> On Aug 31, 2021, at 3:05 PM, Chuck Lever <chuck.lever@xxxxxxxxxx> wrote:
>
> This enables the XDR decoder to figure out how the payload sink
> buffer needs to be aligned before setting up the RDMA Reads.
> Then re-alignment of large RDMA Read payloads can be avoided.
Should say "can be avoided" but isn't yet. All of these patches
are pre-requisites to that behavior.
> Signed-off-by: Chuck Lever <chuck.lever@xxxxxxxxxx>
> ---
> include/linux/sunrpc/svc_rdma.h | 6 +
> include/trace/events/rpcrdma.h | 26 ++++++
> net/sunrpc/xprtrdma/svc_rdma_recvfrom.c | 17 +++-
> net/sunrpc/xprtrdma/svc_rdma_rw.c | 139 ++++++++++++++++++++++++++++---
> 4 files changed, 169 insertions(+), 19 deletions(-)
>
> diff --git a/include/linux/sunrpc/svc_rdma.h b/include/linux/sunrpc/svc_rdma.h
> index f660244cc8ba..8d80a759a909 100644
> --- a/include/linux/sunrpc/svc_rdma.h
> +++ b/include/linux/sunrpc/svc_rdma.h
> @@ -192,6 +192,12 @@ extern int svc_rdma_send_reply_chunk(struct svcxprt_rdma *rdma,
> extern int svc_rdma_process_read_list(struct svcxprt_rdma *rdma,
> struct svc_rqst *rqstp,
> struct svc_rdma_recv_ctxt *head);
> +extern void svc_rdma_prepare_read_chunk(struct svc_rqst *rqstp,
> + struct svc_rdma_recv_ctxt *head);
> +extern int svc_rdma_pull_read_chunk(struct svcxprt_rdma *rdma,
> + struct svc_rqst *rqstp,
> + struct svc_rdma_recv_ctxt *ctxt,
> + unsigned int offset, unsigned int length);
>
> /* svc_rdma_sendto.c */
> extern void svc_rdma_send_ctxts_destroy(struct svcxprt_rdma *rdma);
> diff --git a/include/trace/events/rpcrdma.h b/include/trace/events/rpcrdma.h
> index 5954ce036173..30440cca321a 100644
> --- a/include/trace/events/rpcrdma.h
> +++ b/include/trace/events/rpcrdma.h
> @@ -2136,6 +2136,32 @@ TRACE_EVENT(svcrdma_sq_post_err,
> )
> );
>
> +TRACE_EVENT(svcrdma_arg_payload,
> + TP_PROTO(
> + const struct svc_rqst *rqstp,
> + unsigned int offset,
> + unsigned int length
> + ),
> +
> + TP_ARGS(rqstp, offset, length),
> +
> + TP_STRUCT__entry(
> + __field(u32, xid)
> + __field(u32, offset)
> + __field(u32, length)
> + ),
> +
> + TP_fast_assign(
> + __entry->xid = __be32_to_cpu(rqstp->rq_xid);
> + __entry->offset = offset_in_page(offset);
> + __entry->length = length;
> + ),
> +
> + TP_printk("xid=0x%08x offset=%u length=%u",
> + __entry->xid, __entry->offset, __entry->length
> + )
> +);
> +
> #endif /* _TRACE_RPCRDMA_H */
>
> #include <trace/define_trace.h>
> diff --git a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
> index 08a620b370ae..cd9c0fb1a470 100644
> --- a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
> +++ b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
> @@ -838,12 +838,13 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
>
> svc_rdma_get_inv_rkey(rdma_xprt, ctxt);
>
> - if (!pcl_is_empty(&ctxt->rc_read_pcl) ||
> - !pcl_is_empty(&ctxt->rc_call_pcl)) {
> + if (!pcl_is_empty(&ctxt->rc_call_pcl) ||
> + ctxt->rc_read_pcl.cl_count > 1) {
> ret = svc_rdma_process_read_list(rdma_xprt, rqstp, ctxt);
> if (ret < 0)
> goto out_readfail;
> - }
> + } else if (ctxt->rc_read_pcl.cl_count == 1)
> + svc_rdma_prepare_read_chunk(rqstp, ctxt);
>
> rqstp->rq_xprt_ctxt = ctxt;
> rqstp->rq_prot = IPPROTO_MAX;
> @@ -887,5 +888,13 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
> int svc_rdma_argument_payload(struct svc_rqst *rqstp, unsigned int offset,
> unsigned int length)
> {
> - return 0;
> + struct svc_rdma_recv_ctxt *ctxt = rqstp->rq_xprt_ctxt;
> + struct svc_xprt *xprt = rqstp->rq_xprt;
> + struct svcxprt_rdma *rdma =
> + container_of(xprt, struct svcxprt_rdma, sc_xprt);
> +
> + if (!pcl_is_empty(&ctxt->rc_call_pcl) ||
> + ctxt->rc_read_pcl.cl_count != 1)
> + return 0;
> + return svc_rdma_pull_read_chunk(rdma, rqstp, ctxt, offset, length);
> }
> diff --git a/net/sunrpc/xprtrdma/svc_rdma_rw.c b/net/sunrpc/xprtrdma/svc_rdma_rw.c
> index 29b7d477891c..5f03dfd2fa03 100644
> --- a/net/sunrpc/xprtrdma/svc_rdma_rw.c
> +++ b/net/sunrpc/xprtrdma/svc_rdma_rw.c
> @@ -707,19 +707,24 @@ static int svc_rdma_build_read_segment(struct svc_rdma_read_info *info,
>
> len = segment->rs_length;
> sge_no = PAGE_ALIGN(info->ri_pageoff + len) >> PAGE_SHIFT;
> +
> + trace_printk("pageoff=%u len=%u sges=%u\n",
> + info->ri_pageoff, len, sge_no);
> +
> ctxt = svc_rdma_get_rw_ctxt(cc->cc_rdma, sge_no);
> if (!ctxt)
> return -ENOMEM;
> ctxt->rw_nents = sge_no;
> + head->rc_page_count += sge_no;
>
> sg = ctxt->rw_sg_table.sgl;
> for (sge_no = 0; sge_no < ctxt->rw_nents; sge_no++) {
> seg_len = min_t(unsigned int, len,
> PAGE_SIZE - info->ri_pageoff);
>
> - if (!info->ri_pageoff)
> - head->rc_page_count++;
> -
> + trace_printk(" page=%p seg_len=%u offset=%u\n",
> + rqstp->rq_pages[info->ri_pageno], seg_len,
> + info->ri_pageoff);
> sg_set_page(sg, rqstp->rq_pages[info->ri_pageno],
> seg_len, info->ri_pageoff);
> sg = sg_next(sg);
> @@ -804,15 +809,14 @@ static int svc_rdma_copy_inline_range(struct svc_rdma_read_info *info,
> unsigned int page_no, numpages;
>
> numpages = PAGE_ALIGN(info->ri_pageoff + remaining) >> PAGE_SHIFT;
> + head->rc_page_count += numpages;
> +
> for (page_no = 0; page_no < numpages; page_no++) {
> unsigned int page_len;
>
> page_len = min_t(unsigned int, remaining,
> PAGE_SIZE - info->ri_pageoff);
>
> - if (!info->ri_pageoff)
> - head->rc_page_count++;
> -
> dst = page_address(rqstp->rq_pages[info->ri_pageno]);
> memcpy(dst + info->ri_pageno, src + offset, page_len);
>
> @@ -1092,15 +1096,8 @@ static noinline int svc_rdma_read_special(struct svc_rdma_read_info *info)
> * @rqstp: set of pages to use as Read sink buffers
> * @head: pages under I/O collect here
> *
> - * The RPC/RDMA protocol assumes that the upper layer's XDR decoders
> - * pull each Read chunk as they decode an incoming RPC message.
> - *
> - * On Linux, however, the server needs to have a fully-constructed RPC
> - * message in rqstp->rq_arg when there is a positive return code from
> - * ->xpo_recvfrom. So the Read list is safety-checked immediately when
> - * it is received, then here the whole Read list is pulled all at once.
> - * The ingress RPC message is fully reconstructed once all associated
> - * RDMA Reads have completed.
> + * Handle complex Read chunk cases fully before svc_rdma_recvfrom()
> + * returns.
> *
> * Return values:
> * %1: all needed RDMA Reads were posted successfully,
> @@ -1159,3 +1156,115 @@ int svc_rdma_process_read_list(struct svcxprt_rdma *rdma,
> svc_rdma_read_info_free(info);
> return ret;
> }
> +
> +/**
> + * svc_rdma_prepare_read_chunk - Prepare rq_arg for Read chunk
> + * @rqstp: set of pages to use as Read sink buffers
> + * @head: pages under I/O collect here
> + *
> + * The Read chunk will be pulled when the upper layer's XDR
> + * decoder calls svc_decode_argument_payload(). In the meantime,
> + * fake up rq_arg.page_len and .len to reflect the size of the
> + * yet-to-be-pulled payload.
> + */
> +void svc_rdma_prepare_read_chunk(struct svc_rqst *rqstp,
> + struct svc_rdma_recv_ctxt *head)
> +{
> + struct svc_rdma_chunk *chunk = pcl_first_chunk(&head->rc_read_pcl);
> + unsigned int length = xdr_align_size(chunk->ch_length);
> + struct xdr_buf *buf = &rqstp->rq_arg;
> +
> + buf->tail[0].iov_base = buf->head[0].iov_base + chunk->ch_position;
> + buf->tail[0].iov_len = buf->head[0].iov_len - chunk->ch_position;
> + buf->head[0].iov_len = chunk->ch_position;
> +
> + buf->page_len = length;
> + buf->len += length;
> + buf->buflen += length;
> +
> + /*
> + * rq_respages starts after the last arg page. Note that we
> + * don't know the offset yet, so add an extra page as slack.
> + */
> + length += PAGE_SIZE * 2 - 1;
> + rqstp->rq_respages = &rqstp->rq_pages[length >> PAGE_SHIFT];
> + rqstp->rq_next_page = rqstp->rq_respages + 1;
> +}
> +
> +/**
> + * svc_rdma_pull_read_chunk - Pull one Read chunk from the client
> + * @rdma: controlling RDMA transport
> + * @rqstp: set of pages to use as Read sink buffers
> + * @head: pages under I/O collect here
> + * @offset: offset of payload in file's page cache
> + * @length: size of payload, in bytes
> + *
> + * Once the upper layer's XDR decoder has decoded the length of
> + * the payload and it's offset, we can be clever about setting up
> + * the RDMA Read sink buffer so that the VFS does not have to
> + * re-align the payload once it is received.
> + *
> + * Caveat: To keep things simple, this is an optimization that is
> + * used only when there is a single Read chunk in the Read
> + * list.
> + *
> + * Return values:
> + * %0: all needed RDMA Reads were posted successfully,
> + * %-EINVAL: client provided the wrong chunk size,
> + * %-ENOMEM: rdma_rw context pool was exhausted,
> + * %-ENOTCONN: posting failed (connection is lost),
> + * %-EIO: rdma_rw initialization failed (DMA mapping, etc).
> + */
> +int svc_rdma_pull_read_chunk(struct svcxprt_rdma *rdma, struct svc_rqst *rqstp,
> + struct svc_rdma_recv_ctxt *head,
> + unsigned int offset, unsigned int length)
> +{
> + struct svc_rdma_read_info *info;
> + struct svc_rdma_chunk_ctxt *cc;
> + struct svc_rdma_chunk *chunk;
> + int ret;
> +
> + trace_svcrdma_arg_payload(rqstp, offset, length);
> +
> + /* Sanity: the Requester must have provided enough
> + * bytes to fill the XDR opaque.
> + */
> + chunk = pcl_first_chunk(&head->rc_read_pcl);
> + if (length > chunk->ch_length)
> + return -EINVAL;
> +
> + info = svc_rdma_read_info_alloc(rdma);
> + if (!info)
> + return -ENOMEM;
> + cc = &info->ri_cc;
> + info->ri_rqst = rqstp;
> + info->ri_readctxt = head;
> + info->ri_pageno = 0;
> + info->ri_pageoff = offset_in_page(offset);
> + info->ri_totalbytes = 0;
> +
> + ret = svc_rdma_build_read_chunk(info, chunk);
> + if (ret < 0)
> + goto out_err;
> + rqstp->rq_arg.pages = &info->ri_rqst->rq_pages[0];
> + rqstp->rq_arg.page_base = offset_in_page(offset);
> + rqstp->rq_arg.buflen += offset_in_page(offset);
> +
> + trace_svcrdma_post_read_chunk(&cc->cc_cid, cc->cc_sqecount);
> + init_completion(&cc->cc_done);
> + ret = svc_rdma_post_chunk_ctxt(cc);
> + if (ret < 0)
> + goto out_err;
> +
> + ret = 0;
> + wait_for_completion(&cc->cc_done);
> + if (cc->cc_status != IB_WC_SUCCESS)
> + ret = -EIO;
> +
> + /* Ensure svc_rdma_recv_ctxt_put() does not try to release pages */
> + head->rc_page_count = 0;
> +
> +out_err:
> + svc_rdma_read_info_free(info);
> + return ret;
> +}
>
>
--
Chuck Lever
