示例#1
0
static inline ssize_t
usdf_cq_copy_soft_entry(void *dst, const struct usdf_cq_soft_entry *src,
			enum fi_cq_format dst_format)
{
	struct fi_cq_entry *ctx_entry;
	struct fi_cq_msg_entry *msg_entry;
	struct fi_cq_data_entry *data_entry;

	switch (dst_format) {
	case FI_CQ_FORMAT_CONTEXT:
		ctx_entry = (struct fi_cq_entry *)dst;
		ctx_entry->op_context = src->cse_context;
		break;
	case FI_CQ_FORMAT_MSG:
		msg_entry = (struct fi_cq_msg_entry *)dst;
		msg_entry->op_context = src->cse_context;
		msg_entry->flags = src->cse_flags;
		msg_entry->len = src->cse_len;
		break;
	case FI_CQ_FORMAT_DATA:
		data_entry = (struct fi_cq_data_entry *)dst;
		data_entry->op_context = src->cse_context;
		data_entry->flags = src->cse_flags;
		data_entry->len = src->cse_len;
		data_entry->buf = src->cse_buf;
		data_entry->data = src->cse_data;
		break;
	default:
		USDF_WARN("unexpected CQ format, internal error\n");
		return -FI_EOPNOTSUPP;
	}

	return FI_SUCCESS;
}
示例#2
0
文件: usdf_cq.c 项目: ORNL/ompi
/*
 * poll a soft CQ
 * This will loop over all the hard CQs within, collecting results.
 * Since this routine is an inline and is always called with format as
 * a constant, I am counting on the compiler optimizing away all the switches
 * on format.
 */
static inline ssize_t
usdf_cq_read_common_soft(struct fid_cq *fcq, void *buf, size_t count,
		enum fi_cq_format format)
{
	struct usdf_cq *cq;
	uint8_t *entry;
	uint8_t *last;
	void *tail;
	size_t entry_len;

	cq = cq_ftou(fcq);
	if (cq->cq_comp.uc_status != 0) {
		return -FI_EAVAIL;
	}

	/* progress... */
	usdf_domain_progress(cq->cq_domain);

	switch (format) {
	case FI_CQ_FORMAT_CONTEXT:
		entry_len = sizeof(struct fi_cq_entry);
		break;
	case FI_CQ_FORMAT_MSG:
		entry_len = sizeof(struct fi_cq_msg_entry);
		break;
	case FI_CQ_FORMAT_DATA:
		entry_len = sizeof(struct fi_cq_data_entry);
		break;
	default:
		USDF_WARN("unexpected CQ format, internal error\n");
		return -FI_EOPNOTSUPP;
	}

	entry = buf;
	last = entry + (entry_len * count);
	tail = cq->c.soft.cq_tail;

	// XXX ... handle error comps
	while (entry < last && tail != cq->c.soft.cq_head) {
		memcpy(entry, tail, entry_len);
		entry += entry_len;

		tail = (uint8_t *)tail + entry_len;
		if (tail == cq->c.soft.cq_end) {
			tail = cq->c.soft.cq_comps;
		}
	}
	cq->c.soft.cq_tail = tail;

	if (entry > (uint8_t *)buf) {
		return (entry - (uint8_t *)buf) / entry_len;
	} else {
		return -FI_EAGAIN;
	}
}
示例#3
0
/* Checks that the given address is actually a sockaddr_in of appropriate
 * length.  "addr_format" is an FI_ constant like FI_SOCKADDR_IN indicating the
 * claimed type of the given address.
 *
 * Returns true if address is actually a sockaddr_in, false otherwise.
 *
 * Upon successful return, "addr" can be safely cast to either
 * "struct sockaddr_in *" or "struct sockaddr *".
 *
 * "addr" should not be NULL.
 */
bool usdf_cm_addr_is_valid_sin(void *addr, size_t addrlen, uint32_t addr_format)
{
	assert(addr != NULL);

	switch (addr_format) {
	case FI_SOCKADDR_IN:
	case FI_SOCKADDR:
		if (addrlen != sizeof(struct sockaddr_in)) {
			USDF_WARN("addrlen is incorrect\n");
			return false;
		}
		if (((struct sockaddr *)addr)->sa_family != AF_INET) {
			USDF_WARN("unknown/unsupported addr_format\n");
			return false;
		}
		return true;
	default:
		USDF_WARN("unknown/unsupported addr_format\n");
		return false;
	}
}
示例#4
0
ssize_t
usdf_msg_inject(struct fid_ep *fep, const void *buf, size_t len,
		fi_addr_t dest_addr)
{
	struct usdf_ep *ep;
	struct usdf_tx *tx;
	struct usdf_msg_qe *wqe;
	struct usdf_domain *udp;

	if (len > USDF_MSG_MAX_INJECT_SIZE) {
		USDF_WARN("cannot inject more than inject_size bytes\n");
		return -EINVAL;
	}

	ep = ep_ftou(fep);
	tx = ep->ep_tx;
	udp = ep->ep_domain;

	if (TAILQ_EMPTY(&tx->t.msg.tx_free_wqe)) {
		return -FI_EAGAIN;
	}

	pthread_spin_lock(&udp->dom_progress_lock);

	wqe = TAILQ_FIRST(&tx->t.msg.tx_free_wqe);
	TAILQ_REMOVE(&tx->t.msg.tx_free_wqe, wqe, ms_link);

	wqe->ms_context = NULL;
	memcpy(wqe->ms_inject_buf, buf, len);
	wqe->ms_iov[0].iov_base = wqe->ms_inject_buf;
	wqe->ms_iov[0].iov_len = len;
	wqe->ms_last_iov = 0;

	wqe->ms_cur_iov = 0;
	wqe->ms_cur_ptr = buf;
	wqe->ms_iov_resid = len;
	wqe->ms_resid = len;
	wqe->ms_length = len;

	/* fi_inject() never signals a completion */
	wqe->ms_signal_comp = 0;

	/* add send to EP, and add EP to TX list if not present */
	TAILQ_INSERT_TAIL(&ep->e.msg.ep_posted_wqe, wqe, ms_link);
	usdf_msg_ep_ready(ep);

	pthread_spin_unlock(&udp->dom_progress_lock);

	usdf_domain_progress(udp);

	return 0;
}
示例#5
0
static inline ssize_t
usdf_cq_copy_cq_entry(void *dst, struct usd_completion *src,
			enum fi_cq_format format)
{
	struct fi_cq_entry *ctx_entry;
	struct fi_cq_msg_entry *msg_entry;
	struct fi_cq_data_entry *data_entry;

	switch (format) {
	case FI_CQ_FORMAT_CONTEXT:
		ctx_entry = (struct fi_cq_entry *)dst;
		ctx_entry->op_context = src->uc_context;
		break;
	case FI_CQ_FORMAT_MSG:
		msg_entry = (struct fi_cq_msg_entry *)dst;
		msg_entry->op_context = src->uc_context;
		msg_entry->flags = usdf_cqe_to_flags(src);
		msg_entry->len = src->uc_bytes;

		usdf_cq_adjust_len(src, &msg_entry->len);

		break;
	case FI_CQ_FORMAT_DATA:
		data_entry = (struct fi_cq_data_entry *)dst;
		data_entry->op_context = src->uc_context;
		data_entry->flags = usdf_cqe_to_flags(src);
		data_entry->len = src->uc_bytes;
		data_entry->buf = 0; /* XXX */
		data_entry->data = 0;

		usdf_cq_adjust_len(src, &data_entry->len);

		break;
	default:
		USDF_WARN("unexpected CQ format, internal error\n");
		return -FI_EOPNOTSUPP;
	}

	return FI_SUCCESS;
}
示例#6
0
/*
 * poll a soft CQ
 * This will loop over all the hard CQs within, collecting results.
 * Since this routine is an inline and is always called with format as
 * a constant, I am counting on the compiler optimizing away all the switches
 * on format.
 */
static inline ssize_t
usdf_cq_read_common_soft(struct fid_cq *fcq, void *buf, size_t count,
		enum fi_cq_format format)
{
	struct usdf_cq *cq;
	uint8_t *entry;
	uint8_t *last;
	struct usdf_cq_soft_entry *tail;
	size_t entry_len;
	ssize_t ret;

	cq = cq_ftou(fcq);
	if (cq->cq_comp.uc_status != 0) {
		return -FI_EAVAIL;
	}

	/* progress... */
	usdf_domain_progress(cq->cq_domain);

	switch (format) {
	case FI_CQ_FORMAT_CONTEXT:
		entry_len = sizeof(struct fi_cq_entry);
		break;
	case FI_CQ_FORMAT_MSG:
		entry_len = sizeof(struct fi_cq_msg_entry);
		break;
	case FI_CQ_FORMAT_DATA:
		entry_len = sizeof(struct fi_cq_data_entry);
		break;
	default:
		USDF_WARN("unexpected CQ format, internal error\n");
		return -FI_EOPNOTSUPP;
	}

	entry = buf;
	last = entry + (entry_len * count);
	tail = cq->c.soft.cq_tail;

	while (entry < last) {
		/* If the head and tail are equal and the last
		 * operation was a read then that means we have an
		 * empty queue.
		 */
		if ((tail == cq->c.soft.cq_head) &&
				(cq->c.soft.cq_last_op == USDF_SOFT_CQ_READ))
			break;

		if (tail->cse_prov_errno > 0) {
			if (entry > (uint8_t *) buf)
				break;
			else
				return -FI_EAVAIL;
		}
		ret = usdf_cq_copy_soft_entry(entry, tail, format);
		if (ret < 0) {
			return ret;
		}
		entry += entry_len;
		tail++;
		if (tail == cq->c.soft.cq_end) {
			tail = cq->c.soft.cq_comps;
		}

		cq->c.soft.cq_last_op = USDF_SOFT_CQ_READ;
	}
	cq->c.soft.cq_tail = tail;

	if (entry > (uint8_t *)buf) {
		return (entry - (uint8_t *)buf) / entry_len;
	} else {
		return -FI_EAGAIN;
	}
}
示例#7
0
static ssize_t
usdf_cq_sread_common_soft(struct fid_cq *fcq, void *buf, size_t count, const void *cond,
		int timeout_ms, enum fi_cq_format format)
{
	struct usdf_cq *cq;
	uint8_t *entry;
	uint8_t *last;
	struct usdf_cq_soft_entry *tail;
	size_t entry_len;
	size_t sleep_time_us;
	size_t time_spent_us = 0;
	ssize_t ret;

	cq = cq_ftou(fcq);

	if (cq->cq_attr.wait_obj == FI_WAIT_NONE)
		return -FI_EOPNOTSUPP;

	sleep_time_us = SREAD_INIT_SLEEP_TIME_US;

	switch (format) {
	case FI_CQ_FORMAT_CONTEXT:
		entry_len = sizeof(struct fi_cq_entry);
		break;
	case FI_CQ_FORMAT_MSG:
		entry_len = sizeof(struct fi_cq_msg_entry);
		break;
	case FI_CQ_FORMAT_DATA:
		entry_len = sizeof(struct fi_cq_data_entry);
		break;
	default:
		USDF_WARN("unexpected CQ format, internal error\n");
		return -FI_EOPNOTSUPP;
	}

	entry = buf;
	last = entry + (entry_len * count);

	while (1) {
		/* progress... */
		usdf_domain_progress(cq->cq_domain);

		tail = cq->c.soft.cq_tail;

		while (entry < last) {
			/* If the head and tail are equal and the last
			 * operation was a read then that means we have an
			 * empty queue.
			 */
			if ((tail == cq->c.soft.cq_head) &&
					(cq->c.soft.cq_last_op ==
					 USDF_SOFT_CQ_READ))
				break;

			if (tail->cse_prov_errno > 0) {
				if (entry > (uint8_t *)buf)
					break;
				else
					return -FI_EAVAIL;
			}

			ret = usdf_cq_copy_soft_entry(entry, tail, format);
			if (ret < 0)
				return ret;

			entry += entry_len;
			tail++;
			if (tail == cq->c.soft.cq_end)
				tail = cq->c.soft.cq_comps;

			cq->c.soft.cq_last_op = USDF_SOFT_CQ_READ;
		}

		if (entry > (uint8_t *)buf) {
			cq->c.soft.cq_tail = tail;
			return (entry - (uint8_t *)buf) / entry_len;
		} else {
			if (timeout_ms >= 0 &&
				(time_spent_us >= 1000 * timeout_ms))
				break;

			usleep(sleep_time_us);
			time_spent_us += sleep_time_us;

			/* exponentially back off up to a limit */
			if (sleep_time_us < SREAD_MAX_SLEEP_TIME_US)
				sleep_time_us *= SREAD_EXP_BASE;
			sleep_time_us = MIN(sleep_time_us,
						SREAD_MAX_SLEEP_TIME_US);
		}
	}

	return -FI_EAGAIN;
}
示例#8
0
static int
usdf_getinfo(uint32_t version, const char *node, const char *service,
	       uint64_t flags, struct fi_info *hints, struct fi_info **info)
{
	struct usdf_usnic_info *dp;
	struct usdf_dev_entry *dep;
	struct usd_device_attrs *dap;
	struct fi_info *fi_first;
	struct fi_info *fi_last;
	struct addrinfo *ai;
	struct sockaddr_in *src;
	struct sockaddr_in *dest;
	enum fi_ep_type ep_type;
	int metric;
	int d;
	int ret;

	USDF_TRACE("\n");

	fi_first = NULL;
	fi_last = NULL;
	ai = NULL;
	src = NULL;
	dest = NULL;

	/*
	 * Get and cache usNIC device info
	 */
	if (__usdf_devinfo == NULL) {
		ret = usdf_get_devinfo();
		if (ret != 0) {
			USDF_WARN("failed to usdf_get_devinfo, ret=%d (%s)\n",
					ret, fi_strerror(-ret));
			if (ret == -FI_ENODEV)
				ret = -FI_ENODATA;
			goto fail;
		}
	}
	dp = __usdf_devinfo;

	if (node != NULL || service != NULL) {
		ret = getaddrinfo(node, service, NULL, &ai);
		if (ret != 0) {
			USDF_DBG("getaddrinfo failed, likely bad node/service specified (%s:%s)\n",
				node, service);
			ret = -errno;
			goto fail;
		}
		if (flags & FI_SOURCE) {
			src = (struct sockaddr_in *)ai->ai_addr;
		} else {
			dest = (struct sockaddr_in *)ai->ai_addr;
		}
	}
	if (hints != NULL) {
		if (dest == NULL && hints->dest_addr != NULL) {
			dest = hints->dest_addr;
		}
		if (src == NULL && hints->src_addr != NULL) {
			src = hints->src_addr;
		}
	}

	for (d = 0; d < dp->uu_num_devs; ++d) {
		dep = &dp->uu_info[d];
		dap = &dep->ue_dattr;

		/* skip this device if it has some problem */
		if (!dep->ue_dev_ok) {
			USDF_DBG("skipping %s/%s\n", dap->uda_devname,
				dap->uda_ifname);
			continue;
		}

		/* See if dest is reachable from this device */
		if (dest != NULL && dest->sin_addr.s_addr != INADDR_ANY) {
			ret = usdf_get_distance(dap,
					dest->sin_addr.s_addr, &metric);
			if (ret != 0) {
				goto fail;
			}
			if (metric == -1) {
				USDF_DBG("dest %s unreachable from %s/%s, skipping\n",
					inet_ntoa(dest->sin_addr),
					dap->uda_devname, dap->uda_ifname);
				continue;
			}
		}

		/* Does this device match requested attributes? */
		if (hints != NULL) {
			ret = usdf_validate_hints(hints, dap);
			if (ret != 0) {
				USDF_DBG("hints do not match for %s/%s, skipping\n",
					dap->uda_devname, dap->uda_ifname);
				continue;
			}

			ep_type = hints->ep_attr ? hints->ep_attr->type :
				  FI_EP_UNSPEC;
		} else {
			ep_type = FI_EP_UNSPEC;
		}

		if (ep_type == FI_EP_DGRAM || ep_type == FI_EP_UNSPEC) {
			ret = usdf_fill_info_dgram(version, hints, src, dest,
					dap, &fi_first, &fi_last);
			if (ret != 0 && ret != -FI_ENODATA) {
				goto fail;
			}
		}

		if (ep_type == FI_EP_MSG || ep_type == FI_EP_UNSPEC) {
			ret = usdf_fill_info_msg(hints, src, dest, dap,
					&fi_first, &fi_last);
			if (ret != 0 && ret != -FI_ENODATA) {
				goto fail;
			}
		}

		if (ep_type == FI_EP_RDM || ep_type == FI_EP_UNSPEC) {
			ret = usdf_fill_info_rdm(hints, src, dest, dap,
					&fi_first, &fi_last);
			if (ret != 0 && ret != -FI_ENODATA) {
				goto fail;
			}
		}
	}

	if (fi_first != NULL) {
		*info = fi_first;
		ret = 0;
	} else {
		ret = -FI_ENODATA;
	}

fail:
	if (ret != 0) {
		fi_freeinfo(fi_first);
	}
	if (ai != NULL) {
		freeaddrinfo(ai);
	}
	if (ret != 0) {
		USDF_INFO("returning %d (%s)\n", ret, fi_strerror(-ret));
	}
	return ret;
}
示例#9
0
static int
usdf_getinfo(uint32_t version, const char *node, const char *service,
	       uint64_t flags, const struct fi_info *hints, struct fi_info **info)
{
	struct usdf_usnic_info *dp;
	struct usdf_dev_entry *dep;
	struct usd_device_attrs *dap;
	struct fi_info *fi_first;
	struct fi_info *fi_last;
	struct addrinfo *ai;
	void *src;
	void *dest;
	enum fi_ep_type ep_type;
	int d;
	int ret;

	USDF_TRACE("\n");

	fi_first = NULL;
	fi_last = NULL;
	ai = NULL;
	src = NULL;
	dest = NULL;

	/*
	 * Get and cache usNIC device info
	 */
	if (__usdf_devinfo == NULL) {
		ret = usdf_get_devinfo();
		if (ret != 0) {
			USDF_WARN("failed to usdf_get_devinfo, ret=%d (%s)\n",
					ret, fi_strerror(-ret));
			if (ret == -FI_ENODEV)
				ret = -FI_ENODATA;
			goto fail;
		}
	}
	dp = __usdf_devinfo;

	/* Check the hints up front and fail if they're invalid. */
	if (hints) {
		ret = usdf_validate_hints(version, hints);
		if (ret) {
			USDF_WARN_SYS(FABRIC, "hints failed to validate\n");
			goto fail;
		}
	}

	/* Get the src and dest if user specified. */
	ret = usdf_handle_node_and_service(node, service, flags,
					   &src, &dest, hints, &ai);
	if (ret) {
		USDF_WARN_SYS(FABRIC, "failed to handle node and service.\n");
		goto fail;
	}

	if (hints != NULL) {
		if (dest == NULL && hints->dest_addr != NULL)
			dest = hints->dest_addr;
		if (src == NULL && hints->src_addr != NULL)
			src = hints->src_addr;
	}

	for (d = 0; d < dp->uu_num_devs; ++d) {
		dep = &dp->uu_info[d];
		dap = &dep->ue_dattr;

		/* If the device has an issue or the hints don't match the
		 * device information, then skip.
		 */
		if (!usdf_check_device(version, hints, src, dest, dep))
			continue;

		if (hints && hints->ep_attr)
			ep_type = hints->ep_attr->type;
		else
			ep_type = FI_EP_UNSPEC;

		if (ep_type == FI_EP_DGRAM || ep_type == FI_EP_UNSPEC) {
			ret = usdf_fill_info_dgram(version, hints, src, dest,
					dap, &fi_first, &fi_last);
			if (ret != 0 && ret != -FI_ENODATA) {
				goto fail;
			}
		}

		if (ep_type == FI_EP_MSG || ep_type == FI_EP_UNSPEC) {
			ret = usdf_fill_info_msg(version, hints, src, dest,
					dap, &fi_first, &fi_last);
			if (ret != 0 && ret != -FI_ENODATA) {
				goto fail;
			}
		}

		if (ep_type == FI_EP_RDM || ep_type == FI_EP_UNSPEC) {
			ret = usdf_fill_info_rdm(version, hints, src, dest,
					dap, &fi_first, &fi_last);
			if (ret != 0 && ret != -FI_ENODATA) {
				goto fail;
			}
		}
	}

	if (fi_first != NULL) {
		*info = fi_first;
		ret = 0;
	} else {
		ret = -FI_ENODATA;
	}


fail:
	if (ai)
		freeaddrinfo(ai);

	if (ret != 0) {
		fi_freeinfo(fi_first);
		USDF_INFO("returning %d (%s)\n", ret, fi_strerror(-ret));
	}

	return ret;
}