xprtrdma: Acquire MRs in rpcrdma_register_external()
Acquiring 64 MRs in rpcrdma_buffer_get() while holding the buffer
pool lock is expensive, and unnecessary because most modern adapters
can transfer 100s of KBs of payload using just a single MR.
Instead, acquire MRs one-at-a-time as chunks are registered, and
return them to rb_mws immediately during deregistration.
Note: commit 539431a437d2 ("xprtrdma: Don't invalidate FRMRs if
registration fails") is reverted: There is now a valid case where
registration can fail (with -ENOMEM) but the QP is still in RTS.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Tested-By: Devesh Sharma <devesh.sharma@avagotech.com>
Reviewed-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
diff --git a/net/sunrpc/xprtrdma/frwr_ops.c b/net/sunrpc/xprtrdma/frwr_ops.c
index a06d9a3..133edf6 100644
--- a/net/sunrpc/xprtrdma/frwr_ops.c
+++ b/net/sunrpc/xprtrdma/frwr_ops.c
@@ -11,6 +11,62 @@
* but most complex memory registration mode.
*/
+/* Normal operation
+ *
+ * A Memory Region is prepared for RDMA READ or WRITE using a FAST_REG
+ * Work Request (frmr_op_map). When the RDMA operation is finished, this
+ * Memory Region is invalidated using a LOCAL_INV Work Request
+ * (frmr_op_unmap).
+ *
+ * Typically these Work Requests are not signaled, and neither are RDMA
+ * SEND Work Requests (with the exception of signaling occasionally to
+ * prevent provider work queue overflows). This greatly reduces HCA
+ * interrupt workload.
+ *
+ * As an optimization, frwr_op_unmap marks MRs INVALID before the
+ * LOCAL_INV WR is posted. If posting succeeds, the MR is placed on
+ * rb_mws immediately so that no work (like managing a linked list
+ * under a spinlock) is needed in the completion upcall.
+ *
+ * But this means that frwr_op_map() can occasionally encounter an MR
+ * that is INVALID but the LOCAL_INV WR has not completed. Work Queue
+ * ordering prevents a subsequent FAST_REG WR from executing against
+ * that MR while it is still being invalidated.
+ */
+
+/* Transport recovery
+ *
+ * ->op_map and the transport connect worker cannot run at the same
+ * time, but ->op_unmap can fire while the transport connect worker
+ * is running. Thus MR recovery is handled in ->op_map, to guarantee
+ * that recovered MRs are owned by a sending RPC, and not one where
+ * ->op_unmap could fire at the same time transport reconnect is
+ * being done.
+ *
+ * When the underlying transport disconnects, MRs are left in one of
+ * three states:
+ *
+ * INVALID: The MR was not in use before the QP entered ERROR state.
+ * (Or, the LOCAL_INV WR has not completed or flushed yet).
+ *
+ * STALE: The MR was being registered or unregistered when the QP
+ * entered ERROR state, and the pending WR was flushed.
+ *
+ * VALID: The MR was registered before the QP entered ERROR state.
+ *
+ * When frwr_op_map encounters STALE and VALID MRs, they are recovered
+ * with ib_dereg_mr and then are re-initialized. Beause MR recovery
+ * allocates fresh resources, it is deferred to a workqueue, and the
+ * recovered MRs are placed back on the rb_mws list when recovery is
+ * complete. frwr_op_map allocates another MR for the current RPC while
+ * the broken MR is reset.
+ *
+ * To ensure that frwr_op_map doesn't encounter an MR that is marked
+ * INVALID but that is about to be flushed due to a previous transport
+ * disconnect, the transport connect worker attempts to drain all
+ * pending send queue WRs before the transport is reconnected.
+ */
+
#include "xprt_rdma.h"
#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
@@ -250,9 +306,9 @@
struct ib_device *device = ia->ri_device;
enum dma_data_direction direction = rpcrdma_data_dir(writing);
struct rpcrdma_mr_seg *seg1 = seg;
- struct rpcrdma_mw *mw = seg1->rl_mw;
- struct rpcrdma_frmr *frmr = &mw->r.frmr;
- struct ib_mr *mr = frmr->fr_mr;
+ struct rpcrdma_mw *mw;
+ struct rpcrdma_frmr *frmr;
+ struct ib_mr *mr;
struct ib_send_wr fastreg_wr, *bad_wr;
u8 key;
int len, pageoff;
@@ -261,12 +317,25 @@
u64 pa;
int page_no;
+ mw = seg1->rl_mw;
+ seg1->rl_mw = NULL;
+ do {
+ if (mw)
+ __frwr_queue_recovery(mw);
+ mw = rpcrdma_get_mw(r_xprt);
+ if (!mw)
+ return -ENOMEM;
+ } while (mw->r.frmr.fr_state != FRMR_IS_INVALID);
+ frmr = &mw->r.frmr;
+ frmr->fr_state = FRMR_IS_VALID;
+
pageoff = offset_in_page(seg1->mr_offset);
seg1->mr_offset -= pageoff; /* start of page */
seg1->mr_len += pageoff;
len = -pageoff;
if (nsegs > ia->ri_max_frmr_depth)
nsegs = ia->ri_max_frmr_depth;
+
for (page_no = i = 0; i < nsegs;) {
rpcrdma_map_one(device, seg, direction);
pa = seg->mr_dma;
@@ -285,8 +354,6 @@
dprintk("RPC: %s: Using frmr %p to map %d segments (%d bytes)\n",
__func__, mw, i, len);
- frmr->fr_state = FRMR_IS_VALID;
-
memset(&fastreg_wr, 0, sizeof(fastreg_wr));
fastreg_wr.wr_id = (unsigned long)(void *)mw;
fastreg_wr.opcode = IB_WR_FAST_REG_MR;
@@ -298,6 +365,7 @@
fastreg_wr.wr.fast_reg.access_flags = writing ?
IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
IB_ACCESS_REMOTE_READ;
+ mr = frmr->fr_mr;
key = (u8)(mr->rkey & 0x000000FF);
ib_update_fast_reg_key(mr, ++key);
fastreg_wr.wr.fast_reg.rkey = mr->rkey;
@@ -307,6 +375,7 @@
if (rc)
goto out_senderr;
+ seg1->rl_mw = mw;
seg1->mr_rkey = mr->rkey;
seg1->mr_base = seg1->mr_dma + pageoff;
seg1->mr_nsegs = i;
@@ -315,10 +384,9 @@
out_senderr:
dprintk("RPC: %s: ib_post_send status %i\n", __func__, rc);
- ib_update_fast_reg_key(mr, --key);
- frmr->fr_state = FRMR_IS_INVALID;
while (i--)
rpcrdma_unmap_one(device, --seg);
+ __frwr_queue_recovery(mw);
return rc;
}
@@ -330,15 +398,19 @@
{
struct rpcrdma_mr_seg *seg1 = seg;
struct rpcrdma_ia *ia = &r_xprt->rx_ia;
+ struct rpcrdma_mw *mw = seg1->rl_mw;
struct ib_send_wr invalidate_wr, *bad_wr;
int rc, nsegs = seg->mr_nsegs;
- seg1->rl_mw->r.frmr.fr_state = FRMR_IS_INVALID;
+ dprintk("RPC: %s: FRMR %p\n", __func__, mw);
+
+ seg1->rl_mw = NULL;
+ mw->r.frmr.fr_state = FRMR_IS_INVALID;
memset(&invalidate_wr, 0, sizeof(invalidate_wr));
- invalidate_wr.wr_id = (unsigned long)(void *)seg1->rl_mw;
+ invalidate_wr.wr_id = (unsigned long)(void *)mw;
invalidate_wr.opcode = IB_WR_LOCAL_INV;
- invalidate_wr.ex.invalidate_rkey = seg1->rl_mw->r.frmr.fr_mr->rkey;
+ invalidate_wr.ex.invalidate_rkey = mw->r.frmr.fr_mr->rkey;
DECR_CQCOUNT(&r_xprt->rx_ep);
while (seg1->mr_nsegs--)
@@ -348,12 +420,13 @@
read_unlock(&ia->ri_qplock);
if (rc)
goto out_err;
+
+ rpcrdma_put_mw(r_xprt, mw);
return nsegs;
out_err:
- /* Force rpcrdma_buffer_get() to retry */
- seg1->rl_mw->r.frmr.fr_state = FRMR_IS_STALE;
dprintk("RPC: %s: ib_post_send status %i\n", __func__, rc);
+ __frwr_queue_recovery(mw);
return nsegs;
}
@@ -400,6 +473,9 @@
{
struct rpcrdma_mw *r;
+ /* Ensure stale MWs for "buf" are no longer in flight */
+ flush_workqueue(frwr_recovery_wq);
+
while (!list_empty(&buf->rb_all)) {
r = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
list_del(&r->mw_all);