/* * Note that the head pointer might be writable by * user processes.Take care to verify it is a sane value.
*/ if (head >= (unsigned)cq->ibcq.cqe) {
head = cq->ibcq.cqe;
next = 0;
} else {
next = head + 1;
}
if (cq->notify == IB_CQ_NEXT_COMP ||
(cq->notify == IB_CQ_SOLICITED &&
(solicited || entry->status != IB_WC_SUCCESS))) { /* * This will cause send_complete() to be called in * another thread.
*/
cq->notify = RVT_CQ_NONE;
cq->triggered++;
queue_work_on(cq->comp_vector_cpu, comp_vector_wq,
&cq->comptask);
}
/* * The completion handler will most likely rearm the notification * and poll for all pending entries. If a new completion entry * is added while we are in this routine, queue_work() * won't call us again until we return so we check triggered to * see if we need to call the handler again.
*/ for (;;) {
u8 triggered = cq->triggered;
/* * IPoIB connected mode assumes the callback is from a * soft IRQ. We simulate this by blocking "bottom halves". * See the implementation for ipoib_cm_handle_tx_wc(), * netif_tx_lock_bh() and netif_tx_lock().
*/
local_bh_disable();
cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
local_bh_enable();
if (cq->triggered == triggered) return;
}
}
/** * rvt_create_cq - create a completion queue * @ibcq: Allocated CQ * @attr: creation attributes * @attrs: uverbs bundle * * Called by ib_create_cq() in the generic verbs code. * * Return: 0 on success
*/ int rvt_create_cq(struct ib_cq *ibcq, conststruct ib_cq_init_attr *attr, struct uverbs_attr_bundle *attrs)
{ struct ib_udata *udata = &attrs->driver_udata; struct ib_device *ibdev = ibcq->device; struct rvt_dev_info *rdi = ib_to_rvt(ibdev); struct rvt_cq *cq = ibcq_to_rvtcq(ibcq); struct rvt_cq_wc *u_wc = NULL; struct rvt_k_cq_wc *k_wc = NULL;
u32 sz; unsignedint entries = attr->cqe; int comp_vector = attr->comp_vector; int err;
if (attr->flags) return -EOPNOTSUPP;
if (entries < 1 || entries > rdi->dparms.props.max_cqe) return -EINVAL;
/* * 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.
*/ if (udata && udata->outlen >= sizeof(__u64)) {
sz = sizeof(struct ib_uverbs_wc) * (entries + 1);
sz += sizeof(*u_wc);
u_wc = vmalloc_user(sz); if (!u_wc) return -ENOMEM;
} else {
sz = sizeof(struct ib_wc) * (entries + 1);
sz += sizeof(*k_wc);
k_wc = vzalloc_node(sz, rdi->dparms.node); if (!k_wc) return -ENOMEM;
}
/* * Return the address of the WC as the offset to mmap. * See rvt_mmap() for details.
*/ if (udata && udata->outlen >= sizeof(__u64)) {
cq->ip = rvt_create_mmap_info(rdi, sz, udata, u_wc); if (IS_ERR(cq->ip)) {
err = PTR_ERR(cq->ip); goto bail_wc;
}
err = ib_copy_to_udata(udata, &cq->ip->offset, sizeof(cq->ip->offset)); if (err) goto bail_ip;
}
if (cq->ip) {
spin_lock_irq(&rdi->pending_lock);
list_add(&cq->ip->pending_mmaps, &rdi->pending_mmaps);
spin_unlock_irq(&rdi->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->rdi = rdi; if (rdi->driver_f.comp_vect_cpu_lookup)
cq->comp_vector_cpu =
rdi->driver_f.comp_vect_cpu_lookup(rdi, comp_vector); else
cq->comp_vector_cpu =
cpumask_first(cpumask_of_node(rdi->dparms.node));
/** * rvt_destroy_cq - destroy a completion queue * @ibcq: the completion queue to destroy. * @udata: user data or NULL for kernel object * * Called by ib_destroy_cq() in the generic verbs code.
*/ int rvt_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
{ struct rvt_cq *cq = ibcq_to_rvtcq(ibcq); struct rvt_dev_info *rdi = cq->rdi;
/** * rvt_req_notify_cq - change the notification type for a completion queue * @ibcq: the completion queue * @notify_flags: the type of notification to request * * This may be called from interrupt context. Also called by * ib_req_notify_cq() in the generic verbs code. * * Return: 0 for success.
*/ int rvt_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags)
{ struct rvt_cq *cq = ibcq_to_rvtcq(ibcq); unsignedlong flags; int ret = 0;
spin_lock_irqsave(&cq->lock, flags); /* * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2).
*/ if (cq->notify != IB_CQ_NEXT_COMP)
cq->notify = notify_flags & IB_CQ_SOLICITED_MASK;
if (notify_flags & IB_CQ_REPORT_MISSED_EVENTS) { if (cq->queue) { if (RDMA_READ_UAPI_ATOMIC(cq->queue->head) !=
RDMA_READ_UAPI_ATOMIC(cq->queue->tail))
ret = 1;
} else { if (cq->kqueue->head != cq->kqueue->tail)
ret = 1;
}
}
spin_unlock_irqrestore(&cq->lock, flags);
return ret;
}
/* * rvt_resize_cq - change the size of the CQ * @ibcq: the completion queue * * Return: 0 for success.
*/ int rvt_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
{ struct rvt_cq *cq = ibcq_to_rvtcq(ibcq);
u32 head, tail, n; int ret;
u32 sz; struct rvt_dev_info *rdi = cq->rdi; struct rvt_cq_wc *u_wc = NULL; struct rvt_cq_wc *old_u_wc = NULL; struct rvt_k_cq_wc *k_wc = NULL; struct rvt_k_cq_wc *old_k_wc = NULL;
if (cqe < 1 || cqe > rdi->dparms.props.max_cqe) return -EINVAL;
/* * Need to use vmalloc() if we want to support large #s of entries.
*/ if (udata && udata->outlen >= sizeof(__u64)) {
sz = sizeof(struct ib_uverbs_wc) * (cqe + 1);
sz += sizeof(*u_wc);
u_wc = vmalloc_user(sz); if (!u_wc) return -ENOMEM;
} else {
sz = sizeof(struct ib_wc) * (cqe + 1);
sz += sizeof(*k_wc);
k_wc = vzalloc_node(sz, rdi->dparms.node); if (!k_wc) return -ENOMEM;
} /* 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.
*/ if (u_wc) {
old_u_wc = cq->queue;
head = RDMA_READ_UAPI_ATOMIC(old_u_wc->head);
tail = RDMA_READ_UAPI_ATOMIC(old_u_wc->tail);
} else {
old_k_wc = cq->kqueue;
head = old_k_wc->head;
tail = old_k_wc->tail;
}
if (head > (u32)cq->ibcq.cqe)
head = (u32)cq->ibcq.cqe; 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 (u_wc)
u_wc->uqueue[n] = old_u_wc->uqueue[tail]; else
k_wc->kqueue[n] = old_k_wc->kqueue[tail]; if (tail == (u32)cq->ibcq.cqe)
tail = 0; else
tail++;
}
cq->ibcq.cqe = cqe; if (u_wc) {
RDMA_WRITE_UAPI_ATOMIC(u_wc->head, n);
RDMA_WRITE_UAPI_ATOMIC(u_wc->tail, 0);
cq->queue = u_wc;
} else {
k_wc->head = n;
k_wc->tail = 0;
cq->kqueue = k_wc;
}
spin_unlock_irq(&cq->lock);
if (u_wc)
vfree(old_u_wc); else
vfree(old_k_wc);
if (cq->ip) { struct rvt_mmap_info *ip = cq->ip;
rvt_update_mmap_info(rdi, ip, sz, u_wc);
/* * Return the offset to mmap. * See rvt_mmap() for details.
*/ if (udata && udata->outlen >= sizeof(__u64)) {
ret = ib_copy_to_udata(udata, &ip->offset, sizeof(ip->offset)); if (ret) return ret;
}
spin_lock_irq(&rdi->pending_lock); if (list_empty(&ip->pending_mmaps))
list_add(&ip->pending_mmaps, &rdi->pending_mmaps);
spin_unlock_irq(&rdi->pending_lock);
}
/** * rvt_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 * * This may be called from interrupt context. Also called by ib_poll_cq() * in the generic verbs code. * * Return: the number of completion entries polled.
*/ int rvt_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
{ struct rvt_cq *cq = ibcq_to_rvtcq(ibcq); struct rvt_k_cq_wc *wc; unsignedlong flags; int npolled;
u32 tail;
/* The kernel can only poll a kernel completion queue */ if (cq->ip) return -EINVAL;
spin_lock_irqsave(&cq->lock, flags);
wc = cq->kqueue;
tail = wc->tail; if (tail > (u32)cq->ibcq.cqe)
tail = (u32)cq->ibcq.cqe; for (npolled = 0; npolled < num_entries; ++npolled, ++entry) { if (tail == wc->head) break; /* The kernel doesn't need a RMB since it has the lock. */
trace_rvt_cq_poll(cq, &wc->kqueue[tail], npolled);
*entry = wc->kqueue[tail]; if (tail >= cq->ibcq.cqe)
tail = 0; else
tail++;
}
wc->tail = tail;
spin_unlock_irqrestore(&cq->lock, flags);
return npolled;
}
/** * rvt_driver_cq_init - Init cq resources on behalf of driver * * Return: 0 on success
*/ int rvt_driver_cq_init(void)
{
comp_vector_wq = alloc_workqueue("%s", WQ_HIGHPRI | WQ_CPU_INTENSIVE,
0, "rdmavt_cq"); if (!comp_vector_wq) return -ENOMEM;
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.