#include <linux/slab.h>
#include <linux/sunrpc/addr.h>
#include <linux/sunrpc/svc_rdma.h>
+#include <linux/log2.h>
#include <asm-generic/barrier.h>
#include <asm/bitops.h>
* internal functions
*/
static void rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc);
+static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf);
static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
static void rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf);
+static void rpcrdma_mr_free(struct rpcrdma_mr *mr);
static struct rpcrdma_regbuf *
rpcrdma_regbuf_alloc(size_t size, enum dma_data_direction direction,
gfp_t flags);
*/
static void rpcrdma_xprt_drain(struct rpcrdma_xprt *r_xprt)
{
- struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
struct rpcrdma_ia *ia = &r_xprt->rx_ia;
/* Flush Receives, then wait for deferred Reply work
* to complete.
*/
ib_drain_rq(ia->ri_id->qp);
- drain_workqueue(buf->rb_completion_wq);
/* Deferred Reply processing might have scheduled
* local invalidations.
struct rpcrdma_ep *ep = &r_xprt->rx_ep;
struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
struct rpcrdma_req *req;
- struct rpcrdma_rep *rep;
- cancel_delayed_work_sync(&buf->rb_refresh_worker);
+ cancel_work_sync(&buf->rb_refresh_worker);
/* This is similar to rpcrdma_ep_destroy, but:
* - Don't cancel the connect worker.
/* The ULP is responsible for ensuring all DMA
* mappings and MRs are gone.
*/
- list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
- rpcrdma_regbuf_dma_unmap(rep->rr_rdmabuf);
+ rpcrdma_reps_destroy(buf);
list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
rpcrdma_regbuf_dma_unmap(req->rl_rdmabuf);
rpcrdma_regbuf_dma_unmap(req->rl_sendbuf);
* completions recently. This is a sign the Send Queue is
* backing up. Cause the caller to pause and try again.
*/
- set_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags);
+ xprt_wait_for_buffer_space(&r_xprt->rx_xprt);
r_xprt->rx_stats.empty_sendctx_q++;
return NULL;
}
/* Paired with READ_ONCE */
smp_store_release(&buf->rb_sc_tail, next_tail);
- if (test_and_clear_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags)) {
- smp_mb__after_atomic();
- xprt_write_space(&sc->sc_xprt->rx_xprt);
- }
+ xprt_write_space(&sc->sc_xprt->rx_xprt);
}
static void
struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
struct rpcrdma_ia *ia = &r_xprt->rx_ia;
unsigned int count;
- LIST_HEAD(free);
- LIST_HEAD(all);
for (count = 0; count < ia->ri_max_segs; count++) {
struct rpcrdma_mr *mr;
int rc;
- mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+ mr = kzalloc(sizeof(*mr), GFP_NOFS);
if (!mr)
break;
mr->mr_xprt = r_xprt;
- list_add(&mr->mr_list, &free);
- list_add(&mr->mr_all, &all);
+ spin_lock(&buf->rb_lock);
+ list_add(&mr->mr_list, &buf->rb_mrs);
+ list_add(&mr->mr_all, &buf->rb_all_mrs);
+ spin_unlock(&buf->rb_lock);
}
- spin_lock(&buf->rb_mrlock);
- list_splice(&free, &buf->rb_mrs);
- list_splice(&all, &buf->rb_all);
r_xprt->rx_stats.mrs_allocated += count;
- spin_unlock(&buf->rb_mrlock);
trace_xprtrdma_createmrs(r_xprt, count);
-
- xprt_write_space(&r_xprt->rx_xprt);
}
static void
rpcrdma_mr_refresh_worker(struct work_struct *work)
{
struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
- rb_refresh_worker.work);
+ rb_refresh_worker);
struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
rx_buf);
rpcrdma_mrs_create(r_xprt);
+ xprt_write_space(&r_xprt->rx_xprt);
}
/**
struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
struct rpcrdma_regbuf *rb;
struct rpcrdma_req *req;
+ size_t maxhdrsize;
req = kzalloc(sizeof(*req), flags);
if (req == NULL)
goto out1;
- rb = rpcrdma_regbuf_alloc(RPCRDMA_HDRBUF_SIZE, DMA_TO_DEVICE, flags);
+ /* Compute maximum header buffer size in bytes */
+ maxhdrsize = rpcrdma_fixed_maxsz + 3 +
+ r_xprt->rx_ia.ri_max_segs * rpcrdma_readchunk_maxsz;
+ maxhdrsize *= sizeof(__be32);
+ rb = rpcrdma_regbuf_alloc(__roundup_pow_of_two(maxhdrsize),
+ DMA_TO_DEVICE, flags);
if (!rb)
goto out2;
req->rl_rdmabuf = rb;
if (!req->rl_recvbuf)
goto out4;
- req->rl_buffer = buffer;
+ INIT_LIST_HEAD(&req->rl_free_mrs);
INIT_LIST_HEAD(&req->rl_registered);
spin_lock(&buffer->rb_lock);
list_add(&req->rl_all, &buffer->rb_allreqs);
return NULL;
}
-static bool rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt, bool temp)
+static struct rpcrdma_rep *rpcrdma_rep_create(struct rpcrdma_xprt *r_xprt,
+ bool temp)
{
- struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
struct rpcrdma_rep *rep;
rep = kzalloc(sizeof(*rep), GFP_KERNEL);
DMA_FROM_DEVICE, GFP_KERNEL);
if (!rep->rr_rdmabuf)
goto out_free;
+
xdr_buf_init(&rep->rr_hdrbuf, rdmab_data(rep->rr_rdmabuf),
rdmab_length(rep->rr_rdmabuf));
-
rep->rr_cqe.done = rpcrdma_wc_receive;
rep->rr_rxprt = r_xprt;
- INIT_WORK(&rep->rr_work, rpcrdma_deferred_completion);
rep->rr_recv_wr.next = NULL;
rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
rep->rr_recv_wr.num_sge = 1;
rep->rr_temp = temp;
-
- spin_lock(&buf->rb_lock);
- list_add(&rep->rr_list, &buf->rb_recv_bufs);
- spin_unlock(&buf->rb_lock);
- return true;
+ return rep;
out_free:
kfree(rep);
out:
- return false;
+ return NULL;
+}
+
+static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
+{
+ rpcrdma_regbuf_free(rep->rr_rdmabuf);
+ kfree(rep);
+}
+
+static struct rpcrdma_rep *rpcrdma_rep_get_locked(struct rpcrdma_buffer *buf)
+{
+ struct llist_node *node;
+
+ /* Calls to llist_del_first are required to be serialized */
+ node = llist_del_first(&buf->rb_free_reps);
+ if (!node)
+ return NULL;
+ return llist_entry(node, struct rpcrdma_rep, rr_node);
+}
+
+static void rpcrdma_rep_put(struct rpcrdma_buffer *buf,
+ struct rpcrdma_rep *rep)
+{
+ if (!rep->rr_temp)
+ llist_add(&rep->rr_node, &buf->rb_free_reps);
+ else
+ rpcrdma_rep_destroy(rep);
+}
+
+static void rpcrdma_reps_destroy(struct rpcrdma_buffer *buf)
+{
+ struct rpcrdma_rep *rep;
+
+ while ((rep = rpcrdma_rep_get_locked(buf)) != NULL)
+ rpcrdma_rep_destroy(rep);
}
/**
struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
int i, rc;
- buf->rb_flags = 0;
buf->rb_max_requests = r_xprt->rx_ep.rep_max_requests;
buf->rb_bc_srv_max_requests = 0;
- spin_lock_init(&buf->rb_mrlock);
spin_lock_init(&buf->rb_lock);
INIT_LIST_HEAD(&buf->rb_mrs);
- INIT_LIST_HEAD(&buf->rb_all);
- INIT_DELAYED_WORK(&buf->rb_refresh_worker,
- rpcrdma_mr_refresh_worker);
+ INIT_LIST_HEAD(&buf->rb_all_mrs);
+ INIT_WORK(&buf->rb_refresh_worker, rpcrdma_mr_refresh_worker);
rpcrdma_mrs_create(r_xprt);
}
buf->rb_credits = 1;
- INIT_LIST_HEAD(&buf->rb_recv_bufs);
+ init_llist_head(&buf->rb_free_reps);
rc = rpcrdma_sendctxs_create(r_xprt);
if (rc)
goto out;
- buf->rb_completion_wq = alloc_workqueue("rpcrdma-%s",
- WQ_MEM_RECLAIM | WQ_HIGHPRI,
- 0,
- r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR]);
- if (!buf->rb_completion_wq) {
- rc = -ENOMEM;
- goto out;
- }
-
return 0;
out:
rpcrdma_buffer_destroy(buf);
return rc;
}
-static void rpcrdma_rep_destroy(struct rpcrdma_rep *rep)
-{
- rpcrdma_regbuf_free(rep->rr_rdmabuf);
- kfree(rep);
-}
-
/**
* rpcrdma_req_destroy - Destroy an rpcrdma_req object
* @req: unused object to be destroyed
* This function assumes that the caller prevents concurrent device
* unload and transport tear-down.
*/
-void
-rpcrdma_req_destroy(struct rpcrdma_req *req)
+void rpcrdma_req_destroy(struct rpcrdma_req *req)
{
list_del(&req->rl_all);
+ while (!list_empty(&req->rl_free_mrs))
+ rpcrdma_mr_free(rpcrdma_mr_pop(&req->rl_free_mrs));
+
rpcrdma_regbuf_free(req->rl_recvbuf);
rpcrdma_regbuf_free(req->rl_sendbuf);
rpcrdma_regbuf_free(req->rl_rdmabuf);
unsigned int count;
count = 0;
- spin_lock(&buf->rb_mrlock);
- while (!list_empty(&buf->rb_all)) {
- mr = list_entry(buf->rb_all.next, struct rpcrdma_mr, mr_all);
+ spin_lock(&buf->rb_lock);
+ while ((mr = list_first_entry_or_null(&buf->rb_all_mrs,
+ struct rpcrdma_mr,
+ mr_all)) != NULL) {
list_del(&mr->mr_all);
-
- spin_unlock(&buf->rb_mrlock);
-
- /* Ensure MW is not on any rl_registered list */
- if (!list_empty(&mr->mr_list))
- list_del(&mr->mr_list);
+ spin_unlock(&buf->rb_lock);
frwr_release_mr(mr);
count++;
- spin_lock(&buf->rb_mrlock);
+ spin_lock(&buf->rb_lock);
}
- spin_unlock(&buf->rb_mrlock);
+ spin_unlock(&buf->rb_lock);
r_xprt->rx_stats.mrs_allocated = 0;
-
- dprintk("RPC: %s: released %u MRs\n", __func__, count);
}
/**
void
rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
{
- cancel_delayed_work_sync(&buf->rb_refresh_worker);
-
- if (buf->rb_completion_wq) {
- destroy_workqueue(buf->rb_completion_wq);
- buf->rb_completion_wq = NULL;
- }
+ cancel_work_sync(&buf->rb_refresh_worker);
rpcrdma_sendctxs_destroy(buf);
-
- while (!list_empty(&buf->rb_recv_bufs)) {
- struct rpcrdma_rep *rep;
-
- rep = list_first_entry(&buf->rb_recv_bufs,
- struct rpcrdma_rep, rr_list);
- list_del(&rep->rr_list);
- rpcrdma_rep_destroy(rep);
- }
+ rpcrdma_reps_destroy(buf);
while (!list_empty(&buf->rb_send_bufs)) {
struct rpcrdma_req *req;
rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
{
struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
- struct rpcrdma_mr *mr = NULL;
-
- spin_lock(&buf->rb_mrlock);
- if (!list_empty(&buf->rb_mrs))
- mr = rpcrdma_mr_pop(&buf->rb_mrs);
- spin_unlock(&buf->rb_mrlock);
+ struct rpcrdma_mr *mr;
- if (!mr)
- goto out_nomrs;
+ spin_lock(&buf->rb_lock);
+ mr = rpcrdma_mr_pop(&buf->rb_mrs);
+ spin_unlock(&buf->rb_lock);
return mr;
-
-out_nomrs:
- trace_xprtrdma_nomrs(r_xprt);
- if (r_xprt->rx_ep.rep_connected != -ENODEV)
- schedule_delayed_work(&buf->rb_refresh_worker, 0);
-
- /* Allow the reply handler and refresh worker to run */
- cond_resched();
-
- return NULL;
-}
-
-static void
-__rpcrdma_mr_put(struct rpcrdma_buffer *buf, struct rpcrdma_mr *mr)
-{
- spin_lock(&buf->rb_mrlock);
- rpcrdma_mr_push(mr, &buf->rb_mrs);
- spin_unlock(&buf->rb_mrlock);
-}
-
-/**
- * rpcrdma_mr_put - Release an rpcrdma_mr object
- * @mr: object to release
- *
- */
-void
-rpcrdma_mr_put(struct rpcrdma_mr *mr)
-{
- __rpcrdma_mr_put(&mr->mr_xprt->rx_buf, mr);
}
/**
- * rpcrdma_mr_unmap_and_put - DMA unmap an MR and release it
- * @mr: object to release
+ * rpcrdma_mr_put - DMA unmap an MR and release it
+ * @mr: MR to release
*
*/
-void
-rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr)
+void rpcrdma_mr_put(struct rpcrdma_mr *mr)
{
struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
mr->mr_sg, mr->mr_nents, mr->mr_dir);
mr->mr_dir = DMA_NONE;
}
- __rpcrdma_mr_put(&r_xprt->rx_buf, mr);
+
+ rpcrdma_mr_push(mr, &mr->mr_req->rl_free_mrs);
+}
+
+static void rpcrdma_mr_free(struct rpcrdma_mr *mr)
+{
+ struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
+ struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
+
+ mr->mr_req = NULL;
+ spin_lock(&buf->rb_lock);
+ rpcrdma_mr_push(mr, &buf->rb_mrs);
+ spin_unlock(&buf->rb_lock);
}
/**
/**
* rpcrdma_buffer_put - Put request/reply buffers back into pool
+ * @buffers: buffer pool
* @req: object to return
*
*/
-void
-rpcrdma_buffer_put(struct rpcrdma_req *req)
+void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req)
{
- struct rpcrdma_buffer *buffers = req->rl_buffer;
- struct rpcrdma_rep *rep = req->rl_reply;
-
+ if (req->rl_reply)
+ rpcrdma_rep_put(buffers, req->rl_reply);
req->rl_reply = NULL;
spin_lock(&buffers->rb_lock);
list_add(&req->rl_list, &buffers->rb_send_bufs);
- if (rep) {
- if (!rep->rr_temp) {
- list_add(&rep->rr_list, &buffers->rb_recv_bufs);
- rep = NULL;
- }
- }
spin_unlock(&buffers->rb_lock);
- if (rep)
- rpcrdma_rep_destroy(rep);
}
-/*
- * Put reply buffers back into pool when not attached to
- * request. This happens in error conditions.
+/**
+ * rpcrdma_recv_buffer_put - Release rpcrdma_rep back to free list
+ * @rep: rep to release
+ *
+ * Used after error conditions.
*/
-void
-rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
+void rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
{
- struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
-
- if (!rep->rr_temp) {
- spin_lock(&buffers->rb_lock);
- list_add(&rep->rr_list, &buffers->rb_recv_bufs);
- spin_unlock(&buffers->rb_lock);
- } else {
- rpcrdma_rep_destroy(rep);
- }
+ rpcrdma_rep_put(&rep->rr_rxprt->rx_buf, rep);
}
/* Returns a pointer to a rpcrdma_regbuf object, or NULL.
struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
int rc;
- if (!ep->rep_send_count ||
- test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) {
+ if (!ep->rep_send_count || kref_read(&req->rl_kref) > 1) {
send_wr->send_flags |= IB_SEND_SIGNALED;
ep->rep_send_count = ep->rep_send_batch;
} else {
{
struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
struct rpcrdma_ep *ep = &r_xprt->rx_ep;
- struct ib_recv_wr *wr, *bad_wr;
+ struct ib_recv_wr *i, *wr, *bad_wr;
+ struct rpcrdma_rep *rep;
int needed, count, rc;
rc = 0;
count = 0;
+
needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
if (ep->rep_receive_count > needed)
goto out;
if (!temp)
needed += RPCRDMA_MAX_RECV_BATCH;
- count = 0;
+ /* fast path: all needed reps can be found on the free list */
wr = NULL;
while (needed) {
- struct rpcrdma_regbuf *rb;
- struct rpcrdma_rep *rep;
-
- spin_lock(&buf->rb_lock);
- rep = list_first_entry_or_null(&buf->rb_recv_bufs,
- struct rpcrdma_rep, rr_list);
- if (likely(rep))
- list_del(&rep->rr_list);
- spin_unlock(&buf->rb_lock);
- if (!rep) {
- if (!rpcrdma_rep_create(r_xprt, temp))
- break;
- continue;
- }
-
- rb = rep->rr_rdmabuf;
- if (!rpcrdma_regbuf_dma_map(r_xprt, rb)) {
- rpcrdma_recv_buffer_put(rep);
+ rep = rpcrdma_rep_get_locked(buf);
+ if (!rep)
+ rep = rpcrdma_rep_create(r_xprt, temp);
+ if (!rep)
break;
- }
- trace_xprtrdma_post_recv(rep->rr_recv_wr.wr_cqe);
rep->rr_recv_wr.next = wr;
wr = &rep->rr_recv_wr;
- ++count;
--needed;
}
- if (!count)
+ if (!wr)
goto out;
+ for (i = wr; i; i = i->next) {
+ rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
+
+ if (!rpcrdma_regbuf_dma_map(r_xprt, rep->rr_rdmabuf))
+ goto release_wrs;
+
+ trace_xprtrdma_post_recv(rep);
+ ++count;
+ }
+
rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
(const struct ib_recv_wr **)&bad_wr);
+out:
+ trace_xprtrdma_post_recvs(r_xprt, count, rc);
if (rc) {
- for (wr = bad_wr; wr; wr = wr->next) {
+ for (wr = bad_wr; wr;) {
struct rpcrdma_rep *rep;
rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
+ wr = wr->next;
rpcrdma_recv_buffer_put(rep);
--count;
}
}
ep->rep_receive_count += count;
-out:
- trace_xprtrdma_post_recvs(r_xprt, count, rc);
+ return;
+
+release_wrs:
+ for (i = wr; i;) {
+ rep = container_of(i, struct rpcrdma_rep, rr_recv_wr);
+ i = i->next;
+ rpcrdma_recv_buffer_put(rep);
+ }
}