diff mbox series

[for-next,v1,08/12] SIW transmit path

Message ID 20190526114156.6827-9-bmt@zurich.ibm.com (mailing list archive)
State Superseded
Headers show
Series SIW: Software iWarp RDMA (siw) driver | expand

Commit Message

Bernard Metzler May 26, 2019, 11:41 a.m. UTC
Signed-off-by: Bernard Metzler <bmt@zurich.ibm.com>
---
 drivers/infiniband/sw/siw/siw_qp_tx.c | 1276 +++++++++++++++++++++++++
 1 file changed, 1276 insertions(+)
 create mode 100644 drivers/infiniband/sw/siw/siw_qp_tx.c

Comments

Leon Romanovsky June 10, 2019, 7:19 a.m. UTC | #1
On Sun, May 26, 2019 at 01:41:52PM +0200, Bernard Metzler wrote:
> Signed-off-by: Bernard Metzler <bmt@zurich.ibm.com>
> ---
>  drivers/infiniband/sw/siw/siw_qp_tx.c | 1276 +++++++++++++++++++++++++
>  1 file changed, 1276 insertions(+)
>  create mode 100644 drivers/infiniband/sw/siw/siw_qp_tx.c
>

I didn't review this patch at all.

Thanks
diff mbox series

Patch

diff --git a/drivers/infiniband/sw/siw/siw_qp_tx.c b/drivers/infiniband/sw/siw/siw_qp_tx.c
new file mode 100644
index 000000000000..4a22c94e55c4
--- /dev/null
+++ b/drivers/infiniband/sw/siw/siw_qp_tx.c
@@ -0,0 +1,1276 @@ 
+// SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
+
+/* Authors: Bernard Metzler <bmt@zurich.ibm.com> */
+/* Copyright (c) 2008-2019, IBM Corporation */
+
+#include <linux/errno.h>
+#include <linux/types.h>
+#include <linux/net.h>
+#include <linux/scatterlist.h>
+#include <linux/highmem.h>
+#include <linux/llist.h>
+#include <net/sock.h>
+#include <net/tcp_states.h>
+#include <net/tcp.h>
+#include <linux/cpu.h>
+
+#include <rdma/iw_cm.h>
+#include <rdma/ib_verbs.h>
+#include <rdma/ib_smi.h>
+#include <rdma/ib_user_verbs.h>
+
+#include "siw.h"
+#include "siw_cm.h"
+#include "siw_verbs.h"
+#include "siw_mem.h"
+#include "siw_debug.h"
+
+#include <linux/kthread.h>
+
+static int siw_crc_txhdr(struct siw_iwarp_tx *ctx)
+{
+	crypto_shash_init(ctx->mpa_crc_hd);
+	return siw_crc_array(ctx->mpa_crc_hd, (u8 *)&ctx->pkt, ctx->ctrl_len);
+}
+
+#define MAX_HDR_INLINE					\
+	(((uint32_t)(sizeof(struct siw_rreq_pkt) -	\
+		     sizeof(struct iwarp_send))) & 0xF8)
+
+static struct page *siw_get_pblpage(struct siw_mem *mem, u64 addr, int *idx)
+{
+	struct siw_pbl *pbl = mem->pbl;
+	u64 offset = addr - mem->va;
+	u64 paddr = siw_pbl_get_buffer(pbl, offset, NULL, idx);
+
+	if (paddr)
+		return virt_to_page(paddr);
+
+	return NULL;
+}
+
+/*
+ * Copy short payload at provided destination payload address
+ */
+static int siw_try_1seg(struct siw_iwarp_tx *c_tx, u64 paddr)
+{
+	struct siw_wqe *wqe = &c_tx->wqe_active;
+	struct siw_sge *sge = &wqe->sqe.sge[0];
+	u32 bytes = sge->length;
+
+	if (bytes > MAX_HDR_INLINE || wqe->sqe.num_sge != 1)
+		return MAX_HDR_INLINE + 1;
+
+	if (!bytes)
+		return 0;
+
+	if (tx_flags(wqe) & SIW_WQE_INLINE) {
+		memcpy((void *)paddr, &wqe->sqe.sge[1], bytes);
+	} else {
+		struct siw_mem *mem = wqe->mem[0];
+
+		if (!mem->mem_obj) {
+			/* Kernel client using kva */
+			memcpy((void *)paddr, (void *)sge->laddr, bytes);
+		} else if (c_tx->in_syscall) {
+			if (copy_from_user((void *)paddr,
+					   (const void __user *)sge->laddr,
+					   bytes))
+				return -EFAULT;
+		} else {
+			unsigned int off = sge->laddr & ~PAGE_MASK;
+			struct page *p;
+			char *buffer;
+			int pbl_idx = 0;
+
+			if (!mem->is_pbl)
+				p = siw_get_upage(mem->umem, sge->laddr);
+			else
+				p = siw_get_pblpage(mem, sge->laddr, &pbl_idx);
+
+			if (unlikely(!p))
+				return -EFAULT;
+
+			buffer = kmap_atomic(p);
+
+			if (likely(PAGE_SIZE - off >= bytes)) {
+				memcpy((void *)paddr, buffer + off, bytes);
+				kunmap_atomic(buffer);
+			} else {
+				unsigned long part = bytes - (PAGE_SIZE - off);
+
+				memcpy((void *)paddr, buffer + off, part);
+				kunmap_atomic(buffer);
+
+				if (!mem->is_pbl)
+					p = siw_get_upage(mem->umem,
+							  sge->laddr + part);
+				else
+					p = siw_get_pblpage(mem,
+							    sge->laddr + part,
+							    &pbl_idx);
+				if (unlikely(!p))
+					return -EFAULT;
+
+				buffer = kmap_atomic(p);
+				memcpy((void *)(paddr + part), buffer,
+				       bytes - part);
+				kunmap_atomic(buffer);
+			}
+		}
+	}
+	return (int)bytes;
+}
+
+#define PKT_FRAGMENTED 1
+#define PKT_COMPLETE 0
+
+/*
+ * siw_qp_prepare_tx()
+ *
+ * Prepare tx state for sending out one fpdu. Builds complete pkt
+ * if no user data or only immediate data are present.
+ *
+ * returns PKT_COMPLETE if complete pkt built, PKT_FRAGMENTED otherwise.
+ */
+static int siw_qp_prepare_tx(struct siw_iwarp_tx *c_tx)
+{
+	struct siw_wqe *wqe = &c_tx->wqe_active;
+	char *crc = NULL;
+	int data = 0;
+
+	switch (tx_type(wqe)) {
+	case SIW_OP_READ:
+	case SIW_OP_READ_LOCAL_INV:
+		memcpy(&c_tx->pkt.ctrl,
+		       &iwarp_pktinfo[RDMAP_RDMA_READ_REQ].ctrl,
+		       sizeof(struct iwarp_ctrl));
+
+		c_tx->pkt.rreq.rsvd = 0;
+		c_tx->pkt.rreq.ddp_qn = htonl(RDMAP_UNTAGGED_QN_RDMA_READ);
+		c_tx->pkt.rreq.ddp_msn =
+			htonl(++c_tx->ddp_msn[RDMAP_UNTAGGED_QN_RDMA_READ]);
+		c_tx->pkt.rreq.ddp_mo = 0;
+		c_tx->pkt.rreq.sink_stag = htonl(wqe->sqe.sge[0].lkey);
+		c_tx->pkt.rreq.sink_to =
+			cpu_to_be64(wqe->sqe.sge[0].laddr);
+		c_tx->pkt.rreq.source_stag = htonl(wqe->sqe.rkey);
+		c_tx->pkt.rreq.source_to = cpu_to_be64(wqe->sqe.raddr);
+		c_tx->pkt.rreq.read_size = htonl(wqe->sqe.sge[0].length);
+
+		c_tx->ctrl_len = sizeof(struct iwarp_rdma_rreq);
+		crc = (char *)&c_tx->pkt.rreq_pkt.crc;
+		break;
+
+	case SIW_OP_SEND:
+		if (tx_flags(wqe) & SIW_WQE_SOLICITED)
+			memcpy(&c_tx->pkt.ctrl,
+			       &iwarp_pktinfo[RDMAP_SEND_SE].ctrl,
+			       sizeof(struct iwarp_ctrl));
+		else
+			memcpy(&c_tx->pkt.ctrl, &iwarp_pktinfo[RDMAP_SEND].ctrl,
+			       sizeof(struct iwarp_ctrl));
+
+		c_tx->pkt.send.ddp_qn = RDMAP_UNTAGGED_QN_SEND;
+		c_tx->pkt.send.ddp_msn =
+			htonl(++c_tx->ddp_msn[RDMAP_UNTAGGED_QN_SEND]);
+		c_tx->pkt.send.ddp_mo = 0;
+
+		c_tx->pkt.send_inv.inval_stag = 0;
+
+		c_tx->ctrl_len = sizeof(struct iwarp_send);
+
+		crc = (char *)&c_tx->pkt.send_pkt.crc;
+		data = siw_try_1seg(c_tx, (u64)crc);
+		break;
+
+	case SIW_OP_SEND_REMOTE_INV:
+		if (tx_flags(wqe) & SIW_WQE_SOLICITED)
+			memcpy(&c_tx->pkt.ctrl,
+			       &iwarp_pktinfo[RDMAP_SEND_SE_INVAL].ctrl,
+			       sizeof(struct iwarp_ctrl));
+		else
+			memcpy(&c_tx->pkt.ctrl,
+			       &iwarp_pktinfo[RDMAP_SEND_INVAL].ctrl,
+			       sizeof(struct iwarp_ctrl));
+
+		c_tx->pkt.send.ddp_qn = RDMAP_UNTAGGED_QN_SEND;
+		c_tx->pkt.send.ddp_msn =
+			htonl(++c_tx->ddp_msn[RDMAP_UNTAGGED_QN_SEND]);
+		c_tx->pkt.send.ddp_mo = 0;
+
+		c_tx->pkt.send_inv.inval_stag = cpu_to_be32(wqe->sqe.rkey);
+
+		c_tx->ctrl_len = sizeof(struct iwarp_send_inv);
+
+		crc = (char *)&c_tx->pkt.send_pkt.crc;
+		data = siw_try_1seg(c_tx, (u64)crc);
+		break;
+
+	case SIW_OP_WRITE:
+		memcpy(&c_tx->pkt.ctrl, &iwarp_pktinfo[RDMAP_RDMA_WRITE].ctrl,
+		       sizeof(struct iwarp_ctrl));
+
+		c_tx->pkt.rwrite.sink_stag = htonl(wqe->sqe.rkey);
+		c_tx->pkt.rwrite.sink_to = cpu_to_be64(wqe->sqe.raddr);
+		c_tx->ctrl_len = sizeof(struct iwarp_rdma_write);
+
+		crc = (char *)&c_tx->pkt.write_pkt.crc;
+		data = siw_try_1seg(c_tx, (u64)crc);
+		break;
+
+	case SIW_OP_READ_RESPONSE:
+		memcpy(&c_tx->pkt.ctrl,
+		       &iwarp_pktinfo[RDMAP_RDMA_READ_RESP].ctrl,
+		       sizeof(struct iwarp_ctrl));
+
+		/* NBO */
+		c_tx->pkt.rresp.sink_stag = cpu_to_be32(wqe->sqe.rkey);
+		c_tx->pkt.rresp.sink_to = cpu_to_be64(wqe->sqe.raddr);
+
+		c_tx->ctrl_len = sizeof(struct iwarp_rdma_rresp);
+
+		crc = (char *)&c_tx->pkt.write_pkt.crc;
+		data = siw_try_1seg(c_tx, (u64)crc);
+		break;
+
+	default:
+		siw_dbg_qp(tx_qp(c_tx), "stale wqe type %d\n", tx_type(wqe));
+		return -EOPNOTSUPP;
+	}
+	if (unlikely(data < 0))
+		return data;
+
+	c_tx->ctrl_sent = 0;
+
+	if (data <= MAX_HDR_INLINE) {
+		if (data) {
+			wqe->processed = data;
+
+			c_tx->pkt.ctrl.mpa_len =
+				htons(c_tx->ctrl_len + data - MPA_HDR_SIZE);
+
+			/* Add pad, if needed */
+			data += -(int)data & 0x3;
+			/* advance CRC location after payload */
+			crc += data;
+			c_tx->ctrl_len += data;
+
+			if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED))
+				c_tx->pkt.c_untagged.ddp_mo = 0;
+			else
+				c_tx->pkt.c_tagged.ddp_to =
+					cpu_to_be64(wqe->sqe.raddr);
+		}
+
+		*(u32 *)crc = 0;
+		/*
+		 * Do complete CRC if enabled and short packet
+		 */
+		if (c_tx->mpa_crc_hd) {
+			if (siw_crc_txhdr(c_tx) != 0)
+				return -EINVAL;
+			crypto_shash_final(c_tx->mpa_crc_hd, (u8 *)crc);
+		}
+		c_tx->ctrl_len += MPA_CRC_SIZE;
+
+		return PKT_COMPLETE;
+	}
+	c_tx->ctrl_len += MPA_CRC_SIZE;
+	c_tx->sge_idx = 0;
+	c_tx->sge_off = 0;
+	c_tx->pbl_idx = 0;
+
+	/*
+	 * Allow direct sending out of user buffer if WR is non signalled
+	 * and payload is over threshold.
+	 * Per RDMA verbs, the application should not change the send buffer
+	 * until the work completed. In iWarp, work completion is only
+	 * local delivery to TCP. TCP may reuse the buffer for
+	 * retransmission. Changing unsent data also breaks the CRC,
+	 * if applied.
+	 */
+	if (c_tx->zcopy_tx && wqe->bytes >= SENDPAGE_THRESH &&
+	    !(tx_flags(wqe) & SIW_WQE_SIGNALLED))
+		c_tx->use_sendpage = 1;
+	else
+		c_tx->use_sendpage = 0;
+
+	return PKT_FRAGMENTED;
+}
+
+/*
+ * Send out one complete control type FPDU, or header of FPDU carrying
+ * data. Used for fixed sized packets like Read.Requests or zero length
+ * SENDs, WRITEs, READ.Responses, or header only.
+ */
+static int siw_tx_ctrl(struct siw_iwarp_tx *c_tx, struct socket *s,
+			      int flags)
+{
+	struct msghdr msg = { .msg_flags = flags };
+	struct kvec iov = { .iov_base =
+				    (char *)&c_tx->pkt.ctrl + c_tx->ctrl_sent,
+			    .iov_len = c_tx->ctrl_len - c_tx->ctrl_sent };
+
+	int rv = kernel_sendmsg(s, &msg, &iov, 1,
+				c_tx->ctrl_len - c_tx->ctrl_sent);
+
+	if (rv >= 0) {
+		c_tx->ctrl_sent += rv;
+
+		if (c_tx->ctrl_sent == c_tx->ctrl_len) {
+			siw_dbg_pkt(&c_tx->pkt.hdr, tx_qp(c_tx),
+				    "HDR/CTRL sent");
+			rv = 0;
+		} else
+			rv = -EAGAIN;
+	}
+	return rv;
+}
+
+/*
+ * 0copy TCP transmit interface: Use do_tcp_sendpages.
+ *
+ * Using sendpage to push page by page appears to be less efficient
+ * than using sendmsg, even if data are copied.
+ *
+ * A general performance limitation might be the extra four bytes
+ * trailer checksum segment to be pushed after user data.
+ */
+static int siw_tcp_sendpages(struct socket *s, struct page **page, int offset,
+			     size_t size)
+{
+	struct sock *sk = s->sk;
+	int i = 0, rv = 0, sent = 0,
+	    flags = MSG_MORE | MSG_DONTWAIT | MSG_SENDPAGE_NOTLAST;
+
+	while (size) {
+		size_t bytes = min_t(size_t, PAGE_SIZE - offset, size);
+
+		if (size + offset <= PAGE_SIZE)
+			flags = MSG_MORE | MSG_DONTWAIT;
+
+		tcp_rate_check_app_limited(sk);
+try_page_again:
+		lock_sock(sk);
+		rv = do_tcp_sendpages(sk, page[i], offset, bytes, flags);
+		release_sock(sk);
+
+		if (rv > 0) {
+			size -= rv;
+			sent += rv;
+			if (rv != bytes) {
+				offset += rv;
+				bytes -= rv;
+				goto try_page_again;
+			}
+			offset = 0;
+		} else {
+			if (rv == -EAGAIN || rv == 0)
+				break;
+			return rv;
+		}
+		i++;
+	}
+	return sent;
+}
+
+/*
+ * siw_0copy_tx()
+ *
+ * Pushes list of pages to TCP socket. If pages from multiple
+ * SGE's, all referenced pages of each SGE are pushed in one
+ * shot.
+ */
+static int siw_0copy_tx(struct socket *s, struct page **page,
+			struct siw_sge *sge, unsigned int offset,
+			unsigned int size)
+{
+	int i = 0, sent = 0, rv;
+	int sge_bytes = min(sge->length - offset, size);
+
+	offset = (sge->laddr + offset) & ~PAGE_MASK;
+
+	while (sent != size) {
+		rv = siw_tcp_sendpages(s, &page[i], offset, sge_bytes);
+		if (rv >= 0) {
+			sent += rv;
+			if (size == sent || sge_bytes > rv)
+				break;
+
+			i += PAGE_ALIGN(sge_bytes + offset) >> PAGE_SHIFT;
+			sge++;
+			sge_bytes = min(sge->length, size - sent);
+			offset = sge->laddr & ~PAGE_MASK;
+		} else {
+			sent = rv;
+			break;
+		}
+	}
+	return sent;
+}
+
+#define MAX_TRAILER (MPA_CRC_SIZE + 4)
+
+static void siw_unmap_pages(struct page **pages, int hdr_len, int num_maps)
+{
+	if (hdr_len) {
+		++pages;
+		--num_maps;
+	}
+	while (num_maps-- > 0) {
+		kunmap(*pages);
+		pages++;
+	}
+}
+
+/*
+ * siw_tx_hdt() tries to push a complete packet to TCP where all
+ * packet fragments are referenced by the elements of one iovec.
+ * For the data portion, each involved page must be referenced by
+ * one extra element. All sge's data can be non-aligned to page
+ * boundaries. Two more elements are referencing iWARP header
+ * and trailer:
+ * MAX_ARRAY = 64KB/PAGE_SIZE + 1 + (2 * (SIW_MAX_SGE - 1) + HDR + TRL
+ */
+#define MAX_ARRAY ((0xffff / PAGE_SIZE) + 1 + (2 * (SIW_MAX_SGE - 1) + 2))
+
+/*
+ * Write out iov referencing hdr, data and trailer of current FPDU.
+ * Update transmit state dependent on write return status
+ */
+static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s)
+{
+	struct siw_wqe *wqe = &c_tx->wqe_active;
+	struct siw_sge *sge = &wqe->sqe.sge[c_tx->sge_idx];
+	struct kvec iov[MAX_ARRAY];
+	struct page *page_array[MAX_ARRAY];
+	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
+
+	int seg = 0, do_crc = c_tx->do_crc, is_kva = 0, rv;
+	unsigned int data_len = c_tx->bytes_unsent, hdr_len = 0, trl_len = 0,
+		     sge_off = c_tx->sge_off, sge_idx = c_tx->sge_idx,
+		     pbl_idx = c_tx->pbl_idx;
+
+	if (c_tx->state == SIW_SEND_HDR) {
+		if (c_tx->use_sendpage) {
+			rv = siw_tx_ctrl(c_tx, s, MSG_DONTWAIT | MSG_MORE);
+			if (rv)
+				goto done;
+
+			c_tx->state = SIW_SEND_DATA;
+		} else {
+			iov[0].iov_base =
+				(char *)&c_tx->pkt.ctrl + c_tx->ctrl_sent;
+			iov[0].iov_len = hdr_len =
+				c_tx->ctrl_len - c_tx->ctrl_sent;
+			seg = 1;
+			siw_dbg_pkt(&c_tx->pkt.hdr, tx_qp(c_tx),
+				    "HDR to send");
+		}
+	}
+
+	wqe->processed += data_len;
+
+	while (data_len) { /* walk the list of SGE's */
+		unsigned int sge_len = min(sge->length - sge_off, data_len);
+		unsigned int fp_off = (sge->laddr + sge_off) & ~PAGE_MASK;
+		struct siw_mem *mem;
+
+		if (!(tx_flags(wqe) & SIW_WQE_INLINE)) {
+			mem = wqe->mem[sge_idx];
+			if (!mem->mem_obj)
+				is_kva = 1;
+		} else {
+			is_kva = 1;
+		}
+		if (is_kva && !c_tx->use_sendpage) {
+			/*
+			 * tx from kernel virtual address: either inline data
+			 * or memory region with assigned kernel buffer
+			 */
+			iov[seg].iov_base = (void *)(sge->laddr + sge_off);
+			iov[seg].iov_len = sge_len;
+
+			if (do_crc)
+				siw_crc_array(c_tx->mpa_crc_hd,
+					      iov[seg].iov_base, sge_len);
+			sge_off += sge_len;
+			data_len -= sge_len;
+			seg++;
+			goto sge_done;
+		}
+
+		while (sge_len) {
+			size_t plen = min((int)PAGE_SIZE - fp_off, sge_len);
+
+			if (!is_kva) {
+				struct page *p;
+
+				if (mem->is_pbl)
+					p = siw_get_pblpage(
+						mem, sge->laddr + sge_off,
+						&pbl_idx);
+				else
+					p = siw_get_upage(mem->umem,
+							  sge->laddr + sge_off);
+				if (unlikely(!p)) {
+					if (hdr_len)
+						seg--;
+					if (!c_tx->use_sendpage && seg) {
+						siw_unmap_pages(page_array,
+								hdr_len, seg);
+					}
+					wqe->processed -= c_tx->bytes_unsent;
+					rv = -EFAULT;
+					goto done_crc;
+				}
+				page_array[seg] = p;
+
+				if (!c_tx->use_sendpage) {
+					iov[seg].iov_base = kmap(p) + fp_off;
+					iov[seg].iov_len = plen;
+				}
+				if (do_crc)
+					siw_crc_page(c_tx->mpa_crc_hd, p,
+						     fp_off, plen);
+			} else {
+				u64 pa = ((sge->laddr + sge_off) & PAGE_MASK);
+
+				page_array[seg] = virt_to_page(pa);
+				if (do_crc)
+					siw_crc_array(
+						c_tx->mpa_crc_hd,
+						(void *)(sge->laddr + sge_off),
+						plen);
+			}
+
+			sge_len -= plen;
+			sge_off += plen;
+			data_len -= plen;
+			fp_off = 0;
+
+			if (++seg > (int)MAX_ARRAY) {
+				siw_dbg_qp(tx_qp(c_tx), "to many fragments\n");
+				if (!is_kva && !c_tx->use_sendpage) {
+					siw_unmap_pages(page_array, hdr_len,
+							seg - 1);
+				}
+				wqe->processed -= c_tx->bytes_unsent;
+				rv = -EMSGSIZE;
+				goto done_crc;
+			}
+		}
+sge_done:
+		/* Update SGE variables at end of SGE */
+		if (sge_off == sge->length &&
+		    (data_len != 0 || wqe->processed < wqe->bytes)) {
+			sge_idx++;
+			sge++;
+			sge_off = 0;
+		}
+	}
+	/* trailer */
+	if (likely(c_tx->state != SIW_SEND_TRAILER)) {
+		iov[seg].iov_base = &c_tx->trailer.pad[4 - c_tx->pad];
+		iov[seg].iov_len = trl_len = MAX_TRAILER - (4 - c_tx->pad);
+	} else {
+		iov[seg].iov_base = &c_tx->trailer.pad[c_tx->ctrl_sent];
+		iov[seg].iov_len = trl_len = MAX_TRAILER - c_tx->ctrl_sent;
+	}
+
+	if (c_tx->pad) {
+		*(u32 *)c_tx->trailer.pad = 0;
+		if (do_crc)
+			siw_crc_array(c_tx->mpa_crc_hd,
+				      (u8 *)&c_tx->trailer.crc - c_tx->pad,
+				      c_tx->pad);
+	}
+	if (!c_tx->mpa_crc_hd)
+		c_tx->trailer.crc = 0;
+	else if (do_crc)
+		crypto_shash_final(c_tx->mpa_crc_hd, (u8 *)&c_tx->trailer.crc);
+
+	data_len = c_tx->bytes_unsent;
+
+	if (c_tx->use_sendpage) {
+		rv = siw_0copy_tx(s, page_array, &wqe->sqe.sge[c_tx->sge_idx],
+				  c_tx->sge_off, data_len);
+		if (rv == data_len) {
+			rv = kernel_sendmsg(s, &msg, &iov[seg], 1, trl_len);
+			if (rv > 0)
+				rv += data_len;
+			else
+				rv = data_len;
+		}
+	} else {
+		rv = kernel_sendmsg(s, &msg, iov, seg + 1,
+				    hdr_len + data_len + trl_len);
+		if (!is_kva)
+			siw_unmap_pages(page_array, hdr_len, seg);
+	}
+	if (rv < (int)hdr_len) {
+		/* Not even complete hdr pushed or negative rv */
+		wqe->processed -= data_len;
+		if (rv >= 0) {
+			c_tx->ctrl_sent += rv;
+			rv = -EAGAIN;
+		}
+		goto done_crc;
+	}
+	rv -= hdr_len;
+
+	if (rv >= (int)data_len) {
+		/* all user data pushed to TCP or no data to push */
+		if (data_len > 0 && wqe->processed < wqe->bytes) {
+			/* Save the current state for next tx */
+			c_tx->sge_idx = sge_idx;
+			c_tx->sge_off = sge_off;
+			c_tx->pbl_idx = pbl_idx;
+		}
+		rv -= data_len;
+
+		if (rv == trl_len) /* all pushed */
+			rv = 0;
+		else {
+			c_tx->state = SIW_SEND_TRAILER;
+			c_tx->ctrl_len = MAX_TRAILER;
+			c_tx->ctrl_sent = rv + 4 - c_tx->pad;
+			c_tx->bytes_unsent = 0;
+			rv = -EAGAIN;
+		}
+
+	} else if (data_len > 0) {
+		/* Maybe some user data pushed to TCP */
+		c_tx->state = SIW_SEND_DATA;
+		wqe->processed -= data_len - rv;
+
+		if (rv) {
+			/*
+			 * Some bytes out. Recompute tx state based
+			 * on old state and bytes pushed
+			 */
+			unsigned int sge_unsent;
+
+			c_tx->bytes_unsent -= rv;
+			sge = &wqe->sqe.sge[c_tx->sge_idx];
+			sge_unsent = sge->length - c_tx->sge_off;
+
+			while (sge_unsent <= rv) {
+				rv -= sge_unsent;
+				c_tx->sge_idx++;
+				c_tx->sge_off = 0;
+				sge++;
+				sge_unsent = sge->length;
+			}
+			c_tx->sge_off += rv;
+		}
+		rv = -EAGAIN;
+	}
+done_crc:
+	c_tx->do_crc = 0;
+done:
+	return rv;
+}
+
+static void siw_update_tcpseg(struct siw_iwarp_tx *c_tx,
+				     struct socket *s)
+{
+	struct tcp_sock *tp = tcp_sk(s->sk);
+
+	if (tp->gso_segs) {
+		if (c_tx->gso_seg_limit == 0)
+			c_tx->tcp_seglen = tp->mss_cache * tp->gso_segs;
+		else
+			c_tx->tcp_seglen =
+				tp->mss_cache *
+				min_t(u16, c_tx->gso_seg_limit, tp->gso_segs);
+	} else {
+		c_tx->tcp_seglen = tp->mss_cache;
+	}
+	/* Loopback may give odd numbers */
+	c_tx->tcp_seglen &= 0xfffffff8;
+}
+
+/*
+ * siw_prepare_fpdu()
+ *
+ * Prepares transmit context to send out one FPDU if FPDU will contain
+ * user data and user data are not immediate data.
+ * Computes maximum FPDU length to fill up TCP MSS if possible.
+ *
+ * @qp:		QP from which to transmit
+ * @wqe:	Current WQE causing transmission
+ *
+ * TODO: Take into account real available sendspace on socket
+ *       to avoid header misalignment due to send pausing within
+ *       fpdu transmission
+ */
+static void siw_prepare_fpdu(struct siw_qp *qp, struct siw_wqe *wqe)
+{
+	struct siw_iwarp_tx *c_tx = &qp->tx_ctx;
+	int data_len;
+
+	c_tx->ctrl_len =
+		iwarp_pktinfo[__rdmap_get_opcode(&c_tx->pkt.ctrl)].hdr_len;
+	c_tx->ctrl_sent = 0;
+
+	/*
+	 * Update target buffer offset if any
+	 */
+	if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED))
+		/* Untagged message */
+		c_tx->pkt.c_untagged.ddp_mo = cpu_to_be32(wqe->processed);
+	else /* Tagged message */
+		c_tx->pkt.c_tagged.ddp_to =
+			cpu_to_be64(wqe->sqe.raddr + wqe->processed);
+
+	data_len = wqe->bytes - wqe->processed;
+	if (data_len + c_tx->ctrl_len + MPA_CRC_SIZE > c_tx->tcp_seglen) {
+		/* Trim DDP payload to fit into current TCP segment */
+		data_len = c_tx->tcp_seglen - (c_tx->ctrl_len + MPA_CRC_SIZE);
+		c_tx->pkt.ctrl.ddp_rdmap_ctrl &= ~DDP_FLAG_LAST;
+		c_tx->pad = 0;
+	} else {
+		c_tx->pkt.ctrl.ddp_rdmap_ctrl |= DDP_FLAG_LAST;
+		c_tx->pad = -data_len & 0x3;
+	}
+	c_tx->bytes_unsent = data_len;
+
+	c_tx->pkt.ctrl.mpa_len =
+		htons(c_tx->ctrl_len + data_len - MPA_HDR_SIZE);
+
+	/*
+	 * Init MPA CRC computation
+	 */
+	if (c_tx->mpa_crc_hd) {
+		siw_crc_txhdr(c_tx);
+		c_tx->do_crc = 1;
+	}
+}
+
+/*
+ * siw_check_sgl_tx()
+ *
+ * Check permissions for a list of SGE's (SGL).
+ * A successful check will have all memory referenced
+ * for transmission resolved and assigned to the WQE.
+ *
+ * @pd:		Protection Domain SGL should belong to
+ * @wqe:	WQE to be checked
+ * @perms:	requested access permissions
+ *
+ */
+
+static int siw_check_sgl_tx(struct ib_pd *pd, struct siw_wqe *wqe,
+			    enum ib_access_flags perms)
+{
+	struct siw_sge *sge = &wqe->sqe.sge[0];
+	int i, len, num_sge = wqe->sqe.num_sge;
+
+	if (unlikely(num_sge > SIW_MAX_SGE))
+		return -EINVAL;
+
+	for (i = 0, len = 0; num_sge; num_sge--, i++, sge++) {
+		/*
+		 * rdma verbs: do not check stag for a zero length sge
+		 */
+		if (sge->length) {
+			int rv = siw_check_sge(pd, sge, &wqe->mem[i], perms, 0,
+					       sge->length);
+
+			if (unlikely(rv != E_ACCESS_OK))
+				return rv;
+		}
+		len += sge->length;
+	}
+	return len;
+}
+
+/*
+ * siw_qp_sq_proc_tx()
+ *
+ * Process one WQE which needs transmission on the wire.
+ */
+static int siw_qp_sq_proc_tx(struct siw_qp *qp, struct siw_wqe *wqe)
+{
+	struct siw_iwarp_tx *c_tx = &qp->tx_ctx;
+	struct socket *s = qp->attrs.sk;
+	int rv = 0, burst_len = qp->tx_ctx.burst;
+	enum rdmap_ecode ecode = RDMAP_ECODE_CATASTROPHIC_STREAM;
+
+	if (unlikely(wqe->wr_status == SIW_WR_IDLE))
+		return 0;
+
+	if (!burst_len)
+		burst_len = SQ_USER_MAXBURST;
+
+	if (wqe->wr_status == SIW_WR_QUEUED) {
+		if (!(wqe->sqe.flags & SIW_WQE_INLINE)) {
+			if (tx_type(wqe) == SIW_OP_READ_RESPONSE)
+				wqe->sqe.num_sge = 1;
+
+			if (tx_type(wqe) != SIW_OP_READ &&
+			    tx_type(wqe) != SIW_OP_READ_LOCAL_INV) {
+				/*
+				 * Reference memory to be tx'd w/o checking
+				 * access for LOCAL_READ permission, since
+				 * not defined in RDMA core.
+				 */
+				rv = siw_check_sgl_tx(qp->pd, wqe, 0);
+				if (rv < 0) {
+					if (tx_type(wqe) ==
+					    SIW_OP_READ_RESPONSE)
+						ecode = siw_rdmap_error(-rv);
+					rv = -EINVAL;
+					goto tx_error;
+				}
+				wqe->bytes = rv;
+			} else {
+				wqe->bytes = 0;
+			}
+		} else {
+			wqe->bytes = wqe->sqe.sge[0].length;
+			if (!qp->kernel_verbs) {
+				if (wqe->bytes > SIW_MAX_INLINE) {
+					rv = -EINVAL;
+					goto tx_error;
+				}
+				wqe->sqe.sge[0].laddr = (u64)&wqe->sqe.sge[1];
+			}
+		}
+		wqe->wr_status = SIW_WR_INPROGRESS;
+		wqe->processed = 0;
+
+		siw_update_tcpseg(c_tx, s);
+
+		rv = siw_qp_prepare_tx(c_tx);
+		if (rv == PKT_FRAGMENTED) {
+			c_tx->state = SIW_SEND_HDR;
+			siw_prepare_fpdu(qp, wqe);
+		} else if (rv == PKT_COMPLETE) {
+			c_tx->state = SIW_SEND_SHORT_FPDU;
+		} else {
+			goto tx_error;
+		}
+	}
+
+next_segment:
+	siw_dbg_qp(qp, "wr type %d, state %d, data %u, sent %u, id %llx\n",
+		   tx_type(wqe), wqe->wr_status, wqe->bytes, wqe->processed,
+		   wqe->sqe.id);
+
+	if (--burst_len == 0) {
+		rv = -EINPROGRESS;
+		goto tx_done;
+	}
+	if (c_tx->state == SIW_SEND_SHORT_FPDU) {
+		enum siw_opcode tx_type = tx_type(wqe);
+		unsigned int msg_flags;
+
+		if (siw_sq_empty(qp) || !siw_tcp_nagle || burst_len == 1)
+			/*
+			 * End current TCP segment, if SQ runs empty,
+			 * or siw_tcp_nagle is not set, or we bail out
+			 * soon due to no burst credit left.
+			 */
+			msg_flags = MSG_DONTWAIT;
+		else
+			msg_flags = MSG_DONTWAIT | MSG_MORE;
+
+		rv = siw_tx_ctrl(c_tx, s, msg_flags);
+
+		if (!rv && tx_type != SIW_OP_READ &&
+		    tx_type != SIW_OP_READ_LOCAL_INV)
+			wqe->processed = wqe->bytes;
+
+		goto tx_done;
+
+	} else {
+		rv = siw_tx_hdt(c_tx, s);
+	}
+	if (!rv) {
+		/*
+		 * One segment sent. Processing completed if last
+		 * segment, Do next segment otherwise.
+		 */
+		if (unlikely(c_tx->tx_suspend)) {
+			/*
+			 * Verbs, 6.4.: Try stopping sending after a full
+			 * DDP segment if the connection goes down
+			 * (== peer halfclose)
+			 */
+			rv = -ECONNABORTED;
+			goto tx_done;
+		}
+		if (c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_LAST) {
+			siw_dbg_qp(qp, "WQE completed\n");
+			goto tx_done;
+		}
+		c_tx->state = SIW_SEND_HDR;
+
+		siw_update_tcpseg(c_tx, s);
+
+		siw_prepare_fpdu(qp, wqe);
+		goto next_segment;
+	}
+tx_done:
+	qp->tx_ctx.burst = burst_len;
+	return rv;
+
+tx_error:
+	if (ecode != RDMAP_ECODE_CATASTROPHIC_STREAM)
+		siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP,
+				   RDMAP_ETYPE_REMOTE_PROTECTION, ecode, 1);
+	else
+		siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP,
+				   RDMAP_ETYPE_CATASTROPHIC,
+				   RDMAP_ECODE_UNSPECIFIED, 1);
+	return rv;
+}
+
+static int siw_fastreg_mr(struct ib_pd *pd, struct siw_sqe *sqe)
+{
+	struct ib_mr *base_mr = (struct ib_mr *)sqe->base_mr;
+	struct siw_device *sdev = to_siw_dev(pd->device);
+	struct siw_mem *mem = siw_mem_id2obj(sdev, sqe->rkey  >> 8);
+	int rv = 0;
+
+	siw_dbg_pd(pd, "STag 0x%08x\n", sqe->rkey);
+
+	if (unlikely(!mem || !base_mr)) {
+		pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey);
+		return -EINVAL;
+	}
+	if (unlikely(base_mr->rkey >> 8 != sqe->rkey  >> 8)) {
+		pr_warn("siw: fastreg: STag 0x%08x: bad MR\n", sqe->rkey);
+		rv = -EINVAL;
+		goto out;
+	}
+	if (unlikely(mem->pd != pd)) {
+		pr_warn("siw: fastreg: PD mismatch\n");
+		rv = -EINVAL;
+		goto out;
+	}
+	if (unlikely(mem->stag_valid)) {
+		pr_warn("siw: fastreg: STag 0x%08x already valid\n", sqe->rkey);
+		rv = -EINVAL;
+		goto out;
+	}
+	/* Refresh STag since user may have changed key part */
+	mem->stag = sqe->rkey;
+	mem->perms = sqe->access;
+
+	siw_dbg_mem(mem, "STag now valid, MR va: 0x%016llx -> 0x%016llx\n",
+		    mem->va, base_mr->iova);
+	mem->va = base_mr->iova;
+	mem->stag_valid = 1;
+out:
+	siw_mem_put(mem);
+	return rv;
+}
+
+static int siw_qp_sq_proc_local(struct siw_qp *qp, struct siw_wqe *wqe)
+{
+	int rv;
+
+	switch (tx_type(wqe)) {
+	case SIW_OP_REG_MR:
+		rv = siw_fastreg_mr(qp->pd, &wqe->sqe);
+		break;
+
+	case SIW_OP_INVAL_STAG:
+		rv = siw_invalidate_stag(qp->pd, wqe->sqe.rkey);
+		break;
+
+	default:
+		rv = -EINVAL;
+	}
+	return rv;
+}
+
+/*
+ * siw_qp_sq_process()
+ *
+ * Core TX path routine for RDMAP/DDP/MPA using a TCP kernel socket.
+ * Sends RDMAP payload for the current SQ WR @wqe of @qp in one or more
+ * MPA FPDUs, each containing a DDP segment.
+ *
+ * SQ processing may occur in user context as a result of posting
+ * new WQE's or from siw_sq_work_handler() context. Processing in
+ * user context is limited to non-kernel verbs users.
+ *
+ * SQ processing may get paused anytime, possibly in the middle of a WR
+ * or FPDU, if insufficient send space is available. SQ processing
+ * gets resumed from siw_sq_work_handler(), if send space becomes
+ * available again.
+ *
+ * Must be called with the QP state read-locked.
+ *
+ * Note:
+ * An outbound RREQ can be satisfied by the corresponding RRESP
+ * _before_ it gets assigned to the ORQ. This happens regularly
+ * in RDMA READ via loopback case. Since both outbound RREQ and
+ * inbound RRESP can be handled by the same CPU, locking the ORQ
+ * is dead-lock prone and thus not an option. With that, the
+ * RREQ gets assigned to the ORQ _before_ being sent - see
+ * siw_activate_tx() - and pulled back in case of send failure.
+ */
+int siw_qp_sq_process(struct siw_qp *qp)
+{
+	struct siw_wqe *wqe = tx_wqe(qp);
+	enum siw_opcode tx_type;
+	unsigned long flags;
+	int rv = 0;
+
+	siw_dbg_qp(qp, "enter for type %d\n", tx_type(wqe));
+
+next_wqe:
+	/*
+	 * Stop QP processing if SQ state changed
+	 */
+	if (unlikely(qp->tx_ctx.tx_suspend)) {
+		siw_dbg_qp(qp, "tx suspended\n");
+		goto done;
+	}
+	tx_type = tx_type(wqe);
+
+	if (tx_type <= SIW_OP_READ_RESPONSE)
+		rv = siw_qp_sq_proc_tx(qp, wqe);
+	else
+		rv = siw_qp_sq_proc_local(qp, wqe);
+
+	if (!rv) {
+		/*
+		 * WQE processing done
+		 */
+		switch (tx_type) {
+		case SIW_OP_SEND:
+		case SIW_OP_SEND_REMOTE_INV:
+		case SIW_OP_WRITE:
+			siw_wqe_put_mem(wqe, tx_type);
+		case SIW_OP_INVAL_STAG:
+		case SIW_OP_REG_MR:
+			if (tx_flags(wqe) & SIW_WQE_SIGNALLED)
+				siw_sqe_complete(qp, &wqe->sqe, wqe->bytes,
+						 SIW_WC_SUCCESS);
+			break;
+
+		case SIW_OP_READ:
+		case SIW_OP_READ_LOCAL_INV:
+			/*
+			 * already enqueued to ORQ queue
+			 */
+			break;
+
+		case SIW_OP_READ_RESPONSE:
+			siw_wqe_put_mem(wqe, tx_type);
+			break;
+
+		default:
+			WARN(1, "undefined WQE type %d\n", tx_type);
+			rv = -EINVAL;
+			goto done;
+		}
+
+		spin_lock_irqsave(&qp->sq_lock, flags);
+		wqe->wr_status = SIW_WR_IDLE;
+		rv = siw_activate_tx(qp);
+		spin_unlock_irqrestore(&qp->sq_lock, flags);
+
+		if (rv <= 0)
+			goto done;
+
+		goto next_wqe;
+
+	} else if (rv == -EAGAIN) {
+		siw_dbg_qp(qp, "sq paused: hd/tr %d of %d, data %d\n",
+			   qp->tx_ctx.ctrl_sent, qp->tx_ctx.ctrl_len,
+			   qp->tx_ctx.bytes_unsent);
+		rv = 0;
+		goto done;
+	} else if (rv == -EINPROGRESS) {
+		rv = siw_sq_start(qp);
+		goto done;
+	} else {
+		/*
+		 * WQE processing failed.
+		 * Verbs 8.3.2:
+		 * o It turns any WQE into a signalled WQE.
+		 * o Local catastrophic error must be surfaced
+		 * o QP must be moved into Terminate state: done by code
+		 *   doing socket state change processing
+		 *
+		 * o TODO: Termination message must be sent.
+		 * o TODO: Implement more precise work completion errors,
+		 *         see enum ib_wc_status in ib_verbs.h
+		 */
+		siw_dbg_qp(qp, "wqe type %d processing failed: %d\n",
+			   tx_type(wqe), rv);
+
+		spin_lock_irqsave(&qp->sq_lock, flags);
+		/*
+		 * RREQ may have already been completed by inbound RRESP!
+		 */
+		if (tx_type == SIW_OP_READ ||
+		    tx_type == SIW_OP_READ_LOCAL_INV) {
+			/* Cleanup pending entry in ORQ */
+			qp->orq_put--;
+			qp->orq[qp->orq_put % qp->attrs.orq_size].flags = 0;
+		}
+		spin_unlock_irqrestore(&qp->sq_lock, flags);
+		/*
+		 * immediately suspends further TX processing
+		 */
+		if (!qp->tx_ctx.tx_suspend)
+			siw_qp_cm_drop(qp, 0);
+
+		switch (tx_type) {
+		case SIW_OP_SEND:
+		case SIW_OP_SEND_REMOTE_INV:
+		case SIW_OP_SEND_WITH_IMM:
+		case SIW_OP_WRITE:
+		case SIW_OP_READ:
+		case SIW_OP_READ_LOCAL_INV:
+			siw_wqe_put_mem(wqe, tx_type);
+		case SIW_OP_INVAL_STAG:
+		case SIW_OP_REG_MR:
+			siw_sqe_complete(qp, &wqe->sqe, wqe->bytes,
+					 SIW_WC_LOC_QP_OP_ERR);
+
+			siw_qp_event(qp, IB_EVENT_QP_FATAL);
+
+			break;
+
+		case SIW_OP_READ_RESPONSE:
+			siw_dbg_qp(qp, "proc. read.response failed: %d\n", rv);
+
+			siw_qp_event(qp, IB_EVENT_QP_REQ_ERR);
+
+			siw_wqe_put_mem(wqe, SIW_OP_READ_RESPONSE);
+
+			break;
+
+		default:
+			WARN(1, "undefined WQE type %d\n", tx_type);
+			rv = -EINVAL;
+		}
+		wqe->wr_status = SIW_WR_IDLE;
+	}
+done:
+	return rv;
+}
+
+static void siw_sq_resume(struct siw_qp *qp)
+{
+	if (down_read_trylock(&qp->state_lock)) {
+		if (likely(qp->attrs.state == SIW_QP_STATE_RTS &&
+			   !qp->tx_ctx.tx_suspend)) {
+			int rv = siw_qp_sq_process(qp);
+
+			up_read(&qp->state_lock);
+
+			if (unlikely(rv < 0)) {
+				siw_dbg_qp(qp, "SQ task failed: err %d\n", rv);
+
+				if (!qp->tx_ctx.tx_suspend)
+					siw_qp_cm_drop(qp, 0);
+			}
+		} else {
+			up_read(&qp->state_lock);
+		}
+	} else {
+		siw_dbg_qp(qp, "Resume SQ while QP locked\n");
+	}
+	siw_qp_put(qp);
+}
+
+struct tx_task_t {
+	struct llist_head active;
+	wait_queue_head_t waiting;
+};
+
+static DEFINE_PER_CPU(struct tx_task_t, tx_task_g);
+
+void siw_stop_tx_thread(int nr_cpu)
+{
+	kthread_stop(siw_tx_thread[nr_cpu]);
+	wake_up(&per_cpu(tx_task_g, nr_cpu).waiting);
+}
+
+int siw_run_sq(void *data)
+{
+	const int nr_cpu = (unsigned int)(long)data;
+	struct llist_node *active;
+	struct siw_qp *qp;
+	struct tx_task_t *tx_task = &per_cpu(tx_task_g, nr_cpu);
+
+	init_llist_head(&tx_task->active);
+	init_waitqueue_head(&tx_task->waiting);
+
+	pr_info("Started siw TX thread on CPU %u\n", nr_cpu);
+
+	while (1) {
+		struct llist_node *fifo_list = NULL;
+
+		wait_event_interruptible(tx_task->waiting,
+					 !llist_empty(&tx_task->active) ||
+						 kthread_should_stop());
+
+		if (kthread_should_stop())
+			break;
+
+		active = llist_del_all(&tx_task->active);
+		/*
+		 * llist_del_all returns a list with newest entry first.
+		 * Re-order list for fairness among QP's.
+		 */
+		while (active) {
+			struct llist_node *tmp = active;
+
+			active = llist_next(active);
+			tmp->next = fifo_list;
+			fifo_list = tmp;
+		}
+		while (fifo_list) {
+			qp = container_of(fifo_list, struct siw_qp, tx_list);
+			fifo_list = llist_next(fifo_list);
+			qp->tx_list.next = NULL;
+
+			siw_sq_resume(qp);
+		}
+	}
+	active = llist_del_all(&tx_task->active);
+	if (active) {
+		llist_for_each_entry(qp, active, tx_list) {
+			qp->tx_list.next = NULL;
+			siw_sq_resume(qp);
+		}
+	}
+	pr_info("Stopped siw TX thread on CPU %u\n", nr_cpu);
+
+	return 0;
+}
+
+int siw_sq_start(struct siw_qp *qp)
+{
+	if (tx_wqe(qp)->wr_status == SIW_WR_IDLE)
+		return 0;
+
+	if (unlikely(!cpu_online(qp->tx_cpu))) {
+		siw_put_tx_cpu(qp->tx_cpu);
+		qp->tx_cpu = siw_get_tx_cpu(qp->sdev);
+		if (qp->tx_cpu < 0) {
+			pr_warn("siw: no tx cpu available\n");
+
+			return -EIO;
+		}
+	}
+	siw_qp_get(qp);
+
+	llist_add(&qp->tx_list, &per_cpu(tx_task_g, qp->tx_cpu).active);
+
+	wake_up(&per_cpu(tx_task_g, qp->tx_cpu).waiting);
+
+	return 0;
+}