verbs.c

{
	int rc;

	rpcrdma_flush_cqs(ep);
	rc = rdma_disconnect(ia->ri_id);
	if (!rc) {
		/* returns without wait if not connected */
		wait_event_interruptible(ep->rep_connect_wait,
							ep->rep_connected != 1);
		dprintk("RPC:       %s: after wait, %sconnected\n", __func__,
			(ep->rep_connected == 1) ? "still " : "dis");
	} else {
		dprintk("RPC:       %s: rdma_disconnect %i\n", __func__, rc);
		ep->rep_connected = rc;
	}
}

static struct rpcrdma_req *
rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
{
	struct rpcrdma_req *req;

	req = kzalloc(sizeof(*req), GFP_KERNEL);
	if (req == NULL)
		return ERR_PTR(-ENOMEM);

	req->rl_buffer = &r_xprt->rx_buf;
	return req;
}

static struct rpcrdma_rep *
rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
{
	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
	struct rpcrdma_rep *rep;
	int rc;

	rc = -ENOMEM;
	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
	if (rep == NULL)
		goto out;

	rep->rr_rdmabuf = rpcrdma_alloc_regbuf(ia, cdata->inline_rsize,
					       GFP_KERNEL);
	if (IS_ERR(rep->rr_rdmabuf)) {
		rc = PTR_ERR(rep->rr_rdmabuf);
		goto out_free;
	}

	rep->rr_device = ia->ri_device;
	rep->rr_rxprt = r_xprt;
	return rep;

out_free:
	kfree(rep);
out:
	return ERR_PTR(rc);
}

int
rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
{
	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
	char *p;
	size_t len;
	int i, rc;

	buf->rb_max_requests = cdata->max_requests;
	spin_lock_init(&buf->rb_lock);

	/* Need to allocate:
	 *   1.  arrays for send and recv pointers
	 *   2.  arrays of struct rpcrdma_req to fill in pointers
	 *   3.  array of struct rpcrdma_rep for replies
	 * Send/recv buffers in req/rep need to be registered
	 */
	len = buf->rb_max_requests *
		(sizeof(struct rpcrdma_req *) + sizeof(struct rpcrdma_rep *));

	p = kzalloc(len, GFP_KERNEL);
	if (p == NULL) {
		dprintk("RPC:       %s: req_t/rep_t/pad kzalloc(%zd) failed\n",
			__func__, len);
		rc = -ENOMEM;
		goto out;
	}
	buf->rb_pool = p;	/* for freeing it later */

	buf->rb_send_bufs = (struct rpcrdma_req **) p;
	p = (char *) &buf->rb_send_bufs[buf->rb_max_requests];
	buf->rb_recv_bufs = (struct rpcrdma_rep **) p;
	p = (char *) &buf->rb_recv_bufs[buf->rb_max_requests];

	rc = ia->ri_ops->ro_init(r_xprt);
	if (rc)
		goto out;

	for (i = 0; i < buf->rb_max_requests; i++) {
		struct rpcrdma_req *req;
		struct rpcrdma_rep *rep;

		req = rpcrdma_create_req(r_xprt);
		if (IS_ERR(req)) {
			dprintk("RPC:       %s: request buffer %d alloc"
				" failed\n", __func__, i);
			rc = PTR_ERR(req);
			goto out;
		}
		buf->rb_send_bufs[i] = req;

		rep = rpcrdma_create_rep(r_xprt);
		if (IS_ERR(rep)) {
			dprintk("RPC:       %s: reply buffer %d alloc failed\n",
				__func__, i);
			rc = PTR_ERR(rep);
			goto out;
		}
		buf->rb_recv_bufs[i] = rep;
	}

	return 0;
out:
	rpcrdma_buffer_destroy(buf);
	return rc;
}

static void
rpcrdma_destroy_rep(struct rpcrdma_ia *ia, struct rpcrdma_rep *rep)
{
	if (!rep)
		return;

	rpcrdma_free_regbuf(ia, rep->rr_rdmabuf);
	kfree(rep);
}

static void
rpcrdma_destroy_req(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
{
	if (!req)
		return;

	rpcrdma_free_regbuf(ia, req->rl_sendbuf);
	rpcrdma_free_regbuf(ia, req->rl_rdmabuf);
	kfree(req);
}

void
rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
{
	struct rpcrdma_ia *ia = rdmab_to_ia(buf);
	int i;

	/* clean up in reverse order from create
	 *   1.  recv mr memory (mr free, then kfree)
	 *   2.  send mr memory (mr free, then kfree)
	 *   3.  MWs
	 */
	dprintk("RPC:       %s: entering\n", __func__);

	for (i = 0; i < buf->rb_max_requests; i++) {
		if (buf->rb_recv_bufs)
			rpcrdma_destroy_rep(ia, buf->rb_recv_bufs[i]);
		if (buf->rb_send_bufs)
			rpcrdma_destroy_req(ia, buf->rb_send_bufs[i]);
	}

	ia->ri_ops->ro_destroy(buf);

	kfree(buf->rb_pool);
}

struct rpcrdma_mw *
rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt)
{
	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
	struct rpcrdma_mw *mw = NULL;
	unsigned long flags;

	spin_lock_irqsave(&buf->rb_lock, flags);
	if (!list_empty(&buf->rb_mws)) {
		mw = list_first_entry(&buf->rb_mws,
				      struct rpcrdma_mw, mw_list);
		list_del_init(&mw->mw_list);
	}
	spin_unlock_irqrestore(&buf->rb_lock, flags);

	if (!mw)
		pr_err("RPC:       %s: no MWs available\n", __func__);
	return mw;
}

void
rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw)
{
	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
	unsigned long flags;

	spin_lock_irqsave(&buf->rb_lock, flags);
	list_add_tail(&mw->mw_list, &buf->rb_mws);
	spin_unlock_irqrestore(&buf->rb_lock, flags);
}

/* "*mw" can be NULL when rpcrdma_buffer_get_mrs() fails, leaving
 * some req segments uninitialized.
 */
static void
rpcrdma_buffer_put_mr(struct rpcrdma_mw **mw, struct rpcrdma_buffer *buf)
{
	if (*mw) {
		list_add_tail(&(*mw)->mw_list, &buf->rb_mws);
		*mw = NULL;
	}
}

/* Cycle mw's back in reverse order, and "spin" them.
 * This delays and scrambles reuse as much as possible.
 */
static void
rpcrdma_buffer_put_mrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
{
	struct rpcrdma_mr_seg *seg = req->rl_segments;
	struct rpcrdma_mr_seg *seg1 = seg;
	int i;

	for (i = 1, seg++; i < RPCRDMA_MAX_SEGS; seg++, i++)
		rpcrdma_buffer_put_mr(&seg->rl_mw, buf);
	rpcrdma_buffer_put_mr(&seg1->rl_mw, buf);
}

static void
rpcrdma_buffer_put_sendbuf(struct rpcrdma_req *req, struct rpcrdma_buffer *buf)
{
	buf->rb_send_bufs[--buf->rb_send_index] = req;
	req->rl_niovs = 0;
	if (req->rl_reply) {
		buf->rb_recv_bufs[--buf->rb_recv_index] = req->rl_reply;
		req->rl_reply = NULL;
	}
}

/* rpcrdma_unmap_one() was already done during deregistration.
 * Redo only the ib_post_send().
 */
static void
rpcrdma_retry_local_inv(struct rpcrdma_mw *r, struct rpcrdma_ia *ia)
{
	struct rpcrdma_xprt *r_xprt =
				container_of(ia, struct rpcrdma_xprt, rx_ia);
	struct ib_send_wr invalidate_wr, *bad_wr;
	int rc;

	dprintk("RPC:       %s: FRMR %p is stale\n", __func__, r);

	/* When this FRMR is re-inserted into rb_mws, it is no longer stale */
	r->r.frmr.fr_state = FRMR_IS_INVALID;

	memset(&invalidate_wr, 0, sizeof(invalidate_wr));
	invalidate_wr.wr_id = (unsigned long)(void *)r;
	invalidate_wr.opcode = IB_WR_LOCAL_INV;
	invalidate_wr.ex.invalidate_rkey = r->r.frmr.fr_mr->rkey;
	DECR_CQCOUNT(&r_xprt->rx_ep);

	dprintk("RPC:       %s: frmr %p invalidating rkey %08x\n",
		__func__, r, r->r.frmr.fr_mr->rkey);

	read_lock(&ia->ri_qplock);
	rc = ib_post_send(ia->ri_id->qp, &invalidate_wr, &bad_wr);
	read_unlock(&ia->ri_qplock);
	if (rc) {
		/* Force rpcrdma_buffer_get() to retry */
		r->r.frmr.fr_state = FRMR_IS_STALE;
		dprintk("RPC:       %s: ib_post_send failed, %i\n",
			__func__, rc);
	}
}

static void
rpcrdma_retry_flushed_linv(struct list_head *stale,
			   struct rpcrdma_buffer *buf)
{
	struct rpcrdma_ia *ia = rdmab_to_ia(buf);
	struct list_head *pos;
	struct rpcrdma_mw *r;
	unsigned long flags;

	list_for_each(pos, stale) {
		r = list_entry(pos, struct rpcrdma_mw, mw_list);
		rpcrdma_retry_local_inv(r, ia);
	}

	spin_lock_irqsave(&buf->rb_lock, flags);
	list_splice_tail(stale, &buf->rb_mws);
	spin_unlock_irqrestore(&buf->rb_lock, flags);
}

static struct rpcrdma_req *
rpcrdma_buffer_get_frmrs(struct rpcrdma_req *req, struct rpcrdma_buffer *buf,
			 struct list_head *stale)
{
	struct rpcrdma_mw *r;
	int i;

	i = RPCRDMA_MAX_SEGS - 1;
	while (!list_empty(&buf->rb_mws)) {
		r = list_entry(buf->rb_mws.next,
			       struct rpcrdma_mw, mw_list);
		list_del(&r->mw_list);
		if (r->r.frmr.fr_state == FRMR_IS_STALE) {
			list_add(&r->mw_list, stale);
			continue;
		}
		req->rl_segments[i].rl_mw = r;
		if (unlikely(i-- == 0))
			return req;	/* Success */
	}

	/* Not enough entries on rb_mws for this req */
	rpcrdma_buffer_put_sendbuf(req, buf);
	rpcrdma_buffer_put_mrs(req, buf);
	return NULL;
}

/*
 * Get a set of request/reply buffers.
 *
 * Reply buffer (if needed) is attached to send buffer upon return.
 * Rule:
 *    rb_send_index and rb_recv_index MUST always be pointing to the
 *    *next* available buffer (non-NULL). They are incremented after
 *    removing buffers, and decremented *before* returning them.
 */
struct rpcrdma_req *
rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
{
	struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
	struct list_head stale;
	struct rpcrdma_req *req;
	unsigned long flags;

	spin_lock_irqsave(&buffers->rb_lock, flags);
	if (buffers->rb_send_index == buffers->rb_max_requests) {
		spin_unlock_irqrestore(&buffers->rb_lock, flags);
		dprintk("RPC:       %s: out of request buffers\n", __func__);
		return ((struct rpcrdma_req *)NULL);
	}

	req = buffers->rb_send_bufs[buffers->rb_send_index];
	if (buffers->rb_send_index < buffers->rb_recv_index) {
		dprintk("RPC:       %s: %d extra receives outstanding (ok)\n",
			__func__,
			buffers->rb_recv_index - buffers->rb_send_index);
		req->rl_reply = NULL;
	} else {
		req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
		buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
	}
	buffers->rb_send_bufs[buffers->rb_send_index++] = NULL;

	INIT_LIST_HEAD(&stale);
	switch (ia->ri_memreg_strategy) {
	case RPCRDMA_FRMR:
		req = rpcrdma_buffer_get_frmrs(req, buffers, &stale);
		break;
	default:
		break;
	}
	spin_unlock_irqrestore(&buffers->rb_lock, flags);
	if (!list_empty(&stale))
		rpcrdma_retry_flushed_linv(&stale, buffers);
	return req;
}

/*
 * Put request/reply buffers back into pool.
 * Pre-decrement counter/array index.
 */
void
rpcrdma_buffer_put(struct rpcrdma_req *req)
{
	struct rpcrdma_buffer *buffers = req->rl_buffer;
	struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
	unsigned long flags;

	spin_lock_irqsave(&buffers->rb_lock, flags);
	rpcrdma_buffer_put_sendbuf(req, buffers);
	switch (ia->ri_memreg_strategy) {
	case RPCRDMA_FRMR:
		rpcrdma_buffer_put_mrs(req, buffers);
		break;
	default:
		break;
	}
	spin_unlock_irqrestore(&buffers->rb_lock, flags);
}

/*
 * Recover reply buffers from pool.
 * This happens when recovering from error conditions.
 * Post-increment counter/array index.
 */
void
rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
{
	struct rpcrdma_buffer *buffers = req->rl_buffer;
	unsigned long flags;

	spin_lock_irqsave(&buffers->rb_lock, flags);
	if (buffers->rb_recv_index < buffers->rb_max_requests) {
		req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
		buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
	}
	spin_unlock_irqrestore(&buffers->rb_lock, flags);
}

/*
 * Put reply buffers back into pool when not attached to
 * request. This happens in error conditions.
 */
void
rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
{
	struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
	unsigned long flags;

	spin_lock_irqsave(&buffers->rb_lock, flags);
	buffers->rb_recv_bufs[--buffers->rb_recv_index] = rep;
	spin_unlock_irqrestore(&buffers->rb_lock, flags);
}

/*
 * Wrappers for internal-use kmalloc memory registration, used by buffer code.
 */

void
rpcrdma_mapping_error(struct rpcrdma_mr_seg *seg)
{
	dprintk("RPC:       map_one: offset %p iova %llx len %zu\n",
		seg->mr_offset,
		(unsigned long long)seg->mr_dma, seg->mr_dmalen);
}

static int
rpcrdma_register_internal(struct rpcrdma_ia *ia, void *va, int len,
				struct ib_mr **mrp, struct ib_sge *iov)
{
	struct ib_phys_buf ipb;
	struct ib_mr *mr;
	int rc;

	/*
	 * All memory passed here was kmalloc'ed, therefore phys-contiguous.
	 */
	iov->addr = ib_dma_map_single(ia->ri_device,
			va, len, DMA_BIDIRECTIONAL);
	if (ib_dma_mapping_error(ia->ri_device, iov->addr))
		return -ENOMEM;

	iov->length = len;

	if (ia->ri_have_dma_lkey) {
		*mrp = NULL;
		iov->lkey = ia->ri_dma_lkey;
		return 0;
	} else if (ia->ri_bind_mem != NULL) {
		*mrp = NULL;
		iov->lkey = ia->ri_bind_mem->lkey;
		return 0;
	}

	ipb.addr = iov->addr;
	ipb.size = iov->length;
	mr = ib_reg_phys_mr(ia->ri_pd, &ipb, 1,
			IB_ACCESS_LOCAL_WRITE, &iov->addr);

	dprintk("RPC:       %s: phys convert: 0x%llx "
			"registered 0x%llx length %d\n",
			__func__, (unsigned long long)ipb.addr,
			(unsigned long long)iov->addr, len);

	if (IS_ERR(mr)) {
		*mrp = NULL;
		rc = PTR_ERR(mr);
		dprintk("RPC:       %s: failed with %i\n", __func__, rc);
	} else {
		*mrp = mr;
		iov->lkey = mr->lkey;
		rc = 0;
	}

	return rc;
}

static int
rpcrdma_deregister_internal(struct rpcrdma_ia *ia,
				struct ib_mr *mr, struct ib_sge *iov)
{
	int rc;

	ib_dma_unmap_single(ia->ri_device,
			    iov->addr, iov->length, DMA_BIDIRECTIONAL);

	if (NULL == mr)
		return 0;

	rc = ib_dereg_mr(mr);
	if (rc)
		dprintk("RPC:       %s: ib_dereg_mr failed %i\n", __func__, rc);
	return rc;
}

/**
 * rpcrdma_alloc_regbuf - kmalloc and register memory for SEND/RECV buffers
 * @ia: controlling rpcrdma_ia
 * @size: size of buffer to be allocated, in bytes
 * @flags: GFP flags
 *
 * Returns pointer to private header of an area of internally
 * registered memory, or an ERR_PTR. The registered buffer follows
 * the end of the private header.
 *
 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
 * receiving the payload of RDMA RECV operations. regbufs are not
 * used for RDMA READ/WRITE operations, thus are registered only for
 * LOCAL access.
 */
struct rpcrdma_regbuf *
rpcrdma_alloc_regbuf(struct rpcrdma_ia *ia, size_t size, gfp_t flags)
{
	struct rpcrdma_regbuf *rb;
	int rc;

	rc = -ENOMEM;
	rb = kmalloc(sizeof(*rb) + size, flags);
	if (rb == NULL)
		goto out;

	rb->rg_size = size;
	rb->rg_owner = NULL;
	rc = rpcrdma_register_internal(ia, rb->rg_base, size,
				       &rb->rg_mr, &rb->rg_iov);
	if (rc)
		goto out_free;

	return rb;

out_free:
	kfree(rb);
out:
	return ERR_PTR(rc);
}

/**
 * rpcrdma_free_regbuf - deregister and free registered buffer
 * @ia: controlling rpcrdma_ia
 * @rb: regbuf to be deregistered and freed
 */
void
rpcrdma_free_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
{
	if (rb) {
		rpcrdma_deregister_internal(ia, rb->rg_mr, &rb->rg_iov);
		kfree(rb);
	}
}

/*
 * Prepost any receive buffer, then post send.
 *
 * Receive buffer is donated to hardware, reclaimed upon recv completion.
 */
int
rpcrdma_ep_post(struct rpcrdma_ia *ia,
		struct rpcrdma_ep *ep,
		struct rpcrdma_req *req)
{
	struct ib_send_wr send_wr, *send_wr_fail;
	struct rpcrdma_rep *rep = req->rl_reply;
	int rc;

	if (rep) {
		rc = rpcrdma_ep_post_recv(ia, ep, rep);
		if (rc)
			goto out;
		req->rl_reply = NULL;
	}

	send_wr.next = NULL;
	send_wr.wr_id = RPCRDMA_IGNORE_COMPLETION;
	send_wr.sg_list = req->rl_send_iov;
	send_wr.num_sge = req->rl_niovs;
	send_wr.opcode = IB_WR_SEND;
	if (send_wr.num_sge == 4)	/* no need to sync any pad (constant) */
		ib_dma_sync_single_for_device(ia->ri_device,
					      req->rl_send_iov[3].addr,
					      req->rl_send_iov[3].length,
					      DMA_TO_DEVICE);
	ib_dma_sync_single_for_device(ia->ri_device,
				      req->rl_send_iov[1].addr,
				      req->rl_send_iov[1].length,
				      DMA_TO_DEVICE);
	ib_dma_sync_single_for_device(ia->ri_device,
				      req->rl_send_iov[0].addr,
				      req->rl_send_iov[0].length,
				      DMA_TO_DEVICE);

	if (DECR_CQCOUNT(ep) > 0)
		send_wr.send_flags = 0;
	else { /* Provider must take a send completion every now and then */
		INIT_CQCOUNT(ep);
		send_wr.send_flags = IB_SEND_SIGNALED;
	}

	rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail);
	if (rc)
		dprintk("RPC:       %s: ib_post_send returned %i\n", __func__,
			rc);
out:
	return rc;
}

/*
 * (Re)post a receive buffer.
 */
int
rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
		     struct rpcrdma_ep *ep,
		     struct rpcrdma_rep *rep)
{
	struct ib_recv_wr recv_wr, *recv_wr_fail;
	int rc;

	recv_wr.next = NULL;
	recv_wr.wr_id = (u64) (unsigned long) rep;
	recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
	recv_wr.num_sge = 1;

	ib_dma_sync_single_for_cpu(ia->ri_device,
				   rdmab_addr(rep->rr_rdmabuf),
				   rdmab_length(rep->rr_rdmabuf),
				   DMA_BIDIRECTIONAL);

	rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);

	if (rc)
		dprintk("RPC:       %s: ib_post_recv returned %i\n", __func__,
			rc);
	return rc;
}

/* How many chunk list items fit within our inline buffers?
 */
unsigned int
rpcrdma_max_segments(struct rpcrdma_xprt *r_xprt)
{
	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
	int bytes, segments;

	bytes = min_t(unsigned int, cdata->inline_wsize, cdata->inline_rsize);
	bytes -= RPCRDMA_HDRLEN_MIN;
	if (bytes < sizeof(struct rpcrdma_segment) * 2) {
		pr_warn("RPC:       %s: inline threshold too small\n",
			__func__);
		return 0;
	}

	segments = 1 << (fls(bytes / sizeof(struct rpcrdma_segment)) - 1);
	dprintk("RPC:       %s: max chunk list size = %d segments\n",
		__func__, segments);
	return segments;
}