Merge tag 'nfs-rdma-4.5' of git://git.linux-nfs.org/projects/anna/nfs-rdma (8d0ed0ca) · Commits · 戴 / test

net/sunrpc/xprtrdma/backchannel.c

+12 −14

Original line number	Diff line number	Diff line
		@@ -15,7 +15,7 @@
		# define RPCDBG_FACILITY RPCDBG_TRANS
		#endif

		#define RPCRDMA_BACKCHANNEL_DEBUG
		#undef RPCRDMA_BACKCHANNEL_DEBUG

		static void rpcrdma_bc_free_rqst(struct rpcrdma_xprt *r_xprt,
		struct rpc_rqst *rqst)
		@@ -42,8 +42,8 @@ static int rpcrdma_bc_setup_rqst(struct rpcrdma_xprt *r_xprt,
		size_t size;

		req = rpcrdma_create_req(r_xprt);
		if (!req)
		return -ENOMEM;
		if (IS_ERR(req))
		return PTR_ERR(req);
		req->rl_backchannel = true;

		size = RPCRDMA_INLINE_WRITE_THRESHOLD(rqst);
		@@ -84,9 +84,7 @@ out_fail:
		static int rpcrdma_bc_setup_reps(struct rpcrdma_xprt *r_xprt,
		unsigned int count)
		{
		struct rpcrdma_buffer *buffers = &r_xprt->rx_buf;
		struct rpcrdma_rep *rep;
		unsigned long flags;
		int rc = 0;

		while (count--) {
		@@ -98,9 +96,7 @@ static int rpcrdma_bc_setup_reps(struct rpcrdma_xprt *r_xprt,
		break;
		}

		spin_lock_irqsave(&buffers->rb_lock, flags);
		list_add(&rep->rr_list, &buffers->rb_recv_bufs);
		spin_unlock_irqrestore(&buffers->rb_lock, flags);
		rpcrdma_recv_buffer_put(rep);
		}

		return rc;
		@@ -140,6 +136,7 @@ int xprt_rdma_bc_setup(struct rpc_xprt *xprt, unsigned int reqs)
		__func__);
		goto out_free;
		}
		dprintk("RPC: %s: new rqst %p\n", __func__, rqst);

		rqst->rq_xprt = &r_xprt->rx_xprt;
		INIT_LIST_HEAD(&rqst->rq_list);
		@@ -220,12 +217,14 @@ int rpcrdma_bc_marshal_reply(struct rpc_rqst *rqst)

		rpclen = rqst->rq_svec[0].iov_len;

		#ifdef RPCRDMA_BACKCHANNEL_DEBUG
		pr_info("RPC: %s: rpclen %zd headerp 0x%p lkey 0x%x\n",
		__func__, rpclen, headerp, rdmab_lkey(req->rl_rdmabuf));
		pr_info("RPC: %s: RPC/RDMA: %*ph\n",
		__func__, (int)RPCRDMA_HDRLEN_MIN, headerp);
		pr_info("RPC: %s: RPC: %*ph\n",
		__func__, (int)rpclen, rqst->rq_svec[0].iov_base);
		#endif

		req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf);
		req->rl_send_iov[0].length = RPCRDMA_HDRLEN_MIN;
		@@ -269,6 +268,9 @@ void xprt_rdma_bc_free_rqst(struct rpc_rqst *rqst)
		{
		struct rpc_xprt *xprt = rqst->rq_xprt;

		dprintk("RPC: %s: freeing rqst %p (req %p)\n",
		__func__, rqst, rpcr_to_rdmar(rqst));

		smp_mb__before_atomic();
		WARN_ON_ONCE(!test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state));
		clear_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state);
		@@ -333,9 +335,7 @@ void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
		struct rpc_rqst, rq_bc_pa_list);
		list_del(&rqst->rq_bc_pa_list);
		spin_unlock(&xprt->bc_pa_lock);
		#ifdef RPCRDMA_BACKCHANNEL_DEBUG
		pr_info("RPC: %s: using rqst %p\n", __func__, rqst);
		#endif
		dprintk("RPC: %s: using rqst %p\n", __func__, rqst);

		/* Prepare rqst */
		rqst->rq_reply_bytes_recvd = 0;
		@@ -355,10 +355,8 @@ void rpcrdma_bc_receive_call(struct rpcrdma_xprt *r_xprt,
		* direction reply.
		*/
		req = rpcr_to_rdmar(rqst);
		#ifdef RPCRDMA_BACKCHANNEL_DEBUG
		pr_info("RPC: %s: attaching rep %p to req %p\n",
		dprintk("RPC: %s: attaching rep %p to req %p\n",
		__func__, rep, req);
		#endif
		req->rl_reply = rep;

		/* Defeat the retransmit detection logic in send_request */

net/sunrpc/xprtrdma/fmr_ops.c

+64 −0

Original line number	Diff line number	Diff line
		@@ -179,6 +179,69 @@ out_maperr:
		return rc;
		}

		static void
		__fmr_dma_unmap(struct rpcrdma_xprt r_xprt, struct rpcrdma_mr_seg seg)
		{
		struct ib_device *device = r_xprt->rx_ia.ri_device;
		struct rpcrdma_mw *mw = seg->rl_mw;
		int nsegs = seg->mr_nsegs;

		seg->rl_mw = NULL;

		while (nsegs--)
		rpcrdma_unmap_one(device, seg++);

		rpcrdma_put_mw(r_xprt, mw);
		}

		/* Invalidate all memory regions that were registered for "req".
		*
		* Sleeps until it is safe for the host CPU to access the
		* previously mapped memory regions.
		*/
		static void
		fmr_op_unmap_sync(struct rpcrdma_xprt r_xprt, struct rpcrdma_req req)
		{
		struct rpcrdma_mr_seg *seg;
		unsigned int i, nchunks;
		struct rpcrdma_mw *mw;
		LIST_HEAD(unmap_list);
		int rc;

		dprintk("RPC: %s: req %p\n", __func__, req);

		/* ORDER: Invalidate all of the req's MRs first
		*
		* ib_unmap_fmr() is slow, so use a single call instead
		* of one call per mapped MR.
		*/
		for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
		seg = &req->rl_segments[i];
		mw = seg->rl_mw;

		list_add(&mw->r.fmr.fmr->list, &unmap_list);

		i += seg->mr_nsegs;
		}
		rc = ib_unmap_fmr(&unmap_list);
		if (rc)
		pr_warn("%s: ib_unmap_fmr failed (%i)\n", __func__, rc);

		/* ORDER: Now DMA unmap all of the req's MRs, and return
		* them to the free MW list.
		*/
		for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
		seg = &req->rl_segments[i];

		__fmr_dma_unmap(r_xprt, seg);

		i += seg->mr_nsegs;
		seg->mr_nsegs = 0;
		}

		req->rl_nchunks = 0;
		}

		/* Use the ib_unmap_fmr() verb to prevent further remote
		* access via RDMA READ or RDMA WRITE.
		*/
		@@ -231,6 +294,7 @@ fmr_op_destroy(struct rpcrdma_buffer *buf)

		const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
		.ro_map = fmr_op_map,
		.ro_unmap_sync = fmr_op_unmap_sync,
		.ro_unmap = fmr_op_unmap,
		.ro_open = fmr_op_open,
		.ro_maxpages = fmr_op_maxpages,

net/sunrpc/xprtrdma/frwr_ops.c

+152 −22

Original line number	Diff line number	Diff line
		@@ -245,12 +245,14 @@ frwr_op_maxpages(struct rpcrdma_xprt *r_xprt)
		rpcrdma_max_segments(r_xprt) * ia->ri_max_frmr_depth);
		}

		/* If FAST_REG or LOCAL_INV failed, indicate the frmr needs to be reset. */
		/* If FAST_REG or LOCAL_INV failed, indicate the frmr needs
		* to be reset.
		*
		* WARNING: Only wr_id and status are reliable at this point
		*/
		static void
		frwr_sendcompletion(struct ib_wc *wc)
		__frwr_sendcompletion_flush(struct ib_wc wc, struct rpcrdma_mw r)
		{
		struct rpcrdma_mw *r;

		if (likely(wc->status == IB_WC_SUCCESS))
		return;

		@@ -261,9 +263,23 @@ frwr_sendcompletion(struct ib_wc *wc)
		else
		pr_warn("RPC: %s: frmr %p error, status %s (%d)\n",
		__func__, r, ib_wc_status_msg(wc->status), wc->status);

		r->r.frmr.fr_state = FRMR_IS_STALE;
		}

		static void
		frwr_sendcompletion(struct ib_wc *wc)
		{
		struct rpcrdma_mw r = (struct rpcrdma_mw )(unsigned long)wc->wr_id;
		struct rpcrdma_frmr *f = &r->r.frmr;

		if (unlikely(wc->status != IB_WC_SUCCESS))
		__frwr_sendcompletion_flush(wc, r);

		if (f->fr_waiter)
		complete(&f->fr_linv_done);
		}

		static int
		frwr_op_init(struct rpcrdma_xprt *r_xprt)
		{
		@@ -319,7 +335,7 @@ frwr_op_map(struct rpcrdma_xprt r_xprt, struct rpcrdma_mr_seg seg,
		struct rpcrdma_mw *mw;
		struct rpcrdma_frmr *frmr;
		struct ib_mr *mr;
		struct ib_reg_wr reg_wr;
		struct ib_reg_wr *reg_wr;
		struct ib_send_wr *bad_wr;
		int rc, i, n, dma_nents;
		u8 key;
		@@ -335,7 +351,9 @@ frwr_op_map(struct rpcrdma_xprt r_xprt, struct rpcrdma_mr_seg seg,
		} while (mw->r.frmr.fr_state != FRMR_IS_INVALID);
		frmr = &mw->r.frmr;
		frmr->fr_state = FRMR_IS_VALID;
		frmr->fr_waiter = false;
		mr = frmr->fr_mr;
		reg_wr = &frmr->fr_regwr;

		if (nsegs > ia->ri_max_frmr_depth)
		nsegs = ia->ri_max_frmr_depth;
		@@ -381,19 +399,19 @@ frwr_op_map(struct rpcrdma_xprt r_xprt, struct rpcrdma_mr_seg seg,
		key = (u8)(mr->rkey & 0x000000FF);
		ib_update_fast_reg_key(mr, ++key);

		reg_wr.wr.next = NULL;
		reg_wr.wr.opcode = IB_WR_REG_MR;
		reg_wr.wr.wr_id = (uintptr_t)mw;
		reg_wr.wr.num_sge = 0;
		reg_wr.wr.send_flags = 0;
		reg_wr.mr = mr;
		reg_wr.key = mr->rkey;
		reg_wr.access = writing ?
		reg_wr->wr.next = NULL;
		reg_wr->wr.opcode = IB_WR_REG_MR;
		reg_wr->wr.wr_id = (uintptr_t)mw;
		reg_wr->wr.num_sge = 0;
		reg_wr->wr.send_flags = 0;
		reg_wr->mr = mr;
		reg_wr->key = mr->rkey;
		reg_wr->access = writing ?
		IB_ACCESS_REMOTE_WRITE \| IB_ACCESS_LOCAL_WRITE :
		IB_ACCESS_REMOTE_READ;

		DECR_CQCOUNT(&r_xprt->rx_ep);
		rc = ib_post_send(ia->ri_id->qp, &reg_wr.wr, &bad_wr);
		rc = ib_post_send(ia->ri_id->qp, &reg_wr->wr, &bad_wr);
		if (rc)
		goto out_senderr;

		@@ -413,6 +431,116 @@ out_senderr:
		return rc;
		}

		static struct ib_send_wr *
		__frwr_prepare_linv_wr(struct rpcrdma_mr_seg *seg)
		{
		struct rpcrdma_mw *mw = seg->rl_mw;
		struct rpcrdma_frmr *f = &mw->r.frmr;
		struct ib_send_wr *invalidate_wr;

		f->fr_waiter = false;
		f->fr_state = FRMR_IS_INVALID;
		invalidate_wr = &f->fr_invwr;

		memset(invalidate_wr, 0, sizeof(*invalidate_wr));
		invalidate_wr->wr_id = (unsigned long)(void *)mw;
		invalidate_wr->opcode = IB_WR_LOCAL_INV;
		invalidate_wr->ex.invalidate_rkey = f->fr_mr->rkey;

		return invalidate_wr;
		}

		static void
		__frwr_dma_unmap(struct rpcrdma_xprt r_xprt, struct rpcrdma_mr_seg seg,
		int rc)
		{
		struct ib_device *device = r_xprt->rx_ia.ri_device;
		struct rpcrdma_mw *mw = seg->rl_mw;
		struct rpcrdma_frmr *f = &mw->r.frmr;

		seg->rl_mw = NULL;

		ib_dma_unmap_sg(device, f->sg, f->sg_nents, seg->mr_dir);

		if (!rc)
		rpcrdma_put_mw(r_xprt, mw);
		else
		__frwr_queue_recovery(mw);
		}

		/* Invalidate all memory regions that were registered for "req".
		*
		* Sleeps until it is safe for the host CPU to access the
		* previously mapped memory regions.
		*/
		static void
		frwr_op_unmap_sync(struct rpcrdma_xprt r_xprt, struct rpcrdma_req req)
		{
		struct ib_send_wr invalidate_wrs, pos, prev, bad_wr;
		struct rpcrdma_ia *ia = &r_xprt->rx_ia;
		struct rpcrdma_mr_seg *seg;
		unsigned int i, nchunks;
		struct rpcrdma_frmr *f;
		int rc;

		dprintk("RPC: %s: req %p\n", __func__, req);

		/* ORDER: Invalidate all of the req's MRs first
		*
		* Chain the LOCAL_INV Work Requests and post them with
		* a single ib_post_send() call.
		*/
		invalidate_wrs = pos = prev = NULL;
		seg = NULL;
		for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
		seg = &req->rl_segments[i];

		pos = __frwr_prepare_linv_wr(seg);

		if (!invalidate_wrs)
		invalidate_wrs = pos;
		else
		prev->next = pos;
		prev = pos;

		i += seg->mr_nsegs;
		}
		f = &seg->rl_mw->r.frmr;

		/* Strong send queue ordering guarantees that when the
		* last WR in the chain completes, all WRs in the chain
		* are complete.
		*/
		f->fr_invwr.send_flags = IB_SEND_SIGNALED;
		f->fr_waiter = true;
		init_completion(&f->fr_linv_done);
		INIT_CQCOUNT(&r_xprt->rx_ep);

		/* Transport disconnect drains the receive CQ before it
		* replaces the QP. The RPC reply handler won't call us
		* unless ri_id->qp is a valid pointer.
		*/
		rc = ib_post_send(ia->ri_id->qp, invalidate_wrs, &bad_wr);
		if (rc)
		pr_warn("%s: ib_post_send failed %i\n", __func__, rc);

		wait_for_completion(&f->fr_linv_done);

		/* ORDER: Now DMA unmap all of the req's MRs, and return
		* them to the free MW list.
		*/
		for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
		seg = &req->rl_segments[i];

		__frwr_dma_unmap(r_xprt, seg, rc);

		i += seg->mr_nsegs;
		seg->mr_nsegs = 0;
		}

		req->rl_nchunks = 0;
		}

		/* Post a LOCAL_INV Work Request to prevent further remote access
		* via RDMA READ or RDMA WRITE.
		*/
		@@ -423,23 +551,24 @@ frwr_op_unmap(struct rpcrdma_xprt r_xprt, struct rpcrdma_mr_seg seg)
		struct rpcrdma_ia *ia = &r_xprt->rx_ia;
		struct rpcrdma_mw *mw = seg1->rl_mw;
		struct rpcrdma_frmr *frmr = &mw->r.frmr;
		struct ib_send_wr invalidate_wr, *bad_wr;
		struct ib_send_wr invalidate_wr, bad_wr;
		int rc, nsegs = seg->mr_nsegs;

		dprintk("RPC: %s: FRMR %p\n", __func__, mw);

		seg1->rl_mw = NULL;
		frmr->fr_state = FRMR_IS_INVALID;
		invalidate_wr = &mw->r.frmr.fr_invwr;

		memset(&invalidate_wr, 0, sizeof(invalidate_wr));
		invalidate_wr.wr_id = (unsigned long)(void *)mw;
		invalidate_wr.opcode = IB_WR_LOCAL_INV;
		invalidate_wr.ex.invalidate_rkey = frmr->fr_mr->rkey;
		memset(invalidate_wr, 0, sizeof(*invalidate_wr));
		invalidate_wr->wr_id = (uintptr_t)mw;
		invalidate_wr->opcode = IB_WR_LOCAL_INV;
		invalidate_wr->ex.invalidate_rkey = frmr->fr_mr->rkey;
		DECR_CQCOUNT(&r_xprt->rx_ep);

		ib_dma_unmap_sg(ia->ri_device, frmr->sg, frmr->sg_nents, seg1->mr_dir);
		read_lock(&ia->ri_qplock);
		rc = ib_post_send(ia->ri_id->qp, &invalidate_wr, &bad_wr);
		rc = ib_post_send(ia->ri_id->qp, invalidate_wr, &bad_wr);
		read_unlock(&ia->ri_qplock);
		if (rc)
		goto out_err;
		@@ -471,6 +600,7 @@ frwr_op_destroy(struct rpcrdma_buffer *buf)

		const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops = {
		.ro_map = frwr_op_map,
		.ro_unmap_sync = frwr_op_unmap_sync,
		.ro_unmap = frwr_op_unmap,
		.ro_open = frwr_op_open,
		.ro_maxpages = frwr_op_maxpages,

net/sunrpc/xprtrdma/physical_ops.c

+13 −0

Original line number	Diff line number	Diff line
		@@ -83,6 +83,18 @@ physical_op_unmap(struct rpcrdma_xprt r_xprt, struct rpcrdma_mr_seg seg)
		return 1;
		}

		/* DMA unmap all memory regions that were mapped for "req".
		*/
		static void
		physical_op_unmap_sync(struct rpcrdma_xprt r_xprt, struct rpcrdma_req req)
		{
		struct ib_device *device = r_xprt->rx_ia.ri_device;
		unsigned int i;

		for (i = 0; req->rl_nchunks; --req->rl_nchunks)
		rpcrdma_unmap_one(device, &req->rl_segments[i++]);
		}

		static void
		physical_op_destroy(struct rpcrdma_buffer *buf)
		{
		@@ -90,6 +102,7 @@ physical_op_destroy(struct rpcrdma_buffer *buf)

		const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops = {
		.ro_map = physical_op_map,
		.ro_unmap_sync = physical_op_unmap_sync,
		.ro_unmap = physical_op_unmap,
		.ro_open = physical_op_open,
		.ro_maxpages = physical_op_maxpages,

net/sunrpc/xprtrdma/rpc_rdma.c

+16 −0

Original line number	Diff line number	Diff line
		@@ -804,6 +804,11 @@ rpcrdma_reply_handler(struct rpcrdma_rep *rep)
		if (req->rl_reply)
		goto out_duplicate;

		/* Sanity checking has passed. We are now committed
		* to complete this transaction.
		*/
		list_del_init(&rqst->rq_list);
		spin_unlock_bh(&xprt->transport_lock);
		dprintk("RPC: %s: reply 0x%p completes request 0x%p\n"
		" RPC request 0x%p xid 0x%08x\n",
		__func__, rep, req, rqst,
		@@ -888,12 +893,23 @@ badheader:
		break;
		}

		/* Invalidate and flush the data payloads before waking the
		* waiting application. This guarantees the memory region is
		* properly fenced from the server before the application
		* accesses the data. It also ensures proper send flow
		* control: waking the next RPC waits until this RPC has
		* relinquished all its Send Queue entries.
		*/
		if (req->rl_nchunks)
		r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt, req);

		credits = be32_to_cpu(headerp->rm_credit);
		if (credits == 0)
		credits = 1; /* don't deadlock */
		else if (credits > r_xprt->rx_buf.rb_max_requests)
		credits = r_xprt->rx_buf.rb_max_requests;

		spin_lock_bh(&xprt->transport_lock);
		cwnd = xprt->cwnd;
		xprt->cwnd = credits << RPC_CWNDSHIFT;
		if (xprt->cwnd > cwnd)

Admin message