From: Tom Tucker Subject: Re: [RFC,PATCH 11/15] knfsd: RDMA transport core Date: Fri, 18 May 2007 15:07:43 -0500 Message-ID: <1179518863.23385.195.camel@trinity.ogc.int> References: <1179510352.23385.123.camel@trinity.ogc.int> <1179515255.6488.136.camel@heimdal.trondhjem.org> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Cc: Neil Brown , Tom Talpey , Linux NFS Mailing List , Peter Leckie , Greg Banks To: Trond Myklebust Return-path: Received: from sc8-sf-mx1-b.sourceforge.net ([10.3.1.91] helo=mail.sourceforge.net) by sc8-sf-list2-new.sourceforge.net with esmtp (Exim 4.43) id 1Hp8js-0007SQ-L5 for nfs@lists.sourceforge.net; Fri, 18 May 2007 13:07:48 -0700 Received: from rrcs-71-42-183-126.sw.biz.rr.com ([71.42.183.126] helo=smtp.opengridcomputing.com) by mail.sourceforge.net with esmtp (Exim 4.44) id 1Hp8jt-0006Aa-Eh for nfs@lists.sourceforge.net; Fri, 18 May 2007 13:07:51 -0700 In-Reply-To: <1179515255.6488.136.camel@heimdal.trondhjem.org> List-Id: "Discussion of NFS under Linux development, interoperability, and testing." List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: nfs-bounces@lists.sourceforge.net Errors-To: nfs-bounces@lists.sourceforge.net 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