IB/qib: Improve SDMA performance

1. The code accepts chunks of messages, and splits the chunk into
   packets when converting packets into sdma queue entries.  Adjacent
   packets will use user buffer pages smartly to avoid pinning the
   same page multiple times.

2. Instead of discarding all the work when SDMA queue is full, the
   work is saved in a pending queue.  Whenever there are enough SDMA
   queue free entries, pending queue is directly put onto SDMA queue.

3. An interrupt handler is used to progress this pending queue.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: CQ Tang <cq.tang@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>

[ Fixed up sparse warnings.  - Roland ]

Signed-off-by: Roland Dreier <roland@purestorage.com>

5 files changed, 728 insertions, 227 deletions

diff --git a/drivers/infiniband/hw/qib/qib.h b/drivers/infiniband/hw/qib/qib.h
index 4a9af795b88f..ae3e4feca718 100644
--- a/drivers/infiniband/hw/qib/qib.h
+++ b/drivers/infiniband/hw/qib/qib.h
@@ -576,11 +576,13 @@ struct qib_pportdata {
 	/* read/write using lock */
 	spinlock_t            sdma_lock ____cacheline_aligned_in_smp;
 	struct list_head      sdma_activelist;
+	struct list_head      sdma_userpending;
 	u64                   sdma_descq_added;
 	u64                   sdma_descq_removed;
 	u16                   sdma_descq_tail;
 	u16                   sdma_descq_head;
 	u8                    sdma_generation;
+	u8                    sdma_intrequest;
 
 	struct tasklet_struct sdma_sw_clean_up_task
 		____cacheline_aligned_in_smp;
@@ -1326,6 +1328,8 @@ int qib_setup_sdma(struct qib_pportdata *);
 void qib_teardown_sdma(struct qib_pportdata *);
 void __qib_sdma_intr(struct qib_pportdata *);
 void qib_sdma_intr(struct qib_pportdata *);
+void qib_user_sdma_send_desc(struct qib_pportdata *dd,
+			struct list_head *pktlist);
 int qib_sdma_verbs_send(struct qib_pportdata *, struct qib_sge_state *,
 			u32, struct qib_verbs_txreq *);
 /* ppd->sdma_lock should be locked before calling this. */
diff --git a/drivers/infiniband/hw/qib/qib_common.h b/drivers/infiniband/hw/qib/qib_common.h
index 4f255b723ffd..5670ace27c63 100644
--- a/drivers/infiniband/hw/qib/qib_common.h
+++ b/drivers/infiniband/hw/qib/qib_common.h
@@ -279,7 +279,7 @@ struct qib_base_info {
  * may not be implemented; the user code must deal with this if it
  * cares, or it must abort after initialization reports the difference.
  */
-#define QIB_USER_SWMINOR 12
+#define QIB_USER_SWMINOR 13
 
 #define QIB_USER_SWVERSION ((QIB_USER_SWMAJOR << 16) | QIB_USER_SWMINOR)
 
@@ -701,7 +701,37 @@ struct qib_message_header {
 	__be32 bth[3];
 	/* fields below this point are in host byte order */
 	struct qib_header iph;
+	/* fields below are simplified, but should match PSM */
+	/* some are accessed by driver when packet spliting is needed */
 	__u8 sub_opcode;
+	__u8 flags;
+	__u16 commidx;
+	__u32 ack_seq_num;
+	__u8 flowid;
+	__u8 hdr_dlen;
+	__u16 mqhdr;
+	__u32 uwords[4];
+};
+
+/* sequence number bits for message */
+union qib_seqnum {
+	struct {
+		__u32 seq:11;
+		__u32 gen:8;
+		__u32 flow:5;
+	};
+	struct {
+		__u32 pkt:16;
+		__u32 msg:8;
+	};
+	__u32 val;
+};
+
+/* qib receiving-dma tid-session-member */
+struct qib_tid_session_member {
+	__u16 tid;
+	__u16 offset;
+	__u16 length;
 };
 
 /* IB - LRH header consts */
diff --git a/drivers/infiniband/hw/qib/qib_file_ops.c b/drivers/infiniband/hw/qib/qib_file_ops.c
index b51a51486cb8..275f247f9fca 100644
--- a/drivers/infiniband/hw/qib/qib_file_ops.c
+++ b/drivers/infiniband/hw/qib/qib_file_ops.c
@@ -1220,7 +1220,7 @@ static int qib_compatible_subctxts(int user_swmajor, int user_swminor)
 			return user_swminor == 3;
 		default:
 			/* >= 4 are compatible (or are expected to be) */
-			return user_swminor >= 4;
+			return user_swminor <= QIB_USER_SWMINOR;
 		}
 	}
 	/* make no promises yet for future major versions */
diff --git a/drivers/infiniband/hw/qib/qib_sdma.c b/drivers/infiniband/hw/qib/qib_sdma.c
index 9b5322d8cd5a..c6d6a54d2e19 100644
--- a/drivers/infiniband/hw/qib/qib_sdma.c
+++ b/drivers/infiniband/hw/qib/qib_sdma.c
@@ -423,8 +423,11 @@ void qib_sdma_intr(struct qib_pportdata *ppd)
 
 void __qib_sdma_intr(struct qib_pportdata *ppd)
 {
-	if (__qib_sdma_running(ppd))
+	if (__qib_sdma_running(ppd)) {
 		qib_sdma_make_progress(ppd);
+		if (!list_empty(&ppd->sdma_userpending))
+			qib_user_sdma_send_desc(ppd, &ppd->sdma_userpending);
+	}
 }
 
 int qib_setup_sdma(struct qib_pportdata *ppd)
@@ -452,6 +455,9 @@ int qib_setup_sdma(struct qib_pportdata *ppd)
 	ppd->sdma_descq_removed = 0;
 	ppd->sdma_descq_added = 0;
 
+	ppd->sdma_intrequest = 0;
+	INIT_LIST_HEAD(&ppd->sdma_userpending);
+
 	INIT_LIST_HEAD(&ppd->sdma_activelist);
 
 	tasklet_init(&ppd->sdma_sw_clean_up_task, sdma_sw_clean_up_task,
diff --git a/drivers/infiniband/hw/qib/qib_user_sdma.c b/drivers/infiniband/hw/qib/qib_user_sdma.c
index 82442085cbe6..d0a0ea0c14d6 100644
--- a/drivers/infiniband/hw/qib/qib_user_sdma.c
+++ b/drivers/infiniband/hw/qib/qib_user_sdma.c
@@ -53,20 +53,36 @@
 #define QIB_USER_SDMA_DRAIN_TIMEOUT 500
 
 struct qib_user_sdma_pkt {
-	u8 naddr;               /* dimension of addr (1..3) ... */
+	struct list_head list;  /* list element */
+
+	u8  tiddma;		/* if this is NEW tid-sdma */
+	u8  largepkt;		/* this is large pkt from kmalloc */
+	u16 frag_size;		/* frag size used by PSM */
+	u16 index;              /* last header index or push index */
+	u16 naddr;              /* dimension of addr (1..3) ... */
+	u16 addrlimit;		/* addr array size */
+	u16 tidsmidx;		/* current tidsm index */
+	u16 tidsmcount;		/* tidsm array item count */
+	u16 payload_size;	/* payload size so far for header */
+	u32 bytes_togo;		/* bytes for processing */
 	u32 counter;            /* sdma pkts queued counter for this entry */
+	struct qib_tid_session_member *tidsm;	/* tid session member array */
+	struct qib_user_sdma_queue *pq;	/* which pq this pkt belongs to */
 	u64 added;              /* global descq number of entries */
 
 	struct {
-		u32 offset;                     /* offset for kvaddr, addr */
-		u32 length;                     /* length in page */
-		u8  put_page;                   /* should we put_page? */
-		u8  dma_mapped;                 /* is page dma_mapped? */
+		u16 offset;                     /* offset for kvaddr, addr */
+		u16 length;                     /* length in page */
+		u16 first_desc;			/* first desc */
+		u16 last_desc;			/* last desc */
+		u16 put_page;                   /* should we put_page? */
+		u16 dma_mapped;                 /* is page dma_mapped? */
+		u16 dma_length;			/* for dma_unmap_page() */
+		u16 padding;
 		struct page *page;              /* may be NULL (coherent mem) */
 		void *kvaddr;                   /* FIXME: only for pio hack */
 		dma_addr_t addr;
 	} addr[4];   /* max pages, any more and we coalesce */
-	struct list_head list;  /* list element */
 };
 
 struct qib_user_sdma_queue {
@@ -77,6 +93,12 @@ struct qib_user_sdma_queue {
 	 */
 	struct list_head sent;
 
+	/*
+	 * Because above list will be accessed by both process and
+	 * signal handler, we need a spinlock for it.
+	 */
+	spinlock_t sent_lock ____cacheline_aligned_in_smp;
+
 	/* headers with expected length are allocated from here... */
 	char header_cache_name[64];
 	struct dma_pool *header_cache;
@@ -88,6 +110,12 @@ struct qib_user_sdma_queue {
 	/* as packets go on the queued queue, they are counted... */
 	u32 counter;
 	u32 sent_counter;
+	/* pending packets, not sending yet */
+	u32 num_pending;
+	/* sending packets, not complete yet */
+	u32 num_sending;
+	/* global descq number of entry of last sending packet */
+	u64 added;
 
 	/* dma page table */
 	struct rb_root dma_pages_root;
@@ -107,8 +135,12 @@ qib_user_sdma_queue_create(struct device *dev, int unit, int ctxt, int sctxt)
 
 	pq->counter = 0;
 	pq->sent_counter = 0;
-	INIT_LIST_HEAD(&pq->sent);
+	pq->num_pending = 0;
+	pq->num_sending = 0;
+	pq->added = 0;
 
+	INIT_LIST_HEAD(&pq->sent);
+	spin_lock_init(&pq->sent_lock);
 	mutex_init(&pq->lock);
 
 	snprintf(pq->pkt_slab_name, sizeof(pq->pkt_slab_name),
@@ -144,34 +176,310 @@ done:
 }
 
 static void qib_user_sdma_init_frag(struct qib_user_sdma_pkt *pkt,
-				    int i, size_t offset, size_t len,
-				    int put_page, int dma_mapped,
-				    struct page *page,
-				    void *kvaddr, dma_addr_t dma_addr)
+				    int i, u16 offset, u16 len,
+				    u16 first_desc, u16 last_desc,
+				    u16 put_page, u16 dma_mapped,
+				    struct page *page, void *kvaddr,
+				    dma_addr_t dma_addr, u16 dma_length)
 {
 	pkt->addr[i].offset = offset;
 	pkt->addr[i].length = len;
+	pkt->addr[i].first_desc = first_desc;
+	pkt->addr[i].last_desc = last_desc;
 	pkt->addr[i].put_page = put_page;
 	pkt->addr[i].dma_mapped = dma_mapped;
 	pkt->addr[i].page = page;
 	pkt->addr[i].kvaddr = kvaddr;
 	pkt->addr[i].addr = dma_addr;
+	pkt->addr[i].dma_length = dma_length;
 }
 
-static void qib_user_sdma_init_header(struct qib_user_sdma_pkt *pkt,
-				      u32 counter, size_t offset,
-				      size_t len, int dma_mapped,
-				      struct page *page,
-				      void *kvaddr, dma_addr_t dma_addr)
+static void *qib_user_sdma_alloc_header(struct qib_user_sdma_queue *pq,
+				size_t len, dma_addr_t *dma_addr)
 {
-	pkt->naddr = 1;
-	pkt->counter = counter;
-	qib_user_sdma_init_frag(pkt, 0, offset, len, 0, dma_mapped, page,
-				kvaddr, dma_addr);
+	void *hdr;
+
+	if (len == QIB_USER_SDMA_EXP_HEADER_LENGTH)
+		hdr = dma_pool_alloc(pq->header_cache, GFP_KERNEL,
+					     dma_addr);
+	else
+		hdr = NULL;
+
+	if (!hdr) {
+		hdr = kmalloc(len, GFP_KERNEL);
+		if (!hdr)
+			return NULL;
+
+		*dma_addr = 0;
+	}
+
+	return hdr;
+}
+
+static int qib_user_sdma_page_to_frags(const struct qib_devdata *dd,
+				       struct qib_user_sdma_queue *pq,
+				       struct qib_user_sdma_pkt *pkt,
+				       struct page *page, u16 put,
+				       u16 offset, u16 len, void *kvaddr)
+{
+	__le16 *pbc16;
+	void *pbcvaddr;
+	struct qib_message_header *hdr;
+	u16 newlen, pbclen, lastdesc, dma_mapped;
+	u32 vcto;
+	union qib_seqnum seqnum;
+	dma_addr_t pbcdaddr;
+	dma_addr_t dma_addr =
+		dma_map_page(&dd->pcidev->dev,
+			page, offset, len, DMA_TO_DEVICE);
+	int ret = 0;
+
+	if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
+		/*
+		 * dma mapping error, pkt has not managed
+		 * this page yet, return the page here so
+		 * the caller can ignore this page.
+		 */
+		if (put) {
+			put_page(page);
+		} else {
+			/* coalesce case */
+			kunmap(page);
+			__free_page(page);
+		}
+		ret = -ENOMEM;
+		goto done;
+	}
+	offset = 0;
+	dma_mapped = 1;
+
+
+next_fragment:
+
+	/*
+	 * In tid-sdma, the transfer length is restricted by
+	 * receiver side current tid page length.
+	 */
+	if (pkt->tiddma && len > pkt->tidsm[pkt->tidsmidx].length)
+		newlen = pkt->tidsm[pkt->tidsmidx].length;
+	else
+		newlen = len;
+
+	/*
+	 * Then the transfer length is restricted by MTU.
+	 * the last descriptor flag is determined by:
+	 * 1. the current packet is at frag size length.
+	 * 2. the current tid page is done if tid-sdma.
+	 * 3. there is no more byte togo if sdma.
+	 */
+	lastdesc = 0;
+	if ((pkt->payload_size + newlen) >= pkt->frag_size) {
+		newlen = pkt->frag_size - pkt->payload_size;
+		lastdesc = 1;
+	} else if (pkt->tiddma) {
+		if (newlen == pkt->tidsm[pkt->tidsmidx].length)
+			lastdesc = 1;
+	} else {
+		if (newlen == pkt->bytes_togo)
+			lastdesc = 1;
+	}
+
+	/* fill the next fragment in this page */
+	qib_user_sdma_init_frag(pkt, pkt->naddr, /* index */
+		offset, newlen,		/* offset, len */
+		0, lastdesc,		/* first last desc */
+		put, dma_mapped,	/* put page, dma mapped */
+		page, kvaddr,		/* struct page, virt addr */
+		dma_addr, len);		/* dma addr, dma length */
+	pkt->bytes_togo -= newlen;
+	pkt->payload_size += newlen;
+	pkt->naddr++;
+	if (pkt->naddr == pkt->addrlimit) {
+		ret = -EFAULT;
+		goto done;
+	}
+
+	/* If there is no more byte togo. (lastdesc==1) */
+	if (pkt->bytes_togo == 0) {
+		/* The packet is done, header is not dma mapped yet.
+		 * it should be from kmalloc */
+		if (!pkt->addr[pkt->index].addr) {
+			pkt->addr[pkt->index].addr =
+				dma_map_single(&dd->pcidev->dev,
+					pkt->addr[pkt->index].kvaddr,
+					pkt->addr[pkt->index].dma_length,
+					DMA_TO_DEVICE);
+			if (dma_mapping_error(&dd->pcidev->dev,
+					pkt->addr[pkt->index].addr)) {
+				ret = -ENOMEM;
+				goto done;
+			}
+			pkt->addr[pkt->index].dma_mapped = 1;
+		}
+
+		goto done;
+	}
+
+	/* If tid-sdma, advance tid info. */
+	if (pkt->tiddma) {
+		pkt->tidsm[pkt->tidsmidx].length -= newlen;
+		if (pkt->tidsm[pkt->tidsmidx].length) {
+			pkt->tidsm[pkt->tidsmidx].offset += newlen;
+		} else {
+			pkt->tidsmidx++;
+			if (pkt->tidsmidx == pkt->tidsmcount) {
+				ret = -EFAULT;
+				goto done;
+			}
+		}
+	}
+
+	/*
+	 * If this is NOT the last descriptor. (newlen==len)
+	 * the current packet is not done yet, but the current
+	 * send side page is done.
+	 */
+	if (lastdesc == 0)
+		goto done;
+
+	/*
+	 * If running this driver under PSM with message size
+	 * fitting into one transfer unit, it is not possible
+	 * to pass this line. otherwise, it is a buggggg.
+	 */
+
+	/*
+	 * Since the current packet is done, and there are more
+	 * bytes togo, we need to create a new sdma header, copying
+	 * from previous sdma header and modify both.
+	 */
+	pbclen = pkt->addr[pkt->index].length;
+	pbcvaddr = qib_user_sdma_alloc_header(pq, pbclen, &pbcdaddr);
+	if (!pbcvaddr) {
+		ret = -ENOMEM;
+		goto done;
+	}
+	/* Copy the previous sdma header to new sdma header */
+	pbc16 = (__le16 *)pkt->addr[pkt->index].kvaddr;
+	memcpy(pbcvaddr, pbc16, pbclen);
+
+	/* Modify the previous sdma header */
+	hdr = (struct qib_message_header *)&pbc16[4];
+
+	/* New pbc length */
+	pbc16[0] = cpu_to_le16(le16_to_cpu(pbc16[0])-(pkt->bytes_togo>>2));
+
+	/* New packet length */
+	hdr->lrh[2] = cpu_to_be16(le16_to_cpu(pbc16[0]));
+
+	if (pkt->tiddma) {
+		/* turn on the header suppression */
+		hdr->iph.pkt_flags =
+			cpu_to_le16(le16_to_cpu(hdr->iph.pkt_flags)|0x2);
+		/* turn off ACK_REQ: 0x04 and EXPECTED_DONE: 0x20 */
+		hdr->flags &= ~(0x04|0x20);
+	} else {
+		/* turn off extra bytes: 20-21 bits */
+		hdr->bth[0] = cpu_to_be32(be32_to_cpu(hdr->bth[0])&0xFFCFFFFF);
+		/* turn off ACK_REQ: 0x04 */
+		hdr->flags &= ~(0x04);
+	}
+
+	/* New kdeth checksum */
+	vcto = le32_to_cpu(hdr->iph.ver_ctxt_tid_offset);
+	hdr->iph.chksum = cpu_to_le16(QIB_LRH_BTH +
+		be16_to_cpu(hdr->lrh[2]) -
+		((vcto>>16)&0xFFFF) - (vcto&0xFFFF) -
+		le16_to_cpu(hdr->iph.pkt_flags));
+
+	/* The packet is done, header is not dma mapped yet.
+	 * it should be from kmalloc */
+	if (!pkt->addr[pkt->index].addr) {
+		pkt->addr[pkt->index].addr =
+			dma_map_single(&dd->pcidev->dev,
+				pkt->addr[pkt->index].kvaddr,
+				pkt->addr[pkt->index].dma_length,
+				DMA_TO_DEVICE);
+		if (dma_mapping_error(&dd->pcidev->dev,
+				pkt->addr[pkt->index].addr)) {
+			ret = -ENOMEM;
+			goto done;
+		}
+		pkt->addr[pkt->index].dma_mapped = 1;
+	}
+
+	/* Modify the new sdma header */
+	pbc16 = (__le16 *)pbcvaddr;
+	hdr = (struct qib_message_header *)&pbc16[4];
+
+	/* New pbc length */
+	pbc16[0] = cpu_to_le16(le16_to_cpu(pbc16[0])-(pkt->payload_size>>2));
+
+	/* New packet length */
+	hdr->lrh[2] = cpu_to_be16(le16_to_cpu(pbc16[0]));
+
+	if (pkt->tiddma) {
+		/* Set new tid and offset for new sdma header */
+		hdr->iph.ver_ctxt_tid_offset = cpu_to_le32(
+			(le32_to_cpu(hdr->iph.ver_ctxt_tid_offset)&0xFF000000) +
+			(pkt->tidsm[pkt->tidsmidx].tid<<QLOGIC_IB_I_TID_SHIFT) +
+			(pkt->tidsm[pkt->tidsmidx].offset>>2));
+	} else {
+		/* Middle protocol new packet offset */
+		hdr->uwords[2] += pkt->payload_size;
+	}
+
+	/* New kdeth checksum */
+	vcto = le32_to_cpu(hdr->iph.ver_ctxt_tid_offset);
+	hdr->iph.chksum = cpu_to_le16(QIB_LRH_BTH +
+		be16_to_cpu(hdr->lrh[2]) -
+		((vcto>>16)&0xFFFF) - (vcto&0xFFFF) -
+		le16_to_cpu(hdr->iph.pkt_flags));
+
+	/* Next sequence number in new sdma header */
+	seqnum.val = be32_to_cpu(hdr->bth[2]);
+	if (pkt->tiddma)
+		seqnum.seq++;
+	else
+		seqnum.pkt++;
+	hdr->bth[2] = cpu_to_be32(seqnum.val);
+
+	/* Init new sdma header. */
+	qib_user_sdma_init_frag(pkt, pkt->naddr, /* index */
+		0, pbclen,		/* offset, len */
+		1, 0,			/* first last desc */
+		0, 0,			/* put page, dma mapped */
+		NULL, pbcvaddr,		/* struct page, virt addr */
+		pbcdaddr, pbclen);	/* dma addr, dma length */
+	pkt->index = pkt->naddr;
+	pkt->payload_size = 0;
+	pkt->naddr++;
+	if (pkt->naddr == pkt->addrlimit) {
+		ret = -EFAULT;
+		goto done;
+	}
+
+	/* Prepare for next fragment in this page */
+	if (newlen != len) {
+		if (dma_mapped) {
+			put = 0;
+			dma_mapped = 0;
+			page = NULL;
+			kvaddr = NULL;
+		}
+		len -= newlen;
+		offset += newlen;
+
+		goto next_fragment;
+	}
+
+done:
+	return ret;
 }
 
 /* we've too many pages in the iovec, coalesce to a single page */
 static int qib_user_sdma_coalesce(const struct qib_devdata *dd,
+				  struct qib_user_sdma_queue *pq,
 				  struct qib_user_sdma_pkt *pkt,
 				  const struct iovec *iov,
 				  unsigned long niov)
@@ -182,7 +490,6 @@ static int qib_user_sdma_coalesce(const struct qib_devdata *dd,
 	char *mpage;
 	int i;
 	int len = 0;
-	dma_addr_t dma_addr;
 
 	if (!page) {
 		ret = -ENOMEM;
@@ -205,17 +512,8 @@ static int qib_user_sdma_coalesce(const struct qib_devdata *dd,
 		len += iov[i].iov_len;
 	}
 
-	dma_addr = dma_map_page(&dd->pcidev->dev, page, 0, len,
-				DMA_TO_DEVICE);
-	if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
-		ret = -ENOMEM;
-		goto free_unmap;
-	}
-
-	qib_user_sdma_init_frag(pkt, 1, 0, len, 0, 1, page, mpage_save,
-				dma_addr);
-	pkt->naddr = 2;
-
+	ret = qib_user_sdma_page_to_frags(dd, pq, pkt,
+			page, 0, 0, len, mpage_save);
 	goto done;
 
 free_unmap:
@@ -238,16 +536,6 @@ static int qib_user_sdma_num_pages(const struct iovec *iov)
 	return 1 + ((epage - spage) >> PAGE_SHIFT);
 }
 
-/*
- * Truncate length to page boundary.
- */
-static int qib_user_sdma_page_length(unsigned long addr, unsigned long len)
-{
-	const unsigned long offset = addr & ~PAGE_MASK;
-
-	return ((offset + len) > PAGE_SIZE) ? (PAGE_SIZE - offset) : len;
-}
-
 static void qib_user_sdma_free_pkt_frag(struct device *dev,
 					struct qib_user_sdma_queue *pq,
 					struct qib_user_sdma_pkt *pkt,
@@ -256,10 +544,11 @@ static void qib_user_sdma_free_pkt_frag(struct device *dev,
 	const int i = frag;
 
 	if (pkt->addr[i].page) {
+		/* only user data has page */
 		if (pkt->addr[i].dma_mapped)
 			dma_unmap_page(dev,
 				       pkt->addr[i].addr,
-				       pkt->addr[i].length,
+				       pkt->addr[i].dma_length,
 				       DMA_TO_DEVICE);
 
 		if (pkt->addr[i].kvaddr)
@@ -269,55 +558,81 @@ static void qib_user_sdma_free_pkt_frag(struct device *dev,
 			put_page(pkt->addr[i].page);
 		else
 			__free_page(pkt->addr[i].page);
-	} else if (pkt->addr[i].kvaddr)
-		/* free coherent mem from cache... */
-		dma_pool_free(pq->header_cache,
+	} else if (pkt->addr[i].kvaddr) {
+		/* for headers */
+		if (pkt->addr[i].dma_mapped) {
+			/* from kmalloc & dma mapped */
+			dma_unmap_single(dev,
+				       pkt->addr[i].addr,
+				       pkt->addr[i].dma_length,
+				       DMA_TO_DEVICE);
+			kfree(pkt->addr[i].kvaddr);
+		} else if (pkt->addr[i].addr) {
+			/* free coherent mem from cache... */
+			dma_pool_free(pq->header_cache,
 			      pkt->addr[i].kvaddr, pkt->addr[i].addr);
+		} else {
+			/* from kmalloc but not dma mapped */
+			kfree(pkt->addr[i].kvaddr);
+		}
+	}
 }
 
 /* return number of pages pinned... */
 static int qib_user_sdma_pin_pages(const struct qib_devdata *dd,
+				   struct qib_user_sdma_queue *pq,
 				   struct qib_user_sdma_pkt *pkt,
 				   unsigned long addr, int tlen, int npages)
 {
-	struct page *pages[2];
-	int j;
-	int ret;
-
-	ret = get_user_pages(current, current->mm, addr,
-			     npages, 0, 1, pages, NULL);
-
-	if (ret != npages) {
-		int i;
-
-		for (i = 0; i < ret; i++)
-			put_page(pages[i]);
-
-		ret = -ENOMEM;
-		goto done;
-	}
+	struct page *pages[8];
+	int i, j;
+	int ret = 0;
 
-	for (j = 0; j < npages; j++) {
-		/* map the pages... */
-		const int flen = qib_user_sdma_page_length(addr, tlen);
-		dma_addr_t dma_addr =
-			dma_map_page(&dd->pcidev->dev,
-				     pages[j], 0, flen, DMA_TO_DEVICE);
-		unsigned long fofs = addr & ~PAGE_MASK;
+	while (npages) {
+		if (npages > 8)
+			j = 8;
+		else
+			j = npages;
 
-		if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
+		ret = get_user_pages(current, current->mm, addr,
+			     j, 0, 1, pages, NULL);
+		if (ret != j) {
+			i = 0;
+			j = ret;
 			ret = -ENOMEM;
-			goto done;
+			goto free_pages;
 		}
 
-		qib_user_sdma_init_frag(pkt, pkt->naddr, fofs, flen, 1, 1,
-					pages[j], kmap(pages[j]), dma_addr);
+		for (i = 0; i < j; i++) {
+			/* map the pages... */
+			unsigned long fofs = addr & ~PAGE_MASK;
+			int flen = ((fofs + tlen) > PAGE_SIZE) ?
+				(PAGE_SIZE - fofs) : tlen;
+
+			ret = qib_user_sdma_page_to_frags(dd, pq, pkt,
+				pages[i], 1, fofs, flen, NULL);
+			if (ret < 0) {
+				/* current page has beed taken
+				 * care of inside above call.
+				 */
+				i++;
+				goto free_pages;
+			}
 
-		pkt->naddr++;
-		addr += flen;
-		tlen -= flen;
+			addr += flen;
+			tlen -= flen;
+		}
+
+		npages -= j;
 	}
 
+	goto done;
+
+	/* if error, return all pages not managed by pkt */
+free_pages:
+	while (i < j)
+		put_page(pages[i++]);
+
 done:
 	return ret;
 }
@@ -335,7 +650,7 @@ static int qib_user_sdma_pin_pkt(const struct qib_devdata *dd,
 		const int npages = qib_user_sdma_num_pages(iov + idx);
 		const unsigned long addr = (unsigned long) iov[idx].iov_base;
 
-		ret = qib_user_sdma_pin_pages(dd, pkt, addr,
+		ret = qib_user_sdma_pin_pages(dd, pq, pkt, addr,
 					      iov[idx].iov_len, npages);
 		if (ret < 0)
 			goto free_pkt;
@@ -344,9 +659,22 @@ static int qib_user_sdma_pin_pkt(const struct qib_devdata *dd,
 	goto done;
 
 free_pkt:
-	for (idx = 0; idx < pkt->naddr; idx++)
+	/* we need to ignore the first entry here */
+	for (idx = 1; idx < pkt->naddr; idx++)
 		qib_user_sdma_free_pkt_frag(&dd->pcidev->dev, pq, pkt, idx);
 
+	/* need to dma unmap the first entry, this is to restore to
+	 * the original state so that caller can free the memory in
+	 * error condition. Caller does not know if dma mapped or not*/
+	if (pkt->addr[0].dma_mapped) {
+		dma_unmap_single(&dd->pcidev->dev,
+		       pkt->addr[0].addr,
+		       pkt->addr[0].dma_length,
+		       DMA_TO_DEVICE);
+		pkt->addr[0].addr = 0;
+		pkt->addr[0].dma_mapped = 0;
+	}
+
 done:
 	return ret;
 }
@@ -359,8 +687,9 @@ static int qib_user_sdma_init_payload(const struct qib_devdata *dd,
 {
 	int ret = 0;
 
-	if (npages >= ARRAY_SIZE(pkt->addr))
-		ret = qib_user_sdma_coalesce(dd, pkt, iov, niov);
+	if (pkt->frag_size == pkt->bytes_togo &&
+			npages >= ARRAY_SIZE(pkt->addr))
+		ret = qib_user_sdma_coalesce(dd, pq, pkt, iov, niov);
 	else
 		ret = qib_user_sdma_pin_pkt(dd, pq, pkt, iov, niov);
 
@@ -380,7 +709,10 @@ static void qib_user_sdma_free_pkt_list(struct device *dev,
 		for (i = 0; i < pkt->naddr; i++)
 			qib_user_sdma_free_pkt_frag(dev, pq, pkt, i);
 
-		kmem_cache_free(pq->pkt_slab, pkt);
+		if (pkt->largepkt)
+			kfree(pkt);
+		else
+			kmem_cache_free(pq->pkt_slab, pkt);
 	}
 	INIT_LIST_HEAD(list);
 }
@@ -393,63 +725,48 @@ static void qib_user_sdma_free_pkt_list(struct device *dev,
  * as, if there is an error we clean it...
  */
 static int qib_user_sdma_queue_pkts(const struct qib_devdata *dd,
+				    struct qib_pportdata *ppd,
 				    struct qib_user_sdma_queue *pq,
-				    struct list_head *list,
 				    const struct iovec *iov,
 				    unsigned long niov,
-				    int maxpkts)
+				    struct list_head *list,
+				    int *maxpkts, int *ndesc)
 {
 	unsigned long idx = 0;
 	int ret = 0;
 	int npkts = 0;
-	struct page *page = NULL;
 	__le32 *pbc;
 	dma_addr_t dma_addr;
 	struct qib_user_sdma_pkt *pkt = NULL;
 	size_t len;
 	size_t nw;
 	u32 counter = pq->counter;
-	int dma_mapped = 0;
+	u16 frag_size;
 
-	while (idx < niov && npkts < maxpkts) {
+	while (idx < niov && npkts < *maxpkts) {
 		const unsigned long addr = (unsigned long) iov[idx].iov_base;
 		const unsigned long idx_save = idx;
 		unsigned pktnw;
 		unsigned pktnwc;
 		int nfrags = 0;
 		int npages = 0;
+		int bytes_togo = 0;
+		int tiddma = 0;
 		int cfur;
 
-		dma_mapped = 0;
 		len = iov[idx].iov_len;
 		nw = len >> 2;
-		page = NULL;
-
-		pkt = kmem_cache_alloc(pq->pkt_slab, GFP_KERNEL);
-		if (!pkt) {
-			ret = -ENOMEM;
-			goto free_list;
-		}
 
 		if (len < QIB_USER_SDMA_MIN_HEADER_LENGTH ||
 		    len > PAGE_SIZE || len & 3 || addr & 3) {
 			ret = -EINVAL;
-			goto free_pkt;
+			goto free_list;
 		}
 
-		if (len == QIB_USER_SDMA_EXP_HEADER_LENGTH)
-			pbc = dma_pool_alloc(pq->header_cache, GFP_KERNEL,
-					     &dma_addr);
-		else
-			pbc = NULL;
-
+		pbc = qib_user_sdma_alloc_header(pq, len, &dma_addr);
 		if (!pbc) {
-			page = alloc_page(GFP_KERNEL);
-			if (!page) {
-				ret = -ENOMEM;
-				goto free_pkt;
-			}
-			pbc = kmap(page);
+			ret = -ENOMEM;
+			goto free_list;
 		}
 
 		cfur = copy_from_user(pbc, iov[idx].iov_base, len);
@@ -474,8 +791,8 @@ static int qib_user_sdma_queue_pkts(const struct qib_devdata *dd,
 		 * we can verify that the packet is consistent with the
 		 * iovec lengths.
 		 */
-		pktnw = le32_to_cpu(*pbc) & QIB_PBC_LENGTH_MASK;
-		if (pktnw < pktnwc || pktnw > pktnwc + (PAGE_SIZE >> 2)) {
+		pktnw = le32_to_cpu(*pbc) & 0xFFFF;
+		if (pktnw < pktnwc) {
 			ret = -EINVAL;
 			goto free_pbc;
 		}
@@ -486,17 +803,14 @@ static int qib_user_sdma_queue_pkts(const struct qib_devdata *dd,
 			const unsigned long faddr =
 				(unsigned long) iov[idx].iov_base;
 
-			if (slen & 3 || faddr & 3 || !slen ||
-			    slen > PAGE_SIZE) {
+			if (slen & 3 || faddr & 3 || !slen) {
 				ret = -EINVAL;
 				goto free_pbc;
 			}
 
-			npages++;
-			if ((faddr & PAGE_MASK) !=
-			    ((faddr + slen - 1) & PAGE_MASK))
-				npages++;
+			npages += qib_user_sdma_num_pages(&iov[idx]);
 
+			bytes_togo += slen;
 			pktnwc += slen >> 2;
 			idx++;
 			nfrags++;
@@ -507,48 +821,139 @@ static int qib_user_sdma_queue_pkts(const struct qib_devdata *dd,
 			goto free_pbc;
 		}
 
-		if (page) {
-			dma_addr = dma_map_page(&dd->pcidev->dev,
-						page, 0, len, DMA_TO_DEVICE);
-			if (dma_mapping_error(&dd->pcidev->dev, dma_addr)) {
+		frag_size = ((le32_to_cpu(*pbc))>>16) & 0xFFFF;
+		if (((frag_size ? frag_size : bytes_togo) + len) >
+						ppd->ibmaxlen) {
+			ret = -EINVAL;
+			goto free_pbc;
+		}
+
+		if (frag_size) {
+			int pktsize, tidsmsize, n;
+
+			n = npages*((2*PAGE_SIZE/frag_size)+1);
+			pktsize = sizeof(*pkt) + sizeof(pkt->addr[0])*n;
+
+			/*
+			 * Determine if this is tid-sdma or just sdma.
+			 */
+			tiddma = (((le32_to_cpu(pbc[7])>>
+				QLOGIC_IB_I_TID_SHIFT)&
+				QLOGIC_IB_I_TID_MASK) !=
+				QLOGIC_IB_I_TID_MASK);
+
+			if (tiddma)
+				tidsmsize = iov[idx].iov_len;
+			else
+				tidsmsize = 0;
+
+			pkt = kmalloc(pktsize+tidsmsize, GFP_KERNEL);
+			if (!pkt) {
 				ret = -ENOMEM;
 				goto free_pbc;
 			}
+			pkt->largepkt = 1;
+			pkt->frag_size = frag_size;
+			pkt->addrlimit = n + ARRAY_SIZE(pkt->addr);
+
+			if (tiddma) {
+				char *tidsm = (char *)pkt + pktsize;
+				cfur = copy_from_user(tidsm,
+					iov[idx].iov_base, tidsmsize);
+				if (cfur) {
+					ret = -EFAULT;
+					goto free_pkt;
+				}
+				pkt->tidsm =
+					(struct qib_tid_session_member *)tidsm;
+				pkt->tidsmcount = tidsmsize/
+					sizeof(struct qib_tid_session_member);
+				pkt->tidsmidx = 0;
+				idx++;
+			}
 
-			dma_mapped = 1;
+			/*
+			 * pbc 'fill1' field is borrowed to pass frag size,
+			 * we need to clear it after picking frag size, the
+			 * hardware requires this field to be zero.
+			 */
+			*pbc = cpu_to_le32(le32_to_cpu(*pbc) & 0x0000FFFF);
+		} else {
+			pkt = kmem_cache_alloc(pq->pkt_slab, GFP_KERNEL);
+			if (!pkt) {
+				ret = -ENOMEM;
+				goto free_pbc;
+			}
+			pkt->largepkt = 0;
+			pkt->frag_size = bytes_togo;
+			pkt->addrlimit = ARRAY_SIZE(pkt->addr);
 		}
-
-		qib_user_sdma_init_header(pkt, counter, 0, len, dma_mapped,
-					  page, pbc, dma_addr);
+		pkt->bytes_togo = bytes_togo;
+		pkt->payload_size = 0;
+		pkt->counter = counter;
+		pkt->tiddma = tiddma;
+
+		/* setup the first header */
+		qib_user_sdma_init_frag(pkt, 0, /* index */
+			0, len,		/* offset, len */
+			1, 0,		/* first last desc */
+			0, 0,		/* put page, dma mapped */
+			NULL, pbc,	/* struct page, virt addr */
+			dma_addr, len);	/* dma addr, dma length */
+		pkt->index = 0;
+		pkt->naddr = 1;
 
 		if (nfrags) {
 			ret = qib_user_sdma_init_payload(dd, pq, pkt,
 							 iov + idx_save + 1,
 							 nfrags, npages);
 			if (ret < 0)
-				goto free_pbc_dma;
+				goto free_pkt;
+		} else {
+			/* since there is no payload, mark the
+			 * header as the last desc. */
+			pkt->addr[0].last_desc = 1;
+
+			if (dma_addr == 0) {
+				/*
+				 * the header is not dma mapped yet.
+				 * it should be from kmalloc.
+				 */
+				dma_addr = dma_map_single(&dd->pcidev->dev,
+					pbc, len, DMA_TO_DEVICE);
+				if (dma_mapping_error(&dd->pcidev->dev,
+								dma_addr)) {
+					ret = -ENOMEM;
+					goto free_pkt;
+				}
+				pkt->addr[0].addr = dma_addr;
+				pkt->addr[0].dma_mapped = 1;
+			}
 		}
 
 		counter++;
 		npkts++;
+		pkt->pq = pq;
+		pkt->index = 0; /* reset index for push on hw */
+		*ndesc += pkt->naddr;
 
 		list_add_tail(&pkt->list, list);
 	}
 
+	*maxpkts = npkts;
 	ret = idx;
 	goto done;
 
-free_pbc_dma:
-	if (dma_mapped)
-		dma_unmap_page(&dd->pcidev->dev, dma_addr, len, DMA_TO_DEVICE);
+free_pkt:
+	if (pkt->largepkt)
+		kfree(pkt);
+	else
+		kmem_cache_free(pq->pkt_slab, pkt);
 free_pbc:
-	if (page) {
-		kunmap(page);
-		__free_page(page);
-	} else
+	if (dma_addr)
 		dma_pool_free(pq->header_cache, pbc, dma_addr);
-free_pkt:
-	kmem_cache_free(pq->pkt_slab, pkt);
+	else
+		kfree(pbc);
 free_list:
 	qib_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, list);
 done:
@@ -569,10 +974,20 @@ static int qib_user_sdma_queue_clean(struct qib_pportdata *ppd,
 	struct list_head free_list;
 	struct qib_user_sdma_pkt *pkt;
 	struct qib_user_sdma_pkt *pkt_prev;
+	unsigned long flags;
 	int ret = 0;
 
+	if (!pq->num_sending)
+		return 0;
+
 	INIT_LIST_HEAD(&free_list);
 
+	/*
+	 * We need this spin lock here because interrupt handler
+	 * might modify this list in qib_user_sdma_send_desc(), also
+	 * we can not get interrupted, otherwise it is a deadlock.
+	 */
+	spin_lock_irqsave(&pq->sent_lock, flags);
 	list_for_each_entry_safe(pkt, pkt_prev, &pq->sent, list) {
 		s64 descd = ppd->sdma_descq_removed - pkt->added;
 
@@ -583,7 +998,9 @@ static int qib_user_sdma_queue_clean(struct qib_pportdata *ppd,
 
 		/* one more packet cleaned */
 		ret++;
+		pq->num_sending--;
 	}
+	spin_unlock_irqrestore(&pq->sent_lock, flags);
 
 	if (!list_empty(&free_list)) {
 		u32 counter;
@@ -627,6 +1044,7 @@ void qib_user_sdma_queue_drain(struct qib_pportdata *ppd,
 			       struct qib_user_sdma_queue *pq)
 {
 	struct qib_devdata *dd = ppd->dd;
+	unsigned long flags;
 	int i;
 
 	if (!pq)
@@ -634,7 +1052,7 @@ void qib_user_sdma_queue_drain(struct qib_pportdata *ppd,
 
 	for (i = 0; i < QIB_USER_SDMA_DRAIN_TIMEOUT; i++) {
 		mutex_lock(&pq->lock);
-		if (list_empty(&pq->sent)) {
+		if (!pq->num_pending && !pq->num_sending) {
 			mutex_unlock(&pq->lock);
 			break;
 		}
@@ -644,29 +1062,44 @@ void qib_user_sdma_queue_drain(struct qib_pportdata *ppd,
 		msleep(10);
 	}
 
-	if (!list_empty(&pq->sent)) {
+	if (pq->num_pending || pq->num_sending) {
+		struct qib_user_sdma_pkt *pkt;
+		struct qib_user_sdma_pkt *pkt_prev;
 		struct list_head free_list;
 
+		mutex_lock(&pq->lock);
+		spin_lock_irqsave(&ppd->sdma_lock, flags);
+		/*
+		 * Since we hold sdma_lock, it is safe without sent_lock.
+		 */
+		if (pq->num_pending) {
+			list_for_each_entry_safe(pkt, pkt_prev,
+					&ppd->sdma_userpending, list) {
+				if (pkt->pq == pq) {
+					list_move_tail(&pkt->list, &pq->sent);
+					pq->num_pending--;
+					pq->num_sending++;
+				}
+			}
+		}
+		spin_unlock_irqrestore(&ppd->sdma_lock, flags);
+
 		qib_dev_err(dd, "user sdma lists not empty: forcing!\n");
 		INIT_LIST_HEAD(&free_list);
-		mutex_lock(&pq->lock);
 		list_splice_init(&pq->sent, &free_list);
+		pq->num_sending = 0;
 		qib_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &free_list);
 		mutex_unlock(&pq->lock);
 	}
 }
 
-static inline __le64 qib_sdma_make_desc0(struct qib_pportdata *ppd,
+static inline __le64 qib_sdma_make_desc0(u8 gen,
 					 u64 addr, u64 dwlen, u64 dwoffset)
 {
-	u8 tmpgen;
-
-	tmpgen = ppd->sdma_generation;
-
 	return cpu_to_le64(/* SDmaPhyAddr[31:0] */
 			   ((addr & 0xfffffffcULL) << 32) |
 			   /* SDmaGeneration[1:0] */
-			   ((tmpgen & 3ULL) << 30) |
+			   ((gen & 3ULL) << 30) |
 			   /* SDmaDwordCount[10:0] */
 			   ((dwlen & 0x7ffULL) << 16) |
 			   /* SDmaBufOffset[12:2] */
@@ -692,7 +1125,7 @@ static inline __le64 qib_sdma_make_desc1(u64 addr)
 
 static void qib_user_sdma_send_frag(struct qib_pportdata *ppd,
 				    struct qib_user_sdma_pkt *pkt, int idx,
-				    unsigned ofs, u16 tail)
+				    unsigned ofs, u16 tail, u8 gen)
 {
 	const u64 addr = (u64) pkt->addr[idx].addr +
 		(u64) pkt->addr[idx].offset;
@@ -702,104 +1135,132 @@ static void qib_user_sdma_send_frag(struct qib_pportdata *ppd,
 
 	descqp = &ppd->sdma_descq[tail].qw[0];
 
-	descq0 = qib_sdma_make_desc0(ppd, addr, dwlen, ofs);
-	if (idx == 0)
+	descq0 = qib_sdma_make_desc0(gen, addr, dwlen, ofs);
+	if (pkt->addr[idx].first_desc)
 		descq0 = qib_sdma_make_first_desc0(descq0);
-	if (idx == pkt->naddr - 1)
+	if (pkt->addr[idx].last_desc) {
 		descq0 = qib_sdma_make_last_desc0(descq0);
+		if (ppd->sdma_intrequest) {
+			descq0 |= cpu_to_le64(1ULL << 15);
+			ppd->sdma_intrequest = 0;
+		}
+	}
 
 	descqp[0] = descq0;
 	descqp[1] = qib_sdma_make_desc1(addr);
 }
 
-/* pq->lock must be held, get packets on the wire... */
-static int qib_user_sdma_push_pkts(struct qib_pportdata *ppd,
-				   struct qib_user_sdma_queue *pq,
-				   struct list_head *pktlist)
+void qib_user_sdma_send_desc(struct qib_pportdata *ppd,
+				struct list_head *pktlist)
 {
 	struct qib_devdata *dd = ppd->dd;
-	int ret = 0;
-	unsigned long flags;
-	u16 tail;
-	u8 generation;
-	u64 descq_added;
-
-	if (list_empty(pktlist))
-		return 0;
+	u16 nfree, nsent;
+	u16 tail, tail_c;
+	u8 gen, gen_c;
 
-	if (unlikely(!(ppd->lflags & QIBL_LINKACTIVE)))
-		return -ECOMM;
-
-	spin_lock_irqsave(&ppd->sdma_lock, flags);
-
-	/* keep a copy for restoring purposes in case of problems */
-	generation = ppd->sdma_generation;
-	descq_added = ppd->sdma_descq_added;
-
-	if (unlikely(!__qib_sdma_running(ppd))) {
-		ret = -ECOMM;
-		goto unlock;
-	}
+	nfree = qib_sdma_descq_freecnt(ppd);
+	if (!nfree)
+		return;
 
-	tail = ppd->sdma_descq_tail;
+retry:
+	nsent = 0;
+	tail_c = tail = ppd->sdma_descq_tail;
+	gen_c = gen = ppd->sdma_generation;
 	while (!list_empty(pktlist)) {
 		struct qib_user_sdma_pkt *pkt =
 			list_entry(pktlist->next, struct qib_user_sdma_pkt,
 				   list);
-		int i;
+		int i, j, c = 0;
 		unsigned ofs = 0;
 		u16 dtail = tail;
 
-		if (pkt->naddr > qib_sdma_descq_freecnt(ppd))
-			goto unlock_check_tail;
-
-		for (i = 0; i < pkt->naddr; i++) {
-			qib_user_sdma_send_frag(ppd, pkt, i, ofs, tail);
+		for (i = pkt->index; i < pkt->naddr && nfree; i++) {
+			qib_user_sdma_send_frag(ppd, pkt, i, ofs, tail, gen);
 			ofs += pkt->addr[i].length >> 2;
 
 			if (++tail == ppd->sdma_descq_cnt) {
 				tail = 0;
-				++ppd->sdma_generation;
+				++gen;
+				ppd->sdma_intrequest = 1;
+			} else if (tail == (ppd->sdma_descq_cnt>>1)) {
+				ppd->sdma_intrequest = 1;
 			}
-		}
+			nfree--;
+			if (pkt->addr[i].last_desc == 0)
+				continue;
 
-		if ((ofs << 2) > ppd->ibmaxlen) {
-			ret = -EMSGSIZE;
-			goto unlock;
-		}
-
-		/*
-		 * If the packet is >= 2KB mtu equivalent, we have to use
-		 * the large buffers, and have to mark each descriptor as
-		 * part of a large buffer packet.
-		 */
-		if (ofs > dd->piosize2kmax_dwords) {
-			for (i = 0; i < pkt->naddr; i++) {
-				ppd->sdma_descq[dtail].qw[0] |=
-					cpu_to_le64(1ULL << 14);
-				if (++dtail == ppd->sdma_descq_cnt)
-					dtail = 0;
+			/*
+			 * If the packet is >= 2KB mtu equivalent, we
+			 * have to use the large buffers, and have to
+			 * mark each descriptor as part of a large
+			 * buffer packet.
+			 */
+			if (ofs > dd->piosize2kmax_dwords) {
+				for (j = pkt->index; j <= i; j++) {
+					ppd->sdma_descq[dtail].qw[0] |=
+						cpu_to_le64(1ULL << 14);
+					if (++dtail == ppd->sdma_descq_cnt)
+						dtail = 0;
+				}
 			}
+			c += i + 1 - pkt->index;
+			pkt->index = i + 1; /* index for next first */
+			tail_c = dtail = tail;
+			gen_c = gen;
+			ofs = 0;  /* reset for next packet */
 		}
 
-		ppd->sdma_descq_added += pkt->naddr;
-		pkt->added = ppd->sdma_descq_added;
-		list_move_tail(&pkt->list, &pq->sent);
-		ret++;
+		ppd->sdma_descq_added += c;
+		nsent += c;
+		if (pkt->index == pkt->naddr) {
+			pkt->added = ppd->sdma_descq_added;
+			pkt->pq->added = pkt->added;
+			pkt->pq->num_pending--;
+			spin_lock(&pkt->pq->sent_lock);
+			pkt->pq->num_sending++;
+			list_move_tail(&pkt->list, &pkt->pq->sent);
+			spin_unlock(&pkt->pq->sent_lock);
+		}
+		if (!nfree || (nsent<<2) > ppd->sdma_descq_cnt)
+			break;
 	}
 
-unlock_check_tail:
 	/* advance the tail on the chip if necessary */
-	if (ppd->sdma_descq_tail != tail)
-		dd->f_sdma_update_tail(ppd, tail);
+	if (ppd->sdma_descq_tail != tail_c) {
+		ppd->sdma_generation = gen_c;
+		dd->f_sdma_update_tail(ppd, tail_c);
+	}
 
-unlock:
-	if (unlikely(ret < 0)) {
-		ppd->sdma_generation = generation;
-		ppd->sdma_descq_added = descq_added;
+	if (nfree && !list_empty(pktlist))
+		goto retry;
+
+	return;
+}
+
+/* pq->lock must be held, get packets on the wire... */
+static int qib_user_sdma_push_pkts(struct qib_pportdata *ppd,
+				 struct qib_user_sdma_queue *pq,
+				 struct list_head *pktlist, int count)
+{
+	int ret = 0;
+	unsigned long flags;
+
+	if (unlikely(!(ppd->lflags & QIBL_LINKACTIVE)))
+		return -ECOMM;
+
+	spin_lock_irqsave(&ppd->sdma_lock, flags);
+
+	if (unlikely(!__qib_sdma_running(ppd))) {
+		ret = -ECOMM;
+		goto unlock;
 	}
-	spin_unlock_irqrestore(&ppd->sdma_lock, flags);
 
+	pq->num_pending += count;
+	list_splice_tail_init(pktlist, &ppd->sdma_userpending);
+	qib_user_sdma_send_desc(ppd, &ppd->sdma_userpending);
+
+unlock:
+	spin_unlock_irqrestore(&ppd->sdma_lock, flags);
 	return ret;
 }
 
@@ -822,19 +1283,23 @@ int qib_user_sdma_writev(struct qib_ctxtdata *rcd,
 	if (!qib_sdma_running(ppd))
 		goto done_unlock;
 
-	if (ppd->sdma_descq_added != ppd->sdma_descq_removed) {
+	/* if I have packets not complete yet */
+	if (pq->added > ppd->sdma_descq_removed)
 		qib_user_sdma_hwqueue_clean(ppd);
+	/* if I have complete packets to be freed */
+	if (pq->num_sending)
 		qib_user_sdma_queue_clean(ppd, pq);
-	}
 
 	while (dim) {
-		const int mxp = 8;
+		int mxp = 8;
+		int ndesc = 0;
 
 		down_write(&current->mm->mmap_sem);
-		ret = qib_user_sdma_queue_pkts(dd, pq, &list, iov, dim, mxp);
+		ret = qib_user_sdma_queue_pkts(dd, ppd, pq,
+				iov, dim, &list, &mxp, &ndesc);
 		up_write(&current->mm->mmap_sem);
 
-		if (ret <= 0)
+		if (ret < 0)
 			goto done_unlock;
 		else {
 			dim -= ret;
@@ -844,24 +1309,20 @@ int qib_user_sdma_writev(struct qib_ctxtdata *rcd,
 		/* force packets onto the sdma hw queue... */
 		if (!list_empty(&list)) {
 			/*
-			 * Lazily clean hw queue.  the 4 is a guess of about
-			 * how many sdma descriptors a packet will take (it
-			 * doesn't have to be perfect).
+			 * Lazily clean hw queue.
 			 */
-			if (qib_sdma_descq_freecnt(ppd) < ret * 4) {
+			if (qib_sdma_descq_freecnt(ppd) < ndesc) {
 				qib_user_sdma_hwqueue_clean(ppd);
-				qib_user_sdma_queue_clean(ppd, pq);
+				if (pq->num_sending)
+					qib_user_sdma_queue_clean(ppd, pq);
 			}
 
-			ret = qib_user_sdma_push_pkts(ppd, pq, &list);
+			ret = qib_user_sdma_push_pkts(ppd, pq, &list, mxp);
 			if (ret < 0)
 				goto done_unlock;
 			else {
-				npkts += ret;
-				pq->counter += ret;
-
-				if (!list_empty(&list))
-					goto done_unlock;
+				npkts += mxp;
+				pq->counter += mxp;
 			}
 		}
 	}