe->length,
wqe->wr.wr.rdma.remote_addr,
wqe->wr.wr.rdma.rkey,
IB_ACCESS_REMOTE_READ)))
goto acc_err;
if (unlikely(!(qp->qp_access_flags &
IB_ACCESS_REMOTE_READ)))
goto acc_err;
qp->r_sge.sge = wqe->sg_list[0];
qp->r_sge.sg_list = wqe->sg_list + 1;
qp->r_sge.num_sge = wqe->wr.num_sge;
break;
case IB_WR_ATOMIC_CMP_AND_SWP:
case IB_WR_ATOMIC_FETCH_AND_ADD:
if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge, sizeof(u64),
wqe->wr.wr.rdma.remote_addr,
wqe->wr.wr.rdma.rkey,
IB_ACCESS_REMOTE_ATOMIC)))
goto acc_err;
/* Perform atomic OP and save result. */
sdata = wqe->wr.wr.atomic.swap;
spin_lock_irqsave(&dev->pending_lock, flags);
qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr;
if (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
*(u64 *) qp->r_sge.sge.vaddr =
qp->r_atomic_data + sdata;
else if (qp->r_atomic_data == wqe->wr.wr.atomic.compare_add)
*(u64 *) qp->r_sge.sge.vaddr = sdata;
spin_unlock_irqrestore(&dev->pending_lock, flags);
*(u64 *) sqp->s_sge.sge.vaddr = qp->r_atomic_data;
goto send_comp;
default:
goto done;
}
sge = &sqp->s_sge.sge;
while (sqp->s_len) {
u32 len = sqp->s_len;
if (len > sge->length)
len = sge->length;
BUG_ON(len == 0);
ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
sge->vaddr += len;
sge->length -= len;
sge->sge_length -= len;
if (sge->sge_length == 0) {
if (--sqp->s_sge.num_sge)
*sge = *sqp->s_sge.sg_list++;
} else if (sge->length == 0 && sge->mr != NULL) {
if (++sge->n >= IPATH_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;
}
sqp->s_len -= len;
}
if (wqe->wr.opcode == IB_WR_RDMA_WRITE ||
wqe->wr.opcode == IB_WR_RDMA_READ)
goto send_comp;
if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
else
wc->opcode = IB_WC_RECV;
wc->wr_id = qp->r_wr_id;
wc->status = IB_WC_SUCCESS;
wc->vendor_err = 0;
wc->byte_len = wqe->length;
wc->qp_num = qp->ibqp.qp_num;
wc->src_qp = qp->remote_qpn;
/* XXX do we know which pkey matched? Only needed for GSI. */
wc->pkey_index = 0;
wc->slid = qp->remote_ah_attr.dlid;
wc->sl = qp->remote_ah_attr.sl;
wc->dlid_path_bits = 0;
/* Signal completion event if the solicited bit is set. */
ipath_cq_enter(to_icq(qp->ibqp.recv_cq), wc,
wqe->wr.send_flags & IB_SEND_SOLICITED);
send_comp:
sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &sqp->s_flags) ||
(wqe->wr.send_flags & IB_SEND_SIGNALED)) {
wc->wr_id = wqe->wr.wr_id;
wc->status = IB_WC_SUCCESS;
wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
wc->vendor_err = 0;
wc->byte_len = wqe->length;
wc->qp_num = sqp->ibqp.qp_num;
wc->src_qp = 0;
wc->pkey_index = 0;
wc->slid = 0;
wc->sl = 0;
wc->dlid_path_bits = 0;
wc->port_num = 0;
ipath_cq_enter(to_icq(sqp->ibqp.send_cq), wc, 0);
}
/* Update s_last now that we are finished with the SWQE */
spin_lock_irqsave(&sqp->s_lock, flags);
if (++sqp->s_last >= sqp->s_size)
sqp->s_last = 0;
spin_unlock_irqrestore(&sqp->s_lock, flags);
goto again;
done:
if (atomic_dec_and_test(&qp->refcount))
wake_up(&qp->wait);
}
/**
* 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_busy))
goto bail;
if (unlikely(qp->remote_ah_attr.dlid == dev->dd->ipath_lid)) {
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);
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_busy);
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 = IPATH_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 = IPATH_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(dev->dd->ipath_lid);
bth0 |= ipath_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_busy);
bail:
return;
}
int ipath_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
int attr_mask, struct ib_udata *udata)
{
struct ipath_ibdev *dev = to_idev(ibqp->device);
struct ipath_qp *qp = to_iqp(ibqp);
enum ib_qp_state cur_state, new_state;
int lastwqe = 0;
int ret;
spin_lock_irq(&qp->s_lock);
cur_state = attr_mask & IB_QP_CUR_STATE ?
attr->cur_qp_state : qp->state;
new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
attr_mask))
goto inval;
if (attr_mask & IB_QP_AV) {
if (attr->ah_attr.dlid == 0 ||
attr->ah_attr.dlid >= IPATH_MULTICAST_LID_BASE)
goto inval;
if ((attr->ah_attr.ah_flags & IB_AH_GRH) &&
(attr->ah_attr.grh.sgid_index > 1))
goto inval;
}
if (attr_mask & IB_QP_PKEY_INDEX)
if (attr->pkey_index >= ipath_get_npkeys(dev->dd))
goto inval;
if (attr_mask & IB_QP_MIN_RNR_TIMER)
if (attr->min_rnr_timer > 31)
goto inval;
if (attr_mask & IB_QP_PORT)
if (attr->port_num == 0 ||
attr->port_num > ibqp->device->phys_port_cnt)
goto inval;
/*
*/
if ((attr_mask & IB_QP_PATH_MTU) &&
(ib_mtu_enum_to_int(attr->path_mtu) == -1 ||
(attr->path_mtu > IB_MTU_2048 && !ipath_mtu4096)))
goto inval;
if (attr_mask & IB_QP_PATH_MIG_STATE)
if (attr->path_mig_state != IB_MIG_MIGRATED &&
attr->path_mig_state != IB_MIG_REARM)
goto inval;
if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
if (attr->max_dest_rd_atomic > IPATH_MAX_RDMA_ATOMIC)
goto inval;
switch (new_state) {
case IB_QPS_RESET:
if (qp->state != IB_QPS_RESET) {
qp->state = IB_QPS_RESET;
spin_lock(&dev->pending_lock);
if (!list_empty(&qp->timerwait))
list_del_init(&qp->timerwait);
if (!list_empty(&qp->piowait))
list_del_init(&qp->piowait);
spin_unlock(&dev->pending_lock);
qp->s_flags &= ~IPATH_S_ANY_WAIT;
spin_unlock_irq(&qp->s_lock);
/* */
tasklet_kill(&qp->s_task);
wait_event(qp->wait_dma, !atomic_read(&qp->s_dma_busy));
spin_lock_irq(&qp->s_lock);
}
ipath_reset_qp(qp, ibqp->qp_type);
break;
case IB_QPS_SQD:
qp->s_draining = qp->s_last != qp->s_cur;
qp->state = new_state;
break;
case IB_QPS_SQE:
if (qp->ibqp.qp_type == IB_QPT_RC)
goto inval;
qp->state = new_state;
break;
case IB_QPS_ERR:
lastwqe = ipath_error_qp(qp, IB_WC_WR_FLUSH_ERR);
break;
default:
qp->state = new_state;
break;
}
if (attr_mask & IB_QP_PKEY_INDEX)
qp->s_pkey_index = attr->pkey_index;
if (attr_mask & IB_QP_DEST_QPN)
qp->remote_qpn = attr->dest_qp_num;
if (attr_mask & IB_QP_SQ_PSN) {
qp->s_psn = qp->s_next_psn = attr->sq_psn;
qp->s_last_psn = qp->s_next_psn - 1;
}
if (attr_mask & IB_QP_RQ_PSN)
qp->r_psn = attr->rq_psn;
if (attr_mask & IB_QP_ACCESS_FLAGS)
qp->qp_access_flags = attr->qp_access_flags;
if (attr_mask & IB_QP_AV) {
qp->remote_ah_attr = attr->ah_attr;
qp->s_dmult = ipath_ib_rate_to_mult(attr->ah_attr.static_rate);
}
if (attr_mask & IB_QP_PATH_MTU)
qp->path_mtu = attr->path_mtu;
if (attr_mask & IB_QP_RETRY_CNT)
qp->s_retry = qp->s_retry_cnt = attr->retry_cnt;
if (attr_mask & IB_QP_RNR_RETRY) {
qp->s_rnr_retry = attr->rnr_retry;
if (qp->s_rnr_retry > 7)
qp->s_rnr_retry = 7;
qp->s_rnr_retry_cnt = qp->s_rnr_retry;
}
if (attr_mask & IB_QP_MIN_RNR_TIMER)
qp->r_min_rnr_timer = attr->min_rnr_timer;
if (attr_mask & IB_QP_TIMEOUT)
qp->timeout = attr->timeout;
if (attr_mask & IB_QP_QKEY)
qp->qkey = attr->qkey;
if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
qp->s_max_rd_atomic = attr->max_rd_atomic;
spin_unlock_irq(&qp->s_lock);
if (lastwqe) {
struct ib_event ev;
ev.device = qp->ibqp.device;
ev.element.qp = &qp->ibqp;
ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
}
ret = 0;
goto bail;
inval:
spin_unlock_irq(&qp->s_lock);
ret = -EINVAL;
bail:
return ret;
}
struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
struct ib_qp_init_attr *init_attr,
struct ib_udata *udata)
{
struct ipath_qp *qp;
int err;
struct ipath_swqe *swq = NULL;
struct ipath_ibdev *dev;
size_t sz;
size_t sg_list_sz;
struct ib_qp *ret;
if (init_attr->create_flags) {
ret = ERR_PTR(-EINVAL);
goto bail;
}
if (init_attr->cap.max_send_sge > ib_ipath_max_sges ||
init_attr->cap.max_send_wr > ib_ipath_max_qp_wrs) {
ret = ERR_PTR(-EINVAL);
goto bail;
}
/* */
if (!init_attr->srq) {
if (init_attr->cap.max_recv_sge > ib_ipath_max_sges ||
init_attr->cap.max_recv_wr > ib_ipath_max_qp_wrs) {
ret = ERR_PTR(-EINVAL);
goto bail;
}
if (init_attr->cap.max_send_sge +
init_attr->cap.max_send_wr +
init_attr->cap.max_recv_sge +
init_attr->cap.max_recv_wr == 0) {
ret = ERR_PTR(-EINVAL);
goto bail;
}
}
switch (init_attr->qp_type) {
case IB_QPT_UC:
case IB_QPT_RC:
case IB_QPT_UD:
case IB_QPT_SMI:
case IB_QPT_GSI:
sz = sizeof(struct ipath_sge) *
init_attr->cap.max_send_sge +
sizeof(struct ipath_swqe);
swq = vmalloc((init_attr->cap.max_send_wr + 1) * sz);
if (swq == NULL) {
ret = ERR_PTR(-ENOMEM);
goto bail;
}
sz = sizeof(*qp);
sg_list_sz = 0;
if (init_attr->srq) {
struct ipath_srq *srq = to_isrq(init_attr->srq);
if (srq->rq.max_sge > 1)
sg_list_sz = sizeof(*qp->r_sg_list) *
(srq->rq.max_sge - 1);
} else if (init_attr->cap.max_recv_sge > 1)
sg_list_sz = sizeof(*qp->r_sg_list) *
(init_attr->cap.max_recv_sge - 1);
qp = kmalloc(sz + sg_list_sz, GFP_KERNEL);
if (!qp) {
ret = ERR_PTR(-ENOMEM);
goto bail_swq;
}
if (sg_list_sz && (init_attr->qp_type == IB_QPT_UD ||
init_attr->qp_type == IB_QPT_SMI ||
init_attr->qp_type == IB_QPT_GSI)) {
qp->r_ud_sg_list = kmalloc(sg_list_sz, GFP_KERNEL);
if (!qp->r_ud_sg_list) {
ret = ERR_PTR(-ENOMEM);
goto bail_qp;
}
} else
qp->r_ud_sg_list = NULL;
if (init_attr->srq) {
sz = 0;
qp->r_rq.size = 0;
qp->r_rq.max_sge = 0;
qp->r_rq.wq = NULL;
init_attr->cap.max_recv_wr = 0;
init_attr->cap.max_recv_sge = 0;
} else {
qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
sizeof(struct ipath_rwqe);
qp->r_rq.wq = vmalloc_user(sizeof(struct ipath_rwq) +
qp->r_rq.size * sz);
if (!qp->r_rq.wq) {
ret = ERR_PTR(-ENOMEM);
goto bail_sg_list;
}
}
/*
*/
spin_lock_init(&qp->s_lock);
spin_lock_init(&qp->r_rq.lock);
atomic_set(&qp->refcount, 0);
init_waitqueue_head(&qp->wait);
init_waitqueue_head(&qp->wait_dma);
tasklet_init(&qp->s_task, ipath_do_send, (unsigned long)qp);
INIT_LIST_HEAD(&qp->piowait);
INIT_LIST_HEAD(&qp->timerwait);
qp->state = IB_QPS_RESET;
qp->s_wq = swq;
qp->s_size = init_attr->cap.max_send_wr + 1;
qp->s_max_sge = init_attr->cap.max_send_sge;
if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
qp->s_flags = IPATH_S_SIGNAL_REQ_WR;
else
qp->s_flags = 0;
dev = to_idev(ibpd->device);
err = ipath_alloc_qpn(&dev->qp_table, qp,
init_attr->qp_type);
if (err) {
ret = ERR_PTR(err);
vfree(qp->r_rq.wq);
goto bail_sg_list;
}
qp->ip = NULL;
qp->s_tx = NULL;
ipath_reset_qp(qp, init_attr->qp_type);
break;
default:
/* */
ret = ERR_PTR(-ENOSYS);
goto bail;
}
init_attr->cap.max_inline_data = 0;
/*
*/
if (udata && udata->outlen >= sizeof(__u64)) {
if (!qp->r_rq.wq) {
__u64 offset = 0;
err = ib_copy_to_udata(udata, &offset,
sizeof(offset));
if (err) {
ret = ERR_PTR(err);
goto bail_ip;
}
} else {
u32 s = sizeof(struct ipath_rwq) +
qp->r_rq.size * sz;
qp->ip =
ipath_create_mmap_info(dev, s,
ibpd->uobject->context,
qp->r_rq.wq);
if (!qp->ip) {
ret = ERR_PTR(-ENOMEM);
goto bail_ip;
}
err = ib_copy_to_udata(udata, &(qp->ip->offset),
sizeof(qp->ip->offset));
if (err) {
ret = ERR_PTR(err);
goto bail_ip;
}
}
}
spin_lock(&dev->n_qps_lock);
if (dev->n_qps_allocated == ib_ipath_max_qps) {
spin_unlock(&dev->n_qps_lock);
ret = ERR_PTR(-ENOMEM);
goto bail_ip;
}
dev->n_qps_allocated++;
spin_unlock(&dev->n_qps_lock);
if (qp->ip) {
spin_lock_irq(&dev->pending_lock);
list_add(&qp->ip->pending_mmaps, &dev->pending_mmaps);
spin_unlock_irq(&dev->pending_lock);
}
ret = &qp->ibqp;
goto bail;
bail_ip:
if (qp->ip)
kref_put(&qp->ip->ref, ipath_release_mmap_info);
else
vfree(qp->r_rq.wq);
ipath_free_qp(&dev->qp_table, qp);
free_qpn(&dev->qp_table, qp->ibqp.qp_num);
bail_sg_list:
kfree(qp->r_ud_sg_list);
bail_qp:
kfree(qp);
bail_swq:
vfree(swq);
bail:
return ret;
}
int ipath_error_qp(struct ipath_qp *qp, enum ib_wc_status err)
{
struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
struct ib_wc wc;
int ret = 0;
if (qp->state == IB_QPS_ERR)
goto bail;
qp->state = IB_QPS_ERR;
spin_lock(&dev->pending_lock);
if (!list_empty(&qp->timerwait))
list_del_init(&qp->timerwait);
if (!list_empty(&qp->piowait))
list_del_init(&qp->piowait);
spin_unlock(&dev->pending_lock);
/* */
if (qp->s_last != qp->s_head)
ipath_schedule_send(qp);
memset(&wc, 0, sizeof(wc));
wc.qp = &qp->ibqp;
wc.opcode = IB_WC_RECV;
if (test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags)) {
wc.wr_id = qp->r_wr_id;
wc.status = err;
ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
}
wc.status = IB_WC_WR_FLUSH_ERR;
if (qp->r_rq.wq) {
struct ipath_rwq *wq;
u32 head;
u32 tail;
spin_lock(&qp->r_rq.lock);
/* */
wq = qp->r_rq.wq;
head = wq->head;
if (head >= qp->r_rq.size)
head = 0;
tail = wq->tail;
if (tail >= qp->r_rq.size)
tail = 0;
while (tail != head) {
wc.wr_id = get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
if (++tail >= qp->r_rq.size)
tail = 0;
ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
}
wq->tail = tail;
spin_unlock(&qp->r_rq.lock);
} else if (qp->ibqp.event_handler)
ret = 1;
bail:
return ret;
}
/**
* ipath_post_rc_send - post RC and UC sends
* @qp: the QP to post on
* @wr: the work request to send
*/
int ipath_post_rc_send(struct ipath_qp *qp, struct ib_send_wr *wr)
{
struct ipath_swqe *wqe;
unsigned long flags;
u32 next;
int i, j;
int acc;
int ret;
/*
* 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 (qp->ibqp.qp_type == IB_QPT_UC) {
if ((unsigned) wr->opcode >= IB_WR_RDMA_READ) {
ret = -EINVAL;
goto bail;
}
} else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
ret = -EINVAL;
goto bail;
} 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))) {
ret = -EINVAL;
goto bail;
}
/* IB spec says that num_sge == 0 is OK. */
if (wr->num_sge > qp->s_max_sge) {
ret = -ENOMEM;
goto bail;
}
spin_lock_irqsave(&qp->s_lock, flags);
next = qp->s_head + 1;
if (next >= qp->s_size)
next = 0;
if (next == qp->s_last) {
spin_unlock_irqrestore(&qp->s_lock, flags);
ret = -EINVAL;
goto bail;
}
wqe = get_swqe_ptr(qp, qp->s_head);
wqe->wr = *wr;
wqe->ssn = qp->s_ssn++;
wqe->sg_list[0].mr = NULL;
wqe->sg_list[0].vaddr = NULL;
wqe->sg_list[0].length = 0;
wqe->sg_list[0].sge_length = 0;
wqe->length = 0;
acc = wr->opcode >= IB_WR_RDMA_READ ? IB_ACCESS_LOCAL_WRITE : 0;
for (i = 0, j = 0; i < wr->num_sge; i++) {
if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
spin_unlock_irqrestore(&qp->s_lock, flags);
ret = -EINVAL;
goto bail;
}
if (wr->sg_list[i].length == 0)
continue;
if (!ipath_lkey_ok(&to_idev(qp->ibqp.device)->lk_table,
&wqe->sg_list[j], &wr->sg_list[i],
acc)) {
spin_unlock_irqrestore(&qp->s_lock, flags);
ret = -EINVAL;
goto bail;
}
wqe->length += wr->sg_list[i].length;
j++;
}
wqe->wr.num_sge = j;
qp->s_head = next;
spin_unlock_irqrestore(&qp->s_lock, flags);
if (qp->ibqp.qp_type == IB_QPT_UC)
ipath_do_uc_send((unsigned long) qp);
else
ipath_do_rc_send((unsigned long) qp);
ret = 0;
bail:
return ret;
}
/**
* ipath_post_srq_receive - post a receive on a shared receive queue
* @ibsrq: the SRQ to post the receive on
* @wr: the list of work requests to post
* @bad_wr: the first WR to cause a problem is put here
*
* This may be called from interrupt context.
*/
int ipath_post_srq_receive(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
struct ib_recv_wr **bad_wr)
{
struct ipath_srq *srq = to_isrq(ibsrq);
struct ipath_ibdev *dev = to_idev(ibsrq->device);
unsigned long flags;
int ret;
for (; wr; wr = wr->next) {
struct ipath_rwqe *wqe;
u32 next;
int i, j;
if (wr->num_sge > srq->rq.max_sge) {
*bad_wr = wr;
ret = -ENOMEM;
goto bail;
}
spin_lock_irqsave(&srq->rq.lock, flags);
next = srq->rq.head + 1;
if (next >= srq->rq.size)
next = 0;
if (next == srq->rq.tail) {
spin_unlock_irqrestore(&srq->rq.lock, flags);
*bad_wr = wr;
ret = -ENOMEM;
goto bail;
}
wqe = get_rwqe_ptr(&srq->rq, srq->rq.head);
wqe->wr_id = wr->wr_id;
wqe->sg_list[0].mr = NULL;
wqe->sg_list[0].vaddr = NULL;
wqe->sg_list[0].length = 0;
wqe->sg_list[0].sge_length = 0;
wqe->length = 0;
for (i = 0, j = 0; i < wr->num_sge; i++) {
/* Check LKEY */
if (to_ipd(srq->ibsrq.pd)->user &&
wr->sg_list[i].lkey == 0) {
spin_unlock_irqrestore(&srq->rq.lock,
flags);
*bad_wr = wr;
ret = -EINVAL;
goto bail;
}
if (wr->sg_list[i].length == 0)
continue;
if (!ipath_lkey_ok(&dev->lk_table,
&wqe->sg_list[j],
&wr->sg_list[i],
IB_ACCESS_LOCAL_WRITE)) {
spin_unlock_irqrestore(&srq->rq.lock,
flags);
*bad_wr = wr;
ret = -EINVAL;
goto bail;
}
wqe->length += wr->sg_list[i].length;
j++;
}
wqe->num_sge = j;
srq->rq.head = next;
spin_unlock_irqrestore(&srq->rq.lock, flags);
}
ret = 0;
bail:
return ret;
}
void ipath_cq_enter(struct ipath_cq *cq, struct ib_wc *entry, int solicited)
{
struct ipath_cq_wc *wc;
unsigned long flags;
u32 head;
u32 next;
spin_lock_irqsave(&cq->lock, flags);
/*
* Note that the head pointer might be writable by user processes.
* Take care to verify it is a sane value.
*/
wc = cq->queue;
head = wc->head;
if (head >= (unsigned) cq->ibcq.cqe) {
head = cq->ibcq.cqe;
next = 0;
} else
next = head + 1;
if (unlikely(next == wc->tail)) {
spin_unlock_irqrestore(&cq->lock, flags);
if (cq->ibcq.event_handler) {
struct ib_event ev;
ev.device = cq->ibcq.device;
ev.element.cq = &cq->ibcq;
ev.event = IB_EVENT_CQ_ERR;
cq->ibcq.event_handler(&ev, cq->ibcq.cq_context);
}
return;
}
if (cq->ip) {
wc->uqueue[head].wr_id = entry->wr_id;
wc->uqueue[head].status = entry->status;
wc->uqueue[head].opcode = entry->opcode;
wc->uqueue[head].vendor_err = entry->vendor_err;
wc->uqueue[head].byte_len = entry->byte_len;
wc->uqueue[head].ex.imm_data = (__u32 __force) entry->ex.imm_data;
wc->uqueue[head].qp_num = entry->qp->qp_num;
wc->uqueue[head].src_qp = entry->src_qp;
wc->uqueue[head].wc_flags = entry->wc_flags;
wc->uqueue[head].pkey_index = entry->pkey_index;
wc->uqueue[head].slid = entry->slid;
wc->uqueue[head].sl = entry->sl;
wc->uqueue[head].dlid_path_bits = entry->dlid_path_bits;
wc->uqueue[head].port_num = entry->port_num;
/* Make sure entry is written before the head index. */
smp_wmb();
} else
wc->kqueue[head] = *entry;
wc->head = next;
if (cq->notify == IB_CQ_NEXT_COMP ||
(cq->notify == IB_CQ_SOLICITED && solicited)) {
cq->notify = IB_CQ_NONE;
cq->triggered++;
/*
* This will cause send_complete() to be called in
* another thread.
*/
tasklet_hi_schedule(&cq->comptask);
}
spin_unlock_irqrestore(&cq->lock, flags);
if (entry->status != IB_WC_SUCCESS)
to_idev(cq->ibcq.device)->n_wqe_errs++;
}
/**
* ipath_ruc_loopback - handle UC and RC lookback requests
* @sqp: the sending QP
*
* This is called from ipath_do_send() to
* forward a WQE addressed to the same HCA.
* Note that although we are single threaded due to the tasklet, we still
* have to protect against post_send(). We don't have to worry about
* receive interrupts since this is a connected protocol and all packets
* will pass through here.
*/
static void ipath_ruc_loopback(struct ipath_qp *sqp)
{
struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
struct ipath_qp *qp;
struct ipath_swqe *wqe;
struct ipath_sge *sge;
unsigned long flags;
struct ib_wc wc;
u64 sdata;
atomic64_t *maddr;
enum ib_wc_status send_status;
/*
* Note that we check the responder QP state after
* checking the requester's state.
*/
qp = ipath_lookup_qpn(&dev->qp_table, sqp->remote_qpn);
spin_lock_irqsave(&sqp->s_lock, flags);
/* Return if we are already busy processing a work request. */
if ((sqp->s_flags & (IPATH_S_BUSY | IPATH_S_ANY_WAIT)) ||
!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_OR_FLUSH_SEND))
goto unlock;
sqp->s_flags |= IPATH_S_BUSY;
again:
if (sqp->s_last == sqp->s_head)
goto clr_busy;
wqe = get_swqe_ptr(sqp, sqp->s_last);
/* Return if it is not OK to start a new work reqeust. */
if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_NEXT_SEND_OK)) {
if (!(ib_ipath_state_ops[sqp->state] & IPATH_FLUSH_SEND))
goto clr_busy;
/* We are in the error state, flush the work request. */
send_status = IB_WC_WR_FLUSH_ERR;
goto flush_send;
}
/*
* We can rely on the entry not changing without the s_lock
* being held until we update s_last.
* We increment s_cur to indicate s_last is in progress.
*/
if (sqp->s_last == sqp->s_cur) {
if (++sqp->s_cur >= sqp->s_size)
sqp->s_cur = 0;
}
spin_unlock_irqrestore(&sqp->s_lock, flags);
if (!qp || !(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_RECV_OK)) {
dev->n_pkt_drops++;
/*
* For RC, the requester would timeout and retry so
* shortcut the timeouts and just signal too many retries.
*/
if (sqp->ibqp.qp_type == IB_QPT_RC)
send_status = IB_WC_RETRY_EXC_ERR;
else
send_status = IB_WC_SUCCESS;
goto serr;
}
memset(&wc, 0, sizeof wc);
send_status = IB_WC_SUCCESS;
sqp->s_sge.sge = wqe->sg_list[0];
sqp->s_sge.sg_list = wqe->sg_list + 1;
sqp->s_sge.num_sge = wqe->wr.num_sge;
sqp->s_len = wqe->length;
switch (wqe->wr.opcode) {
case IB_WR_SEND_WITH_IMM:
wc.wc_flags = IB_WC_WITH_IMM;
wc.imm_data = wqe->wr.ex.imm_data;
/* FALLTHROUGH */
case IB_WR_SEND:
if (!ipath_get_rwqe(qp, 0))
goto rnr_nak;
break;
case IB_WR_RDMA_WRITE_WITH_IMM:
if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
goto inv_err;
wc.wc_flags = IB_WC_WITH_IMM;
wc.imm_data = wqe->wr.ex.imm_data;
if (!ipath_get_rwqe(qp, 1))
goto rnr_nak;
/* FALLTHROUGH */
case IB_WR_RDMA_WRITE:
if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
goto inv_err;
if (wqe->length == 0)
break;
if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, wqe->length,
wqe->wr.wr.rdma.remote_addr,
wqe->wr.wr.rdma.rkey,
IB_ACCESS_REMOTE_WRITE)))
goto acc_err;
break;
case IB_WR_RDMA_READ:
if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
goto inv_err;
if (unlikely(!ipath_rkey_ok(qp, &sqp->s_sge, wqe->length,
wqe->wr.wr.rdma.remote_addr,
wqe->wr.wr.rdma.rkey,
IB_ACCESS_REMOTE_READ)))
goto acc_err;
qp->r_sge.sge = wqe->sg_list[0];
qp->r_sge.sg_list = wqe->sg_list + 1;
qp->r_sge.num_sge = wqe->wr.num_sge;
break;
case IB_WR_ATOMIC_CMP_AND_SWP:
case IB_WR_ATOMIC_FETCH_AND_ADD:
if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
goto inv_err;
if (unlikely(!ipath_rkey_ok(qp, &qp->r_sge, sizeof(u64),
wqe->wr.wr.atomic.remote_addr,
wqe->wr.wr.atomic.rkey,
IB_ACCESS_REMOTE_ATOMIC)))
goto acc_err;
/* Perform atomic OP and save result. */
maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
sdata = wqe->wr.wr.atomic.compare_add;
*(u64 *) sqp->s_sge.sge.vaddr =
(wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
(u64) atomic64_add_return(sdata, maddr) - sdata :
(u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
sdata, wqe->wr.wr.atomic.swap);
goto send_comp;
default:
send_status = IB_WC_LOC_QP_OP_ERR;
goto serr;
}
sge = &sqp->s_sge.sge;
while (sqp->s_len) {
u32 len = sqp->s_len;
if (len > sge->length)
len = sge->length;
if (len > sge->sge_length)
len = sge->sge_length;
BUG_ON(len == 0);
ipath_copy_sge(&qp->r_sge, sge->vaddr, len);
sge->vaddr += len;
sge->length -= len;
sge->sge_length -= len;
if (sge->sge_length == 0) {
if (--sqp->s_sge.num_sge)
*sge = *sqp->s_sge.sg_list++;
} else if (sge->length == 0 && sge->mr != NULL) {
if (++sge->n >= IPATH_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;
}
sqp->s_len -= len;
}
if (!test_and_clear_bit(IPATH_R_WRID_VALID, &qp->r_aflags))
goto send_comp;
if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
else
wc.opcode = IB_WC_RECV;
wc.wr_id = qp->r_wr_id;
wc.status = IB_WC_SUCCESS;
wc.byte_len = wqe->length;
wc.qp = &qp->ibqp;
wc.src_qp = qp->remote_qpn;
wc.slid = qp->remote_ah_attr.dlid;
wc.sl = qp->remote_ah_attr.sl;
wc.port_num = 1;
/* Signal completion event if the solicited bit is set. */
ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
wqe->wr.send_flags & IB_SEND_SOLICITED);
send_comp:
spin_lock_irqsave(&sqp->s_lock, flags);
flush_send:
sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
ipath_send_complete(sqp, wqe, send_status);
goto again;
rnr_nak:
/* Handle RNR NAK */
if (qp->ibqp.qp_type == IB_QPT_UC)
goto send_comp;
/*
* Note: we don't need the s_lock held since the BUSY flag
* makes this single threaded.
*/
if (sqp->s_rnr_retry == 0) {
send_status = IB_WC_RNR_RETRY_EXC_ERR;
goto serr;
}
if (sqp->s_rnr_retry_cnt < 7)
sqp->s_rnr_retry--;
spin_lock_irqsave(&sqp->s_lock, flags);
if (!(ib_ipath_state_ops[sqp->state] & IPATH_PROCESS_RECV_OK))
goto clr_busy;
sqp->s_flags |= IPATH_S_WAITING;
dev->n_rnr_naks++;
sqp->s_rnr_timeout = ib_ipath_rnr_table[qp->r_min_rnr_timer];
ipath_insert_rnr_queue(sqp);
goto clr_busy;
inv_err:
send_status = IB_WC_REM_INV_REQ_ERR;
wc.status = IB_WC_LOC_QP_OP_ERR;
goto err;
acc_err:
send_status = IB_WC_REM_ACCESS_ERR;
wc.status = IB_WC_LOC_PROT_ERR;
err:
/* responder goes to error state */
ipath_rc_error(qp, wc.status);
serr:
spin_lock_irqsave(&sqp->s_lock, flags);
ipath_send_complete(sqp, wqe, send_status);
if (sqp->ibqp.qp_type == IB_QPT_RC) {
int lastwqe = ipath_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
sqp->s_flags &= ~IPATH_S_BUSY;
spin_unlock_irqrestore(&sqp->s_lock, flags);
if (lastwqe) {
struct ib_event ev;
ev.device = sqp->ibqp.device;
ev.element.qp = &sqp->ibqp;
ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
}
goto done;
}
clr_busy:
sqp->s_flags &= ~IPATH_S_BUSY;
unlock:
spin_unlock_irqrestore(&sqp->s_lock, flags);
done:
if (qp && atomic_dec_and_test(&qp->refcount))
wake_up(&qp->wait);
}
