IB/ipath: User mode send DMA
authorArthur Jones <arthur.jones@qlogic.com>
Thu, 17 Apr 2008 04:09:31 +0000 (21:09 -0700)
committerRoland Dreier <rolandd@cisco.com>
Thu, 17 Apr 2008 04:09:31 +0000 (21:09 -0700)
A new file which allows the IBA7220 send DMA engine to be used from
userland.  The routines here are not linked in yet, that will happen in
a follow-on patch...

Signed-off-by: Arthur Jones <arthur.jones@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/ipath/ipath_user_sdma.c [new file with mode: 0644]

diff --git a/drivers/infiniband/hw/ipath/ipath_user_sdma.c b/drivers/infiniband/hw/ipath/ipath_user_sdma.c
new file mode 100644 (file)
index 0000000..86e0169
--- /dev/null
@@ -0,0 +1,879 @@
+/*
+ * Copyright (c) 2007, 2008 QLogic Corporation. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include <linux/mm.h>
+#include <linux/types.h>
+#include <linux/device.h>
+#include <linux/dmapool.h>
+#include <linux/slab.h>
+#include <linux/list.h>
+#include <linux/highmem.h>
+#include <linux/io.h>
+#include <linux/uio.h>
+#include <linux/rbtree.h>
+#include <linux/spinlock.h>
+#include <linux/delay.h>
+
+#include "ipath_kernel.h"
+#include "ipath_user_sdma.h"
+
+/* minimum size of header */
+#define IPATH_USER_SDMA_MIN_HEADER_LENGTH      64
+/* expected size of headers (for dma_pool) */
+#define IPATH_USER_SDMA_EXP_HEADER_LENGTH      64
+/* length mask in PBC (lower 11 bits) */
+#define IPATH_PBC_LENGTH_MASK                  ((1 << 11) - 1)
+
+struct ipath_user_sdma_pkt {
+       u8 naddr;               /* dimension of addr (1..3) ... */
+       u32 counter;            /* sdma pkts queued counter for this entry */
+       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? */
+               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 ipath_user_sdma_queue {
+       /*
+        * pkts sent to dma engine are queued on this
+        * list head.  the type of the elements of this
+        * list are struct ipath_user_sdma_pkt...
+        */
+       struct list_head sent;
+
+       /* headers with expected length are allocated from here... */
+       char header_cache_name[64];
+       struct dma_pool *header_cache;
+
+       /* packets are allocated from the slab cache... */
+       char pkt_slab_name[64];
+       struct kmem_cache *pkt_slab;
+
+       /* as packets go on the queued queue, they are counted... */
+       u32 counter;
+       u32 sent_counter;
+
+       /* dma page table */
+       struct rb_root dma_pages_root;
+
+       /* protect everything above... */
+       struct mutex lock;
+};
+
+struct ipath_user_sdma_queue *
+ipath_user_sdma_queue_create(struct device *dev, int unit, int port, int sport)
+{
+       struct ipath_user_sdma_queue *pq =
+               kmalloc(sizeof(struct ipath_user_sdma_queue), GFP_KERNEL);
+
+       if (!pq)
+               goto done;
+
+       pq->counter = 0;
+       pq->sent_counter = 0;
+       INIT_LIST_HEAD(&pq->sent);
+
+       mutex_init(&pq->lock);
+
+       snprintf(pq->pkt_slab_name, sizeof(pq->pkt_slab_name),
+                "ipath-user-sdma-pkts-%u-%02u.%02u", unit, port, sport);
+       pq->pkt_slab = kmem_cache_create(pq->pkt_slab_name,
+                                        sizeof(struct ipath_user_sdma_pkt),
+                                        0, 0, NULL);
+
+       if (!pq->pkt_slab)
+               goto err_kfree;
+
+       snprintf(pq->header_cache_name, sizeof(pq->header_cache_name),
+                "ipath-user-sdma-headers-%u-%02u.%02u", unit, port, sport);
+       pq->header_cache = dma_pool_create(pq->header_cache_name,
+                                          dev,
+                                          IPATH_USER_SDMA_EXP_HEADER_LENGTH,
+                                          4, 0);
+       if (!pq->header_cache)
+               goto err_slab;
+
+       pq->dma_pages_root = RB_ROOT;
+
+       goto done;
+
+err_slab:
+       kmem_cache_destroy(pq->pkt_slab);
+err_kfree:
+       kfree(pq);
+       pq = NULL;
+
+done:
+       return pq;
+}
+
+static void ipath_user_sdma_init_frag(struct ipath_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)
+{
+       pkt->addr[i].offset = offset;
+       pkt->addr[i].length = len;
+       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;
+}
+
+static void ipath_user_sdma_init_header(struct ipath_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)
+{
+       pkt->naddr = 1;
+       pkt->counter = counter;
+       ipath_user_sdma_init_frag(pkt, 0, offset, len, 0, dma_mapped, page,
+                                 kvaddr, dma_addr);
+}
+
+/* we've too many pages in the iovec, coalesce to a single page */
+static int ipath_user_sdma_coalesce(const struct ipath_devdata *dd,
+                                   struct ipath_user_sdma_pkt *pkt,
+                                   const struct iovec *iov,
+                                   unsigned long niov) {
+       int ret = 0;
+       struct page *page = alloc_page(GFP_KERNEL);
+       void *mpage_save;
+       char *mpage;
+       int i;
+       int len = 0;
+       dma_addr_t dma_addr;
+
+       if (!page) {
+               ret = -ENOMEM;
+               goto done;
+       }
+
+       mpage = kmap(page);
+       mpage_save = mpage;
+       for (i = 0; i < niov; i++) {
+               int cfur;
+
+               cfur = copy_from_user(mpage,
+                                     iov[i].iov_base, iov[i].iov_len);
+               if (cfur) {
+                       ret = -EFAULT;
+                       goto free_unmap;
+               }
+
+               mpage += iov[i].iov_len;
+               len += iov[i].iov_len;
+       }
+
+       dma_addr = dma_map_page(&dd->pcidev->dev, page, 0, len,
+                               DMA_TO_DEVICE);
+       if (dma_mapping_error(dma_addr)) {
+               ret = -ENOMEM;
+               goto free_unmap;
+       }
+
+       ipath_user_sdma_init_frag(pkt, 1, 0, len, 0, 1, page, mpage_save,
+                                 dma_addr);
+       pkt->naddr = 2;
+
+       goto done;
+
+free_unmap:
+       kunmap(page);
+       __free_page(page);
+done:
+       return ret;
+}
+
+/* how many pages in this iovec element? */
+static int ipath_user_sdma_num_pages(const struct iovec *iov)
+{
+       const unsigned long addr  = (unsigned long) iov->iov_base;
+       const unsigned long  len  = iov->iov_len;
+       const unsigned long spage = addr & PAGE_MASK;
+       const unsigned long epage = (addr + len - 1) & PAGE_MASK;
+
+       return 1 + ((epage - spage) >> PAGE_SHIFT);
+}
+
+/* truncate length to page boundry */
+static int ipath_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 ipath_user_sdma_free_pkt_frag(struct device *dev,
+                                         struct ipath_user_sdma_queue *pq,
+                                         struct ipath_user_sdma_pkt *pkt,
+                                         int frag)
+{
+       const int i = frag;
+
+       if (pkt->addr[i].page) {
+               if (pkt->addr[i].dma_mapped)
+                       dma_unmap_page(dev,
+                                      pkt->addr[i].addr,
+                                      pkt->addr[i].length,
+                                      DMA_TO_DEVICE);
+
+               if (pkt->addr[i].kvaddr)
+                       kunmap(pkt->addr[i].page);
+
+               if (pkt->addr[i].put_page)
+                       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,
+                             pkt->addr[i].kvaddr, pkt->addr[i].addr);
+}
+
+/* return number of pages pinned... */
+static int ipath_user_sdma_pin_pages(const struct ipath_devdata *dd,
+                                    struct ipath_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;
+       }
+
+       for (j = 0; j < npages; j++) {
+               /* map the pages... */
+               const int flen =
+                       ipath_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;
+
+               if (dma_mapping_error(dma_addr)) {
+                       ret = -ENOMEM;
+                       goto done;
+               }
+
+               ipath_user_sdma_init_frag(pkt, pkt->naddr, fofs, flen, 1, 1,
+                                         pages[j], kmap(pages[j]),
+                                         dma_addr);
+
+               pkt->naddr++;
+               addr += flen;
+               tlen -= flen;
+       }
+
+done:
+       return ret;
+}
+
+static int ipath_user_sdma_pin_pkt(const struct ipath_devdata *dd,
+                                  struct ipath_user_sdma_queue *pq,
+                                  struct ipath_user_sdma_pkt *pkt,
+                                  const struct iovec *iov,
+                                  unsigned long niov)
+{
+       int ret = 0;
+       unsigned long idx;
+
+       for (idx = 0; idx < niov; idx++) {
+               const int npages = ipath_user_sdma_num_pages(iov + idx);
+               const unsigned long addr = (unsigned long) iov[idx].iov_base;
+
+               ret = ipath_user_sdma_pin_pages(dd, pkt,
+                                               addr, iov[idx].iov_len,
+                                               npages);
+               if (ret < 0)
+                       goto free_pkt;
+       }
+
+       goto done;
+
+free_pkt:
+       for (idx = 0; idx < pkt->naddr; idx++)
+               ipath_user_sdma_free_pkt_frag(&dd->pcidev->dev, pq, pkt, idx);
+
+done:
+       return ret;
+}
+
+static int ipath_user_sdma_init_payload(const struct ipath_devdata *dd,
+                                       struct ipath_user_sdma_queue *pq,
+                                       struct ipath_user_sdma_pkt *pkt,
+                                       const struct iovec *iov,
+                                       unsigned long niov, int npages)
+{
+       int ret = 0;
+
+       if (npages >= ARRAY_SIZE(pkt->addr))
+               ret = ipath_user_sdma_coalesce(dd, pkt, iov, niov);
+       else
+               ret = ipath_user_sdma_pin_pkt(dd, pq, pkt, iov, niov);
+
+       return ret;
+}
+
+/* free a packet list -- return counter value of last packet */
+static void ipath_user_sdma_free_pkt_list(struct device *dev,
+                                         struct ipath_user_sdma_queue *pq,
+                                         struct list_head *list)
+{
+       struct ipath_user_sdma_pkt *pkt, *pkt_next;
+
+       list_for_each_entry_safe(pkt, pkt_next, list, list) {
+               int i;
+
+               for (i = 0; i < pkt->naddr; i++)
+                       ipath_user_sdma_free_pkt_frag(dev, pq, pkt, i);
+
+               kmem_cache_free(pq->pkt_slab, pkt);
+       }
+}
+
+/*
+ * copy headers, coalesce etc -- pq->lock must be held
+ *
+ * we queue all the packets to list, returning the
+ * number of bytes total.  list must be empty initially,
+ * as, if there is an error we clean it...
+ */
+static int ipath_user_sdma_queue_pkts(const struct ipath_devdata *dd,
+                                     struct ipath_user_sdma_queue *pq,
+                                     struct list_head *list,
+                                     const struct iovec *iov,
+                                     unsigned long niov,
+                                     int maxpkts)
+{
+       unsigned long idx = 0;
+       int ret = 0;
+       int npkts = 0;
+       struct page *page = NULL;
+       __le32 *pbc;
+       dma_addr_t dma_addr;
+       struct ipath_user_sdma_pkt *pkt = NULL;
+       size_t len;
+       size_t nw;
+       u32 counter = pq->counter;
+       int dma_mapped = 0;
+
+       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 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 < IPATH_USER_SDMA_MIN_HEADER_LENGTH ||
+                   len > PAGE_SIZE || len & 3 || addr & 3) {
+                       ret = -EINVAL;
+                       goto free_pkt;
+               }
+
+               if (len == IPATH_USER_SDMA_EXP_HEADER_LENGTH)
+                       pbc = dma_pool_alloc(pq->header_cache, GFP_KERNEL,
+                                            &dma_addr);
+               else
+                       pbc = NULL;
+
+               if (!pbc) {
+                       page = alloc_page(GFP_KERNEL);
+                       if (!page) {
+                               ret = -ENOMEM;
+                               goto free_pkt;
+                       }
+                       pbc = kmap(page);
+               }
+
+               cfur = copy_from_user(pbc, iov[idx].iov_base, len);
+               if (cfur) {
+                       ret = -EFAULT;
+                       goto free_pbc;
+               }
+
+               /*
+                * this assignment is a bit strange.  it's because the
+                * the pbc counts the number of 32 bit words in the full
+                * packet _except_ the first word of the pbc itself...
+                */
+               pktnwc = nw - 1;
+
+               /*
+                * pktnw computation yields the number of 32 bit words
+                * that the caller has indicated in the PBC.  note that
+                * this is one less than the total number of words that
+                * goes to the send DMA engine as the first 32 bit word
+                * of the PBC itself is not counted.  Armed with this count,
+                * we can verify that the packet is consistent with the
+                * iovec lengths.
+                */
+               pktnw = le32_to_cpu(*pbc) & IPATH_PBC_LENGTH_MASK;
+               if (pktnw < pktnwc || pktnw > pktnwc + (PAGE_SIZE >> 2)) {
+                       ret = -EINVAL;
+                       goto free_pbc;
+               }
+
+
+               idx++;
+               while (pktnwc < pktnw && idx < niov) {
+                       const size_t slen = iov[idx].iov_len;
+                       const unsigned long faddr =
+                               (unsigned long) iov[idx].iov_base;
+
+                       if (slen & 3 || faddr & 3 || !slen ||
+                           slen > PAGE_SIZE) {
+                               ret = -EINVAL;
+                               goto free_pbc;
+                       }
+
+                       npages++;
+                       if ((faddr & PAGE_MASK) !=
+                           ((faddr + slen - 1) & PAGE_MASK))
+                               npages++;
+
+                       pktnwc += slen >> 2;
+                       idx++;
+                       nfrags++;
+               }
+
+               if (pktnwc != pktnw) {
+                       ret = -EINVAL;
+                       goto free_pbc;
+               }
+
+               if (page) {
+                       dma_addr = dma_map_page(&dd->pcidev->dev,
+                                               page, 0, len, DMA_TO_DEVICE);
+                       if (dma_mapping_error(dma_addr)) {
+                               ret = -ENOMEM;
+                               goto free_pbc;
+                       }
+
+                       dma_mapped = 1;
+               }
+
+               ipath_user_sdma_init_header(pkt, counter, 0, len, dma_mapped,
+                                           page, pbc, dma_addr);
+
+               if (nfrags) {
+                       ret = ipath_user_sdma_init_payload(dd, pq, pkt,
+                                                          iov + idx_save + 1,
+                                                          nfrags, npages);
+                       if (ret < 0)
+                               goto free_pbc_dma;
+               }
+
+               counter++;
+               npkts++;
+
+               list_add_tail(&pkt->list, list);
+       }
+
+       ret = idx;
+       goto done;
+
+free_pbc_dma:
+       if (dma_mapped)
+               dma_unmap_page(&dd->pcidev->dev, dma_addr, len, DMA_TO_DEVICE);
+free_pbc:
+       if (page) {
+               kunmap(page);
+               __free_page(page);
+       } else
+               dma_pool_free(pq->header_cache, pbc, dma_addr);
+free_pkt:
+       kmem_cache_free(pq->pkt_slab, pkt);
+free_list:
+       ipath_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, list);
+done:
+       return ret;
+}
+
+static void ipath_user_sdma_set_complete_counter(struct ipath_user_sdma_queue *pq,
+                                                u32 c)
+{
+       pq->sent_counter = c;
+}
+
+/* try to clean out queue -- needs pq->lock */
+static int ipath_user_sdma_queue_clean(const struct ipath_devdata *dd,
+                                      struct ipath_user_sdma_queue *pq)
+{
+       struct list_head free_list;
+       struct ipath_user_sdma_pkt *pkt;
+       struct ipath_user_sdma_pkt *pkt_prev;
+       int ret = 0;
+
+       INIT_LIST_HEAD(&free_list);
+
+       list_for_each_entry_safe(pkt, pkt_prev, &pq->sent, list) {
+               s64 descd = dd->ipath_sdma_descq_removed - pkt->added;
+
+               if (descd < 0)
+                       break;
+
+               list_move_tail(&pkt->list, &free_list);
+
+               /* one more packet cleaned */
+               ret++;
+       }
+
+       if (!list_empty(&free_list)) {
+               u32 counter;
+
+               pkt = list_entry(free_list.prev,
+                                struct ipath_user_sdma_pkt, list);
+               counter = pkt->counter;
+
+               ipath_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &free_list);
+               ipath_user_sdma_set_complete_counter(pq, counter);
+       }
+
+       return ret;
+}
+
+void ipath_user_sdma_queue_destroy(struct ipath_user_sdma_queue *pq)
+{
+       if (!pq)
+               return;
+
+       kmem_cache_destroy(pq->pkt_slab);
+       dma_pool_destroy(pq->header_cache);
+       kfree(pq);
+}
+
+/* clean descriptor queue, returns > 0 if some elements cleaned */
+static int ipath_user_sdma_hwqueue_clean(struct ipath_devdata *dd)
+{
+       int ret;
+       unsigned long flags;
+
+       spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
+       ret = ipath_sdma_make_progress(dd);
+       spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
+
+       return ret;
+}
+
+/* we're in close, drain packets so that we can cleanup successfully... */
+void ipath_user_sdma_queue_drain(struct ipath_devdata *dd,
+                                struct ipath_user_sdma_queue *pq)
+{
+       int i;
+
+       if (!pq)
+               return;
+
+       for (i = 0; i < 100; i++) {
+               mutex_lock(&pq->lock);
+               if (list_empty(&pq->sent)) {
+                       mutex_unlock(&pq->lock);
+                       break;
+               }
+               ipath_user_sdma_hwqueue_clean(dd);
+               ipath_user_sdma_queue_clean(dd, pq);
+               mutex_unlock(&pq->lock);
+               msleep(10);
+       }
+
+       if (!list_empty(&pq->sent)) {
+               struct list_head free_list;
+
+               printk(KERN_INFO "drain: lists not empty: forcing!\n");
+               INIT_LIST_HEAD(&free_list);
+               mutex_lock(&pq->lock);
+               list_splice_init(&pq->sent, &free_list);
+               ipath_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &free_list);
+               mutex_unlock(&pq->lock);
+       }
+}
+
+static inline __le64 ipath_sdma_make_desc0(struct ipath_devdata *dd,
+                                          u64 addr, u64 dwlen, u64 dwoffset)
+{
+       return cpu_to_le64(/* SDmaPhyAddr[31:0] */
+                          ((addr & 0xfffffffcULL) << 32) |
+                          /* SDmaGeneration[1:0] */
+                          ((dd->ipath_sdma_generation & 3ULL) << 30) |
+                          /* SDmaDwordCount[10:0] */
+                          ((dwlen & 0x7ffULL) << 16) |
+                          /* SDmaBufOffset[12:2] */
+                          (dwoffset & 0x7ffULL));
+}
+
+static inline __le64 ipath_sdma_make_first_desc0(__le64 descq)
+{
+       return descq | __constant_cpu_to_le64(1ULL << 12);
+}
+
+static inline __le64 ipath_sdma_make_last_desc0(__le64 descq)
+{
+                                             /* last */  /* dma head */
+       return descq | __constant_cpu_to_le64(1ULL << 11 | 1ULL << 13);
+}
+
+static inline __le64 ipath_sdma_make_desc1(u64 addr)
+{
+       /* SDmaPhyAddr[47:32] */
+       return cpu_to_le64(addr >> 32);
+}
+
+static void ipath_user_sdma_send_frag(struct ipath_devdata *dd,
+                                     struct ipath_user_sdma_pkt *pkt, int idx,
+                                     unsigned ofs, u16 tail)
+{
+       const u64 addr = (u64) pkt->addr[idx].addr +
+               (u64) pkt->addr[idx].offset;
+       const u64 dwlen = (u64) pkt->addr[idx].length / 4;
+       __le64 *descqp;
+       __le64 descq0;
+
+       descqp = &dd->ipath_sdma_descq[tail].qw[0];
+
+       descq0 = ipath_sdma_make_desc0(dd, addr, dwlen, ofs);
+       if (idx == 0)
+               descq0 = ipath_sdma_make_first_desc0(descq0);
+       if (idx == pkt->naddr - 1)
+               descq0 = ipath_sdma_make_last_desc0(descq0);
+
+       descqp[0] = descq0;
+       descqp[1] = ipath_sdma_make_desc1(addr);
+}
+
+/* pq->lock must be held, get packets on the wire... */
+static int ipath_user_sdma_push_pkts(struct ipath_devdata *dd,
+                                    struct ipath_user_sdma_queue *pq,
+                                    struct list_head *pktlist)
+{
+       int ret = 0;
+       unsigned long flags;
+       u16 tail;
+
+       if (list_empty(pktlist))
+               return 0;
+
+       if (unlikely(!(dd->ipath_flags & IPATH_LINKACTIVE)))
+               return -ECOMM;
+
+       spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
+
+       if (unlikely(dd->ipath_sdma_status & IPATH_SDMA_ABORT_MASK)) {
+               ret = -ECOMM;
+               goto unlock;
+       }
+
+       tail = dd->ipath_sdma_descq_tail;
+       while (!list_empty(pktlist)) {
+               struct ipath_user_sdma_pkt *pkt =
+                       list_entry(pktlist->next, struct ipath_user_sdma_pkt,
+                                  list);
+               int i;
+               unsigned ofs = 0;
+               u16 dtail = tail;
+
+               if (pkt->naddr > ipath_sdma_descq_freecnt(dd))
+                       goto unlock_check_tail;
+
+               for (i = 0; i < pkt->naddr; i++) {
+                       ipath_user_sdma_send_frag(dd, pkt, i, ofs, tail);
+                       ofs += pkt->addr[i].length >> 2;
+
+                       if (++tail == dd->ipath_sdma_descq_cnt) {
+                               tail = 0;
+                               ++dd->ipath_sdma_generation;
+                       }
+               }
+
+               if ((ofs<<2) > dd->ipath_ibmaxlen) {
+                       ipath_dbg("packet size %X > ibmax %X, fail\n",
+                               ofs<<2, dd->ipath_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 >= IPATH_SMALLBUF_DWORDS) {
+                       for (i = 0; i < pkt->naddr; i++) {
+                               dd->ipath_sdma_descq[dtail].qw[0] |=
+                                       __constant_cpu_to_le64(1ULL << 14);
+                               if (++dtail == dd->ipath_sdma_descq_cnt)
+                                       dtail = 0;
+                       }
+               }
+
+               dd->ipath_sdma_descq_added += pkt->naddr;
+               pkt->added = dd->ipath_sdma_descq_added;
+               list_move_tail(&pkt->list, &pq->sent);
+               ret++;
+       }
+
+unlock_check_tail:
+       /* advance the tail on the chip if necessary */
+       if (dd->ipath_sdma_descq_tail != tail) {
+               wmb();
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail, tail);
+               dd->ipath_sdma_descq_tail = tail;
+       }
+
+unlock:
+       spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
+
+       return ret;
+}
+
+int ipath_user_sdma_writev(struct ipath_devdata *dd,
+                          struct ipath_user_sdma_queue *pq,
+                          const struct iovec *iov,
+                          unsigned long dim)
+{
+       int ret = 0;
+       struct list_head list;
+       int npkts = 0;
+
+       INIT_LIST_HEAD(&list);
+
+       mutex_lock(&pq->lock);
+
+       if (dd->ipath_sdma_descq_added != dd->ipath_sdma_descq_removed) {
+               ipath_user_sdma_hwqueue_clean(dd);
+               ipath_user_sdma_queue_clean(dd, pq);
+       }
+
+       while (dim) {
+               const int mxp = 8;
+
+               down_write(&current->mm->mmap_sem);
+               ret = ipath_user_sdma_queue_pkts(dd, pq, &list, iov, dim, mxp);
+               up_write(&current->mm->mmap_sem);
+
+               if (ret <= 0)
+                       goto done_unlock;
+               else {
+                       dim -= ret;
+                       iov += ret;
+               }
+
+               /* 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).
+                        */
+                       if (ipath_sdma_descq_freecnt(dd) < ret * 4) {
+                               ipath_user_sdma_hwqueue_clean(dd);
+                               ipath_user_sdma_queue_clean(dd, pq);
+                       }
+
+                       ret = ipath_user_sdma_push_pkts(dd, pq, &list);
+                       if (ret < 0)
+                               goto done_unlock;
+                       else {
+                               npkts += ret;
+                               pq->counter += ret;
+
+                               if (!list_empty(&list))
+                                       goto done_unlock;
+                       }
+               }
+       }
+
+done_unlock:
+       if (!list_empty(&list))
+               ipath_user_sdma_free_pkt_list(&dd->pcidev->dev, pq, &list);
+       mutex_unlock(&pq->lock);
+
+       return (ret < 0) ? ret : npkts;
+}
+
+int ipath_user_sdma_make_progress(struct ipath_devdata *dd,
+                                 struct ipath_user_sdma_queue *pq)
+{
+       int ret = 0;
+
+       mutex_lock(&pq->lock);
+       ipath_user_sdma_hwqueue_clean(dd);
+       ret = ipath_user_sdma_queue_clean(dd, pq);
+       mutex_unlock(&pq->lock);
+
+       return ret;
+}
+
+u32 ipath_user_sdma_complete_counter(const struct ipath_user_sdma_queue *pq)
+{
+       return pq->sent_counter;
+}
+
+u32 ipath_user_sdma_inflight_counter(struct ipath_user_sdma_queue *pq)
+{
+       return pq->counter;
+}
+