RE: [PATCH v3 05/13] SoftiWarp application interface

["Steve Wise" <swise-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>]

> 
> Signed-off-by: Bernard Metzler <bmt-OA+xvbQnYDHMbYB6QlFGEg@public.gmane.org>
> ---
>  drivers/infiniband/sw/siw/siw_ae.c    |  120 +++
>  drivers/infiniband/sw/siw/siw_verbs.c | 1876
> +++++++++++++++++++++++++++++++++
>  drivers/infiniband/sw/siw/siw_verbs.h |  119 +++
>  include/uapi/rdma/siw_user.h          |  216 ++++
>  4 files changed, 2331 insertions(+)
>  create mode 100644 drivers/infiniband/sw/siw/siw_ae.c
>  create mode 100644 drivers/infiniband/sw/siw/siw_verbs.c
>  create mode 100644 drivers/infiniband/sw/siw/siw_verbs.h
>  create mode 100644 include/uapi/rdma/siw_user.h
> 
> diff --git a/drivers/infiniband/sw/siw/siw_ae.c
> b/drivers/infiniband/sw/siw/siw_ae.c
> new file mode 100644
> index 000000000000..eac55ce7f56d
> --- /dev/null
> +++ b/drivers/infiniband/sw/siw/siw_ae.c
> @@ -0,0 +1,120 @@
> +/*
> + * Software iWARP device driver
> + *
> + * Authors: Bernard Metzler <bmt-OA+xvbQnYDHMbYB6QlFGEg@public.gmane.org>
> + *
> + * Copyright (c) 2008-2017, IBM Corporation
> + *
> + * This software is available to you under a choice of one of two
> + * licenses. You may choose to be licensed under the terms of the GNU
> + * General Public License (GPL) Version 2, available from the file
> + * COPYING in the main directory of this source tree, or the
> + * BSD license below:
> + *
> + *   Redistribution and use in source and binary forms, with or
> + *   without modification, are permitted provided that the following
> + *   conditions are met:
> + *
> + *   - Redistributions of source code must retain the above copyright
notice,
> + *     this list of conditions and the following disclaimer.
> + *
> + *   - Redistributions in binary form must reproduce the above copyright
> + *     notice, this list of conditions and the following disclaimer in
the
> + *     documentation and/or other materials provided with the
distribution.
> + *
> + *   - Neither the name of IBM nor the names of its contributors may be
> + *     used to endorse or promote products derived from this software
without
> + *     specific prior written permission.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
> + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
> OF
> + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
> + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
> HOLDERS
> + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
> + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR
> IN
> + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
> THE
> + * SOFTWARE.
> + */
> +
> +#include <linux/errno.h>
> +#include <linux/types.h>
> +#include <linux/net.h>
> +#include <linux/scatterlist.h>
> +#include <linux/highmem.h>
> +#include <net/sock.h>
> +#include <net/tcp_states.h>
> +#include <net/tcp.h>
> +
> +#include <rdma/iw_cm.h>
> +#include <rdma/ib_verbs.h>
> +#include <rdma/ib_smi.h>
> +#include <rdma/ib_user_verbs.h>
> +
> +#include "siw.h"
> +#include "siw_obj.h"
> +#include "siw_cm.h"
> +
> +void siw_qp_event(struct siw_qp *qp, enum ib_event_type etype)
> +{
> +	struct ib_event event;
> +	struct ib_qp	*base_qp = &qp->base_qp;
> +
> +	/*
> +	 * Do not report asynchronous errors on QP which gets
> +	 * destroyed via verbs interface (siw_destroy_qp())
> +	 */
> +	if (qp->attrs.flags & SIW_QP_IN_DESTROY)
> +		return;
> +
> +	event.event = etype;
> +	event.device = base_qp->device;
> +	event.element.qp = base_qp;
> +
> +	if (base_qp->event_handler) {
> +		siw_dbg_qp(qp, "reporting event %d\n", etype);
> +		(*base_qp->event_handler)(&event, base_qp->qp_context);
> +	}
> +}
> +
> +void siw_cq_event(struct siw_cq *cq, enum ib_event_type etype)
> +{
> +	struct ib_event event;
> +	struct ib_cq	*base_cq = &cq->base_cq;
> +
> +	event.event = etype;
> +	event.device = base_cq->device;
> +	event.element.cq = base_cq;
> +
> +	if (base_cq->event_handler) {
> +		siw_dbg(cq->hdr.sdev, "reporting CQ event %d\n", etype);
> +		(*base_cq->event_handler)(&event, base_cq->cq_context);
> +	}
> +}
> +

The rdma provider must ensure that event upcalls are serialized per object.
Is this being done somewhere?  See "Callbacks" in
Documentation/infiniband/core_locking.txt.



> +void siw_srq_event(struct siw_srq *srq, enum ib_event_type etype)
> +{
> +	struct ib_event event;
> +	struct ib_srq	*base_srq = &srq->base_srq;
> +
> +	event.event = etype;
> +	event.device = base_srq->device;
> +	event.element.srq = base_srq;
> +
> +	if (base_srq->event_handler) {
> +		siw_dbg(srq->pd->hdr.sdev, "reporting SRQ event %d\n",
> etype);
> +		(*base_srq->event_handler)(&event, base_srq->srq_context);
> +	}
> +}
> +
> +void siw_port_event(struct siw_device *sdev, u8 port, enum ib_event_type
> etype)
> +{
> +	struct ib_event event;
> +
> +	event.event = etype;
> +	event.device = &sdev->base_dev;
> +	event.element.port_num = port;
> +
> +	siw_dbg(sdev, "reporting port event %d\n", etype);
> +
> +	ib_dispatch_event(&event);
> +}
> diff --git a/drivers/infiniband/sw/siw/siw_verbs.c
> b/drivers/infiniband/sw/siw/siw_verbs.c
> new file mode 100644
> index 000000000000..667b004e5c6d
> --- /dev/null
> +++ b/drivers/infiniband/sw/siw/siw_verbs.c
> @@ -0,0 +1,1876 @@
> +/*
> + * Software iWARP device driver
> + *
> + * Authors: Bernard Metzler <bmt-OA+xvbQnYDHMbYB6QlFGEg@public.gmane.org>
> + *
> + * Copyright (c) 2008-2017, IBM Corporation
> + *
> + * This software is available to you under a choice of one of two
> + * licenses. You may choose to be licensed under the terms of the GNU
> + * General Public License (GPL) Version 2, available from the file
> + * COPYING in the main directory of this source tree, or the
> + * BSD license below:
> + *
> + *   Redistribution and use in source and binary forms, with or
> + *   without modification, are permitted provided that the following
> + *   conditions are met:
> + *
> + *   - Redistributions of source code must retain the above copyright
notice,
> + *     this list of conditions and the following disclaimer.
> + *
> + *   - Redistributions in binary form must reproduce the above copyright
> + *     notice, this list of conditions and the following disclaimer in
the
> + *     documentation and/or other materials provided with the
distribution.
> + *
> + *   - Neither the name of IBM nor the names of its contributors may be
> + *     used to endorse or promote products derived from this software
without
> + *     specific prior written permission.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
> + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
> OF
> + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
> + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
> HOLDERS
> + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
> + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR
> IN
> + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
> THE
> + * SOFTWARE.
> + */
> +
> +#include <linux/errno.h>
> +#include <linux/types.h>
> +#include <linux/uaccess.h>
> +#include <linux/vmalloc.h>
> +
> +#include <rdma/iw_cm.h>
> +#include <rdma/ib_verbs.h>
> +#include <rdma/ib_smi.h>
> +#include <rdma/ib_user_verbs.h>
> +
> +#include "siw.h"
> +#include "siw_verbs.h"
> +#include "siw_obj.h"
> +#include "siw_cm.h"
> +
> +static int ib_qp_state_to_siw_qp_state[IB_QPS_ERR+1] = {
> +	[IB_QPS_RESET]	= SIW_QP_STATE_IDLE,
> +	[IB_QPS_INIT]	= SIW_QP_STATE_IDLE,
> +	[IB_QPS_RTR]	= SIW_QP_STATE_RTR,
> +	[IB_QPS_RTS]	= SIW_QP_STATE_RTS,
> +	[IB_QPS_SQD]	= SIW_QP_STATE_CLOSING,
> +	[IB_QPS_SQE]	= SIW_QP_STATE_TERMINATE,
> +	[IB_QPS_ERR]	= SIW_QP_STATE_ERROR
> +};
> +
> +static struct siw_pd *siw_pd_base2siw(struct ib_pd *base_pd)
> +{
> +	return container_of(base_pd, struct siw_pd, base_pd);
> +}
> +
> +static struct siw_ucontext *siw_ctx_base2siw(struct ib_ucontext
*base_ctx)
> +{
> +	return container_of(base_ctx, struct siw_ucontext, ib_ucontext);
> +}
> +
> +static struct siw_cq *siw_cq_base2siw(struct ib_cq *base_cq)
> +{
> +	return container_of(base_cq, struct siw_cq, base_cq);
> +}
> +
> +static struct siw_srq *siw_srq_base2siw(struct ib_srq *base_srq)
> +{
> +	return container_of(base_srq, struct siw_srq, base_srq);
> +}
> +
> +static u32 siw_insert_uobj(struct siw_ucontext *uctx, void *vaddr, u32
size)
> +{
> +	struct siw_uobj *uobj;
> +	u32	key;
> +
> +	uobj = kzalloc(sizeof(*uobj), GFP_KERNEL);
> +	if (!uobj)
> +		return SIW_INVAL_UOBJ_KEY;
> +
> +	size = PAGE_ALIGN(size);
> +
> +	spin_lock(&uctx->uobj_lock);
> +
> +	if (list_empty(&uctx->uobj_list))
> +		uctx->uobj_key = 0;
> +
> +	key = uctx->uobj_key;
> +	if (key > SIW_MAX_UOBJ_KEY) {
> +		spin_unlock(&uctx->uobj_lock);
> +		kfree(uobj);
> +		return SIW_INVAL_UOBJ_KEY;
> +	}
> +	uobj->key = key;
> +	uobj->size = size;
> +	uobj->addr = vaddr;
> +
> +	uctx->uobj_key += size; /* advance for next object */
> +
> +	list_add_tail(&uobj->list, &uctx->uobj_list);
> +
> +	spin_unlock(&uctx->uobj_lock);
> +
> +	return key;
> +}
> +
> +static struct siw_uobj *siw_remove_uobj(struct siw_ucontext *uctx, u32
key,
> +					u32 size)
> +{
> +	struct list_head *pos, *nxt;
> +
> +	spin_lock(&uctx->uobj_lock);
> +
> +	list_for_each_safe(pos, nxt, &uctx->uobj_list) {
> +		struct siw_uobj *uobj = list_entry(pos, struct siw_uobj,
list);
> +
> +		if (uobj->key == key && uobj->size == size) {
> +			list_del(&uobj->list);
> +			spin_unlock(&uctx->uobj_lock);
> +			return uobj;
> +		}
> +	}
> +	spin_unlock(&uctx->uobj_lock);
> +
> +	return NULL;
> +}
> +
> +int siw_mmap(struct ib_ucontext *ctx, struct vm_area_struct *vma)
> +{
> +	struct siw_ucontext     *uctx = siw_ctx_base2siw(ctx);
> +	struct siw_uobj         *uobj;
> +	u32     key = vma->vm_pgoff << PAGE_SHIFT;
> +	int     size = vma->vm_end - vma->vm_start;
> +	int     rv = -EINVAL;
> +
> +	/*
> +	 * Must be page aligned
> +	 */
> +	if (vma->vm_start & (PAGE_SIZE - 1)) {
> +		pr_warn("map not page aligned\n");
> +		goto out;
> +	}
> +
> +	uobj = siw_remove_uobj(uctx, key, size);
> +	if (!uobj) {
> +		pr_warn("mmap lookup failed: %u, %d\n", key, size);
> +		goto out;
> +	}
> +	rv = remap_vmalloc_range(vma, uobj->addr, 0);
> +	if (rv)
> +		pr_warn("remap_vmalloc_range failed: %u, %d\n", key, size);
> +
> +	kfree(uobj);
> +out:
> +	return rv;
> +}
> +
> +struct ib_ucontext *siw_alloc_ucontext(struct ib_device *base_dev,
> +				       struct ib_udata *udata)
> +{
> +	struct siw_ucontext *ctx = NULL;
> +	struct siw_device *sdev = siw_dev_base2siw(base_dev);
> +	int rv;
> +
> +	if (atomic_inc_return(&sdev->num_ctx) > SIW_MAX_CONTEXT) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
> +	if (!ctx) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	spin_lock_init(&ctx->uobj_lock);
> +	INIT_LIST_HEAD(&ctx->uobj_list);
> +	ctx->uobj_key = 0;
> +
> +	ctx->sdev = sdev;
> +	if (udata) {
> +		struct siw_uresp_alloc_ctx uresp;
> +
> +		memset(&uresp, 0, sizeof(uresp));
> +		uresp.dev_id = sdev->attrs.vendor_part_id;
> +
> +		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
> +		if (rv)
> +			goto err_out;
> +	}
> +	siw_dbg(sdev, "success. now %d context(s)\n",
> +		atomic_read(&sdev->num_ctx));
> +
> +	return &ctx->ib_ucontext;
> +
> +err_out:
> +	kfree(ctx);
> +
> +	atomic_dec(&sdev->num_ctx);
> +	siw_dbg(sdev, "failure %d. now %d context(s)\n",
> +		rv, atomic_read(&sdev->num_ctx));
> +
> +	return ERR_PTR(rv);
> +}
> +
> +int siw_dealloc_ucontext(struct ib_ucontext *base_ctx)
> +{
> +	struct siw_ucontext *ctx = siw_ctx_base2siw(base_ctx);
> +
> +	atomic_dec(&ctx->sdev->num_ctx);
> +	kfree(ctx);
> +	return 0;
> +}
> +
> +int siw_query_device(struct ib_device *base_dev, struct ib_device_attr
*attr,
> +		     struct ib_udata *unused)
> +{
> +	struct siw_device *sdev = siw_dev_base2siw(base_dev);
> +	/*
> +	 * A process context is needed to report avail memory resources.
> +	 */
> +	if (in_interrupt())
> +		return -EINVAL;
> +
> +	memset(attr, 0, sizeof(*attr));
> +
> +	attr->max_mr_size = rlimit(RLIMIT_MEMLOCK); /* per process */
> +	attr->vendor_id = sdev->attrs.vendor_id;
> +	attr->vendor_part_id = sdev->attrs.vendor_part_id;
> +	attr->max_qp = sdev->attrs.max_qp;
> +	attr->max_qp_wr = sdev->attrs.max_qp_wr;
> +
> +	attr->max_qp_rd_atom = sdev->attrs.max_ord;
> +	attr->max_qp_init_rd_atom = sdev->attrs.max_ird;
> +	attr->max_res_rd_atom = sdev->attrs.max_qp * sdev->attrs.max_ird;
> +	attr->device_cap_flags = sdev->attrs.cap_flags;
> +	attr->max_sge = sdev->attrs.max_sge;
> +	attr->max_sge_rd = sdev->attrs.max_sge_rd;
> +	attr->max_cq = sdev->attrs.max_cq;
> +	attr->max_cqe = sdev->attrs.max_cqe;
> +	attr->max_mr = sdev->attrs.max_mr;
> +	attr->max_pd = sdev->attrs.max_pd;
> +	attr->max_mw = sdev->attrs.max_mw;
> +	attr->max_fmr = sdev->attrs.max_fmr;
> +	attr->max_srq = sdev->attrs.max_srq;
> +	attr->max_srq_wr = sdev->attrs.max_srq_wr;
> +	attr->max_srq_sge = sdev->attrs.max_srq_sge;
> +	attr->max_fast_reg_page_list_len = SIW_MAX_SGE_PBL;
> +	attr->page_size_cap = PAGE_SIZE;
> +	/* Revisit if RFC 7306 gets supported */
> +	attr->atomic_cap = 0;
> +
> +	memcpy(&attr->sys_image_guid, sdev->netdev->dev_addr, 6);
> +
> +	return 0;
> +}
> +
> +/*
> + * Approximate translation of real MTU for IB.
> + */
> +static inline enum ib_mtu siw_mtu_net2base(unsigned short mtu)
> +{
> +	if (mtu >= 4096)
> +		return IB_MTU_4096;
> +	if (mtu >= 2048)
> +		return IB_MTU_2048;
> +	if (mtu >= 1024)
> +		return IB_MTU_1024;
> +	if (mtu >= 512)
> +		return IB_MTU_512;
> +	if (mtu >= 256)
> +		return IB_MTU_256;
> +	return IB_MTU_4096;
> +}
> +
> +int siw_query_port(struct ib_device *base_dev, u8 port,
> +		     struct ib_port_attr *attr)
> +{
> +	struct siw_device *sdev = siw_dev_base2siw(base_dev);
> +
> +	memset(attr, 0, sizeof(*attr));
> +
> +	attr->state = sdev->state;
> +	attr->max_mtu = siw_mtu_net2base(sdev->netdev->mtu);
> +	attr->active_mtu = attr->max_mtu;
> +	attr->gid_tbl_len = 1;
> +	attr->port_cap_flags = IB_PORT_CM_SUP;
> +	attr->port_cap_flags |= IB_PORT_DEVICE_MGMT_SUP;
> +	attr->max_msg_sz = -1;
> +	attr->pkey_tbl_len = 1;
> +	attr->active_width = 2;
> +	attr->active_speed = 2;
> +	attr->phys_state = sdev->state == IB_PORT_ACTIVE ? 5 : 3;
> +	/*
> +	 * All zero
> +	 *
> +	 * attr->lid = 0;
> +	 * attr->bad_pkey_cntr = 0;
> +	 * attr->qkey_viol_cntr = 0;
> +	 * attr->sm_lid = 0;
> +	 * attr->lmc = 0;
> +	 * attr->max_vl_num = 0;
> +	 * attr->sm_sl = 0;
> +	 * attr->subnet_timeout = 0;
> +	 * attr->init_type_repy = 0;
> +	 */
> +	return 0;
> +}
> +
> +int siw_get_port_immutable(struct ib_device *base_dev, u8 port,
> +			   struct ib_port_immutable *port_immutable)
> +{
> +	struct ib_port_attr attr;
> +	int rv = siw_query_port(base_dev, port, &attr);
> +
> +	if (rv)
> +		return rv;
> +
> +	port_immutable->pkey_tbl_len = attr.pkey_tbl_len;
> +	port_immutable->gid_tbl_len = attr.gid_tbl_len;
> +	port_immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
> +
> +	return 0;
> +}
> +
> +int siw_query_pkey(struct ib_device *base_dev, u8 port, u16 idx, u16
*pkey)
> +{
> +	/* Report the default pkey */
> +	*pkey = 0xffff;
> +	return 0;
> +}
> +
> +int siw_query_gid(struct ib_device *base_dev, u8 port, int idx,
> +		   union ib_gid *gid)
> +{
> +	struct siw_device *sdev = siw_dev_base2siw(base_dev);
> +
> +	/* subnet_prefix == interface_id == 0; */
> +	memset(gid, 0, sizeof(*gid));
> +	memcpy(&gid->raw[0], sdev->netdev->dev_addr, 6);
> +
> +	return 0;
> +}
> +
> +struct ib_pd *siw_alloc_pd(struct ib_device *base_dev,
> +			   struct ib_ucontext *context, struct ib_udata
*udata)
> +{
> +	struct siw_pd *pd = NULL;
> +	struct siw_device *sdev  = siw_dev_base2siw(base_dev);
> +	int rv;
> +
> +	if (atomic_inc_return(&sdev->num_pd) > SIW_MAX_PD) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	pd = kmalloc(sizeof(*pd), GFP_KERNEL);
> +	if (!pd) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	rv = siw_pd_add(sdev, pd);
> +	if (rv) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	if (context) {
> +		if (ib_copy_to_udata(udata, &pd->hdr.id,
sizeof(pd->hdr.id))) {
> +			rv = -EFAULT;
> +			goto err_out_idr;
> +		}
> +	}
> +	siw_dbg(sdev, "success. now %d pd's(s)\n",
> +		atomic_read(&sdev->num_pd));
> +	return &pd->base_pd;
> +
> +err_out_idr:
> +	siw_remove_obj(&sdev->idr_lock, &sdev->pd_idr, &pd->hdr);
> +err_out:
> +	kfree(pd);
> +	atomic_dec(&sdev->num_pd);
> +	siw_dbg(sdev, "failure %d. now %d pd's(s)\n",
> +		rv, atomic_read(&sdev->num_pd));
> +
> +	return ERR_PTR(rv);
> +}
> +
> +int siw_dealloc_pd(struct ib_pd *base_pd)
> +{
> +	struct siw_pd *pd = siw_pd_base2siw(base_pd);
> +	struct siw_device *sdev = siw_dev_base2siw(base_pd->device);
> +
> +	siw_remove_obj(&sdev->idr_lock, &sdev->pd_idr, &pd->hdr);
> +	siw_pd_put(pd);
> +
> +	return 0;
> +}
> +
> +void siw_qp_get_ref(struct ib_qp *base_qp)
> +{
> +	struct siw_qp	*qp = siw_qp_base2siw(base_qp);
> +
> +	siw_dbg_qp(qp, "get user reference\n");
> +	siw_qp_get(qp);
> +}
> +
> +void siw_qp_put_ref(struct ib_qp *base_qp)
> +{
> +	struct siw_qp	*qp = siw_qp_base2siw(base_qp);
> +
> +	siw_dbg_qp(qp, "put user reference\n");
> +	siw_qp_put(qp);
> +}
> +
> +int siw_no_mad(struct ib_device *base_dev, int flags, u8 port,
> +	       const struct ib_wc *wc, const struct ib_grh *grh,
> +	       const struct ib_mad_hdr *in_mad, size_t in_mad_size,
> +	       struct ib_mad_hdr *out_mad, size_t *out_mad_size,
> +	       u16 *outmad_pkey_index)
> +{
> +	return -EOPNOTSUPP;
> +}
> +
> +/*
> + * siw_create_qp()
> + *
> + * Create QP of requested size on given device.
> + *
> + * @base_pd:	Base PD contained in siw PD
> + * @attrs:	Initial QP attributes.
> + * @udata:	used to provide QP ID, SQ and RQ size back to user.
> + */
> +
> +struct ib_qp *siw_create_qp(struct ib_pd *base_pd,
> +			    struct ib_qp_init_attr *attrs,
> +			    struct ib_udata *udata)
> +{
> +	struct siw_qp		*qp = NULL;
> +	struct siw_pd		*pd = siw_pd_base2siw(base_pd);
> +	struct ib_device	 *base_dev = base_pd->device;
> +	struct siw_device	*sdev = siw_dev_base2siw(base_dev);
> +	struct siw_cq		*scq = NULL, *rcq = NULL;
> +
> +	unsigned long flags;
> +	int num_sqe, num_rqe, rv = 0;
> +
> +	siw_dbg(sdev, "create new qp\n");
> +
> +	if (atomic_inc_return(&sdev->num_qp) > SIW_MAX_QP) {
> +		siw_dbg(sdev, "too many qp's\n");
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	if (attrs->qp_type != IB_QPT_RC) {
> +		siw_dbg(sdev, "only rc qp's supported\n");
> +		rv = -EINVAL;
> +		goto err_out;
> +	}
> +	if ((attrs->cap.max_send_wr > SIW_MAX_QP_WR) ||
> +	    (attrs->cap.max_recv_wr > SIW_MAX_QP_WR) ||
> +	    (attrs->cap.max_send_sge > SIW_MAX_SGE)  ||
> +	    (attrs->cap.max_recv_sge > SIW_MAX_SGE)) {
> +		siw_dbg(sdev, "qp size error\n");
> +		rv = -EINVAL;
> +		goto err_out;
> +	}
> +	if (attrs->cap.max_inline_data > SIW_MAX_INLINE) {
> +		siw_dbg(sdev, "max inline send: %d > %d\n",
> +			attrs->cap.max_inline_data, (int)SIW_MAX_INLINE);
> +		rv = -EINVAL;
> +		goto err_out;
> +	}
> +	/*
> +	 * NOTE: we allow for zero element SQ and RQ WQE's SGL's
> +	 * but not for a QP unable to hold any WQE (SQ + RQ)
> +	 */
> +	if (attrs->cap.max_send_wr + attrs->cap.max_recv_wr == 0) {
> +		siw_dbg(sdev, "qp must have send or receive queue\n");
> +		rv = -EINVAL;
> +		goto err_out;
> +	}
> +
> +	scq = siw_cq_id2obj(sdev, ((struct siw_cq
*)attrs->send_cq)->hdr.id);
> +	rcq = siw_cq_id2obj(sdev, ((struct siw_cq
*)attrs->recv_cq)->hdr.id);
> +
> +	if (!scq || (!rcq && !attrs->srq)) {
> +		siw_dbg(sdev, "send cq or receive cq invalid\n");
> +		rv = -EINVAL;
> +		goto err_out;
> +	}
> +	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
> +	if (!qp) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +
> +	init_rwsem(&qp->state_lock);
> +	spin_lock_init(&qp->sq_lock);
> +	spin_lock_init(&qp->rq_lock);
> +	spin_lock_init(&qp->orq_lock);
> +
> +	init_waitqueue_head(&qp->tx_ctx.waitq);
> +
> +	if (!base_pd->uobject)
> +		qp->kernel_verbs = 1;
> +
> +	rv = siw_qp_add(sdev, qp);
> +	if (rv)
> +		goto err_out;
> +
> +	num_sqe = roundup_pow_of_two(attrs->cap.max_send_wr);
> +	num_rqe = roundup_pow_of_two(attrs->cap.max_recv_wr);
> +
> +	if (qp->kernel_verbs)
> +		qp->sendq = vzalloc(num_sqe * sizeof(struct siw_sqe));
> +	else
> +		qp->sendq = vmalloc_user(num_sqe * sizeof(struct siw_sqe));
> +
> +	if (qp->sendq == NULL) {
> +		siw_dbg_qp(qp, "send queue size %d alloc failed\n",
num_sqe);
> +		rv = -ENOMEM;
> +		goto err_out_idr;
> +	}
> +	if (attrs->sq_sig_type != IB_SIGNAL_REQ_WR) {
> +		if (attrs->sq_sig_type == IB_SIGNAL_ALL_WR)
> +			qp->attrs.flags |= SIW_SIGNAL_ALL_WR;
> +		else {
> +			rv = -EINVAL;
> +			goto err_out_idr;
> +		}
> +	}
> +	qp->pd  = pd;
> +	qp->scq = scq;
> +	qp->rcq = rcq;
> +
> +	if (attrs->srq) {
> +		/*
> +		 * SRQ support.
> +		 * Verbs 6.3.7: ignore RQ size, if SRQ present
> +		 * Verbs 6.3.5: do not check PD of SRQ against PD of QP
> +		 */
> +		qp->srq = siw_srq_base2siw(attrs->srq);
> +		qp->attrs.rq_size = 0;
> +		siw_dbg_qp(qp, "[SRQ 0x%p] attached\n", qp->srq);
> +	} else if (num_rqe) {
> +		if (qp->kernel_verbs)
> +			qp->recvq = vzalloc(num_rqe * sizeof(struct
siw_rqe));
> +		else
> +			qp->recvq = vmalloc_user(num_rqe *
> +						 sizeof(struct siw_rqe));
> +
> +		if (qp->recvq == NULL) {
> +			siw_dbg_qp(qp, "recv queue size %d alloc failed\n",
> +				   num_rqe);
> +			rv = -ENOMEM;
> +			goto err_out_idr;
> +		}
> +
> +		qp->attrs.rq_size = num_rqe;
> +	}
> +	qp->attrs.sq_size = num_sqe;
> +	qp->attrs.sq_max_sges = attrs->cap.max_send_sge;
> +	/*
> +	 * ofed has no max_send_sge_rdmawrite
> +	 */
> +	qp->attrs.sq_max_sges_rdmaw = attrs->cap.max_send_sge;
> +	qp->attrs.rq_max_sges = attrs->cap.max_recv_sge;
> +
> +	qp->attrs.state = SIW_QP_STATE_IDLE;
> +
> +	if (udata) {
> +		struct siw_uresp_create_qp uresp;
> +		struct siw_ucontext *ctx;
> +
> +		memset(&uresp, 0, sizeof(uresp));
> +		ctx = siw_ctx_base2siw(base_pd->uobject->context);
> +
> +		uresp.sq_key = uresp.rq_key = SIW_INVAL_UOBJ_KEY;
> +		uresp.num_sqe = num_sqe;
> +		uresp.num_rqe = num_rqe;
> +		uresp.qp_id = QP_ID(qp);
> +
> +		if (qp->sendq) {
> +			uresp.sq_key = siw_insert_uobj(ctx, qp->sendq,
> +					num_sqe * sizeof(struct siw_sqe));
> +			if (uresp.sq_key > SIW_MAX_UOBJ_KEY)
> +				siw_dbg_qp(qp, "preparing mmap sq
failed\n");
> +		}
> +		if (qp->recvq) {
> +			uresp.rq_key = siw_insert_uobj(ctx, qp->recvq,
> +					num_rqe * sizeof(struct siw_rqe));
> +			if (uresp.rq_key > SIW_MAX_UOBJ_KEY)
> +				siw_dbg_qp(qp, "preparing mmap rq
failed\n");
> +		}
> +		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
> +		if (rv)
> +			goto err_out_idr;
> +	}
> +	qp->tx_cpu = siw_get_tx_cpu(sdev);
> +	if (qp->tx_cpu < 0) {
> +		rv = -EINVAL;
> +		goto err_out_idr;
> +	}
> +	atomic_set(&qp->tx_ctx.in_use, 0);
> +
> +	qp->base_qp.qp_num = QP_ID(qp);
> +
> +	siw_pd_get(pd);
> +
> +	INIT_LIST_HEAD(&qp->devq);
> +	spin_lock_irqsave(&sdev->idr_lock, flags);
> +	list_add_tail(&qp->devq, &sdev->qp_list);
> +	spin_unlock_irqrestore(&sdev->idr_lock, flags);
> +
> +	return &qp->base_qp;
> +
> +err_out_idr:
> +	siw_remove_obj(&sdev->idr_lock, &sdev->qp_idr, &qp->hdr);
> +err_out:
> +	if (scq)
> +		siw_cq_put(scq);
> +	if (rcq)
> +		siw_cq_put(rcq);
> +
> +	if (qp) {
> +		if (qp->sendq)
> +			vfree(qp->sendq);
> +		if (qp->recvq)
> +			vfree(qp->recvq);
> +		kfree(qp);
> +	}
> +	atomic_dec(&sdev->num_qp);
> +
> +	return ERR_PTR(rv);
> +}
> +
> +/*
> + * Minimum siw_query_qp() verb interface.
> + *
> + * @qp_attr_mask is not used but all available information is provided
> + */
> +int siw_query_qp(struct ib_qp *base_qp, struct ib_qp_attr *qp_attr,
> +		 int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
> +{
> +	struct siw_qp *qp;
> +	struct siw_device *sdev;
> +
> +	if (base_qp && qp_attr && qp_init_attr) {
> +		qp = siw_qp_base2siw(base_qp);
> +		sdev = siw_dev_base2siw(base_qp->device);
> +	} else
> +		return -EINVAL;
> +
> +	qp_attr->cap.max_inline_data = SIW_MAX_INLINE;
> +	qp_attr->cap.max_send_wr = qp->attrs.sq_size;
> +	qp_attr->cap.max_send_sge = qp->attrs.sq_max_sges;
> +	qp_attr->cap.max_recv_wr = qp->attrs.rq_size;
> +	qp_attr->cap.max_recv_sge = qp->attrs.rq_max_sges;
> +	qp_attr->path_mtu = siw_mtu_net2base(sdev->netdev->mtu);
> +	qp_attr->max_rd_atomic = qp->attrs.irq_size;
> +	qp_attr->max_dest_rd_atomic = qp->attrs.orq_size;
> +
> +	qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE |
> +			IB_ACCESS_REMOTE_WRITE |
> IB_ACCESS_REMOTE_READ;
> +
> +	qp_init_attr->qp_type = base_qp->qp_type;
> +	qp_init_attr->send_cq = base_qp->send_cq;
> +	qp_init_attr->recv_cq = base_qp->recv_cq;
> +	qp_init_attr->srq = base_qp->srq;
> +
> +	qp_init_attr->cap = qp_attr->cap;
> +
> +	return 0;
> +}
> +
> +int siw_verbs_modify_qp(struct ib_qp *base_qp, struct ib_qp_attr *attr,
> +			int attr_mask, struct ib_udata *udata)
> +{
> +	struct siw_qp_attrs	new_attrs;
> +	enum siw_qp_attr_mask	siw_attr_mask = 0;
> +	struct siw_qp		*qp = siw_qp_base2siw(base_qp);
> +	int			rv = 0;
> +
> +	if (!attr_mask)
> +		return 0;
> +
> +	memset(&new_attrs, 0, sizeof(new_attrs));
> +
> +	if (attr_mask & IB_QP_ACCESS_FLAGS) {
> +
> +		siw_attr_mask = SIW_QP_ATTR_ACCESS_FLAGS;
> +
> +		if (attr->qp_access_flags & IB_ACCESS_REMOTE_READ)
> +			new_attrs.flags |= SIW_RDMA_READ_ENABLED;
> +		if (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE)
> +			new_attrs.flags |= SIW_RDMA_WRITE_ENABLED;
> +		if (attr->qp_access_flags & IB_ACCESS_MW_BIND)
> +			new_attrs.flags |= SIW_RDMA_BIND_ENABLED;
> +	}
> +	if (attr_mask & IB_QP_STATE) {
> +		siw_dbg_qp(qp, "desired ib qp state: %s\n",
> +			   ib_qp_state_to_string[attr->qp_state]);
> +
> +		new_attrs.state = ib_qp_state_to_siw_qp_state[attr-
> >qp_state];
> +
> +		if (new_attrs.state > SIW_QP_STATE_RTS)
> +			qp->tx_ctx.tx_suspend = 1;
> +
> +		siw_attr_mask |= SIW_QP_ATTR_STATE;
> +	}
> +	if (!attr_mask)
> +		goto out;
> +
> +	down_write(&qp->state_lock);
> +
> +	rv = siw_qp_modify(qp, &new_attrs, siw_attr_mask);
> +
> +	up_write(&qp->state_lock);
> +out:
> +	return rv;
> +}
> +
> +int siw_destroy_qp(struct ib_qp *base_qp)
> +{
> +	struct siw_qp		*qp = siw_qp_base2siw(base_qp);
> +	struct siw_qp_attrs	qp_attrs;
> +
> +	siw_dbg_qp(qp, "state %d, cep 0x%p\n", qp->attrs.state, qp->cep);
> +
> +	/*
> +	 * Mark QP as in process of destruction to prevent from
> +	 * any async callbacks to RDMA core
> +	 */
> +	qp->attrs.flags |= SIW_QP_IN_DESTROY;
> +	qp->rx_ctx.rx_suspend = 1;
> +
> +	down_write(&qp->state_lock);
> +
> +	qp_attrs.state = SIW_QP_STATE_ERROR;
> +	(void)siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE);
> +
> +	if (qp->cep) {
> +		siw_cep_put(qp->cep);
> +		qp->cep = NULL;
> +	}
> +
> +	up_write(&qp->state_lock);
> +
> +	kfree(qp->rx_ctx.mpa_crc_hd);
> +	kfree(qp->tx_ctx.mpa_crc_hd);
> +
> +	/* Drop references */
> +	siw_cq_put(qp->scq);
> +	siw_cq_put(qp->rcq);
> +	siw_pd_put(qp->pd);
> +	qp->scq = qp->rcq = NULL;
> +
> +	siw_qp_put(qp);
> +
> +	return 0;
> +}
> +
> +/*
> + * siw_copy_sgl()
> + *
> + * Copy SGL from RDMA core representation to local
> + * representation.
> + */
> +static inline void siw_copy_sgl(struct ib_sge *sge, struct siw_sge
*siw_sge,
> +			       int num_sge)
> +{
> +	while (num_sge--) {
> +		siw_sge->laddr = sge->addr;
> +		siw_sge->length = sge->length;
> +		siw_sge->lkey = sge->lkey;
> +
> +		siw_sge++; sge++;
> +	}
> +}
> +
> +/*
> + * siw_copy_inline_sgl()
> + *
> + * Prepare sgl of inlined data for sending. For userland callers
> + * function checks if given buffer addresses and len's are within
> + * process context bounds.
> + * Data from all provided sge's are copied together into the wqe,
> + * referenced by a single sge.
> + */
> +static int siw_copy_inline_sgl(struct ib_send_wr *core_wr, struct siw_sqe
*sqe)
> +{
> +	struct ib_sge	*core_sge = core_wr->sg_list;
> +	void		*kbuf	 = &sqe->sge[1];
> +	int		num_sge	 = core_wr->num_sge,
> +			bytes	 = 0;
> +
> +	sqe->sge[0].laddr = (u64)kbuf;
> +	sqe->sge[0].lkey = 0;
> +
> +	while (num_sge--) {
> +		if (!core_sge->length) {
> +			core_sge++;
> +			continue;
> +		}
> +		bytes += core_sge->length;
> +		if (bytes > SIW_MAX_INLINE) {
> +			bytes = -EINVAL;
> +			break;
> +		}
> +		memcpy(kbuf, (void *)(uintptr_t)core_sge->addr,
> +		       core_sge->length);
> +
> +		kbuf += core_sge->length;
> +		core_sge++;
> +	}
> +	sqe->sge[0].length = bytes > 0 ? bytes : 0;
> +	sqe->num_sge = bytes > 0 ? 1 : 0;
> +
> +	return bytes;
> +}
> +
> +/*
> + * siw_post_send()
> + *
> + * Post a list of S-WR's to a SQ.
> + *
> + * @base_qp:	Base QP contained in siw QP
> + * @wr:		Null terminated list of user WR's
> + * @bad_wr:	Points to failing WR in case of synchronous failure.
> + */
> +int siw_post_send(struct ib_qp *base_qp, struct ib_send_wr *wr,
> +		  struct ib_send_wr **bad_wr)
> +{
> +	struct siw_qp	*qp = siw_qp_base2siw(base_qp);
> +	struct siw_wqe	*wqe = tx_wqe(qp);
> +
> +	unsigned long flags;
> +	int rv = 0;
> +
> +	siw_dbg_qp(qp, "state %d\n", qp->attrs.state);
> +
> +	/*
> +	 * Try to acquire QP state lock. Must be non-blocking
> +	 * to accommodate kernel clients needs.
> +	 */
> +	if (!down_read_trylock(&qp->state_lock)) {
> +		*bad_wr = wr;
> +		return -ENOTCONN;
> +	}
> +

Under what conditions does down_read_trylock() return 0?  This seems like a
kernel ULP that has multiple threads posting to the sq might get an error
due to lock contention?  

> +	if (unlikely(qp->attrs.state != SIW_QP_STATE_RTS)) {
> +		up_read(&qp->state_lock);
> +		*bad_wr = wr;
> +		return -ENOTCONN;
> +	}

NOTCONN is inappropriate here too.  Also, I relaxed iw_cxgb4 to support
ib_drain_qp().  Basically if the kernel QP is in ERROR,  a post will be
completed via a CQE inserted with status IB_WC_WR_FLUSH_ERR.  See how
ib_drain_qp() works for details.  The reason I bring this up, is that both
iser and nvmf ULPs use ib_drain_sq/rq/qp() so you might want to support
them.  There is an alternate solution though.  ib_drain_sq/rq() allows the
low level driver to have its own drain implementation.  See those functions
for details.


> +	if (wr && !qp->kernel_verbs) {
> +		siw_dbg_qp(qp, "wr must be empty for user mapped sq\n");
> +		up_read(&qp->state_lock);
> +		*bad_wr = wr;
> +		return -EINVAL;
> +	}
> +
> +	spin_lock_irqsave(&qp->sq_lock, flags);
> +
> +	while (wr) {
> +		u32 idx = qp->sq_put % qp->attrs.sq_size;
> +		struct siw_sqe *sqe = &qp->sendq[idx];
> +
> +		if (sqe->flags) {
> +			siw_dbg_qp(qp, "sq full\n");
> +			rv = -ENOMEM;
> +			break;
> +		}
> +		if (wr->num_sge > qp->attrs.sq_max_sges) {
> +			siw_dbg_qp(qp, "too many sge's: %d\n", wr-
> >num_sge);
> +			rv = -EINVAL;
> +			break;
> +		}
> +		sqe->id = wr->wr_id;
> +
> +		if ((wr->send_flags & IB_SEND_SIGNALED) ||
> +		    (qp->attrs.flags & SIW_SIGNAL_ALL_WR))
> +			sqe->flags |= SIW_WQE_SIGNALLED;
> +
> +		if (wr->send_flags & IB_SEND_FENCE)
> +			sqe->flags |= SIW_WQE_READ_FENCE;
> +
> +		switch (wr->opcode) {
> +
> +		case IB_WR_SEND:
> +		case IB_WR_SEND_WITH_INV:
> +			if (wr->send_flags & IB_SEND_SOLICITED)
> +				sqe->flags |= SIW_WQE_SOLICITED;
> +
> +			if (!(wr->send_flags & IB_SEND_INLINE)) {
> +				siw_copy_sgl(wr->sg_list, sqe->sge,
> +					     wr->num_sge);
> +				sqe->num_sge = wr->num_sge;
> +			} else {
> +				rv = siw_copy_inline_sgl(wr, sqe);
> +				if (rv <= 0) {
> +					rv = -EINVAL;
> +					break;
> +				}
> +				sqe->flags |= SIW_WQE_INLINE;
> +				sqe->num_sge = 1;
> +			}
> +			if (wr->opcode == IB_WR_SEND)
> +				sqe->opcode = SIW_OP_SEND;
> +			else {
> +				sqe->opcode = SIW_OP_SEND_REMOTE_INV;
> +				sqe->rkey = wr->ex.invalidate_rkey;
> +			}
> +			break;
> +
> +		case IB_WR_RDMA_READ_WITH_INV:
> +		case IB_WR_RDMA_READ:
> +			/*
> +			 * OFED WR restricts RREAD sink to SGL containing
> +			 * 1 SGE only. we could relax to SGL with multiple
> +			 * elements referring the SAME ltag or even sending
> +			 * a private per-rreq tag referring to a checked
> +			 * local sgl with MULTIPLE ltag's. would be easy
> +			 * to do...
> +			 */
> +			if (unlikely(wr->num_sge != 1)) {
> +				rv = -EINVAL;
> +				break;
> +			}
> +			siw_copy_sgl(wr->sg_list, &sqe->sge[0], 1);
> +			/*
> +			 * NOTE: zero length RREAD is allowed!
> +			 */
> +			sqe->raddr	= rdma_wr(wr)->remote_addr;
> +			sqe->rkey	= rdma_wr(wr)->rkey;
> +			sqe->num_sge	= 1;
> +
> +			if (wr->opcode == IB_WR_RDMA_READ)
> +				sqe->opcode = SIW_OP_READ;
> +			else
> +				sqe->opcode = SIW_OP_READ_LOCAL_INV;
> +			break;
> +
> +		case IB_WR_RDMA_WRITE:
> +			if (!(wr->send_flags & IB_SEND_INLINE)) {
> +				siw_copy_sgl(wr->sg_list, &sqe->sge[0],
> +					     wr->num_sge);
> +				sqe->num_sge = wr->num_sge;
> +			} else {
> +				rv = siw_copy_inline_sgl(wr, sqe);
> +				if (unlikely(rv < 0)) {
> +					rv = -EINVAL;
> +					break;
> +				}
> +				sqe->flags |= SIW_WQE_INLINE;
> +				sqe->num_sge = 1;
> +			}
> +			sqe->raddr	= rdma_wr(wr)->remote_addr;
> +			sqe->rkey	= rdma_wr(wr)->rkey;
> +			sqe->opcode	= SIW_OP_WRITE;
> +
> +			break;
> +
> +		case IB_WR_REG_MR:
> +			sqe->base_mr = (uint64_t)reg_wr(wr)->mr;
> +			sqe->rkey = reg_wr(wr)->key;
> +			sqe->access = SIW_MEM_LREAD;
> +			if (reg_wr(wr)->access & IB_ACCESS_LOCAL_WRITE)
> +				sqe->access |= SIW_MEM_LWRITE;
> +			if (reg_wr(wr)->access & IB_ACCESS_REMOTE_WRITE)
> +				sqe->access |= SIW_MEM_RWRITE;
> +			if (reg_wr(wr)->access & IB_ACCESS_REMOTE_READ)
> +				sqe->access |= SIW_MEM_RREAD;
> +			sqe->opcode = SIW_OP_REG_MR;
> +
> +			break;
> +
> +		case IB_WR_LOCAL_INV:
> +			sqe->rkey = wr->ex.invalidate_rkey;
> +			sqe->opcode = SIW_OP_INVAL_STAG;
> +
> +			break;
> +
> +		default:
> +			siw_dbg_qp(qp, "ib wr type %d unsupported\n",
> +				   wr->opcode);
> +			rv = -EINVAL;
> +			break;
> +		}
> +		siw_dbg_qp(qp, "opcode %d, flags 0x%x\n",
> +			   sqe->opcode, sqe->flags);
> +
> +		if (unlikely(rv < 0))
> +			break;
> +
> +		/* make SQE only vaild after completely written */
> +		smp_wmb();
> +		sqe->flags |= SIW_WQE_VALID;
> +
> +		qp->sq_put++;
> +		wr = wr->next;
> +	}
> +
> +	/*
> +	 * Send directly if SQ processing is not in progress.
> +	 * Eventual immediate errors (rv < 0) do not affect the involved
> +	 * RI resources (Verbs, 8.3.1) and thus do not prevent from SQ
> +	 * processing, if new work is already pending. But rv must be passed
> +	 * to caller.
> +	 */
> +	if (wqe->wr_status != SIW_WR_IDLE) {
> +		spin_unlock_irqrestore(&qp->sq_lock, flags);
> +		goto skip_direct_sending;
> +	}
> +	rv = siw_activate_tx(qp);
> +	spin_unlock_irqrestore(&qp->sq_lock, flags);
> +
> +	if (rv <= 0)
> +		goto skip_direct_sending;
> +
> +	if (qp->kernel_verbs) {
> +		rv = siw_sq_start(qp);
> +	} else {
> +		qp->tx_ctx.in_syscall = 1;
> +
> +		if (siw_qp_sq_process(qp) != 0 && !(qp->tx_ctx.tx_suspend))
> +			siw_qp_cm_drop(qp, 0);
> +
> +		qp->tx_ctx.in_syscall = 0;
> +	}
> +skip_direct_sending:
> +
> +	up_read(&qp->state_lock);
> +
> +	if (rv >= 0)
> +		return 0;
> +	/*
> +	 * Immediate error
> +	 */
> +	siw_dbg_qp(qp, "error %d\n", rv);
> +
> +	*bad_wr = wr;
> +	return rv;
> +}
> +
> +/*
> + * siw_post_receive()
> + *
> + * Post a list of R-WR's to a RQ.
> + *
> + * @base_qp:	Base QP contained in siw QP
> + * @wr:		Null terminated list of user WR's
> + * @bad_wr:	Points to failing WR in case of synchronous failure.
> + */
> +int siw_post_receive(struct ib_qp *base_qp, struct ib_recv_wr *wr,
> +		     struct ib_recv_wr **bad_wr)
> +{
> +	struct siw_qp	*qp = siw_qp_base2siw(base_qp);
> +	int rv = 0;
> +
> +	siw_dbg_qp(qp, "state %d\n", qp->attrs.state);
> +
> +	if (qp->srq) {
> +		*bad_wr = wr;
> +		return -EOPNOTSUPP; /* what else from errno.h? */
> +	}
> +	/*
> +	 * Try to acquire QP state lock. Must be non-blocking
> +	 * to accommodate kernel clients needs.
> +	 */
> +	if (!down_read_trylock(&qp->state_lock)) {
> +		*bad_wr = wr;
> +		return -ENOTCONN;
> +	}

Same comment here on trylock...


> +	if (!qp->kernel_verbs) {
> +		siw_dbg_qp(qp, "no kernel post_recv for user mapped sq\n");
> +		up_read(&qp->state_lock);
> +		*bad_wr = wr;
> +		return -EINVAL;
> +	}
> +	if (qp->attrs.state > SIW_QP_STATE_RTS) {
> +		up_read(&qp->state_lock);
> +		*bad_wr = wr;
> +		return -EINVAL;
> +	}
> +	while (wr) {
> +		u32 idx = qp->rq_put % qp->attrs.rq_size;
> +		struct siw_rqe *rqe = &qp->recvq[idx];
> +
> +		if (rqe->flags) {
> +			siw_dbg_qp(qp, "rq full\n");
> +			rv = -ENOMEM;
> +			break;
> +		}
> +		if (wr->num_sge > qp->attrs.rq_max_sges) {
> +			siw_dbg_qp(qp, "too many sge's: %d\n", wr-
> >num_sge);
> +			rv = -EINVAL;
> +			break;
> +		}
> +		rqe->id = wr->wr_id;
> +		rqe->num_sge = wr->num_sge;
> +		siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
> +
> +		/* make sure RQE is completely written before valid */
> +		smp_wmb();
> +
> +		rqe->flags = SIW_WQE_VALID;
> +
> +		qp->rq_put++;
> +		wr = wr->next;
> +	}
> +	if (rv < 0) {
> +		siw_dbg_qp(qp, "error %d\n", rv);
> +		*bad_wr = wr;
> +	}
> +	up_read(&qp->state_lock);
> +
> +	return rv > 0 ? 0 : rv;
> +}
> +
> +int siw_destroy_cq(struct ib_cq *base_cq)
> +{
> +	struct siw_cq		*cq  = siw_cq_base2siw(base_cq);
> +	struct ib_device	*base_dev = base_cq->device;
> +	struct siw_device	*sdev = siw_dev_base2siw(base_dev);
> +
> +	siw_cq_flush(cq);
> +
> +	siw_remove_obj(&sdev->idr_lock, &sdev->cq_idr, &cq->hdr);
> +	siw_cq_put(cq);
> +
> +	return 0;
> +}
> +
> +/*
> + * siw_create_cq()
> + *
> + * Create CQ of requested size on given device.
> + *
> + * @base_dev:	RDMA device contained in siw device
> + * @size:	maximum number of CQE's allowed.
> + * @ib_context: user context.
> + * @udata:	used to provide CQ ID back to user.
> + */
> +
> +struct ib_cq *siw_create_cq(struct ib_device *base_dev,
> +			    const struct ib_cq_init_attr *attr,
> +			    struct ib_ucontext *ib_context,
> +			    struct ib_udata *udata)
> +{
> +	struct siw_cq			*cq = NULL;
> +	struct siw_device		*sdev = siw_dev_base2siw(base_dev);
> +	struct siw_uresp_create_cq	uresp;
> +	int rv, size = attr->cqe;
> +
> +	if (!base_dev) {
> +		rv = -ENODEV;
> +		goto err_out;
> +	}
> +	if (atomic_inc_return(&sdev->num_cq) > SIW_MAX_CQ) {
> +		siw_dbg(sdev, "too many cq's\n");
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	if (size < 1 || size > SIW_MAX_CQE) {
> +		siw_dbg(sdev, "cq size error: %d\n", size);
> +		rv = -EINVAL;
> +		goto err_out;
> +	}
> +	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
> +	if (!cq) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	size = roundup_pow_of_two(size);
> +	cq->base_cq.cqe = size;
> +	cq->num_cqe = size;
> +
> +	if (!ib_context) {
> +		cq->kernel_verbs = 1;
> +		cq->queue = vzalloc(size * sizeof(struct siw_cqe)
> +				+ sizeof(struct siw_cq_ctrl));
> +	} else
> +		cq->queue = vmalloc_user(size * sizeof(struct siw_cqe)
> +				+ sizeof(struct siw_cq_ctrl));
> +
> +	if (cq->queue == NULL) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +
> +	rv = siw_cq_add(sdev, cq);
> +	if (rv)
> +		goto err_out;
> +
> +	spin_lock_init(&cq->lock);
> +
> +	cq->notify = &((struct siw_cq_ctrl *)&cq->queue[size])->notify;
> +
> +	if (!cq->kernel_verbs) {
> +		struct siw_ucontext *ctx = siw_ctx_base2siw(ib_context);
> +
> +		uresp.cq_key = siw_insert_uobj(ctx, cq->queue,
> +					size * sizeof(struct siw_cqe) +
> +					sizeof(struct siw_cq_ctrl));
> +
> +		if (uresp.cq_key > SIW_MAX_UOBJ_KEY)
> +			siw_dbg(sdev, "[CQ %d]: preparing mmap rq failed\n",
> +				OBJ_ID(cq));
> +
> +		uresp.cq_id = OBJ_ID(cq);
> +		uresp.num_cqe = size;
> +
> +		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
> +		if (rv)
> +			goto err_out_idr;
> +	}
> +	return &cq->base_cq;
> +
> +err_out_idr:
> +	siw_remove_obj(&sdev->idr_lock, &sdev->cq_idr, &cq->hdr);
> +err_out:
> +	siw_dbg(sdev, "cq creation failed: %d", rv);
> +
> +	if (cq && cq->queue)
> +		vfree(cq->queue);
> +
> +	kfree(cq);
> +	atomic_dec(&sdev->num_cq);
> +
> +	return ERR_PTR(rv);
> +}
> +
> +/*
> + * siw_poll_cq()
> + *
> + * Reap CQ entries if available and copy work completion status into
> + * array of WC's provided by caller. Returns number of reaped CQE's.
> + *
> + * @base_cq:	Base CQ contained in siw CQ.
> + * @num_cqe:	Maximum number of CQE's to reap.
> + * @wc:		Array of work completions to be filled by siw.
> + */
> +int siw_poll_cq(struct ib_cq *base_cq, int num_cqe, struct ib_wc *wc)
> +{
> +	struct siw_cq		*cq  = siw_cq_base2siw(base_cq);
> +	int			i;
> +
> +	for (i = 0; i < num_cqe; i++) {
> +		if (!(siw_reap_cqe(cq, wc)))
> +			break;
> +		wc++;
> +	}
> +	return i;
> +}
> +
> +/*
> + * siw_req_notify_cq()
> + *
> + * Request notification for new CQE's added to that CQ.
> + * Defined flags:
> + * o SIW_CQ_NOTIFY_SOLICITED lets siw trigger a notification
> + *   event if a WQE with notification flag set enters the CQ
> + * o SIW_CQ_NOTIFY_NEXT_COMP lets siw trigger a notification
> + *   event if a WQE enters the CQ.
> + * o IB_CQ_REPORT_MISSED_EVENTS: return value will provide the
> + *   number of not reaped CQE's regardless of its notification
> + *   type and current or new CQ notification settings.
> + *
> + * @base_cq:	Base CQ contained in siw CQ.
> + * @flags:	Requested notification flags.
> + */
> +int siw_req_notify_cq(struct ib_cq *base_cq, enum ib_cq_notify_flags
flags)
> +{
> +	struct siw_cq	 *cq  = siw_cq_base2siw(base_cq);
> +
> +	siw_dbg(cq->hdr.sdev, "[CQ %d]: flags: 0x%8x\n", OBJ_ID(cq), flags);
> +
> +	if ((flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED)
> +		/* CQ event for next solicited completion */
> +		smp_store_mb(*cq->notify, SIW_NOTIFY_SOLICITED);
> +	else
> +		/* CQ event for any signalled completion */
> +		smp_store_mb(*cq->notify, SIW_NOTIFY_ALL);
> +
> +	if (flags & IB_CQ_REPORT_MISSED_EVENTS)
> +		return cq->cq_put - cq->cq_get;
> +
> +	return 0;
> +}
> +
> +/*
> + * siw_dereg_mr()
> + *
> + * Release Memory Region.
> + *
> + * TODO: Update function if Memory Windows are supported by siw:
> + *       Is OFED core checking for MW dependencies for current
> + *       MR before calling MR deregistration?.
> + *
> + * @base_mr:     Base MR contained in siw MR.
> + */
> +int siw_dereg_mr(struct ib_mr *base_mr)
> +{
> +	struct siw_mr *mr;
> +	struct siw_device *sdev = siw_dev_base2siw(base_mr->device);
> +
> +	mr = siw_mr_base2siw(base_mr);
> +
> +	siw_dbg(sdev, "[MEM %d]: deregister mr, #ref's %d\n",
> +		mr->mem.hdr.id, refcount_read(&mr->mem.hdr.ref));
> +
> +	mr->mem.stag_valid = 0;
> +
> +	siw_remove_obj(&sdev->idr_lock, &sdev->mem_idr, &mr->mem.hdr);
> +	siw_mem_put(&mr->mem);
> +
> +	return 0;
> +}
> +
> +static struct siw_mr *siw_create_mr(struct siw_device *sdev, void
*mem_obj,
> +				    u64 start, u64 len, int rights)
> +{
> +	struct siw_mr *mr = kzalloc(sizeof(*mr), GFP_KERNEL);
> +	unsigned long flags;
> +
> +	if (!mr)
> +		return NULL;
> +
> +	mr->mem.stag_valid = 0;
> +
> +	if (siw_mem_add(sdev, &mr->mem) < 0) {
> +		kfree(mr);
> +		return NULL;
> +	}
> +	siw_dbg(sdev, "[MEM %d]: new mr, object 0x%p\n",
> +		mr->mem.hdr.id, mem_obj);
> +
> +	mr->base_mr.lkey = mr->base_mr.rkey = mr->mem.hdr.id << 8;
> +
> +	mr->mem.va  = start;
> +	mr->mem.len = len;
> +	mr->mem.mr  = NULL;
> +	mr->mem.perms = SIW_MEM_LREAD | /* not selectable in RDMA core
> */
> +		(rights & IB_ACCESS_REMOTE_READ  ? SIW_MEM_RREAD  : 0) |
> +		(rights & IB_ACCESS_LOCAL_WRITE  ? SIW_MEM_LWRITE : 0) |
> +		(rights & IB_ACCESS_REMOTE_WRITE ? SIW_MEM_RWRITE :
> 0);
> +
> +	mr->mem_obj = mem_obj;
> +
> +	INIT_LIST_HEAD(&mr->devq);
> +	spin_lock_irqsave(&sdev->idr_lock, flags);
> +	list_add_tail(&mr->devq, &sdev->mr_list);
> +	spin_unlock_irqrestore(&sdev->idr_lock, flags);
> +
> +	return mr;
> +}
> +
> +/*
> + * siw_reg_user_mr()
> + *
> + * Register Memory Region.
> + *
> + * @base_pd:	Base PD contained in siw PD.
> + * @start:	starting address of MR (virtual address)
> + * @len:	len of MR
> + * @rnic_va:	not used by siw
> + * @rights:	MR access rights
> + * @udata:	user buffer to communicate STag and Key.
> + */
> +struct ib_mr *siw_reg_user_mr(struct ib_pd *base_pd, u64 start, u64 len,
> +			      u64 rnic_va, int rights, struct ib_udata
*udata)
> +{
> +	struct siw_mr		*mr = NULL;
> +	struct siw_pd		*pd = siw_pd_base2siw(base_pd);
> +	struct siw_umem		*umem = NULL;
> +	struct siw_ureq_reg_mr	ureq;
> +	struct siw_uresp_reg_mr	uresp;
> +	struct siw_device	*sdev = pd->hdr.sdev;
> +
> +	unsigned long mem_limit = rlimit(RLIMIT_MEMLOCK);
> +	int rv;
> +
> +	siw_dbg(sdev, "[PD %d]: start: 0x%016llx, va: 0x%016llx, len:
%llu\n",
> +		OBJ_ID(pd), (unsigned long long)start,
> +		(unsigned long long)rnic_va, (unsigned long long)len);
> +
> +	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
> +		siw_dbg(sdev, "[PD %d]: too many mr's\n", OBJ_ID(pd));
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	if (!len) {
> +		rv = -EINVAL;
> +		goto err_out;
> +	}
> +	if (mem_limit != RLIM_INFINITY) {
> +		unsigned long num_pages =
> +			(PAGE_ALIGN(len + (start & ~PAGE_MASK))) >>
> PAGE_SHIFT;
> +		mem_limit >>= PAGE_SHIFT;
> +
> +		if (num_pages > mem_limit - current->mm->locked_vm) {
> +			siw_dbg(sdev,
> +				"[PD %d]: pages req %lu, max %lu, lock
%lu\n",
> +				OBJ_ID(pd), num_pages, mem_limit,
> +				current->mm->locked_vm);
> +			rv = -ENOMEM;
> +			goto err_out;
> +		}
> +	}
> +	umem = siw_umem_get(start, len);
> +	if (IS_ERR(umem)) {
> +		rv = PTR_ERR(umem);
> +		siw_dbg(sdev, "[PD %d]: getting user memory failed: %d\n",
> +			OBJ_ID(pd), rv);
> +		umem = NULL;
> +		goto err_out;
> +	}
> +	mr = siw_create_mr(sdev, umem, start, len, rights);
> +	if (!mr) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	if (udata) {
> +		rv = ib_copy_from_udata(&ureq, udata, sizeof(ureq));
> +		if (rv)
> +			goto err_out;
> +
> +		mr->base_mr.lkey |= ureq.stag_key;
> +		mr->base_mr.rkey |= ureq.stag_key;
> +		uresp.stag = mr->base_mr.lkey;
> +
> +		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
> +		if (rv)
> +			goto err_out;
> +	}
> +	mr->pd = pd;
> +	siw_pd_get(pd);
> +
> +	mr->mem.stag_valid = 1;
> +
> +	return &mr->base_mr;
> +
> +err_out:
> +	if (mr) {
> +		siw_remove_obj(&sdev->idr_lock, &sdev->mem_idr, &mr-
> >mem.hdr);
> +		siw_mem_put(&mr->mem);
> +		umem = NULL;
> +	} else
> +		atomic_dec(&sdev->num_mr);
> +
> +	if (umem)
> +		siw_umem_release(umem);
> +
> +	return ERR_PTR(rv);
> +}
> +
> +struct ib_mr *siw_alloc_mr(struct ib_pd *base_pd, enum ib_mr_type
mr_type,
> +			   u32 max_sge)
> +{
> +	struct siw_mr	*mr;
> +	struct siw_pd	*pd = siw_pd_base2siw(base_pd);
> +	struct siw_device *sdev = pd->hdr.sdev;
> +	struct siw_pbl	*pbl = NULL;
> +	int rv;
> +
> +	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
> +		siw_dbg(sdev, "[PD %d]: too many mr's\n", OBJ_ID(pd));
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	if (mr_type != IB_MR_TYPE_MEM_REG) {
> +		siw_dbg(sdev, "[PD %d]: mr type %d unsupported\n",
> +			OBJ_ID(pd), mr_type);
> +		rv = -EOPNOTSUPP;
> +		goto err_out;
> +	}
> +	if (max_sge > SIW_MAX_SGE_PBL) {
> +		siw_dbg(sdev, "[PD %d]: too many sge's: %d\n",
> +			OBJ_ID(pd), max_sge);
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	pbl = siw_pbl_alloc(max_sge);
> +	if (IS_ERR(pbl)) {
> +		rv = PTR_ERR(pbl);
> +		siw_dbg(sdev, "[PD %d]: pbl allocation failed: %d\n",
> +			OBJ_ID(pd), rv);
> +		pbl = NULL;
> +		goto err_out;
> +	}
> +	mr = siw_create_mr(sdev, pbl, 0, max_sge * PAGE_SIZE, 0);
> +	if (!mr) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	mr->mem.is_pbl = 1;
> +	mr->pd = pd;
> +	siw_pd_get(pd);
> +
> +	siw_dbg(sdev, "[PD %d], [MEM %d]: success\n",
> +		OBJ_ID(pd), OBJ_ID(&mr->mem));
> +
> +	return &mr->base_mr;
> +
> +err_out:
> +	if (pbl)
> +		siw_pbl_free(pbl);
> +
> +	siw_dbg(sdev, "[PD %d]: failed: %d\n", OBJ_ID(pd), rv);
> +
> +	atomic_dec(&sdev->num_mr);
> +
> +	return ERR_PTR(rv);
> +}
> +
> +/* Just used to count number of pages being mapped */
> +static int siw_set_pbl_page(struct ib_mr *base_mr, u64 buf_addr)
> +{
> +	return 0;
> +}
> +
> +int siw_map_mr_sg(struct ib_mr *base_mr, struct scatterlist *sl, int
num_sle,
> +		  unsigned int *sg_off)
> +{
> +	struct scatterlist *slp;
> +	struct siw_mr *mr = siw_mr_base2siw(base_mr);
> +	struct siw_pbl *pbl = mr->pbl;
> +	struct siw_pble *pble = pbl->pbe;
> +	u64 pbl_size;
> +	int i, rv;
> +
> +	if (!pbl) {
> +		siw_dbg(mr->mem.hdr.sdev, "[MEM %d]: no pbl allocated\n",
> +			OBJ_ID(&mr->mem));
> +		return -EINVAL;
> +	}
> +	if (pbl->max_buf < num_sle) {
> +		siw_dbg(mr->mem.hdr.sdev, "[MEM %d]: too many sge's:
> %d>%d\n",
> +			OBJ_ID(&mr->mem), mr->pbl->max_buf, num_sle);
> +		return -ENOMEM;
> +	}
> +
> +	for_each_sg(sl, slp, num_sle, i) {
> +		if (sg_dma_len(slp) == 0) {
> +			siw_dbg(mr->mem.hdr.sdev, "[MEM %d]: empty
> sge\n",
> +				OBJ_ID(&mr->mem));
> +			return -EINVAL;
> +		}
> +		if (i == 0) {
> +			pble->addr = sg_dma_address(slp);
> +			pble->size = sg_dma_len(slp);
> +			pble->pbl_off = 0;
> +			pbl_size = pble->size;
> +			pbl->num_buf = 1;
> +
> +			continue;
> +		}
> +		/* Merge PBL entries if adjacent */
> +		if (pble->addr + pble->size == sg_dma_address(slp))
> +			pble->size += sg_dma_len(slp);
> +		else {
> +			pble++;
> +			pbl->num_buf++;
> +			pble->addr = sg_dma_address(slp);
> +			pble->size = sg_dma_len(slp);
> +			pble->pbl_off = pbl_size;
> +		}
> +		pbl_size += sg_dma_len(slp);
> +
> +		siw_dbg(mr->mem.hdr.sdev,
> +			"[MEM %d]: sge[%d], size %llu, addr %p, total
%llu\n",
> +			OBJ_ID(&mr->mem), i, pble->size, (void *)pble->addr,
> +			pbl_size);
> +	}
> +	rv = ib_sg_to_pages(base_mr, sl, num_sle, sg_off, siw_set_pbl_page);
> +	if (rv > 0) {
> +		mr->mem.len = base_mr->length;
> +		mr->mem.va = base_mr->iova;
> +		siw_dbg(mr->mem.hdr.sdev,
> +			"[MEM %d]: %llu byte, %u SLE into %u entries\n",
> +			OBJ_ID(&mr->mem), mr->mem.len, num_sle, pbl-
> >num_buf);
> +	}
> +	return rv;
> +}
> +
> +/*
> + * siw_get_dma_mr()
> + *
> + * Create a (empty) DMA memory region, where no umem is attached.
> + */
> +struct ib_mr *siw_get_dma_mr(struct ib_pd *base_pd, int rights)
> +{
> +	struct siw_mr *mr;
> +	struct siw_pd *pd = siw_pd_base2siw(base_pd);
> +	struct siw_device *sdev = pd->hdr.sdev;
> +	int rv;
> +
> +	if (atomic_inc_return(&sdev->num_mr) > SIW_MAX_MR) {
> +		siw_dbg(sdev, "[PD %d]: too many mr's\n", OBJ_ID(pd));
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	mr = siw_create_mr(sdev, NULL, 0, ULONG_MAX, rights);
> +	if (!mr) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	mr->mem.stag_valid = 1;
> +
> +	mr->pd = pd;
> +	siw_pd_get(pd);
> +
> +	siw_dbg(sdev, "[PD %d], [MEM %d]: success\n",
> +		OBJ_ID(pd), OBJ_ID(&mr->mem));
> +
> +	return &mr->base_mr;
> +
> +err_out:
> +	atomic_dec(&sdev->num_mr);
> +
> +	return ERR_PTR(rv);
> +}
> +
> +/*
> + * siw_create_srq()
> + *
> + * Create Shared Receive Queue of attributes @init_attrs
> + * within protection domain given by @base_pd.
> + *
> + * @base_pd:	Base PD contained in siw PD.
> + * @init_attrs:	SRQ init attributes.
> + * @udata:	not used by siw.
> + */
> +struct ib_srq *siw_create_srq(struct ib_pd *base_pd,
> +			      struct ib_srq_init_attr *init_attrs,
> +			      struct ib_udata *udata)
> +{
> +	struct siw_srq		*srq = NULL;
> +	struct ib_srq_attr	*attrs = &init_attrs->attr;
> +	struct siw_pd		*pd = siw_pd_base2siw(base_pd);
> +	struct siw_device	*sdev = pd->hdr.sdev;
> +
> +	int kernel_verbs = base_pd->uobject ? 0 : 1;
> +	int rv;
> +
> +	if (atomic_inc_return(&sdev->num_srq) > SIW_MAX_SRQ) {
> +		siw_dbg(sdev, "too many srq's\n");
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	if (attrs->max_wr == 0 || attrs->max_wr > SIW_MAX_SRQ_WR ||
> +	    attrs->max_sge > SIW_MAX_SGE || attrs->srq_limit >
attrs->max_wr)
> {
> +		rv = -EINVAL;
> +		goto err_out;
> +	}
> +
> +	srq = kzalloc(sizeof(*srq), GFP_KERNEL);
> +	if (!srq) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +
> +	srq->max_sge = attrs->max_sge;
> +	srq->num_rqe = roundup_pow_of_two(attrs->max_wr);
> +	atomic_set(&srq->space, srq->num_rqe);
> +
> +	srq->limit = attrs->srq_limit;
> +	if (srq->limit)
> +		srq->armed = 1;
> +
> +	if (kernel_verbs)
> +		srq->recvq = vzalloc(srq->num_rqe * sizeof(struct siw_rqe));
> +	else
> +		srq->recvq = vmalloc_user(srq->num_rqe *
> +					  sizeof(struct siw_rqe));
> +
> +	if (srq->recvq == NULL) {
> +		rv = -ENOMEM;
> +		goto err_out;
> +	}
> +	if (kernel_verbs) {
> +		srq->kernel_verbs = 1;
> +	} else if (udata) {
> +		struct siw_uresp_create_srq uresp;
> +		struct siw_ucontext *ctx;
> +
> +		memset(&uresp, 0, sizeof(uresp));
> +		ctx = siw_ctx_base2siw(base_pd->uobject->context);
> +
> +		uresp.num_rqe = srq->num_rqe;
> +		uresp.srq_key = siw_insert_uobj(ctx, srq->recvq,
> +					srq->num_rqe * sizeof(struct
siw_rqe));
> +
> +		if (uresp.srq_key > SIW_MAX_UOBJ_KEY)
> +			siw_dbg(sdev, "preparing mmap srq failed\n");
> +
> +		rv = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
> +		if (rv)
> +			goto err_out;
> +	}
> +	srq->pd	= pd;
> +	siw_pd_get(pd);
> +
> +	spin_lock_init(&srq->lock);
> +
> +	siw_dbg(sdev, "[SRQ 0x%p]: success\n", srq);
> +
> +	return &srq->base_srq;
> +
> +err_out:
> +	if (srq) {
> +		if (srq->recvq)
> +			vfree(srq->recvq);
> +		kfree(srq);
> +	}
> +	atomic_dec(&sdev->num_srq);
> +
> +	return ERR_PTR(rv);
> +}
> +
> +/*
> + * siw_modify_srq()
> + *
> + * Modify SRQ. The caller may resize SRQ and/or set/reset notification
> + * limit and (re)arm IB_EVENT_SRQ_LIMIT_REACHED notification.
> + *
> + * NOTE: it is unclear if RDMA core allows for changing the MAX_SGE
> + * parameter. siw_modify_srq() does not check the attrs->max_sge param.
> + */
> +int siw_modify_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs,
> +		   enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
> +{
> +	struct siw_srq	*srq = siw_srq_base2siw(base_srq);
> +	unsigned long	flags;
> +	int rv = 0;
> +
> +	spin_lock_irqsave(&srq->lock, flags);
> +
> +	if (attr_mask & IB_SRQ_MAX_WR) {
> +		/* resize request not yet supported */
> +		rv = -EOPNOTSUPP;
> +		goto out;
> +	}
> +	if (attr_mask & IB_SRQ_LIMIT) {
> +		if (attrs->srq_limit) {
> +			if (unlikely(attrs->srq_limit > srq->num_rqe)) {
> +				rv = -EINVAL;
> +				goto out;
> +			}
> +			srq->armed = 1;
> +		} else
> +			srq->armed = 0;
> +
> +		srq->limit = attrs->srq_limit;
> +	}
> +out:
> +	spin_unlock_irqrestore(&srq->lock, flags);
> +
> +	return rv;
> +}
> +
> +/*
> + * siw_query_srq()
> + *
> + * Query SRQ attributes.
> + */
> +int siw_query_srq(struct ib_srq *base_srq, struct ib_srq_attr *attrs)
> +{
> +	struct siw_srq	*srq = siw_srq_base2siw(base_srq);
> +	unsigned long	flags;
> +
> +	spin_lock_irqsave(&srq->lock, flags);
> +
> +	attrs->max_wr = srq->num_rqe;
> +	attrs->max_sge = srq->max_sge;
> +	attrs->srq_limit = srq->limit;
> +
> +	spin_unlock_irqrestore(&srq->lock, flags);
> +
> +	return 0;
> +}
> +
> +/*
> + * siw_destroy_srq()
> + *
> + * Destroy SRQ.
> + * It is assumed that the SRQ is not referenced by any
> + * QP anymore - the code trusts the RDMA core environment to keep track
> + * of QP references.
> + */
> +int siw_destroy_srq(struct ib_srq *base_srq)
> +{
> +	struct siw_srq		*srq = siw_srq_base2siw(base_srq);
> +	struct siw_device	*sdev = srq->pd->hdr.sdev;
> +
> +	siw_pd_put(srq->pd);
> +
> +	vfree(srq->recvq);
> +	kfree(srq);
> +
> +	atomic_dec(&sdev->num_srq);
> +
> +	return 0;
> +}
> +
> +/*
> + * siw_post_srq_recv()
> + *
> + * Post a list of receive queue elements to SRQ.
> + * NOTE: The function does not check or lock a certain SRQ state
> + *       during the post operation. The code simply trusts the
> + *       RDMA core environment.
> + *
> + * @base_srq:	Base SRQ contained in siw SRQ
> + * @wr:		List of R-WR's
> + * @bad_wr:	Updated to failing WR if posting fails.
> + */
> +int siw_post_srq_recv(struct ib_srq *base_srq, struct ib_recv_wr *wr,
> +		      struct ib_recv_wr **bad_wr)
> +{
> +	struct siw_srq	*srq = siw_srq_base2siw(base_srq);
> +	int rv = 0;
> +
> +	if (!srq->kernel_verbs) {
> +		siw_dbg(srq->pd->hdr.sdev,
> +			"[SRQ 0x%p]: no kernel post_recv for mapped srq\n",
> +			srq);
> +		rv = -EINVAL;
> +		goto out;
> +	}
> +	while (wr) {
> +		u32 idx = srq->rq_put % srq->num_rqe;
> +		struct siw_rqe *rqe = &srq->recvq[idx];
> +
> +		if (rqe->flags) {
> +			siw_dbg(srq->pd->hdr.sdev, "[SRQ 0x%p]: srq full\n",
> +				srq);
> +			rv = -ENOMEM;
> +			break;
> +		}
> +		if (wr->num_sge > srq->max_sge) {
> +			siw_dbg(srq->pd->hdr.sdev,
> +				"[SRQ 0x%p]: too many sge's: %d\n",
> +				srq, wr->num_sge);
> +			rv = -EINVAL;
> +			break;
> +		}
> +		rqe->id = wr->wr_id;
> +		rqe->num_sge = wr->num_sge;
> +		siw_copy_sgl(wr->sg_list, rqe->sge, wr->num_sge);
> +
> +		/* Make sure S-RQE is completely written before valid */
> +		smp_wmb();
> +
> +		rqe->flags = SIW_WQE_VALID;
> +
> +		srq->rq_put++;
> +		wr = wr->next;
> +	}
> +out:
> +	if (unlikely(rv < 0)) {
> +		siw_dbg(srq->pd->hdr.sdev, "[SRQ 0x%p]: error %d\n", srq,
rv);
> +		*bad_wr = wr;
> +	}
> +	return rv;
> +}
> diff --git a/drivers/infiniband/sw/siw/siw_verbs.h
> b/drivers/infiniband/sw/siw/siw_verbs.h
> new file mode 100644
> index 000000000000..829f603923cc
> --- /dev/null
> +++ b/drivers/infiniband/sw/siw/siw_verbs.h
> @@ -0,0 +1,119 @@
> +/*
> + * Software iWARP device driver
> + *
> + * Authors: Bernard Metzler <bmt-OA+xvbQnYDHMbYB6QlFGEg@public.gmane.org>
> + *
> + * Copyright (c) 2008-2017, IBM Corporation
> + *
> + * This software is available to you under a choice of one of two
> + * licenses. You may choose to be licensed under the terms of the GNU
> + * General Public License (GPL) Version 2, available from the file
> + * COPYING in the main directory of this source tree, or the
> + * BSD license below:
> + *
> + *   Redistribution and use in source and binary forms, with or
> + *   without modification, are permitted provided that the following
> + *   conditions are met:
> + *
> + *   - Redistributions of source code must retain the above copyright
notice,
> + *     this list of conditions and the following disclaimer.
> + *
> + *   - Redistributions in binary form must reproduce the above copyright
> + *     notice, this list of conditions and the following disclaimer in
the
> + *     documentation and/or other materials provided with the
distribution.
> + *
> + *   - Neither the name of IBM nor the names of its contributors may be
> + *     used to endorse or promote products derived from this software
without
> + *     specific prior written permission.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
> + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
> OF
> + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
> + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
> HOLDERS
> + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
> + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR
> IN
> + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
> THE
> + * SOFTWARE.
> + */
> +
> +#ifndef _SIW_VERBS_H
> +#define _SIW_VERBS_H
> +
> +#include <linux/errno.h>
> +
> +#include <rdma/iw_cm.h>
> +#include <rdma/ib_verbs.h>
> +#include <rdma/ib_smi.h>
> +#include <rdma/ib_user_verbs.h>
> +
> +#include "siw.h"
> +#include "siw_cm.h"
> +
> +
> +extern struct ib_ucontext *siw_alloc_ucontext(struct ib_device *ibdev,
> +					      struct ib_udata *udata);
> +extern int siw_dealloc_ucontext(struct ib_ucontext *ucontext);
> +extern int siw_query_port(struct ib_device *ibdev, u8 port,
> +			  struct ib_port_attr *attr);
> +extern int siw_get_port_immutable(struct ib_device *ibdev, u8 port,
> +				  struct ib_port_immutable *port_imm);
> +extern int siw_query_device(struct ib_device *ibdev,
> +			    struct ib_device_attr *attr,
> +			    struct ib_udata *udata);
> +extern struct ib_cq *siw_create_cq(struct ib_device *ibdev,
> +				   const struct ib_cq_init_attr *attr,
> +				   struct ib_ucontext *ucontext,
> +				   struct ib_udata *udata);
> +extern int siw_no_mad(struct ib_device *base_dev, int flags, u8 port,
> +		      const struct ib_wc *wc, const struct ib_grh *grh,
> +		      const struct ib_mad_hdr *in_mad, size_t in_mad_size,
> +		      struct ib_mad_hdr *out_mad, size_t *out_mad_size,
> +		      u16 *outmad_pkey_index);
> +extern int siw_query_port(struct ib_device *ibdev, u8 port,
> +			  struct ib_port_attr *attr);
> +extern int siw_query_pkey(struct ib_device *ibdev, u8 port,
> +			  u16 idx, u16 *pkey);
> +extern int siw_query_gid(struct ib_device *ibdev, u8 port, int idx,
> +			 union ib_gid *gid);
> +extern struct ib_pd *siw_alloc_pd(struct ib_device *ibdev,
> +				  struct ib_ucontext *ucontext,
> +				  struct ib_udata *udata);
> +extern int siw_dealloc_pd(struct ib_pd *pd);
> +extern struct ib_qp *siw_create_qp(struct ib_pd *pd,
> +				  struct ib_qp_init_attr *attr,
> +				   struct ib_udata *udata);
> +extern int siw_query_qp(struct ib_qp *ofa_qp, struct ib_qp_attr *qp_attr,
> +			int qp_attr_mask, struct ib_qp_init_attr
*qp_init_attr);
> +extern int siw_verbs_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr
*attr,
> +			       int attr_mask, struct ib_udata *udata);
> +extern int siw_destroy_qp(struct ib_qp *ibqp);
> +extern int siw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
> +			 struct ib_send_wr **bad_wr);
> +extern int siw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
> +			    struct ib_recv_wr **bad_wr);
> +extern int siw_destroy_cq(struct ib_cq *ibcq);
> +extern int siw_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc
*wc);
> +extern int siw_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags
> flags);
> +extern struct ib_mr *siw_reg_user_mr(struct ib_pd *ibpd, u64 start, u64
len,
> +				     u64 rnic_va, int rights,
> +				     struct ib_udata *udata);
> +extern struct ib_mr *siw_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type
> mr_type,
> +				  u32 max_sge);
> +extern struct ib_mr *siw_get_dma_mr(struct ib_pd *ibpd, int rights);
> +extern int siw_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sl,
> +			 int num_sle, unsigned int *sg_off);
> +extern int siw_dereg_mr(struct ib_mr *ibmr);
> +extern struct ib_srq *siw_create_srq(struct ib_pd *ibpd,
> +				     struct ib_srq_init_attr *attr,
> +				     struct ib_udata *udata);
> +extern int siw_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
> +			  enum ib_srq_attr_mask mask, struct ib_udata
*udata);
> +extern int siw_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr);
> +extern int siw_destroy_srq(struct ib_srq *ibsrq);
> +extern int siw_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
> +			     struct ib_recv_wr **bad_wr);
> +extern int siw_mmap(struct ib_ucontext *ibctx, struct vm_area_struct
*vma);
> +
> +extern const struct dma_map_ops siw_dma_generic_ops;
> +
> +#endif
> diff --git a/include/uapi/rdma/siw_user.h b/include/uapi/rdma/siw_user.h
> new file mode 100644
> index 000000000000..85cdc3e2c83d
> --- /dev/null
> +++ b/include/uapi/rdma/siw_user.h
> @@ -0,0 +1,216 @@
> +/*
> + * Software iWARP device driver for Linux
> + *
> + * Authors: Bernard Metzler <bmt-OA+xvbQnYDHMbYB6QlFGEg@public.gmane.org>
> + *
> + * Copyright (c) 2008-2017, IBM Corporation
> + *
> + * This software is available to you under a choice of one of two
> + * licenses.  You may choose to be licensed under the terms of the GNU
> + * General Public License (GPL) Version 2, available from the file
> + * COPYING in the main directory of this source tree, or the
> + * BSD license below:
> + *
> + *   Redistribution and use in source and binary forms, with or
> + *   without modification, are permitted provided that the following
> + *   conditions are met:
> + *
> + *   - Redistributions of source code must retain the above copyright
notice,
> + *     this list of conditions and the following disclaimer.
> + *
> + *   - Redistributions in binary form must reproduce the above copyright
> + *     notice, this list of conditions and the following disclaimer in
the
> + *     documentation and/or other materials provided with the
distribution.
> + *
> + *   - Neither the name of IBM nor the names of its contributors may be
> + *     used to endorse or promote products derived from this software
without
> + *     specific prior written permission.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
> + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
> OF
> + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
> + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
> HOLDERS
> + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
> + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR
> IN
> + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
> THE
> + * SOFTWARE.
> + */
> +
> +#ifndef _SIW_USER_H
> +#define _SIW_USER_H
> +
> +#include <linux/types.h>
> +
> +/*Common string that is matched to accept the device by the user
library*/
> +#define SIW_NODE_DESC_COMMON	"Software iWARP stack"
> +
> +#define SIW_IBDEV_PREFIX "siw_"
> +
> +#define VERSION_ID_SOFTIWARP	2
> +
> +#define SIW_MAX_SGE		6
> +#define SIW_MAX_UOBJ_KEY	0xffffff
> +#define SIW_INVAL_UOBJ_KEY	(SIW_MAX_UOBJ_KEY + 1)
> +
> +struct siw_uresp_create_cq {
> +	__u32	cq_id;
> +	__u32	num_cqe;
> +	__u32	cq_key;
> +};
> +
> +struct siw_uresp_create_qp {
> +	__u32	qp_id;
> +	__u32	num_sqe;
> +	__u32	num_rqe;
> +	__u32	sq_key;
> +	__u32	rq_key;
> +};
> +
> +struct siw_ureq_reg_mr {
> +	__u8	stag_key;
> +	__u8	reserved[3];
> +};
> +
> +struct siw_uresp_reg_mr {
> +	__u32	stag;
> +};
> +
> +struct siw_uresp_create_srq {
> +	__u32	num_rqe;
> +	__u32	srq_key;
> +};
> +
> +struct siw_uresp_alloc_ctx {
> +	__u32	dev_id;
> +};
> +
> +enum siw_opcode {
> +	SIW_OP_WRITE		= 0,
> +	SIW_OP_READ		= 1,
> +	SIW_OP_READ_LOCAL_INV	= 2,
> +	SIW_OP_SEND		= 3,
> +	SIW_OP_SEND_WITH_IMM	= 4,
> +	SIW_OP_SEND_REMOTE_INV	= 5,
> +
> +	/* Unsupported */
> +	SIW_OP_FETCH_AND_ADD	= 6,
> +	SIW_OP_COMP_AND_SWAP	= 7,
> +
> +	SIW_OP_RECEIVE		= 8,
> +	/* provider internal SQE */
> +	SIW_OP_READ_RESPONSE	= 9,
> +	/*
> +	 * below opcodes valid for
> +	 * in-kernel clients only
> +	 */
> +	SIW_OP_INVAL_STAG	= 10,
> +	SIW_OP_REG_MR		= 11,
> +	SIW_NUM_OPCODES		= 12
> +};
> +
> +/* Keep it same as ibv_sge to allow for memcpy */
> +struct siw_sge {
> +	__u64	laddr;
> +	__u32	length;
> +	__u32	lkey;
> +};
> +
> +/*
> + * Inline data are kept within the work request itself occupying
> + * the space of sge[1] .. sge[n]. Therefore, inline data cannot be
> + * supported if SIW_MAX_SGE is below 2 elements.
> + */
> +#define SIW_MAX_INLINE	(sizeof(struct siw_sge) * (SIW_MAX_SGE - 1))
> +
> +#if SIW_MAX_SGE < 2
> +#error "SIW_MAX_SGE must be at least 2"
> +#endif

This could be done with BUILD_BUG_ON().

> +
> +enum siw_wqe_flags {
> +	SIW_WQE_VALID           = 1,
> +	SIW_WQE_INLINE          = (1 << 1),
> +	SIW_WQE_SIGNALLED       = (1 << 2),
> +	SIW_WQE_SOLICITED       = (1 << 3),
> +	SIW_WQE_READ_FENCE	= (1 << 4),
> +	SIW_WQE_COMPLETED       = (1 << 5)
> +};
> +
> +/* Send Queue Element */
> +struct siw_sqe {
> +	__u64	id;
> +	__u16	flags;
> +	__u8	num_sge;
> +	/* Contains enum siw_opcode values */
> +	__u8	opcode;
> +	__u32	rkey;
> +	union {
> +		__u64	raddr;
> +		__u64	base_mr;
> +	};
> +	union {
> +		struct siw_sge	sge[SIW_MAX_SGE];
> +		__u32	access;
> +	};
> +};
> +
> +/* Receive Queue Element */
> +struct siw_rqe {
> +	__u64	id;
> +	__u16	flags;
> +	__u8	num_sge;
> +	/*
> +	 * only used by kernel driver,
> +	 * ignored if set by user
> +	 */
> +	__u8	opcode;
> +	__u32	unused;
> +	struct siw_sge	sge[SIW_MAX_SGE];
> +};
> +
> +enum siw_notify_flags {
> +	SIW_NOTIFY_NOT			= (0),
> +	SIW_NOTIFY_SOLICITED		= (1 << 0),
> +	SIW_NOTIFY_NEXT_COMPLETION	= (1 << 1),
> +	SIW_NOTIFY_MISSED_EVENTS	= (1 << 2),
> +	SIW_NOTIFY_ALL = SIW_NOTIFY_SOLICITED |
> +			SIW_NOTIFY_NEXT_COMPLETION |
> +			SIW_NOTIFY_MISSED_EVENTS
> +};
> +
> +enum siw_wc_status {
> +	SIW_WC_SUCCESS		= 0,
> +	SIW_WC_LOC_LEN_ERR	= 1,
> +	SIW_WC_LOC_PROT_ERR	= 2,
> +	SIW_WC_LOC_QP_OP_ERR	= 3,
> +	SIW_WC_WR_FLUSH_ERR	= 4,
> +	SIW_WC_BAD_RESP_ERR	= 5,
> +	SIW_WC_LOC_ACCESS_ERR	= 6,
> +	SIW_WC_REM_ACCESS_ERR	= 7,
> +	SIW_WC_REM_INV_REQ_ERR	= 8,
> +	SIW_WC_GENERAL_ERR	= 9,
> +	SIW_NUM_WC_STATUS	= 10
> +};
> +
> +struct siw_cqe {
> +	__u64	id;
> +	__u8	flags;
> +	__u8	opcode;
> +	__u16	status;
> +	__u32	bytes;
> +	__u64	imm_data;
> +	/* QP number or QP pointer */
> +	union {
> +		void	*qp;
> +		__u64	qp_id;
> +	};
> +};
> +
> +/*
> + * Shared structure between user and kernel
> + * to control CQ arming.
> + */
> +struct siw_cq_ctrl {
> +	enum siw_notify_flags	notify;
> +};
> +
> +#endif
> --
> 2.13.6
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
> the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html