/** rvt_query_srq - query srq data
* @ibsrq: srq to query
* @attr: return info in attr
*
* Return: always 0
*/
int rvt_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
{
struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
attr->max_wr = srq->rq.size - 1;
attr->max_sge = srq->rq.max_sge;
attr->srq_limit = srq->limit;
return 0;
}
/**
* rvt_destroy_srq - destory an srq
* @ibsrq: srq object to destroy
*
* Return always 0
*/
int rvt_destroy_srq(struct ib_srq *ibsrq)
{
struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
struct rvt_dev_info *dev = ib_to_rvt(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, rvt_release_mmap_info);
else
vfree(srq->rq.wq);
kfree(srq);
return 0;
}
/**
* rvt_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: A pointer to the first WR to cause a problem is put here
*
* This may be called from interrupt context.
*
* Return: 0 on success else errno
*/
int rvt_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
struct ib_recv_wr **bad_wr)
{
struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
struct rvt_rwq *wq;
unsigned long flags;
for (; wr; wr = wr->next) {
struct rvt_rwqe *wqe;
u32 next;
int i;
if ((unsigned)wr->num_sge > srq->rq.max_sge) {
*bad_wr = wr;
return -EINVAL;
}
spin_lock_irqsave(&srq->rq.lock, flags);
wq = srq->rq.wq;
next = wq->head + 1;
if (next >= srq->rq.size)
next = 0;
if (next == wq->tail) {
spin_unlock_irqrestore(&srq->rq.lock, flags);
*bad_wr = wr;
return -ENOMEM;
}
wqe = rvt_get_rwqe_ptr(&srq->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(&srq->rq.lock, flags);
}
return 0;
}
/**
* qib_get_rwqe - copy the next RWQE into the QP's RWQE
* @qp: the QP
* @wr_id_only: update qp->r_wr_id only, not qp->r_sge
*
* Return -1 if there is a local error, 0 if no RWQE is available,
* otherwise return 1.
*
* Can be called from interrupt level.
*/
int qib_get_rwqe(struct rvt_qp *qp, int wr_id_only)
{
unsigned long flags;
struct rvt_rq *rq;
struct rvt_rwq *wq;
struct rvt_srq *srq;
struct rvt_rwqe *wqe;
void (*handler)(struct ib_event *, void *);
u32 tail;
int ret;
if (qp->ibqp.srq) {
srq = ibsrq_to_rvtsrq(qp->ibqp.srq);
handler = srq->ibsrq.event_handler;
rq = &srq->rq;
} else {
srq = NULL;
handler = NULL;
rq = &qp->r_rq;
}
spin_lock_irqsave(&rq->lock, flags);
if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
ret = 0;
goto unlock;
}
wq = rq->wq;
tail = wq->tail;
/* Validate tail before using it since it is user writable. */
if (tail >= rq->size)
tail = 0;
if (unlikely(tail == wq->head)) {
ret = 0;
goto unlock;
}
/* Make sure entry is read after head index is read. */
smp_rmb();
wqe = rvt_get_rwqe_ptr(rq, tail);
/*
* Even though we update the tail index in memory, the verbs
* consumer is not supposed to post more entries until a
* completion is generated.
*/
if (++tail >= rq->size)
tail = 0;
wq->tail = tail;
if (!wr_id_only && !qib_init_sge(qp, wqe)) {
ret = -1;
goto unlock;
}
qp->r_wr_id = wqe->wr_id;
ret = 1;
set_bit(RVT_R_WRID_VALID, &qp->r_aflags);
if (handler) {
u32 n;
/*
* Validate head pointer value and compute
* the number of remaining WQEs.
*/
n = wq->head;
if (n >= rq->size)
n = 0;
if (n < tail)
n += rq->size - tail;
else
n -= tail;
if (n < srq->limit) {
struct ib_event ev;
srq->limit = 0;
spin_unlock_irqrestore(&rq->lock, flags);
ev.device = qp->ibqp.device;
ev.element.srq = qp->ibqp.srq;
ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
handler(&ev, srq->ibsrq.srq_context);
goto bail;
}
}
unlock:
spin_unlock_irqrestore(&rq->lock, flags);
bail:
return ret;
}
/**
* rvt_create_qp - create a queue pair for a device
* @ibpd: the protection domain who's device we create the queue pair for
* @init_attr: the attributes of the queue pair
* @udata: user data for libibverbs.so
*
* Queue pair creation is mostly an rvt issue. However, drivers have their own
* unique idea of what queue pair numbers mean. For instance there is a reserved
* range for PSM.
*
* Return: the queue pair on success, otherwise returns an errno.
*
* Called by the ib_create_qp() core verbs function.
*/
struct ib_qp *rvt_create_qp(struct ib_pd *ibpd,
struct ib_qp_init_attr *init_attr,
struct ib_udata *udata)
{
struct rvt_qp *qp;
int err;
struct rvt_swqe *swq = NULL;
size_t sz;
size_t sg_list_sz;
struct ib_qp *ret = ERR_PTR(-ENOMEM);
struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device);
void *priv = NULL;
gfp_t gfp;
if (!rdi)
return ERR_PTR(-EINVAL);
if (init_attr->cap.max_send_sge > rdi->dparms.props.max_sge ||
init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr ||
init_attr->create_flags & ~(IB_QP_CREATE_USE_GFP_NOIO))
return ERR_PTR(-EINVAL);
/* GFP_NOIO is applicable to RC QP's only */
if (init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO &&
init_attr->qp_type != IB_QPT_RC)
return ERR_PTR(-EINVAL);
gfp = init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO ?
GFP_NOIO : GFP_KERNEL;
/* Check receive queue parameters if no SRQ is specified. */
if (!init_attr->srq) {
if (init_attr->cap.max_recv_sge > rdi->dparms.props.max_sge ||
init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr)
return ERR_PTR(-EINVAL);
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)
return ERR_PTR(-EINVAL);
}
switch (init_attr->qp_type) {
case IB_QPT_SMI:
case IB_QPT_GSI:
if (init_attr->port_num == 0 ||
init_attr->port_num > ibpd->device->phys_port_cnt)
return ERR_PTR(-EINVAL);
case IB_QPT_UC:
case IB_QPT_RC:
case IB_QPT_UD:
sz = sizeof(struct rvt_sge) *
init_attr->cap.max_send_sge +
sizeof(struct rvt_swqe);
if (gfp == GFP_NOIO)
swq = __vmalloc(
(init_attr->cap.max_send_wr + 1) * sz,
gfp, PAGE_KERNEL);
else
swq = vmalloc_node(
(init_attr->cap.max_send_wr + 1) * sz,
rdi->dparms.node);
if (!swq)
return ERR_PTR(-ENOMEM);
sz = sizeof(*qp);
sg_list_sz = 0;
if (init_attr->srq) {
struct rvt_srq *srq = ibsrq_to_rvtsrq(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 = kzalloc_node(sz + sg_list_sz, gfp, rdi->dparms.node);
if (!qp)
goto bail_swq;
RCU_INIT_POINTER(qp->next, NULL);
/*
* Driver needs to set up it's private QP structure and do any
* initialization that is needed.
*/
priv = rdi->driver_f.qp_priv_alloc(rdi, qp, gfp);
if (!priv)
goto bail_qp;
qp->priv = priv;
qp->timeout_jiffies =
usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
1000UL);
if (init_attr->srq) {
sz = 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 rvt_rwqe);
if (udata)
qp->r_rq.wq = vmalloc_user(
sizeof(struct rvt_rwq) +
qp->r_rq.size * sz);
else if (gfp == GFP_NOIO)
qp->r_rq.wq = __vmalloc(
sizeof(struct rvt_rwq) +
qp->r_rq.size * sz,
gfp, PAGE_KERNEL);
else
qp->r_rq.wq = vmalloc_node(
sizeof(struct rvt_rwq) +
qp->r_rq.size * sz,
rdi->dparms.node);
if (!qp->r_rq.wq)
goto bail_driver_priv;
}
/*
* ib_create_qp() will initialize qp->ibqp
* except for qp->ibqp.qp_num.
*/
spin_lock_init(&qp->r_lock);
spin_lock_init(&qp->s_hlock);
spin_lock_init(&qp->s_lock);
spin_lock_init(&qp->r_rq.lock);
atomic_set(&qp->refcount, 0);
init_waitqueue_head(&qp->wait);
init_timer(&qp->s_timer);
qp->s_timer.data = (unsigned long)qp;
INIT_LIST_HEAD(&qp->rspwait);
qp->state = IB_QPS_RESET;
qp->s_wq = swq;
qp->s_size = init_attr->cap.max_send_wr + 1;
qp->s_avail = init_attr->cap.max_send_wr;
qp->s_max_sge = init_attr->cap.max_send_sge;
if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
qp->s_flags = RVT_S_SIGNAL_REQ_WR;
err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table,
init_attr->qp_type,
init_attr->port_num, gfp);
if (err < 0) {
ret = ERR_PTR(err);
goto bail_rq_wq;
}
qp->ibqp.qp_num = err;
qp->port_num = init_attr->port_num;
rvt_reset_qp(rdi, qp, init_attr->qp_type);
break;
default:
/* Don't support raw QPs */
return ERR_PTR(-EINVAL);
}
init_attr->cap.max_inline_data = 0;
/*
* Return the address of the RWQ as the offset to mmap.
* See rvt_mmap() for details.
*/
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_qpn;
}
} else {
u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz;
qp->ip = rvt_create_mmap_info(rdi, s,
ibpd->uobject->context,
qp->r_rq.wq);
if (!qp->ip) {
ret = ERR_PTR(-ENOMEM);
goto bail_qpn;
}
err = ib_copy_to_udata(udata, &qp->ip->offset,
sizeof(qp->ip->offset));
if (err) {
ret = ERR_PTR(err);
goto bail_ip;
}
}
qp->pid = current->pid;
}
spin_lock(&rdi->n_qps_lock);
if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) {
spin_unlock(&rdi->n_qps_lock);
ret = ERR_PTR(-ENOMEM);
goto bail_ip;
}
rdi->n_qps_allocated++;
/*
* Maintain a busy_jiffies variable that will be added to the timeout
* period in mod_retry_timer and add_retry_timer. This busy jiffies
* is scaled by the number of rc qps created for the device to reduce
* the number of timeouts occurring when there is a large number of
* qps. busy_jiffies is incremented every rc qp scaling interval.
* The scaling interval is selected based on extensive performance
* evaluation of targeted workloads.
*/
if (init_attr->qp_type == IB_QPT_RC) {
rdi->n_rc_qps++;
rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
}
spin_unlock(&rdi->n_qps_lock);
if (qp->ip) {
spin_lock_irq(&rdi->pending_lock);
list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps);
spin_unlock_irq(&rdi->pending_lock);
}
ret = &qp->ibqp;
/*
* We have our QP and its good, now keep track of what types of opcodes
* can be processed on this QP. We do this by keeping track of what the
* 3 high order bits of the opcode are.
*/
switch (init_attr->qp_type) {
case IB_QPT_SMI:
case IB_QPT_GSI:
case IB_QPT_UD:
qp->allowed_ops = IB_OPCODE_UD_SEND_ONLY & RVT_OPCODE_QP_MASK;
break;
case IB_QPT_RC:
qp->allowed_ops = IB_OPCODE_RC_SEND_ONLY & RVT_OPCODE_QP_MASK;
break;
case IB_QPT_UC:
qp->allowed_ops = IB_OPCODE_UC_SEND_ONLY & RVT_OPCODE_QP_MASK;
break;
default:
ret = ERR_PTR(-EINVAL);
goto bail_ip;
}
return ret;
bail_ip:
kref_put(&qp->ip->ref, rvt_release_mmap_info);
bail_qpn:
free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
bail_rq_wq:
vfree(qp->r_rq.wq);
bail_driver_priv:
rdi->driver_f.qp_priv_free(rdi, qp);
bail_qp:
kfree(qp);
bail_swq:
vfree(swq);
return ret;
}
/**
* rvt_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
*
* Return: 0 on success
*/
int rvt_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
enum ib_srq_attr_mask attr_mask,
struct ib_udata *udata)
{
struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
struct rvt_dev_info *dev = ib_to_rvt(ibsrq->device);
struct rvt_rwq *wq;
int ret = 0;
if (attr_mask & IB_SRQ_MAX_WR) {
struct rvt_rwq *owq;
struct rvt_rwqe *p;
u32 sz, size, n, head, tail;
/* Check that the requested sizes are below the limits. */
if ((attr->max_wr > dev->dparms.props.max_srq_wr) ||
((attr_mask & IB_SRQ_LIMIT) ?
attr->srq_limit : srq->limit) > attr->max_wr)
return -EINVAL;
sz = sizeof(struct rvt_rwqe) +
srq->rq.max_sge * sizeof(struct ib_sge);
size = attr->max_wr + 1;
wq = udata ?
vmalloc_user(sizeof(struct rvt_rwq) + size * sz) :
vzalloc_node(sizeof(struct rvt_rwq) + size * sz,
dev->dparms.node);
if (!wq)
return -ENOMEM;
/* 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 rvt_rwqe *wqe;
int i;
wqe = rvt_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 rvt_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 rvt_mmap_info *ip = srq->ip;
struct rvt_dev_info *dev = ib_to_rvt(srq->ibsrq.device);
u32 s = sizeof(struct rvt_rwq) + size * sz;
rvt_update_mmap_info(dev, ip, s, wq);
/*
* Return the offset to mmap.
* See rvt_mmap() for details.
*/
if (udata && udata->inlen >= sizeof(__u64)) {
ret = ib_copy_to_udata(udata, &ip->offset,
sizeof(ip->offset));
if (ret)
return ret;
}
/*
* 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);
}
return ret;
bail_unlock:
spin_unlock_irq(&srq->rq.lock);
bail_free:
vfree(wq);
return ret;
}
