#define P9_RDMA_IRD 0
#define P9_RDMA_ORD 0
#define P9_RDMA_TIMEOUT 30000 /* 30 seconds */
-#define P9_RDMA_MAXSIZE (4*4096) /* Min SGE is 4, so we can
- * safely advertise a maxsize
- * of 64k */
+#define P9_RDMA_MAXSIZE (1024*1024) /* 1MB */
/**
* struct p9_trans_rdma - RDMA transport instance
* @sq_depth: The depth of the Send Queue
* @sq_sem: Semaphore for the SQ
* @rq_depth: The depth of the Receive Queue.
- * @rq_count: Count of requests in the Receive Queue.
+ * @rq_sem: Semaphore for the RQ
+ * @excess_rc : Amount of posted Receive Contexts without a pending request.
+ * See rdma_request()
* @addr: The remote peer's address
* @req_lock: Protects the active request list
* @cm_done: Completion event for connection management tracking
int sq_depth;
struct semaphore sq_sem;
int rq_depth;
- atomic_t rq_count;
+ struct semaphore rq_sem;
+ atomic_t excess_rc;
struct sockaddr_in addr;
spinlock_t req_lock;
if (!req)
goto err_out;
+ /* Check that we have not yet received a reply for this request.
+ */
+ if (unlikely(req->rc)) {
+ pr_err("Duplicate reply for request %d", tag);
+ goto err_out;
+ }
+
req->rc = c->rc;
req->status = REQ_STATUS_RCVD;
p9_client_cb(client, req);
switch (c->wc_op) {
case IB_WC_RECV:
- atomic_dec(&rdma->rq_count);
handle_recv(client, rdma, c, wc.status, wc.byte_len);
+ up(&rdma->rq_sem);
break;
case IB_WC_SEND:
struct p9_rdma_context *c = NULL;
struct p9_rdma_context *rpl_context = NULL;
+ /* When an error occurs between posting the recv and the send,
+ * there will be a receive context posted without a pending request.
+ * Since there is no way to "un-post" it, we remember it and skip
+ * post_recv() for the next request.
+ * So here,
+ * see if we are this `next request' and need to absorb an excess rc.
+ * If yes, then drop and free our own, and do not recv_post().
+ **/
+ if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
+ if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
+ /* Got one ! */
+ kfree(req->rc);
+ req->rc = NULL;
+ goto dont_need_post_recv;
+ } else {
+ /* We raced and lost. */
+ atomic_inc(&rdma->excess_rc);
+ }
+ }
+
/* Allocate an fcall for the reply */
rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
if (!rpl_context) {
err = -ENOMEM;
- goto err_close;
- }
-
- /*
- * If the request has a buffer, steal it, otherwise
- * allocate a new one. Typically, requests should already
- * have receive buffers allocated and just swap them around
- */
- if (!req->rc) {
- req->rc = kmalloc(sizeof(struct p9_fcall)+client->msize,
- GFP_NOFS);
- if (req->rc) {
- req->rc->sdata = (char *) req->rc +
- sizeof(struct p9_fcall);
- req->rc->capacity = client->msize;
- }
+ goto recv_error;
}
rpl_context->rc = req->rc;
- if (!rpl_context->rc) {
- err = -ENOMEM;
- goto err_free2;
- }
/*
* Post a receive buffer for this request. We need to ensure
* outstanding request, so we must keep a count to avoid
* overflowing the RQ.
*/
- if (atomic_inc_return(&rdma->rq_count) <= rdma->rq_depth) {
- err = post_recv(client, rpl_context);
- if (err)
- goto err_free1;
- } else
- atomic_dec(&rdma->rq_count);
+ if (down_interruptible(&rdma->rq_sem)) {
+ err = -EINTR;
+ goto recv_error;
+ }
+ err = post_recv(client, rpl_context);
+ if (err) {
+ p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
+ goto recv_error;
+ }
/* remove posted receive buffer from request structure */
req->rc = NULL;
+dont_need_post_recv:
/* Post the request */
c = kmalloc(sizeof *c, GFP_NOFS);
if (!c) {
err = -ENOMEM;
- goto err_free1;
+ goto send_error;
}
c->req = req;
c->busa = ib_dma_map_single(rdma->cm_id->device,
c->req->tc->sdata, c->req->tc->size,
DMA_TO_DEVICE);
- if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
- goto error;
+ if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
+ err = -EIO;
+ goto send_error;
+ }
sge.addr = c->busa;
sge.length = c->req->tc->size;
wr.sg_list = &sge;
wr.num_sge = 1;
- if (down_interruptible(&rdma->sq_sem))
- goto error;
+ if (down_interruptible(&rdma->sq_sem)) {
+ err = -EINTR;
+ goto send_error;
+ }
- return ib_post_send(rdma->qp, &wr, &bad_wr);
+ err = ib_post_send(rdma->qp, &wr, &bad_wr);
+ if (err)
+ goto send_error;
- error:
+ /* Success */
+ return 0;
+
+ /* Handle errors that happened during or while preparing the send: */
+ send_error:
kfree(c);
- kfree(rpl_context->rc);
- kfree(rpl_context);
- p9_debug(P9_DEBUG_ERROR, "EIO\n");
- return -EIO;
- err_free1:
- kfree(rpl_context->rc);
- err_free2:
+ p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
+
+ /* Ach.
+ * We did recv_post(), but not send. We have one recv_post in excess.
+ */
+ atomic_inc(&rdma->excess_rc);
+ return err;
+
+ /* Handle errors that happened during or while preparing post_recv(): */
+ recv_error:
kfree(rpl_context);
- err_close:
spin_lock_irqsave(&rdma->req_lock, flags);
if (rdma->state < P9_RDMA_CLOSING) {
rdma->state = P9_RDMA_CLOSING;
spin_lock_init(&rdma->req_lock);
init_completion(&rdma->cm_done);
sema_init(&rdma->sq_sem, rdma->sq_depth);
- atomic_set(&rdma->rq_count, 0);
+ sema_init(&rdma->rq_sem, rdma->rq_depth);
+ atomic_set(&rdma->excess_rc, 0);
return rdma;
}
return 1;
}
+/* A request has been fully flushed without a reply.
+ * That means we have posted one buffer in excess.
+ */
+static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
+{
+ struct p9_trans_rdma *rdma = client->trans;
+
+ atomic_inc(&rdma->excess_rc);
+ return 0;
+}
+
/**
* trans_create_rdma - Transport method for creating atransport instance
* @client: client instance
projects / ~andy / linux / blobdiff