Re: [PATCH v1 10/19] svcrdma: Persistently allocate and DMA-map Receive buffers

[Bruce Fields <bfields@fieldses.org>]

On Wed, May 09, 2018 at 05:31:17PM -0400, Chuck Lever wrote:
> 
> 
> > On May 9, 2018, at 5:18 PM, J. Bruce Fields <bfields@fieldses.org> wrote:
> > 
> > On Mon, May 07, 2018 at 03:27:43PM -0400, Chuck Lever wrote:
> >> The current Receive path uses an array of pages which are allocated
> >> and DMA mapped when each Receive WR is posted, and then handed off
> >> to the upper layer in rqstp::rq_arg. The page flip releases unused
> >> pages in the rq_pages pagelist. This mechanism introduces a
> >> significant amount of overhead.
> >> 
> >> So instead, kmalloc the Receive buffer, and leave it DMA-mapped
> >> while the transport remains connected. This confers a number of
> >> benefits:
> >> 
> >> * Each Receive WR requires only one receive SGE, no matter how large
> >>  the inline threshold is. This helps the server-side NFS/RDMA
> >>  transport operate on less capable RDMA devices.
> >> 
> >> * The Receive buffer is left allocated and mapped all the time. This
> >>  relieves svc_rdma_post_recv from the overhead of allocating and
> >>  DMA-mapping a fresh buffer.
> > 
> > Dumb question: does that mean the buffer could still change if the
> > client does something weird?  (So could the xdr decoding code see data
> > change out from under it?)
> 
> No, once the Receive completes, the HCA no longer has access
> to the buffer. Leaving it DMA-mapped is not the same as leaving
> it posted or leaving it registered; it just means that the I/O
> subsystem has the buffer prepared for more I/O.
> 
> The current scheme works like this:
> 
> repeat:
> Allocate a page
> DMA-map the page
> Post a Recv WR for that page
> Recv WR completes
> DMA-unmap the page
> XDR decoding and stuff
> Free the page
> goto repeat
> 
> And I want it to go faster:
> 
> kmalloc a buffer
> DMA-map the buffer
> repeat:
> Post a Recv for that buffer
> Recv WR completes
> DMA-sync the buffer
> XDR decoding and stuff
> goto repeat
> 
> On disconnect, the Recv flushes. The transport tear-down logic
> takes care of DMA-unmapping and freeing the buffer.

Got it, thanks!


--b.

> 
> 
> > --b.
> > 
> >> 
> >> * svc_rdma_wc_receive no longer has to DMA unmap the Receive buffer.
> >>  It has to DMA sync only the number of bytes that were received.
> >> 
> >> * svc_rdma_build_arg_xdr no longer has to free a page in rq_pages
> >>  for each page in the Receive buffer, making it a constant-time
> >>  function.
> >> 
> >> * The Receive buffer is now plugged directly into the rq_arg's
> >>  head[0].iov_vec, and can be larger than a page without spilling
> >>  over into rq_arg's page list. This enables simplification of
> >>  the RDMA Read path in subsequent patches.
> >> 
> >> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
> >> ---
> >> include/linux/sunrpc/svc_rdma.h          |    4 -
> >> net/sunrpc/xprtrdma/svc_rdma_recvfrom.c  |  168 +++++++++++-------------------
> >> net/sunrpc/xprtrdma/svc_rdma_rw.c        |   32 ++----
> >> net/sunrpc/xprtrdma/svc_rdma_sendto.c    |    5 -
> >> net/sunrpc/xprtrdma/svc_rdma_transport.c |    2 
> >> 5 files changed, 75 insertions(+), 136 deletions(-)
> >> 
> >> diff --git a/include/linux/sunrpc/svc_rdma.h b/include/linux/sunrpc/svc_rdma.h
> >> index f0bd0b6d..01baabf 100644
> >> --- a/include/linux/sunrpc/svc_rdma.h
> >> +++ b/include/linux/sunrpc/svc_rdma.h
> >> @@ -148,12 +148,12 @@ struct svc_rdma_recv_ctxt {
> >> 	struct list_head	rc_list;
> >> 	struct ib_recv_wr	rc_recv_wr;
> >> 	struct ib_cqe		rc_cqe;
> >> +	struct ib_sge		rc_recv_sge;
> >> +	void			*rc_recv_buf;
> >> 	struct xdr_buf		rc_arg;
> >> 	u32			rc_byte_len;
> >> 	unsigned int		rc_page_count;
> >> 	unsigned int		rc_hdr_count;
> >> -	struct ib_sge		rc_sges[1 +
> >> -					RPCRDMA_MAX_INLINE_THRESH / PAGE_SIZE];
> >> 	struct page		*rc_pages[RPCSVC_MAXPAGES];
> >> };
> >> 
> >> diff --git a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
> >> index d9fef52..d4ccd1c 100644
> >> --- a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
> >> +++ b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
> >> @@ -117,6 +117,43 @@
> >> 					rc_list);
> >> }
> >> 
> >> +static struct svc_rdma_recv_ctxt *
> >> +svc_rdma_recv_ctxt_alloc(struct svcxprt_rdma *rdma)
> >> +{
> >> +	struct svc_rdma_recv_ctxt *ctxt;
> >> +	dma_addr_t addr;
> >> +	void *buffer;
> >> +
> >> +	ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
> >> +	if (!ctxt)
> >> +		goto fail0;
> >> +	buffer = kmalloc(rdma->sc_max_req_size, GFP_KERNEL);
> >> +	if (!buffer)
> >> +		goto fail1;
> >> +	addr = ib_dma_map_single(rdma->sc_pd->device, buffer,
> >> +				 rdma->sc_max_req_size, DMA_FROM_DEVICE);
> >> +	if (ib_dma_mapping_error(rdma->sc_pd->device, addr))
> >> +		goto fail2;
> >> +
> >> +	ctxt->rc_recv_wr.next = NULL;
> >> +	ctxt->rc_recv_wr.wr_cqe = &ctxt->rc_cqe;
> >> +	ctxt->rc_recv_wr.sg_list = &ctxt->rc_recv_sge;
> >> +	ctxt->rc_recv_wr.num_sge = 1;
> >> +	ctxt->rc_cqe.done = svc_rdma_wc_receive;
> >> +	ctxt->rc_recv_sge.addr = addr;
> >> +	ctxt->rc_recv_sge.length = rdma->sc_max_req_size;
> >> +	ctxt->rc_recv_sge.lkey = rdma->sc_pd->local_dma_lkey;
> >> +	ctxt->rc_recv_buf = buffer;
> >> +	return ctxt;
> >> +
> >> +fail2:
> >> +	kfree(buffer);
> >> +fail1:
> >> +	kfree(ctxt);
> >> +fail0:
> >> +	return NULL;
> >> +}
> >> +
> >> /**
> >>  * svc_rdma_recv_ctxts_destroy - Release all recv_ctxt's for an xprt
> >>  * @rdma: svcxprt_rdma being torn down
> >> @@ -128,6 +165,11 @@ void svc_rdma_recv_ctxts_destroy(struct svcxprt_rdma *rdma)
> >> 
> >> 	while ((ctxt = svc_rdma_next_recv_ctxt(&rdma->sc_recv_ctxts))) {
> >> 		list_del(&ctxt->rc_list);
> >> +		ib_dma_unmap_single(rdma->sc_pd->device,
> >> +				    ctxt->rc_recv_sge.addr,
> >> +				    ctxt->rc_recv_sge.length,
> >> +				    DMA_FROM_DEVICE);
> >> +		kfree(ctxt->rc_recv_buf);
> >> 		kfree(ctxt);
> >> 	}
> >> }
> >> @@ -145,32 +187,18 @@ void svc_rdma_recv_ctxts_destroy(struct svcxprt_rdma *rdma)
> >> 	spin_unlock(&rdma->sc_recv_lock);
> >> 
> >> out:
> >> -	ctxt->rc_recv_wr.num_sge = 0;
> >> 	ctxt->rc_page_count = 0;
> >> 	return ctxt;
> >> 
> >> out_empty:
> >> 	spin_unlock(&rdma->sc_recv_lock);
> >> 
> >> -	ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
> >> +	ctxt = svc_rdma_recv_ctxt_alloc(rdma);
> >> 	if (!ctxt)
> >> 		return NULL;
> >> 	goto out;
> >> }
> >> 
> >> -static void svc_rdma_recv_ctxt_unmap(struct svcxprt_rdma *rdma,
> >> -				     struct svc_rdma_recv_ctxt *ctxt)
> >> -{
> >> -	struct ib_device *device = rdma->sc_cm_id->device;
> >> -	int i;
> >> -
> >> -	for (i = 0; i < ctxt->rc_recv_wr.num_sge; i++)
> >> -		ib_dma_unmap_page(device,
> >> -				  ctxt->rc_sges[i].addr,
> >> -				  ctxt->rc_sges[i].length,
> >> -				  DMA_FROM_DEVICE);
> >> -}
> >> -
> >> /**
> >>  * svc_rdma_recv_ctxt_put - Return recv_ctxt to free list
> >>  * @rdma: controlling svcxprt_rdma
> >> @@ -191,46 +219,14 @@ void svc_rdma_recv_ctxt_put(struct svcxprt_rdma *rdma,
> >> 
> >> static int svc_rdma_post_recv(struct svcxprt_rdma *rdma)
> >> {
> >> -	struct ib_device *device = rdma->sc_cm_id->device;
> >> 	struct svc_rdma_recv_ctxt *ctxt;
> >> 	struct ib_recv_wr *bad_recv_wr;
> >> -	int sge_no, buflen, ret;
> >> -	struct page *page;
> >> -	dma_addr_t pa;
> >> +	int ret;
> >> 
> >> 	ctxt = svc_rdma_recv_ctxt_get(rdma);
> >> 	if (!ctxt)
> >> 		return -ENOMEM;
> >> 
> >> -	buflen = 0;
> >> -	ctxt->rc_cqe.done = svc_rdma_wc_receive;
> >> -	for (sge_no = 0; buflen < rdma->sc_max_req_size; sge_no++) {
> >> -		if (sge_no >= rdma->sc_max_sge) {
> >> -			pr_err("svcrdma: Too many sges (%d)\n", sge_no);
> >> -			goto err_put_ctxt;
> >> -		}
> >> -
> >> -		page = alloc_page(GFP_KERNEL);
> >> -		if (!page)
> >> -			goto err_put_ctxt;
> >> -		ctxt->rc_pages[sge_no] = page;
> >> -		ctxt->rc_page_count++;
> >> -
> >> -		pa = ib_dma_map_page(device, ctxt->rc_pages[sge_no],
> >> -				     0, PAGE_SIZE, DMA_FROM_DEVICE);
> >> -		if (ib_dma_mapping_error(device, pa))
> >> -			goto err_put_ctxt;
> >> -		ctxt->rc_sges[sge_no].addr = pa;
> >> -		ctxt->rc_sges[sge_no].length = PAGE_SIZE;
> >> -		ctxt->rc_sges[sge_no].lkey = rdma->sc_pd->local_dma_lkey;
> >> -		ctxt->rc_recv_wr.num_sge++;
> >> -
> >> -		buflen += PAGE_SIZE;
> >> -	}
> >> -	ctxt->rc_recv_wr.next = NULL;
> >> -	ctxt->rc_recv_wr.sg_list = &ctxt->rc_sges[0];
> >> -	ctxt->rc_recv_wr.wr_cqe = &ctxt->rc_cqe;
> >> -
> >> 	svc_xprt_get(&rdma->sc_xprt);
> >> 	ret = ib_post_recv(rdma->sc_qp, &ctxt->rc_recv_wr, &bad_recv_wr);
> >> 	trace_svcrdma_post_recv(&ctxt->rc_recv_wr, ret);
> >> @@ -238,12 +234,7 @@ static int svc_rdma_post_recv(struct svcxprt_rdma *rdma)
> >> 		goto err_post;
> >> 	return 0;
> >> 
> >> -err_put_ctxt:
> >> -	svc_rdma_recv_ctxt_unmap(rdma, ctxt);
> >> -	svc_rdma_recv_ctxt_put(rdma, ctxt);
> >> -	return -ENOMEM;
> >> err_post:
> >> -	svc_rdma_recv_ctxt_unmap(rdma, ctxt);
> >> 	svc_rdma_recv_ctxt_put(rdma, ctxt);
> >> 	svc_xprt_put(&rdma->sc_xprt);
> >> 	return ret;
> >> @@ -289,7 +280,6 @@ static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
> >> 
> >> 	/* WARNING: Only wc->wr_cqe and wc->status are reliable */
> >> 	ctxt = container_of(cqe, struct svc_rdma_recv_ctxt, rc_cqe);
> >> -	svc_rdma_recv_ctxt_unmap(rdma, ctxt);
> >> 
> >> 	if (wc->status != IB_WC_SUCCESS)
> >> 		goto flushed;
> >> @@ -299,6 +289,10 @@ static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
> >> 
> >> 	/* All wc fields are now known to be valid */
> >> 	ctxt->rc_byte_len = wc->byte_len;
> >> +	ib_dma_sync_single_for_cpu(rdma->sc_pd->device,
> >> +				   ctxt->rc_recv_sge.addr,
> >> +				   wc->byte_len, DMA_FROM_DEVICE);
> >> +
> >> 	spin_lock(&rdma->sc_rq_dto_lock);
> >> 	list_add_tail(&ctxt->rc_list, &rdma->sc_rq_dto_q);
> >> 	spin_unlock(&rdma->sc_rq_dto_lock);
> >> @@ -339,64 +333,22 @@ void svc_rdma_flush_recv_queues(struct svcxprt_rdma *rdma)
> >> 	}
> >> }
> >> 
> >> -/*
> >> - * Replace the pages in the rq_argpages array with the pages from the SGE in
> >> - * the RDMA_RECV completion. The SGL should contain full pages up until the
> >> - * last one.
> >> - */
> >> static void svc_rdma_build_arg_xdr(struct svc_rqst *rqstp,
> >> 				   struct svc_rdma_recv_ctxt *ctxt)
> >> {
> >> -	struct page *page;
> >> -	int sge_no;
> >> -	u32 len;
> >> -
> >> -	/* The reply path assumes the Call's transport header resides
> >> -	 * in rqstp->rq_pages[0].
> >> -	 */
> >> -	page = ctxt->rc_pages[0];
> >> -	put_page(rqstp->rq_pages[0]);
> >> -	rqstp->rq_pages[0] = page;
> >> -
> >> -	/* Set up the XDR head */
> >> -	rqstp->rq_arg.head[0].iov_base = page_address(page);
> >> -	rqstp->rq_arg.head[0].iov_len =
> >> -		min_t(size_t, ctxt->rc_byte_len, ctxt->rc_sges[0].length);
> >> -	rqstp->rq_arg.len = ctxt->rc_byte_len;
> >> -	rqstp->rq_arg.buflen = ctxt->rc_byte_len;
> >> -
> >> -	/* Compute bytes past head in the SGL */
> >> -	len = ctxt->rc_byte_len - rqstp->rq_arg.head[0].iov_len;
> >> -
> >> -	/* If data remains, store it in the pagelist */
> >> -	rqstp->rq_arg.page_len = len;
> >> -	rqstp->rq_arg.page_base = 0;
> >> -
> >> -	sge_no = 1;
> >> -	while (len && sge_no < ctxt->rc_recv_wr.num_sge) {
> >> -		page = ctxt->rc_pages[sge_no];
> >> -		put_page(rqstp->rq_pages[sge_no]);
> >> -		rqstp->rq_pages[sge_no] = page;
> >> -		len -= min_t(u32, len, ctxt->rc_sges[sge_no].length);
> >> -		sge_no++;
> >> -	}
> >> -	ctxt->rc_hdr_count = sge_no;
> >> -	rqstp->rq_respages = &rqstp->rq_pages[sge_no];
> >> +	struct xdr_buf *arg = &rqstp->rq_arg;
> >> +
> >> +	arg->head[0].iov_base = ctxt->rc_recv_buf;
> >> +	arg->head[0].iov_len = ctxt->rc_byte_len;
> >> +	arg->tail[0].iov_base = NULL;
> >> +	arg->tail[0].iov_len = 0;
> >> +	arg->page_len = 0;
> >> +	arg->page_base = 0;
> >> +	arg->buflen = ctxt->rc_byte_len;
> >> +	arg->len = ctxt->rc_byte_len;
> >> +
> >> +	rqstp->rq_respages = &rqstp->rq_pages[0];
> >> 	rqstp->rq_next_page = rqstp->rq_respages + 1;
> >> -
> >> -	/* If not all pages were used from the SGL, free the remaining ones */
> >> -	while (sge_no < ctxt->rc_recv_wr.num_sge) {
> >> -		page = ctxt->rc_pages[sge_no++];
> >> -		put_page(page);
> >> -	}
> >> -
> >> -	/* @ctxt's pages have all been released or moved to @rqstp->rq_pages.
> >> -	 */
> >> -	ctxt->rc_page_count = 0;
> >> -
> >> -	/* Set up tail */
> >> -	rqstp->rq_arg.tail[0].iov_base = NULL;
> >> -	rqstp->rq_arg.tail[0].iov_len = 0;
> >> }
> >> 
> >> /* This accommodates the largest possible Write chunk,
> >> diff --git a/net/sunrpc/xprtrdma/svc_rdma_rw.c b/net/sunrpc/xprtrdma/svc_rdma_rw.c
> >> index 8242aa3..ce3ea84 100644
> >> --- a/net/sunrpc/xprtrdma/svc_rdma_rw.c
> >> +++ b/net/sunrpc/xprtrdma/svc_rdma_rw.c
> >> @@ -718,15 +718,14 @@ static int svc_rdma_build_normal_read_chunk(struct svc_rqst *rqstp,
> >> 	struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
> >> 	int ret;
> >> 
> >> -	info->ri_pageno = head->rc_hdr_count;
> >> -	info->ri_pageoff = 0;
> >> -
> >> 	ret = svc_rdma_build_read_chunk(rqstp, info, p);
> >> 	if (ret < 0)
> >> 		goto out;
> >> 
> >> 	trace_svcrdma_encode_read(info->ri_chunklen, info->ri_position);
> >> 
> >> +	head->rc_hdr_count = 0;
> >> +
> >> 	/* Split the Receive buffer between the head and tail
> >> 	 * buffers at Read chunk's position. XDR roundup of the
> >> 	 * chunk is not included in either the pagelist or in
> >> @@ -775,9 +774,6 @@ static int svc_rdma_build_pz_read_chunk(struct svc_rqst *rqstp,
> >> 	struct svc_rdma_recv_ctxt *head = info->ri_readctxt;
> >> 	int ret;
> >> 
> >> -	info->ri_pageno = head->rc_hdr_count - 1;
> >> -	info->ri_pageoff = offset_in_page(head->rc_byte_len);
> >> -
> >> 	ret = svc_rdma_build_read_chunk(rqstp, info, p);
> >> 	if (ret < 0)
> >> 		goto out;
> >> @@ -787,20 +783,13 @@ static int svc_rdma_build_pz_read_chunk(struct svc_rqst *rqstp,
> >> 	head->rc_arg.len += info->ri_chunklen;
> >> 	head->rc_arg.buflen += info->ri_chunklen;
> >> 
> >> -	if (head->rc_arg.buflen <= head->rc_sges[0].length) {
> >> -		/* Transport header and RPC message fit entirely
> >> -		 * in page where head iovec resides.
> >> -		 */
> >> -		head->rc_arg.head[0].iov_len = info->ri_chunklen;
> >> -	} else {
> >> -		/* Transport header and part of RPC message reside
> >> -		 * in the head iovec's page.
> >> -		 */
> >> -		head->rc_arg.head[0].iov_len =
> >> -			head->rc_sges[0].length - head->rc_byte_len;
> >> -		head->rc_arg.page_len =
> >> -			info->ri_chunklen - head->rc_arg.head[0].iov_len;
> >> -	}
> >> +	head->rc_hdr_count = 1;
> >> +	head->rc_arg.head[0].iov_base = page_address(head->rc_pages[0]);
> >> +	head->rc_arg.head[0].iov_len = min_t(size_t, PAGE_SIZE,
> >> +					     info->ri_chunklen);
> >> +
> >> +	head->rc_arg.page_len = info->ri_chunklen -
> >> +				head->rc_arg.head[0].iov_len;
> >> 
> >> out:
> >> 	return ret;
> >> @@ -834,7 +823,6 @@ int svc_rdma_recv_read_chunk(struct svcxprt_rdma *rdma, struct svc_rqst *rqstp,
> >> 	 * head->rc_arg. Pages involved with RDMA Read I/O are
> >> 	 * transferred there.
> >> 	 */
> >> -	head->rc_page_count = head->rc_hdr_count;
> >> 	head->rc_arg.head[0] = rqstp->rq_arg.head[0];
> >> 	head->rc_arg.tail[0] = rqstp->rq_arg.tail[0];
> >> 	head->rc_arg.pages = head->rc_pages;
> >> @@ -847,6 +835,8 @@ int svc_rdma_recv_read_chunk(struct svcxprt_rdma *rdma, struct svc_rqst *rqstp,
> >> 	if (!info)
> >> 		return -ENOMEM;
> >> 	info->ri_readctxt = head;
> >> +	info->ri_pageno = 0;
> >> +	info->ri_pageoff = 0;
> >> 
> >> 	info->ri_position = be32_to_cpup(p + 1);
> >> 	if (info->ri_position)
> >> diff --git a/net/sunrpc/xprtrdma/svc_rdma_sendto.c b/net/sunrpc/xprtrdma/svc_rdma_sendto.c
> >> index cbbde70..b27b597 100644
> >> --- a/net/sunrpc/xprtrdma/svc_rdma_sendto.c
> >> +++ b/net/sunrpc/xprtrdma/svc_rdma_sendto.c
> >> @@ -629,10 +629,7 @@ int svc_rdma_sendto(struct svc_rqst *rqstp)
> >> 	struct page *res_page;
> >> 	int ret;
> >> 
> >> -	/* Find the call's chunk lists to decide how to send the reply.
> >> -	 * Receive places the Call's xprt header at the start of page 0.
> >> -	 */
> >> -	rdma_argp = page_address(rqstp->rq_pages[0]);
> >> +	rdma_argp = rctxt->rc_recv_buf;
> >> 	svc_rdma_get_write_arrays(rdma_argp, &wr_lst, &rp_ch);
> >> 
> >> 	/* Create the RDMA response header. xprt->xpt_mutex,
> >> diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c
> >> index 20abd3a..333c432 100644
> >> --- a/net/sunrpc/xprtrdma/svc_rdma_transport.c
> >> +++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c
> >> @@ -670,7 +670,7 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
> >> 	qp_attr.cap.max_send_wr = newxprt->sc_sq_depth - ctxts;
> >> 	qp_attr.cap.max_recv_wr = rq_depth;
> >> 	qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
> >> -	qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
> >> +	qp_attr.cap.max_recv_sge = 1;
> >> 	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
> >> 	qp_attr.qp_type = IB_QPT_RC;
> >> 	qp_attr.send_cq = newxprt->sc_sq_cq;
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> > the body of a message to majordomo@vger.kernel.org
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 
> --
> Chuck Lever
> 
>