Re: [RFC,PATCH 11/15] knfsd: RDMA transport core

[Tom Tucker <tom@opengridcomputing.com>]

On Fri, 2007-05-18 at 15:07 -0400, Trond Myklebust wrote:
[...snip...]
> +		     xprt->sc_ctxt_max);
> > +
> > +	spin_lock_irqsave(&xprt->sc_ctxt_lock, flags);
> 
> Why do you need an irqsafe spinlock to protect this list?
> 

The svc_rdma_put_context function is called from interrupt context. This
function adds free contexts back to this list.

> > +	while (xprt->sc_ctxt_cnt < target) {
> > +		xprt->sc_ctxt_cnt ++;
> > +		spin_unlock_irqrestore(&xprt->sc_ctxt_lock, flags);
> > +
> > +		ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
> > +
> > +		spin_lock_irqsave(&xprt->sc_ctxt_lock, flags);
> 
> You've now dropped the spinlock. How can you know that the condition
> xprt->sc_ctxt_cnt <= target is still valid?
> 

I increment the sc_ctxt_cnt with the lock held. If I'm at the limit, the
competing thread will find it equal -- correct? 

> > +		if (ctxt) {
> > +			at_least_one = 1;
> > +			ctxt->next = xprt->sc_ctxt_head;
> > +			xprt->sc_ctxt_head = ctxt;
> > +		} else {
> > +			/* kmalloc failed...give up for now */
> > +			xprt->sc_ctxt_cnt --;
> > +			break;
> > +		}
> > +	}
> > +	spin_unlock_irqrestore(&xprt->sc_ctxt_lock, flags);
> > +
> > +	return at_least_one;
> > +}
> > +
> > +struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
> > +{
> > +	struct svc_rdma_op_ctxt *ctxt;
> > +	unsigned long flags;
> > +
> > +	while (1) {
> > +		spin_lock_irqsave(&xprt->sc_ctxt_lock, flags);
> > +		if (unlikely(xprt->sc_ctxt_head == NULL)) {
> > +			/* Try to bump my cache. */
> > +			spin_unlock_irqrestore(&xprt->sc_ctxt_lock, flags);
> > +
> > +			if (rdma_bump_context_cache(xprt))
> > +				continue;
> > +
> > +			printk(KERN_INFO "svcrdma: sleeping waiting for context "
> > +			       "memory on xprt=%p\n",
> > +			       xprt);
> > +			schedule_timeout_uninterruptible(msecs_to_jiffies(500));
> 
> 							   (HZ >> 1)
> However this is rather naughty: you are tying up an nfsd thread that
> could be put to doing useful work elsewhere.
> 

True, it could be processing requests for another client.

Here was my dilemma. I can actually predict what the worst case number
is so that I'm guaranteed not to fail. The problem is that this is a
very big number and I wanted to have a footprint driven by load as
opposed to worst-case. This led to the code above. 

As an intermediate approach, I believe that I need about 4 contexts to
be guaranteed I can process a request. I could attempt to acquire them
all (at the top of svc_rdma_recvfrom) and cache them in the rqstp
structure. If I couldn't get them, I would just return 0 (EAGAIN). This
would be marginally wasteful on a given request, but would avoid ever
sleeping for a context.

Comments?

> > +			continue;
> > +		}
> > +		ctxt = xprt->sc_ctxt_head;
> > +		xprt->sc_ctxt_head = ctxt->next;
> > +		spin_unlock_irqrestore(&xprt->sc_ctxt_lock, flags);
> > +		ctxt->xprt = xprt;
> > +		INIT_LIST_HEAD(&ctxt->dto_q);
> > +		break;
> > +	}
> > +	ctxt->count = 0;
> > +	return ctxt;
> > +}
> > +
> > +void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
> > +{
> > +	unsigned long flags;
> > +	struct svcxprt_rdma *xprt;
> > +	int i;
> > +
> > +	BUG_ON(!ctxt);
> > +	xprt = ctxt->xprt;
> > +	if (free_pages) {
> > +		for (i=0; i < ctxt->count; i++)
> > +			put_page(ctxt->pages[i]);
> > +	}
> > +
> > +	for (i=0; i < ctxt->count; i++) {
> > +		dma_unmap_single(xprt->sc_cm_id->device->dma_device,
> > +				 ctxt->sge[i].addr,
> > +				 ctxt->sge[i].length,
> > +				 ctxt->direction);
> > +	}
> > +	spin_lock_irqsave(&xprt->sc_ctxt_lock, flags);
> > +	ctxt->next = xprt->sc_ctxt_head;
> > +	xprt->sc_ctxt_head = ctxt;
> > +	spin_unlock_irqrestore(&xprt->sc_ctxt_lock, flags);
> > +}
> > +
> > +/* ib_cq event handler */
> > +static void cq_event_handler(struct ib_event *event, void *context)
> > +{
> > +	struct svcxprt_rdma *xprt = (struct svcxprt_rdma *)context;
> > +	printk(KERN_INFO "svcrdma: received CQ event id=%d, context=%p\n",
> > +	       event->event, context);
> > +	set_bit(SK_CLOSE, &xprt->sc_xprt.sk_flags);
> > +}
> > +
> > +/* QP event handler */
> > +static void qp_event_handler(struct ib_event *event, void *context)
> > +{
> > +	struct svcxprt_rdma *xprt = context;
> > +
> > +        switch (event->event) {
> > +        /* These are considered benign events */
> > +        case IB_EVENT_PATH_MIG:
> > +        case IB_EVENT_COMM_EST:
> > +        case IB_EVENT_SQ_DRAINED:
> > +        case IB_EVENT_QP_LAST_WQE_REACHED:
> > +                printk(KERN_INFO "svcrdma: QP event %d received for QP=%p\n",
> > +                       event->event, event->element.qp);
> > +                break;
> > +        /* These are considered fatal events */
> > +        case IB_EVENT_PATH_MIG_ERR:
> > +        case IB_EVENT_QP_FATAL:
> > +        case IB_EVENT_QP_REQ_ERR:
> > +        case IB_EVENT_QP_ACCESS_ERR:
> > +        case IB_EVENT_DEVICE_FATAL:
> > +        default:
> > +                printk(KERN_ERR "svcrdma: QP ERROR event %d received for QP=%p, "
> > +		       "closing transport\n",
> > +                       event->event, event->element.qp);
> > +                set_bit(SK_CLOSE, &xprt->sc_xprt.sk_flags);
> > +                break;
> > +        }
> > +}
> > +
> > +/*
> > + * Data Transfer Operation Tasklet
> > + *
> > + * Walks a list of transports with I/O pending, removing entries as
> > + * they are added to the server's I/O pending list.
> > + */
> > +static void dto_tasklet_func(unsigned long data)
> > +{
> > +	struct svcxprt_rdma *xprt;
> > +	unsigned long flags;
> > +
> > +	spin_lock_irqsave(&dto_lock, flags);
> > +	while (!list_empty(&dto_xprt_q)) {
> > +		xprt = list_entry(dto_xprt_q.next, struct svcxprt_rdma, sc_dto_q);
> > +		list_del_init(&xprt->sc_dto_q);
> > +		spin_unlock_irqrestore(&dto_lock, flags);
> > +		if (0==test_bit(SK_DEAD, &xprt->sc_xprt.sk_flags)) {
> > +			/* Serialize with svc_rdma_recvfrom which will also
> > +			 * enqueue the transport
> > +			 */
> > +			set_bit(SK_DATA, &xprt->sc_xprt.sk_flags);
> > +			svc_sock_enqueue(&xprt->sc_xprt);
> > +		}
> > +		spin_lock_irqsave(&dto_lock, flags);
> > +	}
> > +	spin_unlock_irqrestore(&dto_lock, flags);
> > +}
> > +
> > +/*
> > + * rq_cq_reap - Process the RQ CQ.
> > + *
> > + * Take all completing WC off the CQE and enqueue the associated DTO context
> > + * on the dto_q for the transport.
> > + */
> > +static void
> > +rq_cq_reap(struct svcxprt_rdma *xprt)
> > +{
> > +	int ret;
> > +	struct ib_wc wc;
> > +	struct svc_rdma_op_ctxt *ctxt = NULL;
> > +	unsigned long flags;
> > +
> > +	rdma_stat_rq_poll ++;
> 
> Is this a global variable? How are you ensuring atomicity above?
> 

It's global and it's really debug code. It should be removed along with
a few of the others. The remaining ones should be atomic_inc.

> > +
> > +	while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
> > +		ctxt = (struct svc_rdma_op_ctxt*)(unsigned long)wc.wr_id;
> > +		ctxt->wc_status = wc.status;
> > +		ctxt->byte_len = wc.byte_len;
> > +		if (wc.status != IB_WC_SUCCESS) {
> > +			DBG_DUMP_WC(__FUNCTION__, &wc);
> > +			/* Close the transport */
> > +			set_bit(SK_CLOSE, &xprt->sc_xprt.sk_flags);
> > +			svc_rdma_put_context(ctxt, 1);
> > +			continue;
> > +		}
> > +		spin_lock_irqsave(&xprt->sc_rq_dto_lock, flags);
> > +		list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
> > +		spin_unlock_irqrestore(&xprt->sc_rq_dto_lock, flags);
> > +	}
> > +
> > +	if (ctxt)
> > +		rdma_stat_rq_prod ++;
> > +}
> > +
> > +/*
> > + * Receive Queue Completion Handler - potentially called on interrupt context.
> > + *
> > + * svc_sock_enqueue and the remainder of the svc core assumes
> > + * uses _bh locks. Since the rq_comp_handler is called on interrupt
> > + * context, we need to refer the handling of the I/O to a tasklet
> > + */
> > +static void
> > +rq_comp_handler(struct ib_cq *cq, void *cq_context)
> > +{
> > +	struct svcxprt_rdma *xprt = cq_context;
> > +	unsigned long flags;
> > +
> > +	ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
> > +	rq_cq_reap(xprt);
> > +
> > +	/*
> > +	 * If this transport is not already on the DTO transport queue,
> > +	 * add it
> > +	 */
> > +	spin_lock_irqsave(&dto_lock, flags);
> > +	if (list_empty(&xprt->sc_dto_q))
> > +		list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
> > +	spin_unlock_irqrestore(&dto_lock, flags);
> > +	tasklet_schedule(&dto_tasklet);
> > +}
> > +
> > +/*
> > + * Send Queue Completion Handler - potentially called on interrupt context.
> > + *
> > + * - Purges the CQ
> > + * - Wakes up threads waiting on SQ WR space
> > + * - Wakes up threads waiting on the ORD throttle
> > + * - Wakes up threads waiting for an RDMA_READ to complete.
> > + */
> > +static void
> > +sq_cq_reap(struct svcxprt_rdma *xprt)
> > +{
> > +	struct svc_rdma_op_ctxt *ctxt = NULL;
> > +	struct ib_wc wc;
> > +	struct ib_cq *cq = xprt->sc_sq_cq;
> > +	int ret;
> > +	
> > +	rdma_stat_sq_poll ++;
> > +
> > +	while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
> > +		ctxt = (struct svc_rdma_op_ctxt*)(unsigned long)wc.wr_id;
> > +		xprt = ctxt->xprt;
> > +
> > +		if (wc.status != IB_WC_SUCCESS) {
> > +			/* Close the transport */
> > +			DBG_DUMP_WC(__FUNCTION__, &wc);
> > +			set_bit(SK_CLOSE, &xprt->sc_xprt.sk_flags);
> > +		}
> > +
> > +		/* Decrement used SQ WR count */
> > +		atomic_dec(&xprt->sc_sq_count);
> > +		wake_up(&xprt->sc_send_wait);
> > +
> > +		switch (ctxt->wr_op) {
> > +		case IB_WR_SEND:
> > +		case IB_WR_RDMA_WRITE:
> > +			svc_rdma_put_context(ctxt,1);
> > +			break;
> > +
> > +		case IB_WR_RDMA_READ:
> > +			if (svcrdma_read_throttle) {
> > +				atomic_dec(&xprt->sc_read_count);
> > +				wake_up(&xprt->sc_read_wait);
> > +			}
> > +			/* 
> > +			 * Set the the RDMA_DONE flag in the context and
> > +			 * wakeup any waiters.
> > +			 */
> > +			set_bit(RDMACTXT_F_READ_DONE, &ctxt->flags);
> > +			wake_up(&ctxt->read_wait);
> > +			break;
> > +
> > +		default:
> > +			printk(KERN_ERR "svcrdma: unexpected completion type, "
> > +			       "opcode=%d, status=%d\n",
> > +			       wc.opcode, wc.status);
> > +			break;
> > +		}
> > +	}
> > +
> > +	if (ctxt)
> > +		rdma_stat_sq_prod ++;
> > +}
> > +
> > +void svc_sq_reap(struct svcxprt_rdma *xprt)
> > +{
> > +	sq_cq_reap(xprt);
> > +}
> > +
> > +void svc_rq_reap(struct svcxprt_rdma *xprt)
> > +{
> > +	rq_cq_reap(xprt);
> > +}
> > +
> > +static void
> > +sq_comp_handler(struct ib_cq *cq, void *cq_context)
> > +{
> > +	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
> > +	sq_cq_reap(cq_context);
> > +}
> > +
> > +static void
> > +create_context_cache(struct svcxprt_rdma *xprt,
> > +		     int ctxt_count, int ctxt_bump, int ctxt_max)
> > +{
> > +	struct svc_rdma_op_ctxt *ctxt;
> > +	int i;
> > +
> > +	xprt->sc_ctxt_max = ctxt_max;
> > +	xprt->sc_ctxt_bump = ctxt_bump;
> > +	xprt->sc_ctxt_cnt = 0;
> > +	xprt->sc_ctxt_head = NULL;
> > +	for (i=0; i < ctxt_count; i++) {
> > +		ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
> > +		if (ctxt) {
> > +			ctxt->next = xprt->sc_ctxt_head;
> > +			xprt->sc_ctxt_head = ctxt;
> > +			xprt->sc_ctxt_cnt ++;
> > +		}
> > +	}
> > +}
> > +
> > +static void destroy_context_cache(struct svc_rdma_op_ctxt *ctxt)
> > +{
> > +	struct svc_rdma_op_ctxt *next;
> > +	if (!ctxt)
> > +		return;
> > +
> > +	do {
> > +		next = ctxt->next;
> > +		kfree(ctxt);
> > +		ctxt = next;
> > +	} while (next);
> > +}
> > +
> > +static struct svcxprt_rdma *rdma_create_xprt(int listener)
> > +{
> > +	struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
> > +
> > +	if (!cma_xprt)
> > +		return NULL;
> > +
> > +	INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
> > +	INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
> > +	INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
> > +	init_waitqueue_head(&cma_xprt->sc_send_wait);
> > +	init_waitqueue_head(&cma_xprt->sc_read_wait);
> > +
> > +	spin_lock_init(&cma_xprt->sc_lock);
> > +	spin_lock_init(&cma_xprt->sc_read_lock);
> > +	spin_lock_init(&cma_xprt->sc_ctxt_lock);
> > +	spin_lock_init(&cma_xprt->sc_rq_dto_lock);
> > +
> > +	cma_xprt->sc_ord = svcrdma_ord;
> > +
> > +	cma_xprt->sc_max_req_size = svcrdma_max_req_size;
> > +	cma_xprt->sc_max_requests = svcrdma_max_requests;
> > +	cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
> > +	atomic_set(&cma_xprt->sc_sq_count,0);
> > +	atomic_set(&cma_xprt->sc_read_count,0);
> > +
> > +	if (!listener) {
> > +		int reqs = cma_xprt->sc_max_requests;
> > +		create_context_cache(cma_xprt,
> > +				     reqs << 1, /* starting size */
> > +				     reqs, 	/* bump amount */
> > +				     reqs +
> > +				     cma_xprt->sc_sq_depth +
> > +				     RPCRDMA_MAX_THREADS); /* max */
> > +
> > +		if (!cma_xprt->sc_ctxt_head) {
> > +			kfree(cma_xprt);
> > +			return NULL;
> > +		}
> > +	}
> > +
> > +	return cma_xprt;
> > +}
> > +
> > +static void svc_rdma_put(struct svc_sock *xprt)
> > +{
> > +	struct svcxprt_rdma *rdma = (struct svcxprt_rdma *)xprt;
> > +
> > +	if (atomic_dec_and_test(&xprt->sk_inuse)) {
> > +		BUG_ON(! test_bit(SK_DEAD, &xprt->sk_flags));
> > +
> > +		printk("svcrdma: Destroying transport %p, cm_id=%p, "
> > +		       "sk_flags=%lx\n",
> > +		       xprt, rdma->sc_cm_id, xprt->sk_flags);
> > +
> > +		rdma_disconnect(rdma->sc_cm_id);
> > +		rdma_destroy_id(rdma->sc_cm_id);
> > +		rdma_destroy_xprt(rdma);
> > +	}
> > +}
> > +
> > +struct page *svc_rdma_get_page(void)
> > +{
> > +	struct page *page;
> > +
> > +	while ((page = alloc_page(GFP_KERNEL))==NULL) {
> > +		/* If we can't get memory, wait a bit and try again */
> > +		printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 jiffies.\n");
> > +		schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
> 							HZ
> See comment above about tying up threads. Also note that you are
> probably better off using __GFP_NOFAIL instead of the loop.
> 

I thought that the __GFP_NOFAIL pool was a precious resource and that a
memory out condition was one of the places you'd actually want to sleep.
I basically copied this approach from the svc_recv implementation in
svcsock.c

> > +	}
> > +	return page;
> > +}
> > +
> > +int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
> > +{
> > +	struct ib_recv_wr recv_wr, *bad_recv_wr;
> > +	struct svc_rdma_op_ctxt *ctxt;
> > +	struct page *page;
> > +	unsigned long pa;
> > +	int sge_no;
> > +	int buflen;
> > +	int ret;
> > +
> > +	ctxt = svc_rdma_get_context(xprt);
> > +	buflen = 0;
> > +	ctxt->direction = DMA_FROM_DEVICE;
> > +	for (sge_no=0; buflen < xprt->sc_max_req_size; sge_no++) {
> > +		BUG_ON(sge_no >= xprt->sc_max_sge);
> > +		page = svc_rdma_get_page();
> > +		ctxt->pages[sge_no] = page;
> > +		pa = ib_dma_map_page(xprt->sc_cm_id->device,
> > +				     page, 0, PAGE_SIZE,
> > +				     DMA_FROM_DEVICE);
> > +		ctxt->sge[sge_no].addr = pa;
> > +		ctxt->sge[sge_no].length = PAGE_SIZE;
> > +		ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
> > +		buflen += PAGE_SIZE;
> > +	}
> > +	ctxt->count = sge_no;
> > +	recv_wr.next = NULL;
> > +	recv_wr.sg_list = &ctxt->sge[0];
> > +	recv_wr.num_sge = ctxt->count;
> > +	recv_wr.wr_id = (u64)(unsigned long)ctxt;
> > +
> > +	ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
> > +	return ret;
> > +}
> > +
> > +
> > +/*
> > + * This function handles the CONNECT_REQUEST event on a listening
> > + * endpoint. It is passed the cma_id for the _new_ connection. The context in
> > + * this cma_id is inherited from the listening cma_id and is the svc_sock
> > + * structure for the listening endpoint.
> > + *
> > + * This function creates a new xprt for the new connection and enqueues it on
> > + * the accept queue for the listent xprt. When the listen thread is kicked, it
> > + * will call the recvfrom method on the listen xprt which will accept the new
> > + * connection.
> > + */
> > +static void handle_connect_req(struct rdma_cm_id *new_cma_id)
> > +{
> > +	struct svcxprt_rdma *listen_xprt = new_cma_id->context;
> > +	struct svcxprt_rdma *newxprt;
> > +
> > +	/* Create a new transport */
> > +	newxprt = rdma_create_xprt(0);
> > +	if (!newxprt) {
> > +		dprintk("svcrdma: failed to create new transport\n");
> > +		return;
> > +	}
> > +	newxprt->sc_cm_id = new_cma_id;
> > +	new_cma_id->context = newxprt;
> > +	dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
> > +		newxprt, newxprt->sc_cm_id, listen_xprt);
> > +
> > +	/* Initialize the new transport */
> > +	newxprt->sc_xprt.sk_server = listen_xprt->sc_xprt.sk_server;
> > +	newxprt->sc_xprt.sk_lastrecv = get_seconds();
> > +	newxprt->sc_xprt.sk_delete = svc_rdma_delete;
> > +	newxprt->sc_xprt.sk_recvfrom = svc_rdma_recvfrom;
> > +	newxprt->sc_xprt.sk_sendto = svc_rdma_sendto;
> > +	newxprt->sc_xprt.sk_put = svc_rdma_put;
> > +	newxprt->sc_xprt.sk_prep_reply_buf = svc_rdma_prep_reply_buf;
> > +	newxprt->sc_xprt.sk_defer = svc_rdma_defer;
> > +	newxprt->sc_xprt.sk_revisit = svc_rdma_revisit;
> > +	newxprt->sc_xprt.sk_pool = NULL;
> > +
> > +	atomic_set(&newxprt->sc_xprt.sk_inuse, 1);
> > +	set_bit(SK_TEMP, &newxprt->sc_xprt.sk_flags);
> > +	INIT_LIST_HEAD(&newxprt->sc_xprt.sk_ready);
> > +	INIT_LIST_HEAD(&newxprt->sc_xprt.sk_list);
> > +	INIT_LIST_HEAD(&newxprt->sc_xprt.sk_deferred);
> > +	spin_lock_init(&newxprt->sc_xprt.sk_defer_lock);
> > +	mutex_init(&newxprt->sc_xprt.sk_mutex);
> > +
> > +	/* Enqueue the new transport on the accept queue of the listening
> > +	 * transport */
> > +	spin_lock_bh(&listen_xprt->sc_lock);
> > +	list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
> > +	spin_unlock_bh(&listen_xprt->sc_lock);
> 
> Why do you need bh-safe spinlocks here?
> 

This list is appended to by a callback running in another context in the
rdma_cm. I actually think that this is a thread and not a tasklet. I
think this is a case of blindly copying the svcsock locking style.

> > +
> > +	listen_xprt->sc_xprt.sk_pool = NULL;
> > +	set_bit(SK_CONN, &listen_xprt->sc_xprt.sk_flags);
> > +	svc_sock_enqueue(&listen_xprt->sc_xprt);
> > +}
> > +
> > +/*
> > + * Handles events generated on the listening endpoint. These events will be
> > + * either be incoming connect requests or adapter removal  events.
> > + * @param cma_id The CMA ID for the listening endpoint
> > + * @event the event being delivered.
> > + */
> > +static int
> > +rdma_listen_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
> > +{
> > +	struct svcxprt_rdma *xprt = cma_id->context;
> > +	int ret = 0;
> > +
> > +	switch (event->event) {
> > +	case RDMA_CM_EVENT_CONNECT_REQUEST:
> > +		dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, event=%d\n",
> > +			cma_id, cma_id->context, event->event);
> > +		handle_connect_req(cma_id);
> > +		break;
> > +
> > +	case RDMA_CM_EVENT_ESTABLISHED:
> > +		/* Accept complete */
> > +		dprintk("svcrdma: Connection completed on LISTEN xprt=%p, cm_id=%p\n",
> > +			xprt, cma_id);
> > +		break;
> > +
> > +	case RDMA_CM_EVENT_DEVICE_REMOVAL:
> > +		dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
> > +			xprt, cma_id);
> > +		if (xprt)
> > +			set_bit(SK_CLOSE, &xprt->sc_xprt.sk_flags);
> > +		break;
> > +
> > +	default:
> > +		dprintk("svcrdma: Unexpected event on listening endpoint %p, event=%d\n",
> > +			cma_id, event->event);
> > +		break;
> > +	}
> > +
> > +	return ret;
> > +}
> > +
> > +static int
> > +rdma_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
> > +{
> > +	struct svcxprt_rdma *xprt = cma_id->context;
> > +	int ret = 0;
> > +
> > +	switch (event->event) {
> > +	case RDMA_CM_EVENT_ESTABLISHED:
> > +		/* Accept complete */
> > +		dprintk("svcrdma: Connection completed on DTO xprt=%p, cm_id=%p\n",
> > +			xprt, cma_id);
> > +		break;
> > +
> > +	case RDMA_CM_EVENT_DISCONNECTED:
> > +		dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
> > +			xprt, cma_id);
> > +		if (xprt) {
> > +			xprt->sc_xprt.sk_pool = NULL;
> > +			set_bit(SK_CLOSE, &xprt->sc_xprt.sk_flags);
> > +			svc_sock_enqueue(&xprt->sc_xprt);
> > +		}
> > +		break;
> > +
> > +	case RDMA_CM_EVENT_DEVICE_REMOVAL:
> > +		dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, event=%d\n",
> > +			cma_id, cma_id->context, event->event);
> > +		if (xprt) {
> > +			xprt->sc_xprt.sk_pool = NULL;
> > +			set_bit(SK_CLOSE, &xprt->sc_xprt.sk_flags);
> > +			svc_sock_enqueue(&xprt->sc_xprt);
> > +		}
> > +		break;
> > +
> > +	default:
> > +		dprintk("svcrdma: Unexpected event on DTO endpoint %p, event=%d\n",
> > +			cma_id, event->event);
> > +		break;
> > +	}
> > +
> > +	return ret;
> > +}
> > +
> > +/*
> > + * Create a listening RDMA service endpoint
> > + * @param serv the RPC service this instance will belong to
> > + * @param protocol the protocol for the instance
> > + * @param sa the address to bind the local interface to
> > + * @return 0 on success, negative value for errors
> > + */
> > +int svc_rdma_create_listen(struct svc_serv *serv, int protocol,
> > +			   struct sockaddr *sa)
> > +{
> > +	struct rdma_cm_id *listen_id;
> > +	struct svcxprt_rdma *cma_xprt;
> > +	struct svc_sock *xprt;
> > +	int ret;
> > +
> > +	dprintk("svcrdma: Creating RDMA socket\n");
> > +
> > +	cma_xprt = rdma_create_xprt(1);
> > +	if (!cma_xprt)
> > +		return -ENOMEM;
> > +
> > +	xprt = &cma_xprt->sc_xprt;
> > +	xprt->sk_delete = svc_rdma_delete;
> > +	xprt->sk_recvfrom = svc_rdma_accept;
> > +	xprt->sk_put = svc_rdma_put;
> > +	xprt->sk_prep_reply_buf = svc_rdma_prep_reply_buf;
> > +	xprt->sk_server = serv;
> > +	xprt->sk_lastrecv = get_seconds();
> > +	INIT_LIST_HEAD(&xprt->sk_ready);
> > +	INIT_LIST_HEAD(&xprt->sk_list);
> > +	INIT_LIST_HEAD(&xprt->sk_deferred);
> > +	spin_lock_init(&xprt->sk_defer_lock);
> > +	mutex_init(&xprt->sk_mutex);
> > +	xprt->sk_pool = NULL;
> > +	atomic_set(&xprt->sk_inuse, 1);
> > +	spin_lock_bh(&serv->sv_lock);
> > +	list_add(&xprt->sk_list, &serv->sv_permsocks);
> > +	spin_unlock_bh(&serv->sv_lock);
> > +	clear_bit(SK_BUSY, &xprt->sk_flags);
> > +
> > +	/*
> > +	 * We shouldn't receive any events (except device removal) on
> > +	 * the id until we submit the listen request.  Any events that
> > +	 * we do receive will get logged as errors and ignored
> > +	 */
> > +	listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
> > +	if (IS_ERR(listen_id)) {
> > +		ret = PTR_ERR(listen_id);
> > +		rdma_destroy_xprt(cma_xprt);
> > +		dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
> > +		return ret;
> > +	}
> > +	ret = rdma_bind_addr(listen_id, sa);
> > +	if (ret) {
> > +		ret = PTR_ERR(listen_id);
> > +		rdma_destroy_xprt(cma_xprt);
> > +		rdma_destroy_id(listen_id);
> > +		dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
> > +		return ret;
> > +	}
> > +	cma_xprt->sc_cm_id = listen_id;
> > +
> > +	/* The xprt is ready to process events at this point */
> > +	ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
> > +	if (ret) {
> > +		ret = PTR_ERR(listen_id);
> > +		rdma_destroy_id(listen_id);
> > +		rdma_destroy_xprt(cma_xprt);
> > +		dprintk("svcrdma: rdma_listen failed = %d\n", ret);
> > +		return ret;
> > +	}
> > +
> > +	return 0;
> > +}
> > +
> > +/*
> > + * This is the sk_recvfrom function for listening endpoints. It's purpose is
> > + * to accept incoming connections. The CMA callback handler has already
> > + * created a new transport and attached the new CMA ID.
> > + *
> > + * There is a queue of pending connections hung on the listening
> > + * transport. This queue contains the new svc_sock structure. This function
> > + * takes svc_sock structures off the accept_q and completes the
> > + * connection.
> > + */
> > +static int
> > +svc_rdma_accept(struct svc_rqst *rqstp)
> > +{
> > +	struct svc_sock *xprt = rqstp->rq_sock;
> > +	struct svcxprt_rdma *listen_xprt;
> > +	struct svcxprt_rdma *newxprt;
> > +	struct rdma_conn_param conn_param;
> > +	struct ib_qp_init_attr qp_attr;
> > +	struct ib_device_attr devattr;
> > +	int ret;
> > +	int i;
> > +
> > +	listen_xprt = (struct svcxprt_rdma*)xprt;
> > +	if (list_empty(&listen_xprt->sc_accept_q)) {
> > +		printk(KERN_INFO
> > +		       "svcrdma: woke-up with no pending connection!\n");
> > +		clear_bit(SK_CONN, &listen_xprt->sc_xprt.sk_flags);
> > +		BUG_ON(test_bit(SK_BUSY, &listen_xprt->sc_xprt.sk_flags)==0);
> > +		clear_bit(SK_BUSY, &listen_xprt->sc_xprt.sk_flags);
> > +		return 0;
> > +	}
> > +
> > +	/* Get the next entry off the accept list */
> > +	spin_lock_bh(&listen_xprt->sc_lock);
> > +	newxprt = list_entry(listen_xprt->sc_accept_q.next,
> > +			     struct svcxprt_rdma, sc_accept_q);
> > +	list_del_init(&newxprt->sc_accept_q);
> > +	spin_unlock_bh(&listen_xprt->sc_lock);
> > +
> > +	dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
> > +		newxprt, newxprt->sc_cm_id);
> > +
> > +	ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
> > +	if (ret) {
> > +		printk(KERN_ERR
> > +		       "svcrdma: could not query device attributes on "
> > +		       "device %p, rc=%d\n", 
> > +		       newxprt->sc_cm_id->device, ret);
> > +		goto errout;
> > +	}
> > +
> > +	/* Qualify the transport resource defaults with the
> > +	 * capabilities of this particular device */
> > +	newxprt->sc_max_sge = min((size_t)devattr.max_sge,
> > +				  (size_t)RPCSVC_MAXPAGES);
> > +	newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
> > +				   (size_t)svcrdma_max_requests);
> > +	newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
> > +
> > +	newxprt->sc_ord =  min((size_t)devattr.max_qp_rd_atom,
> > +			       (size_t)svcrdma_ord);
> > +	spin_lock_bh(&rqstp->rq_server->sv_lock);
> > +	list_add(&newxprt->sc_xprt.sk_list, &rqstp->rq_server->sv_tempsocks);
> > +	rqstp->rq_server->sv_tmpcnt ++;
> > +	spin_unlock_bh(&rqstp->rq_server->sv_lock);
> > +
> > +	newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
> > +	if (IS_ERR(newxprt->sc_pd)) {
> > +		printk(KERN_ERR
> > +		       "svcrdma: error creating PD for connect request\n");
> > +		ret = PTR_ERR(newxprt->sc_pd);
> > +		goto errout;
> > +	}
> > +	newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
> > +					 sq_comp_handler,
> > +					 cq_event_handler,
> > +					 newxprt,
> > +					 newxprt->sc_sq_depth);
> > +	if (IS_ERR(newxprt->sc_sq_cq)) {
> > +		printk(KERN_ERR
> > +		       "svcrdma: error creating SQ CQ for connect request\n");
> > +		ret = PTR_ERR(newxprt->sc_sq_cq);
> > +		goto errout;
> > +	}
> > +	newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
> > +					 rq_comp_handler,
> > +					 cq_event_handler,
> > +					 newxprt,
> > +					 newxprt->sc_max_requests);
> > +	if (IS_ERR(newxprt->sc_rq_cq)) {
> > +		printk(KERN_ERR
> > +		       "svcrdma: error creating RQ CQ for connect request\n");
> > +		ret = PTR_ERR(newxprt->sc_rq_cq);
> > +		goto errout;
> > +	}
> > +
> > +	memset(&qp_attr, 0, sizeof qp_attr);
> > +	qp_attr.event_handler = qp_event_handler;
> > +	qp_attr.qp_context = newxprt;
> > +	qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
> > +	qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
> > +	qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
> > +	qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
> > +	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
> > +	qp_attr.qp_type = IB_QPT_RC;
> > +	qp_attr.send_cq = newxprt->sc_sq_cq;
> > +	qp_attr.recv_cq = newxprt->sc_rq_cq;
> > +	printk("newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
> > +	       "cm_id->device=%p, sc_pd->device=%p\n"
> > +	       "qp_attr.cap.max_send_wr = %d\n"
> > +	       "qp_attr.cap.max_recv_wr = %d\n"
> > +	       "qp_attr.cap.max_send_sge = %d\n"
> > +	       "qp_attr.cap.max_recv_sge = %d\n",
> > +	       newxprt->sc_cm_id, newxprt->sc_pd,
> > +	       newxprt->sc_cm_id->device, newxprt->sc_pd->device,
> > +	       qp_attr.cap.max_send_wr,
> > +	       qp_attr.cap.max_recv_wr,
> > +	       qp_attr.cap.max_send_sge,
> > +	       qp_attr.cap.max_recv_sge);
> > +
> > +	ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
> > +	if (ret) {
> > +		/* 
> > +		 * XXX: This is a hack. We need a xx_request_qp interface
> > +		 * that will adjust the qp_attr's with a best-effort
> > +		 * number 
> > +		 */
> > +		qp_attr.cap.max_send_sge -= 2;
> > +		qp_attr.cap.max_recv_sge -= 2;
> > +		ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
> > +		if (ret) {
> > +			printk(KERN_ERR "svcrdma: failed to create QP, ret=%d\n", ret);
> > +			goto errout;
> > +		}
> > +		newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
> > +		newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
> > +		newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
> > +		newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
> > +	}
> > +	newxprt->sc_qp = newxprt->sc_cm_id->qp;
> > +	DBG_DUMP_QP(__FUNCTION__, newxprt->sc_qp, &qp_attr);
> > +
> > +	/* Register all of physical memory */
> > +	newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
> > +					    IB_ACCESS_LOCAL_WRITE |
> > +					    IB_ACCESS_REMOTE_WRITE);
> > +	if (IS_ERR(newxprt->sc_phys_mr)) {
> > +		ret = PTR_ERR(newxprt->sc_phys_mr);
> > +		printk(KERN_ERR
> > +		       "svcrdma: Failed to create DMA MR ret=%d\n", ret);
> > +		goto errout;
> > +	}
> > +
> > +	/* Post receive buffers */
> > +	for (i=0; i < newxprt->sc_max_requests; i++)
> > +		if ((ret = svc_rdma_post_recv(newxprt))) {
> > +			printk(KERN_ERR
> > +			       "svcrdma: failure posting receive buffers\n");
> > +			goto errout;
> > +		}
> > +
> > +	/* Swap out the handler */
> > +	newxprt->sc_cm_id->event_handler = rdma_cma_handler;
> > +
> > +	/* We will get a getattr request from the client before we see
> > +	 * the connect complete event because DTO's run on tasklets,
> > +	 * and connection events run on threads
> > +	*/
> > +	clear_bit(SK_BUSY, &newxprt->sc_xprt.sk_flags);
> > +
> > +	/* Accept Connection */
> > +	memset(&conn_param, 0, sizeof conn_param);
> > +	conn_param.responder_resources = 0;
> > +	conn_param.initiator_depth = newxprt->sc_ord;
> > +	ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
> > +	if (ret) {
> > +		printk(KERN_ERR
> > +		       "svcrdma: failed to accept new connection, ret=%d\n",
> > +		       ret);
> > +		goto errout;
> > +	}
> > +
> > +	printk("svcrdma: new connection %p accepted with the following "
> > +		"attributes:\n"
> > +		"\tlocal_ip        : %d.%d.%d.%d\n"
> > +		"\tlocal_port	   : %d\n"
> > +		"\tremote_ip       : %d.%d.%d.%d\n"
> > +		"\tremote_port     : %d\n"
> > +		"\tmax_sge         : %d\n"
> > +		"\tsq_depth        : %d\n"
> > +		"\tmax_requests    : %d\n"
> > +		"\tread throttle   : %s\n"
> > +		"\tord             : %d\n",
> > +		newxprt,
> > +		NIPQUAD(((struct sockaddr_in*)&newxprt->sc_cm_id->
> > +			 route.addr.src_addr)->sin_addr.s_addr),
> > +		ntohs(((struct sockaddr_in*)&newxprt->sc_cm_id->
> > +		       route.addr.src_addr)->sin_port),
> > +		NIPQUAD(((struct sockaddr_in*)&newxprt->sc_cm_id->
> > +			 route.addr.dst_addr)->sin_addr.s_addr),
> > +		ntohs(((struct sockaddr_in*)&newxprt->sc_cm_id->
> > +		       route.addr.dst_addr)->sin_port),
> > +		newxprt->sc_max_sge,
> > +		newxprt->sc_sq_depth,
> > +		newxprt->sc_max_requests,
> > +		(svcrdma_read_throttle?"TRUE":"FALSE"),
> > +		newxprt->sc_ord);
> > +
> > +	spin_lock_bh(&listen_xprt->sc_lock);
> > +	if (list_empty(&listen_xprt->sc_accept_q))
> > +		clear_bit(SK_CONN, &listen_xprt->sc_xprt.sk_flags);
> > +	spin_unlock_bh(&listen_xprt->sc_lock);
> > +	listen_xprt->sc_xprt.sk_pool = NULL;
> > +	BUG_ON(test_bit(SK_BUSY, &listen_xprt->sc_xprt.sk_flags)==0);
> > +	clear_bit(SK_BUSY, &listen_xprt->sc_xprt.sk_flags);
> > +	svc_sock_enqueue(&listen_xprt->sc_xprt);
> > +
> > +	ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
> > +	ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
> > +	return ret;
> > +
> > + errout:
> > +	printk(KERN_ERR "svcrdma: failure accepting new connection rc=%d.\n", 
> > +	       ret);
> > +	BUG_ON(test_bit(SK_BUSY, &listen_xprt->sc_xprt.sk_flags)==0);
> > +	clear_bit(SK_BUSY, &listen_xprt->sc_xprt.sk_flags);
> > +	clear_bit(SK_CONN, &listen_xprt->sc_xprt.sk_flags);
> > +	rdma_destroy_id(newxprt->sc_cm_id);
> > +	rdma_destroy_xprt(newxprt);
> > +	return 0; /* ret; */
> > +}
> > +
> > +static void svc_rdma_delete(struct svc_sock *xprt)
> > +{
> > +	struct svc_serv *serv = xprt->sk_server;
> > +
> > +	spin_lock_bh(&serv->sv_lock);
> > +	if (!test_and_set_bit(SK_DETACHED, &xprt->sk_flags))
> > +		list_del_init(&xprt->sk_list);
> > +
> > +	if (!test_and_set_bit(SK_DEAD, &xprt->sk_flags)) {
> > +		BUG_ON(atomic_read(&xprt->sk_inuse)<2);
> > +		atomic_dec(&xprt->sk_inuse);
> > +		if (test_bit(SK_TEMP, &xprt->sk_flags))
> > +			serv->sv_tmpcnt--;
> > +	}
> > +	spin_unlock_bh(&serv->sv_lock);
> > +}
> > +
> > +static void rdma_destroy_xprt(struct svcxprt_rdma *xprt)
> > +{
> > +	if (xprt->sc_qp)
> > +		ib_destroy_qp(xprt->sc_qp);
> > +
> > +	if (xprt->sc_sq_cq)
> > +		ib_destroy_cq(xprt->sc_sq_cq);
> > +
> > +	if (xprt->sc_rq_cq)
> > +		ib_destroy_cq(xprt->sc_rq_cq);
> > +
> > +	if (xprt->sc_pd)
> > +		ib_dealloc_pd(xprt->sc_pd);
> > +
> > +	destroy_context_cache(xprt->sc_ctxt_head);
> > +
> > +	if (xprt->sc_xprt.sk_info_authunix != NULL)
> > +		svcauth_unix_info_release(xprt->sc_xprt.sk_info_authunix);
> > +
> > +	kfree(xprt);
> > +}
> > +
> > +int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
> > +{
> > +	struct ib_send_wr *bad_wr;
> > +	int ret;
> > +
> > +	if (test_bit(SK_CLOSE, &xprt->sc_xprt.sk_flags))
> > +		return 0;
> > +
> > +	BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
> > +	BUG_ON(((struct svc_rdma_op_ctxt*)(unsigned long)wr->wr_id)->wr_op !=
> > +		wr->opcode);
> > +	/* If the SQ is full, wait until an SQ entry is available */
> > +	while (1) {
> > +		spin_lock_bh(&xprt->sc_lock);
> > +		if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
> > +			spin_unlock_bh(&xprt->sc_lock);
> > +			rdma_stat_sq_starve ++;
> > +			/* First see if we can opportunistically reap some SQ WR */
> > +			sq_cq_reap(xprt);
> > +
> > +			/* Wait until SQ WR available if SQ still full*/
> > +			wait_event(xprt->sc_send_wait,
> > +				   atomic_read(&xprt->sc_sq_count) < xprt->sc_sq_depth);
> > +			continue;
> > +		}
> > +		/* Bumped used SQ WR count and post */
> > +		ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
> > +		if (!ret)
> > +			atomic_inc(&xprt->sc_sq_count);
> > +		else {
> > +			printk(KERN_ERR "svcrdma: failed to post SQ WR rc=%d, "
> > +			       "sc_sq_count=%d, sc_sq_depth=%d\n",
> > +			       ret, atomic_read(&xprt->sc_sq_count),
> > +			       xprt->sc_sq_depth);
> > +		}
> > +		spin_unlock_bh(&xprt->sc_lock);
> > +		break;
> > +	}
> > +
> > +	return ret;
> > +}
> > +
> > +int svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
> > +		    enum rpcrdma_errcode err)
> > +{
> > +	struct ib_send_wr err_wr;
> > +	struct ib_sge sge;
> > +	struct page *p;
> > +	struct svc_rdma_op_ctxt *ctxt;
> > +	u32 *va;
> > +	int length;
> > +	int ret;
> > +
> > +	p = svc_rdma_get_page();
> > +	va = page_address(p);
> > +
> > +	/* XDR encode error */
> > +	length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
> > +
> > +	/* Prepare SGE for local address */
> > +	sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
> > +				   p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
> > +	sge.lkey = xprt->sc_phys_mr->lkey;
> > +	sge.length = length;
> > +
> > +	ctxt = svc_rdma_get_context(xprt);
> > +	ctxt->count = 1;
> > +	ctxt->pages[0] = p;
> > +
> > +	/* Prepare SEND WR */
> > +	memset(&err_wr, 0, sizeof err_wr);
> > +	ctxt->wr_op = IB_WR_SEND;
> > +	err_wr.wr_id = (unsigned long)ctxt;
> > +	err_wr.sg_list = &sge;
> > +	err_wr.num_sge = 1;
> > +	err_wr.opcode = IB_WR_SEND;
> > +	err_wr.send_flags = IB_SEND_SIGNALED;
> > +
> > +	/* Post It */
> > +	ret = svc_rdma_send(xprt, &err_wr);
> > +	if (ret) {
> > +		dprintk("svcrdma: Error posting send = %d\n", ret);
> > +		svc_rdma_put_context(ctxt,1);
> > +	}
> > +
> > +	return ret;
> > +}
> > +
> > +/*
> > + * Setup the reply buffer for the svc_process function to write the
> > + * RPC into.
> > + */
> > +static int svc_rdma_prep_reply_buf(struct svc_rqst *rqstp)
> > +{
> > +	struct kvec *resv = &rqstp->rq_res.head[0];
> > +
> > +	/* setup response xdr_buf.
> > +	 * Initially it has just one page
> > +	 */
> > +	rqstp->rq_resused = 1;
> > +	resv->iov_base = page_address(rqstp->rq_respages[0]);
> > +	resv->iov_len = 0;
> > +	rqstp->rq_res.pages = rqstp->rq_respages+1;
> > +	rqstp->rq_res.len = 0;
> > +	rqstp->rq_res.page_base = 0;
> > +	rqstp->rq_res.page_len = 0;
> > +	rqstp->rq_res.buflen = PAGE_SIZE;
> > +	rqstp->rq_res.tail[0].iov_base = NULL;
> > +	rqstp->rq_res.tail[0].iov_len = 0;
> > +
> > +	return 0;
> > +}
> > +
> > +/*
> > + * This request cannot be handled right now. Allocate a structure to
> > + * keep it's state pending completion processing. To accomplish this, the
> > + * function creates an svc_rdma_op_ctxt that looks like a receive completion and
> > + * enqueues it on the svc_sock's deferred request list.  When*
> > + * svc_rdma_recvfrom is subsequently called, it first checks if there is a
> > + * deferred RPC and if there is:
> > + * - Takes the deferred request off the deferred request queue
> > + * - Extracts the svc_rdma_op_ctxt from the deferred request structure
> > + * - Frees the deferred request structure
> > + * - Skips the ib_cq_poll call and processes the svc_rdma_op_ctxt as if it had
> > + *   just come out of an WR pulled from the CQ.
> > + */
> > +static struct cache_deferred_req *
> > +svc_rdma_defer(struct cache_req *req)
> > +{
> > +	struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
> > +	struct svcxprt_rdma *xprt;
> > +	struct svc_rdma_deferred_req *dr;
> > +
> > +	dprintk("svcrdma: deferring request on \n"
> > +		"  rqstp=%p\n"
> > +		"  rqstp->rq_arg.len=%d\n", 
> > +		rqstp, 
> > +		rqstp->rq_arg.len);
> > +
> > +	/* if more than a page, give up FIXME */
> > +	if (rqstp->rq_arg.page_len)
> > +		return NULL;
> > +	BUG_ON(rqstp->rq_deferred);
> > +	xprt = (struct svcxprt_rdma*)rqstp->rq_sock;
> > + retry:
> > +	dr = kmalloc(sizeof(struct svc_rdma_deferred_req), GFP_KERNEL);
> > +	if (!dr) {
> > +		printk(KERN_INFO "svcrdma: sleeping waiting for memory\n");
> > +		schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
> > +		goto retry;
> > +	}
> > +	dr->req.handle.owner = rqstp->rq_server;
> > +	dr->req.prot = rqstp->rq_prot;
> > +	dr->req.addr = rqstp->rq_addr;
> > +	dr->req.daddr = rqstp->rq_daddr;
> > +	dr->req.argslen = rqstp->rq_arg.len >> 2;
> > +	dr->arg_page = rqstp->rq_pages[0];
> > +	dr->arg_len = rqstp->rq_arg.len;
> > +	rqstp->rq_pages[0] = svc_rdma_get_page();
> > +
> > +	atomic_inc(&rqstp->rq_sock->sk_inuse);
> > +	dr->req.svsk = rqstp->rq_sock;
> > +	dr->req.handle.revisit = rqstp->rq_sock->sk_revisit;
> > +
> > +	return &dr->req.handle;
> > +}
> > +
> > +/*
> > + * This is called by the cache code when it either gets an answer from
> > + * a user-mode daemon or gives up...as indicated by 'too_many'
> > + */
> > +static void svc_rdma_revisit(struct cache_deferred_req *dreq, int too_many)
> > +{
> > +	struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
> > +	struct svc_serv *serv = dreq->owner;
> > +	struct svc_sock *svsk;
> > +
> > +	if (unlikely(too_many)) {
> > +		printk(KERN_INFO "svcrdma: giving up on deferred request "
> > +		       "on svc_sock=%p, too many outstanding\n", dr->svsk);
> > +		dr->svsk->sk_put(dr->svsk);
> > +		kfree(dr);
> > +		return;
> > +	}
> > +	svsk = dr->svsk;
> > +	dprintk("svcrdma: revisit deferred RPC on xprt=%p\n", svsk);
> > +	dr->svsk = NULL;
> > +	spin_lock_bh(&serv->sv_lock);
> > +	list_add(&dr->handle.recent, &svsk->sk_deferred);
> > +	spin_unlock_bh(&serv->sv_lock);
> > +	svsk->sk_pool = NULL;
> > +	set_bit(SK_DEFERRED, &svsk->sk_flags);
> > +	svc_sock_enqueue(svsk);
> > +	svsk->sk_put(svsk);
> > +}
> > +
> > 
> > 
> > -------------------------------------------------------------------------
> > This SF.net email is sponsored by DB2 Express
> > Download DB2 Express C - the FREE version of DB2 express and take
> > control of your XML. No limits. Just data. Click to get it now.
> > http://sourceforge.net/powerbar/db2/
> > _______________________________________________
> > NFS maillist  -  NFS@lists.sourceforge.net
> > https://lists.sourceforge.net/lists/listinfo/nfs
> 


-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/
_______________________________________________
NFS maillist  -  NFS@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/nfs