return n;
}
-/*
- * Create read/write chunk lists, and reply chunks, for RDMA
- *
- * Assume check against THRESHOLD has been done, and chunks are required.
- * Assume only encoding one list entry for read|write chunks. The NFSv3
- * protocol is simple enough to allow this as it only has a single "bulk
- * result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The
- * RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.)
- *
- * When used for a single reply chunk (which is a special write
- * chunk used for the entire reply, rather than just the data), it
- * is used primarily for READDIR and READLINK which would otherwise
- * be severely size-limited by a small rdma inline read max. The server
- * response will come back as an RDMA Write, followed by a message
- * of type RDMA_NOMSG carrying the xid and length. As a result, reply
- * chunks do not provide data alignment, however they do not require
- * "fixup" (moving the response to the upper layer buffer) either.
- *
- * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64):
- *
- * Read chunklist (a linked list):
- * N elements, position P (same P for all chunks of same arg!):
- * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0
- *
- * Write chunklist (a list of (one) counted array):
- * N elements:
- * 1 - N - HLOO - HLOO - ... - HLOO - 0
- *
- * Reply chunk (a counted array):
- * N elements:
- * 1 - N - HLOO - HLOO - ... - HLOO
- *
- * Returns positive RPC/RDMA header size, or negative errno.
- */
-
-static ssize_t
-rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target,
- struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type)
-{
- struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
- struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
- int n, nsegs, nchunks = 0;
- unsigned int pos;
- struct rpcrdma_mr_seg *seg = req->rl_segments;
- struct rpcrdma_read_chunk *cur_rchunk = NULL;
- struct rpcrdma_write_array *warray = NULL;
- struct rpcrdma_write_chunk *cur_wchunk = NULL;
- __be32 *iptr = headerp->rm_body.rm_chunks;
- int (*map)(struct rpcrdma_xprt *, struct rpcrdma_mr_seg *, int, bool);
-
- if (type == rpcrdma_readch || type == rpcrdma_areadch) {
- /* a read chunk - server will RDMA Read our memory */
- cur_rchunk = (struct rpcrdma_read_chunk *) iptr;
- } else {
- /* a write or reply chunk - server will RDMA Write our memory */
- *iptr++ = xdr_zero; /* encode a NULL read chunk list */
- if (type == rpcrdma_replych)
- *iptr++ = xdr_zero; /* a NULL write chunk list */
- warray = (struct rpcrdma_write_array *) iptr;
- cur_wchunk = (struct rpcrdma_write_chunk *) (warray + 1);
- }
-
- if (type == rpcrdma_replych || type == rpcrdma_areadch)
- pos = 0;
- else
- pos = target->head[0].iov_len;
-
- nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS);
- if (nsegs < 0)
- return nsegs;
-
- map = r_xprt->rx_ia.ri_ops->ro_map;
- do {
- n = map(r_xprt, seg, nsegs, cur_wchunk != NULL);
- if (n <= 0)
- goto out;
- if (cur_rchunk) { /* read */
- cur_rchunk->rc_discrim = xdr_one;
- /* all read chunks have the same "position" */
- cur_rchunk->rc_position = cpu_to_be32(pos);
- cur_rchunk->rc_target.rs_handle =
- cpu_to_be32(seg->mr_rkey);
- cur_rchunk->rc_target.rs_length =
- cpu_to_be32(seg->mr_len);
- xdr_encode_hyper(
- (__be32 *)&cur_rchunk->rc_target.rs_offset,
- seg->mr_base);
- dprintk("RPC: %s: read chunk "
- "elem %d@0x%llx:0x%x pos %u (%s)\n", __func__,
- seg->mr_len, (unsigned long long)seg->mr_base,
- seg->mr_rkey, pos, n < nsegs ? "more" : "last");
- cur_rchunk++;
- r_xprt->rx_stats.read_chunk_count++;
- } else { /* write/reply */
- cur_wchunk->wc_target.rs_handle =
- cpu_to_be32(seg->mr_rkey);
- cur_wchunk->wc_target.rs_length =
- cpu_to_be32(seg->mr_len);
- xdr_encode_hyper(
- (__be32 *)&cur_wchunk->wc_target.rs_offset,
- seg->mr_base);
- dprintk("RPC: %s: %s chunk "
- "elem %d@0x%llx:0x%x (%s)\n", __func__,
- (type == rpcrdma_replych) ? "reply" : "write",
- seg->mr_len, (unsigned long long)seg->mr_base,
- seg->mr_rkey, n < nsegs ? "more" : "last");
- cur_wchunk++;
- if (type == rpcrdma_replych)
- r_xprt->rx_stats.reply_chunk_count++;
- else
- r_xprt->rx_stats.write_chunk_count++;
- r_xprt->rx_stats.total_rdma_request += seg->mr_len;
- }
- nchunks++;
- seg += n;
- nsegs -= n;
- } while (nsegs);
-
- /* success. all failures return above */
- req->rl_nchunks = nchunks;
-
- /*
- * finish off header. If write, marshal discrim and nchunks.
- */
- if (cur_rchunk) {
- iptr = (__be32 *) cur_rchunk;
- *iptr++ = xdr_zero; /* finish the read chunk list */
- *iptr++ = xdr_zero; /* encode a NULL write chunk list */
- *iptr++ = xdr_zero; /* encode a NULL reply chunk */
- } else {
- warray->wc_discrim = xdr_one;
- warray->wc_nchunks = cpu_to_be32(nchunks);
- iptr = (__be32 *) cur_wchunk;
- if (type == rpcrdma_writech) {
- *iptr++ = xdr_zero; /* finish the write chunk list */
- *iptr++ = xdr_zero; /* encode a NULL reply chunk */
- }
- }
-
- /*
- * Return header size.
- */
- return (unsigned char *)iptr - (unsigned char *)headerp;
-
-out:
- for (pos = 0; nchunks--;)
- pos += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt,
- &req->rl_segments[pos]);
- return n;
-}
-
static inline __be32 *
xdr_encode_rdma_segment(__be32 *iptr, struct rpcrdma_mr_seg *seg)
{
projects / deliverable / linux.git / commitdiff
